A   OÀ¯     H‚B     G³SkPTeSep  4 202621:40:44HOST64sm_89







.version 9.0
.target sm_89
.address_size 64

.global .align 1 .b8 kmask_iq2xs[8] = {1, 2, 4, 8, 16, 32, 64, 128};
.global .align 1 .b8 ksigns_iq2xs[128] = {0, 129, 130, 3, 132, 5, 6, 135, 136, 9, 10, 139, 12, 141, 142, 15, 144, 17, 18, 147, 20, 149, 150, 23, 24, 153, 154, 27, 156, 29, 30, 159, 160, 33, 34, 163, 36, 165, 166, 39, 40, 169, 170, 43, 172, 45, 46, 175, 48, 177, 178, 51, 180, 53, 54, 183, 184, 57, 58, 187, 60, 189, 190, 63, 192, 65, 66, 195, 68, 197, 198, 71, 72, 201, 202, 75, 204, 77, 78, 207, 80, 209, 210, 83, 212, 85, 86, 215, 216, 89, 90, 219, 92, 221, 222, 95, 96, 225, 226, 99, 228, 101, 102, 231, 232, 105, 106, 235, 108, 237, 238, 111, 240, 113, 114, 243, 116, 245, 246, 119, 120, 249, 250, 123, 252, 125, 126, 255};
.global .align 8 .b8 ksigns64[1024] = {0, 0, 0, 0, 0, 0, 0, 0, 255, 0, 0, 0, 0, 0, 0, 255, 0, 255, 0, 0, 0, 0, 0, 255, 255, 255, 0, 0, 0, 0, 0, 0, 0, 0, 255, 0, 0, 0, 0, 255, 255, 0, 255, 0, 0, 0, 0, 0, 0, 255, 255, 0, 0, 0, 0, 0, 255, 255, 255, 0, 0, 0, 0, 255, 0, 0, 0, 255, 0, 0, 0, 255, 255, 0, 0, 255, 0, 0, 0, 0, 0, 255, 0, 255, 0, 0, 0, 0, 255, 255, 0, 255, 0, 0, 0, 255, 0, 0, 255, 255, 0, 0, 0, 0, 255, 0, 255, 255, 0, 0, 0, 255, 0, 255, 255, 255, 0, 0, 0, 255, 255, 255, 255, 255, 0, 0, 0, 0, 0, 0, 0, 0, 255, 0, 0, 255, 255, 0, 0, 0, 255, 0, 0, 0, 0, 255, 0, 0, 255, 0, 0, 0, 255, 255, 0, 0, 255, 0, 0, 255, 0, 0, 255, 0, 255, 0, 0, 0, 255, 0, 255, 0, 255, 0, 0, 255, 0, 255, 255, 0, 255, 0, 0, 255, 255, 255, 255, 0, 255, 0, 0, 0, 0, 0, 0, 255, 255, 0, 0, 0, 255, 0, 0, 255, 255, 0, 0, 255, 0, 255, 0, 255, 255, 0, 0, 255, 255, 255, 0, 255, 255, 0, 0, 0, 0, 0, 255, 255, 255, 0, 0, 255, 255, 0, 255, 255, 255, 0, 0, 0, 0, 255, 255, 255, 255, 0, 0, 0, 255, 255, 255, 255, 255, 0, 0, 255, 0, 0, 0, 0, 0, 255, 0, 255, 255, 0, 0, 0, 0, 255, 0, 0, 0, 255, 0, 0, 0, 255, 0, 0, 255, 255, 0, 0, 0, 255, 0, 255, 0, 0, 255, 0, 0, 255, 0, 0, 255, 0, 255, 0, 0, 255, 0, 255, 0, 255, 255, 0, 0, 255, 0, 255, 255, 255, 255, 0, 0, 255, 0, 0, 0, 0, 0, 255, 0, 255, 0, 0, 255, 0, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 255, 255, 0, 255, 0, 255, 0, 0, 0, 0, 255, 255, 0, 255, 0, 255, 255, 0, 255, 255, 0, 255, 0, 0, 0, 255, 255, 255, 0, 255, 0, 0, 255, 255, 255, 255, 0, 255, 0, 255, 0, 0, 0, 0, 255, 255, 0, 0, 255, 0, 0, 0, 255, 255, 0, 255, 0, 255, 0, 0, 255, 255, 0, 255, 255, 255, 0, 0, 255, 255, 0, 0, 0, 0, 255, 0, 255, 255, 0, 255, 255, 0, 255, 0, 255, 255, 0, 0, 0, 255, 255, 0, 255, 255, 0, 0, 255, 255, 255, 0, 255, 255, 0, 255, 0, 0, 0, 255, 255, 255, 0, 255, 255, 0, 0, 255, 255, 255, 0, 0, 0, 255, 0, 255, 255, 255, 0, 0, 255, 255, 0, 255, 255, 255, 0, 255, 0, 0, 255, 255, 255, 255, 0, 0, 255, 0, 255, 255, 255, 255, 0, 255, 0, 255, 255, 255, 255, 255, 0, 255, 255, 255, 255, 255, 255, 255, 0, 0, 0, 0, 0, 0, 0, 0, 255, 255, 255, 0, 0, 0, 0, 0, 255, 0, 0, 255, 0, 0, 0, 0, 255, 0, 255, 255, 0, 0, 0, 0, 255, 255, 0, 0, 255, 0, 0, 0, 255, 0, 255, 0, 255, 0, 0, 0, 255, 255, 0, 255, 255, 0, 0, 0, 255, 255, 255, 255, 255, 0, 0, 0, 255, 0, 0, 0, 0, 255, 0, 0, 255, 0, 255, 0, 0, 255, 0, 0, 255, 255, 0, 255, 0, 255, 0, 0, 255, 255, 255, 255, 0, 255, 0, 0, 255, 0, 0, 0, 255, 255, 0, 0, 255, 255, 255, 0, 255, 255, 0, 0, 255, 0, 0, 255, 255, 255, 0, 0, 255, 0, 255, 255, 255, 255, 0, 0, 255, 255, 0, 0, 0, 0, 255, 0, 255, 0, 255, 0, 0, 0, 255, 0, 255, 255, 0, 255, 0, 0, 255, 0, 255, 255, 255, 255, 0, 0, 255, 0, 255, 0, 0, 0, 255, 0, 255, 0, 255, 255, 255, 0, 255, 0, 255, 0, 255, 0, 0, 255, 255, 0, 255, 0, 255, 0, 255, 255, 255, 0, 255, 0, 255, 255, 0, 0, 0, 255, 255, 0, 255, 255, 255, 0, 0, 255, 255, 0, 255, 0, 0, 255, 0, 255, 255, 0, 255, 0, 255, 255, 0, 255, 255, 0, 255, 255, 0, 0, 255, 255, 255, 0, 255, 0, 255, 0, 255, 255, 255, 0, 255, 255, 0, 255, 255, 255, 255, 0, 255, 255, 255, 255, 255, 255, 255, 0, 255, 0, 0, 0, 0, 0, 0, 255, 255, 0, 255, 0, 0, 0, 0, 255, 255, 255, 0, 255, 0, 0, 0, 255, 255, 255, 255, 255, 0, 0, 0, 255, 255, 0, 0, 0, 255, 0, 0, 255, 255, 255, 255, 0, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 255, 255, 255, 0, 0, 255, 255, 255, 0, 0, 0, 255, 0, 255, 255, 255, 255, 0, 0, 255, 0, 255, 255, 0, 0, 255, 0, 255, 0, 255, 255, 0, 255, 255, 0, 255, 0, 255, 255, 255, 0, 0, 255, 255, 0, 255, 255, 0, 255, 0, 255, 255, 0, 255, 255, 255, 0, 255, 255, 255, 0, 255, 255, 255, 255, 255, 255, 255, 0, 255, 255, 0, 0, 0, 0, 0, 255, 255, 255, 255, 255, 0, 0, 0, 255, 255, 255, 0, 0, 255, 0, 0, 255, 255, 255, 0, 255, 255, 0, 0, 255, 255, 255, 255, 0, 0, 255, 0, 255, 255, 255, 0, 255, 0, 255, 0, 255, 255, 255, 255, 0, 255, 255, 0, 255, 255, 255, 255, 255, 255, 255, 0, 255, 255, 255, 0, 0, 0, 0, 255, 255, 255, 255, 0, 255, 0, 0, 255, 255, 255, 255, 255, 0, 255, 0, 255, 255, 255, 255, 255, 255, 255, 0, 255, 255, 255, 255, 0, 0, 0, 255, 255, 255, 255, 255, 255, 255, 0, 255, 255, 255, 255, 255, 0, 0, 255, 255, 255, 255, 255, 255, 0, 255, 255, 255, 255, 255, 255, 255, 255};
.global .align 8 .b8 iq2xxs_grid[2048] = {8, 8, 8, 8, 8, 8, 8, 8, 43, 8, 8, 8, 8, 8, 8, 8, 25, 25, 8, 8, 8, 8, 8, 8, 8, 43, 8, 8, 8, 8, 8, 8, 43, 43, 8, 8, 8, 8, 8, 8, 25, 8, 25, 8, 8, 8, 8, 8, 8, 25, 25, 8, 8, 8, 8, 8, 8, 8, 43, 8, 8, 8, 8, 8, 43, 8, 43, 8, 8, 8, 8, 8, 8, 43, 43, 8, 8, 8, 8, 8, 43, 43, 43, 8, 8, 8, 8, 8, 25, 8, 8, 25, 8, 8, 8, 8, 8, 25, 8, 25, 8, 8, 8, 8, 8, 8, 25, 25, 8, 8, 8, 8, 8, 43, 25, 25, 8, 8, 8, 8, 25, 8, 43, 25, 8, 8, 8, 8, 8, 25, 43, 25, 8, 8, 8, 8, 8, 8, 8, 43, 8, 8, 8, 8, 43, 8, 8, 43, 8, 8, 8, 8, 43, 43, 8, 43, 8, 8, 8, 8, 43, 8, 43, 43, 8, 8, 8, 8, 25, 8, 8, 8, 25, 8, 8, 8, 8, 25, 8, 8, 25, 8, 8, 8, 8, 8, 25, 8, 25, 8, 8, 8, 25, 25, 25, 8, 25, 8, 8, 8, 8, 8, 8, 25, 25, 8, 8, 8, 8, 25, 8, 43, 25, 8, 8, 8, 8, 43, 25, 43, 25, 8, 8, 8, 8, 8, 8, 8, 43, 8, 8, 8, 43, 8, 8, 8, 43, 8, 8, 8, 43, 8, 43, 8, 43, 8, 8, 8, 43, 8, 8, 43, 43, 8, 8, 8, 25, 8, 8, 8, 8, 25, 8, 8, 8, 25, 8, 8, 8, 25, 8, 8, 8, 8, 25, 8, 8, 25, 8, 8, 25, 8, 43, 8, 8, 25, 8, 8, 8, 25, 43, 8, 8, 25, 8, 8, 8, 8, 8, 25, 8, 25, 8, 8, 43, 8, 8, 25, 8, 25, 8, 8, 8, 43, 8, 25, 8, 25, 8, 8, 8, 8, 43, 25, 8, 25, 8, 8, 25, 8, 8, 43, 8, 25, 8, 8, 8, 25, 8, 43, 8, 25, 8, 8, 8, 8, 25, 43, 8, 25, 8, 8, 8, 25, 43, 43, 8, 25, 8, 8, 8, 8, 8, 8, 25, 25, 8, 8, 43, 8, 8, 8, 25, 25, 8, 8, 8, 43, 8, 8, 25, 25, 8, 8, 8, 8, 43, 8, 25, 25, 8, 8, 43, 25, 8, 25, 25, 25, 8, 8, 25, 43, 43, 25, 25, 25, 8, 8, 8, 8, 8, 43, 25, 25, 8, 8, 25, 8, 25, 43, 25, 25, 8, 8, 25, 43, 8, 8, 43, 25, 8, 8, 8, 8, 25, 8, 43, 25, 8, 8, 8, 8, 8, 25, 43, 25, 8, 8, 8, 25, 8, 43, 43, 25, 8, 8, 8, 25, 43, 43, 43, 25, 8, 8, 8, 8, 8, 8, 8, 43, 8, 8, 25, 25, 8, 8, 8, 43, 8, 8, 8, 43, 8, 8, 8, 43, 8, 8, 8, 25, 25, 8, 8, 43, 8, 8, 8, 43, 43, 8, 8, 43, 8, 8, 25, 8, 8, 25, 8, 43, 8, 8, 8, 25, 8, 25, 8, 43, 8, 8, 8, 8, 25, 25, 8, 43, 8, 8, 43, 8, 25, 25, 8, 43, 8, 8, 8, 43, 8, 43, 8, 43, 8, 8, 8, 25, 8, 8, 25, 43, 8, 8, 8, 8, 8, 25, 25, 43, 8, 8, 43, 8, 8, 8, 43, 43, 8, 8, 8, 25, 25, 8, 43, 43, 8, 8, 25, 8, 8, 8, 8, 8, 25, 8, 8, 25, 8, 8, 8, 8, 25, 8, 8, 8, 25, 8, 8, 8, 25, 8, 25, 8, 43, 8, 8, 8, 25, 8, 8, 8, 8, 25, 8, 8, 25, 8, 8, 8, 43, 25, 8, 8, 25, 8, 8, 25, 8, 43, 8, 8, 25, 8, 8, 8, 25, 43, 8, 8, 25, 8, 25, 25, 25, 43, 8, 8, 25, 8, 8, 8, 8, 8, 25, 8, 25, 8, 8, 43, 8, 8, 25, 8, 25, 8, 8, 8, 43, 8, 25, 8, 25, 8, 8, 8, 25, 25, 25, 8, 25, 8, 43, 43, 25, 25, 25, 8, 25, 8, 8, 8, 8, 43, 25, 8, 25, 8, 8, 25, 43, 8, 43, 8, 25, 8, 25, 25, 8, 25, 43, 8, 25, 8, 8, 8, 8, 8, 8, 25, 25, 8, 8, 43, 8, 8, 8, 25, 25, 8, 8, 8, 43, 8, 8, 25, 25, 8, 25, 25, 43, 8, 8, 25, 25, 8, 25, 43, 8, 25, 8, 25, 25, 8, 8, 8, 8, 43, 8, 25, 25, 8, 8, 43, 25, 8, 25, 25, 25, 8, 43, 8, 43, 25, 25, 25, 25, 8, 8, 8, 8, 8, 43, 25, 25, 8, 43, 25, 25, 8, 43, 25, 25, 8, 25, 8, 8, 8, 8, 43, 25, 8, 8, 25, 8, 8, 8, 43, 25, 8, 8, 8, 25, 8, 8, 43, 25, 8, 8, 8, 8, 25, 8, 43, 25, 8, 25, 8, 8, 43, 8, 43, 25, 8, 8, 8, 8, 8, 25, 43, 25, 8, 25, 25, 8, 8, 25, 43, 25, 8, 8, 8, 43, 43, 25, 43, 25, 8, 25, 8, 25, 25, 43, 43, 25, 8, 8, 8, 8, 8, 8, 8, 43, 8, 43, 8, 8, 8, 8, 8, 43, 8, 43, 43, 8, 8, 8, 8, 43, 8, 8, 25, 8, 25, 8, 8, 43, 8, 25, 8, 43, 25, 8, 8, 43, 8, 8, 8, 8, 43, 8, 8, 43, 8, 43, 8, 8, 43, 8, 8, 43, 8, 25, 43, 43, 8, 25, 8, 43, 8, 8, 43, 8, 25, 25, 8, 43, 8, 8, 8, 8, 8, 43, 8, 43, 8, 43, 8, 8, 8, 43, 8, 43, 8, 25, 8, 8, 8, 8, 25, 43, 8, 8, 25, 8, 8, 8, 25, 43, 8, 8, 8, 25, 8, 8, 25, 43, 8, 8, 8, 8, 25, 8, 25, 43, 8, 43, 25, 25, 25, 8, 25, 43, 8, 8, 8, 8, 8, 25, 25, 43, 8, 25, 8, 8, 25, 25, 25, 43, 8, 8, 25, 43, 25, 25, 25, 43, 8, 8, 8, 25, 43, 43, 25, 43, 8, 8, 43, 8, 8, 8, 43, 43, 8, 8, 8, 43, 8, 8, 43, 43, 8, 8, 25, 25, 43, 8, 43, 43, 8, 8, 25, 8, 25, 43, 43, 43, 8, 25, 8, 8, 8, 8, 8, 8, 25, 8, 25, 8, 8, 8, 8, 8, 25, 8, 8, 25, 8, 8, 8, 8, 25, 8, 43, 25, 8, 8, 8, 8, 25, 25, 8, 43, 8, 8, 8, 8, 25, 8, 25, 43, 8, 8, 8, 8, 25, 8, 8, 8, 25, 8, 8, 8, 25, 8, 43, 8, 25, 8, 8, 8, 25, 43, 25, 25, 25, 8, 8, 8, 25, 8, 8, 43, 25, 8, 8, 8, 25, 25, 8, 8, 43, 8, 8, 8, 25, 8, 25, 8, 43, 8, 8, 8, 25, 8, 8, 25, 43, 8, 8, 8, 25, 8, 8, 8, 8, 25, 8, 8, 25, 8, 8, 43, 8, 25, 8, 8, 25, 25, 8, 43, 25, 25, 8, 8, 25, 8, 8, 8, 43, 25, 8, 8, 25, 25, 25, 8, 43, 25, 8, 8, 25, 25, 8, 8, 8, 43, 8, 8, 25, 8, 8, 25, 8, 43, 8, 8, 25, 8, 43, 8, 25, 43, 8, 8, 25, 43, 25, 25, 25, 43, 8, 8, 25, 8, 43, 43, 25, 43, 8, 8, 25, 8, 8, 8, 8, 8, 25, 8, 25, 8, 43, 8, 8, 8, 25, 8, 25, 8, 8, 43, 8, 8, 25, 8, 25, 8, 8, 8, 43, 8, 25, 8, 25, 25, 43, 25, 43, 8, 25, 8, 25, 43, 8, 25, 8, 25, 25, 8, 25, 8, 25, 43, 8, 25, 25, 8, 25, 8, 8, 8, 8, 43, 25, 8, 25, 25, 8, 8, 8, 8, 43, 8, 25, 8, 25, 8, 8, 8, 43, 8, 25, 8, 8, 25, 8, 8, 43, 8, 25, 8, 8, 8, 25, 8, 43, 8, 25, 25, 25, 8, 25, 8, 43, 8, 25, 8, 8, 8, 8, 25, 43, 8, 25, 8, 43, 25, 25, 25, 43, 8, 25, 25, 8, 43, 25, 25, 43, 8, 25, 43, 8, 8, 43, 25, 43, 8, 25, 25, 25, 8, 25, 43, 43, 8, 25, 8, 8, 25, 43, 43, 43, 8, 25, 8, 8, 8, 8, 8, 8, 25, 25, 8, 43, 8, 8, 8, 8, 25, 25, 25, 8, 25, 8, 8, 8, 25, 25, 25, 43, 25, 8, 8, 8, 25, 25, 8, 8, 43, 8, 8, 8, 25, 25, 8, 8, 8, 43, 8, 8, 25, 25, 8, 43, 8, 43, 8, 8, 25, 25, 8, 25, 8, 8, 25, 8, 25, 25, 43, 8, 8, 25, 25, 8, 25, 25, 8, 25, 43, 43, 25, 8, 25, 25, 25, 8, 25, 43, 43, 8, 25, 25, 8, 8, 25, 43, 8, 25, 25, 25, 43, 8, 25, 43, 8, 25, 25, 25, 43, 43, 8, 8, 25, 25, 25, 25, 25, 8, 8, 8, 43, 25, 25, 25, 8, 25, 25, 25, 43, 25, 25, 25, 8, 8, 8, 8, 8, 43, 25, 25, 25, 8, 25, 8, 8, 43, 25, 25, 25, 43, 25, 8, 8, 43, 25, 25, 8, 25, 43, 25, 8, 43, 25, 25, 8, 8, 8, 25, 25, 43, 25, 25, 8, 43, 8, 8, 43, 43, 25, 25, 8, 25, 8, 8, 8, 8, 43, 25, 8, 8, 25, 8, 8, 8, 43, 25, 8, 8, 8, 25, 8, 8, 43, 25, 8, 43, 43, 25, 8, 8, 43, 25, 8, 8, 8, 8, 25, 8, 43, 25, 25, 25, 25, 25, 25, 8, 43, 25, 8, 43, 25, 8, 43, 8, 43, 25, 8, 8, 43, 25, 43, 8, 43, 25, 8, 8, 8, 8, 8, 25, 43, 25, 25, 25, 8, 8, 8, 25, 43, 25, 8, 8, 25, 8, 25, 25, 43, 25, 43, 8, 25, 8, 25, 25, 43, 25, 8, 25, 8, 43, 25, 25, 43, 25, 43, 8, 8, 25, 8, 43, 43, 25, 8, 8, 8, 8, 8, 8, 8, 43, 43, 8, 8, 8, 8, 8, 8, 43, 43, 43, 8, 8, 8, 8, 8, 43, 25, 8, 8, 25, 8, 8, 8, 43, 43, 8, 8, 43, 8, 8, 8, 43, 8, 25, 8, 8, 25, 8, 8, 43, 8, 43, 25, 8, 25, 8, 8, 43, 8, 8, 8, 25, 25, 8, 8, 43, 25, 8, 25, 8, 43, 8, 8, 43, 25, 8, 8, 8, 8, 25, 8, 43, 8, 25, 8, 8, 8, 25, 8, 43, 8, 8, 25, 8, 8, 25, 8, 43, 25, 25, 25, 8, 8, 25, 8, 43, 8, 8, 8, 25, 8, 25, 8, 43, 8, 8, 43, 25, 8, 25, 8, 43, 8, 8, 8, 8, 25, 25, 8, 43, 43, 25, 8, 25, 25, 25, 8, 43, 8, 25, 25, 43, 25, 25, 8, 43, 25, 43, 8, 8, 43, 25, 8, 43, 8, 8, 8, 25, 43, 25, 8, 43, 8, 8, 43, 25, 43, 25, 8, 43, 43, 8, 8, 8, 8, 43, 8, 43, 8, 25, 8, 8, 25, 43, 8, 43, 25, 8, 25, 8, 43, 43, 8, 43, 8, 25, 8, 8, 8, 8, 25, 43, 8, 8, 25, 8, 8, 8, 25, 43, 8, 25, 43, 8, 8, 8, 25, 43, 8, 8, 8, 25, 8, 8, 25, 43, 25, 8, 43, 43, 8, 8, 25, 43, 43, 25, 25, 8, 25, 8, 25, 43, 8, 8, 8, 43, 25, 8, 25, 43, 25, 25, 8, 25, 43, 8, 25, 43, 8, 8, 8, 8, 8, 25, 25, 43, 43, 8, 43, 8, 8, 25, 25, 43, 8, 25, 8, 25, 8, 25, 25, 43, 25, 8, 25, 25, 25, 25, 25, 43, 25, 8, 8, 43, 8, 43, 25, 43, 8, 8, 43, 8, 25, 43, 25, 43, 43, 8, 8, 8, 8, 8, 43, 43, 8, 8, 25, 25, 8, 8, 43, 43, 25, 25, 8, 43, 8, 8, 43, 43, 25, 43, 8, 8, 25, 8, 43, 43, 8, 8, 8, 8, 43, 8, 43, 43, 8, 43, 25, 8, 8, 25, 43, 43, 8, 8, 25, 25, 8, 43, 43, 43, 8, 25, 8, 8, 25, 43, 43, 43};
.global .align 8 .b8 iq2xs_grid[4096] = {8, 8, 8, 8, 8, 8, 8, 8, 43, 8, 8, 8, 8, 8, 8, 8, 25, 25, 8, 8, 8, 8, 8, 8, 8, 43, 8, 8, 8, 8, 8, 8, 43, 43, 8, 8, 8, 8, 8, 8, 25, 8, 25, 8, 8, 8, 8, 8, 8, 25, 25, 8, 8, 8, 8, 8, 43, 25, 25, 8, 8, 8, 8, 8, 25, 43, 25, 8, 8, 8, 8, 8, 8, 8, 43, 8, 8, 8, 8, 8, 43, 8, 43, 8, 8, 8, 8, 8, 25, 25, 43, 8, 8, 8, 8, 8, 8, 43, 43, 8, 8, 8, 8, 8, 25, 8, 8, 25, 8, 8, 8, 8, 8, 25, 8, 25, 8, 8, 8, 8, 43, 25, 8, 25, 8, 8, 8, 8, 25, 43, 8, 25, 8, 8, 8, 8, 8, 8, 25, 25, 8, 8, 8, 8, 43, 8, 25, 25, 8, 8, 8, 8, 25, 25, 25, 25, 8, 8, 8, 8, 8, 43, 25, 25, 8, 8, 8, 8, 25, 8, 43, 25, 8, 8, 8, 8, 8, 25, 43, 25, 8, 8, 8, 8, 8, 8, 8, 43, 8, 8, 8, 8, 43, 8, 8, 43, 8, 8, 8, 8, 25, 25, 8, 43, 8, 8, 8, 8, 8, 43, 8, 43, 8, 8, 8, 8, 25, 8, 25, 43, 8, 8, 8, 8, 8, 25, 25, 43, 8, 8, 8, 8, 25, 43, 25, 43, 8, 8, 8, 8, 8, 8, 43, 43, 8, 8, 8, 8, 25, 8, 8, 8, 25, 8, 8, 8, 8, 25, 8, 8, 25, 8, 8, 8, 43, 25, 8, 8, 25, 8, 8, 8, 25, 43, 8, 8, 25, 8, 8, 8, 8, 8, 25, 8, 25, 8, 8, 8, 43, 8, 25, 8, 25, 8, 8, 8, 25, 25, 25, 8, 25, 8, 8, 8, 8, 43, 25, 8, 25, 8, 8, 8, 43, 43, 25, 8, 25, 8, 8, 8, 25, 8, 43, 8, 25, 8, 8, 8, 8, 25, 43, 8, 25, 8, 8, 8, 8, 8, 8, 25, 25, 8, 8, 8, 43, 8, 8, 25, 25, 8, 8, 8, 25, 25, 8, 25, 25, 8, 8, 8, 8, 43, 8, 25, 25, 8, 8, 8, 25, 8, 25, 25, 25, 8, 8, 8, 8, 25, 25, 25, 25, 8, 8, 8, 8, 8, 43, 25, 25, 8, 8, 8, 8, 43, 43, 25, 25, 8, 8, 8, 25, 8, 8, 43, 25, 8, 8, 8, 8, 25, 8, 43, 25, 8, 8, 8, 8, 8, 25, 43, 25, 8, 8, 8, 8, 8, 8, 8, 43, 8, 8, 8, 43, 8, 8, 8, 43, 8, 8, 8, 25, 25, 8, 8, 43, 8, 8, 8, 8, 43, 8, 8, 43, 8, 8, 8, 25, 8, 25, 8, 43, 8, 8, 8, 8, 25, 25, 8, 43, 8, 8, 8, 8, 8, 43, 8, 43, 8, 8, 8, 25, 8, 8, 25, 43, 8, 8, 8, 8, 25, 8, 25, 43, 8, 8, 8, 8, 8, 25, 25, 43, 8, 8, 8, 25, 25, 25, 25, 43, 8, 8, 8, 8, 8, 8, 43, 43, 8, 8, 8, 43, 43, 8, 43, 43, 8, 8, 8, 25, 8, 8, 8, 8, 25, 8, 8, 8, 25, 8, 8, 8, 25, 8, 8, 43, 25, 8, 8, 8, 25, 8, 8, 25, 43, 8, 8, 8, 25, 8, 8, 8, 8, 25, 8, 8, 25, 8, 8, 43, 8, 25, 8, 8, 25, 8, 8, 25, 25, 25, 8, 8, 25, 8, 8, 8, 43, 25, 8, 8, 25, 8, 8, 25, 8, 43, 8, 8, 25, 8, 8, 8, 25, 43, 8, 8, 25, 8, 8, 8, 8, 8, 25, 8, 25, 8, 8, 43, 8, 8, 25, 8, 25, 8, 8, 25, 25, 8, 25, 8, 25, 8, 8, 8, 43, 8, 25, 8, 25, 8, 8, 25, 8, 25, 25, 8, 25, 8, 8, 8, 25, 25, 25, 8, 25, 8, 8, 43, 25, 25, 25, 8, 25, 8, 8, 8, 8, 43, 25, 8, 25, 8, 8, 25, 8, 8, 43, 8, 25, 8, 8, 8, 25, 8, 43, 8, 25, 8, 8, 8, 8, 25, 43, 8, 25, 8, 8, 8, 8, 8, 8, 25, 25, 8, 8, 43, 8, 8, 8, 25, 25, 8, 8, 25, 25, 8, 8, 25, 25, 8, 8, 8, 43, 8, 8, 25, 25, 8, 8, 25, 8, 25, 8, 25, 25, 8, 8, 8, 25, 25, 8, 25, 25, 8, 8, 8, 8, 43, 8, 25, 25, 8, 8, 25, 8, 8, 25, 25, 25, 8, 8, 8, 25, 8, 25, 25, 25, 8, 8, 8, 8, 25, 25, 25, 25, 8, 8, 25, 8, 43, 25, 25, 25, 8, 8, 8, 8, 8, 43, 25, 25, 8, 8, 25, 8, 8, 8, 43, 25, 8, 8, 8, 25, 8, 8, 43, 25, 8, 8, 8, 8, 25, 8, 43, 25, 8, 8, 43, 25, 43, 8, 43, 25, 8, 8, 8, 8, 8, 25, 43, 25, 8, 8, 43, 8, 8, 25, 43, 25, 8, 8, 8, 25, 8, 43, 43, 25, 8, 8, 8, 8, 8, 8, 8, 43, 8, 8, 43, 8, 8, 8, 8, 43, 8, 8, 25, 25, 8, 8, 8, 43, 8, 8, 8, 43, 8, 8, 8, 43, 8, 8, 43, 43, 8, 8, 8, 43, 8, 8, 25, 8, 25, 8, 8, 43, 8, 8, 8, 25, 25, 8, 8, 43, 8, 8, 8, 8, 43, 8, 8, 43, 8, 8, 25, 25, 43, 8, 8, 43, 8, 8, 25, 8, 8, 25, 8, 43, 8, 8, 8, 25, 8, 25, 8, 43, 8, 8, 8, 8, 25, 25, 8, 43, 8, 8, 8, 43, 25, 25, 8, 43, 8, 8, 8, 8, 8, 43, 8, 43, 8, 8, 8, 8, 43, 43, 8, 43, 8, 8, 43, 43, 43, 43, 8, 43, 8, 8, 25, 8, 8, 8, 25, 43, 8, 8, 8, 25, 8, 8, 25, 43, 8, 8, 8, 8, 25, 8, 25, 43, 8, 8, 8, 8, 8, 25, 25, 43, 8, 8, 25, 8, 8, 43, 25, 43, 8, 8, 25, 43, 8, 43, 25, 43, 8, 8, 8, 8, 8, 8, 43, 43, 8, 8, 8, 8, 43, 8, 43, 43, 8, 8, 8, 43, 43, 8, 43, 43, 8, 8, 43, 25, 25, 43, 43, 43, 8, 8, 8, 8, 43, 43, 43, 43, 8, 8, 25, 8, 8, 8, 8, 8, 25, 8, 8, 25, 8, 8, 8, 8, 25, 8, 43, 25, 8, 8, 8, 8, 25, 8, 25, 43, 8, 8, 8, 8, 25, 8, 8, 8, 25, 8, 8, 8, 25, 8, 43, 8, 25, 8, 8, 8, 25, 8, 25, 25, 25, 8, 8, 8, 25, 8, 8, 43, 25, 8, 8, 8, 25, 8, 25, 8, 43, 8, 8, 8, 25, 8, 8, 25, 43, 8, 8, 8, 25, 8, 8, 8, 8, 25, 8, 8, 25, 8, 43, 8, 8, 25, 8, 8, 25, 8, 25, 25, 8, 25, 8, 8, 25, 8, 8, 43, 8, 25, 8, 8, 25, 8, 25, 8, 25, 25, 8, 8, 25, 8, 8, 25, 25, 25, 8, 8, 25, 8, 8, 8, 43, 25, 8, 8, 25, 8, 43, 43, 43, 25, 8, 8, 25, 8, 25, 8, 8, 43, 8, 8, 25, 8, 8, 25, 8, 43, 8, 8, 25, 8, 8, 8, 25, 43, 8, 8, 25, 8, 8, 8, 8, 8, 25, 8, 25, 8, 43, 8, 8, 8, 25, 8, 25, 8, 25, 25, 8, 8, 25, 8, 25, 8, 8, 43, 8, 8, 25, 8, 25, 8, 25, 8, 25, 8, 25, 8, 25, 8, 8, 25, 25, 8, 25, 8, 25, 8, 8, 8, 43, 8, 25, 8, 25, 8, 25, 8, 8, 25, 25, 8, 25, 8, 8, 25, 8, 25, 25, 8, 25, 8, 8, 8, 25, 25, 25, 8, 25, 8, 8, 8, 8, 43, 25, 8, 25, 8, 8, 25, 25, 43, 25, 8, 25, 8, 43, 25, 25, 43, 25, 8, 25, 8, 25, 8, 8, 8, 43, 8, 25, 8, 8, 25, 8, 8, 43, 8, 25, 8, 43, 25, 8, 8, 43, 8, 25, 8, 8, 8, 25, 8, 43, 8, 25, 8, 8, 8, 8, 25, 43, 8, 25, 8, 8, 8, 43, 25, 43, 8, 25, 8, 8, 8, 8, 8, 8, 25, 25, 8, 43, 8, 8, 8, 8, 25, 25, 8, 25, 25, 8, 8, 8, 25, 25, 8, 8, 43, 8, 8, 8, 25, 25, 8, 25, 8, 25, 8, 8, 25, 25, 8, 8, 25, 25, 8, 8, 25, 25, 8, 8, 8, 43, 8, 8, 25, 25, 8, 25, 8, 8, 25, 8, 25, 25, 8, 8, 25, 8, 25, 8, 25, 25, 8, 25, 43, 8, 25, 8, 25, 25, 8, 8, 8, 25, 25, 8, 25, 25, 8, 8, 25, 43, 25, 8, 25, 25, 8, 8, 8, 8, 43, 8, 25, 25, 8, 25, 8, 8, 8, 25, 25, 25, 8, 8, 25, 8, 8, 25, 25, 25, 8, 8, 8, 25, 8, 25, 25, 25, 8, 8, 8, 8, 25, 25, 25, 25, 8, 8, 8, 8, 8, 43, 25, 25, 8, 8, 25, 25, 8, 43, 25, 25, 8, 25, 43, 8, 25, 43, 25, 25, 8, 25, 8, 8, 8, 8, 43, 25, 8, 8, 25, 8, 8, 8, 43, 25, 8, 8, 8, 25, 8, 8, 43, 25, 8, 43, 8, 25, 8, 8, 43, 25, 8, 8, 8, 8, 25, 8, 43, 25, 8, 8, 25, 25, 25, 8, 43, 25, 8, 43, 25, 8, 43, 8, 43, 25, 8, 8, 8, 8, 8, 25, 43, 25, 8, 25, 25, 8, 8, 25, 43, 25, 8, 43, 25, 43, 25, 25, 43, 25, 8, 25, 8, 25, 25, 43, 43, 25, 8, 25, 43, 43, 43, 43, 43, 25, 8, 8, 8, 8, 8, 8, 8, 43, 8, 43, 8, 8, 8, 8, 8, 43, 8, 25, 25, 8, 8, 8, 8, 43, 8, 8, 43, 8, 8, 8, 8, 43, 8, 43, 43, 8, 8, 8, 8, 43, 8, 25, 8, 25, 8, 8, 8, 43, 8, 8, 25, 25, 8, 8, 8, 43, 8, 8, 8, 43, 8, 8, 8, 43, 8, 25, 8, 8, 25, 8, 8, 43, 8, 8, 25, 8, 25, 8, 8, 43, 8, 8, 8, 25, 25, 8, 8, 43, 8, 8, 8, 8, 43, 8, 8, 43, 8, 8, 8, 43, 43, 8, 8, 43, 8, 25, 8, 8, 8, 25, 8, 43, 8, 8, 25, 8, 8, 25, 8, 43, 8, 8, 8, 25, 8, 25, 8, 43, 8, 8, 8, 8, 25, 25, 8, 43, 8, 8, 43, 8, 25, 25, 8, 43, 8, 25, 25, 43, 25, 25, 8, 43, 8, 8, 8, 8, 8, 43, 8, 43, 8, 43, 8, 43, 8, 43, 8, 43, 8, 8, 8, 8, 43, 43, 8, 43, 8, 8, 43, 43, 43, 43, 8, 43, 8, 25, 8, 8, 8, 8, 25, 43, 8, 8, 25, 8, 8, 8, 25, 43, 8, 8, 8, 25, 8, 8, 25, 43, 8, 25, 43, 43, 8, 8, 25, 43, 8, 8, 8, 8, 25, 8, 25, 43, 8, 8, 8, 8, 8, 25, 25, 43, 8, 25, 8, 8, 25, 25, 25, 43, 8, 43, 8, 25, 25, 25, 25, 43, 8, 25, 43, 25, 43, 25, 25, 43, 8, 25, 8, 8, 8, 43, 25, 43, 8, 43, 43, 25, 8, 43, 25, 43, 8, 43, 25, 43, 43, 43, 25, 43, 8, 8, 8, 8, 8, 8, 43, 43, 8, 8, 43, 8, 8, 8, 43, 43, 8, 43, 43, 8, 8, 8, 43, 43, 8, 8, 8, 43, 8, 8, 43, 43, 8, 25, 25, 25, 25, 8, 43, 43, 8, 8, 43, 8, 43, 8, 43, 43, 8, 43, 8, 43, 43, 8, 43, 43, 8, 8, 43, 43, 25, 25, 43, 43, 8, 8, 8, 25, 43, 25, 43, 43, 8, 8, 43, 8, 8, 43, 43, 43, 8, 8, 8, 43, 8, 43, 43, 43, 8, 43, 8, 8, 43, 43, 43, 43, 8, 8, 43, 8, 43, 43, 43, 43, 8, 43, 43, 8, 43, 43, 43, 43, 8, 25, 8, 8, 8, 8, 8, 8, 25, 8, 25, 8, 8, 8, 8, 8, 25, 43, 25, 8, 8, 8, 8, 8, 25, 25, 43, 8, 8, 8, 8, 8, 25, 8, 8, 25, 8, 8, 8, 8, 25, 43, 8, 25, 8, 8, 8, 8, 25, 25, 25, 25, 8, 8, 8, 8, 25, 8, 43, 25, 8, 8, 8, 8, 25, 25, 8, 43, 8, 8, 8, 8, 25, 8, 25, 43, 8, 8, 8, 8, 25, 8, 8, 8, 25, 8, 8, 8, 25, 43, 8, 8, 25, 8, 8, 8, 25, 25, 25, 8, 25, 8, 8, 8, 25, 8, 43, 8, 25, 8, 8, 8, 25, 43, 43, 8, 25, 8, 8, 8, 25, 25, 8, 25, 25, 8, 8, 8, 25, 8, 25, 25, 25, 8, 8, 8, 25, 8, 8, 43, 25, 8, 8, 8, 25, 25, 25, 43, 25, 8, 8, 8, 25, 25, 8, 8, 43, 8, 8, 8, 25, 8, 25, 8, 43, 8, 8, 8, 25, 8, 8, 25, 43, 8, 8, 8, 25, 8, 8, 8, 8, 25, 8, 8, 25, 43, 8, 8, 8, 25, 8, 8, 25, 25, 25, 8, 8, 25, 8, 8, 25, 8, 43, 8, 8, 25, 8, 8, 25, 25, 8, 25, 8, 25, 8, 8, 25, 8, 25, 25, 8, 25, 8, 8, 25, 8, 8, 43, 8, 25, 8, 8, 25, 25, 8, 8, 25, 25, 8, 8, 25, 8, 25, 8, 25, 25, 8, 8, 25, 8, 8, 25, 25, 25, 8, 8, 25, 8, 8, 8, 43, 25, 8, 8, 25, 25, 25, 8, 43, 25, 8, 8, 25, 43, 8, 43, 43, 25, 8, 8, 25, 25, 8, 8, 8, 43, 8, 8, 25, 8, 25, 8, 8, 43, 8, 8, 25, 8, 8, 25, 8, 43, 8, 8, 25, 43, 8, 25, 8, 43, 8, 8, 25, 25, 43, 43, 8, 43, 8, 8, 25, 8, 8, 8, 25, 43, 8, 8, 25, 8, 8, 8, 8, 8, 25, 8, 25, 43, 8, 8, 8, 8, 25, 8, 25, 25, 25, 8, 8, 8, 25, 8, 25, 8, 43, 8, 8, 8, 25, 8, 25, 25, 8, 25, 8, 8, 25, 8, 25, 8, 25, 25, 8, 8, 25, 8, 25, 25, 43, 25, 8, 8, 25, 8, 25, 8, 8, 43, 8, 8, 25, 8, 25, 25, 8, 8, 25, 8, 25, 8, 25, 8, 25, 8, 25, 8, 25, 8, 25, 8, 8, 25, 25, 8, 25, 8, 25, 8, 8, 8, 43, 8, 25, 8, 25, 8, 25, 25, 43, 8, 25, 8, 25, 25, 8, 8, 8, 25, 25, 8, 25, 8, 25, 8, 8, 25, 25, 8, 25, 8, 8, 25, 8, 25, 25, 8, 25, 8, 25, 43, 8, 25, 25, 8, 25, 8, 8, 8, 25, 25, 25, 8, 25, 43, 43, 25, 43, 25, 25, 8, 25, 8, 8, 8, 8, 43, 25, 8, 25, 43, 43, 8, 8, 43, 25, 8, 25, 8, 25, 8, 25, 43, 25, 8, 25, 8, 8, 25, 25, 43, 25, 8, 25, 25, 8, 8, 8, 8, 43, 8, 25, 8, 25, 8, 8, 8, 43, 8, 25, 8, 8, 25, 8, 8, 43, 8, 25, 8, 8, 8, 25, 8, 43, 8, 25, 25, 25, 8, 25, 8, 43, 8, 25, 8, 25, 25, 25, 8, 43, 8, 25, 43, 8, 43, 25, 8, 43, 8, 25, 8, 8, 8, 8, 25, 43, 8, 25, 25, 8, 25, 8, 25, 43, 8, 25, 8, 25, 8, 25, 25, 43, 8, 25, 8, 8, 25, 25, 25, 43, 8, 25, 25, 43, 43, 25, 25, 43, 8, 25, 8, 25, 8, 8, 43, 43, 8, 25, 8, 8, 8, 8, 8, 8, 25, 25, 43, 8, 8, 8, 8, 8, 25, 25, 25, 25, 8, 8, 8, 8, 25, 25, 8, 43, 8, 8, 8, 8, 25, 25, 25, 8, 25, 8, 8, 8, 25, 25, 8, 25, 25, 8, 8, 8, 25, 25, 8, 8, 43, 8, 8, 8, 25, 25, 8, 43, 43, 8, 8, 8, 25, 25, 25, 8, 8, 25, 8, 8, 25, 25, 8, 25, 8, 25, 8, 8, 25, 25, 8, 8, 25, 25, 8, 8, 25, 25, 8, 8, 8, 43, 8, 8, 25, 25, 25, 8, 8, 8, 25, 8, 25, 25, 8, 25, 8, 8, 25, 8, 25, 25, 8, 8, 25, 8, 25, 8, 25, 25, 25, 25, 25, 8, 25, 8, 25, 25, 8, 8, 8, 25, 25, 8, 25, 25, 43, 8, 8, 25, 25, 8, 25, 25, 8, 8, 8, 8, 43, 8, 25, 25, 8, 25, 8, 25, 43, 8, 25, 25, 43, 43, 43, 43, 43, 8, 25, 25, 25, 8, 8, 8, 8, 25, 25, 25, 8, 25, 8, 8, 8, 25, 25, 25, 8, 8, 25, 8, 8, 25, 25, 25, 25, 8, 43, 8, 8, 25, 25, 25, 8, 8, 8, 25, 8, 25, 25, 25, 8, 8, 43, 25, 8, 25, 25, 25, 25, 8, 8, 43, 8, 25, 25, 25, 25, 8, 43, 43, 8, 25, 25, 25, 8, 8, 8, 8, 25, 25, 25, 25, 8, 43, 8, 8, 25, 25, 25, 25, 8, 8, 8, 43, 25, 25, 25, 25, 8, 43, 8, 43, 25, 25, 25, 25, 25, 8, 43, 8, 43, 25, 25, 25, 8, 43, 43, 25, 43, 25, 25, 25, 25, 8, 43, 43, 43, 25, 25, 25, 8, 8, 8, 8, 8, 43, 25, 25, 8, 25, 25, 8, 8, 43, 25, 25, 25, 8, 8, 25, 8, 43, 25, 25, 8, 8, 25, 25, 8, 43, 25, 25, 25, 43, 25, 43, 8, 43, 25, 25, 43, 43, 25, 8, 25, 43, 25, 25, 8, 8, 8, 25, 25, 43, 25, 25, 43, 8, 8, 25, 25, 43, 25, 25, 25, 25, 8, 43, 43, 43, 25, 25, 25, 8, 8, 8, 8, 8, 43, 25, 8, 25, 8, 8, 8, 8, 43, 25, 8, 8, 25, 8, 8, 8, 43, 25, 8, 8, 8, 25, 8, 8, 43, 25, 8, 25, 25, 25, 8, 8, 43, 25, 43, 8, 43, 25, 8, 8, 43, 25, 43, 25, 8, 43, 8, 8, 43, 25, 25, 43, 43, 43, 8, 8, 43, 25, 8, 8, 8, 8, 25, 8, 43, 25, 8, 25, 43, 8, 43, 8, 43, 25, 43, 43, 8, 25, 43, 8, 43, 25, 43, 8, 25, 43, 43, 8, 43, 25, 8, 8, 8, 8, 8, 25, 43, 25, 43, 25, 25, 8, 8, 25, 43, 25, 8, 8, 25, 8, 25, 25, 43, 25, 8, 8, 8, 25, 25, 25, 43, 25, 25, 25, 8, 25, 25, 25, 43, 25, 8, 25, 43, 43, 25, 25, 43, 25, 25, 8, 8, 8, 8, 43, 43, 25, 43, 43, 43, 25, 8, 43, 43, 25, 25, 25, 43, 8, 25, 43, 43, 25, 43, 25, 8, 8, 43, 43, 43, 25, 8, 25, 25, 25, 43, 43, 43, 25, 43, 8, 43, 25, 43, 43, 43, 25, 8, 8, 8, 8, 8, 8, 8, 43, 43, 8, 8, 8, 8, 8, 8, 43, 25, 25, 8, 8, 8, 8, 8, 43, 8, 43, 8, 8, 8, 8, 8, 43, 25, 8, 25, 8, 8, 8, 8, 43, 8, 25, 25, 8, 8, 8, 8, 43, 8, 8, 43, 8, 8, 8, 8, 43, 43, 43, 43, 8, 8, 8, 8, 43, 25, 8, 8, 25, 8, 8, 8, 43, 8, 25, 8, 25, 8, 8, 8, 43, 8, 8, 25, 25, 8, 8, 8, 43, 8, 8, 8, 43, 8, 8, 8, 43, 43, 8, 8, 43, 8, 8, 8, 43, 8, 43, 43, 43, 8, 8, 8, 43, 43, 43, 43, 43, 8, 8, 8, 43, 25, 8, 8, 8, 25, 8, 8, 43, 8, 25, 8, 8, 25, 8, 8, 43, 43, 25, 8, 8, 25, 8, 8, 43, 8, 8, 25, 8, 25, 8, 8, 43, 8, 8, 8, 25, 25, 8, 8, 43, 25, 8, 25, 25, 25, 8, 8, 43, 25, 43, 25, 25, 25, 8, 8, 43, 8, 8, 8, 8, 43, 8, 8, 43, 8, 8, 43, 8, 43, 8, 8, 43, 8, 8, 8, 43, 43, 8, 8, 43, 43, 8, 8, 43, 43, 8, 8, 43, 8, 8, 43, 43, 43, 8, 8, 43, 8, 43, 43, 43, 43, 8, 8, 43, 25, 8, 8, 8, 8, 25, 8, 43, 8, 25, 8, 8, 8, 25, 8, 43, 8, 8, 25, 8, 8, 25, 8, 43, 43, 8, 25, 8, 8, 25, 8, 43, 25, 25, 25, 8, 8, 25, 8, 43, 8, 8, 8, 25, 8, 25, 8, 43, 8, 8, 43, 25, 8, 25, 8, 43, 25, 43, 8, 43, 8, 25, 8, 43, 8, 8, 8, 8, 25, 25, 8, 43, 8, 25, 8, 25, 25, 25, 8, 43, 25, 25, 43, 43, 25, 25, 8, 43, 8, 43, 25, 8, 43, 25, 8, 43, 43, 43, 43, 25, 43, 25, 8, 43, 8, 8, 8, 8, 8, 43, 8, 43, 8, 43, 8, 8, 8, 43, 8, 43, 25, 25, 43, 8, 8, 43, 8, 43, 43, 43, 25, 25, 8, 43, 8, 43, 8, 8, 8, 43, 8, 43, 8, 43, 43, 8, 8, 43, 8, 43, 8, 43, 8, 43, 43, 43, 8, 43, 8, 43, 43, 25, 8, 8, 25, 43, 8, 43, 43, 8, 43, 8, 43, 43, 8, 43, 8, 8, 8, 43, 43, 43, 8, 43, 8, 43, 8, 43, 43, 43, 8, 43, 43, 25, 25, 43, 43, 43, 8, 43, 8, 43, 43, 43, 43, 43, 8, 43, 25, 8, 8, 8, 8, 8, 25, 43, 8, 25, 8, 8, 8, 8, 25, 43, 8, 8, 25, 8, 8, 8, 25, 43, 8, 8, 8, 25, 8, 8, 25, 43, 43, 25, 25, 25, 8, 8, 25, 43, 8, 25, 8, 43, 8, 8, 25, 43, 8, 8, 8, 8, 25, 8, 25, 43, 43, 8, 43, 8, 25, 8, 25, 43, 8, 25, 43, 25, 25, 8, 25, 43, 43, 25, 25, 25, 43, 8, 25, 43, 25, 43, 8, 43, 43, 8, 25, 43, 8, 8, 8, 8, 8, 25, 25, 43, 25, 25, 8, 8, 8, 25, 25, 43, 8, 25, 8, 25, 8, 25, 25, 43, 8, 8, 25, 25, 8, 25, 25, 43, 8, 43, 25, 25, 8, 25, 25, 43, 25, 43, 43, 8, 25, 25, 25, 43, 8, 8, 25, 43, 25, 25, 25, 43, 43, 8, 25, 43, 25, 25, 25, 43, 25, 8, 8, 25, 43, 25, 25, 43, 25, 8, 25, 25, 8, 43, 25, 43, 43, 25, 43, 43, 8, 43, 25, 43, 25, 43, 8, 25, 25, 43, 25, 43, 25, 25, 25, 8, 43, 43, 25, 43, 8, 8, 43, 25, 43, 43, 25, 43, 8, 8, 8, 8, 8, 8, 43, 43, 43, 8, 8, 8, 8, 8, 43, 43, 8, 43, 8, 8, 8, 8, 43, 43, 43, 43, 8, 8, 8, 8, 43, 43, 8, 8, 43, 8, 8, 8, 43, 43, 43, 43, 43, 8, 8, 8, 43, 43, 8, 8, 43, 43, 8, 8, 43, 43, 25, 8, 25, 25, 25, 8, 43, 43, 25, 43, 25, 25, 25, 8, 43, 43, 43, 25, 43, 43, 25, 8, 43, 43, 8, 8, 8, 8, 43, 8, 43, 43, 43, 8, 8, 8, 43, 8, 43, 43, 8, 43, 8, 8, 43, 8, 43, 43, 43, 43, 43, 8, 43, 8, 43, 43, 8, 8, 8, 43, 43, 8, 43, 43, 8, 8, 43, 43, 43, 8, 43, 43, 8, 8, 8, 25, 8, 25, 43, 43, 25, 25, 25, 43, 8, 25, 43, 43, 25, 25, 43, 25, 43, 25, 43, 43, 8, 43, 25, 43, 43, 25, 43, 43, 43, 43, 8, 8, 8, 43, 43, 43, 8, 8, 43, 8, 8, 43, 43, 43, 43, 8, 43, 8, 8, 43, 43, 43, 8, 43, 43, 8, 8, 43, 43, 43, 8, 8, 43, 43, 8, 43, 43, 43, 8, 43, 43, 43, 8, 43, 43, 43, 8, 25, 8, 8, 25, 43, 43, 43, 8, 25, 8, 43, 25, 43, 43, 43, 43, 25, 8, 43, 25, 43, 43, 43, 8, 43, 43, 8, 43, 43, 43, 43, 43, 43, 43, 8, 43, 43, 43, 43, 25, 8, 25, 43, 43, 43, 43, 43, 43, 43, 43, 43, 43, 43, 43, 43};
.global .align 8 .b8 iq2s_grid[8192] = {8, 8, 8, 8, 8, 8, 8, 8, 43, 8, 8, 8, 8, 8, 8, 8, 25, 25, 8, 8, 8, 8, 8, 8, 8, 43, 8, 8, 8, 8, 8, 8, 43, 43, 8, 8, 8, 8, 8, 8, 25, 8, 25, 8, 8, 8, 8, 8, 8, 25, 25, 8, 8, 8, 8, 8, 43, 25, 25, 8, 8, 8, 8, 8, 25, 43, 25, 8, 8, 8, 8, 8, 8, 8, 43, 8, 8, 8, 8, 8, 43, 8, 43, 8, 8, 8, 8, 8, 25, 25, 43, 8, 8, 8, 8, 8, 8, 43, 43, 8, 8, 8, 8, 8, 25, 8, 8, 25, 8, 8, 8, 8, 8, 25, 8, 25, 8, 8, 8, 8, 43, 25, 8, 25, 8, 8, 8, 8, 25, 43, 8, 25, 8, 8, 8, 8, 8, 8, 25, 25, 8, 8, 8, 8, 43, 8, 25, 25, 8, 8, 8, 8, 25, 25, 25, 25, 8, 8, 8, 8, 8, 43, 25, 25, 8, 8, 8, 8, 25, 8, 43, 25, 8, 8, 8, 8, 8, 25, 43, 25, 8, 8, 8, 8, 43, 25, 43, 25, 8, 8, 8, 8, 25, 43, 43, 25, 8, 8, 8, 8, 8, 8, 8, 43, 8, 8, 8, 8, 43, 8, 8, 43, 8, 8, 8, 8, 25, 25, 8, 43, 8, 8, 8, 8, 8, 43, 8, 43, 8, 8, 8, 8, 25, 8, 25, 43, 8, 8, 8, 8, 8, 25, 25, 43, 8, 8, 8, 8, 8, 8, 43, 43, 8, 8, 8, 8, 25, 25, 43, 43, 8, 8, 8, 8, 43, 43, 43, 43, 8, 8, 8, 8, 25, 8, 8, 8, 25, 8, 8, 8, 8, 25, 8, 8, 25, 8, 8, 8, 43, 25, 8, 8, 25, 8, 8, 8, 25, 43, 8, 8, 25, 8, 8, 8, 8, 8, 25, 8, 25, 8, 8, 8, 43, 8, 25, 8, 25, 8, 8, 8, 25, 25, 25, 8, 25, 8, 8, 8, 8, 43, 25, 8, 25, 8, 8, 8, 25, 8, 43, 8, 25, 8, 8, 8, 8, 25, 43, 8, 25, 8, 8, 8, 8, 8, 8, 25, 25, 8, 8, 8, 43, 8, 8, 25, 25, 8, 8, 8, 25, 25, 8, 25, 25, 8, 8, 8, 8, 43, 8, 25, 25, 8, 8, 8, 25, 8, 25, 25, 25, 8, 8, 8, 8, 25, 25, 25, 25, 8, 8, 8, 43, 25, 25, 25, 25, 8, 8, 8, 25, 43, 25, 25, 25, 8, 8, 8, 8, 8, 43, 25, 25, 8, 8, 8, 25, 25, 43, 25, 25, 8, 8, 8, 8, 43, 43, 25, 25, 8, 8, 8, 25, 8, 8, 43, 25, 8, 8, 8, 8, 25, 8, 43, 25, 8, 8, 8, 8, 8, 25, 43, 25, 8, 8, 8, 43, 8, 25, 43, 25, 8, 8, 8, 25, 25, 25, 43, 25, 8, 8, 8, 25, 8, 43, 43, 25, 8, 8, 8, 8, 25, 43, 43, 25, 8, 8, 8, 8, 8, 8, 8, 43, 8, 8, 8, 43, 8, 8, 8, 43, 8, 8, 8, 25, 25, 8, 8, 43, 8, 8, 8, 8, 43, 8, 8, 43, 8, 8, 8, 25, 8, 25, 8, 43, 8, 8, 8, 8, 25, 25, 8, 43, 8, 8, 8, 8, 8, 43, 8, 43, 8, 8, 8, 43, 43, 43, 8, 43, 8, 8, 8, 25, 8, 8, 25, 43, 8, 8, 8, 8, 25, 8, 25, 43, 8, 8, 8, 43, 25, 8, 25, 43, 8, 8, 8, 25, 43, 8, 25, 43, 8, 8, 8, 8, 8, 25, 25, 43, 8, 8, 8, 25, 25, 25, 25, 43, 8, 8, 8, 8, 8, 8, 43, 43, 8, 8, 8, 25, 25, 8, 43, 43, 8, 8, 8, 43, 43, 8, 43, 43, 8, 8, 8, 8, 25, 25, 43, 43, 8, 8, 8, 43, 8, 43, 43, 43, 8, 8, 8, 25, 8, 8, 8, 8, 25, 8, 8, 8, 25, 8, 8, 8, 25, 8, 8, 43, 25, 8, 8, 8, 25, 8, 8, 25, 43, 8, 8, 8, 25, 8, 8, 8, 8, 25, 8, 8, 25, 8, 8, 43, 8, 25, 8, 8, 25, 8, 8, 25, 25, 25, 8, 8, 25, 8, 8, 8, 43, 25, 8, 8, 25, 8, 8, 25, 8, 43, 8, 8, 25, 8, 8, 8, 25, 43, 8, 8, 25, 8, 8, 43, 25, 43, 8, 8, 25, 8, 8, 25, 43, 43, 8, 8, 25, 8, 8, 8, 8, 8, 25, 8, 25, 8, 8, 43, 8, 8, 25, 8, 25, 8, 8, 25, 25, 8, 25, 8, 25, 8, 8, 8, 43, 8, 25, 8, 25, 8, 8, 43, 43, 8, 25, 8, 25, 8, 8, 25, 8, 25, 25, 8, 25, 8, 8, 8, 25, 25, 25, 8, 25, 8, 8, 43, 25, 25, 25, 8, 25, 8, 8, 25, 43, 25, 25, 8, 25, 8, 8, 8, 8, 43, 25, 8, 25, 8, 8, 43, 8, 43, 25, 8, 25, 8, 8, 25, 25, 43, 25, 8, 25, 8, 8, 25, 8, 8, 43, 8, 25, 8, 8, 8, 25, 8, 43, 8, 25, 8, 8, 43, 25, 8, 43, 8, 25, 8, 8, 25, 43, 8, 43, 8, 25, 8, 8, 8, 8, 25, 43, 8, 25, 8, 8, 25, 25, 25, 43, 8, 25, 8, 8, 8, 43, 25, 43, 8, 25, 8, 8, 25, 8, 43, 43, 8, 25, 8, 8, 8, 25, 43, 43, 8, 25, 8, 8, 8, 8, 8, 8, 25, 25, 8, 8, 43, 8, 8, 8, 25, 25, 8, 8, 25, 25, 8, 8, 25, 25, 8, 8, 8, 43, 8, 8, 25, 25, 8, 8, 43, 43, 8, 8, 25, 25, 8, 8, 25, 8, 25, 8, 25, 25, 8, 8, 8, 25, 25, 8, 25, 25, 8, 8, 43, 25, 25, 8, 25, 25, 8, 8, 25, 43, 25, 8, 25, 25, 8, 8, 8, 8, 43, 8, 25, 25, 8, 8, 25, 25, 43, 8, 25, 25, 8, 8, 8, 43, 43, 8, 25, 25, 8, 8, 25, 8, 8, 25, 25, 25, 8, 8, 8, 25, 8, 25, 25, 25, 8, 8, 43, 25, 8, 25, 25, 25, 8, 8, 25, 43, 8, 25, 25, 25, 8, 8, 8, 8, 25, 25, 25, 25, 8, 8, 43, 8, 25, 25, 25, 25, 8, 8, 25, 25, 25, 25, 25, 25, 8, 8, 8, 43, 25, 25, 25, 25, 8, 8, 25, 8, 43, 25, 25, 25, 8, 8, 8, 25, 43, 25, 25, 25, 8, 8, 8, 8, 8, 43, 25, 25, 8, 8, 43, 8, 8, 43, 25, 25, 8, 8, 25, 25, 8, 43, 25, 25, 8, 8, 8, 43, 8, 43, 25, 25, 8, 8, 25, 8, 25, 43, 25, 25, 8, 8, 8, 25, 25, 43, 25, 25, 8, 8, 8, 8, 43, 43, 25, 25, 8, 8, 25, 8, 8, 8, 43, 25, 8, 8, 8, 25, 8, 8, 43, 25, 8, 8, 43, 25, 8, 8, 43, 25, 8, 8, 25, 43, 8, 8, 43, 25, 8, 8, 8, 8, 25, 8, 43, 25, 8, 8, 25, 25, 25, 8, 43, 25, 8, 8, 8, 8, 8, 25, 43, 25, 8, 8, 25, 25, 8, 25, 43, 25, 8, 8, 8, 43, 8, 25, 43, 25, 8, 8, 25, 8, 25, 25, 43, 25, 8, 8, 8, 25, 25, 25, 43, 25, 8, 8, 8, 8, 43, 25, 43, 25, 8, 8, 25, 8, 8, 43, 43, 25, 8, 8, 8, 25, 8, 43, 43, 25, 8, 8, 8, 8, 25, 43, 43, 25, 8, 8, 8, 8, 8, 8, 8, 43, 8, 8, 43, 8, 8, 8, 8, 43, 8, 8, 25, 25, 8, 8, 8, 43, 8, 8, 8, 43, 8, 8, 8, 43, 8, 8, 25, 8, 25, 8, 8, 43, 8, 8, 8, 25, 25, 8, 8, 43, 8, 8, 43, 25, 25, 8, 8, 43, 8, 8, 25, 43, 25, 8, 8, 43, 8, 8, 8, 8, 43, 8, 8, 43, 8, 8, 25, 25, 43, 8, 8, 43, 8, 8, 43, 43, 43, 8, 8, 43, 8, 8, 25, 8, 8, 25, 8, 43, 8, 8, 8, 25, 8, 25, 8, 43, 8, 8, 43, 25, 8, 25, 8, 43, 8, 8, 25, 43, 8, 25, 8, 43, 8, 8, 8, 8, 25, 25, 8, 43, 8, 8, 43, 8, 25, 25, 8, 43, 8, 8, 25, 25, 25, 25, 8, 43, 8, 8, 8, 43, 25, 25, 8, 43, 8, 8, 25, 8, 43, 25, 8, 43, 8, 8, 8, 25, 43, 25, 8, 43, 8, 8, 8, 8, 8, 43, 8, 43, 8, 8, 25, 25, 8, 43, 8, 43, 8, 8, 8, 25, 25, 43, 8, 43, 8, 8, 43, 43, 43, 43, 8, 43, 8, 8, 25, 8, 8, 8, 25, 43, 8, 8, 8, 25, 8, 8, 25, 43, 8, 8, 8, 8, 25, 8, 25, 43, 8, 8, 43, 8, 25, 8, 25, 43, 8, 8, 25, 25, 25, 8, 25, 43, 8, 8, 8, 43, 25, 8, 25, 43, 8, 8, 25, 8, 43, 8, 25, 43, 8, 8, 8, 8, 8, 25, 25, 43, 8, 8, 25, 25, 8, 25, 25, 43, 8, 8, 8, 43, 8, 25, 25, 43, 8, 8, 25, 8, 25, 25, 25, 43, 8, 8, 8, 25, 25, 25, 25, 43, 8, 8, 8, 8, 43, 25, 25, 43, 8, 8, 25, 8, 8, 43, 25, 43, 8, 8, 8, 8, 25, 43, 25, 43, 8, 8, 8, 8, 8, 8, 43, 43, 8, 8, 25, 8, 25, 8, 43, 43, 8, 8, 8, 25, 25, 8, 43, 43, 8, 8, 43, 8, 43, 8, 43, 43, 8, 8, 8, 43, 43, 8, 43, 43, 8, 8, 43, 43, 43, 8, 43, 43, 8, 8, 8, 8, 25, 25, 43, 43, 8, 8, 25, 43, 25, 43, 43, 43, 8, 8, 25, 8, 8, 8, 8, 8, 25, 8, 8, 25, 8, 8, 8, 8, 25, 8, 43, 25, 8, 8, 8, 8, 25, 8, 25, 43, 8, 8, 8, 8, 25, 8, 8, 8, 25, 8, 8, 8, 25, 8, 43, 8, 25, 8, 8, 8, 25, 8, 25, 25, 25, 8, 8, 8, 25, 8, 8, 43, 25, 8, 8, 8, 25, 8, 25, 8, 43, 8, 8, 8, 25, 8, 8, 25, 43, 8, 8, 8, 25, 8, 43, 25, 43, 8, 8, 8, 25, 8, 8, 8, 8, 25, 8, 8, 25, 8, 43, 8, 8, 25, 8, 8, 25, 8, 25, 25, 8, 25, 8, 8, 25, 8, 8, 43, 8, 25, 8, 8, 25, 8, 25, 8, 25, 25, 8, 8, 25, 8, 8, 25, 25, 25, 8, 8, 25, 8, 43, 25, 25, 25, 8, 8, 25, 8, 25, 43, 25, 25, 8, 8, 25, 8, 8, 8, 43, 25, 8, 8, 25, 8, 43, 8, 43, 25, 8, 8, 25, 8, 25, 25, 43, 25, 8, 8, 25, 8, 8, 43, 43, 25, 8, 8, 25, 8, 25, 8, 8, 43, 8, 8, 25, 8, 8, 25, 8, 43, 8, 8, 25, 8, 43, 25, 8, 43, 8, 8, 25, 8, 8, 8, 25, 43, 8, 8, 25, 8, 25, 25, 25, 43, 8, 8, 25, 8, 8, 43, 25, 43, 8, 8, 25, 8, 25, 8, 43, 43, 8, 8, 25, 8, 8, 25, 43, 43, 8, 8, 25, 8, 8, 8, 8, 8, 25, 8, 25, 8, 43, 8, 8, 8, 25, 8, 25, 8, 25, 25, 8, 8, 25, 8, 25, 8, 8, 43, 8, 8, 25, 8, 25, 8, 43, 43, 8, 8, 25, 8, 25, 8, 25, 8, 25, 8, 25, 8, 25, 8, 8, 25, 25, 8, 25, 8, 25, 8, 43, 25, 25, 8, 25, 8, 25, 8, 25, 43, 25, 8, 25, 8, 25, 8, 8, 8, 43, 8, 25, 8, 25, 8, 43, 8, 43, 8, 25, 8, 25, 8, 25, 25, 43, 8, 25, 8, 25, 8, 8, 43, 43, 8, 25, 8, 25, 8, 25, 8, 8, 25, 25, 8, 25, 8, 8, 25, 8, 25, 25, 8, 25, 8, 43, 25, 8, 25, 25, 8, 25, 8, 25, 43, 8, 25, 25, 8, 25, 8, 8, 8, 25, 25, 25, 8, 25, 8, 43, 8, 25, 25, 25, 8, 25, 8, 25, 25, 25, 25, 25, 8, 25, 8, 8, 43, 25, 25, 25, 8, 25, 8, 25, 8, 43, 25, 25, 8, 25, 8, 8, 25, 43, 25, 25, 8, 25, 8, 8, 8, 8, 43, 25, 8, 25, 8, 43, 8, 8, 43, 25, 8, 25, 8, 25, 25, 8, 43, 25, 8, 25, 8, 8, 43, 8, 43, 25, 8, 25, 8, 25, 8, 25, 43, 25, 8, 25, 8, 8, 25, 25, 43, 25, 8, 25, 8, 25, 8, 8, 8, 43, 8, 25, 8, 8, 25, 8, 8, 43, 8, 25, 8, 25, 43, 8, 8, 43, 8, 25, 8, 8, 8, 25, 8, 43, 8, 25, 8, 25, 25, 25, 8, 43, 8, 25, 8, 25, 8, 43, 8, 43, 8, 25, 8, 8, 25, 43, 8, 43, 8, 25, 8, 8, 8, 8, 25, 43, 8, 25, 8, 25, 25, 8, 25, 43, 8, 25, 8, 25, 8, 25, 25, 43, 8, 25, 8, 8, 25, 25, 25, 43, 8, 25, 8, 25, 8, 8, 43, 43, 8, 25, 8, 8, 25, 8, 43, 43, 8, 25, 8, 8, 8, 25, 43, 43, 8, 25, 8, 8, 8, 8, 8, 8, 25, 25, 8, 43, 8, 8, 8, 8, 25, 25, 8, 25, 25, 8, 8, 8, 25, 25, 8, 8, 43, 8, 8, 8, 25, 25, 8, 25, 8, 25, 8, 8, 25, 25, 8, 8, 25, 25, 8, 8, 25, 25, 8, 43, 25, 25, 8, 8, 25, 25, 8, 25, 43, 25, 8, 8, 25, 25, 8, 8, 8, 43, 8, 8, 25, 25, 8, 25, 25, 43, 8, 8, 25, 25, 8, 8, 43, 43, 8, 8, 25, 25, 8, 25, 8, 8, 25, 8, 25, 25, 8, 8, 25, 8, 25, 8, 25, 25, 8, 43, 25, 8, 25, 8, 25, 25, 8, 25, 43, 8, 25, 8, 25, 25, 8, 8, 8, 25, 25, 8, 25, 25, 8, 43, 8, 25, 25, 8, 25, 25, 8, 25, 25, 25, 25, 8, 25, 25, 8, 8, 43, 25, 25, 8, 25, 25, 8, 25, 8, 43, 25, 8, 25, 25, 8, 8, 25, 43, 25, 8, 25, 25, 8, 8, 8, 8, 43, 8, 25, 25, 8, 43, 8, 8, 43, 8, 25, 25, 8, 25, 25, 8, 43, 8, 25, 25, 8, 8, 43, 8, 43, 8, 25, 25, 8, 25, 8, 25, 43, 8, 25, 25, 8, 8, 25, 25, 43, 8, 25, 25, 8, 8, 8, 43, 43, 8, 25, 25, 8, 25, 8, 8, 8, 25, 25, 25, 8, 8, 25, 8, 8, 25, 25, 25, 8, 43, 25, 8, 8, 25, 25, 25, 8, 25, 43, 8, 8, 25, 25, 25, 8, 8, 8, 25, 8, 25, 25, 25, 8, 43, 8, 25, 8, 25, 25, 25, 8, 25, 25, 25, 8, 25, 25, 25, 8, 8, 43, 25, 8, 25, 25, 25, 8, 25, 8, 43, 8, 25, 25, 25, 8, 8, 25, 43, 8, 25, 25, 25, 8, 8, 8, 8, 25, 25, 25, 25, 8, 43, 8, 8, 25, 25, 25, 25, 8, 25, 25, 8, 25, 25, 25, 25, 8, 8, 43, 8, 25, 25, 25, 25, 8, 25, 8, 25, 25, 25, 25, 25, 8, 8, 25, 25, 25, 25, 25, 25, 8, 8, 8, 43, 25, 25, 25, 25, 8, 25, 8, 8, 43, 25, 25, 25, 8, 8, 25, 8, 43, 25, 25, 25, 8, 8, 8, 25, 43, 25, 25, 25, 8, 8, 8, 8, 8, 43, 25, 25, 8, 25, 25, 8, 8, 43, 25, 25, 8, 8, 43, 8, 8, 43, 25, 25, 8, 25, 8, 25, 8, 43, 25, 25, 8, 8, 25, 25, 8, 43, 25, 25, 8, 8, 8, 43, 8, 43, 25, 25, 8, 25, 8, 8, 25, 43, 25, 25, 8, 8, 25, 8, 25, 43, 25, 25, 8, 8, 8, 25, 25, 43, 25, 25, 8, 8, 8, 8, 43, 43, 25, 25, 8, 43, 43, 43, 43, 43, 25, 25, 8, 25, 8, 8, 8, 8, 43, 25, 8, 8, 25, 8, 8, 8, 43, 25, 8, 43, 25, 8, 8, 8, 43, 25, 8, 25, 43, 8, 8, 8, 43, 25, 8, 8, 8, 25, 8, 8, 43, 25, 8, 25, 25, 25, 8, 8, 43, 25, 8, 8, 43, 25, 8, 8, 43, 25, 8, 25, 8, 43, 8, 8, 43, 25, 8, 8, 8, 8, 25, 8, 43, 25, 8, 43, 8, 8, 25, 8, 43, 25, 8, 25, 25, 8, 25, 8, 43, 25, 8, 8, 43, 8, 25, 8, 43, 25, 8, 25, 8, 25, 25, 8, 43, 25, 8, 8, 25, 25, 25, 8, 43, 25, 8, 8, 8, 43, 25, 8, 43, 25, 8, 25, 8, 8, 43, 8, 43, 25, 8, 8, 25, 8, 43, 8, 43, 25, 8, 8, 8, 8, 8, 25, 43, 25, 8, 43, 8, 8, 8, 25, 43, 25, 8, 25, 25, 8, 8, 25, 43, 25, 8, 8, 43, 8, 8, 25, 43, 25, 8, 25, 8, 25, 8, 25, 43, 25, 8, 8, 25, 25, 8, 25, 43, 25, 8, 8, 8, 43, 8, 25, 43, 25, 8, 25, 8, 8, 25, 25, 43, 25, 8, 8, 25, 8, 25, 25, 43, 25, 8, 8, 8, 25, 25, 25, 43, 25, 8, 25, 43, 43, 25, 25, 43, 25, 8, 43, 8, 43, 43, 25, 43, 25, 8, 8, 25, 8, 8, 43, 43, 25, 8, 8, 8, 25, 8, 43, 43, 25, 8, 8, 8, 8, 25, 43, 43, 25, 8, 43, 25, 25, 25, 43, 43, 25, 8, 8, 8, 8, 8, 8, 8, 43, 8, 43, 8, 8, 8, 8, 8, 43, 8, 25, 25, 8, 8, 8, 8, 43, 8, 8, 43, 8, 8, 8, 8, 43, 8, 25, 8, 25, 8, 8, 8, 43, 8, 8, 25, 25, 8, 8, 8, 43, 8, 43, 25, 25, 8, 8, 8, 43, 8, 25, 43, 25, 8, 8, 8, 43, 8, 8, 8, 43, 8, 8, 8, 43, 8, 25, 25, 43, 8, 8, 8, 43, 8, 43, 43, 43, 8, 8, 8, 43, 8, 25, 8, 8, 25, 8, 8, 43, 8, 8, 25, 8, 25, 8, 8, 43, 8, 8, 8, 25, 25, 8, 8, 43, 8, 43, 8, 25, 25, 8, 8, 43, 8, 25, 25, 25, 25, 8, 8, 43, 8, 8, 25, 43, 25, 8, 8, 43, 8, 8, 8, 8, 43, 8, 8, 43, 8, 43, 43, 8, 43, 8, 8, 43, 8, 8, 25, 25, 43, 8, 8, 43, 8, 43, 43, 43, 43, 8, 8, 43, 8, 25, 8, 8, 8, 25, 8, 43, 8, 8, 25, 8, 8, 25, 8, 43, 8, 8, 8, 25, 8, 25, 8, 43, 8, 43, 8, 25, 8, 25, 8, 43, 8, 25, 25, 25, 8, 25, 8, 43, 8, 25, 8, 43, 8, 25, 8, 43, 8, 8, 8, 8, 25, 25, 8, 43, 8, 43, 8, 8, 25, 25, 8, 43, 8, 25, 25, 8, 25, 25, 8, 43, 8, 25, 8, 25, 25, 25, 8, 43, 8, 8, 25, 25, 25, 25, 8, 43, 8, 8, 8, 43, 25, 25, 8, 43, 8, 25, 8, 8, 43, 25, 8, 43, 8, 8, 25, 8, 43, 25, 8, 43, 8, 8, 8, 25, 43, 25, 8, 43, 8, 8, 8, 8, 8, 43, 8, 43, 8, 43, 43, 8, 8, 43, 8, 43, 8, 43, 8, 43, 8, 43, 8, 43, 8, 8, 43, 43, 8, 43, 8, 43, 8, 43, 43, 43, 8, 43, 8, 43, 8, 8, 25, 8, 25, 43, 8, 43, 8, 8, 8, 25, 25, 43, 8, 43, 8, 8, 43, 8, 43, 43, 8, 43, 8, 43, 43, 8, 43, 43, 8, 43, 8, 8, 43, 43, 43, 43, 8, 43, 8, 25, 8, 8, 8, 8, 25, 43, 8, 8, 25, 8, 8, 8, 25, 43, 8, 43, 25, 8, 8, 8, 25, 43, 8, 25, 43, 8, 8, 8, 25, 43, 8, 8, 8, 25, 8, 8, 25, 43, 8, 25, 25, 25, 8, 8, 25, 43, 8, 8, 43, 25, 8, 8, 25, 43, 8, 25, 8, 43, 8, 8, 25, 43, 8, 8, 25, 43, 8, 8, 25, 43, 8, 8, 8, 8, 25, 8, 25, 43, 8, 43, 8, 8, 25, 8, 25, 43, 8, 25, 25, 8, 25, 8, 25, 43, 8, 8, 43, 8, 25, 8, 25, 43, 8, 25, 8, 25, 25, 8, 25, 43, 8, 8, 25, 25, 25, 8, 25, 43, 8, 8, 8, 43, 25, 8, 25, 43, 8, 25, 8, 8, 43, 8, 25, 43, 8, 8, 25, 8, 43, 8, 25, 43, 8, 8, 8, 25, 43, 8, 25, 43, 8, 8, 8, 8, 8, 25, 25, 43, 8, 25, 25, 8, 8, 25, 25, 43, 8, 8, 43, 8, 8, 25, 25, 43, 8, 25, 8, 25, 8, 25, 25, 43, 8, 8, 25, 25, 8, 25, 25, 43, 8, 8, 8, 43, 8, 25, 25, 43, 8, 25, 8, 8, 25, 25, 25, 43, 8, 8, 25, 8, 25, 25, 25, 43, 8, 8, 8, 25, 25, 25, 25, 43, 8, 43, 25, 43, 25, 25, 25, 43, 8, 8, 8, 8, 43, 25, 25, 43, 8, 25, 8, 8, 8, 43, 25, 43, 8, 8, 25, 8, 8, 43, 25, 43, 8, 8, 8, 25, 8, 43, 25, 43, 8, 8, 8, 8, 25, 43, 25, 43, 8, 25, 43, 25, 25, 43, 25, 43, 8, 8, 8, 8, 8, 8, 43, 43, 8, 25, 25, 8, 8, 8, 43, 43, 8, 25, 8, 25, 8, 8, 43, 43, 8, 8, 25, 25, 8, 8, 43, 43, 8, 25, 8, 8, 25, 8, 43, 43, 8, 8, 25, 8, 25, 8, 43, 43, 8, 8, 8, 25, 25, 8, 43, 43, 8, 43, 43, 8, 43, 8, 43, 43, 8, 43, 43, 43, 43, 8, 43, 43, 8, 25, 8, 8, 8, 25, 43, 43, 8, 8, 25, 8, 8, 25, 43, 43, 8, 8, 8, 25, 8, 25, 43, 43, 8, 25, 25, 25, 43, 25, 43, 43, 8, 43, 43, 8, 8, 43, 43, 43, 8, 43, 8, 43, 8, 43, 43, 43, 8, 8, 25, 43, 25, 43, 43, 43, 8, 8, 43, 8, 43, 43, 43, 43, 8, 43, 43, 8, 43, 43, 43, 43, 8, 25, 8, 8, 8, 8, 8, 8, 25, 8, 25, 8, 8, 8, 8, 8, 25, 43, 25, 8, 8, 8, 8, 8, 25, 25, 43, 8, 8, 8, 8, 8, 25, 8, 8, 25, 8, 8, 8, 8, 25, 43, 8, 25, 8, 8, 8, 8, 25, 25, 25, 25, 8, 8, 8, 8, 25, 8, 43, 25, 8, 8, 8, 8, 25, 43, 43, 25, 8, 8, 8, 8, 25, 25, 8, 43, 8, 8, 8, 8, 25, 8, 25, 43, 8, 8, 8, 8, 25, 43, 25, 43, 8, 8, 8, 8, 25, 8, 8, 8, 25, 8, 8, 8, 25, 43, 8, 8, 25, 8, 8, 8, 25, 25, 25, 8, 25, 8, 8, 8, 25, 8, 43, 8, 25, 8, 8, 8, 25, 43, 43, 8, 25, 8, 8, 8, 25, 25, 8, 25, 25, 8, 8, 8, 25, 8, 25, 25, 25, 8, 8, 8, 25, 43, 25, 25, 25, 8, 8, 8, 25, 25, 43, 25, 25, 8, 8, 8, 25, 8, 8, 43, 25, 8, 8, 8, 25, 43, 8, 43, 25, 8, 8, 8, 25, 25, 25, 43, 25, 8, 8, 8, 25, 25, 8, 8, 43, 8, 8, 8, 25, 8, 25, 8, 43, 8, 8, 8, 25, 8, 8, 25, 43, 8, 8, 8, 25, 25, 25, 25, 43, 8, 8, 8, 25, 8, 43, 25, 43, 8, 8, 8, 25, 25, 8, 43, 43, 8, 8, 8, 25, 8, 25, 43, 43, 8, 8, 8, 25, 8, 8, 8, 8, 25, 8, 8, 25, 43, 8, 8, 8, 25, 8, 8, 25, 25, 25, 8, 8, 25, 8, 8, 25, 8, 43, 8, 8, 25, 8, 8, 25, 25, 8, 25, 8, 25, 8, 8, 25, 8, 25, 25, 8, 25, 8, 8, 25, 43, 25, 25, 8, 25, 8, 8, 25, 25, 43, 25, 8, 25, 8, 8, 25, 8, 8, 43, 8, 25, 8, 8, 25, 43, 8, 43, 8, 25, 8, 8, 25, 25, 25, 43, 8, 25, 8, 8, 25, 25, 8, 8, 25, 25, 8, 8, 25, 8, 25, 8, 25, 25, 8, 8, 25, 43, 25, 8, 25, 25, 8, 8, 25, 25, 43, 8, 25, 25, 8, 8, 25, 8, 8, 25, 25, 25, 8, 8, 25, 43, 8, 25, 25, 25, 8, 8, 25, 25, 25, 25, 25, 25, 8, 8, 25, 8, 43, 25, 25, 25, 8, 8, 25, 25, 8, 43, 25, 25, 8, 8, 25, 8, 25, 43, 25, 25, 8, 8, 25, 8, 8, 8, 43, 25, 8, 8, 25, 43, 8, 8, 43, 25, 8, 8, 25, 25, 25, 8, 43, 25, 8, 8, 25, 8, 43, 8, 43, 25, 8, 8, 25, 25, 8, 25, 43, 25, 8, 8, 25, 8, 25, 25, 43, 25, 8, 8, 25, 8, 8, 43, 43, 25, 8, 8, 25, 25, 8, 8, 8, 43, 8, 8, 25, 8, 25, 8, 8, 43, 8, 8, 25, 8, 8, 25, 8, 43, 8, 8, 25, 43, 8, 25, 8, 43, 8, 8, 25, 25, 25, 25, 8, 43, 8, 8, 25, 8, 43, 25, 8, 43, 8, 8, 25, 8, 25, 43, 8, 43, 8, 8, 25, 8, 8, 8, 25, 43, 8, 8, 25, 25, 25, 8, 25, 43, 8, 8, 25, 8, 43, 8, 25, 43, 8, 8, 25, 25, 8, 25, 25, 43, 8, 8, 25, 8, 25, 25, 25, 43, 8, 8, 25, 8, 8, 43, 25, 43, 8, 8, 25, 25, 8, 8, 43, 43, 8, 8, 25, 8, 25, 8, 43, 43, 8, 8, 25, 8, 8, 8, 8, 8, 25, 8, 25, 43, 8, 8, 8, 8, 25, 8, 25, 25, 25, 8, 8, 8, 25, 8, 25, 8, 43, 8, 8, 8, 25, 8, 25, 43, 43, 8, 8, 8, 25, 8, 25, 25, 8, 25, 8, 8, 25, 8, 25, 8, 25, 25, 8, 8, 25, 8, 25, 43, 25, 25, 8, 8, 25, 8, 25, 25, 43, 25, 8, 8, 25, 8, 25, 8, 8, 43, 8, 8, 25, 8, 25, 43, 8, 43, 8, 8, 25, 8, 25, 25, 25, 43, 8, 8, 25, 8, 25, 8, 43, 43, 8, 8, 25, 8, 25, 25, 8, 8, 25, 8, 25, 8, 25, 8, 25, 8, 25, 8, 25, 8, 25, 43, 25, 8, 25, 8, 25, 8, 25, 25, 43, 8, 25, 8, 25, 8, 25, 8, 8, 25, 25, 8, 25, 8, 25, 43, 8, 25, 25, 8, 25, 8, 25, 25, 25, 25, 25, 8, 25, 8, 25, 8, 43, 25, 25, 8, 25, 8, 25, 25, 8, 43, 25, 8, 25, 8, 25, 8, 25, 43, 25, 8, 25, 8, 25, 8, 8, 8, 43, 8, 25, 8, 25, 43, 8, 8, 43, 8, 25, 8, 25, 25, 25, 8, 43, 8, 25, 8, 25, 8, 43, 8, 43, 8, 25, 8, 25, 25, 8, 25, 43, 8, 25, 8, 25, 8, 25, 25, 43, 8, 25, 8, 25, 8, 8, 43, 43, 8, 25, 8, 25, 25, 8, 8, 8, 25, 25, 8, 25, 8, 25, 8, 8, 25, 25, 8, 25, 43, 25, 8, 8, 25, 25, 8, 25, 25, 43, 8, 8, 25, 25, 8, 25, 8, 8, 25, 8, 25, 25, 8, 25, 43, 8, 25, 8, 25, 25, 8, 25, 25, 25, 25, 8, 25, 25, 8, 25, 8, 43, 25, 8, 25, 25, 8, 25, 25, 8, 43, 8, 25, 25, 8, 25, 8, 25, 43, 8, 25, 25, 8, 25, 8, 8, 8, 25, 25, 25, 8, 25, 43, 8, 8, 25, 25, 25, 8, 25, 25, 25, 8, 25, 25, 25, 8, 25, 8, 43, 8, 25, 25, 25, 8, 25, 25, 8, 25, 25, 25, 25, 8, 25, 8, 25, 25, 25, 25, 25, 8, 25, 8, 8, 43, 25, 25, 25, 8, 25, 43, 43, 43, 25, 25, 25, 8, 25, 25, 8, 8, 43, 25, 25, 8, 25, 8, 25, 8, 43, 25, 25, 8, 25, 8, 8, 25, 43, 25, 25, 8, 25, 8, 8, 8, 8, 43, 25, 8, 25, 43, 8, 8, 8, 43, 25, 8, 25, 25, 25, 8, 8, 43, 25, 8, 25, 8, 43, 8, 8, 43, 25, 8, 25, 25, 8, 25, 8, 43, 25, 8, 25, 8, 25, 25, 8, 43, 25, 8, 25, 8, 8, 43, 8, 43, 25, 8, 25, 25, 8, 8, 25, 43, 25, 8, 25, 8, 25, 8, 25, 43, 25, 8, 25, 8, 8, 25, 25, 43, 25, 8, 25, 8, 8, 8, 43, 43, 25, 8, 25, 25, 25, 43, 43, 43, 25, 8, 25, 25, 8, 8, 8, 8, 43, 8, 25, 8, 25, 8, 8, 8, 43, 8, 25, 25, 43, 8, 8, 8, 43, 8, 25, 8, 8, 25, 8, 8, 43, 8, 25, 43, 8, 25, 8, 8, 43, 8, 25, 25, 25, 25, 8, 8, 43, 8, 25, 8, 43, 25, 8, 8, 43, 8, 25, 25, 8, 43, 8, 8, 43, 8, 25, 8, 25, 43, 8, 8, 43, 8, 25, 8, 8, 8, 25, 8, 43, 8, 25, 43, 8, 8, 25, 8, 43, 8, 25, 25, 25, 8, 25, 8, 43, 8, 25, 8, 43, 8, 25, 8, 43, 8, 25, 25, 8, 25, 25, 8, 43, 8, 25, 8, 25, 25, 25, 8, 43, 8, 25, 8, 8, 43, 25, 8, 43, 8, 25, 8, 25, 8, 43, 8, 43, 8, 25, 8, 8, 25, 43, 8, 43, 8, 25, 8, 8, 8, 8, 25, 43, 8, 25, 43, 8, 8, 8, 25, 43, 8, 25, 25, 25, 8, 8, 25, 43, 8, 25, 8, 43, 8, 8, 25, 43, 8, 25, 25, 8, 25, 8, 25, 43, 8, 25, 8, 25, 25, 8, 25, 43, 8, 25, 8, 8, 43, 8, 25, 43, 8, 25, 25, 8, 8, 25, 25, 43, 8, 25, 8, 25, 8, 25, 25, 43, 8, 25, 8, 8, 25, 25, 25, 43, 8, 25, 8, 8, 8, 43, 25, 43, 8, 25, 43, 25, 25, 43, 25, 43, 8, 25, 25, 8, 8, 8, 43, 43, 8, 25, 8, 25, 8, 8, 43, 43, 8, 25, 8, 8, 25, 8, 43, 43, 8, 25, 8, 8, 8, 25, 43, 43, 8, 25, 8, 8, 8, 8, 8, 8, 25, 25, 43, 8, 8, 8, 8, 8, 25, 25, 25, 25, 8, 8, 8, 8, 25, 25, 8, 43, 8, 8, 8, 8, 25, 25, 25, 8, 25, 8, 8, 8, 25, 25, 8, 25, 25, 8, 8, 8, 25, 25, 43, 25, 25, 8, 8, 8, 25, 25, 25, 43, 25, 8, 8, 8, 25, 25, 8, 8, 43, 8, 8, 8, 25, 25, 43, 8, 43, 8, 8, 8, 25, 25, 25, 25, 43, 8, 8, 8, 25, 25, 8, 43, 43, 8, 8, 8, 25, 25, 25, 8, 8, 25, 8, 8, 25, 25, 8, 25, 8, 25, 8, 8, 25, 25, 43, 25, 8, 25, 8, 8, 25, 25, 25, 43, 8, 25, 8, 8, 25, 25, 8, 8, 25, 25, 8, 8, 25, 25, 43, 8, 25, 25, 8, 8, 25, 25, 25, 25, 25, 25, 8, 8, 25, 25, 8, 43, 25, 25, 8, 8, 25, 25, 25, 8, 43, 25, 8, 8, 25, 25, 8, 25, 43, 25, 8, 8, 25, 25, 8, 8, 8, 43, 8, 8, 25, 25, 43, 8, 8, 43, 8, 8, 25, 25, 25, 25, 8, 43, 8, 8, 25, 25, 8, 43, 8, 43, 8, 8, 25, 25, 25, 8, 25, 43, 8, 8, 25, 25, 8, 25, 25, 43, 8, 8, 25, 25, 25, 8, 8, 8, 25, 8, 25, 25, 8, 25, 8, 8, 25, 8, 25, 25, 43, 25, 8, 8, 25, 8, 25, 25, 25, 43, 8, 8, 25, 8, 25, 25, 8, 8, 25, 8, 25, 8, 25, 25, 43, 8, 25, 8, 25, 8, 25, 25, 25, 25, 25, 8, 25, 8, 25, 25, 8, 43, 25, 8, 25, 8, 25, 25, 25, 8, 43, 8, 25, 8, 25, 25, 8, 25, 43, 8, 25, 8, 25, 25, 8, 8, 8, 25, 25, 8, 25, 25, 43, 8, 8, 25, 25, 8, 25, 25, 25, 25, 8, 25, 25, 8, 25, 25, 8, 43, 8, 25, 25, 8, 25, 25, 25, 8, 25, 25, 25, 8, 25, 25, 8, 25, 25, 25, 25, 8, 25, 25, 8, 8, 43, 25, 25, 8, 25, 25, 25, 8, 8, 43, 25, 8, 25, 25, 8, 25, 8, 43, 25, 8, 25, 25, 8, 8, 25, 43, 25, 8, 25, 25, 8, 8, 8, 8, 43, 8, 25, 25, 25, 25, 8, 8, 43, 8, 25, 25, 8, 43, 8, 8, 43, 8, 25, 25, 25, 8, 25, 8, 43, 8, 25, 25, 8, 25, 25, 8, 43, 8, 25, 25, 8, 8, 43, 8, 43, 8, 25, 25, 25, 8, 8, 25, 43, 8, 25, 25, 8, 25, 8, 25, 43, 8, 25, 25, 8, 8, 25, 25, 43, 8, 25, 25, 25, 43, 43, 25, 43, 8, 25, 25, 8, 8, 8, 43, 43, 8, 25, 25, 25, 8, 8, 8, 8, 25, 25, 25, 8, 25, 8, 8, 8, 25, 25, 25, 43, 25, 8, 8, 8, 25, 25, 25, 25, 43, 8, 8, 8, 25, 25, 25, 8, 8, 25, 8, 8, 25, 25, 25, 43, 8, 25, 8, 8, 25, 25, 25, 25, 25, 25, 8, 8, 25, 25, 25, 8, 43, 25, 8, 8, 25, 25, 25, 25, 8, 43, 8, 8, 25, 25, 25, 8, 25, 43, 8, 8, 25, 25, 25, 8, 8, 8, 25, 8, 25, 25, 25, 43, 8, 8, 25, 8, 25, 25, 25, 25, 25, 8, 25, 8, 25, 25, 25, 8, 43, 8, 25, 8, 25, 25, 25, 25, 8, 25, 25, 8, 25, 25, 25, 8, 25, 25, 25, 8, 25, 25, 25, 8, 8, 43, 25, 8, 25, 25, 25, 25, 8, 8, 43, 8, 25, 25, 25, 8, 25, 8, 43, 8, 25, 25, 25, 8, 8, 25, 43, 8, 25, 25, 25, 8, 8, 8, 8, 25, 25, 25, 25, 43, 8, 8, 8, 25, 25, 25, 25, 25, 25, 8, 8, 25, 25, 25, 25, 8, 43, 8, 8, 25, 25, 25, 25, 25, 8, 25, 8, 25, 25, 25, 25, 8, 25, 25, 8, 25, 25, 25, 25, 8, 8, 43, 8, 25, 25, 25, 25, 25, 8, 8, 25, 25, 25, 25, 25, 8, 25, 8, 25, 25, 25, 25, 25, 8, 8, 25, 25, 25, 25, 25, 25, 8, 8, 8, 43, 25, 25, 25, 25, 25, 8, 8, 8, 43, 25, 25, 25, 8, 25, 8, 8, 43, 25, 25, 25, 8, 8, 25, 8, 43, 25, 25, 25, 43, 25, 43, 8, 43, 25, 25, 25, 8, 8, 8, 25, 43, 25, 25, 25, 8, 8, 8, 8, 8, 43, 25, 25, 43, 8, 8, 8, 8, 43, 25, 25, 25, 25, 8, 8, 8, 43, 25, 25, 8, 43, 8, 8, 8, 43, 25, 25, 25, 8, 25, 8, 8, 43, 25, 25, 8, 25, 25, 8, 8, 43, 25, 25, 8, 8, 43, 8, 8, 43, 25, 25, 25, 8, 8, 25, 8, 43, 25, 25, 8, 25, 8, 25, 8, 43, 25, 25, 8, 8, 25, 25, 8, 43, 25, 25, 43, 43, 25, 25, 8, 43, 25, 25, 8, 8, 8, 43, 8, 43, 25, 25, 25, 8, 8, 8, 25, 43, 25, 25, 8, 25, 8, 8, 25, 43, 25, 25, 8, 8, 25, 8, 25, 43, 25, 25, 8, 8, 8, 25, 25, 43, 25, 25, 8, 8, 8, 8, 43, 43, 25, 25, 25, 43, 25, 8, 43, 43, 25, 25, 25, 25, 8, 43, 43, 43, 25, 25, 8, 43, 43, 43, 43, 43, 25, 25, 25, 8, 8, 8, 8, 8, 43, 25, 8, 25, 8, 8, 8, 8, 43, 25, 43, 25, 8, 8, 8, 8, 43, 25, 8, 8, 25, 8, 8, 8, 43, 25, 43, 8, 25, 8, 8, 8, 43, 25, 25, 25, 25, 8, 8, 8, 43, 25, 8, 43, 25, 8, 8, 8, 43, 25, 25, 8, 43, 8, 8, 8, 43, 25, 8, 25, 43, 8, 8, 8, 43, 25, 8, 8, 8, 25, 8, 8, 43, 25, 25, 25, 8, 25, 8, 8, 43, 25, 8, 43, 8, 25, 8, 8, 43, 25, 25, 8, 25, 25, 8, 8, 43, 25, 8, 25, 25, 25, 8, 8, 43, 25, 8, 8, 43, 25, 8, 8, 43, 25, 8, 25, 8, 43, 8, 8, 43, 25, 8, 8, 25, 43, 8, 8, 43, 25, 8, 8, 8, 8, 25, 8, 43, 25, 43, 8, 8, 8, 25, 8, 43, 25, 25, 25, 8, 8, 25, 8, 43, 25, 8, 43, 8, 8, 25, 8, 43, 25, 25, 8, 25, 8, 25, 8, 43, 25, 8, 25, 25, 8, 25, 8, 43, 25, 8, 8, 43, 8, 25, 8, 43, 25, 25, 8, 8, 25, 25, 8, 43, 25, 8, 25, 8, 25, 25, 8, 43, 25, 8, 8, 25, 25, 25, 8, 43, 25, 8, 8, 8, 43, 25, 8, 43, 25, 25, 43, 25, 43, 25, 8, 43, 25, 8, 25, 8, 8, 43, 8, 43, 25, 8, 8, 25, 8, 43, 8, 43, 25, 8, 8, 8, 25, 43, 8, 43, 25, 43, 25, 25, 25, 43, 8, 43, 25, 25, 8, 43, 43, 43, 8, 43, 25, 8, 8, 8, 8, 8, 25, 43, 25, 25, 25, 8, 8, 8, 25, 43, 25, 8, 43, 8, 8, 8, 25, 43, 25, 25, 8, 25, 8, 8, 25, 43, 25, 8, 25, 25, 8, 8, 25, 43, 25, 8, 8, 43, 8, 8, 25, 43, 25, 25, 8, 8, 25, 8, 25, 43, 25, 8, 25, 8, 25, 8, 25, 43, 25, 8, 8, 25, 25, 8, 25, 43, 25, 8, 8, 8, 43, 8, 25, 43, 25, 25, 8, 8, 8, 25, 25, 43, 25, 8, 25, 8, 8, 25, 25, 43, 25, 8, 8, 25, 8, 25, 25, 43, 25, 8, 8, 8, 25, 25, 25, 43, 25, 43, 43, 8, 25, 25, 25, 43, 25, 8, 43, 43, 25, 25, 25, 43, 25, 43, 8, 25, 43, 25, 25, 43, 25, 8, 8, 8, 8, 43, 25, 43, 25, 8, 25, 25, 43, 43, 25, 43, 25, 25, 8, 8, 8, 8, 43, 43, 25, 8, 25, 8, 8, 8, 43, 43, 25, 8, 8, 25, 8, 8, 43, 43, 25, 25, 25, 43, 25, 8, 43, 43, 25, 8, 43, 25, 43, 8, 43, 43, 25, 8, 8, 8, 8, 25, 43, 43, 25, 43, 43, 43, 8, 25, 43, 43, 25, 43, 8, 8, 25, 43, 43, 43, 25, 25, 8, 43, 43, 43, 43, 43, 25, 8, 8, 8, 8, 8, 8, 8, 43, 43, 8, 8, 8, 8, 8, 8, 43, 25, 25, 8, 8, 8, 8, 8, 43, 8, 43, 8, 8, 8, 8, 8, 43, 25, 8, 25, 8, 8, 8, 8, 43, 8, 25, 25, 8, 8, 8, 8, 43, 25, 43, 25, 8, 8, 8, 8, 43, 8, 8, 43, 8, 8, 8, 8, 43, 25, 25, 43, 8, 8, 8, 8, 43, 25, 8, 8, 25, 8, 8, 8, 43, 8, 25, 8, 25, 8, 8, 8, 43, 8, 8, 25, 25, 8, 8, 8, 43, 43, 8, 25, 25, 8, 8, 8, 43, 25, 25, 25, 25, 8, 8, 8, 43, 8, 43, 25, 25, 8, 8, 8, 43, 25, 8, 43, 25, 8, 8, 8, 43, 8, 8, 8, 43, 8, 8, 8, 43, 25, 25, 8, 43, 8, 8, 8, 43, 25, 8, 25, 43, 8, 8, 8, 43, 8, 25, 25, 43, 8, 8, 8, 43, 25, 8, 8, 8, 25, 8, 8, 43, 8, 25, 8, 8, 25, 8, 8, 43, 25, 43, 8, 8, 25, 8, 8, 43, 8, 8, 25, 8, 25, 8, 8, 43, 43, 8, 25, 8, 25, 8, 8, 43, 25, 25, 25, 8, 25, 8, 8, 43, 8, 43, 25, 8, 25, 8, 8, 43, 25, 8, 43, 8, 25, 8, 8, 43, 8, 25, 43, 8, 25, 8, 8, 43, 8, 8, 8, 25, 25, 8, 8, 43, 43, 8, 8, 25, 25, 8, 8, 43, 25, 25, 8, 25, 25, 8, 8, 43, 8, 43, 8, 25, 25, 8, 8, 43, 25, 8, 25, 25, 25, 8, 8, 43, 8, 25, 25, 25, 25, 8, 8, 43, 25, 8, 8, 43, 25, 8, 8, 43, 8, 25, 8, 43, 25, 8, 8, 43, 8, 8, 25, 43, 25, 8, 8, 43, 25, 43, 43, 43, 25, 8, 8, 43, 8, 8, 8, 8, 43, 8, 8, 43, 25, 25, 8, 8, 43, 8, 8, 43, 43, 43, 8, 8, 43, 8, 8, 43, 25, 8, 25, 8, 43, 8, 8, 43, 8, 25, 25, 8, 43, 8, 8, 43, 25, 8, 8, 25, 43, 8, 8, 43, 8, 25, 8, 25, 43, 8, 8, 43, 8, 8, 25, 25, 43, 8, 8, 43, 25, 8, 8, 8, 8, 25, 8, 43, 8, 25, 8, 8, 8, 25, 8, 43, 43, 25, 8, 8, 8, 25, 8, 43, 25, 43, 8, 8, 8, 25, 8, 43, 8, 8, 25, 8, 8, 25, 8, 43, 43, 8, 25, 8, 8, 25, 8, 43, 25, 25, 25, 8, 8, 25, 8, 43, 8, 43, 25, 8, 8, 25, 8, 43, 25, 8, 43, 8, 8, 25, 8, 43, 8, 8, 8, 25, 8, 25, 8, 43, 43, 8, 8, 25, 8, 25, 8, 43, 25, 25, 8, 25, 8, 25, 8, 43, 8, 43, 8, 25, 8, 25, 8, 43, 25, 8, 25, 25, 8, 25, 8, 43, 8, 25, 25, 25, 8, 25, 8, 43, 8, 8, 43, 25, 8, 25, 8, 43, 25, 8, 8, 43, 8, 25, 8, 43, 8, 25, 8, 43, 8, 25, 8, 43, 8, 8, 25, 43, 8, 25, 8, 43, 8, 8, 8, 8, 25, 25, 8, 43, 43, 8, 8, 8, 25, 25, 8, 43, 25, 25, 8, 8, 25, 25, 8, 43, 8, 43, 8, 8, 25, 25, 8, 43, 25, 8, 25, 8, 25, 25, 8, 43, 8, 25, 25, 8, 25, 25, 8, 43, 8, 8, 43, 8, 25, 25, 8, 43, 25, 8, 8, 25, 25, 25, 8, 43, 8, 25, 8, 25, 25, 25, 8, 43, 8, 8, 25, 25, 25, 25, 8, 43, 8, 8, 8, 43, 25, 25, 8, 43, 43, 43, 8, 43, 25, 25, 8, 43, 25, 8, 8, 8, 43, 25, 8, 43, 8, 25, 8, 8, 43, 25, 8, 43, 8, 8, 25, 8, 43, 25, 8, 43, 25, 43, 43, 8, 43, 25, 8, 43, 8, 8, 8, 25, 43, 25, 8, 43, 8, 8, 8, 8, 8, 43, 8, 43, 25, 25, 8, 8, 8, 43, 8, 43, 25, 8, 25, 8, 8, 43, 8, 43, 8, 25, 25, 8, 8, 43, 8, 43, 25, 8, 8, 25, 8, 43, 8, 43, 8, 25, 8, 25, 8, 43, 8, 43, 8, 8, 25, 25, 8, 43, 8, 43, 43, 8, 43, 43, 8, 43, 8, 43, 25, 8, 8, 8, 25, 43, 8, 43, 8, 25, 8, 8, 25, 43, 8, 43, 8, 8, 8, 25, 25, 43, 8, 43, 25, 25, 43, 25, 25, 43, 8, 43, 43, 8, 43, 8, 43, 43, 8, 43, 8, 43, 25, 25, 43, 43, 8, 43, 43, 43, 25, 25, 43, 43, 8, 43, 43, 8, 8, 43, 43, 43, 8, 43, 43, 8, 43, 43, 43, 43, 8, 43, 25, 8, 8, 8, 8, 8, 25, 43, 8, 25, 8, 8, 8, 8, 25, 43, 25, 43, 8, 8, 8, 8, 25, 43, 8, 8, 25, 8, 8, 8, 25, 43, 43, 8, 25, 8, 8, 8, 25, 43, 25, 25, 25, 8, 8, 8, 25, 43, 8, 43, 25, 8, 8, 8, 25, 43, 8, 25, 43, 8, 8, 8, 25, 43, 8, 8, 8, 25, 8, 8, 25, 43, 43, 8, 8, 25, 8, 8, 25, 43, 25, 25, 8, 25, 8, 8, 25, 43, 8, 43, 8, 25, 8, 8, 25, 43, 25, 8, 25, 25, 8, 8, 25, 43, 8, 25, 25, 25, 8, 8, 25, 43, 8, 8, 43, 25, 8, 8, 25, 43, 25, 8, 8, 43, 8, 8, 25, 43, 8, 25, 8, 43, 8, 8, 25, 43, 8, 8, 25, 43, 8, 8, 25, 43, 8, 8, 8, 8, 25, 8, 25, 43, 25, 25, 8, 8, 25, 8, 25, 43, 25, 8, 25, 8, 25, 8, 25, 43, 8, 25, 25, 8, 25, 8, 25, 43, 25, 8, 8, 25, 25, 8, 25, 43, 8, 25, 8, 25, 25, 8, 25, 43, 8, 8, 25, 25, 25, 8, 25, 43, 43, 43, 25, 25, 25, 8, 25, 43, 25, 8, 8, 8, 43, 8, 25, 43, 8, 25, 8, 8, 43, 8, 25, 43, 8, 8, 25, 8, 43, 8, 25, 43, 8, 8, 8, 25, 43, 8, 25, 43, 43, 25, 43, 43, 43, 8, 25, 43, 8, 8, 8, 8, 8, 25, 25, 43, 43, 8, 8, 8, 8, 25, 25, 43, 25, 25, 8, 8, 8, 25, 25, 43, 8, 43, 8, 8, 8, 25, 25, 43, 25, 8, 25, 8, 8, 25, 25, 43, 8, 25, 25, 8, 8, 25, 25, 43, 8, 8, 43, 8, 8, 25, 25, 43, 25, 8, 8, 25, 8, 25, 25, 43, 8, 25, 8, 25, 8, 25, 25, 43, 8, 8, 25, 25, 8, 25, 25, 43, 8, 8, 8, 43, 8, 25, 25, 43, 43, 25, 25, 43, 8, 25, 25, 43, 25, 8, 8, 8, 25, 25, 25, 43, 8, 25, 8, 8, 25, 25, 25, 43, 8, 8, 25, 8, 25, 25, 25, 43, 8, 8, 8, 25, 25, 25, 25, 43, 8, 43, 25, 43, 25, 25, 25, 43, 25, 8, 43, 43, 25, 25, 25, 43, 8, 8, 8, 8, 43, 25, 25, 43, 43, 25, 8, 25, 43, 25, 25, 43, 8, 25, 43, 25, 43, 25, 25, 43, 25, 8, 8, 8, 8, 43, 25, 43, 8, 25, 8, 8, 8, 43, 25, 43, 8, 8, 25, 8, 8, 43, 25, 43, 43, 25, 43, 8, 8, 43, 25, 43, 8, 8, 8, 25, 8, 43, 25, 43, 25, 43, 43, 43, 8, 43, 25, 43, 8, 8, 8, 8, 25, 43, 25, 43, 25, 43, 8, 25, 25, 43, 25, 43, 43, 8, 25, 25, 25, 43, 25, 43, 8, 8, 25, 43, 43, 43, 25, 43, 8, 8, 8, 8, 8, 8, 43, 43, 25, 25, 8, 8, 8, 8, 43, 43, 43, 43, 8, 8, 8, 8, 43, 43, 8, 25, 25, 8, 8, 8, 43, 43, 43, 8, 43, 8, 8, 8, 43, 43, 43, 43, 43, 8, 8, 8, 43, 43, 25, 8, 8, 25, 8, 8, 43, 43, 8, 25, 8, 25, 8, 8, 43, 43, 8, 8, 25, 25, 8, 8, 43, 43, 43, 8, 43, 43, 8, 8, 43, 43, 43, 43, 43, 43, 8, 8, 43, 43, 8, 8, 8, 25, 25, 8, 43, 43, 25, 25, 43, 25, 25, 8, 43, 43, 43, 8, 8, 8, 43, 8, 43, 43, 43, 43, 8, 8, 43, 8, 43, 43, 43, 8, 43, 8, 43, 8, 43, 43, 8, 43, 43, 8, 43, 8, 43, 43, 43, 43, 43, 8, 43, 8, 43, 43, 43, 8, 8, 43, 43, 8, 43, 43, 8, 43, 8, 43, 43, 8, 43, 43, 43, 43, 8, 43, 43, 8, 43, 43, 8, 43, 43, 43, 43, 8, 43, 43, 25, 8, 8, 8, 8, 25, 43, 43, 8, 25, 8, 8, 8, 25, 43, 43, 8, 8, 25, 8, 8, 25, 43, 43, 8, 8, 8, 25, 8, 25, 43, 43, 25, 43, 8, 43, 8, 25, 43, 43, 8, 25, 43, 43, 8, 25, 43, 43, 8, 8, 8, 8, 25, 25, 43, 43, 25, 43, 25, 8, 25, 25, 43, 43, 25, 8, 25, 25, 43, 25, 43, 43, 43, 43, 8, 8, 8, 43, 43, 43, 8, 43, 43, 8, 8, 43, 43, 43, 43, 8, 43, 43, 8, 43, 43, 43, 8, 25, 25, 25, 25, 43, 43, 43, 43, 25, 8, 43, 25, 43, 43, 43, 8, 43, 8, 8, 43, 43, 43, 43, 43, 43, 8, 8, 43, 43, 43, 43, 8, 8, 43, 8, 43, 43, 43, 43, 43, 8, 43, 8, 43, 43, 43, 43, 8, 43, 43, 8, 43, 43, 43, 43, 8, 43, 8, 43, 43, 43, 43, 43, 43, 43, 43, 43, 43, 43, 43, 43};
.global .align 4 .b8 iq3xxs_grid[1024] = {4, 4, 4, 4, 20, 4, 4, 4, 36, 4, 4, 4, 12, 12, 4, 4, 28, 12, 4, 4, 62, 12, 4, 4, 4, 20, 4, 4, 20, 20, 4, 4, 12, 28, 4, 4, 20, 36, 4, 4, 28, 62, 4, 4, 44, 62, 4, 4, 12, 4, 12, 4, 28, 4, 12, 4, 4, 12, 12, 4, 20, 12, 12, 4, 12, 20, 12, 4, 44, 20, 12, 4, 4, 28, 12, 4, 20, 28, 12, 4, 12, 36, 12, 4, 36, 44, 12, 4, 4, 62, 12, 4, 4, 4, 20, 4, 20, 4, 20, 4, 36, 4, 20, 4, 12, 12, 20, 4, 4, 20, 20, 4, 20, 20, 20, 4, 12, 28, 20, 4, 28, 28, 20, 4, 62, 28, 20, 4, 12, 44, 20, 4, 62, 44, 20, 4, 44, 62, 20, 4, 12, 4, 28, 4, 62, 4, 28, 4, 4, 12, 28, 4, 20, 12, 28, 4, 44, 20, 28, 4, 4, 62, 28, 4, 28, 12, 36, 4, 62, 28, 36, 4, 36, 36, 36, 4, 62, 44, 36, 4, 28, 62, 36, 4, 44, 62, 36, 4, 12, 4, 44, 4, 62, 4, 44, 4, 20, 28, 44, 4, 20, 44, 44, 4, 44, 28, 52, 4, 36, 52, 52, 4, 4, 12, 62, 4, 36, 12, 62, 4, 52, 12, 62, 4, 28, 36, 62, 4, 12, 52, 62, 4, 12, 4, 4, 12, 28, 4, 4, 12, 4, 12, 4, 12, 20, 12, 4, 12, 12, 20, 4, 12, 28, 20, 4, 12, 4, 28, 4, 12, 20, 28, 4, 12, 36, 28, 4, 12, 62, 36, 4, 12, 4, 44, 4, 12, 4, 4, 12, 12, 20, 4, 12, 12, 12, 12, 12, 12, 4, 20, 12, 12, 20, 20, 12, 12, 12, 4, 20, 12, 28, 4, 20, 12, 4, 12, 20, 12, 20, 12, 20, 12, 12, 20, 20, 12, 4, 28, 20, 12, 20, 62, 20, 12, 4, 4, 28, 12, 20, 4, 28, 12, 4, 20, 28, 12, 12, 28, 28, 12, 52, 36, 28, 12, 52, 52, 28, 12, 12, 4, 36, 12, 44, 4, 36, 12, 4, 44, 36, 12, 4, 20, 44, 12, 36, 20, 44, 12, 52, 36, 44, 12, 12, 62, 44, 12, 44, 4, 52, 12, 20, 20, 62, 12, 4, 36, 62, 12, 4, 4, 4, 20, 20, 4, 4, 20, 12, 12, 4, 20, 28, 12, 4, 20, 4, 20, 4, 20, 20, 20, 4, 20, 52, 20, 4, 20, 12, 28, 4, 20, 20, 36, 4, 20, 12, 4, 12, 20, 28, 4, 12, 20, 44, 4, 12, 20, 4, 12, 12, 20, 20, 12, 12, 20, 12, 20, 12, 20, 4, 28, 12, 20, 28, 52, 12, 20, 62, 52, 12, 20, 4, 62, 12, 20, 4, 4, 20, 20, 20, 4, 20, 20, 12, 12, 20, 20, 62, 12, 20, 20, 4, 20, 20, 20, 20, 20, 20, 20, 62, 28, 20, 20, 4, 36, 20, 20, 44, 44, 20, 20, 12, 4, 28, 20, 4, 12, 28, 20, 36, 12, 28, 20, 4, 62, 28, 20, 36, 62, 28, 20, 44, 28, 36, 20, 28, 44, 36, 20, 28, 4, 44, 20, 62, 20, 44, 20, 12, 36, 44, 20, 36, 62, 44, 20, 12, 4, 62, 20, 28, 4, 62, 20, 52, 12, 62, 20, 44, 36, 62, 20, 12, 4, 4, 28, 4, 12, 4, 28, 20, 12, 4, 28, 12, 20, 4, 28, 28, 20, 4, 28, 4, 44, 4, 28, 44, 52, 4, 28, 20, 62, 4, 28, 4, 4, 12, 28, 20, 4, 12, 28, 4, 20, 12, 28, 12, 28, 12, 28, 36, 36, 12, 28, 52, 36, 12, 28, 12, 4, 20, 28, 28, 4, 20, 28, 4, 12, 20, 28, 44, 20, 20, 28, 20, 44, 20, 28, 20, 62, 20, 28, 12, 12, 28, 28, 28, 28, 28, 28, 4, 28, 36, 28, 62, 36, 36, 28, 20, 62, 36, 28, 4, 4, 44, 28, 52, 4, 44, 28, 20, 20, 44, 28, 44, 44, 44, 28, 36, 12, 52, 28, 52, 28, 52, 28, 28, 52, 52, 28, 28, 28, 62, 28, 4, 52, 62, 28, 36, 4, 4, 36, 62, 12, 4, 36, 44, 28, 4, 36, 62, 28, 4, 36, 28, 44, 4, 36, 62, 44, 4, 36, 36, 62, 12, 36, 4, 20, 20, 36, 62, 28, 20, 36, 4, 36, 20, 36, 4, 52, 20, 36, 52, 52, 20, 36, 62, 4, 28, 36, 44, 36, 28, 36, 36, 4, 36, 36, 12, 44, 36, 36, 36, 52, 36, 36, 44, 20, 44, 36, 28, 36, 44, 36, 4, 62, 44, 36, 44, 4, 62, 36, 4, 12, 62, 36, 20, 12, 62, 36, 4, 28, 62, 36, 20, 12, 4, 44, 12, 36, 4, 44, 4, 62, 4, 44, 4, 4, 12, 44, 52, 4, 12, 44, 52, 20, 12, 44, 44, 44, 12, 44, 36, 12, 20, 44, 20, 28, 20, 44, 20, 62, 20, 44, 20, 4, 28, 44, 28, 44, 28, 44, 4, 12, 36, 44, 28, 20, 36, 44, 62, 20, 36, 44, 20, 62, 36, 44, 20, 4, 44, 44, 12, 28, 44, 44, 4, 44, 52, 44, 36, 20, 62, 44, 20, 36, 62, 44, 36, 20, 4, 52, 36, 36, 4, 52, 52, 36, 4, 52, 36, 52, 4, 52, 12, 20, 12, 52, 12, 52, 12, 52, 62, 12, 20, 52, 36, 52, 20, 52, 4, 28, 28, 52, 52, 28, 28, 52, 36, 36, 36, 52, 44, 4, 44, 52, 20, 44, 44, 52, 28, 28, 52, 52, 28, 4, 62, 52, 12, 20, 62, 52, 28, 4, 4, 62, 44, 4, 4, 62, 62, 4, 4, 62, 4, 12, 4, 62, 20, 28, 4, 62, 20, 44, 4, 62, 52, 20, 12, 62, 4, 36, 12, 62, 20, 12, 20, 62, 44, 36, 20, 62, 20, 44, 20, 62, 4, 4, 28, 62, 44, 12, 28, 62, 28, 28, 28, 62, 4, 52, 28, 62, 12, 20, 36, 62, 12, 36, 36, 62, 4, 4, 44, 62, 20, 4, 44, 62, 36, 20, 44, 62, 4, 28, 52, 62};
.global .align 4 .b8 iq3s_grid[2048] = {1, 1, 1, 1, 3, 1, 1, 1, 5, 1, 1, 1, 11, 1, 1, 1, 15, 1, 1, 1, 1, 3, 1, 1, 3, 3, 1, 1, 5, 3, 1, 1, 9, 3, 1, 1, 13, 3, 1, 1, 1, 5, 1, 1, 3, 5, 1, 1, 11, 5, 1, 1, 7, 7, 1, 1, 1, 9, 1, 1, 5, 9, 1, 1, 11, 9, 1, 1, 15, 9, 1, 1, 3, 11, 1, 1, 7, 11, 1, 1, 1, 13, 1, 1, 5, 13, 1, 1, 3, 15, 1, 1, 9, 15, 1, 1, 15, 15, 1, 1, 1, 1, 3, 1, 3, 1, 3, 1, 5, 1, 3, 1, 9, 1, 3, 1, 1, 3, 3, 1, 3, 3, 3, 1, 11, 3, 3, 1, 1, 5, 3, 1, 7, 5, 3, 1, 15, 5, 3, 1, 3, 7, 3, 1, 11, 7, 3, 1, 9, 9, 3, 1, 3, 13, 3, 1, 11, 13, 3, 1, 5, 15, 3, 1, 1, 1, 5, 1, 3, 1, 5, 1, 11, 1, 5, 1, 15, 1, 5, 1, 1, 3, 5, 1, 7, 3, 5, 1, 13, 3, 5, 1, 3, 5, 5, 1, 11, 5, 5, 1, 1, 7, 5, 1, 9, 7, 5, 1, 5, 9, 5, 1, 11, 9, 5, 1, 15, 9, 5, 1, 3, 11, 5, 1, 7, 11, 5, 1, 1, 15, 5, 1, 7, 15, 5, 1, 7, 1, 7, 1, 3, 3, 7, 1, 11, 3, 7, 1, 1, 5, 7, 1, 5, 5, 7, 1, 3, 7, 7, 1, 7, 7, 7, 1, 13, 7, 7, 1, 9, 9, 7, 1, 1, 11, 7, 1, 5, 11, 7, 1, 15, 13, 7, 1, 3, 15, 7, 1, 11, 15, 7, 1, 1, 1, 9, 1, 7, 3, 9, 1, 15, 3, 9, 1, 3, 5, 9, 1, 9, 5, 9, 1, 5, 7, 9, 1, 1, 9, 9, 1, 7, 9, 9, 1, 3, 11, 9, 1, 1, 15, 9, 1, 5, 1, 11, 1, 9, 1, 11, 1, 1, 5, 11, 1, 5, 5, 11, 1, 13, 5, 11, 1, 7, 7, 11, 1, 3, 9, 11, 1, 11, 9, 11, 1, 15, 9, 11, 1, 13, 13, 11, 1, 7, 15, 11, 1, 13, 1, 13, 1, 3, 3, 13, 1, 7, 3, 13, 1, 3, 7, 13, 1, 5, 11, 13, 1, 3, 15, 13, 1, 1, 1, 15, 1, 5, 1, 15, 1, 9, 1, 15, 1, 1, 5, 15, 1, 5, 5, 15, 1, 13, 5, 15, 1, 7, 7, 15, 1, 1, 11, 15, 1, 9, 11, 15, 1, 1, 1, 1, 3, 3, 1, 1, 3, 5, 1, 1, 3, 9, 1, 1, 3, 1, 3, 1, 3, 3, 3, 1, 3, 7, 3, 1, 3, 11, 3, 1, 3, 15, 3, 1, 3, 1, 5, 1, 3, 5, 5, 1, 3, 3, 7, 1, 3, 9, 7, 1, 3, 13, 7, 1, 3, 9, 11, 1, 3, 13, 11, 1, 3, 3, 13, 1, 3, 5, 15, 1, 3, 1, 1, 3, 3, 3, 1, 3, 3, 7, 1, 3, 3, 13, 1, 3, 3, 1, 3, 3, 3, 9, 3, 3, 3, 3, 5, 3, 3, 1, 7, 3, 3, 7, 7, 3, 3, 3, 9, 3, 3, 1, 11, 3, 3, 5, 11, 3, 3, 1, 15, 3, 3, 13, 15, 3, 3, 1, 1, 5, 3, 5, 3, 5, 3, 11, 3, 5, 3, 15, 3, 5, 3, 1, 5, 5, 3, 9, 5, 5, 3, 5, 7, 5, 3, 1, 9, 5, 3, 7, 9, 5, 3, 11, 11, 5, 3, 1, 13, 5, 3, 5, 15, 5, 3, 3, 1, 7, 3, 9, 1, 7, 3, 15, 1, 7, 3, 1, 3, 7, 3, 7, 3, 7, 3, 3, 5, 7, 3, 15, 5, 7, 3, 1, 7, 7, 3, 9, 7, 7, 3, 3, 9, 7, 3, 5, 13, 7, 3, 1, 15, 7, 3, 7, 1, 9, 3, 11, 1, 9, 3, 5, 3, 9, 3, 9, 3, 9, 3, 3, 7, 9, 3, 7, 7, 9, 3, 5, 9, 9, 3, 13, 9, 9, 3, 1, 11, 9, 3, 9, 11, 9, 3, 3, 1, 11, 3, 1, 3, 11, 3, 7, 3, 11, 3, 3, 5, 11, 3, 1, 7, 11, 3, 5, 7, 11, 3, 3, 11, 11, 3, 1, 5, 13, 3, 9, 5, 13, 3, 15, 5, 13, 3, 9, 9, 13, 3, 13, 9, 13, 3, 3, 1, 15, 3, 7, 1, 15, 3, 1, 3, 15, 3, 5, 3, 15, 3, 3, 5, 15, 3, 11, 7, 15, 3, 3, 9, 15, 3, 5, 13, 15, 3, 1, 15, 15, 3, 1, 1, 1, 5, 3, 1, 1, 5, 7, 1, 1, 5, 11, 1, 1, 5, 15, 1, 1, 5, 1, 3, 1, 5, 5, 3, 1, 5, 9, 3, 1, 5, 13, 3, 1, 5, 3, 5, 1, 5, 7, 5, 1, 5, 15, 5, 1, 5, 1, 7, 1, 5, 5, 7, 1, 5, 3, 9, 1, 5, 7, 9, 1, 5, 11, 9, 1, 5, 1, 11, 1, 5, 5, 11, 1, 5, 15, 13, 1, 5, 1, 15, 1, 5, 7, 15, 1, 5, 11, 15, 1, 5, 1, 1, 3, 5, 5, 1, 3, 5, 1, 3, 3, 5, 7, 3, 3, 5, 15, 3, 3, 5, 5, 5, 3, 5, 11, 5, 3, 5, 3, 7, 3, 5, 9, 7, 3, 5, 5, 9, 3, 5, 3, 11, 3, 5, 3, 1, 5, 5, 9, 1, 5, 5, 15, 1, 5, 5, 3, 5, 5, 5, 7, 5, 5, 5, 1, 7, 5, 5, 15, 7, 5, 5, 3, 9, 5, 5, 7, 11, 5, 5, 15, 11, 5, 5, 3, 15, 5, 5, 9, 15, 5, 5, 1, 1, 7, 5, 5, 1, 7, 5, 11, 1, 7, 5, 3, 3, 7, 5, 5, 5, 7, 5, 9, 5, 7, 5, 3, 7, 7, 5, 7, 7, 7, 5, 5, 9, 7, 5, 1, 11, 7, 5, 13, 13, 7, 5, 3, 1, 9, 5, 15, 1, 9, 5, 1, 5, 9, 5, 7, 5, 9, 5, 5, 7, 9, 5, 11, 7, 9, 5, 3, 9, 9, 5, 5, 15, 9, 5, 11, 15, 9, 5, 9, 1, 11, 5, 3, 3, 11, 5, 5, 5, 11, 5, 15, 7, 11, 5, 1, 9, 11, 5, 7, 11, 11, 5, 1, 15, 11, 5, 1, 1, 13, 5, 5, 1, 13, 5, 15, 1, 13, 5, 3, 5, 13, 5, 11, 11, 13, 5, 3, 13, 13, 5, 11, 1, 15, 5, 3, 3, 15, 5, 13, 5, 15, 5, 1, 7, 15, 5, 7, 9, 15, 5, 1, 11, 15, 5, 5, 1, 1, 7, 3, 3, 1, 7, 7, 3, 1, 7, 11, 3, 1, 7, 15, 3, 1, 7, 5, 5, 1, 7, 3, 7, 1, 7, 7, 7, 1, 7, 11, 7, 1, 7, 5, 9, 1, 7, 9, 9, 1, 7, 15, 9, 1, 7, 3, 11, 1, 7, 7, 13, 1, 7, 3, 15, 1, 7, 3, 1, 3, 7, 7, 1, 3, 7, 11, 1, 3, 7, 9, 3, 3, 7, 3, 5, 3, 7, 7, 5, 3, 7, 1, 9, 3, 7, 1, 13, 3, 7, 5, 15, 3, 7, 13, 15, 3, 7, 1, 1, 5, 7, 5, 3, 5, 7, 1, 5, 5, 7, 5, 7, 5, 7, 9, 7, 5, 7, 1, 11, 5, 7, 3, 1, 7, 7, 1, 3, 7, 7, 9, 3, 7, 7, 3, 5, 7, 7, 7, 5, 7, 7, 15, 5, 7, 7, 1, 7, 7, 7, 3, 9, 7, 7, 7, 9, 7, 7, 15, 9, 7, 7, 11, 11, 7, 7, 7, 15, 7, 7, 7, 1, 9, 7, 3, 3, 9, 7, 13, 3, 9, 7, 5, 5, 9, 7, 3, 7, 9, 7, 5, 11, 9, 7, 1, 13, 9, 7, 9, 13, 9, 7, 3, 1, 11, 7, 1, 3, 11, 7, 5, 3, 11, 7, 11, 5, 11, 7, 5, 7, 11, 7, 9, 9, 11, 7, 13, 11, 11, 7, 7, 15, 11, 7, 13, 3, 13, 7, 3, 9, 13, 7, 3, 1, 15, 7, 7, 1, 15, 7, 1, 5, 15, 7, 5, 5, 15, 7, 11, 7, 15, 7, 1, 1, 1, 9, 9, 1, 1, 9, 5, 3, 1, 9, 1, 5, 1, 9, 9, 5, 1, 9, 15, 5, 1, 9, 5, 7, 1, 9, 3, 9, 1, 9, 1, 11, 1, 9, 1, 15, 1, 9, 5, 1, 3, 9, 15, 1, 3, 9, 3, 3, 3, 9, 7, 3, 3, 9, 5, 5, 3, 9, 1, 7, 3, 9, 11, 7, 3, 9, 7, 9, 3, 9, 3, 11, 3, 9, 11, 11, 3, 9, 3, 1, 5, 9, 7, 1, 5, 9, 1, 3, 5, 9, 11, 3, 5, 9, 3, 5, 5, 9, 7, 7, 5, 9, 1, 9, 5, 9, 15, 11, 5, 9, 5, 13, 5, 9, 1, 15, 5, 9, 9, 1, 7, 9, 3, 3, 7, 9, 7, 3, 7, 9, 1, 5, 7, 9, 5, 5, 7, 9, 3, 7, 7, 9, 11, 7, 7, 9, 1, 1, 9, 9, 5, 1, 9, 9, 9, 5, 9, 9, 15, 7, 9, 9, 1, 9, 9, 9, 3, 15, 9, 9, 11, 1, 11, 9, 15, 1, 11, 9, 3, 5, 11, 9, 5, 13, 11, 9, 7, 3, 13, 9, 9, 7, 13, 9, 1, 13, 13, 9, 1, 3, 15, 9, 11, 3, 15, 9, 1, 7, 15, 9, 7, 9, 15, 9, 3, 11, 15, 9, 5, 1, 1, 11, 1, 3, 1, 11, 9, 3, 1, 11, 5, 5, 1, 11, 1, 9, 1, 11, 9, 9, 1, 11, 15, 9, 1, 11, 5, 11, 1, 11, 13, 13, 1, 11, 9, 15, 1, 11, 3, 1, 3, 11, 7, 1, 3, 11, 11, 1, 3, 11, 5, 3, 3, 11, 3, 5, 3, 11, 5, 7, 3, 11, 5, 15, 3, 11, 1, 1, 5, 11, 3, 3, 5, 11, 7, 5, 5, 11, 1, 7, 5, 11, 13, 7, 5, 11, 7, 11, 5, 11, 5, 1, 7, 11, 15, 1, 7, 11, 1, 3, 7, 11, 15, 5, 7, 11, 9, 9, 7, 11, 3, 11, 7, 11, 11, 13, 7, 11, 7, 15, 7, 11, 3, 1, 9, 11, 9, 1, 9, 11, 1, 5, 9, 11, 5, 7, 9, 11, 13, 9, 9, 11, 5, 3, 11, 11, 13, 5, 11, 11, 3, 11, 11, 11, 7, 11, 11, 11, 5, 9, 13, 11, 5, 1, 15, 11, 9, 1, 15, 11, 5, 5, 15, 11, 3, 3, 1, 13, 7, 3, 1, 13, 11, 3, 1, 13, 3, 7, 1, 13, 7, 7, 1, 13, 1, 13, 1, 13, 1, 1, 3, 13, 1, 5, 3, 13, 15, 5, 3, 13, 9, 13, 3, 13, 5, 3, 5, 13, 9, 7, 5, 13, 5, 9, 5, 13, 11, 11, 5, 13, 5, 13, 5, 13, 1, 15, 5, 13, 1, 1, 7, 13, 9, 3, 7, 13, 3, 5, 7, 13, 1, 9, 7, 13, 11, 5, 9, 13, 7, 9, 9, 13, 5, 13, 9, 13, 1, 1, 11, 13, 7, 1, 11, 13, 9, 7, 11, 13, 1, 13, 11, 13, 11, 1, 13, 13, 1, 9, 13, 13, 3, 3, 15, 13, 7, 3, 15, 13, 1, 1, 1, 15, 9, 1, 1, 15, 15, 1, 1, 15, 1, 5, 1, 15, 5, 5, 1, 15, 13, 7, 1, 15, 1, 9, 1, 15, 9, 11, 1, 15, 5, 13, 1, 15, 5, 1, 3, 15, 3, 3, 3, 15, 9, 5, 3, 15, 7, 9, 3, 15, 11, 9, 3, 15, 3, 1, 5, 15, 9, 1, 5, 15, 1, 3, 5, 15, 13, 3, 5, 15, 3, 5, 5, 15, 1, 7, 5, 15, 3, 11, 5, 15, 5, 1, 7, 15, 5, 7, 7, 15, 11, 7, 7, 15, 7, 11, 7, 15, 3, 1, 9, 15, 11, 1, 9, 15, 7, 3, 9, 15, 1, 5, 9, 15, 1, 11, 9, 15, 5, 5, 11, 15, 5, 9, 11, 15, 5, 1, 13, 15, 3, 7, 13, 15, 1, 1, 15, 15};
.global .align 1 .b8 kvalues_iq4nl[16] = {129, 152, 173, 191, 207, 221, 234, 246, 1, 13, 25, 38, 53, 69, 89, 113};
.global .align 1 .b8 kvalues_fp4[16] = {0, 1, 2, 3, 4, 6, 8, 12, 0, 255, 254, 253, 252, 250, 248, 244};
.global .align 4 .b8 iq1s_grid_gpu[8192] = {0, 0, 0, 0, 2, 0, 0, 0, 1, 1, 0, 0, 0, 2, 0, 0, 2, 2, 0, 0, 1, 0, 1, 0, 1, 1, 1, 0, 0, 0, 2, 0, 2, 0, 2, 0, 0, 2, 2, 0, 2, 2, 2, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 1, 2, 1, 1, 1, 1, 1, 2, 1, 0, 0, 0, 2, 2, 0, 0, 2, 0, 2, 0, 2, 2, 2, 0, 2, 1, 1, 1, 2, 0, 0, 2, 2, 2, 0, 2, 2, 0, 2, 2, 2, 2, 2, 2, 2, 16, 1, 0, 0, 17, 1, 0, 0, 17, 0, 1, 0, 16, 1, 1, 0, 18, 1, 1, 0, 17, 2, 1, 0, 18, 2, 1, 0, 17, 1, 2, 0, 17, 0, 0, 1, 18, 1, 0, 1, 17, 2, 0, 1, 18, 0, 1, 1, 17, 1, 1, 1, 18, 2, 1, 1, 17, 0, 2, 1, 16, 1, 2, 1, 18, 1, 2, 1, 16, 2, 2, 1, 17, 1, 0, 2, 17, 0, 1, 2, 16, 1, 1, 2, 18, 1, 1, 2, 17, 1, 2, 2, 32, 0, 0, 0, 34, 0, 0, 0, 32, 2, 0, 0, 34, 2, 0, 0, 33, 1, 1, 0, 32, 0, 2, 0, 34, 0, 2, 0, 32, 2, 2, 0, 34, 2, 2, 0, 33, 1, 0, 1, 33, 0, 1, 1, 33, 2, 1, 1, 32, 1, 2, 1, 33, 2, 2, 1, 32, 0, 0, 2, 34, 0, 0, 2, 32, 2, 0, 2, 34, 2, 0, 2, 33, 0, 1, 2, 33, 1, 1, 2, 33, 2, 1, 2, 32, 0, 2, 2, 34, 0, 2, 2, 32, 2, 2, 2, 34, 2, 2, 2, 1, 16, 1, 0, 0, 17, 1, 0, 2, 17, 1, 0, 1, 17, 2, 0, 1, 16, 0, 1, 1, 18, 0, 1, 1, 17, 1, 1, 2, 18, 1, 1, 0, 17, 2, 1, 1, 17, 2, 1, 1, 16, 1, 2, 1, 18, 1, 2, 1, 17, 2, 2, 17, 16, 0, 0, 16, 17, 0, 0, 17, 17, 0, 0, 18, 17, 0, 0, 17, 17, 1, 0, 16, 18, 1, 0, 18, 18, 1, 0, 17, 18, 2, 0, 16, 16, 0, 1, 17, 17, 0, 1, 18, 18, 0, 1, 16, 16, 1, 1, 17, 16, 1, 1, 16, 17, 1, 1, 17, 17, 1, 1, 18, 17, 1, 1, 17, 18, 1, 1, 16, 16, 2, 1, 18, 16, 2, 1, 17, 17, 2, 1, 16, 18, 2, 1, 18, 18, 2, 1, 17, 16, 0, 2, 17, 16, 1, 2, 17, 17, 1, 2, 16, 18, 1, 2, 18, 18, 1, 2, 17, 16, 2, 2, 16, 17, 2, 2, 17, 17, 2, 2, 18, 17, 2, 2, 17, 18, 2, 2, 32, 17, 1, 0, 33, 18, 1, 0, 33, 16, 0, 1, 32, 17, 0, 1, 32, 16, 1, 1, 34, 16, 1, 1, 33, 17, 1, 1, 32, 18, 1, 1, 32, 16, 2, 1, 33, 16, 2, 1, 34, 17, 2, 1, 33, 18, 2, 1, 33, 17, 0, 2, 33, 16, 1, 2, 32, 17, 1, 2, 33, 18, 1, 2, 0, 32, 0, 0, 2, 32, 0, 0, 0, 34, 0, 0, 2, 34, 0, 0, 1, 33, 1, 0, 0, 32, 2, 0, 2, 32, 2, 0, 0, 34, 2, 0, 2, 34, 2, 0, 1, 33, 0, 1, 1, 32, 1, 1, 2, 33, 1, 1, 1, 33, 2, 1, 0, 32, 0, 2, 2, 32, 0, 2, 0, 34, 0, 2, 2, 34, 0, 2, 1, 33, 1, 2, 0, 32, 2, 2, 2, 32, 2, 2, 0, 34, 2, 2, 2, 34, 2, 2, 17, 33, 0, 0, 17, 32, 1, 0, 16, 33, 1, 0, 17, 34, 1, 0, 16, 33, 2, 0, 17, 33, 2, 0, 17, 32, 0, 1, 16, 32, 1, 1, 17, 32, 1, 1, 17, 33, 1, 1, 17, 32, 2, 1, 16, 33, 2, 1, 17, 34, 2, 1, 17, 32, 1, 2, 16, 33, 1, 2, 18, 33, 1, 2, 17, 34, 1, 2, 17, 33, 2, 2, 32, 32, 0, 0, 34, 32, 0, 0, 32, 34, 0, 0, 34, 34, 0, 0, 33, 33, 1, 0, 32, 32, 2, 0, 34, 32, 2, 0, 32, 34, 2, 0, 34, 34, 2, 0, 33, 33, 0, 1, 33, 32, 1, 1, 33, 34, 1, 1, 33, 32, 2, 1, 33, 33, 2, 1, 32, 32, 0, 2, 34, 32, 0, 2, 33, 33, 0, 2, 32, 34, 0, 2, 34, 34, 0, 2, 33, 33, 1, 2, 32, 32, 2, 2, 34, 32, 2, 2, 32, 34, 2, 2, 34, 34, 2, 2, 0, 0, 17, 0, 1, 0, 17, 0, 0, 1, 17, 0, 1, 2, 17, 0, 0, 1, 18, 0, 1, 1, 18, 0, 1, 0, 16, 1, 0, 1, 16, 1, 0, 0, 17, 1, 1, 1, 17, 1, 0, 2, 17, 1, 1, 0, 18, 1, 0, 1, 18, 1, 1, 1, 18, 1, 1, 2, 18, 1, 1, 0, 17, 2, 0, 1, 17, 2, 2, 1, 17, 2, 1, 0, 18, 2, 1, 1, 18, 2, 17, 0, 16, 0, 16, 1, 16, 0, 18, 1, 16, 0, 17, 2, 16, 0, 16, 0, 17, 0, 18, 0, 17, 0, 17, 1, 17, 0, 16, 2, 17, 0, 17, 0, 18, 0, 16, 1, 18, 0, 17, 2, 18, 0, 17, 1, 16, 1, 18, 2, 16, 1, 16, 0, 17, 1, 17, 0, 17, 1, 18, 0, 17, 1, 16, 1, 17, 1, 17, 1, 17, 1, 18, 1, 17, 1, 17, 2, 17, 1, 16, 0, 18, 1, 17, 1, 18, 1, 16, 1, 16, 2, 18, 0, 17, 2, 17, 1, 17, 2, 17, 0, 18, 2, 16, 1, 18, 2, 33, 0, 17, 0, 32, 1, 17, 0, 34, 1, 17, 0, 33, 1, 18, 0, 32, 0, 16, 1, 34, 1, 16, 1, 33, 2, 16, 1, 34, 0, 17, 1, 33, 1, 17, 1, 32, 2, 17, 1, 34, 2, 17, 1, 32, 1, 18, 1, 34, 1, 18, 1, 33, 1, 16, 2, 33, 0, 17, 2, 32, 1, 17, 2, 34, 1, 17, 2, 33, 1, 18, 2, 1, 16, 16, 0, 2, 17, 16, 0, 1, 18, 16, 0, 0, 17, 17, 0, 1, 17, 17, 0, 0, 18, 17, 0, 1, 18, 17, 0, 1, 16, 18, 0, 2, 17, 18, 0, 1, 16, 16, 1, 1, 17, 16, 1, 2, 17, 16, 1, 0, 18, 16, 1, 2, 18, 16, 1, 1, 16, 17, 1, 0, 17, 17, 1, 1, 17, 17, 1, 2, 17, 17, 1, 1, 18, 17, 1, 2, 16, 18, 1, 1, 17, 18, 1, 0, 18, 18, 1, 0, 17, 16, 2, 1, 18, 16, 2, 0, 16, 17, 2, 0, 17, 17, 2, 1, 17, 17, 2, 0, 18, 17, 2, 1, 18, 17, 2, 2, 18, 17, 2, 1, 16, 18, 2, 0, 17, 18, 2, 1, 17, 18, 2, 1, 18, 18, 2, 18, 16, 16, 0, 17, 17, 16, 0, 18, 18, 16, 0, 17, 16, 17, 0, 16, 17, 17, 0, 17, 17, 17, 0, 18, 17, 17, 0, 17, 18, 17, 0, 16, 16, 18, 0, 18, 16, 18, 0, 17, 17, 18, 0, 16, 18, 18, 0, 18, 18, 18, 0, 17, 16, 16, 1, 16, 17, 16, 1, 17, 17, 16, 1, 18, 17, 16, 1, 17, 16, 17, 1, 18, 16, 17, 1, 16, 17, 17, 1, 17, 17, 17, 1, 18, 17, 17, 1, 17, 18, 17, 1, 18, 18, 17, 1, 17, 16, 18, 1, 16, 17, 18, 1, 17, 17, 18, 1, 18, 17, 18, 1, 17, 18, 18, 1, 16, 16, 16, 2, 18, 16, 16, 2, 16, 17, 16, 2, 17, 17, 16, 2, 16, 18, 16, 2, 18, 18, 16, 2, 16, 16, 17, 2, 17, 16, 17, 2, 16, 17, 17, 2, 17, 17, 17, 2, 18, 17, 17, 2, 17, 18, 17, 2, 18, 18, 17, 2, 16, 16, 18, 2, 18, 16, 18, 2, 17, 17, 18, 2, 33, 16, 16, 0, 32, 17, 16, 0, 33, 17, 16, 0, 34, 17, 16, 0, 33, 17, 17, 0, 34, 17, 17, 0, 32, 18, 17, 0, 34, 18, 17, 0, 33, 16, 18, 0, 34, 17, 18, 0, 32, 16, 16, 1, 34, 16, 16, 1, 32, 17, 16, 1, 33, 17, 16, 1, 32, 18, 16, 1, 34, 18, 16, 1, 33, 16, 17, 1, 33, 17, 17, 1, 34, 17, 17, 1, 32, 18, 17, 1, 33, 18, 17, 1, 33, 16, 18, 1, 32, 17, 18, 1, 33, 17, 18, 1, 32, 18, 18, 1, 33, 18, 18, 1, 34, 18, 18, 1, 34, 17, 16, 2, 34, 18, 16, 2, 34, 16, 17, 2, 33, 17, 17, 2, 32, 17, 18, 2, 33, 18, 18, 2, 1, 32, 17, 0, 2, 33, 17, 0, 1, 33, 18, 0, 1, 32, 16, 1, 0, 33, 16, 1, 2, 33, 16, 1, 1, 34, 16, 1, 0, 32, 17, 1, 1, 33, 17, 1, 0, 34, 17, 1, 2, 34, 17, 1, 0, 32, 18, 1, 1, 32, 18, 1, 0, 33, 18, 1, 2, 33, 18, 1, 1, 34, 18, 1, 1, 33, 16, 2, 1, 32, 17, 2, 0, 33, 17, 2, 1, 33, 18, 2, 16, 32, 17, 0, 18, 32, 17, 0, 17, 33, 17, 0, 18, 34, 17, 0, 17, 32, 18, 0, 17, 33, 18, 0, 18, 32, 16, 1, 16, 33, 16, 1, 17, 33, 16, 1, 16, 34, 16, 1, 17, 32, 17, 1, 16, 33, 17, 1, 17, 33, 17, 1, 18, 33, 17, 1, 17, 34, 17, 1, 18, 34, 17, 1, 16, 32, 18, 1, 17, 33, 18, 1, 18, 34, 18, 1, 17, 34, 16, 2, 17, 32, 17, 2, 18, 32, 17, 2, 17, 33, 17, 2, 16, 34, 17, 2, 17, 32, 18, 2, 18, 33, 18, 2, 17, 34, 18, 2, 33, 34, 16, 0, 34, 33, 17, 0, 32, 33, 18, 0, 34, 33, 18, 0, 32, 33, 16, 1, 34, 33, 16, 1, 33, 34, 16, 1, 32, 32, 17, 1, 34, 32, 17, 1, 33, 33, 17, 1, 32, 34, 17, 1, 33, 32, 18, 1, 34, 33, 18, 1, 33, 34, 18, 1, 33, 33, 16, 2, 33, 32, 17, 2, 34, 33, 17, 2, 34, 34, 17, 2, 0, 0, 32, 0, 2, 0, 32, 0, 0, 2, 32, 0, 2, 2, 32, 0, 1, 1, 33, 0, 0, 0, 34, 0, 2, 0, 34, 0, 1, 1, 34, 0, 0, 2, 34, 0, 2, 2, 34, 0, 1, 1, 32, 1, 1, 0, 33, 1, 1, 2, 33, 1, 1, 0, 34, 1, 1, 1, 34, 1, 0, 0, 32, 2, 2, 0, 32, 2, 0, 2, 32, 2, 2, 2, 32, 2, 1, 1, 33, 2, 0, 0, 34, 2, 2, 0, 34, 2, 1, 1, 34, 2, 0, 2, 34, 2, 2, 2, 34, 2, 17, 1, 32, 0, 17, 0, 33, 0, 16, 1, 33, 0, 17, 2, 33, 0, 17, 1, 34, 0, 18, 0, 32, 1, 16, 1, 32, 1, 17, 2, 32, 1, 17, 1, 33, 1, 16, 2, 33, 1, 18, 2, 33, 1, 17, 0, 34, 1, 16, 1, 34, 1, 17, 1, 34, 1, 18, 1, 34, 1, 17, 1, 32, 2, 16, 0, 33, 2, 18, 1, 33, 2, 17, 2, 33, 2, 17, 1, 34, 2, 33, 0, 32, 0, 32, 2, 32, 0, 34, 2, 32, 0, 33, 0, 33, 0, 33, 1, 33, 0, 32, 0, 34, 0, 34, 0, 34, 0, 32, 2, 34, 0, 34, 2, 34, 0, 33, 1, 32, 1, 33, 0, 33, 1, 34, 1, 33, 1, 33, 2, 33, 1, 33, 1, 34, 1, 33, 0, 32, 2, 32, 2, 32, 2, 34, 2, 32, 2, 33, 0, 33, 2, 33, 1, 33, 2, 32, 0, 34, 2, 34, 0, 34, 2, 32, 2, 34, 2, 34, 2, 34, 2, 1, 17, 32, 0, 0, 17, 33, 0, 2, 17, 33, 0, 1, 18, 33, 0, 1, 17, 34, 0, 0, 17, 32, 1, 1, 17, 32, 1, 2, 17, 32, 1, 1, 18, 32, 1, 2, 16, 33, 1, 1, 17, 33, 1, 0, 18, 33, 1, 2, 18, 33, 1, 2, 17, 34, 1, 1, 17, 32, 2, 1, 16, 33, 2, 0, 17, 33, 2, 1, 18, 33, 2, 1, 16, 34, 2, 1, 17, 34, 2, 17, 18, 32, 0, 17, 17, 33, 0, 17, 16, 34, 0, 17, 18, 34, 0, 16, 16, 32, 1, 17, 17, 32, 1, 16, 18, 32, 1, 17, 16, 33, 1, 16, 17, 33, 1, 17, 17, 33, 1, 17, 18, 33, 1, 18, 16, 34, 1, 17, 17, 34, 1, 16, 18, 34, 1, 17, 18, 32, 2, 16, 16, 33, 2, 16, 17, 33, 2, 17, 17, 33, 2, 16, 18, 33, 2, 18, 18, 33, 2, 17, 16, 34, 2, 16, 17, 34, 2, 18, 17, 34, 2, 17, 18, 34, 2, 33, 17, 32, 0, 32, 16, 33, 0, 34, 16, 33, 0, 33, 18, 33, 0, 33, 17, 34, 0, 33, 16, 32, 1, 33, 18, 32, 1, 33, 17, 33, 1, 32, 16, 34, 1, 33, 16, 34, 1, 33, 18, 34, 1, 32, 17, 32, 2, 34, 17, 32, 2, 32, 16, 33, 2, 34, 18, 33, 2, 0, 32, 32, 0, 2, 32, 32, 0, 0, 34, 32, 0, 2, 34, 32, 0, 1, 33, 33, 0, 0, 32, 34, 0, 2, 32, 34, 0, 0, 34, 34, 0, 2, 34, 34, 0, 1, 33, 32, 1, 1, 32, 33, 1, 0, 33, 33, 1, 1, 33, 34, 1, 0, 32, 32, 2, 2, 32, 32, 2, 0, 34, 32, 2, 2, 34, 32, 2, 0, 32, 34, 2, 2, 32, 34, 2, 0, 34, 34, 2, 2, 34, 34, 2, 17, 34, 32, 0, 17, 32, 33, 0, 16, 33, 33, 0, 17, 34, 33, 0, 17, 33, 34, 0, 18, 33, 32, 1, 17, 34, 32, 1, 18, 32, 33, 1, 17, 33, 33, 1, 17, 32, 34, 1, 16, 33, 34, 1, 18, 33, 34, 1, 17, 34, 34, 1, 17, 33, 32, 2, 16, 32, 33, 2, 18, 33, 33, 2, 17, 34, 33, 2, 16, 33, 34, 2, 17, 33, 34, 2, 32, 32, 32, 0, 34, 32, 32, 0, 32, 34, 32, 0, 34, 34, 32, 0, 32, 32, 34, 0, 34, 32, 34, 0, 32, 34, 34, 0, 34, 34, 34, 0, 33, 33, 32, 1, 33, 32, 33, 1, 34, 33, 33, 1, 33, 34, 33, 1, 33, 33, 34, 1, 32, 32, 32, 2, 34, 32, 32, 2, 32, 34, 32, 2, 34, 34, 32, 2, 33, 33, 33, 2, 32, 32, 34, 2, 34, 32, 34, 2, 32, 34, 34, 2, 34, 34, 34, 2, 1, 1, 0, 16, 1, 0, 1, 16, 2, 1, 1, 16, 1, 1, 2, 16, 1, 2, 0, 17, 2, 0, 1, 17, 1, 1, 1, 17, 0, 2, 1, 17, 2, 2, 1, 17, 1, 0, 2, 17, 0, 1, 2, 17, 2, 1, 2, 17, 0, 1, 1, 18, 1, 2, 1, 18, 1, 0, 2, 18, 2, 1, 2, 18, 16, 0, 0, 16, 17, 0, 0, 16, 16, 1, 0, 16, 18, 1, 0, 16, 17, 2, 0, 16, 18, 0, 1, 16, 17, 1, 1, 16, 18, 1, 1, 16, 16, 2, 1, 16, 18, 2, 1, 16, 17, 0, 2, 16, 18, 1, 2, 16, 17, 2, 2, 16, 17, 1, 0, 17, 16, 2, 0, 17, 18, 2, 0, 17, 17, 0, 1, 17, 16, 1, 1, 17, 17, 1, 1, 17, 18, 1, 1, 17, 17, 2, 1, 17, 18, 2, 1, 17, 17, 1, 2, 17, 16, 2, 2, 17, 18, 2, 2, 17, 17, 0, 0, 18, 16, 1, 0, 18, 18, 1, 0, 18, 16, 0, 1, 18, 18, 0, 1, 18, 17, 1, 1, 18, 16, 0, 2, 18, 17, 0, 2, 18, 18, 0, 2, 18, 33, 1, 0, 16, 33, 0, 1, 16, 32, 1, 1, 16, 34, 1, 1, 16, 33, 1, 2, 16, 33, 0, 0, 17, 34, 0, 1, 17, 33, 1, 1, 17, 34, 2, 1, 17, 32, 1, 2, 17, 33, 2, 2, 17, 33, 2, 0, 18, 32, 1, 1, 18, 33, 1, 2, 18, 1, 16, 0, 16, 1, 17, 1, 16, 1, 18, 1, 16, 1, 18, 2, 16, 1, 17, 0, 17, 0, 18, 0, 17, 2, 18, 0, 17, 1, 16, 1, 17, 0, 17, 1, 17, 1, 17, 1, 17, 2, 17, 1, 17, 1, 16, 2, 17, 2, 16, 2, 17, 1, 17, 2, 17, 0, 18, 2, 17, 2, 18, 2, 17, 1, 16, 0, 18, 2, 17, 0, 18, 1, 18, 0, 18, 0, 16, 1, 18, 2, 16, 1, 18, 1, 17, 1, 18, 0, 16, 2, 18, 1, 16, 2, 18, 1, 18, 2, 18, 17, 16, 0, 16, 18, 16, 0, 16, 17, 17, 0, 16, 18, 18, 0, 16, 17, 16, 1, 16, 16, 17, 1, 16, 17, 17, 1, 16, 18, 17, 1, 16, 17, 18, 1, 16, 16, 16, 2, 16, 17, 17, 2, 16, 18, 18, 2, 16, 17, 16, 0, 17, 16, 17, 0, 17, 17, 17, 0, 17, 18, 17, 0, 17, 17, 18, 0, 17, 16, 16, 1, 17, 17, 16, 1, 17, 16, 17, 1, 17, 17, 17, 1, 17, 18, 17, 1, 17, 16, 18, 1, 17, 17, 18, 1, 17, 17, 16, 2, 17, 16, 17, 2, 17, 17, 17, 2, 17, 18, 17, 2, 17, 17, 18, 2, 17, 18, 16, 0, 18, 16, 17, 0, 18, 17, 17, 0, 18, 16, 18, 0, 18, 17, 16, 1, 18, 16, 17, 1, 18, 17, 17, 1, 18, 18, 17, 1, 18, 17, 18, 1, 18, 18, 18, 1, 18, 17, 17, 2, 18, 16, 18, 2, 18, 18, 18, 2, 18, 33, 16, 0, 16, 33, 17, 0, 16, 33, 18, 0, 16, 32, 17, 1, 16, 33, 17, 1, 16, 32, 18, 1, 16, 34, 18, 1, 16, 33, 16, 2, 16, 32, 17, 2, 16, 33, 18, 2, 16, 32, 16, 0, 17, 34, 16, 0, 17, 33, 17, 0, 17, 32, 18, 0, 17, 32, 16, 1, 17, 33, 16, 1, 17, 34, 16, 1, 17, 33, 17, 1, 17, 34, 17, 1, 17, 33, 18, 1, 17, 34, 16, 2, 17, 33, 17, 2, 17, 32, 18, 2, 17, 33, 16, 0, 18, 33, 17, 0, 18, 34, 18, 0, 18, 32, 17, 1, 18, 33, 17, 1, 18, 33, 16, 2, 18, 32, 17, 2, 18, 34, 17, 2, 18, 1, 33, 0, 16, 1, 32, 1, 16, 1, 33, 1, 16, 2, 34, 1, 16, 1, 33, 2, 16, 2, 32, 0, 17, 1, 34, 0, 17, 0, 32, 1, 17, 1, 33, 1, 17, 0, 34, 1, 17, 1, 32, 2, 17, 0, 33, 2, 17, 2, 33, 2, 17, 1, 34, 2, 17, 1, 33, 0, 18, 1, 32, 1, 18, 0, 33, 1, 18, 2, 33, 1, 18, 1, 34, 1, 18, 1, 33, 2, 18, 17, 32, 0, 16, 17, 33, 0, 16, 18, 33, 0, 16, 18, 34, 0, 16, 16, 32, 1, 16, 16, 33, 1, 16, 17, 33, 1, 16, 16, 34, 1, 16, 17, 32, 2, 16, 16, 33, 2, 16, 18, 33, 2, 16, 16, 32, 0, 17, 17, 33, 0, 17, 18, 34, 0, 17, 17, 32, 1, 17, 18, 32, 1, 17, 16, 33, 1, 17, 17, 33, 1, 17, 18, 33, 1, 17, 17, 34, 1, 17, 16, 32, 2, 17, 18, 32, 2, 17, 17, 33, 2, 17, 18, 33, 2, 17, 18, 34, 2, 17, 18, 33, 0, 18, 17, 34, 0, 18, 18, 32, 1, 18, 17, 33, 1, 18, 18, 33, 1, 18, 16, 34, 1, 18, 17, 32, 2, 18, 16, 33, 2, 18, 18, 33, 2, 18, 17, 34, 2, 18, 34, 33, 1, 16, 32, 33, 0, 17, 34, 33, 0, 17, 33, 34, 0, 17, 33, 33, 1, 17, 32, 34, 1, 17, 34, 34, 1, 17, 32, 33, 2, 17, 33, 34, 2, 17, 32, 33, 1, 18, 33, 33, 2, 18, 1, 0, 16, 16, 0, 1, 16, 16, 1, 1, 16, 16, 2, 1, 16, 16, 1, 2, 16, 16, 2, 0, 17, 16, 1, 1, 17, 16, 2, 2, 17, 16, 1, 0, 18, 16, 0, 1, 18, 16, 1, 2, 18, 16, 0, 0, 16, 17, 1, 1, 16, 17, 0, 2, 16, 17, 1, 0, 17, 17, 0, 1, 17, 17, 1, 1, 17, 17, 2, 1, 17, 17, 1, 2, 17, 17, 1, 1, 18, 17, 0, 2, 18, 17, 2, 1, 16, 18, 1, 2, 16, 18, 1, 1, 17, 18, 0, 2, 17, 18, 0, 0, 18, 18, 1, 0, 18, 18, 2, 1, 18, 18, 1, 2, 18, 18, 17, 1, 16, 16, 16, 2, 16, 16, 17, 2, 16, 16, 18, 2, 16, 16, 17, 0, 17, 16, 16, 1, 17, 16, 17, 1, 17, 16, 18, 1, 17, 16, 16, 2, 17, 16, 17, 2, 17, 16, 16, 0, 18, 16, 17, 1, 18, 16, 18, 1, 18, 16, 16, 2, 18, 16, 18, 2, 18, 16, 17, 0, 16, 17, 16, 1, 16, 17, 17, 1, 16, 17, 18, 1, 16, 17, 17, 2, 16, 17, 16, 0, 17, 17, 17, 0, 17, 17, 18, 0, 17, 17, 16, 1, 17, 17, 17, 1, 17, 17, 18, 1, 17, 17, 16, 2, 17, 17, 17, 2, 17, 17, 18, 2, 17, 17, 17, 0, 18, 17, 16, 1, 18, 17, 17, 1, 18, 17, 18, 1, 18, 17, 17, 2, 18, 17, 18, 0, 16, 18, 17, 1, 16, 18, 17, 0, 17, 18, 16, 1, 17, 18, 17, 1, 17, 18, 18, 1, 17, 18, 17, 2, 17, 18, 16, 0, 18, 18, 17, 1, 18, 18, 18, 2, 18, 18, 33, 0, 16, 16, 34, 1, 16, 16, 34, 0, 17, 16, 33, 1, 17, 16, 34, 2, 17, 16, 33, 0, 18, 16, 32, 1, 18, 16, 34, 0, 16, 17, 33, 1, 16, 17, 34, 2, 16, 17, 33, 0, 17, 17, 32, 1, 17, 17, 33, 1, 17, 17, 34, 1, 17, 17, 33, 2, 17, 17, 34, 0, 18, 17, 33, 1, 18, 17, 33, 1, 16, 18, 32, 0, 17, 18, 34, 0, 17, 18, 33, 1, 17, 18, 33, 2, 17, 18, 34, 2, 17, 18, 32, 1, 18, 18, 0, 17, 16, 16, 1, 17, 16, 16, 1, 16, 17, 16, 0, 17, 17, 16, 1, 17, 17, 16, 2, 17, 17, 16, 0, 18, 17, 16, 1, 18, 17, 16, 1, 16, 18, 16, 1, 17, 18, 16, 0, 18, 18, 16, 2, 18, 18, 16, 1, 16, 16, 17, 0, 17, 16, 17, 1, 17, 16, 17, 2, 17, 16, 17, 1, 18, 16, 17, 2, 18, 16, 17, 0, 16, 17, 17, 1, 16, 17, 17, 0, 17, 17, 17, 1, 17, 17, 17, 2, 17, 17, 17, 0, 18, 17, 17, 1, 18, 17, 17, 2, 18, 17, 17, 1, 16, 18, 17, 2, 16, 18, 17, 0, 17, 18, 17, 1, 17, 18, 17, 2, 17, 18, 17, 1, 18, 18, 17, 0, 16, 16, 18, 0, 18, 16, 18, 2, 18, 16, 18, 1, 16, 17, 18, 0, 17, 17, 18, 1, 17, 17, 18, 2, 17, 17, 18, 1, 18, 17, 18, 1, 16, 18, 18, 0, 17, 18, 18, 1, 17, 18, 18, 2, 18, 18, 18, 17, 16, 16, 16, 18, 16, 16, 16, 16, 17, 16, 16, 17, 17, 16, 16, 18, 17, 16, 16, 17, 18, 16, 16, 16, 16, 17, 16, 17, 16, 17, 16, 18, 16, 17, 16, 16, 17, 17, 16, 17, 17, 17, 16, 18, 17, 17, 16, 17, 18, 17, 16, 18, 18, 17, 16, 17, 16, 18, 16, 16, 17, 18, 16, 17, 17, 18, 16, 18, 17, 18, 16, 17, 18, 18, 16, 16, 16, 16, 17, 17, 16, 16, 17, 18, 16, 16, 17, 16, 17, 16, 17, 17, 17, 16, 17, 18, 17, 16, 17, 16, 18, 16, 17, 17, 18, 16, 17, 16, 16, 17, 17, 17, 16, 17, 17, 18, 16, 17, 17, 16, 17, 17, 17, 17, 17, 17, 17, 18, 17, 17, 17, 16, 18, 17, 17, 17, 18, 17, 17, 18, 18, 17, 17, 16, 16, 18, 17, 17, 16, 18, 17, 16, 17, 18, 17, 17, 17, 18, 17, 18, 17, 18, 17, 16, 18, 18, 17, 17, 18, 18, 17, 18, 18, 18, 17, 17, 16, 16, 18, 16, 17, 16, 18, 17, 17, 16, 18, 17, 18, 16, 18, 18, 18, 16, 18, 16, 16, 17, 18, 17, 16, 17, 18, 16, 17, 17, 18, 17, 17, 17, 18, 18, 17, 17, 18, 16, 18, 17, 18, 17, 18, 17, 18, 17, 16, 18, 18, 16, 17, 18, 18, 17, 17, 18, 18, 18, 17, 18, 18, 17, 18, 18, 18, 32, 16, 16, 16, 33, 16, 16, 16, 34, 16, 16, 16, 32, 17, 16, 16, 34, 17, 16, 16, 32, 18, 16, 16, 33, 18, 16, 16, 33, 16, 17, 16, 32, 17, 17, 16, 33, 17, 17, 16, 32, 18, 17, 16, 33, 18, 17, 16, 32, 16, 18, 16, 33, 16, 18, 16, 34, 16, 18, 16, 32, 17, 18, 16, 33, 17, 18, 16, 34, 17, 18, 16, 32, 18, 18, 16, 33, 18, 18, 16, 33, 16, 16, 17, 33, 17, 16, 17, 34, 17, 16, 17, 32, 18, 16, 17, 33, 18, 16, 17, 34, 18, 16, 17, 32, 16, 17, 17, 33, 16, 17, 17, 34, 16, 17, 17, 32, 17, 17, 17, 33, 17, 17, 17, 34, 17, 17, 17, 32, 18, 17, 17, 33, 18, 17, 17, 34, 18, 17, 17, 33, 16, 18, 17, 32, 17, 18, 17, 33, 17, 18, 17, 33, 18, 18, 17, 34, 16, 16, 18, 33, 17, 16, 18, 34, 17, 16, 18, 32, 18, 16, 18, 33, 18, 16, 18, 34, 18, 16, 18, 33, 16, 17, 18, 33, 17, 17, 18, 34, 18, 17, 18, 34, 16, 18, 18, 33, 17, 18, 18, 34, 17, 18, 18, 32, 18, 18, 18, 33, 18, 18, 18, 0, 33, 16, 16, 1, 33, 16, 16, 2, 33, 16, 16, 1, 34, 16, 16, 0, 32, 17, 16, 1, 33, 17, 16, 0, 34, 17, 16, 1, 32, 18, 16, 2, 34, 18, 16, 1, 33, 16, 17, 0, 34, 16, 17, 2, 34, 16, 17, 1, 32, 17, 17, 0, 33, 17, 17, 1, 33, 17, 17, 2, 33, 17, 17, 0, 34, 17, 17, 1, 34, 17, 17, 0, 32, 18, 17, 2, 32, 18, 17, 0, 33, 18, 17, 1, 33, 18, 17, 2, 32, 16, 18, 1, 34, 16, 18, 0, 32, 17, 18, 2, 32, 17, 18, 1, 33, 17, 18, 0, 34, 17, 18, 1, 32, 18, 18, 1, 34, 18, 18, 17, 32, 16, 16, 18, 32, 16, 16, 17, 33, 16, 16, 18, 34, 16, 16, 17, 32, 17, 16, 16, 33, 17, 16, 17, 33, 17, 16, 18, 33, 17, 16, 17, 34, 17, 16, 17, 33, 18, 16, 17, 32, 16, 17, 16, 33, 16, 17, 17, 33, 16, 17, 18, 33, 16, 17, 17, 34, 16, 17, 16, 32, 17, 17, 17, 32, 17, 17, 18, 32, 17, 17, 16, 33, 17, 17, 17, 33, 17, 17, 18, 33, 17, 17, 16, 34, 17, 17, 17, 34, 17, 17, 18, 34, 17, 17, 17, 32, 18, 17, 16, 33, 18, 17, 17, 33, 18, 17, 18, 33, 18, 17, 17, 34, 18, 17, 17, 32, 16, 18, 17, 33, 16, 18, 17, 34, 16, 18, 17, 32, 17, 18, 16, 33, 17, 18, 17, 33, 17, 18, 18, 33, 17, 18, 16, 34, 17, 18, 17, 34, 17, 18, 17, 33, 18, 18, 32, 33, 16, 16, 32, 34, 16, 16, 33, 33, 17, 16, 34, 34, 17, 16, 32, 32, 18, 16, 33, 33, 18, 16, 34, 33, 18, 16, 33, 34, 18, 16, 33, 33, 16, 17, 32, 34, 16, 17, 33, 34, 16, 17, 33, 32, 17, 17, 33, 33, 17, 17, 34, 33, 17, 17, 32, 34, 17, 17, 33, 34, 17, 17, 34, 32, 18, 17, 33, 33, 18, 17, 32, 34, 18, 17, 34, 34, 18, 17, 33, 32, 16, 18, 34, 34, 16, 18, 34, 32, 17, 18, 33, 33, 17, 18, 34, 33, 17, 18, 32, 34, 17, 18, 34, 34, 17, 18, 33, 32, 18, 18, 1, 1, 32, 16, 0, 1, 33, 16, 2, 1, 33, 16, 1, 2, 33, 16, 1, 1, 34, 16, 0, 1, 32, 17, 0, 0, 33, 17, 1, 1, 33, 17, 2, 1, 33, 17, 0, 2, 33, 17, 2, 2, 33, 17, 1, 0, 34, 17, 0, 1, 34, 17, 2, 1, 34, 17, 1, 2, 34, 17, 1, 0, 32, 18, 2, 1, 33, 18, 1, 1, 34, 18, 17, 0, 32, 16, 16, 1, 32, 16, 18, 1, 32, 16, 17, 2, 32, 16, 18, 0, 33, 16, 17, 1, 33, 16, 17, 0, 34, 16, 18, 0, 34, 16, 18, 1, 34, 16, 17, 2, 34, 16, 17, 1, 32, 17, 17, 2, 32, 17, 17, 0, 33, 17, 17, 1, 33, 17, 18, 1, 33, 17, 17, 2, 33, 17, 17, 1, 34, 17, 18, 1, 34, 17, 18, 2, 34, 17, 16, 1, 32, 18, 18, 2, 32, 18, 18, 0, 33, 18, 17, 1, 33, 18, 17, 0, 34, 18, 18, 1, 34, 18, 17, 2, 34, 18, 33, 0, 33, 16, 34, 1, 33, 16, 33, 2, 33, 16, 32, 0, 32, 17, 33, 0, 32, 17, 34, 1, 32, 17, 33, 1, 33, 17, 34, 1, 33, 17, 32, 2, 33, 17, 32, 0, 34, 17, 33, 1, 32, 18, 33, 0, 33, 18, 34, 1, 33, 18, 33, 1, 34, 18, 1, 16, 33, 16, 2, 16, 33, 16, 1, 17, 33, 16, 2, 17, 33, 16, 2, 18, 33, 16, 1, 16, 34, 16, 2, 17, 34, 16, 1, 18, 34, 16, 0, 16, 32, 17, 2, 16, 32, 17, 1, 17, 32, 17, 0, 18, 32, 17, 2, 18, 32, 17, 1, 16, 33, 17, 0, 17, 33, 17, 1, 17, 33, 17, 2, 17, 33, 17, 1, 18, 33, 17, 2, 18, 33, 17, 0, 16, 34, 17, 2, 16, 34, 17, 1, 17, 34, 17, 0, 17, 32, 18, 1, 17, 32, 18, 1, 18, 32, 18, 0, 16, 33, 18, 2, 16, 33, 18, 0, 17, 33, 18, 1, 17, 33, 18, 2, 17, 33, 18, 0, 18, 33, 18, 2, 18, 33, 18, 1, 16, 34, 18, 0, 17, 34, 18, 1, 18, 34, 18, 17, 17, 32, 16, 16, 18, 32, 16, 18, 18, 32, 16, 17, 16, 33, 16, 17, 17, 33, 16, 18, 17, 33, 16, 17, 18, 33, 16, 16, 17, 32, 17, 17, 17, 32, 17, 18, 17, 32, 17, 17, 18, 32, 17, 16, 16, 33, 17, 17, 16, 33, 17, 16, 17, 33, 17, 17, 17, 33, 17, 18, 17, 33, 17, 17, 18, 33, 17, 17, 16, 34, 17, 16, 17, 34, 17, 17, 17, 34, 17, 18, 17, 34, 17, 17, 18, 34, 17, 18, 17, 32, 18, 17, 18, 32, 18, 18, 18, 32, 18, 17, 16, 33, 18, 17, 17, 33, 18, 18, 17, 33, 18, 17, 18, 33, 18, 18, 18, 33, 18, 18, 16, 34, 18, 17, 17, 34, 18, 18, 17, 34, 18, 16, 18, 34, 18, 34, 16, 32, 16, 33, 18, 32, 16, 33, 17, 33, 16, 32, 16, 34, 16, 34, 17, 34, 16, 32, 18, 34, 16, 33, 18, 34, 16, 32, 16, 32, 17, 33, 17, 32, 17, 32, 18, 32, 17, 34, 18, 32, 17, 33, 16, 33, 17, 32, 17, 33, 17, 33, 17, 33, 17, 34, 17, 33, 17, 32, 18, 33, 17, 34, 18, 33, 17, 32, 16, 34, 17, 33, 17, 34, 17, 32, 18, 34, 17, 32, 16, 32, 18, 34, 16, 32, 18, 33, 17, 32, 18, 34, 18, 32, 18, 32, 17, 33, 18, 34, 17, 33, 18, 32, 18, 33, 18, 33, 18, 33, 18, 32, 16, 34, 18, 32, 17, 34, 18, 34, 17, 34, 18, 34, 18, 34, 18, 2, 33, 33, 16, 1, 34, 33, 16, 1, 33, 34, 16, 1, 32, 32, 17, 2, 32, 33, 17, 1, 33, 33, 17, 2, 34, 33, 17, 1, 32, 34, 17, 1, 34, 34, 17, 1, 33, 32, 18, 1, 32, 33, 18, 0, 34, 33, 18, 2, 33, 34, 18, 17, 32, 32, 16, 16, 33, 32, 16, 16, 32, 33, 16, 17, 33, 33, 16, 17, 32, 34, 16, 16, 33, 34, 16, 18, 33, 34, 16, 17, 34, 34, 16, 16, 32, 32, 17, 17, 32, 32, 17, 17, 33, 32, 17, 18, 33, 32, 17, 16, 34, 32, 17, 17, 32, 33, 17, 16, 33, 33, 17, 17, 33, 33, 17, 18, 33, 33, 17, 17, 34, 33, 17, 16, 32, 34, 17, 17, 33, 34, 17, 18, 34, 34, 17, 18, 32, 32, 18, 16, 33, 32, 18, 18, 34, 32, 18, 17, 33, 33, 18, 17, 32, 34, 18, 16, 33, 34, 18, 17, 33, 34, 18, 17, 34, 34, 18, 33, 32, 33, 16, 34, 33, 33, 16, 32, 34, 33, 16, 33, 32, 32, 17, 32, 33, 32, 17, 33, 34, 32, 17, 32, 32, 33, 17, 33, 33, 33, 17, 32, 34, 33, 17, 34, 34, 33, 17, 32, 33, 34, 17, 33, 33, 34, 17, 33, 34, 34, 17, 34, 33, 32, 18, 32, 33, 33, 18, 32, 34, 33, 18, 34, 34, 33, 18, 34, 33, 34, 18, 0, 0, 0, 32, 2, 0, 0, 32, 0, 2, 0, 32, 2, 2, 0, 32, 0, 0, 2, 32, 2, 0, 2, 32, 0, 2, 2, 32, 2, 2, 2, 32, 1, 1, 0, 33, 0, 0, 1, 33, 1, 0, 1, 33, 0, 1, 1, 33, 2, 1, 1, 33, 1, 2, 1, 33, 1, 1, 2, 33, 0, 0, 0, 34, 2, 0, 0, 34, 0, 2, 0, 34, 2, 2, 0, 34, 1, 1, 1, 34, 0, 0, 2, 34, 2, 0, 2, 34, 0, 2, 2, 34, 2, 2, 2, 34, 17, 1, 0, 32, 17, 0, 1, 32, 16, 1, 1, 32, 18, 1, 1, 32, 17, 2, 1, 32, 17, 1, 2, 32, 17, 0, 0, 33, 16, 1, 0, 33, 17, 2, 0, 33, 16, 0, 1, 33, 18, 0, 1, 33, 17, 1, 1, 33, 18, 1, 1, 33, 16, 2, 1, 33, 17, 2, 1, 33, 16, 1, 2, 33, 18, 1, 2, 33, 17, 2, 2, 33, 17, 1, 0, 34, 17, 2, 0, 34, 16, 1, 1, 34, 18, 1, 1, 34, 17, 2, 1, 34, 17, 1, 2, 34, 32, 0, 0, 32, 34, 0, 0, 32, 32, 2, 0, 32, 34, 2, 0, 32, 33, 1, 1, 32, 32, 0, 2, 32, 34, 0, 2, 32, 32, 2, 2, 32, 34, 2, 2, 32, 33, 0, 1, 33, 32, 1, 1, 33, 33, 2, 1, 33, 33, 1, 2, 33, 32, 0, 0, 34, 34, 0, 0, 34, 32, 2, 0, 34, 34, 2, 0, 34, 33, 1, 1, 34, 32, 0, 2, 34, 34, 0, 2, 34, 32, 2, 2, 34, 34, 2, 2, 34, 0, 17, 1, 32, 1, 18, 1, 32, 1, 16, 0, 33, 0, 17, 0, 33, 1, 16, 1, 33, 1, 17, 1, 33, 2, 18, 1, 33, 1, 16, 2, 33, 0, 17, 2, 33, 1, 18, 2, 33, 0, 17, 1, 34, 1, 18, 1, 34, 17, 16, 0, 32, 17, 18, 0, 32, 18, 16, 1, 32, 17, 17, 1, 32, 18, 18, 1, 32, 18, 17, 2, 32, 17, 18, 2, 32, 16, 16, 0, 33, 17, 16, 0, 33, 17, 17, 0, 33, 16, 18, 0, 33, 17, 16, 1, 33, 16, 17, 1, 33, 17, 17, 1, 33, 18, 17, 1, 33, 17, 18, 1, 33, 18, 18, 1, 33, 17, 17, 2, 33, 18, 17, 2, 33, 16, 18, 2, 33, 18, 18, 2, 33, 17, 16, 0, 34, 16, 17, 0, 34, 18, 17, 0, 34, 17, 18, 0, 34, 16, 16, 1, 34, 18, 16, 1, 34, 17, 17, 1, 34, 16, 18, 1, 34, 18, 17, 2, 34, 33, 16, 1, 32, 34, 17, 1, 32, 33, 18, 1, 32, 33, 17, 2, 32, 33, 16, 0, 33, 32, 17, 0, 33, 33, 18, 0, 33, 34, 18, 0, 33, 32, 16, 1, 33, 33, 17, 1, 33, 33, 18, 1, 33, 34, 18, 1, 33, 33, 16, 2, 33, 34, 17, 2, 33, 34, 18, 2, 33, 33, 17, 0, 34, 33, 16, 1, 34, 34, 18, 1, 34, 32, 17, 2, 34, 0, 32, 0, 32, 2, 32, 0, 32, 0, 34, 0, 32, 2, 34, 0, 32, 1, 33, 1, 32, 0, 32, 2, 32, 2, 32, 2, 32, 0, 34, 2, 32, 2, 34, 2, 32, 1, 32, 0, 33, 1, 33, 0, 33, 1, 32, 1, 33, 0, 33, 1, 33, 1, 34, 1, 33, 1, 33, 2, 33, 1, 34, 2, 33, 0, 32, 0, 34, 2, 32, 0, 34, 0, 34, 0, 34, 2, 34, 0, 34, 1, 33, 1, 34, 0, 32, 2, 34, 2, 32, 2, 34, 0, 34, 2, 34, 2, 34, 2, 34, 17, 33, 0, 32, 18, 33, 0, 32, 17, 32, 1, 32, 16, 33, 1, 32, 18, 33, 1, 32, 17, 33, 2, 32, 17, 32, 0, 33, 16, 33, 0, 33, 18, 33, 0, 33, 17, 34, 0, 33, 16, 32, 1, 33, 18, 32, 1, 33, 17, 33, 1, 33, 18, 34, 1, 33, 17, 32, 2, 33, 16, 33, 2, 33, 17, 33, 0, 34, 18, 33, 1, 34, 17, 34, 1, 34, 17, 33, 2, 34, 32, 32, 0, 32, 34, 32, 0, 32, 32, 34, 0, 32, 34, 34, 0, 32, 33, 33, 1, 32, 32, 32, 2, 32, 34, 32, 2, 32, 32, 34, 2, 32, 34, 34, 2, 32, 33, 33, 0, 33, 33, 32, 1, 33, 32, 33, 1, 33, 34, 33, 1, 33, 32, 32, 0, 34, 34, 32, 0, 34, 32, 34, 0, 34, 34, 34, 0, 34, 33, 33, 1, 34, 32, 32, 2, 34, 34, 32, 2, 34, 32, 34, 2, 34, 34, 34, 2, 34, 1, 1, 16, 32, 1, 0, 17, 32, 2, 1, 17, 32, 0, 2, 17, 32, 1, 2, 17, 32, 1, 1, 18, 32, 1, 0, 16, 33, 2, 1, 16, 33, 1, 2, 16, 33, 1, 1, 17, 33, 0, 2, 17, 33, 2, 2, 17, 33, 1, 2, 18, 33, 2, 2, 18, 33, 1, 1, 16, 34, 1, 0, 17, 34, 0, 1, 17, 34, 2, 1, 17, 34, 1, 2, 17, 34, 1, 1, 18, 34, 17, 0, 16, 32, 16, 1, 16, 32, 18, 1, 16, 32, 17, 2, 16, 32, 16, 0, 17, 32, 17, 1, 17, 32, 16, 2, 17, 32, 18, 2, 17, 32, 17, 0, 18, 32, 16, 1, 18, 32, 18, 1, 18, 32, 17, 2, 18, 32, 16, 0, 16, 33, 17, 1, 16, 33, 16, 0, 17, 33, 17, 0, 17, 33, 16, 1, 17, 33, 17, 1, 17, 33, 18, 1, 17, 33, 17, 2, 17, 33, 18, 0, 18, 33, 17, 1, 18, 33, 16, 1, 16, 34, 18, 1, 16, 34, 18, 0, 17, 34, 17, 1, 17, 34, 16, 2, 17, 34, 17, 0, 18, 34, 16, 1, 18, 34, 18, 1, 18, 34, 17, 2, 18, 34, 33, 1, 16, 32, 33, 0, 17, 32, 32, 1, 17, 32, 33, 2, 17, 32, 33, 1, 18, 32, 32, 1, 16, 33, 34, 1, 16, 33, 33, 2, 16, 33, 32, 0, 17, 33, 34, 0, 17, 33, 33, 1, 17, 33, 32, 2, 17, 33, 34, 1, 18, 33, 33, 2, 18, 33, 33, 1, 16, 34, 32, 1, 17, 34, 34, 1, 17, 34, 33, 2, 18, 34, 1, 16, 16, 32, 0, 17, 16, 32, 2, 17, 16, 32, 0, 16, 17, 32, 1, 17, 17, 32, 0, 18, 17, 32, 2, 17, 18, 32, 0, 16, 16, 33, 2, 18, 16, 33, 1, 16, 17, 33, 0, 17, 17, 33, 1, 17, 17, 33, 2, 17, 17, 33, 0, 18, 17, 33, 1, 18, 17, 33, 0, 16, 18, 33, 1, 16, 18, 33, 2, 16, 18, 33, 1, 17, 18, 33, 0, 17, 16, 34, 2, 17, 16, 34, 2, 16, 17, 34, 0, 17, 17, 34, 1, 17, 17, 34, 0, 18, 17, 34, 1, 16, 18, 34, 1, 18, 18, 34, 16, 16, 16, 32, 17, 17, 16, 32, 16, 18, 16, 32, 18, 18, 16, 32, 16, 16, 17, 32, 17, 16, 17, 32, 16, 17, 17, 32, 17, 17, 17, 32, 18, 17, 17, 32, 17, 18, 17, 32, 17, 16, 18, 32, 17, 17, 18, 32, 17, 18, 18, 32, 18, 18, 18, 32, 17, 16, 16, 33, 16, 17, 16, 33, 17, 17, 16, 33, 18, 17, 16, 33, 17, 18, 16, 33, 16, 16, 17, 33, 17, 16, 17, 33, 18, 16, 17, 33, 16, 17, 17, 33, 17, 17, 17, 33, 18, 17, 17, 33, 16, 18, 17, 33, 17, 18, 17, 33, 18, 18, 17, 33, 17, 16, 18, 33, 16, 17, 18, 33, 17, 17, 18, 33, 18, 17, 18, 33, 17, 18, 18, 33, 17, 16, 16, 34, 17, 17, 16, 34, 16, 18, 16, 34, 17, 16, 17, 34, 18, 16, 17, 34, 16, 17, 17, 34, 17, 17, 17, 34, 18, 17, 17, 34, 17, 18, 17, 34, 18, 18, 17, 34, 16, 16, 18, 34, 18, 16, 18, 34, 17, 17, 18, 34, 16, 18, 18, 34, 18, 18, 18, 34, 33, 16, 16, 32, 32, 17, 16, 32, 32, 16, 17, 32, 33, 17, 17, 32, 33, 18, 17, 32, 32, 16, 18, 32, 34, 17, 18, 32, 33, 18, 18, 32, 33, 17, 16, 33, 32, 18, 16, 33, 33, 18, 16, 33, 33, 16, 17, 33, 34, 16, 17, 33, 33, 17, 17, 33, 34, 17, 17, 33, 33, 18, 17, 33, 33, 17, 18, 33, 32, 18, 18, 33, 34, 16, 16, 34, 32, 17, 16, 34, 33, 18, 16, 34, 34, 18, 16, 34, 34, 16, 17, 34, 32, 17, 17, 34, 33, 17, 17, 34, 32, 17, 18, 34, 34, 17, 18, 34, 33, 18, 18, 34, 1, 33, 16, 32, 2, 33, 17, 32, 1, 34, 17, 32, 1, 33, 18, 32, 1, 32, 16, 33, 2, 33, 16, 33, 0, 32, 17, 33, 2, 32, 17, 33, 1, 33, 17, 33, 2, 33, 17, 33, 2, 34, 17, 33, 0, 33, 18, 33, 1, 33, 18, 33, 1, 33, 16, 34, 1, 32, 17, 34, 2, 33, 17, 34, 1, 34, 17, 34, 1, 33, 18, 34, 16, 33, 16, 32, 18, 33, 16, 32, 17, 34, 16, 32, 16, 32, 17, 32, 18, 32, 17, 32, 17, 33, 17, 32, 16, 34, 17, 32, 18, 34, 17, 32, 16, 32, 18, 32, 17, 32, 18, 32, 16, 33, 18, 32, 18, 33, 18, 32, 16, 32, 16, 33, 18, 32, 16, 33, 17, 33, 16, 33, 16, 34, 16, 33, 18, 34, 16, 33, 17, 32, 17, 33, 16, 33, 17, 33, 17, 33, 17, 33, 18, 33, 17, 33, 17, 34, 17, 33, 18, 32, 18, 33, 17, 33, 18, 33, 18, 33, 18, 33, 18, 34, 18, 33, 17, 32, 16, 34, 16, 33, 16, 34, 16, 32, 17, 34, 18, 32, 17, 34, 17, 33, 17, 34, 18, 34, 17, 34, 17, 32, 18, 34, 18, 33, 18, 34, 33, 33, 16, 32, 33, 33, 17, 32, 33, 33, 18, 32, 32, 33, 16, 33, 34, 33, 16, 33, 33, 34, 16, 33, 32, 32, 17, 33, 33, 33, 17, 33, 32, 34, 17, 33, 33, 32, 18, 33, 33, 33, 16, 34, 33, 32, 17, 34, 32, 33, 17, 34, 33, 33, 17, 34, 34, 33, 17, 34, 0, 0, 32, 32, 2, 0, 32, 32, 0, 2, 32, 32, 2, 2, 32, 32, 1, 1, 33, 32, 0, 0, 34, 32, 2, 0, 34, 32, 0, 2, 34, 32, 2, 2, 34, 32, 1, 1, 32, 33, 1, 0, 33, 33, 0, 1, 33, 33, 2, 1, 33, 33, 1, 2, 33, 33, 0, 0, 32, 34, 2, 0, 32, 34, 0, 2, 32, 34, 2, 2, 32, 34, 1, 1, 33, 34, 0, 0, 34, 34, 2, 0, 34, 34, 0, 2, 34, 34, 2, 2, 34, 34, 17, 1, 32, 32, 17, 2, 32, 32, 17, 0, 33, 32, 16, 1, 33, 32, 18, 1, 33, 32, 17, 2, 33, 32, 18, 2, 33, 32, 18, 1, 32, 33, 17, 2, 32, 33, 17, 0, 33, 33, 17, 1, 33, 33, 16, 2, 33, 33, 18, 2, 33, 33, 17, 0, 34, 33, 16, 1, 34, 33, 17, 1, 32, 34, 16, 0, 33, 34, 18, 0, 33, 34, 18, 1, 33, 34, 17, 2, 33, 34, 34, 0, 32, 32, 32, 2, 32, 32, 34, 2, 32, 32, 32, 0, 33, 32, 33, 2, 33, 32, 34, 0, 34, 32, 32, 2, 34, 32, 34, 2, 34, 32, 33, 1, 32, 33, 33, 0, 33, 33, 34, 1, 33, 33, 33, 2, 33, 33, 33, 1, 34, 33, 32, 0, 32, 34, 34, 0, 32, 34, 32, 2, 32, 34, 34, 2, 32, 34, 33, 1, 33, 34, 32, 0, 34, 34, 34, 0, 34, 34, 32, 2, 34, 34, 34, 2, 34, 34, 1, 18, 33, 32, 1, 17, 34, 32, 1, 16, 32, 33, 0, 17, 32, 33, 0, 16, 33, 33, 0, 17, 33, 33, 1, 17, 33, 33, 0, 18, 33, 33, 2, 18, 33, 33, 1, 16, 34, 33, 1, 17, 34, 33, 2, 17, 34, 33, 0, 18, 34, 33, 1, 18, 34, 33, 1, 17, 32, 34, 18, 17, 32, 32, 17, 18, 32, 32, 16, 16, 33, 32, 18, 16, 33, 32, 17, 17, 33, 32, 16, 18, 33, 32, 18, 17, 34, 32, 17, 18, 34, 32, 18, 16, 32, 33, 17, 17, 32, 33, 17, 16, 33, 33, 16, 17, 33, 33, 17, 17, 33, 33, 18, 17, 33, 33, 17, 18, 33, 33, 17, 17, 34, 33, 18, 18, 34, 33, 17, 16, 32, 34, 16, 17, 32, 34, 17, 17, 32, 34, 18, 17, 32, 34, 17, 18, 32, 34, 18, 16, 33, 34, 17, 17, 33, 34, 16, 18, 33, 34, 33, 17, 32, 32, 33, 16, 33, 32, 34, 17, 33, 32, 34, 18, 33, 32, 33, 16, 34, 32, 33, 17, 34, 32, 32, 17, 32, 33, 34, 17, 32, 33, 34, 18, 32, 33, 34, 16, 33, 33, 33, 17, 33, 33, 34, 17, 33, 33, 32, 18, 33, 33, 32, 16, 34, 33, 34, 16, 34, 33, 34, 17, 32, 34, 32, 16, 33, 34, 33, 17, 33, 34, 34, 17, 33, 34, 33, 18, 33, 34, 33, 16, 34, 34, 32, 17, 34, 34, 34, 17, 34, 34, 0, 32, 32, 32, 2, 32, 32, 32, 0, 34, 32, 32, 2, 34, 32, 32, 0, 32, 34, 32, 2, 32, 34, 32, 0, 34, 34, 32, 2, 34, 34, 32, 1, 32, 33, 33, 0, 33, 33, 33, 2, 33, 33, 33, 1, 34, 33, 33, 0, 32, 32, 34, 2, 32, 32, 34, 0, 34, 32, 34, 2, 34, 32, 34, 1, 33, 33, 34, 0, 32, 34, 34, 2, 32, 34, 34, 0, 34, 34, 34, 2, 34, 34, 34, 17, 33, 32, 32, 16, 33, 33, 32, 17, 34, 33, 32, 17, 32, 34, 32, 17, 33, 34, 32, 17, 32, 32, 33, 16, 32, 33, 33, 17, 33, 33, 33, 18, 34, 33, 33, 17, 32, 34, 33, 18, 33, 34, 33, 17, 34, 34, 33, 16, 32, 33, 34, 18, 33, 33, 34, 32, 32, 32, 32, 34, 32, 32, 32, 32, 34, 32, 32, 34, 34, 32, 32, 32, 32, 34, 32, 34, 32, 34, 32, 32, 34, 34, 32, 34, 34, 34, 32, 33, 32, 33, 33, 32, 33, 33, 33, 34, 33, 33, 33, 32, 32, 32, 34, 34, 32, 32, 34, 32, 34, 32, 34, 34, 34, 32, 34, 33, 33, 33, 34, 32, 32, 34, 34, 34, 32, 34, 34, 32, 34, 34, 34, 34, 34, 34, 34};







































































































































































































































































































































.entry _Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj(
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2,
.param .align 8 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3[48],
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5,
.param .align 4 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6[12],
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15
)
.maxntid 256, 1, 1
.minnctapersm 1
{
.local .align 8 .b8 __local_depot0[16];
.reg .b64 %SP;
.reg .b64 %SPL;
.reg .pred %p<48>;
.reg .b16 %rs<9>;
.reg .f32 %f<169>;
.reg .b32 %r<441>;
.reg .b64 %rd<84>;


mov.u64 %SPL, __local_depot0;
ld.param.u64 %rd9, [_Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0];
ld.param.u64 %rd10, [_Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1];
ld.param.u64 %rd11, [_Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2];
ld.param.u32 %r21, [_Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5];
ld.param.u32 %r26, [_Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8];
ld.param.u32 %r27, [_Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9];
ld.param.u32 %r29, [_Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11];
ld.param.u32 %r30, [_Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12];
ld.param.u32 %r33, [_Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14];
ld.param.u32 %r32, [_Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15];
ld.param.u32 %r24, [_Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+8];
ld.param.u32 %r23, [_Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+4];
ld.param.u32 %r22, [_Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6];
ld.param.u64 %rd13, [_Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+8];
add.u64 %rd1, %SPL, 0;
add.u64 %rd2, %SPL, 8;
mov.u32 %r1, %tid.y;
setp.ge.u32 %p1, %r1, %r33;
@%p1 bra $L__BB0_29;

mov.u32 %r2, %ctaid.y;
cvta.to.global.u64 %rd20, %rd11;
mad.lo.s32 %r34, %r1, %r32, %r2;
mul.wide.u32 %rd21, %r34, 4;
add.s64 %rd22, %rd20, %rd21;
ld.global.nc.u32 %r3, [%rd22];
mov.f32 %f162, 0f00000000;
st.local.v2.f32 [%rd1], {%f162, %f162};
st.local.v2.f32 [%rd2], {%f162, %f162};
mov.u32 %r4, %tid.x;
shr.u32 %r440, %r4, 2;
shr.u32 %r35, %r21, 7;
setp.ge.u32 %p2, %r440, %r35;
mov.f32 %f163, %f162;
mov.f32 %f164, %f162;
mov.f32 %f165, %f162;
@%p2 bra $L__BB0_9;

mul.lo.s32 %r36, %r1, %r27;
cvt.u64.u32 %rd23, %r36;
and.b32 %r37, %r4, 3;
cvt.u64.u32 %rd24, %r37;
add.s64 %rd25, %rd23, %rd24;
mul.hi.u32 %r38, %r2, %r22;
add.s32 %r39, %r38, %r2;
shr.u32 %r40, %r39, %r23;
mul.lo.s32 %r41, %r40, %r24;
sub.s32 %r42, %r2, %r41;
mul.lo.s32 %r43, %r42, %r30;
cvt.u64.u32 %rd26, %r43;
add.s64 %rd4, %rd25, %rd26;
mov.f32 %f162, 0f00000000;
cvta.to.global.u64 %rd27, %rd9;
cvta.to.global.u64 %rd28, %rd10;

$L__BB0_3:
shl.b32 %r77, %r440, 2;
cvt.s64.s32 %rd29, %r77;
mad.lo.s32 %r78, %r3, %r29, %r440;
add.s64 %rd30, %rd4, %rd29;
mul.lo.s64 %rd31, %rd30, 36;
mov.u32 %r79, %ctaid.x;
shl.b32 %r80, %r79, 1;
mad.lo.s32 %r81, %r80, %r26, %r78;
cvt.s64.s32 %rd5, %r81;
mul.wide.s32 %rd32, %r81, 18;
add.s64 %rd33, %rd27, %rd32;
ld.global.nc.u16 %rs1, [%rd33];

{ cvt.f32.f16 %f42, %rs1;}


shl.b32 %r83, %r4, 1;
and.b32 %r84, %r83, 6;
mul.wide.u32 %rd34, %r84, 2;
add.s64 %rd35, %rd33, %rd34;
ld.global.nc.u16 %rs2, [%rd35+2];
cvt.s32.s16 %r85, %rs2;
add.s64 %rd36, %rd28, %rd31;
ld.global.nc.u32 %r7, [%rd36+4];
ld.global.nc.u32 %r8, [%rd36+8];
ld.global.nc.u32 %r9, [%rd36+12];
ld.global.nc.u32 %r10, [%rd36+16];
and.b32 %r86, %r85, 30583;
mov.u32 %r87, 286261504;
prmt.b32 %r88, %r87, %r87, %r86;
shr.u32 %r89, %r85, 2;
and.b32 %r90, %r89, 30583;
prmt.b32 %r91, %r87, %r87, %r90;
and.b32 %r92, %r88, 30583;
mov.u32 %r93, 511;
prmt.b32 %r94, %r93, %r93, %r92;
and.b32 %r95, %r91, 30583;
prmt.b32 %r96, %r93, %r93, %r95;
shr.u32 %r97, %r88, 16;
and.b32 %r98, %r97, 30583;
prmt.b32 %r99, %r93, %r93, %r98;
shr.u32 %r100, %r91, 16;
and.b32 %r101, %r100, 30583;
prmt.b32 %r102, %r93, %r93, %r101;
mov.u32 %r103, 21520;
prmt.b32 %r45, %r94, %r96, %r103;
mov.u32 %r104, 30258;
prmt.b32 %r49, %r94, %r96, %r104;
prmt.b32 %r53, %r99, %r102, %r103;
prmt.b32 %r57, %r99, %r102, %r104;
mov.u32 %r47, 0;

dp4a.s32.s32 %r44, %r45, %r7, %r47;


dp4a.s32.s32 %r48, %r49, %r8, %r44;


dp4a.s32.s32 %r52, %r53, %r9, %r48;


dp4a.s32.s32 %r56, %r57, %r10, %r52;

ld.global.nc.u16 %rs3, [%rd35+4];
cvt.s32.s16 %r105, %rs3;
ld.global.nc.u32 %r11, [%rd36+20];
ld.global.nc.u32 %r12, [%rd36+24];
ld.global.nc.u32 %r13, [%rd36+28];
ld.global.nc.u32 %r14, [%rd36+32];
and.b32 %r106, %r105, 30583;
prmt.b32 %r107, %r87, %r87, %r106;
shr.u32 %r108, %r105, 2;
and.b32 %r109, %r108, 30583;
prmt.b32 %r110, %r87, %r87, %r109;
and.b32 %r111, %r107, 30583;
prmt.b32 %r112, %r93, %r93, %r111;
and.b32 %r113, %r110, 30583;
prmt.b32 %r114, %r93, %r93, %r113;
shr.u32 %r115, %r107, 16;
and.b32 %r116, %r115, 30583;
prmt.b32 %r117, %r93, %r93, %r116;
shr.u32 %r118, %r110, 16;
and.b32 %r119, %r118, 30583;
prmt.b32 %r120, %r93, %r93, %r119;
prmt.b32 %r61, %r112, %r114, %r103;
prmt.b32 %r65, %r112, %r114, %r104;
prmt.b32 %r69, %r117, %r120, %r103;
prmt.b32 %r73, %r117, %r120, %r104;

dp4a.s32.s32 %r60, %r61, %r11, %r56;


dp4a.s32.s32 %r64, %r65, %r12, %r60;


dp4a.s32.s32 %r68, %r69, %r13, %r64;


dp4a.s32.s32 %r72, %r73, %r14, %r68;

ld.global.nc.u32 %r15, [%rd36];

{.reg .f16 low,high;
mov.b32 {low,high},%r15;
cvt.f32.f16 %f43, low;}


mul.ftz.f32 %f44, %f42, %f43;
cvt.rn.f32.s32 %f45, %r72;
fma.rn.ftz.f32 %f165, %f44, %f45, %f165;
setp.eq.s64 %p3, %rd13, 0;
@%p3 bra $L__BB0_5;

cvta.to.global.u64 %rd37, %rd13;
mul.lo.s64 %rd38, %rd5, 18;
add.s64 %rd39, %rd37, %rd38;
ld.global.u16 %rs4, [%rd39];

{ cvt.f32.f16 %f46, %rs4;}


add.s64 %rd41, %rd39, %rd34;
ld.global.s16 %r157, [%rd41+2];
and.b32 %r158, %r157, 30583;
prmt.b32 %r160, %r87, %r87, %r158;
shr.u32 %r161, %r157, 2;
and.b32 %r162, %r161, 30583;
prmt.b32 %r163, %r87, %r87, %r162;
and.b32 %r164, %r160, 30583;
prmt.b32 %r166, %r93, %r93, %r164;
and.b32 %r167, %r163, 30583;
prmt.b32 %r168, %r93, %r93, %r167;
shr.u32 %r169, %r160, 16;
and.b32 %r170, %r169, 30583;
prmt.b32 %r171, %r93, %r93, %r170;
shr.u32 %r172, %r163, 16;
and.b32 %r173, %r172, 30583;
prmt.b32 %r174, %r93, %r93, %r173;
prmt.b32 %r122, %r166, %r168, %r103;
prmt.b32 %r126, %r166, %r168, %r104;
prmt.b32 %r130, %r171, %r174, %r103;
prmt.b32 %r134, %r171, %r174, %r104;

dp4a.s32.s32 %r121, %r122, %r7, %r47;


dp4a.s32.s32 %r125, %r126, %r8, %r121;


dp4a.s32.s32 %r129, %r130, %r9, %r125;


dp4a.s32.s32 %r133, %r134, %r10, %r129;

ld.global.s16 %r177, [%rd41+4];
and.b32 %r178, %r177, 30583;
prmt.b32 %r179, %r87, %r87, %r178;
shr.u32 %r180, %r177, 2;
and.b32 %r181, %r180, 30583;
prmt.b32 %r182, %r87, %r87, %r181;
and.b32 %r183, %r179, 30583;
prmt.b32 %r184, %r93, %r93, %r183;
and.b32 %r185, %r182, 30583;
prmt.b32 %r186, %r93, %r93, %r185;
shr.u32 %r187, %r179, 16;
and.b32 %r188, %r187, 30583;
prmt.b32 %r189, %r93, %r93, %r188;
shr.u32 %r190, %r182, 16;
and.b32 %r191, %r190, 30583;
prmt.b32 %r192, %r93, %r93, %r191;
prmt.b32 %r138, %r184, %r186, %r103;
prmt.b32 %r142, %r184, %r186, %r104;
prmt.b32 %r146, %r189, %r192, %r103;
prmt.b32 %r150, %r189, %r192, %r104;

dp4a.s32.s32 %r137, %r138, %r11, %r133;


dp4a.s32.s32 %r141, %r142, %r12, %r137;


dp4a.s32.s32 %r145, %r146, %r13, %r141;


dp4a.s32.s32 %r149, %r150, %r14, %r145;


{.reg .f16 low,high;
mov.b32 {low,high},%r15;
cvt.f32.f16 %f47, low;}


mul.ftz.f32 %f48, %f46, %f47;
cvt.rn.f32.s32 %f49, %r149;
fma.rn.ftz.f32 %f163, %f48, %f49, %f163;

$L__BB0_5:
add.s32 %r230, %r81, %r26;
mul.wide.s32 %rd43, %r230, 18;
add.s64 %rd44, %rd27, %rd43;
ld.global.nc.u16 %rs5, [%rd44];

{ cvt.f32.f16 %f50, %rs5;}


add.s64 %rd46, %rd44, %rd34;
ld.global.nc.u16 %rs6, [%rd46+2];
cvt.s32.s16 %r234, %rs6;
and.b32 %r235, %r234, 30583;
prmt.b32 %r237, %r87, %r87, %r235;
shr.u32 %r238, %r234, 2;
and.b32 %r239, %r238, 30583;
prmt.b32 %r240, %r87, %r87, %r239;
and.b32 %r241, %r237, 30583;
prmt.b32 %r243, %r93, %r93, %r241;
and.b32 %r244, %r240, 30583;
prmt.b32 %r245, %r93, %r93, %r244;
shr.u32 %r246, %r237, 16;
and.b32 %r247, %r246, 30583;
prmt.b32 %r248, %r93, %r93, %r247;
shr.u32 %r249, %r240, 16;
and.b32 %r250, %r249, 30583;
prmt.b32 %r251, %r93, %r93, %r250;
prmt.b32 %r194, %r243, %r245, %r103;
prmt.b32 %r198, %r243, %r245, %r104;
prmt.b32 %r202, %r248, %r251, %r103;
prmt.b32 %r206, %r248, %r251, %r104;

dp4a.s32.s32 %r193, %r194, %r7, %r47;


dp4a.s32.s32 %r197, %r198, %r8, %r193;


dp4a.s32.s32 %r201, %r202, %r9, %r197;


dp4a.s32.s32 %r205, %r206, %r10, %r201;

ld.global.nc.u16 %rs7, [%rd46+4];
cvt.s32.s16 %r254, %rs7;
and.b32 %r255, %r254, 30583;
prmt.b32 %r256, %r87, %r87, %r255;
shr.u32 %r257, %r254, 2;
and.b32 %r258, %r257, 30583;
prmt.b32 %r259, %r87, %r87, %r258;
and.b32 %r260, %r256, 30583;
prmt.b32 %r261, %r93, %r93, %r260;
and.b32 %r262, %r259, 30583;
prmt.b32 %r263, %r93, %r93, %r262;
shr.u32 %r264, %r256, 16;
and.b32 %r265, %r264, 30583;
prmt.b32 %r266, %r93, %r93, %r265;
shr.u32 %r267, %r259, 16;
and.b32 %r268, %r267, 30583;
prmt.b32 %r269, %r93, %r93, %r268;
prmt.b32 %r210, %r261, %r263, %r103;
prmt.b32 %r214, %r261, %r263, %r104;
prmt.b32 %r218, %r266, %r269, %r103;
prmt.b32 %r222, %r266, %r269, %r104;

dp4a.s32.s32 %r209, %r210, %r11, %r205;


dp4a.s32.s32 %r213, %r214, %r12, %r209;


dp4a.s32.s32 %r217, %r218, %r13, %r213;


dp4a.s32.s32 %r221, %r222, %r14, %r217;


{.reg .f16 low,high;
mov.b32 {low,high},%r15;
cvt.f32.f16 %f51, low;}


mul.ftz.f32 %f52, %f50, %f51;
cvt.rn.f32.s32 %f53, %r221;
fma.rn.ftz.f32 %f164, %f52, %f53, %f164;
@%p3 bra $L__BB0_7;

cvta.to.global.u64 %rd47, %rd13;
add.s64 %rd49, %rd47, %rd43;
ld.global.u16 %rs8, [%rd49];

{ cvt.f32.f16 %f54, %rs8;}


add.s64 %rd51, %rd49, %rd34;
ld.global.s16 %r311, [%rd51+2];
and.b32 %r312, %r311, 30583;
mov.u32 %r313, 286261504;
prmt.b32 %r314, %r313, %r313, %r312;
shr.u32 %r315, %r311, 2;
and.b32 %r316, %r315, 30583;
prmt.b32 %r317, %r313, %r313, %r316;
and.b32 %r318, %r314, 30583;
mov.u32 %r319, 511;
prmt.b32 %r320, %r319, %r319, %r318;
and.b32 %r321, %r317, 30583;
prmt.b32 %r322, %r319, %r319, %r321;
shr.u32 %r323, %r314, 16;
and.b32 %r324, %r323, 30583;
prmt.b32 %r325, %r319, %r319, %r324;
shr.u32 %r326, %r317, 16;
and.b32 %r327, %r326, 30583;
prmt.b32 %r328, %r319, %r319, %r327;
mov.u32 %r329, 21520;
prmt.b32 %r271, %r320, %r322, %r329;
mov.u32 %r330, 30258;
prmt.b32 %r275, %r320, %r322, %r330;
prmt.b32 %r279, %r325, %r328, %r329;
prmt.b32 %r283, %r325, %r328, %r330;
mov.u32 %r273, 0;

dp4a.s32.s32 %r270, %r271, %r7, %r273;


dp4a.s32.s32 %r274, %r275, %r8, %r270;


dp4a.s32.s32 %r278, %r279, %r9, %r274;


dp4a.s32.s32 %r282, %r283, %r10, %r278;

ld.global.s16 %r331, [%rd51+4];
and.b32 %r332, %r331, 30583;
prmt.b32 %r333, %r313, %r313, %r332;
shr.u32 %r334, %r331, 2;
and.b32 %r335, %r334, 30583;
prmt.b32 %r336, %r313, %r313, %r335;
and.b32 %r337, %r333, 30583;
prmt.b32 %r338, %r319, %r319, %r337;
and.b32 %r339, %r336, 30583;
prmt.b32 %r340, %r319, %r319, %r339;
shr.u32 %r341, %r333, 16;
and.b32 %r342, %r341, 30583;
prmt.b32 %r343, %r319, %r319, %r342;
shr.u32 %r344, %r336, 16;
and.b32 %r345, %r344, 30583;
prmt.b32 %r346, %r319, %r319, %r345;
prmt.b32 %r287, %r338, %r340, %r329;
prmt.b32 %r291, %r338, %r340, %r330;
prmt.b32 %r295, %r343, %r346, %r329;
prmt.b32 %r299, %r343, %r346, %r330;

dp4a.s32.s32 %r286, %r287, %r11, %r282;


dp4a.s32.s32 %r290, %r291, %r12, %r286;


dp4a.s32.s32 %r294, %r295, %r13, %r290;


dp4a.s32.s32 %r298, %r299, %r14, %r294;


{.reg .f16 low,high;
mov.b32 {low,high},%r15;
cvt.f32.f16 %f55, low;}


mul.ftz.f32 %f56, %f54, %f55;
cvt.rn.f32.s32 %f57, %r298;
fma.rn.ftz.f32 %f162, %f56, %f57, %f162;

$L__BB0_7:
add.s32 %r440, %r440, 8;
setp.lt.u32 %p5, %r440, %r35;
@%p5 bra $L__BB0_3;

add.u64 %rd76, %SPL, 0;
add.u64 %rd65, %SPL, 8;
st.local.v2.f32 [%rd76], {%f165, %f164};
st.local.v2.f32 [%rd65], {%f163, %f162};

$L__BB0_9:
mov.b32 %r17, %f165;
setp.ne.s64 %p6, %rd13, 0;
@%p6 bra $L__BB0_11;
bra.uni $L__BB0_10;

$L__BB0_11:
add.u64 %rd82, %SPL, 0;
add.u64 %rd74, %SPL, 8;
mov.u32 %r374, 31;
mov.u32 %r375, 16;
mov.u32 %r376, -1;
shfl.sync.bfly.b32 %r377|%p17, %r17, %r375, %r374, %r376;
mov.b32 %f78, %r377;
add.ftz.f32 %f79, %f165, %f78;
mov.b32 %r378, %f79;
mov.u32 %r379, 8;
shfl.sync.bfly.b32 %r380|%p18, %r378, %r379, %r374, %r376;
mov.b32 %f80, %r380;
add.ftz.f32 %f81, %f79, %f80;
mov.b32 %r381, %f81;
mov.u32 %r382, 4;
shfl.sync.bfly.b32 %r383|%p19, %r381, %r382, %r374, %r376;
mov.b32 %f82, %r383;
add.ftz.f32 %f83, %f81, %f82;
mov.b32 %r384, %f83;
mov.u32 %r385, 2;
shfl.sync.bfly.b32 %r386|%p20, %r384, %r385, %r374, %r376;
mov.b32 %f84, %r386;
add.ftz.f32 %f85, %f83, %f84;
mov.b32 %r387, %f85;
mov.u32 %r388, 1;
shfl.sync.bfly.b32 %r389|%p21, %r387, %r388, %r374, %r376;
mov.b32 %f86, %r389;
add.ftz.f32 %f87, %f85, %f86;
st.local.f32 [%rd82], %f87;
mov.b32 %r390, %f163;
shfl.sync.bfly.b32 %r391|%p22, %r390, %r375, %r374, %r376;
mov.b32 %f88, %r391;
add.ftz.f32 %f89, %f163, %f88;
mov.b32 %r392, %f89;
shfl.sync.bfly.b32 %r393|%p23, %r392, %r379, %r374, %r376;
mov.b32 %f90, %r393;
add.ftz.f32 %f91, %f89, %f90;
mov.b32 %r394, %f91;
shfl.sync.bfly.b32 %r395|%p24, %r394, %r382, %r374, %r376;
mov.b32 %f92, %r395;
add.ftz.f32 %f93, %f91, %f92;
mov.b32 %r396, %f93;
shfl.sync.bfly.b32 %r397|%p25, %r396, %r385, %r374, %r376;
mov.b32 %f94, %r397;
add.ftz.f32 %f95, %f93, %f94;
mov.b32 %r398, %f95;
shfl.sync.bfly.b32 %r399|%p26, %r398, %r388, %r374, %r376;
mov.b32 %f96, %r399;
add.ftz.f32 %f97, %f95, %f96;
st.local.f32 [%rd74], %f97;
mov.b32 %r400, %f164;
shfl.sync.bfly.b32 %r401|%p27, %r400, %r375, %r374, %r376;
mov.b32 %f98, %r401;
add.ftz.f32 %f99, %f164, %f98;
mov.b32 %r402, %f99;
shfl.sync.bfly.b32 %r403|%p28, %r402, %r379, %r374, %r376;
mov.b32 %f100, %r403;
add.ftz.f32 %f101, %f99, %f100;
mov.b32 %r404, %f101;
shfl.sync.bfly.b32 %r405|%p29, %r404, %r382, %r374, %r376;
mov.b32 %f102, %r405;
add.ftz.f32 %f103, %f101, %f102;
mov.b32 %r406, %f103;
shfl.sync.bfly.b32 %r407|%p30, %r406, %r385, %r374, %r376;
mov.b32 %f104, %r407;
add.ftz.f32 %f105, %f103, %f104;
mov.b32 %r408, %f105;
shfl.sync.bfly.b32 %r409|%p31, %r408, %r388, %r374, %r376;
mov.b32 %f106, %r409;
add.ftz.f32 %f107, %f105, %f106;
st.local.f32 [%rd82+4], %f107;
mov.b32 %r410, %f162;
shfl.sync.bfly.b32 %r411|%p32, %r410, %r375, %r374, %r376;
mov.b32 %f108, %r411;
add.ftz.f32 %f109, %f162, %f108;
mov.b32 %r412, %f109;
shfl.sync.bfly.b32 %r413|%p33, %r412, %r379, %r374, %r376;
mov.b32 %f110, %r413;
add.ftz.f32 %f111, %f109, %f110;
mov.b32 %r414, %f111;
shfl.sync.bfly.b32 %r415|%p34, %r414, %r382, %r374, %r376;
mov.b32 %f112, %r415;
add.ftz.f32 %f113, %f111, %f112;
mov.b32 %r416, %f113;
shfl.sync.bfly.b32 %r417|%p35, %r416, %r385, %r374, %r376;
mov.b32 %f114, %r417;
add.ftz.f32 %f115, %f113, %f114;
mov.b32 %r418, %f115;
shfl.sync.bfly.b32 %r419|%p36, %r418, %r388, %r374, %r376;
mov.b32 %f116, %r419;
add.ftz.f32 %f117, %f115, %f116;
st.local.f32 [%rd74+4], %f117;
bra.uni $L__BB0_12;

$L__BB0_10:
add.u64 %rd78, %SPL, 0;
mov.u32 %r348, 31;
mov.u32 %r349, 16;
mov.u32 %r350, -1;
shfl.sync.bfly.b32 %r351|%p7, %r17, %r349, %r348, %r350;
mov.b32 %f58, %r351;
add.ftz.f32 %f59, %f165, %f58;
mov.b32 %r352, %f59;
mov.u32 %r353, 8;
shfl.sync.bfly.b32 %r354|%p8, %r352, %r353, %r348, %r350;
mov.b32 %f60, %r354;
add.ftz.f32 %f61, %f59, %f60;
mov.b32 %r355, %f61;
mov.u32 %r356, 4;
shfl.sync.bfly.b32 %r357|%p9, %r355, %r356, %r348, %r350;
mov.b32 %f62, %r357;
add.ftz.f32 %f63, %f61, %f62;
mov.b32 %r358, %f63;
mov.u32 %r359, 2;
shfl.sync.bfly.b32 %r360|%p10, %r358, %r359, %r348, %r350;
mov.b32 %f64, %r360;
add.ftz.f32 %f65, %f63, %f64;
mov.b32 %r361, %f65;
mov.u32 %r362, 1;
shfl.sync.bfly.b32 %r363|%p11, %r361, %r362, %r348, %r350;
mov.b32 %f66, %r363;
add.ftz.f32 %f67, %f65, %f66;
st.local.f32 [%rd78], %f67;
mov.b32 %r364, %f164;
shfl.sync.bfly.b32 %r365|%p12, %r364, %r349, %r348, %r350;
mov.b32 %f68, %r365;
add.ftz.f32 %f69, %f164, %f68;
mov.b32 %r366, %f69;
shfl.sync.bfly.b32 %r367|%p13, %r366, %r353, %r348, %r350;
mov.b32 %f70, %r367;
add.ftz.f32 %f71, %f69, %f70;
mov.b32 %r368, %f71;
shfl.sync.bfly.b32 %r369|%p14, %r368, %r356, %r348, %r350;
mov.b32 %f72, %r369;
add.ftz.f32 %f73, %f71, %f72;
mov.b32 %r370, %f73;
shfl.sync.bfly.b32 %r371|%p15, %r370, %r359, %r348, %r350;
mov.b32 %f74, %r371;
add.ftz.f32 %f75, %f73, %f74;
mov.b32 %r372, %f75;
shfl.sync.bfly.b32 %r373|%p16, %r372, %r362, %r348, %r350;
mov.b32 %f76, %r373;
add.ftz.f32 %f77, %f75, %f76;
st.local.f32 [%rd78+4], %f77;

$L__BB0_12:
ld.param.u32 %r429, [_Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7];
setp.gt.u32 %p37, %r4, 1;
mov.u32 %r421, %ctaid.x;
shl.b32 %r422, %r421, 1;
add.s32 %r18, %r422, %r4;
setp.ge.u32 %p38, %r18, %r429;
or.pred %p39, %p37, %p38;
@%p39 bra $L__BB0_29;

add.u64 %rd80, %SPL, 0;
ld.param.u64 %rd67, [_Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3];
ld.param.u32 %r430, [_Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13];
cvt.u64.u32 %rd7, %r4;
mul.wide.u32 %rd52, %r4, 4;
add.s64 %rd53, %rd80, %rd52;
ld.local.f32 %f168, [%rd53];
mad.lo.s32 %r19, %r3, %r430, %r18;
setp.eq.s64 %p40, %rd67, 0;
@%p40 bra $L__BB0_15;

ld.param.u64 %rd68, [_Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3];
cvta.to.global.u64 %rd54, %rd68;
mul.wide.u32 %rd55, %r19, 4;
add.s64 %rd56, %rd54, %rd55;
ld.global.f32 %f118, [%rd56];
add.ftz.f32 %f168, %f168, %f118;

$L__BB0_15:
setp.eq.s64 %p41, %rd13, 0;
@%p41 bra $L__BB0_28;

ld.param.u64 %rd72, [_Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+16];
add.u64 %rd70, %SPL, 8;
shl.b64 %rd57, %rd7, 2;
add.s64 %rd58, %rd70, %rd57;
ld.local.f32 %f24, [%rd58];
setp.eq.s64 %p42, %rd72, 0;
@%p42 bra $L__BB0_18;

ld.param.u64 %rd73, [_Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+16];
cvta.to.global.u64 %rd59, %rd73;
mul.wide.u32 %rd60, %r19, 4;
add.s64 %rd61, %rd59, %rd60;
ld.global.f32 %f119, [%rd61];
add.ftz.f32 %f24, %f24, %f119;

$L__BB0_18:
ld.param.u32 %r435, [_Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.gt.s32 %p43, %r435, 2;
@%p43 bra $L__BB0_22;

ld.param.u32 %r438, [_Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.eq.s32 %p46, %r438, 1;
@%p46 bra $L__BB0_26;

ld.param.u32 %r439, [_Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.eq.s32 %p47, %r439, 2;
@%p47 bra $L__BB0_21;
bra.uni $L__BB0_27;

$L__BB0_21:
mul.ftz.f32 %f139, %f24, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f140, %f139;
add.ftz.f32 %f141, %f140, 0f3F800000;
div.approx.ftz.f32 %f142, %f24, %f141;
mul.ftz.f32 %f168, %f168, %f142;
bra.uni $L__BB0_28;

$L__BB0_22:
ld.param.u32 %r436, [_Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.eq.s32 %p44, %r436, 3;
@%p44 bra $L__BB0_25;

ld.param.u32 %r437, [_Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.ne.s32 %p45, %r437, 6;
@%p45 bra $L__BB0_27;

ld.param.f32 %f151, [_Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+44];
min.ftz.f32 %f120, %f24, %f151;
neg.ftz.f32 %f121, %f151;
min.ftz.f32 %f122, %f168, %f151;
max.ftz.f32 %f123, %f122, %f121;
mul.ftz.f32 %f124, %f120, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f125, %f124;
add.ftz.f32 %f126, %f125, 0f3F800000;
div.approx.ftz.f32 %f127, %f120, %f126;
mul.ftz.f32 %f168, %f123, %f127;
bra.uni $L__BB0_28;

$L__BB0_27:
mul.ftz.f32 %f168, %f168, %f24;
bra.uni $L__BB0_28;

$L__BB0_26:
mul.ftz.f32 %f143, %f24, 0f3F000000;
mul.ftz.f32 %f144, %f24, 0f3D372713;
fma.rn.ftz.f32 %f145, %f24, %f144, 0f3F800000;
mul.ftz.f32 %f146, %f24, 0f3F4C422A;
mul.ftz.f32 %f147, %f146, %f145;
tanh.approx.f32 %f148, %f147;
add.ftz.f32 %f149, %f148, 0f3F800000;
mul.ftz.f32 %f150, %f143, %f149;
mul.ftz.f32 %f168, %f168, %f150;
bra.uni $L__BB0_28;

$L__BB0_25:
mov.f32 %f128, 0f40E00000;
min.ftz.f32 %f129, %f24, %f128;
min.ftz.f32 %f130, %f168, %f128;
mov.f32 %f131, 0fC0E00000;
max.ftz.f32 %f132, %f130, %f131;
mul.ftz.f32 %f133, %f129, 0f3FD9DB23;
mul.ftz.f32 %f134, %f133, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f135, %f134;
add.ftz.f32 %f136, %f135, 0f3F800000;
div.approx.ftz.f32 %f137, %f129, %f136;
add.ftz.f32 %f138, %f132, 0f3F800000;
mul.ftz.f32 %f168, %f138, %f137;

$L__BB0_28:
ld.param.u64 %rd69, [_Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4];
ld.param.u32 %r434, [_Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10];
ld.param.u32 %r433, [_Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13];
mov.u32 %r432, %ctaid.y;
mov.u32 %r431, %tid.y;
mad.lo.s32 %r427, %r431, %r434, %r18;
mad.lo.s32 %r428, %r432, %r433, %r427;
cvta.to.global.u64 %rd62, %rd69;
mul.wide.u32 %rd63, %r428, 4;
add.s64 %rd64, %rd62, %rd63;
st.global.f32 [%rd64], %f168;

$L__BB0_29:
ret;

}
.entry _Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj(
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2,
.param .align 8 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3[48],
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5,
.param .align 4 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6[12],
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15
)
.maxntid 256, 1, 1
.minnctapersm 1
{
.local .align 8 .b8 __local_depot1[8];
.reg .b64 %SP;
.reg .b64 %SPL;
.reg .pred %p<17>;
.reg .b16 %rs<7>;
.reg .f32 %f<44>;
.reg .b32 %r<214>;
.reg .b64 %rd<50>;


mov.u64 %SPL, __local_depot1;
ld.param.u64 %rd17, [_Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0];
ld.param.u64 %rd18, [_Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1];
ld.param.u64 %rd19, [_Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2];
ld.param.u64 %rd20, [_Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4];
ld.param.u32 %r12, [_Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5];
ld.param.u32 %r16, [_Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7];
ld.param.u32 %r17, [_Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8];
ld.param.u32 %r18, [_Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9];
ld.param.u32 %r19, [_Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10];
ld.param.u32 %r20, [_Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11];
ld.param.u32 %r21, [_Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12];
ld.param.u32 %r22, [_Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13];
ld.param.u32 %r24, [_Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14];
ld.param.u32 %r23, [_Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15];
ld.param.u32 %r15, [_Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+8];
ld.param.u32 %r14, [_Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+4];
ld.param.u32 %r13, [_Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6];
add.u64 %rd1, %SPL, 0;
mov.u32 %r25, %ctaid.x;
shl.b32 %r1, %r25, 1;
mov.u32 %r2, %tid.y;
setp.ge.u32 %p1, %r2, %r24;
@%p1 bra $L__BB1_7;

mov.f32 %f42, 0f00000000;
st.local.v2.f32 [%rd1], {%f42, %f42};
mov.u32 %r3, %tid.x;
shr.u32 %r213, %r3, 2;
shr.u32 %r27, %r12, 7;
setp.ge.u32 %p2, %r213, %r27;
mov.f32 %f43, %f42;
@%p2 bra $L__BB1_5;

mov.u32 %r28, %ctaid.y;
mad.lo.s32 %r29, %r2, %r23, %r28;
cvta.to.global.u64 %rd22, %rd19;
mul.wide.u32 %rd23, %r29, 4;
add.s64 %rd24, %rd22, %rd23;
ld.global.nc.u32 %r30, [%rd24];
mul.lo.s32 %r31, %r30, %r20;
mul.lo.s32 %r33, %r2, %r18;
cvt.u64.u32 %rd25, %r33;
and.b32 %r35, %r3, 3;
cvt.u64.u32 %rd26, %r35;
add.s64 %rd27, %rd25, %rd26;
mul.hi.u32 %r36, %r28, %r13;
add.s32 %r37, %r36, %r28;
shr.u32 %r38, %r37, %r14;
mul.lo.s32 %r39, %r38, %r15;
sub.s32 %r40, %r28, %r39;
mul.lo.s32 %r41, %r40, %r21;
cvt.u64.u32 %rd28, %r41;
add.s64 %rd3, %rd27, %rd28;
shl.b32 %r212, %r213, 2;
mul.wide.u32 %rd29, %r35, 4;
mad.lo.s32 %r42, %r1, %r17, %r213;
add.s32 %r43, %r42, %r31;
mul.wide.s32 %rd30, %r43, 18;
add.s64 %rd31, %rd29, %rd30;
cvta.to.global.u64 %rd32, %rd17;
add.s64 %rd33, %rd32, %rd31;
add.s64 %rd49, %rd33, 4;
add.s32 %r44, %r213, %r31;
add.s32 %r45, %r1, 1;
mad.lo.s32 %r46, %r17, %r45, %r44;
mul.wide.s32 %rd34, %r46, 18;
add.s64 %rd35, %rd29, %rd34;
add.s64 %rd36, %rd32, %rd35;
add.s64 %rd48, %rd36, 4;
add.s64 %rd47, %rd32, %rd34;
add.s64 %rd46, %rd32, %rd30;
cvta.to.global.u64 %rd8, %rd18;
mov.f32 %f42, 0f00000000;

$L__BB1_3:
cvt.s64.s32 %rd37, %r212;
add.s64 %rd38, %rd3, %rd37;
mul.lo.s64 %rd39, %rd38, 36;
ld.global.nc.u16 %rs1, [%rd46];

{ cvt.f32.f16 %f11, %rs1;}


ld.global.nc.u16 %rs3, [%rd49+-2];
cvt.s32.s16 %r113, %rs3;
add.s64 %rd40, %rd8, %rd39;
ld.global.nc.u32 %r49, [%rd40+4];
ld.global.nc.u32 %r53, [%rd40+8];
ld.global.nc.u32 %r57, [%rd40+12];
ld.global.nc.u32 %r61, [%rd40+16];
and.b32 %r114, %r113, 30583;
mov.u32 %r115, 286261504;
prmt.b32 %r116, %r115, %r115, %r114;
shr.u32 %r117, %r113, 2;
and.b32 %r118, %r117, 30583;
prmt.b32 %r119, %r115, %r115, %r118;
and.b32 %r120, %r116, 30583;
mov.u32 %r121, 511;
prmt.b32 %r122, %r121, %r121, %r120;
and.b32 %r123, %r119, 30583;
prmt.b32 %r124, %r121, %r121, %r123;
shr.u32 %r125, %r116, 16;
and.b32 %r126, %r125, 30583;
prmt.b32 %r127, %r121, %r121, %r126;
shr.u32 %r128, %r119, 16;
and.b32 %r129, %r128, 30583;
prmt.b32 %r130, %r121, %r121, %r129;
mov.u32 %r131, 21520;
prmt.b32 %r48, %r122, %r124, %r131;
mov.u32 %r132, 30258;
prmt.b32 %r52, %r122, %r124, %r132;
prmt.b32 %r56, %r127, %r130, %r131;
prmt.b32 %r60, %r127, %r130, %r132;
mov.u32 %r83, 0;

dp4a.s32.s32 %r47, %r48, %r49, %r83;


dp4a.s32.s32 %r51, %r52, %r53, %r47;


dp4a.s32.s32 %r55, %r56, %r57, %r51;


dp4a.s32.s32 %r59, %r60, %r61, %r55;

ld.global.nc.u16 %rs4, [%rd49];
cvt.s32.s16 %r133, %rs4;
ld.global.nc.u32 %r65, [%rd40+20];
ld.global.nc.u32 %r69, [%rd40+24];
ld.global.nc.u32 %r73, [%rd40+28];
ld.global.nc.u32 %r77, [%rd40+32];
and.b32 %r134, %r133, 30583;
prmt.b32 %r135, %r115, %r115, %r134;
shr.u32 %r136, %r133, 2;
and.b32 %r137, %r136, 30583;
prmt.b32 %r138, %r115, %r115, %r137;
and.b32 %r139, %r135, 30583;
prmt.b32 %r140, %r121, %r121, %r139;
and.b32 %r141, %r138, 30583;
prmt.b32 %r142, %r121, %r121, %r141;
shr.u32 %r143, %r135, 16;
and.b32 %r144, %r143, 30583;
prmt.b32 %r145, %r121, %r121, %r144;
shr.u32 %r146, %r138, 16;
and.b32 %r147, %r146, 30583;
prmt.b32 %r148, %r121, %r121, %r147;
prmt.b32 %r64, %r140, %r142, %r131;
prmt.b32 %r68, %r140, %r142, %r132;
prmt.b32 %r72, %r145, %r148, %r131;
prmt.b32 %r76, %r145, %r148, %r132;

dp4a.s32.s32 %r63, %r64, %r65, %r59;


dp4a.s32.s32 %r67, %r68, %r69, %r63;


dp4a.s32.s32 %r71, %r72, %r73, %r67;


dp4a.s32.s32 %r75, %r76, %r77, %r71;

ld.global.nc.u32 %r79, [%rd40];

{.reg .f16 low,high;
mov.b32 {low,high},%r79;
cvt.f32.f16 %f12, low;}


mul.ftz.f32 %f15, %f11, %f12;
cvt.rn.f32.s32 %f16, %r75;
fma.rn.ftz.f32 %f43, %f15, %f16, %f43;
ld.global.nc.u16 %rs2, [%rd47];

{ cvt.f32.f16 %f13, %rs2;}


ld.global.nc.u16 %rs5, [%rd48+-2];
cvt.s32.s16 %r149, %rs5;
and.b32 %r150, %r149, 30583;
prmt.b32 %r151, %r115, %r115, %r150;
shr.u32 %r152, %r149, 2;
and.b32 %r153, %r152, 30583;
prmt.b32 %r154, %r115, %r115, %r153;
and.b32 %r155, %r151, 30583;
prmt.b32 %r156, %r121, %r121, %r155;
and.b32 %r157, %r154, 30583;
prmt.b32 %r158, %r121, %r121, %r157;
shr.u32 %r159, %r151, 16;
and.b32 %r160, %r159, 30583;
prmt.b32 %r161, %r121, %r121, %r160;
shr.u32 %r162, %r154, 16;
and.b32 %r163, %r162, 30583;
prmt.b32 %r164, %r121, %r121, %r163;
prmt.b32 %r81, %r156, %r158, %r131;
prmt.b32 %r85, %r156, %r158, %r132;
prmt.b32 %r89, %r161, %r164, %r131;
prmt.b32 %r93, %r161, %r164, %r132;

dp4a.s32.s32 %r80, %r81, %r49, %r83;


dp4a.s32.s32 %r84, %r85, %r53, %r80;


dp4a.s32.s32 %r88, %r89, %r57, %r84;


dp4a.s32.s32 %r92, %r93, %r61, %r88;

ld.global.nc.u16 %rs6, [%rd48];
cvt.s32.s16 %r165, %rs6;
and.b32 %r166, %r165, 30583;
prmt.b32 %r167, %r115, %r115, %r166;
shr.u32 %r168, %r165, 2;
and.b32 %r169, %r168, 30583;
prmt.b32 %r170, %r115, %r115, %r169;
and.b32 %r171, %r167, 30583;
prmt.b32 %r172, %r121, %r121, %r171;
and.b32 %r173, %r170, 30583;
prmt.b32 %r174, %r121, %r121, %r173;
shr.u32 %r175, %r167, 16;
and.b32 %r176, %r175, 30583;
prmt.b32 %r177, %r121, %r121, %r176;
shr.u32 %r178, %r170, 16;
and.b32 %r179, %r178, 30583;
prmt.b32 %r180, %r121, %r121, %r179;
prmt.b32 %r97, %r172, %r174, %r131;
prmt.b32 %r101, %r172, %r174, %r132;
prmt.b32 %r105, %r177, %r180, %r131;
prmt.b32 %r109, %r177, %r180, %r132;

dp4a.s32.s32 %r96, %r97, %r65, %r92;


dp4a.s32.s32 %r100, %r101, %r69, %r96;


dp4a.s32.s32 %r104, %r105, %r73, %r100;


dp4a.s32.s32 %r108, %r109, %r77, %r104;


{.reg .f16 low,high;
mov.b32 {low,high},%r79;
cvt.f32.f16 %f14, low;}


mul.ftz.f32 %f17, %f13, %f14;
cvt.rn.f32.s32 %f18, %r108;
fma.rn.ftz.f32 %f42, %f17, %f18, %f42;
add.s32 %r212, %r212, 32;
add.s64 %rd49, %rd49, 144;
add.s64 %rd48, %rd48, 144;
add.s64 %rd47, %rd47, 144;
add.s64 %rd46, %rd46, 144;
add.s32 %r213, %r213, 8;
setp.lt.u32 %p3, %r213, %r27;
@%p3 bra $L__BB1_3;

st.local.v2.f32 [%rd1], {%f43, %f42};

$L__BB1_5:
mov.b32 %r181, %f43;
mov.u32 %r182, 31;
mov.u32 %r183, 16;
mov.u32 %r184, -1;
shfl.sync.bfly.b32 %r185|%p4, %r181, %r183, %r182, %r184;
mov.b32 %f19, %r185;
add.ftz.f32 %f20, %f43, %f19;
mov.b32 %r186, %f20;
mov.u32 %r187, 8;
shfl.sync.bfly.b32 %r188|%p5, %r186, %r187, %r182, %r184;
mov.b32 %f21, %r188;
add.ftz.f32 %f22, %f20, %f21;
mov.b32 %r189, %f22;
mov.u32 %r190, 4;
shfl.sync.bfly.b32 %r191|%p6, %r189, %r190, %r182, %r184;
mov.b32 %f23, %r191;
add.ftz.f32 %f24, %f22, %f23;
mov.b32 %r192, %f24;
mov.u32 %r193, 2;
shfl.sync.bfly.b32 %r194|%p7, %r192, %r193, %r182, %r184;
mov.b32 %f25, %r194;
add.ftz.f32 %f26, %f24, %f25;
mov.b32 %r195, %f26;
mov.u32 %r196, 1;
shfl.sync.bfly.b32 %r197|%p8, %r195, %r196, %r182, %r184;
mov.b32 %f27, %r197;
add.ftz.f32 %f28, %f26, %f27;
st.local.f32 [%rd1], %f28;
mov.b32 %r198, %f42;
shfl.sync.bfly.b32 %r199|%p9, %r198, %r183, %r182, %r184;
mov.b32 %f29, %r199;
add.ftz.f32 %f30, %f42, %f29;
mov.b32 %r200, %f30;
shfl.sync.bfly.b32 %r201|%p10, %r200, %r187, %r182, %r184;
mov.b32 %f31, %r201;
add.ftz.f32 %f32, %f30, %f31;
mov.b32 %r202, %f32;
shfl.sync.bfly.b32 %r203|%p11, %r202, %r190, %r182, %r184;
mov.b32 %f33, %r203;
add.ftz.f32 %f34, %f32, %f33;
mov.b32 %r204, %f34;
shfl.sync.bfly.b32 %r205|%p12, %r204, %r193, %r182, %r184;
mov.b32 %f35, %r205;
add.ftz.f32 %f36, %f34, %f35;
mov.b32 %r206, %f36;
shfl.sync.bfly.b32 %r207|%p13, %r206, %r196, %r182, %r184;
mov.b32 %f37, %r207;
add.ftz.f32 %f38, %f36, %f37;
st.local.f32 [%rd1+4], %f38;
add.s32 %r11, %r1, %r3;
setp.ge.u32 %p14, %r11, %r16;
setp.gt.u32 %p15, %r3, 1;
or.pred %p16, %p15, %p14;
@%p16 bra $L__BB1_7;

mul.wide.u32 %rd41, %r3, 4;
add.s64 %rd42, %rd1, %rd41;
ld.local.f32 %f39, [%rd42];
mov.u32 %r208, %ctaid.y;
mad.lo.s32 %r210, %r2, %r19, %r11;
mad.lo.s32 %r211, %r208, %r22, %r210;
cvta.to.global.u64 %rd43, %rd20;
mul.wide.u32 %rd44, %r211, 4;
add.s64 %rd45, %rd43, %rd44;
st.global.f32 [%rd45], %f39;

$L__BB1_7:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<320>;
.reg .b16 %rs<17>;
.reg .f32 %f<1005>;
.reg .b32 %r<1417>;
.reg .b64 %rd<94>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate[1536];

ld.param.u64 %rd24, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd25, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd26, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u32 %r47, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r51, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r52, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r56, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r57, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r58, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r62, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r63, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r64, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r60, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r59, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r54, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r53, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r50, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r49, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r48, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
ld.param.u64 %rd29, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+16];
ld.param.u64 %rd28, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+8];
ld.param.u64 %rd27, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3];
mov.u32 %r1, %tid.y;
shl.b32 %r65, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r65, %r2;
mov.u32 %r66, %ctaid.x;
shl.b32 %r4, %r66, 2;
mov.u32 %r5, %ctaid.y;
setp.eq.s64 %p1, %rd26, 0;
@%p1 bra $L__BB2_2;

cvta.to.global.u64 %rd33, %rd26;
mul.wide.u32 %rd34, %r5, 4;
add.s64 %rd35, %rd33, %rd34;
ld.global.nc.u32 %r1413, [%rd35];
mul.hi.u32 %r67, %r5, %r48;
add.s32 %r68, %r67, %r5;
shr.u32 %r69, %r68, %r49;
mul.lo.s32 %r70, %r69, %r50;
sub.s32 %r1414, %r5, %r70;
bra.uni $L__BB2_3;

$L__BB2_2:
mul.hi.u32 %r71, %r5, %r53;
add.s32 %r72, %r71, %r5;
shr.u32 %r1413, %r72, %r54;
mov.u32 %r1414, %r5;

$L__BB2_3:
mov.u32 %r11, %ctaid.z;
selp.b32 %r12, %r5, %r1413, %p1;
setp.ne.s32 %p3, %r1, 0;
setp.gt.u32 %p4, %r2, 3;
or.pred %p5, %p4, %p3;
add.s32 %r73, %r4, %r2;
setp.ge.u32 %p6, %r73, %r52;
mov.f32 %f26, 0f00000000;
or.pred %p7, %p5, %p6;
mov.f32 %f4, %f26;
mov.f32 %f5, %f26;
@%p7 bra $L__BB2_8;

setp.eq.s64 %p8, %rd27, 0;
mov.f32 %f4, 0f00000000;
mov.f32 %f5, %f4;
@%p8 bra $L__BB2_6;

cvta.to.global.u64 %rd36, %rd27;
mul.lo.s32 %r74, %r11, %r64;
cvt.u64.u32 %rd37, %r74;
mul.lo.s32 %r75, %r12, %r58;
cvt.u64.u32 %rd38, %r75;
cvt.u64.u32 %rd39, %r2;
cvt.s64.s32 %rd40, %r4;
add.s64 %rd41, %rd40, %rd39;
add.s64 %rd42, %rd41, %rd37;
add.s64 %rd43, %rd42, %rd38;
shl.b64 %rd44, %rd43, 2;
add.s64 %rd45, %rd36, %rd44;
ld.global.f32 %f5, [%rd45];

$L__BB2_6:
setp.eq.s64 %p9, %rd29, 0;
setp.eq.s64 %p10, %rd28, 0;
or.pred %p11, %p9, %p10;
@%p11 bra $L__BB2_8;

cvta.to.global.u64 %rd46, %rd29;
mul.lo.s32 %r76, %r11, %r64;
cvt.u64.u32 %rd47, %r76;
mul.lo.s32 %r77, %r12, %r58;
cvt.u64.u32 %rd48, %r77;
cvt.u64.u32 %rd49, %r2;
cvt.s64.s32 %rd50, %r4;
add.s64 %rd51, %rd50, %rd49;
add.s64 %rd52, %rd51, %rd47;
add.s64 %rd53, %rd52, %rd48;
shl.b64 %rd54, %rd53, 2;
add.s64 %rd55, %rd46, %rd54;
ld.global.f32 %f4, [%rd55];

$L__BB2_8:
shr.s32 %r78, %r3, 31;
shr.u32 %r79, %r78, 30;
add.s32 %r80, %r3, %r79;
shr.s32 %r1416, %r80, 2;
shr.u32 %r81, %r47, 7;
setp.ge.s32 %p12, %r1416, %r81;
mov.f32 %f27, %f26;
mov.f32 %f28, %f26;
mov.f32 %f29, %f26;
mov.f32 %f30, %f26;
mov.f32 %f31, %f26;
mov.f32 %f32, %f26;
mov.f32 %f33, %f26;
@%p12 bra $L__BB2_19;

mul.hi.u32 %r82, %r11, %r59;
add.s32 %r83, %r82, %r11;
shr.u32 %r84, %r83, %r60;
mul.lo.s32 %r85, %r84, %r62;
mul.lo.s32 %r86, %r1413, %r56;
mul.lo.s32 %r87, %r1414, %r57;
cvt.u64.u32 %rd56, %r87;
and.b32 %r91, %r80, -4;
sub.s32 %r92, %r3, %r91;
shl.b32 %r93, %r92, 1;
cvt.s64.s32 %rd57, %r92;
mul.lo.s32 %r94, %r11, %r63;
cvt.u64.u32 %rd58, %r94;
add.s64 %rd59, %rd58, %rd57;
add.s64 %rd4, %rd59, %rd56;
shl.b32 %r1415, %r1416, 2;
mad.lo.s32 %r95, %r4, %r51, %r1416;
add.s32 %r96, %r95, %r85;
add.s32 %r97, %r96, %r86;
mul.wide.s32 %rd5, %r97, 18;
mul.wide.s32 %rd60, %r93, 2;
add.s64 %rd61, %rd5, %rd60;
add.s64 %rd6, %rd61, 4;
add.s32 %r98, %r1416, %r85;
add.s32 %r17, %r98, %r86;
add.s32 %r99, %r4, 1;
mad.lo.s32 %r100, %r51, %r99, %r17;
mul.wide.s32 %rd62, %r100, 18;
add.s64 %rd63, %rd62, %rd60;
add.s64 %rd7, %rd63, 2;
add.s32 %r101, %r4, 2;
mad.lo.s32 %r102, %r51, %r101, %r17;
mul.wide.s32 %rd64, %r102, 18;
add.s64 %rd65, %rd64, %rd60;
add.s64 %rd8, %rd65, 2;
add.s32 %r103, %r4, 3;
mad.lo.s32 %r104, %r51, %r103, %r17;
mul.wide.s32 %rd66, %r104, 18;
add.s64 %rd67, %rd66, %rd60;
add.s64 %rd9, %rd67, 2;
mul.wide.s32 %rd10, %r51, 18;
cvta.to.global.u64 %rd92, %rd28;
cvta.to.global.u64 %rd93, %rd24;
mov.f32 %f26, 0f00000000;
cvta.to.global.u64 %rd68, %rd25;
mov.f32 %f27, %f26;
mov.f32 %f28, %f26;
mov.f32 %f29, %f26;
mov.f32 %f30, %f26;
mov.f32 %f31, %f26;
mov.f32 %f32, %f26;
mov.f32 %f33, %f26;

$L__BB2_10:
add.s64 %rd15, %rd93, %rd5;
ld.global.nc.u16 %rs1, [%rd15];

{ cvt.f32.f16 %f101, %rs1;}


cvt.u64.u32 %rd69, %r1415;
add.s64 %rd70, %rd4, %rd69;
mul.lo.s64 %rd71, %rd70, 36;
add.s64 %rd72, %rd93, %rd6;
ld.global.nc.u16 %rs2, [%rd72+-2];
cvt.s32.s16 %r138, %rs2;
add.s64 %rd73, %rd68, %rd71;
ld.global.nc.u32 %r20, [%rd73+4];
ld.global.nc.u32 %r21, [%rd73+8];
ld.global.nc.u32 %r22, [%rd73+12];
ld.global.nc.u32 %r23, [%rd73+16];
and.b32 %r139, %r138, 30583;
mov.u32 %r140, 286261504;
prmt.b32 %r141, %r140, %r140, %r139;
shr.u32 %r142, %r138, 2;
and.b32 %r143, %r142, 30583;
prmt.b32 %r144, %r140, %r140, %r143;
and.b32 %r145, %r141, 30583;
mov.u32 %r146, 511;
prmt.b32 %r147, %r146, %r146, %r145;
and.b32 %r148, %r144, 30583;
prmt.b32 %r149, %r146, %r146, %r148;
shr.u32 %r150, %r141, 16;
and.b32 %r151, %r150, 30583;
prmt.b32 %r152, %r146, %r146, %r151;
shr.u32 %r153, %r144, 16;
and.b32 %r154, %r153, 30583;
prmt.b32 %r155, %r146, %r146, %r154;
mov.u32 %r156, 21520;
prmt.b32 %r106, %r147, %r149, %r156;
mov.u32 %r157, 30258;
prmt.b32 %r110, %r147, %r149, %r157;
prmt.b32 %r114, %r152, %r155, %r156;
prmt.b32 %r118, %r152, %r155, %r157;
mov.u32 %r108, 0;

dp4a.s32.s32 %r105, %r106, %r20, %r108;


dp4a.s32.s32 %r109, %r110, %r21, %r105;


dp4a.s32.s32 %r113, %r114, %r22, %r109;


dp4a.s32.s32 %r117, %r118, %r23, %r113;

ld.global.nc.u16 %rs3, [%rd72];
cvt.s32.s16 %r158, %rs3;
ld.global.nc.u32 %r24, [%rd73+20];
ld.global.nc.u32 %r25, [%rd73+24];
ld.global.nc.u32 %r26, [%rd73+28];
ld.global.nc.u32 %r27, [%rd73+32];
and.b32 %r159, %r158, 30583;
prmt.b32 %r160, %r140, %r140, %r159;
shr.u32 %r161, %r158, 2;
and.b32 %r162, %r161, 30583;
prmt.b32 %r163, %r140, %r140, %r162;
and.b32 %r164, %r160, 30583;
prmt.b32 %r165, %r146, %r146, %r164;
and.b32 %r166, %r163, 30583;
prmt.b32 %r167, %r146, %r146, %r166;
shr.u32 %r168, %r160, 16;
and.b32 %r169, %r168, 30583;
prmt.b32 %r170, %r146, %r146, %r169;
shr.u32 %r171, %r163, 16;
and.b32 %r172, %r171, 30583;
prmt.b32 %r173, %r146, %r146, %r172;
prmt.b32 %r122, %r165, %r167, %r156;
prmt.b32 %r126, %r165, %r167, %r157;
prmt.b32 %r130, %r170, %r173, %r156;
prmt.b32 %r134, %r170, %r173, %r157;

dp4a.s32.s32 %r121, %r122, %r24, %r117;


dp4a.s32.s32 %r125, %r126, %r25, %r121;


dp4a.s32.s32 %r129, %r130, %r26, %r125;


dp4a.s32.s32 %r133, %r134, %r27, %r129;

ld.global.nc.u32 %r28, [%rd73];

{.reg .f16 low,high;
mov.b32 {low,high},%r28;
cvt.f32.f16 %f102, low;}


mul.ftz.f32 %f103, %f101, %f102;
cvt.rn.f32.s32 %f104, %r133;
fma.rn.ftz.f32 %f29, %f103, %f104, %f29;
setp.eq.s64 %p13, %rd28, 0;
@%p13 bra $L__BB2_12;

add.s64 %rd74, %rd92, %rd5;
ld.global.u16 %rs4, [%rd74];

{ cvt.f32.f16 %f105, %rs4;}


add.s64 %rd75, %rd92, %rd6;
ld.global.s16 %r207, [%rd75+-2];
and.b32 %r208, %r207, 30583;
prmt.b32 %r210, %r140, %r140, %r208;
shr.u32 %r211, %r207, 2;
and.b32 %r212, %r211, 30583;
prmt.b32 %r213, %r140, %r140, %r212;
and.b32 %r214, %r210, 30583;
prmt.b32 %r216, %r146, %r146, %r214;
and.b32 %r217, %r213, 30583;
prmt.b32 %r218, %r146, %r146, %r217;
shr.u32 %r219, %r210, 16;
and.b32 %r220, %r219, 30583;
prmt.b32 %r221, %r146, %r146, %r220;
shr.u32 %r222, %r213, 16;
and.b32 %r223, %r222, 30583;
prmt.b32 %r224, %r146, %r146, %r223;
prmt.b32 %r175, %r216, %r218, %r156;
prmt.b32 %r179, %r216, %r218, %r157;
prmt.b32 %r183, %r221, %r224, %r156;
prmt.b32 %r187, %r221, %r224, %r157;

dp4a.s32.s32 %r174, %r175, %r20, %r108;


dp4a.s32.s32 %r178, %r179, %r21, %r174;


dp4a.s32.s32 %r182, %r183, %r22, %r178;


dp4a.s32.s32 %r186, %r187, %r23, %r182;

ld.global.s16 %r227, [%rd75];
and.b32 %r228, %r227, 30583;
prmt.b32 %r229, %r140, %r140, %r228;
shr.u32 %r230, %r227, 2;
and.b32 %r231, %r230, 30583;
prmt.b32 %r232, %r140, %r140, %r231;
and.b32 %r233, %r229, 30583;
prmt.b32 %r234, %r146, %r146, %r233;
and.b32 %r235, %r232, 30583;
prmt.b32 %r236, %r146, %r146, %r235;
shr.u32 %r237, %r229, 16;
and.b32 %r238, %r237, 30583;
prmt.b32 %r239, %r146, %r146, %r238;
shr.u32 %r240, %r232, 16;
and.b32 %r241, %r240, 30583;
prmt.b32 %r242, %r146, %r146, %r241;
prmt.b32 %r191, %r234, %r236, %r156;
prmt.b32 %r195, %r234, %r236, %r157;
prmt.b32 %r199, %r239, %r242, %r156;
prmt.b32 %r203, %r239, %r242, %r157;

dp4a.s32.s32 %r190, %r191, %r24, %r186;


dp4a.s32.s32 %r194, %r195, %r25, %r190;


dp4a.s32.s32 %r198, %r199, %r26, %r194;


dp4a.s32.s32 %r202, %r203, %r27, %r198;


{.reg .f16 low,high;
mov.b32 {low,high},%r28;
cvt.f32.f16 %f106, low;}


mul.ftz.f32 %f107, %f105, %f106;
cvt.rn.f32.s32 %f108, %r202;
fma.rn.ftz.f32 %f33, %f107, %f108, %f33;

$L__BB2_12:
add.s64 %rd90, %rd93, %rd6;
add.s64 %rd89, %rd90, -2;
add.s64 %rd17, %rd15, %rd10;
ld.global.nc.u16 %rs5, [%rd17];

{ cvt.f32.f16 %f109, %rs5;}


add.s64 %rd18, %rd89, %rd10;
ld.global.nc.u16 %rs6, [%rd18];
cvt.s32.s16 %r276, %rs6;
and.b32 %r277, %r276, 30583;
prmt.b32 %r279, %r140, %r140, %r277;
shr.u32 %r280, %r276, 2;
and.b32 %r281, %r280, 30583;
prmt.b32 %r282, %r140, %r140, %r281;
and.b32 %r283, %r279, 30583;
prmt.b32 %r285, %r146, %r146, %r283;
and.b32 %r286, %r282, 30583;
prmt.b32 %r287, %r146, %r146, %r286;
shr.u32 %r288, %r279, 16;
and.b32 %r289, %r288, 30583;
prmt.b32 %r290, %r146, %r146, %r289;
shr.u32 %r291, %r282, 16;
and.b32 %r292, %r291, 30583;
prmt.b32 %r293, %r146, %r146, %r292;
prmt.b32 %r244, %r285, %r287, %r156;
prmt.b32 %r248, %r285, %r287, %r157;
prmt.b32 %r252, %r290, %r293, %r156;
prmt.b32 %r256, %r290, %r293, %r157;

dp4a.s32.s32 %r243, %r244, %r20, %r108;


dp4a.s32.s32 %r247, %r248, %r21, %r243;


dp4a.s32.s32 %r251, %r252, %r22, %r247;


dp4a.s32.s32 %r255, %r256, %r23, %r251;

ld.global.nc.u16 %rs7, [%rd18+2];
cvt.s32.s16 %r296, %rs7;
and.b32 %r297, %r296, 30583;
prmt.b32 %r298, %r140, %r140, %r297;
shr.u32 %r299, %r296, 2;
and.b32 %r300, %r299, 30583;
prmt.b32 %r301, %r140, %r140, %r300;
and.b32 %r302, %r298, 30583;
prmt.b32 %r303, %r146, %r146, %r302;
and.b32 %r304, %r301, 30583;
prmt.b32 %r305, %r146, %r146, %r304;
shr.u32 %r306, %r298, 16;
and.b32 %r307, %r306, 30583;
prmt.b32 %r308, %r146, %r146, %r307;
shr.u32 %r309, %r301, 16;
and.b32 %r310, %r309, 30583;
prmt.b32 %r311, %r146, %r146, %r310;
prmt.b32 %r260, %r303, %r305, %r156;
prmt.b32 %r264, %r303, %r305, %r157;
prmt.b32 %r268, %r308, %r311, %r156;
prmt.b32 %r272, %r308, %r311, %r157;

dp4a.s32.s32 %r259, %r260, %r24, %r255;


dp4a.s32.s32 %r263, %r264, %r25, %r259;


dp4a.s32.s32 %r267, %r268, %r26, %r263;


dp4a.s32.s32 %r271, %r272, %r27, %r267;


{.reg .f16 low,high;
mov.b32 {low,high},%r28;
cvt.f32.f16 %f110, low;}


mul.ftz.f32 %f111, %f109, %f110;
cvt.rn.f32.s32 %f112, %r271;
fma.rn.ftz.f32 %f28, %f111, %f112, %f28;
@%p13 bra $L__BB2_14;

or.b32 %r347, %r4, 1;
mad.lo.s32 %r348, %r51, %r347, %r17;
mul.wide.s32 %rd76, %r348, 18;
add.s64 %rd77, %rd92, %rd76;
ld.global.u16 %rs8, [%rd77];

{ cvt.f32.f16 %f113, %rs8;}


add.s64 %rd78, %rd92, %rd7;
ld.global.s16 %r349, [%rd78];
and.b32 %r350, %r349, 30583;
mov.u32 %r351, 286261504;
prmt.b32 %r352, %r351, %r351, %r350;
shr.u32 %r353, %r349, 2;
and.b32 %r354, %r353, 30583;
prmt.b32 %r355, %r351, %r351, %r354;
and.b32 %r356, %r352, 30583;
mov.u32 %r357, 511;
prmt.b32 %r358, %r357, %r357, %r356;
and.b32 %r359, %r355, 30583;
prmt.b32 %r360, %r357, %r357, %r359;
shr.u32 %r361, %r352, 16;
and.b32 %r362, %r361, 30583;
prmt.b32 %r363, %r357, %r357, %r362;
shr.u32 %r364, %r355, 16;
and.b32 %r365, %r364, 30583;
prmt.b32 %r366, %r357, %r357, %r365;
mov.u32 %r367, 21520;
prmt.b32 %r313, %r358, %r360, %r367;
mov.u32 %r368, 30258;
prmt.b32 %r317, %r358, %r360, %r368;
prmt.b32 %r321, %r363, %r366, %r367;
prmt.b32 %r325, %r363, %r366, %r368;
mov.u32 %r315, 0;

dp4a.s32.s32 %r312, %r313, %r20, %r315;


dp4a.s32.s32 %r316, %r317, %r21, %r312;


dp4a.s32.s32 %r320, %r321, %r22, %r316;


dp4a.s32.s32 %r324, %r325, %r23, %r320;

ld.global.s16 %r369, [%rd78+2];
and.b32 %r370, %r369, 30583;
prmt.b32 %r371, %r351, %r351, %r370;
shr.u32 %r372, %r369, 2;
and.b32 %r373, %r372, 30583;
prmt.b32 %r374, %r351, %r351, %r373;
and.b32 %r375, %r371, 30583;
prmt.b32 %r376, %r357, %r357, %r375;
and.b32 %r377, %r374, 30583;
prmt.b32 %r378, %r357, %r357, %r377;
shr.u32 %r379, %r371, 16;
and.b32 %r380, %r379, 30583;
prmt.b32 %r381, %r357, %r357, %r380;
shr.u32 %r382, %r374, 16;
and.b32 %r383, %r382, 30583;
prmt.b32 %r384, %r357, %r357, %r383;
prmt.b32 %r329, %r376, %r378, %r367;
prmt.b32 %r333, %r376, %r378, %r368;
prmt.b32 %r337, %r381, %r384, %r367;
prmt.b32 %r341, %r381, %r384, %r368;

dp4a.s32.s32 %r328, %r329, %r24, %r324;


dp4a.s32.s32 %r332, %r333, %r25, %r328;


dp4a.s32.s32 %r336, %r337, %r26, %r332;


dp4a.s32.s32 %r340, %r341, %r27, %r336;


{.reg .f16 low,high;
mov.b32 {low,high},%r28;
cvt.f32.f16 %f114, low;}


mul.ftz.f32 %f115, %f113, %f114;
cvt.rn.f32.s32 %f116, %r340;
fma.rn.ftz.f32 %f32, %f115, %f116, %f32;

$L__BB2_14:
add.s64 %rd19, %rd17, %rd10;
ld.global.nc.u16 %rs9, [%rd19];

{ cvt.f32.f16 %f117, %rs9;}


add.s64 %rd20, %rd18, %rd10;
ld.global.nc.u16 %rs10, [%rd20];
cvt.s32.s16 %r418, %rs10;
and.b32 %r419, %r418, 30583;
mov.u32 %r420, 286261504;
prmt.b32 %r421, %r420, %r420, %r419;
shr.u32 %r422, %r418, 2;
and.b32 %r423, %r422, 30583;
prmt.b32 %r424, %r420, %r420, %r423;
and.b32 %r425, %r421, 30583;
mov.u32 %r426, 511;
prmt.b32 %r427, %r426, %r426, %r425;
and.b32 %r428, %r424, 30583;
prmt.b32 %r429, %r426, %r426, %r428;
shr.u32 %r430, %r421, 16;
and.b32 %r431, %r430, 30583;
prmt.b32 %r432, %r426, %r426, %r431;
shr.u32 %r433, %r424, 16;
and.b32 %r434, %r433, 30583;
prmt.b32 %r435, %r426, %r426, %r434;
mov.u32 %r436, 21520;
prmt.b32 %r386, %r427, %r429, %r436;
mov.u32 %r437, 30258;
prmt.b32 %r390, %r427, %r429, %r437;
prmt.b32 %r394, %r432, %r435, %r436;
prmt.b32 %r398, %r432, %r435, %r437;
mov.u32 %r388, 0;

dp4a.s32.s32 %r385, %r386, %r20, %r388;


dp4a.s32.s32 %r389, %r390, %r21, %r385;


dp4a.s32.s32 %r393, %r394, %r22, %r389;


dp4a.s32.s32 %r397, %r398, %r23, %r393;

ld.global.nc.u16 %rs11, [%rd20+2];
cvt.s32.s16 %r438, %rs11;
and.b32 %r439, %r438, 30583;
prmt.b32 %r440, %r420, %r420, %r439;
shr.u32 %r441, %r438, 2;
and.b32 %r442, %r441, 30583;
prmt.b32 %r443, %r420, %r420, %r442;
and.b32 %r444, %r440, 30583;
prmt.b32 %r445, %r426, %r426, %r444;
and.b32 %r446, %r443, 30583;
prmt.b32 %r447, %r426, %r426, %r446;
shr.u32 %r448, %r440, 16;
and.b32 %r449, %r448, 30583;
prmt.b32 %r450, %r426, %r426, %r449;
shr.u32 %r451, %r443, 16;
and.b32 %r452, %r451, 30583;
prmt.b32 %r453, %r426, %r426, %r452;
prmt.b32 %r402, %r445, %r447, %r436;
prmt.b32 %r406, %r445, %r447, %r437;
prmt.b32 %r410, %r450, %r453, %r436;
prmt.b32 %r414, %r450, %r453, %r437;

dp4a.s32.s32 %r401, %r402, %r24, %r397;


dp4a.s32.s32 %r405, %r406, %r25, %r401;


dp4a.s32.s32 %r409, %r410, %r26, %r405;


dp4a.s32.s32 %r413, %r414, %r27, %r409;


{.reg .f16 low,high;
mov.b32 {low,high},%r28;
cvt.f32.f16 %f118, low;}


mul.ftz.f32 %f119, %f117, %f118;
cvt.rn.f32.s32 %f120, %r413;
fma.rn.ftz.f32 %f27, %f119, %f120, %f27;
@%p13 bra $L__BB2_16;

or.b32 %r489, %r4, 2;
mad.lo.s32 %r490, %r51, %r489, %r17;
mul.wide.s32 %rd79, %r490, 18;
add.s64 %rd80, %rd92, %rd79;
ld.global.u16 %rs12, [%rd80];

{ cvt.f32.f16 %f121, %rs12;}


add.s64 %rd81, %rd92, %rd8;
ld.global.s16 %r491, [%rd81];
and.b32 %r492, %r491, 30583;
prmt.b32 %r494, %r420, %r420, %r492;
shr.u32 %r495, %r491, 2;
and.b32 %r496, %r495, 30583;
prmt.b32 %r497, %r420, %r420, %r496;
and.b32 %r498, %r494, 30583;
prmt.b32 %r500, %r426, %r426, %r498;
and.b32 %r501, %r497, 30583;
prmt.b32 %r502, %r426, %r426, %r501;
shr.u32 %r503, %r494, 16;
and.b32 %r504, %r503, 30583;
prmt.b32 %r505, %r426, %r426, %r504;
shr.u32 %r506, %r497, 16;
and.b32 %r507, %r506, 30583;
prmt.b32 %r508, %r426, %r426, %r507;
prmt.b32 %r455, %r500, %r502, %r436;
prmt.b32 %r459, %r500, %r502, %r437;
prmt.b32 %r463, %r505, %r508, %r436;
prmt.b32 %r467, %r505, %r508, %r437;

dp4a.s32.s32 %r454, %r455, %r20, %r388;


dp4a.s32.s32 %r458, %r459, %r21, %r454;


dp4a.s32.s32 %r462, %r463, %r22, %r458;


dp4a.s32.s32 %r466, %r467, %r23, %r462;

ld.global.s16 %r511, [%rd81+2];
and.b32 %r512, %r511, 30583;
prmt.b32 %r513, %r420, %r420, %r512;
shr.u32 %r514, %r511, 2;
and.b32 %r515, %r514, 30583;
prmt.b32 %r516, %r420, %r420, %r515;
and.b32 %r517, %r513, 30583;
prmt.b32 %r518, %r426, %r426, %r517;
and.b32 %r519, %r516, 30583;
prmt.b32 %r520, %r426, %r426, %r519;
shr.u32 %r521, %r513, 16;
and.b32 %r522, %r521, 30583;
prmt.b32 %r523, %r426, %r426, %r522;
shr.u32 %r524, %r516, 16;
and.b32 %r525, %r524, 30583;
prmt.b32 %r526, %r426, %r426, %r525;
prmt.b32 %r471, %r518, %r520, %r436;
prmt.b32 %r475, %r518, %r520, %r437;
prmt.b32 %r479, %r523, %r526, %r436;
prmt.b32 %r483, %r523, %r526, %r437;

dp4a.s32.s32 %r470, %r471, %r24, %r466;


dp4a.s32.s32 %r474, %r475, %r25, %r470;


dp4a.s32.s32 %r478, %r479, %r26, %r474;


dp4a.s32.s32 %r482, %r483, %r27, %r478;


{.reg .f16 low,high;
mov.b32 {low,high},%r28;
cvt.f32.f16 %f122, low;}


mul.ftz.f32 %f123, %f121, %f122;
cvt.rn.f32.s32 %f124, %r482;
fma.rn.ftz.f32 %f31, %f123, %f124, %f31;

$L__BB2_16:
add.s64 %rd82, %rd19, %rd10;
ld.global.nc.u16 %rs13, [%rd82];

{ cvt.f32.f16 %f125, %rs13;}


add.s64 %rd83, %rd20, %rd10;
ld.global.nc.u16 %rs14, [%rd83];
cvt.s32.s16 %r560, %rs14;
and.b32 %r561, %r560, 30583;
prmt.b32 %r563, %r420, %r420, %r561;
shr.u32 %r564, %r560, 2;
and.b32 %r565, %r564, 30583;
prmt.b32 %r566, %r420, %r420, %r565;
and.b32 %r567, %r563, 30583;
prmt.b32 %r569, %r426, %r426, %r567;
and.b32 %r570, %r566, 30583;
prmt.b32 %r571, %r426, %r426, %r570;
shr.u32 %r572, %r563, 16;
and.b32 %r573, %r572, 30583;
prmt.b32 %r574, %r426, %r426, %r573;
shr.u32 %r575, %r566, 16;
and.b32 %r576, %r575, 30583;
prmt.b32 %r577, %r426, %r426, %r576;
prmt.b32 %r528, %r569, %r571, %r436;
prmt.b32 %r532, %r569, %r571, %r437;
prmt.b32 %r536, %r574, %r577, %r436;
prmt.b32 %r540, %r574, %r577, %r437;

dp4a.s32.s32 %r527, %r528, %r20, %r388;


dp4a.s32.s32 %r531, %r532, %r21, %r527;


dp4a.s32.s32 %r535, %r536, %r22, %r531;


dp4a.s32.s32 %r539, %r540, %r23, %r535;

ld.global.nc.u16 %rs15, [%rd83+2];
cvt.s32.s16 %r580, %rs15;
and.b32 %r581, %r580, 30583;
prmt.b32 %r582, %r420, %r420, %r581;
shr.u32 %r583, %r580, 2;
and.b32 %r584, %r583, 30583;
prmt.b32 %r585, %r420, %r420, %r584;
and.b32 %r586, %r582, 30583;
prmt.b32 %r587, %r426, %r426, %r586;
and.b32 %r588, %r585, 30583;
prmt.b32 %r589, %r426, %r426, %r588;
shr.u32 %r590, %r582, 16;
and.b32 %r591, %r590, 30583;
prmt.b32 %r592, %r426, %r426, %r591;
shr.u32 %r593, %r585, 16;
and.b32 %r594, %r593, 30583;
prmt.b32 %r595, %r426, %r426, %r594;
prmt.b32 %r544, %r587, %r589, %r436;
prmt.b32 %r548, %r587, %r589, %r437;
prmt.b32 %r552, %r592, %r595, %r436;
prmt.b32 %r556, %r592, %r595, %r437;

dp4a.s32.s32 %r543, %r544, %r24, %r539;


dp4a.s32.s32 %r547, %r548, %r25, %r543;


dp4a.s32.s32 %r551, %r552, %r26, %r547;


dp4a.s32.s32 %r555, %r556, %r27, %r551;


{.reg .f16 low,high;
mov.b32 {low,high},%r28;
cvt.f32.f16 %f126, low;}


mul.ftz.f32 %f127, %f125, %f126;
cvt.rn.f32.s32 %f128, %r555;
fma.rn.ftz.f32 %f26, %f127, %f128, %f26;
@%p13 bra $L__BB2_18;

or.b32 %r631, %r4, 3;
mad.lo.s32 %r632, %r51, %r631, %r17;
mul.wide.s32 %rd84, %r632, 18;
add.s64 %rd85, %rd92, %rd84;
ld.global.u16 %rs16, [%rd85];

{ cvt.f32.f16 %f129, %rs16;}


add.s64 %rd86, %rd92, %rd9;
ld.global.s16 %r633, [%rd86];
and.b32 %r634, %r633, 30583;
mov.u32 %r635, 286261504;
prmt.b32 %r636, %r635, %r635, %r634;
shr.u32 %r637, %r633, 2;
and.b32 %r638, %r637, 30583;
prmt.b32 %r639, %r635, %r635, %r638;
and.b32 %r640, %r636, 30583;
mov.u32 %r641, 511;
prmt.b32 %r642, %r641, %r641, %r640;
and.b32 %r643, %r639, 30583;
prmt.b32 %r644, %r641, %r641, %r643;
shr.u32 %r645, %r636, 16;
and.b32 %r646, %r645, 30583;
prmt.b32 %r647, %r641, %r641, %r646;
shr.u32 %r648, %r639, 16;
and.b32 %r649, %r648, 30583;
prmt.b32 %r650, %r641, %r641, %r649;
mov.u32 %r651, 21520;
prmt.b32 %r597, %r642, %r644, %r651;
mov.u32 %r652, 30258;
prmt.b32 %r601, %r642, %r644, %r652;
prmt.b32 %r605, %r647, %r650, %r651;
prmt.b32 %r609, %r647, %r650, %r652;
mov.u32 %r599, 0;

dp4a.s32.s32 %r596, %r597, %r20, %r599;


dp4a.s32.s32 %r600, %r601, %r21, %r596;


dp4a.s32.s32 %r604, %r605, %r22, %r600;


dp4a.s32.s32 %r608, %r609, %r23, %r604;

ld.global.s16 %r653, [%rd86+2];
and.b32 %r654, %r653, 30583;
prmt.b32 %r655, %r635, %r635, %r654;
shr.u32 %r656, %r653, 2;
and.b32 %r657, %r656, 30583;
prmt.b32 %r658, %r635, %r635, %r657;
and.b32 %r659, %r655, 30583;
prmt.b32 %r660, %r641, %r641, %r659;
and.b32 %r661, %r658, 30583;
prmt.b32 %r662, %r641, %r641, %r661;
shr.u32 %r663, %r655, 16;
and.b32 %r664, %r663, 30583;
prmt.b32 %r665, %r641, %r641, %r664;
shr.u32 %r666, %r658, 16;
and.b32 %r667, %r666, 30583;
prmt.b32 %r668, %r641, %r641, %r667;
prmt.b32 %r613, %r660, %r662, %r651;
prmt.b32 %r617, %r660, %r662, %r652;
prmt.b32 %r621, %r665, %r668, %r651;
prmt.b32 %r625, %r665, %r668, %r652;

dp4a.s32.s32 %r612, %r613, %r24, %r608;


dp4a.s32.s32 %r616, %r617, %r25, %r612;


dp4a.s32.s32 %r620, %r621, %r26, %r616;


dp4a.s32.s32 %r624, %r625, %r27, %r620;


{.reg .f16 low,high;
mov.b32 {low,high},%r28;
cvt.f32.f16 %f130, low;}


mul.ftz.f32 %f131, %f129, %f130;
cvt.rn.f32.s32 %f132, %r624;
fma.rn.ftz.f32 %f30, %f131, %f132, %f30;

$L__BB2_18:
add.s32 %r1415, %r1415, 128;
add.s64 %rd93, %rd93, 576;
add.s64 %rd92, %rd92, 576;
add.s32 %r1416, %r1416, 32;
setp.lt.s32 %p17, %r1416, %r81;
@%p17 bra $L__BB2_10;

$L__BB2_19:
mov.u32 %r1402, %tid.y;
setp.eq.s32 %p18, %r1402, 0;
@%p18 bra $L__BB2_23;

mov.u32 %r1403, %tid.y;
setp.ne.s64 %p19, %rd28, 0;
add.s32 %r31, %r1403, -1;
shl.b32 %r673, %r1403, 9;
mov.u32 %r674, _ZZ13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r675, %r674, %r673;
shl.b32 %r676, %r2, 2;
add.s32 %r677, %r675, %r676;
st.shared.f32 [%r677+-512], %f29;
@%p19 bra $L__BB2_22;
bra.uni $L__BB2_21;

$L__BB2_22:
shl.b32 %r679, %r31, 9;
mov.u32 %r680, _ZZ13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r681, %r680, %r679;
add.s32 %r683, %r681, %r676;
st.shared.f32 [%r683], %f33;
add.s32 %r1399, %r677, -512;
st.shared.f32 [%r1399+128], %f28;
st.shared.f32 [%r683+128], %f32;
add.s32 %r1400, %r677, -512;
st.shared.f32 [%r1400+256], %f27;
st.shared.f32 [%r683+256], %f31;
add.s32 %r1401, %r677, -512;
st.shared.f32 [%r1401+384], %f26;
st.shared.f32 [%r683+384], %f30;
bra.uni $L__BB2_23;

$L__BB2_21:
add.s32 %r1396, %r677, -512;
st.shared.f32 [%r1396+128], %f28;
add.s32 %r1397, %r677, -512;
st.shared.f32 [%r1397+256], %f27;
add.s32 %r1398, %r677, -512;
st.shared.f32 [%r1398+384], %f26;

$L__BB2_23:
mov.u32 %r1404, %tid.y;
setp.ne.s32 %p319, %r1404, 0;
bar.sync 0;
@%p319 bra $L__BB2_78;
bra.uni $L__BB2_24;

$L__BB2_78:
ret;

$L__BB2_24:
ld.param.f32 %f977, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
ld.param.u64 %rd91, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r1407, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r1406, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
mov.u32 %r1405, %ctaid.z;
setp.ne.s64 %p21, %rd28, 0;
mov.u32 %r686, %ctaid.y;
mad.lo.s32 %r689, %r686, %r1407, %r4;
mad.lo.s32 %r690, %r1405, %r1406, %r689;
neg.ftz.f32 %f35, %f977;
shl.b32 %r692, %r2, 2;
mov.u32 %r693, _ZZ13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r33, %r693, %r692;
cvta.to.global.u64 %rd87, %rd91;
mul.wide.u32 %rd88, %r690, 4;
add.s64 %rd23, %rd87, %rd88;
@%p21 bra $L__BB2_33;
bra.uni $L__BB2_25;

$L__BB2_33:
ld.param.u32 %r1408, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
mov.u32 %r780, _ZZ13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r35, %r780, %r692;
setp.gt.s32 %p54, %r1408, 2;
@%p54 bra $L__BB2_44;

ld.param.u32 %r1411, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p57, %r1411, 1;
@%p57 bra $L__BB2_62;

ld.param.u32 %r1412, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p58, %r1412, 2;
@%p58 bra $L__BB2_36;
bra.uni $L__BB2_70;

$L__BB2_36:
ld.shared.f32 %f533, [%r33];
add.ftz.f32 %f534, %f533, %f29;
ld.shared.f32 %f535, [%r35];
add.ftz.f32 %f536, %f535, %f33;
ld.shared.f32 %f537, [%r33+512];
add.ftz.f32 %f538, %f537, %f534;
ld.shared.f32 %f539, [%r35+512];
add.ftz.f32 %f540, %f539, %f536;
ld.shared.f32 %f541, [%r33+1024];
add.ftz.f32 %f542, %f541, %f538;
ld.shared.f32 %f543, [%r35+1024];
add.ftz.f32 %f544, %f543, %f540;
mov.b32 %r1027, %f542;
mov.u32 %r1028, 31;
mov.u32 %r1029, 16;
mov.u32 %r1030, -1;
shfl.sync.bfly.b32 %r1031|%p163, %r1027, %r1029, %r1028, %r1030;
mov.b32 %f545, %r1031;
add.ftz.f32 %f546, %f542, %f545;
mov.b32 %r1032, %f546;
mov.u32 %r1033, 8;
shfl.sync.bfly.b32 %r1034|%p164, %r1032, %r1033, %r1028, %r1030;
mov.b32 %f547, %r1034;
add.ftz.f32 %f548, %f546, %f547;
mov.b32 %r1035, %f548;
mov.u32 %r1036, 4;
shfl.sync.bfly.b32 %r1037|%p165, %r1035, %r1036, %r1028, %r1030;
mov.b32 %f549, %r1037;
add.ftz.f32 %f550, %f548, %f549;
mov.b32 %r1038, %f550;
mov.u32 %r1039, 2;
shfl.sync.bfly.b32 %r1040|%p166, %r1038, %r1039, %r1028, %r1030;
mov.b32 %f551, %r1040;
add.ftz.f32 %f552, %f550, %f551;
mov.b32 %r1041, %f552;
mov.u32 %r1042, 1;
shfl.sync.bfly.b32 %r1043|%p167, %r1041, %r1042, %r1028, %r1030;
mov.b32 %f553, %r1043;
add.ftz.f32 %f56, %f552, %f553;
mov.b32 %r1044, %f544;
shfl.sync.bfly.b32 %r1045|%p168, %r1044, %r1029, %r1028, %r1030;
mov.b32 %f554, %r1045;
add.ftz.f32 %f555, %f544, %f554;
mov.b32 %r1046, %f555;
shfl.sync.bfly.b32 %r1047|%p169, %r1046, %r1033, %r1028, %r1030;
mov.b32 %f556, %r1047;
add.ftz.f32 %f557, %f555, %f556;
mov.b32 %r1048, %f557;
shfl.sync.bfly.b32 %r1049|%p170, %r1048, %r1036, %r1028, %r1030;
mov.b32 %f558, %r1049;
add.ftz.f32 %f559, %f557, %f558;
mov.b32 %r1050, %f559;
shfl.sync.bfly.b32 %r1051|%p171, %r1050, %r1039, %r1028, %r1030;
mov.b32 %f560, %r1051;
add.ftz.f32 %f561, %f559, %f560;
mov.b32 %r1052, %f561;
shfl.sync.bfly.b32 %r1053|%p172, %r1052, %r1042, %r1028, %r1030;
mov.b32 %f562, %r1053;
add.ftz.f32 %f57, %f561, %f562;
setp.ne.s32 %p173, %r2, 0;
setp.ge.u32 %p174, %r4, %r52;
or.pred %p175, %p173, %p174;
@%p175 bra $L__BB2_38;

add.ftz.f32 %f563, %f5, %f56;
add.ftz.f32 %f564, %f4, %f57;
mul.ftz.f32 %f565, %f564, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f566, %f565;
add.ftz.f32 %f567, %f566, 0f3F800000;
div.approx.ftz.f32 %f568, %f564, %f567;
mul.ftz.f32 %f569, %f563, %f568;
st.global.f32 [%rd23], %f569;

$L__BB2_38:
ld.shared.f32 %f570, [%r33+128];
add.ftz.f32 %f571, %f570, %f28;
ld.shared.f32 %f572, [%r35+128];
add.ftz.f32 %f573, %f572, %f32;
ld.shared.f32 %f574, [%r33+640];
add.ftz.f32 %f575, %f574, %f571;
ld.shared.f32 %f576, [%r35+640];
add.ftz.f32 %f577, %f576, %f573;
ld.shared.f32 %f578, [%r33+1152];
add.ftz.f32 %f579, %f578, %f575;
ld.shared.f32 %f580, [%r35+1152];
add.ftz.f32 %f581, %f580, %f577;
mov.b32 %r1057, %f579;
shfl.sync.bfly.b32 %r1061|%p176, %r1057, %r1029, %r1028, %r1030;
mov.b32 %f582, %r1061;
add.ftz.f32 %f583, %f579, %f582;
mov.b32 %r1062, %f583;
shfl.sync.bfly.b32 %r1064|%p177, %r1062, %r1033, %r1028, %r1030;
mov.b32 %f584, %r1064;
add.ftz.f32 %f585, %f583, %f584;
mov.b32 %r1065, %f585;
shfl.sync.bfly.b32 %r1067|%p178, %r1065, %r1036, %r1028, %r1030;
mov.b32 %f586, %r1067;
add.ftz.f32 %f587, %f585, %f586;
mov.b32 %r1068, %f587;
shfl.sync.bfly.b32 %r1070|%p179, %r1068, %r1039, %r1028, %r1030;
mov.b32 %f588, %r1070;
add.ftz.f32 %f589, %f587, %f588;
mov.b32 %r1071, %f589;
shfl.sync.bfly.b32 %r1073|%p180, %r1071, %r1042, %r1028, %r1030;
mov.b32 %f590, %r1073;
add.ftz.f32 %f58, %f589, %f590;
mov.b32 %r1074, %f581;
shfl.sync.bfly.b32 %r1075|%p181, %r1074, %r1029, %r1028, %r1030;
mov.b32 %f591, %r1075;
add.ftz.f32 %f592, %f581, %f591;
mov.b32 %r1076, %f592;
shfl.sync.bfly.b32 %r1077|%p182, %r1076, %r1033, %r1028, %r1030;
mov.b32 %f593, %r1077;
add.ftz.f32 %f594, %f592, %f593;
mov.b32 %r1078, %f594;
shfl.sync.bfly.b32 %r1079|%p183, %r1078, %r1036, %r1028, %r1030;
mov.b32 %f595, %r1079;
add.ftz.f32 %f596, %f594, %f595;
mov.b32 %r1080, %f596;
shfl.sync.bfly.b32 %r1081|%p184, %r1080, %r1039, %r1028, %r1030;
mov.b32 %f597, %r1081;
add.ftz.f32 %f598, %f596, %f597;
mov.b32 %r1082, %f598;
shfl.sync.bfly.b32 %r1083|%p185, %r1082, %r1042, %r1028, %r1030;
mov.b32 %f599, %r1083;
add.ftz.f32 %f59, %f598, %f599;
setp.ne.s32 %p186, %r2, 1;
or.b32 %r1087, %r4, 1;
setp.ge.u32 %p187, %r1087, %r52;
or.pred %p188, %p186, %p187;
@%p188 bra $L__BB2_40;

add.ftz.f32 %f600, %f5, %f58;
add.ftz.f32 %f601, %f4, %f59;
mul.ftz.f32 %f602, %f601, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f603, %f602;
add.ftz.f32 %f604, %f603, 0f3F800000;
div.approx.ftz.f32 %f605, %f601, %f604;
mul.ftz.f32 %f606, %f600, %f605;
st.global.f32 [%rd23+4], %f606;

$L__BB2_40:
ld.shared.f32 %f607, [%r33+256];
add.ftz.f32 %f608, %f607, %f27;
ld.shared.f32 %f609, [%r35+256];
add.ftz.f32 %f610, %f609, %f31;
ld.shared.f32 %f611, [%r33+768];
add.ftz.f32 %f612, %f611, %f608;
ld.shared.f32 %f613, [%r35+768];
add.ftz.f32 %f614, %f613, %f610;
ld.shared.f32 %f615, [%r33+1280];
add.ftz.f32 %f616, %f615, %f612;
ld.shared.f32 %f617, [%r35+1280];
add.ftz.f32 %f618, %f617, %f614;
mov.b32 %r1088, %f616;
mov.u32 %r1089, 31;
mov.u32 %r1090, 16;
mov.u32 %r1091, -1;
shfl.sync.bfly.b32 %r1092|%p189, %r1088, %r1090, %r1089, %r1091;
mov.b32 %f619, %r1092;
add.ftz.f32 %f620, %f616, %f619;
mov.b32 %r1093, %f620;
mov.u32 %r1094, 8;
shfl.sync.bfly.b32 %r1095|%p190, %r1093, %r1094, %r1089, %r1091;
mov.b32 %f621, %r1095;
add.ftz.f32 %f622, %f620, %f621;
mov.b32 %r1096, %f622;
mov.u32 %r1097, 4;
shfl.sync.bfly.b32 %r1098|%p191, %r1096, %r1097, %r1089, %r1091;
mov.b32 %f623, %r1098;
add.ftz.f32 %f624, %f622, %f623;
mov.b32 %r1099, %f624;
mov.u32 %r1100, 2;
shfl.sync.bfly.b32 %r1101|%p192, %r1099, %r1100, %r1089, %r1091;
mov.b32 %f625, %r1101;
add.ftz.f32 %f626, %f624, %f625;
mov.b32 %r1102, %f626;
mov.u32 %r1103, 1;
shfl.sync.bfly.b32 %r1104|%p193, %r1102, %r1103, %r1089, %r1091;
mov.b32 %f627, %r1104;
add.ftz.f32 %f60, %f626, %f627;
mov.b32 %r1105, %f618;
shfl.sync.bfly.b32 %r1106|%p194, %r1105, %r1090, %r1089, %r1091;
mov.b32 %f628, %r1106;
add.ftz.f32 %f629, %f618, %f628;
mov.b32 %r1107, %f629;
shfl.sync.bfly.b32 %r1108|%p195, %r1107, %r1094, %r1089, %r1091;
mov.b32 %f630, %r1108;
add.ftz.f32 %f631, %f629, %f630;
mov.b32 %r1109, %f631;
shfl.sync.bfly.b32 %r1110|%p196, %r1109, %r1097, %r1089, %r1091;
mov.b32 %f632, %r1110;
add.ftz.f32 %f633, %f631, %f632;
mov.b32 %r1111, %f633;
shfl.sync.bfly.b32 %r1112|%p197, %r1111, %r1100, %r1089, %r1091;
mov.b32 %f634, %r1112;
add.ftz.f32 %f635, %f633, %f634;
mov.b32 %r1113, %f635;
shfl.sync.bfly.b32 %r1114|%p198, %r1113, %r1103, %r1089, %r1091;
mov.b32 %f636, %r1114;
add.ftz.f32 %f61, %f635, %f636;
setp.ne.s32 %p199, %r2, 2;
or.b32 %r1118, %r4, 2;
setp.ge.u32 %p200, %r1118, %r52;
or.pred %p201, %p199, %p200;
@%p201 bra $L__BB2_42;

add.ftz.f32 %f637, %f5, %f60;
add.ftz.f32 %f638, %f4, %f61;
mul.ftz.f32 %f639, %f638, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f640, %f639;
add.ftz.f32 %f641, %f640, 0f3F800000;
div.approx.ftz.f32 %f642, %f638, %f641;
mul.ftz.f32 %f643, %f637, %f642;
st.global.f32 [%rd23+8], %f643;

$L__BB2_42:
ld.shared.f32 %f644, [%r33+384];
add.ftz.f32 %f645, %f644, %f26;
ld.shared.f32 %f646, [%r35+384];
add.ftz.f32 %f647, %f646, %f30;
ld.shared.f32 %f648, [%r33+896];
add.ftz.f32 %f649, %f648, %f645;
ld.shared.f32 %f650, [%r35+896];
add.ftz.f32 %f651, %f650, %f647;
ld.shared.f32 %f652, [%r33+1408];
add.ftz.f32 %f653, %f652, %f649;
ld.shared.f32 %f654, [%r35+1408];
add.ftz.f32 %f655, %f654, %f651;
mov.b32 %r1119, %f653;
shfl.sync.bfly.b32 %r1123|%p202, %r1119, %r1090, %r1089, %r1091;
mov.b32 %f656, %r1123;
add.ftz.f32 %f657, %f653, %f656;
mov.b32 %r1124, %f657;
shfl.sync.bfly.b32 %r1126|%p203, %r1124, %r1094, %r1089, %r1091;
mov.b32 %f658, %r1126;
add.ftz.f32 %f659, %f657, %f658;
mov.b32 %r1127, %f659;
shfl.sync.bfly.b32 %r1129|%p204, %r1127, %r1097, %r1089, %r1091;
mov.b32 %f660, %r1129;
add.ftz.f32 %f661, %f659, %f660;
mov.b32 %r1130, %f661;
shfl.sync.bfly.b32 %r1132|%p205, %r1130, %r1100, %r1089, %r1091;
mov.b32 %f662, %r1132;
add.ftz.f32 %f663, %f661, %f662;
mov.b32 %r1133, %f663;
shfl.sync.bfly.b32 %r1135|%p206, %r1133, %r1103, %r1089, %r1091;
mov.b32 %f664, %r1135;
add.ftz.f32 %f62, %f663, %f664;
mov.b32 %r1136, %f655;
shfl.sync.bfly.b32 %r1137|%p207, %r1136, %r1090, %r1089, %r1091;
mov.b32 %f665, %r1137;
add.ftz.f32 %f666, %f655, %f665;
mov.b32 %r1138, %f666;
shfl.sync.bfly.b32 %r1139|%p208, %r1138, %r1094, %r1089, %r1091;
mov.b32 %f667, %r1139;
add.ftz.f32 %f668, %f666, %f667;
mov.b32 %r1140, %f668;
shfl.sync.bfly.b32 %r1141|%p209, %r1140, %r1097, %r1089, %r1091;
mov.b32 %f669, %r1141;
add.ftz.f32 %f670, %f668, %f669;
mov.b32 %r1142, %f670;
shfl.sync.bfly.b32 %r1143|%p210, %r1142, %r1100, %r1089, %r1091;
mov.b32 %f671, %r1143;
add.ftz.f32 %f672, %f670, %f671;
mov.b32 %r1144, %f672;
shfl.sync.bfly.b32 %r1145|%p211, %r1144, %r1103, %r1089, %r1091;
mov.b32 %f673, %r1145;
add.ftz.f32 %f63, %f672, %f673;
setp.ne.s32 %p212, %r2, 3;
or.b32 %r1149, %r4, 3;
setp.ge.u32 %p213, %r1149, %r52;
or.pred %p214, %p212, %p213;
@%p214 bra $L__BB2_78;

add.ftz.f32 %f674, %f5, %f62;
add.ftz.f32 %f675, %f4, %f63;
mul.ftz.f32 %f676, %f675, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f677, %f676;
add.ftz.f32 %f678, %f677, 0f3F800000;
div.approx.ftz.f32 %f679, %f675, %f678;
mul.ftz.f32 %f680, %f674, %f679;
st.global.f32 [%rd23+12], %f680;
bra.uni $L__BB2_78;

$L__BB2_25:
ld.shared.f32 %f133, [%r33];
add.ftz.f32 %f134, %f133, %f29;
ld.shared.f32 %f135, [%r33+512];
add.ftz.f32 %f136, %f135, %f134;
ld.shared.f32 %f137, [%r33+1024];
add.ftz.f32 %f138, %f137, %f136;
mov.b32 %r694, %f138;
mov.u32 %r695, 31;
mov.u32 %r696, 16;
mov.u32 %r697, -1;
shfl.sync.bfly.b32 %r698|%p22, %r694, %r696, %r695, %r697;
mov.b32 %f139, %r698;
add.ftz.f32 %f140, %f138, %f139;
mov.b32 %r699, %f140;
mov.u32 %r700, 8;
shfl.sync.bfly.b32 %r701|%p23, %r699, %r700, %r695, %r697;
mov.b32 %f141, %r701;
add.ftz.f32 %f142, %f140, %f141;
mov.b32 %r702, %f142;
mov.u32 %r703, 4;
shfl.sync.bfly.b32 %r704|%p24, %r702, %r703, %r695, %r697;
mov.b32 %f143, %r704;
add.ftz.f32 %f144, %f142, %f143;
mov.b32 %r705, %f144;
mov.u32 %r706, 2;
shfl.sync.bfly.b32 %r707|%p25, %r705, %r706, %r695, %r697;
mov.b32 %f145, %r707;
add.ftz.f32 %f146, %f144, %f145;
mov.b32 %r708, %f146;
mov.u32 %r709, 1;
shfl.sync.bfly.b32 %r710|%p26, %r708, %r709, %r695, %r697;
mov.b32 %f147, %r710;
add.ftz.f32 %f36, %f146, %f147;
setp.ne.s32 %p27, %r2, 0;
setp.ge.u32 %p28, %r4, %r52;
or.pred %p29, %p27, %p28;
@%p29 bra $L__BB2_27;

add.ftz.f32 %f148, %f5, %f36;
st.global.f32 [%rd23], %f148;

$L__BB2_27:
ld.shared.f32 %f149, [%r33+128];
add.ftz.f32 %f150, %f149, %f28;
ld.shared.f32 %f151, [%r33+640];
add.ftz.f32 %f152, %f151, %f150;
ld.shared.f32 %f153, [%r33+1152];
add.ftz.f32 %f154, %f153, %f152;
mov.b32 %r714, %f154;
shfl.sync.bfly.b32 %r718|%p30, %r714, %r696, %r695, %r697;
mov.b32 %f155, %r718;
add.ftz.f32 %f156, %f154, %f155;
mov.b32 %r719, %f156;
shfl.sync.bfly.b32 %r721|%p31, %r719, %r700, %r695, %r697;
mov.b32 %f157, %r721;
add.ftz.f32 %f158, %f156, %f157;
mov.b32 %r722, %f158;
shfl.sync.bfly.b32 %r724|%p32, %r722, %r703, %r695, %r697;
mov.b32 %f159, %r724;
add.ftz.f32 %f160, %f158, %f159;
mov.b32 %r725, %f160;
shfl.sync.bfly.b32 %r727|%p33, %r725, %r706, %r695, %r697;
mov.b32 %f161, %r727;
add.ftz.f32 %f162, %f160, %f161;
mov.b32 %r728, %f162;
shfl.sync.bfly.b32 %r730|%p34, %r728, %r709, %r695, %r697;
mov.b32 %f163, %r730;
add.ftz.f32 %f37, %f162, %f163;
setp.ne.s32 %p35, %r2, 1;
or.b32 %r734, %r4, 1;
setp.ge.u32 %p36, %r734, %r52;
or.pred %p37, %p35, %p36;
@%p37 bra $L__BB2_29;

add.ftz.f32 %f164, %f5, %f37;
st.global.f32 [%rd23+4], %f164;

$L__BB2_29:
ld.shared.f32 %f165, [%r33+256];
add.ftz.f32 %f166, %f165, %f27;
ld.shared.f32 %f167, [%r33+768];
add.ftz.f32 %f168, %f167, %f166;
ld.shared.f32 %f169, [%r33+1280];
add.ftz.f32 %f170, %f169, %f168;
mov.b32 %r735, %f170;
mov.u32 %r736, 31;
mov.u32 %r737, 16;
mov.u32 %r738, -1;
shfl.sync.bfly.b32 %r739|%p38, %r735, %r737, %r736, %r738;
mov.b32 %f171, %r739;
add.ftz.f32 %f172, %f170, %f171;
mov.b32 %r740, %f172;
mov.u32 %r741, 8;
shfl.sync.bfly.b32 %r742|%p39, %r740, %r741, %r736, %r738;
mov.b32 %f173, %r742;
add.ftz.f32 %f174, %f172, %f173;
mov.b32 %r743, %f174;
mov.u32 %r744, 4;
shfl.sync.bfly.b32 %r745|%p40, %r743, %r744, %r736, %r738;
mov.b32 %f175, %r745;
add.ftz.f32 %f176, %f174, %f175;
mov.b32 %r746, %f176;
mov.u32 %r747, 2;
shfl.sync.bfly.b32 %r748|%p41, %r746, %r747, %r736, %r738;
mov.b32 %f177, %r748;
add.ftz.f32 %f178, %f176, %f177;
mov.b32 %r749, %f178;
mov.u32 %r750, 1;
shfl.sync.bfly.b32 %r751|%p42, %r749, %r750, %r736, %r738;
mov.b32 %f179, %r751;
add.ftz.f32 %f38, %f178, %f179;
setp.ne.s32 %p43, %r2, 2;
or.b32 %r755, %r4, 2;
setp.ge.u32 %p44, %r755, %r52;
or.pred %p45, %p43, %p44;
@%p45 bra $L__BB2_31;

add.ftz.f32 %f180, %f5, %f38;
st.global.f32 [%rd23+8], %f180;

$L__BB2_31:
ld.shared.f32 %f181, [%r33+384];
add.ftz.f32 %f182, %f181, %f26;
ld.shared.f32 %f183, [%r33+896];
add.ftz.f32 %f184, %f183, %f182;
ld.shared.f32 %f185, [%r33+1408];
add.ftz.f32 %f186, %f185, %f184;
mov.b32 %r756, %f186;
shfl.sync.bfly.b32 %r760|%p46, %r756, %r737, %r736, %r738;
mov.b32 %f187, %r760;
add.ftz.f32 %f188, %f186, %f187;
mov.b32 %r761, %f188;
shfl.sync.bfly.b32 %r763|%p47, %r761, %r741, %r736, %r738;
mov.b32 %f189, %r763;
add.ftz.f32 %f190, %f188, %f189;
mov.b32 %r764, %f190;
shfl.sync.bfly.b32 %r766|%p48, %r764, %r744, %r736, %r738;
mov.b32 %f191, %r766;
add.ftz.f32 %f192, %f190, %f191;
mov.b32 %r767, %f192;
shfl.sync.bfly.b32 %r769|%p49, %r767, %r747, %r736, %r738;
mov.b32 %f193, %r769;
add.ftz.f32 %f194, %f192, %f193;
mov.b32 %r770, %f194;
shfl.sync.bfly.b32 %r772|%p50, %r770, %r750, %r736, %r738;
mov.b32 %f195, %r772;
add.ftz.f32 %f39, %f194, %f195;
setp.ne.s32 %p51, %r2, 3;
or.b32 %r776, %r4, 3;
setp.ge.u32 %p52, %r776, %r52;
or.pred %p53, %p51, %p52;
@%p53 bra $L__BB2_78;

add.ftz.f32 %f196, %f5, %f39;
st.global.f32 [%rd23+12], %f196;
bra.uni $L__BB2_78;

$L__BB2_44:
ld.param.u32 %r1409, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p55, %r1409, 3;
@%p55 bra $L__BB2_54;

ld.param.u32 %r1410, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.ne.s32 %p56, %r1410, 6;
@%p56 bra $L__BB2_70;

ld.shared.f32 %f197, [%r33];
add.ftz.f32 %f198, %f197, %f29;
ld.shared.f32 %f199, [%r35];
add.ftz.f32 %f200, %f199, %f33;
ld.shared.f32 %f201, [%r33+512];
add.ftz.f32 %f202, %f201, %f198;
ld.shared.f32 %f203, [%r35+512];
add.ftz.f32 %f204, %f203, %f200;
ld.shared.f32 %f205, [%r33+1024];
add.ftz.f32 %f206, %f205, %f202;
ld.shared.f32 %f207, [%r35+1024];
add.ftz.f32 %f208, %f207, %f204;
mov.b32 %r781, %f206;
mov.u32 %r782, 31;
mov.u32 %r783, 16;
mov.u32 %r784, -1;
shfl.sync.bfly.b32 %r785|%p59, %r781, %r783, %r782, %r784;
mov.b32 %f209, %r785;
add.ftz.f32 %f210, %f206, %f209;
mov.b32 %r786, %f210;
mov.u32 %r787, 8;
shfl.sync.bfly.b32 %r788|%p60, %r786, %r787, %r782, %r784;
mov.b32 %f211, %r788;
add.ftz.f32 %f212, %f210, %f211;
mov.b32 %r789, %f212;
mov.u32 %r790, 4;
shfl.sync.bfly.b32 %r791|%p61, %r789, %r790, %r782, %r784;
mov.b32 %f213, %r791;
add.ftz.f32 %f214, %f212, %f213;
mov.b32 %r792, %f214;
mov.u32 %r793, 2;
shfl.sync.bfly.b32 %r794|%p62, %r792, %r793, %r782, %r784;
mov.b32 %f215, %r794;
add.ftz.f32 %f216, %f214, %f215;
mov.b32 %r795, %f216;
mov.u32 %r796, 1;
shfl.sync.bfly.b32 %r797|%p63, %r795, %r796, %r782, %r784;
mov.b32 %f217, %r797;
add.ftz.f32 %f40, %f216, %f217;
mov.b32 %r798, %f208;
shfl.sync.bfly.b32 %r799|%p64, %r798, %r783, %r782, %r784;
mov.b32 %f218, %r799;
add.ftz.f32 %f219, %f208, %f218;
mov.b32 %r800, %f219;
shfl.sync.bfly.b32 %r801|%p65, %r800, %r787, %r782, %r784;
mov.b32 %f220, %r801;
add.ftz.f32 %f221, %f219, %f220;
mov.b32 %r802, %f221;
shfl.sync.bfly.b32 %r803|%p66, %r802, %r790, %r782, %r784;
mov.b32 %f222, %r803;
add.ftz.f32 %f223, %f221, %f222;
mov.b32 %r804, %f223;
shfl.sync.bfly.b32 %r805|%p67, %r804, %r793, %r782, %r784;
mov.b32 %f224, %r805;
add.ftz.f32 %f225, %f223, %f224;
mov.b32 %r806, %f225;
shfl.sync.bfly.b32 %r807|%p68, %r806, %r796, %r782, %r784;
mov.b32 %f226, %r807;
add.ftz.f32 %f41, %f225, %f226;
setp.ne.s32 %p69, %r2, 0;
setp.ge.u32 %p70, %r4, %r52;
or.pred %p71, %p69, %p70;
@%p71 bra $L__BB2_48;

ld.param.f32 %f981, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f227, %f5, %f40;
add.ftz.f32 %f228, %f4, %f41;
min.ftz.f32 %f229, %f228, %f981;
min.ftz.f32 %f230, %f227, %f981;
max.ftz.f32 %f231, %f230, %f35;
mul.ftz.f32 %f232, %f229, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f233, %f232;
add.ftz.f32 %f234, %f233, 0f3F800000;
div.approx.ftz.f32 %f235, %f229, %f234;
mul.ftz.f32 %f236, %f231, %f235;
st.global.f32 [%rd23], %f236;

$L__BB2_48:
ld.shared.f32 %f237, [%r33+128];
add.ftz.f32 %f238, %f237, %f28;
ld.shared.f32 %f239, [%r35+128];
add.ftz.f32 %f240, %f239, %f32;
ld.shared.f32 %f241, [%r33+640];
add.ftz.f32 %f242, %f241, %f238;
ld.shared.f32 %f243, [%r35+640];
add.ftz.f32 %f244, %f243, %f240;
ld.shared.f32 %f245, [%r33+1152];
add.ftz.f32 %f246, %f245, %f242;
ld.shared.f32 %f247, [%r35+1152];
add.ftz.f32 %f248, %f247, %f244;
mov.b32 %r811, %f246;
shfl.sync.bfly.b32 %r815|%p72, %r811, %r783, %r782, %r784;
mov.b32 %f249, %r815;
add.ftz.f32 %f250, %f246, %f249;
mov.b32 %r816, %f250;
shfl.sync.bfly.b32 %r818|%p73, %r816, %r787, %r782, %r784;
mov.b32 %f251, %r818;
add.ftz.f32 %f252, %f250, %f251;
mov.b32 %r819, %f252;
shfl.sync.bfly.b32 %r821|%p74, %r819, %r790, %r782, %r784;
mov.b32 %f253, %r821;
add.ftz.f32 %f254, %f252, %f253;
mov.b32 %r822, %f254;
shfl.sync.bfly.b32 %r824|%p75, %r822, %r793, %r782, %r784;
mov.b32 %f255, %r824;
add.ftz.f32 %f256, %f254, %f255;
mov.b32 %r825, %f256;
shfl.sync.bfly.b32 %r827|%p76, %r825, %r796, %r782, %r784;
mov.b32 %f257, %r827;
add.ftz.f32 %f42, %f256, %f257;
mov.b32 %r828, %f248;
shfl.sync.bfly.b32 %r829|%p77, %r828, %r783, %r782, %r784;
mov.b32 %f258, %r829;
add.ftz.f32 %f259, %f248, %f258;
mov.b32 %r830, %f259;
shfl.sync.bfly.b32 %r831|%p78, %r830, %r787, %r782, %r784;
mov.b32 %f260, %r831;
add.ftz.f32 %f261, %f259, %f260;
mov.b32 %r832, %f261;
shfl.sync.bfly.b32 %r833|%p79, %r832, %r790, %r782, %r784;
mov.b32 %f262, %r833;
add.ftz.f32 %f263, %f261, %f262;
mov.b32 %r834, %f263;
shfl.sync.bfly.b32 %r835|%p80, %r834, %r793, %r782, %r784;
mov.b32 %f264, %r835;
add.ftz.f32 %f265, %f263, %f264;
mov.b32 %r836, %f265;
shfl.sync.bfly.b32 %r837|%p81, %r836, %r796, %r782, %r784;
mov.b32 %f266, %r837;
add.ftz.f32 %f43, %f265, %f266;
setp.ne.s32 %p82, %r2, 1;
or.b32 %r841, %r4, 1;
setp.ge.u32 %p83, %r841, %r52;
or.pred %p84, %p82, %p83;
@%p84 bra $L__BB2_50;

ld.param.f32 %f980, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f267, %f5, %f42;
add.ftz.f32 %f268, %f4, %f43;
min.ftz.f32 %f269, %f268, %f980;
min.ftz.f32 %f270, %f267, %f980;
max.ftz.f32 %f271, %f270, %f35;
mul.ftz.f32 %f272, %f269, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f273, %f272;
add.ftz.f32 %f274, %f273, 0f3F800000;
div.approx.ftz.f32 %f275, %f269, %f274;
mul.ftz.f32 %f276, %f271, %f275;
st.global.f32 [%rd23+4], %f276;

$L__BB2_50:
ld.shared.f32 %f277, [%r33+256];
add.ftz.f32 %f278, %f277, %f27;
ld.shared.f32 %f279, [%r35+256];
add.ftz.f32 %f280, %f279, %f31;
ld.shared.f32 %f281, [%r33+768];
add.ftz.f32 %f282, %f281, %f278;
ld.shared.f32 %f283, [%r35+768];
add.ftz.f32 %f284, %f283, %f280;
ld.shared.f32 %f285, [%r33+1280];
add.ftz.f32 %f286, %f285, %f282;
ld.shared.f32 %f287, [%r35+1280];
add.ftz.f32 %f288, %f287, %f284;
mov.b32 %r842, %f286;
mov.u32 %r843, 31;
mov.u32 %r844, 16;
mov.u32 %r845, -1;
shfl.sync.bfly.b32 %r846|%p85, %r842, %r844, %r843, %r845;
mov.b32 %f289, %r846;
add.ftz.f32 %f290, %f286, %f289;
mov.b32 %r847, %f290;
mov.u32 %r848, 8;
shfl.sync.bfly.b32 %r849|%p86, %r847, %r848, %r843, %r845;
mov.b32 %f291, %r849;
add.ftz.f32 %f292, %f290, %f291;
mov.b32 %r850, %f292;
mov.u32 %r851, 4;
shfl.sync.bfly.b32 %r852|%p87, %r850, %r851, %r843, %r845;
mov.b32 %f293, %r852;
add.ftz.f32 %f294, %f292, %f293;
mov.b32 %r853, %f294;
mov.u32 %r854, 2;
shfl.sync.bfly.b32 %r855|%p88, %r853, %r854, %r843, %r845;
mov.b32 %f295, %r855;
add.ftz.f32 %f296, %f294, %f295;
mov.b32 %r856, %f296;
mov.u32 %r857, 1;
shfl.sync.bfly.b32 %r858|%p89, %r856, %r857, %r843, %r845;
mov.b32 %f297, %r858;
add.ftz.f32 %f44, %f296, %f297;
mov.b32 %r859, %f288;
shfl.sync.bfly.b32 %r860|%p90, %r859, %r844, %r843, %r845;
mov.b32 %f298, %r860;
add.ftz.f32 %f299, %f288, %f298;
mov.b32 %r861, %f299;
shfl.sync.bfly.b32 %r862|%p91, %r861, %r848, %r843, %r845;
mov.b32 %f300, %r862;
add.ftz.f32 %f301, %f299, %f300;
mov.b32 %r863, %f301;
shfl.sync.bfly.b32 %r864|%p92, %r863, %r851, %r843, %r845;
mov.b32 %f302, %r864;
add.ftz.f32 %f303, %f301, %f302;
mov.b32 %r865, %f303;
shfl.sync.bfly.b32 %r866|%p93, %r865, %r854, %r843, %r845;
mov.b32 %f304, %r866;
add.ftz.f32 %f305, %f303, %f304;
mov.b32 %r867, %f305;
shfl.sync.bfly.b32 %r868|%p94, %r867, %r857, %r843, %r845;
mov.b32 %f306, %r868;
add.ftz.f32 %f45, %f305, %f306;
setp.ne.s32 %p95, %r2, 2;
or.b32 %r872, %r4, 2;
setp.ge.u32 %p96, %r872, %r52;
or.pred %p97, %p95, %p96;
@%p97 bra $L__BB2_52;

ld.param.f32 %f979, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f307, %f5, %f44;
add.ftz.f32 %f308, %f4, %f45;
min.ftz.f32 %f309, %f308, %f979;
min.ftz.f32 %f310, %f307, %f979;
max.ftz.f32 %f311, %f310, %f35;
mul.ftz.f32 %f312, %f309, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f313, %f312;
add.ftz.f32 %f314, %f313, 0f3F800000;
div.approx.ftz.f32 %f315, %f309, %f314;
mul.ftz.f32 %f316, %f311, %f315;
st.global.f32 [%rd23+8], %f316;

$L__BB2_52:
ld.shared.f32 %f317, [%r33+384];
add.ftz.f32 %f318, %f317, %f26;
ld.shared.f32 %f319, [%r35+384];
add.ftz.f32 %f320, %f319, %f30;
ld.shared.f32 %f321, [%r33+896];
add.ftz.f32 %f322, %f321, %f318;
ld.shared.f32 %f323, [%r35+896];
add.ftz.f32 %f324, %f323, %f320;
ld.shared.f32 %f325, [%r33+1408];
add.ftz.f32 %f326, %f325, %f322;
ld.shared.f32 %f327, [%r35+1408];
add.ftz.f32 %f328, %f327, %f324;
mov.b32 %r873, %f326;
shfl.sync.bfly.b32 %r877|%p98, %r873, %r844, %r843, %r845;
mov.b32 %f329, %r877;
add.ftz.f32 %f330, %f326, %f329;
mov.b32 %r878, %f330;
shfl.sync.bfly.b32 %r880|%p99, %r878, %r848, %r843, %r845;
mov.b32 %f331, %r880;
add.ftz.f32 %f332, %f330, %f331;
mov.b32 %r881, %f332;
shfl.sync.bfly.b32 %r883|%p100, %r881, %r851, %r843, %r845;
mov.b32 %f333, %r883;
add.ftz.f32 %f334, %f332, %f333;
mov.b32 %r884, %f334;
shfl.sync.bfly.b32 %r886|%p101, %r884, %r854, %r843, %r845;
mov.b32 %f335, %r886;
add.ftz.f32 %f336, %f334, %f335;
mov.b32 %r887, %f336;
shfl.sync.bfly.b32 %r889|%p102, %r887, %r857, %r843, %r845;
mov.b32 %f337, %r889;
add.ftz.f32 %f46, %f336, %f337;
mov.b32 %r890, %f328;
shfl.sync.bfly.b32 %r891|%p103, %r890, %r844, %r843, %r845;
mov.b32 %f338, %r891;
add.ftz.f32 %f339, %f328, %f338;
mov.b32 %r892, %f339;
shfl.sync.bfly.b32 %r893|%p104, %r892, %r848, %r843, %r845;
mov.b32 %f340, %r893;
add.ftz.f32 %f341, %f339, %f340;
mov.b32 %r894, %f341;
shfl.sync.bfly.b32 %r895|%p105, %r894, %r851, %r843, %r845;
mov.b32 %f342, %r895;
add.ftz.f32 %f343, %f341, %f342;
mov.b32 %r896, %f343;
shfl.sync.bfly.b32 %r897|%p106, %r896, %r854, %r843, %r845;
mov.b32 %f344, %r897;
add.ftz.f32 %f345, %f343, %f344;
mov.b32 %r898, %f345;
shfl.sync.bfly.b32 %r899|%p107, %r898, %r857, %r843, %r845;
mov.b32 %f346, %r899;
add.ftz.f32 %f47, %f345, %f346;
setp.ne.s32 %p108, %r2, 3;
or.b32 %r903, %r4, 3;
setp.ge.u32 %p109, %r903, %r52;
or.pred %p110, %p108, %p109;
@%p110 bra $L__BB2_78;

ld.param.f32 %f978, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f347, %f5, %f46;
add.ftz.f32 %f348, %f4, %f47;
min.ftz.f32 %f349, %f348, %f978;
min.ftz.f32 %f350, %f347, %f978;
max.ftz.f32 %f351, %f350, %f35;
mul.ftz.f32 %f352, %f349, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f353, %f352;
add.ftz.f32 %f354, %f353, 0f3F800000;
div.approx.ftz.f32 %f355, %f349, %f354;
mul.ftz.f32 %f356, %f351, %f355;
st.global.f32 [%rd23+12], %f356;
bra.uni $L__BB2_78;

$L__BB2_70:
ld.shared.f32 %f845, [%r33];
add.ftz.f32 %f846, %f845, %f29;
ld.shared.f32 %f847, [%r35];
add.ftz.f32 %f848, %f847, %f33;
ld.shared.f32 %f849, [%r33+512];
add.ftz.f32 %f850, %f849, %f846;
ld.shared.f32 %f851, [%r35+512];
add.ftz.f32 %f852, %f851, %f848;
ld.shared.f32 %f853, [%r33+1024];
add.ftz.f32 %f854, %f853, %f850;
ld.shared.f32 %f855, [%r35+1024];
add.ftz.f32 %f856, %f855, %f852;
mov.b32 %r1273, %f854;
mov.u32 %r1274, 31;
mov.u32 %r1275, 16;
mov.u32 %r1276, -1;
shfl.sync.bfly.b32 %r1277|%p267, %r1273, %r1275, %r1274, %r1276;
mov.b32 %f857, %r1277;
add.ftz.f32 %f858, %f854, %f857;
mov.b32 %r1278, %f858;
mov.u32 %r1279, 8;
shfl.sync.bfly.b32 %r1280|%p268, %r1278, %r1279, %r1274, %r1276;
mov.b32 %f859, %r1280;
add.ftz.f32 %f860, %f858, %f859;
mov.b32 %r1281, %f860;
mov.u32 %r1282, 4;
shfl.sync.bfly.b32 %r1283|%p269, %r1281, %r1282, %r1274, %r1276;
mov.b32 %f861, %r1283;
add.ftz.f32 %f862, %f860, %f861;
mov.b32 %r1284, %f862;
mov.u32 %r1285, 2;
shfl.sync.bfly.b32 %r1286|%p270, %r1284, %r1285, %r1274, %r1276;
mov.b32 %f863, %r1286;
add.ftz.f32 %f864, %f862, %f863;
mov.b32 %r1287, %f864;
mov.u32 %r1288, 1;
shfl.sync.bfly.b32 %r1289|%p271, %r1287, %r1288, %r1274, %r1276;
mov.b32 %f865, %r1289;
add.ftz.f32 %f72, %f864, %f865;
mov.b32 %r1290, %f856;
shfl.sync.bfly.b32 %r1291|%p272, %r1290, %r1275, %r1274, %r1276;
mov.b32 %f866, %r1291;
add.ftz.f32 %f867, %f856, %f866;
mov.b32 %r1292, %f867;
shfl.sync.bfly.b32 %r1293|%p273, %r1292, %r1279, %r1274, %r1276;
mov.b32 %f868, %r1293;
add.ftz.f32 %f869, %f867, %f868;
mov.b32 %r1294, %f869;
shfl.sync.bfly.b32 %r1295|%p274, %r1294, %r1282, %r1274, %r1276;
mov.b32 %f870, %r1295;
add.ftz.f32 %f871, %f869, %f870;
mov.b32 %r1296, %f871;
shfl.sync.bfly.b32 %r1297|%p275, %r1296, %r1285, %r1274, %r1276;
mov.b32 %f872, %r1297;
add.ftz.f32 %f873, %f871, %f872;
mov.b32 %r1298, %f873;
shfl.sync.bfly.b32 %r1299|%p276, %r1298, %r1288, %r1274, %r1276;
mov.b32 %f874, %r1299;
add.ftz.f32 %f73, %f873, %f874;
setp.ne.s32 %p277, %r2, 0;
setp.ge.u32 %p278, %r4, %r52;
or.pred %p279, %p277, %p278;
@%p279 bra $L__BB2_72;

add.ftz.f32 %f875, %f5, %f72;
add.ftz.f32 %f876, %f4, %f73;
mul.ftz.f32 %f877, %f875, %f876;
st.global.f32 [%rd23], %f877;

$L__BB2_72:
ld.shared.f32 %f878, [%r33+128];
add.ftz.f32 %f879, %f878, %f28;
ld.shared.f32 %f880, [%r35+128];
add.ftz.f32 %f881, %f880, %f32;
ld.shared.f32 %f882, [%r33+640];
add.ftz.f32 %f883, %f882, %f879;
ld.shared.f32 %f884, [%r35+640];
add.ftz.f32 %f885, %f884, %f881;
ld.shared.f32 %f886, [%r33+1152];
add.ftz.f32 %f887, %f886, %f883;
ld.shared.f32 %f888, [%r35+1152];
add.ftz.f32 %f889, %f888, %f885;
mov.b32 %r1303, %f887;
shfl.sync.bfly.b32 %r1307|%p280, %r1303, %r1275, %r1274, %r1276;
mov.b32 %f890, %r1307;
add.ftz.f32 %f891, %f887, %f890;
mov.b32 %r1308, %f891;
shfl.sync.bfly.b32 %r1310|%p281, %r1308, %r1279, %r1274, %r1276;
mov.b32 %f892, %r1310;
add.ftz.f32 %f893, %f891, %f892;
mov.b32 %r1311, %f893;
shfl.sync.bfly.b32 %r1313|%p282, %r1311, %r1282, %r1274, %r1276;
mov.b32 %f894, %r1313;
add.ftz.f32 %f895, %f893, %f894;
mov.b32 %r1314, %f895;
shfl.sync.bfly.b32 %r1316|%p283, %r1314, %r1285, %r1274, %r1276;
mov.b32 %f896, %r1316;
add.ftz.f32 %f897, %f895, %f896;
mov.b32 %r1317, %f897;
shfl.sync.bfly.b32 %r1319|%p284, %r1317, %r1288, %r1274, %r1276;
mov.b32 %f898, %r1319;
add.ftz.f32 %f74, %f897, %f898;
mov.b32 %r1320, %f889;
shfl.sync.bfly.b32 %r1321|%p285, %r1320, %r1275, %r1274, %r1276;
mov.b32 %f899, %r1321;
add.ftz.f32 %f900, %f889, %f899;
mov.b32 %r1322, %f900;
shfl.sync.bfly.b32 %r1323|%p286, %r1322, %r1279, %r1274, %r1276;
mov.b32 %f901, %r1323;
add.ftz.f32 %f902, %f900, %f901;
mov.b32 %r1324, %f902;
shfl.sync.bfly.b32 %r1325|%p287, %r1324, %r1282, %r1274, %r1276;
mov.b32 %f903, %r1325;
add.ftz.f32 %f904, %f902, %f903;
mov.b32 %r1326, %f904;
shfl.sync.bfly.b32 %r1327|%p288, %r1326, %r1285, %r1274, %r1276;
mov.b32 %f905, %r1327;
add.ftz.f32 %f906, %f904, %f905;
mov.b32 %r1328, %f906;
shfl.sync.bfly.b32 %r1329|%p289, %r1328, %r1288, %r1274, %r1276;
mov.b32 %f907, %r1329;
add.ftz.f32 %f75, %f906, %f907;
setp.ne.s32 %p290, %r2, 1;
or.b32 %r1333, %r4, 1;
setp.ge.u32 %p291, %r1333, %r52;
or.pred %p292, %p290, %p291;
@%p292 bra $L__BB2_74;

add.ftz.f32 %f908, %f5, %f74;
add.ftz.f32 %f909, %f4, %f75;
mul.ftz.f32 %f910, %f908, %f909;
st.global.f32 [%rd23+4], %f910;

$L__BB2_74:
ld.shared.f32 %f911, [%r33+256];
add.ftz.f32 %f912, %f911, %f27;
ld.shared.f32 %f913, [%r35+256];
add.ftz.f32 %f914, %f913, %f31;
ld.shared.f32 %f915, [%r33+768];
add.ftz.f32 %f916, %f915, %f912;
ld.shared.f32 %f917, [%r35+768];
add.ftz.f32 %f918, %f917, %f914;
ld.shared.f32 %f919, [%r33+1280];
add.ftz.f32 %f920, %f919, %f916;
ld.shared.f32 %f921, [%r35+1280];
add.ftz.f32 %f922, %f921, %f918;
mov.b32 %r1334, %f920;
mov.u32 %r1335, 31;
mov.u32 %r1336, 16;
mov.u32 %r1337, -1;
shfl.sync.bfly.b32 %r1338|%p293, %r1334, %r1336, %r1335, %r1337;
mov.b32 %f923, %r1338;
add.ftz.f32 %f924, %f920, %f923;
mov.b32 %r1339, %f924;
mov.u32 %r1340, 8;
shfl.sync.bfly.b32 %r1341|%p294, %r1339, %r1340, %r1335, %r1337;
mov.b32 %f925, %r1341;
add.ftz.f32 %f926, %f924, %f925;
mov.b32 %r1342, %f926;
mov.u32 %r1343, 4;
shfl.sync.bfly.b32 %r1344|%p295, %r1342, %r1343, %r1335, %r1337;
mov.b32 %f927, %r1344;
add.ftz.f32 %f928, %f926, %f927;
mov.b32 %r1345, %f928;
mov.u32 %r1346, 2;
shfl.sync.bfly.b32 %r1347|%p296, %r1345, %r1346, %r1335, %r1337;
mov.b32 %f929, %r1347;
add.ftz.f32 %f930, %f928, %f929;
mov.b32 %r1348, %f930;
mov.u32 %r1349, 1;
shfl.sync.bfly.b32 %r1350|%p297, %r1348, %r1349, %r1335, %r1337;
mov.b32 %f931, %r1350;
add.ftz.f32 %f76, %f930, %f931;
mov.b32 %r1351, %f922;
shfl.sync.bfly.b32 %r1352|%p298, %r1351, %r1336, %r1335, %r1337;
mov.b32 %f932, %r1352;
add.ftz.f32 %f933, %f922, %f932;
mov.b32 %r1353, %f933;
shfl.sync.bfly.b32 %r1354|%p299, %r1353, %r1340, %r1335, %r1337;
mov.b32 %f934, %r1354;
add.ftz.f32 %f935, %f933, %f934;
mov.b32 %r1355, %f935;
shfl.sync.bfly.b32 %r1356|%p300, %r1355, %r1343, %r1335, %r1337;
mov.b32 %f936, %r1356;
add.ftz.f32 %f937, %f935, %f936;
mov.b32 %r1357, %f937;
shfl.sync.bfly.b32 %r1358|%p301, %r1357, %r1346, %r1335, %r1337;
mov.b32 %f938, %r1358;
add.ftz.f32 %f939, %f937, %f938;
mov.b32 %r1359, %f939;
shfl.sync.bfly.b32 %r1360|%p302, %r1359, %r1349, %r1335, %r1337;
mov.b32 %f940, %r1360;
add.ftz.f32 %f77, %f939, %f940;
setp.ne.s32 %p303, %r2, 2;
or.b32 %r1364, %r4, 2;
setp.ge.u32 %p304, %r1364, %r52;
or.pred %p305, %p303, %p304;
@%p305 bra $L__BB2_76;

add.ftz.f32 %f941, %f5, %f76;
add.ftz.f32 %f942, %f4, %f77;
mul.ftz.f32 %f943, %f941, %f942;
st.global.f32 [%rd23+8], %f943;

$L__BB2_76:
ld.shared.f32 %f944, [%r33+384];
add.ftz.f32 %f945, %f944, %f26;
ld.shared.f32 %f946, [%r35+384];
add.ftz.f32 %f947, %f946, %f30;
ld.shared.f32 %f948, [%r33+896];
add.ftz.f32 %f949, %f948, %f945;
ld.shared.f32 %f950, [%r35+896];
add.ftz.f32 %f951, %f950, %f947;
ld.shared.f32 %f952, [%r33+1408];
add.ftz.f32 %f953, %f952, %f949;
ld.shared.f32 %f954, [%r35+1408];
add.ftz.f32 %f955, %f954, %f951;
mov.b32 %r1365, %f953;
shfl.sync.bfly.b32 %r1369|%p306, %r1365, %r1336, %r1335, %r1337;
mov.b32 %f956, %r1369;
add.ftz.f32 %f957, %f953, %f956;
mov.b32 %r1370, %f957;
shfl.sync.bfly.b32 %r1372|%p307, %r1370, %r1340, %r1335, %r1337;
mov.b32 %f958, %r1372;
add.ftz.f32 %f959, %f957, %f958;
mov.b32 %r1373, %f959;
shfl.sync.bfly.b32 %r1375|%p308, %r1373, %r1343, %r1335, %r1337;
mov.b32 %f960, %r1375;
add.ftz.f32 %f961, %f959, %f960;
mov.b32 %r1376, %f961;
shfl.sync.bfly.b32 %r1378|%p309, %r1376, %r1346, %r1335, %r1337;
mov.b32 %f962, %r1378;
add.ftz.f32 %f963, %f961, %f962;
mov.b32 %r1379, %f963;
shfl.sync.bfly.b32 %r1381|%p310, %r1379, %r1349, %r1335, %r1337;
mov.b32 %f964, %r1381;
add.ftz.f32 %f78, %f963, %f964;
mov.b32 %r1382, %f955;
shfl.sync.bfly.b32 %r1383|%p311, %r1382, %r1336, %r1335, %r1337;
mov.b32 %f965, %r1383;
add.ftz.f32 %f966, %f955, %f965;
mov.b32 %r1384, %f966;
shfl.sync.bfly.b32 %r1385|%p312, %r1384, %r1340, %r1335, %r1337;
mov.b32 %f967, %r1385;
add.ftz.f32 %f968, %f966, %f967;
mov.b32 %r1386, %f968;
shfl.sync.bfly.b32 %r1387|%p313, %r1386, %r1343, %r1335, %r1337;
mov.b32 %f969, %r1387;
add.ftz.f32 %f970, %f968, %f969;
mov.b32 %r1388, %f970;
shfl.sync.bfly.b32 %r1389|%p314, %r1388, %r1346, %r1335, %r1337;
mov.b32 %f971, %r1389;
add.ftz.f32 %f972, %f970, %f971;
mov.b32 %r1390, %f972;
shfl.sync.bfly.b32 %r1391|%p315, %r1390, %r1349, %r1335, %r1337;
mov.b32 %f973, %r1391;
add.ftz.f32 %f79, %f972, %f973;
setp.ne.s32 %p316, %r2, 3;
or.b32 %r1395, %r4, 3;
setp.ge.u32 %p317, %r1395, %r52;
or.pred %p318, %p316, %p317;
@%p318 bra $L__BB2_78;

add.ftz.f32 %f974, %f5, %f78;
add.ftz.f32 %f975, %f4, %f79;
mul.ftz.f32 %f976, %f974, %f975;
st.global.f32 [%rd23+12], %f976;
bra.uni $L__BB2_78;

$L__BB2_62:
ld.shared.f32 %f681, [%r33];
add.ftz.f32 %f682, %f681, %f29;
ld.shared.f32 %f683, [%r35];
add.ftz.f32 %f684, %f683, %f33;
ld.shared.f32 %f685, [%r33+512];
add.ftz.f32 %f686, %f685, %f682;
ld.shared.f32 %f687, [%r35+512];
add.ftz.f32 %f688, %f687, %f684;
ld.shared.f32 %f689, [%r33+1024];
add.ftz.f32 %f690, %f689, %f686;
ld.shared.f32 %f691, [%r35+1024];
add.ftz.f32 %f692, %f691, %f688;
mov.b32 %r1150, %f690;
mov.u32 %r1151, 31;
mov.u32 %r1152, 16;
mov.u32 %r1153, -1;
shfl.sync.bfly.b32 %r1154|%p215, %r1150, %r1152, %r1151, %r1153;
mov.b32 %f693, %r1154;
add.ftz.f32 %f694, %f690, %f693;
mov.b32 %r1155, %f694;
mov.u32 %r1156, 8;
shfl.sync.bfly.b32 %r1157|%p216, %r1155, %r1156, %r1151, %r1153;
mov.b32 %f695, %r1157;
add.ftz.f32 %f696, %f694, %f695;
mov.b32 %r1158, %f696;
mov.u32 %r1159, 4;
shfl.sync.bfly.b32 %r1160|%p217, %r1158, %r1159, %r1151, %r1153;
mov.b32 %f697, %r1160;
add.ftz.f32 %f698, %f696, %f697;
mov.b32 %r1161, %f698;
mov.u32 %r1162, 2;
shfl.sync.bfly.b32 %r1163|%p218, %r1161, %r1162, %r1151, %r1153;
mov.b32 %f699, %r1163;
add.ftz.f32 %f700, %f698, %f699;
mov.b32 %r1164, %f700;
mov.u32 %r1165, 1;
shfl.sync.bfly.b32 %r1166|%p219, %r1164, %r1165, %r1151, %r1153;
mov.b32 %f701, %r1166;
add.ftz.f32 %f64, %f700, %f701;
mov.b32 %r1167, %f692;
shfl.sync.bfly.b32 %r1168|%p220, %r1167, %r1152, %r1151, %r1153;
mov.b32 %f702, %r1168;
add.ftz.f32 %f703, %f692, %f702;
mov.b32 %r1169, %f703;
shfl.sync.bfly.b32 %r1170|%p221, %r1169, %r1156, %r1151, %r1153;
mov.b32 %f704, %r1170;
add.ftz.f32 %f705, %f703, %f704;
mov.b32 %r1171, %f705;
shfl.sync.bfly.b32 %r1172|%p222, %r1171, %r1159, %r1151, %r1153;
mov.b32 %f706, %r1172;
add.ftz.f32 %f707, %f705, %f706;
mov.b32 %r1173, %f707;
shfl.sync.bfly.b32 %r1174|%p223, %r1173, %r1162, %r1151, %r1153;
mov.b32 %f708, %r1174;
add.ftz.f32 %f709, %f707, %f708;
mov.b32 %r1175, %f709;
shfl.sync.bfly.b32 %r1176|%p224, %r1175, %r1165, %r1151, %r1153;
mov.b32 %f710, %r1176;
add.ftz.f32 %f65, %f709, %f710;
setp.ne.s32 %p225, %r2, 0;
setp.ge.u32 %p226, %r4, %r52;
or.pred %p227, %p225, %p226;
@%p227 bra $L__BB2_64;

add.ftz.f32 %f711, %f5, %f64;
add.ftz.f32 %f712, %f4, %f65;
mul.ftz.f32 %f713, %f712, 0f3F000000;
mul.ftz.f32 %f714, %f712, 0f3F4C422A;
mul.ftz.f32 %f715, %f712, 0f3D372713;
fma.rn.ftz.f32 %f716, %f712, %f715, 0f3F800000;
mul.ftz.f32 %f717, %f714, %f716;
tanh.approx.f32 %f718, %f717;
add.ftz.f32 %f719, %f718, 0f3F800000;
mul.ftz.f32 %f720, %f713, %f719;
mul.ftz.f32 %f721, %f711, %f720;
st.global.f32 [%rd23], %f721;

$L__BB2_64:
ld.shared.f32 %f722, [%r33+128];
add.ftz.f32 %f723, %f722, %f28;
ld.shared.f32 %f724, [%r35+128];
add.ftz.f32 %f725, %f724, %f32;
ld.shared.f32 %f726, [%r33+640];
add.ftz.f32 %f727, %f726, %f723;
ld.shared.f32 %f728, [%r35+640];
add.ftz.f32 %f729, %f728, %f725;
ld.shared.f32 %f730, [%r33+1152];
add.ftz.f32 %f731, %f730, %f727;
ld.shared.f32 %f732, [%r35+1152];
add.ftz.f32 %f733, %f732, %f729;
mov.b32 %r1180, %f731;
shfl.sync.bfly.b32 %r1184|%p228, %r1180, %r1152, %r1151, %r1153;
mov.b32 %f734, %r1184;
add.ftz.f32 %f735, %f731, %f734;
mov.b32 %r1185, %f735;
shfl.sync.bfly.b32 %r1187|%p229, %r1185, %r1156, %r1151, %r1153;
mov.b32 %f736, %r1187;
add.ftz.f32 %f737, %f735, %f736;
mov.b32 %r1188, %f737;
shfl.sync.bfly.b32 %r1190|%p230, %r1188, %r1159, %r1151, %r1153;
mov.b32 %f738, %r1190;
add.ftz.f32 %f739, %f737, %f738;
mov.b32 %r1191, %f739;
shfl.sync.bfly.b32 %r1193|%p231, %r1191, %r1162, %r1151, %r1153;
mov.b32 %f740, %r1193;
add.ftz.f32 %f741, %f739, %f740;
mov.b32 %r1194, %f741;
shfl.sync.bfly.b32 %r1196|%p232, %r1194, %r1165, %r1151, %r1153;
mov.b32 %f742, %r1196;
add.ftz.f32 %f66, %f741, %f742;
mov.b32 %r1197, %f733;
shfl.sync.bfly.b32 %r1198|%p233, %r1197, %r1152, %r1151, %r1153;
mov.b32 %f743, %r1198;
add.ftz.f32 %f744, %f733, %f743;
mov.b32 %r1199, %f744;
shfl.sync.bfly.b32 %r1200|%p234, %r1199, %r1156, %r1151, %r1153;
mov.b32 %f745, %r1200;
add.ftz.f32 %f746, %f744, %f745;
mov.b32 %r1201, %f746;
shfl.sync.bfly.b32 %r1202|%p235, %r1201, %r1159, %r1151, %r1153;
mov.b32 %f747, %r1202;
add.ftz.f32 %f748, %f746, %f747;
mov.b32 %r1203, %f748;
shfl.sync.bfly.b32 %r1204|%p236, %r1203, %r1162, %r1151, %r1153;
mov.b32 %f749, %r1204;
add.ftz.f32 %f750, %f748, %f749;
mov.b32 %r1205, %f750;
shfl.sync.bfly.b32 %r1206|%p237, %r1205, %r1165, %r1151, %r1153;
mov.b32 %f751, %r1206;
add.ftz.f32 %f67, %f750, %f751;
setp.ne.s32 %p238, %r2, 1;
or.b32 %r1210, %r4, 1;
setp.ge.u32 %p239, %r1210, %r52;
or.pred %p240, %p238, %p239;
@%p240 bra $L__BB2_66;

add.ftz.f32 %f752, %f5, %f66;
add.ftz.f32 %f753, %f4, %f67;
mul.ftz.f32 %f754, %f753, 0f3F000000;
mul.ftz.f32 %f755, %f753, 0f3F4C422A;
mul.ftz.f32 %f756, %f753, 0f3D372713;
fma.rn.ftz.f32 %f757, %f753, %f756, 0f3F800000;
mul.ftz.f32 %f758, %f755, %f757;
tanh.approx.f32 %f759, %f758;
add.ftz.f32 %f760, %f759, 0f3F800000;
mul.ftz.f32 %f761, %f754, %f760;
mul.ftz.f32 %f762, %f752, %f761;
st.global.f32 [%rd23+4], %f762;

$L__BB2_66:
ld.shared.f32 %f763, [%r33+256];
add.ftz.f32 %f764, %f763, %f27;
ld.shared.f32 %f765, [%r35+256];
add.ftz.f32 %f766, %f765, %f31;
ld.shared.f32 %f767, [%r33+768];
add.ftz.f32 %f768, %f767, %f764;
ld.shared.f32 %f769, [%r35+768];
add.ftz.f32 %f770, %f769, %f766;
ld.shared.f32 %f771, [%r33+1280];
add.ftz.f32 %f772, %f771, %f768;
ld.shared.f32 %f773, [%r35+1280];
add.ftz.f32 %f774, %f773, %f770;
mov.b32 %r1211, %f772;
mov.u32 %r1212, 31;
mov.u32 %r1213, 16;
mov.u32 %r1214, -1;
shfl.sync.bfly.b32 %r1215|%p241, %r1211, %r1213, %r1212, %r1214;
mov.b32 %f775, %r1215;
add.ftz.f32 %f776, %f772, %f775;
mov.b32 %r1216, %f776;
mov.u32 %r1217, 8;
shfl.sync.bfly.b32 %r1218|%p242, %r1216, %r1217, %r1212, %r1214;
mov.b32 %f777, %r1218;
add.ftz.f32 %f778, %f776, %f777;
mov.b32 %r1219, %f778;
mov.u32 %r1220, 4;
shfl.sync.bfly.b32 %r1221|%p243, %r1219, %r1220, %r1212, %r1214;
mov.b32 %f779, %r1221;
add.ftz.f32 %f780, %f778, %f779;
mov.b32 %r1222, %f780;
mov.u32 %r1223, 2;
shfl.sync.bfly.b32 %r1224|%p244, %r1222, %r1223, %r1212, %r1214;
mov.b32 %f781, %r1224;
add.ftz.f32 %f782, %f780, %f781;
mov.b32 %r1225, %f782;
mov.u32 %r1226, 1;
shfl.sync.bfly.b32 %r1227|%p245, %r1225, %r1226, %r1212, %r1214;
mov.b32 %f783, %r1227;
add.ftz.f32 %f68, %f782, %f783;
mov.b32 %r1228, %f774;
shfl.sync.bfly.b32 %r1229|%p246, %r1228, %r1213, %r1212, %r1214;
mov.b32 %f784, %r1229;
add.ftz.f32 %f785, %f774, %f784;
mov.b32 %r1230, %f785;
shfl.sync.bfly.b32 %r1231|%p247, %r1230, %r1217, %r1212, %r1214;
mov.b32 %f786, %r1231;
add.ftz.f32 %f787, %f785, %f786;
mov.b32 %r1232, %f787;
shfl.sync.bfly.b32 %r1233|%p248, %r1232, %r1220, %r1212, %r1214;
mov.b32 %f788, %r1233;
add.ftz.f32 %f789, %f787, %f788;
mov.b32 %r1234, %f789;
shfl.sync.bfly.b32 %r1235|%p249, %r1234, %r1223, %r1212, %r1214;
mov.b32 %f790, %r1235;
add.ftz.f32 %f791, %f789, %f790;
mov.b32 %r1236, %f791;
shfl.sync.bfly.b32 %r1237|%p250, %r1236, %r1226, %r1212, %r1214;
mov.b32 %f792, %r1237;
add.ftz.f32 %f69, %f791, %f792;
setp.ne.s32 %p251, %r2, 2;
or.b32 %r1241, %r4, 2;
setp.ge.u32 %p252, %r1241, %r52;
or.pred %p253, %p251, %p252;
@%p253 bra $L__BB2_68;

add.ftz.f32 %f793, %f5, %f68;
add.ftz.f32 %f794, %f4, %f69;
mul.ftz.f32 %f795, %f794, 0f3F000000;
mul.ftz.f32 %f796, %f794, 0f3F4C422A;
mul.ftz.f32 %f797, %f794, 0f3D372713;
fma.rn.ftz.f32 %f798, %f794, %f797, 0f3F800000;
mul.ftz.f32 %f799, %f796, %f798;
tanh.approx.f32 %f800, %f799;
add.ftz.f32 %f801, %f800, 0f3F800000;
mul.ftz.f32 %f802, %f795, %f801;
mul.ftz.f32 %f803, %f793, %f802;
st.global.f32 [%rd23+8], %f803;

$L__BB2_68:
ld.shared.f32 %f804, [%r33+384];
add.ftz.f32 %f805, %f804, %f26;
ld.shared.f32 %f806, [%r35+384];
add.ftz.f32 %f807, %f806, %f30;
ld.shared.f32 %f808, [%r33+896];
add.ftz.f32 %f809, %f808, %f805;
ld.shared.f32 %f810, [%r35+896];
add.ftz.f32 %f811, %f810, %f807;
ld.shared.f32 %f812, [%r33+1408];
add.ftz.f32 %f813, %f812, %f809;
ld.shared.f32 %f814, [%r35+1408];
add.ftz.f32 %f815, %f814, %f811;
mov.b32 %r1242, %f813;
shfl.sync.bfly.b32 %r1246|%p254, %r1242, %r1213, %r1212, %r1214;
mov.b32 %f816, %r1246;
add.ftz.f32 %f817, %f813, %f816;
mov.b32 %r1247, %f817;
shfl.sync.bfly.b32 %r1249|%p255, %r1247, %r1217, %r1212, %r1214;
mov.b32 %f818, %r1249;
add.ftz.f32 %f819, %f817, %f818;
mov.b32 %r1250, %f819;
shfl.sync.bfly.b32 %r1252|%p256, %r1250, %r1220, %r1212, %r1214;
mov.b32 %f820, %r1252;
add.ftz.f32 %f821, %f819, %f820;
mov.b32 %r1253, %f821;
shfl.sync.bfly.b32 %r1255|%p257, %r1253, %r1223, %r1212, %r1214;
mov.b32 %f822, %r1255;
add.ftz.f32 %f823, %f821, %f822;
mov.b32 %r1256, %f823;
shfl.sync.bfly.b32 %r1258|%p258, %r1256, %r1226, %r1212, %r1214;
mov.b32 %f824, %r1258;
add.ftz.f32 %f70, %f823, %f824;
mov.b32 %r1259, %f815;
shfl.sync.bfly.b32 %r1260|%p259, %r1259, %r1213, %r1212, %r1214;
mov.b32 %f825, %r1260;
add.ftz.f32 %f826, %f815, %f825;
mov.b32 %r1261, %f826;
shfl.sync.bfly.b32 %r1262|%p260, %r1261, %r1217, %r1212, %r1214;
mov.b32 %f827, %r1262;
add.ftz.f32 %f828, %f826, %f827;
mov.b32 %r1263, %f828;
shfl.sync.bfly.b32 %r1264|%p261, %r1263, %r1220, %r1212, %r1214;
mov.b32 %f829, %r1264;
add.ftz.f32 %f830, %f828, %f829;
mov.b32 %r1265, %f830;
shfl.sync.bfly.b32 %r1266|%p262, %r1265, %r1223, %r1212, %r1214;
mov.b32 %f831, %r1266;
add.ftz.f32 %f832, %f830, %f831;
mov.b32 %r1267, %f832;
shfl.sync.bfly.b32 %r1268|%p263, %r1267, %r1226, %r1212, %r1214;
mov.b32 %f833, %r1268;
add.ftz.f32 %f71, %f832, %f833;
setp.ne.s32 %p264, %r2, 3;
or.b32 %r1272, %r4, 3;
setp.ge.u32 %p265, %r1272, %r52;
or.pred %p266, %p264, %p265;
@%p266 bra $L__BB2_78;

add.ftz.f32 %f834, %f5, %f70;
add.ftz.f32 %f835, %f4, %f71;
mul.ftz.f32 %f836, %f835, 0f3F000000;
mul.ftz.f32 %f837, %f835, 0f3F4C422A;
mul.ftz.f32 %f838, %f835, 0f3D372713;
fma.rn.ftz.f32 %f839, %f835, %f838, 0f3F800000;
mul.ftz.f32 %f840, %f837, %f839;
tanh.approx.f32 %f841, %f840;
add.ftz.f32 %f842, %f841, 0f3F800000;
mul.ftz.f32 %f843, %f836, %f842;
mul.ftz.f32 %f844, %f834, %f843;
st.global.f32 [%rd23+12], %f844;
bra.uni $L__BB2_78;

$L__BB2_54:
ld.shared.f32 %f357, [%r33];
add.ftz.f32 %f358, %f357, %f29;
ld.shared.f32 %f359, [%r35];
add.ftz.f32 %f360, %f359, %f33;
ld.shared.f32 %f361, [%r33+512];
add.ftz.f32 %f362, %f361, %f358;
ld.shared.f32 %f363, [%r35+512];
add.ftz.f32 %f364, %f363, %f360;
ld.shared.f32 %f365, [%r33+1024];
add.ftz.f32 %f366, %f365, %f362;
ld.shared.f32 %f367, [%r35+1024];
add.ftz.f32 %f368, %f367, %f364;
mov.b32 %r904, %f366;
mov.u32 %r905, 31;
mov.u32 %r906, 16;
mov.u32 %r907, -1;
shfl.sync.bfly.b32 %r908|%p111, %r904, %r906, %r905, %r907;
mov.b32 %f369, %r908;
add.ftz.f32 %f370, %f366, %f369;
mov.b32 %r909, %f370;
mov.u32 %r910, 8;
shfl.sync.bfly.b32 %r911|%p112, %r909, %r910, %r905, %r907;
mov.b32 %f371, %r911;
add.ftz.f32 %f372, %f370, %f371;
mov.b32 %r912, %f372;
mov.u32 %r913, 4;
shfl.sync.bfly.b32 %r914|%p113, %r912, %r913, %r905, %r907;
mov.b32 %f373, %r914;
add.ftz.f32 %f374, %f372, %f373;
mov.b32 %r915, %f374;
mov.u32 %r916, 2;
shfl.sync.bfly.b32 %r917|%p114, %r915, %r916, %r905, %r907;
mov.b32 %f375, %r917;
add.ftz.f32 %f376, %f374, %f375;
mov.b32 %r918, %f376;
mov.u32 %r919, 1;
shfl.sync.bfly.b32 %r920|%p115, %r918, %r919, %r905, %r907;
mov.b32 %f377, %r920;
add.ftz.f32 %f48, %f376, %f377;
mov.b32 %r921, %f368;
shfl.sync.bfly.b32 %r922|%p116, %r921, %r906, %r905, %r907;
mov.b32 %f378, %r922;
add.ftz.f32 %f379, %f368, %f378;
mov.b32 %r923, %f379;
shfl.sync.bfly.b32 %r924|%p117, %r923, %r910, %r905, %r907;
mov.b32 %f380, %r924;
add.ftz.f32 %f381, %f379, %f380;
mov.b32 %r925, %f381;
shfl.sync.bfly.b32 %r926|%p118, %r925, %r913, %r905, %r907;
mov.b32 %f382, %r926;
add.ftz.f32 %f383, %f381, %f382;
mov.b32 %r927, %f383;
shfl.sync.bfly.b32 %r928|%p119, %r927, %r916, %r905, %r907;
mov.b32 %f384, %r928;
add.ftz.f32 %f385, %f383, %f384;
mov.b32 %r929, %f385;
shfl.sync.bfly.b32 %r930|%p120, %r929, %r919, %r905, %r907;
mov.b32 %f386, %r930;
add.ftz.f32 %f49, %f385, %f386;
setp.ne.s32 %p121, %r2, 0;
setp.ge.u32 %p122, %r4, %r52;
or.pred %p123, %p121, %p122;
@%p123 bra $L__BB2_56;

add.ftz.f32 %f387, %f5, %f48;
add.ftz.f32 %f388, %f4, %f49;
mov.f32 %f389, 0f40E00000;
min.ftz.f32 %f390, %f388, %f389;
min.ftz.f32 %f391, %f387, %f389;
mov.f32 %f392, 0fC0E00000;
max.ftz.f32 %f393, %f391, %f392;
mul.ftz.f32 %f394, %f390, 0f3FD9DB23;
mul.ftz.f32 %f395, %f394, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f396, %f395;
add.ftz.f32 %f397, %f396, 0f3F800000;
div.approx.ftz.f32 %f398, %f390, %f397;
add.ftz.f32 %f399, %f393, 0f3F800000;
mul.ftz.f32 %f400, %f399, %f398;
st.global.f32 [%rd23], %f400;

$L__BB2_56:
ld.shared.f32 %f401, [%r33+128];
add.ftz.f32 %f402, %f401, %f28;
ld.shared.f32 %f403, [%r35+128];
add.ftz.f32 %f404, %f403, %f32;
ld.shared.f32 %f405, [%r33+640];
add.ftz.f32 %f406, %f405, %f402;
ld.shared.f32 %f407, [%r35+640];
add.ftz.f32 %f408, %f407, %f404;
ld.shared.f32 %f409, [%r33+1152];
add.ftz.f32 %f410, %f409, %f406;
ld.shared.f32 %f411, [%r35+1152];
add.ftz.f32 %f412, %f411, %f408;
mov.b32 %r934, %f410;
shfl.sync.bfly.b32 %r938|%p124, %r934, %r906, %r905, %r907;
mov.b32 %f413, %r938;
add.ftz.f32 %f414, %f410, %f413;
mov.b32 %r939, %f414;
shfl.sync.bfly.b32 %r941|%p125, %r939, %r910, %r905, %r907;
mov.b32 %f415, %r941;
add.ftz.f32 %f416, %f414, %f415;
mov.b32 %r942, %f416;
shfl.sync.bfly.b32 %r944|%p126, %r942, %r913, %r905, %r907;
mov.b32 %f417, %r944;
add.ftz.f32 %f418, %f416, %f417;
mov.b32 %r945, %f418;
shfl.sync.bfly.b32 %r947|%p127, %r945, %r916, %r905, %r907;
mov.b32 %f419, %r947;
add.ftz.f32 %f420, %f418, %f419;
mov.b32 %r948, %f420;
shfl.sync.bfly.b32 %r950|%p128, %r948, %r919, %r905, %r907;
mov.b32 %f421, %r950;
add.ftz.f32 %f50, %f420, %f421;
mov.b32 %r951, %f412;
shfl.sync.bfly.b32 %r952|%p129, %r951, %r906, %r905, %r907;
mov.b32 %f422, %r952;
add.ftz.f32 %f423, %f412, %f422;
mov.b32 %r953, %f423;
shfl.sync.bfly.b32 %r954|%p130, %r953, %r910, %r905, %r907;
mov.b32 %f424, %r954;
add.ftz.f32 %f425, %f423, %f424;
mov.b32 %r955, %f425;
shfl.sync.bfly.b32 %r956|%p131, %r955, %r913, %r905, %r907;
mov.b32 %f426, %r956;
add.ftz.f32 %f427, %f425, %f426;
mov.b32 %r957, %f427;
shfl.sync.bfly.b32 %r958|%p132, %r957, %r916, %r905, %r907;
mov.b32 %f428, %r958;
add.ftz.f32 %f429, %f427, %f428;
mov.b32 %r959, %f429;
shfl.sync.bfly.b32 %r960|%p133, %r959, %r919, %r905, %r907;
mov.b32 %f430, %r960;
add.ftz.f32 %f51, %f429, %f430;
setp.ne.s32 %p134, %r2, 1;
or.b32 %r964, %r4, 1;
setp.ge.u32 %p135, %r964, %r52;
or.pred %p136, %p134, %p135;
@%p136 bra $L__BB2_58;

add.ftz.f32 %f431, %f5, %f50;
add.ftz.f32 %f432, %f4, %f51;
mov.f32 %f433, 0f40E00000;
min.ftz.f32 %f434, %f432, %f433;
min.ftz.f32 %f435, %f431, %f433;
mov.f32 %f436, 0fC0E00000;
max.ftz.f32 %f437, %f435, %f436;
mul.ftz.f32 %f438, %f434, 0f3FD9DB23;
mul.ftz.f32 %f439, %f438, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f440, %f439;
add.ftz.f32 %f441, %f440, 0f3F800000;
div.approx.ftz.f32 %f442, %f434, %f441;
add.ftz.f32 %f443, %f437, 0f3F800000;
mul.ftz.f32 %f444, %f443, %f442;
st.global.f32 [%rd23+4], %f444;

$L__BB2_58:
ld.shared.f32 %f445, [%r33+256];
add.ftz.f32 %f446, %f445, %f27;
ld.shared.f32 %f447, [%r35+256];
add.ftz.f32 %f448, %f447, %f31;
ld.shared.f32 %f449, [%r33+768];
add.ftz.f32 %f450, %f449, %f446;
ld.shared.f32 %f451, [%r35+768];
add.ftz.f32 %f452, %f451, %f448;
ld.shared.f32 %f453, [%r33+1280];
add.ftz.f32 %f454, %f453, %f450;
ld.shared.f32 %f455, [%r35+1280];
add.ftz.f32 %f456, %f455, %f452;
mov.b32 %r965, %f454;
mov.u32 %r966, 31;
mov.u32 %r967, 16;
mov.u32 %r968, -1;
shfl.sync.bfly.b32 %r969|%p137, %r965, %r967, %r966, %r968;
mov.b32 %f457, %r969;
add.ftz.f32 %f458, %f454, %f457;
mov.b32 %r970, %f458;
mov.u32 %r971, 8;
shfl.sync.bfly.b32 %r972|%p138, %r970, %r971, %r966, %r968;
mov.b32 %f459, %r972;
add.ftz.f32 %f460, %f458, %f459;
mov.b32 %r973, %f460;
mov.u32 %r974, 4;
shfl.sync.bfly.b32 %r975|%p139, %r973, %r974, %r966, %r968;
mov.b32 %f461, %r975;
add.ftz.f32 %f462, %f460, %f461;
mov.b32 %r976, %f462;
mov.u32 %r977, 2;
shfl.sync.bfly.b32 %r978|%p140, %r976, %r977, %r966, %r968;
mov.b32 %f463, %r978;
add.ftz.f32 %f464, %f462, %f463;
mov.b32 %r979, %f464;
mov.u32 %r980, 1;
shfl.sync.bfly.b32 %r981|%p141, %r979, %r980, %r966, %r968;
mov.b32 %f465, %r981;
add.ftz.f32 %f52, %f464, %f465;
mov.b32 %r982, %f456;
shfl.sync.bfly.b32 %r983|%p142, %r982, %r967, %r966, %r968;
mov.b32 %f466, %r983;
add.ftz.f32 %f467, %f456, %f466;
mov.b32 %r984, %f467;
shfl.sync.bfly.b32 %r985|%p143, %r984, %r971, %r966, %r968;
mov.b32 %f468, %r985;
add.ftz.f32 %f469, %f467, %f468;
mov.b32 %r986, %f469;
shfl.sync.bfly.b32 %r987|%p144, %r986, %r974, %r966, %r968;
mov.b32 %f470, %r987;
add.ftz.f32 %f471, %f469, %f470;
mov.b32 %r988, %f471;
shfl.sync.bfly.b32 %r989|%p145, %r988, %r977, %r966, %r968;
mov.b32 %f472, %r989;
add.ftz.f32 %f473, %f471, %f472;
mov.b32 %r990, %f473;
shfl.sync.bfly.b32 %r991|%p146, %r990, %r980, %r966, %r968;
mov.b32 %f474, %r991;
add.ftz.f32 %f53, %f473, %f474;
setp.ne.s32 %p147, %r2, 2;
or.b32 %r995, %r4, 2;
setp.ge.u32 %p148, %r995, %r52;
or.pred %p149, %p147, %p148;
@%p149 bra $L__BB2_60;

add.ftz.f32 %f475, %f5, %f52;
add.ftz.f32 %f476, %f4, %f53;
mov.f32 %f477, 0f40E00000;
min.ftz.f32 %f478, %f476, %f477;
min.ftz.f32 %f479, %f475, %f477;
mov.f32 %f480, 0fC0E00000;
max.ftz.f32 %f481, %f479, %f480;
mul.ftz.f32 %f482, %f478, 0f3FD9DB23;
mul.ftz.f32 %f483, %f482, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f484, %f483;
add.ftz.f32 %f485, %f484, 0f3F800000;
div.approx.ftz.f32 %f486, %f478, %f485;
add.ftz.f32 %f487, %f481, 0f3F800000;
mul.ftz.f32 %f488, %f487, %f486;
st.global.f32 [%rd23+8], %f488;

$L__BB2_60:
ld.shared.f32 %f489, [%r33+384];
add.ftz.f32 %f490, %f489, %f26;
ld.shared.f32 %f491, [%r35+384];
add.ftz.f32 %f492, %f491, %f30;
ld.shared.f32 %f493, [%r33+896];
add.ftz.f32 %f494, %f493, %f490;
ld.shared.f32 %f495, [%r35+896];
add.ftz.f32 %f496, %f495, %f492;
ld.shared.f32 %f497, [%r33+1408];
add.ftz.f32 %f498, %f497, %f494;
ld.shared.f32 %f499, [%r35+1408];
add.ftz.f32 %f500, %f499, %f496;
mov.b32 %r996, %f498;
shfl.sync.bfly.b32 %r1000|%p150, %r996, %r967, %r966, %r968;
mov.b32 %f501, %r1000;
add.ftz.f32 %f502, %f498, %f501;
mov.b32 %r1001, %f502;
shfl.sync.bfly.b32 %r1003|%p151, %r1001, %r971, %r966, %r968;
mov.b32 %f503, %r1003;
add.ftz.f32 %f504, %f502, %f503;
mov.b32 %r1004, %f504;
shfl.sync.bfly.b32 %r1006|%p152, %r1004, %r974, %r966, %r968;
mov.b32 %f505, %r1006;
add.ftz.f32 %f506, %f504, %f505;
mov.b32 %r1007, %f506;
shfl.sync.bfly.b32 %r1009|%p153, %r1007, %r977, %r966, %r968;
mov.b32 %f507, %r1009;
add.ftz.f32 %f508, %f506, %f507;
mov.b32 %r1010, %f508;
shfl.sync.bfly.b32 %r1012|%p154, %r1010, %r980, %r966, %r968;
mov.b32 %f509, %r1012;
add.ftz.f32 %f54, %f508, %f509;
mov.b32 %r1013, %f500;
shfl.sync.bfly.b32 %r1014|%p155, %r1013, %r967, %r966, %r968;
mov.b32 %f510, %r1014;
add.ftz.f32 %f511, %f500, %f510;
mov.b32 %r1015, %f511;
shfl.sync.bfly.b32 %r1016|%p156, %r1015, %r971, %r966, %r968;
mov.b32 %f512, %r1016;
add.ftz.f32 %f513, %f511, %f512;
mov.b32 %r1017, %f513;
shfl.sync.bfly.b32 %r1018|%p157, %r1017, %r974, %r966, %r968;
mov.b32 %f514, %r1018;
add.ftz.f32 %f515, %f513, %f514;
mov.b32 %r1019, %f515;
shfl.sync.bfly.b32 %r1020|%p158, %r1019, %r977, %r966, %r968;
mov.b32 %f516, %r1020;
add.ftz.f32 %f517, %f515, %f516;
mov.b32 %r1021, %f517;
shfl.sync.bfly.b32 %r1022|%p159, %r1021, %r980, %r966, %r968;
mov.b32 %f518, %r1022;
add.ftz.f32 %f55, %f517, %f518;
setp.ne.s32 %p160, %r2, 3;
or.b32 %r1026, %r4, 3;
setp.ge.u32 %p161, %r1026, %r52;
or.pred %p162, %p160, %p161;
@%p162 bra $L__BB2_78;

add.ftz.f32 %f519, %f5, %f54;
add.ftz.f32 %f520, %f4, %f55;
mov.f32 %f521, 0f40E00000;
min.ftz.f32 %f522, %f520, %f521;
min.ftz.f32 %f523, %f519, %f521;
mov.f32 %f524, 0fC0E00000;
max.ftz.f32 %f525, %f523, %f524;
mul.ftz.f32 %f526, %f522, 0f3FD9DB23;
mul.ftz.f32 %f527, %f526, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f528, %f527;
add.ftz.f32 %f529, %f528, 0f3F800000;
div.approx.ftz.f32 %f530, %f522, %f529;
add.ftz.f32 %f531, %f525, 0f3F800000;
mul.ftz.f32 %f532, %f531, %f530;
st.global.f32 [%rd23+12], %f532;
bra.uni $L__BB2_78;

}
.entry _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<38>;
.reg .b16 %rs<13>;
.reg .f32 %f<109>;
.reg .b32 %r<418>;
.reg .b64 %rd<39>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

ld.param.u64 %rd10, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd11, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd12, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd13, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r19, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
mov.u32 %r1, %tid.y;
mov.u32 %r2, %tid.x;
mov.u32 %r37, %ctaid.x;
shl.b32 %r3, %r37, 2;
shr.u32 %r4, %r36, 7;
mov.u32 %r415, %ctaid.y;
setp.eq.s64 %p1, %rd12, 0;
@%p1 bra $L__BB3_2;

cvta.to.global.u64 %rd14, %rd12;
mul.wide.u32 %rd15, %r415, 4;
add.s64 %rd16, %rd14, %rd15;
ld.global.nc.u32 %r414, [%rd16];
mul.hi.u32 %r38, %r415, %r19;
add.s32 %r39, %r38, %r415;
shr.u32 %r40, %r39, %r20;
mul.lo.s32 %r41, %r40, %r21;
sub.s32 %r415, %r415, %r41;
bra.uni $L__BB3_3;

$L__BB3_2:
mul.hi.u32 %r42, %r415, %r24;
add.s32 %r43, %r42, %r415;
shr.u32 %r414, %r43, %r25;

$L__BB3_3:
shl.b32 %r44, %r1, 5;
add.s32 %r45, %r44, %r2;
shr.s32 %r46, %r45, 31;
shr.u32 %r47, %r46, 30;
add.s32 %r48, %r45, %r47;
shr.s32 %r417, %r48, 2;
setp.ge.s32 %p2, %r417, %r4;
mov.f32 %f105, 0f00000000;
mov.f32 %f106, %f105;
mov.f32 %f107, %f105;
mov.f32 %f108, %f105;
@%p2 bra $L__BB3_6;

mov.u32 %r49, %ctaid.z;
mul.lo.s32 %r50, %r415, %r28;
cvt.u64.u32 %rd17, %r50;
mul.lo.s32 %r51, %r49, %r34;
cvt.u64.u32 %rd18, %r51;
and.b32 %r59, %r48, -4;
sub.s32 %r60, %r45, %r59;
cvt.s64.s32 %rd19, %r60;
add.s64 %rd20, %rd18, %rd19;
add.s64 %rd1, %rd20, %rd17;
shl.b32 %r61, %r60, 1;
shl.b32 %r416, %r417, 2;
mad.lo.s32 %r62, %r3, %r22, %r417;
mul.hi.u32 %r63, %r49, %r30;
add.s32 %r64, %r63, %r49;
shr.u32 %r65, %r64, %r31;
mad.lo.s32 %r66, %r65, %r33, %r62;
mad.lo.s32 %r67, %r414, %r27, %r66;
mul.wide.s32 %rd2, %r67, 18;
mul.wide.s32 %rd21, %r61, 2;
add.s64 %rd22, %rd2, %rd21;
add.s64 %rd3, %rd22, 4;
mul.wide.s32 %rd4, %r22, 18;
cvta.to.global.u64 %rd38, %rd10;
cvta.to.global.u64 %rd6, %rd11;
mov.f32 %f105, 0f00000000;

$L__BB3_5:
cvt.u64.u32 %rd23, %r416;
add.s64 %rd24, %rd1, %rd23;
mul.lo.s64 %rd25, %rd24, 36;
add.s64 %rd26, %rd38, %rd2;
ld.global.nc.u16 %rs1, [%rd26];

{ cvt.f32.f16 %f25, %rs1;}


add.s64 %rd27, %rd38, %rd3;
add.s64 %rd28, %rd27, -2;
ld.global.nc.u16 %rs5, [%rd27+-2];
cvt.s32.s16 %r200, %rs5;
add.s64 %rd29, %rd6, %rd25;
ld.global.nc.u32 %r70, [%rd29+4];
ld.global.nc.u32 %r74, [%rd29+8];
ld.global.nc.u32 %r78, [%rd29+12];
ld.global.nc.u32 %r82, [%rd29+16];
and.b32 %r201, %r200, 30583;
mov.u32 %r202, 286261504;
prmt.b32 %r203, %r202, %r202, %r201;
shr.u32 %r204, %r200, 2;
and.b32 %r205, %r204, 30583;
prmt.b32 %r206, %r202, %r202, %r205;
and.b32 %r207, %r203, 30583;
mov.u32 %r208, 511;
prmt.b32 %r209, %r208, %r208, %r207;
and.b32 %r210, %r206, 30583;
prmt.b32 %r211, %r208, %r208, %r210;
shr.u32 %r212, %r203, 16;
and.b32 %r213, %r212, 30583;
prmt.b32 %r214, %r208, %r208, %r213;
shr.u32 %r215, %r206, 16;
and.b32 %r216, %r215, 30583;
prmt.b32 %r217, %r208, %r208, %r216;
mov.u32 %r218, 21520;
prmt.b32 %r69, %r209, %r211, %r218;
mov.u32 %r219, 30258;
prmt.b32 %r73, %r209, %r211, %r219;
prmt.b32 %r77, %r214, %r217, %r218;
prmt.b32 %r81, %r214, %r217, %r219;
mov.u32 %r170, 0;

dp4a.s32.s32 %r68, %r69, %r70, %r170;


dp4a.s32.s32 %r72, %r73, %r74, %r68;


dp4a.s32.s32 %r76, %r77, %r78, %r72;


dp4a.s32.s32 %r80, %r81, %r82, %r76;

ld.global.nc.u16 %rs6, [%rd27];
cvt.s32.s16 %r220, %rs6;
ld.global.nc.u32 %r86, [%rd29+20];
ld.global.nc.u32 %r90, [%rd29+24];
ld.global.nc.u32 %r94, [%rd29+28];
ld.global.nc.u32 %r98, [%rd29+32];
and.b32 %r221, %r220, 30583;
prmt.b32 %r222, %r202, %r202, %r221;
shr.u32 %r223, %r220, 2;
and.b32 %r224, %r223, 30583;
prmt.b32 %r225, %r202, %r202, %r224;
and.b32 %r226, %r222, 30583;
prmt.b32 %r227, %r208, %r208, %r226;
and.b32 %r228, %r225, 30583;
prmt.b32 %r229, %r208, %r208, %r228;
shr.u32 %r230, %r222, 16;
and.b32 %r231, %r230, 30583;
prmt.b32 %r232, %r208, %r208, %r231;
shr.u32 %r233, %r225, 16;
and.b32 %r234, %r233, 30583;
prmt.b32 %r235, %r208, %r208, %r234;
prmt.b32 %r85, %r227, %r229, %r218;
prmt.b32 %r89, %r227, %r229, %r219;
prmt.b32 %r93, %r232, %r235, %r218;
prmt.b32 %r97, %r232, %r235, %r219;

dp4a.s32.s32 %r84, %r85, %r86, %r80;


dp4a.s32.s32 %r88, %r89, %r90, %r84;


dp4a.s32.s32 %r92, %r93, %r94, %r88;


dp4a.s32.s32 %r96, %r97, %r98, %r92;

ld.global.nc.u32 %r100, [%rd29];

{.reg .f16 low,high;
mov.b32 {low,high},%r100;
cvt.f32.f16 %f26, low;}


mul.ftz.f32 %f33, %f25, %f26;
cvt.rn.f32.s32 %f34, %r96;
fma.rn.ftz.f32 %f108, %f33, %f34, %f108;
add.s64 %rd30, %rd26, %rd4;
ld.global.nc.u16 %rs2, [%rd30];

{ cvt.f32.f16 %f27, %rs2;}


add.s64 %rd31, %rd28, %rd4;
ld.global.nc.u16 %rs7, [%rd31];
cvt.s32.s16 %r236, %rs7;
and.b32 %r237, %r236, 30583;
prmt.b32 %r238, %r202, %r202, %r237;
shr.u32 %r239, %r236, 2;
and.b32 %r240, %r239, 30583;
prmt.b32 %r241, %r202, %r202, %r240;
and.b32 %r242, %r238, 30583;
prmt.b32 %r243, %r208, %r208, %r242;
and.b32 %r244, %r241, 30583;
prmt.b32 %r245, %r208, %r208, %r244;
shr.u32 %r246, %r238, 16;
and.b32 %r247, %r246, 30583;
prmt.b32 %r248, %r208, %r208, %r247;
shr.u32 %r249, %r241, 16;
and.b32 %r250, %r249, 30583;
prmt.b32 %r251, %r208, %r208, %r250;
prmt.b32 %r102, %r243, %r245, %r218;
prmt.b32 %r106, %r243, %r245, %r219;
prmt.b32 %r110, %r248, %r251, %r218;
prmt.b32 %r114, %r248, %r251, %r219;

dp4a.s32.s32 %r101, %r102, %r70, %r170;


dp4a.s32.s32 %r105, %r106, %r74, %r101;


dp4a.s32.s32 %r109, %r110, %r78, %r105;


dp4a.s32.s32 %r113, %r114, %r82, %r109;

ld.global.nc.u16 %rs8, [%rd31+2];
cvt.s32.s16 %r252, %rs8;
and.b32 %r253, %r252, 30583;
prmt.b32 %r254, %r202, %r202, %r253;
shr.u32 %r255, %r252, 2;
and.b32 %r256, %r255, 30583;
prmt.b32 %r257, %r202, %r202, %r256;
and.b32 %r258, %r254, 30583;
prmt.b32 %r259, %r208, %r208, %r258;
and.b32 %r260, %r257, 30583;
prmt.b32 %r261, %r208, %r208, %r260;
shr.u32 %r262, %r254, 16;
and.b32 %r263, %r262, 30583;
prmt.b32 %r264, %r208, %r208, %r263;
shr.u32 %r265, %r257, 16;
and.b32 %r266, %r265, 30583;
prmt.b32 %r267, %r208, %r208, %r266;
prmt.b32 %r118, %r259, %r261, %r218;
prmt.b32 %r122, %r259, %r261, %r219;
prmt.b32 %r126, %r264, %r267, %r218;
prmt.b32 %r130, %r264, %r267, %r219;

dp4a.s32.s32 %r117, %r118, %r86, %r113;


dp4a.s32.s32 %r121, %r122, %r90, %r117;


dp4a.s32.s32 %r125, %r126, %r94, %r121;


dp4a.s32.s32 %r129, %r130, %r98, %r125;


{.reg .f16 low,high;
mov.b32 {low,high},%r100;
cvt.f32.f16 %f28, low;}


mul.ftz.f32 %f35, %f27, %f28;
cvt.rn.f32.s32 %f36, %r129;
fma.rn.ftz.f32 %f107, %f35, %f36, %f107;
add.s64 %rd32, %rd30, %rd4;
ld.global.nc.u16 %rs3, [%rd32];

{ cvt.f32.f16 %f29, %rs3;}


add.s64 %rd33, %rd31, %rd4;
ld.global.nc.u16 %rs9, [%rd33];
cvt.s32.s16 %r268, %rs9;
and.b32 %r269, %r268, 30583;
prmt.b32 %r270, %r202, %r202, %r269;
shr.u32 %r271, %r268, 2;
and.b32 %r272, %r271, 30583;
prmt.b32 %r273, %r202, %r202, %r272;
and.b32 %r274, %r270, 30583;
prmt.b32 %r275, %r208, %r208, %r274;
and.b32 %r276, %r273, 30583;
prmt.b32 %r277, %r208, %r208, %r276;
shr.u32 %r278, %r270, 16;
and.b32 %r279, %r278, 30583;
prmt.b32 %r280, %r208, %r208, %r279;
shr.u32 %r281, %r273, 16;
and.b32 %r282, %r281, 30583;
prmt.b32 %r283, %r208, %r208, %r282;
prmt.b32 %r135, %r275, %r277, %r218;
prmt.b32 %r139, %r275, %r277, %r219;
prmt.b32 %r143, %r280, %r283, %r218;
prmt.b32 %r147, %r280, %r283, %r219;

dp4a.s32.s32 %r134, %r135, %r70, %r170;


dp4a.s32.s32 %r138, %r139, %r74, %r134;


dp4a.s32.s32 %r142, %r143, %r78, %r138;


dp4a.s32.s32 %r146, %r147, %r82, %r142;

ld.global.nc.u16 %rs10, [%rd33+2];
cvt.s32.s16 %r284, %rs10;
and.b32 %r285, %r284, 30583;
prmt.b32 %r286, %r202, %r202, %r285;
shr.u32 %r287, %r284, 2;
and.b32 %r288, %r287, 30583;
prmt.b32 %r289, %r202, %r202, %r288;
and.b32 %r290, %r286, 30583;
prmt.b32 %r291, %r208, %r208, %r290;
and.b32 %r292, %r289, 30583;
prmt.b32 %r293, %r208, %r208, %r292;
shr.u32 %r294, %r286, 16;
and.b32 %r295, %r294, 30583;
prmt.b32 %r296, %r208, %r208, %r295;
shr.u32 %r297, %r289, 16;
and.b32 %r298, %r297, 30583;
prmt.b32 %r299, %r208, %r208, %r298;
prmt.b32 %r151, %r291, %r293, %r218;
prmt.b32 %r155, %r291, %r293, %r219;
prmt.b32 %r159, %r296, %r299, %r218;
prmt.b32 %r163, %r296, %r299, %r219;

dp4a.s32.s32 %r150, %r151, %r86, %r146;


dp4a.s32.s32 %r154, %r155, %r90, %r150;


dp4a.s32.s32 %r158, %r159, %r94, %r154;


dp4a.s32.s32 %r162, %r163, %r98, %r158;


{.reg .f16 low,high;
mov.b32 {low,high},%r100;
cvt.f32.f16 %f30, low;}


mul.ftz.f32 %f37, %f29, %f30;
cvt.rn.f32.s32 %f38, %r162;
fma.rn.ftz.f32 %f106, %f37, %f38, %f106;
add.s64 %rd34, %rd32, %rd4;
ld.global.nc.u16 %rs4, [%rd34];

{ cvt.f32.f16 %f31, %rs4;}


add.s64 %rd35, %rd33, %rd4;
ld.global.nc.u16 %rs11, [%rd35];
cvt.s32.s16 %r300, %rs11;
and.b32 %r301, %r300, 30583;
prmt.b32 %r302, %r202, %r202, %r301;
shr.u32 %r303, %r300, 2;
and.b32 %r304, %r303, 30583;
prmt.b32 %r305, %r202, %r202, %r304;
and.b32 %r306, %r302, 30583;
prmt.b32 %r307, %r208, %r208, %r306;
and.b32 %r308, %r305, 30583;
prmt.b32 %r309, %r208, %r208, %r308;
shr.u32 %r310, %r302, 16;
and.b32 %r311, %r310, 30583;
prmt.b32 %r312, %r208, %r208, %r311;
shr.u32 %r313, %r305, 16;
and.b32 %r314, %r313, 30583;
prmt.b32 %r315, %r208, %r208, %r314;
prmt.b32 %r168, %r307, %r309, %r218;
prmt.b32 %r172, %r307, %r309, %r219;
prmt.b32 %r176, %r312, %r315, %r218;
prmt.b32 %r180, %r312, %r315, %r219;

dp4a.s32.s32 %r167, %r168, %r70, %r170;


dp4a.s32.s32 %r171, %r172, %r74, %r167;


dp4a.s32.s32 %r175, %r176, %r78, %r171;


dp4a.s32.s32 %r179, %r180, %r82, %r175;

ld.global.nc.u16 %rs12, [%rd35+2];
cvt.s32.s16 %r316, %rs12;
and.b32 %r317, %r316, 30583;
prmt.b32 %r318, %r202, %r202, %r317;
shr.u32 %r319, %r316, 2;
and.b32 %r320, %r319, 30583;
prmt.b32 %r321, %r202, %r202, %r320;
and.b32 %r322, %r318, 30583;
prmt.b32 %r323, %r208, %r208, %r322;
and.b32 %r324, %r321, 30583;
prmt.b32 %r325, %r208, %r208, %r324;
shr.u32 %r326, %r318, 16;
and.b32 %r327, %r326, 30583;
prmt.b32 %r328, %r208, %r208, %r327;
shr.u32 %r329, %r321, 16;
and.b32 %r330, %r329, 30583;
prmt.b32 %r331, %r208, %r208, %r330;
prmt.b32 %r184, %r323, %r325, %r218;
prmt.b32 %r188, %r323, %r325, %r219;
prmt.b32 %r192, %r328, %r331, %r218;
prmt.b32 %r196, %r328, %r331, %r219;

dp4a.s32.s32 %r183, %r184, %r86, %r179;


dp4a.s32.s32 %r187, %r188, %r90, %r183;


dp4a.s32.s32 %r191, %r192, %r94, %r187;


dp4a.s32.s32 %r195, %r196, %r98, %r191;


{.reg .f16 low,high;
mov.b32 {low,high},%r100;
cvt.f32.f16 %f32, low;}


mul.ftz.f32 %f39, %f31, %f32;
cvt.rn.f32.s32 %f40, %r195;
fma.rn.ftz.f32 %f105, %f39, %f40, %f105;
add.s32 %r416, %r416, 128;
add.s64 %rd38, %rd38, 576;
add.s32 %r417, %r417, 32;
setp.lt.s32 %p3, %r417, %r4;
@%p3 bra $L__BB3_5;

$L__BB3_6:
setp.eq.s32 %p4, %r1, 0;
@%p4 bra $L__BB3_8;

shl.b32 %r332, %r1, 9;
mov.u32 %r333, _ZZ13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r334, %r333, %r332;
shl.b32 %r335, %r2, 2;
add.s32 %r336, %r334, %r335;
st.shared.f32 [%r336+-512], %f108;
st.shared.f32 [%r336+-384], %f107;
st.shared.f32 [%r336+-256], %f106;
st.shared.f32 [%r336+-128], %f105;

$L__BB3_8:
setp.ne.s32 %p5, %r1, 0;
bar.sync 0;
@%p5 bra $L__BB3_17;

mov.u32 %r337, %ctaid.z;
mov.u32 %r338, %ctaid.y;
mad.lo.s32 %r339, %r338, %r29, %r3;
mad.lo.s32 %r340, %r337, %r35, %r339;
shl.b32 %r341, %r2, 2;
mov.u32 %r342, _ZZ13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r17, %r342, %r341;
mov.u32 %r343, 2;
ld.shared.f32 %f41, [%r17];
add.ftz.f32 %f42, %f41, %f108;
ld.shared.f32 %f43, [%r17+512];
add.ftz.f32 %f44, %f43, %f42;
ld.shared.f32 %f45, [%r17+1024];
add.ftz.f32 %f46, %f45, %f44;
mov.b32 %r344, %f46;
mov.u32 %r345, 31;
mov.u32 %r346, 16;
mov.u32 %r347, -1;
shfl.sync.bfly.b32 %r348|%p6, %r344, %r346, %r345, %r347;
mov.b32 %f47, %r348;
add.ftz.f32 %f48, %f46, %f47;
mov.b32 %r349, %f48;
mov.u32 %r350, 8;
shfl.sync.bfly.b32 %r351|%p7, %r349, %r350, %r345, %r347;
mov.b32 %f49, %r351;
add.ftz.f32 %f50, %f48, %f49;
mov.b32 %r352, %f50;
mov.u32 %r353, 4;
shfl.sync.bfly.b32 %r354|%p8, %r352, %r353, %r345, %r347;
mov.b32 %f51, %r354;
add.ftz.f32 %f52, %f50, %f51;
mov.b32 %r355, %f52;
shfl.sync.bfly.b32 %r356|%p9, %r355, %r343, %r345, %r347;
mov.b32 %f53, %r356;
add.ftz.f32 %f54, %f52, %f53;
mov.b32 %r357, %f54;
mov.u32 %r358, 1;
shfl.sync.bfly.b32 %r359|%p10, %r357, %r358, %r345, %r347;
mov.b32 %f55, %r359;
add.ftz.f32 %f13, %f54, %f55;
setp.ge.u32 %p11, %r3, %r23;
setp.ne.s32 %p12, %r2, 0;
cvta.to.global.u64 %rd36, %rd13;
mul.wide.u32 %rd37, %r340, 4;
add.s64 %rd9, %rd36, %rd37;
or.pred %p13, %p12, %p11;
@%p13 bra $L__BB3_11;

st.global.f32 [%rd9], %f13;

$L__BB3_11:
ld.shared.f32 %f56, [%r17+128];
add.ftz.f32 %f57, %f56, %f107;
ld.shared.f32 %f58, [%r17+640];
add.ftz.f32 %f59, %f58, %f57;
ld.shared.f32 %f60, [%r17+1152];
add.ftz.f32 %f61, %f60, %f59;
mov.b32 %r360, %f61;
shfl.sync.bfly.b32 %r364|%p14, %r360, %r346, %r345, %r347;
mov.b32 %f62, %r364;
add.ftz.f32 %f63, %f61, %f62;
mov.b32 %r365, %f63;
shfl.sync.bfly.b32 %r367|%p15, %r365, %r350, %r345, %r347;
mov.b32 %f64, %r367;
add.ftz.f32 %f65, %f63, %f64;
mov.b32 %r368, %f65;
shfl.sync.bfly.b32 %r370|%p16, %r368, %r353, %r345, %r347;
mov.b32 %f66, %r370;
add.ftz.f32 %f67, %f65, %f66;
mov.b32 %r371, %f67;
shfl.sync.bfly.b32 %r373|%p17, %r371, %r343, %r345, %r347;
mov.b32 %f68, %r373;
add.ftz.f32 %f69, %f67, %f68;
mov.b32 %r374, %f69;
shfl.sync.bfly.b32 %r376|%p18, %r374, %r358, %r345, %r347;
mov.b32 %f70, %r376;
add.ftz.f32 %f14, %f69, %f70;
add.s32 %r377, %r3, 1;
setp.ge.u32 %p19, %r377, %r23;
setp.ne.s32 %p20, %r2, 1;
or.pred %p21, %p20, %p19;
@%p21 bra $L__BB3_13;

st.global.f32 [%rd9+4], %f14;

$L__BB3_13:
ld.shared.f32 %f71, [%r17+256];
add.ftz.f32 %f72, %f71, %f106;
ld.shared.f32 %f73, [%r17+768];
add.ftz.f32 %f74, %f73, %f72;
ld.shared.f32 %f75, [%r17+1280];
add.ftz.f32 %f76, %f75, %f74;
mov.b32 %r378, %f76;
mov.u32 %r379, 31;
mov.u32 %r380, 16;
mov.u32 %r381, -1;
shfl.sync.bfly.b32 %r382|%p22, %r378, %r380, %r379, %r381;
mov.b32 %f77, %r382;
add.ftz.f32 %f78, %f76, %f77;
mov.b32 %r383, %f78;
mov.u32 %r384, 8;
shfl.sync.bfly.b32 %r385|%p23, %r383, %r384, %r379, %r381;
mov.b32 %f79, %r385;
add.ftz.f32 %f80, %f78, %f79;
mov.b32 %r386, %f80;
mov.u32 %r387, 4;
shfl.sync.bfly.b32 %r388|%p24, %r386, %r387, %r379, %r381;
mov.b32 %f81, %r388;
add.ftz.f32 %f82, %f80, %f81;
mov.b32 %r389, %f82;
mov.u32 %r390, 2;
shfl.sync.bfly.b32 %r391|%p25, %r389, %r390, %r379, %r381;
mov.b32 %f83, %r391;
add.ftz.f32 %f84, %f82, %f83;
mov.b32 %r392, %f84;
mov.u32 %r393, 1;
shfl.sync.bfly.b32 %r394|%p26, %r392, %r393, %r379, %r381;
mov.b32 %f85, %r394;
add.ftz.f32 %f15, %f84, %f85;
add.s32 %r395, %r3, 2;
setp.ge.u32 %p27, %r395, %r23;
setp.ne.s32 %p28, %r2, 2;
or.pred %p29, %p28, %p27;
@%p29 bra $L__BB3_15;

st.global.f32 [%rd9+8], %f15;

$L__BB3_15:
ld.shared.f32 %f86, [%r17+384];
add.ftz.f32 %f87, %f86, %f105;
ld.shared.f32 %f88, [%r17+896];
add.ftz.f32 %f89, %f88, %f87;
ld.shared.f32 %f90, [%r17+1408];
add.ftz.f32 %f91, %f90, %f89;
mov.b32 %r396, %f91;
shfl.sync.bfly.b32 %r400|%p30, %r396, %r380, %r379, %r381;
mov.b32 %f92, %r400;
add.ftz.f32 %f93, %f91, %f92;
mov.b32 %r401, %f93;
shfl.sync.bfly.b32 %r403|%p31, %r401, %r384, %r379, %r381;
mov.b32 %f94, %r403;
add.ftz.f32 %f95, %f93, %f94;
mov.b32 %r404, %f95;
shfl.sync.bfly.b32 %r406|%p32, %r404, %r387, %r379, %r381;
mov.b32 %f96, %r406;
add.ftz.f32 %f97, %f95, %f96;
mov.b32 %r407, %f97;
shfl.sync.bfly.b32 %r409|%p33, %r407, %r390, %r379, %r381;
mov.b32 %f98, %r409;
add.ftz.f32 %f99, %f97, %f98;
mov.b32 %r410, %f99;
shfl.sync.bfly.b32 %r412|%p34, %r410, %r393, %r379, %r381;
mov.b32 %f100, %r412;
add.ftz.f32 %f16, %f99, %f100;
add.s32 %r413, %r3, 3;
setp.ge.u32 %p35, %r413, %r23;
setp.ne.s32 %p36, %r2, 3;
or.pred %p37, %p36, %p35;
@%p37 bra $L__BB3_17;

st.global.f32 [%rd9+12], %f16;

$L__BB3_17:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<34>;
.reg .b16 %rs<5>;
.reg .f32 %f<120>;
.reg .b32 %r<295>;
.reg .b64 %rd<64>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[384];

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate[384];

ld.param.u64 %rd12, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd13, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd14, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd20, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
ld.param.f32 %f31, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+16];
ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+8];
ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3];
mov.u32 %r1, %tid.y;
mov.u32 %r2, %ctaid.y;
setp.eq.s64 %p2, %rd14, 0;
@%p2 bra $L__BB4_2;

cvta.to.global.u64 %rd21, %rd14;
mul.wide.u32 %rd22, %r2, 4;
add.s64 %rd23, %rd21, %rd22;
ld.global.nc.u32 %r291, [%rd23];
mul.hi.u32 %r44, %r2, %r28;
add.s32 %r45, %r44, %r2;
shr.u32 %r46, %r45, %r29;
mul.lo.s32 %r47, %r46, %r30;
sub.s32 %r292, %r2, %r47;
bra.uni $L__BB4_3;

$L__BB4_2:
mul.hi.u32 %r48, %r2, %r32;
add.s32 %r49, %r48, %r2;
shr.u32 %r291, %r49, %r33;
mov.u32 %r292, %r2;

$L__BB4_3:
mov.u32 %r50, %tid.x;
or.b32 %r51, %r50, %r1;
setp.ne.s32 %p3, %r51, 0;
mov.f32 %f113, 0f00000000;
mov.f32 %f109, %f113;
mov.f32 %f110, %f113;
@%p3 bra $L__BB4_8;

setp.eq.s64 %p4, %rd15, 0;
mov.f32 %f109, 0f00000000;
mov.f32 %f110, %f109;
@%p4 bra $L__BB4_6;

cvta.to.global.u64 %rd24, %rd15;
mov.u32 %r52, %ctaid.z;
mul.lo.s32 %r53, %r52, %r43;
cvt.u64.u32 %rd25, %r53;
selp.b32 %r54, %r2, %r291, %p2;
mul.lo.s32 %r55, %r54, %r37;
cvt.u64.u32 %rd26, %r55;
mov.u32 %r56, %ctaid.x;
cvt.s64.s32 %rd27, %r56;
cvt.u64.u32 %rd28, %r50;
add.s64 %rd29, %rd28, %rd27;
add.s64 %rd30, %rd29, %rd25;
add.s64 %rd31, %rd30, %rd26;
shl.b64 %rd32, %rd31, 2;
add.s64 %rd33, %rd24, %rd32;
ld.global.f32 %f110, [%rd33];

$L__BB4_6:
setp.eq.s64 %p6, %rd17, 0;
setp.eq.s64 %p7, %rd16, 0;
or.pred %p8, %p6, %p7;
@%p8 bra $L__BB4_8;

cvta.to.global.u64 %rd34, %rd17;
mov.u32 %r58, %ctaid.z;
mul.lo.s32 %r59, %r58, %r43;
cvt.u64.u32 %rd35, %r59;
selp.b32 %r60, %r2, %r291, %p2;
mul.lo.s32 %r61, %r60, %r37;
cvt.u64.u32 %rd36, %r61;
mov.u32 %r62, %ctaid.x;
cvt.s64.s32 %rd37, %r62;
cvt.u64.u32 %rd38, %r50;
add.s64 %rd39, %rd38, %rd37;
add.s64 %rd40, %rd39, %rd35;
add.s64 %rd41, %rd40, %rd36;
shl.b64 %rd42, %rd41, 2;
add.s64 %rd43, %rd34, %rd42;
ld.global.f32 %f109, [%rd43];

$L__BB4_8:
shl.b32 %r65, %r1, 5;
add.s32 %r66, %r65, %r50;
shr.s32 %r67, %r66, 31;
shr.u32 %r68, %r67, 30;
add.s32 %r69, %r66, %r68;
shr.s32 %r294, %r69, 2;
shr.u32 %r70, %r27, 7;
setp.ge.s32 %p10, %r294, %r70;
mov.f32 %f115, %f113;
@%p10 bra $L__BB4_13;

mov.u32 %r71, %ctaid.z;
mul.lo.s32 %r72, %r292, %r36;
cvt.u64.u32 %rd44, %r72;
and.b32 %r79, %r69, -4;
sub.s32 %r80, %r66, %r79;
shl.b32 %r81, %r80, 1;
cvt.s64.s32 %rd45, %r80;
mul.lo.s32 %r82, %r71, %r42;
cvt.u64.u32 %rd46, %r82;
add.s64 %rd47, %rd46, %rd45;
add.s64 %rd4, %rd47, %rd44;
shl.b32 %r293, %r294, 2;
mul.wide.s32 %rd48, %r81, 2;
or.b64 %rd5, %rd48, 2;
mul.hi.u32 %r83, %r71, %r38;
add.s32 %r84, %r83, %r71;
shr.u32 %r85, %r84, %r39;
mad.lo.s32 %r86, %r85, %r41, %r294;
mad.lo.s32 %r87, %r291, %r35, %r86;
mov.u32 %r88, %ctaid.x;
mad.lo.s32 %r89, %r88, %r31, %r87;
mul.wide.s32 %rd49, %r89, 18;
cvta.to.global.u64 %rd50, %rd16;
add.s64 %rd63, %rd50, %rd49;
cvta.to.global.u64 %rd51, %rd12;
add.s64 %rd62, %rd51, %rd49;
mov.f32 %f115, 0f00000000;
cvta.to.global.u64 %rd52, %rd13;
mov.f32 %f113, %f115;

$L__BB4_10:
ld.global.nc.u16 %rs1, [%rd62];

{ cvt.f32.f16 %f40, %rs1;}


cvt.u64.u32 %rd53, %r293;
add.s64 %rd54, %rd4, %rd53;
mul.lo.s64 %rd55, %rd54, 36;
add.s64 %rd56, %rd62, %rd5;
ld.global.nc.u16 %rs2, [%rd56];
cvt.s32.s16 %r123, %rs2;
add.s64 %rd57, %rd52, %rd55;
ld.global.nc.u32 %r12, [%rd57+4];
ld.global.nc.u32 %r13, [%rd57+8];
ld.global.nc.u32 %r14, [%rd57+12];
ld.global.nc.u32 %r15, [%rd57+16];
and.b32 %r124, %r123, 30583;
mov.u32 %r125, 286261504;
prmt.b32 %r126, %r125, %r125, %r124;
shr.u32 %r127, %r123, 2;
and.b32 %r128, %r127, 30583;
prmt.b32 %r129, %r125, %r125, %r128;
and.b32 %r130, %r126, 30583;
mov.u32 %r131, 511;
prmt.b32 %r132, %r131, %r131, %r130;
and.b32 %r133, %r129, 30583;
prmt.b32 %r134, %r131, %r131, %r133;
shr.u32 %r135, %r126, 16;
and.b32 %r136, %r135, 30583;
prmt.b32 %r137, %r131, %r131, %r136;
shr.u32 %r138, %r129, 16;
and.b32 %r139, %r138, 30583;
prmt.b32 %r140, %r131, %r131, %r139;
mov.u32 %r141, 21520;
prmt.b32 %r91, %r132, %r134, %r141;
mov.u32 %r142, 30258;
prmt.b32 %r95, %r132, %r134, %r142;
prmt.b32 %r99, %r137, %r140, %r141;
prmt.b32 %r103, %r137, %r140, %r142;
mov.u32 %r93, 0;

dp4a.s32.s32 %r90, %r91, %r12, %r93;


dp4a.s32.s32 %r94, %r95, %r13, %r90;


dp4a.s32.s32 %r98, %r99, %r14, %r94;


dp4a.s32.s32 %r102, %r103, %r15, %r98;

ld.global.nc.u16 %rs3, [%rd56+2];
cvt.s32.s16 %r143, %rs3;
ld.global.nc.u32 %r16, [%rd57+20];
ld.global.nc.u32 %r17, [%rd57+24];
ld.global.nc.u32 %r18, [%rd57+28];
ld.global.nc.u32 %r19, [%rd57+32];
and.b32 %r144, %r143, 30583;
prmt.b32 %r145, %r125, %r125, %r144;
shr.u32 %r146, %r143, 2;
and.b32 %r147, %r146, 30583;
prmt.b32 %r148, %r125, %r125, %r147;
and.b32 %r149, %r145, 30583;
prmt.b32 %r150, %r131, %r131, %r149;
and.b32 %r151, %r148, 30583;
prmt.b32 %r152, %r131, %r131, %r151;
shr.u32 %r153, %r145, 16;
and.b32 %r154, %r153, 30583;
prmt.b32 %r155, %r131, %r131, %r154;
shr.u32 %r156, %r148, 16;
and.b32 %r157, %r156, 30583;
prmt.b32 %r158, %r131, %r131, %r157;
prmt.b32 %r107, %r150, %r152, %r141;
prmt.b32 %r111, %r150, %r152, %r142;
prmt.b32 %r115, %r155, %r158, %r141;
prmt.b32 %r119, %r155, %r158, %r142;

dp4a.s32.s32 %r106, %r107, %r16, %r102;


dp4a.s32.s32 %r110, %r111, %r17, %r106;


dp4a.s32.s32 %r114, %r115, %r18, %r110;


dp4a.s32.s32 %r118, %r119, %r19, %r114;

ld.global.nc.u32 %r20, [%rd57];

{.reg .f16 low,high;
mov.b32 {low,high},%r20;
cvt.f32.f16 %f41, low;}


mul.ftz.f32 %f42, %f40, %f41;
cvt.rn.f32.s32 %f43, %r118;
fma.rn.ftz.f32 %f115, %f42, %f43, %f115;
setp.eq.s64 %p11, %rd16, 0;
@%p11 bra $L__BB4_12;

ld.global.u16 %rs4, [%rd63];

{ cvt.f32.f16 %f44, %rs4;}


add.s64 %rd58, %rd63, %rd5;
ld.global.s16 %r192, [%rd58];
and.b32 %r193, %r192, 30583;
prmt.b32 %r195, %r125, %r125, %r193;
shr.u32 %r196, %r192, 2;
and.b32 %r197, %r196, 30583;
prmt.b32 %r198, %r125, %r125, %r197;
and.b32 %r199, %r195, 30583;
prmt.b32 %r201, %r131, %r131, %r199;
and.b32 %r202, %r198, 30583;
prmt.b32 %r203, %r131, %r131, %r202;
shr.u32 %r204, %r195, 16;
and.b32 %r205, %r204, 30583;
prmt.b32 %r206, %r131, %r131, %r205;
shr.u32 %r207, %r198, 16;
and.b32 %r208, %r207, 30583;
prmt.b32 %r209, %r131, %r131, %r208;
prmt.b32 %r160, %r201, %r203, %r141;
prmt.b32 %r164, %r201, %r203, %r142;
prmt.b32 %r168, %r206, %r209, %r141;
prmt.b32 %r172, %r206, %r209, %r142;

dp4a.s32.s32 %r159, %r160, %r12, %r93;


dp4a.s32.s32 %r163, %r164, %r13, %r159;


dp4a.s32.s32 %r167, %r168, %r14, %r163;


dp4a.s32.s32 %r171, %r172, %r15, %r167;

ld.global.s16 %r212, [%rd58+2];
and.b32 %r213, %r212, 30583;
prmt.b32 %r214, %r125, %r125, %r213;
shr.u32 %r215, %r212, 2;
and.b32 %r216, %r215, 30583;
prmt.b32 %r217, %r125, %r125, %r216;
and.b32 %r218, %r214, 30583;
prmt.b32 %r219, %r131, %r131, %r218;
and.b32 %r220, %r217, 30583;
prmt.b32 %r221, %r131, %r131, %r220;
shr.u32 %r222, %r214, 16;
and.b32 %r223, %r222, 30583;
prmt.b32 %r224, %r131, %r131, %r223;
shr.u32 %r225, %r217, 16;
and.b32 %r226, %r225, 30583;
prmt.b32 %r227, %r131, %r131, %r226;
prmt.b32 %r176, %r219, %r221, %r141;
prmt.b32 %r180, %r219, %r221, %r142;
prmt.b32 %r184, %r224, %r227, %r141;
prmt.b32 %r188, %r224, %r227, %r142;

dp4a.s32.s32 %r175, %r176, %r16, %r171;


dp4a.s32.s32 %r179, %r180, %r17, %r175;


dp4a.s32.s32 %r183, %r184, %r18, %r179;


dp4a.s32.s32 %r187, %r188, %r19, %r183;


{.reg .f16 low,high;
mov.b32 {low,high},%r20;
cvt.f32.f16 %f45, low;}


mul.ftz.f32 %f46, %f44, %f45;
cvt.rn.f32.s32 %f47, %r187;
fma.rn.ftz.f32 %f113, %f46, %f47, %f113;

$L__BB4_12:
add.s32 %r293, %r293, 128;
add.s64 %rd63, %rd63, 576;
add.s64 %rd62, %rd62, 576;
add.s32 %r294, %r294, 32;
setp.lt.s32 %p12, %r294, %r70;
@%p12 bra $L__BB4_10;

$L__BB4_13:
setp.eq.s32 %p13, %r1, 0;
@%p13 bra $L__BB4_16;

setp.eq.s64 %p14, %rd16, 0;
add.s32 %r23, %r1, -1;
shl.b32 %r232, %r1, 7;
mov.u32 %r233, _ZZ13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r234, %r233, %r232;
shl.b32 %r235, %r50, 2;
add.s32 %r236, %r234, %r235;
st.shared.f32 [%r236+-128], %f115;
@%p14 bra $L__BB4_16;

shl.b32 %r238, %r23, 7;
mov.u32 %r239, _ZZ13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r240, %r239, %r238;
add.s32 %r242, %r240, %r235;
st.shared.f32 [%r242], %f113;

$L__BB4_16:
setp.ne.s32 %p15, %r1, 0;
bar.sync 0;
@%p15 bra $L__BB4_35;

setp.ne.s64 %p16, %rd16, 0;
shl.b32 %r245, %r50, 2;
mov.u32 %r246, _ZZ13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r24, %r246, %r245;
ld.shared.f32 %f48, [%r24];
add.ftz.f32 %f13, %f115, %f48;
@%p16 bra $L__BB4_19;
bra.uni $L__BB4_18;

$L__BB4_19:
mov.u32 %r249, _ZZ13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r250, %r249, %r245;
ld.shared.f32 %f52, [%r250];
add.ftz.f32 %f53, %f113, %f52;
ld.shared.f32 %f54, [%r24+128];
add.ftz.f32 %f55, %f13, %f54;
ld.shared.f32 %f56, [%r250+128];
add.ftz.f32 %f57, %f53, %f56;
ld.shared.f32 %f58, [%r24+256];
add.ftz.f32 %f117, %f55, %f58;
ld.shared.f32 %f59, [%r250+256];
add.ftz.f32 %f113, %f57, %f59;
bra.uni $L__BB4_20;

$L__BB4_18:
ld.shared.f32 %f49, [%r24+128];
add.ftz.f32 %f50, %f13, %f49;
ld.shared.f32 %f51, [%r24+256];
add.ftz.f32 %f117, %f50, %f51;

$L__BB4_20:
mov.b32 %r251, %f117;
mov.u32 %r252, 31;
mov.u32 %r253, 16;
mov.u32 %r254, -1;
shfl.sync.bfly.b32 %r255|%p17, %r251, %r253, %r252, %r254;
mov.b32 %f60, %r255;
add.ftz.f32 %f61, %f117, %f60;
mov.b32 %r256, %f61;
mov.u32 %r257, 8;
shfl.sync.bfly.b32 %r258|%p18, %r256, %r257, %r252, %r254;
mov.b32 %f62, %r258;
add.ftz.f32 %f63, %f61, %f62;
mov.b32 %r259, %f63;
mov.u32 %r260, 4;
shfl.sync.bfly.b32 %r261|%p19, %r259, %r260, %r252, %r254;
mov.b32 %f64, %r261;
add.ftz.f32 %f65, %f63, %f64;
mov.b32 %r262, %f65;
mov.u32 %r263, 2;
shfl.sync.bfly.b32 %r264|%p20, %r262, %r263, %r252, %r254;
mov.b32 %f66, %r264;
add.ftz.f32 %f19, %f65, %f66;
mov.b32 %r265, %f19;
mov.u32 %r266, 1;
shfl.sync.bfly.b32 %r25|%p1, %r265, %r266, %r252, %r254;
setp.eq.s64 %p21, %rd16, 0;
@%p21 bra $L__BB4_22;

mov.b32 %r267, %f113;
shfl.sync.bfly.b32 %r271|%p22, %r267, %r253, %r252, %r254;
mov.b32 %f67, %r271;
add.ftz.f32 %f68, %f113, %f67;
mov.b32 %r272, %f68;
shfl.sync.bfly.b32 %r274|%p23, %r272, %r257, %r252, %r254;
mov.b32 %f69, %r274;
add.ftz.f32 %f70, %f68, %f69;
mov.b32 %r275, %f70;
shfl.sync.bfly.b32 %r277|%p24, %r275, %r260, %r252, %r254;
mov.b32 %f71, %r277;
add.ftz.f32 %f72, %f70, %f71;
mov.b32 %r278, %f72;
shfl.sync.bfly.b32 %r280|%p25, %r278, %r263, %r252, %r254;
mov.b32 %f73, %r280;
add.ftz.f32 %f74, %f72, %f73;
mov.b32 %r281, %f74;
shfl.sync.bfly.b32 %r283|%p26, %r281, %r266, %r252, %r254;
mov.b32 %f75, %r283;
add.ftz.f32 %f113, %f74, %f75;

$L__BB4_22:
setp.ne.s32 %p27, %r50, 0;
mov.b32 %f76, %r25;
add.ftz.f32 %f22, %f19, %f76;
@%p27 bra $L__BB4_35;

add.ftz.f32 %f119, %f110, %f22;
@%p21 bra $L__BB4_34;

add.ftz.f32 %f24, %f109, %f113;
setp.gt.s32 %p29, %r26, 2;
@%p29 bra $L__BB4_28;

setp.eq.s32 %p32, %r26, 1;
@%p32 bra $L__BB4_32;

setp.eq.s32 %p33, %r26, 2;
@%p33 bra $L__BB4_27;
bra.uni $L__BB4_33;

$L__BB4_27:
mul.ftz.f32 %f96, %f24, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f97, %f96;
add.ftz.f32 %f98, %f97, 0f3F800000;
div.approx.ftz.f32 %f99, %f24, %f98;
mul.ftz.f32 %f119, %f119, %f99;
bra.uni $L__BB4_34;

$L__BB4_28:
setp.eq.s32 %p30, %r26, 3;
@%p30 bra $L__BB4_31;

setp.ne.s32 %p31, %r26, 6;
@%p31 bra $L__BB4_33;

min.ftz.f32 %f77, %f24, %f31;
neg.ftz.f32 %f78, %f31;
min.ftz.f32 %f79, %f119, %f31;
max.ftz.f32 %f80, %f79, %f78;
mul.ftz.f32 %f81, %f77, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f82, %f81;
add.ftz.f32 %f83, %f82, 0f3F800000;
div.approx.ftz.f32 %f84, %f77, %f83;
mul.ftz.f32 %f119, %f80, %f84;
bra.uni $L__BB4_34;

$L__BB4_32:
mul.ftz.f32 %f100, %f24, 0f3F000000;
mul.ftz.f32 %f101, %f24, 0f3D372713;
fma.rn.ftz.f32 %f102, %f24, %f101, 0f3F800000;
mul.ftz.f32 %f103, %f24, 0f3F4C422A;
mul.ftz.f32 %f104, %f103, %f102;
tanh.approx.f32 %f105, %f104;
add.ftz.f32 %f106, %f105, 0f3F800000;
mul.ftz.f32 %f107, %f100, %f106;
mul.ftz.f32 %f119, %f119, %f107;
bra.uni $L__BB4_34;

$L__BB4_31:
mov.f32 %f85, 0f40E00000;
min.ftz.f32 %f86, %f24, %f85;
min.ftz.f32 %f87, %f119, %f85;
mov.f32 %f88, 0fC0E00000;
max.ftz.f32 %f89, %f87, %f88;
mul.ftz.f32 %f90, %f86, 0f3FD9DB23;
mul.ftz.f32 %f91, %f90, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f92, %f91;
add.ftz.f32 %f93, %f92, 0f3F800000;
div.approx.ftz.f32 %f94, %f86, %f93;
add.ftz.f32 %f95, %f89, 0f3F800000;
mul.ftz.f32 %f119, %f95, %f94;
bra.uni $L__BB4_34;

$L__BB4_33:
mul.ftz.f32 %f119, %f119, %f24;

$L__BB4_34:
cvta.to.global.u64 %rd59, %rd20;
mov.u32 %r286, %ctaid.z;
mov.u32 %r287, %ctaid.y;
mov.u32 %r288, %ctaid.x;
mad.lo.s32 %r289, %r287, %r37, %r288;
mad.lo.s32 %r290, %r286, %r43, %r289;
mul.wide.u32 %rd60, %r290, 4;
add.s64 %rd61, %rd59, %rd60;
st.global.f32 [%rd61], %f119;

$L__BB4_35:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<14>;
.reg .b16 %rs<10>;
.reg .f32 %f<43>;
.reg .b32 %r<311>;
.reg .b64 %rd<43>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[384];

ld.param.u64 %rd11, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd12, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd9, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd10, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
cvta.to.global.u64 %rd1, %rd12;
cvta.to.global.u64 %rd2, %rd11;
mov.u32 %r1, %tid.y;
shl.b32 %r44, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r44, %r2;
mov.u32 %r4, %ctaid.x;
shr.u32 %r5, %r27, 7;
mov.u32 %r6, %ctaid.y;
setp.eq.s64 %p1, %rd9, 0;
@%p1 bra $L__BB5_2;

cvta.to.global.u64 %rd13, %rd9;
mul.wide.u32 %rd14, %r6, 4;
add.s64 %rd15, %rd13, %rd14;
ld.global.nc.u32 %r306, [%rd15];
mul.hi.u32 %r45, %r6, %r28;
add.s32 %r46, %r45, %r6;
shr.u32 %r47, %r46, %r29;
mul.lo.s32 %r48, %r47, %r30;
sub.s32 %r307, %r6, %r48;
bra.uni $L__BB5_3;

$L__BB5_2:
mul.hi.u32 %r49, %r6, %r32;
add.s32 %r50, %r49, %r6;
shr.u32 %r306, %r50, %r33;
mov.u32 %r307, %r6;

$L__BB5_3:
mov.u32 %r12, %ctaid.z;
mul.hi.u32 %r51, %r12, %r38;
add.s32 %r52, %r51, %r12;
shr.u32 %r53, %r52, %r39;
mul.lo.s32 %r13, %r53, %r41;
mul.lo.s32 %r14, %r306, %r35;
mul.lo.s32 %r15, %r4, %r31;
shr.s32 %r54, %r3, 31;
shr.u32 %r55, %r54, 30;
add.s32 %r56, %r3, %r55;
shr.s32 %r310, %r56, 2;
setp.ge.s32 %p2, %r310, %r5;
mov.f32 %f42, 0f00000000;
@%p2 bra $L__BB5_9;

mul.lo.s32 %r57, %r307, %r36;
cvt.u64.u32 %rd16, %r57;
and.b32 %r61, %r56, -4;
sub.s32 %r62, %r3, %r61;
cvt.s64.s32 %rd17, %r62;
mul.lo.s32 %r63, %r12, %r42;
cvt.u64.u32 %rd18, %r63;
add.s64 %rd19, %rd18, %rd17;
add.s64 %rd3, %rd19, %rd16;
shl.b32 %r64, %r62, 1;
cvt.s64.s32 %rd4, %r64;
not.b32 %r65, %r310;
add.s32 %r17, %r5, %r65;
and.b32 %r66, %r17, 32;
setp.ne.s32 %p3, %r66, 0;
mov.f32 %f42, 0f00000000;
@%p3 bra $L__BB5_6;

shl.b32 %r100, %r310, 2;
cvt.u64.u32 %rd20, %r100;
add.s32 %r101, %r14, %r15;
add.s32 %r102, %r101, %r13;
add.s32 %r103, %r102, %r310;
mul.wide.s32 %rd21, %r103, 18;
add.s64 %rd22, %rd2, %rd21;
ld.global.nc.u16 %rs1, [%rd22];

{ cvt.f32.f16 %f11, %rs1;}


shl.b64 %rd23, %rd4, 1;
add.s64 %rd24, %rd22, %rd23;
add.s64 %rd25, %rd3, %rd20;
mul.lo.s64 %rd26, %rd25, 36;
ld.global.nc.u16 %rs2, [%rd24+2];
cvt.s32.s16 %r104, %rs2;
add.s64 %rd27, %rd1, %rd26;
ld.global.nc.u32 %r69, [%rd27+4];
ld.global.nc.u32 %r73, [%rd27+8];
ld.global.nc.u32 %r77, [%rd27+12];
ld.global.nc.u32 %r81, [%rd27+16];
and.b32 %r105, %r104, 30583;
mov.u32 %r106, 286261504;
prmt.b32 %r107, %r106, %r106, %r105;
shr.u32 %r108, %r104, 2;
and.b32 %r109, %r108, 30583;
prmt.b32 %r110, %r106, %r106, %r109;
and.b32 %r111, %r107, 30583;
mov.u32 %r112, 511;
prmt.b32 %r113, %r112, %r112, %r111;
and.b32 %r114, %r110, 30583;
prmt.b32 %r115, %r112, %r112, %r114;
shr.u32 %r116, %r107, 16;
and.b32 %r117, %r116, 30583;
prmt.b32 %r118, %r112, %r112, %r117;
shr.u32 %r119, %r110, 16;
and.b32 %r120, %r119, 30583;
prmt.b32 %r121, %r112, %r112, %r120;
mov.u32 %r122, 21520;
prmt.b32 %r68, %r113, %r115, %r122;
mov.u32 %r123, 30258;
prmt.b32 %r72, %r113, %r115, %r123;
prmt.b32 %r76, %r118, %r121, %r122;
prmt.b32 %r80, %r118, %r121, %r123;
mov.u32 %r70, 0;

dp4a.s32.s32 %r67, %r68, %r69, %r70;


dp4a.s32.s32 %r71, %r72, %r73, %r67;


dp4a.s32.s32 %r75, %r76, %r77, %r71;


dp4a.s32.s32 %r79, %r80, %r81, %r75;

ld.global.nc.u16 %rs3, [%rd24+4];
cvt.s32.s16 %r124, %rs3;
ld.global.nc.u32 %r85, [%rd27+20];
ld.global.nc.u32 %r89, [%rd27+24];
ld.global.nc.u32 %r93, [%rd27+28];
ld.global.nc.u32 %r97, [%rd27+32];
and.b32 %r125, %r124, 30583;
prmt.b32 %r126, %r106, %r106, %r125;
shr.u32 %r127, %r124, 2;
and.b32 %r128, %r127, 30583;
prmt.b32 %r129, %r106, %r106, %r128;
and.b32 %r130, %r126, 30583;
prmt.b32 %r131, %r112, %r112, %r130;
and.b32 %r132, %r129, 30583;
prmt.b32 %r133, %r112, %r112, %r132;
shr.u32 %r134, %r126, 16;
and.b32 %r135, %r134, 30583;
prmt.b32 %r136, %r112, %r112, %r135;
shr.u32 %r137, %r129, 16;
and.b32 %r138, %r137, 30583;
prmt.b32 %r139, %r112, %r112, %r138;
prmt.b32 %r84, %r131, %r133, %r122;
prmt.b32 %r88, %r131, %r133, %r123;
prmt.b32 %r92, %r136, %r139, %r122;
prmt.b32 %r96, %r136, %r139, %r123;

dp4a.s32.s32 %r83, %r84, %r85, %r79;


dp4a.s32.s32 %r87, %r88, %r89, %r83;


dp4a.s32.s32 %r91, %r92, %r93, %r87;


dp4a.s32.s32 %r95, %r96, %r97, %r91;

ld.global.nc.u32 %r99, [%rd27];

{.reg .f16 low,high;
mov.b32 {low,high},%r99;
cvt.f32.f16 %f12, low;}


mul.ftz.f32 %f13, %f11, %f12;
cvt.rn.f32.s32 %f14, %r95;
fma.rn.ftz.f32 %f42, %f13, %f14, 0f00000000;
add.s32 %r310, %r310, 32;

$L__BB5_6:
and.b32 %r140, %r17, -32;
setp.eq.s32 %p4, %r140, 0;
@%p4 bra $L__BB5_9;

add.s32 %r141, %r310, %r13;
add.s32 %r142, %r141, %r14;
add.s32 %r143, %r142, %r15;
mul.wide.s32 %rd28, %r143, 18;
add.s64 %rd42, %rd2, %rd28;
shl.b32 %r144, %r310, 2;
add.s32 %r309, %r144, 128;
shl.b64 %rd29, %rd4, 1;
add.s64 %rd6, %rd29, 2;

$L__BB5_8:
add.s32 %r211, %r309, -128;
cvt.u64.u32 %rd30, %r211;
ld.global.nc.u16 %rs4, [%rd42];

{ cvt.f32.f16 %f15, %rs4;}


add.s64 %rd31, %rd3, %rd30;
mul.lo.s64 %rd32, %rd31, 36;
add.s64 %rd33, %rd42, %rd6;
ld.global.nc.u16 %rs6, [%rd33];
cvt.s32.s16 %r212, %rs6;
add.s64 %rd34, %rd1, %rd32;
ld.global.nc.u32 %r147, [%rd34+4];
ld.global.nc.u32 %r151, [%rd34+8];
ld.global.nc.u32 %r155, [%rd34+12];
ld.global.nc.u32 %r159, [%rd34+16];
and.b32 %r213, %r212, 30583;
mov.u32 %r214, 286261504;
prmt.b32 %r215, %r214, %r214, %r213;
shr.u32 %r216, %r212, 2;
and.b32 %r217, %r216, 30583;
prmt.b32 %r218, %r214, %r214, %r217;
and.b32 %r219, %r215, 30583;
mov.u32 %r220, 511;
prmt.b32 %r221, %r220, %r220, %r219;
and.b32 %r222, %r218, 30583;
prmt.b32 %r223, %r220, %r220, %r222;
shr.u32 %r224, %r215, 16;
and.b32 %r225, %r224, 30583;
prmt.b32 %r226, %r220, %r220, %r225;
shr.u32 %r227, %r218, 16;
and.b32 %r228, %r227, 30583;
prmt.b32 %r229, %r220, %r220, %r228;
mov.u32 %r230, 21520;
prmt.b32 %r146, %r221, %r223, %r230;
mov.u32 %r231, 30258;
prmt.b32 %r150, %r221, %r223, %r231;
prmt.b32 %r154, %r226, %r229, %r230;
prmt.b32 %r158, %r226, %r229, %r231;
mov.u32 %r181, 0;

dp4a.s32.s32 %r145, %r146, %r147, %r181;


dp4a.s32.s32 %r149, %r150, %r151, %r145;


dp4a.s32.s32 %r153, %r154, %r155, %r149;


dp4a.s32.s32 %r157, %r158, %r159, %r153;

ld.global.nc.u16 %rs7, [%rd33+2];
cvt.s32.s16 %r232, %rs7;
ld.global.nc.u32 %r163, [%rd34+20];
ld.global.nc.u32 %r167, [%rd34+24];
ld.global.nc.u32 %r171, [%rd34+28];
ld.global.nc.u32 %r175, [%rd34+32];
and.b32 %r233, %r232, 30583;
prmt.b32 %r234, %r214, %r214, %r233;
shr.u32 %r235, %r232, 2;
and.b32 %r236, %r235, 30583;
prmt.b32 %r237, %r214, %r214, %r236;
and.b32 %r238, %r234, 30583;
prmt.b32 %r239, %r220, %r220, %r238;
and.b32 %r240, %r237, 30583;
prmt.b32 %r241, %r220, %r220, %r240;
shr.u32 %r242, %r234, 16;
and.b32 %r243, %r242, 30583;
prmt.b32 %r244, %r220, %r220, %r243;
shr.u32 %r245, %r237, 16;
and.b32 %r246, %r245, 30583;
prmt.b32 %r247, %r220, %r220, %r246;
prmt.b32 %r162, %r239, %r241, %r230;
prmt.b32 %r166, %r239, %r241, %r231;
prmt.b32 %r170, %r244, %r247, %r230;
prmt.b32 %r174, %r244, %r247, %r231;

dp4a.s32.s32 %r161, %r162, %r163, %r157;


dp4a.s32.s32 %r165, %r166, %r167, %r161;


dp4a.s32.s32 %r169, %r170, %r171, %r165;


dp4a.s32.s32 %r173, %r174, %r175, %r169;

ld.global.nc.u32 %r177, [%rd34];

{.reg .f16 low,high;
mov.b32 {low,high},%r177;
cvt.f32.f16 %f16, low;}


mul.ftz.f32 %f19, %f15, %f16;
cvt.rn.f32.s32 %f20, %r173;
fma.rn.ftz.f32 %f21, %f19, %f20, %f42;
ld.global.nc.u16 %rs5, [%rd42+576];

{ cvt.f32.f16 %f17, %rs5;}


cvt.u64.u32 %rd35, %r309;
add.s64 %rd36, %rd3, %rd35;
mul.lo.s64 %rd37, %rd36, 36;
ld.global.nc.u16 %rs8, [%rd33+576];
cvt.s32.s16 %r248, %rs8;
add.s64 %rd38, %rd1, %rd37;
ld.global.nc.u32 %r180, [%rd38+4];
ld.global.nc.u32 %r184, [%rd38+8];
ld.global.nc.u32 %r188, [%rd38+12];
ld.global.nc.u32 %r192, [%rd38+16];
and.b32 %r249, %r248, 30583;
prmt.b32 %r250, %r214, %r214, %r249;
shr.u32 %r251, %r248, 2;
and.b32 %r252, %r251, 30583;
prmt.b32 %r253, %r214, %r214, %r252;
and.b32 %r254, %r250, 30583;
prmt.b32 %r255, %r220, %r220, %r254;
and.b32 %r256, %r253, 30583;
prmt.b32 %r257, %r220, %r220, %r256;
shr.u32 %r258, %r250, 16;
and.b32 %r259, %r258, 30583;
prmt.b32 %r260, %r220, %r220, %r259;
shr.u32 %r261, %r253, 16;
and.b32 %r262, %r261, 30583;
prmt.b32 %r263, %r220, %r220, %r262;
prmt.b32 %r179, %r255, %r257, %r230;
prmt.b32 %r183, %r255, %r257, %r231;
prmt.b32 %r187, %r260, %r263, %r230;
prmt.b32 %r191, %r260, %r263, %r231;

dp4a.s32.s32 %r178, %r179, %r180, %r181;


dp4a.s32.s32 %r182, %r183, %r184, %r178;


dp4a.s32.s32 %r186, %r187, %r188, %r182;


dp4a.s32.s32 %r190, %r191, %r192, %r186;

ld.global.nc.u16 %rs9, [%rd33+578];
cvt.s32.s16 %r264, %rs9;
ld.global.nc.u32 %r196, [%rd38+20];
ld.global.nc.u32 %r200, [%rd38+24];
ld.global.nc.u32 %r204, [%rd38+28];
ld.global.nc.u32 %r208, [%rd38+32];
and.b32 %r265, %r264, 30583;
prmt.b32 %r266, %r214, %r214, %r265;
shr.u32 %r267, %r264, 2;
and.b32 %r268, %r267, 30583;
prmt.b32 %r269, %r214, %r214, %r268;
and.b32 %r270, %r266, 30583;
prmt.b32 %r271, %r220, %r220, %r270;
and.b32 %r272, %r269, 30583;
prmt.b32 %r273, %r220, %r220, %r272;
shr.u32 %r274, %r266, 16;
and.b32 %r275, %r274, 30583;
prmt.b32 %r276, %r220, %r220, %r275;
shr.u32 %r277, %r269, 16;
and.b32 %r278, %r277, 30583;
prmt.b32 %r279, %r220, %r220, %r278;
prmt.b32 %r195, %r271, %r273, %r230;
prmt.b32 %r199, %r271, %r273, %r231;
prmt.b32 %r203, %r276, %r279, %r230;
prmt.b32 %r207, %r276, %r279, %r231;

dp4a.s32.s32 %r194, %r195, %r196, %r190;


dp4a.s32.s32 %r198, %r199, %r200, %r194;


dp4a.s32.s32 %r202, %r203, %r204, %r198;


dp4a.s32.s32 %r206, %r207, %r208, %r202;

ld.global.nc.u32 %r210, [%rd38];

{.reg .f16 low,high;
mov.b32 {low,high},%r210;
cvt.f32.f16 %f18, low;}


mul.ftz.f32 %f22, %f17, %f18;
cvt.rn.f32.s32 %f23, %r206;
fma.rn.ftz.f32 %f42, %f22, %f23, %f21;
add.s64 %rd42, %rd42, 1152;
add.s32 %r309, %r309, 256;
add.s32 %r310, %r310, 64;
setp.lt.s32 %p5, %r310, %r5;
@%p5 bra $L__BB5_8;

$L__BB5_9:
setp.eq.s32 %p6, %r1, 0;
@%p6 bra $L__BB5_11;

shl.b32 %r280, %r1, 7;
mov.u32 %r281, _ZZ13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r282, %r281, %r280;
shl.b32 %r283, %r2, 2;
add.s32 %r284, %r282, %r283;
st.shared.f32 [%r284+-128], %f42;

$L__BB5_11:
setp.ne.s32 %p7, %r1, 0;
bar.sync 0;
@%p7 bra $L__BB5_14;

mad.lo.s32 %r285, %r6, %r37, %r4;
mad.lo.s32 %r26, %r12, %r43, %r285;
shl.b32 %r286, %r2, 2;
mov.u32 %r287, _ZZ13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r288, %r287, %r286;
mov.u32 %r289, 2;
ld.shared.f32 %f24, [%r288];
add.ftz.f32 %f25, %f42, %f24;
ld.shared.f32 %f26, [%r288+128];
add.ftz.f32 %f27, %f25, %f26;
ld.shared.f32 %f28, [%r288+256];
add.ftz.f32 %f29, %f27, %f28;
mov.b32 %r290, %f29;
mov.u32 %r291, 31;
mov.u32 %r292, 16;
mov.u32 %r293, -1;
shfl.sync.bfly.b32 %r294|%p8, %r290, %r292, %r291, %r293;
mov.b32 %f30, %r294;
add.ftz.f32 %f31, %f29, %f30;
mov.b32 %r295, %f31;
mov.u32 %r296, 8;
shfl.sync.bfly.b32 %r297|%p9, %r295, %r296, %r291, %r293;
mov.b32 %f32, %r297;
add.ftz.f32 %f33, %f31, %f32;
mov.b32 %r298, %f33;
mov.u32 %r299, 4;
shfl.sync.bfly.b32 %r300|%p10, %r298, %r299, %r291, %r293;
mov.b32 %f34, %r300;
add.ftz.f32 %f35, %f33, %f34;
mov.b32 %r301, %f35;
shfl.sync.bfly.b32 %r302|%p11, %r301, %r289, %r291, %r293;
mov.b32 %f36, %r302;
add.ftz.f32 %f37, %f35, %f36;
mov.b32 %r303, %f37;
mov.u32 %r304, 1;
shfl.sync.bfly.b32 %r305|%p12, %r303, %r304, %r291, %r293;
mov.b32 %f38, %r305;
add.ftz.f32 %f7, %f37, %f38;
setp.ne.s32 %p13, %r2, 0;
@%p13 bra $L__BB5_14;

cvta.to.global.u64 %rd39, %rd10;
mul.wide.u32 %rd40, %r26, 4;
add.s64 %rd41, %rd39, %rd40;
st.global.f32 [%rd41], %f7;

$L__BB5_14:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type41ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type41ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type41ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type41ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type41ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type41ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type41ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type41ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type41ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<35>;
.reg .b16 %rs<9>;
.reg .f32 %f<109>;
.reg .b32 %r<363>;
.reg .b64 %rd<56>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type41ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type41ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type41ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type41ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r14, [_Z13mul_mat_vec_qIL9ggml_type41ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r15, [_Z13mul_mat_vec_qIL9ggml_type41ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r16, [_Z13mul_mat_vec_qIL9ggml_type41ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r17, [_Z13mul_mat_vec_qIL9ggml_type41ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type41ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type41ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type41ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type41ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type41ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type41ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type41ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type41ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r19, [_Z13mul_mat_vec_qIL9ggml_type41ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r18, [_Z13mul_mat_vec_qIL9ggml_type41ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r30, %r1, 5;
mov.u32 %r31, %tid.x;
add.s32 %r2, %r30, %r31;
mov.u32 %r32, %ctaid.x;
shl.b32 %r3, %r32, 1;
shr.s32 %r33, %r2, 31;
shr.u32 %r34, %r33, 30;
add.s32 %r35, %r2, %r34;
shr.s32 %r362, %r35, 2;
shr.u32 %r36, %r14, 7;
setp.ge.s32 %p3, %r362, %r36;
mov.f32 %f105, 0f00000000;
mov.f32 %f106, %f105;
mov.f32 %f107, %f105;
mov.f32 %f108, %f105;
@%p3 bra $L__BB6_3;

mov.u32 %r37, %ctaid.y;
mov.u32 %r38, %ctaid.z;
mul.hi.u32 %r39, %r38, %r24;
add.s32 %r40, %r39, %r38;
shr.u32 %r41, %r40, %r25;
mul.lo.s32 %r42, %r41, %r27;
mul.hi.u32 %r43, %r37, %r18;
add.s32 %r44, %r43, %r37;
shr.u32 %r45, %r44, %r19;
mul.lo.s32 %r46, %r45, %r21;
mul.lo.s32 %r47, %r37, %r22;
cvt.u64.u32 %rd20, %r47;
mul.lo.s32 %r48, %r38, %r28;
cvt.u64.u32 %rd21, %r48;
and.b32 %r52, %r35, -4;
sub.s32 %r53, %r2, %r52;
cvt.s64.s32 %rd22, %r53;
add.s64 %rd23, %rd20, %rd22;
add.s64 %rd1, %rd23, %rd21;
shl.b32 %r54, %r53, 1;
add.s32 %r55, %r3, 1;
shl.b32 %r360, %r362, 2;
add.s32 %r361, %r16, %r360;
mad.lo.s32 %r56, %r3, %r15, %r362;
add.s32 %r57, %r56, %r42;
add.s32 %r58, %r57, %r46;
mul.wide.s32 %rd24, %r58, 18;
mul.wide.s32 %rd25, %r54, 2;
add.s64 %rd26, %rd24, %rd25;
cvta.to.global.u64 %rd27, %rd17;
add.s64 %rd28, %rd27, %rd26;
add.s64 %rd55, %rd28, 4;
add.s32 %r59, %r362, %r42;
add.s32 %r60, %r59, %r46;
mad.lo.s32 %r61, %r55, %r15, %r60;
mul.wide.s32 %rd29, %r61, 18;
add.s64 %rd30, %rd29, %rd25;
add.s64 %rd31, %rd27, %rd30;
add.s64 %rd54, %rd31, 4;
add.s64 %rd53, %rd27, %rd29;
add.s64 %rd52, %rd27, %rd24;
cvta.to.global.u64 %rd6, %rd18;
mov.f32 %f105, 0f00000000;

$L__BB6_2:
cvt.u64.u32 %rd32, %r360;
add.s64 %rd33, %rd1, %rd32;
mul.lo.s64 %rd34, %rd33, 36;
ld.global.nc.u16 %rs1, [%rd52];

{ cvt.f32.f16 %f25, %rs1;}


ld.global.nc.u16 %rs5, [%rd55+-2];
cvt.s32.s16 %r194, %rs5;
add.s64 %rd35, %rd6, %rd34;
ld.global.nc.u32 %r64, [%rd35+4];
ld.global.nc.u32 %r68, [%rd35+8];
ld.global.nc.u32 %r72, [%rd35+12];
ld.global.nc.u32 %r76, [%rd35+16];
and.b32 %r195, %r194, 30583;
mov.u32 %r196, 286261504;
prmt.b32 %r197, %r196, %r196, %r195;
shr.u32 %r198, %r194, 2;
and.b32 %r199, %r198, 30583;
prmt.b32 %r200, %r196, %r196, %r199;
and.b32 %r201, %r197, 30583;
mov.u32 %r202, 511;
prmt.b32 %r203, %r202, %r202, %r201;
and.b32 %r204, %r200, 30583;
prmt.b32 %r205, %r202, %r202, %r204;
shr.u32 %r206, %r197, 16;
and.b32 %r207, %r206, 30583;
prmt.b32 %r208, %r202, %r202, %r207;
shr.u32 %r209, %r200, 16;
and.b32 %r210, %r209, 30583;
prmt.b32 %r211, %r202, %r202, %r210;
mov.u32 %r212, 21520;
prmt.b32 %r129, %r203, %r205, %r212;
mov.u32 %r213, 30258;
prmt.b32 %r133, %r203, %r205, %r213;
prmt.b32 %r137, %r208, %r211, %r212;
prmt.b32 %r141, %r208, %r211, %r213;
mov.u32 %r164, 0;

dp4a.s32.s32 %r62, %r129, %r64, %r164;


dp4a.s32.s32 %r66, %r133, %r68, %r62;


dp4a.s32.s32 %r70, %r137, %r72, %r66;


dp4a.s32.s32 %r74, %r141, %r76, %r70;

ld.global.nc.u16 %rs6, [%rd55];
cvt.s32.s16 %r214, %rs6;
ld.global.nc.u32 %r80, [%rd35+20];
ld.global.nc.u32 %r84, [%rd35+24];
ld.global.nc.u32 %r88, [%rd35+28];
ld.global.nc.u32 %r92, [%rd35+32];
and.b32 %r215, %r214, 30583;
prmt.b32 %r216, %r196, %r196, %r215;
shr.u32 %r217, %r214, 2;
and.b32 %r218, %r217, 30583;
prmt.b32 %r219, %r196, %r196, %r218;
and.b32 %r220, %r216, 30583;
prmt.b32 %r221, %r202, %r202, %r220;
and.b32 %r222, %r219, 30583;
prmt.b32 %r223, %r202, %r202, %r222;
shr.u32 %r224, %r216, 16;
and.b32 %r225, %r224, 30583;
prmt.b32 %r226, %r202, %r202, %r225;
shr.u32 %r227, %r219, 16;
and.b32 %r228, %r227, 30583;
prmt.b32 %r229, %r202, %r202, %r228;
prmt.b32 %r145, %r221, %r223, %r212;
prmt.b32 %r149, %r221, %r223, %r213;
prmt.b32 %r153, %r226, %r229, %r212;
prmt.b32 %r157, %r226, %r229, %r213;

dp4a.s32.s32 %r78, %r145, %r80, %r74;


dp4a.s32.s32 %r82, %r149, %r84, %r78;


dp4a.s32.s32 %r86, %r153, %r88, %r82;


dp4a.s32.s32 %r90, %r157, %r92, %r86;

ld.global.nc.u32 %r94, [%rd35];

{.reg .f16 low,high;
mov.b32 {low,high},%r94;
cvt.f32.f16 %f26, low;}


mul.ftz.f32 %f33, %f25, %f26;
cvt.rn.f32.s32 %f34, %r90;
fma.rn.ftz.f32 %f108, %f33, %f34, %f108;
ld.global.nc.u16 %rs2, [%rd53];

{ cvt.f32.f16 %f27, %rs2;}


ld.global.nc.u16 %rs7, [%rd54+-2];
cvt.s32.s16 %r230, %rs7;
and.b32 %r231, %r230, 30583;
prmt.b32 %r232, %r196, %r196, %r231;
shr.u32 %r233, %r230, 2;
and.b32 %r234, %r233, 30583;
prmt.b32 %r235, %r196, %r196, %r234;
and.b32 %r236, %r232, 30583;
prmt.b32 %r237, %r202, %r202, %r236;
and.b32 %r238, %r235, 30583;
prmt.b32 %r239, %r202, %r202, %r238;
shr.u32 %r240, %r232, 16;
and.b32 %r241, %r240, 30583;
prmt.b32 %r242, %r202, %r202, %r241;
shr.u32 %r243, %r235, 16;
and.b32 %r244, %r243, 30583;
prmt.b32 %r245, %r202, %r202, %r244;
prmt.b32 %r162, %r237, %r239, %r212;
prmt.b32 %r166, %r237, %r239, %r213;
prmt.b32 %r170, %r242, %r245, %r212;
prmt.b32 %r174, %r242, %r245, %r213;

dp4a.s32.s32 %r95, %r162, %r64, %r164;


dp4a.s32.s32 %r99, %r166, %r68, %r95;


dp4a.s32.s32 %r103, %r170, %r72, %r99;


dp4a.s32.s32 %r107, %r174, %r76, %r103;

ld.global.nc.u16 %rs8, [%rd54];
cvt.s32.s16 %r246, %rs8;
and.b32 %r247, %r246, 30583;
prmt.b32 %r248, %r196, %r196, %r247;
shr.u32 %r249, %r246, 2;
and.b32 %r250, %r249, 30583;
prmt.b32 %r251, %r196, %r196, %r250;
and.b32 %r252, %r248, 30583;
prmt.b32 %r253, %r202, %r202, %r252;
and.b32 %r254, %r251, 30583;
prmt.b32 %r255, %r202, %r202, %r254;
shr.u32 %r256, %r248, 16;
and.b32 %r257, %r256, 30583;
prmt.b32 %r258, %r202, %r202, %r257;
shr.u32 %r259, %r251, 16;
and.b32 %r260, %r259, 30583;
prmt.b32 %r261, %r202, %r202, %r260;
prmt.b32 %r178, %r253, %r255, %r212;
prmt.b32 %r182, %r253, %r255, %r213;
prmt.b32 %r186, %r258, %r261, %r212;
prmt.b32 %r190, %r258, %r261, %r213;

dp4a.s32.s32 %r111, %r178, %r80, %r107;


dp4a.s32.s32 %r115, %r182, %r84, %r111;


dp4a.s32.s32 %r119, %r186, %r88, %r115;


dp4a.s32.s32 %r123, %r190, %r92, %r119;


{.reg .f16 low,high;
mov.b32 {low,high},%r94;
cvt.f32.f16 %f28, low;}


mul.ftz.f32 %f35, %f27, %f28;
cvt.rn.f32.s32 %f36, %r123;
fma.rn.ftz.f32 %f107, %f35, %f36, %f107;
cvt.u64.u32 %rd36, %r361;
add.s64 %rd37, %rd1, %rd36;
mul.lo.s64 %rd38, %rd37, 36;

{ cvt.f32.f16 %f29, %rs1;}


add.s64 %rd39, %rd6, %rd38;
ld.global.nc.u32 %r130, [%rd39+4];
ld.global.nc.u32 %r134, [%rd39+8];
ld.global.nc.u32 %r138, [%rd39+12];
ld.global.nc.u32 %r142, [%rd39+16];

dp4a.s32.s32 %r128, %r129, %r130, %r164;


dp4a.s32.s32 %r132, %r133, %r134, %r128;


dp4a.s32.s32 %r136, %r137, %r138, %r132;


dp4a.s32.s32 %r140, %r141, %r142, %r136;

ld.global.nc.u32 %r146, [%rd39+20];
ld.global.nc.u32 %r150, [%rd39+24];
ld.global.nc.u32 %r154, [%rd39+28];
ld.global.nc.u32 %r158, [%rd39+32];

dp4a.s32.s32 %r144, %r145, %r146, %r140;


dp4a.s32.s32 %r148, %r149, %r150, %r144;


dp4a.s32.s32 %r152, %r153, %r154, %r148;


dp4a.s32.s32 %r156, %r157, %r158, %r152;

ld.global.nc.u32 %r160, [%rd39];

{.reg .f16 low,high;
mov.b32 {low,high},%r160;
cvt.f32.f16 %f30, low;}


mul.ftz.f32 %f37, %f29, %f30;
cvt.rn.f32.s32 %f38, %r156;
fma.rn.ftz.f32 %f106, %f37, %f38, %f106;

{ cvt.f32.f16 %f31, %rs2;}



dp4a.s32.s32 %r161, %r162, %r130, %r164;


dp4a.s32.s32 %r165, %r166, %r134, %r161;


dp4a.s32.s32 %r169, %r170, %r138, %r165;


dp4a.s32.s32 %r173, %r174, %r142, %r169;


dp4a.s32.s32 %r177, %r178, %r146, %r173;


dp4a.s32.s32 %r181, %r182, %r150, %r177;


dp4a.s32.s32 %r185, %r186, %r154, %r181;


dp4a.s32.s32 %r189, %r190, %r158, %r185;


{.reg .f16 low,high;
mov.b32 {low,high},%r160;
cvt.f32.f16 %f32, low;}


mul.ftz.f32 %f39, %f31, %f32;
cvt.rn.f32.s32 %f40, %r189;
fma.rn.ftz.f32 %f105, %f39, %f40, %f105;
add.s32 %r361, %r361, 128;
add.s32 %r360, %r360, 128;
add.s64 %rd55, %rd55, 576;
add.s64 %rd54, %rd54, 576;
add.s64 %rd53, %rd53, 576;
add.s64 %rd52, %rd52, 576;
add.s32 %r362, %r362, 32;
setp.lt.s32 %p4, %r362, %r36;
@%p4 bra $L__BB6_2;

$L__BB6_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB6_5;

shl.b32 %r263, %r1, 9;
mov.u32 %r264, _ZZ13mul_mat_vec_qIL9ggml_type41ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r265, %r264, %r263;
shl.b32 %r266, %r31, 2;
add.s32 %r267, %r265, %r266;
st.shared.f32 [%r267+-512], %f108;
st.shared.f32 [%r267+-384], %f107;
st.shared.f32 [%r267+-256], %f106;
st.shared.f32 [%r267+-128], %f105;

$L__BB6_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB6_14;

mov.u32 %r268, %ctaid.z;
mov.u32 %r269, %ctaid.y;
mad.lo.s32 %r270, %r269, %r23, %r3;
mad.lo.s32 %r271, %r268, %r29, %r270;
cvt.u64.u32 %rd15, %r271;
shl.b32 %r273, %r31, 2;
mov.u32 %r274, _ZZ13mul_mat_vec_qIL9ggml_type41ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r275, %r274, %r273;
mov.u32 %r276, 2;
ld.shared.f32 %f41, [%r275];
add.ftz.f32 %f42, %f41, %f108;
ld.shared.f32 %f43, [%r275+512];
add.ftz.f32 %f44, %f43, %f42;
ld.shared.f32 %f45, [%r275+1024];
add.ftz.f32 %f46, %f45, %f44;
mov.b32 %r277, %f46;
mov.u32 %r278, 31;
mov.u32 %r279, 16;
mov.u32 %r280, -1;
shfl.sync.bfly.b32 %r281|%p7, %r277, %r279, %r278, %r280;
mov.b32 %f47, %r281;
add.ftz.f32 %f48, %f46, %f47;
mov.b32 %r282, %f48;
mov.u32 %r283, 8;
shfl.sync.bfly.b32 %r284|%p8, %r282, %r283, %r278, %r280;
mov.b32 %f49, %r284;
add.ftz.f32 %f50, %f48, %f49;
mov.b32 %r285, %f50;
mov.u32 %r286, 4;
shfl.sync.bfly.b32 %r287|%p9, %r285, %r286, %r278, %r280;
mov.b32 %f51, %r287;
add.ftz.f32 %f52, %f50, %f51;
mov.b32 %r288, %f52;
shfl.sync.bfly.b32 %r289|%p10, %r288, %r276, %r278, %r280;
mov.b32 %f53, %r289;
add.ftz.f32 %f54, %f52, %f53;
mov.b32 %r290, %f54;
mov.u32 %r291, 1;
shfl.sync.bfly.b32 %r292|%p11, %r290, %r291, %r278, %r280;
mov.b32 %f55, %r292;
add.ftz.f32 %f13, %f54, %f55;
setp.eq.s32 %p12, %r31, 0;
setp.lt.u32 %p13, %r3, %r17;
and.pred %p1, %p12, %p13;
cvta.to.global.u64 %rd40, %rd19;
mul.wide.u32 %rd41, %r271, 4;
add.s64 %rd16, %rd40, %rd41;
not.pred %p14, %p1;
@%p14 bra $L__BB6_8;

st.global.f32 [%rd16], %f13;

$L__BB6_8:
ld.shared.f32 %f56, [%r275+128];
add.ftz.f32 %f57, %f56, %f107;
ld.shared.f32 %f58, [%r275+640];
add.ftz.f32 %f59, %f58, %f57;
ld.shared.f32 %f60, [%r275+1152];
add.ftz.f32 %f61, %f60, %f59;
mov.b32 %r298, %f61;
shfl.sync.bfly.b32 %r302|%p15, %r298, %r279, %r278, %r280;
mov.b32 %f62, %r302;
add.ftz.f32 %f63, %f61, %f62;
mov.b32 %r303, %f63;
shfl.sync.bfly.b32 %r305|%p16, %r303, %r283, %r278, %r280;
mov.b32 %f64, %r305;
add.ftz.f32 %f65, %f63, %f64;
mov.b32 %r306, %f65;
shfl.sync.bfly.b32 %r308|%p17, %r306, %r286, %r278, %r280;
mov.b32 %f66, %r308;
add.ftz.f32 %f67, %f65, %f66;
mov.b32 %r309, %f67;
shfl.sync.bfly.b32 %r310|%p18, %r309, %r276, %r278, %r280;
mov.b32 %f68, %r310;
add.ftz.f32 %f69, %f67, %f68;
mov.b32 %r311, %f69;
shfl.sync.bfly.b32 %r313|%p19, %r311, %r291, %r278, %r280;
mov.b32 %f70, %r313;
add.ftz.f32 %f14, %f69, %f70;
setp.eq.s32 %p20, %r31, 1;
or.b32 %r316, %r3, 1;
setp.lt.u32 %p21, %r316, %r17;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB6_10;

st.global.f32 [%rd16+4], %f14;

$L__BB6_10:
mov.u32 %r321, 2;
ld.shared.f32 %f71, [%r275+256];
add.ftz.f32 %f72, %f71, %f106;
ld.shared.f32 %f73, [%r275+768];
add.ftz.f32 %f74, %f73, %f72;
ld.shared.f32 %f75, [%r275+1280];
add.ftz.f32 %f76, %f75, %f74;
mov.b32 %r322, %f76;
mov.u32 %r323, 31;
mov.u32 %r324, 16;
mov.u32 %r325, -1;
shfl.sync.bfly.b32 %r326|%p23, %r322, %r324, %r323, %r325;
mov.b32 %f77, %r326;
add.ftz.f32 %f78, %f76, %f77;
mov.b32 %r327, %f78;
mov.u32 %r328, 8;
shfl.sync.bfly.b32 %r329|%p24, %r327, %r328, %r323, %r325;
mov.b32 %f79, %r329;
add.ftz.f32 %f80, %f78, %f79;
mov.b32 %r330, %f80;
mov.u32 %r331, 4;
shfl.sync.bfly.b32 %r332|%p25, %r330, %r331, %r323, %r325;
mov.b32 %f81, %r332;
add.ftz.f32 %f82, %f80, %f81;
mov.b32 %r333, %f82;
shfl.sync.bfly.b32 %r334|%p26, %r333, %r321, %r323, %r325;
mov.b32 %f83, %r334;
add.ftz.f32 %f84, %f82, %f83;
mov.b32 %r335, %f84;
mov.u32 %r336, 1;
shfl.sync.bfly.b32 %r337|%p27, %r335, %r336, %r323, %r325;
mov.b32 %f85, %r337;
add.ftz.f32 %f15, %f84, %f85;
@%p14 bra $L__BB6_12;

cvt.u64.u32 %rd42, %r17;
add.s64 %rd43, %rd42, %rd15;
shl.b64 %rd45, %rd43, 2;
add.s64 %rd46, %rd40, %rd45;
st.global.f32 [%rd46], %f15;

$L__BB6_12:
ld.shared.f32 %f86, [%r275+384];
add.ftz.f32 %f87, %f86, %f105;
ld.shared.f32 %f88, [%r275+896];
add.ftz.f32 %f89, %f88, %f87;
ld.shared.f32 %f90, [%r275+1408];
add.ftz.f32 %f91, %f90, %f89;
mov.b32 %r343, %f91;
shfl.sync.bfly.b32 %r347|%p29, %r343, %r324, %r323, %r325;
mov.b32 %f92, %r347;
add.ftz.f32 %f93, %f91, %f92;
mov.b32 %r348, %f93;
shfl.sync.bfly.b32 %r350|%p30, %r348, %r328, %r323, %r325;
mov.b32 %f94, %r350;
add.ftz.f32 %f95, %f93, %f94;
mov.b32 %r351, %f95;
shfl.sync.bfly.b32 %r353|%p31, %r351, %r331, %r323, %r325;
mov.b32 %f96, %r353;
add.ftz.f32 %f97, %f95, %f96;
mov.b32 %r354, %f97;
shfl.sync.bfly.b32 %r355|%p32, %r354, %r321, %r323, %r325;
mov.b32 %f98, %r355;
add.ftz.f32 %f99, %f97, %f98;
mov.b32 %r356, %f99;
shfl.sync.bfly.b32 %r358|%p33, %r356, %r336, %r323, %r325;
mov.b32 %f100, %r358;
add.ftz.f32 %f16, %f99, %f100;
@%p22 bra $L__BB6_14;

add.s32 %r359, %r17, 1;
cvt.u64.u32 %rd47, %r359;
add.s64 %rd48, %rd47, %rd15;
shl.b64 %rd50, %rd48, 2;
add.s64 %rd51, %rd40, %rd50;
st.global.f32 [%rd51], %f16;

$L__BB6_14:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type41ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type41ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type41ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type41ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type41ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type41ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type41ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type41ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type41ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<50>;
.reg .b16 %rs<11>;
.reg .f32 %f<163>;
.reg .b32 %r<498>;
.reg .b64 %rd<68>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type41ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[2304];

ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type41ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type41ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type41ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r15, [_Z13mul_mat_vec_qIL9ggml_type41ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r16, [_Z13mul_mat_vec_qIL9ggml_type41ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r17, [_Z13mul_mat_vec_qIL9ggml_type41ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r18, [_Z13mul_mat_vec_qIL9ggml_type41ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type41ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type41ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type41ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type41ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type41ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type41ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type41ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type41ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type41ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r19, [_Z13mul_mat_vec_qIL9ggml_type41ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r31, %r1, 5;
mov.u32 %r32, %tid.x;
add.s32 %r33, %r31, %r32;
shr.s32 %r34, %r33, 31;
shr.u32 %r35, %r34, 30;
add.s32 %r36, %r33, %r35;
shr.s32 %r497, %r36, 2;
shr.u32 %r37, %r15, 7;
setp.ge.s32 %p2, %r497, %r37;
mov.f32 %f157, 0f00000000;
mov.f32 %f158, %f157;
mov.f32 %f159, %f157;
mov.f32 %f160, %f157;
mov.f32 %f161, %f157;
mov.f32 %f162, %f157;
@%p2 bra $L__BB7_3;

mov.u32 %r38, %ctaid.x;
shl.b32 %r39, %r38, 1;
mov.u32 %r40, %ctaid.y;
mov.u32 %r41, %ctaid.z;
mul.hi.u32 %r42, %r41, %r25;
add.s32 %r43, %r42, %r41;
shr.u32 %r44, %r43, %r26;
mul.lo.s32 %r45, %r44, %r28;
mul.hi.u32 %r46, %r40, %r19;
add.s32 %r47, %r46, %r40;
shr.u32 %r48, %r47, %r20;
mul.lo.s32 %r49, %r48, %r22;
mul.lo.s32 %r50, %r40, %r23;
cvt.u64.u32 %rd20, %r50;
mul.lo.s32 %r51, %r41, %r29;
cvt.u64.u32 %rd21, %r51;
and.b32 %r59, %r36, -4;
sub.s32 %r60, %r33, %r59;
cvt.s64.s32 %rd22, %r60;
add.s64 %rd23, %rd20, %rd22;
add.s64 %rd1, %rd23, %rd21;
shl.b32 %r61, %r60, 1;
or.b32 %r62, %r39, 1;
shl.b32 %r63, %r17, 1;
shl.b32 %r494, %r497, 2;
add.s32 %r496, %r494, %r63;
add.s32 %r495, %r17, %r494;
mad.lo.s32 %r64, %r39, %r16, %r497;
add.s32 %r65, %r64, %r45;
add.s32 %r66, %r65, %r49;
mul.wide.s32 %rd24, %r66, 18;
mul.wide.s32 %rd25, %r61, 2;
add.s64 %rd26, %rd24, %rd25;
cvta.to.global.u64 %rd27, %rd17;
add.s64 %rd28, %rd27, %rd26;
add.s64 %rd67, %rd28, 4;
add.s32 %r67, %r497, %r45;
add.s32 %r68, %r67, %r49;
mad.lo.s32 %r69, %r62, %r16, %r68;
mul.wide.s32 %rd29, %r69, 18;
add.s64 %rd30, %rd29, %rd25;
add.s64 %rd31, %rd27, %rd30;
add.s64 %rd66, %rd31, 4;
add.s64 %rd65, %rd27, %rd29;
add.s64 %rd64, %rd27, %rd24;
cvta.to.global.u64 %rd6, %rd18;
mov.f32 %f157, 0f00000000;

$L__BB7_2:
cvt.u64.u32 %rd32, %r494;
add.s64 %rd33, %rd1, %rd32;
mul.lo.s64 %rd34, %rd33, 36;
ld.global.nc.u16 %rs1, [%rd64];

{ cvt.f32.f16 %f37, %rs1;}


ld.global.nc.u16 %rs7, [%rd67+-2];
cvt.s32.s16 %r268, %rs7;
add.s64 %rd35, %rd6, %rd34;
ld.global.nc.u32 %r72, [%rd35+4];
ld.global.nc.u32 %r76, [%rd35+8];
ld.global.nc.u32 %r80, [%rd35+12];
ld.global.nc.u32 %r84, [%rd35+16];
and.b32 %r269, %r268, 30583;
mov.u32 %r270, 286261504;
prmt.b32 %r271, %r270, %r270, %r269;
shr.u32 %r272, %r268, 2;
and.b32 %r273, %r272, 30583;
prmt.b32 %r274, %r270, %r270, %r273;
and.b32 %r275, %r271, 30583;
mov.u32 %r276, 511;
prmt.b32 %r277, %r276, %r276, %r275;
and.b32 %r278, %r274, 30583;
prmt.b32 %r279, %r276, %r276, %r278;
shr.u32 %r280, %r271, 16;
and.b32 %r281, %r280, 30583;
prmt.b32 %r282, %r276, %r276, %r281;
shr.u32 %r283, %r274, 16;
and.b32 %r284, %r283, 30583;
prmt.b32 %r285, %r276, %r276, %r284;
mov.u32 %r286, 21520;
prmt.b32 %r203, %r277, %r279, %r286;
mov.u32 %r287, 30258;
prmt.b32 %r207, %r277, %r279, %r287;
prmt.b32 %r211, %r282, %r285, %r286;
prmt.b32 %r215, %r282, %r285, %r287;
mov.u32 %r238, 0;

dp4a.s32.s32 %r70, %r203, %r72, %r238;


dp4a.s32.s32 %r74, %r207, %r76, %r70;


dp4a.s32.s32 %r78, %r211, %r80, %r74;


dp4a.s32.s32 %r82, %r215, %r84, %r78;

ld.global.nc.u16 %rs8, [%rd67];
cvt.s32.s16 %r288, %rs8;
ld.global.nc.u32 %r88, [%rd35+20];
ld.global.nc.u32 %r92, [%rd35+24];
ld.global.nc.u32 %r96, [%rd35+28];
ld.global.nc.u32 %r100, [%rd35+32];
and.b32 %r289, %r288, 30583;
prmt.b32 %r290, %r270, %r270, %r289;
shr.u32 %r291, %r288, 2;
and.b32 %r292, %r291, 30583;
prmt.b32 %r293, %r270, %r270, %r292;
and.b32 %r294, %r290, 30583;
prmt.b32 %r295, %r276, %r276, %r294;
and.b32 %r296, %r293, 30583;
prmt.b32 %r297, %r276, %r276, %r296;
shr.u32 %r298, %r290, 16;
and.b32 %r299, %r298, 30583;
prmt.b32 %r300, %r276, %r276, %r299;
shr.u32 %r301, %r293, 16;
and.b32 %r302, %r301, 30583;
prmt.b32 %r303, %r276, %r276, %r302;
prmt.b32 %r219, %r295, %r297, %r286;
prmt.b32 %r223, %r295, %r297, %r287;
prmt.b32 %r227, %r300, %r303, %r286;
prmt.b32 %r231, %r300, %r303, %r287;

dp4a.s32.s32 %r86, %r219, %r88, %r82;


dp4a.s32.s32 %r90, %r223, %r92, %r86;


dp4a.s32.s32 %r94, %r227, %r96, %r90;


dp4a.s32.s32 %r98, %r231, %r100, %r94;

ld.global.nc.u32 %r102, [%rd35];

{.reg .f16 low,high;
mov.b32 {low,high},%r102;
cvt.f32.f16 %f38, low;}


mul.ftz.f32 %f49, %f37, %f38;
cvt.rn.f32.s32 %f50, %r98;
fma.rn.ftz.f32 %f162, %f49, %f50, %f162;
ld.global.nc.u16 %rs2, [%rd65];

{ cvt.f32.f16 %f39, %rs2;}


ld.global.nc.u16 %rs9, [%rd66+-2];
cvt.s32.s16 %r304, %rs9;
and.b32 %r305, %r304, 30583;
prmt.b32 %r306, %r270, %r270, %r305;
shr.u32 %r307, %r304, 2;
and.b32 %r308, %r307, 30583;
prmt.b32 %r309, %r270, %r270, %r308;
and.b32 %r310, %r306, 30583;
prmt.b32 %r311, %r276, %r276, %r310;
and.b32 %r312, %r309, 30583;
prmt.b32 %r313, %r276, %r276, %r312;
shr.u32 %r314, %r306, 16;
and.b32 %r315, %r314, 30583;
prmt.b32 %r316, %r276, %r276, %r315;
shr.u32 %r317, %r309, 16;
and.b32 %r318, %r317, 30583;
prmt.b32 %r319, %r276, %r276, %r318;
prmt.b32 %r236, %r311, %r313, %r286;
prmt.b32 %r240, %r311, %r313, %r287;
prmt.b32 %r244, %r316, %r319, %r286;
prmt.b32 %r248, %r316, %r319, %r287;

dp4a.s32.s32 %r103, %r236, %r72, %r238;


dp4a.s32.s32 %r107, %r240, %r76, %r103;


dp4a.s32.s32 %r111, %r244, %r80, %r107;


dp4a.s32.s32 %r115, %r248, %r84, %r111;

ld.global.nc.u16 %rs10, [%rd66];
cvt.s32.s16 %r320, %rs10;
and.b32 %r321, %r320, 30583;
prmt.b32 %r322, %r270, %r270, %r321;
shr.u32 %r323, %r320, 2;
and.b32 %r324, %r323, 30583;
prmt.b32 %r325, %r270, %r270, %r324;
and.b32 %r326, %r322, 30583;
prmt.b32 %r327, %r276, %r276, %r326;
and.b32 %r328, %r325, 30583;
prmt.b32 %r329, %r276, %r276, %r328;
shr.u32 %r330, %r322, 16;
and.b32 %r331, %r330, 30583;
prmt.b32 %r332, %r276, %r276, %r331;
shr.u32 %r333, %r325, 16;
and.b32 %r334, %r333, 30583;
prmt.b32 %r335, %r276, %r276, %r334;
prmt.b32 %r252, %r327, %r329, %r286;
prmt.b32 %r256, %r327, %r329, %r287;
prmt.b32 %r260, %r332, %r335, %r286;
prmt.b32 %r264, %r332, %r335, %r287;

dp4a.s32.s32 %r119, %r252, %r88, %r115;


dp4a.s32.s32 %r123, %r256, %r92, %r119;


dp4a.s32.s32 %r127, %r260, %r96, %r123;


dp4a.s32.s32 %r131, %r264, %r100, %r127;


{.reg .f16 low,high;
mov.b32 {low,high},%r102;
cvt.f32.f16 %f40, low;}


mul.ftz.f32 %f51, %f39, %f40;
cvt.rn.f32.s32 %f52, %r131;
fma.rn.ftz.f32 %f161, %f51, %f52, %f161;
cvt.u64.u32 %rd36, %r495;
add.s64 %rd37, %rd1, %rd36;
mul.lo.s64 %rd38, %rd37, 36;

{ cvt.f32.f16 %f41, %rs1;}


add.s64 %rd39, %rd6, %rd38;
ld.global.nc.u32 %r138, [%rd39+4];
ld.global.nc.u32 %r142, [%rd39+8];
ld.global.nc.u32 %r146, [%rd39+12];
ld.global.nc.u32 %r150, [%rd39+16];

dp4a.s32.s32 %r136, %r203, %r138, %r238;


dp4a.s32.s32 %r140, %r207, %r142, %r136;


dp4a.s32.s32 %r144, %r211, %r146, %r140;


dp4a.s32.s32 %r148, %r215, %r150, %r144;

ld.global.nc.u32 %r154, [%rd39+20];
ld.global.nc.u32 %r158, [%rd39+24];
ld.global.nc.u32 %r162, [%rd39+28];
ld.global.nc.u32 %r166, [%rd39+32];

dp4a.s32.s32 %r152, %r219, %r154, %r148;


dp4a.s32.s32 %r156, %r223, %r158, %r152;


dp4a.s32.s32 %r160, %r227, %r162, %r156;


dp4a.s32.s32 %r164, %r231, %r166, %r160;

ld.global.nc.u32 %r168, [%rd39];

{.reg .f16 low,high;
mov.b32 {low,high},%r168;
cvt.f32.f16 %f42, low;}


mul.ftz.f32 %f53, %f41, %f42;
cvt.rn.f32.s32 %f54, %r164;
fma.rn.ftz.f32 %f160, %f53, %f54, %f160;

{ cvt.f32.f16 %f43, %rs2;}



dp4a.s32.s32 %r169, %r236, %r138, %r238;


dp4a.s32.s32 %r173, %r240, %r142, %r169;


dp4a.s32.s32 %r177, %r244, %r146, %r173;


dp4a.s32.s32 %r181, %r248, %r150, %r177;


dp4a.s32.s32 %r185, %r252, %r154, %r181;


dp4a.s32.s32 %r189, %r256, %r158, %r185;


dp4a.s32.s32 %r193, %r260, %r162, %r189;


dp4a.s32.s32 %r197, %r264, %r166, %r193;


{.reg .f16 low,high;
mov.b32 {low,high},%r168;
cvt.f32.f16 %f44, low;}


mul.ftz.f32 %f55, %f43, %f44;
cvt.rn.f32.s32 %f56, %r197;
fma.rn.ftz.f32 %f159, %f55, %f56, %f159;
cvt.u64.u32 %rd40, %r496;
add.s64 %rd41, %rd1, %rd40;
mul.lo.s64 %rd42, %rd41, 36;

{ cvt.f32.f16 %f45, %rs1;}


add.s64 %rd43, %rd6, %rd42;
ld.global.nc.u32 %r204, [%rd43+4];
ld.global.nc.u32 %r208, [%rd43+8];
ld.global.nc.u32 %r212, [%rd43+12];
ld.global.nc.u32 %r216, [%rd43+16];

dp4a.s32.s32 %r202, %r203, %r204, %r238;


dp4a.s32.s32 %r206, %r207, %r208, %r202;


dp4a.s32.s32 %r210, %r211, %r212, %r206;


dp4a.s32.s32 %r214, %r215, %r216, %r210;

ld.global.nc.u32 %r220, [%rd43+20];
ld.global.nc.u32 %r224, [%rd43+24];
ld.global.nc.u32 %r228, [%rd43+28];
ld.global.nc.u32 %r232, [%rd43+32];

dp4a.s32.s32 %r218, %r219, %r220, %r214;


dp4a.s32.s32 %r222, %r223, %r224, %r218;


dp4a.s32.s32 %r226, %r227, %r228, %r222;


dp4a.s32.s32 %r230, %r231, %r232, %r226;

ld.global.nc.u32 %r234, [%rd43];

{.reg .f16 low,high;
mov.b32 {low,high},%r234;
cvt.f32.f16 %f46, low;}


mul.ftz.f32 %f57, %f45, %f46;
cvt.rn.f32.s32 %f58, %r230;
fma.rn.ftz.f32 %f158, %f57, %f58, %f158;

{ cvt.f32.f16 %f47, %rs2;}



dp4a.s32.s32 %r235, %r236, %r204, %r238;


dp4a.s32.s32 %r239, %r240, %r208, %r235;


dp4a.s32.s32 %r243, %r244, %r212, %r239;


dp4a.s32.s32 %r247, %r248, %r216, %r243;


dp4a.s32.s32 %r251, %r252, %r220, %r247;


dp4a.s32.s32 %r255, %r256, %r224, %r251;


dp4a.s32.s32 %r259, %r260, %r228, %r255;


dp4a.s32.s32 %r263, %r264, %r232, %r259;


{.reg .f16 low,high;
mov.b32 {low,high},%r234;
cvt.f32.f16 %f48, low;}


mul.ftz.f32 %f59, %f47, %f48;
cvt.rn.f32.s32 %f60, %r263;
fma.rn.ftz.f32 %f157, %f59, %f60, %f157;
add.s32 %r496, %r496, 128;
add.s32 %r495, %r495, 128;
add.s32 %r494, %r494, 128;
add.s64 %rd67, %rd67, 576;
add.s64 %rd66, %rd66, 576;
add.s64 %rd65, %rd65, 576;
add.s64 %rd64, %rd64, 576;
add.s32 %r497, %r497, 32;
setp.lt.s32 %p3, %r497, %r37;
@%p3 bra $L__BB7_2;

$L__BB7_3:
setp.eq.s32 %p4, %r1, 0;
@%p4 bra $L__BB7_5;

mov.u32 %r337, _ZZ13mul_mat_vec_qIL9ggml_type41ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r338, %r1, 768, %r337;
shl.b32 %r339, %r32, 2;
add.s32 %r340, %r338, %r339;
st.shared.f32 [%r340+-768], %f162;
st.shared.f32 [%r340+-640], %f161;
st.shared.f32 [%r340+-512], %f160;
st.shared.f32 [%r340+-384], %f159;
st.shared.f32 [%r340+-256], %f158;
st.shared.f32 [%r340+-128], %f157;

$L__BB7_5:
setp.ne.s32 %p5, %r1, 0;
bar.sync 0;
@%p5 bra $L__BB7_18;

mov.u32 %r341, %ctaid.z;
mov.u32 %r342, %ctaid.y;
mov.u32 %r343, %ctaid.x;
shl.b32 %r344, %r343, 1;
mov.u32 %r345, 1;
mad.lo.s32 %r346, %r342, %r24, %r344;
mad.lo.s32 %r347, %r341, %r30, %r346;
cvt.u64.u32 %rd15, %r347;
shl.b32 %r349, %r32, 2;
mov.u32 %r350, _ZZ13mul_mat_vec_qIL9ggml_type41ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r351, %r350, %r349;
mov.u32 %r352, 2;
ld.shared.f32 %f61, [%r351];
add.ftz.f32 %f62, %f61, %f162;
ld.shared.f32 %f63, [%r351+768];
add.ftz.f32 %f64, %f63, %f62;
ld.shared.f32 %f65, [%r351+1536];
add.ftz.f32 %f66, %f65, %f64;
mov.b32 %r353, %f66;
mov.u32 %r354, 31;
mov.u32 %r355, 16;
mov.u32 %r356, -1;
shfl.sync.bfly.b32 %r357|%p6, %r353, %r355, %r354, %r356;
mov.b32 %f67, %r357;
add.ftz.f32 %f68, %f66, %f67;
mov.b32 %r358, %f68;
mov.u32 %r359, 8;
shfl.sync.bfly.b32 %r360|%p7, %r358, %r359, %r354, %r356;
mov.b32 %f69, %r360;
add.ftz.f32 %f70, %f68, %f69;
mov.b32 %r361, %f70;
mov.u32 %r362, 4;
shfl.sync.bfly.b32 %r363|%p8, %r361, %r362, %r354, %r356;
mov.b32 %f71, %r363;
add.ftz.f32 %f72, %f70, %f71;
mov.b32 %r364, %f72;
shfl.sync.bfly.b32 %r365|%p9, %r364, %r352, %r354, %r356;
mov.b32 %f73, %r365;
add.ftz.f32 %f74, %f72, %f73;
mov.b32 %r366, %f74;
shfl.sync.bfly.b32 %r367|%p10, %r366, %r345, %r354, %r356;
mov.b32 %f75, %r367;
add.ftz.f32 %f19, %f74, %f75;
setp.ne.s32 %p11, %r32, 0;
setp.ge.u32 %p12, %r344, %r18;
or.pred %p13, %p11, %p12;
@%p13 bra $L__BB7_8;

cvta.to.global.u64 %rd44, %rd19;
mul.wide.u32 %rd45, %r347, 4;
add.s64 %rd46, %rd44, %rd45;
st.global.f32 [%rd46], %f19;

$L__BB7_8:
ld.shared.f32 %f76, [%r351+128];
add.ftz.f32 %f77, %f76, %f161;
ld.shared.f32 %f78, [%r351+896];
add.ftz.f32 %f79, %f78, %f77;
ld.shared.f32 %f80, [%r351+1664];
add.ftz.f32 %f81, %f80, %f79;
mov.b32 %r379, %f81;
shfl.sync.bfly.b32 %r383|%p14, %r379, %r355, %r354, %r356;
mov.b32 %f82, %r383;
add.ftz.f32 %f83, %f81, %f82;
mov.b32 %r384, %f83;
shfl.sync.bfly.b32 %r386|%p15, %r384, %r359, %r354, %r356;
mov.b32 %f84, %r386;
add.ftz.f32 %f85, %f83, %f84;
mov.b32 %r387, %f85;
shfl.sync.bfly.b32 %r389|%p16, %r387, %r362, %r354, %r356;
mov.b32 %f86, %r389;
add.ftz.f32 %f87, %f85, %f86;
mov.b32 %r390, %f87;
shfl.sync.bfly.b32 %r391|%p17, %r390, %r352, %r354, %r356;
mov.b32 %f88, %r391;
add.ftz.f32 %f89, %f87, %f88;
mov.b32 %r392, %f89;
shfl.sync.bfly.b32 %r394|%p18, %r392, %r345, %r354, %r356;
mov.b32 %f90, %r394;
add.ftz.f32 %f20, %f89, %f90;
setp.eq.s32 %p19, %r32, 1;
or.b32 %r397, %r344, 1;
setp.lt.u32 %p20, %r397, %r18;
and.pred %p1, %p19, %p20;
not.pred %p21, %p1;
@%p21 bra $L__BB7_10;

cvta.to.global.u64 %rd47, %rd19;
mul.wide.u32 %rd48, %r347, 4;
add.s64 %rd49, %rd47, %rd48;
st.global.f32 [%rd49+4], %f20;

$L__BB7_10:
mov.u32 %r408, 2;
ld.shared.f32 %f91, [%r351+256];
add.ftz.f32 %f92, %f91, %f160;
ld.shared.f32 %f93, [%r351+1024];
add.ftz.f32 %f94, %f93, %f92;
ld.shared.f32 %f95, [%r351+1792];
add.ftz.f32 %f96, %f95, %f94;
mov.b32 %r409, %f96;
mov.u32 %r410, 31;
mov.u32 %r411, 16;
mov.u32 %r412, -1;
shfl.sync.bfly.b32 %r413|%p22, %r409, %r411, %r410, %r412;
mov.b32 %f97, %r413;
add.ftz.f32 %f98, %f96, %f97;
mov.b32 %r414, %f98;
mov.u32 %r415, 8;
shfl.sync.bfly.b32 %r416|%p23, %r414, %r415, %r410, %r412;
mov.b32 %f99, %r416;
add.ftz.f32 %f100, %f98, %f99;
mov.b32 %r417, %f100;
mov.u32 %r418, 4;
shfl.sync.bfly.b32 %r419|%p24, %r417, %r418, %r410, %r412;
mov.b32 %f101, %r419;
add.ftz.f32 %f102, %f100, %f101;
mov.b32 %r420, %f102;
shfl.sync.bfly.b32 %r421|%p25, %r420, %r408, %r410, %r412;
mov.b32 %f103, %r421;
add.ftz.f32 %f104, %f102, %f103;
mov.b32 %r422, %f104;
mov.u32 %r423, 1;
shfl.sync.bfly.b32 %r424|%p26, %r422, %r423, %r410, %r412;
mov.b32 %f105, %r424;
add.ftz.f32 %f21, %f104, %f105;
@%p13 bra $L__BB7_12;

cvt.u64.u32 %rd50, %r18;
add.s64 %rd51, %rd50, %rd15;
cvta.to.global.u64 %rd52, %rd19;
shl.b64 %rd53, %rd51, 2;
add.s64 %rd54, %rd52, %rd53;
st.global.f32 [%rd54], %f21;

$L__BB7_12:
ld.shared.f32 %f106, [%r351+384];
add.ftz.f32 %f107, %f106, %f159;
ld.shared.f32 %f108, [%r351+1152];
add.ftz.f32 %f109, %f108, %f107;
ld.shared.f32 %f110, [%r351+1920];
add.ftz.f32 %f111, %f110, %f109;
mov.b32 %r432, %f111;
shfl.sync.bfly.b32 %r436|%p30, %r432, %r411, %r410, %r412;
mov.b32 %f112, %r436;
add.ftz.f32 %f113, %f111, %f112;
mov.b32 %r437, %f113;
shfl.sync.bfly.b32 %r439|%p31, %r437, %r415, %r410, %r412;
mov.b32 %f114, %r439;
add.ftz.f32 %f115, %f113, %f114;
mov.b32 %r440, %f115;
shfl.sync.bfly.b32 %r442|%p32, %r440, %r418, %r410, %r412;
mov.b32 %f116, %r442;
add.ftz.f32 %f117, %f115, %f116;
mov.b32 %r443, %f117;
shfl.sync.bfly.b32 %r444|%p33, %r443, %r408, %r410, %r412;
mov.b32 %f118, %r444;
add.ftz.f32 %f119, %f117, %f118;
mov.b32 %r445, %f119;
shfl.sync.bfly.b32 %r447|%p34, %r445, %r423, %r410, %r412;
mov.b32 %f120, %r447;
add.ftz.f32 %f22, %f119, %f120;
@%p21 bra $L__BB7_14;

add.s32 %r448, %r18, 1;
cvt.u64.u32 %rd55, %r448;
add.s64 %rd56, %rd55, %rd15;
cvta.to.global.u64 %rd57, %rd19;
shl.b64 %rd58, %rd56, 2;
add.s64 %rd59, %rd57, %rd58;
st.global.f32 [%rd59], %f22;

$L__BB7_14:
shl.b32 %r449, %r18, 1;
mov.u32 %r450, 1;
mov.u32 %r455, 2;
ld.shared.f32 %f121, [%r351+512];
add.ftz.f32 %f122, %f121, %f158;
ld.shared.f32 %f123, [%r351+1280];
add.ftz.f32 %f124, %f123, %f122;
ld.shared.f32 %f125, [%r351+2048];
add.ftz.f32 %f126, %f125, %f124;
mov.b32 %r456, %f126;
mov.u32 %r457, 31;
mov.u32 %r458, 16;
mov.u32 %r459, -1;
shfl.sync.bfly.b32 %r460|%p36, %r456, %r458, %r457, %r459;
mov.b32 %f127, %r460;
add.ftz.f32 %f128, %f126, %f127;
mov.b32 %r461, %f128;
mov.u32 %r462, 8;
shfl.sync.bfly.b32 %r463|%p37, %r461, %r462, %r457, %r459;
mov.b32 %f129, %r463;
add.ftz.f32 %f130, %f128, %f129;
mov.b32 %r464, %f130;
mov.u32 %r465, 4;
shfl.sync.bfly.b32 %r466|%p38, %r464, %r465, %r457, %r459;
mov.b32 %f131, %r466;
add.ftz.f32 %f132, %f130, %f131;
mov.b32 %r467, %f132;
shfl.sync.bfly.b32 %r468|%p39, %r467, %r455, %r457, %r459;
mov.b32 %f133, %r468;
add.ftz.f32 %f134, %f132, %f133;
mov.b32 %r469, %f134;
shfl.sync.bfly.b32 %r470|%p40, %r469, %r450, %r457, %r459;
mov.b32 %f135, %r470;
add.ftz.f32 %f23, %f134, %f135;
cvt.u64.u32 %rd60, %r449;
add.s64 %rd61, %rd60, %rd15;
cvta.to.global.u64 %rd62, %rd19;
shl.b64 %rd63, %rd61, 2;
add.s64 %rd16, %rd62, %rd63;
@%p13 bra $L__BB7_16;

st.global.f32 [%rd16], %f23;

$L__BB7_16:
ld.shared.f32 %f136, [%r351+640];
add.ftz.f32 %f137, %f136, %f157;
ld.shared.f32 %f138, [%r351+1408];
add.ftz.f32 %f139, %f138, %f137;
ld.shared.f32 %f140, [%r351+2176];
add.ftz.f32 %f141, %f140, %f139;
mov.b32 %r478, %f141;
shfl.sync.bfly.b32 %r482|%p44, %r478, %r458, %r457, %r459;
mov.b32 %f142, %r482;
add.ftz.f32 %f143, %f141, %f142;
mov.b32 %r483, %f143;
shfl.sync.bfly.b32 %r485|%p45, %r483, %r462, %r457, %r459;
mov.b32 %f144, %r485;
add.ftz.f32 %f145, %f143, %f144;
mov.b32 %r486, %f145;
shfl.sync.bfly.b32 %r488|%p46, %r486, %r465, %r457, %r459;
mov.b32 %f146, %r488;
add.ftz.f32 %f147, %f145, %f146;
mov.b32 %r489, %f147;
shfl.sync.bfly.b32 %r490|%p47, %r489, %r455, %r457, %r459;
mov.b32 %f148, %r490;
add.ftz.f32 %f149, %f147, %f148;
mov.b32 %r491, %f149;
shfl.sync.bfly.b32 %r493|%p48, %r491, %r450, %r457, %r459;
mov.b32 %f150, %r493;
add.ftz.f32 %f24, %f149, %f150;
@%p21 bra $L__BB7_18;

st.global.f32 [%rd16+4], %f24;

$L__BB7_18:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type41ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type41ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type41ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type41ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type41ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type41ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type41ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type41ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type41ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<69>;
.reg .b16 %rs<13>;
.reg .f32 %f<217>;
.reg .b32 %r<647>;
.reg .b64 %rd<86>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type41ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[3072];

ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type41ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type41ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type41ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r19, [_Z13mul_mat_vec_qIL9ggml_type41ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type41ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type41ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type41ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type41ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type41ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type41ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type41ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type41ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type41ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type41ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type41ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type41ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type41ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r35, %tid.y;
shl.b32 %r36, %r35, 5;
mov.u32 %r37, %tid.x;
add.s32 %r1, %r36, %r37;
shr.s32 %r38, %r1, 31;
shr.u32 %r39, %r38, 30;
add.s32 %r40, %r1, %r39;
shr.s32 %r646, %r40, 2;
shr.u32 %r41, %r19, 7;
setp.ge.s32 %p1, %r646, %r41;
mov.f32 %f209, 0f00000000;
mov.f32 %f210, %f209;
mov.f32 %f211, %f209;
mov.f32 %f212, %f209;
mov.f32 %f213, %f209;
mov.f32 %f214, %f209;
mov.f32 %f215, %f209;
mov.f32 %f216, %f209;
@%p1 bra $L__BB8_3;

mov.u32 %r42, %ctaid.y;
mov.u32 %r43, %ctaid.z;
mul.hi.u32 %r44, %r43, %r29;
add.s32 %r45, %r44, %r43;
shr.u32 %r46, %r45, %r30;
mul.lo.s32 %r47, %r46, %r32;
mul.hi.u32 %r48, %r42, %r23;
add.s32 %r49, %r48, %r42;
shr.u32 %r50, %r49, %r24;
mul.lo.s32 %r51, %r50, %r26;
mul.lo.s32 %r52, %r42, %r27;
cvt.u64.u32 %rd19, %r52;
and.b32 %r56, %r40, -4;
sub.s32 %r57, %r1, %r56;
cvt.s64.s32 %rd20, %r57;
add.s64 %rd21, %rd19, %rd20;
mul.lo.s32 %r58, %r43, %r33;
cvt.u64.u32 %rd22, %r58;
add.s64 %rd1, %rd21, %rd22;
shl.b32 %r59, %r57, 1;
mov.u32 %r60, %ctaid.x;
shl.b32 %r61, %r60, 1;
or.b32 %r62, %r61, 1;
shl.b32 %r63, %r21, 1;
shl.b32 %r642, %r646, 2;
add.s32 %r645, %r642, %r63;
mad.lo.s32 %r644, %r21, 3, %r642;
add.s32 %r643, %r21, %r642;
mad.lo.s32 %r64, %r61, %r20, %r646;
add.s32 %r65, %r64, %r47;
add.s32 %r66, %r65, %r51;
mul.wide.s32 %rd23, %r66, 18;
mul.wide.s32 %rd24, %r59, 2;
add.s64 %rd25, %rd23, %rd24;
cvta.to.global.u64 %rd26, %rd16;
add.s64 %rd27, %rd26, %rd25;
add.s64 %rd85, %rd27, 4;
add.s32 %r67, %r646, %r47;
add.s32 %r68, %r67, %r51;
mad.lo.s32 %r69, %r62, %r20, %r68;
mul.wide.s32 %rd28, %r69, 18;
add.s64 %rd29, %rd28, %rd24;
add.s64 %rd30, %rd26, %rd29;
add.s64 %rd84, %rd30, 4;
add.s64 %rd83, %rd26, %rd28;
add.s64 %rd82, %rd26, %rd23;
cvta.to.global.u64 %rd6, %rd17;
mov.f32 %f209, 0f00000000;

$L__BB8_2:
cvt.u64.u32 %rd31, %r642;
add.s64 %rd32, %rd1, %rd31;
mul.lo.s64 %rd33, %rd32, 36;
ld.global.nc.u16 %rs1, [%rd82];

{ cvt.f32.f16 %f49, %rs1;}


ld.global.nc.u16 %rs9, [%rd85+-2];
cvt.s32.s16 %r334, %rs9;
add.s64 %rd34, %rd6, %rd33;
ld.global.nc.u32 %r72, [%rd34+4];
ld.global.nc.u32 %r76, [%rd34+8];
ld.global.nc.u32 %r80, [%rd34+12];
ld.global.nc.u32 %r84, [%rd34+16];
and.b32 %r335, %r334, 30583;
mov.u32 %r336, 286261504;
prmt.b32 %r337, %r336, %r336, %r335;
shr.u32 %r338, %r334, 2;
and.b32 %r339, %r338, 30583;
prmt.b32 %r340, %r336, %r336, %r339;
and.b32 %r341, %r337, 30583;
mov.u32 %r342, 511;
prmt.b32 %r343, %r342, %r342, %r341;
and.b32 %r344, %r340, 30583;
prmt.b32 %r345, %r342, %r342, %r344;
shr.u32 %r346, %r337, 16;
and.b32 %r347, %r346, 30583;
prmt.b32 %r348, %r342, %r342, %r347;
shr.u32 %r349, %r340, 16;
and.b32 %r350, %r349, 30583;
prmt.b32 %r351, %r342, %r342, %r350;
mov.u32 %r352, 21520;
prmt.b32 %r269, %r343, %r345, %r352;
mov.u32 %r353, 30258;
prmt.b32 %r273, %r343, %r345, %r353;
prmt.b32 %r277, %r348, %r351, %r352;
prmt.b32 %r281, %r348, %r351, %r353;
mov.u32 %r304, 0;

dp4a.s32.s32 %r70, %r269, %r72, %r304;


dp4a.s32.s32 %r74, %r273, %r76, %r70;


dp4a.s32.s32 %r78, %r277, %r80, %r74;


dp4a.s32.s32 %r82, %r281, %r84, %r78;

ld.global.nc.u16 %rs10, [%rd85];
cvt.s32.s16 %r354, %rs10;
ld.global.nc.u32 %r88, [%rd34+20];
ld.global.nc.u32 %r92, [%rd34+24];
ld.global.nc.u32 %r96, [%rd34+28];
ld.global.nc.u32 %r100, [%rd34+32];
and.b32 %r355, %r354, 30583;
prmt.b32 %r356, %r336, %r336, %r355;
shr.u32 %r357, %r354, 2;
and.b32 %r358, %r357, 30583;
prmt.b32 %r359, %r336, %r336, %r358;
and.b32 %r360, %r356, 30583;
prmt.b32 %r361, %r342, %r342, %r360;
and.b32 %r362, %r359, 30583;
prmt.b32 %r363, %r342, %r342, %r362;
shr.u32 %r364, %r356, 16;
and.b32 %r365, %r364, 30583;
prmt.b32 %r366, %r342, %r342, %r365;
shr.u32 %r367, %r359, 16;
and.b32 %r368, %r367, 30583;
prmt.b32 %r369, %r342, %r342, %r368;
prmt.b32 %r285, %r361, %r363, %r352;
prmt.b32 %r289, %r361, %r363, %r353;
prmt.b32 %r293, %r366, %r369, %r352;
prmt.b32 %r297, %r366, %r369, %r353;

dp4a.s32.s32 %r86, %r285, %r88, %r82;


dp4a.s32.s32 %r90, %r289, %r92, %r86;


dp4a.s32.s32 %r94, %r293, %r96, %r90;


dp4a.s32.s32 %r98, %r297, %r100, %r94;

ld.global.nc.u32 %r102, [%rd34];

{.reg .f16 low,high;
mov.b32 {low,high},%r102;
cvt.f32.f16 %f50, low;}


mul.ftz.f32 %f65, %f49, %f50;
cvt.rn.f32.s32 %f66, %r98;
fma.rn.ftz.f32 %f216, %f65, %f66, %f216;
ld.global.nc.u16 %rs2, [%rd83];

{ cvt.f32.f16 %f51, %rs2;}


ld.global.nc.u16 %rs11, [%rd84+-2];
cvt.s32.s16 %r370, %rs11;
and.b32 %r371, %r370, 30583;
prmt.b32 %r372, %r336, %r336, %r371;
shr.u32 %r373, %r370, 2;
and.b32 %r374, %r373, 30583;
prmt.b32 %r375, %r336, %r336, %r374;
and.b32 %r376, %r372, 30583;
prmt.b32 %r377, %r342, %r342, %r376;
and.b32 %r378, %r375, 30583;
prmt.b32 %r379, %r342, %r342, %r378;
shr.u32 %r380, %r372, 16;
and.b32 %r381, %r380, 30583;
prmt.b32 %r382, %r342, %r342, %r381;
shr.u32 %r383, %r375, 16;
and.b32 %r384, %r383, 30583;
prmt.b32 %r385, %r342, %r342, %r384;
prmt.b32 %r302, %r377, %r379, %r352;
prmt.b32 %r306, %r377, %r379, %r353;
prmt.b32 %r310, %r382, %r385, %r352;
prmt.b32 %r314, %r382, %r385, %r353;

dp4a.s32.s32 %r103, %r302, %r72, %r304;


dp4a.s32.s32 %r107, %r306, %r76, %r103;


dp4a.s32.s32 %r111, %r310, %r80, %r107;


dp4a.s32.s32 %r115, %r314, %r84, %r111;

ld.global.nc.u16 %rs12, [%rd84];
cvt.s32.s16 %r386, %rs12;
and.b32 %r387, %r386, 30583;
prmt.b32 %r388, %r336, %r336, %r387;
shr.u32 %r389, %r386, 2;
and.b32 %r390, %r389, 30583;
prmt.b32 %r391, %r336, %r336, %r390;
and.b32 %r392, %r388, 30583;
prmt.b32 %r393, %r342, %r342, %r392;
and.b32 %r394, %r391, 30583;
prmt.b32 %r395, %r342, %r342, %r394;
shr.u32 %r396, %r388, 16;
and.b32 %r397, %r396, 30583;
prmt.b32 %r398, %r342, %r342, %r397;
shr.u32 %r399, %r391, 16;
and.b32 %r400, %r399, 30583;
prmt.b32 %r401, %r342, %r342, %r400;
prmt.b32 %r318, %r393, %r395, %r352;
prmt.b32 %r322, %r393, %r395, %r353;
prmt.b32 %r326, %r398, %r401, %r352;
prmt.b32 %r330, %r398, %r401, %r353;

dp4a.s32.s32 %r119, %r318, %r88, %r115;


dp4a.s32.s32 %r123, %r322, %r92, %r119;


dp4a.s32.s32 %r127, %r326, %r96, %r123;


dp4a.s32.s32 %r131, %r330, %r100, %r127;


{.reg .f16 low,high;
mov.b32 {low,high},%r102;
cvt.f32.f16 %f52, low;}


mul.ftz.f32 %f67, %f51, %f52;
cvt.rn.f32.s32 %f68, %r131;
fma.rn.ftz.f32 %f215, %f67, %f68, %f215;
cvt.u64.u32 %rd35, %r643;
add.s64 %rd36, %rd1, %rd35;
mul.lo.s64 %rd37, %rd36, 36;

{ cvt.f32.f16 %f53, %rs1;}


add.s64 %rd38, %rd6, %rd37;
ld.global.nc.u32 %r138, [%rd38+4];
ld.global.nc.u32 %r142, [%rd38+8];
ld.global.nc.u32 %r146, [%rd38+12];
ld.global.nc.u32 %r150, [%rd38+16];

dp4a.s32.s32 %r136, %r269, %r138, %r304;


dp4a.s32.s32 %r140, %r273, %r142, %r136;


dp4a.s32.s32 %r144, %r277, %r146, %r140;


dp4a.s32.s32 %r148, %r281, %r150, %r144;

ld.global.nc.u32 %r154, [%rd38+20];
ld.global.nc.u32 %r158, [%rd38+24];
ld.global.nc.u32 %r162, [%rd38+28];
ld.global.nc.u32 %r166, [%rd38+32];

dp4a.s32.s32 %r152, %r285, %r154, %r148;


dp4a.s32.s32 %r156, %r289, %r158, %r152;


dp4a.s32.s32 %r160, %r293, %r162, %r156;


dp4a.s32.s32 %r164, %r297, %r166, %r160;

ld.global.nc.u32 %r168, [%rd38];

{.reg .f16 low,high;
mov.b32 {low,high},%r168;
cvt.f32.f16 %f54, low;}


mul.ftz.f32 %f69, %f53, %f54;
cvt.rn.f32.s32 %f70, %r164;
fma.rn.ftz.f32 %f214, %f69, %f70, %f214;

{ cvt.f32.f16 %f55, %rs2;}



dp4a.s32.s32 %r169, %r302, %r138, %r304;


dp4a.s32.s32 %r173, %r306, %r142, %r169;


dp4a.s32.s32 %r177, %r310, %r146, %r173;


dp4a.s32.s32 %r181, %r314, %r150, %r177;


dp4a.s32.s32 %r185, %r318, %r154, %r181;


dp4a.s32.s32 %r189, %r322, %r158, %r185;


dp4a.s32.s32 %r193, %r326, %r162, %r189;


dp4a.s32.s32 %r197, %r330, %r166, %r193;


{.reg .f16 low,high;
mov.b32 {low,high},%r168;
cvt.f32.f16 %f56, low;}


mul.ftz.f32 %f71, %f55, %f56;
cvt.rn.f32.s32 %f72, %r197;
fma.rn.ftz.f32 %f213, %f71, %f72, %f213;
cvt.u64.u32 %rd39, %r645;
add.s64 %rd40, %rd1, %rd39;
mul.lo.s64 %rd41, %rd40, 36;

{ cvt.f32.f16 %f57, %rs1;}


add.s64 %rd42, %rd6, %rd41;
ld.global.nc.u32 %r204, [%rd42+4];
ld.global.nc.u32 %r208, [%rd42+8];
ld.global.nc.u32 %r212, [%rd42+12];
ld.global.nc.u32 %r216, [%rd42+16];

dp4a.s32.s32 %r202, %r269, %r204, %r304;


dp4a.s32.s32 %r206, %r273, %r208, %r202;


dp4a.s32.s32 %r210, %r277, %r212, %r206;


dp4a.s32.s32 %r214, %r281, %r216, %r210;

ld.global.nc.u32 %r220, [%rd42+20];
ld.global.nc.u32 %r224, [%rd42+24];
ld.global.nc.u32 %r228, [%rd42+28];
ld.global.nc.u32 %r232, [%rd42+32];

dp4a.s32.s32 %r218, %r285, %r220, %r214;


dp4a.s32.s32 %r222, %r289, %r224, %r218;


dp4a.s32.s32 %r226, %r293, %r228, %r222;


dp4a.s32.s32 %r230, %r297, %r232, %r226;

ld.global.nc.u32 %r234, [%rd42];

{.reg .f16 low,high;
mov.b32 {low,high},%r234;
cvt.f32.f16 %f58, low;}


mul.ftz.f32 %f73, %f57, %f58;
cvt.rn.f32.s32 %f74, %r230;
fma.rn.ftz.f32 %f212, %f73, %f74, %f212;

{ cvt.f32.f16 %f59, %rs2;}



dp4a.s32.s32 %r235, %r302, %r204, %r304;


dp4a.s32.s32 %r239, %r306, %r208, %r235;


dp4a.s32.s32 %r243, %r310, %r212, %r239;


dp4a.s32.s32 %r247, %r314, %r216, %r243;


dp4a.s32.s32 %r251, %r318, %r220, %r247;


dp4a.s32.s32 %r255, %r322, %r224, %r251;


dp4a.s32.s32 %r259, %r326, %r228, %r255;


dp4a.s32.s32 %r263, %r330, %r232, %r259;


{.reg .f16 low,high;
mov.b32 {low,high},%r234;
cvt.f32.f16 %f60, low;}


mul.ftz.f32 %f75, %f59, %f60;
cvt.rn.f32.s32 %f76, %r263;
fma.rn.ftz.f32 %f211, %f75, %f76, %f211;
cvt.u64.u32 %rd43, %r644;
add.s64 %rd44, %rd1, %rd43;
mul.lo.s64 %rd45, %rd44, 36;

{ cvt.f32.f16 %f61, %rs1;}


add.s64 %rd46, %rd6, %rd45;
ld.global.nc.u32 %r270, [%rd46+4];
ld.global.nc.u32 %r274, [%rd46+8];
ld.global.nc.u32 %r278, [%rd46+12];
ld.global.nc.u32 %r282, [%rd46+16];

dp4a.s32.s32 %r268, %r269, %r270, %r304;


dp4a.s32.s32 %r272, %r273, %r274, %r268;


dp4a.s32.s32 %r276, %r277, %r278, %r272;


dp4a.s32.s32 %r280, %r281, %r282, %r276;

ld.global.nc.u32 %r286, [%rd46+20];
ld.global.nc.u32 %r290, [%rd46+24];
ld.global.nc.u32 %r294, [%rd46+28];
ld.global.nc.u32 %r298, [%rd46+32];

dp4a.s32.s32 %r284, %r285, %r286, %r280;


dp4a.s32.s32 %r288, %r289, %r290, %r284;


dp4a.s32.s32 %r292, %r293, %r294, %r288;


dp4a.s32.s32 %r296, %r297, %r298, %r292;

ld.global.nc.u32 %r300, [%rd46];

{.reg .f16 low,high;
mov.b32 {low,high},%r300;
cvt.f32.f16 %f62, low;}


mul.ftz.f32 %f77, %f61, %f62;
cvt.rn.f32.s32 %f78, %r296;
fma.rn.ftz.f32 %f210, %f77, %f78, %f210;

{ cvt.f32.f16 %f63, %rs2;}



dp4a.s32.s32 %r301, %r302, %r270, %r304;


dp4a.s32.s32 %r305, %r306, %r274, %r301;


dp4a.s32.s32 %r309, %r310, %r278, %r305;


dp4a.s32.s32 %r313, %r314, %r282, %r309;


dp4a.s32.s32 %r317, %r318, %r286, %r313;


dp4a.s32.s32 %r321, %r322, %r290, %r317;


dp4a.s32.s32 %r325, %r326, %r294, %r321;


dp4a.s32.s32 %r329, %r330, %r298, %r325;


{.reg .f16 low,high;
mov.b32 {low,high},%r300;
cvt.f32.f16 %f64, low;}


mul.ftz.f32 %f79, %f63, %f64;
cvt.rn.f32.s32 %f80, %r329;
fma.rn.ftz.f32 %f209, %f79, %f80, %f209;
add.s32 %r645, %r645, 128;
add.s32 %r644, %r644, 128;
add.s32 %r643, %r643, 128;
add.s32 %r642, %r642, 128;
add.s64 %rd85, %rd85, 576;
add.s64 %rd84, %rd84, 576;
add.s64 %rd83, %rd83, 576;
add.s64 %rd82, %rd82, 576;
add.s32 %r646, %r646, 32;
setp.lt.s32 %p2, %r646, %r41;
@%p2 bra $L__BB8_2;

$L__BB8_3:
setp.eq.s32 %p3, %r35, 0;
@%p3 bra $L__BB8_5;

shl.b32 %r405, %r35, 10;
mov.u32 %r406, _ZZ13mul_mat_vec_qIL9ggml_type41ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r407, %r406, %r405;
shl.b32 %r408, %r37, 2;
add.s32 %r409, %r407, %r408;
st.shared.f32 [%r409+-1024], %f216;
st.shared.f32 [%r409+-896], %f215;
st.shared.f32 [%r409+-768], %f214;
st.shared.f32 [%r409+-640], %f213;
st.shared.f32 [%r409+-512], %f212;
st.shared.f32 [%r409+-384], %f211;
st.shared.f32 [%r409+-256], %f210;
st.shared.f32 [%r409+-128], %f209;

$L__BB8_5:
setp.ne.s32 %p4, %r35, 0;
bar.sync 0;
@%p4 bra $L__BB8_22;

mov.u32 %r411, %ctaid.x;
shl.b32 %r412, %r411, 1;
mov.u32 %r413, 1;
shl.b32 %r415, %r37, 2;
mov.u32 %r416, _ZZ13mul_mat_vec_qIL9ggml_type41ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r417, %r416, %r415;
mov.u32 %r418, 2;
ld.shared.f32 %f81, [%r417];
add.ftz.f32 %f82, %f81, %f216;
ld.shared.f32 %f83, [%r417+1024];
add.ftz.f32 %f84, %f83, %f82;
ld.shared.f32 %f85, [%r417+2048];
add.ftz.f32 %f86, %f85, %f84;
mov.b32 %r419, %f86;
mov.u32 %r420, 31;
mov.u32 %r421, 16;
mov.u32 %r422, -1;
shfl.sync.bfly.b32 %r423|%p5, %r419, %r421, %r420, %r422;
mov.b32 %f87, %r423;
add.ftz.f32 %f88, %f86, %f87;
mov.b32 %r424, %f88;
mov.u32 %r425, 8;
shfl.sync.bfly.b32 %r426|%p6, %r424, %r425, %r420, %r422;
mov.b32 %f89, %r426;
add.ftz.f32 %f90, %f88, %f89;
mov.b32 %r427, %f90;
mov.u32 %r428, 4;
shfl.sync.bfly.b32 %r429|%p7, %r427, %r428, %r420, %r422;
mov.b32 %f91, %r429;
add.ftz.f32 %f92, %f90, %f91;
mov.b32 %r430, %f92;
shfl.sync.bfly.b32 %r431|%p8, %r430, %r418, %r420, %r422;
mov.b32 %f93, %r431;
add.ftz.f32 %f94, %f92, %f93;
mov.b32 %r432, %f94;
shfl.sync.bfly.b32 %r433|%p9, %r432, %r413, %r420, %r422;
mov.b32 %f95, %r433;
add.ftz.f32 %f25, %f94, %f95;
setp.ne.s32 %p10, %r37, 0;
setp.ge.u32 %p11, %r412, %r22;
or.pred %p12, %p10, %p11;
@%p12 bra $L__BB8_8;

cvta.to.global.u64 %rd47, %rd18;
mov.u32 %r434, %ctaid.y;
mad.lo.s32 %r437, %r434, %r28, %r412;
mov.u32 %r438, %ctaid.z;
mad.lo.s32 %r439, %r438, %r34, %r437;
mul.wide.u32 %rd48, %r439, 4;
add.s64 %rd49, %rd47, %rd48;
st.global.f32 [%rd49], %f25;

$L__BB8_8:
ld.shared.f32 %f96, [%r417+128];
add.ftz.f32 %f97, %f96, %f215;
ld.shared.f32 %f98, [%r417+1152];
add.ftz.f32 %f99, %f98, %f97;
ld.shared.f32 %f100, [%r417+2176];
add.ftz.f32 %f101, %f100, %f99;
mov.b32 %r445, %f101;
shfl.sync.bfly.b32 %r449|%p13, %r445, %r421, %r420, %r422;
mov.b32 %f102, %r449;
add.ftz.f32 %f103, %f101, %f102;
mov.b32 %r450, %f103;
shfl.sync.bfly.b32 %r452|%p14, %r450, %r425, %r420, %r422;
mov.b32 %f104, %r452;
add.ftz.f32 %f105, %f103, %f104;
mov.b32 %r453, %f105;
shfl.sync.bfly.b32 %r455|%p15, %r453, %r428, %r420, %r422;
mov.b32 %f106, %r455;
add.ftz.f32 %f107, %f105, %f106;
mov.b32 %r456, %f107;
shfl.sync.bfly.b32 %r457|%p16, %r456, %r418, %r420, %r422;
mov.b32 %f108, %r457;
add.ftz.f32 %f109, %f107, %f108;
mov.b32 %r458, %f109;
shfl.sync.bfly.b32 %r460|%p17, %r458, %r413, %r420, %r422;
mov.b32 %f110, %r460;
add.ftz.f32 %f26, %f109, %f110;
setp.ne.s32 %p18, %r37, 1;
or.b32 %r463, %r412, 1;
setp.ge.u32 %p19, %r463, %r22;
or.pred %p20, %p18, %p19;
@%p20 bra $L__BB8_10;

cvta.to.global.u64 %rd50, %rd18;
mov.u32 %r464, %ctaid.y;
mad.lo.s32 %r467, %r464, %r28, %r412;
mov.u32 %r468, %ctaid.z;
mad.lo.s32 %r469, %r468, %r34, %r467;
mul.wide.u32 %rd51, %r469, 4;
add.s64 %rd52, %rd50, %rd51;
st.global.f32 [%rd52+4], %f26;

$L__BB8_10:
mov.u32 %r474, 2;
ld.shared.f32 %f111, [%r417+256];
add.ftz.f32 %f112, %f111, %f214;
ld.shared.f32 %f113, [%r417+1280];
add.ftz.f32 %f114, %f113, %f112;
ld.shared.f32 %f115, [%r417+2304];
add.ftz.f32 %f116, %f115, %f114;
mov.b32 %r475, %f116;
mov.u32 %r476, 31;
mov.u32 %r477, 16;
mov.u32 %r478, -1;
shfl.sync.bfly.b32 %r479|%p21, %r475, %r477, %r476, %r478;
mov.b32 %f117, %r479;
add.ftz.f32 %f118, %f116, %f117;
mov.b32 %r480, %f118;
mov.u32 %r481, 8;
shfl.sync.bfly.b32 %r482|%p22, %r480, %r481, %r476, %r478;
mov.b32 %f119, %r482;
add.ftz.f32 %f120, %f118, %f119;
mov.b32 %r483, %f120;
mov.u32 %r484, 4;
shfl.sync.bfly.b32 %r485|%p23, %r483, %r484, %r476, %r478;
mov.b32 %f121, %r485;
add.ftz.f32 %f122, %f120, %f121;
mov.b32 %r486, %f122;
shfl.sync.bfly.b32 %r487|%p24, %r486, %r474, %r476, %r478;
mov.b32 %f123, %r487;
add.ftz.f32 %f124, %f122, %f123;
mov.b32 %r488, %f124;
mov.u32 %r489, 1;
shfl.sync.bfly.b32 %r490|%p25, %r488, %r489, %r476, %r478;
mov.b32 %f125, %r490;
add.ftz.f32 %f27, %f124, %f125;
@%p12 bra $L__BB8_12;

cvt.u64.u32 %rd53, %r22;
mov.u32 %r493, %ctaid.y;
mad.lo.s32 %r496, %r493, %r28, %r412;
mov.u32 %r497, %ctaid.z;
mad.lo.s32 %r498, %r497, %r34, %r496;
cvt.u64.u32 %rd54, %r498;
add.s64 %rd55, %rd53, %rd54;
cvta.to.global.u64 %rd56, %rd18;
shl.b64 %rd57, %rd55, 2;
add.s64 %rd58, %rd56, %rd57;
st.global.f32 [%rd58], %f27;

$L__BB8_12:
ld.shared.f32 %f126, [%r417+384];
add.ftz.f32 %f127, %f126, %f213;
ld.shared.f32 %f128, [%r417+1408];
add.ftz.f32 %f129, %f128, %f127;
ld.shared.f32 %f130, [%r417+2432];
add.ftz.f32 %f131, %f130, %f129;
mov.b32 %r504, %f131;
shfl.sync.bfly.b32 %r508|%p29, %r504, %r477, %r476, %r478;
mov.b32 %f132, %r508;
add.ftz.f32 %f133, %f131, %f132;
mov.b32 %r509, %f133;
shfl.sync.bfly.b32 %r511|%p30, %r509, %r481, %r476, %r478;
mov.b32 %f134, %r511;
add.ftz.f32 %f135, %f133, %f134;
mov.b32 %r512, %f135;
shfl.sync.bfly.b32 %r514|%p31, %r512, %r484, %r476, %r478;
mov.b32 %f136, %r514;
add.ftz.f32 %f137, %f135, %f136;
mov.b32 %r515, %f137;
shfl.sync.bfly.b32 %r516|%p32, %r515, %r474, %r476, %r478;
mov.b32 %f138, %r516;
add.ftz.f32 %f139, %f137, %f138;
mov.b32 %r517, %f139;
shfl.sync.bfly.b32 %r519|%p33, %r517, %r489, %r476, %r478;
mov.b32 %f140, %r519;
add.ftz.f32 %f28, %f139, %f140;
@%p20 bra $L__BB8_14;

add.s32 %r523, %r22, 1;
cvt.u64.u32 %rd59, %r523;
mov.u32 %r524, %ctaid.y;
mad.lo.s32 %r527, %r524, %r28, %r412;
mov.u32 %r528, %ctaid.z;
mad.lo.s32 %r529, %r528, %r34, %r527;
cvt.u64.u32 %rd60, %r529;
add.s64 %rd61, %rd59, %rd60;
cvta.to.global.u64 %rd62, %rd18;
shl.b64 %rd63, %rd61, 2;
add.s64 %rd64, %rd62, %rd63;
st.global.f32 [%rd64], %f28;

$L__BB8_14:
shl.b32 %r530, %r22, 1;
mov.u32 %r531, 1;
mov.u32 %r536, 2;
ld.shared.f32 %f141, [%r417+512];
add.ftz.f32 %f142, %f141, %f212;
ld.shared.f32 %f143, [%r417+1536];
add.ftz.f32 %f144, %f143, %f142;
ld.shared.f32 %f145, [%r417+2560];
add.ftz.f32 %f146, %f145, %f144;
mov.b32 %r537, %f146;
mov.u32 %r538, 31;
mov.u32 %r539, 16;
mov.u32 %r540, -1;
shfl.sync.bfly.b32 %r541|%p37, %r537, %r539, %r538, %r540;
mov.b32 %f147, %r541;
add.ftz.f32 %f148, %f146, %f147;
mov.b32 %r542, %f148;
mov.u32 %r543, 8;
shfl.sync.bfly.b32 %r544|%p38, %r542, %r543, %r538, %r540;
mov.b32 %f149, %r544;
add.ftz.f32 %f150, %f148, %f149;
mov.b32 %r545, %f150;
mov.u32 %r546, 4;
shfl.sync.bfly.b32 %r547|%p39, %r545, %r546, %r538, %r540;
mov.b32 %f151, %r547;
add.ftz.f32 %f152, %f150, %f151;
mov.b32 %r548, %f152;
shfl.sync.bfly.b32 %r549|%p40, %r548, %r536, %r538, %r540;
mov.b32 %f153, %r549;
add.ftz.f32 %f154, %f152, %f153;
mov.b32 %r550, %f154;
shfl.sync.bfly.b32 %r551|%p41, %r550, %r531, %r538, %r540;
mov.b32 %f155, %r551;
add.ftz.f32 %f29, %f154, %f155;
cvt.u64.u32 %rd65, %r530;
mov.u32 %r552, %ctaid.y;
mad.lo.s32 %r555, %r552, %r28, %r412;
mov.u32 %r556, %ctaid.z;
mad.lo.s32 %r557, %r556, %r34, %r555;
cvt.u64.u32 %rd66, %r557;
add.s64 %rd67, %rd65, %rd66;
cvta.to.global.u64 %rd68, %rd18;
shl.b64 %rd69, %rd67, 2;
add.s64 %rd15, %rd68, %rd69;
@%p12 bra $L__BB8_16;

st.global.f32 [%rd15], %f29;

$L__BB8_16:
ld.shared.f32 %f156, [%r417+640];
add.ftz.f32 %f157, %f156, %f211;
ld.shared.f32 %f158, [%r417+1664];
add.ftz.f32 %f159, %f158, %f157;
ld.shared.f32 %f160, [%r417+2688];
add.ftz.f32 %f161, %f160, %f159;
mov.b32 %r563, %f161;
shfl.sync.bfly.b32 %r567|%p45, %r563, %r539, %r538, %r540;
mov.b32 %f162, %r567;
add.ftz.f32 %f163, %f161, %f162;
mov.b32 %r568, %f163;
shfl.sync.bfly.b32 %r570|%p46, %r568, %r543, %r538, %r540;
mov.b32 %f164, %r570;
add.ftz.f32 %f165, %f163, %f164;
mov.b32 %r571, %f165;
shfl.sync.bfly.b32 %r573|%p47, %r571, %r546, %r538, %r540;
mov.b32 %f166, %r573;
add.ftz.f32 %f167, %f165, %f166;
mov.b32 %r574, %f167;
shfl.sync.bfly.b32 %r575|%p48, %r574, %r536, %r538, %r540;
mov.b32 %f168, %r575;
add.ftz.f32 %f169, %f167, %f168;
mov.b32 %r576, %f169;
shfl.sync.bfly.b32 %r578|%p49, %r576, %r531, %r538, %r540;
mov.b32 %f170, %r578;
add.ftz.f32 %f30, %f169, %f170;
@%p20 bra $L__BB8_18;

st.global.f32 [%rd15+4], %f30;

$L__BB8_18:
mul.lo.s32 %r18, %r22, 3;
mov.u32 %r586, 2;
ld.shared.f32 %f171, [%r417+768];
add.ftz.f32 %f172, %f171, %f210;
ld.shared.f32 %f173, [%r417+1792];
add.ftz.f32 %f174, %f173, %f172;
ld.shared.f32 %f175, [%r417+2816];
add.ftz.f32 %f176, %f175, %f174;
mov.b32 %r587, %f176;
mov.u32 %r588, 31;
mov.u32 %r589, 16;
mov.u32 %r590, -1;
shfl.sync.bfly.b32 %r591|%p53, %r587, %r589, %r588, %r590;
mov.b32 %f177, %r591;
add.ftz.f32 %f178, %f176, %f177;
mov.b32 %r592, %f178;
mov.u32 %r593, 8;
shfl.sync.bfly.b32 %r594|%p54, %r592, %r593, %r588, %r590;
mov.b32 %f179, %r594;
add.ftz.f32 %f180, %f178, %f179;
mov.b32 %r595, %f180;
mov.u32 %r596, 4;
shfl.sync.bfly.b32 %r597|%p55, %r595, %r596, %r588, %r590;
mov.b32 %f181, %r597;
add.ftz.f32 %f182, %f180, %f181;
mov.b32 %r598, %f182;
shfl.sync.bfly.b32 %r599|%p56, %r598, %r586, %r588, %r590;
mov.b32 %f183, %r599;
add.ftz.f32 %f184, %f182, %f183;
mov.b32 %r600, %f184;
mov.u32 %r601, 1;
shfl.sync.bfly.b32 %r602|%p57, %r600, %r601, %r588, %r590;
mov.b32 %f185, %r602;
add.ftz.f32 %f31, %f184, %f185;
@%p12 bra $L__BB8_20;

cvt.u64.u32 %rd70, %r18;
add.s64 %rd72, %rd70, %rd66;
shl.b64 %rd74, %rd72, 2;
add.s64 %rd75, %rd68, %rd74;
st.global.f32 [%rd75], %f31;

$L__BB8_20:
ld.shared.f32 %f186, [%r417+896];
add.ftz.f32 %f187, %f186, %f209;
ld.shared.f32 %f188, [%r417+1920];
add.ftz.f32 %f189, %f188, %f187;
ld.shared.f32 %f190, [%r417+2944];
add.ftz.f32 %f191, %f190, %f189;
mov.b32 %r616, %f191;
shfl.sync.bfly.b32 %r620|%p61, %r616, %r589, %r588, %r590;
mov.b32 %f192, %r620;
add.ftz.f32 %f193, %f191, %f192;
mov.b32 %r621, %f193;
shfl.sync.bfly.b32 %r623|%p62, %r621, %r593, %r588, %r590;
mov.b32 %f194, %r623;
add.ftz.f32 %f195, %f193, %f194;
mov.b32 %r624, %f195;
shfl.sync.bfly.b32 %r626|%p63, %r624, %r596, %r588, %r590;
mov.b32 %f196, %r626;
add.ftz.f32 %f197, %f195, %f196;
mov.b32 %r627, %f197;
shfl.sync.bfly.b32 %r628|%p64, %r627, %r586, %r588, %r590;
mov.b32 %f198, %r628;
add.ftz.f32 %f199, %f197, %f198;
mov.b32 %r629, %f199;
shfl.sync.bfly.b32 %r631|%p65, %r629, %r601, %r588, %r590;
mov.b32 %f200, %r631;
add.ftz.f32 %f32, %f199, %f200;
@%p20 bra $L__BB8_22;

add.s32 %r635, %r18, 1;
cvt.u64.u32 %rd76, %r635;
add.s64 %rd78, %rd76, %rd66;
shl.b64 %rd80, %rd78, 2;
add.s64 %rd81, %rd68, %rd80;
st.global.f32 [%rd81], %f32;

$L__BB8_22:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type41ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type41ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type41ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type41ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type41ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type41ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type41ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type41ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type41ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<78>;
.reg .b16 %rs<15>;
.reg .f32 %f<231>;
.reg .b32 %r<761>;
.reg .b64 %rd<96>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type41ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1280];

ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type41ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type41ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type41ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type41ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type41ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type41ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type41ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type41ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type41ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type41ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type41ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type41ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type41ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type41ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type41ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type41ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type41ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r37, %tid.y;
shl.b32 %r38, %r37, 5;
mov.u32 %r39, %tid.x;
add.s32 %r40, %r38, %r39;
shr.s32 %r41, %r40, 31;
shr.u32 %r42, %r41, 30;
add.s32 %r43, %r40, %r42;
shr.s32 %r760, %r43, 2;
shr.u32 %r44, %r21, 7;
setp.ge.s32 %p2, %r760, %r44;
mov.f32 %f221, 0f00000000;
mov.f32 %f222, %f221;
mov.f32 %f223, %f221;
mov.f32 %f224, %f221;
mov.f32 %f225, %f221;
mov.f32 %f226, %f221;
mov.f32 %f227, %f221;
mov.f32 %f228, %f221;
mov.f32 %f229, %f221;
mov.f32 %f230, %f221;
@%p2 bra $L__BB9_3;

mov.u32 %r45, %ctaid.y;
mov.u32 %r46, %ctaid.z;
mul.hi.u32 %r47, %r46, %r31;
add.s32 %r48, %r47, %r46;
shr.u32 %r49, %r48, %r32;
mul.lo.s32 %r50, %r49, %r34;
mul.hi.u32 %r51, %r45, %r25;
add.s32 %r52, %r51, %r45;
shr.u32 %r53, %r52, %r26;
mul.lo.s32 %r54, %r53, %r28;
mul.lo.s32 %r55, %r45, %r29;
cvt.u64.u32 %rd20, %r55;
mul.lo.s32 %r56, %r46, %r35;
cvt.u64.u32 %rd21, %r56;
and.b32 %r64, %r43, -4;
sub.s32 %r65, %r40, %r64;
cvt.s64.s32 %rd22, %r65;
add.s64 %rd23, %rd20, %rd22;
add.s64 %rd1, %rd23, %rd21;
shl.b32 %r66, %r65, 1;
mov.u32 %r67, %ctaid.x;
shl.b32 %r68, %r67, 1;
or.b32 %r69, %r68, 1;
shl.b32 %r70, %r23, 1;
shl.b32 %r755, %r760, 2;
add.s32 %r759, %r755, %r70;
mad.lo.s32 %r758, %r23, 3, %r755;
add.s32 %r71, %r760, %r23;
shl.b32 %r757, %r71, 2;
add.s32 %r756, %r23, %r755;
mad.lo.s32 %r72, %r68, %r22, %r760;
add.s32 %r73, %r72, %r50;
add.s32 %r74, %r73, %r54;
mul.wide.s32 %rd24, %r74, 18;
mul.wide.s32 %rd25, %r66, 2;
add.s64 %rd26, %rd24, %rd25;
cvta.to.global.u64 %rd27, %rd17;
add.s64 %rd28, %rd27, %rd26;
add.s64 %rd95, %rd28, 4;
add.s32 %r75, %r760, %r50;
add.s32 %r76, %r75, %r54;
mad.lo.s32 %r77, %r69, %r22, %r76;
mul.wide.s32 %rd29, %r77, 18;
add.s64 %rd30, %rd29, %rd25;
add.s64 %rd31, %rd27, %rd30;
add.s64 %rd94, %rd31, 4;
add.s64 %rd93, %rd27, %rd29;
add.s64 %rd92, %rd27, %rd24;
cvta.to.global.u64 %rd6, %rd18;
mov.f32 %f221, 0f00000000;

$L__BB9_2:
cvt.u64.u32 %rd32, %r755;
add.s64 %rd33, %rd1, %rd32;
mul.lo.s64 %rd34, %rd33, 36;
ld.global.nc.u16 %rs1, [%rd92];

{ cvt.f32.f16 %f61, %rs1;}


ld.global.nc.u16 %rs11, [%rd95+-2];
cvt.s32.s16 %r408, %rs11;
add.s64 %rd35, %rd6, %rd34;
ld.global.nc.u32 %r80, [%rd35+4];
ld.global.nc.u32 %r84, [%rd35+8];
ld.global.nc.u32 %r88, [%rd35+12];
ld.global.nc.u32 %r92, [%rd35+16];
and.b32 %r409, %r408, 30583;
mov.u32 %r410, 286261504;
prmt.b32 %r411, %r410, %r410, %r409;
shr.u32 %r412, %r408, 2;
and.b32 %r413, %r412, 30583;
prmt.b32 %r414, %r410, %r410, %r413;
and.b32 %r415, %r411, 30583;
mov.u32 %r416, 511;
prmt.b32 %r417, %r416, %r416, %r415;
and.b32 %r418, %r414, 30583;
prmt.b32 %r419, %r416, %r416, %r418;
shr.u32 %r420, %r411, 16;
and.b32 %r421, %r420, 30583;
prmt.b32 %r422, %r416, %r416, %r421;
shr.u32 %r423, %r414, 16;
and.b32 %r424, %r423, 30583;
prmt.b32 %r425, %r416, %r416, %r424;
mov.u32 %r426, 21520;
prmt.b32 %r343, %r417, %r419, %r426;
mov.u32 %r427, 30258;
prmt.b32 %r347, %r417, %r419, %r427;
prmt.b32 %r351, %r422, %r425, %r426;
prmt.b32 %r355, %r422, %r425, %r427;
mov.u32 %r378, 0;

dp4a.s32.s32 %r78, %r343, %r80, %r378;


dp4a.s32.s32 %r82, %r347, %r84, %r78;


dp4a.s32.s32 %r86, %r351, %r88, %r82;


dp4a.s32.s32 %r90, %r355, %r92, %r86;

ld.global.nc.u16 %rs12, [%rd95];
cvt.s32.s16 %r428, %rs12;
ld.global.nc.u32 %r96, [%rd35+20];
ld.global.nc.u32 %r100, [%rd35+24];
ld.global.nc.u32 %r104, [%rd35+28];
ld.global.nc.u32 %r108, [%rd35+32];
and.b32 %r429, %r428, 30583;
prmt.b32 %r430, %r410, %r410, %r429;
shr.u32 %r431, %r428, 2;
and.b32 %r432, %r431, 30583;
prmt.b32 %r433, %r410, %r410, %r432;
and.b32 %r434, %r430, 30583;
prmt.b32 %r435, %r416, %r416, %r434;
and.b32 %r436, %r433, 30583;
prmt.b32 %r437, %r416, %r416, %r436;
shr.u32 %r438, %r430, 16;
and.b32 %r439, %r438, 30583;
prmt.b32 %r440, %r416, %r416, %r439;
shr.u32 %r441, %r433, 16;
and.b32 %r442, %r441, 30583;
prmt.b32 %r443, %r416, %r416, %r442;
prmt.b32 %r359, %r435, %r437, %r426;
prmt.b32 %r363, %r435, %r437, %r427;
prmt.b32 %r367, %r440, %r443, %r426;
prmt.b32 %r371, %r440, %r443, %r427;

dp4a.s32.s32 %r94, %r359, %r96, %r90;


dp4a.s32.s32 %r98, %r363, %r100, %r94;


dp4a.s32.s32 %r102, %r367, %r104, %r98;


dp4a.s32.s32 %r106, %r371, %r108, %r102;

ld.global.nc.u32 %r110, [%rd35];

{.reg .f16 low,high;
mov.b32 {low,high},%r110;
cvt.f32.f16 %f62, low;}


mul.ftz.f32 %f81, %f61, %f62;
cvt.rn.f32.s32 %f82, %r106;
fma.rn.ftz.f32 %f230, %f81, %f82, %f230;
ld.global.nc.u16 %rs2, [%rd93];

{ cvt.f32.f16 %f63, %rs2;}


ld.global.nc.u16 %rs13, [%rd94+-2];
cvt.s32.s16 %r444, %rs13;
and.b32 %r445, %r444, 30583;
prmt.b32 %r446, %r410, %r410, %r445;
shr.u32 %r447, %r444, 2;
and.b32 %r448, %r447, 30583;
prmt.b32 %r449, %r410, %r410, %r448;
and.b32 %r450, %r446, 30583;
prmt.b32 %r451, %r416, %r416, %r450;
and.b32 %r452, %r449, 30583;
prmt.b32 %r453, %r416, %r416, %r452;
shr.u32 %r454, %r446, 16;
and.b32 %r455, %r454, 30583;
prmt.b32 %r456, %r416, %r416, %r455;
shr.u32 %r457, %r449, 16;
and.b32 %r458, %r457, 30583;
prmt.b32 %r459, %r416, %r416, %r458;
prmt.b32 %r376, %r451, %r453, %r426;
prmt.b32 %r380, %r451, %r453, %r427;
prmt.b32 %r384, %r456, %r459, %r426;
prmt.b32 %r388, %r456, %r459, %r427;

dp4a.s32.s32 %r111, %r376, %r80, %r378;


dp4a.s32.s32 %r115, %r380, %r84, %r111;


dp4a.s32.s32 %r119, %r384, %r88, %r115;


dp4a.s32.s32 %r123, %r388, %r92, %r119;

ld.global.nc.u16 %rs14, [%rd94];
cvt.s32.s16 %r460, %rs14;
and.b32 %r461, %r460, 30583;
prmt.b32 %r462, %r410, %r410, %r461;
shr.u32 %r463, %r460, 2;
and.b32 %r464, %r463, 30583;
prmt.b32 %r465, %r410, %r410, %r464;
and.b32 %r466, %r462, 30583;
prmt.b32 %r467, %r416, %r416, %r466;
and.b32 %r468, %r465, 30583;
prmt.b32 %r469, %r416, %r416, %r468;
shr.u32 %r470, %r462, 16;
and.b32 %r471, %r470, 30583;
prmt.b32 %r472, %r416, %r416, %r471;
shr.u32 %r473, %r465, 16;
and.b32 %r474, %r473, 30583;
prmt.b32 %r475, %r416, %r416, %r474;
prmt.b32 %r392, %r467, %r469, %r426;
prmt.b32 %r396, %r467, %r469, %r427;
prmt.b32 %r400, %r472, %r475, %r426;
prmt.b32 %r404, %r472, %r475, %r427;

dp4a.s32.s32 %r127, %r392, %r96, %r123;


dp4a.s32.s32 %r131, %r396, %r100, %r127;


dp4a.s32.s32 %r135, %r400, %r104, %r131;


dp4a.s32.s32 %r139, %r404, %r108, %r135;


{.reg .f16 low,high;
mov.b32 {low,high},%r110;
cvt.f32.f16 %f64, low;}


mul.ftz.f32 %f83, %f63, %f64;
cvt.rn.f32.s32 %f84, %r139;
fma.rn.ftz.f32 %f229, %f83, %f84, %f229;
cvt.u64.u32 %rd36, %r756;
add.s64 %rd37, %rd1, %rd36;
mul.lo.s64 %rd38, %rd37, 36;

{ cvt.f32.f16 %f65, %rs1;}


add.s64 %rd39, %rd6, %rd38;
ld.global.nc.u32 %r146, [%rd39+4];
ld.global.nc.u32 %r150, [%rd39+8];
ld.global.nc.u32 %r154, [%rd39+12];
ld.global.nc.u32 %r158, [%rd39+16];

dp4a.s32.s32 %r144, %r343, %r146, %r378;


dp4a.s32.s32 %r148, %r347, %r150, %r144;


dp4a.s32.s32 %r152, %r351, %r154, %r148;


dp4a.s32.s32 %r156, %r355, %r158, %r152;

ld.global.nc.u32 %r162, [%rd39+20];
ld.global.nc.u32 %r166, [%rd39+24];
ld.global.nc.u32 %r170, [%rd39+28];
ld.global.nc.u32 %r174, [%rd39+32];

dp4a.s32.s32 %r160, %r359, %r162, %r156;


dp4a.s32.s32 %r164, %r363, %r166, %r160;


dp4a.s32.s32 %r168, %r367, %r170, %r164;


dp4a.s32.s32 %r172, %r371, %r174, %r168;

ld.global.nc.u32 %r176, [%rd39];

{.reg .f16 low,high;
mov.b32 {low,high},%r176;
cvt.f32.f16 %f66, low;}


mul.ftz.f32 %f85, %f65, %f66;
cvt.rn.f32.s32 %f86, %r172;
fma.rn.ftz.f32 %f228, %f85, %f86, %f228;

{ cvt.f32.f16 %f67, %rs2;}



dp4a.s32.s32 %r177, %r376, %r146, %r378;


dp4a.s32.s32 %r181, %r380, %r150, %r177;


dp4a.s32.s32 %r185, %r384, %r154, %r181;


dp4a.s32.s32 %r189, %r388, %r158, %r185;


dp4a.s32.s32 %r193, %r392, %r162, %r189;


dp4a.s32.s32 %r197, %r396, %r166, %r193;


dp4a.s32.s32 %r201, %r400, %r170, %r197;


dp4a.s32.s32 %r205, %r404, %r174, %r201;


{.reg .f16 low,high;
mov.b32 {low,high},%r176;
cvt.f32.f16 %f68, low;}


mul.ftz.f32 %f87, %f67, %f68;
cvt.rn.f32.s32 %f88, %r205;
fma.rn.ftz.f32 %f227, %f87, %f88, %f227;
cvt.u64.u32 %rd40, %r759;
add.s64 %rd41, %rd1, %rd40;
mul.lo.s64 %rd42, %rd41, 36;

{ cvt.f32.f16 %f69, %rs1;}


add.s64 %rd43, %rd6, %rd42;
ld.global.nc.u32 %r212, [%rd43+4];
ld.global.nc.u32 %r216, [%rd43+8];
ld.global.nc.u32 %r220, [%rd43+12];
ld.global.nc.u32 %r224, [%rd43+16];

dp4a.s32.s32 %r210, %r343, %r212, %r378;


dp4a.s32.s32 %r214, %r347, %r216, %r210;


dp4a.s32.s32 %r218, %r351, %r220, %r214;


dp4a.s32.s32 %r222, %r355, %r224, %r218;

ld.global.nc.u32 %r228, [%rd43+20];
ld.global.nc.u32 %r232, [%rd43+24];
ld.global.nc.u32 %r236, [%rd43+28];
ld.global.nc.u32 %r240, [%rd43+32];

dp4a.s32.s32 %r226, %r359, %r228, %r222;


dp4a.s32.s32 %r230, %r363, %r232, %r226;


dp4a.s32.s32 %r234, %r367, %r236, %r230;


dp4a.s32.s32 %r238, %r371, %r240, %r234;

ld.global.nc.u32 %r242, [%rd43];

{.reg .f16 low,high;
mov.b32 {low,high},%r242;
cvt.f32.f16 %f70, low;}


mul.ftz.f32 %f89, %f69, %f70;
cvt.rn.f32.s32 %f90, %r238;
fma.rn.ftz.f32 %f226, %f89, %f90, %f226;

{ cvt.f32.f16 %f71, %rs2;}



dp4a.s32.s32 %r243, %r376, %r212, %r378;


dp4a.s32.s32 %r247, %r380, %r216, %r243;


dp4a.s32.s32 %r251, %r384, %r220, %r247;


dp4a.s32.s32 %r255, %r388, %r224, %r251;


dp4a.s32.s32 %r259, %r392, %r228, %r255;


dp4a.s32.s32 %r263, %r396, %r232, %r259;


dp4a.s32.s32 %r267, %r400, %r236, %r263;


dp4a.s32.s32 %r271, %r404, %r240, %r267;


{.reg .f16 low,high;
mov.b32 {low,high},%r242;
cvt.f32.f16 %f72, low;}


mul.ftz.f32 %f91, %f71, %f72;
cvt.rn.f32.s32 %f92, %r271;
fma.rn.ftz.f32 %f225, %f91, %f92, %f225;
cvt.u64.u32 %rd44, %r758;
add.s64 %rd45, %rd1, %rd44;
mul.lo.s64 %rd46, %rd45, 36;

{ cvt.f32.f16 %f73, %rs1;}


add.s64 %rd47, %rd6, %rd46;
ld.global.nc.u32 %r278, [%rd47+4];
ld.global.nc.u32 %r282, [%rd47+8];
ld.global.nc.u32 %r286, [%rd47+12];
ld.global.nc.u32 %r290, [%rd47+16];

dp4a.s32.s32 %r276, %r343, %r278, %r378;


dp4a.s32.s32 %r280, %r347, %r282, %r276;


dp4a.s32.s32 %r284, %r351, %r286, %r280;


dp4a.s32.s32 %r288, %r355, %r290, %r284;

ld.global.nc.u32 %r294, [%rd47+20];
ld.global.nc.u32 %r298, [%rd47+24];
ld.global.nc.u32 %r302, [%rd47+28];
ld.global.nc.u32 %r306, [%rd47+32];

dp4a.s32.s32 %r292, %r359, %r294, %r288;


dp4a.s32.s32 %r296, %r363, %r298, %r292;


dp4a.s32.s32 %r300, %r367, %r302, %r296;


dp4a.s32.s32 %r304, %r371, %r306, %r300;

ld.global.nc.u32 %r308, [%rd47];

{.reg .f16 low,high;
mov.b32 {low,high},%r308;
cvt.f32.f16 %f74, low;}


mul.ftz.f32 %f93, %f73, %f74;
cvt.rn.f32.s32 %f94, %r304;
fma.rn.ftz.f32 %f224, %f93, %f94, %f224;

{ cvt.f32.f16 %f75, %rs2;}



dp4a.s32.s32 %r309, %r376, %r278, %r378;


dp4a.s32.s32 %r313, %r380, %r282, %r309;


dp4a.s32.s32 %r317, %r384, %r286, %r313;


dp4a.s32.s32 %r321, %r388, %r290, %r317;


dp4a.s32.s32 %r325, %r392, %r294, %r321;


dp4a.s32.s32 %r329, %r396, %r298, %r325;


dp4a.s32.s32 %r333, %r400, %r302, %r329;


dp4a.s32.s32 %r337, %r404, %r306, %r333;


{.reg .f16 low,high;
mov.b32 {low,high},%r308;
cvt.f32.f16 %f76, low;}


mul.ftz.f32 %f95, %f75, %f76;
cvt.rn.f32.s32 %f96, %r337;
fma.rn.ftz.f32 %f223, %f95, %f96, %f223;
cvt.u64.u32 %rd48, %r757;
add.s64 %rd49, %rd1, %rd48;
mul.lo.s64 %rd50, %rd49, 36;

{ cvt.f32.f16 %f77, %rs1;}


add.s64 %rd51, %rd6, %rd50;
ld.global.nc.u32 %r344, [%rd51+4];
ld.global.nc.u32 %r348, [%rd51+8];
ld.global.nc.u32 %r352, [%rd51+12];
ld.global.nc.u32 %r356, [%rd51+16];

dp4a.s32.s32 %r342, %r343, %r344, %r378;


dp4a.s32.s32 %r346, %r347, %r348, %r342;


dp4a.s32.s32 %r350, %r351, %r352, %r346;


dp4a.s32.s32 %r354, %r355, %r356, %r350;

ld.global.nc.u32 %r360, [%rd51+20];
ld.global.nc.u32 %r364, [%rd51+24];
ld.global.nc.u32 %r368, [%rd51+28];
ld.global.nc.u32 %r372, [%rd51+32];

dp4a.s32.s32 %r358, %r359, %r360, %r354;


dp4a.s32.s32 %r362, %r363, %r364, %r358;


dp4a.s32.s32 %r366, %r367, %r368, %r362;


dp4a.s32.s32 %r370, %r371, %r372, %r366;

ld.global.nc.u32 %r374, [%rd51];

{.reg .f16 low,high;
mov.b32 {low,high},%r374;
cvt.f32.f16 %f78, low;}


mul.ftz.f32 %f97, %f77, %f78;
cvt.rn.f32.s32 %f98, %r370;
fma.rn.ftz.f32 %f222, %f97, %f98, %f222;

{ cvt.f32.f16 %f79, %rs2;}



dp4a.s32.s32 %r375, %r376, %r344, %r378;


dp4a.s32.s32 %r379, %r380, %r348, %r375;


dp4a.s32.s32 %r383, %r384, %r352, %r379;


dp4a.s32.s32 %r387, %r388, %r356, %r383;


dp4a.s32.s32 %r391, %r392, %r360, %r387;


dp4a.s32.s32 %r395, %r396, %r364, %r391;


dp4a.s32.s32 %r399, %r400, %r368, %r395;


dp4a.s32.s32 %r403, %r404, %r372, %r399;


{.reg .f16 low,high;
mov.b32 {low,high},%r374;
cvt.f32.f16 %f80, low;}


mul.ftz.f32 %f99, %f79, %f80;
cvt.rn.f32.s32 %f100, %r403;
fma.rn.ftz.f32 %f221, %f99, %f100, %f221;
add.s32 %r759, %r759, 64;
add.s32 %r758, %r758, 64;
add.s32 %r757, %r757, 64;
add.s32 %r756, %r756, 64;
add.s32 %r755, %r755, 64;
add.s64 %rd95, %rd95, 288;
add.s64 %rd94, %rd94, 288;
add.s64 %rd93, %rd93, 288;
add.s64 %rd92, %rd92, 288;
add.s32 %r760, %r760, 16;
setp.lt.s32 %p3, %r760, %r44;
@%p3 bra $L__BB9_2;

$L__BB9_3:
setp.eq.s32 %p4, %r37, 0;
@%p4 bra $L__BB9_5;

mov.u32 %r479, _ZZ13mul_mat_vec_qIL9ggml_type41ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r480, %r37, 1280, %r479;
shl.b32 %r481, %r39, 2;
add.s32 %r482, %r480, %r481;
st.shared.f32 [%r482+-1280], %f230;
st.shared.f32 [%r482+-1152], %f229;
st.shared.f32 [%r482+-1024], %f228;
st.shared.f32 [%r482+-896], %f227;
st.shared.f32 [%r482+-768], %f226;
st.shared.f32 [%r482+-640], %f225;
st.shared.f32 [%r482+-512], %f224;
st.shared.f32 [%r482+-384], %f223;
st.shared.f32 [%r482+-256], %f222;
st.shared.f32 [%r482+-128], %f221;

$L__BB9_5:
setp.ne.s32 %p5, %r37, 0;
bar.sync 0;
@%p5 bra $L__BB9_26;

mov.u32 %r484, %ctaid.x;
shl.b32 %r485, %r484, 1;
mov.u32 %r486, 1;
shl.b32 %r488, %r39, 2;
mov.u32 %r489, _ZZ13mul_mat_vec_qIL9ggml_type41ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r490, %r489, %r488;
mov.u32 %r491, 2;
ld.shared.f32 %f101, [%r490];
add.ftz.f32 %f102, %f101, %f230;
mov.b32 %r492, %f102;
mov.u32 %r493, 31;
mov.u32 %r494, 16;
mov.u32 %r495, -1;
shfl.sync.bfly.b32 %r496|%p6, %r492, %r494, %r493, %r495;
mov.b32 %f103, %r496;
add.ftz.f32 %f104, %f102, %f103;
mov.b32 %r497, %f104;
mov.u32 %r498, 8;
shfl.sync.bfly.b32 %r499|%p7, %r497, %r498, %r493, %r495;
mov.b32 %f105, %r499;
add.ftz.f32 %f106, %f104, %f105;
mov.b32 %r500, %f106;
mov.u32 %r501, 4;
shfl.sync.bfly.b32 %r502|%p8, %r500, %r501, %r493, %r495;
mov.b32 %f107, %r502;
add.ftz.f32 %f108, %f106, %f107;
mov.b32 %r503, %f108;
shfl.sync.bfly.b32 %r504|%p9, %r503, %r491, %r493, %r495;
mov.b32 %f109, %r504;
add.ftz.f32 %f110, %f108, %f109;
mov.b32 %r505, %f110;
shfl.sync.bfly.b32 %r506|%p10, %r505, %r486, %r493, %r495;
mov.b32 %f111, %r506;
add.ftz.f32 %f31, %f110, %f111;
setp.ne.s32 %p11, %r39, 0;
setp.ge.u32 %p12, %r485, %r24;
or.pred %p13, %p11, %p12;
@%p13 bra $L__BB9_8;

cvta.to.global.u64 %rd52, %rd19;
mov.u32 %r507, %ctaid.y;
mad.lo.s32 %r510, %r507, %r30, %r485;
mov.u32 %r511, %ctaid.z;
mad.lo.s32 %r512, %r511, %r36, %r510;
mul.wide.u32 %rd53, %r512, 4;
add.s64 %rd54, %rd52, %rd53;
st.global.f32 [%rd54], %f31;

$L__BB9_8:
ld.shared.f32 %f112, [%r490+128];
add.ftz.f32 %f113, %f112, %f229;
mov.b32 %r518, %f113;
shfl.sync.bfly.b32 %r522|%p14, %r518, %r494, %r493, %r495;
mov.b32 %f114, %r522;
add.ftz.f32 %f115, %f113, %f114;
mov.b32 %r523, %f115;
shfl.sync.bfly.b32 %r525|%p15, %r523, %r498, %r493, %r495;
mov.b32 %f116, %r525;
add.ftz.f32 %f117, %f115, %f116;
mov.b32 %r526, %f117;
shfl.sync.bfly.b32 %r528|%p16, %r526, %r501, %r493, %r495;
mov.b32 %f118, %r528;
add.ftz.f32 %f119, %f117, %f118;
mov.b32 %r529, %f119;
shfl.sync.bfly.b32 %r530|%p17, %r529, %r491, %r493, %r495;
mov.b32 %f120, %r530;
add.ftz.f32 %f121, %f119, %f120;
mov.b32 %r531, %f121;
shfl.sync.bfly.b32 %r533|%p18, %r531, %r486, %r493, %r495;
mov.b32 %f122, %r533;
add.ftz.f32 %f32, %f121, %f122;
setp.eq.s32 %p19, %r39, 1;
or.b32 %r536, %r485, 1;
setp.lt.u32 %p20, %r536, %r24;
and.pred %p1, %p19, %p20;
not.pred %p21, %p1;
@%p21 bra $L__BB9_10;

cvta.to.global.u64 %rd55, %rd19;
mov.u32 %r537, %ctaid.y;
mad.lo.s32 %r540, %r537, %r30, %r485;
mov.u32 %r541, %ctaid.z;
mad.lo.s32 %r542, %r541, %r36, %r540;
mul.wide.u32 %rd56, %r542, 4;
add.s64 %rd57, %rd55, %rd56;
st.global.f32 [%rd57+4], %f32;

$L__BB9_10:
mov.u32 %r547, 2;
ld.shared.f32 %f123, [%r490+256];
add.ftz.f32 %f124, %f123, %f228;
mov.b32 %r548, %f124;
mov.u32 %r549, 31;
mov.u32 %r550, 16;
mov.u32 %r551, -1;
shfl.sync.bfly.b32 %r552|%p22, %r548, %r550, %r549, %r551;
mov.b32 %f125, %r552;
add.ftz.f32 %f126, %f124, %f125;
mov.b32 %r553, %f126;
mov.u32 %r554, 8;
shfl.sync.bfly.b32 %r555|%p23, %r553, %r554, %r549, %r551;
mov.b32 %f127, %r555;
add.ftz.f32 %f128, %f126, %f127;
mov.b32 %r556, %f128;
mov.u32 %r557, 4;
shfl.sync.bfly.b32 %r558|%p24, %r556, %r557, %r549, %r551;
mov.b32 %f129, %r558;
add.ftz.f32 %f130, %f128, %f129;
mov.b32 %r559, %f130;
shfl.sync.bfly.b32 %r560|%p25, %r559, %r547, %r549, %r551;
mov.b32 %f131, %r560;
add.ftz.f32 %f132, %f130, %f131;
mov.b32 %r561, %f132;
mov.u32 %r562, 1;
shfl.sync.bfly.b32 %r563|%p26, %r561, %r562, %r549, %r551;
mov.b32 %f133, %r563;
add.ftz.f32 %f33, %f132, %f133;
@%p13 bra $L__BB9_12;

cvt.u64.u32 %rd58, %r24;
mov.u32 %r566, %ctaid.y;
mad.lo.s32 %r569, %r566, %r30, %r485;
mov.u32 %r570, %ctaid.z;
mad.lo.s32 %r571, %r570, %r36, %r569;
cvt.u64.u32 %rd59, %r571;
add.s64 %rd60, %rd58, %rd59;
cvta.to.global.u64 %rd61, %rd19;
shl.b64 %rd62, %rd60, 2;
add.s64 %rd63, %rd61, %rd62;
st.global.f32 [%rd63], %f33;

$L__BB9_12:
ld.shared.f32 %f134, [%r490+384];
add.ftz.f32 %f135, %f134, %f227;
mov.b32 %r577, %f135;
shfl.sync.bfly.b32 %r581|%p30, %r577, %r550, %r549, %r551;
mov.b32 %f136, %r581;
add.ftz.f32 %f137, %f135, %f136;
mov.b32 %r582, %f137;
shfl.sync.bfly.b32 %r584|%p31, %r582, %r554, %r549, %r551;
mov.b32 %f138, %r584;
add.ftz.f32 %f139, %f137, %f138;
mov.b32 %r585, %f139;
shfl.sync.bfly.b32 %r587|%p32, %r585, %r557, %r549, %r551;
mov.b32 %f140, %r587;
add.ftz.f32 %f141, %f139, %f140;
mov.b32 %r588, %f141;
shfl.sync.bfly.b32 %r589|%p33, %r588, %r547, %r549, %r551;
mov.b32 %f142, %r589;
add.ftz.f32 %f143, %f141, %f142;
mov.b32 %r590, %f143;
shfl.sync.bfly.b32 %r592|%p34, %r590, %r562, %r549, %r551;
mov.b32 %f144, %r592;
add.ftz.f32 %f34, %f143, %f144;
@%p21 bra $L__BB9_14;

add.s32 %r593, %r24, 1;
cvt.u64.u32 %rd64, %r593;
mov.u32 %r594, %ctaid.y;
mad.lo.s32 %r597, %r594, %r30, %r485;
mov.u32 %r598, %ctaid.z;
mad.lo.s32 %r599, %r598, %r36, %r597;
cvt.u64.u32 %rd65, %r599;
add.s64 %rd66, %rd64, %rd65;
cvta.to.global.u64 %rd67, %rd19;
shl.b64 %rd68, %rd66, 2;
add.s64 %rd69, %rd67, %rd68;
st.global.f32 [%rd69], %f34;

$L__BB9_14:
shl.b32 %r600, %r24, 1;
mov.u32 %r601, 1;
mov.u32 %r606, 2;
ld.shared.f32 %f145, [%r490+512];
add.ftz.f32 %f146, %f145, %f226;
mov.b32 %r607, %f146;
mov.u32 %r608, 31;
mov.u32 %r609, 16;
mov.u32 %r610, -1;
shfl.sync.bfly.b32 %r611|%p36, %r607, %r609, %r608, %r610;
mov.b32 %f147, %r611;
add.ftz.f32 %f148, %f146, %f147;
mov.b32 %r612, %f148;
mov.u32 %r613, 8;
shfl.sync.bfly.b32 %r614|%p37, %r612, %r613, %r608, %r610;
mov.b32 %f149, %r614;
add.ftz.f32 %f150, %f148, %f149;
mov.b32 %r615, %f150;
mov.u32 %r616, 4;
shfl.sync.bfly.b32 %r617|%p38, %r615, %r616, %r608, %r610;
mov.b32 %f151, %r617;
add.ftz.f32 %f152, %f150, %f151;
mov.b32 %r618, %f152;
shfl.sync.bfly.b32 %r619|%p39, %r618, %r606, %r608, %r610;
mov.b32 %f153, %r619;
add.ftz.f32 %f154, %f152, %f153;
mov.b32 %r620, %f154;
shfl.sync.bfly.b32 %r621|%p40, %r620, %r601, %r608, %r610;
mov.b32 %f155, %r621;
add.ftz.f32 %f35, %f154, %f155;
cvt.u64.u32 %rd70, %r600;
mov.u32 %r622, %ctaid.y;
mad.lo.s32 %r625, %r622, %r30, %r485;
mov.u32 %r626, %ctaid.z;
mad.lo.s32 %r627, %r626, %r36, %r625;
cvt.u64.u32 %rd71, %r627;
add.s64 %rd72, %rd70, %rd71;
cvta.to.global.u64 %rd73, %rd19;
shl.b64 %rd74, %rd72, 2;
add.s64 %rd15, %rd73, %rd74;
@%p13 bra $L__BB9_16;

st.global.f32 [%rd15], %f35;

$L__BB9_16:
ld.shared.f32 %f156, [%r490+640];
add.ftz.f32 %f157, %f156, %f225;
mov.b32 %r633, %f157;
shfl.sync.bfly.b32 %r637|%p44, %r633, %r609, %r608, %r610;
mov.b32 %f158, %r637;
add.ftz.f32 %f159, %f157, %f158;
mov.b32 %r638, %f159;
shfl.sync.bfly.b32 %r640|%p45, %r638, %r613, %r608, %r610;
mov.b32 %f160, %r640;
add.ftz.f32 %f161, %f159, %f160;
mov.b32 %r641, %f161;
shfl.sync.bfly.b32 %r643|%p46, %r641, %r616, %r608, %r610;
mov.b32 %f162, %r643;
add.ftz.f32 %f163, %f161, %f162;
mov.b32 %r644, %f163;
shfl.sync.bfly.b32 %r645|%p47, %r644, %r606, %r608, %r610;
mov.b32 %f164, %r645;
add.ftz.f32 %f165, %f163, %f164;
mov.b32 %r646, %f165;
shfl.sync.bfly.b32 %r648|%p48, %r646, %r601, %r608, %r610;
mov.b32 %f166, %r648;
add.ftz.f32 %f36, %f165, %f166;
@%p21 bra $L__BB9_18;

st.global.f32 [%rd15+4], %f36;

$L__BB9_18:
mul.lo.s32 %r20, %r24, 3;
mov.u32 %r653, 2;
ld.shared.f32 %f167, [%r490+768];
add.ftz.f32 %f168, %f167, %f224;
mov.b32 %r654, %f168;
mov.u32 %r655, 31;
mov.u32 %r656, 16;
mov.u32 %r657, -1;
shfl.sync.bfly.b32 %r658|%p50, %r654, %r656, %r655, %r657;
mov.b32 %f169, %r658;
add.ftz.f32 %f170, %f168, %f169;
mov.b32 %r659, %f170;
mov.u32 %r660, 8;
shfl.sync.bfly.b32 %r661|%p51, %r659, %r660, %r655, %r657;
mov.b32 %f171, %r661;
add.ftz.f32 %f172, %f170, %f171;
mov.b32 %r662, %f172;
mov.u32 %r663, 4;
shfl.sync.bfly.b32 %r664|%p52, %r662, %r663, %r655, %r657;
mov.b32 %f173, %r664;
add.ftz.f32 %f174, %f172, %f173;
mov.b32 %r665, %f174;
shfl.sync.bfly.b32 %r666|%p53, %r665, %r653, %r655, %r657;
mov.b32 %f175, %r666;
add.ftz.f32 %f176, %f174, %f175;
mov.b32 %r667, %f176;
mov.u32 %r668, 1;
shfl.sync.bfly.b32 %r669|%p54, %r667, %r668, %r655, %r657;
mov.b32 %f177, %r669;
add.ftz.f32 %f37, %f176, %f177;
@%p13 bra $L__BB9_20;

cvt.u64.u32 %rd75, %r20;
add.s64 %rd77, %rd75, %rd71;
shl.b64 %rd79, %rd77, 2;
add.s64 %rd80, %rd73, %rd79;
st.global.f32 [%rd80], %f37;

$L__BB9_20:
ld.shared.f32 %f178, [%r490+896];
add.ftz.f32 %f179, %f178, %f223;
mov.b32 %r683, %f179;
shfl.sync.bfly.b32 %r687|%p58, %r683, %r656, %r655, %r657;
mov.b32 %f180, %r687;
add.ftz.f32 %f181, %f179, %f180;
mov.b32 %r688, %f181;
shfl.sync.bfly.b32 %r690|%p59, %r688, %r660, %r655, %r657;
mov.b32 %f182, %r690;
add.ftz.f32 %f183, %f181, %f182;
mov.b32 %r691, %f183;
shfl.sync.bfly.b32 %r693|%p60, %r691, %r663, %r655, %r657;
mov.b32 %f184, %r693;
add.ftz.f32 %f185, %f183, %f184;
mov.b32 %r694, %f185;
shfl.sync.bfly.b32 %r695|%p61, %r694, %r653, %r655, %r657;
mov.b32 %f186, %r695;
add.ftz.f32 %f187, %f185, %f186;
mov.b32 %r696, %f187;
shfl.sync.bfly.b32 %r698|%p62, %r696, %r668, %r655, %r657;
mov.b32 %f188, %r698;
add.ftz.f32 %f38, %f187, %f188;
@%p21 bra $L__BB9_22;

add.s32 %r699, %r20, 1;
cvt.u64.u32 %rd81, %r699;
add.s64 %rd83, %rd81, %rd71;
shl.b64 %rd85, %rd83, 2;
add.s64 %rd86, %rd73, %rd85;
st.global.f32 [%rd86], %f38;

$L__BB9_22:
shl.b32 %r706, %r24, 2;
mov.u32 %r707, 2;
ld.shared.f32 %f189, [%r490+1024];
add.ftz.f32 %f190, %f189, %f222;
mov.b32 %r712, %f190;
mov.u32 %r713, 31;
mov.u32 %r714, 16;
mov.u32 %r715, -1;
shfl.sync.bfly.b32 %r716|%p64, %r712, %r714, %r713, %r715;
mov.b32 %f191, %r716;
add.ftz.f32 %f192, %f190, %f191;
mov.b32 %r717, %f192;
mov.u32 %r718, 8;
shfl.sync.bfly.b32 %r719|%p65, %r717, %r718, %r713, %r715;
mov.b32 %f193, %r719;
add.ftz.f32 %f194, %f192, %f193;
mov.b32 %r720, %f194;
mov.u32 %r721, 4;
shfl.sync.bfly.b32 %r722|%p66, %r720, %r721, %r713, %r715;
mov.b32 %f195, %r722;
add.ftz.f32 %f196, %f194, %f195;
mov.b32 %r723, %f196;
shfl.sync.bfly.b32 %r724|%p67, %r723, %r707, %r713, %r715;
mov.b32 %f197, %r724;
add.ftz.f32 %f198, %f196, %f197;
mov.b32 %r725, %f198;
mov.u32 %r726, 1;
shfl.sync.bfly.b32 %r727|%p68, %r725, %r726, %r713, %r715;
mov.b32 %f199, %r727;
add.ftz.f32 %f39, %f198, %f199;
cvt.u64.u32 %rd87, %r706;
add.s64 %rd89, %rd87, %rd71;
shl.b64 %rd91, %rd89, 2;
add.s64 %rd16, %rd73, %rd91;
@%p13 bra $L__BB9_24;

st.global.f32 [%rd16], %f39;

$L__BB9_24:
ld.shared.f32 %f200, [%r490+1152];
add.ftz.f32 %f201, %f200, %f221;
mov.b32 %r739, %f201;
shfl.sync.bfly.b32 %r743|%p72, %r739, %r714, %r713, %r715;
mov.b32 %f202, %r743;
add.ftz.f32 %f203, %f201, %f202;
mov.b32 %r744, %f203;
shfl.sync.bfly.b32 %r746|%p73, %r744, %r718, %r713, %r715;
mov.b32 %f204, %r746;
add.ftz.f32 %f205, %f203, %f204;
mov.b32 %r747, %f205;
shfl.sync.bfly.b32 %r749|%p74, %r747, %r721, %r713, %r715;
mov.b32 %f206, %r749;
add.ftz.f32 %f207, %f205, %f206;
mov.b32 %r750, %f207;
shfl.sync.bfly.b32 %r751|%p75, %r750, %r707, %r713, %r715;
mov.b32 %f208, %r751;
add.ftz.f32 %f209, %f207, %f208;
mov.b32 %r752, %f209;
shfl.sync.bfly.b32 %r754|%p76, %r752, %r726, %r713, %r715;
mov.b32 %f210, %r754;
add.ftz.f32 %f40, %f209, %f210;
@%p21 bra $L__BB9_26;

st.global.f32 [%rd16+4], %f40;

$L__BB9_26:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type41ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type41ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type41ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type41ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type41ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type41ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type41ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type41ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type41ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<92>;
.reg .b16 %rs<17>;
.reg .f32 %f<277>;
.reg .b32 %r<889>;
.reg .b64 %rd<112>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type41ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type41ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type41ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type41ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type41ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type41ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type41ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type41ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type41ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type41ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type41ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type41ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type41ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type41ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type41ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type41ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type41ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type41ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r41, %tid.y;
shl.b32 %r42, %r41, 5;
mov.u32 %r43, %tid.x;
add.s32 %r44, %r42, %r43;
shr.s32 %r45, %r44, 31;
shr.u32 %r46, %r45, 30;
add.s32 %r47, %r44, %r46;
shr.s32 %r888, %r47, 2;
shr.u32 %r48, %r25, 7;
setp.ge.s32 %p2, %r888, %r48;
mov.f32 %f265, 0f00000000;
mov.f32 %f266, %f265;
mov.f32 %f267, %f265;
mov.f32 %f268, %f265;
mov.f32 %f269, %f265;
mov.f32 %f270, %f265;
mov.f32 %f271, %f265;
mov.f32 %f272, %f265;
mov.f32 %f273, %f265;
mov.f32 %f274, %f265;
mov.f32 %f275, %f265;
mov.f32 %f276, %f265;
@%p2 bra $L__BB10_3;

mov.u32 %r49, %ctaid.y;
mov.u32 %r50, %ctaid.z;
mul.hi.u32 %r51, %r50, %r35;
add.s32 %r52, %r51, %r50;
shr.u32 %r53, %r52, %r36;
mul.lo.s32 %r54, %r53, %r38;
mul.hi.u32 %r55, %r49, %r29;
add.s32 %r56, %r55, %r49;
shr.u32 %r57, %r56, %r30;
mul.lo.s32 %r58, %r57, %r32;
mul.lo.s32 %r59, %r49, %r33;
cvt.u64.u32 %rd20, %r59;
mul.lo.s32 %r60, %r50, %r39;
cvt.u64.u32 %rd21, %r60;
and.b32 %r68, %r47, -4;
sub.s32 %r69, %r44, %r68;
mov.u32 %r70, %ctaid.x;
shl.b32 %r71, %r70, 1;
or.b32 %r72, %r71, 1;
shl.b32 %r73, %r69, 1;
cvt.s64.s32 %rd22, %r69;
add.s64 %rd23, %rd20, %rd22;
add.s64 %rd1, %rd23, %rd21;
shl.b32 %r74, %r27, 1;
shl.b32 %r882, %r888, 2;
add.s32 %r887, %r882, %r74;
mad.lo.s32 %r886, %r27, 3, %r882;
add.s32 %r75, %r888, %r27;
shl.b32 %r885, %r75, 2;
mad.lo.s32 %r884, %r27, 5, %r882;
add.s32 %r883, %r27, %r882;
mad.lo.s32 %r76, %r71, %r26, %r888;
add.s32 %r77, %r76, %r54;
add.s32 %r78, %r77, %r58;
mul.wide.s32 %rd24, %r78, 18;
mul.wide.s32 %rd25, %r73, 2;
add.s64 %rd26, %rd24, %rd25;
cvta.to.global.u64 %rd27, %rd17;
add.s64 %rd28, %rd27, %rd26;
add.s64 %rd111, %rd28, 4;
add.s32 %r79, %r888, %r54;
add.s32 %r80, %r79, %r58;
mad.lo.s32 %r81, %r72, %r26, %r80;
mul.wide.s32 %rd29, %r81, 18;
add.s64 %rd30, %rd29, %rd25;
add.s64 %rd31, %rd27, %rd30;
add.s64 %rd110, %rd31, 4;
add.s64 %rd109, %rd27, %rd29;
add.s64 %rd108, %rd27, %rd24;
cvta.to.global.u64 %rd6, %rd18;
mov.f32 %f265, 0f00000000;

$L__BB10_2:
cvt.u64.u32 %rd32, %r882;
add.s64 %rd33, %rd1, %rd32;
mul.lo.s64 %rd34, %rd33, 36;
ld.global.nc.u16 %rs1, [%rd108];

{ cvt.f32.f16 %f73, %rs1;}


ld.global.nc.u16 %rs13, [%rd111+-2];
cvt.s32.s16 %r478, %rs13;
add.s64 %rd35, %rd6, %rd34;
ld.global.nc.u32 %r84, [%rd35+4];
ld.global.nc.u32 %r88, [%rd35+8];
ld.global.nc.u32 %r92, [%rd35+12];
ld.global.nc.u32 %r96, [%rd35+16];
and.b32 %r479, %r478, 30583;
mov.u32 %r480, 286261504;
prmt.b32 %r481, %r480, %r480, %r479;
shr.u32 %r482, %r478, 2;
and.b32 %r483, %r482, 30583;
prmt.b32 %r484, %r480, %r480, %r483;
and.b32 %r485, %r481, 30583;
mov.u32 %r486, 511;
prmt.b32 %r487, %r486, %r486, %r485;
and.b32 %r488, %r484, 30583;
prmt.b32 %r489, %r486, %r486, %r488;
shr.u32 %r490, %r481, 16;
and.b32 %r491, %r490, 30583;
prmt.b32 %r492, %r486, %r486, %r491;
shr.u32 %r493, %r484, 16;
and.b32 %r494, %r493, 30583;
prmt.b32 %r495, %r486, %r486, %r494;
mov.u32 %r496, 21520;
prmt.b32 %r413, %r487, %r489, %r496;
mov.u32 %r497, 30258;
prmt.b32 %r417, %r487, %r489, %r497;
prmt.b32 %r421, %r492, %r495, %r496;
prmt.b32 %r425, %r492, %r495, %r497;
mov.u32 %r448, 0;

dp4a.s32.s32 %r82, %r413, %r84, %r448;


dp4a.s32.s32 %r86, %r417, %r88, %r82;


dp4a.s32.s32 %r90, %r421, %r92, %r86;


dp4a.s32.s32 %r94, %r425, %r96, %r90;

ld.global.nc.u16 %rs14, [%rd111];
cvt.s32.s16 %r498, %rs14;
ld.global.nc.u32 %r100, [%rd35+20];
ld.global.nc.u32 %r104, [%rd35+24];
ld.global.nc.u32 %r108, [%rd35+28];
ld.global.nc.u32 %r112, [%rd35+32];
and.b32 %r499, %r498, 30583;
prmt.b32 %r500, %r480, %r480, %r499;
shr.u32 %r501, %r498, 2;
and.b32 %r502, %r501, 30583;
prmt.b32 %r503, %r480, %r480, %r502;
and.b32 %r504, %r500, 30583;
prmt.b32 %r505, %r486, %r486, %r504;
and.b32 %r506, %r503, 30583;
prmt.b32 %r507, %r486, %r486, %r506;
shr.u32 %r508, %r500, 16;
and.b32 %r509, %r508, 30583;
prmt.b32 %r510, %r486, %r486, %r509;
shr.u32 %r511, %r503, 16;
and.b32 %r512, %r511, 30583;
prmt.b32 %r513, %r486, %r486, %r512;
prmt.b32 %r429, %r505, %r507, %r496;
prmt.b32 %r433, %r505, %r507, %r497;
prmt.b32 %r437, %r510, %r513, %r496;
prmt.b32 %r441, %r510, %r513, %r497;

dp4a.s32.s32 %r98, %r429, %r100, %r94;


dp4a.s32.s32 %r102, %r433, %r104, %r98;


dp4a.s32.s32 %r106, %r437, %r108, %r102;


dp4a.s32.s32 %r110, %r441, %r112, %r106;

ld.global.nc.u32 %r114, [%rd35];

{.reg .f16 low,high;
mov.b32 {low,high},%r114;
cvt.f32.f16 %f74, low;}


mul.ftz.f32 %f97, %f73, %f74;
cvt.rn.f32.s32 %f98, %r110;
fma.rn.ftz.f32 %f276, %f97, %f98, %f276;
ld.global.nc.u16 %rs2, [%rd109];

{ cvt.f32.f16 %f75, %rs2;}


ld.global.nc.u16 %rs15, [%rd110+-2];
cvt.s32.s16 %r514, %rs15;
and.b32 %r515, %r514, 30583;
prmt.b32 %r516, %r480, %r480, %r515;
shr.u32 %r517, %r514, 2;
and.b32 %r518, %r517, 30583;
prmt.b32 %r519, %r480, %r480, %r518;
and.b32 %r520, %r516, 30583;
prmt.b32 %r521, %r486, %r486, %r520;
and.b32 %r522, %r519, 30583;
prmt.b32 %r523, %r486, %r486, %r522;
shr.u32 %r524, %r516, 16;
and.b32 %r525, %r524, 30583;
prmt.b32 %r526, %r486, %r486, %r525;
shr.u32 %r527, %r519, 16;
and.b32 %r528, %r527, 30583;
prmt.b32 %r529, %r486, %r486, %r528;
prmt.b32 %r446, %r521, %r523, %r496;
prmt.b32 %r450, %r521, %r523, %r497;
prmt.b32 %r454, %r526, %r529, %r496;
prmt.b32 %r458, %r526, %r529, %r497;

dp4a.s32.s32 %r115, %r446, %r84, %r448;


dp4a.s32.s32 %r119, %r450, %r88, %r115;


dp4a.s32.s32 %r123, %r454, %r92, %r119;


dp4a.s32.s32 %r127, %r458, %r96, %r123;

ld.global.nc.u16 %rs16, [%rd110];
cvt.s32.s16 %r530, %rs16;
and.b32 %r531, %r530, 30583;
prmt.b32 %r532, %r480, %r480, %r531;
shr.u32 %r533, %r530, 2;
and.b32 %r534, %r533, 30583;
prmt.b32 %r535, %r480, %r480, %r534;
and.b32 %r536, %r532, 30583;
prmt.b32 %r537, %r486, %r486, %r536;
and.b32 %r538, %r535, 30583;
prmt.b32 %r539, %r486, %r486, %r538;
shr.u32 %r540, %r532, 16;
and.b32 %r541, %r540, 30583;
prmt.b32 %r542, %r486, %r486, %r541;
shr.u32 %r543, %r535, 16;
and.b32 %r544, %r543, 30583;
prmt.b32 %r545, %r486, %r486, %r544;
prmt.b32 %r462, %r537, %r539, %r496;
prmt.b32 %r466, %r537, %r539, %r497;
prmt.b32 %r470, %r542, %r545, %r496;
prmt.b32 %r474, %r542, %r545, %r497;

dp4a.s32.s32 %r131, %r462, %r100, %r127;


dp4a.s32.s32 %r135, %r466, %r104, %r131;


dp4a.s32.s32 %r139, %r470, %r108, %r135;


dp4a.s32.s32 %r143, %r474, %r112, %r139;


{.reg .f16 low,high;
mov.b32 {low,high},%r114;
cvt.f32.f16 %f76, low;}


mul.ftz.f32 %f99, %f75, %f76;
cvt.rn.f32.s32 %f100, %r143;
fma.rn.ftz.f32 %f275, %f99, %f100, %f275;
cvt.u64.u32 %rd36, %r883;
add.s64 %rd37, %rd1, %rd36;
mul.lo.s64 %rd38, %rd37, 36;

{ cvt.f32.f16 %f77, %rs1;}


add.s64 %rd39, %rd6, %rd38;
ld.global.nc.u32 %r150, [%rd39+4];
ld.global.nc.u32 %r154, [%rd39+8];
ld.global.nc.u32 %r158, [%rd39+12];
ld.global.nc.u32 %r162, [%rd39+16];

dp4a.s32.s32 %r148, %r413, %r150, %r448;


dp4a.s32.s32 %r152, %r417, %r154, %r148;


dp4a.s32.s32 %r156, %r421, %r158, %r152;


dp4a.s32.s32 %r160, %r425, %r162, %r156;

ld.global.nc.u32 %r166, [%rd39+20];
ld.global.nc.u32 %r170, [%rd39+24];
ld.global.nc.u32 %r174, [%rd39+28];
ld.global.nc.u32 %r178, [%rd39+32];

dp4a.s32.s32 %r164, %r429, %r166, %r160;


dp4a.s32.s32 %r168, %r433, %r170, %r164;


dp4a.s32.s32 %r172, %r437, %r174, %r168;


dp4a.s32.s32 %r176, %r441, %r178, %r172;

ld.global.nc.u32 %r180, [%rd39];

{.reg .f16 low,high;
mov.b32 {low,high},%r180;
cvt.f32.f16 %f78, low;}


mul.ftz.f32 %f101, %f77, %f78;
cvt.rn.f32.s32 %f102, %r176;
fma.rn.ftz.f32 %f274, %f101, %f102, %f274;

{ cvt.f32.f16 %f79, %rs2;}



dp4a.s32.s32 %r181, %r446, %r150, %r448;


dp4a.s32.s32 %r185, %r450, %r154, %r181;


dp4a.s32.s32 %r189, %r454, %r158, %r185;


dp4a.s32.s32 %r193, %r458, %r162, %r189;


dp4a.s32.s32 %r197, %r462, %r166, %r193;


dp4a.s32.s32 %r201, %r466, %r170, %r197;


dp4a.s32.s32 %r205, %r470, %r174, %r201;


dp4a.s32.s32 %r209, %r474, %r178, %r205;


{.reg .f16 low,high;
mov.b32 {low,high},%r180;
cvt.f32.f16 %f80, low;}


mul.ftz.f32 %f103, %f79, %f80;
cvt.rn.f32.s32 %f104, %r209;
fma.rn.ftz.f32 %f273, %f103, %f104, %f273;
cvt.u64.u32 %rd40, %r887;
add.s64 %rd41, %rd1, %rd40;
mul.lo.s64 %rd42, %rd41, 36;

{ cvt.f32.f16 %f81, %rs1;}


add.s64 %rd43, %rd6, %rd42;
ld.global.nc.u32 %r216, [%rd43+4];
ld.global.nc.u32 %r220, [%rd43+8];
ld.global.nc.u32 %r224, [%rd43+12];
ld.global.nc.u32 %r228, [%rd43+16];

dp4a.s32.s32 %r214, %r413, %r216, %r448;


dp4a.s32.s32 %r218, %r417, %r220, %r214;


dp4a.s32.s32 %r222, %r421, %r224, %r218;


dp4a.s32.s32 %r226, %r425, %r228, %r222;

ld.global.nc.u32 %r232, [%rd43+20];
ld.global.nc.u32 %r236, [%rd43+24];
ld.global.nc.u32 %r240, [%rd43+28];
ld.global.nc.u32 %r244, [%rd43+32];

dp4a.s32.s32 %r230, %r429, %r232, %r226;


dp4a.s32.s32 %r234, %r433, %r236, %r230;


dp4a.s32.s32 %r238, %r437, %r240, %r234;


dp4a.s32.s32 %r242, %r441, %r244, %r238;

ld.global.nc.u32 %r246, [%rd43];

{.reg .f16 low,high;
mov.b32 {low,high},%r246;
cvt.f32.f16 %f82, low;}


mul.ftz.f32 %f105, %f81, %f82;
cvt.rn.f32.s32 %f106, %r242;
fma.rn.ftz.f32 %f272, %f105, %f106, %f272;

{ cvt.f32.f16 %f83, %rs2;}



dp4a.s32.s32 %r247, %r446, %r216, %r448;


dp4a.s32.s32 %r251, %r450, %r220, %r247;


dp4a.s32.s32 %r255, %r454, %r224, %r251;


dp4a.s32.s32 %r259, %r458, %r228, %r255;


dp4a.s32.s32 %r263, %r462, %r232, %r259;


dp4a.s32.s32 %r267, %r466, %r236, %r263;


dp4a.s32.s32 %r271, %r470, %r240, %r267;


dp4a.s32.s32 %r275, %r474, %r244, %r271;


{.reg .f16 low,high;
mov.b32 {low,high},%r246;
cvt.f32.f16 %f84, low;}


mul.ftz.f32 %f107, %f83, %f84;
cvt.rn.f32.s32 %f108, %r275;
fma.rn.ftz.f32 %f271, %f107, %f108, %f271;
cvt.u64.u32 %rd44, %r886;
add.s64 %rd45, %rd1, %rd44;
mul.lo.s64 %rd46, %rd45, 36;

{ cvt.f32.f16 %f85, %rs1;}


add.s64 %rd47, %rd6, %rd46;
ld.global.nc.u32 %r282, [%rd47+4];
ld.global.nc.u32 %r286, [%rd47+8];
ld.global.nc.u32 %r290, [%rd47+12];
ld.global.nc.u32 %r294, [%rd47+16];

dp4a.s32.s32 %r280, %r413, %r282, %r448;


dp4a.s32.s32 %r284, %r417, %r286, %r280;


dp4a.s32.s32 %r288, %r421, %r290, %r284;


dp4a.s32.s32 %r292, %r425, %r294, %r288;

ld.global.nc.u32 %r298, [%rd47+20];
ld.global.nc.u32 %r302, [%rd47+24];
ld.global.nc.u32 %r306, [%rd47+28];
ld.global.nc.u32 %r310, [%rd47+32];

dp4a.s32.s32 %r296, %r429, %r298, %r292;


dp4a.s32.s32 %r300, %r433, %r302, %r296;


dp4a.s32.s32 %r304, %r437, %r306, %r300;


dp4a.s32.s32 %r308, %r441, %r310, %r304;

ld.global.nc.u32 %r312, [%rd47];

{.reg .f16 low,high;
mov.b32 {low,high},%r312;
cvt.f32.f16 %f86, low;}


mul.ftz.f32 %f109, %f85, %f86;
cvt.rn.f32.s32 %f110, %r308;
fma.rn.ftz.f32 %f270, %f109, %f110, %f270;

{ cvt.f32.f16 %f87, %rs2;}



dp4a.s32.s32 %r313, %r446, %r282, %r448;


dp4a.s32.s32 %r317, %r450, %r286, %r313;


dp4a.s32.s32 %r321, %r454, %r290, %r317;


dp4a.s32.s32 %r325, %r458, %r294, %r321;


dp4a.s32.s32 %r329, %r462, %r298, %r325;


dp4a.s32.s32 %r333, %r466, %r302, %r329;


dp4a.s32.s32 %r337, %r470, %r306, %r333;


dp4a.s32.s32 %r341, %r474, %r310, %r337;


{.reg .f16 low,high;
mov.b32 {low,high},%r312;
cvt.f32.f16 %f88, low;}


mul.ftz.f32 %f111, %f87, %f88;
cvt.rn.f32.s32 %f112, %r341;
fma.rn.ftz.f32 %f269, %f111, %f112, %f269;
cvt.u64.u32 %rd48, %r885;
add.s64 %rd49, %rd1, %rd48;
mul.lo.s64 %rd50, %rd49, 36;

{ cvt.f32.f16 %f89, %rs1;}


add.s64 %rd51, %rd6, %rd50;
ld.global.nc.u32 %r348, [%rd51+4];
ld.global.nc.u32 %r352, [%rd51+8];
ld.global.nc.u32 %r356, [%rd51+12];
ld.global.nc.u32 %r360, [%rd51+16];

dp4a.s32.s32 %r346, %r413, %r348, %r448;


dp4a.s32.s32 %r350, %r417, %r352, %r346;


dp4a.s32.s32 %r354, %r421, %r356, %r350;


dp4a.s32.s32 %r358, %r425, %r360, %r354;

ld.global.nc.u32 %r364, [%rd51+20];
ld.global.nc.u32 %r368, [%rd51+24];
ld.global.nc.u32 %r372, [%rd51+28];
ld.global.nc.u32 %r376, [%rd51+32];

dp4a.s32.s32 %r362, %r429, %r364, %r358;


dp4a.s32.s32 %r366, %r433, %r368, %r362;


dp4a.s32.s32 %r370, %r437, %r372, %r366;


dp4a.s32.s32 %r374, %r441, %r376, %r370;

ld.global.nc.u32 %r378, [%rd51];

{.reg .f16 low,high;
mov.b32 {low,high},%r378;
cvt.f32.f16 %f90, low;}


mul.ftz.f32 %f113, %f89, %f90;
cvt.rn.f32.s32 %f114, %r374;
fma.rn.ftz.f32 %f268, %f113, %f114, %f268;

{ cvt.f32.f16 %f91, %rs2;}



dp4a.s32.s32 %r379, %r446, %r348, %r448;


dp4a.s32.s32 %r383, %r450, %r352, %r379;


dp4a.s32.s32 %r387, %r454, %r356, %r383;


dp4a.s32.s32 %r391, %r458, %r360, %r387;


dp4a.s32.s32 %r395, %r462, %r364, %r391;


dp4a.s32.s32 %r399, %r466, %r368, %r395;


dp4a.s32.s32 %r403, %r470, %r372, %r399;


dp4a.s32.s32 %r407, %r474, %r376, %r403;


{.reg .f16 low,high;
mov.b32 {low,high},%r378;
cvt.f32.f16 %f92, low;}


mul.ftz.f32 %f115, %f91, %f92;
cvt.rn.f32.s32 %f116, %r407;
fma.rn.ftz.f32 %f267, %f115, %f116, %f267;
cvt.u64.u32 %rd52, %r884;
add.s64 %rd53, %rd1, %rd52;
mul.lo.s64 %rd54, %rd53, 36;

{ cvt.f32.f16 %f93, %rs1;}


add.s64 %rd55, %rd6, %rd54;
ld.global.nc.u32 %r414, [%rd55+4];
ld.global.nc.u32 %r418, [%rd55+8];
ld.global.nc.u32 %r422, [%rd55+12];
ld.global.nc.u32 %r426, [%rd55+16];

dp4a.s32.s32 %r412, %r413, %r414, %r448;


dp4a.s32.s32 %r416, %r417, %r418, %r412;


dp4a.s32.s32 %r420, %r421, %r422, %r416;


dp4a.s32.s32 %r424, %r425, %r426, %r420;

ld.global.nc.u32 %r430, [%rd55+20];
ld.global.nc.u32 %r434, [%rd55+24];
ld.global.nc.u32 %r438, [%rd55+28];
ld.global.nc.u32 %r442, [%rd55+32];

dp4a.s32.s32 %r428, %r429, %r430, %r424;


dp4a.s32.s32 %r432, %r433, %r434, %r428;


dp4a.s32.s32 %r436, %r437, %r438, %r432;


dp4a.s32.s32 %r440, %r441, %r442, %r436;

ld.global.nc.u32 %r444, [%rd55];

{.reg .f16 low,high;
mov.b32 {low,high},%r444;
cvt.f32.f16 %f94, low;}


mul.ftz.f32 %f117, %f93, %f94;
cvt.rn.f32.s32 %f118, %r440;
fma.rn.ftz.f32 %f266, %f117, %f118, %f266;

{ cvt.f32.f16 %f95, %rs2;}



dp4a.s32.s32 %r445, %r446, %r414, %r448;


dp4a.s32.s32 %r449, %r450, %r418, %r445;


dp4a.s32.s32 %r453, %r454, %r422, %r449;


dp4a.s32.s32 %r457, %r458, %r426, %r453;


dp4a.s32.s32 %r461, %r462, %r430, %r457;


dp4a.s32.s32 %r465, %r466, %r434, %r461;


dp4a.s32.s32 %r469, %r470, %r438, %r465;


dp4a.s32.s32 %r473, %r474, %r442, %r469;


{.reg .f16 low,high;
mov.b32 {low,high},%r444;
cvt.f32.f16 %f96, low;}


mul.ftz.f32 %f119, %f95, %f96;
cvt.rn.f32.s32 %f120, %r473;
fma.rn.ftz.f32 %f265, %f119, %f120, %f265;
add.s32 %r887, %r887, 64;
add.s32 %r886, %r886, 64;
add.s32 %r885, %r885, 64;
add.s32 %r884, %r884, 64;
add.s32 %r883, %r883, 64;
add.s32 %r882, %r882, 64;
add.s64 %rd111, %rd111, 288;
add.s64 %rd110, %rd110, 288;
add.s64 %rd109, %rd109, 288;
add.s64 %rd108, %rd108, 288;
add.s32 %r888, %r888, 16;
setp.lt.s32 %p3, %r888, %r48;
@%p3 bra $L__BB10_2;

$L__BB10_3:
setp.eq.s32 %p4, %r41, 0;
@%p4 bra $L__BB10_5;

mov.u32 %r549, _ZZ13mul_mat_vec_qIL9ggml_type41ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r550, %r41, 1536, %r549;
shl.b32 %r551, %r43, 2;
add.s32 %r552, %r550, %r551;
st.shared.f32 [%r552+-1536], %f276;
st.shared.f32 [%r552+-1408], %f275;
st.shared.f32 [%r552+-1280], %f274;
st.shared.f32 [%r552+-1152], %f273;
st.shared.f32 [%r552+-1024], %f272;
st.shared.f32 [%r552+-896], %f271;
st.shared.f32 [%r552+-768], %f270;
st.shared.f32 [%r552+-640], %f269;
st.shared.f32 [%r552+-512], %f268;
st.shared.f32 [%r552+-384], %f267;
st.shared.f32 [%r552+-256], %f266;
st.shared.f32 [%r552+-128], %f265;

$L__BB10_5:
setp.ne.s32 %p5, %r41, 0;
bar.sync 0;
@%p5 bra $L__BB10_30;

mov.u32 %r554, %ctaid.x;
shl.b32 %r555, %r554, 1;
mov.u32 %r556, 1;
shl.b32 %r558, %r43, 2;
mov.u32 %r559, _ZZ13mul_mat_vec_qIL9ggml_type41ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r560, %r559, %r558;
mov.u32 %r561, 2;
ld.shared.f32 %f121, [%r560];
add.ftz.f32 %f122, %f121, %f276;
mov.b32 %r562, %f122;
mov.u32 %r563, 31;
mov.u32 %r564, 16;
mov.u32 %r565, -1;
shfl.sync.bfly.b32 %r566|%p6, %r562, %r564, %r563, %r565;
mov.b32 %f123, %r566;
add.ftz.f32 %f124, %f122, %f123;
mov.b32 %r567, %f124;
mov.u32 %r568, 8;
shfl.sync.bfly.b32 %r569|%p7, %r567, %r568, %r563, %r565;
mov.b32 %f125, %r569;
add.ftz.f32 %f126, %f124, %f125;
mov.b32 %r570, %f126;
mov.u32 %r571, 4;
shfl.sync.bfly.b32 %r572|%p8, %r570, %r571, %r563, %r565;
mov.b32 %f127, %r572;
add.ftz.f32 %f128, %f126, %f127;
mov.b32 %r573, %f128;
shfl.sync.bfly.b32 %r574|%p9, %r573, %r561, %r563, %r565;
mov.b32 %f129, %r574;
add.ftz.f32 %f130, %f128, %f129;
mov.b32 %r575, %f130;
shfl.sync.bfly.b32 %r576|%p10, %r575, %r556, %r563, %r565;
mov.b32 %f131, %r576;
add.ftz.f32 %f37, %f130, %f131;
setp.ne.s32 %p11, %r43, 0;
setp.ge.u32 %p12, %r555, %r28;
or.pred %p13, %p11, %p12;
@%p13 bra $L__BB10_8;

cvta.to.global.u64 %rd56, %rd19;
mov.u32 %r577, %ctaid.y;
mad.lo.s32 %r580, %r577, %r34, %r555;
mov.u32 %r581, %ctaid.z;
mad.lo.s32 %r582, %r581, %r40, %r580;
mul.wide.u32 %rd57, %r582, 4;
add.s64 %rd58, %rd56, %rd57;
st.global.f32 [%rd58], %f37;

$L__BB10_8:
ld.shared.f32 %f132, [%r560+128];
add.ftz.f32 %f133, %f132, %f275;
mov.b32 %r588, %f133;
shfl.sync.bfly.b32 %r592|%p14, %r588, %r564, %r563, %r565;
mov.b32 %f134, %r592;
add.ftz.f32 %f135, %f133, %f134;
mov.b32 %r593, %f135;
shfl.sync.bfly.b32 %r595|%p15, %r593, %r568, %r563, %r565;
mov.b32 %f136, %r595;
add.ftz.f32 %f137, %f135, %f136;
mov.b32 %r596, %f137;
shfl.sync.bfly.b32 %r598|%p16, %r596, %r571, %r563, %r565;
mov.b32 %f138, %r598;
add.ftz.f32 %f139, %f137, %f138;
mov.b32 %r599, %f139;
shfl.sync.bfly.b32 %r600|%p17, %r599, %r561, %r563, %r565;
mov.b32 %f140, %r600;
add.ftz.f32 %f141, %f139, %f140;
mov.b32 %r601, %f141;
shfl.sync.bfly.b32 %r603|%p18, %r601, %r556, %r563, %r565;
mov.b32 %f142, %r603;
add.ftz.f32 %f38, %f141, %f142;
setp.eq.s32 %p19, %r43, 1;
or.b32 %r606, %r555, 1;
setp.lt.u32 %p20, %r606, %r28;
and.pred %p1, %p19, %p20;
not.pred %p21, %p1;
@%p21 bra $L__BB10_10;

cvta.to.global.u64 %rd59, %rd19;
mov.u32 %r607, %ctaid.y;
mad.lo.s32 %r610, %r607, %r34, %r555;
mov.u32 %r611, %ctaid.z;
mad.lo.s32 %r612, %r611, %r40, %r610;
mul.wide.u32 %rd60, %r612, 4;
add.s64 %rd61, %rd59, %rd60;
st.global.f32 [%rd61+4], %f38;

$L__BB10_10:
mov.u32 %r617, 2;
ld.shared.f32 %f143, [%r560+256];
add.ftz.f32 %f144, %f143, %f274;
mov.b32 %r618, %f144;
mov.u32 %r619, 31;
mov.u32 %r620, 16;
mov.u32 %r621, -1;
shfl.sync.bfly.b32 %r622|%p22, %r618, %r620, %r619, %r621;
mov.b32 %f145, %r622;
add.ftz.f32 %f146, %f144, %f145;
mov.b32 %r623, %f146;
mov.u32 %r624, 8;
shfl.sync.bfly.b32 %r625|%p23, %r623, %r624, %r619, %r621;
mov.b32 %f147, %r625;
add.ftz.f32 %f148, %f146, %f147;
mov.b32 %r626, %f148;
mov.u32 %r627, 4;
shfl.sync.bfly.b32 %r628|%p24, %r626, %r627, %r619, %r621;
mov.b32 %f149, %r628;
add.ftz.f32 %f150, %f148, %f149;
mov.b32 %r629, %f150;
shfl.sync.bfly.b32 %r630|%p25, %r629, %r617, %r619, %r621;
mov.b32 %f151, %r630;
add.ftz.f32 %f152, %f150, %f151;
mov.b32 %r631, %f152;
mov.u32 %r632, 1;
shfl.sync.bfly.b32 %r633|%p26, %r631, %r632, %r619, %r621;
mov.b32 %f153, %r633;
add.ftz.f32 %f39, %f152, %f153;
@%p13 bra $L__BB10_12;

cvt.u64.u32 %rd62, %r28;
mov.u32 %r636, %ctaid.y;
mad.lo.s32 %r639, %r636, %r34, %r555;
mov.u32 %r640, %ctaid.z;
mad.lo.s32 %r641, %r640, %r40, %r639;
cvt.u64.u32 %rd63, %r641;
add.s64 %rd64, %rd62, %rd63;
cvta.to.global.u64 %rd65, %rd19;
shl.b64 %rd66, %rd64, 2;
add.s64 %rd67, %rd65, %rd66;
st.global.f32 [%rd67], %f39;

$L__BB10_12:
ld.shared.f32 %f154, [%r560+384];
add.ftz.f32 %f155, %f154, %f273;
mov.b32 %r647, %f155;
shfl.sync.bfly.b32 %r651|%p30, %r647, %r620, %r619, %r621;
mov.b32 %f156, %r651;
add.ftz.f32 %f157, %f155, %f156;
mov.b32 %r652, %f157;
shfl.sync.bfly.b32 %r654|%p31, %r652, %r624, %r619, %r621;
mov.b32 %f158, %r654;
add.ftz.f32 %f159, %f157, %f158;
mov.b32 %r655, %f159;
shfl.sync.bfly.b32 %r657|%p32, %r655, %r627, %r619, %r621;
mov.b32 %f160, %r657;
add.ftz.f32 %f161, %f159, %f160;
mov.b32 %r658, %f161;
shfl.sync.bfly.b32 %r659|%p33, %r658, %r617, %r619, %r621;
mov.b32 %f162, %r659;
add.ftz.f32 %f163, %f161, %f162;
mov.b32 %r660, %f163;
shfl.sync.bfly.b32 %r662|%p34, %r660, %r632, %r619, %r621;
mov.b32 %f164, %r662;
add.ftz.f32 %f40, %f163, %f164;
@%p21 bra $L__BB10_14;

add.s32 %r663, %r28, 1;
cvt.u64.u32 %rd68, %r663;
mov.u32 %r664, %ctaid.y;
mad.lo.s32 %r667, %r664, %r34, %r555;
mov.u32 %r668, %ctaid.z;
mad.lo.s32 %r669, %r668, %r40, %r667;
cvt.u64.u32 %rd69, %r669;
add.s64 %rd70, %rd68, %rd69;
cvta.to.global.u64 %rd71, %rd19;
shl.b64 %rd72, %rd70, 2;
add.s64 %rd73, %rd71, %rd72;
st.global.f32 [%rd73], %f40;

$L__BB10_14:
shl.b32 %r670, %r28, 1;
mov.u32 %r671, 1;
mov.u32 %r676, 2;
ld.shared.f32 %f165, [%r560+512];
add.ftz.f32 %f166, %f165, %f272;
mov.b32 %r677, %f166;
mov.u32 %r678, 31;
mov.u32 %r679, 16;
mov.u32 %r680, -1;
shfl.sync.bfly.b32 %r681|%p36, %r677, %r679, %r678, %r680;
mov.b32 %f167, %r681;
add.ftz.f32 %f168, %f166, %f167;
mov.b32 %r682, %f168;
mov.u32 %r683, 8;
shfl.sync.bfly.b32 %r684|%p37, %r682, %r683, %r678, %r680;
mov.b32 %f169, %r684;
add.ftz.f32 %f170, %f168, %f169;
mov.b32 %r685, %f170;
mov.u32 %r686, 4;
shfl.sync.bfly.b32 %r687|%p38, %r685, %r686, %r678, %r680;
mov.b32 %f171, %r687;
add.ftz.f32 %f172, %f170, %f171;
mov.b32 %r688, %f172;
shfl.sync.bfly.b32 %r689|%p39, %r688, %r676, %r678, %r680;
mov.b32 %f173, %r689;
add.ftz.f32 %f174, %f172, %f173;
mov.b32 %r690, %f174;
shfl.sync.bfly.b32 %r691|%p40, %r690, %r671, %r678, %r680;
mov.b32 %f175, %r691;
add.ftz.f32 %f41, %f174, %f175;
cvt.u64.u32 %rd74, %r670;
mov.u32 %r692, %ctaid.y;
mad.lo.s32 %r695, %r692, %r34, %r555;
mov.u32 %r696, %ctaid.z;
mad.lo.s32 %r697, %r696, %r40, %r695;
cvt.u64.u32 %rd75, %r697;
add.s64 %rd76, %rd74, %rd75;
cvta.to.global.u64 %rd77, %rd19;
shl.b64 %rd78, %rd76, 2;
add.s64 %rd15, %rd77, %rd78;
@%p13 bra $L__BB10_16;

st.global.f32 [%rd15], %f41;

$L__BB10_16:
ld.shared.f32 %f176, [%r560+640];
add.ftz.f32 %f177, %f176, %f271;
mov.b32 %r703, %f177;
shfl.sync.bfly.b32 %r707|%p44, %r703, %r679, %r678, %r680;
mov.b32 %f178, %r707;
add.ftz.f32 %f179, %f177, %f178;
mov.b32 %r708, %f179;
shfl.sync.bfly.b32 %r710|%p45, %r708, %r683, %r678, %r680;
mov.b32 %f180, %r710;
add.ftz.f32 %f181, %f179, %f180;
mov.b32 %r711, %f181;
shfl.sync.bfly.b32 %r713|%p46, %r711, %r686, %r678, %r680;
mov.b32 %f182, %r713;
add.ftz.f32 %f183, %f181, %f182;
mov.b32 %r714, %f183;
shfl.sync.bfly.b32 %r715|%p47, %r714, %r676, %r678, %r680;
mov.b32 %f184, %r715;
add.ftz.f32 %f185, %f183, %f184;
mov.b32 %r716, %f185;
shfl.sync.bfly.b32 %r718|%p48, %r716, %r671, %r678, %r680;
mov.b32 %f186, %r718;
add.ftz.f32 %f42, %f185, %f186;
@%p21 bra $L__BB10_18;

st.global.f32 [%rd15+4], %f42;

$L__BB10_18:
mul.lo.s32 %r23, %r28, 3;
mov.u32 %r723, 2;
ld.shared.f32 %f187, [%r560+768];
add.ftz.f32 %f188, %f187, %f270;
mov.b32 %r724, %f188;
mov.u32 %r725, 31;
mov.u32 %r726, 16;
mov.u32 %r727, -1;
shfl.sync.bfly.b32 %r728|%p50, %r724, %r726, %r725, %r727;
mov.b32 %f189, %r728;
add.ftz.f32 %f190, %f188, %f189;
mov.b32 %r729, %f190;
mov.u32 %r730, 8;
shfl.sync.bfly.b32 %r731|%p51, %r729, %r730, %r725, %r727;
mov.b32 %f191, %r731;
add.ftz.f32 %f192, %f190, %f191;
mov.b32 %r732, %f192;
mov.u32 %r733, 4;
shfl.sync.bfly.b32 %r734|%p52, %r732, %r733, %r725, %r727;
mov.b32 %f193, %r734;
add.ftz.f32 %f194, %f192, %f193;
mov.b32 %r735, %f194;
shfl.sync.bfly.b32 %r736|%p53, %r735, %r723, %r725, %r727;
mov.b32 %f195, %r736;
add.ftz.f32 %f196, %f194, %f195;
mov.b32 %r737, %f196;
mov.u32 %r738, 1;
shfl.sync.bfly.b32 %r739|%p54, %r737, %r738, %r725, %r727;
mov.b32 %f197, %r739;
add.ftz.f32 %f43, %f196, %f197;
@%p13 bra $L__BB10_20;

cvt.u64.u32 %rd79, %r23;
add.s64 %rd81, %rd79, %rd75;
shl.b64 %rd83, %rd81, 2;
add.s64 %rd84, %rd77, %rd83;
st.global.f32 [%rd84], %f43;

$L__BB10_20:
ld.shared.f32 %f198, [%r560+896];
add.ftz.f32 %f199, %f198, %f269;
mov.b32 %r753, %f199;
shfl.sync.bfly.b32 %r757|%p58, %r753, %r726, %r725, %r727;
mov.b32 %f200, %r757;
add.ftz.f32 %f201, %f199, %f200;
mov.b32 %r758, %f201;
shfl.sync.bfly.b32 %r760|%p59, %r758, %r730, %r725, %r727;
mov.b32 %f202, %r760;
add.ftz.f32 %f203, %f201, %f202;
mov.b32 %r761, %f203;
shfl.sync.bfly.b32 %r763|%p60, %r761, %r733, %r725, %r727;
mov.b32 %f204, %r763;
add.ftz.f32 %f205, %f203, %f204;
mov.b32 %r764, %f205;
shfl.sync.bfly.b32 %r765|%p61, %r764, %r723, %r725, %r727;
mov.b32 %f206, %r765;
add.ftz.f32 %f207, %f205, %f206;
mov.b32 %r766, %f207;
shfl.sync.bfly.b32 %r768|%p62, %r766, %r738, %r725, %r727;
mov.b32 %f208, %r768;
add.ftz.f32 %f44, %f207, %f208;
@%p21 bra $L__BB10_22;

add.s32 %r769, %r23, 1;
cvt.u64.u32 %rd85, %r769;
add.s64 %rd87, %rd85, %rd75;
shl.b64 %rd89, %rd87, 2;
add.s64 %rd90, %rd77, %rd89;
st.global.f32 [%rd90], %f44;

$L__BB10_22:
shl.b32 %r776, %r28, 2;
mov.u32 %r777, 2;
ld.shared.f32 %f209, [%r560+1024];
add.ftz.f32 %f210, %f209, %f268;
mov.b32 %r782, %f210;
mov.u32 %r783, 31;
mov.u32 %r784, 16;
mov.u32 %r785, -1;
shfl.sync.bfly.b32 %r786|%p64, %r782, %r784, %r783, %r785;
mov.b32 %f211, %r786;
add.ftz.f32 %f212, %f210, %f211;
mov.b32 %r787, %f212;
mov.u32 %r788, 8;
shfl.sync.bfly.b32 %r789|%p65, %r787, %r788, %r783, %r785;
mov.b32 %f213, %r789;
add.ftz.f32 %f214, %f212, %f213;
mov.b32 %r790, %f214;
mov.u32 %r791, 4;
shfl.sync.bfly.b32 %r792|%p66, %r790, %r791, %r783, %r785;
mov.b32 %f215, %r792;
add.ftz.f32 %f216, %f214, %f215;
mov.b32 %r793, %f216;
shfl.sync.bfly.b32 %r794|%p67, %r793, %r777, %r783, %r785;
mov.b32 %f217, %r794;
add.ftz.f32 %f218, %f216, %f217;
mov.b32 %r795, %f218;
mov.u32 %r796, 1;
shfl.sync.bfly.b32 %r797|%p68, %r795, %r796, %r783, %r785;
mov.b32 %f219, %r797;
add.ftz.f32 %f45, %f218, %f219;
cvt.u64.u32 %rd91, %r776;
add.s64 %rd93, %rd91, %rd75;
shl.b64 %rd95, %rd93, 2;
add.s64 %rd16, %rd77, %rd95;
@%p13 bra $L__BB10_24;

st.global.f32 [%rd16], %f45;

$L__BB10_24:
ld.shared.f32 %f220, [%r560+1152];
add.ftz.f32 %f221, %f220, %f267;
mov.b32 %r809, %f221;
shfl.sync.bfly.b32 %r813|%p72, %r809, %r784, %r783, %r785;
mov.b32 %f222, %r813;
add.ftz.f32 %f223, %f221, %f222;
mov.b32 %r814, %f223;
shfl.sync.bfly.b32 %r816|%p73, %r814, %r788, %r783, %r785;
mov.b32 %f224, %r816;
add.ftz.f32 %f225, %f223, %f224;
mov.b32 %r817, %f225;
shfl.sync.bfly.b32 %r819|%p74, %r817, %r791, %r783, %r785;
mov.b32 %f226, %r819;
add.ftz.f32 %f227, %f225, %f226;
mov.b32 %r820, %f227;
shfl.sync.bfly.b32 %r821|%p75, %r820, %r777, %r783, %r785;
mov.b32 %f228, %r821;
add.ftz.f32 %f229, %f227, %f228;
mov.b32 %r822, %f229;
shfl.sync.bfly.b32 %r824|%p76, %r822, %r796, %r783, %r785;
mov.b32 %f230, %r824;
add.ftz.f32 %f46, %f229, %f230;
@%p21 bra $L__BB10_26;

st.global.f32 [%rd16+4], %f46;

$L__BB10_26:
mul.lo.s32 %r24, %r28, 5;
mov.u32 %r829, 2;
ld.shared.f32 %f231, [%r560+1280];
add.ftz.f32 %f232, %f231, %f266;
mov.b32 %r830, %f232;
mov.u32 %r831, 31;
mov.u32 %r832, 16;
mov.u32 %r833, -1;
shfl.sync.bfly.b32 %r834|%p78, %r830, %r832, %r831, %r833;
mov.b32 %f233, %r834;
add.ftz.f32 %f234, %f232, %f233;
mov.b32 %r835, %f234;
mov.u32 %r836, 8;
shfl.sync.bfly.b32 %r837|%p79, %r835, %r836, %r831, %r833;
mov.b32 %f235, %r837;
add.ftz.f32 %f236, %f234, %f235;
mov.b32 %r838, %f236;
mov.u32 %r839, 4;
shfl.sync.bfly.b32 %r840|%p80, %r838, %r839, %r831, %r833;
mov.b32 %f237, %r840;
add.ftz.f32 %f238, %f236, %f237;
mov.b32 %r841, %f238;
shfl.sync.bfly.b32 %r842|%p81, %r841, %r829, %r831, %r833;
mov.b32 %f239, %r842;
add.ftz.f32 %f240, %f238, %f239;
mov.b32 %r843, %f240;
mov.u32 %r844, 1;
shfl.sync.bfly.b32 %r845|%p82, %r843, %r844, %r831, %r833;
mov.b32 %f241, %r845;
add.ftz.f32 %f47, %f240, %f241;
@%p13 bra $L__BB10_28;

cvt.u64.u32 %rd96, %r24;
add.s64 %rd98, %rd96, %rd75;
shl.b64 %rd100, %rd98, 2;
add.s64 %rd101, %rd77, %rd100;
st.global.f32 [%rd101], %f47;

$L__BB10_28:
ld.shared.f32 %f242, [%r560+1408];
add.ftz.f32 %f243, %f242, %f265;
mov.b32 %r859, %f243;
shfl.sync.bfly.b32 %r863|%p86, %r859, %r832, %r831, %r833;
mov.b32 %f244, %r863;
add.ftz.f32 %f245, %f243, %f244;
mov.b32 %r864, %f245;
shfl.sync.bfly.b32 %r866|%p87, %r864, %r836, %r831, %r833;
mov.b32 %f246, %r866;
add.ftz.f32 %f247, %f245, %f246;
mov.b32 %r867, %f247;
shfl.sync.bfly.b32 %r869|%p88, %r867, %r839, %r831, %r833;
mov.b32 %f248, %r869;
add.ftz.f32 %f249, %f247, %f248;
mov.b32 %r870, %f249;
shfl.sync.bfly.b32 %r871|%p89, %r870, %r829, %r831, %r833;
mov.b32 %f250, %r871;
add.ftz.f32 %f251, %f249, %f250;
mov.b32 %r872, %f251;
shfl.sync.bfly.b32 %r874|%p90, %r872, %r844, %r831, %r833;
mov.b32 %f252, %r874;
add.ftz.f32 %f48, %f251, %f252;
@%p21 bra $L__BB10_30;

add.s32 %r875, %r24, 1;
cvt.u64.u32 %rd102, %r875;
add.s64 %rd104, %rd102, %rd75;
shl.b64 %rd106, %rd104, 2;
add.s64 %rd107, %rd77, %rd106;
st.global.f32 [%rd107], %f48;

$L__BB10_30:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type41ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type41ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type41ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type41ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type41ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type41ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type41ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type41ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type41ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<95>;
.reg .b16 %rs<19>;
.reg .f32 %f<323>;
.reg .b32 %r<877>;
.reg .b64 %rd<111>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type41ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1792];

ld.param.u64 %rd20, [_Z13mul_mat_vec_qIL9ggml_type41ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd21, [_Z13mul_mat_vec_qIL9ggml_type41ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd22, [_Z13mul_mat_vec_qIL9ggml_type41ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r48, [_Z13mul_mat_vec_qIL9ggml_type41ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type41ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type41ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type41ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type41ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type41ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type41ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r45, [_Z13mul_mat_vec_qIL9ggml_type41ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r46, [_Z13mul_mat_vec_qIL9ggml_type41ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r47, [_Z13mul_mat_vec_qIL9ggml_type41ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type41ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type41ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type41ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type41ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r49, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r49, %r2;
mov.u32 %r50, %ctaid.x;
shl.b32 %r4, %r50, 1;
shr.s32 %r51, %r3, 31;
shr.u32 %r52, %r51, 30;
add.s32 %r53, %r3, %r52;
shr.s32 %r876, %r53, 2;
shr.u32 %r6, %r48, 7;
setp.ge.s32 %p3, %r876, %r6;
mov.f32 %f309, 0f00000000;
mov.f32 %f310, %f309;
mov.f32 %f311, %f309;
mov.f32 %f312, %f309;
mov.f32 %f313, %f309;
mov.f32 %f314, %f309;
mov.f32 %f315, %f309;
mov.f32 %f316, %f309;
mov.f32 %f317, %f309;
mov.f32 %f318, %f309;
mov.f32 %f319, %f309;
mov.f32 %f320, %f309;
mov.f32 %f321, %f309;
mov.f32 %f322, %f309;
@%p3 bra $L__BB11_3;

mov.u32 %r54, %ctaid.y;
mov.u32 %r55, %ctaid.z;
mul.hi.u32 %r56, %r55, %r42;
add.s32 %r57, %r56, %r55;
shr.u32 %r58, %r57, %r43;
mul.lo.s32 %r59, %r58, %r45;
mul.hi.u32 %r60, %r54, %r36;
add.s32 %r61, %r60, %r54;
shr.u32 %r62, %r61, %r37;
mul.lo.s32 %r63, %r62, %r39;
mul.lo.s32 %r64, %r54, %r40;
cvt.u64.u32 %rd23, %r64;
add.s32 %r65, %r4, 1;
and.b32 %r69, %r53, -4;
sub.s32 %r70, %r3, %r69;
shl.b32 %r71, %r70, 1;
cvt.s64.s32 %rd24, %r70;
add.s64 %rd25, %rd23, %rd24;
mul.lo.s32 %r72, %r55, %r46;
cvt.u64.u32 %rd26, %r72;
add.s64 %rd1, %rd25, %rd26;
shl.b32 %r73, %r34, 1;
shl.b32 %r869, %r876, 2;
add.s32 %r875, %r869, %r73;
mad.lo.s32 %r874, %r34, 3, %r869;
add.s32 %r74, %r876, %r34;
shl.b32 %r873, %r74, 2;
mad.lo.s32 %r872, %r34, 5, %r869;
mad.lo.s32 %r871, %r34, 6, %r869;
add.s32 %r870, %r34, %r869;
mad.lo.s32 %r75, %r4, %r33, %r876;
add.s32 %r76, %r75, %r59;
add.s32 %r77, %r76, %r63;
mul.wide.s32 %rd27, %r77, 18;
mul.wide.s32 %rd28, %r71, 2;
add.s64 %rd29, %rd27, %rd28;
cvta.to.global.u64 %rd30, %rd20;
add.s64 %rd31, %rd30, %rd29;
add.s64 %rd110, %rd31, 4;
add.s32 %r78, %r876, %r59;
add.s32 %r79, %r78, %r63;
mad.lo.s32 %r80, %r65, %r33, %r79;
mul.wide.s32 %rd32, %r80, 18;
add.s64 %rd33, %rd32, %rd28;
add.s64 %rd34, %rd30, %rd33;
add.s64 %rd109, %rd34, 4;
add.s64 %rd108, %rd30, %rd32;
add.s64 %rd107, %rd30, %rd27;
cvta.to.global.u64 %rd6, %rd21;
mov.f32 %f309, 0f00000000;

$L__BB11_2:
cvt.u64.u32 %rd35, %r869;
add.s64 %rd36, %rd1, %rd35;
mul.lo.s64 %rd37, %rd36, 36;
ld.global.nc.u16 %rs1, [%rd107];

{ cvt.f32.f16 %f85, %rs1;}


ld.global.nc.u16 %rs15, [%rd110+-2];
cvt.s32.s16 %r543, %rs15;
add.s64 %rd38, %rd6, %rd37;
ld.global.nc.u32 %r83, [%rd38+4];
ld.global.nc.u32 %r87, [%rd38+8];
ld.global.nc.u32 %r91, [%rd38+12];
ld.global.nc.u32 %r95, [%rd38+16];
and.b32 %r544, %r543, 30583;
mov.u32 %r545, 286261504;
prmt.b32 %r546, %r545, %r545, %r544;
shr.u32 %r547, %r543, 2;
and.b32 %r548, %r547, 30583;
prmt.b32 %r549, %r545, %r545, %r548;
and.b32 %r550, %r546, 30583;
mov.u32 %r551, 511;
prmt.b32 %r552, %r551, %r551, %r550;
and.b32 %r553, %r549, 30583;
prmt.b32 %r554, %r551, %r551, %r553;
shr.u32 %r555, %r546, 16;
and.b32 %r556, %r555, 30583;
prmt.b32 %r557, %r551, %r551, %r556;
shr.u32 %r558, %r549, 16;
and.b32 %r559, %r558, 30583;
prmt.b32 %r560, %r551, %r551, %r559;
mov.u32 %r561, 21520;
prmt.b32 %r478, %r552, %r554, %r561;
mov.u32 %r562, 30258;
prmt.b32 %r482, %r552, %r554, %r562;
prmt.b32 %r486, %r557, %r560, %r561;
prmt.b32 %r490, %r557, %r560, %r562;
mov.u32 %r513, 0;

dp4a.s32.s32 %r81, %r478, %r83, %r513;


dp4a.s32.s32 %r85, %r482, %r87, %r81;


dp4a.s32.s32 %r89, %r486, %r91, %r85;


dp4a.s32.s32 %r93, %r490, %r95, %r89;

ld.global.nc.u16 %rs16, [%rd110];
cvt.s32.s16 %r563, %rs16;
ld.global.nc.u32 %r99, [%rd38+20];
ld.global.nc.u32 %r103, [%rd38+24];
ld.global.nc.u32 %r107, [%rd38+28];
ld.global.nc.u32 %r111, [%rd38+32];
and.b32 %r564, %r563, 30583;
prmt.b32 %r565, %r545, %r545, %r564;
shr.u32 %r566, %r563, 2;
and.b32 %r567, %r566, 30583;
prmt.b32 %r568, %r545, %r545, %r567;
and.b32 %r569, %r565, 30583;
prmt.b32 %r570, %r551, %r551, %r569;
and.b32 %r571, %r568, 30583;
prmt.b32 %r572, %r551, %r551, %r571;
shr.u32 %r573, %r565, 16;
and.b32 %r574, %r573, 30583;
prmt.b32 %r575, %r551, %r551, %r574;
shr.u32 %r576, %r568, 16;
and.b32 %r577, %r576, 30583;
prmt.b32 %r578, %r551, %r551, %r577;
prmt.b32 %r494, %r570, %r572, %r561;
prmt.b32 %r498, %r570, %r572, %r562;
prmt.b32 %r502, %r575, %r578, %r561;
prmt.b32 %r506, %r575, %r578, %r562;

dp4a.s32.s32 %r97, %r494, %r99, %r93;


dp4a.s32.s32 %r101, %r498, %r103, %r97;


dp4a.s32.s32 %r105, %r502, %r107, %r101;


dp4a.s32.s32 %r109, %r506, %r111, %r105;

ld.global.nc.u32 %r113, [%rd38];

{.reg .f16 low,high;
mov.b32 {low,high},%r113;
cvt.f32.f16 %f86, low;}


mul.ftz.f32 %f113, %f85, %f86;
cvt.rn.f32.s32 %f114, %r109;
fma.rn.ftz.f32 %f322, %f113, %f114, %f322;
ld.global.nc.u16 %rs2, [%rd108];

{ cvt.f32.f16 %f87, %rs2;}


ld.global.nc.u16 %rs17, [%rd109+-2];
cvt.s32.s16 %r579, %rs17;
and.b32 %r580, %r579, 30583;
prmt.b32 %r581, %r545, %r545, %r580;
shr.u32 %r582, %r579, 2;
and.b32 %r583, %r582, 30583;
prmt.b32 %r584, %r545, %r545, %r583;
and.b32 %r585, %r581, 30583;
prmt.b32 %r586, %r551, %r551, %r585;
and.b32 %r587, %r584, 30583;
prmt.b32 %r588, %r551, %r551, %r587;
shr.u32 %r589, %r581, 16;
and.b32 %r590, %r589, 30583;
prmt.b32 %r591, %r551, %r551, %r590;
shr.u32 %r592, %r584, 16;
and.b32 %r593, %r592, 30583;
prmt.b32 %r594, %r551, %r551, %r593;
prmt.b32 %r511, %r586, %r588, %r561;
prmt.b32 %r515, %r586, %r588, %r562;
prmt.b32 %r519, %r591, %r594, %r561;
prmt.b32 %r523, %r591, %r594, %r562;

dp4a.s32.s32 %r114, %r511, %r83, %r513;


dp4a.s32.s32 %r118, %r515, %r87, %r114;


dp4a.s32.s32 %r122, %r519, %r91, %r118;


dp4a.s32.s32 %r126, %r523, %r95, %r122;

ld.global.nc.u16 %rs18, [%rd109];
cvt.s32.s16 %r595, %rs18;
and.b32 %r596, %r595, 30583;
prmt.b32 %r597, %r545, %r545, %r596;
shr.u32 %r598, %r595, 2;
and.b32 %r599, %r598, 30583;
prmt.b32 %r600, %r545, %r545, %r599;
and.b32 %r601, %r597, 30583;
prmt.b32 %r602, %r551, %r551, %r601;
and.b32 %r603, %r600, 30583;
prmt.b32 %r604, %r551, %r551, %r603;
shr.u32 %r605, %r597, 16;
and.b32 %r606, %r605, 30583;
prmt.b32 %r607, %r551, %r551, %r606;
shr.u32 %r608, %r600, 16;
and.b32 %r609, %r608, 30583;
prmt.b32 %r610, %r551, %r551, %r609;
prmt.b32 %r527, %r602, %r604, %r561;
prmt.b32 %r531, %r602, %r604, %r562;
prmt.b32 %r535, %r607, %r610, %r561;
prmt.b32 %r539, %r607, %r610, %r562;

dp4a.s32.s32 %r130, %r527, %r99, %r126;


dp4a.s32.s32 %r134, %r531, %r103, %r130;


dp4a.s32.s32 %r138, %r535, %r107, %r134;


dp4a.s32.s32 %r142, %r539, %r111, %r138;


{.reg .f16 low,high;
mov.b32 {low,high},%r113;
cvt.f32.f16 %f88, low;}


mul.ftz.f32 %f115, %f87, %f88;
cvt.rn.f32.s32 %f116, %r142;
fma.rn.ftz.f32 %f321, %f115, %f116, %f321;
cvt.u64.u32 %rd39, %r870;
add.s64 %rd40, %rd1, %rd39;
mul.lo.s64 %rd41, %rd40, 36;

{ cvt.f32.f16 %f89, %rs1;}


add.s64 %rd42, %rd6, %rd41;
ld.global.nc.u32 %r149, [%rd42+4];
ld.global.nc.u32 %r153, [%rd42+8];
ld.global.nc.u32 %r157, [%rd42+12];
ld.global.nc.u32 %r161, [%rd42+16];

dp4a.s32.s32 %r147, %r478, %r149, %r513;


dp4a.s32.s32 %r151, %r482, %r153, %r147;


dp4a.s32.s32 %r155, %r486, %r157, %r151;


dp4a.s32.s32 %r159, %r490, %r161, %r155;

ld.global.nc.u32 %r165, [%rd42+20];
ld.global.nc.u32 %r169, [%rd42+24];
ld.global.nc.u32 %r173, [%rd42+28];
ld.global.nc.u32 %r177, [%rd42+32];

dp4a.s32.s32 %r163, %r494, %r165, %r159;


dp4a.s32.s32 %r167, %r498, %r169, %r163;


dp4a.s32.s32 %r171, %r502, %r173, %r167;


dp4a.s32.s32 %r175, %r506, %r177, %r171;

ld.global.nc.u32 %r179, [%rd42];

{.reg .f16 low,high;
mov.b32 {low,high},%r179;
cvt.f32.f16 %f90, low;}


mul.ftz.f32 %f117, %f89, %f90;
cvt.rn.f32.s32 %f118, %r175;
fma.rn.ftz.f32 %f320, %f117, %f118, %f320;

{ cvt.f32.f16 %f91, %rs2;}



dp4a.s32.s32 %r180, %r511, %r149, %r513;


dp4a.s32.s32 %r184, %r515, %r153, %r180;


dp4a.s32.s32 %r188, %r519, %r157, %r184;


dp4a.s32.s32 %r192, %r523, %r161, %r188;


dp4a.s32.s32 %r196, %r527, %r165, %r192;


dp4a.s32.s32 %r200, %r531, %r169, %r196;


dp4a.s32.s32 %r204, %r535, %r173, %r200;


dp4a.s32.s32 %r208, %r539, %r177, %r204;


{.reg .f16 low,high;
mov.b32 {low,high},%r179;
cvt.f32.f16 %f92, low;}


mul.ftz.f32 %f119, %f91, %f92;
cvt.rn.f32.s32 %f120, %r208;
fma.rn.ftz.f32 %f319, %f119, %f120, %f319;
cvt.u64.u32 %rd43, %r875;
add.s64 %rd44, %rd1, %rd43;
mul.lo.s64 %rd45, %rd44, 36;

{ cvt.f32.f16 %f93, %rs1;}


add.s64 %rd46, %rd6, %rd45;
ld.global.nc.u32 %r215, [%rd46+4];
ld.global.nc.u32 %r219, [%rd46+8];
ld.global.nc.u32 %r223, [%rd46+12];
ld.global.nc.u32 %r227, [%rd46+16];

dp4a.s32.s32 %r213, %r478, %r215, %r513;


dp4a.s32.s32 %r217, %r482, %r219, %r213;


dp4a.s32.s32 %r221, %r486, %r223, %r217;


dp4a.s32.s32 %r225, %r490, %r227, %r221;

ld.global.nc.u32 %r231, [%rd46+20];
ld.global.nc.u32 %r235, [%rd46+24];
ld.global.nc.u32 %r239, [%rd46+28];
ld.global.nc.u32 %r243, [%rd46+32];

dp4a.s32.s32 %r229, %r494, %r231, %r225;


dp4a.s32.s32 %r233, %r498, %r235, %r229;


dp4a.s32.s32 %r237, %r502, %r239, %r233;


dp4a.s32.s32 %r241, %r506, %r243, %r237;

ld.global.nc.u32 %r245, [%rd46];

{.reg .f16 low,high;
mov.b32 {low,high},%r245;
cvt.f32.f16 %f94, low;}


mul.ftz.f32 %f121, %f93, %f94;
cvt.rn.f32.s32 %f122, %r241;
fma.rn.ftz.f32 %f318, %f121, %f122, %f318;

{ cvt.f32.f16 %f95, %rs2;}



dp4a.s32.s32 %r246, %r511, %r215, %r513;


dp4a.s32.s32 %r250, %r515, %r219, %r246;


dp4a.s32.s32 %r254, %r519, %r223, %r250;


dp4a.s32.s32 %r258, %r523, %r227, %r254;


dp4a.s32.s32 %r262, %r527, %r231, %r258;


dp4a.s32.s32 %r266, %r531, %r235, %r262;


dp4a.s32.s32 %r270, %r535, %r239, %r266;


dp4a.s32.s32 %r274, %r539, %r243, %r270;


{.reg .f16 low,high;
mov.b32 {low,high},%r245;
cvt.f32.f16 %f96, low;}


mul.ftz.f32 %f123, %f95, %f96;
cvt.rn.f32.s32 %f124, %r274;
fma.rn.ftz.f32 %f317, %f123, %f124, %f317;
cvt.u64.u32 %rd47, %r874;
add.s64 %rd48, %rd1, %rd47;
mul.lo.s64 %rd49, %rd48, 36;

{ cvt.f32.f16 %f97, %rs1;}


add.s64 %rd50, %rd6, %rd49;
ld.global.nc.u32 %r281, [%rd50+4];
ld.global.nc.u32 %r285, [%rd50+8];
ld.global.nc.u32 %r289, [%rd50+12];
ld.global.nc.u32 %r293, [%rd50+16];

dp4a.s32.s32 %r279, %r478, %r281, %r513;


dp4a.s32.s32 %r283, %r482, %r285, %r279;


dp4a.s32.s32 %r287, %r486, %r289, %r283;


dp4a.s32.s32 %r291, %r490, %r293, %r287;

ld.global.nc.u32 %r297, [%rd50+20];
ld.global.nc.u32 %r301, [%rd50+24];
ld.global.nc.u32 %r305, [%rd50+28];
ld.global.nc.u32 %r309, [%rd50+32];

dp4a.s32.s32 %r295, %r494, %r297, %r291;


dp4a.s32.s32 %r299, %r498, %r301, %r295;


dp4a.s32.s32 %r303, %r502, %r305, %r299;


dp4a.s32.s32 %r307, %r506, %r309, %r303;

ld.global.nc.u32 %r311, [%rd50];

{.reg .f16 low,high;
mov.b32 {low,high},%r311;
cvt.f32.f16 %f98, low;}


mul.ftz.f32 %f125, %f97, %f98;
cvt.rn.f32.s32 %f126, %r307;
fma.rn.ftz.f32 %f316, %f125, %f126, %f316;

{ cvt.f32.f16 %f99, %rs2;}



dp4a.s32.s32 %r312, %r511, %r281, %r513;


dp4a.s32.s32 %r316, %r515, %r285, %r312;


dp4a.s32.s32 %r320, %r519, %r289, %r316;


dp4a.s32.s32 %r324, %r523, %r293, %r320;


dp4a.s32.s32 %r328, %r527, %r297, %r324;


dp4a.s32.s32 %r332, %r531, %r301, %r328;


dp4a.s32.s32 %r336, %r535, %r305, %r332;


dp4a.s32.s32 %r340, %r539, %r309, %r336;


{.reg .f16 low,high;
mov.b32 {low,high},%r311;
cvt.f32.f16 %f100, low;}


mul.ftz.f32 %f127, %f99, %f100;
cvt.rn.f32.s32 %f128, %r340;
fma.rn.ftz.f32 %f315, %f127, %f128, %f315;
cvt.u64.u32 %rd51, %r873;
add.s64 %rd52, %rd1, %rd51;
mul.lo.s64 %rd53, %rd52, 36;

{ cvt.f32.f16 %f101, %rs1;}


add.s64 %rd54, %rd6, %rd53;
ld.global.nc.u32 %r347, [%rd54+4];
ld.global.nc.u32 %r351, [%rd54+8];
ld.global.nc.u32 %r355, [%rd54+12];
ld.global.nc.u32 %r359, [%rd54+16];

dp4a.s32.s32 %r345, %r478, %r347, %r513;


dp4a.s32.s32 %r349, %r482, %r351, %r345;


dp4a.s32.s32 %r353, %r486, %r355, %r349;


dp4a.s32.s32 %r357, %r490, %r359, %r353;

ld.global.nc.u32 %r363, [%rd54+20];
ld.global.nc.u32 %r367, [%rd54+24];
ld.global.nc.u32 %r371, [%rd54+28];
ld.global.nc.u32 %r375, [%rd54+32];

dp4a.s32.s32 %r361, %r494, %r363, %r357;


dp4a.s32.s32 %r365, %r498, %r367, %r361;


dp4a.s32.s32 %r369, %r502, %r371, %r365;


dp4a.s32.s32 %r373, %r506, %r375, %r369;

ld.global.nc.u32 %r377, [%rd54];

{.reg .f16 low,high;
mov.b32 {low,high},%r377;
cvt.f32.f16 %f102, low;}


mul.ftz.f32 %f129, %f101, %f102;
cvt.rn.f32.s32 %f130, %r373;
fma.rn.ftz.f32 %f314, %f129, %f130, %f314;

{ cvt.f32.f16 %f103, %rs2;}



dp4a.s32.s32 %r378, %r511, %r347, %r513;


dp4a.s32.s32 %r382, %r515, %r351, %r378;


dp4a.s32.s32 %r386, %r519, %r355, %r382;


dp4a.s32.s32 %r390, %r523, %r359, %r386;


dp4a.s32.s32 %r394, %r527, %r363, %r390;


dp4a.s32.s32 %r398, %r531, %r367, %r394;


dp4a.s32.s32 %r402, %r535, %r371, %r398;


dp4a.s32.s32 %r406, %r539, %r375, %r402;


{.reg .f16 low,high;
mov.b32 {low,high},%r377;
cvt.f32.f16 %f104, low;}


mul.ftz.f32 %f131, %f103, %f104;
cvt.rn.f32.s32 %f132, %r406;
fma.rn.ftz.f32 %f313, %f131, %f132, %f313;
cvt.u64.u32 %rd55, %r872;
add.s64 %rd56, %rd1, %rd55;
mul.lo.s64 %rd57, %rd56, 36;

{ cvt.f32.f16 %f105, %rs1;}


add.s64 %rd58, %rd6, %rd57;
ld.global.nc.u32 %r413, [%rd58+4];
ld.global.nc.u32 %r417, [%rd58+8];
ld.global.nc.u32 %r421, [%rd58+12];
ld.global.nc.u32 %r425, [%rd58+16];

dp4a.s32.s32 %r411, %r478, %r413, %r513;


dp4a.s32.s32 %r415, %r482, %r417, %r411;


dp4a.s32.s32 %r419, %r486, %r421, %r415;


dp4a.s32.s32 %r423, %r490, %r425, %r419;

ld.global.nc.u32 %r429, [%rd58+20];
ld.global.nc.u32 %r433, [%rd58+24];
ld.global.nc.u32 %r437, [%rd58+28];
ld.global.nc.u32 %r441, [%rd58+32];

dp4a.s32.s32 %r427, %r494, %r429, %r423;


dp4a.s32.s32 %r431, %r498, %r433, %r427;


dp4a.s32.s32 %r435, %r502, %r437, %r431;


dp4a.s32.s32 %r439, %r506, %r441, %r435;

ld.global.nc.u32 %r443, [%rd58];

{.reg .f16 low,high;
mov.b32 {low,high},%r443;
cvt.f32.f16 %f106, low;}


mul.ftz.f32 %f133, %f105, %f106;
cvt.rn.f32.s32 %f134, %r439;
fma.rn.ftz.f32 %f312, %f133, %f134, %f312;

{ cvt.f32.f16 %f107, %rs2;}



dp4a.s32.s32 %r444, %r511, %r413, %r513;


dp4a.s32.s32 %r448, %r515, %r417, %r444;


dp4a.s32.s32 %r452, %r519, %r421, %r448;


dp4a.s32.s32 %r456, %r523, %r425, %r452;


dp4a.s32.s32 %r460, %r527, %r429, %r456;


dp4a.s32.s32 %r464, %r531, %r433, %r460;


dp4a.s32.s32 %r468, %r535, %r437, %r464;


dp4a.s32.s32 %r472, %r539, %r441, %r468;


{.reg .f16 low,high;
mov.b32 {low,high},%r443;
cvt.f32.f16 %f108, low;}


mul.ftz.f32 %f135, %f107, %f108;
cvt.rn.f32.s32 %f136, %r472;
fma.rn.ftz.f32 %f311, %f135, %f136, %f311;
cvt.u64.u32 %rd59, %r871;
add.s64 %rd60, %rd1, %rd59;
mul.lo.s64 %rd61, %rd60, 36;

{ cvt.f32.f16 %f109, %rs1;}


add.s64 %rd62, %rd6, %rd61;
ld.global.nc.u32 %r479, [%rd62+4];
ld.global.nc.u32 %r483, [%rd62+8];
ld.global.nc.u32 %r487, [%rd62+12];
ld.global.nc.u32 %r491, [%rd62+16];

dp4a.s32.s32 %r477, %r478, %r479, %r513;


dp4a.s32.s32 %r481, %r482, %r483, %r477;


dp4a.s32.s32 %r485, %r486, %r487, %r481;


dp4a.s32.s32 %r489, %r490, %r491, %r485;

ld.global.nc.u32 %r495, [%rd62+20];
ld.global.nc.u32 %r499, [%rd62+24];
ld.global.nc.u32 %r503, [%rd62+28];
ld.global.nc.u32 %r507, [%rd62+32];

dp4a.s32.s32 %r493, %r494, %r495, %r489;


dp4a.s32.s32 %r497, %r498, %r499, %r493;


dp4a.s32.s32 %r501, %r502, %r503, %r497;


dp4a.s32.s32 %r505, %r506, %r507, %r501;

ld.global.nc.u32 %r509, [%rd62];

{.reg .f16 low,high;
mov.b32 {low,high},%r509;
cvt.f32.f16 %f110, low;}


mul.ftz.f32 %f137, %f109, %f110;
cvt.rn.f32.s32 %f138, %r505;
fma.rn.ftz.f32 %f310, %f137, %f138, %f310;

{ cvt.f32.f16 %f111, %rs2;}



dp4a.s32.s32 %r510, %r511, %r479, %r513;


dp4a.s32.s32 %r514, %r515, %r483, %r510;


dp4a.s32.s32 %r518, %r519, %r487, %r514;


dp4a.s32.s32 %r522, %r523, %r491, %r518;


dp4a.s32.s32 %r526, %r527, %r495, %r522;


dp4a.s32.s32 %r530, %r531, %r499, %r526;


dp4a.s32.s32 %r534, %r535, %r503, %r530;


dp4a.s32.s32 %r538, %r539, %r507, %r534;


{.reg .f16 low,high;
mov.b32 {low,high},%r509;
cvt.f32.f16 %f112, low;}


mul.ftz.f32 %f139, %f111, %f112;
cvt.rn.f32.s32 %f140, %r538;
fma.rn.ftz.f32 %f309, %f139, %f140, %f309;
add.s32 %r875, %r875, 64;
add.s32 %r874, %r874, 64;
add.s32 %r873, %r873, 64;
add.s32 %r872, %r872, 64;
add.s32 %r871, %r871, 64;
add.s32 %r870, %r870, 64;
add.s32 %r869, %r869, 64;
add.s64 %rd110, %rd110, 288;
add.s64 %rd109, %rd109, 288;
add.s64 %rd108, %rd108, 288;
add.s64 %rd107, %rd107, 288;
add.s32 %r876, %r876, 16;
setp.lt.s32 %p4, %r876, %r6;
@%p4 bra $L__BB11_2;

$L__BB11_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB11_5;

mov.u32 %r611, _ZZ13mul_mat_vec_qIL9ggml_type41ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r612, %r1, 1792, %r611;
shl.b32 %r613, %r2, 2;
add.s32 %r614, %r612, %r613;
st.shared.f32 [%r614+-1792], %f322;
st.shared.f32 [%r614+-1664], %f321;
st.shared.f32 [%r614+-1536], %f320;
st.shared.f32 [%r614+-1408], %f319;
st.shared.f32 [%r614+-1280], %f318;
st.shared.f32 [%r614+-1152], %f317;
st.shared.f32 [%r614+-1024], %f316;
st.shared.f32 [%r614+-896], %f315;
st.shared.f32 [%r614+-768], %f314;
st.shared.f32 [%r614+-640], %f313;
st.shared.f32 [%r614+-512], %f312;
st.shared.f32 [%r614+-384], %f311;
st.shared.f32 [%r614+-256], %f310;
st.shared.f32 [%r614+-128], %f309;

$L__BB11_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB11_34;

mov.u32 %r615, %ctaid.z;
mov.u32 %r616, %ctaid.y;
mad.lo.s32 %r617, %r616, %r41, %r4;
mad.lo.s32 %r618, %r615, %r47, %r617;
cvt.u64.u32 %rd15, %r618;
shl.b32 %r619, %r2, 2;
mov.u32 %r620, _ZZ13mul_mat_vec_qIL9ggml_type41ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r30, %r620, %r619;
mov.u32 %r621, 2;
ld.shared.f32 %f141, [%r30];
add.ftz.f32 %f142, %f141, %f322;
mov.b32 %r622, %f142;
mov.u32 %r623, 31;
mov.u32 %r624, 16;
mov.u32 %r625, -1;
shfl.sync.bfly.b32 %r626|%p7, %r622, %r624, %r623, %r625;
mov.b32 %f143, %r626;
add.ftz.f32 %f144, %f142, %f143;
mov.b32 %r627, %f144;
mov.u32 %r628, 8;
shfl.sync.bfly.b32 %r629|%p8, %r627, %r628, %r623, %r625;
mov.b32 %f145, %r629;
add.ftz.f32 %f146, %f144, %f145;
mov.b32 %r630, %f146;
mov.u32 %r631, 4;
shfl.sync.bfly.b32 %r632|%p9, %r630, %r631, %r623, %r625;
mov.b32 %f147, %r632;
add.ftz.f32 %f148, %f146, %f147;
mov.b32 %r633, %f148;
shfl.sync.bfly.b32 %r634|%p10, %r633, %r621, %r623, %r625;
mov.b32 %f149, %r634;
add.ftz.f32 %f150, %f148, %f149;
mov.b32 %r635, %f150;
mov.u32 %r636, 1;
shfl.sync.bfly.b32 %r637|%p11, %r635, %r636, %r623, %r625;
mov.b32 %f151, %r637;
add.ftz.f32 %f43, %f150, %f151;
setp.lt.u32 %p12, %r4, %r35;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
cvta.to.global.u64 %rd63, %rd22;
mul.wide.u32 %rd64, %r618, 4;
add.s64 %rd16, %rd63, %rd64;
not.pred %p14, %p1;
@%p14 bra $L__BB11_8;

st.global.f32 [%rd16], %f43;

$L__BB11_8:
ld.shared.f32 %f152, [%r30+128];
add.ftz.f32 %f153, %f152, %f321;
mov.b32 %r638, %f153;
shfl.sync.bfly.b32 %r642|%p15, %r638, %r624, %r623, %r625;
mov.b32 %f154, %r642;
add.ftz.f32 %f155, %f153, %f154;
mov.b32 %r643, %f155;
shfl.sync.bfly.b32 %r645|%p16, %r643, %r628, %r623, %r625;
mov.b32 %f156, %r645;
add.ftz.f32 %f157, %f155, %f156;
mov.b32 %r646, %f157;
shfl.sync.bfly.b32 %r648|%p17, %r646, %r631, %r623, %r625;
mov.b32 %f158, %r648;
add.ftz.f32 %f159, %f157, %f158;
mov.b32 %r649, %f159;
shfl.sync.bfly.b32 %r651|%p18, %r649, %r621, %r623, %r625;
mov.b32 %f160, %r651;
add.ftz.f32 %f161, %f159, %f160;
mov.b32 %r652, %f161;
shfl.sync.bfly.b32 %r654|%p19, %r652, %r636, %r623, %r625;
mov.b32 %f162, %r654;
add.ftz.f32 %f44, %f161, %f162;
setp.eq.s32 %p20, %r2, 1;
or.b32 %r658, %r4, 1;
setp.lt.u32 %p21, %r658, %r35;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB11_10;

st.global.f32 [%rd16+4], %f44;

$L__BB11_10:
ld.shared.f32 %f163, [%r30+256];
add.ftz.f32 %f164, %f163, %f320;
mov.b32 %r659, %f164;
mov.u32 %r660, 31;
mov.u32 %r661, 16;
mov.u32 %r662, -1;
shfl.sync.bfly.b32 %r663|%p23, %r659, %r661, %r660, %r662;
mov.b32 %f165, %r663;
add.ftz.f32 %f166, %f164, %f165;
mov.b32 %r664, %f166;
mov.u32 %r665, 8;
shfl.sync.bfly.b32 %r666|%p24, %r664, %r665, %r660, %r662;
mov.b32 %f167, %r666;
add.ftz.f32 %f168, %f166, %f167;
mov.b32 %r667, %f168;
mov.u32 %r668, 4;
shfl.sync.bfly.b32 %r669|%p25, %r667, %r668, %r660, %r662;
mov.b32 %f169, %r669;
add.ftz.f32 %f170, %f168, %f169;
mov.b32 %r670, %f170;
mov.u32 %r671, 2;
shfl.sync.bfly.b32 %r672|%p26, %r670, %r671, %r660, %r662;
mov.b32 %f171, %r672;
add.ftz.f32 %f172, %f170, %f171;
mov.b32 %r673, %f172;
mov.u32 %r674, 1;
shfl.sync.bfly.b32 %r675|%p27, %r673, %r674, %r660, %r662;
mov.b32 %f173, %r675;
add.ftz.f32 %f45, %f172, %f173;
@%p14 bra $L__BB11_12;

cvt.u64.u32 %rd65, %r35;
add.s64 %rd66, %rd65, %rd15;
shl.b64 %rd68, %rd66, 2;
add.s64 %rd69, %rd63, %rd68;
st.global.f32 [%rd69], %f45;

$L__BB11_12:
ld.shared.f32 %f174, [%r30+384];
add.ftz.f32 %f175, %f174, %f319;
mov.b32 %r676, %f175;
shfl.sync.bfly.b32 %r680|%p29, %r676, %r661, %r660, %r662;
mov.b32 %f176, %r680;
add.ftz.f32 %f177, %f175, %f176;
mov.b32 %r681, %f177;
shfl.sync.bfly.b32 %r683|%p30, %r681, %r665, %r660, %r662;
mov.b32 %f178, %r683;
add.ftz.f32 %f179, %f177, %f178;
mov.b32 %r684, %f179;
shfl.sync.bfly.b32 %r686|%p31, %r684, %r668, %r660, %r662;
mov.b32 %f180, %r686;
add.ftz.f32 %f181, %f179, %f180;
mov.b32 %r687, %f181;
shfl.sync.bfly.b32 %r689|%p32, %r687, %r671, %r660, %r662;
mov.b32 %f182, %r689;
add.ftz.f32 %f183, %f181, %f182;
mov.b32 %r690, %f183;
shfl.sync.bfly.b32 %r692|%p33, %r690, %r674, %r660, %r662;
mov.b32 %f184, %r692;
add.ftz.f32 %f46, %f183, %f184;
@%p22 bra $L__BB11_14;

add.s32 %r693, %r35, 1;
cvt.u64.u32 %rd70, %r693;
add.s64 %rd71, %rd70, %rd15;
shl.b64 %rd73, %rd71, 2;
add.s64 %rd74, %rd63, %rd73;
st.global.f32 [%rd74], %f46;

$L__BB11_14:
shl.b32 %r694, %r35, 1;
mov.u32 %r695, 1;
ld.shared.f32 %f185, [%r30+512];
add.ftz.f32 %f186, %f185, %f318;
mov.b32 %r696, %f186;
mov.u32 %r697, 31;
mov.u32 %r698, 16;
mov.u32 %r699, -1;
shfl.sync.bfly.b32 %r700|%p35, %r696, %r698, %r697, %r699;
mov.b32 %f187, %r700;
add.ftz.f32 %f188, %f186, %f187;
mov.b32 %r701, %f188;
mov.u32 %r702, 8;
shfl.sync.bfly.b32 %r703|%p36, %r701, %r702, %r697, %r699;
mov.b32 %f189, %r703;
add.ftz.f32 %f190, %f188, %f189;
mov.b32 %r704, %f190;
mov.u32 %r705, 4;
shfl.sync.bfly.b32 %r706|%p37, %r704, %r705, %r697, %r699;
mov.b32 %f191, %r706;
add.ftz.f32 %f192, %f190, %f191;
mov.b32 %r707, %f192;
mov.u32 %r708, 2;
shfl.sync.bfly.b32 %r709|%p38, %r707, %r708, %r697, %r699;
mov.b32 %f193, %r709;
add.ftz.f32 %f194, %f192, %f193;
mov.b32 %r710, %f194;
shfl.sync.bfly.b32 %r711|%p39, %r710, %r695, %r697, %r699;
mov.b32 %f195, %r711;
add.ftz.f32 %f47, %f194, %f195;
cvt.u64.u32 %rd75, %r694;
add.s64 %rd76, %rd75, %rd15;
shl.b64 %rd78, %rd76, 2;
add.s64 %rd17, %rd63, %rd78;
@%p14 bra $L__BB11_16;

st.global.f32 [%rd17], %f47;

$L__BB11_16:
ld.shared.f32 %f196, [%r30+640];
add.ftz.f32 %f197, %f196, %f317;
mov.b32 %r712, %f197;
shfl.sync.bfly.b32 %r716|%p41, %r712, %r698, %r697, %r699;
mov.b32 %f198, %r716;
add.ftz.f32 %f199, %f197, %f198;
mov.b32 %r717, %f199;
shfl.sync.bfly.b32 %r719|%p42, %r717, %r702, %r697, %r699;
mov.b32 %f200, %r719;
add.ftz.f32 %f201, %f199, %f200;
mov.b32 %r720, %f201;
shfl.sync.bfly.b32 %r722|%p43, %r720, %r705, %r697, %r699;
mov.b32 %f202, %r722;
add.ftz.f32 %f203, %f201, %f202;
mov.b32 %r723, %f203;
shfl.sync.bfly.b32 %r725|%p44, %r723, %r708, %r697, %r699;
mov.b32 %f204, %r725;
add.ftz.f32 %f205, %f203, %f204;
mov.b32 %r726, %f205;
shfl.sync.bfly.b32 %r728|%p45, %r726, %r695, %r697, %r699;
mov.b32 %f206, %r728;
add.ftz.f32 %f48, %f205, %f206;
@%p22 bra $L__BB11_18;

st.global.f32 [%rd17+4], %f48;

$L__BB11_18:
mul.lo.s32 %r31, %r35, 3;
ld.shared.f32 %f207, [%r30+768];
add.ftz.f32 %f208, %f207, %f316;
mov.b32 %r729, %f208;
mov.u32 %r730, 31;
mov.u32 %r731, 16;
mov.u32 %r732, -1;
shfl.sync.bfly.b32 %r733|%p47, %r729, %r731, %r730, %r732;
mov.b32 %f209, %r733;
add.ftz.f32 %f210, %f208, %f209;
mov.b32 %r734, %f210;
mov.u32 %r735, 8;
shfl.sync.bfly.b32 %r736|%p48, %r734, %r735, %r730, %r732;
mov.b32 %f211, %r736;
add.ftz.f32 %f212, %f210, %f211;
mov.b32 %r737, %f212;
mov.u32 %r738, 4;
shfl.sync.bfly.b32 %r739|%p49, %r737, %r738, %r730, %r732;
mov.b32 %f213, %r739;
add.ftz.f32 %f214, %f212, %f213;
mov.b32 %r740, %f214;
mov.u32 %r741, 2;
shfl.sync.bfly.b32 %r742|%p50, %r740, %r741, %r730, %r732;
mov.b32 %f215, %r742;
add.ftz.f32 %f216, %f214, %f215;
mov.b32 %r743, %f216;
mov.u32 %r744, 1;
shfl.sync.bfly.b32 %r745|%p51, %r743, %r744, %r730, %r732;
mov.b32 %f217, %r745;
add.ftz.f32 %f49, %f216, %f217;
@%p14 bra $L__BB11_20;

cvt.u64.u32 %rd79, %r31;
add.s64 %rd80, %rd79, %rd15;
shl.b64 %rd82, %rd80, 2;
add.s64 %rd83, %rd63, %rd82;
st.global.f32 [%rd83], %f49;

$L__BB11_20:
ld.shared.f32 %f218, [%r30+896];
add.ftz.f32 %f219, %f218, %f315;
mov.b32 %r746, %f219;
shfl.sync.bfly.b32 %r750|%p53, %r746, %r731, %r730, %r732;
mov.b32 %f220, %r750;
add.ftz.f32 %f221, %f219, %f220;
mov.b32 %r751, %f221;
shfl.sync.bfly.b32 %r753|%p54, %r751, %r735, %r730, %r732;
mov.b32 %f222, %r753;
add.ftz.f32 %f223, %f221, %f222;
mov.b32 %r754, %f223;
shfl.sync.bfly.b32 %r756|%p55, %r754, %r738, %r730, %r732;
mov.b32 %f224, %r756;
add.ftz.f32 %f225, %f223, %f224;
mov.b32 %r757, %f225;
shfl.sync.bfly.b32 %r759|%p56, %r757, %r741, %r730, %r732;
mov.b32 %f226, %r759;
add.ftz.f32 %f227, %f225, %f226;
mov.b32 %r760, %f227;
shfl.sync.bfly.b32 %r762|%p57, %r760, %r744, %r730, %r732;
mov.b32 %f228, %r762;
add.ftz.f32 %f50, %f227, %f228;
@%p22 bra $L__BB11_22;

add.s32 %r763, %r31, 1;
cvt.u64.u32 %rd84, %r763;
add.s64 %rd85, %rd84, %rd15;
shl.b64 %rd87, %rd85, 2;
add.s64 %rd88, %rd63, %rd87;
st.global.f32 [%rd88], %f50;

$L__BB11_22:
shl.b32 %r764, %r35, 2;
mov.u32 %r765, 2;
ld.shared.f32 %f229, [%r30+1024];
add.ftz.f32 %f230, %f229, %f314;
mov.b32 %r766, %f230;
mov.u32 %r767, 31;
mov.u32 %r768, 16;
mov.u32 %r769, -1;
shfl.sync.bfly.b32 %r770|%p59, %r766, %r768, %r767, %r769;
mov.b32 %f231, %r770;
add.ftz.f32 %f232, %f230, %f231;
mov.b32 %r771, %f232;
mov.u32 %r772, 8;
shfl.sync.bfly.b32 %r773|%p60, %r771, %r772, %r767, %r769;
mov.b32 %f233, %r773;
add.ftz.f32 %f234, %f232, %f233;
mov.b32 %r774, %f234;
mov.u32 %r775, 4;
shfl.sync.bfly.b32 %r776|%p61, %r774, %r775, %r767, %r769;
mov.b32 %f235, %r776;
add.ftz.f32 %f236, %f234, %f235;
mov.b32 %r777, %f236;
shfl.sync.bfly.b32 %r778|%p62, %r777, %r765, %r767, %r769;
mov.b32 %f237, %r778;
add.ftz.f32 %f238, %f236, %f237;
mov.b32 %r779, %f238;
mov.u32 %r780, 1;
shfl.sync.bfly.b32 %r781|%p63, %r779, %r780, %r767, %r769;
mov.b32 %f239, %r781;
add.ftz.f32 %f51, %f238, %f239;
cvt.u64.u32 %rd89, %r764;
add.s64 %rd90, %rd89, %rd15;
shl.b64 %rd92, %rd90, 2;
add.s64 %rd18, %rd63, %rd92;
@%p14 bra $L__BB11_24;

st.global.f32 [%rd18], %f51;

$L__BB11_24:
ld.shared.f32 %f240, [%r30+1152];
add.ftz.f32 %f241, %f240, %f313;
mov.b32 %r782, %f241;
shfl.sync.bfly.b32 %r786|%p65, %r782, %r768, %r767, %r769;
mov.b32 %f242, %r786;
add.ftz.f32 %f243, %f241, %f242;
mov.b32 %r787, %f243;
shfl.sync.bfly.b32 %r789|%p66, %r787, %r772, %r767, %r769;
mov.b32 %f244, %r789;
add.ftz.f32 %f245, %f243, %f244;
mov.b32 %r790, %f245;
shfl.sync.bfly.b32 %r792|%p67, %r790, %r775, %r767, %r769;
mov.b32 %f246, %r792;
add.ftz.f32 %f247, %f245, %f246;
mov.b32 %r793, %f247;
shfl.sync.bfly.b32 %r795|%p68, %r793, %r765, %r767, %r769;
mov.b32 %f248, %r795;
add.ftz.f32 %f249, %f247, %f248;
mov.b32 %r796, %f249;
shfl.sync.bfly.b32 %r798|%p69, %r796, %r780, %r767, %r769;
mov.b32 %f250, %r798;
add.ftz.f32 %f52, %f249, %f250;
@%p22 bra $L__BB11_26;

st.global.f32 [%rd18+4], %f52;

$L__BB11_26:
mul.lo.s32 %r32, %r35, 5;
ld.shared.f32 %f251, [%r30+1280];
add.ftz.f32 %f252, %f251, %f312;
mov.b32 %r799, %f252;
mov.u32 %r800, 31;
mov.u32 %r801, 16;
mov.u32 %r802, -1;
shfl.sync.bfly.b32 %r803|%p71, %r799, %r801, %r800, %r802;
mov.b32 %f253, %r803;
add.ftz.f32 %f254, %f252, %f253;
mov.b32 %r804, %f254;
mov.u32 %r805, 8;
shfl.sync.bfly.b32 %r806|%p72, %r804, %r805, %r800, %r802;
mov.b32 %f255, %r806;
add.ftz.f32 %f256, %f254, %f255;
mov.b32 %r807, %f256;
mov.u32 %r808, 4;
shfl.sync.bfly.b32 %r809|%p73, %r807, %r808, %r800, %r802;
mov.b32 %f257, %r809;
add.ftz.f32 %f258, %f256, %f257;
mov.b32 %r810, %f258;
mov.u32 %r811, 2;
shfl.sync.bfly.b32 %r812|%p74, %r810, %r811, %r800, %r802;
mov.b32 %f259, %r812;
add.ftz.f32 %f260, %f258, %f259;
mov.b32 %r813, %f260;
mov.u32 %r814, 1;
shfl.sync.bfly.b32 %r815|%p75, %r813, %r814, %r800, %r802;
mov.b32 %f261, %r815;
add.ftz.f32 %f53, %f260, %f261;
@%p14 bra $L__BB11_28;

cvt.u64.u32 %rd93, %r32;
add.s64 %rd94, %rd93, %rd15;
shl.b64 %rd96, %rd94, 2;
add.s64 %rd97, %rd63, %rd96;
st.global.f32 [%rd97], %f53;

$L__BB11_28:
ld.shared.f32 %f262, [%r30+1408];
add.ftz.f32 %f263, %f262, %f311;
mov.b32 %r816, %f263;
shfl.sync.bfly.b32 %r820|%p77, %r816, %r801, %r800, %r802;
mov.b32 %f264, %r820;
add.ftz.f32 %f265, %f263, %f264;
mov.b32 %r821, %f265;
shfl.sync.bfly.b32 %r823|%p78, %r821, %r805, %r800, %r802;
mov.b32 %f266, %r823;
add.ftz.f32 %f267, %f265, %f266;
mov.b32 %r824, %f267;
shfl.sync.bfly.b32 %r826|%p79, %r824, %r808, %r800, %r802;
mov.b32 %f268, %r826;
add.ftz.f32 %f269, %f267, %f268;
mov.b32 %r827, %f269;
shfl.sync.bfly.b32 %r829|%p80, %r827, %r811, %r800, %r802;
mov.b32 %f270, %r829;
add.ftz.f32 %f271, %f269, %f270;
mov.b32 %r830, %f271;
shfl.sync.bfly.b32 %r832|%p81, %r830, %r814, %r800, %r802;
mov.b32 %f272, %r832;
add.ftz.f32 %f54, %f271, %f272;
@%p22 bra $L__BB11_30;

add.s32 %r833, %r32, 1;
cvt.u64.u32 %rd98, %r833;
add.s64 %rd99, %rd98, %rd15;
shl.b64 %rd101, %rd99, 2;
add.s64 %rd102, %rd63, %rd101;
st.global.f32 [%rd102], %f54;

$L__BB11_30:
mul.lo.s32 %r834, %r35, 6;
ld.shared.f32 %f273, [%r30+1536];
add.ftz.f32 %f274, %f273, %f310;
mov.b32 %r835, %f274;
mov.u32 %r836, 31;
mov.u32 %r837, 16;
mov.u32 %r838, -1;
shfl.sync.bfly.b32 %r839|%p83, %r835, %r837, %r836, %r838;
mov.b32 %f275, %r839;
add.ftz.f32 %f276, %f274, %f275;
mov.b32 %r840, %f276;
mov.u32 %r841, 8;
shfl.sync.bfly.b32 %r842|%p84, %r840, %r841, %r836, %r838;
mov.b32 %f277, %r842;
add.ftz.f32 %f278, %f276, %f277;
mov.b32 %r843, %f278;
mov.u32 %r844, 4;
shfl.sync.bfly.b32 %r845|%p85, %r843, %r844, %r836, %r838;
mov.b32 %f279, %r845;
add.ftz.f32 %f280, %f278, %f279;
mov.b32 %r846, %f280;
mov.u32 %r847, 2;
shfl.sync.bfly.b32 %r848|%p86, %r846, %r847, %r836, %r838;
mov.b32 %f281, %r848;
add.ftz.f32 %f282, %f280, %f281;
mov.b32 %r849, %f282;
mov.u32 %r850, 1;
shfl.sync.bfly.b32 %r851|%p87, %r849, %r850, %r836, %r838;
mov.b32 %f283, %r851;
add.ftz.f32 %f55, %f282, %f283;
cvt.u64.u32 %rd103, %r834;
add.s64 %rd104, %rd103, %rd15;
shl.b64 %rd106, %rd104, 2;
add.s64 %rd19, %rd63, %rd106;
@%p14 bra $L__BB11_32;

st.global.f32 [%rd19], %f55;

$L__BB11_32:
ld.shared.f32 %f284, [%r30+1664];
add.ftz.f32 %f285, %f284, %f309;
mov.b32 %r852, %f285;
shfl.sync.bfly.b32 %r856|%p89, %r852, %r837, %r836, %r838;
mov.b32 %f286, %r856;
add.ftz.f32 %f287, %f285, %f286;
mov.b32 %r857, %f287;
shfl.sync.bfly.b32 %r859|%p90, %r857, %r841, %r836, %r838;
mov.b32 %f288, %r859;
add.ftz.f32 %f289, %f287, %f288;
mov.b32 %r860, %f289;
shfl.sync.bfly.b32 %r862|%p91, %r860, %r844, %r836, %r838;
mov.b32 %f290, %r862;
add.ftz.f32 %f291, %f289, %f290;
mov.b32 %r863, %f291;
shfl.sync.bfly.b32 %r865|%p92, %r863, %r847, %r836, %r838;
mov.b32 %f292, %r865;
add.ftz.f32 %f293, %f291, %f292;
mov.b32 %r866, %f293;
shfl.sync.bfly.b32 %r868|%p93, %r866, %r850, %r836, %r838;
mov.b32 %f294, %r868;
add.ftz.f32 %f56, %f293, %f294;
@%p22 bra $L__BB11_34;

st.global.f32 [%rd19+4], %f56;

$L__BB11_34:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type41ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type41ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type41ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type41ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type41ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type41ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type41ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type41ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type41ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type41ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<107>;
.reg .b16 %rs<21>;
.reg .f32 %f<369>;
.reg .b32 %r<984>;
.reg .b64 %rd<125>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type41ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[2048];

ld.param.u64 %rd20, [_Z13mul_mat_vec_qIL9ggml_type41ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd21, [_Z13mul_mat_vec_qIL9ggml_type41ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd22, [_Z13mul_mat_vec_qIL9ggml_type41ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r52, [_Z13mul_mat_vec_qIL9ggml_type41ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type41ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type41ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type41ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type41ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type41ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r45, [_Z13mul_mat_vec_qIL9ggml_type41ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r49, [_Z13mul_mat_vec_qIL9ggml_type41ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r50, [_Z13mul_mat_vec_qIL9ggml_type41ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r51, [_Z13mul_mat_vec_qIL9ggml_type41ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r47, [_Z13mul_mat_vec_qIL9ggml_type41ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r46, [_Z13mul_mat_vec_qIL9ggml_type41ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type41ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type41ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r53, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r53, %r2;
mov.u32 %r54, %ctaid.x;
shl.b32 %r4, %r54, 1;
shr.s32 %r55, %r3, 31;
shr.u32 %r56, %r55, 30;
add.s32 %r57, %r3, %r56;
shr.s32 %r983, %r57, 2;
shr.u32 %r6, %r52, 7;
setp.ge.s32 %p3, %r983, %r6;
mov.f32 %f353, 0f00000000;
mov.f32 %f354, %f353;
mov.f32 %f355, %f353;
mov.f32 %f356, %f353;
mov.f32 %f357, %f353;
mov.f32 %f358, %f353;
mov.f32 %f359, %f353;
mov.f32 %f360, %f353;
mov.f32 %f361, %f353;
mov.f32 %f362, %f353;
mov.f32 %f363, %f353;
mov.f32 %f364, %f353;
mov.f32 %f365, %f353;
mov.f32 %f366, %f353;
mov.f32 %f367, %f353;
mov.f32 %f368, %f353;
@%p3 bra $L__BB12_3;

mov.u32 %r58, %ctaid.y;
mov.u32 %r59, %ctaid.z;
mul.hi.u32 %r60, %r59, %r46;
add.s32 %r61, %r60, %r59;
shr.u32 %r62, %r61, %r47;
mul.lo.s32 %r63, %r62, %r49;
mul.hi.u32 %r64, %r58, %r40;
add.s32 %r65, %r64, %r58;
shr.u32 %r66, %r65, %r41;
mul.lo.s32 %r67, %r66, %r43;
mul.lo.s32 %r68, %r58, %r44;
cvt.u64.u32 %rd23, %r68;
add.s32 %r69, %r4, 1;
and.b32 %r73, %r57, -4;
sub.s32 %r74, %r3, %r73;
shl.b32 %r75, %r74, 1;
cvt.s64.s32 %rd24, %r74;
add.s64 %rd25, %rd23, %rd24;
mul.lo.s32 %r76, %r59, %r50;
cvt.u64.u32 %rd26, %r76;
add.s64 %rd1, %rd25, %rd26;
shl.b32 %r77, %r38, 1;
shl.b32 %r975, %r983, 2;
add.s32 %r982, %r975, %r77;
mad.lo.s32 %r981, %r38, 3, %r975;
add.s32 %r78, %r983, %r38;
shl.b32 %r980, %r78, 2;
mad.lo.s32 %r979, %r38, 5, %r975;
mad.lo.s32 %r978, %r38, 6, %r975;
mad.lo.s32 %r977, %r38, 7, %r975;
add.s32 %r976, %r38, %r975;
mad.lo.s32 %r79, %r4, %r37, %r983;
add.s32 %r80, %r79, %r63;
add.s32 %r81, %r80, %r67;
mul.wide.s32 %rd27, %r81, 18;
mul.wide.s32 %rd28, %r75, 2;
add.s64 %rd29, %rd27, %rd28;
cvta.to.global.u64 %rd30, %rd20;
add.s64 %rd31, %rd30, %rd29;
add.s64 %rd124, %rd31, 4;
add.s32 %r82, %r983, %r63;
add.s32 %r83, %r82, %r67;
mad.lo.s32 %r84, %r69, %r37, %r83;
mul.wide.s32 %rd32, %r84, 18;
add.s64 %rd33, %rd32, %rd28;
add.s64 %rd34, %rd30, %rd33;
add.s64 %rd123, %rd34, 4;
add.s64 %rd122, %rd30, %rd32;
add.s64 %rd121, %rd30, %rd27;
cvta.to.global.u64 %rd6, %rd21;
mov.f32 %f353, 0f00000000;

$L__BB12_2:
cvt.u64.u32 %rd35, %r975;
add.s64 %rd36, %rd1, %rd35;
mul.lo.s64 %rd37, %rd36, 36;
ld.global.nc.u16 %rs1, [%rd121];

{ cvt.f32.f16 %f97, %rs1;}


ld.global.nc.u16 %rs17, [%rd124+-2];
cvt.s32.s16 %r613, %rs17;
add.s64 %rd38, %rd6, %rd37;
ld.global.nc.u32 %r87, [%rd38+4];
ld.global.nc.u32 %r91, [%rd38+8];
ld.global.nc.u32 %r95, [%rd38+12];
ld.global.nc.u32 %r99, [%rd38+16];
and.b32 %r614, %r613, 30583;
mov.u32 %r615, 286261504;
prmt.b32 %r616, %r615, %r615, %r614;
shr.u32 %r617, %r613, 2;
and.b32 %r618, %r617, 30583;
prmt.b32 %r619, %r615, %r615, %r618;
and.b32 %r620, %r616, 30583;
mov.u32 %r621, 511;
prmt.b32 %r622, %r621, %r621, %r620;
and.b32 %r623, %r619, 30583;
prmt.b32 %r624, %r621, %r621, %r623;
shr.u32 %r625, %r616, 16;
and.b32 %r626, %r625, 30583;
prmt.b32 %r627, %r621, %r621, %r626;
shr.u32 %r628, %r619, 16;
and.b32 %r629, %r628, 30583;
prmt.b32 %r630, %r621, %r621, %r629;
mov.u32 %r631, 21520;
prmt.b32 %r548, %r622, %r624, %r631;
mov.u32 %r632, 30258;
prmt.b32 %r552, %r622, %r624, %r632;
prmt.b32 %r556, %r627, %r630, %r631;
prmt.b32 %r560, %r627, %r630, %r632;
mov.u32 %r583, 0;

dp4a.s32.s32 %r85, %r548, %r87, %r583;


dp4a.s32.s32 %r89, %r552, %r91, %r85;


dp4a.s32.s32 %r93, %r556, %r95, %r89;


dp4a.s32.s32 %r97, %r560, %r99, %r93;

ld.global.nc.u16 %rs18, [%rd124];
cvt.s32.s16 %r633, %rs18;
ld.global.nc.u32 %r103, [%rd38+20];
ld.global.nc.u32 %r107, [%rd38+24];
ld.global.nc.u32 %r111, [%rd38+28];
ld.global.nc.u32 %r115, [%rd38+32];
and.b32 %r634, %r633, 30583;
prmt.b32 %r635, %r615, %r615, %r634;
shr.u32 %r636, %r633, 2;
and.b32 %r637, %r636, 30583;
prmt.b32 %r638, %r615, %r615, %r637;
and.b32 %r639, %r635, 30583;
prmt.b32 %r640, %r621, %r621, %r639;
and.b32 %r641, %r638, 30583;
prmt.b32 %r642, %r621, %r621, %r641;
shr.u32 %r643, %r635, 16;
and.b32 %r644, %r643, 30583;
prmt.b32 %r645, %r621, %r621, %r644;
shr.u32 %r646, %r638, 16;
and.b32 %r647, %r646, 30583;
prmt.b32 %r648, %r621, %r621, %r647;
prmt.b32 %r564, %r640, %r642, %r631;
prmt.b32 %r568, %r640, %r642, %r632;
prmt.b32 %r572, %r645, %r648, %r631;
prmt.b32 %r576, %r645, %r648, %r632;

dp4a.s32.s32 %r101, %r564, %r103, %r97;


dp4a.s32.s32 %r105, %r568, %r107, %r101;


dp4a.s32.s32 %r109, %r572, %r111, %r105;


dp4a.s32.s32 %r113, %r576, %r115, %r109;

ld.global.nc.u32 %r117, [%rd38];

{.reg .f16 low,high;
mov.b32 {low,high},%r117;
cvt.f32.f16 %f98, low;}


mul.ftz.f32 %f129, %f97, %f98;
cvt.rn.f32.s32 %f130, %r113;
fma.rn.ftz.f32 %f368, %f129, %f130, %f368;
ld.global.nc.u16 %rs2, [%rd122];

{ cvt.f32.f16 %f99, %rs2;}


ld.global.nc.u16 %rs19, [%rd123+-2];
cvt.s32.s16 %r649, %rs19;
and.b32 %r650, %r649, 30583;
prmt.b32 %r651, %r615, %r615, %r650;
shr.u32 %r652, %r649, 2;
and.b32 %r653, %r652, 30583;
prmt.b32 %r654, %r615, %r615, %r653;
and.b32 %r655, %r651, 30583;
prmt.b32 %r656, %r621, %r621, %r655;
and.b32 %r657, %r654, 30583;
prmt.b32 %r658, %r621, %r621, %r657;
shr.u32 %r659, %r651, 16;
and.b32 %r660, %r659, 30583;
prmt.b32 %r661, %r621, %r621, %r660;
shr.u32 %r662, %r654, 16;
and.b32 %r663, %r662, 30583;
prmt.b32 %r664, %r621, %r621, %r663;
prmt.b32 %r581, %r656, %r658, %r631;
prmt.b32 %r585, %r656, %r658, %r632;
prmt.b32 %r589, %r661, %r664, %r631;
prmt.b32 %r593, %r661, %r664, %r632;

dp4a.s32.s32 %r118, %r581, %r87, %r583;


dp4a.s32.s32 %r122, %r585, %r91, %r118;


dp4a.s32.s32 %r126, %r589, %r95, %r122;


dp4a.s32.s32 %r130, %r593, %r99, %r126;

ld.global.nc.u16 %rs20, [%rd123];
cvt.s32.s16 %r665, %rs20;
and.b32 %r666, %r665, 30583;
prmt.b32 %r667, %r615, %r615, %r666;
shr.u32 %r668, %r665, 2;
and.b32 %r669, %r668, 30583;
prmt.b32 %r670, %r615, %r615, %r669;
and.b32 %r671, %r667, 30583;
prmt.b32 %r672, %r621, %r621, %r671;
and.b32 %r673, %r670, 30583;
prmt.b32 %r674, %r621, %r621, %r673;
shr.u32 %r675, %r667, 16;
and.b32 %r676, %r675, 30583;
prmt.b32 %r677, %r621, %r621, %r676;
shr.u32 %r678, %r670, 16;
and.b32 %r679, %r678, 30583;
prmt.b32 %r680, %r621, %r621, %r679;
prmt.b32 %r597, %r672, %r674, %r631;
prmt.b32 %r601, %r672, %r674, %r632;
prmt.b32 %r605, %r677, %r680, %r631;
prmt.b32 %r609, %r677, %r680, %r632;

dp4a.s32.s32 %r134, %r597, %r103, %r130;


dp4a.s32.s32 %r138, %r601, %r107, %r134;


dp4a.s32.s32 %r142, %r605, %r111, %r138;


dp4a.s32.s32 %r146, %r609, %r115, %r142;


{.reg .f16 low,high;
mov.b32 {low,high},%r117;
cvt.f32.f16 %f100, low;}


mul.ftz.f32 %f131, %f99, %f100;
cvt.rn.f32.s32 %f132, %r146;
fma.rn.ftz.f32 %f367, %f131, %f132, %f367;
cvt.u64.u32 %rd39, %r976;
add.s64 %rd40, %rd1, %rd39;
mul.lo.s64 %rd41, %rd40, 36;

{ cvt.f32.f16 %f101, %rs1;}


add.s64 %rd42, %rd6, %rd41;
ld.global.nc.u32 %r153, [%rd42+4];
ld.global.nc.u32 %r157, [%rd42+8];
ld.global.nc.u32 %r161, [%rd42+12];
ld.global.nc.u32 %r165, [%rd42+16];

dp4a.s32.s32 %r151, %r548, %r153, %r583;


dp4a.s32.s32 %r155, %r552, %r157, %r151;


dp4a.s32.s32 %r159, %r556, %r161, %r155;


dp4a.s32.s32 %r163, %r560, %r165, %r159;

ld.global.nc.u32 %r169, [%rd42+20];
ld.global.nc.u32 %r173, [%rd42+24];
ld.global.nc.u32 %r177, [%rd42+28];
ld.global.nc.u32 %r181, [%rd42+32];

dp4a.s32.s32 %r167, %r564, %r169, %r163;


dp4a.s32.s32 %r171, %r568, %r173, %r167;


dp4a.s32.s32 %r175, %r572, %r177, %r171;


dp4a.s32.s32 %r179, %r576, %r181, %r175;

ld.global.nc.u32 %r183, [%rd42];

{.reg .f16 low,high;
mov.b32 {low,high},%r183;
cvt.f32.f16 %f102, low;}


mul.ftz.f32 %f133, %f101, %f102;
cvt.rn.f32.s32 %f134, %r179;
fma.rn.ftz.f32 %f366, %f133, %f134, %f366;

{ cvt.f32.f16 %f103, %rs2;}



dp4a.s32.s32 %r184, %r581, %r153, %r583;


dp4a.s32.s32 %r188, %r585, %r157, %r184;


dp4a.s32.s32 %r192, %r589, %r161, %r188;


dp4a.s32.s32 %r196, %r593, %r165, %r192;


dp4a.s32.s32 %r200, %r597, %r169, %r196;


dp4a.s32.s32 %r204, %r601, %r173, %r200;


dp4a.s32.s32 %r208, %r605, %r177, %r204;


dp4a.s32.s32 %r212, %r609, %r181, %r208;


{.reg .f16 low,high;
mov.b32 {low,high},%r183;
cvt.f32.f16 %f104, low;}


mul.ftz.f32 %f135, %f103, %f104;
cvt.rn.f32.s32 %f136, %r212;
fma.rn.ftz.f32 %f365, %f135, %f136, %f365;
cvt.u64.u32 %rd43, %r982;
add.s64 %rd44, %rd1, %rd43;
mul.lo.s64 %rd45, %rd44, 36;

{ cvt.f32.f16 %f105, %rs1;}


add.s64 %rd46, %rd6, %rd45;
ld.global.nc.u32 %r219, [%rd46+4];
ld.global.nc.u32 %r223, [%rd46+8];
ld.global.nc.u32 %r227, [%rd46+12];
ld.global.nc.u32 %r231, [%rd46+16];

dp4a.s32.s32 %r217, %r548, %r219, %r583;


dp4a.s32.s32 %r221, %r552, %r223, %r217;


dp4a.s32.s32 %r225, %r556, %r227, %r221;


dp4a.s32.s32 %r229, %r560, %r231, %r225;

ld.global.nc.u32 %r235, [%rd46+20];
ld.global.nc.u32 %r239, [%rd46+24];
ld.global.nc.u32 %r243, [%rd46+28];
ld.global.nc.u32 %r247, [%rd46+32];

dp4a.s32.s32 %r233, %r564, %r235, %r229;


dp4a.s32.s32 %r237, %r568, %r239, %r233;


dp4a.s32.s32 %r241, %r572, %r243, %r237;


dp4a.s32.s32 %r245, %r576, %r247, %r241;

ld.global.nc.u32 %r249, [%rd46];

{.reg .f16 low,high;
mov.b32 {low,high},%r249;
cvt.f32.f16 %f106, low;}


mul.ftz.f32 %f137, %f105, %f106;
cvt.rn.f32.s32 %f138, %r245;
fma.rn.ftz.f32 %f364, %f137, %f138, %f364;

{ cvt.f32.f16 %f107, %rs2;}



dp4a.s32.s32 %r250, %r581, %r219, %r583;


dp4a.s32.s32 %r254, %r585, %r223, %r250;


dp4a.s32.s32 %r258, %r589, %r227, %r254;


dp4a.s32.s32 %r262, %r593, %r231, %r258;


dp4a.s32.s32 %r266, %r597, %r235, %r262;


dp4a.s32.s32 %r270, %r601, %r239, %r266;


dp4a.s32.s32 %r274, %r605, %r243, %r270;


dp4a.s32.s32 %r278, %r609, %r247, %r274;


{.reg .f16 low,high;
mov.b32 {low,high},%r249;
cvt.f32.f16 %f108, low;}


mul.ftz.f32 %f139, %f107, %f108;
cvt.rn.f32.s32 %f140, %r278;
fma.rn.ftz.f32 %f363, %f139, %f140, %f363;
cvt.u64.u32 %rd47, %r981;
add.s64 %rd48, %rd1, %rd47;
mul.lo.s64 %rd49, %rd48, 36;

{ cvt.f32.f16 %f109, %rs1;}


add.s64 %rd50, %rd6, %rd49;
ld.global.nc.u32 %r285, [%rd50+4];
ld.global.nc.u32 %r289, [%rd50+8];
ld.global.nc.u32 %r293, [%rd50+12];
ld.global.nc.u32 %r297, [%rd50+16];

dp4a.s32.s32 %r283, %r548, %r285, %r583;


dp4a.s32.s32 %r287, %r552, %r289, %r283;


dp4a.s32.s32 %r291, %r556, %r293, %r287;


dp4a.s32.s32 %r295, %r560, %r297, %r291;

ld.global.nc.u32 %r301, [%rd50+20];
ld.global.nc.u32 %r305, [%rd50+24];
ld.global.nc.u32 %r309, [%rd50+28];
ld.global.nc.u32 %r313, [%rd50+32];

dp4a.s32.s32 %r299, %r564, %r301, %r295;


dp4a.s32.s32 %r303, %r568, %r305, %r299;


dp4a.s32.s32 %r307, %r572, %r309, %r303;


dp4a.s32.s32 %r311, %r576, %r313, %r307;

ld.global.nc.u32 %r315, [%rd50];

{.reg .f16 low,high;
mov.b32 {low,high},%r315;
cvt.f32.f16 %f110, low;}


mul.ftz.f32 %f141, %f109, %f110;
cvt.rn.f32.s32 %f142, %r311;
fma.rn.ftz.f32 %f362, %f141, %f142, %f362;

{ cvt.f32.f16 %f111, %rs2;}



dp4a.s32.s32 %r316, %r581, %r285, %r583;


dp4a.s32.s32 %r320, %r585, %r289, %r316;


dp4a.s32.s32 %r324, %r589, %r293, %r320;


dp4a.s32.s32 %r328, %r593, %r297, %r324;


dp4a.s32.s32 %r332, %r597, %r301, %r328;


dp4a.s32.s32 %r336, %r601, %r305, %r332;


dp4a.s32.s32 %r340, %r605, %r309, %r336;


dp4a.s32.s32 %r344, %r609, %r313, %r340;


{.reg .f16 low,high;
mov.b32 {low,high},%r315;
cvt.f32.f16 %f112, low;}


mul.ftz.f32 %f143, %f111, %f112;
cvt.rn.f32.s32 %f144, %r344;
fma.rn.ftz.f32 %f361, %f143, %f144, %f361;
cvt.u64.u32 %rd51, %r980;
add.s64 %rd52, %rd1, %rd51;
mul.lo.s64 %rd53, %rd52, 36;

{ cvt.f32.f16 %f113, %rs1;}


add.s64 %rd54, %rd6, %rd53;
ld.global.nc.u32 %r351, [%rd54+4];
ld.global.nc.u32 %r355, [%rd54+8];
ld.global.nc.u32 %r359, [%rd54+12];
ld.global.nc.u32 %r363, [%rd54+16];

dp4a.s32.s32 %r349, %r548, %r351, %r583;


dp4a.s32.s32 %r353, %r552, %r355, %r349;


dp4a.s32.s32 %r357, %r556, %r359, %r353;


dp4a.s32.s32 %r361, %r560, %r363, %r357;

ld.global.nc.u32 %r367, [%rd54+20];
ld.global.nc.u32 %r371, [%rd54+24];
ld.global.nc.u32 %r375, [%rd54+28];
ld.global.nc.u32 %r379, [%rd54+32];

dp4a.s32.s32 %r365, %r564, %r367, %r361;


dp4a.s32.s32 %r369, %r568, %r371, %r365;


dp4a.s32.s32 %r373, %r572, %r375, %r369;


dp4a.s32.s32 %r377, %r576, %r379, %r373;

ld.global.nc.u32 %r381, [%rd54];

{.reg .f16 low,high;
mov.b32 {low,high},%r381;
cvt.f32.f16 %f114, low;}


mul.ftz.f32 %f145, %f113, %f114;
cvt.rn.f32.s32 %f146, %r377;
fma.rn.ftz.f32 %f360, %f145, %f146, %f360;

{ cvt.f32.f16 %f115, %rs2;}



dp4a.s32.s32 %r382, %r581, %r351, %r583;


dp4a.s32.s32 %r386, %r585, %r355, %r382;


dp4a.s32.s32 %r390, %r589, %r359, %r386;


dp4a.s32.s32 %r394, %r593, %r363, %r390;


dp4a.s32.s32 %r398, %r597, %r367, %r394;


dp4a.s32.s32 %r402, %r601, %r371, %r398;


dp4a.s32.s32 %r406, %r605, %r375, %r402;


dp4a.s32.s32 %r410, %r609, %r379, %r406;


{.reg .f16 low,high;
mov.b32 {low,high},%r381;
cvt.f32.f16 %f116, low;}


mul.ftz.f32 %f147, %f115, %f116;
cvt.rn.f32.s32 %f148, %r410;
fma.rn.ftz.f32 %f359, %f147, %f148, %f359;
cvt.u64.u32 %rd55, %r979;
add.s64 %rd56, %rd1, %rd55;
mul.lo.s64 %rd57, %rd56, 36;

{ cvt.f32.f16 %f117, %rs1;}


add.s64 %rd58, %rd6, %rd57;
ld.global.nc.u32 %r417, [%rd58+4];
ld.global.nc.u32 %r421, [%rd58+8];
ld.global.nc.u32 %r425, [%rd58+12];
ld.global.nc.u32 %r429, [%rd58+16];

dp4a.s32.s32 %r415, %r548, %r417, %r583;


dp4a.s32.s32 %r419, %r552, %r421, %r415;


dp4a.s32.s32 %r423, %r556, %r425, %r419;


dp4a.s32.s32 %r427, %r560, %r429, %r423;

ld.global.nc.u32 %r433, [%rd58+20];
ld.global.nc.u32 %r437, [%rd58+24];
ld.global.nc.u32 %r441, [%rd58+28];
ld.global.nc.u32 %r445, [%rd58+32];

dp4a.s32.s32 %r431, %r564, %r433, %r427;


dp4a.s32.s32 %r435, %r568, %r437, %r431;


dp4a.s32.s32 %r439, %r572, %r441, %r435;


dp4a.s32.s32 %r443, %r576, %r445, %r439;

ld.global.nc.u32 %r447, [%rd58];

{.reg .f16 low,high;
mov.b32 {low,high},%r447;
cvt.f32.f16 %f118, low;}


mul.ftz.f32 %f149, %f117, %f118;
cvt.rn.f32.s32 %f150, %r443;
fma.rn.ftz.f32 %f358, %f149, %f150, %f358;

{ cvt.f32.f16 %f119, %rs2;}



dp4a.s32.s32 %r448, %r581, %r417, %r583;


dp4a.s32.s32 %r452, %r585, %r421, %r448;


dp4a.s32.s32 %r456, %r589, %r425, %r452;


dp4a.s32.s32 %r460, %r593, %r429, %r456;


dp4a.s32.s32 %r464, %r597, %r433, %r460;


dp4a.s32.s32 %r468, %r601, %r437, %r464;


dp4a.s32.s32 %r472, %r605, %r441, %r468;


dp4a.s32.s32 %r476, %r609, %r445, %r472;


{.reg .f16 low,high;
mov.b32 {low,high},%r447;
cvt.f32.f16 %f120, low;}


mul.ftz.f32 %f151, %f119, %f120;
cvt.rn.f32.s32 %f152, %r476;
fma.rn.ftz.f32 %f357, %f151, %f152, %f357;
cvt.u64.u32 %rd59, %r978;
add.s64 %rd60, %rd1, %rd59;
mul.lo.s64 %rd61, %rd60, 36;

{ cvt.f32.f16 %f121, %rs1;}


add.s64 %rd62, %rd6, %rd61;
ld.global.nc.u32 %r483, [%rd62+4];
ld.global.nc.u32 %r487, [%rd62+8];
ld.global.nc.u32 %r491, [%rd62+12];
ld.global.nc.u32 %r495, [%rd62+16];

dp4a.s32.s32 %r481, %r548, %r483, %r583;


dp4a.s32.s32 %r485, %r552, %r487, %r481;


dp4a.s32.s32 %r489, %r556, %r491, %r485;


dp4a.s32.s32 %r493, %r560, %r495, %r489;

ld.global.nc.u32 %r499, [%rd62+20];
ld.global.nc.u32 %r503, [%rd62+24];
ld.global.nc.u32 %r507, [%rd62+28];
ld.global.nc.u32 %r511, [%rd62+32];

dp4a.s32.s32 %r497, %r564, %r499, %r493;


dp4a.s32.s32 %r501, %r568, %r503, %r497;


dp4a.s32.s32 %r505, %r572, %r507, %r501;


dp4a.s32.s32 %r509, %r576, %r511, %r505;

ld.global.nc.u32 %r513, [%rd62];

{.reg .f16 low,high;
mov.b32 {low,high},%r513;
cvt.f32.f16 %f122, low;}


mul.ftz.f32 %f153, %f121, %f122;
cvt.rn.f32.s32 %f154, %r509;
fma.rn.ftz.f32 %f356, %f153, %f154, %f356;

{ cvt.f32.f16 %f123, %rs2;}



dp4a.s32.s32 %r514, %r581, %r483, %r583;


dp4a.s32.s32 %r518, %r585, %r487, %r514;


dp4a.s32.s32 %r522, %r589, %r491, %r518;


dp4a.s32.s32 %r526, %r593, %r495, %r522;


dp4a.s32.s32 %r530, %r597, %r499, %r526;


dp4a.s32.s32 %r534, %r601, %r503, %r530;


dp4a.s32.s32 %r538, %r605, %r507, %r534;


dp4a.s32.s32 %r542, %r609, %r511, %r538;


{.reg .f16 low,high;
mov.b32 {low,high},%r513;
cvt.f32.f16 %f124, low;}


mul.ftz.f32 %f155, %f123, %f124;
cvt.rn.f32.s32 %f156, %r542;
fma.rn.ftz.f32 %f355, %f155, %f156, %f355;
cvt.u64.u32 %rd63, %r977;
add.s64 %rd64, %rd1, %rd63;
mul.lo.s64 %rd65, %rd64, 36;

{ cvt.f32.f16 %f125, %rs1;}


add.s64 %rd66, %rd6, %rd65;
ld.global.nc.u32 %r549, [%rd66+4];
ld.global.nc.u32 %r553, [%rd66+8];
ld.global.nc.u32 %r557, [%rd66+12];
ld.global.nc.u32 %r561, [%rd66+16];

dp4a.s32.s32 %r547, %r548, %r549, %r583;


dp4a.s32.s32 %r551, %r552, %r553, %r547;


dp4a.s32.s32 %r555, %r556, %r557, %r551;


dp4a.s32.s32 %r559, %r560, %r561, %r555;

ld.global.nc.u32 %r565, [%rd66+20];
ld.global.nc.u32 %r569, [%rd66+24];
ld.global.nc.u32 %r573, [%rd66+28];
ld.global.nc.u32 %r577, [%rd66+32];

dp4a.s32.s32 %r563, %r564, %r565, %r559;


dp4a.s32.s32 %r567, %r568, %r569, %r563;


dp4a.s32.s32 %r571, %r572, %r573, %r567;


dp4a.s32.s32 %r575, %r576, %r577, %r571;

ld.global.nc.u32 %r579, [%rd66];

{.reg .f16 low,high;
mov.b32 {low,high},%r579;
cvt.f32.f16 %f126, low;}


mul.ftz.f32 %f157, %f125, %f126;
cvt.rn.f32.s32 %f158, %r575;
fma.rn.ftz.f32 %f354, %f157, %f158, %f354;

{ cvt.f32.f16 %f127, %rs2;}



dp4a.s32.s32 %r580, %r581, %r549, %r583;


dp4a.s32.s32 %r584, %r585, %r553, %r580;


dp4a.s32.s32 %r588, %r589, %r557, %r584;


dp4a.s32.s32 %r592, %r593, %r561, %r588;


dp4a.s32.s32 %r596, %r597, %r565, %r592;


dp4a.s32.s32 %r600, %r601, %r569, %r596;


dp4a.s32.s32 %r604, %r605, %r573, %r600;


dp4a.s32.s32 %r608, %r609, %r577, %r604;


{.reg .f16 low,high;
mov.b32 {low,high},%r579;
cvt.f32.f16 %f128, low;}


mul.ftz.f32 %f159, %f127, %f128;
cvt.rn.f32.s32 %f160, %r608;
fma.rn.ftz.f32 %f353, %f159, %f160, %f353;
add.s32 %r982, %r982, 64;
add.s32 %r981, %r981, 64;
add.s32 %r980, %r980, 64;
add.s32 %r979, %r979, 64;
add.s32 %r978, %r978, 64;
add.s32 %r977, %r977, 64;
add.s32 %r976, %r976, 64;
add.s32 %r975, %r975, 64;
add.s64 %rd124, %rd124, 288;
add.s64 %rd123, %rd123, 288;
add.s64 %rd122, %rd122, 288;
add.s64 %rd121, %rd121, 288;
add.s32 %r983, %r983, 16;
setp.lt.s32 %p4, %r983, %r6;
@%p4 bra $L__BB12_2;

$L__BB12_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB12_5;

shl.b32 %r681, %r1, 11;
mov.u32 %r682, _ZZ13mul_mat_vec_qIL9ggml_type41ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r683, %r682, %r681;
shl.b32 %r684, %r2, 2;
add.s32 %r685, %r683, %r684;
st.shared.f32 [%r685+-2048], %f368;
st.shared.f32 [%r685+-1920], %f367;
st.shared.f32 [%r685+-1792], %f366;
st.shared.f32 [%r685+-1664], %f365;
st.shared.f32 [%r685+-1536], %f364;
st.shared.f32 [%r685+-1408], %f363;
st.shared.f32 [%r685+-1280], %f362;
st.shared.f32 [%r685+-1152], %f361;
st.shared.f32 [%r685+-1024], %f360;
st.shared.f32 [%r685+-896], %f359;
st.shared.f32 [%r685+-768], %f358;
st.shared.f32 [%r685+-640], %f357;
st.shared.f32 [%r685+-512], %f356;
st.shared.f32 [%r685+-384], %f355;
st.shared.f32 [%r685+-256], %f354;
st.shared.f32 [%r685+-128], %f353;

$L__BB12_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB12_38;

mov.u32 %r686, %ctaid.z;
mov.u32 %r687, %ctaid.y;
mad.lo.s32 %r688, %r687, %r45, %r4;
mad.lo.s32 %r689, %r686, %r51, %r688;
cvt.u64.u32 %rd15, %r689;
shl.b32 %r690, %r2, 2;
mov.u32 %r691, _ZZ13mul_mat_vec_qIL9ggml_type41ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r33, %r691, %r690;
mov.u32 %r692, 2;
ld.shared.f32 %f161, [%r33];
add.ftz.f32 %f162, %f161, %f368;
mov.b32 %r693, %f162;
mov.u32 %r694, 31;
mov.u32 %r695, 16;
mov.u32 %r696, -1;
shfl.sync.bfly.b32 %r697|%p7, %r693, %r695, %r694, %r696;
mov.b32 %f163, %r697;
add.ftz.f32 %f164, %f162, %f163;
mov.b32 %r698, %f164;
mov.u32 %r699, 8;
shfl.sync.bfly.b32 %r700|%p8, %r698, %r699, %r694, %r696;
mov.b32 %f165, %r700;
add.ftz.f32 %f166, %f164, %f165;
mov.b32 %r701, %f166;
mov.u32 %r702, 4;
shfl.sync.bfly.b32 %r703|%p9, %r701, %r702, %r694, %r696;
mov.b32 %f167, %r703;
add.ftz.f32 %f168, %f166, %f167;
mov.b32 %r704, %f168;
shfl.sync.bfly.b32 %r705|%p10, %r704, %r692, %r694, %r696;
mov.b32 %f169, %r705;
add.ftz.f32 %f170, %f168, %f169;
mov.b32 %r706, %f170;
mov.u32 %r707, 1;
shfl.sync.bfly.b32 %r708|%p11, %r706, %r707, %r694, %r696;
mov.b32 %f171, %r708;
add.ftz.f32 %f49, %f170, %f171;
setp.lt.u32 %p12, %r4, %r39;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
cvta.to.global.u64 %rd67, %rd22;
mul.wide.u32 %rd68, %r689, 4;
add.s64 %rd16, %rd67, %rd68;
not.pred %p14, %p1;
@%p14 bra $L__BB12_8;

st.global.f32 [%rd16], %f49;

$L__BB12_8:
ld.shared.f32 %f172, [%r33+128];
add.ftz.f32 %f173, %f172, %f367;
mov.b32 %r709, %f173;
shfl.sync.bfly.b32 %r713|%p15, %r709, %r695, %r694, %r696;
mov.b32 %f174, %r713;
add.ftz.f32 %f175, %f173, %f174;
mov.b32 %r714, %f175;
shfl.sync.bfly.b32 %r716|%p16, %r714, %r699, %r694, %r696;
mov.b32 %f176, %r716;
add.ftz.f32 %f177, %f175, %f176;
mov.b32 %r717, %f177;
shfl.sync.bfly.b32 %r719|%p17, %r717, %r702, %r694, %r696;
mov.b32 %f178, %r719;
add.ftz.f32 %f179, %f177, %f178;
mov.b32 %r720, %f179;
shfl.sync.bfly.b32 %r722|%p18, %r720, %r692, %r694, %r696;
mov.b32 %f180, %r722;
add.ftz.f32 %f181, %f179, %f180;
mov.b32 %r723, %f181;
shfl.sync.bfly.b32 %r725|%p19, %r723, %r707, %r694, %r696;
mov.b32 %f182, %r725;
add.ftz.f32 %f50, %f181, %f182;
setp.eq.s32 %p20, %r2, 1;
or.b32 %r729, %r4, 1;
setp.lt.u32 %p21, %r729, %r39;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB12_10;

st.global.f32 [%rd16+4], %f50;

$L__BB12_10:
ld.shared.f32 %f183, [%r33+256];
add.ftz.f32 %f184, %f183, %f366;
mov.b32 %r730, %f184;
mov.u32 %r731, 31;
mov.u32 %r732, 16;
mov.u32 %r733, -1;
shfl.sync.bfly.b32 %r734|%p23, %r730, %r732, %r731, %r733;
mov.b32 %f185, %r734;
add.ftz.f32 %f186, %f184, %f185;
mov.b32 %r735, %f186;
mov.u32 %r736, 8;
shfl.sync.bfly.b32 %r737|%p24, %r735, %r736, %r731, %r733;
mov.b32 %f187, %r737;
add.ftz.f32 %f188, %f186, %f187;
mov.b32 %r738, %f188;
mov.u32 %r739, 4;
shfl.sync.bfly.b32 %r740|%p25, %r738, %r739, %r731, %r733;
mov.b32 %f189, %r740;
add.ftz.f32 %f190, %f188, %f189;
mov.b32 %r741, %f190;
mov.u32 %r742, 2;
shfl.sync.bfly.b32 %r743|%p26, %r741, %r742, %r731, %r733;
mov.b32 %f191, %r743;
add.ftz.f32 %f192, %f190, %f191;
mov.b32 %r744, %f192;
mov.u32 %r745, 1;
shfl.sync.bfly.b32 %r746|%p27, %r744, %r745, %r731, %r733;
mov.b32 %f193, %r746;
add.ftz.f32 %f51, %f192, %f193;
@%p14 bra $L__BB12_12;

cvt.u64.u32 %rd69, %r39;
add.s64 %rd70, %rd69, %rd15;
shl.b64 %rd72, %rd70, 2;
add.s64 %rd73, %rd67, %rd72;
st.global.f32 [%rd73], %f51;

$L__BB12_12:
ld.shared.f32 %f194, [%r33+384];
add.ftz.f32 %f195, %f194, %f365;
mov.b32 %r747, %f195;
shfl.sync.bfly.b32 %r751|%p29, %r747, %r732, %r731, %r733;
mov.b32 %f196, %r751;
add.ftz.f32 %f197, %f195, %f196;
mov.b32 %r752, %f197;
shfl.sync.bfly.b32 %r754|%p30, %r752, %r736, %r731, %r733;
mov.b32 %f198, %r754;
add.ftz.f32 %f199, %f197, %f198;
mov.b32 %r755, %f199;
shfl.sync.bfly.b32 %r757|%p31, %r755, %r739, %r731, %r733;
mov.b32 %f200, %r757;
add.ftz.f32 %f201, %f199, %f200;
mov.b32 %r758, %f201;
shfl.sync.bfly.b32 %r760|%p32, %r758, %r742, %r731, %r733;
mov.b32 %f202, %r760;
add.ftz.f32 %f203, %f201, %f202;
mov.b32 %r761, %f203;
shfl.sync.bfly.b32 %r763|%p33, %r761, %r745, %r731, %r733;
mov.b32 %f204, %r763;
add.ftz.f32 %f52, %f203, %f204;
@%p22 bra $L__BB12_14;

add.s32 %r764, %r39, 1;
cvt.u64.u32 %rd74, %r764;
add.s64 %rd75, %rd74, %rd15;
shl.b64 %rd77, %rd75, 2;
add.s64 %rd78, %rd67, %rd77;
st.global.f32 [%rd78], %f52;

$L__BB12_14:
shl.b32 %r765, %r39, 1;
mov.u32 %r766, 1;
ld.shared.f32 %f205, [%r33+512];
add.ftz.f32 %f206, %f205, %f364;
mov.b32 %r767, %f206;
mov.u32 %r768, 31;
mov.u32 %r769, 16;
mov.u32 %r770, -1;
shfl.sync.bfly.b32 %r771|%p35, %r767, %r769, %r768, %r770;
mov.b32 %f207, %r771;
add.ftz.f32 %f208, %f206, %f207;
mov.b32 %r772, %f208;
mov.u32 %r773, 8;
shfl.sync.bfly.b32 %r774|%p36, %r772, %r773, %r768, %r770;
mov.b32 %f209, %r774;
add.ftz.f32 %f210, %f208, %f209;
mov.b32 %r775, %f210;
mov.u32 %r776, 4;
shfl.sync.bfly.b32 %r777|%p37, %r775, %r776, %r768, %r770;
mov.b32 %f211, %r777;
add.ftz.f32 %f212, %f210, %f211;
mov.b32 %r778, %f212;
mov.u32 %r779, 2;
shfl.sync.bfly.b32 %r780|%p38, %r778, %r779, %r768, %r770;
mov.b32 %f213, %r780;
add.ftz.f32 %f214, %f212, %f213;
mov.b32 %r781, %f214;
shfl.sync.bfly.b32 %r782|%p39, %r781, %r766, %r768, %r770;
mov.b32 %f215, %r782;
add.ftz.f32 %f53, %f214, %f215;
cvt.u64.u32 %rd79, %r765;
add.s64 %rd80, %rd79, %rd15;
shl.b64 %rd82, %rd80, 2;
add.s64 %rd17, %rd67, %rd82;
@%p14 bra $L__BB12_16;

st.global.f32 [%rd17], %f53;

$L__BB12_16:
ld.shared.f32 %f216, [%r33+640];
add.ftz.f32 %f217, %f216, %f363;
mov.b32 %r783, %f217;
shfl.sync.bfly.b32 %r787|%p41, %r783, %r769, %r768, %r770;
mov.b32 %f218, %r787;
add.ftz.f32 %f219, %f217, %f218;
mov.b32 %r788, %f219;
shfl.sync.bfly.b32 %r790|%p42, %r788, %r773, %r768, %r770;
mov.b32 %f220, %r790;
add.ftz.f32 %f221, %f219, %f220;
mov.b32 %r791, %f221;
shfl.sync.bfly.b32 %r793|%p43, %r791, %r776, %r768, %r770;
mov.b32 %f222, %r793;
add.ftz.f32 %f223, %f221, %f222;
mov.b32 %r794, %f223;
shfl.sync.bfly.b32 %r796|%p44, %r794, %r779, %r768, %r770;
mov.b32 %f224, %r796;
add.ftz.f32 %f225, %f223, %f224;
mov.b32 %r797, %f225;
shfl.sync.bfly.b32 %r799|%p45, %r797, %r766, %r768, %r770;
mov.b32 %f226, %r799;
add.ftz.f32 %f54, %f225, %f226;
@%p22 bra $L__BB12_18;

st.global.f32 [%rd17+4], %f54;

$L__BB12_18:
mul.lo.s32 %r34, %r39, 3;
ld.shared.f32 %f227, [%r33+768];
add.ftz.f32 %f228, %f227, %f362;
mov.b32 %r800, %f228;
mov.u32 %r801, 31;
mov.u32 %r802, 16;
mov.u32 %r803, -1;
shfl.sync.bfly.b32 %r804|%p47, %r800, %r802, %r801, %r803;
mov.b32 %f229, %r804;
add.ftz.f32 %f230, %f228, %f229;
mov.b32 %r805, %f230;
mov.u32 %r806, 8;
shfl.sync.bfly.b32 %r807|%p48, %r805, %r806, %r801, %r803;
mov.b32 %f231, %r807;
add.ftz.f32 %f232, %f230, %f231;
mov.b32 %r808, %f232;
mov.u32 %r809, 4;
shfl.sync.bfly.b32 %r810|%p49, %r808, %r809, %r801, %r803;
mov.b32 %f233, %r810;
add.ftz.f32 %f234, %f232, %f233;
mov.b32 %r811, %f234;
mov.u32 %r812, 2;
shfl.sync.bfly.b32 %r813|%p50, %r811, %r812, %r801, %r803;
mov.b32 %f235, %r813;
add.ftz.f32 %f236, %f234, %f235;
mov.b32 %r814, %f236;
mov.u32 %r815, 1;
shfl.sync.bfly.b32 %r816|%p51, %r814, %r815, %r801, %r803;
mov.b32 %f237, %r816;
add.ftz.f32 %f55, %f236, %f237;
@%p14 bra $L__BB12_20;

cvt.u64.u32 %rd83, %r34;
add.s64 %rd84, %rd83, %rd15;
shl.b64 %rd86, %rd84, 2;
add.s64 %rd87, %rd67, %rd86;
st.global.f32 [%rd87], %f55;

$L__BB12_20:
ld.shared.f32 %f238, [%r33+896];
add.ftz.f32 %f239, %f238, %f361;
mov.b32 %r817, %f239;
shfl.sync.bfly.b32 %r821|%p53, %r817, %r802, %r801, %r803;
mov.b32 %f240, %r821;
add.ftz.f32 %f241, %f239, %f240;
mov.b32 %r822, %f241;
shfl.sync.bfly.b32 %r824|%p54, %r822, %r806, %r801, %r803;
mov.b32 %f242, %r824;
add.ftz.f32 %f243, %f241, %f242;
mov.b32 %r825, %f243;
shfl.sync.bfly.b32 %r827|%p55, %r825, %r809, %r801, %r803;
mov.b32 %f244, %r827;
add.ftz.f32 %f245, %f243, %f244;
mov.b32 %r828, %f245;
shfl.sync.bfly.b32 %r830|%p56, %r828, %r812, %r801, %r803;
mov.b32 %f246, %r830;
add.ftz.f32 %f247, %f245, %f246;
mov.b32 %r831, %f247;
shfl.sync.bfly.b32 %r833|%p57, %r831, %r815, %r801, %r803;
mov.b32 %f248, %r833;
add.ftz.f32 %f56, %f247, %f248;
@%p22 bra $L__BB12_22;

add.s32 %r834, %r34, 1;
cvt.u64.u32 %rd88, %r834;
add.s64 %rd89, %rd88, %rd15;
shl.b64 %rd91, %rd89, 2;
add.s64 %rd92, %rd67, %rd91;
st.global.f32 [%rd92], %f56;

$L__BB12_22:
shl.b32 %r835, %r39, 2;
mov.u32 %r836, 2;
ld.shared.f32 %f249, [%r33+1024];
add.ftz.f32 %f250, %f249, %f360;
mov.b32 %r837, %f250;
mov.u32 %r838, 31;
mov.u32 %r839, 16;
mov.u32 %r840, -1;
shfl.sync.bfly.b32 %r841|%p59, %r837, %r839, %r838, %r840;
mov.b32 %f251, %r841;
add.ftz.f32 %f252, %f250, %f251;
mov.b32 %r842, %f252;
mov.u32 %r843, 8;
shfl.sync.bfly.b32 %r844|%p60, %r842, %r843, %r838, %r840;
mov.b32 %f253, %r844;
add.ftz.f32 %f254, %f252, %f253;
mov.b32 %r845, %f254;
mov.u32 %r846, 4;
shfl.sync.bfly.b32 %r847|%p61, %r845, %r846, %r838, %r840;
mov.b32 %f255, %r847;
add.ftz.f32 %f256, %f254, %f255;
mov.b32 %r848, %f256;
shfl.sync.bfly.b32 %r849|%p62, %r848, %r836, %r838, %r840;
mov.b32 %f257, %r849;
add.ftz.f32 %f258, %f256, %f257;
mov.b32 %r850, %f258;
mov.u32 %r851, 1;
shfl.sync.bfly.b32 %r852|%p63, %r850, %r851, %r838, %r840;
mov.b32 %f259, %r852;
add.ftz.f32 %f57, %f258, %f259;
cvt.u64.u32 %rd93, %r835;
add.s64 %rd94, %rd93, %rd15;
shl.b64 %rd96, %rd94, 2;
add.s64 %rd18, %rd67, %rd96;
@%p14 bra $L__BB12_24;

st.global.f32 [%rd18], %f57;

$L__BB12_24:
ld.shared.f32 %f260, [%r33+1152];
add.ftz.f32 %f261, %f260, %f359;
mov.b32 %r853, %f261;
shfl.sync.bfly.b32 %r857|%p65, %r853, %r839, %r838, %r840;
mov.b32 %f262, %r857;
add.ftz.f32 %f263, %f261, %f262;
mov.b32 %r858, %f263;
shfl.sync.bfly.b32 %r860|%p66, %r858, %r843, %r838, %r840;
mov.b32 %f264, %r860;
add.ftz.f32 %f265, %f263, %f264;
mov.b32 %r861, %f265;
shfl.sync.bfly.b32 %r863|%p67, %r861, %r846, %r838, %r840;
mov.b32 %f266, %r863;
add.ftz.f32 %f267, %f265, %f266;
mov.b32 %r864, %f267;
shfl.sync.bfly.b32 %r866|%p68, %r864, %r836, %r838, %r840;
mov.b32 %f268, %r866;
add.ftz.f32 %f269, %f267, %f268;
mov.b32 %r867, %f269;
shfl.sync.bfly.b32 %r869|%p69, %r867, %r851, %r838, %r840;
mov.b32 %f270, %r869;
add.ftz.f32 %f58, %f269, %f270;
@%p22 bra $L__BB12_26;

st.global.f32 [%rd18+4], %f58;

$L__BB12_26:
mul.lo.s32 %r35, %r39, 5;
ld.shared.f32 %f271, [%r33+1280];
add.ftz.f32 %f272, %f271, %f358;
mov.b32 %r870, %f272;
mov.u32 %r871, 31;
mov.u32 %r872, 16;
mov.u32 %r873, -1;
shfl.sync.bfly.b32 %r874|%p71, %r870, %r872, %r871, %r873;
mov.b32 %f273, %r874;
add.ftz.f32 %f274, %f272, %f273;
mov.b32 %r875, %f274;
mov.u32 %r876, 8;
shfl.sync.bfly.b32 %r877|%p72, %r875, %r876, %r871, %r873;
mov.b32 %f275, %r877;
add.ftz.f32 %f276, %f274, %f275;
mov.b32 %r878, %f276;
mov.u32 %r879, 4;
shfl.sync.bfly.b32 %r880|%p73, %r878, %r879, %r871, %r873;
mov.b32 %f277, %r880;
add.ftz.f32 %f278, %f276, %f277;
mov.b32 %r881, %f278;
mov.u32 %r882, 2;
shfl.sync.bfly.b32 %r883|%p74, %r881, %r882, %r871, %r873;
mov.b32 %f279, %r883;
add.ftz.f32 %f280, %f278, %f279;
mov.b32 %r884, %f280;
mov.u32 %r885, 1;
shfl.sync.bfly.b32 %r886|%p75, %r884, %r885, %r871, %r873;
mov.b32 %f281, %r886;
add.ftz.f32 %f59, %f280, %f281;
@%p14 bra $L__BB12_28;

cvt.u64.u32 %rd97, %r35;
add.s64 %rd98, %rd97, %rd15;
shl.b64 %rd100, %rd98, 2;
add.s64 %rd101, %rd67, %rd100;
st.global.f32 [%rd101], %f59;

$L__BB12_28:
ld.shared.f32 %f282, [%r33+1408];
add.ftz.f32 %f283, %f282, %f357;
mov.b32 %r887, %f283;
shfl.sync.bfly.b32 %r891|%p77, %r887, %r872, %r871, %r873;
mov.b32 %f284, %r891;
add.ftz.f32 %f285, %f283, %f284;
mov.b32 %r892, %f285;
shfl.sync.bfly.b32 %r894|%p78, %r892, %r876, %r871, %r873;
mov.b32 %f286, %r894;
add.ftz.f32 %f287, %f285, %f286;
mov.b32 %r895, %f287;
shfl.sync.bfly.b32 %r897|%p79, %r895, %r879, %r871, %r873;
mov.b32 %f288, %r897;
add.ftz.f32 %f289, %f287, %f288;
mov.b32 %r898, %f289;
shfl.sync.bfly.b32 %r900|%p80, %r898, %r882, %r871, %r873;
mov.b32 %f290, %r900;
add.ftz.f32 %f291, %f289, %f290;
mov.b32 %r901, %f291;
shfl.sync.bfly.b32 %r903|%p81, %r901, %r885, %r871, %r873;
mov.b32 %f292, %r903;
add.ftz.f32 %f60, %f291, %f292;
@%p22 bra $L__BB12_30;

add.s32 %r904, %r35, 1;
cvt.u64.u32 %rd102, %r904;
add.s64 %rd103, %rd102, %rd15;
shl.b64 %rd105, %rd103, 2;
add.s64 %rd106, %rd67, %rd105;
st.global.f32 [%rd106], %f60;

$L__BB12_30:
mul.lo.s32 %r905, %r39, 6;
ld.shared.f32 %f293, [%r33+1536];
add.ftz.f32 %f294, %f293, %f356;
mov.b32 %r906, %f294;
mov.u32 %r907, 31;
mov.u32 %r908, 16;
mov.u32 %r909, -1;
shfl.sync.bfly.b32 %r910|%p83, %r906, %r908, %r907, %r909;
mov.b32 %f295, %r910;
add.ftz.f32 %f296, %f294, %f295;
mov.b32 %r911, %f296;
mov.u32 %r912, 8;
shfl.sync.bfly.b32 %r913|%p84, %r911, %r912, %r907, %r909;
mov.b32 %f297, %r913;
add.ftz.f32 %f298, %f296, %f297;
mov.b32 %r914, %f298;
mov.u32 %r915, 4;
shfl.sync.bfly.b32 %r916|%p85, %r914, %r915, %r907, %r909;
mov.b32 %f299, %r916;
add.ftz.f32 %f300, %f298, %f299;
mov.b32 %r917, %f300;
mov.u32 %r918, 2;
shfl.sync.bfly.b32 %r919|%p86, %r917, %r918, %r907, %r909;
mov.b32 %f301, %r919;
add.ftz.f32 %f302, %f300, %f301;
mov.b32 %r920, %f302;
mov.u32 %r921, 1;
shfl.sync.bfly.b32 %r922|%p87, %r920, %r921, %r907, %r909;
mov.b32 %f303, %r922;
add.ftz.f32 %f61, %f302, %f303;
cvt.u64.u32 %rd107, %r905;
add.s64 %rd108, %rd107, %rd15;
shl.b64 %rd110, %rd108, 2;
add.s64 %rd19, %rd67, %rd110;
@%p14 bra $L__BB12_32;

st.global.f32 [%rd19], %f61;

$L__BB12_32:
ld.shared.f32 %f304, [%r33+1664];
add.ftz.f32 %f305, %f304, %f355;
mov.b32 %r923, %f305;
shfl.sync.bfly.b32 %r927|%p89, %r923, %r908, %r907, %r909;
mov.b32 %f306, %r927;
add.ftz.f32 %f307, %f305, %f306;
mov.b32 %r928, %f307;
shfl.sync.bfly.b32 %r930|%p90, %r928, %r912, %r907, %r909;
mov.b32 %f308, %r930;
add.ftz.f32 %f309, %f307, %f308;
mov.b32 %r931, %f309;
shfl.sync.bfly.b32 %r933|%p91, %r931, %r915, %r907, %r909;
mov.b32 %f310, %r933;
add.ftz.f32 %f311, %f309, %f310;
mov.b32 %r934, %f311;
shfl.sync.bfly.b32 %r936|%p92, %r934, %r918, %r907, %r909;
mov.b32 %f312, %r936;
add.ftz.f32 %f313, %f311, %f312;
mov.b32 %r937, %f313;
shfl.sync.bfly.b32 %r939|%p93, %r937, %r921, %r907, %r909;
mov.b32 %f314, %r939;
add.ftz.f32 %f62, %f313, %f314;
@%p22 bra $L__BB12_34;

st.global.f32 [%rd19+4], %f62;

$L__BB12_34:
mul.lo.s32 %r36, %r39, 7;
ld.shared.f32 %f315, [%r33+1792];
add.ftz.f32 %f316, %f315, %f354;
mov.b32 %r940, %f316;
mov.u32 %r941, 31;
mov.u32 %r942, 16;
mov.u32 %r943, -1;
shfl.sync.bfly.b32 %r944|%p95, %r940, %r942, %r941, %r943;
mov.b32 %f317, %r944;
add.ftz.f32 %f318, %f316, %f317;
mov.b32 %r945, %f318;
mov.u32 %r946, 8;
shfl.sync.bfly.b32 %r947|%p96, %r945, %r946, %r941, %r943;
mov.b32 %f319, %r947;
add.ftz.f32 %f320, %f318, %f319;
mov.b32 %r948, %f320;
mov.u32 %r949, 4;
shfl.sync.bfly.b32 %r950|%p97, %r948, %r949, %r941, %r943;
mov.b32 %f321, %r950;
add.ftz.f32 %f322, %f320, %f321;
mov.b32 %r951, %f322;
mov.u32 %r952, 2;
shfl.sync.bfly.b32 %r953|%p98, %r951, %r952, %r941, %r943;
mov.b32 %f323, %r953;
add.ftz.f32 %f324, %f322, %f323;
mov.b32 %r954, %f324;
mov.u32 %r955, 1;
shfl.sync.bfly.b32 %r956|%p99, %r954, %r955, %r941, %r943;
mov.b32 %f325, %r956;
add.ftz.f32 %f63, %f324, %f325;
@%p14 bra $L__BB12_36;

cvt.u64.u32 %rd111, %r36;
add.s64 %rd112, %rd111, %rd15;
shl.b64 %rd114, %rd112, 2;
add.s64 %rd115, %rd67, %rd114;
st.global.f32 [%rd115], %f63;

$L__BB12_36:
ld.shared.f32 %f326, [%r33+1920];
add.ftz.f32 %f327, %f326, %f353;
mov.b32 %r957, %f327;
shfl.sync.bfly.b32 %r961|%p101, %r957, %r942, %r941, %r943;
mov.b32 %f328, %r961;
add.ftz.f32 %f329, %f327, %f328;
mov.b32 %r962, %f329;
shfl.sync.bfly.b32 %r964|%p102, %r962, %r946, %r941, %r943;
mov.b32 %f330, %r964;
add.ftz.f32 %f331, %f329, %f330;
mov.b32 %r965, %f331;
shfl.sync.bfly.b32 %r967|%p103, %r965, %r949, %r941, %r943;
mov.b32 %f332, %r967;
add.ftz.f32 %f333, %f331, %f332;
mov.b32 %r968, %f333;
shfl.sync.bfly.b32 %r970|%p104, %r968, %r952, %r941, %r943;
mov.b32 %f334, %r970;
add.ftz.f32 %f335, %f333, %f334;
mov.b32 %r971, %f335;
shfl.sync.bfly.b32 %r973|%p105, %r971, %r955, %r941, %r943;
mov.b32 %f336, %r973;
add.ftz.f32 %f64, %f335, %f336;
@%p22 bra $L__BB12_38;

add.s32 %r974, %r36, 1;
cvt.u64.u32 %rd116, %r974;
add.s64 %rd117, %rd116, %rd15;
shl.b64 %rd119, %rd117, 2;
add.s64 %rd120, %rd67, %rd119;
st.global.f32 [%rd120], %f64;

$L__BB12_38:
ret;

}
.entry _Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj(
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2,
.param .align 8 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3[48],
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5,
.param .align 4 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6[12],
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15
)
.maxntid 256, 1, 1
.minnctapersm 1
{
.local .align 8 .b8 __local_depot13[16];
.reg .b64 %SP;
.reg .b64 %SPL;
.reg .pred %p<48>;
.reg .b16 %rs<13>;
.reg .f32 %f<169>;
.reg .b32 %r<397>;
.reg .b64 %rd<85>;


mov.u64 %SPL, __local_depot13;
ld.param.u64 %rd10, [_Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0];
ld.param.u64 %rd11, [_Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1];
ld.param.u64 %rd12, [_Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2];
ld.param.u32 %r21, [_Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5];
ld.param.u32 %r26, [_Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8];
ld.param.u32 %r27, [_Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9];
ld.param.u32 %r29, [_Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11];
ld.param.u32 %r30, [_Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12];
ld.param.u32 %r33, [_Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14];
ld.param.u32 %r32, [_Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15];
ld.param.u32 %r24, [_Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+8];
ld.param.u32 %r23, [_Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+4];
ld.param.u32 %r22, [_Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6];
ld.param.u64 %rd14, [_Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+8];
add.u64 %rd1, %SPL, 0;
add.u64 %rd2, %SPL, 8;
mov.u32 %r1, %tid.y;
setp.ge.u32 %p1, %r1, %r33;
@%p1 bra $L__BB13_29;

mov.u32 %r2, %ctaid.y;
cvta.to.global.u64 %rd21, %rd12;
mad.lo.s32 %r34, %r1, %r32, %r2;
mul.wide.u32 %rd22, %r34, 4;
add.s64 %rd23, %rd21, %rd22;
ld.global.nc.u32 %r3, [%rd23];
mov.f32 %f162, 0f00000000;
st.local.v2.f32 [%rd1], {%f162, %f162};
st.local.v2.f32 [%rd2], {%f162, %f162};
mov.u32 %r4, %tid.x;
shr.u32 %r396, %r4, 1;
shr.u32 %r35, %r21, 6;
setp.ge.u32 %p2, %r396, %r35;
mov.f32 %f163, %f162;
mov.f32 %f164, %f162;
mov.f32 %f165, %f162;
@%p2 bra $L__BB13_9;

mul.lo.s32 %r36, %r1, %r27;
cvt.u64.u32 %rd24, %r36;
and.b32 %r37, %r4, 1;
cvt.u64.u32 %rd25, %r37;
add.s64 %rd26, %rd24, %rd25;
mul.hi.u32 %r38, %r2, %r22;
add.s32 %r39, %r38, %r2;
shr.u32 %r40, %r39, %r23;
mul.lo.s32 %r41, %r40, %r24;
sub.s32 %r42, %r2, %r41;
mul.lo.s32 %r43, %r42, %r30;
cvt.u64.u32 %rd27, %r43;
add.s64 %rd4, %rd26, %rd27;
mov.f32 %f162, 0f00000000;
cvta.to.global.u64 %rd28, %rd10;
cvta.to.global.u64 %rd29, %rd11;

$L__BB13_3:
shl.b32 %r77, %r396, 1;
cvt.s64.s32 %rd30, %r77;
mad.lo.s32 %r78, %r3, %r29, %r396;
add.s64 %rd31, %rd4, %rd30;
mul.lo.s64 %rd32, %rd31, 36;
mov.u32 %r79, %ctaid.x;
shl.b32 %r80, %r79, 1;
mad.lo.s32 %r81, %r80, %r26, %r78;
cvt.s64.s32 %rd5, %r81;
mul.wide.s32 %rd33, %r81, 18;
add.s64 %rd34, %rd28, %rd33;
ld.global.nc.u16 %rs1, [%rd34];

{ cvt.f32.f16 %f42, %rs1;}


shl.b32 %r83, %r4, 2;
and.b32 %r84, %r83, 4;
mul.wide.u32 %rd35, %r84, 2;
add.s64 %rd36, %rd34, %rd35;
ld.global.nc.u16 %rs2, [%rd36+2];
cvt.s32.s16 %r85, %rs2;
add.s64 %rd37, %rd29, %rd32;
ld.global.nc.u32 %r7, [%rd37+4];
ld.global.nc.u32 %r8, [%rd37+8];
and.b32 %r86, %r85, 30583;
mov.u32 %r87, 33620223;
prmt.b32 %r88, %r87, %r87, %r86;
shr.u32 %r89, %r85, 2;
and.b32 %r90, %r89, 30583;
prmt.b32 %r91, %r87, %r87, %r90;
mov.u32 %r92, 20800;
prmt.b32 %r46, %r88, %r91, %r92;
mov.u32 %r93, 29538;
prmt.b32 %r50, %r88, %r91, %r93;
mov.u32 %r47, 0;

dp4a.s32.s32 %r44, %r7, %r46, %r47;


dp4a.s32.s32 %r48, %r8, %r50, %r44;

ld.global.nc.u16 %rs3, [%rd36+4];
cvt.s32.s16 %r94, %rs3;
ld.global.nc.u32 %r9, [%rd37+12];
ld.global.nc.u32 %r10, [%rd37+16];
and.b32 %r95, %r94, 30583;
prmt.b32 %r96, %r87, %r87, %r95;
shr.u32 %r97, %r94, 2;
and.b32 %r98, %r97, 30583;
prmt.b32 %r99, %r87, %r87, %r98;
prmt.b32 %r54, %r96, %r99, %r92;
prmt.b32 %r58, %r96, %r99, %r93;

dp4a.s32.s32 %r52, %r9, %r54, %r48;


dp4a.s32.s32 %r56, %r10, %r58, %r52;

ld.global.nc.u16 %rs4, [%rd36+6];
cvt.s32.s16 %r100, %rs4;
ld.global.nc.u32 %r11, [%rd37+20];
ld.global.nc.u32 %r12, [%rd37+24];
and.b32 %r101, %r100, 30583;
prmt.b32 %r102, %r87, %r87, %r101;
shr.u32 %r103, %r100, 2;
and.b32 %r104, %r103, 30583;
prmt.b32 %r105, %r87, %r87, %r104;
prmt.b32 %r62, %r102, %r105, %r92;
prmt.b32 %r66, %r102, %r105, %r93;

dp4a.s32.s32 %r60, %r11, %r62, %r56;


dp4a.s32.s32 %r64, %r12, %r66, %r60;

ld.global.nc.u16 %rs5, [%rd36+8];
cvt.s32.s16 %r106, %rs5;
ld.global.nc.u32 %r13, [%rd37+28];
ld.global.nc.u32 %r14, [%rd37+32];
and.b32 %r107, %r106, 30583;
prmt.b32 %r108, %r87, %r87, %r107;
shr.u32 %r109, %r106, 2;
and.b32 %r110, %r109, 30583;
prmt.b32 %r111, %r87, %r87, %r110;
prmt.b32 %r70, %r108, %r111, %r92;
prmt.b32 %r74, %r108, %r111, %r93;

dp4a.s32.s32 %r68, %r13, %r70, %r64;


dp4a.s32.s32 %r72, %r14, %r74, %r68;

ld.global.nc.u32 %r15, [%rd37];

{.reg .f16 low,high;
mov.b32 {low,high},%r15;
cvt.f32.f16 %f43, low;}


mul.ftz.f32 %f44, %f42, %f43;
cvt.rn.f32.s32 %f45, %r72;
fma.rn.ftz.f32 %f165, %f44, %f45, %f165;
setp.eq.s64 %p3, %rd14, 0;
@%p3 bra $L__BB13_5;

cvta.to.global.u64 %rd38, %rd14;
mul.lo.s64 %rd39, %rd5, 18;
add.s64 %rd40, %rd38, %rd39;
ld.global.u16 %rs6, [%rd40];

{ cvt.f32.f16 %f46, %rs6;}


add.s64 %rd42, %rd40, %rd35;
ld.global.s16 %r148, [%rd42+2];
and.b32 %r149, %r148, 30583;
prmt.b32 %r151, %r87, %r87, %r149;
shr.u32 %r152, %r148, 2;
and.b32 %r153, %r152, 30583;
prmt.b32 %r154, %r87, %r87, %r153;
prmt.b32 %r114, %r151, %r154, %r92;
prmt.b32 %r118, %r151, %r154, %r93;

dp4a.s32.s32 %r112, %r7, %r114, %r47;


dp4a.s32.s32 %r116, %r8, %r118, %r112;

ld.global.s16 %r157, [%rd42+4];
and.b32 %r158, %r157, 30583;
prmt.b32 %r159, %r87, %r87, %r158;
shr.u32 %r160, %r157, 2;
and.b32 %r161, %r160, 30583;
prmt.b32 %r162, %r87, %r87, %r161;
prmt.b32 %r122, %r159, %r162, %r92;
prmt.b32 %r126, %r159, %r162, %r93;

dp4a.s32.s32 %r120, %r9, %r122, %r116;


dp4a.s32.s32 %r124, %r10, %r126, %r120;

ld.global.s16 %r163, [%rd42+6];
and.b32 %r164, %r163, 30583;
prmt.b32 %r165, %r87, %r87, %r164;
shr.u32 %r166, %r163, 2;
and.b32 %r167, %r166, 30583;
prmt.b32 %r168, %r87, %r87, %r167;
prmt.b32 %r130, %r165, %r168, %r92;
prmt.b32 %r134, %r165, %r168, %r93;

dp4a.s32.s32 %r128, %r11, %r130, %r124;


dp4a.s32.s32 %r132, %r12, %r134, %r128;

ld.global.s16 %r169, [%rd42+8];
and.b32 %r170, %r169, 30583;
prmt.b32 %r171, %r87, %r87, %r170;
shr.u32 %r172, %r169, 2;
and.b32 %r173, %r172, 30583;
prmt.b32 %r174, %r87, %r87, %r173;
prmt.b32 %r138, %r171, %r174, %r92;
prmt.b32 %r142, %r171, %r174, %r93;

dp4a.s32.s32 %r136, %r13, %r138, %r132;


dp4a.s32.s32 %r140, %r14, %r142, %r136;


{.reg .f16 low,high;
mov.b32 {low,high},%r15;
cvt.f32.f16 %f47, low;}


mul.ftz.f32 %f48, %f46, %f47;
cvt.rn.f32.s32 %f49, %r140;
fma.rn.ftz.f32 %f163, %f48, %f49, %f163;

$L__BB13_5:
cvt.u32.u64 %r208, %rd5;
add.s32 %r209, %r208, %r26;
cvt.s64.s32 %rd6, %r209;
mul.wide.s32 %rd44, %r209, 18;
add.s64 %rd45, %rd28, %rd44;
ld.global.nc.u16 %rs7, [%rd45];

{ cvt.f32.f16 %f50, %rs7;}


add.s64 %rd47, %rd45, %rd35;
ld.global.nc.u16 %rs8, [%rd47+2];
cvt.s32.s16 %r213, %rs8;
and.b32 %r214, %r213, 30583;
prmt.b32 %r216, %r87, %r87, %r214;
shr.u32 %r217, %r213, 2;
and.b32 %r218, %r217, 30583;
prmt.b32 %r219, %r87, %r87, %r218;
prmt.b32 %r177, %r216, %r219, %r92;
prmt.b32 %r181, %r216, %r219, %r93;

dp4a.s32.s32 %r175, %r7, %r177, %r47;


dp4a.s32.s32 %r179, %r8, %r181, %r175;

ld.global.nc.u16 %rs9, [%rd47+4];
cvt.s32.s16 %r222, %rs9;
and.b32 %r223, %r222, 30583;
prmt.b32 %r224, %r87, %r87, %r223;
shr.u32 %r225, %r222, 2;
and.b32 %r226, %r225, 30583;
prmt.b32 %r227, %r87, %r87, %r226;
prmt.b32 %r185, %r224, %r227, %r92;
prmt.b32 %r189, %r224, %r227, %r93;

dp4a.s32.s32 %r183, %r9, %r185, %r179;


dp4a.s32.s32 %r187, %r10, %r189, %r183;

ld.global.nc.u16 %rs10, [%rd47+6];
cvt.s32.s16 %r228, %rs10;
and.b32 %r229, %r228, 30583;
prmt.b32 %r230, %r87, %r87, %r229;
shr.u32 %r231, %r228, 2;
and.b32 %r232, %r231, 30583;
prmt.b32 %r233, %r87, %r87, %r232;
prmt.b32 %r193, %r230, %r233, %r92;
prmt.b32 %r197, %r230, %r233, %r93;

dp4a.s32.s32 %r191, %r11, %r193, %r187;


dp4a.s32.s32 %r195, %r12, %r197, %r191;

ld.global.nc.u16 %rs11, [%rd47+8];
cvt.s32.s16 %r234, %rs11;
and.b32 %r235, %r234, 30583;
prmt.b32 %r236, %r87, %r87, %r235;
shr.u32 %r237, %r234, 2;
and.b32 %r238, %r237, 30583;
prmt.b32 %r239, %r87, %r87, %r238;
prmt.b32 %r201, %r236, %r239, %r92;
prmt.b32 %r205, %r236, %r239, %r93;

dp4a.s32.s32 %r199, %r13, %r201, %r195;


dp4a.s32.s32 %r203, %r14, %r205, %r199;


{.reg .f16 low,high;
mov.b32 {low,high},%r15;
cvt.f32.f16 %f51, low;}


mul.ftz.f32 %f52, %f50, %f51;
cvt.rn.f32.s32 %f53, %r203;
fma.rn.ftz.f32 %f164, %f52, %f53, %f164;
@%p3 bra $L__BB13_7;

cvta.to.global.u64 %rd48, %rd14;
mul.lo.s64 %rd49, %rd6, 18;
add.s64 %rd50, %rd48, %rd49;
ld.global.u16 %rs12, [%rd50];

{ cvt.f32.f16 %f54, %rs12;}


add.s64 %rd52, %rd50, %rd35;
ld.global.s16 %r276, [%rd52+2];
and.b32 %r277, %r276, 30583;
mov.u32 %r278, 33620223;
prmt.b32 %r279, %r278, %r278, %r277;
shr.u32 %r280, %r276, 2;
and.b32 %r281, %r280, 30583;
prmt.b32 %r282, %r278, %r278, %r281;
mov.u32 %r283, 20800;
prmt.b32 %r242, %r279, %r282, %r283;
mov.u32 %r284, 29538;
prmt.b32 %r246, %r279, %r282, %r284;
mov.u32 %r243, 0;

dp4a.s32.s32 %r240, %r7, %r242, %r243;


dp4a.s32.s32 %r244, %r8, %r246, %r240;

ld.global.s16 %r285, [%rd52+4];
and.b32 %r286, %r285, 30583;
prmt.b32 %r287, %r278, %r278, %r286;
shr.u32 %r288, %r285, 2;
and.b32 %r289, %r288, 30583;
prmt.b32 %r290, %r278, %r278, %r289;
prmt.b32 %r250, %r287, %r290, %r283;
prmt.b32 %r254, %r287, %r290, %r284;

dp4a.s32.s32 %r248, %r9, %r250, %r244;


dp4a.s32.s32 %r252, %r10, %r254, %r248;

ld.global.s16 %r291, [%rd52+6];
and.b32 %r292, %r291, 30583;
prmt.b32 %r293, %r278, %r278, %r292;
shr.u32 %r294, %r291, 2;
and.b32 %r295, %r294, 30583;
prmt.b32 %r296, %r278, %r278, %r295;
prmt.b32 %r258, %r293, %r296, %r283;
prmt.b32 %r262, %r293, %r296, %r284;

dp4a.s32.s32 %r256, %r11, %r258, %r252;


dp4a.s32.s32 %r260, %r12, %r262, %r256;

ld.global.s16 %r297, [%rd52+8];
and.b32 %r298, %r297, 30583;
prmt.b32 %r299, %r278, %r278, %r298;
shr.u32 %r300, %r297, 2;
and.b32 %r301, %r300, 30583;
prmt.b32 %r302, %r278, %r278, %r301;
prmt.b32 %r266, %r299, %r302, %r283;
prmt.b32 %r270, %r299, %r302, %r284;

dp4a.s32.s32 %r264, %r13, %r266, %r260;


dp4a.s32.s32 %r268, %r14, %r270, %r264;


{.reg .f16 low,high;
mov.b32 {low,high},%r15;
cvt.f32.f16 %f55, low;}


mul.ftz.f32 %f56, %f54, %f55;
cvt.rn.f32.s32 %f57, %r268;
fma.rn.ftz.f32 %f162, %f56, %f57, %f162;

$L__BB13_7:
add.s32 %r396, %r396, 16;
setp.lt.u32 %p5, %r396, %r35;
@%p5 bra $L__BB13_3;

add.u64 %rd73, %SPL, 8;
add.u64 %rd71, %SPL, 0;
st.local.v2.f32 [%rd71], {%f165, %f164};
st.local.v2.f32 [%rd73], {%f163, %f162};

$L__BB13_9:
mov.b32 %r17, %f165;
setp.ne.s64 %p6, %rd14, 0;
@%p6 bra $L__BB13_11;
bra.uni $L__BB13_10;

$L__BB13_11:
add.u64 %rd83, %SPL, 8;
add.u64 %rd81, %SPL, 0;
mov.u32 %r330, 31;
mov.u32 %r331, 16;
mov.u32 %r332, -1;
shfl.sync.bfly.b32 %r333|%p17, %r17, %r331, %r330, %r332;
mov.b32 %f78, %r333;
add.ftz.f32 %f79, %f165, %f78;
mov.b32 %r334, %f79;
mov.u32 %r335, 8;
shfl.sync.bfly.b32 %r336|%p18, %r334, %r335, %r330, %r332;
mov.b32 %f80, %r336;
add.ftz.f32 %f81, %f79, %f80;
mov.b32 %r337, %f81;
mov.u32 %r338, 4;
shfl.sync.bfly.b32 %r339|%p19, %r337, %r338, %r330, %r332;
mov.b32 %f82, %r339;
add.ftz.f32 %f83, %f81, %f82;
mov.b32 %r340, %f83;
mov.u32 %r341, 2;
shfl.sync.bfly.b32 %r342|%p20, %r340, %r341, %r330, %r332;
mov.b32 %f84, %r342;
add.ftz.f32 %f85, %f83, %f84;
mov.b32 %r343, %f85;
mov.u32 %r344, 1;
shfl.sync.bfly.b32 %r345|%p21, %r343, %r344, %r330, %r332;
mov.b32 %f86, %r345;
add.ftz.f32 %f87, %f85, %f86;
st.local.f32 [%rd81], %f87;
mov.b32 %r346, %f163;
shfl.sync.bfly.b32 %r347|%p22, %r346, %r331, %r330, %r332;
mov.b32 %f88, %r347;
add.ftz.f32 %f89, %f163, %f88;
mov.b32 %r348, %f89;
shfl.sync.bfly.b32 %r349|%p23, %r348, %r335, %r330, %r332;
mov.b32 %f90, %r349;
add.ftz.f32 %f91, %f89, %f90;
mov.b32 %r350, %f91;
shfl.sync.bfly.b32 %r351|%p24, %r350, %r338, %r330, %r332;
mov.b32 %f92, %r351;
add.ftz.f32 %f93, %f91, %f92;
mov.b32 %r352, %f93;
shfl.sync.bfly.b32 %r353|%p25, %r352, %r341, %r330, %r332;
mov.b32 %f94, %r353;
add.ftz.f32 %f95, %f93, %f94;
mov.b32 %r354, %f95;
shfl.sync.bfly.b32 %r355|%p26, %r354, %r344, %r330, %r332;
mov.b32 %f96, %r355;
add.ftz.f32 %f97, %f95, %f96;
st.local.f32 [%rd83], %f97;
mov.b32 %r356, %f164;
shfl.sync.bfly.b32 %r357|%p27, %r356, %r331, %r330, %r332;
mov.b32 %f98, %r357;
add.ftz.f32 %f99, %f164, %f98;
mov.b32 %r358, %f99;
shfl.sync.bfly.b32 %r359|%p28, %r358, %r335, %r330, %r332;
mov.b32 %f100, %r359;
add.ftz.f32 %f101, %f99, %f100;
mov.b32 %r360, %f101;
shfl.sync.bfly.b32 %r361|%p29, %r360, %r338, %r330, %r332;
mov.b32 %f102, %r361;
add.ftz.f32 %f103, %f101, %f102;
mov.b32 %r362, %f103;
shfl.sync.bfly.b32 %r363|%p30, %r362, %r341, %r330, %r332;
mov.b32 %f104, %r363;
add.ftz.f32 %f105, %f103, %f104;
mov.b32 %r364, %f105;
shfl.sync.bfly.b32 %r365|%p31, %r364, %r344, %r330, %r332;
mov.b32 %f106, %r365;
add.ftz.f32 %f107, %f105, %f106;
st.local.f32 [%rd81+4], %f107;
mov.b32 %r366, %f162;
shfl.sync.bfly.b32 %r367|%p32, %r366, %r331, %r330, %r332;
mov.b32 %f108, %r367;
add.ftz.f32 %f109, %f162, %f108;
mov.b32 %r368, %f109;
shfl.sync.bfly.b32 %r369|%p33, %r368, %r335, %r330, %r332;
mov.b32 %f110, %r369;
add.ftz.f32 %f111, %f109, %f110;
mov.b32 %r370, %f111;
shfl.sync.bfly.b32 %r371|%p34, %r370, %r338, %r330, %r332;
mov.b32 %f112, %r371;
add.ftz.f32 %f113, %f111, %f112;
mov.b32 %r372, %f113;
shfl.sync.bfly.b32 %r373|%p35, %r372, %r341, %r330, %r332;
mov.b32 %f114, %r373;
add.ftz.f32 %f115, %f113, %f114;
mov.b32 %r374, %f115;
shfl.sync.bfly.b32 %r375|%p36, %r374, %r344, %r330, %r332;
mov.b32 %f116, %r375;
add.ftz.f32 %f117, %f115, %f116;
st.local.f32 [%rd83+4], %f117;
bra.uni $L__BB13_12;

$L__BB13_10:
add.u64 %rd75, %SPL, 0;
mov.u32 %r304, 31;
mov.u32 %r305, 16;
mov.u32 %r306, -1;
shfl.sync.bfly.b32 %r307|%p7, %r17, %r305, %r304, %r306;
mov.b32 %f58, %r307;
add.ftz.f32 %f59, %f165, %f58;
mov.b32 %r308, %f59;
mov.u32 %r309, 8;
shfl.sync.bfly.b32 %r310|%p8, %r308, %r309, %r304, %r306;
mov.b32 %f60, %r310;
add.ftz.f32 %f61, %f59, %f60;
mov.b32 %r311, %f61;
mov.u32 %r312, 4;
shfl.sync.bfly.b32 %r313|%p9, %r311, %r312, %r304, %r306;
mov.b32 %f62, %r313;
add.ftz.f32 %f63, %f61, %f62;
mov.b32 %r314, %f63;
mov.u32 %r315, 2;
shfl.sync.bfly.b32 %r316|%p10, %r314, %r315, %r304, %r306;
mov.b32 %f64, %r316;
add.ftz.f32 %f65, %f63, %f64;
mov.b32 %r317, %f65;
mov.u32 %r318, 1;
shfl.sync.bfly.b32 %r319|%p11, %r317, %r318, %r304, %r306;
mov.b32 %f66, %r319;
add.ftz.f32 %f67, %f65, %f66;
st.local.f32 [%rd75], %f67;
mov.b32 %r320, %f164;
shfl.sync.bfly.b32 %r321|%p12, %r320, %r305, %r304, %r306;
mov.b32 %f68, %r321;
add.ftz.f32 %f69, %f164, %f68;
mov.b32 %r322, %f69;
shfl.sync.bfly.b32 %r323|%p13, %r322, %r309, %r304, %r306;
mov.b32 %f70, %r323;
add.ftz.f32 %f71, %f69, %f70;
mov.b32 %r324, %f71;
shfl.sync.bfly.b32 %r325|%p14, %r324, %r312, %r304, %r306;
mov.b32 %f72, %r325;
add.ftz.f32 %f73, %f71, %f72;
mov.b32 %r326, %f73;
shfl.sync.bfly.b32 %r327|%p15, %r326, %r315, %r304, %r306;
mov.b32 %f74, %r327;
add.ftz.f32 %f75, %f73, %f74;
mov.b32 %r328, %f75;
shfl.sync.bfly.b32 %r329|%p16, %r328, %r318, %r304, %r306;
mov.b32 %f76, %r329;
add.ftz.f32 %f77, %f75, %f76;
st.local.f32 [%rd75+4], %f77;

$L__BB13_12:
ld.param.u32 %r385, [_Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7];
setp.gt.u32 %p37, %r4, 1;
mov.u32 %r377, %ctaid.x;
shl.b32 %r378, %r377, 1;
add.s32 %r18, %r378, %r4;
setp.ge.u32 %p38, %r18, %r385;
or.pred %p39, %p37, %p38;
@%p39 bra $L__BB13_29;

add.u64 %rd77, %SPL, 0;
ld.param.u64 %rd66, [_Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3];
ld.param.u32 %r386, [_Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13];
cvt.u64.u32 %rd8, %r4;
mul.wide.u32 %rd53, %r4, 4;
add.s64 %rd54, %rd77, %rd53;
ld.local.f32 %f168, [%rd54];
mad.lo.s32 %r19, %r3, %r386, %r18;
setp.eq.s64 %p40, %rd66, 0;
@%p40 bra $L__BB13_15;

ld.param.u64 %rd67, [_Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3];
cvta.to.global.u64 %rd55, %rd67;
mul.wide.u32 %rd56, %r19, 4;
add.s64 %rd57, %rd55, %rd56;
ld.global.f32 %f118, [%rd57];
add.ftz.f32 %f168, %f168, %f118;

$L__BB13_15:
setp.eq.s64 %p41, %rd14, 0;
@%p41 bra $L__BB13_28;

add.u64 %rd79, %SPL, 8;
ld.param.u64 %rd69, [_Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+16];
shl.b64 %rd58, %rd8, 2;
add.s64 %rd59, %rd79, %rd58;
ld.local.f32 %f24, [%rd59];
setp.eq.s64 %p42, %rd69, 0;
@%p42 bra $L__BB13_18;

ld.param.u64 %rd70, [_Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+16];
cvta.to.global.u64 %rd60, %rd70;
mul.wide.u32 %rd61, %r19, 4;
add.s64 %rd62, %rd60, %rd61;
ld.global.f32 %f119, [%rd62];
add.ftz.f32 %f24, %f24, %f119;

$L__BB13_18:
ld.param.u32 %r391, [_Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.gt.s32 %p43, %r391, 2;
@%p43 bra $L__BB13_22;

ld.param.u32 %r394, [_Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.eq.s32 %p46, %r394, 1;
@%p46 bra $L__BB13_26;

ld.param.u32 %r395, [_Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.eq.s32 %p47, %r395, 2;
@%p47 bra $L__BB13_21;
bra.uni $L__BB13_27;

$L__BB13_21:
mul.ftz.f32 %f139, %f24, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f140, %f139;
add.ftz.f32 %f141, %f140, 0f3F800000;
div.approx.ftz.f32 %f142, %f24, %f141;
mul.ftz.f32 %f168, %f168, %f142;
bra.uni $L__BB13_28;

$L__BB13_22:
ld.param.u32 %r392, [_Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.eq.s32 %p44, %r392, 3;
@%p44 bra $L__BB13_25;

ld.param.u32 %r393, [_Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.ne.s32 %p45, %r393, 6;
@%p45 bra $L__BB13_27;

ld.param.f32 %f151, [_Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+44];
min.ftz.f32 %f120, %f24, %f151;
neg.ftz.f32 %f121, %f151;
min.ftz.f32 %f122, %f168, %f151;
max.ftz.f32 %f123, %f122, %f121;
mul.ftz.f32 %f124, %f120, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f125, %f124;
add.ftz.f32 %f126, %f125, 0f3F800000;
div.approx.ftz.f32 %f127, %f120, %f126;
mul.ftz.f32 %f168, %f123, %f127;
bra.uni $L__BB13_28;

$L__BB13_27:
mul.ftz.f32 %f168, %f168, %f24;
bra.uni $L__BB13_28;

$L__BB13_26:
mul.ftz.f32 %f143, %f24, 0f3F000000;
mul.ftz.f32 %f144, %f24, 0f3D372713;
fma.rn.ftz.f32 %f145, %f24, %f144, 0f3F800000;
mul.ftz.f32 %f146, %f24, 0f3F4C422A;
mul.ftz.f32 %f147, %f146, %f145;
tanh.approx.f32 %f148, %f147;
add.ftz.f32 %f149, %f148, 0f3F800000;
mul.ftz.f32 %f150, %f143, %f149;
mul.ftz.f32 %f168, %f168, %f150;
bra.uni $L__BB13_28;

$L__BB13_25:
mov.f32 %f128, 0f40E00000;
min.ftz.f32 %f129, %f24, %f128;
min.ftz.f32 %f130, %f168, %f128;
mov.f32 %f131, 0fC0E00000;
max.ftz.f32 %f132, %f130, %f131;
mul.ftz.f32 %f133, %f129, 0f3FD9DB23;
mul.ftz.f32 %f134, %f133, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f135, %f134;
add.ftz.f32 %f136, %f135, 0f3F800000;
div.approx.ftz.f32 %f137, %f129, %f136;
add.ftz.f32 %f138, %f132, 0f3F800000;
mul.ftz.f32 %f168, %f138, %f137;

$L__BB13_28:
ld.param.u64 %rd68, [_Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4];
ld.param.u32 %r390, [_Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10];
ld.param.u32 %r389, [_Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13];
mov.u32 %r388, %ctaid.y;
mov.u32 %r387, %tid.y;
mad.lo.s32 %r383, %r387, %r390, %r18;
mad.lo.s32 %r384, %r388, %r389, %r383;
cvta.to.global.u64 %rd63, %rd68;
mul.wide.u32 %rd64, %r384, 4;
add.s64 %rd65, %rd63, %rd64;
st.global.f32 [%rd65], %f168;

$L__BB13_29:
ret;

}
.entry _Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj(
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2,
.param .align 8 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3[48],
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5,
.param .align 4 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6[12],
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15
)
.maxntid 256, 1, 1
.minnctapersm 1
{
.local .align 8 .b8 __local_depot14[8];
.reg .b64 %SP;
.reg .b64 %SPL;
.reg .pred %p<17>;
.reg .b16 %rs<11>;
.reg .f32 %f<44>;
.reg .b32 %r<195>;
.reg .b64 %rd<54>;


mov.u64 %SPL, __local_depot14;
ld.param.u64 %rd17, [_Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0];
ld.param.u64 %rd18, [_Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1];
ld.param.u64 %rd19, [_Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2];
ld.param.u64 %rd20, [_Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4];
ld.param.u32 %r8, [_Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5];
ld.param.u32 %r12, [_Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7];
ld.param.u32 %r13, [_Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8];
ld.param.u32 %r14, [_Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9];
ld.param.u32 %r15, [_Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10];
ld.param.u32 %r16, [_Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11];
ld.param.u32 %r17, [_Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12];
ld.param.u32 %r18, [_Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13];
ld.param.u32 %r20, [_Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14];
ld.param.u32 %r19, [_Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15];
ld.param.u32 %r11, [_Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+8];
ld.param.u32 %r10, [_Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+4];
ld.param.u32 %r9, [_Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6];
add.u64 %rd1, %SPL, 0;
mov.u32 %r21, %ctaid.x;
shl.b32 %r1, %r21, 1;
mov.u32 %r22, %tid.y;
setp.ge.u32 %p1, %r22, %r20;
@%p1 bra $L__BB14_7;

mov.f32 %f42, 0f00000000;
st.local.v2.f32 [%rd1], {%f42, %f42};
mov.u32 %r2, %tid.x;
shr.u32 %r194, %r2, 1;
shr.u32 %r24, %r8, 6;
setp.ge.u32 %p2, %r194, %r24;
mov.f32 %f43, %f42;
@%p2 bra $L__BB14_5;

mov.u32 %r25, %ctaid.y;
mad.lo.s32 %r27, %r22, %r19, %r25;
cvta.to.global.u64 %rd22, %rd19;
mul.wide.u32 %rd23, %r27, 4;
add.s64 %rd24, %rd22, %rd23;
ld.global.nc.u32 %r28, [%rd24];
mul.lo.s32 %r29, %r28, %r16;
mul.lo.s32 %r30, %r22, %r14;
and.b32 %r32, %r2, 1;
shl.b32 %r33, %r2, 3;
cvt.u64.u32 %rd25, %r33;
and.b64 %rd26, %rd25, 8;
mad.lo.s32 %r34, %r1, %r13, %r194;
add.s32 %r35, %r34, %r29;
mul.wide.s32 %rd27, %r35, 18;
add.s64 %rd28, %rd26, %rd27;
cvta.to.global.u64 %rd29, %rd17;
add.s64 %rd30, %rd29, %rd28;
add.s64 %rd53, %rd30, 8;
add.s64 %rd52, %rd29, %rd27;
add.s32 %r36, %r194, %r29;
add.s32 %r37, %r1, 1;
mad.lo.s32 %r38, %r13, %r37, %r36;
mul.wide.s32 %rd31, %r38, 18;
add.s64 %rd32, %rd29, %rd31;
add.s64 %rd51, %rd32, 4;
add.s64 %rd33, %rd26, %rd31;
add.s64 %rd34, %rd29, %rd33;
add.s64 %rd50, %rd34, 8;
mul.wide.u32 %rd35, %r30, 36;
mul.hi.u32 %r39, %r25, %r9;
add.s32 %r40, %r39, %r25;
shr.u32 %r41, %r40, %r10;
mul.lo.s32 %r42, %r41, %r11;
sub.s32 %r43, %r25, %r42;
mul.lo.s32 %r44, %r43, %r17;
mul.wide.u32 %rd36, %r44, 36;
add.s64 %rd37, %rd35, %rd36;
mul.wide.s32 %rd38, %r32, 36;
add.s64 %rd39, %rd37, %rd38;
shl.b32 %r45, %r194, 1;
mul.wide.s32 %rd40, %r45, 36;
add.s64 %rd41, %rd39, %rd40;
cvta.to.global.u64 %rd42, %rd18;
add.s64 %rd43, %rd42, %rd41;
add.s64 %rd49, %rd43, 16;
mov.f32 %f42, 0f00000000;

$L__BB14_3:
ld.global.nc.u16 %rs1, [%rd52];

{ cvt.f32.f16 %f11, %rs1;}


ld.global.nc.u16 %rs3, [%rd53+-6];
cvt.s32.s16 %r112, %rs3;
ld.global.nc.u32 %r47, [%rd49+-12];
ld.global.nc.u32 %r51, [%rd49+-8];
and.b32 %r113, %r112, 30583;
mov.u32 %r114, 33620223;
prmt.b32 %r115, %r114, %r114, %r113;
shr.u32 %r116, %r112, 2;
and.b32 %r117, %r116, 30583;
prmt.b32 %r118, %r114, %r114, %r117;
mov.u32 %r119, 20800;
prmt.b32 %r48, %r115, %r118, %r119;
mov.u32 %r120, 29538;
prmt.b32 %r52, %r115, %r118, %r120;
mov.u32 %r82, 0;

dp4a.s32.s32 %r46, %r47, %r48, %r82;


dp4a.s32.s32 %r50, %r51, %r52, %r46;

ld.global.nc.u16 %rs4, [%rd53+-4];
cvt.s32.s16 %r121, %rs4;
ld.global.nc.u32 %r55, [%rd49+-4];
ld.global.nc.u32 %r59, [%rd49];
and.b32 %r122, %r121, 30583;
prmt.b32 %r123, %r114, %r114, %r122;
shr.u32 %r124, %r121, 2;
and.b32 %r125, %r124, 30583;
prmt.b32 %r126, %r114, %r114, %r125;
prmt.b32 %r56, %r123, %r126, %r119;
prmt.b32 %r60, %r123, %r126, %r120;

dp4a.s32.s32 %r54, %r55, %r56, %r50;


dp4a.s32.s32 %r58, %r59, %r60, %r54;

ld.global.nc.u16 %rs5, [%rd53+-2];
cvt.s32.s16 %r127, %rs5;
ld.global.nc.u32 %r63, [%rd49+4];
ld.global.nc.u32 %r67, [%rd49+8];
and.b32 %r128, %r127, 30583;
prmt.b32 %r129, %r114, %r114, %r128;
shr.u32 %r130, %r127, 2;
and.b32 %r131, %r130, 30583;
prmt.b32 %r132, %r114, %r114, %r131;
prmt.b32 %r64, %r129, %r132, %r119;
prmt.b32 %r68, %r129, %r132, %r120;

dp4a.s32.s32 %r62, %r63, %r64, %r58;


dp4a.s32.s32 %r66, %r67, %r68, %r62;

ld.global.nc.u16 %rs6, [%rd53];
cvt.s32.s16 %r133, %rs6;
ld.global.nc.u32 %r71, [%rd49+12];
ld.global.nc.u32 %r75, [%rd49+16];
and.b32 %r134, %r133, 30583;
prmt.b32 %r135, %r114, %r114, %r134;
shr.u32 %r136, %r133, 2;
and.b32 %r137, %r136, 30583;
prmt.b32 %r138, %r114, %r114, %r137;
prmt.b32 %r72, %r135, %r138, %r119;
prmt.b32 %r76, %r135, %r138, %r120;

dp4a.s32.s32 %r70, %r71, %r72, %r66;


dp4a.s32.s32 %r74, %r75, %r76, %r70;

ld.global.nc.u32 %r78, [%rd49+-16];

{.reg .f16 low,high;
mov.b32 {low,high},%r78;
cvt.f32.f16 %f12, low;}


mul.ftz.f32 %f15, %f11, %f12;
cvt.rn.f32.s32 %f16, %r74;
fma.rn.ftz.f32 %f43, %f15, %f16, %f43;
ld.global.nc.u16 %rs2, [%rd51+-4];

{ cvt.f32.f16 %f13, %rs2;}


ld.global.nc.u16 %rs7, [%rd50+-6];
cvt.s32.s16 %r139, %rs7;
and.b32 %r140, %r139, 30583;
prmt.b32 %r141, %r114, %r114, %r140;
shr.u32 %r142, %r139, 2;
and.b32 %r143, %r142, 30583;
prmt.b32 %r144, %r114, %r114, %r143;
prmt.b32 %r81, %r141, %r144, %r119;
prmt.b32 %r85, %r141, %r144, %r120;

dp4a.s32.s32 %r79, %r47, %r81, %r82;


dp4a.s32.s32 %r83, %r51, %r85, %r79;

ld.global.nc.u16 %rs8, [%rd50+-4];
cvt.s32.s16 %r145, %rs8;
and.b32 %r146, %r145, 30583;
prmt.b32 %r147, %r114, %r114, %r146;
shr.u32 %r148, %r145, 2;
and.b32 %r149, %r148, 30583;
prmt.b32 %r150, %r114, %r114, %r149;
prmt.b32 %r89, %r147, %r150, %r119;
prmt.b32 %r93, %r147, %r150, %r120;

dp4a.s32.s32 %r87, %r55, %r89, %r83;


dp4a.s32.s32 %r91, %r59, %r93, %r87;

ld.global.nc.u16 %rs9, [%rd50+-2];
cvt.s32.s16 %r151, %rs9;
and.b32 %r152, %r151, 30583;
prmt.b32 %r153, %r114, %r114, %r152;
shr.u32 %r154, %r151, 2;
and.b32 %r155, %r154, 30583;
prmt.b32 %r156, %r114, %r114, %r155;
prmt.b32 %r97, %r153, %r156, %r119;
prmt.b32 %r101, %r153, %r156, %r120;

dp4a.s32.s32 %r95, %r63, %r97, %r91;


dp4a.s32.s32 %r99, %r67, %r101, %r95;

ld.global.nc.u16 %rs10, [%rd50];
cvt.s32.s16 %r157, %rs10;
and.b32 %r158, %r157, 30583;
prmt.b32 %r159, %r114, %r114, %r158;
shr.u32 %r160, %r157, 2;
and.b32 %r161, %r160, 30583;
prmt.b32 %r162, %r114, %r114, %r161;
prmt.b32 %r105, %r159, %r162, %r119;
prmt.b32 %r109, %r159, %r162, %r120;

dp4a.s32.s32 %r103, %r71, %r105, %r99;


dp4a.s32.s32 %r107, %r75, %r109, %r103;


{.reg .f16 low,high;
mov.b32 {low,high},%r78;
cvt.f32.f16 %f14, low;}


mul.ftz.f32 %f17, %f13, %f14;
cvt.rn.f32.s32 %f18, %r107;
fma.rn.ftz.f32 %f42, %f17, %f18, %f42;
add.s64 %rd53, %rd53, 288;
add.s64 %rd52, %rd52, 288;
add.s64 %rd51, %rd51, 288;
add.s64 %rd50, %rd50, 288;
add.s64 %rd49, %rd49, 1152;
add.s32 %r194, %r194, 16;
setp.lt.u32 %p3, %r194, %r24;
@%p3 bra $L__BB14_3;

st.local.v2.f32 [%rd1], {%f43, %f42};

$L__BB14_5:
mov.b32 %r163, %f43;
mov.u32 %r164, 31;
mov.u32 %r165, 16;
mov.u32 %r166, -1;
shfl.sync.bfly.b32 %r167|%p4, %r163, %r165, %r164, %r166;
mov.b32 %f19, %r167;
add.ftz.f32 %f20, %f43, %f19;
mov.b32 %r168, %f20;
mov.u32 %r169, 8;
shfl.sync.bfly.b32 %r170|%p5, %r168, %r169, %r164, %r166;
mov.b32 %f21, %r170;
add.ftz.f32 %f22, %f20, %f21;
mov.b32 %r171, %f22;
mov.u32 %r172, 4;
shfl.sync.bfly.b32 %r173|%p6, %r171, %r172, %r164, %r166;
mov.b32 %f23, %r173;
add.ftz.f32 %f24, %f22, %f23;
mov.b32 %r174, %f24;
mov.u32 %r175, 2;
shfl.sync.bfly.b32 %r176|%p7, %r174, %r175, %r164, %r166;
mov.b32 %f25, %r176;
add.ftz.f32 %f26, %f24, %f25;
mov.b32 %r177, %f26;
mov.u32 %r178, 1;
shfl.sync.bfly.b32 %r179|%p8, %r177, %r178, %r164, %r166;
mov.b32 %f27, %r179;
add.ftz.f32 %f28, %f26, %f27;
st.local.f32 [%rd1], %f28;
mov.b32 %r180, %f42;
shfl.sync.bfly.b32 %r181|%p9, %r180, %r165, %r164, %r166;
mov.b32 %f29, %r181;
add.ftz.f32 %f30, %f42, %f29;
mov.b32 %r182, %f30;
shfl.sync.bfly.b32 %r183|%p10, %r182, %r169, %r164, %r166;
mov.b32 %f31, %r183;
add.ftz.f32 %f32, %f30, %f31;
mov.b32 %r184, %f32;
shfl.sync.bfly.b32 %r185|%p11, %r184, %r172, %r164, %r166;
mov.b32 %f33, %r185;
add.ftz.f32 %f34, %f32, %f33;
mov.b32 %r186, %f34;
shfl.sync.bfly.b32 %r187|%p12, %r186, %r175, %r164, %r166;
mov.b32 %f35, %r187;
add.ftz.f32 %f36, %f34, %f35;
mov.b32 %r188, %f36;
shfl.sync.bfly.b32 %r189|%p13, %r188, %r178, %r164, %r166;
mov.b32 %f37, %r189;
add.ftz.f32 %f38, %f36, %f37;
st.local.f32 [%rd1+4], %f38;
add.s32 %r7, %r1, %r2;
setp.ge.u32 %p14, %r7, %r12;
setp.gt.u32 %p15, %r2, 1;
or.pred %p16, %p15, %p14;
@%p16 bra $L__BB14_7;

mul.wide.u32 %rd44, %r2, 4;
add.s64 %rd45, %rd1, %rd44;
ld.local.f32 %f39, [%rd45];
mov.u32 %r190, %ctaid.y;
mad.lo.s32 %r192, %r22, %r15, %r7;
mad.lo.s32 %r193, %r190, %r18, %r192;
cvta.to.global.u64 %rd46, %rd20;
mul.wide.u32 %rd47, %r193, 4;
add.s64 %rd48, %rd46, %rd47;
st.global.f32 [%rd48], %f39;

$L__BB14_7:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<320>;
.reg .b16 %rs<25>;
.reg .f32 %f<1005>;
.reg .b32 %r<1339>;
.reg .b64 %rd<93>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate[1536];

ld.param.u64 %rd25, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd26, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd27, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u32 %r47, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r51, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r52, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r56, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r57, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r58, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r62, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r63, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r64, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r60, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r59, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r54, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r53, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r50, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r49, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r48, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
ld.param.u64 %rd30, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+16];
ld.param.u64 %rd29, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+8];
ld.param.u64 %rd28, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3];
mov.u32 %r1, %tid.y;
shl.b32 %r65, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r65, %r2;
mov.u32 %r66, %ctaid.x;
shl.b32 %r4, %r66, 2;
mov.u32 %r5, %ctaid.y;
setp.eq.s64 %p1, %rd27, 0;
@%p1 bra $L__BB15_2;

cvta.to.global.u64 %rd34, %rd27;
mul.wide.u32 %rd35, %r5, 4;
add.s64 %rd36, %rd34, %rd35;
ld.global.nc.u32 %r1335, [%rd36];
mul.hi.u32 %r67, %r5, %r48;
add.s32 %r68, %r67, %r5;
shr.u32 %r69, %r68, %r49;
mul.lo.s32 %r70, %r69, %r50;
sub.s32 %r1336, %r5, %r70;
bra.uni $L__BB15_3;

$L__BB15_2:
mul.hi.u32 %r71, %r5, %r53;
add.s32 %r72, %r71, %r5;
shr.u32 %r1335, %r72, %r54;
mov.u32 %r1336, %r5;

$L__BB15_3:
mov.u32 %r11, %ctaid.z;
selp.b32 %r12, %r5, %r1335, %p1;
setp.ne.s32 %p3, %r1, 0;
setp.gt.u32 %p4, %r2, 3;
or.pred %p5, %p4, %p3;
add.s32 %r73, %r4, %r2;
setp.ge.u32 %p6, %r73, %r52;
mov.f32 %f26, 0f00000000;
or.pred %p7, %p5, %p6;
mov.f32 %f4, %f26;
mov.f32 %f5, %f26;
@%p7 bra $L__BB15_8;

setp.eq.s64 %p8, %rd28, 0;
mov.f32 %f4, 0f00000000;
mov.f32 %f5, %f4;
@%p8 bra $L__BB15_6;

cvta.to.global.u64 %rd37, %rd28;
mul.lo.s32 %r74, %r11, %r64;
cvt.u64.u32 %rd38, %r74;
mul.lo.s32 %r75, %r12, %r58;
cvt.u64.u32 %rd39, %r75;
cvt.u64.u32 %rd40, %r2;
cvt.s64.s32 %rd41, %r4;
add.s64 %rd42, %rd41, %rd40;
add.s64 %rd43, %rd42, %rd38;
add.s64 %rd44, %rd43, %rd39;
shl.b64 %rd45, %rd44, 2;
add.s64 %rd46, %rd37, %rd45;
ld.global.f32 %f5, [%rd46];

$L__BB15_6:
setp.eq.s64 %p9, %rd30, 0;
setp.eq.s64 %p10, %rd29, 0;
or.pred %p11, %p9, %p10;
@%p11 bra $L__BB15_8;

cvta.to.global.u64 %rd47, %rd30;
mul.lo.s32 %r76, %r11, %r64;
cvt.u64.u32 %rd48, %r76;
mul.lo.s32 %r77, %r12, %r58;
cvt.u64.u32 %rd49, %r77;
cvt.u64.u32 %rd50, %r2;
cvt.s64.s32 %rd51, %r4;
add.s64 %rd52, %rd51, %rd50;
add.s64 %rd53, %rd52, %rd48;
add.s64 %rd54, %rd53, %rd49;
shl.b64 %rd55, %rd54, 2;
add.s64 %rd56, %rd47, %rd55;
ld.global.f32 %f4, [%rd56];

$L__BB15_8:
shr.u32 %r78, %r3, 31;
add.s32 %r79, %r3, %r78;
shr.s32 %r1338, %r79, 1;
shr.u32 %r80, %r47, 6;
setp.ge.s32 %p12, %r1338, %r80;
mov.f32 %f27, %f26;
mov.f32 %f28, %f26;
mov.f32 %f29, %f26;
mov.f32 %f30, %f26;
mov.f32 %f31, %f26;
mov.f32 %f32, %f26;
mov.f32 %f33, %f26;
@%p12 bra $L__BB15_19;

mul.hi.u32 %r81, %r11, %r59;
add.s32 %r82, %r81, %r11;
shr.u32 %r83, %r82, %r60;
mul.lo.s32 %r84, %r83, %r62;
mul.lo.s32 %r85, %r1335, %r56;
mul.lo.s32 %r86, %r1336, %r57;
cvt.u64.u32 %rd57, %r86;
and.b32 %r89, %r79, -2;
sub.s32 %r90, %r3, %r89;
shl.b32 %r91, %r90, 2;
cvt.s64.s32 %rd58, %r90;
mul.lo.s32 %r92, %r11, %r63;
cvt.u64.u32 %rd59, %r92;
add.s64 %rd60, %rd59, %rd58;
add.s64 %rd4, %rd60, %rd57;
shl.b32 %r1337, %r1338, 1;
mad.lo.s32 %r93, %r4, %r51, %r1338;
add.s32 %r94, %r93, %r84;
add.s32 %r95, %r94, %r85;
mul.wide.s32 %rd5, %r95, 18;
mul.wide.s32 %rd61, %r91, 2;
add.s64 %rd62, %rd5, %rd61;
add.s64 %rd6, %rd62, 8;
add.s32 %r96, %r1338, %r84;
add.s32 %r17, %r96, %r85;
add.s32 %r97, %r4, 1;
mad.lo.s32 %r98, %r51, %r97, %r17;
mul.wide.s32 %rd63, %r98, 18;
add.s64 %rd64, %rd63, %rd61;
add.s64 %rd7, %rd64, 4;
add.s32 %r99, %r4, 3;
mad.lo.s32 %r100, %r51, %r99, %r17;
mul.wide.s32 %rd8, %r100, 18;
add.s64 %rd65, %rd8, %rd61;
add.s64 %rd9, %rd65, 8;
add.s32 %r101, %r4, 2;
mad.lo.s32 %r102, %r51, %r101, %r17;
mul.wide.s32 %rd66, %r102, 18;
add.s64 %rd67, %rd66, %rd61;
add.s64 %rd10, %rd67, 4;
mul.wide.s32 %rd11, %r51, 18;
cvta.to.global.u64 %rd91, %rd29;
cvta.to.global.u64 %rd92, %rd25;
mov.f32 %f26, 0f00000000;
cvta.to.global.u64 %rd68, %rd26;
mov.f32 %f27, %f26;
mov.f32 %f28, %f26;
mov.f32 %f29, %f26;
mov.f32 %f30, %f26;
mov.f32 %f31, %f26;
mov.f32 %f32, %f26;
mov.f32 %f33, %f26;

$L__BB15_10:
cvt.u64.u32 %rd69, %r1337;
add.s64 %rd70, %rd4, %rd69;
mul.lo.s64 %rd71, %rd70, 36;
add.s64 %rd16, %rd92, %rd5;
ld.global.nc.u16 %rs1, [%rd16];

{ cvt.f32.f16 %f101, %rs1;}


add.s64 %rd72, %rd92, %rd6;
ld.global.nc.u16 %rs2, [%rd72+-6];
cvt.s32.s16 %r136, %rs2;
add.s64 %rd73, %rd68, %rd71;
ld.global.nc.u32 %r20, [%rd73+4];
ld.global.nc.u32 %r21, [%rd73+8];
and.b32 %r137, %r136, 30583;
mov.u32 %r138, 33620223;
prmt.b32 %r139, %r138, %r138, %r137;
shr.u32 %r140, %r136, 2;
and.b32 %r141, %r140, 30583;
prmt.b32 %r142, %r138, %r138, %r141;
mov.u32 %r143, 20800;
prmt.b32 %r105, %r139, %r142, %r143;
mov.u32 %r144, 29538;
prmt.b32 %r109, %r139, %r142, %r144;
mov.u32 %r106, 0;

dp4a.s32.s32 %r103, %r20, %r105, %r106;


dp4a.s32.s32 %r107, %r21, %r109, %r103;

ld.global.nc.u16 %rs3, [%rd72+-4];
cvt.s32.s16 %r145, %rs3;
ld.global.nc.u32 %r22, [%rd73+12];
ld.global.nc.u32 %r23, [%rd73+16];
and.b32 %r146, %r145, 30583;
prmt.b32 %r147, %r138, %r138, %r146;
shr.u32 %r148, %r145, 2;
and.b32 %r149, %r148, 30583;
prmt.b32 %r150, %r138, %r138, %r149;
prmt.b32 %r113, %r147, %r150, %r143;
prmt.b32 %r117, %r147, %r150, %r144;

dp4a.s32.s32 %r111, %r22, %r113, %r107;


dp4a.s32.s32 %r115, %r23, %r117, %r111;

ld.global.nc.u16 %rs4, [%rd72+-2];
cvt.s32.s16 %r151, %rs4;
ld.global.nc.u32 %r24, [%rd73+20];
ld.global.nc.u32 %r25, [%rd73+24];
and.b32 %r152, %r151, 30583;
prmt.b32 %r153, %r138, %r138, %r152;
shr.u32 %r154, %r151, 2;
and.b32 %r155, %r154, 30583;
prmt.b32 %r156, %r138, %r138, %r155;
prmt.b32 %r121, %r153, %r156, %r143;
prmt.b32 %r125, %r153, %r156, %r144;

dp4a.s32.s32 %r119, %r24, %r121, %r115;


dp4a.s32.s32 %r123, %r25, %r125, %r119;

ld.global.nc.u16 %rs5, [%rd72];
cvt.s32.s16 %r157, %rs5;
ld.global.nc.u32 %r26, [%rd73+28];
ld.global.nc.u32 %r27, [%rd73+32];
and.b32 %r158, %r157, 30583;
prmt.b32 %r159, %r138, %r138, %r158;
shr.u32 %r160, %r157, 2;
and.b32 %r161, %r160, 30583;
prmt.b32 %r162, %r138, %r138, %r161;
prmt.b32 %r129, %r159, %r162, %r143;
prmt.b32 %r133, %r159, %r162, %r144;

dp4a.s32.s32 %r127, %r26, %r129, %r123;


dp4a.s32.s32 %r131, %r27, %r133, %r127;

ld.global.nc.u32 %r28, [%rd73];

{.reg .f16 low,high;
mov.b32 {low,high},%r28;
cvt.f32.f16 %f102, low;}


mul.ftz.f32 %f103, %f101, %f102;
cvt.rn.f32.s32 %f104, %r131;
fma.rn.ftz.f32 %f29, %f103, %f104, %f29;
setp.eq.s64 %p13, %rd29, 0;
@%p13 bra $L__BB15_12;

add.s64 %rd74, %rd91, %rd5;
ld.global.u16 %rs6, [%rd74];

{ cvt.f32.f16 %f105, %rs6;}


add.s64 %rd75, %rd91, %rd6;
ld.global.s16 %r196, [%rd75+-6];
and.b32 %r197, %r196, 30583;
prmt.b32 %r199, %r138, %r138, %r197;
shr.u32 %r200, %r196, 2;
and.b32 %r201, %r200, 30583;
prmt.b32 %r202, %r138, %r138, %r201;
prmt.b32 %r165, %r199, %r202, %r143;
prmt.b32 %r169, %r199, %r202, %r144;

dp4a.s32.s32 %r163, %r20, %r165, %r106;


dp4a.s32.s32 %r167, %r21, %r169, %r163;

ld.global.s16 %r205, [%rd75+-4];
and.b32 %r206, %r205, 30583;
prmt.b32 %r207, %r138, %r138, %r206;
shr.u32 %r208, %r205, 2;
and.b32 %r209, %r208, 30583;
prmt.b32 %r210, %r138, %r138, %r209;
prmt.b32 %r173, %r207, %r210, %r143;
prmt.b32 %r177, %r207, %r210, %r144;

dp4a.s32.s32 %r171, %r22, %r173, %r167;


dp4a.s32.s32 %r175, %r23, %r177, %r171;

ld.global.s16 %r211, [%rd75+-2];
and.b32 %r212, %r211, 30583;
prmt.b32 %r213, %r138, %r138, %r212;
shr.u32 %r214, %r211, 2;
and.b32 %r215, %r214, 30583;
prmt.b32 %r216, %r138, %r138, %r215;
prmt.b32 %r181, %r213, %r216, %r143;
prmt.b32 %r185, %r213, %r216, %r144;

dp4a.s32.s32 %r179, %r24, %r181, %r175;


dp4a.s32.s32 %r183, %r25, %r185, %r179;

ld.global.s16 %r217, [%rd75];
and.b32 %r218, %r217, 30583;
prmt.b32 %r219, %r138, %r138, %r218;
shr.u32 %r220, %r217, 2;
and.b32 %r221, %r220, 30583;
prmt.b32 %r222, %r138, %r138, %r221;
prmt.b32 %r189, %r219, %r222, %r143;
prmt.b32 %r193, %r219, %r222, %r144;

dp4a.s32.s32 %r187, %r26, %r189, %r183;


dp4a.s32.s32 %r191, %r27, %r193, %r187;


{.reg .f16 low,high;
mov.b32 {low,high},%r28;
cvt.f32.f16 %f106, low;}


mul.ftz.f32 %f107, %f105, %f106;
cvt.rn.f32.s32 %f108, %r191;
fma.rn.ftz.f32 %f33, %f107, %f108, %f33;

$L__BB15_12:
add.s64 %rd89, %rd92, %rd6;
add.s64 %rd88, %rd89, -6;
add.s64 %rd18, %rd16, %rd11;
ld.global.nc.u16 %rs7, [%rd18];

{ cvt.f32.f16 %f109, %rs7;}


add.s64 %rd19, %rd88, %rd11;
ld.global.nc.u16 %rs8, [%rd19];
cvt.s32.s16 %r256, %rs8;
and.b32 %r257, %r256, 30583;
prmt.b32 %r259, %r138, %r138, %r257;
shr.u32 %r260, %r256, 2;
and.b32 %r261, %r260, 30583;
prmt.b32 %r262, %r138, %r138, %r261;
prmt.b32 %r225, %r259, %r262, %r143;
prmt.b32 %r229, %r259, %r262, %r144;

dp4a.s32.s32 %r223, %r20, %r225, %r106;


dp4a.s32.s32 %r227, %r21, %r229, %r223;

ld.global.nc.u16 %rs9, [%rd19+2];
cvt.s32.s16 %r265, %rs9;
and.b32 %r266, %r265, 30583;
prmt.b32 %r267, %r138, %r138, %r266;
shr.u32 %r268, %r265, 2;
and.b32 %r269, %r268, 30583;
prmt.b32 %r270, %r138, %r138, %r269;
prmt.b32 %r233, %r267, %r270, %r143;
prmt.b32 %r237, %r267, %r270, %r144;

dp4a.s32.s32 %r231, %r22, %r233, %r227;


dp4a.s32.s32 %r235, %r23, %r237, %r231;

ld.global.nc.u16 %rs10, [%rd19+4];
cvt.s32.s16 %r271, %rs10;
and.b32 %r272, %r271, 30583;
prmt.b32 %r273, %r138, %r138, %r272;
shr.u32 %r274, %r271, 2;
and.b32 %r275, %r274, 30583;
prmt.b32 %r276, %r138, %r138, %r275;
prmt.b32 %r241, %r273, %r276, %r143;
prmt.b32 %r245, %r273, %r276, %r144;

dp4a.s32.s32 %r239, %r24, %r241, %r235;


dp4a.s32.s32 %r243, %r25, %r245, %r239;

ld.global.nc.u16 %rs11, [%rd19+6];
cvt.s32.s16 %r277, %rs11;
and.b32 %r278, %r277, 30583;
prmt.b32 %r279, %r138, %r138, %r278;
shr.u32 %r280, %r277, 2;
and.b32 %r281, %r280, 30583;
prmt.b32 %r282, %r138, %r138, %r281;
prmt.b32 %r249, %r279, %r282, %r143;
prmt.b32 %r253, %r279, %r282, %r144;

dp4a.s32.s32 %r247, %r26, %r249, %r243;


dp4a.s32.s32 %r251, %r27, %r253, %r247;


{.reg .f16 low,high;
mov.b32 {low,high},%r28;
cvt.f32.f16 %f110, low;}


mul.ftz.f32 %f111, %f109, %f110;
cvt.rn.f32.s32 %f112, %r251;
fma.rn.ftz.f32 %f28, %f111, %f112, %f28;
@%p13 bra $L__BB15_14;

or.b32 %r318, %r4, 1;
mad.lo.s32 %r319, %r51, %r318, %r17;
mul.wide.s32 %rd76, %r319, 18;
add.s64 %rd77, %rd91, %rd76;
ld.global.u16 %rs12, [%rd77];

{ cvt.f32.f16 %f113, %rs12;}


add.s64 %rd78, %rd91, %rd7;
ld.global.s16 %r320, [%rd78+-2];
and.b32 %r321, %r320, 30583;
mov.u32 %r322, 33620223;
prmt.b32 %r323, %r322, %r322, %r321;
shr.u32 %r324, %r320, 2;
and.b32 %r325, %r324, 30583;
prmt.b32 %r326, %r322, %r322, %r325;
mov.u32 %r327, 20800;
prmt.b32 %r285, %r323, %r326, %r327;
mov.u32 %r328, 29538;
prmt.b32 %r289, %r323, %r326, %r328;
mov.u32 %r286, 0;

dp4a.s32.s32 %r283, %r20, %r285, %r286;


dp4a.s32.s32 %r287, %r21, %r289, %r283;

ld.global.s16 %r329, [%rd78];
and.b32 %r330, %r329, 30583;
prmt.b32 %r331, %r322, %r322, %r330;
shr.u32 %r332, %r329, 2;
and.b32 %r333, %r332, 30583;
prmt.b32 %r334, %r322, %r322, %r333;
prmt.b32 %r293, %r331, %r334, %r327;
prmt.b32 %r297, %r331, %r334, %r328;

dp4a.s32.s32 %r291, %r22, %r293, %r287;


dp4a.s32.s32 %r295, %r23, %r297, %r291;

ld.global.s16 %r335, [%rd78+2];
and.b32 %r336, %r335, 30583;
prmt.b32 %r337, %r322, %r322, %r336;
shr.u32 %r338, %r335, 2;
and.b32 %r339, %r338, 30583;
prmt.b32 %r340, %r322, %r322, %r339;
prmt.b32 %r301, %r337, %r340, %r327;
prmt.b32 %r305, %r337, %r340, %r328;

dp4a.s32.s32 %r299, %r24, %r301, %r295;


dp4a.s32.s32 %r303, %r25, %r305, %r299;

ld.global.s16 %r341, [%rd78+4];
and.b32 %r342, %r341, 30583;
prmt.b32 %r343, %r322, %r322, %r342;
shr.u32 %r344, %r341, 2;
and.b32 %r345, %r344, 30583;
prmt.b32 %r346, %r322, %r322, %r345;
prmt.b32 %r309, %r343, %r346, %r327;
prmt.b32 %r313, %r343, %r346, %r328;

dp4a.s32.s32 %r307, %r26, %r309, %r303;


dp4a.s32.s32 %r311, %r27, %r313, %r307;


{.reg .f16 low,high;
mov.b32 {low,high},%r28;
cvt.f32.f16 %f114, low;}


mul.ftz.f32 %f115, %f113, %f114;
cvt.rn.f32.s32 %f116, %r311;
fma.rn.ftz.f32 %f32, %f115, %f116, %f32;

$L__BB15_14:
add.s64 %rd20, %rd18, %rd11;
ld.global.nc.u16 %rs13, [%rd20];

{ cvt.f32.f16 %f117, %rs13;}


add.s64 %rd21, %rd19, %rd11;
ld.global.nc.u16 %rs14, [%rd21];
cvt.s32.s16 %r380, %rs14;
and.b32 %r381, %r380, 30583;
mov.u32 %r382, 33620223;
prmt.b32 %r383, %r382, %r382, %r381;
shr.u32 %r384, %r380, 2;
and.b32 %r385, %r384, 30583;
prmt.b32 %r386, %r382, %r382, %r385;
mov.u32 %r387, 20800;
prmt.b32 %r349, %r383, %r386, %r387;
mov.u32 %r388, 29538;
prmt.b32 %r353, %r383, %r386, %r388;
mov.u32 %r350, 0;

dp4a.s32.s32 %r347, %r20, %r349, %r350;


dp4a.s32.s32 %r351, %r21, %r353, %r347;

ld.global.nc.u16 %rs15, [%rd21+2];
cvt.s32.s16 %r389, %rs15;
and.b32 %r390, %r389, 30583;
prmt.b32 %r391, %r382, %r382, %r390;
shr.u32 %r392, %r389, 2;
and.b32 %r393, %r392, 30583;
prmt.b32 %r394, %r382, %r382, %r393;
prmt.b32 %r357, %r391, %r394, %r387;
prmt.b32 %r361, %r391, %r394, %r388;

dp4a.s32.s32 %r355, %r22, %r357, %r351;


dp4a.s32.s32 %r359, %r23, %r361, %r355;

ld.global.nc.u16 %rs16, [%rd21+4];
cvt.s32.s16 %r395, %rs16;
and.b32 %r396, %r395, 30583;
prmt.b32 %r397, %r382, %r382, %r396;
shr.u32 %r398, %r395, 2;
and.b32 %r399, %r398, 30583;
prmt.b32 %r400, %r382, %r382, %r399;
prmt.b32 %r365, %r397, %r400, %r387;
prmt.b32 %r369, %r397, %r400, %r388;

dp4a.s32.s32 %r363, %r24, %r365, %r359;


dp4a.s32.s32 %r367, %r25, %r369, %r363;

ld.global.nc.u16 %rs17, [%rd21+6];
cvt.s32.s16 %r401, %rs17;
and.b32 %r402, %r401, 30583;
prmt.b32 %r403, %r382, %r382, %r402;
shr.u32 %r404, %r401, 2;
and.b32 %r405, %r404, 30583;
prmt.b32 %r406, %r382, %r382, %r405;
prmt.b32 %r373, %r403, %r406, %r387;
prmt.b32 %r377, %r403, %r406, %r388;

dp4a.s32.s32 %r371, %r26, %r373, %r367;


dp4a.s32.s32 %r375, %r27, %r377, %r371;


{.reg .f16 low,high;
mov.b32 {low,high},%r28;
cvt.f32.f16 %f118, low;}


mul.ftz.f32 %f119, %f117, %f118;
cvt.rn.f32.s32 %f120, %r375;
fma.rn.ftz.f32 %f27, %f119, %f120, %f27;
@%p13 bra $L__BB15_16;

or.b32 %r442, %r4, 2;
mad.lo.s32 %r443, %r51, %r442, %r17;
mul.wide.s32 %rd79, %r443, 18;
add.s64 %rd80, %rd91, %rd79;
ld.global.u16 %rs18, [%rd80];

{ cvt.f32.f16 %f121, %rs18;}


add.s64 %rd81, %rd91, %rd10;
ld.global.s16 %r444, [%rd81+-2];
and.b32 %r445, %r444, 30583;
prmt.b32 %r447, %r382, %r382, %r445;
shr.u32 %r448, %r444, 2;
and.b32 %r449, %r448, 30583;
prmt.b32 %r450, %r382, %r382, %r449;
prmt.b32 %r409, %r447, %r450, %r387;
prmt.b32 %r413, %r447, %r450, %r388;

dp4a.s32.s32 %r407, %r20, %r409, %r350;


dp4a.s32.s32 %r411, %r21, %r413, %r407;

ld.global.s16 %r453, [%rd81];
and.b32 %r454, %r453, 30583;
prmt.b32 %r455, %r382, %r382, %r454;
shr.u32 %r456, %r453, 2;
and.b32 %r457, %r456, 30583;
prmt.b32 %r458, %r382, %r382, %r457;
prmt.b32 %r417, %r455, %r458, %r387;
prmt.b32 %r421, %r455, %r458, %r388;

dp4a.s32.s32 %r415, %r22, %r417, %r411;


dp4a.s32.s32 %r419, %r23, %r421, %r415;

ld.global.s16 %r459, [%rd81+2];
and.b32 %r460, %r459, 30583;
prmt.b32 %r461, %r382, %r382, %r460;
shr.u32 %r462, %r459, 2;
and.b32 %r463, %r462, 30583;
prmt.b32 %r464, %r382, %r382, %r463;
prmt.b32 %r425, %r461, %r464, %r387;
prmt.b32 %r429, %r461, %r464, %r388;

dp4a.s32.s32 %r423, %r24, %r425, %r419;


dp4a.s32.s32 %r427, %r25, %r429, %r423;

ld.global.s16 %r465, [%rd81+4];
and.b32 %r466, %r465, 30583;
prmt.b32 %r467, %r382, %r382, %r466;
shr.u32 %r468, %r465, 2;
and.b32 %r469, %r468, 30583;
prmt.b32 %r470, %r382, %r382, %r469;
prmt.b32 %r433, %r467, %r470, %r387;
prmt.b32 %r437, %r467, %r470, %r388;

dp4a.s32.s32 %r431, %r26, %r433, %r427;


dp4a.s32.s32 %r435, %r27, %r437, %r431;


{.reg .f16 low,high;
mov.b32 {low,high},%r28;
cvt.f32.f16 %f122, low;}


mul.ftz.f32 %f123, %f121, %f122;
cvt.rn.f32.s32 %f124, %r435;
fma.rn.ftz.f32 %f31, %f123, %f124, %f31;

$L__BB15_16:
add.s64 %rd82, %rd20, %rd11;
ld.global.nc.u16 %rs19, [%rd82];

{ cvt.f32.f16 %f125, %rs19;}


add.s64 %rd83, %rd21, %rd11;
ld.global.nc.u16 %rs20, [%rd83];
cvt.s32.s16 %r504, %rs20;
and.b32 %r505, %r504, 30583;
prmt.b32 %r507, %r382, %r382, %r505;
shr.u32 %r508, %r504, 2;
and.b32 %r509, %r508, 30583;
prmt.b32 %r510, %r382, %r382, %r509;
prmt.b32 %r473, %r507, %r510, %r387;
prmt.b32 %r477, %r507, %r510, %r388;

dp4a.s32.s32 %r471, %r20, %r473, %r350;


dp4a.s32.s32 %r475, %r21, %r477, %r471;

ld.global.nc.u16 %rs21, [%rd83+2];
cvt.s32.s16 %r513, %rs21;
and.b32 %r514, %r513, 30583;
prmt.b32 %r515, %r382, %r382, %r514;
shr.u32 %r516, %r513, 2;
and.b32 %r517, %r516, 30583;
prmt.b32 %r518, %r382, %r382, %r517;
prmt.b32 %r481, %r515, %r518, %r387;
prmt.b32 %r485, %r515, %r518, %r388;

dp4a.s32.s32 %r479, %r22, %r481, %r475;


dp4a.s32.s32 %r483, %r23, %r485, %r479;

ld.global.nc.u16 %rs22, [%rd83+4];
cvt.s32.s16 %r519, %rs22;
and.b32 %r520, %r519, 30583;
prmt.b32 %r521, %r382, %r382, %r520;
shr.u32 %r522, %r519, 2;
and.b32 %r523, %r522, 30583;
prmt.b32 %r524, %r382, %r382, %r523;
prmt.b32 %r489, %r521, %r524, %r387;
prmt.b32 %r493, %r521, %r524, %r388;

dp4a.s32.s32 %r487, %r24, %r489, %r483;


dp4a.s32.s32 %r491, %r25, %r493, %r487;

ld.global.nc.u16 %rs23, [%rd83+6];
cvt.s32.s16 %r525, %rs23;
and.b32 %r526, %r525, 30583;
prmt.b32 %r527, %r382, %r382, %r526;
shr.u32 %r528, %r525, 2;
and.b32 %r529, %r528, 30583;
prmt.b32 %r530, %r382, %r382, %r529;
prmt.b32 %r497, %r527, %r530, %r387;
prmt.b32 %r501, %r527, %r530, %r388;

dp4a.s32.s32 %r495, %r26, %r497, %r491;


dp4a.s32.s32 %r499, %r27, %r501, %r495;


{.reg .f16 low,high;
mov.b32 {low,high},%r28;
cvt.f32.f16 %f126, low;}


mul.ftz.f32 %f127, %f125, %f126;
cvt.rn.f32.s32 %f128, %r499;
fma.rn.ftz.f32 %f26, %f127, %f128, %f26;
@%p13 bra $L__BB15_18;

add.s64 %rd84, %rd91, %rd8;
ld.global.u16 %rs24, [%rd84];

{ cvt.f32.f16 %f129, %rs24;}


add.s64 %rd85, %rd91, %rd9;
ld.global.s16 %r564, [%rd85+-6];
and.b32 %r565, %r564, 30583;
mov.u32 %r566, 33620223;
prmt.b32 %r567, %r566, %r566, %r565;
shr.u32 %r568, %r564, 2;
and.b32 %r569, %r568, 30583;
prmt.b32 %r570, %r566, %r566, %r569;
mov.u32 %r571, 20800;
prmt.b32 %r533, %r567, %r570, %r571;
mov.u32 %r572, 29538;
prmt.b32 %r537, %r567, %r570, %r572;
mov.u32 %r534, 0;

dp4a.s32.s32 %r531, %r20, %r533, %r534;


dp4a.s32.s32 %r535, %r21, %r537, %r531;

ld.global.s16 %r573, [%rd85+-4];
and.b32 %r574, %r573, 30583;
prmt.b32 %r575, %r566, %r566, %r574;
shr.u32 %r576, %r573, 2;
and.b32 %r577, %r576, 30583;
prmt.b32 %r578, %r566, %r566, %r577;
prmt.b32 %r541, %r575, %r578, %r571;
prmt.b32 %r545, %r575, %r578, %r572;

dp4a.s32.s32 %r539, %r22, %r541, %r535;


dp4a.s32.s32 %r543, %r23, %r545, %r539;

ld.global.s16 %r579, [%rd85+-2];
and.b32 %r580, %r579, 30583;
prmt.b32 %r581, %r566, %r566, %r580;
shr.u32 %r582, %r579, 2;
and.b32 %r583, %r582, 30583;
prmt.b32 %r584, %r566, %r566, %r583;
prmt.b32 %r549, %r581, %r584, %r571;
prmt.b32 %r553, %r581, %r584, %r572;

dp4a.s32.s32 %r547, %r24, %r549, %r543;


dp4a.s32.s32 %r551, %r25, %r553, %r547;

ld.global.s16 %r585, [%rd85];
and.b32 %r586, %r585, 30583;
prmt.b32 %r587, %r566, %r566, %r586;
shr.u32 %r588, %r585, 2;
and.b32 %r589, %r588, 30583;
prmt.b32 %r590, %r566, %r566, %r589;
prmt.b32 %r557, %r587, %r590, %r571;
prmt.b32 %r561, %r587, %r590, %r572;

dp4a.s32.s32 %r555, %r26, %r557, %r551;


dp4a.s32.s32 %r559, %r27, %r561, %r555;


{.reg .f16 low,high;
mov.b32 {low,high},%r28;
cvt.f32.f16 %f130, low;}


mul.ftz.f32 %f131, %f129, %f130;
cvt.rn.f32.s32 %f132, %r559;
fma.rn.ftz.f32 %f30, %f131, %f132, %f30;

$L__BB15_18:
add.s32 %r1337, %r1337, 128;
add.s64 %rd92, %rd92, 1152;
add.s64 %rd91, %rd91, 1152;
add.s32 %r1338, %r1338, 64;
setp.lt.s32 %p17, %r1338, %r80;
@%p17 bra $L__BB15_10;

$L__BB15_19:
mov.u32 %r1324, %tid.y;
setp.eq.s32 %p18, %r1324, 0;
@%p18 bra $L__BB15_23;

mov.u32 %r1325, %tid.y;
setp.ne.s64 %p19, %rd29, 0;
add.s32 %r31, %r1325, -1;
shl.b32 %r595, %r1325, 9;
mov.u32 %r596, _ZZ13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r597, %r596, %r595;
shl.b32 %r598, %r2, 2;
add.s32 %r599, %r597, %r598;
st.shared.f32 [%r599+-512], %f29;
@%p19 bra $L__BB15_22;
bra.uni $L__BB15_21;

$L__BB15_22:
shl.b32 %r601, %r31, 9;
mov.u32 %r602, _ZZ13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r603, %r602, %r601;
add.s32 %r605, %r603, %r598;
st.shared.f32 [%r605], %f33;
add.s32 %r1321, %r599, -512;
st.shared.f32 [%r1321+128], %f28;
st.shared.f32 [%r605+128], %f32;
add.s32 %r1322, %r599, -512;
st.shared.f32 [%r1322+256], %f27;
st.shared.f32 [%r605+256], %f31;
add.s32 %r1323, %r599, -512;
st.shared.f32 [%r1323+384], %f26;
st.shared.f32 [%r605+384], %f30;
bra.uni $L__BB15_23;

$L__BB15_21:
add.s32 %r1318, %r599, -512;
st.shared.f32 [%r1318+128], %f28;
add.s32 %r1319, %r599, -512;
st.shared.f32 [%r1319+256], %f27;
add.s32 %r1320, %r599, -512;
st.shared.f32 [%r1320+384], %f26;

$L__BB15_23:
mov.u32 %r1326, %tid.y;
setp.ne.s32 %p319, %r1326, 0;
bar.sync 0;
@%p319 bra $L__BB15_78;
bra.uni $L__BB15_24;

$L__BB15_78:
ret;

$L__BB15_24:
ld.param.f32 %f977, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
ld.param.u64 %rd90, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r1329, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r1328, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
mov.u32 %r1327, %ctaid.z;
setp.ne.s64 %p21, %rd29, 0;
mov.u32 %r608, %ctaid.y;
mad.lo.s32 %r611, %r608, %r1329, %r4;
mad.lo.s32 %r612, %r1327, %r1328, %r611;
neg.ftz.f32 %f35, %f977;
shl.b32 %r614, %r2, 2;
mov.u32 %r615, _ZZ13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r33, %r615, %r614;
cvta.to.global.u64 %rd86, %rd90;
mul.wide.u32 %rd87, %r612, 4;
add.s64 %rd24, %rd86, %rd87;
@%p21 bra $L__BB15_33;
bra.uni $L__BB15_25;

$L__BB15_33:
ld.param.u32 %r1330, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
mov.u32 %r702, _ZZ13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r35, %r702, %r614;
setp.gt.s32 %p54, %r1330, 2;
@%p54 bra $L__BB15_44;

ld.param.u32 %r1333, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p57, %r1333, 1;
@%p57 bra $L__BB15_62;

ld.param.u32 %r1334, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p58, %r1334, 2;
@%p58 bra $L__BB15_36;
bra.uni $L__BB15_70;

$L__BB15_36:
ld.shared.f32 %f533, [%r33];
add.ftz.f32 %f534, %f533, %f29;
ld.shared.f32 %f535, [%r35];
add.ftz.f32 %f536, %f535, %f33;
ld.shared.f32 %f537, [%r33+512];
add.ftz.f32 %f538, %f537, %f534;
ld.shared.f32 %f539, [%r35+512];
add.ftz.f32 %f540, %f539, %f536;
ld.shared.f32 %f541, [%r33+1024];
add.ftz.f32 %f542, %f541, %f538;
ld.shared.f32 %f543, [%r35+1024];
add.ftz.f32 %f544, %f543, %f540;
mov.b32 %r949, %f542;
mov.u32 %r950, 31;
mov.u32 %r951, 16;
mov.u32 %r952, -1;
shfl.sync.bfly.b32 %r953|%p163, %r949, %r951, %r950, %r952;
mov.b32 %f545, %r953;
add.ftz.f32 %f546, %f542, %f545;
mov.b32 %r954, %f546;
mov.u32 %r955, 8;
shfl.sync.bfly.b32 %r956|%p164, %r954, %r955, %r950, %r952;
mov.b32 %f547, %r956;
add.ftz.f32 %f548, %f546, %f547;
mov.b32 %r957, %f548;
mov.u32 %r958, 4;
shfl.sync.bfly.b32 %r959|%p165, %r957, %r958, %r950, %r952;
mov.b32 %f549, %r959;
add.ftz.f32 %f550, %f548, %f549;
mov.b32 %r960, %f550;
mov.u32 %r961, 2;
shfl.sync.bfly.b32 %r962|%p166, %r960, %r961, %r950, %r952;
mov.b32 %f551, %r962;
add.ftz.f32 %f552, %f550, %f551;
mov.b32 %r963, %f552;
mov.u32 %r964, 1;
shfl.sync.bfly.b32 %r965|%p167, %r963, %r964, %r950, %r952;
mov.b32 %f553, %r965;
add.ftz.f32 %f56, %f552, %f553;
mov.b32 %r966, %f544;
shfl.sync.bfly.b32 %r967|%p168, %r966, %r951, %r950, %r952;
mov.b32 %f554, %r967;
add.ftz.f32 %f555, %f544, %f554;
mov.b32 %r968, %f555;
shfl.sync.bfly.b32 %r969|%p169, %r968, %r955, %r950, %r952;
mov.b32 %f556, %r969;
add.ftz.f32 %f557, %f555, %f556;
mov.b32 %r970, %f557;
shfl.sync.bfly.b32 %r971|%p170, %r970, %r958, %r950, %r952;
mov.b32 %f558, %r971;
add.ftz.f32 %f559, %f557, %f558;
mov.b32 %r972, %f559;
shfl.sync.bfly.b32 %r973|%p171, %r972, %r961, %r950, %r952;
mov.b32 %f560, %r973;
add.ftz.f32 %f561, %f559, %f560;
mov.b32 %r974, %f561;
shfl.sync.bfly.b32 %r975|%p172, %r974, %r964, %r950, %r952;
mov.b32 %f562, %r975;
add.ftz.f32 %f57, %f561, %f562;
setp.ne.s32 %p173, %r2, 0;
setp.ge.u32 %p174, %r4, %r52;
or.pred %p175, %p173, %p174;
@%p175 bra $L__BB15_38;

add.ftz.f32 %f563, %f5, %f56;
add.ftz.f32 %f564, %f4, %f57;
mul.ftz.f32 %f565, %f564, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f566, %f565;
add.ftz.f32 %f567, %f566, 0f3F800000;
div.approx.ftz.f32 %f568, %f564, %f567;
mul.ftz.f32 %f569, %f563, %f568;
st.global.f32 [%rd24], %f569;

$L__BB15_38:
ld.shared.f32 %f570, [%r33+128];
add.ftz.f32 %f571, %f570, %f28;
ld.shared.f32 %f572, [%r35+128];
add.ftz.f32 %f573, %f572, %f32;
ld.shared.f32 %f574, [%r33+640];
add.ftz.f32 %f575, %f574, %f571;
ld.shared.f32 %f576, [%r35+640];
add.ftz.f32 %f577, %f576, %f573;
ld.shared.f32 %f578, [%r33+1152];
add.ftz.f32 %f579, %f578, %f575;
ld.shared.f32 %f580, [%r35+1152];
add.ftz.f32 %f581, %f580, %f577;
mov.b32 %r979, %f579;
shfl.sync.bfly.b32 %r983|%p176, %r979, %r951, %r950, %r952;
mov.b32 %f582, %r983;
add.ftz.f32 %f583, %f579, %f582;
mov.b32 %r984, %f583;
shfl.sync.bfly.b32 %r986|%p177, %r984, %r955, %r950, %r952;
mov.b32 %f584, %r986;
add.ftz.f32 %f585, %f583, %f584;
mov.b32 %r987, %f585;
shfl.sync.bfly.b32 %r989|%p178, %r987, %r958, %r950, %r952;
mov.b32 %f586, %r989;
add.ftz.f32 %f587, %f585, %f586;
mov.b32 %r990, %f587;
shfl.sync.bfly.b32 %r992|%p179, %r990, %r961, %r950, %r952;
mov.b32 %f588, %r992;
add.ftz.f32 %f589, %f587, %f588;
mov.b32 %r993, %f589;
shfl.sync.bfly.b32 %r995|%p180, %r993, %r964, %r950, %r952;
mov.b32 %f590, %r995;
add.ftz.f32 %f58, %f589, %f590;
mov.b32 %r996, %f581;
shfl.sync.bfly.b32 %r997|%p181, %r996, %r951, %r950, %r952;
mov.b32 %f591, %r997;
add.ftz.f32 %f592, %f581, %f591;
mov.b32 %r998, %f592;
shfl.sync.bfly.b32 %r999|%p182, %r998, %r955, %r950, %r952;
mov.b32 %f593, %r999;
add.ftz.f32 %f594, %f592, %f593;
mov.b32 %r1000, %f594;
shfl.sync.bfly.b32 %r1001|%p183, %r1000, %r958, %r950, %r952;
mov.b32 %f595, %r1001;
add.ftz.f32 %f596, %f594, %f595;
mov.b32 %r1002, %f596;
shfl.sync.bfly.b32 %r1003|%p184, %r1002, %r961, %r950, %r952;
mov.b32 %f597, %r1003;
add.ftz.f32 %f598, %f596, %f597;
mov.b32 %r1004, %f598;
shfl.sync.bfly.b32 %r1005|%p185, %r1004, %r964, %r950, %r952;
mov.b32 %f599, %r1005;
add.ftz.f32 %f59, %f598, %f599;
setp.ne.s32 %p186, %r2, 1;
or.b32 %r1009, %r4, 1;
setp.ge.u32 %p187, %r1009, %r52;
or.pred %p188, %p186, %p187;
@%p188 bra $L__BB15_40;

add.ftz.f32 %f600, %f5, %f58;
add.ftz.f32 %f601, %f4, %f59;
mul.ftz.f32 %f602, %f601, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f603, %f602;
add.ftz.f32 %f604, %f603, 0f3F800000;
div.approx.ftz.f32 %f605, %f601, %f604;
mul.ftz.f32 %f606, %f600, %f605;
st.global.f32 [%rd24+4], %f606;

$L__BB15_40:
ld.shared.f32 %f607, [%r33+256];
add.ftz.f32 %f608, %f607, %f27;
ld.shared.f32 %f609, [%r35+256];
add.ftz.f32 %f610, %f609, %f31;
ld.shared.f32 %f611, [%r33+768];
add.ftz.f32 %f612, %f611, %f608;
ld.shared.f32 %f613, [%r35+768];
add.ftz.f32 %f614, %f613, %f610;
ld.shared.f32 %f615, [%r33+1280];
add.ftz.f32 %f616, %f615, %f612;
ld.shared.f32 %f617, [%r35+1280];
add.ftz.f32 %f618, %f617, %f614;
mov.b32 %r1010, %f616;
mov.u32 %r1011, 31;
mov.u32 %r1012, 16;
mov.u32 %r1013, -1;
shfl.sync.bfly.b32 %r1014|%p189, %r1010, %r1012, %r1011, %r1013;
mov.b32 %f619, %r1014;
add.ftz.f32 %f620, %f616, %f619;
mov.b32 %r1015, %f620;
mov.u32 %r1016, 8;
shfl.sync.bfly.b32 %r1017|%p190, %r1015, %r1016, %r1011, %r1013;
mov.b32 %f621, %r1017;
add.ftz.f32 %f622, %f620, %f621;
mov.b32 %r1018, %f622;
mov.u32 %r1019, 4;
shfl.sync.bfly.b32 %r1020|%p191, %r1018, %r1019, %r1011, %r1013;
mov.b32 %f623, %r1020;
add.ftz.f32 %f624, %f622, %f623;
mov.b32 %r1021, %f624;
mov.u32 %r1022, 2;
shfl.sync.bfly.b32 %r1023|%p192, %r1021, %r1022, %r1011, %r1013;
mov.b32 %f625, %r1023;
add.ftz.f32 %f626, %f624, %f625;
mov.b32 %r1024, %f626;
mov.u32 %r1025, 1;
shfl.sync.bfly.b32 %r1026|%p193, %r1024, %r1025, %r1011, %r1013;
mov.b32 %f627, %r1026;
add.ftz.f32 %f60, %f626, %f627;
mov.b32 %r1027, %f618;
shfl.sync.bfly.b32 %r1028|%p194, %r1027, %r1012, %r1011, %r1013;
mov.b32 %f628, %r1028;
add.ftz.f32 %f629, %f618, %f628;
mov.b32 %r1029, %f629;
shfl.sync.bfly.b32 %r1030|%p195, %r1029, %r1016, %r1011, %r1013;
mov.b32 %f630, %r1030;
add.ftz.f32 %f631, %f629, %f630;
mov.b32 %r1031, %f631;
shfl.sync.bfly.b32 %r1032|%p196, %r1031, %r1019, %r1011, %r1013;
mov.b32 %f632, %r1032;
add.ftz.f32 %f633, %f631, %f632;
mov.b32 %r1033, %f633;
shfl.sync.bfly.b32 %r1034|%p197, %r1033, %r1022, %r1011, %r1013;
mov.b32 %f634, %r1034;
add.ftz.f32 %f635, %f633, %f634;
mov.b32 %r1035, %f635;
shfl.sync.bfly.b32 %r1036|%p198, %r1035, %r1025, %r1011, %r1013;
mov.b32 %f636, %r1036;
add.ftz.f32 %f61, %f635, %f636;
setp.ne.s32 %p199, %r2, 2;
or.b32 %r1040, %r4, 2;
setp.ge.u32 %p200, %r1040, %r52;
or.pred %p201, %p199, %p200;
@%p201 bra $L__BB15_42;

add.ftz.f32 %f637, %f5, %f60;
add.ftz.f32 %f638, %f4, %f61;
mul.ftz.f32 %f639, %f638, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f640, %f639;
add.ftz.f32 %f641, %f640, 0f3F800000;
div.approx.ftz.f32 %f642, %f638, %f641;
mul.ftz.f32 %f643, %f637, %f642;
st.global.f32 [%rd24+8], %f643;

$L__BB15_42:
ld.shared.f32 %f644, [%r33+384];
add.ftz.f32 %f645, %f644, %f26;
ld.shared.f32 %f646, [%r35+384];
add.ftz.f32 %f647, %f646, %f30;
ld.shared.f32 %f648, [%r33+896];
add.ftz.f32 %f649, %f648, %f645;
ld.shared.f32 %f650, [%r35+896];
add.ftz.f32 %f651, %f650, %f647;
ld.shared.f32 %f652, [%r33+1408];
add.ftz.f32 %f653, %f652, %f649;
ld.shared.f32 %f654, [%r35+1408];
add.ftz.f32 %f655, %f654, %f651;
mov.b32 %r1041, %f653;
shfl.sync.bfly.b32 %r1045|%p202, %r1041, %r1012, %r1011, %r1013;
mov.b32 %f656, %r1045;
add.ftz.f32 %f657, %f653, %f656;
mov.b32 %r1046, %f657;
shfl.sync.bfly.b32 %r1048|%p203, %r1046, %r1016, %r1011, %r1013;
mov.b32 %f658, %r1048;
add.ftz.f32 %f659, %f657, %f658;
mov.b32 %r1049, %f659;
shfl.sync.bfly.b32 %r1051|%p204, %r1049, %r1019, %r1011, %r1013;
mov.b32 %f660, %r1051;
add.ftz.f32 %f661, %f659, %f660;
mov.b32 %r1052, %f661;
shfl.sync.bfly.b32 %r1054|%p205, %r1052, %r1022, %r1011, %r1013;
mov.b32 %f662, %r1054;
add.ftz.f32 %f663, %f661, %f662;
mov.b32 %r1055, %f663;
shfl.sync.bfly.b32 %r1057|%p206, %r1055, %r1025, %r1011, %r1013;
mov.b32 %f664, %r1057;
add.ftz.f32 %f62, %f663, %f664;
mov.b32 %r1058, %f655;
shfl.sync.bfly.b32 %r1059|%p207, %r1058, %r1012, %r1011, %r1013;
mov.b32 %f665, %r1059;
add.ftz.f32 %f666, %f655, %f665;
mov.b32 %r1060, %f666;
shfl.sync.bfly.b32 %r1061|%p208, %r1060, %r1016, %r1011, %r1013;
mov.b32 %f667, %r1061;
add.ftz.f32 %f668, %f666, %f667;
mov.b32 %r1062, %f668;
shfl.sync.bfly.b32 %r1063|%p209, %r1062, %r1019, %r1011, %r1013;
mov.b32 %f669, %r1063;
add.ftz.f32 %f670, %f668, %f669;
mov.b32 %r1064, %f670;
shfl.sync.bfly.b32 %r1065|%p210, %r1064, %r1022, %r1011, %r1013;
mov.b32 %f671, %r1065;
add.ftz.f32 %f672, %f670, %f671;
mov.b32 %r1066, %f672;
shfl.sync.bfly.b32 %r1067|%p211, %r1066, %r1025, %r1011, %r1013;
mov.b32 %f673, %r1067;
add.ftz.f32 %f63, %f672, %f673;
setp.ne.s32 %p212, %r2, 3;
or.b32 %r1071, %r4, 3;
setp.ge.u32 %p213, %r1071, %r52;
or.pred %p214, %p212, %p213;
@%p214 bra $L__BB15_78;

add.ftz.f32 %f674, %f5, %f62;
add.ftz.f32 %f675, %f4, %f63;
mul.ftz.f32 %f676, %f675, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f677, %f676;
add.ftz.f32 %f678, %f677, 0f3F800000;
div.approx.ftz.f32 %f679, %f675, %f678;
mul.ftz.f32 %f680, %f674, %f679;
st.global.f32 [%rd24+12], %f680;
bra.uni $L__BB15_78;

$L__BB15_25:
ld.shared.f32 %f133, [%r33];
add.ftz.f32 %f134, %f133, %f29;
ld.shared.f32 %f135, [%r33+512];
add.ftz.f32 %f136, %f135, %f134;
ld.shared.f32 %f137, [%r33+1024];
add.ftz.f32 %f138, %f137, %f136;
mov.b32 %r616, %f138;
mov.u32 %r617, 31;
mov.u32 %r618, 16;
mov.u32 %r619, -1;
shfl.sync.bfly.b32 %r620|%p22, %r616, %r618, %r617, %r619;
mov.b32 %f139, %r620;
add.ftz.f32 %f140, %f138, %f139;
mov.b32 %r621, %f140;
mov.u32 %r622, 8;
shfl.sync.bfly.b32 %r623|%p23, %r621, %r622, %r617, %r619;
mov.b32 %f141, %r623;
add.ftz.f32 %f142, %f140, %f141;
mov.b32 %r624, %f142;
mov.u32 %r625, 4;
shfl.sync.bfly.b32 %r626|%p24, %r624, %r625, %r617, %r619;
mov.b32 %f143, %r626;
add.ftz.f32 %f144, %f142, %f143;
mov.b32 %r627, %f144;
mov.u32 %r628, 2;
shfl.sync.bfly.b32 %r629|%p25, %r627, %r628, %r617, %r619;
mov.b32 %f145, %r629;
add.ftz.f32 %f146, %f144, %f145;
mov.b32 %r630, %f146;
mov.u32 %r631, 1;
shfl.sync.bfly.b32 %r632|%p26, %r630, %r631, %r617, %r619;
mov.b32 %f147, %r632;
add.ftz.f32 %f36, %f146, %f147;
setp.ne.s32 %p27, %r2, 0;
setp.ge.u32 %p28, %r4, %r52;
or.pred %p29, %p27, %p28;
@%p29 bra $L__BB15_27;

add.ftz.f32 %f148, %f5, %f36;
st.global.f32 [%rd24], %f148;

$L__BB15_27:
ld.shared.f32 %f149, [%r33+128];
add.ftz.f32 %f150, %f149, %f28;
ld.shared.f32 %f151, [%r33+640];
add.ftz.f32 %f152, %f151, %f150;
ld.shared.f32 %f153, [%r33+1152];
add.ftz.f32 %f154, %f153, %f152;
mov.b32 %r636, %f154;
shfl.sync.bfly.b32 %r640|%p30, %r636, %r618, %r617, %r619;
mov.b32 %f155, %r640;
add.ftz.f32 %f156, %f154, %f155;
mov.b32 %r641, %f156;
shfl.sync.bfly.b32 %r643|%p31, %r641, %r622, %r617, %r619;
mov.b32 %f157, %r643;
add.ftz.f32 %f158, %f156, %f157;
mov.b32 %r644, %f158;
shfl.sync.bfly.b32 %r646|%p32, %r644, %r625, %r617, %r619;
mov.b32 %f159, %r646;
add.ftz.f32 %f160, %f158, %f159;
mov.b32 %r647, %f160;
shfl.sync.bfly.b32 %r649|%p33, %r647, %r628, %r617, %r619;
mov.b32 %f161, %r649;
add.ftz.f32 %f162, %f160, %f161;
mov.b32 %r650, %f162;
shfl.sync.bfly.b32 %r652|%p34, %r650, %r631, %r617, %r619;
mov.b32 %f163, %r652;
add.ftz.f32 %f37, %f162, %f163;
setp.ne.s32 %p35, %r2, 1;
or.b32 %r656, %r4, 1;
setp.ge.u32 %p36, %r656, %r52;
or.pred %p37, %p35, %p36;
@%p37 bra $L__BB15_29;

add.ftz.f32 %f164, %f5, %f37;
st.global.f32 [%rd24+4], %f164;

$L__BB15_29:
ld.shared.f32 %f165, [%r33+256];
add.ftz.f32 %f166, %f165, %f27;
ld.shared.f32 %f167, [%r33+768];
add.ftz.f32 %f168, %f167, %f166;
ld.shared.f32 %f169, [%r33+1280];
add.ftz.f32 %f170, %f169, %f168;
mov.b32 %r657, %f170;
mov.u32 %r658, 31;
mov.u32 %r659, 16;
mov.u32 %r660, -1;
shfl.sync.bfly.b32 %r661|%p38, %r657, %r659, %r658, %r660;
mov.b32 %f171, %r661;
add.ftz.f32 %f172, %f170, %f171;
mov.b32 %r662, %f172;
mov.u32 %r663, 8;
shfl.sync.bfly.b32 %r664|%p39, %r662, %r663, %r658, %r660;
mov.b32 %f173, %r664;
add.ftz.f32 %f174, %f172, %f173;
mov.b32 %r665, %f174;
mov.u32 %r666, 4;
shfl.sync.bfly.b32 %r667|%p40, %r665, %r666, %r658, %r660;
mov.b32 %f175, %r667;
add.ftz.f32 %f176, %f174, %f175;
mov.b32 %r668, %f176;
mov.u32 %r669, 2;
shfl.sync.bfly.b32 %r670|%p41, %r668, %r669, %r658, %r660;
mov.b32 %f177, %r670;
add.ftz.f32 %f178, %f176, %f177;
mov.b32 %r671, %f178;
mov.u32 %r672, 1;
shfl.sync.bfly.b32 %r673|%p42, %r671, %r672, %r658, %r660;
mov.b32 %f179, %r673;
add.ftz.f32 %f38, %f178, %f179;
setp.ne.s32 %p43, %r2, 2;
or.b32 %r677, %r4, 2;
setp.ge.u32 %p44, %r677, %r52;
or.pred %p45, %p43, %p44;
@%p45 bra $L__BB15_31;

add.ftz.f32 %f180, %f5, %f38;
st.global.f32 [%rd24+8], %f180;

$L__BB15_31:
ld.shared.f32 %f181, [%r33+384];
add.ftz.f32 %f182, %f181, %f26;
ld.shared.f32 %f183, [%r33+896];
add.ftz.f32 %f184, %f183, %f182;
ld.shared.f32 %f185, [%r33+1408];
add.ftz.f32 %f186, %f185, %f184;
mov.b32 %r678, %f186;
shfl.sync.bfly.b32 %r682|%p46, %r678, %r659, %r658, %r660;
mov.b32 %f187, %r682;
add.ftz.f32 %f188, %f186, %f187;
mov.b32 %r683, %f188;
shfl.sync.bfly.b32 %r685|%p47, %r683, %r663, %r658, %r660;
mov.b32 %f189, %r685;
add.ftz.f32 %f190, %f188, %f189;
mov.b32 %r686, %f190;
shfl.sync.bfly.b32 %r688|%p48, %r686, %r666, %r658, %r660;
mov.b32 %f191, %r688;
add.ftz.f32 %f192, %f190, %f191;
mov.b32 %r689, %f192;
shfl.sync.bfly.b32 %r691|%p49, %r689, %r669, %r658, %r660;
mov.b32 %f193, %r691;
add.ftz.f32 %f194, %f192, %f193;
mov.b32 %r692, %f194;
shfl.sync.bfly.b32 %r694|%p50, %r692, %r672, %r658, %r660;
mov.b32 %f195, %r694;
add.ftz.f32 %f39, %f194, %f195;
setp.ne.s32 %p51, %r2, 3;
or.b32 %r698, %r4, 3;
setp.ge.u32 %p52, %r698, %r52;
or.pred %p53, %p51, %p52;
@%p53 bra $L__BB15_78;

add.ftz.f32 %f196, %f5, %f39;
st.global.f32 [%rd24+12], %f196;
bra.uni $L__BB15_78;

$L__BB15_44:
ld.param.u32 %r1331, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p55, %r1331, 3;
@%p55 bra $L__BB15_54;

ld.param.u32 %r1332, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.ne.s32 %p56, %r1332, 6;
@%p56 bra $L__BB15_70;

ld.shared.f32 %f197, [%r33];
add.ftz.f32 %f198, %f197, %f29;
ld.shared.f32 %f199, [%r35];
add.ftz.f32 %f200, %f199, %f33;
ld.shared.f32 %f201, [%r33+512];
add.ftz.f32 %f202, %f201, %f198;
ld.shared.f32 %f203, [%r35+512];
add.ftz.f32 %f204, %f203, %f200;
ld.shared.f32 %f205, [%r33+1024];
add.ftz.f32 %f206, %f205, %f202;
ld.shared.f32 %f207, [%r35+1024];
add.ftz.f32 %f208, %f207, %f204;
mov.b32 %r703, %f206;
mov.u32 %r704, 31;
mov.u32 %r705, 16;
mov.u32 %r706, -1;
shfl.sync.bfly.b32 %r707|%p59, %r703, %r705, %r704, %r706;
mov.b32 %f209, %r707;
add.ftz.f32 %f210, %f206, %f209;
mov.b32 %r708, %f210;
mov.u32 %r709, 8;
shfl.sync.bfly.b32 %r710|%p60, %r708, %r709, %r704, %r706;
mov.b32 %f211, %r710;
add.ftz.f32 %f212, %f210, %f211;
mov.b32 %r711, %f212;
mov.u32 %r712, 4;
shfl.sync.bfly.b32 %r713|%p61, %r711, %r712, %r704, %r706;
mov.b32 %f213, %r713;
add.ftz.f32 %f214, %f212, %f213;
mov.b32 %r714, %f214;
mov.u32 %r715, 2;
shfl.sync.bfly.b32 %r716|%p62, %r714, %r715, %r704, %r706;
mov.b32 %f215, %r716;
add.ftz.f32 %f216, %f214, %f215;
mov.b32 %r717, %f216;
mov.u32 %r718, 1;
shfl.sync.bfly.b32 %r719|%p63, %r717, %r718, %r704, %r706;
mov.b32 %f217, %r719;
add.ftz.f32 %f40, %f216, %f217;
mov.b32 %r720, %f208;
shfl.sync.bfly.b32 %r721|%p64, %r720, %r705, %r704, %r706;
mov.b32 %f218, %r721;
add.ftz.f32 %f219, %f208, %f218;
mov.b32 %r722, %f219;
shfl.sync.bfly.b32 %r723|%p65, %r722, %r709, %r704, %r706;
mov.b32 %f220, %r723;
add.ftz.f32 %f221, %f219, %f220;
mov.b32 %r724, %f221;
shfl.sync.bfly.b32 %r725|%p66, %r724, %r712, %r704, %r706;
mov.b32 %f222, %r725;
add.ftz.f32 %f223, %f221, %f222;
mov.b32 %r726, %f223;
shfl.sync.bfly.b32 %r727|%p67, %r726, %r715, %r704, %r706;
mov.b32 %f224, %r727;
add.ftz.f32 %f225, %f223, %f224;
mov.b32 %r728, %f225;
shfl.sync.bfly.b32 %r729|%p68, %r728, %r718, %r704, %r706;
mov.b32 %f226, %r729;
add.ftz.f32 %f41, %f225, %f226;
setp.ne.s32 %p69, %r2, 0;
setp.ge.u32 %p70, %r4, %r52;
or.pred %p71, %p69, %p70;
@%p71 bra $L__BB15_48;

ld.param.f32 %f981, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f227, %f5, %f40;
add.ftz.f32 %f228, %f4, %f41;
min.ftz.f32 %f229, %f228, %f981;
min.ftz.f32 %f230, %f227, %f981;
max.ftz.f32 %f231, %f230, %f35;
mul.ftz.f32 %f232, %f229, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f233, %f232;
add.ftz.f32 %f234, %f233, 0f3F800000;
div.approx.ftz.f32 %f235, %f229, %f234;
mul.ftz.f32 %f236, %f231, %f235;
st.global.f32 [%rd24], %f236;

$L__BB15_48:
ld.shared.f32 %f237, [%r33+128];
add.ftz.f32 %f238, %f237, %f28;
ld.shared.f32 %f239, [%r35+128];
add.ftz.f32 %f240, %f239, %f32;
ld.shared.f32 %f241, [%r33+640];
add.ftz.f32 %f242, %f241, %f238;
ld.shared.f32 %f243, [%r35+640];
add.ftz.f32 %f244, %f243, %f240;
ld.shared.f32 %f245, [%r33+1152];
add.ftz.f32 %f246, %f245, %f242;
ld.shared.f32 %f247, [%r35+1152];
add.ftz.f32 %f248, %f247, %f244;
mov.b32 %r733, %f246;
shfl.sync.bfly.b32 %r737|%p72, %r733, %r705, %r704, %r706;
mov.b32 %f249, %r737;
add.ftz.f32 %f250, %f246, %f249;
mov.b32 %r738, %f250;
shfl.sync.bfly.b32 %r740|%p73, %r738, %r709, %r704, %r706;
mov.b32 %f251, %r740;
add.ftz.f32 %f252, %f250, %f251;
mov.b32 %r741, %f252;
shfl.sync.bfly.b32 %r743|%p74, %r741, %r712, %r704, %r706;
mov.b32 %f253, %r743;
add.ftz.f32 %f254, %f252, %f253;
mov.b32 %r744, %f254;
shfl.sync.bfly.b32 %r746|%p75, %r744, %r715, %r704, %r706;
mov.b32 %f255, %r746;
add.ftz.f32 %f256, %f254, %f255;
mov.b32 %r747, %f256;
shfl.sync.bfly.b32 %r749|%p76, %r747, %r718, %r704, %r706;
mov.b32 %f257, %r749;
add.ftz.f32 %f42, %f256, %f257;
mov.b32 %r750, %f248;
shfl.sync.bfly.b32 %r751|%p77, %r750, %r705, %r704, %r706;
mov.b32 %f258, %r751;
add.ftz.f32 %f259, %f248, %f258;
mov.b32 %r752, %f259;
shfl.sync.bfly.b32 %r753|%p78, %r752, %r709, %r704, %r706;
mov.b32 %f260, %r753;
add.ftz.f32 %f261, %f259, %f260;
mov.b32 %r754, %f261;
shfl.sync.bfly.b32 %r755|%p79, %r754, %r712, %r704, %r706;
mov.b32 %f262, %r755;
add.ftz.f32 %f263, %f261, %f262;
mov.b32 %r756, %f263;
shfl.sync.bfly.b32 %r757|%p80, %r756, %r715, %r704, %r706;
mov.b32 %f264, %r757;
add.ftz.f32 %f265, %f263, %f264;
mov.b32 %r758, %f265;
shfl.sync.bfly.b32 %r759|%p81, %r758, %r718, %r704, %r706;
mov.b32 %f266, %r759;
add.ftz.f32 %f43, %f265, %f266;
setp.ne.s32 %p82, %r2, 1;
or.b32 %r763, %r4, 1;
setp.ge.u32 %p83, %r763, %r52;
or.pred %p84, %p82, %p83;
@%p84 bra $L__BB15_50;

ld.param.f32 %f980, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f267, %f5, %f42;
add.ftz.f32 %f268, %f4, %f43;
min.ftz.f32 %f269, %f268, %f980;
min.ftz.f32 %f270, %f267, %f980;
max.ftz.f32 %f271, %f270, %f35;
mul.ftz.f32 %f272, %f269, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f273, %f272;
add.ftz.f32 %f274, %f273, 0f3F800000;
div.approx.ftz.f32 %f275, %f269, %f274;
mul.ftz.f32 %f276, %f271, %f275;
st.global.f32 [%rd24+4], %f276;

$L__BB15_50:
ld.shared.f32 %f277, [%r33+256];
add.ftz.f32 %f278, %f277, %f27;
ld.shared.f32 %f279, [%r35+256];
add.ftz.f32 %f280, %f279, %f31;
ld.shared.f32 %f281, [%r33+768];
add.ftz.f32 %f282, %f281, %f278;
ld.shared.f32 %f283, [%r35+768];
add.ftz.f32 %f284, %f283, %f280;
ld.shared.f32 %f285, [%r33+1280];
add.ftz.f32 %f286, %f285, %f282;
ld.shared.f32 %f287, [%r35+1280];
add.ftz.f32 %f288, %f287, %f284;
mov.b32 %r764, %f286;
mov.u32 %r765, 31;
mov.u32 %r766, 16;
mov.u32 %r767, -1;
shfl.sync.bfly.b32 %r768|%p85, %r764, %r766, %r765, %r767;
mov.b32 %f289, %r768;
add.ftz.f32 %f290, %f286, %f289;
mov.b32 %r769, %f290;
mov.u32 %r770, 8;
shfl.sync.bfly.b32 %r771|%p86, %r769, %r770, %r765, %r767;
mov.b32 %f291, %r771;
add.ftz.f32 %f292, %f290, %f291;
mov.b32 %r772, %f292;
mov.u32 %r773, 4;
shfl.sync.bfly.b32 %r774|%p87, %r772, %r773, %r765, %r767;
mov.b32 %f293, %r774;
add.ftz.f32 %f294, %f292, %f293;
mov.b32 %r775, %f294;
mov.u32 %r776, 2;
shfl.sync.bfly.b32 %r777|%p88, %r775, %r776, %r765, %r767;
mov.b32 %f295, %r777;
add.ftz.f32 %f296, %f294, %f295;
mov.b32 %r778, %f296;
mov.u32 %r779, 1;
shfl.sync.bfly.b32 %r780|%p89, %r778, %r779, %r765, %r767;
mov.b32 %f297, %r780;
add.ftz.f32 %f44, %f296, %f297;
mov.b32 %r781, %f288;
shfl.sync.bfly.b32 %r782|%p90, %r781, %r766, %r765, %r767;
mov.b32 %f298, %r782;
add.ftz.f32 %f299, %f288, %f298;
mov.b32 %r783, %f299;
shfl.sync.bfly.b32 %r784|%p91, %r783, %r770, %r765, %r767;
mov.b32 %f300, %r784;
add.ftz.f32 %f301, %f299, %f300;
mov.b32 %r785, %f301;
shfl.sync.bfly.b32 %r786|%p92, %r785, %r773, %r765, %r767;
mov.b32 %f302, %r786;
add.ftz.f32 %f303, %f301, %f302;
mov.b32 %r787, %f303;
shfl.sync.bfly.b32 %r788|%p93, %r787, %r776, %r765, %r767;
mov.b32 %f304, %r788;
add.ftz.f32 %f305, %f303, %f304;
mov.b32 %r789, %f305;
shfl.sync.bfly.b32 %r790|%p94, %r789, %r779, %r765, %r767;
mov.b32 %f306, %r790;
add.ftz.f32 %f45, %f305, %f306;
setp.ne.s32 %p95, %r2, 2;
or.b32 %r794, %r4, 2;
setp.ge.u32 %p96, %r794, %r52;
or.pred %p97, %p95, %p96;
@%p97 bra $L__BB15_52;

ld.param.f32 %f979, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f307, %f5, %f44;
add.ftz.f32 %f308, %f4, %f45;
min.ftz.f32 %f309, %f308, %f979;
min.ftz.f32 %f310, %f307, %f979;
max.ftz.f32 %f311, %f310, %f35;
mul.ftz.f32 %f312, %f309, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f313, %f312;
add.ftz.f32 %f314, %f313, 0f3F800000;
div.approx.ftz.f32 %f315, %f309, %f314;
mul.ftz.f32 %f316, %f311, %f315;
st.global.f32 [%rd24+8], %f316;

$L__BB15_52:
ld.shared.f32 %f317, [%r33+384];
add.ftz.f32 %f318, %f317, %f26;
ld.shared.f32 %f319, [%r35+384];
add.ftz.f32 %f320, %f319, %f30;
ld.shared.f32 %f321, [%r33+896];
add.ftz.f32 %f322, %f321, %f318;
ld.shared.f32 %f323, [%r35+896];
add.ftz.f32 %f324, %f323, %f320;
ld.shared.f32 %f325, [%r33+1408];
add.ftz.f32 %f326, %f325, %f322;
ld.shared.f32 %f327, [%r35+1408];
add.ftz.f32 %f328, %f327, %f324;
mov.b32 %r795, %f326;
shfl.sync.bfly.b32 %r799|%p98, %r795, %r766, %r765, %r767;
mov.b32 %f329, %r799;
add.ftz.f32 %f330, %f326, %f329;
mov.b32 %r800, %f330;
shfl.sync.bfly.b32 %r802|%p99, %r800, %r770, %r765, %r767;
mov.b32 %f331, %r802;
add.ftz.f32 %f332, %f330, %f331;
mov.b32 %r803, %f332;
shfl.sync.bfly.b32 %r805|%p100, %r803, %r773, %r765, %r767;
mov.b32 %f333, %r805;
add.ftz.f32 %f334, %f332, %f333;
mov.b32 %r806, %f334;
shfl.sync.bfly.b32 %r808|%p101, %r806, %r776, %r765, %r767;
mov.b32 %f335, %r808;
add.ftz.f32 %f336, %f334, %f335;
mov.b32 %r809, %f336;
shfl.sync.bfly.b32 %r811|%p102, %r809, %r779, %r765, %r767;
mov.b32 %f337, %r811;
add.ftz.f32 %f46, %f336, %f337;
mov.b32 %r812, %f328;
shfl.sync.bfly.b32 %r813|%p103, %r812, %r766, %r765, %r767;
mov.b32 %f338, %r813;
add.ftz.f32 %f339, %f328, %f338;
mov.b32 %r814, %f339;
shfl.sync.bfly.b32 %r815|%p104, %r814, %r770, %r765, %r767;
mov.b32 %f340, %r815;
add.ftz.f32 %f341, %f339, %f340;
mov.b32 %r816, %f341;
shfl.sync.bfly.b32 %r817|%p105, %r816, %r773, %r765, %r767;
mov.b32 %f342, %r817;
add.ftz.f32 %f343, %f341, %f342;
mov.b32 %r818, %f343;
shfl.sync.bfly.b32 %r819|%p106, %r818, %r776, %r765, %r767;
mov.b32 %f344, %r819;
add.ftz.f32 %f345, %f343, %f344;
mov.b32 %r820, %f345;
shfl.sync.bfly.b32 %r821|%p107, %r820, %r779, %r765, %r767;
mov.b32 %f346, %r821;
add.ftz.f32 %f47, %f345, %f346;
setp.ne.s32 %p108, %r2, 3;
or.b32 %r825, %r4, 3;
setp.ge.u32 %p109, %r825, %r52;
or.pred %p110, %p108, %p109;
@%p110 bra $L__BB15_78;

ld.param.f32 %f978, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f347, %f5, %f46;
add.ftz.f32 %f348, %f4, %f47;
min.ftz.f32 %f349, %f348, %f978;
min.ftz.f32 %f350, %f347, %f978;
max.ftz.f32 %f351, %f350, %f35;
mul.ftz.f32 %f352, %f349, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f353, %f352;
add.ftz.f32 %f354, %f353, 0f3F800000;
div.approx.ftz.f32 %f355, %f349, %f354;
mul.ftz.f32 %f356, %f351, %f355;
st.global.f32 [%rd24+12], %f356;
bra.uni $L__BB15_78;

$L__BB15_70:
ld.shared.f32 %f845, [%r33];
add.ftz.f32 %f846, %f845, %f29;
ld.shared.f32 %f847, [%r35];
add.ftz.f32 %f848, %f847, %f33;
ld.shared.f32 %f849, [%r33+512];
add.ftz.f32 %f850, %f849, %f846;
ld.shared.f32 %f851, [%r35+512];
add.ftz.f32 %f852, %f851, %f848;
ld.shared.f32 %f853, [%r33+1024];
add.ftz.f32 %f854, %f853, %f850;
ld.shared.f32 %f855, [%r35+1024];
add.ftz.f32 %f856, %f855, %f852;
mov.b32 %r1195, %f854;
mov.u32 %r1196, 31;
mov.u32 %r1197, 16;
mov.u32 %r1198, -1;
shfl.sync.bfly.b32 %r1199|%p267, %r1195, %r1197, %r1196, %r1198;
mov.b32 %f857, %r1199;
add.ftz.f32 %f858, %f854, %f857;
mov.b32 %r1200, %f858;
mov.u32 %r1201, 8;
shfl.sync.bfly.b32 %r1202|%p268, %r1200, %r1201, %r1196, %r1198;
mov.b32 %f859, %r1202;
add.ftz.f32 %f860, %f858, %f859;
mov.b32 %r1203, %f860;
mov.u32 %r1204, 4;
shfl.sync.bfly.b32 %r1205|%p269, %r1203, %r1204, %r1196, %r1198;
mov.b32 %f861, %r1205;
add.ftz.f32 %f862, %f860, %f861;
mov.b32 %r1206, %f862;
mov.u32 %r1207, 2;
shfl.sync.bfly.b32 %r1208|%p270, %r1206, %r1207, %r1196, %r1198;
mov.b32 %f863, %r1208;
add.ftz.f32 %f864, %f862, %f863;
mov.b32 %r1209, %f864;
mov.u32 %r1210, 1;
shfl.sync.bfly.b32 %r1211|%p271, %r1209, %r1210, %r1196, %r1198;
mov.b32 %f865, %r1211;
add.ftz.f32 %f72, %f864, %f865;
mov.b32 %r1212, %f856;
shfl.sync.bfly.b32 %r1213|%p272, %r1212, %r1197, %r1196, %r1198;
mov.b32 %f866, %r1213;
add.ftz.f32 %f867, %f856, %f866;
mov.b32 %r1214, %f867;
shfl.sync.bfly.b32 %r1215|%p273, %r1214, %r1201, %r1196, %r1198;
mov.b32 %f868, %r1215;
add.ftz.f32 %f869, %f867, %f868;
mov.b32 %r1216, %f869;
shfl.sync.bfly.b32 %r1217|%p274, %r1216, %r1204, %r1196, %r1198;
mov.b32 %f870, %r1217;
add.ftz.f32 %f871, %f869, %f870;
mov.b32 %r1218, %f871;
shfl.sync.bfly.b32 %r1219|%p275, %r1218, %r1207, %r1196, %r1198;
mov.b32 %f872, %r1219;
add.ftz.f32 %f873, %f871, %f872;
mov.b32 %r1220, %f873;
shfl.sync.bfly.b32 %r1221|%p276, %r1220, %r1210, %r1196, %r1198;
mov.b32 %f874, %r1221;
add.ftz.f32 %f73, %f873, %f874;
setp.ne.s32 %p277, %r2, 0;
setp.ge.u32 %p278, %r4, %r52;
or.pred %p279, %p277, %p278;
@%p279 bra $L__BB15_72;

add.ftz.f32 %f875, %f5, %f72;
add.ftz.f32 %f876, %f4, %f73;
mul.ftz.f32 %f877, %f875, %f876;
st.global.f32 [%rd24], %f877;

$L__BB15_72:
ld.shared.f32 %f878, [%r33+128];
add.ftz.f32 %f879, %f878, %f28;
ld.shared.f32 %f880, [%r35+128];
add.ftz.f32 %f881, %f880, %f32;
ld.shared.f32 %f882, [%r33+640];
add.ftz.f32 %f883, %f882, %f879;
ld.shared.f32 %f884, [%r35+640];
add.ftz.f32 %f885, %f884, %f881;
ld.shared.f32 %f886, [%r33+1152];
add.ftz.f32 %f887, %f886, %f883;
ld.shared.f32 %f888, [%r35+1152];
add.ftz.f32 %f889, %f888, %f885;
mov.b32 %r1225, %f887;
shfl.sync.bfly.b32 %r1229|%p280, %r1225, %r1197, %r1196, %r1198;
mov.b32 %f890, %r1229;
add.ftz.f32 %f891, %f887, %f890;
mov.b32 %r1230, %f891;
shfl.sync.bfly.b32 %r1232|%p281, %r1230, %r1201, %r1196, %r1198;
mov.b32 %f892, %r1232;
add.ftz.f32 %f893, %f891, %f892;
mov.b32 %r1233, %f893;
shfl.sync.bfly.b32 %r1235|%p282, %r1233, %r1204, %r1196, %r1198;
mov.b32 %f894, %r1235;
add.ftz.f32 %f895, %f893, %f894;
mov.b32 %r1236, %f895;
shfl.sync.bfly.b32 %r1238|%p283, %r1236, %r1207, %r1196, %r1198;
mov.b32 %f896, %r1238;
add.ftz.f32 %f897, %f895, %f896;
mov.b32 %r1239, %f897;
shfl.sync.bfly.b32 %r1241|%p284, %r1239, %r1210, %r1196, %r1198;
mov.b32 %f898, %r1241;
add.ftz.f32 %f74, %f897, %f898;
mov.b32 %r1242, %f889;
shfl.sync.bfly.b32 %r1243|%p285, %r1242, %r1197, %r1196, %r1198;
mov.b32 %f899, %r1243;
add.ftz.f32 %f900, %f889, %f899;
mov.b32 %r1244, %f900;
shfl.sync.bfly.b32 %r1245|%p286, %r1244, %r1201, %r1196, %r1198;
mov.b32 %f901, %r1245;
add.ftz.f32 %f902, %f900, %f901;
mov.b32 %r1246, %f902;
shfl.sync.bfly.b32 %r1247|%p287, %r1246, %r1204, %r1196, %r1198;
mov.b32 %f903, %r1247;
add.ftz.f32 %f904, %f902, %f903;
mov.b32 %r1248, %f904;
shfl.sync.bfly.b32 %r1249|%p288, %r1248, %r1207, %r1196, %r1198;
mov.b32 %f905, %r1249;
add.ftz.f32 %f906, %f904, %f905;
mov.b32 %r1250, %f906;
shfl.sync.bfly.b32 %r1251|%p289, %r1250, %r1210, %r1196, %r1198;
mov.b32 %f907, %r1251;
add.ftz.f32 %f75, %f906, %f907;
setp.ne.s32 %p290, %r2, 1;
or.b32 %r1255, %r4, 1;
setp.ge.u32 %p291, %r1255, %r52;
or.pred %p292, %p290, %p291;
@%p292 bra $L__BB15_74;

add.ftz.f32 %f908, %f5, %f74;
add.ftz.f32 %f909, %f4, %f75;
mul.ftz.f32 %f910, %f908, %f909;
st.global.f32 [%rd24+4], %f910;

$L__BB15_74:
ld.shared.f32 %f911, [%r33+256];
add.ftz.f32 %f912, %f911, %f27;
ld.shared.f32 %f913, [%r35+256];
add.ftz.f32 %f914, %f913, %f31;
ld.shared.f32 %f915, [%r33+768];
add.ftz.f32 %f916, %f915, %f912;
ld.shared.f32 %f917, [%r35+768];
add.ftz.f32 %f918, %f917, %f914;
ld.shared.f32 %f919, [%r33+1280];
add.ftz.f32 %f920, %f919, %f916;
ld.shared.f32 %f921, [%r35+1280];
add.ftz.f32 %f922, %f921, %f918;
mov.b32 %r1256, %f920;
mov.u32 %r1257, 31;
mov.u32 %r1258, 16;
mov.u32 %r1259, -1;
shfl.sync.bfly.b32 %r1260|%p293, %r1256, %r1258, %r1257, %r1259;
mov.b32 %f923, %r1260;
add.ftz.f32 %f924, %f920, %f923;
mov.b32 %r1261, %f924;
mov.u32 %r1262, 8;
shfl.sync.bfly.b32 %r1263|%p294, %r1261, %r1262, %r1257, %r1259;
mov.b32 %f925, %r1263;
add.ftz.f32 %f926, %f924, %f925;
mov.b32 %r1264, %f926;
mov.u32 %r1265, 4;
shfl.sync.bfly.b32 %r1266|%p295, %r1264, %r1265, %r1257, %r1259;
mov.b32 %f927, %r1266;
add.ftz.f32 %f928, %f926, %f927;
mov.b32 %r1267, %f928;
mov.u32 %r1268, 2;
shfl.sync.bfly.b32 %r1269|%p296, %r1267, %r1268, %r1257, %r1259;
mov.b32 %f929, %r1269;
add.ftz.f32 %f930, %f928, %f929;
mov.b32 %r1270, %f930;
mov.u32 %r1271, 1;
shfl.sync.bfly.b32 %r1272|%p297, %r1270, %r1271, %r1257, %r1259;
mov.b32 %f931, %r1272;
add.ftz.f32 %f76, %f930, %f931;
mov.b32 %r1273, %f922;
shfl.sync.bfly.b32 %r1274|%p298, %r1273, %r1258, %r1257, %r1259;
mov.b32 %f932, %r1274;
add.ftz.f32 %f933, %f922, %f932;
mov.b32 %r1275, %f933;
shfl.sync.bfly.b32 %r1276|%p299, %r1275, %r1262, %r1257, %r1259;
mov.b32 %f934, %r1276;
add.ftz.f32 %f935, %f933, %f934;
mov.b32 %r1277, %f935;
shfl.sync.bfly.b32 %r1278|%p300, %r1277, %r1265, %r1257, %r1259;
mov.b32 %f936, %r1278;
add.ftz.f32 %f937, %f935, %f936;
mov.b32 %r1279, %f937;
shfl.sync.bfly.b32 %r1280|%p301, %r1279, %r1268, %r1257, %r1259;
mov.b32 %f938, %r1280;
add.ftz.f32 %f939, %f937, %f938;
mov.b32 %r1281, %f939;
shfl.sync.bfly.b32 %r1282|%p302, %r1281, %r1271, %r1257, %r1259;
mov.b32 %f940, %r1282;
add.ftz.f32 %f77, %f939, %f940;
setp.ne.s32 %p303, %r2, 2;
or.b32 %r1286, %r4, 2;
setp.ge.u32 %p304, %r1286, %r52;
or.pred %p305, %p303, %p304;
@%p305 bra $L__BB15_76;

add.ftz.f32 %f941, %f5, %f76;
add.ftz.f32 %f942, %f4, %f77;
mul.ftz.f32 %f943, %f941, %f942;
st.global.f32 [%rd24+8], %f943;

$L__BB15_76:
ld.shared.f32 %f944, [%r33+384];
add.ftz.f32 %f945, %f944, %f26;
ld.shared.f32 %f946, [%r35+384];
add.ftz.f32 %f947, %f946, %f30;
ld.shared.f32 %f948, [%r33+896];
add.ftz.f32 %f949, %f948, %f945;
ld.shared.f32 %f950, [%r35+896];
add.ftz.f32 %f951, %f950, %f947;
ld.shared.f32 %f952, [%r33+1408];
add.ftz.f32 %f953, %f952, %f949;
ld.shared.f32 %f954, [%r35+1408];
add.ftz.f32 %f955, %f954, %f951;
mov.b32 %r1287, %f953;
shfl.sync.bfly.b32 %r1291|%p306, %r1287, %r1258, %r1257, %r1259;
mov.b32 %f956, %r1291;
add.ftz.f32 %f957, %f953, %f956;
mov.b32 %r1292, %f957;
shfl.sync.bfly.b32 %r1294|%p307, %r1292, %r1262, %r1257, %r1259;
mov.b32 %f958, %r1294;
add.ftz.f32 %f959, %f957, %f958;
mov.b32 %r1295, %f959;
shfl.sync.bfly.b32 %r1297|%p308, %r1295, %r1265, %r1257, %r1259;
mov.b32 %f960, %r1297;
add.ftz.f32 %f961, %f959, %f960;
mov.b32 %r1298, %f961;
shfl.sync.bfly.b32 %r1300|%p309, %r1298, %r1268, %r1257, %r1259;
mov.b32 %f962, %r1300;
add.ftz.f32 %f963, %f961, %f962;
mov.b32 %r1301, %f963;
shfl.sync.bfly.b32 %r1303|%p310, %r1301, %r1271, %r1257, %r1259;
mov.b32 %f964, %r1303;
add.ftz.f32 %f78, %f963, %f964;
mov.b32 %r1304, %f955;
shfl.sync.bfly.b32 %r1305|%p311, %r1304, %r1258, %r1257, %r1259;
mov.b32 %f965, %r1305;
add.ftz.f32 %f966, %f955, %f965;
mov.b32 %r1306, %f966;
shfl.sync.bfly.b32 %r1307|%p312, %r1306, %r1262, %r1257, %r1259;
mov.b32 %f967, %r1307;
add.ftz.f32 %f968, %f966, %f967;
mov.b32 %r1308, %f968;
shfl.sync.bfly.b32 %r1309|%p313, %r1308, %r1265, %r1257, %r1259;
mov.b32 %f969, %r1309;
add.ftz.f32 %f970, %f968, %f969;
mov.b32 %r1310, %f970;
shfl.sync.bfly.b32 %r1311|%p314, %r1310, %r1268, %r1257, %r1259;
mov.b32 %f971, %r1311;
add.ftz.f32 %f972, %f970, %f971;
mov.b32 %r1312, %f972;
shfl.sync.bfly.b32 %r1313|%p315, %r1312, %r1271, %r1257, %r1259;
mov.b32 %f973, %r1313;
add.ftz.f32 %f79, %f972, %f973;
setp.ne.s32 %p316, %r2, 3;
or.b32 %r1317, %r4, 3;
setp.ge.u32 %p317, %r1317, %r52;
or.pred %p318, %p316, %p317;
@%p318 bra $L__BB15_78;

add.ftz.f32 %f974, %f5, %f78;
add.ftz.f32 %f975, %f4, %f79;
mul.ftz.f32 %f976, %f974, %f975;
st.global.f32 [%rd24+12], %f976;
bra.uni $L__BB15_78;

$L__BB15_62:
ld.shared.f32 %f681, [%r33];
add.ftz.f32 %f682, %f681, %f29;
ld.shared.f32 %f683, [%r35];
add.ftz.f32 %f684, %f683, %f33;
ld.shared.f32 %f685, [%r33+512];
add.ftz.f32 %f686, %f685, %f682;
ld.shared.f32 %f687, [%r35+512];
add.ftz.f32 %f688, %f687, %f684;
ld.shared.f32 %f689, [%r33+1024];
add.ftz.f32 %f690, %f689, %f686;
ld.shared.f32 %f691, [%r35+1024];
add.ftz.f32 %f692, %f691, %f688;
mov.b32 %r1072, %f690;
mov.u32 %r1073, 31;
mov.u32 %r1074, 16;
mov.u32 %r1075, -1;
shfl.sync.bfly.b32 %r1076|%p215, %r1072, %r1074, %r1073, %r1075;
mov.b32 %f693, %r1076;
add.ftz.f32 %f694, %f690, %f693;
mov.b32 %r1077, %f694;
mov.u32 %r1078, 8;
shfl.sync.bfly.b32 %r1079|%p216, %r1077, %r1078, %r1073, %r1075;
mov.b32 %f695, %r1079;
add.ftz.f32 %f696, %f694, %f695;
mov.b32 %r1080, %f696;
mov.u32 %r1081, 4;
shfl.sync.bfly.b32 %r1082|%p217, %r1080, %r1081, %r1073, %r1075;
mov.b32 %f697, %r1082;
add.ftz.f32 %f698, %f696, %f697;
mov.b32 %r1083, %f698;
mov.u32 %r1084, 2;
shfl.sync.bfly.b32 %r1085|%p218, %r1083, %r1084, %r1073, %r1075;
mov.b32 %f699, %r1085;
add.ftz.f32 %f700, %f698, %f699;
mov.b32 %r1086, %f700;
mov.u32 %r1087, 1;
shfl.sync.bfly.b32 %r1088|%p219, %r1086, %r1087, %r1073, %r1075;
mov.b32 %f701, %r1088;
add.ftz.f32 %f64, %f700, %f701;
mov.b32 %r1089, %f692;
shfl.sync.bfly.b32 %r1090|%p220, %r1089, %r1074, %r1073, %r1075;
mov.b32 %f702, %r1090;
add.ftz.f32 %f703, %f692, %f702;
mov.b32 %r1091, %f703;
shfl.sync.bfly.b32 %r1092|%p221, %r1091, %r1078, %r1073, %r1075;
mov.b32 %f704, %r1092;
add.ftz.f32 %f705, %f703, %f704;
mov.b32 %r1093, %f705;
shfl.sync.bfly.b32 %r1094|%p222, %r1093, %r1081, %r1073, %r1075;
mov.b32 %f706, %r1094;
add.ftz.f32 %f707, %f705, %f706;
mov.b32 %r1095, %f707;
shfl.sync.bfly.b32 %r1096|%p223, %r1095, %r1084, %r1073, %r1075;
mov.b32 %f708, %r1096;
add.ftz.f32 %f709, %f707, %f708;
mov.b32 %r1097, %f709;
shfl.sync.bfly.b32 %r1098|%p224, %r1097, %r1087, %r1073, %r1075;
mov.b32 %f710, %r1098;
add.ftz.f32 %f65, %f709, %f710;
setp.ne.s32 %p225, %r2, 0;
setp.ge.u32 %p226, %r4, %r52;
or.pred %p227, %p225, %p226;
@%p227 bra $L__BB15_64;

add.ftz.f32 %f711, %f5, %f64;
add.ftz.f32 %f712, %f4, %f65;
mul.ftz.f32 %f713, %f712, 0f3F000000;
mul.ftz.f32 %f714, %f712, 0f3F4C422A;
mul.ftz.f32 %f715, %f712, 0f3D372713;
fma.rn.ftz.f32 %f716, %f712, %f715, 0f3F800000;
mul.ftz.f32 %f717, %f714, %f716;
tanh.approx.f32 %f718, %f717;
add.ftz.f32 %f719, %f718, 0f3F800000;
mul.ftz.f32 %f720, %f713, %f719;
mul.ftz.f32 %f721, %f711, %f720;
st.global.f32 [%rd24], %f721;

$L__BB15_64:
ld.shared.f32 %f722, [%r33+128];
add.ftz.f32 %f723, %f722, %f28;
ld.shared.f32 %f724, [%r35+128];
add.ftz.f32 %f725, %f724, %f32;
ld.shared.f32 %f726, [%r33+640];
add.ftz.f32 %f727, %f726, %f723;
ld.shared.f32 %f728, [%r35+640];
add.ftz.f32 %f729, %f728, %f725;
ld.shared.f32 %f730, [%r33+1152];
add.ftz.f32 %f731, %f730, %f727;
ld.shared.f32 %f732, [%r35+1152];
add.ftz.f32 %f733, %f732, %f729;
mov.b32 %r1102, %f731;
shfl.sync.bfly.b32 %r1106|%p228, %r1102, %r1074, %r1073, %r1075;
mov.b32 %f734, %r1106;
add.ftz.f32 %f735, %f731, %f734;
mov.b32 %r1107, %f735;
shfl.sync.bfly.b32 %r1109|%p229, %r1107, %r1078, %r1073, %r1075;
mov.b32 %f736, %r1109;
add.ftz.f32 %f737, %f735, %f736;
mov.b32 %r1110, %f737;
shfl.sync.bfly.b32 %r1112|%p230, %r1110, %r1081, %r1073, %r1075;
mov.b32 %f738, %r1112;
add.ftz.f32 %f739, %f737, %f738;
mov.b32 %r1113, %f739;
shfl.sync.bfly.b32 %r1115|%p231, %r1113, %r1084, %r1073, %r1075;
mov.b32 %f740, %r1115;
add.ftz.f32 %f741, %f739, %f740;
mov.b32 %r1116, %f741;
shfl.sync.bfly.b32 %r1118|%p232, %r1116, %r1087, %r1073, %r1075;
mov.b32 %f742, %r1118;
add.ftz.f32 %f66, %f741, %f742;
mov.b32 %r1119, %f733;
shfl.sync.bfly.b32 %r1120|%p233, %r1119, %r1074, %r1073, %r1075;
mov.b32 %f743, %r1120;
add.ftz.f32 %f744, %f733, %f743;
mov.b32 %r1121, %f744;
shfl.sync.bfly.b32 %r1122|%p234, %r1121, %r1078, %r1073, %r1075;
mov.b32 %f745, %r1122;
add.ftz.f32 %f746, %f744, %f745;
mov.b32 %r1123, %f746;
shfl.sync.bfly.b32 %r1124|%p235, %r1123, %r1081, %r1073, %r1075;
mov.b32 %f747, %r1124;
add.ftz.f32 %f748, %f746, %f747;
mov.b32 %r1125, %f748;
shfl.sync.bfly.b32 %r1126|%p236, %r1125, %r1084, %r1073, %r1075;
mov.b32 %f749, %r1126;
add.ftz.f32 %f750, %f748, %f749;
mov.b32 %r1127, %f750;
shfl.sync.bfly.b32 %r1128|%p237, %r1127, %r1087, %r1073, %r1075;
mov.b32 %f751, %r1128;
add.ftz.f32 %f67, %f750, %f751;
setp.ne.s32 %p238, %r2, 1;
or.b32 %r1132, %r4, 1;
setp.ge.u32 %p239, %r1132, %r52;
or.pred %p240, %p238, %p239;
@%p240 bra $L__BB15_66;

add.ftz.f32 %f752, %f5, %f66;
add.ftz.f32 %f753, %f4, %f67;
mul.ftz.f32 %f754, %f753, 0f3F000000;
mul.ftz.f32 %f755, %f753, 0f3F4C422A;
mul.ftz.f32 %f756, %f753, 0f3D372713;
fma.rn.ftz.f32 %f757, %f753, %f756, 0f3F800000;
mul.ftz.f32 %f758, %f755, %f757;
tanh.approx.f32 %f759, %f758;
add.ftz.f32 %f760, %f759, 0f3F800000;
mul.ftz.f32 %f761, %f754, %f760;
mul.ftz.f32 %f762, %f752, %f761;
st.global.f32 [%rd24+4], %f762;

$L__BB15_66:
ld.shared.f32 %f763, [%r33+256];
add.ftz.f32 %f764, %f763, %f27;
ld.shared.f32 %f765, [%r35+256];
add.ftz.f32 %f766, %f765, %f31;
ld.shared.f32 %f767, [%r33+768];
add.ftz.f32 %f768, %f767, %f764;
ld.shared.f32 %f769, [%r35+768];
add.ftz.f32 %f770, %f769, %f766;
ld.shared.f32 %f771, [%r33+1280];
add.ftz.f32 %f772, %f771, %f768;
ld.shared.f32 %f773, [%r35+1280];
add.ftz.f32 %f774, %f773, %f770;
mov.b32 %r1133, %f772;
mov.u32 %r1134, 31;
mov.u32 %r1135, 16;
mov.u32 %r1136, -1;
shfl.sync.bfly.b32 %r1137|%p241, %r1133, %r1135, %r1134, %r1136;
mov.b32 %f775, %r1137;
add.ftz.f32 %f776, %f772, %f775;
mov.b32 %r1138, %f776;
mov.u32 %r1139, 8;
shfl.sync.bfly.b32 %r1140|%p242, %r1138, %r1139, %r1134, %r1136;
mov.b32 %f777, %r1140;
add.ftz.f32 %f778, %f776, %f777;
mov.b32 %r1141, %f778;
mov.u32 %r1142, 4;
shfl.sync.bfly.b32 %r1143|%p243, %r1141, %r1142, %r1134, %r1136;
mov.b32 %f779, %r1143;
add.ftz.f32 %f780, %f778, %f779;
mov.b32 %r1144, %f780;
mov.u32 %r1145, 2;
shfl.sync.bfly.b32 %r1146|%p244, %r1144, %r1145, %r1134, %r1136;
mov.b32 %f781, %r1146;
add.ftz.f32 %f782, %f780, %f781;
mov.b32 %r1147, %f782;
mov.u32 %r1148, 1;
shfl.sync.bfly.b32 %r1149|%p245, %r1147, %r1148, %r1134, %r1136;
mov.b32 %f783, %r1149;
add.ftz.f32 %f68, %f782, %f783;
mov.b32 %r1150, %f774;
shfl.sync.bfly.b32 %r1151|%p246, %r1150, %r1135, %r1134, %r1136;
mov.b32 %f784, %r1151;
add.ftz.f32 %f785, %f774, %f784;
mov.b32 %r1152, %f785;
shfl.sync.bfly.b32 %r1153|%p247, %r1152, %r1139, %r1134, %r1136;
mov.b32 %f786, %r1153;
add.ftz.f32 %f787, %f785, %f786;
mov.b32 %r1154, %f787;
shfl.sync.bfly.b32 %r1155|%p248, %r1154, %r1142, %r1134, %r1136;
mov.b32 %f788, %r1155;
add.ftz.f32 %f789, %f787, %f788;
mov.b32 %r1156, %f789;
shfl.sync.bfly.b32 %r1157|%p249, %r1156, %r1145, %r1134, %r1136;
mov.b32 %f790, %r1157;
add.ftz.f32 %f791, %f789, %f790;
mov.b32 %r1158, %f791;
shfl.sync.bfly.b32 %r1159|%p250, %r1158, %r1148, %r1134, %r1136;
mov.b32 %f792, %r1159;
add.ftz.f32 %f69, %f791, %f792;
setp.ne.s32 %p251, %r2, 2;
or.b32 %r1163, %r4, 2;
setp.ge.u32 %p252, %r1163, %r52;
or.pred %p253, %p251, %p252;
@%p253 bra $L__BB15_68;

add.ftz.f32 %f793, %f5, %f68;
add.ftz.f32 %f794, %f4, %f69;
mul.ftz.f32 %f795, %f794, 0f3F000000;
mul.ftz.f32 %f796, %f794, 0f3F4C422A;
mul.ftz.f32 %f797, %f794, 0f3D372713;
fma.rn.ftz.f32 %f798, %f794, %f797, 0f3F800000;
mul.ftz.f32 %f799, %f796, %f798;
tanh.approx.f32 %f800, %f799;
add.ftz.f32 %f801, %f800, 0f3F800000;
mul.ftz.f32 %f802, %f795, %f801;
mul.ftz.f32 %f803, %f793, %f802;
st.global.f32 [%rd24+8], %f803;

$L__BB15_68:
ld.shared.f32 %f804, [%r33+384];
add.ftz.f32 %f805, %f804, %f26;
ld.shared.f32 %f806, [%r35+384];
add.ftz.f32 %f807, %f806, %f30;
ld.shared.f32 %f808, [%r33+896];
add.ftz.f32 %f809, %f808, %f805;
ld.shared.f32 %f810, [%r35+896];
add.ftz.f32 %f811, %f810, %f807;
ld.shared.f32 %f812, [%r33+1408];
add.ftz.f32 %f813, %f812, %f809;
ld.shared.f32 %f814, [%r35+1408];
add.ftz.f32 %f815, %f814, %f811;
mov.b32 %r1164, %f813;
shfl.sync.bfly.b32 %r1168|%p254, %r1164, %r1135, %r1134, %r1136;
mov.b32 %f816, %r1168;
add.ftz.f32 %f817, %f813, %f816;
mov.b32 %r1169, %f817;
shfl.sync.bfly.b32 %r1171|%p255, %r1169, %r1139, %r1134, %r1136;
mov.b32 %f818, %r1171;
add.ftz.f32 %f819, %f817, %f818;
mov.b32 %r1172, %f819;
shfl.sync.bfly.b32 %r1174|%p256, %r1172, %r1142, %r1134, %r1136;
mov.b32 %f820, %r1174;
add.ftz.f32 %f821, %f819, %f820;
mov.b32 %r1175, %f821;
shfl.sync.bfly.b32 %r1177|%p257, %r1175, %r1145, %r1134, %r1136;
mov.b32 %f822, %r1177;
add.ftz.f32 %f823, %f821, %f822;
mov.b32 %r1178, %f823;
shfl.sync.bfly.b32 %r1180|%p258, %r1178, %r1148, %r1134, %r1136;
mov.b32 %f824, %r1180;
add.ftz.f32 %f70, %f823, %f824;
mov.b32 %r1181, %f815;
shfl.sync.bfly.b32 %r1182|%p259, %r1181, %r1135, %r1134, %r1136;
mov.b32 %f825, %r1182;
add.ftz.f32 %f826, %f815, %f825;
mov.b32 %r1183, %f826;
shfl.sync.bfly.b32 %r1184|%p260, %r1183, %r1139, %r1134, %r1136;
mov.b32 %f827, %r1184;
add.ftz.f32 %f828, %f826, %f827;
mov.b32 %r1185, %f828;
shfl.sync.bfly.b32 %r1186|%p261, %r1185, %r1142, %r1134, %r1136;
mov.b32 %f829, %r1186;
add.ftz.f32 %f830, %f828, %f829;
mov.b32 %r1187, %f830;
shfl.sync.bfly.b32 %r1188|%p262, %r1187, %r1145, %r1134, %r1136;
mov.b32 %f831, %r1188;
add.ftz.f32 %f832, %f830, %f831;
mov.b32 %r1189, %f832;
shfl.sync.bfly.b32 %r1190|%p263, %r1189, %r1148, %r1134, %r1136;
mov.b32 %f833, %r1190;
add.ftz.f32 %f71, %f832, %f833;
setp.ne.s32 %p264, %r2, 3;
or.b32 %r1194, %r4, 3;
setp.ge.u32 %p265, %r1194, %r52;
or.pred %p266, %p264, %p265;
@%p266 bra $L__BB15_78;

add.ftz.f32 %f834, %f5, %f70;
add.ftz.f32 %f835, %f4, %f71;
mul.ftz.f32 %f836, %f835, 0f3F000000;
mul.ftz.f32 %f837, %f835, 0f3F4C422A;
mul.ftz.f32 %f838, %f835, 0f3D372713;
fma.rn.ftz.f32 %f839, %f835, %f838, 0f3F800000;
mul.ftz.f32 %f840, %f837, %f839;
tanh.approx.f32 %f841, %f840;
add.ftz.f32 %f842, %f841, 0f3F800000;
mul.ftz.f32 %f843, %f836, %f842;
mul.ftz.f32 %f844, %f834, %f843;
st.global.f32 [%rd24+12], %f844;
bra.uni $L__BB15_78;

$L__BB15_54:
ld.shared.f32 %f357, [%r33];
add.ftz.f32 %f358, %f357, %f29;
ld.shared.f32 %f359, [%r35];
add.ftz.f32 %f360, %f359, %f33;
ld.shared.f32 %f361, [%r33+512];
add.ftz.f32 %f362, %f361, %f358;
ld.shared.f32 %f363, [%r35+512];
add.ftz.f32 %f364, %f363, %f360;
ld.shared.f32 %f365, [%r33+1024];
add.ftz.f32 %f366, %f365, %f362;
ld.shared.f32 %f367, [%r35+1024];
add.ftz.f32 %f368, %f367, %f364;
mov.b32 %r826, %f366;
mov.u32 %r827, 31;
mov.u32 %r828, 16;
mov.u32 %r829, -1;
shfl.sync.bfly.b32 %r830|%p111, %r826, %r828, %r827, %r829;
mov.b32 %f369, %r830;
add.ftz.f32 %f370, %f366, %f369;
mov.b32 %r831, %f370;
mov.u32 %r832, 8;
shfl.sync.bfly.b32 %r833|%p112, %r831, %r832, %r827, %r829;
mov.b32 %f371, %r833;
add.ftz.f32 %f372, %f370, %f371;
mov.b32 %r834, %f372;
mov.u32 %r835, 4;
shfl.sync.bfly.b32 %r836|%p113, %r834, %r835, %r827, %r829;
mov.b32 %f373, %r836;
add.ftz.f32 %f374, %f372, %f373;
mov.b32 %r837, %f374;
mov.u32 %r838, 2;
shfl.sync.bfly.b32 %r839|%p114, %r837, %r838, %r827, %r829;
mov.b32 %f375, %r839;
add.ftz.f32 %f376, %f374, %f375;
mov.b32 %r840, %f376;
mov.u32 %r841, 1;
shfl.sync.bfly.b32 %r842|%p115, %r840, %r841, %r827, %r829;
mov.b32 %f377, %r842;
add.ftz.f32 %f48, %f376, %f377;
mov.b32 %r843, %f368;
shfl.sync.bfly.b32 %r844|%p116, %r843, %r828, %r827, %r829;
mov.b32 %f378, %r844;
add.ftz.f32 %f379, %f368, %f378;
mov.b32 %r845, %f379;
shfl.sync.bfly.b32 %r846|%p117, %r845, %r832, %r827, %r829;
mov.b32 %f380, %r846;
add.ftz.f32 %f381, %f379, %f380;
mov.b32 %r847, %f381;
shfl.sync.bfly.b32 %r848|%p118, %r847, %r835, %r827, %r829;
mov.b32 %f382, %r848;
add.ftz.f32 %f383, %f381, %f382;
mov.b32 %r849, %f383;
shfl.sync.bfly.b32 %r850|%p119, %r849, %r838, %r827, %r829;
mov.b32 %f384, %r850;
add.ftz.f32 %f385, %f383, %f384;
mov.b32 %r851, %f385;
shfl.sync.bfly.b32 %r852|%p120, %r851, %r841, %r827, %r829;
mov.b32 %f386, %r852;
add.ftz.f32 %f49, %f385, %f386;
setp.ne.s32 %p121, %r2, 0;
setp.ge.u32 %p122, %r4, %r52;
or.pred %p123, %p121, %p122;
@%p123 bra $L__BB15_56;

add.ftz.f32 %f387, %f5, %f48;
add.ftz.f32 %f388, %f4, %f49;
mov.f32 %f389, 0f40E00000;
min.ftz.f32 %f390, %f388, %f389;
min.ftz.f32 %f391, %f387, %f389;
mov.f32 %f392, 0fC0E00000;
max.ftz.f32 %f393, %f391, %f392;
mul.ftz.f32 %f394, %f390, 0f3FD9DB23;
mul.ftz.f32 %f395, %f394, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f396, %f395;
add.ftz.f32 %f397, %f396, 0f3F800000;
div.approx.ftz.f32 %f398, %f390, %f397;
add.ftz.f32 %f399, %f393, 0f3F800000;
mul.ftz.f32 %f400, %f399, %f398;
st.global.f32 [%rd24], %f400;

$L__BB15_56:
ld.shared.f32 %f401, [%r33+128];
add.ftz.f32 %f402, %f401, %f28;
ld.shared.f32 %f403, [%r35+128];
add.ftz.f32 %f404, %f403, %f32;
ld.shared.f32 %f405, [%r33+640];
add.ftz.f32 %f406, %f405, %f402;
ld.shared.f32 %f407, [%r35+640];
add.ftz.f32 %f408, %f407, %f404;
ld.shared.f32 %f409, [%r33+1152];
add.ftz.f32 %f410, %f409, %f406;
ld.shared.f32 %f411, [%r35+1152];
add.ftz.f32 %f412, %f411, %f408;
mov.b32 %r856, %f410;
shfl.sync.bfly.b32 %r860|%p124, %r856, %r828, %r827, %r829;
mov.b32 %f413, %r860;
add.ftz.f32 %f414, %f410, %f413;
mov.b32 %r861, %f414;
shfl.sync.bfly.b32 %r863|%p125, %r861, %r832, %r827, %r829;
mov.b32 %f415, %r863;
add.ftz.f32 %f416, %f414, %f415;
mov.b32 %r864, %f416;
shfl.sync.bfly.b32 %r866|%p126, %r864, %r835, %r827, %r829;
mov.b32 %f417, %r866;
add.ftz.f32 %f418, %f416, %f417;
mov.b32 %r867, %f418;
shfl.sync.bfly.b32 %r869|%p127, %r867, %r838, %r827, %r829;
mov.b32 %f419, %r869;
add.ftz.f32 %f420, %f418, %f419;
mov.b32 %r870, %f420;
shfl.sync.bfly.b32 %r872|%p128, %r870, %r841, %r827, %r829;
mov.b32 %f421, %r872;
add.ftz.f32 %f50, %f420, %f421;
mov.b32 %r873, %f412;
shfl.sync.bfly.b32 %r874|%p129, %r873, %r828, %r827, %r829;
mov.b32 %f422, %r874;
add.ftz.f32 %f423, %f412, %f422;
mov.b32 %r875, %f423;
shfl.sync.bfly.b32 %r876|%p130, %r875, %r832, %r827, %r829;
mov.b32 %f424, %r876;
add.ftz.f32 %f425, %f423, %f424;
mov.b32 %r877, %f425;
shfl.sync.bfly.b32 %r878|%p131, %r877, %r835, %r827, %r829;
mov.b32 %f426, %r878;
add.ftz.f32 %f427, %f425, %f426;
mov.b32 %r879, %f427;
shfl.sync.bfly.b32 %r880|%p132, %r879, %r838, %r827, %r829;
mov.b32 %f428, %r880;
add.ftz.f32 %f429, %f427, %f428;
mov.b32 %r881, %f429;
shfl.sync.bfly.b32 %r882|%p133, %r881, %r841, %r827, %r829;
mov.b32 %f430, %r882;
add.ftz.f32 %f51, %f429, %f430;
setp.ne.s32 %p134, %r2, 1;
or.b32 %r886, %r4, 1;
setp.ge.u32 %p135, %r886, %r52;
or.pred %p136, %p134, %p135;
@%p136 bra $L__BB15_58;

add.ftz.f32 %f431, %f5, %f50;
add.ftz.f32 %f432, %f4, %f51;
mov.f32 %f433, 0f40E00000;
min.ftz.f32 %f434, %f432, %f433;
min.ftz.f32 %f435, %f431, %f433;
mov.f32 %f436, 0fC0E00000;
max.ftz.f32 %f437, %f435, %f436;
mul.ftz.f32 %f438, %f434, 0f3FD9DB23;
mul.ftz.f32 %f439, %f438, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f440, %f439;
add.ftz.f32 %f441, %f440, 0f3F800000;
div.approx.ftz.f32 %f442, %f434, %f441;
add.ftz.f32 %f443, %f437, 0f3F800000;
mul.ftz.f32 %f444, %f443, %f442;
st.global.f32 [%rd24+4], %f444;

$L__BB15_58:
ld.shared.f32 %f445, [%r33+256];
add.ftz.f32 %f446, %f445, %f27;
ld.shared.f32 %f447, [%r35+256];
add.ftz.f32 %f448, %f447, %f31;
ld.shared.f32 %f449, [%r33+768];
add.ftz.f32 %f450, %f449, %f446;
ld.shared.f32 %f451, [%r35+768];
add.ftz.f32 %f452, %f451, %f448;
ld.shared.f32 %f453, [%r33+1280];
add.ftz.f32 %f454, %f453, %f450;
ld.shared.f32 %f455, [%r35+1280];
add.ftz.f32 %f456, %f455, %f452;
mov.b32 %r887, %f454;
mov.u32 %r888, 31;
mov.u32 %r889, 16;
mov.u32 %r890, -1;
shfl.sync.bfly.b32 %r891|%p137, %r887, %r889, %r888, %r890;
mov.b32 %f457, %r891;
add.ftz.f32 %f458, %f454, %f457;
mov.b32 %r892, %f458;
mov.u32 %r893, 8;
shfl.sync.bfly.b32 %r894|%p138, %r892, %r893, %r888, %r890;
mov.b32 %f459, %r894;
add.ftz.f32 %f460, %f458, %f459;
mov.b32 %r895, %f460;
mov.u32 %r896, 4;
shfl.sync.bfly.b32 %r897|%p139, %r895, %r896, %r888, %r890;
mov.b32 %f461, %r897;
add.ftz.f32 %f462, %f460, %f461;
mov.b32 %r898, %f462;
mov.u32 %r899, 2;
shfl.sync.bfly.b32 %r900|%p140, %r898, %r899, %r888, %r890;
mov.b32 %f463, %r900;
add.ftz.f32 %f464, %f462, %f463;
mov.b32 %r901, %f464;
mov.u32 %r902, 1;
shfl.sync.bfly.b32 %r903|%p141, %r901, %r902, %r888, %r890;
mov.b32 %f465, %r903;
add.ftz.f32 %f52, %f464, %f465;
mov.b32 %r904, %f456;
shfl.sync.bfly.b32 %r905|%p142, %r904, %r889, %r888, %r890;
mov.b32 %f466, %r905;
add.ftz.f32 %f467, %f456, %f466;
mov.b32 %r906, %f467;
shfl.sync.bfly.b32 %r907|%p143, %r906, %r893, %r888, %r890;
mov.b32 %f468, %r907;
add.ftz.f32 %f469, %f467, %f468;
mov.b32 %r908, %f469;
shfl.sync.bfly.b32 %r909|%p144, %r908, %r896, %r888, %r890;
mov.b32 %f470, %r909;
add.ftz.f32 %f471, %f469, %f470;
mov.b32 %r910, %f471;
shfl.sync.bfly.b32 %r911|%p145, %r910, %r899, %r888, %r890;
mov.b32 %f472, %r911;
add.ftz.f32 %f473, %f471, %f472;
mov.b32 %r912, %f473;
shfl.sync.bfly.b32 %r913|%p146, %r912, %r902, %r888, %r890;
mov.b32 %f474, %r913;
add.ftz.f32 %f53, %f473, %f474;
setp.ne.s32 %p147, %r2, 2;
or.b32 %r917, %r4, 2;
setp.ge.u32 %p148, %r917, %r52;
or.pred %p149, %p147, %p148;
@%p149 bra $L__BB15_60;

add.ftz.f32 %f475, %f5, %f52;
add.ftz.f32 %f476, %f4, %f53;
mov.f32 %f477, 0f40E00000;
min.ftz.f32 %f478, %f476, %f477;
min.ftz.f32 %f479, %f475, %f477;
mov.f32 %f480, 0fC0E00000;
max.ftz.f32 %f481, %f479, %f480;
mul.ftz.f32 %f482, %f478, 0f3FD9DB23;
mul.ftz.f32 %f483, %f482, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f484, %f483;
add.ftz.f32 %f485, %f484, 0f3F800000;
div.approx.ftz.f32 %f486, %f478, %f485;
add.ftz.f32 %f487, %f481, 0f3F800000;
mul.ftz.f32 %f488, %f487, %f486;
st.global.f32 [%rd24+8], %f488;

$L__BB15_60:
ld.shared.f32 %f489, [%r33+384];
add.ftz.f32 %f490, %f489, %f26;
ld.shared.f32 %f491, [%r35+384];
add.ftz.f32 %f492, %f491, %f30;
ld.shared.f32 %f493, [%r33+896];
add.ftz.f32 %f494, %f493, %f490;
ld.shared.f32 %f495, [%r35+896];
add.ftz.f32 %f496, %f495, %f492;
ld.shared.f32 %f497, [%r33+1408];
add.ftz.f32 %f498, %f497, %f494;
ld.shared.f32 %f499, [%r35+1408];
add.ftz.f32 %f500, %f499, %f496;
mov.b32 %r918, %f498;
shfl.sync.bfly.b32 %r922|%p150, %r918, %r889, %r888, %r890;
mov.b32 %f501, %r922;
add.ftz.f32 %f502, %f498, %f501;
mov.b32 %r923, %f502;
shfl.sync.bfly.b32 %r925|%p151, %r923, %r893, %r888, %r890;
mov.b32 %f503, %r925;
add.ftz.f32 %f504, %f502, %f503;
mov.b32 %r926, %f504;
shfl.sync.bfly.b32 %r928|%p152, %r926, %r896, %r888, %r890;
mov.b32 %f505, %r928;
add.ftz.f32 %f506, %f504, %f505;
mov.b32 %r929, %f506;
shfl.sync.bfly.b32 %r931|%p153, %r929, %r899, %r888, %r890;
mov.b32 %f507, %r931;
add.ftz.f32 %f508, %f506, %f507;
mov.b32 %r932, %f508;
shfl.sync.bfly.b32 %r934|%p154, %r932, %r902, %r888, %r890;
mov.b32 %f509, %r934;
add.ftz.f32 %f54, %f508, %f509;
mov.b32 %r935, %f500;
shfl.sync.bfly.b32 %r936|%p155, %r935, %r889, %r888, %r890;
mov.b32 %f510, %r936;
add.ftz.f32 %f511, %f500, %f510;
mov.b32 %r937, %f511;
shfl.sync.bfly.b32 %r938|%p156, %r937, %r893, %r888, %r890;
mov.b32 %f512, %r938;
add.ftz.f32 %f513, %f511, %f512;
mov.b32 %r939, %f513;
shfl.sync.bfly.b32 %r940|%p157, %r939, %r896, %r888, %r890;
mov.b32 %f514, %r940;
add.ftz.f32 %f515, %f513, %f514;
mov.b32 %r941, %f515;
shfl.sync.bfly.b32 %r942|%p158, %r941, %r899, %r888, %r890;
mov.b32 %f516, %r942;
add.ftz.f32 %f517, %f515, %f516;
mov.b32 %r943, %f517;
shfl.sync.bfly.b32 %r944|%p159, %r943, %r902, %r888, %r890;
mov.b32 %f518, %r944;
add.ftz.f32 %f55, %f517, %f518;
setp.ne.s32 %p160, %r2, 3;
or.b32 %r948, %r4, 3;
setp.ge.u32 %p161, %r948, %r52;
or.pred %p162, %p160, %p161;
@%p162 bra $L__BB15_78;

add.ftz.f32 %f519, %f5, %f54;
add.ftz.f32 %f520, %f4, %f55;
mov.f32 %f521, 0f40E00000;
min.ftz.f32 %f522, %f520, %f521;
min.ftz.f32 %f523, %f519, %f521;
mov.f32 %f524, 0fC0E00000;
max.ftz.f32 %f525, %f523, %f524;
mul.ftz.f32 %f526, %f522, 0f3FD9DB23;
mul.ftz.f32 %f527, %f526, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f528, %f527;
add.ftz.f32 %f529, %f528, 0f3F800000;
div.approx.ftz.f32 %f530, %f522, %f529;
add.ftz.f32 %f531, %f525, 0f3F800000;
mul.ftz.f32 %f532, %f531, %f530;
st.global.f32 [%rd24+12], %f532;
bra.uni $L__BB15_78;

}
.entry _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<38>;
.reg .b16 %rs<21>;
.reg .f32 %f<109>;
.reg .b32 %r<380>;
.reg .b64 %rd<39>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

ld.param.u64 %rd10, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd11, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd12, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd13, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
mov.u32 %r1, %tid.y;
shl.b32 %r38, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r38, %r2;
mov.u32 %r39, %ctaid.x;
shl.b32 %r4, %r39, 2;
mov.u32 %r377, %ctaid.y;
setp.eq.s64 %p1, %rd12, 0;
@%p1 bra $L__BB16_2;

cvta.to.global.u64 %rd14, %rd12;
mul.wide.u32 %rd15, %r377, 4;
add.s64 %rd16, %rd14, %rd15;
ld.global.nc.u32 %r376, [%rd16];
mul.hi.u32 %r40, %r377, %r21;
add.s32 %r41, %r40, %r377;
shr.u32 %r42, %r41, %r22;
mul.lo.s32 %r43, %r42, %r23;
sub.s32 %r377, %r377, %r43;
bra.uni $L__BB16_3;

$L__BB16_2:
mul.hi.u32 %r44, %r377, %r26;
add.s32 %r45, %r44, %r377;
shr.u32 %r376, %r45, %r27;

$L__BB16_3:
shr.u32 %r46, %r3, 31;
add.s32 %r47, %r3, %r46;
shr.s32 %r379, %r47, 1;
shr.u32 %r48, %r20, 6;
setp.ge.s32 %p2, %r379, %r48;
mov.f32 %f105, 0f00000000;
mov.f32 %f106, %f105;
mov.f32 %f107, %f105;
mov.f32 %f108, %f105;
@%p2 bra $L__BB16_6;

mov.u32 %r49, %ctaid.z;
mul.lo.s32 %r50, %r377, %r30;
cvt.u64.u32 %rd17, %r50;
and.b32 %r53, %r47, -2;
sub.s32 %r54, %r3, %r53;
cvt.s64.s32 %rd18, %r54;
mul.lo.s32 %r55, %r49, %r36;
cvt.u64.u32 %rd19, %r55;
add.s64 %rd20, %rd19, %rd18;
add.s64 %rd1, %rd20, %rd17;
shl.b32 %r56, %r54, 2;
shl.b32 %r378, %r379, 1;
mad.lo.s32 %r57, %r4, %r24, %r379;
mul.hi.u32 %r58, %r49, %r32;
add.s32 %r59, %r58, %r49;
shr.u32 %r60, %r59, %r33;
mad.lo.s32 %r61, %r60, %r35, %r57;
mad.lo.s32 %r62, %r376, %r29, %r61;
mul.wide.s32 %rd2, %r62, 18;
mul.wide.s32 %rd21, %r56, 2;
add.s64 %rd22, %rd2, %rd21;
add.s64 %rd3, %rd22, 8;
mul.wide.s32 %rd4, %r24, 18;
cvta.to.global.u64 %rd38, %rd10;
cvta.to.global.u64 %rd6, %rd11;
mov.f32 %f105, 0f00000000;

$L__BB16_5:
cvt.u64.u32 %rd23, %r378;
add.s64 %rd24, %rd1, %rd23;
mul.lo.s64 %rd25, %rd24, 36;
add.s64 %rd26, %rd38, %rd2;
ld.global.nc.u16 %rs1, [%rd26];

{ cvt.f32.f16 %f25, %rs1;}


add.s64 %rd27, %rd38, %rd3;
add.s64 %rd28, %rd27, -6;
ld.global.nc.u16 %rs5, [%rd27+-6];
cvt.s32.s16 %r195, %rs5;
add.s64 %rd29, %rd6, %rd25;
ld.global.nc.u32 %r64, [%rd29+4];
ld.global.nc.u32 %r68, [%rd29+8];
and.b32 %r196, %r195, 30583;
mov.u32 %r197, 33620223;
prmt.b32 %r198, %r197, %r197, %r196;
shr.u32 %r199, %r195, 2;
and.b32 %r200, %r199, 30583;
prmt.b32 %r201, %r197, %r197, %r200;
mov.u32 %r202, 20800;
prmt.b32 %r65, %r198, %r201, %r202;
mov.u32 %r203, 29538;
prmt.b32 %r69, %r198, %r201, %r203;
mov.u32 %r165, 0;

dp4a.s32.s32 %r63, %r64, %r65, %r165;


dp4a.s32.s32 %r67, %r68, %r69, %r63;

ld.global.nc.u16 %rs6, [%rd27+-4];
cvt.s32.s16 %r204, %rs6;
ld.global.nc.u32 %r72, [%rd29+12];
ld.global.nc.u32 %r76, [%rd29+16];
and.b32 %r205, %r204, 30583;
prmt.b32 %r206, %r197, %r197, %r205;
shr.u32 %r207, %r204, 2;
and.b32 %r208, %r207, 30583;
prmt.b32 %r209, %r197, %r197, %r208;
prmt.b32 %r73, %r206, %r209, %r202;
prmt.b32 %r77, %r206, %r209, %r203;

dp4a.s32.s32 %r71, %r72, %r73, %r67;


dp4a.s32.s32 %r75, %r76, %r77, %r71;

ld.global.nc.u16 %rs7, [%rd27+-2];
cvt.s32.s16 %r210, %rs7;
ld.global.nc.u32 %r80, [%rd29+20];
ld.global.nc.u32 %r84, [%rd29+24];
and.b32 %r211, %r210, 30583;
prmt.b32 %r212, %r197, %r197, %r211;
shr.u32 %r213, %r210, 2;
and.b32 %r214, %r213, 30583;
prmt.b32 %r215, %r197, %r197, %r214;
prmt.b32 %r81, %r212, %r215, %r202;
prmt.b32 %r85, %r212, %r215, %r203;

dp4a.s32.s32 %r79, %r80, %r81, %r75;


dp4a.s32.s32 %r83, %r84, %r85, %r79;

ld.global.nc.u16 %rs8, [%rd27];
cvt.s32.s16 %r216, %rs8;
ld.global.nc.u32 %r88, [%rd29+28];
ld.global.nc.u32 %r92, [%rd29+32];
and.b32 %r217, %r216, 30583;
prmt.b32 %r218, %r197, %r197, %r217;
shr.u32 %r219, %r216, 2;
and.b32 %r220, %r219, 30583;
prmt.b32 %r221, %r197, %r197, %r220;
prmt.b32 %r89, %r218, %r221, %r202;
prmt.b32 %r93, %r218, %r221, %r203;

dp4a.s32.s32 %r87, %r88, %r89, %r83;


dp4a.s32.s32 %r91, %r92, %r93, %r87;

ld.global.nc.u32 %r95, [%rd29];

{.reg .f16 low,high;
mov.b32 {low,high},%r95;
cvt.f32.f16 %f26, low;}


mul.ftz.f32 %f33, %f25, %f26;
cvt.rn.f32.s32 %f34, %r91;
fma.rn.ftz.f32 %f108, %f33, %f34, %f108;
add.s64 %rd30, %rd26, %rd4;
ld.global.nc.u16 %rs2, [%rd30];

{ cvt.f32.f16 %f27, %rs2;}


add.s64 %rd31, %rd28, %rd4;
ld.global.nc.u16 %rs9, [%rd31];
cvt.s32.s16 %r222, %rs9;
and.b32 %r223, %r222, 30583;
prmt.b32 %r224, %r197, %r197, %r223;
shr.u32 %r225, %r222, 2;
and.b32 %r226, %r225, 30583;
prmt.b32 %r227, %r197, %r197, %r226;
prmt.b32 %r98, %r224, %r227, %r202;
prmt.b32 %r102, %r224, %r227, %r203;

dp4a.s32.s32 %r96, %r64, %r98, %r165;


dp4a.s32.s32 %r100, %r68, %r102, %r96;

ld.global.nc.u16 %rs10, [%rd31+2];
cvt.s32.s16 %r228, %rs10;
and.b32 %r229, %r228, 30583;
prmt.b32 %r230, %r197, %r197, %r229;
shr.u32 %r231, %r228, 2;
and.b32 %r232, %r231, 30583;
prmt.b32 %r233, %r197, %r197, %r232;
prmt.b32 %r106, %r230, %r233, %r202;
prmt.b32 %r110, %r230, %r233, %r203;

dp4a.s32.s32 %r104, %r72, %r106, %r100;


dp4a.s32.s32 %r108, %r76, %r110, %r104;

ld.global.nc.u16 %rs11, [%rd31+4];
cvt.s32.s16 %r234, %rs11;
and.b32 %r235, %r234, 30583;
prmt.b32 %r236, %r197, %r197, %r235;
shr.u32 %r237, %r234, 2;
and.b32 %r238, %r237, 30583;
prmt.b32 %r239, %r197, %r197, %r238;
prmt.b32 %r114, %r236, %r239, %r202;
prmt.b32 %r118, %r236, %r239, %r203;

dp4a.s32.s32 %r112, %r80, %r114, %r108;


dp4a.s32.s32 %r116, %r84, %r118, %r112;

ld.global.nc.u16 %rs12, [%rd31+6];
cvt.s32.s16 %r240, %rs12;
and.b32 %r241, %r240, 30583;
prmt.b32 %r242, %r197, %r197, %r241;
shr.u32 %r243, %r240, 2;
and.b32 %r244, %r243, 30583;
prmt.b32 %r245, %r197, %r197, %r244;
prmt.b32 %r122, %r242, %r245, %r202;
prmt.b32 %r126, %r242, %r245, %r203;

dp4a.s32.s32 %r120, %r88, %r122, %r116;


dp4a.s32.s32 %r124, %r92, %r126, %r120;


{.reg .f16 low,high;
mov.b32 {low,high},%r95;
cvt.f32.f16 %f28, low;}


mul.ftz.f32 %f35, %f27, %f28;
cvt.rn.f32.s32 %f36, %r124;
fma.rn.ftz.f32 %f107, %f35, %f36, %f107;
add.s64 %rd32, %rd30, %rd4;
ld.global.nc.u16 %rs3, [%rd32];

{ cvt.f32.f16 %f29, %rs3;}


add.s64 %rd33, %rd31, %rd4;
ld.global.nc.u16 %rs13, [%rd33];
cvt.s32.s16 %r246, %rs13;
and.b32 %r247, %r246, 30583;
prmt.b32 %r248, %r197, %r197, %r247;
shr.u32 %r249, %r246, 2;
and.b32 %r250, %r249, 30583;
prmt.b32 %r251, %r197, %r197, %r250;
prmt.b32 %r131, %r248, %r251, %r202;
prmt.b32 %r135, %r248, %r251, %r203;

dp4a.s32.s32 %r129, %r64, %r131, %r165;


dp4a.s32.s32 %r133, %r68, %r135, %r129;

ld.global.nc.u16 %rs14, [%rd33+2];
cvt.s32.s16 %r252, %rs14;
and.b32 %r253, %r252, 30583;
prmt.b32 %r254, %r197, %r197, %r253;
shr.u32 %r255, %r252, 2;
and.b32 %r256, %r255, 30583;
prmt.b32 %r257, %r197, %r197, %r256;
prmt.b32 %r139, %r254, %r257, %r202;
prmt.b32 %r143, %r254, %r257, %r203;

dp4a.s32.s32 %r137, %r72, %r139, %r133;


dp4a.s32.s32 %r141, %r76, %r143, %r137;

ld.global.nc.u16 %rs15, [%rd33+4];
cvt.s32.s16 %r258, %rs15;
and.b32 %r259, %r258, 30583;
prmt.b32 %r260, %r197, %r197, %r259;
shr.u32 %r261, %r258, 2;
and.b32 %r262, %r261, 30583;
prmt.b32 %r263, %r197, %r197, %r262;
prmt.b32 %r147, %r260, %r263, %r202;
prmt.b32 %r151, %r260, %r263, %r203;

dp4a.s32.s32 %r145, %r80, %r147, %r141;


dp4a.s32.s32 %r149, %r84, %r151, %r145;

ld.global.nc.u16 %rs16, [%rd33+6];
cvt.s32.s16 %r264, %rs16;
and.b32 %r265, %r264, 30583;
prmt.b32 %r266, %r197, %r197, %r265;
shr.u32 %r267, %r264, 2;
and.b32 %r268, %r267, 30583;
prmt.b32 %r269, %r197, %r197, %r268;
prmt.b32 %r155, %r266, %r269, %r202;
prmt.b32 %r159, %r266, %r269, %r203;

dp4a.s32.s32 %r153, %r88, %r155, %r149;


dp4a.s32.s32 %r157, %r92, %r159, %r153;


{.reg .f16 low,high;
mov.b32 {low,high},%r95;
cvt.f32.f16 %f30, low;}


mul.ftz.f32 %f37, %f29, %f30;
cvt.rn.f32.s32 %f38, %r157;
fma.rn.ftz.f32 %f106, %f37, %f38, %f106;
add.s64 %rd34, %rd32, %rd4;
ld.global.nc.u16 %rs4, [%rd34];

{ cvt.f32.f16 %f31, %rs4;}


add.s64 %rd35, %rd33, %rd4;
ld.global.nc.u16 %rs17, [%rd35];
cvt.s32.s16 %r270, %rs17;
and.b32 %r271, %r270, 30583;
prmt.b32 %r272, %r197, %r197, %r271;
shr.u32 %r273, %r270, 2;
and.b32 %r274, %r273, 30583;
prmt.b32 %r275, %r197, %r197, %r274;
prmt.b32 %r164, %r272, %r275, %r202;
prmt.b32 %r168, %r272, %r275, %r203;

dp4a.s32.s32 %r162, %r64, %r164, %r165;


dp4a.s32.s32 %r166, %r68, %r168, %r162;

ld.global.nc.u16 %rs18, [%rd35+2];
cvt.s32.s16 %r276, %rs18;
and.b32 %r277, %r276, 30583;
prmt.b32 %r278, %r197, %r197, %r277;
shr.u32 %r279, %r276, 2;
and.b32 %r280, %r279, 30583;
prmt.b32 %r281, %r197, %r197, %r280;
prmt.b32 %r172, %r278, %r281, %r202;
prmt.b32 %r176, %r278, %r281, %r203;

dp4a.s32.s32 %r170, %r72, %r172, %r166;


dp4a.s32.s32 %r174, %r76, %r176, %r170;

ld.global.nc.u16 %rs19, [%rd35+4];
cvt.s32.s16 %r282, %rs19;
and.b32 %r283, %r282, 30583;
prmt.b32 %r284, %r197, %r197, %r283;
shr.u32 %r285, %r282, 2;
and.b32 %r286, %r285, 30583;
prmt.b32 %r287, %r197, %r197, %r286;
prmt.b32 %r180, %r284, %r287, %r202;
prmt.b32 %r184, %r284, %r287, %r203;

dp4a.s32.s32 %r178, %r80, %r180, %r174;


dp4a.s32.s32 %r182, %r84, %r184, %r178;

ld.global.nc.u16 %rs20, [%rd35+6];
cvt.s32.s16 %r288, %rs20;
and.b32 %r289, %r288, 30583;
prmt.b32 %r290, %r197, %r197, %r289;
shr.u32 %r291, %r288, 2;
and.b32 %r292, %r291, 30583;
prmt.b32 %r293, %r197, %r197, %r292;
prmt.b32 %r188, %r290, %r293, %r202;
prmt.b32 %r192, %r290, %r293, %r203;

dp4a.s32.s32 %r186, %r88, %r188, %r182;


dp4a.s32.s32 %r190, %r92, %r192, %r186;


{.reg .f16 low,high;
mov.b32 {low,high},%r95;
cvt.f32.f16 %f32, low;}


mul.ftz.f32 %f39, %f31, %f32;
cvt.rn.f32.s32 %f40, %r190;
fma.rn.ftz.f32 %f105, %f39, %f40, %f105;
add.s32 %r378, %r378, 128;
add.s64 %rd38, %rd38, 1152;
add.s32 %r379, %r379, 64;
setp.lt.s32 %p3, %r379, %r48;
@%p3 bra $L__BB16_5;

$L__BB16_6:
setp.eq.s32 %p4, %r1, 0;
@%p4 bra $L__BB16_8;

shl.b32 %r294, %r1, 9;
mov.u32 %r295, _ZZ13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r296, %r295, %r294;
shl.b32 %r297, %r2, 2;
add.s32 %r298, %r296, %r297;
st.shared.f32 [%r298+-512], %f108;
st.shared.f32 [%r298+-384], %f107;
st.shared.f32 [%r298+-256], %f106;
st.shared.f32 [%r298+-128], %f105;

$L__BB16_8:
setp.ne.s32 %p5, %r1, 0;
bar.sync 0;
@%p5 bra $L__BB16_17;

mov.u32 %r299, %ctaid.z;
mov.u32 %r300, %ctaid.y;
mad.lo.s32 %r301, %r300, %r31, %r4;
mad.lo.s32 %r302, %r299, %r37, %r301;
shl.b32 %r303, %r2, 2;
mov.u32 %r304, _ZZ13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r18, %r304, %r303;
mov.u32 %r305, 2;
ld.shared.f32 %f41, [%r18];
add.ftz.f32 %f42, %f41, %f108;
ld.shared.f32 %f43, [%r18+512];
add.ftz.f32 %f44, %f43, %f42;
ld.shared.f32 %f45, [%r18+1024];
add.ftz.f32 %f46, %f45, %f44;
mov.b32 %r306, %f46;
mov.u32 %r307, 31;
mov.u32 %r308, 16;
mov.u32 %r309, -1;
shfl.sync.bfly.b32 %r310|%p6, %r306, %r308, %r307, %r309;
mov.b32 %f47, %r310;
add.ftz.f32 %f48, %f46, %f47;
mov.b32 %r311, %f48;
mov.u32 %r312, 8;
shfl.sync.bfly.b32 %r313|%p7, %r311, %r312, %r307, %r309;
mov.b32 %f49, %r313;
add.ftz.f32 %f50, %f48, %f49;
mov.b32 %r314, %f50;
mov.u32 %r315, 4;
shfl.sync.bfly.b32 %r316|%p8, %r314, %r315, %r307, %r309;
mov.b32 %f51, %r316;
add.ftz.f32 %f52, %f50, %f51;
mov.b32 %r317, %f52;
shfl.sync.bfly.b32 %r318|%p9, %r317, %r305, %r307, %r309;
mov.b32 %f53, %r318;
add.ftz.f32 %f54, %f52, %f53;
mov.b32 %r319, %f54;
mov.u32 %r320, 1;
shfl.sync.bfly.b32 %r321|%p10, %r319, %r320, %r307, %r309;
mov.b32 %f55, %r321;
add.ftz.f32 %f13, %f54, %f55;
setp.ge.u32 %p11, %r4, %r25;
setp.ne.s32 %p12, %r2, 0;
cvta.to.global.u64 %rd36, %rd13;
mul.wide.u32 %rd37, %r302, 4;
add.s64 %rd9, %rd36, %rd37;
or.pred %p13, %p12, %p11;
@%p13 bra $L__BB16_11;

st.global.f32 [%rd9], %f13;

$L__BB16_11:
ld.shared.f32 %f56, [%r18+128];
add.ftz.f32 %f57, %f56, %f107;
ld.shared.f32 %f58, [%r18+640];
add.ftz.f32 %f59, %f58, %f57;
ld.shared.f32 %f60, [%r18+1152];
add.ftz.f32 %f61, %f60, %f59;
mov.b32 %r322, %f61;
shfl.sync.bfly.b32 %r326|%p14, %r322, %r308, %r307, %r309;
mov.b32 %f62, %r326;
add.ftz.f32 %f63, %f61, %f62;
mov.b32 %r327, %f63;
shfl.sync.bfly.b32 %r329|%p15, %r327, %r312, %r307, %r309;
mov.b32 %f64, %r329;
add.ftz.f32 %f65, %f63, %f64;
mov.b32 %r330, %f65;
shfl.sync.bfly.b32 %r332|%p16, %r330, %r315, %r307, %r309;
mov.b32 %f66, %r332;
add.ftz.f32 %f67, %f65, %f66;
mov.b32 %r333, %f67;
shfl.sync.bfly.b32 %r335|%p17, %r333, %r305, %r307, %r309;
mov.b32 %f68, %r335;
add.ftz.f32 %f69, %f67, %f68;
mov.b32 %r336, %f69;
shfl.sync.bfly.b32 %r338|%p18, %r336, %r320, %r307, %r309;
mov.b32 %f70, %r338;
add.ftz.f32 %f14, %f69, %f70;
add.s32 %r339, %r4, 1;
setp.ge.u32 %p19, %r339, %r25;
setp.ne.s32 %p20, %r2, 1;
or.pred %p21, %p20, %p19;
@%p21 bra $L__BB16_13;

st.global.f32 [%rd9+4], %f14;

$L__BB16_13:
ld.shared.f32 %f71, [%r18+256];
add.ftz.f32 %f72, %f71, %f106;
ld.shared.f32 %f73, [%r18+768];
add.ftz.f32 %f74, %f73, %f72;
ld.shared.f32 %f75, [%r18+1280];
add.ftz.f32 %f76, %f75, %f74;
mov.b32 %r340, %f76;
mov.u32 %r341, 31;
mov.u32 %r342, 16;
mov.u32 %r343, -1;
shfl.sync.bfly.b32 %r344|%p22, %r340, %r342, %r341, %r343;
mov.b32 %f77, %r344;
add.ftz.f32 %f78, %f76, %f77;
mov.b32 %r345, %f78;
mov.u32 %r346, 8;
shfl.sync.bfly.b32 %r347|%p23, %r345, %r346, %r341, %r343;
mov.b32 %f79, %r347;
add.ftz.f32 %f80, %f78, %f79;
mov.b32 %r348, %f80;
mov.u32 %r349, 4;
shfl.sync.bfly.b32 %r350|%p24, %r348, %r349, %r341, %r343;
mov.b32 %f81, %r350;
add.ftz.f32 %f82, %f80, %f81;
mov.b32 %r351, %f82;
mov.u32 %r352, 2;
shfl.sync.bfly.b32 %r353|%p25, %r351, %r352, %r341, %r343;
mov.b32 %f83, %r353;
add.ftz.f32 %f84, %f82, %f83;
mov.b32 %r354, %f84;
mov.u32 %r355, 1;
shfl.sync.bfly.b32 %r356|%p26, %r354, %r355, %r341, %r343;
mov.b32 %f85, %r356;
add.ftz.f32 %f15, %f84, %f85;
add.s32 %r357, %r4, 2;
setp.ge.u32 %p27, %r357, %r25;
setp.ne.s32 %p28, %r2, 2;
or.pred %p29, %p28, %p27;
@%p29 bra $L__BB16_15;

st.global.f32 [%rd9+8], %f15;

$L__BB16_15:
ld.shared.f32 %f86, [%r18+384];
add.ftz.f32 %f87, %f86, %f105;
ld.shared.f32 %f88, [%r18+896];
add.ftz.f32 %f89, %f88, %f87;
ld.shared.f32 %f90, [%r18+1408];
add.ftz.f32 %f91, %f90, %f89;
mov.b32 %r358, %f91;
shfl.sync.bfly.b32 %r362|%p30, %r358, %r342, %r341, %r343;
mov.b32 %f92, %r362;
add.ftz.f32 %f93, %f91, %f92;
mov.b32 %r363, %f93;
shfl.sync.bfly.b32 %r365|%p31, %r363, %r346, %r341, %r343;
mov.b32 %f94, %r365;
add.ftz.f32 %f95, %f93, %f94;
mov.b32 %r366, %f95;
shfl.sync.bfly.b32 %r368|%p32, %r366, %r349, %r341, %r343;
mov.b32 %f96, %r368;
add.ftz.f32 %f97, %f95, %f96;
mov.b32 %r369, %f97;
shfl.sync.bfly.b32 %r371|%p33, %r369, %r352, %r341, %r343;
mov.b32 %f98, %r371;
add.ftz.f32 %f99, %f97, %f98;
mov.b32 %r372, %f99;
shfl.sync.bfly.b32 %r374|%p34, %r372, %r355, %r341, %r343;
mov.b32 %f100, %r374;
add.ftz.f32 %f16, %f99, %f100;
add.s32 %r375, %r4, 3;
setp.ge.u32 %p35, %r375, %r25;
setp.ne.s32 %p36, %r2, 3;
or.pred %p37, %p36, %p35;
@%p37 bra $L__BB16_17;

st.global.f32 [%rd9+12], %f16;

$L__BB16_17:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<34>;
.reg .b16 %rs<7>;
.reg .f32 %f<120>;
.reg .b32 %r<277>;
.reg .b64 %rd<72>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[384];

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate[384];

ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd25, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r45, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
ld.param.f32 %f31, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
ld.param.u64 %rd22, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+16];
ld.param.u64 %rd21, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+8];
ld.param.u64 %rd20, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3];
mov.u32 %r1, %tid.y;
mov.u32 %r2, %ctaid.y;
setp.eq.s64 %p2, %rd19, 0;
@%p2 bra $L__BB17_2;

cvta.to.global.u64 %rd26, %rd19;
mul.wide.u32 %rd27, %r2, 4;
add.s64 %rd28, %rd26, %rd27;
ld.global.nc.u32 %r273, [%rd28];
mul.hi.u32 %r46, %r2, %r30;
add.s32 %r47, %r46, %r2;
shr.u32 %r48, %r47, %r31;
mul.lo.s32 %r49, %r48, %r32;
sub.s32 %r274, %r2, %r49;
bra.uni $L__BB17_3;

$L__BB17_2:
mul.hi.u32 %r50, %r2, %r34;
add.s32 %r51, %r50, %r2;
shr.u32 %r273, %r51, %r35;
mov.u32 %r274, %r2;

$L__BB17_3:
mov.u32 %r52, %tid.x;
or.b32 %r53, %r52, %r1;
setp.ne.s32 %p3, %r53, 0;
mov.f32 %f113, 0f00000000;
mov.f32 %f109, %f113;
mov.f32 %f110, %f113;
@%p3 bra $L__BB17_8;

setp.eq.s64 %p4, %rd20, 0;
mov.f32 %f109, 0f00000000;
mov.f32 %f110, %f109;
@%p4 bra $L__BB17_6;

cvta.to.global.u64 %rd29, %rd20;
mov.u32 %r54, %ctaid.z;
mul.lo.s32 %r55, %r54, %r45;
cvt.u64.u32 %rd30, %r55;
selp.b32 %r56, %r2, %r273, %p2;
mul.lo.s32 %r57, %r56, %r39;
cvt.u64.u32 %rd31, %r57;
mov.u32 %r58, %ctaid.x;
cvt.s64.s32 %rd32, %r58;
cvt.u64.u32 %rd33, %r52;
add.s64 %rd34, %rd33, %rd32;
add.s64 %rd35, %rd34, %rd30;
add.s64 %rd36, %rd35, %rd31;
shl.b64 %rd37, %rd36, 2;
add.s64 %rd38, %rd29, %rd37;
ld.global.f32 %f110, [%rd38];

$L__BB17_6:
setp.eq.s64 %p6, %rd22, 0;
setp.eq.s64 %p7, %rd21, 0;
or.pred %p8, %p6, %p7;
@%p8 bra $L__BB17_8;

cvta.to.global.u64 %rd39, %rd22;
mov.u32 %r60, %ctaid.z;
mul.lo.s32 %r61, %r60, %r45;
cvt.u64.u32 %rd40, %r61;
selp.b32 %r62, %r2, %r273, %p2;
mul.lo.s32 %r63, %r62, %r39;
cvt.u64.u32 %rd41, %r63;
mov.u32 %r64, %ctaid.x;
cvt.s64.s32 %rd42, %r64;
cvt.u64.u32 %rd43, %r52;
add.s64 %rd44, %rd43, %rd42;
add.s64 %rd45, %rd44, %rd40;
add.s64 %rd46, %rd45, %rd41;
shl.b64 %rd47, %rd46, 2;
add.s64 %rd48, %rd39, %rd47;
ld.global.f32 %f109, [%rd48];

$L__BB17_8:
shl.b32 %r67, %r1, 5;
add.s32 %r68, %r67, %r52;
shr.u32 %r69, %r68, 31;
add.s32 %r70, %r68, %r69;
shr.s32 %r276, %r70, 1;
shr.u32 %r71, %r29, 6;
setp.ge.s32 %p10, %r276, %r71;
mov.f32 %f115, %f113;
@%p10 bra $L__BB17_13;

mov.u32 %r72, %ctaid.z;
mul.lo.s32 %r73, %r274, %r38;
cvt.u64.u32 %rd49, %r73;
and.b32 %r79, %r70, -2;
sub.s32 %r80, %r68, %r79;
shl.b32 %r81, %r80, 2;
cvt.s64.s32 %rd50, %r80;
mul.lo.s32 %r82, %r72, %r44;
cvt.u64.u32 %rd51, %r82;
add.s64 %rd52, %rd51, %rd50;
add.s64 %rd4, %rd52, %rd49;
shl.b32 %r275, %r276, 1;
mul.hi.u32 %r83, %r72, %r40;
add.s32 %r84, %r83, %r72;
shr.u32 %r85, %r84, %r41;
mad.lo.s32 %r86, %r85, %r43, %r276;
mad.lo.s32 %r87, %r273, %r37, %r86;
mov.u32 %r88, %ctaid.x;
mad.lo.s32 %r89, %r88, %r33, %r87;
mul.wide.s32 %rd53, %r89, 18;
mul.wide.s32 %rd54, %r81, 2;
add.s64 %rd55, %rd53, %rd54;
cvta.to.global.u64 %rd56, %rd21;
add.s64 %rd57, %rd56, %rd55;
add.s64 %rd71, %rd57, 4;
cvta.to.global.u64 %rd58, %rd17;
add.s64 %rd59, %rd58, %rd55;
add.s64 %rd70, %rd59, 8;
add.s64 %rd69, %rd56, %rd53;
add.s64 %rd68, %rd58, %rd53;
mov.f32 %f115, 0f00000000;
cvta.to.global.u64 %rd60, %rd18;
mov.f32 %f113, %f115;

$L__BB17_10:
ld.global.nc.u16 %rs1, [%rd68];

{ cvt.f32.f16 %f40, %rs1;}


cvt.u64.u32 %rd61, %r275;
add.s64 %rd62, %rd4, %rd61;
mul.lo.s64 %rd63, %rd62, 36;
ld.global.nc.u16 %rs2, [%rd70+-6];
cvt.s32.s16 %r123, %rs2;
add.s64 %rd64, %rd60, %rd63;
ld.global.nc.u32 %r14, [%rd64+4];
ld.global.nc.u32 %r15, [%rd64+8];
and.b32 %r124, %r123, 30583;
mov.u32 %r125, 33620223;
prmt.b32 %r126, %r125, %r125, %r124;
shr.u32 %r127, %r123, 2;
and.b32 %r128, %r127, 30583;
prmt.b32 %r129, %r125, %r125, %r128;
mov.u32 %r130, 20800;
prmt.b32 %r92, %r126, %r129, %r130;
mov.u32 %r131, 29538;
prmt.b32 %r96, %r126, %r129, %r131;
mov.u32 %r93, 0;

dp4a.s32.s32 %r90, %r14, %r92, %r93;


dp4a.s32.s32 %r94, %r15, %r96, %r90;

ld.global.nc.u16 %rs3, [%rd70+-4];
cvt.s32.s16 %r132, %rs3;
ld.global.nc.u32 %r16, [%rd64+12];
ld.global.nc.u32 %r17, [%rd64+16];
and.b32 %r133, %r132, 30583;
prmt.b32 %r134, %r125, %r125, %r133;
shr.u32 %r135, %r132, 2;
and.b32 %r136, %r135, 30583;
prmt.b32 %r137, %r125, %r125, %r136;
prmt.b32 %r100, %r134, %r137, %r130;
prmt.b32 %r104, %r134, %r137, %r131;

dp4a.s32.s32 %r98, %r16, %r100, %r94;


dp4a.s32.s32 %r102, %r17, %r104, %r98;

ld.global.nc.u16 %rs4, [%rd70+-2];
cvt.s32.s16 %r138, %rs4;
ld.global.nc.u32 %r18, [%rd64+20];
ld.global.nc.u32 %r19, [%rd64+24];
and.b32 %r139, %r138, 30583;
prmt.b32 %r140, %r125, %r125, %r139;
shr.u32 %r141, %r138, 2;
and.b32 %r142, %r141, 30583;
prmt.b32 %r143, %r125, %r125, %r142;
prmt.b32 %r108, %r140, %r143, %r130;
prmt.b32 %r112, %r140, %r143, %r131;

dp4a.s32.s32 %r106, %r18, %r108, %r102;


dp4a.s32.s32 %r110, %r19, %r112, %r106;

ld.global.nc.u16 %rs5, [%rd70];
cvt.s32.s16 %r144, %rs5;
ld.global.nc.u32 %r20, [%rd64+28];
ld.global.nc.u32 %r21, [%rd64+32];
and.b32 %r145, %r144, 30583;
prmt.b32 %r146, %r125, %r125, %r145;
shr.u32 %r147, %r144, 2;
and.b32 %r148, %r147, 30583;
prmt.b32 %r149, %r125, %r125, %r148;
prmt.b32 %r116, %r146, %r149, %r130;
prmt.b32 %r120, %r146, %r149, %r131;

dp4a.s32.s32 %r114, %r20, %r116, %r110;


dp4a.s32.s32 %r118, %r21, %r120, %r114;

ld.global.nc.u32 %r22, [%rd64];

{.reg .f16 low,high;
mov.b32 {low,high},%r22;
cvt.f32.f16 %f41, low;}


mul.ftz.f32 %f42, %f40, %f41;
cvt.rn.f32.s32 %f43, %r118;
fma.rn.ftz.f32 %f115, %f42, %f43, %f115;
setp.eq.s64 %p11, %rd21, 0;
@%p11 bra $L__BB17_12;

ld.global.u16 %rs6, [%rd69];

{ cvt.f32.f16 %f44, %rs6;}


ld.global.s16 %r183, [%rd71+-2];
and.b32 %r184, %r183, 30583;
prmt.b32 %r186, %r125, %r125, %r184;
shr.u32 %r187, %r183, 2;
and.b32 %r188, %r187, 30583;
prmt.b32 %r189, %r125, %r125, %r188;
prmt.b32 %r152, %r186, %r189, %r130;
prmt.b32 %r156, %r186, %r189, %r131;

dp4a.s32.s32 %r150, %r14, %r152, %r93;


dp4a.s32.s32 %r154, %r15, %r156, %r150;

ld.global.s16 %r192, [%rd71];
and.b32 %r193, %r192, 30583;
prmt.b32 %r194, %r125, %r125, %r193;
shr.u32 %r195, %r192, 2;
and.b32 %r196, %r195, 30583;
prmt.b32 %r197, %r125, %r125, %r196;
prmt.b32 %r160, %r194, %r197, %r130;
prmt.b32 %r164, %r194, %r197, %r131;

dp4a.s32.s32 %r158, %r16, %r160, %r154;


dp4a.s32.s32 %r162, %r17, %r164, %r158;

ld.global.s16 %r198, [%rd71+2];
and.b32 %r199, %r198, 30583;
prmt.b32 %r200, %r125, %r125, %r199;
shr.u32 %r201, %r198, 2;
and.b32 %r202, %r201, 30583;
prmt.b32 %r203, %r125, %r125, %r202;
prmt.b32 %r168, %r200, %r203, %r130;
prmt.b32 %r172, %r200, %r203, %r131;

dp4a.s32.s32 %r166, %r18, %r168, %r162;


dp4a.s32.s32 %r170, %r19, %r172, %r166;

ld.global.s16 %r204, [%rd71+4];
and.b32 %r205, %r204, 30583;
prmt.b32 %r206, %r125, %r125, %r205;
shr.u32 %r207, %r204, 2;
and.b32 %r208, %r207, 30583;
prmt.b32 %r209, %r125, %r125, %r208;
prmt.b32 %r176, %r206, %r209, %r130;
prmt.b32 %r180, %r206, %r209, %r131;

dp4a.s32.s32 %r174, %r20, %r176, %r170;


dp4a.s32.s32 %r178, %r21, %r180, %r174;


{.reg .f16 low,high;
mov.b32 {low,high},%r22;
cvt.f32.f16 %f45, low;}


mul.ftz.f32 %f46, %f44, %f45;
cvt.rn.f32.s32 %f47, %r178;
fma.rn.ftz.f32 %f113, %f46, %f47, %f113;

$L__BB17_12:
add.s32 %r275, %r275, 128;
add.s64 %rd71, %rd71, 1152;
add.s64 %rd70, %rd70, 1152;
add.s64 %rd69, %rd69, 1152;
add.s64 %rd68, %rd68, 1152;
add.s32 %r276, %r276, 64;
setp.lt.s32 %p12, %r276, %r71;
@%p12 bra $L__BB17_10;

$L__BB17_13:
setp.eq.s32 %p13, %r1, 0;
@%p13 bra $L__BB17_16;

setp.eq.s64 %p14, %rd21, 0;
add.s32 %r25, %r1, -1;
shl.b32 %r214, %r1, 7;
mov.u32 %r215, _ZZ13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r216, %r215, %r214;
shl.b32 %r217, %r52, 2;
add.s32 %r218, %r216, %r217;
st.shared.f32 [%r218+-128], %f115;
@%p14 bra $L__BB17_16;

shl.b32 %r220, %r25, 7;
mov.u32 %r221, _ZZ13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r222, %r221, %r220;
add.s32 %r224, %r222, %r217;
st.shared.f32 [%r224], %f113;

$L__BB17_16:
setp.ne.s32 %p15, %r1, 0;
bar.sync 0;
@%p15 bra $L__BB17_35;

setp.ne.s64 %p16, %rd21, 0;
shl.b32 %r227, %r52, 2;
mov.u32 %r228, _ZZ13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r26, %r228, %r227;
ld.shared.f32 %f48, [%r26];
add.ftz.f32 %f13, %f115, %f48;
@%p16 bra $L__BB17_19;
bra.uni $L__BB17_18;

$L__BB17_19:
mov.u32 %r231, _ZZ13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r232, %r231, %r227;
ld.shared.f32 %f52, [%r232];
add.ftz.f32 %f53, %f113, %f52;
ld.shared.f32 %f54, [%r26+128];
add.ftz.f32 %f55, %f13, %f54;
ld.shared.f32 %f56, [%r232+128];
add.ftz.f32 %f57, %f53, %f56;
ld.shared.f32 %f58, [%r26+256];
add.ftz.f32 %f117, %f55, %f58;
ld.shared.f32 %f59, [%r232+256];
add.ftz.f32 %f113, %f57, %f59;
bra.uni $L__BB17_20;

$L__BB17_18:
ld.shared.f32 %f49, [%r26+128];
add.ftz.f32 %f50, %f13, %f49;
ld.shared.f32 %f51, [%r26+256];
add.ftz.f32 %f117, %f50, %f51;

$L__BB17_20:
mov.b32 %r233, %f117;
mov.u32 %r234, 31;
mov.u32 %r235, 16;
mov.u32 %r236, -1;
shfl.sync.bfly.b32 %r237|%p17, %r233, %r235, %r234, %r236;
mov.b32 %f60, %r237;
add.ftz.f32 %f61, %f117, %f60;
mov.b32 %r238, %f61;
mov.u32 %r239, 8;
shfl.sync.bfly.b32 %r240|%p18, %r238, %r239, %r234, %r236;
mov.b32 %f62, %r240;
add.ftz.f32 %f63, %f61, %f62;
mov.b32 %r241, %f63;
mov.u32 %r242, 4;
shfl.sync.bfly.b32 %r243|%p19, %r241, %r242, %r234, %r236;
mov.b32 %f64, %r243;
add.ftz.f32 %f65, %f63, %f64;
mov.b32 %r244, %f65;
mov.u32 %r245, 2;
shfl.sync.bfly.b32 %r246|%p20, %r244, %r245, %r234, %r236;
mov.b32 %f66, %r246;
add.ftz.f32 %f19, %f65, %f66;
mov.b32 %r247, %f19;
mov.u32 %r248, 1;
shfl.sync.bfly.b32 %r27|%p1, %r247, %r248, %r234, %r236;
setp.eq.s64 %p21, %rd21, 0;
@%p21 bra $L__BB17_22;

mov.b32 %r249, %f113;
shfl.sync.bfly.b32 %r253|%p22, %r249, %r235, %r234, %r236;
mov.b32 %f67, %r253;
add.ftz.f32 %f68, %f113, %f67;
mov.b32 %r254, %f68;
shfl.sync.bfly.b32 %r256|%p23, %r254, %r239, %r234, %r236;
mov.b32 %f69, %r256;
add.ftz.f32 %f70, %f68, %f69;
mov.b32 %r257, %f70;
shfl.sync.bfly.b32 %r259|%p24, %r257, %r242, %r234, %r236;
mov.b32 %f71, %r259;
add.ftz.f32 %f72, %f70, %f71;
mov.b32 %r260, %f72;
shfl.sync.bfly.b32 %r262|%p25, %r260, %r245, %r234, %r236;
mov.b32 %f73, %r262;
add.ftz.f32 %f74, %f72, %f73;
mov.b32 %r263, %f74;
shfl.sync.bfly.b32 %r265|%p26, %r263, %r248, %r234, %r236;
mov.b32 %f75, %r265;
add.ftz.f32 %f113, %f74, %f75;

$L__BB17_22:
setp.ne.s32 %p27, %r52, 0;
mov.b32 %f76, %r27;
add.ftz.f32 %f22, %f19, %f76;
@%p27 bra $L__BB17_35;

add.ftz.f32 %f119, %f110, %f22;
@%p21 bra $L__BB17_34;

add.ftz.f32 %f24, %f109, %f113;
setp.gt.s32 %p29, %r28, 2;
@%p29 bra $L__BB17_28;

setp.eq.s32 %p32, %r28, 1;
@%p32 bra $L__BB17_32;

setp.eq.s32 %p33, %r28, 2;
@%p33 bra $L__BB17_27;
bra.uni $L__BB17_33;

$L__BB17_27:
mul.ftz.f32 %f96, %f24, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f97, %f96;
add.ftz.f32 %f98, %f97, 0f3F800000;
div.approx.ftz.f32 %f99, %f24, %f98;
mul.ftz.f32 %f119, %f119, %f99;
bra.uni $L__BB17_34;

$L__BB17_28:
setp.eq.s32 %p30, %r28, 3;
@%p30 bra $L__BB17_31;

setp.ne.s32 %p31, %r28, 6;
@%p31 bra $L__BB17_33;

min.ftz.f32 %f77, %f24, %f31;
neg.ftz.f32 %f78, %f31;
min.ftz.f32 %f79, %f119, %f31;
max.ftz.f32 %f80, %f79, %f78;
mul.ftz.f32 %f81, %f77, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f82, %f81;
add.ftz.f32 %f83, %f82, 0f3F800000;
div.approx.ftz.f32 %f84, %f77, %f83;
mul.ftz.f32 %f119, %f80, %f84;
bra.uni $L__BB17_34;

$L__BB17_32:
mul.ftz.f32 %f100, %f24, 0f3F000000;
mul.ftz.f32 %f101, %f24, 0f3D372713;
fma.rn.ftz.f32 %f102, %f24, %f101, 0f3F800000;
mul.ftz.f32 %f103, %f24, 0f3F4C422A;
mul.ftz.f32 %f104, %f103, %f102;
tanh.approx.f32 %f105, %f104;
add.ftz.f32 %f106, %f105, 0f3F800000;
mul.ftz.f32 %f107, %f100, %f106;
mul.ftz.f32 %f119, %f119, %f107;
bra.uni $L__BB17_34;

$L__BB17_31:
mov.f32 %f85, 0f40E00000;
min.ftz.f32 %f86, %f24, %f85;
min.ftz.f32 %f87, %f119, %f85;
mov.f32 %f88, 0fC0E00000;
max.ftz.f32 %f89, %f87, %f88;
mul.ftz.f32 %f90, %f86, 0f3FD9DB23;
mul.ftz.f32 %f91, %f90, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f92, %f91;
add.ftz.f32 %f93, %f92, 0f3F800000;
div.approx.ftz.f32 %f94, %f86, %f93;
add.ftz.f32 %f95, %f89, 0f3F800000;
mul.ftz.f32 %f119, %f95, %f94;
bra.uni $L__BB17_34;

$L__BB17_33:
mul.ftz.f32 %f119, %f119, %f24;

$L__BB17_34:
cvta.to.global.u64 %rd65, %rd25;
mov.u32 %r268, %ctaid.z;
mov.u32 %r269, %ctaid.y;
mov.u32 %r270, %ctaid.x;
mad.lo.s32 %r271, %r269, %r39, %r270;
mad.lo.s32 %r272, %r268, %r45, %r271;
mul.wide.u32 %rd66, %r272, 4;
add.s64 %rd67, %rd65, %rd66;
st.global.f32 [%rd67], %f119;

$L__BB17_35:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<14>;
.reg .b16 %rs<16>;
.reg .f32 %f<43>;
.reg .b32 %r<285>;
.reg .b64 %rd<43>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[384];

ld.param.u64 %rd11, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd12, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd9, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd10, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
cvta.to.global.u64 %rd1, %rd12;
cvta.to.global.u64 %rd2, %rd11;
mov.u32 %r1, %tid.y;
shl.b32 %r43, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r43, %r2;
shr.u32 %r4, %r26, 6;
mov.u32 %r281, %ctaid.y;
setp.eq.s64 %p1, %rd9, 0;
@%p1 bra $L__BB18_2;

cvta.to.global.u64 %rd13, %rd9;
mul.wide.u32 %rd14, %r281, 4;
add.s64 %rd15, %rd13, %rd14;
ld.global.nc.u32 %r280, [%rd15];
mul.hi.u32 %r44, %r281, %r27;
add.s32 %r45, %r44, %r281;
shr.u32 %r46, %r45, %r28;
mul.lo.s32 %r47, %r46, %r29;
sub.s32 %r281, %r281, %r47;
bra.uni $L__BB18_3;

$L__BB18_2:
mul.hi.u32 %r48, %r281, %r31;
add.s32 %r49, %r48, %r281;
shr.u32 %r280, %r49, %r32;

$L__BB18_3:
mov.u32 %r11, %ctaid.z;
mul.hi.u32 %r50, %r11, %r37;
add.s32 %r51, %r50, %r11;
shr.u32 %r52, %r51, %r38;
mul.lo.s32 %r12, %r52, %r40;
mul.lo.s32 %r13, %r280, %r34;
mov.u32 %r53, %ctaid.x;
mul.lo.s32 %r14, %r53, %r30;
shr.u32 %r54, %r3, 31;
add.s32 %r55, %r3, %r54;
shr.s32 %r284, %r55, 1;
setp.ge.s32 %p2, %r284, %r4;
mov.f32 %f42, 0f00000000;
@%p2 bra $L__BB18_9;

mul.lo.s32 %r56, %r281, %r35;
cvt.u64.u32 %rd16, %r56;
and.b32 %r59, %r55, -2;
sub.s32 %r60, %r3, %r59;
cvt.s64.s32 %rd17, %r60;
mul.lo.s32 %r61, %r11, %r41;
cvt.u64.u32 %rd18, %r61;
add.s64 %rd19, %rd18, %rd17;
add.s64 %rd3, %rd19, %rd16;
shl.b32 %r62, %r60, 2;
cvt.s64.s32 %rd4, %r62;
not.b32 %r63, %r284;
add.s32 %r16, %r4, %r63;
and.b32 %r64, %r16, 64;
setp.ne.s32 %p3, %r64, 0;
mov.f32 %f42, 0f00000000;
@%p3 bra $L__BB18_6;

shl.b32 %r98, %r284, 1;
cvt.u64.u32 %rd20, %r98;
add.s32 %r99, %r13, %r14;
add.s32 %r100, %r99, %r12;
add.s32 %r101, %r100, %r284;
mul.wide.s32 %rd21, %r101, 18;
add.s64 %rd22, %rd2, %rd21;
ld.global.nc.u16 %rs1, [%rd22];

{ cvt.f32.f16 %f11, %rs1;}


shl.b64 %rd23, %rd4, 1;
add.s64 %rd24, %rd22, %rd23;
add.s64 %rd25, %rd3, %rd20;
mul.lo.s64 %rd26, %rd25, 36;
ld.global.nc.u16 %rs2, [%rd24+2];
cvt.s32.s16 %r102, %rs2;
add.s64 %rd27, %rd1, %rd26;
ld.global.nc.u32 %r66, [%rd27+4];
ld.global.nc.u32 %r70, [%rd27+8];
and.b32 %r103, %r102, 30583;
mov.u32 %r104, 33620223;
prmt.b32 %r105, %r104, %r104, %r103;
shr.u32 %r106, %r102, 2;
and.b32 %r107, %r106, 30583;
prmt.b32 %r108, %r104, %r104, %r107;
mov.u32 %r109, 20800;
prmt.b32 %r67, %r105, %r108, %r109;
mov.u32 %r110, 29538;
prmt.b32 %r71, %r105, %r108, %r110;
mov.u32 %r68, 0;

dp4a.s32.s32 %r65, %r66, %r67, %r68;


dp4a.s32.s32 %r69, %r70, %r71, %r65;

ld.global.nc.u16 %rs3, [%rd24+4];
cvt.s32.s16 %r111, %rs3;
ld.global.nc.u32 %r74, [%rd27+12];
ld.global.nc.u32 %r78, [%rd27+16];
and.b32 %r112, %r111, 30583;
prmt.b32 %r113, %r104, %r104, %r112;
shr.u32 %r114, %r111, 2;
and.b32 %r115, %r114, 30583;
prmt.b32 %r116, %r104, %r104, %r115;
prmt.b32 %r75, %r113, %r116, %r109;
prmt.b32 %r79, %r113, %r116, %r110;

dp4a.s32.s32 %r73, %r74, %r75, %r69;


dp4a.s32.s32 %r77, %r78, %r79, %r73;

ld.global.nc.u16 %rs4, [%rd24+6];
cvt.s32.s16 %r117, %rs4;
ld.global.nc.u32 %r82, [%rd27+20];
ld.global.nc.u32 %r86, [%rd27+24];
and.b32 %r118, %r117, 30583;
prmt.b32 %r119, %r104, %r104, %r118;
shr.u32 %r120, %r117, 2;
and.b32 %r121, %r120, 30583;
prmt.b32 %r122, %r104, %r104, %r121;
prmt.b32 %r83, %r119, %r122, %r109;
prmt.b32 %r87, %r119, %r122, %r110;

dp4a.s32.s32 %r81, %r82, %r83, %r77;


dp4a.s32.s32 %r85, %r86, %r87, %r81;

ld.global.nc.u16 %rs5, [%rd24+8];
cvt.s32.s16 %r123, %rs5;
ld.global.nc.u32 %r90, [%rd27+28];
ld.global.nc.u32 %r94, [%rd27+32];
and.b32 %r124, %r123, 30583;
prmt.b32 %r125, %r104, %r104, %r124;
shr.u32 %r126, %r123, 2;
and.b32 %r127, %r126, 30583;
prmt.b32 %r128, %r104, %r104, %r127;
prmt.b32 %r91, %r125, %r128, %r109;
prmt.b32 %r95, %r125, %r128, %r110;

dp4a.s32.s32 %r89, %r90, %r91, %r85;


dp4a.s32.s32 %r93, %r94, %r95, %r89;

ld.global.nc.u32 %r97, [%rd27];

{.reg .f16 low,high;
mov.b32 {low,high},%r97;
cvt.f32.f16 %f12, low;}


mul.ftz.f32 %f13, %f11, %f12;
cvt.rn.f32.s32 %f14, %r93;
fma.rn.ftz.f32 %f42, %f13, %f14, 0f00000000;
add.s32 %r284, %r284, 64;

$L__BB18_6:
and.b32 %r129, %r16, -64;
setp.eq.s32 %p4, %r129, 0;
@%p4 bra $L__BB18_9;

add.s32 %r130, %r284, %r12;
add.s32 %r131, %r130, %r13;
add.s32 %r132, %r131, %r14;
mul.wide.s32 %rd28, %r132, 18;
add.s64 %rd42, %rd2, %rd28;
shl.b32 %r133, %r284, 1;
add.s32 %r283, %r133, 128;
shl.b64 %rd29, %rd4, 1;
add.s64 %rd6, %rd29, 4;

$L__BB18_8:
add.s32 %r200, %r283, -128;
cvt.u64.u32 %rd30, %r200;
ld.global.nc.u16 %rs6, [%rd42];

{ cvt.f32.f16 %f15, %rs6;}


add.s64 %rd31, %rd3, %rd30;
mul.lo.s64 %rd32, %rd31, 36;
add.s64 %rd33, %rd42, %rd6;
ld.global.nc.u16 %rs8, [%rd33+-2];
cvt.s32.s16 %r201, %rs8;
add.s64 %rd34, %rd1, %rd32;
ld.global.nc.u32 %r135, [%rd34+4];
ld.global.nc.u32 %r139, [%rd34+8];
and.b32 %r202, %r201, 30583;
mov.u32 %r203, 33620223;
prmt.b32 %r204, %r203, %r203, %r202;
shr.u32 %r205, %r201, 2;
and.b32 %r206, %r205, 30583;
prmt.b32 %r207, %r203, %r203, %r206;
mov.u32 %r208, 20800;
prmt.b32 %r136, %r204, %r207, %r208;
mov.u32 %r209, 29538;
prmt.b32 %r140, %r204, %r207, %r209;
mov.u32 %r170, 0;

dp4a.s32.s32 %r134, %r135, %r136, %r170;


dp4a.s32.s32 %r138, %r139, %r140, %r134;

ld.global.nc.u16 %rs9, [%rd33];
cvt.s32.s16 %r210, %rs9;
ld.global.nc.u32 %r143, [%rd34+12];
ld.global.nc.u32 %r147, [%rd34+16];
and.b32 %r211, %r210, 30583;
prmt.b32 %r212, %r203, %r203, %r211;
shr.u32 %r213, %r210, 2;
and.b32 %r214, %r213, 30583;
prmt.b32 %r215, %r203, %r203, %r214;
prmt.b32 %r144, %r212, %r215, %r208;
prmt.b32 %r148, %r212, %r215, %r209;

dp4a.s32.s32 %r142, %r143, %r144, %r138;


dp4a.s32.s32 %r146, %r147, %r148, %r142;

ld.global.nc.u16 %rs10, [%rd33+2];
cvt.s32.s16 %r216, %rs10;
ld.global.nc.u32 %r151, [%rd34+20];
ld.global.nc.u32 %r155, [%rd34+24];
and.b32 %r217, %r216, 30583;
prmt.b32 %r218, %r203, %r203, %r217;
shr.u32 %r219, %r216, 2;
and.b32 %r220, %r219, 30583;
prmt.b32 %r221, %r203, %r203, %r220;
prmt.b32 %r152, %r218, %r221, %r208;
prmt.b32 %r156, %r218, %r221, %r209;

dp4a.s32.s32 %r150, %r151, %r152, %r146;


dp4a.s32.s32 %r154, %r155, %r156, %r150;

ld.global.nc.u16 %rs11, [%rd33+4];
cvt.s32.s16 %r222, %rs11;
ld.global.nc.u32 %r159, [%rd34+28];
ld.global.nc.u32 %r163, [%rd34+32];
and.b32 %r223, %r222, 30583;
prmt.b32 %r224, %r203, %r203, %r223;
shr.u32 %r225, %r222, 2;
and.b32 %r226, %r225, 30583;
prmt.b32 %r227, %r203, %r203, %r226;
prmt.b32 %r160, %r224, %r227, %r208;
prmt.b32 %r164, %r224, %r227, %r209;

dp4a.s32.s32 %r158, %r159, %r160, %r154;


dp4a.s32.s32 %r162, %r163, %r164, %r158;

ld.global.nc.u32 %r166, [%rd34];

{.reg .f16 low,high;
mov.b32 {low,high},%r166;
cvt.f32.f16 %f16, low;}


mul.ftz.f32 %f19, %f15, %f16;
cvt.rn.f32.s32 %f20, %r162;
fma.rn.ftz.f32 %f21, %f19, %f20, %f42;
ld.global.nc.u16 %rs7, [%rd42+1152];

{ cvt.f32.f16 %f17, %rs7;}


cvt.u64.u32 %rd35, %r283;
add.s64 %rd36, %rd3, %rd35;
mul.lo.s64 %rd37, %rd36, 36;
ld.global.nc.u16 %rs12, [%rd33+1150];
cvt.s32.s16 %r228, %rs12;
add.s64 %rd38, %rd1, %rd37;
ld.global.nc.u32 %r168, [%rd38+4];
ld.global.nc.u32 %r172, [%rd38+8];
and.b32 %r229, %r228, 30583;
prmt.b32 %r230, %r203, %r203, %r229;
shr.u32 %r231, %r228, 2;
and.b32 %r232, %r231, 30583;
prmt.b32 %r233, %r203, %r203, %r232;
prmt.b32 %r169, %r230, %r233, %r208;
prmt.b32 %r173, %r230, %r233, %r209;

dp4a.s32.s32 %r167, %r168, %r169, %r170;


dp4a.s32.s32 %r171, %r172, %r173, %r167;

ld.global.nc.u16 %rs13, [%rd33+1152];
cvt.s32.s16 %r234, %rs13;
ld.global.nc.u32 %r176, [%rd38+12];
ld.global.nc.u32 %r180, [%rd38+16];
and.b32 %r235, %r234, 30583;
prmt.b32 %r236, %r203, %r203, %r235;
shr.u32 %r237, %r234, 2;
and.b32 %r238, %r237, 30583;
prmt.b32 %r239, %r203, %r203, %r238;
prmt.b32 %r177, %r236, %r239, %r208;
prmt.b32 %r181, %r236, %r239, %r209;

dp4a.s32.s32 %r175, %r176, %r177, %r171;


dp4a.s32.s32 %r179, %r180, %r181, %r175;

ld.global.nc.u16 %rs14, [%rd33+1154];
cvt.s32.s16 %r240, %rs14;
ld.global.nc.u32 %r184, [%rd38+20];
ld.global.nc.u32 %r188, [%rd38+24];
and.b32 %r241, %r240, 30583;
prmt.b32 %r242, %r203, %r203, %r241;
shr.u32 %r243, %r240, 2;
and.b32 %r244, %r243, 30583;
prmt.b32 %r245, %r203, %r203, %r244;
prmt.b32 %r185, %r242, %r245, %r208;
prmt.b32 %r189, %r242, %r245, %r209;

dp4a.s32.s32 %r183, %r184, %r185, %r179;


dp4a.s32.s32 %r187, %r188, %r189, %r183;

ld.global.nc.u16 %rs15, [%rd33+1156];
cvt.s32.s16 %r246, %rs15;
ld.global.nc.u32 %r192, [%rd38+28];
ld.global.nc.u32 %r196, [%rd38+32];
and.b32 %r247, %r246, 30583;
prmt.b32 %r248, %r203, %r203, %r247;
shr.u32 %r249, %r246, 2;
and.b32 %r250, %r249, 30583;
prmt.b32 %r251, %r203, %r203, %r250;
prmt.b32 %r193, %r248, %r251, %r208;
prmt.b32 %r197, %r248, %r251, %r209;

dp4a.s32.s32 %r191, %r192, %r193, %r187;


dp4a.s32.s32 %r195, %r196, %r197, %r191;

ld.global.nc.u32 %r199, [%rd38];

{.reg .f16 low,high;
mov.b32 {low,high},%r199;
cvt.f32.f16 %f18, low;}


mul.ftz.f32 %f22, %f17, %f18;
cvt.rn.f32.s32 %f23, %r195;
fma.rn.ftz.f32 %f42, %f22, %f23, %f21;
add.s64 %rd42, %rd42, 2304;
add.s32 %r283, %r283, 256;
add.s32 %r284, %r284, 128;
setp.lt.s32 %p5, %r284, %r4;
@%p5 bra $L__BB18_8;

$L__BB18_9:
setp.eq.s32 %p6, %r1, 0;
@%p6 bra $L__BB18_11;

shl.b32 %r252, %r1, 7;
mov.u32 %r253, _ZZ13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r254, %r253, %r252;
shl.b32 %r255, %r2, 2;
add.s32 %r256, %r254, %r255;
st.shared.f32 [%r256+-128], %f42;

$L__BB18_11:
setp.ne.s32 %p7, %r1, 0;
bar.sync 0;
@%p7 bra $L__BB18_14;

mov.u32 %r257, %ctaid.y;
mad.lo.s32 %r259, %r257, %r36, %r53;
mad.lo.s32 %r25, %r11, %r42, %r259;
shl.b32 %r260, %r2, 2;
mov.u32 %r261, _ZZ13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r262, %r261, %r260;
mov.u32 %r263, 2;
ld.shared.f32 %f24, [%r262];
add.ftz.f32 %f25, %f42, %f24;
ld.shared.f32 %f26, [%r262+128];
add.ftz.f32 %f27, %f25, %f26;
ld.shared.f32 %f28, [%r262+256];
add.ftz.f32 %f29, %f27, %f28;
mov.b32 %r264, %f29;
mov.u32 %r265, 31;
mov.u32 %r266, 16;
mov.u32 %r267, -1;
shfl.sync.bfly.b32 %r268|%p8, %r264, %r266, %r265, %r267;
mov.b32 %f30, %r268;
add.ftz.f32 %f31, %f29, %f30;
mov.b32 %r269, %f31;
mov.u32 %r270, 8;
shfl.sync.bfly.b32 %r271|%p9, %r269, %r270, %r265, %r267;
mov.b32 %f32, %r271;
add.ftz.f32 %f33, %f31, %f32;
mov.b32 %r272, %f33;
mov.u32 %r273, 4;
shfl.sync.bfly.b32 %r274|%p10, %r272, %r273, %r265, %r267;
mov.b32 %f34, %r274;
add.ftz.f32 %f35, %f33, %f34;
mov.b32 %r275, %f35;
shfl.sync.bfly.b32 %r276|%p11, %r275, %r263, %r265, %r267;
mov.b32 %f36, %r276;
add.ftz.f32 %f37, %f35, %f36;
mov.b32 %r277, %f37;
mov.u32 %r278, 1;
shfl.sync.bfly.b32 %r279|%p12, %r277, %r278, %r265, %r267;
mov.b32 %f38, %r279;
add.ftz.f32 %f7, %f37, %f38;
setp.ne.s32 %p13, %r2, 0;
@%p13 bra $L__BB18_14;

cvta.to.global.u64 %rd39, %rd10;
mul.wide.u32 %rd40, %r25, 4;
add.s64 %rd41, %rd39, %rd40;
st.global.f32 [%rd41], %f7;

$L__BB18_14:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type42ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type42ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type42ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type42ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type42ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type42ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type42ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type42ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type42ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<36>;
.reg .b16 %rs<13>;
.reg .f32 %f<109>;
.reg .b32 %r<350>;
.reg .b64 %rd<56>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type42ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type42ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type42ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type42ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r13, [_Z13mul_mat_vec_qIL9ggml_type42ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r14, [_Z13mul_mat_vec_qIL9ggml_type42ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r15, [_Z13mul_mat_vec_qIL9ggml_type42ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r16, [_Z13mul_mat_vec_qIL9ggml_type42ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type42ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type42ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type42ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type42ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type42ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type42ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type42ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type42ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r18, [_Z13mul_mat_vec_qIL9ggml_type42ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r17, [_Z13mul_mat_vec_qIL9ggml_type42ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r29, %r1, 5;
mov.u32 %r30, %tid.x;
mov.u32 %r31, %ctaid.x;
shl.b32 %r2, %r31, 1;
add.s32 %r32, %r29, %r30;
shr.u32 %r33, %r32, 31;
add.s32 %r34, %r32, %r33;
shr.s32 %r349, %r34, 1;
shr.u32 %r35, %r13, 6;
setp.ge.s32 %p2, %r349, %r35;
mov.f32 %f105, 0f00000000;
mov.f32 %f106, %f105;
mov.f32 %f107, %f105;
mov.f32 %f108, %f105;
@%p2 bra $L__BB19_3;

mov.u32 %r36, %ctaid.y;
mov.u32 %r37, %ctaid.z;
mul.hi.u32 %r38, %r37, %r23;
add.s32 %r39, %r38, %r37;
shr.u32 %r40, %r39, %r24;
mul.lo.s32 %r41, %r40, %r26;
mul.hi.u32 %r42, %r36, %r17;
add.s32 %r43, %r42, %r36;
shr.u32 %r44, %r43, %r18;
mul.lo.s32 %r45, %r44, %r20;
mul.lo.s32 %r46, %r36, %r21;
cvt.u64.u32 %rd20, %r46;
mul.lo.s32 %r47, %r37, %r27;
cvt.u64.u32 %rd21, %r47;
and.b32 %r54, %r34, -2;
sub.s32 %r55, %r32, %r54;
cvt.s64.s32 %rd22, %r55;
add.s64 %rd23, %rd20, %rd22;
add.s64 %rd1, %rd23, %rd21;
shl.b32 %r56, %r55, 2;
add.s32 %r57, %r2, 1;
shl.b32 %r347, %r349, 1;
add.s32 %r348, %r15, %r347;
mad.lo.s32 %r58, %r2, %r14, %r349;
add.s32 %r59, %r58, %r41;
add.s32 %r60, %r59, %r45;
mul.wide.s32 %rd24, %r60, 18;
mul.wide.s32 %rd25, %r56, 2;
add.s64 %rd26, %rd24, %rd25;
cvta.to.global.u64 %rd27, %rd17;
add.s64 %rd28, %rd27, %rd26;
add.s64 %rd55, %rd28, 8;
add.s64 %rd54, %rd27, %rd24;
add.s32 %r61, %r349, %r41;
add.s32 %r62, %r61, %r45;
mad.lo.s32 %r63, %r57, %r14, %r62;
mul.wide.s32 %rd29, %r63, 18;
add.s64 %rd30, %rd29, %rd25;
add.s64 %rd31, %rd27, %rd30;
add.s64 %rd53, %rd31, 4;
add.s64 %rd52, %rd27, %rd29;
cvta.to.global.u64 %rd6, %rd18;
mov.f32 %f105, 0f00000000;

$L__BB19_2:
cvt.u64.u32 %rd32, %r347;
add.s64 %rd33, %rd1, %rd32;
mul.lo.s64 %rd34, %rd33, 36;
ld.global.nc.u16 %rs1, [%rd54];

{ cvt.f32.f16 %f25, %rs1;}


ld.global.nc.u16 %rs5, [%rd55+-6];
cvt.s32.s16 %r196, %rs5;
add.s64 %rd35, %rd6, %rd34;
ld.global.nc.u32 %r65, [%rd35+4];
ld.global.nc.u32 %r69, [%rd35+8];
and.b32 %r197, %r196, 30583;
mov.u32 %r198, 33620223;
prmt.b32 %r199, %r198, %r198, %r197;
shr.u32 %r200, %r196, 2;
and.b32 %r201, %r200, 30583;
prmt.b32 %r202, %r198, %r198, %r201;
mov.u32 %r203, 20800;
prmt.b32 %r132, %r199, %r202, %r203;
mov.u32 %r204, 29538;
prmt.b32 %r136, %r199, %r202, %r204;
mov.u32 %r166, 0;

dp4a.s32.s32 %r64, %r65, %r132, %r166;


dp4a.s32.s32 %r68, %r69, %r136, %r64;

ld.global.nc.u16 %rs6, [%rd55+-4];
cvt.s32.s16 %r205, %rs6;
ld.global.nc.u32 %r73, [%rd35+12];
ld.global.nc.u32 %r77, [%rd35+16];
and.b32 %r206, %r205, 30583;
prmt.b32 %r207, %r198, %r198, %r206;
shr.u32 %r208, %r205, 2;
and.b32 %r209, %r208, 30583;
prmt.b32 %r210, %r198, %r198, %r209;
prmt.b32 %r140, %r207, %r210, %r203;
prmt.b32 %r144, %r207, %r210, %r204;

dp4a.s32.s32 %r72, %r73, %r140, %r68;


dp4a.s32.s32 %r76, %r77, %r144, %r72;

ld.global.nc.u16 %rs7, [%rd55+-2];
cvt.s32.s16 %r211, %rs7;
ld.global.nc.u32 %r81, [%rd35+20];
ld.global.nc.u32 %r85, [%rd35+24];
and.b32 %r212, %r211, 30583;
prmt.b32 %r213, %r198, %r198, %r212;
shr.u32 %r214, %r211, 2;
and.b32 %r215, %r214, 30583;
prmt.b32 %r216, %r198, %r198, %r215;
prmt.b32 %r148, %r213, %r216, %r203;
prmt.b32 %r152, %r213, %r216, %r204;

dp4a.s32.s32 %r80, %r81, %r148, %r76;


dp4a.s32.s32 %r84, %r85, %r152, %r80;

ld.global.nc.u16 %rs8, [%rd55];
cvt.s32.s16 %r217, %rs8;
ld.global.nc.u32 %r89, [%rd35+28];
ld.global.nc.u32 %r93, [%rd35+32];
and.b32 %r218, %r217, 30583;
prmt.b32 %r219, %r198, %r198, %r218;
shr.u32 %r220, %r217, 2;
and.b32 %r221, %r220, 30583;
prmt.b32 %r222, %r198, %r198, %r221;
prmt.b32 %r156, %r219, %r222, %r203;
prmt.b32 %r160, %r219, %r222, %r204;

dp4a.s32.s32 %r88, %r89, %r156, %r84;


dp4a.s32.s32 %r92, %r93, %r160, %r88;

ld.global.nc.u32 %r96, [%rd35];

{.reg .f16 low,high;
mov.b32 {low,high},%r96;
cvt.f32.f16 %f26, low;}


mul.ftz.f32 %f33, %f25, %f26;
cvt.rn.f32.s32 %f34, %r92;
fma.rn.ftz.f32 %f108, %f33, %f34, %f108;
ld.global.nc.u16 %rs2, [%rd52];

{ cvt.f32.f16 %f27, %rs2;}


ld.global.nc.u16 %rs9, [%rd53+-2];
cvt.s32.s16 %r223, %rs9;
and.b32 %r224, %r223, 30583;
prmt.b32 %r225, %r198, %r198, %r224;
shr.u32 %r226, %r223, 2;
and.b32 %r227, %r226, 30583;
prmt.b32 %r228, %r198, %r198, %r227;
prmt.b32 %r165, %r225, %r228, %r203;
prmt.b32 %r169, %r225, %r228, %r204;

dp4a.s32.s32 %r97, %r65, %r165, %r166;


dp4a.s32.s32 %r101, %r69, %r169, %r97;

ld.global.nc.u16 %rs10, [%rd53];
cvt.s32.s16 %r229, %rs10;
and.b32 %r230, %r229, 30583;
prmt.b32 %r231, %r198, %r198, %r230;
shr.u32 %r232, %r229, 2;
and.b32 %r233, %r232, 30583;
prmt.b32 %r234, %r198, %r198, %r233;
prmt.b32 %r173, %r231, %r234, %r203;
prmt.b32 %r177, %r231, %r234, %r204;

dp4a.s32.s32 %r105, %r73, %r173, %r101;


dp4a.s32.s32 %r109, %r77, %r177, %r105;

ld.global.nc.u16 %rs11, [%rd53+2];
cvt.s32.s16 %r235, %rs11;
and.b32 %r236, %r235, 30583;
prmt.b32 %r237, %r198, %r198, %r236;
shr.u32 %r238, %r235, 2;
and.b32 %r239, %r238, 30583;
prmt.b32 %r240, %r198, %r198, %r239;
prmt.b32 %r181, %r237, %r240, %r203;
prmt.b32 %r185, %r237, %r240, %r204;

dp4a.s32.s32 %r113, %r81, %r181, %r109;


dp4a.s32.s32 %r117, %r85, %r185, %r113;

ld.global.nc.u16 %rs12, [%rd53+4];
cvt.s32.s16 %r241, %rs12;
and.b32 %r242, %r241, 30583;
prmt.b32 %r243, %r198, %r198, %r242;
shr.u32 %r244, %r241, 2;
and.b32 %r245, %r244, 30583;
prmt.b32 %r246, %r198, %r198, %r245;
prmt.b32 %r189, %r243, %r246, %r203;
prmt.b32 %r193, %r243, %r246, %r204;

dp4a.s32.s32 %r121, %r89, %r189, %r117;


dp4a.s32.s32 %r125, %r93, %r193, %r121;


{.reg .f16 low,high;
mov.b32 {low,high},%r96;
cvt.f32.f16 %f28, low;}


mul.ftz.f32 %f35, %f27, %f28;
cvt.rn.f32.s32 %f36, %r125;
fma.rn.ftz.f32 %f107, %f35, %f36, %f107;
cvt.u64.u32 %rd36, %r348;
add.s64 %rd37, %rd1, %rd36;
mul.lo.s64 %rd38, %rd37, 36;

{ cvt.f32.f16 %f29, %rs1;}


add.s64 %rd39, %rd6, %rd38;
ld.global.nc.u32 %r131, [%rd39+4];
ld.global.nc.u32 %r135, [%rd39+8];

dp4a.s32.s32 %r130, %r131, %r132, %r166;


dp4a.s32.s32 %r134, %r135, %r136, %r130;

ld.global.nc.u32 %r139, [%rd39+12];
ld.global.nc.u32 %r143, [%rd39+16];

dp4a.s32.s32 %r138, %r139, %r140, %r134;


dp4a.s32.s32 %r142, %r143, %r144, %r138;

ld.global.nc.u32 %r147, [%rd39+20];
ld.global.nc.u32 %r151, [%rd39+24];

dp4a.s32.s32 %r146, %r147, %r148, %r142;


dp4a.s32.s32 %r150, %r151, %r152, %r146;

ld.global.nc.u32 %r155, [%rd39+28];
ld.global.nc.u32 %r159, [%rd39+32];

dp4a.s32.s32 %r154, %r155, %r156, %r150;


dp4a.s32.s32 %r158, %r159, %r160, %r154;

ld.global.nc.u32 %r162, [%rd39];

{.reg .f16 low,high;
mov.b32 {low,high},%r162;
cvt.f32.f16 %f30, low;}


mul.ftz.f32 %f37, %f29, %f30;
cvt.rn.f32.s32 %f38, %r158;
fma.rn.ftz.f32 %f106, %f37, %f38, %f106;

{ cvt.f32.f16 %f31, %rs2;}



dp4a.s32.s32 %r163, %r131, %r165, %r166;


dp4a.s32.s32 %r167, %r135, %r169, %r163;


dp4a.s32.s32 %r171, %r139, %r173, %r167;


dp4a.s32.s32 %r175, %r143, %r177, %r171;


dp4a.s32.s32 %r179, %r147, %r181, %r175;


dp4a.s32.s32 %r183, %r151, %r185, %r179;


dp4a.s32.s32 %r187, %r155, %r189, %r183;


dp4a.s32.s32 %r191, %r159, %r193, %r187;


{.reg .f16 low,high;
mov.b32 {low,high},%r162;
cvt.f32.f16 %f32, low;}


mul.ftz.f32 %f39, %f31, %f32;
cvt.rn.f32.s32 %f40, %r191;
fma.rn.ftz.f32 %f105, %f39, %f40, %f105;
add.s32 %r348, %r348, 128;
add.s32 %r347, %r347, 128;
add.s64 %rd55, %rd55, 1152;
add.s64 %rd54, %rd54, 1152;
add.s64 %rd53, %rd53, 1152;
add.s64 %rd52, %rd52, 1152;
add.s32 %r349, %r349, 64;
setp.lt.s32 %p3, %r349, %r35;
@%p3 bra $L__BB19_2;

$L__BB19_3:
setp.eq.s32 %p4, %r1, 0;
@%p4 bra $L__BB19_5;

shl.b32 %r248, %r1, 9;
mov.u32 %r249, _ZZ13mul_mat_vec_qIL9ggml_type42ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r250, %r249, %r248;
shl.b32 %r251, %r30, 2;
add.s32 %r252, %r250, %r251;
st.shared.f32 [%r252+-512], %f108;
st.shared.f32 [%r252+-384], %f107;
st.shared.f32 [%r252+-256], %f106;
st.shared.f32 [%r252+-128], %f105;

$L__BB19_5:
setp.ne.s32 %p5, %r1, 0;
bar.sync 0;
@%p5 bra $L__BB19_14;

mov.u32 %r253, %ctaid.z;
mov.u32 %r254, %ctaid.y;
mad.lo.s32 %r255, %r254, %r22, %r2;
mad.lo.s32 %r256, %r253, %r28, %r255;
cvt.u64.u32 %rd15, %r256;
shl.b32 %r258, %r30, 2;
mov.u32 %r259, _ZZ13mul_mat_vec_qIL9ggml_type42ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r260, %r259, %r258;
mov.u32 %r261, 2;
ld.shared.f32 %f41, [%r260];
add.ftz.f32 %f42, %f41, %f108;
ld.shared.f32 %f43, [%r260+512];
add.ftz.f32 %f44, %f43, %f42;
ld.shared.f32 %f45, [%r260+1024];
add.ftz.f32 %f46, %f45, %f44;
mov.b32 %r262, %f46;
mov.u32 %r263, 31;
mov.u32 %r264, 16;
mov.u32 %r265, -1;
shfl.sync.bfly.b32 %r266|%p6, %r262, %r264, %r263, %r265;
mov.b32 %f47, %r266;
add.ftz.f32 %f48, %f46, %f47;
mov.b32 %r267, %f48;
mov.u32 %r268, 8;
shfl.sync.bfly.b32 %r269|%p7, %r267, %r268, %r263, %r265;
mov.b32 %f49, %r269;
add.ftz.f32 %f50, %f48, %f49;
mov.b32 %r270, %f50;
mov.u32 %r271, 4;
shfl.sync.bfly.b32 %r272|%p8, %r270, %r271, %r263, %r265;
mov.b32 %f51, %r272;
add.ftz.f32 %f52, %f50, %f51;
mov.b32 %r273, %f52;
shfl.sync.bfly.b32 %r274|%p9, %r273, %r261, %r263, %r265;
mov.b32 %f53, %r274;
add.ftz.f32 %f54, %f52, %f53;
mov.b32 %r275, %f54;
mov.u32 %r276, 1;
shfl.sync.bfly.b32 %r277|%p10, %r275, %r276, %r263, %r265;
mov.b32 %f55, %r277;
add.ftz.f32 %f13, %f54, %f55;
setp.ne.s32 %p11, %r30, 0;
setp.ge.u32 %p12, %r2, %r16;
cvta.to.global.u64 %rd40, %rd19;
mul.wide.u32 %rd41, %r256, 4;
add.s64 %rd16, %rd40, %rd41;
or.pred %p13, %p11, %p12;
@%p13 bra $L__BB19_8;

st.global.f32 [%rd16], %f13;

$L__BB19_8:
ld.shared.f32 %f56, [%r260+128];
add.ftz.f32 %f57, %f56, %f107;
ld.shared.f32 %f58, [%r260+640];
add.ftz.f32 %f59, %f58, %f57;
ld.shared.f32 %f60, [%r260+1152];
add.ftz.f32 %f61, %f60, %f59;
mov.b32 %r283, %f61;
shfl.sync.bfly.b32 %r287|%p14, %r283, %r264, %r263, %r265;
mov.b32 %f62, %r287;
add.ftz.f32 %f63, %f61, %f62;
mov.b32 %r288, %f63;
shfl.sync.bfly.b32 %r290|%p15, %r288, %r268, %r263, %r265;
mov.b32 %f64, %r290;
add.ftz.f32 %f65, %f63, %f64;
mov.b32 %r291, %f65;
shfl.sync.bfly.b32 %r293|%p16, %r291, %r271, %r263, %r265;
mov.b32 %f66, %r293;
add.ftz.f32 %f67, %f65, %f66;
mov.b32 %r294, %f67;
shfl.sync.bfly.b32 %r295|%p17, %r294, %r261, %r263, %r265;
mov.b32 %f68, %r295;
add.ftz.f32 %f69, %f67, %f68;
mov.b32 %r296, %f69;
shfl.sync.bfly.b32 %r298|%p18, %r296, %r276, %r263, %r265;
mov.b32 %f70, %r298;
add.ftz.f32 %f14, %f69, %f70;
setp.eq.s32 %p19, %r30, 1;
or.b32 %r301, %r2, 1;
setp.lt.u32 %p20, %r301, %r16;
and.pred %p1, %p19, %p20;
not.pred %p21, %p1;
@%p21 bra $L__BB19_10;

st.global.f32 [%rd16+4], %f14;

$L__BB19_10:
mov.u32 %r306, 2;
ld.shared.f32 %f71, [%r260+256];
add.ftz.f32 %f72, %f71, %f106;
ld.shared.f32 %f73, [%r260+768];
add.ftz.f32 %f74, %f73, %f72;
ld.shared.f32 %f75, [%r260+1280];
add.ftz.f32 %f76, %f75, %f74;
mov.b32 %r307, %f76;
mov.u32 %r308, 31;
mov.u32 %r309, 16;
mov.u32 %r310, -1;
shfl.sync.bfly.b32 %r311|%p22, %r307, %r309, %r308, %r310;
mov.b32 %f77, %r311;
add.ftz.f32 %f78, %f76, %f77;
mov.b32 %r312, %f78;
mov.u32 %r313, 8;
shfl.sync.bfly.b32 %r314|%p23, %r312, %r313, %r308, %r310;
mov.b32 %f79, %r314;
add.ftz.f32 %f80, %f78, %f79;
mov.b32 %r315, %f80;
mov.u32 %r316, 4;
shfl.sync.bfly.b32 %r317|%p24, %r315, %r316, %r308, %r310;
mov.b32 %f81, %r317;
add.ftz.f32 %f82, %f80, %f81;
mov.b32 %r318, %f82;
shfl.sync.bfly.b32 %r319|%p25, %r318, %r306, %r308, %r310;
mov.b32 %f83, %r319;
add.ftz.f32 %f84, %f82, %f83;
mov.b32 %r320, %f84;
mov.u32 %r321, 1;
shfl.sync.bfly.b32 %r322|%p26, %r320, %r321, %r308, %r310;
mov.b32 %f85, %r322;
add.ftz.f32 %f15, %f84, %f85;
@%p13 bra $L__BB19_12;

cvt.u64.u32 %rd42, %r16;
add.s64 %rd43, %rd42, %rd15;
shl.b64 %rd45, %rd43, 2;
add.s64 %rd46, %rd40, %rd45;
st.global.f32 [%rd46], %f15;

$L__BB19_12:
ld.shared.f32 %f86, [%r260+384];
add.ftz.f32 %f87, %f86, %f105;
ld.shared.f32 %f88, [%r260+896];
add.ftz.f32 %f89, %f88, %f87;
ld.shared.f32 %f90, [%r260+1408];
add.ftz.f32 %f91, %f90, %f89;
mov.b32 %r330, %f91;
shfl.sync.bfly.b32 %r334|%p30, %r330, %r309, %r308, %r310;
mov.b32 %f92, %r334;
add.ftz.f32 %f93, %f91, %f92;
mov.b32 %r335, %f93;
shfl.sync.bfly.b32 %r337|%p31, %r335, %r313, %r308, %r310;
mov.b32 %f94, %r337;
add.ftz.f32 %f95, %f93, %f94;
mov.b32 %r338, %f95;
shfl.sync.bfly.b32 %r340|%p32, %r338, %r316, %r308, %r310;
mov.b32 %f96, %r340;
add.ftz.f32 %f97, %f95, %f96;
mov.b32 %r341, %f97;
shfl.sync.bfly.b32 %r342|%p33, %r341, %r306, %r308, %r310;
mov.b32 %f98, %r342;
add.ftz.f32 %f99, %f97, %f98;
mov.b32 %r343, %f99;
shfl.sync.bfly.b32 %r345|%p34, %r343, %r321, %r308, %r310;
mov.b32 %f100, %r345;
add.ftz.f32 %f16, %f99, %f100;
@%p21 bra $L__BB19_14;

add.s32 %r346, %r16, 1;
cvt.u64.u32 %rd47, %r346;
add.s64 %rd48, %rd47, %rd15;
shl.b64 %rd50, %rd48, 2;
add.s64 %rd51, %rd40, %rd50;
st.global.f32 [%rd51], %f16;

$L__BB19_14:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type42ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type42ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type42ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type42ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type42ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type42ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type42ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type42ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type42ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<50>;
.reg .b16 %rs<15>;
.reg .f32 %f<163>;
.reg .b32 %r<493>;
.reg .b64 %rd<70>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type42ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[2304];

ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type42ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type42ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type42ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r15, [_Z13mul_mat_vec_qIL9ggml_type42ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r16, [_Z13mul_mat_vec_qIL9ggml_type42ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r17, [_Z13mul_mat_vec_qIL9ggml_type42ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r18, [_Z13mul_mat_vec_qIL9ggml_type42ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type42ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type42ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type42ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type42ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type42ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type42ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type42ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type42ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type42ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r19, [_Z13mul_mat_vec_qIL9ggml_type42ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r31, %r1, 5;
mov.u32 %r32, %tid.x;
add.s32 %r33, %r31, %r32;
shr.u32 %r34, %r33, 31;
add.s32 %r35, %r33, %r34;
shr.s32 %r492, %r35, 1;
shr.u32 %r36, %r15, 6;
setp.ge.s32 %p2, %r492, %r36;
mov.f32 %f157, 0f00000000;
mov.f32 %f158, %f157;
mov.f32 %f159, %f157;
mov.f32 %f160, %f157;
mov.f32 %f161, %f157;
mov.f32 %f162, %f157;
@%p2 bra $L__BB20_3;

mov.u32 %r37, %ctaid.x;
shl.b32 %r38, %r37, 1;
mov.u32 %r39, %ctaid.y;
mov.u32 %r40, %ctaid.z;
mul.hi.u32 %r41, %r40, %r25;
add.s32 %r42, %r41, %r40;
shr.u32 %r43, %r42, %r26;
mul.lo.s32 %r44, %r43, %r28;
mul.hi.u32 %r45, %r39, %r19;
add.s32 %r46, %r45, %r39;
shr.u32 %r47, %r46, %r20;
mul.lo.s32 %r48, %r47, %r22;
mul.lo.s32 %r49, %r39, %r23;
cvt.u64.u32 %rd19, %r49;
mul.lo.s32 %r50, %r40, %r29;
cvt.u64.u32 %rd20, %r50;
and.b32 %r57, %r35, -2;
sub.s32 %r58, %r33, %r57;
cvt.s64.s32 %rd21, %r58;
add.s64 %rd22, %rd19, %rd21;
add.s64 %rd1, %rd22, %rd20;
shl.b32 %r59, %r58, 2;
or.b32 %r60, %r38, 1;
shl.b32 %r61, %r17, 1;
shl.b32 %r489, %r492, 1;
add.s32 %r491, %r489, %r61;
add.s32 %r490, %r17, %r489;
mad.lo.s32 %r62, %r38, %r16, %r492;
add.s32 %r63, %r62, %r44;
add.s32 %r64, %r63, %r48;
mul.wide.s32 %rd23, %r64, 18;
mul.wide.s32 %rd24, %r59, 2;
add.s64 %rd25, %rd23, %rd24;
cvta.to.global.u64 %rd26, %rd16;
add.s64 %rd27, %rd26, %rd25;
add.s64 %rd69, %rd27, 8;
add.s64 %rd68, %rd26, %rd23;
add.s32 %r65, %r492, %r44;
add.s32 %r66, %r65, %r48;
mad.lo.s32 %r67, %r60, %r16, %r66;
mul.wide.s32 %rd28, %r67, 18;
add.s64 %rd29, %rd28, %rd24;
add.s64 %rd30, %rd26, %rd29;
add.s64 %rd67, %rd30, 4;
add.s64 %rd66, %rd26, %rd28;
cvta.to.global.u64 %rd6, %rd17;
mov.f32 %f157, 0f00000000;

$L__BB20_2:
cvt.u64.u32 %rd31, %r489;
add.s64 %rd32, %rd1, %rd31;
mul.lo.s64 %rd33, %rd32, 36;
ld.global.nc.u16 %rs1, [%rd68];

{ cvt.f32.f16 %f37, %rs1;}


ld.global.nc.u16 %rs7, [%rd69+-6];
cvt.s32.s16 %r266, %rs7;
add.s64 %rd34, %rd6, %rd33;
ld.global.nc.u32 %r69, [%rd34+4];
ld.global.nc.u32 %r73, [%rd34+8];
and.b32 %r267, %r266, 30583;
mov.u32 %r268, 33620223;
prmt.b32 %r269, %r268, %r268, %r267;
shr.u32 %r270, %r266, 2;
and.b32 %r271, %r270, 30583;
prmt.b32 %r272, %r268, %r268, %r271;
mov.u32 %r273, 20800;
prmt.b32 %r202, %r269, %r272, %r273;
mov.u32 %r274, 29538;
prmt.b32 %r206, %r269, %r272, %r274;
mov.u32 %r236, 0;

dp4a.s32.s32 %r68, %r69, %r202, %r236;


dp4a.s32.s32 %r72, %r73, %r206, %r68;

ld.global.nc.u16 %rs8, [%rd69+-4];
cvt.s32.s16 %r275, %rs8;
ld.global.nc.u32 %r77, [%rd34+12];
ld.global.nc.u32 %r81, [%rd34+16];
and.b32 %r276, %r275, 30583;
prmt.b32 %r277, %r268, %r268, %r276;
shr.u32 %r278, %r275, 2;
and.b32 %r279, %r278, 30583;
prmt.b32 %r280, %r268, %r268, %r279;
prmt.b32 %r210, %r277, %r280, %r273;
prmt.b32 %r214, %r277, %r280, %r274;

dp4a.s32.s32 %r76, %r77, %r210, %r72;


dp4a.s32.s32 %r80, %r81, %r214, %r76;

ld.global.nc.u16 %rs9, [%rd69+-2];
cvt.s32.s16 %r281, %rs9;
ld.global.nc.u32 %r85, [%rd34+20];
ld.global.nc.u32 %r89, [%rd34+24];
and.b32 %r282, %r281, 30583;
prmt.b32 %r283, %r268, %r268, %r282;
shr.u32 %r284, %r281, 2;
and.b32 %r285, %r284, 30583;
prmt.b32 %r286, %r268, %r268, %r285;
prmt.b32 %r218, %r283, %r286, %r273;
prmt.b32 %r222, %r283, %r286, %r274;

dp4a.s32.s32 %r84, %r85, %r218, %r80;


dp4a.s32.s32 %r88, %r89, %r222, %r84;

ld.global.nc.u16 %rs10, [%rd69];
cvt.s32.s16 %r287, %rs10;
ld.global.nc.u32 %r93, [%rd34+28];
ld.global.nc.u32 %r97, [%rd34+32];
and.b32 %r288, %r287, 30583;
prmt.b32 %r289, %r268, %r268, %r288;
shr.u32 %r290, %r287, 2;
and.b32 %r291, %r290, 30583;
prmt.b32 %r292, %r268, %r268, %r291;
prmt.b32 %r226, %r289, %r292, %r273;
prmt.b32 %r230, %r289, %r292, %r274;

dp4a.s32.s32 %r92, %r93, %r226, %r88;


dp4a.s32.s32 %r96, %r97, %r230, %r92;

ld.global.nc.u32 %r100, [%rd34];

{.reg .f16 low,high;
mov.b32 {low,high},%r100;
cvt.f32.f16 %f38, low;}


mul.ftz.f32 %f49, %f37, %f38;
cvt.rn.f32.s32 %f50, %r96;
fma.rn.ftz.f32 %f162, %f49, %f50, %f162;
ld.global.nc.u16 %rs2, [%rd66];

{ cvt.f32.f16 %f39, %rs2;}


ld.global.nc.u16 %rs11, [%rd67+-2];
cvt.s32.s16 %r293, %rs11;
and.b32 %r294, %r293, 30583;
prmt.b32 %r295, %r268, %r268, %r294;
shr.u32 %r296, %r293, 2;
and.b32 %r297, %r296, 30583;
prmt.b32 %r298, %r268, %r268, %r297;
prmt.b32 %r235, %r295, %r298, %r273;
prmt.b32 %r239, %r295, %r298, %r274;

dp4a.s32.s32 %r101, %r69, %r235, %r236;


dp4a.s32.s32 %r105, %r73, %r239, %r101;

ld.global.nc.u16 %rs12, [%rd67];
cvt.s32.s16 %r299, %rs12;
and.b32 %r300, %r299, 30583;
prmt.b32 %r301, %r268, %r268, %r300;
shr.u32 %r302, %r299, 2;
and.b32 %r303, %r302, 30583;
prmt.b32 %r304, %r268, %r268, %r303;
prmt.b32 %r243, %r301, %r304, %r273;
prmt.b32 %r247, %r301, %r304, %r274;

dp4a.s32.s32 %r109, %r77, %r243, %r105;


dp4a.s32.s32 %r113, %r81, %r247, %r109;

ld.global.nc.u16 %rs13, [%rd67+2];
cvt.s32.s16 %r305, %rs13;
and.b32 %r306, %r305, 30583;
prmt.b32 %r307, %r268, %r268, %r306;
shr.u32 %r308, %r305, 2;
and.b32 %r309, %r308, 30583;
prmt.b32 %r310, %r268, %r268, %r309;
prmt.b32 %r251, %r307, %r310, %r273;
prmt.b32 %r255, %r307, %r310, %r274;

dp4a.s32.s32 %r117, %r85, %r251, %r113;


dp4a.s32.s32 %r121, %r89, %r255, %r117;

ld.global.nc.u16 %rs14, [%rd67+4];
cvt.s32.s16 %r311, %rs14;
and.b32 %r312, %r311, 30583;
prmt.b32 %r313, %r268, %r268, %r312;
shr.u32 %r314, %r311, 2;
and.b32 %r315, %r314, 30583;
prmt.b32 %r316, %r268, %r268, %r315;
prmt.b32 %r259, %r313, %r316, %r273;
prmt.b32 %r263, %r313, %r316, %r274;

dp4a.s32.s32 %r125, %r93, %r259, %r121;


dp4a.s32.s32 %r129, %r97, %r263, %r125;


{.reg .f16 low,high;
mov.b32 {low,high},%r100;
cvt.f32.f16 %f40, low;}


mul.ftz.f32 %f51, %f39, %f40;
cvt.rn.f32.s32 %f52, %r129;
fma.rn.ftz.f32 %f161, %f51, %f52, %f161;
cvt.u64.u32 %rd35, %r490;
add.s64 %rd36, %rd1, %rd35;
mul.lo.s64 %rd37, %rd36, 36;

{ cvt.f32.f16 %f41, %rs1;}


add.s64 %rd38, %rd6, %rd37;
ld.global.nc.u32 %r135, [%rd38+4];
ld.global.nc.u32 %r139, [%rd38+8];

dp4a.s32.s32 %r134, %r135, %r202, %r236;


dp4a.s32.s32 %r138, %r139, %r206, %r134;

ld.global.nc.u32 %r143, [%rd38+12];
ld.global.nc.u32 %r147, [%rd38+16];

dp4a.s32.s32 %r142, %r143, %r210, %r138;


dp4a.s32.s32 %r146, %r147, %r214, %r142;

ld.global.nc.u32 %r151, [%rd38+20];
ld.global.nc.u32 %r155, [%rd38+24];

dp4a.s32.s32 %r150, %r151, %r218, %r146;


dp4a.s32.s32 %r154, %r155, %r222, %r150;

ld.global.nc.u32 %r159, [%rd38+28];
ld.global.nc.u32 %r163, [%rd38+32];

dp4a.s32.s32 %r158, %r159, %r226, %r154;


dp4a.s32.s32 %r162, %r163, %r230, %r158;

ld.global.nc.u32 %r166, [%rd38];

{.reg .f16 low,high;
mov.b32 {low,high},%r166;
cvt.f32.f16 %f42, low;}


mul.ftz.f32 %f53, %f41, %f42;
cvt.rn.f32.s32 %f54, %r162;
fma.rn.ftz.f32 %f160, %f53, %f54, %f160;

{ cvt.f32.f16 %f43, %rs2;}



dp4a.s32.s32 %r167, %r135, %r235, %r236;


dp4a.s32.s32 %r171, %r139, %r239, %r167;


dp4a.s32.s32 %r175, %r143, %r243, %r171;


dp4a.s32.s32 %r179, %r147, %r247, %r175;


dp4a.s32.s32 %r183, %r151, %r251, %r179;


dp4a.s32.s32 %r187, %r155, %r255, %r183;


dp4a.s32.s32 %r191, %r159, %r259, %r187;


dp4a.s32.s32 %r195, %r163, %r263, %r191;


{.reg .f16 low,high;
mov.b32 {low,high},%r166;
cvt.f32.f16 %f44, low;}


mul.ftz.f32 %f55, %f43, %f44;
cvt.rn.f32.s32 %f56, %r195;
fma.rn.ftz.f32 %f159, %f55, %f56, %f159;
cvt.u64.u32 %rd39, %r491;
add.s64 %rd40, %rd1, %rd39;
mul.lo.s64 %rd41, %rd40, 36;

{ cvt.f32.f16 %f45, %rs1;}


add.s64 %rd42, %rd6, %rd41;
ld.global.nc.u32 %r201, [%rd42+4];
ld.global.nc.u32 %r205, [%rd42+8];

dp4a.s32.s32 %r200, %r201, %r202, %r236;


dp4a.s32.s32 %r204, %r205, %r206, %r200;

ld.global.nc.u32 %r209, [%rd42+12];
ld.global.nc.u32 %r213, [%rd42+16];

dp4a.s32.s32 %r208, %r209, %r210, %r204;


dp4a.s32.s32 %r212, %r213, %r214, %r208;

ld.global.nc.u32 %r217, [%rd42+20];
ld.global.nc.u32 %r221, [%rd42+24];

dp4a.s32.s32 %r216, %r217, %r218, %r212;


dp4a.s32.s32 %r220, %r221, %r222, %r216;

ld.global.nc.u32 %r225, [%rd42+28];
ld.global.nc.u32 %r229, [%rd42+32];

dp4a.s32.s32 %r224, %r225, %r226, %r220;


dp4a.s32.s32 %r228, %r229, %r230, %r224;

ld.global.nc.u32 %r232, [%rd42];

{.reg .f16 low,high;
mov.b32 {low,high},%r232;
cvt.f32.f16 %f46, low;}


mul.ftz.f32 %f57, %f45, %f46;
cvt.rn.f32.s32 %f58, %r228;
fma.rn.ftz.f32 %f158, %f57, %f58, %f158;

{ cvt.f32.f16 %f47, %rs2;}



dp4a.s32.s32 %r233, %r201, %r235, %r236;


dp4a.s32.s32 %r237, %r205, %r239, %r233;


dp4a.s32.s32 %r241, %r209, %r243, %r237;


dp4a.s32.s32 %r245, %r213, %r247, %r241;


dp4a.s32.s32 %r249, %r217, %r251, %r245;


dp4a.s32.s32 %r253, %r221, %r255, %r249;


dp4a.s32.s32 %r257, %r225, %r259, %r253;


dp4a.s32.s32 %r261, %r229, %r263, %r257;


{.reg .f16 low,high;
mov.b32 {low,high},%r232;
cvt.f32.f16 %f48, low;}


mul.ftz.f32 %f59, %f47, %f48;
cvt.rn.f32.s32 %f60, %r261;
fma.rn.ftz.f32 %f157, %f59, %f60, %f157;
add.s32 %r491, %r491, 128;
add.s32 %r490, %r490, 128;
add.s32 %r489, %r489, 128;
add.s64 %rd69, %rd69, 1152;
add.s64 %rd68, %rd68, 1152;
add.s64 %rd67, %rd67, 1152;
add.s64 %rd66, %rd66, 1152;
add.s32 %r492, %r492, 64;
setp.lt.s32 %p3, %r492, %r36;
@%p3 bra $L__BB20_2;

$L__BB20_3:
setp.eq.s32 %p4, %r1, 0;
@%p4 bra $L__BB20_5;

mov.u32 %r319, _ZZ13mul_mat_vec_qIL9ggml_type42ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r320, %r1, 768, %r319;
shl.b32 %r321, %r32, 2;
add.s32 %r322, %r320, %r321;
st.shared.f32 [%r322+-768], %f162;
st.shared.f32 [%r322+-640], %f161;
st.shared.f32 [%r322+-512], %f160;
st.shared.f32 [%r322+-384], %f159;
st.shared.f32 [%r322+-256], %f158;
st.shared.f32 [%r322+-128], %f157;

$L__BB20_5:
setp.ne.s32 %p5, %r1, 0;
bar.sync 0;
@%p5 bra $L__BB20_18;

mov.u32 %r324, %ctaid.x;
shl.b32 %r325, %r324, 1;
mov.u32 %r326, 1;
shl.b32 %r328, %r32, 2;
mov.u32 %r329, _ZZ13mul_mat_vec_qIL9ggml_type42ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r330, %r329, %r328;
mov.u32 %r331, 2;
ld.shared.f32 %f61, [%r330];
add.ftz.f32 %f62, %f61, %f162;
ld.shared.f32 %f63, [%r330+768];
add.ftz.f32 %f64, %f63, %f62;
ld.shared.f32 %f65, [%r330+1536];
add.ftz.f32 %f66, %f65, %f64;
mov.b32 %r332, %f66;
mov.u32 %r333, 31;
mov.u32 %r334, 16;
mov.u32 %r335, -1;
shfl.sync.bfly.b32 %r336|%p6, %r332, %r334, %r333, %r335;
mov.b32 %f67, %r336;
add.ftz.f32 %f68, %f66, %f67;
mov.b32 %r337, %f68;
mov.u32 %r338, 8;
shfl.sync.bfly.b32 %r339|%p7, %r337, %r338, %r333, %r335;
mov.b32 %f69, %r339;
add.ftz.f32 %f70, %f68, %f69;
mov.b32 %r340, %f70;
mov.u32 %r341, 4;
shfl.sync.bfly.b32 %r342|%p8, %r340, %r341, %r333, %r335;
mov.b32 %f71, %r342;
add.ftz.f32 %f72, %f70, %f71;
mov.b32 %r343, %f72;
shfl.sync.bfly.b32 %r344|%p9, %r343, %r331, %r333, %r335;
mov.b32 %f73, %r344;
add.ftz.f32 %f74, %f72, %f73;
mov.b32 %r345, %f74;
shfl.sync.bfly.b32 %r346|%p10, %r345, %r326, %r333, %r335;
mov.b32 %f75, %r346;
add.ftz.f32 %f19, %f74, %f75;
setp.ne.s32 %p11, %r32, 0;
setp.ge.u32 %p12, %r325, %r18;
or.pred %p13, %p11, %p12;
@%p13 bra $L__BB20_8;

cvta.to.global.u64 %rd43, %rd18;
mov.u32 %r347, %ctaid.y;
mad.lo.s32 %r350, %r347, %r24, %r325;
mov.u32 %r351, %ctaid.z;
mad.lo.s32 %r352, %r351, %r30, %r350;
mul.wide.u32 %rd44, %r352, 4;
add.s64 %rd45, %rd43, %rd44;
st.global.f32 [%rd45], %f19;

$L__BB20_8:
ld.shared.f32 %f76, [%r330+128];
add.ftz.f32 %f77, %f76, %f161;
ld.shared.f32 %f78, [%r330+896];
add.ftz.f32 %f79, %f78, %f77;
ld.shared.f32 %f80, [%r330+1664];
add.ftz.f32 %f81, %f80, %f79;
mov.b32 %r358, %f81;
shfl.sync.bfly.b32 %r362|%p14, %r358, %r334, %r333, %r335;
mov.b32 %f82, %r362;
add.ftz.f32 %f83, %f81, %f82;
mov.b32 %r363, %f83;
shfl.sync.bfly.b32 %r365|%p15, %r363, %r338, %r333, %r335;
mov.b32 %f84, %r365;
add.ftz.f32 %f85, %f83, %f84;
mov.b32 %r366, %f85;
shfl.sync.bfly.b32 %r368|%p16, %r366, %r341, %r333, %r335;
mov.b32 %f86, %r368;
add.ftz.f32 %f87, %f85, %f86;
mov.b32 %r369, %f87;
shfl.sync.bfly.b32 %r370|%p17, %r369, %r331, %r333, %r335;
mov.b32 %f88, %r370;
add.ftz.f32 %f89, %f87, %f88;
mov.b32 %r371, %f89;
shfl.sync.bfly.b32 %r373|%p18, %r371, %r326, %r333, %r335;
mov.b32 %f90, %r373;
add.ftz.f32 %f20, %f89, %f90;
setp.eq.s32 %p19, %r32, 1;
or.b32 %r376, %r325, 1;
setp.lt.u32 %p20, %r376, %r18;
and.pred %p1, %p19, %p20;
not.pred %p21, %p1;
@%p21 bra $L__BB20_10;

cvta.to.global.u64 %rd46, %rd18;
mov.u32 %r377, %ctaid.y;
mad.lo.s32 %r380, %r377, %r24, %r325;
mov.u32 %r381, %ctaid.z;
mad.lo.s32 %r382, %r381, %r30, %r380;
mul.wide.u32 %rd47, %r382, 4;
add.s64 %rd48, %rd46, %rd47;
st.global.f32 [%rd48+4], %f20;

$L__BB20_10:
mov.u32 %r387, 2;
ld.shared.f32 %f91, [%r330+256];
add.ftz.f32 %f92, %f91, %f160;
ld.shared.f32 %f93, [%r330+1024];
add.ftz.f32 %f94, %f93, %f92;
ld.shared.f32 %f95, [%r330+1792];
add.ftz.f32 %f96, %f95, %f94;
mov.b32 %r388, %f96;
mov.u32 %r389, 31;
mov.u32 %r390, 16;
mov.u32 %r391, -1;
shfl.sync.bfly.b32 %r392|%p22, %r388, %r390, %r389, %r391;
mov.b32 %f97, %r392;
add.ftz.f32 %f98, %f96, %f97;
mov.b32 %r393, %f98;
mov.u32 %r394, 8;
shfl.sync.bfly.b32 %r395|%p23, %r393, %r394, %r389, %r391;
mov.b32 %f99, %r395;
add.ftz.f32 %f100, %f98, %f99;
mov.b32 %r396, %f100;
mov.u32 %r397, 4;
shfl.sync.bfly.b32 %r398|%p24, %r396, %r397, %r389, %r391;
mov.b32 %f101, %r398;
add.ftz.f32 %f102, %f100, %f101;
mov.b32 %r399, %f102;
shfl.sync.bfly.b32 %r400|%p25, %r399, %r387, %r389, %r391;
mov.b32 %f103, %r400;
add.ftz.f32 %f104, %f102, %f103;
mov.b32 %r401, %f104;
mov.u32 %r402, 1;
shfl.sync.bfly.b32 %r403|%p26, %r401, %r402, %r389, %r391;
mov.b32 %f105, %r403;
add.ftz.f32 %f21, %f104, %f105;
@%p13 bra $L__BB20_12;

cvt.u64.u32 %rd49, %r18;
mov.u32 %r406, %ctaid.y;
mad.lo.s32 %r409, %r406, %r24, %r325;
mov.u32 %r410, %ctaid.z;
mad.lo.s32 %r411, %r410, %r30, %r409;
cvt.u64.u32 %rd50, %r411;
add.s64 %rd51, %rd49, %rd50;
cvta.to.global.u64 %rd52, %rd18;
shl.b64 %rd53, %rd51, 2;
add.s64 %rd54, %rd52, %rd53;
st.global.f32 [%rd54], %f21;

$L__BB20_12:
ld.shared.f32 %f106, [%r330+384];
add.ftz.f32 %f107, %f106, %f159;
ld.shared.f32 %f108, [%r330+1152];
add.ftz.f32 %f109, %f108, %f107;
ld.shared.f32 %f110, [%r330+1920];
add.ftz.f32 %f111, %f110, %f109;
mov.b32 %r417, %f111;
shfl.sync.bfly.b32 %r421|%p30, %r417, %r390, %r389, %r391;
mov.b32 %f112, %r421;
add.ftz.f32 %f113, %f111, %f112;
mov.b32 %r422, %f113;
shfl.sync.bfly.b32 %r424|%p31, %r422, %r394, %r389, %r391;
mov.b32 %f114, %r424;
add.ftz.f32 %f115, %f113, %f114;
mov.b32 %r425, %f115;
shfl.sync.bfly.b32 %r427|%p32, %r425, %r397, %r389, %r391;
mov.b32 %f116, %r427;
add.ftz.f32 %f117, %f115, %f116;
mov.b32 %r428, %f117;
shfl.sync.bfly.b32 %r429|%p33, %r428, %r387, %r389, %r391;
mov.b32 %f118, %r429;
add.ftz.f32 %f119, %f117, %f118;
mov.b32 %r430, %f119;
shfl.sync.bfly.b32 %r432|%p34, %r430, %r402, %r389, %r391;
mov.b32 %f120, %r432;
add.ftz.f32 %f22, %f119, %f120;
@%p21 bra $L__BB20_14;

add.s32 %r433, %r18, 1;
cvt.u64.u32 %rd55, %r433;
mov.u32 %r434, %ctaid.y;
mad.lo.s32 %r437, %r434, %r24, %r325;
mov.u32 %r438, %ctaid.z;
mad.lo.s32 %r439, %r438, %r30, %r437;
cvt.u64.u32 %rd56, %r439;
add.s64 %rd57, %rd55, %rd56;
cvta.to.global.u64 %rd58, %rd18;
shl.b64 %rd59, %rd57, 2;
add.s64 %rd60, %rd58, %rd59;
st.global.f32 [%rd60], %f22;

$L__BB20_14:
shl.b32 %r440, %r18, 1;
mov.u32 %r441, 1;
mov.u32 %r446, 2;
ld.shared.f32 %f121, [%r330+512];
add.ftz.f32 %f122, %f121, %f158;
ld.shared.f32 %f123, [%r330+1280];
add.ftz.f32 %f124, %f123, %f122;
ld.shared.f32 %f125, [%r330+2048];
add.ftz.f32 %f126, %f125, %f124;
mov.b32 %r447, %f126;
mov.u32 %r448, 31;
mov.u32 %r449, 16;
mov.u32 %r450, -1;
shfl.sync.bfly.b32 %r451|%p36, %r447, %r449, %r448, %r450;
mov.b32 %f127, %r451;
add.ftz.f32 %f128, %f126, %f127;
mov.b32 %r452, %f128;
mov.u32 %r453, 8;
shfl.sync.bfly.b32 %r454|%p37, %r452, %r453, %r448, %r450;
mov.b32 %f129, %r454;
add.ftz.f32 %f130, %f128, %f129;
mov.b32 %r455, %f130;
mov.u32 %r456, 4;
shfl.sync.bfly.b32 %r457|%p38, %r455, %r456, %r448, %r450;
mov.b32 %f131, %r457;
add.ftz.f32 %f132, %f130, %f131;
mov.b32 %r458, %f132;
shfl.sync.bfly.b32 %r459|%p39, %r458, %r446, %r448, %r450;
mov.b32 %f133, %r459;
add.ftz.f32 %f134, %f132, %f133;
mov.b32 %r460, %f134;
shfl.sync.bfly.b32 %r461|%p40, %r460, %r441, %r448, %r450;
mov.b32 %f135, %r461;
add.ftz.f32 %f23, %f134, %f135;
cvt.u64.u32 %rd61, %r440;
mov.u32 %r462, %ctaid.y;
mad.lo.s32 %r465, %r462, %r24, %r325;
mov.u32 %r466, %ctaid.z;
mad.lo.s32 %r467, %r466, %r30, %r465;
cvt.u64.u32 %rd62, %r467;
add.s64 %rd63, %rd61, %rd62;
cvta.to.global.u64 %rd64, %rd18;
shl.b64 %rd65, %rd63, 2;
add.s64 %rd15, %rd64, %rd65;
@%p13 bra $L__BB20_16;

st.global.f32 [%rd15], %f23;

$L__BB20_16:
ld.shared.f32 %f136, [%r330+640];
add.ftz.f32 %f137, %f136, %f157;
ld.shared.f32 %f138, [%r330+1408];
add.ftz.f32 %f139, %f138, %f137;
ld.shared.f32 %f140, [%r330+2176];
add.ftz.f32 %f141, %f140, %f139;
mov.b32 %r473, %f141;
shfl.sync.bfly.b32 %r477|%p44, %r473, %r449, %r448, %r450;
mov.b32 %f142, %r477;
add.ftz.f32 %f143, %f141, %f142;
mov.b32 %r478, %f143;
shfl.sync.bfly.b32 %r480|%p45, %r478, %r453, %r448, %r450;
mov.b32 %f144, %r480;
add.ftz.f32 %f145, %f143, %f144;
mov.b32 %r481, %f145;
shfl.sync.bfly.b32 %r483|%p46, %r481, %r456, %r448, %r450;
mov.b32 %f146, %r483;
add.ftz.f32 %f147, %f145, %f146;
mov.b32 %r484, %f147;
shfl.sync.bfly.b32 %r485|%p47, %r484, %r446, %r448, %r450;
mov.b32 %f148, %r485;
add.ftz.f32 %f149, %f147, %f148;
mov.b32 %r486, %f149;
shfl.sync.bfly.b32 %r488|%p48, %r486, %r441, %r448, %r450;
mov.b32 %f150, %r488;
add.ftz.f32 %f24, %f149, %f150;
@%p21 bra $L__BB20_18;

st.global.f32 [%rd15+4], %f24;

$L__BB20_18:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type42ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type42ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type42ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type42ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type42ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type42ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type42ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type42ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type42ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<69>;
.reg .b16 %rs<17>;
.reg .f32 %f<217>;
.reg .b32 %r<628>;
.reg .b64 %rd<86>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type42ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[3072];

ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type42ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type42ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type42ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r19, [_Z13mul_mat_vec_qIL9ggml_type42ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type42ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type42ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type42ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type42ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type42ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type42ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type42ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type42ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type42ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type42ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type42ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type42ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type42ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r35, %tid.y;
shl.b32 %r36, %r35, 5;
mov.u32 %r37, %tid.x;
add.s32 %r1, %r36, %r37;
shr.u32 %r38, %r1, 31;
add.s32 %r39, %r1, %r38;
shr.s32 %r627, %r39, 1;
shr.u32 %r40, %r19, 6;
setp.ge.s32 %p1, %r627, %r40;
mov.f32 %f209, 0f00000000;
mov.f32 %f210, %f209;
mov.f32 %f211, %f209;
mov.f32 %f212, %f209;
mov.f32 %f213, %f209;
mov.f32 %f214, %f209;
mov.f32 %f215, %f209;
mov.f32 %f216, %f209;
@%p1 bra $L__BB21_3;

mov.u32 %r41, %ctaid.y;
mov.u32 %r42, %ctaid.z;
mul.hi.u32 %r43, %r42, %r29;
add.s32 %r44, %r43, %r42;
shr.u32 %r45, %r44, %r30;
mul.lo.s32 %r46, %r45, %r32;
mul.hi.u32 %r47, %r41, %r23;
add.s32 %r48, %r47, %r41;
shr.u32 %r49, %r48, %r24;
mul.lo.s32 %r50, %r49, %r26;
mul.lo.s32 %r51, %r41, %r27;
cvt.u64.u32 %rd19, %r51;
mul.lo.s32 %r52, %r42, %r33;
cvt.u64.u32 %rd20, %r52;
and.b32 %r55, %r39, -2;
sub.s32 %r56, %r1, %r55;
cvt.s64.s32 %rd21, %r56;
add.s64 %rd22, %rd19, %rd21;
add.s64 %rd1, %rd22, %rd20;
shl.b32 %r57, %r56, 2;
mov.u32 %r58, %ctaid.x;
shl.b32 %r59, %r58, 1;
or.b32 %r60, %r59, 1;
shl.b32 %r61, %r21, 1;
shl.b32 %r623, %r627, 1;
add.s32 %r626, %r623, %r61;
mad.lo.s32 %r625, %r21, 3, %r623;
add.s32 %r624, %r21, %r623;
mad.lo.s32 %r62, %r59, %r20, %r627;
add.s32 %r63, %r62, %r46;
add.s32 %r64, %r63, %r50;
mul.wide.s32 %rd23, %r64, 18;
mul.wide.s32 %rd24, %r57, 2;
add.s64 %rd25, %rd23, %rd24;
cvta.to.global.u64 %rd26, %rd16;
add.s64 %rd27, %rd26, %rd25;
add.s64 %rd85, %rd27, 8;
add.s64 %rd84, %rd26, %rd23;
add.s32 %r65, %r627, %r46;
add.s32 %r66, %r65, %r50;
mad.lo.s32 %r67, %r60, %r20, %r66;
mul.wide.s32 %rd28, %r67, 18;
add.s64 %rd29, %rd28, %rd24;
add.s64 %rd30, %rd26, %rd29;
add.s64 %rd83, %rd30, 4;
add.s64 %rd82, %rd26, %rd28;
cvta.to.global.u64 %rd6, %rd17;
mov.f32 %f209, 0f00000000;

$L__BB21_2:
cvt.u64.u32 %rd31, %r623;
add.s64 %rd32, %rd1, %rd31;
mul.lo.s64 %rd33, %rd32, 36;
ld.global.nc.u16 %rs1, [%rd84];

{ cvt.f32.f16 %f49, %rs1;}


ld.global.nc.u16 %rs9, [%rd85+-6];
cvt.s32.s16 %r332, %rs9;
add.s64 %rd34, %rd6, %rd33;
ld.global.nc.u32 %r69, [%rd34+4];
ld.global.nc.u32 %r73, [%rd34+8];
and.b32 %r333, %r332, 30583;
mov.u32 %r334, 33620223;
prmt.b32 %r335, %r334, %r334, %r333;
shr.u32 %r336, %r332, 2;
and.b32 %r337, %r336, 30583;
prmt.b32 %r338, %r334, %r334, %r337;
mov.u32 %r339, 20800;
prmt.b32 %r268, %r335, %r338, %r339;
mov.u32 %r340, 29538;
prmt.b32 %r272, %r335, %r338, %r340;
mov.u32 %r302, 0;

dp4a.s32.s32 %r68, %r69, %r268, %r302;


dp4a.s32.s32 %r72, %r73, %r272, %r68;

ld.global.nc.u16 %rs10, [%rd85+-4];
cvt.s32.s16 %r341, %rs10;
ld.global.nc.u32 %r77, [%rd34+12];
ld.global.nc.u32 %r81, [%rd34+16];
and.b32 %r342, %r341, 30583;
prmt.b32 %r343, %r334, %r334, %r342;
shr.u32 %r344, %r341, 2;
and.b32 %r345, %r344, 30583;
prmt.b32 %r346, %r334, %r334, %r345;
prmt.b32 %r276, %r343, %r346, %r339;
prmt.b32 %r280, %r343, %r346, %r340;

dp4a.s32.s32 %r76, %r77, %r276, %r72;


dp4a.s32.s32 %r80, %r81, %r280, %r76;

ld.global.nc.u16 %rs11, [%rd85+-2];
cvt.s32.s16 %r347, %rs11;
ld.global.nc.u32 %r85, [%rd34+20];
ld.global.nc.u32 %r89, [%rd34+24];
and.b32 %r348, %r347, 30583;
prmt.b32 %r349, %r334, %r334, %r348;
shr.u32 %r350, %r347, 2;
and.b32 %r351, %r350, 30583;
prmt.b32 %r352, %r334, %r334, %r351;
prmt.b32 %r284, %r349, %r352, %r339;
prmt.b32 %r288, %r349, %r352, %r340;

dp4a.s32.s32 %r84, %r85, %r284, %r80;


dp4a.s32.s32 %r88, %r89, %r288, %r84;

ld.global.nc.u16 %rs12, [%rd85];
cvt.s32.s16 %r353, %rs12;
ld.global.nc.u32 %r93, [%rd34+28];
ld.global.nc.u32 %r97, [%rd34+32];
and.b32 %r354, %r353, 30583;
prmt.b32 %r355, %r334, %r334, %r354;
shr.u32 %r356, %r353, 2;
and.b32 %r357, %r356, 30583;
prmt.b32 %r358, %r334, %r334, %r357;
prmt.b32 %r292, %r355, %r358, %r339;
prmt.b32 %r296, %r355, %r358, %r340;

dp4a.s32.s32 %r92, %r93, %r292, %r88;


dp4a.s32.s32 %r96, %r97, %r296, %r92;

ld.global.nc.u32 %r100, [%rd34];

{.reg .f16 low,high;
mov.b32 {low,high},%r100;
cvt.f32.f16 %f50, low;}


mul.ftz.f32 %f65, %f49, %f50;
cvt.rn.f32.s32 %f66, %r96;
fma.rn.ftz.f32 %f216, %f65, %f66, %f216;
ld.global.nc.u16 %rs2, [%rd82];

{ cvt.f32.f16 %f51, %rs2;}


ld.global.nc.u16 %rs13, [%rd83+-2];
cvt.s32.s16 %r359, %rs13;
and.b32 %r360, %r359, 30583;
prmt.b32 %r361, %r334, %r334, %r360;
shr.u32 %r362, %r359, 2;
and.b32 %r363, %r362, 30583;
prmt.b32 %r364, %r334, %r334, %r363;
prmt.b32 %r301, %r361, %r364, %r339;
prmt.b32 %r305, %r361, %r364, %r340;

dp4a.s32.s32 %r101, %r69, %r301, %r302;


dp4a.s32.s32 %r105, %r73, %r305, %r101;

ld.global.nc.u16 %rs14, [%rd83];
cvt.s32.s16 %r365, %rs14;
and.b32 %r366, %r365, 30583;
prmt.b32 %r367, %r334, %r334, %r366;
shr.u32 %r368, %r365, 2;
and.b32 %r369, %r368, 30583;
prmt.b32 %r370, %r334, %r334, %r369;
prmt.b32 %r309, %r367, %r370, %r339;
prmt.b32 %r313, %r367, %r370, %r340;

dp4a.s32.s32 %r109, %r77, %r309, %r105;


dp4a.s32.s32 %r113, %r81, %r313, %r109;

ld.global.nc.u16 %rs15, [%rd83+2];
cvt.s32.s16 %r371, %rs15;
and.b32 %r372, %r371, 30583;
prmt.b32 %r373, %r334, %r334, %r372;
shr.u32 %r374, %r371, 2;
and.b32 %r375, %r374, 30583;
prmt.b32 %r376, %r334, %r334, %r375;
prmt.b32 %r317, %r373, %r376, %r339;
prmt.b32 %r321, %r373, %r376, %r340;

dp4a.s32.s32 %r117, %r85, %r317, %r113;


dp4a.s32.s32 %r121, %r89, %r321, %r117;

ld.global.nc.u16 %rs16, [%rd83+4];
cvt.s32.s16 %r377, %rs16;
and.b32 %r378, %r377, 30583;
prmt.b32 %r379, %r334, %r334, %r378;
shr.u32 %r380, %r377, 2;
and.b32 %r381, %r380, 30583;
prmt.b32 %r382, %r334, %r334, %r381;
prmt.b32 %r325, %r379, %r382, %r339;
prmt.b32 %r329, %r379, %r382, %r340;

dp4a.s32.s32 %r125, %r93, %r325, %r121;


dp4a.s32.s32 %r129, %r97, %r329, %r125;


{.reg .f16 low,high;
mov.b32 {low,high},%r100;
cvt.f32.f16 %f52, low;}


mul.ftz.f32 %f67, %f51, %f52;
cvt.rn.f32.s32 %f68, %r129;
fma.rn.ftz.f32 %f215, %f67, %f68, %f215;
cvt.u64.u32 %rd35, %r624;
add.s64 %rd36, %rd1, %rd35;
mul.lo.s64 %rd37, %rd36, 36;

{ cvt.f32.f16 %f53, %rs1;}


add.s64 %rd38, %rd6, %rd37;
ld.global.nc.u32 %r135, [%rd38+4];
ld.global.nc.u32 %r139, [%rd38+8];

dp4a.s32.s32 %r134, %r135, %r268, %r302;


dp4a.s32.s32 %r138, %r139, %r272, %r134;

ld.global.nc.u32 %r143, [%rd38+12];
ld.global.nc.u32 %r147, [%rd38+16];

dp4a.s32.s32 %r142, %r143, %r276, %r138;


dp4a.s32.s32 %r146, %r147, %r280, %r142;

ld.global.nc.u32 %r151, [%rd38+20];
ld.global.nc.u32 %r155, [%rd38+24];

dp4a.s32.s32 %r150, %r151, %r284, %r146;


dp4a.s32.s32 %r154, %r155, %r288, %r150;

ld.global.nc.u32 %r159, [%rd38+28];
ld.global.nc.u32 %r163, [%rd38+32];

dp4a.s32.s32 %r158, %r159, %r292, %r154;


dp4a.s32.s32 %r162, %r163, %r296, %r158;

ld.global.nc.u32 %r166, [%rd38];

{.reg .f16 low,high;
mov.b32 {low,high},%r166;
cvt.f32.f16 %f54, low;}


mul.ftz.f32 %f69, %f53, %f54;
cvt.rn.f32.s32 %f70, %r162;
fma.rn.ftz.f32 %f214, %f69, %f70, %f214;

{ cvt.f32.f16 %f55, %rs2;}



dp4a.s32.s32 %r167, %r135, %r301, %r302;


dp4a.s32.s32 %r171, %r139, %r305, %r167;


dp4a.s32.s32 %r175, %r143, %r309, %r171;


dp4a.s32.s32 %r179, %r147, %r313, %r175;


dp4a.s32.s32 %r183, %r151, %r317, %r179;


dp4a.s32.s32 %r187, %r155, %r321, %r183;


dp4a.s32.s32 %r191, %r159, %r325, %r187;


dp4a.s32.s32 %r195, %r163, %r329, %r191;


{.reg .f16 low,high;
mov.b32 {low,high},%r166;
cvt.f32.f16 %f56, low;}


mul.ftz.f32 %f71, %f55, %f56;
cvt.rn.f32.s32 %f72, %r195;
fma.rn.ftz.f32 %f213, %f71, %f72, %f213;
cvt.u64.u32 %rd39, %r626;
add.s64 %rd40, %rd1, %rd39;
mul.lo.s64 %rd41, %rd40, 36;

{ cvt.f32.f16 %f57, %rs1;}


add.s64 %rd42, %rd6, %rd41;
ld.global.nc.u32 %r201, [%rd42+4];
ld.global.nc.u32 %r205, [%rd42+8];

dp4a.s32.s32 %r200, %r201, %r268, %r302;


dp4a.s32.s32 %r204, %r205, %r272, %r200;

ld.global.nc.u32 %r209, [%rd42+12];
ld.global.nc.u32 %r213, [%rd42+16];

dp4a.s32.s32 %r208, %r209, %r276, %r204;


dp4a.s32.s32 %r212, %r213, %r280, %r208;

ld.global.nc.u32 %r217, [%rd42+20];
ld.global.nc.u32 %r221, [%rd42+24];

dp4a.s32.s32 %r216, %r217, %r284, %r212;


dp4a.s32.s32 %r220, %r221, %r288, %r216;

ld.global.nc.u32 %r225, [%rd42+28];
ld.global.nc.u32 %r229, [%rd42+32];

dp4a.s32.s32 %r224, %r225, %r292, %r220;


dp4a.s32.s32 %r228, %r229, %r296, %r224;

ld.global.nc.u32 %r232, [%rd42];

{.reg .f16 low,high;
mov.b32 {low,high},%r232;
cvt.f32.f16 %f58, low;}


mul.ftz.f32 %f73, %f57, %f58;
cvt.rn.f32.s32 %f74, %r228;
fma.rn.ftz.f32 %f212, %f73, %f74, %f212;

{ cvt.f32.f16 %f59, %rs2;}



dp4a.s32.s32 %r233, %r201, %r301, %r302;


dp4a.s32.s32 %r237, %r205, %r305, %r233;


dp4a.s32.s32 %r241, %r209, %r309, %r237;


dp4a.s32.s32 %r245, %r213, %r313, %r241;


dp4a.s32.s32 %r249, %r217, %r317, %r245;


dp4a.s32.s32 %r253, %r221, %r321, %r249;


dp4a.s32.s32 %r257, %r225, %r325, %r253;


dp4a.s32.s32 %r261, %r229, %r329, %r257;


{.reg .f16 low,high;
mov.b32 {low,high},%r232;
cvt.f32.f16 %f60, low;}


mul.ftz.f32 %f75, %f59, %f60;
cvt.rn.f32.s32 %f76, %r261;
fma.rn.ftz.f32 %f211, %f75, %f76, %f211;
cvt.u64.u32 %rd43, %r625;
add.s64 %rd44, %rd1, %rd43;
mul.lo.s64 %rd45, %rd44, 36;

{ cvt.f32.f16 %f61, %rs1;}


add.s64 %rd46, %rd6, %rd45;
ld.global.nc.u32 %r267, [%rd46+4];
ld.global.nc.u32 %r271, [%rd46+8];

dp4a.s32.s32 %r266, %r267, %r268, %r302;


dp4a.s32.s32 %r270, %r271, %r272, %r266;

ld.global.nc.u32 %r275, [%rd46+12];
ld.global.nc.u32 %r279, [%rd46+16];

dp4a.s32.s32 %r274, %r275, %r276, %r270;


dp4a.s32.s32 %r278, %r279, %r280, %r274;

ld.global.nc.u32 %r283, [%rd46+20];
ld.global.nc.u32 %r287, [%rd46+24];

dp4a.s32.s32 %r282, %r283, %r284, %r278;


dp4a.s32.s32 %r286, %r287, %r288, %r282;

ld.global.nc.u32 %r291, [%rd46+28];
ld.global.nc.u32 %r295, [%rd46+32];

dp4a.s32.s32 %r290, %r291, %r292, %r286;


dp4a.s32.s32 %r294, %r295, %r296, %r290;

ld.global.nc.u32 %r298, [%rd46];

{.reg .f16 low,high;
mov.b32 {low,high},%r298;
cvt.f32.f16 %f62, low;}


mul.ftz.f32 %f77, %f61, %f62;
cvt.rn.f32.s32 %f78, %r294;
fma.rn.ftz.f32 %f210, %f77, %f78, %f210;

{ cvt.f32.f16 %f63, %rs2;}



dp4a.s32.s32 %r299, %r267, %r301, %r302;


dp4a.s32.s32 %r303, %r271, %r305, %r299;


dp4a.s32.s32 %r307, %r275, %r309, %r303;


dp4a.s32.s32 %r311, %r279, %r313, %r307;


dp4a.s32.s32 %r315, %r283, %r317, %r311;


dp4a.s32.s32 %r319, %r287, %r321, %r315;


dp4a.s32.s32 %r323, %r291, %r325, %r319;


dp4a.s32.s32 %r327, %r295, %r329, %r323;


{.reg .f16 low,high;
mov.b32 {low,high},%r298;
cvt.f32.f16 %f64, low;}


mul.ftz.f32 %f79, %f63, %f64;
cvt.rn.f32.s32 %f80, %r327;
fma.rn.ftz.f32 %f209, %f79, %f80, %f209;
add.s32 %r626, %r626, 128;
add.s32 %r625, %r625, 128;
add.s32 %r624, %r624, 128;
add.s32 %r623, %r623, 128;
add.s64 %rd85, %rd85, 1152;
add.s64 %rd84, %rd84, 1152;
add.s64 %rd83, %rd83, 1152;
add.s64 %rd82, %rd82, 1152;
add.s32 %r627, %r627, 64;
setp.lt.s32 %p2, %r627, %r40;
@%p2 bra $L__BB21_2;

$L__BB21_3:
setp.eq.s32 %p3, %r35, 0;
@%p3 bra $L__BB21_5;

shl.b32 %r386, %r35, 10;
mov.u32 %r387, _ZZ13mul_mat_vec_qIL9ggml_type42ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r388, %r387, %r386;
shl.b32 %r389, %r37, 2;
add.s32 %r390, %r388, %r389;
st.shared.f32 [%r390+-1024], %f216;
st.shared.f32 [%r390+-896], %f215;
st.shared.f32 [%r390+-768], %f214;
st.shared.f32 [%r390+-640], %f213;
st.shared.f32 [%r390+-512], %f212;
st.shared.f32 [%r390+-384], %f211;
st.shared.f32 [%r390+-256], %f210;
st.shared.f32 [%r390+-128], %f209;

$L__BB21_5:
setp.ne.s32 %p4, %r35, 0;
bar.sync 0;
@%p4 bra $L__BB21_22;

mov.u32 %r392, %ctaid.x;
shl.b32 %r393, %r392, 1;
mov.u32 %r394, 1;
shl.b32 %r396, %r37, 2;
mov.u32 %r397, _ZZ13mul_mat_vec_qIL9ggml_type42ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r398, %r397, %r396;
mov.u32 %r399, 2;
ld.shared.f32 %f81, [%r398];
add.ftz.f32 %f82, %f81, %f216;
ld.shared.f32 %f83, [%r398+1024];
add.ftz.f32 %f84, %f83, %f82;
ld.shared.f32 %f85, [%r398+2048];
add.ftz.f32 %f86, %f85, %f84;
mov.b32 %r400, %f86;
mov.u32 %r401, 31;
mov.u32 %r402, 16;
mov.u32 %r403, -1;
shfl.sync.bfly.b32 %r404|%p5, %r400, %r402, %r401, %r403;
mov.b32 %f87, %r404;
add.ftz.f32 %f88, %f86, %f87;
mov.b32 %r405, %f88;
mov.u32 %r406, 8;
shfl.sync.bfly.b32 %r407|%p6, %r405, %r406, %r401, %r403;
mov.b32 %f89, %r407;
add.ftz.f32 %f90, %f88, %f89;
mov.b32 %r408, %f90;
mov.u32 %r409, 4;
shfl.sync.bfly.b32 %r410|%p7, %r408, %r409, %r401, %r403;
mov.b32 %f91, %r410;
add.ftz.f32 %f92, %f90, %f91;
mov.b32 %r411, %f92;
shfl.sync.bfly.b32 %r412|%p8, %r411, %r399, %r401, %r403;
mov.b32 %f93, %r412;
add.ftz.f32 %f94, %f92, %f93;
mov.b32 %r413, %f94;
shfl.sync.bfly.b32 %r414|%p9, %r413, %r394, %r401, %r403;
mov.b32 %f95, %r414;
add.ftz.f32 %f25, %f94, %f95;
setp.ne.s32 %p10, %r37, 0;
setp.ge.u32 %p11, %r393, %r22;
or.pred %p12, %p10, %p11;
@%p12 bra $L__BB21_8;

cvta.to.global.u64 %rd47, %rd18;
mov.u32 %r415, %ctaid.y;
mad.lo.s32 %r418, %r415, %r28, %r393;
mov.u32 %r419, %ctaid.z;
mad.lo.s32 %r420, %r419, %r34, %r418;
mul.wide.u32 %rd48, %r420, 4;
add.s64 %rd49, %rd47, %rd48;
st.global.f32 [%rd49], %f25;

$L__BB21_8:
ld.shared.f32 %f96, [%r398+128];
add.ftz.f32 %f97, %f96, %f215;
ld.shared.f32 %f98, [%r398+1152];
add.ftz.f32 %f99, %f98, %f97;
ld.shared.f32 %f100, [%r398+2176];
add.ftz.f32 %f101, %f100, %f99;
mov.b32 %r426, %f101;
shfl.sync.bfly.b32 %r430|%p13, %r426, %r402, %r401, %r403;
mov.b32 %f102, %r430;
add.ftz.f32 %f103, %f101, %f102;
mov.b32 %r431, %f103;
shfl.sync.bfly.b32 %r433|%p14, %r431, %r406, %r401, %r403;
mov.b32 %f104, %r433;
add.ftz.f32 %f105, %f103, %f104;
mov.b32 %r434, %f105;
shfl.sync.bfly.b32 %r436|%p15, %r434, %r409, %r401, %r403;
mov.b32 %f106, %r436;
add.ftz.f32 %f107, %f105, %f106;
mov.b32 %r437, %f107;
shfl.sync.bfly.b32 %r438|%p16, %r437, %r399, %r401, %r403;
mov.b32 %f108, %r438;
add.ftz.f32 %f109, %f107, %f108;
mov.b32 %r439, %f109;
shfl.sync.bfly.b32 %r441|%p17, %r439, %r394, %r401, %r403;
mov.b32 %f110, %r441;
add.ftz.f32 %f26, %f109, %f110;
setp.ne.s32 %p18, %r37, 1;
or.b32 %r444, %r393, 1;
setp.ge.u32 %p19, %r444, %r22;
or.pred %p20, %p18, %p19;
@%p20 bra $L__BB21_10;

cvta.to.global.u64 %rd50, %rd18;
mov.u32 %r445, %ctaid.y;
mad.lo.s32 %r448, %r445, %r28, %r393;
mov.u32 %r449, %ctaid.z;
mad.lo.s32 %r450, %r449, %r34, %r448;
mul.wide.u32 %rd51, %r450, 4;
add.s64 %rd52, %rd50, %rd51;
st.global.f32 [%rd52+4], %f26;

$L__BB21_10:
mov.u32 %r455, 2;
ld.shared.f32 %f111, [%r398+256];
add.ftz.f32 %f112, %f111, %f214;
ld.shared.f32 %f113, [%r398+1280];
add.ftz.f32 %f114, %f113, %f112;
ld.shared.f32 %f115, [%r398+2304];
add.ftz.f32 %f116, %f115, %f114;
mov.b32 %r456, %f116;
mov.u32 %r457, 31;
mov.u32 %r458, 16;
mov.u32 %r459, -1;
shfl.sync.bfly.b32 %r460|%p21, %r456, %r458, %r457, %r459;
mov.b32 %f117, %r460;
add.ftz.f32 %f118, %f116, %f117;
mov.b32 %r461, %f118;
mov.u32 %r462, 8;
shfl.sync.bfly.b32 %r463|%p22, %r461, %r462, %r457, %r459;
mov.b32 %f119, %r463;
add.ftz.f32 %f120, %f118, %f119;
mov.b32 %r464, %f120;
mov.u32 %r465, 4;
shfl.sync.bfly.b32 %r466|%p23, %r464, %r465, %r457, %r459;
mov.b32 %f121, %r466;
add.ftz.f32 %f122, %f120, %f121;
mov.b32 %r467, %f122;
shfl.sync.bfly.b32 %r468|%p24, %r467, %r455, %r457, %r459;
mov.b32 %f123, %r468;
add.ftz.f32 %f124, %f122, %f123;
mov.b32 %r469, %f124;
mov.u32 %r470, 1;
shfl.sync.bfly.b32 %r471|%p25, %r469, %r470, %r457, %r459;
mov.b32 %f125, %r471;
add.ftz.f32 %f27, %f124, %f125;
@%p12 bra $L__BB21_12;

cvt.u64.u32 %rd53, %r22;
mov.u32 %r474, %ctaid.y;
mad.lo.s32 %r477, %r474, %r28, %r393;
mov.u32 %r478, %ctaid.z;
mad.lo.s32 %r479, %r478, %r34, %r477;
cvt.u64.u32 %rd54, %r479;
add.s64 %rd55, %rd53, %rd54;
cvta.to.global.u64 %rd56, %rd18;
shl.b64 %rd57, %rd55, 2;
add.s64 %rd58, %rd56, %rd57;
st.global.f32 [%rd58], %f27;

$L__BB21_12:
ld.shared.f32 %f126, [%r398+384];
add.ftz.f32 %f127, %f126, %f213;
ld.shared.f32 %f128, [%r398+1408];
add.ftz.f32 %f129, %f128, %f127;
ld.shared.f32 %f130, [%r398+2432];
add.ftz.f32 %f131, %f130, %f129;
mov.b32 %r485, %f131;
shfl.sync.bfly.b32 %r489|%p29, %r485, %r458, %r457, %r459;
mov.b32 %f132, %r489;
add.ftz.f32 %f133, %f131, %f132;
mov.b32 %r490, %f133;
shfl.sync.bfly.b32 %r492|%p30, %r490, %r462, %r457, %r459;
mov.b32 %f134, %r492;
add.ftz.f32 %f135, %f133, %f134;
mov.b32 %r493, %f135;
shfl.sync.bfly.b32 %r495|%p31, %r493, %r465, %r457, %r459;
mov.b32 %f136, %r495;
add.ftz.f32 %f137, %f135, %f136;
mov.b32 %r496, %f137;
shfl.sync.bfly.b32 %r497|%p32, %r496, %r455, %r457, %r459;
mov.b32 %f138, %r497;
add.ftz.f32 %f139, %f137, %f138;
mov.b32 %r498, %f139;
shfl.sync.bfly.b32 %r500|%p33, %r498, %r470, %r457, %r459;
mov.b32 %f140, %r500;
add.ftz.f32 %f28, %f139, %f140;
@%p20 bra $L__BB21_14;

add.s32 %r504, %r22, 1;
cvt.u64.u32 %rd59, %r504;
mov.u32 %r505, %ctaid.y;
mad.lo.s32 %r508, %r505, %r28, %r393;
mov.u32 %r509, %ctaid.z;
mad.lo.s32 %r510, %r509, %r34, %r508;
cvt.u64.u32 %rd60, %r510;
add.s64 %rd61, %rd59, %rd60;
cvta.to.global.u64 %rd62, %rd18;
shl.b64 %rd63, %rd61, 2;
add.s64 %rd64, %rd62, %rd63;
st.global.f32 [%rd64], %f28;

$L__BB21_14:
shl.b32 %r511, %r22, 1;
mov.u32 %r512, 1;
mov.u32 %r517, 2;
ld.shared.f32 %f141, [%r398+512];
add.ftz.f32 %f142, %f141, %f212;
ld.shared.f32 %f143, [%r398+1536];
add.ftz.f32 %f144, %f143, %f142;
ld.shared.f32 %f145, [%r398+2560];
add.ftz.f32 %f146, %f145, %f144;
mov.b32 %r518, %f146;
mov.u32 %r519, 31;
mov.u32 %r520, 16;
mov.u32 %r521, -1;
shfl.sync.bfly.b32 %r522|%p37, %r518, %r520, %r519, %r521;
mov.b32 %f147, %r522;
add.ftz.f32 %f148, %f146, %f147;
mov.b32 %r523, %f148;
mov.u32 %r524, 8;
shfl.sync.bfly.b32 %r525|%p38, %r523, %r524, %r519, %r521;
mov.b32 %f149, %r525;
add.ftz.f32 %f150, %f148, %f149;
mov.b32 %r526, %f150;
mov.u32 %r527, 4;
shfl.sync.bfly.b32 %r528|%p39, %r526, %r527, %r519, %r521;
mov.b32 %f151, %r528;
add.ftz.f32 %f152, %f150, %f151;
mov.b32 %r529, %f152;
shfl.sync.bfly.b32 %r530|%p40, %r529, %r517, %r519, %r521;
mov.b32 %f153, %r530;
add.ftz.f32 %f154, %f152, %f153;
mov.b32 %r531, %f154;
shfl.sync.bfly.b32 %r532|%p41, %r531, %r512, %r519, %r521;
mov.b32 %f155, %r532;
add.ftz.f32 %f29, %f154, %f155;
cvt.u64.u32 %rd65, %r511;
mov.u32 %r533, %ctaid.y;
mad.lo.s32 %r536, %r533, %r28, %r393;
mov.u32 %r537, %ctaid.z;
mad.lo.s32 %r538, %r537, %r34, %r536;
cvt.u64.u32 %rd66, %r538;
add.s64 %rd67, %rd65, %rd66;
cvta.to.global.u64 %rd68, %rd18;
shl.b64 %rd69, %rd67, 2;
add.s64 %rd15, %rd68, %rd69;
@%p12 bra $L__BB21_16;

st.global.f32 [%rd15], %f29;

$L__BB21_16:
ld.shared.f32 %f156, [%r398+640];
add.ftz.f32 %f157, %f156, %f211;
ld.shared.f32 %f158, [%r398+1664];
add.ftz.f32 %f159, %f158, %f157;
ld.shared.f32 %f160, [%r398+2688];
add.ftz.f32 %f161, %f160, %f159;
mov.b32 %r544, %f161;
shfl.sync.bfly.b32 %r548|%p45, %r544, %r520, %r519, %r521;
mov.b32 %f162, %r548;
add.ftz.f32 %f163, %f161, %f162;
mov.b32 %r549, %f163;
shfl.sync.bfly.b32 %r551|%p46, %r549, %r524, %r519, %r521;
mov.b32 %f164, %r551;
add.ftz.f32 %f165, %f163, %f164;
mov.b32 %r552, %f165;
shfl.sync.bfly.b32 %r554|%p47, %r552, %r527, %r519, %r521;
mov.b32 %f166, %r554;
add.ftz.f32 %f167, %f165, %f166;
mov.b32 %r555, %f167;
shfl.sync.bfly.b32 %r556|%p48, %r555, %r517, %r519, %r521;
mov.b32 %f168, %r556;
add.ftz.f32 %f169, %f167, %f168;
mov.b32 %r557, %f169;
shfl.sync.bfly.b32 %r559|%p49, %r557, %r512, %r519, %r521;
mov.b32 %f170, %r559;
add.ftz.f32 %f30, %f169, %f170;
@%p20 bra $L__BB21_18;

st.global.f32 [%rd15+4], %f30;

$L__BB21_18:
mul.lo.s32 %r18, %r22, 3;
mov.u32 %r567, 2;
ld.shared.f32 %f171, [%r398+768];
add.ftz.f32 %f172, %f171, %f210;
ld.shared.f32 %f173, [%r398+1792];
add.ftz.f32 %f174, %f173, %f172;
ld.shared.f32 %f175, [%r398+2816];
add.ftz.f32 %f176, %f175, %f174;
mov.b32 %r568, %f176;
mov.u32 %r569, 31;
mov.u32 %r570, 16;
mov.u32 %r571, -1;
shfl.sync.bfly.b32 %r572|%p53, %r568, %r570, %r569, %r571;
mov.b32 %f177, %r572;
add.ftz.f32 %f178, %f176, %f177;
mov.b32 %r573, %f178;
mov.u32 %r574, 8;
shfl.sync.bfly.b32 %r575|%p54, %r573, %r574, %r569, %r571;
mov.b32 %f179, %r575;
add.ftz.f32 %f180, %f178, %f179;
mov.b32 %r576, %f180;
mov.u32 %r577, 4;
shfl.sync.bfly.b32 %r578|%p55, %r576, %r577, %r569, %r571;
mov.b32 %f181, %r578;
add.ftz.f32 %f182, %f180, %f181;
mov.b32 %r579, %f182;
shfl.sync.bfly.b32 %r580|%p56, %r579, %r567, %r569, %r571;
mov.b32 %f183, %r580;
add.ftz.f32 %f184, %f182, %f183;
mov.b32 %r581, %f184;
mov.u32 %r582, 1;
shfl.sync.bfly.b32 %r583|%p57, %r581, %r582, %r569, %r571;
mov.b32 %f185, %r583;
add.ftz.f32 %f31, %f184, %f185;
@%p12 bra $L__BB21_20;

cvt.u64.u32 %rd70, %r18;
add.s64 %rd72, %rd70, %rd66;
shl.b64 %rd74, %rd72, 2;
add.s64 %rd75, %rd68, %rd74;
st.global.f32 [%rd75], %f31;

$L__BB21_20:
ld.shared.f32 %f186, [%r398+896];
add.ftz.f32 %f187, %f186, %f209;
ld.shared.f32 %f188, [%r398+1920];
add.ftz.f32 %f189, %f188, %f187;
ld.shared.f32 %f190, [%r398+2944];
add.ftz.f32 %f191, %f190, %f189;
mov.b32 %r597, %f191;
shfl.sync.bfly.b32 %r601|%p61, %r597, %r570, %r569, %r571;
mov.b32 %f192, %r601;
add.ftz.f32 %f193, %f191, %f192;
mov.b32 %r602, %f193;
shfl.sync.bfly.b32 %r604|%p62, %r602, %r574, %r569, %r571;
mov.b32 %f194, %r604;
add.ftz.f32 %f195, %f193, %f194;
mov.b32 %r605, %f195;
shfl.sync.bfly.b32 %r607|%p63, %r605, %r577, %r569, %r571;
mov.b32 %f196, %r607;
add.ftz.f32 %f197, %f195, %f196;
mov.b32 %r608, %f197;
shfl.sync.bfly.b32 %r609|%p64, %r608, %r567, %r569, %r571;
mov.b32 %f198, %r609;
add.ftz.f32 %f199, %f197, %f198;
mov.b32 %r610, %f199;
shfl.sync.bfly.b32 %r612|%p65, %r610, %r582, %r569, %r571;
mov.b32 %f200, %r612;
add.ftz.f32 %f32, %f199, %f200;
@%p20 bra $L__BB21_22;

add.s32 %r616, %r18, 1;
cvt.u64.u32 %rd76, %r616;
add.s64 %rd78, %rd76, %rd66;
shl.b64 %rd80, %rd78, 2;
add.s64 %rd81, %rd68, %rd80;
st.global.f32 [%rd81], %f32;

$L__BB21_22:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type42ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type42ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type42ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type42ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type42ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type42ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type42ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type42ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type42ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<78>;
.reg .b16 %rs<19>;
.reg .f32 %f<231>;
.reg .b32 %r<742>;
.reg .b64 %rd<96>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type42ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1280];

ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type42ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type42ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type42ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type42ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type42ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type42ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type42ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type42ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type42ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type42ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type42ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type42ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type42ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type42ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type42ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type42ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type42ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r37, %tid.y;
shl.b32 %r38, %r37, 5;
mov.u32 %r39, %tid.x;
add.s32 %r40, %r38, %r39;
shr.u32 %r41, %r40, 31;
add.s32 %r42, %r40, %r41;
shr.s32 %r741, %r42, 1;
shr.u32 %r43, %r21, 6;
setp.ge.s32 %p2, %r741, %r43;
mov.f32 %f221, 0f00000000;
mov.f32 %f222, %f221;
mov.f32 %f223, %f221;
mov.f32 %f224, %f221;
mov.f32 %f225, %f221;
mov.f32 %f226, %f221;
mov.f32 %f227, %f221;
mov.f32 %f228, %f221;
mov.f32 %f229, %f221;
mov.f32 %f230, %f221;
@%p2 bra $L__BB22_3;

mov.u32 %r44, %ctaid.y;
mov.u32 %r45, %ctaid.z;
mul.hi.u32 %r46, %r45, %r31;
add.s32 %r47, %r46, %r45;
shr.u32 %r48, %r47, %r32;
mul.lo.s32 %r49, %r48, %r34;
mul.hi.u32 %r50, %r44, %r25;
add.s32 %r51, %r50, %r44;
shr.u32 %r52, %r51, %r26;
mul.lo.s32 %r53, %r52, %r28;
mul.lo.s32 %r54, %r44, %r29;
cvt.u64.u32 %rd20, %r54;
mul.lo.s32 %r55, %r45, %r35;
cvt.u64.u32 %rd21, %r55;
and.b32 %r62, %r42, -2;
sub.s32 %r63, %r40, %r62;
cvt.s64.s32 %rd22, %r63;
add.s64 %rd23, %rd20, %rd22;
add.s64 %rd1, %rd23, %rd21;
shl.b32 %r64, %r63, 2;
mov.u32 %r65, %ctaid.x;
shl.b32 %r66, %r65, 1;
or.b32 %r67, %r66, 1;
shl.b32 %r68, %r23, 1;
shl.b32 %r736, %r741, 1;
add.s32 %r740, %r736, %r68;
mad.lo.s32 %r739, %r23, 3, %r736;
shl.b32 %r69, %r23, 2;
add.s32 %r738, %r736, %r69;
add.s32 %r737, %r23, %r736;
mad.lo.s32 %r70, %r66, %r22, %r741;
add.s32 %r71, %r70, %r49;
add.s32 %r72, %r71, %r53;
mul.wide.s32 %rd24, %r72, 18;
mul.wide.s32 %rd25, %r64, 2;
add.s64 %rd26, %rd24, %rd25;
cvta.to.global.u64 %rd27, %rd17;
add.s64 %rd28, %rd27, %rd26;
add.s64 %rd95, %rd28, 8;
add.s64 %rd94, %rd27, %rd24;
add.s32 %r73, %r741, %r49;
add.s32 %r74, %r73, %r53;
mad.lo.s32 %r75, %r67, %r22, %r74;
mul.wide.s32 %rd29, %r75, 18;
add.s64 %rd30, %rd29, %rd25;
add.s64 %rd31, %rd27, %rd30;
add.s64 %rd93, %rd31, 4;
add.s64 %rd92, %rd27, %rd29;
cvta.to.global.u64 %rd6, %rd18;
mov.f32 %f221, 0f00000000;

$L__BB22_2:
cvt.u64.u32 %rd32, %r736;
add.s64 %rd33, %rd1, %rd32;
mul.lo.s64 %rd34, %rd33, 36;
ld.global.nc.u16 %rs1, [%rd94];

{ cvt.f32.f16 %f61, %rs1;}


ld.global.nc.u16 %rs11, [%rd95+-6];
cvt.s32.s16 %r406, %rs11;
add.s64 %rd35, %rd6, %rd34;
ld.global.nc.u32 %r77, [%rd35+4];
ld.global.nc.u32 %r81, [%rd35+8];
and.b32 %r407, %r406, 30583;
mov.u32 %r408, 33620223;
prmt.b32 %r409, %r408, %r408, %r407;
shr.u32 %r410, %r406, 2;
and.b32 %r411, %r410, 30583;
prmt.b32 %r412, %r408, %r408, %r411;
mov.u32 %r413, 20800;
prmt.b32 %r342, %r409, %r412, %r413;
mov.u32 %r414, 29538;
prmt.b32 %r346, %r409, %r412, %r414;
mov.u32 %r376, 0;

dp4a.s32.s32 %r76, %r77, %r342, %r376;


dp4a.s32.s32 %r80, %r81, %r346, %r76;

ld.global.nc.u16 %rs12, [%rd95+-4];
cvt.s32.s16 %r415, %rs12;
ld.global.nc.u32 %r85, [%rd35+12];
ld.global.nc.u32 %r89, [%rd35+16];
and.b32 %r416, %r415, 30583;
prmt.b32 %r417, %r408, %r408, %r416;
shr.u32 %r418, %r415, 2;
and.b32 %r419, %r418, 30583;
prmt.b32 %r420, %r408, %r408, %r419;
prmt.b32 %r350, %r417, %r420, %r413;
prmt.b32 %r354, %r417, %r420, %r414;

dp4a.s32.s32 %r84, %r85, %r350, %r80;


dp4a.s32.s32 %r88, %r89, %r354, %r84;

ld.global.nc.u16 %rs13, [%rd95+-2];
cvt.s32.s16 %r421, %rs13;
ld.global.nc.u32 %r93, [%rd35+20];
ld.global.nc.u32 %r97, [%rd35+24];
and.b32 %r422, %r421, 30583;
prmt.b32 %r423, %r408, %r408, %r422;
shr.u32 %r424, %r421, 2;
and.b32 %r425, %r424, 30583;
prmt.b32 %r426, %r408, %r408, %r425;
prmt.b32 %r358, %r423, %r426, %r413;
prmt.b32 %r362, %r423, %r426, %r414;

dp4a.s32.s32 %r92, %r93, %r358, %r88;


dp4a.s32.s32 %r96, %r97, %r362, %r92;

ld.global.nc.u16 %rs14, [%rd95];
cvt.s32.s16 %r427, %rs14;
ld.global.nc.u32 %r101, [%rd35+28];
ld.global.nc.u32 %r105, [%rd35+32];
and.b32 %r428, %r427, 30583;
prmt.b32 %r429, %r408, %r408, %r428;
shr.u32 %r430, %r427, 2;
and.b32 %r431, %r430, 30583;
prmt.b32 %r432, %r408, %r408, %r431;
prmt.b32 %r366, %r429, %r432, %r413;
prmt.b32 %r370, %r429, %r432, %r414;

dp4a.s32.s32 %r100, %r101, %r366, %r96;


dp4a.s32.s32 %r104, %r105, %r370, %r100;

ld.global.nc.u32 %r108, [%rd35];

{.reg .f16 low,high;
mov.b32 {low,high},%r108;
cvt.f32.f16 %f62, low;}


mul.ftz.f32 %f81, %f61, %f62;
cvt.rn.f32.s32 %f82, %r104;
fma.rn.ftz.f32 %f230, %f81, %f82, %f230;
ld.global.nc.u16 %rs2, [%rd92];

{ cvt.f32.f16 %f63, %rs2;}


ld.global.nc.u16 %rs15, [%rd93+-2];
cvt.s32.s16 %r433, %rs15;
and.b32 %r434, %r433, 30583;
prmt.b32 %r435, %r408, %r408, %r434;
shr.u32 %r436, %r433, 2;
and.b32 %r437, %r436, 30583;
prmt.b32 %r438, %r408, %r408, %r437;
prmt.b32 %r375, %r435, %r438, %r413;
prmt.b32 %r379, %r435, %r438, %r414;

dp4a.s32.s32 %r109, %r77, %r375, %r376;


dp4a.s32.s32 %r113, %r81, %r379, %r109;

ld.global.nc.u16 %rs16, [%rd93];
cvt.s32.s16 %r439, %rs16;
and.b32 %r440, %r439, 30583;
prmt.b32 %r441, %r408, %r408, %r440;
shr.u32 %r442, %r439, 2;
and.b32 %r443, %r442, 30583;
prmt.b32 %r444, %r408, %r408, %r443;
prmt.b32 %r383, %r441, %r444, %r413;
prmt.b32 %r387, %r441, %r444, %r414;

dp4a.s32.s32 %r117, %r85, %r383, %r113;


dp4a.s32.s32 %r121, %r89, %r387, %r117;

ld.global.nc.u16 %rs17, [%rd93+2];
cvt.s32.s16 %r445, %rs17;
and.b32 %r446, %r445, 30583;
prmt.b32 %r447, %r408, %r408, %r446;
shr.u32 %r448, %r445, 2;
and.b32 %r449, %r448, 30583;
prmt.b32 %r450, %r408, %r408, %r449;
prmt.b32 %r391, %r447, %r450, %r413;
prmt.b32 %r395, %r447, %r450, %r414;

dp4a.s32.s32 %r125, %r93, %r391, %r121;


dp4a.s32.s32 %r129, %r97, %r395, %r125;

ld.global.nc.u16 %rs18, [%rd93+4];
cvt.s32.s16 %r451, %rs18;
and.b32 %r452, %r451, 30583;
prmt.b32 %r453, %r408, %r408, %r452;
shr.u32 %r454, %r451, 2;
and.b32 %r455, %r454, 30583;
prmt.b32 %r456, %r408, %r408, %r455;
prmt.b32 %r399, %r453, %r456, %r413;
prmt.b32 %r403, %r453, %r456, %r414;

dp4a.s32.s32 %r133, %r101, %r399, %r129;


dp4a.s32.s32 %r137, %r105, %r403, %r133;


{.reg .f16 low,high;
mov.b32 {low,high},%r108;
cvt.f32.f16 %f64, low;}


mul.ftz.f32 %f83, %f63, %f64;
cvt.rn.f32.s32 %f84, %r137;
fma.rn.ftz.f32 %f229, %f83, %f84, %f229;
cvt.u64.u32 %rd36, %r737;
add.s64 %rd37, %rd1, %rd36;
mul.lo.s64 %rd38, %rd37, 36;

{ cvt.f32.f16 %f65, %rs1;}


add.s64 %rd39, %rd6, %rd38;
ld.global.nc.u32 %r143, [%rd39+4];
ld.global.nc.u32 %r147, [%rd39+8];

dp4a.s32.s32 %r142, %r143, %r342, %r376;


dp4a.s32.s32 %r146, %r147, %r346, %r142;

ld.global.nc.u32 %r151, [%rd39+12];
ld.global.nc.u32 %r155, [%rd39+16];

dp4a.s32.s32 %r150, %r151, %r350, %r146;


dp4a.s32.s32 %r154, %r155, %r354, %r150;

ld.global.nc.u32 %r159, [%rd39+20];
ld.global.nc.u32 %r163, [%rd39+24];

dp4a.s32.s32 %r158, %r159, %r358, %r154;


dp4a.s32.s32 %r162, %r163, %r362, %r158;

ld.global.nc.u32 %r167, [%rd39+28];
ld.global.nc.u32 %r171, [%rd39+32];

dp4a.s32.s32 %r166, %r167, %r366, %r162;


dp4a.s32.s32 %r170, %r171, %r370, %r166;

ld.global.nc.u32 %r174, [%rd39];

{.reg .f16 low,high;
mov.b32 {low,high},%r174;
cvt.f32.f16 %f66, low;}


mul.ftz.f32 %f85, %f65, %f66;
cvt.rn.f32.s32 %f86, %r170;
fma.rn.ftz.f32 %f228, %f85, %f86, %f228;

{ cvt.f32.f16 %f67, %rs2;}



dp4a.s32.s32 %r175, %r143, %r375, %r376;


dp4a.s32.s32 %r179, %r147, %r379, %r175;


dp4a.s32.s32 %r183, %r151, %r383, %r179;


dp4a.s32.s32 %r187, %r155, %r387, %r183;


dp4a.s32.s32 %r191, %r159, %r391, %r187;


dp4a.s32.s32 %r195, %r163, %r395, %r191;


dp4a.s32.s32 %r199, %r167, %r399, %r195;


dp4a.s32.s32 %r203, %r171, %r403, %r199;


{.reg .f16 low,high;
mov.b32 {low,high},%r174;
cvt.f32.f16 %f68, low;}


mul.ftz.f32 %f87, %f67, %f68;
cvt.rn.f32.s32 %f88, %r203;
fma.rn.ftz.f32 %f227, %f87, %f88, %f227;
cvt.u64.u32 %rd40, %r740;
add.s64 %rd41, %rd1, %rd40;
mul.lo.s64 %rd42, %rd41, 36;

{ cvt.f32.f16 %f69, %rs1;}


add.s64 %rd43, %rd6, %rd42;
ld.global.nc.u32 %r209, [%rd43+4];
ld.global.nc.u32 %r213, [%rd43+8];

dp4a.s32.s32 %r208, %r209, %r342, %r376;


dp4a.s32.s32 %r212, %r213, %r346, %r208;

ld.global.nc.u32 %r217, [%rd43+12];
ld.global.nc.u32 %r221, [%rd43+16];

dp4a.s32.s32 %r216, %r217, %r350, %r212;


dp4a.s32.s32 %r220, %r221, %r354, %r216;

ld.global.nc.u32 %r225, [%rd43+20];
ld.global.nc.u32 %r229, [%rd43+24];

dp4a.s32.s32 %r224, %r225, %r358, %r220;


dp4a.s32.s32 %r228, %r229, %r362, %r224;

ld.global.nc.u32 %r233, [%rd43+28];
ld.global.nc.u32 %r237, [%rd43+32];

dp4a.s32.s32 %r232, %r233, %r366, %r228;


dp4a.s32.s32 %r236, %r237, %r370, %r232;

ld.global.nc.u32 %r240, [%rd43];

{.reg .f16 low,high;
mov.b32 {low,high},%r240;
cvt.f32.f16 %f70, low;}


mul.ftz.f32 %f89, %f69, %f70;
cvt.rn.f32.s32 %f90, %r236;
fma.rn.ftz.f32 %f226, %f89, %f90, %f226;

{ cvt.f32.f16 %f71, %rs2;}



dp4a.s32.s32 %r241, %r209, %r375, %r376;


dp4a.s32.s32 %r245, %r213, %r379, %r241;


dp4a.s32.s32 %r249, %r217, %r383, %r245;


dp4a.s32.s32 %r253, %r221, %r387, %r249;


dp4a.s32.s32 %r257, %r225, %r391, %r253;


dp4a.s32.s32 %r261, %r229, %r395, %r257;


dp4a.s32.s32 %r265, %r233, %r399, %r261;


dp4a.s32.s32 %r269, %r237, %r403, %r265;


{.reg .f16 low,high;
mov.b32 {low,high},%r240;
cvt.f32.f16 %f72, low;}


mul.ftz.f32 %f91, %f71, %f72;
cvt.rn.f32.s32 %f92, %r269;
fma.rn.ftz.f32 %f225, %f91, %f92, %f225;
cvt.u64.u32 %rd44, %r739;
add.s64 %rd45, %rd1, %rd44;
mul.lo.s64 %rd46, %rd45, 36;

{ cvt.f32.f16 %f73, %rs1;}


add.s64 %rd47, %rd6, %rd46;
ld.global.nc.u32 %r275, [%rd47+4];
ld.global.nc.u32 %r279, [%rd47+8];

dp4a.s32.s32 %r274, %r275, %r342, %r376;


dp4a.s32.s32 %r278, %r279, %r346, %r274;

ld.global.nc.u32 %r283, [%rd47+12];
ld.global.nc.u32 %r287, [%rd47+16];

dp4a.s32.s32 %r282, %r283, %r350, %r278;


dp4a.s32.s32 %r286, %r287, %r354, %r282;

ld.global.nc.u32 %r291, [%rd47+20];
ld.global.nc.u32 %r295, [%rd47+24];

dp4a.s32.s32 %r290, %r291, %r358, %r286;


dp4a.s32.s32 %r294, %r295, %r362, %r290;

ld.global.nc.u32 %r299, [%rd47+28];
ld.global.nc.u32 %r303, [%rd47+32];

dp4a.s32.s32 %r298, %r299, %r366, %r294;


dp4a.s32.s32 %r302, %r303, %r370, %r298;

ld.global.nc.u32 %r306, [%rd47];

{.reg .f16 low,high;
mov.b32 {low,high},%r306;
cvt.f32.f16 %f74, low;}


mul.ftz.f32 %f93, %f73, %f74;
cvt.rn.f32.s32 %f94, %r302;
fma.rn.ftz.f32 %f224, %f93, %f94, %f224;

{ cvt.f32.f16 %f75, %rs2;}



dp4a.s32.s32 %r307, %r275, %r375, %r376;


dp4a.s32.s32 %r311, %r279, %r379, %r307;


dp4a.s32.s32 %r315, %r283, %r383, %r311;


dp4a.s32.s32 %r319, %r287, %r387, %r315;


dp4a.s32.s32 %r323, %r291, %r391, %r319;


dp4a.s32.s32 %r327, %r295, %r395, %r323;


dp4a.s32.s32 %r331, %r299, %r399, %r327;


dp4a.s32.s32 %r335, %r303, %r403, %r331;


{.reg .f16 low,high;
mov.b32 {low,high},%r306;
cvt.f32.f16 %f76, low;}


mul.ftz.f32 %f95, %f75, %f76;
cvt.rn.f32.s32 %f96, %r335;
fma.rn.ftz.f32 %f223, %f95, %f96, %f223;
cvt.u64.u32 %rd48, %r738;
add.s64 %rd49, %rd1, %rd48;
mul.lo.s64 %rd50, %rd49, 36;

{ cvt.f32.f16 %f77, %rs1;}


add.s64 %rd51, %rd6, %rd50;
ld.global.nc.u32 %r341, [%rd51+4];
ld.global.nc.u32 %r345, [%rd51+8];

dp4a.s32.s32 %r340, %r341, %r342, %r376;


dp4a.s32.s32 %r344, %r345, %r346, %r340;

ld.global.nc.u32 %r349, [%rd51+12];
ld.global.nc.u32 %r353, [%rd51+16];

dp4a.s32.s32 %r348, %r349, %r350, %r344;


dp4a.s32.s32 %r352, %r353, %r354, %r348;

ld.global.nc.u32 %r357, [%rd51+20];
ld.global.nc.u32 %r361, [%rd51+24];

dp4a.s32.s32 %r356, %r357, %r358, %r352;


dp4a.s32.s32 %r360, %r361, %r362, %r356;

ld.global.nc.u32 %r365, [%rd51+28];
ld.global.nc.u32 %r369, [%rd51+32];

dp4a.s32.s32 %r364, %r365, %r366, %r360;


dp4a.s32.s32 %r368, %r369, %r370, %r364;

ld.global.nc.u32 %r372, [%rd51];

{.reg .f16 low,high;
mov.b32 {low,high},%r372;
cvt.f32.f16 %f78, low;}


mul.ftz.f32 %f97, %f77, %f78;
cvt.rn.f32.s32 %f98, %r368;
fma.rn.ftz.f32 %f222, %f97, %f98, %f222;

{ cvt.f32.f16 %f79, %rs2;}



dp4a.s32.s32 %r373, %r341, %r375, %r376;


dp4a.s32.s32 %r377, %r345, %r379, %r373;


dp4a.s32.s32 %r381, %r349, %r383, %r377;


dp4a.s32.s32 %r385, %r353, %r387, %r381;


dp4a.s32.s32 %r389, %r357, %r391, %r385;


dp4a.s32.s32 %r393, %r361, %r395, %r389;


dp4a.s32.s32 %r397, %r365, %r399, %r393;


dp4a.s32.s32 %r401, %r369, %r403, %r397;


{.reg .f16 low,high;
mov.b32 {low,high},%r372;
cvt.f32.f16 %f80, low;}


mul.ftz.f32 %f99, %f79, %f80;
cvt.rn.f32.s32 %f100, %r401;
fma.rn.ftz.f32 %f221, %f99, %f100, %f221;
add.s32 %r740, %r740, 64;
add.s32 %r739, %r739, 64;
add.s32 %r738, %r738, 64;
add.s32 %r737, %r737, 64;
add.s32 %r736, %r736, 64;
add.s64 %rd95, %rd95, 576;
add.s64 %rd94, %rd94, 576;
add.s64 %rd93, %rd93, 576;
add.s64 %rd92, %rd92, 576;
add.s32 %r741, %r741, 32;
setp.lt.s32 %p3, %r741, %r43;
@%p3 bra $L__BB22_2;

$L__BB22_3:
setp.eq.s32 %p4, %r37, 0;
@%p4 bra $L__BB22_5;

mov.u32 %r460, _ZZ13mul_mat_vec_qIL9ggml_type42ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r461, %r37, 1280, %r460;
shl.b32 %r462, %r39, 2;
add.s32 %r463, %r461, %r462;
st.shared.f32 [%r463+-1280], %f230;
st.shared.f32 [%r463+-1152], %f229;
st.shared.f32 [%r463+-1024], %f228;
st.shared.f32 [%r463+-896], %f227;
st.shared.f32 [%r463+-768], %f226;
st.shared.f32 [%r463+-640], %f225;
st.shared.f32 [%r463+-512], %f224;
st.shared.f32 [%r463+-384], %f223;
st.shared.f32 [%r463+-256], %f222;
st.shared.f32 [%r463+-128], %f221;

$L__BB22_5:
setp.ne.s32 %p5, %r37, 0;
bar.sync 0;
@%p5 bra $L__BB22_26;

mov.u32 %r465, %ctaid.x;
shl.b32 %r466, %r465, 1;
mov.u32 %r467, 1;
shl.b32 %r469, %r39, 2;
mov.u32 %r470, _ZZ13mul_mat_vec_qIL9ggml_type42ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r471, %r470, %r469;
mov.u32 %r472, 2;
ld.shared.f32 %f101, [%r471];
add.ftz.f32 %f102, %f101, %f230;
mov.b32 %r473, %f102;
mov.u32 %r474, 31;
mov.u32 %r475, 16;
mov.u32 %r476, -1;
shfl.sync.bfly.b32 %r477|%p6, %r473, %r475, %r474, %r476;
mov.b32 %f103, %r477;
add.ftz.f32 %f104, %f102, %f103;
mov.b32 %r478, %f104;
mov.u32 %r479, 8;
shfl.sync.bfly.b32 %r480|%p7, %r478, %r479, %r474, %r476;
mov.b32 %f105, %r480;
add.ftz.f32 %f106, %f104, %f105;
mov.b32 %r481, %f106;
mov.u32 %r482, 4;
shfl.sync.bfly.b32 %r483|%p8, %r481, %r482, %r474, %r476;
mov.b32 %f107, %r483;
add.ftz.f32 %f108, %f106, %f107;
mov.b32 %r484, %f108;
shfl.sync.bfly.b32 %r485|%p9, %r484, %r472, %r474, %r476;
mov.b32 %f109, %r485;
add.ftz.f32 %f110, %f108, %f109;
mov.b32 %r486, %f110;
shfl.sync.bfly.b32 %r487|%p10, %r486, %r467, %r474, %r476;
mov.b32 %f111, %r487;
add.ftz.f32 %f31, %f110, %f111;
setp.ne.s32 %p11, %r39, 0;
setp.ge.u32 %p12, %r466, %r24;
or.pred %p13, %p11, %p12;
@%p13 bra $L__BB22_8;

cvta.to.global.u64 %rd52, %rd19;
mov.u32 %r488, %ctaid.y;
mad.lo.s32 %r491, %r488, %r30, %r466;
mov.u32 %r492, %ctaid.z;
mad.lo.s32 %r493, %r492, %r36, %r491;
mul.wide.u32 %rd53, %r493, 4;
add.s64 %rd54, %rd52, %rd53;
st.global.f32 [%rd54], %f31;

$L__BB22_8:
ld.shared.f32 %f112, [%r471+128];
add.ftz.f32 %f113, %f112, %f229;
mov.b32 %r499, %f113;
shfl.sync.bfly.b32 %r503|%p14, %r499, %r475, %r474, %r476;
mov.b32 %f114, %r503;
add.ftz.f32 %f115, %f113, %f114;
mov.b32 %r504, %f115;
shfl.sync.bfly.b32 %r506|%p15, %r504, %r479, %r474, %r476;
mov.b32 %f116, %r506;
add.ftz.f32 %f117, %f115, %f116;
mov.b32 %r507, %f117;
shfl.sync.bfly.b32 %r509|%p16, %r507, %r482, %r474, %r476;
mov.b32 %f118, %r509;
add.ftz.f32 %f119, %f117, %f118;
mov.b32 %r510, %f119;
shfl.sync.bfly.b32 %r511|%p17, %r510, %r472, %r474, %r476;
mov.b32 %f120, %r511;
add.ftz.f32 %f121, %f119, %f120;
mov.b32 %r512, %f121;
shfl.sync.bfly.b32 %r514|%p18, %r512, %r467, %r474, %r476;
mov.b32 %f122, %r514;
add.ftz.f32 %f32, %f121, %f122;
setp.eq.s32 %p19, %r39, 1;
or.b32 %r517, %r466, 1;
setp.lt.u32 %p20, %r517, %r24;
and.pred %p1, %p19, %p20;
not.pred %p21, %p1;
@%p21 bra $L__BB22_10;

cvta.to.global.u64 %rd55, %rd19;
mov.u32 %r518, %ctaid.y;
mad.lo.s32 %r521, %r518, %r30, %r466;
mov.u32 %r522, %ctaid.z;
mad.lo.s32 %r523, %r522, %r36, %r521;
mul.wide.u32 %rd56, %r523, 4;
add.s64 %rd57, %rd55, %rd56;
st.global.f32 [%rd57+4], %f32;

$L__BB22_10:
mov.u32 %r528, 2;
ld.shared.f32 %f123, [%r471+256];
add.ftz.f32 %f124, %f123, %f228;
mov.b32 %r529, %f124;
mov.u32 %r530, 31;
mov.u32 %r531, 16;
mov.u32 %r532, -1;
shfl.sync.bfly.b32 %r533|%p22, %r529, %r531, %r530, %r532;
mov.b32 %f125, %r533;
add.ftz.f32 %f126, %f124, %f125;
mov.b32 %r534, %f126;
mov.u32 %r535, 8;
shfl.sync.bfly.b32 %r536|%p23, %r534, %r535, %r530, %r532;
mov.b32 %f127, %r536;
add.ftz.f32 %f128, %f126, %f127;
mov.b32 %r537, %f128;
mov.u32 %r538, 4;
shfl.sync.bfly.b32 %r539|%p24, %r537, %r538, %r530, %r532;
mov.b32 %f129, %r539;
add.ftz.f32 %f130, %f128, %f129;
mov.b32 %r540, %f130;
shfl.sync.bfly.b32 %r541|%p25, %r540, %r528, %r530, %r532;
mov.b32 %f131, %r541;
add.ftz.f32 %f132, %f130, %f131;
mov.b32 %r542, %f132;
mov.u32 %r543, 1;
shfl.sync.bfly.b32 %r544|%p26, %r542, %r543, %r530, %r532;
mov.b32 %f133, %r544;
add.ftz.f32 %f33, %f132, %f133;
@%p13 bra $L__BB22_12;

cvt.u64.u32 %rd58, %r24;
mov.u32 %r547, %ctaid.y;
mad.lo.s32 %r550, %r547, %r30, %r466;
mov.u32 %r551, %ctaid.z;
mad.lo.s32 %r552, %r551, %r36, %r550;
cvt.u64.u32 %rd59, %r552;
add.s64 %rd60, %rd58, %rd59;
cvta.to.global.u64 %rd61, %rd19;
shl.b64 %rd62, %rd60, 2;
add.s64 %rd63, %rd61, %rd62;
st.global.f32 [%rd63], %f33;

$L__BB22_12:
ld.shared.f32 %f134, [%r471+384];
add.ftz.f32 %f135, %f134, %f227;
mov.b32 %r558, %f135;
shfl.sync.bfly.b32 %r562|%p30, %r558, %r531, %r530, %r532;
mov.b32 %f136, %r562;
add.ftz.f32 %f137, %f135, %f136;
mov.b32 %r563, %f137;
shfl.sync.bfly.b32 %r565|%p31, %r563, %r535, %r530, %r532;
mov.b32 %f138, %r565;
add.ftz.f32 %f139, %f137, %f138;
mov.b32 %r566, %f139;
shfl.sync.bfly.b32 %r568|%p32, %r566, %r538, %r530, %r532;
mov.b32 %f140, %r568;
add.ftz.f32 %f141, %f139, %f140;
mov.b32 %r569, %f141;
shfl.sync.bfly.b32 %r570|%p33, %r569, %r528, %r530, %r532;
mov.b32 %f142, %r570;
add.ftz.f32 %f143, %f141, %f142;
mov.b32 %r571, %f143;
shfl.sync.bfly.b32 %r573|%p34, %r571, %r543, %r530, %r532;
mov.b32 %f144, %r573;
add.ftz.f32 %f34, %f143, %f144;
@%p21 bra $L__BB22_14;

add.s32 %r574, %r24, 1;
cvt.u64.u32 %rd64, %r574;
mov.u32 %r575, %ctaid.y;
mad.lo.s32 %r578, %r575, %r30, %r466;
mov.u32 %r579, %ctaid.z;
mad.lo.s32 %r580, %r579, %r36, %r578;
cvt.u64.u32 %rd65, %r580;
add.s64 %rd66, %rd64, %rd65;
cvta.to.global.u64 %rd67, %rd19;
shl.b64 %rd68, %rd66, 2;
add.s64 %rd69, %rd67, %rd68;
st.global.f32 [%rd69], %f34;

$L__BB22_14:
shl.b32 %r581, %r24, 1;
mov.u32 %r582, 1;
mov.u32 %r587, 2;
ld.shared.f32 %f145, [%r471+512];
add.ftz.f32 %f146, %f145, %f226;
mov.b32 %r588, %f146;
mov.u32 %r589, 31;
mov.u32 %r590, 16;
mov.u32 %r591, -1;
shfl.sync.bfly.b32 %r592|%p36, %r588, %r590, %r589, %r591;
mov.b32 %f147, %r592;
add.ftz.f32 %f148, %f146, %f147;
mov.b32 %r593, %f148;
mov.u32 %r594, 8;
shfl.sync.bfly.b32 %r595|%p37, %r593, %r594, %r589, %r591;
mov.b32 %f149, %r595;
add.ftz.f32 %f150, %f148, %f149;
mov.b32 %r596, %f150;
mov.u32 %r597, 4;
shfl.sync.bfly.b32 %r598|%p38, %r596, %r597, %r589, %r591;
mov.b32 %f151, %r598;
add.ftz.f32 %f152, %f150, %f151;
mov.b32 %r599, %f152;
shfl.sync.bfly.b32 %r600|%p39, %r599, %r587, %r589, %r591;
mov.b32 %f153, %r600;
add.ftz.f32 %f154, %f152, %f153;
mov.b32 %r601, %f154;
shfl.sync.bfly.b32 %r602|%p40, %r601, %r582, %r589, %r591;
mov.b32 %f155, %r602;
add.ftz.f32 %f35, %f154, %f155;
cvt.u64.u32 %rd70, %r581;
mov.u32 %r603, %ctaid.y;
mad.lo.s32 %r606, %r603, %r30, %r466;
mov.u32 %r607, %ctaid.z;
mad.lo.s32 %r608, %r607, %r36, %r606;
cvt.u64.u32 %rd71, %r608;
add.s64 %rd72, %rd70, %rd71;
cvta.to.global.u64 %rd73, %rd19;
shl.b64 %rd74, %rd72, 2;
add.s64 %rd15, %rd73, %rd74;
@%p13 bra $L__BB22_16;

st.global.f32 [%rd15], %f35;

$L__BB22_16:
ld.shared.f32 %f156, [%r471+640];
add.ftz.f32 %f157, %f156, %f225;
mov.b32 %r614, %f157;
shfl.sync.bfly.b32 %r618|%p44, %r614, %r590, %r589, %r591;
mov.b32 %f158, %r618;
add.ftz.f32 %f159, %f157, %f158;
mov.b32 %r619, %f159;
shfl.sync.bfly.b32 %r621|%p45, %r619, %r594, %r589, %r591;
mov.b32 %f160, %r621;
add.ftz.f32 %f161, %f159, %f160;
mov.b32 %r622, %f161;
shfl.sync.bfly.b32 %r624|%p46, %r622, %r597, %r589, %r591;
mov.b32 %f162, %r624;
add.ftz.f32 %f163, %f161, %f162;
mov.b32 %r625, %f163;
shfl.sync.bfly.b32 %r626|%p47, %r625, %r587, %r589, %r591;
mov.b32 %f164, %r626;
add.ftz.f32 %f165, %f163, %f164;
mov.b32 %r627, %f165;
shfl.sync.bfly.b32 %r629|%p48, %r627, %r582, %r589, %r591;
mov.b32 %f166, %r629;
add.ftz.f32 %f36, %f165, %f166;
@%p21 bra $L__BB22_18;

st.global.f32 [%rd15+4], %f36;

$L__BB22_18:
mul.lo.s32 %r20, %r24, 3;
mov.u32 %r634, 2;
ld.shared.f32 %f167, [%r471+768];
add.ftz.f32 %f168, %f167, %f224;
mov.b32 %r635, %f168;
mov.u32 %r636, 31;
mov.u32 %r637, 16;
mov.u32 %r638, -1;
shfl.sync.bfly.b32 %r639|%p50, %r635, %r637, %r636, %r638;
mov.b32 %f169, %r639;
add.ftz.f32 %f170, %f168, %f169;
mov.b32 %r640, %f170;
mov.u32 %r641, 8;
shfl.sync.bfly.b32 %r642|%p51, %r640, %r641, %r636, %r638;
mov.b32 %f171, %r642;
add.ftz.f32 %f172, %f170, %f171;
mov.b32 %r643, %f172;
mov.u32 %r644, 4;
shfl.sync.bfly.b32 %r645|%p52, %r643, %r644, %r636, %r638;
mov.b32 %f173, %r645;
add.ftz.f32 %f174, %f172, %f173;
mov.b32 %r646, %f174;
shfl.sync.bfly.b32 %r647|%p53, %r646, %r634, %r636, %r638;
mov.b32 %f175, %r647;
add.ftz.f32 %f176, %f174, %f175;
mov.b32 %r648, %f176;
mov.u32 %r649, 1;
shfl.sync.bfly.b32 %r650|%p54, %r648, %r649, %r636, %r638;
mov.b32 %f177, %r650;
add.ftz.f32 %f37, %f176, %f177;
@%p13 bra $L__BB22_20;

cvt.u64.u32 %rd75, %r20;
add.s64 %rd77, %rd75, %rd71;
shl.b64 %rd79, %rd77, 2;
add.s64 %rd80, %rd73, %rd79;
st.global.f32 [%rd80], %f37;

$L__BB22_20:
ld.shared.f32 %f178, [%r471+896];
add.ftz.f32 %f179, %f178, %f223;
mov.b32 %r664, %f179;
shfl.sync.bfly.b32 %r668|%p58, %r664, %r637, %r636, %r638;
mov.b32 %f180, %r668;
add.ftz.f32 %f181, %f179, %f180;
mov.b32 %r669, %f181;
shfl.sync.bfly.b32 %r671|%p59, %r669, %r641, %r636, %r638;
mov.b32 %f182, %r671;
add.ftz.f32 %f183, %f181, %f182;
mov.b32 %r672, %f183;
shfl.sync.bfly.b32 %r674|%p60, %r672, %r644, %r636, %r638;
mov.b32 %f184, %r674;
add.ftz.f32 %f185, %f183, %f184;
mov.b32 %r675, %f185;
shfl.sync.bfly.b32 %r676|%p61, %r675, %r634, %r636, %r638;
mov.b32 %f186, %r676;
add.ftz.f32 %f187, %f185, %f186;
mov.b32 %r677, %f187;
shfl.sync.bfly.b32 %r679|%p62, %r677, %r649, %r636, %r638;
mov.b32 %f188, %r679;
add.ftz.f32 %f38, %f187, %f188;
@%p21 bra $L__BB22_22;

add.s32 %r680, %r20, 1;
cvt.u64.u32 %rd81, %r680;
add.s64 %rd83, %rd81, %rd71;
shl.b64 %rd85, %rd83, 2;
add.s64 %rd86, %rd73, %rd85;
st.global.f32 [%rd86], %f38;

$L__BB22_22:
shl.b32 %r687, %r24, 2;
mov.u32 %r688, 2;
ld.shared.f32 %f189, [%r471+1024];
add.ftz.f32 %f190, %f189, %f222;
mov.b32 %r693, %f190;
mov.u32 %r694, 31;
mov.u32 %r695, 16;
mov.u32 %r696, -1;
shfl.sync.bfly.b32 %r697|%p64, %r693, %r695, %r694, %r696;
mov.b32 %f191, %r697;
add.ftz.f32 %f192, %f190, %f191;
mov.b32 %r698, %f192;
mov.u32 %r699, 8;
shfl.sync.bfly.b32 %r700|%p65, %r698, %r699, %r694, %r696;
mov.b32 %f193, %r700;
add.ftz.f32 %f194, %f192, %f193;
mov.b32 %r701, %f194;
mov.u32 %r702, 4;
shfl.sync.bfly.b32 %r703|%p66, %r701, %r702, %r694, %r696;
mov.b32 %f195, %r703;
add.ftz.f32 %f196, %f194, %f195;
mov.b32 %r704, %f196;
shfl.sync.bfly.b32 %r705|%p67, %r704, %r688, %r694, %r696;
mov.b32 %f197, %r705;
add.ftz.f32 %f198, %f196, %f197;
mov.b32 %r706, %f198;
mov.u32 %r707, 1;
shfl.sync.bfly.b32 %r708|%p68, %r706, %r707, %r694, %r696;
mov.b32 %f199, %r708;
add.ftz.f32 %f39, %f198, %f199;
cvt.u64.u32 %rd87, %r687;
add.s64 %rd89, %rd87, %rd71;
shl.b64 %rd91, %rd89, 2;
add.s64 %rd16, %rd73, %rd91;
@%p13 bra $L__BB22_24;

st.global.f32 [%rd16], %f39;

$L__BB22_24:
ld.shared.f32 %f200, [%r471+1152];
add.ftz.f32 %f201, %f200, %f221;
mov.b32 %r720, %f201;
shfl.sync.bfly.b32 %r724|%p72, %r720, %r695, %r694, %r696;
mov.b32 %f202, %r724;
add.ftz.f32 %f203, %f201, %f202;
mov.b32 %r725, %f203;
shfl.sync.bfly.b32 %r727|%p73, %r725, %r699, %r694, %r696;
mov.b32 %f204, %r727;
add.ftz.f32 %f205, %f203, %f204;
mov.b32 %r728, %f205;
shfl.sync.bfly.b32 %r730|%p74, %r728, %r702, %r694, %r696;
mov.b32 %f206, %r730;
add.ftz.f32 %f207, %f205, %f206;
mov.b32 %r731, %f207;
shfl.sync.bfly.b32 %r732|%p75, %r731, %r688, %r694, %r696;
mov.b32 %f208, %r732;
add.ftz.f32 %f209, %f207, %f208;
mov.b32 %r733, %f209;
shfl.sync.bfly.b32 %r735|%p76, %r733, %r707, %r694, %r696;
mov.b32 %f210, %r735;
add.ftz.f32 %f40, %f209, %f210;
@%p21 bra $L__BB22_26;

st.global.f32 [%rd16+4], %f40;

$L__BB22_26:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type42ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type42ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type42ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type42ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type42ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type42ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type42ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type42ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type42ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<92>;
.reg .b16 %rs<21>;
.reg .f32 %f<277>;
.reg .b32 %r<870>;
.reg .b64 %rd<112>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type42ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type42ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type42ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type42ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type42ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type42ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type42ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type42ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type42ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type42ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type42ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type42ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type42ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type42ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type42ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type42ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type42ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type42ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r41, %tid.y;
shl.b32 %r42, %r41, 5;
mov.u32 %r43, %tid.x;
add.s32 %r44, %r42, %r43;
shr.u32 %r45, %r44, 31;
add.s32 %r46, %r44, %r45;
shr.s32 %r869, %r46, 1;
shr.u32 %r47, %r25, 6;
setp.ge.s32 %p2, %r869, %r47;
mov.f32 %f265, 0f00000000;
mov.f32 %f266, %f265;
mov.f32 %f267, %f265;
mov.f32 %f268, %f265;
mov.f32 %f269, %f265;
mov.f32 %f270, %f265;
mov.f32 %f271, %f265;
mov.f32 %f272, %f265;
mov.f32 %f273, %f265;
mov.f32 %f274, %f265;
mov.f32 %f275, %f265;
mov.f32 %f276, %f265;
@%p2 bra $L__BB23_3;

mov.u32 %r48, %ctaid.y;
mov.u32 %r49, %ctaid.z;
mul.hi.u32 %r50, %r49, %r35;
add.s32 %r51, %r50, %r49;
shr.u32 %r52, %r51, %r36;
mul.lo.s32 %r53, %r52, %r38;
mul.hi.u32 %r54, %r48, %r29;
add.s32 %r55, %r54, %r48;
shr.u32 %r56, %r55, %r30;
mul.lo.s32 %r57, %r56, %r32;
mul.lo.s32 %r58, %r48, %r33;
cvt.u64.u32 %rd20, %r58;
mul.lo.s32 %r59, %r49, %r39;
cvt.u64.u32 %rd21, %r59;
and.b32 %r66, %r46, -2;
sub.s32 %r67, %r44, %r66;
cvt.s64.s32 %rd22, %r67;
add.s64 %rd23, %rd20, %rd22;
add.s64 %rd1, %rd23, %rd21;
shl.b32 %r68, %r67, 2;
mov.u32 %r69, %ctaid.x;
shl.b32 %r70, %r69, 1;
or.b32 %r71, %r70, 1;
shl.b32 %r72, %r27, 1;
shl.b32 %r863, %r869, 1;
add.s32 %r868, %r863, %r72;
mad.lo.s32 %r867, %r27, 3, %r863;
shl.b32 %r73, %r27, 2;
add.s32 %r866, %r863, %r73;
mad.lo.s32 %r865, %r27, 5, %r863;
add.s32 %r864, %r27, %r863;
mad.lo.s32 %r74, %r70, %r26, %r869;
add.s32 %r75, %r74, %r53;
add.s32 %r76, %r75, %r57;
mul.wide.s32 %rd24, %r76, 18;
mul.wide.s32 %rd25, %r68, 2;
add.s64 %rd26, %rd24, %rd25;
cvta.to.global.u64 %rd27, %rd17;
add.s64 %rd28, %rd27, %rd26;
add.s64 %rd111, %rd28, 8;
add.s64 %rd110, %rd27, %rd24;
add.s32 %r77, %r869, %r53;
add.s32 %r78, %r77, %r57;
mad.lo.s32 %r79, %r71, %r26, %r78;
mul.wide.s32 %rd29, %r79, 18;
add.s64 %rd30, %rd29, %rd25;
add.s64 %rd31, %rd27, %rd30;
add.s64 %rd109, %rd31, 4;
add.s64 %rd108, %rd27, %rd29;
cvta.to.global.u64 %rd6, %rd18;
mov.f32 %f265, 0f00000000;

$L__BB23_2:
cvt.u64.u32 %rd32, %r863;
add.s64 %rd33, %rd1, %rd32;
mul.lo.s64 %rd34, %rd33, 36;
ld.global.nc.u16 %rs1, [%rd110];

{ cvt.f32.f16 %f73, %rs1;}


ld.global.nc.u16 %rs13, [%rd111+-6];
cvt.s32.s16 %r476, %rs13;
add.s64 %rd35, %rd6, %rd34;
ld.global.nc.u32 %r81, [%rd35+4];
ld.global.nc.u32 %r85, [%rd35+8];
and.b32 %r477, %r476, 30583;
mov.u32 %r478, 33620223;
prmt.b32 %r479, %r478, %r478, %r477;
shr.u32 %r480, %r476, 2;
and.b32 %r481, %r480, 30583;
prmt.b32 %r482, %r478, %r478, %r481;
mov.u32 %r483, 20800;
prmt.b32 %r412, %r479, %r482, %r483;
mov.u32 %r484, 29538;
prmt.b32 %r416, %r479, %r482, %r484;
mov.u32 %r446, 0;

dp4a.s32.s32 %r80, %r81, %r412, %r446;


dp4a.s32.s32 %r84, %r85, %r416, %r80;

ld.global.nc.u16 %rs14, [%rd111+-4];
cvt.s32.s16 %r485, %rs14;
ld.global.nc.u32 %r89, [%rd35+12];
ld.global.nc.u32 %r93, [%rd35+16];
and.b32 %r486, %r485, 30583;
prmt.b32 %r487, %r478, %r478, %r486;
shr.u32 %r488, %r485, 2;
and.b32 %r489, %r488, 30583;
prmt.b32 %r490, %r478, %r478, %r489;
prmt.b32 %r420, %r487, %r490, %r483;
prmt.b32 %r424, %r487, %r490, %r484;

dp4a.s32.s32 %r88, %r89, %r420, %r84;


dp4a.s32.s32 %r92, %r93, %r424, %r88;

ld.global.nc.u16 %rs15, [%rd111+-2];
cvt.s32.s16 %r491, %rs15;
ld.global.nc.u32 %r97, [%rd35+20];
ld.global.nc.u32 %r101, [%rd35+24];
and.b32 %r492, %r491, 30583;
prmt.b32 %r493, %r478, %r478, %r492;
shr.u32 %r494, %r491, 2;
and.b32 %r495, %r494, 30583;
prmt.b32 %r496, %r478, %r478, %r495;
prmt.b32 %r428, %r493, %r496, %r483;
prmt.b32 %r432, %r493, %r496, %r484;

dp4a.s32.s32 %r96, %r97, %r428, %r92;


dp4a.s32.s32 %r100, %r101, %r432, %r96;

ld.global.nc.u16 %rs16, [%rd111];
cvt.s32.s16 %r497, %rs16;
ld.global.nc.u32 %r105, [%rd35+28];
ld.global.nc.u32 %r109, [%rd35+32];
and.b32 %r498, %r497, 30583;
prmt.b32 %r499, %r478, %r478, %r498;
shr.u32 %r500, %r497, 2;
and.b32 %r501, %r500, 30583;
prmt.b32 %r502, %r478, %r478, %r501;
prmt.b32 %r436, %r499, %r502, %r483;
prmt.b32 %r440, %r499, %r502, %r484;

dp4a.s32.s32 %r104, %r105, %r436, %r100;


dp4a.s32.s32 %r108, %r109, %r440, %r104;

ld.global.nc.u32 %r112, [%rd35];

{.reg .f16 low,high;
mov.b32 {low,high},%r112;
cvt.f32.f16 %f74, low;}


mul.ftz.f32 %f97, %f73, %f74;
cvt.rn.f32.s32 %f98, %r108;
fma.rn.ftz.f32 %f276, %f97, %f98, %f276;
ld.global.nc.u16 %rs2, [%rd108];

{ cvt.f32.f16 %f75, %rs2;}


ld.global.nc.u16 %rs17, [%rd109+-2];
cvt.s32.s16 %r503, %rs17;
and.b32 %r504, %r503, 30583;
prmt.b32 %r505, %r478, %r478, %r504;
shr.u32 %r506, %r503, 2;
and.b32 %r507, %r506, 30583;
prmt.b32 %r508, %r478, %r478, %r507;
prmt.b32 %r445, %r505, %r508, %r483;
prmt.b32 %r449, %r505, %r508, %r484;

dp4a.s32.s32 %r113, %r81, %r445, %r446;


dp4a.s32.s32 %r117, %r85, %r449, %r113;

ld.global.nc.u16 %rs18, [%rd109];
cvt.s32.s16 %r509, %rs18;
and.b32 %r510, %r509, 30583;
prmt.b32 %r511, %r478, %r478, %r510;
shr.u32 %r512, %r509, 2;
and.b32 %r513, %r512, 30583;
prmt.b32 %r514, %r478, %r478, %r513;
prmt.b32 %r453, %r511, %r514, %r483;
prmt.b32 %r457, %r511, %r514, %r484;

dp4a.s32.s32 %r121, %r89, %r453, %r117;


dp4a.s32.s32 %r125, %r93, %r457, %r121;

ld.global.nc.u16 %rs19, [%rd109+2];
cvt.s32.s16 %r515, %rs19;
and.b32 %r516, %r515, 30583;
prmt.b32 %r517, %r478, %r478, %r516;
shr.u32 %r518, %r515, 2;
and.b32 %r519, %r518, 30583;
prmt.b32 %r520, %r478, %r478, %r519;
prmt.b32 %r461, %r517, %r520, %r483;
prmt.b32 %r465, %r517, %r520, %r484;

dp4a.s32.s32 %r129, %r97, %r461, %r125;


dp4a.s32.s32 %r133, %r101, %r465, %r129;

ld.global.nc.u16 %rs20, [%rd109+4];
cvt.s32.s16 %r521, %rs20;
and.b32 %r522, %r521, 30583;
prmt.b32 %r523, %r478, %r478, %r522;
shr.u32 %r524, %r521, 2;
and.b32 %r525, %r524, 30583;
prmt.b32 %r526, %r478, %r478, %r525;
prmt.b32 %r469, %r523, %r526, %r483;
prmt.b32 %r473, %r523, %r526, %r484;

dp4a.s32.s32 %r137, %r105, %r469, %r133;


dp4a.s32.s32 %r141, %r109, %r473, %r137;


{.reg .f16 low,high;
mov.b32 {low,high},%r112;
cvt.f32.f16 %f76, low;}


mul.ftz.f32 %f99, %f75, %f76;
cvt.rn.f32.s32 %f100, %r141;
fma.rn.ftz.f32 %f275, %f99, %f100, %f275;
cvt.u64.u32 %rd36, %r864;
add.s64 %rd37, %rd1, %rd36;
mul.lo.s64 %rd38, %rd37, 36;

{ cvt.f32.f16 %f77, %rs1;}


add.s64 %rd39, %rd6, %rd38;
ld.global.nc.u32 %r147, [%rd39+4];
ld.global.nc.u32 %r151, [%rd39+8];

dp4a.s32.s32 %r146, %r147, %r412, %r446;


dp4a.s32.s32 %r150, %r151, %r416, %r146;

ld.global.nc.u32 %r155, [%rd39+12];
ld.global.nc.u32 %r159, [%rd39+16];

dp4a.s32.s32 %r154, %r155, %r420, %r150;


dp4a.s32.s32 %r158, %r159, %r424, %r154;

ld.global.nc.u32 %r163, [%rd39+20];
ld.global.nc.u32 %r167, [%rd39+24];

dp4a.s32.s32 %r162, %r163, %r428, %r158;


dp4a.s32.s32 %r166, %r167, %r432, %r162;

ld.global.nc.u32 %r171, [%rd39+28];
ld.global.nc.u32 %r175, [%rd39+32];

dp4a.s32.s32 %r170, %r171, %r436, %r166;


dp4a.s32.s32 %r174, %r175, %r440, %r170;

ld.global.nc.u32 %r178, [%rd39];

{.reg .f16 low,high;
mov.b32 {low,high},%r178;
cvt.f32.f16 %f78, low;}


mul.ftz.f32 %f101, %f77, %f78;
cvt.rn.f32.s32 %f102, %r174;
fma.rn.ftz.f32 %f274, %f101, %f102, %f274;

{ cvt.f32.f16 %f79, %rs2;}



dp4a.s32.s32 %r179, %r147, %r445, %r446;


dp4a.s32.s32 %r183, %r151, %r449, %r179;


dp4a.s32.s32 %r187, %r155, %r453, %r183;


dp4a.s32.s32 %r191, %r159, %r457, %r187;


dp4a.s32.s32 %r195, %r163, %r461, %r191;


dp4a.s32.s32 %r199, %r167, %r465, %r195;


dp4a.s32.s32 %r203, %r171, %r469, %r199;


dp4a.s32.s32 %r207, %r175, %r473, %r203;


{.reg .f16 low,high;
mov.b32 {low,high},%r178;
cvt.f32.f16 %f80, low;}


mul.ftz.f32 %f103, %f79, %f80;
cvt.rn.f32.s32 %f104, %r207;
fma.rn.ftz.f32 %f273, %f103, %f104, %f273;
cvt.u64.u32 %rd40, %r868;
add.s64 %rd41, %rd1, %rd40;
mul.lo.s64 %rd42, %rd41, 36;

{ cvt.f32.f16 %f81, %rs1;}


add.s64 %rd43, %rd6, %rd42;
ld.global.nc.u32 %r213, [%rd43+4];
ld.global.nc.u32 %r217, [%rd43+8];

dp4a.s32.s32 %r212, %r213, %r412, %r446;


dp4a.s32.s32 %r216, %r217, %r416, %r212;

ld.global.nc.u32 %r221, [%rd43+12];
ld.global.nc.u32 %r225, [%rd43+16];

dp4a.s32.s32 %r220, %r221, %r420, %r216;


dp4a.s32.s32 %r224, %r225, %r424, %r220;

ld.global.nc.u32 %r229, [%rd43+20];
ld.global.nc.u32 %r233, [%rd43+24];

dp4a.s32.s32 %r228, %r229, %r428, %r224;


dp4a.s32.s32 %r232, %r233, %r432, %r228;

ld.global.nc.u32 %r237, [%rd43+28];
ld.global.nc.u32 %r241, [%rd43+32];

dp4a.s32.s32 %r236, %r237, %r436, %r232;


dp4a.s32.s32 %r240, %r241, %r440, %r236;

ld.global.nc.u32 %r244, [%rd43];

{.reg .f16 low,high;
mov.b32 {low,high},%r244;
cvt.f32.f16 %f82, low;}


mul.ftz.f32 %f105, %f81, %f82;
cvt.rn.f32.s32 %f106, %r240;
fma.rn.ftz.f32 %f272, %f105, %f106, %f272;

{ cvt.f32.f16 %f83, %rs2;}



dp4a.s32.s32 %r245, %r213, %r445, %r446;


dp4a.s32.s32 %r249, %r217, %r449, %r245;


dp4a.s32.s32 %r253, %r221, %r453, %r249;


dp4a.s32.s32 %r257, %r225, %r457, %r253;


dp4a.s32.s32 %r261, %r229, %r461, %r257;


dp4a.s32.s32 %r265, %r233, %r465, %r261;


dp4a.s32.s32 %r269, %r237, %r469, %r265;


dp4a.s32.s32 %r273, %r241, %r473, %r269;


{.reg .f16 low,high;
mov.b32 {low,high},%r244;
cvt.f32.f16 %f84, low;}


mul.ftz.f32 %f107, %f83, %f84;
cvt.rn.f32.s32 %f108, %r273;
fma.rn.ftz.f32 %f271, %f107, %f108, %f271;
cvt.u64.u32 %rd44, %r867;
add.s64 %rd45, %rd1, %rd44;
mul.lo.s64 %rd46, %rd45, 36;

{ cvt.f32.f16 %f85, %rs1;}


add.s64 %rd47, %rd6, %rd46;
ld.global.nc.u32 %r279, [%rd47+4];
ld.global.nc.u32 %r283, [%rd47+8];

dp4a.s32.s32 %r278, %r279, %r412, %r446;


dp4a.s32.s32 %r282, %r283, %r416, %r278;

ld.global.nc.u32 %r287, [%rd47+12];
ld.global.nc.u32 %r291, [%rd47+16];

dp4a.s32.s32 %r286, %r287, %r420, %r282;


dp4a.s32.s32 %r290, %r291, %r424, %r286;

ld.global.nc.u32 %r295, [%rd47+20];
ld.global.nc.u32 %r299, [%rd47+24];

dp4a.s32.s32 %r294, %r295, %r428, %r290;


dp4a.s32.s32 %r298, %r299, %r432, %r294;

ld.global.nc.u32 %r303, [%rd47+28];
ld.global.nc.u32 %r307, [%rd47+32];

dp4a.s32.s32 %r302, %r303, %r436, %r298;


dp4a.s32.s32 %r306, %r307, %r440, %r302;

ld.global.nc.u32 %r310, [%rd47];

{.reg .f16 low,high;
mov.b32 {low,high},%r310;
cvt.f32.f16 %f86, low;}


mul.ftz.f32 %f109, %f85, %f86;
cvt.rn.f32.s32 %f110, %r306;
fma.rn.ftz.f32 %f270, %f109, %f110, %f270;

{ cvt.f32.f16 %f87, %rs2;}



dp4a.s32.s32 %r311, %r279, %r445, %r446;


dp4a.s32.s32 %r315, %r283, %r449, %r311;


dp4a.s32.s32 %r319, %r287, %r453, %r315;


dp4a.s32.s32 %r323, %r291, %r457, %r319;


dp4a.s32.s32 %r327, %r295, %r461, %r323;


dp4a.s32.s32 %r331, %r299, %r465, %r327;


dp4a.s32.s32 %r335, %r303, %r469, %r331;


dp4a.s32.s32 %r339, %r307, %r473, %r335;


{.reg .f16 low,high;
mov.b32 {low,high},%r310;
cvt.f32.f16 %f88, low;}


mul.ftz.f32 %f111, %f87, %f88;
cvt.rn.f32.s32 %f112, %r339;
fma.rn.ftz.f32 %f269, %f111, %f112, %f269;
cvt.u64.u32 %rd48, %r866;
add.s64 %rd49, %rd1, %rd48;
mul.lo.s64 %rd50, %rd49, 36;

{ cvt.f32.f16 %f89, %rs1;}


add.s64 %rd51, %rd6, %rd50;
ld.global.nc.u32 %r345, [%rd51+4];
ld.global.nc.u32 %r349, [%rd51+8];

dp4a.s32.s32 %r344, %r345, %r412, %r446;


dp4a.s32.s32 %r348, %r349, %r416, %r344;

ld.global.nc.u32 %r353, [%rd51+12];
ld.global.nc.u32 %r357, [%rd51+16];

dp4a.s32.s32 %r352, %r353, %r420, %r348;


dp4a.s32.s32 %r356, %r357, %r424, %r352;

ld.global.nc.u32 %r361, [%rd51+20];
ld.global.nc.u32 %r365, [%rd51+24];

dp4a.s32.s32 %r360, %r361, %r428, %r356;


dp4a.s32.s32 %r364, %r365, %r432, %r360;

ld.global.nc.u32 %r369, [%rd51+28];
ld.global.nc.u32 %r373, [%rd51+32];

dp4a.s32.s32 %r368, %r369, %r436, %r364;


dp4a.s32.s32 %r372, %r373, %r440, %r368;

ld.global.nc.u32 %r376, [%rd51];

{.reg .f16 low,high;
mov.b32 {low,high},%r376;
cvt.f32.f16 %f90, low;}


mul.ftz.f32 %f113, %f89, %f90;
cvt.rn.f32.s32 %f114, %r372;
fma.rn.ftz.f32 %f268, %f113, %f114, %f268;

{ cvt.f32.f16 %f91, %rs2;}



dp4a.s32.s32 %r377, %r345, %r445, %r446;


dp4a.s32.s32 %r381, %r349, %r449, %r377;


dp4a.s32.s32 %r385, %r353, %r453, %r381;


dp4a.s32.s32 %r389, %r357, %r457, %r385;


dp4a.s32.s32 %r393, %r361, %r461, %r389;


dp4a.s32.s32 %r397, %r365, %r465, %r393;


dp4a.s32.s32 %r401, %r369, %r469, %r397;


dp4a.s32.s32 %r405, %r373, %r473, %r401;


{.reg .f16 low,high;
mov.b32 {low,high},%r376;
cvt.f32.f16 %f92, low;}


mul.ftz.f32 %f115, %f91, %f92;
cvt.rn.f32.s32 %f116, %r405;
fma.rn.ftz.f32 %f267, %f115, %f116, %f267;
cvt.u64.u32 %rd52, %r865;
add.s64 %rd53, %rd1, %rd52;
mul.lo.s64 %rd54, %rd53, 36;

{ cvt.f32.f16 %f93, %rs1;}


add.s64 %rd55, %rd6, %rd54;
ld.global.nc.u32 %r411, [%rd55+4];
ld.global.nc.u32 %r415, [%rd55+8];

dp4a.s32.s32 %r410, %r411, %r412, %r446;


dp4a.s32.s32 %r414, %r415, %r416, %r410;

ld.global.nc.u32 %r419, [%rd55+12];
ld.global.nc.u32 %r423, [%rd55+16];

dp4a.s32.s32 %r418, %r419, %r420, %r414;


dp4a.s32.s32 %r422, %r423, %r424, %r418;

ld.global.nc.u32 %r427, [%rd55+20];
ld.global.nc.u32 %r431, [%rd55+24];

dp4a.s32.s32 %r426, %r427, %r428, %r422;


dp4a.s32.s32 %r430, %r431, %r432, %r426;

ld.global.nc.u32 %r435, [%rd55+28];
ld.global.nc.u32 %r439, [%rd55+32];

dp4a.s32.s32 %r434, %r435, %r436, %r430;


dp4a.s32.s32 %r438, %r439, %r440, %r434;

ld.global.nc.u32 %r442, [%rd55];

{.reg .f16 low,high;
mov.b32 {low,high},%r442;
cvt.f32.f16 %f94, low;}


mul.ftz.f32 %f117, %f93, %f94;
cvt.rn.f32.s32 %f118, %r438;
fma.rn.ftz.f32 %f266, %f117, %f118, %f266;

{ cvt.f32.f16 %f95, %rs2;}



dp4a.s32.s32 %r443, %r411, %r445, %r446;


dp4a.s32.s32 %r447, %r415, %r449, %r443;


dp4a.s32.s32 %r451, %r419, %r453, %r447;


dp4a.s32.s32 %r455, %r423, %r457, %r451;


dp4a.s32.s32 %r459, %r427, %r461, %r455;


dp4a.s32.s32 %r463, %r431, %r465, %r459;


dp4a.s32.s32 %r467, %r435, %r469, %r463;


dp4a.s32.s32 %r471, %r439, %r473, %r467;


{.reg .f16 low,high;
mov.b32 {low,high},%r442;
cvt.f32.f16 %f96, low;}


mul.ftz.f32 %f119, %f95, %f96;
cvt.rn.f32.s32 %f120, %r471;
fma.rn.ftz.f32 %f265, %f119, %f120, %f265;
add.s32 %r868, %r868, 64;
add.s32 %r867, %r867, 64;
add.s32 %r866, %r866, 64;
add.s32 %r865, %r865, 64;
add.s32 %r864, %r864, 64;
add.s32 %r863, %r863, 64;
add.s64 %rd111, %rd111, 576;
add.s64 %rd110, %rd110, 576;
add.s64 %rd109, %rd109, 576;
add.s64 %rd108, %rd108, 576;
add.s32 %r869, %r869, 32;
setp.lt.s32 %p3, %r869, %r47;
@%p3 bra $L__BB23_2;

$L__BB23_3:
setp.eq.s32 %p4, %r41, 0;
@%p4 bra $L__BB23_5;

mov.u32 %r530, _ZZ13mul_mat_vec_qIL9ggml_type42ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r531, %r41, 1536, %r530;
shl.b32 %r532, %r43, 2;
add.s32 %r533, %r531, %r532;
st.shared.f32 [%r533+-1536], %f276;
st.shared.f32 [%r533+-1408], %f275;
st.shared.f32 [%r533+-1280], %f274;
st.shared.f32 [%r533+-1152], %f273;
st.shared.f32 [%r533+-1024], %f272;
st.shared.f32 [%r533+-896], %f271;
st.shared.f32 [%r533+-768], %f270;
st.shared.f32 [%r533+-640], %f269;
st.shared.f32 [%r533+-512], %f268;
st.shared.f32 [%r533+-384], %f267;
st.shared.f32 [%r533+-256], %f266;
st.shared.f32 [%r533+-128], %f265;

$L__BB23_5:
setp.ne.s32 %p5, %r41, 0;
bar.sync 0;
@%p5 bra $L__BB23_30;

mov.u32 %r535, %ctaid.x;
shl.b32 %r536, %r535, 1;
mov.u32 %r537, 1;
shl.b32 %r539, %r43, 2;
mov.u32 %r540, _ZZ13mul_mat_vec_qIL9ggml_type42ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r541, %r540, %r539;
mov.u32 %r542, 2;
ld.shared.f32 %f121, [%r541];
add.ftz.f32 %f122, %f121, %f276;
mov.b32 %r543, %f122;
mov.u32 %r544, 31;
mov.u32 %r545, 16;
mov.u32 %r546, -1;
shfl.sync.bfly.b32 %r547|%p6, %r543, %r545, %r544, %r546;
mov.b32 %f123, %r547;
add.ftz.f32 %f124, %f122, %f123;
mov.b32 %r548, %f124;
mov.u32 %r549, 8;
shfl.sync.bfly.b32 %r550|%p7, %r548, %r549, %r544, %r546;
mov.b32 %f125, %r550;
add.ftz.f32 %f126, %f124, %f125;
mov.b32 %r551, %f126;
mov.u32 %r552, 4;
shfl.sync.bfly.b32 %r553|%p8, %r551, %r552, %r544, %r546;
mov.b32 %f127, %r553;
add.ftz.f32 %f128, %f126, %f127;
mov.b32 %r554, %f128;
shfl.sync.bfly.b32 %r555|%p9, %r554, %r542, %r544, %r546;
mov.b32 %f129, %r555;
add.ftz.f32 %f130, %f128, %f129;
mov.b32 %r556, %f130;
shfl.sync.bfly.b32 %r557|%p10, %r556, %r537, %r544, %r546;
mov.b32 %f131, %r557;
add.ftz.f32 %f37, %f130, %f131;
setp.ne.s32 %p11, %r43, 0;
setp.ge.u32 %p12, %r536, %r28;
or.pred %p13, %p11, %p12;
@%p13 bra $L__BB23_8;

cvta.to.global.u64 %rd56, %rd19;
mov.u32 %r558, %ctaid.y;
mad.lo.s32 %r561, %r558, %r34, %r536;
mov.u32 %r562, %ctaid.z;
mad.lo.s32 %r563, %r562, %r40, %r561;
mul.wide.u32 %rd57, %r563, 4;
add.s64 %rd58, %rd56, %rd57;
st.global.f32 [%rd58], %f37;

$L__BB23_8:
ld.shared.f32 %f132, [%r541+128];
add.ftz.f32 %f133, %f132, %f275;
mov.b32 %r569, %f133;
shfl.sync.bfly.b32 %r573|%p14, %r569, %r545, %r544, %r546;
mov.b32 %f134, %r573;
add.ftz.f32 %f135, %f133, %f134;
mov.b32 %r574, %f135;
shfl.sync.bfly.b32 %r576|%p15, %r574, %r549, %r544, %r546;
mov.b32 %f136, %r576;
add.ftz.f32 %f137, %f135, %f136;
mov.b32 %r577, %f137;
shfl.sync.bfly.b32 %r579|%p16, %r577, %r552, %r544, %r546;
mov.b32 %f138, %r579;
add.ftz.f32 %f139, %f137, %f138;
mov.b32 %r580, %f139;
shfl.sync.bfly.b32 %r581|%p17, %r580, %r542, %r544, %r546;
mov.b32 %f140, %r581;
add.ftz.f32 %f141, %f139, %f140;
mov.b32 %r582, %f141;
shfl.sync.bfly.b32 %r584|%p18, %r582, %r537, %r544, %r546;
mov.b32 %f142, %r584;
add.ftz.f32 %f38, %f141, %f142;
setp.eq.s32 %p19, %r43, 1;
or.b32 %r587, %r536, 1;
setp.lt.u32 %p20, %r587, %r28;
and.pred %p1, %p19, %p20;
not.pred %p21, %p1;
@%p21 bra $L__BB23_10;

cvta.to.global.u64 %rd59, %rd19;
mov.u32 %r588, %ctaid.y;
mad.lo.s32 %r591, %r588, %r34, %r536;
mov.u32 %r592, %ctaid.z;
mad.lo.s32 %r593, %r592, %r40, %r591;
mul.wide.u32 %rd60, %r593, 4;
add.s64 %rd61, %rd59, %rd60;
st.global.f32 [%rd61+4], %f38;

$L__BB23_10:
mov.u32 %r598, 2;
ld.shared.f32 %f143, [%r541+256];
add.ftz.f32 %f144, %f143, %f274;
mov.b32 %r599, %f144;
mov.u32 %r600, 31;
mov.u32 %r601, 16;
mov.u32 %r602, -1;
shfl.sync.bfly.b32 %r603|%p22, %r599, %r601, %r600, %r602;
mov.b32 %f145, %r603;
add.ftz.f32 %f146, %f144, %f145;
mov.b32 %r604, %f146;
mov.u32 %r605, 8;
shfl.sync.bfly.b32 %r606|%p23, %r604, %r605, %r600, %r602;
mov.b32 %f147, %r606;
add.ftz.f32 %f148, %f146, %f147;
mov.b32 %r607, %f148;
mov.u32 %r608, 4;
shfl.sync.bfly.b32 %r609|%p24, %r607, %r608, %r600, %r602;
mov.b32 %f149, %r609;
add.ftz.f32 %f150, %f148, %f149;
mov.b32 %r610, %f150;
shfl.sync.bfly.b32 %r611|%p25, %r610, %r598, %r600, %r602;
mov.b32 %f151, %r611;
add.ftz.f32 %f152, %f150, %f151;
mov.b32 %r612, %f152;
mov.u32 %r613, 1;
shfl.sync.bfly.b32 %r614|%p26, %r612, %r613, %r600, %r602;
mov.b32 %f153, %r614;
add.ftz.f32 %f39, %f152, %f153;
@%p13 bra $L__BB23_12;

cvt.u64.u32 %rd62, %r28;
mov.u32 %r617, %ctaid.y;
mad.lo.s32 %r620, %r617, %r34, %r536;
mov.u32 %r621, %ctaid.z;
mad.lo.s32 %r622, %r621, %r40, %r620;
cvt.u64.u32 %rd63, %r622;
add.s64 %rd64, %rd62, %rd63;
cvta.to.global.u64 %rd65, %rd19;
shl.b64 %rd66, %rd64, 2;
add.s64 %rd67, %rd65, %rd66;
st.global.f32 [%rd67], %f39;

$L__BB23_12:
ld.shared.f32 %f154, [%r541+384];
add.ftz.f32 %f155, %f154, %f273;
mov.b32 %r628, %f155;
shfl.sync.bfly.b32 %r632|%p30, %r628, %r601, %r600, %r602;
mov.b32 %f156, %r632;
add.ftz.f32 %f157, %f155, %f156;
mov.b32 %r633, %f157;
shfl.sync.bfly.b32 %r635|%p31, %r633, %r605, %r600, %r602;
mov.b32 %f158, %r635;
add.ftz.f32 %f159, %f157, %f158;
mov.b32 %r636, %f159;
shfl.sync.bfly.b32 %r638|%p32, %r636, %r608, %r600, %r602;
mov.b32 %f160, %r638;
add.ftz.f32 %f161, %f159, %f160;
mov.b32 %r639, %f161;
shfl.sync.bfly.b32 %r640|%p33, %r639, %r598, %r600, %r602;
mov.b32 %f162, %r640;
add.ftz.f32 %f163, %f161, %f162;
mov.b32 %r641, %f163;
shfl.sync.bfly.b32 %r643|%p34, %r641, %r613, %r600, %r602;
mov.b32 %f164, %r643;
add.ftz.f32 %f40, %f163, %f164;
@%p21 bra $L__BB23_14;

add.s32 %r644, %r28, 1;
cvt.u64.u32 %rd68, %r644;
mov.u32 %r645, %ctaid.y;
mad.lo.s32 %r648, %r645, %r34, %r536;
mov.u32 %r649, %ctaid.z;
mad.lo.s32 %r650, %r649, %r40, %r648;
cvt.u64.u32 %rd69, %r650;
add.s64 %rd70, %rd68, %rd69;
cvta.to.global.u64 %rd71, %rd19;
shl.b64 %rd72, %rd70, 2;
add.s64 %rd73, %rd71, %rd72;
st.global.f32 [%rd73], %f40;

$L__BB23_14:
shl.b32 %r651, %r28, 1;
mov.u32 %r652, 1;
mov.u32 %r657, 2;
ld.shared.f32 %f165, [%r541+512];
add.ftz.f32 %f166, %f165, %f272;
mov.b32 %r658, %f166;
mov.u32 %r659, 31;
mov.u32 %r660, 16;
mov.u32 %r661, -1;
shfl.sync.bfly.b32 %r662|%p36, %r658, %r660, %r659, %r661;
mov.b32 %f167, %r662;
add.ftz.f32 %f168, %f166, %f167;
mov.b32 %r663, %f168;
mov.u32 %r664, 8;
shfl.sync.bfly.b32 %r665|%p37, %r663, %r664, %r659, %r661;
mov.b32 %f169, %r665;
add.ftz.f32 %f170, %f168, %f169;
mov.b32 %r666, %f170;
mov.u32 %r667, 4;
shfl.sync.bfly.b32 %r668|%p38, %r666, %r667, %r659, %r661;
mov.b32 %f171, %r668;
add.ftz.f32 %f172, %f170, %f171;
mov.b32 %r669, %f172;
shfl.sync.bfly.b32 %r670|%p39, %r669, %r657, %r659, %r661;
mov.b32 %f173, %r670;
add.ftz.f32 %f174, %f172, %f173;
mov.b32 %r671, %f174;
shfl.sync.bfly.b32 %r672|%p40, %r671, %r652, %r659, %r661;
mov.b32 %f175, %r672;
add.ftz.f32 %f41, %f174, %f175;
cvt.u64.u32 %rd74, %r651;
mov.u32 %r673, %ctaid.y;
mad.lo.s32 %r676, %r673, %r34, %r536;
mov.u32 %r677, %ctaid.z;
mad.lo.s32 %r678, %r677, %r40, %r676;
cvt.u64.u32 %rd75, %r678;
add.s64 %rd76, %rd74, %rd75;
cvta.to.global.u64 %rd77, %rd19;
shl.b64 %rd78, %rd76, 2;
add.s64 %rd15, %rd77, %rd78;
@%p13 bra $L__BB23_16;

st.global.f32 [%rd15], %f41;

$L__BB23_16:
ld.shared.f32 %f176, [%r541+640];
add.ftz.f32 %f177, %f176, %f271;
mov.b32 %r684, %f177;
shfl.sync.bfly.b32 %r688|%p44, %r684, %r660, %r659, %r661;
mov.b32 %f178, %r688;
add.ftz.f32 %f179, %f177, %f178;
mov.b32 %r689, %f179;
shfl.sync.bfly.b32 %r691|%p45, %r689, %r664, %r659, %r661;
mov.b32 %f180, %r691;
add.ftz.f32 %f181, %f179, %f180;
mov.b32 %r692, %f181;
shfl.sync.bfly.b32 %r694|%p46, %r692, %r667, %r659, %r661;
mov.b32 %f182, %r694;
add.ftz.f32 %f183, %f181, %f182;
mov.b32 %r695, %f183;
shfl.sync.bfly.b32 %r696|%p47, %r695, %r657, %r659, %r661;
mov.b32 %f184, %r696;
add.ftz.f32 %f185, %f183, %f184;
mov.b32 %r697, %f185;
shfl.sync.bfly.b32 %r699|%p48, %r697, %r652, %r659, %r661;
mov.b32 %f186, %r699;
add.ftz.f32 %f42, %f185, %f186;
@%p21 bra $L__BB23_18;

st.global.f32 [%rd15+4], %f42;

$L__BB23_18:
mul.lo.s32 %r23, %r28, 3;
mov.u32 %r704, 2;
ld.shared.f32 %f187, [%r541+768];
add.ftz.f32 %f188, %f187, %f270;
mov.b32 %r705, %f188;
mov.u32 %r706, 31;
mov.u32 %r707, 16;
mov.u32 %r708, -1;
shfl.sync.bfly.b32 %r709|%p50, %r705, %r707, %r706, %r708;
mov.b32 %f189, %r709;
add.ftz.f32 %f190, %f188, %f189;
mov.b32 %r710, %f190;
mov.u32 %r711, 8;
shfl.sync.bfly.b32 %r712|%p51, %r710, %r711, %r706, %r708;
mov.b32 %f191, %r712;
add.ftz.f32 %f192, %f190, %f191;
mov.b32 %r713, %f192;
mov.u32 %r714, 4;
shfl.sync.bfly.b32 %r715|%p52, %r713, %r714, %r706, %r708;
mov.b32 %f193, %r715;
add.ftz.f32 %f194, %f192, %f193;
mov.b32 %r716, %f194;
shfl.sync.bfly.b32 %r717|%p53, %r716, %r704, %r706, %r708;
mov.b32 %f195, %r717;
add.ftz.f32 %f196, %f194, %f195;
mov.b32 %r718, %f196;
mov.u32 %r719, 1;
shfl.sync.bfly.b32 %r720|%p54, %r718, %r719, %r706, %r708;
mov.b32 %f197, %r720;
add.ftz.f32 %f43, %f196, %f197;
@%p13 bra $L__BB23_20;

cvt.u64.u32 %rd79, %r23;
add.s64 %rd81, %rd79, %rd75;
shl.b64 %rd83, %rd81, 2;
add.s64 %rd84, %rd77, %rd83;
st.global.f32 [%rd84], %f43;

$L__BB23_20:
ld.shared.f32 %f198, [%r541+896];
add.ftz.f32 %f199, %f198, %f269;
mov.b32 %r734, %f199;
shfl.sync.bfly.b32 %r738|%p58, %r734, %r707, %r706, %r708;
mov.b32 %f200, %r738;
add.ftz.f32 %f201, %f199, %f200;
mov.b32 %r739, %f201;
shfl.sync.bfly.b32 %r741|%p59, %r739, %r711, %r706, %r708;
mov.b32 %f202, %r741;
add.ftz.f32 %f203, %f201, %f202;
mov.b32 %r742, %f203;
shfl.sync.bfly.b32 %r744|%p60, %r742, %r714, %r706, %r708;
mov.b32 %f204, %r744;
add.ftz.f32 %f205, %f203, %f204;
mov.b32 %r745, %f205;
shfl.sync.bfly.b32 %r746|%p61, %r745, %r704, %r706, %r708;
mov.b32 %f206, %r746;
add.ftz.f32 %f207, %f205, %f206;
mov.b32 %r747, %f207;
shfl.sync.bfly.b32 %r749|%p62, %r747, %r719, %r706, %r708;
mov.b32 %f208, %r749;
add.ftz.f32 %f44, %f207, %f208;
@%p21 bra $L__BB23_22;

add.s32 %r750, %r23, 1;
cvt.u64.u32 %rd85, %r750;
add.s64 %rd87, %rd85, %rd75;
shl.b64 %rd89, %rd87, 2;
add.s64 %rd90, %rd77, %rd89;
st.global.f32 [%rd90], %f44;

$L__BB23_22:
shl.b32 %r757, %r28, 2;
mov.u32 %r758, 2;
ld.shared.f32 %f209, [%r541+1024];
add.ftz.f32 %f210, %f209, %f268;
mov.b32 %r763, %f210;
mov.u32 %r764, 31;
mov.u32 %r765, 16;
mov.u32 %r766, -1;
shfl.sync.bfly.b32 %r767|%p64, %r763, %r765, %r764, %r766;
mov.b32 %f211, %r767;
add.ftz.f32 %f212, %f210, %f211;
mov.b32 %r768, %f212;
mov.u32 %r769, 8;
shfl.sync.bfly.b32 %r770|%p65, %r768, %r769, %r764, %r766;
mov.b32 %f213, %r770;
add.ftz.f32 %f214, %f212, %f213;
mov.b32 %r771, %f214;
mov.u32 %r772, 4;
shfl.sync.bfly.b32 %r773|%p66, %r771, %r772, %r764, %r766;
mov.b32 %f215, %r773;
add.ftz.f32 %f216, %f214, %f215;
mov.b32 %r774, %f216;
shfl.sync.bfly.b32 %r775|%p67, %r774, %r758, %r764, %r766;
mov.b32 %f217, %r775;
add.ftz.f32 %f218, %f216, %f217;
mov.b32 %r776, %f218;
mov.u32 %r777, 1;
shfl.sync.bfly.b32 %r778|%p68, %r776, %r777, %r764, %r766;
mov.b32 %f219, %r778;
add.ftz.f32 %f45, %f218, %f219;
cvt.u64.u32 %rd91, %r757;
add.s64 %rd93, %rd91, %rd75;
shl.b64 %rd95, %rd93, 2;
add.s64 %rd16, %rd77, %rd95;
@%p13 bra $L__BB23_24;

st.global.f32 [%rd16], %f45;

$L__BB23_24:
ld.shared.f32 %f220, [%r541+1152];
add.ftz.f32 %f221, %f220, %f267;
mov.b32 %r790, %f221;
shfl.sync.bfly.b32 %r794|%p72, %r790, %r765, %r764, %r766;
mov.b32 %f222, %r794;
add.ftz.f32 %f223, %f221, %f222;
mov.b32 %r795, %f223;
shfl.sync.bfly.b32 %r797|%p73, %r795, %r769, %r764, %r766;
mov.b32 %f224, %r797;
add.ftz.f32 %f225, %f223, %f224;
mov.b32 %r798, %f225;
shfl.sync.bfly.b32 %r800|%p74, %r798, %r772, %r764, %r766;
mov.b32 %f226, %r800;
add.ftz.f32 %f227, %f225, %f226;
mov.b32 %r801, %f227;
shfl.sync.bfly.b32 %r802|%p75, %r801, %r758, %r764, %r766;
mov.b32 %f228, %r802;
add.ftz.f32 %f229, %f227, %f228;
mov.b32 %r803, %f229;
shfl.sync.bfly.b32 %r805|%p76, %r803, %r777, %r764, %r766;
mov.b32 %f230, %r805;
add.ftz.f32 %f46, %f229, %f230;
@%p21 bra $L__BB23_26;

st.global.f32 [%rd16+4], %f46;

$L__BB23_26:
mul.lo.s32 %r24, %r28, 5;
mov.u32 %r810, 2;
ld.shared.f32 %f231, [%r541+1280];
add.ftz.f32 %f232, %f231, %f266;
mov.b32 %r811, %f232;
mov.u32 %r812, 31;
mov.u32 %r813, 16;
mov.u32 %r814, -1;
shfl.sync.bfly.b32 %r815|%p78, %r811, %r813, %r812, %r814;
mov.b32 %f233, %r815;
add.ftz.f32 %f234, %f232, %f233;
mov.b32 %r816, %f234;
mov.u32 %r817, 8;
shfl.sync.bfly.b32 %r818|%p79, %r816, %r817, %r812, %r814;
mov.b32 %f235, %r818;
add.ftz.f32 %f236, %f234, %f235;
mov.b32 %r819, %f236;
mov.u32 %r820, 4;
shfl.sync.bfly.b32 %r821|%p80, %r819, %r820, %r812, %r814;
mov.b32 %f237, %r821;
add.ftz.f32 %f238, %f236, %f237;
mov.b32 %r822, %f238;
shfl.sync.bfly.b32 %r823|%p81, %r822, %r810, %r812, %r814;
mov.b32 %f239, %r823;
add.ftz.f32 %f240, %f238, %f239;
mov.b32 %r824, %f240;
mov.u32 %r825, 1;
shfl.sync.bfly.b32 %r826|%p82, %r824, %r825, %r812, %r814;
mov.b32 %f241, %r826;
add.ftz.f32 %f47, %f240, %f241;
@%p13 bra $L__BB23_28;

cvt.u64.u32 %rd96, %r24;
add.s64 %rd98, %rd96, %rd75;
shl.b64 %rd100, %rd98, 2;
add.s64 %rd101, %rd77, %rd100;
st.global.f32 [%rd101], %f47;

$L__BB23_28:
ld.shared.f32 %f242, [%r541+1408];
add.ftz.f32 %f243, %f242, %f265;
mov.b32 %r840, %f243;
shfl.sync.bfly.b32 %r844|%p86, %r840, %r813, %r812, %r814;
mov.b32 %f244, %r844;
add.ftz.f32 %f245, %f243, %f244;
mov.b32 %r845, %f245;
shfl.sync.bfly.b32 %r847|%p87, %r845, %r817, %r812, %r814;
mov.b32 %f246, %r847;
add.ftz.f32 %f247, %f245, %f246;
mov.b32 %r848, %f247;
shfl.sync.bfly.b32 %r850|%p88, %r848, %r820, %r812, %r814;
mov.b32 %f248, %r850;
add.ftz.f32 %f249, %f247, %f248;
mov.b32 %r851, %f249;
shfl.sync.bfly.b32 %r852|%p89, %r851, %r810, %r812, %r814;
mov.b32 %f250, %r852;
add.ftz.f32 %f251, %f249, %f250;
mov.b32 %r853, %f251;
shfl.sync.bfly.b32 %r855|%p90, %r853, %r825, %r812, %r814;
mov.b32 %f252, %r855;
add.ftz.f32 %f48, %f251, %f252;
@%p21 bra $L__BB23_30;

add.s32 %r856, %r24, 1;
cvt.u64.u32 %rd102, %r856;
add.s64 %rd104, %rd102, %rd75;
shl.b64 %rd106, %rd104, 2;
add.s64 %rd107, %rd77, %rd106;
st.global.f32 [%rd107], %f48;

$L__BB23_30:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type42ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type42ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type42ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type42ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type42ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type42ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type42ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type42ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type42ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<95>;
.reg .b16 %rs<23>;
.reg .f32 %f<323>;
.reg .b32 %r<858>;
.reg .b64 %rd<111>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type42ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1792];

ld.param.u64 %rd20, [_Z13mul_mat_vec_qIL9ggml_type42ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd21, [_Z13mul_mat_vec_qIL9ggml_type42ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd22, [_Z13mul_mat_vec_qIL9ggml_type42ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r48, [_Z13mul_mat_vec_qIL9ggml_type42ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type42ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type42ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type42ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type42ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type42ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type42ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r45, [_Z13mul_mat_vec_qIL9ggml_type42ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r46, [_Z13mul_mat_vec_qIL9ggml_type42ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r47, [_Z13mul_mat_vec_qIL9ggml_type42ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type42ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type42ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type42ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type42ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r49, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r49, %r2;
mov.u32 %r50, %ctaid.x;
shl.b32 %r4, %r50, 1;
shr.u32 %r51, %r3, 31;
add.s32 %r52, %r3, %r51;
shr.s32 %r857, %r52, 1;
shr.u32 %r6, %r48, 6;
setp.ge.s32 %p3, %r857, %r6;
mov.f32 %f309, 0f00000000;
mov.f32 %f310, %f309;
mov.f32 %f311, %f309;
mov.f32 %f312, %f309;
mov.f32 %f313, %f309;
mov.f32 %f314, %f309;
mov.f32 %f315, %f309;
mov.f32 %f316, %f309;
mov.f32 %f317, %f309;
mov.f32 %f318, %f309;
mov.f32 %f319, %f309;
mov.f32 %f320, %f309;
mov.f32 %f321, %f309;
mov.f32 %f322, %f309;
@%p3 bra $L__BB24_3;

mov.u32 %r53, %ctaid.y;
mov.u32 %r54, %ctaid.z;
mul.hi.u32 %r55, %r54, %r42;
add.s32 %r56, %r55, %r54;
shr.u32 %r57, %r56, %r43;
mul.lo.s32 %r58, %r57, %r45;
mul.hi.u32 %r59, %r53, %r36;
add.s32 %r60, %r59, %r53;
shr.u32 %r61, %r60, %r37;
mul.lo.s32 %r62, %r61, %r39;
mul.lo.s32 %r63, %r53, %r40;
cvt.u64.u32 %rd23, %r63;
add.s32 %r64, %r4, 1;
and.b32 %r67, %r52, -2;
sub.s32 %r68, %r3, %r67;
shl.b32 %r69, %r68, 2;
cvt.s64.s32 %rd24, %r68;
add.s64 %rd25, %rd23, %rd24;
mul.lo.s32 %r70, %r54, %r46;
cvt.u64.u32 %rd26, %r70;
add.s64 %rd1, %rd25, %rd26;
shl.b32 %r71, %r34, 1;
shl.b32 %r850, %r857, 1;
add.s32 %r856, %r850, %r71;
mad.lo.s32 %r855, %r34, 3, %r850;
shl.b32 %r72, %r34, 2;
add.s32 %r854, %r850, %r72;
mad.lo.s32 %r853, %r34, 5, %r850;
mad.lo.s32 %r852, %r34, 6, %r850;
add.s32 %r851, %r34, %r850;
mad.lo.s32 %r73, %r4, %r33, %r857;
add.s32 %r74, %r73, %r58;
add.s32 %r75, %r74, %r62;
mul.wide.s32 %rd27, %r75, 18;
mul.wide.s32 %rd28, %r69, 2;
add.s64 %rd29, %rd27, %rd28;
cvta.to.global.u64 %rd30, %rd20;
add.s64 %rd31, %rd30, %rd29;
add.s64 %rd110, %rd31, 8;
add.s64 %rd109, %rd30, %rd27;
add.s32 %r76, %r857, %r58;
add.s32 %r77, %r76, %r62;
mad.lo.s32 %r78, %r64, %r33, %r77;
mul.wide.s32 %rd32, %r78, 18;
add.s64 %rd33, %rd32, %rd28;
add.s64 %rd34, %rd30, %rd33;
add.s64 %rd108, %rd34, 4;
add.s64 %rd107, %rd30, %rd32;
cvta.to.global.u64 %rd6, %rd21;
mov.f32 %f309, 0f00000000;

$L__BB24_2:
cvt.u64.u32 %rd35, %r850;
add.s64 %rd36, %rd1, %rd35;
mul.lo.s64 %rd37, %rd36, 36;
ld.global.nc.u16 %rs1, [%rd109];

{ cvt.f32.f16 %f85, %rs1;}


ld.global.nc.u16 %rs15, [%rd110+-6];
cvt.s32.s16 %r541, %rs15;
add.s64 %rd38, %rd6, %rd37;
ld.global.nc.u32 %r80, [%rd38+4];
ld.global.nc.u32 %r84, [%rd38+8];
and.b32 %r542, %r541, 30583;
mov.u32 %r543, 33620223;
prmt.b32 %r544, %r543, %r543, %r542;
shr.u32 %r545, %r541, 2;
and.b32 %r546, %r545, 30583;
prmt.b32 %r547, %r543, %r543, %r546;
mov.u32 %r548, 20800;
prmt.b32 %r477, %r544, %r547, %r548;
mov.u32 %r549, 29538;
prmt.b32 %r481, %r544, %r547, %r549;
mov.u32 %r511, 0;

dp4a.s32.s32 %r79, %r80, %r477, %r511;


dp4a.s32.s32 %r83, %r84, %r481, %r79;

ld.global.nc.u16 %rs16, [%rd110+-4];
cvt.s32.s16 %r550, %rs16;
ld.global.nc.u32 %r88, [%rd38+12];
ld.global.nc.u32 %r92, [%rd38+16];
and.b32 %r551, %r550, 30583;
prmt.b32 %r552, %r543, %r543, %r551;
shr.u32 %r553, %r550, 2;
and.b32 %r554, %r553, 30583;
prmt.b32 %r555, %r543, %r543, %r554;
prmt.b32 %r485, %r552, %r555, %r548;
prmt.b32 %r489, %r552, %r555, %r549;

dp4a.s32.s32 %r87, %r88, %r485, %r83;


dp4a.s32.s32 %r91, %r92, %r489, %r87;

ld.global.nc.u16 %rs17, [%rd110+-2];
cvt.s32.s16 %r556, %rs17;
ld.global.nc.u32 %r96, [%rd38+20];
ld.global.nc.u32 %r100, [%rd38+24];
and.b32 %r557, %r556, 30583;
prmt.b32 %r558, %r543, %r543, %r557;
shr.u32 %r559, %r556, 2;
and.b32 %r560, %r559, 30583;
prmt.b32 %r561, %r543, %r543, %r560;
prmt.b32 %r493, %r558, %r561, %r548;
prmt.b32 %r497, %r558, %r561, %r549;

dp4a.s32.s32 %r95, %r96, %r493, %r91;


dp4a.s32.s32 %r99, %r100, %r497, %r95;

ld.global.nc.u16 %rs18, [%rd110];
cvt.s32.s16 %r562, %rs18;
ld.global.nc.u32 %r104, [%rd38+28];
ld.global.nc.u32 %r108, [%rd38+32];
and.b32 %r563, %r562, 30583;
prmt.b32 %r564, %r543, %r543, %r563;
shr.u32 %r565, %r562, 2;
and.b32 %r566, %r565, 30583;
prmt.b32 %r567, %r543, %r543, %r566;
prmt.b32 %r501, %r564, %r567, %r548;
prmt.b32 %r505, %r564, %r567, %r549;

dp4a.s32.s32 %r103, %r104, %r501, %r99;


dp4a.s32.s32 %r107, %r108, %r505, %r103;

ld.global.nc.u32 %r111, [%rd38];

{.reg .f16 low,high;
mov.b32 {low,high},%r111;
cvt.f32.f16 %f86, low;}


mul.ftz.f32 %f113, %f85, %f86;
cvt.rn.f32.s32 %f114, %r107;
fma.rn.ftz.f32 %f322, %f113, %f114, %f322;
ld.global.nc.u16 %rs2, [%rd107];

{ cvt.f32.f16 %f87, %rs2;}


ld.global.nc.u16 %rs19, [%rd108+-2];
cvt.s32.s16 %r568, %rs19;
and.b32 %r569, %r568, 30583;
prmt.b32 %r570, %r543, %r543, %r569;
shr.u32 %r571, %r568, 2;
and.b32 %r572, %r571, 30583;
prmt.b32 %r573, %r543, %r543, %r572;
prmt.b32 %r510, %r570, %r573, %r548;
prmt.b32 %r514, %r570, %r573, %r549;

dp4a.s32.s32 %r112, %r80, %r510, %r511;


dp4a.s32.s32 %r116, %r84, %r514, %r112;

ld.global.nc.u16 %rs20, [%rd108];
cvt.s32.s16 %r574, %rs20;
and.b32 %r575, %r574, 30583;
prmt.b32 %r576, %r543, %r543, %r575;
shr.u32 %r577, %r574, 2;
and.b32 %r578, %r577, 30583;
prmt.b32 %r579, %r543, %r543, %r578;
prmt.b32 %r518, %r576, %r579, %r548;
prmt.b32 %r522, %r576, %r579, %r549;

dp4a.s32.s32 %r120, %r88, %r518, %r116;


dp4a.s32.s32 %r124, %r92, %r522, %r120;

ld.global.nc.u16 %rs21, [%rd108+2];
cvt.s32.s16 %r580, %rs21;
and.b32 %r581, %r580, 30583;
prmt.b32 %r582, %r543, %r543, %r581;
shr.u32 %r583, %r580, 2;
and.b32 %r584, %r583, 30583;
prmt.b32 %r585, %r543, %r543, %r584;
prmt.b32 %r526, %r582, %r585, %r548;
prmt.b32 %r530, %r582, %r585, %r549;

dp4a.s32.s32 %r128, %r96, %r526, %r124;


dp4a.s32.s32 %r132, %r100, %r530, %r128;

ld.global.nc.u16 %rs22, [%rd108+4];
cvt.s32.s16 %r586, %rs22;
and.b32 %r587, %r586, 30583;
prmt.b32 %r588, %r543, %r543, %r587;
shr.u32 %r589, %r586, 2;
and.b32 %r590, %r589, 30583;
prmt.b32 %r591, %r543, %r543, %r590;
prmt.b32 %r534, %r588, %r591, %r548;
prmt.b32 %r538, %r588, %r591, %r549;

dp4a.s32.s32 %r136, %r104, %r534, %r132;


dp4a.s32.s32 %r140, %r108, %r538, %r136;


{.reg .f16 low,high;
mov.b32 {low,high},%r111;
cvt.f32.f16 %f88, low;}


mul.ftz.f32 %f115, %f87, %f88;
cvt.rn.f32.s32 %f116, %r140;
fma.rn.ftz.f32 %f321, %f115, %f116, %f321;
cvt.u64.u32 %rd39, %r851;
add.s64 %rd40, %rd1, %rd39;
mul.lo.s64 %rd41, %rd40, 36;

{ cvt.f32.f16 %f89, %rs1;}


add.s64 %rd42, %rd6, %rd41;
ld.global.nc.u32 %r146, [%rd42+4];
ld.global.nc.u32 %r150, [%rd42+8];

dp4a.s32.s32 %r145, %r146, %r477, %r511;


dp4a.s32.s32 %r149, %r150, %r481, %r145;

ld.global.nc.u32 %r154, [%rd42+12];
ld.global.nc.u32 %r158, [%rd42+16];

dp4a.s32.s32 %r153, %r154, %r485, %r149;


dp4a.s32.s32 %r157, %r158, %r489, %r153;

ld.global.nc.u32 %r162, [%rd42+20];
ld.global.nc.u32 %r166, [%rd42+24];

dp4a.s32.s32 %r161, %r162, %r493, %r157;


dp4a.s32.s32 %r165, %r166, %r497, %r161;

ld.global.nc.u32 %r170, [%rd42+28];
ld.global.nc.u32 %r174, [%rd42+32];

dp4a.s32.s32 %r169, %r170, %r501, %r165;


dp4a.s32.s32 %r173, %r174, %r505, %r169;

ld.global.nc.u32 %r177, [%rd42];

{.reg .f16 low,high;
mov.b32 {low,high},%r177;
cvt.f32.f16 %f90, low;}


mul.ftz.f32 %f117, %f89, %f90;
cvt.rn.f32.s32 %f118, %r173;
fma.rn.ftz.f32 %f320, %f117, %f118, %f320;

{ cvt.f32.f16 %f91, %rs2;}



dp4a.s32.s32 %r178, %r146, %r510, %r511;


dp4a.s32.s32 %r182, %r150, %r514, %r178;


dp4a.s32.s32 %r186, %r154, %r518, %r182;


dp4a.s32.s32 %r190, %r158, %r522, %r186;


dp4a.s32.s32 %r194, %r162, %r526, %r190;


dp4a.s32.s32 %r198, %r166, %r530, %r194;


dp4a.s32.s32 %r202, %r170, %r534, %r198;


dp4a.s32.s32 %r206, %r174, %r538, %r202;


{.reg .f16 low,high;
mov.b32 {low,high},%r177;
cvt.f32.f16 %f92, low;}


mul.ftz.f32 %f119, %f91, %f92;
cvt.rn.f32.s32 %f120, %r206;
fma.rn.ftz.f32 %f319, %f119, %f120, %f319;
cvt.u64.u32 %rd43, %r856;
add.s64 %rd44, %rd1, %rd43;
mul.lo.s64 %rd45, %rd44, 36;

{ cvt.f32.f16 %f93, %rs1;}


add.s64 %rd46, %rd6, %rd45;
ld.global.nc.u32 %r212, [%rd46+4];
ld.global.nc.u32 %r216, [%rd46+8];

dp4a.s32.s32 %r211, %r212, %r477, %r511;


dp4a.s32.s32 %r215, %r216, %r481, %r211;

ld.global.nc.u32 %r220, [%rd46+12];
ld.global.nc.u32 %r224, [%rd46+16];

dp4a.s32.s32 %r219, %r220, %r485, %r215;


dp4a.s32.s32 %r223, %r224, %r489, %r219;

ld.global.nc.u32 %r228, [%rd46+20];
ld.global.nc.u32 %r232, [%rd46+24];

dp4a.s32.s32 %r227, %r228, %r493, %r223;


dp4a.s32.s32 %r231, %r232, %r497, %r227;

ld.global.nc.u32 %r236, [%rd46+28];
ld.global.nc.u32 %r240, [%rd46+32];

dp4a.s32.s32 %r235, %r236, %r501, %r231;


dp4a.s32.s32 %r239, %r240, %r505, %r235;

ld.global.nc.u32 %r243, [%rd46];

{.reg .f16 low,high;
mov.b32 {low,high},%r243;
cvt.f32.f16 %f94, low;}


mul.ftz.f32 %f121, %f93, %f94;
cvt.rn.f32.s32 %f122, %r239;
fma.rn.ftz.f32 %f318, %f121, %f122, %f318;

{ cvt.f32.f16 %f95, %rs2;}



dp4a.s32.s32 %r244, %r212, %r510, %r511;


dp4a.s32.s32 %r248, %r216, %r514, %r244;


dp4a.s32.s32 %r252, %r220, %r518, %r248;


dp4a.s32.s32 %r256, %r224, %r522, %r252;


dp4a.s32.s32 %r260, %r228, %r526, %r256;


dp4a.s32.s32 %r264, %r232, %r530, %r260;


dp4a.s32.s32 %r268, %r236, %r534, %r264;


dp4a.s32.s32 %r272, %r240, %r538, %r268;


{.reg .f16 low,high;
mov.b32 {low,high},%r243;
cvt.f32.f16 %f96, low;}


mul.ftz.f32 %f123, %f95, %f96;
cvt.rn.f32.s32 %f124, %r272;
fma.rn.ftz.f32 %f317, %f123, %f124, %f317;
cvt.u64.u32 %rd47, %r855;
add.s64 %rd48, %rd1, %rd47;
mul.lo.s64 %rd49, %rd48, 36;

{ cvt.f32.f16 %f97, %rs1;}


add.s64 %rd50, %rd6, %rd49;
ld.global.nc.u32 %r278, [%rd50+4];
ld.global.nc.u32 %r282, [%rd50+8];

dp4a.s32.s32 %r277, %r278, %r477, %r511;


dp4a.s32.s32 %r281, %r282, %r481, %r277;

ld.global.nc.u32 %r286, [%rd50+12];
ld.global.nc.u32 %r290, [%rd50+16];

dp4a.s32.s32 %r285, %r286, %r485, %r281;


dp4a.s32.s32 %r289, %r290, %r489, %r285;

ld.global.nc.u32 %r294, [%rd50+20];
ld.global.nc.u32 %r298, [%rd50+24];

dp4a.s32.s32 %r293, %r294, %r493, %r289;


dp4a.s32.s32 %r297, %r298, %r497, %r293;

ld.global.nc.u32 %r302, [%rd50+28];
ld.global.nc.u32 %r306, [%rd50+32];

dp4a.s32.s32 %r301, %r302, %r501, %r297;


dp4a.s32.s32 %r305, %r306, %r505, %r301;

ld.global.nc.u32 %r309, [%rd50];

{.reg .f16 low,high;
mov.b32 {low,high},%r309;
cvt.f32.f16 %f98, low;}


mul.ftz.f32 %f125, %f97, %f98;
cvt.rn.f32.s32 %f126, %r305;
fma.rn.ftz.f32 %f316, %f125, %f126, %f316;

{ cvt.f32.f16 %f99, %rs2;}



dp4a.s32.s32 %r310, %r278, %r510, %r511;


dp4a.s32.s32 %r314, %r282, %r514, %r310;


dp4a.s32.s32 %r318, %r286, %r518, %r314;


dp4a.s32.s32 %r322, %r290, %r522, %r318;


dp4a.s32.s32 %r326, %r294, %r526, %r322;


dp4a.s32.s32 %r330, %r298, %r530, %r326;


dp4a.s32.s32 %r334, %r302, %r534, %r330;


dp4a.s32.s32 %r338, %r306, %r538, %r334;


{.reg .f16 low,high;
mov.b32 {low,high},%r309;
cvt.f32.f16 %f100, low;}


mul.ftz.f32 %f127, %f99, %f100;
cvt.rn.f32.s32 %f128, %r338;
fma.rn.ftz.f32 %f315, %f127, %f128, %f315;
cvt.u64.u32 %rd51, %r854;
add.s64 %rd52, %rd1, %rd51;
mul.lo.s64 %rd53, %rd52, 36;

{ cvt.f32.f16 %f101, %rs1;}


add.s64 %rd54, %rd6, %rd53;
ld.global.nc.u32 %r344, [%rd54+4];
ld.global.nc.u32 %r348, [%rd54+8];

dp4a.s32.s32 %r343, %r344, %r477, %r511;


dp4a.s32.s32 %r347, %r348, %r481, %r343;

ld.global.nc.u32 %r352, [%rd54+12];
ld.global.nc.u32 %r356, [%rd54+16];

dp4a.s32.s32 %r351, %r352, %r485, %r347;


dp4a.s32.s32 %r355, %r356, %r489, %r351;

ld.global.nc.u32 %r360, [%rd54+20];
ld.global.nc.u32 %r364, [%rd54+24];

dp4a.s32.s32 %r359, %r360, %r493, %r355;


dp4a.s32.s32 %r363, %r364, %r497, %r359;

ld.global.nc.u32 %r368, [%rd54+28];
ld.global.nc.u32 %r372, [%rd54+32];

dp4a.s32.s32 %r367, %r368, %r501, %r363;


dp4a.s32.s32 %r371, %r372, %r505, %r367;

ld.global.nc.u32 %r375, [%rd54];

{.reg .f16 low,high;
mov.b32 {low,high},%r375;
cvt.f32.f16 %f102, low;}


mul.ftz.f32 %f129, %f101, %f102;
cvt.rn.f32.s32 %f130, %r371;
fma.rn.ftz.f32 %f314, %f129, %f130, %f314;

{ cvt.f32.f16 %f103, %rs2;}



dp4a.s32.s32 %r376, %r344, %r510, %r511;


dp4a.s32.s32 %r380, %r348, %r514, %r376;


dp4a.s32.s32 %r384, %r352, %r518, %r380;


dp4a.s32.s32 %r388, %r356, %r522, %r384;


dp4a.s32.s32 %r392, %r360, %r526, %r388;


dp4a.s32.s32 %r396, %r364, %r530, %r392;


dp4a.s32.s32 %r400, %r368, %r534, %r396;


dp4a.s32.s32 %r404, %r372, %r538, %r400;


{.reg .f16 low,high;
mov.b32 {low,high},%r375;
cvt.f32.f16 %f104, low;}


mul.ftz.f32 %f131, %f103, %f104;
cvt.rn.f32.s32 %f132, %r404;
fma.rn.ftz.f32 %f313, %f131, %f132, %f313;
cvt.u64.u32 %rd55, %r853;
add.s64 %rd56, %rd1, %rd55;
mul.lo.s64 %rd57, %rd56, 36;

{ cvt.f32.f16 %f105, %rs1;}


add.s64 %rd58, %rd6, %rd57;
ld.global.nc.u32 %r410, [%rd58+4];
ld.global.nc.u32 %r414, [%rd58+8];

dp4a.s32.s32 %r409, %r410, %r477, %r511;


dp4a.s32.s32 %r413, %r414, %r481, %r409;

ld.global.nc.u32 %r418, [%rd58+12];
ld.global.nc.u32 %r422, [%rd58+16];

dp4a.s32.s32 %r417, %r418, %r485, %r413;


dp4a.s32.s32 %r421, %r422, %r489, %r417;

ld.global.nc.u32 %r426, [%rd58+20];
ld.global.nc.u32 %r430, [%rd58+24];

dp4a.s32.s32 %r425, %r426, %r493, %r421;


dp4a.s32.s32 %r429, %r430, %r497, %r425;

ld.global.nc.u32 %r434, [%rd58+28];
ld.global.nc.u32 %r438, [%rd58+32];

dp4a.s32.s32 %r433, %r434, %r501, %r429;


dp4a.s32.s32 %r437, %r438, %r505, %r433;

ld.global.nc.u32 %r441, [%rd58];

{.reg .f16 low,high;
mov.b32 {low,high},%r441;
cvt.f32.f16 %f106, low;}


mul.ftz.f32 %f133, %f105, %f106;
cvt.rn.f32.s32 %f134, %r437;
fma.rn.ftz.f32 %f312, %f133, %f134, %f312;

{ cvt.f32.f16 %f107, %rs2;}



dp4a.s32.s32 %r442, %r410, %r510, %r511;


dp4a.s32.s32 %r446, %r414, %r514, %r442;


dp4a.s32.s32 %r450, %r418, %r518, %r446;


dp4a.s32.s32 %r454, %r422, %r522, %r450;


dp4a.s32.s32 %r458, %r426, %r526, %r454;


dp4a.s32.s32 %r462, %r430, %r530, %r458;


dp4a.s32.s32 %r466, %r434, %r534, %r462;


dp4a.s32.s32 %r470, %r438, %r538, %r466;


{.reg .f16 low,high;
mov.b32 {low,high},%r441;
cvt.f32.f16 %f108, low;}


mul.ftz.f32 %f135, %f107, %f108;
cvt.rn.f32.s32 %f136, %r470;
fma.rn.ftz.f32 %f311, %f135, %f136, %f311;
cvt.u64.u32 %rd59, %r852;
add.s64 %rd60, %rd1, %rd59;
mul.lo.s64 %rd61, %rd60, 36;

{ cvt.f32.f16 %f109, %rs1;}


add.s64 %rd62, %rd6, %rd61;
ld.global.nc.u32 %r476, [%rd62+4];
ld.global.nc.u32 %r480, [%rd62+8];

dp4a.s32.s32 %r475, %r476, %r477, %r511;


dp4a.s32.s32 %r479, %r480, %r481, %r475;

ld.global.nc.u32 %r484, [%rd62+12];
ld.global.nc.u32 %r488, [%rd62+16];

dp4a.s32.s32 %r483, %r484, %r485, %r479;


dp4a.s32.s32 %r487, %r488, %r489, %r483;

ld.global.nc.u32 %r492, [%rd62+20];
ld.global.nc.u32 %r496, [%rd62+24];

dp4a.s32.s32 %r491, %r492, %r493, %r487;


dp4a.s32.s32 %r495, %r496, %r497, %r491;

ld.global.nc.u32 %r500, [%rd62+28];
ld.global.nc.u32 %r504, [%rd62+32];

dp4a.s32.s32 %r499, %r500, %r501, %r495;


dp4a.s32.s32 %r503, %r504, %r505, %r499;

ld.global.nc.u32 %r507, [%rd62];

{.reg .f16 low,high;
mov.b32 {low,high},%r507;
cvt.f32.f16 %f110, low;}


mul.ftz.f32 %f137, %f109, %f110;
cvt.rn.f32.s32 %f138, %r503;
fma.rn.ftz.f32 %f310, %f137, %f138, %f310;

{ cvt.f32.f16 %f111, %rs2;}



dp4a.s32.s32 %r508, %r476, %r510, %r511;


dp4a.s32.s32 %r512, %r480, %r514, %r508;


dp4a.s32.s32 %r516, %r484, %r518, %r512;


dp4a.s32.s32 %r520, %r488, %r522, %r516;


dp4a.s32.s32 %r524, %r492, %r526, %r520;


dp4a.s32.s32 %r528, %r496, %r530, %r524;


dp4a.s32.s32 %r532, %r500, %r534, %r528;


dp4a.s32.s32 %r536, %r504, %r538, %r532;


{.reg .f16 low,high;
mov.b32 {low,high},%r507;
cvt.f32.f16 %f112, low;}


mul.ftz.f32 %f139, %f111, %f112;
cvt.rn.f32.s32 %f140, %r536;
fma.rn.ftz.f32 %f309, %f139, %f140, %f309;
add.s32 %r856, %r856, 64;
add.s32 %r855, %r855, 64;
add.s32 %r854, %r854, 64;
add.s32 %r853, %r853, 64;
add.s32 %r852, %r852, 64;
add.s32 %r851, %r851, 64;
add.s32 %r850, %r850, 64;
add.s64 %rd110, %rd110, 576;
add.s64 %rd109, %rd109, 576;
add.s64 %rd108, %rd108, 576;
add.s64 %rd107, %rd107, 576;
add.s32 %r857, %r857, 32;
setp.lt.s32 %p4, %r857, %r6;
@%p4 bra $L__BB24_2;

$L__BB24_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB24_5;

mov.u32 %r592, _ZZ13mul_mat_vec_qIL9ggml_type42ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r593, %r1, 1792, %r592;
shl.b32 %r594, %r2, 2;
add.s32 %r595, %r593, %r594;
st.shared.f32 [%r595+-1792], %f322;
st.shared.f32 [%r595+-1664], %f321;
st.shared.f32 [%r595+-1536], %f320;
st.shared.f32 [%r595+-1408], %f319;
st.shared.f32 [%r595+-1280], %f318;
st.shared.f32 [%r595+-1152], %f317;
st.shared.f32 [%r595+-1024], %f316;
st.shared.f32 [%r595+-896], %f315;
st.shared.f32 [%r595+-768], %f314;
st.shared.f32 [%r595+-640], %f313;
st.shared.f32 [%r595+-512], %f312;
st.shared.f32 [%r595+-384], %f311;
st.shared.f32 [%r595+-256], %f310;
st.shared.f32 [%r595+-128], %f309;

$L__BB24_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB24_34;

mov.u32 %r596, %ctaid.z;
mov.u32 %r597, %ctaid.y;
mad.lo.s32 %r598, %r597, %r41, %r4;
mad.lo.s32 %r599, %r596, %r47, %r598;
cvt.u64.u32 %rd15, %r599;
shl.b32 %r600, %r2, 2;
mov.u32 %r601, _ZZ13mul_mat_vec_qIL9ggml_type42ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r30, %r601, %r600;
mov.u32 %r602, 2;
ld.shared.f32 %f141, [%r30];
add.ftz.f32 %f142, %f141, %f322;
mov.b32 %r603, %f142;
mov.u32 %r604, 31;
mov.u32 %r605, 16;
mov.u32 %r606, -1;
shfl.sync.bfly.b32 %r607|%p7, %r603, %r605, %r604, %r606;
mov.b32 %f143, %r607;
add.ftz.f32 %f144, %f142, %f143;
mov.b32 %r608, %f144;
mov.u32 %r609, 8;
shfl.sync.bfly.b32 %r610|%p8, %r608, %r609, %r604, %r606;
mov.b32 %f145, %r610;
add.ftz.f32 %f146, %f144, %f145;
mov.b32 %r611, %f146;
mov.u32 %r612, 4;
shfl.sync.bfly.b32 %r613|%p9, %r611, %r612, %r604, %r606;
mov.b32 %f147, %r613;
add.ftz.f32 %f148, %f146, %f147;
mov.b32 %r614, %f148;
shfl.sync.bfly.b32 %r615|%p10, %r614, %r602, %r604, %r606;
mov.b32 %f149, %r615;
add.ftz.f32 %f150, %f148, %f149;
mov.b32 %r616, %f150;
mov.u32 %r617, 1;
shfl.sync.bfly.b32 %r618|%p11, %r616, %r617, %r604, %r606;
mov.b32 %f151, %r618;
add.ftz.f32 %f43, %f150, %f151;
setp.lt.u32 %p12, %r4, %r35;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
cvta.to.global.u64 %rd63, %rd22;
mul.wide.u32 %rd64, %r599, 4;
add.s64 %rd16, %rd63, %rd64;
not.pred %p14, %p1;
@%p14 bra $L__BB24_8;

st.global.f32 [%rd16], %f43;

$L__BB24_8:
ld.shared.f32 %f152, [%r30+128];
add.ftz.f32 %f153, %f152, %f321;
mov.b32 %r619, %f153;
shfl.sync.bfly.b32 %r623|%p15, %r619, %r605, %r604, %r606;
mov.b32 %f154, %r623;
add.ftz.f32 %f155, %f153, %f154;
mov.b32 %r624, %f155;
shfl.sync.bfly.b32 %r626|%p16, %r624, %r609, %r604, %r606;
mov.b32 %f156, %r626;
add.ftz.f32 %f157, %f155, %f156;
mov.b32 %r627, %f157;
shfl.sync.bfly.b32 %r629|%p17, %r627, %r612, %r604, %r606;
mov.b32 %f158, %r629;
add.ftz.f32 %f159, %f157, %f158;
mov.b32 %r630, %f159;
shfl.sync.bfly.b32 %r632|%p18, %r630, %r602, %r604, %r606;
mov.b32 %f160, %r632;
add.ftz.f32 %f161, %f159, %f160;
mov.b32 %r633, %f161;
shfl.sync.bfly.b32 %r635|%p19, %r633, %r617, %r604, %r606;
mov.b32 %f162, %r635;
add.ftz.f32 %f44, %f161, %f162;
setp.eq.s32 %p20, %r2, 1;
or.b32 %r639, %r4, 1;
setp.lt.u32 %p21, %r639, %r35;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB24_10;

st.global.f32 [%rd16+4], %f44;

$L__BB24_10:
ld.shared.f32 %f163, [%r30+256];
add.ftz.f32 %f164, %f163, %f320;
mov.b32 %r640, %f164;
mov.u32 %r641, 31;
mov.u32 %r642, 16;
mov.u32 %r643, -1;
shfl.sync.bfly.b32 %r644|%p23, %r640, %r642, %r641, %r643;
mov.b32 %f165, %r644;
add.ftz.f32 %f166, %f164, %f165;
mov.b32 %r645, %f166;
mov.u32 %r646, 8;
shfl.sync.bfly.b32 %r647|%p24, %r645, %r646, %r641, %r643;
mov.b32 %f167, %r647;
add.ftz.f32 %f168, %f166, %f167;
mov.b32 %r648, %f168;
mov.u32 %r649, 4;
shfl.sync.bfly.b32 %r650|%p25, %r648, %r649, %r641, %r643;
mov.b32 %f169, %r650;
add.ftz.f32 %f170, %f168, %f169;
mov.b32 %r651, %f170;
mov.u32 %r652, 2;
shfl.sync.bfly.b32 %r653|%p26, %r651, %r652, %r641, %r643;
mov.b32 %f171, %r653;
add.ftz.f32 %f172, %f170, %f171;
mov.b32 %r654, %f172;
mov.u32 %r655, 1;
shfl.sync.bfly.b32 %r656|%p27, %r654, %r655, %r641, %r643;
mov.b32 %f173, %r656;
add.ftz.f32 %f45, %f172, %f173;
@%p14 bra $L__BB24_12;

cvt.u64.u32 %rd65, %r35;
add.s64 %rd66, %rd65, %rd15;
shl.b64 %rd68, %rd66, 2;
add.s64 %rd69, %rd63, %rd68;
st.global.f32 [%rd69], %f45;

$L__BB24_12:
ld.shared.f32 %f174, [%r30+384];
add.ftz.f32 %f175, %f174, %f319;
mov.b32 %r657, %f175;
shfl.sync.bfly.b32 %r661|%p29, %r657, %r642, %r641, %r643;
mov.b32 %f176, %r661;
add.ftz.f32 %f177, %f175, %f176;
mov.b32 %r662, %f177;
shfl.sync.bfly.b32 %r664|%p30, %r662, %r646, %r641, %r643;
mov.b32 %f178, %r664;
add.ftz.f32 %f179, %f177, %f178;
mov.b32 %r665, %f179;
shfl.sync.bfly.b32 %r667|%p31, %r665, %r649, %r641, %r643;
mov.b32 %f180, %r667;
add.ftz.f32 %f181, %f179, %f180;
mov.b32 %r668, %f181;
shfl.sync.bfly.b32 %r670|%p32, %r668, %r652, %r641, %r643;
mov.b32 %f182, %r670;
add.ftz.f32 %f183, %f181, %f182;
mov.b32 %r671, %f183;
shfl.sync.bfly.b32 %r673|%p33, %r671, %r655, %r641, %r643;
mov.b32 %f184, %r673;
add.ftz.f32 %f46, %f183, %f184;
@%p22 bra $L__BB24_14;

add.s32 %r674, %r35, 1;
cvt.u64.u32 %rd70, %r674;
add.s64 %rd71, %rd70, %rd15;
shl.b64 %rd73, %rd71, 2;
add.s64 %rd74, %rd63, %rd73;
st.global.f32 [%rd74], %f46;

$L__BB24_14:
shl.b32 %r675, %r35, 1;
mov.u32 %r676, 1;
ld.shared.f32 %f185, [%r30+512];
add.ftz.f32 %f186, %f185, %f318;
mov.b32 %r677, %f186;
mov.u32 %r678, 31;
mov.u32 %r679, 16;
mov.u32 %r680, -1;
shfl.sync.bfly.b32 %r681|%p35, %r677, %r679, %r678, %r680;
mov.b32 %f187, %r681;
add.ftz.f32 %f188, %f186, %f187;
mov.b32 %r682, %f188;
mov.u32 %r683, 8;
shfl.sync.bfly.b32 %r684|%p36, %r682, %r683, %r678, %r680;
mov.b32 %f189, %r684;
add.ftz.f32 %f190, %f188, %f189;
mov.b32 %r685, %f190;
mov.u32 %r686, 4;
shfl.sync.bfly.b32 %r687|%p37, %r685, %r686, %r678, %r680;
mov.b32 %f191, %r687;
add.ftz.f32 %f192, %f190, %f191;
mov.b32 %r688, %f192;
mov.u32 %r689, 2;
shfl.sync.bfly.b32 %r690|%p38, %r688, %r689, %r678, %r680;
mov.b32 %f193, %r690;
add.ftz.f32 %f194, %f192, %f193;
mov.b32 %r691, %f194;
shfl.sync.bfly.b32 %r692|%p39, %r691, %r676, %r678, %r680;
mov.b32 %f195, %r692;
add.ftz.f32 %f47, %f194, %f195;
cvt.u64.u32 %rd75, %r675;
add.s64 %rd76, %rd75, %rd15;
shl.b64 %rd78, %rd76, 2;
add.s64 %rd17, %rd63, %rd78;
@%p14 bra $L__BB24_16;

st.global.f32 [%rd17], %f47;

$L__BB24_16:
ld.shared.f32 %f196, [%r30+640];
add.ftz.f32 %f197, %f196, %f317;
mov.b32 %r693, %f197;
shfl.sync.bfly.b32 %r697|%p41, %r693, %r679, %r678, %r680;
mov.b32 %f198, %r697;
add.ftz.f32 %f199, %f197, %f198;
mov.b32 %r698, %f199;
shfl.sync.bfly.b32 %r700|%p42, %r698, %r683, %r678, %r680;
mov.b32 %f200, %r700;
add.ftz.f32 %f201, %f199, %f200;
mov.b32 %r701, %f201;
shfl.sync.bfly.b32 %r703|%p43, %r701, %r686, %r678, %r680;
mov.b32 %f202, %r703;
add.ftz.f32 %f203, %f201, %f202;
mov.b32 %r704, %f203;
shfl.sync.bfly.b32 %r706|%p44, %r704, %r689, %r678, %r680;
mov.b32 %f204, %r706;
add.ftz.f32 %f205, %f203, %f204;
mov.b32 %r707, %f205;
shfl.sync.bfly.b32 %r709|%p45, %r707, %r676, %r678, %r680;
mov.b32 %f206, %r709;
add.ftz.f32 %f48, %f205, %f206;
@%p22 bra $L__BB24_18;

st.global.f32 [%rd17+4], %f48;

$L__BB24_18:
mul.lo.s32 %r31, %r35, 3;
ld.shared.f32 %f207, [%r30+768];
add.ftz.f32 %f208, %f207, %f316;
mov.b32 %r710, %f208;
mov.u32 %r711, 31;
mov.u32 %r712, 16;
mov.u32 %r713, -1;
shfl.sync.bfly.b32 %r714|%p47, %r710, %r712, %r711, %r713;
mov.b32 %f209, %r714;
add.ftz.f32 %f210, %f208, %f209;
mov.b32 %r715, %f210;
mov.u32 %r716, 8;
shfl.sync.bfly.b32 %r717|%p48, %r715, %r716, %r711, %r713;
mov.b32 %f211, %r717;
add.ftz.f32 %f212, %f210, %f211;
mov.b32 %r718, %f212;
mov.u32 %r719, 4;
shfl.sync.bfly.b32 %r720|%p49, %r718, %r719, %r711, %r713;
mov.b32 %f213, %r720;
add.ftz.f32 %f214, %f212, %f213;
mov.b32 %r721, %f214;
mov.u32 %r722, 2;
shfl.sync.bfly.b32 %r723|%p50, %r721, %r722, %r711, %r713;
mov.b32 %f215, %r723;
add.ftz.f32 %f216, %f214, %f215;
mov.b32 %r724, %f216;
mov.u32 %r725, 1;
shfl.sync.bfly.b32 %r726|%p51, %r724, %r725, %r711, %r713;
mov.b32 %f217, %r726;
add.ftz.f32 %f49, %f216, %f217;
@%p14 bra $L__BB24_20;

cvt.u64.u32 %rd79, %r31;
add.s64 %rd80, %rd79, %rd15;
shl.b64 %rd82, %rd80, 2;
add.s64 %rd83, %rd63, %rd82;
st.global.f32 [%rd83], %f49;

$L__BB24_20:
ld.shared.f32 %f218, [%r30+896];
add.ftz.f32 %f219, %f218, %f315;
mov.b32 %r727, %f219;
shfl.sync.bfly.b32 %r731|%p53, %r727, %r712, %r711, %r713;
mov.b32 %f220, %r731;
add.ftz.f32 %f221, %f219, %f220;
mov.b32 %r732, %f221;
shfl.sync.bfly.b32 %r734|%p54, %r732, %r716, %r711, %r713;
mov.b32 %f222, %r734;
add.ftz.f32 %f223, %f221, %f222;
mov.b32 %r735, %f223;
shfl.sync.bfly.b32 %r737|%p55, %r735, %r719, %r711, %r713;
mov.b32 %f224, %r737;
add.ftz.f32 %f225, %f223, %f224;
mov.b32 %r738, %f225;
shfl.sync.bfly.b32 %r740|%p56, %r738, %r722, %r711, %r713;
mov.b32 %f226, %r740;
add.ftz.f32 %f227, %f225, %f226;
mov.b32 %r741, %f227;
shfl.sync.bfly.b32 %r743|%p57, %r741, %r725, %r711, %r713;
mov.b32 %f228, %r743;
add.ftz.f32 %f50, %f227, %f228;
@%p22 bra $L__BB24_22;

add.s32 %r744, %r31, 1;
cvt.u64.u32 %rd84, %r744;
add.s64 %rd85, %rd84, %rd15;
shl.b64 %rd87, %rd85, 2;
add.s64 %rd88, %rd63, %rd87;
st.global.f32 [%rd88], %f50;

$L__BB24_22:
shl.b32 %r745, %r35, 2;
mov.u32 %r746, 2;
ld.shared.f32 %f229, [%r30+1024];
add.ftz.f32 %f230, %f229, %f314;
mov.b32 %r747, %f230;
mov.u32 %r748, 31;
mov.u32 %r749, 16;
mov.u32 %r750, -1;
shfl.sync.bfly.b32 %r751|%p59, %r747, %r749, %r748, %r750;
mov.b32 %f231, %r751;
add.ftz.f32 %f232, %f230, %f231;
mov.b32 %r752, %f232;
mov.u32 %r753, 8;
shfl.sync.bfly.b32 %r754|%p60, %r752, %r753, %r748, %r750;
mov.b32 %f233, %r754;
add.ftz.f32 %f234, %f232, %f233;
mov.b32 %r755, %f234;
mov.u32 %r756, 4;
shfl.sync.bfly.b32 %r757|%p61, %r755, %r756, %r748, %r750;
mov.b32 %f235, %r757;
add.ftz.f32 %f236, %f234, %f235;
mov.b32 %r758, %f236;
shfl.sync.bfly.b32 %r759|%p62, %r758, %r746, %r748, %r750;
mov.b32 %f237, %r759;
add.ftz.f32 %f238, %f236, %f237;
mov.b32 %r760, %f238;
mov.u32 %r761, 1;
shfl.sync.bfly.b32 %r762|%p63, %r760, %r761, %r748, %r750;
mov.b32 %f239, %r762;
add.ftz.f32 %f51, %f238, %f239;
cvt.u64.u32 %rd89, %r745;
add.s64 %rd90, %rd89, %rd15;
shl.b64 %rd92, %rd90, 2;
add.s64 %rd18, %rd63, %rd92;
@%p14 bra $L__BB24_24;

st.global.f32 [%rd18], %f51;

$L__BB24_24:
ld.shared.f32 %f240, [%r30+1152];
add.ftz.f32 %f241, %f240, %f313;
mov.b32 %r763, %f241;
shfl.sync.bfly.b32 %r767|%p65, %r763, %r749, %r748, %r750;
mov.b32 %f242, %r767;
add.ftz.f32 %f243, %f241, %f242;
mov.b32 %r768, %f243;
shfl.sync.bfly.b32 %r770|%p66, %r768, %r753, %r748, %r750;
mov.b32 %f244, %r770;
add.ftz.f32 %f245, %f243, %f244;
mov.b32 %r771, %f245;
shfl.sync.bfly.b32 %r773|%p67, %r771, %r756, %r748, %r750;
mov.b32 %f246, %r773;
add.ftz.f32 %f247, %f245, %f246;
mov.b32 %r774, %f247;
shfl.sync.bfly.b32 %r776|%p68, %r774, %r746, %r748, %r750;
mov.b32 %f248, %r776;
add.ftz.f32 %f249, %f247, %f248;
mov.b32 %r777, %f249;
shfl.sync.bfly.b32 %r779|%p69, %r777, %r761, %r748, %r750;
mov.b32 %f250, %r779;
add.ftz.f32 %f52, %f249, %f250;
@%p22 bra $L__BB24_26;

st.global.f32 [%rd18+4], %f52;

$L__BB24_26:
mul.lo.s32 %r32, %r35, 5;
ld.shared.f32 %f251, [%r30+1280];
add.ftz.f32 %f252, %f251, %f312;
mov.b32 %r780, %f252;
mov.u32 %r781, 31;
mov.u32 %r782, 16;
mov.u32 %r783, -1;
shfl.sync.bfly.b32 %r784|%p71, %r780, %r782, %r781, %r783;
mov.b32 %f253, %r784;
add.ftz.f32 %f254, %f252, %f253;
mov.b32 %r785, %f254;
mov.u32 %r786, 8;
shfl.sync.bfly.b32 %r787|%p72, %r785, %r786, %r781, %r783;
mov.b32 %f255, %r787;
add.ftz.f32 %f256, %f254, %f255;
mov.b32 %r788, %f256;
mov.u32 %r789, 4;
shfl.sync.bfly.b32 %r790|%p73, %r788, %r789, %r781, %r783;
mov.b32 %f257, %r790;
add.ftz.f32 %f258, %f256, %f257;
mov.b32 %r791, %f258;
mov.u32 %r792, 2;
shfl.sync.bfly.b32 %r793|%p74, %r791, %r792, %r781, %r783;
mov.b32 %f259, %r793;
add.ftz.f32 %f260, %f258, %f259;
mov.b32 %r794, %f260;
mov.u32 %r795, 1;
shfl.sync.bfly.b32 %r796|%p75, %r794, %r795, %r781, %r783;
mov.b32 %f261, %r796;
add.ftz.f32 %f53, %f260, %f261;
@%p14 bra $L__BB24_28;

cvt.u64.u32 %rd93, %r32;
add.s64 %rd94, %rd93, %rd15;
shl.b64 %rd96, %rd94, 2;
add.s64 %rd97, %rd63, %rd96;
st.global.f32 [%rd97], %f53;

$L__BB24_28:
ld.shared.f32 %f262, [%r30+1408];
add.ftz.f32 %f263, %f262, %f311;
mov.b32 %r797, %f263;
shfl.sync.bfly.b32 %r801|%p77, %r797, %r782, %r781, %r783;
mov.b32 %f264, %r801;
add.ftz.f32 %f265, %f263, %f264;
mov.b32 %r802, %f265;
shfl.sync.bfly.b32 %r804|%p78, %r802, %r786, %r781, %r783;
mov.b32 %f266, %r804;
add.ftz.f32 %f267, %f265, %f266;
mov.b32 %r805, %f267;
shfl.sync.bfly.b32 %r807|%p79, %r805, %r789, %r781, %r783;
mov.b32 %f268, %r807;
add.ftz.f32 %f269, %f267, %f268;
mov.b32 %r808, %f269;
shfl.sync.bfly.b32 %r810|%p80, %r808, %r792, %r781, %r783;
mov.b32 %f270, %r810;
add.ftz.f32 %f271, %f269, %f270;
mov.b32 %r811, %f271;
shfl.sync.bfly.b32 %r813|%p81, %r811, %r795, %r781, %r783;
mov.b32 %f272, %r813;
add.ftz.f32 %f54, %f271, %f272;
@%p22 bra $L__BB24_30;

add.s32 %r814, %r32, 1;
cvt.u64.u32 %rd98, %r814;
add.s64 %rd99, %rd98, %rd15;
shl.b64 %rd101, %rd99, 2;
add.s64 %rd102, %rd63, %rd101;
st.global.f32 [%rd102], %f54;

$L__BB24_30:
mul.lo.s32 %r815, %r35, 6;
ld.shared.f32 %f273, [%r30+1536];
add.ftz.f32 %f274, %f273, %f310;
mov.b32 %r816, %f274;
mov.u32 %r817, 31;
mov.u32 %r818, 16;
mov.u32 %r819, -1;
shfl.sync.bfly.b32 %r820|%p83, %r816, %r818, %r817, %r819;
mov.b32 %f275, %r820;
add.ftz.f32 %f276, %f274, %f275;
mov.b32 %r821, %f276;
mov.u32 %r822, 8;
shfl.sync.bfly.b32 %r823|%p84, %r821, %r822, %r817, %r819;
mov.b32 %f277, %r823;
add.ftz.f32 %f278, %f276, %f277;
mov.b32 %r824, %f278;
mov.u32 %r825, 4;
shfl.sync.bfly.b32 %r826|%p85, %r824, %r825, %r817, %r819;
mov.b32 %f279, %r826;
add.ftz.f32 %f280, %f278, %f279;
mov.b32 %r827, %f280;
mov.u32 %r828, 2;
shfl.sync.bfly.b32 %r829|%p86, %r827, %r828, %r817, %r819;
mov.b32 %f281, %r829;
add.ftz.f32 %f282, %f280, %f281;
mov.b32 %r830, %f282;
mov.u32 %r831, 1;
shfl.sync.bfly.b32 %r832|%p87, %r830, %r831, %r817, %r819;
mov.b32 %f283, %r832;
add.ftz.f32 %f55, %f282, %f283;
cvt.u64.u32 %rd103, %r815;
add.s64 %rd104, %rd103, %rd15;
shl.b64 %rd106, %rd104, 2;
add.s64 %rd19, %rd63, %rd106;
@%p14 bra $L__BB24_32;

st.global.f32 [%rd19], %f55;

$L__BB24_32:
ld.shared.f32 %f284, [%r30+1664];
add.ftz.f32 %f285, %f284, %f309;
mov.b32 %r833, %f285;
shfl.sync.bfly.b32 %r837|%p89, %r833, %r818, %r817, %r819;
mov.b32 %f286, %r837;
add.ftz.f32 %f287, %f285, %f286;
mov.b32 %r838, %f287;
shfl.sync.bfly.b32 %r840|%p90, %r838, %r822, %r817, %r819;
mov.b32 %f288, %r840;
add.ftz.f32 %f289, %f287, %f288;
mov.b32 %r841, %f289;
shfl.sync.bfly.b32 %r843|%p91, %r841, %r825, %r817, %r819;
mov.b32 %f290, %r843;
add.ftz.f32 %f291, %f289, %f290;
mov.b32 %r844, %f291;
shfl.sync.bfly.b32 %r846|%p92, %r844, %r828, %r817, %r819;
mov.b32 %f292, %r846;
add.ftz.f32 %f293, %f291, %f292;
mov.b32 %r847, %f293;
shfl.sync.bfly.b32 %r849|%p93, %r847, %r831, %r817, %r819;
mov.b32 %f294, %r849;
add.ftz.f32 %f56, %f293, %f294;
@%p22 bra $L__BB24_34;

st.global.f32 [%rd19+4], %f56;

$L__BB24_34:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type42ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type42ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type42ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type42ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type42ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type42ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type42ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type42ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type42ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type42ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<107>;
.reg .b16 %rs<25>;
.reg .f32 %f<369>;
.reg .b32 %r<965>;
.reg .b64 %rd<125>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type42ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[2048];

ld.param.u64 %rd20, [_Z13mul_mat_vec_qIL9ggml_type42ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd21, [_Z13mul_mat_vec_qIL9ggml_type42ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd22, [_Z13mul_mat_vec_qIL9ggml_type42ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r52, [_Z13mul_mat_vec_qIL9ggml_type42ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type42ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type42ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type42ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type42ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type42ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r45, [_Z13mul_mat_vec_qIL9ggml_type42ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r49, [_Z13mul_mat_vec_qIL9ggml_type42ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r50, [_Z13mul_mat_vec_qIL9ggml_type42ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r51, [_Z13mul_mat_vec_qIL9ggml_type42ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r47, [_Z13mul_mat_vec_qIL9ggml_type42ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r46, [_Z13mul_mat_vec_qIL9ggml_type42ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type42ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type42ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r53, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r53, %r2;
mov.u32 %r54, %ctaid.x;
shl.b32 %r4, %r54, 1;
shr.u32 %r55, %r3, 31;
add.s32 %r56, %r3, %r55;
shr.s32 %r964, %r56, 1;
shr.u32 %r6, %r52, 6;
setp.ge.s32 %p3, %r964, %r6;
mov.f32 %f353, 0f00000000;
mov.f32 %f354, %f353;
mov.f32 %f355, %f353;
mov.f32 %f356, %f353;
mov.f32 %f357, %f353;
mov.f32 %f358, %f353;
mov.f32 %f359, %f353;
mov.f32 %f360, %f353;
mov.f32 %f361, %f353;
mov.f32 %f362, %f353;
mov.f32 %f363, %f353;
mov.f32 %f364, %f353;
mov.f32 %f365, %f353;
mov.f32 %f366, %f353;
mov.f32 %f367, %f353;
mov.f32 %f368, %f353;
@%p3 bra $L__BB25_3;

mov.u32 %r57, %ctaid.y;
mov.u32 %r58, %ctaid.z;
mul.hi.u32 %r59, %r58, %r46;
add.s32 %r60, %r59, %r58;
shr.u32 %r61, %r60, %r47;
mul.lo.s32 %r62, %r61, %r49;
mul.hi.u32 %r63, %r57, %r40;
add.s32 %r64, %r63, %r57;
shr.u32 %r65, %r64, %r41;
mul.lo.s32 %r66, %r65, %r43;
mul.lo.s32 %r67, %r57, %r44;
cvt.u64.u32 %rd23, %r67;
add.s32 %r68, %r4, 1;
and.b32 %r71, %r56, -2;
sub.s32 %r72, %r3, %r71;
shl.b32 %r73, %r72, 2;
cvt.s64.s32 %rd24, %r72;
add.s64 %rd25, %rd23, %rd24;
mul.lo.s32 %r74, %r58, %r50;
cvt.u64.u32 %rd26, %r74;
add.s64 %rd1, %rd25, %rd26;
shl.b32 %r75, %r38, 1;
shl.b32 %r956, %r964, 1;
add.s32 %r963, %r956, %r75;
mad.lo.s32 %r962, %r38, 3, %r956;
shl.b32 %r76, %r38, 2;
add.s32 %r961, %r956, %r76;
mad.lo.s32 %r960, %r38, 5, %r956;
mad.lo.s32 %r959, %r38, 6, %r956;
mad.lo.s32 %r958, %r38, 7, %r956;
add.s32 %r957, %r38, %r956;
mad.lo.s32 %r77, %r4, %r37, %r964;
add.s32 %r78, %r77, %r62;
add.s32 %r79, %r78, %r66;
mul.wide.s32 %rd27, %r79, 18;
mul.wide.s32 %rd28, %r73, 2;
add.s64 %rd29, %rd27, %rd28;
cvta.to.global.u64 %rd30, %rd20;
add.s64 %rd31, %rd30, %rd29;
add.s64 %rd124, %rd31, 8;
add.s64 %rd123, %rd30, %rd27;
add.s32 %r80, %r964, %r62;
add.s32 %r81, %r80, %r66;
mad.lo.s32 %r82, %r68, %r37, %r81;
mul.wide.s32 %rd32, %r82, 18;
add.s64 %rd33, %rd32, %rd28;
add.s64 %rd34, %rd30, %rd33;
add.s64 %rd122, %rd34, 4;
add.s64 %rd121, %rd30, %rd32;
cvta.to.global.u64 %rd6, %rd21;
mov.f32 %f353, 0f00000000;

$L__BB25_2:
cvt.u64.u32 %rd35, %r956;
add.s64 %rd36, %rd1, %rd35;
mul.lo.s64 %rd37, %rd36, 36;
ld.global.nc.u16 %rs1, [%rd123];

{ cvt.f32.f16 %f97, %rs1;}


ld.global.nc.u16 %rs17, [%rd124+-6];
cvt.s32.s16 %r611, %rs17;
add.s64 %rd38, %rd6, %rd37;
ld.global.nc.u32 %r84, [%rd38+4];
ld.global.nc.u32 %r88, [%rd38+8];
and.b32 %r612, %r611, 30583;
mov.u32 %r613, 33620223;
prmt.b32 %r614, %r613, %r613, %r612;
shr.u32 %r615, %r611, 2;
and.b32 %r616, %r615, 30583;
prmt.b32 %r617, %r613, %r613, %r616;
mov.u32 %r618, 20800;
prmt.b32 %r547, %r614, %r617, %r618;
mov.u32 %r619, 29538;
prmt.b32 %r551, %r614, %r617, %r619;
mov.u32 %r581, 0;

dp4a.s32.s32 %r83, %r84, %r547, %r581;


dp4a.s32.s32 %r87, %r88, %r551, %r83;

ld.global.nc.u16 %rs18, [%rd124+-4];
cvt.s32.s16 %r620, %rs18;
ld.global.nc.u32 %r92, [%rd38+12];
ld.global.nc.u32 %r96, [%rd38+16];
and.b32 %r621, %r620, 30583;
prmt.b32 %r622, %r613, %r613, %r621;
shr.u32 %r623, %r620, 2;
and.b32 %r624, %r623, 30583;
prmt.b32 %r625, %r613, %r613, %r624;
prmt.b32 %r555, %r622, %r625, %r618;
prmt.b32 %r559, %r622, %r625, %r619;

dp4a.s32.s32 %r91, %r92, %r555, %r87;


dp4a.s32.s32 %r95, %r96, %r559, %r91;

ld.global.nc.u16 %rs19, [%rd124+-2];
cvt.s32.s16 %r626, %rs19;
ld.global.nc.u32 %r100, [%rd38+20];
ld.global.nc.u32 %r104, [%rd38+24];
and.b32 %r627, %r626, 30583;
prmt.b32 %r628, %r613, %r613, %r627;
shr.u32 %r629, %r626, 2;
and.b32 %r630, %r629, 30583;
prmt.b32 %r631, %r613, %r613, %r630;
prmt.b32 %r563, %r628, %r631, %r618;
prmt.b32 %r567, %r628, %r631, %r619;

dp4a.s32.s32 %r99, %r100, %r563, %r95;


dp4a.s32.s32 %r103, %r104, %r567, %r99;

ld.global.nc.u16 %rs20, [%rd124];
cvt.s32.s16 %r632, %rs20;
ld.global.nc.u32 %r108, [%rd38+28];
ld.global.nc.u32 %r112, [%rd38+32];
and.b32 %r633, %r632, 30583;
prmt.b32 %r634, %r613, %r613, %r633;
shr.u32 %r635, %r632, 2;
and.b32 %r636, %r635, 30583;
prmt.b32 %r637, %r613, %r613, %r636;
prmt.b32 %r571, %r634, %r637, %r618;
prmt.b32 %r575, %r634, %r637, %r619;

dp4a.s32.s32 %r107, %r108, %r571, %r103;


dp4a.s32.s32 %r111, %r112, %r575, %r107;

ld.global.nc.u32 %r115, [%rd38];

{.reg .f16 low,high;
mov.b32 {low,high},%r115;
cvt.f32.f16 %f98, low;}


mul.ftz.f32 %f129, %f97, %f98;
cvt.rn.f32.s32 %f130, %r111;
fma.rn.ftz.f32 %f368, %f129, %f130, %f368;
ld.global.nc.u16 %rs2, [%rd121];

{ cvt.f32.f16 %f99, %rs2;}


ld.global.nc.u16 %rs21, [%rd122+-2];
cvt.s32.s16 %r638, %rs21;
and.b32 %r639, %r638, 30583;
prmt.b32 %r640, %r613, %r613, %r639;
shr.u32 %r641, %r638, 2;
and.b32 %r642, %r641, 30583;
prmt.b32 %r643, %r613, %r613, %r642;
prmt.b32 %r580, %r640, %r643, %r618;
prmt.b32 %r584, %r640, %r643, %r619;

dp4a.s32.s32 %r116, %r84, %r580, %r581;


dp4a.s32.s32 %r120, %r88, %r584, %r116;

ld.global.nc.u16 %rs22, [%rd122];
cvt.s32.s16 %r644, %rs22;
and.b32 %r645, %r644, 30583;
prmt.b32 %r646, %r613, %r613, %r645;
shr.u32 %r647, %r644, 2;
and.b32 %r648, %r647, 30583;
prmt.b32 %r649, %r613, %r613, %r648;
prmt.b32 %r588, %r646, %r649, %r618;
prmt.b32 %r592, %r646, %r649, %r619;

dp4a.s32.s32 %r124, %r92, %r588, %r120;


dp4a.s32.s32 %r128, %r96, %r592, %r124;

ld.global.nc.u16 %rs23, [%rd122+2];
cvt.s32.s16 %r650, %rs23;
and.b32 %r651, %r650, 30583;
prmt.b32 %r652, %r613, %r613, %r651;
shr.u32 %r653, %r650, 2;
and.b32 %r654, %r653, 30583;
prmt.b32 %r655, %r613, %r613, %r654;
prmt.b32 %r596, %r652, %r655, %r618;
prmt.b32 %r600, %r652, %r655, %r619;

dp4a.s32.s32 %r132, %r100, %r596, %r128;


dp4a.s32.s32 %r136, %r104, %r600, %r132;

ld.global.nc.u16 %rs24, [%rd122+4];
cvt.s32.s16 %r656, %rs24;
and.b32 %r657, %r656, 30583;
prmt.b32 %r658, %r613, %r613, %r657;
shr.u32 %r659, %r656, 2;
and.b32 %r660, %r659, 30583;
prmt.b32 %r661, %r613, %r613, %r660;
prmt.b32 %r604, %r658, %r661, %r618;
prmt.b32 %r608, %r658, %r661, %r619;

dp4a.s32.s32 %r140, %r108, %r604, %r136;


dp4a.s32.s32 %r144, %r112, %r608, %r140;


{.reg .f16 low,high;
mov.b32 {low,high},%r115;
cvt.f32.f16 %f100, low;}


mul.ftz.f32 %f131, %f99, %f100;
cvt.rn.f32.s32 %f132, %r144;
fma.rn.ftz.f32 %f367, %f131, %f132, %f367;
cvt.u64.u32 %rd39, %r957;
add.s64 %rd40, %rd1, %rd39;
mul.lo.s64 %rd41, %rd40, 36;

{ cvt.f32.f16 %f101, %rs1;}


add.s64 %rd42, %rd6, %rd41;
ld.global.nc.u32 %r150, [%rd42+4];
ld.global.nc.u32 %r154, [%rd42+8];

dp4a.s32.s32 %r149, %r150, %r547, %r581;


dp4a.s32.s32 %r153, %r154, %r551, %r149;

ld.global.nc.u32 %r158, [%rd42+12];
ld.global.nc.u32 %r162, [%rd42+16];

dp4a.s32.s32 %r157, %r158, %r555, %r153;


dp4a.s32.s32 %r161, %r162, %r559, %r157;

ld.global.nc.u32 %r166, [%rd42+20];
ld.global.nc.u32 %r170, [%rd42+24];

dp4a.s32.s32 %r165, %r166, %r563, %r161;


dp4a.s32.s32 %r169, %r170, %r567, %r165;

ld.global.nc.u32 %r174, [%rd42+28];
ld.global.nc.u32 %r178, [%rd42+32];

dp4a.s32.s32 %r173, %r174, %r571, %r169;


dp4a.s32.s32 %r177, %r178, %r575, %r173;

ld.global.nc.u32 %r181, [%rd42];

{.reg .f16 low,high;
mov.b32 {low,high},%r181;
cvt.f32.f16 %f102, low;}


mul.ftz.f32 %f133, %f101, %f102;
cvt.rn.f32.s32 %f134, %r177;
fma.rn.ftz.f32 %f366, %f133, %f134, %f366;

{ cvt.f32.f16 %f103, %rs2;}



dp4a.s32.s32 %r182, %r150, %r580, %r581;


dp4a.s32.s32 %r186, %r154, %r584, %r182;


dp4a.s32.s32 %r190, %r158, %r588, %r186;


dp4a.s32.s32 %r194, %r162, %r592, %r190;


dp4a.s32.s32 %r198, %r166, %r596, %r194;


dp4a.s32.s32 %r202, %r170, %r600, %r198;


dp4a.s32.s32 %r206, %r174, %r604, %r202;


dp4a.s32.s32 %r210, %r178, %r608, %r206;


{.reg .f16 low,high;
mov.b32 {low,high},%r181;
cvt.f32.f16 %f104, low;}


mul.ftz.f32 %f135, %f103, %f104;
cvt.rn.f32.s32 %f136, %r210;
fma.rn.ftz.f32 %f365, %f135, %f136, %f365;
cvt.u64.u32 %rd43, %r963;
add.s64 %rd44, %rd1, %rd43;
mul.lo.s64 %rd45, %rd44, 36;

{ cvt.f32.f16 %f105, %rs1;}


add.s64 %rd46, %rd6, %rd45;
ld.global.nc.u32 %r216, [%rd46+4];
ld.global.nc.u32 %r220, [%rd46+8];

dp4a.s32.s32 %r215, %r216, %r547, %r581;


dp4a.s32.s32 %r219, %r220, %r551, %r215;

ld.global.nc.u32 %r224, [%rd46+12];
ld.global.nc.u32 %r228, [%rd46+16];

dp4a.s32.s32 %r223, %r224, %r555, %r219;


dp4a.s32.s32 %r227, %r228, %r559, %r223;

ld.global.nc.u32 %r232, [%rd46+20];
ld.global.nc.u32 %r236, [%rd46+24];

dp4a.s32.s32 %r231, %r232, %r563, %r227;


dp4a.s32.s32 %r235, %r236, %r567, %r231;

ld.global.nc.u32 %r240, [%rd46+28];
ld.global.nc.u32 %r244, [%rd46+32];

dp4a.s32.s32 %r239, %r240, %r571, %r235;


dp4a.s32.s32 %r243, %r244, %r575, %r239;

ld.global.nc.u32 %r247, [%rd46];

{.reg .f16 low,high;
mov.b32 {low,high},%r247;
cvt.f32.f16 %f106, low;}


mul.ftz.f32 %f137, %f105, %f106;
cvt.rn.f32.s32 %f138, %r243;
fma.rn.ftz.f32 %f364, %f137, %f138, %f364;

{ cvt.f32.f16 %f107, %rs2;}



dp4a.s32.s32 %r248, %r216, %r580, %r581;


dp4a.s32.s32 %r252, %r220, %r584, %r248;


dp4a.s32.s32 %r256, %r224, %r588, %r252;


dp4a.s32.s32 %r260, %r228, %r592, %r256;


dp4a.s32.s32 %r264, %r232, %r596, %r260;


dp4a.s32.s32 %r268, %r236, %r600, %r264;


dp4a.s32.s32 %r272, %r240, %r604, %r268;


dp4a.s32.s32 %r276, %r244, %r608, %r272;


{.reg .f16 low,high;
mov.b32 {low,high},%r247;
cvt.f32.f16 %f108, low;}


mul.ftz.f32 %f139, %f107, %f108;
cvt.rn.f32.s32 %f140, %r276;
fma.rn.ftz.f32 %f363, %f139, %f140, %f363;
cvt.u64.u32 %rd47, %r962;
add.s64 %rd48, %rd1, %rd47;
mul.lo.s64 %rd49, %rd48, 36;

{ cvt.f32.f16 %f109, %rs1;}


add.s64 %rd50, %rd6, %rd49;
ld.global.nc.u32 %r282, [%rd50+4];
ld.global.nc.u32 %r286, [%rd50+8];

dp4a.s32.s32 %r281, %r282, %r547, %r581;


dp4a.s32.s32 %r285, %r286, %r551, %r281;

ld.global.nc.u32 %r290, [%rd50+12];
ld.global.nc.u32 %r294, [%rd50+16];

dp4a.s32.s32 %r289, %r290, %r555, %r285;


dp4a.s32.s32 %r293, %r294, %r559, %r289;

ld.global.nc.u32 %r298, [%rd50+20];
ld.global.nc.u32 %r302, [%rd50+24];

dp4a.s32.s32 %r297, %r298, %r563, %r293;


dp4a.s32.s32 %r301, %r302, %r567, %r297;

ld.global.nc.u32 %r306, [%rd50+28];
ld.global.nc.u32 %r310, [%rd50+32];

dp4a.s32.s32 %r305, %r306, %r571, %r301;


dp4a.s32.s32 %r309, %r310, %r575, %r305;

ld.global.nc.u32 %r313, [%rd50];

{.reg .f16 low,high;
mov.b32 {low,high},%r313;
cvt.f32.f16 %f110, low;}


mul.ftz.f32 %f141, %f109, %f110;
cvt.rn.f32.s32 %f142, %r309;
fma.rn.ftz.f32 %f362, %f141, %f142, %f362;

{ cvt.f32.f16 %f111, %rs2;}



dp4a.s32.s32 %r314, %r282, %r580, %r581;


dp4a.s32.s32 %r318, %r286, %r584, %r314;


dp4a.s32.s32 %r322, %r290, %r588, %r318;


dp4a.s32.s32 %r326, %r294, %r592, %r322;


dp4a.s32.s32 %r330, %r298, %r596, %r326;


dp4a.s32.s32 %r334, %r302, %r600, %r330;


dp4a.s32.s32 %r338, %r306, %r604, %r334;


dp4a.s32.s32 %r342, %r310, %r608, %r338;


{.reg .f16 low,high;
mov.b32 {low,high},%r313;
cvt.f32.f16 %f112, low;}


mul.ftz.f32 %f143, %f111, %f112;
cvt.rn.f32.s32 %f144, %r342;
fma.rn.ftz.f32 %f361, %f143, %f144, %f361;
cvt.u64.u32 %rd51, %r961;
add.s64 %rd52, %rd1, %rd51;
mul.lo.s64 %rd53, %rd52, 36;

{ cvt.f32.f16 %f113, %rs1;}


add.s64 %rd54, %rd6, %rd53;
ld.global.nc.u32 %r348, [%rd54+4];
ld.global.nc.u32 %r352, [%rd54+8];

dp4a.s32.s32 %r347, %r348, %r547, %r581;


dp4a.s32.s32 %r351, %r352, %r551, %r347;

ld.global.nc.u32 %r356, [%rd54+12];
ld.global.nc.u32 %r360, [%rd54+16];

dp4a.s32.s32 %r355, %r356, %r555, %r351;


dp4a.s32.s32 %r359, %r360, %r559, %r355;

ld.global.nc.u32 %r364, [%rd54+20];
ld.global.nc.u32 %r368, [%rd54+24];

dp4a.s32.s32 %r363, %r364, %r563, %r359;


dp4a.s32.s32 %r367, %r368, %r567, %r363;

ld.global.nc.u32 %r372, [%rd54+28];
ld.global.nc.u32 %r376, [%rd54+32];

dp4a.s32.s32 %r371, %r372, %r571, %r367;


dp4a.s32.s32 %r375, %r376, %r575, %r371;

ld.global.nc.u32 %r379, [%rd54];

{.reg .f16 low,high;
mov.b32 {low,high},%r379;
cvt.f32.f16 %f114, low;}


mul.ftz.f32 %f145, %f113, %f114;
cvt.rn.f32.s32 %f146, %r375;
fma.rn.ftz.f32 %f360, %f145, %f146, %f360;

{ cvt.f32.f16 %f115, %rs2;}



dp4a.s32.s32 %r380, %r348, %r580, %r581;


dp4a.s32.s32 %r384, %r352, %r584, %r380;


dp4a.s32.s32 %r388, %r356, %r588, %r384;


dp4a.s32.s32 %r392, %r360, %r592, %r388;


dp4a.s32.s32 %r396, %r364, %r596, %r392;


dp4a.s32.s32 %r400, %r368, %r600, %r396;


dp4a.s32.s32 %r404, %r372, %r604, %r400;


dp4a.s32.s32 %r408, %r376, %r608, %r404;


{.reg .f16 low,high;
mov.b32 {low,high},%r379;
cvt.f32.f16 %f116, low;}


mul.ftz.f32 %f147, %f115, %f116;
cvt.rn.f32.s32 %f148, %r408;
fma.rn.ftz.f32 %f359, %f147, %f148, %f359;
cvt.u64.u32 %rd55, %r960;
add.s64 %rd56, %rd1, %rd55;
mul.lo.s64 %rd57, %rd56, 36;

{ cvt.f32.f16 %f117, %rs1;}


add.s64 %rd58, %rd6, %rd57;
ld.global.nc.u32 %r414, [%rd58+4];
ld.global.nc.u32 %r418, [%rd58+8];

dp4a.s32.s32 %r413, %r414, %r547, %r581;


dp4a.s32.s32 %r417, %r418, %r551, %r413;

ld.global.nc.u32 %r422, [%rd58+12];
ld.global.nc.u32 %r426, [%rd58+16];

dp4a.s32.s32 %r421, %r422, %r555, %r417;


dp4a.s32.s32 %r425, %r426, %r559, %r421;

ld.global.nc.u32 %r430, [%rd58+20];
ld.global.nc.u32 %r434, [%rd58+24];

dp4a.s32.s32 %r429, %r430, %r563, %r425;


dp4a.s32.s32 %r433, %r434, %r567, %r429;

ld.global.nc.u32 %r438, [%rd58+28];
ld.global.nc.u32 %r442, [%rd58+32];

dp4a.s32.s32 %r437, %r438, %r571, %r433;


dp4a.s32.s32 %r441, %r442, %r575, %r437;

ld.global.nc.u32 %r445, [%rd58];

{.reg .f16 low,high;
mov.b32 {low,high},%r445;
cvt.f32.f16 %f118, low;}


mul.ftz.f32 %f149, %f117, %f118;
cvt.rn.f32.s32 %f150, %r441;
fma.rn.ftz.f32 %f358, %f149, %f150, %f358;

{ cvt.f32.f16 %f119, %rs2;}



dp4a.s32.s32 %r446, %r414, %r580, %r581;


dp4a.s32.s32 %r450, %r418, %r584, %r446;


dp4a.s32.s32 %r454, %r422, %r588, %r450;


dp4a.s32.s32 %r458, %r426, %r592, %r454;


dp4a.s32.s32 %r462, %r430, %r596, %r458;


dp4a.s32.s32 %r466, %r434, %r600, %r462;


dp4a.s32.s32 %r470, %r438, %r604, %r466;


dp4a.s32.s32 %r474, %r442, %r608, %r470;


{.reg .f16 low,high;
mov.b32 {low,high},%r445;
cvt.f32.f16 %f120, low;}


mul.ftz.f32 %f151, %f119, %f120;
cvt.rn.f32.s32 %f152, %r474;
fma.rn.ftz.f32 %f357, %f151, %f152, %f357;
cvt.u64.u32 %rd59, %r959;
add.s64 %rd60, %rd1, %rd59;
mul.lo.s64 %rd61, %rd60, 36;

{ cvt.f32.f16 %f121, %rs1;}


add.s64 %rd62, %rd6, %rd61;
ld.global.nc.u32 %r480, [%rd62+4];
ld.global.nc.u32 %r484, [%rd62+8];

dp4a.s32.s32 %r479, %r480, %r547, %r581;


dp4a.s32.s32 %r483, %r484, %r551, %r479;

ld.global.nc.u32 %r488, [%rd62+12];
ld.global.nc.u32 %r492, [%rd62+16];

dp4a.s32.s32 %r487, %r488, %r555, %r483;


dp4a.s32.s32 %r491, %r492, %r559, %r487;

ld.global.nc.u32 %r496, [%rd62+20];
ld.global.nc.u32 %r500, [%rd62+24];

dp4a.s32.s32 %r495, %r496, %r563, %r491;


dp4a.s32.s32 %r499, %r500, %r567, %r495;

ld.global.nc.u32 %r504, [%rd62+28];
ld.global.nc.u32 %r508, [%rd62+32];

dp4a.s32.s32 %r503, %r504, %r571, %r499;


dp4a.s32.s32 %r507, %r508, %r575, %r503;

ld.global.nc.u32 %r511, [%rd62];

{.reg .f16 low,high;
mov.b32 {low,high},%r511;
cvt.f32.f16 %f122, low;}


mul.ftz.f32 %f153, %f121, %f122;
cvt.rn.f32.s32 %f154, %r507;
fma.rn.ftz.f32 %f356, %f153, %f154, %f356;

{ cvt.f32.f16 %f123, %rs2;}



dp4a.s32.s32 %r512, %r480, %r580, %r581;


dp4a.s32.s32 %r516, %r484, %r584, %r512;


dp4a.s32.s32 %r520, %r488, %r588, %r516;


dp4a.s32.s32 %r524, %r492, %r592, %r520;


dp4a.s32.s32 %r528, %r496, %r596, %r524;


dp4a.s32.s32 %r532, %r500, %r600, %r528;


dp4a.s32.s32 %r536, %r504, %r604, %r532;


dp4a.s32.s32 %r540, %r508, %r608, %r536;


{.reg .f16 low,high;
mov.b32 {low,high},%r511;
cvt.f32.f16 %f124, low;}


mul.ftz.f32 %f155, %f123, %f124;
cvt.rn.f32.s32 %f156, %r540;
fma.rn.ftz.f32 %f355, %f155, %f156, %f355;
cvt.u64.u32 %rd63, %r958;
add.s64 %rd64, %rd1, %rd63;
mul.lo.s64 %rd65, %rd64, 36;

{ cvt.f32.f16 %f125, %rs1;}


add.s64 %rd66, %rd6, %rd65;
ld.global.nc.u32 %r546, [%rd66+4];
ld.global.nc.u32 %r550, [%rd66+8];

dp4a.s32.s32 %r545, %r546, %r547, %r581;


dp4a.s32.s32 %r549, %r550, %r551, %r545;

ld.global.nc.u32 %r554, [%rd66+12];
ld.global.nc.u32 %r558, [%rd66+16];

dp4a.s32.s32 %r553, %r554, %r555, %r549;


dp4a.s32.s32 %r557, %r558, %r559, %r553;

ld.global.nc.u32 %r562, [%rd66+20];
ld.global.nc.u32 %r566, [%rd66+24];

dp4a.s32.s32 %r561, %r562, %r563, %r557;


dp4a.s32.s32 %r565, %r566, %r567, %r561;

ld.global.nc.u32 %r570, [%rd66+28];
ld.global.nc.u32 %r574, [%rd66+32];

dp4a.s32.s32 %r569, %r570, %r571, %r565;


dp4a.s32.s32 %r573, %r574, %r575, %r569;

ld.global.nc.u32 %r577, [%rd66];

{.reg .f16 low,high;
mov.b32 {low,high},%r577;
cvt.f32.f16 %f126, low;}


mul.ftz.f32 %f157, %f125, %f126;
cvt.rn.f32.s32 %f158, %r573;
fma.rn.ftz.f32 %f354, %f157, %f158, %f354;

{ cvt.f32.f16 %f127, %rs2;}



dp4a.s32.s32 %r578, %r546, %r580, %r581;


dp4a.s32.s32 %r582, %r550, %r584, %r578;


dp4a.s32.s32 %r586, %r554, %r588, %r582;


dp4a.s32.s32 %r590, %r558, %r592, %r586;


dp4a.s32.s32 %r594, %r562, %r596, %r590;


dp4a.s32.s32 %r598, %r566, %r600, %r594;


dp4a.s32.s32 %r602, %r570, %r604, %r598;


dp4a.s32.s32 %r606, %r574, %r608, %r602;


{.reg .f16 low,high;
mov.b32 {low,high},%r577;
cvt.f32.f16 %f128, low;}


mul.ftz.f32 %f159, %f127, %f128;
cvt.rn.f32.s32 %f160, %r606;
fma.rn.ftz.f32 %f353, %f159, %f160, %f353;
add.s32 %r963, %r963, 64;
add.s32 %r962, %r962, 64;
add.s32 %r961, %r961, 64;
add.s32 %r960, %r960, 64;
add.s32 %r959, %r959, 64;
add.s32 %r958, %r958, 64;
add.s32 %r957, %r957, 64;
add.s32 %r956, %r956, 64;
add.s64 %rd124, %rd124, 576;
add.s64 %rd123, %rd123, 576;
add.s64 %rd122, %rd122, 576;
add.s64 %rd121, %rd121, 576;
add.s32 %r964, %r964, 32;
setp.lt.s32 %p4, %r964, %r6;
@%p4 bra $L__BB25_2;

$L__BB25_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB25_5;

shl.b32 %r662, %r1, 11;
mov.u32 %r663, _ZZ13mul_mat_vec_qIL9ggml_type42ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r664, %r663, %r662;
shl.b32 %r665, %r2, 2;
add.s32 %r666, %r664, %r665;
st.shared.f32 [%r666+-2048], %f368;
st.shared.f32 [%r666+-1920], %f367;
st.shared.f32 [%r666+-1792], %f366;
st.shared.f32 [%r666+-1664], %f365;
st.shared.f32 [%r666+-1536], %f364;
st.shared.f32 [%r666+-1408], %f363;
st.shared.f32 [%r666+-1280], %f362;
st.shared.f32 [%r666+-1152], %f361;
st.shared.f32 [%r666+-1024], %f360;
st.shared.f32 [%r666+-896], %f359;
st.shared.f32 [%r666+-768], %f358;
st.shared.f32 [%r666+-640], %f357;
st.shared.f32 [%r666+-512], %f356;
st.shared.f32 [%r666+-384], %f355;
st.shared.f32 [%r666+-256], %f354;
st.shared.f32 [%r666+-128], %f353;

$L__BB25_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB25_38;

mov.u32 %r667, %ctaid.z;
mov.u32 %r668, %ctaid.y;
mad.lo.s32 %r669, %r668, %r45, %r4;
mad.lo.s32 %r670, %r667, %r51, %r669;
cvt.u64.u32 %rd15, %r670;
shl.b32 %r671, %r2, 2;
mov.u32 %r672, _ZZ13mul_mat_vec_qIL9ggml_type42ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r33, %r672, %r671;
mov.u32 %r673, 2;
ld.shared.f32 %f161, [%r33];
add.ftz.f32 %f162, %f161, %f368;
mov.b32 %r674, %f162;
mov.u32 %r675, 31;
mov.u32 %r676, 16;
mov.u32 %r677, -1;
shfl.sync.bfly.b32 %r678|%p7, %r674, %r676, %r675, %r677;
mov.b32 %f163, %r678;
add.ftz.f32 %f164, %f162, %f163;
mov.b32 %r679, %f164;
mov.u32 %r680, 8;
shfl.sync.bfly.b32 %r681|%p8, %r679, %r680, %r675, %r677;
mov.b32 %f165, %r681;
add.ftz.f32 %f166, %f164, %f165;
mov.b32 %r682, %f166;
mov.u32 %r683, 4;
shfl.sync.bfly.b32 %r684|%p9, %r682, %r683, %r675, %r677;
mov.b32 %f167, %r684;
add.ftz.f32 %f168, %f166, %f167;
mov.b32 %r685, %f168;
shfl.sync.bfly.b32 %r686|%p10, %r685, %r673, %r675, %r677;
mov.b32 %f169, %r686;
add.ftz.f32 %f170, %f168, %f169;
mov.b32 %r687, %f170;
mov.u32 %r688, 1;
shfl.sync.bfly.b32 %r689|%p11, %r687, %r688, %r675, %r677;
mov.b32 %f171, %r689;
add.ftz.f32 %f49, %f170, %f171;
setp.lt.u32 %p12, %r4, %r39;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
cvta.to.global.u64 %rd67, %rd22;
mul.wide.u32 %rd68, %r670, 4;
add.s64 %rd16, %rd67, %rd68;
not.pred %p14, %p1;
@%p14 bra $L__BB25_8;

st.global.f32 [%rd16], %f49;

$L__BB25_8:
ld.shared.f32 %f172, [%r33+128];
add.ftz.f32 %f173, %f172, %f367;
mov.b32 %r690, %f173;
shfl.sync.bfly.b32 %r694|%p15, %r690, %r676, %r675, %r677;
mov.b32 %f174, %r694;
add.ftz.f32 %f175, %f173, %f174;
mov.b32 %r695, %f175;
shfl.sync.bfly.b32 %r697|%p16, %r695, %r680, %r675, %r677;
mov.b32 %f176, %r697;
add.ftz.f32 %f177, %f175, %f176;
mov.b32 %r698, %f177;
shfl.sync.bfly.b32 %r700|%p17, %r698, %r683, %r675, %r677;
mov.b32 %f178, %r700;
add.ftz.f32 %f179, %f177, %f178;
mov.b32 %r701, %f179;
shfl.sync.bfly.b32 %r703|%p18, %r701, %r673, %r675, %r677;
mov.b32 %f180, %r703;
add.ftz.f32 %f181, %f179, %f180;
mov.b32 %r704, %f181;
shfl.sync.bfly.b32 %r706|%p19, %r704, %r688, %r675, %r677;
mov.b32 %f182, %r706;
add.ftz.f32 %f50, %f181, %f182;
setp.eq.s32 %p20, %r2, 1;
or.b32 %r710, %r4, 1;
setp.lt.u32 %p21, %r710, %r39;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB25_10;

st.global.f32 [%rd16+4], %f50;

$L__BB25_10:
ld.shared.f32 %f183, [%r33+256];
add.ftz.f32 %f184, %f183, %f366;
mov.b32 %r711, %f184;
mov.u32 %r712, 31;
mov.u32 %r713, 16;
mov.u32 %r714, -1;
shfl.sync.bfly.b32 %r715|%p23, %r711, %r713, %r712, %r714;
mov.b32 %f185, %r715;
add.ftz.f32 %f186, %f184, %f185;
mov.b32 %r716, %f186;
mov.u32 %r717, 8;
shfl.sync.bfly.b32 %r718|%p24, %r716, %r717, %r712, %r714;
mov.b32 %f187, %r718;
add.ftz.f32 %f188, %f186, %f187;
mov.b32 %r719, %f188;
mov.u32 %r720, 4;
shfl.sync.bfly.b32 %r721|%p25, %r719, %r720, %r712, %r714;
mov.b32 %f189, %r721;
add.ftz.f32 %f190, %f188, %f189;
mov.b32 %r722, %f190;
mov.u32 %r723, 2;
shfl.sync.bfly.b32 %r724|%p26, %r722, %r723, %r712, %r714;
mov.b32 %f191, %r724;
add.ftz.f32 %f192, %f190, %f191;
mov.b32 %r725, %f192;
mov.u32 %r726, 1;
shfl.sync.bfly.b32 %r727|%p27, %r725, %r726, %r712, %r714;
mov.b32 %f193, %r727;
add.ftz.f32 %f51, %f192, %f193;
@%p14 bra $L__BB25_12;

cvt.u64.u32 %rd69, %r39;
add.s64 %rd70, %rd69, %rd15;
shl.b64 %rd72, %rd70, 2;
add.s64 %rd73, %rd67, %rd72;
st.global.f32 [%rd73], %f51;

$L__BB25_12:
ld.shared.f32 %f194, [%r33+384];
add.ftz.f32 %f195, %f194, %f365;
mov.b32 %r728, %f195;
shfl.sync.bfly.b32 %r732|%p29, %r728, %r713, %r712, %r714;
mov.b32 %f196, %r732;
add.ftz.f32 %f197, %f195, %f196;
mov.b32 %r733, %f197;
shfl.sync.bfly.b32 %r735|%p30, %r733, %r717, %r712, %r714;
mov.b32 %f198, %r735;
add.ftz.f32 %f199, %f197, %f198;
mov.b32 %r736, %f199;
shfl.sync.bfly.b32 %r738|%p31, %r736, %r720, %r712, %r714;
mov.b32 %f200, %r738;
add.ftz.f32 %f201, %f199, %f200;
mov.b32 %r739, %f201;
shfl.sync.bfly.b32 %r741|%p32, %r739, %r723, %r712, %r714;
mov.b32 %f202, %r741;
add.ftz.f32 %f203, %f201, %f202;
mov.b32 %r742, %f203;
shfl.sync.bfly.b32 %r744|%p33, %r742, %r726, %r712, %r714;
mov.b32 %f204, %r744;
add.ftz.f32 %f52, %f203, %f204;
@%p22 bra $L__BB25_14;

add.s32 %r745, %r39, 1;
cvt.u64.u32 %rd74, %r745;
add.s64 %rd75, %rd74, %rd15;
shl.b64 %rd77, %rd75, 2;
add.s64 %rd78, %rd67, %rd77;
st.global.f32 [%rd78], %f52;

$L__BB25_14:
shl.b32 %r746, %r39, 1;
mov.u32 %r747, 1;
ld.shared.f32 %f205, [%r33+512];
add.ftz.f32 %f206, %f205, %f364;
mov.b32 %r748, %f206;
mov.u32 %r749, 31;
mov.u32 %r750, 16;
mov.u32 %r751, -1;
shfl.sync.bfly.b32 %r752|%p35, %r748, %r750, %r749, %r751;
mov.b32 %f207, %r752;
add.ftz.f32 %f208, %f206, %f207;
mov.b32 %r753, %f208;
mov.u32 %r754, 8;
shfl.sync.bfly.b32 %r755|%p36, %r753, %r754, %r749, %r751;
mov.b32 %f209, %r755;
add.ftz.f32 %f210, %f208, %f209;
mov.b32 %r756, %f210;
mov.u32 %r757, 4;
shfl.sync.bfly.b32 %r758|%p37, %r756, %r757, %r749, %r751;
mov.b32 %f211, %r758;
add.ftz.f32 %f212, %f210, %f211;
mov.b32 %r759, %f212;
mov.u32 %r760, 2;
shfl.sync.bfly.b32 %r761|%p38, %r759, %r760, %r749, %r751;
mov.b32 %f213, %r761;
add.ftz.f32 %f214, %f212, %f213;
mov.b32 %r762, %f214;
shfl.sync.bfly.b32 %r763|%p39, %r762, %r747, %r749, %r751;
mov.b32 %f215, %r763;
add.ftz.f32 %f53, %f214, %f215;
cvt.u64.u32 %rd79, %r746;
add.s64 %rd80, %rd79, %rd15;
shl.b64 %rd82, %rd80, 2;
add.s64 %rd17, %rd67, %rd82;
@%p14 bra $L__BB25_16;

st.global.f32 [%rd17], %f53;

$L__BB25_16:
ld.shared.f32 %f216, [%r33+640];
add.ftz.f32 %f217, %f216, %f363;
mov.b32 %r764, %f217;
shfl.sync.bfly.b32 %r768|%p41, %r764, %r750, %r749, %r751;
mov.b32 %f218, %r768;
add.ftz.f32 %f219, %f217, %f218;
mov.b32 %r769, %f219;
shfl.sync.bfly.b32 %r771|%p42, %r769, %r754, %r749, %r751;
mov.b32 %f220, %r771;
add.ftz.f32 %f221, %f219, %f220;
mov.b32 %r772, %f221;
shfl.sync.bfly.b32 %r774|%p43, %r772, %r757, %r749, %r751;
mov.b32 %f222, %r774;
add.ftz.f32 %f223, %f221, %f222;
mov.b32 %r775, %f223;
shfl.sync.bfly.b32 %r777|%p44, %r775, %r760, %r749, %r751;
mov.b32 %f224, %r777;
add.ftz.f32 %f225, %f223, %f224;
mov.b32 %r778, %f225;
shfl.sync.bfly.b32 %r780|%p45, %r778, %r747, %r749, %r751;
mov.b32 %f226, %r780;
add.ftz.f32 %f54, %f225, %f226;
@%p22 bra $L__BB25_18;

st.global.f32 [%rd17+4], %f54;

$L__BB25_18:
mul.lo.s32 %r34, %r39, 3;
ld.shared.f32 %f227, [%r33+768];
add.ftz.f32 %f228, %f227, %f362;
mov.b32 %r781, %f228;
mov.u32 %r782, 31;
mov.u32 %r783, 16;
mov.u32 %r784, -1;
shfl.sync.bfly.b32 %r785|%p47, %r781, %r783, %r782, %r784;
mov.b32 %f229, %r785;
add.ftz.f32 %f230, %f228, %f229;
mov.b32 %r786, %f230;
mov.u32 %r787, 8;
shfl.sync.bfly.b32 %r788|%p48, %r786, %r787, %r782, %r784;
mov.b32 %f231, %r788;
add.ftz.f32 %f232, %f230, %f231;
mov.b32 %r789, %f232;
mov.u32 %r790, 4;
shfl.sync.bfly.b32 %r791|%p49, %r789, %r790, %r782, %r784;
mov.b32 %f233, %r791;
add.ftz.f32 %f234, %f232, %f233;
mov.b32 %r792, %f234;
mov.u32 %r793, 2;
shfl.sync.bfly.b32 %r794|%p50, %r792, %r793, %r782, %r784;
mov.b32 %f235, %r794;
add.ftz.f32 %f236, %f234, %f235;
mov.b32 %r795, %f236;
mov.u32 %r796, 1;
shfl.sync.bfly.b32 %r797|%p51, %r795, %r796, %r782, %r784;
mov.b32 %f237, %r797;
add.ftz.f32 %f55, %f236, %f237;
@%p14 bra $L__BB25_20;

cvt.u64.u32 %rd83, %r34;
add.s64 %rd84, %rd83, %rd15;
shl.b64 %rd86, %rd84, 2;
add.s64 %rd87, %rd67, %rd86;
st.global.f32 [%rd87], %f55;

$L__BB25_20:
ld.shared.f32 %f238, [%r33+896];
add.ftz.f32 %f239, %f238, %f361;
mov.b32 %r798, %f239;
shfl.sync.bfly.b32 %r802|%p53, %r798, %r783, %r782, %r784;
mov.b32 %f240, %r802;
add.ftz.f32 %f241, %f239, %f240;
mov.b32 %r803, %f241;
shfl.sync.bfly.b32 %r805|%p54, %r803, %r787, %r782, %r784;
mov.b32 %f242, %r805;
add.ftz.f32 %f243, %f241, %f242;
mov.b32 %r806, %f243;
shfl.sync.bfly.b32 %r808|%p55, %r806, %r790, %r782, %r784;
mov.b32 %f244, %r808;
add.ftz.f32 %f245, %f243, %f244;
mov.b32 %r809, %f245;
shfl.sync.bfly.b32 %r811|%p56, %r809, %r793, %r782, %r784;
mov.b32 %f246, %r811;
add.ftz.f32 %f247, %f245, %f246;
mov.b32 %r812, %f247;
shfl.sync.bfly.b32 %r814|%p57, %r812, %r796, %r782, %r784;
mov.b32 %f248, %r814;
add.ftz.f32 %f56, %f247, %f248;
@%p22 bra $L__BB25_22;

add.s32 %r815, %r34, 1;
cvt.u64.u32 %rd88, %r815;
add.s64 %rd89, %rd88, %rd15;
shl.b64 %rd91, %rd89, 2;
add.s64 %rd92, %rd67, %rd91;
st.global.f32 [%rd92], %f56;

$L__BB25_22:
shl.b32 %r816, %r39, 2;
mov.u32 %r817, 2;
ld.shared.f32 %f249, [%r33+1024];
add.ftz.f32 %f250, %f249, %f360;
mov.b32 %r818, %f250;
mov.u32 %r819, 31;
mov.u32 %r820, 16;
mov.u32 %r821, -1;
shfl.sync.bfly.b32 %r822|%p59, %r818, %r820, %r819, %r821;
mov.b32 %f251, %r822;
add.ftz.f32 %f252, %f250, %f251;
mov.b32 %r823, %f252;
mov.u32 %r824, 8;
shfl.sync.bfly.b32 %r825|%p60, %r823, %r824, %r819, %r821;
mov.b32 %f253, %r825;
add.ftz.f32 %f254, %f252, %f253;
mov.b32 %r826, %f254;
mov.u32 %r827, 4;
shfl.sync.bfly.b32 %r828|%p61, %r826, %r827, %r819, %r821;
mov.b32 %f255, %r828;
add.ftz.f32 %f256, %f254, %f255;
mov.b32 %r829, %f256;
shfl.sync.bfly.b32 %r830|%p62, %r829, %r817, %r819, %r821;
mov.b32 %f257, %r830;
add.ftz.f32 %f258, %f256, %f257;
mov.b32 %r831, %f258;
mov.u32 %r832, 1;
shfl.sync.bfly.b32 %r833|%p63, %r831, %r832, %r819, %r821;
mov.b32 %f259, %r833;
add.ftz.f32 %f57, %f258, %f259;
cvt.u64.u32 %rd93, %r816;
add.s64 %rd94, %rd93, %rd15;
shl.b64 %rd96, %rd94, 2;
add.s64 %rd18, %rd67, %rd96;
@%p14 bra $L__BB25_24;

st.global.f32 [%rd18], %f57;

$L__BB25_24:
ld.shared.f32 %f260, [%r33+1152];
add.ftz.f32 %f261, %f260, %f359;
mov.b32 %r834, %f261;
shfl.sync.bfly.b32 %r838|%p65, %r834, %r820, %r819, %r821;
mov.b32 %f262, %r838;
add.ftz.f32 %f263, %f261, %f262;
mov.b32 %r839, %f263;
shfl.sync.bfly.b32 %r841|%p66, %r839, %r824, %r819, %r821;
mov.b32 %f264, %r841;
add.ftz.f32 %f265, %f263, %f264;
mov.b32 %r842, %f265;
shfl.sync.bfly.b32 %r844|%p67, %r842, %r827, %r819, %r821;
mov.b32 %f266, %r844;
add.ftz.f32 %f267, %f265, %f266;
mov.b32 %r845, %f267;
shfl.sync.bfly.b32 %r847|%p68, %r845, %r817, %r819, %r821;
mov.b32 %f268, %r847;
add.ftz.f32 %f269, %f267, %f268;
mov.b32 %r848, %f269;
shfl.sync.bfly.b32 %r850|%p69, %r848, %r832, %r819, %r821;
mov.b32 %f270, %r850;
add.ftz.f32 %f58, %f269, %f270;
@%p22 bra $L__BB25_26;

st.global.f32 [%rd18+4], %f58;

$L__BB25_26:
mul.lo.s32 %r35, %r39, 5;
ld.shared.f32 %f271, [%r33+1280];
add.ftz.f32 %f272, %f271, %f358;
mov.b32 %r851, %f272;
mov.u32 %r852, 31;
mov.u32 %r853, 16;
mov.u32 %r854, -1;
shfl.sync.bfly.b32 %r855|%p71, %r851, %r853, %r852, %r854;
mov.b32 %f273, %r855;
add.ftz.f32 %f274, %f272, %f273;
mov.b32 %r856, %f274;
mov.u32 %r857, 8;
shfl.sync.bfly.b32 %r858|%p72, %r856, %r857, %r852, %r854;
mov.b32 %f275, %r858;
add.ftz.f32 %f276, %f274, %f275;
mov.b32 %r859, %f276;
mov.u32 %r860, 4;
shfl.sync.bfly.b32 %r861|%p73, %r859, %r860, %r852, %r854;
mov.b32 %f277, %r861;
add.ftz.f32 %f278, %f276, %f277;
mov.b32 %r862, %f278;
mov.u32 %r863, 2;
shfl.sync.bfly.b32 %r864|%p74, %r862, %r863, %r852, %r854;
mov.b32 %f279, %r864;
add.ftz.f32 %f280, %f278, %f279;
mov.b32 %r865, %f280;
mov.u32 %r866, 1;
shfl.sync.bfly.b32 %r867|%p75, %r865, %r866, %r852, %r854;
mov.b32 %f281, %r867;
add.ftz.f32 %f59, %f280, %f281;
@%p14 bra $L__BB25_28;

cvt.u64.u32 %rd97, %r35;
add.s64 %rd98, %rd97, %rd15;
shl.b64 %rd100, %rd98, 2;
add.s64 %rd101, %rd67, %rd100;
st.global.f32 [%rd101], %f59;

$L__BB25_28:
ld.shared.f32 %f282, [%r33+1408];
add.ftz.f32 %f283, %f282, %f357;
mov.b32 %r868, %f283;
shfl.sync.bfly.b32 %r872|%p77, %r868, %r853, %r852, %r854;
mov.b32 %f284, %r872;
add.ftz.f32 %f285, %f283, %f284;
mov.b32 %r873, %f285;
shfl.sync.bfly.b32 %r875|%p78, %r873, %r857, %r852, %r854;
mov.b32 %f286, %r875;
add.ftz.f32 %f287, %f285, %f286;
mov.b32 %r876, %f287;
shfl.sync.bfly.b32 %r878|%p79, %r876, %r860, %r852, %r854;
mov.b32 %f288, %r878;
add.ftz.f32 %f289, %f287, %f288;
mov.b32 %r879, %f289;
shfl.sync.bfly.b32 %r881|%p80, %r879, %r863, %r852, %r854;
mov.b32 %f290, %r881;
add.ftz.f32 %f291, %f289, %f290;
mov.b32 %r882, %f291;
shfl.sync.bfly.b32 %r884|%p81, %r882, %r866, %r852, %r854;
mov.b32 %f292, %r884;
add.ftz.f32 %f60, %f291, %f292;
@%p22 bra $L__BB25_30;

add.s32 %r885, %r35, 1;
cvt.u64.u32 %rd102, %r885;
add.s64 %rd103, %rd102, %rd15;
shl.b64 %rd105, %rd103, 2;
add.s64 %rd106, %rd67, %rd105;
st.global.f32 [%rd106], %f60;

$L__BB25_30:
mul.lo.s32 %r886, %r39, 6;
ld.shared.f32 %f293, [%r33+1536];
add.ftz.f32 %f294, %f293, %f356;
mov.b32 %r887, %f294;
mov.u32 %r888, 31;
mov.u32 %r889, 16;
mov.u32 %r890, -1;
shfl.sync.bfly.b32 %r891|%p83, %r887, %r889, %r888, %r890;
mov.b32 %f295, %r891;
add.ftz.f32 %f296, %f294, %f295;
mov.b32 %r892, %f296;
mov.u32 %r893, 8;
shfl.sync.bfly.b32 %r894|%p84, %r892, %r893, %r888, %r890;
mov.b32 %f297, %r894;
add.ftz.f32 %f298, %f296, %f297;
mov.b32 %r895, %f298;
mov.u32 %r896, 4;
shfl.sync.bfly.b32 %r897|%p85, %r895, %r896, %r888, %r890;
mov.b32 %f299, %r897;
add.ftz.f32 %f300, %f298, %f299;
mov.b32 %r898, %f300;
mov.u32 %r899, 2;
shfl.sync.bfly.b32 %r900|%p86, %r898, %r899, %r888, %r890;
mov.b32 %f301, %r900;
add.ftz.f32 %f302, %f300, %f301;
mov.b32 %r901, %f302;
mov.u32 %r902, 1;
shfl.sync.bfly.b32 %r903|%p87, %r901, %r902, %r888, %r890;
mov.b32 %f303, %r903;
add.ftz.f32 %f61, %f302, %f303;
cvt.u64.u32 %rd107, %r886;
add.s64 %rd108, %rd107, %rd15;
shl.b64 %rd110, %rd108, 2;
add.s64 %rd19, %rd67, %rd110;
@%p14 bra $L__BB25_32;

st.global.f32 [%rd19], %f61;

$L__BB25_32:
ld.shared.f32 %f304, [%r33+1664];
add.ftz.f32 %f305, %f304, %f355;
mov.b32 %r904, %f305;
shfl.sync.bfly.b32 %r908|%p89, %r904, %r889, %r888, %r890;
mov.b32 %f306, %r908;
add.ftz.f32 %f307, %f305, %f306;
mov.b32 %r909, %f307;
shfl.sync.bfly.b32 %r911|%p90, %r909, %r893, %r888, %r890;
mov.b32 %f308, %r911;
add.ftz.f32 %f309, %f307, %f308;
mov.b32 %r912, %f309;
shfl.sync.bfly.b32 %r914|%p91, %r912, %r896, %r888, %r890;
mov.b32 %f310, %r914;
add.ftz.f32 %f311, %f309, %f310;
mov.b32 %r915, %f311;
shfl.sync.bfly.b32 %r917|%p92, %r915, %r899, %r888, %r890;
mov.b32 %f312, %r917;
add.ftz.f32 %f313, %f311, %f312;
mov.b32 %r918, %f313;
shfl.sync.bfly.b32 %r920|%p93, %r918, %r902, %r888, %r890;
mov.b32 %f314, %r920;
add.ftz.f32 %f62, %f313, %f314;
@%p22 bra $L__BB25_34;

st.global.f32 [%rd19+4], %f62;

$L__BB25_34:
mul.lo.s32 %r36, %r39, 7;
ld.shared.f32 %f315, [%r33+1792];
add.ftz.f32 %f316, %f315, %f354;
mov.b32 %r921, %f316;
mov.u32 %r922, 31;
mov.u32 %r923, 16;
mov.u32 %r924, -1;
shfl.sync.bfly.b32 %r925|%p95, %r921, %r923, %r922, %r924;
mov.b32 %f317, %r925;
add.ftz.f32 %f318, %f316, %f317;
mov.b32 %r926, %f318;
mov.u32 %r927, 8;
shfl.sync.bfly.b32 %r928|%p96, %r926, %r927, %r922, %r924;
mov.b32 %f319, %r928;
add.ftz.f32 %f320, %f318, %f319;
mov.b32 %r929, %f320;
mov.u32 %r930, 4;
shfl.sync.bfly.b32 %r931|%p97, %r929, %r930, %r922, %r924;
mov.b32 %f321, %r931;
add.ftz.f32 %f322, %f320, %f321;
mov.b32 %r932, %f322;
mov.u32 %r933, 2;
shfl.sync.bfly.b32 %r934|%p98, %r932, %r933, %r922, %r924;
mov.b32 %f323, %r934;
add.ftz.f32 %f324, %f322, %f323;
mov.b32 %r935, %f324;
mov.u32 %r936, 1;
shfl.sync.bfly.b32 %r937|%p99, %r935, %r936, %r922, %r924;
mov.b32 %f325, %r937;
add.ftz.f32 %f63, %f324, %f325;
@%p14 bra $L__BB25_36;

cvt.u64.u32 %rd111, %r36;
add.s64 %rd112, %rd111, %rd15;
shl.b64 %rd114, %rd112, 2;
add.s64 %rd115, %rd67, %rd114;
st.global.f32 [%rd115], %f63;

$L__BB25_36:
ld.shared.f32 %f326, [%r33+1920];
add.ftz.f32 %f327, %f326, %f353;
mov.b32 %r938, %f327;
shfl.sync.bfly.b32 %r942|%p101, %r938, %r923, %r922, %r924;
mov.b32 %f328, %r942;
add.ftz.f32 %f329, %f327, %f328;
mov.b32 %r943, %f329;
shfl.sync.bfly.b32 %r945|%p102, %r943, %r927, %r922, %r924;
mov.b32 %f330, %r945;
add.ftz.f32 %f331, %f329, %f330;
mov.b32 %r946, %f331;
shfl.sync.bfly.b32 %r948|%p103, %r946, %r930, %r922, %r924;
mov.b32 %f332, %r948;
add.ftz.f32 %f333, %f331, %f332;
mov.b32 %r949, %f333;
shfl.sync.bfly.b32 %r951|%p104, %r949, %r933, %r922, %r924;
mov.b32 %f334, %r951;
add.ftz.f32 %f335, %f333, %f334;
mov.b32 %r952, %f335;
shfl.sync.bfly.b32 %r954|%p105, %r952, %r936, %r922, %r924;
mov.b32 %f336, %r954;
add.ftz.f32 %f64, %f335, %f336;
@%p22 bra $L__BB25_38;

add.s32 %r955, %r36, 1;
cvt.u64.u32 %rd116, %r955;
add.s64 %rd117, %rd116, %rd15;
shl.b64 %rd119, %rd117, 2;
add.s64 %rd120, %rd67, %rd119;
st.global.f32 [%rd120], %f64;

$L__BB25_38:
ret;

}
.entry _Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj(
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2,
.param .align 8 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3[48],
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5,
.param .align 4 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6[12],
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15
)
.maxntid 256, 1, 1
.minnctapersm 1
{
.local .align 8 .b8 __local_depot26[16];
.reg .b64 %SP;
.reg .b64 %SPL;
.reg .pred %p<48>;
.reg .b16 %rs<21>;
.reg .f32 %f<177>;
.reg .b32 %r<236>;
.reg .b64 %rd<82>;


mov.u64 %SPL, __local_depot26;
ld.param.u64 %rd23, [_Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0];
ld.param.u64 %rd24, [_Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1];
ld.param.u64 %rd25, [_Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2];
ld.param.u32 %r29, [_Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5];
ld.param.u32 %r22, [_Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8];
ld.param.u32 %r23, [_Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9];
ld.param.u32 %r25, [_Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11];
ld.param.u32 %r26, [_Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12];
ld.param.u32 %r30, [_Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14];
ld.param.u32 %r28, [_Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15];
ld.param.u32 %r20, [_Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+8];
ld.param.u32 %r19, [_Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+4];
ld.param.u32 %r18, [_Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6];
ld.param.u32 %r17, [_Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
ld.param.u64 %rd28, [_Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+16];
ld.param.u64 %rd27, [_Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+8];
ld.param.u64 %rd26, [_Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3];
add.u64 %rd1, %SPL, 0;
add.u64 %rd2, %SPL, 8;
shr.u32 %r1, %r29, 5;
mov.u32 %r2, %tid.y;
setp.ge.u32 %p1, %r2, %r30;
@%p1 bra $L__BB26_29;

mov.u32 %r3, %ctaid.y;
cvta.to.global.u64 %rd34, %rd25;
mad.lo.s32 %r31, %r2, %r28, %r3;
mul.wide.u32 %rd35, %r31, 4;
add.s64 %rd36, %rd34, %rd35;
ld.global.nc.u32 %r4, [%rd36];
mov.f32 %f170, 0f00000000;
st.local.v2.f32 [%rd1], {%f170, %f170};
st.local.v2.f32 [%rd2], {%f170, %f170};
mov.u32 %r5, %tid.x;
shr.u32 %r235, %r5, 1;
setp.ge.u32 %p2, %r235, %r1;
mov.f32 %f171, %f170;
mov.f32 %f172, %f170;
mov.f32 %f173, %f170;
@%p2 bra $L__BB26_9;

mov.u32 %r32, %ctaid.x;
shl.b32 %r33, %r32, 1;
cvta.to.global.u64 %rd37, %rd24;
mul.hi.u32 %r34, %r3, %r18;
add.s32 %r35, %r34, %r3;
shr.u32 %r36, %r35, %r19;
mul.lo.s32 %r37, %r36, %r20;
sub.s32 %r38, %r3, %r37;
mul.lo.s32 %r39, %r38, %r26;
cvt.u64.u32 %rd38, %r39;
mul.lo.s32 %r40, %r2, %r23;
cvt.u64.u32 %rd39, %r40;
mul.lo.s32 %r41, %r4, %r25;
cvt.u64.u32 %rd40, %r5;
shr.u64 %rd41, %rd40, 1;
mul.lo.s64 %rd42, %rd41, 36;
mul.wide.u32 %rd43, %r40, 36;
add.s64 %rd44, %rd42, %rd43;
add.s64 %rd4, %rd37, %rd44;
mul.wide.u32 %rd81, %r39, 36;
and.b32 %r42, %r5, 1;
mul.wide.u32 %rd45, %r42, 8;
mad.lo.s32 %r43, %r33, %r22, %r235;
add.s32 %r44, %r43, %r41;
mul.wide.s32 %rd46, %r44, 18;
add.s64 %rd47, %rd45, %rd46;
add.s64 %rd6, %rd47, 8;
add.s64 %rd48, %rd41, %rd39;
add.s64 %rd49, %rd48, %rd38;
mul.lo.s64 %rd50, %rd49, 36;
add.s64 %rd51, %rd50, %rd45;
add.s64 %rd52, %rd37, %rd51;
add.s64 %rd79, %rd52, 20;
or.b32 %r45, %r33, 1;
add.s32 %r46, %r235, %r41;
mad.lo.s32 %r47, %r22, %r45, %r46;
mul.wide.s32 %rd8, %r47, 18;
add.s64 %rd9, %rd45, %rd8;
cvta.to.global.u64 %rd80, %rd23;
cvta.to.global.u64 %rd78, %rd27;
mov.f32 %f170, 0f00000000;

$L__BB26_3:
add.s64 %rd53, %rd4, %rd81;
add.s64 %rd54, %rd80, %rd6;
ld.global.nc.u16 %rs2, [%rd54+-6];
ld.global.nc.u16 %rs3, [%rd54+-4];
mov.b32 %r66, {%rs2, %rs3};
ld.global.nc.u32 %r8, [%rd79+-16];
ld.global.nc.u32 %r9, [%rd79];
ld.global.nc.u16 %rs4, [%rd54+-2];
ld.global.nc.u16 %rs5, [%rd54];
mov.b32 %r67, {%rs4, %rs5};
ld.global.nc.u32 %r10, [%rd79+-12];
ld.global.nc.u32 %r11, [%rd79+4];
mad.lo.s32 %r73, %r4, %r25, %r43;
mul.wide.s32 %rd55, %r73, 18;
add.s64 %rd56, %rd80, %rd55;
ld.global.nc.u16 %rs1, [%rd56];

{ cvt.f32.f16 %f42, %rs1;}


and.b32 %r49, %r66, 252645135;
shr.u32 %r74, %r66, 4;
and.b32 %r53, %r74, 252645135;
mov.u32 %r51, 0;

dp4a.s32.s32 %r48, %r49, %r8, %r51;


dp4a.s32.s32 %r52, %r53, %r9, %r48;

and.b32 %r57, %r67, 252645135;
shr.u32 %r75, %r67, 4;
and.b32 %r61, %r75, 252645135;

dp4a.s32.s32 %r56, %r57, %r10, %r52;


dp4a.s32.s32 %r60, %r61, %r11, %r56;

ld.global.nc.u32 %r12, [%rd53];

{.reg .f16 low,high;
mov.b32 {low,high},%r12;
cvt.f32.f16 %f43, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r12;
cvt.f32.f16 %f44, high;}


cvt.rn.f32.s32 %f45, %r60;
mul.ftz.f32 %f46, %f44, 0fC0800000;
fma.rn.ftz.f32 %f47, %f43, %f45, %f46;
fma.rn.ftz.f32 %f173, %f42, %f47, %f173;
setp.eq.s64 %p3, %rd27, 0;
@%p3 bra $L__BB26_5;

add.s64 %rd57, %rd78, %rd6;
ld.global.u16 %rs7, [%rd57+-6];
ld.global.u16 %rs8, [%rd57+-4];
mov.b32 %r94, {%rs7, %rs8};
ld.global.u16 %rs9, [%rd57+-2];
ld.global.u16 %rs10, [%rd57];
mov.b32 %r95, {%rs9, %rs10};
shr.u32 %r97, %r5, 1;
mad.lo.s32 %r100, %r33, %r22, %r97;
mad.lo.s32 %r101, %r4, %r25, %r100;
mul.wide.s32 %rd58, %r101, 18;
add.s64 %rd59, %rd78, %rd58;
ld.global.u16 %rs6, [%rd59];

{ cvt.f32.f16 %f48, %rs6;}


and.b32 %r77, %r94, 252645135;
shr.u32 %r102, %r94, 4;
and.b32 %r81, %r102, 252645135;

dp4a.s32.s32 %r76, %r77, %r8, %r51;


dp4a.s32.s32 %r80, %r81, %r9, %r76;

and.b32 %r85, %r95, 252645135;
shr.u32 %r103, %r95, 4;
and.b32 %r89, %r103, 252645135;

dp4a.s32.s32 %r84, %r85, %r10, %r80;


dp4a.s32.s32 %r88, %r89, %r11, %r84;


{.reg .f16 low,high;
mov.b32 {low,high},%r12;
cvt.f32.f16 %f49, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r12;
cvt.f32.f16 %f50, high;}


cvt.rn.f32.s32 %f51, %r88;
mul.ftz.f32 %f52, %f50, 0fC0800000;
fma.rn.ftz.f32 %f53, %f49, %f51, %f52;
fma.rn.ftz.f32 %f171, %f48, %f53, %f171;

$L__BB26_5:
add.s64 %rd60, %rd80, %rd9;
ld.global.nc.u16 %rs12, [%rd60+2];
ld.global.nc.u16 %rs13, [%rd60+4];
mov.b32 %r122, {%rs12, %rs13};
ld.global.nc.u16 %rs14, [%rd60+6];
ld.global.nc.u16 %rs15, [%rd60+8];
mov.b32 %r123, {%rs14, %rs15};
add.s64 %rd61, %rd80, %rd8;
ld.global.nc.u16 %rs11, [%rd61];

{ cvt.f32.f16 %f54, %rs11;}


and.b32 %r105, %r122, 252645135;
shr.u32 %r124, %r122, 4;
and.b32 %r109, %r124, 252645135;

dp4a.s32.s32 %r104, %r105, %r8, %r51;


dp4a.s32.s32 %r108, %r109, %r9, %r104;

and.b32 %r113, %r123, 252645135;
shr.u32 %r125, %r123, 4;
and.b32 %r117, %r125, 252645135;

dp4a.s32.s32 %r112, %r113, %r10, %r108;


dp4a.s32.s32 %r116, %r117, %r11, %r112;


{.reg .f16 low,high;
mov.b32 {low,high},%r12;
cvt.f32.f16 %f55, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r12;
cvt.f32.f16 %f56, high;}


cvt.rn.f32.s32 %f57, %r116;
mul.ftz.f32 %f58, %f56, 0fC0800000;
fma.rn.ftz.f32 %f59, %f55, %f57, %f58;
fma.rn.ftz.f32 %f172, %f54, %f59, %f172;
@%p3 bra $L__BB26_7;

add.s64 %rd62, %rd78, %rd9;
ld.global.u16 %rs17, [%rd62+2];
ld.global.u16 %rs18, [%rd62+4];
mov.b32 %r144, {%rs17, %rs18};
ld.global.u16 %rs19, [%rd62+6];
ld.global.u16 %rs20, [%rd62+8];
mov.b32 %r145, {%rs19, %rs20};
add.s64 %rd63, %rd78, %rd8;
ld.global.u16 %rs16, [%rd63];

{ cvt.f32.f16 %f60, %rs16;}


and.b32 %r127, %r144, 252645135;
shr.u32 %r146, %r144, 4;
and.b32 %r131, %r146, 252645135;
mov.u32 %r129, 0;

dp4a.s32.s32 %r126, %r127, %r8, %r129;


dp4a.s32.s32 %r130, %r131, %r9, %r126;

and.b32 %r135, %r145, 252645135;
shr.u32 %r147, %r145, 4;
and.b32 %r139, %r147, 252645135;

dp4a.s32.s32 %r134, %r135, %r10, %r130;


dp4a.s32.s32 %r138, %r139, %r11, %r134;


{.reg .f16 low,high;
mov.b32 {low,high},%r12;
cvt.f32.f16 %f61, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r12;
cvt.f32.f16 %f62, high;}


cvt.rn.f32.s32 %f63, %r138;
mul.ftz.f32 %f64, %f62, 0fC0800000;
fma.rn.ftz.f32 %f65, %f61, %f63, %f64;
fma.rn.ftz.f32 %f170, %f60, %f65, %f170;

$L__BB26_7:
add.s64 %rd81, %rd81, 576;
add.s64 %rd80, %rd80, 288;
add.s64 %rd79, %rd79, 576;
add.s64 %rd78, %rd78, 288;
add.s32 %r235, %r235, 16;
setp.lt.u32 %p5, %r235, %r1;
@%p5 bra $L__BB26_3;

st.local.v2.f32 [%rd1], {%f173, %f172};
st.local.v2.f32 [%rd2], {%f171, %f170};

$L__BB26_9:
mov.b32 %r14, %f173;
setp.ne.s64 %p6, %rd27, 0;
@%p6 bra $L__BB26_11;
bra.uni $L__BB26_10;

$L__BB26_11:
mov.u32 %r174, 31;
mov.u32 %r175, 16;
mov.u32 %r176, -1;
shfl.sync.bfly.b32 %r177|%p17, %r14, %r175, %r174, %r176;
mov.b32 %f86, %r177;
add.ftz.f32 %f87, %f173, %f86;
mov.b32 %r178, %f87;
mov.u32 %r179, 8;
shfl.sync.bfly.b32 %r180|%p18, %r178, %r179, %r174, %r176;
mov.b32 %f88, %r180;
add.ftz.f32 %f89, %f87, %f88;
mov.b32 %r181, %f89;
mov.u32 %r182, 4;
shfl.sync.bfly.b32 %r183|%p19, %r181, %r182, %r174, %r176;
mov.b32 %f90, %r183;
add.ftz.f32 %f91, %f89, %f90;
mov.b32 %r184, %f91;
mov.u32 %r185, 2;
shfl.sync.bfly.b32 %r186|%p20, %r184, %r185, %r174, %r176;
mov.b32 %f92, %r186;
add.ftz.f32 %f93, %f91, %f92;
mov.b32 %r187, %f93;
mov.u32 %r188, 1;
shfl.sync.bfly.b32 %r189|%p21, %r187, %r188, %r174, %r176;
mov.b32 %f94, %r189;
add.ftz.f32 %f95, %f93, %f94;
st.local.f32 [%rd1], %f95;
mov.b32 %r190, %f171;
shfl.sync.bfly.b32 %r191|%p22, %r190, %r175, %r174, %r176;
mov.b32 %f96, %r191;
add.ftz.f32 %f97, %f171, %f96;
mov.b32 %r192, %f97;
shfl.sync.bfly.b32 %r193|%p23, %r192, %r179, %r174, %r176;
mov.b32 %f98, %r193;
add.ftz.f32 %f99, %f97, %f98;
mov.b32 %r194, %f99;
shfl.sync.bfly.b32 %r195|%p24, %r194, %r182, %r174, %r176;
mov.b32 %f100, %r195;
add.ftz.f32 %f101, %f99, %f100;
mov.b32 %r196, %f101;
shfl.sync.bfly.b32 %r197|%p25, %r196, %r185, %r174, %r176;
mov.b32 %f102, %r197;
add.ftz.f32 %f103, %f101, %f102;
mov.b32 %r198, %f103;
shfl.sync.bfly.b32 %r199|%p26, %r198, %r188, %r174, %r176;
mov.b32 %f104, %r199;
add.ftz.f32 %f105, %f103, %f104;
st.local.f32 [%rd2], %f105;
mov.b32 %r200, %f172;
shfl.sync.bfly.b32 %r201|%p27, %r200, %r175, %r174, %r176;
mov.b32 %f106, %r201;
add.ftz.f32 %f107, %f172, %f106;
mov.b32 %r202, %f107;
shfl.sync.bfly.b32 %r203|%p28, %r202, %r179, %r174, %r176;
mov.b32 %f108, %r203;
add.ftz.f32 %f109, %f107, %f108;
mov.b32 %r204, %f109;
shfl.sync.bfly.b32 %r205|%p29, %r204, %r182, %r174, %r176;
mov.b32 %f110, %r205;
add.ftz.f32 %f111, %f109, %f110;
mov.b32 %r206, %f111;
shfl.sync.bfly.b32 %r207|%p30, %r206, %r185, %r174, %r176;
mov.b32 %f112, %r207;
add.ftz.f32 %f113, %f111, %f112;
mov.b32 %r208, %f113;
shfl.sync.bfly.b32 %r209|%p31, %r208, %r188, %r174, %r176;
mov.b32 %f114, %r209;
add.ftz.f32 %f115, %f113, %f114;
st.local.f32 [%rd1+4], %f115;
mov.b32 %r210, %f170;
shfl.sync.bfly.b32 %r211|%p32, %r210, %r175, %r174, %r176;
mov.b32 %f116, %r211;
add.ftz.f32 %f117, %f170, %f116;
mov.b32 %r212, %f117;
shfl.sync.bfly.b32 %r213|%p33, %r212, %r179, %r174, %r176;
mov.b32 %f118, %r213;
add.ftz.f32 %f119, %f117, %f118;
mov.b32 %r214, %f119;
shfl.sync.bfly.b32 %r215|%p34, %r214, %r182, %r174, %r176;
mov.b32 %f120, %r215;
add.ftz.f32 %f121, %f119, %f120;
mov.b32 %r216, %f121;
shfl.sync.bfly.b32 %r217|%p35, %r216, %r185, %r174, %r176;
mov.b32 %f122, %r217;
add.ftz.f32 %f123, %f121, %f122;
mov.b32 %r218, %f123;
shfl.sync.bfly.b32 %r219|%p36, %r218, %r188, %r174, %r176;
mov.b32 %f124, %r219;
add.ftz.f32 %f125, %f123, %f124;
st.local.f32 [%rd2+4], %f125;
bra.uni $L__BB26_12;

$L__BB26_10:
mov.u32 %r148, 31;
mov.u32 %r149, 16;
mov.u32 %r150, -1;
shfl.sync.bfly.b32 %r151|%p7, %r14, %r149, %r148, %r150;
mov.b32 %f66, %r151;
add.ftz.f32 %f67, %f173, %f66;
mov.b32 %r152, %f67;
mov.u32 %r153, 8;
shfl.sync.bfly.b32 %r154|%p8, %r152, %r153, %r148, %r150;
mov.b32 %f68, %r154;
add.ftz.f32 %f69, %f67, %f68;
mov.b32 %r155, %f69;
mov.u32 %r156, 4;
shfl.sync.bfly.b32 %r157|%p9, %r155, %r156, %r148, %r150;
mov.b32 %f70, %r157;
add.ftz.f32 %f71, %f69, %f70;
mov.b32 %r158, %f71;
mov.u32 %r159, 2;
shfl.sync.bfly.b32 %r160|%p10, %r158, %r159, %r148, %r150;
mov.b32 %f72, %r160;
add.ftz.f32 %f73, %f71, %f72;
mov.b32 %r161, %f73;
mov.u32 %r162, 1;
shfl.sync.bfly.b32 %r163|%p11, %r161, %r162, %r148, %r150;
mov.b32 %f74, %r163;
add.ftz.f32 %f75, %f73, %f74;
st.local.f32 [%rd1], %f75;
mov.b32 %r164, %f172;
shfl.sync.bfly.b32 %r165|%p12, %r164, %r149, %r148, %r150;
mov.b32 %f76, %r165;
add.ftz.f32 %f77, %f172, %f76;
mov.b32 %r166, %f77;
shfl.sync.bfly.b32 %r167|%p13, %r166, %r153, %r148, %r150;
mov.b32 %f78, %r167;
add.ftz.f32 %f79, %f77, %f78;
mov.b32 %r168, %f79;
shfl.sync.bfly.b32 %r169|%p14, %r168, %r156, %r148, %r150;
mov.b32 %f80, %r169;
add.ftz.f32 %f81, %f79, %f80;
mov.b32 %r170, %f81;
shfl.sync.bfly.b32 %r171|%p15, %r170, %r159, %r148, %r150;
mov.b32 %f82, %r171;
add.ftz.f32 %f83, %f81, %f82;
mov.b32 %r172, %f83;
shfl.sync.bfly.b32 %r173|%p16, %r172, %r162, %r148, %r150;
mov.b32 %f84, %r173;
add.ftz.f32 %f85, %f83, %f84;
st.local.f32 [%rd1+4], %f85;

$L__BB26_12:
ld.param.u32 %r229, [_Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7];
setp.gt.u32 %p37, %r5, 1;
mov.u32 %r221, %ctaid.x;
shl.b32 %r222, %r221, 1;
add.s32 %r15, %r222, %r5;
setp.ge.u32 %p38, %r15, %r229;
or.pred %p39, %p37, %p38;
@%p39 bra $L__BB26_29;

ld.param.u32 %r230, [_Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13];
cvt.u64.u32 %rd21, %r5;
mul.wide.u32 %rd64, %r5, 4;
add.s64 %rd65, %rd1, %rd64;
ld.local.f32 %f176, [%rd65];
mad.lo.s32 %r16, %r4, %r230, %r15;
setp.eq.s64 %p40, %rd26, 0;
@%p40 bra $L__BB26_15;

cvta.to.global.u64 %rd66, %rd26;
mul.wide.u32 %rd67, %r16, 4;
add.s64 %rd68, %rd66, %rd67;
ld.global.f32 %f126, [%rd68];
add.ftz.f32 %f176, %f176, %f126;

$L__BB26_15:
setp.eq.s64 %p41, %rd27, 0;
@%p41 bra $L__BB26_28;

shl.b64 %rd69, %rd21, 2;
add.s64 %rd70, %rd2, %rd69;
ld.local.f32 %f24, [%rd70];
setp.eq.s64 %p42, %rd28, 0;
@%p42 bra $L__BB26_18;

cvta.to.global.u64 %rd71, %rd28;
mul.wide.u32 %rd72, %r16, 4;
add.s64 %rd73, %rd71, %rd72;
ld.global.f32 %f127, [%rd73];
add.ftz.f32 %f24, %f24, %f127;

$L__BB26_18:
setp.gt.s32 %p43, %r17, 2;
@%p43 bra $L__BB26_22;

setp.eq.s32 %p46, %r17, 1;
@%p46 bra $L__BB26_26;

setp.eq.s32 %p47, %r17, 2;
@%p47 bra $L__BB26_21;
bra.uni $L__BB26_27;

$L__BB26_21:
mul.ftz.f32 %f147, %f24, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f148, %f147;
add.ftz.f32 %f149, %f148, 0f3F800000;
div.approx.ftz.f32 %f150, %f24, %f149;
mul.ftz.f32 %f176, %f176, %f150;
bra.uni $L__BB26_28;

$L__BB26_22:
setp.eq.s32 %p44, %r17, 3;
@%p44 bra $L__BB26_25;

setp.ne.s32 %p45, %r17, 6;
@%p45 bra $L__BB26_27;

ld.param.f32 %f159, [_Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+44];
min.ftz.f32 %f128, %f24, %f159;
neg.ftz.f32 %f129, %f159;
min.ftz.f32 %f130, %f176, %f159;
max.ftz.f32 %f131, %f130, %f129;
mul.ftz.f32 %f132, %f128, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f133, %f132;
add.ftz.f32 %f134, %f133, 0f3F800000;
div.approx.ftz.f32 %f135, %f128, %f134;
mul.ftz.f32 %f176, %f131, %f135;
bra.uni $L__BB26_28;

$L__BB26_27:
mul.ftz.f32 %f176, %f176, %f24;
bra.uni $L__BB26_28;

$L__BB26_26:
mul.ftz.f32 %f151, %f24, 0f3F000000;
mul.ftz.f32 %f152, %f24, 0f3D372713;
fma.rn.ftz.f32 %f153, %f24, %f152, 0f3F800000;
mul.ftz.f32 %f154, %f24, 0f3F4C422A;
mul.ftz.f32 %f155, %f154, %f153;
tanh.approx.f32 %f156, %f155;
add.ftz.f32 %f157, %f156, 0f3F800000;
mul.ftz.f32 %f158, %f151, %f157;
mul.ftz.f32 %f176, %f176, %f158;
bra.uni $L__BB26_28;

$L__BB26_25:
mov.f32 %f136, 0f40E00000;
min.ftz.f32 %f137, %f24, %f136;
min.ftz.f32 %f138, %f176, %f136;
mov.f32 %f139, 0fC0E00000;
max.ftz.f32 %f140, %f138, %f139;
mul.ftz.f32 %f141, %f137, 0f3FD9DB23;
mul.ftz.f32 %f142, %f141, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f143, %f142;
add.ftz.f32 %f144, %f143, 0f3F800000;
div.approx.ftz.f32 %f145, %f137, %f144;
add.ftz.f32 %f146, %f140, 0f3F800000;
mul.ftz.f32 %f176, %f146, %f145;

$L__BB26_28:
ld.param.u64 %rd77, [_Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4];
ld.param.u32 %r234, [_Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10];
ld.param.u32 %r233, [_Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13];
mov.u32 %r232, %ctaid.y;
mov.u32 %r231, %tid.y;
mad.lo.s32 %r227, %r231, %r234, %r15;
mad.lo.s32 %r228, %r232, %r233, %r227;
cvta.to.global.u64 %rd74, %rd77;
mul.wide.u32 %rd75, %r228, 4;
add.s64 %rd76, %rd74, %rd75;
st.global.f32 [%rd76], %f176;

$L__BB26_29:
ret;

}
.entry _Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj(
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2,
.param .align 8 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3[48],
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5,
.param .align 4 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6[12],
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15
)
.maxntid 256, 1, 1
.minnctapersm 1
{
.local .align 8 .b8 __local_depot27[8];
.reg .b64 %SP;
.reg .b64 %SPL;
.reg .pred %p<17>;
.reg .b16 %rs<11>;
.reg .f32 %f<50>;
.reg .b32 %r<117>;
.reg .b64 %rd<54>;


mov.u64 %SPL, __local_depot27;
ld.param.u64 %rd17, [_Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0];
ld.param.u64 %rd18, [_Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1];
ld.param.u64 %rd19, [_Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2];
ld.param.u64 %rd20, [_Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4];
ld.param.u32 %r9, [_Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5];
ld.param.u32 %r13, [_Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7];
ld.param.u32 %r14, [_Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8];
ld.param.u32 %r15, [_Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9];
ld.param.u32 %r16, [_Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10];
ld.param.u32 %r17, [_Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11];
ld.param.u32 %r18, [_Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12];
ld.param.u32 %r19, [_Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13];
ld.param.u32 %r21, [_Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14];
ld.param.u32 %r20, [_Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15];
ld.param.u32 %r12, [_Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+8];
ld.param.u32 %r11, [_Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+4];
ld.param.u32 %r10, [_Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6];
add.u64 %rd1, %SPL, 0;
mov.u32 %r22, %ctaid.x;
shl.b32 %r1, %r22, 1;
mov.u32 %r2, %tid.y;
setp.ge.u32 %p1, %r2, %r21;
@%p1 bra $L__BB27_7;

mov.f32 %f48, 0f00000000;
st.local.v2.f32 [%rd1], {%f48, %f48};
mov.u32 %r3, %tid.x;
shr.u32 %r116, %r3, 1;
shr.u32 %r23, %r9, 5;
setp.ge.u32 %p2, %r116, %r23;
mov.f32 %f49, %f48;
@%p2 bra $L__BB27_5;

mov.u32 %r24, %ctaid.y;
cvta.to.global.u64 %rd22, %rd18;
mul.hi.u32 %r25, %r24, %r10;
add.s32 %r26, %r25, %r24;
shr.u32 %r27, %r26, %r11;
mul.lo.s32 %r28, %r27, %r12;
sub.s32 %r29, %r24, %r28;
mul.lo.s32 %r30, %r29, %r18;
cvt.u64.u32 %rd23, %r30;
mul.lo.s32 %r31, %r2, %r15;
cvt.u64.u32 %rd24, %r31;
mad.lo.s32 %r32, %r2, %r20, %r24;
cvta.to.global.u64 %rd25, %rd19;
mul.wide.u32 %rd26, %r32, 4;
add.s64 %rd27, %rd25, %rd26;
ld.global.nc.u32 %r33, [%rd27];
mul.lo.s32 %r34, %r33, %r17;
cvt.u64.u32 %rd28, %r3;
shr.u64 %rd29, %rd28, 1;
mul.lo.s64 %rd30, %rd29, 36;
mul.wide.u32 %rd31, %r31, 36;
add.s64 %rd32, %rd30, %rd31;
add.s64 %rd3, %rd22, %rd32;
mul.wide.u32 %rd53, %r30, 36;
and.b32 %r35, %r3, 1;
mul.wide.u32 %rd33, %r35, 8;
mad.lo.s32 %r36, %r1, %r14, %r116;
add.s32 %r37, %r36, %r34;
mul.wide.s32 %rd5, %r37, 18;
add.s64 %rd34, %rd33, %rd5;
add.s64 %rd6, %rd34, 8;
add.s64 %rd35, %rd29, %rd24;
add.s64 %rd36, %rd35, %rd23;
mul.lo.s64 %rd37, %rd36, 36;
add.s64 %rd38, %rd37, %rd33;
add.s64 %rd39, %rd22, %rd38;
add.s64 %rd51, %rd39, 20;
add.s32 %r38, %r1, 1;
add.s32 %r39, %r116, %r34;
mad.lo.s32 %r40, %r14, %r38, %r39;
mul.wide.s32 %rd8, %r40, 18;
add.s64 %rd40, %rd33, %rd8;
add.s64 %rd9, %rd40, 8;
cvta.to.global.u64 %rd52, %rd17;
mov.f32 %f48, 0f00000000;

$L__BB27_3:
add.s64 %rd41, %rd3, %rd53;
add.s64 %rd42, %rd52, %rd6;
ld.global.nc.u16 %rs3, [%rd42+-6];
ld.global.nc.u16 %rs4, [%rd42+-4];
mov.b32 %r77, {%rs3, %rs4};
ld.global.nc.u32 %r43, [%rd51+-16];
ld.global.nc.u32 %r47, [%rd51];
ld.global.nc.u16 %rs5, [%rd42+-2];
ld.global.nc.u16 %rs6, [%rd42];
mov.b32 %r78, {%rs5, %rs6};
ld.global.nc.u32 %r51, [%rd51+-12];
ld.global.nc.u32 %r55, [%rd51+4];
add.s64 %rd43, %rd52, %rd5;
ld.global.nc.u16 %rs1, [%rd43];

{ cvt.f32.f16 %f11, %rs1;}


and.b32 %r42, %r77, 252645135;
shr.u32 %r79, %r77, 4;
and.b32 %r46, %r79, 252645135;
mov.u32 %r62, 0;

dp4a.s32.s32 %r41, %r42, %r43, %r62;


dp4a.s32.s32 %r45, %r46, %r47, %r41;

and.b32 %r50, %r78, 252645135;
shr.u32 %r80, %r78, 4;
and.b32 %r54, %r80, 252645135;

dp4a.s32.s32 %r49, %r50, %r51, %r45;


dp4a.s32.s32 %r53, %r54, %r55, %r49;

ld.global.nc.u32 %r57, [%rd41];

{.reg .f16 low,high;
mov.b32 {low,high},%r57;
cvt.f32.f16 %f12, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r57;
cvt.f32.f16 %f13, high;}


cvt.rn.f32.s32 %f17, %r53;
mul.ftz.f32 %f18, %f12, %f17;
mul.ftz.f32 %f19, %f13, 0f40800000;
sub.ftz.f32 %f20, %f18, %f19;
fma.rn.ftz.f32 %f49, %f11, %f20, %f49;
add.s64 %rd44, %rd52, %rd9;
ld.global.nc.u16 %rs7, [%rd44+-6];
ld.global.nc.u16 %rs8, [%rd44+-4];
mov.b32 %r81, {%rs7, %rs8};
ld.global.nc.u16 %rs9, [%rd44+-2];
ld.global.nc.u16 %rs10, [%rd44];
mov.b32 %r82, {%rs9, %rs10};
add.s64 %rd45, %rd52, %rd8;
ld.global.nc.u16 %rs2, [%rd45];

{ cvt.f32.f16 %f14, %rs2;}


and.b32 %r60, %r81, 252645135;
shr.u32 %r83, %r81, 4;
and.b32 %r64, %r83, 252645135;

dp4a.s32.s32 %r59, %r60, %r43, %r62;


dp4a.s32.s32 %r63, %r64, %r47, %r59;

and.b32 %r68, %r82, 252645135;
shr.u32 %r84, %r82, 4;
and.b32 %r72, %r84, 252645135;

dp4a.s32.s32 %r67, %r68, %r51, %r63;


dp4a.s32.s32 %r71, %r72, %r55, %r67;


{.reg .f16 low,high;
mov.b32 {low,high},%r57;
cvt.f32.f16 %f15, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r57;
cvt.f32.f16 %f16, high;}


cvt.rn.f32.s32 %f21, %r71;
mul.ftz.f32 %f22, %f15, %f21;
mul.ftz.f32 %f23, %f16, 0f40800000;
sub.ftz.f32 %f24, %f22, %f23;
fma.rn.ftz.f32 %f48, %f14, %f24, %f48;
add.s64 %rd53, %rd53, 576;
add.s64 %rd52, %rd52, 288;
add.s64 %rd51, %rd51, 576;
add.s32 %r116, %r116, 16;
setp.lt.u32 %p3, %r116, %r23;
@%p3 bra $L__BB27_3;

st.local.v2.f32 [%rd1], {%f49, %f48};

$L__BB27_5:
mov.b32 %r85, %f49;
mov.u32 %r86, 31;
mov.u32 %r87, 16;
mov.u32 %r88, -1;
shfl.sync.bfly.b32 %r89|%p4, %r85, %r87, %r86, %r88;
mov.b32 %f25, %r89;
add.ftz.f32 %f26, %f49, %f25;
mov.b32 %r90, %f26;
mov.u32 %r91, 8;
shfl.sync.bfly.b32 %r92|%p5, %r90, %r91, %r86, %r88;
mov.b32 %f27, %r92;
add.ftz.f32 %f28, %f26, %f27;
mov.b32 %r93, %f28;
mov.u32 %r94, 4;
shfl.sync.bfly.b32 %r95|%p6, %r93, %r94, %r86, %r88;
mov.b32 %f29, %r95;
add.ftz.f32 %f30, %f28, %f29;
mov.b32 %r96, %f30;
mov.u32 %r97, 2;
shfl.sync.bfly.b32 %r98|%p7, %r96, %r97, %r86, %r88;
mov.b32 %f31, %r98;
add.ftz.f32 %f32, %f30, %f31;
mov.b32 %r99, %f32;
mov.u32 %r100, 1;
shfl.sync.bfly.b32 %r101|%p8, %r99, %r100, %r86, %r88;
mov.b32 %f33, %r101;
add.ftz.f32 %f34, %f32, %f33;
st.local.f32 [%rd1], %f34;
mov.b32 %r102, %f48;
shfl.sync.bfly.b32 %r103|%p9, %r102, %r87, %r86, %r88;
mov.b32 %f35, %r103;
add.ftz.f32 %f36, %f48, %f35;
mov.b32 %r104, %f36;
shfl.sync.bfly.b32 %r105|%p10, %r104, %r91, %r86, %r88;
mov.b32 %f37, %r105;
add.ftz.f32 %f38, %f36, %f37;
mov.b32 %r106, %f38;
shfl.sync.bfly.b32 %r107|%p11, %r106, %r94, %r86, %r88;
mov.b32 %f39, %r107;
add.ftz.f32 %f40, %f38, %f39;
mov.b32 %r108, %f40;
shfl.sync.bfly.b32 %r109|%p12, %r108, %r97, %r86, %r88;
mov.b32 %f41, %r109;
add.ftz.f32 %f42, %f40, %f41;
mov.b32 %r110, %f42;
shfl.sync.bfly.b32 %r111|%p13, %r110, %r100, %r86, %r88;
mov.b32 %f43, %r111;
add.ftz.f32 %f44, %f42, %f43;
st.local.f32 [%rd1+4], %f44;
add.s32 %r8, %r1, %r3;
setp.ge.u32 %p14, %r8, %r13;
setp.gt.u32 %p15, %r3, 1;
or.pred %p16, %p15, %p14;
@%p16 bra $L__BB27_7;

mul.wide.u32 %rd46, %r3, 4;
add.s64 %rd47, %rd1, %rd46;
ld.local.f32 %f45, [%rd47];
mov.u32 %r112, %ctaid.y;
mad.lo.s32 %r114, %r2, %r16, %r8;
mad.lo.s32 %r115, %r112, %r19, %r114;
cvta.to.global.u64 %rd48, %rd20;
mul.wide.u32 %rd49, %r115, 4;
add.s64 %rd50, %rd48, %rd49;
st.global.f32 [%rd50], %f45;

$L__BB27_7:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<320>;
.reg .b16 %rs<41>;
.reg .f32 %f<1021>;
.reg .b32 %r<1032>;
.reg .b64 %rd<100>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate[1536];

ld.param.u64 %rd25, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd26, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd27, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r48, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r49, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r50, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r54, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r55, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r56, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r52, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r51, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r46, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r45, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
ld.param.u64 %rd30, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+16];
ld.param.u64 %rd29, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+8];
ld.param.u64 %rd28, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3];
mov.u32 %r1, %tid.y;
shl.b32 %r57, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r57, %r2;
mov.u32 %r58, %ctaid.x;
shl.b32 %r4, %r58, 2;
mov.u32 %r5, %ctaid.y;
setp.eq.s64 %p1, %rd27, 0;
@%p1 bra $L__BB28_2;

cvta.to.global.u64 %rd34, %rd27;
mul.wide.u32 %rd35, %r5, 4;
add.s64 %rd36, %rd34, %rd35;
ld.global.nc.u32 %r1029, [%rd36];
mul.hi.u32 %r59, %r5, %r40;
add.s32 %r60, %r59, %r5;
shr.u32 %r61, %r60, %r41;
mul.lo.s32 %r62, %r61, %r42;
sub.s32 %r1030, %r5, %r62;
bra.uni $L__BB28_3;

$L__BB28_2:
mul.hi.u32 %r63, %r5, %r45;
add.s32 %r64, %r63, %r5;
shr.u32 %r1029, %r64, %r46;
mov.u32 %r1030, %r5;

$L__BB28_3:
mov.u32 %r11, %ctaid.z;
selp.b32 %r12, %r5, %r1029, %p1;
setp.ne.s32 %p3, %r1, 0;
setp.gt.u32 %p4, %r2, 3;
or.pred %p5, %p4, %p3;
add.s32 %r65, %r4, %r2;
setp.ge.u32 %p6, %r65, %r44;
mov.f32 %f27, 0f00000000;
or.pred %p7, %p5, %p6;
mov.f32 %f4, %f27;
mov.f32 %f5, %f27;
@%p7 bra $L__BB28_8;

setp.eq.s64 %p8, %rd28, 0;
mov.f32 %f4, 0f00000000;
mov.f32 %f5, %f4;
@%p8 bra $L__BB28_6;

cvta.to.global.u64 %rd37, %rd28;
mul.lo.s32 %r66, %r11, %r56;
cvt.u64.u32 %rd38, %r66;
mul.lo.s32 %r67, %r12, %r50;
cvt.u64.u32 %rd39, %r67;
cvt.u64.u32 %rd40, %r2;
cvt.s64.s32 %rd41, %r4;
add.s64 %rd42, %rd41, %rd40;
add.s64 %rd43, %rd42, %rd38;
add.s64 %rd44, %rd43, %rd39;
shl.b64 %rd45, %rd44, 2;
add.s64 %rd46, %rd37, %rd45;
ld.global.f32 %f5, [%rd46];

$L__BB28_6:
setp.eq.s64 %p9, %rd30, 0;
setp.eq.s64 %p10, %rd29, 0;
or.pred %p11, %p9, %p10;
@%p11 bra $L__BB28_8;

cvta.to.global.u64 %rd47, %rd30;
mul.lo.s32 %r68, %r11, %r56;
cvt.u64.u32 %rd48, %r68;
mul.lo.s32 %r69, %r12, %r50;
cvt.u64.u32 %rd49, %r69;
cvt.u64.u32 %rd50, %r2;
cvt.s64.s32 %rd51, %r4;
add.s64 %rd52, %rd51, %rd50;
add.s64 %rd53, %rd52, %rd48;
add.s64 %rd54, %rd53, %rd49;
shl.b64 %rd55, %rd54, 2;
add.s64 %rd56, %rd47, %rd55;
ld.global.f32 %f4, [%rd56];

$L__BB28_8:
mul.lo.s32 %r70, %r11, %r55;
cvt.u64.u32 %rd57, %r70;
mul.lo.s32 %r71, %r1030, %r49;
cvt.u64.u32 %rd58, %r71;
add.s64 %rd4, %rd58, %rd57;
mul.hi.u32 %r72, %r11, %r51;
add.s32 %r73, %r72, %r11;
shr.u32 %r13, %r73, %r52;
shr.u32 %r74, %r3, 31;
add.s32 %r75, %r3, %r74;
shr.s32 %r1031, %r75, 1;
shr.u32 %r76, %r39, 5;
setp.ge.s32 %p12, %r1031, %r76;
mov.f32 %f28, %f27;
mov.f32 %f29, %f27;
mov.f32 %f30, %f27;
mov.f32 %f31, %f27;
mov.f32 %f32, %f27;
mov.f32 %f33, %f27;
mov.f32 %f34, %f27;
@%p12 bra $L__BB28_19;

mul.lo.s32 %r77, %r13, %r54;
mul.lo.s32 %r78, %r1029, %r48;
and.b32 %r81, %r75, -2;
sub.s32 %r82, %r3, %r81;
shl.b32 %r83, %r82, 1;
shl.b32 %r84, %r82, 2;
or.b32 %r85, %r84, 1;
cvt.s64.s32 %rd5, %r83;
sub.s32 %r86, %r85, %r83;
cvt.s64.s32 %rd6, %r86;
add.s32 %r87, %r1031, %r77;
add.s32 %r15, %r87, %r78;
add.s32 %r88, %r4, 1;
mad.lo.s32 %r89, %r43, %r88, %r15;
mul.wide.s32 %rd59, %r89, 18;
mad.lo.s32 %r90, %r4, %r43, %r1031;
add.s32 %r91, %r90, %r77;
add.s32 %r92, %r91, %r78;
mul.wide.s32 %rd7, %r92, 18;
mul.wide.s32 %rd60, %r84, 2;
add.s64 %rd61, %rd7, %rd60;
add.s64 %rd8, %rd61, 4;
add.s64 %rd62, %rd59, %rd60;
add.s64 %rd9, %rd62, 2;
add.s32 %r93, %r4, 2;
mad.lo.s32 %r94, %r43, %r93, %r15;
mul.wide.s32 %rd63, %r94, 18;
add.s64 %rd64, %rd63, %rd60;
add.s64 %rd10, %rd64, 2;
add.s32 %r95, %r4, 3;
mad.lo.s32 %r96, %r43, %r95, %r15;
mul.wide.s32 %rd65, %r96, 18;
add.s64 %rd11, %rd65, %rd60;
cvta.to.global.u64 %rd99, %rd29;
cvta.to.global.u64 %rd98, %rd25;
mov.f32 %f27, 0f00000000;
cvta.to.global.u64 %rd66, %rd26;
shl.b64 %rd73, %rd5, 2;
shl.b64 %rd75, %rd6, 2;
mov.f32 %f28, %f27;
mov.f32 %f29, %f27;
mov.f32 %f30, %f27;
mov.f32 %f31, %f27;
mov.f32 %f32, %f27;
mov.f32 %f33, %f27;
mov.f32 %f34, %f27;

$L__BB28_10:
cvt.u64.u32 %rd67, %r1031;
add.s64 %rd68, %rd4, %rd67;
mul.lo.s64 %rd69, %rd68, 36;
add.s64 %rd70, %rd66, %rd69;
add.s64 %rd71, %rd70, 4;
add.s64 %rd72, %rd98, %rd8;
add.s64 %rd16, %rd72, -2;
ld.global.nc.u16 %rs2, [%rd72+-2];
ld.global.nc.u16 %rs3, [%rd72];
mov.b32 %r115, {%rs2, %rs3};
add.s64 %rd74, %rd71, %rd73;
ld.global.nc.u32 %r17, [%rd74];
ld.global.nc.u32 %r18, [%rd74+16];
ld.global.nc.u16 %rs4, [%rd72+2];
ld.global.nc.u16 %rs5, [%rd72+4];
mov.b32 %r116, {%rs4, %rs5};
add.s64 %rd76, %rd71, %rd75;
ld.global.nc.u32 %r19, [%rd76];
ld.global.nc.u32 %r20, [%rd74+20];
add.s64 %rd17, %rd98, %rd7;
ld.global.nc.u16 %rs1, [%rd17];

{ cvt.f32.f16 %f101, %rs1;}


and.b32 %r98, %r115, 252645135;
shr.u32 %r117, %r115, 4;
and.b32 %r102, %r117, 252645135;
mov.u32 %r100, 0;

dp4a.s32.s32 %r97, %r98, %r17, %r100;


dp4a.s32.s32 %r101, %r102, %r18, %r97;

and.b32 %r106, %r116, 252645135;
shr.u32 %r118, %r116, 4;
and.b32 %r110, %r118, 252645135;

dp4a.s32.s32 %r105, %r106, %r19, %r101;


dp4a.s32.s32 %r109, %r110, %r20, %r105;

ld.global.nc.u32 %r21, [%rd70];

{.reg .f16 low,high;
mov.b32 {low,high},%r21;
cvt.f32.f16 %f102, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r21;
cvt.f32.f16 %f103, high;}


cvt.rn.f32.s32 %f104, %r109;
mul.ftz.f32 %f105, %f103, 0fC0800000;
fma.rn.ftz.f32 %f106, %f102, %f104, %f105;
fma.rn.ftz.f32 %f30, %f101, %f106, %f30;
setp.eq.s64 %p13, %rd29, 0;
@%p13 bra $L__BB28_12;

add.s64 %rd77, %rd99, %rd8;
ld.global.u16 %rs7, [%rd77+-2];
ld.global.u16 %rs8, [%rd77];
mov.b32 %r137, {%rs7, %rs8};
ld.global.u16 %rs9, [%rd77+2];
ld.global.u16 %rs10, [%rd77+4];
mov.b32 %r138, {%rs9, %rs10};
add.s64 %rd78, %rd99, %rd7;
ld.global.u16 %rs6, [%rd78];

{ cvt.f32.f16 %f107, %rs6;}


and.b32 %r120, %r137, 252645135;
shr.u32 %r139, %r137, 4;
and.b32 %r124, %r139, 252645135;

dp4a.s32.s32 %r119, %r120, %r17, %r100;


dp4a.s32.s32 %r123, %r124, %r18, %r119;

and.b32 %r128, %r138, 252645135;
shr.u32 %r140, %r138, 4;
and.b32 %r132, %r140, 252645135;

dp4a.s32.s32 %r127, %r128, %r19, %r123;


dp4a.s32.s32 %r131, %r132, %r20, %r127;


{.reg .f16 low,high;
mov.b32 {low,high},%r21;
cvt.f32.f16 %f108, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r21;
cvt.f32.f16 %f109, high;}


cvt.rn.f32.s32 %f110, %r131;
mul.ftz.f32 %f111, %f109, 0fC0800000;
fma.rn.ftz.f32 %f112, %f108, %f110, %f111;
fma.rn.ftz.f32 %f34, %f107, %f112, %f34;

$L__BB28_12:
mul.wide.s32 %rd79, %r43, 18;
add.s64 %rd18, %rd16, %rd79;
ld.global.nc.u16 %rs12, [%rd18];
ld.global.nc.u16 %rs13, [%rd18+2];
mov.b32 %r159, {%rs12, %rs13};
ld.global.nc.u16 %rs14, [%rd18+4];
ld.global.nc.u16 %rs15, [%rd18+6];
mov.b32 %r160, {%rs14, %rs15};
add.s64 %rd19, %rd17, %rd79;
ld.global.nc.u16 %rs11, [%rd19];

{ cvt.f32.f16 %f113, %rs11;}


and.b32 %r142, %r159, 252645135;
shr.u32 %r161, %r159, 4;
and.b32 %r146, %r161, 252645135;

dp4a.s32.s32 %r141, %r142, %r17, %r100;


dp4a.s32.s32 %r145, %r146, %r18, %r141;

and.b32 %r150, %r160, 252645135;
shr.u32 %r162, %r160, 4;
and.b32 %r154, %r162, 252645135;

dp4a.s32.s32 %r149, %r150, %r19, %r145;


dp4a.s32.s32 %r153, %r154, %r20, %r149;


{.reg .f16 low,high;
mov.b32 {low,high},%r21;
cvt.f32.f16 %f114, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r21;
cvt.f32.f16 %f115, high;}


cvt.rn.f32.s32 %f116, %r153;
mul.ftz.f32 %f117, %f115, 0fC0800000;
fma.rn.ftz.f32 %f118, %f114, %f116, %f117;
fma.rn.ftz.f32 %f29, %f113, %f118, %f29;
@%p13 bra $L__BB28_14;

add.s64 %rd80, %rd99, %rd9;
ld.global.u16 %rs17, [%rd80];
ld.global.u16 %rs18, [%rd80+2];
mov.b32 %r181, {%rs17, %rs18};
ld.global.u16 %rs19, [%rd80+4];
ld.global.u16 %rs20, [%rd80+6];
mov.b32 %r182, {%rs19, %rs20};
or.b32 %r185, %r4, 1;
mad.lo.s32 %r186, %r43, %r185, %r15;
mul.wide.s32 %rd81, %r186, 18;
add.s64 %rd82, %rd99, %rd81;
ld.global.u16 %rs16, [%rd82];

{ cvt.f32.f16 %f119, %rs16;}


and.b32 %r164, %r181, 252645135;
shr.u32 %r187, %r181, 4;
and.b32 %r168, %r187, 252645135;
mov.u32 %r166, 0;

dp4a.s32.s32 %r163, %r164, %r17, %r166;


dp4a.s32.s32 %r167, %r168, %r18, %r163;

and.b32 %r172, %r182, 252645135;
shr.u32 %r188, %r182, 4;
and.b32 %r176, %r188, 252645135;

dp4a.s32.s32 %r171, %r172, %r19, %r167;


dp4a.s32.s32 %r175, %r176, %r20, %r171;


{.reg .f16 low,high;
mov.b32 {low,high},%r21;
cvt.f32.f16 %f120, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r21;
cvt.f32.f16 %f121, high;}


cvt.rn.f32.s32 %f122, %r175;
mul.ftz.f32 %f123, %f121, 0fC0800000;
fma.rn.ftz.f32 %f124, %f120, %f122, %f123;
fma.rn.ftz.f32 %f33, %f119, %f124, %f33;

$L__BB28_14:
mul.wide.s32 %rd97, %r43, 18;
add.s64 %rd20, %rd18, %rd97;
ld.global.nc.u16 %rs22, [%rd20];
ld.global.nc.u16 %rs23, [%rd20+2];
mov.b32 %r207, {%rs22, %rs23};
ld.global.nc.u16 %rs24, [%rd20+4];
ld.global.nc.u16 %rs25, [%rd20+6];
mov.b32 %r208, {%rs24, %rs25};
add.s64 %rd21, %rd19, %rd97;
ld.global.nc.u16 %rs21, [%rd21];

{ cvt.f32.f16 %f125, %rs21;}


and.b32 %r190, %r207, 252645135;
shr.u32 %r209, %r207, 4;
and.b32 %r194, %r209, 252645135;
mov.u32 %r192, 0;

dp4a.s32.s32 %r189, %r190, %r17, %r192;


dp4a.s32.s32 %r193, %r194, %r18, %r189;

and.b32 %r198, %r208, 252645135;
shr.u32 %r210, %r208, 4;
and.b32 %r202, %r210, 252645135;

dp4a.s32.s32 %r197, %r198, %r19, %r193;


dp4a.s32.s32 %r201, %r202, %r20, %r197;


{.reg .f16 low,high;
mov.b32 {low,high},%r21;
cvt.f32.f16 %f126, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r21;
cvt.f32.f16 %f127, high;}


cvt.rn.f32.s32 %f128, %r201;
mul.ftz.f32 %f129, %f127, 0fC0800000;
fma.rn.ftz.f32 %f130, %f126, %f128, %f129;
fma.rn.ftz.f32 %f28, %f125, %f130, %f28;
@%p13 bra $L__BB28_16;

add.s64 %rd84, %rd99, %rd10;
ld.global.u16 %rs27, [%rd84];
ld.global.u16 %rs28, [%rd84+2];
mov.b32 %r229, {%rs27, %rs28};
ld.global.u16 %rs29, [%rd84+4];
ld.global.u16 %rs30, [%rd84+6];
mov.b32 %r230, {%rs29, %rs30};
or.b32 %r233, %r4, 2;
mad.lo.s32 %r234, %r43, %r233, %r15;
mul.wide.s32 %rd85, %r234, 18;
add.s64 %rd86, %rd99, %rd85;
ld.global.u16 %rs26, [%rd86];

{ cvt.f32.f16 %f131, %rs26;}


and.b32 %r212, %r229, 252645135;
shr.u32 %r235, %r229, 4;
and.b32 %r216, %r235, 252645135;

dp4a.s32.s32 %r211, %r212, %r17, %r192;


dp4a.s32.s32 %r215, %r216, %r18, %r211;

and.b32 %r220, %r230, 252645135;
shr.u32 %r236, %r230, 4;
and.b32 %r224, %r236, 252645135;

dp4a.s32.s32 %r219, %r220, %r19, %r215;


dp4a.s32.s32 %r223, %r224, %r20, %r219;


{.reg .f16 low,high;
mov.b32 {low,high},%r21;
cvt.f32.f16 %f132, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r21;
cvt.f32.f16 %f133, high;}


cvt.rn.f32.s32 %f134, %r223;
mul.ftz.f32 %f135, %f133, 0fC0800000;
fma.rn.ftz.f32 %f136, %f132, %f134, %f135;
fma.rn.ftz.f32 %f32, %f131, %f136, %f32;

$L__BB28_16:
mul.wide.s32 %rd95, %r43, 18;
add.s64 %rd88, %rd20, %rd95;
ld.global.nc.u16 %rs32, [%rd88];
ld.global.nc.u16 %rs33, [%rd88+2];
mov.b32 %r255, {%rs32, %rs33};
ld.global.nc.u16 %rs34, [%rd88+4];
ld.global.nc.u16 %rs35, [%rd88+6];
mov.b32 %r256, {%rs34, %rs35};
add.s64 %rd89, %rd21, %rd95;
ld.global.nc.u16 %rs31, [%rd89];

{ cvt.f32.f16 %f137, %rs31;}


and.b32 %r238, %r255, 252645135;
shr.u32 %r257, %r255, 4;
and.b32 %r242, %r257, 252645135;

dp4a.s32.s32 %r237, %r238, %r17, %r192;


dp4a.s32.s32 %r241, %r242, %r18, %r237;

and.b32 %r246, %r256, 252645135;
shr.u32 %r258, %r256, 4;
and.b32 %r250, %r258, 252645135;

dp4a.s32.s32 %r245, %r246, %r19, %r241;


dp4a.s32.s32 %r249, %r250, %r20, %r245;


{.reg .f16 low,high;
mov.b32 {low,high},%r21;
cvt.f32.f16 %f138, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r21;
cvt.f32.f16 %f139, high;}


cvt.rn.f32.s32 %f140, %r249;
mul.ftz.f32 %f141, %f139, 0fC0800000;
fma.rn.ftz.f32 %f142, %f138, %f140, %f141;
fma.rn.ftz.f32 %f27, %f137, %f142, %f27;
@%p13 bra $L__BB28_18;

add.s64 %rd90, %rd99, %rd11;
ld.global.u16 %rs37, [%rd90+2];
ld.global.u16 %rs38, [%rd90+4];
mov.b32 %r277, {%rs37, %rs38};
ld.global.u16 %rs39, [%rd90+6];
ld.global.u16 %rs40, [%rd90+8];
mov.b32 %r278, {%rs39, %rs40};
or.b32 %r281, %r4, 3;
mad.lo.s32 %r282, %r43, %r281, %r15;
mul.wide.s32 %rd91, %r282, 18;
add.s64 %rd92, %rd99, %rd91;
ld.global.u16 %rs36, [%rd92];

{ cvt.f32.f16 %f143, %rs36;}


and.b32 %r260, %r277, 252645135;
shr.u32 %r283, %r277, 4;
and.b32 %r264, %r283, 252645135;
mov.u32 %r262, 0;

dp4a.s32.s32 %r259, %r260, %r17, %r262;


dp4a.s32.s32 %r263, %r264, %r18, %r259;

and.b32 %r268, %r278, 252645135;
shr.u32 %r284, %r278, 4;
and.b32 %r272, %r284, 252645135;

dp4a.s32.s32 %r267, %r268, %r19, %r263;


dp4a.s32.s32 %r271, %r272, %r20, %r267;


{.reg .f16 low,high;
mov.b32 {low,high},%r21;
cvt.f32.f16 %f144, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r21;
cvt.f32.f16 %f145, high;}


cvt.rn.f32.s32 %f146, %r271;
mul.ftz.f32 %f147, %f145, 0fC0800000;
fma.rn.ftz.f32 %f148, %f144, %f146, %f147;
fma.rn.ftz.f32 %f31, %f143, %f148, %f31;

$L__BB28_18:
add.s64 %rd99, %rd99, 1152;
add.s64 %rd98, %rd98, 1152;
add.s32 %r1031, %r1031, 64;
setp.lt.s32 %p17, %r1031, %r76;
@%p17 bra $L__BB28_10;

$L__BB28_19:
mov.u32 %r1018, %tid.y;
setp.eq.s32 %p18, %r1018, 0;
@%p18 bra $L__BB28_23;

mov.u32 %r1019, %tid.y;
setp.ne.s64 %p19, %rd29, 0;
add.s32 %r23, %r1019, -1;
shl.b32 %r289, %r1019, 9;
mov.u32 %r290, _ZZ13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r291, %r290, %r289;
shl.b32 %r292, %r2, 2;
add.s32 %r293, %r291, %r292;
st.shared.f32 [%r293+-512], %f30;
@%p19 bra $L__BB28_22;
bra.uni $L__BB28_21;

$L__BB28_22:
shl.b32 %r295, %r23, 9;
mov.u32 %r296, _ZZ13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r297, %r296, %r295;
add.s32 %r299, %r297, %r292;
st.shared.f32 [%r299], %f34;
add.s32 %r1015, %r293, -512;
st.shared.f32 [%r1015+128], %f29;
st.shared.f32 [%r299+128], %f33;
add.s32 %r1016, %r293, -512;
st.shared.f32 [%r1016+256], %f28;
st.shared.f32 [%r299+256], %f32;
add.s32 %r1017, %r293, -512;
st.shared.f32 [%r1017+384], %f27;
st.shared.f32 [%r299+384], %f31;
bra.uni $L__BB28_23;

$L__BB28_21:
add.s32 %r1012, %r293, -512;
st.shared.f32 [%r1012+128], %f29;
add.s32 %r1013, %r293, -512;
st.shared.f32 [%r1013+256], %f28;
add.s32 %r1014, %r293, -512;
st.shared.f32 [%r1014+384], %f27;

$L__BB28_23:
mov.u32 %r1020, %tid.y;
setp.ne.s32 %p319, %r1020, 0;
bar.sync 0;
@%p319 bra $L__BB28_78;
bra.uni $L__BB28_24;

$L__BB28_78:
ret;

$L__BB28_24:
ld.param.u64 %rd96, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.f32 %f993, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
ld.param.u32 %r1023, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r1022, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
mov.u32 %r1021, %ctaid.z;
setp.ne.s64 %p21, %rd29, 0;
mov.u32 %r302, %ctaid.y;
mad.lo.s32 %r305, %r302, %r1023, %r4;
mad.lo.s32 %r306, %r1021, %r1022, %r305;
neg.ftz.f32 %f35, %f993;
shl.b32 %r308, %r2, 2;
mov.u32 %r309, _ZZ13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r25, %r309, %r308;
cvta.to.global.u64 %rd93, %rd96;
mul.wide.u32 %rd94, %r306, 4;
add.s64 %rd24, %rd93, %rd94;
@%p21 bra $L__BB28_33;
bra.uni $L__BB28_25;

$L__BB28_33:
ld.param.u32 %r1024, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
mov.u32 %r396, _ZZ13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r27, %r396, %r308;
setp.gt.s32 %p54, %r1024, 2;
@%p54 bra $L__BB28_44;

ld.param.u32 %r1027, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p57, %r1027, 1;
@%p57 bra $L__BB28_62;

ld.param.u32 %r1028, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p58, %r1028, 2;
@%p58 bra $L__BB28_36;
bra.uni $L__BB28_70;

$L__BB28_36:
ld.shared.f32 %f549, [%r25];
add.ftz.f32 %f550, %f549, %f30;
ld.shared.f32 %f551, [%r27];
add.ftz.f32 %f552, %f551, %f34;
ld.shared.f32 %f553, [%r25+512];
add.ftz.f32 %f554, %f553, %f550;
ld.shared.f32 %f555, [%r27+512];
add.ftz.f32 %f556, %f555, %f552;
ld.shared.f32 %f557, [%r25+1024];
add.ftz.f32 %f558, %f557, %f554;
ld.shared.f32 %f559, [%r27+1024];
add.ftz.f32 %f560, %f559, %f556;
mov.b32 %r643, %f558;
mov.u32 %r644, 31;
mov.u32 %r645, 16;
mov.u32 %r646, -1;
shfl.sync.bfly.b32 %r647|%p163, %r643, %r645, %r644, %r646;
mov.b32 %f561, %r647;
add.ftz.f32 %f562, %f558, %f561;
mov.b32 %r648, %f562;
mov.u32 %r649, 8;
shfl.sync.bfly.b32 %r650|%p164, %r648, %r649, %r644, %r646;
mov.b32 %f563, %r650;
add.ftz.f32 %f564, %f562, %f563;
mov.b32 %r651, %f564;
mov.u32 %r652, 4;
shfl.sync.bfly.b32 %r653|%p165, %r651, %r652, %r644, %r646;
mov.b32 %f565, %r653;
add.ftz.f32 %f566, %f564, %f565;
mov.b32 %r654, %f566;
mov.u32 %r655, 2;
shfl.sync.bfly.b32 %r656|%p166, %r654, %r655, %r644, %r646;
mov.b32 %f567, %r656;
add.ftz.f32 %f568, %f566, %f567;
mov.b32 %r657, %f568;
mov.u32 %r658, 1;
shfl.sync.bfly.b32 %r659|%p167, %r657, %r658, %r644, %r646;
mov.b32 %f569, %r659;
add.ftz.f32 %f56, %f568, %f569;
mov.b32 %r660, %f560;
shfl.sync.bfly.b32 %r661|%p168, %r660, %r645, %r644, %r646;
mov.b32 %f570, %r661;
add.ftz.f32 %f571, %f560, %f570;
mov.b32 %r662, %f571;
shfl.sync.bfly.b32 %r663|%p169, %r662, %r649, %r644, %r646;
mov.b32 %f572, %r663;
add.ftz.f32 %f573, %f571, %f572;
mov.b32 %r664, %f573;
shfl.sync.bfly.b32 %r665|%p170, %r664, %r652, %r644, %r646;
mov.b32 %f574, %r665;
add.ftz.f32 %f575, %f573, %f574;
mov.b32 %r666, %f575;
shfl.sync.bfly.b32 %r667|%p171, %r666, %r655, %r644, %r646;
mov.b32 %f576, %r667;
add.ftz.f32 %f577, %f575, %f576;
mov.b32 %r668, %f577;
shfl.sync.bfly.b32 %r669|%p172, %r668, %r658, %r644, %r646;
mov.b32 %f578, %r669;
add.ftz.f32 %f57, %f577, %f578;
setp.ne.s32 %p173, %r2, 0;
setp.ge.u32 %p174, %r4, %r44;
or.pred %p175, %p173, %p174;
@%p175 bra $L__BB28_38;

add.ftz.f32 %f579, %f5, %f56;
add.ftz.f32 %f580, %f4, %f57;
mul.ftz.f32 %f581, %f580, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f582, %f581;
add.ftz.f32 %f583, %f582, 0f3F800000;
div.approx.ftz.f32 %f584, %f580, %f583;
mul.ftz.f32 %f585, %f579, %f584;
st.global.f32 [%rd24], %f585;

$L__BB28_38:
ld.shared.f32 %f586, [%r25+128];
add.ftz.f32 %f587, %f586, %f29;
ld.shared.f32 %f588, [%r27+128];
add.ftz.f32 %f589, %f588, %f33;
ld.shared.f32 %f590, [%r25+640];
add.ftz.f32 %f591, %f590, %f587;
ld.shared.f32 %f592, [%r27+640];
add.ftz.f32 %f593, %f592, %f589;
ld.shared.f32 %f594, [%r25+1152];
add.ftz.f32 %f595, %f594, %f591;
ld.shared.f32 %f596, [%r27+1152];
add.ftz.f32 %f597, %f596, %f593;
mov.b32 %r673, %f595;
shfl.sync.bfly.b32 %r677|%p176, %r673, %r645, %r644, %r646;
mov.b32 %f598, %r677;
add.ftz.f32 %f599, %f595, %f598;
mov.b32 %r678, %f599;
shfl.sync.bfly.b32 %r680|%p177, %r678, %r649, %r644, %r646;
mov.b32 %f600, %r680;
add.ftz.f32 %f601, %f599, %f600;
mov.b32 %r681, %f601;
shfl.sync.bfly.b32 %r683|%p178, %r681, %r652, %r644, %r646;
mov.b32 %f602, %r683;
add.ftz.f32 %f603, %f601, %f602;
mov.b32 %r684, %f603;
shfl.sync.bfly.b32 %r686|%p179, %r684, %r655, %r644, %r646;
mov.b32 %f604, %r686;
add.ftz.f32 %f605, %f603, %f604;
mov.b32 %r687, %f605;
shfl.sync.bfly.b32 %r689|%p180, %r687, %r658, %r644, %r646;
mov.b32 %f606, %r689;
add.ftz.f32 %f58, %f605, %f606;
mov.b32 %r690, %f597;
shfl.sync.bfly.b32 %r691|%p181, %r690, %r645, %r644, %r646;
mov.b32 %f607, %r691;
add.ftz.f32 %f608, %f597, %f607;
mov.b32 %r692, %f608;
shfl.sync.bfly.b32 %r693|%p182, %r692, %r649, %r644, %r646;
mov.b32 %f609, %r693;
add.ftz.f32 %f610, %f608, %f609;
mov.b32 %r694, %f610;
shfl.sync.bfly.b32 %r695|%p183, %r694, %r652, %r644, %r646;
mov.b32 %f611, %r695;
add.ftz.f32 %f612, %f610, %f611;
mov.b32 %r696, %f612;
shfl.sync.bfly.b32 %r697|%p184, %r696, %r655, %r644, %r646;
mov.b32 %f613, %r697;
add.ftz.f32 %f614, %f612, %f613;
mov.b32 %r698, %f614;
shfl.sync.bfly.b32 %r699|%p185, %r698, %r658, %r644, %r646;
mov.b32 %f615, %r699;
add.ftz.f32 %f59, %f614, %f615;
setp.ne.s32 %p186, %r2, 1;
or.b32 %r703, %r4, 1;
setp.ge.u32 %p187, %r703, %r44;
or.pred %p188, %p186, %p187;
@%p188 bra $L__BB28_40;

add.ftz.f32 %f616, %f5, %f58;
add.ftz.f32 %f617, %f4, %f59;
mul.ftz.f32 %f618, %f617, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f619, %f618;
add.ftz.f32 %f620, %f619, 0f3F800000;
div.approx.ftz.f32 %f621, %f617, %f620;
mul.ftz.f32 %f622, %f616, %f621;
st.global.f32 [%rd24+4], %f622;

$L__BB28_40:
ld.shared.f32 %f623, [%r25+256];
add.ftz.f32 %f624, %f623, %f28;
ld.shared.f32 %f625, [%r27+256];
add.ftz.f32 %f626, %f625, %f32;
ld.shared.f32 %f627, [%r25+768];
add.ftz.f32 %f628, %f627, %f624;
ld.shared.f32 %f629, [%r27+768];
add.ftz.f32 %f630, %f629, %f626;
ld.shared.f32 %f631, [%r25+1280];
add.ftz.f32 %f632, %f631, %f628;
ld.shared.f32 %f633, [%r27+1280];
add.ftz.f32 %f634, %f633, %f630;
mov.b32 %r704, %f632;
mov.u32 %r705, 31;
mov.u32 %r706, 16;
mov.u32 %r707, -1;
shfl.sync.bfly.b32 %r708|%p189, %r704, %r706, %r705, %r707;
mov.b32 %f635, %r708;
add.ftz.f32 %f636, %f632, %f635;
mov.b32 %r709, %f636;
mov.u32 %r710, 8;
shfl.sync.bfly.b32 %r711|%p190, %r709, %r710, %r705, %r707;
mov.b32 %f637, %r711;
add.ftz.f32 %f638, %f636, %f637;
mov.b32 %r712, %f638;
mov.u32 %r713, 4;
shfl.sync.bfly.b32 %r714|%p191, %r712, %r713, %r705, %r707;
mov.b32 %f639, %r714;
add.ftz.f32 %f640, %f638, %f639;
mov.b32 %r715, %f640;
mov.u32 %r716, 2;
shfl.sync.bfly.b32 %r717|%p192, %r715, %r716, %r705, %r707;
mov.b32 %f641, %r717;
add.ftz.f32 %f642, %f640, %f641;
mov.b32 %r718, %f642;
mov.u32 %r719, 1;
shfl.sync.bfly.b32 %r720|%p193, %r718, %r719, %r705, %r707;
mov.b32 %f643, %r720;
add.ftz.f32 %f60, %f642, %f643;
mov.b32 %r721, %f634;
shfl.sync.bfly.b32 %r722|%p194, %r721, %r706, %r705, %r707;
mov.b32 %f644, %r722;
add.ftz.f32 %f645, %f634, %f644;
mov.b32 %r723, %f645;
shfl.sync.bfly.b32 %r724|%p195, %r723, %r710, %r705, %r707;
mov.b32 %f646, %r724;
add.ftz.f32 %f647, %f645, %f646;
mov.b32 %r725, %f647;
shfl.sync.bfly.b32 %r726|%p196, %r725, %r713, %r705, %r707;
mov.b32 %f648, %r726;
add.ftz.f32 %f649, %f647, %f648;
mov.b32 %r727, %f649;
shfl.sync.bfly.b32 %r728|%p197, %r727, %r716, %r705, %r707;
mov.b32 %f650, %r728;
add.ftz.f32 %f651, %f649, %f650;
mov.b32 %r729, %f651;
shfl.sync.bfly.b32 %r730|%p198, %r729, %r719, %r705, %r707;
mov.b32 %f652, %r730;
add.ftz.f32 %f61, %f651, %f652;
setp.ne.s32 %p199, %r2, 2;
or.b32 %r734, %r4, 2;
setp.ge.u32 %p200, %r734, %r44;
or.pred %p201, %p199, %p200;
@%p201 bra $L__BB28_42;

add.ftz.f32 %f653, %f5, %f60;
add.ftz.f32 %f654, %f4, %f61;
mul.ftz.f32 %f655, %f654, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f656, %f655;
add.ftz.f32 %f657, %f656, 0f3F800000;
div.approx.ftz.f32 %f658, %f654, %f657;
mul.ftz.f32 %f659, %f653, %f658;
st.global.f32 [%rd24+8], %f659;

$L__BB28_42:
ld.shared.f32 %f660, [%r25+384];
add.ftz.f32 %f661, %f660, %f27;
ld.shared.f32 %f662, [%r27+384];
add.ftz.f32 %f663, %f662, %f31;
ld.shared.f32 %f664, [%r25+896];
add.ftz.f32 %f665, %f664, %f661;
ld.shared.f32 %f666, [%r27+896];
add.ftz.f32 %f667, %f666, %f663;
ld.shared.f32 %f668, [%r25+1408];
add.ftz.f32 %f669, %f668, %f665;
ld.shared.f32 %f670, [%r27+1408];
add.ftz.f32 %f671, %f670, %f667;
mov.b32 %r735, %f669;
shfl.sync.bfly.b32 %r739|%p202, %r735, %r706, %r705, %r707;
mov.b32 %f672, %r739;
add.ftz.f32 %f673, %f669, %f672;
mov.b32 %r740, %f673;
shfl.sync.bfly.b32 %r742|%p203, %r740, %r710, %r705, %r707;
mov.b32 %f674, %r742;
add.ftz.f32 %f675, %f673, %f674;
mov.b32 %r743, %f675;
shfl.sync.bfly.b32 %r745|%p204, %r743, %r713, %r705, %r707;
mov.b32 %f676, %r745;
add.ftz.f32 %f677, %f675, %f676;
mov.b32 %r746, %f677;
shfl.sync.bfly.b32 %r748|%p205, %r746, %r716, %r705, %r707;
mov.b32 %f678, %r748;
add.ftz.f32 %f679, %f677, %f678;
mov.b32 %r749, %f679;
shfl.sync.bfly.b32 %r751|%p206, %r749, %r719, %r705, %r707;
mov.b32 %f680, %r751;
add.ftz.f32 %f62, %f679, %f680;
mov.b32 %r752, %f671;
shfl.sync.bfly.b32 %r753|%p207, %r752, %r706, %r705, %r707;
mov.b32 %f681, %r753;
add.ftz.f32 %f682, %f671, %f681;
mov.b32 %r754, %f682;
shfl.sync.bfly.b32 %r755|%p208, %r754, %r710, %r705, %r707;
mov.b32 %f683, %r755;
add.ftz.f32 %f684, %f682, %f683;
mov.b32 %r756, %f684;
shfl.sync.bfly.b32 %r757|%p209, %r756, %r713, %r705, %r707;
mov.b32 %f685, %r757;
add.ftz.f32 %f686, %f684, %f685;
mov.b32 %r758, %f686;
shfl.sync.bfly.b32 %r759|%p210, %r758, %r716, %r705, %r707;
mov.b32 %f687, %r759;
add.ftz.f32 %f688, %f686, %f687;
mov.b32 %r760, %f688;
shfl.sync.bfly.b32 %r761|%p211, %r760, %r719, %r705, %r707;
mov.b32 %f689, %r761;
add.ftz.f32 %f63, %f688, %f689;
setp.ne.s32 %p212, %r2, 3;
or.b32 %r765, %r4, 3;
setp.ge.u32 %p213, %r765, %r44;
or.pred %p214, %p212, %p213;
@%p214 bra $L__BB28_78;

add.ftz.f32 %f690, %f5, %f62;
add.ftz.f32 %f691, %f4, %f63;
mul.ftz.f32 %f692, %f691, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f693, %f692;
add.ftz.f32 %f694, %f693, 0f3F800000;
div.approx.ftz.f32 %f695, %f691, %f694;
mul.ftz.f32 %f696, %f690, %f695;
st.global.f32 [%rd24+12], %f696;
bra.uni $L__BB28_78;

$L__BB28_25:
ld.shared.f32 %f149, [%r25];
add.ftz.f32 %f150, %f149, %f30;
ld.shared.f32 %f151, [%r25+512];
add.ftz.f32 %f152, %f151, %f150;
ld.shared.f32 %f153, [%r25+1024];
add.ftz.f32 %f154, %f153, %f152;
mov.b32 %r310, %f154;
mov.u32 %r311, 31;
mov.u32 %r312, 16;
mov.u32 %r313, -1;
shfl.sync.bfly.b32 %r314|%p22, %r310, %r312, %r311, %r313;
mov.b32 %f155, %r314;
add.ftz.f32 %f156, %f154, %f155;
mov.b32 %r315, %f156;
mov.u32 %r316, 8;
shfl.sync.bfly.b32 %r317|%p23, %r315, %r316, %r311, %r313;
mov.b32 %f157, %r317;
add.ftz.f32 %f158, %f156, %f157;
mov.b32 %r318, %f158;
mov.u32 %r319, 4;
shfl.sync.bfly.b32 %r320|%p24, %r318, %r319, %r311, %r313;
mov.b32 %f159, %r320;
add.ftz.f32 %f160, %f158, %f159;
mov.b32 %r321, %f160;
mov.u32 %r322, 2;
shfl.sync.bfly.b32 %r323|%p25, %r321, %r322, %r311, %r313;
mov.b32 %f161, %r323;
add.ftz.f32 %f162, %f160, %f161;
mov.b32 %r324, %f162;
mov.u32 %r325, 1;
shfl.sync.bfly.b32 %r326|%p26, %r324, %r325, %r311, %r313;
mov.b32 %f163, %r326;
add.ftz.f32 %f36, %f162, %f163;
setp.ne.s32 %p27, %r2, 0;
setp.ge.u32 %p28, %r4, %r44;
or.pred %p29, %p27, %p28;
@%p29 bra $L__BB28_27;

add.ftz.f32 %f164, %f5, %f36;
st.global.f32 [%rd24], %f164;

$L__BB28_27:
ld.shared.f32 %f165, [%r25+128];
add.ftz.f32 %f166, %f165, %f29;
ld.shared.f32 %f167, [%r25+640];
add.ftz.f32 %f168, %f167, %f166;
ld.shared.f32 %f169, [%r25+1152];
add.ftz.f32 %f170, %f169, %f168;
mov.b32 %r330, %f170;
shfl.sync.bfly.b32 %r334|%p30, %r330, %r312, %r311, %r313;
mov.b32 %f171, %r334;
add.ftz.f32 %f172, %f170, %f171;
mov.b32 %r335, %f172;
shfl.sync.bfly.b32 %r337|%p31, %r335, %r316, %r311, %r313;
mov.b32 %f173, %r337;
add.ftz.f32 %f174, %f172, %f173;
mov.b32 %r338, %f174;
shfl.sync.bfly.b32 %r340|%p32, %r338, %r319, %r311, %r313;
mov.b32 %f175, %r340;
add.ftz.f32 %f176, %f174, %f175;
mov.b32 %r341, %f176;
shfl.sync.bfly.b32 %r343|%p33, %r341, %r322, %r311, %r313;
mov.b32 %f177, %r343;
add.ftz.f32 %f178, %f176, %f177;
mov.b32 %r344, %f178;
shfl.sync.bfly.b32 %r346|%p34, %r344, %r325, %r311, %r313;
mov.b32 %f179, %r346;
add.ftz.f32 %f37, %f178, %f179;
setp.ne.s32 %p35, %r2, 1;
or.b32 %r350, %r4, 1;
setp.ge.u32 %p36, %r350, %r44;
or.pred %p37, %p35, %p36;
@%p37 bra $L__BB28_29;

add.ftz.f32 %f180, %f5, %f37;
st.global.f32 [%rd24+4], %f180;

$L__BB28_29:
ld.shared.f32 %f181, [%r25+256];
add.ftz.f32 %f182, %f181, %f28;
ld.shared.f32 %f183, [%r25+768];
add.ftz.f32 %f184, %f183, %f182;
ld.shared.f32 %f185, [%r25+1280];
add.ftz.f32 %f186, %f185, %f184;
mov.b32 %r351, %f186;
mov.u32 %r352, 31;
mov.u32 %r353, 16;
mov.u32 %r354, -1;
shfl.sync.bfly.b32 %r355|%p38, %r351, %r353, %r352, %r354;
mov.b32 %f187, %r355;
add.ftz.f32 %f188, %f186, %f187;
mov.b32 %r356, %f188;
mov.u32 %r357, 8;
shfl.sync.bfly.b32 %r358|%p39, %r356, %r357, %r352, %r354;
mov.b32 %f189, %r358;
add.ftz.f32 %f190, %f188, %f189;
mov.b32 %r359, %f190;
mov.u32 %r360, 4;
shfl.sync.bfly.b32 %r361|%p40, %r359, %r360, %r352, %r354;
mov.b32 %f191, %r361;
add.ftz.f32 %f192, %f190, %f191;
mov.b32 %r362, %f192;
mov.u32 %r363, 2;
shfl.sync.bfly.b32 %r364|%p41, %r362, %r363, %r352, %r354;
mov.b32 %f193, %r364;
add.ftz.f32 %f194, %f192, %f193;
mov.b32 %r365, %f194;
mov.u32 %r366, 1;
shfl.sync.bfly.b32 %r367|%p42, %r365, %r366, %r352, %r354;
mov.b32 %f195, %r367;
add.ftz.f32 %f38, %f194, %f195;
setp.ne.s32 %p43, %r2, 2;
or.b32 %r371, %r4, 2;
setp.ge.u32 %p44, %r371, %r44;
or.pred %p45, %p43, %p44;
@%p45 bra $L__BB28_31;

add.ftz.f32 %f196, %f5, %f38;
st.global.f32 [%rd24+8], %f196;

$L__BB28_31:
ld.shared.f32 %f197, [%r25+384];
add.ftz.f32 %f198, %f197, %f27;
ld.shared.f32 %f199, [%r25+896];
add.ftz.f32 %f200, %f199, %f198;
ld.shared.f32 %f201, [%r25+1408];
add.ftz.f32 %f202, %f201, %f200;
mov.b32 %r372, %f202;
shfl.sync.bfly.b32 %r376|%p46, %r372, %r353, %r352, %r354;
mov.b32 %f203, %r376;
add.ftz.f32 %f204, %f202, %f203;
mov.b32 %r377, %f204;
shfl.sync.bfly.b32 %r379|%p47, %r377, %r357, %r352, %r354;
mov.b32 %f205, %r379;
add.ftz.f32 %f206, %f204, %f205;
mov.b32 %r380, %f206;
shfl.sync.bfly.b32 %r382|%p48, %r380, %r360, %r352, %r354;
mov.b32 %f207, %r382;
add.ftz.f32 %f208, %f206, %f207;
mov.b32 %r383, %f208;
shfl.sync.bfly.b32 %r385|%p49, %r383, %r363, %r352, %r354;
mov.b32 %f209, %r385;
add.ftz.f32 %f210, %f208, %f209;
mov.b32 %r386, %f210;
shfl.sync.bfly.b32 %r388|%p50, %r386, %r366, %r352, %r354;
mov.b32 %f211, %r388;
add.ftz.f32 %f39, %f210, %f211;
setp.ne.s32 %p51, %r2, 3;
or.b32 %r392, %r4, 3;
setp.ge.u32 %p52, %r392, %r44;
or.pred %p53, %p51, %p52;
@%p53 bra $L__BB28_78;

add.ftz.f32 %f212, %f5, %f39;
st.global.f32 [%rd24+12], %f212;
bra.uni $L__BB28_78;

$L__BB28_44:
ld.param.u32 %r1025, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p55, %r1025, 3;
@%p55 bra $L__BB28_54;

ld.param.u32 %r1026, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.ne.s32 %p56, %r1026, 6;
@%p56 bra $L__BB28_70;

ld.shared.f32 %f213, [%r25];
add.ftz.f32 %f214, %f213, %f30;
ld.shared.f32 %f215, [%r27];
add.ftz.f32 %f216, %f215, %f34;
ld.shared.f32 %f217, [%r25+512];
add.ftz.f32 %f218, %f217, %f214;
ld.shared.f32 %f219, [%r27+512];
add.ftz.f32 %f220, %f219, %f216;
ld.shared.f32 %f221, [%r25+1024];
add.ftz.f32 %f222, %f221, %f218;
ld.shared.f32 %f223, [%r27+1024];
add.ftz.f32 %f224, %f223, %f220;
mov.b32 %r397, %f222;
mov.u32 %r398, 31;
mov.u32 %r399, 16;
mov.u32 %r400, -1;
shfl.sync.bfly.b32 %r401|%p59, %r397, %r399, %r398, %r400;
mov.b32 %f225, %r401;
add.ftz.f32 %f226, %f222, %f225;
mov.b32 %r402, %f226;
mov.u32 %r403, 8;
shfl.sync.bfly.b32 %r404|%p60, %r402, %r403, %r398, %r400;
mov.b32 %f227, %r404;
add.ftz.f32 %f228, %f226, %f227;
mov.b32 %r405, %f228;
mov.u32 %r406, 4;
shfl.sync.bfly.b32 %r407|%p61, %r405, %r406, %r398, %r400;
mov.b32 %f229, %r407;
add.ftz.f32 %f230, %f228, %f229;
mov.b32 %r408, %f230;
mov.u32 %r409, 2;
shfl.sync.bfly.b32 %r410|%p62, %r408, %r409, %r398, %r400;
mov.b32 %f231, %r410;
add.ftz.f32 %f232, %f230, %f231;
mov.b32 %r411, %f232;
mov.u32 %r412, 1;
shfl.sync.bfly.b32 %r413|%p63, %r411, %r412, %r398, %r400;
mov.b32 %f233, %r413;
add.ftz.f32 %f40, %f232, %f233;
mov.b32 %r414, %f224;
shfl.sync.bfly.b32 %r415|%p64, %r414, %r399, %r398, %r400;
mov.b32 %f234, %r415;
add.ftz.f32 %f235, %f224, %f234;
mov.b32 %r416, %f235;
shfl.sync.bfly.b32 %r417|%p65, %r416, %r403, %r398, %r400;
mov.b32 %f236, %r417;
add.ftz.f32 %f237, %f235, %f236;
mov.b32 %r418, %f237;
shfl.sync.bfly.b32 %r419|%p66, %r418, %r406, %r398, %r400;
mov.b32 %f238, %r419;
add.ftz.f32 %f239, %f237, %f238;
mov.b32 %r420, %f239;
shfl.sync.bfly.b32 %r421|%p67, %r420, %r409, %r398, %r400;
mov.b32 %f240, %r421;
add.ftz.f32 %f241, %f239, %f240;
mov.b32 %r422, %f241;
shfl.sync.bfly.b32 %r423|%p68, %r422, %r412, %r398, %r400;
mov.b32 %f242, %r423;
add.ftz.f32 %f41, %f241, %f242;
setp.ne.s32 %p69, %r2, 0;
setp.ge.u32 %p70, %r4, %r44;
or.pred %p71, %p69, %p70;
@%p71 bra $L__BB28_48;

ld.param.f32 %f997, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f243, %f5, %f40;
add.ftz.f32 %f244, %f4, %f41;
min.ftz.f32 %f245, %f244, %f997;
min.ftz.f32 %f246, %f243, %f997;
max.ftz.f32 %f247, %f246, %f35;
mul.ftz.f32 %f248, %f245, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f249, %f248;
add.ftz.f32 %f250, %f249, 0f3F800000;
div.approx.ftz.f32 %f251, %f245, %f250;
mul.ftz.f32 %f252, %f247, %f251;
st.global.f32 [%rd24], %f252;

$L__BB28_48:
ld.shared.f32 %f253, [%r25+128];
add.ftz.f32 %f254, %f253, %f29;
ld.shared.f32 %f255, [%r27+128];
add.ftz.f32 %f256, %f255, %f33;
ld.shared.f32 %f257, [%r25+640];
add.ftz.f32 %f258, %f257, %f254;
ld.shared.f32 %f259, [%r27+640];
add.ftz.f32 %f260, %f259, %f256;
ld.shared.f32 %f261, [%r25+1152];
add.ftz.f32 %f262, %f261, %f258;
ld.shared.f32 %f263, [%r27+1152];
add.ftz.f32 %f264, %f263, %f260;
mov.b32 %r427, %f262;
shfl.sync.bfly.b32 %r431|%p72, %r427, %r399, %r398, %r400;
mov.b32 %f265, %r431;
add.ftz.f32 %f266, %f262, %f265;
mov.b32 %r432, %f266;
shfl.sync.bfly.b32 %r434|%p73, %r432, %r403, %r398, %r400;
mov.b32 %f267, %r434;
add.ftz.f32 %f268, %f266, %f267;
mov.b32 %r435, %f268;
shfl.sync.bfly.b32 %r437|%p74, %r435, %r406, %r398, %r400;
mov.b32 %f269, %r437;
add.ftz.f32 %f270, %f268, %f269;
mov.b32 %r438, %f270;
shfl.sync.bfly.b32 %r440|%p75, %r438, %r409, %r398, %r400;
mov.b32 %f271, %r440;
add.ftz.f32 %f272, %f270, %f271;
mov.b32 %r441, %f272;
shfl.sync.bfly.b32 %r443|%p76, %r441, %r412, %r398, %r400;
mov.b32 %f273, %r443;
add.ftz.f32 %f42, %f272, %f273;
mov.b32 %r444, %f264;
shfl.sync.bfly.b32 %r445|%p77, %r444, %r399, %r398, %r400;
mov.b32 %f274, %r445;
add.ftz.f32 %f275, %f264, %f274;
mov.b32 %r446, %f275;
shfl.sync.bfly.b32 %r447|%p78, %r446, %r403, %r398, %r400;
mov.b32 %f276, %r447;
add.ftz.f32 %f277, %f275, %f276;
mov.b32 %r448, %f277;
shfl.sync.bfly.b32 %r449|%p79, %r448, %r406, %r398, %r400;
mov.b32 %f278, %r449;
add.ftz.f32 %f279, %f277, %f278;
mov.b32 %r450, %f279;
shfl.sync.bfly.b32 %r451|%p80, %r450, %r409, %r398, %r400;
mov.b32 %f280, %r451;
add.ftz.f32 %f281, %f279, %f280;
mov.b32 %r452, %f281;
shfl.sync.bfly.b32 %r453|%p81, %r452, %r412, %r398, %r400;
mov.b32 %f282, %r453;
add.ftz.f32 %f43, %f281, %f282;
setp.ne.s32 %p82, %r2, 1;
or.b32 %r457, %r4, 1;
setp.ge.u32 %p83, %r457, %r44;
or.pred %p84, %p82, %p83;
@%p84 bra $L__BB28_50;

ld.param.f32 %f996, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f283, %f5, %f42;
add.ftz.f32 %f284, %f4, %f43;
min.ftz.f32 %f285, %f284, %f996;
min.ftz.f32 %f286, %f283, %f996;
max.ftz.f32 %f287, %f286, %f35;
mul.ftz.f32 %f288, %f285, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f289, %f288;
add.ftz.f32 %f290, %f289, 0f3F800000;
div.approx.ftz.f32 %f291, %f285, %f290;
mul.ftz.f32 %f292, %f287, %f291;
st.global.f32 [%rd24+4], %f292;

$L__BB28_50:
ld.shared.f32 %f293, [%r25+256];
add.ftz.f32 %f294, %f293, %f28;
ld.shared.f32 %f295, [%r27+256];
add.ftz.f32 %f296, %f295, %f32;
ld.shared.f32 %f297, [%r25+768];
add.ftz.f32 %f298, %f297, %f294;
ld.shared.f32 %f299, [%r27+768];
add.ftz.f32 %f300, %f299, %f296;
ld.shared.f32 %f301, [%r25+1280];
add.ftz.f32 %f302, %f301, %f298;
ld.shared.f32 %f303, [%r27+1280];
add.ftz.f32 %f304, %f303, %f300;
mov.b32 %r458, %f302;
mov.u32 %r459, 31;
mov.u32 %r460, 16;
mov.u32 %r461, -1;
shfl.sync.bfly.b32 %r462|%p85, %r458, %r460, %r459, %r461;
mov.b32 %f305, %r462;
add.ftz.f32 %f306, %f302, %f305;
mov.b32 %r463, %f306;
mov.u32 %r464, 8;
shfl.sync.bfly.b32 %r465|%p86, %r463, %r464, %r459, %r461;
mov.b32 %f307, %r465;
add.ftz.f32 %f308, %f306, %f307;
mov.b32 %r466, %f308;
mov.u32 %r467, 4;
shfl.sync.bfly.b32 %r468|%p87, %r466, %r467, %r459, %r461;
mov.b32 %f309, %r468;
add.ftz.f32 %f310, %f308, %f309;
mov.b32 %r469, %f310;
mov.u32 %r470, 2;
shfl.sync.bfly.b32 %r471|%p88, %r469, %r470, %r459, %r461;
mov.b32 %f311, %r471;
add.ftz.f32 %f312, %f310, %f311;
mov.b32 %r472, %f312;
mov.u32 %r473, 1;
shfl.sync.bfly.b32 %r474|%p89, %r472, %r473, %r459, %r461;
mov.b32 %f313, %r474;
add.ftz.f32 %f44, %f312, %f313;
mov.b32 %r475, %f304;
shfl.sync.bfly.b32 %r476|%p90, %r475, %r460, %r459, %r461;
mov.b32 %f314, %r476;
add.ftz.f32 %f315, %f304, %f314;
mov.b32 %r477, %f315;
shfl.sync.bfly.b32 %r478|%p91, %r477, %r464, %r459, %r461;
mov.b32 %f316, %r478;
add.ftz.f32 %f317, %f315, %f316;
mov.b32 %r479, %f317;
shfl.sync.bfly.b32 %r480|%p92, %r479, %r467, %r459, %r461;
mov.b32 %f318, %r480;
add.ftz.f32 %f319, %f317, %f318;
mov.b32 %r481, %f319;
shfl.sync.bfly.b32 %r482|%p93, %r481, %r470, %r459, %r461;
mov.b32 %f320, %r482;
add.ftz.f32 %f321, %f319, %f320;
mov.b32 %r483, %f321;
shfl.sync.bfly.b32 %r484|%p94, %r483, %r473, %r459, %r461;
mov.b32 %f322, %r484;
add.ftz.f32 %f45, %f321, %f322;
setp.ne.s32 %p95, %r2, 2;
or.b32 %r488, %r4, 2;
setp.ge.u32 %p96, %r488, %r44;
or.pred %p97, %p95, %p96;
@%p97 bra $L__BB28_52;

ld.param.f32 %f995, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f323, %f5, %f44;
add.ftz.f32 %f324, %f4, %f45;
min.ftz.f32 %f325, %f324, %f995;
min.ftz.f32 %f326, %f323, %f995;
max.ftz.f32 %f327, %f326, %f35;
mul.ftz.f32 %f328, %f325, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f329, %f328;
add.ftz.f32 %f330, %f329, 0f3F800000;
div.approx.ftz.f32 %f331, %f325, %f330;
mul.ftz.f32 %f332, %f327, %f331;
st.global.f32 [%rd24+8], %f332;

$L__BB28_52:
ld.shared.f32 %f333, [%r25+384];
add.ftz.f32 %f334, %f333, %f27;
ld.shared.f32 %f335, [%r27+384];
add.ftz.f32 %f336, %f335, %f31;
ld.shared.f32 %f337, [%r25+896];
add.ftz.f32 %f338, %f337, %f334;
ld.shared.f32 %f339, [%r27+896];
add.ftz.f32 %f340, %f339, %f336;
ld.shared.f32 %f341, [%r25+1408];
add.ftz.f32 %f342, %f341, %f338;
ld.shared.f32 %f343, [%r27+1408];
add.ftz.f32 %f344, %f343, %f340;
mov.b32 %r489, %f342;
shfl.sync.bfly.b32 %r493|%p98, %r489, %r460, %r459, %r461;
mov.b32 %f345, %r493;
add.ftz.f32 %f346, %f342, %f345;
mov.b32 %r494, %f346;
shfl.sync.bfly.b32 %r496|%p99, %r494, %r464, %r459, %r461;
mov.b32 %f347, %r496;
add.ftz.f32 %f348, %f346, %f347;
mov.b32 %r497, %f348;
shfl.sync.bfly.b32 %r499|%p100, %r497, %r467, %r459, %r461;
mov.b32 %f349, %r499;
add.ftz.f32 %f350, %f348, %f349;
mov.b32 %r500, %f350;
shfl.sync.bfly.b32 %r502|%p101, %r500, %r470, %r459, %r461;
mov.b32 %f351, %r502;
add.ftz.f32 %f352, %f350, %f351;
mov.b32 %r503, %f352;
shfl.sync.bfly.b32 %r505|%p102, %r503, %r473, %r459, %r461;
mov.b32 %f353, %r505;
add.ftz.f32 %f46, %f352, %f353;
mov.b32 %r506, %f344;
shfl.sync.bfly.b32 %r507|%p103, %r506, %r460, %r459, %r461;
mov.b32 %f354, %r507;
add.ftz.f32 %f355, %f344, %f354;
mov.b32 %r508, %f355;
shfl.sync.bfly.b32 %r509|%p104, %r508, %r464, %r459, %r461;
mov.b32 %f356, %r509;
add.ftz.f32 %f357, %f355, %f356;
mov.b32 %r510, %f357;
shfl.sync.bfly.b32 %r511|%p105, %r510, %r467, %r459, %r461;
mov.b32 %f358, %r511;
add.ftz.f32 %f359, %f357, %f358;
mov.b32 %r512, %f359;
shfl.sync.bfly.b32 %r513|%p106, %r512, %r470, %r459, %r461;
mov.b32 %f360, %r513;
add.ftz.f32 %f361, %f359, %f360;
mov.b32 %r514, %f361;
shfl.sync.bfly.b32 %r515|%p107, %r514, %r473, %r459, %r461;
mov.b32 %f362, %r515;
add.ftz.f32 %f47, %f361, %f362;
setp.ne.s32 %p108, %r2, 3;
or.b32 %r519, %r4, 3;
setp.ge.u32 %p109, %r519, %r44;
or.pred %p110, %p108, %p109;
@%p110 bra $L__BB28_78;

ld.param.f32 %f994, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f363, %f5, %f46;
add.ftz.f32 %f364, %f4, %f47;
min.ftz.f32 %f365, %f364, %f994;
min.ftz.f32 %f366, %f363, %f994;
max.ftz.f32 %f367, %f366, %f35;
mul.ftz.f32 %f368, %f365, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f369, %f368;
add.ftz.f32 %f370, %f369, 0f3F800000;
div.approx.ftz.f32 %f371, %f365, %f370;
mul.ftz.f32 %f372, %f367, %f371;
st.global.f32 [%rd24+12], %f372;
bra.uni $L__BB28_78;

$L__BB28_70:
ld.shared.f32 %f861, [%r25];
add.ftz.f32 %f862, %f861, %f30;
ld.shared.f32 %f863, [%r27];
add.ftz.f32 %f864, %f863, %f34;
ld.shared.f32 %f865, [%r25+512];
add.ftz.f32 %f866, %f865, %f862;
ld.shared.f32 %f867, [%r27+512];
add.ftz.f32 %f868, %f867, %f864;
ld.shared.f32 %f869, [%r25+1024];
add.ftz.f32 %f870, %f869, %f866;
ld.shared.f32 %f871, [%r27+1024];
add.ftz.f32 %f872, %f871, %f868;
mov.b32 %r889, %f870;
mov.u32 %r890, 31;
mov.u32 %r891, 16;
mov.u32 %r892, -1;
shfl.sync.bfly.b32 %r893|%p267, %r889, %r891, %r890, %r892;
mov.b32 %f873, %r893;
add.ftz.f32 %f874, %f870, %f873;
mov.b32 %r894, %f874;
mov.u32 %r895, 8;
shfl.sync.bfly.b32 %r896|%p268, %r894, %r895, %r890, %r892;
mov.b32 %f875, %r896;
add.ftz.f32 %f876, %f874, %f875;
mov.b32 %r897, %f876;
mov.u32 %r898, 4;
shfl.sync.bfly.b32 %r899|%p269, %r897, %r898, %r890, %r892;
mov.b32 %f877, %r899;
add.ftz.f32 %f878, %f876, %f877;
mov.b32 %r900, %f878;
mov.u32 %r901, 2;
shfl.sync.bfly.b32 %r902|%p270, %r900, %r901, %r890, %r892;
mov.b32 %f879, %r902;
add.ftz.f32 %f880, %f878, %f879;
mov.b32 %r903, %f880;
mov.u32 %r904, 1;
shfl.sync.bfly.b32 %r905|%p271, %r903, %r904, %r890, %r892;
mov.b32 %f881, %r905;
add.ftz.f32 %f72, %f880, %f881;
mov.b32 %r906, %f872;
shfl.sync.bfly.b32 %r907|%p272, %r906, %r891, %r890, %r892;
mov.b32 %f882, %r907;
add.ftz.f32 %f883, %f872, %f882;
mov.b32 %r908, %f883;
shfl.sync.bfly.b32 %r909|%p273, %r908, %r895, %r890, %r892;
mov.b32 %f884, %r909;
add.ftz.f32 %f885, %f883, %f884;
mov.b32 %r910, %f885;
shfl.sync.bfly.b32 %r911|%p274, %r910, %r898, %r890, %r892;
mov.b32 %f886, %r911;
add.ftz.f32 %f887, %f885, %f886;
mov.b32 %r912, %f887;
shfl.sync.bfly.b32 %r913|%p275, %r912, %r901, %r890, %r892;
mov.b32 %f888, %r913;
add.ftz.f32 %f889, %f887, %f888;
mov.b32 %r914, %f889;
shfl.sync.bfly.b32 %r915|%p276, %r914, %r904, %r890, %r892;
mov.b32 %f890, %r915;
add.ftz.f32 %f73, %f889, %f890;
setp.ne.s32 %p277, %r2, 0;
setp.ge.u32 %p278, %r4, %r44;
or.pred %p279, %p277, %p278;
@%p279 bra $L__BB28_72;

add.ftz.f32 %f891, %f5, %f72;
add.ftz.f32 %f892, %f4, %f73;
mul.ftz.f32 %f893, %f891, %f892;
st.global.f32 [%rd24], %f893;

$L__BB28_72:
ld.shared.f32 %f894, [%r25+128];
add.ftz.f32 %f895, %f894, %f29;
ld.shared.f32 %f896, [%r27+128];
add.ftz.f32 %f897, %f896, %f33;
ld.shared.f32 %f898, [%r25+640];
add.ftz.f32 %f899, %f898, %f895;
ld.shared.f32 %f900, [%r27+640];
add.ftz.f32 %f901, %f900, %f897;
ld.shared.f32 %f902, [%r25+1152];
add.ftz.f32 %f903, %f902, %f899;
ld.shared.f32 %f904, [%r27+1152];
add.ftz.f32 %f905, %f904, %f901;
mov.b32 %r919, %f903;
shfl.sync.bfly.b32 %r923|%p280, %r919, %r891, %r890, %r892;
mov.b32 %f906, %r923;
add.ftz.f32 %f907, %f903, %f906;
mov.b32 %r924, %f907;
shfl.sync.bfly.b32 %r926|%p281, %r924, %r895, %r890, %r892;
mov.b32 %f908, %r926;
add.ftz.f32 %f909, %f907, %f908;
mov.b32 %r927, %f909;
shfl.sync.bfly.b32 %r929|%p282, %r927, %r898, %r890, %r892;
mov.b32 %f910, %r929;
add.ftz.f32 %f911, %f909, %f910;
mov.b32 %r930, %f911;
shfl.sync.bfly.b32 %r932|%p283, %r930, %r901, %r890, %r892;
mov.b32 %f912, %r932;
add.ftz.f32 %f913, %f911, %f912;
mov.b32 %r933, %f913;
shfl.sync.bfly.b32 %r935|%p284, %r933, %r904, %r890, %r892;
mov.b32 %f914, %r935;
add.ftz.f32 %f74, %f913, %f914;
mov.b32 %r936, %f905;
shfl.sync.bfly.b32 %r937|%p285, %r936, %r891, %r890, %r892;
mov.b32 %f915, %r937;
add.ftz.f32 %f916, %f905, %f915;
mov.b32 %r938, %f916;
shfl.sync.bfly.b32 %r939|%p286, %r938, %r895, %r890, %r892;
mov.b32 %f917, %r939;
add.ftz.f32 %f918, %f916, %f917;
mov.b32 %r940, %f918;
shfl.sync.bfly.b32 %r941|%p287, %r940, %r898, %r890, %r892;
mov.b32 %f919, %r941;
add.ftz.f32 %f920, %f918, %f919;
mov.b32 %r942, %f920;
shfl.sync.bfly.b32 %r943|%p288, %r942, %r901, %r890, %r892;
mov.b32 %f921, %r943;
add.ftz.f32 %f922, %f920, %f921;
mov.b32 %r944, %f922;
shfl.sync.bfly.b32 %r945|%p289, %r944, %r904, %r890, %r892;
mov.b32 %f923, %r945;
add.ftz.f32 %f75, %f922, %f923;
setp.ne.s32 %p290, %r2, 1;
or.b32 %r949, %r4, 1;
setp.ge.u32 %p291, %r949, %r44;
or.pred %p292, %p290, %p291;
@%p292 bra $L__BB28_74;

add.ftz.f32 %f924, %f5, %f74;
add.ftz.f32 %f925, %f4, %f75;
mul.ftz.f32 %f926, %f924, %f925;
st.global.f32 [%rd24+4], %f926;

$L__BB28_74:
ld.shared.f32 %f927, [%r25+256];
add.ftz.f32 %f928, %f927, %f28;
ld.shared.f32 %f929, [%r27+256];
add.ftz.f32 %f930, %f929, %f32;
ld.shared.f32 %f931, [%r25+768];
add.ftz.f32 %f932, %f931, %f928;
ld.shared.f32 %f933, [%r27+768];
add.ftz.f32 %f934, %f933, %f930;
ld.shared.f32 %f935, [%r25+1280];
add.ftz.f32 %f936, %f935, %f932;
ld.shared.f32 %f937, [%r27+1280];
add.ftz.f32 %f938, %f937, %f934;
mov.b32 %r950, %f936;
mov.u32 %r951, 31;
mov.u32 %r952, 16;
mov.u32 %r953, -1;
shfl.sync.bfly.b32 %r954|%p293, %r950, %r952, %r951, %r953;
mov.b32 %f939, %r954;
add.ftz.f32 %f940, %f936, %f939;
mov.b32 %r955, %f940;
mov.u32 %r956, 8;
shfl.sync.bfly.b32 %r957|%p294, %r955, %r956, %r951, %r953;
mov.b32 %f941, %r957;
add.ftz.f32 %f942, %f940, %f941;
mov.b32 %r958, %f942;
mov.u32 %r959, 4;
shfl.sync.bfly.b32 %r960|%p295, %r958, %r959, %r951, %r953;
mov.b32 %f943, %r960;
add.ftz.f32 %f944, %f942, %f943;
mov.b32 %r961, %f944;
mov.u32 %r962, 2;
shfl.sync.bfly.b32 %r963|%p296, %r961, %r962, %r951, %r953;
mov.b32 %f945, %r963;
add.ftz.f32 %f946, %f944, %f945;
mov.b32 %r964, %f946;
mov.u32 %r965, 1;
shfl.sync.bfly.b32 %r966|%p297, %r964, %r965, %r951, %r953;
mov.b32 %f947, %r966;
add.ftz.f32 %f76, %f946, %f947;
mov.b32 %r967, %f938;
shfl.sync.bfly.b32 %r968|%p298, %r967, %r952, %r951, %r953;
mov.b32 %f948, %r968;
add.ftz.f32 %f949, %f938, %f948;
mov.b32 %r969, %f949;
shfl.sync.bfly.b32 %r970|%p299, %r969, %r956, %r951, %r953;
mov.b32 %f950, %r970;
add.ftz.f32 %f951, %f949, %f950;
mov.b32 %r971, %f951;
shfl.sync.bfly.b32 %r972|%p300, %r971, %r959, %r951, %r953;
mov.b32 %f952, %r972;
add.ftz.f32 %f953, %f951, %f952;
mov.b32 %r973, %f953;
shfl.sync.bfly.b32 %r974|%p301, %r973, %r962, %r951, %r953;
mov.b32 %f954, %r974;
add.ftz.f32 %f955, %f953, %f954;
mov.b32 %r975, %f955;
shfl.sync.bfly.b32 %r976|%p302, %r975, %r965, %r951, %r953;
mov.b32 %f956, %r976;
add.ftz.f32 %f77, %f955, %f956;
setp.ne.s32 %p303, %r2, 2;
or.b32 %r980, %r4, 2;
setp.ge.u32 %p304, %r980, %r44;
or.pred %p305, %p303, %p304;
@%p305 bra $L__BB28_76;

add.ftz.f32 %f957, %f5, %f76;
add.ftz.f32 %f958, %f4, %f77;
mul.ftz.f32 %f959, %f957, %f958;
st.global.f32 [%rd24+8], %f959;

$L__BB28_76:
ld.shared.f32 %f960, [%r25+384];
add.ftz.f32 %f961, %f960, %f27;
ld.shared.f32 %f962, [%r27+384];
add.ftz.f32 %f963, %f962, %f31;
ld.shared.f32 %f964, [%r25+896];
add.ftz.f32 %f965, %f964, %f961;
ld.shared.f32 %f966, [%r27+896];
add.ftz.f32 %f967, %f966, %f963;
ld.shared.f32 %f968, [%r25+1408];
add.ftz.f32 %f969, %f968, %f965;
ld.shared.f32 %f970, [%r27+1408];
add.ftz.f32 %f971, %f970, %f967;
mov.b32 %r981, %f969;
shfl.sync.bfly.b32 %r985|%p306, %r981, %r952, %r951, %r953;
mov.b32 %f972, %r985;
add.ftz.f32 %f973, %f969, %f972;
mov.b32 %r986, %f973;
shfl.sync.bfly.b32 %r988|%p307, %r986, %r956, %r951, %r953;
mov.b32 %f974, %r988;
add.ftz.f32 %f975, %f973, %f974;
mov.b32 %r989, %f975;
shfl.sync.bfly.b32 %r991|%p308, %r989, %r959, %r951, %r953;
mov.b32 %f976, %r991;
add.ftz.f32 %f977, %f975, %f976;
mov.b32 %r992, %f977;
shfl.sync.bfly.b32 %r994|%p309, %r992, %r962, %r951, %r953;
mov.b32 %f978, %r994;
add.ftz.f32 %f979, %f977, %f978;
mov.b32 %r995, %f979;
shfl.sync.bfly.b32 %r997|%p310, %r995, %r965, %r951, %r953;
mov.b32 %f980, %r997;
add.ftz.f32 %f78, %f979, %f980;
mov.b32 %r998, %f971;
shfl.sync.bfly.b32 %r999|%p311, %r998, %r952, %r951, %r953;
mov.b32 %f981, %r999;
add.ftz.f32 %f982, %f971, %f981;
mov.b32 %r1000, %f982;
shfl.sync.bfly.b32 %r1001|%p312, %r1000, %r956, %r951, %r953;
mov.b32 %f983, %r1001;
add.ftz.f32 %f984, %f982, %f983;
mov.b32 %r1002, %f984;
shfl.sync.bfly.b32 %r1003|%p313, %r1002, %r959, %r951, %r953;
mov.b32 %f985, %r1003;
add.ftz.f32 %f986, %f984, %f985;
mov.b32 %r1004, %f986;
shfl.sync.bfly.b32 %r1005|%p314, %r1004, %r962, %r951, %r953;
mov.b32 %f987, %r1005;
add.ftz.f32 %f988, %f986, %f987;
mov.b32 %r1006, %f988;
shfl.sync.bfly.b32 %r1007|%p315, %r1006, %r965, %r951, %r953;
mov.b32 %f989, %r1007;
add.ftz.f32 %f79, %f988, %f989;
setp.ne.s32 %p316, %r2, 3;
or.b32 %r1011, %r4, 3;
setp.ge.u32 %p317, %r1011, %r44;
or.pred %p318, %p316, %p317;
@%p318 bra $L__BB28_78;

add.ftz.f32 %f990, %f5, %f78;
add.ftz.f32 %f991, %f4, %f79;
mul.ftz.f32 %f992, %f990, %f991;
st.global.f32 [%rd24+12], %f992;
bra.uni $L__BB28_78;

$L__BB28_62:
ld.shared.f32 %f697, [%r25];
add.ftz.f32 %f698, %f697, %f30;
ld.shared.f32 %f699, [%r27];
add.ftz.f32 %f700, %f699, %f34;
ld.shared.f32 %f701, [%r25+512];
add.ftz.f32 %f702, %f701, %f698;
ld.shared.f32 %f703, [%r27+512];
add.ftz.f32 %f704, %f703, %f700;
ld.shared.f32 %f705, [%r25+1024];
add.ftz.f32 %f706, %f705, %f702;
ld.shared.f32 %f707, [%r27+1024];
add.ftz.f32 %f708, %f707, %f704;
mov.b32 %r766, %f706;
mov.u32 %r767, 31;
mov.u32 %r768, 16;
mov.u32 %r769, -1;
shfl.sync.bfly.b32 %r770|%p215, %r766, %r768, %r767, %r769;
mov.b32 %f709, %r770;
add.ftz.f32 %f710, %f706, %f709;
mov.b32 %r771, %f710;
mov.u32 %r772, 8;
shfl.sync.bfly.b32 %r773|%p216, %r771, %r772, %r767, %r769;
mov.b32 %f711, %r773;
add.ftz.f32 %f712, %f710, %f711;
mov.b32 %r774, %f712;
mov.u32 %r775, 4;
shfl.sync.bfly.b32 %r776|%p217, %r774, %r775, %r767, %r769;
mov.b32 %f713, %r776;
add.ftz.f32 %f714, %f712, %f713;
mov.b32 %r777, %f714;
mov.u32 %r778, 2;
shfl.sync.bfly.b32 %r779|%p218, %r777, %r778, %r767, %r769;
mov.b32 %f715, %r779;
add.ftz.f32 %f716, %f714, %f715;
mov.b32 %r780, %f716;
mov.u32 %r781, 1;
shfl.sync.bfly.b32 %r782|%p219, %r780, %r781, %r767, %r769;
mov.b32 %f717, %r782;
add.ftz.f32 %f64, %f716, %f717;
mov.b32 %r783, %f708;
shfl.sync.bfly.b32 %r784|%p220, %r783, %r768, %r767, %r769;
mov.b32 %f718, %r784;
add.ftz.f32 %f719, %f708, %f718;
mov.b32 %r785, %f719;
shfl.sync.bfly.b32 %r786|%p221, %r785, %r772, %r767, %r769;
mov.b32 %f720, %r786;
add.ftz.f32 %f721, %f719, %f720;
mov.b32 %r787, %f721;
shfl.sync.bfly.b32 %r788|%p222, %r787, %r775, %r767, %r769;
mov.b32 %f722, %r788;
add.ftz.f32 %f723, %f721, %f722;
mov.b32 %r789, %f723;
shfl.sync.bfly.b32 %r790|%p223, %r789, %r778, %r767, %r769;
mov.b32 %f724, %r790;
add.ftz.f32 %f725, %f723, %f724;
mov.b32 %r791, %f725;
shfl.sync.bfly.b32 %r792|%p224, %r791, %r781, %r767, %r769;
mov.b32 %f726, %r792;
add.ftz.f32 %f65, %f725, %f726;
setp.ne.s32 %p225, %r2, 0;
setp.ge.u32 %p226, %r4, %r44;
or.pred %p227, %p225, %p226;
@%p227 bra $L__BB28_64;

add.ftz.f32 %f727, %f5, %f64;
add.ftz.f32 %f728, %f4, %f65;
mul.ftz.f32 %f729, %f728, 0f3F000000;
mul.ftz.f32 %f730, %f728, 0f3F4C422A;
mul.ftz.f32 %f731, %f728, 0f3D372713;
fma.rn.ftz.f32 %f732, %f728, %f731, 0f3F800000;
mul.ftz.f32 %f733, %f730, %f732;
tanh.approx.f32 %f734, %f733;
add.ftz.f32 %f735, %f734, 0f3F800000;
mul.ftz.f32 %f736, %f729, %f735;
mul.ftz.f32 %f737, %f727, %f736;
st.global.f32 [%rd24], %f737;

$L__BB28_64:
ld.shared.f32 %f738, [%r25+128];
add.ftz.f32 %f739, %f738, %f29;
ld.shared.f32 %f740, [%r27+128];
add.ftz.f32 %f741, %f740, %f33;
ld.shared.f32 %f742, [%r25+640];
add.ftz.f32 %f743, %f742, %f739;
ld.shared.f32 %f744, [%r27+640];
add.ftz.f32 %f745, %f744, %f741;
ld.shared.f32 %f746, [%r25+1152];
add.ftz.f32 %f747, %f746, %f743;
ld.shared.f32 %f748, [%r27+1152];
add.ftz.f32 %f749, %f748, %f745;
mov.b32 %r796, %f747;
shfl.sync.bfly.b32 %r800|%p228, %r796, %r768, %r767, %r769;
mov.b32 %f750, %r800;
add.ftz.f32 %f751, %f747, %f750;
mov.b32 %r801, %f751;
shfl.sync.bfly.b32 %r803|%p229, %r801, %r772, %r767, %r769;
mov.b32 %f752, %r803;
add.ftz.f32 %f753, %f751, %f752;
mov.b32 %r804, %f753;
shfl.sync.bfly.b32 %r806|%p230, %r804, %r775, %r767, %r769;
mov.b32 %f754, %r806;
add.ftz.f32 %f755, %f753, %f754;
mov.b32 %r807, %f755;
shfl.sync.bfly.b32 %r809|%p231, %r807, %r778, %r767, %r769;
mov.b32 %f756, %r809;
add.ftz.f32 %f757, %f755, %f756;
mov.b32 %r810, %f757;
shfl.sync.bfly.b32 %r812|%p232, %r810, %r781, %r767, %r769;
mov.b32 %f758, %r812;
add.ftz.f32 %f66, %f757, %f758;
mov.b32 %r813, %f749;
shfl.sync.bfly.b32 %r814|%p233, %r813, %r768, %r767, %r769;
mov.b32 %f759, %r814;
add.ftz.f32 %f760, %f749, %f759;
mov.b32 %r815, %f760;
shfl.sync.bfly.b32 %r816|%p234, %r815, %r772, %r767, %r769;
mov.b32 %f761, %r816;
add.ftz.f32 %f762, %f760, %f761;
mov.b32 %r817, %f762;
shfl.sync.bfly.b32 %r818|%p235, %r817, %r775, %r767, %r769;
mov.b32 %f763, %r818;
add.ftz.f32 %f764, %f762, %f763;
mov.b32 %r819, %f764;
shfl.sync.bfly.b32 %r820|%p236, %r819, %r778, %r767, %r769;
mov.b32 %f765, %r820;
add.ftz.f32 %f766, %f764, %f765;
mov.b32 %r821, %f766;
shfl.sync.bfly.b32 %r822|%p237, %r821, %r781, %r767, %r769;
mov.b32 %f767, %r822;
add.ftz.f32 %f67, %f766, %f767;
setp.ne.s32 %p238, %r2, 1;
or.b32 %r826, %r4, 1;
setp.ge.u32 %p239, %r826, %r44;
or.pred %p240, %p238, %p239;
@%p240 bra $L__BB28_66;

add.ftz.f32 %f768, %f5, %f66;
add.ftz.f32 %f769, %f4, %f67;
mul.ftz.f32 %f770, %f769, 0f3F000000;
mul.ftz.f32 %f771, %f769, 0f3F4C422A;
mul.ftz.f32 %f772, %f769, 0f3D372713;
fma.rn.ftz.f32 %f773, %f769, %f772, 0f3F800000;
mul.ftz.f32 %f774, %f771, %f773;
tanh.approx.f32 %f775, %f774;
add.ftz.f32 %f776, %f775, 0f3F800000;
mul.ftz.f32 %f777, %f770, %f776;
mul.ftz.f32 %f778, %f768, %f777;
st.global.f32 [%rd24+4], %f778;

$L__BB28_66:
ld.shared.f32 %f779, [%r25+256];
add.ftz.f32 %f780, %f779, %f28;
ld.shared.f32 %f781, [%r27+256];
add.ftz.f32 %f782, %f781, %f32;
ld.shared.f32 %f783, [%r25+768];
add.ftz.f32 %f784, %f783, %f780;
ld.shared.f32 %f785, [%r27+768];
add.ftz.f32 %f786, %f785, %f782;
ld.shared.f32 %f787, [%r25+1280];
add.ftz.f32 %f788, %f787, %f784;
ld.shared.f32 %f789, [%r27+1280];
add.ftz.f32 %f790, %f789, %f786;
mov.b32 %r827, %f788;
mov.u32 %r828, 31;
mov.u32 %r829, 16;
mov.u32 %r830, -1;
shfl.sync.bfly.b32 %r831|%p241, %r827, %r829, %r828, %r830;
mov.b32 %f791, %r831;
add.ftz.f32 %f792, %f788, %f791;
mov.b32 %r832, %f792;
mov.u32 %r833, 8;
shfl.sync.bfly.b32 %r834|%p242, %r832, %r833, %r828, %r830;
mov.b32 %f793, %r834;
add.ftz.f32 %f794, %f792, %f793;
mov.b32 %r835, %f794;
mov.u32 %r836, 4;
shfl.sync.bfly.b32 %r837|%p243, %r835, %r836, %r828, %r830;
mov.b32 %f795, %r837;
add.ftz.f32 %f796, %f794, %f795;
mov.b32 %r838, %f796;
mov.u32 %r839, 2;
shfl.sync.bfly.b32 %r840|%p244, %r838, %r839, %r828, %r830;
mov.b32 %f797, %r840;
add.ftz.f32 %f798, %f796, %f797;
mov.b32 %r841, %f798;
mov.u32 %r842, 1;
shfl.sync.bfly.b32 %r843|%p245, %r841, %r842, %r828, %r830;
mov.b32 %f799, %r843;
add.ftz.f32 %f68, %f798, %f799;
mov.b32 %r844, %f790;
shfl.sync.bfly.b32 %r845|%p246, %r844, %r829, %r828, %r830;
mov.b32 %f800, %r845;
add.ftz.f32 %f801, %f790, %f800;
mov.b32 %r846, %f801;
shfl.sync.bfly.b32 %r847|%p247, %r846, %r833, %r828, %r830;
mov.b32 %f802, %r847;
add.ftz.f32 %f803, %f801, %f802;
mov.b32 %r848, %f803;
shfl.sync.bfly.b32 %r849|%p248, %r848, %r836, %r828, %r830;
mov.b32 %f804, %r849;
add.ftz.f32 %f805, %f803, %f804;
mov.b32 %r850, %f805;
shfl.sync.bfly.b32 %r851|%p249, %r850, %r839, %r828, %r830;
mov.b32 %f806, %r851;
add.ftz.f32 %f807, %f805, %f806;
mov.b32 %r852, %f807;
shfl.sync.bfly.b32 %r853|%p250, %r852, %r842, %r828, %r830;
mov.b32 %f808, %r853;
add.ftz.f32 %f69, %f807, %f808;
setp.ne.s32 %p251, %r2, 2;
or.b32 %r857, %r4, 2;
setp.ge.u32 %p252, %r857, %r44;
or.pred %p253, %p251, %p252;
@%p253 bra $L__BB28_68;

add.ftz.f32 %f809, %f5, %f68;
add.ftz.f32 %f810, %f4, %f69;
mul.ftz.f32 %f811, %f810, 0f3F000000;
mul.ftz.f32 %f812, %f810, 0f3F4C422A;
mul.ftz.f32 %f813, %f810, 0f3D372713;
fma.rn.ftz.f32 %f814, %f810, %f813, 0f3F800000;
mul.ftz.f32 %f815, %f812, %f814;
tanh.approx.f32 %f816, %f815;
add.ftz.f32 %f817, %f816, 0f3F800000;
mul.ftz.f32 %f818, %f811, %f817;
mul.ftz.f32 %f819, %f809, %f818;
st.global.f32 [%rd24+8], %f819;

$L__BB28_68:
ld.shared.f32 %f820, [%r25+384];
add.ftz.f32 %f821, %f820, %f27;
ld.shared.f32 %f822, [%r27+384];
add.ftz.f32 %f823, %f822, %f31;
ld.shared.f32 %f824, [%r25+896];
add.ftz.f32 %f825, %f824, %f821;
ld.shared.f32 %f826, [%r27+896];
add.ftz.f32 %f827, %f826, %f823;
ld.shared.f32 %f828, [%r25+1408];
add.ftz.f32 %f829, %f828, %f825;
ld.shared.f32 %f830, [%r27+1408];
add.ftz.f32 %f831, %f830, %f827;
mov.b32 %r858, %f829;
shfl.sync.bfly.b32 %r862|%p254, %r858, %r829, %r828, %r830;
mov.b32 %f832, %r862;
add.ftz.f32 %f833, %f829, %f832;
mov.b32 %r863, %f833;
shfl.sync.bfly.b32 %r865|%p255, %r863, %r833, %r828, %r830;
mov.b32 %f834, %r865;
add.ftz.f32 %f835, %f833, %f834;
mov.b32 %r866, %f835;
shfl.sync.bfly.b32 %r868|%p256, %r866, %r836, %r828, %r830;
mov.b32 %f836, %r868;
add.ftz.f32 %f837, %f835, %f836;
mov.b32 %r869, %f837;
shfl.sync.bfly.b32 %r871|%p257, %r869, %r839, %r828, %r830;
mov.b32 %f838, %r871;
add.ftz.f32 %f839, %f837, %f838;
mov.b32 %r872, %f839;
shfl.sync.bfly.b32 %r874|%p258, %r872, %r842, %r828, %r830;
mov.b32 %f840, %r874;
add.ftz.f32 %f70, %f839, %f840;
mov.b32 %r875, %f831;
shfl.sync.bfly.b32 %r876|%p259, %r875, %r829, %r828, %r830;
mov.b32 %f841, %r876;
add.ftz.f32 %f842, %f831, %f841;
mov.b32 %r877, %f842;
shfl.sync.bfly.b32 %r878|%p260, %r877, %r833, %r828, %r830;
mov.b32 %f843, %r878;
add.ftz.f32 %f844, %f842, %f843;
mov.b32 %r879, %f844;
shfl.sync.bfly.b32 %r880|%p261, %r879, %r836, %r828, %r830;
mov.b32 %f845, %r880;
add.ftz.f32 %f846, %f844, %f845;
mov.b32 %r881, %f846;
shfl.sync.bfly.b32 %r882|%p262, %r881, %r839, %r828, %r830;
mov.b32 %f847, %r882;
add.ftz.f32 %f848, %f846, %f847;
mov.b32 %r883, %f848;
shfl.sync.bfly.b32 %r884|%p263, %r883, %r842, %r828, %r830;
mov.b32 %f849, %r884;
add.ftz.f32 %f71, %f848, %f849;
setp.ne.s32 %p264, %r2, 3;
or.b32 %r888, %r4, 3;
setp.ge.u32 %p265, %r888, %r44;
or.pred %p266, %p264, %p265;
@%p266 bra $L__BB28_78;

add.ftz.f32 %f850, %f5, %f70;
add.ftz.f32 %f851, %f4, %f71;
mul.ftz.f32 %f852, %f851, 0f3F000000;
mul.ftz.f32 %f853, %f851, 0f3F4C422A;
mul.ftz.f32 %f854, %f851, 0f3D372713;
fma.rn.ftz.f32 %f855, %f851, %f854, 0f3F800000;
mul.ftz.f32 %f856, %f853, %f855;
tanh.approx.f32 %f857, %f856;
add.ftz.f32 %f858, %f857, 0f3F800000;
mul.ftz.f32 %f859, %f852, %f858;
mul.ftz.f32 %f860, %f850, %f859;
st.global.f32 [%rd24+12], %f860;
bra.uni $L__BB28_78;

$L__BB28_54:
ld.shared.f32 %f373, [%r25];
add.ftz.f32 %f374, %f373, %f30;
ld.shared.f32 %f375, [%r27];
add.ftz.f32 %f376, %f375, %f34;
ld.shared.f32 %f377, [%r25+512];
add.ftz.f32 %f378, %f377, %f374;
ld.shared.f32 %f379, [%r27+512];
add.ftz.f32 %f380, %f379, %f376;
ld.shared.f32 %f381, [%r25+1024];
add.ftz.f32 %f382, %f381, %f378;
ld.shared.f32 %f383, [%r27+1024];
add.ftz.f32 %f384, %f383, %f380;
mov.b32 %r520, %f382;
mov.u32 %r521, 31;
mov.u32 %r522, 16;
mov.u32 %r523, -1;
shfl.sync.bfly.b32 %r524|%p111, %r520, %r522, %r521, %r523;
mov.b32 %f385, %r524;
add.ftz.f32 %f386, %f382, %f385;
mov.b32 %r525, %f386;
mov.u32 %r526, 8;
shfl.sync.bfly.b32 %r527|%p112, %r525, %r526, %r521, %r523;
mov.b32 %f387, %r527;
add.ftz.f32 %f388, %f386, %f387;
mov.b32 %r528, %f388;
mov.u32 %r529, 4;
shfl.sync.bfly.b32 %r530|%p113, %r528, %r529, %r521, %r523;
mov.b32 %f389, %r530;
add.ftz.f32 %f390, %f388, %f389;
mov.b32 %r531, %f390;
mov.u32 %r532, 2;
shfl.sync.bfly.b32 %r533|%p114, %r531, %r532, %r521, %r523;
mov.b32 %f391, %r533;
add.ftz.f32 %f392, %f390, %f391;
mov.b32 %r534, %f392;
mov.u32 %r535, 1;
shfl.sync.bfly.b32 %r536|%p115, %r534, %r535, %r521, %r523;
mov.b32 %f393, %r536;
add.ftz.f32 %f48, %f392, %f393;
mov.b32 %r537, %f384;
shfl.sync.bfly.b32 %r538|%p116, %r537, %r522, %r521, %r523;
mov.b32 %f394, %r538;
add.ftz.f32 %f395, %f384, %f394;
mov.b32 %r539, %f395;
shfl.sync.bfly.b32 %r540|%p117, %r539, %r526, %r521, %r523;
mov.b32 %f396, %r540;
add.ftz.f32 %f397, %f395, %f396;
mov.b32 %r541, %f397;
shfl.sync.bfly.b32 %r542|%p118, %r541, %r529, %r521, %r523;
mov.b32 %f398, %r542;
add.ftz.f32 %f399, %f397, %f398;
mov.b32 %r543, %f399;
shfl.sync.bfly.b32 %r544|%p119, %r543, %r532, %r521, %r523;
mov.b32 %f400, %r544;
add.ftz.f32 %f401, %f399, %f400;
mov.b32 %r545, %f401;
shfl.sync.bfly.b32 %r546|%p120, %r545, %r535, %r521, %r523;
mov.b32 %f402, %r546;
add.ftz.f32 %f49, %f401, %f402;
setp.ne.s32 %p121, %r2, 0;
setp.ge.u32 %p122, %r4, %r44;
or.pred %p123, %p121, %p122;
@%p123 bra $L__BB28_56;

add.ftz.f32 %f403, %f5, %f48;
add.ftz.f32 %f404, %f4, %f49;
mov.f32 %f405, 0f40E00000;
min.ftz.f32 %f406, %f404, %f405;
min.ftz.f32 %f407, %f403, %f405;
mov.f32 %f408, 0fC0E00000;
max.ftz.f32 %f409, %f407, %f408;
mul.ftz.f32 %f410, %f406, 0f3FD9DB23;
mul.ftz.f32 %f411, %f410, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f412, %f411;
add.ftz.f32 %f413, %f412, 0f3F800000;
div.approx.ftz.f32 %f414, %f406, %f413;
add.ftz.f32 %f415, %f409, 0f3F800000;
mul.ftz.f32 %f416, %f415, %f414;
st.global.f32 [%rd24], %f416;

$L__BB28_56:
ld.shared.f32 %f417, [%r25+128];
add.ftz.f32 %f418, %f417, %f29;
ld.shared.f32 %f419, [%r27+128];
add.ftz.f32 %f420, %f419, %f33;
ld.shared.f32 %f421, [%r25+640];
add.ftz.f32 %f422, %f421, %f418;
ld.shared.f32 %f423, [%r27+640];
add.ftz.f32 %f424, %f423, %f420;
ld.shared.f32 %f425, [%r25+1152];
add.ftz.f32 %f426, %f425, %f422;
ld.shared.f32 %f427, [%r27+1152];
add.ftz.f32 %f428, %f427, %f424;
mov.b32 %r550, %f426;
shfl.sync.bfly.b32 %r554|%p124, %r550, %r522, %r521, %r523;
mov.b32 %f429, %r554;
add.ftz.f32 %f430, %f426, %f429;
mov.b32 %r555, %f430;
shfl.sync.bfly.b32 %r557|%p125, %r555, %r526, %r521, %r523;
mov.b32 %f431, %r557;
add.ftz.f32 %f432, %f430, %f431;
mov.b32 %r558, %f432;
shfl.sync.bfly.b32 %r560|%p126, %r558, %r529, %r521, %r523;
mov.b32 %f433, %r560;
add.ftz.f32 %f434, %f432, %f433;
mov.b32 %r561, %f434;
shfl.sync.bfly.b32 %r563|%p127, %r561, %r532, %r521, %r523;
mov.b32 %f435, %r563;
add.ftz.f32 %f436, %f434, %f435;
mov.b32 %r564, %f436;
shfl.sync.bfly.b32 %r566|%p128, %r564, %r535, %r521, %r523;
mov.b32 %f437, %r566;
add.ftz.f32 %f50, %f436, %f437;
mov.b32 %r567, %f428;
shfl.sync.bfly.b32 %r568|%p129, %r567, %r522, %r521, %r523;
mov.b32 %f438, %r568;
add.ftz.f32 %f439, %f428, %f438;
mov.b32 %r569, %f439;
shfl.sync.bfly.b32 %r570|%p130, %r569, %r526, %r521, %r523;
mov.b32 %f440, %r570;
add.ftz.f32 %f441, %f439, %f440;
mov.b32 %r571, %f441;
shfl.sync.bfly.b32 %r572|%p131, %r571, %r529, %r521, %r523;
mov.b32 %f442, %r572;
add.ftz.f32 %f443, %f441, %f442;
mov.b32 %r573, %f443;
shfl.sync.bfly.b32 %r574|%p132, %r573, %r532, %r521, %r523;
mov.b32 %f444, %r574;
add.ftz.f32 %f445, %f443, %f444;
mov.b32 %r575, %f445;
shfl.sync.bfly.b32 %r576|%p133, %r575, %r535, %r521, %r523;
mov.b32 %f446, %r576;
add.ftz.f32 %f51, %f445, %f446;
setp.ne.s32 %p134, %r2, 1;
or.b32 %r580, %r4, 1;
setp.ge.u32 %p135, %r580, %r44;
or.pred %p136, %p134, %p135;
@%p136 bra $L__BB28_58;

add.ftz.f32 %f447, %f5, %f50;
add.ftz.f32 %f448, %f4, %f51;
mov.f32 %f449, 0f40E00000;
min.ftz.f32 %f450, %f448, %f449;
min.ftz.f32 %f451, %f447, %f449;
mov.f32 %f452, 0fC0E00000;
max.ftz.f32 %f453, %f451, %f452;
mul.ftz.f32 %f454, %f450, 0f3FD9DB23;
mul.ftz.f32 %f455, %f454, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f456, %f455;
add.ftz.f32 %f457, %f456, 0f3F800000;
div.approx.ftz.f32 %f458, %f450, %f457;
add.ftz.f32 %f459, %f453, 0f3F800000;
mul.ftz.f32 %f460, %f459, %f458;
st.global.f32 [%rd24+4], %f460;

$L__BB28_58:
ld.shared.f32 %f461, [%r25+256];
add.ftz.f32 %f462, %f461, %f28;
ld.shared.f32 %f463, [%r27+256];
add.ftz.f32 %f464, %f463, %f32;
ld.shared.f32 %f465, [%r25+768];
add.ftz.f32 %f466, %f465, %f462;
ld.shared.f32 %f467, [%r27+768];
add.ftz.f32 %f468, %f467, %f464;
ld.shared.f32 %f469, [%r25+1280];
add.ftz.f32 %f470, %f469, %f466;
ld.shared.f32 %f471, [%r27+1280];
add.ftz.f32 %f472, %f471, %f468;
mov.b32 %r581, %f470;
mov.u32 %r582, 31;
mov.u32 %r583, 16;
mov.u32 %r584, -1;
shfl.sync.bfly.b32 %r585|%p137, %r581, %r583, %r582, %r584;
mov.b32 %f473, %r585;
add.ftz.f32 %f474, %f470, %f473;
mov.b32 %r586, %f474;
mov.u32 %r587, 8;
shfl.sync.bfly.b32 %r588|%p138, %r586, %r587, %r582, %r584;
mov.b32 %f475, %r588;
add.ftz.f32 %f476, %f474, %f475;
mov.b32 %r589, %f476;
mov.u32 %r590, 4;
shfl.sync.bfly.b32 %r591|%p139, %r589, %r590, %r582, %r584;
mov.b32 %f477, %r591;
add.ftz.f32 %f478, %f476, %f477;
mov.b32 %r592, %f478;
mov.u32 %r593, 2;
shfl.sync.bfly.b32 %r594|%p140, %r592, %r593, %r582, %r584;
mov.b32 %f479, %r594;
add.ftz.f32 %f480, %f478, %f479;
mov.b32 %r595, %f480;
mov.u32 %r596, 1;
shfl.sync.bfly.b32 %r597|%p141, %r595, %r596, %r582, %r584;
mov.b32 %f481, %r597;
add.ftz.f32 %f52, %f480, %f481;
mov.b32 %r598, %f472;
shfl.sync.bfly.b32 %r599|%p142, %r598, %r583, %r582, %r584;
mov.b32 %f482, %r599;
add.ftz.f32 %f483, %f472, %f482;
mov.b32 %r600, %f483;
shfl.sync.bfly.b32 %r601|%p143, %r600, %r587, %r582, %r584;
mov.b32 %f484, %r601;
add.ftz.f32 %f485, %f483, %f484;
mov.b32 %r602, %f485;
shfl.sync.bfly.b32 %r603|%p144, %r602, %r590, %r582, %r584;
mov.b32 %f486, %r603;
add.ftz.f32 %f487, %f485, %f486;
mov.b32 %r604, %f487;
shfl.sync.bfly.b32 %r605|%p145, %r604, %r593, %r582, %r584;
mov.b32 %f488, %r605;
add.ftz.f32 %f489, %f487, %f488;
mov.b32 %r606, %f489;
shfl.sync.bfly.b32 %r607|%p146, %r606, %r596, %r582, %r584;
mov.b32 %f490, %r607;
add.ftz.f32 %f53, %f489, %f490;
setp.ne.s32 %p147, %r2, 2;
or.b32 %r611, %r4, 2;
setp.ge.u32 %p148, %r611, %r44;
or.pred %p149, %p147, %p148;
@%p149 bra $L__BB28_60;

add.ftz.f32 %f491, %f5, %f52;
add.ftz.f32 %f492, %f4, %f53;
mov.f32 %f493, 0f40E00000;
min.ftz.f32 %f494, %f492, %f493;
min.ftz.f32 %f495, %f491, %f493;
mov.f32 %f496, 0fC0E00000;
max.ftz.f32 %f497, %f495, %f496;
mul.ftz.f32 %f498, %f494, 0f3FD9DB23;
mul.ftz.f32 %f499, %f498, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f500, %f499;
add.ftz.f32 %f501, %f500, 0f3F800000;
div.approx.ftz.f32 %f502, %f494, %f501;
add.ftz.f32 %f503, %f497, 0f3F800000;
mul.ftz.f32 %f504, %f503, %f502;
st.global.f32 [%rd24+8], %f504;

$L__BB28_60:
ld.shared.f32 %f505, [%r25+384];
add.ftz.f32 %f506, %f505, %f27;
ld.shared.f32 %f507, [%r27+384];
add.ftz.f32 %f508, %f507, %f31;
ld.shared.f32 %f509, [%r25+896];
add.ftz.f32 %f510, %f509, %f506;
ld.shared.f32 %f511, [%r27+896];
add.ftz.f32 %f512, %f511, %f508;
ld.shared.f32 %f513, [%r25+1408];
add.ftz.f32 %f514, %f513, %f510;
ld.shared.f32 %f515, [%r27+1408];
add.ftz.f32 %f516, %f515, %f512;
mov.b32 %r612, %f514;
shfl.sync.bfly.b32 %r616|%p150, %r612, %r583, %r582, %r584;
mov.b32 %f517, %r616;
add.ftz.f32 %f518, %f514, %f517;
mov.b32 %r617, %f518;
shfl.sync.bfly.b32 %r619|%p151, %r617, %r587, %r582, %r584;
mov.b32 %f519, %r619;
add.ftz.f32 %f520, %f518, %f519;
mov.b32 %r620, %f520;
shfl.sync.bfly.b32 %r622|%p152, %r620, %r590, %r582, %r584;
mov.b32 %f521, %r622;
add.ftz.f32 %f522, %f520, %f521;
mov.b32 %r623, %f522;
shfl.sync.bfly.b32 %r625|%p153, %r623, %r593, %r582, %r584;
mov.b32 %f523, %r625;
add.ftz.f32 %f524, %f522, %f523;
mov.b32 %r626, %f524;
shfl.sync.bfly.b32 %r628|%p154, %r626, %r596, %r582, %r584;
mov.b32 %f525, %r628;
add.ftz.f32 %f54, %f524, %f525;
mov.b32 %r629, %f516;
shfl.sync.bfly.b32 %r630|%p155, %r629, %r583, %r582, %r584;
mov.b32 %f526, %r630;
add.ftz.f32 %f527, %f516, %f526;
mov.b32 %r631, %f527;
shfl.sync.bfly.b32 %r632|%p156, %r631, %r587, %r582, %r584;
mov.b32 %f528, %r632;
add.ftz.f32 %f529, %f527, %f528;
mov.b32 %r633, %f529;
shfl.sync.bfly.b32 %r634|%p157, %r633, %r590, %r582, %r584;
mov.b32 %f530, %r634;
add.ftz.f32 %f531, %f529, %f530;
mov.b32 %r635, %f531;
shfl.sync.bfly.b32 %r636|%p158, %r635, %r593, %r582, %r584;
mov.b32 %f532, %r636;
add.ftz.f32 %f533, %f531, %f532;
mov.b32 %r637, %f533;
shfl.sync.bfly.b32 %r638|%p159, %r637, %r596, %r582, %r584;
mov.b32 %f534, %r638;
add.ftz.f32 %f55, %f533, %f534;
setp.ne.s32 %p160, %r2, 3;
or.b32 %r642, %r4, 3;
setp.ge.u32 %p161, %r642, %r44;
or.pred %p162, %p160, %p161;
@%p162 bra $L__BB28_78;

add.ftz.f32 %f535, %f5, %f54;
add.ftz.f32 %f536, %f4, %f55;
mov.f32 %f537, 0f40E00000;
min.ftz.f32 %f538, %f536, %f537;
min.ftz.f32 %f539, %f535, %f537;
mov.f32 %f540, 0fC0E00000;
max.ftz.f32 %f541, %f539, %f540;
mul.ftz.f32 %f542, %f538, 0f3FD9DB23;
mul.ftz.f32 %f543, %f542, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f544, %f543;
add.ftz.f32 %f545, %f544, 0f3F800000;
div.approx.ftz.f32 %f546, %f538, %f545;
add.ftz.f32 %f547, %f541, 0f3F800000;
mul.ftz.f32 %f548, %f547, %f546;
st.global.f32 [%rd24+12], %f548;
bra.uni $L__BB28_78;

}
.entry _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<38>;
.reg .b16 %rs<21>;
.reg .f32 %f<121>;
.reg .b32 %r<235>;
.reg .b64 %rd<44>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

ld.param.u64 %rd12, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd13, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd14, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r17, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r19, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r18, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
mov.u32 %r1, %tid.y;
shl.b32 %r35, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r35, %r2;
mov.u32 %r36, %ctaid.x;
shl.b32 %r4, %r36, 2;
mov.u32 %r233, %ctaid.y;
setp.eq.s64 %p1, %rd14, 0;
@%p1 bra $L__BB29_2;

cvta.to.global.u64 %rd16, %rd14;
mul.wide.u32 %rd17, %r233, 4;
add.s64 %rd18, %rd16, %rd17;
ld.global.nc.u32 %r232, [%rd18];
mul.hi.u32 %r37, %r233, %r18;
add.s32 %r38, %r37, %r233;
shr.u32 %r39, %r38, %r19;
mul.lo.s32 %r40, %r39, %r20;
sub.s32 %r233, %r233, %r40;
bra.uni $L__BB29_3;

$L__BB29_2:
mul.hi.u32 %r41, %r233, %r23;
add.s32 %r42, %r41, %r233;
shr.u32 %r232, %r42, %r24;

$L__BB29_3:
mov.u32 %r11, %ctaid.z;
shr.u32 %r43, %r3, 31;
add.s32 %r44, %r3, %r43;
shr.s32 %r234, %r44, 1;
shr.u32 %r45, %r17, 5;
setp.ge.s32 %p2, %r234, %r45;
mov.f32 %f117, 0f00000000;
mov.f32 %f118, %f117;
mov.f32 %f119, %f117;
mov.f32 %f120, %f117;
@%p2 bra $L__BB29_6;

and.b32 %r48, %r44, -2;
sub.s32 %r49, %r3, %r48;
shl.b32 %r50, %r49, 1;
shl.b32 %r51, %r49, 2;
or.b32 %r52, %r51, 1;
cvt.s64.s32 %rd1, %r50;
sub.s32 %r53, %r52, %r50;
cvt.s64.s32 %rd2, %r53;
mad.lo.s32 %r54, %r4, %r21, %r234;
mul.hi.u32 %r55, %r11, %r29;
add.s32 %r56, %r55, %r11;
shr.u32 %r57, %r56, %r30;
mad.lo.s32 %r58, %r57, %r32, %r54;
mad.lo.s32 %r59, %r232, %r26, %r58;
mul.wide.s32 %rd3, %r59, 18;
mul.wide.s32 %rd19, %r51, 2;
add.s64 %rd20, %rd3, %rd19;
add.s64 %rd4, %rd20, 4;
mul.wide.s32 %rd5, %r21, 18;
cvta.to.global.u64 %rd6, %rd13;
cvta.to.global.u64 %rd43, %rd12;
mul.lo.s32 %r60, %r233, %r27;
cvt.u64.u32 %rd21, %r60;
mul.lo.s32 %r61, %r11, %r33;
cvt.u64.u32 %rd22, %r61;
add.s64 %rd8, %rd21, %rd22;
mov.f32 %f117, 0f00000000;
shl.b64 %rd30, %rd1, 2;
shl.b64 %rd32, %rd2, 2;

$L__BB29_5:
cvt.u64.u32 %rd23, %r234;
add.s64 %rd24, %rd8, %rd23;
mul.lo.s64 %rd25, %rd24, 36;
add.s64 %rd26, %rd6, %rd25;
add.s64 %rd27, %rd26, 4;
add.s64 %rd28, %rd43, %rd4;
add.s64 %rd29, %rd28, -2;
ld.global.nc.u16 %rs5, [%rd28+-2];
ld.global.nc.u16 %rs6, [%rd28];
mov.b32 %r134, {%rs5, %rs6};
add.s64 %rd31, %rd27, %rd30;
ld.global.nc.u32 %r64, [%rd31];
ld.global.nc.u32 %r68, [%rd31+16];
ld.global.nc.u16 %rs7, [%rd28+2];
ld.global.nc.u16 %rs8, [%rd28+4];
mov.b32 %r135, {%rs7, %rs8};
add.s64 %rd33, %rd27, %rd32;
ld.global.nc.u32 %r72, [%rd33];
ld.global.nc.u32 %r76, [%rd31+20];
add.s64 %rd34, %rd43, %rd3;
ld.global.nc.u16 %rs1, [%rd34];

{ cvt.f32.f16 %f25, %rs1;}


and.b32 %r63, %r134, 252645135;
shr.u32 %r136, %r134, 4;
and.b32 %r67, %r136, 252645135;
mov.u32 %r119, 0;

dp4a.s32.s32 %r62, %r63, %r64, %r119;


dp4a.s32.s32 %r66, %r67, %r68, %r62;

and.b32 %r71, %r135, 252645135;
shr.u32 %r137, %r135, 4;
and.b32 %r75, %r137, 252645135;

dp4a.s32.s32 %r70, %r71, %r72, %r66;


dp4a.s32.s32 %r74, %r75, %r76, %r70;

ld.global.nc.u32 %r78, [%rd26];

{.reg .f16 low,high;
mov.b32 {low,high},%r78;
cvt.f32.f16 %f26, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r78;
cvt.f32.f16 %f27, high;}


cvt.rn.f32.s32 %f37, %r74;
mul.ftz.f32 %f38, %f26, %f37;
mul.ftz.f32 %f39, %f27, 0f40800000;
sub.ftz.f32 %f40, %f38, %f39;
fma.rn.ftz.f32 %f120, %f25, %f40, %f120;
add.s64 %rd35, %rd29, %rd5;
ld.global.nc.u16 %rs9, [%rd35];
ld.global.nc.u16 %rs10, [%rd35+2];
mov.b32 %r138, {%rs9, %rs10};
ld.global.nc.u16 %rs11, [%rd35+4];
ld.global.nc.u16 %rs12, [%rd35+6];
mov.b32 %r139, {%rs11, %rs12};
add.s64 %rd36, %rd34, %rd5;
ld.global.nc.u16 %rs2, [%rd36];

{ cvt.f32.f16 %f28, %rs2;}


and.b32 %r81, %r138, 252645135;
shr.u32 %r140, %r138, 4;
and.b32 %r85, %r140, 252645135;

dp4a.s32.s32 %r80, %r81, %r64, %r119;


dp4a.s32.s32 %r84, %r85, %r68, %r80;

and.b32 %r89, %r139, 252645135;
shr.u32 %r141, %r139, 4;
and.b32 %r93, %r141, 252645135;

dp4a.s32.s32 %r88, %r89, %r72, %r84;


dp4a.s32.s32 %r92, %r93, %r76, %r88;


{.reg .f16 low,high;
mov.b32 {low,high},%r78;
cvt.f32.f16 %f29, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r78;
cvt.f32.f16 %f30, high;}


cvt.rn.f32.s32 %f41, %r92;
mul.ftz.f32 %f42, %f29, %f41;
mul.ftz.f32 %f43, %f30, 0f40800000;
sub.ftz.f32 %f44, %f42, %f43;
fma.rn.ftz.f32 %f119, %f28, %f44, %f119;
add.s64 %rd37, %rd35, %rd5;
ld.global.nc.u16 %rs13, [%rd37];
ld.global.nc.u16 %rs14, [%rd37+2];
mov.b32 %r142, {%rs13, %rs14};
ld.global.nc.u16 %rs15, [%rd37+4];
ld.global.nc.u16 %rs16, [%rd37+6];
mov.b32 %r143, {%rs15, %rs16};
add.s64 %rd38, %rd36, %rd5;
ld.global.nc.u16 %rs3, [%rd38];

{ cvt.f32.f16 %f31, %rs3;}


and.b32 %r99, %r142, 252645135;
shr.u32 %r144, %r142, 4;
and.b32 %r103, %r144, 252645135;

dp4a.s32.s32 %r98, %r99, %r64, %r119;


dp4a.s32.s32 %r102, %r103, %r68, %r98;

and.b32 %r107, %r143, 252645135;
shr.u32 %r145, %r143, 4;
and.b32 %r111, %r145, 252645135;

dp4a.s32.s32 %r106, %r107, %r72, %r102;


dp4a.s32.s32 %r110, %r111, %r76, %r106;


{.reg .f16 low,high;
mov.b32 {low,high},%r78;
cvt.f32.f16 %f32, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r78;
cvt.f32.f16 %f33, high;}


cvt.rn.f32.s32 %f45, %r110;
mul.ftz.f32 %f46, %f32, %f45;
mul.ftz.f32 %f47, %f33, 0f40800000;
sub.ftz.f32 %f48, %f46, %f47;
fma.rn.ftz.f32 %f118, %f31, %f48, %f118;
add.s64 %rd39, %rd37, %rd5;
ld.global.nc.u16 %rs17, [%rd39];
ld.global.nc.u16 %rs18, [%rd39+2];
mov.b32 %r146, {%rs17, %rs18};
ld.global.nc.u16 %rs19, [%rd39+4];
ld.global.nc.u16 %rs20, [%rd39+6];
mov.b32 %r147, {%rs19, %rs20};
add.s64 %rd40, %rd38, %rd5;
ld.global.nc.u16 %rs4, [%rd40];

{ cvt.f32.f16 %f34, %rs4;}


and.b32 %r117, %r146, 252645135;
shr.u32 %r148, %r146, 4;
and.b32 %r121, %r148, 252645135;

dp4a.s32.s32 %r116, %r117, %r64, %r119;


dp4a.s32.s32 %r120, %r121, %r68, %r116;

and.b32 %r125, %r147, 252645135;
shr.u32 %r149, %r147, 4;
and.b32 %r129, %r149, 252645135;

dp4a.s32.s32 %r124, %r125, %r72, %r120;


dp4a.s32.s32 %r128, %r129, %r76, %r124;


{.reg .f16 low,high;
mov.b32 {low,high},%r78;
cvt.f32.f16 %f35, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r78;
cvt.f32.f16 %f36, high;}


cvt.rn.f32.s32 %f49, %r128;
mul.ftz.f32 %f50, %f35, %f49;
mul.ftz.f32 %f51, %f36, 0f40800000;
sub.ftz.f32 %f52, %f50, %f51;
fma.rn.ftz.f32 %f117, %f34, %f52, %f117;
add.s64 %rd43, %rd43, 1152;
add.s32 %r234, %r234, 64;
setp.lt.s32 %p3, %r234, %r45;
@%p3 bra $L__BB29_5;

$L__BB29_6:
setp.eq.s32 %p4, %r1, 0;
@%p4 bra $L__BB29_8;

shl.b32 %r151, %r1, 9;
mov.u32 %r152, _ZZ13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r153, %r152, %r151;
shl.b32 %r154, %r2, 2;
add.s32 %r155, %r153, %r154;
st.shared.f32 [%r155+-512], %f120;
st.shared.f32 [%r155+-384], %f119;
st.shared.f32 [%r155+-256], %f118;
st.shared.f32 [%r155+-128], %f117;

$L__BB29_8:
setp.ne.s32 %p5, %r1, 0;
bar.sync 0;
@%p5 bra $L__BB29_17;

mov.u32 %r156, %ctaid.y;
mad.lo.s32 %r157, %r156, %r28, %r4;
mad.lo.s32 %r158, %r11, %r34, %r157;
shl.b32 %r159, %r2, 2;
mov.u32 %r160, _ZZ13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r15, %r160, %r159;
mov.u32 %r161, 2;
ld.shared.f32 %f53, [%r15];
add.ftz.f32 %f54, %f53, %f120;
ld.shared.f32 %f55, [%r15+512];
add.ftz.f32 %f56, %f55, %f54;
ld.shared.f32 %f57, [%r15+1024];
add.ftz.f32 %f58, %f57, %f56;
mov.b32 %r162, %f58;
mov.u32 %r163, 31;
mov.u32 %r164, 16;
mov.u32 %r165, -1;
shfl.sync.bfly.b32 %r166|%p6, %r162, %r164, %r163, %r165;
mov.b32 %f59, %r166;
add.ftz.f32 %f60, %f58, %f59;
mov.b32 %r167, %f60;
mov.u32 %r168, 8;
shfl.sync.bfly.b32 %r169|%p7, %r167, %r168, %r163, %r165;
mov.b32 %f61, %r169;
add.ftz.f32 %f62, %f60, %f61;
mov.b32 %r170, %f62;
mov.u32 %r171, 4;
shfl.sync.bfly.b32 %r172|%p8, %r170, %r171, %r163, %r165;
mov.b32 %f63, %r172;
add.ftz.f32 %f64, %f62, %f63;
mov.b32 %r173, %f64;
shfl.sync.bfly.b32 %r174|%p9, %r173, %r161, %r163, %r165;
mov.b32 %f65, %r174;
add.ftz.f32 %f66, %f64, %f65;
mov.b32 %r175, %f66;
mov.u32 %r176, 1;
shfl.sync.bfly.b32 %r177|%p10, %r175, %r176, %r163, %r165;
mov.b32 %f67, %r177;
add.ftz.f32 %f13, %f66, %f67;
setp.ge.u32 %p11, %r4, %r22;
setp.ne.s32 %p12, %r2, 0;
cvta.to.global.u64 %rd41, %rd15;
mul.wide.u32 %rd42, %r158, 4;
add.s64 %rd11, %rd41, %rd42;
or.pred %p13, %p12, %p11;
@%p13 bra $L__BB29_11;

st.global.f32 [%rd11], %f13;

$L__BB29_11:
ld.shared.f32 %f68, [%r15+128];
add.ftz.f32 %f69, %f68, %f119;
ld.shared.f32 %f70, [%r15+640];
add.ftz.f32 %f71, %f70, %f69;
ld.shared.f32 %f72, [%r15+1152];
add.ftz.f32 %f73, %f72, %f71;
mov.b32 %r178, %f73;
shfl.sync.bfly.b32 %r182|%p14, %r178, %r164, %r163, %r165;
mov.b32 %f74, %r182;
add.ftz.f32 %f75, %f73, %f74;
mov.b32 %r183, %f75;
shfl.sync.bfly.b32 %r185|%p15, %r183, %r168, %r163, %r165;
mov.b32 %f76, %r185;
add.ftz.f32 %f77, %f75, %f76;
mov.b32 %r186, %f77;
shfl.sync.bfly.b32 %r188|%p16, %r186, %r171, %r163, %r165;
mov.b32 %f78, %r188;
add.ftz.f32 %f79, %f77, %f78;
mov.b32 %r189, %f79;
shfl.sync.bfly.b32 %r191|%p17, %r189, %r161, %r163, %r165;
mov.b32 %f80, %r191;
add.ftz.f32 %f81, %f79, %f80;
mov.b32 %r192, %f81;
shfl.sync.bfly.b32 %r194|%p18, %r192, %r176, %r163, %r165;
mov.b32 %f82, %r194;
add.ftz.f32 %f14, %f81, %f82;
add.s32 %r195, %r4, 1;
setp.ge.u32 %p19, %r195, %r22;
setp.ne.s32 %p20, %r2, 1;
or.pred %p21, %p20, %p19;
@%p21 bra $L__BB29_13;

st.global.f32 [%rd11+4], %f14;

$L__BB29_13:
ld.shared.f32 %f83, [%r15+256];
add.ftz.f32 %f84, %f83, %f118;
ld.shared.f32 %f85, [%r15+768];
add.ftz.f32 %f86, %f85, %f84;
ld.shared.f32 %f87, [%r15+1280];
add.ftz.f32 %f88, %f87, %f86;
mov.b32 %r196, %f88;
mov.u32 %r197, 31;
mov.u32 %r198, 16;
mov.u32 %r199, -1;
shfl.sync.bfly.b32 %r200|%p22, %r196, %r198, %r197, %r199;
mov.b32 %f89, %r200;
add.ftz.f32 %f90, %f88, %f89;
mov.b32 %r201, %f90;
mov.u32 %r202, 8;
shfl.sync.bfly.b32 %r203|%p23, %r201, %r202, %r197, %r199;
mov.b32 %f91, %r203;
add.ftz.f32 %f92, %f90, %f91;
mov.b32 %r204, %f92;
mov.u32 %r205, 4;
shfl.sync.bfly.b32 %r206|%p24, %r204, %r205, %r197, %r199;
mov.b32 %f93, %r206;
add.ftz.f32 %f94, %f92, %f93;
mov.b32 %r207, %f94;
mov.u32 %r208, 2;
shfl.sync.bfly.b32 %r209|%p25, %r207, %r208, %r197, %r199;
mov.b32 %f95, %r209;
add.ftz.f32 %f96, %f94, %f95;
mov.b32 %r210, %f96;
mov.u32 %r211, 1;
shfl.sync.bfly.b32 %r212|%p26, %r210, %r211, %r197, %r199;
mov.b32 %f97, %r212;
add.ftz.f32 %f15, %f96, %f97;
add.s32 %r213, %r4, 2;
setp.ge.u32 %p27, %r213, %r22;
setp.ne.s32 %p28, %r2, 2;
or.pred %p29, %p28, %p27;
@%p29 bra $L__BB29_15;

st.global.f32 [%rd11+8], %f15;

$L__BB29_15:
ld.shared.f32 %f98, [%r15+384];
add.ftz.f32 %f99, %f98, %f117;
ld.shared.f32 %f100, [%r15+896];
add.ftz.f32 %f101, %f100, %f99;
ld.shared.f32 %f102, [%r15+1408];
add.ftz.f32 %f103, %f102, %f101;
mov.b32 %r214, %f103;
shfl.sync.bfly.b32 %r218|%p30, %r214, %r198, %r197, %r199;
mov.b32 %f104, %r218;
add.ftz.f32 %f105, %f103, %f104;
mov.b32 %r219, %f105;
shfl.sync.bfly.b32 %r221|%p31, %r219, %r202, %r197, %r199;
mov.b32 %f106, %r221;
add.ftz.f32 %f107, %f105, %f106;
mov.b32 %r222, %f107;
shfl.sync.bfly.b32 %r224|%p32, %r222, %r205, %r197, %r199;
mov.b32 %f108, %r224;
add.ftz.f32 %f109, %f107, %f108;
mov.b32 %r225, %f109;
shfl.sync.bfly.b32 %r227|%p33, %r225, %r208, %r197, %r199;
mov.b32 %f110, %r227;
add.ftz.f32 %f111, %f109, %f110;
mov.b32 %r228, %f111;
shfl.sync.bfly.b32 %r230|%p34, %r228, %r211, %r197, %r199;
mov.b32 %f112, %r230;
add.ftz.f32 %f16, %f111, %f112;
add.s32 %r231, %r4, 3;
setp.ge.u32 %p35, %r231, %r22;
setp.ne.s32 %p36, %r2, 3;
or.pred %p37, %p36, %p35;
@%p37 bra $L__BB29_17;

st.global.f32 [%rd11+12], %f16;

$L__BB29_17:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<33>;
.reg .b16 %rs<11>;
.reg .f32 %f<124>;
.reg .b32 %r<187>;
.reg .b64 %rd<69>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[384];

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate[384];

ld.param.u64 %rd11, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd12, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd13, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
ld.param.f32 %f31, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+16];
ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+8];
ld.param.u64 %rd14, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3];
mov.u32 %r1, %tid.y;
shl.b32 %r43, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r43, %r2;
shr.u32 %r4, %r42, 5;
mov.u32 %r5, %ctaid.y;
setp.eq.s64 %p2, %rd13, 0;
@%p2 bra $L__BB30_2;

cvta.to.global.u64 %rd20, %rd13;
mul.wide.u32 %rd21, %r5, 4;
add.s64 %rd22, %rd20, %rd21;
ld.global.nc.u32 %r184, [%rd22];
mul.hi.u32 %r44, %r5, %r26;
add.s32 %r45, %r44, %r5;
shr.u32 %r46, %r45, %r27;
mul.lo.s32 %r47, %r46, %r28;
sub.s32 %r185, %r5, %r47;
bra.uni $L__BB30_3;

$L__BB30_2:
mul.hi.u32 %r48, %r5, %r30;
add.s32 %r49, %r48, %r5;
shr.u32 %r184, %r49, %r31;
mov.u32 %r185, %r5;

$L__BB30_3:
mov.u32 %r11, %ctaid.z;
selp.b32 %r12, %r5, %r184, %p2;
or.b32 %r50, %r2, %r1;
setp.ne.s32 %p4, %r50, 0;
mov.f32 %f117, 0f00000000;
mov.f32 %f113, %f117;
mov.f32 %f114, %f117;
@%p4 bra $L__BB30_8;

setp.eq.s64 %p5, %rd14, 0;
mov.f32 %f113, 0f00000000;
mov.f32 %f114, %f113;
@%p5 bra $L__BB30_6;

cvta.to.global.u64 %rd23, %rd14;
mul.lo.s32 %r51, %r11, %r41;
cvt.u64.u32 %rd24, %r51;
mul.lo.s32 %r52, %r12, %r35;
cvt.u64.u32 %rd25, %r52;
mov.u32 %r53, %ctaid.x;
cvt.s64.s32 %rd26, %r53;
cvt.u64.u32 %rd27, %r2;
add.s64 %rd28, %rd27, %rd26;
add.s64 %rd29, %rd28, %rd24;
add.s64 %rd30, %rd29, %rd25;
shl.b64 %rd31, %rd30, 2;
add.s64 %rd32, %rd23, %rd31;
ld.global.f32 %f114, [%rd32];

$L__BB30_6:
setp.eq.s64 %p6, %rd16, 0;
setp.eq.s64 %p7, %rd15, 0;
or.pred %p8, %p6, %p7;
@%p8 bra $L__BB30_8;

cvta.to.global.u64 %rd33, %rd16;
mul.lo.s32 %r55, %r11, %r41;
cvt.u64.u32 %rd34, %r55;
mul.lo.s32 %r56, %r12, %r35;
cvt.u64.u32 %rd35, %r56;
mov.u32 %r57, %ctaid.x;
cvt.s64.s32 %rd36, %r57;
cvt.u64.u32 %rd37, %r2;
add.s64 %rd38, %rd37, %rd36;
add.s64 %rd39, %rd38, %rd34;
add.s64 %rd40, %rd39, %rd35;
shl.b64 %rd41, %rd40, 2;
add.s64 %rd42, %rd33, %rd41;
ld.global.f32 %f113, [%rd42];

$L__BB30_8:
shr.u32 %r59, %r3, 31;
add.s32 %r60, %r3, %r59;
shr.s32 %r186, %r60, 1;
setp.ge.s32 %p9, %r186, %r4;
mov.f32 %f119, %f117;
@%p9 bra $L__BB30_13;

and.b32 %r63, %r60, -2;
sub.s32 %r14, %r3, %r63;
shl.b32 %r64, %r14, 2;
mul.hi.u32 %r65, %r11, %r36;
add.s32 %r66, %r65, %r11;
shr.u32 %r67, %r66, %r37;
mad.lo.s32 %r68, %r67, %r39, %r186;
mad.lo.s32 %r69, %r184, %r33, %r68;
mov.u32 %r70, %ctaid.x;
mad.lo.s32 %r71, %r70, %r29, %r69;
mul.wide.s32 %rd68, %r71, 18;
mul.wide.s32 %rd43, %r64, 2;
cvta.to.global.u64 %rd44, %rd11;
add.s64 %rd45, %rd44, %rd43;
add.s64 %rd5, %rd45, 4;
cvta.to.global.u64 %rd46, %rd15;
add.s64 %rd47, %rd46, %rd43;
add.s64 %rd6, %rd47, 2;
cvta.to.global.u64 %rd7, %rd12;
mul.lo.s32 %r72, %r185, %r34;
cvt.u64.u32 %rd48, %r72;
mul.lo.s32 %r73, %r11, %r40;
cvt.u64.u32 %rd49, %r73;
add.s64 %rd8, %rd48, %rd49;
mov.f32 %f119, 0f00000000;
mov.f32 %f117, %f119;

$L__BB30_10:
cvt.u64.u32 %rd50, %r186;
add.s64 %rd51, %rd8, %rd50;
mul.lo.s64 %rd52, %rd51, 36;
add.s64 %rd53, %rd7, %rd52;
add.s64 %rd54, %rd53, 4;
add.s64 %rd55, %rd5, %rd68;
ld.global.nc.u16 %rs2, [%rd55+-2];
ld.global.nc.u16 %rs3, [%rd55];
mov.b32 %r92, {%rs2, %rs3};
shl.b32 %r93, %r14, 1;
mul.wide.s32 %rd56, %r93, 4;
add.s64 %rd57, %rd54, %rd56;
ld.global.nc.u32 %r16, [%rd57];
ld.global.nc.u32 %r17, [%rd57+16];
ld.global.nc.u16 %rs4, [%rd55+2];
ld.global.nc.u16 %rs5, [%rd55+4];
mov.b32 %r94, {%rs4, %rs5};
or.b32 %r96, %r64, 1;
sub.s32 %r97, %r96, %r93;
mul.wide.s32 %rd58, %r97, 4;
add.s64 %rd59, %rd54, %rd58;
ld.global.nc.u32 %r18, [%rd59];
ld.global.nc.u32 %r19, [%rd57+20];
add.s64 %rd61, %rd44, %rd68;
ld.global.nc.u16 %rs1, [%rd61];

{ cvt.f32.f16 %f40, %rs1;}


and.b32 %r75, %r92, 252645135;
shr.u32 %r98, %r92, 4;
and.b32 %r79, %r98, 252645135;
mov.u32 %r77, 0;

dp4a.s32.s32 %r74, %r75, %r16, %r77;


dp4a.s32.s32 %r78, %r79, %r17, %r74;

and.b32 %r83, %r94, 252645135;
shr.u32 %r99, %r94, 4;
and.b32 %r87, %r99, 252645135;

dp4a.s32.s32 %r82, %r83, %r18, %r78;


dp4a.s32.s32 %r86, %r87, %r19, %r82;

ld.global.nc.u32 %r20, [%rd53];

{.reg .f16 low,high;
mov.b32 {low,high},%r20;
cvt.f32.f16 %f41, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r20;
cvt.f32.f16 %f42, high;}


cvt.rn.f32.s32 %f43, %r86;
mul.ftz.f32 %f44, %f42, 0fC0800000;
fma.rn.ftz.f32 %f45, %f41, %f43, %f44;
fma.rn.ftz.f32 %f119, %f40, %f45, %f119;
setp.eq.s64 %p10, %rd15, 0;
@%p10 bra $L__BB30_12;

add.s64 %rd62, %rd6, %rd68;
ld.global.u16 %rs7, [%rd62];
ld.global.u16 %rs8, [%rd62+2];
mov.b32 %r118, {%rs7, %rs8};
ld.global.u16 %rs9, [%rd62+4];
ld.global.u16 %rs10, [%rd62+6];
mov.b32 %r119, {%rs9, %rs10};
add.s64 %rd64, %rd46, %rd68;
ld.global.u16 %rs6, [%rd64];

{ cvt.f32.f16 %f46, %rs6;}


and.b32 %r101, %r118, 252645135;
shr.u32 %r120, %r118, 4;
and.b32 %r105, %r120, 252645135;

dp4a.s32.s32 %r100, %r101, %r16, %r77;


dp4a.s32.s32 %r104, %r105, %r17, %r100;

and.b32 %r109, %r119, 252645135;
shr.u32 %r121, %r119, 4;
and.b32 %r113, %r121, 252645135;

dp4a.s32.s32 %r108, %r109, %r18, %r104;


dp4a.s32.s32 %r112, %r113, %r19, %r108;


{.reg .f16 low,high;
mov.b32 {low,high},%r20;
cvt.f32.f16 %f47, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r20;
cvt.f32.f16 %f48, high;}


cvt.rn.f32.s32 %f49, %r112;
mul.ftz.f32 %f50, %f48, 0fC0800000;
fma.rn.ftz.f32 %f51, %f47, %f49, %f50;
fma.rn.ftz.f32 %f117, %f46, %f51, %f117;

$L__BB30_12:
add.s64 %rd68, %rd68, 1152;
add.s32 %r186, %r186, 64;
setp.lt.s32 %p11, %r186, %r4;
@%p11 bra $L__BB30_10;

$L__BB30_13:
setp.eq.s32 %p12, %r1, 0;
@%p12 bra $L__BB30_16;

setp.eq.s64 %p13, %rd15, 0;
add.s32 %r22, %r1, -1;
shl.b32 %r125, %r1, 7;
mov.u32 %r126, _ZZ13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r127, %r126, %r125;
shl.b32 %r128, %r2, 2;
add.s32 %r129, %r127, %r128;
st.shared.f32 [%r129+-128], %f119;
@%p13 bra $L__BB30_16;

shl.b32 %r131, %r22, 7;
mov.u32 %r132, _ZZ13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r133, %r132, %r131;
add.s32 %r135, %r133, %r128;
st.shared.f32 [%r135], %f117;

$L__BB30_16:
setp.ne.s32 %p14, %r1, 0;
bar.sync 0;
@%p14 bra $L__BB30_35;

setp.ne.s64 %p15, %rd15, 0;
shl.b32 %r138, %r2, 2;
mov.u32 %r139, _ZZ13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r23, %r139, %r138;
ld.shared.f32 %f52, [%r23];
add.ftz.f32 %f13, %f119, %f52;
@%p15 bra $L__BB30_19;
bra.uni $L__BB30_18;

$L__BB30_19:
mov.u32 %r142, _ZZ13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r143, %r142, %r138;
ld.shared.f32 %f56, [%r143];
add.ftz.f32 %f57, %f117, %f56;
ld.shared.f32 %f58, [%r23+128];
add.ftz.f32 %f59, %f13, %f58;
ld.shared.f32 %f60, [%r143+128];
add.ftz.f32 %f61, %f57, %f60;
ld.shared.f32 %f62, [%r23+256];
add.ftz.f32 %f121, %f59, %f62;
ld.shared.f32 %f63, [%r143+256];
add.ftz.f32 %f117, %f61, %f63;
bra.uni $L__BB30_20;

$L__BB30_18:
ld.shared.f32 %f53, [%r23+128];
add.ftz.f32 %f54, %f13, %f53;
ld.shared.f32 %f55, [%r23+256];
add.ftz.f32 %f121, %f54, %f55;

$L__BB30_20:
mov.b32 %r144, %f121;
mov.u32 %r145, 31;
mov.u32 %r146, 16;
mov.u32 %r147, -1;
shfl.sync.bfly.b32 %r148|%p16, %r144, %r146, %r145, %r147;
mov.b32 %f64, %r148;
add.ftz.f32 %f65, %f121, %f64;
mov.b32 %r149, %f65;
mov.u32 %r150, 8;
shfl.sync.bfly.b32 %r151|%p17, %r149, %r150, %r145, %r147;
mov.b32 %f66, %r151;
add.ftz.f32 %f67, %f65, %f66;
mov.b32 %r152, %f67;
mov.u32 %r153, 4;
shfl.sync.bfly.b32 %r154|%p18, %r152, %r153, %r145, %r147;
mov.b32 %f68, %r154;
add.ftz.f32 %f69, %f67, %f68;
mov.b32 %r155, %f69;
mov.u32 %r156, 2;
shfl.sync.bfly.b32 %r157|%p19, %r155, %r156, %r145, %r147;
mov.b32 %f70, %r157;
add.ftz.f32 %f19, %f69, %f70;
mov.b32 %r158, %f19;
mov.u32 %r159, 1;
shfl.sync.bfly.b32 %r24|%p1, %r158, %r159, %r145, %r147;
setp.eq.s64 %p20, %rd15, 0;
@%p20 bra $L__BB30_22;

mov.b32 %r160, %f117;
shfl.sync.bfly.b32 %r164|%p21, %r160, %r146, %r145, %r147;
mov.b32 %f71, %r164;
add.ftz.f32 %f72, %f117, %f71;
mov.b32 %r165, %f72;
shfl.sync.bfly.b32 %r167|%p22, %r165, %r150, %r145, %r147;
mov.b32 %f73, %r167;
add.ftz.f32 %f74, %f72, %f73;
mov.b32 %r168, %f74;
shfl.sync.bfly.b32 %r170|%p23, %r168, %r153, %r145, %r147;
mov.b32 %f75, %r170;
add.ftz.f32 %f76, %f74, %f75;
mov.b32 %r171, %f76;
shfl.sync.bfly.b32 %r173|%p24, %r171, %r156, %r145, %r147;
mov.b32 %f77, %r173;
add.ftz.f32 %f78, %f76, %f77;
mov.b32 %r174, %f78;
shfl.sync.bfly.b32 %r176|%p25, %r174, %r159, %r145, %r147;
mov.b32 %f79, %r176;
add.ftz.f32 %f117, %f78, %f79;

$L__BB30_22:
setp.ne.s32 %p26, %r2, 0;
mov.b32 %f80, %r24;
add.ftz.f32 %f22, %f19, %f80;
@%p26 bra $L__BB30_35;

add.ftz.f32 %f123, %f114, %f22;
@%p20 bra $L__BB30_34;

add.ftz.f32 %f24, %f113, %f117;
setp.gt.s32 %p28, %r25, 2;
@%p28 bra $L__BB30_28;

setp.eq.s32 %p31, %r25, 1;
@%p31 bra $L__BB30_32;

setp.eq.s32 %p32, %r25, 2;
@%p32 bra $L__BB30_27;
bra.uni $L__BB30_33;

$L__BB30_27:
mul.ftz.f32 %f100, %f24, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f101, %f100;
add.ftz.f32 %f102, %f101, 0f3F800000;
div.approx.ftz.f32 %f103, %f24, %f102;
mul.ftz.f32 %f123, %f123, %f103;
bra.uni $L__BB30_34;

$L__BB30_28:
setp.eq.s32 %p29, %r25, 3;
@%p29 bra $L__BB30_31;

setp.ne.s32 %p30, %r25, 6;
@%p30 bra $L__BB30_33;

min.ftz.f32 %f81, %f24, %f31;
neg.ftz.f32 %f82, %f31;
min.ftz.f32 %f83, %f123, %f31;
max.ftz.f32 %f84, %f83, %f82;
mul.ftz.f32 %f85, %f81, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f86, %f85;
add.ftz.f32 %f87, %f86, 0f3F800000;
div.approx.ftz.f32 %f88, %f81, %f87;
mul.ftz.f32 %f123, %f84, %f88;
bra.uni $L__BB30_34;

$L__BB30_32:
mul.ftz.f32 %f104, %f24, 0f3F000000;
mul.ftz.f32 %f105, %f24, 0f3D372713;
fma.rn.ftz.f32 %f106, %f24, %f105, 0f3F800000;
mul.ftz.f32 %f107, %f24, 0f3F4C422A;
mul.ftz.f32 %f108, %f107, %f106;
tanh.approx.f32 %f109, %f108;
add.ftz.f32 %f110, %f109, 0f3F800000;
mul.ftz.f32 %f111, %f104, %f110;
mul.ftz.f32 %f123, %f123, %f111;
bra.uni $L__BB30_34;

$L__BB30_31:
mov.f32 %f89, 0f40E00000;
min.ftz.f32 %f90, %f24, %f89;
min.ftz.f32 %f91, %f123, %f89;
mov.f32 %f92, 0fC0E00000;
max.ftz.f32 %f93, %f91, %f92;
mul.ftz.f32 %f94, %f90, 0f3FD9DB23;
mul.ftz.f32 %f95, %f94, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f96, %f95;
add.ftz.f32 %f97, %f96, 0f3F800000;
div.approx.ftz.f32 %f98, %f90, %f97;
add.ftz.f32 %f99, %f93, 0f3F800000;
mul.ftz.f32 %f123, %f99, %f98;
bra.uni $L__BB30_34;

$L__BB30_33:
mul.ftz.f32 %f123, %f123, %f24;

$L__BB30_34:
cvta.to.global.u64 %rd65, %rd19;
mov.u32 %r180, %ctaid.y;
mov.u32 %r181, %ctaid.x;
mad.lo.s32 %r182, %r180, %r35, %r181;
mad.lo.s32 %r183, %r11, %r41, %r182;
mul.wide.u32 %rd66, %r183, 4;
add.s64 %rd67, %rd65, %rd66;
st.global.f32 [%rd67], %f123;

$L__BB30_35:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<14>;
.reg .b16 %rs<16>;
.reg .f32 %f<51>;
.reg .b32 %r<172>;
.reg .b64 %rd<56>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[384];

ld.param.u64 %rd12, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd13, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd10, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd11, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
cvta.to.global.u64 %rd1, %rd13;
cvta.to.global.u64 %rd55, %rd12;
mov.u32 %r1, %tid.y;
shl.b32 %r43, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r43, %r2;
mov.u32 %r4, %ctaid.x;
shr.u32 %r5, %r26, 5;
mov.u32 %r6, %ctaid.y;
setp.eq.s64 %p1, %rd10, 0;
@%p1 bra $L__BB31_2;

cvta.to.global.u64 %rd14, %rd10;
mul.wide.u32 %rd15, %r6, 4;
add.s64 %rd16, %rd14, %rd15;
ld.global.nc.u32 %r168, [%rd16];
mul.hi.u32 %r44, %r6, %r27;
add.s32 %r45, %r44, %r6;
shr.u32 %r46, %r45, %r28;
mul.lo.s32 %r47, %r46, %r29;
sub.s32 %r169, %r6, %r47;
bra.uni $L__BB31_3;

$L__BB31_2:
mul.hi.u32 %r48, %r6, %r31;
add.s32 %r49, %r48, %r6;
shr.u32 %r168, %r49, %r32;
mov.u32 %r169, %r6;

$L__BB31_3:
mov.u32 %r12, %ctaid.z;
mul.hi.u32 %r50, %r12, %r37;
add.s32 %r51, %r50, %r12;
shr.u32 %r52, %r51, %r38;
mul.lo.s32 %r53, %r12, %r41;
cvt.u64.u32 %rd17, %r53;
mul.lo.s32 %r54, %r169, %r35;
cvt.u64.u32 %rd18, %r54;
add.s64 %rd3, %rd18, %rd17;
mul.lo.s32 %r13, %r52, %r40;
mul.lo.s32 %r14, %r168, %r34;
mul.lo.s32 %r15, %r4, %r30;
shr.u32 %r55, %r3, 31;
add.s32 %r56, %r3, %r55;
shr.s32 %r171, %r56, 1;
setp.ge.s32 %p2, %r171, %r5;
mov.f32 %f50, 0f00000000;
@%p2 bra $L__BB31_9;

and.b32 %r59, %r56, -2;
sub.s32 %r17, %r3, %r59;
shl.b32 %r18, %r17, 1;
not.b32 %r60, %r171;
add.s32 %r19, %r5, %r60;
and.b32 %r61, %r19, 64;
setp.ne.s32 %p3, %r61, 0;
mov.f32 %f50, 0f00000000;
@%p3 bra $L__BB31_6;

cvt.u64.u32 %rd19, %r171;
add.s64 %rd20, %rd3, %rd19;
add.s32 %r80, %r14, %r15;
add.s32 %r81, %r80, %r13;
add.s32 %r82, %r81, %r171;
mul.wide.s32 %rd21, %r82, 18;
add.s64 %rd22, %rd55, %rd21;
shl.b32 %r83, %r17, 2;
mul.wide.s32 %rd23, %r83, 2;
add.s64 %rd24, %rd22, %rd23;
mul.lo.s64 %rd25, %rd20, 36;
add.s64 %rd26, %rd1, %rd25;
add.s64 %rd27, %rd26, 4;
ld.global.nc.u16 %rs2, [%rd24+2];
or.b32 %r84, %r83, 1;
ld.global.nc.u16 %rs3, [%rd24+4];
mov.b32 %r85, {%rs2, %rs3};
mul.wide.s32 %rd28, %r18, 4;
add.s64 %rd29, %rd27, %rd28;
ld.global.nc.u32 %r64, [%rd29];
ld.global.nc.u32 %r68, [%rd29+16];
sub.s32 %r86, %r84, %r18;
ld.global.nc.u16 %rs4, [%rd24+6];
ld.global.nc.u16 %rs5, [%rd24+8];
mov.b32 %r87, {%rs4, %rs5};
mul.wide.s32 %rd30, %r86, 4;
add.s64 %rd31, %rd27, %rd30;
ld.global.nc.u32 %r72, [%rd31];
ld.global.nc.u32 %r76, [%rd29+20];
ld.global.nc.u16 %rs1, [%rd22];

{ cvt.f32.f16 %f11, %rs1;}


and.b32 %r63, %r85, 252645135;
shr.u32 %r88, %r85, 4;
and.b32 %r67, %r88, 252645135;
mov.u32 %r65, 0;

dp4a.s32.s32 %r62, %r63, %r64, %r65;


dp4a.s32.s32 %r66, %r67, %r68, %r62;

and.b32 %r71, %r87, 252645135;
shr.u32 %r89, %r87, 4;
and.b32 %r75, %r89, 252645135;

dp4a.s32.s32 %r70, %r71, %r72, %r66;


dp4a.s32.s32 %r74, %r75, %r76, %r70;

ld.global.nc.u32 %r78, [%rd26];

{.reg .f16 low,high;
mov.b32 {low,high},%r78;
cvt.f32.f16 %f12, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r78;
cvt.f32.f16 %f13, high;}


cvt.rn.f32.s32 %f14, %r74;
mul.ftz.f32 %f15, %f13, 0fC0800000;
fma.rn.ftz.f32 %f16, %f12, %f14, %f15;
fma.rn.ftz.f32 %f50, %f11, %f16, 0f00000000;
add.s32 %r171, %r171, 64;

$L__BB31_6:
and.b32 %r90, %r19, -64;
setp.eq.s32 %p4, %r90, 0;
@%p4 bra $L__BB31_9;

shl.b32 %r91, %r17, 2;
or.b32 %r92, %r91, 1;
cvt.s64.s32 %rd4, %r18;
sub.s32 %r93, %r92, %r18;
cvt.s64.s32 %rd5, %r93;
add.s32 %r94, %r171, %r13;
add.s32 %r95, %r94, %r14;
add.s32 %r96, %r95, %r15;
mul.wide.s32 %rd6, %r96, 18;
mul.wide.s32 %rd32, %r91, 2;
add.s64 %rd33, %rd6, %rd32;
add.s64 %rd7, %rd33, 2;
shl.b64 %rd40, %rd4, 2;
shl.b64 %rd42, %rd5, 2;

$L__BB31_8:
cvt.u64.u32 %rd34, %r171;
add.s64 %rd35, %rd3, %rd34;
mul.lo.s64 %rd36, %rd35, 36;
add.s64 %rd37, %rd1, %rd36;
add.s64 %rd38, %rd37, 4;
add.s64 %rd39, %rd55, %rd7;
ld.global.nc.u16 %rs8, [%rd39];
ld.global.nc.u16 %rs9, [%rd39+2];
mov.b32 %r133, {%rs8, %rs9};
add.s64 %rd41, %rd38, %rd40;
ld.global.nc.u32 %r99, [%rd41];
ld.global.nc.u32 %r103, [%rd41+16];
ld.global.nc.u16 %rs10, [%rd39+4];
ld.global.nc.u16 %rs11, [%rd39+6];
mov.b32 %r134, {%rs10, %rs11};
add.s64 %rd43, %rd38, %rd42;
ld.global.nc.u32 %r107, [%rd43];
ld.global.nc.u32 %r111, [%rd41+20];
add.s64 %rd44, %rd55, %rd6;
ld.global.nc.u16 %rs6, [%rd44];

{ cvt.f32.f16 %f17, %rs6;}


and.b32 %r98, %r133, 252645135;
shr.u32 %r135, %r133, 4;
and.b32 %r102, %r135, 252645135;
mov.u32 %r118, 0;

dp4a.s32.s32 %r97, %r98, %r99, %r118;


dp4a.s32.s32 %r101, %r102, %r103, %r97;

and.b32 %r106, %r134, 252645135;
shr.u32 %r136, %r134, 4;
and.b32 %r110, %r136, 252645135;

dp4a.s32.s32 %r105, %r106, %r107, %r101;


dp4a.s32.s32 %r109, %r110, %r111, %r105;

ld.global.nc.u32 %r113, [%rd37];

{.reg .f16 low,high;
mov.b32 {low,high},%r113;
cvt.f32.f16 %f18, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r113;
cvt.f32.f16 %f19, high;}


cvt.rn.f32.s32 %f23, %r109;
mul.ftz.f32 %f24, %f18, %f23;
mul.ftz.f32 %f25, %f19, 0f40800000;
sub.ftz.f32 %f26, %f24, %f25;
fma.rn.ftz.f32 %f27, %f17, %f26, %f50;
add.s32 %r137, %r171, 64;
cvt.u64.u32 %rd45, %r137;
add.s64 %rd46, %rd3, %rd45;
mul.lo.s64 %rd47, %rd46, 36;
add.s64 %rd48, %rd1, %rd47;
add.s64 %rd49, %rd48, 4;
ld.global.nc.u16 %rs12, [%rd39+1152];
ld.global.nc.u16 %rs13, [%rd39+1154];
mov.b32 %r138, {%rs12, %rs13};
add.s64 %rd50, %rd49, %rd40;
ld.global.nc.u32 %r117, [%rd50];
ld.global.nc.u32 %r121, [%rd50+16];
ld.global.nc.u16 %rs14, [%rd39+1156];
ld.global.nc.u16 %rs15, [%rd39+1158];
mov.b32 %r139, {%rs14, %rs15};
add.s64 %rd51, %rd49, %rd42;
ld.global.nc.u32 %r125, [%rd51];
ld.global.nc.u32 %r129, [%rd50+20];
ld.global.nc.u16 %rs7, [%rd44+1152];

{ cvt.f32.f16 %f20, %rs7;}


and.b32 %r116, %r138, 252645135;
shr.u32 %r140, %r138, 4;
and.b32 %r120, %r140, 252645135;

dp4a.s32.s32 %r115, %r116, %r117, %r118;


dp4a.s32.s32 %r119, %r120, %r121, %r115;

and.b32 %r124, %r139, 252645135;
shr.u32 %r141, %r139, 4;
and.b32 %r128, %r141, 252645135;

dp4a.s32.s32 %r123, %r124, %r125, %r119;


dp4a.s32.s32 %r127, %r128, %r129, %r123;

ld.global.nc.u32 %r131, [%rd48];

{.reg .f16 low,high;
mov.b32 {low,high},%r131;
cvt.f32.f16 %f21, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r131;
cvt.f32.f16 %f22, high;}


cvt.rn.f32.s32 %f28, %r127;
mul.ftz.f32 %f29, %f21, %f28;
mul.ftz.f32 %f30, %f22, 0f40800000;
sub.ftz.f32 %f31, %f29, %f30;
fma.rn.ftz.f32 %f50, %f20, %f31, %f27;
add.s64 %rd55, %rd55, 2304;
add.s32 %r171, %r171, 128;
setp.lt.s32 %p5, %r171, %r5;
@%p5 bra $L__BB31_8;

$L__BB31_9:
setp.eq.s32 %p6, %r1, 0;
@%p6 bra $L__BB31_11;

shl.b32 %r142, %r1, 7;
mov.u32 %r143, _ZZ13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r144, %r143, %r142;
shl.b32 %r145, %r2, 2;
add.s32 %r146, %r144, %r145;
st.shared.f32 [%r146+-128], %f50;

$L__BB31_11:
setp.ne.s32 %p7, %r1, 0;
bar.sync 0;
@%p7 bra $L__BB31_14;

mad.lo.s32 %r147, %r6, %r36, %r4;
mad.lo.s32 %r25, %r12, %r42, %r147;
shl.b32 %r148, %r2, 2;
mov.u32 %r149, _ZZ13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r150, %r149, %r148;
mov.u32 %r151, 2;
ld.shared.f32 %f32, [%r150];
add.ftz.f32 %f33, %f50, %f32;
ld.shared.f32 %f34, [%r150+128];
add.ftz.f32 %f35, %f33, %f34;
ld.shared.f32 %f36, [%r150+256];
add.ftz.f32 %f37, %f35, %f36;
mov.b32 %r152, %f37;
mov.u32 %r153, 31;
mov.u32 %r154, 16;
mov.u32 %r155, -1;
shfl.sync.bfly.b32 %r156|%p8, %r152, %r154, %r153, %r155;
mov.b32 %f38, %r156;
add.ftz.f32 %f39, %f37, %f38;
mov.b32 %r157, %f39;
mov.u32 %r158, 8;
shfl.sync.bfly.b32 %r159|%p9, %r157, %r158, %r153, %r155;
mov.b32 %f40, %r159;
add.ftz.f32 %f41, %f39, %f40;
mov.b32 %r160, %f41;
mov.u32 %r161, 4;
shfl.sync.bfly.b32 %r162|%p10, %r160, %r161, %r153, %r155;
mov.b32 %f42, %r162;
add.ftz.f32 %f43, %f41, %f42;
mov.b32 %r163, %f43;
shfl.sync.bfly.b32 %r164|%p11, %r163, %r151, %r153, %r155;
mov.b32 %f44, %r164;
add.ftz.f32 %f45, %f43, %f44;
mov.b32 %r165, %f45;
mov.u32 %r166, 1;
shfl.sync.bfly.b32 %r167|%p12, %r165, %r166, %r153, %r155;
mov.b32 %f46, %r167;
add.ftz.f32 %f7, %f45, %f46;
setp.ne.s32 %p13, %r2, 0;
@%p13 bra $L__BB31_14;

cvta.to.global.u64 %rd52, %rd11;
mul.wide.u32 %rd53, %r25, 4;
add.s64 %rd54, %rd52, %rd53;
st.global.f32 [%rd54], %f7;

$L__BB31_14:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<33>;
.reg .b16 %rs<11>;
.reg .f32 %f<124>;
.reg .b32 %r<187>;
.reg .b64 %rd<69>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[384];

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate[384];

ld.param.u64 %rd11, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd12, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd13, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
ld.param.f32 %f31, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+16];
ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+8];
ld.param.u64 %rd14, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3];
mov.u32 %r1, %tid.y;
shl.b32 %r43, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r43, %r2;
shr.u32 %r4, %r42, 5;
mov.u32 %r5, %ctaid.y;
setp.eq.s64 %p2, %rd13, 0;
@%p2 bra $L__BB32_2;

cvta.to.global.u64 %rd20, %rd13;
mul.wide.u32 %rd21, %r5, 4;
add.s64 %rd22, %rd20, %rd21;
ld.global.nc.u32 %r184, [%rd22];
mul.hi.u32 %r44, %r5, %r26;
add.s32 %r45, %r44, %r5;
shr.u32 %r46, %r45, %r27;
mul.lo.s32 %r47, %r46, %r28;
sub.s32 %r185, %r5, %r47;
bra.uni $L__BB32_3;

$L__BB32_2:
mul.hi.u32 %r48, %r5, %r30;
add.s32 %r49, %r48, %r5;
shr.u32 %r184, %r49, %r31;
mov.u32 %r185, %r5;

$L__BB32_3:
mov.u32 %r11, %ctaid.z;
selp.b32 %r12, %r5, %r184, %p2;
or.b32 %r50, %r2, %r1;
setp.ne.s32 %p4, %r50, 0;
mov.f32 %f117, 0f00000000;
mov.f32 %f113, %f117;
mov.f32 %f114, %f117;
@%p4 bra $L__BB32_8;

setp.eq.s64 %p5, %rd14, 0;
mov.f32 %f113, 0f00000000;
mov.f32 %f114, %f113;
@%p5 bra $L__BB32_6;

cvta.to.global.u64 %rd23, %rd14;
mul.lo.s32 %r51, %r11, %r41;
cvt.u64.u32 %rd24, %r51;
mul.lo.s32 %r52, %r12, %r35;
cvt.u64.u32 %rd25, %r52;
mov.u32 %r53, %ctaid.x;
cvt.s64.s32 %rd26, %r53;
cvt.u64.u32 %rd27, %r2;
add.s64 %rd28, %rd27, %rd26;
add.s64 %rd29, %rd28, %rd24;
add.s64 %rd30, %rd29, %rd25;
shl.b64 %rd31, %rd30, 2;
add.s64 %rd32, %rd23, %rd31;
ld.global.f32 %f114, [%rd32];

$L__BB32_6:
setp.eq.s64 %p6, %rd16, 0;
setp.eq.s64 %p7, %rd15, 0;
or.pred %p8, %p6, %p7;
@%p8 bra $L__BB32_8;

cvta.to.global.u64 %rd33, %rd16;
mul.lo.s32 %r55, %r11, %r41;
cvt.u64.u32 %rd34, %r55;
mul.lo.s32 %r56, %r12, %r35;
cvt.u64.u32 %rd35, %r56;
mov.u32 %r57, %ctaid.x;
cvt.s64.s32 %rd36, %r57;
cvt.u64.u32 %rd37, %r2;
add.s64 %rd38, %rd37, %rd36;
add.s64 %rd39, %rd38, %rd34;
add.s64 %rd40, %rd39, %rd35;
shl.b64 %rd41, %rd40, 2;
add.s64 %rd42, %rd33, %rd41;
ld.global.f32 %f113, [%rd42];

$L__BB32_8:
shr.u32 %r59, %r3, 31;
add.s32 %r60, %r3, %r59;
shr.s32 %r186, %r60, 1;
setp.ge.s32 %p9, %r186, %r4;
mov.f32 %f119, %f117;
@%p9 bra $L__BB32_13;

and.b32 %r63, %r60, -2;
sub.s32 %r14, %r3, %r63;
shl.b32 %r64, %r14, 2;
mul.hi.u32 %r65, %r11, %r36;
add.s32 %r66, %r65, %r11;
shr.u32 %r67, %r66, %r37;
mad.lo.s32 %r68, %r67, %r39, %r186;
mad.lo.s32 %r69, %r184, %r33, %r68;
mov.u32 %r70, %ctaid.x;
mad.lo.s32 %r71, %r70, %r29, %r69;
mul.wide.s32 %rd68, %r71, 18;
mul.wide.s32 %rd43, %r64, 2;
cvta.to.global.u64 %rd44, %rd11;
add.s64 %rd45, %rd44, %rd43;
add.s64 %rd5, %rd45, 4;
cvta.to.global.u64 %rd46, %rd15;
add.s64 %rd47, %rd46, %rd43;
add.s64 %rd6, %rd47, 2;
cvta.to.global.u64 %rd7, %rd12;
mul.lo.s32 %r72, %r185, %r34;
cvt.u64.u32 %rd48, %r72;
mul.lo.s32 %r73, %r11, %r40;
cvt.u64.u32 %rd49, %r73;
add.s64 %rd8, %rd48, %rd49;
mov.f32 %f119, 0f00000000;
mov.f32 %f117, %f119;

$L__BB32_10:
cvt.u64.u32 %rd50, %r186;
add.s64 %rd51, %rd8, %rd50;
mul.lo.s64 %rd52, %rd51, 36;
add.s64 %rd53, %rd7, %rd52;
add.s64 %rd54, %rd53, 4;
add.s64 %rd55, %rd5, %rd68;
ld.global.nc.u16 %rs2, [%rd55+-2];
ld.global.nc.u16 %rs3, [%rd55];
mov.b32 %r92, {%rs2, %rs3};
shl.b32 %r93, %r14, 1;
mul.wide.s32 %rd56, %r93, 4;
add.s64 %rd57, %rd54, %rd56;
ld.global.nc.u32 %r16, [%rd57];
ld.global.nc.u32 %r17, [%rd57+16];
ld.global.nc.u16 %rs4, [%rd55+2];
ld.global.nc.u16 %rs5, [%rd55+4];
mov.b32 %r94, {%rs4, %rs5};
or.b32 %r96, %r64, 1;
sub.s32 %r97, %r96, %r93;
mul.wide.s32 %rd58, %r97, 4;
add.s64 %rd59, %rd54, %rd58;
ld.global.nc.u32 %r18, [%rd59];
ld.global.nc.u32 %r19, [%rd57+20];
add.s64 %rd61, %rd44, %rd68;
ld.global.nc.u16 %rs1, [%rd61];

{ cvt.f32.f16 %f40, %rs1;}


and.b32 %r75, %r92, 252645135;
shr.u32 %r98, %r92, 4;
and.b32 %r79, %r98, 252645135;
mov.u32 %r77, 0;

dp4a.s32.s32 %r74, %r75, %r16, %r77;


dp4a.s32.s32 %r78, %r79, %r17, %r74;

and.b32 %r83, %r94, 252645135;
shr.u32 %r99, %r94, 4;
and.b32 %r87, %r99, 252645135;

dp4a.s32.s32 %r82, %r83, %r18, %r78;


dp4a.s32.s32 %r86, %r87, %r19, %r82;

ld.global.nc.u32 %r20, [%rd53];

{.reg .f16 low,high;
mov.b32 {low,high},%r20;
cvt.f32.f16 %f41, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r20;
cvt.f32.f16 %f42, high;}


cvt.rn.f32.s32 %f43, %r86;
mul.ftz.f32 %f44, %f42, 0fC0800000;
fma.rn.ftz.f32 %f45, %f41, %f43, %f44;
fma.rn.ftz.f32 %f119, %f40, %f45, %f119;
setp.eq.s64 %p10, %rd15, 0;
@%p10 bra $L__BB32_12;

add.s64 %rd62, %rd6, %rd68;
ld.global.u16 %rs7, [%rd62];
ld.global.u16 %rs8, [%rd62+2];
mov.b32 %r118, {%rs7, %rs8};
ld.global.u16 %rs9, [%rd62+4];
ld.global.u16 %rs10, [%rd62+6];
mov.b32 %r119, {%rs9, %rs10};
add.s64 %rd64, %rd46, %rd68;
ld.global.u16 %rs6, [%rd64];

{ cvt.f32.f16 %f46, %rs6;}


and.b32 %r101, %r118, 252645135;
shr.u32 %r120, %r118, 4;
and.b32 %r105, %r120, 252645135;

dp4a.s32.s32 %r100, %r101, %r16, %r77;


dp4a.s32.s32 %r104, %r105, %r17, %r100;

and.b32 %r109, %r119, 252645135;
shr.u32 %r121, %r119, 4;
and.b32 %r113, %r121, 252645135;

dp4a.s32.s32 %r108, %r109, %r18, %r104;


dp4a.s32.s32 %r112, %r113, %r19, %r108;


{.reg .f16 low,high;
mov.b32 {low,high},%r20;
cvt.f32.f16 %f47, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r20;
cvt.f32.f16 %f48, high;}


cvt.rn.f32.s32 %f49, %r112;
mul.ftz.f32 %f50, %f48, 0fC0800000;
fma.rn.ftz.f32 %f51, %f47, %f49, %f50;
fma.rn.ftz.f32 %f117, %f46, %f51, %f117;

$L__BB32_12:
add.s64 %rd68, %rd68, 1152;
add.s32 %r186, %r186, 64;
setp.lt.s32 %p11, %r186, %r4;
@%p11 bra $L__BB32_10;

$L__BB32_13:
setp.eq.s32 %p12, %r1, 0;
@%p12 bra $L__BB32_16;

setp.eq.s64 %p13, %rd15, 0;
add.s32 %r22, %r1, -1;
shl.b32 %r125, %r1, 7;
mov.u32 %r126, _ZZ13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r127, %r126, %r125;
shl.b32 %r128, %r2, 2;
add.s32 %r129, %r127, %r128;
st.shared.f32 [%r129+-128], %f119;
@%p13 bra $L__BB32_16;

shl.b32 %r131, %r22, 7;
mov.u32 %r132, _ZZ13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r133, %r132, %r131;
add.s32 %r135, %r133, %r128;
st.shared.f32 [%r135], %f117;

$L__BB32_16:
setp.ne.s32 %p14, %r1, 0;
bar.sync 0;
@%p14 bra $L__BB32_35;

setp.ne.s64 %p15, %rd15, 0;
shl.b32 %r138, %r2, 2;
mov.u32 %r139, _ZZ13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r23, %r139, %r138;
ld.shared.f32 %f52, [%r23];
add.ftz.f32 %f13, %f119, %f52;
@%p15 bra $L__BB32_19;
bra.uni $L__BB32_18;

$L__BB32_19:
mov.u32 %r142, _ZZ13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r143, %r142, %r138;
ld.shared.f32 %f56, [%r143];
add.ftz.f32 %f57, %f117, %f56;
ld.shared.f32 %f58, [%r23+128];
add.ftz.f32 %f59, %f13, %f58;
ld.shared.f32 %f60, [%r143+128];
add.ftz.f32 %f61, %f57, %f60;
ld.shared.f32 %f62, [%r23+256];
add.ftz.f32 %f121, %f59, %f62;
ld.shared.f32 %f63, [%r143+256];
add.ftz.f32 %f117, %f61, %f63;
bra.uni $L__BB32_20;

$L__BB32_18:
ld.shared.f32 %f53, [%r23+128];
add.ftz.f32 %f54, %f13, %f53;
ld.shared.f32 %f55, [%r23+256];
add.ftz.f32 %f121, %f54, %f55;

$L__BB32_20:
mov.b32 %r144, %f121;
mov.u32 %r145, 31;
mov.u32 %r146, 16;
mov.u32 %r147, -1;
shfl.sync.bfly.b32 %r148|%p16, %r144, %r146, %r145, %r147;
mov.b32 %f64, %r148;
add.ftz.f32 %f65, %f121, %f64;
mov.b32 %r149, %f65;
mov.u32 %r150, 8;
shfl.sync.bfly.b32 %r151|%p17, %r149, %r150, %r145, %r147;
mov.b32 %f66, %r151;
add.ftz.f32 %f67, %f65, %f66;
mov.b32 %r152, %f67;
mov.u32 %r153, 4;
shfl.sync.bfly.b32 %r154|%p18, %r152, %r153, %r145, %r147;
mov.b32 %f68, %r154;
add.ftz.f32 %f69, %f67, %f68;
mov.b32 %r155, %f69;
mov.u32 %r156, 2;
shfl.sync.bfly.b32 %r157|%p19, %r155, %r156, %r145, %r147;
mov.b32 %f70, %r157;
add.ftz.f32 %f19, %f69, %f70;
mov.b32 %r158, %f19;
mov.u32 %r159, 1;
shfl.sync.bfly.b32 %r24|%p1, %r158, %r159, %r145, %r147;
setp.eq.s64 %p20, %rd15, 0;
@%p20 bra $L__BB32_22;

mov.b32 %r160, %f117;
shfl.sync.bfly.b32 %r164|%p21, %r160, %r146, %r145, %r147;
mov.b32 %f71, %r164;
add.ftz.f32 %f72, %f117, %f71;
mov.b32 %r165, %f72;
shfl.sync.bfly.b32 %r167|%p22, %r165, %r150, %r145, %r147;
mov.b32 %f73, %r167;
add.ftz.f32 %f74, %f72, %f73;
mov.b32 %r168, %f74;
shfl.sync.bfly.b32 %r170|%p23, %r168, %r153, %r145, %r147;
mov.b32 %f75, %r170;
add.ftz.f32 %f76, %f74, %f75;
mov.b32 %r171, %f76;
shfl.sync.bfly.b32 %r173|%p24, %r171, %r156, %r145, %r147;
mov.b32 %f77, %r173;
add.ftz.f32 %f78, %f76, %f77;
mov.b32 %r174, %f78;
shfl.sync.bfly.b32 %r176|%p25, %r174, %r159, %r145, %r147;
mov.b32 %f79, %r176;
add.ftz.f32 %f117, %f78, %f79;

$L__BB32_22:
setp.ne.s32 %p26, %r2, 0;
mov.b32 %f80, %r24;
add.ftz.f32 %f22, %f19, %f80;
@%p26 bra $L__BB32_35;

add.ftz.f32 %f123, %f114, %f22;
@%p20 bra $L__BB32_34;

add.ftz.f32 %f24, %f113, %f117;
setp.gt.s32 %p28, %r25, 2;
@%p28 bra $L__BB32_28;

setp.eq.s32 %p31, %r25, 1;
@%p31 bra $L__BB32_32;

setp.eq.s32 %p32, %r25, 2;
@%p32 bra $L__BB32_27;
bra.uni $L__BB32_33;

$L__BB32_27:
mul.ftz.f32 %f100, %f24, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f101, %f100;
add.ftz.f32 %f102, %f101, 0f3F800000;
div.approx.ftz.f32 %f103, %f24, %f102;
mul.ftz.f32 %f123, %f123, %f103;
bra.uni $L__BB32_34;

$L__BB32_28:
setp.eq.s32 %p29, %r25, 3;
@%p29 bra $L__BB32_31;

setp.ne.s32 %p30, %r25, 6;
@%p30 bra $L__BB32_33;

min.ftz.f32 %f81, %f24, %f31;
neg.ftz.f32 %f82, %f31;
min.ftz.f32 %f83, %f123, %f31;
max.ftz.f32 %f84, %f83, %f82;
mul.ftz.f32 %f85, %f81, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f86, %f85;
add.ftz.f32 %f87, %f86, 0f3F800000;
div.approx.ftz.f32 %f88, %f81, %f87;
mul.ftz.f32 %f123, %f84, %f88;
bra.uni $L__BB32_34;

$L__BB32_32:
mul.ftz.f32 %f104, %f24, 0f3F000000;
mul.ftz.f32 %f105, %f24, 0f3D372713;
fma.rn.ftz.f32 %f106, %f24, %f105, 0f3F800000;
mul.ftz.f32 %f107, %f24, 0f3F4C422A;
mul.ftz.f32 %f108, %f107, %f106;
tanh.approx.f32 %f109, %f108;
add.ftz.f32 %f110, %f109, 0f3F800000;
mul.ftz.f32 %f111, %f104, %f110;
mul.ftz.f32 %f123, %f123, %f111;
bra.uni $L__BB32_34;

$L__BB32_31:
mov.f32 %f89, 0f40E00000;
min.ftz.f32 %f90, %f24, %f89;
min.ftz.f32 %f91, %f123, %f89;
mov.f32 %f92, 0fC0E00000;
max.ftz.f32 %f93, %f91, %f92;
mul.ftz.f32 %f94, %f90, 0f3FD9DB23;
mul.ftz.f32 %f95, %f94, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f96, %f95;
add.ftz.f32 %f97, %f96, 0f3F800000;
div.approx.ftz.f32 %f98, %f90, %f97;
add.ftz.f32 %f99, %f93, 0f3F800000;
mul.ftz.f32 %f123, %f99, %f98;
bra.uni $L__BB32_34;

$L__BB32_33:
mul.ftz.f32 %f123, %f123, %f24;

$L__BB32_34:
cvta.to.global.u64 %rd65, %rd19;
mov.u32 %r180, %ctaid.y;
mov.u32 %r181, %ctaid.x;
mad.lo.s32 %r182, %r180, %r35, %r181;
mad.lo.s32 %r183, %r11, %r41, %r182;
mul.wide.u32 %rd66, %r183, 4;
add.s64 %rd67, %rd65, %rd66;
st.global.f32 [%rd67], %f123;

$L__BB32_35:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<14>;
.reg .b16 %rs<16>;
.reg .f32 %f<51>;
.reg .b32 %r<172>;
.reg .b64 %rd<56>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[384];

ld.param.u64 %rd12, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd13, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd10, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd11, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
cvta.to.global.u64 %rd1, %rd13;
cvta.to.global.u64 %rd55, %rd12;
mov.u32 %r1, %tid.y;
shl.b32 %r43, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r43, %r2;
mov.u32 %r4, %ctaid.x;
shr.u32 %r5, %r26, 5;
mov.u32 %r6, %ctaid.y;
setp.eq.s64 %p1, %rd10, 0;
@%p1 bra $L__BB33_2;

cvta.to.global.u64 %rd14, %rd10;
mul.wide.u32 %rd15, %r6, 4;
add.s64 %rd16, %rd14, %rd15;
ld.global.nc.u32 %r168, [%rd16];
mul.hi.u32 %r44, %r6, %r27;
add.s32 %r45, %r44, %r6;
shr.u32 %r46, %r45, %r28;
mul.lo.s32 %r47, %r46, %r29;
sub.s32 %r169, %r6, %r47;
bra.uni $L__BB33_3;

$L__BB33_2:
mul.hi.u32 %r48, %r6, %r31;
add.s32 %r49, %r48, %r6;
shr.u32 %r168, %r49, %r32;
mov.u32 %r169, %r6;

$L__BB33_3:
mov.u32 %r12, %ctaid.z;
mul.hi.u32 %r50, %r12, %r37;
add.s32 %r51, %r50, %r12;
shr.u32 %r52, %r51, %r38;
mul.lo.s32 %r53, %r12, %r41;
cvt.u64.u32 %rd17, %r53;
mul.lo.s32 %r54, %r169, %r35;
cvt.u64.u32 %rd18, %r54;
add.s64 %rd3, %rd18, %rd17;
mul.lo.s32 %r13, %r52, %r40;
mul.lo.s32 %r14, %r168, %r34;
mul.lo.s32 %r15, %r4, %r30;
shr.u32 %r55, %r3, 31;
add.s32 %r56, %r3, %r55;
shr.s32 %r171, %r56, 1;
setp.ge.s32 %p2, %r171, %r5;
mov.f32 %f50, 0f00000000;
@%p2 bra $L__BB33_9;

and.b32 %r59, %r56, -2;
sub.s32 %r17, %r3, %r59;
shl.b32 %r18, %r17, 1;
not.b32 %r60, %r171;
add.s32 %r19, %r5, %r60;
and.b32 %r61, %r19, 64;
setp.ne.s32 %p3, %r61, 0;
mov.f32 %f50, 0f00000000;
@%p3 bra $L__BB33_6;

cvt.u64.u32 %rd19, %r171;
add.s64 %rd20, %rd3, %rd19;
add.s32 %r80, %r14, %r15;
add.s32 %r81, %r80, %r13;
add.s32 %r82, %r81, %r171;
mul.wide.s32 %rd21, %r82, 18;
add.s64 %rd22, %rd55, %rd21;
shl.b32 %r83, %r17, 2;
mul.wide.s32 %rd23, %r83, 2;
add.s64 %rd24, %rd22, %rd23;
mul.lo.s64 %rd25, %rd20, 36;
add.s64 %rd26, %rd1, %rd25;
add.s64 %rd27, %rd26, 4;
ld.global.nc.u16 %rs2, [%rd24+2];
or.b32 %r84, %r83, 1;
ld.global.nc.u16 %rs3, [%rd24+4];
mov.b32 %r85, {%rs2, %rs3};
mul.wide.s32 %rd28, %r18, 4;
add.s64 %rd29, %rd27, %rd28;
ld.global.nc.u32 %r64, [%rd29];
ld.global.nc.u32 %r68, [%rd29+16];
sub.s32 %r86, %r84, %r18;
ld.global.nc.u16 %rs4, [%rd24+6];
ld.global.nc.u16 %rs5, [%rd24+8];
mov.b32 %r87, {%rs4, %rs5};
mul.wide.s32 %rd30, %r86, 4;
add.s64 %rd31, %rd27, %rd30;
ld.global.nc.u32 %r72, [%rd31];
ld.global.nc.u32 %r76, [%rd29+20];
ld.global.nc.u16 %rs1, [%rd22];

{ cvt.f32.f16 %f11, %rs1;}


and.b32 %r63, %r85, 252645135;
shr.u32 %r88, %r85, 4;
and.b32 %r67, %r88, 252645135;
mov.u32 %r65, 0;

dp4a.s32.s32 %r62, %r63, %r64, %r65;


dp4a.s32.s32 %r66, %r67, %r68, %r62;

and.b32 %r71, %r87, 252645135;
shr.u32 %r89, %r87, 4;
and.b32 %r75, %r89, 252645135;

dp4a.s32.s32 %r70, %r71, %r72, %r66;


dp4a.s32.s32 %r74, %r75, %r76, %r70;

ld.global.nc.u32 %r78, [%rd26];

{.reg .f16 low,high;
mov.b32 {low,high},%r78;
cvt.f32.f16 %f12, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r78;
cvt.f32.f16 %f13, high;}


cvt.rn.f32.s32 %f14, %r74;
mul.ftz.f32 %f15, %f13, 0fC0800000;
fma.rn.ftz.f32 %f16, %f12, %f14, %f15;
fma.rn.ftz.f32 %f50, %f11, %f16, 0f00000000;
add.s32 %r171, %r171, 64;

$L__BB33_6:
and.b32 %r90, %r19, -64;
setp.eq.s32 %p4, %r90, 0;
@%p4 bra $L__BB33_9;

shl.b32 %r91, %r17, 2;
or.b32 %r92, %r91, 1;
cvt.s64.s32 %rd4, %r18;
sub.s32 %r93, %r92, %r18;
cvt.s64.s32 %rd5, %r93;
add.s32 %r94, %r171, %r13;
add.s32 %r95, %r94, %r14;
add.s32 %r96, %r95, %r15;
mul.wide.s32 %rd6, %r96, 18;
mul.wide.s32 %rd32, %r91, 2;
add.s64 %rd33, %rd6, %rd32;
add.s64 %rd7, %rd33, 2;
shl.b64 %rd40, %rd4, 2;
shl.b64 %rd42, %rd5, 2;

$L__BB33_8:
cvt.u64.u32 %rd34, %r171;
add.s64 %rd35, %rd3, %rd34;
mul.lo.s64 %rd36, %rd35, 36;
add.s64 %rd37, %rd1, %rd36;
add.s64 %rd38, %rd37, 4;
add.s64 %rd39, %rd55, %rd7;
ld.global.nc.u16 %rs8, [%rd39];
ld.global.nc.u16 %rs9, [%rd39+2];
mov.b32 %r133, {%rs8, %rs9};
add.s64 %rd41, %rd38, %rd40;
ld.global.nc.u32 %r99, [%rd41];
ld.global.nc.u32 %r103, [%rd41+16];
ld.global.nc.u16 %rs10, [%rd39+4];
ld.global.nc.u16 %rs11, [%rd39+6];
mov.b32 %r134, {%rs10, %rs11};
add.s64 %rd43, %rd38, %rd42;
ld.global.nc.u32 %r107, [%rd43];
ld.global.nc.u32 %r111, [%rd41+20];
add.s64 %rd44, %rd55, %rd6;
ld.global.nc.u16 %rs6, [%rd44];

{ cvt.f32.f16 %f17, %rs6;}


and.b32 %r98, %r133, 252645135;
shr.u32 %r135, %r133, 4;
and.b32 %r102, %r135, 252645135;
mov.u32 %r118, 0;

dp4a.s32.s32 %r97, %r98, %r99, %r118;


dp4a.s32.s32 %r101, %r102, %r103, %r97;

and.b32 %r106, %r134, 252645135;
shr.u32 %r136, %r134, 4;
and.b32 %r110, %r136, 252645135;

dp4a.s32.s32 %r105, %r106, %r107, %r101;


dp4a.s32.s32 %r109, %r110, %r111, %r105;

ld.global.nc.u32 %r113, [%rd37];

{.reg .f16 low,high;
mov.b32 {low,high},%r113;
cvt.f32.f16 %f18, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r113;
cvt.f32.f16 %f19, high;}


cvt.rn.f32.s32 %f23, %r109;
mul.ftz.f32 %f24, %f18, %f23;
mul.ftz.f32 %f25, %f19, 0f40800000;
sub.ftz.f32 %f26, %f24, %f25;
fma.rn.ftz.f32 %f27, %f17, %f26, %f50;
add.s32 %r137, %r171, 64;
cvt.u64.u32 %rd45, %r137;
add.s64 %rd46, %rd3, %rd45;
mul.lo.s64 %rd47, %rd46, 36;
add.s64 %rd48, %rd1, %rd47;
add.s64 %rd49, %rd48, 4;
ld.global.nc.u16 %rs12, [%rd39+1152];
ld.global.nc.u16 %rs13, [%rd39+1154];
mov.b32 %r138, {%rs12, %rs13};
add.s64 %rd50, %rd49, %rd40;
ld.global.nc.u32 %r117, [%rd50];
ld.global.nc.u32 %r121, [%rd50+16];
ld.global.nc.u16 %rs14, [%rd39+1156];
ld.global.nc.u16 %rs15, [%rd39+1158];
mov.b32 %r139, {%rs14, %rs15};
add.s64 %rd51, %rd49, %rd42;
ld.global.nc.u32 %r125, [%rd51];
ld.global.nc.u32 %r129, [%rd50+20];
ld.global.nc.u16 %rs7, [%rd44+1152];

{ cvt.f32.f16 %f20, %rs7;}


and.b32 %r116, %r138, 252645135;
shr.u32 %r140, %r138, 4;
and.b32 %r120, %r140, 252645135;

dp4a.s32.s32 %r115, %r116, %r117, %r118;


dp4a.s32.s32 %r119, %r120, %r121, %r115;

and.b32 %r124, %r139, 252645135;
shr.u32 %r141, %r139, 4;
and.b32 %r128, %r141, 252645135;

dp4a.s32.s32 %r123, %r124, %r125, %r119;


dp4a.s32.s32 %r127, %r128, %r129, %r123;

ld.global.nc.u32 %r131, [%rd48];

{.reg .f16 low,high;
mov.b32 {low,high},%r131;
cvt.f32.f16 %f21, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r131;
cvt.f32.f16 %f22, high;}


cvt.rn.f32.s32 %f28, %r127;
mul.ftz.f32 %f29, %f21, %f28;
mul.ftz.f32 %f30, %f22, 0f40800000;
sub.ftz.f32 %f31, %f29, %f30;
fma.rn.ftz.f32 %f50, %f20, %f31, %f27;
add.s64 %rd55, %rd55, 2304;
add.s32 %r171, %r171, 128;
setp.lt.s32 %p5, %r171, %r5;
@%p5 bra $L__BB33_8;

$L__BB33_9:
setp.eq.s32 %p6, %r1, 0;
@%p6 bra $L__BB33_11;

shl.b32 %r142, %r1, 7;
mov.u32 %r143, _ZZ13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r144, %r143, %r142;
shl.b32 %r145, %r2, 2;
add.s32 %r146, %r144, %r145;
st.shared.f32 [%r146+-128], %f50;

$L__BB33_11:
setp.ne.s32 %p7, %r1, 0;
bar.sync 0;
@%p7 bra $L__BB33_14;

mad.lo.s32 %r147, %r6, %r36, %r4;
mad.lo.s32 %r25, %r12, %r42, %r147;
shl.b32 %r148, %r2, 2;
mov.u32 %r149, _ZZ13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r150, %r149, %r148;
mov.u32 %r151, 2;
ld.shared.f32 %f32, [%r150];
add.ftz.f32 %f33, %f50, %f32;
ld.shared.f32 %f34, [%r150+128];
add.ftz.f32 %f35, %f33, %f34;
ld.shared.f32 %f36, [%r150+256];
add.ftz.f32 %f37, %f35, %f36;
mov.b32 %r152, %f37;
mov.u32 %r153, 31;
mov.u32 %r154, 16;
mov.u32 %r155, -1;
shfl.sync.bfly.b32 %r156|%p8, %r152, %r154, %r153, %r155;
mov.b32 %f38, %r156;
add.ftz.f32 %f39, %f37, %f38;
mov.b32 %r157, %f39;
mov.u32 %r158, 8;
shfl.sync.bfly.b32 %r159|%p9, %r157, %r158, %r153, %r155;
mov.b32 %f40, %r159;
add.ftz.f32 %f41, %f39, %f40;
mov.b32 %r160, %f41;
mov.u32 %r161, 4;
shfl.sync.bfly.b32 %r162|%p10, %r160, %r161, %r153, %r155;
mov.b32 %f42, %r162;
add.ftz.f32 %f43, %f41, %f42;
mov.b32 %r163, %f43;
shfl.sync.bfly.b32 %r164|%p11, %r163, %r151, %r153, %r155;
mov.b32 %f44, %r164;
add.ftz.f32 %f45, %f43, %f44;
mov.b32 %r165, %f45;
mov.u32 %r166, 1;
shfl.sync.bfly.b32 %r167|%p12, %r165, %r166, %r153, %r155;
mov.b32 %f46, %r167;
add.ftz.f32 %f7, %f45, %f46;
setp.ne.s32 %p13, %r2, 0;
@%p13 bra $L__BB33_14;

cvta.to.global.u64 %rd52, %rd11;
mul.wide.u32 %rd53, %r25, 4;
add.s64 %rd54, %rd52, %rd53;
st.global.f32 [%rd54], %f7;

$L__BB33_14:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type2ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type2ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type2ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type2ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type2ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type2ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type2ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type2ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type2ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<35>;
.reg .b16 %rs<13>;
.reg .f32 %f<121>;
.reg .b32 %r<223>;
.reg .b64 %rd<53>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type2ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type2ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type2ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type2ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r15, [_Z13mul_mat_vec_qIL9ggml_type2ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r16, [_Z13mul_mat_vec_qIL9ggml_type2ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r17, [_Z13mul_mat_vec_qIL9ggml_type2ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r18, [_Z13mul_mat_vec_qIL9ggml_type2ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type2ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type2ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type2ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type2ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type2ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type2ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type2ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type2ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type2ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r19, [_Z13mul_mat_vec_qIL9ggml_type2ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
cvta.to.global.u64 %rd1, %rd17;
mov.u32 %r1, %tid.y;
shl.b32 %r31, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r31, %r2;
mov.u32 %r32, %ctaid.x;
shl.b32 %r4, %r32, 1;
mov.u32 %r5, %ctaid.y;
mov.u32 %r6, %ctaid.z;
shr.u32 %r33, %r3, 31;
add.s32 %r34, %r3, %r33;
shr.s32 %r222, %r34, 1;
shr.u32 %r35, %r15, 5;
setp.ge.s32 %p3, %r222, %r35;
mov.f32 %f117, 0f00000000;
mov.f32 %f118, %f117;
mov.f32 %f119, %f117;
mov.f32 %f120, %f117;
@%p3 bra $L__BB34_3;

mul.lo.s32 %r36, %r5, %r23;
cvt.u64.u32 %rd18, %r36;
mul.lo.s32 %r37, %r6, %r29;
cvt.u64.u32 %rd19, %r37;
add.s64 %rd2, %rd19, %rd18;
mul.hi.u32 %r38, %r6, %r25;
add.s32 %r39, %r38, %r6;
shr.u32 %r40, %r39, %r26;
mul.lo.s32 %r41, %r40, %r28;
mul.hi.u32 %r42, %r5, %r19;
add.s32 %r43, %r42, %r5;
shr.u32 %r44, %r43, %r20;
mul.lo.s32 %r45, %r44, %r22;
cvta.to.global.u64 %rd52, %rd15;
and.b32 %r48, %r34, -2;
sub.s32 %r49, %r3, %r48;
shl.b32 %r50, %r49, 1;
shl.b32 %r51, %r49, 2;
or.b32 %r52, %r51, 1;
cvt.s64.s32 %rd4, %r50;
sub.s32 %r53, %r52, %r50;
cvt.s64.s32 %rd5, %r53;
add.s32 %r221, %r222, %r17;
add.s32 %r54, %r222, %r41;
add.s32 %r55, %r54, %r45;
add.s32 %r56, %r4, 1;
mad.lo.s32 %r57, %r16, %r56, %r55;
mul.wide.s32 %rd6, %r57, 18;
mad.lo.s32 %r58, %r4, %r16, %r222;
add.s32 %r59, %r58, %r41;
add.s32 %r60, %r59, %r45;
mul.wide.s32 %rd7, %r60, 18;
mul.wide.s32 %rd20, %r51, 2;
add.s64 %rd21, %rd7, %rd20;
add.s64 %rd8, %rd21, 4;
add.s64 %rd22, %rd6, %rd20;
add.s64 %rd9, %rd22, 2;
cvta.to.global.u64 %rd10, %rd16;
mov.f32 %f117, 0f00000000;
shl.b64 %rd29, %rd4, 2;
shl.b64 %rd31, %rd5, 2;

$L__BB34_2:
cvt.u64.u32 %rd23, %r222;
add.s64 %rd24, %rd2, %rd23;
mul.lo.s64 %rd25, %rd24, 36;
add.s64 %rd26, %rd10, %rd25;
add.s64 %rd27, %rd26, 4;
add.s64 %rd28, %rd52, %rd8;
ld.global.nc.u16 %rs5, [%rd28+-2];
ld.global.nc.u16 %rs6, [%rd28];
mov.b32 %r133, {%rs5, %rs6};
add.s64 %rd30, %rd27, %rd29;
ld.global.nc.u32 %r63, [%rd30];
ld.global.nc.u32 %r67, [%rd30+16];
ld.global.nc.u16 %rs7, [%rd28+2];
ld.global.nc.u16 %rs8, [%rd28+4];
mov.b32 %r134, {%rs7, %rs8};
add.s64 %rd32, %rd27, %rd31;
ld.global.nc.u32 %r71, [%rd32];
ld.global.nc.u32 %r75, [%rd30+20];
add.s64 %rd33, %rd52, %rd7;
ld.global.nc.u16 %rs1, [%rd33];

{ cvt.f32.f16 %f25, %rs1;}


and.b32 %r98, %r133, 252645135;
shr.u32 %r135, %r133, 4;
and.b32 %r102, %r135, 252645135;
mov.u32 %r118, 0;

dp4a.s32.s32 %r61, %r98, %r63, %r118;


dp4a.s32.s32 %r65, %r102, %r67, %r61;

and.b32 %r106, %r134, 252645135;
shr.u32 %r136, %r134, 4;
and.b32 %r110, %r136, 252645135;

dp4a.s32.s32 %r69, %r106, %r71, %r65;


dp4a.s32.s32 %r73, %r110, %r75, %r69;

ld.global.nc.u32 %r77, [%rd26];

{.reg .f16 low,high;
mov.b32 {low,high},%r77;
cvt.f32.f16 %f26, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r77;
cvt.f32.f16 %f27, high;}


cvt.rn.f32.s32 %f37, %r73;
mul.ftz.f32 %f38, %f26, %f37;
mul.ftz.f32 %f39, %f27, 0f40800000;
sub.ftz.f32 %f40, %f38, %f39;
fma.rn.ftz.f32 %f120, %f25, %f40, %f120;
add.s64 %rd34, %rd52, %rd9;
ld.global.nc.u16 %rs9, [%rd34];
ld.global.nc.u16 %rs10, [%rd34+2];
mov.b32 %r137, {%rs9, %rs10};
ld.global.nc.u16 %rs11, [%rd34+4];
ld.global.nc.u16 %rs12, [%rd34+6];
mov.b32 %r138, {%rs11, %rs12};
add.s64 %rd35, %rd52, %rd6;
ld.global.nc.u16 %rs2, [%rd35];

{ cvt.f32.f16 %f28, %rs2;}


and.b32 %r116, %r137, 252645135;
shr.u32 %r139, %r137, 4;
and.b32 %r120, %r139, 252645135;

dp4a.s32.s32 %r79, %r116, %r63, %r118;


dp4a.s32.s32 %r83, %r120, %r67, %r79;

and.b32 %r124, %r138, 252645135;
shr.u32 %r140, %r138, 4;
and.b32 %r128, %r140, 252645135;

dp4a.s32.s32 %r87, %r124, %r71, %r83;


dp4a.s32.s32 %r91, %r128, %r75, %r87;


{.reg .f16 low,high;
mov.b32 {low,high},%r77;
cvt.f32.f16 %f29, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r77;
cvt.f32.f16 %f30, high;}


cvt.rn.f32.s32 %f41, %r91;
mul.ftz.f32 %f42, %f29, %f41;
mul.ftz.f32 %f43, %f30, 0f40800000;
sub.ftz.f32 %f44, %f42, %f43;
fma.rn.ftz.f32 %f119, %f28, %f44, %f119;
cvt.u64.u32 %rd36, %r221;
add.s64 %rd37, %rd2, %rd36;
mul.lo.s64 %rd38, %rd37, 36;
add.s64 %rd39, %rd10, %rd38;
add.s64 %rd40, %rd39, 4;
add.s64 %rd41, %rd40, %rd29;
ld.global.nc.u32 %r99, [%rd41];
ld.global.nc.u32 %r103, [%rd41+16];
add.s64 %rd42, %rd40, %rd31;
ld.global.nc.u32 %r107, [%rd42];
ld.global.nc.u32 %r111, [%rd41+20];

{ cvt.f32.f16 %f31, %rs1;}



dp4a.s32.s32 %r97, %r98, %r99, %r118;


dp4a.s32.s32 %r101, %r102, %r103, %r97;


dp4a.s32.s32 %r105, %r106, %r107, %r101;


dp4a.s32.s32 %r109, %r110, %r111, %r105;

ld.global.nc.u32 %r113, [%rd39];

{.reg .f16 low,high;
mov.b32 {low,high},%r113;
cvt.f32.f16 %f32, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r113;
cvt.f32.f16 %f33, high;}


cvt.rn.f32.s32 %f45, %r109;
mul.ftz.f32 %f46, %f32, %f45;
mul.ftz.f32 %f47, %f33, 0f40800000;
sub.ftz.f32 %f48, %f46, %f47;
fma.rn.ftz.f32 %f118, %f31, %f48, %f118;

{ cvt.f32.f16 %f34, %rs2;}



dp4a.s32.s32 %r115, %r116, %r99, %r118;


dp4a.s32.s32 %r119, %r120, %r103, %r115;


dp4a.s32.s32 %r123, %r124, %r107, %r119;


dp4a.s32.s32 %r127, %r128, %r111, %r123;


{.reg .f16 low,high;
mov.b32 {low,high},%r113;
cvt.f32.f16 %f35, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r113;
cvt.f32.f16 %f36, high;}


cvt.rn.f32.s32 %f49, %r127;
mul.ftz.f32 %f50, %f35, %f49;
mul.ftz.f32 %f51, %f36, 0f40800000;
sub.ftz.f32 %f52, %f50, %f51;
fma.rn.ftz.f32 %f117, %f34, %f52, %f117;
add.s32 %r221, %r221, 64;
add.s64 %rd52, %rd52, 1152;
add.s32 %r222, %r222, 64;
setp.lt.s32 %p4, %r222, %r35;
@%p4 bra $L__BB34_2;

$L__BB34_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB34_5;

shl.b32 %r142, %r1, 9;
mov.u32 %r143, _ZZ13mul_mat_vec_qIL9ggml_type2ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r144, %r143, %r142;
shl.b32 %r145, %r2, 2;
add.s32 %r146, %r144, %r145;
st.shared.f32 [%r146+-512], %f120;
st.shared.f32 [%r146+-384], %f119;
st.shared.f32 [%r146+-256], %f118;
st.shared.f32 [%r146+-128], %f117;

$L__BB34_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB34_14;

mad.lo.s32 %r147, %r5, %r24, %r4;
mad.lo.s32 %r148, %r6, %r30, %r147;
cvt.u64.u32 %rd13, %r148;
shl.b32 %r149, %r2, 2;
mov.u32 %r150, _ZZ13mul_mat_vec_qIL9ggml_type2ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r13, %r150, %r149;
mov.u32 %r151, 2;
ld.shared.f32 %f53, [%r13];
add.ftz.f32 %f54, %f53, %f120;
ld.shared.f32 %f55, [%r13+512];
add.ftz.f32 %f56, %f55, %f54;
ld.shared.f32 %f57, [%r13+1024];
add.ftz.f32 %f58, %f57, %f56;
mov.b32 %r152, %f58;
mov.u32 %r153, 31;
mov.u32 %r154, 16;
mov.u32 %r155, -1;
shfl.sync.bfly.b32 %r156|%p7, %r152, %r154, %r153, %r155;
mov.b32 %f59, %r156;
add.ftz.f32 %f60, %f58, %f59;
mov.b32 %r157, %f60;
mov.u32 %r158, 8;
shfl.sync.bfly.b32 %r159|%p8, %r157, %r158, %r153, %r155;
mov.b32 %f61, %r159;
add.ftz.f32 %f62, %f60, %f61;
mov.b32 %r160, %f62;
mov.u32 %r161, 4;
shfl.sync.bfly.b32 %r162|%p9, %r160, %r161, %r153, %r155;
mov.b32 %f63, %r162;
add.ftz.f32 %f64, %f62, %f63;
mov.b32 %r163, %f64;
shfl.sync.bfly.b32 %r164|%p10, %r163, %r151, %r153, %r155;
mov.b32 %f65, %r164;
add.ftz.f32 %f66, %f64, %f65;
mov.b32 %r165, %f66;
mov.u32 %r166, 1;
shfl.sync.bfly.b32 %r167|%p11, %r165, %r166, %r153, %r155;
mov.b32 %f67, %r167;
add.ftz.f32 %f13, %f66, %f67;
setp.lt.u32 %p12, %r4, %r18;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
mul.wide.u32 %rd43, %r148, 4;
add.s64 %rd14, %rd1, %rd43;
not.pred %p14, %p1;
@%p14 bra $L__BB34_8;

st.global.f32 [%rd14], %f13;

$L__BB34_8:
ld.shared.f32 %f68, [%r13+128];
add.ftz.f32 %f69, %f68, %f119;
ld.shared.f32 %f70, [%r13+640];
add.ftz.f32 %f71, %f70, %f69;
ld.shared.f32 %f72, [%r13+1152];
add.ftz.f32 %f73, %f72, %f71;
mov.b32 %r168, %f73;
shfl.sync.bfly.b32 %r172|%p15, %r168, %r154, %r153, %r155;
mov.b32 %f74, %r172;
add.ftz.f32 %f75, %f73, %f74;
mov.b32 %r173, %f75;
shfl.sync.bfly.b32 %r175|%p16, %r173, %r158, %r153, %r155;
mov.b32 %f76, %r175;
add.ftz.f32 %f77, %f75, %f76;
mov.b32 %r176, %f77;
shfl.sync.bfly.b32 %r178|%p17, %r176, %r161, %r153, %r155;
mov.b32 %f78, %r178;
add.ftz.f32 %f79, %f77, %f78;
mov.b32 %r179, %f79;
shfl.sync.bfly.b32 %r181|%p18, %r179, %r151, %r153, %r155;
mov.b32 %f80, %r181;
add.ftz.f32 %f81, %f79, %f80;
mov.b32 %r182, %f81;
shfl.sync.bfly.b32 %r184|%p19, %r182, %r166, %r153, %r155;
mov.b32 %f82, %r184;
add.ftz.f32 %f14, %f81, %f82;
add.s32 %r185, %r4, 1;
setp.lt.u32 %p20, %r185, %r18;
setp.eq.s32 %p21, %r2, 1;
and.pred %p2, %p21, %p20;
not.pred %p22, %p2;
@%p22 bra $L__BB34_10;

st.global.f32 [%rd14+4], %f14;

$L__BB34_10:
ld.shared.f32 %f83, [%r13+256];
add.ftz.f32 %f84, %f83, %f118;
ld.shared.f32 %f85, [%r13+768];
add.ftz.f32 %f86, %f85, %f84;
ld.shared.f32 %f87, [%r13+1280];
add.ftz.f32 %f88, %f87, %f86;
mov.b32 %r186, %f88;
mov.u32 %r187, 31;
mov.u32 %r188, 16;
mov.u32 %r189, -1;
shfl.sync.bfly.b32 %r190|%p23, %r186, %r188, %r187, %r189;
mov.b32 %f89, %r190;
add.ftz.f32 %f90, %f88, %f89;
mov.b32 %r191, %f90;
mov.u32 %r192, 8;
shfl.sync.bfly.b32 %r193|%p24, %r191, %r192, %r187, %r189;
mov.b32 %f91, %r193;
add.ftz.f32 %f92, %f90, %f91;
mov.b32 %r194, %f92;
mov.u32 %r195, 4;
shfl.sync.bfly.b32 %r196|%p25, %r194, %r195, %r187, %r189;
mov.b32 %f93, %r196;
add.ftz.f32 %f94, %f92, %f93;
mov.b32 %r197, %f94;
mov.u32 %r198, 2;
shfl.sync.bfly.b32 %r199|%p26, %r197, %r198, %r187, %r189;
mov.b32 %f95, %r199;
add.ftz.f32 %f96, %f94, %f95;
mov.b32 %r200, %f96;
mov.u32 %r201, 1;
shfl.sync.bfly.b32 %r202|%p27, %r200, %r201, %r187, %r189;
mov.b32 %f97, %r202;
add.ftz.f32 %f15, %f96, %f97;
@%p14 bra $L__BB34_12;

cvt.u64.u32 %rd44, %r18;
add.s64 %rd45, %rd44, %rd13;
shl.b64 %rd46, %rd45, 2;
add.s64 %rd47, %rd1, %rd46;
st.global.f32 [%rd47], %f15;

$L__BB34_12:
ld.shared.f32 %f98, [%r13+384];
add.ftz.f32 %f99, %f98, %f117;
ld.shared.f32 %f100, [%r13+896];
add.ftz.f32 %f101, %f100, %f99;
ld.shared.f32 %f102, [%r13+1408];
add.ftz.f32 %f103, %f102, %f101;
mov.b32 %r203, %f103;
shfl.sync.bfly.b32 %r207|%p29, %r203, %r188, %r187, %r189;
mov.b32 %f104, %r207;
add.ftz.f32 %f105, %f103, %f104;
mov.b32 %r208, %f105;
shfl.sync.bfly.b32 %r210|%p30, %r208, %r192, %r187, %r189;
mov.b32 %f106, %r210;
add.ftz.f32 %f107, %f105, %f106;
mov.b32 %r211, %f107;
shfl.sync.bfly.b32 %r213|%p31, %r211, %r195, %r187, %r189;
mov.b32 %f108, %r213;
add.ftz.f32 %f109, %f107, %f108;
mov.b32 %r214, %f109;
shfl.sync.bfly.b32 %r216|%p32, %r214, %r198, %r187, %r189;
mov.b32 %f110, %r216;
add.ftz.f32 %f111, %f109, %f110;
mov.b32 %r217, %f111;
shfl.sync.bfly.b32 %r219|%p33, %r217, %r201, %r187, %r189;
mov.b32 %f112, %r219;
add.ftz.f32 %f16, %f111, %f112;
@%p22 bra $L__BB34_14;

add.s32 %r220, %r18, 1;
cvt.u64.u32 %rd48, %r220;
add.s64 %rd49, %rd48, %rd13;
shl.b64 %rd50, %rd49, 2;
add.s64 %rd51, %rd1, %rd50;
st.global.f32 [%rd51], %f16;

$L__BB34_14:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type2ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type2ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type2ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type2ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type2ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type2ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type2ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type2ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type2ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<47>;
.reg .b16 %rs<15>;
.reg .f32 %f<181>;
.reg .b32 %r<298>;
.reg .b64 %rd<64>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type2ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[2304];

ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type2ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type2ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type2ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r18, [_Z13mul_mat_vec_qIL9ggml_type2ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r19, [_Z13mul_mat_vec_qIL9ggml_type2ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type2ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type2ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type2ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type2ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type2ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type2ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type2ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type2ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type2ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type2ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type2ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type2ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
cvta.to.global.u64 %rd1, %rd18;
mov.u32 %r1, %tid.y;
shl.b32 %r34, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r34, %r2;
mov.u32 %r35, %ctaid.x;
shl.b32 %r4, %r35, 1;
mov.u32 %r5, %ctaid.y;
mov.u32 %r6, %ctaid.z;
shr.u32 %r36, %r3, 31;
add.s32 %r37, %r3, %r36;
shr.s32 %r297, %r37, 1;
shr.u32 %r38, %r18, 5;
setp.ge.s32 %p3, %r297, %r38;
mov.f32 %f175, 0f00000000;
mov.f32 %f176, %f175;
mov.f32 %f177, %f175;
mov.f32 %f178, %f175;
mov.f32 %f179, %f175;
mov.f32 %f180, %f175;
@%p3 bra $L__BB35_3;

mul.lo.s32 %r39, %r5, %r26;
cvt.u64.u32 %rd19, %r39;
mul.lo.s32 %r40, %r6, %r32;
cvt.u64.u32 %rd20, %r40;
add.s64 %rd2, %rd20, %rd19;
mul.hi.u32 %r41, %r6, %r28;
add.s32 %r42, %r41, %r6;
shr.u32 %r43, %r42, %r29;
mul.lo.s32 %r44, %r43, %r31;
mul.hi.u32 %r45, %r5, %r22;
add.s32 %r46, %r45, %r5;
shr.u32 %r47, %r46, %r23;
mul.lo.s32 %r48, %r47, %r25;
cvta.to.global.u64 %rd63, %rd16;
and.b32 %r51, %r37, -2;
sub.s32 %r52, %r3, %r51;
shl.b32 %r53, %r52, 1;
shl.b32 %r54, %r52, 2;
or.b32 %r55, %r54, 1;
cvt.s64.s32 %rd4, %r53;
sub.s32 %r56, %r55, %r53;
cvt.s64.s32 %rd5, %r56;
shl.b32 %r57, %r20, 1;
add.s32 %r296, %r297, %r57;
add.s32 %r295, %r297, %r20;
add.s32 %r58, %r297, %r44;
add.s32 %r59, %r58, %r48;
add.s32 %r60, %r4, 1;
mad.lo.s32 %r61, %r19, %r60, %r59;
mul.wide.s32 %rd6, %r61, 18;
mad.lo.s32 %r62, %r4, %r19, %r297;
add.s32 %r63, %r62, %r44;
add.s32 %r64, %r63, %r48;
mul.wide.s32 %rd7, %r64, 18;
mul.wide.s32 %rd21, %r54, 2;
add.s64 %rd22, %rd7, %rd21;
add.s64 %rd8, %rd22, 4;
add.s64 %rd23, %rd6, %rd21;
add.s64 %rd9, %rd23, 2;
cvta.to.global.u64 %rd10, %rd17;
mov.f32 %f175, 0f00000000;
shl.b64 %rd30, %rd4, 2;
shl.b64 %rd32, %rd5, 2;

$L__BB35_2:
cvt.u64.u32 %rd24, %r297;
add.s64 %rd25, %rd2, %rd24;
mul.lo.s64 %rd26, %rd25, 36;
add.s64 %rd27, %rd10, %rd26;
add.s64 %rd28, %rd27, 4;
add.s64 %rd29, %rd63, %rd8;
ld.global.nc.u16 %rs7, [%rd29+-2];
ld.global.nc.u16 %rs8, [%rd29];
mov.b32 %r173, {%rs7, %rs8};
add.s64 %rd31, %rd28, %rd30;
ld.global.nc.u32 %r67, [%rd31];
ld.global.nc.u32 %r71, [%rd31+16];
ld.global.nc.u16 %rs9, [%rd29+2];
ld.global.nc.u16 %rs10, [%rd29+4];
mov.b32 %r174, {%rs9, %rs10};
add.s64 %rd33, %rd28, %rd32;
ld.global.nc.u32 %r75, [%rd33];
ld.global.nc.u32 %r79, [%rd31+20];
add.s64 %rd34, %rd63, %rd7;
ld.global.nc.u16 %rs1, [%rd34];

{ cvt.f32.f16 %f37, %rs1;}


and.b32 %r138, %r173, 252645135;
shr.u32 %r175, %r173, 4;
and.b32 %r142, %r175, 252645135;
mov.u32 %r158, 0;

dp4a.s32.s32 %r65, %r138, %r67, %r158;


dp4a.s32.s32 %r69, %r142, %r71, %r65;

and.b32 %r146, %r174, 252645135;
shr.u32 %r176, %r174, 4;
and.b32 %r150, %r176, 252645135;

dp4a.s32.s32 %r73, %r146, %r75, %r69;


dp4a.s32.s32 %r77, %r150, %r79, %r73;

ld.global.nc.u32 %r81, [%rd27];

{.reg .f16 low,high;
mov.b32 {low,high},%r81;
cvt.f32.f16 %f38, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r81;
cvt.f32.f16 %f39, high;}


cvt.rn.f32.s32 %f55, %r77;
mul.ftz.f32 %f56, %f38, %f55;
mul.ftz.f32 %f57, %f39, 0f40800000;
sub.ftz.f32 %f58, %f56, %f57;
fma.rn.ftz.f32 %f180, %f37, %f58, %f180;
add.s64 %rd35, %rd63, %rd9;
ld.global.nc.u16 %rs11, [%rd35];
ld.global.nc.u16 %rs12, [%rd35+2];
mov.b32 %r177, {%rs11, %rs12};
ld.global.nc.u16 %rs13, [%rd35+4];
ld.global.nc.u16 %rs14, [%rd35+6];
mov.b32 %r178, {%rs13, %rs14};
add.s64 %rd36, %rd63, %rd6;
ld.global.nc.u16 %rs2, [%rd36];

{ cvt.f32.f16 %f40, %rs2;}


and.b32 %r156, %r177, 252645135;
shr.u32 %r179, %r177, 4;
and.b32 %r160, %r179, 252645135;

dp4a.s32.s32 %r83, %r156, %r67, %r158;


dp4a.s32.s32 %r87, %r160, %r71, %r83;

and.b32 %r164, %r178, 252645135;
shr.u32 %r180, %r178, 4;
and.b32 %r168, %r180, 252645135;

dp4a.s32.s32 %r91, %r164, %r75, %r87;


dp4a.s32.s32 %r95, %r168, %r79, %r91;


{.reg .f16 low,high;
mov.b32 {low,high},%r81;
cvt.f32.f16 %f41, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r81;
cvt.f32.f16 %f42, high;}


cvt.rn.f32.s32 %f59, %r95;
mul.ftz.f32 %f60, %f41, %f59;
mul.ftz.f32 %f61, %f42, 0f40800000;
sub.ftz.f32 %f62, %f60, %f61;
fma.rn.ftz.f32 %f179, %f40, %f62, %f179;
cvt.u64.u32 %rd37, %r295;
add.s64 %rd38, %rd2, %rd37;
mul.lo.s64 %rd39, %rd38, 36;
add.s64 %rd40, %rd10, %rd39;
add.s64 %rd41, %rd40, 4;
add.s64 %rd42, %rd41, %rd30;
ld.global.nc.u32 %r103, [%rd42];
ld.global.nc.u32 %r107, [%rd42+16];
add.s64 %rd43, %rd41, %rd32;
ld.global.nc.u32 %r111, [%rd43];
ld.global.nc.u32 %r115, [%rd42+20];

{ cvt.f32.f16 %f43, %rs1;}



dp4a.s32.s32 %r101, %r138, %r103, %r158;


dp4a.s32.s32 %r105, %r142, %r107, %r101;


dp4a.s32.s32 %r109, %r146, %r111, %r105;


dp4a.s32.s32 %r113, %r150, %r115, %r109;

ld.global.nc.u32 %r117, [%rd40];

{.reg .f16 low,high;
mov.b32 {low,high},%r117;
cvt.f32.f16 %f44, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r117;
cvt.f32.f16 %f45, high;}


cvt.rn.f32.s32 %f63, %r113;
mul.ftz.f32 %f64, %f44, %f63;
mul.ftz.f32 %f65, %f45, 0f40800000;
sub.ftz.f32 %f66, %f64, %f65;
fma.rn.ftz.f32 %f178, %f43, %f66, %f178;

{ cvt.f32.f16 %f46, %rs2;}



dp4a.s32.s32 %r119, %r156, %r103, %r158;


dp4a.s32.s32 %r123, %r160, %r107, %r119;


dp4a.s32.s32 %r127, %r164, %r111, %r123;


dp4a.s32.s32 %r131, %r168, %r115, %r127;


{.reg .f16 low,high;
mov.b32 {low,high},%r117;
cvt.f32.f16 %f47, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r117;
cvt.f32.f16 %f48, high;}


cvt.rn.f32.s32 %f67, %r131;
mul.ftz.f32 %f68, %f47, %f67;
mul.ftz.f32 %f69, %f48, 0f40800000;
sub.ftz.f32 %f70, %f68, %f69;
fma.rn.ftz.f32 %f177, %f46, %f70, %f177;
cvt.u64.u32 %rd44, %r296;
add.s64 %rd45, %rd2, %rd44;
mul.lo.s64 %rd46, %rd45, 36;
add.s64 %rd47, %rd10, %rd46;
add.s64 %rd48, %rd47, 4;
add.s64 %rd49, %rd48, %rd30;
ld.global.nc.u32 %r139, [%rd49];
ld.global.nc.u32 %r143, [%rd49+16];
add.s64 %rd50, %rd48, %rd32;
ld.global.nc.u32 %r147, [%rd50];
ld.global.nc.u32 %r151, [%rd49+20];

{ cvt.f32.f16 %f49, %rs1;}



dp4a.s32.s32 %r137, %r138, %r139, %r158;


dp4a.s32.s32 %r141, %r142, %r143, %r137;


dp4a.s32.s32 %r145, %r146, %r147, %r141;


dp4a.s32.s32 %r149, %r150, %r151, %r145;

ld.global.nc.u32 %r153, [%rd47];

{.reg .f16 low,high;
mov.b32 {low,high},%r153;
cvt.f32.f16 %f50, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r153;
cvt.f32.f16 %f51, high;}


cvt.rn.f32.s32 %f71, %r149;
mul.ftz.f32 %f72, %f50, %f71;
mul.ftz.f32 %f73, %f51, 0f40800000;
sub.ftz.f32 %f74, %f72, %f73;
fma.rn.ftz.f32 %f176, %f49, %f74, %f176;

{ cvt.f32.f16 %f52, %rs2;}



dp4a.s32.s32 %r155, %r156, %r139, %r158;


dp4a.s32.s32 %r159, %r160, %r143, %r155;


dp4a.s32.s32 %r163, %r164, %r147, %r159;


dp4a.s32.s32 %r167, %r168, %r151, %r163;


{.reg .f16 low,high;
mov.b32 {low,high},%r153;
cvt.f32.f16 %f53, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r153;
cvt.f32.f16 %f54, high;}


cvt.rn.f32.s32 %f75, %r167;
mul.ftz.f32 %f76, %f53, %f75;
mul.ftz.f32 %f77, %f54, 0f40800000;
sub.ftz.f32 %f78, %f76, %f77;
fma.rn.ftz.f32 %f175, %f52, %f78, %f175;
add.s32 %r296, %r296, 64;
add.s32 %r295, %r295, 64;
add.s64 %rd63, %rd63, 1152;
add.s32 %r297, %r297, 64;
setp.lt.s32 %p4, %r297, %r38;
@%p4 bra $L__BB35_2;

$L__BB35_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB35_5;

mov.u32 %r182, _ZZ13mul_mat_vec_qIL9ggml_type2ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r183, %r1, 768, %r182;
shl.b32 %r184, %r2, 2;
add.s32 %r185, %r183, %r184;
st.shared.f32 [%r185+-768], %f180;
st.shared.f32 [%r185+-640], %f179;
st.shared.f32 [%r185+-512], %f178;
st.shared.f32 [%r185+-384], %f177;
st.shared.f32 [%r185+-256], %f176;
st.shared.f32 [%r185+-128], %f175;

$L__BB35_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB35_18;

mad.lo.s32 %r186, %r5, %r27, %r4;
mad.lo.s32 %r187, %r6, %r33, %r186;
cvt.u64.u32 %rd13, %r187;
shl.b32 %r188, %r2, 2;
mov.u32 %r189, _ZZ13mul_mat_vec_qIL9ggml_type2ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r16, %r189, %r188;
mov.u32 %r190, 2;
ld.shared.f32 %f79, [%r16];
add.ftz.f32 %f80, %f79, %f180;
ld.shared.f32 %f81, [%r16+768];
add.ftz.f32 %f82, %f81, %f80;
ld.shared.f32 %f83, [%r16+1536];
add.ftz.f32 %f84, %f83, %f82;
mov.b32 %r191, %f84;
mov.u32 %r192, 31;
mov.u32 %r193, 16;
mov.u32 %r194, -1;
shfl.sync.bfly.b32 %r195|%p7, %r191, %r193, %r192, %r194;
mov.b32 %f85, %r195;
add.ftz.f32 %f86, %f84, %f85;
mov.b32 %r196, %f86;
mov.u32 %r197, 8;
shfl.sync.bfly.b32 %r198|%p8, %r196, %r197, %r192, %r194;
mov.b32 %f87, %r198;
add.ftz.f32 %f88, %f86, %f87;
mov.b32 %r199, %f88;
mov.u32 %r200, 4;
shfl.sync.bfly.b32 %r201|%p9, %r199, %r200, %r192, %r194;
mov.b32 %f89, %r201;
add.ftz.f32 %f90, %f88, %f89;
mov.b32 %r202, %f90;
shfl.sync.bfly.b32 %r203|%p10, %r202, %r190, %r192, %r194;
mov.b32 %f91, %r203;
add.ftz.f32 %f92, %f90, %f91;
mov.b32 %r204, %f92;
mov.u32 %r205, 1;
shfl.sync.bfly.b32 %r206|%p11, %r204, %r205, %r192, %r194;
mov.b32 %f93, %r206;
add.ftz.f32 %f19, %f92, %f93;
setp.lt.u32 %p12, %r4, %r21;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
mul.wide.u32 %rd51, %r187, 4;
add.s64 %rd14, %rd1, %rd51;
not.pred %p14, %p1;
@%p14 bra $L__BB35_8;

st.global.f32 [%rd14], %f19;

$L__BB35_8:
ld.shared.f32 %f94, [%r16+128];
add.ftz.f32 %f95, %f94, %f179;
ld.shared.f32 %f96, [%r16+896];
add.ftz.f32 %f97, %f96, %f95;
ld.shared.f32 %f98, [%r16+1664];
add.ftz.f32 %f99, %f98, %f97;
mov.b32 %r207, %f99;
shfl.sync.bfly.b32 %r211|%p15, %r207, %r193, %r192, %r194;
mov.b32 %f100, %r211;
add.ftz.f32 %f101, %f99, %f100;
mov.b32 %r212, %f101;
shfl.sync.bfly.b32 %r214|%p16, %r212, %r197, %r192, %r194;
mov.b32 %f102, %r214;
add.ftz.f32 %f103, %f101, %f102;
mov.b32 %r215, %f103;
shfl.sync.bfly.b32 %r217|%p17, %r215, %r200, %r192, %r194;
mov.b32 %f104, %r217;
add.ftz.f32 %f105, %f103, %f104;
mov.b32 %r218, %f105;
shfl.sync.bfly.b32 %r220|%p18, %r218, %r190, %r192, %r194;
mov.b32 %f106, %r220;
add.ftz.f32 %f107, %f105, %f106;
mov.b32 %r221, %f107;
shfl.sync.bfly.b32 %r223|%p19, %r221, %r205, %r192, %r194;
mov.b32 %f108, %r223;
add.ftz.f32 %f20, %f107, %f108;
add.s32 %r224, %r4, 1;
setp.lt.u32 %p20, %r224, %r21;
setp.eq.s32 %p21, %r2, 1;
and.pred %p2, %p21, %p20;
not.pred %p22, %p2;
@%p22 bra $L__BB35_10;

st.global.f32 [%rd14+4], %f20;

$L__BB35_10:
ld.shared.f32 %f109, [%r16+256];
add.ftz.f32 %f110, %f109, %f178;
ld.shared.f32 %f111, [%r16+1024];
add.ftz.f32 %f112, %f111, %f110;
ld.shared.f32 %f113, [%r16+1792];
add.ftz.f32 %f114, %f113, %f112;
mov.b32 %r225, %f114;
mov.u32 %r226, 31;
mov.u32 %r227, 16;
mov.u32 %r228, -1;
shfl.sync.bfly.b32 %r229|%p23, %r225, %r227, %r226, %r228;
mov.b32 %f115, %r229;
add.ftz.f32 %f116, %f114, %f115;
mov.b32 %r230, %f116;
mov.u32 %r231, 8;
shfl.sync.bfly.b32 %r232|%p24, %r230, %r231, %r226, %r228;
mov.b32 %f117, %r232;
add.ftz.f32 %f118, %f116, %f117;
mov.b32 %r233, %f118;
mov.u32 %r234, 4;
shfl.sync.bfly.b32 %r235|%p25, %r233, %r234, %r226, %r228;
mov.b32 %f119, %r235;
add.ftz.f32 %f120, %f118, %f119;
mov.b32 %r236, %f120;
mov.u32 %r237, 2;
shfl.sync.bfly.b32 %r238|%p26, %r236, %r237, %r226, %r228;
mov.b32 %f121, %r238;
add.ftz.f32 %f122, %f120, %f121;
mov.b32 %r239, %f122;
mov.u32 %r240, 1;
shfl.sync.bfly.b32 %r241|%p27, %r239, %r240, %r226, %r228;
mov.b32 %f123, %r241;
add.ftz.f32 %f21, %f122, %f123;
@%p14 bra $L__BB35_12;

cvt.u64.u32 %rd52, %r21;
add.s64 %rd53, %rd52, %rd13;
shl.b64 %rd54, %rd53, 2;
add.s64 %rd55, %rd1, %rd54;
st.global.f32 [%rd55], %f21;

$L__BB35_12:
ld.shared.f32 %f124, [%r16+384];
add.ftz.f32 %f125, %f124, %f177;
ld.shared.f32 %f126, [%r16+1152];
add.ftz.f32 %f127, %f126, %f125;
ld.shared.f32 %f128, [%r16+1920];
add.ftz.f32 %f129, %f128, %f127;
mov.b32 %r242, %f129;
shfl.sync.bfly.b32 %r246|%p29, %r242, %r227, %r226, %r228;
mov.b32 %f130, %r246;
add.ftz.f32 %f131, %f129, %f130;
mov.b32 %r247, %f131;
shfl.sync.bfly.b32 %r249|%p30, %r247, %r231, %r226, %r228;
mov.b32 %f132, %r249;
add.ftz.f32 %f133, %f131, %f132;
mov.b32 %r250, %f133;
shfl.sync.bfly.b32 %r252|%p31, %r250, %r234, %r226, %r228;
mov.b32 %f134, %r252;
add.ftz.f32 %f135, %f133, %f134;
mov.b32 %r253, %f135;
shfl.sync.bfly.b32 %r255|%p32, %r253, %r237, %r226, %r228;
mov.b32 %f136, %r255;
add.ftz.f32 %f137, %f135, %f136;
mov.b32 %r256, %f137;
shfl.sync.bfly.b32 %r258|%p33, %r256, %r240, %r226, %r228;
mov.b32 %f138, %r258;
add.ftz.f32 %f22, %f137, %f138;
@%p22 bra $L__BB35_14;

add.s32 %r259, %r21, 1;
cvt.u64.u32 %rd56, %r259;
add.s64 %rd57, %rd56, %rd13;
shl.b64 %rd58, %rd57, 2;
add.s64 %rd59, %rd1, %rd58;
st.global.f32 [%rd59], %f22;

$L__BB35_14:
shl.b32 %r260, %r21, 1;
mov.u32 %r261, 1;
ld.shared.f32 %f139, [%r16+512];
add.ftz.f32 %f140, %f139, %f176;
ld.shared.f32 %f141, [%r16+1280];
add.ftz.f32 %f142, %f141, %f140;
ld.shared.f32 %f143, [%r16+2048];
add.ftz.f32 %f144, %f143, %f142;
mov.b32 %r262, %f144;
mov.u32 %r263, 31;
mov.u32 %r264, 16;
mov.u32 %r265, -1;
shfl.sync.bfly.b32 %r266|%p35, %r262, %r264, %r263, %r265;
mov.b32 %f145, %r266;
add.ftz.f32 %f146, %f144, %f145;
mov.b32 %r267, %f146;
mov.u32 %r268, 8;
shfl.sync.bfly.b32 %r269|%p36, %r267, %r268, %r263, %r265;
mov.b32 %f147, %r269;
add.ftz.f32 %f148, %f146, %f147;
mov.b32 %r270, %f148;
mov.u32 %r271, 4;
shfl.sync.bfly.b32 %r272|%p37, %r270, %r271, %r263, %r265;
mov.b32 %f149, %r272;
add.ftz.f32 %f150, %f148, %f149;
mov.b32 %r273, %f150;
mov.u32 %r274, 2;
shfl.sync.bfly.b32 %r275|%p38, %r273, %r274, %r263, %r265;
mov.b32 %f151, %r275;
add.ftz.f32 %f152, %f150, %f151;
mov.b32 %r276, %f152;
shfl.sync.bfly.b32 %r277|%p39, %r276, %r261, %r263, %r265;
mov.b32 %f153, %r277;
add.ftz.f32 %f23, %f152, %f153;
cvt.u64.u32 %rd60, %r260;
add.s64 %rd61, %rd60, %rd13;
shl.b64 %rd62, %rd61, 2;
add.s64 %rd15, %rd1, %rd62;
@%p14 bra $L__BB35_16;

st.global.f32 [%rd15], %f23;

$L__BB35_16:
ld.shared.f32 %f154, [%r16+640];
add.ftz.f32 %f155, %f154, %f175;
ld.shared.f32 %f156, [%r16+1408];
add.ftz.f32 %f157, %f156, %f155;
ld.shared.f32 %f158, [%r16+2176];
add.ftz.f32 %f159, %f158, %f157;
mov.b32 %r278, %f159;
shfl.sync.bfly.b32 %r282|%p41, %r278, %r264, %r263, %r265;
mov.b32 %f160, %r282;
add.ftz.f32 %f161, %f159, %f160;
mov.b32 %r283, %f161;
shfl.sync.bfly.b32 %r285|%p42, %r283, %r268, %r263, %r265;
mov.b32 %f162, %r285;
add.ftz.f32 %f163, %f161, %f162;
mov.b32 %r286, %f163;
shfl.sync.bfly.b32 %r288|%p43, %r286, %r271, %r263, %r265;
mov.b32 %f164, %r288;
add.ftz.f32 %f165, %f163, %f164;
mov.b32 %r289, %f165;
shfl.sync.bfly.b32 %r291|%p44, %r289, %r274, %r263, %r265;
mov.b32 %f166, %r291;
add.ftz.f32 %f167, %f165, %f166;
mov.b32 %r292, %f167;
shfl.sync.bfly.b32 %r294|%p45, %r292, %r261, %r263, %r265;
mov.b32 %f168, %r294;
add.ftz.f32 %f24, %f167, %f168;
@%p22 bra $L__BB35_18;

st.global.f32 [%rd15+4], %f24;

$L__BB35_18:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type2ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type2ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type2ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type2ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type2ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type2ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type2ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type2ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type2ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<59>;
.reg .b16 %rs<17>;
.reg .f32 %f<241>;
.reg .b32 %r<377>;
.reg .b64 %rd<79>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type2ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[3072];

ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type2ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type2ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type2ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type2ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type2ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type2ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type2ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type2ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type2ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type2ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type2ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type2ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type2ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type2ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type2ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type2ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type2ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
cvta.to.global.u64 %rd1, %rd18;
mov.u32 %r1, %tid.y;
shl.b32 %r38, %r1, 5;
mov.u32 %r39, %tid.x;
add.s32 %r2, %r38, %r39;
mov.u32 %r40, %ctaid.x;
shl.b32 %r3, %r40, 1;
mov.u32 %r4, %ctaid.y;
mov.u32 %r5, %ctaid.z;
shr.u32 %r41, %r2, 31;
add.s32 %r42, %r2, %r41;
shr.s32 %r376, %r42, 1;
shr.u32 %r7, %r37, 5;
setp.ge.s32 %p3, %r376, %r7;
mov.f32 %f233, 0f00000000;
mov.f32 %f234, %f233;
mov.f32 %f235, %f233;
mov.f32 %f236, %f233;
mov.f32 %f237, %f233;
mov.f32 %f238, %f233;
mov.f32 %f239, %f233;
mov.f32 %f240, %f233;
@%p3 bra $L__BB36_3;

mul.lo.s32 %r43, %r4, %r29;
cvt.u64.u32 %rd19, %r43;
mul.lo.s32 %r44, %r5, %r35;
cvt.u64.u32 %rd20, %r44;
add.s64 %rd2, %rd20, %rd19;
mul.hi.u32 %r45, %r5, %r31;
add.s32 %r46, %r45, %r5;
shr.u32 %r47, %r46, %r32;
mul.lo.s32 %r48, %r47, %r34;
mul.hi.u32 %r49, %r4, %r25;
add.s32 %r50, %r49, %r4;
shr.u32 %r51, %r50, %r26;
mul.lo.s32 %r52, %r51, %r28;
cvta.to.global.u64 %rd78, %rd16;
and.b32 %r55, %r42, -2;
sub.s32 %r56, %r2, %r55;
shl.b32 %r57, %r56, 1;
shl.b32 %r58, %r56, 2;
or.b32 %r59, %r58, 1;
cvt.s64.s32 %rd4, %r57;
sub.s32 %r60, %r59, %r57;
cvt.s64.s32 %rd5, %r60;
shl.b32 %r61, %r23, 1;
add.s32 %r375, %r376, %r61;
mad.lo.s32 %r374, %r23, 3, %r376;
add.s32 %r373, %r376, %r23;
add.s32 %r62, %r376, %r48;
add.s32 %r63, %r62, %r52;
add.s32 %r64, %r3, 1;
mad.lo.s32 %r65, %r22, %r64, %r63;
mul.wide.s32 %rd6, %r65, 18;
mad.lo.s32 %r66, %r3, %r22, %r376;
add.s32 %r67, %r66, %r48;
add.s32 %r68, %r67, %r52;
mul.wide.s32 %rd7, %r68, 18;
mul.wide.s32 %rd21, %r58, 2;
add.s64 %rd22, %rd7, %rd21;
add.s64 %rd8, %rd22, 4;
add.s64 %rd23, %rd6, %rd21;
add.s64 %rd9, %rd23, 2;
cvta.to.global.u64 %rd10, %rd17;
mov.f32 %f233, 0f00000000;
shl.b64 %rd30, %rd4, 2;
shl.b64 %rd32, %rd5, 2;

$L__BB36_2:
cvt.u64.u32 %rd24, %r376;
add.s64 %rd25, %rd2, %rd24;
mul.lo.s64 %rd26, %rd25, 36;
add.s64 %rd27, %rd10, %rd26;
add.s64 %rd28, %rd27, 4;
add.s64 %rd29, %rd78, %rd8;
ld.global.nc.u16 %rs9, [%rd29+-2];
ld.global.nc.u16 %rs10, [%rd29];
mov.b32 %r213, {%rs9, %rs10};
add.s64 %rd31, %rd28, %rd30;
ld.global.nc.u32 %r71, [%rd31];
ld.global.nc.u32 %r75, [%rd31+16];
ld.global.nc.u16 %rs11, [%rd29+2];
ld.global.nc.u16 %rs12, [%rd29+4];
mov.b32 %r214, {%rs11, %rs12};
add.s64 %rd33, %rd28, %rd32;
ld.global.nc.u32 %r79, [%rd33];
ld.global.nc.u32 %r83, [%rd31+20];
add.s64 %rd34, %rd78, %rd7;
ld.global.nc.u16 %rs1, [%rd34];

{ cvt.f32.f16 %f49, %rs1;}


and.b32 %r178, %r213, 252645135;
shr.u32 %r215, %r213, 4;
and.b32 %r182, %r215, 252645135;
mov.u32 %r198, 0;

dp4a.s32.s32 %r69, %r178, %r71, %r198;


dp4a.s32.s32 %r73, %r182, %r75, %r69;

and.b32 %r186, %r214, 252645135;
shr.u32 %r216, %r214, 4;
and.b32 %r190, %r216, 252645135;

dp4a.s32.s32 %r77, %r186, %r79, %r73;


dp4a.s32.s32 %r81, %r190, %r83, %r77;

ld.global.nc.u32 %r85, [%rd27];

{.reg .f16 low,high;
mov.b32 {low,high},%r85;
cvt.f32.f16 %f50, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r85;
cvt.f32.f16 %f51, high;}


cvt.rn.f32.s32 %f73, %r81;
mul.ftz.f32 %f74, %f50, %f73;
mul.ftz.f32 %f75, %f51, 0f40800000;
sub.ftz.f32 %f76, %f74, %f75;
fma.rn.ftz.f32 %f240, %f49, %f76, %f240;
add.s64 %rd35, %rd78, %rd9;
ld.global.nc.u16 %rs13, [%rd35];
ld.global.nc.u16 %rs14, [%rd35+2];
mov.b32 %r217, {%rs13, %rs14};
ld.global.nc.u16 %rs15, [%rd35+4];
ld.global.nc.u16 %rs16, [%rd35+6];
mov.b32 %r218, {%rs15, %rs16};
add.s64 %rd36, %rd78, %rd6;
ld.global.nc.u16 %rs2, [%rd36];

{ cvt.f32.f16 %f52, %rs2;}


and.b32 %r196, %r217, 252645135;
shr.u32 %r219, %r217, 4;
and.b32 %r200, %r219, 252645135;

dp4a.s32.s32 %r87, %r196, %r71, %r198;


dp4a.s32.s32 %r91, %r200, %r75, %r87;

and.b32 %r204, %r218, 252645135;
shr.u32 %r220, %r218, 4;
and.b32 %r208, %r220, 252645135;

dp4a.s32.s32 %r95, %r204, %r79, %r91;


dp4a.s32.s32 %r99, %r208, %r83, %r95;


{.reg .f16 low,high;
mov.b32 {low,high},%r85;
cvt.f32.f16 %f53, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r85;
cvt.f32.f16 %f54, high;}


cvt.rn.f32.s32 %f77, %r99;
mul.ftz.f32 %f78, %f53, %f77;
mul.ftz.f32 %f79, %f54, 0f40800000;
sub.ftz.f32 %f80, %f78, %f79;
fma.rn.ftz.f32 %f239, %f52, %f80, %f239;
cvt.u64.u32 %rd37, %r373;
add.s64 %rd38, %rd2, %rd37;
mul.lo.s64 %rd39, %rd38, 36;
add.s64 %rd40, %rd10, %rd39;
add.s64 %rd41, %rd40, 4;
add.s64 %rd42, %rd41, %rd30;
ld.global.nc.u32 %r107, [%rd42];
ld.global.nc.u32 %r111, [%rd42+16];
add.s64 %rd43, %rd41, %rd32;
ld.global.nc.u32 %r115, [%rd43];
ld.global.nc.u32 %r119, [%rd42+20];

{ cvt.f32.f16 %f55, %rs1;}



dp4a.s32.s32 %r105, %r178, %r107, %r198;


dp4a.s32.s32 %r109, %r182, %r111, %r105;


dp4a.s32.s32 %r113, %r186, %r115, %r109;


dp4a.s32.s32 %r117, %r190, %r119, %r113;

ld.global.nc.u32 %r121, [%rd40];

{.reg .f16 low,high;
mov.b32 {low,high},%r121;
cvt.f32.f16 %f56, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r121;
cvt.f32.f16 %f57, high;}


cvt.rn.f32.s32 %f81, %r117;
mul.ftz.f32 %f82, %f56, %f81;
mul.ftz.f32 %f83, %f57, 0f40800000;
sub.ftz.f32 %f84, %f82, %f83;
fma.rn.ftz.f32 %f238, %f55, %f84, %f238;

{ cvt.f32.f16 %f58, %rs2;}



dp4a.s32.s32 %r123, %r196, %r107, %r198;


dp4a.s32.s32 %r127, %r200, %r111, %r123;


dp4a.s32.s32 %r131, %r204, %r115, %r127;


dp4a.s32.s32 %r135, %r208, %r119, %r131;


{.reg .f16 low,high;
mov.b32 {low,high},%r121;
cvt.f32.f16 %f59, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r121;
cvt.f32.f16 %f60, high;}


cvt.rn.f32.s32 %f85, %r135;
mul.ftz.f32 %f86, %f59, %f85;
mul.ftz.f32 %f87, %f60, 0f40800000;
sub.ftz.f32 %f88, %f86, %f87;
fma.rn.ftz.f32 %f237, %f58, %f88, %f237;
cvt.u64.u32 %rd44, %r375;
add.s64 %rd45, %rd2, %rd44;
mul.lo.s64 %rd46, %rd45, 36;
add.s64 %rd47, %rd10, %rd46;
add.s64 %rd48, %rd47, 4;
add.s64 %rd49, %rd48, %rd30;
ld.global.nc.u32 %r143, [%rd49];
ld.global.nc.u32 %r147, [%rd49+16];
add.s64 %rd50, %rd48, %rd32;
ld.global.nc.u32 %r151, [%rd50];
ld.global.nc.u32 %r155, [%rd49+20];

{ cvt.f32.f16 %f61, %rs1;}



dp4a.s32.s32 %r141, %r178, %r143, %r198;


dp4a.s32.s32 %r145, %r182, %r147, %r141;


dp4a.s32.s32 %r149, %r186, %r151, %r145;


dp4a.s32.s32 %r153, %r190, %r155, %r149;

ld.global.nc.u32 %r157, [%rd47];

{.reg .f16 low,high;
mov.b32 {low,high},%r157;
cvt.f32.f16 %f62, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r157;
cvt.f32.f16 %f63, high;}


cvt.rn.f32.s32 %f89, %r153;
mul.ftz.f32 %f90, %f62, %f89;
mul.ftz.f32 %f91, %f63, 0f40800000;
sub.ftz.f32 %f92, %f90, %f91;
fma.rn.ftz.f32 %f236, %f61, %f92, %f236;

{ cvt.f32.f16 %f64, %rs2;}



dp4a.s32.s32 %r159, %r196, %r143, %r198;


dp4a.s32.s32 %r163, %r200, %r147, %r159;


dp4a.s32.s32 %r167, %r204, %r151, %r163;


dp4a.s32.s32 %r171, %r208, %r155, %r167;


{.reg .f16 low,high;
mov.b32 {low,high},%r157;
cvt.f32.f16 %f65, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r157;
cvt.f32.f16 %f66, high;}


cvt.rn.f32.s32 %f93, %r171;
mul.ftz.f32 %f94, %f65, %f93;
mul.ftz.f32 %f95, %f66, 0f40800000;
sub.ftz.f32 %f96, %f94, %f95;
fma.rn.ftz.f32 %f235, %f64, %f96, %f235;
cvt.u64.u32 %rd51, %r374;
add.s64 %rd52, %rd2, %rd51;
mul.lo.s64 %rd53, %rd52, 36;
add.s64 %rd54, %rd10, %rd53;
add.s64 %rd55, %rd54, 4;
add.s64 %rd56, %rd55, %rd30;
ld.global.nc.u32 %r179, [%rd56];
ld.global.nc.u32 %r183, [%rd56+16];
add.s64 %rd57, %rd55, %rd32;
ld.global.nc.u32 %r187, [%rd57];
ld.global.nc.u32 %r191, [%rd56+20];

{ cvt.f32.f16 %f67, %rs1;}



dp4a.s32.s32 %r177, %r178, %r179, %r198;


dp4a.s32.s32 %r181, %r182, %r183, %r177;


dp4a.s32.s32 %r185, %r186, %r187, %r181;


dp4a.s32.s32 %r189, %r190, %r191, %r185;

ld.global.nc.u32 %r193, [%rd54];

{.reg .f16 low,high;
mov.b32 {low,high},%r193;
cvt.f32.f16 %f68, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r193;
cvt.f32.f16 %f69, high;}


cvt.rn.f32.s32 %f97, %r189;
mul.ftz.f32 %f98, %f68, %f97;
mul.ftz.f32 %f99, %f69, 0f40800000;
sub.ftz.f32 %f100, %f98, %f99;
fma.rn.ftz.f32 %f234, %f67, %f100, %f234;

{ cvt.f32.f16 %f70, %rs2;}



dp4a.s32.s32 %r195, %r196, %r179, %r198;


dp4a.s32.s32 %r199, %r200, %r183, %r195;


dp4a.s32.s32 %r203, %r204, %r187, %r199;


dp4a.s32.s32 %r207, %r208, %r191, %r203;


{.reg .f16 low,high;
mov.b32 {low,high},%r193;
cvt.f32.f16 %f71, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r193;
cvt.f32.f16 %f72, high;}


cvt.rn.f32.s32 %f101, %r207;
mul.ftz.f32 %f102, %f71, %f101;
mul.ftz.f32 %f103, %f72, 0f40800000;
sub.ftz.f32 %f104, %f102, %f103;
fma.rn.ftz.f32 %f233, %f70, %f104, %f233;
add.s32 %r375, %r375, 64;
add.s32 %r374, %r374, 64;
add.s32 %r373, %r373, 64;
add.s64 %rd78, %rd78, 1152;
add.s32 %r376, %r376, 64;
setp.lt.s32 %p4, %r376, %r7;
@%p4 bra $L__BB36_2;

$L__BB36_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB36_5;

shl.b32 %r222, %r1, 10;
mov.u32 %r223, _ZZ13mul_mat_vec_qIL9ggml_type2ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r224, %r223, %r222;
shl.b32 %r225, %r39, 2;
add.s32 %r226, %r224, %r225;
st.shared.f32 [%r226+-1024], %f240;
st.shared.f32 [%r226+-896], %f239;
st.shared.f32 [%r226+-768], %f238;
st.shared.f32 [%r226+-640], %f237;
st.shared.f32 [%r226+-512], %f236;
st.shared.f32 [%r226+-384], %f235;
st.shared.f32 [%r226+-256], %f234;
st.shared.f32 [%r226+-128], %f233;

$L__BB36_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB36_22;

mad.lo.s32 %r227, %r4, %r30, %r3;
mad.lo.s32 %r228, %r5, %r36, %r227;
cvt.u64.u32 %rd13, %r228;
shl.b32 %r230, %r39, 2;
mov.u32 %r231, _ZZ13mul_mat_vec_qIL9ggml_type2ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r19, %r231, %r230;
mov.u32 %r232, 2;
ld.shared.f32 %f105, [%r19];
add.ftz.f32 %f106, %f105, %f240;
ld.shared.f32 %f107, [%r19+1024];
add.ftz.f32 %f108, %f107, %f106;
ld.shared.f32 %f109, [%r19+2048];
add.ftz.f32 %f110, %f109, %f108;
mov.b32 %r233, %f110;
mov.u32 %r234, 31;
mov.u32 %r235, 16;
mov.u32 %r236, -1;
shfl.sync.bfly.b32 %r237|%p7, %r233, %r235, %r234, %r236;
mov.b32 %f111, %r237;
add.ftz.f32 %f112, %f110, %f111;
mov.b32 %r238, %f112;
mov.u32 %r239, 8;
shfl.sync.bfly.b32 %r240|%p8, %r238, %r239, %r234, %r236;
mov.b32 %f113, %r240;
add.ftz.f32 %f114, %f112, %f113;
mov.b32 %r241, %f114;
mov.u32 %r242, 4;
shfl.sync.bfly.b32 %r243|%p9, %r241, %r242, %r234, %r236;
mov.b32 %f115, %r243;
add.ftz.f32 %f116, %f114, %f115;
mov.b32 %r244, %f116;
shfl.sync.bfly.b32 %r245|%p10, %r244, %r232, %r234, %r236;
mov.b32 %f117, %r245;
add.ftz.f32 %f118, %f116, %f117;
mov.b32 %r246, %f118;
mov.u32 %r247, 1;
shfl.sync.bfly.b32 %r248|%p11, %r246, %r247, %r234, %r236;
mov.b32 %f119, %r248;
add.ftz.f32 %f25, %f118, %f119;
setp.eq.s32 %p12, %r39, 0;
setp.lt.u32 %p13, %r3, %r24;
and.pred %p1, %p12, %p13;
mul.wide.u32 %rd58, %r228, 4;
add.s64 %rd14, %rd1, %rd58;
not.pred %p14, %p1;
@%p14 bra $L__BB36_8;

st.global.f32 [%rd14], %f25;

$L__BB36_8:
ld.shared.f32 %f120, [%r19+128];
add.ftz.f32 %f121, %f120, %f239;
ld.shared.f32 %f122, [%r19+1152];
add.ftz.f32 %f123, %f122, %f121;
ld.shared.f32 %f124, [%r19+2176];
add.ftz.f32 %f125, %f124, %f123;
mov.b32 %r249, %f125;
shfl.sync.bfly.b32 %r253|%p15, %r249, %r235, %r234, %r236;
mov.b32 %f126, %r253;
add.ftz.f32 %f127, %f125, %f126;
mov.b32 %r254, %f127;
shfl.sync.bfly.b32 %r256|%p16, %r254, %r239, %r234, %r236;
mov.b32 %f128, %r256;
add.ftz.f32 %f129, %f127, %f128;
mov.b32 %r257, %f129;
shfl.sync.bfly.b32 %r259|%p17, %r257, %r242, %r234, %r236;
mov.b32 %f130, %r259;
add.ftz.f32 %f131, %f129, %f130;
mov.b32 %r260, %f131;
shfl.sync.bfly.b32 %r262|%p18, %r260, %r232, %r234, %r236;
mov.b32 %f132, %r262;
add.ftz.f32 %f133, %f131, %f132;
mov.b32 %r263, %f133;
shfl.sync.bfly.b32 %r265|%p19, %r263, %r247, %r234, %r236;
mov.b32 %f134, %r265;
add.ftz.f32 %f26, %f133, %f134;
setp.eq.s32 %p20, %r39, 1;
add.s32 %r267, %r3, 1;
setp.lt.u32 %p21, %r267, %r24;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB36_10;

st.global.f32 [%rd14+4], %f26;

$L__BB36_10:
ld.shared.f32 %f135, [%r19+256];
add.ftz.f32 %f136, %f135, %f238;
ld.shared.f32 %f137, [%r19+1280];
add.ftz.f32 %f138, %f137, %f136;
ld.shared.f32 %f139, [%r19+2304];
add.ftz.f32 %f140, %f139, %f138;
mov.b32 %r268, %f140;
mov.u32 %r269, 31;
mov.u32 %r270, 16;
mov.u32 %r271, -1;
shfl.sync.bfly.b32 %r272|%p23, %r268, %r270, %r269, %r271;
mov.b32 %f141, %r272;
add.ftz.f32 %f142, %f140, %f141;
mov.b32 %r273, %f142;
mov.u32 %r274, 8;
shfl.sync.bfly.b32 %r275|%p24, %r273, %r274, %r269, %r271;
mov.b32 %f143, %r275;
add.ftz.f32 %f144, %f142, %f143;
mov.b32 %r276, %f144;
mov.u32 %r277, 4;
shfl.sync.bfly.b32 %r278|%p25, %r276, %r277, %r269, %r271;
mov.b32 %f145, %r278;
add.ftz.f32 %f146, %f144, %f145;
mov.b32 %r279, %f146;
mov.u32 %r280, 2;
shfl.sync.bfly.b32 %r281|%p26, %r279, %r280, %r269, %r271;
mov.b32 %f147, %r281;
add.ftz.f32 %f148, %f146, %f147;
mov.b32 %r282, %f148;
mov.u32 %r283, 1;
shfl.sync.bfly.b32 %r284|%p27, %r282, %r283, %r269, %r271;
mov.b32 %f149, %r284;
add.ftz.f32 %f27, %f148, %f149;
@%p14 bra $L__BB36_12;

cvt.u64.u32 %rd59, %r24;
add.s64 %rd60, %rd59, %rd13;
shl.b64 %rd61, %rd60, 2;
add.s64 %rd62, %rd1, %rd61;
st.global.f32 [%rd62], %f27;

$L__BB36_12:
ld.shared.f32 %f150, [%r19+384];
add.ftz.f32 %f151, %f150, %f237;
ld.shared.f32 %f152, [%r19+1408];
add.ftz.f32 %f153, %f152, %f151;
ld.shared.f32 %f154, [%r19+2432];
add.ftz.f32 %f155, %f154, %f153;
mov.b32 %r285, %f155;
shfl.sync.bfly.b32 %r289|%p29, %r285, %r270, %r269, %r271;
mov.b32 %f156, %r289;
add.ftz.f32 %f157, %f155, %f156;
mov.b32 %r290, %f157;
shfl.sync.bfly.b32 %r292|%p30, %r290, %r274, %r269, %r271;
mov.b32 %f158, %r292;
add.ftz.f32 %f159, %f157, %f158;
mov.b32 %r293, %f159;
shfl.sync.bfly.b32 %r295|%p31, %r293, %r277, %r269, %r271;
mov.b32 %f160, %r295;
add.ftz.f32 %f161, %f159, %f160;
mov.b32 %r296, %f161;
shfl.sync.bfly.b32 %r298|%p32, %r296, %r280, %r269, %r271;
mov.b32 %f162, %r298;
add.ftz.f32 %f163, %f161, %f162;
mov.b32 %r299, %f163;
shfl.sync.bfly.b32 %r301|%p33, %r299, %r283, %r269, %r271;
mov.b32 %f164, %r301;
add.ftz.f32 %f28, %f163, %f164;
@%p22 bra $L__BB36_14;

add.s32 %r302, %r24, 1;
cvt.u64.u32 %rd63, %r302;
add.s64 %rd64, %rd63, %rd13;
shl.b64 %rd65, %rd64, 2;
add.s64 %rd66, %rd1, %rd65;
st.global.f32 [%rd66], %f28;

$L__BB36_14:
shl.b32 %r303, %r24, 1;
mov.u32 %r304, 1;
ld.shared.f32 %f165, [%r19+512];
add.ftz.f32 %f166, %f165, %f236;
ld.shared.f32 %f167, [%r19+1536];
add.ftz.f32 %f168, %f167, %f166;
ld.shared.f32 %f169, [%r19+2560];
add.ftz.f32 %f170, %f169, %f168;
mov.b32 %r305, %f170;
mov.u32 %r306, 31;
mov.u32 %r307, 16;
mov.u32 %r308, -1;
shfl.sync.bfly.b32 %r309|%p35, %r305, %r307, %r306, %r308;
mov.b32 %f171, %r309;
add.ftz.f32 %f172, %f170, %f171;
mov.b32 %r310, %f172;
mov.u32 %r311, 8;
shfl.sync.bfly.b32 %r312|%p36, %r310, %r311, %r306, %r308;
mov.b32 %f173, %r312;
add.ftz.f32 %f174, %f172, %f173;
mov.b32 %r313, %f174;
mov.u32 %r314, 4;
shfl.sync.bfly.b32 %r315|%p37, %r313, %r314, %r306, %r308;
mov.b32 %f175, %r315;
add.ftz.f32 %f176, %f174, %f175;
mov.b32 %r316, %f176;
mov.u32 %r317, 2;
shfl.sync.bfly.b32 %r318|%p38, %r316, %r317, %r306, %r308;
mov.b32 %f177, %r318;
add.ftz.f32 %f178, %f176, %f177;
mov.b32 %r319, %f178;
shfl.sync.bfly.b32 %r320|%p39, %r319, %r304, %r306, %r308;
mov.b32 %f179, %r320;
add.ftz.f32 %f29, %f178, %f179;
cvt.u64.u32 %rd67, %r303;
add.s64 %rd68, %rd67, %rd13;
shl.b64 %rd69, %rd68, 2;
add.s64 %rd15, %rd1, %rd69;
@%p14 bra $L__BB36_16;

st.global.f32 [%rd15], %f29;

$L__BB36_16:
ld.shared.f32 %f180, [%r19+640];
add.ftz.f32 %f181, %f180, %f235;
ld.shared.f32 %f182, [%r19+1664];
add.ftz.f32 %f183, %f182, %f181;
ld.shared.f32 %f184, [%r19+2688];
add.ftz.f32 %f185, %f184, %f183;
mov.b32 %r321, %f185;
shfl.sync.bfly.b32 %r325|%p41, %r321, %r307, %r306, %r308;
mov.b32 %f186, %r325;
add.ftz.f32 %f187, %f185, %f186;
mov.b32 %r326, %f187;
shfl.sync.bfly.b32 %r328|%p42, %r326, %r311, %r306, %r308;
mov.b32 %f188, %r328;
add.ftz.f32 %f189, %f187, %f188;
mov.b32 %r329, %f189;
shfl.sync.bfly.b32 %r331|%p43, %r329, %r314, %r306, %r308;
mov.b32 %f190, %r331;
add.ftz.f32 %f191, %f189, %f190;
mov.b32 %r332, %f191;
shfl.sync.bfly.b32 %r334|%p44, %r332, %r317, %r306, %r308;
mov.b32 %f192, %r334;
add.ftz.f32 %f193, %f191, %f192;
mov.b32 %r335, %f193;
shfl.sync.bfly.b32 %r337|%p45, %r335, %r304, %r306, %r308;
mov.b32 %f194, %r337;
add.ftz.f32 %f30, %f193, %f194;
@%p22 bra $L__BB36_18;

st.global.f32 [%rd15+4], %f30;

$L__BB36_18:
mul.lo.s32 %r21, %r24, 3;
ld.shared.f32 %f195, [%r19+768];
add.ftz.f32 %f196, %f195, %f234;
ld.shared.f32 %f197, [%r19+1792];
add.ftz.f32 %f198, %f197, %f196;
ld.shared.f32 %f199, [%r19+2816];
add.ftz.f32 %f200, %f199, %f198;
mov.b32 %r338, %f200;
mov.u32 %r339, 31;
mov.u32 %r340, 16;
mov.u32 %r341, -1;
shfl.sync.bfly.b32 %r342|%p47, %r338, %r340, %r339, %r341;
mov.b32 %f201, %r342;
add.ftz.f32 %f202, %f200, %f201;
mov.b32 %r343, %f202;
mov.u32 %r344, 8;
shfl.sync.bfly.b32 %r345|%p48, %r343, %r344, %r339, %r341;
mov.b32 %f203, %r345;
add.ftz.f32 %f204, %f202, %f203;
mov.b32 %r346, %f204;
mov.u32 %r347, 4;
shfl.sync.bfly.b32 %r348|%p49, %r346, %r347, %r339, %r341;
mov.b32 %f205, %r348;
add.ftz.f32 %f206, %f204, %f205;
mov.b32 %r349, %f206;
mov.u32 %r350, 2;
shfl.sync.bfly.b32 %r351|%p50, %r349, %r350, %r339, %r341;
mov.b32 %f207, %r351;
add.ftz.f32 %f208, %f206, %f207;
mov.b32 %r352, %f208;
mov.u32 %r353, 1;
shfl.sync.bfly.b32 %r354|%p51, %r352, %r353, %r339, %r341;
mov.b32 %f209, %r354;
add.ftz.f32 %f31, %f208, %f209;
@%p14 bra $L__BB36_20;

cvt.u64.u32 %rd70, %r21;
add.s64 %rd71, %rd70, %rd13;
shl.b64 %rd72, %rd71, 2;
add.s64 %rd73, %rd1, %rd72;
st.global.f32 [%rd73], %f31;

$L__BB36_20:
ld.shared.f32 %f210, [%r19+896];
add.ftz.f32 %f211, %f210, %f233;
ld.shared.f32 %f212, [%r19+1920];
add.ftz.f32 %f213, %f212, %f211;
ld.shared.f32 %f214, [%r19+2944];
add.ftz.f32 %f215, %f214, %f213;
mov.b32 %r355, %f215;
shfl.sync.bfly.b32 %r359|%p53, %r355, %r340, %r339, %r341;
mov.b32 %f216, %r359;
add.ftz.f32 %f217, %f215, %f216;
mov.b32 %r360, %f217;
shfl.sync.bfly.b32 %r362|%p54, %r360, %r344, %r339, %r341;
mov.b32 %f218, %r362;
add.ftz.f32 %f219, %f217, %f218;
mov.b32 %r363, %f219;
shfl.sync.bfly.b32 %r365|%p55, %r363, %r347, %r339, %r341;
mov.b32 %f220, %r365;
add.ftz.f32 %f221, %f219, %f220;
mov.b32 %r366, %f221;
shfl.sync.bfly.b32 %r368|%p56, %r366, %r350, %r339, %r341;
mov.b32 %f222, %r368;
add.ftz.f32 %f223, %f221, %f222;
mov.b32 %r369, %f223;
shfl.sync.bfly.b32 %r371|%p57, %r369, %r353, %r339, %r341;
mov.b32 %f224, %r371;
add.ftz.f32 %f32, %f223, %f224;
@%p22 bra $L__BB36_22;

add.s32 %r372, %r21, 1;
cvt.u64.u32 %rd74, %r372;
add.s64 %rd75, %rd74, %rd13;
shl.b64 %rd76, %rd75, 2;
add.s64 %rd77, %rd1, %rd76;
st.global.f32 [%rd77], %f32;

$L__BB36_22:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type2ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type2ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type2ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type2ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type2ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type2ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type2ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type2ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type2ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<71>;
.reg .b16 %rs<19>;
.reg .f32 %f<261>;
.reg .b32 %r<454>;
.reg .b64 %rd<96>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type2ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1280];

ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type2ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type2ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type2ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type2ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type2ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type2ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type2ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type2ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type2ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type2ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type2ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type2ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type2ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type2ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type2ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type2ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type2ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r38, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r38, %r2;
mov.u32 %r39, %ctaid.x;
shl.b32 %r4, %r39, 1;
shr.u32 %r40, %r3, 31;
add.s32 %r41, %r3, %r40;
shr.s32 %r453, %r41, 1;
shr.u32 %r42, %r22, 5;
setp.ge.s32 %p3, %r453, %r42;
mov.f32 %f251, 0f00000000;
mov.f32 %f252, %f251;
mov.f32 %f253, %f251;
mov.f32 %f254, %f251;
mov.f32 %f255, %f251;
mov.f32 %f256, %f251;
mov.f32 %f257, %f251;
mov.f32 %f258, %f251;
mov.f32 %f259, %f251;
mov.f32 %f260, %f251;
@%p3 bra $L__BB37_3;

mov.u32 %r43, %ctaid.y;
mov.u32 %r44, %ctaid.z;
mul.lo.s32 %r45, %r43, %r30;
cvt.u64.u32 %rd19, %r45;
mul.lo.s32 %r46, %r44, %r36;
cvt.u64.u32 %rd20, %r46;
add.s64 %rd1, %rd20, %rd19;
mul.hi.u32 %r47, %r44, %r32;
add.s32 %r48, %r47, %r44;
shr.u32 %r49, %r48, %r33;
mul.lo.s32 %r50, %r49, %r35;
mul.hi.u32 %r51, %r43, %r26;
add.s32 %r52, %r51, %r43;
shr.u32 %r53, %r52, %r27;
mul.lo.s32 %r54, %r53, %r29;
cvta.to.global.u64 %rd95, %rd16;
and.b32 %r57, %r41, -2;
sub.s32 %r58, %r3, %r57;
shl.b32 %r59, %r58, 1;
shl.b32 %r60, %r58, 2;
or.b32 %r61, %r60, 1;
cvt.s64.s32 %rd3, %r59;
sub.s32 %r62, %r61, %r59;
cvt.s64.s32 %rd4, %r62;
shl.b32 %r63, %r24, 1;
add.s32 %r452, %r453, %r63;
mad.lo.s32 %r451, %r24, 3, %r453;
shl.b32 %r64, %r24, 2;
add.s32 %r450, %r453, %r64;
add.s32 %r449, %r453, %r24;
add.s32 %r65, %r453, %r50;
add.s32 %r66, %r65, %r54;
add.s32 %r67, %r4, 1;
mad.lo.s32 %r68, %r23, %r67, %r66;
mul.wide.s32 %rd5, %r68, 18;
mad.lo.s32 %r69, %r4, %r23, %r453;
add.s32 %r70, %r69, %r50;
add.s32 %r71, %r70, %r54;
mul.wide.s32 %rd6, %r71, 18;
mul.wide.s32 %rd21, %r60, 2;
add.s64 %rd22, %rd6, %rd21;
add.s64 %rd7, %rd22, 4;
add.s64 %rd23, %rd5, %rd21;
add.s64 %rd8, %rd23, 2;
cvta.to.global.u64 %rd9, %rd17;
mov.f32 %f251, 0f00000000;
shl.b64 %rd30, %rd3, 2;
shl.b64 %rd32, %rd4, 2;

$L__BB37_2:
cvt.u64.u32 %rd24, %r453;
add.s64 %rd25, %rd1, %rd24;
mul.lo.s64 %rd26, %rd25, 36;
add.s64 %rd27, %rd9, %rd26;
add.s64 %rd28, %rd27, 4;
add.s64 %rd29, %rd95, %rd7;
ld.global.nc.u16 %rs11, [%rd29+-2];
ld.global.nc.u16 %rs12, [%rd29];
mov.b32 %r252, {%rs11, %rs12};
add.s64 %rd31, %rd28, %rd30;
ld.global.nc.u32 %r74, [%rd31];
ld.global.nc.u32 %r78, [%rd31+16];
ld.global.nc.u16 %rs13, [%rd29+2];
ld.global.nc.u16 %rs14, [%rd29+4];
mov.b32 %r253, {%rs13, %rs14};
add.s64 %rd33, %rd28, %rd32;
ld.global.nc.u32 %r82, [%rd33];
ld.global.nc.u32 %r86, [%rd31+20];
add.s64 %rd34, %rd95, %rd6;
ld.global.nc.u16 %rs1, [%rd34];

{ cvt.f32.f16 %f61, %rs1;}


and.b32 %r217, %r252, 252645135;
shr.u32 %r254, %r252, 4;
and.b32 %r221, %r254, 252645135;
mov.u32 %r237, 0;

dp4a.s32.s32 %r72, %r217, %r74, %r237;


dp4a.s32.s32 %r76, %r221, %r78, %r72;

and.b32 %r225, %r253, 252645135;
shr.u32 %r255, %r253, 4;
and.b32 %r229, %r255, 252645135;

dp4a.s32.s32 %r80, %r225, %r82, %r76;


dp4a.s32.s32 %r84, %r229, %r86, %r80;

ld.global.nc.u32 %r88, [%rd27];

{.reg .f16 low,high;
mov.b32 {low,high},%r88;
cvt.f32.f16 %f62, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r88;
cvt.f32.f16 %f63, high;}


cvt.rn.f32.s32 %f91, %r84;
mul.ftz.f32 %f92, %f62, %f91;
mul.ftz.f32 %f93, %f63, 0f40800000;
sub.ftz.f32 %f94, %f92, %f93;
fma.rn.ftz.f32 %f260, %f61, %f94, %f260;
add.s64 %rd35, %rd95, %rd8;
ld.global.nc.u16 %rs15, [%rd35];
ld.global.nc.u16 %rs16, [%rd35+2];
mov.b32 %r256, {%rs15, %rs16};
ld.global.nc.u16 %rs17, [%rd35+4];
ld.global.nc.u16 %rs18, [%rd35+6];
mov.b32 %r257, {%rs17, %rs18};
add.s64 %rd36, %rd95, %rd5;
ld.global.nc.u16 %rs2, [%rd36];

{ cvt.f32.f16 %f64, %rs2;}


and.b32 %r235, %r256, 252645135;
shr.u32 %r258, %r256, 4;
and.b32 %r239, %r258, 252645135;

dp4a.s32.s32 %r90, %r235, %r74, %r237;


dp4a.s32.s32 %r94, %r239, %r78, %r90;

and.b32 %r243, %r257, 252645135;
shr.u32 %r259, %r257, 4;
and.b32 %r247, %r259, 252645135;

dp4a.s32.s32 %r98, %r243, %r82, %r94;


dp4a.s32.s32 %r102, %r247, %r86, %r98;


{.reg .f16 low,high;
mov.b32 {low,high},%r88;
cvt.f32.f16 %f65, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r88;
cvt.f32.f16 %f66, high;}


cvt.rn.f32.s32 %f95, %r102;
mul.ftz.f32 %f96, %f65, %f95;
mul.ftz.f32 %f97, %f66, 0f40800000;
sub.ftz.f32 %f98, %f96, %f97;
fma.rn.ftz.f32 %f259, %f64, %f98, %f259;
cvt.u64.u32 %rd37, %r449;
add.s64 %rd38, %rd1, %rd37;
mul.lo.s64 %rd39, %rd38, 36;
add.s64 %rd40, %rd9, %rd39;
add.s64 %rd41, %rd40, 4;
add.s64 %rd42, %rd41, %rd30;
ld.global.nc.u32 %r110, [%rd42];
ld.global.nc.u32 %r114, [%rd42+16];
add.s64 %rd43, %rd41, %rd32;
ld.global.nc.u32 %r118, [%rd43];
ld.global.nc.u32 %r122, [%rd42+20];

{ cvt.f32.f16 %f67, %rs1;}



dp4a.s32.s32 %r108, %r217, %r110, %r237;


dp4a.s32.s32 %r112, %r221, %r114, %r108;


dp4a.s32.s32 %r116, %r225, %r118, %r112;


dp4a.s32.s32 %r120, %r229, %r122, %r116;

ld.global.nc.u32 %r124, [%rd40];

{.reg .f16 low,high;
mov.b32 {low,high},%r124;
cvt.f32.f16 %f68, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r124;
cvt.f32.f16 %f69, high;}


cvt.rn.f32.s32 %f99, %r120;
mul.ftz.f32 %f100, %f68, %f99;
mul.ftz.f32 %f101, %f69, 0f40800000;
sub.ftz.f32 %f102, %f100, %f101;
fma.rn.ftz.f32 %f258, %f67, %f102, %f258;

{ cvt.f32.f16 %f70, %rs2;}



dp4a.s32.s32 %r126, %r235, %r110, %r237;


dp4a.s32.s32 %r130, %r239, %r114, %r126;


dp4a.s32.s32 %r134, %r243, %r118, %r130;


dp4a.s32.s32 %r138, %r247, %r122, %r134;


{.reg .f16 low,high;
mov.b32 {low,high},%r124;
cvt.f32.f16 %f71, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r124;
cvt.f32.f16 %f72, high;}


cvt.rn.f32.s32 %f103, %r138;
mul.ftz.f32 %f104, %f71, %f103;
mul.ftz.f32 %f105, %f72, 0f40800000;
sub.ftz.f32 %f106, %f104, %f105;
fma.rn.ftz.f32 %f257, %f70, %f106, %f257;
cvt.u64.u32 %rd44, %r452;
add.s64 %rd45, %rd1, %rd44;
mul.lo.s64 %rd46, %rd45, 36;
add.s64 %rd47, %rd9, %rd46;
add.s64 %rd48, %rd47, 4;
add.s64 %rd49, %rd48, %rd30;
ld.global.nc.u32 %r146, [%rd49];
ld.global.nc.u32 %r150, [%rd49+16];
add.s64 %rd50, %rd48, %rd32;
ld.global.nc.u32 %r154, [%rd50];
ld.global.nc.u32 %r158, [%rd49+20];

{ cvt.f32.f16 %f73, %rs1;}



dp4a.s32.s32 %r144, %r217, %r146, %r237;


dp4a.s32.s32 %r148, %r221, %r150, %r144;


dp4a.s32.s32 %r152, %r225, %r154, %r148;


dp4a.s32.s32 %r156, %r229, %r158, %r152;

ld.global.nc.u32 %r160, [%rd47];

{.reg .f16 low,high;
mov.b32 {low,high},%r160;
cvt.f32.f16 %f74, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r160;
cvt.f32.f16 %f75, high;}


cvt.rn.f32.s32 %f107, %r156;
mul.ftz.f32 %f108, %f74, %f107;
mul.ftz.f32 %f109, %f75, 0f40800000;
sub.ftz.f32 %f110, %f108, %f109;
fma.rn.ftz.f32 %f256, %f73, %f110, %f256;

{ cvt.f32.f16 %f76, %rs2;}



dp4a.s32.s32 %r162, %r235, %r146, %r237;


dp4a.s32.s32 %r166, %r239, %r150, %r162;


dp4a.s32.s32 %r170, %r243, %r154, %r166;


dp4a.s32.s32 %r174, %r247, %r158, %r170;


{.reg .f16 low,high;
mov.b32 {low,high},%r160;
cvt.f32.f16 %f77, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r160;
cvt.f32.f16 %f78, high;}


cvt.rn.f32.s32 %f111, %r174;
mul.ftz.f32 %f112, %f77, %f111;
mul.ftz.f32 %f113, %f78, 0f40800000;
sub.ftz.f32 %f114, %f112, %f113;
fma.rn.ftz.f32 %f255, %f76, %f114, %f255;
cvt.u64.u32 %rd51, %r451;
add.s64 %rd52, %rd1, %rd51;
mul.lo.s64 %rd53, %rd52, 36;
add.s64 %rd54, %rd9, %rd53;
add.s64 %rd55, %rd54, 4;
add.s64 %rd56, %rd55, %rd30;
ld.global.nc.u32 %r182, [%rd56];
ld.global.nc.u32 %r186, [%rd56+16];
add.s64 %rd57, %rd55, %rd32;
ld.global.nc.u32 %r190, [%rd57];
ld.global.nc.u32 %r194, [%rd56+20];

{ cvt.f32.f16 %f79, %rs1;}



dp4a.s32.s32 %r180, %r217, %r182, %r237;


dp4a.s32.s32 %r184, %r221, %r186, %r180;


dp4a.s32.s32 %r188, %r225, %r190, %r184;


dp4a.s32.s32 %r192, %r229, %r194, %r188;

ld.global.nc.u32 %r196, [%rd54];

{.reg .f16 low,high;
mov.b32 {low,high},%r196;
cvt.f32.f16 %f80, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r196;
cvt.f32.f16 %f81, high;}


cvt.rn.f32.s32 %f115, %r192;
mul.ftz.f32 %f116, %f80, %f115;
mul.ftz.f32 %f117, %f81, 0f40800000;
sub.ftz.f32 %f118, %f116, %f117;
fma.rn.ftz.f32 %f254, %f79, %f118, %f254;

{ cvt.f32.f16 %f82, %rs2;}



dp4a.s32.s32 %r198, %r235, %r182, %r237;


dp4a.s32.s32 %r202, %r239, %r186, %r198;


dp4a.s32.s32 %r206, %r243, %r190, %r202;


dp4a.s32.s32 %r210, %r247, %r194, %r206;


{.reg .f16 low,high;
mov.b32 {low,high},%r196;
cvt.f32.f16 %f83, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r196;
cvt.f32.f16 %f84, high;}


cvt.rn.f32.s32 %f119, %r210;
mul.ftz.f32 %f120, %f83, %f119;
mul.ftz.f32 %f121, %f84, 0f40800000;
sub.ftz.f32 %f122, %f120, %f121;
fma.rn.ftz.f32 %f253, %f82, %f122, %f253;
cvt.u64.u32 %rd58, %r450;
add.s64 %rd59, %rd1, %rd58;
mul.lo.s64 %rd60, %rd59, 36;
add.s64 %rd61, %rd9, %rd60;
add.s64 %rd62, %rd61, 4;
add.s64 %rd63, %rd62, %rd30;
ld.global.nc.u32 %r218, [%rd63];
ld.global.nc.u32 %r222, [%rd63+16];
add.s64 %rd64, %rd62, %rd32;
ld.global.nc.u32 %r226, [%rd64];
ld.global.nc.u32 %r230, [%rd63+20];

{ cvt.f32.f16 %f85, %rs1;}



dp4a.s32.s32 %r216, %r217, %r218, %r237;


dp4a.s32.s32 %r220, %r221, %r222, %r216;


dp4a.s32.s32 %r224, %r225, %r226, %r220;


dp4a.s32.s32 %r228, %r229, %r230, %r224;

ld.global.nc.u32 %r232, [%rd61];

{.reg .f16 low,high;
mov.b32 {low,high},%r232;
cvt.f32.f16 %f86, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r232;
cvt.f32.f16 %f87, high;}


cvt.rn.f32.s32 %f123, %r228;
mul.ftz.f32 %f124, %f86, %f123;
mul.ftz.f32 %f125, %f87, 0f40800000;
sub.ftz.f32 %f126, %f124, %f125;
fma.rn.ftz.f32 %f252, %f85, %f126, %f252;

{ cvt.f32.f16 %f88, %rs2;}



dp4a.s32.s32 %r234, %r235, %r218, %r237;


dp4a.s32.s32 %r238, %r239, %r222, %r234;


dp4a.s32.s32 %r242, %r243, %r226, %r238;


dp4a.s32.s32 %r246, %r247, %r230, %r242;


{.reg .f16 low,high;
mov.b32 {low,high},%r232;
cvt.f32.f16 %f89, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r232;
cvt.f32.f16 %f90, high;}


cvt.rn.f32.s32 %f127, %r246;
mul.ftz.f32 %f128, %f89, %f127;
mul.ftz.f32 %f129, %f90, 0f40800000;
sub.ftz.f32 %f130, %f128, %f129;
fma.rn.ftz.f32 %f251, %f88, %f130, %f251;
add.s32 %r452, %r452, 32;
add.s32 %r451, %r451, 32;
add.s32 %r450, %r450, 32;
add.s32 %r449, %r449, 32;
add.s64 %rd95, %rd95, 576;
add.s32 %r453, %r453, 32;
setp.lt.s32 %p4, %r453, %r42;
@%p4 bra $L__BB37_2;

$L__BB37_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB37_5;

mov.u32 %r261, _ZZ13mul_mat_vec_qIL9ggml_type2ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r262, %r1, 1280, %r261;
shl.b32 %r263, %r2, 2;
add.s32 %r264, %r262, %r263;
st.shared.f32 [%r264+-1280], %f260;
st.shared.f32 [%r264+-1152], %f259;
st.shared.f32 [%r264+-1024], %f258;
st.shared.f32 [%r264+-896], %f257;
st.shared.f32 [%r264+-768], %f256;
st.shared.f32 [%r264+-640], %f255;
st.shared.f32 [%r264+-512], %f254;
st.shared.f32 [%r264+-384], %f253;
st.shared.f32 [%r264+-256], %f252;
st.shared.f32 [%r264+-128], %f251;

$L__BB37_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB37_26;

mov.u32 %r265, %ctaid.z;
mov.u32 %r266, %ctaid.y;
mad.lo.s32 %r267, %r266, %r31, %r4;
mad.lo.s32 %r268, %r265, %r37, %r267;
cvt.u64.u32 %rd12, %r268;
shl.b32 %r269, %r2, 2;
mov.u32 %r270, _ZZ13mul_mat_vec_qIL9ggml_type2ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r20, %r270, %r269;
mov.u32 %r271, 2;
ld.shared.f32 %f131, [%r20];
add.ftz.f32 %f132, %f131, %f260;
mov.b32 %r272, %f132;
mov.u32 %r273, 31;
mov.u32 %r274, 16;
mov.u32 %r275, -1;
shfl.sync.bfly.b32 %r276|%p7, %r272, %r274, %r273, %r275;
mov.b32 %f133, %r276;
add.ftz.f32 %f134, %f132, %f133;
mov.b32 %r277, %f134;
mov.u32 %r278, 8;
shfl.sync.bfly.b32 %r279|%p8, %r277, %r278, %r273, %r275;
mov.b32 %f135, %r279;
add.ftz.f32 %f136, %f134, %f135;
mov.b32 %r280, %f136;
mov.u32 %r281, 4;
shfl.sync.bfly.b32 %r282|%p9, %r280, %r281, %r273, %r275;
mov.b32 %f137, %r282;
add.ftz.f32 %f138, %f136, %f137;
mov.b32 %r283, %f138;
shfl.sync.bfly.b32 %r284|%p10, %r283, %r271, %r273, %r275;
mov.b32 %f139, %r284;
add.ftz.f32 %f140, %f138, %f139;
mov.b32 %r285, %f140;
mov.u32 %r286, 1;
shfl.sync.bfly.b32 %r287|%p11, %r285, %r286, %r273, %r275;
mov.b32 %f141, %r287;
add.ftz.f32 %f31, %f140, %f141;
setp.lt.u32 %p12, %r4, %r25;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
cvta.to.global.u64 %rd65, %rd18;
mul.wide.u32 %rd66, %r268, 4;
add.s64 %rd13, %rd65, %rd66;
not.pred %p14, %p1;
@%p14 bra $L__BB37_8;

st.global.f32 [%rd13], %f31;

$L__BB37_8:
ld.shared.f32 %f142, [%r20+128];
add.ftz.f32 %f143, %f142, %f259;
mov.b32 %r288, %f143;
shfl.sync.bfly.b32 %r292|%p15, %r288, %r274, %r273, %r275;
mov.b32 %f144, %r292;
add.ftz.f32 %f145, %f143, %f144;
mov.b32 %r293, %f145;
shfl.sync.bfly.b32 %r295|%p16, %r293, %r278, %r273, %r275;
mov.b32 %f146, %r295;
add.ftz.f32 %f147, %f145, %f146;
mov.b32 %r296, %f147;
shfl.sync.bfly.b32 %r298|%p17, %r296, %r281, %r273, %r275;
mov.b32 %f148, %r298;
add.ftz.f32 %f149, %f147, %f148;
mov.b32 %r299, %f149;
shfl.sync.bfly.b32 %r301|%p18, %r299, %r271, %r273, %r275;
mov.b32 %f150, %r301;
add.ftz.f32 %f151, %f149, %f150;
mov.b32 %r302, %f151;
shfl.sync.bfly.b32 %r304|%p19, %r302, %r286, %r273, %r275;
mov.b32 %f152, %r304;
add.ftz.f32 %f32, %f151, %f152;
setp.eq.s32 %p20, %r2, 1;
or.b32 %r308, %r4, 1;
setp.lt.u32 %p21, %r308, %r25;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB37_10;

st.global.f32 [%rd13+4], %f32;

$L__BB37_10:
ld.shared.f32 %f153, [%r20+256];
add.ftz.f32 %f154, %f153, %f258;
mov.b32 %r309, %f154;
mov.u32 %r310, 31;
mov.u32 %r311, 16;
mov.u32 %r312, -1;
shfl.sync.bfly.b32 %r313|%p23, %r309, %r311, %r310, %r312;
mov.b32 %f155, %r313;
add.ftz.f32 %f156, %f154, %f155;
mov.b32 %r314, %f156;
mov.u32 %r315, 8;
shfl.sync.bfly.b32 %r316|%p24, %r314, %r315, %r310, %r312;
mov.b32 %f157, %r316;
add.ftz.f32 %f158, %f156, %f157;
mov.b32 %r317, %f158;
mov.u32 %r318, 4;
shfl.sync.bfly.b32 %r319|%p25, %r317, %r318, %r310, %r312;
mov.b32 %f159, %r319;
add.ftz.f32 %f160, %f158, %f159;
mov.b32 %r320, %f160;
mov.u32 %r321, 2;
shfl.sync.bfly.b32 %r322|%p26, %r320, %r321, %r310, %r312;
mov.b32 %f161, %r322;
add.ftz.f32 %f162, %f160, %f161;
mov.b32 %r323, %f162;
mov.u32 %r324, 1;
shfl.sync.bfly.b32 %r325|%p27, %r323, %r324, %r310, %r312;
mov.b32 %f163, %r325;
add.ftz.f32 %f33, %f162, %f163;
@%p14 bra $L__BB37_12;

cvt.u64.u32 %rd67, %r25;
add.s64 %rd68, %rd67, %rd12;
shl.b64 %rd70, %rd68, 2;
add.s64 %rd71, %rd65, %rd70;
st.global.f32 [%rd71], %f33;

$L__BB37_12:
ld.shared.f32 %f164, [%r20+384];
add.ftz.f32 %f165, %f164, %f257;
mov.b32 %r326, %f165;
shfl.sync.bfly.b32 %r330|%p29, %r326, %r311, %r310, %r312;
mov.b32 %f166, %r330;
add.ftz.f32 %f167, %f165, %f166;
mov.b32 %r331, %f167;
shfl.sync.bfly.b32 %r333|%p30, %r331, %r315, %r310, %r312;
mov.b32 %f168, %r333;
add.ftz.f32 %f169, %f167, %f168;
mov.b32 %r334, %f169;
shfl.sync.bfly.b32 %r336|%p31, %r334, %r318, %r310, %r312;
mov.b32 %f170, %r336;
add.ftz.f32 %f171, %f169, %f170;
mov.b32 %r337, %f171;
shfl.sync.bfly.b32 %r339|%p32, %r337, %r321, %r310, %r312;
mov.b32 %f172, %r339;
add.ftz.f32 %f173, %f171, %f172;
mov.b32 %r340, %f173;
shfl.sync.bfly.b32 %r342|%p33, %r340, %r324, %r310, %r312;
mov.b32 %f174, %r342;
add.ftz.f32 %f34, %f173, %f174;
@%p22 bra $L__BB37_14;

add.s32 %r343, %r25, 1;
cvt.u64.u32 %rd72, %r343;
add.s64 %rd73, %rd72, %rd12;
shl.b64 %rd75, %rd73, 2;
add.s64 %rd76, %rd65, %rd75;
st.global.f32 [%rd76], %f34;

$L__BB37_14:
shl.b32 %r344, %r25, 1;
mov.u32 %r345, 1;
ld.shared.f32 %f175, [%r20+512];
add.ftz.f32 %f176, %f175, %f256;
mov.b32 %r346, %f176;
mov.u32 %r347, 31;
mov.u32 %r348, 16;
mov.u32 %r349, -1;
shfl.sync.bfly.b32 %r350|%p35, %r346, %r348, %r347, %r349;
mov.b32 %f177, %r350;
add.ftz.f32 %f178, %f176, %f177;
mov.b32 %r351, %f178;
mov.u32 %r352, 8;
shfl.sync.bfly.b32 %r353|%p36, %r351, %r352, %r347, %r349;
mov.b32 %f179, %r353;
add.ftz.f32 %f180, %f178, %f179;
mov.b32 %r354, %f180;
mov.u32 %r355, 4;
shfl.sync.bfly.b32 %r356|%p37, %r354, %r355, %r347, %r349;
mov.b32 %f181, %r356;
add.ftz.f32 %f182, %f180, %f181;
mov.b32 %r357, %f182;
mov.u32 %r358, 2;
shfl.sync.bfly.b32 %r359|%p38, %r357, %r358, %r347, %r349;
mov.b32 %f183, %r359;
add.ftz.f32 %f184, %f182, %f183;
mov.b32 %r360, %f184;
shfl.sync.bfly.b32 %r361|%p39, %r360, %r345, %r347, %r349;
mov.b32 %f185, %r361;
add.ftz.f32 %f35, %f184, %f185;
cvt.u64.u32 %rd77, %r344;
add.s64 %rd78, %rd77, %rd12;
shl.b64 %rd80, %rd78, 2;
add.s64 %rd14, %rd65, %rd80;
@%p14 bra $L__BB37_16;

st.global.f32 [%rd14], %f35;

$L__BB37_16:
ld.shared.f32 %f186, [%r20+640];
add.ftz.f32 %f187, %f186, %f255;
mov.b32 %r362, %f187;
shfl.sync.bfly.b32 %r366|%p41, %r362, %r348, %r347, %r349;
mov.b32 %f188, %r366;
add.ftz.f32 %f189, %f187, %f188;
mov.b32 %r367, %f189;
shfl.sync.bfly.b32 %r369|%p42, %r367, %r352, %r347, %r349;
mov.b32 %f190, %r369;
add.ftz.f32 %f191, %f189, %f190;
mov.b32 %r370, %f191;
shfl.sync.bfly.b32 %r372|%p43, %r370, %r355, %r347, %r349;
mov.b32 %f192, %r372;
add.ftz.f32 %f193, %f191, %f192;
mov.b32 %r373, %f193;
shfl.sync.bfly.b32 %r375|%p44, %r373, %r358, %r347, %r349;
mov.b32 %f194, %r375;
add.ftz.f32 %f195, %f193, %f194;
mov.b32 %r376, %f195;
shfl.sync.bfly.b32 %r378|%p45, %r376, %r345, %r347, %r349;
mov.b32 %f196, %r378;
add.ftz.f32 %f36, %f195, %f196;
@%p22 bra $L__BB37_18;

st.global.f32 [%rd14+4], %f36;

$L__BB37_18:
mul.lo.s32 %r21, %r25, 3;
ld.shared.f32 %f197, [%r20+768];
add.ftz.f32 %f198, %f197, %f254;
mov.b32 %r379, %f198;
mov.u32 %r380, 31;
mov.u32 %r381, 16;
mov.u32 %r382, -1;
shfl.sync.bfly.b32 %r383|%p47, %r379, %r381, %r380, %r382;
mov.b32 %f199, %r383;
add.ftz.f32 %f200, %f198, %f199;
mov.b32 %r384, %f200;
mov.u32 %r385, 8;
shfl.sync.bfly.b32 %r386|%p48, %r384, %r385, %r380, %r382;
mov.b32 %f201, %r386;
add.ftz.f32 %f202, %f200, %f201;
mov.b32 %r387, %f202;
mov.u32 %r388, 4;
shfl.sync.bfly.b32 %r389|%p49, %r387, %r388, %r380, %r382;
mov.b32 %f203, %r389;
add.ftz.f32 %f204, %f202, %f203;
mov.b32 %r390, %f204;
mov.u32 %r391, 2;
shfl.sync.bfly.b32 %r392|%p50, %r390, %r391, %r380, %r382;
mov.b32 %f205, %r392;
add.ftz.f32 %f206, %f204, %f205;
mov.b32 %r393, %f206;
mov.u32 %r394, 1;
shfl.sync.bfly.b32 %r395|%p51, %r393, %r394, %r380, %r382;
mov.b32 %f207, %r395;
add.ftz.f32 %f37, %f206, %f207;
@%p14 bra $L__BB37_20;

cvt.u64.u32 %rd81, %r21;
add.s64 %rd82, %rd81, %rd12;
shl.b64 %rd84, %rd82, 2;
add.s64 %rd85, %rd65, %rd84;
st.global.f32 [%rd85], %f37;

$L__BB37_20:
ld.shared.f32 %f208, [%r20+896];
add.ftz.f32 %f209, %f208, %f253;
mov.b32 %r396, %f209;
shfl.sync.bfly.b32 %r400|%p53, %r396, %r381, %r380, %r382;
mov.b32 %f210, %r400;
add.ftz.f32 %f211, %f209, %f210;
mov.b32 %r401, %f211;
shfl.sync.bfly.b32 %r403|%p54, %r401, %r385, %r380, %r382;
mov.b32 %f212, %r403;
add.ftz.f32 %f213, %f211, %f212;
mov.b32 %r404, %f213;
shfl.sync.bfly.b32 %r406|%p55, %r404, %r388, %r380, %r382;
mov.b32 %f214, %r406;
add.ftz.f32 %f215, %f213, %f214;
mov.b32 %r407, %f215;
shfl.sync.bfly.b32 %r409|%p56, %r407, %r391, %r380, %r382;
mov.b32 %f216, %r409;
add.ftz.f32 %f217, %f215, %f216;
mov.b32 %r410, %f217;
shfl.sync.bfly.b32 %r412|%p57, %r410, %r394, %r380, %r382;
mov.b32 %f218, %r412;
add.ftz.f32 %f38, %f217, %f218;
@%p22 bra $L__BB37_22;

add.s32 %r413, %r21, 1;
cvt.u64.u32 %rd86, %r413;
add.s64 %rd87, %rd86, %rd12;
shl.b64 %rd89, %rd87, 2;
add.s64 %rd90, %rd65, %rd89;
st.global.f32 [%rd90], %f38;

$L__BB37_22:
shl.b32 %r414, %r25, 2;
mov.u32 %r415, 2;
ld.shared.f32 %f219, [%r20+1024];
add.ftz.f32 %f220, %f219, %f252;
mov.b32 %r416, %f220;
mov.u32 %r417, 31;
mov.u32 %r418, 16;
mov.u32 %r419, -1;
shfl.sync.bfly.b32 %r420|%p59, %r416, %r418, %r417, %r419;
mov.b32 %f221, %r420;
add.ftz.f32 %f222, %f220, %f221;
mov.b32 %r421, %f222;
mov.u32 %r422, 8;
shfl.sync.bfly.b32 %r423|%p60, %r421, %r422, %r417, %r419;
mov.b32 %f223, %r423;
add.ftz.f32 %f224, %f222, %f223;
mov.b32 %r424, %f224;
mov.u32 %r425, 4;
shfl.sync.bfly.b32 %r426|%p61, %r424, %r425, %r417, %r419;
mov.b32 %f225, %r426;
add.ftz.f32 %f226, %f224, %f225;
mov.b32 %r427, %f226;
shfl.sync.bfly.b32 %r428|%p62, %r427, %r415, %r417, %r419;
mov.b32 %f227, %r428;
add.ftz.f32 %f228, %f226, %f227;
mov.b32 %r429, %f228;
mov.u32 %r430, 1;
shfl.sync.bfly.b32 %r431|%p63, %r429, %r430, %r417, %r419;
mov.b32 %f229, %r431;
add.ftz.f32 %f39, %f228, %f229;
cvt.u64.u32 %rd91, %r414;
add.s64 %rd92, %rd91, %rd12;
shl.b64 %rd94, %rd92, 2;
add.s64 %rd15, %rd65, %rd94;
@%p14 bra $L__BB37_24;

st.global.f32 [%rd15], %f39;

$L__BB37_24:
ld.shared.f32 %f230, [%r20+1152];
add.ftz.f32 %f231, %f230, %f251;
mov.b32 %r432, %f231;
shfl.sync.bfly.b32 %r436|%p65, %r432, %r418, %r417, %r419;
mov.b32 %f232, %r436;
add.ftz.f32 %f233, %f231, %f232;
mov.b32 %r437, %f233;
shfl.sync.bfly.b32 %r439|%p66, %r437, %r422, %r417, %r419;
mov.b32 %f234, %r439;
add.ftz.f32 %f235, %f233, %f234;
mov.b32 %r440, %f235;
shfl.sync.bfly.b32 %r442|%p67, %r440, %r425, %r417, %r419;
mov.b32 %f236, %r442;
add.ftz.f32 %f237, %f235, %f236;
mov.b32 %r443, %f237;
shfl.sync.bfly.b32 %r445|%p68, %r443, %r415, %r417, %r419;
mov.b32 %f238, %r445;
add.ftz.f32 %f239, %f237, %f238;
mov.b32 %r446, %f239;
shfl.sync.bfly.b32 %r448|%p69, %r446, %r430, %r417, %r419;
mov.b32 %f240, %r448;
add.ftz.f32 %f40, %f239, %f240;
@%p22 bra $L__BB37_26;

st.global.f32 [%rd15+4], %f40;

$L__BB37_26:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type2ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type2ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type2ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type2ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type2ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type2ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type2ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type2ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type2ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<83>;
.reg .b16 %rs<21>;
.reg .f32 %f<313>;
.reg .b32 %r<541>;
.reg .b64 %rd<116>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type2ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type2ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type2ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type2ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type2ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type2ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type2ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type2ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type2ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type2ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type2ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type2ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type2ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type2ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type2ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type2ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type2ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type2ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r43, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r43, %r2;
mov.u32 %r44, %ctaid.x;
shl.b32 %r4, %r44, 1;
shr.u32 %r45, %r3, 31;
add.s32 %r46, %r3, %r45;
shr.s32 %r540, %r46, 1;
shr.u32 %r6, %r42, 5;
setp.ge.s32 %p3, %r540, %r6;
mov.f32 %f301, 0f00000000;
mov.f32 %f302, %f301;
mov.f32 %f303, %f301;
mov.f32 %f304, %f301;
mov.f32 %f305, %f301;
mov.f32 %f306, %f301;
mov.f32 %f307, %f301;
mov.f32 %f308, %f301;
mov.f32 %f309, %f301;
mov.f32 %f310, %f301;
mov.f32 %f311, %f301;
mov.f32 %f312, %f301;
@%p3 bra $L__BB38_3;

mov.u32 %r47, %ctaid.y;
mov.u32 %r48, %ctaid.z;
mul.lo.s32 %r49, %r47, %r34;
cvt.u64.u32 %rd18, %r49;
mul.lo.s32 %r50, %r48, %r40;
cvt.u64.u32 %rd19, %r50;
add.s64 %rd1, %rd19, %rd18;
mul.hi.u32 %r51, %r48, %r36;
add.s32 %r52, %r51, %r48;
shr.u32 %r53, %r52, %r37;
mul.lo.s32 %r54, %r53, %r39;
mul.hi.u32 %r55, %r47, %r30;
add.s32 %r56, %r55, %r47;
shr.u32 %r57, %r56, %r31;
mul.lo.s32 %r58, %r57, %r33;
cvta.to.global.u64 %rd115, %rd15;
and.b32 %r61, %r46, -2;
sub.s32 %r62, %r3, %r61;
shl.b32 %r63, %r62, 1;
shl.b32 %r64, %r62, 2;
or.b32 %r65, %r64, 1;
cvt.s64.s32 %rd3, %r63;
sub.s32 %r66, %r65, %r63;
cvt.s64.s32 %rd4, %r66;
shl.b32 %r67, %r28, 1;
add.s32 %r539, %r540, %r67;
mad.lo.s32 %r538, %r28, 3, %r540;
shl.b32 %r68, %r28, 2;
add.s32 %r537, %r540, %r68;
mad.lo.s32 %r536, %r28, 5, %r540;
add.s32 %r535, %r540, %r28;
add.s32 %r69, %r540, %r54;
add.s32 %r70, %r69, %r58;
add.s32 %r71, %r4, 1;
mad.lo.s32 %r72, %r27, %r71, %r70;
mul.wide.s32 %rd5, %r72, 18;
mad.lo.s32 %r73, %r4, %r27, %r540;
add.s32 %r74, %r73, %r54;
add.s32 %r75, %r74, %r58;
mul.wide.s32 %rd6, %r75, 18;
mul.wide.s32 %rd20, %r64, 2;
add.s64 %rd21, %rd6, %rd20;
add.s64 %rd7, %rd21, 4;
add.s64 %rd22, %rd5, %rd20;
add.s64 %rd8, %rd22, 2;
cvta.to.global.u64 %rd9, %rd16;
mov.f32 %f301, 0f00000000;
shl.b64 %rd29, %rd3, 2;
shl.b64 %rd31, %rd4, 2;

$L__BB38_2:
cvt.u64.u32 %rd23, %r540;
add.s64 %rd24, %rd1, %rd23;
mul.lo.s64 %rd25, %rd24, 36;
add.s64 %rd26, %rd9, %rd25;
add.s64 %rd27, %rd26, 4;
add.s64 %rd28, %rd115, %rd7;
ld.global.nc.u16 %rs13, [%rd28+-2];
ld.global.nc.u16 %rs14, [%rd28];
mov.b32 %r292, {%rs13, %rs14};
add.s64 %rd30, %rd27, %rd29;
ld.global.nc.u32 %r78, [%rd30];
ld.global.nc.u32 %r82, [%rd30+16];
ld.global.nc.u16 %rs15, [%rd28+2];
ld.global.nc.u16 %rs16, [%rd28+4];
mov.b32 %r293, {%rs15, %rs16};
add.s64 %rd32, %rd27, %rd31;
ld.global.nc.u32 %r86, [%rd32];
ld.global.nc.u32 %r90, [%rd30+20];
add.s64 %rd33, %rd115, %rd6;
ld.global.nc.u16 %rs1, [%rd33];

{ cvt.f32.f16 %f73, %rs1;}


and.b32 %r257, %r292, 252645135;
shr.u32 %r294, %r292, 4;
and.b32 %r261, %r294, 252645135;
mov.u32 %r277, 0;

dp4a.s32.s32 %r76, %r257, %r78, %r277;


dp4a.s32.s32 %r80, %r261, %r82, %r76;

and.b32 %r265, %r293, 252645135;
shr.u32 %r295, %r293, 4;
and.b32 %r269, %r295, 252645135;

dp4a.s32.s32 %r84, %r265, %r86, %r80;


dp4a.s32.s32 %r88, %r269, %r90, %r84;

ld.global.nc.u32 %r92, [%rd26];

{.reg .f16 low,high;
mov.b32 {low,high},%r92;
cvt.f32.f16 %f74, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r92;
cvt.f32.f16 %f75, high;}


cvt.rn.f32.s32 %f109, %r88;
mul.ftz.f32 %f110, %f74, %f109;
mul.ftz.f32 %f111, %f75, 0f40800000;
sub.ftz.f32 %f112, %f110, %f111;
fma.rn.ftz.f32 %f312, %f73, %f112, %f312;
add.s64 %rd34, %rd115, %rd8;
ld.global.nc.u16 %rs17, [%rd34];
ld.global.nc.u16 %rs18, [%rd34+2];
mov.b32 %r296, {%rs17, %rs18};
ld.global.nc.u16 %rs19, [%rd34+4];
ld.global.nc.u16 %rs20, [%rd34+6];
mov.b32 %r297, {%rs19, %rs20};
add.s64 %rd35, %rd115, %rd5;
ld.global.nc.u16 %rs2, [%rd35];

{ cvt.f32.f16 %f76, %rs2;}


and.b32 %r275, %r296, 252645135;
shr.u32 %r298, %r296, 4;
and.b32 %r279, %r298, 252645135;

dp4a.s32.s32 %r94, %r275, %r78, %r277;


dp4a.s32.s32 %r98, %r279, %r82, %r94;

and.b32 %r283, %r297, 252645135;
shr.u32 %r299, %r297, 4;
and.b32 %r287, %r299, 252645135;

dp4a.s32.s32 %r102, %r283, %r86, %r98;


dp4a.s32.s32 %r106, %r287, %r90, %r102;


{.reg .f16 low,high;
mov.b32 {low,high},%r92;
cvt.f32.f16 %f77, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r92;
cvt.f32.f16 %f78, high;}


cvt.rn.f32.s32 %f113, %r106;
mul.ftz.f32 %f114, %f77, %f113;
mul.ftz.f32 %f115, %f78, 0f40800000;
sub.ftz.f32 %f116, %f114, %f115;
fma.rn.ftz.f32 %f311, %f76, %f116, %f311;
cvt.u64.u32 %rd36, %r535;
add.s64 %rd37, %rd1, %rd36;
mul.lo.s64 %rd38, %rd37, 36;
add.s64 %rd39, %rd9, %rd38;
add.s64 %rd40, %rd39, 4;
add.s64 %rd41, %rd40, %rd29;
ld.global.nc.u32 %r114, [%rd41];
ld.global.nc.u32 %r118, [%rd41+16];
add.s64 %rd42, %rd40, %rd31;
ld.global.nc.u32 %r122, [%rd42];
ld.global.nc.u32 %r126, [%rd41+20];

{ cvt.f32.f16 %f79, %rs1;}



dp4a.s32.s32 %r112, %r257, %r114, %r277;


dp4a.s32.s32 %r116, %r261, %r118, %r112;


dp4a.s32.s32 %r120, %r265, %r122, %r116;


dp4a.s32.s32 %r124, %r269, %r126, %r120;

ld.global.nc.u32 %r128, [%rd39];

{.reg .f16 low,high;
mov.b32 {low,high},%r128;
cvt.f32.f16 %f80, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r128;
cvt.f32.f16 %f81, high;}


cvt.rn.f32.s32 %f117, %r124;
mul.ftz.f32 %f118, %f80, %f117;
mul.ftz.f32 %f119, %f81, 0f40800000;
sub.ftz.f32 %f120, %f118, %f119;
fma.rn.ftz.f32 %f310, %f79, %f120, %f310;

{ cvt.f32.f16 %f82, %rs2;}



dp4a.s32.s32 %r130, %r275, %r114, %r277;


dp4a.s32.s32 %r134, %r279, %r118, %r130;


dp4a.s32.s32 %r138, %r283, %r122, %r134;


dp4a.s32.s32 %r142, %r287, %r126, %r138;


{.reg .f16 low,high;
mov.b32 {low,high},%r128;
cvt.f32.f16 %f83, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r128;
cvt.f32.f16 %f84, high;}


cvt.rn.f32.s32 %f121, %r142;
mul.ftz.f32 %f122, %f83, %f121;
mul.ftz.f32 %f123, %f84, 0f40800000;
sub.ftz.f32 %f124, %f122, %f123;
fma.rn.ftz.f32 %f309, %f82, %f124, %f309;
cvt.u64.u32 %rd43, %r539;
add.s64 %rd44, %rd1, %rd43;
mul.lo.s64 %rd45, %rd44, 36;
add.s64 %rd46, %rd9, %rd45;
add.s64 %rd47, %rd46, 4;
add.s64 %rd48, %rd47, %rd29;
ld.global.nc.u32 %r150, [%rd48];
ld.global.nc.u32 %r154, [%rd48+16];
add.s64 %rd49, %rd47, %rd31;
ld.global.nc.u32 %r158, [%rd49];
ld.global.nc.u32 %r162, [%rd48+20];

{ cvt.f32.f16 %f85, %rs1;}



dp4a.s32.s32 %r148, %r257, %r150, %r277;


dp4a.s32.s32 %r152, %r261, %r154, %r148;


dp4a.s32.s32 %r156, %r265, %r158, %r152;


dp4a.s32.s32 %r160, %r269, %r162, %r156;

ld.global.nc.u32 %r164, [%rd46];

{.reg .f16 low,high;
mov.b32 {low,high},%r164;
cvt.f32.f16 %f86, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r164;
cvt.f32.f16 %f87, high;}


cvt.rn.f32.s32 %f125, %r160;
mul.ftz.f32 %f126, %f86, %f125;
mul.ftz.f32 %f127, %f87, 0f40800000;
sub.ftz.f32 %f128, %f126, %f127;
fma.rn.ftz.f32 %f308, %f85, %f128, %f308;

{ cvt.f32.f16 %f88, %rs2;}



dp4a.s32.s32 %r166, %r275, %r150, %r277;


dp4a.s32.s32 %r170, %r279, %r154, %r166;


dp4a.s32.s32 %r174, %r283, %r158, %r170;


dp4a.s32.s32 %r178, %r287, %r162, %r174;


{.reg .f16 low,high;
mov.b32 {low,high},%r164;
cvt.f32.f16 %f89, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r164;
cvt.f32.f16 %f90, high;}


cvt.rn.f32.s32 %f129, %r178;
mul.ftz.f32 %f130, %f89, %f129;
mul.ftz.f32 %f131, %f90, 0f40800000;
sub.ftz.f32 %f132, %f130, %f131;
fma.rn.ftz.f32 %f307, %f88, %f132, %f307;
cvt.u64.u32 %rd50, %r538;
add.s64 %rd51, %rd1, %rd50;
mul.lo.s64 %rd52, %rd51, 36;
add.s64 %rd53, %rd9, %rd52;
add.s64 %rd54, %rd53, 4;
add.s64 %rd55, %rd54, %rd29;
ld.global.nc.u32 %r186, [%rd55];
ld.global.nc.u32 %r190, [%rd55+16];
add.s64 %rd56, %rd54, %rd31;
ld.global.nc.u32 %r194, [%rd56];
ld.global.nc.u32 %r198, [%rd55+20];

{ cvt.f32.f16 %f91, %rs1;}



dp4a.s32.s32 %r184, %r257, %r186, %r277;


dp4a.s32.s32 %r188, %r261, %r190, %r184;


dp4a.s32.s32 %r192, %r265, %r194, %r188;


dp4a.s32.s32 %r196, %r269, %r198, %r192;

ld.global.nc.u32 %r200, [%rd53];

{.reg .f16 low,high;
mov.b32 {low,high},%r200;
cvt.f32.f16 %f92, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r200;
cvt.f32.f16 %f93, high;}


cvt.rn.f32.s32 %f133, %r196;
mul.ftz.f32 %f134, %f92, %f133;
mul.ftz.f32 %f135, %f93, 0f40800000;
sub.ftz.f32 %f136, %f134, %f135;
fma.rn.ftz.f32 %f306, %f91, %f136, %f306;

{ cvt.f32.f16 %f94, %rs2;}



dp4a.s32.s32 %r202, %r275, %r186, %r277;


dp4a.s32.s32 %r206, %r279, %r190, %r202;


dp4a.s32.s32 %r210, %r283, %r194, %r206;


dp4a.s32.s32 %r214, %r287, %r198, %r210;


{.reg .f16 low,high;
mov.b32 {low,high},%r200;
cvt.f32.f16 %f95, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r200;
cvt.f32.f16 %f96, high;}


cvt.rn.f32.s32 %f137, %r214;
mul.ftz.f32 %f138, %f95, %f137;
mul.ftz.f32 %f139, %f96, 0f40800000;
sub.ftz.f32 %f140, %f138, %f139;
fma.rn.ftz.f32 %f305, %f94, %f140, %f305;
cvt.u64.u32 %rd57, %r537;
add.s64 %rd58, %rd1, %rd57;
mul.lo.s64 %rd59, %rd58, 36;
add.s64 %rd60, %rd9, %rd59;
add.s64 %rd61, %rd60, 4;
add.s64 %rd62, %rd61, %rd29;
ld.global.nc.u32 %r222, [%rd62];
ld.global.nc.u32 %r226, [%rd62+16];
add.s64 %rd63, %rd61, %rd31;
ld.global.nc.u32 %r230, [%rd63];
ld.global.nc.u32 %r234, [%rd62+20];

{ cvt.f32.f16 %f97, %rs1;}



dp4a.s32.s32 %r220, %r257, %r222, %r277;


dp4a.s32.s32 %r224, %r261, %r226, %r220;


dp4a.s32.s32 %r228, %r265, %r230, %r224;


dp4a.s32.s32 %r232, %r269, %r234, %r228;

ld.global.nc.u32 %r236, [%rd60];

{.reg .f16 low,high;
mov.b32 {low,high},%r236;
cvt.f32.f16 %f98, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r236;
cvt.f32.f16 %f99, high;}


cvt.rn.f32.s32 %f141, %r232;
mul.ftz.f32 %f142, %f98, %f141;
mul.ftz.f32 %f143, %f99, 0f40800000;
sub.ftz.f32 %f144, %f142, %f143;
fma.rn.ftz.f32 %f304, %f97, %f144, %f304;

{ cvt.f32.f16 %f100, %rs2;}



dp4a.s32.s32 %r238, %r275, %r222, %r277;


dp4a.s32.s32 %r242, %r279, %r226, %r238;


dp4a.s32.s32 %r246, %r283, %r230, %r242;


dp4a.s32.s32 %r250, %r287, %r234, %r246;


{.reg .f16 low,high;
mov.b32 {low,high},%r236;
cvt.f32.f16 %f101, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r236;
cvt.f32.f16 %f102, high;}


cvt.rn.f32.s32 %f145, %r250;
mul.ftz.f32 %f146, %f101, %f145;
mul.ftz.f32 %f147, %f102, 0f40800000;
sub.ftz.f32 %f148, %f146, %f147;
fma.rn.ftz.f32 %f303, %f100, %f148, %f303;
cvt.u64.u32 %rd64, %r536;
add.s64 %rd65, %rd1, %rd64;
mul.lo.s64 %rd66, %rd65, 36;
add.s64 %rd67, %rd9, %rd66;
add.s64 %rd68, %rd67, 4;
add.s64 %rd69, %rd68, %rd29;
ld.global.nc.u32 %r258, [%rd69];
ld.global.nc.u32 %r262, [%rd69+16];
add.s64 %rd70, %rd68, %rd31;
ld.global.nc.u32 %r266, [%rd70];
ld.global.nc.u32 %r270, [%rd69+20];

{ cvt.f32.f16 %f103, %rs1;}



dp4a.s32.s32 %r256, %r257, %r258, %r277;


dp4a.s32.s32 %r260, %r261, %r262, %r256;


dp4a.s32.s32 %r264, %r265, %r266, %r260;


dp4a.s32.s32 %r268, %r269, %r270, %r264;

ld.global.nc.u32 %r272, [%rd67];

{.reg .f16 low,high;
mov.b32 {low,high},%r272;
cvt.f32.f16 %f104, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r272;
cvt.f32.f16 %f105, high;}


cvt.rn.f32.s32 %f149, %r268;
mul.ftz.f32 %f150, %f104, %f149;
mul.ftz.f32 %f151, %f105, 0f40800000;
sub.ftz.f32 %f152, %f150, %f151;
fma.rn.ftz.f32 %f302, %f103, %f152, %f302;

{ cvt.f32.f16 %f106, %rs2;}



dp4a.s32.s32 %r274, %r275, %r258, %r277;


dp4a.s32.s32 %r278, %r279, %r262, %r274;


dp4a.s32.s32 %r282, %r283, %r266, %r278;


dp4a.s32.s32 %r286, %r287, %r270, %r282;


{.reg .f16 low,high;
mov.b32 {low,high},%r272;
cvt.f32.f16 %f107, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r272;
cvt.f32.f16 %f108, high;}


cvt.rn.f32.s32 %f153, %r286;
mul.ftz.f32 %f154, %f107, %f153;
mul.ftz.f32 %f155, %f108, 0f40800000;
sub.ftz.f32 %f156, %f154, %f155;
fma.rn.ftz.f32 %f301, %f106, %f156, %f301;
add.s32 %r539, %r539, 32;
add.s32 %r538, %r538, 32;
add.s32 %r537, %r537, 32;
add.s32 %r536, %r536, 32;
add.s32 %r535, %r535, 32;
add.s64 %rd115, %rd115, 576;
add.s32 %r540, %r540, 32;
setp.lt.s32 %p4, %r540, %r6;
@%p4 bra $L__BB38_2;

$L__BB38_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB38_5;

mov.u32 %r300, _ZZ13mul_mat_vec_qIL9ggml_type2ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r301, %r1, 1536, %r300;
shl.b32 %r302, %r2, 2;
add.s32 %r303, %r301, %r302;
st.shared.f32 [%r303+-1536], %f312;
st.shared.f32 [%r303+-1408], %f311;
st.shared.f32 [%r303+-1280], %f310;
st.shared.f32 [%r303+-1152], %f309;
st.shared.f32 [%r303+-1024], %f308;
st.shared.f32 [%r303+-896], %f307;
st.shared.f32 [%r303+-768], %f306;
st.shared.f32 [%r303+-640], %f305;
st.shared.f32 [%r303+-512], %f304;
st.shared.f32 [%r303+-384], %f303;
st.shared.f32 [%r303+-256], %f302;
st.shared.f32 [%r303+-128], %f301;

$L__BB38_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB38_30;

mov.u32 %r304, %ctaid.z;
mov.u32 %r305, %ctaid.y;
mad.lo.s32 %r306, %r305, %r35, %r4;
mad.lo.s32 %r307, %r304, %r41, %r306;
cvt.u64.u32 %rd12, %r307;
shl.b32 %r308, %r2, 2;
mov.u32 %r309, _ZZ13mul_mat_vec_qIL9ggml_type2ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r24, %r309, %r308;
mov.u32 %r310, 2;
ld.shared.f32 %f157, [%r24];
add.ftz.f32 %f158, %f157, %f312;
mov.b32 %r311, %f158;
mov.u32 %r312, 31;
mov.u32 %r313, 16;
mov.u32 %r314, -1;
shfl.sync.bfly.b32 %r315|%p7, %r311, %r313, %r312, %r314;
mov.b32 %f159, %r315;
add.ftz.f32 %f160, %f158, %f159;
mov.b32 %r316, %f160;
mov.u32 %r317, 8;
shfl.sync.bfly.b32 %r318|%p8, %r316, %r317, %r312, %r314;
mov.b32 %f161, %r318;
add.ftz.f32 %f162, %f160, %f161;
mov.b32 %r319, %f162;
mov.u32 %r320, 4;
shfl.sync.bfly.b32 %r321|%p9, %r319, %r320, %r312, %r314;
mov.b32 %f163, %r321;
add.ftz.f32 %f164, %f162, %f163;
mov.b32 %r322, %f164;
shfl.sync.bfly.b32 %r323|%p10, %r322, %r310, %r312, %r314;
mov.b32 %f165, %r323;
add.ftz.f32 %f166, %f164, %f165;
mov.b32 %r324, %f166;
mov.u32 %r325, 1;
shfl.sync.bfly.b32 %r326|%p11, %r324, %r325, %r312, %r314;
mov.b32 %f167, %r326;
add.ftz.f32 %f37, %f166, %f167;
setp.lt.u32 %p12, %r4, %r29;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
not.pred %p14, %p1;
@%p14 bra $L__BB38_8;

cvta.to.global.u64 %rd71, %rd17;
mul.wide.u32 %rd72, %r307, 4;
add.s64 %rd73, %rd71, %rd72;
st.global.f32 [%rd73], %f37;

$L__BB38_8:
ld.shared.f32 %f168, [%r24+128];
add.ftz.f32 %f169, %f168, %f311;
mov.b32 %r333, %f169;
shfl.sync.bfly.b32 %r337|%p15, %r333, %r313, %r312, %r314;
mov.b32 %f170, %r337;
add.ftz.f32 %f171, %f169, %f170;
mov.b32 %r338, %f171;
shfl.sync.bfly.b32 %r340|%p16, %r338, %r317, %r312, %r314;
mov.b32 %f172, %r340;
add.ftz.f32 %f173, %f171, %f172;
mov.b32 %r341, %f173;
shfl.sync.bfly.b32 %r343|%p17, %r341, %r320, %r312, %r314;
mov.b32 %f174, %r343;
add.ftz.f32 %f175, %f173, %f174;
mov.b32 %r344, %f175;
shfl.sync.bfly.b32 %r346|%p18, %r344, %r310, %r312, %r314;
mov.b32 %f176, %r346;
add.ftz.f32 %f177, %f175, %f176;
mov.b32 %r347, %f177;
shfl.sync.bfly.b32 %r349|%p19, %r347, %r325, %r312, %r314;
mov.b32 %f178, %r349;
add.ftz.f32 %f38, %f177, %f178;
setp.eq.s32 %p20, %r2, 1;
or.b32 %r353, %r4, 1;
setp.lt.u32 %p21, %r353, %r29;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB38_10;

cvta.to.global.u64 %rd74, %rd17;
mul.wide.u32 %rd75, %r307, 4;
add.s64 %rd76, %rd74, %rd75;
st.global.f32 [%rd76+4], %f38;

$L__BB38_10:
ld.shared.f32 %f179, [%r24+256];
add.ftz.f32 %f180, %f179, %f310;
mov.b32 %r360, %f180;
mov.u32 %r361, 31;
mov.u32 %r362, 16;
mov.u32 %r363, -1;
shfl.sync.bfly.b32 %r364|%p23, %r360, %r362, %r361, %r363;
mov.b32 %f181, %r364;
add.ftz.f32 %f182, %f180, %f181;
mov.b32 %r365, %f182;
mov.u32 %r366, 8;
shfl.sync.bfly.b32 %r367|%p24, %r365, %r366, %r361, %r363;
mov.b32 %f183, %r367;
add.ftz.f32 %f184, %f182, %f183;
mov.b32 %r368, %f184;
mov.u32 %r369, 4;
shfl.sync.bfly.b32 %r370|%p25, %r368, %r369, %r361, %r363;
mov.b32 %f185, %r370;
add.ftz.f32 %f186, %f184, %f185;
mov.b32 %r371, %f186;
mov.u32 %r372, 2;
shfl.sync.bfly.b32 %r373|%p26, %r371, %r372, %r361, %r363;
mov.b32 %f187, %r373;
add.ftz.f32 %f188, %f186, %f187;
mov.b32 %r374, %f188;
mov.u32 %r375, 1;
shfl.sync.bfly.b32 %r376|%p27, %r374, %r375, %r361, %r363;
mov.b32 %f189, %r376;
add.ftz.f32 %f39, %f188, %f189;
@%p14 bra $L__BB38_12;

cvt.u64.u32 %rd77, %r29;
add.s64 %rd78, %rd77, %rd12;
cvta.to.global.u64 %rd79, %rd17;
shl.b64 %rd80, %rd78, 2;
add.s64 %rd81, %rd79, %rd80;
st.global.f32 [%rd81], %f39;

$L__BB38_12:
ld.shared.f32 %f190, [%r24+384];
add.ftz.f32 %f191, %f190, %f309;
mov.b32 %r377, %f191;
shfl.sync.bfly.b32 %r381|%p29, %r377, %r362, %r361, %r363;
mov.b32 %f192, %r381;
add.ftz.f32 %f193, %f191, %f192;
mov.b32 %r382, %f193;
shfl.sync.bfly.b32 %r384|%p30, %r382, %r366, %r361, %r363;
mov.b32 %f194, %r384;
add.ftz.f32 %f195, %f193, %f194;
mov.b32 %r385, %f195;
shfl.sync.bfly.b32 %r387|%p31, %r385, %r369, %r361, %r363;
mov.b32 %f196, %r387;
add.ftz.f32 %f197, %f195, %f196;
mov.b32 %r388, %f197;
shfl.sync.bfly.b32 %r390|%p32, %r388, %r372, %r361, %r363;
mov.b32 %f198, %r390;
add.ftz.f32 %f199, %f197, %f198;
mov.b32 %r391, %f199;
shfl.sync.bfly.b32 %r393|%p33, %r391, %r375, %r361, %r363;
mov.b32 %f200, %r393;
add.ftz.f32 %f40, %f199, %f200;
@%p22 bra $L__BB38_14;

add.s32 %r394, %r29, 1;
cvt.u64.u32 %rd82, %r394;
add.s64 %rd83, %rd82, %rd12;
cvta.to.global.u64 %rd84, %rd17;
shl.b64 %rd85, %rd83, 2;
add.s64 %rd86, %rd84, %rd85;
st.global.f32 [%rd86], %f40;

$L__BB38_14:
shl.b32 %r395, %r29, 1;
mov.u32 %r396, 1;
ld.shared.f32 %f201, [%r24+512];
add.ftz.f32 %f202, %f201, %f308;
mov.b32 %r397, %f202;
mov.u32 %r398, 31;
mov.u32 %r399, 16;
mov.u32 %r400, -1;
shfl.sync.bfly.b32 %r401|%p35, %r397, %r399, %r398, %r400;
mov.b32 %f203, %r401;
add.ftz.f32 %f204, %f202, %f203;
mov.b32 %r402, %f204;
mov.u32 %r403, 8;
shfl.sync.bfly.b32 %r404|%p36, %r402, %r403, %r398, %r400;
mov.b32 %f205, %r404;
add.ftz.f32 %f206, %f204, %f205;
mov.b32 %r405, %f206;
mov.u32 %r406, 4;
shfl.sync.bfly.b32 %r407|%p37, %r405, %r406, %r398, %r400;
mov.b32 %f207, %r407;
add.ftz.f32 %f208, %f206, %f207;
mov.b32 %r408, %f208;
mov.u32 %r409, 2;
shfl.sync.bfly.b32 %r410|%p38, %r408, %r409, %r398, %r400;
mov.b32 %f209, %r410;
add.ftz.f32 %f210, %f208, %f209;
mov.b32 %r411, %f210;
shfl.sync.bfly.b32 %r412|%p39, %r411, %r396, %r398, %r400;
mov.b32 %f211, %r412;
add.ftz.f32 %f41, %f210, %f211;
cvt.u64.u32 %rd87, %r395;
add.s64 %rd88, %rd87, %rd12;
cvta.to.global.u64 %rd89, %rd17;
shl.b64 %rd90, %rd88, 2;
add.s64 %rd13, %rd89, %rd90;
@%p14 bra $L__BB38_16;

st.global.f32 [%rd13], %f41;

$L__BB38_16:
ld.shared.f32 %f212, [%r24+640];
add.ftz.f32 %f213, %f212, %f307;
mov.b32 %r413, %f213;
shfl.sync.bfly.b32 %r417|%p41, %r413, %r399, %r398, %r400;
mov.b32 %f214, %r417;
add.ftz.f32 %f215, %f213, %f214;
mov.b32 %r418, %f215;
shfl.sync.bfly.b32 %r420|%p42, %r418, %r403, %r398, %r400;
mov.b32 %f216, %r420;
add.ftz.f32 %f217, %f215, %f216;
mov.b32 %r421, %f217;
shfl.sync.bfly.b32 %r423|%p43, %r421, %r406, %r398, %r400;
mov.b32 %f218, %r423;
add.ftz.f32 %f219, %f217, %f218;
mov.b32 %r424, %f219;
shfl.sync.bfly.b32 %r426|%p44, %r424, %r409, %r398, %r400;
mov.b32 %f220, %r426;
add.ftz.f32 %f221, %f219, %f220;
mov.b32 %r427, %f221;
shfl.sync.bfly.b32 %r429|%p45, %r427, %r396, %r398, %r400;
mov.b32 %f222, %r429;
add.ftz.f32 %f42, %f221, %f222;
@%p22 bra $L__BB38_18;

st.global.f32 [%rd13+4], %f42;

$L__BB38_18:
mul.lo.s32 %r25, %r29, 3;
ld.shared.f32 %f223, [%r24+768];
add.ftz.f32 %f224, %f223, %f306;
mov.b32 %r430, %f224;
mov.u32 %r431, 31;
mov.u32 %r432, 16;
mov.u32 %r433, -1;
shfl.sync.bfly.b32 %r434|%p47, %r430, %r432, %r431, %r433;
mov.b32 %f225, %r434;
add.ftz.f32 %f226, %f224, %f225;
mov.b32 %r435, %f226;
mov.u32 %r436, 8;
shfl.sync.bfly.b32 %r437|%p48, %r435, %r436, %r431, %r433;
mov.b32 %f227, %r437;
add.ftz.f32 %f228, %f226, %f227;
mov.b32 %r438, %f228;
mov.u32 %r439, 4;
shfl.sync.bfly.b32 %r440|%p49, %r438, %r439, %r431, %r433;
mov.b32 %f229, %r440;
add.ftz.f32 %f230, %f228, %f229;
mov.b32 %r441, %f230;
mov.u32 %r442, 2;
shfl.sync.bfly.b32 %r443|%p50, %r441, %r442, %r431, %r433;
mov.b32 %f231, %r443;
add.ftz.f32 %f232, %f230, %f231;
mov.b32 %r444, %f232;
mov.u32 %r445, 1;
shfl.sync.bfly.b32 %r446|%p51, %r444, %r445, %r431, %r433;
mov.b32 %f233, %r446;
add.ftz.f32 %f43, %f232, %f233;
@%p14 bra $L__BB38_20;

cvt.u64.u32 %rd91, %r25;
add.s64 %rd92, %rd91, %rd12;
shl.b64 %rd94, %rd92, 2;
add.s64 %rd95, %rd89, %rd94;
st.global.f32 [%rd95], %f43;

$L__BB38_20:
ld.shared.f32 %f234, [%r24+896];
add.ftz.f32 %f235, %f234, %f305;
mov.b32 %r447, %f235;
shfl.sync.bfly.b32 %r451|%p53, %r447, %r432, %r431, %r433;
mov.b32 %f236, %r451;
add.ftz.f32 %f237, %f235, %f236;
mov.b32 %r452, %f237;
shfl.sync.bfly.b32 %r454|%p54, %r452, %r436, %r431, %r433;
mov.b32 %f238, %r454;
add.ftz.f32 %f239, %f237, %f238;
mov.b32 %r455, %f239;
shfl.sync.bfly.b32 %r457|%p55, %r455, %r439, %r431, %r433;
mov.b32 %f240, %r457;
add.ftz.f32 %f241, %f239, %f240;
mov.b32 %r458, %f241;
shfl.sync.bfly.b32 %r460|%p56, %r458, %r442, %r431, %r433;
mov.b32 %f242, %r460;
add.ftz.f32 %f243, %f241, %f242;
mov.b32 %r461, %f243;
shfl.sync.bfly.b32 %r463|%p57, %r461, %r445, %r431, %r433;
mov.b32 %f244, %r463;
add.ftz.f32 %f44, %f243, %f244;
@%p22 bra $L__BB38_22;

add.s32 %r464, %r25, 1;
cvt.u64.u32 %rd96, %r464;
add.s64 %rd97, %rd96, %rd12;
shl.b64 %rd99, %rd97, 2;
add.s64 %rd100, %rd89, %rd99;
st.global.f32 [%rd100], %f44;

$L__BB38_22:
shl.b32 %r465, %r29, 2;
mov.u32 %r466, 2;
ld.shared.f32 %f245, [%r24+1024];
add.ftz.f32 %f246, %f245, %f304;
mov.b32 %r467, %f246;
mov.u32 %r468, 31;
mov.u32 %r469, 16;
mov.u32 %r470, -1;
shfl.sync.bfly.b32 %r471|%p59, %r467, %r469, %r468, %r470;
mov.b32 %f247, %r471;
add.ftz.f32 %f248, %f246, %f247;
mov.b32 %r472, %f248;
mov.u32 %r473, 8;
shfl.sync.bfly.b32 %r474|%p60, %r472, %r473, %r468, %r470;
mov.b32 %f249, %r474;
add.ftz.f32 %f250, %f248, %f249;
mov.b32 %r475, %f250;
mov.u32 %r476, 4;
shfl.sync.bfly.b32 %r477|%p61, %r475, %r476, %r468, %r470;
mov.b32 %f251, %r477;
add.ftz.f32 %f252, %f250, %f251;
mov.b32 %r478, %f252;
shfl.sync.bfly.b32 %r479|%p62, %r478, %r466, %r468, %r470;
mov.b32 %f253, %r479;
add.ftz.f32 %f254, %f252, %f253;
mov.b32 %r480, %f254;
mov.u32 %r481, 1;
shfl.sync.bfly.b32 %r482|%p63, %r480, %r481, %r468, %r470;
mov.b32 %f255, %r482;
add.ftz.f32 %f45, %f254, %f255;
cvt.u64.u32 %rd101, %r465;
add.s64 %rd102, %rd101, %rd12;
shl.b64 %rd104, %rd102, 2;
add.s64 %rd14, %rd89, %rd104;
@%p14 bra $L__BB38_24;

st.global.f32 [%rd14], %f45;

$L__BB38_24:
ld.shared.f32 %f256, [%r24+1152];
add.ftz.f32 %f257, %f256, %f303;
mov.b32 %r483, %f257;
shfl.sync.bfly.b32 %r487|%p65, %r483, %r469, %r468, %r470;
mov.b32 %f258, %r487;
add.ftz.f32 %f259, %f257, %f258;
mov.b32 %r488, %f259;
shfl.sync.bfly.b32 %r490|%p66, %r488, %r473, %r468, %r470;
mov.b32 %f260, %r490;
add.ftz.f32 %f261, %f259, %f260;
mov.b32 %r491, %f261;
shfl.sync.bfly.b32 %r493|%p67, %r491, %r476, %r468, %r470;
mov.b32 %f262, %r493;
add.ftz.f32 %f263, %f261, %f262;
mov.b32 %r494, %f263;
shfl.sync.bfly.b32 %r496|%p68, %r494, %r466, %r468, %r470;
mov.b32 %f264, %r496;
add.ftz.f32 %f265, %f263, %f264;
mov.b32 %r497, %f265;
shfl.sync.bfly.b32 %r499|%p69, %r497, %r481, %r468, %r470;
mov.b32 %f266, %r499;
add.ftz.f32 %f46, %f265, %f266;
@%p22 bra $L__BB38_26;

st.global.f32 [%rd14+4], %f46;

$L__BB38_26:
mul.lo.s32 %r26, %r29, 5;
ld.shared.f32 %f267, [%r24+1280];
add.ftz.f32 %f268, %f267, %f302;
mov.b32 %r500, %f268;
mov.u32 %r501, 31;
mov.u32 %r502, 16;
mov.u32 %r503, -1;
shfl.sync.bfly.b32 %r504|%p71, %r500, %r502, %r501, %r503;
mov.b32 %f269, %r504;
add.ftz.f32 %f270, %f268, %f269;
mov.b32 %r505, %f270;
mov.u32 %r506, 8;
shfl.sync.bfly.b32 %r507|%p72, %r505, %r506, %r501, %r503;
mov.b32 %f271, %r507;
add.ftz.f32 %f272, %f270, %f271;
mov.b32 %r508, %f272;
mov.u32 %r509, 4;
shfl.sync.bfly.b32 %r510|%p73, %r508, %r509, %r501, %r503;
mov.b32 %f273, %r510;
add.ftz.f32 %f274, %f272, %f273;
mov.b32 %r511, %f274;
mov.u32 %r512, 2;
shfl.sync.bfly.b32 %r513|%p74, %r511, %r512, %r501, %r503;
mov.b32 %f275, %r513;
add.ftz.f32 %f276, %f274, %f275;
mov.b32 %r514, %f276;
mov.u32 %r515, 1;
shfl.sync.bfly.b32 %r516|%p75, %r514, %r515, %r501, %r503;
mov.b32 %f277, %r516;
add.ftz.f32 %f47, %f276, %f277;
@%p14 bra $L__BB38_28;

cvt.u64.u32 %rd105, %r26;
add.s64 %rd106, %rd105, %rd12;
shl.b64 %rd108, %rd106, 2;
add.s64 %rd109, %rd89, %rd108;
st.global.f32 [%rd109], %f47;

$L__BB38_28:
ld.shared.f32 %f278, [%r24+1408];
add.ftz.f32 %f279, %f278, %f301;
mov.b32 %r517, %f279;
shfl.sync.bfly.b32 %r521|%p77, %r517, %r502, %r501, %r503;
mov.b32 %f280, %r521;
add.ftz.f32 %f281, %f279, %f280;
mov.b32 %r522, %f281;
shfl.sync.bfly.b32 %r524|%p78, %r522, %r506, %r501, %r503;
mov.b32 %f282, %r524;
add.ftz.f32 %f283, %f281, %f282;
mov.b32 %r525, %f283;
shfl.sync.bfly.b32 %r527|%p79, %r525, %r509, %r501, %r503;
mov.b32 %f284, %r527;
add.ftz.f32 %f285, %f283, %f284;
mov.b32 %r528, %f285;
shfl.sync.bfly.b32 %r530|%p80, %r528, %r512, %r501, %r503;
mov.b32 %f286, %r530;
add.ftz.f32 %f287, %f285, %f286;
mov.b32 %r531, %f287;
shfl.sync.bfly.b32 %r533|%p81, %r531, %r515, %r501, %r503;
mov.b32 %f288, %r533;
add.ftz.f32 %f48, %f287, %f288;
@%p22 bra $L__BB38_30;

add.s32 %r534, %r26, 1;
cvt.u64.u32 %rd110, %r534;
add.s64 %rd111, %rd110, %rd12;
shl.b64 %rd113, %rd111, 2;
add.s64 %rd114, %rd89, %rd113;
st.global.f32 [%rd114], %f48;

$L__BB38_30:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type2ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type2ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type2ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type2ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type2ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type2ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type2ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type2ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type2ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<95>;
.reg .b16 %rs<23>;
.reg .f32 %f<365>;
.reg .b32 %r<604>;
.reg .b64 %rd<125>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type2ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1792];

ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type2ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type2ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type2ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r45, [_Z13mul_mat_vec_qIL9ggml_type2ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type2ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type2ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type2ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type2ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type2ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type2ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type2ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type2ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type2ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type2ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type2ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type2ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type2ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r46, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r46, %r2;
mov.u32 %r47, %ctaid.x;
shl.b32 %r4, %r47, 1;
shr.u32 %r48, %r3, 31;
add.s32 %r49, %r3, %r48;
shr.s32 %r603, %r49, 1;
shr.u32 %r6, %r45, 5;
setp.ge.s32 %p3, %r603, %r6;
mov.f32 %f351, 0f00000000;
mov.f32 %f352, %f351;
mov.f32 %f353, %f351;
mov.f32 %f354, %f351;
mov.f32 %f355, %f351;
mov.f32 %f356, %f351;
mov.f32 %f357, %f351;
mov.f32 %f358, %f351;
mov.f32 %f359, %f351;
mov.f32 %f360, %f351;
mov.f32 %f361, %f351;
mov.f32 %f362, %f351;
mov.f32 %f363, %f351;
mov.f32 %f364, %f351;
@%p3 bra $L__BB39_3;

mov.u32 %r50, %ctaid.y;
mov.u32 %r51, %ctaid.z;
mul.lo.s32 %r52, %r50, %r37;
cvt.u64.u32 %rd20, %r52;
mul.lo.s32 %r53, %r51, %r43;
cvt.u64.u32 %rd21, %r53;
add.s64 %rd1, %rd21, %rd20;
mul.hi.u32 %r54, %r51, %r39;
add.s32 %r55, %r54, %r51;
shr.u32 %r56, %r55, %r40;
mul.lo.s32 %r57, %r56, %r42;
mul.hi.u32 %r58, %r50, %r33;
add.s32 %r59, %r58, %r50;
shr.u32 %r60, %r59, %r34;
mul.lo.s32 %r61, %r60, %r36;
cvta.to.global.u64 %rd124, %rd17;
and.b32 %r64, %r49, -2;
sub.s32 %r65, %r3, %r64;
shl.b32 %r66, %r65, 1;
shl.b32 %r67, %r65, 2;
or.b32 %r68, %r67, 1;
cvt.s64.s32 %rd3, %r66;
sub.s32 %r69, %r68, %r66;
cvt.s64.s32 %rd4, %r69;
shl.b32 %r70, %r31, 1;
add.s32 %r602, %r603, %r70;
mad.lo.s32 %r601, %r31, 3, %r603;
shl.b32 %r71, %r31, 2;
add.s32 %r600, %r603, %r71;
mad.lo.s32 %r599, %r31, 5, %r603;
mad.lo.s32 %r598, %r31, 6, %r603;
add.s32 %r597, %r603, %r31;
add.s32 %r72, %r603, %r57;
add.s32 %r73, %r72, %r61;
add.s32 %r74, %r4, 1;
mad.lo.s32 %r75, %r30, %r74, %r73;
mul.wide.s32 %rd5, %r75, 18;
mad.lo.s32 %r76, %r4, %r30, %r603;
add.s32 %r77, %r76, %r57;
add.s32 %r78, %r77, %r61;
mul.wide.s32 %rd6, %r78, 18;
mul.wide.s32 %rd22, %r67, 2;
add.s64 %rd23, %rd6, %rd22;
add.s64 %rd7, %rd23, 4;
add.s64 %rd24, %rd5, %rd22;
add.s64 %rd8, %rd24, 2;
cvta.to.global.u64 %rd9, %rd18;
mov.f32 %f351, 0f00000000;
shl.b64 %rd31, %rd3, 2;
shl.b64 %rd33, %rd4, 2;

$L__BB39_2:
cvt.u64.u32 %rd25, %r603;
add.s64 %rd26, %rd1, %rd25;
mul.lo.s64 %rd27, %rd26, 36;
add.s64 %rd28, %rd9, %rd27;
add.s64 %rd29, %rd28, 4;
add.s64 %rd30, %rd124, %rd7;
ld.global.nc.u16 %rs15, [%rd30+-2];
ld.global.nc.u16 %rs16, [%rd30];
mov.b32 %r331, {%rs15, %rs16};
add.s64 %rd32, %rd29, %rd31;
ld.global.nc.u32 %r81, [%rd32];
ld.global.nc.u32 %r85, [%rd32+16];
ld.global.nc.u16 %rs17, [%rd30+2];
ld.global.nc.u16 %rs18, [%rd30+4];
mov.b32 %r332, {%rs17, %rs18};
add.s64 %rd34, %rd29, %rd33;
ld.global.nc.u32 %r89, [%rd34];
ld.global.nc.u32 %r93, [%rd32+20];
add.s64 %rd35, %rd124, %rd6;
ld.global.nc.u16 %rs1, [%rd35];

{ cvt.f32.f16 %f85, %rs1;}


and.b32 %r296, %r331, 252645135;
shr.u32 %r333, %r331, 4;
and.b32 %r300, %r333, 252645135;
mov.u32 %r316, 0;

dp4a.s32.s32 %r79, %r296, %r81, %r316;


dp4a.s32.s32 %r83, %r300, %r85, %r79;

and.b32 %r304, %r332, 252645135;
shr.u32 %r334, %r332, 4;
and.b32 %r308, %r334, 252645135;

dp4a.s32.s32 %r87, %r304, %r89, %r83;


dp4a.s32.s32 %r91, %r308, %r93, %r87;

ld.global.nc.u32 %r95, [%rd28];

{.reg .f16 low,high;
mov.b32 {low,high},%r95;
cvt.f32.f16 %f86, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r95;
cvt.f32.f16 %f87, high;}


cvt.rn.f32.s32 %f127, %r91;
mul.ftz.f32 %f128, %f86, %f127;
mul.ftz.f32 %f129, %f87, 0f40800000;
sub.ftz.f32 %f130, %f128, %f129;
fma.rn.ftz.f32 %f364, %f85, %f130, %f364;
add.s64 %rd36, %rd124, %rd8;
ld.global.nc.u16 %rs19, [%rd36];
ld.global.nc.u16 %rs20, [%rd36+2];
mov.b32 %r335, {%rs19, %rs20};
ld.global.nc.u16 %rs21, [%rd36+4];
ld.global.nc.u16 %rs22, [%rd36+6];
mov.b32 %r336, {%rs21, %rs22};
add.s64 %rd37, %rd124, %rd5;
ld.global.nc.u16 %rs2, [%rd37];

{ cvt.f32.f16 %f88, %rs2;}


and.b32 %r314, %r335, 252645135;
shr.u32 %r337, %r335, 4;
and.b32 %r318, %r337, 252645135;

dp4a.s32.s32 %r97, %r314, %r81, %r316;


dp4a.s32.s32 %r101, %r318, %r85, %r97;

and.b32 %r322, %r336, 252645135;
shr.u32 %r338, %r336, 4;
and.b32 %r326, %r338, 252645135;

dp4a.s32.s32 %r105, %r322, %r89, %r101;


dp4a.s32.s32 %r109, %r326, %r93, %r105;


{.reg .f16 low,high;
mov.b32 {low,high},%r95;
cvt.f32.f16 %f89, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r95;
cvt.f32.f16 %f90, high;}


cvt.rn.f32.s32 %f131, %r109;
mul.ftz.f32 %f132, %f89, %f131;
mul.ftz.f32 %f133, %f90, 0f40800000;
sub.ftz.f32 %f134, %f132, %f133;
fma.rn.ftz.f32 %f363, %f88, %f134, %f363;
cvt.u64.u32 %rd38, %r597;
add.s64 %rd39, %rd1, %rd38;
mul.lo.s64 %rd40, %rd39, 36;
add.s64 %rd41, %rd9, %rd40;
add.s64 %rd42, %rd41, 4;
add.s64 %rd43, %rd42, %rd31;
ld.global.nc.u32 %r117, [%rd43];
ld.global.nc.u32 %r121, [%rd43+16];
add.s64 %rd44, %rd42, %rd33;
ld.global.nc.u32 %r125, [%rd44];
ld.global.nc.u32 %r129, [%rd43+20];

{ cvt.f32.f16 %f91, %rs1;}



dp4a.s32.s32 %r115, %r296, %r117, %r316;


dp4a.s32.s32 %r119, %r300, %r121, %r115;


dp4a.s32.s32 %r123, %r304, %r125, %r119;


dp4a.s32.s32 %r127, %r308, %r129, %r123;

ld.global.nc.u32 %r131, [%rd41];

{.reg .f16 low,high;
mov.b32 {low,high},%r131;
cvt.f32.f16 %f92, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r131;
cvt.f32.f16 %f93, high;}


cvt.rn.f32.s32 %f135, %r127;
mul.ftz.f32 %f136, %f92, %f135;
mul.ftz.f32 %f137, %f93, 0f40800000;
sub.ftz.f32 %f138, %f136, %f137;
fma.rn.ftz.f32 %f362, %f91, %f138, %f362;

{ cvt.f32.f16 %f94, %rs2;}



dp4a.s32.s32 %r133, %r314, %r117, %r316;


dp4a.s32.s32 %r137, %r318, %r121, %r133;


dp4a.s32.s32 %r141, %r322, %r125, %r137;


dp4a.s32.s32 %r145, %r326, %r129, %r141;


{.reg .f16 low,high;
mov.b32 {low,high},%r131;
cvt.f32.f16 %f95, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r131;
cvt.f32.f16 %f96, high;}


cvt.rn.f32.s32 %f139, %r145;
mul.ftz.f32 %f140, %f95, %f139;
mul.ftz.f32 %f141, %f96, 0f40800000;
sub.ftz.f32 %f142, %f140, %f141;
fma.rn.ftz.f32 %f361, %f94, %f142, %f361;
cvt.u64.u32 %rd45, %r602;
add.s64 %rd46, %rd1, %rd45;
mul.lo.s64 %rd47, %rd46, 36;
add.s64 %rd48, %rd9, %rd47;
add.s64 %rd49, %rd48, 4;
add.s64 %rd50, %rd49, %rd31;
ld.global.nc.u32 %r153, [%rd50];
ld.global.nc.u32 %r157, [%rd50+16];
add.s64 %rd51, %rd49, %rd33;
ld.global.nc.u32 %r161, [%rd51];
ld.global.nc.u32 %r165, [%rd50+20];

{ cvt.f32.f16 %f97, %rs1;}



dp4a.s32.s32 %r151, %r296, %r153, %r316;


dp4a.s32.s32 %r155, %r300, %r157, %r151;


dp4a.s32.s32 %r159, %r304, %r161, %r155;


dp4a.s32.s32 %r163, %r308, %r165, %r159;

ld.global.nc.u32 %r167, [%rd48];

{.reg .f16 low,high;
mov.b32 {low,high},%r167;
cvt.f32.f16 %f98, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r167;
cvt.f32.f16 %f99, high;}


cvt.rn.f32.s32 %f143, %r163;
mul.ftz.f32 %f144, %f98, %f143;
mul.ftz.f32 %f145, %f99, 0f40800000;
sub.ftz.f32 %f146, %f144, %f145;
fma.rn.ftz.f32 %f360, %f97, %f146, %f360;

{ cvt.f32.f16 %f100, %rs2;}



dp4a.s32.s32 %r169, %r314, %r153, %r316;


dp4a.s32.s32 %r173, %r318, %r157, %r169;


dp4a.s32.s32 %r177, %r322, %r161, %r173;


dp4a.s32.s32 %r181, %r326, %r165, %r177;


{.reg .f16 low,high;
mov.b32 {low,high},%r167;
cvt.f32.f16 %f101, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r167;
cvt.f32.f16 %f102, high;}


cvt.rn.f32.s32 %f147, %r181;
mul.ftz.f32 %f148, %f101, %f147;
mul.ftz.f32 %f149, %f102, 0f40800000;
sub.ftz.f32 %f150, %f148, %f149;
fma.rn.ftz.f32 %f359, %f100, %f150, %f359;
cvt.u64.u32 %rd52, %r601;
add.s64 %rd53, %rd1, %rd52;
mul.lo.s64 %rd54, %rd53, 36;
add.s64 %rd55, %rd9, %rd54;
add.s64 %rd56, %rd55, 4;
add.s64 %rd57, %rd56, %rd31;
ld.global.nc.u32 %r189, [%rd57];
ld.global.nc.u32 %r193, [%rd57+16];
add.s64 %rd58, %rd56, %rd33;
ld.global.nc.u32 %r197, [%rd58];
ld.global.nc.u32 %r201, [%rd57+20];

{ cvt.f32.f16 %f103, %rs1;}



dp4a.s32.s32 %r187, %r296, %r189, %r316;


dp4a.s32.s32 %r191, %r300, %r193, %r187;


dp4a.s32.s32 %r195, %r304, %r197, %r191;


dp4a.s32.s32 %r199, %r308, %r201, %r195;

ld.global.nc.u32 %r203, [%rd55];

{.reg .f16 low,high;
mov.b32 {low,high},%r203;
cvt.f32.f16 %f104, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r203;
cvt.f32.f16 %f105, high;}


cvt.rn.f32.s32 %f151, %r199;
mul.ftz.f32 %f152, %f104, %f151;
mul.ftz.f32 %f153, %f105, 0f40800000;
sub.ftz.f32 %f154, %f152, %f153;
fma.rn.ftz.f32 %f358, %f103, %f154, %f358;

{ cvt.f32.f16 %f106, %rs2;}



dp4a.s32.s32 %r205, %r314, %r189, %r316;


dp4a.s32.s32 %r209, %r318, %r193, %r205;


dp4a.s32.s32 %r213, %r322, %r197, %r209;


dp4a.s32.s32 %r217, %r326, %r201, %r213;


{.reg .f16 low,high;
mov.b32 {low,high},%r203;
cvt.f32.f16 %f107, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r203;
cvt.f32.f16 %f108, high;}


cvt.rn.f32.s32 %f155, %r217;
mul.ftz.f32 %f156, %f107, %f155;
mul.ftz.f32 %f157, %f108, 0f40800000;
sub.ftz.f32 %f158, %f156, %f157;
fma.rn.ftz.f32 %f357, %f106, %f158, %f357;
cvt.u64.u32 %rd59, %r600;
add.s64 %rd60, %rd1, %rd59;
mul.lo.s64 %rd61, %rd60, 36;
add.s64 %rd62, %rd9, %rd61;
add.s64 %rd63, %rd62, 4;
add.s64 %rd64, %rd63, %rd31;
ld.global.nc.u32 %r225, [%rd64];
ld.global.nc.u32 %r229, [%rd64+16];
add.s64 %rd65, %rd63, %rd33;
ld.global.nc.u32 %r233, [%rd65];
ld.global.nc.u32 %r237, [%rd64+20];

{ cvt.f32.f16 %f109, %rs1;}



dp4a.s32.s32 %r223, %r296, %r225, %r316;


dp4a.s32.s32 %r227, %r300, %r229, %r223;


dp4a.s32.s32 %r231, %r304, %r233, %r227;


dp4a.s32.s32 %r235, %r308, %r237, %r231;

ld.global.nc.u32 %r239, [%rd62];

{.reg .f16 low,high;
mov.b32 {low,high},%r239;
cvt.f32.f16 %f110, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r239;
cvt.f32.f16 %f111, high;}


cvt.rn.f32.s32 %f159, %r235;
mul.ftz.f32 %f160, %f110, %f159;
mul.ftz.f32 %f161, %f111, 0f40800000;
sub.ftz.f32 %f162, %f160, %f161;
fma.rn.ftz.f32 %f356, %f109, %f162, %f356;

{ cvt.f32.f16 %f112, %rs2;}



dp4a.s32.s32 %r241, %r314, %r225, %r316;


dp4a.s32.s32 %r245, %r318, %r229, %r241;


dp4a.s32.s32 %r249, %r322, %r233, %r245;


dp4a.s32.s32 %r253, %r326, %r237, %r249;


{.reg .f16 low,high;
mov.b32 {low,high},%r239;
cvt.f32.f16 %f113, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r239;
cvt.f32.f16 %f114, high;}


cvt.rn.f32.s32 %f163, %r253;
mul.ftz.f32 %f164, %f113, %f163;
mul.ftz.f32 %f165, %f114, 0f40800000;
sub.ftz.f32 %f166, %f164, %f165;
fma.rn.ftz.f32 %f355, %f112, %f166, %f355;
cvt.u64.u32 %rd66, %r599;
add.s64 %rd67, %rd1, %rd66;
mul.lo.s64 %rd68, %rd67, 36;
add.s64 %rd69, %rd9, %rd68;
add.s64 %rd70, %rd69, 4;
add.s64 %rd71, %rd70, %rd31;
ld.global.nc.u32 %r261, [%rd71];
ld.global.nc.u32 %r265, [%rd71+16];
add.s64 %rd72, %rd70, %rd33;
ld.global.nc.u32 %r269, [%rd72];
ld.global.nc.u32 %r273, [%rd71+20];

{ cvt.f32.f16 %f115, %rs1;}



dp4a.s32.s32 %r259, %r296, %r261, %r316;


dp4a.s32.s32 %r263, %r300, %r265, %r259;


dp4a.s32.s32 %r267, %r304, %r269, %r263;


dp4a.s32.s32 %r271, %r308, %r273, %r267;

ld.global.nc.u32 %r275, [%rd69];

{.reg .f16 low,high;
mov.b32 {low,high},%r275;
cvt.f32.f16 %f116, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r275;
cvt.f32.f16 %f117, high;}


cvt.rn.f32.s32 %f167, %r271;
mul.ftz.f32 %f168, %f116, %f167;
mul.ftz.f32 %f169, %f117, 0f40800000;
sub.ftz.f32 %f170, %f168, %f169;
fma.rn.ftz.f32 %f354, %f115, %f170, %f354;

{ cvt.f32.f16 %f118, %rs2;}



dp4a.s32.s32 %r277, %r314, %r261, %r316;


dp4a.s32.s32 %r281, %r318, %r265, %r277;


dp4a.s32.s32 %r285, %r322, %r269, %r281;


dp4a.s32.s32 %r289, %r326, %r273, %r285;


{.reg .f16 low,high;
mov.b32 {low,high},%r275;
cvt.f32.f16 %f119, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r275;
cvt.f32.f16 %f120, high;}


cvt.rn.f32.s32 %f171, %r289;
mul.ftz.f32 %f172, %f119, %f171;
mul.ftz.f32 %f173, %f120, 0f40800000;
sub.ftz.f32 %f174, %f172, %f173;
fma.rn.ftz.f32 %f353, %f118, %f174, %f353;
cvt.u64.u32 %rd73, %r598;
add.s64 %rd74, %rd1, %rd73;
mul.lo.s64 %rd75, %rd74, 36;
add.s64 %rd76, %rd9, %rd75;
add.s64 %rd77, %rd76, 4;
add.s64 %rd78, %rd77, %rd31;
ld.global.nc.u32 %r297, [%rd78];
ld.global.nc.u32 %r301, [%rd78+16];
add.s64 %rd79, %rd77, %rd33;
ld.global.nc.u32 %r305, [%rd79];
ld.global.nc.u32 %r309, [%rd78+20];

{ cvt.f32.f16 %f121, %rs1;}



dp4a.s32.s32 %r295, %r296, %r297, %r316;


dp4a.s32.s32 %r299, %r300, %r301, %r295;


dp4a.s32.s32 %r303, %r304, %r305, %r299;


dp4a.s32.s32 %r307, %r308, %r309, %r303;

ld.global.nc.u32 %r311, [%rd76];

{.reg .f16 low,high;
mov.b32 {low,high},%r311;
cvt.f32.f16 %f122, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r311;
cvt.f32.f16 %f123, high;}


cvt.rn.f32.s32 %f175, %r307;
mul.ftz.f32 %f176, %f122, %f175;
mul.ftz.f32 %f177, %f123, 0f40800000;
sub.ftz.f32 %f178, %f176, %f177;
fma.rn.ftz.f32 %f352, %f121, %f178, %f352;

{ cvt.f32.f16 %f124, %rs2;}



dp4a.s32.s32 %r313, %r314, %r297, %r316;


dp4a.s32.s32 %r317, %r318, %r301, %r313;


dp4a.s32.s32 %r321, %r322, %r305, %r317;


dp4a.s32.s32 %r325, %r326, %r309, %r321;


{.reg .f16 low,high;
mov.b32 {low,high},%r311;
cvt.f32.f16 %f125, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r311;
cvt.f32.f16 %f126, high;}


cvt.rn.f32.s32 %f179, %r325;
mul.ftz.f32 %f180, %f125, %f179;
mul.ftz.f32 %f181, %f126, 0f40800000;
sub.ftz.f32 %f182, %f180, %f181;
fma.rn.ftz.f32 %f351, %f124, %f182, %f351;
add.s32 %r602, %r602, 32;
add.s32 %r601, %r601, 32;
add.s32 %r600, %r600, 32;
add.s32 %r599, %r599, 32;
add.s32 %r598, %r598, 32;
add.s32 %r597, %r597, 32;
add.s64 %rd124, %rd124, 576;
add.s32 %r603, %r603, 32;
setp.lt.s32 %p4, %r603, %r6;
@%p4 bra $L__BB39_2;

$L__BB39_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB39_5;

mov.u32 %r339, _ZZ13mul_mat_vec_qIL9ggml_type2ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r340, %r1, 1792, %r339;
shl.b32 %r341, %r2, 2;
add.s32 %r342, %r340, %r341;
st.shared.f32 [%r342+-1792], %f364;
st.shared.f32 [%r342+-1664], %f363;
st.shared.f32 [%r342+-1536], %f362;
st.shared.f32 [%r342+-1408], %f361;
st.shared.f32 [%r342+-1280], %f360;
st.shared.f32 [%r342+-1152], %f359;
st.shared.f32 [%r342+-1024], %f358;
st.shared.f32 [%r342+-896], %f357;
st.shared.f32 [%r342+-768], %f356;
st.shared.f32 [%r342+-640], %f355;
st.shared.f32 [%r342+-512], %f354;
st.shared.f32 [%r342+-384], %f353;
st.shared.f32 [%r342+-256], %f352;
st.shared.f32 [%r342+-128], %f351;

$L__BB39_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB39_34;

mov.u32 %r343, %ctaid.z;
mov.u32 %r344, %ctaid.y;
mad.lo.s32 %r345, %r344, %r38, %r4;
mad.lo.s32 %r346, %r343, %r44, %r345;
cvt.u64.u32 %rd12, %r346;
shl.b32 %r347, %r2, 2;
mov.u32 %r348, _ZZ13mul_mat_vec_qIL9ggml_type2ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r27, %r348, %r347;
mov.u32 %r349, 2;
ld.shared.f32 %f183, [%r27];
add.ftz.f32 %f184, %f183, %f364;
mov.b32 %r350, %f184;
mov.u32 %r351, 31;
mov.u32 %r352, 16;
mov.u32 %r353, -1;
shfl.sync.bfly.b32 %r354|%p7, %r350, %r352, %r351, %r353;
mov.b32 %f185, %r354;
add.ftz.f32 %f186, %f184, %f185;
mov.b32 %r355, %f186;
mov.u32 %r356, 8;
shfl.sync.bfly.b32 %r357|%p8, %r355, %r356, %r351, %r353;
mov.b32 %f187, %r357;
add.ftz.f32 %f188, %f186, %f187;
mov.b32 %r358, %f188;
mov.u32 %r359, 4;
shfl.sync.bfly.b32 %r360|%p9, %r358, %r359, %r351, %r353;
mov.b32 %f189, %r360;
add.ftz.f32 %f190, %f188, %f189;
mov.b32 %r361, %f190;
shfl.sync.bfly.b32 %r362|%p10, %r361, %r349, %r351, %r353;
mov.b32 %f191, %r362;
add.ftz.f32 %f192, %f190, %f191;
mov.b32 %r363, %f192;
mov.u32 %r364, 1;
shfl.sync.bfly.b32 %r365|%p11, %r363, %r364, %r351, %r353;
mov.b32 %f193, %r365;
add.ftz.f32 %f43, %f192, %f193;
setp.lt.u32 %p12, %r4, %r32;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
cvta.to.global.u64 %rd80, %rd19;
mul.wide.u32 %rd81, %r346, 4;
add.s64 %rd13, %rd80, %rd81;
not.pred %p14, %p1;
@%p14 bra $L__BB39_8;

st.global.f32 [%rd13], %f43;

$L__BB39_8:
ld.shared.f32 %f194, [%r27+128];
add.ftz.f32 %f195, %f194, %f363;
mov.b32 %r366, %f195;
shfl.sync.bfly.b32 %r370|%p15, %r366, %r352, %r351, %r353;
mov.b32 %f196, %r370;
add.ftz.f32 %f197, %f195, %f196;
mov.b32 %r371, %f197;
shfl.sync.bfly.b32 %r373|%p16, %r371, %r356, %r351, %r353;
mov.b32 %f198, %r373;
add.ftz.f32 %f199, %f197, %f198;
mov.b32 %r374, %f199;
shfl.sync.bfly.b32 %r376|%p17, %r374, %r359, %r351, %r353;
mov.b32 %f200, %r376;
add.ftz.f32 %f201, %f199, %f200;
mov.b32 %r377, %f201;
shfl.sync.bfly.b32 %r379|%p18, %r377, %r349, %r351, %r353;
mov.b32 %f202, %r379;
add.ftz.f32 %f203, %f201, %f202;
mov.b32 %r380, %f203;
shfl.sync.bfly.b32 %r382|%p19, %r380, %r364, %r351, %r353;
mov.b32 %f204, %r382;
add.ftz.f32 %f44, %f203, %f204;
setp.eq.s32 %p20, %r2, 1;
or.b32 %r386, %r4, 1;
setp.lt.u32 %p21, %r386, %r32;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB39_10;

st.global.f32 [%rd13+4], %f44;

$L__BB39_10:
ld.shared.f32 %f205, [%r27+256];
add.ftz.f32 %f206, %f205, %f362;
mov.b32 %r387, %f206;
mov.u32 %r388, 31;
mov.u32 %r389, 16;
mov.u32 %r390, -1;
shfl.sync.bfly.b32 %r391|%p23, %r387, %r389, %r388, %r390;
mov.b32 %f207, %r391;
add.ftz.f32 %f208, %f206, %f207;
mov.b32 %r392, %f208;
mov.u32 %r393, 8;
shfl.sync.bfly.b32 %r394|%p24, %r392, %r393, %r388, %r390;
mov.b32 %f209, %r394;
add.ftz.f32 %f210, %f208, %f209;
mov.b32 %r395, %f210;
mov.u32 %r396, 4;
shfl.sync.bfly.b32 %r397|%p25, %r395, %r396, %r388, %r390;
mov.b32 %f211, %r397;
add.ftz.f32 %f212, %f210, %f211;
mov.b32 %r398, %f212;
mov.u32 %r399, 2;
shfl.sync.bfly.b32 %r400|%p26, %r398, %r399, %r388, %r390;
mov.b32 %f213, %r400;
add.ftz.f32 %f214, %f212, %f213;
mov.b32 %r401, %f214;
mov.u32 %r402, 1;
shfl.sync.bfly.b32 %r403|%p27, %r401, %r402, %r388, %r390;
mov.b32 %f215, %r403;
add.ftz.f32 %f45, %f214, %f215;
@%p14 bra $L__BB39_12;

cvt.u64.u32 %rd82, %r32;
add.s64 %rd83, %rd82, %rd12;
shl.b64 %rd85, %rd83, 2;
add.s64 %rd86, %rd80, %rd85;
st.global.f32 [%rd86], %f45;

$L__BB39_12:
ld.shared.f32 %f216, [%r27+384];
add.ftz.f32 %f217, %f216, %f361;
mov.b32 %r404, %f217;
shfl.sync.bfly.b32 %r408|%p29, %r404, %r389, %r388, %r390;
mov.b32 %f218, %r408;
add.ftz.f32 %f219, %f217, %f218;
mov.b32 %r409, %f219;
shfl.sync.bfly.b32 %r411|%p30, %r409, %r393, %r388, %r390;
mov.b32 %f220, %r411;
add.ftz.f32 %f221, %f219, %f220;
mov.b32 %r412, %f221;
shfl.sync.bfly.b32 %r414|%p31, %r412, %r396, %r388, %r390;
mov.b32 %f222, %r414;
add.ftz.f32 %f223, %f221, %f222;
mov.b32 %r415, %f223;
shfl.sync.bfly.b32 %r417|%p32, %r415, %r399, %r388, %r390;
mov.b32 %f224, %r417;
add.ftz.f32 %f225, %f223, %f224;
mov.b32 %r418, %f225;
shfl.sync.bfly.b32 %r420|%p33, %r418, %r402, %r388, %r390;
mov.b32 %f226, %r420;
add.ftz.f32 %f46, %f225, %f226;
@%p22 bra $L__BB39_14;

add.s32 %r421, %r32, 1;
cvt.u64.u32 %rd87, %r421;
add.s64 %rd88, %rd87, %rd12;
shl.b64 %rd90, %rd88, 2;
add.s64 %rd91, %rd80, %rd90;
st.global.f32 [%rd91], %f46;

$L__BB39_14:
shl.b32 %r422, %r32, 1;
mov.u32 %r423, 1;
ld.shared.f32 %f227, [%r27+512];
add.ftz.f32 %f228, %f227, %f360;
mov.b32 %r424, %f228;
mov.u32 %r425, 31;
mov.u32 %r426, 16;
mov.u32 %r427, -1;
shfl.sync.bfly.b32 %r428|%p35, %r424, %r426, %r425, %r427;
mov.b32 %f229, %r428;
add.ftz.f32 %f230, %f228, %f229;
mov.b32 %r429, %f230;
mov.u32 %r430, 8;
shfl.sync.bfly.b32 %r431|%p36, %r429, %r430, %r425, %r427;
mov.b32 %f231, %r431;
add.ftz.f32 %f232, %f230, %f231;
mov.b32 %r432, %f232;
mov.u32 %r433, 4;
shfl.sync.bfly.b32 %r434|%p37, %r432, %r433, %r425, %r427;
mov.b32 %f233, %r434;
add.ftz.f32 %f234, %f232, %f233;
mov.b32 %r435, %f234;
mov.u32 %r436, 2;
shfl.sync.bfly.b32 %r437|%p38, %r435, %r436, %r425, %r427;
mov.b32 %f235, %r437;
add.ftz.f32 %f236, %f234, %f235;
mov.b32 %r438, %f236;
shfl.sync.bfly.b32 %r439|%p39, %r438, %r423, %r425, %r427;
mov.b32 %f237, %r439;
add.ftz.f32 %f47, %f236, %f237;
cvt.u64.u32 %rd92, %r422;
add.s64 %rd93, %rd92, %rd12;
shl.b64 %rd95, %rd93, 2;
add.s64 %rd14, %rd80, %rd95;
@%p14 bra $L__BB39_16;

st.global.f32 [%rd14], %f47;

$L__BB39_16:
ld.shared.f32 %f238, [%r27+640];
add.ftz.f32 %f239, %f238, %f359;
mov.b32 %r440, %f239;
shfl.sync.bfly.b32 %r444|%p41, %r440, %r426, %r425, %r427;
mov.b32 %f240, %r444;
add.ftz.f32 %f241, %f239, %f240;
mov.b32 %r445, %f241;
shfl.sync.bfly.b32 %r447|%p42, %r445, %r430, %r425, %r427;
mov.b32 %f242, %r447;
add.ftz.f32 %f243, %f241, %f242;
mov.b32 %r448, %f243;
shfl.sync.bfly.b32 %r450|%p43, %r448, %r433, %r425, %r427;
mov.b32 %f244, %r450;
add.ftz.f32 %f245, %f243, %f244;
mov.b32 %r451, %f245;
shfl.sync.bfly.b32 %r453|%p44, %r451, %r436, %r425, %r427;
mov.b32 %f246, %r453;
add.ftz.f32 %f247, %f245, %f246;
mov.b32 %r454, %f247;
shfl.sync.bfly.b32 %r456|%p45, %r454, %r423, %r425, %r427;
mov.b32 %f248, %r456;
add.ftz.f32 %f48, %f247, %f248;
@%p22 bra $L__BB39_18;

st.global.f32 [%rd14+4], %f48;

$L__BB39_18:
mul.lo.s32 %r28, %r32, 3;
ld.shared.f32 %f249, [%r27+768];
add.ftz.f32 %f250, %f249, %f358;
mov.b32 %r457, %f250;
mov.u32 %r458, 31;
mov.u32 %r459, 16;
mov.u32 %r460, -1;
shfl.sync.bfly.b32 %r461|%p47, %r457, %r459, %r458, %r460;
mov.b32 %f251, %r461;
add.ftz.f32 %f252, %f250, %f251;
mov.b32 %r462, %f252;
mov.u32 %r463, 8;
shfl.sync.bfly.b32 %r464|%p48, %r462, %r463, %r458, %r460;
mov.b32 %f253, %r464;
add.ftz.f32 %f254, %f252, %f253;
mov.b32 %r465, %f254;
mov.u32 %r466, 4;
shfl.sync.bfly.b32 %r467|%p49, %r465, %r466, %r458, %r460;
mov.b32 %f255, %r467;
add.ftz.f32 %f256, %f254, %f255;
mov.b32 %r468, %f256;
mov.u32 %r469, 2;
shfl.sync.bfly.b32 %r470|%p50, %r468, %r469, %r458, %r460;
mov.b32 %f257, %r470;
add.ftz.f32 %f258, %f256, %f257;
mov.b32 %r471, %f258;
mov.u32 %r472, 1;
shfl.sync.bfly.b32 %r473|%p51, %r471, %r472, %r458, %r460;
mov.b32 %f259, %r473;
add.ftz.f32 %f49, %f258, %f259;
@%p14 bra $L__BB39_20;

cvt.u64.u32 %rd96, %r28;
add.s64 %rd97, %rd96, %rd12;
shl.b64 %rd99, %rd97, 2;
add.s64 %rd100, %rd80, %rd99;
st.global.f32 [%rd100], %f49;

$L__BB39_20:
ld.shared.f32 %f260, [%r27+896];
add.ftz.f32 %f261, %f260, %f357;
mov.b32 %r474, %f261;
shfl.sync.bfly.b32 %r478|%p53, %r474, %r459, %r458, %r460;
mov.b32 %f262, %r478;
add.ftz.f32 %f263, %f261, %f262;
mov.b32 %r479, %f263;
shfl.sync.bfly.b32 %r481|%p54, %r479, %r463, %r458, %r460;
mov.b32 %f264, %r481;
add.ftz.f32 %f265, %f263, %f264;
mov.b32 %r482, %f265;
shfl.sync.bfly.b32 %r484|%p55, %r482, %r466, %r458, %r460;
mov.b32 %f266, %r484;
add.ftz.f32 %f267, %f265, %f266;
mov.b32 %r485, %f267;
shfl.sync.bfly.b32 %r487|%p56, %r485, %r469, %r458, %r460;
mov.b32 %f268, %r487;
add.ftz.f32 %f269, %f267, %f268;
mov.b32 %r488, %f269;
shfl.sync.bfly.b32 %r490|%p57, %r488, %r472, %r458, %r460;
mov.b32 %f270, %r490;
add.ftz.f32 %f50, %f269, %f270;
@%p22 bra $L__BB39_22;

add.s32 %r491, %r28, 1;
cvt.u64.u32 %rd101, %r491;
add.s64 %rd102, %rd101, %rd12;
shl.b64 %rd104, %rd102, 2;
add.s64 %rd105, %rd80, %rd104;
st.global.f32 [%rd105], %f50;

$L__BB39_22:
shl.b32 %r492, %r32, 2;
mov.u32 %r493, 2;
ld.shared.f32 %f271, [%r27+1024];
add.ftz.f32 %f272, %f271, %f356;
mov.b32 %r494, %f272;
mov.u32 %r495, 31;
mov.u32 %r496, 16;
mov.u32 %r497, -1;
shfl.sync.bfly.b32 %r498|%p59, %r494, %r496, %r495, %r497;
mov.b32 %f273, %r498;
add.ftz.f32 %f274, %f272, %f273;
mov.b32 %r499, %f274;
mov.u32 %r500, 8;
shfl.sync.bfly.b32 %r501|%p60, %r499, %r500, %r495, %r497;
mov.b32 %f275, %r501;
add.ftz.f32 %f276, %f274, %f275;
mov.b32 %r502, %f276;
mov.u32 %r503, 4;
shfl.sync.bfly.b32 %r504|%p61, %r502, %r503, %r495, %r497;
mov.b32 %f277, %r504;
add.ftz.f32 %f278, %f276, %f277;
mov.b32 %r505, %f278;
shfl.sync.bfly.b32 %r506|%p62, %r505, %r493, %r495, %r497;
mov.b32 %f279, %r506;
add.ftz.f32 %f280, %f278, %f279;
mov.b32 %r507, %f280;
mov.u32 %r508, 1;
shfl.sync.bfly.b32 %r509|%p63, %r507, %r508, %r495, %r497;
mov.b32 %f281, %r509;
add.ftz.f32 %f51, %f280, %f281;
cvt.u64.u32 %rd106, %r492;
add.s64 %rd107, %rd106, %rd12;
shl.b64 %rd109, %rd107, 2;
add.s64 %rd15, %rd80, %rd109;
@%p14 bra $L__BB39_24;

st.global.f32 [%rd15], %f51;

$L__BB39_24:
ld.shared.f32 %f282, [%r27+1152];
add.ftz.f32 %f283, %f282, %f355;
mov.b32 %r510, %f283;
shfl.sync.bfly.b32 %r514|%p65, %r510, %r496, %r495, %r497;
mov.b32 %f284, %r514;
add.ftz.f32 %f285, %f283, %f284;
mov.b32 %r515, %f285;
shfl.sync.bfly.b32 %r517|%p66, %r515, %r500, %r495, %r497;
mov.b32 %f286, %r517;
add.ftz.f32 %f287, %f285, %f286;
mov.b32 %r518, %f287;
shfl.sync.bfly.b32 %r520|%p67, %r518, %r503, %r495, %r497;
mov.b32 %f288, %r520;
add.ftz.f32 %f289, %f287, %f288;
mov.b32 %r521, %f289;
shfl.sync.bfly.b32 %r523|%p68, %r521, %r493, %r495, %r497;
mov.b32 %f290, %r523;
add.ftz.f32 %f291, %f289, %f290;
mov.b32 %r524, %f291;
shfl.sync.bfly.b32 %r526|%p69, %r524, %r508, %r495, %r497;
mov.b32 %f292, %r526;
add.ftz.f32 %f52, %f291, %f292;
@%p22 bra $L__BB39_26;

st.global.f32 [%rd15+4], %f52;

$L__BB39_26:
mul.lo.s32 %r29, %r32, 5;
ld.shared.f32 %f293, [%r27+1280];
add.ftz.f32 %f294, %f293, %f354;
mov.b32 %r527, %f294;
mov.u32 %r528, 31;
mov.u32 %r529, 16;
mov.u32 %r530, -1;
shfl.sync.bfly.b32 %r531|%p71, %r527, %r529, %r528, %r530;
mov.b32 %f295, %r531;
add.ftz.f32 %f296, %f294, %f295;
mov.b32 %r532, %f296;
mov.u32 %r533, 8;
shfl.sync.bfly.b32 %r534|%p72, %r532, %r533, %r528, %r530;
mov.b32 %f297, %r534;
add.ftz.f32 %f298, %f296, %f297;
mov.b32 %r535, %f298;
mov.u32 %r536, 4;
shfl.sync.bfly.b32 %r537|%p73, %r535, %r536, %r528, %r530;
mov.b32 %f299, %r537;
add.ftz.f32 %f300, %f298, %f299;
mov.b32 %r538, %f300;
mov.u32 %r539, 2;
shfl.sync.bfly.b32 %r540|%p74, %r538, %r539, %r528, %r530;
mov.b32 %f301, %r540;
add.ftz.f32 %f302, %f300, %f301;
mov.b32 %r541, %f302;
mov.u32 %r542, 1;
shfl.sync.bfly.b32 %r543|%p75, %r541, %r542, %r528, %r530;
mov.b32 %f303, %r543;
add.ftz.f32 %f53, %f302, %f303;
@%p14 bra $L__BB39_28;

cvt.u64.u32 %rd110, %r29;
add.s64 %rd111, %rd110, %rd12;
shl.b64 %rd113, %rd111, 2;
add.s64 %rd114, %rd80, %rd113;
st.global.f32 [%rd114], %f53;

$L__BB39_28:
ld.shared.f32 %f304, [%r27+1408];
add.ftz.f32 %f305, %f304, %f353;
mov.b32 %r544, %f305;
shfl.sync.bfly.b32 %r548|%p77, %r544, %r529, %r528, %r530;
mov.b32 %f306, %r548;
add.ftz.f32 %f307, %f305, %f306;
mov.b32 %r549, %f307;
shfl.sync.bfly.b32 %r551|%p78, %r549, %r533, %r528, %r530;
mov.b32 %f308, %r551;
add.ftz.f32 %f309, %f307, %f308;
mov.b32 %r552, %f309;
shfl.sync.bfly.b32 %r554|%p79, %r552, %r536, %r528, %r530;
mov.b32 %f310, %r554;
add.ftz.f32 %f311, %f309, %f310;
mov.b32 %r555, %f311;
shfl.sync.bfly.b32 %r557|%p80, %r555, %r539, %r528, %r530;
mov.b32 %f312, %r557;
add.ftz.f32 %f313, %f311, %f312;
mov.b32 %r558, %f313;
shfl.sync.bfly.b32 %r560|%p81, %r558, %r542, %r528, %r530;
mov.b32 %f314, %r560;
add.ftz.f32 %f54, %f313, %f314;
@%p22 bra $L__BB39_30;

add.s32 %r561, %r29, 1;
cvt.u64.u32 %rd115, %r561;
add.s64 %rd116, %rd115, %rd12;
shl.b64 %rd118, %rd116, 2;
add.s64 %rd119, %rd80, %rd118;
st.global.f32 [%rd119], %f54;

$L__BB39_30:
mul.lo.s32 %r562, %r32, 6;
ld.shared.f32 %f315, [%r27+1536];
add.ftz.f32 %f316, %f315, %f352;
mov.b32 %r563, %f316;
mov.u32 %r564, 31;
mov.u32 %r565, 16;
mov.u32 %r566, -1;
shfl.sync.bfly.b32 %r567|%p83, %r563, %r565, %r564, %r566;
mov.b32 %f317, %r567;
add.ftz.f32 %f318, %f316, %f317;
mov.b32 %r568, %f318;
mov.u32 %r569, 8;
shfl.sync.bfly.b32 %r570|%p84, %r568, %r569, %r564, %r566;
mov.b32 %f319, %r570;
add.ftz.f32 %f320, %f318, %f319;
mov.b32 %r571, %f320;
mov.u32 %r572, 4;
shfl.sync.bfly.b32 %r573|%p85, %r571, %r572, %r564, %r566;
mov.b32 %f321, %r573;
add.ftz.f32 %f322, %f320, %f321;
mov.b32 %r574, %f322;
mov.u32 %r575, 2;
shfl.sync.bfly.b32 %r576|%p86, %r574, %r575, %r564, %r566;
mov.b32 %f323, %r576;
add.ftz.f32 %f324, %f322, %f323;
mov.b32 %r577, %f324;
mov.u32 %r578, 1;
shfl.sync.bfly.b32 %r579|%p87, %r577, %r578, %r564, %r566;
mov.b32 %f325, %r579;
add.ftz.f32 %f55, %f324, %f325;
cvt.u64.u32 %rd120, %r562;
add.s64 %rd121, %rd120, %rd12;
shl.b64 %rd123, %rd121, 2;
add.s64 %rd16, %rd80, %rd123;
@%p14 bra $L__BB39_32;

st.global.f32 [%rd16], %f55;

$L__BB39_32:
ld.shared.f32 %f326, [%r27+1664];
add.ftz.f32 %f327, %f326, %f351;
mov.b32 %r580, %f327;
shfl.sync.bfly.b32 %r584|%p89, %r580, %r565, %r564, %r566;
mov.b32 %f328, %r584;
add.ftz.f32 %f329, %f327, %f328;
mov.b32 %r585, %f329;
shfl.sync.bfly.b32 %r587|%p90, %r585, %r569, %r564, %r566;
mov.b32 %f330, %r587;
add.ftz.f32 %f331, %f329, %f330;
mov.b32 %r588, %f331;
shfl.sync.bfly.b32 %r590|%p91, %r588, %r572, %r564, %r566;
mov.b32 %f332, %r590;
add.ftz.f32 %f333, %f331, %f332;
mov.b32 %r591, %f333;
shfl.sync.bfly.b32 %r593|%p92, %r591, %r575, %r564, %r566;
mov.b32 %f334, %r593;
add.ftz.f32 %f335, %f333, %f334;
mov.b32 %r594, %f335;
shfl.sync.bfly.b32 %r596|%p93, %r594, %r578, %r564, %r566;
mov.b32 %f336, %r596;
add.ftz.f32 %f56, %f335, %f336;
@%p22 bra $L__BB39_34;

st.global.f32 [%rd16+4], %f56;

$L__BB39_34:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type2ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type2ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type2ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type2ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type2ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type2ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type2ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type2ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type2ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type2ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<107>;
.reg .b16 %rs<25>;
.reg .f32 %f<417>;
.reg .b32 %r<681>;
.reg .b64 %rd<142>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type2ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[2048];

ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type2ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type2ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type2ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r49, [_Z13mul_mat_vec_qIL9ggml_type2ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type2ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type2ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type2ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type2ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type2ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type2ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r46, [_Z13mul_mat_vec_qIL9ggml_type2ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r47, [_Z13mul_mat_vec_qIL9ggml_type2ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r48, [_Z13mul_mat_vec_qIL9ggml_type2ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type2ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type2ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type2ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type2ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r50, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r50, %r2;
mov.u32 %r51, %ctaid.x;
shl.b32 %r4, %r51, 1;
shr.u32 %r52, %r3, 31;
add.s32 %r53, %r3, %r52;
shr.s32 %r680, %r53, 1;
shr.u32 %r6, %r49, 5;
setp.ge.s32 %p3, %r680, %r6;
mov.f32 %f401, 0f00000000;
mov.f32 %f402, %f401;
mov.f32 %f403, %f401;
mov.f32 %f404, %f401;
mov.f32 %f405, %f401;
mov.f32 %f406, %f401;
mov.f32 %f407, %f401;
mov.f32 %f408, %f401;
mov.f32 %f409, %f401;
mov.f32 %f410, %f401;
mov.f32 %f411, %f401;
mov.f32 %f412, %f401;
mov.f32 %f413, %f401;
mov.f32 %f414, %f401;
mov.f32 %f415, %f401;
mov.f32 %f416, %f401;
@%p3 bra $L__BB40_3;

mov.u32 %r54, %ctaid.y;
mov.u32 %r55, %ctaid.z;
mul.lo.s32 %r56, %r54, %r41;
cvt.u64.u32 %rd20, %r56;
mul.lo.s32 %r57, %r55, %r47;
cvt.u64.u32 %rd21, %r57;
add.s64 %rd1, %rd21, %rd20;
mul.hi.u32 %r58, %r55, %r43;
add.s32 %r59, %r58, %r55;
shr.u32 %r60, %r59, %r44;
mul.lo.s32 %r61, %r60, %r46;
mul.hi.u32 %r62, %r54, %r37;
add.s32 %r63, %r62, %r54;
shr.u32 %r64, %r63, %r38;
mul.lo.s32 %r65, %r64, %r40;
cvta.to.global.u64 %rd141, %rd17;
and.b32 %r68, %r53, -2;
sub.s32 %r69, %r3, %r68;
shl.b32 %r70, %r69, 1;
shl.b32 %r71, %r69, 2;
or.b32 %r72, %r71, 1;
cvt.s64.s32 %rd3, %r70;
sub.s32 %r73, %r72, %r70;
cvt.s64.s32 %rd4, %r73;
shl.b32 %r74, %r35, 1;
add.s32 %r679, %r680, %r74;
mad.lo.s32 %r678, %r35, 3, %r680;
shl.b32 %r75, %r35, 2;
add.s32 %r677, %r680, %r75;
mad.lo.s32 %r676, %r35, 5, %r680;
mad.lo.s32 %r675, %r35, 6, %r680;
mad.lo.s32 %r674, %r35, 7, %r680;
add.s32 %r673, %r680, %r35;
add.s32 %r76, %r680, %r61;
add.s32 %r77, %r76, %r65;
add.s32 %r78, %r4, 1;
mad.lo.s32 %r79, %r34, %r78, %r77;
mul.wide.s32 %rd5, %r79, 18;
mad.lo.s32 %r80, %r4, %r34, %r680;
add.s32 %r81, %r80, %r61;
add.s32 %r82, %r81, %r65;
mul.wide.s32 %rd6, %r82, 18;
mul.wide.s32 %rd22, %r71, 2;
add.s64 %rd23, %rd6, %rd22;
add.s64 %rd7, %rd23, 4;
add.s64 %rd24, %rd5, %rd22;
add.s64 %rd8, %rd24, 2;
cvta.to.global.u64 %rd9, %rd18;
mov.f32 %f401, 0f00000000;
shl.b64 %rd31, %rd3, 2;
shl.b64 %rd33, %rd4, 2;

$L__BB40_2:
cvt.u64.u32 %rd25, %r680;
add.s64 %rd26, %rd1, %rd25;
mul.lo.s64 %rd27, %rd26, 36;
add.s64 %rd28, %rd9, %rd27;
add.s64 %rd29, %rd28, 4;
add.s64 %rd30, %rd141, %rd7;
ld.global.nc.u16 %rs17, [%rd30+-2];
ld.global.nc.u16 %rs18, [%rd30];
mov.b32 %r371, {%rs17, %rs18};
add.s64 %rd32, %rd29, %rd31;
ld.global.nc.u32 %r85, [%rd32];
ld.global.nc.u32 %r89, [%rd32+16];
ld.global.nc.u16 %rs19, [%rd30+2];
ld.global.nc.u16 %rs20, [%rd30+4];
mov.b32 %r372, {%rs19, %rs20};
add.s64 %rd34, %rd29, %rd33;
ld.global.nc.u32 %r93, [%rd34];
ld.global.nc.u32 %r97, [%rd32+20];
add.s64 %rd35, %rd141, %rd6;
ld.global.nc.u16 %rs1, [%rd35];

{ cvt.f32.f16 %f97, %rs1;}


and.b32 %r336, %r371, 252645135;
shr.u32 %r373, %r371, 4;
and.b32 %r340, %r373, 252645135;
mov.u32 %r356, 0;

dp4a.s32.s32 %r83, %r336, %r85, %r356;


dp4a.s32.s32 %r87, %r340, %r89, %r83;

and.b32 %r344, %r372, 252645135;
shr.u32 %r374, %r372, 4;
and.b32 %r348, %r374, 252645135;

dp4a.s32.s32 %r91, %r344, %r93, %r87;


dp4a.s32.s32 %r95, %r348, %r97, %r91;

ld.global.nc.u32 %r99, [%rd28];

{.reg .f16 low,high;
mov.b32 {low,high},%r99;
cvt.f32.f16 %f98, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r99;
cvt.f32.f16 %f99, high;}


cvt.rn.f32.s32 %f145, %r95;
mul.ftz.f32 %f146, %f98, %f145;
mul.ftz.f32 %f147, %f99, 0f40800000;
sub.ftz.f32 %f148, %f146, %f147;
fma.rn.ftz.f32 %f416, %f97, %f148, %f416;
add.s64 %rd36, %rd141, %rd8;
ld.global.nc.u16 %rs21, [%rd36];
ld.global.nc.u16 %rs22, [%rd36+2];
mov.b32 %r375, {%rs21, %rs22};
ld.global.nc.u16 %rs23, [%rd36+4];
ld.global.nc.u16 %rs24, [%rd36+6];
mov.b32 %r376, {%rs23, %rs24};
add.s64 %rd37, %rd141, %rd5;
ld.global.nc.u16 %rs2, [%rd37];

{ cvt.f32.f16 %f100, %rs2;}


and.b32 %r354, %r375, 252645135;
shr.u32 %r377, %r375, 4;
and.b32 %r358, %r377, 252645135;

dp4a.s32.s32 %r101, %r354, %r85, %r356;


dp4a.s32.s32 %r105, %r358, %r89, %r101;

and.b32 %r362, %r376, 252645135;
shr.u32 %r378, %r376, 4;
and.b32 %r366, %r378, 252645135;

dp4a.s32.s32 %r109, %r362, %r93, %r105;


dp4a.s32.s32 %r113, %r366, %r97, %r109;


{.reg .f16 low,high;
mov.b32 {low,high},%r99;
cvt.f32.f16 %f101, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r99;
cvt.f32.f16 %f102, high;}


cvt.rn.f32.s32 %f149, %r113;
mul.ftz.f32 %f150, %f101, %f149;
mul.ftz.f32 %f151, %f102, 0f40800000;
sub.ftz.f32 %f152, %f150, %f151;
fma.rn.ftz.f32 %f415, %f100, %f152, %f415;
cvt.u64.u32 %rd38, %r673;
add.s64 %rd39, %rd1, %rd38;
mul.lo.s64 %rd40, %rd39, 36;
add.s64 %rd41, %rd9, %rd40;
add.s64 %rd42, %rd41, 4;
add.s64 %rd43, %rd42, %rd31;
ld.global.nc.u32 %r121, [%rd43];
ld.global.nc.u32 %r125, [%rd43+16];
add.s64 %rd44, %rd42, %rd33;
ld.global.nc.u32 %r129, [%rd44];
ld.global.nc.u32 %r133, [%rd43+20];

{ cvt.f32.f16 %f103, %rs1;}



dp4a.s32.s32 %r119, %r336, %r121, %r356;


dp4a.s32.s32 %r123, %r340, %r125, %r119;


dp4a.s32.s32 %r127, %r344, %r129, %r123;


dp4a.s32.s32 %r131, %r348, %r133, %r127;

ld.global.nc.u32 %r135, [%rd41];

{.reg .f16 low,high;
mov.b32 {low,high},%r135;
cvt.f32.f16 %f104, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r135;
cvt.f32.f16 %f105, high;}


cvt.rn.f32.s32 %f153, %r131;
mul.ftz.f32 %f154, %f104, %f153;
mul.ftz.f32 %f155, %f105, 0f40800000;
sub.ftz.f32 %f156, %f154, %f155;
fma.rn.ftz.f32 %f414, %f103, %f156, %f414;

{ cvt.f32.f16 %f106, %rs2;}



dp4a.s32.s32 %r137, %r354, %r121, %r356;


dp4a.s32.s32 %r141, %r358, %r125, %r137;


dp4a.s32.s32 %r145, %r362, %r129, %r141;


dp4a.s32.s32 %r149, %r366, %r133, %r145;


{.reg .f16 low,high;
mov.b32 {low,high},%r135;
cvt.f32.f16 %f107, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r135;
cvt.f32.f16 %f108, high;}


cvt.rn.f32.s32 %f157, %r149;
mul.ftz.f32 %f158, %f107, %f157;
mul.ftz.f32 %f159, %f108, 0f40800000;
sub.ftz.f32 %f160, %f158, %f159;
fma.rn.ftz.f32 %f413, %f106, %f160, %f413;
cvt.u64.u32 %rd45, %r679;
add.s64 %rd46, %rd1, %rd45;
mul.lo.s64 %rd47, %rd46, 36;
add.s64 %rd48, %rd9, %rd47;
add.s64 %rd49, %rd48, 4;
add.s64 %rd50, %rd49, %rd31;
ld.global.nc.u32 %r157, [%rd50];
ld.global.nc.u32 %r161, [%rd50+16];
add.s64 %rd51, %rd49, %rd33;
ld.global.nc.u32 %r165, [%rd51];
ld.global.nc.u32 %r169, [%rd50+20];

{ cvt.f32.f16 %f109, %rs1;}



dp4a.s32.s32 %r155, %r336, %r157, %r356;


dp4a.s32.s32 %r159, %r340, %r161, %r155;


dp4a.s32.s32 %r163, %r344, %r165, %r159;


dp4a.s32.s32 %r167, %r348, %r169, %r163;

ld.global.nc.u32 %r171, [%rd48];

{.reg .f16 low,high;
mov.b32 {low,high},%r171;
cvt.f32.f16 %f110, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r171;
cvt.f32.f16 %f111, high;}


cvt.rn.f32.s32 %f161, %r167;
mul.ftz.f32 %f162, %f110, %f161;
mul.ftz.f32 %f163, %f111, 0f40800000;
sub.ftz.f32 %f164, %f162, %f163;
fma.rn.ftz.f32 %f412, %f109, %f164, %f412;

{ cvt.f32.f16 %f112, %rs2;}



dp4a.s32.s32 %r173, %r354, %r157, %r356;


dp4a.s32.s32 %r177, %r358, %r161, %r173;


dp4a.s32.s32 %r181, %r362, %r165, %r177;


dp4a.s32.s32 %r185, %r366, %r169, %r181;


{.reg .f16 low,high;
mov.b32 {low,high},%r171;
cvt.f32.f16 %f113, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r171;
cvt.f32.f16 %f114, high;}


cvt.rn.f32.s32 %f165, %r185;
mul.ftz.f32 %f166, %f113, %f165;
mul.ftz.f32 %f167, %f114, 0f40800000;
sub.ftz.f32 %f168, %f166, %f167;
fma.rn.ftz.f32 %f411, %f112, %f168, %f411;
cvt.u64.u32 %rd52, %r678;
add.s64 %rd53, %rd1, %rd52;
mul.lo.s64 %rd54, %rd53, 36;
add.s64 %rd55, %rd9, %rd54;
add.s64 %rd56, %rd55, 4;
add.s64 %rd57, %rd56, %rd31;
ld.global.nc.u32 %r193, [%rd57];
ld.global.nc.u32 %r197, [%rd57+16];
add.s64 %rd58, %rd56, %rd33;
ld.global.nc.u32 %r201, [%rd58];
ld.global.nc.u32 %r205, [%rd57+20];

{ cvt.f32.f16 %f115, %rs1;}



dp4a.s32.s32 %r191, %r336, %r193, %r356;


dp4a.s32.s32 %r195, %r340, %r197, %r191;


dp4a.s32.s32 %r199, %r344, %r201, %r195;


dp4a.s32.s32 %r203, %r348, %r205, %r199;

ld.global.nc.u32 %r207, [%rd55];

{.reg .f16 low,high;
mov.b32 {low,high},%r207;
cvt.f32.f16 %f116, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r207;
cvt.f32.f16 %f117, high;}


cvt.rn.f32.s32 %f169, %r203;
mul.ftz.f32 %f170, %f116, %f169;
mul.ftz.f32 %f171, %f117, 0f40800000;
sub.ftz.f32 %f172, %f170, %f171;
fma.rn.ftz.f32 %f410, %f115, %f172, %f410;

{ cvt.f32.f16 %f118, %rs2;}



dp4a.s32.s32 %r209, %r354, %r193, %r356;


dp4a.s32.s32 %r213, %r358, %r197, %r209;


dp4a.s32.s32 %r217, %r362, %r201, %r213;


dp4a.s32.s32 %r221, %r366, %r205, %r217;


{.reg .f16 low,high;
mov.b32 {low,high},%r207;
cvt.f32.f16 %f119, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r207;
cvt.f32.f16 %f120, high;}


cvt.rn.f32.s32 %f173, %r221;
mul.ftz.f32 %f174, %f119, %f173;
mul.ftz.f32 %f175, %f120, 0f40800000;
sub.ftz.f32 %f176, %f174, %f175;
fma.rn.ftz.f32 %f409, %f118, %f176, %f409;
cvt.u64.u32 %rd59, %r677;
add.s64 %rd60, %rd1, %rd59;
mul.lo.s64 %rd61, %rd60, 36;
add.s64 %rd62, %rd9, %rd61;
add.s64 %rd63, %rd62, 4;
add.s64 %rd64, %rd63, %rd31;
ld.global.nc.u32 %r229, [%rd64];
ld.global.nc.u32 %r233, [%rd64+16];
add.s64 %rd65, %rd63, %rd33;
ld.global.nc.u32 %r237, [%rd65];
ld.global.nc.u32 %r241, [%rd64+20];

{ cvt.f32.f16 %f121, %rs1;}



dp4a.s32.s32 %r227, %r336, %r229, %r356;


dp4a.s32.s32 %r231, %r340, %r233, %r227;


dp4a.s32.s32 %r235, %r344, %r237, %r231;


dp4a.s32.s32 %r239, %r348, %r241, %r235;

ld.global.nc.u32 %r243, [%rd62];

{.reg .f16 low,high;
mov.b32 {low,high},%r243;
cvt.f32.f16 %f122, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r243;
cvt.f32.f16 %f123, high;}


cvt.rn.f32.s32 %f177, %r239;
mul.ftz.f32 %f178, %f122, %f177;
mul.ftz.f32 %f179, %f123, 0f40800000;
sub.ftz.f32 %f180, %f178, %f179;
fma.rn.ftz.f32 %f408, %f121, %f180, %f408;

{ cvt.f32.f16 %f124, %rs2;}



dp4a.s32.s32 %r245, %r354, %r229, %r356;


dp4a.s32.s32 %r249, %r358, %r233, %r245;


dp4a.s32.s32 %r253, %r362, %r237, %r249;


dp4a.s32.s32 %r257, %r366, %r241, %r253;


{.reg .f16 low,high;
mov.b32 {low,high},%r243;
cvt.f32.f16 %f125, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r243;
cvt.f32.f16 %f126, high;}


cvt.rn.f32.s32 %f181, %r257;
mul.ftz.f32 %f182, %f125, %f181;
mul.ftz.f32 %f183, %f126, 0f40800000;
sub.ftz.f32 %f184, %f182, %f183;
fma.rn.ftz.f32 %f407, %f124, %f184, %f407;
cvt.u64.u32 %rd66, %r676;
add.s64 %rd67, %rd1, %rd66;
mul.lo.s64 %rd68, %rd67, 36;
add.s64 %rd69, %rd9, %rd68;
add.s64 %rd70, %rd69, 4;
add.s64 %rd71, %rd70, %rd31;
ld.global.nc.u32 %r265, [%rd71];
ld.global.nc.u32 %r269, [%rd71+16];
add.s64 %rd72, %rd70, %rd33;
ld.global.nc.u32 %r273, [%rd72];
ld.global.nc.u32 %r277, [%rd71+20];

{ cvt.f32.f16 %f127, %rs1;}



dp4a.s32.s32 %r263, %r336, %r265, %r356;


dp4a.s32.s32 %r267, %r340, %r269, %r263;


dp4a.s32.s32 %r271, %r344, %r273, %r267;


dp4a.s32.s32 %r275, %r348, %r277, %r271;

ld.global.nc.u32 %r279, [%rd69];

{.reg .f16 low,high;
mov.b32 {low,high},%r279;
cvt.f32.f16 %f128, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r279;
cvt.f32.f16 %f129, high;}


cvt.rn.f32.s32 %f185, %r275;
mul.ftz.f32 %f186, %f128, %f185;
mul.ftz.f32 %f187, %f129, 0f40800000;
sub.ftz.f32 %f188, %f186, %f187;
fma.rn.ftz.f32 %f406, %f127, %f188, %f406;

{ cvt.f32.f16 %f130, %rs2;}



dp4a.s32.s32 %r281, %r354, %r265, %r356;


dp4a.s32.s32 %r285, %r358, %r269, %r281;


dp4a.s32.s32 %r289, %r362, %r273, %r285;


dp4a.s32.s32 %r293, %r366, %r277, %r289;


{.reg .f16 low,high;
mov.b32 {low,high},%r279;
cvt.f32.f16 %f131, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r279;
cvt.f32.f16 %f132, high;}


cvt.rn.f32.s32 %f189, %r293;
mul.ftz.f32 %f190, %f131, %f189;
mul.ftz.f32 %f191, %f132, 0f40800000;
sub.ftz.f32 %f192, %f190, %f191;
fma.rn.ftz.f32 %f405, %f130, %f192, %f405;
cvt.u64.u32 %rd73, %r675;
add.s64 %rd74, %rd1, %rd73;
mul.lo.s64 %rd75, %rd74, 36;
add.s64 %rd76, %rd9, %rd75;
add.s64 %rd77, %rd76, 4;
add.s64 %rd78, %rd77, %rd31;
ld.global.nc.u32 %r301, [%rd78];
ld.global.nc.u32 %r305, [%rd78+16];
add.s64 %rd79, %rd77, %rd33;
ld.global.nc.u32 %r309, [%rd79];
ld.global.nc.u32 %r313, [%rd78+20];

{ cvt.f32.f16 %f133, %rs1;}



dp4a.s32.s32 %r299, %r336, %r301, %r356;


dp4a.s32.s32 %r303, %r340, %r305, %r299;


dp4a.s32.s32 %r307, %r344, %r309, %r303;


dp4a.s32.s32 %r311, %r348, %r313, %r307;

ld.global.nc.u32 %r315, [%rd76];

{.reg .f16 low,high;
mov.b32 {low,high},%r315;
cvt.f32.f16 %f134, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r315;
cvt.f32.f16 %f135, high;}


cvt.rn.f32.s32 %f193, %r311;
mul.ftz.f32 %f194, %f134, %f193;
mul.ftz.f32 %f195, %f135, 0f40800000;
sub.ftz.f32 %f196, %f194, %f195;
fma.rn.ftz.f32 %f404, %f133, %f196, %f404;

{ cvt.f32.f16 %f136, %rs2;}



dp4a.s32.s32 %r317, %r354, %r301, %r356;


dp4a.s32.s32 %r321, %r358, %r305, %r317;


dp4a.s32.s32 %r325, %r362, %r309, %r321;


dp4a.s32.s32 %r329, %r366, %r313, %r325;


{.reg .f16 low,high;
mov.b32 {low,high},%r315;
cvt.f32.f16 %f137, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r315;
cvt.f32.f16 %f138, high;}


cvt.rn.f32.s32 %f197, %r329;
mul.ftz.f32 %f198, %f137, %f197;
mul.ftz.f32 %f199, %f138, 0f40800000;
sub.ftz.f32 %f200, %f198, %f199;
fma.rn.ftz.f32 %f403, %f136, %f200, %f403;
cvt.u64.u32 %rd80, %r674;
add.s64 %rd81, %rd1, %rd80;
mul.lo.s64 %rd82, %rd81, 36;
add.s64 %rd83, %rd9, %rd82;
add.s64 %rd84, %rd83, 4;
add.s64 %rd85, %rd84, %rd31;
ld.global.nc.u32 %r337, [%rd85];
ld.global.nc.u32 %r341, [%rd85+16];
add.s64 %rd86, %rd84, %rd33;
ld.global.nc.u32 %r345, [%rd86];
ld.global.nc.u32 %r349, [%rd85+20];

{ cvt.f32.f16 %f139, %rs1;}



dp4a.s32.s32 %r335, %r336, %r337, %r356;


dp4a.s32.s32 %r339, %r340, %r341, %r335;


dp4a.s32.s32 %r343, %r344, %r345, %r339;


dp4a.s32.s32 %r347, %r348, %r349, %r343;

ld.global.nc.u32 %r351, [%rd83];

{.reg .f16 low,high;
mov.b32 {low,high},%r351;
cvt.f32.f16 %f140, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r351;
cvt.f32.f16 %f141, high;}


cvt.rn.f32.s32 %f201, %r347;
mul.ftz.f32 %f202, %f140, %f201;
mul.ftz.f32 %f203, %f141, 0f40800000;
sub.ftz.f32 %f204, %f202, %f203;
fma.rn.ftz.f32 %f402, %f139, %f204, %f402;

{ cvt.f32.f16 %f142, %rs2;}



dp4a.s32.s32 %r353, %r354, %r337, %r356;


dp4a.s32.s32 %r357, %r358, %r341, %r353;


dp4a.s32.s32 %r361, %r362, %r345, %r357;


dp4a.s32.s32 %r365, %r366, %r349, %r361;


{.reg .f16 low,high;
mov.b32 {low,high},%r351;
cvt.f32.f16 %f143, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r351;
cvt.f32.f16 %f144, high;}


cvt.rn.f32.s32 %f205, %r365;
mul.ftz.f32 %f206, %f143, %f205;
mul.ftz.f32 %f207, %f144, 0f40800000;
sub.ftz.f32 %f208, %f206, %f207;
fma.rn.ftz.f32 %f401, %f142, %f208, %f401;
add.s32 %r679, %r679, 32;
add.s32 %r678, %r678, 32;
add.s32 %r677, %r677, 32;
add.s32 %r676, %r676, 32;
add.s32 %r675, %r675, 32;
add.s32 %r674, %r674, 32;
add.s32 %r673, %r673, 32;
add.s64 %rd141, %rd141, 576;
add.s32 %r680, %r680, 32;
setp.lt.s32 %p4, %r680, %r6;
@%p4 bra $L__BB40_2;

$L__BB40_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB40_5;

shl.b32 %r379, %r1, 11;
mov.u32 %r380, _ZZ13mul_mat_vec_qIL9ggml_type2ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r381, %r380, %r379;
shl.b32 %r382, %r2, 2;
add.s32 %r383, %r381, %r382;
st.shared.f32 [%r383+-2048], %f416;
st.shared.f32 [%r383+-1920], %f415;
st.shared.f32 [%r383+-1792], %f414;
st.shared.f32 [%r383+-1664], %f413;
st.shared.f32 [%r383+-1536], %f412;
st.shared.f32 [%r383+-1408], %f411;
st.shared.f32 [%r383+-1280], %f410;
st.shared.f32 [%r383+-1152], %f409;
st.shared.f32 [%r383+-1024], %f408;
st.shared.f32 [%r383+-896], %f407;
st.shared.f32 [%r383+-768], %f406;
st.shared.f32 [%r383+-640], %f405;
st.shared.f32 [%r383+-512], %f404;
st.shared.f32 [%r383+-384], %f403;
st.shared.f32 [%r383+-256], %f402;
st.shared.f32 [%r383+-128], %f401;

$L__BB40_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB40_38;

mov.u32 %r384, %ctaid.z;
mov.u32 %r385, %ctaid.y;
mad.lo.s32 %r386, %r385, %r42, %r4;
mad.lo.s32 %r387, %r384, %r48, %r386;
cvt.u64.u32 %rd12, %r387;
shl.b32 %r388, %r2, 2;
mov.u32 %r389, _ZZ13mul_mat_vec_qIL9ggml_type2ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r30, %r389, %r388;
mov.u32 %r390, 2;
ld.shared.f32 %f209, [%r30];
add.ftz.f32 %f210, %f209, %f416;
mov.b32 %r391, %f210;
mov.u32 %r392, 31;
mov.u32 %r393, 16;
mov.u32 %r394, -1;
shfl.sync.bfly.b32 %r395|%p7, %r391, %r393, %r392, %r394;
mov.b32 %f211, %r395;
add.ftz.f32 %f212, %f210, %f211;
mov.b32 %r396, %f212;
mov.u32 %r397, 8;
shfl.sync.bfly.b32 %r398|%p8, %r396, %r397, %r392, %r394;
mov.b32 %f213, %r398;
add.ftz.f32 %f214, %f212, %f213;
mov.b32 %r399, %f214;
mov.u32 %r400, 4;
shfl.sync.bfly.b32 %r401|%p9, %r399, %r400, %r392, %r394;
mov.b32 %f215, %r401;
add.ftz.f32 %f216, %f214, %f215;
mov.b32 %r402, %f216;
shfl.sync.bfly.b32 %r403|%p10, %r402, %r390, %r392, %r394;
mov.b32 %f217, %r403;
add.ftz.f32 %f218, %f216, %f217;
mov.b32 %r404, %f218;
mov.u32 %r405, 1;
shfl.sync.bfly.b32 %r406|%p11, %r404, %r405, %r392, %r394;
mov.b32 %f219, %r406;
add.ftz.f32 %f49, %f218, %f219;
setp.lt.u32 %p12, %r4, %r36;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
cvta.to.global.u64 %rd87, %rd19;
mul.wide.u32 %rd88, %r387, 4;
add.s64 %rd13, %rd87, %rd88;
not.pred %p14, %p1;
@%p14 bra $L__BB40_8;

st.global.f32 [%rd13], %f49;

$L__BB40_8:
ld.shared.f32 %f220, [%r30+128];
add.ftz.f32 %f221, %f220, %f415;
mov.b32 %r407, %f221;
shfl.sync.bfly.b32 %r411|%p15, %r407, %r393, %r392, %r394;
mov.b32 %f222, %r411;
add.ftz.f32 %f223, %f221, %f222;
mov.b32 %r412, %f223;
shfl.sync.bfly.b32 %r414|%p16, %r412, %r397, %r392, %r394;
mov.b32 %f224, %r414;
add.ftz.f32 %f225, %f223, %f224;
mov.b32 %r415, %f225;
shfl.sync.bfly.b32 %r417|%p17, %r415, %r400, %r392, %r394;
mov.b32 %f226, %r417;
add.ftz.f32 %f227, %f225, %f226;
mov.b32 %r418, %f227;
shfl.sync.bfly.b32 %r420|%p18, %r418, %r390, %r392, %r394;
mov.b32 %f228, %r420;
add.ftz.f32 %f229, %f227, %f228;
mov.b32 %r421, %f229;
shfl.sync.bfly.b32 %r423|%p19, %r421, %r405, %r392, %r394;
mov.b32 %f230, %r423;
add.ftz.f32 %f50, %f229, %f230;
setp.eq.s32 %p20, %r2, 1;
or.b32 %r427, %r4, 1;
setp.lt.u32 %p21, %r427, %r36;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB40_10;

st.global.f32 [%rd13+4], %f50;

$L__BB40_10:
ld.shared.f32 %f231, [%r30+256];
add.ftz.f32 %f232, %f231, %f414;
mov.b32 %r428, %f232;
mov.u32 %r429, 31;
mov.u32 %r430, 16;
mov.u32 %r431, -1;
shfl.sync.bfly.b32 %r432|%p23, %r428, %r430, %r429, %r431;
mov.b32 %f233, %r432;
add.ftz.f32 %f234, %f232, %f233;
mov.b32 %r433, %f234;
mov.u32 %r434, 8;
shfl.sync.bfly.b32 %r435|%p24, %r433, %r434, %r429, %r431;
mov.b32 %f235, %r435;
add.ftz.f32 %f236, %f234, %f235;
mov.b32 %r436, %f236;
mov.u32 %r437, 4;
shfl.sync.bfly.b32 %r438|%p25, %r436, %r437, %r429, %r431;
mov.b32 %f237, %r438;
add.ftz.f32 %f238, %f236, %f237;
mov.b32 %r439, %f238;
mov.u32 %r440, 2;
shfl.sync.bfly.b32 %r441|%p26, %r439, %r440, %r429, %r431;
mov.b32 %f239, %r441;
add.ftz.f32 %f240, %f238, %f239;
mov.b32 %r442, %f240;
mov.u32 %r443, 1;
shfl.sync.bfly.b32 %r444|%p27, %r442, %r443, %r429, %r431;
mov.b32 %f241, %r444;
add.ftz.f32 %f51, %f240, %f241;
@%p14 bra $L__BB40_12;

cvt.u64.u32 %rd89, %r36;
add.s64 %rd90, %rd89, %rd12;
shl.b64 %rd92, %rd90, 2;
add.s64 %rd93, %rd87, %rd92;
st.global.f32 [%rd93], %f51;

$L__BB40_12:
ld.shared.f32 %f242, [%r30+384];
add.ftz.f32 %f243, %f242, %f413;
mov.b32 %r445, %f243;
shfl.sync.bfly.b32 %r449|%p29, %r445, %r430, %r429, %r431;
mov.b32 %f244, %r449;
add.ftz.f32 %f245, %f243, %f244;
mov.b32 %r450, %f245;
shfl.sync.bfly.b32 %r452|%p30, %r450, %r434, %r429, %r431;
mov.b32 %f246, %r452;
add.ftz.f32 %f247, %f245, %f246;
mov.b32 %r453, %f247;
shfl.sync.bfly.b32 %r455|%p31, %r453, %r437, %r429, %r431;
mov.b32 %f248, %r455;
add.ftz.f32 %f249, %f247, %f248;
mov.b32 %r456, %f249;
shfl.sync.bfly.b32 %r458|%p32, %r456, %r440, %r429, %r431;
mov.b32 %f250, %r458;
add.ftz.f32 %f251, %f249, %f250;
mov.b32 %r459, %f251;
shfl.sync.bfly.b32 %r461|%p33, %r459, %r443, %r429, %r431;
mov.b32 %f252, %r461;
add.ftz.f32 %f52, %f251, %f252;
@%p22 bra $L__BB40_14;

add.s32 %r462, %r36, 1;
cvt.u64.u32 %rd94, %r462;
add.s64 %rd95, %rd94, %rd12;
shl.b64 %rd97, %rd95, 2;
add.s64 %rd98, %rd87, %rd97;
st.global.f32 [%rd98], %f52;

$L__BB40_14:
shl.b32 %r463, %r36, 1;
mov.u32 %r464, 1;
ld.shared.f32 %f253, [%r30+512];
add.ftz.f32 %f254, %f253, %f412;
mov.b32 %r465, %f254;
mov.u32 %r466, 31;
mov.u32 %r467, 16;
mov.u32 %r468, -1;
shfl.sync.bfly.b32 %r469|%p35, %r465, %r467, %r466, %r468;
mov.b32 %f255, %r469;
add.ftz.f32 %f256, %f254, %f255;
mov.b32 %r470, %f256;
mov.u32 %r471, 8;
shfl.sync.bfly.b32 %r472|%p36, %r470, %r471, %r466, %r468;
mov.b32 %f257, %r472;
add.ftz.f32 %f258, %f256, %f257;
mov.b32 %r473, %f258;
mov.u32 %r474, 4;
shfl.sync.bfly.b32 %r475|%p37, %r473, %r474, %r466, %r468;
mov.b32 %f259, %r475;
add.ftz.f32 %f260, %f258, %f259;
mov.b32 %r476, %f260;
mov.u32 %r477, 2;
shfl.sync.bfly.b32 %r478|%p38, %r476, %r477, %r466, %r468;
mov.b32 %f261, %r478;
add.ftz.f32 %f262, %f260, %f261;
mov.b32 %r479, %f262;
shfl.sync.bfly.b32 %r480|%p39, %r479, %r464, %r466, %r468;
mov.b32 %f263, %r480;
add.ftz.f32 %f53, %f262, %f263;
cvt.u64.u32 %rd99, %r463;
add.s64 %rd100, %rd99, %rd12;
shl.b64 %rd102, %rd100, 2;
add.s64 %rd14, %rd87, %rd102;
@%p14 bra $L__BB40_16;

st.global.f32 [%rd14], %f53;

$L__BB40_16:
ld.shared.f32 %f264, [%r30+640];
add.ftz.f32 %f265, %f264, %f411;
mov.b32 %r481, %f265;
shfl.sync.bfly.b32 %r485|%p41, %r481, %r467, %r466, %r468;
mov.b32 %f266, %r485;
add.ftz.f32 %f267, %f265, %f266;
mov.b32 %r486, %f267;
shfl.sync.bfly.b32 %r488|%p42, %r486, %r471, %r466, %r468;
mov.b32 %f268, %r488;
add.ftz.f32 %f269, %f267, %f268;
mov.b32 %r489, %f269;
shfl.sync.bfly.b32 %r491|%p43, %r489, %r474, %r466, %r468;
mov.b32 %f270, %r491;
add.ftz.f32 %f271, %f269, %f270;
mov.b32 %r492, %f271;
shfl.sync.bfly.b32 %r494|%p44, %r492, %r477, %r466, %r468;
mov.b32 %f272, %r494;
add.ftz.f32 %f273, %f271, %f272;
mov.b32 %r495, %f273;
shfl.sync.bfly.b32 %r497|%p45, %r495, %r464, %r466, %r468;
mov.b32 %f274, %r497;
add.ftz.f32 %f54, %f273, %f274;
@%p22 bra $L__BB40_18;

st.global.f32 [%rd14+4], %f54;

$L__BB40_18:
mul.lo.s32 %r31, %r36, 3;
ld.shared.f32 %f275, [%r30+768];
add.ftz.f32 %f276, %f275, %f410;
mov.b32 %r498, %f276;
mov.u32 %r499, 31;
mov.u32 %r500, 16;
mov.u32 %r501, -1;
shfl.sync.bfly.b32 %r502|%p47, %r498, %r500, %r499, %r501;
mov.b32 %f277, %r502;
add.ftz.f32 %f278, %f276, %f277;
mov.b32 %r503, %f278;
mov.u32 %r504, 8;
shfl.sync.bfly.b32 %r505|%p48, %r503, %r504, %r499, %r501;
mov.b32 %f279, %r505;
add.ftz.f32 %f280, %f278, %f279;
mov.b32 %r506, %f280;
mov.u32 %r507, 4;
shfl.sync.bfly.b32 %r508|%p49, %r506, %r507, %r499, %r501;
mov.b32 %f281, %r508;
add.ftz.f32 %f282, %f280, %f281;
mov.b32 %r509, %f282;
mov.u32 %r510, 2;
shfl.sync.bfly.b32 %r511|%p50, %r509, %r510, %r499, %r501;
mov.b32 %f283, %r511;
add.ftz.f32 %f284, %f282, %f283;
mov.b32 %r512, %f284;
mov.u32 %r513, 1;
shfl.sync.bfly.b32 %r514|%p51, %r512, %r513, %r499, %r501;
mov.b32 %f285, %r514;
add.ftz.f32 %f55, %f284, %f285;
@%p14 bra $L__BB40_20;

cvt.u64.u32 %rd103, %r31;
add.s64 %rd104, %rd103, %rd12;
shl.b64 %rd106, %rd104, 2;
add.s64 %rd107, %rd87, %rd106;
st.global.f32 [%rd107], %f55;

$L__BB40_20:
ld.shared.f32 %f286, [%r30+896];
add.ftz.f32 %f287, %f286, %f409;
mov.b32 %r515, %f287;
shfl.sync.bfly.b32 %r519|%p53, %r515, %r500, %r499, %r501;
mov.b32 %f288, %r519;
add.ftz.f32 %f289, %f287, %f288;
mov.b32 %r520, %f289;
shfl.sync.bfly.b32 %r522|%p54, %r520, %r504, %r499, %r501;
mov.b32 %f290, %r522;
add.ftz.f32 %f291, %f289, %f290;
mov.b32 %r523, %f291;
shfl.sync.bfly.b32 %r525|%p55, %r523, %r507, %r499, %r501;
mov.b32 %f292, %r525;
add.ftz.f32 %f293, %f291, %f292;
mov.b32 %r526, %f293;
shfl.sync.bfly.b32 %r528|%p56, %r526, %r510, %r499, %r501;
mov.b32 %f294, %r528;
add.ftz.f32 %f295, %f293, %f294;
mov.b32 %r529, %f295;
shfl.sync.bfly.b32 %r531|%p57, %r529, %r513, %r499, %r501;
mov.b32 %f296, %r531;
add.ftz.f32 %f56, %f295, %f296;
@%p22 bra $L__BB40_22;

add.s32 %r532, %r31, 1;
cvt.u64.u32 %rd108, %r532;
add.s64 %rd109, %rd108, %rd12;
shl.b64 %rd111, %rd109, 2;
add.s64 %rd112, %rd87, %rd111;
st.global.f32 [%rd112], %f56;

$L__BB40_22:
shl.b32 %r533, %r36, 2;
mov.u32 %r534, 2;
ld.shared.f32 %f297, [%r30+1024];
add.ftz.f32 %f298, %f297, %f408;
mov.b32 %r535, %f298;
mov.u32 %r536, 31;
mov.u32 %r537, 16;
mov.u32 %r538, -1;
shfl.sync.bfly.b32 %r539|%p59, %r535, %r537, %r536, %r538;
mov.b32 %f299, %r539;
add.ftz.f32 %f300, %f298, %f299;
mov.b32 %r540, %f300;
mov.u32 %r541, 8;
shfl.sync.bfly.b32 %r542|%p60, %r540, %r541, %r536, %r538;
mov.b32 %f301, %r542;
add.ftz.f32 %f302, %f300, %f301;
mov.b32 %r543, %f302;
mov.u32 %r544, 4;
shfl.sync.bfly.b32 %r545|%p61, %r543, %r544, %r536, %r538;
mov.b32 %f303, %r545;
add.ftz.f32 %f304, %f302, %f303;
mov.b32 %r546, %f304;
shfl.sync.bfly.b32 %r547|%p62, %r546, %r534, %r536, %r538;
mov.b32 %f305, %r547;
add.ftz.f32 %f306, %f304, %f305;
mov.b32 %r548, %f306;
mov.u32 %r549, 1;
shfl.sync.bfly.b32 %r550|%p63, %r548, %r549, %r536, %r538;
mov.b32 %f307, %r550;
add.ftz.f32 %f57, %f306, %f307;
cvt.u64.u32 %rd113, %r533;
add.s64 %rd114, %rd113, %rd12;
shl.b64 %rd116, %rd114, 2;
add.s64 %rd15, %rd87, %rd116;
@%p14 bra $L__BB40_24;

st.global.f32 [%rd15], %f57;

$L__BB40_24:
ld.shared.f32 %f308, [%r30+1152];
add.ftz.f32 %f309, %f308, %f407;
mov.b32 %r551, %f309;
shfl.sync.bfly.b32 %r555|%p65, %r551, %r537, %r536, %r538;
mov.b32 %f310, %r555;
add.ftz.f32 %f311, %f309, %f310;
mov.b32 %r556, %f311;
shfl.sync.bfly.b32 %r558|%p66, %r556, %r541, %r536, %r538;
mov.b32 %f312, %r558;
add.ftz.f32 %f313, %f311, %f312;
mov.b32 %r559, %f313;
shfl.sync.bfly.b32 %r561|%p67, %r559, %r544, %r536, %r538;
mov.b32 %f314, %r561;
add.ftz.f32 %f315, %f313, %f314;
mov.b32 %r562, %f315;
shfl.sync.bfly.b32 %r564|%p68, %r562, %r534, %r536, %r538;
mov.b32 %f316, %r564;
add.ftz.f32 %f317, %f315, %f316;
mov.b32 %r565, %f317;
shfl.sync.bfly.b32 %r567|%p69, %r565, %r549, %r536, %r538;
mov.b32 %f318, %r567;
add.ftz.f32 %f58, %f317, %f318;
@%p22 bra $L__BB40_26;

st.global.f32 [%rd15+4], %f58;

$L__BB40_26:
mul.lo.s32 %r32, %r36, 5;
ld.shared.f32 %f319, [%r30+1280];
add.ftz.f32 %f320, %f319, %f406;
mov.b32 %r568, %f320;
mov.u32 %r569, 31;
mov.u32 %r570, 16;
mov.u32 %r571, -1;
shfl.sync.bfly.b32 %r572|%p71, %r568, %r570, %r569, %r571;
mov.b32 %f321, %r572;
add.ftz.f32 %f322, %f320, %f321;
mov.b32 %r573, %f322;
mov.u32 %r574, 8;
shfl.sync.bfly.b32 %r575|%p72, %r573, %r574, %r569, %r571;
mov.b32 %f323, %r575;
add.ftz.f32 %f324, %f322, %f323;
mov.b32 %r576, %f324;
mov.u32 %r577, 4;
shfl.sync.bfly.b32 %r578|%p73, %r576, %r577, %r569, %r571;
mov.b32 %f325, %r578;
add.ftz.f32 %f326, %f324, %f325;
mov.b32 %r579, %f326;
mov.u32 %r580, 2;
shfl.sync.bfly.b32 %r581|%p74, %r579, %r580, %r569, %r571;
mov.b32 %f327, %r581;
add.ftz.f32 %f328, %f326, %f327;
mov.b32 %r582, %f328;
mov.u32 %r583, 1;
shfl.sync.bfly.b32 %r584|%p75, %r582, %r583, %r569, %r571;
mov.b32 %f329, %r584;
add.ftz.f32 %f59, %f328, %f329;
@%p14 bra $L__BB40_28;

cvt.u64.u32 %rd117, %r32;
add.s64 %rd118, %rd117, %rd12;
shl.b64 %rd120, %rd118, 2;
add.s64 %rd121, %rd87, %rd120;
st.global.f32 [%rd121], %f59;

$L__BB40_28:
ld.shared.f32 %f330, [%r30+1408];
add.ftz.f32 %f331, %f330, %f405;
mov.b32 %r585, %f331;
shfl.sync.bfly.b32 %r589|%p77, %r585, %r570, %r569, %r571;
mov.b32 %f332, %r589;
add.ftz.f32 %f333, %f331, %f332;
mov.b32 %r590, %f333;
shfl.sync.bfly.b32 %r592|%p78, %r590, %r574, %r569, %r571;
mov.b32 %f334, %r592;
add.ftz.f32 %f335, %f333, %f334;
mov.b32 %r593, %f335;
shfl.sync.bfly.b32 %r595|%p79, %r593, %r577, %r569, %r571;
mov.b32 %f336, %r595;
add.ftz.f32 %f337, %f335, %f336;
mov.b32 %r596, %f337;
shfl.sync.bfly.b32 %r598|%p80, %r596, %r580, %r569, %r571;
mov.b32 %f338, %r598;
add.ftz.f32 %f339, %f337, %f338;
mov.b32 %r599, %f339;
shfl.sync.bfly.b32 %r601|%p81, %r599, %r583, %r569, %r571;
mov.b32 %f340, %r601;
add.ftz.f32 %f60, %f339, %f340;
@%p22 bra $L__BB40_30;

add.s32 %r602, %r32, 1;
cvt.u64.u32 %rd122, %r602;
add.s64 %rd123, %rd122, %rd12;
shl.b64 %rd125, %rd123, 2;
add.s64 %rd126, %rd87, %rd125;
st.global.f32 [%rd126], %f60;

$L__BB40_30:
mul.lo.s32 %r603, %r36, 6;
ld.shared.f32 %f341, [%r30+1536];
add.ftz.f32 %f342, %f341, %f404;
mov.b32 %r604, %f342;
mov.u32 %r605, 31;
mov.u32 %r606, 16;
mov.u32 %r607, -1;
shfl.sync.bfly.b32 %r608|%p83, %r604, %r606, %r605, %r607;
mov.b32 %f343, %r608;
add.ftz.f32 %f344, %f342, %f343;
mov.b32 %r609, %f344;
mov.u32 %r610, 8;
shfl.sync.bfly.b32 %r611|%p84, %r609, %r610, %r605, %r607;
mov.b32 %f345, %r611;
add.ftz.f32 %f346, %f344, %f345;
mov.b32 %r612, %f346;
mov.u32 %r613, 4;
shfl.sync.bfly.b32 %r614|%p85, %r612, %r613, %r605, %r607;
mov.b32 %f347, %r614;
add.ftz.f32 %f348, %f346, %f347;
mov.b32 %r615, %f348;
mov.u32 %r616, 2;
shfl.sync.bfly.b32 %r617|%p86, %r615, %r616, %r605, %r607;
mov.b32 %f349, %r617;
add.ftz.f32 %f350, %f348, %f349;
mov.b32 %r618, %f350;
mov.u32 %r619, 1;
shfl.sync.bfly.b32 %r620|%p87, %r618, %r619, %r605, %r607;
mov.b32 %f351, %r620;
add.ftz.f32 %f61, %f350, %f351;
cvt.u64.u32 %rd127, %r603;
add.s64 %rd128, %rd127, %rd12;
shl.b64 %rd130, %rd128, 2;
add.s64 %rd16, %rd87, %rd130;
@%p14 bra $L__BB40_32;

st.global.f32 [%rd16], %f61;

$L__BB40_32:
ld.shared.f32 %f352, [%r30+1664];
add.ftz.f32 %f353, %f352, %f403;
mov.b32 %r621, %f353;
shfl.sync.bfly.b32 %r625|%p89, %r621, %r606, %r605, %r607;
mov.b32 %f354, %r625;
add.ftz.f32 %f355, %f353, %f354;
mov.b32 %r626, %f355;
shfl.sync.bfly.b32 %r628|%p90, %r626, %r610, %r605, %r607;
mov.b32 %f356, %r628;
add.ftz.f32 %f357, %f355, %f356;
mov.b32 %r629, %f357;
shfl.sync.bfly.b32 %r631|%p91, %r629, %r613, %r605, %r607;
mov.b32 %f358, %r631;
add.ftz.f32 %f359, %f357, %f358;
mov.b32 %r632, %f359;
shfl.sync.bfly.b32 %r634|%p92, %r632, %r616, %r605, %r607;
mov.b32 %f360, %r634;
add.ftz.f32 %f361, %f359, %f360;
mov.b32 %r635, %f361;
shfl.sync.bfly.b32 %r637|%p93, %r635, %r619, %r605, %r607;
mov.b32 %f362, %r637;
add.ftz.f32 %f62, %f361, %f362;
@%p22 bra $L__BB40_34;

st.global.f32 [%rd16+4], %f62;

$L__BB40_34:
mul.lo.s32 %r33, %r36, 7;
ld.shared.f32 %f363, [%r30+1792];
add.ftz.f32 %f364, %f363, %f402;
mov.b32 %r638, %f364;
mov.u32 %r639, 31;
mov.u32 %r640, 16;
mov.u32 %r641, -1;
shfl.sync.bfly.b32 %r642|%p95, %r638, %r640, %r639, %r641;
mov.b32 %f365, %r642;
add.ftz.f32 %f366, %f364, %f365;
mov.b32 %r643, %f366;
mov.u32 %r644, 8;
shfl.sync.bfly.b32 %r645|%p96, %r643, %r644, %r639, %r641;
mov.b32 %f367, %r645;
add.ftz.f32 %f368, %f366, %f367;
mov.b32 %r646, %f368;
mov.u32 %r647, 4;
shfl.sync.bfly.b32 %r648|%p97, %r646, %r647, %r639, %r641;
mov.b32 %f369, %r648;
add.ftz.f32 %f370, %f368, %f369;
mov.b32 %r649, %f370;
mov.u32 %r650, 2;
shfl.sync.bfly.b32 %r651|%p98, %r649, %r650, %r639, %r641;
mov.b32 %f371, %r651;
add.ftz.f32 %f372, %f370, %f371;
mov.b32 %r652, %f372;
mov.u32 %r653, 1;
shfl.sync.bfly.b32 %r654|%p99, %r652, %r653, %r639, %r641;
mov.b32 %f373, %r654;
add.ftz.f32 %f63, %f372, %f373;
@%p14 bra $L__BB40_36;

cvt.u64.u32 %rd131, %r33;
add.s64 %rd132, %rd131, %rd12;
shl.b64 %rd134, %rd132, 2;
add.s64 %rd135, %rd87, %rd134;
st.global.f32 [%rd135], %f63;

$L__BB40_36:
ld.shared.f32 %f374, [%r30+1920];
add.ftz.f32 %f375, %f374, %f401;
mov.b32 %r655, %f375;
shfl.sync.bfly.b32 %r659|%p101, %r655, %r640, %r639, %r641;
mov.b32 %f376, %r659;
add.ftz.f32 %f377, %f375, %f376;
mov.b32 %r660, %f377;
shfl.sync.bfly.b32 %r662|%p102, %r660, %r644, %r639, %r641;
mov.b32 %f378, %r662;
add.ftz.f32 %f379, %f377, %f378;
mov.b32 %r663, %f379;
shfl.sync.bfly.b32 %r665|%p103, %r663, %r647, %r639, %r641;
mov.b32 %f380, %r665;
add.ftz.f32 %f381, %f379, %f380;
mov.b32 %r666, %f381;
shfl.sync.bfly.b32 %r668|%p104, %r666, %r650, %r639, %r641;
mov.b32 %f382, %r668;
add.ftz.f32 %f383, %f381, %f382;
mov.b32 %r669, %f383;
shfl.sync.bfly.b32 %r671|%p105, %r669, %r653, %r639, %r641;
mov.b32 %f384, %r671;
add.ftz.f32 %f64, %f383, %f384;
@%p22 bra $L__BB40_38;

add.s32 %r672, %r33, 1;
cvt.u64.u32 %rd136, %r672;
add.s64 %rd137, %rd136, %rd12;
shl.b64 %rd139, %rd137, 2;
add.s64 %rd140, %rd87, %rd139;
st.global.f32 [%rd140], %f64;

$L__BB40_38:
ret;

}
.entry _Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj(
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2,
.param .align 8 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3[48],
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5,
.param .align 4 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6[12],
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15
)
.maxntid 256, 1, 1
.minnctapersm 1
{
.local .align 8 .b8 __local_depot41[16];
.reg .b64 %SP;
.reg .b64 %SPL;
.reg .pred %p<48>;
.reg .f32 %f<177>;
.reg .b32 %r<234>;
.reg .b64 %rd<80>;


mov.u64 %SPL, __local_depot41;
ld.param.u64 %rd24, [_Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0];
ld.param.u64 %rd25, [_Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1];
ld.param.u64 %rd26, [_Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2];
ld.param.u64 %rd32, [_Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4];
ld.param.u32 %r29, [_Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5];
ld.param.u32 %r22, [_Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8];
ld.param.u32 %r23, [_Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9];
ld.param.u32 %r25, [_Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11];
ld.param.u32 %r26, [_Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12];
ld.param.u32 %r27, [_Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13];
ld.param.u32 %r30, [_Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14];
ld.param.u32 %r28, [_Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15];
ld.param.u32 %r20, [_Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+8];
ld.param.u32 %r19, [_Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+4];
ld.param.u32 %r18, [_Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6];
ld.param.u32 %r17, [_Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
ld.param.u64 %rd29, [_Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+16];
ld.param.u64 %rd28, [_Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+8];
ld.param.u64 %rd27, [_Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3];
add.u64 %rd1, %SPL, 0;
add.u64 %rd2, %SPL, 8;
shr.u32 %r1, %r29, 5;
mov.u32 %r2, %tid.y;
setp.ge.u32 %p1, %r2, %r30;
@%p1 bra $L__BB41_29;

mov.u32 %r3, %ctaid.y;
cvta.to.global.u64 %rd35, %rd26;
mad.lo.s32 %r31, %r2, %r28, %r3;
mul.wide.u32 %rd36, %r31, 4;
add.s64 %rd37, %rd35, %rd36;
ld.global.nc.u32 %r4, [%rd37];
mov.f32 %f170, 0f00000000;
st.local.v2.f32 [%rd1], {%f170, %f170};
st.local.v2.f32 [%rd2], {%f170, %f170};
mov.u32 %r5, %tid.x;
shr.u32 %r233, %r5, 1;
setp.ge.u32 %p2, %r233, %r1;
mov.f32 %f171, %f170;
mov.f32 %f172, %f170;
mov.f32 %f173, %f170;
@%p2 bra $L__BB41_9;

mov.u32 %r32, %ctaid.x;
shl.b32 %r33, %r32, 1;
cvta.to.global.u64 %rd38, %rd25;
mul.hi.u32 %r34, %r3, %r18;
add.s32 %r35, %r34, %r3;
shr.u32 %r36, %r35, %r19;
mul.lo.s32 %r37, %r36, %r20;
sub.s32 %r38, %r3, %r37;
mul.lo.s32 %r39, %r38, %r26;
cvt.u64.u32 %rd39, %r39;
mul.lo.s32 %r40, %r2, %r23;
cvt.u64.u32 %rd40, %r40;
mul.lo.s32 %r41, %r4, %r25;
cvt.u64.u32 %rd41, %r5;
shr.u64 %rd42, %rd41, 1;
mul.lo.s64 %rd43, %rd42, 36;
mul.wide.u32 %rd44, %r40, 36;
add.s64 %rd45, %rd43, %rd44;
add.s64 %rd4, %rd38, %rd45;
mul.wide.u32 %rd79, %r39, 36;
and.b32 %r42, %r5, 1;
mul.wide.u32 %rd46, %r42, 8;
mad.lo.s32 %r43, %r33, %r22, %r233;
add.s32 %r44, %r43, %r41;
mul.wide.s32 %rd6, %r44, 20;
add.s64 %rd47, %rd46, %rd6;
add.s64 %rd7, %rd47, 8;
add.s64 %rd48, %rd42, %rd40;
add.s64 %rd49, %rd48, %rd39;
mul.lo.s64 %rd50, %rd49, 36;
add.s64 %rd51, %rd50, %rd46;
add.s64 %rd52, %rd38, %rd51;
add.s64 %rd77, %rd52, 20;
or.b32 %r45, %r33, 1;
add.s32 %r46, %r233, %r41;
mad.lo.s32 %r47, %r22, %r45, %r46;
mul.wide.s32 %rd9, %r47, 20;
add.s64 %rd53, %rd46, %rd9;
add.s64 %rd10, %rd53, 4;
cvta.to.global.u64 %rd78, %rd24;
cvta.to.global.u64 %rd76, %rd28;
mov.f32 %f170, 0f00000000;

$L__BB41_3:
add.s64 %rd54, %rd4, %rd79;
ld.global.nc.u32 %r8, [%rd77+-16];
ld.global.nc.u32 %r9, [%rd77];
add.s64 %rd55, %rd78, %rd7;
ld.global.nc.u32 %r69, [%rd55];
ld.global.nc.u32 %r10, [%rd77+-12];
ld.global.nc.u32 %r11, [%rd77+4];
ld.global.nc.u32 %r70, [%rd55+-4];
and.b32 %r49, %r70, 252645135;
shr.u32 %r71, %r70, 4;
and.b32 %r53, %r71, 252645135;
mov.u32 %r51, 0;

dp4a.s32.s32 %r48, %r49, %r8, %r51;


dp4a.s32.s32 %r52, %r53, %r9, %r48;

and.b32 %r57, %r69, 252645135;
shr.u32 %r72, %r69, 4;
and.b32 %r61, %r72, 252645135;

dp4a.s32.s32 %r56, %r57, %r10, %r52;


dp4a.s32.s32 %r60, %r61, %r11, %r56;

add.s64 %rd56, %rd78, %rd6;
ld.global.nc.u32 %r65, [%rd56];
ld.global.nc.u32 %r12, [%rd54];

{mul.f16x2 %r64,%r65,%r12;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r64;
cvt.f32.f16 %f42, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r64;
cvt.f32.f16 %f43, high;}


cvt.rn.f32.s32 %f44, %r60;
mov.f32 %f45, 0f40000000;
div.approx.ftz.f32 %f46, %f43, %f45;
fma.rn.ftz.f32 %f47, %f42, %f44, %f46;
add.ftz.f32 %f173, %f47, %f173;
setp.eq.s64 %p3, %rd28, 0;
@%p3 bra $L__BB41_5;

add.s64 %rd57, %rd76, %rd7;
ld.global.u32 %r94, [%rd57];
ld.global.u32 %r95, [%rd57+-4];
and.b32 %r74, %r95, 252645135;
shr.u32 %r96, %r95, 4;
and.b32 %r78, %r96, 252645135;

dp4a.s32.s32 %r73, %r74, %r8, %r51;


dp4a.s32.s32 %r77, %r78, %r9, %r73;

and.b32 %r82, %r94, 252645135;
shr.u32 %r97, %r94, 4;
and.b32 %r86, %r97, 252645135;

dp4a.s32.s32 %r81, %r82, %r10, %r77;


dp4a.s32.s32 %r85, %r86, %r11, %r81;

add.s64 %rd58, %rd76, %rd6;
ld.global.u32 %r90, [%rd58];

{mul.f16x2 %r89,%r90,%r12;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r89;
cvt.f32.f16 %f48, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r89;
cvt.f32.f16 %f49, high;}


cvt.rn.f32.s32 %f50, %r85;
div.approx.ftz.f32 %f52, %f49, %f45;
fma.rn.ftz.f32 %f53, %f48, %f50, %f52;
add.ftz.f32 %f171, %f53, %f171;

$L__BB41_5:
add.s64 %rd59, %rd78, %rd10;
ld.global.nc.u32 %r119, [%rd59+4];
ld.global.nc.u32 %r120, [%rd59];
and.b32 %r99, %r120, 252645135;
shr.u32 %r121, %r120, 4;
and.b32 %r103, %r121, 252645135;

dp4a.s32.s32 %r98, %r99, %r8, %r51;


dp4a.s32.s32 %r102, %r103, %r9, %r98;

and.b32 %r107, %r119, 252645135;
shr.u32 %r122, %r119, 4;
and.b32 %r111, %r122, 252645135;

dp4a.s32.s32 %r106, %r107, %r10, %r102;


dp4a.s32.s32 %r110, %r111, %r11, %r106;

add.s64 %rd60, %rd78, %rd9;
ld.global.nc.u32 %r115, [%rd60];

{mul.f16x2 %r114,%r115,%r12;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r114;
cvt.f32.f16 %f54, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r114;
cvt.f32.f16 %f55, high;}


cvt.rn.f32.s32 %f56, %r110;
div.approx.ftz.f32 %f58, %f55, %f45;
fma.rn.ftz.f32 %f59, %f54, %f56, %f58;
add.ftz.f32 %f172, %f59, %f172;
@%p3 bra $L__BB41_7;

add.s64 %rd61, %rd76, %rd10;
ld.global.u32 %r144, [%rd61+4];
ld.global.u32 %r145, [%rd61];
and.b32 %r124, %r145, 252645135;
shr.u32 %r146, %r145, 4;
and.b32 %r128, %r146, 252645135;
mov.u32 %r126, 0;

dp4a.s32.s32 %r123, %r124, %r8, %r126;


dp4a.s32.s32 %r127, %r128, %r9, %r123;

and.b32 %r132, %r144, 252645135;
shr.u32 %r147, %r144, 4;
and.b32 %r136, %r147, 252645135;

dp4a.s32.s32 %r131, %r132, %r10, %r127;


dp4a.s32.s32 %r135, %r136, %r11, %r131;

add.s64 %rd62, %rd76, %rd9;
ld.global.u32 %r140, [%rd62];

{mul.f16x2 %r139,%r140,%r12;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r139;
cvt.f32.f16 %f60, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r139;
cvt.f32.f16 %f61, high;}


cvt.rn.f32.s32 %f62, %r135;
mov.f32 %f63, 0f40000000;
div.approx.ftz.f32 %f64, %f61, %f63;
fma.rn.ftz.f32 %f65, %f60, %f62, %f64;
add.ftz.f32 %f170, %f65, %f170;

$L__BB41_7:
add.s64 %rd79, %rd79, 576;
add.s64 %rd78, %rd78, 320;
add.s64 %rd77, %rd77, 576;
add.s64 %rd76, %rd76, 320;
add.s32 %r233, %r233, 16;
setp.lt.u32 %p5, %r233, %r1;
@%p5 bra $L__BB41_3;

st.local.v2.f32 [%rd1], {%f173, %f172};
st.local.v2.f32 [%rd2], {%f171, %f170};

$L__BB41_9:
mov.b32 %r14, %f173;
setp.ne.s64 %p6, %rd28, 0;
@%p6 bra $L__BB41_11;
bra.uni $L__BB41_10;

$L__BB41_11:
mov.u32 %r174, 31;
mov.u32 %r175, 16;
mov.u32 %r176, -1;
shfl.sync.bfly.b32 %r177|%p17, %r14, %r175, %r174, %r176;
mov.b32 %f86, %r177;
add.ftz.f32 %f87, %f173, %f86;
mov.b32 %r178, %f87;
mov.u32 %r179, 8;
shfl.sync.bfly.b32 %r180|%p18, %r178, %r179, %r174, %r176;
mov.b32 %f88, %r180;
add.ftz.f32 %f89, %f87, %f88;
mov.b32 %r181, %f89;
mov.u32 %r182, 4;
shfl.sync.bfly.b32 %r183|%p19, %r181, %r182, %r174, %r176;
mov.b32 %f90, %r183;
add.ftz.f32 %f91, %f89, %f90;
mov.b32 %r184, %f91;
mov.u32 %r185, 2;
shfl.sync.bfly.b32 %r186|%p20, %r184, %r185, %r174, %r176;
mov.b32 %f92, %r186;
add.ftz.f32 %f93, %f91, %f92;
mov.b32 %r187, %f93;
mov.u32 %r188, 1;
shfl.sync.bfly.b32 %r189|%p21, %r187, %r188, %r174, %r176;
mov.b32 %f94, %r189;
add.ftz.f32 %f95, %f93, %f94;
st.local.f32 [%rd1], %f95;
mov.b32 %r190, %f171;
shfl.sync.bfly.b32 %r191|%p22, %r190, %r175, %r174, %r176;
mov.b32 %f96, %r191;
add.ftz.f32 %f97, %f171, %f96;
mov.b32 %r192, %f97;
shfl.sync.bfly.b32 %r193|%p23, %r192, %r179, %r174, %r176;
mov.b32 %f98, %r193;
add.ftz.f32 %f99, %f97, %f98;
mov.b32 %r194, %f99;
shfl.sync.bfly.b32 %r195|%p24, %r194, %r182, %r174, %r176;
mov.b32 %f100, %r195;
add.ftz.f32 %f101, %f99, %f100;
mov.b32 %r196, %f101;
shfl.sync.bfly.b32 %r197|%p25, %r196, %r185, %r174, %r176;
mov.b32 %f102, %r197;
add.ftz.f32 %f103, %f101, %f102;
mov.b32 %r198, %f103;
shfl.sync.bfly.b32 %r199|%p26, %r198, %r188, %r174, %r176;
mov.b32 %f104, %r199;
add.ftz.f32 %f105, %f103, %f104;
st.local.f32 [%rd2], %f105;
mov.b32 %r200, %f172;
shfl.sync.bfly.b32 %r201|%p27, %r200, %r175, %r174, %r176;
mov.b32 %f106, %r201;
add.ftz.f32 %f107, %f172, %f106;
mov.b32 %r202, %f107;
shfl.sync.bfly.b32 %r203|%p28, %r202, %r179, %r174, %r176;
mov.b32 %f108, %r203;
add.ftz.f32 %f109, %f107, %f108;
mov.b32 %r204, %f109;
shfl.sync.bfly.b32 %r205|%p29, %r204, %r182, %r174, %r176;
mov.b32 %f110, %r205;
add.ftz.f32 %f111, %f109, %f110;
mov.b32 %r206, %f111;
shfl.sync.bfly.b32 %r207|%p30, %r206, %r185, %r174, %r176;
mov.b32 %f112, %r207;
add.ftz.f32 %f113, %f111, %f112;
mov.b32 %r208, %f113;
shfl.sync.bfly.b32 %r209|%p31, %r208, %r188, %r174, %r176;
mov.b32 %f114, %r209;
add.ftz.f32 %f115, %f113, %f114;
st.local.f32 [%rd1+4], %f115;
mov.b32 %r210, %f170;
shfl.sync.bfly.b32 %r211|%p32, %r210, %r175, %r174, %r176;
mov.b32 %f116, %r211;
add.ftz.f32 %f117, %f170, %f116;
mov.b32 %r212, %f117;
shfl.sync.bfly.b32 %r213|%p33, %r212, %r179, %r174, %r176;
mov.b32 %f118, %r213;
add.ftz.f32 %f119, %f117, %f118;
mov.b32 %r214, %f119;
shfl.sync.bfly.b32 %r215|%p34, %r214, %r182, %r174, %r176;
mov.b32 %f120, %r215;
add.ftz.f32 %f121, %f119, %f120;
mov.b32 %r216, %f121;
shfl.sync.bfly.b32 %r217|%p35, %r216, %r185, %r174, %r176;
mov.b32 %f122, %r217;
add.ftz.f32 %f123, %f121, %f122;
mov.b32 %r218, %f123;
shfl.sync.bfly.b32 %r219|%p36, %r218, %r188, %r174, %r176;
mov.b32 %f124, %r219;
add.ftz.f32 %f125, %f123, %f124;
st.local.f32 [%rd2+4], %f125;
bra.uni $L__BB41_12;

$L__BB41_10:
mov.u32 %r148, 31;
mov.u32 %r149, 16;
mov.u32 %r150, -1;
shfl.sync.bfly.b32 %r151|%p7, %r14, %r149, %r148, %r150;
mov.b32 %f66, %r151;
add.ftz.f32 %f67, %f173, %f66;
mov.b32 %r152, %f67;
mov.u32 %r153, 8;
shfl.sync.bfly.b32 %r154|%p8, %r152, %r153, %r148, %r150;
mov.b32 %f68, %r154;
add.ftz.f32 %f69, %f67, %f68;
mov.b32 %r155, %f69;
mov.u32 %r156, 4;
shfl.sync.bfly.b32 %r157|%p9, %r155, %r156, %r148, %r150;
mov.b32 %f70, %r157;
add.ftz.f32 %f71, %f69, %f70;
mov.b32 %r158, %f71;
mov.u32 %r159, 2;
shfl.sync.bfly.b32 %r160|%p10, %r158, %r159, %r148, %r150;
mov.b32 %f72, %r160;
add.ftz.f32 %f73, %f71, %f72;
mov.b32 %r161, %f73;
mov.u32 %r162, 1;
shfl.sync.bfly.b32 %r163|%p11, %r161, %r162, %r148, %r150;
mov.b32 %f74, %r163;
add.ftz.f32 %f75, %f73, %f74;
st.local.f32 [%rd1], %f75;
mov.b32 %r164, %f172;
shfl.sync.bfly.b32 %r165|%p12, %r164, %r149, %r148, %r150;
mov.b32 %f76, %r165;
add.ftz.f32 %f77, %f172, %f76;
mov.b32 %r166, %f77;
shfl.sync.bfly.b32 %r167|%p13, %r166, %r153, %r148, %r150;
mov.b32 %f78, %r167;
add.ftz.f32 %f79, %f77, %f78;
mov.b32 %r168, %f79;
shfl.sync.bfly.b32 %r169|%p14, %r168, %r156, %r148, %r150;
mov.b32 %f80, %r169;
add.ftz.f32 %f81, %f79, %f80;
mov.b32 %r170, %f81;
shfl.sync.bfly.b32 %r171|%p15, %r170, %r159, %r148, %r150;
mov.b32 %f82, %r171;
add.ftz.f32 %f83, %f81, %f82;
mov.b32 %r172, %f83;
shfl.sync.bfly.b32 %r173|%p16, %r172, %r162, %r148, %r150;
mov.b32 %f84, %r173;
add.ftz.f32 %f85, %f83, %f84;
st.local.f32 [%rd1+4], %f85;

$L__BB41_12:
ld.param.u32 %r229, [_Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7];
setp.gt.u32 %p37, %r5, 1;
mov.u32 %r221, %ctaid.x;
shl.b32 %r222, %r221, 1;
add.s32 %r15, %r222, %r5;
setp.ge.u32 %p38, %r15, %r229;
or.pred %p39, %p37, %p38;
@%p39 bra $L__BB41_29;

cvt.u64.u32 %rd22, %r5;
mul.wide.u32 %rd63, %r5, 4;
add.s64 %rd64, %rd1, %rd63;
ld.local.f32 %f176, [%rd64];
mad.lo.s32 %r16, %r4, %r27, %r15;
setp.eq.s64 %p40, %rd27, 0;
@%p40 bra $L__BB41_15;

cvta.to.global.u64 %rd65, %rd27;
mul.wide.u32 %rd66, %r16, 4;
add.s64 %rd67, %rd65, %rd66;
ld.global.f32 %f126, [%rd67];
add.ftz.f32 %f176, %f176, %f126;

$L__BB41_15:
setp.eq.s64 %p41, %rd28, 0;
@%p41 bra $L__BB41_28;

shl.b64 %rd68, %rd22, 2;
add.s64 %rd69, %rd2, %rd68;
ld.local.f32 %f24, [%rd69];
setp.eq.s64 %p42, %rd29, 0;
@%p42 bra $L__BB41_18;

cvta.to.global.u64 %rd70, %rd29;
mul.wide.u32 %rd71, %r16, 4;
add.s64 %rd72, %rd70, %rd71;
ld.global.f32 %f127, [%rd72];
add.ftz.f32 %f24, %f24, %f127;

$L__BB41_18:
setp.gt.s32 %p43, %r17, 2;
@%p43 bra $L__BB41_22;

setp.eq.s32 %p46, %r17, 1;
@%p46 bra $L__BB41_26;

setp.eq.s32 %p47, %r17, 2;
@%p47 bra $L__BB41_21;
bra.uni $L__BB41_27;

$L__BB41_21:
mul.ftz.f32 %f147, %f24, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f148, %f147;
add.ftz.f32 %f149, %f148, 0f3F800000;
div.approx.ftz.f32 %f150, %f24, %f149;
mul.ftz.f32 %f176, %f176, %f150;
bra.uni $L__BB41_28;

$L__BB41_22:
setp.eq.s32 %p44, %r17, 3;
@%p44 bra $L__BB41_25;

setp.ne.s32 %p45, %r17, 6;
@%p45 bra $L__BB41_27;

ld.param.f32 %f159, [_Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+44];
min.ftz.f32 %f128, %f24, %f159;
neg.ftz.f32 %f129, %f159;
min.ftz.f32 %f130, %f176, %f159;
max.ftz.f32 %f131, %f130, %f129;
mul.ftz.f32 %f132, %f128, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f133, %f132;
add.ftz.f32 %f134, %f133, 0f3F800000;
div.approx.ftz.f32 %f135, %f128, %f134;
mul.ftz.f32 %f176, %f131, %f135;
bra.uni $L__BB41_28;

$L__BB41_27:
mul.ftz.f32 %f176, %f176, %f24;
bra.uni $L__BB41_28;

$L__BB41_26:
mul.ftz.f32 %f151, %f24, 0f3F000000;
mul.ftz.f32 %f152, %f24, 0f3D372713;
fma.rn.ftz.f32 %f153, %f24, %f152, 0f3F800000;
mul.ftz.f32 %f154, %f24, 0f3F4C422A;
mul.ftz.f32 %f155, %f154, %f153;
tanh.approx.f32 %f156, %f155;
add.ftz.f32 %f157, %f156, 0f3F800000;
mul.ftz.f32 %f158, %f151, %f157;
mul.ftz.f32 %f176, %f176, %f158;
bra.uni $L__BB41_28;

$L__BB41_25:
mov.f32 %f136, 0f40E00000;
min.ftz.f32 %f137, %f24, %f136;
min.ftz.f32 %f138, %f176, %f136;
mov.f32 %f139, 0fC0E00000;
max.ftz.f32 %f140, %f138, %f139;
mul.ftz.f32 %f141, %f137, 0f3FD9DB23;
mul.ftz.f32 %f142, %f141, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f143, %f142;
add.ftz.f32 %f144, %f143, 0f3F800000;
div.approx.ftz.f32 %f145, %f137, %f144;
add.ftz.f32 %f146, %f140, 0f3F800000;
mul.ftz.f32 %f176, %f146, %f145;

$L__BB41_28:
ld.param.u32 %r232, [_Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10];
mov.u32 %r231, %ctaid.y;
mov.u32 %r230, %tid.y;
mad.lo.s32 %r227, %r230, %r232, %r15;
mad.lo.s32 %r228, %r231, %r27, %r227;
cvta.to.global.u64 %rd73, %rd32;
mul.wide.u32 %rd74, %r228, 4;
add.s64 %rd75, %rd73, %rd74;
st.global.f32 [%rd75], %f176;

$L__BB41_29:
ret;

}
.entry _Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj(
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2,
.param .align 8 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3[48],
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5,
.param .align 4 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6[12],
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15
)
.maxntid 256, 1, 1
.minnctapersm 1
{
.local .align 8 .b8 __local_depot42[8];
.reg .b64 %SP;
.reg .b64 %SPL;
.reg .pred %p<17>;
.reg .f32 %f<47>;
.reg .b32 %r<122>;
.reg .b64 %rd<54>;


mov.u64 %SPL, __local_depot42;
ld.param.u64 %rd17, [_Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0];
ld.param.u64 %rd18, [_Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1];
ld.param.u64 %rd19, [_Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2];
ld.param.u64 %rd20, [_Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4];
ld.param.u32 %r20, [_Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5];
ld.param.u32 %r12, [_Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7];
ld.param.u32 %r13, [_Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8];
ld.param.u32 %r14, [_Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9];
ld.param.u32 %r15, [_Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10];
ld.param.u32 %r16, [_Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11];
ld.param.u32 %r17, [_Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12];
ld.param.u32 %r18, [_Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13];
ld.param.u32 %r21, [_Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14];
ld.param.u32 %r19, [_Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15];
ld.param.u32 %r11, [_Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+8];
ld.param.u32 %r10, [_Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+4];
ld.param.u32 %r9, [_Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6];
add.u64 %rd1, %SPL, 0;
mov.u32 %r22, %ctaid.x;
shl.b32 %r1, %r22, 1;
shr.u32 %r2, %r20, 5;
mov.u32 %r3, %tid.y;
setp.ge.u32 %p1, %r3, %r21;
@%p1 bra $L__BB42_7;

mov.f32 %f45, 0f00000000;
st.local.v2.f32 [%rd1], {%f45, %f45};
mov.u32 %r4, %tid.x;
shr.u32 %r121, %r4, 1;
setp.ge.u32 %p2, %r121, %r2;
mov.f32 %f46, %f45;
@%p2 bra $L__BB42_5;

mov.u32 %r23, %ctaid.y;
cvta.to.global.u64 %rd22, %rd18;
mul.hi.u32 %r24, %r23, %r9;
add.s32 %r25, %r24, %r23;
shr.u32 %r26, %r25, %r10;
mul.lo.s32 %r27, %r26, %r11;
sub.s32 %r28, %r23, %r27;
mul.lo.s32 %r29, %r28, %r17;
cvt.u64.u32 %rd23, %r29;
mul.lo.s32 %r30, %r3, %r14;
cvt.u64.u32 %rd24, %r30;
mad.lo.s32 %r31, %r3, %r19, %r23;
cvta.to.global.u64 %rd25, %rd19;
mul.wide.u32 %rd26, %r31, 4;
add.s64 %rd27, %rd25, %rd26;
ld.global.nc.u32 %r32, [%rd27];
mul.lo.s32 %r33, %r32, %r16;
cvt.u64.u32 %rd28, %r4;
shr.u64 %rd29, %rd28, 1;
mul.lo.s64 %rd30, %rd29, 36;
mul.wide.u32 %rd31, %r30, 36;
add.s64 %rd32, %rd30, %rd31;
add.s64 %rd3, %rd22, %rd32;
mul.wide.u32 %rd53, %r29, 36;
and.b32 %r34, %r4, 1;
mul.wide.u32 %rd33, %r34, 8;
mad.lo.s32 %r35, %r1, %r13, %r121;
add.s32 %r36, %r35, %r33;
mul.wide.s32 %rd5, %r36, 20;
add.s64 %rd34, %rd33, %rd5;
add.s64 %rd6, %rd34, 8;
add.s64 %rd35, %rd29, %rd24;
add.s64 %rd36, %rd35, %rd23;
mul.lo.s64 %rd37, %rd36, 36;
add.s64 %rd38, %rd37, %rd33;
add.s64 %rd39, %rd22, %rd38;
add.s64 %rd51, %rd39, 20;
add.s32 %r37, %r1, 1;
add.s32 %r38, %r121, %r33;
mad.lo.s32 %r39, %r13, %r37, %r38;
mul.wide.s32 %rd8, %r39, 20;
add.s64 %rd40, %rd33, %rd8;
add.s64 %rd9, %rd40, 8;
cvta.to.global.u64 %rd52, %rd17;
mov.f32 %f45, 0f00000000;

$L__BB42_3:
add.s64 %rd41, %rd3, %rd53;
ld.global.nc.u32 %r42, [%rd51+-16];
ld.global.nc.u32 %r46, [%rd51];
add.s64 %rd42, %rd52, %rd6;
ld.global.nc.u32 %r82, [%rd42];
ld.global.nc.u32 %r50, [%rd51+-12];
ld.global.nc.u32 %r54, [%rd51+4];
ld.global.nc.u32 %r83, [%rd42+-4];
and.b32 %r41, %r83, 252645135;
shr.u32 %r84, %r83, 4;
and.b32 %r45, %r84, 252645135;
mov.u32 %r64, 0;

dp4a.s32.s32 %r40, %r41, %r42, %r64;


dp4a.s32.s32 %r44, %r45, %r46, %r40;

and.b32 %r49, %r82, 252645135;
shr.u32 %r85, %r82, 4;
and.b32 %r53, %r85, 252645135;

dp4a.s32.s32 %r48, %r49, %r50, %r44;


dp4a.s32.s32 %r52, %r53, %r54, %r48;

add.s64 %rd43, %rd52, %rd5;
ld.global.nc.u32 %r57, [%rd43];
ld.global.nc.u32 %r58, [%rd41];

{mul.f16x2 %r56,%r57,%r58;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r56;
cvt.f32.f16 %f11, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r56;
cvt.f32.f16 %f12, high;}


cvt.rn.f32.s32 %f15, %r52;
mov.f32 %f16, 0f40000000;
div.approx.ftz.f32 %f17, %f12, %f16;
fma.rn.ftz.f32 %f18, %f11, %f15, %f17;
add.ftz.f32 %f46, %f18, %f46;
add.s64 %rd44, %rd52, %rd9;
ld.global.nc.u32 %r86, [%rd44];
ld.global.nc.u32 %r87, [%rd44+-4];
and.b32 %r62, %r87, 252645135;
shr.u32 %r88, %r87, 4;
and.b32 %r66, %r88, 252645135;

dp4a.s32.s32 %r61, %r62, %r42, %r64;


dp4a.s32.s32 %r65, %r66, %r46, %r61;

and.b32 %r70, %r86, 252645135;
shr.u32 %r89, %r86, 4;
and.b32 %r74, %r89, 252645135;

dp4a.s32.s32 %r69, %r70, %r50, %r65;


dp4a.s32.s32 %r73, %r74, %r54, %r69;

add.s64 %rd45, %rd52, %rd8;
ld.global.nc.u32 %r78, [%rd45];

{mul.f16x2 %r77,%r78,%r58;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r77;
cvt.f32.f16 %f13, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r77;
cvt.f32.f16 %f14, high;}


cvt.rn.f32.s32 %f19, %r73;
div.approx.ftz.f32 %f20, %f14, %f16;
fma.rn.ftz.f32 %f21, %f13, %f19, %f20;
add.ftz.f32 %f45, %f21, %f45;
add.s64 %rd53, %rd53, 576;
add.s64 %rd52, %rd52, 320;
add.s64 %rd51, %rd51, 576;
add.s32 %r121, %r121, 16;
setp.lt.u32 %p3, %r121, %r2;
@%p3 bra $L__BB42_3;

st.local.v2.f32 [%rd1], {%f46, %f45};

$L__BB42_5:
mov.b32 %r90, %f46;
mov.u32 %r91, 31;
mov.u32 %r92, 16;
mov.u32 %r93, -1;
shfl.sync.bfly.b32 %r94|%p4, %r90, %r92, %r91, %r93;
mov.b32 %f22, %r94;
add.ftz.f32 %f23, %f46, %f22;
mov.b32 %r95, %f23;
mov.u32 %r96, 8;
shfl.sync.bfly.b32 %r97|%p5, %r95, %r96, %r91, %r93;
mov.b32 %f24, %r97;
add.ftz.f32 %f25, %f23, %f24;
mov.b32 %r98, %f25;
mov.u32 %r99, 4;
shfl.sync.bfly.b32 %r100|%p6, %r98, %r99, %r91, %r93;
mov.b32 %f26, %r100;
add.ftz.f32 %f27, %f25, %f26;
mov.b32 %r101, %f27;
mov.u32 %r102, 2;
shfl.sync.bfly.b32 %r103|%p7, %r101, %r102, %r91, %r93;
mov.b32 %f28, %r103;
add.ftz.f32 %f29, %f27, %f28;
mov.b32 %r104, %f29;
mov.u32 %r105, 1;
shfl.sync.bfly.b32 %r106|%p8, %r104, %r105, %r91, %r93;
mov.b32 %f30, %r106;
add.ftz.f32 %f31, %f29, %f30;
st.local.f32 [%rd1], %f31;
mov.b32 %r107, %f45;
shfl.sync.bfly.b32 %r108|%p9, %r107, %r92, %r91, %r93;
mov.b32 %f32, %r108;
add.ftz.f32 %f33, %f45, %f32;
mov.b32 %r109, %f33;
shfl.sync.bfly.b32 %r110|%p10, %r109, %r96, %r91, %r93;
mov.b32 %f34, %r110;
add.ftz.f32 %f35, %f33, %f34;
mov.b32 %r111, %f35;
shfl.sync.bfly.b32 %r112|%p11, %r111, %r99, %r91, %r93;
mov.b32 %f36, %r112;
add.ftz.f32 %f37, %f35, %f36;
mov.b32 %r113, %f37;
shfl.sync.bfly.b32 %r114|%p12, %r113, %r102, %r91, %r93;
mov.b32 %f38, %r114;
add.ftz.f32 %f39, %f37, %f38;
mov.b32 %r115, %f39;
shfl.sync.bfly.b32 %r116|%p13, %r115, %r105, %r91, %r93;
mov.b32 %f40, %r116;
add.ftz.f32 %f41, %f39, %f40;
st.local.f32 [%rd1+4], %f41;
add.s32 %r8, %r1, %r4;
setp.ge.u32 %p14, %r8, %r12;
setp.gt.u32 %p15, %r4, 1;
or.pred %p16, %p15, %p14;
@%p16 bra $L__BB42_7;

mul.wide.u32 %rd46, %r4, 4;
add.s64 %rd47, %rd1, %rd46;
ld.local.f32 %f42, [%rd47];
mov.u32 %r117, %ctaid.y;
mad.lo.s32 %r119, %r3, %r15, %r8;
mad.lo.s32 %r120, %r117, %r18, %r119;
cvta.to.global.u64 %rd48, %rd20;
mul.wide.u32 %rd49, %r120, 4;
add.s64 %rd50, %rd48, %rd49;
st.global.f32 [%rd50], %f42;

$L__BB42_7:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<320>;
.reg .f32 %f<1021>;
.reg .b32 %r<1053>;
.reg .b64 %rd<96>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate[1536];

ld.param.u64 %rd24, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd25, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd26, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r48, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r49, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r50, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r54, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r55, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r56, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r52, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r51, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r46, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r45, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
ld.param.u64 %rd29, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+16];
ld.param.u64 %rd28, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+8];
ld.param.u64 %rd27, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3];
mov.u32 %r1, %tid.y;
shl.b32 %r57, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r57, %r2;
mov.u32 %r58, %ctaid.x;
shl.b32 %r4, %r58, 2;
mov.u32 %r5, %ctaid.y;
setp.eq.s64 %p1, %rd26, 0;
@%p1 bra $L__BB43_2;

cvta.to.global.u64 %rd33, %rd26;
mul.wide.u32 %rd34, %r5, 4;
add.s64 %rd35, %rd33, %rd34;
ld.global.nc.u32 %r1050, [%rd35];
mul.hi.u32 %r59, %r5, %r40;
add.s32 %r60, %r59, %r5;
shr.u32 %r61, %r60, %r41;
mul.lo.s32 %r62, %r61, %r42;
sub.s32 %r1051, %r5, %r62;
bra.uni $L__BB43_3;

$L__BB43_2:
mul.hi.u32 %r63, %r5, %r45;
add.s32 %r64, %r63, %r5;
shr.u32 %r1050, %r64, %r46;
mov.u32 %r1051, %r5;

$L__BB43_3:
mov.u32 %r11, %ctaid.z;
selp.b32 %r12, %r5, %r1050, %p1;
setp.ne.s32 %p3, %r1, 0;
setp.gt.u32 %p4, %r2, 3;
or.pred %p5, %p4, %p3;
add.s32 %r65, %r4, %r2;
setp.ge.u32 %p6, %r65, %r44;
mov.f32 %f27, 0f00000000;
or.pred %p7, %p5, %p6;
mov.f32 %f4, %f27;
mov.f32 %f5, %f27;
@%p7 bra $L__BB43_8;

setp.eq.s64 %p8, %rd27, 0;
mov.f32 %f4, 0f00000000;
mov.f32 %f5, %f4;
@%p8 bra $L__BB43_6;

cvta.to.global.u64 %rd36, %rd27;
mul.lo.s32 %r66, %r11, %r56;
cvt.u64.u32 %rd37, %r66;
mul.lo.s32 %r67, %r12, %r50;
cvt.u64.u32 %rd38, %r67;
cvt.u64.u32 %rd39, %r2;
cvt.s64.s32 %rd40, %r4;
add.s64 %rd41, %rd40, %rd39;
add.s64 %rd42, %rd41, %rd37;
add.s64 %rd43, %rd42, %rd38;
shl.b64 %rd44, %rd43, 2;
add.s64 %rd45, %rd36, %rd44;
ld.global.f32 %f5, [%rd45];

$L__BB43_6:
setp.eq.s64 %p9, %rd29, 0;
setp.eq.s64 %p10, %rd28, 0;
or.pred %p11, %p9, %p10;
@%p11 bra $L__BB43_8;

cvta.to.global.u64 %rd46, %rd29;
mul.lo.s32 %r68, %r11, %r56;
cvt.u64.u32 %rd47, %r68;
mul.lo.s32 %r69, %r12, %r50;
cvt.u64.u32 %rd48, %r69;
cvt.u64.u32 %rd49, %r2;
cvt.s64.s32 %rd50, %r4;
add.s64 %rd51, %rd50, %rd49;
add.s64 %rd52, %rd51, %rd47;
add.s64 %rd53, %rd52, %rd48;
shl.b64 %rd54, %rd53, 2;
add.s64 %rd55, %rd46, %rd54;
ld.global.f32 %f4, [%rd55];

$L__BB43_8:
mul.lo.s32 %r70, %r11, %r55;
cvt.u64.u32 %rd56, %r70;
mul.lo.s32 %r71, %r1051, %r49;
cvt.u64.u32 %rd57, %r71;
add.s64 %rd4, %rd57, %rd56;
mul.hi.u32 %r72, %r11, %r51;
add.s32 %r73, %r72, %r11;
shr.u32 %r13, %r73, %r52;
shr.u32 %r74, %r3, 31;
add.s32 %r75, %r3, %r74;
shr.s32 %r1052, %r75, 1;
shr.u32 %r76, %r39, 5;
setp.ge.s32 %p12, %r1052, %r76;
mov.f32 %f28, %f27;
mov.f32 %f29, %f27;
mov.f32 %f30, %f27;
mov.f32 %f31, %f27;
mov.f32 %f32, %f27;
mov.f32 %f33, %f27;
mov.f32 %f34, %f27;
@%p12 bra $L__BB43_19;

mul.lo.s32 %r77, %r13, %r54;
mul.lo.s32 %r78, %r1050, %r48;
and.b32 %r81, %r75, 2147483646;
sub.s32 %r82, %r3, %r81;
shl.b32 %r83, %r82, 1;
cvt.s64.s32 %rd5, %r83;
add.s32 %r84, %r1052, %r77;
add.s32 %r15, %r84, %r78;
add.s32 %r85, %r4, 1;
mad.lo.s32 %r86, %r43, %r85, %r15;
mul.wide.s32 %rd58, %r86, 20;
mad.lo.s32 %r87, %r4, %r43, %r1052;
add.s32 %r88, %r87, %r77;
add.s32 %r89, %r88, %r78;
mul.wide.s32 %rd6, %r89, 20;
mul.wide.s32 %rd59, %r83, 4;
add.s64 %rd60, %rd6, %rd59;
add.s64 %rd7, %rd60, 8;
add.s64 %rd61, %rd58, %rd59;
add.s64 %rd8, %rd61, 4;
add.s32 %r90, %r4, 2;
mad.lo.s32 %r91, %r43, %r90, %r15;
mul.wide.s32 %rd62, %r91, 20;
add.s64 %rd63, %rd62, %rd59;
add.s64 %rd9, %rd63, 4;
add.s32 %r92, %r4, 3;
mad.lo.s32 %r93, %r43, %r92, %r15;
mul.wide.s32 %rd64, %r93, 20;
add.s64 %rd10, %rd64, %rd59;
cvta.to.global.u64 %rd95, %rd28;
cvta.to.global.u64 %rd94, %rd24;
mov.f32 %f27, 0f00000000;
cvta.to.global.u64 %rd65, %rd25;
shl.b64 %rd70, %rd5, 2;
mov.f32 %f28, %f27;
mov.f32 %f29, %f27;
mov.f32 %f30, %f27;
mov.f32 %f31, %f27;
mov.f32 %f32, %f27;
mov.f32 %f33, %f27;
mov.f32 %f34, %f27;

$L__BB43_10:
cvt.u64.u32 %rd66, %r1052;
add.s64 %rd67, %rd4, %rd66;
mul.lo.s64 %rd68, %rd67, 36;
add.s64 %rd69, %rd65, %rd68;
add.s64 %rd71, %rd69, %rd70;
add.s64 %rd72, %rd94, %rd7;
add.s64 %rd15, %rd72, -4;
ld.global.nc.u32 %r17, [%rd71+4];
ld.global.nc.u32 %r18, [%rd71+20];
ld.global.nc.u32 %r115, [%rd72];
ld.global.nc.u32 %r19, [%rd71+8];
ld.global.nc.u32 %r20, [%rd71+24];
ld.global.nc.u32 %r116, [%rd72+-4];
and.b32 %r95, %r116, 252645135;
shr.u32 %r117, %r116, 4;
and.b32 %r99, %r117, 252645135;
mov.u32 %r97, 0;

dp4a.s32.s32 %r94, %r95, %r17, %r97;


dp4a.s32.s32 %r98, %r99, %r18, %r94;

and.b32 %r103, %r115, 252645135;
shr.u32 %r118, %r115, 4;
and.b32 %r107, %r118, 252645135;

dp4a.s32.s32 %r102, %r103, %r19, %r98;


dp4a.s32.s32 %r106, %r107, %r20, %r102;

add.s64 %rd16, %rd94, %rd6;
ld.global.nc.u32 %r111, [%rd16];
ld.global.nc.u32 %r21, [%rd69];

{mul.f16x2 %r110,%r111,%r21;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r110;
cvt.f32.f16 %f101, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r110;
cvt.f32.f16 %f102, high;}


cvt.rn.f32.s32 %f103, %r106;
mov.f32 %f104, 0f40000000;
div.approx.ftz.f32 %f105, %f102, %f104;
fma.rn.ftz.f32 %f106, %f101, %f103, %f105;
add.ftz.f32 %f30, %f106, %f30;
setp.eq.s64 %p13, %rd28, 0;
@%p13 bra $L__BB43_12;

add.s64 %rd73, %rd95, %rd7;
ld.global.u32 %r140, [%rd73];
ld.global.u32 %r141, [%rd73+-4];
and.b32 %r120, %r141, 252645135;
shr.u32 %r142, %r141, 4;
and.b32 %r124, %r142, 252645135;

dp4a.s32.s32 %r119, %r120, %r17, %r97;


dp4a.s32.s32 %r123, %r124, %r18, %r119;

and.b32 %r128, %r140, 252645135;
shr.u32 %r143, %r140, 4;
and.b32 %r132, %r143, 252645135;

dp4a.s32.s32 %r127, %r128, %r19, %r123;


dp4a.s32.s32 %r131, %r132, %r20, %r127;

add.s64 %rd74, %rd95, %rd6;
ld.global.u32 %r136, [%rd74];

{mul.f16x2 %r135,%r136,%r21;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r135;
cvt.f32.f16 %f107, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r135;
cvt.f32.f16 %f108, high;}


cvt.rn.f32.s32 %f109, %r131;
div.approx.ftz.f32 %f111, %f108, %f104;
fma.rn.ftz.f32 %f112, %f107, %f109, %f111;
add.ftz.f32 %f34, %f112, %f34;

$L__BB43_12:
mul.wide.s32 %rd75, %r43, 20;
add.s64 %rd17, %rd15, %rd75;
ld.global.nc.u32 %r165, [%rd17+4];
ld.global.nc.u32 %r166, [%rd17];
and.b32 %r145, %r166, 252645135;
shr.u32 %r167, %r166, 4;
and.b32 %r149, %r167, 252645135;

dp4a.s32.s32 %r144, %r145, %r17, %r97;


dp4a.s32.s32 %r148, %r149, %r18, %r144;

and.b32 %r153, %r165, 252645135;
shr.u32 %r168, %r165, 4;
and.b32 %r157, %r168, 252645135;

dp4a.s32.s32 %r152, %r153, %r19, %r148;


dp4a.s32.s32 %r156, %r157, %r20, %r152;

add.s64 %rd18, %rd16, %rd75;
ld.global.nc.u32 %r161, [%rd18];

{mul.f16x2 %r160,%r161,%r21;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r160;
cvt.f32.f16 %f113, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r160;
cvt.f32.f16 %f114, high;}


cvt.rn.f32.s32 %f115, %r156;
div.approx.ftz.f32 %f117, %f114, %f104;
fma.rn.ftz.f32 %f118, %f113, %f115, %f117;
add.ftz.f32 %f29, %f118, %f29;
@%p13 bra $L__BB43_14;

add.s64 %rd76, %rd95, %rd8;
ld.global.u32 %r190, [%rd76+4];
ld.global.u32 %r191, [%rd76];
and.b32 %r170, %r191, 252645135;
shr.u32 %r192, %r191, 4;
and.b32 %r174, %r192, 252645135;
mov.u32 %r172, 0;

dp4a.s32.s32 %r169, %r170, %r17, %r172;


dp4a.s32.s32 %r173, %r174, %r18, %r169;

and.b32 %r178, %r190, 252645135;
shr.u32 %r193, %r190, 4;
and.b32 %r182, %r193, 252645135;

dp4a.s32.s32 %r177, %r178, %r19, %r173;


dp4a.s32.s32 %r181, %r182, %r20, %r177;

or.b32 %r196, %r4, 1;
mad.lo.s32 %r197, %r43, %r196, %r15;
mul.wide.s32 %rd77, %r197, 20;
add.s64 %rd78, %rd95, %rd77;
ld.global.u32 %r186, [%rd78];

{mul.f16x2 %r185,%r186,%r21;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r185;
cvt.f32.f16 %f119, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r185;
cvt.f32.f16 %f120, high;}


cvt.rn.f32.s32 %f121, %r181;
mov.f32 %f122, 0f40000000;
div.approx.ftz.f32 %f123, %f120, %f122;
fma.rn.ftz.f32 %f124, %f119, %f121, %f123;
add.ftz.f32 %f33, %f124, %f33;

$L__BB43_14:
mul.wide.s32 %rd93, %r43, 20;
add.s64 %rd19, %rd17, %rd93;
ld.global.nc.u32 %r219, [%rd19+4];
ld.global.nc.u32 %r220, [%rd19];
and.b32 %r199, %r220, 252645135;
shr.u32 %r221, %r220, 4;
and.b32 %r203, %r221, 252645135;
mov.u32 %r201, 0;

dp4a.s32.s32 %r198, %r199, %r17, %r201;


dp4a.s32.s32 %r202, %r203, %r18, %r198;

and.b32 %r207, %r219, 252645135;
shr.u32 %r222, %r219, 4;
and.b32 %r211, %r222, 252645135;

dp4a.s32.s32 %r206, %r207, %r19, %r202;


dp4a.s32.s32 %r210, %r211, %r20, %r206;

add.s64 %rd20, %rd18, %rd93;
ld.global.nc.u32 %r215, [%rd20];

{mul.f16x2 %r214,%r215,%r21;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r214;
cvt.f32.f16 %f125, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r214;
cvt.f32.f16 %f126, high;}


cvt.rn.f32.s32 %f127, %r210;
mov.f32 %f128, 0f40000000;
div.approx.ftz.f32 %f129, %f126, %f128;
fma.rn.ftz.f32 %f130, %f125, %f127, %f129;
add.ftz.f32 %f28, %f130, %f28;
@%p13 bra $L__BB43_16;

add.s64 %rd80, %rd95, %rd9;
ld.global.u32 %r244, [%rd80+4];
ld.global.u32 %r245, [%rd80];
and.b32 %r224, %r245, 252645135;
shr.u32 %r246, %r245, 4;
and.b32 %r228, %r246, 252645135;

dp4a.s32.s32 %r223, %r224, %r17, %r201;


dp4a.s32.s32 %r227, %r228, %r18, %r223;

and.b32 %r232, %r244, 252645135;
shr.u32 %r247, %r244, 4;
and.b32 %r236, %r247, 252645135;

dp4a.s32.s32 %r231, %r232, %r19, %r227;


dp4a.s32.s32 %r235, %r236, %r20, %r231;

or.b32 %r250, %r4, 2;
mad.lo.s32 %r251, %r43, %r250, %r15;
mul.wide.s32 %rd81, %r251, 20;
add.s64 %rd82, %rd95, %rd81;
ld.global.u32 %r240, [%rd82];

{mul.f16x2 %r239,%r240,%r21;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r239;
cvt.f32.f16 %f131, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r239;
cvt.f32.f16 %f132, high;}


cvt.rn.f32.s32 %f133, %r235;
div.approx.ftz.f32 %f135, %f132, %f128;
fma.rn.ftz.f32 %f136, %f131, %f133, %f135;
add.ftz.f32 %f32, %f136, %f32;

$L__BB43_16:
mul.wide.s32 %rd91, %r43, 20;
add.s64 %rd84, %rd19, %rd91;
ld.global.nc.u32 %r273, [%rd84+4];
ld.global.nc.u32 %r274, [%rd84];
and.b32 %r253, %r274, 252645135;
shr.u32 %r275, %r274, 4;
and.b32 %r257, %r275, 252645135;

dp4a.s32.s32 %r252, %r253, %r17, %r201;


dp4a.s32.s32 %r256, %r257, %r18, %r252;

and.b32 %r261, %r273, 252645135;
shr.u32 %r276, %r273, 4;
and.b32 %r265, %r276, 252645135;

dp4a.s32.s32 %r260, %r261, %r19, %r256;


dp4a.s32.s32 %r264, %r265, %r20, %r260;

add.s64 %rd85, %rd20, %rd91;
ld.global.nc.u32 %r269, [%rd85];

{mul.f16x2 %r268,%r269,%r21;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r268;
cvt.f32.f16 %f137, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r268;
cvt.f32.f16 %f138, high;}


cvt.rn.f32.s32 %f139, %r264;
div.approx.ftz.f32 %f141, %f138, %f128;
fma.rn.ftz.f32 %f142, %f137, %f139, %f141;
add.ftz.f32 %f27, %f142, %f27;
@%p13 bra $L__BB43_18;

add.s64 %rd86, %rd95, %rd10;
ld.global.u32 %r298, [%rd86+8];
ld.global.u32 %r299, [%rd86+4];
and.b32 %r278, %r299, 252645135;
shr.u32 %r300, %r299, 4;
and.b32 %r282, %r300, 252645135;
mov.u32 %r280, 0;

dp4a.s32.s32 %r277, %r278, %r17, %r280;


dp4a.s32.s32 %r281, %r282, %r18, %r277;

and.b32 %r286, %r298, 252645135;
shr.u32 %r301, %r298, 4;
and.b32 %r290, %r301, 252645135;

dp4a.s32.s32 %r285, %r286, %r19, %r281;


dp4a.s32.s32 %r289, %r290, %r20, %r285;

or.b32 %r304, %r4, 3;
mad.lo.s32 %r305, %r43, %r304, %r15;
mul.wide.s32 %rd87, %r305, 20;
add.s64 %rd88, %rd95, %rd87;
ld.global.u32 %r294, [%rd88];

{mul.f16x2 %r293,%r294,%r21;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r293;
cvt.f32.f16 %f143, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r293;
cvt.f32.f16 %f144, high;}


cvt.rn.f32.s32 %f145, %r289;
mov.f32 %f146, 0f40000000;
div.approx.ftz.f32 %f147, %f144, %f146;
fma.rn.ftz.f32 %f148, %f143, %f145, %f147;
add.ftz.f32 %f31, %f148, %f31;

$L__BB43_18:
add.s64 %rd95, %rd95, 1280;
add.s64 %rd94, %rd94, 1280;
add.s32 %r1052, %r1052, 64;
setp.lt.s32 %p17, %r1052, %r76;
@%p17 bra $L__BB43_10;

$L__BB43_19:
mov.u32 %r1039, %tid.y;
setp.eq.s32 %p18, %r1039, 0;
@%p18 bra $L__BB43_23;

mov.u32 %r1040, %tid.y;
setp.ne.s64 %p19, %rd28, 0;
add.s32 %r23, %r1040, -1;
shl.b32 %r310, %r1040, 9;
mov.u32 %r311, _ZZ13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r312, %r311, %r310;
shl.b32 %r313, %r2, 2;
add.s32 %r314, %r312, %r313;
st.shared.f32 [%r314+-512], %f30;
@%p19 bra $L__BB43_22;
bra.uni $L__BB43_21;

$L__BB43_22:
shl.b32 %r316, %r23, 9;
mov.u32 %r317, _ZZ13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r318, %r317, %r316;
add.s32 %r320, %r318, %r313;
st.shared.f32 [%r320], %f34;
add.s32 %r1036, %r314, -512;
st.shared.f32 [%r1036+128], %f29;
st.shared.f32 [%r320+128], %f33;
add.s32 %r1037, %r314, -512;
st.shared.f32 [%r1037+256], %f28;
st.shared.f32 [%r320+256], %f32;
add.s32 %r1038, %r314, -512;
st.shared.f32 [%r1038+384], %f27;
st.shared.f32 [%r320+384], %f31;
bra.uni $L__BB43_23;

$L__BB43_21:
add.s32 %r1033, %r314, -512;
st.shared.f32 [%r1033+128], %f29;
add.s32 %r1034, %r314, -512;
st.shared.f32 [%r1034+256], %f28;
add.s32 %r1035, %r314, -512;
st.shared.f32 [%r1035+384], %f27;

$L__BB43_23:
mov.u32 %r1041, %tid.y;
setp.ne.s32 %p319, %r1041, 0;
bar.sync 0;
@%p319 bra $L__BB43_78;
bra.uni $L__BB43_24;

$L__BB43_78:
ret;

$L__BB43_24:
ld.param.u64 %rd92, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.f32 %f993, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
ld.param.u32 %r1044, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r1043, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
mov.u32 %r1042, %ctaid.z;
setp.ne.s64 %p21, %rd28, 0;
mov.u32 %r323, %ctaid.y;
mad.lo.s32 %r326, %r323, %r1044, %r4;
mad.lo.s32 %r327, %r1042, %r1043, %r326;
neg.ftz.f32 %f35, %f993;
shl.b32 %r329, %r2, 2;
mov.u32 %r330, _ZZ13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r25, %r330, %r329;
cvta.to.global.u64 %rd89, %rd92;
mul.wide.u32 %rd90, %r327, 4;
add.s64 %rd23, %rd89, %rd90;
@%p21 bra $L__BB43_33;
bra.uni $L__BB43_25;

$L__BB43_33:
ld.param.u32 %r1045, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
mov.u32 %r417, _ZZ13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r27, %r417, %r329;
setp.gt.s32 %p54, %r1045, 2;
@%p54 bra $L__BB43_44;

ld.param.u32 %r1048, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p57, %r1048, 1;
@%p57 bra $L__BB43_62;

ld.param.u32 %r1049, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p58, %r1049, 2;
@%p58 bra $L__BB43_36;
bra.uni $L__BB43_70;

$L__BB43_36:
ld.shared.f32 %f549, [%r25];
add.ftz.f32 %f550, %f549, %f30;
ld.shared.f32 %f551, [%r27];
add.ftz.f32 %f552, %f551, %f34;
ld.shared.f32 %f553, [%r25+512];
add.ftz.f32 %f554, %f553, %f550;
ld.shared.f32 %f555, [%r27+512];
add.ftz.f32 %f556, %f555, %f552;
ld.shared.f32 %f557, [%r25+1024];
add.ftz.f32 %f558, %f557, %f554;
ld.shared.f32 %f559, [%r27+1024];
add.ftz.f32 %f560, %f559, %f556;
mov.b32 %r664, %f558;
mov.u32 %r665, 31;
mov.u32 %r666, 16;
mov.u32 %r667, -1;
shfl.sync.bfly.b32 %r668|%p163, %r664, %r666, %r665, %r667;
mov.b32 %f561, %r668;
add.ftz.f32 %f562, %f558, %f561;
mov.b32 %r669, %f562;
mov.u32 %r670, 8;
shfl.sync.bfly.b32 %r671|%p164, %r669, %r670, %r665, %r667;
mov.b32 %f563, %r671;
add.ftz.f32 %f564, %f562, %f563;
mov.b32 %r672, %f564;
mov.u32 %r673, 4;
shfl.sync.bfly.b32 %r674|%p165, %r672, %r673, %r665, %r667;
mov.b32 %f565, %r674;
add.ftz.f32 %f566, %f564, %f565;
mov.b32 %r675, %f566;
mov.u32 %r676, 2;
shfl.sync.bfly.b32 %r677|%p166, %r675, %r676, %r665, %r667;
mov.b32 %f567, %r677;
add.ftz.f32 %f568, %f566, %f567;
mov.b32 %r678, %f568;
mov.u32 %r679, 1;
shfl.sync.bfly.b32 %r680|%p167, %r678, %r679, %r665, %r667;
mov.b32 %f569, %r680;
add.ftz.f32 %f56, %f568, %f569;
mov.b32 %r681, %f560;
shfl.sync.bfly.b32 %r682|%p168, %r681, %r666, %r665, %r667;
mov.b32 %f570, %r682;
add.ftz.f32 %f571, %f560, %f570;
mov.b32 %r683, %f571;
shfl.sync.bfly.b32 %r684|%p169, %r683, %r670, %r665, %r667;
mov.b32 %f572, %r684;
add.ftz.f32 %f573, %f571, %f572;
mov.b32 %r685, %f573;
shfl.sync.bfly.b32 %r686|%p170, %r685, %r673, %r665, %r667;
mov.b32 %f574, %r686;
add.ftz.f32 %f575, %f573, %f574;
mov.b32 %r687, %f575;
shfl.sync.bfly.b32 %r688|%p171, %r687, %r676, %r665, %r667;
mov.b32 %f576, %r688;
add.ftz.f32 %f577, %f575, %f576;
mov.b32 %r689, %f577;
shfl.sync.bfly.b32 %r690|%p172, %r689, %r679, %r665, %r667;
mov.b32 %f578, %r690;
add.ftz.f32 %f57, %f577, %f578;
setp.ne.s32 %p173, %r2, 0;
setp.ge.u32 %p174, %r4, %r44;
or.pred %p175, %p173, %p174;
@%p175 bra $L__BB43_38;

add.ftz.f32 %f579, %f5, %f56;
add.ftz.f32 %f580, %f4, %f57;
mul.ftz.f32 %f581, %f580, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f582, %f581;
add.ftz.f32 %f583, %f582, 0f3F800000;
div.approx.ftz.f32 %f584, %f580, %f583;
mul.ftz.f32 %f585, %f579, %f584;
st.global.f32 [%rd23], %f585;

$L__BB43_38:
ld.shared.f32 %f586, [%r25+128];
add.ftz.f32 %f587, %f586, %f29;
ld.shared.f32 %f588, [%r27+128];
add.ftz.f32 %f589, %f588, %f33;
ld.shared.f32 %f590, [%r25+640];
add.ftz.f32 %f591, %f590, %f587;
ld.shared.f32 %f592, [%r27+640];
add.ftz.f32 %f593, %f592, %f589;
ld.shared.f32 %f594, [%r25+1152];
add.ftz.f32 %f595, %f594, %f591;
ld.shared.f32 %f596, [%r27+1152];
add.ftz.f32 %f597, %f596, %f593;
mov.b32 %r694, %f595;
shfl.sync.bfly.b32 %r698|%p176, %r694, %r666, %r665, %r667;
mov.b32 %f598, %r698;
add.ftz.f32 %f599, %f595, %f598;
mov.b32 %r699, %f599;
shfl.sync.bfly.b32 %r701|%p177, %r699, %r670, %r665, %r667;
mov.b32 %f600, %r701;
add.ftz.f32 %f601, %f599, %f600;
mov.b32 %r702, %f601;
shfl.sync.bfly.b32 %r704|%p178, %r702, %r673, %r665, %r667;
mov.b32 %f602, %r704;
add.ftz.f32 %f603, %f601, %f602;
mov.b32 %r705, %f603;
shfl.sync.bfly.b32 %r707|%p179, %r705, %r676, %r665, %r667;
mov.b32 %f604, %r707;
add.ftz.f32 %f605, %f603, %f604;
mov.b32 %r708, %f605;
shfl.sync.bfly.b32 %r710|%p180, %r708, %r679, %r665, %r667;
mov.b32 %f606, %r710;
add.ftz.f32 %f58, %f605, %f606;
mov.b32 %r711, %f597;
shfl.sync.bfly.b32 %r712|%p181, %r711, %r666, %r665, %r667;
mov.b32 %f607, %r712;
add.ftz.f32 %f608, %f597, %f607;
mov.b32 %r713, %f608;
shfl.sync.bfly.b32 %r714|%p182, %r713, %r670, %r665, %r667;
mov.b32 %f609, %r714;
add.ftz.f32 %f610, %f608, %f609;
mov.b32 %r715, %f610;
shfl.sync.bfly.b32 %r716|%p183, %r715, %r673, %r665, %r667;
mov.b32 %f611, %r716;
add.ftz.f32 %f612, %f610, %f611;
mov.b32 %r717, %f612;
shfl.sync.bfly.b32 %r718|%p184, %r717, %r676, %r665, %r667;
mov.b32 %f613, %r718;
add.ftz.f32 %f614, %f612, %f613;
mov.b32 %r719, %f614;
shfl.sync.bfly.b32 %r720|%p185, %r719, %r679, %r665, %r667;
mov.b32 %f615, %r720;
add.ftz.f32 %f59, %f614, %f615;
setp.ne.s32 %p186, %r2, 1;
or.b32 %r724, %r4, 1;
setp.ge.u32 %p187, %r724, %r44;
or.pred %p188, %p186, %p187;
@%p188 bra $L__BB43_40;

add.ftz.f32 %f616, %f5, %f58;
add.ftz.f32 %f617, %f4, %f59;
mul.ftz.f32 %f618, %f617, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f619, %f618;
add.ftz.f32 %f620, %f619, 0f3F800000;
div.approx.ftz.f32 %f621, %f617, %f620;
mul.ftz.f32 %f622, %f616, %f621;
st.global.f32 [%rd23+4], %f622;

$L__BB43_40:
ld.shared.f32 %f623, [%r25+256];
add.ftz.f32 %f624, %f623, %f28;
ld.shared.f32 %f625, [%r27+256];
add.ftz.f32 %f626, %f625, %f32;
ld.shared.f32 %f627, [%r25+768];
add.ftz.f32 %f628, %f627, %f624;
ld.shared.f32 %f629, [%r27+768];
add.ftz.f32 %f630, %f629, %f626;
ld.shared.f32 %f631, [%r25+1280];
add.ftz.f32 %f632, %f631, %f628;
ld.shared.f32 %f633, [%r27+1280];
add.ftz.f32 %f634, %f633, %f630;
mov.b32 %r725, %f632;
mov.u32 %r726, 31;
mov.u32 %r727, 16;
mov.u32 %r728, -1;
shfl.sync.bfly.b32 %r729|%p189, %r725, %r727, %r726, %r728;
mov.b32 %f635, %r729;
add.ftz.f32 %f636, %f632, %f635;
mov.b32 %r730, %f636;
mov.u32 %r731, 8;
shfl.sync.bfly.b32 %r732|%p190, %r730, %r731, %r726, %r728;
mov.b32 %f637, %r732;
add.ftz.f32 %f638, %f636, %f637;
mov.b32 %r733, %f638;
mov.u32 %r734, 4;
shfl.sync.bfly.b32 %r735|%p191, %r733, %r734, %r726, %r728;
mov.b32 %f639, %r735;
add.ftz.f32 %f640, %f638, %f639;
mov.b32 %r736, %f640;
mov.u32 %r737, 2;
shfl.sync.bfly.b32 %r738|%p192, %r736, %r737, %r726, %r728;
mov.b32 %f641, %r738;
add.ftz.f32 %f642, %f640, %f641;
mov.b32 %r739, %f642;
mov.u32 %r740, 1;
shfl.sync.bfly.b32 %r741|%p193, %r739, %r740, %r726, %r728;
mov.b32 %f643, %r741;
add.ftz.f32 %f60, %f642, %f643;
mov.b32 %r742, %f634;
shfl.sync.bfly.b32 %r743|%p194, %r742, %r727, %r726, %r728;
mov.b32 %f644, %r743;
add.ftz.f32 %f645, %f634, %f644;
mov.b32 %r744, %f645;
shfl.sync.bfly.b32 %r745|%p195, %r744, %r731, %r726, %r728;
mov.b32 %f646, %r745;
add.ftz.f32 %f647, %f645, %f646;
mov.b32 %r746, %f647;
shfl.sync.bfly.b32 %r747|%p196, %r746, %r734, %r726, %r728;
mov.b32 %f648, %r747;
add.ftz.f32 %f649, %f647, %f648;
mov.b32 %r748, %f649;
shfl.sync.bfly.b32 %r749|%p197, %r748, %r737, %r726, %r728;
mov.b32 %f650, %r749;
add.ftz.f32 %f651, %f649, %f650;
mov.b32 %r750, %f651;
shfl.sync.bfly.b32 %r751|%p198, %r750, %r740, %r726, %r728;
mov.b32 %f652, %r751;
add.ftz.f32 %f61, %f651, %f652;
setp.ne.s32 %p199, %r2, 2;
or.b32 %r755, %r4, 2;
setp.ge.u32 %p200, %r755, %r44;
or.pred %p201, %p199, %p200;
@%p201 bra $L__BB43_42;

add.ftz.f32 %f653, %f5, %f60;
add.ftz.f32 %f654, %f4, %f61;
mul.ftz.f32 %f655, %f654, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f656, %f655;
add.ftz.f32 %f657, %f656, 0f3F800000;
div.approx.ftz.f32 %f658, %f654, %f657;
mul.ftz.f32 %f659, %f653, %f658;
st.global.f32 [%rd23+8], %f659;

$L__BB43_42:
ld.shared.f32 %f660, [%r25+384];
add.ftz.f32 %f661, %f660, %f27;
ld.shared.f32 %f662, [%r27+384];
add.ftz.f32 %f663, %f662, %f31;
ld.shared.f32 %f664, [%r25+896];
add.ftz.f32 %f665, %f664, %f661;
ld.shared.f32 %f666, [%r27+896];
add.ftz.f32 %f667, %f666, %f663;
ld.shared.f32 %f668, [%r25+1408];
add.ftz.f32 %f669, %f668, %f665;
ld.shared.f32 %f670, [%r27+1408];
add.ftz.f32 %f671, %f670, %f667;
mov.b32 %r756, %f669;
shfl.sync.bfly.b32 %r760|%p202, %r756, %r727, %r726, %r728;
mov.b32 %f672, %r760;
add.ftz.f32 %f673, %f669, %f672;
mov.b32 %r761, %f673;
shfl.sync.bfly.b32 %r763|%p203, %r761, %r731, %r726, %r728;
mov.b32 %f674, %r763;
add.ftz.f32 %f675, %f673, %f674;
mov.b32 %r764, %f675;
shfl.sync.bfly.b32 %r766|%p204, %r764, %r734, %r726, %r728;
mov.b32 %f676, %r766;
add.ftz.f32 %f677, %f675, %f676;
mov.b32 %r767, %f677;
shfl.sync.bfly.b32 %r769|%p205, %r767, %r737, %r726, %r728;
mov.b32 %f678, %r769;
add.ftz.f32 %f679, %f677, %f678;
mov.b32 %r770, %f679;
shfl.sync.bfly.b32 %r772|%p206, %r770, %r740, %r726, %r728;
mov.b32 %f680, %r772;
add.ftz.f32 %f62, %f679, %f680;
mov.b32 %r773, %f671;
shfl.sync.bfly.b32 %r774|%p207, %r773, %r727, %r726, %r728;
mov.b32 %f681, %r774;
add.ftz.f32 %f682, %f671, %f681;
mov.b32 %r775, %f682;
shfl.sync.bfly.b32 %r776|%p208, %r775, %r731, %r726, %r728;
mov.b32 %f683, %r776;
add.ftz.f32 %f684, %f682, %f683;
mov.b32 %r777, %f684;
shfl.sync.bfly.b32 %r778|%p209, %r777, %r734, %r726, %r728;
mov.b32 %f685, %r778;
add.ftz.f32 %f686, %f684, %f685;
mov.b32 %r779, %f686;
shfl.sync.bfly.b32 %r780|%p210, %r779, %r737, %r726, %r728;
mov.b32 %f687, %r780;
add.ftz.f32 %f688, %f686, %f687;
mov.b32 %r781, %f688;
shfl.sync.bfly.b32 %r782|%p211, %r781, %r740, %r726, %r728;
mov.b32 %f689, %r782;
add.ftz.f32 %f63, %f688, %f689;
setp.ne.s32 %p212, %r2, 3;
or.b32 %r786, %r4, 3;
setp.ge.u32 %p213, %r786, %r44;
or.pred %p214, %p212, %p213;
@%p214 bra $L__BB43_78;

add.ftz.f32 %f690, %f5, %f62;
add.ftz.f32 %f691, %f4, %f63;
mul.ftz.f32 %f692, %f691, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f693, %f692;
add.ftz.f32 %f694, %f693, 0f3F800000;
div.approx.ftz.f32 %f695, %f691, %f694;
mul.ftz.f32 %f696, %f690, %f695;
st.global.f32 [%rd23+12], %f696;
bra.uni $L__BB43_78;

$L__BB43_25:
ld.shared.f32 %f149, [%r25];
add.ftz.f32 %f150, %f149, %f30;
ld.shared.f32 %f151, [%r25+512];
add.ftz.f32 %f152, %f151, %f150;
ld.shared.f32 %f153, [%r25+1024];
add.ftz.f32 %f154, %f153, %f152;
mov.b32 %r331, %f154;
mov.u32 %r332, 31;
mov.u32 %r333, 16;
mov.u32 %r334, -1;
shfl.sync.bfly.b32 %r335|%p22, %r331, %r333, %r332, %r334;
mov.b32 %f155, %r335;
add.ftz.f32 %f156, %f154, %f155;
mov.b32 %r336, %f156;
mov.u32 %r337, 8;
shfl.sync.bfly.b32 %r338|%p23, %r336, %r337, %r332, %r334;
mov.b32 %f157, %r338;
add.ftz.f32 %f158, %f156, %f157;
mov.b32 %r339, %f158;
mov.u32 %r340, 4;
shfl.sync.bfly.b32 %r341|%p24, %r339, %r340, %r332, %r334;
mov.b32 %f159, %r341;
add.ftz.f32 %f160, %f158, %f159;
mov.b32 %r342, %f160;
mov.u32 %r343, 2;
shfl.sync.bfly.b32 %r344|%p25, %r342, %r343, %r332, %r334;
mov.b32 %f161, %r344;
add.ftz.f32 %f162, %f160, %f161;
mov.b32 %r345, %f162;
mov.u32 %r346, 1;
shfl.sync.bfly.b32 %r347|%p26, %r345, %r346, %r332, %r334;
mov.b32 %f163, %r347;
add.ftz.f32 %f36, %f162, %f163;
setp.ne.s32 %p27, %r2, 0;
setp.ge.u32 %p28, %r4, %r44;
or.pred %p29, %p27, %p28;
@%p29 bra $L__BB43_27;

add.ftz.f32 %f164, %f5, %f36;
st.global.f32 [%rd23], %f164;

$L__BB43_27:
ld.shared.f32 %f165, [%r25+128];
add.ftz.f32 %f166, %f165, %f29;
ld.shared.f32 %f167, [%r25+640];
add.ftz.f32 %f168, %f167, %f166;
ld.shared.f32 %f169, [%r25+1152];
add.ftz.f32 %f170, %f169, %f168;
mov.b32 %r351, %f170;
shfl.sync.bfly.b32 %r355|%p30, %r351, %r333, %r332, %r334;
mov.b32 %f171, %r355;
add.ftz.f32 %f172, %f170, %f171;
mov.b32 %r356, %f172;
shfl.sync.bfly.b32 %r358|%p31, %r356, %r337, %r332, %r334;
mov.b32 %f173, %r358;
add.ftz.f32 %f174, %f172, %f173;
mov.b32 %r359, %f174;
shfl.sync.bfly.b32 %r361|%p32, %r359, %r340, %r332, %r334;
mov.b32 %f175, %r361;
add.ftz.f32 %f176, %f174, %f175;
mov.b32 %r362, %f176;
shfl.sync.bfly.b32 %r364|%p33, %r362, %r343, %r332, %r334;
mov.b32 %f177, %r364;
add.ftz.f32 %f178, %f176, %f177;
mov.b32 %r365, %f178;
shfl.sync.bfly.b32 %r367|%p34, %r365, %r346, %r332, %r334;
mov.b32 %f179, %r367;
add.ftz.f32 %f37, %f178, %f179;
setp.ne.s32 %p35, %r2, 1;
or.b32 %r371, %r4, 1;
setp.ge.u32 %p36, %r371, %r44;
or.pred %p37, %p35, %p36;
@%p37 bra $L__BB43_29;

add.ftz.f32 %f180, %f5, %f37;
st.global.f32 [%rd23+4], %f180;

$L__BB43_29:
ld.shared.f32 %f181, [%r25+256];
add.ftz.f32 %f182, %f181, %f28;
ld.shared.f32 %f183, [%r25+768];
add.ftz.f32 %f184, %f183, %f182;
ld.shared.f32 %f185, [%r25+1280];
add.ftz.f32 %f186, %f185, %f184;
mov.b32 %r372, %f186;
mov.u32 %r373, 31;
mov.u32 %r374, 16;
mov.u32 %r375, -1;
shfl.sync.bfly.b32 %r376|%p38, %r372, %r374, %r373, %r375;
mov.b32 %f187, %r376;
add.ftz.f32 %f188, %f186, %f187;
mov.b32 %r377, %f188;
mov.u32 %r378, 8;
shfl.sync.bfly.b32 %r379|%p39, %r377, %r378, %r373, %r375;
mov.b32 %f189, %r379;
add.ftz.f32 %f190, %f188, %f189;
mov.b32 %r380, %f190;
mov.u32 %r381, 4;
shfl.sync.bfly.b32 %r382|%p40, %r380, %r381, %r373, %r375;
mov.b32 %f191, %r382;
add.ftz.f32 %f192, %f190, %f191;
mov.b32 %r383, %f192;
mov.u32 %r384, 2;
shfl.sync.bfly.b32 %r385|%p41, %r383, %r384, %r373, %r375;
mov.b32 %f193, %r385;
add.ftz.f32 %f194, %f192, %f193;
mov.b32 %r386, %f194;
mov.u32 %r387, 1;
shfl.sync.bfly.b32 %r388|%p42, %r386, %r387, %r373, %r375;
mov.b32 %f195, %r388;
add.ftz.f32 %f38, %f194, %f195;
setp.ne.s32 %p43, %r2, 2;
or.b32 %r392, %r4, 2;
setp.ge.u32 %p44, %r392, %r44;
or.pred %p45, %p43, %p44;
@%p45 bra $L__BB43_31;

add.ftz.f32 %f196, %f5, %f38;
st.global.f32 [%rd23+8], %f196;

$L__BB43_31:
ld.shared.f32 %f197, [%r25+384];
add.ftz.f32 %f198, %f197, %f27;
ld.shared.f32 %f199, [%r25+896];
add.ftz.f32 %f200, %f199, %f198;
ld.shared.f32 %f201, [%r25+1408];
add.ftz.f32 %f202, %f201, %f200;
mov.b32 %r393, %f202;
shfl.sync.bfly.b32 %r397|%p46, %r393, %r374, %r373, %r375;
mov.b32 %f203, %r397;
add.ftz.f32 %f204, %f202, %f203;
mov.b32 %r398, %f204;
shfl.sync.bfly.b32 %r400|%p47, %r398, %r378, %r373, %r375;
mov.b32 %f205, %r400;
add.ftz.f32 %f206, %f204, %f205;
mov.b32 %r401, %f206;
shfl.sync.bfly.b32 %r403|%p48, %r401, %r381, %r373, %r375;
mov.b32 %f207, %r403;
add.ftz.f32 %f208, %f206, %f207;
mov.b32 %r404, %f208;
shfl.sync.bfly.b32 %r406|%p49, %r404, %r384, %r373, %r375;
mov.b32 %f209, %r406;
add.ftz.f32 %f210, %f208, %f209;
mov.b32 %r407, %f210;
shfl.sync.bfly.b32 %r409|%p50, %r407, %r387, %r373, %r375;
mov.b32 %f211, %r409;
add.ftz.f32 %f39, %f210, %f211;
setp.ne.s32 %p51, %r2, 3;
or.b32 %r413, %r4, 3;
setp.ge.u32 %p52, %r413, %r44;
or.pred %p53, %p51, %p52;
@%p53 bra $L__BB43_78;

add.ftz.f32 %f212, %f5, %f39;
st.global.f32 [%rd23+12], %f212;
bra.uni $L__BB43_78;

$L__BB43_44:
ld.param.u32 %r1046, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p55, %r1046, 3;
@%p55 bra $L__BB43_54;

ld.param.u32 %r1047, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.ne.s32 %p56, %r1047, 6;
@%p56 bra $L__BB43_70;

ld.shared.f32 %f213, [%r25];
add.ftz.f32 %f214, %f213, %f30;
ld.shared.f32 %f215, [%r27];
add.ftz.f32 %f216, %f215, %f34;
ld.shared.f32 %f217, [%r25+512];
add.ftz.f32 %f218, %f217, %f214;
ld.shared.f32 %f219, [%r27+512];
add.ftz.f32 %f220, %f219, %f216;
ld.shared.f32 %f221, [%r25+1024];
add.ftz.f32 %f222, %f221, %f218;
ld.shared.f32 %f223, [%r27+1024];
add.ftz.f32 %f224, %f223, %f220;
mov.b32 %r418, %f222;
mov.u32 %r419, 31;
mov.u32 %r420, 16;
mov.u32 %r421, -1;
shfl.sync.bfly.b32 %r422|%p59, %r418, %r420, %r419, %r421;
mov.b32 %f225, %r422;
add.ftz.f32 %f226, %f222, %f225;
mov.b32 %r423, %f226;
mov.u32 %r424, 8;
shfl.sync.bfly.b32 %r425|%p60, %r423, %r424, %r419, %r421;
mov.b32 %f227, %r425;
add.ftz.f32 %f228, %f226, %f227;
mov.b32 %r426, %f228;
mov.u32 %r427, 4;
shfl.sync.bfly.b32 %r428|%p61, %r426, %r427, %r419, %r421;
mov.b32 %f229, %r428;
add.ftz.f32 %f230, %f228, %f229;
mov.b32 %r429, %f230;
mov.u32 %r430, 2;
shfl.sync.bfly.b32 %r431|%p62, %r429, %r430, %r419, %r421;
mov.b32 %f231, %r431;
add.ftz.f32 %f232, %f230, %f231;
mov.b32 %r432, %f232;
mov.u32 %r433, 1;
shfl.sync.bfly.b32 %r434|%p63, %r432, %r433, %r419, %r421;
mov.b32 %f233, %r434;
add.ftz.f32 %f40, %f232, %f233;
mov.b32 %r435, %f224;
shfl.sync.bfly.b32 %r436|%p64, %r435, %r420, %r419, %r421;
mov.b32 %f234, %r436;
add.ftz.f32 %f235, %f224, %f234;
mov.b32 %r437, %f235;
shfl.sync.bfly.b32 %r438|%p65, %r437, %r424, %r419, %r421;
mov.b32 %f236, %r438;
add.ftz.f32 %f237, %f235, %f236;
mov.b32 %r439, %f237;
shfl.sync.bfly.b32 %r440|%p66, %r439, %r427, %r419, %r421;
mov.b32 %f238, %r440;
add.ftz.f32 %f239, %f237, %f238;
mov.b32 %r441, %f239;
shfl.sync.bfly.b32 %r442|%p67, %r441, %r430, %r419, %r421;
mov.b32 %f240, %r442;
add.ftz.f32 %f241, %f239, %f240;
mov.b32 %r443, %f241;
shfl.sync.bfly.b32 %r444|%p68, %r443, %r433, %r419, %r421;
mov.b32 %f242, %r444;
add.ftz.f32 %f41, %f241, %f242;
setp.ne.s32 %p69, %r2, 0;
setp.ge.u32 %p70, %r4, %r44;
or.pred %p71, %p69, %p70;
@%p71 bra $L__BB43_48;

ld.param.f32 %f997, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f243, %f5, %f40;
add.ftz.f32 %f244, %f4, %f41;
min.ftz.f32 %f245, %f244, %f997;
min.ftz.f32 %f246, %f243, %f997;
max.ftz.f32 %f247, %f246, %f35;
mul.ftz.f32 %f248, %f245, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f249, %f248;
add.ftz.f32 %f250, %f249, 0f3F800000;
div.approx.ftz.f32 %f251, %f245, %f250;
mul.ftz.f32 %f252, %f247, %f251;
st.global.f32 [%rd23], %f252;

$L__BB43_48:
ld.shared.f32 %f253, [%r25+128];
add.ftz.f32 %f254, %f253, %f29;
ld.shared.f32 %f255, [%r27+128];
add.ftz.f32 %f256, %f255, %f33;
ld.shared.f32 %f257, [%r25+640];
add.ftz.f32 %f258, %f257, %f254;
ld.shared.f32 %f259, [%r27+640];
add.ftz.f32 %f260, %f259, %f256;
ld.shared.f32 %f261, [%r25+1152];
add.ftz.f32 %f262, %f261, %f258;
ld.shared.f32 %f263, [%r27+1152];
add.ftz.f32 %f264, %f263, %f260;
mov.b32 %r448, %f262;
shfl.sync.bfly.b32 %r452|%p72, %r448, %r420, %r419, %r421;
mov.b32 %f265, %r452;
add.ftz.f32 %f266, %f262, %f265;
mov.b32 %r453, %f266;
shfl.sync.bfly.b32 %r455|%p73, %r453, %r424, %r419, %r421;
mov.b32 %f267, %r455;
add.ftz.f32 %f268, %f266, %f267;
mov.b32 %r456, %f268;
shfl.sync.bfly.b32 %r458|%p74, %r456, %r427, %r419, %r421;
mov.b32 %f269, %r458;
add.ftz.f32 %f270, %f268, %f269;
mov.b32 %r459, %f270;
shfl.sync.bfly.b32 %r461|%p75, %r459, %r430, %r419, %r421;
mov.b32 %f271, %r461;
add.ftz.f32 %f272, %f270, %f271;
mov.b32 %r462, %f272;
shfl.sync.bfly.b32 %r464|%p76, %r462, %r433, %r419, %r421;
mov.b32 %f273, %r464;
add.ftz.f32 %f42, %f272, %f273;
mov.b32 %r465, %f264;
shfl.sync.bfly.b32 %r466|%p77, %r465, %r420, %r419, %r421;
mov.b32 %f274, %r466;
add.ftz.f32 %f275, %f264, %f274;
mov.b32 %r467, %f275;
shfl.sync.bfly.b32 %r468|%p78, %r467, %r424, %r419, %r421;
mov.b32 %f276, %r468;
add.ftz.f32 %f277, %f275, %f276;
mov.b32 %r469, %f277;
shfl.sync.bfly.b32 %r470|%p79, %r469, %r427, %r419, %r421;
mov.b32 %f278, %r470;
add.ftz.f32 %f279, %f277, %f278;
mov.b32 %r471, %f279;
shfl.sync.bfly.b32 %r472|%p80, %r471, %r430, %r419, %r421;
mov.b32 %f280, %r472;
add.ftz.f32 %f281, %f279, %f280;
mov.b32 %r473, %f281;
shfl.sync.bfly.b32 %r474|%p81, %r473, %r433, %r419, %r421;
mov.b32 %f282, %r474;
add.ftz.f32 %f43, %f281, %f282;
setp.ne.s32 %p82, %r2, 1;
or.b32 %r478, %r4, 1;
setp.ge.u32 %p83, %r478, %r44;
or.pred %p84, %p82, %p83;
@%p84 bra $L__BB43_50;

ld.param.f32 %f996, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f283, %f5, %f42;
add.ftz.f32 %f284, %f4, %f43;
min.ftz.f32 %f285, %f284, %f996;
min.ftz.f32 %f286, %f283, %f996;
max.ftz.f32 %f287, %f286, %f35;
mul.ftz.f32 %f288, %f285, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f289, %f288;
add.ftz.f32 %f290, %f289, 0f3F800000;
div.approx.ftz.f32 %f291, %f285, %f290;
mul.ftz.f32 %f292, %f287, %f291;
st.global.f32 [%rd23+4], %f292;

$L__BB43_50:
ld.shared.f32 %f293, [%r25+256];
add.ftz.f32 %f294, %f293, %f28;
ld.shared.f32 %f295, [%r27+256];
add.ftz.f32 %f296, %f295, %f32;
ld.shared.f32 %f297, [%r25+768];
add.ftz.f32 %f298, %f297, %f294;
ld.shared.f32 %f299, [%r27+768];
add.ftz.f32 %f300, %f299, %f296;
ld.shared.f32 %f301, [%r25+1280];
add.ftz.f32 %f302, %f301, %f298;
ld.shared.f32 %f303, [%r27+1280];
add.ftz.f32 %f304, %f303, %f300;
mov.b32 %r479, %f302;
mov.u32 %r480, 31;
mov.u32 %r481, 16;
mov.u32 %r482, -1;
shfl.sync.bfly.b32 %r483|%p85, %r479, %r481, %r480, %r482;
mov.b32 %f305, %r483;
add.ftz.f32 %f306, %f302, %f305;
mov.b32 %r484, %f306;
mov.u32 %r485, 8;
shfl.sync.bfly.b32 %r486|%p86, %r484, %r485, %r480, %r482;
mov.b32 %f307, %r486;
add.ftz.f32 %f308, %f306, %f307;
mov.b32 %r487, %f308;
mov.u32 %r488, 4;
shfl.sync.bfly.b32 %r489|%p87, %r487, %r488, %r480, %r482;
mov.b32 %f309, %r489;
add.ftz.f32 %f310, %f308, %f309;
mov.b32 %r490, %f310;
mov.u32 %r491, 2;
shfl.sync.bfly.b32 %r492|%p88, %r490, %r491, %r480, %r482;
mov.b32 %f311, %r492;
add.ftz.f32 %f312, %f310, %f311;
mov.b32 %r493, %f312;
mov.u32 %r494, 1;
shfl.sync.bfly.b32 %r495|%p89, %r493, %r494, %r480, %r482;
mov.b32 %f313, %r495;
add.ftz.f32 %f44, %f312, %f313;
mov.b32 %r496, %f304;
shfl.sync.bfly.b32 %r497|%p90, %r496, %r481, %r480, %r482;
mov.b32 %f314, %r497;
add.ftz.f32 %f315, %f304, %f314;
mov.b32 %r498, %f315;
shfl.sync.bfly.b32 %r499|%p91, %r498, %r485, %r480, %r482;
mov.b32 %f316, %r499;
add.ftz.f32 %f317, %f315, %f316;
mov.b32 %r500, %f317;
shfl.sync.bfly.b32 %r501|%p92, %r500, %r488, %r480, %r482;
mov.b32 %f318, %r501;
add.ftz.f32 %f319, %f317, %f318;
mov.b32 %r502, %f319;
shfl.sync.bfly.b32 %r503|%p93, %r502, %r491, %r480, %r482;
mov.b32 %f320, %r503;
add.ftz.f32 %f321, %f319, %f320;
mov.b32 %r504, %f321;
shfl.sync.bfly.b32 %r505|%p94, %r504, %r494, %r480, %r482;
mov.b32 %f322, %r505;
add.ftz.f32 %f45, %f321, %f322;
setp.ne.s32 %p95, %r2, 2;
or.b32 %r509, %r4, 2;
setp.ge.u32 %p96, %r509, %r44;
or.pred %p97, %p95, %p96;
@%p97 bra $L__BB43_52;

ld.param.f32 %f995, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f323, %f5, %f44;
add.ftz.f32 %f324, %f4, %f45;
min.ftz.f32 %f325, %f324, %f995;
min.ftz.f32 %f326, %f323, %f995;
max.ftz.f32 %f327, %f326, %f35;
mul.ftz.f32 %f328, %f325, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f329, %f328;
add.ftz.f32 %f330, %f329, 0f3F800000;
div.approx.ftz.f32 %f331, %f325, %f330;
mul.ftz.f32 %f332, %f327, %f331;
st.global.f32 [%rd23+8], %f332;

$L__BB43_52:
ld.shared.f32 %f333, [%r25+384];
add.ftz.f32 %f334, %f333, %f27;
ld.shared.f32 %f335, [%r27+384];
add.ftz.f32 %f336, %f335, %f31;
ld.shared.f32 %f337, [%r25+896];
add.ftz.f32 %f338, %f337, %f334;
ld.shared.f32 %f339, [%r27+896];
add.ftz.f32 %f340, %f339, %f336;
ld.shared.f32 %f341, [%r25+1408];
add.ftz.f32 %f342, %f341, %f338;
ld.shared.f32 %f343, [%r27+1408];
add.ftz.f32 %f344, %f343, %f340;
mov.b32 %r510, %f342;
shfl.sync.bfly.b32 %r514|%p98, %r510, %r481, %r480, %r482;
mov.b32 %f345, %r514;
add.ftz.f32 %f346, %f342, %f345;
mov.b32 %r515, %f346;
shfl.sync.bfly.b32 %r517|%p99, %r515, %r485, %r480, %r482;
mov.b32 %f347, %r517;
add.ftz.f32 %f348, %f346, %f347;
mov.b32 %r518, %f348;
shfl.sync.bfly.b32 %r520|%p100, %r518, %r488, %r480, %r482;
mov.b32 %f349, %r520;
add.ftz.f32 %f350, %f348, %f349;
mov.b32 %r521, %f350;
shfl.sync.bfly.b32 %r523|%p101, %r521, %r491, %r480, %r482;
mov.b32 %f351, %r523;
add.ftz.f32 %f352, %f350, %f351;
mov.b32 %r524, %f352;
shfl.sync.bfly.b32 %r526|%p102, %r524, %r494, %r480, %r482;
mov.b32 %f353, %r526;
add.ftz.f32 %f46, %f352, %f353;
mov.b32 %r527, %f344;
shfl.sync.bfly.b32 %r528|%p103, %r527, %r481, %r480, %r482;
mov.b32 %f354, %r528;
add.ftz.f32 %f355, %f344, %f354;
mov.b32 %r529, %f355;
shfl.sync.bfly.b32 %r530|%p104, %r529, %r485, %r480, %r482;
mov.b32 %f356, %r530;
add.ftz.f32 %f357, %f355, %f356;
mov.b32 %r531, %f357;
shfl.sync.bfly.b32 %r532|%p105, %r531, %r488, %r480, %r482;
mov.b32 %f358, %r532;
add.ftz.f32 %f359, %f357, %f358;
mov.b32 %r533, %f359;
shfl.sync.bfly.b32 %r534|%p106, %r533, %r491, %r480, %r482;
mov.b32 %f360, %r534;
add.ftz.f32 %f361, %f359, %f360;
mov.b32 %r535, %f361;
shfl.sync.bfly.b32 %r536|%p107, %r535, %r494, %r480, %r482;
mov.b32 %f362, %r536;
add.ftz.f32 %f47, %f361, %f362;
setp.ne.s32 %p108, %r2, 3;
or.b32 %r540, %r4, 3;
setp.ge.u32 %p109, %r540, %r44;
or.pred %p110, %p108, %p109;
@%p110 bra $L__BB43_78;

ld.param.f32 %f994, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f363, %f5, %f46;
add.ftz.f32 %f364, %f4, %f47;
min.ftz.f32 %f365, %f364, %f994;
min.ftz.f32 %f366, %f363, %f994;
max.ftz.f32 %f367, %f366, %f35;
mul.ftz.f32 %f368, %f365, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f369, %f368;
add.ftz.f32 %f370, %f369, 0f3F800000;
div.approx.ftz.f32 %f371, %f365, %f370;
mul.ftz.f32 %f372, %f367, %f371;
st.global.f32 [%rd23+12], %f372;
bra.uni $L__BB43_78;

$L__BB43_70:
ld.shared.f32 %f861, [%r25];
add.ftz.f32 %f862, %f861, %f30;
ld.shared.f32 %f863, [%r27];
add.ftz.f32 %f864, %f863, %f34;
ld.shared.f32 %f865, [%r25+512];
add.ftz.f32 %f866, %f865, %f862;
ld.shared.f32 %f867, [%r27+512];
add.ftz.f32 %f868, %f867, %f864;
ld.shared.f32 %f869, [%r25+1024];
add.ftz.f32 %f870, %f869, %f866;
ld.shared.f32 %f871, [%r27+1024];
add.ftz.f32 %f872, %f871, %f868;
mov.b32 %r910, %f870;
mov.u32 %r911, 31;
mov.u32 %r912, 16;
mov.u32 %r913, -1;
shfl.sync.bfly.b32 %r914|%p267, %r910, %r912, %r911, %r913;
mov.b32 %f873, %r914;
add.ftz.f32 %f874, %f870, %f873;
mov.b32 %r915, %f874;
mov.u32 %r916, 8;
shfl.sync.bfly.b32 %r917|%p268, %r915, %r916, %r911, %r913;
mov.b32 %f875, %r917;
add.ftz.f32 %f876, %f874, %f875;
mov.b32 %r918, %f876;
mov.u32 %r919, 4;
shfl.sync.bfly.b32 %r920|%p269, %r918, %r919, %r911, %r913;
mov.b32 %f877, %r920;
add.ftz.f32 %f878, %f876, %f877;
mov.b32 %r921, %f878;
mov.u32 %r922, 2;
shfl.sync.bfly.b32 %r923|%p270, %r921, %r922, %r911, %r913;
mov.b32 %f879, %r923;
add.ftz.f32 %f880, %f878, %f879;
mov.b32 %r924, %f880;
mov.u32 %r925, 1;
shfl.sync.bfly.b32 %r926|%p271, %r924, %r925, %r911, %r913;
mov.b32 %f881, %r926;
add.ftz.f32 %f72, %f880, %f881;
mov.b32 %r927, %f872;
shfl.sync.bfly.b32 %r928|%p272, %r927, %r912, %r911, %r913;
mov.b32 %f882, %r928;
add.ftz.f32 %f883, %f872, %f882;
mov.b32 %r929, %f883;
shfl.sync.bfly.b32 %r930|%p273, %r929, %r916, %r911, %r913;
mov.b32 %f884, %r930;
add.ftz.f32 %f885, %f883, %f884;
mov.b32 %r931, %f885;
shfl.sync.bfly.b32 %r932|%p274, %r931, %r919, %r911, %r913;
mov.b32 %f886, %r932;
add.ftz.f32 %f887, %f885, %f886;
mov.b32 %r933, %f887;
shfl.sync.bfly.b32 %r934|%p275, %r933, %r922, %r911, %r913;
mov.b32 %f888, %r934;
add.ftz.f32 %f889, %f887, %f888;
mov.b32 %r935, %f889;
shfl.sync.bfly.b32 %r936|%p276, %r935, %r925, %r911, %r913;
mov.b32 %f890, %r936;
add.ftz.f32 %f73, %f889, %f890;
setp.ne.s32 %p277, %r2, 0;
setp.ge.u32 %p278, %r4, %r44;
or.pred %p279, %p277, %p278;
@%p279 bra $L__BB43_72;

add.ftz.f32 %f891, %f5, %f72;
add.ftz.f32 %f892, %f4, %f73;
mul.ftz.f32 %f893, %f891, %f892;
st.global.f32 [%rd23], %f893;

$L__BB43_72:
ld.shared.f32 %f894, [%r25+128];
add.ftz.f32 %f895, %f894, %f29;
ld.shared.f32 %f896, [%r27+128];
add.ftz.f32 %f897, %f896, %f33;
ld.shared.f32 %f898, [%r25+640];
add.ftz.f32 %f899, %f898, %f895;
ld.shared.f32 %f900, [%r27+640];
add.ftz.f32 %f901, %f900, %f897;
ld.shared.f32 %f902, [%r25+1152];
add.ftz.f32 %f903, %f902, %f899;
ld.shared.f32 %f904, [%r27+1152];
add.ftz.f32 %f905, %f904, %f901;
mov.b32 %r940, %f903;
shfl.sync.bfly.b32 %r944|%p280, %r940, %r912, %r911, %r913;
mov.b32 %f906, %r944;
add.ftz.f32 %f907, %f903, %f906;
mov.b32 %r945, %f907;
shfl.sync.bfly.b32 %r947|%p281, %r945, %r916, %r911, %r913;
mov.b32 %f908, %r947;
add.ftz.f32 %f909, %f907, %f908;
mov.b32 %r948, %f909;
shfl.sync.bfly.b32 %r950|%p282, %r948, %r919, %r911, %r913;
mov.b32 %f910, %r950;
add.ftz.f32 %f911, %f909, %f910;
mov.b32 %r951, %f911;
shfl.sync.bfly.b32 %r953|%p283, %r951, %r922, %r911, %r913;
mov.b32 %f912, %r953;
add.ftz.f32 %f913, %f911, %f912;
mov.b32 %r954, %f913;
shfl.sync.bfly.b32 %r956|%p284, %r954, %r925, %r911, %r913;
mov.b32 %f914, %r956;
add.ftz.f32 %f74, %f913, %f914;
mov.b32 %r957, %f905;
shfl.sync.bfly.b32 %r958|%p285, %r957, %r912, %r911, %r913;
mov.b32 %f915, %r958;
add.ftz.f32 %f916, %f905, %f915;
mov.b32 %r959, %f916;
shfl.sync.bfly.b32 %r960|%p286, %r959, %r916, %r911, %r913;
mov.b32 %f917, %r960;
add.ftz.f32 %f918, %f916, %f917;
mov.b32 %r961, %f918;
shfl.sync.bfly.b32 %r962|%p287, %r961, %r919, %r911, %r913;
mov.b32 %f919, %r962;
add.ftz.f32 %f920, %f918, %f919;
mov.b32 %r963, %f920;
shfl.sync.bfly.b32 %r964|%p288, %r963, %r922, %r911, %r913;
mov.b32 %f921, %r964;
add.ftz.f32 %f922, %f920, %f921;
mov.b32 %r965, %f922;
shfl.sync.bfly.b32 %r966|%p289, %r965, %r925, %r911, %r913;
mov.b32 %f923, %r966;
add.ftz.f32 %f75, %f922, %f923;
setp.ne.s32 %p290, %r2, 1;
or.b32 %r970, %r4, 1;
setp.ge.u32 %p291, %r970, %r44;
or.pred %p292, %p290, %p291;
@%p292 bra $L__BB43_74;

add.ftz.f32 %f924, %f5, %f74;
add.ftz.f32 %f925, %f4, %f75;
mul.ftz.f32 %f926, %f924, %f925;
st.global.f32 [%rd23+4], %f926;

$L__BB43_74:
ld.shared.f32 %f927, [%r25+256];
add.ftz.f32 %f928, %f927, %f28;
ld.shared.f32 %f929, [%r27+256];
add.ftz.f32 %f930, %f929, %f32;
ld.shared.f32 %f931, [%r25+768];
add.ftz.f32 %f932, %f931, %f928;
ld.shared.f32 %f933, [%r27+768];
add.ftz.f32 %f934, %f933, %f930;
ld.shared.f32 %f935, [%r25+1280];
add.ftz.f32 %f936, %f935, %f932;
ld.shared.f32 %f937, [%r27+1280];
add.ftz.f32 %f938, %f937, %f934;
mov.b32 %r971, %f936;
mov.u32 %r972, 31;
mov.u32 %r973, 16;
mov.u32 %r974, -1;
shfl.sync.bfly.b32 %r975|%p293, %r971, %r973, %r972, %r974;
mov.b32 %f939, %r975;
add.ftz.f32 %f940, %f936, %f939;
mov.b32 %r976, %f940;
mov.u32 %r977, 8;
shfl.sync.bfly.b32 %r978|%p294, %r976, %r977, %r972, %r974;
mov.b32 %f941, %r978;
add.ftz.f32 %f942, %f940, %f941;
mov.b32 %r979, %f942;
mov.u32 %r980, 4;
shfl.sync.bfly.b32 %r981|%p295, %r979, %r980, %r972, %r974;
mov.b32 %f943, %r981;
add.ftz.f32 %f944, %f942, %f943;
mov.b32 %r982, %f944;
mov.u32 %r983, 2;
shfl.sync.bfly.b32 %r984|%p296, %r982, %r983, %r972, %r974;
mov.b32 %f945, %r984;
add.ftz.f32 %f946, %f944, %f945;
mov.b32 %r985, %f946;
mov.u32 %r986, 1;
shfl.sync.bfly.b32 %r987|%p297, %r985, %r986, %r972, %r974;
mov.b32 %f947, %r987;
add.ftz.f32 %f76, %f946, %f947;
mov.b32 %r988, %f938;
shfl.sync.bfly.b32 %r989|%p298, %r988, %r973, %r972, %r974;
mov.b32 %f948, %r989;
add.ftz.f32 %f949, %f938, %f948;
mov.b32 %r990, %f949;
shfl.sync.bfly.b32 %r991|%p299, %r990, %r977, %r972, %r974;
mov.b32 %f950, %r991;
add.ftz.f32 %f951, %f949, %f950;
mov.b32 %r992, %f951;
shfl.sync.bfly.b32 %r993|%p300, %r992, %r980, %r972, %r974;
mov.b32 %f952, %r993;
add.ftz.f32 %f953, %f951, %f952;
mov.b32 %r994, %f953;
shfl.sync.bfly.b32 %r995|%p301, %r994, %r983, %r972, %r974;
mov.b32 %f954, %r995;
add.ftz.f32 %f955, %f953, %f954;
mov.b32 %r996, %f955;
shfl.sync.bfly.b32 %r997|%p302, %r996, %r986, %r972, %r974;
mov.b32 %f956, %r997;
add.ftz.f32 %f77, %f955, %f956;
setp.ne.s32 %p303, %r2, 2;
or.b32 %r1001, %r4, 2;
setp.ge.u32 %p304, %r1001, %r44;
or.pred %p305, %p303, %p304;
@%p305 bra $L__BB43_76;

add.ftz.f32 %f957, %f5, %f76;
add.ftz.f32 %f958, %f4, %f77;
mul.ftz.f32 %f959, %f957, %f958;
st.global.f32 [%rd23+8], %f959;

$L__BB43_76:
ld.shared.f32 %f960, [%r25+384];
add.ftz.f32 %f961, %f960, %f27;
ld.shared.f32 %f962, [%r27+384];
add.ftz.f32 %f963, %f962, %f31;
ld.shared.f32 %f964, [%r25+896];
add.ftz.f32 %f965, %f964, %f961;
ld.shared.f32 %f966, [%r27+896];
add.ftz.f32 %f967, %f966, %f963;
ld.shared.f32 %f968, [%r25+1408];
add.ftz.f32 %f969, %f968, %f965;
ld.shared.f32 %f970, [%r27+1408];
add.ftz.f32 %f971, %f970, %f967;
mov.b32 %r1002, %f969;
shfl.sync.bfly.b32 %r1006|%p306, %r1002, %r973, %r972, %r974;
mov.b32 %f972, %r1006;
add.ftz.f32 %f973, %f969, %f972;
mov.b32 %r1007, %f973;
shfl.sync.bfly.b32 %r1009|%p307, %r1007, %r977, %r972, %r974;
mov.b32 %f974, %r1009;
add.ftz.f32 %f975, %f973, %f974;
mov.b32 %r1010, %f975;
shfl.sync.bfly.b32 %r1012|%p308, %r1010, %r980, %r972, %r974;
mov.b32 %f976, %r1012;
add.ftz.f32 %f977, %f975, %f976;
mov.b32 %r1013, %f977;
shfl.sync.bfly.b32 %r1015|%p309, %r1013, %r983, %r972, %r974;
mov.b32 %f978, %r1015;
add.ftz.f32 %f979, %f977, %f978;
mov.b32 %r1016, %f979;
shfl.sync.bfly.b32 %r1018|%p310, %r1016, %r986, %r972, %r974;
mov.b32 %f980, %r1018;
add.ftz.f32 %f78, %f979, %f980;
mov.b32 %r1019, %f971;
shfl.sync.bfly.b32 %r1020|%p311, %r1019, %r973, %r972, %r974;
mov.b32 %f981, %r1020;
add.ftz.f32 %f982, %f971, %f981;
mov.b32 %r1021, %f982;
shfl.sync.bfly.b32 %r1022|%p312, %r1021, %r977, %r972, %r974;
mov.b32 %f983, %r1022;
add.ftz.f32 %f984, %f982, %f983;
mov.b32 %r1023, %f984;
shfl.sync.bfly.b32 %r1024|%p313, %r1023, %r980, %r972, %r974;
mov.b32 %f985, %r1024;
add.ftz.f32 %f986, %f984, %f985;
mov.b32 %r1025, %f986;
shfl.sync.bfly.b32 %r1026|%p314, %r1025, %r983, %r972, %r974;
mov.b32 %f987, %r1026;
add.ftz.f32 %f988, %f986, %f987;
mov.b32 %r1027, %f988;
shfl.sync.bfly.b32 %r1028|%p315, %r1027, %r986, %r972, %r974;
mov.b32 %f989, %r1028;
add.ftz.f32 %f79, %f988, %f989;
setp.ne.s32 %p316, %r2, 3;
or.b32 %r1032, %r4, 3;
setp.ge.u32 %p317, %r1032, %r44;
or.pred %p318, %p316, %p317;
@%p318 bra $L__BB43_78;

add.ftz.f32 %f990, %f5, %f78;
add.ftz.f32 %f991, %f4, %f79;
mul.ftz.f32 %f992, %f990, %f991;
st.global.f32 [%rd23+12], %f992;
bra.uni $L__BB43_78;

$L__BB43_62:
ld.shared.f32 %f697, [%r25];
add.ftz.f32 %f698, %f697, %f30;
ld.shared.f32 %f699, [%r27];
add.ftz.f32 %f700, %f699, %f34;
ld.shared.f32 %f701, [%r25+512];
add.ftz.f32 %f702, %f701, %f698;
ld.shared.f32 %f703, [%r27+512];
add.ftz.f32 %f704, %f703, %f700;
ld.shared.f32 %f705, [%r25+1024];
add.ftz.f32 %f706, %f705, %f702;
ld.shared.f32 %f707, [%r27+1024];
add.ftz.f32 %f708, %f707, %f704;
mov.b32 %r787, %f706;
mov.u32 %r788, 31;
mov.u32 %r789, 16;
mov.u32 %r790, -1;
shfl.sync.bfly.b32 %r791|%p215, %r787, %r789, %r788, %r790;
mov.b32 %f709, %r791;
add.ftz.f32 %f710, %f706, %f709;
mov.b32 %r792, %f710;
mov.u32 %r793, 8;
shfl.sync.bfly.b32 %r794|%p216, %r792, %r793, %r788, %r790;
mov.b32 %f711, %r794;
add.ftz.f32 %f712, %f710, %f711;
mov.b32 %r795, %f712;
mov.u32 %r796, 4;
shfl.sync.bfly.b32 %r797|%p217, %r795, %r796, %r788, %r790;
mov.b32 %f713, %r797;
add.ftz.f32 %f714, %f712, %f713;
mov.b32 %r798, %f714;
mov.u32 %r799, 2;
shfl.sync.bfly.b32 %r800|%p218, %r798, %r799, %r788, %r790;
mov.b32 %f715, %r800;
add.ftz.f32 %f716, %f714, %f715;
mov.b32 %r801, %f716;
mov.u32 %r802, 1;
shfl.sync.bfly.b32 %r803|%p219, %r801, %r802, %r788, %r790;
mov.b32 %f717, %r803;
add.ftz.f32 %f64, %f716, %f717;
mov.b32 %r804, %f708;
shfl.sync.bfly.b32 %r805|%p220, %r804, %r789, %r788, %r790;
mov.b32 %f718, %r805;
add.ftz.f32 %f719, %f708, %f718;
mov.b32 %r806, %f719;
shfl.sync.bfly.b32 %r807|%p221, %r806, %r793, %r788, %r790;
mov.b32 %f720, %r807;
add.ftz.f32 %f721, %f719, %f720;
mov.b32 %r808, %f721;
shfl.sync.bfly.b32 %r809|%p222, %r808, %r796, %r788, %r790;
mov.b32 %f722, %r809;
add.ftz.f32 %f723, %f721, %f722;
mov.b32 %r810, %f723;
shfl.sync.bfly.b32 %r811|%p223, %r810, %r799, %r788, %r790;
mov.b32 %f724, %r811;
add.ftz.f32 %f725, %f723, %f724;
mov.b32 %r812, %f725;
shfl.sync.bfly.b32 %r813|%p224, %r812, %r802, %r788, %r790;
mov.b32 %f726, %r813;
add.ftz.f32 %f65, %f725, %f726;
setp.ne.s32 %p225, %r2, 0;
setp.ge.u32 %p226, %r4, %r44;
or.pred %p227, %p225, %p226;
@%p227 bra $L__BB43_64;

add.ftz.f32 %f727, %f5, %f64;
add.ftz.f32 %f728, %f4, %f65;
mul.ftz.f32 %f729, %f728, 0f3F000000;
mul.ftz.f32 %f730, %f728, 0f3F4C422A;
mul.ftz.f32 %f731, %f728, 0f3D372713;
fma.rn.ftz.f32 %f732, %f728, %f731, 0f3F800000;
mul.ftz.f32 %f733, %f730, %f732;
tanh.approx.f32 %f734, %f733;
add.ftz.f32 %f735, %f734, 0f3F800000;
mul.ftz.f32 %f736, %f729, %f735;
mul.ftz.f32 %f737, %f727, %f736;
st.global.f32 [%rd23], %f737;

$L__BB43_64:
ld.shared.f32 %f738, [%r25+128];
add.ftz.f32 %f739, %f738, %f29;
ld.shared.f32 %f740, [%r27+128];
add.ftz.f32 %f741, %f740, %f33;
ld.shared.f32 %f742, [%r25+640];
add.ftz.f32 %f743, %f742, %f739;
ld.shared.f32 %f744, [%r27+640];
add.ftz.f32 %f745, %f744, %f741;
ld.shared.f32 %f746, [%r25+1152];
add.ftz.f32 %f747, %f746, %f743;
ld.shared.f32 %f748, [%r27+1152];
add.ftz.f32 %f749, %f748, %f745;
mov.b32 %r817, %f747;
shfl.sync.bfly.b32 %r821|%p228, %r817, %r789, %r788, %r790;
mov.b32 %f750, %r821;
add.ftz.f32 %f751, %f747, %f750;
mov.b32 %r822, %f751;
shfl.sync.bfly.b32 %r824|%p229, %r822, %r793, %r788, %r790;
mov.b32 %f752, %r824;
add.ftz.f32 %f753, %f751, %f752;
mov.b32 %r825, %f753;
shfl.sync.bfly.b32 %r827|%p230, %r825, %r796, %r788, %r790;
mov.b32 %f754, %r827;
add.ftz.f32 %f755, %f753, %f754;
mov.b32 %r828, %f755;
shfl.sync.bfly.b32 %r830|%p231, %r828, %r799, %r788, %r790;
mov.b32 %f756, %r830;
add.ftz.f32 %f757, %f755, %f756;
mov.b32 %r831, %f757;
shfl.sync.bfly.b32 %r833|%p232, %r831, %r802, %r788, %r790;
mov.b32 %f758, %r833;
add.ftz.f32 %f66, %f757, %f758;
mov.b32 %r834, %f749;
shfl.sync.bfly.b32 %r835|%p233, %r834, %r789, %r788, %r790;
mov.b32 %f759, %r835;
add.ftz.f32 %f760, %f749, %f759;
mov.b32 %r836, %f760;
shfl.sync.bfly.b32 %r837|%p234, %r836, %r793, %r788, %r790;
mov.b32 %f761, %r837;
add.ftz.f32 %f762, %f760, %f761;
mov.b32 %r838, %f762;
shfl.sync.bfly.b32 %r839|%p235, %r838, %r796, %r788, %r790;
mov.b32 %f763, %r839;
add.ftz.f32 %f764, %f762, %f763;
mov.b32 %r840, %f764;
shfl.sync.bfly.b32 %r841|%p236, %r840, %r799, %r788, %r790;
mov.b32 %f765, %r841;
add.ftz.f32 %f766, %f764, %f765;
mov.b32 %r842, %f766;
shfl.sync.bfly.b32 %r843|%p237, %r842, %r802, %r788, %r790;
mov.b32 %f767, %r843;
add.ftz.f32 %f67, %f766, %f767;
setp.ne.s32 %p238, %r2, 1;
or.b32 %r847, %r4, 1;
setp.ge.u32 %p239, %r847, %r44;
or.pred %p240, %p238, %p239;
@%p240 bra $L__BB43_66;

add.ftz.f32 %f768, %f5, %f66;
add.ftz.f32 %f769, %f4, %f67;
mul.ftz.f32 %f770, %f769, 0f3F000000;
mul.ftz.f32 %f771, %f769, 0f3F4C422A;
mul.ftz.f32 %f772, %f769, 0f3D372713;
fma.rn.ftz.f32 %f773, %f769, %f772, 0f3F800000;
mul.ftz.f32 %f774, %f771, %f773;
tanh.approx.f32 %f775, %f774;
add.ftz.f32 %f776, %f775, 0f3F800000;
mul.ftz.f32 %f777, %f770, %f776;
mul.ftz.f32 %f778, %f768, %f777;
st.global.f32 [%rd23+4], %f778;

$L__BB43_66:
ld.shared.f32 %f779, [%r25+256];
add.ftz.f32 %f780, %f779, %f28;
ld.shared.f32 %f781, [%r27+256];
add.ftz.f32 %f782, %f781, %f32;
ld.shared.f32 %f783, [%r25+768];
add.ftz.f32 %f784, %f783, %f780;
ld.shared.f32 %f785, [%r27+768];
add.ftz.f32 %f786, %f785, %f782;
ld.shared.f32 %f787, [%r25+1280];
add.ftz.f32 %f788, %f787, %f784;
ld.shared.f32 %f789, [%r27+1280];
add.ftz.f32 %f790, %f789, %f786;
mov.b32 %r848, %f788;
mov.u32 %r849, 31;
mov.u32 %r850, 16;
mov.u32 %r851, -1;
shfl.sync.bfly.b32 %r852|%p241, %r848, %r850, %r849, %r851;
mov.b32 %f791, %r852;
add.ftz.f32 %f792, %f788, %f791;
mov.b32 %r853, %f792;
mov.u32 %r854, 8;
shfl.sync.bfly.b32 %r855|%p242, %r853, %r854, %r849, %r851;
mov.b32 %f793, %r855;
add.ftz.f32 %f794, %f792, %f793;
mov.b32 %r856, %f794;
mov.u32 %r857, 4;
shfl.sync.bfly.b32 %r858|%p243, %r856, %r857, %r849, %r851;
mov.b32 %f795, %r858;
add.ftz.f32 %f796, %f794, %f795;
mov.b32 %r859, %f796;
mov.u32 %r860, 2;
shfl.sync.bfly.b32 %r861|%p244, %r859, %r860, %r849, %r851;
mov.b32 %f797, %r861;
add.ftz.f32 %f798, %f796, %f797;
mov.b32 %r862, %f798;
mov.u32 %r863, 1;
shfl.sync.bfly.b32 %r864|%p245, %r862, %r863, %r849, %r851;
mov.b32 %f799, %r864;
add.ftz.f32 %f68, %f798, %f799;
mov.b32 %r865, %f790;
shfl.sync.bfly.b32 %r866|%p246, %r865, %r850, %r849, %r851;
mov.b32 %f800, %r866;
add.ftz.f32 %f801, %f790, %f800;
mov.b32 %r867, %f801;
shfl.sync.bfly.b32 %r868|%p247, %r867, %r854, %r849, %r851;
mov.b32 %f802, %r868;
add.ftz.f32 %f803, %f801, %f802;
mov.b32 %r869, %f803;
shfl.sync.bfly.b32 %r870|%p248, %r869, %r857, %r849, %r851;
mov.b32 %f804, %r870;
add.ftz.f32 %f805, %f803, %f804;
mov.b32 %r871, %f805;
shfl.sync.bfly.b32 %r872|%p249, %r871, %r860, %r849, %r851;
mov.b32 %f806, %r872;
add.ftz.f32 %f807, %f805, %f806;
mov.b32 %r873, %f807;
shfl.sync.bfly.b32 %r874|%p250, %r873, %r863, %r849, %r851;
mov.b32 %f808, %r874;
add.ftz.f32 %f69, %f807, %f808;
setp.ne.s32 %p251, %r2, 2;
or.b32 %r878, %r4, 2;
setp.ge.u32 %p252, %r878, %r44;
or.pred %p253, %p251, %p252;
@%p253 bra $L__BB43_68;

add.ftz.f32 %f809, %f5, %f68;
add.ftz.f32 %f810, %f4, %f69;
mul.ftz.f32 %f811, %f810, 0f3F000000;
mul.ftz.f32 %f812, %f810, 0f3F4C422A;
mul.ftz.f32 %f813, %f810, 0f3D372713;
fma.rn.ftz.f32 %f814, %f810, %f813, 0f3F800000;
mul.ftz.f32 %f815, %f812, %f814;
tanh.approx.f32 %f816, %f815;
add.ftz.f32 %f817, %f816, 0f3F800000;
mul.ftz.f32 %f818, %f811, %f817;
mul.ftz.f32 %f819, %f809, %f818;
st.global.f32 [%rd23+8], %f819;

$L__BB43_68:
ld.shared.f32 %f820, [%r25+384];
add.ftz.f32 %f821, %f820, %f27;
ld.shared.f32 %f822, [%r27+384];
add.ftz.f32 %f823, %f822, %f31;
ld.shared.f32 %f824, [%r25+896];
add.ftz.f32 %f825, %f824, %f821;
ld.shared.f32 %f826, [%r27+896];
add.ftz.f32 %f827, %f826, %f823;
ld.shared.f32 %f828, [%r25+1408];
add.ftz.f32 %f829, %f828, %f825;
ld.shared.f32 %f830, [%r27+1408];
add.ftz.f32 %f831, %f830, %f827;
mov.b32 %r879, %f829;
shfl.sync.bfly.b32 %r883|%p254, %r879, %r850, %r849, %r851;
mov.b32 %f832, %r883;
add.ftz.f32 %f833, %f829, %f832;
mov.b32 %r884, %f833;
shfl.sync.bfly.b32 %r886|%p255, %r884, %r854, %r849, %r851;
mov.b32 %f834, %r886;
add.ftz.f32 %f835, %f833, %f834;
mov.b32 %r887, %f835;
shfl.sync.bfly.b32 %r889|%p256, %r887, %r857, %r849, %r851;
mov.b32 %f836, %r889;
add.ftz.f32 %f837, %f835, %f836;
mov.b32 %r890, %f837;
shfl.sync.bfly.b32 %r892|%p257, %r890, %r860, %r849, %r851;
mov.b32 %f838, %r892;
add.ftz.f32 %f839, %f837, %f838;
mov.b32 %r893, %f839;
shfl.sync.bfly.b32 %r895|%p258, %r893, %r863, %r849, %r851;
mov.b32 %f840, %r895;
add.ftz.f32 %f70, %f839, %f840;
mov.b32 %r896, %f831;
shfl.sync.bfly.b32 %r897|%p259, %r896, %r850, %r849, %r851;
mov.b32 %f841, %r897;
add.ftz.f32 %f842, %f831, %f841;
mov.b32 %r898, %f842;
shfl.sync.bfly.b32 %r899|%p260, %r898, %r854, %r849, %r851;
mov.b32 %f843, %r899;
add.ftz.f32 %f844, %f842, %f843;
mov.b32 %r900, %f844;
shfl.sync.bfly.b32 %r901|%p261, %r900, %r857, %r849, %r851;
mov.b32 %f845, %r901;
add.ftz.f32 %f846, %f844, %f845;
mov.b32 %r902, %f846;
shfl.sync.bfly.b32 %r903|%p262, %r902, %r860, %r849, %r851;
mov.b32 %f847, %r903;
add.ftz.f32 %f848, %f846, %f847;
mov.b32 %r904, %f848;
shfl.sync.bfly.b32 %r905|%p263, %r904, %r863, %r849, %r851;
mov.b32 %f849, %r905;
add.ftz.f32 %f71, %f848, %f849;
setp.ne.s32 %p264, %r2, 3;
or.b32 %r909, %r4, 3;
setp.ge.u32 %p265, %r909, %r44;
or.pred %p266, %p264, %p265;
@%p266 bra $L__BB43_78;

add.ftz.f32 %f850, %f5, %f70;
add.ftz.f32 %f851, %f4, %f71;
mul.ftz.f32 %f852, %f851, 0f3F000000;
mul.ftz.f32 %f853, %f851, 0f3F4C422A;
mul.ftz.f32 %f854, %f851, 0f3D372713;
fma.rn.ftz.f32 %f855, %f851, %f854, 0f3F800000;
mul.ftz.f32 %f856, %f853, %f855;
tanh.approx.f32 %f857, %f856;
add.ftz.f32 %f858, %f857, 0f3F800000;
mul.ftz.f32 %f859, %f852, %f858;
mul.ftz.f32 %f860, %f850, %f859;
st.global.f32 [%rd23+12], %f860;
bra.uni $L__BB43_78;

$L__BB43_54:
ld.shared.f32 %f373, [%r25];
add.ftz.f32 %f374, %f373, %f30;
ld.shared.f32 %f375, [%r27];
add.ftz.f32 %f376, %f375, %f34;
ld.shared.f32 %f377, [%r25+512];
add.ftz.f32 %f378, %f377, %f374;
ld.shared.f32 %f379, [%r27+512];
add.ftz.f32 %f380, %f379, %f376;
ld.shared.f32 %f381, [%r25+1024];
add.ftz.f32 %f382, %f381, %f378;
ld.shared.f32 %f383, [%r27+1024];
add.ftz.f32 %f384, %f383, %f380;
mov.b32 %r541, %f382;
mov.u32 %r542, 31;
mov.u32 %r543, 16;
mov.u32 %r544, -1;
shfl.sync.bfly.b32 %r545|%p111, %r541, %r543, %r542, %r544;
mov.b32 %f385, %r545;
add.ftz.f32 %f386, %f382, %f385;
mov.b32 %r546, %f386;
mov.u32 %r547, 8;
shfl.sync.bfly.b32 %r548|%p112, %r546, %r547, %r542, %r544;
mov.b32 %f387, %r548;
add.ftz.f32 %f388, %f386, %f387;
mov.b32 %r549, %f388;
mov.u32 %r550, 4;
shfl.sync.bfly.b32 %r551|%p113, %r549, %r550, %r542, %r544;
mov.b32 %f389, %r551;
add.ftz.f32 %f390, %f388, %f389;
mov.b32 %r552, %f390;
mov.u32 %r553, 2;
shfl.sync.bfly.b32 %r554|%p114, %r552, %r553, %r542, %r544;
mov.b32 %f391, %r554;
add.ftz.f32 %f392, %f390, %f391;
mov.b32 %r555, %f392;
mov.u32 %r556, 1;
shfl.sync.bfly.b32 %r557|%p115, %r555, %r556, %r542, %r544;
mov.b32 %f393, %r557;
add.ftz.f32 %f48, %f392, %f393;
mov.b32 %r558, %f384;
shfl.sync.bfly.b32 %r559|%p116, %r558, %r543, %r542, %r544;
mov.b32 %f394, %r559;
add.ftz.f32 %f395, %f384, %f394;
mov.b32 %r560, %f395;
shfl.sync.bfly.b32 %r561|%p117, %r560, %r547, %r542, %r544;
mov.b32 %f396, %r561;
add.ftz.f32 %f397, %f395, %f396;
mov.b32 %r562, %f397;
shfl.sync.bfly.b32 %r563|%p118, %r562, %r550, %r542, %r544;
mov.b32 %f398, %r563;
add.ftz.f32 %f399, %f397, %f398;
mov.b32 %r564, %f399;
shfl.sync.bfly.b32 %r565|%p119, %r564, %r553, %r542, %r544;
mov.b32 %f400, %r565;
add.ftz.f32 %f401, %f399, %f400;
mov.b32 %r566, %f401;
shfl.sync.bfly.b32 %r567|%p120, %r566, %r556, %r542, %r544;
mov.b32 %f402, %r567;
add.ftz.f32 %f49, %f401, %f402;
setp.ne.s32 %p121, %r2, 0;
setp.ge.u32 %p122, %r4, %r44;
or.pred %p123, %p121, %p122;
@%p123 bra $L__BB43_56;

add.ftz.f32 %f403, %f5, %f48;
add.ftz.f32 %f404, %f4, %f49;
mov.f32 %f405, 0f40E00000;
min.ftz.f32 %f406, %f404, %f405;
min.ftz.f32 %f407, %f403, %f405;
mov.f32 %f408, 0fC0E00000;
max.ftz.f32 %f409, %f407, %f408;
mul.ftz.f32 %f410, %f406, 0f3FD9DB23;
mul.ftz.f32 %f411, %f410, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f412, %f411;
add.ftz.f32 %f413, %f412, 0f3F800000;
div.approx.ftz.f32 %f414, %f406, %f413;
add.ftz.f32 %f415, %f409, 0f3F800000;
mul.ftz.f32 %f416, %f415, %f414;
st.global.f32 [%rd23], %f416;

$L__BB43_56:
ld.shared.f32 %f417, [%r25+128];
add.ftz.f32 %f418, %f417, %f29;
ld.shared.f32 %f419, [%r27+128];
add.ftz.f32 %f420, %f419, %f33;
ld.shared.f32 %f421, [%r25+640];
add.ftz.f32 %f422, %f421, %f418;
ld.shared.f32 %f423, [%r27+640];
add.ftz.f32 %f424, %f423, %f420;
ld.shared.f32 %f425, [%r25+1152];
add.ftz.f32 %f426, %f425, %f422;
ld.shared.f32 %f427, [%r27+1152];
add.ftz.f32 %f428, %f427, %f424;
mov.b32 %r571, %f426;
shfl.sync.bfly.b32 %r575|%p124, %r571, %r543, %r542, %r544;
mov.b32 %f429, %r575;
add.ftz.f32 %f430, %f426, %f429;
mov.b32 %r576, %f430;
shfl.sync.bfly.b32 %r578|%p125, %r576, %r547, %r542, %r544;
mov.b32 %f431, %r578;
add.ftz.f32 %f432, %f430, %f431;
mov.b32 %r579, %f432;
shfl.sync.bfly.b32 %r581|%p126, %r579, %r550, %r542, %r544;
mov.b32 %f433, %r581;
add.ftz.f32 %f434, %f432, %f433;
mov.b32 %r582, %f434;
shfl.sync.bfly.b32 %r584|%p127, %r582, %r553, %r542, %r544;
mov.b32 %f435, %r584;
add.ftz.f32 %f436, %f434, %f435;
mov.b32 %r585, %f436;
shfl.sync.bfly.b32 %r587|%p128, %r585, %r556, %r542, %r544;
mov.b32 %f437, %r587;
add.ftz.f32 %f50, %f436, %f437;
mov.b32 %r588, %f428;
shfl.sync.bfly.b32 %r589|%p129, %r588, %r543, %r542, %r544;
mov.b32 %f438, %r589;
add.ftz.f32 %f439, %f428, %f438;
mov.b32 %r590, %f439;
shfl.sync.bfly.b32 %r591|%p130, %r590, %r547, %r542, %r544;
mov.b32 %f440, %r591;
add.ftz.f32 %f441, %f439, %f440;
mov.b32 %r592, %f441;
shfl.sync.bfly.b32 %r593|%p131, %r592, %r550, %r542, %r544;
mov.b32 %f442, %r593;
add.ftz.f32 %f443, %f441, %f442;
mov.b32 %r594, %f443;
shfl.sync.bfly.b32 %r595|%p132, %r594, %r553, %r542, %r544;
mov.b32 %f444, %r595;
add.ftz.f32 %f445, %f443, %f444;
mov.b32 %r596, %f445;
shfl.sync.bfly.b32 %r597|%p133, %r596, %r556, %r542, %r544;
mov.b32 %f446, %r597;
add.ftz.f32 %f51, %f445, %f446;
setp.ne.s32 %p134, %r2, 1;
or.b32 %r601, %r4, 1;
setp.ge.u32 %p135, %r601, %r44;
or.pred %p136, %p134, %p135;
@%p136 bra $L__BB43_58;

add.ftz.f32 %f447, %f5, %f50;
add.ftz.f32 %f448, %f4, %f51;
mov.f32 %f449, 0f40E00000;
min.ftz.f32 %f450, %f448, %f449;
min.ftz.f32 %f451, %f447, %f449;
mov.f32 %f452, 0fC0E00000;
max.ftz.f32 %f453, %f451, %f452;
mul.ftz.f32 %f454, %f450, 0f3FD9DB23;
mul.ftz.f32 %f455, %f454, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f456, %f455;
add.ftz.f32 %f457, %f456, 0f3F800000;
div.approx.ftz.f32 %f458, %f450, %f457;
add.ftz.f32 %f459, %f453, 0f3F800000;
mul.ftz.f32 %f460, %f459, %f458;
st.global.f32 [%rd23+4], %f460;

$L__BB43_58:
ld.shared.f32 %f461, [%r25+256];
add.ftz.f32 %f462, %f461, %f28;
ld.shared.f32 %f463, [%r27+256];
add.ftz.f32 %f464, %f463, %f32;
ld.shared.f32 %f465, [%r25+768];
add.ftz.f32 %f466, %f465, %f462;
ld.shared.f32 %f467, [%r27+768];
add.ftz.f32 %f468, %f467, %f464;
ld.shared.f32 %f469, [%r25+1280];
add.ftz.f32 %f470, %f469, %f466;
ld.shared.f32 %f471, [%r27+1280];
add.ftz.f32 %f472, %f471, %f468;
mov.b32 %r602, %f470;
mov.u32 %r603, 31;
mov.u32 %r604, 16;
mov.u32 %r605, -1;
shfl.sync.bfly.b32 %r606|%p137, %r602, %r604, %r603, %r605;
mov.b32 %f473, %r606;
add.ftz.f32 %f474, %f470, %f473;
mov.b32 %r607, %f474;
mov.u32 %r608, 8;
shfl.sync.bfly.b32 %r609|%p138, %r607, %r608, %r603, %r605;
mov.b32 %f475, %r609;
add.ftz.f32 %f476, %f474, %f475;
mov.b32 %r610, %f476;
mov.u32 %r611, 4;
shfl.sync.bfly.b32 %r612|%p139, %r610, %r611, %r603, %r605;
mov.b32 %f477, %r612;
add.ftz.f32 %f478, %f476, %f477;
mov.b32 %r613, %f478;
mov.u32 %r614, 2;
shfl.sync.bfly.b32 %r615|%p140, %r613, %r614, %r603, %r605;
mov.b32 %f479, %r615;
add.ftz.f32 %f480, %f478, %f479;
mov.b32 %r616, %f480;
mov.u32 %r617, 1;
shfl.sync.bfly.b32 %r618|%p141, %r616, %r617, %r603, %r605;
mov.b32 %f481, %r618;
add.ftz.f32 %f52, %f480, %f481;
mov.b32 %r619, %f472;
shfl.sync.bfly.b32 %r620|%p142, %r619, %r604, %r603, %r605;
mov.b32 %f482, %r620;
add.ftz.f32 %f483, %f472, %f482;
mov.b32 %r621, %f483;
shfl.sync.bfly.b32 %r622|%p143, %r621, %r608, %r603, %r605;
mov.b32 %f484, %r622;
add.ftz.f32 %f485, %f483, %f484;
mov.b32 %r623, %f485;
shfl.sync.bfly.b32 %r624|%p144, %r623, %r611, %r603, %r605;
mov.b32 %f486, %r624;
add.ftz.f32 %f487, %f485, %f486;
mov.b32 %r625, %f487;
shfl.sync.bfly.b32 %r626|%p145, %r625, %r614, %r603, %r605;
mov.b32 %f488, %r626;
add.ftz.f32 %f489, %f487, %f488;
mov.b32 %r627, %f489;
shfl.sync.bfly.b32 %r628|%p146, %r627, %r617, %r603, %r605;
mov.b32 %f490, %r628;
add.ftz.f32 %f53, %f489, %f490;
setp.ne.s32 %p147, %r2, 2;
or.b32 %r632, %r4, 2;
setp.ge.u32 %p148, %r632, %r44;
or.pred %p149, %p147, %p148;
@%p149 bra $L__BB43_60;

add.ftz.f32 %f491, %f5, %f52;
add.ftz.f32 %f492, %f4, %f53;
mov.f32 %f493, 0f40E00000;
min.ftz.f32 %f494, %f492, %f493;
min.ftz.f32 %f495, %f491, %f493;
mov.f32 %f496, 0fC0E00000;
max.ftz.f32 %f497, %f495, %f496;
mul.ftz.f32 %f498, %f494, 0f3FD9DB23;
mul.ftz.f32 %f499, %f498, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f500, %f499;
add.ftz.f32 %f501, %f500, 0f3F800000;
div.approx.ftz.f32 %f502, %f494, %f501;
add.ftz.f32 %f503, %f497, 0f3F800000;
mul.ftz.f32 %f504, %f503, %f502;
st.global.f32 [%rd23+8], %f504;

$L__BB43_60:
ld.shared.f32 %f505, [%r25+384];
add.ftz.f32 %f506, %f505, %f27;
ld.shared.f32 %f507, [%r27+384];
add.ftz.f32 %f508, %f507, %f31;
ld.shared.f32 %f509, [%r25+896];
add.ftz.f32 %f510, %f509, %f506;
ld.shared.f32 %f511, [%r27+896];
add.ftz.f32 %f512, %f511, %f508;
ld.shared.f32 %f513, [%r25+1408];
add.ftz.f32 %f514, %f513, %f510;
ld.shared.f32 %f515, [%r27+1408];
add.ftz.f32 %f516, %f515, %f512;
mov.b32 %r633, %f514;
shfl.sync.bfly.b32 %r637|%p150, %r633, %r604, %r603, %r605;
mov.b32 %f517, %r637;
add.ftz.f32 %f518, %f514, %f517;
mov.b32 %r638, %f518;
shfl.sync.bfly.b32 %r640|%p151, %r638, %r608, %r603, %r605;
mov.b32 %f519, %r640;
add.ftz.f32 %f520, %f518, %f519;
mov.b32 %r641, %f520;
shfl.sync.bfly.b32 %r643|%p152, %r641, %r611, %r603, %r605;
mov.b32 %f521, %r643;
add.ftz.f32 %f522, %f520, %f521;
mov.b32 %r644, %f522;
shfl.sync.bfly.b32 %r646|%p153, %r644, %r614, %r603, %r605;
mov.b32 %f523, %r646;
add.ftz.f32 %f524, %f522, %f523;
mov.b32 %r647, %f524;
shfl.sync.bfly.b32 %r649|%p154, %r647, %r617, %r603, %r605;
mov.b32 %f525, %r649;
add.ftz.f32 %f54, %f524, %f525;
mov.b32 %r650, %f516;
shfl.sync.bfly.b32 %r651|%p155, %r650, %r604, %r603, %r605;
mov.b32 %f526, %r651;
add.ftz.f32 %f527, %f516, %f526;
mov.b32 %r652, %f527;
shfl.sync.bfly.b32 %r653|%p156, %r652, %r608, %r603, %r605;
mov.b32 %f528, %r653;
add.ftz.f32 %f529, %f527, %f528;
mov.b32 %r654, %f529;
shfl.sync.bfly.b32 %r655|%p157, %r654, %r611, %r603, %r605;
mov.b32 %f530, %r655;
add.ftz.f32 %f531, %f529, %f530;
mov.b32 %r656, %f531;
shfl.sync.bfly.b32 %r657|%p158, %r656, %r614, %r603, %r605;
mov.b32 %f532, %r657;
add.ftz.f32 %f533, %f531, %f532;
mov.b32 %r658, %f533;
shfl.sync.bfly.b32 %r659|%p159, %r658, %r617, %r603, %r605;
mov.b32 %f534, %r659;
add.ftz.f32 %f55, %f533, %f534;
setp.ne.s32 %p160, %r2, 3;
or.b32 %r663, %r4, 3;
setp.ge.u32 %p161, %r663, %r44;
or.pred %p162, %p160, %p161;
@%p162 bra $L__BB43_78;

add.ftz.f32 %f535, %f5, %f54;
add.ftz.f32 %f536, %f4, %f55;
mov.f32 %f537, 0f40E00000;
min.ftz.f32 %f538, %f536, %f537;
min.ftz.f32 %f539, %f535, %f537;
mov.f32 %f540, 0fC0E00000;
max.ftz.f32 %f541, %f539, %f540;
mul.ftz.f32 %f542, %f538, 0f3FD9DB23;
mul.ftz.f32 %f543, %f542, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f544, %f543;
add.ftz.f32 %f545, %f544, 0f3F800000;
div.approx.ftz.f32 %f546, %f538, %f545;
add.ftz.f32 %f547, %f541, 0f3F800000;
mul.ftz.f32 %f548, %f547, %f546;
st.global.f32 [%rd23+12], %f548;
bra.uni $L__BB43_78;

}
.entry _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<38>;
.reg .f32 %f<114>;
.reg .b32 %r<244>;
.reg .b64 %rd<40>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

ld.param.u64 %rd11, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd12, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd13, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd14, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r17, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r19, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r18, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
mov.u32 %r1, %tid.y;
shl.b32 %r35, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r35, %r2;
mov.u32 %r36, %ctaid.x;
shl.b32 %r4, %r36, 2;
mov.u32 %r242, %ctaid.y;
setp.eq.s64 %p1, %rd13, 0;
@%p1 bra $L__BB44_2;

cvta.to.global.u64 %rd15, %rd13;
mul.wide.u32 %rd16, %r242, 4;
add.s64 %rd17, %rd15, %rd16;
ld.global.nc.u32 %r241, [%rd17];
mul.hi.u32 %r37, %r242, %r18;
add.s32 %r38, %r37, %r242;
shr.u32 %r39, %r38, %r19;
mul.lo.s32 %r40, %r39, %r20;
sub.s32 %r242, %r242, %r40;
bra.uni $L__BB44_3;

$L__BB44_2:
mul.hi.u32 %r41, %r242, %r23;
add.s32 %r42, %r41, %r242;
shr.u32 %r241, %r42, %r24;

$L__BB44_3:
mov.u32 %r11, %ctaid.z;
shr.u32 %r43, %r3, 31;
add.s32 %r44, %r3, %r43;
shr.s32 %r243, %r44, 1;
shr.u32 %r45, %r17, 5;
setp.ge.s32 %p2, %r243, %r45;
mov.f32 %f110, 0f00000000;
mov.f32 %f111, %f110;
mov.f32 %f112, %f110;
mov.f32 %f113, %f110;
@%p2 bra $L__BB44_6;

and.b32 %r48, %r44, 2147483646;
sub.s32 %r49, %r3, %r48;
shl.b32 %r50, %r49, 1;
cvt.s64.s32 %rd1, %r50;
mad.lo.s32 %r51, %r4, %r21, %r243;
mul.hi.u32 %r52, %r11, %r29;
add.s32 %r53, %r52, %r11;
shr.u32 %r54, %r53, %r30;
mad.lo.s32 %r55, %r54, %r32, %r51;
mad.lo.s32 %r56, %r241, %r26, %r55;
mul.wide.s32 %rd2, %r56, 20;
mul.wide.s32 %rd18, %r50, 4;
add.s64 %rd19, %rd2, %rd18;
add.s64 %rd3, %rd19, 8;
mul.wide.s32 %rd4, %r21, 20;
cvta.to.global.u64 %rd39, %rd11;
cvta.to.global.u64 %rd6, %rd12;
mul.lo.s32 %r57, %r242, %r27;
cvt.u64.u32 %rd20, %r57;
mul.lo.s32 %r58, %r11, %r33;
cvt.u64.u32 %rd21, %r58;
add.s64 %rd7, %rd20, %rd21;
mov.f32 %f110, 0f00000000;
shl.b64 %rd26, %rd1, 2;

$L__BB44_5:
cvt.u64.u32 %rd22, %r243;
add.s64 %rd23, %rd7, %rd22;
mul.lo.s64 %rd24, %rd23, 36;
add.s64 %rd25, %rd6, %rd24;
add.s64 %rd27, %rd25, %rd26;
add.s64 %rd28, %rd39, %rd3;
add.s64 %rd29, %rd28, -4;
ld.global.nc.u32 %r61, [%rd27+4];
ld.global.nc.u32 %r65, [%rd27+20];
ld.global.nc.u32 %r143, [%rd28];
ld.global.nc.u32 %r69, [%rd27+8];
ld.global.nc.u32 %r73, [%rd27+24];
ld.global.nc.u32 %r144, [%rd28+-4];
and.b32 %r60, %r144, 252645135;
shr.u32 %r145, %r144, 4;
and.b32 %r64, %r145, 252645135;
mov.u32 %r125, 0;

dp4a.s32.s32 %r59, %r60, %r61, %r125;


dp4a.s32.s32 %r63, %r64, %r65, %r59;

and.b32 %r68, %r143, 252645135;
shr.u32 %r146, %r143, 4;
and.b32 %r72, %r146, 252645135;

dp4a.s32.s32 %r67, %r68, %r69, %r63;


dp4a.s32.s32 %r71, %r72, %r73, %r67;

add.s64 %rd30, %rd39, %rd2;
ld.global.nc.u32 %r76, [%rd30];
ld.global.nc.u32 %r77, [%rd25];

{mul.f16x2 %r75,%r76,%r77;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r75;
cvt.f32.f16 %f25, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r75;
cvt.f32.f16 %f26, high;}


cvt.rn.f32.s32 %f33, %r71;
mov.f32 %f34, 0f40000000;
div.approx.ftz.f32 %f35, %f26, %f34;
fma.rn.ftz.f32 %f36, %f25, %f33, %f35;
add.ftz.f32 %f113, %f36, %f113;
add.s64 %rd31, %rd29, %rd4;
ld.global.nc.u32 %r147, [%rd31+4];
ld.global.nc.u32 %r148, [%rd31];
and.b32 %r81, %r148, 252645135;
shr.u32 %r149, %r148, 4;
and.b32 %r85, %r149, 252645135;

dp4a.s32.s32 %r80, %r81, %r61, %r125;


dp4a.s32.s32 %r84, %r85, %r65, %r80;

and.b32 %r89, %r147, 252645135;
shr.u32 %r150, %r147, 4;
and.b32 %r93, %r150, 252645135;

dp4a.s32.s32 %r88, %r89, %r69, %r84;


dp4a.s32.s32 %r92, %r93, %r73, %r88;

add.s64 %rd32, %rd30, %rd4;
ld.global.nc.u32 %r97, [%rd32];

{mul.f16x2 %r96,%r97,%r77;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r96;
cvt.f32.f16 %f27, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r96;
cvt.f32.f16 %f28, high;}


cvt.rn.f32.s32 %f37, %r92;
div.approx.ftz.f32 %f38, %f28, %f34;
fma.rn.ftz.f32 %f39, %f27, %f37, %f38;
add.ftz.f32 %f112, %f39, %f112;
add.s64 %rd33, %rd31, %rd4;
ld.global.nc.u32 %r151, [%rd33+4];
ld.global.nc.u32 %r152, [%rd33];
and.b32 %r102, %r152, 252645135;
shr.u32 %r153, %r152, 4;
and.b32 %r106, %r153, 252645135;

dp4a.s32.s32 %r101, %r102, %r61, %r125;


dp4a.s32.s32 %r105, %r106, %r65, %r101;

and.b32 %r110, %r151, 252645135;
shr.u32 %r154, %r151, 4;
and.b32 %r114, %r154, 252645135;

dp4a.s32.s32 %r109, %r110, %r69, %r105;


dp4a.s32.s32 %r113, %r114, %r73, %r109;

add.s64 %rd34, %rd32, %rd4;
ld.global.nc.u32 %r118, [%rd34];

{mul.f16x2 %r117,%r118,%r77;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r117;
cvt.f32.f16 %f29, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r117;
cvt.f32.f16 %f30, high;}


cvt.rn.f32.s32 %f40, %r113;
div.approx.ftz.f32 %f41, %f30, %f34;
fma.rn.ftz.f32 %f42, %f29, %f40, %f41;
add.ftz.f32 %f111, %f42, %f111;
add.s64 %rd35, %rd33, %rd4;
ld.global.nc.u32 %r155, [%rd35+4];
ld.global.nc.u32 %r156, [%rd35];
and.b32 %r123, %r156, 252645135;
shr.u32 %r157, %r156, 4;
and.b32 %r127, %r157, 252645135;

dp4a.s32.s32 %r122, %r123, %r61, %r125;


dp4a.s32.s32 %r126, %r127, %r65, %r122;

and.b32 %r131, %r155, 252645135;
shr.u32 %r158, %r155, 4;
and.b32 %r135, %r158, 252645135;

dp4a.s32.s32 %r130, %r131, %r69, %r126;


dp4a.s32.s32 %r134, %r135, %r73, %r130;

add.s64 %rd36, %rd34, %rd4;
ld.global.nc.u32 %r139, [%rd36];

{mul.f16x2 %r138,%r139,%r77;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r138;
cvt.f32.f16 %f31, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r138;
cvt.f32.f16 %f32, high;}


cvt.rn.f32.s32 %f43, %r134;
div.approx.ftz.f32 %f44, %f32, %f34;
fma.rn.ftz.f32 %f45, %f31, %f43, %f44;
add.ftz.f32 %f110, %f45, %f110;
add.s64 %rd39, %rd39, 1280;
add.s32 %r243, %r243, 64;
setp.lt.s32 %p3, %r243, %r45;
@%p3 bra $L__BB44_5;

$L__BB44_6:
setp.eq.s32 %p4, %r1, 0;
@%p4 bra $L__BB44_8;

shl.b32 %r160, %r1, 9;
mov.u32 %r161, _ZZ13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r162, %r161, %r160;
shl.b32 %r163, %r2, 2;
add.s32 %r164, %r162, %r163;
st.shared.f32 [%r164+-512], %f113;
st.shared.f32 [%r164+-384], %f112;
st.shared.f32 [%r164+-256], %f111;
st.shared.f32 [%r164+-128], %f110;

$L__BB44_8:
setp.ne.s32 %p5, %r1, 0;
bar.sync 0;
@%p5 bra $L__BB44_17;

mov.u32 %r165, %ctaid.y;
mad.lo.s32 %r166, %r165, %r28, %r4;
mad.lo.s32 %r167, %r11, %r34, %r166;
shl.b32 %r168, %r2, 2;
mov.u32 %r169, _ZZ13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r15, %r169, %r168;
mov.u32 %r170, 2;
ld.shared.f32 %f46, [%r15];
add.ftz.f32 %f47, %f46, %f113;
ld.shared.f32 %f48, [%r15+512];
add.ftz.f32 %f49, %f48, %f47;
ld.shared.f32 %f50, [%r15+1024];
add.ftz.f32 %f51, %f50, %f49;
mov.b32 %r171, %f51;
mov.u32 %r172, 31;
mov.u32 %r173, 16;
mov.u32 %r174, -1;
shfl.sync.bfly.b32 %r175|%p6, %r171, %r173, %r172, %r174;
mov.b32 %f52, %r175;
add.ftz.f32 %f53, %f51, %f52;
mov.b32 %r176, %f53;
mov.u32 %r177, 8;
shfl.sync.bfly.b32 %r178|%p7, %r176, %r177, %r172, %r174;
mov.b32 %f54, %r178;
add.ftz.f32 %f55, %f53, %f54;
mov.b32 %r179, %f55;
mov.u32 %r180, 4;
shfl.sync.bfly.b32 %r181|%p8, %r179, %r180, %r172, %r174;
mov.b32 %f56, %r181;
add.ftz.f32 %f57, %f55, %f56;
mov.b32 %r182, %f57;
shfl.sync.bfly.b32 %r183|%p9, %r182, %r170, %r172, %r174;
mov.b32 %f58, %r183;
add.ftz.f32 %f59, %f57, %f58;
mov.b32 %r184, %f59;
mov.u32 %r185, 1;
shfl.sync.bfly.b32 %r186|%p10, %r184, %r185, %r172, %r174;
mov.b32 %f60, %r186;
add.ftz.f32 %f13, %f59, %f60;
setp.ge.u32 %p11, %r4, %r22;
setp.ne.s32 %p12, %r2, 0;
cvta.to.global.u64 %rd37, %rd14;
mul.wide.u32 %rd38, %r167, 4;
add.s64 %rd10, %rd37, %rd38;
or.pred %p13, %p12, %p11;
@%p13 bra $L__BB44_11;

st.global.f32 [%rd10], %f13;

$L__BB44_11:
ld.shared.f32 %f61, [%r15+128];
add.ftz.f32 %f62, %f61, %f112;
ld.shared.f32 %f63, [%r15+640];
add.ftz.f32 %f64, %f63, %f62;
ld.shared.f32 %f65, [%r15+1152];
add.ftz.f32 %f66, %f65, %f64;
mov.b32 %r187, %f66;
shfl.sync.bfly.b32 %r191|%p14, %r187, %r173, %r172, %r174;
mov.b32 %f67, %r191;
add.ftz.f32 %f68, %f66, %f67;
mov.b32 %r192, %f68;
shfl.sync.bfly.b32 %r194|%p15, %r192, %r177, %r172, %r174;
mov.b32 %f69, %r194;
add.ftz.f32 %f70, %f68, %f69;
mov.b32 %r195, %f70;
shfl.sync.bfly.b32 %r197|%p16, %r195, %r180, %r172, %r174;
mov.b32 %f71, %r197;
add.ftz.f32 %f72, %f70, %f71;
mov.b32 %r198, %f72;
shfl.sync.bfly.b32 %r200|%p17, %r198, %r170, %r172, %r174;
mov.b32 %f73, %r200;
add.ftz.f32 %f74, %f72, %f73;
mov.b32 %r201, %f74;
shfl.sync.bfly.b32 %r203|%p18, %r201, %r185, %r172, %r174;
mov.b32 %f75, %r203;
add.ftz.f32 %f14, %f74, %f75;
add.s32 %r204, %r4, 1;
setp.ge.u32 %p19, %r204, %r22;
setp.ne.s32 %p20, %r2, 1;
or.pred %p21, %p20, %p19;
@%p21 bra $L__BB44_13;

st.global.f32 [%rd10+4], %f14;

$L__BB44_13:
ld.shared.f32 %f76, [%r15+256];
add.ftz.f32 %f77, %f76, %f111;
ld.shared.f32 %f78, [%r15+768];
add.ftz.f32 %f79, %f78, %f77;
ld.shared.f32 %f80, [%r15+1280];
add.ftz.f32 %f81, %f80, %f79;
mov.b32 %r205, %f81;
mov.u32 %r206, 31;
mov.u32 %r207, 16;
mov.u32 %r208, -1;
shfl.sync.bfly.b32 %r209|%p22, %r205, %r207, %r206, %r208;
mov.b32 %f82, %r209;
add.ftz.f32 %f83, %f81, %f82;
mov.b32 %r210, %f83;
mov.u32 %r211, 8;
shfl.sync.bfly.b32 %r212|%p23, %r210, %r211, %r206, %r208;
mov.b32 %f84, %r212;
add.ftz.f32 %f85, %f83, %f84;
mov.b32 %r213, %f85;
mov.u32 %r214, 4;
shfl.sync.bfly.b32 %r215|%p24, %r213, %r214, %r206, %r208;
mov.b32 %f86, %r215;
add.ftz.f32 %f87, %f85, %f86;
mov.b32 %r216, %f87;
mov.u32 %r217, 2;
shfl.sync.bfly.b32 %r218|%p25, %r216, %r217, %r206, %r208;
mov.b32 %f88, %r218;
add.ftz.f32 %f89, %f87, %f88;
mov.b32 %r219, %f89;
mov.u32 %r220, 1;
shfl.sync.bfly.b32 %r221|%p26, %r219, %r220, %r206, %r208;
mov.b32 %f90, %r221;
add.ftz.f32 %f15, %f89, %f90;
add.s32 %r222, %r4, 2;
setp.ge.u32 %p27, %r222, %r22;
setp.ne.s32 %p28, %r2, 2;
or.pred %p29, %p28, %p27;
@%p29 bra $L__BB44_15;

st.global.f32 [%rd10+8], %f15;

$L__BB44_15:
ld.shared.f32 %f91, [%r15+384];
add.ftz.f32 %f92, %f91, %f110;
ld.shared.f32 %f93, [%r15+896];
add.ftz.f32 %f94, %f93, %f92;
ld.shared.f32 %f95, [%r15+1408];
add.ftz.f32 %f96, %f95, %f94;
mov.b32 %r223, %f96;
shfl.sync.bfly.b32 %r227|%p30, %r223, %r207, %r206, %r208;
mov.b32 %f97, %r227;
add.ftz.f32 %f98, %f96, %f97;
mov.b32 %r228, %f98;
shfl.sync.bfly.b32 %r230|%p31, %r228, %r211, %r206, %r208;
mov.b32 %f99, %r230;
add.ftz.f32 %f100, %f98, %f99;
mov.b32 %r231, %f100;
shfl.sync.bfly.b32 %r233|%p32, %r231, %r214, %r206, %r208;
mov.b32 %f101, %r233;
add.ftz.f32 %f102, %f100, %f101;
mov.b32 %r234, %f102;
shfl.sync.bfly.b32 %r236|%p33, %r234, %r217, %r206, %r208;
mov.b32 %f103, %r236;
add.ftz.f32 %f104, %f102, %f103;
mov.b32 %r237, %f104;
shfl.sync.bfly.b32 %r239|%p34, %r237, %r220, %r206, %r208;
mov.b32 %f105, %r239;
add.ftz.f32 %f16, %f104, %f105;
add.s32 %r240, %r4, 3;
setp.ge.u32 %p35, %r240, %r22;
setp.ne.s32 %p36, %r2, 3;
or.pred %p37, %p36, %p35;
@%p37 bra $L__BB44_17;

st.global.f32 [%rd10+12], %f16;

$L__BB44_17:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<34>;
.reg .f32 %f<124>;
.reg .b32 %r<202>;
.reg .b64 %rd<63>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[384];

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate[384];

ld.param.u64 %rd11, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd12, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd13, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
ld.param.f32 %f31, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
ld.param.u32 %r19, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+16];
ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+8];
ld.param.u64 %rd14, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3];
setp.eq.s64 %p2, %rd13, 0;
@%p2 bra $L__BB45_2;

cvta.to.global.u64 %rd20, %rd13;
mov.u32 %r37, %ctaid.y;
mul.wide.u32 %rd21, %r37, 4;
add.s64 %rd22, %rd20, %rd21;
ld.global.nc.u32 %r199, [%rd22];
mul.hi.u32 %r38, %r37, %r21;
add.s32 %r39, %r38, %r37;
shr.u32 %r40, %r39, %r22;
mul.lo.s32 %r41, %r40, %r23;
sub.s32 %r200, %r37, %r41;
bra.uni $L__BB45_3;

$L__BB45_2:
mov.u32 %r200, %ctaid.y;
mul.hi.u32 %r42, %r200, %r25;
add.s32 %r43, %r42, %r200;
shr.u32 %r199, %r43, %r26;

$L__BB45_3:
mov.u32 %r44, %tid.y;
mov.u32 %r45, %tid.x;
or.b32 %r46, %r45, %r44;
setp.ne.s32 %p3, %r46, 0;
mov.f32 %f117, 0f00000000;
mov.f32 %f113, %f117;
mov.f32 %f114, %f117;
@%p3 bra $L__BB45_8;

setp.eq.s64 %p4, %rd14, 0;
mov.f32 %f113, 0f00000000;
mov.f32 %f114, %f113;
@%p4 bra $L__BB45_6;

cvta.to.global.u64 %rd23, %rd14;
mov.u32 %r47, %ctaid.z;
mul.lo.s32 %r48, %r47, %r36;
cvt.u64.u32 %rd24, %r48;
mov.u32 %r49, %ctaid.y;
selp.b32 %r50, %r49, %r199, %p2;
mul.lo.s32 %r51, %r50, %r30;
cvt.u64.u32 %rd25, %r51;
mov.u32 %r52, %ctaid.x;
cvt.s64.s32 %rd26, %r52;
cvt.u64.u32 %rd27, %r45;
add.s64 %rd28, %rd27, %rd26;
add.s64 %rd29, %rd28, %rd24;
add.s64 %rd30, %rd29, %rd25;
shl.b64 %rd31, %rd30, 2;
add.s64 %rd32, %rd23, %rd31;
ld.global.f32 %f114, [%rd32];

$L__BB45_6:
setp.eq.s64 %p6, %rd16, 0;
setp.eq.s64 %p7, %rd15, 0;
or.pred %p8, %p6, %p7;
@%p8 bra $L__BB45_8;

cvta.to.global.u64 %rd33, %rd16;
mov.u32 %r54, %ctaid.z;
mul.lo.s32 %r55, %r54, %r36;
cvt.u64.u32 %rd34, %r55;
mov.u32 %r56, %ctaid.y;
selp.b32 %r57, %r56, %r199, %p2;
mul.lo.s32 %r58, %r57, %r30;
cvt.u64.u32 %rd35, %r58;
mov.u32 %r59, %ctaid.x;
cvt.s64.s32 %rd36, %r59;
cvt.u64.u32 %rd37, %r45;
add.s64 %rd38, %rd37, %rd36;
add.s64 %rd39, %rd38, %rd34;
add.s64 %rd40, %rd39, %rd35;
shl.b64 %rd41, %rd40, 2;
add.s64 %rd42, %rd33, %rd41;
ld.global.f32 %f113, [%rd42];

$L__BB45_8:
shl.b32 %r62, %r44, 5;
add.s32 %r64, %r62, %r45;
shr.u32 %r65, %r64, 31;
add.s32 %r66, %r64, %r65;
shr.s32 %r201, %r66, 1;
shr.u32 %r67, %r20, 5;
setp.ge.s32 %p10, %r201, %r67;
mov.f32 %f119, %f117;
@%p10 bra $L__BB45_13;

mov.u32 %r68, %ctaid.z;
and.b32 %r75, %r66, 2147483646;
sub.s32 %r76, %r64, %r75;
shl.b32 %r8, %r76, 1;
mul.hi.u32 %r77, %r68, %r31;
add.s32 %r78, %r77, %r68;
shr.u32 %r79, %r78, %r32;
mad.lo.s32 %r80, %r79, %r34, %r201;
mad.lo.s32 %r81, %r199, %r28, %r80;
mov.u32 %r82, %ctaid.x;
mad.lo.s32 %r83, %r82, %r24, %r81;
mul.wide.s32 %rd43, %r83, 20;
cvta.to.global.u64 %rd44, %rd15;
add.s64 %rd62, %rd44, %rd43;
cvta.to.global.u64 %rd45, %rd11;
add.s64 %rd61, %rd45, %rd43;
mul.lo.s32 %r84, %r200, %r29;
cvt.u64.u32 %rd46, %r84;
mul.lo.s32 %r85, %r68, %r35;
cvt.u64.u32 %rd47, %r85;
add.s64 %rd6, %rd46, %rd47;
mov.f32 %f119, 0f00000000;
cvta.to.global.u64 %rd48, %rd12;
mov.f32 %f117, %f119;

$L__BB45_10:
cvt.u64.u32 %rd49, %r201;
add.s64 %rd50, %rd6, %rd49;
mul.lo.s64 %rd51, %rd50, 36;
add.s64 %rd52, %rd48, %rd51;
mul.wide.s32 %rd53, %r8, 4;
add.s64 %rd54, %rd52, %rd53;
add.s64 %rd55, %rd61, %rd53;
ld.global.nc.u32 %r10, [%rd54+4];
ld.global.nc.u32 %r11, [%rd54+20];
ld.global.nc.u32 %r107, [%rd55+8];
ld.global.nc.u32 %r12, [%rd54+8];
ld.global.nc.u32 %r13, [%rd54+24];
ld.global.nc.u32 %r108, [%rd55+4];
and.b32 %r87, %r108, 252645135;
shr.u32 %r109, %r108, 4;
and.b32 %r91, %r109, 252645135;
mov.u32 %r89, 0;

dp4a.s32.s32 %r86, %r87, %r10, %r89;


dp4a.s32.s32 %r90, %r91, %r11, %r86;

and.b32 %r95, %r107, 252645135;
shr.u32 %r110, %r107, 4;
and.b32 %r99, %r110, 252645135;

dp4a.s32.s32 %r94, %r95, %r12, %r90;


dp4a.s32.s32 %r98, %r99, %r13, %r94;

ld.global.nc.u32 %r103, [%rd61];
ld.global.nc.u32 %r14, [%rd52];

{mul.f16x2 %r102,%r103,%r14;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r102;
cvt.f32.f16 %f40, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r102;
cvt.f32.f16 %f41, high;}


cvt.rn.f32.s32 %f42, %r98;
mov.f32 %f43, 0f40000000;
div.approx.ftz.f32 %f44, %f41, %f43;
fma.rn.ftz.f32 %f45, %f40, %f42, %f44;
add.ftz.f32 %f119, %f119, %f45;
setp.eq.s64 %p11, %rd15, 0;
@%p11 bra $L__BB45_12;

add.s64 %rd57, %rd62, %rd53;
ld.global.u32 %r132, [%rd57+8];
ld.global.u32 %r133, [%rd57+4];
and.b32 %r112, %r133, 252645135;
shr.u32 %r134, %r133, 4;
and.b32 %r116, %r134, 252645135;

dp4a.s32.s32 %r111, %r112, %r10, %r89;


dp4a.s32.s32 %r115, %r116, %r11, %r111;

and.b32 %r120, %r132, 252645135;
shr.u32 %r135, %r132, 4;
and.b32 %r124, %r135, 252645135;

dp4a.s32.s32 %r119, %r120, %r12, %r115;


dp4a.s32.s32 %r123, %r124, %r13, %r119;

ld.global.u32 %r128, [%rd62];

{mul.f16x2 %r127,%r128,%r14;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r127;
cvt.f32.f16 %f46, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r127;
cvt.f32.f16 %f47, high;}


cvt.rn.f32.s32 %f48, %r123;
div.approx.ftz.f32 %f50, %f47, %f43;
fma.rn.ftz.f32 %f51, %f46, %f48, %f50;
add.ftz.f32 %f117, %f117, %f51;

$L__BB45_12:
add.s64 %rd62, %rd62, 1280;
add.s64 %rd61, %rd61, 1280;
add.s32 %r201, %r201, 64;
setp.lt.s32 %p12, %r201, %r67;
@%p12 bra $L__BB45_10;

$L__BB45_13:
setp.eq.s32 %p13, %r44, 0;
@%p13 bra $L__BB45_16;

setp.eq.s64 %p14, %rd15, 0;
add.s32 %r16, %r44, -1;
shl.b32 %r140, %r44, 7;
mov.u32 %r141, _ZZ13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r142, %r141, %r140;
shl.b32 %r143, %r45, 2;
add.s32 %r144, %r142, %r143;
st.shared.f32 [%r144+-128], %f119;
@%p14 bra $L__BB45_16;

shl.b32 %r146, %r16, 7;
mov.u32 %r147, _ZZ13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r148, %r147, %r146;
add.s32 %r150, %r148, %r143;
st.shared.f32 [%r150], %f117;

$L__BB45_16:
setp.ne.s32 %p15, %r44, 0;
bar.sync 0;
@%p15 bra $L__BB45_35;

setp.ne.s64 %p16, %rd15, 0;
shl.b32 %r153, %r45, 2;
mov.u32 %r154, _ZZ13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r17, %r154, %r153;
ld.shared.f32 %f52, [%r17];
add.ftz.f32 %f13, %f119, %f52;
@%p16 bra $L__BB45_19;
bra.uni $L__BB45_18;

$L__BB45_19:
mov.u32 %r157, _ZZ13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r158, %r157, %r153;
ld.shared.f32 %f56, [%r158];
add.ftz.f32 %f57, %f117, %f56;
ld.shared.f32 %f58, [%r17+128];
add.ftz.f32 %f59, %f13, %f58;
ld.shared.f32 %f60, [%r158+128];
add.ftz.f32 %f61, %f57, %f60;
ld.shared.f32 %f62, [%r17+256];
add.ftz.f32 %f121, %f59, %f62;
ld.shared.f32 %f63, [%r158+256];
add.ftz.f32 %f117, %f61, %f63;
bra.uni $L__BB45_20;

$L__BB45_18:
ld.shared.f32 %f53, [%r17+128];
add.ftz.f32 %f54, %f13, %f53;
ld.shared.f32 %f55, [%r17+256];
add.ftz.f32 %f121, %f54, %f55;

$L__BB45_20:
mov.b32 %r159, %f121;
mov.u32 %r160, 31;
mov.u32 %r161, 16;
mov.u32 %r162, -1;
shfl.sync.bfly.b32 %r163|%p17, %r159, %r161, %r160, %r162;
mov.b32 %f64, %r163;
add.ftz.f32 %f65, %f121, %f64;
mov.b32 %r164, %f65;
mov.u32 %r165, 8;
shfl.sync.bfly.b32 %r166|%p18, %r164, %r165, %r160, %r162;
mov.b32 %f66, %r166;
add.ftz.f32 %f67, %f65, %f66;
mov.b32 %r167, %f67;
mov.u32 %r168, 4;
shfl.sync.bfly.b32 %r169|%p19, %r167, %r168, %r160, %r162;
mov.b32 %f68, %r169;
add.ftz.f32 %f69, %f67, %f68;
mov.b32 %r170, %f69;
mov.u32 %r171, 2;
shfl.sync.bfly.b32 %r172|%p20, %r170, %r171, %r160, %r162;
mov.b32 %f70, %r172;
add.ftz.f32 %f19, %f69, %f70;
mov.b32 %r173, %f19;
mov.u32 %r174, 1;
shfl.sync.bfly.b32 %r18|%p1, %r173, %r174, %r160, %r162;
setp.eq.s64 %p21, %rd15, 0;
@%p21 bra $L__BB45_22;

mov.b32 %r175, %f117;
shfl.sync.bfly.b32 %r179|%p22, %r175, %r161, %r160, %r162;
mov.b32 %f71, %r179;
add.ftz.f32 %f72, %f117, %f71;
mov.b32 %r180, %f72;
shfl.sync.bfly.b32 %r182|%p23, %r180, %r165, %r160, %r162;
mov.b32 %f73, %r182;
add.ftz.f32 %f74, %f72, %f73;
mov.b32 %r183, %f74;
shfl.sync.bfly.b32 %r185|%p24, %r183, %r168, %r160, %r162;
mov.b32 %f75, %r185;
add.ftz.f32 %f76, %f74, %f75;
mov.b32 %r186, %f76;
shfl.sync.bfly.b32 %r188|%p25, %r186, %r171, %r160, %r162;
mov.b32 %f77, %r188;
add.ftz.f32 %f78, %f76, %f77;
mov.b32 %r189, %f78;
shfl.sync.bfly.b32 %r191|%p26, %r189, %r174, %r160, %r162;
mov.b32 %f79, %r191;
add.ftz.f32 %f117, %f78, %f79;

$L__BB45_22:
setp.ne.s32 %p27, %r45, 0;
mov.b32 %f80, %r18;
add.ftz.f32 %f22, %f19, %f80;
@%p27 bra $L__BB45_35;

add.ftz.f32 %f123, %f114, %f22;
@%p21 bra $L__BB45_34;

add.ftz.f32 %f24, %f113, %f117;
setp.gt.s32 %p29, %r19, 2;
@%p29 bra $L__BB45_28;

setp.eq.s32 %p32, %r19, 1;
@%p32 bra $L__BB45_32;

setp.eq.s32 %p33, %r19, 2;
@%p33 bra $L__BB45_27;
bra.uni $L__BB45_33;

$L__BB45_27:
mul.ftz.f32 %f100, %f24, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f101, %f100;
add.ftz.f32 %f102, %f101, 0f3F800000;
div.approx.ftz.f32 %f103, %f24, %f102;
mul.ftz.f32 %f123, %f123, %f103;
bra.uni $L__BB45_34;

$L__BB45_28:
setp.eq.s32 %p30, %r19, 3;
@%p30 bra $L__BB45_31;

setp.ne.s32 %p31, %r19, 6;
@%p31 bra $L__BB45_33;

min.ftz.f32 %f81, %f24, %f31;
neg.ftz.f32 %f82, %f31;
min.ftz.f32 %f83, %f123, %f31;
max.ftz.f32 %f84, %f83, %f82;
mul.ftz.f32 %f85, %f81, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f86, %f85;
add.ftz.f32 %f87, %f86, 0f3F800000;
div.approx.ftz.f32 %f88, %f81, %f87;
mul.ftz.f32 %f123, %f84, %f88;
bra.uni $L__BB45_34;

$L__BB45_32:
mul.ftz.f32 %f104, %f24, 0f3F000000;
mul.ftz.f32 %f105, %f24, 0f3D372713;
fma.rn.ftz.f32 %f106, %f24, %f105, 0f3F800000;
mul.ftz.f32 %f107, %f24, 0f3F4C422A;
mul.ftz.f32 %f108, %f107, %f106;
tanh.approx.f32 %f109, %f108;
add.ftz.f32 %f110, %f109, 0f3F800000;
mul.ftz.f32 %f111, %f104, %f110;
mul.ftz.f32 %f123, %f123, %f111;
bra.uni $L__BB45_34;

$L__BB45_31:
mov.f32 %f89, 0f40E00000;
min.ftz.f32 %f90, %f24, %f89;
min.ftz.f32 %f91, %f123, %f89;
mov.f32 %f92, 0fC0E00000;
max.ftz.f32 %f93, %f91, %f92;
mul.ftz.f32 %f94, %f90, 0f3FD9DB23;
mul.ftz.f32 %f95, %f94, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f96, %f95;
add.ftz.f32 %f97, %f96, 0f3F800000;
div.approx.ftz.f32 %f98, %f90, %f97;
add.ftz.f32 %f99, %f93, 0f3F800000;
mul.ftz.f32 %f123, %f99, %f98;
bra.uni $L__BB45_34;

$L__BB45_33:
mul.ftz.f32 %f123, %f123, %f24;

$L__BB45_34:
cvta.to.global.u64 %rd58, %rd19;
mov.u32 %r194, %ctaid.z;
mov.u32 %r195, %ctaid.y;
mov.u32 %r196, %ctaid.x;
mad.lo.s32 %r197, %r195, %r30, %r196;
mad.lo.s32 %r198, %r194, %r36, %r197;
mul.wide.u32 %rd59, %r198, 4;
add.s64 %rd60, %rd58, %rd59;
st.global.f32 [%rd60], %f123;

$L__BB45_35:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<14>;
.reg .f32 %f<48>;
.reg .b32 %r<176>;
.reg .b64 %rd<45>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[384];

ld.param.u64 %rd11, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd12, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd9, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd10, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
cvta.to.global.u64 %rd1, %rd12;
cvta.to.global.u64 %rd2, %rd11;
mov.u32 %r1, %tid.y;
shl.b32 %r41, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r41, %r2;
shr.u32 %r4, %r24, 5;
mov.u32 %r5, %ctaid.y;
setp.eq.s64 %p1, %rd9, 0;
@%p1 bra $L__BB46_2;

cvta.to.global.u64 %rd13, %rd9;
mul.wide.u32 %rd14, %r5, 4;
add.s64 %rd15, %rd13, %rd14;
ld.global.nc.u32 %r172, [%rd15];
mul.hi.u32 %r42, %r5, %r25;
add.s32 %r43, %r42, %r5;
shr.u32 %r44, %r43, %r26;
mul.lo.s32 %r45, %r44, %r27;
sub.s32 %r173, %r5, %r45;
bra.uni $L__BB46_3;

$L__BB46_2:
mul.hi.u32 %r46, %r5, %r29;
add.s32 %r47, %r46, %r5;
shr.u32 %r172, %r47, %r30;
mov.u32 %r173, %r5;

$L__BB46_3:
mov.u32 %r11, %ctaid.z;
mul.hi.u32 %r48, %r11, %r35;
add.s32 %r49, %r48, %r11;
shr.u32 %r50, %r49, %r36;
mul.lo.s32 %r51, %r11, %r39;
cvt.u64.u32 %rd16, %r51;
mul.lo.s32 %r52, %r173, %r33;
cvt.u64.u32 %rd17, %r52;
add.s64 %rd3, %rd17, %rd16;
mul.lo.s32 %r12, %r50, %r38;
mul.lo.s32 %r13, %r172, %r32;
mov.u32 %r53, %ctaid.x;
mul.lo.s32 %r14, %r53, %r28;
shr.u32 %r54, %r3, 31;
add.s32 %r55, %r3, %r54;
shr.s32 %r175, %r55, 1;
setp.ge.s32 %p2, %r175, %r4;
mov.f32 %f47, 0f00000000;
@%p2 bra $L__BB46_9;

and.b32 %r58, %r55, 2147483646;
sub.s32 %r59, %r3, %r58;
shl.b32 %r16, %r59, 1;
not.b32 %r60, %r175;
add.s32 %r17, %r4, %r60;
and.b32 %r61, %r17, 64;
setp.ne.s32 %p3, %r61, 0;
mov.f32 %f47, 0f00000000;
@%p3 bra $L__BB46_6;

cvt.u64.u32 %rd18, %r175;
add.s64 %rd19, %rd3, %rd18;
add.s32 %r83, %r13, %r14;
add.s32 %r84, %r83, %r12;
add.s32 %r85, %r84, %r175;
mul.wide.s32 %rd20, %r85, 20;
add.s64 %rd21, %rd2, %rd20;
mul.wide.s32 %rd22, %r16, 4;
add.s64 %rd23, %rd21, %rd22;
mul.lo.s64 %rd24, %rd19, 36;
add.s64 %rd25, %rd1, %rd24;
add.s64 %rd26, %rd25, %rd22;
ld.global.nc.u32 %r64, [%rd26+4];
ld.global.nc.u32 %r68, [%rd26+20];
ld.global.nc.u32 %r86, [%rd23+8];
ld.global.nc.u32 %r72, [%rd26+8];
ld.global.nc.u32 %r76, [%rd26+24];
ld.global.nc.u32 %r87, [%rd23+4];
and.b32 %r63, %r87, 252645135;
shr.u32 %r88, %r87, 4;
and.b32 %r67, %r88, 252645135;
mov.u32 %r65, 0;

dp4a.s32.s32 %r62, %r63, %r64, %r65;


dp4a.s32.s32 %r66, %r67, %r68, %r62;

and.b32 %r71, %r86, 252645135;
shr.u32 %r89, %r86, 4;
and.b32 %r75, %r89, 252645135;

dp4a.s32.s32 %r70, %r71, %r72, %r66;


dp4a.s32.s32 %r74, %r75, %r76, %r70;

ld.global.nc.u32 %r79, [%rd21];
ld.global.nc.u32 %r80, [%rd25];

{mul.f16x2 %r78,%r79,%r80;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r78;
cvt.f32.f16 %f11, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r78;
cvt.f32.f16 %f12, high;}


cvt.rn.f32.s32 %f13, %r74;
mov.f32 %f14, 0f40000000;
div.approx.ftz.f32 %f15, %f12, %f14;
fma.rn.ftz.f32 %f16, %f11, %f13, %f15;
add.ftz.f32 %f47, %f16, 0f00000000;
add.s32 %r175, %r175, 64;

$L__BB46_6:
and.b32 %r90, %r17, -64;
setp.eq.s32 %p4, %r90, 0;
@%p4 bra $L__BB46_9;

cvt.s64.s32 %rd4, %r16;
add.s32 %r91, %r175, %r12;
add.s32 %r92, %r91, %r13;
add.s32 %r93, %r92, %r14;
mul.wide.s32 %rd27, %r93, 20;
add.s64 %rd44, %rd2, %rd27;
mul.wide.s32 %rd28, %r16, 4;
add.s64 %rd6, %rd28, 4;
shl.b64 %rd33, %rd4, 2;

$L__BB46_8:
cvt.u64.u32 %rd29, %r175;
add.s64 %rd30, %rd3, %rd29;
mul.lo.s64 %rd31, %rd30, 36;
add.s64 %rd32, %rd1, %rd31;
add.s64 %rd34, %rd32, %rd33;
add.s64 %rd35, %rd44, %rd6;
ld.global.nc.u32 %r96, [%rd34+4];
ld.global.nc.u32 %r100, [%rd34+20];
ld.global.nc.u32 %r136, [%rd35+4];
ld.global.nc.u32 %r104, [%rd34+8];
ld.global.nc.u32 %r108, [%rd34+24];
ld.global.nc.u32 %r137, [%rd35];
and.b32 %r95, %r137, 252645135;
shr.u32 %r138, %r137, 4;
and.b32 %r99, %r138, 252645135;
mov.u32 %r118, 0;

dp4a.s32.s32 %r94, %r95, %r96, %r118;


dp4a.s32.s32 %r98, %r99, %r100, %r94;

and.b32 %r103, %r136, 252645135;
shr.u32 %r139, %r136, 4;
and.b32 %r107, %r139, 252645135;

dp4a.s32.s32 %r102, %r103, %r104, %r98;


dp4a.s32.s32 %r106, %r107, %r108, %r102;

ld.global.nc.u32 %r111, [%rd44];
ld.global.nc.u32 %r112, [%rd32];

{mul.f16x2 %r110,%r111,%r112;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r110;
cvt.f32.f16 %f17, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r110;
cvt.f32.f16 %f18, high;}


cvt.rn.f32.s32 %f21, %r106;
mov.f32 %f22, 0f40000000;
div.approx.ftz.f32 %f23, %f18, %f22;
fma.rn.ftz.f32 %f24, %f17, %f21, %f23;
add.ftz.f32 %f25, %f47, %f24;
add.s32 %r140, %r175, 64;
cvt.u64.u32 %rd36, %r140;
add.s64 %rd37, %rd3, %rd36;
mul.lo.s64 %rd38, %rd37, 36;
add.s64 %rd39, %rd1, %rd38;
add.s64 %rd40, %rd39, %rd33;
ld.global.nc.u32 %r117, [%rd40+4];
ld.global.nc.u32 %r121, [%rd40+20];
ld.global.nc.u32 %r141, [%rd35+1284];
ld.global.nc.u32 %r125, [%rd40+8];
ld.global.nc.u32 %r129, [%rd40+24];
ld.global.nc.u32 %r142, [%rd35+1280];
and.b32 %r116, %r142, 252645135;
shr.u32 %r143, %r142, 4;
and.b32 %r120, %r143, 252645135;

dp4a.s32.s32 %r115, %r116, %r117, %r118;


dp4a.s32.s32 %r119, %r120, %r121, %r115;

and.b32 %r124, %r141, 252645135;
shr.u32 %r144, %r141, 4;
and.b32 %r128, %r144, 252645135;

dp4a.s32.s32 %r123, %r124, %r125, %r119;


dp4a.s32.s32 %r127, %r128, %r129, %r123;

ld.global.nc.u32 %r132, [%rd44+1280];
ld.global.nc.u32 %r133, [%rd39];

{mul.f16x2 %r131,%r132,%r133;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r131;
cvt.f32.f16 %f19, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r131;
cvt.f32.f16 %f20, high;}


cvt.rn.f32.s32 %f26, %r127;
div.approx.ftz.f32 %f27, %f20, %f22;
fma.rn.ftz.f32 %f28, %f19, %f26, %f27;
add.ftz.f32 %f47, %f25, %f28;
add.s64 %rd44, %rd44, 2560;
add.s32 %r175, %r175, 128;
setp.lt.s32 %p5, %r175, %r4;
@%p5 bra $L__BB46_8;

$L__BB46_9:
setp.eq.s32 %p6, %r1, 0;
@%p6 bra $L__BB46_11;

shl.b32 %r145, %r1, 7;
mov.u32 %r146, _ZZ13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r147, %r146, %r145;
shl.b32 %r148, %r2, 2;
add.s32 %r149, %r147, %r148;
st.shared.f32 [%r149+-128], %f47;

$L__BB46_11:
setp.ne.s32 %p7, %r1, 0;
bar.sync 0;
@%p7 bra $L__BB46_14;

mad.lo.s32 %r151, %r5, %r34, %r53;
mad.lo.s32 %r23, %r11, %r40, %r151;
shl.b32 %r152, %r2, 2;
mov.u32 %r153, _ZZ13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r154, %r153, %r152;
mov.u32 %r155, 2;
ld.shared.f32 %f29, [%r154];
add.ftz.f32 %f30, %f47, %f29;
ld.shared.f32 %f31, [%r154+128];
add.ftz.f32 %f32, %f30, %f31;
ld.shared.f32 %f33, [%r154+256];
add.ftz.f32 %f34, %f32, %f33;
mov.b32 %r156, %f34;
mov.u32 %r157, 31;
mov.u32 %r158, 16;
mov.u32 %r159, -1;
shfl.sync.bfly.b32 %r160|%p8, %r156, %r158, %r157, %r159;
mov.b32 %f35, %r160;
add.ftz.f32 %f36, %f34, %f35;
mov.b32 %r161, %f36;
mov.u32 %r162, 8;
shfl.sync.bfly.b32 %r163|%p9, %r161, %r162, %r157, %r159;
mov.b32 %f37, %r163;
add.ftz.f32 %f38, %f36, %f37;
mov.b32 %r164, %f38;
mov.u32 %r165, 4;
shfl.sync.bfly.b32 %r166|%p10, %r164, %r165, %r157, %r159;
mov.b32 %f39, %r166;
add.ftz.f32 %f40, %f38, %f39;
mov.b32 %r167, %f40;
shfl.sync.bfly.b32 %r168|%p11, %r167, %r155, %r157, %r159;
mov.b32 %f41, %r168;
add.ftz.f32 %f42, %f40, %f41;
mov.b32 %r169, %f42;
mov.u32 %r170, 1;
shfl.sync.bfly.b32 %r171|%p12, %r169, %r170, %r157, %r159;
mov.b32 %f43, %r171;
add.ftz.f32 %f7, %f42, %f43;
setp.ne.s32 %p13, %r2, 0;
@%p13 bra $L__BB46_14;

cvta.to.global.u64 %rd41, %rd10;
mul.wide.u32 %rd42, %r23, 4;
add.s64 %rd43, %rd41, %rd42;
st.global.f32 [%rd43], %f7;

$L__BB46_14:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<34>;
.reg .f32 %f<124>;
.reg .b32 %r<202>;
.reg .b64 %rd<63>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[384];

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate[384];

ld.param.u64 %rd11, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd12, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd13, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
ld.param.f32 %f31, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
ld.param.u32 %r19, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+16];
ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+8];
ld.param.u64 %rd14, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3];
setp.eq.s64 %p2, %rd13, 0;
@%p2 bra $L__BB47_2;

cvta.to.global.u64 %rd20, %rd13;
mov.u32 %r37, %ctaid.y;
mul.wide.u32 %rd21, %r37, 4;
add.s64 %rd22, %rd20, %rd21;
ld.global.nc.u32 %r199, [%rd22];
mul.hi.u32 %r38, %r37, %r21;
add.s32 %r39, %r38, %r37;
shr.u32 %r40, %r39, %r22;
mul.lo.s32 %r41, %r40, %r23;
sub.s32 %r200, %r37, %r41;
bra.uni $L__BB47_3;

$L__BB47_2:
mov.u32 %r200, %ctaid.y;
mul.hi.u32 %r42, %r200, %r25;
add.s32 %r43, %r42, %r200;
shr.u32 %r199, %r43, %r26;

$L__BB47_3:
mov.u32 %r44, %tid.y;
mov.u32 %r45, %tid.x;
or.b32 %r46, %r45, %r44;
setp.ne.s32 %p3, %r46, 0;
mov.f32 %f117, 0f00000000;
mov.f32 %f113, %f117;
mov.f32 %f114, %f117;
@%p3 bra $L__BB47_8;

setp.eq.s64 %p4, %rd14, 0;
mov.f32 %f113, 0f00000000;
mov.f32 %f114, %f113;
@%p4 bra $L__BB47_6;

cvta.to.global.u64 %rd23, %rd14;
mov.u32 %r47, %ctaid.z;
mul.lo.s32 %r48, %r47, %r36;
cvt.u64.u32 %rd24, %r48;
mov.u32 %r49, %ctaid.y;
selp.b32 %r50, %r49, %r199, %p2;
mul.lo.s32 %r51, %r50, %r30;
cvt.u64.u32 %rd25, %r51;
mov.u32 %r52, %ctaid.x;
cvt.s64.s32 %rd26, %r52;
cvt.u64.u32 %rd27, %r45;
add.s64 %rd28, %rd27, %rd26;
add.s64 %rd29, %rd28, %rd24;
add.s64 %rd30, %rd29, %rd25;
shl.b64 %rd31, %rd30, 2;
add.s64 %rd32, %rd23, %rd31;
ld.global.f32 %f114, [%rd32];

$L__BB47_6:
setp.eq.s64 %p6, %rd16, 0;
setp.eq.s64 %p7, %rd15, 0;
or.pred %p8, %p6, %p7;
@%p8 bra $L__BB47_8;

cvta.to.global.u64 %rd33, %rd16;
mov.u32 %r54, %ctaid.z;
mul.lo.s32 %r55, %r54, %r36;
cvt.u64.u32 %rd34, %r55;
mov.u32 %r56, %ctaid.y;
selp.b32 %r57, %r56, %r199, %p2;
mul.lo.s32 %r58, %r57, %r30;
cvt.u64.u32 %rd35, %r58;
mov.u32 %r59, %ctaid.x;
cvt.s64.s32 %rd36, %r59;
cvt.u64.u32 %rd37, %r45;
add.s64 %rd38, %rd37, %rd36;
add.s64 %rd39, %rd38, %rd34;
add.s64 %rd40, %rd39, %rd35;
shl.b64 %rd41, %rd40, 2;
add.s64 %rd42, %rd33, %rd41;
ld.global.f32 %f113, [%rd42];

$L__BB47_8:
shl.b32 %r62, %r44, 5;
add.s32 %r64, %r62, %r45;
shr.u32 %r65, %r64, 31;
add.s32 %r66, %r64, %r65;
shr.s32 %r201, %r66, 1;
shr.u32 %r67, %r20, 5;
setp.ge.s32 %p10, %r201, %r67;
mov.f32 %f119, %f117;
@%p10 bra $L__BB47_13;

mov.u32 %r68, %ctaid.z;
and.b32 %r75, %r66, 2147483646;
sub.s32 %r76, %r64, %r75;
shl.b32 %r8, %r76, 1;
mul.hi.u32 %r77, %r68, %r31;
add.s32 %r78, %r77, %r68;
shr.u32 %r79, %r78, %r32;
mad.lo.s32 %r80, %r79, %r34, %r201;
mad.lo.s32 %r81, %r199, %r28, %r80;
mov.u32 %r82, %ctaid.x;
mad.lo.s32 %r83, %r82, %r24, %r81;
mul.wide.s32 %rd43, %r83, 20;
cvta.to.global.u64 %rd44, %rd15;
add.s64 %rd62, %rd44, %rd43;
cvta.to.global.u64 %rd45, %rd11;
add.s64 %rd61, %rd45, %rd43;
mul.lo.s32 %r84, %r200, %r29;
cvt.u64.u32 %rd46, %r84;
mul.lo.s32 %r85, %r68, %r35;
cvt.u64.u32 %rd47, %r85;
add.s64 %rd6, %rd46, %rd47;
mov.f32 %f119, 0f00000000;
cvta.to.global.u64 %rd48, %rd12;
mov.f32 %f117, %f119;

$L__BB47_10:
cvt.u64.u32 %rd49, %r201;
add.s64 %rd50, %rd6, %rd49;
mul.lo.s64 %rd51, %rd50, 36;
add.s64 %rd52, %rd48, %rd51;
mul.wide.s32 %rd53, %r8, 4;
add.s64 %rd54, %rd52, %rd53;
add.s64 %rd55, %rd61, %rd53;
ld.global.nc.u32 %r10, [%rd54+4];
ld.global.nc.u32 %r11, [%rd54+20];
ld.global.nc.u32 %r107, [%rd55+8];
ld.global.nc.u32 %r12, [%rd54+8];
ld.global.nc.u32 %r13, [%rd54+24];
ld.global.nc.u32 %r108, [%rd55+4];
and.b32 %r87, %r108, 252645135;
shr.u32 %r109, %r108, 4;
and.b32 %r91, %r109, 252645135;
mov.u32 %r89, 0;

dp4a.s32.s32 %r86, %r87, %r10, %r89;


dp4a.s32.s32 %r90, %r91, %r11, %r86;

and.b32 %r95, %r107, 252645135;
shr.u32 %r110, %r107, 4;
and.b32 %r99, %r110, 252645135;

dp4a.s32.s32 %r94, %r95, %r12, %r90;


dp4a.s32.s32 %r98, %r99, %r13, %r94;

ld.global.nc.u32 %r103, [%rd61];
ld.global.nc.u32 %r14, [%rd52];

{mul.f16x2 %r102,%r103,%r14;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r102;
cvt.f32.f16 %f40, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r102;
cvt.f32.f16 %f41, high;}


cvt.rn.f32.s32 %f42, %r98;
mov.f32 %f43, 0f40000000;
div.approx.ftz.f32 %f44, %f41, %f43;
fma.rn.ftz.f32 %f45, %f40, %f42, %f44;
add.ftz.f32 %f119, %f119, %f45;
setp.eq.s64 %p11, %rd15, 0;
@%p11 bra $L__BB47_12;

add.s64 %rd57, %rd62, %rd53;
ld.global.u32 %r132, [%rd57+8];
ld.global.u32 %r133, [%rd57+4];
and.b32 %r112, %r133, 252645135;
shr.u32 %r134, %r133, 4;
and.b32 %r116, %r134, 252645135;

dp4a.s32.s32 %r111, %r112, %r10, %r89;


dp4a.s32.s32 %r115, %r116, %r11, %r111;

and.b32 %r120, %r132, 252645135;
shr.u32 %r135, %r132, 4;
and.b32 %r124, %r135, 252645135;

dp4a.s32.s32 %r119, %r120, %r12, %r115;


dp4a.s32.s32 %r123, %r124, %r13, %r119;

ld.global.u32 %r128, [%rd62];

{mul.f16x2 %r127,%r128,%r14;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r127;
cvt.f32.f16 %f46, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r127;
cvt.f32.f16 %f47, high;}


cvt.rn.f32.s32 %f48, %r123;
div.approx.ftz.f32 %f50, %f47, %f43;
fma.rn.ftz.f32 %f51, %f46, %f48, %f50;
add.ftz.f32 %f117, %f117, %f51;

$L__BB47_12:
add.s64 %rd62, %rd62, 1280;
add.s64 %rd61, %rd61, 1280;
add.s32 %r201, %r201, 64;
setp.lt.s32 %p12, %r201, %r67;
@%p12 bra $L__BB47_10;

$L__BB47_13:
setp.eq.s32 %p13, %r44, 0;
@%p13 bra $L__BB47_16;

setp.eq.s64 %p14, %rd15, 0;
add.s32 %r16, %r44, -1;
shl.b32 %r140, %r44, 7;
mov.u32 %r141, _ZZ13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r142, %r141, %r140;
shl.b32 %r143, %r45, 2;
add.s32 %r144, %r142, %r143;
st.shared.f32 [%r144+-128], %f119;
@%p14 bra $L__BB47_16;

shl.b32 %r146, %r16, 7;
mov.u32 %r147, _ZZ13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r148, %r147, %r146;
add.s32 %r150, %r148, %r143;
st.shared.f32 [%r150], %f117;

$L__BB47_16:
setp.ne.s32 %p15, %r44, 0;
bar.sync 0;
@%p15 bra $L__BB47_35;

setp.ne.s64 %p16, %rd15, 0;
shl.b32 %r153, %r45, 2;
mov.u32 %r154, _ZZ13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r17, %r154, %r153;
ld.shared.f32 %f52, [%r17];
add.ftz.f32 %f13, %f119, %f52;
@%p16 bra $L__BB47_19;
bra.uni $L__BB47_18;

$L__BB47_19:
mov.u32 %r157, _ZZ13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r158, %r157, %r153;
ld.shared.f32 %f56, [%r158];
add.ftz.f32 %f57, %f117, %f56;
ld.shared.f32 %f58, [%r17+128];
add.ftz.f32 %f59, %f13, %f58;
ld.shared.f32 %f60, [%r158+128];
add.ftz.f32 %f61, %f57, %f60;
ld.shared.f32 %f62, [%r17+256];
add.ftz.f32 %f121, %f59, %f62;
ld.shared.f32 %f63, [%r158+256];
add.ftz.f32 %f117, %f61, %f63;
bra.uni $L__BB47_20;

$L__BB47_18:
ld.shared.f32 %f53, [%r17+128];
add.ftz.f32 %f54, %f13, %f53;
ld.shared.f32 %f55, [%r17+256];
add.ftz.f32 %f121, %f54, %f55;

$L__BB47_20:
mov.b32 %r159, %f121;
mov.u32 %r160, 31;
mov.u32 %r161, 16;
mov.u32 %r162, -1;
shfl.sync.bfly.b32 %r163|%p17, %r159, %r161, %r160, %r162;
mov.b32 %f64, %r163;
add.ftz.f32 %f65, %f121, %f64;
mov.b32 %r164, %f65;
mov.u32 %r165, 8;
shfl.sync.bfly.b32 %r166|%p18, %r164, %r165, %r160, %r162;
mov.b32 %f66, %r166;
add.ftz.f32 %f67, %f65, %f66;
mov.b32 %r167, %f67;
mov.u32 %r168, 4;
shfl.sync.bfly.b32 %r169|%p19, %r167, %r168, %r160, %r162;
mov.b32 %f68, %r169;
add.ftz.f32 %f69, %f67, %f68;
mov.b32 %r170, %f69;
mov.u32 %r171, 2;
shfl.sync.bfly.b32 %r172|%p20, %r170, %r171, %r160, %r162;
mov.b32 %f70, %r172;
add.ftz.f32 %f19, %f69, %f70;
mov.b32 %r173, %f19;
mov.u32 %r174, 1;
shfl.sync.bfly.b32 %r18|%p1, %r173, %r174, %r160, %r162;
setp.eq.s64 %p21, %rd15, 0;
@%p21 bra $L__BB47_22;

mov.b32 %r175, %f117;
shfl.sync.bfly.b32 %r179|%p22, %r175, %r161, %r160, %r162;
mov.b32 %f71, %r179;
add.ftz.f32 %f72, %f117, %f71;
mov.b32 %r180, %f72;
shfl.sync.bfly.b32 %r182|%p23, %r180, %r165, %r160, %r162;
mov.b32 %f73, %r182;
add.ftz.f32 %f74, %f72, %f73;
mov.b32 %r183, %f74;
shfl.sync.bfly.b32 %r185|%p24, %r183, %r168, %r160, %r162;
mov.b32 %f75, %r185;
add.ftz.f32 %f76, %f74, %f75;
mov.b32 %r186, %f76;
shfl.sync.bfly.b32 %r188|%p25, %r186, %r171, %r160, %r162;
mov.b32 %f77, %r188;
add.ftz.f32 %f78, %f76, %f77;
mov.b32 %r189, %f78;
shfl.sync.bfly.b32 %r191|%p26, %r189, %r174, %r160, %r162;
mov.b32 %f79, %r191;
add.ftz.f32 %f117, %f78, %f79;

$L__BB47_22:
setp.ne.s32 %p27, %r45, 0;
mov.b32 %f80, %r18;
add.ftz.f32 %f22, %f19, %f80;
@%p27 bra $L__BB47_35;

add.ftz.f32 %f123, %f114, %f22;
@%p21 bra $L__BB47_34;

add.ftz.f32 %f24, %f113, %f117;
setp.gt.s32 %p29, %r19, 2;
@%p29 bra $L__BB47_28;

setp.eq.s32 %p32, %r19, 1;
@%p32 bra $L__BB47_32;

setp.eq.s32 %p33, %r19, 2;
@%p33 bra $L__BB47_27;
bra.uni $L__BB47_33;

$L__BB47_27:
mul.ftz.f32 %f100, %f24, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f101, %f100;
add.ftz.f32 %f102, %f101, 0f3F800000;
div.approx.ftz.f32 %f103, %f24, %f102;
mul.ftz.f32 %f123, %f123, %f103;
bra.uni $L__BB47_34;

$L__BB47_28:
setp.eq.s32 %p30, %r19, 3;
@%p30 bra $L__BB47_31;

setp.ne.s32 %p31, %r19, 6;
@%p31 bra $L__BB47_33;

min.ftz.f32 %f81, %f24, %f31;
neg.ftz.f32 %f82, %f31;
min.ftz.f32 %f83, %f123, %f31;
max.ftz.f32 %f84, %f83, %f82;
mul.ftz.f32 %f85, %f81, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f86, %f85;
add.ftz.f32 %f87, %f86, 0f3F800000;
div.approx.ftz.f32 %f88, %f81, %f87;
mul.ftz.f32 %f123, %f84, %f88;
bra.uni $L__BB47_34;

$L__BB47_32:
mul.ftz.f32 %f104, %f24, 0f3F000000;
mul.ftz.f32 %f105, %f24, 0f3D372713;
fma.rn.ftz.f32 %f106, %f24, %f105, 0f3F800000;
mul.ftz.f32 %f107, %f24, 0f3F4C422A;
mul.ftz.f32 %f108, %f107, %f106;
tanh.approx.f32 %f109, %f108;
add.ftz.f32 %f110, %f109, 0f3F800000;
mul.ftz.f32 %f111, %f104, %f110;
mul.ftz.f32 %f123, %f123, %f111;
bra.uni $L__BB47_34;

$L__BB47_31:
mov.f32 %f89, 0f40E00000;
min.ftz.f32 %f90, %f24, %f89;
min.ftz.f32 %f91, %f123, %f89;
mov.f32 %f92, 0fC0E00000;
max.ftz.f32 %f93, %f91, %f92;
mul.ftz.f32 %f94, %f90, 0f3FD9DB23;
mul.ftz.f32 %f95, %f94, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f96, %f95;
add.ftz.f32 %f97, %f96, 0f3F800000;
div.approx.ftz.f32 %f98, %f90, %f97;
add.ftz.f32 %f99, %f93, 0f3F800000;
mul.ftz.f32 %f123, %f99, %f98;
bra.uni $L__BB47_34;

$L__BB47_33:
mul.ftz.f32 %f123, %f123, %f24;

$L__BB47_34:
cvta.to.global.u64 %rd58, %rd19;
mov.u32 %r194, %ctaid.z;
mov.u32 %r195, %ctaid.y;
mov.u32 %r196, %ctaid.x;
mad.lo.s32 %r197, %r195, %r30, %r196;
mad.lo.s32 %r198, %r194, %r36, %r197;
mul.wide.u32 %rd59, %r198, 4;
add.s64 %rd60, %rd58, %rd59;
st.global.f32 [%rd60], %f123;

$L__BB47_35:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<14>;
.reg .f32 %f<48>;
.reg .b32 %r<176>;
.reg .b64 %rd<45>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[384];

ld.param.u64 %rd11, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd12, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd9, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd10, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
cvta.to.global.u64 %rd1, %rd12;
cvta.to.global.u64 %rd2, %rd11;
mov.u32 %r1, %tid.y;
shl.b32 %r41, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r41, %r2;
shr.u32 %r4, %r24, 5;
mov.u32 %r5, %ctaid.y;
setp.eq.s64 %p1, %rd9, 0;
@%p1 bra $L__BB48_2;

cvta.to.global.u64 %rd13, %rd9;
mul.wide.u32 %rd14, %r5, 4;
add.s64 %rd15, %rd13, %rd14;
ld.global.nc.u32 %r172, [%rd15];
mul.hi.u32 %r42, %r5, %r25;
add.s32 %r43, %r42, %r5;
shr.u32 %r44, %r43, %r26;
mul.lo.s32 %r45, %r44, %r27;
sub.s32 %r173, %r5, %r45;
bra.uni $L__BB48_3;

$L__BB48_2:
mul.hi.u32 %r46, %r5, %r29;
add.s32 %r47, %r46, %r5;
shr.u32 %r172, %r47, %r30;
mov.u32 %r173, %r5;

$L__BB48_3:
mov.u32 %r11, %ctaid.z;
mul.hi.u32 %r48, %r11, %r35;
add.s32 %r49, %r48, %r11;
shr.u32 %r50, %r49, %r36;
mul.lo.s32 %r51, %r11, %r39;
cvt.u64.u32 %rd16, %r51;
mul.lo.s32 %r52, %r173, %r33;
cvt.u64.u32 %rd17, %r52;
add.s64 %rd3, %rd17, %rd16;
mul.lo.s32 %r12, %r50, %r38;
mul.lo.s32 %r13, %r172, %r32;
mov.u32 %r53, %ctaid.x;
mul.lo.s32 %r14, %r53, %r28;
shr.u32 %r54, %r3, 31;
add.s32 %r55, %r3, %r54;
shr.s32 %r175, %r55, 1;
setp.ge.s32 %p2, %r175, %r4;
mov.f32 %f47, 0f00000000;
@%p2 bra $L__BB48_9;

and.b32 %r58, %r55, 2147483646;
sub.s32 %r59, %r3, %r58;
shl.b32 %r16, %r59, 1;
not.b32 %r60, %r175;
add.s32 %r17, %r4, %r60;
and.b32 %r61, %r17, 64;
setp.ne.s32 %p3, %r61, 0;
mov.f32 %f47, 0f00000000;
@%p3 bra $L__BB48_6;

cvt.u64.u32 %rd18, %r175;
add.s64 %rd19, %rd3, %rd18;
add.s32 %r83, %r13, %r14;
add.s32 %r84, %r83, %r12;
add.s32 %r85, %r84, %r175;
mul.wide.s32 %rd20, %r85, 20;
add.s64 %rd21, %rd2, %rd20;
mul.wide.s32 %rd22, %r16, 4;
add.s64 %rd23, %rd21, %rd22;
mul.lo.s64 %rd24, %rd19, 36;
add.s64 %rd25, %rd1, %rd24;
add.s64 %rd26, %rd25, %rd22;
ld.global.nc.u32 %r64, [%rd26+4];
ld.global.nc.u32 %r68, [%rd26+20];
ld.global.nc.u32 %r86, [%rd23+8];
ld.global.nc.u32 %r72, [%rd26+8];
ld.global.nc.u32 %r76, [%rd26+24];
ld.global.nc.u32 %r87, [%rd23+4];
and.b32 %r63, %r87, 252645135;
shr.u32 %r88, %r87, 4;
and.b32 %r67, %r88, 252645135;
mov.u32 %r65, 0;

dp4a.s32.s32 %r62, %r63, %r64, %r65;


dp4a.s32.s32 %r66, %r67, %r68, %r62;

and.b32 %r71, %r86, 252645135;
shr.u32 %r89, %r86, 4;
and.b32 %r75, %r89, 252645135;

dp4a.s32.s32 %r70, %r71, %r72, %r66;


dp4a.s32.s32 %r74, %r75, %r76, %r70;

ld.global.nc.u32 %r79, [%rd21];
ld.global.nc.u32 %r80, [%rd25];

{mul.f16x2 %r78,%r79,%r80;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r78;
cvt.f32.f16 %f11, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r78;
cvt.f32.f16 %f12, high;}


cvt.rn.f32.s32 %f13, %r74;
mov.f32 %f14, 0f40000000;
div.approx.ftz.f32 %f15, %f12, %f14;
fma.rn.ftz.f32 %f16, %f11, %f13, %f15;
add.ftz.f32 %f47, %f16, 0f00000000;
add.s32 %r175, %r175, 64;

$L__BB48_6:
and.b32 %r90, %r17, -64;
setp.eq.s32 %p4, %r90, 0;
@%p4 bra $L__BB48_9;

cvt.s64.s32 %rd4, %r16;
add.s32 %r91, %r175, %r12;
add.s32 %r92, %r91, %r13;
add.s32 %r93, %r92, %r14;
mul.wide.s32 %rd27, %r93, 20;
add.s64 %rd44, %rd2, %rd27;
mul.wide.s32 %rd28, %r16, 4;
add.s64 %rd6, %rd28, 4;
shl.b64 %rd33, %rd4, 2;

$L__BB48_8:
cvt.u64.u32 %rd29, %r175;
add.s64 %rd30, %rd3, %rd29;
mul.lo.s64 %rd31, %rd30, 36;
add.s64 %rd32, %rd1, %rd31;
add.s64 %rd34, %rd32, %rd33;
add.s64 %rd35, %rd44, %rd6;
ld.global.nc.u32 %r96, [%rd34+4];
ld.global.nc.u32 %r100, [%rd34+20];
ld.global.nc.u32 %r136, [%rd35+4];
ld.global.nc.u32 %r104, [%rd34+8];
ld.global.nc.u32 %r108, [%rd34+24];
ld.global.nc.u32 %r137, [%rd35];
and.b32 %r95, %r137, 252645135;
shr.u32 %r138, %r137, 4;
and.b32 %r99, %r138, 252645135;
mov.u32 %r118, 0;

dp4a.s32.s32 %r94, %r95, %r96, %r118;


dp4a.s32.s32 %r98, %r99, %r100, %r94;

and.b32 %r103, %r136, 252645135;
shr.u32 %r139, %r136, 4;
and.b32 %r107, %r139, 252645135;

dp4a.s32.s32 %r102, %r103, %r104, %r98;


dp4a.s32.s32 %r106, %r107, %r108, %r102;

ld.global.nc.u32 %r111, [%rd44];
ld.global.nc.u32 %r112, [%rd32];

{mul.f16x2 %r110,%r111,%r112;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r110;
cvt.f32.f16 %f17, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r110;
cvt.f32.f16 %f18, high;}


cvt.rn.f32.s32 %f21, %r106;
mov.f32 %f22, 0f40000000;
div.approx.ftz.f32 %f23, %f18, %f22;
fma.rn.ftz.f32 %f24, %f17, %f21, %f23;
add.ftz.f32 %f25, %f47, %f24;
add.s32 %r140, %r175, 64;
cvt.u64.u32 %rd36, %r140;
add.s64 %rd37, %rd3, %rd36;
mul.lo.s64 %rd38, %rd37, 36;
add.s64 %rd39, %rd1, %rd38;
add.s64 %rd40, %rd39, %rd33;
ld.global.nc.u32 %r117, [%rd40+4];
ld.global.nc.u32 %r121, [%rd40+20];
ld.global.nc.u32 %r141, [%rd35+1284];
ld.global.nc.u32 %r125, [%rd40+8];
ld.global.nc.u32 %r129, [%rd40+24];
ld.global.nc.u32 %r142, [%rd35+1280];
and.b32 %r116, %r142, 252645135;
shr.u32 %r143, %r142, 4;
and.b32 %r120, %r143, 252645135;

dp4a.s32.s32 %r115, %r116, %r117, %r118;


dp4a.s32.s32 %r119, %r120, %r121, %r115;

and.b32 %r124, %r141, 252645135;
shr.u32 %r144, %r141, 4;
and.b32 %r128, %r144, 252645135;

dp4a.s32.s32 %r123, %r124, %r125, %r119;


dp4a.s32.s32 %r127, %r128, %r129, %r123;

ld.global.nc.u32 %r132, [%rd44+1280];
ld.global.nc.u32 %r133, [%rd39];

{mul.f16x2 %r131,%r132,%r133;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r131;
cvt.f32.f16 %f19, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r131;
cvt.f32.f16 %f20, high;}


cvt.rn.f32.s32 %f26, %r127;
div.approx.ftz.f32 %f27, %f20, %f22;
fma.rn.ftz.f32 %f28, %f19, %f26, %f27;
add.ftz.f32 %f47, %f25, %f28;
add.s64 %rd44, %rd44, 2560;
add.s32 %r175, %r175, 128;
setp.lt.s32 %p5, %r175, %r4;
@%p5 bra $L__BB48_8;

$L__BB48_9:
setp.eq.s32 %p6, %r1, 0;
@%p6 bra $L__BB48_11;

shl.b32 %r145, %r1, 7;
mov.u32 %r146, _ZZ13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r147, %r146, %r145;
shl.b32 %r148, %r2, 2;
add.s32 %r149, %r147, %r148;
st.shared.f32 [%r149+-128], %f47;

$L__BB48_11:
setp.ne.s32 %p7, %r1, 0;
bar.sync 0;
@%p7 bra $L__BB48_14;

mad.lo.s32 %r151, %r5, %r34, %r53;
mad.lo.s32 %r23, %r11, %r40, %r151;
shl.b32 %r152, %r2, 2;
mov.u32 %r153, _ZZ13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r154, %r153, %r152;
mov.u32 %r155, 2;
ld.shared.f32 %f29, [%r154];
add.ftz.f32 %f30, %f47, %f29;
ld.shared.f32 %f31, [%r154+128];
add.ftz.f32 %f32, %f30, %f31;
ld.shared.f32 %f33, [%r154+256];
add.ftz.f32 %f34, %f32, %f33;
mov.b32 %r156, %f34;
mov.u32 %r157, 31;
mov.u32 %r158, 16;
mov.u32 %r159, -1;
shfl.sync.bfly.b32 %r160|%p8, %r156, %r158, %r157, %r159;
mov.b32 %f35, %r160;
add.ftz.f32 %f36, %f34, %f35;
mov.b32 %r161, %f36;
mov.u32 %r162, 8;
shfl.sync.bfly.b32 %r163|%p9, %r161, %r162, %r157, %r159;
mov.b32 %f37, %r163;
add.ftz.f32 %f38, %f36, %f37;
mov.b32 %r164, %f38;
mov.u32 %r165, 4;
shfl.sync.bfly.b32 %r166|%p10, %r164, %r165, %r157, %r159;
mov.b32 %f39, %r166;
add.ftz.f32 %f40, %f38, %f39;
mov.b32 %r167, %f40;
shfl.sync.bfly.b32 %r168|%p11, %r167, %r155, %r157, %r159;
mov.b32 %f41, %r168;
add.ftz.f32 %f42, %f40, %f41;
mov.b32 %r169, %f42;
mov.u32 %r170, 1;
shfl.sync.bfly.b32 %r171|%p12, %r169, %r170, %r157, %r159;
mov.b32 %f43, %r171;
add.ftz.f32 %f7, %f42, %f43;
setp.ne.s32 %p13, %r2, 0;
@%p13 bra $L__BB48_14;

cvta.to.global.u64 %rd41, %rd10;
mul.wide.u32 %rd42, %r23, 4;
add.s64 %rd43, %rd41, %rd42;
st.global.f32 [%rd43], %f7;

$L__BB48_14:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type3ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type3ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type3ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type3ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type3ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type3ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type3ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type3ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type3ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<35>;
.reg .f32 %f<114>;
.reg .b32 %r<239>;
.reg .b64 %rd<48>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type3ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

ld.param.u64 %rd14, [_Z13mul_mat_vec_qIL9ggml_type3ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type3ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type3ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r13, [_Z13mul_mat_vec_qIL9ggml_type3ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r14, [_Z13mul_mat_vec_qIL9ggml_type3ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r15, [_Z13mul_mat_vec_qIL9ggml_type3ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r16, [_Z13mul_mat_vec_qIL9ggml_type3ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type3ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type3ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type3ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type3ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type3ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type3ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type3ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type3ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r18, [_Z13mul_mat_vec_qIL9ggml_type3ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r17, [_Z13mul_mat_vec_qIL9ggml_type3ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
cvta.to.global.u64 %rd1, %rd16;
mov.u32 %r1, %tid.y;
shl.b32 %r29, %r1, 5;
mov.u32 %r30, %tid.x;
add.s32 %r2, %r29, %r30;
mov.u32 %r31, %ctaid.x;
shl.b32 %r3, %r31, 1;
shr.u32 %r32, %r2, 31;
add.s32 %r33, %r2, %r32;
shr.s32 %r238, %r33, 1;
shr.u32 %r34, %r13, 5;
setp.ge.s32 %p3, %r238, %r34;
mov.f32 %f110, 0f00000000;
mov.f32 %f111, %f110;
mov.f32 %f112, %f110;
mov.f32 %f113, %f110;
@%p3 bra $L__BB49_3;

mov.u32 %r35, %ctaid.y;
mov.u32 %r36, %ctaid.z;
mul.hi.u32 %r37, %r36, %r23;
add.s32 %r38, %r37, %r36;
shr.u32 %r39, %r38, %r24;
mul.lo.s32 %r40, %r39, %r26;
mul.hi.u32 %r41, %r35, %r17;
add.s32 %r42, %r41, %r35;
shr.u32 %r43, %r42, %r18;
mul.lo.s32 %r44, %r43, %r20;
cvta.to.global.u64 %rd17, %rd14;
and.b32 %r47, %r33, 2147483646;
sub.s32 %r48, %r2, %r47;
shl.b32 %r49, %r48, 1;
cvt.s64.s32 %rd2, %r49;
add.s32 %r237, %r238, %r15;
add.s32 %r50, %r238, %r40;
add.s32 %r51, %r50, %r44;
add.s32 %r52, %r3, 1;
mad.lo.s32 %r53, %r14, %r52, %r51;
mul.wide.s32 %rd18, %r53, 20;
add.s64 %rd19, %rd17, %rd18;
add.s64 %rd47, %rd19, 8;
mul.wide.s32 %rd4, %r49, 4;
mad.lo.s32 %r54, %r3, %r14, %r238;
add.s32 %r55, %r54, %r40;
add.s32 %r56, %r55, %r44;
mul.wide.s32 %rd20, %r56, 20;
add.s64 %rd21, %rd17, %rd20;
add.s64 %rd46, %rd21, 8;
mul.lo.s32 %r57, %r35, %r21;
cvt.u64.u32 %rd22, %r57;
mul.lo.s32 %r58, %r36, %r27;
cvt.u64.u32 %rd23, %r58;
add.s64 %rd6, %rd23, %rd22;
cvta.to.global.u64 %rd7, %rd15;
mov.f32 %f110, 0f00000000;
shl.b64 %rd28, %rd2, 2;

$L__BB49_2:
cvt.u64.u32 %rd24, %r238;
add.s64 %rd25, %rd6, %rd24;
mul.lo.s64 %rd26, %rd25, 36;
add.s64 %rd27, %rd7, %rd26;
add.s64 %rd29, %rd27, %rd28;
ld.global.nc.u32 %r61, [%rd29+4];
ld.global.nc.u32 %r65, [%rd29+20];
add.s64 %rd30, %rd46, %rd4;
ld.global.nc.u32 %r143, [%rd30];
ld.global.nc.u32 %r69, [%rd29+8];
ld.global.nc.u32 %r73, [%rd29+24];
ld.global.nc.u32 %r144, [%rd30+-4];
and.b32 %r102, %r144, 252645135;
shr.u32 %r145, %r144, 4;
and.b32 %r106, %r145, 252645135;
mov.u32 %r125, 0;

dp4a.s32.s32 %r59, %r102, %r61, %r125;


dp4a.s32.s32 %r63, %r106, %r65, %r59;

and.b32 %r110, %r143, 252645135;
shr.u32 %r146, %r143, 4;
and.b32 %r114, %r146, 252645135;

dp4a.s32.s32 %r67, %r110, %r69, %r63;


dp4a.s32.s32 %r71, %r114, %r73, %r67;

ld.global.nc.u32 %r76, [%rd46+-8];
ld.global.nc.u32 %r77, [%rd27];

{mul.f16x2 %r75,%r76,%r77;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r75;
cvt.f32.f16 %f25, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r75;
cvt.f32.f16 %f26, high;}


cvt.rn.f32.s32 %f33, %r71;
mov.f32 %f34, 0f40000000;
div.approx.ftz.f32 %f35, %f26, %f34;
fma.rn.ftz.f32 %f36, %f25, %f33, %f35;
add.ftz.f32 %f113, %f36, %f113;
add.s64 %rd31, %rd47, %rd4;
ld.global.nc.u32 %r147, [%rd31];
ld.global.nc.u32 %r148, [%rd31+-4];
and.b32 %r123, %r148, 252645135;
shr.u32 %r149, %r148, 4;
and.b32 %r127, %r149, 252645135;

dp4a.s32.s32 %r80, %r123, %r61, %r125;


dp4a.s32.s32 %r84, %r127, %r65, %r80;

and.b32 %r131, %r147, 252645135;
shr.u32 %r150, %r147, 4;
and.b32 %r135, %r150, 252645135;

dp4a.s32.s32 %r88, %r131, %r69, %r84;


dp4a.s32.s32 %r92, %r135, %r73, %r88;

ld.global.nc.u32 %r97, [%rd47+-8];

{mul.f16x2 %r96,%r97,%r77;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r96;
cvt.f32.f16 %f27, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r96;
cvt.f32.f16 %f28, high;}


cvt.rn.f32.s32 %f37, %r92;
div.approx.ftz.f32 %f38, %f28, %f34;
fma.rn.ftz.f32 %f39, %f27, %f37, %f38;
add.ftz.f32 %f112, %f39, %f112;
cvt.u64.u32 %rd32, %r237;
add.s64 %rd33, %rd6, %rd32;
mul.lo.s64 %rd34, %rd33, 36;
add.s64 %rd35, %rd7, %rd34;
add.s64 %rd36, %rd35, %rd28;
ld.global.nc.u32 %r103, [%rd36+4];
ld.global.nc.u32 %r107, [%rd36+20];
ld.global.nc.u32 %r111, [%rd36+8];
ld.global.nc.u32 %r115, [%rd36+24];

dp4a.s32.s32 %r101, %r102, %r103, %r125;


dp4a.s32.s32 %r105, %r106, %r107, %r101;


dp4a.s32.s32 %r109, %r110, %r111, %r105;


dp4a.s32.s32 %r113, %r114, %r115, %r109;

ld.global.nc.u32 %r119, [%rd35];

{mul.f16x2 %r117,%r76,%r119;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r117;
cvt.f32.f16 %f29, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r117;
cvt.f32.f16 %f30, high;}


cvt.rn.f32.s32 %f40, %r113;
div.approx.ftz.f32 %f41, %f30, %f34;
fma.rn.ftz.f32 %f42, %f29, %f40, %f41;
add.ftz.f32 %f111, %f42, %f111;

dp4a.s32.s32 %r122, %r123, %r103, %r125;


dp4a.s32.s32 %r126, %r127, %r107, %r122;


dp4a.s32.s32 %r130, %r131, %r111, %r126;


dp4a.s32.s32 %r134, %r135, %r115, %r130;


{mul.f16x2 %r138,%r97,%r119;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r138;
cvt.f32.f16 %f31, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r138;
cvt.f32.f16 %f32, high;}


cvt.rn.f32.s32 %f43, %r134;
div.approx.ftz.f32 %f44, %f32, %f34;
fma.rn.ftz.f32 %f45, %f31, %f43, %f44;
add.ftz.f32 %f110, %f45, %f110;
add.s32 %r237, %r237, 64;
add.s64 %rd47, %rd47, 1280;
add.s64 %rd46, %rd46, 1280;
add.s32 %r238, %r238, 64;
setp.lt.s32 %p4, %r238, %r34;
@%p4 bra $L__BB49_2;

$L__BB49_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB49_5;

shl.b32 %r152, %r1, 9;
mov.u32 %r153, _ZZ13mul_mat_vec_qIL9ggml_type3ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r154, %r153, %r152;
shl.b32 %r155, %r30, 2;
add.s32 %r156, %r154, %r155;
st.shared.f32 [%r156+-512], %f113;
st.shared.f32 [%r156+-384], %f112;
st.shared.f32 [%r156+-256], %f111;
st.shared.f32 [%r156+-128], %f110;

$L__BB49_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB49_14;

mov.u32 %r157, %ctaid.z;
mov.u32 %r158, %ctaid.y;
mad.lo.s32 %r159, %r158, %r22, %r3;
mad.lo.s32 %r160, %r157, %r28, %r159;
cvt.u64.u32 %rd12, %r160;
shl.b32 %r162, %r30, 2;
mov.u32 %r163, _ZZ13mul_mat_vec_qIL9ggml_type3ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r11, %r163, %r162;
mov.u32 %r164, 2;
ld.shared.f32 %f46, [%r11];
add.ftz.f32 %f47, %f46, %f113;
ld.shared.f32 %f48, [%r11+512];
add.ftz.f32 %f49, %f48, %f47;
ld.shared.f32 %f50, [%r11+1024];
add.ftz.f32 %f51, %f50, %f49;
mov.b32 %r165, %f51;
mov.u32 %r166, 31;
mov.u32 %r167, 16;
mov.u32 %r168, -1;
shfl.sync.bfly.b32 %r169|%p7, %r165, %r167, %r166, %r168;
mov.b32 %f52, %r169;
add.ftz.f32 %f53, %f51, %f52;
mov.b32 %r170, %f53;
mov.u32 %r171, 8;
shfl.sync.bfly.b32 %r172|%p8, %r170, %r171, %r166, %r168;
mov.b32 %f54, %r172;
add.ftz.f32 %f55, %f53, %f54;
mov.b32 %r173, %f55;
mov.u32 %r174, 4;
shfl.sync.bfly.b32 %r175|%p9, %r173, %r174, %r166, %r168;
mov.b32 %f56, %r175;
add.ftz.f32 %f57, %f55, %f56;
mov.b32 %r176, %f57;
shfl.sync.bfly.b32 %r177|%p10, %r176, %r164, %r166, %r168;
mov.b32 %f58, %r177;
add.ftz.f32 %f59, %f57, %f58;
mov.b32 %r178, %f59;
mov.u32 %r179, 1;
shfl.sync.bfly.b32 %r180|%p11, %r178, %r179, %r166, %r168;
mov.b32 %f60, %r180;
add.ftz.f32 %f13, %f59, %f60;
setp.eq.s32 %p12, %r30, 0;
setp.lt.u32 %p13, %r3, %r16;
and.pred %p1, %p12, %p13;
mul.wide.u32 %rd37, %r160, 4;
add.s64 %rd13, %rd1, %rd37;
not.pred %p14, %p1;
@%p14 bra $L__BB49_8;

st.global.f32 [%rd13], %f13;

$L__BB49_8:
ld.shared.f32 %f61, [%r11+128];
add.ftz.f32 %f62, %f61, %f112;
ld.shared.f32 %f63, [%r11+640];
add.ftz.f32 %f64, %f63, %f62;
ld.shared.f32 %f65, [%r11+1152];
add.ftz.f32 %f66, %f65, %f64;
mov.b32 %r181, %f66;
shfl.sync.bfly.b32 %r185|%p15, %r181, %r167, %r166, %r168;
mov.b32 %f67, %r185;
add.ftz.f32 %f68, %f66, %f67;
mov.b32 %r186, %f68;
shfl.sync.bfly.b32 %r188|%p16, %r186, %r171, %r166, %r168;
mov.b32 %f69, %r188;
add.ftz.f32 %f70, %f68, %f69;
mov.b32 %r189, %f70;
shfl.sync.bfly.b32 %r191|%p17, %r189, %r174, %r166, %r168;
mov.b32 %f71, %r191;
add.ftz.f32 %f72, %f70, %f71;
mov.b32 %r192, %f72;
shfl.sync.bfly.b32 %r194|%p18, %r192, %r164, %r166, %r168;
mov.b32 %f73, %r194;
add.ftz.f32 %f74, %f72, %f73;
mov.b32 %r195, %f74;
shfl.sync.bfly.b32 %r197|%p19, %r195, %r179, %r166, %r168;
mov.b32 %f75, %r197;
add.ftz.f32 %f14, %f74, %f75;
setp.eq.s32 %p20, %r30, 1;
or.b32 %r201, %r3, 1;
setp.lt.u32 %p21, %r201, %r16;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB49_10;

st.global.f32 [%rd13+4], %f14;

$L__BB49_10:
ld.shared.f32 %f76, [%r11+256];
add.ftz.f32 %f77, %f76, %f111;
ld.shared.f32 %f78, [%r11+768];
add.ftz.f32 %f79, %f78, %f77;
ld.shared.f32 %f80, [%r11+1280];
add.ftz.f32 %f81, %f80, %f79;
mov.b32 %r202, %f81;
mov.u32 %r203, 31;
mov.u32 %r204, 16;
mov.u32 %r205, -1;
shfl.sync.bfly.b32 %r206|%p23, %r202, %r204, %r203, %r205;
mov.b32 %f82, %r206;
add.ftz.f32 %f83, %f81, %f82;
mov.b32 %r207, %f83;
mov.u32 %r208, 8;
shfl.sync.bfly.b32 %r209|%p24, %r207, %r208, %r203, %r205;
mov.b32 %f84, %r209;
add.ftz.f32 %f85, %f83, %f84;
mov.b32 %r210, %f85;
mov.u32 %r211, 4;
shfl.sync.bfly.b32 %r212|%p25, %r210, %r211, %r203, %r205;
mov.b32 %f86, %r212;
add.ftz.f32 %f87, %f85, %f86;
mov.b32 %r213, %f87;
mov.u32 %r214, 2;
shfl.sync.bfly.b32 %r215|%p26, %r213, %r214, %r203, %r205;
mov.b32 %f88, %r215;
add.ftz.f32 %f89, %f87, %f88;
mov.b32 %r216, %f89;
mov.u32 %r217, 1;
shfl.sync.bfly.b32 %r218|%p27, %r216, %r217, %r203, %r205;
mov.b32 %f90, %r218;
add.ftz.f32 %f15, %f89, %f90;
@%p14 bra $L__BB49_12;

cvt.u64.u32 %rd38, %r16;
add.s64 %rd39, %rd38, %rd12;
shl.b64 %rd40, %rd39, 2;
add.s64 %rd41, %rd1, %rd40;
st.global.f32 [%rd41], %f15;

$L__BB49_12:
ld.shared.f32 %f91, [%r11+384];
add.ftz.f32 %f92, %f91, %f110;
ld.shared.f32 %f93, [%r11+896];
add.ftz.f32 %f94, %f93, %f92;
ld.shared.f32 %f95, [%r11+1408];
add.ftz.f32 %f96, %f95, %f94;
mov.b32 %r219, %f96;
shfl.sync.bfly.b32 %r223|%p29, %r219, %r204, %r203, %r205;
mov.b32 %f97, %r223;
add.ftz.f32 %f98, %f96, %f97;
mov.b32 %r224, %f98;
shfl.sync.bfly.b32 %r226|%p30, %r224, %r208, %r203, %r205;
mov.b32 %f99, %r226;
add.ftz.f32 %f100, %f98, %f99;
mov.b32 %r227, %f100;
shfl.sync.bfly.b32 %r229|%p31, %r227, %r211, %r203, %r205;
mov.b32 %f101, %r229;
add.ftz.f32 %f102, %f100, %f101;
mov.b32 %r230, %f102;
shfl.sync.bfly.b32 %r232|%p32, %r230, %r214, %r203, %r205;
mov.b32 %f103, %r232;
add.ftz.f32 %f104, %f102, %f103;
mov.b32 %r233, %f104;
shfl.sync.bfly.b32 %r235|%p33, %r233, %r217, %r203, %r205;
mov.b32 %f105, %r235;
add.ftz.f32 %f16, %f104, %f105;
@%p22 bra $L__BB49_14;

add.s32 %r236, %r16, 1;
cvt.u64.u32 %rd42, %r236;
add.s64 %rd43, %rd42, %rd12;
shl.b64 %rd44, %rd43, 2;
add.s64 %rd45, %rd1, %rd44;
st.global.f32 [%rd45], %f16;

$L__BB49_14:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type3ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type3ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type3ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type3ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type3ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type3ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type3ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type3ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type3ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<47>;
.reg .f32 %f<170>;
.reg .b32 %r<319>;
.reg .b64 %rd<60>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type3ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[2304];

ld.param.u64 %rd14, [_Z13mul_mat_vec_qIL9ggml_type3ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type3ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type3ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type3ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r17, [_Z13mul_mat_vec_qIL9ggml_type3ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r18, [_Z13mul_mat_vec_qIL9ggml_type3ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r19, [_Z13mul_mat_vec_qIL9ggml_type3ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type3ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type3ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type3ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type3ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type3ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type3ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type3ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type3ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type3ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type3ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r33, %r1, 5;
mov.u32 %r34, %tid.x;
add.s32 %r2, %r33, %r34;
mov.u32 %r3, %ctaid.y;
mov.u32 %r4, %ctaid.z;
shr.u32 %r35, %r2, 31;
add.s32 %r36, %r2, %r35;
shr.s32 %r318, %r36, 1;
shr.u32 %r6, %r32, 5;
setp.ge.s32 %p3, %r318, %r6;
mov.f32 %f164, 0f00000000;
mov.f32 %f165, %f164;
mov.f32 %f166, %f164;
mov.f32 %f167, %f164;
mov.f32 %f168, %f164;
mov.f32 %f169, %f164;
@%p3 bra $L__BB50_3;

mul.lo.s32 %r37, %r3, %r24;
cvt.u64.u32 %rd17, %r37;
mul.lo.s32 %r38, %r4, %r30;
cvt.u64.u32 %rd18, %r38;
add.s64 %rd1, %rd18, %rd17;
mul.hi.u32 %r39, %r4, %r26;
add.s32 %r40, %r39, %r4;
shr.u32 %r41, %r40, %r27;
mul.lo.s32 %r42, %r41, %r29;
mul.hi.u32 %r43, %r3, %r20;
add.s32 %r44, %r43, %r3;
shr.u32 %r45, %r44, %r21;
mul.lo.s32 %r46, %r45, %r23;
cvta.to.global.u64 %rd19, %rd14;
and.b32 %r49, %r36, 2147483646;
sub.s32 %r50, %r2, %r49;
shl.b32 %r51, %r50, 1;
mov.u32 %r52, %ctaid.x;
shl.b32 %r53, %r52, 1;
cvt.s64.s32 %rd2, %r51;
shl.b32 %r54, %r18, 1;
add.s32 %r317, %r318, %r54;
add.s32 %r316, %r318, %r18;
add.s32 %r55, %r318, %r42;
add.s32 %r56, %r55, %r46;
or.b32 %r57, %r53, 1;
mad.lo.s32 %r58, %r17, %r57, %r56;
mul.wide.s32 %rd20, %r58, 20;
add.s64 %rd21, %rd19, %rd20;
add.s64 %rd59, %rd21, 8;
mul.wide.s32 %rd4, %r51, 4;
mad.lo.s32 %r59, %r53, %r17, %r318;
add.s32 %r60, %r59, %r42;
add.s32 %r61, %r60, %r46;
mul.wide.s32 %rd22, %r61, 20;
add.s64 %rd23, %rd19, %rd22;
add.s64 %rd58, %rd23, 8;
cvta.to.global.u64 %rd6, %rd15;
mov.f32 %f164, 0f00000000;
shl.b64 %rd28, %rd2, 2;

$L__BB50_2:
cvt.u64.u32 %rd24, %r318;
add.s64 %rd25, %rd1, %rd24;
mul.lo.s64 %rd26, %rd25, 36;
add.s64 %rd27, %rd6, %rd26;
add.s64 %rd29, %rd27, %rd28;
ld.global.nc.u32 %r64, [%rd29+4];
ld.global.nc.u32 %r68, [%rd29+20];
add.s64 %rd30, %rd58, %rd4;
ld.global.nc.u32 %r188, [%rd30];
ld.global.nc.u32 %r72, [%rd29+8];
ld.global.nc.u32 %r76, [%rd29+24];
ld.global.nc.u32 %r189, [%rd30+-4];
and.b32 %r147, %r189, 252645135;
shr.u32 %r190, %r189, 4;
and.b32 %r151, %r190, 252645135;
mov.u32 %r170, 0;

dp4a.s32.s32 %r62, %r147, %r64, %r170;


dp4a.s32.s32 %r66, %r151, %r68, %r62;

and.b32 %r155, %r188, 252645135;
shr.u32 %r191, %r188, 4;
and.b32 %r159, %r191, 252645135;

dp4a.s32.s32 %r70, %r155, %r72, %r66;


dp4a.s32.s32 %r74, %r159, %r76, %r70;

ld.global.nc.u32 %r79, [%rd58+-8];
ld.global.nc.u32 %r80, [%rd27];

{mul.f16x2 %r78,%r79,%r80;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r78;
cvt.f32.f16 %f37, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r78;
cvt.f32.f16 %f38, high;}


cvt.rn.f32.s32 %f49, %r74;
mov.f32 %f50, 0f40000000;
div.approx.ftz.f32 %f51, %f38, %f50;
fma.rn.ftz.f32 %f52, %f37, %f49, %f51;
add.ftz.f32 %f169, %f52, %f169;
add.s64 %rd31, %rd59, %rd4;
ld.global.nc.u32 %r192, [%rd31];
ld.global.nc.u32 %r193, [%rd31+-4];
and.b32 %r168, %r193, 252645135;
shr.u32 %r194, %r193, 4;
and.b32 %r172, %r194, 252645135;

dp4a.s32.s32 %r83, %r168, %r64, %r170;


dp4a.s32.s32 %r87, %r172, %r68, %r83;

and.b32 %r176, %r192, 252645135;
shr.u32 %r195, %r192, 4;
and.b32 %r180, %r195, 252645135;

dp4a.s32.s32 %r91, %r176, %r72, %r87;


dp4a.s32.s32 %r95, %r180, %r76, %r91;

ld.global.nc.u32 %r100, [%rd59+-8];

{mul.f16x2 %r99,%r100,%r80;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r99;
cvt.f32.f16 %f39, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r99;
cvt.f32.f16 %f40, high;}


cvt.rn.f32.s32 %f53, %r95;
div.approx.ftz.f32 %f54, %f40, %f50;
fma.rn.ftz.f32 %f55, %f39, %f53, %f54;
add.ftz.f32 %f168, %f55, %f168;
cvt.u64.u32 %rd32, %r316;
add.s64 %rd33, %rd1, %rd32;
mul.lo.s64 %rd34, %rd33, 36;
add.s64 %rd35, %rd6, %rd34;
add.s64 %rd36, %rd35, %rd28;
ld.global.nc.u32 %r106, [%rd36+4];
ld.global.nc.u32 %r110, [%rd36+20];
ld.global.nc.u32 %r114, [%rd36+8];
ld.global.nc.u32 %r118, [%rd36+24];

dp4a.s32.s32 %r104, %r147, %r106, %r170;


dp4a.s32.s32 %r108, %r151, %r110, %r104;


dp4a.s32.s32 %r112, %r155, %r114, %r108;


dp4a.s32.s32 %r116, %r159, %r118, %r112;

ld.global.nc.u32 %r122, [%rd35];

{mul.f16x2 %r120,%r79,%r122;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r120;
cvt.f32.f16 %f41, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r120;
cvt.f32.f16 %f42, high;}


cvt.rn.f32.s32 %f56, %r116;
div.approx.ftz.f32 %f57, %f42, %f50;
fma.rn.ftz.f32 %f58, %f41, %f56, %f57;
add.ftz.f32 %f167, %f58, %f167;

dp4a.s32.s32 %r125, %r168, %r106, %r170;


dp4a.s32.s32 %r129, %r172, %r110, %r125;


dp4a.s32.s32 %r133, %r176, %r114, %r129;


dp4a.s32.s32 %r137, %r180, %r118, %r133;


{mul.f16x2 %r141,%r100,%r122;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r141;
cvt.f32.f16 %f43, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r141;
cvt.f32.f16 %f44, high;}


cvt.rn.f32.s32 %f59, %r137;
div.approx.ftz.f32 %f60, %f44, %f50;
fma.rn.ftz.f32 %f61, %f43, %f59, %f60;
add.ftz.f32 %f166, %f61, %f166;
cvt.u64.u32 %rd37, %r317;
add.s64 %rd38, %rd1, %rd37;
mul.lo.s64 %rd39, %rd38, 36;
add.s64 %rd40, %rd6, %rd39;
add.s64 %rd41, %rd40, %rd28;
ld.global.nc.u32 %r148, [%rd41+4];
ld.global.nc.u32 %r152, [%rd41+20];
ld.global.nc.u32 %r156, [%rd41+8];
ld.global.nc.u32 %r160, [%rd41+24];

dp4a.s32.s32 %r146, %r147, %r148, %r170;


dp4a.s32.s32 %r150, %r151, %r152, %r146;


dp4a.s32.s32 %r154, %r155, %r156, %r150;


dp4a.s32.s32 %r158, %r159, %r160, %r154;

ld.global.nc.u32 %r164, [%rd40];

{mul.f16x2 %r162,%r79,%r164;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r162;
cvt.f32.f16 %f45, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r162;
cvt.f32.f16 %f46, high;}


cvt.rn.f32.s32 %f62, %r158;
div.approx.ftz.f32 %f63, %f46, %f50;
fma.rn.ftz.f32 %f64, %f45, %f62, %f63;
add.ftz.f32 %f165, %f64, %f165;

dp4a.s32.s32 %r167, %r168, %r148, %r170;


dp4a.s32.s32 %r171, %r172, %r152, %r167;


dp4a.s32.s32 %r175, %r176, %r156, %r171;


dp4a.s32.s32 %r179, %r180, %r160, %r175;


{mul.f16x2 %r183,%r100,%r164;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r183;
cvt.f32.f16 %f47, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r183;
cvt.f32.f16 %f48, high;}


cvt.rn.f32.s32 %f65, %r179;
div.approx.ftz.f32 %f66, %f48, %f50;
fma.rn.ftz.f32 %f67, %f47, %f65, %f66;
add.ftz.f32 %f164, %f67, %f164;
add.s32 %r317, %r317, 64;
add.s32 %r316, %r316, 64;
add.s64 %rd59, %rd59, 1280;
add.s64 %rd58, %rd58, 1280;
add.s32 %r318, %r318, 64;
setp.lt.s32 %p4, %r318, %r6;
@%p4 bra $L__BB50_2;

$L__BB50_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB50_5;

mov.u32 %r197, _ZZ13mul_mat_vec_qIL9ggml_type3ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r198, %r1, 768, %r197;
shl.b32 %r199, %r34, 2;
add.s32 %r200, %r198, %r199;
st.shared.f32 [%r200+-768], %f169;
st.shared.f32 [%r200+-640], %f168;
st.shared.f32 [%r200+-512], %f167;
st.shared.f32 [%r200+-384], %f166;
st.shared.f32 [%r200+-256], %f165;
st.shared.f32 [%r200+-128], %f164;

$L__BB50_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB50_18;

mov.u32 %r201, %ctaid.x;
shl.b32 %r202, %r201, 1;
mov.u32 %r203, 1;
mad.lo.s32 %r204, %r3, %r25, %r202;
mad.lo.s32 %r205, %r4, %r31, %r204;
cvt.u64.u32 %rd11, %r205;
shl.b32 %r207, %r34, 2;
mov.u32 %r208, _ZZ13mul_mat_vec_qIL9ggml_type3ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r15, %r208, %r207;
mov.u32 %r209, 2;
ld.shared.f32 %f68, [%r15];
add.ftz.f32 %f69, %f68, %f169;
ld.shared.f32 %f70, [%r15+768];
add.ftz.f32 %f71, %f70, %f69;
ld.shared.f32 %f72, [%r15+1536];
add.ftz.f32 %f73, %f72, %f71;
mov.b32 %r210, %f73;
mov.u32 %r211, 31;
mov.u32 %r212, 16;
mov.u32 %r213, -1;
shfl.sync.bfly.b32 %r214|%p7, %r210, %r212, %r211, %r213;
mov.b32 %f74, %r214;
add.ftz.f32 %f75, %f73, %f74;
mov.b32 %r215, %f75;
mov.u32 %r216, 8;
shfl.sync.bfly.b32 %r217|%p8, %r215, %r216, %r211, %r213;
mov.b32 %f76, %r217;
add.ftz.f32 %f77, %f75, %f76;
mov.b32 %r218, %f77;
mov.u32 %r219, 4;
shfl.sync.bfly.b32 %r220|%p9, %r218, %r219, %r211, %r213;
mov.b32 %f78, %r220;
add.ftz.f32 %f79, %f77, %f78;
mov.b32 %r221, %f79;
shfl.sync.bfly.b32 %r222|%p10, %r221, %r209, %r211, %r213;
mov.b32 %f80, %r222;
add.ftz.f32 %f81, %f79, %f80;
mov.b32 %r223, %f81;
shfl.sync.bfly.b32 %r224|%p11, %r223, %r203, %r211, %r213;
mov.b32 %f82, %r224;
add.ftz.f32 %f19, %f81, %f82;
setp.eq.s32 %p12, %r34, 0;
setp.lt.u32 %p13, %r202, %r19;
and.pred %p1, %p12, %p13;
cvta.to.global.u64 %rd42, %rd16;
mul.wide.u32 %rd43, %r205, 4;
add.s64 %rd12, %rd42, %rd43;
not.pred %p14, %p1;
@%p14 bra $L__BB50_8;

st.global.f32 [%rd12], %f19;

$L__BB50_8:
ld.shared.f32 %f83, [%r15+128];
add.ftz.f32 %f84, %f83, %f168;
ld.shared.f32 %f85, [%r15+896];
add.ftz.f32 %f86, %f85, %f84;
ld.shared.f32 %f87, [%r15+1664];
add.ftz.f32 %f88, %f87, %f86;
mov.b32 %r225, %f88;
shfl.sync.bfly.b32 %r229|%p15, %r225, %r212, %r211, %r213;
mov.b32 %f89, %r229;
add.ftz.f32 %f90, %f88, %f89;
mov.b32 %r230, %f90;
shfl.sync.bfly.b32 %r232|%p16, %r230, %r216, %r211, %r213;
mov.b32 %f91, %r232;
add.ftz.f32 %f92, %f90, %f91;
mov.b32 %r233, %f92;
shfl.sync.bfly.b32 %r235|%p17, %r233, %r219, %r211, %r213;
mov.b32 %f93, %r235;
add.ftz.f32 %f94, %f92, %f93;
mov.b32 %r236, %f94;
shfl.sync.bfly.b32 %r238|%p18, %r236, %r209, %r211, %r213;
mov.b32 %f95, %r238;
add.ftz.f32 %f96, %f94, %f95;
mov.b32 %r239, %f96;
shfl.sync.bfly.b32 %r241|%p19, %r239, %r203, %r211, %r213;
mov.b32 %f97, %r241;
add.ftz.f32 %f20, %f96, %f97;
setp.eq.s32 %p20, %r34, 1;
or.b32 %r245, %r202, 1;
setp.lt.u32 %p21, %r245, %r19;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB50_10;

st.global.f32 [%rd12+4], %f20;

$L__BB50_10:
ld.shared.f32 %f98, [%r15+256];
add.ftz.f32 %f99, %f98, %f167;
ld.shared.f32 %f100, [%r15+1024];
add.ftz.f32 %f101, %f100, %f99;
ld.shared.f32 %f102, [%r15+1792];
add.ftz.f32 %f103, %f102, %f101;
mov.b32 %r246, %f103;
mov.u32 %r247, 31;
mov.u32 %r248, 16;
mov.u32 %r249, -1;
shfl.sync.bfly.b32 %r250|%p23, %r246, %r248, %r247, %r249;
mov.b32 %f104, %r250;
add.ftz.f32 %f105, %f103, %f104;
mov.b32 %r251, %f105;
mov.u32 %r252, 8;
shfl.sync.bfly.b32 %r253|%p24, %r251, %r252, %r247, %r249;
mov.b32 %f106, %r253;
add.ftz.f32 %f107, %f105, %f106;
mov.b32 %r254, %f107;
mov.u32 %r255, 4;
shfl.sync.bfly.b32 %r256|%p25, %r254, %r255, %r247, %r249;
mov.b32 %f108, %r256;
add.ftz.f32 %f109, %f107, %f108;
mov.b32 %r257, %f109;
mov.u32 %r258, 2;
shfl.sync.bfly.b32 %r259|%p26, %r257, %r258, %r247, %r249;
mov.b32 %f110, %r259;
add.ftz.f32 %f111, %f109, %f110;
mov.b32 %r260, %f111;
mov.u32 %r261, 1;
shfl.sync.bfly.b32 %r262|%p27, %r260, %r261, %r247, %r249;
mov.b32 %f112, %r262;
add.ftz.f32 %f21, %f111, %f112;
@%p14 bra $L__BB50_12;

cvt.u64.u32 %rd44, %r19;
add.s64 %rd45, %rd44, %rd11;
shl.b64 %rd47, %rd45, 2;
add.s64 %rd48, %rd42, %rd47;
st.global.f32 [%rd48], %f21;

$L__BB50_12:
ld.shared.f32 %f113, [%r15+384];
add.ftz.f32 %f114, %f113, %f166;
ld.shared.f32 %f115, [%r15+1152];
add.ftz.f32 %f116, %f115, %f114;
ld.shared.f32 %f117, [%r15+1920];
add.ftz.f32 %f118, %f117, %f116;
mov.b32 %r263, %f118;
shfl.sync.bfly.b32 %r267|%p29, %r263, %r248, %r247, %r249;
mov.b32 %f119, %r267;
add.ftz.f32 %f120, %f118, %f119;
mov.b32 %r268, %f120;
shfl.sync.bfly.b32 %r270|%p30, %r268, %r252, %r247, %r249;
mov.b32 %f121, %r270;
add.ftz.f32 %f122, %f120, %f121;
mov.b32 %r271, %f122;
shfl.sync.bfly.b32 %r273|%p31, %r271, %r255, %r247, %r249;
mov.b32 %f123, %r273;
add.ftz.f32 %f124, %f122, %f123;
mov.b32 %r274, %f124;
shfl.sync.bfly.b32 %r276|%p32, %r274, %r258, %r247, %r249;
mov.b32 %f125, %r276;
add.ftz.f32 %f126, %f124, %f125;
mov.b32 %r277, %f126;
shfl.sync.bfly.b32 %r279|%p33, %r277, %r261, %r247, %r249;
mov.b32 %f127, %r279;
add.ftz.f32 %f22, %f126, %f127;
@%p22 bra $L__BB50_14;

add.s32 %r280, %r19, 1;
cvt.u64.u32 %rd49, %r280;
add.s64 %rd50, %rd49, %rd11;
shl.b64 %rd52, %rd50, 2;
add.s64 %rd53, %rd42, %rd52;
st.global.f32 [%rd53], %f22;

$L__BB50_14:
shl.b32 %r281, %r19, 1;
mov.u32 %r282, 1;
ld.shared.f32 %f128, [%r15+512];
add.ftz.f32 %f129, %f128, %f165;
ld.shared.f32 %f130, [%r15+1280];
add.ftz.f32 %f131, %f130, %f129;
ld.shared.f32 %f132, [%r15+2048];
add.ftz.f32 %f133, %f132, %f131;
mov.b32 %r283, %f133;
mov.u32 %r284, 31;
mov.u32 %r285, 16;
mov.u32 %r286, -1;
shfl.sync.bfly.b32 %r287|%p35, %r283, %r285, %r284, %r286;
mov.b32 %f134, %r287;
add.ftz.f32 %f135, %f133, %f134;
mov.b32 %r288, %f135;
mov.u32 %r289, 8;
shfl.sync.bfly.b32 %r290|%p36, %r288, %r289, %r284, %r286;
mov.b32 %f136, %r290;
add.ftz.f32 %f137, %f135, %f136;
mov.b32 %r291, %f137;
mov.u32 %r292, 4;
shfl.sync.bfly.b32 %r293|%p37, %r291, %r292, %r284, %r286;
mov.b32 %f138, %r293;
add.ftz.f32 %f139, %f137, %f138;
mov.b32 %r294, %f139;
mov.u32 %r295, 2;
shfl.sync.bfly.b32 %r296|%p38, %r294, %r295, %r284, %r286;
mov.b32 %f140, %r296;
add.ftz.f32 %f141, %f139, %f140;
mov.b32 %r297, %f141;
shfl.sync.bfly.b32 %r298|%p39, %r297, %r282, %r284, %r286;
mov.b32 %f142, %r298;
add.ftz.f32 %f23, %f141, %f142;
cvt.u64.u32 %rd54, %r281;
add.s64 %rd55, %rd54, %rd11;
shl.b64 %rd57, %rd55, 2;
add.s64 %rd13, %rd42, %rd57;
@%p14 bra $L__BB50_16;

st.global.f32 [%rd13], %f23;

$L__BB50_16:
ld.shared.f32 %f143, [%r15+640];
add.ftz.f32 %f144, %f143, %f164;
ld.shared.f32 %f145, [%r15+1408];
add.ftz.f32 %f146, %f145, %f144;
ld.shared.f32 %f147, [%r15+2176];
add.ftz.f32 %f148, %f147, %f146;
mov.b32 %r299, %f148;
shfl.sync.bfly.b32 %r303|%p41, %r299, %r285, %r284, %r286;
mov.b32 %f149, %r303;
add.ftz.f32 %f150, %f148, %f149;
mov.b32 %r304, %f150;
shfl.sync.bfly.b32 %r306|%p42, %r304, %r289, %r284, %r286;
mov.b32 %f151, %r306;
add.ftz.f32 %f152, %f150, %f151;
mov.b32 %r307, %f152;
shfl.sync.bfly.b32 %r309|%p43, %r307, %r292, %r284, %r286;
mov.b32 %f153, %r309;
add.ftz.f32 %f154, %f152, %f153;
mov.b32 %r310, %f154;
shfl.sync.bfly.b32 %r312|%p44, %r310, %r295, %r284, %r286;
mov.b32 %f155, %r312;
add.ftz.f32 %f156, %f154, %f155;
mov.b32 %r313, %f156;
shfl.sync.bfly.b32 %r315|%p45, %r313, %r282, %r284, %r286;
mov.b32 %f157, %r315;
add.ftz.f32 %f24, %f156, %f157;
@%p22 bra $L__BB50_18;

st.global.f32 [%rd13+4], %f24;

$L__BB50_18:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type3ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type3ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type3ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type3ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type3ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type3ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type3ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type3ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type3ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<59>;
.reg .f32 %f<226>;
.reg .b32 %r<402>;
.reg .b64 %rd<75>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type3ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[3072];

ld.param.u64 %rd14, [_Z13mul_mat_vec_qIL9ggml_type3ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type3ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type3ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type3ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type3ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type3ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type3ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type3ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type3ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type3ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type3ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type3ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type3ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type3ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type3ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type3ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type3ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r36, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r36, %r2;
mov.u32 %r37, %ctaid.x;
shl.b32 %r4, %r37, 1;
shr.u32 %r38, %r3, 31;
add.s32 %r39, %r3, %r38;
shr.s32 %r401, %r39, 1;
shr.u32 %r6, %r35, 5;
setp.ge.s32 %p3, %r401, %r6;
mov.f32 %f218, 0f00000000;
mov.f32 %f219, %f218;
mov.f32 %f220, %f218;
mov.f32 %f221, %f218;
mov.f32 %f222, %f218;
mov.f32 %f223, %f218;
mov.f32 %f224, %f218;
mov.f32 %f225, %f218;
@%p3 bra $L__BB51_3;

mov.u32 %r40, %ctaid.y;
mov.u32 %r41, %ctaid.z;
mul.lo.s32 %r42, %r40, %r27;
cvt.u64.u32 %rd17, %r42;
mul.lo.s32 %r43, %r41, %r33;
cvt.u64.u32 %rd18, %r43;
add.s64 %rd1, %rd18, %rd17;
mul.hi.u32 %r44, %r41, %r29;
add.s32 %r45, %r44, %r41;
shr.u32 %r46, %r45, %r30;
mul.lo.s32 %r47, %r46, %r32;
mul.hi.u32 %r48, %r40, %r23;
add.s32 %r49, %r48, %r40;
shr.u32 %r50, %r49, %r24;
mul.lo.s32 %r51, %r50, %r26;
cvta.to.global.u64 %rd19, %rd14;
and.b32 %r54, %r39, 2147483646;
sub.s32 %r55, %r3, %r54;
shl.b32 %r56, %r55, 1;
cvt.s64.s32 %rd2, %r56;
shl.b32 %r57, %r21, 1;
add.s32 %r400, %r401, %r57;
mad.lo.s32 %r399, %r21, 3, %r401;
add.s32 %r398, %r401, %r21;
add.s32 %r58, %r401, %r47;
add.s32 %r59, %r58, %r51;
add.s32 %r60, %r4, 1;
mad.lo.s32 %r61, %r20, %r60, %r59;
mul.wide.s32 %rd20, %r61, 20;
add.s64 %rd21, %rd19, %rd20;
add.s64 %rd74, %rd21, 8;
mul.wide.s32 %rd4, %r56, 4;
mad.lo.s32 %r62, %r4, %r20, %r401;
add.s32 %r63, %r62, %r47;
add.s32 %r64, %r63, %r51;
mul.wide.s32 %rd22, %r64, 20;
add.s64 %rd23, %rd19, %rd22;
add.s64 %rd73, %rd23, 8;
cvta.to.global.u64 %rd6, %rd15;
mov.f32 %f218, 0f00000000;
shl.b64 %rd28, %rd2, 2;

$L__BB51_2:
cvt.u64.u32 %rd24, %r401;
add.s64 %rd25, %rd1, %rd24;
mul.lo.s64 %rd26, %rd25, 36;
add.s64 %rd27, %rd6, %rd26;
add.s64 %rd29, %rd27, %rd28;
ld.global.nc.u32 %r67, [%rd29+4];
ld.global.nc.u32 %r71, [%rd29+20];
add.s64 %rd30, %rd73, %rd4;
ld.global.nc.u32 %r233, [%rd30];
ld.global.nc.u32 %r75, [%rd29+8];
ld.global.nc.u32 %r79, [%rd29+24];
ld.global.nc.u32 %r234, [%rd30+-4];
and.b32 %r192, %r234, 252645135;
shr.u32 %r235, %r234, 4;
and.b32 %r196, %r235, 252645135;
mov.u32 %r215, 0;

dp4a.s32.s32 %r65, %r192, %r67, %r215;


dp4a.s32.s32 %r69, %r196, %r71, %r65;

and.b32 %r200, %r233, 252645135;
shr.u32 %r236, %r233, 4;
and.b32 %r204, %r236, 252645135;

dp4a.s32.s32 %r73, %r200, %r75, %r69;


dp4a.s32.s32 %r77, %r204, %r79, %r73;

ld.global.nc.u32 %r82, [%rd73+-8];
ld.global.nc.u32 %r83, [%rd27];

{mul.f16x2 %r81,%r82,%r83;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r81;
cvt.f32.f16 %f49, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r81;
cvt.f32.f16 %f50, high;}


cvt.rn.f32.s32 %f65, %r77;
mov.f32 %f66, 0f40000000;
div.approx.ftz.f32 %f67, %f50, %f66;
fma.rn.ftz.f32 %f68, %f49, %f65, %f67;
add.ftz.f32 %f225, %f68, %f225;
add.s64 %rd31, %rd74, %rd4;
ld.global.nc.u32 %r237, [%rd31];
ld.global.nc.u32 %r238, [%rd31+-4];
and.b32 %r213, %r238, 252645135;
shr.u32 %r239, %r238, 4;
and.b32 %r217, %r239, 252645135;

dp4a.s32.s32 %r86, %r213, %r67, %r215;


dp4a.s32.s32 %r90, %r217, %r71, %r86;

and.b32 %r221, %r237, 252645135;
shr.u32 %r240, %r237, 4;
and.b32 %r225, %r240, 252645135;

dp4a.s32.s32 %r94, %r221, %r75, %r90;


dp4a.s32.s32 %r98, %r225, %r79, %r94;

ld.global.nc.u32 %r103, [%rd74+-8];

{mul.f16x2 %r102,%r103,%r83;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r102;
cvt.f32.f16 %f51, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r102;
cvt.f32.f16 %f52, high;}


cvt.rn.f32.s32 %f69, %r98;
div.approx.ftz.f32 %f70, %f52, %f66;
fma.rn.ftz.f32 %f71, %f51, %f69, %f70;
add.ftz.f32 %f224, %f71, %f224;
cvt.u64.u32 %rd32, %r398;
add.s64 %rd33, %rd1, %rd32;
mul.lo.s64 %rd34, %rd33, 36;
add.s64 %rd35, %rd6, %rd34;
add.s64 %rd36, %rd35, %rd28;
ld.global.nc.u32 %r109, [%rd36+4];
ld.global.nc.u32 %r113, [%rd36+20];
ld.global.nc.u32 %r117, [%rd36+8];
ld.global.nc.u32 %r121, [%rd36+24];

dp4a.s32.s32 %r107, %r192, %r109, %r215;


dp4a.s32.s32 %r111, %r196, %r113, %r107;


dp4a.s32.s32 %r115, %r200, %r117, %r111;


dp4a.s32.s32 %r119, %r204, %r121, %r115;

ld.global.nc.u32 %r125, [%rd35];

{mul.f16x2 %r123,%r82,%r125;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r123;
cvt.f32.f16 %f53, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r123;
cvt.f32.f16 %f54, high;}


cvt.rn.f32.s32 %f72, %r119;
div.approx.ftz.f32 %f73, %f54, %f66;
fma.rn.ftz.f32 %f74, %f53, %f72, %f73;
add.ftz.f32 %f223, %f74, %f223;

dp4a.s32.s32 %r128, %r213, %r109, %r215;


dp4a.s32.s32 %r132, %r217, %r113, %r128;


dp4a.s32.s32 %r136, %r221, %r117, %r132;


dp4a.s32.s32 %r140, %r225, %r121, %r136;


{mul.f16x2 %r144,%r103,%r125;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r144;
cvt.f32.f16 %f55, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r144;
cvt.f32.f16 %f56, high;}


cvt.rn.f32.s32 %f75, %r140;
div.approx.ftz.f32 %f76, %f56, %f66;
fma.rn.ftz.f32 %f77, %f55, %f75, %f76;
add.ftz.f32 %f222, %f77, %f222;
cvt.u64.u32 %rd37, %r400;
add.s64 %rd38, %rd1, %rd37;
mul.lo.s64 %rd39, %rd38, 36;
add.s64 %rd40, %rd6, %rd39;
add.s64 %rd41, %rd40, %rd28;
ld.global.nc.u32 %r151, [%rd41+4];
ld.global.nc.u32 %r155, [%rd41+20];
ld.global.nc.u32 %r159, [%rd41+8];
ld.global.nc.u32 %r163, [%rd41+24];

dp4a.s32.s32 %r149, %r192, %r151, %r215;


dp4a.s32.s32 %r153, %r196, %r155, %r149;


dp4a.s32.s32 %r157, %r200, %r159, %r153;


dp4a.s32.s32 %r161, %r204, %r163, %r157;

ld.global.nc.u32 %r167, [%rd40];

{mul.f16x2 %r165,%r82,%r167;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r165;
cvt.f32.f16 %f57, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r165;
cvt.f32.f16 %f58, high;}


cvt.rn.f32.s32 %f78, %r161;
div.approx.ftz.f32 %f79, %f58, %f66;
fma.rn.ftz.f32 %f80, %f57, %f78, %f79;
add.ftz.f32 %f221, %f80, %f221;

dp4a.s32.s32 %r170, %r213, %r151, %r215;


dp4a.s32.s32 %r174, %r217, %r155, %r170;


dp4a.s32.s32 %r178, %r221, %r159, %r174;


dp4a.s32.s32 %r182, %r225, %r163, %r178;


{mul.f16x2 %r186,%r103,%r167;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r186;
cvt.f32.f16 %f59, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r186;
cvt.f32.f16 %f60, high;}


cvt.rn.f32.s32 %f81, %r182;
div.approx.ftz.f32 %f82, %f60, %f66;
fma.rn.ftz.f32 %f83, %f59, %f81, %f82;
add.ftz.f32 %f220, %f83, %f220;
cvt.u64.u32 %rd42, %r399;
add.s64 %rd43, %rd1, %rd42;
mul.lo.s64 %rd44, %rd43, 36;
add.s64 %rd45, %rd6, %rd44;
add.s64 %rd46, %rd45, %rd28;
ld.global.nc.u32 %r193, [%rd46+4];
ld.global.nc.u32 %r197, [%rd46+20];
ld.global.nc.u32 %r201, [%rd46+8];
ld.global.nc.u32 %r205, [%rd46+24];

dp4a.s32.s32 %r191, %r192, %r193, %r215;


dp4a.s32.s32 %r195, %r196, %r197, %r191;


dp4a.s32.s32 %r199, %r200, %r201, %r195;


dp4a.s32.s32 %r203, %r204, %r205, %r199;

ld.global.nc.u32 %r209, [%rd45];

{mul.f16x2 %r207,%r82,%r209;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r207;
cvt.f32.f16 %f61, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r207;
cvt.f32.f16 %f62, high;}


cvt.rn.f32.s32 %f84, %r203;
div.approx.ftz.f32 %f85, %f62, %f66;
fma.rn.ftz.f32 %f86, %f61, %f84, %f85;
add.ftz.f32 %f219, %f86, %f219;

dp4a.s32.s32 %r212, %r213, %r193, %r215;


dp4a.s32.s32 %r216, %r217, %r197, %r212;


dp4a.s32.s32 %r220, %r221, %r201, %r216;


dp4a.s32.s32 %r224, %r225, %r205, %r220;


{mul.f16x2 %r228,%r103,%r209;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r228;
cvt.f32.f16 %f63, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r228;
cvt.f32.f16 %f64, high;}


cvt.rn.f32.s32 %f87, %r224;
div.approx.ftz.f32 %f88, %f64, %f66;
fma.rn.ftz.f32 %f89, %f63, %f87, %f88;
add.ftz.f32 %f218, %f89, %f218;
add.s32 %r400, %r400, 64;
add.s32 %r399, %r399, 64;
add.s32 %r398, %r398, 64;
add.s64 %rd74, %rd74, 1280;
add.s64 %rd73, %rd73, 1280;
add.s32 %r401, %r401, 64;
setp.lt.s32 %p4, %r401, %r6;
@%p4 bra $L__BB51_2;

$L__BB51_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB51_5;

shl.b32 %r241, %r1, 10;
mov.u32 %r242, _ZZ13mul_mat_vec_qIL9ggml_type3ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r243, %r242, %r241;
shl.b32 %r244, %r2, 2;
add.s32 %r245, %r243, %r244;
st.shared.f32 [%r245+-1024], %f225;
st.shared.f32 [%r245+-896], %f224;
st.shared.f32 [%r245+-768], %f223;
st.shared.f32 [%r245+-640], %f222;
st.shared.f32 [%r245+-512], %f221;
st.shared.f32 [%r245+-384], %f220;
st.shared.f32 [%r245+-256], %f219;
st.shared.f32 [%r245+-128], %f218;

$L__BB51_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB51_22;

mov.u32 %r246, %ctaid.z;
mov.u32 %r247, %ctaid.y;
mad.lo.s32 %r248, %r247, %r28, %r4;
mad.lo.s32 %r249, %r246, %r34, %r248;
cvt.u64.u32 %rd11, %r249;
shl.b32 %r250, %r2, 2;
mov.u32 %r251, _ZZ13mul_mat_vec_qIL9ggml_type3ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r252, %r251, %r250;
mov.u32 %r253, 2;
ld.shared.f32 %f90, [%r252];
add.ftz.f32 %f91, %f90, %f225;
ld.shared.f32 %f92, [%r252+1024];
add.ftz.f32 %f93, %f92, %f91;
ld.shared.f32 %f94, [%r252+2048];
add.ftz.f32 %f95, %f94, %f93;
mov.b32 %r254, %f95;
mov.u32 %r255, 31;
mov.u32 %r256, 16;
mov.u32 %r257, -1;
shfl.sync.bfly.b32 %r258|%p7, %r254, %r256, %r255, %r257;
mov.b32 %f96, %r258;
add.ftz.f32 %f97, %f95, %f96;
mov.b32 %r259, %f97;
mov.u32 %r260, 8;
shfl.sync.bfly.b32 %r261|%p8, %r259, %r260, %r255, %r257;
mov.b32 %f98, %r261;
add.ftz.f32 %f99, %f97, %f98;
mov.b32 %r262, %f99;
mov.u32 %r263, 4;
shfl.sync.bfly.b32 %r264|%p9, %r262, %r263, %r255, %r257;
mov.b32 %f100, %r264;
add.ftz.f32 %f101, %f99, %f100;
mov.b32 %r265, %f101;
shfl.sync.bfly.b32 %r266|%p10, %r265, %r253, %r255, %r257;
mov.b32 %f102, %r266;
add.ftz.f32 %f103, %f101, %f102;
mov.b32 %r267, %f103;
mov.u32 %r268, 1;
shfl.sync.bfly.b32 %r269|%p11, %r267, %r268, %r255, %r257;
mov.b32 %f104, %r269;
add.ftz.f32 %f25, %f103, %f104;
setp.lt.u32 %p12, %r4, %r22;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
cvta.to.global.u64 %rd47, %rd16;
mul.wide.u32 %rd48, %r249, 4;
add.s64 %rd12, %rd47, %rd48;
not.pred %p14, %p1;
@%p14 bra $L__BB51_8;

st.global.f32 [%rd12], %f25;

$L__BB51_8:
ld.shared.f32 %f105, [%r252+128];
add.ftz.f32 %f106, %f105, %f224;
ld.shared.f32 %f107, [%r252+1152];
add.ftz.f32 %f108, %f107, %f106;
ld.shared.f32 %f109, [%r252+2176];
add.ftz.f32 %f110, %f109, %f108;
mov.b32 %r274, %f110;
shfl.sync.bfly.b32 %r278|%p15, %r274, %r256, %r255, %r257;
mov.b32 %f111, %r278;
add.ftz.f32 %f112, %f110, %f111;
mov.b32 %r279, %f112;
shfl.sync.bfly.b32 %r281|%p16, %r279, %r260, %r255, %r257;
mov.b32 %f113, %r281;
add.ftz.f32 %f114, %f112, %f113;
mov.b32 %r282, %f114;
shfl.sync.bfly.b32 %r284|%p17, %r282, %r263, %r255, %r257;
mov.b32 %f115, %r284;
add.ftz.f32 %f116, %f114, %f115;
mov.b32 %r285, %f116;
shfl.sync.bfly.b32 %r286|%p18, %r285, %r253, %r255, %r257;
mov.b32 %f117, %r286;
add.ftz.f32 %f118, %f116, %f117;
mov.b32 %r287, %f118;
shfl.sync.bfly.b32 %r289|%p19, %r287, %r268, %r255, %r257;
mov.b32 %f119, %r289;
add.ftz.f32 %f26, %f118, %f119;
setp.eq.s32 %p20, %r2, 1;
or.b32 %r292, %r4, 1;
setp.lt.u32 %p21, %r292, %r22;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB51_10;

st.global.f32 [%rd12+4], %f26;

$L__BB51_10:
ld.shared.f32 %f120, [%r252+256];
add.ftz.f32 %f121, %f120, %f223;
ld.shared.f32 %f122, [%r252+1280];
add.ftz.f32 %f123, %f122, %f121;
ld.shared.f32 %f124, [%r252+2304];
add.ftz.f32 %f125, %f124, %f123;
mov.b32 %r293, %f125;
mov.u32 %r294, 31;
mov.u32 %r295, 16;
mov.u32 %r296, -1;
shfl.sync.bfly.b32 %r297|%p23, %r293, %r295, %r294, %r296;
mov.b32 %f126, %r297;
add.ftz.f32 %f127, %f125, %f126;
mov.b32 %r298, %f127;
mov.u32 %r299, 8;
shfl.sync.bfly.b32 %r300|%p24, %r298, %r299, %r294, %r296;
mov.b32 %f128, %r300;
add.ftz.f32 %f129, %f127, %f128;
mov.b32 %r301, %f129;
mov.u32 %r302, 4;
shfl.sync.bfly.b32 %r303|%p25, %r301, %r302, %r294, %r296;
mov.b32 %f130, %r303;
add.ftz.f32 %f131, %f129, %f130;
mov.b32 %r304, %f131;
mov.u32 %r305, 2;
shfl.sync.bfly.b32 %r306|%p26, %r304, %r305, %r294, %r296;
mov.b32 %f132, %r306;
add.ftz.f32 %f133, %f131, %f132;
mov.b32 %r307, %f133;
mov.u32 %r308, 1;
shfl.sync.bfly.b32 %r309|%p27, %r307, %r308, %r294, %r296;
mov.b32 %f134, %r309;
add.ftz.f32 %f27, %f133, %f134;
@%p14 bra $L__BB51_12;

cvt.u64.u32 %rd49, %r22;
add.s64 %rd50, %rd49, %rd11;
shl.b64 %rd52, %rd50, 2;
add.s64 %rd53, %rd47, %rd52;
st.global.f32 [%rd53], %f27;

$L__BB51_12:
ld.shared.f32 %f135, [%r252+384];
add.ftz.f32 %f136, %f135, %f222;
ld.shared.f32 %f137, [%r252+1408];
add.ftz.f32 %f138, %f137, %f136;
ld.shared.f32 %f139, [%r252+2432];
add.ftz.f32 %f140, %f139, %f138;
mov.b32 %r310, %f140;
shfl.sync.bfly.b32 %r314|%p29, %r310, %r295, %r294, %r296;
mov.b32 %f141, %r314;
add.ftz.f32 %f142, %f140, %f141;
mov.b32 %r315, %f142;
shfl.sync.bfly.b32 %r317|%p30, %r315, %r299, %r294, %r296;
mov.b32 %f143, %r317;
add.ftz.f32 %f144, %f142, %f143;
mov.b32 %r318, %f144;
shfl.sync.bfly.b32 %r320|%p31, %r318, %r302, %r294, %r296;
mov.b32 %f145, %r320;
add.ftz.f32 %f146, %f144, %f145;
mov.b32 %r321, %f146;
shfl.sync.bfly.b32 %r323|%p32, %r321, %r305, %r294, %r296;
mov.b32 %f147, %r323;
add.ftz.f32 %f148, %f146, %f147;
mov.b32 %r324, %f148;
shfl.sync.bfly.b32 %r326|%p33, %r324, %r308, %r294, %r296;
mov.b32 %f149, %r326;
add.ftz.f32 %f28, %f148, %f149;
@%p22 bra $L__BB51_14;

add.s32 %r327, %r22, 1;
cvt.u64.u32 %rd54, %r327;
add.s64 %rd55, %rd54, %rd11;
shl.b64 %rd57, %rd55, 2;
add.s64 %rd58, %rd47, %rd57;
st.global.f32 [%rd58], %f28;

$L__BB51_14:
shl.b32 %r328, %r22, 1;
mov.u32 %r329, 1;
ld.shared.f32 %f150, [%r252+512];
add.ftz.f32 %f151, %f150, %f221;
ld.shared.f32 %f152, [%r252+1536];
add.ftz.f32 %f153, %f152, %f151;
ld.shared.f32 %f154, [%r252+2560];
add.ftz.f32 %f155, %f154, %f153;
mov.b32 %r330, %f155;
mov.u32 %r331, 31;
mov.u32 %r332, 16;
mov.u32 %r333, -1;
shfl.sync.bfly.b32 %r334|%p35, %r330, %r332, %r331, %r333;
mov.b32 %f156, %r334;
add.ftz.f32 %f157, %f155, %f156;
mov.b32 %r335, %f157;
mov.u32 %r336, 8;
shfl.sync.bfly.b32 %r337|%p36, %r335, %r336, %r331, %r333;
mov.b32 %f158, %r337;
add.ftz.f32 %f159, %f157, %f158;
mov.b32 %r338, %f159;
mov.u32 %r339, 4;
shfl.sync.bfly.b32 %r340|%p37, %r338, %r339, %r331, %r333;
mov.b32 %f160, %r340;
add.ftz.f32 %f161, %f159, %f160;
mov.b32 %r341, %f161;
mov.u32 %r342, 2;
shfl.sync.bfly.b32 %r343|%p38, %r341, %r342, %r331, %r333;
mov.b32 %f162, %r343;
add.ftz.f32 %f163, %f161, %f162;
mov.b32 %r344, %f163;
shfl.sync.bfly.b32 %r345|%p39, %r344, %r329, %r331, %r333;
mov.b32 %f164, %r345;
add.ftz.f32 %f29, %f163, %f164;
cvt.u64.u32 %rd59, %r328;
add.s64 %rd60, %rd59, %rd11;
shl.b64 %rd62, %rd60, 2;
add.s64 %rd13, %rd47, %rd62;
@%p14 bra $L__BB51_16;

st.global.f32 [%rd13], %f29;

$L__BB51_16:
ld.shared.f32 %f165, [%r252+640];
add.ftz.f32 %f166, %f165, %f220;
ld.shared.f32 %f167, [%r252+1664];
add.ftz.f32 %f168, %f167, %f166;
ld.shared.f32 %f169, [%r252+2688];
add.ftz.f32 %f170, %f169, %f168;
mov.b32 %r346, %f170;
shfl.sync.bfly.b32 %r350|%p41, %r346, %r332, %r331, %r333;
mov.b32 %f171, %r350;
add.ftz.f32 %f172, %f170, %f171;
mov.b32 %r351, %f172;
shfl.sync.bfly.b32 %r353|%p42, %r351, %r336, %r331, %r333;
mov.b32 %f173, %r353;
add.ftz.f32 %f174, %f172, %f173;
mov.b32 %r354, %f174;
shfl.sync.bfly.b32 %r356|%p43, %r354, %r339, %r331, %r333;
mov.b32 %f175, %r356;
add.ftz.f32 %f176, %f174, %f175;
mov.b32 %r357, %f176;
shfl.sync.bfly.b32 %r359|%p44, %r357, %r342, %r331, %r333;
mov.b32 %f177, %r359;
add.ftz.f32 %f178, %f176, %f177;
mov.b32 %r360, %f178;
shfl.sync.bfly.b32 %r362|%p45, %r360, %r329, %r331, %r333;
mov.b32 %f179, %r362;
add.ftz.f32 %f30, %f178, %f179;
@%p22 bra $L__BB51_18;

st.global.f32 [%rd13+4], %f30;

$L__BB51_18:
mul.lo.s32 %r19, %r22, 3;
ld.shared.f32 %f180, [%r252+768];
add.ftz.f32 %f181, %f180, %f219;
ld.shared.f32 %f182, [%r252+1792];
add.ftz.f32 %f183, %f182, %f181;
ld.shared.f32 %f184, [%r252+2816];
add.ftz.f32 %f185, %f184, %f183;
mov.b32 %r363, %f185;
mov.u32 %r364, 31;
mov.u32 %r365, 16;
mov.u32 %r366, -1;
shfl.sync.bfly.b32 %r367|%p47, %r363, %r365, %r364, %r366;
mov.b32 %f186, %r367;
add.ftz.f32 %f187, %f185, %f186;
mov.b32 %r368, %f187;
mov.u32 %r369, 8;
shfl.sync.bfly.b32 %r370|%p48, %r368, %r369, %r364, %r366;
mov.b32 %f188, %r370;
add.ftz.f32 %f189, %f187, %f188;
mov.b32 %r371, %f189;
mov.u32 %r372, 4;
shfl.sync.bfly.b32 %r373|%p49, %r371, %r372, %r364, %r366;
mov.b32 %f190, %r373;
add.ftz.f32 %f191, %f189, %f190;
mov.b32 %r374, %f191;
mov.u32 %r375, 2;
shfl.sync.bfly.b32 %r376|%p50, %r374, %r375, %r364, %r366;
mov.b32 %f192, %r376;
add.ftz.f32 %f193, %f191, %f192;
mov.b32 %r377, %f193;
mov.u32 %r378, 1;
shfl.sync.bfly.b32 %r379|%p51, %r377, %r378, %r364, %r366;
mov.b32 %f194, %r379;
add.ftz.f32 %f31, %f193, %f194;
@%p14 bra $L__BB51_20;

cvt.u64.u32 %rd63, %r19;
add.s64 %rd64, %rd63, %rd11;
shl.b64 %rd66, %rd64, 2;
add.s64 %rd67, %rd47, %rd66;
st.global.f32 [%rd67], %f31;

$L__BB51_20:
ld.shared.f32 %f195, [%r252+896];
add.ftz.f32 %f196, %f195, %f218;
ld.shared.f32 %f197, [%r252+1920];
add.ftz.f32 %f198, %f197, %f196;
ld.shared.f32 %f199, [%r252+2944];
add.ftz.f32 %f200, %f199, %f198;
mov.b32 %r380, %f200;
shfl.sync.bfly.b32 %r384|%p53, %r380, %r365, %r364, %r366;
mov.b32 %f201, %r384;
add.ftz.f32 %f202, %f200, %f201;
mov.b32 %r385, %f202;
shfl.sync.bfly.b32 %r387|%p54, %r385, %r369, %r364, %r366;
mov.b32 %f203, %r387;
add.ftz.f32 %f204, %f202, %f203;
mov.b32 %r388, %f204;
shfl.sync.bfly.b32 %r390|%p55, %r388, %r372, %r364, %r366;
mov.b32 %f205, %r390;
add.ftz.f32 %f206, %f204, %f205;
mov.b32 %r391, %f206;
shfl.sync.bfly.b32 %r393|%p56, %r391, %r375, %r364, %r366;
mov.b32 %f207, %r393;
add.ftz.f32 %f208, %f206, %f207;
mov.b32 %r394, %f208;
shfl.sync.bfly.b32 %r396|%p57, %r394, %r378, %r364, %r366;
mov.b32 %f209, %r396;
add.ftz.f32 %f32, %f208, %f209;
@%p22 bra $L__BB51_22;

add.s32 %r397, %r19, 1;
cvt.u64.u32 %rd68, %r397;
add.s64 %rd69, %rd68, %rd11;
shl.b64 %rd71, %rd69, 2;
add.s64 %rd72, %rd47, %rd71;
st.global.f32 [%rd72], %f32;

$L__BB51_22:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type3ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type3ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type3ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type3ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type3ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type3ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type3ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type3ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type3ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<71>;
.reg .f32 %f<242>;
.reg .b32 %r<480>;
.reg .b64 %rd<85>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type3ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1280];

ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type3ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type3ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type3ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type3ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type3ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type3ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type3ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type3ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type3ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type3ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type3ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type3ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type3ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type3ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type3ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type3ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type3ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r39, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r39, %r2;
mov.u32 %r40, %ctaid.x;
shl.b32 %r4, %r40, 1;
shr.u32 %r41, %r3, 31;
add.s32 %r42, %r3, %r41;
shr.s32 %r479, %r42, 1;
shr.u32 %r6, %r38, 5;
setp.ge.s32 %p3, %r479, %r6;
mov.f32 %f232, 0f00000000;
mov.f32 %f233, %f232;
mov.f32 %f234, %f232;
mov.f32 %f235, %f232;
mov.f32 %f236, %f232;
mov.f32 %f237, %f232;
mov.f32 %f238, %f232;
mov.f32 %f239, %f232;
mov.f32 %f240, %f232;
mov.f32 %f241, %f232;
@%p3 bra $L__BB52_3;

mov.u32 %r43, %ctaid.y;
mov.u32 %r44, %ctaid.z;
mul.lo.s32 %r45, %r43, %r30;
cvt.u64.u32 %rd18, %r45;
mul.lo.s32 %r46, %r44, %r36;
cvt.u64.u32 %rd19, %r46;
add.s64 %rd1, %rd19, %rd18;
mul.hi.u32 %r47, %r44, %r32;
add.s32 %r48, %r47, %r44;
shr.u32 %r49, %r48, %r33;
mul.lo.s32 %r50, %r49, %r35;
mul.hi.u32 %r51, %r43, %r26;
add.s32 %r52, %r51, %r43;
shr.u32 %r53, %r52, %r27;
mul.lo.s32 %r54, %r53, %r29;
cvta.to.global.u64 %rd20, %rd15;
and.b32 %r57, %r42, 2147483646;
sub.s32 %r58, %r3, %r57;
shl.b32 %r59, %r58, 1;
cvt.s64.s32 %rd2, %r59;
shl.b32 %r60, %r24, 1;
add.s32 %r478, %r479, %r60;
mad.lo.s32 %r477, %r24, 3, %r479;
shl.b32 %r61, %r24, 2;
add.s32 %r476, %r479, %r61;
add.s32 %r475, %r479, %r24;
add.s32 %r62, %r479, %r50;
add.s32 %r63, %r62, %r54;
add.s32 %r64, %r4, 1;
mad.lo.s32 %r65, %r23, %r64, %r63;
mul.wide.s32 %rd21, %r65, 20;
add.s64 %rd22, %rd20, %rd21;
add.s64 %rd84, %rd22, 8;
mul.wide.s32 %rd4, %r59, 4;
mad.lo.s32 %r66, %r4, %r23, %r479;
add.s32 %r67, %r66, %r50;
add.s32 %r68, %r67, %r54;
mul.wide.s32 %rd23, %r68, 20;
add.s64 %rd24, %rd20, %rd23;
add.s64 %rd83, %rd24, 8;
cvta.to.global.u64 %rd6, %rd16;
mov.f32 %f232, 0f00000000;
shl.b64 %rd29, %rd2, 2;

$L__BB52_2:
cvt.u64.u32 %rd25, %r479;
add.s64 %rd26, %rd1, %rd25;
mul.lo.s64 %rd27, %rd26, 36;
add.s64 %rd28, %rd6, %rd27;
add.s64 %rd30, %rd28, %rd29;
ld.global.nc.u32 %r71, [%rd30+4];
ld.global.nc.u32 %r75, [%rd30+20];
add.s64 %rd31, %rd83, %rd4;
ld.global.nc.u32 %r279, [%rd31];
ld.global.nc.u32 %r79, [%rd30+8];
ld.global.nc.u32 %r83, [%rd30+24];
ld.global.nc.u32 %r280, [%rd31+-4];
and.b32 %r238, %r280, 252645135;
shr.u32 %r281, %r280, 4;
and.b32 %r242, %r281, 252645135;
mov.u32 %r261, 0;

dp4a.s32.s32 %r69, %r238, %r71, %r261;


dp4a.s32.s32 %r73, %r242, %r75, %r69;

and.b32 %r246, %r279, 252645135;
shr.u32 %r282, %r279, 4;
and.b32 %r250, %r282, 252645135;

dp4a.s32.s32 %r77, %r246, %r79, %r73;


dp4a.s32.s32 %r81, %r250, %r83, %r77;

ld.global.nc.u32 %r86, [%rd83+-8];
ld.global.nc.u32 %r87, [%rd28];

{mul.f16x2 %r85,%r86,%r87;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r85;
cvt.f32.f16 %f61, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r85;
cvt.f32.f16 %f62, high;}


cvt.rn.f32.s32 %f81, %r81;
mov.f32 %f82, 0f40000000;
div.approx.ftz.f32 %f83, %f62, %f82;
fma.rn.ftz.f32 %f84, %f61, %f81, %f83;
add.ftz.f32 %f241, %f84, %f241;
add.s64 %rd32, %rd84, %rd4;
ld.global.nc.u32 %r283, [%rd32];
ld.global.nc.u32 %r284, [%rd32+-4];
and.b32 %r259, %r284, 252645135;
shr.u32 %r285, %r284, 4;
and.b32 %r263, %r285, 252645135;

dp4a.s32.s32 %r90, %r259, %r71, %r261;


dp4a.s32.s32 %r94, %r263, %r75, %r90;

and.b32 %r267, %r283, 252645135;
shr.u32 %r286, %r283, 4;
and.b32 %r271, %r286, 252645135;

dp4a.s32.s32 %r98, %r267, %r79, %r94;


dp4a.s32.s32 %r102, %r271, %r83, %r98;

ld.global.nc.u32 %r107, [%rd84+-8];

{mul.f16x2 %r106,%r107,%r87;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r106;
cvt.f32.f16 %f63, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r106;
cvt.f32.f16 %f64, high;}


cvt.rn.f32.s32 %f85, %r102;
div.approx.ftz.f32 %f86, %f64, %f82;
fma.rn.ftz.f32 %f87, %f63, %f85, %f86;
add.ftz.f32 %f240, %f87, %f240;
cvt.u64.u32 %rd33, %r475;
add.s64 %rd34, %rd1, %rd33;
mul.lo.s64 %rd35, %rd34, 36;
add.s64 %rd36, %rd6, %rd35;
add.s64 %rd37, %rd36, %rd29;
ld.global.nc.u32 %r113, [%rd37+4];
ld.global.nc.u32 %r117, [%rd37+20];
ld.global.nc.u32 %r121, [%rd37+8];
ld.global.nc.u32 %r125, [%rd37+24];

dp4a.s32.s32 %r111, %r238, %r113, %r261;


dp4a.s32.s32 %r115, %r242, %r117, %r111;


dp4a.s32.s32 %r119, %r246, %r121, %r115;


dp4a.s32.s32 %r123, %r250, %r125, %r119;

ld.global.nc.u32 %r129, [%rd36];

{mul.f16x2 %r127,%r86,%r129;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r127;
cvt.f32.f16 %f65, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r127;
cvt.f32.f16 %f66, high;}


cvt.rn.f32.s32 %f88, %r123;
div.approx.ftz.f32 %f89, %f66, %f82;
fma.rn.ftz.f32 %f90, %f65, %f88, %f89;
add.ftz.f32 %f239, %f90, %f239;

dp4a.s32.s32 %r132, %r259, %r113, %r261;


dp4a.s32.s32 %r136, %r263, %r117, %r132;


dp4a.s32.s32 %r140, %r267, %r121, %r136;


dp4a.s32.s32 %r144, %r271, %r125, %r140;


{mul.f16x2 %r148,%r107,%r129;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r148;
cvt.f32.f16 %f67, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r148;
cvt.f32.f16 %f68, high;}


cvt.rn.f32.s32 %f91, %r144;
div.approx.ftz.f32 %f92, %f68, %f82;
fma.rn.ftz.f32 %f93, %f67, %f91, %f92;
add.ftz.f32 %f238, %f93, %f238;
cvt.u64.u32 %rd38, %r478;
add.s64 %rd39, %rd1, %rd38;
mul.lo.s64 %rd40, %rd39, 36;
add.s64 %rd41, %rd6, %rd40;
add.s64 %rd42, %rd41, %rd29;
ld.global.nc.u32 %r155, [%rd42+4];
ld.global.nc.u32 %r159, [%rd42+20];
ld.global.nc.u32 %r163, [%rd42+8];
ld.global.nc.u32 %r167, [%rd42+24];

dp4a.s32.s32 %r153, %r238, %r155, %r261;


dp4a.s32.s32 %r157, %r242, %r159, %r153;


dp4a.s32.s32 %r161, %r246, %r163, %r157;


dp4a.s32.s32 %r165, %r250, %r167, %r161;

ld.global.nc.u32 %r171, [%rd41];

{mul.f16x2 %r169,%r86,%r171;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r169;
cvt.f32.f16 %f69, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r169;
cvt.f32.f16 %f70, high;}


cvt.rn.f32.s32 %f94, %r165;
div.approx.ftz.f32 %f95, %f70, %f82;
fma.rn.ftz.f32 %f96, %f69, %f94, %f95;
add.ftz.f32 %f237, %f96, %f237;

dp4a.s32.s32 %r174, %r259, %r155, %r261;


dp4a.s32.s32 %r178, %r263, %r159, %r174;


dp4a.s32.s32 %r182, %r267, %r163, %r178;


dp4a.s32.s32 %r186, %r271, %r167, %r182;


{mul.f16x2 %r190,%r107,%r171;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r190;
cvt.f32.f16 %f71, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r190;
cvt.f32.f16 %f72, high;}


cvt.rn.f32.s32 %f97, %r186;
div.approx.ftz.f32 %f98, %f72, %f82;
fma.rn.ftz.f32 %f99, %f71, %f97, %f98;
add.ftz.f32 %f236, %f99, %f236;
cvt.u64.u32 %rd43, %r477;
add.s64 %rd44, %rd1, %rd43;
mul.lo.s64 %rd45, %rd44, 36;
add.s64 %rd46, %rd6, %rd45;
add.s64 %rd47, %rd46, %rd29;
ld.global.nc.u32 %r197, [%rd47+4];
ld.global.nc.u32 %r201, [%rd47+20];
ld.global.nc.u32 %r205, [%rd47+8];
ld.global.nc.u32 %r209, [%rd47+24];

dp4a.s32.s32 %r195, %r238, %r197, %r261;


dp4a.s32.s32 %r199, %r242, %r201, %r195;


dp4a.s32.s32 %r203, %r246, %r205, %r199;


dp4a.s32.s32 %r207, %r250, %r209, %r203;

ld.global.nc.u32 %r213, [%rd46];

{mul.f16x2 %r211,%r86,%r213;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r211;
cvt.f32.f16 %f73, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r211;
cvt.f32.f16 %f74, high;}


cvt.rn.f32.s32 %f100, %r207;
div.approx.ftz.f32 %f101, %f74, %f82;
fma.rn.ftz.f32 %f102, %f73, %f100, %f101;
add.ftz.f32 %f235, %f102, %f235;

dp4a.s32.s32 %r216, %r259, %r197, %r261;


dp4a.s32.s32 %r220, %r263, %r201, %r216;


dp4a.s32.s32 %r224, %r267, %r205, %r220;


dp4a.s32.s32 %r228, %r271, %r209, %r224;


{mul.f16x2 %r232,%r107,%r213;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r232;
cvt.f32.f16 %f75, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r232;
cvt.f32.f16 %f76, high;}


cvt.rn.f32.s32 %f103, %r228;
div.approx.ftz.f32 %f104, %f76, %f82;
fma.rn.ftz.f32 %f105, %f75, %f103, %f104;
add.ftz.f32 %f234, %f105, %f234;
cvt.u64.u32 %rd48, %r476;
add.s64 %rd49, %rd1, %rd48;
mul.lo.s64 %rd50, %rd49, 36;
add.s64 %rd51, %rd6, %rd50;
add.s64 %rd52, %rd51, %rd29;
ld.global.nc.u32 %r239, [%rd52+4];
ld.global.nc.u32 %r243, [%rd52+20];
ld.global.nc.u32 %r247, [%rd52+8];
ld.global.nc.u32 %r251, [%rd52+24];

dp4a.s32.s32 %r237, %r238, %r239, %r261;


dp4a.s32.s32 %r241, %r242, %r243, %r237;


dp4a.s32.s32 %r245, %r246, %r247, %r241;


dp4a.s32.s32 %r249, %r250, %r251, %r245;

ld.global.nc.u32 %r255, [%rd51];

{mul.f16x2 %r253,%r86,%r255;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r253;
cvt.f32.f16 %f77, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r253;
cvt.f32.f16 %f78, high;}


cvt.rn.f32.s32 %f106, %r249;
div.approx.ftz.f32 %f107, %f78, %f82;
fma.rn.ftz.f32 %f108, %f77, %f106, %f107;
add.ftz.f32 %f233, %f108, %f233;

dp4a.s32.s32 %r258, %r259, %r239, %r261;


dp4a.s32.s32 %r262, %r263, %r243, %r258;


dp4a.s32.s32 %r266, %r267, %r247, %r262;


dp4a.s32.s32 %r270, %r271, %r251, %r266;


{mul.f16x2 %r274,%r107,%r255;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r274;
cvt.f32.f16 %f79, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r274;
cvt.f32.f16 %f80, high;}


cvt.rn.f32.s32 %f109, %r270;
div.approx.ftz.f32 %f110, %f80, %f82;
fma.rn.ftz.f32 %f111, %f79, %f109, %f110;
add.ftz.f32 %f232, %f111, %f232;
add.s32 %r478, %r478, 32;
add.s32 %r477, %r477, 32;
add.s32 %r476, %r476, 32;
add.s32 %r475, %r475, 32;
add.s64 %rd84, %rd84, 640;
add.s64 %rd83, %rd83, 640;
add.s32 %r479, %r479, 32;
setp.lt.s32 %p4, %r479, %r6;
@%p4 bra $L__BB52_2;

$L__BB52_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB52_5;

mov.u32 %r287, _ZZ13mul_mat_vec_qIL9ggml_type3ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r288, %r1, 1280, %r287;
shl.b32 %r289, %r2, 2;
add.s32 %r290, %r288, %r289;
st.shared.f32 [%r290+-1280], %f241;
st.shared.f32 [%r290+-1152], %f240;
st.shared.f32 [%r290+-1024], %f239;
st.shared.f32 [%r290+-896], %f238;
st.shared.f32 [%r290+-768], %f237;
st.shared.f32 [%r290+-640], %f236;
st.shared.f32 [%r290+-512], %f235;
st.shared.f32 [%r290+-384], %f234;
st.shared.f32 [%r290+-256], %f233;
st.shared.f32 [%r290+-128], %f232;

$L__BB52_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB52_26;

mov.u32 %r291, %ctaid.z;
mov.u32 %r292, %ctaid.y;
mad.lo.s32 %r293, %r292, %r31, %r4;
mad.lo.s32 %r294, %r291, %r37, %r293;
cvt.u64.u32 %rd11, %r294;
shl.b32 %r295, %r2, 2;
mov.u32 %r296, _ZZ13mul_mat_vec_qIL9ggml_type3ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r21, %r296, %r295;
mov.u32 %r297, 2;
ld.shared.f32 %f112, [%r21];
add.ftz.f32 %f113, %f112, %f241;
mov.b32 %r298, %f113;
mov.u32 %r299, 31;
mov.u32 %r300, 16;
mov.u32 %r301, -1;
shfl.sync.bfly.b32 %r302|%p7, %r298, %r300, %r299, %r301;
mov.b32 %f114, %r302;
add.ftz.f32 %f115, %f113, %f114;
mov.b32 %r303, %f115;
mov.u32 %r304, 8;
shfl.sync.bfly.b32 %r305|%p8, %r303, %r304, %r299, %r301;
mov.b32 %f116, %r305;
add.ftz.f32 %f117, %f115, %f116;
mov.b32 %r306, %f117;
mov.u32 %r307, 4;
shfl.sync.bfly.b32 %r308|%p9, %r306, %r307, %r299, %r301;
mov.b32 %f118, %r308;
add.ftz.f32 %f119, %f117, %f118;
mov.b32 %r309, %f119;
shfl.sync.bfly.b32 %r310|%p10, %r309, %r297, %r299, %r301;
mov.b32 %f120, %r310;
add.ftz.f32 %f121, %f119, %f120;
mov.b32 %r311, %f121;
mov.u32 %r312, 1;
shfl.sync.bfly.b32 %r313|%p11, %r311, %r312, %r299, %r301;
mov.b32 %f122, %r313;
add.ftz.f32 %f31, %f121, %f122;
setp.lt.u32 %p12, %r4, %r25;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
cvta.to.global.u64 %rd53, %rd17;
mul.wide.u32 %rd54, %r294, 4;
add.s64 %rd12, %rd53, %rd54;
not.pred %p14, %p1;
@%p14 bra $L__BB52_8;

st.global.f32 [%rd12], %f31;

$L__BB52_8:
ld.shared.f32 %f123, [%r21+128];
add.ftz.f32 %f124, %f123, %f240;
mov.b32 %r314, %f124;
shfl.sync.bfly.b32 %r318|%p15, %r314, %r300, %r299, %r301;
mov.b32 %f125, %r318;
add.ftz.f32 %f126, %f124, %f125;
mov.b32 %r319, %f126;
shfl.sync.bfly.b32 %r321|%p16, %r319, %r304, %r299, %r301;
mov.b32 %f127, %r321;
add.ftz.f32 %f128, %f126, %f127;
mov.b32 %r322, %f128;
shfl.sync.bfly.b32 %r324|%p17, %r322, %r307, %r299, %r301;
mov.b32 %f129, %r324;
add.ftz.f32 %f130, %f128, %f129;
mov.b32 %r325, %f130;
shfl.sync.bfly.b32 %r327|%p18, %r325, %r297, %r299, %r301;
mov.b32 %f131, %r327;
add.ftz.f32 %f132, %f130, %f131;
mov.b32 %r328, %f132;
shfl.sync.bfly.b32 %r330|%p19, %r328, %r312, %r299, %r301;
mov.b32 %f133, %r330;
add.ftz.f32 %f32, %f132, %f133;
setp.eq.s32 %p20, %r2, 1;
or.b32 %r334, %r4, 1;
setp.lt.u32 %p21, %r334, %r25;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB52_10;

st.global.f32 [%rd12+4], %f32;

$L__BB52_10:
ld.shared.f32 %f134, [%r21+256];
add.ftz.f32 %f135, %f134, %f239;
mov.b32 %r335, %f135;
mov.u32 %r336, 31;
mov.u32 %r337, 16;
mov.u32 %r338, -1;
shfl.sync.bfly.b32 %r339|%p23, %r335, %r337, %r336, %r338;
mov.b32 %f136, %r339;
add.ftz.f32 %f137, %f135, %f136;
mov.b32 %r340, %f137;
mov.u32 %r341, 8;
shfl.sync.bfly.b32 %r342|%p24, %r340, %r341, %r336, %r338;
mov.b32 %f138, %r342;
add.ftz.f32 %f139, %f137, %f138;
mov.b32 %r343, %f139;
mov.u32 %r344, 4;
shfl.sync.bfly.b32 %r345|%p25, %r343, %r344, %r336, %r338;
mov.b32 %f140, %r345;
add.ftz.f32 %f141, %f139, %f140;
mov.b32 %r346, %f141;
mov.u32 %r347, 2;
shfl.sync.bfly.b32 %r348|%p26, %r346, %r347, %r336, %r338;
mov.b32 %f142, %r348;
add.ftz.f32 %f143, %f141, %f142;
mov.b32 %r349, %f143;
mov.u32 %r350, 1;
shfl.sync.bfly.b32 %r351|%p27, %r349, %r350, %r336, %r338;
mov.b32 %f144, %r351;
add.ftz.f32 %f33, %f143, %f144;
@%p14 bra $L__BB52_12;

cvt.u64.u32 %rd55, %r25;
add.s64 %rd56, %rd55, %rd11;
shl.b64 %rd58, %rd56, 2;
add.s64 %rd59, %rd53, %rd58;
st.global.f32 [%rd59], %f33;

$L__BB52_12:
ld.shared.f32 %f145, [%r21+384];
add.ftz.f32 %f146, %f145, %f238;
mov.b32 %r352, %f146;
shfl.sync.bfly.b32 %r356|%p29, %r352, %r337, %r336, %r338;
mov.b32 %f147, %r356;
add.ftz.f32 %f148, %f146, %f147;
mov.b32 %r357, %f148;
shfl.sync.bfly.b32 %r359|%p30, %r357, %r341, %r336, %r338;
mov.b32 %f149, %r359;
add.ftz.f32 %f150, %f148, %f149;
mov.b32 %r360, %f150;
shfl.sync.bfly.b32 %r362|%p31, %r360, %r344, %r336, %r338;
mov.b32 %f151, %r362;
add.ftz.f32 %f152, %f150, %f151;
mov.b32 %r363, %f152;
shfl.sync.bfly.b32 %r365|%p32, %r363, %r347, %r336, %r338;
mov.b32 %f153, %r365;
add.ftz.f32 %f154, %f152, %f153;
mov.b32 %r366, %f154;
shfl.sync.bfly.b32 %r368|%p33, %r366, %r350, %r336, %r338;
mov.b32 %f155, %r368;
add.ftz.f32 %f34, %f154, %f155;
@%p22 bra $L__BB52_14;

add.s32 %r369, %r25, 1;
cvt.u64.u32 %rd60, %r369;
add.s64 %rd61, %rd60, %rd11;
shl.b64 %rd63, %rd61, 2;
add.s64 %rd64, %rd53, %rd63;
st.global.f32 [%rd64], %f34;

$L__BB52_14:
shl.b32 %r370, %r25, 1;
mov.u32 %r371, 1;
ld.shared.f32 %f156, [%r21+512];
add.ftz.f32 %f157, %f156, %f237;
mov.b32 %r372, %f157;
mov.u32 %r373, 31;
mov.u32 %r374, 16;
mov.u32 %r375, -1;
shfl.sync.bfly.b32 %r376|%p35, %r372, %r374, %r373, %r375;
mov.b32 %f158, %r376;
add.ftz.f32 %f159, %f157, %f158;
mov.b32 %r377, %f159;
mov.u32 %r378, 8;
shfl.sync.bfly.b32 %r379|%p36, %r377, %r378, %r373, %r375;
mov.b32 %f160, %r379;
add.ftz.f32 %f161, %f159, %f160;
mov.b32 %r380, %f161;
mov.u32 %r381, 4;
shfl.sync.bfly.b32 %r382|%p37, %r380, %r381, %r373, %r375;
mov.b32 %f162, %r382;
add.ftz.f32 %f163, %f161, %f162;
mov.b32 %r383, %f163;
mov.u32 %r384, 2;
shfl.sync.bfly.b32 %r385|%p38, %r383, %r384, %r373, %r375;
mov.b32 %f164, %r385;
add.ftz.f32 %f165, %f163, %f164;
mov.b32 %r386, %f165;
shfl.sync.bfly.b32 %r387|%p39, %r386, %r371, %r373, %r375;
mov.b32 %f166, %r387;
add.ftz.f32 %f35, %f165, %f166;
cvt.u64.u32 %rd65, %r370;
add.s64 %rd66, %rd65, %rd11;
shl.b64 %rd68, %rd66, 2;
add.s64 %rd13, %rd53, %rd68;
@%p14 bra $L__BB52_16;

st.global.f32 [%rd13], %f35;

$L__BB52_16:
ld.shared.f32 %f167, [%r21+640];
add.ftz.f32 %f168, %f167, %f236;
mov.b32 %r388, %f168;
shfl.sync.bfly.b32 %r392|%p41, %r388, %r374, %r373, %r375;
mov.b32 %f169, %r392;
add.ftz.f32 %f170, %f168, %f169;
mov.b32 %r393, %f170;
shfl.sync.bfly.b32 %r395|%p42, %r393, %r378, %r373, %r375;
mov.b32 %f171, %r395;
add.ftz.f32 %f172, %f170, %f171;
mov.b32 %r396, %f172;
shfl.sync.bfly.b32 %r398|%p43, %r396, %r381, %r373, %r375;
mov.b32 %f173, %r398;
add.ftz.f32 %f174, %f172, %f173;
mov.b32 %r399, %f174;
shfl.sync.bfly.b32 %r401|%p44, %r399, %r384, %r373, %r375;
mov.b32 %f175, %r401;
add.ftz.f32 %f176, %f174, %f175;
mov.b32 %r402, %f176;
shfl.sync.bfly.b32 %r404|%p45, %r402, %r371, %r373, %r375;
mov.b32 %f177, %r404;
add.ftz.f32 %f36, %f176, %f177;
@%p22 bra $L__BB52_18;

st.global.f32 [%rd13+4], %f36;

$L__BB52_18:
mul.lo.s32 %r22, %r25, 3;
ld.shared.f32 %f178, [%r21+768];
add.ftz.f32 %f179, %f178, %f235;
mov.b32 %r405, %f179;
mov.u32 %r406, 31;
mov.u32 %r407, 16;
mov.u32 %r408, -1;
shfl.sync.bfly.b32 %r409|%p47, %r405, %r407, %r406, %r408;
mov.b32 %f180, %r409;
add.ftz.f32 %f181, %f179, %f180;
mov.b32 %r410, %f181;
mov.u32 %r411, 8;
shfl.sync.bfly.b32 %r412|%p48, %r410, %r411, %r406, %r408;
mov.b32 %f182, %r412;
add.ftz.f32 %f183, %f181, %f182;
mov.b32 %r413, %f183;
mov.u32 %r414, 4;
shfl.sync.bfly.b32 %r415|%p49, %r413, %r414, %r406, %r408;
mov.b32 %f184, %r415;
add.ftz.f32 %f185, %f183, %f184;
mov.b32 %r416, %f185;
mov.u32 %r417, 2;
shfl.sync.bfly.b32 %r418|%p50, %r416, %r417, %r406, %r408;
mov.b32 %f186, %r418;
add.ftz.f32 %f187, %f185, %f186;
mov.b32 %r419, %f187;
mov.u32 %r420, 1;
shfl.sync.bfly.b32 %r421|%p51, %r419, %r420, %r406, %r408;
mov.b32 %f188, %r421;
add.ftz.f32 %f37, %f187, %f188;
@%p14 bra $L__BB52_20;

cvt.u64.u32 %rd69, %r22;
add.s64 %rd70, %rd69, %rd11;
shl.b64 %rd72, %rd70, 2;
add.s64 %rd73, %rd53, %rd72;
st.global.f32 [%rd73], %f37;

$L__BB52_20:
ld.shared.f32 %f189, [%r21+896];
add.ftz.f32 %f190, %f189, %f234;
mov.b32 %r422, %f190;
shfl.sync.bfly.b32 %r426|%p53, %r422, %r407, %r406, %r408;
mov.b32 %f191, %r426;
add.ftz.f32 %f192, %f190, %f191;
mov.b32 %r427, %f192;
shfl.sync.bfly.b32 %r429|%p54, %r427, %r411, %r406, %r408;
mov.b32 %f193, %r429;
add.ftz.f32 %f194, %f192, %f193;
mov.b32 %r430, %f194;
shfl.sync.bfly.b32 %r432|%p55, %r430, %r414, %r406, %r408;
mov.b32 %f195, %r432;
add.ftz.f32 %f196, %f194, %f195;
mov.b32 %r433, %f196;
shfl.sync.bfly.b32 %r435|%p56, %r433, %r417, %r406, %r408;
mov.b32 %f197, %r435;
add.ftz.f32 %f198, %f196, %f197;
mov.b32 %r436, %f198;
shfl.sync.bfly.b32 %r438|%p57, %r436, %r420, %r406, %r408;
mov.b32 %f199, %r438;
add.ftz.f32 %f38, %f198, %f199;
@%p22 bra $L__BB52_22;

add.s32 %r439, %r22, 1;
cvt.u64.u32 %rd74, %r439;
add.s64 %rd75, %rd74, %rd11;
shl.b64 %rd77, %rd75, 2;
add.s64 %rd78, %rd53, %rd77;
st.global.f32 [%rd78], %f38;

$L__BB52_22:
shl.b32 %r440, %r25, 2;
mov.u32 %r441, 2;
ld.shared.f32 %f200, [%r21+1024];
add.ftz.f32 %f201, %f200, %f233;
mov.b32 %r442, %f201;
mov.u32 %r443, 31;
mov.u32 %r444, 16;
mov.u32 %r445, -1;
shfl.sync.bfly.b32 %r446|%p59, %r442, %r444, %r443, %r445;
mov.b32 %f202, %r446;
add.ftz.f32 %f203, %f201, %f202;
mov.b32 %r447, %f203;
mov.u32 %r448, 8;
shfl.sync.bfly.b32 %r449|%p60, %r447, %r448, %r443, %r445;
mov.b32 %f204, %r449;
add.ftz.f32 %f205, %f203, %f204;
mov.b32 %r450, %f205;
mov.u32 %r451, 4;
shfl.sync.bfly.b32 %r452|%p61, %r450, %r451, %r443, %r445;
mov.b32 %f206, %r452;
add.ftz.f32 %f207, %f205, %f206;
mov.b32 %r453, %f207;
shfl.sync.bfly.b32 %r454|%p62, %r453, %r441, %r443, %r445;
mov.b32 %f208, %r454;
add.ftz.f32 %f209, %f207, %f208;
mov.b32 %r455, %f209;
mov.u32 %r456, 1;
shfl.sync.bfly.b32 %r457|%p63, %r455, %r456, %r443, %r445;
mov.b32 %f210, %r457;
add.ftz.f32 %f39, %f209, %f210;
cvt.u64.u32 %rd79, %r440;
add.s64 %rd80, %rd79, %rd11;
shl.b64 %rd82, %rd80, 2;
add.s64 %rd14, %rd53, %rd82;
@%p14 bra $L__BB52_24;

st.global.f32 [%rd14], %f39;

$L__BB52_24:
ld.shared.f32 %f211, [%r21+1152];
add.ftz.f32 %f212, %f211, %f232;
mov.b32 %r458, %f212;
shfl.sync.bfly.b32 %r462|%p65, %r458, %r444, %r443, %r445;
mov.b32 %f213, %r462;
add.ftz.f32 %f214, %f212, %f213;
mov.b32 %r463, %f214;
shfl.sync.bfly.b32 %r465|%p66, %r463, %r448, %r443, %r445;
mov.b32 %f215, %r465;
add.ftz.f32 %f216, %f214, %f215;
mov.b32 %r466, %f216;
shfl.sync.bfly.b32 %r468|%p67, %r466, %r451, %r443, %r445;
mov.b32 %f217, %r468;
add.ftz.f32 %f218, %f216, %f217;
mov.b32 %r469, %f218;
shfl.sync.bfly.b32 %r471|%p68, %r469, %r441, %r443, %r445;
mov.b32 %f219, %r471;
add.ftz.f32 %f220, %f218, %f219;
mov.b32 %r472, %f220;
shfl.sync.bfly.b32 %r474|%p69, %r472, %r456, %r443, %r445;
mov.b32 %f221, %r474;
add.ftz.f32 %f40, %f220, %f221;
@%p22 bra $L__BB52_26;

st.global.f32 [%rd14+4], %f40;

$L__BB52_26:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type3ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type3ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type3ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type3ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type3ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type3ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type3ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type3ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type3ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<83>;
.reg .f32 %f<290>;
.reg .b32 %r<574>;
.reg .b64 %rd<103>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type3ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

ld.param.u64 %rd14, [_Z13mul_mat_vec_qIL9ggml_type3ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type3ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type3ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type3ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type3ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type3ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type3ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type3ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type3ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type3ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type3ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type3ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type3ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type3ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type3ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type3ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type3ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r43, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r43, %r2;
mov.u32 %r44, %ctaid.x;
shl.b32 %r4, %r44, 1;
shr.u32 %r45, %r3, 31;
add.s32 %r46, %r3, %r45;
shr.s32 %r573, %r46, 1;
shr.u32 %r6, %r42, 5;
setp.ge.s32 %p3, %r573, %r6;
mov.f32 %f278, 0f00000000;
mov.f32 %f279, %f278;
mov.f32 %f280, %f278;
mov.f32 %f281, %f278;
mov.f32 %f282, %f278;
mov.f32 %f283, %f278;
mov.f32 %f284, %f278;
mov.f32 %f285, %f278;
mov.f32 %f286, %f278;
mov.f32 %f287, %f278;
mov.f32 %f288, %f278;
mov.f32 %f289, %f278;
@%p3 bra $L__BB53_3;

mov.u32 %r47, %ctaid.y;
mov.u32 %r48, %ctaid.z;
mul.lo.s32 %r49, %r47, %r34;
cvt.u64.u32 %rd17, %r49;
mul.lo.s32 %r50, %r48, %r40;
cvt.u64.u32 %rd18, %r50;
add.s64 %rd1, %rd18, %rd17;
mul.hi.u32 %r51, %r48, %r36;
add.s32 %r52, %r51, %r48;
shr.u32 %r53, %r52, %r37;
mul.lo.s32 %r54, %r53, %r39;
mul.hi.u32 %r55, %r47, %r30;
add.s32 %r56, %r55, %r47;
shr.u32 %r57, %r56, %r31;
mul.lo.s32 %r58, %r57, %r33;
cvta.to.global.u64 %rd19, %rd14;
and.b32 %r61, %r46, 2147483646;
sub.s32 %r62, %r3, %r61;
shl.b32 %r63, %r62, 1;
cvt.s64.s32 %rd2, %r63;
shl.b32 %r64, %r28, 1;
add.s32 %r572, %r573, %r64;
mad.lo.s32 %r571, %r28, 3, %r573;
shl.b32 %r65, %r28, 2;
add.s32 %r570, %r573, %r65;
mad.lo.s32 %r569, %r28, 5, %r573;
add.s32 %r568, %r573, %r28;
add.s32 %r66, %r573, %r54;
add.s32 %r67, %r66, %r58;
add.s32 %r68, %r4, 1;
mad.lo.s32 %r69, %r27, %r68, %r67;
mul.wide.s32 %rd20, %r69, 20;
add.s64 %rd21, %rd19, %rd20;
add.s64 %rd102, %rd21, 8;
mul.wide.s32 %rd4, %r63, 4;
mad.lo.s32 %r70, %r4, %r27, %r573;
add.s32 %r71, %r70, %r54;
add.s32 %r72, %r71, %r58;
mul.wide.s32 %rd22, %r72, 20;
add.s64 %rd23, %rd19, %rd22;
add.s64 %rd101, %rd23, 8;
cvta.to.global.u64 %rd6, %rd15;
mov.f32 %f278, 0f00000000;
shl.b64 %rd28, %rd2, 2;

$L__BB53_2:
cvt.u64.u32 %rd24, %r573;
add.s64 %rd25, %rd1, %rd24;
mul.lo.s64 %rd26, %rd25, 36;
add.s64 %rd27, %rd6, %rd26;
add.s64 %rd29, %rd27, %rd28;
ld.global.nc.u32 %r75, [%rd29+4];
ld.global.nc.u32 %r79, [%rd29+20];
add.s64 %rd30, %rd101, %rd4;
ld.global.nc.u32 %r325, [%rd30];
ld.global.nc.u32 %r83, [%rd29+8];
ld.global.nc.u32 %r87, [%rd29+24];
ld.global.nc.u32 %r326, [%rd30+-4];
and.b32 %r284, %r326, 252645135;
shr.u32 %r327, %r326, 4;
and.b32 %r288, %r327, 252645135;
mov.u32 %r307, 0;

dp4a.s32.s32 %r73, %r284, %r75, %r307;


dp4a.s32.s32 %r77, %r288, %r79, %r73;

and.b32 %r292, %r325, 252645135;
shr.u32 %r328, %r325, 4;
and.b32 %r296, %r328, 252645135;

dp4a.s32.s32 %r81, %r292, %r83, %r77;


dp4a.s32.s32 %r85, %r296, %r87, %r81;

ld.global.nc.u32 %r90, [%rd101+-8];
ld.global.nc.u32 %r91, [%rd27];

{mul.f16x2 %r89,%r90,%r91;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r89;
cvt.f32.f16 %f73, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r89;
cvt.f32.f16 %f74, high;}


cvt.rn.f32.s32 %f97, %r85;
mov.f32 %f98, 0f40000000;
div.approx.ftz.f32 %f99, %f74, %f98;
fma.rn.ftz.f32 %f100, %f73, %f97, %f99;
add.ftz.f32 %f289, %f100, %f289;
add.s64 %rd31, %rd102, %rd4;
ld.global.nc.u32 %r329, [%rd31];
ld.global.nc.u32 %r330, [%rd31+-4];
and.b32 %r305, %r330, 252645135;
shr.u32 %r331, %r330, 4;
and.b32 %r309, %r331, 252645135;

dp4a.s32.s32 %r94, %r305, %r75, %r307;


dp4a.s32.s32 %r98, %r309, %r79, %r94;

and.b32 %r313, %r329, 252645135;
shr.u32 %r332, %r329, 4;
and.b32 %r317, %r332, 252645135;

dp4a.s32.s32 %r102, %r313, %r83, %r98;


dp4a.s32.s32 %r106, %r317, %r87, %r102;

ld.global.nc.u32 %r111, [%rd102+-8];

{mul.f16x2 %r110,%r111,%r91;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r110;
cvt.f32.f16 %f75, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r110;
cvt.f32.f16 %f76, high;}


cvt.rn.f32.s32 %f101, %r106;
div.approx.ftz.f32 %f102, %f76, %f98;
fma.rn.ftz.f32 %f103, %f75, %f101, %f102;
add.ftz.f32 %f288, %f103, %f288;
cvt.u64.u32 %rd32, %r568;
add.s64 %rd33, %rd1, %rd32;
mul.lo.s64 %rd34, %rd33, 36;
add.s64 %rd35, %rd6, %rd34;
add.s64 %rd36, %rd35, %rd28;
ld.global.nc.u32 %r117, [%rd36+4];
ld.global.nc.u32 %r121, [%rd36+20];
ld.global.nc.u32 %r125, [%rd36+8];
ld.global.nc.u32 %r129, [%rd36+24];

dp4a.s32.s32 %r115, %r284, %r117, %r307;


dp4a.s32.s32 %r119, %r288, %r121, %r115;


dp4a.s32.s32 %r123, %r292, %r125, %r119;


dp4a.s32.s32 %r127, %r296, %r129, %r123;

ld.global.nc.u32 %r133, [%rd35];

{mul.f16x2 %r131,%r90,%r133;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r131;
cvt.f32.f16 %f77, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r131;
cvt.f32.f16 %f78, high;}


cvt.rn.f32.s32 %f104, %r127;
div.approx.ftz.f32 %f105, %f78, %f98;
fma.rn.ftz.f32 %f106, %f77, %f104, %f105;
add.ftz.f32 %f287, %f106, %f287;

dp4a.s32.s32 %r136, %r305, %r117, %r307;


dp4a.s32.s32 %r140, %r309, %r121, %r136;


dp4a.s32.s32 %r144, %r313, %r125, %r140;


dp4a.s32.s32 %r148, %r317, %r129, %r144;


{mul.f16x2 %r152,%r111,%r133;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r152;
cvt.f32.f16 %f79, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r152;
cvt.f32.f16 %f80, high;}


cvt.rn.f32.s32 %f107, %r148;
div.approx.ftz.f32 %f108, %f80, %f98;
fma.rn.ftz.f32 %f109, %f79, %f107, %f108;
add.ftz.f32 %f286, %f109, %f286;
cvt.u64.u32 %rd37, %r572;
add.s64 %rd38, %rd1, %rd37;
mul.lo.s64 %rd39, %rd38, 36;
add.s64 %rd40, %rd6, %rd39;
add.s64 %rd41, %rd40, %rd28;
ld.global.nc.u32 %r159, [%rd41+4];
ld.global.nc.u32 %r163, [%rd41+20];
ld.global.nc.u32 %r167, [%rd41+8];
ld.global.nc.u32 %r171, [%rd41+24];

dp4a.s32.s32 %r157, %r284, %r159, %r307;


dp4a.s32.s32 %r161, %r288, %r163, %r157;


dp4a.s32.s32 %r165, %r292, %r167, %r161;


dp4a.s32.s32 %r169, %r296, %r171, %r165;

ld.global.nc.u32 %r175, [%rd40];

{mul.f16x2 %r173,%r90,%r175;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r173;
cvt.f32.f16 %f81, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r173;
cvt.f32.f16 %f82, high;}


cvt.rn.f32.s32 %f110, %r169;
div.approx.ftz.f32 %f111, %f82, %f98;
fma.rn.ftz.f32 %f112, %f81, %f110, %f111;
add.ftz.f32 %f285, %f112, %f285;

dp4a.s32.s32 %r178, %r305, %r159, %r307;


dp4a.s32.s32 %r182, %r309, %r163, %r178;


dp4a.s32.s32 %r186, %r313, %r167, %r182;


dp4a.s32.s32 %r190, %r317, %r171, %r186;


{mul.f16x2 %r194,%r111,%r175;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r194;
cvt.f32.f16 %f83, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r194;
cvt.f32.f16 %f84, high;}


cvt.rn.f32.s32 %f113, %r190;
div.approx.ftz.f32 %f114, %f84, %f98;
fma.rn.ftz.f32 %f115, %f83, %f113, %f114;
add.ftz.f32 %f284, %f115, %f284;
cvt.u64.u32 %rd42, %r571;
add.s64 %rd43, %rd1, %rd42;
mul.lo.s64 %rd44, %rd43, 36;
add.s64 %rd45, %rd6, %rd44;
add.s64 %rd46, %rd45, %rd28;
ld.global.nc.u32 %r201, [%rd46+4];
ld.global.nc.u32 %r205, [%rd46+20];
ld.global.nc.u32 %r209, [%rd46+8];
ld.global.nc.u32 %r213, [%rd46+24];

dp4a.s32.s32 %r199, %r284, %r201, %r307;


dp4a.s32.s32 %r203, %r288, %r205, %r199;


dp4a.s32.s32 %r207, %r292, %r209, %r203;


dp4a.s32.s32 %r211, %r296, %r213, %r207;

ld.global.nc.u32 %r217, [%rd45];

{mul.f16x2 %r215,%r90,%r217;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r215;
cvt.f32.f16 %f85, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r215;
cvt.f32.f16 %f86, high;}


cvt.rn.f32.s32 %f116, %r211;
div.approx.ftz.f32 %f117, %f86, %f98;
fma.rn.ftz.f32 %f118, %f85, %f116, %f117;
add.ftz.f32 %f283, %f118, %f283;

dp4a.s32.s32 %r220, %r305, %r201, %r307;


dp4a.s32.s32 %r224, %r309, %r205, %r220;


dp4a.s32.s32 %r228, %r313, %r209, %r224;


dp4a.s32.s32 %r232, %r317, %r213, %r228;


{mul.f16x2 %r236,%r111,%r217;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r236;
cvt.f32.f16 %f87, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r236;
cvt.f32.f16 %f88, high;}


cvt.rn.f32.s32 %f119, %r232;
div.approx.ftz.f32 %f120, %f88, %f98;
fma.rn.ftz.f32 %f121, %f87, %f119, %f120;
add.ftz.f32 %f282, %f121, %f282;
cvt.u64.u32 %rd47, %r570;
add.s64 %rd48, %rd1, %rd47;
mul.lo.s64 %rd49, %rd48, 36;
add.s64 %rd50, %rd6, %rd49;
add.s64 %rd51, %rd50, %rd28;
ld.global.nc.u32 %r243, [%rd51+4];
ld.global.nc.u32 %r247, [%rd51+20];
ld.global.nc.u32 %r251, [%rd51+8];
ld.global.nc.u32 %r255, [%rd51+24];

dp4a.s32.s32 %r241, %r284, %r243, %r307;


dp4a.s32.s32 %r245, %r288, %r247, %r241;


dp4a.s32.s32 %r249, %r292, %r251, %r245;


dp4a.s32.s32 %r253, %r296, %r255, %r249;

ld.global.nc.u32 %r259, [%rd50];

{mul.f16x2 %r257,%r90,%r259;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r257;
cvt.f32.f16 %f89, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r257;
cvt.f32.f16 %f90, high;}


cvt.rn.f32.s32 %f122, %r253;
div.approx.ftz.f32 %f123, %f90, %f98;
fma.rn.ftz.f32 %f124, %f89, %f122, %f123;
add.ftz.f32 %f281, %f124, %f281;

dp4a.s32.s32 %r262, %r305, %r243, %r307;


dp4a.s32.s32 %r266, %r309, %r247, %r262;


dp4a.s32.s32 %r270, %r313, %r251, %r266;


dp4a.s32.s32 %r274, %r317, %r255, %r270;


{mul.f16x2 %r278,%r111,%r259;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r278;
cvt.f32.f16 %f91, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r278;
cvt.f32.f16 %f92, high;}


cvt.rn.f32.s32 %f125, %r274;
div.approx.ftz.f32 %f126, %f92, %f98;
fma.rn.ftz.f32 %f127, %f91, %f125, %f126;
add.ftz.f32 %f280, %f127, %f280;
cvt.u64.u32 %rd52, %r569;
add.s64 %rd53, %rd1, %rd52;
mul.lo.s64 %rd54, %rd53, 36;
add.s64 %rd55, %rd6, %rd54;
add.s64 %rd56, %rd55, %rd28;
ld.global.nc.u32 %r285, [%rd56+4];
ld.global.nc.u32 %r289, [%rd56+20];
ld.global.nc.u32 %r293, [%rd56+8];
ld.global.nc.u32 %r297, [%rd56+24];

dp4a.s32.s32 %r283, %r284, %r285, %r307;


dp4a.s32.s32 %r287, %r288, %r289, %r283;


dp4a.s32.s32 %r291, %r292, %r293, %r287;


dp4a.s32.s32 %r295, %r296, %r297, %r291;

ld.global.nc.u32 %r301, [%rd55];

{mul.f16x2 %r299,%r90,%r301;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r299;
cvt.f32.f16 %f93, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r299;
cvt.f32.f16 %f94, high;}


cvt.rn.f32.s32 %f128, %r295;
div.approx.ftz.f32 %f129, %f94, %f98;
fma.rn.ftz.f32 %f130, %f93, %f128, %f129;
add.ftz.f32 %f279, %f130, %f279;

dp4a.s32.s32 %r304, %r305, %r285, %r307;


dp4a.s32.s32 %r308, %r309, %r289, %r304;


dp4a.s32.s32 %r312, %r313, %r293, %r308;


dp4a.s32.s32 %r316, %r317, %r297, %r312;


{mul.f16x2 %r320,%r111,%r301;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r320;
cvt.f32.f16 %f95, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r320;
cvt.f32.f16 %f96, high;}


cvt.rn.f32.s32 %f131, %r316;
div.approx.ftz.f32 %f132, %f96, %f98;
fma.rn.ftz.f32 %f133, %f95, %f131, %f132;
add.ftz.f32 %f278, %f133, %f278;
add.s32 %r572, %r572, 32;
add.s32 %r571, %r571, 32;
add.s32 %r570, %r570, 32;
add.s32 %r569, %r569, 32;
add.s32 %r568, %r568, 32;
add.s64 %rd102, %rd102, 640;
add.s64 %rd101, %rd101, 640;
add.s32 %r573, %r573, 32;
setp.lt.s32 %p4, %r573, %r6;
@%p4 bra $L__BB53_2;

$L__BB53_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB53_5;

mov.u32 %r333, _ZZ13mul_mat_vec_qIL9ggml_type3ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r334, %r1, 1536, %r333;
shl.b32 %r335, %r2, 2;
add.s32 %r336, %r334, %r335;
st.shared.f32 [%r336+-1536], %f289;
st.shared.f32 [%r336+-1408], %f288;
st.shared.f32 [%r336+-1280], %f287;
st.shared.f32 [%r336+-1152], %f286;
st.shared.f32 [%r336+-1024], %f285;
st.shared.f32 [%r336+-896], %f284;
st.shared.f32 [%r336+-768], %f283;
st.shared.f32 [%r336+-640], %f282;
st.shared.f32 [%r336+-512], %f281;
st.shared.f32 [%r336+-384], %f280;
st.shared.f32 [%r336+-256], %f279;
st.shared.f32 [%r336+-128], %f278;

$L__BB53_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB53_30;

mov.u32 %r337, %ctaid.z;
mov.u32 %r338, %ctaid.y;
mad.lo.s32 %r339, %r338, %r35, %r4;
mad.lo.s32 %r340, %r337, %r41, %r339;
cvt.u64.u32 %rd11, %r340;
shl.b32 %r341, %r2, 2;
mov.u32 %r342, _ZZ13mul_mat_vec_qIL9ggml_type3ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r24, %r342, %r341;
mov.u32 %r343, 2;
ld.shared.f32 %f134, [%r24];
add.ftz.f32 %f135, %f134, %f289;
mov.b32 %r344, %f135;
mov.u32 %r345, 31;
mov.u32 %r346, 16;
mov.u32 %r347, -1;
shfl.sync.bfly.b32 %r348|%p7, %r344, %r346, %r345, %r347;
mov.b32 %f136, %r348;
add.ftz.f32 %f137, %f135, %f136;
mov.b32 %r349, %f137;
mov.u32 %r350, 8;
shfl.sync.bfly.b32 %r351|%p8, %r349, %r350, %r345, %r347;
mov.b32 %f138, %r351;
add.ftz.f32 %f139, %f137, %f138;
mov.b32 %r352, %f139;
mov.u32 %r353, 4;
shfl.sync.bfly.b32 %r354|%p9, %r352, %r353, %r345, %r347;
mov.b32 %f140, %r354;
add.ftz.f32 %f141, %f139, %f140;
mov.b32 %r355, %f141;
shfl.sync.bfly.b32 %r356|%p10, %r355, %r343, %r345, %r347;
mov.b32 %f142, %r356;
add.ftz.f32 %f143, %f141, %f142;
mov.b32 %r357, %f143;
mov.u32 %r358, 1;
shfl.sync.bfly.b32 %r359|%p11, %r357, %r358, %r345, %r347;
mov.b32 %f144, %r359;
add.ftz.f32 %f37, %f143, %f144;
setp.lt.u32 %p12, %r4, %r29;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
not.pred %p14, %p1;
@%p14 bra $L__BB53_8;

cvta.to.global.u64 %rd57, %rd16;
mul.wide.u32 %rd58, %r340, 4;
add.s64 %rd59, %rd57, %rd58;
st.global.f32 [%rd59], %f37;

$L__BB53_8:
ld.shared.f32 %f145, [%r24+128];
add.ftz.f32 %f146, %f145, %f288;
mov.b32 %r366, %f146;
shfl.sync.bfly.b32 %r370|%p15, %r366, %r346, %r345, %r347;
mov.b32 %f147, %r370;
add.ftz.f32 %f148, %f146, %f147;
mov.b32 %r371, %f148;
shfl.sync.bfly.b32 %r373|%p16, %r371, %r350, %r345, %r347;
mov.b32 %f149, %r373;
add.ftz.f32 %f150, %f148, %f149;
mov.b32 %r374, %f150;
shfl.sync.bfly.b32 %r376|%p17, %r374, %r353, %r345, %r347;
mov.b32 %f151, %r376;
add.ftz.f32 %f152, %f150, %f151;
mov.b32 %r377, %f152;
shfl.sync.bfly.b32 %r379|%p18, %r377, %r343, %r345, %r347;
mov.b32 %f153, %r379;
add.ftz.f32 %f154, %f152, %f153;
mov.b32 %r380, %f154;
shfl.sync.bfly.b32 %r382|%p19, %r380, %r358, %r345, %r347;
mov.b32 %f155, %r382;
add.ftz.f32 %f38, %f154, %f155;
setp.eq.s32 %p20, %r2, 1;
or.b32 %r386, %r4, 1;
setp.lt.u32 %p21, %r386, %r29;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB53_10;

cvta.to.global.u64 %rd60, %rd16;
mul.wide.u32 %rd61, %r340, 4;
add.s64 %rd62, %rd60, %rd61;
st.global.f32 [%rd62+4], %f38;

$L__BB53_10:
ld.shared.f32 %f156, [%r24+256];
add.ftz.f32 %f157, %f156, %f287;
mov.b32 %r393, %f157;
mov.u32 %r394, 31;
mov.u32 %r395, 16;
mov.u32 %r396, -1;
shfl.sync.bfly.b32 %r397|%p23, %r393, %r395, %r394, %r396;
mov.b32 %f158, %r397;
add.ftz.f32 %f159, %f157, %f158;
mov.b32 %r398, %f159;
mov.u32 %r399, 8;
shfl.sync.bfly.b32 %r400|%p24, %r398, %r399, %r394, %r396;
mov.b32 %f160, %r400;
add.ftz.f32 %f161, %f159, %f160;
mov.b32 %r401, %f161;
mov.u32 %r402, 4;
shfl.sync.bfly.b32 %r403|%p25, %r401, %r402, %r394, %r396;
mov.b32 %f162, %r403;
add.ftz.f32 %f163, %f161, %f162;
mov.b32 %r404, %f163;
mov.u32 %r405, 2;
shfl.sync.bfly.b32 %r406|%p26, %r404, %r405, %r394, %r396;
mov.b32 %f164, %r406;
add.ftz.f32 %f165, %f163, %f164;
mov.b32 %r407, %f165;
mov.u32 %r408, 1;
shfl.sync.bfly.b32 %r409|%p27, %r407, %r408, %r394, %r396;
mov.b32 %f166, %r409;
add.ftz.f32 %f39, %f165, %f166;
@%p14 bra $L__BB53_12;

cvt.u64.u32 %rd63, %r29;
add.s64 %rd64, %rd63, %rd11;
cvta.to.global.u64 %rd65, %rd16;
shl.b64 %rd66, %rd64, 2;
add.s64 %rd67, %rd65, %rd66;
st.global.f32 [%rd67], %f39;

$L__BB53_12:
ld.shared.f32 %f167, [%r24+384];
add.ftz.f32 %f168, %f167, %f286;
mov.b32 %r410, %f168;
shfl.sync.bfly.b32 %r414|%p29, %r410, %r395, %r394, %r396;
mov.b32 %f169, %r414;
add.ftz.f32 %f170, %f168, %f169;
mov.b32 %r415, %f170;
shfl.sync.bfly.b32 %r417|%p30, %r415, %r399, %r394, %r396;
mov.b32 %f171, %r417;
add.ftz.f32 %f172, %f170, %f171;
mov.b32 %r418, %f172;
shfl.sync.bfly.b32 %r420|%p31, %r418, %r402, %r394, %r396;
mov.b32 %f173, %r420;
add.ftz.f32 %f174, %f172, %f173;
mov.b32 %r421, %f174;
shfl.sync.bfly.b32 %r423|%p32, %r421, %r405, %r394, %r396;
mov.b32 %f175, %r423;
add.ftz.f32 %f176, %f174, %f175;
mov.b32 %r424, %f176;
shfl.sync.bfly.b32 %r426|%p33, %r424, %r408, %r394, %r396;
mov.b32 %f177, %r426;
add.ftz.f32 %f40, %f176, %f177;
@%p22 bra $L__BB53_14;

add.s32 %r427, %r29, 1;
cvt.u64.u32 %rd68, %r427;
add.s64 %rd69, %rd68, %rd11;
cvta.to.global.u64 %rd70, %rd16;
shl.b64 %rd71, %rd69, 2;
add.s64 %rd72, %rd70, %rd71;
st.global.f32 [%rd72], %f40;

$L__BB53_14:
shl.b32 %r428, %r29, 1;
mov.u32 %r429, 1;
ld.shared.f32 %f178, [%r24+512];
add.ftz.f32 %f179, %f178, %f285;
mov.b32 %r430, %f179;
mov.u32 %r431, 31;
mov.u32 %r432, 16;
mov.u32 %r433, -1;
shfl.sync.bfly.b32 %r434|%p35, %r430, %r432, %r431, %r433;
mov.b32 %f180, %r434;
add.ftz.f32 %f181, %f179, %f180;
mov.b32 %r435, %f181;
mov.u32 %r436, 8;
shfl.sync.bfly.b32 %r437|%p36, %r435, %r436, %r431, %r433;
mov.b32 %f182, %r437;
add.ftz.f32 %f183, %f181, %f182;
mov.b32 %r438, %f183;
mov.u32 %r439, 4;
shfl.sync.bfly.b32 %r440|%p37, %r438, %r439, %r431, %r433;
mov.b32 %f184, %r440;
add.ftz.f32 %f185, %f183, %f184;
mov.b32 %r441, %f185;
mov.u32 %r442, 2;
shfl.sync.bfly.b32 %r443|%p38, %r441, %r442, %r431, %r433;
mov.b32 %f186, %r443;
add.ftz.f32 %f187, %f185, %f186;
mov.b32 %r444, %f187;
shfl.sync.bfly.b32 %r445|%p39, %r444, %r429, %r431, %r433;
mov.b32 %f188, %r445;
add.ftz.f32 %f41, %f187, %f188;
cvt.u64.u32 %rd73, %r428;
add.s64 %rd74, %rd73, %rd11;
cvta.to.global.u64 %rd75, %rd16;
shl.b64 %rd76, %rd74, 2;
add.s64 %rd12, %rd75, %rd76;
@%p14 bra $L__BB53_16;

st.global.f32 [%rd12], %f41;

$L__BB53_16:
ld.shared.f32 %f189, [%r24+640];
add.ftz.f32 %f190, %f189, %f284;
mov.b32 %r446, %f190;
shfl.sync.bfly.b32 %r450|%p41, %r446, %r432, %r431, %r433;
mov.b32 %f191, %r450;
add.ftz.f32 %f192, %f190, %f191;
mov.b32 %r451, %f192;
shfl.sync.bfly.b32 %r453|%p42, %r451, %r436, %r431, %r433;
mov.b32 %f193, %r453;
add.ftz.f32 %f194, %f192, %f193;
mov.b32 %r454, %f194;
shfl.sync.bfly.b32 %r456|%p43, %r454, %r439, %r431, %r433;
mov.b32 %f195, %r456;
add.ftz.f32 %f196, %f194, %f195;
mov.b32 %r457, %f196;
shfl.sync.bfly.b32 %r459|%p44, %r457, %r442, %r431, %r433;
mov.b32 %f197, %r459;
add.ftz.f32 %f198, %f196, %f197;
mov.b32 %r460, %f198;
shfl.sync.bfly.b32 %r462|%p45, %r460, %r429, %r431, %r433;
mov.b32 %f199, %r462;
add.ftz.f32 %f42, %f198, %f199;
@%p22 bra $L__BB53_18;

st.global.f32 [%rd12+4], %f42;

$L__BB53_18:
mul.lo.s32 %r25, %r29, 3;
ld.shared.f32 %f200, [%r24+768];
add.ftz.f32 %f201, %f200, %f283;
mov.b32 %r463, %f201;
mov.u32 %r464, 31;
mov.u32 %r465, 16;
mov.u32 %r466, -1;
shfl.sync.bfly.b32 %r467|%p47, %r463, %r465, %r464, %r466;
mov.b32 %f202, %r467;
add.ftz.f32 %f203, %f201, %f202;
mov.b32 %r468, %f203;
mov.u32 %r469, 8;
shfl.sync.bfly.b32 %r470|%p48, %r468, %r469, %r464, %r466;
mov.b32 %f204, %r470;
add.ftz.f32 %f205, %f203, %f204;
mov.b32 %r471, %f205;
mov.u32 %r472, 4;
shfl.sync.bfly.b32 %r473|%p49, %r471, %r472, %r464, %r466;
mov.b32 %f206, %r473;
add.ftz.f32 %f207, %f205, %f206;
mov.b32 %r474, %f207;
mov.u32 %r475, 2;
shfl.sync.bfly.b32 %r476|%p50, %r474, %r475, %r464, %r466;
mov.b32 %f208, %r476;
add.ftz.f32 %f209, %f207, %f208;
mov.b32 %r477, %f209;
mov.u32 %r478, 1;
shfl.sync.bfly.b32 %r479|%p51, %r477, %r478, %r464, %r466;
mov.b32 %f210, %r479;
add.ftz.f32 %f43, %f209, %f210;
@%p14 bra $L__BB53_20;

cvt.u64.u32 %rd77, %r25;
add.s64 %rd78, %rd77, %rd11;
shl.b64 %rd80, %rd78, 2;
add.s64 %rd81, %rd75, %rd80;
st.global.f32 [%rd81], %f43;

$L__BB53_20:
ld.shared.f32 %f211, [%r24+896];
add.ftz.f32 %f212, %f211, %f282;
mov.b32 %r480, %f212;
shfl.sync.bfly.b32 %r484|%p53, %r480, %r465, %r464, %r466;
mov.b32 %f213, %r484;
add.ftz.f32 %f214, %f212, %f213;
mov.b32 %r485, %f214;
shfl.sync.bfly.b32 %r487|%p54, %r485, %r469, %r464, %r466;
mov.b32 %f215, %r487;
add.ftz.f32 %f216, %f214, %f215;
mov.b32 %r488, %f216;
shfl.sync.bfly.b32 %r490|%p55, %r488, %r472, %r464, %r466;
mov.b32 %f217, %r490;
add.ftz.f32 %f218, %f216, %f217;
mov.b32 %r491, %f218;
shfl.sync.bfly.b32 %r493|%p56, %r491, %r475, %r464, %r466;
mov.b32 %f219, %r493;
add.ftz.f32 %f220, %f218, %f219;
mov.b32 %r494, %f220;
shfl.sync.bfly.b32 %r496|%p57, %r494, %r478, %r464, %r466;
mov.b32 %f221, %r496;
add.ftz.f32 %f44, %f220, %f221;
@%p22 bra $L__BB53_22;

add.s32 %r497, %r25, 1;
cvt.u64.u32 %rd82, %r497;
add.s64 %rd83, %rd82, %rd11;
shl.b64 %rd85, %rd83, 2;
add.s64 %rd86, %rd75, %rd85;
st.global.f32 [%rd86], %f44;

$L__BB53_22:
shl.b32 %r498, %r29, 2;
mov.u32 %r499, 2;
ld.shared.f32 %f222, [%r24+1024];
add.ftz.f32 %f223, %f222, %f281;
mov.b32 %r500, %f223;
mov.u32 %r501, 31;
mov.u32 %r502, 16;
mov.u32 %r503, -1;
shfl.sync.bfly.b32 %r504|%p59, %r500, %r502, %r501, %r503;
mov.b32 %f224, %r504;
add.ftz.f32 %f225, %f223, %f224;
mov.b32 %r505, %f225;
mov.u32 %r506, 8;
shfl.sync.bfly.b32 %r507|%p60, %r505, %r506, %r501, %r503;
mov.b32 %f226, %r507;
add.ftz.f32 %f227, %f225, %f226;
mov.b32 %r508, %f227;
mov.u32 %r509, 4;
shfl.sync.bfly.b32 %r510|%p61, %r508, %r509, %r501, %r503;
mov.b32 %f228, %r510;
add.ftz.f32 %f229, %f227, %f228;
mov.b32 %r511, %f229;
shfl.sync.bfly.b32 %r512|%p62, %r511, %r499, %r501, %r503;
mov.b32 %f230, %r512;
add.ftz.f32 %f231, %f229, %f230;
mov.b32 %r513, %f231;
mov.u32 %r514, 1;
shfl.sync.bfly.b32 %r515|%p63, %r513, %r514, %r501, %r503;
mov.b32 %f232, %r515;
add.ftz.f32 %f45, %f231, %f232;
cvt.u64.u32 %rd87, %r498;
add.s64 %rd88, %rd87, %rd11;
shl.b64 %rd90, %rd88, 2;
add.s64 %rd13, %rd75, %rd90;
@%p14 bra $L__BB53_24;

st.global.f32 [%rd13], %f45;

$L__BB53_24:
ld.shared.f32 %f233, [%r24+1152];
add.ftz.f32 %f234, %f233, %f280;
mov.b32 %r516, %f234;
shfl.sync.bfly.b32 %r520|%p65, %r516, %r502, %r501, %r503;
mov.b32 %f235, %r520;
add.ftz.f32 %f236, %f234, %f235;
mov.b32 %r521, %f236;
shfl.sync.bfly.b32 %r523|%p66, %r521, %r506, %r501, %r503;
mov.b32 %f237, %r523;
add.ftz.f32 %f238, %f236, %f237;
mov.b32 %r524, %f238;
shfl.sync.bfly.b32 %r526|%p67, %r524, %r509, %r501, %r503;
mov.b32 %f239, %r526;
add.ftz.f32 %f240, %f238, %f239;
mov.b32 %r527, %f240;
shfl.sync.bfly.b32 %r529|%p68, %r527, %r499, %r501, %r503;
mov.b32 %f241, %r529;
add.ftz.f32 %f242, %f240, %f241;
mov.b32 %r530, %f242;
shfl.sync.bfly.b32 %r532|%p69, %r530, %r514, %r501, %r503;
mov.b32 %f243, %r532;
add.ftz.f32 %f46, %f242, %f243;
@%p22 bra $L__BB53_26;

st.global.f32 [%rd13+4], %f46;

$L__BB53_26:
mul.lo.s32 %r26, %r29, 5;
ld.shared.f32 %f244, [%r24+1280];
add.ftz.f32 %f245, %f244, %f279;
mov.b32 %r533, %f245;
mov.u32 %r534, 31;
mov.u32 %r535, 16;
mov.u32 %r536, -1;
shfl.sync.bfly.b32 %r537|%p71, %r533, %r535, %r534, %r536;
mov.b32 %f246, %r537;
add.ftz.f32 %f247, %f245, %f246;
mov.b32 %r538, %f247;
mov.u32 %r539, 8;
shfl.sync.bfly.b32 %r540|%p72, %r538, %r539, %r534, %r536;
mov.b32 %f248, %r540;
add.ftz.f32 %f249, %f247, %f248;
mov.b32 %r541, %f249;
mov.u32 %r542, 4;
shfl.sync.bfly.b32 %r543|%p73, %r541, %r542, %r534, %r536;
mov.b32 %f250, %r543;
add.ftz.f32 %f251, %f249, %f250;
mov.b32 %r544, %f251;
mov.u32 %r545, 2;
shfl.sync.bfly.b32 %r546|%p74, %r544, %r545, %r534, %r536;
mov.b32 %f252, %r546;
add.ftz.f32 %f253, %f251, %f252;
mov.b32 %r547, %f253;
mov.u32 %r548, 1;
shfl.sync.bfly.b32 %r549|%p75, %r547, %r548, %r534, %r536;
mov.b32 %f254, %r549;
add.ftz.f32 %f47, %f253, %f254;
@%p14 bra $L__BB53_28;

cvt.u64.u32 %rd91, %r26;
add.s64 %rd92, %rd91, %rd11;
shl.b64 %rd94, %rd92, 2;
add.s64 %rd95, %rd75, %rd94;
st.global.f32 [%rd95], %f47;

$L__BB53_28:
ld.shared.f32 %f255, [%r24+1408];
add.ftz.f32 %f256, %f255, %f278;
mov.b32 %r550, %f256;
shfl.sync.bfly.b32 %r554|%p77, %r550, %r535, %r534, %r536;
mov.b32 %f257, %r554;
add.ftz.f32 %f258, %f256, %f257;
mov.b32 %r555, %f258;
shfl.sync.bfly.b32 %r557|%p78, %r555, %r539, %r534, %r536;
mov.b32 %f259, %r557;
add.ftz.f32 %f260, %f258, %f259;
mov.b32 %r558, %f260;
shfl.sync.bfly.b32 %r560|%p79, %r558, %r542, %r534, %r536;
mov.b32 %f261, %r560;
add.ftz.f32 %f262, %f260, %f261;
mov.b32 %r561, %f262;
shfl.sync.bfly.b32 %r563|%p80, %r561, %r545, %r534, %r536;
mov.b32 %f263, %r563;
add.ftz.f32 %f264, %f262, %f263;
mov.b32 %r564, %f264;
shfl.sync.bfly.b32 %r566|%p81, %r564, %r548, %r534, %r536;
mov.b32 %f265, %r566;
add.ftz.f32 %f48, %f264, %f265;
@%p22 bra $L__BB53_30;

add.s32 %r567, %r26, 1;
cvt.u64.u32 %rd96, %r567;
add.s64 %rd97, %rd96, %rd11;
shl.b64 %rd99, %rd97, 2;
add.s64 %rd100, %rd75, %rd99;
st.global.f32 [%rd100], %f48;

$L__BB53_30:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type3ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type3ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type3ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type3ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type3ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type3ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type3ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type3ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type3ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<95>;
.reg .f32 %f<338>;
.reg .b32 %r<643>;
.reg .b64 %rd<110>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type3ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1792];

ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type3ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type3ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type3ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r45, [_Z13mul_mat_vec_qIL9ggml_type3ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type3ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type3ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type3ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type3ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type3ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type3ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type3ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type3ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type3ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type3ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type3ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type3ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type3ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r46, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r46, %r2;
mov.u32 %r47, %ctaid.x;
shl.b32 %r4, %r47, 1;
shr.u32 %r48, %r3, 31;
add.s32 %r49, %r3, %r48;
shr.s32 %r642, %r49, 1;
shr.u32 %r6, %r45, 5;
setp.ge.s32 %p3, %r642, %r6;
mov.f32 %f324, 0f00000000;
mov.f32 %f325, %f324;
mov.f32 %f326, %f324;
mov.f32 %f327, %f324;
mov.f32 %f328, %f324;
mov.f32 %f329, %f324;
mov.f32 %f330, %f324;
mov.f32 %f331, %f324;
mov.f32 %f332, %f324;
mov.f32 %f333, %f324;
mov.f32 %f334, %f324;
mov.f32 %f335, %f324;
mov.f32 %f336, %f324;
mov.f32 %f337, %f324;
@%p3 bra $L__BB54_3;

mov.u32 %r50, %ctaid.y;
mov.u32 %r51, %ctaid.z;
mul.lo.s32 %r52, %r50, %r37;
cvt.u64.u32 %rd19, %r52;
mul.lo.s32 %r53, %r51, %r43;
cvt.u64.u32 %rd20, %r53;
add.s64 %rd1, %rd20, %rd19;
mul.hi.u32 %r54, %r51, %r39;
add.s32 %r55, %r54, %r51;
shr.u32 %r56, %r55, %r40;
mul.lo.s32 %r57, %r56, %r42;
mul.hi.u32 %r58, %r50, %r33;
add.s32 %r59, %r58, %r50;
shr.u32 %r60, %r59, %r34;
mul.lo.s32 %r61, %r60, %r36;
cvta.to.global.u64 %rd21, %rd16;
and.b32 %r64, %r49, 2147483646;
sub.s32 %r65, %r3, %r64;
shl.b32 %r66, %r65, 1;
cvt.s64.s32 %rd2, %r66;
shl.b32 %r67, %r31, 1;
add.s32 %r641, %r642, %r67;
mad.lo.s32 %r640, %r31, 3, %r642;
shl.b32 %r68, %r31, 2;
add.s32 %r639, %r642, %r68;
mad.lo.s32 %r638, %r31, 5, %r642;
mad.lo.s32 %r637, %r31, 6, %r642;
add.s32 %r636, %r642, %r31;
add.s32 %r69, %r642, %r57;
add.s32 %r70, %r69, %r61;
add.s32 %r71, %r4, 1;
mad.lo.s32 %r72, %r30, %r71, %r70;
mul.wide.s32 %rd22, %r72, 20;
add.s64 %rd23, %rd21, %rd22;
add.s64 %rd109, %rd23, 8;
mul.wide.s32 %rd4, %r66, 4;
mad.lo.s32 %r73, %r4, %r30, %r642;
add.s32 %r74, %r73, %r57;
add.s32 %r75, %r74, %r61;
mul.wide.s32 %rd24, %r75, 20;
add.s64 %rd25, %rd21, %rd24;
add.s64 %rd108, %rd25, 8;
cvta.to.global.u64 %rd6, %rd17;
mov.f32 %f324, 0f00000000;
shl.b64 %rd30, %rd2, 2;

$L__BB54_2:
cvt.u64.u32 %rd26, %r642;
add.s64 %rd27, %rd1, %rd26;
mul.lo.s64 %rd28, %rd27, 36;
add.s64 %rd29, %rd6, %rd28;
add.s64 %rd31, %rd29, %rd30;
ld.global.nc.u32 %r78, [%rd31+4];
ld.global.nc.u32 %r82, [%rd31+20];
add.s64 %rd32, %rd108, %rd4;
ld.global.nc.u32 %r370, [%rd32];
ld.global.nc.u32 %r86, [%rd31+8];
ld.global.nc.u32 %r90, [%rd31+24];
ld.global.nc.u32 %r371, [%rd32+-4];
and.b32 %r329, %r371, 252645135;
shr.u32 %r372, %r371, 4;
and.b32 %r333, %r372, 252645135;
mov.u32 %r352, 0;

dp4a.s32.s32 %r76, %r329, %r78, %r352;


dp4a.s32.s32 %r80, %r333, %r82, %r76;

and.b32 %r337, %r370, 252645135;
shr.u32 %r373, %r370, 4;
and.b32 %r341, %r373, 252645135;

dp4a.s32.s32 %r84, %r337, %r86, %r80;


dp4a.s32.s32 %r88, %r341, %r90, %r84;

ld.global.nc.u32 %r93, [%rd108+-8];
ld.global.nc.u32 %r94, [%rd29];

{mul.f16x2 %r92,%r93,%r94;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r92;
cvt.f32.f16 %f85, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r92;
cvt.f32.f16 %f86, high;}


cvt.rn.f32.s32 %f113, %r88;
mov.f32 %f114, 0f40000000;
div.approx.ftz.f32 %f115, %f86, %f114;
fma.rn.ftz.f32 %f116, %f85, %f113, %f115;
add.ftz.f32 %f337, %f116, %f337;
add.s64 %rd33, %rd109, %rd4;
ld.global.nc.u32 %r374, [%rd33];
ld.global.nc.u32 %r375, [%rd33+-4];
and.b32 %r350, %r375, 252645135;
shr.u32 %r376, %r375, 4;
and.b32 %r354, %r376, 252645135;

dp4a.s32.s32 %r97, %r350, %r78, %r352;


dp4a.s32.s32 %r101, %r354, %r82, %r97;

and.b32 %r358, %r374, 252645135;
shr.u32 %r377, %r374, 4;
and.b32 %r362, %r377, 252645135;

dp4a.s32.s32 %r105, %r358, %r86, %r101;


dp4a.s32.s32 %r109, %r362, %r90, %r105;

ld.global.nc.u32 %r114, [%rd109+-8];

{mul.f16x2 %r113,%r114,%r94;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r113;
cvt.f32.f16 %f87, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r113;
cvt.f32.f16 %f88, high;}


cvt.rn.f32.s32 %f117, %r109;
div.approx.ftz.f32 %f118, %f88, %f114;
fma.rn.ftz.f32 %f119, %f87, %f117, %f118;
add.ftz.f32 %f336, %f119, %f336;
cvt.u64.u32 %rd34, %r636;
add.s64 %rd35, %rd1, %rd34;
mul.lo.s64 %rd36, %rd35, 36;
add.s64 %rd37, %rd6, %rd36;
add.s64 %rd38, %rd37, %rd30;
ld.global.nc.u32 %r120, [%rd38+4];
ld.global.nc.u32 %r124, [%rd38+20];
ld.global.nc.u32 %r128, [%rd38+8];
ld.global.nc.u32 %r132, [%rd38+24];

dp4a.s32.s32 %r118, %r329, %r120, %r352;


dp4a.s32.s32 %r122, %r333, %r124, %r118;


dp4a.s32.s32 %r126, %r337, %r128, %r122;


dp4a.s32.s32 %r130, %r341, %r132, %r126;

ld.global.nc.u32 %r136, [%rd37];

{mul.f16x2 %r134,%r93,%r136;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r134;
cvt.f32.f16 %f89, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r134;
cvt.f32.f16 %f90, high;}


cvt.rn.f32.s32 %f120, %r130;
div.approx.ftz.f32 %f121, %f90, %f114;
fma.rn.ftz.f32 %f122, %f89, %f120, %f121;
add.ftz.f32 %f335, %f122, %f335;

dp4a.s32.s32 %r139, %r350, %r120, %r352;


dp4a.s32.s32 %r143, %r354, %r124, %r139;


dp4a.s32.s32 %r147, %r358, %r128, %r143;


dp4a.s32.s32 %r151, %r362, %r132, %r147;


{mul.f16x2 %r155,%r114,%r136;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r155;
cvt.f32.f16 %f91, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r155;
cvt.f32.f16 %f92, high;}


cvt.rn.f32.s32 %f123, %r151;
div.approx.ftz.f32 %f124, %f92, %f114;
fma.rn.ftz.f32 %f125, %f91, %f123, %f124;
add.ftz.f32 %f334, %f125, %f334;
cvt.u64.u32 %rd39, %r641;
add.s64 %rd40, %rd1, %rd39;
mul.lo.s64 %rd41, %rd40, 36;
add.s64 %rd42, %rd6, %rd41;
add.s64 %rd43, %rd42, %rd30;
ld.global.nc.u32 %r162, [%rd43+4];
ld.global.nc.u32 %r166, [%rd43+20];
ld.global.nc.u32 %r170, [%rd43+8];
ld.global.nc.u32 %r174, [%rd43+24];

dp4a.s32.s32 %r160, %r329, %r162, %r352;


dp4a.s32.s32 %r164, %r333, %r166, %r160;


dp4a.s32.s32 %r168, %r337, %r170, %r164;


dp4a.s32.s32 %r172, %r341, %r174, %r168;

ld.global.nc.u32 %r178, [%rd42];

{mul.f16x2 %r176,%r93,%r178;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r176;
cvt.f32.f16 %f93, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r176;
cvt.f32.f16 %f94, high;}


cvt.rn.f32.s32 %f126, %r172;
div.approx.ftz.f32 %f127, %f94, %f114;
fma.rn.ftz.f32 %f128, %f93, %f126, %f127;
add.ftz.f32 %f333, %f128, %f333;

dp4a.s32.s32 %r181, %r350, %r162, %r352;


dp4a.s32.s32 %r185, %r354, %r166, %r181;


dp4a.s32.s32 %r189, %r358, %r170, %r185;


dp4a.s32.s32 %r193, %r362, %r174, %r189;


{mul.f16x2 %r197,%r114,%r178;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r197;
cvt.f32.f16 %f95, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r197;
cvt.f32.f16 %f96, high;}


cvt.rn.f32.s32 %f129, %r193;
div.approx.ftz.f32 %f130, %f96, %f114;
fma.rn.ftz.f32 %f131, %f95, %f129, %f130;
add.ftz.f32 %f332, %f131, %f332;
cvt.u64.u32 %rd44, %r640;
add.s64 %rd45, %rd1, %rd44;
mul.lo.s64 %rd46, %rd45, 36;
add.s64 %rd47, %rd6, %rd46;
add.s64 %rd48, %rd47, %rd30;
ld.global.nc.u32 %r204, [%rd48+4];
ld.global.nc.u32 %r208, [%rd48+20];
ld.global.nc.u32 %r212, [%rd48+8];
ld.global.nc.u32 %r216, [%rd48+24];

dp4a.s32.s32 %r202, %r329, %r204, %r352;


dp4a.s32.s32 %r206, %r333, %r208, %r202;


dp4a.s32.s32 %r210, %r337, %r212, %r206;


dp4a.s32.s32 %r214, %r341, %r216, %r210;

ld.global.nc.u32 %r220, [%rd47];

{mul.f16x2 %r218,%r93,%r220;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r218;
cvt.f32.f16 %f97, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r218;
cvt.f32.f16 %f98, high;}


cvt.rn.f32.s32 %f132, %r214;
div.approx.ftz.f32 %f133, %f98, %f114;
fma.rn.ftz.f32 %f134, %f97, %f132, %f133;
add.ftz.f32 %f331, %f134, %f331;

dp4a.s32.s32 %r223, %r350, %r204, %r352;


dp4a.s32.s32 %r227, %r354, %r208, %r223;


dp4a.s32.s32 %r231, %r358, %r212, %r227;


dp4a.s32.s32 %r235, %r362, %r216, %r231;


{mul.f16x2 %r239,%r114,%r220;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r239;
cvt.f32.f16 %f99, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r239;
cvt.f32.f16 %f100, high;}


cvt.rn.f32.s32 %f135, %r235;
div.approx.ftz.f32 %f136, %f100, %f114;
fma.rn.ftz.f32 %f137, %f99, %f135, %f136;
add.ftz.f32 %f330, %f137, %f330;
cvt.u64.u32 %rd49, %r639;
add.s64 %rd50, %rd1, %rd49;
mul.lo.s64 %rd51, %rd50, 36;
add.s64 %rd52, %rd6, %rd51;
add.s64 %rd53, %rd52, %rd30;
ld.global.nc.u32 %r246, [%rd53+4];
ld.global.nc.u32 %r250, [%rd53+20];
ld.global.nc.u32 %r254, [%rd53+8];
ld.global.nc.u32 %r258, [%rd53+24];

dp4a.s32.s32 %r244, %r329, %r246, %r352;


dp4a.s32.s32 %r248, %r333, %r250, %r244;


dp4a.s32.s32 %r252, %r337, %r254, %r248;


dp4a.s32.s32 %r256, %r341, %r258, %r252;

ld.global.nc.u32 %r262, [%rd52];

{mul.f16x2 %r260,%r93,%r262;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r260;
cvt.f32.f16 %f101, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r260;
cvt.f32.f16 %f102, high;}


cvt.rn.f32.s32 %f138, %r256;
div.approx.ftz.f32 %f139, %f102, %f114;
fma.rn.ftz.f32 %f140, %f101, %f138, %f139;
add.ftz.f32 %f329, %f140, %f329;

dp4a.s32.s32 %r265, %r350, %r246, %r352;


dp4a.s32.s32 %r269, %r354, %r250, %r265;


dp4a.s32.s32 %r273, %r358, %r254, %r269;


dp4a.s32.s32 %r277, %r362, %r258, %r273;


{mul.f16x2 %r281,%r114,%r262;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r281;
cvt.f32.f16 %f103, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r281;
cvt.f32.f16 %f104, high;}


cvt.rn.f32.s32 %f141, %r277;
div.approx.ftz.f32 %f142, %f104, %f114;
fma.rn.ftz.f32 %f143, %f103, %f141, %f142;
add.ftz.f32 %f328, %f143, %f328;
cvt.u64.u32 %rd54, %r638;
add.s64 %rd55, %rd1, %rd54;
mul.lo.s64 %rd56, %rd55, 36;
add.s64 %rd57, %rd6, %rd56;
add.s64 %rd58, %rd57, %rd30;
ld.global.nc.u32 %r288, [%rd58+4];
ld.global.nc.u32 %r292, [%rd58+20];
ld.global.nc.u32 %r296, [%rd58+8];
ld.global.nc.u32 %r300, [%rd58+24];

dp4a.s32.s32 %r286, %r329, %r288, %r352;


dp4a.s32.s32 %r290, %r333, %r292, %r286;


dp4a.s32.s32 %r294, %r337, %r296, %r290;


dp4a.s32.s32 %r298, %r341, %r300, %r294;

ld.global.nc.u32 %r304, [%rd57];

{mul.f16x2 %r302,%r93,%r304;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r302;
cvt.f32.f16 %f105, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r302;
cvt.f32.f16 %f106, high;}


cvt.rn.f32.s32 %f144, %r298;
div.approx.ftz.f32 %f145, %f106, %f114;
fma.rn.ftz.f32 %f146, %f105, %f144, %f145;
add.ftz.f32 %f327, %f146, %f327;

dp4a.s32.s32 %r307, %r350, %r288, %r352;


dp4a.s32.s32 %r311, %r354, %r292, %r307;


dp4a.s32.s32 %r315, %r358, %r296, %r311;


dp4a.s32.s32 %r319, %r362, %r300, %r315;


{mul.f16x2 %r323,%r114,%r304;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r323;
cvt.f32.f16 %f107, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r323;
cvt.f32.f16 %f108, high;}


cvt.rn.f32.s32 %f147, %r319;
div.approx.ftz.f32 %f148, %f108, %f114;
fma.rn.ftz.f32 %f149, %f107, %f147, %f148;
add.ftz.f32 %f326, %f149, %f326;
cvt.u64.u32 %rd59, %r637;
add.s64 %rd60, %rd1, %rd59;
mul.lo.s64 %rd61, %rd60, 36;
add.s64 %rd62, %rd6, %rd61;
add.s64 %rd63, %rd62, %rd30;
ld.global.nc.u32 %r330, [%rd63+4];
ld.global.nc.u32 %r334, [%rd63+20];
ld.global.nc.u32 %r338, [%rd63+8];
ld.global.nc.u32 %r342, [%rd63+24];

dp4a.s32.s32 %r328, %r329, %r330, %r352;


dp4a.s32.s32 %r332, %r333, %r334, %r328;


dp4a.s32.s32 %r336, %r337, %r338, %r332;


dp4a.s32.s32 %r340, %r341, %r342, %r336;

ld.global.nc.u32 %r346, [%rd62];

{mul.f16x2 %r344,%r93,%r346;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r344;
cvt.f32.f16 %f109, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r344;
cvt.f32.f16 %f110, high;}


cvt.rn.f32.s32 %f150, %r340;
div.approx.ftz.f32 %f151, %f110, %f114;
fma.rn.ftz.f32 %f152, %f109, %f150, %f151;
add.ftz.f32 %f325, %f152, %f325;

dp4a.s32.s32 %r349, %r350, %r330, %r352;


dp4a.s32.s32 %r353, %r354, %r334, %r349;


dp4a.s32.s32 %r357, %r358, %r338, %r353;


dp4a.s32.s32 %r361, %r362, %r342, %r357;


{mul.f16x2 %r365,%r114,%r346;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r365;
cvt.f32.f16 %f111, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r365;
cvt.f32.f16 %f112, high;}


cvt.rn.f32.s32 %f153, %r361;
div.approx.ftz.f32 %f154, %f112, %f114;
fma.rn.ftz.f32 %f155, %f111, %f153, %f154;
add.ftz.f32 %f324, %f155, %f324;
add.s32 %r641, %r641, 32;
add.s32 %r640, %r640, 32;
add.s32 %r639, %r639, 32;
add.s32 %r638, %r638, 32;
add.s32 %r637, %r637, 32;
add.s32 %r636, %r636, 32;
add.s64 %rd109, %rd109, 640;
add.s64 %rd108, %rd108, 640;
add.s32 %r642, %r642, 32;
setp.lt.s32 %p4, %r642, %r6;
@%p4 bra $L__BB54_2;

$L__BB54_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB54_5;

mov.u32 %r378, _ZZ13mul_mat_vec_qIL9ggml_type3ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r379, %r1, 1792, %r378;
shl.b32 %r380, %r2, 2;
add.s32 %r381, %r379, %r380;
st.shared.f32 [%r381+-1792], %f337;
st.shared.f32 [%r381+-1664], %f336;
st.shared.f32 [%r381+-1536], %f335;
st.shared.f32 [%r381+-1408], %f334;
st.shared.f32 [%r381+-1280], %f333;
st.shared.f32 [%r381+-1152], %f332;
st.shared.f32 [%r381+-1024], %f331;
st.shared.f32 [%r381+-896], %f330;
st.shared.f32 [%r381+-768], %f329;
st.shared.f32 [%r381+-640], %f328;
st.shared.f32 [%r381+-512], %f327;
st.shared.f32 [%r381+-384], %f326;
st.shared.f32 [%r381+-256], %f325;
st.shared.f32 [%r381+-128], %f324;

$L__BB54_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB54_34;

mov.u32 %r382, %ctaid.z;
mov.u32 %r383, %ctaid.y;
mad.lo.s32 %r384, %r383, %r38, %r4;
mad.lo.s32 %r385, %r382, %r44, %r384;
cvt.u64.u32 %rd11, %r385;
shl.b32 %r386, %r2, 2;
mov.u32 %r387, _ZZ13mul_mat_vec_qIL9ggml_type3ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r27, %r387, %r386;
mov.u32 %r388, 2;
ld.shared.f32 %f156, [%r27];
add.ftz.f32 %f157, %f156, %f337;
mov.b32 %r389, %f157;
mov.u32 %r390, 31;
mov.u32 %r391, 16;
mov.u32 %r392, -1;
shfl.sync.bfly.b32 %r393|%p7, %r389, %r391, %r390, %r392;
mov.b32 %f158, %r393;
add.ftz.f32 %f159, %f157, %f158;
mov.b32 %r394, %f159;
mov.u32 %r395, 8;
shfl.sync.bfly.b32 %r396|%p8, %r394, %r395, %r390, %r392;
mov.b32 %f160, %r396;
add.ftz.f32 %f161, %f159, %f160;
mov.b32 %r397, %f161;
mov.u32 %r398, 4;
shfl.sync.bfly.b32 %r399|%p9, %r397, %r398, %r390, %r392;
mov.b32 %f162, %r399;
add.ftz.f32 %f163, %f161, %f162;
mov.b32 %r400, %f163;
shfl.sync.bfly.b32 %r401|%p10, %r400, %r388, %r390, %r392;
mov.b32 %f164, %r401;
add.ftz.f32 %f165, %f163, %f164;
mov.b32 %r402, %f165;
mov.u32 %r403, 1;
shfl.sync.bfly.b32 %r404|%p11, %r402, %r403, %r390, %r392;
mov.b32 %f166, %r404;
add.ftz.f32 %f43, %f165, %f166;
setp.lt.u32 %p12, %r4, %r32;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
cvta.to.global.u64 %rd64, %rd18;
mul.wide.u32 %rd65, %r385, 4;
add.s64 %rd12, %rd64, %rd65;
not.pred %p14, %p1;
@%p14 bra $L__BB54_8;

st.global.f32 [%rd12], %f43;

$L__BB54_8:
ld.shared.f32 %f167, [%r27+128];
add.ftz.f32 %f168, %f167, %f336;
mov.b32 %r405, %f168;
shfl.sync.bfly.b32 %r409|%p15, %r405, %r391, %r390, %r392;
mov.b32 %f169, %r409;
add.ftz.f32 %f170, %f168, %f169;
mov.b32 %r410, %f170;
shfl.sync.bfly.b32 %r412|%p16, %r410, %r395, %r390, %r392;
mov.b32 %f171, %r412;
add.ftz.f32 %f172, %f170, %f171;
mov.b32 %r413, %f172;
shfl.sync.bfly.b32 %r415|%p17, %r413, %r398, %r390, %r392;
mov.b32 %f173, %r415;
add.ftz.f32 %f174, %f172, %f173;
mov.b32 %r416, %f174;
shfl.sync.bfly.b32 %r418|%p18, %r416, %r388, %r390, %r392;
mov.b32 %f175, %r418;
add.ftz.f32 %f176, %f174, %f175;
mov.b32 %r419, %f176;
shfl.sync.bfly.b32 %r421|%p19, %r419, %r403, %r390, %r392;
mov.b32 %f177, %r421;
add.ftz.f32 %f44, %f176, %f177;
setp.eq.s32 %p20, %r2, 1;
or.b32 %r425, %r4, 1;
setp.lt.u32 %p21, %r425, %r32;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB54_10;

st.global.f32 [%rd12+4], %f44;

$L__BB54_10:
ld.shared.f32 %f178, [%r27+256];
add.ftz.f32 %f179, %f178, %f335;
mov.b32 %r426, %f179;
mov.u32 %r427, 31;
mov.u32 %r428, 16;
mov.u32 %r429, -1;
shfl.sync.bfly.b32 %r430|%p23, %r426, %r428, %r427, %r429;
mov.b32 %f180, %r430;
add.ftz.f32 %f181, %f179, %f180;
mov.b32 %r431, %f181;
mov.u32 %r432, 8;
shfl.sync.bfly.b32 %r433|%p24, %r431, %r432, %r427, %r429;
mov.b32 %f182, %r433;
add.ftz.f32 %f183, %f181, %f182;
mov.b32 %r434, %f183;
mov.u32 %r435, 4;
shfl.sync.bfly.b32 %r436|%p25, %r434, %r435, %r427, %r429;
mov.b32 %f184, %r436;
add.ftz.f32 %f185, %f183, %f184;
mov.b32 %r437, %f185;
mov.u32 %r438, 2;
shfl.sync.bfly.b32 %r439|%p26, %r437, %r438, %r427, %r429;
mov.b32 %f186, %r439;
add.ftz.f32 %f187, %f185, %f186;
mov.b32 %r440, %f187;
mov.u32 %r441, 1;
shfl.sync.bfly.b32 %r442|%p27, %r440, %r441, %r427, %r429;
mov.b32 %f188, %r442;
add.ftz.f32 %f45, %f187, %f188;
@%p14 bra $L__BB54_12;

cvt.u64.u32 %rd66, %r32;
add.s64 %rd67, %rd66, %rd11;
shl.b64 %rd69, %rd67, 2;
add.s64 %rd70, %rd64, %rd69;
st.global.f32 [%rd70], %f45;

$L__BB54_12:
ld.shared.f32 %f189, [%r27+384];
add.ftz.f32 %f190, %f189, %f334;
mov.b32 %r443, %f190;
shfl.sync.bfly.b32 %r447|%p29, %r443, %r428, %r427, %r429;
mov.b32 %f191, %r447;
add.ftz.f32 %f192, %f190, %f191;
mov.b32 %r448, %f192;
shfl.sync.bfly.b32 %r450|%p30, %r448, %r432, %r427, %r429;
mov.b32 %f193, %r450;
add.ftz.f32 %f194, %f192, %f193;
mov.b32 %r451, %f194;
shfl.sync.bfly.b32 %r453|%p31, %r451, %r435, %r427, %r429;
mov.b32 %f195, %r453;
add.ftz.f32 %f196, %f194, %f195;
mov.b32 %r454, %f196;
shfl.sync.bfly.b32 %r456|%p32, %r454, %r438, %r427, %r429;
mov.b32 %f197, %r456;
add.ftz.f32 %f198, %f196, %f197;
mov.b32 %r457, %f198;
shfl.sync.bfly.b32 %r459|%p33, %r457, %r441, %r427, %r429;
mov.b32 %f199, %r459;
add.ftz.f32 %f46, %f198, %f199;
@%p22 bra $L__BB54_14;

add.s32 %r460, %r32, 1;
cvt.u64.u32 %rd71, %r460;
add.s64 %rd72, %rd71, %rd11;
shl.b64 %rd74, %rd72, 2;
add.s64 %rd75, %rd64, %rd74;
st.global.f32 [%rd75], %f46;

$L__BB54_14:
shl.b32 %r461, %r32, 1;
mov.u32 %r462, 1;
ld.shared.f32 %f200, [%r27+512];
add.ftz.f32 %f201, %f200, %f333;
mov.b32 %r463, %f201;
mov.u32 %r464, 31;
mov.u32 %r465, 16;
mov.u32 %r466, -1;
shfl.sync.bfly.b32 %r467|%p35, %r463, %r465, %r464, %r466;
mov.b32 %f202, %r467;
add.ftz.f32 %f203, %f201, %f202;
mov.b32 %r468, %f203;
mov.u32 %r469, 8;
shfl.sync.bfly.b32 %r470|%p36, %r468, %r469, %r464, %r466;
mov.b32 %f204, %r470;
add.ftz.f32 %f205, %f203, %f204;
mov.b32 %r471, %f205;
mov.u32 %r472, 4;
shfl.sync.bfly.b32 %r473|%p37, %r471, %r472, %r464, %r466;
mov.b32 %f206, %r473;
add.ftz.f32 %f207, %f205, %f206;
mov.b32 %r474, %f207;
mov.u32 %r475, 2;
shfl.sync.bfly.b32 %r476|%p38, %r474, %r475, %r464, %r466;
mov.b32 %f208, %r476;
add.ftz.f32 %f209, %f207, %f208;
mov.b32 %r477, %f209;
shfl.sync.bfly.b32 %r478|%p39, %r477, %r462, %r464, %r466;
mov.b32 %f210, %r478;
add.ftz.f32 %f47, %f209, %f210;
cvt.u64.u32 %rd76, %r461;
add.s64 %rd77, %rd76, %rd11;
shl.b64 %rd79, %rd77, 2;
add.s64 %rd13, %rd64, %rd79;
@%p14 bra $L__BB54_16;

st.global.f32 [%rd13], %f47;

$L__BB54_16:
ld.shared.f32 %f211, [%r27+640];
add.ftz.f32 %f212, %f211, %f332;
mov.b32 %r479, %f212;
shfl.sync.bfly.b32 %r483|%p41, %r479, %r465, %r464, %r466;
mov.b32 %f213, %r483;
add.ftz.f32 %f214, %f212, %f213;
mov.b32 %r484, %f214;
shfl.sync.bfly.b32 %r486|%p42, %r484, %r469, %r464, %r466;
mov.b32 %f215, %r486;
add.ftz.f32 %f216, %f214, %f215;
mov.b32 %r487, %f216;
shfl.sync.bfly.b32 %r489|%p43, %r487, %r472, %r464, %r466;
mov.b32 %f217, %r489;
add.ftz.f32 %f218, %f216, %f217;
mov.b32 %r490, %f218;
shfl.sync.bfly.b32 %r492|%p44, %r490, %r475, %r464, %r466;
mov.b32 %f219, %r492;
add.ftz.f32 %f220, %f218, %f219;
mov.b32 %r493, %f220;
shfl.sync.bfly.b32 %r495|%p45, %r493, %r462, %r464, %r466;
mov.b32 %f221, %r495;
add.ftz.f32 %f48, %f220, %f221;
@%p22 bra $L__BB54_18;

st.global.f32 [%rd13+4], %f48;

$L__BB54_18:
mul.lo.s32 %r28, %r32, 3;
ld.shared.f32 %f222, [%r27+768];
add.ftz.f32 %f223, %f222, %f331;
mov.b32 %r496, %f223;
mov.u32 %r497, 31;
mov.u32 %r498, 16;
mov.u32 %r499, -1;
shfl.sync.bfly.b32 %r500|%p47, %r496, %r498, %r497, %r499;
mov.b32 %f224, %r500;
add.ftz.f32 %f225, %f223, %f224;
mov.b32 %r501, %f225;
mov.u32 %r502, 8;
shfl.sync.bfly.b32 %r503|%p48, %r501, %r502, %r497, %r499;
mov.b32 %f226, %r503;
add.ftz.f32 %f227, %f225, %f226;
mov.b32 %r504, %f227;
mov.u32 %r505, 4;
shfl.sync.bfly.b32 %r506|%p49, %r504, %r505, %r497, %r499;
mov.b32 %f228, %r506;
add.ftz.f32 %f229, %f227, %f228;
mov.b32 %r507, %f229;
mov.u32 %r508, 2;
shfl.sync.bfly.b32 %r509|%p50, %r507, %r508, %r497, %r499;
mov.b32 %f230, %r509;
add.ftz.f32 %f231, %f229, %f230;
mov.b32 %r510, %f231;
mov.u32 %r511, 1;
shfl.sync.bfly.b32 %r512|%p51, %r510, %r511, %r497, %r499;
mov.b32 %f232, %r512;
add.ftz.f32 %f49, %f231, %f232;
@%p14 bra $L__BB54_20;

cvt.u64.u32 %rd80, %r28;
add.s64 %rd81, %rd80, %rd11;
shl.b64 %rd83, %rd81, 2;
add.s64 %rd84, %rd64, %rd83;
st.global.f32 [%rd84], %f49;

$L__BB54_20:
ld.shared.f32 %f233, [%r27+896];
add.ftz.f32 %f234, %f233, %f330;
mov.b32 %r513, %f234;
shfl.sync.bfly.b32 %r517|%p53, %r513, %r498, %r497, %r499;
mov.b32 %f235, %r517;
add.ftz.f32 %f236, %f234, %f235;
mov.b32 %r518, %f236;
shfl.sync.bfly.b32 %r520|%p54, %r518, %r502, %r497, %r499;
mov.b32 %f237, %r520;
add.ftz.f32 %f238, %f236, %f237;
mov.b32 %r521, %f238;
shfl.sync.bfly.b32 %r523|%p55, %r521, %r505, %r497, %r499;
mov.b32 %f239, %r523;
add.ftz.f32 %f240, %f238, %f239;
mov.b32 %r524, %f240;
shfl.sync.bfly.b32 %r526|%p56, %r524, %r508, %r497, %r499;
mov.b32 %f241, %r526;
add.ftz.f32 %f242, %f240, %f241;
mov.b32 %r527, %f242;
shfl.sync.bfly.b32 %r529|%p57, %r527, %r511, %r497, %r499;
mov.b32 %f243, %r529;
add.ftz.f32 %f50, %f242, %f243;
@%p22 bra $L__BB54_22;

add.s32 %r530, %r28, 1;
cvt.u64.u32 %rd85, %r530;
add.s64 %rd86, %rd85, %rd11;
shl.b64 %rd88, %rd86, 2;
add.s64 %rd89, %rd64, %rd88;
st.global.f32 [%rd89], %f50;

$L__BB54_22:
shl.b32 %r531, %r32, 2;
mov.u32 %r532, 2;
ld.shared.f32 %f244, [%r27+1024];
add.ftz.f32 %f245, %f244, %f329;
mov.b32 %r533, %f245;
mov.u32 %r534, 31;
mov.u32 %r535, 16;
mov.u32 %r536, -1;
shfl.sync.bfly.b32 %r537|%p59, %r533, %r535, %r534, %r536;
mov.b32 %f246, %r537;
add.ftz.f32 %f247, %f245, %f246;
mov.b32 %r538, %f247;
mov.u32 %r539, 8;
shfl.sync.bfly.b32 %r540|%p60, %r538, %r539, %r534, %r536;
mov.b32 %f248, %r540;
add.ftz.f32 %f249, %f247, %f248;
mov.b32 %r541, %f249;
mov.u32 %r542, 4;
shfl.sync.bfly.b32 %r543|%p61, %r541, %r542, %r534, %r536;
mov.b32 %f250, %r543;
add.ftz.f32 %f251, %f249, %f250;
mov.b32 %r544, %f251;
shfl.sync.bfly.b32 %r545|%p62, %r544, %r532, %r534, %r536;
mov.b32 %f252, %r545;
add.ftz.f32 %f253, %f251, %f252;
mov.b32 %r546, %f253;
mov.u32 %r547, 1;
shfl.sync.bfly.b32 %r548|%p63, %r546, %r547, %r534, %r536;
mov.b32 %f254, %r548;
add.ftz.f32 %f51, %f253, %f254;
cvt.u64.u32 %rd90, %r531;
add.s64 %rd91, %rd90, %rd11;
shl.b64 %rd93, %rd91, 2;
add.s64 %rd14, %rd64, %rd93;
@%p14 bra $L__BB54_24;

st.global.f32 [%rd14], %f51;

$L__BB54_24:
ld.shared.f32 %f255, [%r27+1152];
add.ftz.f32 %f256, %f255, %f328;
mov.b32 %r549, %f256;
shfl.sync.bfly.b32 %r553|%p65, %r549, %r535, %r534, %r536;
mov.b32 %f257, %r553;
add.ftz.f32 %f258, %f256, %f257;
mov.b32 %r554, %f258;
shfl.sync.bfly.b32 %r556|%p66, %r554, %r539, %r534, %r536;
mov.b32 %f259, %r556;
add.ftz.f32 %f260, %f258, %f259;
mov.b32 %r557, %f260;
shfl.sync.bfly.b32 %r559|%p67, %r557, %r542, %r534, %r536;
mov.b32 %f261, %r559;
add.ftz.f32 %f262, %f260, %f261;
mov.b32 %r560, %f262;
shfl.sync.bfly.b32 %r562|%p68, %r560, %r532, %r534, %r536;
mov.b32 %f263, %r562;
add.ftz.f32 %f264, %f262, %f263;
mov.b32 %r563, %f264;
shfl.sync.bfly.b32 %r565|%p69, %r563, %r547, %r534, %r536;
mov.b32 %f265, %r565;
add.ftz.f32 %f52, %f264, %f265;
@%p22 bra $L__BB54_26;

st.global.f32 [%rd14+4], %f52;

$L__BB54_26:
mul.lo.s32 %r29, %r32, 5;
ld.shared.f32 %f266, [%r27+1280];
add.ftz.f32 %f267, %f266, %f327;
mov.b32 %r566, %f267;
mov.u32 %r567, 31;
mov.u32 %r568, 16;
mov.u32 %r569, -1;
shfl.sync.bfly.b32 %r570|%p71, %r566, %r568, %r567, %r569;
mov.b32 %f268, %r570;
add.ftz.f32 %f269, %f267, %f268;
mov.b32 %r571, %f269;
mov.u32 %r572, 8;
shfl.sync.bfly.b32 %r573|%p72, %r571, %r572, %r567, %r569;
mov.b32 %f270, %r573;
add.ftz.f32 %f271, %f269, %f270;
mov.b32 %r574, %f271;
mov.u32 %r575, 4;
shfl.sync.bfly.b32 %r576|%p73, %r574, %r575, %r567, %r569;
mov.b32 %f272, %r576;
add.ftz.f32 %f273, %f271, %f272;
mov.b32 %r577, %f273;
mov.u32 %r578, 2;
shfl.sync.bfly.b32 %r579|%p74, %r577, %r578, %r567, %r569;
mov.b32 %f274, %r579;
add.ftz.f32 %f275, %f273, %f274;
mov.b32 %r580, %f275;
mov.u32 %r581, 1;
shfl.sync.bfly.b32 %r582|%p75, %r580, %r581, %r567, %r569;
mov.b32 %f276, %r582;
add.ftz.f32 %f53, %f275, %f276;
@%p14 bra $L__BB54_28;

cvt.u64.u32 %rd94, %r29;
add.s64 %rd95, %rd94, %rd11;
shl.b64 %rd97, %rd95, 2;
add.s64 %rd98, %rd64, %rd97;
st.global.f32 [%rd98], %f53;

$L__BB54_28:
ld.shared.f32 %f277, [%r27+1408];
add.ftz.f32 %f278, %f277, %f326;
mov.b32 %r583, %f278;
shfl.sync.bfly.b32 %r587|%p77, %r583, %r568, %r567, %r569;
mov.b32 %f279, %r587;
add.ftz.f32 %f280, %f278, %f279;
mov.b32 %r588, %f280;
shfl.sync.bfly.b32 %r590|%p78, %r588, %r572, %r567, %r569;
mov.b32 %f281, %r590;
add.ftz.f32 %f282, %f280, %f281;
mov.b32 %r591, %f282;
shfl.sync.bfly.b32 %r593|%p79, %r591, %r575, %r567, %r569;
mov.b32 %f283, %r593;
add.ftz.f32 %f284, %f282, %f283;
mov.b32 %r594, %f284;
shfl.sync.bfly.b32 %r596|%p80, %r594, %r578, %r567, %r569;
mov.b32 %f285, %r596;
add.ftz.f32 %f286, %f284, %f285;
mov.b32 %r597, %f286;
shfl.sync.bfly.b32 %r599|%p81, %r597, %r581, %r567, %r569;
mov.b32 %f287, %r599;
add.ftz.f32 %f54, %f286, %f287;
@%p22 bra $L__BB54_30;

add.s32 %r600, %r29, 1;
cvt.u64.u32 %rd99, %r600;
add.s64 %rd100, %rd99, %rd11;
shl.b64 %rd102, %rd100, 2;
add.s64 %rd103, %rd64, %rd102;
st.global.f32 [%rd103], %f54;

$L__BB54_30:
mul.lo.s32 %r601, %r32, 6;
ld.shared.f32 %f288, [%r27+1536];
add.ftz.f32 %f289, %f288, %f325;
mov.b32 %r602, %f289;
mov.u32 %r603, 31;
mov.u32 %r604, 16;
mov.u32 %r605, -1;
shfl.sync.bfly.b32 %r606|%p83, %r602, %r604, %r603, %r605;
mov.b32 %f290, %r606;
add.ftz.f32 %f291, %f289, %f290;
mov.b32 %r607, %f291;
mov.u32 %r608, 8;
shfl.sync.bfly.b32 %r609|%p84, %r607, %r608, %r603, %r605;
mov.b32 %f292, %r609;
add.ftz.f32 %f293, %f291, %f292;
mov.b32 %r610, %f293;
mov.u32 %r611, 4;
shfl.sync.bfly.b32 %r612|%p85, %r610, %r611, %r603, %r605;
mov.b32 %f294, %r612;
add.ftz.f32 %f295, %f293, %f294;
mov.b32 %r613, %f295;
mov.u32 %r614, 2;
shfl.sync.bfly.b32 %r615|%p86, %r613, %r614, %r603, %r605;
mov.b32 %f296, %r615;
add.ftz.f32 %f297, %f295, %f296;
mov.b32 %r616, %f297;
mov.u32 %r617, 1;
shfl.sync.bfly.b32 %r618|%p87, %r616, %r617, %r603, %r605;
mov.b32 %f298, %r618;
add.ftz.f32 %f55, %f297, %f298;
cvt.u64.u32 %rd104, %r601;
add.s64 %rd105, %rd104, %rd11;
shl.b64 %rd107, %rd105, 2;
add.s64 %rd15, %rd64, %rd107;
@%p14 bra $L__BB54_32;

st.global.f32 [%rd15], %f55;

$L__BB54_32:
ld.shared.f32 %f299, [%r27+1664];
add.ftz.f32 %f300, %f299, %f324;
mov.b32 %r619, %f300;
shfl.sync.bfly.b32 %r623|%p89, %r619, %r604, %r603, %r605;
mov.b32 %f301, %r623;
add.ftz.f32 %f302, %f300, %f301;
mov.b32 %r624, %f302;
shfl.sync.bfly.b32 %r626|%p90, %r624, %r608, %r603, %r605;
mov.b32 %f303, %r626;
add.ftz.f32 %f304, %f302, %f303;
mov.b32 %r627, %f304;
shfl.sync.bfly.b32 %r629|%p91, %r627, %r611, %r603, %r605;
mov.b32 %f305, %r629;
add.ftz.f32 %f306, %f304, %f305;
mov.b32 %r630, %f306;
shfl.sync.bfly.b32 %r632|%p92, %r630, %r614, %r603, %r605;
mov.b32 %f307, %r632;
add.ftz.f32 %f308, %f306, %f307;
mov.b32 %r633, %f308;
shfl.sync.bfly.b32 %r635|%p93, %r633, %r617, %r603, %r605;
mov.b32 %f309, %r635;
add.ftz.f32 %f56, %f308, %f309;
@%p22 bra $L__BB54_34;

st.global.f32 [%rd15+4], %f56;

$L__BB54_34:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type3ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type3ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type3ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type3ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type3ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type3ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type3ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type3ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type3ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type3ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<107>;
.reg .f32 %f<386>;
.reg .b32 %r<726>;
.reg .b64 %rd<125>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type3ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[2048];

ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type3ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type3ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type3ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r49, [_Z13mul_mat_vec_qIL9ggml_type3ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type3ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type3ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type3ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type3ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type3ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type3ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r46, [_Z13mul_mat_vec_qIL9ggml_type3ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r47, [_Z13mul_mat_vec_qIL9ggml_type3ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r48, [_Z13mul_mat_vec_qIL9ggml_type3ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type3ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type3ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type3ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type3ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r50, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r50, %r2;
mov.u32 %r51, %ctaid.x;
shl.b32 %r4, %r51, 1;
shr.u32 %r52, %r3, 31;
add.s32 %r53, %r3, %r52;
shr.s32 %r725, %r53, 1;
shr.u32 %r6, %r49, 5;
setp.ge.s32 %p3, %r725, %r6;
mov.f32 %f370, 0f00000000;
mov.f32 %f371, %f370;
mov.f32 %f372, %f370;
mov.f32 %f373, %f370;
mov.f32 %f374, %f370;
mov.f32 %f375, %f370;
mov.f32 %f376, %f370;
mov.f32 %f377, %f370;
mov.f32 %f378, %f370;
mov.f32 %f379, %f370;
mov.f32 %f380, %f370;
mov.f32 %f381, %f370;
mov.f32 %f382, %f370;
mov.f32 %f383, %f370;
mov.f32 %f384, %f370;
mov.f32 %f385, %f370;
@%p3 bra $L__BB55_3;

mov.u32 %r54, %ctaid.y;
mov.u32 %r55, %ctaid.z;
mul.lo.s32 %r56, %r54, %r41;
cvt.u64.u32 %rd19, %r56;
mul.lo.s32 %r57, %r55, %r47;
cvt.u64.u32 %rd20, %r57;
add.s64 %rd1, %rd20, %rd19;
mul.hi.u32 %r58, %r55, %r43;
add.s32 %r59, %r58, %r55;
shr.u32 %r60, %r59, %r44;
mul.lo.s32 %r61, %r60, %r46;
mul.hi.u32 %r62, %r54, %r37;
add.s32 %r63, %r62, %r54;
shr.u32 %r64, %r63, %r38;
mul.lo.s32 %r65, %r64, %r40;
cvta.to.global.u64 %rd21, %rd16;
and.b32 %r68, %r53, 2147483646;
sub.s32 %r69, %r3, %r68;
shl.b32 %r70, %r69, 1;
cvt.s64.s32 %rd2, %r70;
shl.b32 %r71, %r35, 1;
add.s32 %r724, %r725, %r71;
mad.lo.s32 %r723, %r35, 3, %r725;
shl.b32 %r72, %r35, 2;
add.s32 %r722, %r725, %r72;
mad.lo.s32 %r721, %r35, 5, %r725;
mad.lo.s32 %r720, %r35, 6, %r725;
mad.lo.s32 %r719, %r35, 7, %r725;
add.s32 %r718, %r725, %r35;
add.s32 %r73, %r725, %r61;
add.s32 %r74, %r73, %r65;
add.s32 %r75, %r4, 1;
mad.lo.s32 %r76, %r34, %r75, %r74;
mul.wide.s32 %rd22, %r76, 20;
add.s64 %rd23, %rd21, %rd22;
add.s64 %rd124, %rd23, 8;
mul.wide.s32 %rd4, %r70, 4;
mad.lo.s32 %r77, %r4, %r34, %r725;
add.s32 %r78, %r77, %r61;
add.s32 %r79, %r78, %r65;
mul.wide.s32 %rd24, %r79, 20;
add.s64 %rd25, %rd21, %rd24;
add.s64 %rd123, %rd25, 8;
cvta.to.global.u64 %rd6, %rd17;
mov.f32 %f370, 0f00000000;
shl.b64 %rd30, %rd2, 2;

$L__BB55_2:
cvt.u64.u32 %rd26, %r725;
add.s64 %rd27, %rd1, %rd26;
mul.lo.s64 %rd28, %rd27, 36;
add.s64 %rd29, %rd6, %rd28;
add.s64 %rd31, %rd29, %rd30;
ld.global.nc.u32 %r82, [%rd31+4];
ld.global.nc.u32 %r86, [%rd31+20];
add.s64 %rd32, %rd123, %rd4;
ld.global.nc.u32 %r416, [%rd32];
ld.global.nc.u32 %r90, [%rd31+8];
ld.global.nc.u32 %r94, [%rd31+24];
ld.global.nc.u32 %r417, [%rd32+-4];
and.b32 %r375, %r417, 252645135;
shr.u32 %r418, %r417, 4;
and.b32 %r379, %r418, 252645135;
mov.u32 %r398, 0;

dp4a.s32.s32 %r80, %r375, %r82, %r398;


dp4a.s32.s32 %r84, %r379, %r86, %r80;

and.b32 %r383, %r416, 252645135;
shr.u32 %r419, %r416, 4;
and.b32 %r387, %r419, 252645135;

dp4a.s32.s32 %r88, %r383, %r90, %r84;


dp4a.s32.s32 %r92, %r387, %r94, %r88;

ld.global.nc.u32 %r97, [%rd123+-8];
ld.global.nc.u32 %r98, [%rd29];

{mul.f16x2 %r96,%r97,%r98;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r96;
cvt.f32.f16 %f97, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r96;
cvt.f32.f16 %f98, high;}


cvt.rn.f32.s32 %f129, %r92;
mov.f32 %f130, 0f40000000;
div.approx.ftz.f32 %f131, %f98, %f130;
fma.rn.ftz.f32 %f132, %f97, %f129, %f131;
add.ftz.f32 %f385, %f132, %f385;
add.s64 %rd33, %rd124, %rd4;
ld.global.nc.u32 %r420, [%rd33];
ld.global.nc.u32 %r421, [%rd33+-4];
and.b32 %r396, %r421, 252645135;
shr.u32 %r422, %r421, 4;
and.b32 %r400, %r422, 252645135;

dp4a.s32.s32 %r101, %r396, %r82, %r398;


dp4a.s32.s32 %r105, %r400, %r86, %r101;

and.b32 %r404, %r420, 252645135;
shr.u32 %r423, %r420, 4;
and.b32 %r408, %r423, 252645135;

dp4a.s32.s32 %r109, %r404, %r90, %r105;


dp4a.s32.s32 %r113, %r408, %r94, %r109;

ld.global.nc.u32 %r118, [%rd124+-8];

{mul.f16x2 %r117,%r118,%r98;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r117;
cvt.f32.f16 %f99, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r117;
cvt.f32.f16 %f100, high;}


cvt.rn.f32.s32 %f133, %r113;
div.approx.ftz.f32 %f134, %f100, %f130;
fma.rn.ftz.f32 %f135, %f99, %f133, %f134;
add.ftz.f32 %f384, %f135, %f384;
cvt.u64.u32 %rd34, %r718;
add.s64 %rd35, %rd1, %rd34;
mul.lo.s64 %rd36, %rd35, 36;
add.s64 %rd37, %rd6, %rd36;
add.s64 %rd38, %rd37, %rd30;
ld.global.nc.u32 %r124, [%rd38+4];
ld.global.nc.u32 %r128, [%rd38+20];
ld.global.nc.u32 %r132, [%rd38+8];
ld.global.nc.u32 %r136, [%rd38+24];

dp4a.s32.s32 %r122, %r375, %r124, %r398;


dp4a.s32.s32 %r126, %r379, %r128, %r122;


dp4a.s32.s32 %r130, %r383, %r132, %r126;


dp4a.s32.s32 %r134, %r387, %r136, %r130;

ld.global.nc.u32 %r140, [%rd37];

{mul.f16x2 %r138,%r97,%r140;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r138;
cvt.f32.f16 %f101, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r138;
cvt.f32.f16 %f102, high;}


cvt.rn.f32.s32 %f136, %r134;
div.approx.ftz.f32 %f137, %f102, %f130;
fma.rn.ftz.f32 %f138, %f101, %f136, %f137;
add.ftz.f32 %f383, %f138, %f383;

dp4a.s32.s32 %r143, %r396, %r124, %r398;


dp4a.s32.s32 %r147, %r400, %r128, %r143;


dp4a.s32.s32 %r151, %r404, %r132, %r147;


dp4a.s32.s32 %r155, %r408, %r136, %r151;


{mul.f16x2 %r159,%r118,%r140;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r159;
cvt.f32.f16 %f103, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r159;
cvt.f32.f16 %f104, high;}


cvt.rn.f32.s32 %f139, %r155;
div.approx.ftz.f32 %f140, %f104, %f130;
fma.rn.ftz.f32 %f141, %f103, %f139, %f140;
add.ftz.f32 %f382, %f141, %f382;
cvt.u64.u32 %rd39, %r724;
add.s64 %rd40, %rd1, %rd39;
mul.lo.s64 %rd41, %rd40, 36;
add.s64 %rd42, %rd6, %rd41;
add.s64 %rd43, %rd42, %rd30;
ld.global.nc.u32 %r166, [%rd43+4];
ld.global.nc.u32 %r170, [%rd43+20];
ld.global.nc.u32 %r174, [%rd43+8];
ld.global.nc.u32 %r178, [%rd43+24];

dp4a.s32.s32 %r164, %r375, %r166, %r398;


dp4a.s32.s32 %r168, %r379, %r170, %r164;


dp4a.s32.s32 %r172, %r383, %r174, %r168;


dp4a.s32.s32 %r176, %r387, %r178, %r172;

ld.global.nc.u32 %r182, [%rd42];

{mul.f16x2 %r180,%r97,%r182;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r180;
cvt.f32.f16 %f105, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r180;
cvt.f32.f16 %f106, high;}


cvt.rn.f32.s32 %f142, %r176;
div.approx.ftz.f32 %f143, %f106, %f130;
fma.rn.ftz.f32 %f144, %f105, %f142, %f143;
add.ftz.f32 %f381, %f144, %f381;

dp4a.s32.s32 %r185, %r396, %r166, %r398;


dp4a.s32.s32 %r189, %r400, %r170, %r185;


dp4a.s32.s32 %r193, %r404, %r174, %r189;


dp4a.s32.s32 %r197, %r408, %r178, %r193;


{mul.f16x2 %r201,%r118,%r182;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r201;
cvt.f32.f16 %f107, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r201;
cvt.f32.f16 %f108, high;}


cvt.rn.f32.s32 %f145, %r197;
div.approx.ftz.f32 %f146, %f108, %f130;
fma.rn.ftz.f32 %f147, %f107, %f145, %f146;
add.ftz.f32 %f380, %f147, %f380;
cvt.u64.u32 %rd44, %r723;
add.s64 %rd45, %rd1, %rd44;
mul.lo.s64 %rd46, %rd45, 36;
add.s64 %rd47, %rd6, %rd46;
add.s64 %rd48, %rd47, %rd30;
ld.global.nc.u32 %r208, [%rd48+4];
ld.global.nc.u32 %r212, [%rd48+20];
ld.global.nc.u32 %r216, [%rd48+8];
ld.global.nc.u32 %r220, [%rd48+24];

dp4a.s32.s32 %r206, %r375, %r208, %r398;


dp4a.s32.s32 %r210, %r379, %r212, %r206;


dp4a.s32.s32 %r214, %r383, %r216, %r210;


dp4a.s32.s32 %r218, %r387, %r220, %r214;

ld.global.nc.u32 %r224, [%rd47];

{mul.f16x2 %r222,%r97,%r224;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r222;
cvt.f32.f16 %f109, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r222;
cvt.f32.f16 %f110, high;}


cvt.rn.f32.s32 %f148, %r218;
div.approx.ftz.f32 %f149, %f110, %f130;
fma.rn.ftz.f32 %f150, %f109, %f148, %f149;
add.ftz.f32 %f379, %f150, %f379;

dp4a.s32.s32 %r227, %r396, %r208, %r398;


dp4a.s32.s32 %r231, %r400, %r212, %r227;


dp4a.s32.s32 %r235, %r404, %r216, %r231;


dp4a.s32.s32 %r239, %r408, %r220, %r235;


{mul.f16x2 %r243,%r118,%r224;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r243;
cvt.f32.f16 %f111, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r243;
cvt.f32.f16 %f112, high;}


cvt.rn.f32.s32 %f151, %r239;
div.approx.ftz.f32 %f152, %f112, %f130;
fma.rn.ftz.f32 %f153, %f111, %f151, %f152;
add.ftz.f32 %f378, %f153, %f378;
cvt.u64.u32 %rd49, %r722;
add.s64 %rd50, %rd1, %rd49;
mul.lo.s64 %rd51, %rd50, 36;
add.s64 %rd52, %rd6, %rd51;
add.s64 %rd53, %rd52, %rd30;
ld.global.nc.u32 %r250, [%rd53+4];
ld.global.nc.u32 %r254, [%rd53+20];
ld.global.nc.u32 %r258, [%rd53+8];
ld.global.nc.u32 %r262, [%rd53+24];

dp4a.s32.s32 %r248, %r375, %r250, %r398;


dp4a.s32.s32 %r252, %r379, %r254, %r248;


dp4a.s32.s32 %r256, %r383, %r258, %r252;


dp4a.s32.s32 %r260, %r387, %r262, %r256;

ld.global.nc.u32 %r266, [%rd52];

{mul.f16x2 %r264,%r97,%r266;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r264;
cvt.f32.f16 %f113, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r264;
cvt.f32.f16 %f114, high;}


cvt.rn.f32.s32 %f154, %r260;
div.approx.ftz.f32 %f155, %f114, %f130;
fma.rn.ftz.f32 %f156, %f113, %f154, %f155;
add.ftz.f32 %f377, %f156, %f377;

dp4a.s32.s32 %r269, %r396, %r250, %r398;


dp4a.s32.s32 %r273, %r400, %r254, %r269;


dp4a.s32.s32 %r277, %r404, %r258, %r273;


dp4a.s32.s32 %r281, %r408, %r262, %r277;


{mul.f16x2 %r285,%r118,%r266;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r285;
cvt.f32.f16 %f115, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r285;
cvt.f32.f16 %f116, high;}


cvt.rn.f32.s32 %f157, %r281;
div.approx.ftz.f32 %f158, %f116, %f130;
fma.rn.ftz.f32 %f159, %f115, %f157, %f158;
add.ftz.f32 %f376, %f159, %f376;
cvt.u64.u32 %rd54, %r721;
add.s64 %rd55, %rd1, %rd54;
mul.lo.s64 %rd56, %rd55, 36;
add.s64 %rd57, %rd6, %rd56;
add.s64 %rd58, %rd57, %rd30;
ld.global.nc.u32 %r292, [%rd58+4];
ld.global.nc.u32 %r296, [%rd58+20];
ld.global.nc.u32 %r300, [%rd58+8];
ld.global.nc.u32 %r304, [%rd58+24];

dp4a.s32.s32 %r290, %r375, %r292, %r398;


dp4a.s32.s32 %r294, %r379, %r296, %r290;


dp4a.s32.s32 %r298, %r383, %r300, %r294;


dp4a.s32.s32 %r302, %r387, %r304, %r298;

ld.global.nc.u32 %r308, [%rd57];

{mul.f16x2 %r306,%r97,%r308;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r306;
cvt.f32.f16 %f117, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r306;
cvt.f32.f16 %f118, high;}


cvt.rn.f32.s32 %f160, %r302;
div.approx.ftz.f32 %f161, %f118, %f130;
fma.rn.ftz.f32 %f162, %f117, %f160, %f161;
add.ftz.f32 %f375, %f162, %f375;

dp4a.s32.s32 %r311, %r396, %r292, %r398;


dp4a.s32.s32 %r315, %r400, %r296, %r311;


dp4a.s32.s32 %r319, %r404, %r300, %r315;


dp4a.s32.s32 %r323, %r408, %r304, %r319;


{mul.f16x2 %r327,%r118,%r308;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r327;
cvt.f32.f16 %f119, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r327;
cvt.f32.f16 %f120, high;}


cvt.rn.f32.s32 %f163, %r323;
div.approx.ftz.f32 %f164, %f120, %f130;
fma.rn.ftz.f32 %f165, %f119, %f163, %f164;
add.ftz.f32 %f374, %f165, %f374;
cvt.u64.u32 %rd59, %r720;
add.s64 %rd60, %rd1, %rd59;
mul.lo.s64 %rd61, %rd60, 36;
add.s64 %rd62, %rd6, %rd61;
add.s64 %rd63, %rd62, %rd30;
ld.global.nc.u32 %r334, [%rd63+4];
ld.global.nc.u32 %r338, [%rd63+20];
ld.global.nc.u32 %r342, [%rd63+8];
ld.global.nc.u32 %r346, [%rd63+24];

dp4a.s32.s32 %r332, %r375, %r334, %r398;


dp4a.s32.s32 %r336, %r379, %r338, %r332;


dp4a.s32.s32 %r340, %r383, %r342, %r336;


dp4a.s32.s32 %r344, %r387, %r346, %r340;

ld.global.nc.u32 %r350, [%rd62];

{mul.f16x2 %r348,%r97,%r350;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r348;
cvt.f32.f16 %f121, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r348;
cvt.f32.f16 %f122, high;}


cvt.rn.f32.s32 %f166, %r344;
div.approx.ftz.f32 %f167, %f122, %f130;
fma.rn.ftz.f32 %f168, %f121, %f166, %f167;
add.ftz.f32 %f373, %f168, %f373;

dp4a.s32.s32 %r353, %r396, %r334, %r398;


dp4a.s32.s32 %r357, %r400, %r338, %r353;


dp4a.s32.s32 %r361, %r404, %r342, %r357;


dp4a.s32.s32 %r365, %r408, %r346, %r361;


{mul.f16x2 %r369,%r118,%r350;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r369;
cvt.f32.f16 %f123, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r369;
cvt.f32.f16 %f124, high;}


cvt.rn.f32.s32 %f169, %r365;
div.approx.ftz.f32 %f170, %f124, %f130;
fma.rn.ftz.f32 %f171, %f123, %f169, %f170;
add.ftz.f32 %f372, %f171, %f372;
cvt.u64.u32 %rd64, %r719;
add.s64 %rd65, %rd1, %rd64;
mul.lo.s64 %rd66, %rd65, 36;
add.s64 %rd67, %rd6, %rd66;
add.s64 %rd68, %rd67, %rd30;
ld.global.nc.u32 %r376, [%rd68+4];
ld.global.nc.u32 %r380, [%rd68+20];
ld.global.nc.u32 %r384, [%rd68+8];
ld.global.nc.u32 %r388, [%rd68+24];

dp4a.s32.s32 %r374, %r375, %r376, %r398;


dp4a.s32.s32 %r378, %r379, %r380, %r374;


dp4a.s32.s32 %r382, %r383, %r384, %r378;


dp4a.s32.s32 %r386, %r387, %r388, %r382;

ld.global.nc.u32 %r392, [%rd67];

{mul.f16x2 %r390,%r97,%r392;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r390;
cvt.f32.f16 %f125, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r390;
cvt.f32.f16 %f126, high;}


cvt.rn.f32.s32 %f172, %r386;
div.approx.ftz.f32 %f173, %f126, %f130;
fma.rn.ftz.f32 %f174, %f125, %f172, %f173;
add.ftz.f32 %f371, %f174, %f371;

dp4a.s32.s32 %r395, %r396, %r376, %r398;


dp4a.s32.s32 %r399, %r400, %r380, %r395;


dp4a.s32.s32 %r403, %r404, %r384, %r399;


dp4a.s32.s32 %r407, %r408, %r388, %r403;


{mul.f16x2 %r411,%r118,%r392;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r411;
cvt.f32.f16 %f127, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r411;
cvt.f32.f16 %f128, high;}


cvt.rn.f32.s32 %f175, %r407;
div.approx.ftz.f32 %f176, %f128, %f130;
fma.rn.ftz.f32 %f177, %f127, %f175, %f176;
add.ftz.f32 %f370, %f177, %f370;
add.s32 %r724, %r724, 32;
add.s32 %r723, %r723, 32;
add.s32 %r722, %r722, 32;
add.s32 %r721, %r721, 32;
add.s32 %r720, %r720, 32;
add.s32 %r719, %r719, 32;
add.s32 %r718, %r718, 32;
add.s64 %rd124, %rd124, 640;
add.s64 %rd123, %rd123, 640;
add.s32 %r725, %r725, 32;
setp.lt.s32 %p4, %r725, %r6;
@%p4 bra $L__BB55_2;

$L__BB55_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB55_5;

shl.b32 %r424, %r1, 11;
mov.u32 %r425, _ZZ13mul_mat_vec_qIL9ggml_type3ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r426, %r425, %r424;
shl.b32 %r427, %r2, 2;
add.s32 %r428, %r426, %r427;
st.shared.f32 [%r428+-2048], %f385;
st.shared.f32 [%r428+-1920], %f384;
st.shared.f32 [%r428+-1792], %f383;
st.shared.f32 [%r428+-1664], %f382;
st.shared.f32 [%r428+-1536], %f381;
st.shared.f32 [%r428+-1408], %f380;
st.shared.f32 [%r428+-1280], %f379;
st.shared.f32 [%r428+-1152], %f378;
st.shared.f32 [%r428+-1024], %f377;
st.shared.f32 [%r428+-896], %f376;
st.shared.f32 [%r428+-768], %f375;
st.shared.f32 [%r428+-640], %f374;
st.shared.f32 [%r428+-512], %f373;
st.shared.f32 [%r428+-384], %f372;
st.shared.f32 [%r428+-256], %f371;
st.shared.f32 [%r428+-128], %f370;

$L__BB55_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB55_38;

mov.u32 %r429, %ctaid.z;
mov.u32 %r430, %ctaid.y;
mad.lo.s32 %r431, %r430, %r42, %r4;
mad.lo.s32 %r432, %r429, %r48, %r431;
cvt.u64.u32 %rd11, %r432;
shl.b32 %r433, %r2, 2;
mov.u32 %r434, _ZZ13mul_mat_vec_qIL9ggml_type3ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r30, %r434, %r433;
mov.u32 %r435, 2;
ld.shared.f32 %f178, [%r30];
add.ftz.f32 %f179, %f178, %f385;
mov.b32 %r436, %f179;
mov.u32 %r437, 31;
mov.u32 %r438, 16;
mov.u32 %r439, -1;
shfl.sync.bfly.b32 %r440|%p7, %r436, %r438, %r437, %r439;
mov.b32 %f180, %r440;
add.ftz.f32 %f181, %f179, %f180;
mov.b32 %r441, %f181;
mov.u32 %r442, 8;
shfl.sync.bfly.b32 %r443|%p8, %r441, %r442, %r437, %r439;
mov.b32 %f182, %r443;
add.ftz.f32 %f183, %f181, %f182;
mov.b32 %r444, %f183;
mov.u32 %r445, 4;
shfl.sync.bfly.b32 %r446|%p9, %r444, %r445, %r437, %r439;
mov.b32 %f184, %r446;
add.ftz.f32 %f185, %f183, %f184;
mov.b32 %r447, %f185;
shfl.sync.bfly.b32 %r448|%p10, %r447, %r435, %r437, %r439;
mov.b32 %f186, %r448;
add.ftz.f32 %f187, %f185, %f186;
mov.b32 %r449, %f187;
mov.u32 %r450, 1;
shfl.sync.bfly.b32 %r451|%p11, %r449, %r450, %r437, %r439;
mov.b32 %f188, %r451;
add.ftz.f32 %f49, %f187, %f188;
setp.lt.u32 %p12, %r4, %r36;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
cvta.to.global.u64 %rd69, %rd18;
mul.wide.u32 %rd70, %r432, 4;
add.s64 %rd12, %rd69, %rd70;
not.pred %p14, %p1;
@%p14 bra $L__BB55_8;

st.global.f32 [%rd12], %f49;

$L__BB55_8:
ld.shared.f32 %f189, [%r30+128];
add.ftz.f32 %f190, %f189, %f384;
mov.b32 %r452, %f190;
shfl.sync.bfly.b32 %r456|%p15, %r452, %r438, %r437, %r439;
mov.b32 %f191, %r456;
add.ftz.f32 %f192, %f190, %f191;
mov.b32 %r457, %f192;
shfl.sync.bfly.b32 %r459|%p16, %r457, %r442, %r437, %r439;
mov.b32 %f193, %r459;
add.ftz.f32 %f194, %f192, %f193;
mov.b32 %r460, %f194;
shfl.sync.bfly.b32 %r462|%p17, %r460, %r445, %r437, %r439;
mov.b32 %f195, %r462;
add.ftz.f32 %f196, %f194, %f195;
mov.b32 %r463, %f196;
shfl.sync.bfly.b32 %r465|%p18, %r463, %r435, %r437, %r439;
mov.b32 %f197, %r465;
add.ftz.f32 %f198, %f196, %f197;
mov.b32 %r466, %f198;
shfl.sync.bfly.b32 %r468|%p19, %r466, %r450, %r437, %r439;
mov.b32 %f199, %r468;
add.ftz.f32 %f50, %f198, %f199;
setp.eq.s32 %p20, %r2, 1;
or.b32 %r472, %r4, 1;
setp.lt.u32 %p21, %r472, %r36;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB55_10;

st.global.f32 [%rd12+4], %f50;

$L__BB55_10:
ld.shared.f32 %f200, [%r30+256];
add.ftz.f32 %f201, %f200, %f383;
mov.b32 %r473, %f201;
mov.u32 %r474, 31;
mov.u32 %r475, 16;
mov.u32 %r476, -1;
shfl.sync.bfly.b32 %r477|%p23, %r473, %r475, %r474, %r476;
mov.b32 %f202, %r477;
add.ftz.f32 %f203, %f201, %f202;
mov.b32 %r478, %f203;
mov.u32 %r479, 8;
shfl.sync.bfly.b32 %r480|%p24, %r478, %r479, %r474, %r476;
mov.b32 %f204, %r480;
add.ftz.f32 %f205, %f203, %f204;
mov.b32 %r481, %f205;
mov.u32 %r482, 4;
shfl.sync.bfly.b32 %r483|%p25, %r481, %r482, %r474, %r476;
mov.b32 %f206, %r483;
add.ftz.f32 %f207, %f205, %f206;
mov.b32 %r484, %f207;
mov.u32 %r485, 2;
shfl.sync.bfly.b32 %r486|%p26, %r484, %r485, %r474, %r476;
mov.b32 %f208, %r486;
add.ftz.f32 %f209, %f207, %f208;
mov.b32 %r487, %f209;
mov.u32 %r488, 1;
shfl.sync.bfly.b32 %r489|%p27, %r487, %r488, %r474, %r476;
mov.b32 %f210, %r489;
add.ftz.f32 %f51, %f209, %f210;
@%p14 bra $L__BB55_12;

cvt.u64.u32 %rd71, %r36;
add.s64 %rd72, %rd71, %rd11;
shl.b64 %rd74, %rd72, 2;
add.s64 %rd75, %rd69, %rd74;
st.global.f32 [%rd75], %f51;

$L__BB55_12:
ld.shared.f32 %f211, [%r30+384];
add.ftz.f32 %f212, %f211, %f382;
mov.b32 %r490, %f212;
shfl.sync.bfly.b32 %r494|%p29, %r490, %r475, %r474, %r476;
mov.b32 %f213, %r494;
add.ftz.f32 %f214, %f212, %f213;
mov.b32 %r495, %f214;
shfl.sync.bfly.b32 %r497|%p30, %r495, %r479, %r474, %r476;
mov.b32 %f215, %r497;
add.ftz.f32 %f216, %f214, %f215;
mov.b32 %r498, %f216;
shfl.sync.bfly.b32 %r500|%p31, %r498, %r482, %r474, %r476;
mov.b32 %f217, %r500;
add.ftz.f32 %f218, %f216, %f217;
mov.b32 %r501, %f218;
shfl.sync.bfly.b32 %r503|%p32, %r501, %r485, %r474, %r476;
mov.b32 %f219, %r503;
add.ftz.f32 %f220, %f218, %f219;
mov.b32 %r504, %f220;
shfl.sync.bfly.b32 %r506|%p33, %r504, %r488, %r474, %r476;
mov.b32 %f221, %r506;
add.ftz.f32 %f52, %f220, %f221;
@%p22 bra $L__BB55_14;

add.s32 %r507, %r36, 1;
cvt.u64.u32 %rd76, %r507;
add.s64 %rd77, %rd76, %rd11;
shl.b64 %rd79, %rd77, 2;
add.s64 %rd80, %rd69, %rd79;
st.global.f32 [%rd80], %f52;

$L__BB55_14:
shl.b32 %r508, %r36, 1;
mov.u32 %r509, 1;
ld.shared.f32 %f222, [%r30+512];
add.ftz.f32 %f223, %f222, %f381;
mov.b32 %r510, %f223;
mov.u32 %r511, 31;
mov.u32 %r512, 16;
mov.u32 %r513, -1;
shfl.sync.bfly.b32 %r514|%p35, %r510, %r512, %r511, %r513;
mov.b32 %f224, %r514;
add.ftz.f32 %f225, %f223, %f224;
mov.b32 %r515, %f225;
mov.u32 %r516, 8;
shfl.sync.bfly.b32 %r517|%p36, %r515, %r516, %r511, %r513;
mov.b32 %f226, %r517;
add.ftz.f32 %f227, %f225, %f226;
mov.b32 %r518, %f227;
mov.u32 %r519, 4;
shfl.sync.bfly.b32 %r520|%p37, %r518, %r519, %r511, %r513;
mov.b32 %f228, %r520;
add.ftz.f32 %f229, %f227, %f228;
mov.b32 %r521, %f229;
mov.u32 %r522, 2;
shfl.sync.bfly.b32 %r523|%p38, %r521, %r522, %r511, %r513;
mov.b32 %f230, %r523;
add.ftz.f32 %f231, %f229, %f230;
mov.b32 %r524, %f231;
shfl.sync.bfly.b32 %r525|%p39, %r524, %r509, %r511, %r513;
mov.b32 %f232, %r525;
add.ftz.f32 %f53, %f231, %f232;
cvt.u64.u32 %rd81, %r508;
add.s64 %rd82, %rd81, %rd11;
shl.b64 %rd84, %rd82, 2;
add.s64 %rd13, %rd69, %rd84;
@%p14 bra $L__BB55_16;

st.global.f32 [%rd13], %f53;

$L__BB55_16:
ld.shared.f32 %f233, [%r30+640];
add.ftz.f32 %f234, %f233, %f380;
mov.b32 %r526, %f234;
shfl.sync.bfly.b32 %r530|%p41, %r526, %r512, %r511, %r513;
mov.b32 %f235, %r530;
add.ftz.f32 %f236, %f234, %f235;
mov.b32 %r531, %f236;
shfl.sync.bfly.b32 %r533|%p42, %r531, %r516, %r511, %r513;
mov.b32 %f237, %r533;
add.ftz.f32 %f238, %f236, %f237;
mov.b32 %r534, %f238;
shfl.sync.bfly.b32 %r536|%p43, %r534, %r519, %r511, %r513;
mov.b32 %f239, %r536;
add.ftz.f32 %f240, %f238, %f239;
mov.b32 %r537, %f240;
shfl.sync.bfly.b32 %r539|%p44, %r537, %r522, %r511, %r513;
mov.b32 %f241, %r539;
add.ftz.f32 %f242, %f240, %f241;
mov.b32 %r540, %f242;
shfl.sync.bfly.b32 %r542|%p45, %r540, %r509, %r511, %r513;
mov.b32 %f243, %r542;
add.ftz.f32 %f54, %f242, %f243;
@%p22 bra $L__BB55_18;

st.global.f32 [%rd13+4], %f54;

$L__BB55_18:
mul.lo.s32 %r31, %r36, 3;
ld.shared.f32 %f244, [%r30+768];
add.ftz.f32 %f245, %f244, %f379;
mov.b32 %r543, %f245;
mov.u32 %r544, 31;
mov.u32 %r545, 16;
mov.u32 %r546, -1;
shfl.sync.bfly.b32 %r547|%p47, %r543, %r545, %r544, %r546;
mov.b32 %f246, %r547;
add.ftz.f32 %f247, %f245, %f246;
mov.b32 %r548, %f247;
mov.u32 %r549, 8;
shfl.sync.bfly.b32 %r550|%p48, %r548, %r549, %r544, %r546;
mov.b32 %f248, %r550;
add.ftz.f32 %f249, %f247, %f248;
mov.b32 %r551, %f249;
mov.u32 %r552, 4;
shfl.sync.bfly.b32 %r553|%p49, %r551, %r552, %r544, %r546;
mov.b32 %f250, %r553;
add.ftz.f32 %f251, %f249, %f250;
mov.b32 %r554, %f251;
mov.u32 %r555, 2;
shfl.sync.bfly.b32 %r556|%p50, %r554, %r555, %r544, %r546;
mov.b32 %f252, %r556;
add.ftz.f32 %f253, %f251, %f252;
mov.b32 %r557, %f253;
mov.u32 %r558, 1;
shfl.sync.bfly.b32 %r559|%p51, %r557, %r558, %r544, %r546;
mov.b32 %f254, %r559;
add.ftz.f32 %f55, %f253, %f254;
@%p14 bra $L__BB55_20;

cvt.u64.u32 %rd85, %r31;
add.s64 %rd86, %rd85, %rd11;
shl.b64 %rd88, %rd86, 2;
add.s64 %rd89, %rd69, %rd88;
st.global.f32 [%rd89], %f55;

$L__BB55_20:
ld.shared.f32 %f255, [%r30+896];
add.ftz.f32 %f256, %f255, %f378;
mov.b32 %r560, %f256;
shfl.sync.bfly.b32 %r564|%p53, %r560, %r545, %r544, %r546;
mov.b32 %f257, %r564;
add.ftz.f32 %f258, %f256, %f257;
mov.b32 %r565, %f258;
shfl.sync.bfly.b32 %r567|%p54, %r565, %r549, %r544, %r546;
mov.b32 %f259, %r567;
add.ftz.f32 %f260, %f258, %f259;
mov.b32 %r568, %f260;
shfl.sync.bfly.b32 %r570|%p55, %r568, %r552, %r544, %r546;
mov.b32 %f261, %r570;
add.ftz.f32 %f262, %f260, %f261;
mov.b32 %r571, %f262;
shfl.sync.bfly.b32 %r573|%p56, %r571, %r555, %r544, %r546;
mov.b32 %f263, %r573;
add.ftz.f32 %f264, %f262, %f263;
mov.b32 %r574, %f264;
shfl.sync.bfly.b32 %r576|%p57, %r574, %r558, %r544, %r546;
mov.b32 %f265, %r576;
add.ftz.f32 %f56, %f264, %f265;
@%p22 bra $L__BB55_22;

add.s32 %r577, %r31, 1;
cvt.u64.u32 %rd90, %r577;
add.s64 %rd91, %rd90, %rd11;
shl.b64 %rd93, %rd91, 2;
add.s64 %rd94, %rd69, %rd93;
st.global.f32 [%rd94], %f56;

$L__BB55_22:
shl.b32 %r578, %r36, 2;
mov.u32 %r579, 2;
ld.shared.f32 %f266, [%r30+1024];
add.ftz.f32 %f267, %f266, %f377;
mov.b32 %r580, %f267;
mov.u32 %r581, 31;
mov.u32 %r582, 16;
mov.u32 %r583, -1;
shfl.sync.bfly.b32 %r584|%p59, %r580, %r582, %r581, %r583;
mov.b32 %f268, %r584;
add.ftz.f32 %f269, %f267, %f268;
mov.b32 %r585, %f269;
mov.u32 %r586, 8;
shfl.sync.bfly.b32 %r587|%p60, %r585, %r586, %r581, %r583;
mov.b32 %f270, %r587;
add.ftz.f32 %f271, %f269, %f270;
mov.b32 %r588, %f271;
mov.u32 %r589, 4;
shfl.sync.bfly.b32 %r590|%p61, %r588, %r589, %r581, %r583;
mov.b32 %f272, %r590;
add.ftz.f32 %f273, %f271, %f272;
mov.b32 %r591, %f273;
shfl.sync.bfly.b32 %r592|%p62, %r591, %r579, %r581, %r583;
mov.b32 %f274, %r592;
add.ftz.f32 %f275, %f273, %f274;
mov.b32 %r593, %f275;
mov.u32 %r594, 1;
shfl.sync.bfly.b32 %r595|%p63, %r593, %r594, %r581, %r583;
mov.b32 %f276, %r595;
add.ftz.f32 %f57, %f275, %f276;
cvt.u64.u32 %rd95, %r578;
add.s64 %rd96, %rd95, %rd11;
shl.b64 %rd98, %rd96, 2;
add.s64 %rd14, %rd69, %rd98;
@%p14 bra $L__BB55_24;

st.global.f32 [%rd14], %f57;

$L__BB55_24:
ld.shared.f32 %f277, [%r30+1152];
add.ftz.f32 %f278, %f277, %f376;
mov.b32 %r596, %f278;
shfl.sync.bfly.b32 %r600|%p65, %r596, %r582, %r581, %r583;
mov.b32 %f279, %r600;
add.ftz.f32 %f280, %f278, %f279;
mov.b32 %r601, %f280;
shfl.sync.bfly.b32 %r603|%p66, %r601, %r586, %r581, %r583;
mov.b32 %f281, %r603;
add.ftz.f32 %f282, %f280, %f281;
mov.b32 %r604, %f282;
shfl.sync.bfly.b32 %r606|%p67, %r604, %r589, %r581, %r583;
mov.b32 %f283, %r606;
add.ftz.f32 %f284, %f282, %f283;
mov.b32 %r607, %f284;
shfl.sync.bfly.b32 %r609|%p68, %r607, %r579, %r581, %r583;
mov.b32 %f285, %r609;
add.ftz.f32 %f286, %f284, %f285;
mov.b32 %r610, %f286;
shfl.sync.bfly.b32 %r612|%p69, %r610, %r594, %r581, %r583;
mov.b32 %f287, %r612;
add.ftz.f32 %f58, %f286, %f287;
@%p22 bra $L__BB55_26;

st.global.f32 [%rd14+4], %f58;

$L__BB55_26:
mul.lo.s32 %r32, %r36, 5;
ld.shared.f32 %f288, [%r30+1280];
add.ftz.f32 %f289, %f288, %f375;
mov.b32 %r613, %f289;
mov.u32 %r614, 31;
mov.u32 %r615, 16;
mov.u32 %r616, -1;
shfl.sync.bfly.b32 %r617|%p71, %r613, %r615, %r614, %r616;
mov.b32 %f290, %r617;
add.ftz.f32 %f291, %f289, %f290;
mov.b32 %r618, %f291;
mov.u32 %r619, 8;
shfl.sync.bfly.b32 %r620|%p72, %r618, %r619, %r614, %r616;
mov.b32 %f292, %r620;
add.ftz.f32 %f293, %f291, %f292;
mov.b32 %r621, %f293;
mov.u32 %r622, 4;
shfl.sync.bfly.b32 %r623|%p73, %r621, %r622, %r614, %r616;
mov.b32 %f294, %r623;
add.ftz.f32 %f295, %f293, %f294;
mov.b32 %r624, %f295;
mov.u32 %r625, 2;
shfl.sync.bfly.b32 %r626|%p74, %r624, %r625, %r614, %r616;
mov.b32 %f296, %r626;
add.ftz.f32 %f297, %f295, %f296;
mov.b32 %r627, %f297;
mov.u32 %r628, 1;
shfl.sync.bfly.b32 %r629|%p75, %r627, %r628, %r614, %r616;
mov.b32 %f298, %r629;
add.ftz.f32 %f59, %f297, %f298;
@%p14 bra $L__BB55_28;

cvt.u64.u32 %rd99, %r32;
add.s64 %rd100, %rd99, %rd11;
shl.b64 %rd102, %rd100, 2;
add.s64 %rd103, %rd69, %rd102;
st.global.f32 [%rd103], %f59;

$L__BB55_28:
ld.shared.f32 %f299, [%r30+1408];
add.ftz.f32 %f300, %f299, %f374;
mov.b32 %r630, %f300;
shfl.sync.bfly.b32 %r634|%p77, %r630, %r615, %r614, %r616;
mov.b32 %f301, %r634;
add.ftz.f32 %f302, %f300, %f301;
mov.b32 %r635, %f302;
shfl.sync.bfly.b32 %r637|%p78, %r635, %r619, %r614, %r616;
mov.b32 %f303, %r637;
add.ftz.f32 %f304, %f302, %f303;
mov.b32 %r638, %f304;
shfl.sync.bfly.b32 %r640|%p79, %r638, %r622, %r614, %r616;
mov.b32 %f305, %r640;
add.ftz.f32 %f306, %f304, %f305;
mov.b32 %r641, %f306;
shfl.sync.bfly.b32 %r643|%p80, %r641, %r625, %r614, %r616;
mov.b32 %f307, %r643;
add.ftz.f32 %f308, %f306, %f307;
mov.b32 %r644, %f308;
shfl.sync.bfly.b32 %r646|%p81, %r644, %r628, %r614, %r616;
mov.b32 %f309, %r646;
add.ftz.f32 %f60, %f308, %f309;
@%p22 bra $L__BB55_30;

add.s32 %r647, %r32, 1;
cvt.u64.u32 %rd104, %r647;
add.s64 %rd105, %rd104, %rd11;
shl.b64 %rd107, %rd105, 2;
add.s64 %rd108, %rd69, %rd107;
st.global.f32 [%rd108], %f60;

$L__BB55_30:
mul.lo.s32 %r648, %r36, 6;
ld.shared.f32 %f310, [%r30+1536];
add.ftz.f32 %f311, %f310, %f373;
mov.b32 %r649, %f311;
mov.u32 %r650, 31;
mov.u32 %r651, 16;
mov.u32 %r652, -1;
shfl.sync.bfly.b32 %r653|%p83, %r649, %r651, %r650, %r652;
mov.b32 %f312, %r653;
add.ftz.f32 %f313, %f311, %f312;
mov.b32 %r654, %f313;
mov.u32 %r655, 8;
shfl.sync.bfly.b32 %r656|%p84, %r654, %r655, %r650, %r652;
mov.b32 %f314, %r656;
add.ftz.f32 %f315, %f313, %f314;
mov.b32 %r657, %f315;
mov.u32 %r658, 4;
shfl.sync.bfly.b32 %r659|%p85, %r657, %r658, %r650, %r652;
mov.b32 %f316, %r659;
add.ftz.f32 %f317, %f315, %f316;
mov.b32 %r660, %f317;
mov.u32 %r661, 2;
shfl.sync.bfly.b32 %r662|%p86, %r660, %r661, %r650, %r652;
mov.b32 %f318, %r662;
add.ftz.f32 %f319, %f317, %f318;
mov.b32 %r663, %f319;
mov.u32 %r664, 1;
shfl.sync.bfly.b32 %r665|%p87, %r663, %r664, %r650, %r652;
mov.b32 %f320, %r665;
add.ftz.f32 %f61, %f319, %f320;
cvt.u64.u32 %rd109, %r648;
add.s64 %rd110, %rd109, %rd11;
shl.b64 %rd112, %rd110, 2;
add.s64 %rd15, %rd69, %rd112;
@%p14 bra $L__BB55_32;

st.global.f32 [%rd15], %f61;

$L__BB55_32:
ld.shared.f32 %f321, [%r30+1664];
add.ftz.f32 %f322, %f321, %f372;
mov.b32 %r666, %f322;
shfl.sync.bfly.b32 %r670|%p89, %r666, %r651, %r650, %r652;
mov.b32 %f323, %r670;
add.ftz.f32 %f324, %f322, %f323;
mov.b32 %r671, %f324;
shfl.sync.bfly.b32 %r673|%p90, %r671, %r655, %r650, %r652;
mov.b32 %f325, %r673;
add.ftz.f32 %f326, %f324, %f325;
mov.b32 %r674, %f326;
shfl.sync.bfly.b32 %r676|%p91, %r674, %r658, %r650, %r652;
mov.b32 %f327, %r676;
add.ftz.f32 %f328, %f326, %f327;
mov.b32 %r677, %f328;
shfl.sync.bfly.b32 %r679|%p92, %r677, %r661, %r650, %r652;
mov.b32 %f329, %r679;
add.ftz.f32 %f330, %f328, %f329;
mov.b32 %r680, %f330;
shfl.sync.bfly.b32 %r682|%p93, %r680, %r664, %r650, %r652;
mov.b32 %f331, %r682;
add.ftz.f32 %f62, %f330, %f331;
@%p22 bra $L__BB55_34;

st.global.f32 [%rd15+4], %f62;

$L__BB55_34:
mul.lo.s32 %r33, %r36, 7;
ld.shared.f32 %f332, [%r30+1792];
add.ftz.f32 %f333, %f332, %f371;
mov.b32 %r683, %f333;
mov.u32 %r684, 31;
mov.u32 %r685, 16;
mov.u32 %r686, -1;
shfl.sync.bfly.b32 %r687|%p95, %r683, %r685, %r684, %r686;
mov.b32 %f334, %r687;
add.ftz.f32 %f335, %f333, %f334;
mov.b32 %r688, %f335;
mov.u32 %r689, 8;
shfl.sync.bfly.b32 %r690|%p96, %r688, %r689, %r684, %r686;
mov.b32 %f336, %r690;
add.ftz.f32 %f337, %f335, %f336;
mov.b32 %r691, %f337;
mov.u32 %r692, 4;
shfl.sync.bfly.b32 %r693|%p97, %r691, %r692, %r684, %r686;
mov.b32 %f338, %r693;
add.ftz.f32 %f339, %f337, %f338;
mov.b32 %r694, %f339;
mov.u32 %r695, 2;
shfl.sync.bfly.b32 %r696|%p98, %r694, %r695, %r684, %r686;
mov.b32 %f340, %r696;
add.ftz.f32 %f341, %f339, %f340;
mov.b32 %r697, %f341;
mov.u32 %r698, 1;
shfl.sync.bfly.b32 %r699|%p99, %r697, %r698, %r684, %r686;
mov.b32 %f342, %r699;
add.ftz.f32 %f63, %f341, %f342;
@%p14 bra $L__BB55_36;

cvt.u64.u32 %rd113, %r33;
add.s64 %rd114, %rd113, %rd11;
shl.b64 %rd116, %rd114, 2;
add.s64 %rd117, %rd69, %rd116;
st.global.f32 [%rd117], %f63;

$L__BB55_36:
ld.shared.f32 %f343, [%r30+1920];
add.ftz.f32 %f344, %f343, %f370;
mov.b32 %r700, %f344;
shfl.sync.bfly.b32 %r704|%p101, %r700, %r685, %r684, %r686;
mov.b32 %f345, %r704;
add.ftz.f32 %f346, %f344, %f345;
mov.b32 %r705, %f346;
shfl.sync.bfly.b32 %r707|%p102, %r705, %r689, %r684, %r686;
mov.b32 %f347, %r707;
add.ftz.f32 %f348, %f346, %f347;
mov.b32 %r708, %f348;
shfl.sync.bfly.b32 %r710|%p103, %r708, %r692, %r684, %r686;
mov.b32 %f349, %r710;
add.ftz.f32 %f350, %f348, %f349;
mov.b32 %r711, %f350;
shfl.sync.bfly.b32 %r713|%p104, %r711, %r695, %r684, %r686;
mov.b32 %f351, %r713;
add.ftz.f32 %f352, %f350, %f351;
mov.b32 %r714, %f352;
shfl.sync.bfly.b32 %r716|%p105, %r714, %r698, %r684, %r686;
mov.b32 %f353, %r716;
add.ftz.f32 %f64, %f352, %f353;
@%p22 bra $L__BB55_38;

add.s32 %r717, %r33, 1;
cvt.u64.u32 %rd118, %r717;
add.s64 %rd119, %rd118, %rd11;
shl.b64 %rd121, %rd119, 2;
add.s64 %rd122, %rd69, %rd121;
st.global.f32 [%rd122], %f64;

$L__BB55_38:
ret;

}
.entry _Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj(
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2,
.param .align 8 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3[48],
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5,
.param .align 4 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6[12],
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15
)
.maxntid 256, 1, 1
.minnctapersm 1
{
.local .align 8 .b8 __local_depot56[16];
.reg .b64 %SP;
.reg .b64 %SPL;
.reg .pred %p<48>;
.reg .b16 %rs<29>;
.reg .f32 %f<177>;
.reg .b32 %r<491>;
.reg .b64 %rd<104>;


mov.u64 %SPL, __local_depot56;
ld.param.u64 %rd21, [_Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0];
ld.param.u64 %rd22, [_Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1];
ld.param.u64 %rd23, [_Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2];
ld.param.u32 %r17, [_Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5];
ld.param.u32 %r22, [_Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8];
ld.param.u32 %r23, [_Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9];
ld.param.u32 %r25, [_Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11];
ld.param.u32 %r26, [_Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12];
ld.param.u32 %r29, [_Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14];
ld.param.u32 %r28, [_Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15];
ld.param.u32 %r20, [_Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+8];
ld.param.u32 %r19, [_Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+4];
ld.param.u32 %r18, [_Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6];
ld.param.u64 %rd25, [_Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+8];
add.u64 %rd1, %SPL, 0;
add.u64 %rd2, %SPL, 8;
mov.u32 %r1, %tid.y;
setp.ge.u32 %p1, %r1, %r29;
@%p1 bra $L__BB56_29;

mov.u32 %r2, %ctaid.y;
cvta.to.global.u64 %rd32, %rd23;
mad.lo.s32 %r30, %r1, %r28, %r2;
mul.wide.u32 %rd33, %r30, 4;
add.s64 %rd34, %rd32, %rd33;
ld.global.nc.u32 %r3, [%rd34];
mov.f32 %f170, 0f00000000;
st.local.v2.f32 [%rd1], {%f170, %f170};
st.local.v2.f32 [%rd2], {%f170, %f170};
mov.u32 %r4, %tid.x;
shr.u32 %r5, %r4, 1;
shr.u32 %r31, %r17, 5;
setp.ge.u32 %p2, %r5, %r31;
mov.f32 %f171, %f170;
mov.f32 %f172, %f170;
mov.f32 %f173, %f170;
@%p2 bra $L__BB56_9;

mul.hi.u32 %r32, %r2, %r18;
add.s32 %r33, %r32, %r2;
shr.u32 %r34, %r33, %r19;
mul.lo.s32 %r35, %r34, %r20;
sub.s32 %r36, %r2, %r35;
mul.lo.s32 %r37, %r36, %r26;
cvt.u64.u32 %rd35, %r37;
cvt.u64.u32 %rd36, %r4;
shr.u64 %rd37, %rd36, 1;
mul.wide.u32 %rd103, %r37, 36;
and.b32 %r38, %r4, 1;
mul.wide.u32 %rd38, %r38, 8;
mul.lo.s32 %r39, %r1, %r23;
cvt.u64.u32 %rd39, %r39;
add.s64 %rd40, %rd37, %rd39;
add.s64 %rd41, %rd40, %rd35;
mul.lo.s64 %rd42, %rd41, 36;
add.s64 %rd43, %rd42, %rd38;
cvta.to.global.u64 %rd44, %rd22;
add.s64 %rd45, %rd44, %rd43;
add.s64 %rd101, %rd45, 20;
mov.u32 %r40, %ctaid.x;
shl.b32 %r41, %r40, 1;
or.b32 %r42, %r41, 1;
mad.lo.s32 %r43, %r3, %r25, %r5;
mad.lo.s32 %r44, %r22, %r42, %r43;
mul.wide.s32 %rd6, %r44, 22;
add.s64 %rd7, %rd38, %rd6;
cvta.to.global.u64 %rd102, %rd21;
cvta.to.global.u64 %rd100, %rd25;
mov.f32 %f170, 0f00000000;
mov.u32 %r490, %r5;

$L__BB56_3:
mad.lo.s32 %r67, %r41, %r22, %r5;
mad.lo.s32 %r68, %r3, %r25, %r67;
mul.wide.s32 %rd46, %r68, 22;
add.s64 %rd47, %rd102, %rd46;
ld.global.nc.u16 %rs2, [%rd47+2];
ld.global.nc.u16 %rs3, [%rd47+4];
mov.b32 %r69, {%rs2, %rs3};
add.s64 %rd49, %rd38, %rd46;
add.s64 %rd50, %rd102, %rd49;
ld.global.nc.u16 %rs4, [%rd50+6];
ld.global.nc.u16 %rs5, [%rd50+8];
mov.b32 %r71, {%rs4, %rs5};
shl.b32 %r72, %r4, 3;
and.b32 %r73, %r72, 8;
shr.s32 %r74, %r69, %r73;
ld.global.nc.u32 %r7, [%rd101+-16];
ld.global.nc.u32 %r8, [%rd101];
ld.global.nc.u16 %rs6, [%rd50+10];
ld.global.nc.u16 %rs7, [%rd50+12];
mov.b32 %r75, {%rs6, %rs7};
shl.b32 %r76, %r4, 2;
and.b32 %r77, %r76, 4;
or.b32 %r78, %r77, 1;
shl.b32 %r79, %r4, 1;
and.b32 %r80, %r79, 2;
sub.s32 %r81, %r78, %r80;
shl.b32 %r82, %r81, 2;
shr.s32 %r83, %r69, %r82;
ld.global.nc.u32 %r9, [%rd101+-12];
ld.global.nc.u32 %r10, [%rd101+4];
ld.global.nc.u16 %rs1, [%rd47];

{ cvt.f32.f16 %f42, %rs1;}


and.b32 %r84, %r71, 252645135;
shl.b32 %r85, %r74, 4;
and.b32 %r86, %r85, 16;
or.b32 %r87, %r86, %r84;
shl.b32 %r88, %r74, 11;
and.b32 %r89, %r88, 4096;
or.b32 %r90, %r87, %r89;
shl.b32 %r91, %r74, 18;
and.b32 %r92, %r91, 1048576;
or.b32 %r93, %r90, %r92;
shl.b32 %r94, %r74, 25;
and.b32 %r95, %r94, 268435456;
or.b32 %r46, %r93, %r95;
mov.u32 %r48, 0;

dp4a.s32.s32 %r45, %r46, %r7, %r48;

shr.u32 %r96, %r71, 4;
and.b32 %r97, %r96, 252645135;
shr.u32 %r98, %r74, 12;
and.b32 %r99, %r98, 16;
or.b32 %r100, %r99, %r97;
shr.u32 %r101, %r74, 5;
and.b32 %r102, %r101, 4096;
or.b32 %r103, %r100, %r102;
shl.b32 %r104, %r74, 2;
and.b32 %r105, %r104, 1048576;
or.b32 %r106, %r103, %r105;
shl.b32 %r107, %r74, 9;
and.b32 %r108, %r107, 268435456;
or.b32 %r50, %r106, %r108;

dp4a.s32.s32 %r49, %r50, %r8, %r45;

and.b32 %r109, %r75, 252645135;
shl.b32 %r110, %r83, 4;
and.b32 %r111, %r110, 16;
or.b32 %r112, %r111, %r109;
shl.b32 %r113, %r83, 11;
and.b32 %r114, %r113, 4096;
or.b32 %r115, %r112, %r114;
shl.b32 %r116, %r83, 18;
and.b32 %r117, %r116, 1048576;
or.b32 %r118, %r115, %r117;
shl.b32 %r119, %r83, 25;
and.b32 %r120, %r119, 268435456;
or.b32 %r54, %r118, %r120;

dp4a.s32.s32 %r53, %r54, %r9, %r49;

shr.u32 %r121, %r75, 4;
and.b32 %r122, %r121, 252645135;
shr.u32 %r123, %r83, 12;
and.b32 %r124, %r123, 16;
or.b32 %r125, %r124, %r122;
shr.u32 %r126, %r83, 5;
and.b32 %r127, %r126, 4096;
or.b32 %r128, %r125, %r127;
shl.b32 %r129, %r83, 2;
and.b32 %r130, %r129, 1048576;
or.b32 %r131, %r128, %r130;
shl.b32 %r132, %r83, 9;
and.b32 %r133, %r132, 268435456;
or.b32 %r58, %r131, %r133;

dp4a.s32.s32 %r57, %r58, %r10, %r53;

mul.lo.s64 %rd53, %rd37, 36;
mul.wide.u32 %rd54, %r39, 36;
add.s64 %rd55, %rd53, %rd54;
add.s64 %rd57, %rd44, %rd55;
add.s64 %rd58, %rd57, %rd103;
ld.global.nc.u32 %r11, [%rd58];

{.reg .f16 low,high;
mov.b32 {low,high},%r11;
cvt.f32.f16 %f43, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r11;
cvt.f32.f16 %f44, high;}


cvt.rn.f32.s32 %f45, %r57;
mul.ftz.f32 %f46, %f44, 0fC1000000;
fma.rn.ftz.f32 %f47, %f43, %f45, %f46;
fma.rn.ftz.f32 %f173, %f42, %f47, %f173;
setp.eq.s64 %p3, %rd25, 0;
@%p3 bra $L__BB56_5;

shl.b32 %r487, %r4, 3;
and.b32 %r486, %r487, 8;
shr.u32 %r155, %r4, 1;
mad.lo.s32 %r158, %r41, %r22, %r155;
mad.lo.s32 %r159, %r3, %r25, %r158;
mul.wide.s32 %rd59, %r159, 22;
add.s64 %rd60, %rd100, %rd59;
ld.global.u16 %rs9, [%rd60+2];
ld.global.u16 %rs10, [%rd60+4];
mov.b32 %r160, {%rs9, %rs10};
add.s64 %rd62, %rd38, %rd59;
add.s64 %rd63, %rd100, %rd62;
ld.global.u16 %rs11, [%rd63+6];
ld.global.u16 %rs12, [%rd63+8];
mov.b32 %r162, {%rs11, %rs12};
shr.s32 %r165, %r160, %r486;
ld.global.u16 %rs13, [%rd63+10];
ld.global.u16 %rs14, [%rd63+12];
mov.b32 %r166, {%rs13, %rs14};
shr.s32 %r174, %r160, %r82;
ld.global.u16 %rs8, [%rd60];

{ cvt.f32.f16 %f48, %rs8;}


and.b32 %r175, %r162, 252645135;
shl.b32 %r176, %r165, 4;
and.b32 %r177, %r176, 16;
or.b32 %r178, %r177, %r175;
shl.b32 %r179, %r165, 11;
and.b32 %r180, %r179, 4096;
or.b32 %r181, %r178, %r180;
shl.b32 %r182, %r165, 18;
and.b32 %r183, %r182, 1048576;
or.b32 %r184, %r181, %r183;
shl.b32 %r185, %r165, 25;
and.b32 %r186, %r185, 268435456;
or.b32 %r137, %r184, %r186;

dp4a.s32.s32 %r136, %r137, %r7, %r48;

shr.u32 %r187, %r162, 4;
and.b32 %r188, %r187, 252645135;
shr.u32 %r189, %r165, 12;
and.b32 %r190, %r189, 16;
or.b32 %r191, %r190, %r188;
shr.u32 %r192, %r165, 5;
and.b32 %r193, %r192, 4096;
or.b32 %r194, %r191, %r193;
shl.b32 %r195, %r165, 2;
and.b32 %r196, %r195, 1048576;
or.b32 %r197, %r194, %r196;
shl.b32 %r198, %r165, 9;
and.b32 %r199, %r198, 268435456;
or.b32 %r141, %r197, %r199;

dp4a.s32.s32 %r140, %r141, %r8, %r136;

and.b32 %r200, %r166, 252645135;
shl.b32 %r201, %r174, 4;
and.b32 %r202, %r201, 16;
or.b32 %r203, %r202, %r200;
shl.b32 %r204, %r174, 11;
and.b32 %r205, %r204, 4096;
or.b32 %r206, %r203, %r205;
shl.b32 %r207, %r174, 18;
and.b32 %r208, %r207, 1048576;
or.b32 %r209, %r206, %r208;
shl.b32 %r210, %r174, 25;
and.b32 %r211, %r210, 268435456;
or.b32 %r145, %r209, %r211;

dp4a.s32.s32 %r144, %r145, %r9, %r140;

shr.u32 %r212, %r166, 4;
and.b32 %r213, %r212, 252645135;
shr.u32 %r214, %r174, 12;
and.b32 %r215, %r214, 16;
or.b32 %r216, %r215, %r213;
shr.u32 %r217, %r174, 5;
and.b32 %r218, %r217, 4096;
or.b32 %r219, %r216, %r218;
shl.b32 %r220, %r174, 2;
and.b32 %r221, %r220, 1048576;
or.b32 %r222, %r219, %r221;
shl.b32 %r223, %r174, 9;
and.b32 %r224, %r223, 268435456;
or.b32 %r149, %r222, %r224;

dp4a.s32.s32 %r148, %r149, %r10, %r144;


{.reg .f16 low,high;
mov.b32 {low,high},%r11;
cvt.f32.f16 %f49, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r11;
cvt.f32.f16 %f50, high;}


cvt.rn.f32.s32 %f51, %r148;
mul.ftz.f32 %f52, %f50, 0fC1000000;
fma.rn.ftz.f32 %f53, %f49, %f51, %f52;
fma.rn.ftz.f32 %f171, %f48, %f53, %f171;

$L__BB56_5:
shl.b32 %r485, %r4, 3;
and.b32 %r484, %r485, 8;
add.s64 %rd64, %rd102, %rd6;
ld.global.nc.u16 %rs16, [%rd64+2];
ld.global.nc.u16 %rs17, [%rd64+4];
mov.b32 %r243, {%rs16, %rs17};
add.s64 %rd65, %rd102, %rd7;
ld.global.nc.u16 %rs18, [%rd65+6];
ld.global.nc.u16 %rs19, [%rd65+8];
mov.b32 %r244, {%rs18, %rs19};
shr.s32 %r248, %r243, %r484;
ld.global.nc.u16 %rs20, [%rd65+10];
ld.global.nc.u16 %rs21, [%rd65+12];
mov.b32 %r249, {%rs20, %rs21};
shr.s32 %r257, %r243, %r82;
ld.global.nc.u16 %rs15, [%rd64];

{ cvt.f32.f16 %f54, %rs15;}


and.b32 %r258, %r244, 252645135;
shl.b32 %r259, %r248, 4;
and.b32 %r260, %r259, 16;
or.b32 %r261, %r260, %r258;
shl.b32 %r262, %r248, 11;
and.b32 %r263, %r262, 4096;
or.b32 %r264, %r261, %r263;
shl.b32 %r265, %r248, 18;
and.b32 %r266, %r265, 1048576;
or.b32 %r267, %r264, %r266;
shl.b32 %r268, %r248, 25;
and.b32 %r269, %r268, 268435456;
or.b32 %r226, %r267, %r269;

dp4a.s32.s32 %r225, %r226, %r7, %r48;

shr.u32 %r270, %r244, 4;
and.b32 %r271, %r270, 252645135;
shr.u32 %r272, %r248, 12;
and.b32 %r273, %r272, 16;
or.b32 %r274, %r273, %r271;
shr.u32 %r275, %r248, 5;
and.b32 %r276, %r275, 4096;
or.b32 %r277, %r274, %r276;
shl.b32 %r278, %r248, 2;
and.b32 %r279, %r278, 1048576;
or.b32 %r280, %r277, %r279;
shl.b32 %r281, %r248, 9;
and.b32 %r282, %r281, 268435456;
or.b32 %r230, %r280, %r282;

dp4a.s32.s32 %r229, %r230, %r8, %r225;

and.b32 %r283, %r249, 252645135;
shl.b32 %r284, %r257, 4;
and.b32 %r285, %r284, 16;
or.b32 %r286, %r285, %r283;
shl.b32 %r287, %r257, 11;
and.b32 %r288, %r287, 4096;
or.b32 %r289, %r286, %r288;
shl.b32 %r290, %r257, 18;
and.b32 %r291, %r290, 1048576;
or.b32 %r292, %r289, %r291;
shl.b32 %r293, %r257, 25;
and.b32 %r294, %r293, 268435456;
or.b32 %r234, %r292, %r294;

dp4a.s32.s32 %r233, %r234, %r9, %r229;

shr.u32 %r295, %r249, 4;
and.b32 %r296, %r295, 252645135;
shr.u32 %r297, %r257, 12;
and.b32 %r298, %r297, 16;
or.b32 %r299, %r298, %r296;
shr.u32 %r300, %r257, 5;
and.b32 %r301, %r300, 4096;
or.b32 %r302, %r299, %r301;
shl.b32 %r303, %r257, 2;
and.b32 %r304, %r303, 1048576;
or.b32 %r305, %r302, %r304;
shl.b32 %r306, %r257, 9;
and.b32 %r307, %r306, 268435456;
or.b32 %r238, %r305, %r307;

dp4a.s32.s32 %r237, %r238, %r10, %r233;


{.reg .f16 low,high;
mov.b32 {low,high},%r11;
cvt.f32.f16 %f55, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r11;
cvt.f32.f16 %f56, high;}


cvt.rn.f32.s32 %f57, %r237;
mul.ftz.f32 %f58, %f56, 0fC1000000;
fma.rn.ftz.f32 %f59, %f55, %f57, %f58;
fma.rn.ftz.f32 %f172, %f54, %f59, %f172;
@%p3 bra $L__BB56_7;

shl.b32 %r489, %r4, 3;
and.b32 %r488, %r489, 8;
add.s64 %rd66, %rd100, %rd6;
ld.global.u16 %rs23, [%rd66+2];
ld.global.u16 %rs24, [%rd66+4];
mov.b32 %r326, {%rs23, %rs24};
add.s64 %rd67, %rd100, %rd7;
ld.global.u16 %rs25, [%rd67+6];
ld.global.u16 %rs26, [%rd67+8];
mov.b32 %r327, {%rs25, %rs26};
shr.s32 %r331, %r326, %r488;
ld.global.u16 %rs27, [%rd67+10];
ld.global.u16 %rs28, [%rd67+12];
mov.b32 %r332, {%rs27, %rs28};
shr.s32 %r340, %r326, %r82;
ld.global.u16 %rs22, [%rd66];

{ cvt.f32.f16 %f60, %rs22;}


and.b32 %r341, %r327, 252645135;
shl.b32 %r342, %r331, 4;
and.b32 %r343, %r342, 16;
or.b32 %r344, %r343, %r341;
shl.b32 %r345, %r331, 11;
and.b32 %r346, %r345, 4096;
or.b32 %r347, %r344, %r346;
shl.b32 %r348, %r331, 18;
and.b32 %r349, %r348, 1048576;
or.b32 %r350, %r347, %r349;
shl.b32 %r351, %r331, 25;
and.b32 %r352, %r351, 268435456;
or.b32 %r309, %r350, %r352;
mov.u32 %r311, 0;

dp4a.s32.s32 %r308, %r309, %r7, %r311;

shr.u32 %r353, %r327, 4;
and.b32 %r354, %r353, 252645135;
shr.u32 %r355, %r331, 12;
and.b32 %r356, %r355, 16;
or.b32 %r357, %r356, %r354;
shr.u32 %r358, %r331, 5;
and.b32 %r359, %r358, 4096;
or.b32 %r360, %r357, %r359;
shl.b32 %r361, %r331, 2;
and.b32 %r362, %r361, 1048576;
or.b32 %r363, %r360, %r362;
shl.b32 %r364, %r331, 9;
and.b32 %r365, %r364, 268435456;
or.b32 %r313, %r363, %r365;

dp4a.s32.s32 %r312, %r313, %r8, %r308;

and.b32 %r366, %r332, 252645135;
shl.b32 %r367, %r340, 4;
and.b32 %r368, %r367, 16;
or.b32 %r369, %r368, %r366;
shl.b32 %r370, %r340, 11;
and.b32 %r371, %r370, 4096;
or.b32 %r372, %r369, %r371;
shl.b32 %r373, %r340, 18;
and.b32 %r374, %r373, 1048576;
or.b32 %r375, %r372, %r374;
shl.b32 %r376, %r340, 25;
and.b32 %r377, %r376, 268435456;
or.b32 %r317, %r375, %r377;

dp4a.s32.s32 %r316, %r317, %r9, %r312;

shr.u32 %r378, %r332, 4;
and.b32 %r379, %r378, 252645135;
shr.u32 %r380, %r340, 12;
and.b32 %r381, %r380, 16;
or.b32 %r382, %r381, %r379;
shr.u32 %r383, %r340, 5;
and.b32 %r384, %r383, 4096;
or.b32 %r385, %r382, %r384;
shl.b32 %r386, %r340, 2;
and.b32 %r387, %r386, 1048576;
or.b32 %r388, %r385, %r387;
shl.b32 %r389, %r340, 9;
and.b32 %r390, %r389, 268435456;
or.b32 %r321, %r388, %r390;

dp4a.s32.s32 %r320, %r321, %r10, %r316;


{.reg .f16 low,high;
mov.b32 {low,high},%r11;
cvt.f32.f16 %f61, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r11;
cvt.f32.f16 %f62, high;}


cvt.rn.f32.s32 %f63, %r320;
mul.ftz.f32 %f64, %f62, 0fC1000000;
fma.rn.ftz.f32 %f65, %f61, %f63, %f64;
fma.rn.ftz.f32 %f170, %f60, %f65, %f170;

$L__BB56_7:
add.s64 %rd103, %rd103, 576;
add.s64 %rd102, %rd102, 352;
add.s64 %rd101, %rd101, 576;
add.s64 %rd100, %rd100, 352;
add.s32 %r490, %r490, 16;
setp.lt.u32 %p5, %r490, %r31;
@%p5 bra $L__BB56_3;

add.u64 %rd92, %SPL, 0;
add.u64 %rd81, %SPL, 8;
st.local.v2.f32 [%rd92], {%f173, %f172};
st.local.v2.f32 [%rd81], {%f171, %f170};

$L__BB56_9:
mov.b32 %r13, %f173;
setp.ne.s64 %p6, %rd25, 0;
@%p6 bra $L__BB56_11;
bra.uni $L__BB56_10;

$L__BB56_11:
add.u64 %rd98, %SPL, 0;
add.u64 %rd90, %SPL, 8;
mov.u32 %r418, 31;
mov.u32 %r419, 16;
mov.u32 %r420, -1;
shfl.sync.bfly.b32 %r421|%p17, %r13, %r419, %r418, %r420;
mov.b32 %f86, %r421;
add.ftz.f32 %f87, %f173, %f86;
mov.b32 %r422, %f87;
mov.u32 %r423, 8;
shfl.sync.bfly.b32 %r424|%p18, %r422, %r423, %r418, %r420;
mov.b32 %f88, %r424;
add.ftz.f32 %f89, %f87, %f88;
mov.b32 %r425, %f89;
mov.u32 %r426, 4;
shfl.sync.bfly.b32 %r427|%p19, %r425, %r426, %r418, %r420;
mov.b32 %f90, %r427;
add.ftz.f32 %f91, %f89, %f90;
mov.b32 %r428, %f91;
mov.u32 %r429, 2;
shfl.sync.bfly.b32 %r430|%p20, %r428, %r429, %r418, %r420;
mov.b32 %f92, %r430;
add.ftz.f32 %f93, %f91, %f92;
mov.b32 %r431, %f93;
mov.u32 %r432, 1;
shfl.sync.bfly.b32 %r433|%p21, %r431, %r432, %r418, %r420;
mov.b32 %f94, %r433;
add.ftz.f32 %f95, %f93, %f94;
st.local.f32 [%rd98], %f95;
mov.b32 %r434, %f171;
shfl.sync.bfly.b32 %r435|%p22, %r434, %r419, %r418, %r420;
mov.b32 %f96, %r435;
add.ftz.f32 %f97, %f171, %f96;
mov.b32 %r436, %f97;
shfl.sync.bfly.b32 %r437|%p23, %r436, %r423, %r418, %r420;
mov.b32 %f98, %r437;
add.ftz.f32 %f99, %f97, %f98;
mov.b32 %r438, %f99;
shfl.sync.bfly.b32 %r439|%p24, %r438, %r426, %r418, %r420;
mov.b32 %f100, %r439;
add.ftz.f32 %f101, %f99, %f100;
mov.b32 %r440, %f101;
shfl.sync.bfly.b32 %r441|%p25, %r440, %r429, %r418, %r420;
mov.b32 %f102, %r441;
add.ftz.f32 %f103, %f101, %f102;
mov.b32 %r442, %f103;
shfl.sync.bfly.b32 %r443|%p26, %r442, %r432, %r418, %r420;
mov.b32 %f104, %r443;
add.ftz.f32 %f105, %f103, %f104;
st.local.f32 [%rd90], %f105;
mov.b32 %r444, %f172;
shfl.sync.bfly.b32 %r445|%p27, %r444, %r419, %r418, %r420;
mov.b32 %f106, %r445;
add.ftz.f32 %f107, %f172, %f106;
mov.b32 %r446, %f107;
shfl.sync.bfly.b32 %r447|%p28, %r446, %r423, %r418, %r420;
mov.b32 %f108, %r447;
add.ftz.f32 %f109, %f107, %f108;
mov.b32 %r448, %f109;
shfl.sync.bfly.b32 %r449|%p29, %r448, %r426, %r418, %r420;
mov.b32 %f110, %r449;
add.ftz.f32 %f111, %f109, %f110;
mov.b32 %r450, %f111;
shfl.sync.bfly.b32 %r451|%p30, %r450, %r429, %r418, %r420;
mov.b32 %f112, %r451;
add.ftz.f32 %f113, %f111, %f112;
mov.b32 %r452, %f113;
shfl.sync.bfly.b32 %r453|%p31, %r452, %r432, %r418, %r420;
mov.b32 %f114, %r453;
add.ftz.f32 %f115, %f113, %f114;
st.local.f32 [%rd98+4], %f115;
mov.b32 %r454, %f170;
shfl.sync.bfly.b32 %r455|%p32, %r454, %r419, %r418, %r420;
mov.b32 %f116, %r455;
add.ftz.f32 %f117, %f170, %f116;
mov.b32 %r456, %f117;
shfl.sync.bfly.b32 %r457|%p33, %r456, %r423, %r418, %r420;
mov.b32 %f118, %r457;
add.ftz.f32 %f119, %f117, %f118;
mov.b32 %r458, %f119;
shfl.sync.bfly.b32 %r459|%p34, %r458, %r426, %r418, %r420;
mov.b32 %f120, %r459;
add.ftz.f32 %f121, %f119, %f120;
mov.b32 %r460, %f121;
shfl.sync.bfly.b32 %r461|%p35, %r460, %r429, %r418, %r420;
mov.b32 %f122, %r461;
add.ftz.f32 %f123, %f121, %f122;
mov.b32 %r462, %f123;
shfl.sync.bfly.b32 %r463|%p36, %r462, %r432, %r418, %r420;
mov.b32 %f124, %r463;
add.ftz.f32 %f125, %f123, %f124;
st.local.f32 [%rd90+4], %f125;
bra.uni $L__BB56_12;

$L__BB56_10:
add.u64 %rd94, %SPL, 0;
mov.u32 %r392, 31;
mov.u32 %r393, 16;
mov.u32 %r394, -1;
shfl.sync.bfly.b32 %r395|%p7, %r13, %r393, %r392, %r394;
mov.b32 %f66, %r395;
add.ftz.f32 %f67, %f173, %f66;
mov.b32 %r396, %f67;
mov.u32 %r397, 8;
shfl.sync.bfly.b32 %r398|%p8, %r396, %r397, %r392, %r394;
mov.b32 %f68, %r398;
add.ftz.f32 %f69, %f67, %f68;
mov.b32 %r399, %f69;
mov.u32 %r400, 4;
shfl.sync.bfly.b32 %r401|%p9, %r399, %r400, %r392, %r394;
mov.b32 %f70, %r401;
add.ftz.f32 %f71, %f69, %f70;
mov.b32 %r402, %f71;
mov.u32 %r403, 2;
shfl.sync.bfly.b32 %r404|%p10, %r402, %r403, %r392, %r394;
mov.b32 %f72, %r404;
add.ftz.f32 %f73, %f71, %f72;
mov.b32 %r405, %f73;
mov.u32 %r406, 1;
shfl.sync.bfly.b32 %r407|%p11, %r405, %r406, %r392, %r394;
mov.b32 %f74, %r407;
add.ftz.f32 %f75, %f73, %f74;
st.local.f32 [%rd94], %f75;
mov.b32 %r408, %f172;
shfl.sync.bfly.b32 %r409|%p12, %r408, %r393, %r392, %r394;
mov.b32 %f76, %r409;
add.ftz.f32 %f77, %f172, %f76;
mov.b32 %r410, %f77;
shfl.sync.bfly.b32 %r411|%p13, %r410, %r397, %r392, %r394;
mov.b32 %f78, %r411;
add.ftz.f32 %f79, %f77, %f78;
mov.b32 %r412, %f79;
shfl.sync.bfly.b32 %r413|%p14, %r412, %r400, %r392, %r394;
mov.b32 %f80, %r413;
add.ftz.f32 %f81, %f79, %f80;
mov.b32 %r414, %f81;
shfl.sync.bfly.b32 %r415|%p15, %r414, %r403, %r392, %r394;
mov.b32 %f82, %r415;
add.ftz.f32 %f83, %f81, %f82;
mov.b32 %r416, %f83;
shfl.sync.bfly.b32 %r417|%p16, %r416, %r406, %r392, %r394;
mov.b32 %f84, %r417;
add.ftz.f32 %f85, %f83, %f84;
st.local.f32 [%rd94+4], %f85;

$L__BB56_12:
ld.param.u32 %r473, [_Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7];
setp.gt.u32 %p37, %r4, 1;
mov.u32 %r465, %ctaid.x;
shl.b32 %r466, %r465, 1;
add.s32 %r14, %r466, %r4;
setp.ge.u32 %p38, %r14, %r473;
or.pred %p39, %p37, %p38;
@%p39 bra $L__BB56_29;

add.u64 %rd96, %SPL, 0;
ld.param.u64 %rd83, [_Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3];
ld.param.u32 %r474, [_Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13];
cvt.u64.u32 %rd19, %r4;
mul.wide.u32 %rd68, %r4, 4;
add.s64 %rd69, %rd96, %rd68;
ld.local.f32 %f176, [%rd69];
mad.lo.s32 %r15, %r3, %r474, %r14;
setp.eq.s64 %p40, %rd83, 0;
@%p40 bra $L__BB56_15;

ld.param.u64 %rd84, [_Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3];
cvta.to.global.u64 %rd70, %rd84;
mul.wide.u32 %rd71, %r15, 4;
add.s64 %rd72, %rd70, %rd71;
ld.global.f32 %f126, [%rd72];
add.ftz.f32 %f176, %f176, %f126;

$L__BB56_15:
setp.eq.s64 %p41, %rd25, 0;
@%p41 bra $L__BB56_28;

ld.param.u64 %rd88, [_Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+16];
add.u64 %rd86, %SPL, 8;
shl.b64 %rd73, %rd19, 2;
add.s64 %rd74, %rd86, %rd73;
ld.local.f32 %f24, [%rd74];
setp.eq.s64 %p42, %rd88, 0;
@%p42 bra $L__BB56_18;

ld.param.u64 %rd89, [_Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+16];
cvta.to.global.u64 %rd75, %rd89;
mul.wide.u32 %rd76, %r15, 4;
add.s64 %rd77, %rd75, %rd76;
ld.global.f32 %f127, [%rd77];
add.ftz.f32 %f24, %f24, %f127;

$L__BB56_18:
ld.param.u32 %r479, [_Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.gt.s32 %p43, %r479, 2;
@%p43 bra $L__BB56_22;

ld.param.u32 %r482, [_Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.eq.s32 %p46, %r482, 1;
@%p46 bra $L__BB56_26;

ld.param.u32 %r483, [_Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.eq.s32 %p47, %r483, 2;
@%p47 bra $L__BB56_21;
bra.uni $L__BB56_27;

$L__BB56_21:
mul.ftz.f32 %f147, %f24, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f148, %f147;
add.ftz.f32 %f149, %f148, 0f3F800000;
div.approx.ftz.f32 %f150, %f24, %f149;
mul.ftz.f32 %f176, %f176, %f150;
bra.uni $L__BB56_28;

$L__BB56_22:
ld.param.u32 %r480, [_Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.eq.s32 %p44, %r480, 3;
@%p44 bra $L__BB56_25;

ld.param.u32 %r481, [_Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.ne.s32 %p45, %r481, 6;
@%p45 bra $L__BB56_27;

ld.param.f32 %f159, [_Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+44];
min.ftz.f32 %f128, %f24, %f159;
neg.ftz.f32 %f129, %f159;
min.ftz.f32 %f130, %f176, %f159;
max.ftz.f32 %f131, %f130, %f129;
mul.ftz.f32 %f132, %f128, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f133, %f132;
add.ftz.f32 %f134, %f133, 0f3F800000;
div.approx.ftz.f32 %f135, %f128, %f134;
mul.ftz.f32 %f176, %f131, %f135;
bra.uni $L__BB56_28;

$L__BB56_27:
mul.ftz.f32 %f176, %f176, %f24;
bra.uni $L__BB56_28;

$L__BB56_26:
mul.ftz.f32 %f151, %f24, 0f3F000000;
mul.ftz.f32 %f152, %f24, 0f3D372713;
fma.rn.ftz.f32 %f153, %f24, %f152, 0f3F800000;
mul.ftz.f32 %f154, %f24, 0f3F4C422A;
mul.ftz.f32 %f155, %f154, %f153;
tanh.approx.f32 %f156, %f155;
add.ftz.f32 %f157, %f156, 0f3F800000;
mul.ftz.f32 %f158, %f151, %f157;
mul.ftz.f32 %f176, %f176, %f158;
bra.uni $L__BB56_28;

$L__BB56_25:
mov.f32 %f136, 0f40E00000;
min.ftz.f32 %f137, %f24, %f136;
min.ftz.f32 %f138, %f176, %f136;
mov.f32 %f139, 0fC0E00000;
max.ftz.f32 %f140, %f138, %f139;
mul.ftz.f32 %f141, %f137, 0f3FD9DB23;
mul.ftz.f32 %f142, %f141, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f143, %f142;
add.ftz.f32 %f144, %f143, 0f3F800000;
div.approx.ftz.f32 %f145, %f137, %f144;
add.ftz.f32 %f146, %f140, 0f3F800000;
mul.ftz.f32 %f176, %f146, %f145;

$L__BB56_28:
ld.param.u64 %rd85, [_Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4];
ld.param.u32 %r478, [_Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10];
ld.param.u32 %r477, [_Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13];
mov.u32 %r476, %ctaid.y;
mov.u32 %r475, %tid.y;
mad.lo.s32 %r471, %r475, %r478, %r14;
mad.lo.s32 %r472, %r476, %r477, %r471;
cvta.to.global.u64 %rd78, %rd85;
mul.wide.u32 %rd79, %r472, 4;
add.s64 %rd80, %rd78, %rd79;
st.global.f32 [%rd80], %f176;

$L__BB56_29:
ret;

}
.entry _Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj(
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2,
.param .align 8 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3[48],
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5,
.param .align 4 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6[12],
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15
)
.maxntid 256, 1, 1
.minnctapersm 1
{
.local .align 8 .b8 __local_depot57[8];
.reg .b64 %SP;
.reg .b64 %SPL;
.reg .pred %p<17>;
.reg .b16 %rs<15>;
.reg .f32 %f<50>;
.reg .b32 %r<229>;
.reg .b64 %rd<54>;


mov.u64 %SPL, __local_depot57;
ld.param.u64 %rd17, [_Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0];
ld.param.u64 %rd18, [_Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1];
ld.param.u64 %rd19, [_Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2];
ld.param.u64 %rd20, [_Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4];
ld.param.u32 %r21, [_Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5];
ld.param.u32 %r13, [_Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7];
ld.param.u32 %r14, [_Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8];
ld.param.u32 %r15, [_Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9];
ld.param.u32 %r16, [_Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10];
ld.param.u32 %r17, [_Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11];
ld.param.u32 %r18, [_Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12];
ld.param.u32 %r19, [_Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13];
ld.param.u32 %r22, [_Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14];
ld.param.u32 %r20, [_Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15];
ld.param.u32 %r12, [_Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+8];
ld.param.u32 %r11, [_Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+4];
ld.param.u32 %r10, [_Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6];
add.u64 %rd1, %SPL, 0;
shr.u32 %r1, %r21, 5;
mov.u32 %r2, %tid.y;
setp.ge.u32 %p1, %r2, %r22;
@%p1 bra $L__BB57_7;

mov.f32 %f48, 0f00000000;
st.local.v2.f32 [%rd1], {%f48, %f48};
mov.u32 %r3, %tid.x;
shr.u32 %r228, %r3, 1;
setp.ge.u32 %p2, %r228, %r1;
mov.f32 %f49, %f48;
@%p2 bra $L__BB57_5;

mov.u32 %r23, %ctaid.y;
mul.hi.u32 %r24, %r23, %r10;
add.s32 %r25, %r24, %r23;
shr.u32 %r26, %r25, %r11;
mul.lo.s32 %r27, %r26, %r12;
sub.s32 %r28, %r23, %r27;
mul.lo.s32 %r29, %r28, %r18;
cvt.u64.u32 %rd22, %r29;
mul.lo.s32 %r30, %r2, %r15;
cvt.u64.u32 %rd23, %r30;
mad.lo.s32 %r31, %r2, %r20, %r23;
cvta.to.global.u64 %rd24, %rd19;
mul.wide.u32 %rd25, %r31, 4;
add.s64 %rd26, %rd24, %rd25;
ld.global.nc.u32 %r32, [%rd26];
mul.lo.s32 %r33, %r32, %r17;
shl.b32 %r34, %r3, 1;
and.b32 %r35, %r34, 2;
mov.u32 %r36, %ctaid.x;
shl.b32 %r37, %r36, 1;
shl.b32 %r38, %r3, 2;
and.b32 %r39, %r38, 4;
or.b32 %r40, %r39, 1;
shl.b32 %r41, %r3, 3;
and.b32 %r5, %r41, 8;
sub.s32 %r42, %r40, %r35;
shl.b32 %r6, %r42, 2;
cvt.u64.u32 %rd27, %r3;
shr.u64 %rd28, %rd27, 1;
mul.lo.s64 %rd29, %rd28, 36;
mul.wide.u32 %rd30, %r30, 36;
add.s64 %rd31, %rd29, %rd30;
cvta.to.global.u64 %rd32, %rd18;
add.s64 %rd3, %rd32, %rd31;
mul.wide.u32 %rd53, %r29, 36;
and.b32 %r43, %r3, 1;
mul.wide.u32 %rd33, %r43, 8;
mad.lo.s32 %r44, %r37, %r14, %r228;
add.s32 %r45, %r44, %r33;
mul.wide.s32 %rd5, %r45, 22;
add.s64 %rd34, %rd33, %rd5;
add.s64 %rd6, %rd34, 12;
add.s64 %rd35, %rd28, %rd23;
add.s64 %rd36, %rd35, %rd22;
mul.lo.s64 %rd37, %rd36, 36;
add.s64 %rd38, %rd37, %rd33;
add.s64 %rd39, %rd32, %rd38;
add.s64 %rd51, %rd39, 20;
or.b32 %r46, %r37, 1;
add.s32 %r47, %r228, %r33;
mad.lo.s32 %r48, %r14, %r46, %r47;
mul.wide.s32 %rd8, %r48, 22;
add.s64 %rd40, %rd33, %rd8;
add.s64 %rd9, %rd40, 12;
cvta.to.global.u64 %rd52, %rd17;
mov.f32 %f48, 0f00000000;

$L__BB57_3:
add.s64 %rd41, %rd3, %rd53;
add.s64 %rd42, %rd52, %rd5;
ld.global.nc.u16 %rs3, [%rd42+2];
ld.global.nc.u16 %rs4, [%rd42+4];
mov.b32 %r85, {%rs3, %rs4};
add.s64 %rd43, %rd52, %rd6;
ld.global.nc.u16 %rs5, [%rd43+-6];
ld.global.nc.u16 %rs6, [%rd43+-4];
mov.b32 %r86, {%rs5, %rs6};
shr.s32 %r87, %r85, %r5;
ld.global.nc.u32 %r51, [%rd51+-16];
ld.global.nc.u32 %r55, [%rd51];
ld.global.nc.u16 %rs7, [%rd43+-2];
ld.global.nc.u16 %rs8, [%rd43];
mov.b32 %r88, {%rs7, %rs8};
shr.s32 %r89, %r85, %r6;
ld.global.nc.u32 %r59, [%rd51+-12];
ld.global.nc.u32 %r63, [%rd51+4];
ld.global.nc.u16 %rs1, [%rd42];

{ cvt.f32.f16 %f11, %rs1;}


and.b32 %r90, %r86, 252645135;
shl.b32 %r91, %r87, 4;
and.b32 %r92, %r91, 16;
or.b32 %r93, %r92, %r90;
shl.b32 %r94, %r87, 11;
and.b32 %r95, %r94, 4096;
or.b32 %r96, %r93, %r95;
shl.b32 %r97, %r87, 18;
and.b32 %r98, %r97, 1048576;
or.b32 %r99, %r96, %r98;
shl.b32 %r100, %r87, 25;
and.b32 %r101, %r100, 268435456;
or.b32 %r50, %r99, %r101;
mov.u32 %r70, 0;

dp4a.s32.s32 %r49, %r50, %r51, %r70;

shr.u32 %r102, %r86, 4;
and.b32 %r103, %r102, 252645135;
shr.u32 %r104, %r87, 12;
and.b32 %r105, %r104, 16;
or.b32 %r106, %r105, %r103;
shr.u32 %r107, %r87, 5;
and.b32 %r108, %r107, 4096;
or.b32 %r109, %r106, %r108;
shl.b32 %r110, %r87, 2;
and.b32 %r111, %r110, 1048576;
or.b32 %r112, %r109, %r111;
shl.b32 %r113, %r87, 9;
and.b32 %r114, %r113, 268435456;
or.b32 %r54, %r112, %r114;

dp4a.s32.s32 %r53, %r54, %r55, %r49;

and.b32 %r115, %r88, 252645135;
shl.b32 %r116, %r89, 4;
and.b32 %r117, %r116, 16;
or.b32 %r118, %r117, %r115;
shl.b32 %r119, %r89, 11;
and.b32 %r120, %r119, 4096;
or.b32 %r121, %r118, %r120;
shl.b32 %r122, %r89, 18;
and.b32 %r123, %r122, 1048576;
or.b32 %r124, %r121, %r123;
shl.b32 %r125, %r89, 25;
and.b32 %r126, %r125, 268435456;
or.b32 %r58, %r124, %r126;

dp4a.s32.s32 %r57, %r58, %r59, %r53;

shr.u32 %r127, %r88, 4;
and.b32 %r128, %r127, 252645135;
shr.u32 %r129, %r89, 12;
and.b32 %r130, %r129, 16;
or.b32 %r131, %r130, %r128;
shr.u32 %r132, %r89, 5;
and.b32 %r133, %r132, 4096;
or.b32 %r134, %r131, %r133;
shl.b32 %r135, %r89, 2;
and.b32 %r136, %r135, 1048576;
or.b32 %r137, %r134, %r136;
shl.b32 %r138, %r89, 9;
and.b32 %r139, %r138, 268435456;
or.b32 %r62, %r137, %r139;

dp4a.s32.s32 %r61, %r62, %r63, %r57;

ld.global.nc.u32 %r65, [%rd41];

{.reg .f16 low,high;
mov.b32 {low,high},%r65;
cvt.f32.f16 %f12, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r65;
cvt.f32.f16 %f13, high;}


cvt.rn.f32.s32 %f17, %r61;
mul.ftz.f32 %f18, %f12, %f17;
mul.ftz.f32 %f19, %f13, 0f41000000;
sub.ftz.f32 %f20, %f18, %f19;
fma.rn.ftz.f32 %f49, %f11, %f20, %f49;
add.s64 %rd44, %rd52, %rd8;
ld.global.nc.u16 %rs9, [%rd44+2];
ld.global.nc.u16 %rs10, [%rd44+4];
mov.b32 %r140, {%rs9, %rs10};
add.s64 %rd45, %rd52, %rd9;
ld.global.nc.u16 %rs11, [%rd45+-6];
ld.global.nc.u16 %rs12, [%rd45+-4];
mov.b32 %r141, {%rs11, %rs12};
shr.s32 %r142, %r140, %r5;
ld.global.nc.u16 %rs13, [%rd45+-2];
ld.global.nc.u16 %rs14, [%rd45];
mov.b32 %r143, {%rs13, %rs14};
shr.s32 %r144, %r140, %r6;
ld.global.nc.u16 %rs2, [%rd44];

{ cvt.f32.f16 %f14, %rs2;}


and.b32 %r145, %r141, 252645135;
shl.b32 %r146, %r142, 4;
and.b32 %r147, %r146, 16;
or.b32 %r148, %r147, %r145;
shl.b32 %r149, %r142, 11;
and.b32 %r150, %r149, 4096;
or.b32 %r151, %r148, %r150;
shl.b32 %r152, %r142, 18;
and.b32 %r153, %r152, 1048576;
or.b32 %r154, %r151, %r153;
shl.b32 %r155, %r142, 25;
and.b32 %r156, %r155, 268435456;
or.b32 %r68, %r154, %r156;

dp4a.s32.s32 %r67, %r68, %r51, %r70;

shr.u32 %r157, %r141, 4;
and.b32 %r158, %r157, 252645135;
shr.u32 %r159, %r142, 12;
and.b32 %r160, %r159, 16;
or.b32 %r161, %r160, %r158;
shr.u32 %r162, %r142, 5;
and.b32 %r163, %r162, 4096;
or.b32 %r164, %r161, %r163;
shl.b32 %r165, %r142, 2;
and.b32 %r166, %r165, 1048576;
or.b32 %r167, %r164, %r166;
shl.b32 %r168, %r142, 9;
and.b32 %r169, %r168, 268435456;
or.b32 %r72, %r167, %r169;

dp4a.s32.s32 %r71, %r72, %r55, %r67;

and.b32 %r170, %r143, 252645135;
shl.b32 %r171, %r144, 4;
and.b32 %r172, %r171, 16;
or.b32 %r173, %r172, %r170;
shl.b32 %r174, %r144, 11;
and.b32 %r175, %r174, 4096;
or.b32 %r176, %r173, %r175;
shl.b32 %r177, %r144, 18;
and.b32 %r178, %r177, 1048576;
or.b32 %r179, %r176, %r178;
shl.b32 %r180, %r144, 25;
and.b32 %r181, %r180, 268435456;
or.b32 %r76, %r179, %r181;

dp4a.s32.s32 %r75, %r76, %r59, %r71;

shr.u32 %r182, %r143, 4;
and.b32 %r183, %r182, 252645135;
shr.u32 %r184, %r144, 12;
and.b32 %r185, %r184, 16;
or.b32 %r186, %r185, %r183;
shr.u32 %r187, %r144, 5;
and.b32 %r188, %r187, 4096;
or.b32 %r189, %r186, %r188;
shl.b32 %r190, %r144, 2;
and.b32 %r191, %r190, 1048576;
or.b32 %r192, %r189, %r191;
shl.b32 %r193, %r144, 9;
and.b32 %r194, %r193, 268435456;
or.b32 %r80, %r192, %r194;

dp4a.s32.s32 %r79, %r80, %r63, %r75;


{.reg .f16 low,high;
mov.b32 {low,high},%r65;
cvt.f32.f16 %f15, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r65;
cvt.f32.f16 %f16, high;}


cvt.rn.f32.s32 %f21, %r79;
mul.ftz.f32 %f22, %f15, %f21;
mul.ftz.f32 %f23, %f16, 0f41000000;
sub.ftz.f32 %f24, %f22, %f23;
fma.rn.ftz.f32 %f48, %f14, %f24, %f48;
add.s64 %rd53, %rd53, 576;
add.s64 %rd52, %rd52, 352;
add.s64 %rd51, %rd51, 576;
add.s32 %r228, %r228, 16;
setp.lt.u32 %p3, %r228, %r1;
@%p3 bra $L__BB57_3;

st.local.v2.f32 [%rd1], {%f49, %f48};

$L__BB57_5:
mov.b32 %r195, %f49;
mov.u32 %r196, 31;
mov.u32 %r197, 16;
mov.u32 %r198, -1;
shfl.sync.bfly.b32 %r199|%p4, %r195, %r197, %r196, %r198;
mov.b32 %f25, %r199;
add.ftz.f32 %f26, %f49, %f25;
mov.b32 %r200, %f26;
mov.u32 %r201, 8;
shfl.sync.bfly.b32 %r202|%p5, %r200, %r201, %r196, %r198;
mov.b32 %f27, %r202;
add.ftz.f32 %f28, %f26, %f27;
mov.b32 %r203, %f28;
mov.u32 %r204, 4;
shfl.sync.bfly.b32 %r205|%p6, %r203, %r204, %r196, %r198;
mov.b32 %f29, %r205;
add.ftz.f32 %f30, %f28, %f29;
mov.b32 %r206, %f30;
mov.u32 %r207, 2;
shfl.sync.bfly.b32 %r208|%p7, %r206, %r207, %r196, %r198;
mov.b32 %f31, %r208;
add.ftz.f32 %f32, %f30, %f31;
mov.b32 %r209, %f32;
mov.u32 %r210, 1;
shfl.sync.bfly.b32 %r211|%p8, %r209, %r210, %r196, %r198;
mov.b32 %f33, %r211;
add.ftz.f32 %f34, %f32, %f33;
st.local.f32 [%rd1], %f34;
mov.b32 %r212, %f48;
shfl.sync.bfly.b32 %r213|%p9, %r212, %r197, %r196, %r198;
mov.b32 %f35, %r213;
add.ftz.f32 %f36, %f48, %f35;
mov.b32 %r214, %f36;
shfl.sync.bfly.b32 %r215|%p10, %r214, %r201, %r196, %r198;
mov.b32 %f37, %r215;
add.ftz.f32 %f38, %f36, %f37;
mov.b32 %r216, %f38;
shfl.sync.bfly.b32 %r217|%p11, %r216, %r204, %r196, %r198;
mov.b32 %f39, %r217;
add.ftz.f32 %f40, %f38, %f39;
mov.b32 %r218, %f40;
shfl.sync.bfly.b32 %r219|%p12, %r218, %r207, %r196, %r198;
mov.b32 %f41, %r219;
add.ftz.f32 %f42, %f40, %f41;
mov.b32 %r220, %f42;
shfl.sync.bfly.b32 %r221|%p13, %r220, %r210, %r196, %r198;
mov.b32 %f43, %r221;
add.ftz.f32 %f44, %f42, %f43;
st.local.f32 [%rd1+4], %f44;
mov.u32 %r222, %ctaid.x;
shl.b32 %r223, %r222, 1;
add.s32 %r9, %r223, %r3;
setp.ge.u32 %p14, %r9, %r13;
setp.gt.u32 %p15, %r3, 1;
or.pred %p16, %p15, %p14;
@%p16 bra $L__BB57_7;

mul.wide.u32 %rd46, %r3, 4;
add.s64 %rd47, %rd1, %rd46;
ld.local.f32 %f45, [%rd47];
mov.u32 %r224, %ctaid.y;
mad.lo.s32 %r226, %r2, %r16, %r9;
mad.lo.s32 %r227, %r224, %r19, %r226;
cvta.to.global.u64 %rd48, %rd20;
mul.wide.u32 %rd49, %r227, 4;
add.s64 %rd50, %rd48, %rd49;
st.global.f32 [%rd50], %f45;

$L__BB57_7:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<320>;
.reg .b16 %rs<57>;
.reg .f32 %f<1021>;
.reg .b32 %r<1478>;
.reg .b64 %rd<96>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate[1536];

ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd20, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r45, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r49, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r50, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r51, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r55, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r56, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r57, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r53, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r52, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r47, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r46, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
ld.param.u64 %rd23, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+16];
ld.param.u64 %rd22, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+8];
ld.param.u64 %rd21, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3];
mov.u32 %r1, %tid.y;
shl.b32 %r58, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r58, %r2;
mov.u32 %r59, %ctaid.x;
shl.b32 %r4, %r59, 2;
mov.u32 %r5, %ctaid.y;
setp.eq.s64 %p1, %rd20, 0;
@%p1 bra $L__BB58_2;

cvta.to.global.u64 %rd27, %rd20;
mul.wide.u32 %rd28, %r5, 4;
add.s64 %rd29, %rd27, %rd28;
ld.global.nc.u32 %r1475, [%rd29];
mul.hi.u32 %r60, %r5, %r41;
add.s32 %r61, %r60, %r5;
shr.u32 %r62, %r61, %r42;
mul.lo.s32 %r63, %r62, %r43;
sub.s32 %r1476, %r5, %r63;
bra.uni $L__BB58_3;

$L__BB58_2:
mul.hi.u32 %r64, %r5, %r46;
add.s32 %r65, %r64, %r5;
shr.u32 %r1475, %r65, %r47;
mov.u32 %r1476, %r5;

$L__BB58_3:
mov.u32 %r11, %ctaid.z;
selp.b32 %r12, %r5, %r1475, %p1;
setp.ne.s32 %p3, %r1, 0;
setp.gt.u32 %p4, %r2, 3;
or.pred %p5, %p4, %p3;
add.s32 %r66, %r4, %r2;
setp.ge.u32 %p6, %r66, %r45;
mov.f32 %f26, 0f00000000;
or.pred %p7, %p5, %p6;
mov.f32 %f4, %f26;
mov.f32 %f5, %f26;
@%p7 bra $L__BB58_8;

setp.eq.s64 %p8, %rd21, 0;
mov.f32 %f4, 0f00000000;
mov.f32 %f5, %f4;
@%p8 bra $L__BB58_6;

cvta.to.global.u64 %rd30, %rd21;
mul.lo.s32 %r67, %r11, %r57;
cvt.u64.u32 %rd31, %r67;
mul.lo.s32 %r68, %r12, %r51;
cvt.u64.u32 %rd32, %r68;
cvt.u64.u32 %rd33, %r2;
cvt.s64.s32 %rd34, %r4;
add.s64 %rd35, %rd34, %rd33;
add.s64 %rd36, %rd35, %rd31;
add.s64 %rd37, %rd36, %rd32;
shl.b64 %rd38, %rd37, 2;
add.s64 %rd39, %rd30, %rd38;
ld.global.f32 %f5, [%rd39];

$L__BB58_6:
setp.eq.s64 %p9, %rd23, 0;
setp.eq.s64 %p10, %rd22, 0;
or.pred %p11, %p9, %p10;
@%p11 bra $L__BB58_8;

cvta.to.global.u64 %rd40, %rd23;
mul.lo.s32 %r69, %r11, %r57;
cvt.u64.u32 %rd41, %r69;
mul.lo.s32 %r70, %r12, %r51;
cvt.u64.u32 %rd42, %r70;
cvt.u64.u32 %rd43, %r2;
cvt.s64.s32 %rd44, %r4;
add.s64 %rd45, %rd44, %rd43;
add.s64 %rd46, %rd45, %rd41;
add.s64 %rd47, %rd46, %rd42;
shl.b64 %rd48, %rd47, 2;
add.s64 %rd49, %rd40, %rd48;
ld.global.f32 %f4, [%rd49];

$L__BB58_8:
shr.u32 %r71, %r3, 31;
add.s32 %r72, %r3, %r71;
shr.s32 %r1477, %r72, 1;
shr.u32 %r73, %r40, 5;
setp.ge.s32 %p12, %r1477, %r73;
mov.f32 %f27, %f26;
mov.f32 %f28, %f26;
mov.f32 %f29, %f26;
mov.f32 %f30, %f26;
mov.f32 %f31, %f26;
mov.f32 %f32, %f26;
mov.f32 %f33, %f26;
@%p12 bra $L__BB58_19;

mul.hi.u32 %r74, %r11, %r52;
add.s32 %r75, %r74, %r11;
shr.u32 %r76, %r75, %r53;
mul.lo.s32 %r77, %r76, %r55;
mul.lo.s32 %r78, %r1475, %r49;
and.b32 %r81, %r72, -2;
sub.s32 %r14, %r3, %r81;
mad.lo.s32 %r82, %r4, %r44, %r1477;
add.s32 %r83, %r82, %r77;
add.s32 %r15, %r83, %r78;
add.s32 %r84, %r1477, %r77;
add.s32 %r16, %r84, %r78;
cvta.to.global.u64 %rd94, %rd22;
cvta.to.global.u64 %rd95, %rd18;
mul.lo.s32 %r85, %r1476, %r50;
cvt.u64.u32 %rd50, %r85;
mul.lo.s32 %r86, %r11, %r56;
cvt.u64.u32 %rd51, %r86;
add.s64 %rd6, %rd50, %rd51;
mov.f32 %f26, 0f00000000;
cvta.to.global.u64 %rd52, %rd19;
mov.f32 %f27, %f26;
mov.f32 %f28, %f26;
mov.f32 %f29, %f26;
mov.f32 %f30, %f26;
mov.f32 %f31, %f26;
mov.f32 %f32, %f26;
mov.f32 %f33, %f26;

$L__BB58_10:
cvt.u64.u32 %rd53, %r1477;
add.s64 %rd54, %rd6, %rd53;
mul.wide.s32 %rd55, %r15, 22;
add.s64 %rd9, %rd95, %rd55;
ld.global.nc.u16 %rs2, [%rd9+2];
ld.global.nc.u16 %rs3, [%rd9+4];
mov.b32 %r105, {%rs2, %rs3};
mul.lo.s64 %rd56, %rd54, 36;
add.s64 %rd57, %rd52, %rd56;
add.s64 %rd58, %rd57, 4;
shl.b32 %r106, %r14, 2;
mul.wide.s32 %rd59, %r106, 2;
add.s64 %rd60, %rd55, %rd59;
add.s64 %rd61, %rd95, %rd60;
add.s64 %rd10, %rd61, 6;
ld.global.nc.u16 %rs4, [%rd61+6];
ld.global.nc.u16 %rs5, [%rd61+8];
mov.b32 %r107, {%rs4, %rs5};
shl.b32 %r108, %r14, 3;
shr.s32 %r109, %r105, %r108;
shl.b32 %r110, %r14, 1;
mul.wide.s32 %rd62, %r110, 4;
add.s64 %rd63, %rd58, %rd62;
ld.global.nc.u32 %r18, [%rd63];
ld.global.nc.u32 %r19, [%rd63+16];
ld.global.nc.u16 %rs6, [%rd61+10];
ld.global.nc.u16 %rs7, [%rd61+12];
mov.b32 %r111, {%rs6, %rs7};
or.b32 %r112, %r106, 1;
sub.s32 %r113, %r112, %r110;
shl.b32 %r114, %r113, 2;
shr.s32 %r115, %r105, %r114;
mul.wide.s32 %rd64, %r113, 4;
add.s64 %rd65, %rd58, %rd64;
ld.global.nc.u32 %r20, [%rd65];
ld.global.nc.u32 %r21, [%rd63+20];
ld.global.nc.u16 %rs1, [%rd9];

{ cvt.f32.f16 %f101, %rs1;}


and.b32 %r116, %r107, 252645135;
shl.b32 %r117, %r109, 4;
and.b32 %r118, %r117, 16;
or.b32 %r119, %r118, %r116;
shl.b32 %r120, %r109, 11;
and.b32 %r121, %r120, 4096;
or.b32 %r122, %r119, %r121;
shl.b32 %r123, %r109, 18;
and.b32 %r124, %r123, 1048576;
or.b32 %r125, %r122, %r124;
shl.b32 %r126, %r109, 25;
and.b32 %r127, %r126, 268435456;
or.b32 %r88, %r125, %r127;
mov.u32 %r90, 0;

dp4a.s32.s32 %r87, %r88, %r18, %r90;

shr.u32 %r128, %r107, 4;
and.b32 %r129, %r128, 252645135;
shr.u32 %r130, %r109, 12;
and.b32 %r131, %r130, 16;
or.b32 %r132, %r131, %r129;
shr.u32 %r133, %r109, 5;
and.b32 %r134, %r133, 4096;
or.b32 %r135, %r132, %r134;
shl.b32 %r136, %r109, 2;
and.b32 %r137, %r136, 1048576;
or.b32 %r138, %r135, %r137;
shl.b32 %r139, %r109, 9;
and.b32 %r140, %r139, 268435456;
or.b32 %r92, %r138, %r140;

dp4a.s32.s32 %r91, %r92, %r19, %r87;

and.b32 %r141, %r111, 252645135;
shl.b32 %r142, %r115, 4;
and.b32 %r143, %r142, 16;
or.b32 %r144, %r143, %r141;
shl.b32 %r145, %r115, 11;
and.b32 %r146, %r145, 4096;
or.b32 %r147, %r144, %r146;
shl.b32 %r148, %r115, 18;
and.b32 %r149, %r148, 1048576;
or.b32 %r150, %r147, %r149;
shl.b32 %r151, %r115, 25;
and.b32 %r152, %r151, 268435456;
or.b32 %r96, %r150, %r152;

dp4a.s32.s32 %r95, %r96, %r20, %r91;

shr.u32 %r153, %r111, 4;
and.b32 %r154, %r153, 252645135;
shr.u32 %r155, %r115, 12;
and.b32 %r156, %r155, 16;
or.b32 %r157, %r156, %r154;
shr.u32 %r158, %r115, 5;
and.b32 %r159, %r158, 4096;
or.b32 %r160, %r157, %r159;
shl.b32 %r161, %r115, 2;
and.b32 %r162, %r161, 1048576;
or.b32 %r163, %r160, %r162;
shl.b32 %r164, %r115, 9;
and.b32 %r165, %r164, 268435456;
or.b32 %r100, %r163, %r165;

dp4a.s32.s32 %r99, %r100, %r21, %r95;

ld.global.nc.u32 %r22, [%rd57];

{.reg .f16 low,high;
mov.b32 {low,high},%r22;
cvt.f32.f16 %f102, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r22;
cvt.f32.f16 %f103, high;}


cvt.rn.f32.s32 %f104, %r99;
mul.ftz.f32 %f105, %f103, 0fC1000000;
fma.rn.ftz.f32 %f106, %f102, %f104, %f105;
fma.rn.ftz.f32 %f29, %f101, %f106, %f29;
setp.eq.s64 %p13, %rd22, 0;
@%p13 bra $L__BB58_12;

add.s64 %rd67, %rd94, %rd55;
ld.global.u16 %rs9, [%rd67+2];
ld.global.u16 %rs10, [%rd67+4];
mov.b32 %r184, {%rs9, %rs10};
add.s64 %rd70, %rd94, %rd60;
ld.global.u16 %rs11, [%rd70+6];
ld.global.u16 %rs12, [%rd70+8];
mov.b32 %r186, {%rs11, %rs12};
shr.s32 %r188, %r184, %r108;
ld.global.u16 %rs13, [%rd70+10];
ld.global.u16 %rs14, [%rd70+12];
mov.b32 %r189, {%rs13, %rs14};
shr.s32 %r194, %r184, %r114;
ld.global.u16 %rs8, [%rd67];

{ cvt.f32.f16 %f107, %rs8;}


and.b32 %r195, %r186, 252645135;
shl.b32 %r196, %r188, 4;
and.b32 %r197, %r196, 16;
or.b32 %r198, %r197, %r195;
shl.b32 %r199, %r188, 11;
and.b32 %r200, %r199, 4096;
or.b32 %r201, %r198, %r200;
shl.b32 %r202, %r188, 18;
and.b32 %r203, %r202, 1048576;
or.b32 %r204, %r201, %r203;
shl.b32 %r205, %r188, 25;
and.b32 %r206, %r205, 268435456;
or.b32 %r167, %r204, %r206;

dp4a.s32.s32 %r166, %r167, %r18, %r90;

shr.u32 %r207, %r186, 4;
and.b32 %r208, %r207, 252645135;
shr.u32 %r209, %r188, 12;
and.b32 %r210, %r209, 16;
or.b32 %r211, %r210, %r208;
shr.u32 %r212, %r188, 5;
and.b32 %r213, %r212, 4096;
or.b32 %r214, %r211, %r213;
shl.b32 %r215, %r188, 2;
and.b32 %r216, %r215, 1048576;
or.b32 %r217, %r214, %r216;
shl.b32 %r218, %r188, 9;
and.b32 %r219, %r218, 268435456;
or.b32 %r171, %r217, %r219;

dp4a.s32.s32 %r170, %r171, %r19, %r166;

and.b32 %r220, %r189, 252645135;
shl.b32 %r221, %r194, 4;
and.b32 %r222, %r221, 16;
or.b32 %r223, %r222, %r220;
shl.b32 %r224, %r194, 11;
and.b32 %r225, %r224, 4096;
or.b32 %r226, %r223, %r225;
shl.b32 %r227, %r194, 18;
and.b32 %r228, %r227, 1048576;
or.b32 %r229, %r226, %r228;
shl.b32 %r230, %r194, 25;
and.b32 %r231, %r230, 268435456;
or.b32 %r175, %r229, %r231;

dp4a.s32.s32 %r174, %r175, %r20, %r170;

shr.u32 %r232, %r189, 4;
and.b32 %r233, %r232, 252645135;
shr.u32 %r234, %r194, 12;
and.b32 %r235, %r234, 16;
or.b32 %r236, %r235, %r233;
shr.u32 %r237, %r194, 5;
and.b32 %r238, %r237, 4096;
or.b32 %r239, %r236, %r238;
shl.b32 %r240, %r194, 2;
and.b32 %r241, %r240, 1048576;
or.b32 %r242, %r239, %r241;
shl.b32 %r243, %r194, 9;
and.b32 %r244, %r243, 268435456;
or.b32 %r179, %r242, %r244;

dp4a.s32.s32 %r178, %r179, %r21, %r174;


{.reg .f16 low,high;
mov.b32 {low,high},%r22;
cvt.f32.f16 %f108, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r22;
cvt.f32.f16 %f109, high;}


cvt.rn.f32.s32 %f110, %r178;
mul.ftz.f32 %f111, %f109, 0fC1000000;
fma.rn.ftz.f32 %f112, %f108, %f110, %f111;
fma.rn.ftz.f32 %f33, %f107, %f112, %f33;

$L__BB58_12:
mul.wide.s32 %rd71, %r44, 22;
add.s64 %rd11, %rd9, %rd71;
ld.global.nc.u16 %rs16, [%rd11+2];
ld.global.nc.u16 %rs17, [%rd11+4];
mov.b32 %r263, {%rs16, %rs17};
add.s64 %rd12, %rd10, %rd71;
ld.global.nc.u16 %rs18, [%rd12];
ld.global.nc.u16 %rs19, [%rd12+2];
mov.b32 %r264, {%rs18, %rs19};
shr.s32 %r266, %r263, %r108;
ld.global.nc.u16 %rs20, [%rd12+4];
ld.global.nc.u16 %rs21, [%rd12+6];
mov.b32 %r267, {%rs20, %rs21};
shr.s32 %r273, %r263, %r114;
ld.global.nc.u16 %rs15, [%rd11];

{ cvt.f32.f16 %f113, %rs15;}


and.b32 %r274, %r264, 252645135;
shl.b32 %r275, %r266, 4;
and.b32 %r276, %r275, 16;
or.b32 %r277, %r276, %r274;
shl.b32 %r278, %r266, 11;
and.b32 %r279, %r278, 4096;
or.b32 %r280, %r277, %r279;
shl.b32 %r281, %r266, 18;
and.b32 %r282, %r281, 1048576;
or.b32 %r283, %r280, %r282;
shl.b32 %r284, %r266, 25;
and.b32 %r285, %r284, 268435456;
or.b32 %r246, %r283, %r285;

dp4a.s32.s32 %r245, %r246, %r18, %r90;

shr.u32 %r286, %r264, 4;
and.b32 %r287, %r286, 252645135;
shr.u32 %r288, %r266, 12;
and.b32 %r289, %r288, 16;
or.b32 %r290, %r289, %r287;
shr.u32 %r291, %r266, 5;
and.b32 %r292, %r291, 4096;
or.b32 %r293, %r290, %r292;
shl.b32 %r294, %r266, 2;
and.b32 %r295, %r294, 1048576;
or.b32 %r296, %r293, %r295;
shl.b32 %r297, %r266, 9;
and.b32 %r298, %r297, 268435456;
or.b32 %r250, %r296, %r298;

dp4a.s32.s32 %r249, %r250, %r19, %r245;

and.b32 %r299, %r267, 252645135;
shl.b32 %r300, %r273, 4;
and.b32 %r301, %r300, 16;
or.b32 %r302, %r301, %r299;
shl.b32 %r303, %r273, 11;
and.b32 %r304, %r303, 4096;
or.b32 %r305, %r302, %r304;
shl.b32 %r306, %r273, 18;
and.b32 %r307, %r306, 1048576;
or.b32 %r308, %r305, %r307;
shl.b32 %r309, %r273, 25;
and.b32 %r310, %r309, 268435456;
or.b32 %r254, %r308, %r310;

dp4a.s32.s32 %r253, %r254, %r20, %r249;

shr.u32 %r311, %r267, 4;
and.b32 %r312, %r311, 252645135;
shr.u32 %r313, %r273, 12;
and.b32 %r314, %r313, 16;
or.b32 %r315, %r314, %r312;
shr.u32 %r316, %r273, 5;
and.b32 %r317, %r316, 4096;
or.b32 %r318, %r315, %r317;
shl.b32 %r319, %r273, 2;
and.b32 %r320, %r319, 1048576;
or.b32 %r321, %r318, %r320;
shl.b32 %r322, %r273, 9;
and.b32 %r323, %r322, 268435456;
or.b32 %r258, %r321, %r323;

dp4a.s32.s32 %r257, %r258, %r21, %r253;


{.reg .f16 low,high;
mov.b32 {low,high},%r22;
cvt.f32.f16 %f114, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r22;
cvt.f32.f16 %f115, high;}


cvt.rn.f32.s32 %f116, %r257;
mul.ftz.f32 %f117, %f115, 0fC1000000;
fma.rn.ftz.f32 %f118, %f114, %f116, %f117;
fma.rn.ftz.f32 %f28, %f113, %f118, %f28;
@%p13 bra $L__BB58_14;

or.b32 %r344, %r4, 1;
mad.lo.s32 %r345, %r44, %r344, %r16;
mul.wide.s32 %rd72, %r345, 22;
add.s64 %rd73, %rd94, %rd72;
ld.global.u16 %rs23, [%rd73+2];
ld.global.u16 %rs24, [%rd73+4];
mov.b32 %r346, {%rs23, %rs24};
add.s64 %rd75, %rd72, %rd59;
add.s64 %rd76, %rd94, %rd75;
ld.global.u16 %rs25, [%rd76+6];
ld.global.u16 %rs26, [%rd76+8];
mov.b32 %r348, {%rs25, %rs26};
shr.s32 %r350, %r346, %r108;
ld.global.u16 %rs27, [%rd76+10];
ld.global.u16 %rs28, [%rd76+12];
mov.b32 %r351, {%rs27, %rs28};
shr.s32 %r356, %r346, %r114;
ld.global.u16 %rs22, [%rd73];

{ cvt.f32.f16 %f119, %rs22;}


and.b32 %r357, %r348, 252645135;
shl.b32 %r358, %r350, 4;
and.b32 %r359, %r358, 16;
or.b32 %r360, %r359, %r357;
shl.b32 %r361, %r350, 11;
and.b32 %r362, %r361, 4096;
or.b32 %r363, %r360, %r362;
shl.b32 %r364, %r350, 18;
and.b32 %r365, %r364, 1048576;
or.b32 %r366, %r363, %r365;
shl.b32 %r367, %r350, 25;
and.b32 %r368, %r367, 268435456;
or.b32 %r325, %r366, %r368;
mov.u32 %r327, 0;

dp4a.s32.s32 %r324, %r325, %r18, %r327;

shr.u32 %r369, %r348, 4;
and.b32 %r370, %r369, 252645135;
shr.u32 %r371, %r350, 12;
and.b32 %r372, %r371, 16;
or.b32 %r373, %r372, %r370;
shr.u32 %r374, %r350, 5;
and.b32 %r375, %r374, 4096;
or.b32 %r376, %r373, %r375;
shl.b32 %r377, %r350, 2;
and.b32 %r378, %r377, 1048576;
or.b32 %r379, %r376, %r378;
shl.b32 %r380, %r350, 9;
and.b32 %r381, %r380, 268435456;
or.b32 %r329, %r379, %r381;

dp4a.s32.s32 %r328, %r329, %r19, %r324;

and.b32 %r382, %r351, 252645135;
shl.b32 %r383, %r356, 4;
and.b32 %r384, %r383, 16;
or.b32 %r385, %r384, %r382;
shl.b32 %r386, %r356, 11;
and.b32 %r387, %r386, 4096;
or.b32 %r388, %r385, %r387;
shl.b32 %r389, %r356, 18;
and.b32 %r390, %r389, 1048576;
or.b32 %r391, %r388, %r390;
shl.b32 %r392, %r356, 25;
and.b32 %r393, %r392, 268435456;
or.b32 %r333, %r391, %r393;

dp4a.s32.s32 %r332, %r333, %r20, %r328;

shr.u32 %r394, %r351, 4;
and.b32 %r395, %r394, 252645135;
shr.u32 %r396, %r356, 12;
and.b32 %r397, %r396, 16;
or.b32 %r398, %r397, %r395;
shr.u32 %r399, %r356, 5;
and.b32 %r400, %r399, 4096;
or.b32 %r401, %r398, %r400;
shl.b32 %r402, %r356, 2;
and.b32 %r403, %r402, 1048576;
or.b32 %r404, %r401, %r403;
shl.b32 %r405, %r356, 9;
and.b32 %r406, %r405, 268435456;
or.b32 %r337, %r404, %r406;

dp4a.s32.s32 %r336, %r337, %r21, %r332;


{.reg .f16 low,high;
mov.b32 {low,high},%r22;
cvt.f32.f16 %f120, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r22;
cvt.f32.f16 %f121, high;}


cvt.rn.f32.s32 %f122, %r336;
mul.ftz.f32 %f123, %f121, 0fC1000000;
fma.rn.ftz.f32 %f124, %f120, %f122, %f123;
fma.rn.ftz.f32 %f32, %f119, %f124, %f32;

$L__BB58_14:
add.s64 %rd13, %rd11, %rd71;
ld.global.nc.u16 %rs30, [%rd13+2];
ld.global.nc.u16 %rs31, [%rd13+4];
mov.b32 %r425, {%rs30, %rs31};
add.s64 %rd14, %rd12, %rd71;
ld.global.nc.u16 %rs32, [%rd14];
ld.global.nc.u16 %rs33, [%rd14+2];
mov.b32 %r426, {%rs32, %rs33};
shr.s32 %r428, %r425, %r108;
ld.global.nc.u16 %rs34, [%rd14+4];
ld.global.nc.u16 %rs35, [%rd14+6];
mov.b32 %r429, {%rs34, %rs35};
shr.s32 %r435, %r425, %r114;
ld.global.nc.u16 %rs29, [%rd13];

{ cvt.f32.f16 %f125, %rs29;}


and.b32 %r436, %r426, 252645135;
shl.b32 %r437, %r428, 4;
and.b32 %r438, %r437, 16;
or.b32 %r439, %r438, %r436;
shl.b32 %r440, %r428, 11;
and.b32 %r441, %r440, 4096;
or.b32 %r442, %r439, %r441;
shl.b32 %r443, %r428, 18;
and.b32 %r444, %r443, 1048576;
or.b32 %r445, %r442, %r444;
shl.b32 %r446, %r428, 25;
and.b32 %r447, %r446, 268435456;
or.b32 %r408, %r445, %r447;
mov.u32 %r410, 0;

dp4a.s32.s32 %r407, %r408, %r18, %r410;

shr.u32 %r448, %r426, 4;
and.b32 %r449, %r448, 252645135;
shr.u32 %r450, %r428, 12;
and.b32 %r451, %r450, 16;
or.b32 %r452, %r451, %r449;
shr.u32 %r453, %r428, 5;
and.b32 %r454, %r453, 4096;
or.b32 %r455, %r452, %r454;
shl.b32 %r456, %r428, 2;
and.b32 %r457, %r456, 1048576;
or.b32 %r458, %r455, %r457;
shl.b32 %r459, %r428, 9;
and.b32 %r460, %r459, 268435456;
or.b32 %r412, %r458, %r460;

dp4a.s32.s32 %r411, %r412, %r19, %r407;

and.b32 %r461, %r429, 252645135;
shl.b32 %r462, %r435, 4;
and.b32 %r463, %r462, 16;
or.b32 %r464, %r463, %r461;
shl.b32 %r465, %r435, 11;
and.b32 %r466, %r465, 4096;
or.b32 %r467, %r464, %r466;
shl.b32 %r468, %r435, 18;
and.b32 %r469, %r468, 1048576;
or.b32 %r470, %r467, %r469;
shl.b32 %r471, %r435, 25;
and.b32 %r472, %r471, 268435456;
or.b32 %r416, %r470, %r472;

dp4a.s32.s32 %r415, %r416, %r20, %r411;

shr.u32 %r473, %r429, 4;
and.b32 %r474, %r473, 252645135;
shr.u32 %r475, %r435, 12;
and.b32 %r476, %r475, 16;
or.b32 %r477, %r476, %r474;
shr.u32 %r478, %r435, 5;
and.b32 %r479, %r478, 4096;
or.b32 %r480, %r477, %r479;
shl.b32 %r481, %r435, 2;
and.b32 %r482, %r481, 1048576;
or.b32 %r483, %r480, %r482;
shl.b32 %r484, %r435, 9;
and.b32 %r485, %r484, 268435456;
or.b32 %r420, %r483, %r485;

dp4a.s32.s32 %r419, %r420, %r21, %r415;


{.reg .f16 low,high;
mov.b32 {low,high},%r22;
cvt.f32.f16 %f126, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r22;
cvt.f32.f16 %f127, high;}


cvt.rn.f32.s32 %f128, %r419;
mul.ftz.f32 %f129, %f127, 0fC1000000;
fma.rn.ftz.f32 %f130, %f126, %f128, %f129;
fma.rn.ftz.f32 %f27, %f125, %f130, %f27;
@%p13 bra $L__BB58_16;

or.b32 %r506, %r4, 2;
mad.lo.s32 %r507, %r44, %r506, %r16;
mul.wide.s32 %rd78, %r507, 22;
add.s64 %rd79, %rd94, %rd78;
ld.global.u16 %rs37, [%rd79+2];
ld.global.u16 %rs38, [%rd79+4];
mov.b32 %r508, {%rs37, %rs38};
add.s64 %rd81, %rd78, %rd59;
add.s64 %rd82, %rd94, %rd81;
ld.global.u16 %rs39, [%rd82+6];
ld.global.u16 %rs40, [%rd82+8];
mov.b32 %r510, {%rs39, %rs40};
shr.s32 %r512, %r508, %r108;
ld.global.u16 %rs41, [%rd82+10];
ld.global.u16 %rs42, [%rd82+12];
mov.b32 %r513, {%rs41, %rs42};
shr.s32 %r518, %r508, %r114;
ld.global.u16 %rs36, [%rd79];

{ cvt.f32.f16 %f131, %rs36;}


and.b32 %r519, %r510, 252645135;
shl.b32 %r520, %r512, 4;
and.b32 %r521, %r520, 16;
or.b32 %r522, %r521, %r519;
shl.b32 %r523, %r512, 11;
and.b32 %r524, %r523, 4096;
or.b32 %r525, %r522, %r524;
shl.b32 %r526, %r512, 18;
and.b32 %r527, %r526, 1048576;
or.b32 %r528, %r525, %r527;
shl.b32 %r529, %r512, 25;
and.b32 %r530, %r529, 268435456;
or.b32 %r487, %r528, %r530;

dp4a.s32.s32 %r486, %r487, %r18, %r410;

shr.u32 %r531, %r510, 4;
and.b32 %r532, %r531, 252645135;
shr.u32 %r533, %r512, 12;
and.b32 %r534, %r533, 16;
or.b32 %r535, %r534, %r532;
shr.u32 %r536, %r512, 5;
and.b32 %r537, %r536, 4096;
or.b32 %r538, %r535, %r537;
shl.b32 %r539, %r512, 2;
and.b32 %r540, %r539, 1048576;
or.b32 %r541, %r538, %r540;
shl.b32 %r542, %r512, 9;
and.b32 %r543, %r542, 268435456;
or.b32 %r491, %r541, %r543;

dp4a.s32.s32 %r490, %r491, %r19, %r486;

and.b32 %r544, %r513, 252645135;
shl.b32 %r545, %r518, 4;
and.b32 %r546, %r545, 16;
or.b32 %r547, %r546, %r544;
shl.b32 %r548, %r518, 11;
and.b32 %r549, %r548, 4096;
or.b32 %r550, %r547, %r549;
shl.b32 %r551, %r518, 18;
and.b32 %r552, %r551, 1048576;
or.b32 %r553, %r550, %r552;
shl.b32 %r554, %r518, 25;
and.b32 %r555, %r554, 268435456;
or.b32 %r495, %r553, %r555;

dp4a.s32.s32 %r494, %r495, %r20, %r490;

shr.u32 %r556, %r513, 4;
and.b32 %r557, %r556, 252645135;
shr.u32 %r558, %r518, 12;
and.b32 %r559, %r558, 16;
or.b32 %r560, %r559, %r557;
shr.u32 %r561, %r518, 5;
and.b32 %r562, %r561, 4096;
or.b32 %r563, %r560, %r562;
shl.b32 %r564, %r518, 2;
and.b32 %r565, %r564, 1048576;
or.b32 %r566, %r563, %r565;
shl.b32 %r567, %r518, 9;
and.b32 %r568, %r567, 268435456;
or.b32 %r499, %r566, %r568;

dp4a.s32.s32 %r498, %r499, %r21, %r494;


{.reg .f16 low,high;
mov.b32 {low,high},%r22;
cvt.f32.f16 %f132, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r22;
cvt.f32.f16 %f133, high;}


cvt.rn.f32.s32 %f134, %r498;
mul.ftz.f32 %f135, %f133, 0fC1000000;
fma.rn.ftz.f32 %f136, %f132, %f134, %f135;
fma.rn.ftz.f32 %f31, %f131, %f136, %f31;

$L__BB58_16:
add.s64 %rd84, %rd13, %rd71;
ld.global.nc.u16 %rs44, [%rd84+2];
ld.global.nc.u16 %rs45, [%rd84+4];
mov.b32 %r587, {%rs44, %rs45};
add.s64 %rd85, %rd14, %rd71;
ld.global.nc.u16 %rs46, [%rd85];
ld.global.nc.u16 %rs47, [%rd85+2];
mov.b32 %r588, {%rs46, %rs47};
shr.s32 %r590, %r587, %r108;
ld.global.nc.u16 %rs48, [%rd85+4];
ld.global.nc.u16 %rs49, [%rd85+6];
mov.b32 %r591, {%rs48, %rs49};
shr.s32 %r597, %r587, %r114;
ld.global.nc.u16 %rs43, [%rd84];

{ cvt.f32.f16 %f137, %rs43;}


and.b32 %r598, %r588, 252645135;
shl.b32 %r599, %r590, 4;
and.b32 %r600, %r599, 16;
or.b32 %r601, %r600, %r598;
shl.b32 %r602, %r590, 11;
and.b32 %r603, %r602, 4096;
or.b32 %r604, %r601, %r603;
shl.b32 %r605, %r590, 18;
and.b32 %r606, %r605, 1048576;
or.b32 %r607, %r604, %r606;
shl.b32 %r608, %r590, 25;
and.b32 %r609, %r608, 268435456;
or.b32 %r570, %r607, %r609;

dp4a.s32.s32 %r569, %r570, %r18, %r410;

shr.u32 %r610, %r588, 4;
and.b32 %r611, %r610, 252645135;
shr.u32 %r612, %r590, 12;
and.b32 %r613, %r612, 16;
or.b32 %r614, %r613, %r611;
shr.u32 %r615, %r590, 5;
and.b32 %r616, %r615, 4096;
or.b32 %r617, %r614, %r616;
shl.b32 %r618, %r590, 2;
and.b32 %r619, %r618, 1048576;
or.b32 %r620, %r617, %r619;
shl.b32 %r621, %r590, 9;
and.b32 %r622, %r621, 268435456;
or.b32 %r574, %r620, %r622;

dp4a.s32.s32 %r573, %r574, %r19, %r569;

and.b32 %r623, %r591, 252645135;
shl.b32 %r624, %r597, 4;
and.b32 %r625, %r624, 16;
or.b32 %r626, %r625, %r623;
shl.b32 %r627, %r597, 11;
and.b32 %r628, %r627, 4096;
or.b32 %r629, %r626, %r628;
shl.b32 %r630, %r597, 18;
and.b32 %r631, %r630, 1048576;
or.b32 %r632, %r629, %r631;
shl.b32 %r633, %r597, 25;
and.b32 %r634, %r633, 268435456;
or.b32 %r578, %r632, %r634;

dp4a.s32.s32 %r577, %r578, %r20, %r573;

shr.u32 %r635, %r591, 4;
and.b32 %r636, %r635, 252645135;
shr.u32 %r637, %r597, 12;
and.b32 %r638, %r637, 16;
or.b32 %r639, %r638, %r636;
shr.u32 %r640, %r597, 5;
and.b32 %r641, %r640, 4096;
or.b32 %r642, %r639, %r641;
shl.b32 %r643, %r597, 2;
and.b32 %r644, %r643, 1048576;
or.b32 %r645, %r642, %r644;
shl.b32 %r646, %r597, 9;
and.b32 %r647, %r646, 268435456;
or.b32 %r582, %r645, %r647;

dp4a.s32.s32 %r581, %r582, %r21, %r577;


{.reg .f16 low,high;
mov.b32 {low,high},%r22;
cvt.f32.f16 %f138, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r22;
cvt.f32.f16 %f139, high;}


cvt.rn.f32.s32 %f140, %r581;
mul.ftz.f32 %f141, %f139, 0fC1000000;
fma.rn.ftz.f32 %f142, %f138, %f140, %f141;
fma.rn.ftz.f32 %f26, %f137, %f142, %f26;
@%p13 bra $L__BB58_18;

or.b32 %r668, %r4, 3;
mad.lo.s32 %r669, %r44, %r668, %r16;
mul.wide.s32 %rd86, %r669, 22;
add.s64 %rd87, %rd94, %rd86;
ld.global.u16 %rs51, [%rd87+2];
ld.global.u16 %rs52, [%rd87+4];
mov.b32 %r670, {%rs51, %rs52};
add.s64 %rd89, %rd86, %rd59;
add.s64 %rd90, %rd94, %rd89;
ld.global.u16 %rs53, [%rd90+6];
ld.global.u16 %rs54, [%rd90+8];
mov.b32 %r672, {%rs53, %rs54};
shr.s32 %r674, %r670, %r108;
ld.global.u16 %rs55, [%rd90+10];
ld.global.u16 %rs56, [%rd90+12];
mov.b32 %r675, {%rs55, %rs56};
shr.s32 %r680, %r670, %r114;
ld.global.u16 %rs50, [%rd87];

{ cvt.f32.f16 %f143, %rs50;}


and.b32 %r681, %r672, 252645135;
shl.b32 %r682, %r674, 4;
and.b32 %r683, %r682, 16;
or.b32 %r684, %r683, %r681;
shl.b32 %r685, %r674, 11;
and.b32 %r686, %r685, 4096;
or.b32 %r687, %r684, %r686;
shl.b32 %r688, %r674, 18;
and.b32 %r689, %r688, 1048576;
or.b32 %r690, %r687, %r689;
shl.b32 %r691, %r674, 25;
and.b32 %r692, %r691, 268435456;
or.b32 %r649, %r690, %r692;
mov.u32 %r651, 0;

dp4a.s32.s32 %r648, %r649, %r18, %r651;

shr.u32 %r693, %r672, 4;
and.b32 %r694, %r693, 252645135;
shr.u32 %r695, %r674, 12;
and.b32 %r696, %r695, 16;
or.b32 %r697, %r696, %r694;
shr.u32 %r698, %r674, 5;
and.b32 %r699, %r698, 4096;
or.b32 %r700, %r697, %r699;
shl.b32 %r701, %r674, 2;
and.b32 %r702, %r701, 1048576;
or.b32 %r703, %r700, %r702;
shl.b32 %r704, %r674, 9;
and.b32 %r705, %r704, 268435456;
or.b32 %r653, %r703, %r705;

dp4a.s32.s32 %r652, %r653, %r19, %r648;

and.b32 %r706, %r675, 252645135;
shl.b32 %r707, %r680, 4;
and.b32 %r708, %r707, 16;
or.b32 %r709, %r708, %r706;
shl.b32 %r710, %r680, 11;
and.b32 %r711, %r710, 4096;
or.b32 %r712, %r709, %r711;
shl.b32 %r713, %r680, 18;
and.b32 %r714, %r713, 1048576;
or.b32 %r715, %r712, %r714;
shl.b32 %r716, %r680, 25;
and.b32 %r717, %r716, 268435456;
or.b32 %r657, %r715, %r717;

dp4a.s32.s32 %r656, %r657, %r20, %r652;

shr.u32 %r718, %r675, 4;
and.b32 %r719, %r718, 252645135;
shr.u32 %r720, %r680, 12;
and.b32 %r721, %r720, 16;
or.b32 %r722, %r721, %r719;
shr.u32 %r723, %r680, 5;
and.b32 %r724, %r723, 4096;
or.b32 %r725, %r722, %r724;
shl.b32 %r726, %r680, 2;
and.b32 %r727, %r726, 1048576;
or.b32 %r728, %r725, %r727;
shl.b32 %r729, %r680, 9;
and.b32 %r730, %r729, 268435456;
or.b32 %r661, %r728, %r730;

dp4a.s32.s32 %r660, %r661, %r21, %r656;


{.reg .f16 low,high;
mov.b32 {low,high},%r22;
cvt.f32.f16 %f144, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r22;
cvt.f32.f16 %f145, high;}


cvt.rn.f32.s32 %f146, %r660;
mul.ftz.f32 %f147, %f145, 0fC1000000;
fma.rn.ftz.f32 %f148, %f144, %f146, %f147;
fma.rn.ftz.f32 %f30, %f143, %f148, %f30;

$L__BB58_18:
add.s64 %rd95, %rd95, 1408;
add.s64 %rd94, %rd94, 1408;
add.s32 %r1477, %r1477, 64;
setp.lt.s32 %p17, %r1477, %r73;
@%p17 bra $L__BB58_10;

$L__BB58_19:
mov.u32 %r1464, %tid.y;
setp.eq.s32 %p18, %r1464, 0;
@%p18 bra $L__BB58_23;

mov.u32 %r1465, %tid.y;
setp.ne.s64 %p19, %rd22, 0;
add.s32 %r24, %r1465, -1;
shl.b32 %r735, %r1465, 9;
mov.u32 %r736, _ZZ13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r737, %r736, %r735;
shl.b32 %r738, %r2, 2;
add.s32 %r739, %r737, %r738;
st.shared.f32 [%r739+-512], %f29;
@%p19 bra $L__BB58_22;
bra.uni $L__BB58_21;

$L__BB58_22:
shl.b32 %r741, %r24, 9;
mov.u32 %r742, _ZZ13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r743, %r742, %r741;
add.s32 %r745, %r743, %r738;
st.shared.f32 [%r745], %f33;
add.s32 %r1461, %r739, -512;
st.shared.f32 [%r1461+128], %f28;
st.shared.f32 [%r745+128], %f32;
add.s32 %r1462, %r739, -512;
st.shared.f32 [%r1462+256], %f27;
st.shared.f32 [%r745+256], %f31;
add.s32 %r1463, %r739, -512;
st.shared.f32 [%r1463+384], %f26;
st.shared.f32 [%r745+384], %f30;
bra.uni $L__BB58_23;

$L__BB58_21:
add.s32 %r1458, %r739, -512;
st.shared.f32 [%r1458+128], %f28;
add.s32 %r1459, %r739, -512;
st.shared.f32 [%r1459+256], %f27;
add.s32 %r1460, %r739, -512;
st.shared.f32 [%r1460+384], %f26;

$L__BB58_23:
mov.u32 %r1466, %tid.y;
setp.ne.s32 %p319, %r1466, 0;
bar.sync 0;
@%p319 bra $L__BB58_78;
bra.uni $L__BB58_24;

$L__BB58_78:
ret;

$L__BB58_24:
ld.param.u64 %rd93, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.f32 %f993, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
ld.param.u32 %r1469, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r1468, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
mov.u32 %r1467, %ctaid.z;
setp.ne.s64 %p21, %rd22, 0;
mov.u32 %r748, %ctaid.y;
mad.lo.s32 %r751, %r748, %r1469, %r4;
mad.lo.s32 %r752, %r1467, %r1468, %r751;
neg.ftz.f32 %f35, %f993;
shl.b32 %r754, %r2, 2;
mov.u32 %r755, _ZZ13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r26, %r755, %r754;
cvta.to.global.u64 %rd91, %rd93;
mul.wide.u32 %rd92, %r752, 4;
add.s64 %rd17, %rd91, %rd92;
@%p21 bra $L__BB58_33;
bra.uni $L__BB58_25;

$L__BB58_33:
ld.param.u32 %r1470, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
mov.u32 %r842, _ZZ13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r28, %r842, %r754;
setp.gt.s32 %p54, %r1470, 2;
@%p54 bra $L__BB58_44;

ld.param.u32 %r1473, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p57, %r1473, 1;
@%p57 bra $L__BB58_62;

ld.param.u32 %r1474, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p58, %r1474, 2;
@%p58 bra $L__BB58_36;
bra.uni $L__BB58_70;

$L__BB58_36:
ld.shared.f32 %f549, [%r26];
add.ftz.f32 %f550, %f549, %f29;
ld.shared.f32 %f551, [%r28];
add.ftz.f32 %f552, %f551, %f33;
ld.shared.f32 %f553, [%r26+512];
add.ftz.f32 %f554, %f553, %f550;
ld.shared.f32 %f555, [%r28+512];
add.ftz.f32 %f556, %f555, %f552;
ld.shared.f32 %f557, [%r26+1024];
add.ftz.f32 %f558, %f557, %f554;
ld.shared.f32 %f559, [%r28+1024];
add.ftz.f32 %f560, %f559, %f556;
mov.b32 %r1089, %f558;
mov.u32 %r1090, 31;
mov.u32 %r1091, 16;
mov.u32 %r1092, -1;
shfl.sync.bfly.b32 %r1093|%p163, %r1089, %r1091, %r1090, %r1092;
mov.b32 %f561, %r1093;
add.ftz.f32 %f562, %f558, %f561;
mov.b32 %r1094, %f562;
mov.u32 %r1095, 8;
shfl.sync.bfly.b32 %r1096|%p164, %r1094, %r1095, %r1090, %r1092;
mov.b32 %f563, %r1096;
add.ftz.f32 %f564, %f562, %f563;
mov.b32 %r1097, %f564;
mov.u32 %r1098, 4;
shfl.sync.bfly.b32 %r1099|%p165, %r1097, %r1098, %r1090, %r1092;
mov.b32 %f565, %r1099;
add.ftz.f32 %f566, %f564, %f565;
mov.b32 %r1100, %f566;
mov.u32 %r1101, 2;
shfl.sync.bfly.b32 %r1102|%p166, %r1100, %r1101, %r1090, %r1092;
mov.b32 %f567, %r1102;
add.ftz.f32 %f568, %f566, %f567;
mov.b32 %r1103, %f568;
mov.u32 %r1104, 1;
shfl.sync.bfly.b32 %r1105|%p167, %r1103, %r1104, %r1090, %r1092;
mov.b32 %f569, %r1105;
add.ftz.f32 %f56, %f568, %f569;
mov.b32 %r1106, %f560;
shfl.sync.bfly.b32 %r1107|%p168, %r1106, %r1091, %r1090, %r1092;
mov.b32 %f570, %r1107;
add.ftz.f32 %f571, %f560, %f570;
mov.b32 %r1108, %f571;
shfl.sync.bfly.b32 %r1109|%p169, %r1108, %r1095, %r1090, %r1092;
mov.b32 %f572, %r1109;
add.ftz.f32 %f573, %f571, %f572;
mov.b32 %r1110, %f573;
shfl.sync.bfly.b32 %r1111|%p170, %r1110, %r1098, %r1090, %r1092;
mov.b32 %f574, %r1111;
add.ftz.f32 %f575, %f573, %f574;
mov.b32 %r1112, %f575;
shfl.sync.bfly.b32 %r1113|%p171, %r1112, %r1101, %r1090, %r1092;
mov.b32 %f576, %r1113;
add.ftz.f32 %f577, %f575, %f576;
mov.b32 %r1114, %f577;
shfl.sync.bfly.b32 %r1115|%p172, %r1114, %r1104, %r1090, %r1092;
mov.b32 %f578, %r1115;
add.ftz.f32 %f57, %f577, %f578;
setp.ne.s32 %p173, %r2, 0;
setp.ge.u32 %p174, %r4, %r45;
or.pred %p175, %p173, %p174;
@%p175 bra $L__BB58_38;

add.ftz.f32 %f579, %f5, %f56;
add.ftz.f32 %f580, %f4, %f57;
mul.ftz.f32 %f581, %f580, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f582, %f581;
add.ftz.f32 %f583, %f582, 0f3F800000;
div.approx.ftz.f32 %f584, %f580, %f583;
mul.ftz.f32 %f585, %f579, %f584;
st.global.f32 [%rd17], %f585;

$L__BB58_38:
ld.shared.f32 %f586, [%r26+128];
add.ftz.f32 %f587, %f586, %f28;
ld.shared.f32 %f588, [%r28+128];
add.ftz.f32 %f589, %f588, %f32;
ld.shared.f32 %f590, [%r26+640];
add.ftz.f32 %f591, %f590, %f587;
ld.shared.f32 %f592, [%r28+640];
add.ftz.f32 %f593, %f592, %f589;
ld.shared.f32 %f594, [%r26+1152];
add.ftz.f32 %f595, %f594, %f591;
ld.shared.f32 %f596, [%r28+1152];
add.ftz.f32 %f597, %f596, %f593;
mov.b32 %r1119, %f595;
shfl.sync.bfly.b32 %r1123|%p176, %r1119, %r1091, %r1090, %r1092;
mov.b32 %f598, %r1123;
add.ftz.f32 %f599, %f595, %f598;
mov.b32 %r1124, %f599;
shfl.sync.bfly.b32 %r1126|%p177, %r1124, %r1095, %r1090, %r1092;
mov.b32 %f600, %r1126;
add.ftz.f32 %f601, %f599, %f600;
mov.b32 %r1127, %f601;
shfl.sync.bfly.b32 %r1129|%p178, %r1127, %r1098, %r1090, %r1092;
mov.b32 %f602, %r1129;
add.ftz.f32 %f603, %f601, %f602;
mov.b32 %r1130, %f603;
shfl.sync.bfly.b32 %r1132|%p179, %r1130, %r1101, %r1090, %r1092;
mov.b32 %f604, %r1132;
add.ftz.f32 %f605, %f603, %f604;
mov.b32 %r1133, %f605;
shfl.sync.bfly.b32 %r1135|%p180, %r1133, %r1104, %r1090, %r1092;
mov.b32 %f606, %r1135;
add.ftz.f32 %f58, %f605, %f606;
mov.b32 %r1136, %f597;
shfl.sync.bfly.b32 %r1137|%p181, %r1136, %r1091, %r1090, %r1092;
mov.b32 %f607, %r1137;
add.ftz.f32 %f608, %f597, %f607;
mov.b32 %r1138, %f608;
shfl.sync.bfly.b32 %r1139|%p182, %r1138, %r1095, %r1090, %r1092;
mov.b32 %f609, %r1139;
add.ftz.f32 %f610, %f608, %f609;
mov.b32 %r1140, %f610;
shfl.sync.bfly.b32 %r1141|%p183, %r1140, %r1098, %r1090, %r1092;
mov.b32 %f611, %r1141;
add.ftz.f32 %f612, %f610, %f611;
mov.b32 %r1142, %f612;
shfl.sync.bfly.b32 %r1143|%p184, %r1142, %r1101, %r1090, %r1092;
mov.b32 %f613, %r1143;
add.ftz.f32 %f614, %f612, %f613;
mov.b32 %r1144, %f614;
shfl.sync.bfly.b32 %r1145|%p185, %r1144, %r1104, %r1090, %r1092;
mov.b32 %f615, %r1145;
add.ftz.f32 %f59, %f614, %f615;
setp.ne.s32 %p186, %r2, 1;
or.b32 %r1149, %r4, 1;
setp.ge.u32 %p187, %r1149, %r45;
or.pred %p188, %p186, %p187;
@%p188 bra $L__BB58_40;

add.ftz.f32 %f616, %f5, %f58;
add.ftz.f32 %f617, %f4, %f59;
mul.ftz.f32 %f618, %f617, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f619, %f618;
add.ftz.f32 %f620, %f619, 0f3F800000;
div.approx.ftz.f32 %f621, %f617, %f620;
mul.ftz.f32 %f622, %f616, %f621;
st.global.f32 [%rd17+4], %f622;

$L__BB58_40:
ld.shared.f32 %f623, [%r26+256];
add.ftz.f32 %f624, %f623, %f27;
ld.shared.f32 %f625, [%r28+256];
add.ftz.f32 %f626, %f625, %f31;
ld.shared.f32 %f627, [%r26+768];
add.ftz.f32 %f628, %f627, %f624;
ld.shared.f32 %f629, [%r28+768];
add.ftz.f32 %f630, %f629, %f626;
ld.shared.f32 %f631, [%r26+1280];
add.ftz.f32 %f632, %f631, %f628;
ld.shared.f32 %f633, [%r28+1280];
add.ftz.f32 %f634, %f633, %f630;
mov.b32 %r1150, %f632;
mov.u32 %r1151, 31;
mov.u32 %r1152, 16;
mov.u32 %r1153, -1;
shfl.sync.bfly.b32 %r1154|%p189, %r1150, %r1152, %r1151, %r1153;
mov.b32 %f635, %r1154;
add.ftz.f32 %f636, %f632, %f635;
mov.b32 %r1155, %f636;
mov.u32 %r1156, 8;
shfl.sync.bfly.b32 %r1157|%p190, %r1155, %r1156, %r1151, %r1153;
mov.b32 %f637, %r1157;
add.ftz.f32 %f638, %f636, %f637;
mov.b32 %r1158, %f638;
mov.u32 %r1159, 4;
shfl.sync.bfly.b32 %r1160|%p191, %r1158, %r1159, %r1151, %r1153;
mov.b32 %f639, %r1160;
add.ftz.f32 %f640, %f638, %f639;
mov.b32 %r1161, %f640;
mov.u32 %r1162, 2;
shfl.sync.bfly.b32 %r1163|%p192, %r1161, %r1162, %r1151, %r1153;
mov.b32 %f641, %r1163;
add.ftz.f32 %f642, %f640, %f641;
mov.b32 %r1164, %f642;
mov.u32 %r1165, 1;
shfl.sync.bfly.b32 %r1166|%p193, %r1164, %r1165, %r1151, %r1153;
mov.b32 %f643, %r1166;
add.ftz.f32 %f60, %f642, %f643;
mov.b32 %r1167, %f634;
shfl.sync.bfly.b32 %r1168|%p194, %r1167, %r1152, %r1151, %r1153;
mov.b32 %f644, %r1168;
add.ftz.f32 %f645, %f634, %f644;
mov.b32 %r1169, %f645;
shfl.sync.bfly.b32 %r1170|%p195, %r1169, %r1156, %r1151, %r1153;
mov.b32 %f646, %r1170;
add.ftz.f32 %f647, %f645, %f646;
mov.b32 %r1171, %f647;
shfl.sync.bfly.b32 %r1172|%p196, %r1171, %r1159, %r1151, %r1153;
mov.b32 %f648, %r1172;
add.ftz.f32 %f649, %f647, %f648;
mov.b32 %r1173, %f649;
shfl.sync.bfly.b32 %r1174|%p197, %r1173, %r1162, %r1151, %r1153;
mov.b32 %f650, %r1174;
add.ftz.f32 %f651, %f649, %f650;
mov.b32 %r1175, %f651;
shfl.sync.bfly.b32 %r1176|%p198, %r1175, %r1165, %r1151, %r1153;
mov.b32 %f652, %r1176;
add.ftz.f32 %f61, %f651, %f652;
setp.ne.s32 %p199, %r2, 2;
or.b32 %r1180, %r4, 2;
setp.ge.u32 %p200, %r1180, %r45;
or.pred %p201, %p199, %p200;
@%p201 bra $L__BB58_42;

add.ftz.f32 %f653, %f5, %f60;
add.ftz.f32 %f654, %f4, %f61;
mul.ftz.f32 %f655, %f654, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f656, %f655;
add.ftz.f32 %f657, %f656, 0f3F800000;
div.approx.ftz.f32 %f658, %f654, %f657;
mul.ftz.f32 %f659, %f653, %f658;
st.global.f32 [%rd17+8], %f659;

$L__BB58_42:
ld.shared.f32 %f660, [%r26+384];
add.ftz.f32 %f661, %f660, %f26;
ld.shared.f32 %f662, [%r28+384];
add.ftz.f32 %f663, %f662, %f30;
ld.shared.f32 %f664, [%r26+896];
add.ftz.f32 %f665, %f664, %f661;
ld.shared.f32 %f666, [%r28+896];
add.ftz.f32 %f667, %f666, %f663;
ld.shared.f32 %f668, [%r26+1408];
add.ftz.f32 %f669, %f668, %f665;
ld.shared.f32 %f670, [%r28+1408];
add.ftz.f32 %f671, %f670, %f667;
mov.b32 %r1181, %f669;
shfl.sync.bfly.b32 %r1185|%p202, %r1181, %r1152, %r1151, %r1153;
mov.b32 %f672, %r1185;
add.ftz.f32 %f673, %f669, %f672;
mov.b32 %r1186, %f673;
shfl.sync.bfly.b32 %r1188|%p203, %r1186, %r1156, %r1151, %r1153;
mov.b32 %f674, %r1188;
add.ftz.f32 %f675, %f673, %f674;
mov.b32 %r1189, %f675;
shfl.sync.bfly.b32 %r1191|%p204, %r1189, %r1159, %r1151, %r1153;
mov.b32 %f676, %r1191;
add.ftz.f32 %f677, %f675, %f676;
mov.b32 %r1192, %f677;
shfl.sync.bfly.b32 %r1194|%p205, %r1192, %r1162, %r1151, %r1153;
mov.b32 %f678, %r1194;
add.ftz.f32 %f679, %f677, %f678;
mov.b32 %r1195, %f679;
shfl.sync.bfly.b32 %r1197|%p206, %r1195, %r1165, %r1151, %r1153;
mov.b32 %f680, %r1197;
add.ftz.f32 %f62, %f679, %f680;
mov.b32 %r1198, %f671;
shfl.sync.bfly.b32 %r1199|%p207, %r1198, %r1152, %r1151, %r1153;
mov.b32 %f681, %r1199;
add.ftz.f32 %f682, %f671, %f681;
mov.b32 %r1200, %f682;
shfl.sync.bfly.b32 %r1201|%p208, %r1200, %r1156, %r1151, %r1153;
mov.b32 %f683, %r1201;
add.ftz.f32 %f684, %f682, %f683;
mov.b32 %r1202, %f684;
shfl.sync.bfly.b32 %r1203|%p209, %r1202, %r1159, %r1151, %r1153;
mov.b32 %f685, %r1203;
add.ftz.f32 %f686, %f684, %f685;
mov.b32 %r1204, %f686;
shfl.sync.bfly.b32 %r1205|%p210, %r1204, %r1162, %r1151, %r1153;
mov.b32 %f687, %r1205;
add.ftz.f32 %f688, %f686, %f687;
mov.b32 %r1206, %f688;
shfl.sync.bfly.b32 %r1207|%p211, %r1206, %r1165, %r1151, %r1153;
mov.b32 %f689, %r1207;
add.ftz.f32 %f63, %f688, %f689;
setp.ne.s32 %p212, %r2, 3;
or.b32 %r1211, %r4, 3;
setp.ge.u32 %p213, %r1211, %r45;
or.pred %p214, %p212, %p213;
@%p214 bra $L__BB58_78;

add.ftz.f32 %f690, %f5, %f62;
add.ftz.f32 %f691, %f4, %f63;
mul.ftz.f32 %f692, %f691, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f693, %f692;
add.ftz.f32 %f694, %f693, 0f3F800000;
div.approx.ftz.f32 %f695, %f691, %f694;
mul.ftz.f32 %f696, %f690, %f695;
st.global.f32 [%rd17+12], %f696;
bra.uni $L__BB58_78;

$L__BB58_25:
ld.shared.f32 %f149, [%r26];
add.ftz.f32 %f150, %f149, %f29;
ld.shared.f32 %f151, [%r26+512];
add.ftz.f32 %f152, %f151, %f150;
ld.shared.f32 %f153, [%r26+1024];
add.ftz.f32 %f154, %f153, %f152;
mov.b32 %r756, %f154;
mov.u32 %r757, 31;
mov.u32 %r758, 16;
mov.u32 %r759, -1;
shfl.sync.bfly.b32 %r760|%p22, %r756, %r758, %r757, %r759;
mov.b32 %f155, %r760;
add.ftz.f32 %f156, %f154, %f155;
mov.b32 %r761, %f156;
mov.u32 %r762, 8;
shfl.sync.bfly.b32 %r763|%p23, %r761, %r762, %r757, %r759;
mov.b32 %f157, %r763;
add.ftz.f32 %f158, %f156, %f157;
mov.b32 %r764, %f158;
mov.u32 %r765, 4;
shfl.sync.bfly.b32 %r766|%p24, %r764, %r765, %r757, %r759;
mov.b32 %f159, %r766;
add.ftz.f32 %f160, %f158, %f159;
mov.b32 %r767, %f160;
mov.u32 %r768, 2;
shfl.sync.bfly.b32 %r769|%p25, %r767, %r768, %r757, %r759;
mov.b32 %f161, %r769;
add.ftz.f32 %f162, %f160, %f161;
mov.b32 %r770, %f162;
mov.u32 %r771, 1;
shfl.sync.bfly.b32 %r772|%p26, %r770, %r771, %r757, %r759;
mov.b32 %f163, %r772;
add.ftz.f32 %f36, %f162, %f163;
setp.ne.s32 %p27, %r2, 0;
setp.ge.u32 %p28, %r4, %r45;
or.pred %p29, %p27, %p28;
@%p29 bra $L__BB58_27;

add.ftz.f32 %f164, %f5, %f36;
st.global.f32 [%rd17], %f164;

$L__BB58_27:
ld.shared.f32 %f165, [%r26+128];
add.ftz.f32 %f166, %f165, %f28;
ld.shared.f32 %f167, [%r26+640];
add.ftz.f32 %f168, %f167, %f166;
ld.shared.f32 %f169, [%r26+1152];
add.ftz.f32 %f170, %f169, %f168;
mov.b32 %r776, %f170;
shfl.sync.bfly.b32 %r780|%p30, %r776, %r758, %r757, %r759;
mov.b32 %f171, %r780;
add.ftz.f32 %f172, %f170, %f171;
mov.b32 %r781, %f172;
shfl.sync.bfly.b32 %r783|%p31, %r781, %r762, %r757, %r759;
mov.b32 %f173, %r783;
add.ftz.f32 %f174, %f172, %f173;
mov.b32 %r784, %f174;
shfl.sync.bfly.b32 %r786|%p32, %r784, %r765, %r757, %r759;
mov.b32 %f175, %r786;
add.ftz.f32 %f176, %f174, %f175;
mov.b32 %r787, %f176;
shfl.sync.bfly.b32 %r789|%p33, %r787, %r768, %r757, %r759;
mov.b32 %f177, %r789;
add.ftz.f32 %f178, %f176, %f177;
mov.b32 %r790, %f178;
shfl.sync.bfly.b32 %r792|%p34, %r790, %r771, %r757, %r759;
mov.b32 %f179, %r792;
add.ftz.f32 %f37, %f178, %f179;
setp.ne.s32 %p35, %r2, 1;
or.b32 %r796, %r4, 1;
setp.ge.u32 %p36, %r796, %r45;
or.pred %p37, %p35, %p36;
@%p37 bra $L__BB58_29;

add.ftz.f32 %f180, %f5, %f37;
st.global.f32 [%rd17+4], %f180;

$L__BB58_29:
ld.shared.f32 %f181, [%r26+256];
add.ftz.f32 %f182, %f181, %f27;
ld.shared.f32 %f183, [%r26+768];
add.ftz.f32 %f184, %f183, %f182;
ld.shared.f32 %f185, [%r26+1280];
add.ftz.f32 %f186, %f185, %f184;
mov.b32 %r797, %f186;
mov.u32 %r798, 31;
mov.u32 %r799, 16;
mov.u32 %r800, -1;
shfl.sync.bfly.b32 %r801|%p38, %r797, %r799, %r798, %r800;
mov.b32 %f187, %r801;
add.ftz.f32 %f188, %f186, %f187;
mov.b32 %r802, %f188;
mov.u32 %r803, 8;
shfl.sync.bfly.b32 %r804|%p39, %r802, %r803, %r798, %r800;
mov.b32 %f189, %r804;
add.ftz.f32 %f190, %f188, %f189;
mov.b32 %r805, %f190;
mov.u32 %r806, 4;
shfl.sync.bfly.b32 %r807|%p40, %r805, %r806, %r798, %r800;
mov.b32 %f191, %r807;
add.ftz.f32 %f192, %f190, %f191;
mov.b32 %r808, %f192;
mov.u32 %r809, 2;
shfl.sync.bfly.b32 %r810|%p41, %r808, %r809, %r798, %r800;
mov.b32 %f193, %r810;
add.ftz.f32 %f194, %f192, %f193;
mov.b32 %r811, %f194;
mov.u32 %r812, 1;
shfl.sync.bfly.b32 %r813|%p42, %r811, %r812, %r798, %r800;
mov.b32 %f195, %r813;
add.ftz.f32 %f38, %f194, %f195;
setp.ne.s32 %p43, %r2, 2;
or.b32 %r817, %r4, 2;
setp.ge.u32 %p44, %r817, %r45;
or.pred %p45, %p43, %p44;
@%p45 bra $L__BB58_31;

add.ftz.f32 %f196, %f5, %f38;
st.global.f32 [%rd17+8], %f196;

$L__BB58_31:
ld.shared.f32 %f197, [%r26+384];
add.ftz.f32 %f198, %f197, %f26;
ld.shared.f32 %f199, [%r26+896];
add.ftz.f32 %f200, %f199, %f198;
ld.shared.f32 %f201, [%r26+1408];
add.ftz.f32 %f202, %f201, %f200;
mov.b32 %r818, %f202;
shfl.sync.bfly.b32 %r822|%p46, %r818, %r799, %r798, %r800;
mov.b32 %f203, %r822;
add.ftz.f32 %f204, %f202, %f203;
mov.b32 %r823, %f204;
shfl.sync.bfly.b32 %r825|%p47, %r823, %r803, %r798, %r800;
mov.b32 %f205, %r825;
add.ftz.f32 %f206, %f204, %f205;
mov.b32 %r826, %f206;
shfl.sync.bfly.b32 %r828|%p48, %r826, %r806, %r798, %r800;
mov.b32 %f207, %r828;
add.ftz.f32 %f208, %f206, %f207;
mov.b32 %r829, %f208;
shfl.sync.bfly.b32 %r831|%p49, %r829, %r809, %r798, %r800;
mov.b32 %f209, %r831;
add.ftz.f32 %f210, %f208, %f209;
mov.b32 %r832, %f210;
shfl.sync.bfly.b32 %r834|%p50, %r832, %r812, %r798, %r800;
mov.b32 %f211, %r834;
add.ftz.f32 %f39, %f210, %f211;
setp.ne.s32 %p51, %r2, 3;
or.b32 %r838, %r4, 3;
setp.ge.u32 %p52, %r838, %r45;
or.pred %p53, %p51, %p52;
@%p53 bra $L__BB58_78;

add.ftz.f32 %f212, %f5, %f39;
st.global.f32 [%rd17+12], %f212;
bra.uni $L__BB58_78;

$L__BB58_44:
ld.param.u32 %r1471, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p55, %r1471, 3;
@%p55 bra $L__BB58_54;

ld.param.u32 %r1472, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.ne.s32 %p56, %r1472, 6;
@%p56 bra $L__BB58_70;

ld.shared.f32 %f213, [%r26];
add.ftz.f32 %f214, %f213, %f29;
ld.shared.f32 %f215, [%r28];
add.ftz.f32 %f216, %f215, %f33;
ld.shared.f32 %f217, [%r26+512];
add.ftz.f32 %f218, %f217, %f214;
ld.shared.f32 %f219, [%r28+512];
add.ftz.f32 %f220, %f219, %f216;
ld.shared.f32 %f221, [%r26+1024];
add.ftz.f32 %f222, %f221, %f218;
ld.shared.f32 %f223, [%r28+1024];
add.ftz.f32 %f224, %f223, %f220;
mov.b32 %r843, %f222;
mov.u32 %r844, 31;
mov.u32 %r845, 16;
mov.u32 %r846, -1;
shfl.sync.bfly.b32 %r847|%p59, %r843, %r845, %r844, %r846;
mov.b32 %f225, %r847;
add.ftz.f32 %f226, %f222, %f225;
mov.b32 %r848, %f226;
mov.u32 %r849, 8;
shfl.sync.bfly.b32 %r850|%p60, %r848, %r849, %r844, %r846;
mov.b32 %f227, %r850;
add.ftz.f32 %f228, %f226, %f227;
mov.b32 %r851, %f228;
mov.u32 %r852, 4;
shfl.sync.bfly.b32 %r853|%p61, %r851, %r852, %r844, %r846;
mov.b32 %f229, %r853;
add.ftz.f32 %f230, %f228, %f229;
mov.b32 %r854, %f230;
mov.u32 %r855, 2;
shfl.sync.bfly.b32 %r856|%p62, %r854, %r855, %r844, %r846;
mov.b32 %f231, %r856;
add.ftz.f32 %f232, %f230, %f231;
mov.b32 %r857, %f232;
mov.u32 %r858, 1;
shfl.sync.bfly.b32 %r859|%p63, %r857, %r858, %r844, %r846;
mov.b32 %f233, %r859;
add.ftz.f32 %f40, %f232, %f233;
mov.b32 %r860, %f224;
shfl.sync.bfly.b32 %r861|%p64, %r860, %r845, %r844, %r846;
mov.b32 %f234, %r861;
add.ftz.f32 %f235, %f224, %f234;
mov.b32 %r862, %f235;
shfl.sync.bfly.b32 %r863|%p65, %r862, %r849, %r844, %r846;
mov.b32 %f236, %r863;
add.ftz.f32 %f237, %f235, %f236;
mov.b32 %r864, %f237;
shfl.sync.bfly.b32 %r865|%p66, %r864, %r852, %r844, %r846;
mov.b32 %f238, %r865;
add.ftz.f32 %f239, %f237, %f238;
mov.b32 %r866, %f239;
shfl.sync.bfly.b32 %r867|%p67, %r866, %r855, %r844, %r846;
mov.b32 %f240, %r867;
add.ftz.f32 %f241, %f239, %f240;
mov.b32 %r868, %f241;
shfl.sync.bfly.b32 %r869|%p68, %r868, %r858, %r844, %r846;
mov.b32 %f242, %r869;
add.ftz.f32 %f41, %f241, %f242;
setp.ne.s32 %p69, %r2, 0;
setp.ge.u32 %p70, %r4, %r45;
or.pred %p71, %p69, %p70;
@%p71 bra $L__BB58_48;

ld.param.f32 %f997, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f243, %f5, %f40;
add.ftz.f32 %f244, %f4, %f41;
min.ftz.f32 %f245, %f244, %f997;
min.ftz.f32 %f246, %f243, %f997;
max.ftz.f32 %f247, %f246, %f35;
mul.ftz.f32 %f248, %f245, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f249, %f248;
add.ftz.f32 %f250, %f249, 0f3F800000;
div.approx.ftz.f32 %f251, %f245, %f250;
mul.ftz.f32 %f252, %f247, %f251;
st.global.f32 [%rd17], %f252;

$L__BB58_48:
ld.shared.f32 %f253, [%r26+128];
add.ftz.f32 %f254, %f253, %f28;
ld.shared.f32 %f255, [%r28+128];
add.ftz.f32 %f256, %f255, %f32;
ld.shared.f32 %f257, [%r26+640];
add.ftz.f32 %f258, %f257, %f254;
ld.shared.f32 %f259, [%r28+640];
add.ftz.f32 %f260, %f259, %f256;
ld.shared.f32 %f261, [%r26+1152];
add.ftz.f32 %f262, %f261, %f258;
ld.shared.f32 %f263, [%r28+1152];
add.ftz.f32 %f264, %f263, %f260;
mov.b32 %r873, %f262;
shfl.sync.bfly.b32 %r877|%p72, %r873, %r845, %r844, %r846;
mov.b32 %f265, %r877;
add.ftz.f32 %f266, %f262, %f265;
mov.b32 %r878, %f266;
shfl.sync.bfly.b32 %r880|%p73, %r878, %r849, %r844, %r846;
mov.b32 %f267, %r880;
add.ftz.f32 %f268, %f266, %f267;
mov.b32 %r881, %f268;
shfl.sync.bfly.b32 %r883|%p74, %r881, %r852, %r844, %r846;
mov.b32 %f269, %r883;
add.ftz.f32 %f270, %f268, %f269;
mov.b32 %r884, %f270;
shfl.sync.bfly.b32 %r886|%p75, %r884, %r855, %r844, %r846;
mov.b32 %f271, %r886;
add.ftz.f32 %f272, %f270, %f271;
mov.b32 %r887, %f272;
shfl.sync.bfly.b32 %r889|%p76, %r887, %r858, %r844, %r846;
mov.b32 %f273, %r889;
add.ftz.f32 %f42, %f272, %f273;
mov.b32 %r890, %f264;
shfl.sync.bfly.b32 %r891|%p77, %r890, %r845, %r844, %r846;
mov.b32 %f274, %r891;
add.ftz.f32 %f275, %f264, %f274;
mov.b32 %r892, %f275;
shfl.sync.bfly.b32 %r893|%p78, %r892, %r849, %r844, %r846;
mov.b32 %f276, %r893;
add.ftz.f32 %f277, %f275, %f276;
mov.b32 %r894, %f277;
shfl.sync.bfly.b32 %r895|%p79, %r894, %r852, %r844, %r846;
mov.b32 %f278, %r895;
add.ftz.f32 %f279, %f277, %f278;
mov.b32 %r896, %f279;
shfl.sync.bfly.b32 %r897|%p80, %r896, %r855, %r844, %r846;
mov.b32 %f280, %r897;
add.ftz.f32 %f281, %f279, %f280;
mov.b32 %r898, %f281;
shfl.sync.bfly.b32 %r899|%p81, %r898, %r858, %r844, %r846;
mov.b32 %f282, %r899;
add.ftz.f32 %f43, %f281, %f282;
setp.ne.s32 %p82, %r2, 1;
or.b32 %r903, %r4, 1;
setp.ge.u32 %p83, %r903, %r45;
or.pred %p84, %p82, %p83;
@%p84 bra $L__BB58_50;

ld.param.f32 %f996, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f283, %f5, %f42;
add.ftz.f32 %f284, %f4, %f43;
min.ftz.f32 %f285, %f284, %f996;
min.ftz.f32 %f286, %f283, %f996;
max.ftz.f32 %f287, %f286, %f35;
mul.ftz.f32 %f288, %f285, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f289, %f288;
add.ftz.f32 %f290, %f289, 0f3F800000;
div.approx.ftz.f32 %f291, %f285, %f290;
mul.ftz.f32 %f292, %f287, %f291;
st.global.f32 [%rd17+4], %f292;

$L__BB58_50:
ld.shared.f32 %f293, [%r26+256];
add.ftz.f32 %f294, %f293, %f27;
ld.shared.f32 %f295, [%r28+256];
add.ftz.f32 %f296, %f295, %f31;
ld.shared.f32 %f297, [%r26+768];
add.ftz.f32 %f298, %f297, %f294;
ld.shared.f32 %f299, [%r28+768];
add.ftz.f32 %f300, %f299, %f296;
ld.shared.f32 %f301, [%r26+1280];
add.ftz.f32 %f302, %f301, %f298;
ld.shared.f32 %f303, [%r28+1280];
add.ftz.f32 %f304, %f303, %f300;
mov.b32 %r904, %f302;
mov.u32 %r905, 31;
mov.u32 %r906, 16;
mov.u32 %r907, -1;
shfl.sync.bfly.b32 %r908|%p85, %r904, %r906, %r905, %r907;
mov.b32 %f305, %r908;
add.ftz.f32 %f306, %f302, %f305;
mov.b32 %r909, %f306;
mov.u32 %r910, 8;
shfl.sync.bfly.b32 %r911|%p86, %r909, %r910, %r905, %r907;
mov.b32 %f307, %r911;
add.ftz.f32 %f308, %f306, %f307;
mov.b32 %r912, %f308;
mov.u32 %r913, 4;
shfl.sync.bfly.b32 %r914|%p87, %r912, %r913, %r905, %r907;
mov.b32 %f309, %r914;
add.ftz.f32 %f310, %f308, %f309;
mov.b32 %r915, %f310;
mov.u32 %r916, 2;
shfl.sync.bfly.b32 %r917|%p88, %r915, %r916, %r905, %r907;
mov.b32 %f311, %r917;
add.ftz.f32 %f312, %f310, %f311;
mov.b32 %r918, %f312;
mov.u32 %r919, 1;
shfl.sync.bfly.b32 %r920|%p89, %r918, %r919, %r905, %r907;
mov.b32 %f313, %r920;
add.ftz.f32 %f44, %f312, %f313;
mov.b32 %r921, %f304;
shfl.sync.bfly.b32 %r922|%p90, %r921, %r906, %r905, %r907;
mov.b32 %f314, %r922;
add.ftz.f32 %f315, %f304, %f314;
mov.b32 %r923, %f315;
shfl.sync.bfly.b32 %r924|%p91, %r923, %r910, %r905, %r907;
mov.b32 %f316, %r924;
add.ftz.f32 %f317, %f315, %f316;
mov.b32 %r925, %f317;
shfl.sync.bfly.b32 %r926|%p92, %r925, %r913, %r905, %r907;
mov.b32 %f318, %r926;
add.ftz.f32 %f319, %f317, %f318;
mov.b32 %r927, %f319;
shfl.sync.bfly.b32 %r928|%p93, %r927, %r916, %r905, %r907;
mov.b32 %f320, %r928;
add.ftz.f32 %f321, %f319, %f320;
mov.b32 %r929, %f321;
shfl.sync.bfly.b32 %r930|%p94, %r929, %r919, %r905, %r907;
mov.b32 %f322, %r930;
add.ftz.f32 %f45, %f321, %f322;
setp.ne.s32 %p95, %r2, 2;
or.b32 %r934, %r4, 2;
setp.ge.u32 %p96, %r934, %r45;
or.pred %p97, %p95, %p96;
@%p97 bra $L__BB58_52;

ld.param.f32 %f995, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f323, %f5, %f44;
add.ftz.f32 %f324, %f4, %f45;
min.ftz.f32 %f325, %f324, %f995;
min.ftz.f32 %f326, %f323, %f995;
max.ftz.f32 %f327, %f326, %f35;
mul.ftz.f32 %f328, %f325, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f329, %f328;
add.ftz.f32 %f330, %f329, 0f3F800000;
div.approx.ftz.f32 %f331, %f325, %f330;
mul.ftz.f32 %f332, %f327, %f331;
st.global.f32 [%rd17+8], %f332;

$L__BB58_52:
ld.shared.f32 %f333, [%r26+384];
add.ftz.f32 %f334, %f333, %f26;
ld.shared.f32 %f335, [%r28+384];
add.ftz.f32 %f336, %f335, %f30;
ld.shared.f32 %f337, [%r26+896];
add.ftz.f32 %f338, %f337, %f334;
ld.shared.f32 %f339, [%r28+896];
add.ftz.f32 %f340, %f339, %f336;
ld.shared.f32 %f341, [%r26+1408];
add.ftz.f32 %f342, %f341, %f338;
ld.shared.f32 %f343, [%r28+1408];
add.ftz.f32 %f344, %f343, %f340;
mov.b32 %r935, %f342;
shfl.sync.bfly.b32 %r939|%p98, %r935, %r906, %r905, %r907;
mov.b32 %f345, %r939;
add.ftz.f32 %f346, %f342, %f345;
mov.b32 %r940, %f346;
shfl.sync.bfly.b32 %r942|%p99, %r940, %r910, %r905, %r907;
mov.b32 %f347, %r942;
add.ftz.f32 %f348, %f346, %f347;
mov.b32 %r943, %f348;
shfl.sync.bfly.b32 %r945|%p100, %r943, %r913, %r905, %r907;
mov.b32 %f349, %r945;
add.ftz.f32 %f350, %f348, %f349;
mov.b32 %r946, %f350;
shfl.sync.bfly.b32 %r948|%p101, %r946, %r916, %r905, %r907;
mov.b32 %f351, %r948;
add.ftz.f32 %f352, %f350, %f351;
mov.b32 %r949, %f352;
shfl.sync.bfly.b32 %r951|%p102, %r949, %r919, %r905, %r907;
mov.b32 %f353, %r951;
add.ftz.f32 %f46, %f352, %f353;
mov.b32 %r952, %f344;
shfl.sync.bfly.b32 %r953|%p103, %r952, %r906, %r905, %r907;
mov.b32 %f354, %r953;
add.ftz.f32 %f355, %f344, %f354;
mov.b32 %r954, %f355;
shfl.sync.bfly.b32 %r955|%p104, %r954, %r910, %r905, %r907;
mov.b32 %f356, %r955;
add.ftz.f32 %f357, %f355, %f356;
mov.b32 %r956, %f357;
shfl.sync.bfly.b32 %r957|%p105, %r956, %r913, %r905, %r907;
mov.b32 %f358, %r957;
add.ftz.f32 %f359, %f357, %f358;
mov.b32 %r958, %f359;
shfl.sync.bfly.b32 %r959|%p106, %r958, %r916, %r905, %r907;
mov.b32 %f360, %r959;
add.ftz.f32 %f361, %f359, %f360;
mov.b32 %r960, %f361;
shfl.sync.bfly.b32 %r961|%p107, %r960, %r919, %r905, %r907;
mov.b32 %f362, %r961;
add.ftz.f32 %f47, %f361, %f362;
setp.ne.s32 %p108, %r2, 3;
or.b32 %r965, %r4, 3;
setp.ge.u32 %p109, %r965, %r45;
or.pred %p110, %p108, %p109;
@%p110 bra $L__BB58_78;

ld.param.f32 %f994, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f363, %f5, %f46;
add.ftz.f32 %f364, %f4, %f47;
min.ftz.f32 %f365, %f364, %f994;
min.ftz.f32 %f366, %f363, %f994;
max.ftz.f32 %f367, %f366, %f35;
mul.ftz.f32 %f368, %f365, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f369, %f368;
add.ftz.f32 %f370, %f369, 0f3F800000;
div.approx.ftz.f32 %f371, %f365, %f370;
mul.ftz.f32 %f372, %f367, %f371;
st.global.f32 [%rd17+12], %f372;
bra.uni $L__BB58_78;

$L__BB58_70:
ld.shared.f32 %f861, [%r26];
add.ftz.f32 %f862, %f861, %f29;
ld.shared.f32 %f863, [%r28];
add.ftz.f32 %f864, %f863, %f33;
ld.shared.f32 %f865, [%r26+512];
add.ftz.f32 %f866, %f865, %f862;
ld.shared.f32 %f867, [%r28+512];
add.ftz.f32 %f868, %f867, %f864;
ld.shared.f32 %f869, [%r26+1024];
add.ftz.f32 %f870, %f869, %f866;
ld.shared.f32 %f871, [%r28+1024];
add.ftz.f32 %f872, %f871, %f868;
mov.b32 %r1335, %f870;
mov.u32 %r1336, 31;
mov.u32 %r1337, 16;
mov.u32 %r1338, -1;
shfl.sync.bfly.b32 %r1339|%p267, %r1335, %r1337, %r1336, %r1338;
mov.b32 %f873, %r1339;
add.ftz.f32 %f874, %f870, %f873;
mov.b32 %r1340, %f874;
mov.u32 %r1341, 8;
shfl.sync.bfly.b32 %r1342|%p268, %r1340, %r1341, %r1336, %r1338;
mov.b32 %f875, %r1342;
add.ftz.f32 %f876, %f874, %f875;
mov.b32 %r1343, %f876;
mov.u32 %r1344, 4;
shfl.sync.bfly.b32 %r1345|%p269, %r1343, %r1344, %r1336, %r1338;
mov.b32 %f877, %r1345;
add.ftz.f32 %f878, %f876, %f877;
mov.b32 %r1346, %f878;
mov.u32 %r1347, 2;
shfl.sync.bfly.b32 %r1348|%p270, %r1346, %r1347, %r1336, %r1338;
mov.b32 %f879, %r1348;
add.ftz.f32 %f880, %f878, %f879;
mov.b32 %r1349, %f880;
mov.u32 %r1350, 1;
shfl.sync.bfly.b32 %r1351|%p271, %r1349, %r1350, %r1336, %r1338;
mov.b32 %f881, %r1351;
add.ftz.f32 %f72, %f880, %f881;
mov.b32 %r1352, %f872;
shfl.sync.bfly.b32 %r1353|%p272, %r1352, %r1337, %r1336, %r1338;
mov.b32 %f882, %r1353;
add.ftz.f32 %f883, %f872, %f882;
mov.b32 %r1354, %f883;
shfl.sync.bfly.b32 %r1355|%p273, %r1354, %r1341, %r1336, %r1338;
mov.b32 %f884, %r1355;
add.ftz.f32 %f885, %f883, %f884;
mov.b32 %r1356, %f885;
shfl.sync.bfly.b32 %r1357|%p274, %r1356, %r1344, %r1336, %r1338;
mov.b32 %f886, %r1357;
add.ftz.f32 %f887, %f885, %f886;
mov.b32 %r1358, %f887;
shfl.sync.bfly.b32 %r1359|%p275, %r1358, %r1347, %r1336, %r1338;
mov.b32 %f888, %r1359;
add.ftz.f32 %f889, %f887, %f888;
mov.b32 %r1360, %f889;
shfl.sync.bfly.b32 %r1361|%p276, %r1360, %r1350, %r1336, %r1338;
mov.b32 %f890, %r1361;
add.ftz.f32 %f73, %f889, %f890;
setp.ne.s32 %p277, %r2, 0;
setp.ge.u32 %p278, %r4, %r45;
or.pred %p279, %p277, %p278;
@%p279 bra $L__BB58_72;

add.ftz.f32 %f891, %f5, %f72;
add.ftz.f32 %f892, %f4, %f73;
mul.ftz.f32 %f893, %f891, %f892;
st.global.f32 [%rd17], %f893;

$L__BB58_72:
ld.shared.f32 %f894, [%r26+128];
add.ftz.f32 %f895, %f894, %f28;
ld.shared.f32 %f896, [%r28+128];
add.ftz.f32 %f897, %f896, %f32;
ld.shared.f32 %f898, [%r26+640];
add.ftz.f32 %f899, %f898, %f895;
ld.shared.f32 %f900, [%r28+640];
add.ftz.f32 %f901, %f900, %f897;
ld.shared.f32 %f902, [%r26+1152];
add.ftz.f32 %f903, %f902, %f899;
ld.shared.f32 %f904, [%r28+1152];
add.ftz.f32 %f905, %f904, %f901;
mov.b32 %r1365, %f903;
shfl.sync.bfly.b32 %r1369|%p280, %r1365, %r1337, %r1336, %r1338;
mov.b32 %f906, %r1369;
add.ftz.f32 %f907, %f903, %f906;
mov.b32 %r1370, %f907;
shfl.sync.bfly.b32 %r1372|%p281, %r1370, %r1341, %r1336, %r1338;
mov.b32 %f908, %r1372;
add.ftz.f32 %f909, %f907, %f908;
mov.b32 %r1373, %f909;
shfl.sync.bfly.b32 %r1375|%p282, %r1373, %r1344, %r1336, %r1338;
mov.b32 %f910, %r1375;
add.ftz.f32 %f911, %f909, %f910;
mov.b32 %r1376, %f911;
shfl.sync.bfly.b32 %r1378|%p283, %r1376, %r1347, %r1336, %r1338;
mov.b32 %f912, %r1378;
add.ftz.f32 %f913, %f911, %f912;
mov.b32 %r1379, %f913;
shfl.sync.bfly.b32 %r1381|%p284, %r1379, %r1350, %r1336, %r1338;
mov.b32 %f914, %r1381;
add.ftz.f32 %f74, %f913, %f914;
mov.b32 %r1382, %f905;
shfl.sync.bfly.b32 %r1383|%p285, %r1382, %r1337, %r1336, %r1338;
mov.b32 %f915, %r1383;
add.ftz.f32 %f916, %f905, %f915;
mov.b32 %r1384, %f916;
shfl.sync.bfly.b32 %r1385|%p286, %r1384, %r1341, %r1336, %r1338;
mov.b32 %f917, %r1385;
add.ftz.f32 %f918, %f916, %f917;
mov.b32 %r1386, %f918;
shfl.sync.bfly.b32 %r1387|%p287, %r1386, %r1344, %r1336, %r1338;
mov.b32 %f919, %r1387;
add.ftz.f32 %f920, %f918, %f919;
mov.b32 %r1388, %f920;
shfl.sync.bfly.b32 %r1389|%p288, %r1388, %r1347, %r1336, %r1338;
mov.b32 %f921, %r1389;
add.ftz.f32 %f922, %f920, %f921;
mov.b32 %r1390, %f922;
shfl.sync.bfly.b32 %r1391|%p289, %r1390, %r1350, %r1336, %r1338;
mov.b32 %f923, %r1391;
add.ftz.f32 %f75, %f922, %f923;
setp.ne.s32 %p290, %r2, 1;
or.b32 %r1395, %r4, 1;
setp.ge.u32 %p291, %r1395, %r45;
or.pred %p292, %p290, %p291;
@%p292 bra $L__BB58_74;

add.ftz.f32 %f924, %f5, %f74;
add.ftz.f32 %f925, %f4, %f75;
mul.ftz.f32 %f926, %f924, %f925;
st.global.f32 [%rd17+4], %f926;

$L__BB58_74:
ld.shared.f32 %f927, [%r26+256];
add.ftz.f32 %f928, %f927, %f27;
ld.shared.f32 %f929, [%r28+256];
add.ftz.f32 %f930, %f929, %f31;
ld.shared.f32 %f931, [%r26+768];
add.ftz.f32 %f932, %f931, %f928;
ld.shared.f32 %f933, [%r28+768];
add.ftz.f32 %f934, %f933, %f930;
ld.shared.f32 %f935, [%r26+1280];
add.ftz.f32 %f936, %f935, %f932;
ld.shared.f32 %f937, [%r28+1280];
add.ftz.f32 %f938, %f937, %f934;
mov.b32 %r1396, %f936;
mov.u32 %r1397, 31;
mov.u32 %r1398, 16;
mov.u32 %r1399, -1;
shfl.sync.bfly.b32 %r1400|%p293, %r1396, %r1398, %r1397, %r1399;
mov.b32 %f939, %r1400;
add.ftz.f32 %f940, %f936, %f939;
mov.b32 %r1401, %f940;
mov.u32 %r1402, 8;
shfl.sync.bfly.b32 %r1403|%p294, %r1401, %r1402, %r1397, %r1399;
mov.b32 %f941, %r1403;
add.ftz.f32 %f942, %f940, %f941;
mov.b32 %r1404, %f942;
mov.u32 %r1405, 4;
shfl.sync.bfly.b32 %r1406|%p295, %r1404, %r1405, %r1397, %r1399;
mov.b32 %f943, %r1406;
add.ftz.f32 %f944, %f942, %f943;
mov.b32 %r1407, %f944;
mov.u32 %r1408, 2;
shfl.sync.bfly.b32 %r1409|%p296, %r1407, %r1408, %r1397, %r1399;
mov.b32 %f945, %r1409;
add.ftz.f32 %f946, %f944, %f945;
mov.b32 %r1410, %f946;
mov.u32 %r1411, 1;
shfl.sync.bfly.b32 %r1412|%p297, %r1410, %r1411, %r1397, %r1399;
mov.b32 %f947, %r1412;
add.ftz.f32 %f76, %f946, %f947;
mov.b32 %r1413, %f938;
shfl.sync.bfly.b32 %r1414|%p298, %r1413, %r1398, %r1397, %r1399;
mov.b32 %f948, %r1414;
add.ftz.f32 %f949, %f938, %f948;
mov.b32 %r1415, %f949;
shfl.sync.bfly.b32 %r1416|%p299, %r1415, %r1402, %r1397, %r1399;
mov.b32 %f950, %r1416;
add.ftz.f32 %f951, %f949, %f950;
mov.b32 %r1417, %f951;
shfl.sync.bfly.b32 %r1418|%p300, %r1417, %r1405, %r1397, %r1399;
mov.b32 %f952, %r1418;
add.ftz.f32 %f953, %f951, %f952;
mov.b32 %r1419, %f953;
shfl.sync.bfly.b32 %r1420|%p301, %r1419, %r1408, %r1397, %r1399;
mov.b32 %f954, %r1420;
add.ftz.f32 %f955, %f953, %f954;
mov.b32 %r1421, %f955;
shfl.sync.bfly.b32 %r1422|%p302, %r1421, %r1411, %r1397, %r1399;
mov.b32 %f956, %r1422;
add.ftz.f32 %f77, %f955, %f956;
setp.ne.s32 %p303, %r2, 2;
or.b32 %r1426, %r4, 2;
setp.ge.u32 %p304, %r1426, %r45;
or.pred %p305, %p303, %p304;
@%p305 bra $L__BB58_76;

add.ftz.f32 %f957, %f5, %f76;
add.ftz.f32 %f958, %f4, %f77;
mul.ftz.f32 %f959, %f957, %f958;
st.global.f32 [%rd17+8], %f959;

$L__BB58_76:
ld.shared.f32 %f960, [%r26+384];
add.ftz.f32 %f961, %f960, %f26;
ld.shared.f32 %f962, [%r28+384];
add.ftz.f32 %f963, %f962, %f30;
ld.shared.f32 %f964, [%r26+896];
add.ftz.f32 %f965, %f964, %f961;
ld.shared.f32 %f966, [%r28+896];
add.ftz.f32 %f967, %f966, %f963;
ld.shared.f32 %f968, [%r26+1408];
add.ftz.f32 %f969, %f968, %f965;
ld.shared.f32 %f970, [%r28+1408];
add.ftz.f32 %f971, %f970, %f967;
mov.b32 %r1427, %f969;
shfl.sync.bfly.b32 %r1431|%p306, %r1427, %r1398, %r1397, %r1399;
mov.b32 %f972, %r1431;
add.ftz.f32 %f973, %f969, %f972;
mov.b32 %r1432, %f973;
shfl.sync.bfly.b32 %r1434|%p307, %r1432, %r1402, %r1397, %r1399;
mov.b32 %f974, %r1434;
add.ftz.f32 %f975, %f973, %f974;
mov.b32 %r1435, %f975;
shfl.sync.bfly.b32 %r1437|%p308, %r1435, %r1405, %r1397, %r1399;
mov.b32 %f976, %r1437;
add.ftz.f32 %f977, %f975, %f976;
mov.b32 %r1438, %f977;
shfl.sync.bfly.b32 %r1440|%p309, %r1438, %r1408, %r1397, %r1399;
mov.b32 %f978, %r1440;
add.ftz.f32 %f979, %f977, %f978;
mov.b32 %r1441, %f979;
shfl.sync.bfly.b32 %r1443|%p310, %r1441, %r1411, %r1397, %r1399;
mov.b32 %f980, %r1443;
add.ftz.f32 %f78, %f979, %f980;
mov.b32 %r1444, %f971;
shfl.sync.bfly.b32 %r1445|%p311, %r1444, %r1398, %r1397, %r1399;
mov.b32 %f981, %r1445;
add.ftz.f32 %f982, %f971, %f981;
mov.b32 %r1446, %f982;
shfl.sync.bfly.b32 %r1447|%p312, %r1446, %r1402, %r1397, %r1399;
mov.b32 %f983, %r1447;
add.ftz.f32 %f984, %f982, %f983;
mov.b32 %r1448, %f984;
shfl.sync.bfly.b32 %r1449|%p313, %r1448, %r1405, %r1397, %r1399;
mov.b32 %f985, %r1449;
add.ftz.f32 %f986, %f984, %f985;
mov.b32 %r1450, %f986;
shfl.sync.bfly.b32 %r1451|%p314, %r1450, %r1408, %r1397, %r1399;
mov.b32 %f987, %r1451;
add.ftz.f32 %f988, %f986, %f987;
mov.b32 %r1452, %f988;
shfl.sync.bfly.b32 %r1453|%p315, %r1452, %r1411, %r1397, %r1399;
mov.b32 %f989, %r1453;
add.ftz.f32 %f79, %f988, %f989;
setp.ne.s32 %p316, %r2, 3;
or.b32 %r1457, %r4, 3;
setp.ge.u32 %p317, %r1457, %r45;
or.pred %p318, %p316, %p317;
@%p318 bra $L__BB58_78;

add.ftz.f32 %f990, %f5, %f78;
add.ftz.f32 %f991, %f4, %f79;
mul.ftz.f32 %f992, %f990, %f991;
st.global.f32 [%rd17+12], %f992;
bra.uni $L__BB58_78;

$L__BB58_62:
ld.shared.f32 %f697, [%r26];
add.ftz.f32 %f698, %f697, %f29;
ld.shared.f32 %f699, [%r28];
add.ftz.f32 %f700, %f699, %f33;
ld.shared.f32 %f701, [%r26+512];
add.ftz.f32 %f702, %f701, %f698;
ld.shared.f32 %f703, [%r28+512];
add.ftz.f32 %f704, %f703, %f700;
ld.shared.f32 %f705, [%r26+1024];
add.ftz.f32 %f706, %f705, %f702;
ld.shared.f32 %f707, [%r28+1024];
add.ftz.f32 %f708, %f707, %f704;
mov.b32 %r1212, %f706;
mov.u32 %r1213, 31;
mov.u32 %r1214, 16;
mov.u32 %r1215, -1;
shfl.sync.bfly.b32 %r1216|%p215, %r1212, %r1214, %r1213, %r1215;
mov.b32 %f709, %r1216;
add.ftz.f32 %f710, %f706, %f709;
mov.b32 %r1217, %f710;
mov.u32 %r1218, 8;
shfl.sync.bfly.b32 %r1219|%p216, %r1217, %r1218, %r1213, %r1215;
mov.b32 %f711, %r1219;
add.ftz.f32 %f712, %f710, %f711;
mov.b32 %r1220, %f712;
mov.u32 %r1221, 4;
shfl.sync.bfly.b32 %r1222|%p217, %r1220, %r1221, %r1213, %r1215;
mov.b32 %f713, %r1222;
add.ftz.f32 %f714, %f712, %f713;
mov.b32 %r1223, %f714;
mov.u32 %r1224, 2;
shfl.sync.bfly.b32 %r1225|%p218, %r1223, %r1224, %r1213, %r1215;
mov.b32 %f715, %r1225;
add.ftz.f32 %f716, %f714, %f715;
mov.b32 %r1226, %f716;
mov.u32 %r1227, 1;
shfl.sync.bfly.b32 %r1228|%p219, %r1226, %r1227, %r1213, %r1215;
mov.b32 %f717, %r1228;
add.ftz.f32 %f64, %f716, %f717;
mov.b32 %r1229, %f708;
shfl.sync.bfly.b32 %r1230|%p220, %r1229, %r1214, %r1213, %r1215;
mov.b32 %f718, %r1230;
add.ftz.f32 %f719, %f708, %f718;
mov.b32 %r1231, %f719;
shfl.sync.bfly.b32 %r1232|%p221, %r1231, %r1218, %r1213, %r1215;
mov.b32 %f720, %r1232;
add.ftz.f32 %f721, %f719, %f720;
mov.b32 %r1233, %f721;
shfl.sync.bfly.b32 %r1234|%p222, %r1233, %r1221, %r1213, %r1215;
mov.b32 %f722, %r1234;
add.ftz.f32 %f723, %f721, %f722;
mov.b32 %r1235, %f723;
shfl.sync.bfly.b32 %r1236|%p223, %r1235, %r1224, %r1213, %r1215;
mov.b32 %f724, %r1236;
add.ftz.f32 %f725, %f723, %f724;
mov.b32 %r1237, %f725;
shfl.sync.bfly.b32 %r1238|%p224, %r1237, %r1227, %r1213, %r1215;
mov.b32 %f726, %r1238;
add.ftz.f32 %f65, %f725, %f726;
setp.ne.s32 %p225, %r2, 0;
setp.ge.u32 %p226, %r4, %r45;
or.pred %p227, %p225, %p226;
@%p227 bra $L__BB58_64;

add.ftz.f32 %f727, %f5, %f64;
add.ftz.f32 %f728, %f4, %f65;
mul.ftz.f32 %f729, %f728, 0f3F000000;
mul.ftz.f32 %f730, %f728, 0f3F4C422A;
mul.ftz.f32 %f731, %f728, 0f3D372713;
fma.rn.ftz.f32 %f732, %f728, %f731, 0f3F800000;
mul.ftz.f32 %f733, %f730, %f732;
tanh.approx.f32 %f734, %f733;
add.ftz.f32 %f735, %f734, 0f3F800000;
mul.ftz.f32 %f736, %f729, %f735;
mul.ftz.f32 %f737, %f727, %f736;
st.global.f32 [%rd17], %f737;

$L__BB58_64:
ld.shared.f32 %f738, [%r26+128];
add.ftz.f32 %f739, %f738, %f28;
ld.shared.f32 %f740, [%r28+128];
add.ftz.f32 %f741, %f740, %f32;
ld.shared.f32 %f742, [%r26+640];
add.ftz.f32 %f743, %f742, %f739;
ld.shared.f32 %f744, [%r28+640];
add.ftz.f32 %f745, %f744, %f741;
ld.shared.f32 %f746, [%r26+1152];
add.ftz.f32 %f747, %f746, %f743;
ld.shared.f32 %f748, [%r28+1152];
add.ftz.f32 %f749, %f748, %f745;
mov.b32 %r1242, %f747;
shfl.sync.bfly.b32 %r1246|%p228, %r1242, %r1214, %r1213, %r1215;
mov.b32 %f750, %r1246;
add.ftz.f32 %f751, %f747, %f750;
mov.b32 %r1247, %f751;
shfl.sync.bfly.b32 %r1249|%p229, %r1247, %r1218, %r1213, %r1215;
mov.b32 %f752, %r1249;
add.ftz.f32 %f753, %f751, %f752;
mov.b32 %r1250, %f753;
shfl.sync.bfly.b32 %r1252|%p230, %r1250, %r1221, %r1213, %r1215;
mov.b32 %f754, %r1252;
add.ftz.f32 %f755, %f753, %f754;
mov.b32 %r1253, %f755;
shfl.sync.bfly.b32 %r1255|%p231, %r1253, %r1224, %r1213, %r1215;
mov.b32 %f756, %r1255;
add.ftz.f32 %f757, %f755, %f756;
mov.b32 %r1256, %f757;
shfl.sync.bfly.b32 %r1258|%p232, %r1256, %r1227, %r1213, %r1215;
mov.b32 %f758, %r1258;
add.ftz.f32 %f66, %f757, %f758;
mov.b32 %r1259, %f749;
shfl.sync.bfly.b32 %r1260|%p233, %r1259, %r1214, %r1213, %r1215;
mov.b32 %f759, %r1260;
add.ftz.f32 %f760, %f749, %f759;
mov.b32 %r1261, %f760;
shfl.sync.bfly.b32 %r1262|%p234, %r1261, %r1218, %r1213, %r1215;
mov.b32 %f761, %r1262;
add.ftz.f32 %f762, %f760, %f761;
mov.b32 %r1263, %f762;
shfl.sync.bfly.b32 %r1264|%p235, %r1263, %r1221, %r1213, %r1215;
mov.b32 %f763, %r1264;
add.ftz.f32 %f764, %f762, %f763;
mov.b32 %r1265, %f764;
shfl.sync.bfly.b32 %r1266|%p236, %r1265, %r1224, %r1213, %r1215;
mov.b32 %f765, %r1266;
add.ftz.f32 %f766, %f764, %f765;
mov.b32 %r1267, %f766;
shfl.sync.bfly.b32 %r1268|%p237, %r1267, %r1227, %r1213, %r1215;
mov.b32 %f767, %r1268;
add.ftz.f32 %f67, %f766, %f767;
setp.ne.s32 %p238, %r2, 1;
or.b32 %r1272, %r4, 1;
setp.ge.u32 %p239, %r1272, %r45;
or.pred %p240, %p238, %p239;
@%p240 bra $L__BB58_66;

add.ftz.f32 %f768, %f5, %f66;
add.ftz.f32 %f769, %f4, %f67;
mul.ftz.f32 %f770, %f769, 0f3F000000;
mul.ftz.f32 %f771, %f769, 0f3F4C422A;
mul.ftz.f32 %f772, %f769, 0f3D372713;
fma.rn.ftz.f32 %f773, %f769, %f772, 0f3F800000;
mul.ftz.f32 %f774, %f771, %f773;
tanh.approx.f32 %f775, %f774;
add.ftz.f32 %f776, %f775, 0f3F800000;
mul.ftz.f32 %f777, %f770, %f776;
mul.ftz.f32 %f778, %f768, %f777;
st.global.f32 [%rd17+4], %f778;

$L__BB58_66:
ld.shared.f32 %f779, [%r26+256];
add.ftz.f32 %f780, %f779, %f27;
ld.shared.f32 %f781, [%r28+256];
add.ftz.f32 %f782, %f781, %f31;
ld.shared.f32 %f783, [%r26+768];
add.ftz.f32 %f784, %f783, %f780;
ld.shared.f32 %f785, [%r28+768];
add.ftz.f32 %f786, %f785, %f782;
ld.shared.f32 %f787, [%r26+1280];
add.ftz.f32 %f788, %f787, %f784;
ld.shared.f32 %f789, [%r28+1280];
add.ftz.f32 %f790, %f789, %f786;
mov.b32 %r1273, %f788;
mov.u32 %r1274, 31;
mov.u32 %r1275, 16;
mov.u32 %r1276, -1;
shfl.sync.bfly.b32 %r1277|%p241, %r1273, %r1275, %r1274, %r1276;
mov.b32 %f791, %r1277;
add.ftz.f32 %f792, %f788, %f791;
mov.b32 %r1278, %f792;
mov.u32 %r1279, 8;
shfl.sync.bfly.b32 %r1280|%p242, %r1278, %r1279, %r1274, %r1276;
mov.b32 %f793, %r1280;
add.ftz.f32 %f794, %f792, %f793;
mov.b32 %r1281, %f794;
mov.u32 %r1282, 4;
shfl.sync.bfly.b32 %r1283|%p243, %r1281, %r1282, %r1274, %r1276;
mov.b32 %f795, %r1283;
add.ftz.f32 %f796, %f794, %f795;
mov.b32 %r1284, %f796;
mov.u32 %r1285, 2;
shfl.sync.bfly.b32 %r1286|%p244, %r1284, %r1285, %r1274, %r1276;
mov.b32 %f797, %r1286;
add.ftz.f32 %f798, %f796, %f797;
mov.b32 %r1287, %f798;
mov.u32 %r1288, 1;
shfl.sync.bfly.b32 %r1289|%p245, %r1287, %r1288, %r1274, %r1276;
mov.b32 %f799, %r1289;
add.ftz.f32 %f68, %f798, %f799;
mov.b32 %r1290, %f790;
shfl.sync.bfly.b32 %r1291|%p246, %r1290, %r1275, %r1274, %r1276;
mov.b32 %f800, %r1291;
add.ftz.f32 %f801, %f790, %f800;
mov.b32 %r1292, %f801;
shfl.sync.bfly.b32 %r1293|%p247, %r1292, %r1279, %r1274, %r1276;
mov.b32 %f802, %r1293;
add.ftz.f32 %f803, %f801, %f802;
mov.b32 %r1294, %f803;
shfl.sync.bfly.b32 %r1295|%p248, %r1294, %r1282, %r1274, %r1276;
mov.b32 %f804, %r1295;
add.ftz.f32 %f805, %f803, %f804;
mov.b32 %r1296, %f805;
shfl.sync.bfly.b32 %r1297|%p249, %r1296, %r1285, %r1274, %r1276;
mov.b32 %f806, %r1297;
add.ftz.f32 %f807, %f805, %f806;
mov.b32 %r1298, %f807;
shfl.sync.bfly.b32 %r1299|%p250, %r1298, %r1288, %r1274, %r1276;
mov.b32 %f808, %r1299;
add.ftz.f32 %f69, %f807, %f808;
setp.ne.s32 %p251, %r2, 2;
or.b32 %r1303, %r4, 2;
setp.ge.u32 %p252, %r1303, %r45;
or.pred %p253, %p251, %p252;
@%p253 bra $L__BB58_68;

add.ftz.f32 %f809, %f5, %f68;
add.ftz.f32 %f810, %f4, %f69;
mul.ftz.f32 %f811, %f810, 0f3F000000;
mul.ftz.f32 %f812, %f810, 0f3F4C422A;
mul.ftz.f32 %f813, %f810, 0f3D372713;
fma.rn.ftz.f32 %f814, %f810, %f813, 0f3F800000;
mul.ftz.f32 %f815, %f812, %f814;
tanh.approx.f32 %f816, %f815;
add.ftz.f32 %f817, %f816, 0f3F800000;
mul.ftz.f32 %f818, %f811, %f817;
mul.ftz.f32 %f819, %f809, %f818;
st.global.f32 [%rd17+8], %f819;

$L__BB58_68:
ld.shared.f32 %f820, [%r26+384];
add.ftz.f32 %f821, %f820, %f26;
ld.shared.f32 %f822, [%r28+384];
add.ftz.f32 %f823, %f822, %f30;
ld.shared.f32 %f824, [%r26+896];
add.ftz.f32 %f825, %f824, %f821;
ld.shared.f32 %f826, [%r28+896];
add.ftz.f32 %f827, %f826, %f823;
ld.shared.f32 %f828, [%r26+1408];
add.ftz.f32 %f829, %f828, %f825;
ld.shared.f32 %f830, [%r28+1408];
add.ftz.f32 %f831, %f830, %f827;
mov.b32 %r1304, %f829;
shfl.sync.bfly.b32 %r1308|%p254, %r1304, %r1275, %r1274, %r1276;
mov.b32 %f832, %r1308;
add.ftz.f32 %f833, %f829, %f832;
mov.b32 %r1309, %f833;
shfl.sync.bfly.b32 %r1311|%p255, %r1309, %r1279, %r1274, %r1276;
mov.b32 %f834, %r1311;
add.ftz.f32 %f835, %f833, %f834;
mov.b32 %r1312, %f835;
shfl.sync.bfly.b32 %r1314|%p256, %r1312, %r1282, %r1274, %r1276;
mov.b32 %f836, %r1314;
add.ftz.f32 %f837, %f835, %f836;
mov.b32 %r1315, %f837;
shfl.sync.bfly.b32 %r1317|%p257, %r1315, %r1285, %r1274, %r1276;
mov.b32 %f838, %r1317;
add.ftz.f32 %f839, %f837, %f838;
mov.b32 %r1318, %f839;
shfl.sync.bfly.b32 %r1320|%p258, %r1318, %r1288, %r1274, %r1276;
mov.b32 %f840, %r1320;
add.ftz.f32 %f70, %f839, %f840;
mov.b32 %r1321, %f831;
shfl.sync.bfly.b32 %r1322|%p259, %r1321, %r1275, %r1274, %r1276;
mov.b32 %f841, %r1322;
add.ftz.f32 %f842, %f831, %f841;
mov.b32 %r1323, %f842;
shfl.sync.bfly.b32 %r1324|%p260, %r1323, %r1279, %r1274, %r1276;
mov.b32 %f843, %r1324;
add.ftz.f32 %f844, %f842, %f843;
mov.b32 %r1325, %f844;
shfl.sync.bfly.b32 %r1326|%p261, %r1325, %r1282, %r1274, %r1276;
mov.b32 %f845, %r1326;
add.ftz.f32 %f846, %f844, %f845;
mov.b32 %r1327, %f846;
shfl.sync.bfly.b32 %r1328|%p262, %r1327, %r1285, %r1274, %r1276;
mov.b32 %f847, %r1328;
add.ftz.f32 %f848, %f846, %f847;
mov.b32 %r1329, %f848;
shfl.sync.bfly.b32 %r1330|%p263, %r1329, %r1288, %r1274, %r1276;
mov.b32 %f849, %r1330;
add.ftz.f32 %f71, %f848, %f849;
setp.ne.s32 %p264, %r2, 3;
or.b32 %r1334, %r4, 3;
setp.ge.u32 %p265, %r1334, %r45;
or.pred %p266, %p264, %p265;
@%p266 bra $L__BB58_78;

add.ftz.f32 %f850, %f5, %f70;
add.ftz.f32 %f851, %f4, %f71;
mul.ftz.f32 %f852, %f851, 0f3F000000;
mul.ftz.f32 %f853, %f851, 0f3F4C422A;
mul.ftz.f32 %f854, %f851, 0f3D372713;
fma.rn.ftz.f32 %f855, %f851, %f854, 0f3F800000;
mul.ftz.f32 %f856, %f853, %f855;
tanh.approx.f32 %f857, %f856;
add.ftz.f32 %f858, %f857, 0f3F800000;
mul.ftz.f32 %f859, %f852, %f858;
mul.ftz.f32 %f860, %f850, %f859;
st.global.f32 [%rd17+12], %f860;
bra.uni $L__BB58_78;

$L__BB58_54:
ld.shared.f32 %f373, [%r26];
add.ftz.f32 %f374, %f373, %f29;
ld.shared.f32 %f375, [%r28];
add.ftz.f32 %f376, %f375, %f33;
ld.shared.f32 %f377, [%r26+512];
add.ftz.f32 %f378, %f377, %f374;
ld.shared.f32 %f379, [%r28+512];
add.ftz.f32 %f380, %f379, %f376;
ld.shared.f32 %f381, [%r26+1024];
add.ftz.f32 %f382, %f381, %f378;
ld.shared.f32 %f383, [%r28+1024];
add.ftz.f32 %f384, %f383, %f380;
mov.b32 %r966, %f382;
mov.u32 %r967, 31;
mov.u32 %r968, 16;
mov.u32 %r969, -1;
shfl.sync.bfly.b32 %r970|%p111, %r966, %r968, %r967, %r969;
mov.b32 %f385, %r970;
add.ftz.f32 %f386, %f382, %f385;
mov.b32 %r971, %f386;
mov.u32 %r972, 8;
shfl.sync.bfly.b32 %r973|%p112, %r971, %r972, %r967, %r969;
mov.b32 %f387, %r973;
add.ftz.f32 %f388, %f386, %f387;
mov.b32 %r974, %f388;
mov.u32 %r975, 4;
shfl.sync.bfly.b32 %r976|%p113, %r974, %r975, %r967, %r969;
mov.b32 %f389, %r976;
add.ftz.f32 %f390, %f388, %f389;
mov.b32 %r977, %f390;
mov.u32 %r978, 2;
shfl.sync.bfly.b32 %r979|%p114, %r977, %r978, %r967, %r969;
mov.b32 %f391, %r979;
add.ftz.f32 %f392, %f390, %f391;
mov.b32 %r980, %f392;
mov.u32 %r981, 1;
shfl.sync.bfly.b32 %r982|%p115, %r980, %r981, %r967, %r969;
mov.b32 %f393, %r982;
add.ftz.f32 %f48, %f392, %f393;
mov.b32 %r983, %f384;
shfl.sync.bfly.b32 %r984|%p116, %r983, %r968, %r967, %r969;
mov.b32 %f394, %r984;
add.ftz.f32 %f395, %f384, %f394;
mov.b32 %r985, %f395;
shfl.sync.bfly.b32 %r986|%p117, %r985, %r972, %r967, %r969;
mov.b32 %f396, %r986;
add.ftz.f32 %f397, %f395, %f396;
mov.b32 %r987, %f397;
shfl.sync.bfly.b32 %r988|%p118, %r987, %r975, %r967, %r969;
mov.b32 %f398, %r988;
add.ftz.f32 %f399, %f397, %f398;
mov.b32 %r989, %f399;
shfl.sync.bfly.b32 %r990|%p119, %r989, %r978, %r967, %r969;
mov.b32 %f400, %r990;
add.ftz.f32 %f401, %f399, %f400;
mov.b32 %r991, %f401;
shfl.sync.bfly.b32 %r992|%p120, %r991, %r981, %r967, %r969;
mov.b32 %f402, %r992;
add.ftz.f32 %f49, %f401, %f402;
setp.ne.s32 %p121, %r2, 0;
setp.ge.u32 %p122, %r4, %r45;
or.pred %p123, %p121, %p122;
@%p123 bra $L__BB58_56;

add.ftz.f32 %f403, %f5, %f48;
add.ftz.f32 %f404, %f4, %f49;
mov.f32 %f405, 0f40E00000;
min.ftz.f32 %f406, %f404, %f405;
min.ftz.f32 %f407, %f403, %f405;
mov.f32 %f408, 0fC0E00000;
max.ftz.f32 %f409, %f407, %f408;
mul.ftz.f32 %f410, %f406, 0f3FD9DB23;
mul.ftz.f32 %f411, %f410, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f412, %f411;
add.ftz.f32 %f413, %f412, 0f3F800000;
div.approx.ftz.f32 %f414, %f406, %f413;
add.ftz.f32 %f415, %f409, 0f3F800000;
mul.ftz.f32 %f416, %f415, %f414;
st.global.f32 [%rd17], %f416;

$L__BB58_56:
ld.shared.f32 %f417, [%r26+128];
add.ftz.f32 %f418, %f417, %f28;
ld.shared.f32 %f419, [%r28+128];
add.ftz.f32 %f420, %f419, %f32;
ld.shared.f32 %f421, [%r26+640];
add.ftz.f32 %f422, %f421, %f418;
ld.shared.f32 %f423, [%r28+640];
add.ftz.f32 %f424, %f423, %f420;
ld.shared.f32 %f425, [%r26+1152];
add.ftz.f32 %f426, %f425, %f422;
ld.shared.f32 %f427, [%r28+1152];
add.ftz.f32 %f428, %f427, %f424;
mov.b32 %r996, %f426;
shfl.sync.bfly.b32 %r1000|%p124, %r996, %r968, %r967, %r969;
mov.b32 %f429, %r1000;
add.ftz.f32 %f430, %f426, %f429;
mov.b32 %r1001, %f430;
shfl.sync.bfly.b32 %r1003|%p125, %r1001, %r972, %r967, %r969;
mov.b32 %f431, %r1003;
add.ftz.f32 %f432, %f430, %f431;
mov.b32 %r1004, %f432;
shfl.sync.bfly.b32 %r1006|%p126, %r1004, %r975, %r967, %r969;
mov.b32 %f433, %r1006;
add.ftz.f32 %f434, %f432, %f433;
mov.b32 %r1007, %f434;
shfl.sync.bfly.b32 %r1009|%p127, %r1007, %r978, %r967, %r969;
mov.b32 %f435, %r1009;
add.ftz.f32 %f436, %f434, %f435;
mov.b32 %r1010, %f436;
shfl.sync.bfly.b32 %r1012|%p128, %r1010, %r981, %r967, %r969;
mov.b32 %f437, %r1012;
add.ftz.f32 %f50, %f436, %f437;
mov.b32 %r1013, %f428;
shfl.sync.bfly.b32 %r1014|%p129, %r1013, %r968, %r967, %r969;
mov.b32 %f438, %r1014;
add.ftz.f32 %f439, %f428, %f438;
mov.b32 %r1015, %f439;
shfl.sync.bfly.b32 %r1016|%p130, %r1015, %r972, %r967, %r969;
mov.b32 %f440, %r1016;
add.ftz.f32 %f441, %f439, %f440;
mov.b32 %r1017, %f441;
shfl.sync.bfly.b32 %r1018|%p131, %r1017, %r975, %r967, %r969;
mov.b32 %f442, %r1018;
add.ftz.f32 %f443, %f441, %f442;
mov.b32 %r1019, %f443;
shfl.sync.bfly.b32 %r1020|%p132, %r1019, %r978, %r967, %r969;
mov.b32 %f444, %r1020;
add.ftz.f32 %f445, %f443, %f444;
mov.b32 %r1021, %f445;
shfl.sync.bfly.b32 %r1022|%p133, %r1021, %r981, %r967, %r969;
mov.b32 %f446, %r1022;
add.ftz.f32 %f51, %f445, %f446;
setp.ne.s32 %p134, %r2, 1;
or.b32 %r1026, %r4, 1;
setp.ge.u32 %p135, %r1026, %r45;
or.pred %p136, %p134, %p135;
@%p136 bra $L__BB58_58;

add.ftz.f32 %f447, %f5, %f50;
add.ftz.f32 %f448, %f4, %f51;
mov.f32 %f449, 0f40E00000;
min.ftz.f32 %f450, %f448, %f449;
min.ftz.f32 %f451, %f447, %f449;
mov.f32 %f452, 0fC0E00000;
max.ftz.f32 %f453, %f451, %f452;
mul.ftz.f32 %f454, %f450, 0f3FD9DB23;
mul.ftz.f32 %f455, %f454, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f456, %f455;
add.ftz.f32 %f457, %f456, 0f3F800000;
div.approx.ftz.f32 %f458, %f450, %f457;
add.ftz.f32 %f459, %f453, 0f3F800000;
mul.ftz.f32 %f460, %f459, %f458;
st.global.f32 [%rd17+4], %f460;

$L__BB58_58:
ld.shared.f32 %f461, [%r26+256];
add.ftz.f32 %f462, %f461, %f27;
ld.shared.f32 %f463, [%r28+256];
add.ftz.f32 %f464, %f463, %f31;
ld.shared.f32 %f465, [%r26+768];
add.ftz.f32 %f466, %f465, %f462;
ld.shared.f32 %f467, [%r28+768];
add.ftz.f32 %f468, %f467, %f464;
ld.shared.f32 %f469, [%r26+1280];
add.ftz.f32 %f470, %f469, %f466;
ld.shared.f32 %f471, [%r28+1280];
add.ftz.f32 %f472, %f471, %f468;
mov.b32 %r1027, %f470;
mov.u32 %r1028, 31;
mov.u32 %r1029, 16;
mov.u32 %r1030, -1;
shfl.sync.bfly.b32 %r1031|%p137, %r1027, %r1029, %r1028, %r1030;
mov.b32 %f473, %r1031;
add.ftz.f32 %f474, %f470, %f473;
mov.b32 %r1032, %f474;
mov.u32 %r1033, 8;
shfl.sync.bfly.b32 %r1034|%p138, %r1032, %r1033, %r1028, %r1030;
mov.b32 %f475, %r1034;
add.ftz.f32 %f476, %f474, %f475;
mov.b32 %r1035, %f476;
mov.u32 %r1036, 4;
shfl.sync.bfly.b32 %r1037|%p139, %r1035, %r1036, %r1028, %r1030;
mov.b32 %f477, %r1037;
add.ftz.f32 %f478, %f476, %f477;
mov.b32 %r1038, %f478;
mov.u32 %r1039, 2;
shfl.sync.bfly.b32 %r1040|%p140, %r1038, %r1039, %r1028, %r1030;
mov.b32 %f479, %r1040;
add.ftz.f32 %f480, %f478, %f479;
mov.b32 %r1041, %f480;
mov.u32 %r1042, 1;
shfl.sync.bfly.b32 %r1043|%p141, %r1041, %r1042, %r1028, %r1030;
mov.b32 %f481, %r1043;
add.ftz.f32 %f52, %f480, %f481;
mov.b32 %r1044, %f472;
shfl.sync.bfly.b32 %r1045|%p142, %r1044, %r1029, %r1028, %r1030;
mov.b32 %f482, %r1045;
add.ftz.f32 %f483, %f472, %f482;
mov.b32 %r1046, %f483;
shfl.sync.bfly.b32 %r1047|%p143, %r1046, %r1033, %r1028, %r1030;
mov.b32 %f484, %r1047;
add.ftz.f32 %f485, %f483, %f484;
mov.b32 %r1048, %f485;
shfl.sync.bfly.b32 %r1049|%p144, %r1048, %r1036, %r1028, %r1030;
mov.b32 %f486, %r1049;
add.ftz.f32 %f487, %f485, %f486;
mov.b32 %r1050, %f487;
shfl.sync.bfly.b32 %r1051|%p145, %r1050, %r1039, %r1028, %r1030;
mov.b32 %f488, %r1051;
add.ftz.f32 %f489, %f487, %f488;
mov.b32 %r1052, %f489;
shfl.sync.bfly.b32 %r1053|%p146, %r1052, %r1042, %r1028, %r1030;
mov.b32 %f490, %r1053;
add.ftz.f32 %f53, %f489, %f490;
setp.ne.s32 %p147, %r2, 2;
or.b32 %r1057, %r4, 2;
setp.ge.u32 %p148, %r1057, %r45;
or.pred %p149, %p147, %p148;
@%p149 bra $L__BB58_60;

add.ftz.f32 %f491, %f5, %f52;
add.ftz.f32 %f492, %f4, %f53;
mov.f32 %f493, 0f40E00000;
min.ftz.f32 %f494, %f492, %f493;
min.ftz.f32 %f495, %f491, %f493;
mov.f32 %f496, 0fC0E00000;
max.ftz.f32 %f497, %f495, %f496;
mul.ftz.f32 %f498, %f494, 0f3FD9DB23;
mul.ftz.f32 %f499, %f498, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f500, %f499;
add.ftz.f32 %f501, %f500, 0f3F800000;
div.approx.ftz.f32 %f502, %f494, %f501;
add.ftz.f32 %f503, %f497, 0f3F800000;
mul.ftz.f32 %f504, %f503, %f502;
st.global.f32 [%rd17+8], %f504;

$L__BB58_60:
ld.shared.f32 %f505, [%r26+384];
add.ftz.f32 %f506, %f505, %f26;
ld.shared.f32 %f507, [%r28+384];
add.ftz.f32 %f508, %f507, %f30;
ld.shared.f32 %f509, [%r26+896];
add.ftz.f32 %f510, %f509, %f506;
ld.shared.f32 %f511, [%r28+896];
add.ftz.f32 %f512, %f511, %f508;
ld.shared.f32 %f513, [%r26+1408];
add.ftz.f32 %f514, %f513, %f510;
ld.shared.f32 %f515, [%r28+1408];
add.ftz.f32 %f516, %f515, %f512;
mov.b32 %r1058, %f514;
shfl.sync.bfly.b32 %r1062|%p150, %r1058, %r1029, %r1028, %r1030;
mov.b32 %f517, %r1062;
add.ftz.f32 %f518, %f514, %f517;
mov.b32 %r1063, %f518;
shfl.sync.bfly.b32 %r1065|%p151, %r1063, %r1033, %r1028, %r1030;
mov.b32 %f519, %r1065;
add.ftz.f32 %f520, %f518, %f519;
mov.b32 %r1066, %f520;
shfl.sync.bfly.b32 %r1068|%p152, %r1066, %r1036, %r1028, %r1030;
mov.b32 %f521, %r1068;
add.ftz.f32 %f522, %f520, %f521;
mov.b32 %r1069, %f522;
shfl.sync.bfly.b32 %r1071|%p153, %r1069, %r1039, %r1028, %r1030;
mov.b32 %f523, %r1071;
add.ftz.f32 %f524, %f522, %f523;
mov.b32 %r1072, %f524;
shfl.sync.bfly.b32 %r1074|%p154, %r1072, %r1042, %r1028, %r1030;
mov.b32 %f525, %r1074;
add.ftz.f32 %f54, %f524, %f525;
mov.b32 %r1075, %f516;
shfl.sync.bfly.b32 %r1076|%p155, %r1075, %r1029, %r1028, %r1030;
mov.b32 %f526, %r1076;
add.ftz.f32 %f527, %f516, %f526;
mov.b32 %r1077, %f527;
shfl.sync.bfly.b32 %r1078|%p156, %r1077, %r1033, %r1028, %r1030;
mov.b32 %f528, %r1078;
add.ftz.f32 %f529, %f527, %f528;
mov.b32 %r1079, %f529;
shfl.sync.bfly.b32 %r1080|%p157, %r1079, %r1036, %r1028, %r1030;
mov.b32 %f530, %r1080;
add.ftz.f32 %f531, %f529, %f530;
mov.b32 %r1081, %f531;
shfl.sync.bfly.b32 %r1082|%p158, %r1081, %r1039, %r1028, %r1030;
mov.b32 %f532, %r1082;
add.ftz.f32 %f533, %f531, %f532;
mov.b32 %r1083, %f533;
shfl.sync.bfly.b32 %r1084|%p159, %r1083, %r1042, %r1028, %r1030;
mov.b32 %f534, %r1084;
add.ftz.f32 %f55, %f533, %f534;
setp.ne.s32 %p160, %r2, 3;
or.b32 %r1088, %r4, 3;
setp.ge.u32 %p161, %r1088, %r45;
or.pred %p162, %p160, %p161;
@%p162 bra $L__BB58_78;

add.ftz.f32 %f535, %f5, %f54;
add.ftz.f32 %f536, %f4, %f55;
mov.f32 %f537, 0f40E00000;
min.ftz.f32 %f538, %f536, %f537;
min.ftz.f32 %f539, %f535, %f537;
mov.f32 %f540, 0fC0E00000;
max.ftz.f32 %f541, %f539, %f540;
mul.ftz.f32 %f542, %f538, 0f3FD9DB23;
mul.ftz.f32 %f543, %f542, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f544, %f543;
add.ftz.f32 %f545, %f544, 0f3F800000;
div.approx.ftz.f32 %f546, %f538, %f545;
add.ftz.f32 %f547, %f541, 0f3F800000;
mul.ftz.f32 %f548, %f547, %f546;
st.global.f32 [%rd17+12], %f548;
bra.uni $L__BB58_78;

}
.entry _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<38>;
.reg .b16 %rs<29>;
.reg .f32 %f<121>;
.reg .b32 %r<441>;
.reg .b64 %rd<49>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

ld.param.u64 %rd13, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd14, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r19, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
mov.u32 %r1, %tid.y;
shl.b32 %r37, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r37, %r2;
mov.u32 %r38, %ctaid.x;
shl.b32 %r4, %r38, 2;
shr.u32 %r5, %r36, 5;
mov.u32 %r439, %ctaid.y;
setp.eq.s64 %p1, %rd15, 0;
@%p1 bra $L__BB59_2;

cvta.to.global.u64 %rd17, %rd15;
mul.wide.u32 %rd18, %r439, 4;
add.s64 %rd19, %rd17, %rd18;
ld.global.nc.u32 %r438, [%rd19];
mul.hi.u32 %r39, %r439, %r19;
add.s32 %r40, %r39, %r439;
shr.u32 %r41, %r40, %r20;
mul.lo.s32 %r42, %r41, %r21;
sub.s32 %r439, %r439, %r42;
bra.uni $L__BB59_3;

$L__BB59_2:
mul.hi.u32 %r43, %r439, %r24;
add.s32 %r44, %r43, %r439;
shr.u32 %r438, %r44, %r25;

$L__BB59_3:
shr.u32 %r45, %r3, 31;
add.s32 %r46, %r3, %r45;
shr.s32 %r440, %r46, 1;
setp.ge.s32 %p2, %r440, %r5;
mov.f32 %f117, 0f00000000;
mov.f32 %f118, %f117;
mov.f32 %f119, %f117;
mov.f32 %f120, %f117;
@%p2 bra $L__BB59_6;

mov.u32 %r47, %ctaid.z;
and.b32 %r50, %r46, -2;
sub.s32 %r51, %r3, %r50;
shl.b32 %r52, %r51, 1;
shl.b32 %r53, %r51, 2;
or.b32 %r54, %r53, 1;
shl.b32 %r13, %r51, 3;
cvt.s64.s32 %rd1, %r52;
sub.s32 %r55, %r54, %r52;
shl.b32 %r14, %r55, 2;
cvt.s64.s32 %rd2, %r55;
mad.lo.s32 %r56, %r4, %r22, %r440;
mul.hi.u32 %r57, %r47, %r30;
add.s32 %r58, %r57, %r47;
shr.u32 %r59, %r58, %r31;
mad.lo.s32 %r60, %r59, %r33, %r56;
mad.lo.s32 %r61, %r438, %r27, %r60;
mul.wide.s32 %rd3, %r61, 22;
mul.wide.s32 %rd20, %r53, 2;
add.s64 %rd21, %rd3, %rd20;
add.s64 %rd4, %rd21, 8;
mul.wide.s32 %rd5, %r22, 22;
add.s64 %rd6, %rd5, 2;
cvta.to.global.u64 %rd7, %rd14;
cvta.to.global.u64 %rd48, %rd13;
mul.lo.s32 %r62, %r439, %r28;
cvt.u64.u32 %rd22, %r62;
mul.lo.s32 %r63, %r47, %r34;
cvt.u64.u32 %rd23, %r63;
add.s64 %rd9, %rd22, %rd23;
mov.f32 %f117, 0f00000000;
shl.b64 %rd33, %rd1, 2;
shl.b64 %rd35, %rd2, 2;

$L__BB59_5:
cvt.u64.u32 %rd24, %r440;
add.s64 %rd25, %rd9, %rd24;
mul.lo.s64 %rd26, %rd25, 36;
add.s64 %rd27, %rd7, %rd26;
add.s64 %rd28, %rd48, %rd3;
add.s64 %rd29, %rd28, 2;
ld.global.nc.u16 %rs5, [%rd28+2];
ld.global.nc.u16 %rs6, [%rd28+4];
mov.b32 %r136, {%rs5, %rs6};
add.s64 %rd30, %rd27, 4;
add.s64 %rd31, %rd48, %rd4;
add.s64 %rd32, %rd31, -2;
ld.global.nc.u16 %rs7, [%rd31+-2];
ld.global.nc.u16 %rs8, [%rd31];
mov.b32 %r137, {%rs7, %rs8};
shr.s32 %r138, %r136, %r13;
add.s64 %rd34, %rd30, %rd33;
ld.global.nc.u32 %r66, [%rd34];
ld.global.nc.u32 %r70, [%rd34+16];
ld.global.nc.u16 %rs9, [%rd31+2];
ld.global.nc.u16 %rs10, [%rd31+4];
mov.b32 %r139, {%rs9, %rs10};
shr.s32 %r140, %r136, %r14;
add.s64 %rd36, %rd30, %rd35;
ld.global.nc.u32 %r74, [%rd36];
ld.global.nc.u32 %r78, [%rd34+20];
ld.global.nc.u16 %rs1, [%rd28];

{ cvt.f32.f16 %f25, %rs1;}


and.b32 %r141, %r137, 252645135;
shl.b32 %r142, %r138, 4;
and.b32 %r143, %r142, 16;
or.b32 %r144, %r143, %r141;
shl.b32 %r145, %r138, 11;
and.b32 %r146, %r145, 4096;
or.b32 %r147, %r144, %r146;
shl.b32 %r148, %r138, 18;
and.b32 %r149, %r148, 1048576;
or.b32 %r150, %r147, %r149;
shl.b32 %r151, %r138, 25;
and.b32 %r152, %r151, 268435456;
or.b32 %r65, %r150, %r152;
mov.u32 %r121, 0;

dp4a.s32.s32 %r64, %r65, %r66, %r121;

shr.u32 %r153, %r137, 4;
and.b32 %r154, %r153, 252645135;
shr.u32 %r155, %r138, 12;
and.b32 %r156, %r155, 16;
or.b32 %r157, %r156, %r154;
shr.u32 %r158, %r138, 5;
and.b32 %r159, %r158, 4096;
or.b32 %r160, %r157, %r159;
shl.b32 %r161, %r138, 2;
and.b32 %r162, %r161, 1048576;
or.b32 %r163, %r160, %r162;
shl.b32 %r164, %r138, 9;
and.b32 %r165, %r164, 268435456;
or.b32 %r69, %r163, %r165;

dp4a.s32.s32 %r68, %r69, %r70, %r64;

and.b32 %r166, %r139, 252645135;
shl.b32 %r167, %r140, 4;
and.b32 %r168, %r167, 16;
or.b32 %r169, %r168, %r166;
shl.b32 %r170, %r140, 11;
and.b32 %r171, %r170, 4096;
or.b32 %r172, %r169, %r171;
shl.b32 %r173, %r140, 18;
and.b32 %r174, %r173, 1048576;
or.b32 %r175, %r172, %r174;
shl.b32 %r176, %r140, 25;
and.b32 %r177, %r176, 268435456;
or.b32 %r73, %r175, %r177;

dp4a.s32.s32 %r72, %r73, %r74, %r68;

shr.u32 %r178, %r139, 4;
and.b32 %r179, %r178, 252645135;
shr.u32 %r180, %r140, 12;
and.b32 %r181, %r180, 16;
or.b32 %r182, %r181, %r179;
shr.u32 %r183, %r140, 5;
and.b32 %r184, %r183, 4096;
or.b32 %r185, %r182, %r184;
shl.b32 %r186, %r140, 2;
and.b32 %r187, %r186, 1048576;
or.b32 %r188, %r185, %r187;
shl.b32 %r189, %r140, 9;
and.b32 %r190, %r189, 268435456;
or.b32 %r77, %r188, %r190;

dp4a.s32.s32 %r76, %r77, %r78, %r72;

ld.global.nc.u32 %r80, [%rd27];

{.reg .f16 low,high;
mov.b32 {low,high},%r80;
cvt.f32.f16 %f26, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r80;
cvt.f32.f16 %f27, high;}


cvt.rn.f32.s32 %f37, %r76;
mul.ftz.f32 %f38, %f26, %f37;
mul.ftz.f32 %f39, %f27, 0f41000000;
sub.ftz.f32 %f40, %f38, %f39;
fma.rn.ftz.f32 %f120, %f25, %f40, %f120;
add.s64 %rd37, %rd29, %rd5;
add.s64 %rd38, %rd37, -2;
add.s64 %rd39, %rd28, %rd5;
ld.global.nc.u16 %rs11, [%rd39+2];
ld.global.nc.u16 %rs12, [%rd39+4];
mov.b32 %r191, {%rs11, %rs12};
add.s64 %rd40, %rd32, %rd5;
ld.global.nc.u16 %rs13, [%rd40];
ld.global.nc.u16 %rs14, [%rd40+2];
mov.b32 %r192, {%rs13, %rs14};
shr.s32 %r193, %r191, %r13;
ld.global.nc.u16 %rs15, [%rd40+4];
ld.global.nc.u16 %rs16, [%rd40+6];
mov.b32 %r194, {%rs15, %rs16};
shr.s32 %r195, %r191, %r14;
ld.global.nc.u16 %rs2, [%rd39];

{ cvt.f32.f16 %f28, %rs2;}


and.b32 %r196, %r192, 252645135;
shl.b32 %r197, %r193, 4;
and.b32 %r198, %r197, 16;
or.b32 %r199, %r198, %r196;
shl.b32 %r200, %r193, 11;
and.b32 %r201, %r200, 4096;
or.b32 %r202, %r199, %r201;
shl.b32 %r203, %r193, 18;
and.b32 %r204, %r203, 1048576;
or.b32 %r205, %r202, %r204;
shl.b32 %r206, %r193, 25;
and.b32 %r207, %r206, 268435456;
or.b32 %r83, %r205, %r207;

dp4a.s32.s32 %r82, %r83, %r66, %r121;

shr.u32 %r208, %r192, 4;
and.b32 %r209, %r208, 252645135;
shr.u32 %r210, %r193, 12;
and.b32 %r211, %r210, 16;
or.b32 %r212, %r211, %r209;
shr.u32 %r213, %r193, 5;
and.b32 %r214, %r213, 4096;
or.b32 %r215, %r212, %r214;
shl.b32 %r216, %r193, 2;
and.b32 %r217, %r216, 1048576;
or.b32 %r218, %r215, %r217;
shl.b32 %r219, %r193, 9;
and.b32 %r220, %r219, 268435456;
or.b32 %r87, %r218, %r220;

dp4a.s32.s32 %r86, %r87, %r70, %r82;

and.b32 %r221, %r194, 252645135;
shl.b32 %r222, %r195, 4;
and.b32 %r223, %r222, 16;
or.b32 %r224, %r223, %r221;
shl.b32 %r225, %r195, 11;
and.b32 %r226, %r225, 4096;
or.b32 %r227, %r224, %r226;
shl.b32 %r228, %r195, 18;
and.b32 %r229, %r228, 1048576;
or.b32 %r230, %r227, %r229;
shl.b32 %r231, %r195, 25;
and.b32 %r232, %r231, 268435456;
or.b32 %r91, %r230, %r232;

dp4a.s32.s32 %r90, %r91, %r74, %r86;

shr.u32 %r233, %r194, 4;
and.b32 %r234, %r233, 252645135;
shr.u32 %r235, %r195, 12;
and.b32 %r236, %r235, 16;
or.b32 %r237, %r236, %r234;
shr.u32 %r238, %r195, 5;
and.b32 %r239, %r238, 4096;
or.b32 %r240, %r237, %r239;
shl.b32 %r241, %r195, 2;
and.b32 %r242, %r241, 1048576;
or.b32 %r243, %r240, %r242;
shl.b32 %r244, %r195, 9;
and.b32 %r245, %r244, 268435456;
or.b32 %r95, %r243, %r245;

dp4a.s32.s32 %r94, %r95, %r78, %r90;


{.reg .f16 low,high;
mov.b32 {low,high},%r80;
cvt.f32.f16 %f29, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r80;
cvt.f32.f16 %f30, high;}


cvt.rn.f32.s32 %f41, %r94;
mul.ftz.f32 %f42, %f29, %f41;
mul.ftz.f32 %f43, %f30, 0f41000000;
sub.ftz.f32 %f44, %f42, %f43;
fma.rn.ftz.f32 %f119, %f28, %f44, %f119;
add.s64 %rd41, %rd38, %rd5;
add.s64 %rd42, %rd38, %rd6;
ld.global.nc.u16 %rs17, [%rd42];
ld.global.nc.u16 %rs18, [%rd42+2];
mov.b32 %r246, {%rs17, %rs18};
add.s64 %rd43, %rd40, %rd5;
ld.global.nc.u16 %rs19, [%rd43];
ld.global.nc.u16 %rs20, [%rd43+2];
mov.b32 %r247, {%rs19, %rs20};
shr.s32 %r248, %r246, %r13;
ld.global.nc.u16 %rs21, [%rd43+4];
ld.global.nc.u16 %rs22, [%rd43+6];
mov.b32 %r249, {%rs21, %rs22};
shr.s32 %r250, %r246, %r14;
ld.global.nc.u16 %rs3, [%rd42+-2];

{ cvt.f32.f16 %f31, %rs3;}


and.b32 %r251, %r247, 252645135;
shl.b32 %r252, %r248, 4;
and.b32 %r253, %r252, 16;
or.b32 %r254, %r253, %r251;
shl.b32 %r255, %r248, 11;
and.b32 %r256, %r255, 4096;
or.b32 %r257, %r254, %r256;
shl.b32 %r258, %r248, 18;
and.b32 %r259, %r258, 1048576;
or.b32 %r260, %r257, %r259;
shl.b32 %r261, %r248, 25;
and.b32 %r262, %r261, 268435456;
or.b32 %r101, %r260, %r262;

dp4a.s32.s32 %r100, %r101, %r66, %r121;

shr.u32 %r263, %r247, 4;
and.b32 %r264, %r263, 252645135;
shr.u32 %r265, %r248, 12;
and.b32 %r266, %r265, 16;
or.b32 %r267, %r266, %r264;
shr.u32 %r268, %r248, 5;
and.b32 %r269, %r268, 4096;
or.b32 %r270, %r267, %r269;
shl.b32 %r271, %r248, 2;
and.b32 %r272, %r271, 1048576;
or.b32 %r273, %r270, %r272;
shl.b32 %r274, %r248, 9;
and.b32 %r275, %r274, 268435456;
or.b32 %r105, %r273, %r275;

dp4a.s32.s32 %r104, %r105, %r70, %r100;

and.b32 %r276, %r249, 252645135;
shl.b32 %r277, %r250, 4;
and.b32 %r278, %r277, 16;
or.b32 %r279, %r278, %r276;
shl.b32 %r280, %r250, 11;
and.b32 %r281, %r280, 4096;
or.b32 %r282, %r279, %r281;
shl.b32 %r283, %r250, 18;
and.b32 %r284, %r283, 1048576;
or.b32 %r285, %r282, %r284;
shl.b32 %r286, %r250, 25;
and.b32 %r287, %r286, 268435456;
or.b32 %r109, %r285, %r287;

dp4a.s32.s32 %r108, %r109, %r74, %r104;

shr.u32 %r288, %r249, 4;
and.b32 %r289, %r288, 252645135;
shr.u32 %r290, %r250, 12;
and.b32 %r291, %r290, 16;
or.b32 %r292, %r291, %r289;
shr.u32 %r293, %r250, 5;
and.b32 %r294, %r293, 4096;
or.b32 %r295, %r292, %r294;
shl.b32 %r296, %r250, 2;
and.b32 %r297, %r296, 1048576;
or.b32 %r298, %r295, %r297;
shl.b32 %r299, %r250, 9;
and.b32 %r300, %r299, 268435456;
or.b32 %r113, %r298, %r300;

dp4a.s32.s32 %r112, %r113, %r78, %r108;


{.reg .f16 low,high;
mov.b32 {low,high},%r80;
cvt.f32.f16 %f32, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r80;
cvt.f32.f16 %f33, high;}


cvt.rn.f32.s32 %f45, %r112;
mul.ftz.f32 %f46, %f32, %f45;
mul.ftz.f32 %f47, %f33, 0f41000000;
sub.ftz.f32 %f48, %f46, %f47;
fma.rn.ftz.f32 %f118, %f31, %f48, %f118;
add.s64 %rd44, %rd41, %rd6;
ld.global.nc.u16 %rs23, [%rd44];
ld.global.nc.u16 %rs24, [%rd44+2];
mov.b32 %r301, {%rs23, %rs24};
add.s64 %rd45, %rd43, %rd5;
ld.global.nc.u16 %rs25, [%rd45];
ld.global.nc.u16 %rs26, [%rd45+2];
mov.b32 %r302, {%rs25, %rs26};
shr.s32 %r303, %r301, %r13;
ld.global.nc.u16 %rs27, [%rd45+4];
ld.global.nc.u16 %rs28, [%rd45+6];
mov.b32 %r304, {%rs27, %rs28};
shr.s32 %r305, %r301, %r14;
ld.global.nc.u16 %rs4, [%rd44+-2];

{ cvt.f32.f16 %f34, %rs4;}


and.b32 %r306, %r302, 252645135;
shl.b32 %r307, %r303, 4;
and.b32 %r308, %r307, 16;
or.b32 %r309, %r308, %r306;
shl.b32 %r310, %r303, 11;
and.b32 %r311, %r310, 4096;
or.b32 %r312, %r309, %r311;
shl.b32 %r313, %r303, 18;
and.b32 %r314, %r313, 1048576;
or.b32 %r315, %r312, %r314;
shl.b32 %r316, %r303, 25;
and.b32 %r317, %r316, 268435456;
or.b32 %r119, %r315, %r317;

dp4a.s32.s32 %r118, %r119, %r66, %r121;

shr.u32 %r318, %r302, 4;
and.b32 %r319, %r318, 252645135;
shr.u32 %r320, %r303, 12;
and.b32 %r321, %r320, 16;
or.b32 %r322, %r321, %r319;
shr.u32 %r323, %r303, 5;
and.b32 %r324, %r323, 4096;
or.b32 %r325, %r322, %r324;
shl.b32 %r326, %r303, 2;
and.b32 %r327, %r326, 1048576;
or.b32 %r328, %r325, %r327;
shl.b32 %r329, %r303, 9;
and.b32 %r330, %r329, 268435456;
or.b32 %r123, %r328, %r330;

dp4a.s32.s32 %r122, %r123, %r70, %r118;

and.b32 %r331, %r304, 252645135;
shl.b32 %r332, %r305, 4;
and.b32 %r333, %r332, 16;
or.b32 %r334, %r333, %r331;
shl.b32 %r335, %r305, 11;
and.b32 %r336, %r335, 4096;
or.b32 %r337, %r334, %r336;
shl.b32 %r338, %r305, 18;
and.b32 %r339, %r338, 1048576;
or.b32 %r340, %r337, %r339;
shl.b32 %r341, %r305, 25;
and.b32 %r342, %r341, 268435456;
or.b32 %r127, %r340, %r342;

dp4a.s32.s32 %r126, %r127, %r74, %r122;

shr.u32 %r343, %r304, 4;
and.b32 %r344, %r343, 252645135;
shr.u32 %r345, %r305, 12;
and.b32 %r346, %r345, 16;
or.b32 %r347, %r346, %r344;
shr.u32 %r348, %r305, 5;
and.b32 %r349, %r348, 4096;
or.b32 %r350, %r347, %r349;
shl.b32 %r351, %r305, 2;
and.b32 %r352, %r351, 1048576;
or.b32 %r353, %r350, %r352;
shl.b32 %r354, %r305, 9;
and.b32 %r355, %r354, 268435456;
or.b32 %r131, %r353, %r355;

dp4a.s32.s32 %r130, %r131, %r78, %r126;


{.reg .f16 low,high;
mov.b32 {low,high},%r80;
cvt.f32.f16 %f35, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r80;
cvt.f32.f16 %f36, high;}


cvt.rn.f32.s32 %f49, %r130;
mul.ftz.f32 %f50, %f35, %f49;
mul.ftz.f32 %f51, %f36, 0f41000000;
sub.ftz.f32 %f52, %f50, %f51;
fma.rn.ftz.f32 %f117, %f34, %f52, %f117;
add.s64 %rd48, %rd48, 1408;
add.s32 %r440, %r440, 64;
setp.lt.s32 %p3, %r440, %r5;
@%p3 bra $L__BB59_5;

$L__BB59_6:
setp.eq.s32 %p4, %r1, 0;
@%p4 bra $L__BB59_8;

shl.b32 %r356, %r1, 9;
mov.u32 %r357, _ZZ13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r358, %r357, %r356;
shl.b32 %r359, %r2, 2;
add.s32 %r360, %r358, %r359;
st.shared.f32 [%r360+-512], %f120;
st.shared.f32 [%r360+-384], %f119;
st.shared.f32 [%r360+-256], %f118;
st.shared.f32 [%r360+-128], %f117;

$L__BB59_8:
setp.ne.s32 %p5, %r1, 0;
bar.sync 0;
@%p5 bra $L__BB59_17;

mov.u32 %r361, %ctaid.z;
mov.u32 %r362, %ctaid.y;
mad.lo.s32 %r363, %r362, %r29, %r4;
mad.lo.s32 %r364, %r361, %r35, %r363;
shl.b32 %r365, %r2, 2;
mov.u32 %r366, _ZZ13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r17, %r366, %r365;
mov.u32 %r367, 2;
ld.shared.f32 %f53, [%r17];
add.ftz.f32 %f54, %f53, %f120;
ld.shared.f32 %f55, [%r17+512];
add.ftz.f32 %f56, %f55, %f54;
ld.shared.f32 %f57, [%r17+1024];
add.ftz.f32 %f58, %f57, %f56;
mov.b32 %r368, %f58;
mov.u32 %r369, 31;
mov.u32 %r370, 16;
mov.u32 %r371, -1;
shfl.sync.bfly.b32 %r372|%p6, %r368, %r370, %r369, %r371;
mov.b32 %f59, %r372;
add.ftz.f32 %f60, %f58, %f59;
mov.b32 %r373, %f60;
mov.u32 %r374, 8;
shfl.sync.bfly.b32 %r375|%p7, %r373, %r374, %r369, %r371;
mov.b32 %f61, %r375;
add.ftz.f32 %f62, %f60, %f61;
mov.b32 %r376, %f62;
mov.u32 %r377, 4;
shfl.sync.bfly.b32 %r378|%p8, %r376, %r377, %r369, %r371;
mov.b32 %f63, %r378;
add.ftz.f32 %f64, %f62, %f63;
mov.b32 %r379, %f64;
shfl.sync.bfly.b32 %r380|%p9, %r379, %r367, %r369, %r371;
mov.b32 %f65, %r380;
add.ftz.f32 %f66, %f64, %f65;
mov.b32 %r381, %f66;
mov.u32 %r382, 1;
shfl.sync.bfly.b32 %r383|%p10, %r381, %r382, %r369, %r371;
mov.b32 %f67, %r383;
add.ftz.f32 %f13, %f66, %f67;
setp.ge.u32 %p11, %r4, %r23;
setp.ne.s32 %p12, %r2, 0;
cvta.to.global.u64 %rd46, %rd16;
mul.wide.u32 %rd47, %r364, 4;
add.s64 %rd12, %rd46, %rd47;
or.pred %p13, %p12, %p11;
@%p13 bra $L__BB59_11;

st.global.f32 [%rd12], %f13;

$L__BB59_11:
ld.shared.f32 %f68, [%r17+128];
add.ftz.f32 %f69, %f68, %f119;
ld.shared.f32 %f70, [%r17+640];
add.ftz.f32 %f71, %f70, %f69;
ld.shared.f32 %f72, [%r17+1152];
add.ftz.f32 %f73, %f72, %f71;
mov.b32 %r384, %f73;
shfl.sync.bfly.b32 %r388|%p14, %r384, %r370, %r369, %r371;
mov.b32 %f74, %r388;
add.ftz.f32 %f75, %f73, %f74;
mov.b32 %r389, %f75;
shfl.sync.bfly.b32 %r391|%p15, %r389, %r374, %r369, %r371;
mov.b32 %f76, %r391;
add.ftz.f32 %f77, %f75, %f76;
mov.b32 %r392, %f77;
shfl.sync.bfly.b32 %r394|%p16, %r392, %r377, %r369, %r371;
mov.b32 %f78, %r394;
add.ftz.f32 %f79, %f77, %f78;
mov.b32 %r395, %f79;
shfl.sync.bfly.b32 %r397|%p17, %r395, %r367, %r369, %r371;
mov.b32 %f80, %r397;
add.ftz.f32 %f81, %f79, %f80;
mov.b32 %r398, %f81;
shfl.sync.bfly.b32 %r400|%p18, %r398, %r382, %r369, %r371;
mov.b32 %f82, %r400;
add.ftz.f32 %f14, %f81, %f82;
add.s32 %r401, %r4, 1;
setp.ge.u32 %p19, %r401, %r23;
setp.ne.s32 %p20, %r2, 1;
or.pred %p21, %p20, %p19;
@%p21 bra $L__BB59_13;

st.global.f32 [%rd12+4], %f14;

$L__BB59_13:
ld.shared.f32 %f83, [%r17+256];
add.ftz.f32 %f84, %f83, %f118;
ld.shared.f32 %f85, [%r17+768];
add.ftz.f32 %f86, %f85, %f84;
ld.shared.f32 %f87, [%r17+1280];
add.ftz.f32 %f88, %f87, %f86;
mov.b32 %r402, %f88;
mov.u32 %r403, 31;
mov.u32 %r404, 16;
mov.u32 %r405, -1;
shfl.sync.bfly.b32 %r406|%p22, %r402, %r404, %r403, %r405;
mov.b32 %f89, %r406;
add.ftz.f32 %f90, %f88, %f89;
mov.b32 %r407, %f90;
mov.u32 %r408, 8;
shfl.sync.bfly.b32 %r409|%p23, %r407, %r408, %r403, %r405;
mov.b32 %f91, %r409;
add.ftz.f32 %f92, %f90, %f91;
mov.b32 %r410, %f92;
mov.u32 %r411, 4;
shfl.sync.bfly.b32 %r412|%p24, %r410, %r411, %r403, %r405;
mov.b32 %f93, %r412;
add.ftz.f32 %f94, %f92, %f93;
mov.b32 %r413, %f94;
mov.u32 %r414, 2;
shfl.sync.bfly.b32 %r415|%p25, %r413, %r414, %r403, %r405;
mov.b32 %f95, %r415;
add.ftz.f32 %f96, %f94, %f95;
mov.b32 %r416, %f96;
mov.u32 %r417, 1;
shfl.sync.bfly.b32 %r418|%p26, %r416, %r417, %r403, %r405;
mov.b32 %f97, %r418;
add.ftz.f32 %f15, %f96, %f97;
add.s32 %r419, %r4, 2;
setp.ge.u32 %p27, %r419, %r23;
setp.ne.s32 %p28, %r2, 2;
or.pred %p29, %p28, %p27;
@%p29 bra $L__BB59_15;

st.global.f32 [%rd12+8], %f15;

$L__BB59_15:
ld.shared.f32 %f98, [%r17+384];
add.ftz.f32 %f99, %f98, %f117;
ld.shared.f32 %f100, [%r17+896];
add.ftz.f32 %f101, %f100, %f99;
ld.shared.f32 %f102, [%r17+1408];
add.ftz.f32 %f103, %f102, %f101;
mov.b32 %r420, %f103;
shfl.sync.bfly.b32 %r424|%p30, %r420, %r404, %r403, %r405;
mov.b32 %f104, %r424;
add.ftz.f32 %f105, %f103, %f104;
mov.b32 %r425, %f105;
shfl.sync.bfly.b32 %r427|%p31, %r425, %r408, %r403, %r405;
mov.b32 %f106, %r427;
add.ftz.f32 %f107, %f105, %f106;
mov.b32 %r428, %f107;
shfl.sync.bfly.b32 %r430|%p32, %r428, %r411, %r403, %r405;
mov.b32 %f108, %r430;
add.ftz.f32 %f109, %f107, %f108;
mov.b32 %r431, %f109;
shfl.sync.bfly.b32 %r433|%p33, %r431, %r414, %r403, %r405;
mov.b32 %f110, %r433;
add.ftz.f32 %f111, %f109, %f110;
mov.b32 %r434, %f111;
shfl.sync.bfly.b32 %r436|%p34, %r434, %r417, %r403, %r405;
mov.b32 %f112, %r436;
add.ftz.f32 %f16, %f111, %f112;
add.s32 %r437, %r4, 3;
setp.ge.u32 %p35, %r437, %r23;
setp.ne.s32 %p36, %r2, 3;
or.pred %p37, %p36, %p35;
@%p37 bra $L__BB59_17;

st.global.f32 [%rd12+12], %f16;

$L__BB59_17:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<33>;
.reg .b16 %rs<15>;
.reg .f32 %f<124>;
.reg .b32 %r<294>;
.reg .b64 %rd<69>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[384];

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate[384];

ld.param.u64 %rd10, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd11, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd12, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
ld.param.f32 %f31, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+16];
ld.param.u64 %rd14, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+8];
ld.param.u64 %rd13, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3];
mov.u32 %r1, %tid.y;
shl.b32 %r43, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r43, %r2;
mov.u32 %r4, %ctaid.x;
mov.u32 %r5, %ctaid.y;
setp.eq.s64 %p2, %rd12, 0;
@%p2 bra $L__BB60_2;

cvta.to.global.u64 %rd19, %rd12;
mul.wide.u32 %rd20, %r5, 4;
add.s64 %rd21, %rd19, %rd20;
ld.global.nc.u32 %r291, [%rd21];
mul.hi.u32 %r44, %r5, %r27;
add.s32 %r45, %r44, %r5;
shr.u32 %r46, %r45, %r28;
mul.lo.s32 %r47, %r46, %r29;
sub.s32 %r292, %r5, %r47;
bra.uni $L__BB60_3;

$L__BB60_2:
mul.hi.u32 %r48, %r5, %r31;
add.s32 %r49, %r48, %r5;
shr.u32 %r291, %r49, %r32;
mov.u32 %r292, %r5;

$L__BB60_3:
selp.b32 %r11, %r5, %r291, %p2;
or.b32 %r50, %r2, %r1;
setp.ne.s32 %p4, %r50, 0;
mov.f32 %f117, 0f00000000;
mov.f32 %f113, %f117;
mov.f32 %f114, %f117;
@%p4 bra $L__BB60_8;

setp.eq.s64 %p5, %rd13, 0;
mov.f32 %f113, 0f00000000;
mov.f32 %f114, %f113;
@%p5 bra $L__BB60_6;

cvta.to.global.u64 %rd22, %rd13;
mov.u32 %r51, %ctaid.z;
mul.lo.s32 %r52, %r51, %r42;
cvt.u64.u32 %rd23, %r52;
mul.lo.s32 %r53, %r11, %r36;
cvt.u64.u32 %rd24, %r53;
cvt.u64.u32 %rd25, %r2;
cvt.s64.s32 %rd26, %r4;
add.s64 %rd27, %rd25, %rd26;
add.s64 %rd28, %rd27, %rd23;
add.s64 %rd29, %rd28, %rd24;
shl.b64 %rd30, %rd29, 2;
add.s64 %rd31, %rd22, %rd30;
ld.global.f32 %f114, [%rd31];

$L__BB60_6:
setp.eq.s64 %p6, %rd15, 0;
setp.eq.s64 %p7, %rd14, 0;
or.pred %p8, %p6, %p7;
@%p8 bra $L__BB60_8;

cvta.to.global.u64 %rd32, %rd15;
mov.u32 %r55, %ctaid.z;
mul.lo.s32 %r56, %r55, %r42;
cvt.u64.u32 %rd33, %r56;
mul.lo.s32 %r57, %r11, %r36;
cvt.u64.u32 %rd34, %r57;
cvt.u64.u32 %rd35, %r2;
cvt.s64.s32 %rd36, %r4;
add.s64 %rd37, %rd35, %rd36;
add.s64 %rd38, %rd37, %rd33;
add.s64 %rd39, %rd38, %rd34;
shl.b64 %rd40, %rd39, 2;
add.s64 %rd41, %rd32, %rd40;
ld.global.f32 %f113, [%rd41];

$L__BB60_8:
shr.u32 %r59, %r26, 5;
shr.u32 %r60, %r3, 31;
add.s32 %r61, %r3, %r60;
shr.s32 %r293, %r61, 1;
setp.ge.s32 %p9, %r293, %r59;
mov.f32 %f119, %f117;
@%p9 bra $L__BB60_13;

mov.u32 %r62, %ctaid.z;
and.b32 %r65, %r61, -2;
sub.s32 %r13, %r3, %r65;
shl.b32 %r66, %r13, 1;
shl.b32 %r67, %r13, 2;
or.b32 %r68, %r67, 1;
sub.s32 %r69, %r68, %r66;
shl.b32 %r14, %r69, 2;
cvt.s64.s32 %rd4, %r69;
mul.hi.u32 %r70, %r62, %r37;
add.s32 %r71, %r70, %r62;
shr.u32 %r72, %r71, %r38;
mad.lo.s32 %r73, %r72, %r40, %r293;
mad.lo.s32 %r74, %r291, %r34, %r73;
mad.lo.s32 %r75, %r4, %r30, %r74;
mul.wide.s32 %rd68, %r75, 22;
mul.wide.s32 %rd42, %r67, 2;
cvta.to.global.u64 %rd43, %rd14;
add.s64 %rd44, %rd43, %rd42;
add.s64 %rd6, %rd44, 6;
mul.lo.s32 %r76, %r292, %r35;
cvt.u64.u32 %rd45, %r76;
mul.lo.s32 %r77, %r62, %r41;
cvt.u64.u32 %rd46, %r77;
add.s64 %rd7, %rd45, %rd46;
mov.f32 %f119, 0f00000000;
cvta.to.global.u64 %rd47, %rd11;
cvta.to.global.u64 %rd50, %rd10;
shl.b64 %rd60, %rd4, 2;
mov.f32 %f117, %f119;

$L__BB60_10:
cvt.u64.u32 %rd48, %r293;
add.s64 %rd49, %rd7, %rd48;
add.s64 %rd51, %rd50, %rd68;
ld.global.nc.u16 %rs2, [%rd51+2];
ld.global.nc.u16 %rs3, [%rd51+4];
mov.b32 %r96, {%rs2, %rs3};
mul.lo.s64 %rd52, %rd49, 36;
add.s64 %rd53, %rd47, %rd52;
add.s64 %rd54, %rd53, 4;
add.s64 %rd56, %rd50, %rd42;
add.s64 %rd57, %rd56, %rd68;
ld.global.nc.u16 %rs4, [%rd57+6];
ld.global.nc.u16 %rs5, [%rd57+8];
mov.b32 %r98, {%rs4, %rs5};
shl.b32 %r99, %r13, 3;
shr.s32 %r100, %r96, %r99;
mul.wide.s32 %rd58, %r66, 4;
add.s64 %rd59, %rd54, %rd58;
ld.global.nc.u32 %r16, [%rd59];
ld.global.nc.u32 %r17, [%rd59+16];
ld.global.nc.u16 %rs6, [%rd57+10];
ld.global.nc.u16 %rs7, [%rd57+12];
mov.b32 %r102, {%rs6, %rs7};
shr.s32 %r103, %r96, %r14;
add.s64 %rd61, %rd54, %rd60;
ld.global.nc.u32 %r18, [%rd61];
ld.global.nc.u32 %r19, [%rd59+20];
ld.global.nc.u16 %rs1, [%rd51];

{ cvt.f32.f16 %f40, %rs1;}


and.b32 %r104, %r98, 252645135;
shl.b32 %r105, %r100, 4;
and.b32 %r106, %r105, 16;
or.b32 %r107, %r106, %r104;
shl.b32 %r108, %r100, 11;
and.b32 %r109, %r108, 4096;
or.b32 %r110, %r107, %r109;
shl.b32 %r111, %r100, 18;
and.b32 %r112, %r111, 1048576;
or.b32 %r113, %r110, %r112;
shl.b32 %r114, %r100, 25;
and.b32 %r115, %r114, 268435456;
or.b32 %r79, %r113, %r115;
mov.u32 %r81, 0;

dp4a.s32.s32 %r78, %r79, %r16, %r81;

shr.u32 %r116, %r98, 4;
and.b32 %r117, %r116, 252645135;
shr.u32 %r118, %r100, 12;
and.b32 %r119, %r118, 16;
or.b32 %r120, %r119, %r117;
shr.u32 %r121, %r100, 5;
and.b32 %r122, %r121, 4096;
or.b32 %r123, %r120, %r122;
shl.b32 %r124, %r100, 2;
and.b32 %r125, %r124, 1048576;
or.b32 %r126, %r123, %r125;
shl.b32 %r127, %r100, 9;
and.b32 %r128, %r127, 268435456;
or.b32 %r83, %r126, %r128;

dp4a.s32.s32 %r82, %r83, %r17, %r78;

and.b32 %r129, %r102, 252645135;
shl.b32 %r130, %r103, 4;
and.b32 %r131, %r130, 16;
or.b32 %r132, %r131, %r129;
shl.b32 %r133, %r103, 11;
and.b32 %r134, %r133, 4096;
or.b32 %r135, %r132, %r134;
shl.b32 %r136, %r103, 18;
and.b32 %r137, %r136, 1048576;
or.b32 %r138, %r135, %r137;
shl.b32 %r139, %r103, 25;
and.b32 %r140, %r139, 268435456;
or.b32 %r87, %r138, %r140;

dp4a.s32.s32 %r86, %r87, %r18, %r82;

shr.u32 %r141, %r102, 4;
and.b32 %r142, %r141, 252645135;
shr.u32 %r143, %r103, 12;
and.b32 %r144, %r143, 16;
or.b32 %r145, %r144, %r142;
shr.u32 %r146, %r103, 5;
and.b32 %r147, %r146, 4096;
or.b32 %r148, %r145, %r147;
shl.b32 %r149, %r103, 2;
and.b32 %r150, %r149, 1048576;
or.b32 %r151, %r148, %r150;
shl.b32 %r152, %r103, 9;
and.b32 %r153, %r152, 268435456;
or.b32 %r91, %r151, %r153;

dp4a.s32.s32 %r90, %r91, %r19, %r86;

ld.global.nc.u32 %r20, [%rd53];

{.reg .f16 low,high;
mov.b32 {low,high},%r20;
cvt.f32.f16 %f41, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r20;
cvt.f32.f16 %f42, high;}


cvt.rn.f32.s32 %f43, %r90;
mul.ftz.f32 %f44, %f42, 0fC1000000;
fma.rn.ftz.f32 %f45, %f41, %f43, %f44;
fma.rn.ftz.f32 %f119, %f40, %f45, %f119;
setp.eq.s64 %p10, %rd14, 0;
@%p10 bra $L__BB60_12;

add.s64 %rd63, %rd43, %rd68;
ld.global.u16 %rs9, [%rd63+2];
ld.global.u16 %rs10, [%rd63+4];
mov.b32 %r172, {%rs9, %rs10};
add.s64 %rd64, %rd6, %rd68;
ld.global.u16 %rs11, [%rd64];
ld.global.u16 %rs12, [%rd64+2];
mov.b32 %r173, {%rs11, %rs12};
shr.s32 %r175, %r172, %r99;
ld.global.u16 %rs13, [%rd64+4];
ld.global.u16 %rs14, [%rd64+6];
mov.b32 %r176, {%rs13, %rs14};
shr.s32 %r177, %r172, %r14;
ld.global.u16 %rs8, [%rd63];

{ cvt.f32.f16 %f46, %rs8;}


and.b32 %r178, %r173, 252645135;
shl.b32 %r179, %r175, 4;
and.b32 %r180, %r179, 16;
or.b32 %r181, %r180, %r178;
shl.b32 %r182, %r175, 11;
and.b32 %r183, %r182, 4096;
or.b32 %r184, %r181, %r183;
shl.b32 %r185, %r175, 18;
and.b32 %r186, %r185, 1048576;
or.b32 %r187, %r184, %r186;
shl.b32 %r188, %r175, 25;
and.b32 %r189, %r188, 268435456;
or.b32 %r155, %r187, %r189;

dp4a.s32.s32 %r154, %r155, %r16, %r81;

shr.u32 %r190, %r173, 4;
and.b32 %r191, %r190, 252645135;
shr.u32 %r192, %r175, 12;
and.b32 %r193, %r192, 16;
or.b32 %r194, %r193, %r191;
shr.u32 %r195, %r175, 5;
and.b32 %r196, %r195, 4096;
or.b32 %r197, %r194, %r196;
shl.b32 %r198, %r175, 2;
and.b32 %r199, %r198, 1048576;
or.b32 %r200, %r197, %r199;
shl.b32 %r201, %r175, 9;
and.b32 %r202, %r201, 268435456;
or.b32 %r159, %r200, %r202;

dp4a.s32.s32 %r158, %r159, %r17, %r154;

and.b32 %r203, %r176, 252645135;
shl.b32 %r204, %r177, 4;
and.b32 %r205, %r204, 16;
or.b32 %r206, %r205, %r203;
shl.b32 %r207, %r177, 11;
and.b32 %r208, %r207, 4096;
or.b32 %r209, %r206, %r208;
shl.b32 %r210, %r177, 18;
and.b32 %r211, %r210, 1048576;
or.b32 %r212, %r209, %r211;
shl.b32 %r213, %r177, 25;
and.b32 %r214, %r213, 268435456;
or.b32 %r163, %r212, %r214;

dp4a.s32.s32 %r162, %r163, %r18, %r158;

shr.u32 %r215, %r176, 4;
and.b32 %r216, %r215, 252645135;
shr.u32 %r217, %r177, 12;
and.b32 %r218, %r217, 16;
or.b32 %r219, %r218, %r216;
shr.u32 %r220, %r177, 5;
and.b32 %r221, %r220, 4096;
or.b32 %r222, %r219, %r221;
shl.b32 %r223, %r177, 2;
and.b32 %r224, %r223, 1048576;
or.b32 %r225, %r222, %r224;
shl.b32 %r226, %r177, 9;
and.b32 %r227, %r226, 268435456;
or.b32 %r167, %r225, %r227;

dp4a.s32.s32 %r166, %r167, %r19, %r162;


{.reg .f16 low,high;
mov.b32 {low,high},%r20;
cvt.f32.f16 %f47, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r20;
cvt.f32.f16 %f48, high;}


cvt.rn.f32.s32 %f49, %r166;
mul.ftz.f32 %f50, %f48, 0fC1000000;
fma.rn.ftz.f32 %f51, %f47, %f49, %f50;
fma.rn.ftz.f32 %f117, %f46, %f51, %f117;

$L__BB60_12:
add.s64 %rd68, %rd68, 1408;
add.s32 %r293, %r293, 64;
setp.lt.s32 %p11, %r293, %r59;
@%p11 bra $L__BB60_10;

$L__BB60_13:
setp.eq.s32 %p12, %r1, 0;
@%p12 bra $L__BB60_16;

setp.eq.s64 %p13, %rd14, 0;
add.s32 %r22, %r1, -1;
shl.b32 %r232, %r1, 7;
mov.u32 %r233, _ZZ13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r234, %r233, %r232;
shl.b32 %r235, %r2, 2;
add.s32 %r236, %r234, %r235;
st.shared.f32 [%r236+-128], %f119;
@%p13 bra $L__BB60_16;

shl.b32 %r238, %r22, 7;
mov.u32 %r239, _ZZ13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r240, %r239, %r238;
add.s32 %r242, %r240, %r235;
st.shared.f32 [%r242], %f117;

$L__BB60_16:
setp.ne.s32 %p14, %r1, 0;
bar.sync 0;
@%p14 bra $L__BB60_35;

setp.ne.s64 %p15, %rd14, 0;
shl.b32 %r245, %r2, 2;
mov.u32 %r246, _ZZ13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r23, %r246, %r245;
ld.shared.f32 %f52, [%r23];
add.ftz.f32 %f13, %f119, %f52;
@%p15 bra $L__BB60_19;
bra.uni $L__BB60_18;

$L__BB60_19:
mov.u32 %r249, _ZZ13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r250, %r249, %r245;
ld.shared.f32 %f56, [%r250];
add.ftz.f32 %f57, %f117, %f56;
ld.shared.f32 %f58, [%r23+128];
add.ftz.f32 %f59, %f13, %f58;
ld.shared.f32 %f60, [%r250+128];
add.ftz.f32 %f61, %f57, %f60;
ld.shared.f32 %f62, [%r23+256];
add.ftz.f32 %f121, %f59, %f62;
ld.shared.f32 %f63, [%r250+256];
add.ftz.f32 %f117, %f61, %f63;
bra.uni $L__BB60_20;

$L__BB60_18:
ld.shared.f32 %f53, [%r23+128];
add.ftz.f32 %f54, %f13, %f53;
ld.shared.f32 %f55, [%r23+256];
add.ftz.f32 %f121, %f54, %f55;

$L__BB60_20:
mov.b32 %r251, %f121;
mov.u32 %r252, 31;
mov.u32 %r253, 16;
mov.u32 %r254, -1;
shfl.sync.bfly.b32 %r255|%p16, %r251, %r253, %r252, %r254;
mov.b32 %f64, %r255;
add.ftz.f32 %f65, %f121, %f64;
mov.b32 %r256, %f65;
mov.u32 %r257, 8;
shfl.sync.bfly.b32 %r258|%p17, %r256, %r257, %r252, %r254;
mov.b32 %f66, %r258;
add.ftz.f32 %f67, %f65, %f66;
mov.b32 %r259, %f67;
mov.u32 %r260, 4;
shfl.sync.bfly.b32 %r261|%p18, %r259, %r260, %r252, %r254;
mov.b32 %f68, %r261;
add.ftz.f32 %f69, %f67, %f68;
mov.b32 %r262, %f69;
mov.u32 %r263, 2;
shfl.sync.bfly.b32 %r264|%p19, %r262, %r263, %r252, %r254;
mov.b32 %f70, %r264;
add.ftz.f32 %f19, %f69, %f70;
mov.b32 %r265, %f19;
mov.u32 %r266, 1;
shfl.sync.bfly.b32 %r24|%p1, %r265, %r266, %r252, %r254;
setp.eq.s64 %p20, %rd14, 0;
@%p20 bra $L__BB60_22;

mov.b32 %r267, %f117;
shfl.sync.bfly.b32 %r271|%p21, %r267, %r253, %r252, %r254;
mov.b32 %f71, %r271;
add.ftz.f32 %f72, %f117, %f71;
mov.b32 %r272, %f72;
shfl.sync.bfly.b32 %r274|%p22, %r272, %r257, %r252, %r254;
mov.b32 %f73, %r274;
add.ftz.f32 %f74, %f72, %f73;
mov.b32 %r275, %f74;
shfl.sync.bfly.b32 %r277|%p23, %r275, %r260, %r252, %r254;
mov.b32 %f75, %r277;
add.ftz.f32 %f76, %f74, %f75;
mov.b32 %r278, %f76;
shfl.sync.bfly.b32 %r280|%p24, %r278, %r263, %r252, %r254;
mov.b32 %f77, %r280;
add.ftz.f32 %f78, %f76, %f77;
mov.b32 %r281, %f78;
shfl.sync.bfly.b32 %r283|%p25, %r281, %r266, %r252, %r254;
mov.b32 %f79, %r283;
add.ftz.f32 %f117, %f78, %f79;

$L__BB60_22:
setp.ne.s32 %p26, %r2, 0;
mov.b32 %f80, %r24;
add.ftz.f32 %f22, %f19, %f80;
@%p26 bra $L__BB60_35;

add.ftz.f32 %f123, %f114, %f22;
@%p20 bra $L__BB60_34;

add.ftz.f32 %f24, %f113, %f117;
setp.gt.s32 %p28, %r25, 2;
@%p28 bra $L__BB60_28;

setp.eq.s32 %p31, %r25, 1;
@%p31 bra $L__BB60_32;

setp.eq.s32 %p32, %r25, 2;
@%p32 bra $L__BB60_27;
bra.uni $L__BB60_33;

$L__BB60_27:
mul.ftz.f32 %f100, %f24, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f101, %f100;
add.ftz.f32 %f102, %f101, 0f3F800000;
div.approx.ftz.f32 %f103, %f24, %f102;
mul.ftz.f32 %f123, %f123, %f103;
bra.uni $L__BB60_34;

$L__BB60_28:
setp.eq.s32 %p29, %r25, 3;
@%p29 bra $L__BB60_31;

setp.ne.s32 %p30, %r25, 6;
@%p30 bra $L__BB60_33;

min.ftz.f32 %f81, %f24, %f31;
neg.ftz.f32 %f82, %f31;
min.ftz.f32 %f83, %f123, %f31;
max.ftz.f32 %f84, %f83, %f82;
mul.ftz.f32 %f85, %f81, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f86, %f85;
add.ftz.f32 %f87, %f86, 0f3F800000;
div.approx.ftz.f32 %f88, %f81, %f87;
mul.ftz.f32 %f123, %f84, %f88;
bra.uni $L__BB60_34;

$L__BB60_32:
mul.ftz.f32 %f104, %f24, 0f3F000000;
mul.ftz.f32 %f105, %f24, 0f3D372713;
fma.rn.ftz.f32 %f106, %f24, %f105, 0f3F800000;
mul.ftz.f32 %f107, %f24, 0f3F4C422A;
mul.ftz.f32 %f108, %f107, %f106;
tanh.approx.f32 %f109, %f108;
add.ftz.f32 %f110, %f109, 0f3F800000;
mul.ftz.f32 %f111, %f104, %f110;
mul.ftz.f32 %f123, %f123, %f111;
bra.uni $L__BB60_34;

$L__BB60_31:
mov.f32 %f89, 0f40E00000;
min.ftz.f32 %f90, %f24, %f89;
min.ftz.f32 %f91, %f123, %f89;
mov.f32 %f92, 0fC0E00000;
max.ftz.f32 %f93, %f91, %f92;
mul.ftz.f32 %f94, %f90, 0f3FD9DB23;
mul.ftz.f32 %f95, %f94, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f96, %f95;
add.ftz.f32 %f97, %f96, 0f3F800000;
div.approx.ftz.f32 %f98, %f90, %f97;
add.ftz.f32 %f99, %f93, 0f3F800000;
mul.ftz.f32 %f123, %f99, %f98;
bra.uni $L__BB60_34;

$L__BB60_33:
mul.ftz.f32 %f123, %f123, %f24;

$L__BB60_34:
cvta.to.global.u64 %rd65, %rd18;
mov.u32 %r286, %ctaid.z;
mov.u32 %r287, %ctaid.y;
mad.lo.s32 %r289, %r287, %r36, %r4;
mad.lo.s32 %r290, %r286, %r42, %r289;
mul.wide.u32 %rd66, %r290, 4;
add.s64 %rd67, %rd65, %rd66;
st.global.f32 [%rd67], %f123;

$L__BB60_35:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<12>;
.reg .b16 %rs<8>;
.reg .f32 %f<30>;
.reg .b32 %r<165>;
.reg .b64 %rd<35>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[384];

ld.param.u64 %rd9, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd10, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd11, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd12, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r18, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r19, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
mov.u32 %r1, %tid.y;
shl.b32 %r35, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r35, %r2;
mov.u32 %r4, %ctaid.y;
setp.eq.s64 %p1, %rd11, 0;
@%p1 bra $L__BB61_2;

cvta.to.global.u64 %rd13, %rd11;
mul.wide.u32 %rd14, %r4, 4;
add.s64 %rd15, %rd13, %rd14;
ld.global.nc.u32 %r162, [%rd15];
mul.hi.u32 %r36, %r4, %r19;
add.s32 %r37, %r36, %r4;
shr.u32 %r38, %r37, %r20;
mul.lo.s32 %r39, %r38, %r21;
sub.s32 %r163, %r4, %r39;
bra.uni $L__BB61_3;

$L__BB61_2:
mul.hi.u32 %r40, %r4, %r23;
add.s32 %r41, %r40, %r4;
shr.u32 %r162, %r41, %r24;
mov.u32 %r163, %r4;

$L__BB61_3:
mov.u32 %r10, %ctaid.z;
shr.u32 %r42, %r3, 31;
add.s32 %r43, %r3, %r42;
shr.s32 %r164, %r43, 1;
shr.u32 %r44, %r18, 5;
setp.ge.s32 %p2, %r164, %r44;
mov.f32 %f29, 0f00000000;
@%p2 bra $L__BB61_6;

and.b32 %r47, %r43, -2;
sub.s32 %r48, %r3, %r47;
shl.b32 %r49, %r48, 1;
shl.b32 %r50, %r48, 2;
or.b32 %r51, %r50, 1;
shl.b32 %r12, %r48, 3;
cvt.s64.s32 %rd1, %r49;
sub.s32 %r52, %r51, %r49;
shl.b32 %r13, %r52, 2;
cvt.s64.s32 %rd2, %r52;
mul.hi.u32 %r53, %r10, %r29;
add.s32 %r54, %r53, %r10;
shr.u32 %r55, %r54, %r30;
mad.lo.s32 %r56, %r55, %r32, %r164;
mad.lo.s32 %r57, %r162, %r26, %r56;
mov.u32 %r58, %ctaid.x;
mad.lo.s32 %r59, %r58, %r22, %r57;
mul.wide.s32 %rd16, %r59, 22;
cvta.to.global.u64 %rd17, %rd9;
add.s64 %rd34, %rd17, %rd16;
mul.wide.s32 %rd18, %r50, 2;
add.s64 %rd4, %rd18, 8;
cvta.to.global.u64 %rd5, %rd10;
mul.lo.s32 %r60, %r163, %r27;
cvt.u64.u32 %rd19, %r60;
mul.lo.s32 %r61, %r10, %r33;
cvt.u64.u32 %rd20, %r61;
add.s64 %rd6, %rd19, %rd20;
mov.f32 %f29, 0f00000000;
shl.b64 %rd27, %rd1, 2;
shl.b64 %rd29, %rd2, 2;

$L__BB61_5:
cvt.u64.u32 %rd21, %r164;
add.s64 %rd22, %rd6, %rd21;
ld.global.nc.u16 %rs2, [%rd34+2];
ld.global.nc.u16 %rs3, [%rd34+4];
mov.b32 %r80, {%rs2, %rs3};
mul.lo.s64 %rd23, %rd22, 36;
add.s64 %rd24, %rd5, %rd23;
add.s64 %rd25, %rd24, 4;
add.s64 %rd26, %rd34, %rd4;
ld.global.nc.u16 %rs4, [%rd26+-2];
ld.global.nc.u16 %rs5, [%rd26];
mov.b32 %r81, {%rs4, %rs5};
shr.s32 %r82, %r80, %r12;
add.s64 %rd28, %rd25, %rd27;
ld.global.nc.u32 %r64, [%rd28];
ld.global.nc.u32 %r68, [%rd28+16];
ld.global.nc.u16 %rs6, [%rd26+2];
ld.global.nc.u16 %rs7, [%rd26+4];
mov.b32 %r83, {%rs6, %rs7};
shr.s32 %r84, %r80, %r13;
add.s64 %rd30, %rd25, %rd29;
ld.global.nc.u32 %r72, [%rd30];
ld.global.nc.u32 %r76, [%rd28+20];
ld.global.nc.u16 %rs1, [%rd34];

{ cvt.f32.f16 %f7, %rs1;}


and.b32 %r85, %r81, 252645135;
shl.b32 %r86, %r82, 4;
and.b32 %r87, %r86, 16;
or.b32 %r88, %r87, %r85;
shl.b32 %r89, %r82, 11;
and.b32 %r90, %r89, 4096;
or.b32 %r91, %r88, %r90;
shl.b32 %r92, %r82, 18;
and.b32 %r93, %r92, 1048576;
or.b32 %r94, %r91, %r93;
shl.b32 %r95, %r82, 25;
and.b32 %r96, %r95, 268435456;
or.b32 %r63, %r94, %r96;
mov.u32 %r65, 0;

dp4a.s32.s32 %r62, %r63, %r64, %r65;

shr.u32 %r97, %r81, 4;
and.b32 %r98, %r97, 252645135;
shr.u32 %r99, %r82, 12;
and.b32 %r100, %r99, 16;
or.b32 %r101, %r100, %r98;
shr.u32 %r102, %r82, 5;
and.b32 %r103, %r102, 4096;
or.b32 %r104, %r101, %r103;
shl.b32 %r105, %r82, 2;
and.b32 %r106, %r105, 1048576;
or.b32 %r107, %r104, %r106;
shl.b32 %r108, %r82, 9;
and.b32 %r109, %r108, 268435456;
or.b32 %r67, %r107, %r109;

dp4a.s32.s32 %r66, %r67, %r68, %r62;

and.b32 %r110, %r83, 252645135;
shl.b32 %r111, %r84, 4;
and.b32 %r112, %r111, 16;
or.b32 %r113, %r112, %r110;
shl.b32 %r114, %r84, 11;
and.b32 %r115, %r114, 4096;
or.b32 %r116, %r113, %r115;
shl.b32 %r117, %r84, 18;
and.b32 %r118, %r117, 1048576;
or.b32 %r119, %r116, %r118;
shl.b32 %r120, %r84, 25;
and.b32 %r121, %r120, 268435456;
or.b32 %r71, %r119, %r121;

dp4a.s32.s32 %r70, %r71, %r72, %r66;

shr.u32 %r122, %r83, 4;
and.b32 %r123, %r122, 252645135;
shr.u32 %r124, %r84, 12;
and.b32 %r125, %r124, 16;
or.b32 %r126, %r125, %r123;
shr.u32 %r127, %r84, 5;
and.b32 %r128, %r127, 4096;
or.b32 %r129, %r126, %r128;
shl.b32 %r130, %r84, 2;
and.b32 %r131, %r130, 1048576;
or.b32 %r132, %r129, %r131;
shl.b32 %r133, %r84, 9;
and.b32 %r134, %r133, 268435456;
or.b32 %r75, %r132, %r134;

dp4a.s32.s32 %r74, %r75, %r76, %r70;

ld.global.nc.u32 %r78, [%rd24];

{.reg .f16 low,high;
mov.b32 {low,high},%r78;
cvt.f32.f16 %f8, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r78;
cvt.f32.f16 %f9, high;}


cvt.rn.f32.s32 %f10, %r74;
mul.ftz.f32 %f11, %f9, 0fC1000000;
fma.rn.ftz.f32 %f12, %f8, %f10, %f11;
fma.rn.ftz.f32 %f29, %f7, %f12, %f29;
add.s64 %rd34, %rd34, 1408;
add.s32 %r164, %r164, 64;
setp.lt.s32 %p3, %r164, %r44;
@%p3 bra $L__BB61_5;

$L__BB61_6:
setp.eq.s32 %p4, %r1, 0;
@%p4 bra $L__BB61_8;

shl.b32 %r135, %r1, 7;
mov.u32 %r136, _ZZ13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r137, %r136, %r135;
shl.b32 %r138, %r2, 2;
add.s32 %r139, %r137, %r138;
st.shared.f32 [%r139+-128], %f29;

$L__BB61_8:
setp.ne.s32 %p5, %r1, 0;
bar.sync 0;
@%p5 bra $L__BB61_11;

mov.u32 %r140, %ctaid.x;
mad.lo.s32 %r141, %r4, %r28, %r140;
mad.lo.s32 %r17, %r10, %r34, %r141;
shl.b32 %r142, %r2, 2;
mov.u32 %r143, _ZZ13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r144, %r143, %r142;
mov.u32 %r145, 2;
ld.shared.f32 %f13, [%r144];
add.ftz.f32 %f14, %f29, %f13;
ld.shared.f32 %f15, [%r144+128];
add.ftz.f32 %f16, %f14, %f15;
ld.shared.f32 %f17, [%r144+256];
add.ftz.f32 %f18, %f16, %f17;
mov.b32 %r146, %f18;
mov.u32 %r147, 31;
mov.u32 %r148, 16;
mov.u32 %r149, -1;
shfl.sync.bfly.b32 %r150|%p6, %r146, %r148, %r147, %r149;
mov.b32 %f19, %r150;
add.ftz.f32 %f20, %f18, %f19;
mov.b32 %r151, %f20;
mov.u32 %r152, 8;
shfl.sync.bfly.b32 %r153|%p7, %r151, %r152, %r147, %r149;
mov.b32 %f21, %r153;
add.ftz.f32 %f22, %f20, %f21;
mov.b32 %r154, %f22;
mov.u32 %r155, 4;
shfl.sync.bfly.b32 %r156|%p8, %r154, %r155, %r147, %r149;
mov.b32 %f23, %r156;
add.ftz.f32 %f24, %f22, %f23;
mov.b32 %r157, %f24;
shfl.sync.bfly.b32 %r158|%p9, %r157, %r145, %r147, %r149;
mov.b32 %f25, %r158;
add.ftz.f32 %f26, %f24, %f25;
mov.b32 %r159, %f26;
mov.u32 %r160, 1;
shfl.sync.bfly.b32 %r161|%p10, %r159, %r160, %r147, %r149;
mov.b32 %f27, %r161;
add.ftz.f32 %f4, %f26, %f27;
setp.ne.s32 %p11, %r2, 0;
@%p11 bra $L__BB61_11;

cvta.to.global.u64 %rd31, %rd12;
mul.wide.u32 %rd32, %r17, 4;
add.s64 %rd33, %rd31, %rd32;
st.global.f32 [%rd33], %f4;

$L__BB61_11:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<33>;
.reg .b16 %rs<15>;
.reg .f32 %f<124>;
.reg .b32 %r<294>;
.reg .b64 %rd<69>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[384];

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate[384];

ld.param.u64 %rd10, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd11, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd12, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
ld.param.f32 %f31, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+16];
ld.param.u64 %rd14, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+8];
ld.param.u64 %rd13, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3];
mov.u32 %r1, %tid.y;
shl.b32 %r43, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r43, %r2;
mov.u32 %r4, %ctaid.x;
mov.u32 %r5, %ctaid.y;
setp.eq.s64 %p2, %rd12, 0;
@%p2 bra $L__BB62_2;

cvta.to.global.u64 %rd19, %rd12;
mul.wide.u32 %rd20, %r5, 4;
add.s64 %rd21, %rd19, %rd20;
ld.global.nc.u32 %r291, [%rd21];
mul.hi.u32 %r44, %r5, %r27;
add.s32 %r45, %r44, %r5;
shr.u32 %r46, %r45, %r28;
mul.lo.s32 %r47, %r46, %r29;
sub.s32 %r292, %r5, %r47;
bra.uni $L__BB62_3;

$L__BB62_2:
mul.hi.u32 %r48, %r5, %r31;
add.s32 %r49, %r48, %r5;
shr.u32 %r291, %r49, %r32;
mov.u32 %r292, %r5;

$L__BB62_3:
selp.b32 %r11, %r5, %r291, %p2;
or.b32 %r50, %r2, %r1;
setp.ne.s32 %p4, %r50, 0;
mov.f32 %f117, 0f00000000;
mov.f32 %f113, %f117;
mov.f32 %f114, %f117;
@%p4 bra $L__BB62_8;

setp.eq.s64 %p5, %rd13, 0;
mov.f32 %f113, 0f00000000;
mov.f32 %f114, %f113;
@%p5 bra $L__BB62_6;

cvta.to.global.u64 %rd22, %rd13;
mov.u32 %r51, %ctaid.z;
mul.lo.s32 %r52, %r51, %r42;
cvt.u64.u32 %rd23, %r52;
mul.lo.s32 %r53, %r11, %r36;
cvt.u64.u32 %rd24, %r53;
cvt.u64.u32 %rd25, %r2;
cvt.s64.s32 %rd26, %r4;
add.s64 %rd27, %rd25, %rd26;
add.s64 %rd28, %rd27, %rd23;
add.s64 %rd29, %rd28, %rd24;
shl.b64 %rd30, %rd29, 2;
add.s64 %rd31, %rd22, %rd30;
ld.global.f32 %f114, [%rd31];

$L__BB62_6:
setp.eq.s64 %p6, %rd15, 0;
setp.eq.s64 %p7, %rd14, 0;
or.pred %p8, %p6, %p7;
@%p8 bra $L__BB62_8;

cvta.to.global.u64 %rd32, %rd15;
mov.u32 %r55, %ctaid.z;
mul.lo.s32 %r56, %r55, %r42;
cvt.u64.u32 %rd33, %r56;
mul.lo.s32 %r57, %r11, %r36;
cvt.u64.u32 %rd34, %r57;
cvt.u64.u32 %rd35, %r2;
cvt.s64.s32 %rd36, %r4;
add.s64 %rd37, %rd35, %rd36;
add.s64 %rd38, %rd37, %rd33;
add.s64 %rd39, %rd38, %rd34;
shl.b64 %rd40, %rd39, 2;
add.s64 %rd41, %rd32, %rd40;
ld.global.f32 %f113, [%rd41];

$L__BB62_8:
shr.u32 %r59, %r26, 5;
shr.u32 %r60, %r3, 31;
add.s32 %r61, %r3, %r60;
shr.s32 %r293, %r61, 1;
setp.ge.s32 %p9, %r293, %r59;
mov.f32 %f119, %f117;
@%p9 bra $L__BB62_13;

mov.u32 %r62, %ctaid.z;
and.b32 %r65, %r61, -2;
sub.s32 %r13, %r3, %r65;
shl.b32 %r66, %r13, 1;
shl.b32 %r67, %r13, 2;
or.b32 %r68, %r67, 1;
sub.s32 %r69, %r68, %r66;
shl.b32 %r14, %r69, 2;
cvt.s64.s32 %rd4, %r69;
mul.hi.u32 %r70, %r62, %r37;
add.s32 %r71, %r70, %r62;
shr.u32 %r72, %r71, %r38;
mad.lo.s32 %r73, %r72, %r40, %r293;
mad.lo.s32 %r74, %r291, %r34, %r73;
mad.lo.s32 %r75, %r4, %r30, %r74;
mul.wide.s32 %rd68, %r75, 22;
mul.wide.s32 %rd42, %r67, 2;
cvta.to.global.u64 %rd43, %rd14;
add.s64 %rd44, %rd43, %rd42;
add.s64 %rd6, %rd44, 6;
mul.lo.s32 %r76, %r292, %r35;
cvt.u64.u32 %rd45, %r76;
mul.lo.s32 %r77, %r62, %r41;
cvt.u64.u32 %rd46, %r77;
add.s64 %rd7, %rd45, %rd46;
mov.f32 %f119, 0f00000000;
cvta.to.global.u64 %rd47, %rd11;
cvta.to.global.u64 %rd50, %rd10;
shl.b64 %rd60, %rd4, 2;
mov.f32 %f117, %f119;

$L__BB62_10:
cvt.u64.u32 %rd48, %r293;
add.s64 %rd49, %rd7, %rd48;
add.s64 %rd51, %rd50, %rd68;
ld.global.nc.u16 %rs2, [%rd51+2];
ld.global.nc.u16 %rs3, [%rd51+4];
mov.b32 %r96, {%rs2, %rs3};
mul.lo.s64 %rd52, %rd49, 36;
add.s64 %rd53, %rd47, %rd52;
add.s64 %rd54, %rd53, 4;
add.s64 %rd56, %rd50, %rd42;
add.s64 %rd57, %rd56, %rd68;
ld.global.nc.u16 %rs4, [%rd57+6];
ld.global.nc.u16 %rs5, [%rd57+8];
mov.b32 %r98, {%rs4, %rs5};
shl.b32 %r99, %r13, 3;
shr.s32 %r100, %r96, %r99;
mul.wide.s32 %rd58, %r66, 4;
add.s64 %rd59, %rd54, %rd58;
ld.global.nc.u32 %r16, [%rd59];
ld.global.nc.u32 %r17, [%rd59+16];
ld.global.nc.u16 %rs6, [%rd57+10];
ld.global.nc.u16 %rs7, [%rd57+12];
mov.b32 %r102, {%rs6, %rs7};
shr.s32 %r103, %r96, %r14;
add.s64 %rd61, %rd54, %rd60;
ld.global.nc.u32 %r18, [%rd61];
ld.global.nc.u32 %r19, [%rd59+20];
ld.global.nc.u16 %rs1, [%rd51];

{ cvt.f32.f16 %f40, %rs1;}


and.b32 %r104, %r98, 252645135;
shl.b32 %r105, %r100, 4;
and.b32 %r106, %r105, 16;
or.b32 %r107, %r106, %r104;
shl.b32 %r108, %r100, 11;
and.b32 %r109, %r108, 4096;
or.b32 %r110, %r107, %r109;
shl.b32 %r111, %r100, 18;
and.b32 %r112, %r111, 1048576;
or.b32 %r113, %r110, %r112;
shl.b32 %r114, %r100, 25;
and.b32 %r115, %r114, 268435456;
or.b32 %r79, %r113, %r115;
mov.u32 %r81, 0;

dp4a.s32.s32 %r78, %r79, %r16, %r81;

shr.u32 %r116, %r98, 4;
and.b32 %r117, %r116, 252645135;
shr.u32 %r118, %r100, 12;
and.b32 %r119, %r118, 16;
or.b32 %r120, %r119, %r117;
shr.u32 %r121, %r100, 5;
and.b32 %r122, %r121, 4096;
or.b32 %r123, %r120, %r122;
shl.b32 %r124, %r100, 2;
and.b32 %r125, %r124, 1048576;
or.b32 %r126, %r123, %r125;
shl.b32 %r127, %r100, 9;
and.b32 %r128, %r127, 268435456;
or.b32 %r83, %r126, %r128;

dp4a.s32.s32 %r82, %r83, %r17, %r78;

and.b32 %r129, %r102, 252645135;
shl.b32 %r130, %r103, 4;
and.b32 %r131, %r130, 16;
or.b32 %r132, %r131, %r129;
shl.b32 %r133, %r103, 11;
and.b32 %r134, %r133, 4096;
or.b32 %r135, %r132, %r134;
shl.b32 %r136, %r103, 18;
and.b32 %r137, %r136, 1048576;
or.b32 %r138, %r135, %r137;
shl.b32 %r139, %r103, 25;
and.b32 %r140, %r139, 268435456;
or.b32 %r87, %r138, %r140;

dp4a.s32.s32 %r86, %r87, %r18, %r82;

shr.u32 %r141, %r102, 4;
and.b32 %r142, %r141, 252645135;
shr.u32 %r143, %r103, 12;
and.b32 %r144, %r143, 16;
or.b32 %r145, %r144, %r142;
shr.u32 %r146, %r103, 5;
and.b32 %r147, %r146, 4096;
or.b32 %r148, %r145, %r147;
shl.b32 %r149, %r103, 2;
and.b32 %r150, %r149, 1048576;
or.b32 %r151, %r148, %r150;
shl.b32 %r152, %r103, 9;
and.b32 %r153, %r152, 268435456;
or.b32 %r91, %r151, %r153;

dp4a.s32.s32 %r90, %r91, %r19, %r86;

ld.global.nc.u32 %r20, [%rd53];

{.reg .f16 low,high;
mov.b32 {low,high},%r20;
cvt.f32.f16 %f41, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r20;
cvt.f32.f16 %f42, high;}


cvt.rn.f32.s32 %f43, %r90;
mul.ftz.f32 %f44, %f42, 0fC1000000;
fma.rn.ftz.f32 %f45, %f41, %f43, %f44;
fma.rn.ftz.f32 %f119, %f40, %f45, %f119;
setp.eq.s64 %p10, %rd14, 0;
@%p10 bra $L__BB62_12;

add.s64 %rd63, %rd43, %rd68;
ld.global.u16 %rs9, [%rd63+2];
ld.global.u16 %rs10, [%rd63+4];
mov.b32 %r172, {%rs9, %rs10};
add.s64 %rd64, %rd6, %rd68;
ld.global.u16 %rs11, [%rd64];
ld.global.u16 %rs12, [%rd64+2];
mov.b32 %r173, {%rs11, %rs12};
shr.s32 %r175, %r172, %r99;
ld.global.u16 %rs13, [%rd64+4];
ld.global.u16 %rs14, [%rd64+6];
mov.b32 %r176, {%rs13, %rs14};
shr.s32 %r177, %r172, %r14;
ld.global.u16 %rs8, [%rd63];

{ cvt.f32.f16 %f46, %rs8;}


and.b32 %r178, %r173, 252645135;
shl.b32 %r179, %r175, 4;
and.b32 %r180, %r179, 16;
or.b32 %r181, %r180, %r178;
shl.b32 %r182, %r175, 11;
and.b32 %r183, %r182, 4096;
or.b32 %r184, %r181, %r183;
shl.b32 %r185, %r175, 18;
and.b32 %r186, %r185, 1048576;
or.b32 %r187, %r184, %r186;
shl.b32 %r188, %r175, 25;
and.b32 %r189, %r188, 268435456;
or.b32 %r155, %r187, %r189;

dp4a.s32.s32 %r154, %r155, %r16, %r81;

shr.u32 %r190, %r173, 4;
and.b32 %r191, %r190, 252645135;
shr.u32 %r192, %r175, 12;
and.b32 %r193, %r192, 16;
or.b32 %r194, %r193, %r191;
shr.u32 %r195, %r175, 5;
and.b32 %r196, %r195, 4096;
or.b32 %r197, %r194, %r196;
shl.b32 %r198, %r175, 2;
and.b32 %r199, %r198, 1048576;
or.b32 %r200, %r197, %r199;
shl.b32 %r201, %r175, 9;
and.b32 %r202, %r201, 268435456;
or.b32 %r159, %r200, %r202;

dp4a.s32.s32 %r158, %r159, %r17, %r154;

and.b32 %r203, %r176, 252645135;
shl.b32 %r204, %r177, 4;
and.b32 %r205, %r204, 16;
or.b32 %r206, %r205, %r203;
shl.b32 %r207, %r177, 11;
and.b32 %r208, %r207, 4096;
or.b32 %r209, %r206, %r208;
shl.b32 %r210, %r177, 18;
and.b32 %r211, %r210, 1048576;
or.b32 %r212, %r209, %r211;
shl.b32 %r213, %r177, 25;
and.b32 %r214, %r213, 268435456;
or.b32 %r163, %r212, %r214;

dp4a.s32.s32 %r162, %r163, %r18, %r158;

shr.u32 %r215, %r176, 4;
and.b32 %r216, %r215, 252645135;
shr.u32 %r217, %r177, 12;
and.b32 %r218, %r217, 16;
or.b32 %r219, %r218, %r216;
shr.u32 %r220, %r177, 5;
and.b32 %r221, %r220, 4096;
or.b32 %r222, %r219, %r221;
shl.b32 %r223, %r177, 2;
and.b32 %r224, %r223, 1048576;
or.b32 %r225, %r222, %r224;
shl.b32 %r226, %r177, 9;
and.b32 %r227, %r226, 268435456;
or.b32 %r167, %r225, %r227;

dp4a.s32.s32 %r166, %r167, %r19, %r162;


{.reg .f16 low,high;
mov.b32 {low,high},%r20;
cvt.f32.f16 %f47, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r20;
cvt.f32.f16 %f48, high;}


cvt.rn.f32.s32 %f49, %r166;
mul.ftz.f32 %f50, %f48, 0fC1000000;
fma.rn.ftz.f32 %f51, %f47, %f49, %f50;
fma.rn.ftz.f32 %f117, %f46, %f51, %f117;

$L__BB62_12:
add.s64 %rd68, %rd68, 1408;
add.s32 %r293, %r293, 64;
setp.lt.s32 %p11, %r293, %r59;
@%p11 bra $L__BB62_10;

$L__BB62_13:
setp.eq.s32 %p12, %r1, 0;
@%p12 bra $L__BB62_16;

setp.eq.s64 %p13, %rd14, 0;
add.s32 %r22, %r1, -1;
shl.b32 %r232, %r1, 7;
mov.u32 %r233, _ZZ13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r234, %r233, %r232;
shl.b32 %r235, %r2, 2;
add.s32 %r236, %r234, %r235;
st.shared.f32 [%r236+-128], %f119;
@%p13 bra $L__BB62_16;

shl.b32 %r238, %r22, 7;
mov.u32 %r239, _ZZ13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r240, %r239, %r238;
add.s32 %r242, %r240, %r235;
st.shared.f32 [%r242], %f117;

$L__BB62_16:
setp.ne.s32 %p14, %r1, 0;
bar.sync 0;
@%p14 bra $L__BB62_35;

setp.ne.s64 %p15, %rd14, 0;
shl.b32 %r245, %r2, 2;
mov.u32 %r246, _ZZ13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r23, %r246, %r245;
ld.shared.f32 %f52, [%r23];
add.ftz.f32 %f13, %f119, %f52;
@%p15 bra $L__BB62_19;
bra.uni $L__BB62_18;

$L__BB62_19:
mov.u32 %r249, _ZZ13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r250, %r249, %r245;
ld.shared.f32 %f56, [%r250];
add.ftz.f32 %f57, %f117, %f56;
ld.shared.f32 %f58, [%r23+128];
add.ftz.f32 %f59, %f13, %f58;
ld.shared.f32 %f60, [%r250+128];
add.ftz.f32 %f61, %f57, %f60;
ld.shared.f32 %f62, [%r23+256];
add.ftz.f32 %f121, %f59, %f62;
ld.shared.f32 %f63, [%r250+256];
add.ftz.f32 %f117, %f61, %f63;
bra.uni $L__BB62_20;

$L__BB62_18:
ld.shared.f32 %f53, [%r23+128];
add.ftz.f32 %f54, %f13, %f53;
ld.shared.f32 %f55, [%r23+256];
add.ftz.f32 %f121, %f54, %f55;

$L__BB62_20:
mov.b32 %r251, %f121;
mov.u32 %r252, 31;
mov.u32 %r253, 16;
mov.u32 %r254, -1;
shfl.sync.bfly.b32 %r255|%p16, %r251, %r253, %r252, %r254;
mov.b32 %f64, %r255;
add.ftz.f32 %f65, %f121, %f64;
mov.b32 %r256, %f65;
mov.u32 %r257, 8;
shfl.sync.bfly.b32 %r258|%p17, %r256, %r257, %r252, %r254;
mov.b32 %f66, %r258;
add.ftz.f32 %f67, %f65, %f66;
mov.b32 %r259, %f67;
mov.u32 %r260, 4;
shfl.sync.bfly.b32 %r261|%p18, %r259, %r260, %r252, %r254;
mov.b32 %f68, %r261;
add.ftz.f32 %f69, %f67, %f68;
mov.b32 %r262, %f69;
mov.u32 %r263, 2;
shfl.sync.bfly.b32 %r264|%p19, %r262, %r263, %r252, %r254;
mov.b32 %f70, %r264;
add.ftz.f32 %f19, %f69, %f70;
mov.b32 %r265, %f19;
mov.u32 %r266, 1;
shfl.sync.bfly.b32 %r24|%p1, %r265, %r266, %r252, %r254;
setp.eq.s64 %p20, %rd14, 0;
@%p20 bra $L__BB62_22;

mov.b32 %r267, %f117;
shfl.sync.bfly.b32 %r271|%p21, %r267, %r253, %r252, %r254;
mov.b32 %f71, %r271;
add.ftz.f32 %f72, %f117, %f71;
mov.b32 %r272, %f72;
shfl.sync.bfly.b32 %r274|%p22, %r272, %r257, %r252, %r254;
mov.b32 %f73, %r274;
add.ftz.f32 %f74, %f72, %f73;
mov.b32 %r275, %f74;
shfl.sync.bfly.b32 %r277|%p23, %r275, %r260, %r252, %r254;
mov.b32 %f75, %r277;
add.ftz.f32 %f76, %f74, %f75;
mov.b32 %r278, %f76;
shfl.sync.bfly.b32 %r280|%p24, %r278, %r263, %r252, %r254;
mov.b32 %f77, %r280;
add.ftz.f32 %f78, %f76, %f77;
mov.b32 %r281, %f78;
shfl.sync.bfly.b32 %r283|%p25, %r281, %r266, %r252, %r254;
mov.b32 %f79, %r283;
add.ftz.f32 %f117, %f78, %f79;

$L__BB62_22:
setp.ne.s32 %p26, %r2, 0;
mov.b32 %f80, %r24;
add.ftz.f32 %f22, %f19, %f80;
@%p26 bra $L__BB62_35;

add.ftz.f32 %f123, %f114, %f22;
@%p20 bra $L__BB62_34;

add.ftz.f32 %f24, %f113, %f117;
setp.gt.s32 %p28, %r25, 2;
@%p28 bra $L__BB62_28;

setp.eq.s32 %p31, %r25, 1;
@%p31 bra $L__BB62_32;

setp.eq.s32 %p32, %r25, 2;
@%p32 bra $L__BB62_27;
bra.uni $L__BB62_33;

$L__BB62_27:
mul.ftz.f32 %f100, %f24, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f101, %f100;
add.ftz.f32 %f102, %f101, 0f3F800000;
div.approx.ftz.f32 %f103, %f24, %f102;
mul.ftz.f32 %f123, %f123, %f103;
bra.uni $L__BB62_34;

$L__BB62_28:
setp.eq.s32 %p29, %r25, 3;
@%p29 bra $L__BB62_31;

setp.ne.s32 %p30, %r25, 6;
@%p30 bra $L__BB62_33;

min.ftz.f32 %f81, %f24, %f31;
neg.ftz.f32 %f82, %f31;
min.ftz.f32 %f83, %f123, %f31;
max.ftz.f32 %f84, %f83, %f82;
mul.ftz.f32 %f85, %f81, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f86, %f85;
add.ftz.f32 %f87, %f86, 0f3F800000;
div.approx.ftz.f32 %f88, %f81, %f87;
mul.ftz.f32 %f123, %f84, %f88;
bra.uni $L__BB62_34;

$L__BB62_32:
mul.ftz.f32 %f104, %f24, 0f3F000000;
mul.ftz.f32 %f105, %f24, 0f3D372713;
fma.rn.ftz.f32 %f106, %f24, %f105, 0f3F800000;
mul.ftz.f32 %f107, %f24, 0f3F4C422A;
mul.ftz.f32 %f108, %f107, %f106;
tanh.approx.f32 %f109, %f108;
add.ftz.f32 %f110, %f109, 0f3F800000;
mul.ftz.f32 %f111, %f104, %f110;
mul.ftz.f32 %f123, %f123, %f111;
bra.uni $L__BB62_34;

$L__BB62_31:
mov.f32 %f89, 0f40E00000;
min.ftz.f32 %f90, %f24, %f89;
min.ftz.f32 %f91, %f123, %f89;
mov.f32 %f92, 0fC0E00000;
max.ftz.f32 %f93, %f91, %f92;
mul.ftz.f32 %f94, %f90, 0f3FD9DB23;
mul.ftz.f32 %f95, %f94, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f96, %f95;
add.ftz.f32 %f97, %f96, 0f3F800000;
div.approx.ftz.f32 %f98, %f90, %f97;
add.ftz.f32 %f99, %f93, 0f3F800000;
mul.ftz.f32 %f123, %f99, %f98;
bra.uni $L__BB62_34;

$L__BB62_33:
mul.ftz.f32 %f123, %f123, %f24;

$L__BB62_34:
cvta.to.global.u64 %rd65, %rd18;
mov.u32 %r286, %ctaid.z;
mov.u32 %r287, %ctaid.y;
mad.lo.s32 %r289, %r287, %r36, %r4;
mad.lo.s32 %r290, %r286, %r42, %r289;
mul.wide.u32 %rd66, %r290, 4;
add.s64 %rd67, %rd65, %rd66;
st.global.f32 [%rd67], %f123;

$L__BB62_35:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<12>;
.reg .b16 %rs<8>;
.reg .f32 %f<30>;
.reg .b32 %r<165>;
.reg .b64 %rd<35>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[384];

ld.param.u64 %rd9, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd10, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd11, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd12, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r18, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r19, [_Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
mov.u32 %r1, %tid.y;
shl.b32 %r35, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r35, %r2;
mov.u32 %r4, %ctaid.y;
setp.eq.s64 %p1, %rd11, 0;
@%p1 bra $L__BB63_2;

cvta.to.global.u64 %rd13, %rd11;
mul.wide.u32 %rd14, %r4, 4;
add.s64 %rd15, %rd13, %rd14;
ld.global.nc.u32 %r162, [%rd15];
mul.hi.u32 %r36, %r4, %r19;
add.s32 %r37, %r36, %r4;
shr.u32 %r38, %r37, %r20;
mul.lo.s32 %r39, %r38, %r21;
sub.s32 %r163, %r4, %r39;
bra.uni $L__BB63_3;

$L__BB63_2:
mul.hi.u32 %r40, %r4, %r23;
add.s32 %r41, %r40, %r4;
shr.u32 %r162, %r41, %r24;
mov.u32 %r163, %r4;

$L__BB63_3:
mov.u32 %r10, %ctaid.z;
shr.u32 %r42, %r3, 31;
add.s32 %r43, %r3, %r42;
shr.s32 %r164, %r43, 1;
shr.u32 %r44, %r18, 5;
setp.ge.s32 %p2, %r164, %r44;
mov.f32 %f29, 0f00000000;
@%p2 bra $L__BB63_6;

and.b32 %r47, %r43, -2;
sub.s32 %r48, %r3, %r47;
shl.b32 %r49, %r48, 1;
shl.b32 %r50, %r48, 2;
or.b32 %r51, %r50, 1;
shl.b32 %r12, %r48, 3;
cvt.s64.s32 %rd1, %r49;
sub.s32 %r52, %r51, %r49;
shl.b32 %r13, %r52, 2;
cvt.s64.s32 %rd2, %r52;
mul.hi.u32 %r53, %r10, %r29;
add.s32 %r54, %r53, %r10;
shr.u32 %r55, %r54, %r30;
mad.lo.s32 %r56, %r55, %r32, %r164;
mad.lo.s32 %r57, %r162, %r26, %r56;
mov.u32 %r58, %ctaid.x;
mad.lo.s32 %r59, %r58, %r22, %r57;
mul.wide.s32 %rd16, %r59, 22;
cvta.to.global.u64 %rd17, %rd9;
add.s64 %rd34, %rd17, %rd16;
mul.wide.s32 %rd18, %r50, 2;
add.s64 %rd4, %rd18, 8;
cvta.to.global.u64 %rd5, %rd10;
mul.lo.s32 %r60, %r163, %r27;
cvt.u64.u32 %rd19, %r60;
mul.lo.s32 %r61, %r10, %r33;
cvt.u64.u32 %rd20, %r61;
add.s64 %rd6, %rd19, %rd20;
mov.f32 %f29, 0f00000000;
shl.b64 %rd27, %rd1, 2;
shl.b64 %rd29, %rd2, 2;

$L__BB63_5:
cvt.u64.u32 %rd21, %r164;
add.s64 %rd22, %rd6, %rd21;
ld.global.nc.u16 %rs2, [%rd34+2];
ld.global.nc.u16 %rs3, [%rd34+4];
mov.b32 %r80, {%rs2, %rs3};
mul.lo.s64 %rd23, %rd22, 36;
add.s64 %rd24, %rd5, %rd23;
add.s64 %rd25, %rd24, 4;
add.s64 %rd26, %rd34, %rd4;
ld.global.nc.u16 %rs4, [%rd26+-2];
ld.global.nc.u16 %rs5, [%rd26];
mov.b32 %r81, {%rs4, %rs5};
shr.s32 %r82, %r80, %r12;
add.s64 %rd28, %rd25, %rd27;
ld.global.nc.u32 %r64, [%rd28];
ld.global.nc.u32 %r68, [%rd28+16];
ld.global.nc.u16 %rs6, [%rd26+2];
ld.global.nc.u16 %rs7, [%rd26+4];
mov.b32 %r83, {%rs6, %rs7};
shr.s32 %r84, %r80, %r13;
add.s64 %rd30, %rd25, %rd29;
ld.global.nc.u32 %r72, [%rd30];
ld.global.nc.u32 %r76, [%rd28+20];
ld.global.nc.u16 %rs1, [%rd34];

{ cvt.f32.f16 %f7, %rs1;}


and.b32 %r85, %r81, 252645135;
shl.b32 %r86, %r82, 4;
and.b32 %r87, %r86, 16;
or.b32 %r88, %r87, %r85;
shl.b32 %r89, %r82, 11;
and.b32 %r90, %r89, 4096;
or.b32 %r91, %r88, %r90;
shl.b32 %r92, %r82, 18;
and.b32 %r93, %r92, 1048576;
or.b32 %r94, %r91, %r93;
shl.b32 %r95, %r82, 25;
and.b32 %r96, %r95, 268435456;
or.b32 %r63, %r94, %r96;
mov.u32 %r65, 0;

dp4a.s32.s32 %r62, %r63, %r64, %r65;

shr.u32 %r97, %r81, 4;
and.b32 %r98, %r97, 252645135;
shr.u32 %r99, %r82, 12;
and.b32 %r100, %r99, 16;
or.b32 %r101, %r100, %r98;
shr.u32 %r102, %r82, 5;
and.b32 %r103, %r102, 4096;
or.b32 %r104, %r101, %r103;
shl.b32 %r105, %r82, 2;
and.b32 %r106, %r105, 1048576;
or.b32 %r107, %r104, %r106;
shl.b32 %r108, %r82, 9;
and.b32 %r109, %r108, 268435456;
or.b32 %r67, %r107, %r109;

dp4a.s32.s32 %r66, %r67, %r68, %r62;

and.b32 %r110, %r83, 252645135;
shl.b32 %r111, %r84, 4;
and.b32 %r112, %r111, 16;
or.b32 %r113, %r112, %r110;
shl.b32 %r114, %r84, 11;
and.b32 %r115, %r114, 4096;
or.b32 %r116, %r113, %r115;
shl.b32 %r117, %r84, 18;
and.b32 %r118, %r117, 1048576;
or.b32 %r119, %r116, %r118;
shl.b32 %r120, %r84, 25;
and.b32 %r121, %r120, 268435456;
or.b32 %r71, %r119, %r121;

dp4a.s32.s32 %r70, %r71, %r72, %r66;

shr.u32 %r122, %r83, 4;
and.b32 %r123, %r122, 252645135;
shr.u32 %r124, %r84, 12;
and.b32 %r125, %r124, 16;
or.b32 %r126, %r125, %r123;
shr.u32 %r127, %r84, 5;
and.b32 %r128, %r127, 4096;
or.b32 %r129, %r126, %r128;
shl.b32 %r130, %r84, 2;
and.b32 %r131, %r130, 1048576;
or.b32 %r132, %r129, %r131;
shl.b32 %r133, %r84, 9;
and.b32 %r134, %r133, 268435456;
or.b32 %r75, %r132, %r134;

dp4a.s32.s32 %r74, %r75, %r76, %r70;

ld.global.nc.u32 %r78, [%rd24];

{.reg .f16 low,high;
mov.b32 {low,high},%r78;
cvt.f32.f16 %f8, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r78;
cvt.f32.f16 %f9, high;}


cvt.rn.f32.s32 %f10, %r74;
mul.ftz.f32 %f11, %f9, 0fC1000000;
fma.rn.ftz.f32 %f12, %f8, %f10, %f11;
fma.rn.ftz.f32 %f29, %f7, %f12, %f29;
add.s64 %rd34, %rd34, 1408;
add.s32 %r164, %r164, 64;
setp.lt.s32 %p3, %r164, %r44;
@%p3 bra $L__BB63_5;

$L__BB63_6:
setp.eq.s32 %p4, %r1, 0;
@%p4 bra $L__BB63_8;

shl.b32 %r135, %r1, 7;
mov.u32 %r136, _ZZ13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r137, %r136, %r135;
shl.b32 %r138, %r2, 2;
add.s32 %r139, %r137, %r138;
st.shared.f32 [%r139+-128], %f29;

$L__BB63_8:
setp.ne.s32 %p5, %r1, 0;
bar.sync 0;
@%p5 bra $L__BB63_11;

mov.u32 %r140, %ctaid.x;
mad.lo.s32 %r141, %r4, %r28, %r140;
mad.lo.s32 %r17, %r10, %r34, %r141;
shl.b32 %r142, %r2, 2;
mov.u32 %r143, _ZZ13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r144, %r143, %r142;
mov.u32 %r145, 2;
ld.shared.f32 %f13, [%r144];
add.ftz.f32 %f14, %f29, %f13;
ld.shared.f32 %f15, [%r144+128];
add.ftz.f32 %f16, %f14, %f15;
ld.shared.f32 %f17, [%r144+256];
add.ftz.f32 %f18, %f16, %f17;
mov.b32 %r146, %f18;
mov.u32 %r147, 31;
mov.u32 %r148, 16;
mov.u32 %r149, -1;
shfl.sync.bfly.b32 %r150|%p6, %r146, %r148, %r147, %r149;
mov.b32 %f19, %r150;
add.ftz.f32 %f20, %f18, %f19;
mov.b32 %r151, %f20;
mov.u32 %r152, 8;
shfl.sync.bfly.b32 %r153|%p7, %r151, %r152, %r147, %r149;
mov.b32 %f21, %r153;
add.ftz.f32 %f22, %f20, %f21;
mov.b32 %r154, %f22;
mov.u32 %r155, 4;
shfl.sync.bfly.b32 %r156|%p8, %r154, %r155, %r147, %r149;
mov.b32 %f23, %r156;
add.ftz.f32 %f24, %f22, %f23;
mov.b32 %r157, %f24;
shfl.sync.bfly.b32 %r158|%p9, %r157, %r145, %r147, %r149;
mov.b32 %f25, %r158;
add.ftz.f32 %f26, %f24, %f25;
mov.b32 %r159, %f26;
mov.u32 %r160, 1;
shfl.sync.bfly.b32 %r161|%p10, %r159, %r160, %r147, %r149;
mov.b32 %f27, %r161;
add.ftz.f32 %f4, %f26, %f27;
setp.ne.s32 %p11, %r2, 0;
@%p11 bra $L__BB63_11;

cvta.to.global.u64 %rd31, %rd12;
mul.wide.u32 %rd32, %r17, 4;
add.s64 %rd33, %rd31, %rd32;
st.global.f32 [%rd33], %f4;

$L__BB63_11:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type6ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type6ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type6ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type6ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type6ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type6ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type6ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type6ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type6ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<35>;
.reg .b16 %rs<17>;
.reg .f32 %f<121>;
.reg .b32 %r<328>;
.reg .b64 %rd<53>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type6ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type6ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type6ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type6ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type6ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r16, [_Z13mul_mat_vec_qIL9ggml_type6ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r17, [_Z13mul_mat_vec_qIL9ggml_type6ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r18, [_Z13mul_mat_vec_qIL9ggml_type6ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type6ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type6ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type6ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type6ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type6ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type6ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type6ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type6ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type6ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r19, [_Z13mul_mat_vec_qIL9ggml_type6ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
cvta.to.global.u64 %rd1, %rd17;
mov.u32 %r1, %tid.y;
shl.b32 %r32, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r32, %r2;
mov.u32 %r33, %ctaid.x;
shl.b32 %r4, %r33, 1;
shr.u32 %r34, %r3, 31;
add.s32 %r35, %r3, %r34;
shr.s32 %r327, %r35, 1;
shr.u32 %r6, %r31, 5;
setp.ge.s32 %p3, %r327, %r6;
mov.f32 %f117, 0f00000000;
mov.f32 %f118, %f117;
mov.f32 %f119, %f117;
mov.f32 %f120, %f117;
@%p3 bra $L__BB64_3;

mov.u32 %r36, %ctaid.y;
mov.u32 %r37, %ctaid.z;
mul.lo.s32 %r38, %r36, %r23;
cvt.u64.u32 %rd18, %r38;
mul.lo.s32 %r39, %r37, %r29;
cvt.u64.u32 %rd19, %r39;
add.s64 %rd2, %rd19, %rd18;
mul.hi.u32 %r40, %r37, %r25;
add.s32 %r41, %r40, %r37;
shr.u32 %r42, %r41, %r26;
mul.lo.s32 %r43, %r42, %r28;
mul.hi.u32 %r44, %r36, %r19;
add.s32 %r45, %r44, %r36;
shr.u32 %r46, %r45, %r20;
mul.lo.s32 %r47, %r46, %r22;
cvta.to.global.u64 %rd52, %rd15;
and.b32 %r50, %r35, -2;
sub.s32 %r51, %r3, %r50;
shl.b32 %r52, %r51, 1;
shl.b32 %r53, %r51, 2;
or.b32 %r54, %r53, 1;
shl.b32 %r7, %r51, 3;
cvt.s64.s32 %rd4, %r52;
sub.s32 %r55, %r54, %r52;
shl.b32 %r8, %r55, 2;
cvt.s64.s32 %rd5, %r55;
add.s32 %r326, %r327, %r17;
mad.lo.s32 %r56, %r4, %r16, %r327;
add.s32 %r57, %r56, %r43;
add.s32 %r58, %r57, %r47;
mul.wide.s32 %rd6, %r58, 22;
mul.wide.s32 %rd20, %r53, 2;
add.s64 %rd21, %rd6, %rd20;
add.s64 %rd7, %rd21, 8;
add.s32 %r59, %r327, %r43;
add.s32 %r60, %r59, %r47;
add.s32 %r61, %r4, 1;
mad.lo.s32 %r62, %r16, %r61, %r60;
mul.wide.s32 %rd8, %r62, 22;
add.s64 %rd22, %rd8, %rd20;
add.s64 %rd9, %rd22, 6;
cvta.to.global.u64 %rd10, %rd16;
mov.f32 %f117, 0f00000000;
shl.b64 %rd30, %rd4, 2;
shl.b64 %rd32, %rd5, 2;

$L__BB64_2:
cvt.u64.u32 %rd23, %r327;
add.s64 %rd24, %rd2, %rd23;
mul.lo.s64 %rd25, %rd24, 36;
add.s64 %rd26, %rd10, %rd25;
add.s64 %rd27, %rd26, 4;
add.s64 %rd28, %rd52, %rd6;
ld.global.nc.u16 %rs5, [%rd28+2];
ld.global.nc.u16 %rs6, [%rd28+4];
mov.b32 %r135, {%rs5, %rs6};
add.s64 %rd29, %rd52, %rd7;
ld.global.nc.u16 %rs7, [%rd29+-2];
ld.global.nc.u16 %rs8, [%rd29];
mov.b32 %r136, {%rs7, %rs8};
shr.s32 %r137, %r135, %r7;
add.s64 %rd31, %rd27, %rd30;
ld.global.nc.u32 %r65, [%rd31];
ld.global.nc.u32 %r69, [%rd31+16];
ld.global.nc.u16 %rs9, [%rd29+2];
ld.global.nc.u16 %rs10, [%rd29+4];
mov.b32 %r138, {%rs9, %rs10};
shr.s32 %r139, %r135, %r8;
add.s64 %rd33, %rd27, %rd32;
ld.global.nc.u32 %r73, [%rd33];
ld.global.nc.u32 %r77, [%rd31+20];
ld.global.nc.u16 %rs1, [%rd28];

{ cvt.f32.f16 %f25, %rs1;}


and.b32 %r140, %r136, 252645135;
shl.b32 %r141, %r137, 4;
and.b32 %r142, %r141, 16;
or.b32 %r143, %r142, %r140;
shl.b32 %r144, %r137, 11;
and.b32 %r145, %r144, 4096;
or.b32 %r146, %r143, %r145;
shl.b32 %r147, %r137, 18;
and.b32 %r148, %r147, 1048576;
or.b32 %r149, %r146, %r148;
shl.b32 %r150, %r137, 25;
and.b32 %r151, %r150, 268435456;
or.b32 %r100, %r149, %r151;
mov.u32 %r120, 0;

dp4a.s32.s32 %r63, %r100, %r65, %r120;

shr.u32 %r152, %r136, 4;
and.b32 %r153, %r152, 252645135;
shr.u32 %r154, %r137, 12;
and.b32 %r155, %r154, 16;
or.b32 %r156, %r155, %r153;
shr.u32 %r157, %r137, 5;
and.b32 %r158, %r157, 4096;
or.b32 %r159, %r156, %r158;
shl.b32 %r160, %r137, 2;
and.b32 %r161, %r160, 1048576;
or.b32 %r162, %r159, %r161;
shl.b32 %r163, %r137, 9;
and.b32 %r164, %r163, 268435456;
or.b32 %r104, %r162, %r164;

dp4a.s32.s32 %r67, %r104, %r69, %r63;

and.b32 %r165, %r138, 252645135;
shl.b32 %r166, %r139, 4;
and.b32 %r167, %r166, 16;
or.b32 %r168, %r167, %r165;
shl.b32 %r169, %r139, 11;
and.b32 %r170, %r169, 4096;
or.b32 %r171, %r168, %r170;
shl.b32 %r172, %r139, 18;
and.b32 %r173, %r172, 1048576;
or.b32 %r174, %r171, %r173;
shl.b32 %r175, %r139, 25;
and.b32 %r176, %r175, 268435456;
or.b32 %r108, %r174, %r176;

dp4a.s32.s32 %r71, %r108, %r73, %r67;

shr.u32 %r177, %r138, 4;
and.b32 %r178, %r177, 252645135;
shr.u32 %r179, %r139, 12;
and.b32 %r180, %r179, 16;
or.b32 %r181, %r180, %r178;
shr.u32 %r182, %r139, 5;
and.b32 %r183, %r182, 4096;
or.b32 %r184, %r181, %r183;
shl.b32 %r185, %r139, 2;
and.b32 %r186, %r185, 1048576;
or.b32 %r187, %r184, %r186;
shl.b32 %r188, %r139, 9;
and.b32 %r189, %r188, 268435456;
or.b32 %r112, %r187, %r189;

dp4a.s32.s32 %r75, %r112, %r77, %r71;

ld.global.nc.u32 %r79, [%rd26];

{.reg .f16 low,high;
mov.b32 {low,high},%r79;
cvt.f32.f16 %f26, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r79;
cvt.f32.f16 %f27, high;}


cvt.rn.f32.s32 %f37, %r75;
mul.ftz.f32 %f38, %f26, %f37;
mul.ftz.f32 %f39, %f27, 0f41000000;
sub.ftz.f32 %f40, %f38, %f39;
fma.rn.ftz.f32 %f120, %f25, %f40, %f120;
add.s64 %rd34, %rd52, %rd8;
ld.global.nc.u16 %rs11, [%rd34+2];
ld.global.nc.u16 %rs12, [%rd34+4];
mov.b32 %r190, {%rs11, %rs12};
add.s64 %rd35, %rd52, %rd9;
ld.global.nc.u16 %rs13, [%rd35];
ld.global.nc.u16 %rs14, [%rd35+2];
mov.b32 %r191, {%rs13, %rs14};
shr.s32 %r192, %r190, %r7;
ld.global.nc.u16 %rs15, [%rd35+4];
ld.global.nc.u16 %rs16, [%rd35+6];
mov.b32 %r193, {%rs15, %rs16};
shr.s32 %r194, %r190, %r8;
ld.global.nc.u16 %rs2, [%rd34];

{ cvt.f32.f16 %f28, %rs2;}


and.b32 %r195, %r191, 252645135;
shl.b32 %r196, %r192, 4;
and.b32 %r197, %r196, 16;
or.b32 %r198, %r197, %r195;
shl.b32 %r199, %r192, 11;
and.b32 %r200, %r199, 4096;
or.b32 %r201, %r198, %r200;
shl.b32 %r202, %r192, 18;
and.b32 %r203, %r202, 1048576;
or.b32 %r204, %r201, %r203;
shl.b32 %r205, %r192, 25;
and.b32 %r206, %r205, 268435456;
or.b32 %r118, %r204, %r206;

dp4a.s32.s32 %r81, %r118, %r65, %r120;

shr.u32 %r207, %r191, 4;
and.b32 %r208, %r207, 252645135;
shr.u32 %r209, %r192, 12;
and.b32 %r210, %r209, 16;
or.b32 %r211, %r210, %r208;
shr.u32 %r212, %r192, 5;
and.b32 %r213, %r212, 4096;
or.b32 %r214, %r211, %r213;
shl.b32 %r215, %r192, 2;
and.b32 %r216, %r215, 1048576;
or.b32 %r217, %r214, %r216;
shl.b32 %r218, %r192, 9;
and.b32 %r219, %r218, 268435456;
or.b32 %r122, %r217, %r219;

dp4a.s32.s32 %r85, %r122, %r69, %r81;

and.b32 %r220, %r193, 252645135;
shl.b32 %r221, %r194, 4;
and.b32 %r222, %r221, 16;
or.b32 %r223, %r222, %r220;
shl.b32 %r224, %r194, 11;
and.b32 %r225, %r224, 4096;
or.b32 %r226, %r223, %r225;
shl.b32 %r227, %r194, 18;
and.b32 %r228, %r227, 1048576;
or.b32 %r229, %r226, %r228;
shl.b32 %r230, %r194, 25;
and.b32 %r231, %r230, 268435456;
or.b32 %r126, %r229, %r231;

dp4a.s32.s32 %r89, %r126, %r73, %r85;

shr.u32 %r232, %r193, 4;
and.b32 %r233, %r232, 252645135;
shr.u32 %r234, %r194, 12;
and.b32 %r235, %r234, 16;
or.b32 %r236, %r235, %r233;
shr.u32 %r237, %r194, 5;
and.b32 %r238, %r237, 4096;
or.b32 %r239, %r236, %r238;
shl.b32 %r240, %r194, 2;
and.b32 %r241, %r240, 1048576;
or.b32 %r242, %r239, %r241;
shl.b32 %r243, %r194, 9;
and.b32 %r244, %r243, 268435456;
or.b32 %r130, %r242, %r244;

dp4a.s32.s32 %r93, %r130, %r77, %r89;


{.reg .f16 low,high;
mov.b32 {low,high},%r79;
cvt.f32.f16 %f29, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r79;
cvt.f32.f16 %f30, high;}


cvt.rn.f32.s32 %f41, %r93;
mul.ftz.f32 %f42, %f29, %f41;
mul.ftz.f32 %f43, %f30, 0f41000000;
sub.ftz.f32 %f44, %f42, %f43;
fma.rn.ftz.f32 %f119, %f28, %f44, %f119;
cvt.u64.u32 %rd36, %r326;
add.s64 %rd37, %rd2, %rd36;
mul.lo.s64 %rd38, %rd37, 36;
add.s64 %rd39, %rd10, %rd38;
add.s64 %rd40, %rd39, 4;
add.s64 %rd41, %rd40, %rd30;
ld.global.nc.u32 %r101, [%rd41];
ld.global.nc.u32 %r105, [%rd41+16];
add.s64 %rd42, %rd40, %rd32;
ld.global.nc.u32 %r109, [%rd42];
ld.global.nc.u32 %r113, [%rd41+20];

{ cvt.f32.f16 %f31, %rs1;}



dp4a.s32.s32 %r99, %r100, %r101, %r120;


dp4a.s32.s32 %r103, %r104, %r105, %r99;


dp4a.s32.s32 %r107, %r108, %r109, %r103;


dp4a.s32.s32 %r111, %r112, %r113, %r107;

ld.global.nc.u32 %r115, [%rd39];

{.reg .f16 low,high;
mov.b32 {low,high},%r115;
cvt.f32.f16 %f32, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r115;
cvt.f32.f16 %f33, high;}


cvt.rn.f32.s32 %f45, %r111;
mul.ftz.f32 %f46, %f32, %f45;
mul.ftz.f32 %f47, %f33, 0f41000000;
sub.ftz.f32 %f48, %f46, %f47;
fma.rn.ftz.f32 %f118, %f31, %f48, %f118;

{ cvt.f32.f16 %f34, %rs2;}



dp4a.s32.s32 %r117, %r118, %r101, %r120;


dp4a.s32.s32 %r121, %r122, %r105, %r117;


dp4a.s32.s32 %r125, %r126, %r109, %r121;


dp4a.s32.s32 %r129, %r130, %r113, %r125;


{.reg .f16 low,high;
mov.b32 {low,high},%r115;
cvt.f32.f16 %f35, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r115;
cvt.f32.f16 %f36, high;}


cvt.rn.f32.s32 %f49, %r129;
mul.ftz.f32 %f50, %f35, %f49;
mul.ftz.f32 %f51, %f36, 0f41000000;
sub.ftz.f32 %f52, %f50, %f51;
fma.rn.ftz.f32 %f117, %f34, %f52, %f117;
add.s32 %r326, %r326, 64;
add.s64 %rd52, %rd52, 1408;
add.s32 %r327, %r327, 64;
setp.lt.s32 %p4, %r327, %r6;
@%p4 bra $L__BB64_2;

$L__BB64_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB64_5;

shl.b32 %r245, %r1, 9;
mov.u32 %r246, _ZZ13mul_mat_vec_qIL9ggml_type6ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r247, %r246, %r245;
shl.b32 %r248, %r2, 2;
add.s32 %r249, %r247, %r248;
st.shared.f32 [%r249+-512], %f120;
st.shared.f32 [%r249+-384], %f119;
st.shared.f32 [%r249+-256], %f118;
st.shared.f32 [%r249+-128], %f117;

$L__BB64_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB64_14;

mov.u32 %r250, %ctaid.z;
mov.u32 %r251, %ctaid.y;
mad.lo.s32 %r252, %r251, %r24, %r4;
mad.lo.s32 %r253, %r250, %r30, %r252;
cvt.u64.u32 %rd13, %r253;
shl.b32 %r254, %r2, 2;
mov.u32 %r255, _ZZ13mul_mat_vec_qIL9ggml_type6ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r14, %r255, %r254;
mov.u32 %r256, 2;
ld.shared.f32 %f53, [%r14];
add.ftz.f32 %f54, %f53, %f120;
ld.shared.f32 %f55, [%r14+512];
add.ftz.f32 %f56, %f55, %f54;
ld.shared.f32 %f57, [%r14+1024];
add.ftz.f32 %f58, %f57, %f56;
mov.b32 %r257, %f58;
mov.u32 %r258, 31;
mov.u32 %r259, 16;
mov.u32 %r260, -1;
shfl.sync.bfly.b32 %r261|%p7, %r257, %r259, %r258, %r260;
mov.b32 %f59, %r261;
add.ftz.f32 %f60, %f58, %f59;
mov.b32 %r262, %f60;
mov.u32 %r263, 8;
shfl.sync.bfly.b32 %r264|%p8, %r262, %r263, %r258, %r260;
mov.b32 %f61, %r264;
add.ftz.f32 %f62, %f60, %f61;
mov.b32 %r265, %f62;
mov.u32 %r266, 4;
shfl.sync.bfly.b32 %r267|%p9, %r265, %r266, %r258, %r260;
mov.b32 %f63, %r267;
add.ftz.f32 %f64, %f62, %f63;
mov.b32 %r268, %f64;
shfl.sync.bfly.b32 %r269|%p10, %r268, %r256, %r258, %r260;
mov.b32 %f65, %r269;
add.ftz.f32 %f66, %f64, %f65;
mov.b32 %r270, %f66;
mov.u32 %r271, 1;
shfl.sync.bfly.b32 %r272|%p11, %r270, %r271, %r258, %r260;
mov.b32 %f67, %r272;
add.ftz.f32 %f13, %f66, %f67;
setp.lt.u32 %p12, %r4, %r18;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
mul.wide.u32 %rd43, %r253, 4;
add.s64 %rd14, %rd1, %rd43;
not.pred %p14, %p1;
@%p14 bra $L__BB64_8;

st.global.f32 [%rd14], %f13;

$L__BB64_8:
ld.shared.f32 %f68, [%r14+128];
add.ftz.f32 %f69, %f68, %f119;
ld.shared.f32 %f70, [%r14+640];
add.ftz.f32 %f71, %f70, %f69;
ld.shared.f32 %f72, [%r14+1152];
add.ftz.f32 %f73, %f72, %f71;
mov.b32 %r273, %f73;
shfl.sync.bfly.b32 %r277|%p15, %r273, %r259, %r258, %r260;
mov.b32 %f74, %r277;
add.ftz.f32 %f75, %f73, %f74;
mov.b32 %r278, %f75;
shfl.sync.bfly.b32 %r280|%p16, %r278, %r263, %r258, %r260;
mov.b32 %f76, %r280;
add.ftz.f32 %f77, %f75, %f76;
mov.b32 %r281, %f77;
shfl.sync.bfly.b32 %r283|%p17, %r281, %r266, %r258, %r260;
mov.b32 %f78, %r283;
add.ftz.f32 %f79, %f77, %f78;
mov.b32 %r284, %f79;
shfl.sync.bfly.b32 %r286|%p18, %r284, %r256, %r258, %r260;
mov.b32 %f80, %r286;
add.ftz.f32 %f81, %f79, %f80;
mov.b32 %r287, %f81;
shfl.sync.bfly.b32 %r289|%p19, %r287, %r271, %r258, %r260;
mov.b32 %f82, %r289;
add.ftz.f32 %f14, %f81, %f82;
add.s32 %r290, %r4, 1;
setp.lt.u32 %p20, %r290, %r18;
setp.eq.s32 %p21, %r2, 1;
and.pred %p2, %p21, %p20;
not.pred %p22, %p2;
@%p22 bra $L__BB64_10;

st.global.f32 [%rd14+4], %f14;

$L__BB64_10:
ld.shared.f32 %f83, [%r14+256];
add.ftz.f32 %f84, %f83, %f118;
ld.shared.f32 %f85, [%r14+768];
add.ftz.f32 %f86, %f85, %f84;
ld.shared.f32 %f87, [%r14+1280];
add.ftz.f32 %f88, %f87, %f86;
mov.b32 %r291, %f88;
mov.u32 %r292, 31;
mov.u32 %r293, 16;
mov.u32 %r294, -1;
shfl.sync.bfly.b32 %r295|%p23, %r291, %r293, %r292, %r294;
mov.b32 %f89, %r295;
add.ftz.f32 %f90, %f88, %f89;
mov.b32 %r296, %f90;
mov.u32 %r297, 8;
shfl.sync.bfly.b32 %r298|%p24, %r296, %r297, %r292, %r294;
mov.b32 %f91, %r298;
add.ftz.f32 %f92, %f90, %f91;
mov.b32 %r299, %f92;
mov.u32 %r300, 4;
shfl.sync.bfly.b32 %r301|%p25, %r299, %r300, %r292, %r294;
mov.b32 %f93, %r301;
add.ftz.f32 %f94, %f92, %f93;
mov.b32 %r302, %f94;
mov.u32 %r303, 2;
shfl.sync.bfly.b32 %r304|%p26, %r302, %r303, %r292, %r294;
mov.b32 %f95, %r304;
add.ftz.f32 %f96, %f94, %f95;
mov.b32 %r305, %f96;
mov.u32 %r306, 1;
shfl.sync.bfly.b32 %r307|%p27, %r305, %r306, %r292, %r294;
mov.b32 %f97, %r307;
add.ftz.f32 %f15, %f96, %f97;
@%p14 bra $L__BB64_12;

cvt.u64.u32 %rd44, %r18;
add.s64 %rd45, %rd44, %rd13;
shl.b64 %rd46, %rd45, 2;
add.s64 %rd47, %rd1, %rd46;
st.global.f32 [%rd47], %f15;

$L__BB64_12:
ld.shared.f32 %f98, [%r14+384];
add.ftz.f32 %f99, %f98, %f117;
ld.shared.f32 %f100, [%r14+896];
add.ftz.f32 %f101, %f100, %f99;
ld.shared.f32 %f102, [%r14+1408];
add.ftz.f32 %f103, %f102, %f101;
mov.b32 %r308, %f103;
shfl.sync.bfly.b32 %r312|%p29, %r308, %r293, %r292, %r294;
mov.b32 %f104, %r312;
add.ftz.f32 %f105, %f103, %f104;
mov.b32 %r313, %f105;
shfl.sync.bfly.b32 %r315|%p30, %r313, %r297, %r292, %r294;
mov.b32 %f106, %r315;
add.ftz.f32 %f107, %f105, %f106;
mov.b32 %r316, %f107;
shfl.sync.bfly.b32 %r318|%p31, %r316, %r300, %r292, %r294;
mov.b32 %f108, %r318;
add.ftz.f32 %f109, %f107, %f108;
mov.b32 %r319, %f109;
shfl.sync.bfly.b32 %r321|%p32, %r319, %r303, %r292, %r294;
mov.b32 %f110, %r321;
add.ftz.f32 %f111, %f109, %f110;
mov.b32 %r322, %f111;
shfl.sync.bfly.b32 %r324|%p33, %r322, %r306, %r292, %r294;
mov.b32 %f112, %r324;
add.ftz.f32 %f16, %f111, %f112;
@%p22 bra $L__BB64_14;

add.s32 %r325, %r18, 1;
cvt.u64.u32 %rd48, %r325;
add.s64 %rd49, %rd48, %rd13;
shl.b64 %rd50, %rd49, 2;
add.s64 %rd51, %rd1, %rd50;
st.global.f32 [%rd51], %f16;

$L__BB64_14:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type6ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type6ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type6ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type6ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type6ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type6ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type6ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type6ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type6ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<47>;
.reg .b16 %rs<19>;
.reg .f32 %f<181>;
.reg .b32 %r<403>;
.reg .b64 %rd<63>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type6ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[2304];

ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type6ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type6ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type6ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type6ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type6ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type6ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type6ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type6ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type6ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type6ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type6ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type6ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type6ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type6ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type6ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type6ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type6ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
cvta.to.global.u64 %rd1, %rd17;
mov.u32 %r1, %tid.y;
shl.b32 %r36, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r36, %r2;
mov.u32 %r37, %ctaid.x;
shl.b32 %r4, %r37, 1;
mov.u32 %r5, %ctaid.y;
mov.u32 %r6, %ctaid.z;
shr.u32 %r38, %r3, 31;
add.s32 %r39, %r3, %r38;
shr.s32 %r402, %r39, 1;
shr.u32 %r40, %r20, 5;
setp.ge.s32 %p3, %r402, %r40;
mov.f32 %f175, 0f00000000;
mov.f32 %f176, %f175;
mov.f32 %f177, %f175;
mov.f32 %f178, %f175;
mov.f32 %f179, %f175;
mov.f32 %f180, %f175;
@%p3 bra $L__BB65_3;

mul.lo.s32 %r41, %r5, %r28;
cvt.u64.u32 %rd18, %r41;
mul.lo.s32 %r42, %r6, %r34;
cvt.u64.u32 %rd19, %r42;
add.s64 %rd2, %rd19, %rd18;
mul.hi.u32 %r43, %r6, %r30;
add.s32 %r44, %r43, %r6;
shr.u32 %r45, %r44, %r31;
mul.lo.s32 %r46, %r45, %r33;
mul.hi.u32 %r47, %r5, %r24;
add.s32 %r48, %r47, %r5;
shr.u32 %r49, %r48, %r25;
mul.lo.s32 %r50, %r49, %r27;
cvta.to.global.u64 %rd62, %rd15;
and.b32 %r53, %r39, -2;
sub.s32 %r8, %r3, %r53;
shl.b32 %r54, %r8, 1;
shl.b32 %r55, %r8, 2;
or.b32 %r56, %r55, 1;
sub.s32 %r57, %r56, %r54;
shl.b32 %r9, %r57, 2;
cvt.s64.s32 %rd4, %r57;
shl.b32 %r58, %r22, 1;
add.s32 %r401, %r402, %r58;
add.s32 %r400, %r402, %r22;
mad.lo.s32 %r59, %r4, %r21, %r402;
add.s32 %r60, %r59, %r46;
add.s32 %r61, %r60, %r50;
mul.wide.s32 %rd5, %r61, 22;
mul.wide.s32 %rd20, %r55, 2;
add.s64 %rd21, %rd5, %rd20;
add.s64 %rd6, %rd21, 8;
add.s32 %r62, %r402, %r46;
add.s32 %r63, %r62, %r50;
add.s32 %r64, %r4, 1;
mad.lo.s32 %r65, %r21, %r64, %r63;
mul.wide.s32 %rd7, %r65, 22;
add.s64 %rd22, %rd7, %rd20;
add.s64 %rd8, %rd22, 6;
cvta.to.global.u64 %rd9, %rd16;
mov.f32 %f175, 0f00000000;
shl.b64 %rd32, %rd4, 2;

$L__BB65_2:
cvt.u64.u32 %rd23, %r402;
add.s64 %rd24, %rd2, %rd23;
mul.lo.s64 %rd25, %rd24, 36;
add.s64 %rd26, %rd9, %rd25;
add.s64 %rd27, %rd26, 4;
add.s64 %rd28, %rd62, %rd5;
ld.global.nc.u16 %rs7, [%rd28+2];
ld.global.nc.u16 %rs8, [%rd28+4];
mov.b32 %r174, {%rs7, %rs8};
add.s64 %rd29, %rd62, %rd6;
ld.global.nc.u16 %rs9, [%rd29+-2];
ld.global.nc.u16 %rs10, [%rd29];
mov.b32 %r175, {%rs9, %rs10};
shl.b32 %r176, %r8, 3;
shr.s32 %r177, %r174, %r176;
mul.wide.s32 %rd30, %r54, 4;
add.s64 %rd31, %rd27, %rd30;
ld.global.nc.u32 %r68, [%rd31];
ld.global.nc.u32 %r72, [%rd31+16];
ld.global.nc.u16 %rs11, [%rd29+2];
ld.global.nc.u16 %rs12, [%rd29+4];
mov.b32 %r179, {%rs11, %rs12};
shr.s32 %r180, %r174, %r9;
add.s64 %rd33, %rd27, %rd32;
ld.global.nc.u32 %r76, [%rd33];
ld.global.nc.u32 %r80, [%rd31+20];
ld.global.nc.u16 %rs1, [%rd28];

{ cvt.f32.f16 %f37, %rs1;}


and.b32 %r181, %r175, 252645135;
shl.b32 %r182, %r177, 4;
and.b32 %r183, %r182, 16;
or.b32 %r184, %r183, %r181;
shl.b32 %r185, %r177, 11;
and.b32 %r186, %r185, 4096;
or.b32 %r187, %r184, %r186;
shl.b32 %r188, %r177, 18;
and.b32 %r189, %r188, 1048576;
or.b32 %r190, %r187, %r189;
shl.b32 %r191, %r177, 25;
and.b32 %r192, %r191, 268435456;
or.b32 %r139, %r190, %r192;
mov.u32 %r159, 0;

dp4a.s32.s32 %r66, %r139, %r68, %r159;

shr.u32 %r193, %r175, 4;
and.b32 %r194, %r193, 252645135;
shr.u32 %r195, %r177, 12;
and.b32 %r196, %r195, 16;
or.b32 %r197, %r196, %r194;
shr.u32 %r198, %r177, 5;
and.b32 %r199, %r198, 4096;
or.b32 %r200, %r197, %r199;
shl.b32 %r201, %r177, 2;
and.b32 %r202, %r201, 1048576;
or.b32 %r203, %r200, %r202;
shl.b32 %r204, %r177, 9;
and.b32 %r205, %r204, 268435456;
or.b32 %r143, %r203, %r205;

dp4a.s32.s32 %r70, %r143, %r72, %r66;

and.b32 %r206, %r179, 252645135;
shl.b32 %r207, %r180, 4;
and.b32 %r208, %r207, 16;
or.b32 %r209, %r208, %r206;
shl.b32 %r210, %r180, 11;
and.b32 %r211, %r210, 4096;
or.b32 %r212, %r209, %r211;
shl.b32 %r213, %r180, 18;
and.b32 %r214, %r213, 1048576;
or.b32 %r215, %r212, %r214;
shl.b32 %r216, %r180, 25;
and.b32 %r217, %r216, 268435456;
or.b32 %r147, %r215, %r217;

dp4a.s32.s32 %r74, %r147, %r76, %r70;

shr.u32 %r218, %r179, 4;
and.b32 %r219, %r218, 252645135;
shr.u32 %r220, %r180, 12;
and.b32 %r221, %r220, 16;
or.b32 %r222, %r221, %r219;
shr.u32 %r223, %r180, 5;
and.b32 %r224, %r223, 4096;
or.b32 %r225, %r222, %r224;
shl.b32 %r226, %r180, 2;
and.b32 %r227, %r226, 1048576;
or.b32 %r228, %r225, %r227;
shl.b32 %r229, %r180, 9;
and.b32 %r230, %r229, 268435456;
or.b32 %r151, %r228, %r230;

dp4a.s32.s32 %r78, %r151, %r80, %r74;

ld.global.nc.u32 %r82, [%rd26];

{.reg .f16 low,high;
mov.b32 {low,high},%r82;
cvt.f32.f16 %f38, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r82;
cvt.f32.f16 %f39, high;}


cvt.rn.f32.s32 %f55, %r78;
mul.ftz.f32 %f56, %f38, %f55;
mul.ftz.f32 %f57, %f39, 0f41000000;
sub.ftz.f32 %f58, %f56, %f57;
fma.rn.ftz.f32 %f180, %f37, %f58, %f180;
add.s64 %rd34, %rd62, %rd7;
ld.global.nc.u16 %rs13, [%rd34+2];
ld.global.nc.u16 %rs14, [%rd34+4];
mov.b32 %r231, {%rs13, %rs14};
add.s64 %rd35, %rd62, %rd8;
ld.global.nc.u16 %rs15, [%rd35];
ld.global.nc.u16 %rs16, [%rd35+2];
mov.b32 %r232, {%rs15, %rs16};
shr.s32 %r233, %r231, %r176;
ld.global.nc.u16 %rs17, [%rd35+4];
ld.global.nc.u16 %rs18, [%rd35+6];
mov.b32 %r234, {%rs17, %rs18};
shr.s32 %r235, %r231, %r9;
ld.global.nc.u16 %rs2, [%rd34];

{ cvt.f32.f16 %f40, %rs2;}


and.b32 %r236, %r232, 252645135;
shl.b32 %r237, %r233, 4;
and.b32 %r238, %r237, 16;
or.b32 %r239, %r238, %r236;
shl.b32 %r240, %r233, 11;
and.b32 %r241, %r240, 4096;
or.b32 %r242, %r239, %r241;
shl.b32 %r243, %r233, 18;
and.b32 %r244, %r243, 1048576;
or.b32 %r245, %r242, %r244;
shl.b32 %r246, %r233, 25;
and.b32 %r247, %r246, 268435456;
or.b32 %r157, %r245, %r247;

dp4a.s32.s32 %r84, %r157, %r68, %r159;

shr.u32 %r248, %r232, 4;
and.b32 %r249, %r248, 252645135;
shr.u32 %r250, %r233, 12;
and.b32 %r251, %r250, 16;
or.b32 %r252, %r251, %r249;
shr.u32 %r253, %r233, 5;
and.b32 %r254, %r253, 4096;
or.b32 %r255, %r252, %r254;
shl.b32 %r256, %r233, 2;
and.b32 %r257, %r256, 1048576;
or.b32 %r258, %r255, %r257;
shl.b32 %r259, %r233, 9;
and.b32 %r260, %r259, 268435456;
or.b32 %r161, %r258, %r260;

dp4a.s32.s32 %r88, %r161, %r72, %r84;

and.b32 %r261, %r234, 252645135;
shl.b32 %r262, %r235, 4;
and.b32 %r263, %r262, 16;
or.b32 %r264, %r263, %r261;
shl.b32 %r265, %r235, 11;
and.b32 %r266, %r265, 4096;
or.b32 %r267, %r264, %r266;
shl.b32 %r268, %r235, 18;
and.b32 %r269, %r268, 1048576;
or.b32 %r270, %r267, %r269;
shl.b32 %r271, %r235, 25;
and.b32 %r272, %r271, 268435456;
or.b32 %r165, %r270, %r272;

dp4a.s32.s32 %r92, %r165, %r76, %r88;

shr.u32 %r273, %r234, 4;
and.b32 %r274, %r273, 252645135;
shr.u32 %r275, %r235, 12;
and.b32 %r276, %r275, 16;
or.b32 %r277, %r276, %r274;
shr.u32 %r278, %r235, 5;
and.b32 %r279, %r278, 4096;
or.b32 %r280, %r277, %r279;
shl.b32 %r281, %r235, 2;
and.b32 %r282, %r281, 1048576;
or.b32 %r283, %r280, %r282;
shl.b32 %r284, %r235, 9;
and.b32 %r285, %r284, 268435456;
or.b32 %r169, %r283, %r285;

dp4a.s32.s32 %r96, %r169, %r80, %r92;


{.reg .f16 low,high;
mov.b32 {low,high},%r82;
cvt.f32.f16 %f41, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r82;
cvt.f32.f16 %f42, high;}


cvt.rn.f32.s32 %f59, %r96;
mul.ftz.f32 %f60, %f41, %f59;
mul.ftz.f32 %f61, %f42, 0f41000000;
sub.ftz.f32 %f62, %f60, %f61;
fma.rn.ftz.f32 %f179, %f40, %f62, %f179;
cvt.u64.u32 %rd36, %r400;
add.s64 %rd37, %rd2, %rd36;
mul.lo.s64 %rd38, %rd37, 36;
add.s64 %rd39, %rd9, %rd38;
add.s64 %rd40, %rd39, 4;
add.s64 %rd41, %rd40, %rd30;
ld.global.nc.u32 %r104, [%rd41];
ld.global.nc.u32 %r108, [%rd41+16];
add.s64 %rd42, %rd40, %rd32;
ld.global.nc.u32 %r112, [%rd42];
ld.global.nc.u32 %r116, [%rd41+20];

{ cvt.f32.f16 %f43, %rs1;}



dp4a.s32.s32 %r102, %r139, %r104, %r159;


dp4a.s32.s32 %r106, %r143, %r108, %r102;


dp4a.s32.s32 %r110, %r147, %r112, %r106;


dp4a.s32.s32 %r114, %r151, %r116, %r110;

ld.global.nc.u32 %r118, [%rd39];

{.reg .f16 low,high;
mov.b32 {low,high},%r118;
cvt.f32.f16 %f44, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r118;
cvt.f32.f16 %f45, high;}


cvt.rn.f32.s32 %f63, %r114;
mul.ftz.f32 %f64, %f44, %f63;
mul.ftz.f32 %f65, %f45, 0f41000000;
sub.ftz.f32 %f66, %f64, %f65;
fma.rn.ftz.f32 %f178, %f43, %f66, %f178;

{ cvt.f32.f16 %f46, %rs2;}



dp4a.s32.s32 %r120, %r157, %r104, %r159;


dp4a.s32.s32 %r124, %r161, %r108, %r120;


dp4a.s32.s32 %r128, %r165, %r112, %r124;


dp4a.s32.s32 %r132, %r169, %r116, %r128;


{.reg .f16 low,high;
mov.b32 {low,high},%r118;
cvt.f32.f16 %f47, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r118;
cvt.f32.f16 %f48, high;}


cvt.rn.f32.s32 %f67, %r132;
mul.ftz.f32 %f68, %f47, %f67;
mul.ftz.f32 %f69, %f48, 0f41000000;
sub.ftz.f32 %f70, %f68, %f69;
fma.rn.ftz.f32 %f177, %f46, %f70, %f177;
cvt.u64.u32 %rd43, %r401;
add.s64 %rd44, %rd2, %rd43;
mul.lo.s64 %rd45, %rd44, 36;
add.s64 %rd46, %rd9, %rd45;
add.s64 %rd47, %rd46, 4;
add.s64 %rd48, %rd47, %rd30;
ld.global.nc.u32 %r140, [%rd48];
ld.global.nc.u32 %r144, [%rd48+16];
add.s64 %rd49, %rd47, %rd32;
ld.global.nc.u32 %r148, [%rd49];
ld.global.nc.u32 %r152, [%rd48+20];

{ cvt.f32.f16 %f49, %rs1;}



dp4a.s32.s32 %r138, %r139, %r140, %r159;


dp4a.s32.s32 %r142, %r143, %r144, %r138;


dp4a.s32.s32 %r146, %r147, %r148, %r142;


dp4a.s32.s32 %r150, %r151, %r152, %r146;

ld.global.nc.u32 %r154, [%rd46];

{.reg .f16 low,high;
mov.b32 {low,high},%r154;
cvt.f32.f16 %f50, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r154;
cvt.f32.f16 %f51, high;}


cvt.rn.f32.s32 %f71, %r150;
mul.ftz.f32 %f72, %f50, %f71;
mul.ftz.f32 %f73, %f51, 0f41000000;
sub.ftz.f32 %f74, %f72, %f73;
fma.rn.ftz.f32 %f176, %f49, %f74, %f176;

{ cvt.f32.f16 %f52, %rs2;}



dp4a.s32.s32 %r156, %r157, %r140, %r159;


dp4a.s32.s32 %r160, %r161, %r144, %r156;


dp4a.s32.s32 %r164, %r165, %r148, %r160;


dp4a.s32.s32 %r168, %r169, %r152, %r164;


{.reg .f16 low,high;
mov.b32 {low,high},%r154;
cvt.f32.f16 %f53, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r154;
cvt.f32.f16 %f54, high;}


cvt.rn.f32.s32 %f75, %r168;
mul.ftz.f32 %f76, %f53, %f75;
mul.ftz.f32 %f77, %f54, 0f41000000;
sub.ftz.f32 %f78, %f76, %f77;
fma.rn.ftz.f32 %f175, %f52, %f78, %f175;
add.s32 %r401, %r401, 64;
add.s32 %r400, %r400, 64;
add.s64 %rd62, %rd62, 1408;
add.s32 %r402, %r402, 64;
setp.lt.s32 %p4, %r402, %r40;
@%p4 bra $L__BB65_2;

$L__BB65_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB65_5;

mov.u32 %r287, _ZZ13mul_mat_vec_qIL9ggml_type6ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r288, %r1, 768, %r287;
shl.b32 %r289, %r2, 2;
add.s32 %r290, %r288, %r289;
st.shared.f32 [%r290+-768], %f180;
st.shared.f32 [%r290+-640], %f179;
st.shared.f32 [%r290+-512], %f178;
st.shared.f32 [%r290+-384], %f177;
st.shared.f32 [%r290+-256], %f176;
st.shared.f32 [%r290+-128], %f175;

$L__BB65_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB65_18;

mad.lo.s32 %r291, %r5, %r29, %r4;
mad.lo.s32 %r292, %r6, %r35, %r291;
cvt.u64.u32 %rd12, %r292;
shl.b32 %r293, %r2, 2;
mov.u32 %r294, _ZZ13mul_mat_vec_qIL9ggml_type6ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r18, %r294, %r293;
mov.u32 %r295, 2;
ld.shared.f32 %f79, [%r18];
add.ftz.f32 %f80, %f79, %f180;
ld.shared.f32 %f81, [%r18+768];
add.ftz.f32 %f82, %f81, %f80;
ld.shared.f32 %f83, [%r18+1536];
add.ftz.f32 %f84, %f83, %f82;
mov.b32 %r296, %f84;
mov.u32 %r297, 31;
mov.u32 %r298, 16;
mov.u32 %r299, -1;
shfl.sync.bfly.b32 %r300|%p7, %r296, %r298, %r297, %r299;
mov.b32 %f85, %r300;
add.ftz.f32 %f86, %f84, %f85;
mov.b32 %r301, %f86;
mov.u32 %r302, 8;
shfl.sync.bfly.b32 %r303|%p8, %r301, %r302, %r297, %r299;
mov.b32 %f87, %r303;
add.ftz.f32 %f88, %f86, %f87;
mov.b32 %r304, %f88;
mov.u32 %r305, 4;
shfl.sync.bfly.b32 %r306|%p9, %r304, %r305, %r297, %r299;
mov.b32 %f89, %r306;
add.ftz.f32 %f90, %f88, %f89;
mov.b32 %r307, %f90;
shfl.sync.bfly.b32 %r308|%p10, %r307, %r295, %r297, %r299;
mov.b32 %f91, %r308;
add.ftz.f32 %f92, %f90, %f91;
mov.b32 %r309, %f92;
mov.u32 %r310, 1;
shfl.sync.bfly.b32 %r311|%p11, %r309, %r310, %r297, %r299;
mov.b32 %f93, %r311;
add.ftz.f32 %f19, %f92, %f93;
setp.lt.u32 %p12, %r4, %r23;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
mul.wide.u32 %rd50, %r292, 4;
add.s64 %rd13, %rd1, %rd50;
not.pred %p14, %p1;
@%p14 bra $L__BB65_8;

st.global.f32 [%rd13], %f19;

$L__BB65_8:
ld.shared.f32 %f94, [%r18+128];
add.ftz.f32 %f95, %f94, %f179;
ld.shared.f32 %f96, [%r18+896];
add.ftz.f32 %f97, %f96, %f95;
ld.shared.f32 %f98, [%r18+1664];
add.ftz.f32 %f99, %f98, %f97;
mov.b32 %r312, %f99;
shfl.sync.bfly.b32 %r316|%p15, %r312, %r298, %r297, %r299;
mov.b32 %f100, %r316;
add.ftz.f32 %f101, %f99, %f100;
mov.b32 %r317, %f101;
shfl.sync.bfly.b32 %r319|%p16, %r317, %r302, %r297, %r299;
mov.b32 %f102, %r319;
add.ftz.f32 %f103, %f101, %f102;
mov.b32 %r320, %f103;
shfl.sync.bfly.b32 %r322|%p17, %r320, %r305, %r297, %r299;
mov.b32 %f104, %r322;
add.ftz.f32 %f105, %f103, %f104;
mov.b32 %r323, %f105;
shfl.sync.bfly.b32 %r325|%p18, %r323, %r295, %r297, %r299;
mov.b32 %f106, %r325;
add.ftz.f32 %f107, %f105, %f106;
mov.b32 %r326, %f107;
shfl.sync.bfly.b32 %r328|%p19, %r326, %r310, %r297, %r299;
mov.b32 %f108, %r328;
add.ftz.f32 %f20, %f107, %f108;
add.s32 %r329, %r4, 1;
setp.lt.u32 %p20, %r329, %r23;
setp.eq.s32 %p21, %r2, 1;
and.pred %p2, %p21, %p20;
not.pred %p22, %p2;
@%p22 bra $L__BB65_10;

st.global.f32 [%rd13+4], %f20;

$L__BB65_10:
ld.shared.f32 %f109, [%r18+256];
add.ftz.f32 %f110, %f109, %f178;
ld.shared.f32 %f111, [%r18+1024];
add.ftz.f32 %f112, %f111, %f110;
ld.shared.f32 %f113, [%r18+1792];
add.ftz.f32 %f114, %f113, %f112;
mov.b32 %r330, %f114;
mov.u32 %r331, 31;
mov.u32 %r332, 16;
mov.u32 %r333, -1;
shfl.sync.bfly.b32 %r334|%p23, %r330, %r332, %r331, %r333;
mov.b32 %f115, %r334;
add.ftz.f32 %f116, %f114, %f115;
mov.b32 %r335, %f116;
mov.u32 %r336, 8;
shfl.sync.bfly.b32 %r337|%p24, %r335, %r336, %r331, %r333;
mov.b32 %f117, %r337;
add.ftz.f32 %f118, %f116, %f117;
mov.b32 %r338, %f118;
mov.u32 %r339, 4;
shfl.sync.bfly.b32 %r340|%p25, %r338, %r339, %r331, %r333;
mov.b32 %f119, %r340;
add.ftz.f32 %f120, %f118, %f119;
mov.b32 %r341, %f120;
mov.u32 %r342, 2;
shfl.sync.bfly.b32 %r343|%p26, %r341, %r342, %r331, %r333;
mov.b32 %f121, %r343;
add.ftz.f32 %f122, %f120, %f121;
mov.b32 %r344, %f122;
mov.u32 %r345, 1;
shfl.sync.bfly.b32 %r346|%p27, %r344, %r345, %r331, %r333;
mov.b32 %f123, %r346;
add.ftz.f32 %f21, %f122, %f123;
@%p14 bra $L__BB65_12;

cvt.u64.u32 %rd51, %r23;
add.s64 %rd52, %rd51, %rd12;
shl.b64 %rd53, %rd52, 2;
add.s64 %rd54, %rd1, %rd53;
st.global.f32 [%rd54], %f21;

$L__BB65_12:
ld.shared.f32 %f124, [%r18+384];
add.ftz.f32 %f125, %f124, %f177;
ld.shared.f32 %f126, [%r18+1152];
add.ftz.f32 %f127, %f126, %f125;
ld.shared.f32 %f128, [%r18+1920];
add.ftz.f32 %f129, %f128, %f127;
mov.b32 %r347, %f129;
shfl.sync.bfly.b32 %r351|%p29, %r347, %r332, %r331, %r333;
mov.b32 %f130, %r351;
add.ftz.f32 %f131, %f129, %f130;
mov.b32 %r352, %f131;
shfl.sync.bfly.b32 %r354|%p30, %r352, %r336, %r331, %r333;
mov.b32 %f132, %r354;
add.ftz.f32 %f133, %f131, %f132;
mov.b32 %r355, %f133;
shfl.sync.bfly.b32 %r357|%p31, %r355, %r339, %r331, %r333;
mov.b32 %f134, %r357;
add.ftz.f32 %f135, %f133, %f134;
mov.b32 %r358, %f135;
shfl.sync.bfly.b32 %r360|%p32, %r358, %r342, %r331, %r333;
mov.b32 %f136, %r360;
add.ftz.f32 %f137, %f135, %f136;
mov.b32 %r361, %f137;
shfl.sync.bfly.b32 %r363|%p33, %r361, %r345, %r331, %r333;
mov.b32 %f138, %r363;
add.ftz.f32 %f22, %f137, %f138;
@%p22 bra $L__BB65_14;

add.s32 %r364, %r23, 1;
cvt.u64.u32 %rd55, %r364;
add.s64 %rd56, %rd55, %rd12;
shl.b64 %rd57, %rd56, 2;
add.s64 %rd58, %rd1, %rd57;
st.global.f32 [%rd58], %f22;

$L__BB65_14:
shl.b32 %r365, %r23, 1;
mov.u32 %r366, 1;
ld.shared.f32 %f139, [%r18+512];
add.ftz.f32 %f140, %f139, %f176;
ld.shared.f32 %f141, [%r18+1280];
add.ftz.f32 %f142, %f141, %f140;
ld.shared.f32 %f143, [%r18+2048];
add.ftz.f32 %f144, %f143, %f142;
mov.b32 %r367, %f144;
mov.u32 %r368, 31;
mov.u32 %r369, 16;
mov.u32 %r370, -1;
shfl.sync.bfly.b32 %r371|%p35, %r367, %r369, %r368, %r370;
mov.b32 %f145, %r371;
add.ftz.f32 %f146, %f144, %f145;
mov.b32 %r372, %f146;
mov.u32 %r373, 8;
shfl.sync.bfly.b32 %r374|%p36, %r372, %r373, %r368, %r370;
mov.b32 %f147, %r374;
add.ftz.f32 %f148, %f146, %f147;
mov.b32 %r375, %f148;
mov.u32 %r376, 4;
shfl.sync.bfly.b32 %r377|%p37, %r375, %r376, %r368, %r370;
mov.b32 %f149, %r377;
add.ftz.f32 %f150, %f148, %f149;
mov.b32 %r378, %f150;
mov.u32 %r379, 2;
shfl.sync.bfly.b32 %r380|%p38, %r378, %r379, %r368, %r370;
mov.b32 %f151, %r380;
add.ftz.f32 %f152, %f150, %f151;
mov.b32 %r381, %f152;
shfl.sync.bfly.b32 %r382|%p39, %r381, %r366, %r368, %r370;
mov.b32 %f153, %r382;
add.ftz.f32 %f23, %f152, %f153;
cvt.u64.u32 %rd59, %r365;
add.s64 %rd60, %rd59, %rd12;
shl.b64 %rd61, %rd60, 2;
add.s64 %rd14, %rd1, %rd61;
@%p14 bra $L__BB65_16;

st.global.f32 [%rd14], %f23;

$L__BB65_16:
ld.shared.f32 %f154, [%r18+640];
add.ftz.f32 %f155, %f154, %f175;
ld.shared.f32 %f156, [%r18+1408];
add.ftz.f32 %f157, %f156, %f155;
ld.shared.f32 %f158, [%r18+2176];
add.ftz.f32 %f159, %f158, %f157;
mov.b32 %r383, %f159;
shfl.sync.bfly.b32 %r387|%p41, %r383, %r369, %r368, %r370;
mov.b32 %f160, %r387;
add.ftz.f32 %f161, %f159, %f160;
mov.b32 %r388, %f161;
shfl.sync.bfly.b32 %r390|%p42, %r388, %r373, %r368, %r370;
mov.b32 %f162, %r390;
add.ftz.f32 %f163, %f161, %f162;
mov.b32 %r391, %f163;
shfl.sync.bfly.b32 %r393|%p43, %r391, %r376, %r368, %r370;
mov.b32 %f164, %r393;
add.ftz.f32 %f165, %f163, %f164;
mov.b32 %r394, %f165;
shfl.sync.bfly.b32 %r396|%p44, %r394, %r379, %r368, %r370;
mov.b32 %f166, %r396;
add.ftz.f32 %f167, %f165, %f166;
mov.b32 %r397, %f167;
shfl.sync.bfly.b32 %r399|%p45, %r397, %r366, %r368, %r370;
mov.b32 %f168, %r399;
add.ftz.f32 %f24, %f167, %f168;
@%p22 bra $L__BB65_18;

st.global.f32 [%rd14+4], %f24;

$L__BB65_18:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type6ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type6ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type6ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type6ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type6ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type6ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type6ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type6ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type6ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<59>;
.reg .b16 %rs<21>;
.reg .f32 %f<241>;
.reg .b32 %r<480>;
.reg .b64 %rd<78>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type6ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[3072];

ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type6ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type6ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type6ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type6ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type6ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type6ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type6ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type6ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type6ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type6ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type6ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type6ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type6ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type6ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type6ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type6ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type6ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
cvta.to.global.u64 %rd1, %rd17;
mov.u32 %r1, %tid.y;
shl.b32 %r40, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r40, %r2;
mov.u32 %r41, %ctaid.x;
shl.b32 %r4, %r41, 1;
mov.u32 %r5, %ctaid.y;
mov.u32 %r6, %ctaid.z;
shr.u32 %r42, %r3, 31;
add.s32 %r43, %r3, %r42;
shr.s32 %r479, %r43, 1;
shr.u32 %r44, %r24, 5;
setp.ge.s32 %p3, %r479, %r44;
mov.f32 %f233, 0f00000000;
mov.f32 %f234, %f233;
mov.f32 %f235, %f233;
mov.f32 %f236, %f233;
mov.f32 %f237, %f233;
mov.f32 %f238, %f233;
mov.f32 %f239, %f233;
mov.f32 %f240, %f233;
@%p3 bra $L__BB66_3;

mul.lo.s32 %r45, %r5, %r32;
cvt.u64.u32 %rd18, %r45;
mul.lo.s32 %r46, %r6, %r38;
cvt.u64.u32 %rd19, %r46;
add.s64 %rd2, %rd19, %rd18;
mul.hi.u32 %r47, %r6, %r34;
add.s32 %r48, %r47, %r6;
shr.u32 %r49, %r48, %r35;
mul.lo.s32 %r50, %r49, %r37;
mul.hi.u32 %r51, %r5, %r28;
add.s32 %r52, %r51, %r5;
shr.u32 %r53, %r52, %r29;
mul.lo.s32 %r54, %r53, %r31;
cvta.to.global.u64 %rd77, %rd15;
and.b32 %r57, %r43, -2;
sub.s32 %r8, %r3, %r57;
shl.b32 %r58, %r8, 1;
shl.b32 %r59, %r8, 2;
or.b32 %r60, %r59, 1;
sub.s32 %r61, %r60, %r58;
shl.b32 %r9, %r61, 2;
cvt.s64.s32 %rd4, %r61;
shl.b32 %r62, %r26, 1;
add.s32 %r478, %r479, %r62;
mad.lo.s32 %r477, %r26, 3, %r479;
add.s32 %r476, %r479, %r26;
mad.lo.s32 %r63, %r4, %r25, %r479;
add.s32 %r64, %r63, %r50;
add.s32 %r65, %r64, %r54;
mul.wide.s32 %rd5, %r65, 22;
mul.wide.s32 %rd20, %r59, 2;
add.s64 %rd21, %rd5, %rd20;
add.s64 %rd6, %rd21, 8;
add.s32 %r66, %r479, %r50;
add.s32 %r67, %r66, %r54;
add.s32 %r68, %r4, 1;
mad.lo.s32 %r69, %r25, %r68, %r67;
mul.wide.s32 %rd7, %r69, 22;
add.s64 %rd22, %rd7, %rd20;
add.s64 %rd8, %rd22, 6;
cvta.to.global.u64 %rd9, %rd16;
mov.f32 %f233, 0f00000000;
shl.b64 %rd32, %rd4, 2;

$L__BB66_2:
cvt.u64.u32 %rd23, %r479;
add.s64 %rd24, %rd2, %rd23;
mul.lo.s64 %rd25, %rd24, 36;
add.s64 %rd26, %rd9, %rd25;
add.s64 %rd27, %rd26, 4;
add.s64 %rd28, %rd77, %rd5;
ld.global.nc.u16 %rs9, [%rd28+2];
ld.global.nc.u16 %rs10, [%rd28+4];
mov.b32 %r214, {%rs9, %rs10};
add.s64 %rd29, %rd77, %rd6;
ld.global.nc.u16 %rs11, [%rd29+-2];
ld.global.nc.u16 %rs12, [%rd29];
mov.b32 %r215, {%rs11, %rs12};
shl.b32 %r216, %r8, 3;
shr.s32 %r217, %r214, %r216;
mul.wide.s32 %rd30, %r58, 4;
add.s64 %rd31, %rd27, %rd30;
ld.global.nc.u32 %r72, [%rd31];
ld.global.nc.u32 %r76, [%rd31+16];
ld.global.nc.u16 %rs13, [%rd29+2];
ld.global.nc.u16 %rs14, [%rd29+4];
mov.b32 %r219, {%rs13, %rs14};
shr.s32 %r220, %r214, %r9;
add.s64 %rd33, %rd27, %rd32;
ld.global.nc.u32 %r80, [%rd33];
ld.global.nc.u32 %r84, [%rd31+20];
ld.global.nc.u16 %rs1, [%rd28];

{ cvt.f32.f16 %f49, %rs1;}


and.b32 %r221, %r215, 252645135;
shl.b32 %r222, %r217, 4;
and.b32 %r223, %r222, 16;
or.b32 %r224, %r223, %r221;
shl.b32 %r225, %r217, 11;
and.b32 %r226, %r225, 4096;
or.b32 %r227, %r224, %r226;
shl.b32 %r228, %r217, 18;
and.b32 %r229, %r228, 1048576;
or.b32 %r230, %r227, %r229;
shl.b32 %r231, %r217, 25;
and.b32 %r232, %r231, 268435456;
or.b32 %r179, %r230, %r232;
mov.u32 %r199, 0;

dp4a.s32.s32 %r70, %r179, %r72, %r199;

shr.u32 %r233, %r215, 4;
and.b32 %r234, %r233, 252645135;
shr.u32 %r235, %r217, 12;
and.b32 %r236, %r235, 16;
or.b32 %r237, %r236, %r234;
shr.u32 %r238, %r217, 5;
and.b32 %r239, %r238, 4096;
or.b32 %r240, %r237, %r239;
shl.b32 %r241, %r217, 2;
and.b32 %r242, %r241, 1048576;
or.b32 %r243, %r240, %r242;
shl.b32 %r244, %r217, 9;
and.b32 %r245, %r244, 268435456;
or.b32 %r183, %r243, %r245;

dp4a.s32.s32 %r74, %r183, %r76, %r70;

and.b32 %r246, %r219, 252645135;
shl.b32 %r247, %r220, 4;
and.b32 %r248, %r247, 16;
or.b32 %r249, %r248, %r246;
shl.b32 %r250, %r220, 11;
and.b32 %r251, %r250, 4096;
or.b32 %r252, %r249, %r251;
shl.b32 %r253, %r220, 18;
and.b32 %r254, %r253, 1048576;
or.b32 %r255, %r252, %r254;
shl.b32 %r256, %r220, 25;
and.b32 %r257, %r256, 268435456;
or.b32 %r187, %r255, %r257;

dp4a.s32.s32 %r78, %r187, %r80, %r74;

shr.u32 %r258, %r219, 4;
and.b32 %r259, %r258, 252645135;
shr.u32 %r260, %r220, 12;
and.b32 %r261, %r260, 16;
or.b32 %r262, %r261, %r259;
shr.u32 %r263, %r220, 5;
and.b32 %r264, %r263, 4096;
or.b32 %r265, %r262, %r264;
shl.b32 %r266, %r220, 2;
and.b32 %r267, %r266, 1048576;
or.b32 %r268, %r265, %r267;
shl.b32 %r269, %r220, 9;
and.b32 %r270, %r269, 268435456;
or.b32 %r191, %r268, %r270;

dp4a.s32.s32 %r82, %r191, %r84, %r78;

ld.global.nc.u32 %r86, [%rd26];

{.reg .f16 low,high;
mov.b32 {low,high},%r86;
cvt.f32.f16 %f50, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r86;
cvt.f32.f16 %f51, high;}


cvt.rn.f32.s32 %f73, %r82;
mul.ftz.f32 %f74, %f50, %f73;
mul.ftz.f32 %f75, %f51, 0f41000000;
sub.ftz.f32 %f76, %f74, %f75;
fma.rn.ftz.f32 %f240, %f49, %f76, %f240;
add.s64 %rd34, %rd77, %rd7;
ld.global.nc.u16 %rs15, [%rd34+2];
ld.global.nc.u16 %rs16, [%rd34+4];
mov.b32 %r271, {%rs15, %rs16};
add.s64 %rd35, %rd77, %rd8;
ld.global.nc.u16 %rs17, [%rd35];
ld.global.nc.u16 %rs18, [%rd35+2];
mov.b32 %r272, {%rs17, %rs18};
shr.s32 %r273, %r271, %r216;
ld.global.nc.u16 %rs19, [%rd35+4];
ld.global.nc.u16 %rs20, [%rd35+6];
mov.b32 %r274, {%rs19, %rs20};
shr.s32 %r275, %r271, %r9;
ld.global.nc.u16 %rs2, [%rd34];

{ cvt.f32.f16 %f52, %rs2;}


and.b32 %r276, %r272, 252645135;
shl.b32 %r277, %r273, 4;
and.b32 %r278, %r277, 16;
or.b32 %r279, %r278, %r276;
shl.b32 %r280, %r273, 11;
and.b32 %r281, %r280, 4096;
or.b32 %r282, %r279, %r281;
shl.b32 %r283, %r273, 18;
and.b32 %r284, %r283, 1048576;
or.b32 %r285, %r282, %r284;
shl.b32 %r286, %r273, 25;
and.b32 %r287, %r286, 268435456;
or.b32 %r197, %r285, %r287;

dp4a.s32.s32 %r88, %r197, %r72, %r199;

shr.u32 %r288, %r272, 4;
and.b32 %r289, %r288, 252645135;
shr.u32 %r290, %r273, 12;
and.b32 %r291, %r290, 16;
or.b32 %r292, %r291, %r289;
shr.u32 %r293, %r273, 5;
and.b32 %r294, %r293, 4096;
or.b32 %r295, %r292, %r294;
shl.b32 %r296, %r273, 2;
and.b32 %r297, %r296, 1048576;
or.b32 %r298, %r295, %r297;
shl.b32 %r299, %r273, 9;
and.b32 %r300, %r299, 268435456;
or.b32 %r201, %r298, %r300;

dp4a.s32.s32 %r92, %r201, %r76, %r88;

and.b32 %r301, %r274, 252645135;
shl.b32 %r302, %r275, 4;
and.b32 %r303, %r302, 16;
or.b32 %r304, %r303, %r301;
shl.b32 %r305, %r275, 11;
and.b32 %r306, %r305, 4096;
or.b32 %r307, %r304, %r306;
shl.b32 %r308, %r275, 18;
and.b32 %r309, %r308, 1048576;
or.b32 %r310, %r307, %r309;
shl.b32 %r311, %r275, 25;
and.b32 %r312, %r311, 268435456;
or.b32 %r205, %r310, %r312;

dp4a.s32.s32 %r96, %r205, %r80, %r92;

shr.u32 %r313, %r274, 4;
and.b32 %r314, %r313, 252645135;
shr.u32 %r315, %r275, 12;
and.b32 %r316, %r315, 16;
or.b32 %r317, %r316, %r314;
shr.u32 %r318, %r275, 5;
and.b32 %r319, %r318, 4096;
or.b32 %r320, %r317, %r319;
shl.b32 %r321, %r275, 2;
and.b32 %r322, %r321, 1048576;
or.b32 %r323, %r320, %r322;
shl.b32 %r324, %r275, 9;
and.b32 %r325, %r324, 268435456;
or.b32 %r209, %r323, %r325;

dp4a.s32.s32 %r100, %r209, %r84, %r96;


{.reg .f16 low,high;
mov.b32 {low,high},%r86;
cvt.f32.f16 %f53, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r86;
cvt.f32.f16 %f54, high;}


cvt.rn.f32.s32 %f77, %r100;
mul.ftz.f32 %f78, %f53, %f77;
mul.ftz.f32 %f79, %f54, 0f41000000;
sub.ftz.f32 %f80, %f78, %f79;
fma.rn.ftz.f32 %f239, %f52, %f80, %f239;
cvt.u64.u32 %rd36, %r476;
add.s64 %rd37, %rd2, %rd36;
mul.lo.s64 %rd38, %rd37, 36;
add.s64 %rd39, %rd9, %rd38;
add.s64 %rd40, %rd39, 4;
add.s64 %rd41, %rd40, %rd30;
ld.global.nc.u32 %r108, [%rd41];
ld.global.nc.u32 %r112, [%rd41+16];
add.s64 %rd42, %rd40, %rd32;
ld.global.nc.u32 %r116, [%rd42];
ld.global.nc.u32 %r120, [%rd41+20];

{ cvt.f32.f16 %f55, %rs1;}



dp4a.s32.s32 %r106, %r179, %r108, %r199;


dp4a.s32.s32 %r110, %r183, %r112, %r106;


dp4a.s32.s32 %r114, %r187, %r116, %r110;


dp4a.s32.s32 %r118, %r191, %r120, %r114;

ld.global.nc.u32 %r122, [%rd39];

{.reg .f16 low,high;
mov.b32 {low,high},%r122;
cvt.f32.f16 %f56, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r122;
cvt.f32.f16 %f57, high;}


cvt.rn.f32.s32 %f81, %r118;
mul.ftz.f32 %f82, %f56, %f81;
mul.ftz.f32 %f83, %f57, 0f41000000;
sub.ftz.f32 %f84, %f82, %f83;
fma.rn.ftz.f32 %f238, %f55, %f84, %f238;

{ cvt.f32.f16 %f58, %rs2;}



dp4a.s32.s32 %r124, %r197, %r108, %r199;


dp4a.s32.s32 %r128, %r201, %r112, %r124;


dp4a.s32.s32 %r132, %r205, %r116, %r128;


dp4a.s32.s32 %r136, %r209, %r120, %r132;


{.reg .f16 low,high;
mov.b32 {low,high},%r122;
cvt.f32.f16 %f59, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r122;
cvt.f32.f16 %f60, high;}


cvt.rn.f32.s32 %f85, %r136;
mul.ftz.f32 %f86, %f59, %f85;
mul.ftz.f32 %f87, %f60, 0f41000000;
sub.ftz.f32 %f88, %f86, %f87;
fma.rn.ftz.f32 %f237, %f58, %f88, %f237;
cvt.u64.u32 %rd43, %r478;
add.s64 %rd44, %rd2, %rd43;
mul.lo.s64 %rd45, %rd44, 36;
add.s64 %rd46, %rd9, %rd45;
add.s64 %rd47, %rd46, 4;
add.s64 %rd48, %rd47, %rd30;
ld.global.nc.u32 %r144, [%rd48];
ld.global.nc.u32 %r148, [%rd48+16];
add.s64 %rd49, %rd47, %rd32;
ld.global.nc.u32 %r152, [%rd49];
ld.global.nc.u32 %r156, [%rd48+20];

{ cvt.f32.f16 %f61, %rs1;}



dp4a.s32.s32 %r142, %r179, %r144, %r199;


dp4a.s32.s32 %r146, %r183, %r148, %r142;


dp4a.s32.s32 %r150, %r187, %r152, %r146;


dp4a.s32.s32 %r154, %r191, %r156, %r150;

ld.global.nc.u32 %r158, [%rd46];

{.reg .f16 low,high;
mov.b32 {low,high},%r158;
cvt.f32.f16 %f62, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r158;
cvt.f32.f16 %f63, high;}


cvt.rn.f32.s32 %f89, %r154;
mul.ftz.f32 %f90, %f62, %f89;
mul.ftz.f32 %f91, %f63, 0f41000000;
sub.ftz.f32 %f92, %f90, %f91;
fma.rn.ftz.f32 %f236, %f61, %f92, %f236;

{ cvt.f32.f16 %f64, %rs2;}



dp4a.s32.s32 %r160, %r197, %r144, %r199;


dp4a.s32.s32 %r164, %r201, %r148, %r160;


dp4a.s32.s32 %r168, %r205, %r152, %r164;


dp4a.s32.s32 %r172, %r209, %r156, %r168;


{.reg .f16 low,high;
mov.b32 {low,high},%r158;
cvt.f32.f16 %f65, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r158;
cvt.f32.f16 %f66, high;}


cvt.rn.f32.s32 %f93, %r172;
mul.ftz.f32 %f94, %f65, %f93;
mul.ftz.f32 %f95, %f66, 0f41000000;
sub.ftz.f32 %f96, %f94, %f95;
fma.rn.ftz.f32 %f235, %f64, %f96, %f235;
cvt.u64.u32 %rd50, %r477;
add.s64 %rd51, %rd2, %rd50;
mul.lo.s64 %rd52, %rd51, 36;
add.s64 %rd53, %rd9, %rd52;
add.s64 %rd54, %rd53, 4;
add.s64 %rd55, %rd54, %rd30;
ld.global.nc.u32 %r180, [%rd55];
ld.global.nc.u32 %r184, [%rd55+16];
add.s64 %rd56, %rd54, %rd32;
ld.global.nc.u32 %r188, [%rd56];
ld.global.nc.u32 %r192, [%rd55+20];

{ cvt.f32.f16 %f67, %rs1;}



dp4a.s32.s32 %r178, %r179, %r180, %r199;


dp4a.s32.s32 %r182, %r183, %r184, %r178;


dp4a.s32.s32 %r186, %r187, %r188, %r182;


dp4a.s32.s32 %r190, %r191, %r192, %r186;

ld.global.nc.u32 %r194, [%rd53];

{.reg .f16 low,high;
mov.b32 {low,high},%r194;
cvt.f32.f16 %f68, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r194;
cvt.f32.f16 %f69, high;}


cvt.rn.f32.s32 %f97, %r190;
mul.ftz.f32 %f98, %f68, %f97;
mul.ftz.f32 %f99, %f69, 0f41000000;
sub.ftz.f32 %f100, %f98, %f99;
fma.rn.ftz.f32 %f234, %f67, %f100, %f234;

{ cvt.f32.f16 %f70, %rs2;}



dp4a.s32.s32 %r196, %r197, %r180, %r199;


dp4a.s32.s32 %r200, %r201, %r184, %r196;


dp4a.s32.s32 %r204, %r205, %r188, %r200;


dp4a.s32.s32 %r208, %r209, %r192, %r204;


{.reg .f16 low,high;
mov.b32 {low,high},%r194;
cvt.f32.f16 %f71, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r194;
cvt.f32.f16 %f72, high;}


cvt.rn.f32.s32 %f101, %r208;
mul.ftz.f32 %f102, %f71, %f101;
mul.ftz.f32 %f103, %f72, 0f41000000;
sub.ftz.f32 %f104, %f102, %f103;
fma.rn.ftz.f32 %f233, %f70, %f104, %f233;
add.s32 %r478, %r478, 64;
add.s32 %r477, %r477, 64;
add.s32 %r476, %r476, 64;
add.s64 %rd77, %rd77, 1408;
add.s32 %r479, %r479, 64;
setp.lt.s32 %p4, %r479, %r44;
@%p4 bra $L__BB66_2;

$L__BB66_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB66_5;

shl.b32 %r327, %r1, 10;
mov.u32 %r328, _ZZ13mul_mat_vec_qIL9ggml_type6ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r329, %r328, %r327;
shl.b32 %r330, %r2, 2;
add.s32 %r331, %r329, %r330;
st.shared.f32 [%r331+-1024], %f240;
st.shared.f32 [%r331+-896], %f239;
st.shared.f32 [%r331+-768], %f238;
st.shared.f32 [%r331+-640], %f237;
st.shared.f32 [%r331+-512], %f236;
st.shared.f32 [%r331+-384], %f235;
st.shared.f32 [%r331+-256], %f234;
st.shared.f32 [%r331+-128], %f233;

$L__BB66_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB66_22;

mad.lo.s32 %r332, %r5, %r33, %r4;
mad.lo.s32 %r333, %r6, %r39, %r332;
cvt.u64.u32 %rd12, %r333;
shl.b32 %r334, %r2, 2;
mov.u32 %r335, _ZZ13mul_mat_vec_qIL9ggml_type6ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r21, %r335, %r334;
mov.u32 %r336, 2;
ld.shared.f32 %f105, [%r21];
add.ftz.f32 %f106, %f105, %f240;
ld.shared.f32 %f107, [%r21+1024];
add.ftz.f32 %f108, %f107, %f106;
ld.shared.f32 %f109, [%r21+2048];
add.ftz.f32 %f110, %f109, %f108;
mov.b32 %r337, %f110;
mov.u32 %r338, 31;
mov.u32 %r339, 16;
mov.u32 %r340, -1;
shfl.sync.bfly.b32 %r341|%p7, %r337, %r339, %r338, %r340;
mov.b32 %f111, %r341;
add.ftz.f32 %f112, %f110, %f111;
mov.b32 %r342, %f112;
mov.u32 %r343, 8;
shfl.sync.bfly.b32 %r344|%p8, %r342, %r343, %r338, %r340;
mov.b32 %f113, %r344;
add.ftz.f32 %f114, %f112, %f113;
mov.b32 %r345, %f114;
mov.u32 %r346, 4;
shfl.sync.bfly.b32 %r347|%p9, %r345, %r346, %r338, %r340;
mov.b32 %f115, %r347;
add.ftz.f32 %f116, %f114, %f115;
mov.b32 %r348, %f116;
shfl.sync.bfly.b32 %r349|%p10, %r348, %r336, %r338, %r340;
mov.b32 %f117, %r349;
add.ftz.f32 %f118, %f116, %f117;
mov.b32 %r350, %f118;
mov.u32 %r351, 1;
shfl.sync.bfly.b32 %r352|%p11, %r350, %r351, %r338, %r340;
mov.b32 %f119, %r352;
add.ftz.f32 %f25, %f118, %f119;
setp.lt.u32 %p12, %r4, %r27;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
mul.wide.u32 %rd57, %r333, 4;
add.s64 %rd13, %rd1, %rd57;
not.pred %p14, %p1;
@%p14 bra $L__BB66_8;

st.global.f32 [%rd13], %f25;

$L__BB66_8:
ld.shared.f32 %f120, [%r21+128];
add.ftz.f32 %f121, %f120, %f239;
ld.shared.f32 %f122, [%r21+1152];
add.ftz.f32 %f123, %f122, %f121;
ld.shared.f32 %f124, [%r21+2176];
add.ftz.f32 %f125, %f124, %f123;
mov.b32 %r353, %f125;
shfl.sync.bfly.b32 %r357|%p15, %r353, %r339, %r338, %r340;
mov.b32 %f126, %r357;
add.ftz.f32 %f127, %f125, %f126;
mov.b32 %r358, %f127;
shfl.sync.bfly.b32 %r360|%p16, %r358, %r343, %r338, %r340;
mov.b32 %f128, %r360;
add.ftz.f32 %f129, %f127, %f128;
mov.b32 %r361, %f129;
shfl.sync.bfly.b32 %r363|%p17, %r361, %r346, %r338, %r340;
mov.b32 %f130, %r363;
add.ftz.f32 %f131, %f129, %f130;
mov.b32 %r364, %f131;
shfl.sync.bfly.b32 %r366|%p18, %r364, %r336, %r338, %r340;
mov.b32 %f132, %r366;
add.ftz.f32 %f133, %f131, %f132;
mov.b32 %r367, %f133;
shfl.sync.bfly.b32 %r369|%p19, %r367, %r351, %r338, %r340;
mov.b32 %f134, %r369;
add.ftz.f32 %f26, %f133, %f134;
add.s32 %r370, %r4, 1;
setp.lt.u32 %p20, %r370, %r27;
setp.eq.s32 %p21, %r2, 1;
and.pred %p2, %p21, %p20;
not.pred %p22, %p2;
@%p22 bra $L__BB66_10;

st.global.f32 [%rd13+4], %f26;

$L__BB66_10:
ld.shared.f32 %f135, [%r21+256];
add.ftz.f32 %f136, %f135, %f238;
ld.shared.f32 %f137, [%r21+1280];
add.ftz.f32 %f138, %f137, %f136;
ld.shared.f32 %f139, [%r21+2304];
add.ftz.f32 %f140, %f139, %f138;
mov.b32 %r371, %f140;
mov.u32 %r372, 31;
mov.u32 %r373, 16;
mov.u32 %r374, -1;
shfl.sync.bfly.b32 %r375|%p23, %r371, %r373, %r372, %r374;
mov.b32 %f141, %r375;
add.ftz.f32 %f142, %f140, %f141;
mov.b32 %r376, %f142;
mov.u32 %r377, 8;
shfl.sync.bfly.b32 %r378|%p24, %r376, %r377, %r372, %r374;
mov.b32 %f143, %r378;
add.ftz.f32 %f144, %f142, %f143;
mov.b32 %r379, %f144;
mov.u32 %r380, 4;
shfl.sync.bfly.b32 %r381|%p25, %r379, %r380, %r372, %r374;
mov.b32 %f145, %r381;
add.ftz.f32 %f146, %f144, %f145;
mov.b32 %r382, %f146;
mov.u32 %r383, 2;
shfl.sync.bfly.b32 %r384|%p26, %r382, %r383, %r372, %r374;
mov.b32 %f147, %r384;
add.ftz.f32 %f148, %f146, %f147;
mov.b32 %r385, %f148;
mov.u32 %r386, 1;
shfl.sync.bfly.b32 %r387|%p27, %r385, %r386, %r372, %r374;
mov.b32 %f149, %r387;
add.ftz.f32 %f27, %f148, %f149;
@%p14 bra $L__BB66_12;

cvt.u64.u32 %rd58, %r27;
add.s64 %rd59, %rd58, %rd12;
shl.b64 %rd60, %rd59, 2;
add.s64 %rd61, %rd1, %rd60;
st.global.f32 [%rd61], %f27;

$L__BB66_12:
ld.shared.f32 %f150, [%r21+384];
add.ftz.f32 %f151, %f150, %f237;
ld.shared.f32 %f152, [%r21+1408];
add.ftz.f32 %f153, %f152, %f151;
ld.shared.f32 %f154, [%r21+2432];
add.ftz.f32 %f155, %f154, %f153;
mov.b32 %r388, %f155;
shfl.sync.bfly.b32 %r392|%p29, %r388, %r373, %r372, %r374;
mov.b32 %f156, %r392;
add.ftz.f32 %f157, %f155, %f156;
mov.b32 %r393, %f157;
shfl.sync.bfly.b32 %r395|%p30, %r393, %r377, %r372, %r374;
mov.b32 %f158, %r395;
add.ftz.f32 %f159, %f157, %f158;
mov.b32 %r396, %f159;
shfl.sync.bfly.b32 %r398|%p31, %r396, %r380, %r372, %r374;
mov.b32 %f160, %r398;
add.ftz.f32 %f161, %f159, %f160;
mov.b32 %r399, %f161;
shfl.sync.bfly.b32 %r401|%p32, %r399, %r383, %r372, %r374;
mov.b32 %f162, %r401;
add.ftz.f32 %f163, %f161, %f162;
mov.b32 %r402, %f163;
shfl.sync.bfly.b32 %r404|%p33, %r402, %r386, %r372, %r374;
mov.b32 %f164, %r404;
add.ftz.f32 %f28, %f163, %f164;
@%p22 bra $L__BB66_14;

add.s32 %r405, %r27, 1;
cvt.u64.u32 %rd62, %r405;
add.s64 %rd63, %rd62, %rd12;
shl.b64 %rd64, %rd63, 2;
add.s64 %rd65, %rd1, %rd64;
st.global.f32 [%rd65], %f28;

$L__BB66_14:
shl.b32 %r406, %r27, 1;
mov.u32 %r407, 1;
ld.shared.f32 %f165, [%r21+512];
add.ftz.f32 %f166, %f165, %f236;
ld.shared.f32 %f167, [%r21+1536];
add.ftz.f32 %f168, %f167, %f166;
ld.shared.f32 %f169, [%r21+2560];
add.ftz.f32 %f170, %f169, %f168;
mov.b32 %r408, %f170;
mov.u32 %r409, 31;
mov.u32 %r410, 16;
mov.u32 %r411, -1;
shfl.sync.bfly.b32 %r412|%p35, %r408, %r410, %r409, %r411;
mov.b32 %f171, %r412;
add.ftz.f32 %f172, %f170, %f171;
mov.b32 %r413, %f172;
mov.u32 %r414, 8;
shfl.sync.bfly.b32 %r415|%p36, %r413, %r414, %r409, %r411;
mov.b32 %f173, %r415;
add.ftz.f32 %f174, %f172, %f173;
mov.b32 %r416, %f174;
mov.u32 %r417, 4;
shfl.sync.bfly.b32 %r418|%p37, %r416, %r417, %r409, %r411;
mov.b32 %f175, %r418;
add.ftz.f32 %f176, %f174, %f175;
mov.b32 %r419, %f176;
mov.u32 %r420, 2;
shfl.sync.bfly.b32 %r421|%p38, %r419, %r420, %r409, %r411;
mov.b32 %f177, %r421;
add.ftz.f32 %f178, %f176, %f177;
mov.b32 %r422, %f178;
shfl.sync.bfly.b32 %r423|%p39, %r422, %r407, %r409, %r411;
mov.b32 %f179, %r423;
add.ftz.f32 %f29, %f178, %f179;
cvt.u64.u32 %rd66, %r406;
add.s64 %rd67, %rd66, %rd12;
shl.b64 %rd68, %rd67, 2;
add.s64 %rd14, %rd1, %rd68;
@%p14 bra $L__BB66_16;

st.global.f32 [%rd14], %f29;

$L__BB66_16:
ld.shared.f32 %f180, [%r21+640];
add.ftz.f32 %f181, %f180, %f235;
ld.shared.f32 %f182, [%r21+1664];
add.ftz.f32 %f183, %f182, %f181;
ld.shared.f32 %f184, [%r21+2688];
add.ftz.f32 %f185, %f184, %f183;
mov.b32 %r424, %f185;
shfl.sync.bfly.b32 %r428|%p41, %r424, %r410, %r409, %r411;
mov.b32 %f186, %r428;
add.ftz.f32 %f187, %f185, %f186;
mov.b32 %r429, %f187;
shfl.sync.bfly.b32 %r431|%p42, %r429, %r414, %r409, %r411;
mov.b32 %f188, %r431;
add.ftz.f32 %f189, %f187, %f188;
mov.b32 %r432, %f189;
shfl.sync.bfly.b32 %r434|%p43, %r432, %r417, %r409, %r411;
mov.b32 %f190, %r434;
add.ftz.f32 %f191, %f189, %f190;
mov.b32 %r435, %f191;
shfl.sync.bfly.b32 %r437|%p44, %r435, %r420, %r409, %r411;
mov.b32 %f192, %r437;
add.ftz.f32 %f193, %f191, %f192;
mov.b32 %r438, %f193;
shfl.sync.bfly.b32 %r440|%p45, %r438, %r407, %r409, %r411;
mov.b32 %f194, %r440;
add.ftz.f32 %f30, %f193, %f194;
@%p22 bra $L__BB66_18;

st.global.f32 [%rd14+4], %f30;

$L__BB66_18:
mul.lo.s32 %r23, %r27, 3;
ld.shared.f32 %f195, [%r21+768];
add.ftz.f32 %f196, %f195, %f234;
ld.shared.f32 %f197, [%r21+1792];
add.ftz.f32 %f198, %f197, %f196;
ld.shared.f32 %f199, [%r21+2816];
add.ftz.f32 %f200, %f199, %f198;
mov.b32 %r441, %f200;
mov.u32 %r442, 31;
mov.u32 %r443, 16;
mov.u32 %r444, -1;
shfl.sync.bfly.b32 %r445|%p47, %r441, %r443, %r442, %r444;
mov.b32 %f201, %r445;
add.ftz.f32 %f202, %f200, %f201;
mov.b32 %r446, %f202;
mov.u32 %r447, 8;
shfl.sync.bfly.b32 %r448|%p48, %r446, %r447, %r442, %r444;
mov.b32 %f203, %r448;
add.ftz.f32 %f204, %f202, %f203;
mov.b32 %r449, %f204;
mov.u32 %r450, 4;
shfl.sync.bfly.b32 %r451|%p49, %r449, %r450, %r442, %r444;
mov.b32 %f205, %r451;
add.ftz.f32 %f206, %f204, %f205;
mov.b32 %r452, %f206;
mov.u32 %r453, 2;
shfl.sync.bfly.b32 %r454|%p50, %r452, %r453, %r442, %r444;
mov.b32 %f207, %r454;
add.ftz.f32 %f208, %f206, %f207;
mov.b32 %r455, %f208;
mov.u32 %r456, 1;
shfl.sync.bfly.b32 %r457|%p51, %r455, %r456, %r442, %r444;
mov.b32 %f209, %r457;
add.ftz.f32 %f31, %f208, %f209;
@%p14 bra $L__BB66_20;

cvt.u64.u32 %rd69, %r23;
add.s64 %rd70, %rd69, %rd12;
shl.b64 %rd71, %rd70, 2;
add.s64 %rd72, %rd1, %rd71;
st.global.f32 [%rd72], %f31;

$L__BB66_20:
ld.shared.f32 %f210, [%r21+896];
add.ftz.f32 %f211, %f210, %f233;
ld.shared.f32 %f212, [%r21+1920];
add.ftz.f32 %f213, %f212, %f211;
ld.shared.f32 %f214, [%r21+2944];
add.ftz.f32 %f215, %f214, %f213;
mov.b32 %r458, %f215;
shfl.sync.bfly.b32 %r462|%p53, %r458, %r443, %r442, %r444;
mov.b32 %f216, %r462;
add.ftz.f32 %f217, %f215, %f216;
mov.b32 %r463, %f217;
shfl.sync.bfly.b32 %r465|%p54, %r463, %r447, %r442, %r444;
mov.b32 %f218, %r465;
add.ftz.f32 %f219, %f217, %f218;
mov.b32 %r466, %f219;
shfl.sync.bfly.b32 %r468|%p55, %r466, %r450, %r442, %r444;
mov.b32 %f220, %r468;
add.ftz.f32 %f221, %f219, %f220;
mov.b32 %r469, %f221;
shfl.sync.bfly.b32 %r471|%p56, %r469, %r453, %r442, %r444;
mov.b32 %f222, %r471;
add.ftz.f32 %f223, %f221, %f222;
mov.b32 %r472, %f223;
shfl.sync.bfly.b32 %r474|%p57, %r472, %r456, %r442, %r444;
mov.b32 %f224, %r474;
add.ftz.f32 %f32, %f223, %f224;
@%p22 bra $L__BB66_22;

add.s32 %r475, %r23, 1;
cvt.u64.u32 %rd73, %r475;
add.s64 %rd74, %rd73, %rd12;
shl.b64 %rd75, %rd74, 2;
add.s64 %rd76, %rd1, %rd75;
st.global.f32 [%rd76], %f32;

$L__BB66_22:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type6ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type6ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type6ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type6ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type6ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type6ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type6ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type6ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type6ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<71>;
.reg .b16 %rs<23>;
.reg .f32 %f<261>;
.reg .b32 %r<559>;
.reg .b64 %rd<90>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type6ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1280];

ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type6ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type6ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type6ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type6ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type6ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type6ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type6ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type6ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type6ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type6ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type6ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type6ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type6ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type6ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type6ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type6ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type6ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
cvta.to.global.u64 %rd1, %rd19;
mov.u32 %r1, %tid.y;
shl.b32 %r41, %r1, 5;
mov.u32 %r42, %tid.x;
add.s32 %r2, %r41, %r42;
mov.u32 %r3, %ctaid.y;
mov.u32 %r4, %ctaid.z;
shr.u32 %r43, %r2, 31;
add.s32 %r44, %r2, %r43;
shr.s32 %r558, %r44, 1;
shr.u32 %r6, %r40, 5;
setp.ge.s32 %p3, %r558, %r6;
mov.f32 %f251, 0f00000000;
mov.f32 %f252, %f251;
mov.f32 %f253, %f251;
mov.f32 %f254, %f251;
mov.f32 %f255, %f251;
mov.f32 %f256, %f251;
mov.f32 %f257, %f251;
mov.f32 %f258, %f251;
mov.f32 %f259, %f251;
mov.f32 %f260, %f251;
@%p3 bra $L__BB67_3;

mov.u32 %r45, %ctaid.x;
shl.b32 %r46, %r45, 1;
mul.lo.s32 %r47, %r3, %r32;
cvt.u64.u32 %rd20, %r47;
mul.lo.s32 %r48, %r4, %r38;
cvt.u64.u32 %rd21, %r48;
add.s64 %rd2, %rd21, %rd20;
mul.hi.u32 %r49, %r4, %r34;
add.s32 %r50, %r49, %r4;
shr.u32 %r51, %r50, %r35;
mul.lo.s32 %r52, %r51, %r37;
mul.hi.u32 %r53, %r3, %r28;
add.s32 %r54, %r53, %r3;
shr.u32 %r55, %r54, %r29;
mul.lo.s32 %r56, %r55, %r31;
cvta.to.global.u64 %rd89, %rd17;
and.b32 %r59, %r44, -2;
sub.s32 %r60, %r2, %r59;
shl.b32 %r61, %r60, 1;
shl.b32 %r62, %r60, 2;
or.b32 %r63, %r62, 1;
shl.b32 %r7, %r60, 3;
cvt.s64.s32 %rd4, %r61;
sub.s32 %r64, %r63, %r61;
shl.b32 %r8, %r64, 2;
cvt.s64.s32 %rd5, %r64;
shl.b32 %r65, %r26, 1;
add.s32 %r557, %r558, %r65;
mad.lo.s32 %r556, %r26, 3, %r558;
shl.b32 %r66, %r26, 2;
add.s32 %r555, %r558, %r66;
add.s32 %r554, %r558, %r26;
mad.lo.s32 %r67, %r46, %r25, %r558;
add.s32 %r68, %r67, %r52;
add.s32 %r69, %r68, %r56;
mul.wide.s32 %rd6, %r69, 22;
mul.wide.s32 %rd22, %r62, 2;
add.s64 %rd23, %rd6, %rd22;
add.s64 %rd7, %rd23, 8;
add.s32 %r70, %r558, %r52;
add.s32 %r71, %r70, %r56;
or.b32 %r72, %r46, 1;
mad.lo.s32 %r73, %r25, %r72, %r71;
mul.wide.s32 %rd8, %r73, 22;
add.s64 %rd24, %rd8, %rd22;
add.s64 %rd9, %rd24, 6;
cvta.to.global.u64 %rd10, %rd18;
mov.f32 %f251, 0f00000000;
shl.b64 %rd32, %rd4, 2;
shl.b64 %rd34, %rd5, 2;

$L__BB67_2:
cvt.u64.u32 %rd25, %r558;
add.s64 %rd26, %rd2, %rd25;
mul.lo.s64 %rd27, %rd26, 36;
add.s64 %rd28, %rd10, %rd27;
add.s64 %rd29, %rd28, 4;
add.s64 %rd30, %rd89, %rd6;
ld.global.nc.u16 %rs11, [%rd30+2];
ld.global.nc.u16 %rs12, [%rd30+4];
mov.b32 %r254, {%rs11, %rs12};
add.s64 %rd31, %rd89, %rd7;
ld.global.nc.u16 %rs13, [%rd31+-2];
ld.global.nc.u16 %rs14, [%rd31];
mov.b32 %r255, {%rs13, %rs14};
shr.s32 %r256, %r254, %r7;
add.s64 %rd33, %rd29, %rd32;
ld.global.nc.u32 %r76, [%rd33];
ld.global.nc.u32 %r80, [%rd33+16];
ld.global.nc.u16 %rs15, [%rd31+2];
ld.global.nc.u16 %rs16, [%rd31+4];
mov.b32 %r257, {%rs15, %rs16};
shr.s32 %r258, %r254, %r8;
add.s64 %rd35, %rd29, %rd34;
ld.global.nc.u32 %r84, [%rd35];
ld.global.nc.u32 %r88, [%rd33+20];
ld.global.nc.u16 %rs1, [%rd30];

{ cvt.f32.f16 %f61, %rs1;}


and.b32 %r259, %r255, 252645135;
shl.b32 %r260, %r256, 4;
and.b32 %r261, %r260, 16;
or.b32 %r262, %r261, %r259;
shl.b32 %r263, %r256, 11;
and.b32 %r264, %r263, 4096;
or.b32 %r265, %r262, %r264;
shl.b32 %r266, %r256, 18;
and.b32 %r267, %r266, 1048576;
or.b32 %r268, %r265, %r267;
shl.b32 %r269, %r256, 25;
and.b32 %r270, %r269, 268435456;
or.b32 %r219, %r268, %r270;
mov.u32 %r239, 0;

dp4a.s32.s32 %r74, %r219, %r76, %r239;

shr.u32 %r271, %r255, 4;
and.b32 %r272, %r271, 252645135;
shr.u32 %r273, %r256, 12;
and.b32 %r274, %r273, 16;
or.b32 %r275, %r274, %r272;
shr.u32 %r276, %r256, 5;
and.b32 %r277, %r276, 4096;
or.b32 %r278, %r275, %r277;
shl.b32 %r279, %r256, 2;
and.b32 %r280, %r279, 1048576;
or.b32 %r281, %r278, %r280;
shl.b32 %r282, %r256, 9;
and.b32 %r283, %r282, 268435456;
or.b32 %r223, %r281, %r283;

dp4a.s32.s32 %r78, %r223, %r80, %r74;

and.b32 %r284, %r257, 252645135;
shl.b32 %r285, %r258, 4;
and.b32 %r286, %r285, 16;
or.b32 %r287, %r286, %r284;
shl.b32 %r288, %r258, 11;
and.b32 %r289, %r288, 4096;
or.b32 %r290, %r287, %r289;
shl.b32 %r291, %r258, 18;
and.b32 %r292, %r291, 1048576;
or.b32 %r293, %r290, %r292;
shl.b32 %r294, %r258, 25;
and.b32 %r295, %r294, 268435456;
or.b32 %r227, %r293, %r295;

dp4a.s32.s32 %r82, %r227, %r84, %r78;

shr.u32 %r296, %r257, 4;
and.b32 %r297, %r296, 252645135;
shr.u32 %r298, %r258, 12;
and.b32 %r299, %r298, 16;
or.b32 %r300, %r299, %r297;
shr.u32 %r301, %r258, 5;
and.b32 %r302, %r301, 4096;
or.b32 %r303, %r300, %r302;
shl.b32 %r304, %r258, 2;
and.b32 %r305, %r304, 1048576;
or.b32 %r306, %r303, %r305;
shl.b32 %r307, %r258, 9;
and.b32 %r308, %r307, 268435456;
or.b32 %r231, %r306, %r308;

dp4a.s32.s32 %r86, %r231, %r88, %r82;

ld.global.nc.u32 %r90, [%rd28];

{.reg .f16 low,high;
mov.b32 {low,high},%r90;
cvt.f32.f16 %f62, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r90;
cvt.f32.f16 %f63, high;}


cvt.rn.f32.s32 %f91, %r86;
mul.ftz.f32 %f92, %f62, %f91;
mul.ftz.f32 %f93, %f63, 0f41000000;
sub.ftz.f32 %f94, %f92, %f93;
fma.rn.ftz.f32 %f260, %f61, %f94, %f260;
add.s64 %rd36, %rd89, %rd8;
ld.global.nc.u16 %rs17, [%rd36+2];
ld.global.nc.u16 %rs18, [%rd36+4];
mov.b32 %r309, {%rs17, %rs18};
add.s64 %rd37, %rd89, %rd9;
ld.global.nc.u16 %rs19, [%rd37];
ld.global.nc.u16 %rs20, [%rd37+2];
mov.b32 %r310, {%rs19, %rs20};
shr.s32 %r311, %r309, %r7;
ld.global.nc.u16 %rs21, [%rd37+4];
ld.global.nc.u16 %rs22, [%rd37+6];
mov.b32 %r312, {%rs21, %rs22};
shr.s32 %r313, %r309, %r8;
ld.global.nc.u16 %rs2, [%rd36];

{ cvt.f32.f16 %f64, %rs2;}


and.b32 %r314, %r310, 252645135;
shl.b32 %r315, %r311, 4;
and.b32 %r316, %r315, 16;
or.b32 %r317, %r316, %r314;
shl.b32 %r318, %r311, 11;
and.b32 %r319, %r318, 4096;
or.b32 %r320, %r317, %r319;
shl.b32 %r321, %r311, 18;
and.b32 %r322, %r321, 1048576;
or.b32 %r323, %r320, %r322;
shl.b32 %r324, %r311, 25;
and.b32 %r325, %r324, 268435456;
or.b32 %r237, %r323, %r325;

dp4a.s32.s32 %r92, %r237, %r76, %r239;

shr.u32 %r326, %r310, 4;
and.b32 %r327, %r326, 252645135;
shr.u32 %r328, %r311, 12;
and.b32 %r329, %r328, 16;
or.b32 %r330, %r329, %r327;
shr.u32 %r331, %r311, 5;
and.b32 %r332, %r331, 4096;
or.b32 %r333, %r330, %r332;
shl.b32 %r334, %r311, 2;
and.b32 %r335, %r334, 1048576;
or.b32 %r336, %r333, %r335;
shl.b32 %r337, %r311, 9;
and.b32 %r338, %r337, 268435456;
or.b32 %r241, %r336, %r338;

dp4a.s32.s32 %r96, %r241, %r80, %r92;

and.b32 %r339, %r312, 252645135;
shl.b32 %r340, %r313, 4;
and.b32 %r341, %r340, 16;
or.b32 %r342, %r341, %r339;
shl.b32 %r343, %r313, 11;
and.b32 %r344, %r343, 4096;
or.b32 %r345, %r342, %r344;
shl.b32 %r346, %r313, 18;
and.b32 %r347, %r346, 1048576;
or.b32 %r348, %r345, %r347;
shl.b32 %r349, %r313, 25;
and.b32 %r350, %r349, 268435456;
or.b32 %r245, %r348, %r350;

dp4a.s32.s32 %r100, %r245, %r84, %r96;

shr.u32 %r351, %r312, 4;
and.b32 %r352, %r351, 252645135;
shr.u32 %r353, %r313, 12;
and.b32 %r354, %r353, 16;
or.b32 %r355, %r354, %r352;
shr.u32 %r356, %r313, 5;
and.b32 %r357, %r356, 4096;
or.b32 %r358, %r355, %r357;
shl.b32 %r359, %r313, 2;
and.b32 %r360, %r359, 1048576;
or.b32 %r361, %r358, %r360;
shl.b32 %r362, %r313, 9;
and.b32 %r363, %r362, 268435456;
or.b32 %r249, %r361, %r363;

dp4a.s32.s32 %r104, %r249, %r88, %r100;


{.reg .f16 low,high;
mov.b32 {low,high},%r90;
cvt.f32.f16 %f65, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r90;
cvt.f32.f16 %f66, high;}


cvt.rn.f32.s32 %f95, %r104;
mul.ftz.f32 %f96, %f65, %f95;
mul.ftz.f32 %f97, %f66, 0f41000000;
sub.ftz.f32 %f98, %f96, %f97;
fma.rn.ftz.f32 %f259, %f64, %f98, %f259;
cvt.u64.u32 %rd38, %r554;
add.s64 %rd39, %rd2, %rd38;
mul.lo.s64 %rd40, %rd39, 36;
add.s64 %rd41, %rd10, %rd40;
add.s64 %rd42, %rd41, 4;
add.s64 %rd43, %rd42, %rd32;
ld.global.nc.u32 %r112, [%rd43];
ld.global.nc.u32 %r116, [%rd43+16];
add.s64 %rd44, %rd42, %rd34;
ld.global.nc.u32 %r120, [%rd44];
ld.global.nc.u32 %r124, [%rd43+20];

{ cvt.f32.f16 %f67, %rs1;}



dp4a.s32.s32 %r110, %r219, %r112, %r239;


dp4a.s32.s32 %r114, %r223, %r116, %r110;


dp4a.s32.s32 %r118, %r227, %r120, %r114;


dp4a.s32.s32 %r122, %r231, %r124, %r118;

ld.global.nc.u32 %r126, [%rd41];

{.reg .f16 low,high;
mov.b32 {low,high},%r126;
cvt.f32.f16 %f68, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r126;
cvt.f32.f16 %f69, high;}


cvt.rn.f32.s32 %f99, %r122;
mul.ftz.f32 %f100, %f68, %f99;
mul.ftz.f32 %f101, %f69, 0f41000000;
sub.ftz.f32 %f102, %f100, %f101;
fma.rn.ftz.f32 %f258, %f67, %f102, %f258;

{ cvt.f32.f16 %f70, %rs2;}



dp4a.s32.s32 %r128, %r237, %r112, %r239;


dp4a.s32.s32 %r132, %r241, %r116, %r128;


dp4a.s32.s32 %r136, %r245, %r120, %r132;


dp4a.s32.s32 %r140, %r249, %r124, %r136;


{.reg .f16 low,high;
mov.b32 {low,high},%r126;
cvt.f32.f16 %f71, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r126;
cvt.f32.f16 %f72, high;}


cvt.rn.f32.s32 %f103, %r140;
mul.ftz.f32 %f104, %f71, %f103;
mul.ftz.f32 %f105, %f72, 0f41000000;
sub.ftz.f32 %f106, %f104, %f105;
fma.rn.ftz.f32 %f257, %f70, %f106, %f257;
cvt.u64.u32 %rd45, %r557;
add.s64 %rd46, %rd2, %rd45;
mul.lo.s64 %rd47, %rd46, 36;
add.s64 %rd48, %rd10, %rd47;
add.s64 %rd49, %rd48, 4;
add.s64 %rd50, %rd49, %rd32;
ld.global.nc.u32 %r148, [%rd50];
ld.global.nc.u32 %r152, [%rd50+16];
add.s64 %rd51, %rd49, %rd34;
ld.global.nc.u32 %r156, [%rd51];
ld.global.nc.u32 %r160, [%rd50+20];

{ cvt.f32.f16 %f73, %rs1;}



dp4a.s32.s32 %r146, %r219, %r148, %r239;


dp4a.s32.s32 %r150, %r223, %r152, %r146;


dp4a.s32.s32 %r154, %r227, %r156, %r150;


dp4a.s32.s32 %r158, %r231, %r160, %r154;

ld.global.nc.u32 %r162, [%rd48];

{.reg .f16 low,high;
mov.b32 {low,high},%r162;
cvt.f32.f16 %f74, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r162;
cvt.f32.f16 %f75, high;}


cvt.rn.f32.s32 %f107, %r158;
mul.ftz.f32 %f108, %f74, %f107;
mul.ftz.f32 %f109, %f75, 0f41000000;
sub.ftz.f32 %f110, %f108, %f109;
fma.rn.ftz.f32 %f256, %f73, %f110, %f256;

{ cvt.f32.f16 %f76, %rs2;}



dp4a.s32.s32 %r164, %r237, %r148, %r239;


dp4a.s32.s32 %r168, %r241, %r152, %r164;


dp4a.s32.s32 %r172, %r245, %r156, %r168;


dp4a.s32.s32 %r176, %r249, %r160, %r172;


{.reg .f16 low,high;
mov.b32 {low,high},%r162;
cvt.f32.f16 %f77, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r162;
cvt.f32.f16 %f78, high;}


cvt.rn.f32.s32 %f111, %r176;
mul.ftz.f32 %f112, %f77, %f111;
mul.ftz.f32 %f113, %f78, 0f41000000;
sub.ftz.f32 %f114, %f112, %f113;
fma.rn.ftz.f32 %f255, %f76, %f114, %f255;
cvt.u64.u32 %rd52, %r556;
add.s64 %rd53, %rd2, %rd52;
mul.lo.s64 %rd54, %rd53, 36;
add.s64 %rd55, %rd10, %rd54;
add.s64 %rd56, %rd55, 4;
add.s64 %rd57, %rd56, %rd32;
ld.global.nc.u32 %r184, [%rd57];
ld.global.nc.u32 %r188, [%rd57+16];
add.s64 %rd58, %rd56, %rd34;
ld.global.nc.u32 %r192, [%rd58];
ld.global.nc.u32 %r196, [%rd57+20];

{ cvt.f32.f16 %f79, %rs1;}



dp4a.s32.s32 %r182, %r219, %r184, %r239;


dp4a.s32.s32 %r186, %r223, %r188, %r182;


dp4a.s32.s32 %r190, %r227, %r192, %r186;


dp4a.s32.s32 %r194, %r231, %r196, %r190;

ld.global.nc.u32 %r198, [%rd55];

{.reg .f16 low,high;
mov.b32 {low,high},%r198;
cvt.f32.f16 %f80, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r198;
cvt.f32.f16 %f81, high;}


cvt.rn.f32.s32 %f115, %r194;
mul.ftz.f32 %f116, %f80, %f115;
mul.ftz.f32 %f117, %f81, 0f41000000;
sub.ftz.f32 %f118, %f116, %f117;
fma.rn.ftz.f32 %f254, %f79, %f118, %f254;

{ cvt.f32.f16 %f82, %rs2;}



dp4a.s32.s32 %r200, %r237, %r184, %r239;


dp4a.s32.s32 %r204, %r241, %r188, %r200;


dp4a.s32.s32 %r208, %r245, %r192, %r204;


dp4a.s32.s32 %r212, %r249, %r196, %r208;


{.reg .f16 low,high;
mov.b32 {low,high},%r198;
cvt.f32.f16 %f83, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r198;
cvt.f32.f16 %f84, high;}


cvt.rn.f32.s32 %f119, %r212;
mul.ftz.f32 %f120, %f83, %f119;
mul.ftz.f32 %f121, %f84, 0f41000000;
sub.ftz.f32 %f122, %f120, %f121;
fma.rn.ftz.f32 %f253, %f82, %f122, %f253;
cvt.u64.u32 %rd59, %r555;
add.s64 %rd60, %rd2, %rd59;
mul.lo.s64 %rd61, %rd60, 36;
add.s64 %rd62, %rd10, %rd61;
add.s64 %rd63, %rd62, 4;
add.s64 %rd64, %rd63, %rd32;
ld.global.nc.u32 %r220, [%rd64];
ld.global.nc.u32 %r224, [%rd64+16];
add.s64 %rd65, %rd63, %rd34;
ld.global.nc.u32 %r228, [%rd65];
ld.global.nc.u32 %r232, [%rd64+20];

{ cvt.f32.f16 %f85, %rs1;}



dp4a.s32.s32 %r218, %r219, %r220, %r239;


dp4a.s32.s32 %r222, %r223, %r224, %r218;


dp4a.s32.s32 %r226, %r227, %r228, %r222;


dp4a.s32.s32 %r230, %r231, %r232, %r226;

ld.global.nc.u32 %r234, [%rd62];

{.reg .f16 low,high;
mov.b32 {low,high},%r234;
cvt.f32.f16 %f86, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r234;
cvt.f32.f16 %f87, high;}


cvt.rn.f32.s32 %f123, %r230;
mul.ftz.f32 %f124, %f86, %f123;
mul.ftz.f32 %f125, %f87, 0f41000000;
sub.ftz.f32 %f126, %f124, %f125;
fma.rn.ftz.f32 %f252, %f85, %f126, %f252;

{ cvt.f32.f16 %f88, %rs2;}



dp4a.s32.s32 %r236, %r237, %r220, %r239;


dp4a.s32.s32 %r240, %r241, %r224, %r236;


dp4a.s32.s32 %r244, %r245, %r228, %r240;


dp4a.s32.s32 %r248, %r249, %r232, %r244;


{.reg .f16 low,high;
mov.b32 {low,high},%r234;
cvt.f32.f16 %f89, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r234;
cvt.f32.f16 %f90, high;}


cvt.rn.f32.s32 %f127, %r248;
mul.ftz.f32 %f128, %f89, %f127;
mul.ftz.f32 %f129, %f90, 0f41000000;
sub.ftz.f32 %f130, %f128, %f129;
fma.rn.ftz.f32 %f251, %f88, %f130, %f251;
add.s32 %r557, %r557, 32;
add.s32 %r556, %r556, 32;
add.s32 %r555, %r555, 32;
add.s32 %r554, %r554, 32;
add.s64 %rd89, %rd89, 704;
add.s32 %r558, %r558, 32;
setp.lt.s32 %p4, %r558, %r6;
@%p4 bra $L__BB67_2;

$L__BB67_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB67_5;

mov.u32 %r365, _ZZ13mul_mat_vec_qIL9ggml_type6ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r366, %r1, 1280, %r365;
shl.b32 %r367, %r42, 2;
add.s32 %r368, %r366, %r367;
st.shared.f32 [%r368+-1280], %f260;
st.shared.f32 [%r368+-1152], %f259;
st.shared.f32 [%r368+-1024], %f258;
st.shared.f32 [%r368+-896], %f257;
st.shared.f32 [%r368+-768], %f256;
st.shared.f32 [%r368+-640], %f255;
st.shared.f32 [%r368+-512], %f254;
st.shared.f32 [%r368+-384], %f253;
st.shared.f32 [%r368+-256], %f252;
st.shared.f32 [%r368+-128], %f251;

$L__BB67_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB67_26;

mov.u32 %r369, %ctaid.x;
shl.b32 %r370, %r369, 1;
mov.u32 %r371, 1;
mad.lo.s32 %r372, %r3, %r33, %r370;
mad.lo.s32 %r373, %r4, %r39, %r372;
cvt.u64.u32 %rd13, %r373;
shl.b32 %r375, %r42, 2;
mov.u32 %r376, _ZZ13mul_mat_vec_qIL9ggml_type6ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r23, %r376, %r375;
mov.u32 %r377, 2;
ld.shared.f32 %f131, [%r23];
add.ftz.f32 %f132, %f131, %f260;
mov.b32 %r378, %f132;
mov.u32 %r379, 31;
mov.u32 %r380, 16;
mov.u32 %r381, -1;
shfl.sync.bfly.b32 %r382|%p7, %r378, %r380, %r379, %r381;
mov.b32 %f133, %r382;
add.ftz.f32 %f134, %f132, %f133;
mov.b32 %r383, %f134;
mov.u32 %r384, 8;
shfl.sync.bfly.b32 %r385|%p8, %r383, %r384, %r379, %r381;
mov.b32 %f135, %r385;
add.ftz.f32 %f136, %f134, %f135;
mov.b32 %r386, %f136;
mov.u32 %r387, 4;
shfl.sync.bfly.b32 %r388|%p9, %r386, %r387, %r379, %r381;
mov.b32 %f137, %r388;
add.ftz.f32 %f138, %f136, %f137;
mov.b32 %r389, %f138;
shfl.sync.bfly.b32 %r390|%p10, %r389, %r377, %r379, %r381;
mov.b32 %f139, %r390;
add.ftz.f32 %f140, %f138, %f139;
mov.b32 %r391, %f140;
shfl.sync.bfly.b32 %r392|%p11, %r391, %r371, %r379, %r381;
mov.b32 %f141, %r392;
add.ftz.f32 %f31, %f140, %f141;
setp.eq.s32 %p12, %r42, 0;
setp.lt.u32 %p13, %r370, %r27;
and.pred %p1, %p12, %p13;
mul.wide.u32 %rd66, %r373, 4;
add.s64 %rd14, %rd1, %rd66;
not.pred %p14, %p1;
@%p14 bra $L__BB67_8;

st.global.f32 [%rd14], %f31;

$L__BB67_8:
ld.shared.f32 %f142, [%r23+128];
add.ftz.f32 %f143, %f142, %f259;
mov.b32 %r393, %f143;
shfl.sync.bfly.b32 %r397|%p15, %r393, %r380, %r379, %r381;
mov.b32 %f144, %r397;
add.ftz.f32 %f145, %f143, %f144;
mov.b32 %r398, %f145;
shfl.sync.bfly.b32 %r400|%p16, %r398, %r384, %r379, %r381;
mov.b32 %f146, %r400;
add.ftz.f32 %f147, %f145, %f146;
mov.b32 %r401, %f147;
shfl.sync.bfly.b32 %r403|%p17, %r401, %r387, %r379, %r381;
mov.b32 %f148, %r403;
add.ftz.f32 %f149, %f147, %f148;
mov.b32 %r404, %f149;
shfl.sync.bfly.b32 %r406|%p18, %r404, %r377, %r379, %r381;
mov.b32 %f150, %r406;
add.ftz.f32 %f151, %f149, %f150;
mov.b32 %r407, %f151;
shfl.sync.bfly.b32 %r409|%p19, %r407, %r371, %r379, %r381;
mov.b32 %f152, %r409;
add.ftz.f32 %f32, %f151, %f152;
setp.eq.s32 %p20, %r42, 1;
or.b32 %r413, %r370, 1;
setp.lt.u32 %p21, %r413, %r27;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB67_10;

st.global.f32 [%rd14+4], %f32;

$L__BB67_10:
ld.shared.f32 %f153, [%r23+256];
add.ftz.f32 %f154, %f153, %f258;
mov.b32 %r414, %f154;
mov.u32 %r415, 31;
mov.u32 %r416, 16;
mov.u32 %r417, -1;
shfl.sync.bfly.b32 %r418|%p23, %r414, %r416, %r415, %r417;
mov.b32 %f155, %r418;
add.ftz.f32 %f156, %f154, %f155;
mov.b32 %r419, %f156;
mov.u32 %r420, 8;
shfl.sync.bfly.b32 %r421|%p24, %r419, %r420, %r415, %r417;
mov.b32 %f157, %r421;
add.ftz.f32 %f158, %f156, %f157;
mov.b32 %r422, %f158;
mov.u32 %r423, 4;
shfl.sync.bfly.b32 %r424|%p25, %r422, %r423, %r415, %r417;
mov.b32 %f159, %r424;
add.ftz.f32 %f160, %f158, %f159;
mov.b32 %r425, %f160;
mov.u32 %r426, 2;
shfl.sync.bfly.b32 %r427|%p26, %r425, %r426, %r415, %r417;
mov.b32 %f161, %r427;
add.ftz.f32 %f162, %f160, %f161;
mov.b32 %r428, %f162;
mov.u32 %r429, 1;
shfl.sync.bfly.b32 %r430|%p27, %r428, %r429, %r415, %r417;
mov.b32 %f163, %r430;
add.ftz.f32 %f33, %f162, %f163;
@%p14 bra $L__BB67_12;

cvt.u64.u32 %rd67, %r27;
add.s64 %rd68, %rd67, %rd13;
shl.b64 %rd69, %rd68, 2;
add.s64 %rd70, %rd1, %rd69;
st.global.f32 [%rd70], %f33;

$L__BB67_12:
ld.shared.f32 %f164, [%r23+384];
add.ftz.f32 %f165, %f164, %f257;
mov.b32 %r431, %f165;
shfl.sync.bfly.b32 %r435|%p29, %r431, %r416, %r415, %r417;
mov.b32 %f166, %r435;
add.ftz.f32 %f167, %f165, %f166;
mov.b32 %r436, %f167;
shfl.sync.bfly.b32 %r438|%p30, %r436, %r420, %r415, %r417;
mov.b32 %f168, %r438;
add.ftz.f32 %f169, %f167, %f168;
mov.b32 %r439, %f169;
shfl.sync.bfly.b32 %r441|%p31, %r439, %r423, %r415, %r417;
mov.b32 %f170, %r441;
add.ftz.f32 %f171, %f169, %f170;
mov.b32 %r442, %f171;
shfl.sync.bfly.b32 %r444|%p32, %r442, %r426, %r415, %r417;
mov.b32 %f172, %r444;
add.ftz.f32 %f173, %f171, %f172;
mov.b32 %r445, %f173;
shfl.sync.bfly.b32 %r447|%p33, %r445, %r429, %r415, %r417;
mov.b32 %f174, %r447;
add.ftz.f32 %f34, %f173, %f174;
@%p22 bra $L__BB67_14;

add.s32 %r448, %r27, 1;
cvt.u64.u32 %rd71, %r448;
add.s64 %rd72, %rd71, %rd13;
shl.b64 %rd73, %rd72, 2;
add.s64 %rd74, %rd1, %rd73;
st.global.f32 [%rd74], %f34;

$L__BB67_14:
shl.b32 %r449, %r27, 1;
mov.u32 %r450, 1;
ld.shared.f32 %f175, [%r23+512];
add.ftz.f32 %f176, %f175, %f256;
mov.b32 %r451, %f176;
mov.u32 %r452, 31;
mov.u32 %r453, 16;
mov.u32 %r454, -1;
shfl.sync.bfly.b32 %r455|%p35, %r451, %r453, %r452, %r454;
mov.b32 %f177, %r455;
add.ftz.f32 %f178, %f176, %f177;
mov.b32 %r456, %f178;
mov.u32 %r457, 8;
shfl.sync.bfly.b32 %r458|%p36, %r456, %r457, %r452, %r454;
mov.b32 %f179, %r458;
add.ftz.f32 %f180, %f178, %f179;
mov.b32 %r459, %f180;
mov.u32 %r460, 4;
shfl.sync.bfly.b32 %r461|%p37, %r459, %r460, %r452, %r454;
mov.b32 %f181, %r461;
add.ftz.f32 %f182, %f180, %f181;
mov.b32 %r462, %f182;
mov.u32 %r463, 2;
shfl.sync.bfly.b32 %r464|%p38, %r462, %r463, %r452, %r454;
mov.b32 %f183, %r464;
add.ftz.f32 %f184, %f182, %f183;
mov.b32 %r465, %f184;
shfl.sync.bfly.b32 %r466|%p39, %r465, %r450, %r452, %r454;
mov.b32 %f185, %r466;
add.ftz.f32 %f35, %f184, %f185;
cvt.u64.u32 %rd75, %r449;
add.s64 %rd76, %rd75, %rd13;
shl.b64 %rd77, %rd76, 2;
add.s64 %rd15, %rd1, %rd77;
@%p14 bra $L__BB67_16;

st.global.f32 [%rd15], %f35;

$L__BB67_16:
ld.shared.f32 %f186, [%r23+640];
add.ftz.f32 %f187, %f186, %f255;
mov.b32 %r467, %f187;
shfl.sync.bfly.b32 %r471|%p41, %r467, %r453, %r452, %r454;
mov.b32 %f188, %r471;
add.ftz.f32 %f189, %f187, %f188;
mov.b32 %r472, %f189;
shfl.sync.bfly.b32 %r474|%p42, %r472, %r457, %r452, %r454;
mov.b32 %f190, %r474;
add.ftz.f32 %f191, %f189, %f190;
mov.b32 %r475, %f191;
shfl.sync.bfly.b32 %r477|%p43, %r475, %r460, %r452, %r454;
mov.b32 %f192, %r477;
add.ftz.f32 %f193, %f191, %f192;
mov.b32 %r478, %f193;
shfl.sync.bfly.b32 %r480|%p44, %r478, %r463, %r452, %r454;
mov.b32 %f194, %r480;
add.ftz.f32 %f195, %f193, %f194;
mov.b32 %r481, %f195;
shfl.sync.bfly.b32 %r483|%p45, %r481, %r450, %r452, %r454;
mov.b32 %f196, %r483;
add.ftz.f32 %f36, %f195, %f196;
@%p22 bra $L__BB67_18;

st.global.f32 [%rd15+4], %f36;

$L__BB67_18:
mul.lo.s32 %r24, %r27, 3;
ld.shared.f32 %f197, [%r23+768];
add.ftz.f32 %f198, %f197, %f254;
mov.b32 %r484, %f198;
mov.u32 %r485, 31;
mov.u32 %r486, 16;
mov.u32 %r487, -1;
shfl.sync.bfly.b32 %r488|%p47, %r484, %r486, %r485, %r487;
mov.b32 %f199, %r488;
add.ftz.f32 %f200, %f198, %f199;
mov.b32 %r489, %f200;
mov.u32 %r490, 8;
shfl.sync.bfly.b32 %r491|%p48, %r489, %r490, %r485, %r487;
mov.b32 %f201, %r491;
add.ftz.f32 %f202, %f200, %f201;
mov.b32 %r492, %f202;
mov.u32 %r493, 4;
shfl.sync.bfly.b32 %r494|%p49, %r492, %r493, %r485, %r487;
mov.b32 %f203, %r494;
add.ftz.f32 %f204, %f202, %f203;
mov.b32 %r495, %f204;
mov.u32 %r496, 2;
shfl.sync.bfly.b32 %r497|%p50, %r495, %r496, %r485, %r487;
mov.b32 %f205, %r497;
add.ftz.f32 %f206, %f204, %f205;
mov.b32 %r498, %f206;
mov.u32 %r499, 1;
shfl.sync.bfly.b32 %r500|%p51, %r498, %r499, %r485, %r487;
mov.b32 %f207, %r500;
add.ftz.f32 %f37, %f206, %f207;
@%p14 bra $L__BB67_20;

cvt.u64.u32 %rd78, %r24;
add.s64 %rd79, %rd78, %rd13;
shl.b64 %rd80, %rd79, 2;
add.s64 %rd81, %rd1, %rd80;
st.global.f32 [%rd81], %f37;

$L__BB67_20:
ld.shared.f32 %f208, [%r23+896];
add.ftz.f32 %f209, %f208, %f253;
mov.b32 %r501, %f209;
shfl.sync.bfly.b32 %r505|%p53, %r501, %r486, %r485, %r487;
mov.b32 %f210, %r505;
add.ftz.f32 %f211, %f209, %f210;
mov.b32 %r506, %f211;
shfl.sync.bfly.b32 %r508|%p54, %r506, %r490, %r485, %r487;
mov.b32 %f212, %r508;
add.ftz.f32 %f213, %f211, %f212;
mov.b32 %r509, %f213;
shfl.sync.bfly.b32 %r511|%p55, %r509, %r493, %r485, %r487;
mov.b32 %f214, %r511;
add.ftz.f32 %f215, %f213, %f214;
mov.b32 %r512, %f215;
shfl.sync.bfly.b32 %r514|%p56, %r512, %r496, %r485, %r487;
mov.b32 %f216, %r514;
add.ftz.f32 %f217, %f215, %f216;
mov.b32 %r515, %f217;
shfl.sync.bfly.b32 %r517|%p57, %r515, %r499, %r485, %r487;
mov.b32 %f218, %r517;
add.ftz.f32 %f38, %f217, %f218;
@%p22 bra $L__BB67_22;

add.s32 %r518, %r24, 1;
cvt.u64.u32 %rd82, %r518;
add.s64 %rd83, %rd82, %rd13;
shl.b64 %rd84, %rd83, 2;
add.s64 %rd85, %rd1, %rd84;
st.global.f32 [%rd85], %f38;

$L__BB67_22:
shl.b32 %r519, %r27, 2;
mov.u32 %r520, 2;
ld.shared.f32 %f219, [%r23+1024];
add.ftz.f32 %f220, %f219, %f252;
mov.b32 %r521, %f220;
mov.u32 %r522, 31;
mov.u32 %r523, 16;
mov.u32 %r524, -1;
shfl.sync.bfly.b32 %r525|%p59, %r521, %r523, %r522, %r524;
mov.b32 %f221, %r525;
add.ftz.f32 %f222, %f220, %f221;
mov.b32 %r526, %f222;
mov.u32 %r527, 8;
shfl.sync.bfly.b32 %r528|%p60, %r526, %r527, %r522, %r524;
mov.b32 %f223, %r528;
add.ftz.f32 %f224, %f222, %f223;
mov.b32 %r529, %f224;
mov.u32 %r530, 4;
shfl.sync.bfly.b32 %r531|%p61, %r529, %r530, %r522, %r524;
mov.b32 %f225, %r531;
add.ftz.f32 %f226, %f224, %f225;
mov.b32 %r532, %f226;
shfl.sync.bfly.b32 %r533|%p62, %r532, %r520, %r522, %r524;
mov.b32 %f227, %r533;
add.ftz.f32 %f228, %f226, %f227;
mov.b32 %r534, %f228;
mov.u32 %r535, 1;
shfl.sync.bfly.b32 %r536|%p63, %r534, %r535, %r522, %r524;
mov.b32 %f229, %r536;
add.ftz.f32 %f39, %f228, %f229;
cvt.u64.u32 %rd86, %r519;
add.s64 %rd87, %rd86, %rd13;
shl.b64 %rd88, %rd87, 2;
add.s64 %rd16, %rd1, %rd88;
@%p14 bra $L__BB67_24;

st.global.f32 [%rd16], %f39;

$L__BB67_24:
ld.shared.f32 %f230, [%r23+1152];
add.ftz.f32 %f231, %f230, %f251;
mov.b32 %r537, %f231;
shfl.sync.bfly.b32 %r541|%p65, %r537, %r523, %r522, %r524;
mov.b32 %f232, %r541;
add.ftz.f32 %f233, %f231, %f232;
mov.b32 %r542, %f233;
shfl.sync.bfly.b32 %r544|%p66, %r542, %r527, %r522, %r524;
mov.b32 %f234, %r544;
add.ftz.f32 %f235, %f233, %f234;
mov.b32 %r545, %f235;
shfl.sync.bfly.b32 %r547|%p67, %r545, %r530, %r522, %r524;
mov.b32 %f236, %r547;
add.ftz.f32 %f237, %f235, %f236;
mov.b32 %r548, %f237;
shfl.sync.bfly.b32 %r550|%p68, %r548, %r520, %r522, %r524;
mov.b32 %f238, %r550;
add.ftz.f32 %f239, %f237, %f238;
mov.b32 %r551, %f239;
shfl.sync.bfly.b32 %r553|%p69, %r551, %r535, %r522, %r524;
mov.b32 %f240, %r553;
add.ftz.f32 %f40, %f239, %f240;
@%p22 bra $L__BB67_26;

st.global.f32 [%rd16+4], %f40;

$L__BB67_26:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type6ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type6ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type6ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type6ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type6ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type6ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type6ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type6ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type6ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<83>;
.reg .b16 %rs<25>;
.reg .f32 %f<313>;
.reg .b32 %r<646>;
.reg .b64 %rd<116>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type6ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type6ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type6ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type6ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type6ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type6ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type6ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type6ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type6ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type6ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type6ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type6ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type6ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type6ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type6ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type6ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type6ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type6ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r44, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r44, %r2;
mov.u32 %r45, %ctaid.x;
shl.b32 %r4, %r45, 1;
shr.u32 %r46, %r3, 31;
add.s32 %r47, %r3, %r46;
shr.s32 %r645, %r47, 1;
shr.u32 %r48, %r28, 5;
setp.ge.s32 %p3, %r645, %r48;
mov.f32 %f301, 0f00000000;
mov.f32 %f302, %f301;
mov.f32 %f303, %f301;
mov.f32 %f304, %f301;
mov.f32 %f305, %f301;
mov.f32 %f306, %f301;
mov.f32 %f307, %f301;
mov.f32 %f308, %f301;
mov.f32 %f309, %f301;
mov.f32 %f310, %f301;
mov.f32 %f311, %f301;
mov.f32 %f312, %f301;
@%p3 bra $L__BB68_3;

mov.u32 %r49, %ctaid.y;
mov.u32 %r50, %ctaid.z;
mul.lo.s32 %r51, %r49, %r36;
cvt.u64.u32 %rd18, %r51;
mul.lo.s32 %r52, %r50, %r42;
cvt.u64.u32 %rd19, %r52;
add.s64 %rd1, %rd19, %rd18;
mul.hi.u32 %r53, %r50, %r38;
add.s32 %r54, %r53, %r50;
shr.u32 %r55, %r54, %r39;
mul.lo.s32 %r56, %r55, %r41;
mul.hi.u32 %r57, %r49, %r32;
add.s32 %r58, %r57, %r49;
shr.u32 %r59, %r58, %r33;
mul.lo.s32 %r60, %r59, %r35;
cvta.to.global.u64 %rd115, %rd15;
and.b32 %r63, %r47, -2;
sub.s32 %r6, %r3, %r63;
shl.b32 %r64, %r6, 1;
shl.b32 %r65, %r6, 2;
or.b32 %r66, %r65, 1;
cvt.s64.s32 %rd3, %r64;
sub.s32 %r67, %r66, %r64;
shl.b32 %r7, %r67, 2;
cvt.s64.s32 %rd4, %r67;
shl.b32 %r68, %r30, 1;
add.s32 %r644, %r645, %r68;
mad.lo.s32 %r643, %r30, 3, %r645;
shl.b32 %r69, %r30, 2;
add.s32 %r642, %r645, %r69;
mad.lo.s32 %r641, %r30, 5, %r645;
add.s32 %r640, %r645, %r30;
mad.lo.s32 %r70, %r4, %r29, %r645;
add.s32 %r71, %r70, %r56;
add.s32 %r72, %r71, %r60;
mul.wide.s32 %rd5, %r72, 22;
mul.wide.s32 %rd20, %r65, 2;
add.s64 %rd21, %rd5, %rd20;
add.s64 %rd6, %rd21, 8;
add.s32 %r73, %r645, %r56;
add.s32 %r74, %r73, %r60;
add.s32 %r75, %r4, 1;
mad.lo.s32 %r76, %r29, %r75, %r74;
mul.wide.s32 %rd7, %r76, 22;
add.s64 %rd22, %rd7, %rd20;
add.s64 %rd8, %rd22, 6;
cvta.to.global.u64 %rd9, %rd16;
mov.f32 %f301, 0f00000000;
shl.b64 %rd30, %rd3, 2;
shl.b64 %rd32, %rd4, 2;

$L__BB68_2:
cvt.u64.u32 %rd23, %r645;
add.s64 %rd24, %rd1, %rd23;
mul.lo.s64 %rd25, %rd24, 36;
add.s64 %rd26, %rd9, %rd25;
add.s64 %rd27, %rd26, 4;
add.s64 %rd28, %rd115, %rd5;
ld.global.nc.u16 %rs13, [%rd28+2];
ld.global.nc.u16 %rs14, [%rd28+4];
mov.b32 %r293, {%rs13, %rs14};
add.s64 %rd29, %rd115, %rd6;
ld.global.nc.u16 %rs15, [%rd29+-2];
ld.global.nc.u16 %rs16, [%rd29];
mov.b32 %r294, {%rs15, %rs16};
shl.b32 %r295, %r6, 3;
shr.s32 %r296, %r293, %r295;
add.s64 %rd31, %rd27, %rd30;
ld.global.nc.u32 %r79, [%rd31];
ld.global.nc.u32 %r83, [%rd31+16];
ld.global.nc.u16 %rs17, [%rd29+2];
ld.global.nc.u16 %rs18, [%rd29+4];
mov.b32 %r297, {%rs17, %rs18};
shr.s32 %r298, %r293, %r7;
add.s64 %rd33, %rd27, %rd32;
ld.global.nc.u32 %r87, [%rd33];
ld.global.nc.u32 %r91, [%rd31+20];
ld.global.nc.u16 %rs1, [%rd28];

{ cvt.f32.f16 %f73, %rs1;}


and.b32 %r299, %r294, 252645135;
shl.b32 %r300, %r296, 4;
and.b32 %r301, %r300, 16;
or.b32 %r302, %r301, %r299;
shl.b32 %r303, %r296, 11;
and.b32 %r304, %r303, 4096;
or.b32 %r305, %r302, %r304;
shl.b32 %r306, %r296, 18;
and.b32 %r307, %r306, 1048576;
or.b32 %r308, %r305, %r307;
shl.b32 %r309, %r296, 25;
and.b32 %r310, %r309, 268435456;
or.b32 %r258, %r308, %r310;
mov.u32 %r278, 0;

dp4a.s32.s32 %r77, %r258, %r79, %r278;

shr.u32 %r311, %r294, 4;
and.b32 %r312, %r311, 252645135;
shr.u32 %r313, %r296, 12;
and.b32 %r314, %r313, 16;
or.b32 %r315, %r314, %r312;
shr.u32 %r316, %r296, 5;
and.b32 %r317, %r316, 4096;
or.b32 %r318, %r315, %r317;
shl.b32 %r319, %r296, 2;
and.b32 %r320, %r319, 1048576;
or.b32 %r321, %r318, %r320;
shl.b32 %r322, %r296, 9;
and.b32 %r323, %r322, 268435456;
or.b32 %r262, %r321, %r323;

dp4a.s32.s32 %r81, %r262, %r83, %r77;

and.b32 %r324, %r297, 252645135;
shl.b32 %r325, %r298, 4;
and.b32 %r326, %r325, 16;
or.b32 %r327, %r326, %r324;
shl.b32 %r328, %r298, 11;
and.b32 %r329, %r328, 4096;
or.b32 %r330, %r327, %r329;
shl.b32 %r331, %r298, 18;
and.b32 %r332, %r331, 1048576;
or.b32 %r333, %r330, %r332;
shl.b32 %r334, %r298, 25;
and.b32 %r335, %r334, 268435456;
or.b32 %r266, %r333, %r335;

dp4a.s32.s32 %r85, %r266, %r87, %r81;

shr.u32 %r336, %r297, 4;
and.b32 %r337, %r336, 252645135;
shr.u32 %r338, %r298, 12;
and.b32 %r339, %r338, 16;
or.b32 %r340, %r339, %r337;
shr.u32 %r341, %r298, 5;
and.b32 %r342, %r341, 4096;
or.b32 %r343, %r340, %r342;
shl.b32 %r344, %r298, 2;
and.b32 %r345, %r344, 1048576;
or.b32 %r346, %r343, %r345;
shl.b32 %r347, %r298, 9;
and.b32 %r348, %r347, 268435456;
or.b32 %r270, %r346, %r348;

dp4a.s32.s32 %r89, %r270, %r91, %r85;

ld.global.nc.u32 %r93, [%rd26];

{.reg .f16 low,high;
mov.b32 {low,high},%r93;
cvt.f32.f16 %f74, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r93;
cvt.f32.f16 %f75, high;}


cvt.rn.f32.s32 %f109, %r89;
mul.ftz.f32 %f110, %f74, %f109;
mul.ftz.f32 %f111, %f75, 0f41000000;
sub.ftz.f32 %f112, %f110, %f111;
fma.rn.ftz.f32 %f312, %f73, %f112, %f312;
add.s64 %rd34, %rd115, %rd7;
ld.global.nc.u16 %rs19, [%rd34+2];
ld.global.nc.u16 %rs20, [%rd34+4];
mov.b32 %r349, {%rs19, %rs20};
add.s64 %rd35, %rd115, %rd8;
ld.global.nc.u16 %rs21, [%rd35];
ld.global.nc.u16 %rs22, [%rd35+2];
mov.b32 %r350, {%rs21, %rs22};
shr.s32 %r351, %r349, %r295;
ld.global.nc.u16 %rs23, [%rd35+4];
ld.global.nc.u16 %rs24, [%rd35+6];
mov.b32 %r352, {%rs23, %rs24};
shr.s32 %r353, %r349, %r7;
ld.global.nc.u16 %rs2, [%rd34];

{ cvt.f32.f16 %f76, %rs2;}


and.b32 %r354, %r350, 252645135;
shl.b32 %r355, %r351, 4;
and.b32 %r356, %r355, 16;
or.b32 %r357, %r356, %r354;
shl.b32 %r358, %r351, 11;
and.b32 %r359, %r358, 4096;
or.b32 %r360, %r357, %r359;
shl.b32 %r361, %r351, 18;
and.b32 %r362, %r361, 1048576;
or.b32 %r363, %r360, %r362;
shl.b32 %r364, %r351, 25;
and.b32 %r365, %r364, 268435456;
or.b32 %r276, %r363, %r365;

dp4a.s32.s32 %r95, %r276, %r79, %r278;

shr.u32 %r366, %r350, 4;
and.b32 %r367, %r366, 252645135;
shr.u32 %r368, %r351, 12;
and.b32 %r369, %r368, 16;
or.b32 %r370, %r369, %r367;
shr.u32 %r371, %r351, 5;
and.b32 %r372, %r371, 4096;
or.b32 %r373, %r370, %r372;
shl.b32 %r374, %r351, 2;
and.b32 %r375, %r374, 1048576;
or.b32 %r376, %r373, %r375;
shl.b32 %r377, %r351, 9;
and.b32 %r378, %r377, 268435456;
or.b32 %r280, %r376, %r378;

dp4a.s32.s32 %r99, %r280, %r83, %r95;

and.b32 %r379, %r352, 252645135;
shl.b32 %r380, %r353, 4;
and.b32 %r381, %r380, 16;
or.b32 %r382, %r381, %r379;
shl.b32 %r383, %r353, 11;
and.b32 %r384, %r383, 4096;
or.b32 %r385, %r382, %r384;
shl.b32 %r386, %r353, 18;
and.b32 %r387, %r386, 1048576;
or.b32 %r388, %r385, %r387;
shl.b32 %r389, %r353, 25;
and.b32 %r390, %r389, 268435456;
or.b32 %r284, %r388, %r390;

dp4a.s32.s32 %r103, %r284, %r87, %r99;

shr.u32 %r391, %r352, 4;
and.b32 %r392, %r391, 252645135;
shr.u32 %r393, %r353, 12;
and.b32 %r394, %r393, 16;
or.b32 %r395, %r394, %r392;
shr.u32 %r396, %r353, 5;
and.b32 %r397, %r396, 4096;
or.b32 %r398, %r395, %r397;
shl.b32 %r399, %r353, 2;
and.b32 %r400, %r399, 1048576;
or.b32 %r401, %r398, %r400;
shl.b32 %r402, %r353, 9;
and.b32 %r403, %r402, 268435456;
or.b32 %r288, %r401, %r403;

dp4a.s32.s32 %r107, %r288, %r91, %r103;


{.reg .f16 low,high;
mov.b32 {low,high},%r93;
cvt.f32.f16 %f77, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r93;
cvt.f32.f16 %f78, high;}


cvt.rn.f32.s32 %f113, %r107;
mul.ftz.f32 %f114, %f77, %f113;
mul.ftz.f32 %f115, %f78, 0f41000000;
sub.ftz.f32 %f116, %f114, %f115;
fma.rn.ftz.f32 %f311, %f76, %f116, %f311;
cvt.u64.u32 %rd36, %r640;
add.s64 %rd37, %rd1, %rd36;
mul.lo.s64 %rd38, %rd37, 36;
add.s64 %rd39, %rd9, %rd38;
add.s64 %rd40, %rd39, 4;
add.s64 %rd41, %rd40, %rd30;
ld.global.nc.u32 %r115, [%rd41];
ld.global.nc.u32 %r119, [%rd41+16];
add.s64 %rd42, %rd40, %rd32;
ld.global.nc.u32 %r123, [%rd42];
ld.global.nc.u32 %r127, [%rd41+20];

{ cvt.f32.f16 %f79, %rs1;}



dp4a.s32.s32 %r113, %r258, %r115, %r278;


dp4a.s32.s32 %r117, %r262, %r119, %r113;


dp4a.s32.s32 %r121, %r266, %r123, %r117;


dp4a.s32.s32 %r125, %r270, %r127, %r121;

ld.global.nc.u32 %r129, [%rd39];

{.reg .f16 low,high;
mov.b32 {low,high},%r129;
cvt.f32.f16 %f80, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r129;
cvt.f32.f16 %f81, high;}


cvt.rn.f32.s32 %f117, %r125;
mul.ftz.f32 %f118, %f80, %f117;
mul.ftz.f32 %f119, %f81, 0f41000000;
sub.ftz.f32 %f120, %f118, %f119;
fma.rn.ftz.f32 %f310, %f79, %f120, %f310;

{ cvt.f32.f16 %f82, %rs2;}



dp4a.s32.s32 %r131, %r276, %r115, %r278;


dp4a.s32.s32 %r135, %r280, %r119, %r131;


dp4a.s32.s32 %r139, %r284, %r123, %r135;


dp4a.s32.s32 %r143, %r288, %r127, %r139;


{.reg .f16 low,high;
mov.b32 {low,high},%r129;
cvt.f32.f16 %f83, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r129;
cvt.f32.f16 %f84, high;}


cvt.rn.f32.s32 %f121, %r143;
mul.ftz.f32 %f122, %f83, %f121;
mul.ftz.f32 %f123, %f84, 0f41000000;
sub.ftz.f32 %f124, %f122, %f123;
fma.rn.ftz.f32 %f309, %f82, %f124, %f309;
cvt.u64.u32 %rd43, %r644;
add.s64 %rd44, %rd1, %rd43;
mul.lo.s64 %rd45, %rd44, 36;
add.s64 %rd46, %rd9, %rd45;
add.s64 %rd47, %rd46, 4;
add.s64 %rd48, %rd47, %rd30;
ld.global.nc.u32 %r151, [%rd48];
ld.global.nc.u32 %r155, [%rd48+16];
add.s64 %rd49, %rd47, %rd32;
ld.global.nc.u32 %r159, [%rd49];
ld.global.nc.u32 %r163, [%rd48+20];

{ cvt.f32.f16 %f85, %rs1;}



dp4a.s32.s32 %r149, %r258, %r151, %r278;


dp4a.s32.s32 %r153, %r262, %r155, %r149;


dp4a.s32.s32 %r157, %r266, %r159, %r153;


dp4a.s32.s32 %r161, %r270, %r163, %r157;

ld.global.nc.u32 %r165, [%rd46];

{.reg .f16 low,high;
mov.b32 {low,high},%r165;
cvt.f32.f16 %f86, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r165;
cvt.f32.f16 %f87, high;}


cvt.rn.f32.s32 %f125, %r161;
mul.ftz.f32 %f126, %f86, %f125;
mul.ftz.f32 %f127, %f87, 0f41000000;
sub.ftz.f32 %f128, %f126, %f127;
fma.rn.ftz.f32 %f308, %f85, %f128, %f308;

{ cvt.f32.f16 %f88, %rs2;}



dp4a.s32.s32 %r167, %r276, %r151, %r278;


dp4a.s32.s32 %r171, %r280, %r155, %r167;


dp4a.s32.s32 %r175, %r284, %r159, %r171;


dp4a.s32.s32 %r179, %r288, %r163, %r175;


{.reg .f16 low,high;
mov.b32 {low,high},%r165;
cvt.f32.f16 %f89, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r165;
cvt.f32.f16 %f90, high;}


cvt.rn.f32.s32 %f129, %r179;
mul.ftz.f32 %f130, %f89, %f129;
mul.ftz.f32 %f131, %f90, 0f41000000;
sub.ftz.f32 %f132, %f130, %f131;
fma.rn.ftz.f32 %f307, %f88, %f132, %f307;
cvt.u64.u32 %rd50, %r643;
add.s64 %rd51, %rd1, %rd50;
mul.lo.s64 %rd52, %rd51, 36;
add.s64 %rd53, %rd9, %rd52;
add.s64 %rd54, %rd53, 4;
add.s64 %rd55, %rd54, %rd30;
ld.global.nc.u32 %r187, [%rd55];
ld.global.nc.u32 %r191, [%rd55+16];
add.s64 %rd56, %rd54, %rd32;
ld.global.nc.u32 %r195, [%rd56];
ld.global.nc.u32 %r199, [%rd55+20];

{ cvt.f32.f16 %f91, %rs1;}



dp4a.s32.s32 %r185, %r258, %r187, %r278;


dp4a.s32.s32 %r189, %r262, %r191, %r185;


dp4a.s32.s32 %r193, %r266, %r195, %r189;


dp4a.s32.s32 %r197, %r270, %r199, %r193;

ld.global.nc.u32 %r201, [%rd53];

{.reg .f16 low,high;
mov.b32 {low,high},%r201;
cvt.f32.f16 %f92, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r201;
cvt.f32.f16 %f93, high;}


cvt.rn.f32.s32 %f133, %r197;
mul.ftz.f32 %f134, %f92, %f133;
mul.ftz.f32 %f135, %f93, 0f41000000;
sub.ftz.f32 %f136, %f134, %f135;
fma.rn.ftz.f32 %f306, %f91, %f136, %f306;

{ cvt.f32.f16 %f94, %rs2;}



dp4a.s32.s32 %r203, %r276, %r187, %r278;


dp4a.s32.s32 %r207, %r280, %r191, %r203;


dp4a.s32.s32 %r211, %r284, %r195, %r207;


dp4a.s32.s32 %r215, %r288, %r199, %r211;


{.reg .f16 low,high;
mov.b32 {low,high},%r201;
cvt.f32.f16 %f95, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r201;
cvt.f32.f16 %f96, high;}


cvt.rn.f32.s32 %f137, %r215;
mul.ftz.f32 %f138, %f95, %f137;
mul.ftz.f32 %f139, %f96, 0f41000000;
sub.ftz.f32 %f140, %f138, %f139;
fma.rn.ftz.f32 %f305, %f94, %f140, %f305;
cvt.u64.u32 %rd57, %r642;
add.s64 %rd58, %rd1, %rd57;
mul.lo.s64 %rd59, %rd58, 36;
add.s64 %rd60, %rd9, %rd59;
add.s64 %rd61, %rd60, 4;
add.s64 %rd62, %rd61, %rd30;
ld.global.nc.u32 %r223, [%rd62];
ld.global.nc.u32 %r227, [%rd62+16];
add.s64 %rd63, %rd61, %rd32;
ld.global.nc.u32 %r231, [%rd63];
ld.global.nc.u32 %r235, [%rd62+20];

{ cvt.f32.f16 %f97, %rs1;}



dp4a.s32.s32 %r221, %r258, %r223, %r278;


dp4a.s32.s32 %r225, %r262, %r227, %r221;


dp4a.s32.s32 %r229, %r266, %r231, %r225;


dp4a.s32.s32 %r233, %r270, %r235, %r229;

ld.global.nc.u32 %r237, [%rd60];

{.reg .f16 low,high;
mov.b32 {low,high},%r237;
cvt.f32.f16 %f98, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r237;
cvt.f32.f16 %f99, high;}


cvt.rn.f32.s32 %f141, %r233;
mul.ftz.f32 %f142, %f98, %f141;
mul.ftz.f32 %f143, %f99, 0f41000000;
sub.ftz.f32 %f144, %f142, %f143;
fma.rn.ftz.f32 %f304, %f97, %f144, %f304;

{ cvt.f32.f16 %f100, %rs2;}



dp4a.s32.s32 %r239, %r276, %r223, %r278;


dp4a.s32.s32 %r243, %r280, %r227, %r239;


dp4a.s32.s32 %r247, %r284, %r231, %r243;


dp4a.s32.s32 %r251, %r288, %r235, %r247;


{.reg .f16 low,high;
mov.b32 {low,high},%r237;
cvt.f32.f16 %f101, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r237;
cvt.f32.f16 %f102, high;}


cvt.rn.f32.s32 %f145, %r251;
mul.ftz.f32 %f146, %f101, %f145;
mul.ftz.f32 %f147, %f102, 0f41000000;
sub.ftz.f32 %f148, %f146, %f147;
fma.rn.ftz.f32 %f303, %f100, %f148, %f303;
cvt.u64.u32 %rd64, %r641;
add.s64 %rd65, %rd1, %rd64;
mul.lo.s64 %rd66, %rd65, 36;
add.s64 %rd67, %rd9, %rd66;
add.s64 %rd68, %rd67, 4;
add.s64 %rd69, %rd68, %rd30;
ld.global.nc.u32 %r259, [%rd69];
ld.global.nc.u32 %r263, [%rd69+16];
add.s64 %rd70, %rd68, %rd32;
ld.global.nc.u32 %r267, [%rd70];
ld.global.nc.u32 %r271, [%rd69+20];

{ cvt.f32.f16 %f103, %rs1;}



dp4a.s32.s32 %r257, %r258, %r259, %r278;


dp4a.s32.s32 %r261, %r262, %r263, %r257;


dp4a.s32.s32 %r265, %r266, %r267, %r261;


dp4a.s32.s32 %r269, %r270, %r271, %r265;

ld.global.nc.u32 %r273, [%rd67];

{.reg .f16 low,high;
mov.b32 {low,high},%r273;
cvt.f32.f16 %f104, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r273;
cvt.f32.f16 %f105, high;}


cvt.rn.f32.s32 %f149, %r269;
mul.ftz.f32 %f150, %f104, %f149;
mul.ftz.f32 %f151, %f105, 0f41000000;
sub.ftz.f32 %f152, %f150, %f151;
fma.rn.ftz.f32 %f302, %f103, %f152, %f302;

{ cvt.f32.f16 %f106, %rs2;}



dp4a.s32.s32 %r275, %r276, %r259, %r278;


dp4a.s32.s32 %r279, %r280, %r263, %r275;


dp4a.s32.s32 %r283, %r284, %r267, %r279;


dp4a.s32.s32 %r287, %r288, %r271, %r283;


{.reg .f16 low,high;
mov.b32 {low,high},%r273;
cvt.f32.f16 %f107, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r273;
cvt.f32.f16 %f108, high;}


cvt.rn.f32.s32 %f153, %r287;
mul.ftz.f32 %f154, %f107, %f153;
mul.ftz.f32 %f155, %f108, 0f41000000;
sub.ftz.f32 %f156, %f154, %f155;
fma.rn.ftz.f32 %f301, %f106, %f156, %f301;
add.s32 %r644, %r644, 32;
add.s32 %r643, %r643, 32;
add.s32 %r642, %r642, 32;
add.s32 %r641, %r641, 32;
add.s32 %r640, %r640, 32;
add.s64 %rd115, %rd115, 704;
add.s32 %r645, %r645, 32;
setp.lt.s32 %p4, %r645, %r48;
@%p4 bra $L__BB68_2;

$L__BB68_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB68_5;

mov.u32 %r405, _ZZ13mul_mat_vec_qIL9ggml_type6ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r406, %r1, 1536, %r405;
shl.b32 %r407, %r2, 2;
add.s32 %r408, %r406, %r407;
st.shared.f32 [%r408+-1536], %f312;
st.shared.f32 [%r408+-1408], %f311;
st.shared.f32 [%r408+-1280], %f310;
st.shared.f32 [%r408+-1152], %f309;
st.shared.f32 [%r408+-1024], %f308;
st.shared.f32 [%r408+-896], %f307;
st.shared.f32 [%r408+-768], %f306;
st.shared.f32 [%r408+-640], %f305;
st.shared.f32 [%r408+-512], %f304;
st.shared.f32 [%r408+-384], %f303;
st.shared.f32 [%r408+-256], %f302;
st.shared.f32 [%r408+-128], %f301;

$L__BB68_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB68_30;

mov.u32 %r409, %ctaid.z;
mov.u32 %r410, %ctaid.y;
mad.lo.s32 %r411, %r410, %r37, %r4;
mad.lo.s32 %r412, %r409, %r43, %r411;
cvt.u64.u32 %rd12, %r412;
shl.b32 %r413, %r2, 2;
mov.u32 %r414, _ZZ13mul_mat_vec_qIL9ggml_type6ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r25, %r414, %r413;
mov.u32 %r415, 2;
ld.shared.f32 %f157, [%r25];
add.ftz.f32 %f158, %f157, %f312;
mov.b32 %r416, %f158;
mov.u32 %r417, 31;
mov.u32 %r418, 16;
mov.u32 %r419, -1;
shfl.sync.bfly.b32 %r420|%p7, %r416, %r418, %r417, %r419;
mov.b32 %f159, %r420;
add.ftz.f32 %f160, %f158, %f159;
mov.b32 %r421, %f160;
mov.u32 %r422, 8;
shfl.sync.bfly.b32 %r423|%p8, %r421, %r422, %r417, %r419;
mov.b32 %f161, %r423;
add.ftz.f32 %f162, %f160, %f161;
mov.b32 %r424, %f162;
mov.u32 %r425, 4;
shfl.sync.bfly.b32 %r426|%p9, %r424, %r425, %r417, %r419;
mov.b32 %f163, %r426;
add.ftz.f32 %f164, %f162, %f163;
mov.b32 %r427, %f164;
shfl.sync.bfly.b32 %r428|%p10, %r427, %r415, %r417, %r419;
mov.b32 %f165, %r428;
add.ftz.f32 %f166, %f164, %f165;
mov.b32 %r429, %f166;
mov.u32 %r430, 1;
shfl.sync.bfly.b32 %r431|%p11, %r429, %r430, %r417, %r419;
mov.b32 %f167, %r431;
add.ftz.f32 %f37, %f166, %f167;
setp.lt.u32 %p12, %r4, %r31;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
not.pred %p14, %p1;
@%p14 bra $L__BB68_8;

cvta.to.global.u64 %rd71, %rd17;
mul.wide.u32 %rd72, %r412, 4;
add.s64 %rd73, %rd71, %rd72;
st.global.f32 [%rd73], %f37;

$L__BB68_8:
ld.shared.f32 %f168, [%r25+128];
add.ftz.f32 %f169, %f168, %f311;
mov.b32 %r438, %f169;
shfl.sync.bfly.b32 %r442|%p15, %r438, %r418, %r417, %r419;
mov.b32 %f170, %r442;
add.ftz.f32 %f171, %f169, %f170;
mov.b32 %r443, %f171;
shfl.sync.bfly.b32 %r445|%p16, %r443, %r422, %r417, %r419;
mov.b32 %f172, %r445;
add.ftz.f32 %f173, %f171, %f172;
mov.b32 %r446, %f173;
shfl.sync.bfly.b32 %r448|%p17, %r446, %r425, %r417, %r419;
mov.b32 %f174, %r448;
add.ftz.f32 %f175, %f173, %f174;
mov.b32 %r449, %f175;
shfl.sync.bfly.b32 %r451|%p18, %r449, %r415, %r417, %r419;
mov.b32 %f176, %r451;
add.ftz.f32 %f177, %f175, %f176;
mov.b32 %r452, %f177;
shfl.sync.bfly.b32 %r454|%p19, %r452, %r430, %r417, %r419;
mov.b32 %f178, %r454;
add.ftz.f32 %f38, %f177, %f178;
setp.eq.s32 %p20, %r2, 1;
or.b32 %r458, %r4, 1;
setp.lt.u32 %p21, %r458, %r31;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB68_10;

cvta.to.global.u64 %rd74, %rd17;
mul.wide.u32 %rd75, %r412, 4;
add.s64 %rd76, %rd74, %rd75;
st.global.f32 [%rd76+4], %f38;

$L__BB68_10:
ld.shared.f32 %f179, [%r25+256];
add.ftz.f32 %f180, %f179, %f310;
mov.b32 %r465, %f180;
mov.u32 %r466, 31;
mov.u32 %r467, 16;
mov.u32 %r468, -1;
shfl.sync.bfly.b32 %r469|%p23, %r465, %r467, %r466, %r468;
mov.b32 %f181, %r469;
add.ftz.f32 %f182, %f180, %f181;
mov.b32 %r470, %f182;
mov.u32 %r471, 8;
shfl.sync.bfly.b32 %r472|%p24, %r470, %r471, %r466, %r468;
mov.b32 %f183, %r472;
add.ftz.f32 %f184, %f182, %f183;
mov.b32 %r473, %f184;
mov.u32 %r474, 4;
shfl.sync.bfly.b32 %r475|%p25, %r473, %r474, %r466, %r468;
mov.b32 %f185, %r475;
add.ftz.f32 %f186, %f184, %f185;
mov.b32 %r476, %f186;
mov.u32 %r477, 2;
shfl.sync.bfly.b32 %r478|%p26, %r476, %r477, %r466, %r468;
mov.b32 %f187, %r478;
add.ftz.f32 %f188, %f186, %f187;
mov.b32 %r479, %f188;
mov.u32 %r480, 1;
shfl.sync.bfly.b32 %r481|%p27, %r479, %r480, %r466, %r468;
mov.b32 %f189, %r481;
add.ftz.f32 %f39, %f188, %f189;
@%p14 bra $L__BB68_12;

cvt.u64.u32 %rd77, %r31;
add.s64 %rd78, %rd77, %rd12;
cvta.to.global.u64 %rd79, %rd17;
shl.b64 %rd80, %rd78, 2;
add.s64 %rd81, %rd79, %rd80;
st.global.f32 [%rd81], %f39;

$L__BB68_12:
ld.shared.f32 %f190, [%r25+384];
add.ftz.f32 %f191, %f190, %f309;
mov.b32 %r482, %f191;
shfl.sync.bfly.b32 %r486|%p29, %r482, %r467, %r466, %r468;
mov.b32 %f192, %r486;
add.ftz.f32 %f193, %f191, %f192;
mov.b32 %r487, %f193;
shfl.sync.bfly.b32 %r489|%p30, %r487, %r471, %r466, %r468;
mov.b32 %f194, %r489;
add.ftz.f32 %f195, %f193, %f194;
mov.b32 %r490, %f195;
shfl.sync.bfly.b32 %r492|%p31, %r490, %r474, %r466, %r468;
mov.b32 %f196, %r492;
add.ftz.f32 %f197, %f195, %f196;
mov.b32 %r493, %f197;
shfl.sync.bfly.b32 %r495|%p32, %r493, %r477, %r466, %r468;
mov.b32 %f198, %r495;
add.ftz.f32 %f199, %f197, %f198;
mov.b32 %r496, %f199;
shfl.sync.bfly.b32 %r498|%p33, %r496, %r480, %r466, %r468;
mov.b32 %f200, %r498;
add.ftz.f32 %f40, %f199, %f200;
@%p22 bra $L__BB68_14;

add.s32 %r499, %r31, 1;
cvt.u64.u32 %rd82, %r499;
add.s64 %rd83, %rd82, %rd12;
cvta.to.global.u64 %rd84, %rd17;
shl.b64 %rd85, %rd83, 2;
add.s64 %rd86, %rd84, %rd85;
st.global.f32 [%rd86], %f40;

$L__BB68_14:
shl.b32 %r500, %r31, 1;
mov.u32 %r501, 1;
ld.shared.f32 %f201, [%r25+512];
add.ftz.f32 %f202, %f201, %f308;
mov.b32 %r502, %f202;
mov.u32 %r503, 31;
mov.u32 %r504, 16;
mov.u32 %r505, -1;
shfl.sync.bfly.b32 %r506|%p35, %r502, %r504, %r503, %r505;
mov.b32 %f203, %r506;
add.ftz.f32 %f204, %f202, %f203;
mov.b32 %r507, %f204;
mov.u32 %r508, 8;
shfl.sync.bfly.b32 %r509|%p36, %r507, %r508, %r503, %r505;
mov.b32 %f205, %r509;
add.ftz.f32 %f206, %f204, %f205;
mov.b32 %r510, %f206;
mov.u32 %r511, 4;
shfl.sync.bfly.b32 %r512|%p37, %r510, %r511, %r503, %r505;
mov.b32 %f207, %r512;
add.ftz.f32 %f208, %f206, %f207;
mov.b32 %r513, %f208;
mov.u32 %r514, 2;
shfl.sync.bfly.b32 %r515|%p38, %r513, %r514, %r503, %r505;
mov.b32 %f209, %r515;
add.ftz.f32 %f210, %f208, %f209;
mov.b32 %r516, %f210;
shfl.sync.bfly.b32 %r517|%p39, %r516, %r501, %r503, %r505;
mov.b32 %f211, %r517;
add.ftz.f32 %f41, %f210, %f211;
cvt.u64.u32 %rd87, %r500;
add.s64 %rd88, %rd87, %rd12;
cvta.to.global.u64 %rd89, %rd17;
shl.b64 %rd90, %rd88, 2;
add.s64 %rd13, %rd89, %rd90;
@%p14 bra $L__BB68_16;

st.global.f32 [%rd13], %f41;

$L__BB68_16:
ld.shared.f32 %f212, [%r25+640];
add.ftz.f32 %f213, %f212, %f307;
mov.b32 %r518, %f213;
shfl.sync.bfly.b32 %r522|%p41, %r518, %r504, %r503, %r505;
mov.b32 %f214, %r522;
add.ftz.f32 %f215, %f213, %f214;
mov.b32 %r523, %f215;
shfl.sync.bfly.b32 %r525|%p42, %r523, %r508, %r503, %r505;
mov.b32 %f216, %r525;
add.ftz.f32 %f217, %f215, %f216;
mov.b32 %r526, %f217;
shfl.sync.bfly.b32 %r528|%p43, %r526, %r511, %r503, %r505;
mov.b32 %f218, %r528;
add.ftz.f32 %f219, %f217, %f218;
mov.b32 %r529, %f219;
shfl.sync.bfly.b32 %r531|%p44, %r529, %r514, %r503, %r505;
mov.b32 %f220, %r531;
add.ftz.f32 %f221, %f219, %f220;
mov.b32 %r532, %f221;
shfl.sync.bfly.b32 %r534|%p45, %r532, %r501, %r503, %r505;
mov.b32 %f222, %r534;
add.ftz.f32 %f42, %f221, %f222;
@%p22 bra $L__BB68_18;

st.global.f32 [%rd13+4], %f42;

$L__BB68_18:
mul.lo.s32 %r26, %r31, 3;
ld.shared.f32 %f223, [%r25+768];
add.ftz.f32 %f224, %f223, %f306;
mov.b32 %r535, %f224;
mov.u32 %r536, 31;
mov.u32 %r537, 16;
mov.u32 %r538, -1;
shfl.sync.bfly.b32 %r539|%p47, %r535, %r537, %r536, %r538;
mov.b32 %f225, %r539;
add.ftz.f32 %f226, %f224, %f225;
mov.b32 %r540, %f226;
mov.u32 %r541, 8;
shfl.sync.bfly.b32 %r542|%p48, %r540, %r541, %r536, %r538;
mov.b32 %f227, %r542;
add.ftz.f32 %f228, %f226, %f227;
mov.b32 %r543, %f228;
mov.u32 %r544, 4;
shfl.sync.bfly.b32 %r545|%p49, %r543, %r544, %r536, %r538;
mov.b32 %f229, %r545;
add.ftz.f32 %f230, %f228, %f229;
mov.b32 %r546, %f230;
mov.u32 %r547, 2;
shfl.sync.bfly.b32 %r548|%p50, %r546, %r547, %r536, %r538;
mov.b32 %f231, %r548;
add.ftz.f32 %f232, %f230, %f231;
mov.b32 %r549, %f232;
mov.u32 %r550, 1;
shfl.sync.bfly.b32 %r551|%p51, %r549, %r550, %r536, %r538;
mov.b32 %f233, %r551;
add.ftz.f32 %f43, %f232, %f233;
@%p14 bra $L__BB68_20;

cvt.u64.u32 %rd91, %r26;
add.s64 %rd92, %rd91, %rd12;
shl.b64 %rd94, %rd92, 2;
add.s64 %rd95, %rd89, %rd94;
st.global.f32 [%rd95], %f43;

$L__BB68_20:
ld.shared.f32 %f234, [%r25+896];
add.ftz.f32 %f235, %f234, %f305;
mov.b32 %r552, %f235;
shfl.sync.bfly.b32 %r556|%p53, %r552, %r537, %r536, %r538;
mov.b32 %f236, %r556;
add.ftz.f32 %f237, %f235, %f236;
mov.b32 %r557, %f237;
shfl.sync.bfly.b32 %r559|%p54, %r557, %r541, %r536, %r538;
mov.b32 %f238, %r559;
add.ftz.f32 %f239, %f237, %f238;
mov.b32 %r560, %f239;
shfl.sync.bfly.b32 %r562|%p55, %r560, %r544, %r536, %r538;
mov.b32 %f240, %r562;
add.ftz.f32 %f241, %f239, %f240;
mov.b32 %r563, %f241;
shfl.sync.bfly.b32 %r565|%p56, %r563, %r547, %r536, %r538;
mov.b32 %f242, %r565;
add.ftz.f32 %f243, %f241, %f242;
mov.b32 %r566, %f243;
shfl.sync.bfly.b32 %r568|%p57, %r566, %r550, %r536, %r538;
mov.b32 %f244, %r568;
add.ftz.f32 %f44, %f243, %f244;
@%p22 bra $L__BB68_22;

add.s32 %r569, %r26, 1;
cvt.u64.u32 %rd96, %r569;
add.s64 %rd97, %rd96, %rd12;
shl.b64 %rd99, %rd97, 2;
add.s64 %rd100, %rd89, %rd99;
st.global.f32 [%rd100], %f44;

$L__BB68_22:
shl.b32 %r570, %r31, 2;
mov.u32 %r571, 2;
ld.shared.f32 %f245, [%r25+1024];
add.ftz.f32 %f246, %f245, %f304;
mov.b32 %r572, %f246;
mov.u32 %r573, 31;
mov.u32 %r574, 16;
mov.u32 %r575, -1;
shfl.sync.bfly.b32 %r576|%p59, %r572, %r574, %r573, %r575;
mov.b32 %f247, %r576;
add.ftz.f32 %f248, %f246, %f247;
mov.b32 %r577, %f248;
mov.u32 %r578, 8;
shfl.sync.bfly.b32 %r579|%p60, %r577, %r578, %r573, %r575;
mov.b32 %f249, %r579;
add.ftz.f32 %f250, %f248, %f249;
mov.b32 %r580, %f250;
mov.u32 %r581, 4;
shfl.sync.bfly.b32 %r582|%p61, %r580, %r581, %r573, %r575;
mov.b32 %f251, %r582;
add.ftz.f32 %f252, %f250, %f251;
mov.b32 %r583, %f252;
shfl.sync.bfly.b32 %r584|%p62, %r583, %r571, %r573, %r575;
mov.b32 %f253, %r584;
add.ftz.f32 %f254, %f252, %f253;
mov.b32 %r585, %f254;
mov.u32 %r586, 1;
shfl.sync.bfly.b32 %r587|%p63, %r585, %r586, %r573, %r575;
mov.b32 %f255, %r587;
add.ftz.f32 %f45, %f254, %f255;
cvt.u64.u32 %rd101, %r570;
add.s64 %rd102, %rd101, %rd12;
shl.b64 %rd104, %rd102, 2;
add.s64 %rd14, %rd89, %rd104;
@%p14 bra $L__BB68_24;

st.global.f32 [%rd14], %f45;

$L__BB68_24:
ld.shared.f32 %f256, [%r25+1152];
add.ftz.f32 %f257, %f256, %f303;
mov.b32 %r588, %f257;
shfl.sync.bfly.b32 %r592|%p65, %r588, %r574, %r573, %r575;
mov.b32 %f258, %r592;
add.ftz.f32 %f259, %f257, %f258;
mov.b32 %r593, %f259;
shfl.sync.bfly.b32 %r595|%p66, %r593, %r578, %r573, %r575;
mov.b32 %f260, %r595;
add.ftz.f32 %f261, %f259, %f260;
mov.b32 %r596, %f261;
shfl.sync.bfly.b32 %r598|%p67, %r596, %r581, %r573, %r575;
mov.b32 %f262, %r598;
add.ftz.f32 %f263, %f261, %f262;
mov.b32 %r599, %f263;
shfl.sync.bfly.b32 %r601|%p68, %r599, %r571, %r573, %r575;
mov.b32 %f264, %r601;
add.ftz.f32 %f265, %f263, %f264;
mov.b32 %r602, %f265;
shfl.sync.bfly.b32 %r604|%p69, %r602, %r586, %r573, %r575;
mov.b32 %f266, %r604;
add.ftz.f32 %f46, %f265, %f266;
@%p22 bra $L__BB68_26;

st.global.f32 [%rd14+4], %f46;

$L__BB68_26:
mul.lo.s32 %r27, %r31, 5;
ld.shared.f32 %f267, [%r25+1280];
add.ftz.f32 %f268, %f267, %f302;
mov.b32 %r605, %f268;
mov.u32 %r606, 31;
mov.u32 %r607, 16;
mov.u32 %r608, -1;
shfl.sync.bfly.b32 %r609|%p71, %r605, %r607, %r606, %r608;
mov.b32 %f269, %r609;
add.ftz.f32 %f270, %f268, %f269;
mov.b32 %r610, %f270;
mov.u32 %r611, 8;
shfl.sync.bfly.b32 %r612|%p72, %r610, %r611, %r606, %r608;
mov.b32 %f271, %r612;
add.ftz.f32 %f272, %f270, %f271;
mov.b32 %r613, %f272;
mov.u32 %r614, 4;
shfl.sync.bfly.b32 %r615|%p73, %r613, %r614, %r606, %r608;
mov.b32 %f273, %r615;
add.ftz.f32 %f274, %f272, %f273;
mov.b32 %r616, %f274;
mov.u32 %r617, 2;
shfl.sync.bfly.b32 %r618|%p74, %r616, %r617, %r606, %r608;
mov.b32 %f275, %r618;
add.ftz.f32 %f276, %f274, %f275;
mov.b32 %r619, %f276;
mov.u32 %r620, 1;
shfl.sync.bfly.b32 %r621|%p75, %r619, %r620, %r606, %r608;
mov.b32 %f277, %r621;
add.ftz.f32 %f47, %f276, %f277;
@%p14 bra $L__BB68_28;

cvt.u64.u32 %rd105, %r27;
add.s64 %rd106, %rd105, %rd12;
shl.b64 %rd108, %rd106, 2;
add.s64 %rd109, %rd89, %rd108;
st.global.f32 [%rd109], %f47;

$L__BB68_28:
ld.shared.f32 %f278, [%r25+1408];
add.ftz.f32 %f279, %f278, %f301;
mov.b32 %r622, %f279;
shfl.sync.bfly.b32 %r626|%p77, %r622, %r607, %r606, %r608;
mov.b32 %f280, %r626;
add.ftz.f32 %f281, %f279, %f280;
mov.b32 %r627, %f281;
shfl.sync.bfly.b32 %r629|%p78, %r627, %r611, %r606, %r608;
mov.b32 %f282, %r629;
add.ftz.f32 %f283, %f281, %f282;
mov.b32 %r630, %f283;
shfl.sync.bfly.b32 %r632|%p79, %r630, %r614, %r606, %r608;
mov.b32 %f284, %r632;
add.ftz.f32 %f285, %f283, %f284;
mov.b32 %r633, %f285;
shfl.sync.bfly.b32 %r635|%p80, %r633, %r617, %r606, %r608;
mov.b32 %f286, %r635;
add.ftz.f32 %f287, %f285, %f286;
mov.b32 %r636, %f287;
shfl.sync.bfly.b32 %r638|%p81, %r636, %r620, %r606, %r608;
mov.b32 %f288, %r638;
add.ftz.f32 %f48, %f287, %f288;
@%p22 bra $L__BB68_30;

add.s32 %r639, %r27, 1;
cvt.u64.u32 %rd110, %r639;
add.s64 %rd111, %rd110, %rd12;
shl.b64 %rd113, %rd111, 2;
add.s64 %rd114, %rd89, %rd113;
st.global.f32 [%rd114], %f48;

$L__BB68_30:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type6ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type6ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type6ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type6ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type6ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type6ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type6ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type6ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type6ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<95>;
.reg .b16 %rs<27>;
.reg .f32 %f<365>;
.reg .b32 %r<708>;
.reg .b64 %rd<125>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type6ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1792];

ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type6ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type6ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type6ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r47, [_Z13mul_mat_vec_qIL9ggml_type6ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type6ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type6ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type6ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type6ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type6ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type6ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type6ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r45, [_Z13mul_mat_vec_qIL9ggml_type6ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r46, [_Z13mul_mat_vec_qIL9ggml_type6ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type6ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type6ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type6ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type6ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r48, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r48, %r2;
mov.u32 %r49, %ctaid.x;
shl.b32 %r4, %r49, 1;
shr.u32 %r50, %r3, 31;
add.s32 %r51, %r3, %r50;
shr.s32 %r707, %r51, 1;
shr.u32 %r6, %r47, 5;
setp.ge.s32 %p3, %r707, %r6;
mov.f32 %f351, 0f00000000;
mov.f32 %f352, %f351;
mov.f32 %f353, %f351;
mov.f32 %f354, %f351;
mov.f32 %f355, %f351;
mov.f32 %f356, %f351;
mov.f32 %f357, %f351;
mov.f32 %f358, %f351;
mov.f32 %f359, %f351;
mov.f32 %f360, %f351;
mov.f32 %f361, %f351;
mov.f32 %f362, %f351;
mov.f32 %f363, %f351;
mov.f32 %f364, %f351;
@%p3 bra $L__BB69_3;

mov.u32 %r52, %ctaid.y;
mov.u32 %r53, %ctaid.z;
mul.lo.s32 %r54, %r52, %r39;
cvt.u64.u32 %rd20, %r54;
mul.lo.s32 %r55, %r53, %r45;
cvt.u64.u32 %rd21, %r55;
add.s64 %rd1, %rd21, %rd20;
mul.hi.u32 %r56, %r53, %r41;
add.s32 %r57, %r56, %r53;
shr.u32 %r58, %r57, %r42;
mul.lo.s32 %r59, %r58, %r44;
mul.hi.u32 %r60, %r52, %r35;
add.s32 %r61, %r60, %r52;
shr.u32 %r62, %r61, %r36;
mul.lo.s32 %r63, %r62, %r38;
cvta.to.global.u64 %rd124, %rd17;
and.b32 %r66, %r51, -2;
sub.s32 %r67, %r3, %r66;
shl.b32 %r68, %r67, 1;
shl.b32 %r69, %r67, 2;
or.b32 %r70, %r69, 1;
shl.b32 %r7, %r67, 3;
cvt.s64.s32 %rd3, %r68;
sub.s32 %r71, %r70, %r68;
shl.b32 %r8, %r71, 2;
cvt.s64.s32 %rd4, %r71;
shl.b32 %r72, %r33, 1;
add.s32 %r706, %r707, %r72;
mad.lo.s32 %r705, %r33, 3, %r707;
shl.b32 %r73, %r33, 2;
add.s32 %r704, %r707, %r73;
mad.lo.s32 %r703, %r33, 5, %r707;
mad.lo.s32 %r702, %r33, 6, %r707;
add.s32 %r701, %r707, %r33;
mad.lo.s32 %r74, %r4, %r32, %r707;
add.s32 %r75, %r74, %r59;
add.s32 %r76, %r75, %r63;
mul.wide.s32 %rd5, %r76, 22;
mul.wide.s32 %rd22, %r69, 2;
add.s64 %rd23, %rd5, %rd22;
add.s64 %rd6, %rd23, 8;
add.s32 %r77, %r707, %r59;
add.s32 %r78, %r77, %r63;
add.s32 %r79, %r4, 1;
mad.lo.s32 %r80, %r32, %r79, %r78;
mul.wide.s32 %rd7, %r80, 22;
add.s64 %rd24, %rd7, %rd22;
add.s64 %rd8, %rd24, 6;
cvta.to.global.u64 %rd9, %rd18;
mov.f32 %f351, 0f00000000;
shl.b64 %rd32, %rd3, 2;
shl.b64 %rd34, %rd4, 2;

$L__BB69_2:
cvt.u64.u32 %rd25, %r707;
add.s64 %rd26, %rd1, %rd25;
mul.lo.s64 %rd27, %rd26, 36;
add.s64 %rd28, %rd9, %rd27;
add.s64 %rd29, %rd28, 4;
add.s64 %rd30, %rd124, %rd5;
ld.global.nc.u16 %rs15, [%rd30+2];
ld.global.nc.u16 %rs16, [%rd30+4];
mov.b32 %r333, {%rs15, %rs16};
add.s64 %rd31, %rd124, %rd6;
ld.global.nc.u16 %rs17, [%rd31+-2];
ld.global.nc.u16 %rs18, [%rd31];
mov.b32 %r334, {%rs17, %rs18};
shr.s32 %r335, %r333, %r7;
add.s64 %rd33, %rd29, %rd32;
ld.global.nc.u32 %r83, [%rd33];
ld.global.nc.u32 %r87, [%rd33+16];
ld.global.nc.u16 %rs19, [%rd31+2];
ld.global.nc.u16 %rs20, [%rd31+4];
mov.b32 %r336, {%rs19, %rs20};
shr.s32 %r337, %r333, %r8;
add.s64 %rd35, %rd29, %rd34;
ld.global.nc.u32 %r91, [%rd35];
ld.global.nc.u32 %r95, [%rd33+20];
ld.global.nc.u16 %rs1, [%rd30];

{ cvt.f32.f16 %f85, %rs1;}


and.b32 %r338, %r334, 252645135;
shl.b32 %r339, %r335, 4;
and.b32 %r340, %r339, 16;
or.b32 %r341, %r340, %r338;
shl.b32 %r342, %r335, 11;
and.b32 %r343, %r342, 4096;
or.b32 %r344, %r341, %r343;
shl.b32 %r345, %r335, 18;
and.b32 %r346, %r345, 1048576;
or.b32 %r347, %r344, %r346;
shl.b32 %r348, %r335, 25;
and.b32 %r349, %r348, 268435456;
or.b32 %r298, %r347, %r349;
mov.u32 %r318, 0;

dp4a.s32.s32 %r81, %r298, %r83, %r318;

shr.u32 %r350, %r334, 4;
and.b32 %r351, %r350, 252645135;
shr.u32 %r352, %r335, 12;
and.b32 %r353, %r352, 16;
or.b32 %r354, %r353, %r351;
shr.u32 %r355, %r335, 5;
and.b32 %r356, %r355, 4096;
or.b32 %r357, %r354, %r356;
shl.b32 %r358, %r335, 2;
and.b32 %r359, %r358, 1048576;
or.b32 %r360, %r357, %r359;
shl.b32 %r361, %r335, 9;
and.b32 %r362, %r361, 268435456;
or.b32 %r302, %r360, %r362;

dp4a.s32.s32 %r85, %r302, %r87, %r81;

and.b32 %r363, %r336, 252645135;
shl.b32 %r364, %r337, 4;
and.b32 %r365, %r364, 16;
or.b32 %r366, %r365, %r363;
shl.b32 %r367, %r337, 11;
and.b32 %r368, %r367, 4096;
or.b32 %r369, %r366, %r368;
shl.b32 %r370, %r337, 18;
and.b32 %r371, %r370, 1048576;
or.b32 %r372, %r369, %r371;
shl.b32 %r373, %r337, 25;
and.b32 %r374, %r373, 268435456;
or.b32 %r306, %r372, %r374;

dp4a.s32.s32 %r89, %r306, %r91, %r85;

shr.u32 %r375, %r336, 4;
and.b32 %r376, %r375, 252645135;
shr.u32 %r377, %r337, 12;
and.b32 %r378, %r377, 16;
or.b32 %r379, %r378, %r376;
shr.u32 %r380, %r337, 5;
and.b32 %r381, %r380, 4096;
or.b32 %r382, %r379, %r381;
shl.b32 %r383, %r337, 2;
and.b32 %r384, %r383, 1048576;
or.b32 %r385, %r382, %r384;
shl.b32 %r386, %r337, 9;
and.b32 %r387, %r386, 268435456;
or.b32 %r310, %r385, %r387;

dp4a.s32.s32 %r93, %r310, %r95, %r89;

ld.global.nc.u32 %r97, [%rd28];

{.reg .f16 low,high;
mov.b32 {low,high},%r97;
cvt.f32.f16 %f86, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r97;
cvt.f32.f16 %f87, high;}


cvt.rn.f32.s32 %f127, %r93;
mul.ftz.f32 %f128, %f86, %f127;
mul.ftz.f32 %f129, %f87, 0f41000000;
sub.ftz.f32 %f130, %f128, %f129;
fma.rn.ftz.f32 %f364, %f85, %f130, %f364;
add.s64 %rd36, %rd124, %rd7;
ld.global.nc.u16 %rs21, [%rd36+2];
ld.global.nc.u16 %rs22, [%rd36+4];
mov.b32 %r388, {%rs21, %rs22};
add.s64 %rd37, %rd124, %rd8;
ld.global.nc.u16 %rs23, [%rd37];
ld.global.nc.u16 %rs24, [%rd37+2];
mov.b32 %r389, {%rs23, %rs24};
shr.s32 %r390, %r388, %r7;
ld.global.nc.u16 %rs25, [%rd37+4];
ld.global.nc.u16 %rs26, [%rd37+6];
mov.b32 %r391, {%rs25, %rs26};
shr.s32 %r392, %r388, %r8;
ld.global.nc.u16 %rs2, [%rd36];

{ cvt.f32.f16 %f88, %rs2;}


and.b32 %r393, %r389, 252645135;
shl.b32 %r394, %r390, 4;
and.b32 %r395, %r394, 16;
or.b32 %r396, %r395, %r393;
shl.b32 %r397, %r390, 11;
and.b32 %r398, %r397, 4096;
or.b32 %r399, %r396, %r398;
shl.b32 %r400, %r390, 18;
and.b32 %r401, %r400, 1048576;
or.b32 %r402, %r399, %r401;
shl.b32 %r403, %r390, 25;
and.b32 %r404, %r403, 268435456;
or.b32 %r316, %r402, %r404;

dp4a.s32.s32 %r99, %r316, %r83, %r318;

shr.u32 %r405, %r389, 4;
and.b32 %r406, %r405, 252645135;
shr.u32 %r407, %r390, 12;
and.b32 %r408, %r407, 16;
or.b32 %r409, %r408, %r406;
shr.u32 %r410, %r390, 5;
and.b32 %r411, %r410, 4096;
or.b32 %r412, %r409, %r411;
shl.b32 %r413, %r390, 2;
and.b32 %r414, %r413, 1048576;
or.b32 %r415, %r412, %r414;
shl.b32 %r416, %r390, 9;
and.b32 %r417, %r416, 268435456;
or.b32 %r320, %r415, %r417;

dp4a.s32.s32 %r103, %r320, %r87, %r99;

and.b32 %r418, %r391, 252645135;
shl.b32 %r419, %r392, 4;
and.b32 %r420, %r419, 16;
or.b32 %r421, %r420, %r418;
shl.b32 %r422, %r392, 11;
and.b32 %r423, %r422, 4096;
or.b32 %r424, %r421, %r423;
shl.b32 %r425, %r392, 18;
and.b32 %r426, %r425, 1048576;
or.b32 %r427, %r424, %r426;
shl.b32 %r428, %r392, 25;
and.b32 %r429, %r428, 268435456;
or.b32 %r324, %r427, %r429;

dp4a.s32.s32 %r107, %r324, %r91, %r103;

shr.u32 %r430, %r391, 4;
and.b32 %r431, %r430, 252645135;
shr.u32 %r432, %r392, 12;
and.b32 %r433, %r432, 16;
or.b32 %r434, %r433, %r431;
shr.u32 %r435, %r392, 5;
and.b32 %r436, %r435, 4096;
or.b32 %r437, %r434, %r436;
shl.b32 %r438, %r392, 2;
and.b32 %r439, %r438, 1048576;
or.b32 %r440, %r437, %r439;
shl.b32 %r441, %r392, 9;
and.b32 %r442, %r441, 268435456;
or.b32 %r328, %r440, %r442;

dp4a.s32.s32 %r111, %r328, %r95, %r107;


{.reg .f16 low,high;
mov.b32 {low,high},%r97;
cvt.f32.f16 %f89, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r97;
cvt.f32.f16 %f90, high;}


cvt.rn.f32.s32 %f131, %r111;
mul.ftz.f32 %f132, %f89, %f131;
mul.ftz.f32 %f133, %f90, 0f41000000;
sub.ftz.f32 %f134, %f132, %f133;
fma.rn.ftz.f32 %f363, %f88, %f134, %f363;
cvt.u64.u32 %rd38, %r701;
add.s64 %rd39, %rd1, %rd38;
mul.lo.s64 %rd40, %rd39, 36;
add.s64 %rd41, %rd9, %rd40;
add.s64 %rd42, %rd41, 4;
add.s64 %rd43, %rd42, %rd32;
ld.global.nc.u32 %r119, [%rd43];
ld.global.nc.u32 %r123, [%rd43+16];
add.s64 %rd44, %rd42, %rd34;
ld.global.nc.u32 %r127, [%rd44];
ld.global.nc.u32 %r131, [%rd43+20];

{ cvt.f32.f16 %f91, %rs1;}



dp4a.s32.s32 %r117, %r298, %r119, %r318;


dp4a.s32.s32 %r121, %r302, %r123, %r117;


dp4a.s32.s32 %r125, %r306, %r127, %r121;


dp4a.s32.s32 %r129, %r310, %r131, %r125;

ld.global.nc.u32 %r133, [%rd41];

{.reg .f16 low,high;
mov.b32 {low,high},%r133;
cvt.f32.f16 %f92, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r133;
cvt.f32.f16 %f93, high;}


cvt.rn.f32.s32 %f135, %r129;
mul.ftz.f32 %f136, %f92, %f135;
mul.ftz.f32 %f137, %f93, 0f41000000;
sub.ftz.f32 %f138, %f136, %f137;
fma.rn.ftz.f32 %f362, %f91, %f138, %f362;

{ cvt.f32.f16 %f94, %rs2;}



dp4a.s32.s32 %r135, %r316, %r119, %r318;


dp4a.s32.s32 %r139, %r320, %r123, %r135;


dp4a.s32.s32 %r143, %r324, %r127, %r139;


dp4a.s32.s32 %r147, %r328, %r131, %r143;


{.reg .f16 low,high;
mov.b32 {low,high},%r133;
cvt.f32.f16 %f95, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r133;
cvt.f32.f16 %f96, high;}


cvt.rn.f32.s32 %f139, %r147;
mul.ftz.f32 %f140, %f95, %f139;
mul.ftz.f32 %f141, %f96, 0f41000000;
sub.ftz.f32 %f142, %f140, %f141;
fma.rn.ftz.f32 %f361, %f94, %f142, %f361;
cvt.u64.u32 %rd45, %r706;
add.s64 %rd46, %rd1, %rd45;
mul.lo.s64 %rd47, %rd46, 36;
add.s64 %rd48, %rd9, %rd47;
add.s64 %rd49, %rd48, 4;
add.s64 %rd50, %rd49, %rd32;
ld.global.nc.u32 %r155, [%rd50];
ld.global.nc.u32 %r159, [%rd50+16];
add.s64 %rd51, %rd49, %rd34;
ld.global.nc.u32 %r163, [%rd51];
ld.global.nc.u32 %r167, [%rd50+20];

{ cvt.f32.f16 %f97, %rs1;}



dp4a.s32.s32 %r153, %r298, %r155, %r318;


dp4a.s32.s32 %r157, %r302, %r159, %r153;


dp4a.s32.s32 %r161, %r306, %r163, %r157;


dp4a.s32.s32 %r165, %r310, %r167, %r161;

ld.global.nc.u32 %r169, [%rd48];

{.reg .f16 low,high;
mov.b32 {low,high},%r169;
cvt.f32.f16 %f98, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r169;
cvt.f32.f16 %f99, high;}


cvt.rn.f32.s32 %f143, %r165;
mul.ftz.f32 %f144, %f98, %f143;
mul.ftz.f32 %f145, %f99, 0f41000000;
sub.ftz.f32 %f146, %f144, %f145;
fma.rn.ftz.f32 %f360, %f97, %f146, %f360;

{ cvt.f32.f16 %f100, %rs2;}



dp4a.s32.s32 %r171, %r316, %r155, %r318;


dp4a.s32.s32 %r175, %r320, %r159, %r171;


dp4a.s32.s32 %r179, %r324, %r163, %r175;


dp4a.s32.s32 %r183, %r328, %r167, %r179;


{.reg .f16 low,high;
mov.b32 {low,high},%r169;
cvt.f32.f16 %f101, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r169;
cvt.f32.f16 %f102, high;}


cvt.rn.f32.s32 %f147, %r183;
mul.ftz.f32 %f148, %f101, %f147;
mul.ftz.f32 %f149, %f102, 0f41000000;
sub.ftz.f32 %f150, %f148, %f149;
fma.rn.ftz.f32 %f359, %f100, %f150, %f359;
cvt.u64.u32 %rd52, %r705;
add.s64 %rd53, %rd1, %rd52;
mul.lo.s64 %rd54, %rd53, 36;
add.s64 %rd55, %rd9, %rd54;
add.s64 %rd56, %rd55, 4;
add.s64 %rd57, %rd56, %rd32;
ld.global.nc.u32 %r191, [%rd57];
ld.global.nc.u32 %r195, [%rd57+16];
add.s64 %rd58, %rd56, %rd34;
ld.global.nc.u32 %r199, [%rd58];
ld.global.nc.u32 %r203, [%rd57+20];

{ cvt.f32.f16 %f103, %rs1;}



dp4a.s32.s32 %r189, %r298, %r191, %r318;


dp4a.s32.s32 %r193, %r302, %r195, %r189;


dp4a.s32.s32 %r197, %r306, %r199, %r193;


dp4a.s32.s32 %r201, %r310, %r203, %r197;

ld.global.nc.u32 %r205, [%rd55];

{.reg .f16 low,high;
mov.b32 {low,high},%r205;
cvt.f32.f16 %f104, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r205;
cvt.f32.f16 %f105, high;}


cvt.rn.f32.s32 %f151, %r201;
mul.ftz.f32 %f152, %f104, %f151;
mul.ftz.f32 %f153, %f105, 0f41000000;
sub.ftz.f32 %f154, %f152, %f153;
fma.rn.ftz.f32 %f358, %f103, %f154, %f358;

{ cvt.f32.f16 %f106, %rs2;}



dp4a.s32.s32 %r207, %r316, %r191, %r318;


dp4a.s32.s32 %r211, %r320, %r195, %r207;


dp4a.s32.s32 %r215, %r324, %r199, %r211;


dp4a.s32.s32 %r219, %r328, %r203, %r215;


{.reg .f16 low,high;
mov.b32 {low,high},%r205;
cvt.f32.f16 %f107, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r205;
cvt.f32.f16 %f108, high;}


cvt.rn.f32.s32 %f155, %r219;
mul.ftz.f32 %f156, %f107, %f155;
mul.ftz.f32 %f157, %f108, 0f41000000;
sub.ftz.f32 %f158, %f156, %f157;
fma.rn.ftz.f32 %f357, %f106, %f158, %f357;
cvt.u64.u32 %rd59, %r704;
add.s64 %rd60, %rd1, %rd59;
mul.lo.s64 %rd61, %rd60, 36;
add.s64 %rd62, %rd9, %rd61;
add.s64 %rd63, %rd62, 4;
add.s64 %rd64, %rd63, %rd32;
ld.global.nc.u32 %r227, [%rd64];
ld.global.nc.u32 %r231, [%rd64+16];
add.s64 %rd65, %rd63, %rd34;
ld.global.nc.u32 %r235, [%rd65];
ld.global.nc.u32 %r239, [%rd64+20];

{ cvt.f32.f16 %f109, %rs1;}



dp4a.s32.s32 %r225, %r298, %r227, %r318;


dp4a.s32.s32 %r229, %r302, %r231, %r225;


dp4a.s32.s32 %r233, %r306, %r235, %r229;


dp4a.s32.s32 %r237, %r310, %r239, %r233;

ld.global.nc.u32 %r241, [%rd62];

{.reg .f16 low,high;
mov.b32 {low,high},%r241;
cvt.f32.f16 %f110, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r241;
cvt.f32.f16 %f111, high;}


cvt.rn.f32.s32 %f159, %r237;
mul.ftz.f32 %f160, %f110, %f159;
mul.ftz.f32 %f161, %f111, 0f41000000;
sub.ftz.f32 %f162, %f160, %f161;
fma.rn.ftz.f32 %f356, %f109, %f162, %f356;

{ cvt.f32.f16 %f112, %rs2;}



dp4a.s32.s32 %r243, %r316, %r227, %r318;


dp4a.s32.s32 %r247, %r320, %r231, %r243;


dp4a.s32.s32 %r251, %r324, %r235, %r247;


dp4a.s32.s32 %r255, %r328, %r239, %r251;


{.reg .f16 low,high;
mov.b32 {low,high},%r241;
cvt.f32.f16 %f113, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r241;
cvt.f32.f16 %f114, high;}


cvt.rn.f32.s32 %f163, %r255;
mul.ftz.f32 %f164, %f113, %f163;
mul.ftz.f32 %f165, %f114, 0f41000000;
sub.ftz.f32 %f166, %f164, %f165;
fma.rn.ftz.f32 %f355, %f112, %f166, %f355;
cvt.u64.u32 %rd66, %r703;
add.s64 %rd67, %rd1, %rd66;
mul.lo.s64 %rd68, %rd67, 36;
add.s64 %rd69, %rd9, %rd68;
add.s64 %rd70, %rd69, 4;
add.s64 %rd71, %rd70, %rd32;
ld.global.nc.u32 %r263, [%rd71];
ld.global.nc.u32 %r267, [%rd71+16];
add.s64 %rd72, %rd70, %rd34;
ld.global.nc.u32 %r271, [%rd72];
ld.global.nc.u32 %r275, [%rd71+20];

{ cvt.f32.f16 %f115, %rs1;}



dp4a.s32.s32 %r261, %r298, %r263, %r318;


dp4a.s32.s32 %r265, %r302, %r267, %r261;


dp4a.s32.s32 %r269, %r306, %r271, %r265;


dp4a.s32.s32 %r273, %r310, %r275, %r269;

ld.global.nc.u32 %r277, [%rd69];

{.reg .f16 low,high;
mov.b32 {low,high},%r277;
cvt.f32.f16 %f116, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r277;
cvt.f32.f16 %f117, high;}


cvt.rn.f32.s32 %f167, %r273;
mul.ftz.f32 %f168, %f116, %f167;
mul.ftz.f32 %f169, %f117, 0f41000000;
sub.ftz.f32 %f170, %f168, %f169;
fma.rn.ftz.f32 %f354, %f115, %f170, %f354;

{ cvt.f32.f16 %f118, %rs2;}



dp4a.s32.s32 %r279, %r316, %r263, %r318;


dp4a.s32.s32 %r283, %r320, %r267, %r279;


dp4a.s32.s32 %r287, %r324, %r271, %r283;


dp4a.s32.s32 %r291, %r328, %r275, %r287;


{.reg .f16 low,high;
mov.b32 {low,high},%r277;
cvt.f32.f16 %f119, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r277;
cvt.f32.f16 %f120, high;}


cvt.rn.f32.s32 %f171, %r291;
mul.ftz.f32 %f172, %f119, %f171;
mul.ftz.f32 %f173, %f120, 0f41000000;
sub.ftz.f32 %f174, %f172, %f173;
fma.rn.ftz.f32 %f353, %f118, %f174, %f353;
cvt.u64.u32 %rd73, %r702;
add.s64 %rd74, %rd1, %rd73;
mul.lo.s64 %rd75, %rd74, 36;
add.s64 %rd76, %rd9, %rd75;
add.s64 %rd77, %rd76, 4;
add.s64 %rd78, %rd77, %rd32;
ld.global.nc.u32 %r299, [%rd78];
ld.global.nc.u32 %r303, [%rd78+16];
add.s64 %rd79, %rd77, %rd34;
ld.global.nc.u32 %r307, [%rd79];
ld.global.nc.u32 %r311, [%rd78+20];

{ cvt.f32.f16 %f121, %rs1;}



dp4a.s32.s32 %r297, %r298, %r299, %r318;


dp4a.s32.s32 %r301, %r302, %r303, %r297;


dp4a.s32.s32 %r305, %r306, %r307, %r301;


dp4a.s32.s32 %r309, %r310, %r311, %r305;

ld.global.nc.u32 %r313, [%rd76];

{.reg .f16 low,high;
mov.b32 {low,high},%r313;
cvt.f32.f16 %f122, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r313;
cvt.f32.f16 %f123, high;}


cvt.rn.f32.s32 %f175, %r309;
mul.ftz.f32 %f176, %f122, %f175;
mul.ftz.f32 %f177, %f123, 0f41000000;
sub.ftz.f32 %f178, %f176, %f177;
fma.rn.ftz.f32 %f352, %f121, %f178, %f352;

{ cvt.f32.f16 %f124, %rs2;}



dp4a.s32.s32 %r315, %r316, %r299, %r318;


dp4a.s32.s32 %r319, %r320, %r303, %r315;


dp4a.s32.s32 %r323, %r324, %r307, %r319;


dp4a.s32.s32 %r327, %r328, %r311, %r323;


{.reg .f16 low,high;
mov.b32 {low,high},%r313;
cvt.f32.f16 %f125, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r313;
cvt.f32.f16 %f126, high;}


cvt.rn.f32.s32 %f179, %r327;
mul.ftz.f32 %f180, %f125, %f179;
mul.ftz.f32 %f181, %f126, 0f41000000;
sub.ftz.f32 %f182, %f180, %f181;
fma.rn.ftz.f32 %f351, %f124, %f182, %f351;
add.s32 %r706, %r706, 32;
add.s32 %r705, %r705, 32;
add.s32 %r704, %r704, 32;
add.s32 %r703, %r703, 32;
add.s32 %r702, %r702, 32;
add.s32 %r701, %r701, 32;
add.s64 %rd124, %rd124, 704;
add.s32 %r707, %r707, 32;
setp.lt.s32 %p4, %r707, %r6;
@%p4 bra $L__BB69_2;

$L__BB69_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB69_5;

mov.u32 %r443, _ZZ13mul_mat_vec_qIL9ggml_type6ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r444, %r1, 1792, %r443;
shl.b32 %r445, %r2, 2;
add.s32 %r446, %r444, %r445;
st.shared.f32 [%r446+-1792], %f364;
st.shared.f32 [%r446+-1664], %f363;
st.shared.f32 [%r446+-1536], %f362;
st.shared.f32 [%r446+-1408], %f361;
st.shared.f32 [%r446+-1280], %f360;
st.shared.f32 [%r446+-1152], %f359;
st.shared.f32 [%r446+-1024], %f358;
st.shared.f32 [%r446+-896], %f357;
st.shared.f32 [%r446+-768], %f356;
st.shared.f32 [%r446+-640], %f355;
st.shared.f32 [%r446+-512], %f354;
st.shared.f32 [%r446+-384], %f353;
st.shared.f32 [%r446+-256], %f352;
st.shared.f32 [%r446+-128], %f351;

$L__BB69_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB69_34;

mov.u32 %r447, %ctaid.z;
mov.u32 %r448, %ctaid.y;
mad.lo.s32 %r449, %r448, %r40, %r4;
mad.lo.s32 %r450, %r447, %r46, %r449;
cvt.u64.u32 %rd12, %r450;
shl.b32 %r451, %r2, 2;
mov.u32 %r452, _ZZ13mul_mat_vec_qIL9ggml_type6ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r29, %r452, %r451;
mov.u32 %r453, 2;
ld.shared.f32 %f183, [%r29];
add.ftz.f32 %f184, %f183, %f364;
mov.b32 %r454, %f184;
mov.u32 %r455, 31;
mov.u32 %r456, 16;
mov.u32 %r457, -1;
shfl.sync.bfly.b32 %r458|%p7, %r454, %r456, %r455, %r457;
mov.b32 %f185, %r458;
add.ftz.f32 %f186, %f184, %f185;
mov.b32 %r459, %f186;
mov.u32 %r460, 8;
shfl.sync.bfly.b32 %r461|%p8, %r459, %r460, %r455, %r457;
mov.b32 %f187, %r461;
add.ftz.f32 %f188, %f186, %f187;
mov.b32 %r462, %f188;
mov.u32 %r463, 4;
shfl.sync.bfly.b32 %r464|%p9, %r462, %r463, %r455, %r457;
mov.b32 %f189, %r464;
add.ftz.f32 %f190, %f188, %f189;
mov.b32 %r465, %f190;
shfl.sync.bfly.b32 %r466|%p10, %r465, %r453, %r455, %r457;
mov.b32 %f191, %r466;
add.ftz.f32 %f192, %f190, %f191;
mov.b32 %r467, %f192;
mov.u32 %r468, 1;
shfl.sync.bfly.b32 %r469|%p11, %r467, %r468, %r455, %r457;
mov.b32 %f193, %r469;
add.ftz.f32 %f43, %f192, %f193;
setp.lt.u32 %p12, %r4, %r34;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
cvta.to.global.u64 %rd80, %rd19;
mul.wide.u32 %rd81, %r450, 4;
add.s64 %rd13, %rd80, %rd81;
not.pred %p14, %p1;
@%p14 bra $L__BB69_8;

st.global.f32 [%rd13], %f43;

$L__BB69_8:
ld.shared.f32 %f194, [%r29+128];
add.ftz.f32 %f195, %f194, %f363;
mov.b32 %r470, %f195;
shfl.sync.bfly.b32 %r474|%p15, %r470, %r456, %r455, %r457;
mov.b32 %f196, %r474;
add.ftz.f32 %f197, %f195, %f196;
mov.b32 %r475, %f197;
shfl.sync.bfly.b32 %r477|%p16, %r475, %r460, %r455, %r457;
mov.b32 %f198, %r477;
add.ftz.f32 %f199, %f197, %f198;
mov.b32 %r478, %f199;
shfl.sync.bfly.b32 %r480|%p17, %r478, %r463, %r455, %r457;
mov.b32 %f200, %r480;
add.ftz.f32 %f201, %f199, %f200;
mov.b32 %r481, %f201;
shfl.sync.bfly.b32 %r483|%p18, %r481, %r453, %r455, %r457;
mov.b32 %f202, %r483;
add.ftz.f32 %f203, %f201, %f202;
mov.b32 %r484, %f203;
shfl.sync.bfly.b32 %r486|%p19, %r484, %r468, %r455, %r457;
mov.b32 %f204, %r486;
add.ftz.f32 %f44, %f203, %f204;
setp.eq.s32 %p20, %r2, 1;
or.b32 %r490, %r4, 1;
setp.lt.u32 %p21, %r490, %r34;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB69_10;

st.global.f32 [%rd13+4], %f44;

$L__BB69_10:
ld.shared.f32 %f205, [%r29+256];
add.ftz.f32 %f206, %f205, %f362;
mov.b32 %r491, %f206;
mov.u32 %r492, 31;
mov.u32 %r493, 16;
mov.u32 %r494, -1;
shfl.sync.bfly.b32 %r495|%p23, %r491, %r493, %r492, %r494;
mov.b32 %f207, %r495;
add.ftz.f32 %f208, %f206, %f207;
mov.b32 %r496, %f208;
mov.u32 %r497, 8;
shfl.sync.bfly.b32 %r498|%p24, %r496, %r497, %r492, %r494;
mov.b32 %f209, %r498;
add.ftz.f32 %f210, %f208, %f209;
mov.b32 %r499, %f210;
mov.u32 %r500, 4;
shfl.sync.bfly.b32 %r501|%p25, %r499, %r500, %r492, %r494;
mov.b32 %f211, %r501;
add.ftz.f32 %f212, %f210, %f211;
mov.b32 %r502, %f212;
mov.u32 %r503, 2;
shfl.sync.bfly.b32 %r504|%p26, %r502, %r503, %r492, %r494;
mov.b32 %f213, %r504;
add.ftz.f32 %f214, %f212, %f213;
mov.b32 %r505, %f214;
mov.u32 %r506, 1;
shfl.sync.bfly.b32 %r507|%p27, %r505, %r506, %r492, %r494;
mov.b32 %f215, %r507;
add.ftz.f32 %f45, %f214, %f215;
@%p14 bra $L__BB69_12;

cvt.u64.u32 %rd82, %r34;
add.s64 %rd83, %rd82, %rd12;
shl.b64 %rd85, %rd83, 2;
add.s64 %rd86, %rd80, %rd85;
st.global.f32 [%rd86], %f45;

$L__BB69_12:
ld.shared.f32 %f216, [%r29+384];
add.ftz.f32 %f217, %f216, %f361;
mov.b32 %r508, %f217;
shfl.sync.bfly.b32 %r512|%p29, %r508, %r493, %r492, %r494;
mov.b32 %f218, %r512;
add.ftz.f32 %f219, %f217, %f218;
mov.b32 %r513, %f219;
shfl.sync.bfly.b32 %r515|%p30, %r513, %r497, %r492, %r494;
mov.b32 %f220, %r515;
add.ftz.f32 %f221, %f219, %f220;
mov.b32 %r516, %f221;
shfl.sync.bfly.b32 %r518|%p31, %r516, %r500, %r492, %r494;
mov.b32 %f222, %r518;
add.ftz.f32 %f223, %f221, %f222;
mov.b32 %r519, %f223;
shfl.sync.bfly.b32 %r521|%p32, %r519, %r503, %r492, %r494;
mov.b32 %f224, %r521;
add.ftz.f32 %f225, %f223, %f224;
mov.b32 %r522, %f225;
shfl.sync.bfly.b32 %r524|%p33, %r522, %r506, %r492, %r494;
mov.b32 %f226, %r524;
add.ftz.f32 %f46, %f225, %f226;
@%p22 bra $L__BB69_14;

add.s32 %r525, %r34, 1;
cvt.u64.u32 %rd87, %r525;
add.s64 %rd88, %rd87, %rd12;
shl.b64 %rd90, %rd88, 2;
add.s64 %rd91, %rd80, %rd90;
st.global.f32 [%rd91], %f46;

$L__BB69_14:
shl.b32 %r526, %r34, 1;
mov.u32 %r527, 1;
ld.shared.f32 %f227, [%r29+512];
add.ftz.f32 %f228, %f227, %f360;
mov.b32 %r528, %f228;
mov.u32 %r529, 31;
mov.u32 %r530, 16;
mov.u32 %r531, -1;
shfl.sync.bfly.b32 %r532|%p35, %r528, %r530, %r529, %r531;
mov.b32 %f229, %r532;
add.ftz.f32 %f230, %f228, %f229;
mov.b32 %r533, %f230;
mov.u32 %r534, 8;
shfl.sync.bfly.b32 %r535|%p36, %r533, %r534, %r529, %r531;
mov.b32 %f231, %r535;
add.ftz.f32 %f232, %f230, %f231;
mov.b32 %r536, %f232;
mov.u32 %r537, 4;
shfl.sync.bfly.b32 %r538|%p37, %r536, %r537, %r529, %r531;
mov.b32 %f233, %r538;
add.ftz.f32 %f234, %f232, %f233;
mov.b32 %r539, %f234;
mov.u32 %r540, 2;
shfl.sync.bfly.b32 %r541|%p38, %r539, %r540, %r529, %r531;
mov.b32 %f235, %r541;
add.ftz.f32 %f236, %f234, %f235;
mov.b32 %r542, %f236;
shfl.sync.bfly.b32 %r543|%p39, %r542, %r527, %r529, %r531;
mov.b32 %f237, %r543;
add.ftz.f32 %f47, %f236, %f237;
cvt.u64.u32 %rd92, %r526;
add.s64 %rd93, %rd92, %rd12;
shl.b64 %rd95, %rd93, 2;
add.s64 %rd14, %rd80, %rd95;
@%p14 bra $L__BB69_16;

st.global.f32 [%rd14], %f47;

$L__BB69_16:
ld.shared.f32 %f238, [%r29+640];
add.ftz.f32 %f239, %f238, %f359;
mov.b32 %r544, %f239;
shfl.sync.bfly.b32 %r548|%p41, %r544, %r530, %r529, %r531;
mov.b32 %f240, %r548;
add.ftz.f32 %f241, %f239, %f240;
mov.b32 %r549, %f241;
shfl.sync.bfly.b32 %r551|%p42, %r549, %r534, %r529, %r531;
mov.b32 %f242, %r551;
add.ftz.f32 %f243, %f241, %f242;
mov.b32 %r552, %f243;
shfl.sync.bfly.b32 %r554|%p43, %r552, %r537, %r529, %r531;
mov.b32 %f244, %r554;
add.ftz.f32 %f245, %f243, %f244;
mov.b32 %r555, %f245;
shfl.sync.bfly.b32 %r557|%p44, %r555, %r540, %r529, %r531;
mov.b32 %f246, %r557;
add.ftz.f32 %f247, %f245, %f246;
mov.b32 %r558, %f247;
shfl.sync.bfly.b32 %r560|%p45, %r558, %r527, %r529, %r531;
mov.b32 %f248, %r560;
add.ftz.f32 %f48, %f247, %f248;
@%p22 bra $L__BB69_18;

st.global.f32 [%rd14+4], %f48;

$L__BB69_18:
mul.lo.s32 %r30, %r34, 3;
ld.shared.f32 %f249, [%r29+768];
add.ftz.f32 %f250, %f249, %f358;
mov.b32 %r561, %f250;
mov.u32 %r562, 31;
mov.u32 %r563, 16;
mov.u32 %r564, -1;
shfl.sync.bfly.b32 %r565|%p47, %r561, %r563, %r562, %r564;
mov.b32 %f251, %r565;
add.ftz.f32 %f252, %f250, %f251;
mov.b32 %r566, %f252;
mov.u32 %r567, 8;
shfl.sync.bfly.b32 %r568|%p48, %r566, %r567, %r562, %r564;
mov.b32 %f253, %r568;
add.ftz.f32 %f254, %f252, %f253;
mov.b32 %r569, %f254;
mov.u32 %r570, 4;
shfl.sync.bfly.b32 %r571|%p49, %r569, %r570, %r562, %r564;
mov.b32 %f255, %r571;
add.ftz.f32 %f256, %f254, %f255;
mov.b32 %r572, %f256;
mov.u32 %r573, 2;
shfl.sync.bfly.b32 %r574|%p50, %r572, %r573, %r562, %r564;
mov.b32 %f257, %r574;
add.ftz.f32 %f258, %f256, %f257;
mov.b32 %r575, %f258;
mov.u32 %r576, 1;
shfl.sync.bfly.b32 %r577|%p51, %r575, %r576, %r562, %r564;
mov.b32 %f259, %r577;
add.ftz.f32 %f49, %f258, %f259;
@%p14 bra $L__BB69_20;

cvt.u64.u32 %rd96, %r30;
add.s64 %rd97, %rd96, %rd12;
shl.b64 %rd99, %rd97, 2;
add.s64 %rd100, %rd80, %rd99;
st.global.f32 [%rd100], %f49;

$L__BB69_20:
ld.shared.f32 %f260, [%r29+896];
add.ftz.f32 %f261, %f260, %f357;
mov.b32 %r578, %f261;
shfl.sync.bfly.b32 %r582|%p53, %r578, %r563, %r562, %r564;
mov.b32 %f262, %r582;
add.ftz.f32 %f263, %f261, %f262;
mov.b32 %r583, %f263;
shfl.sync.bfly.b32 %r585|%p54, %r583, %r567, %r562, %r564;
mov.b32 %f264, %r585;
add.ftz.f32 %f265, %f263, %f264;
mov.b32 %r586, %f265;
shfl.sync.bfly.b32 %r588|%p55, %r586, %r570, %r562, %r564;
mov.b32 %f266, %r588;
add.ftz.f32 %f267, %f265, %f266;
mov.b32 %r589, %f267;
shfl.sync.bfly.b32 %r591|%p56, %r589, %r573, %r562, %r564;
mov.b32 %f268, %r591;
add.ftz.f32 %f269, %f267, %f268;
mov.b32 %r592, %f269;
shfl.sync.bfly.b32 %r594|%p57, %r592, %r576, %r562, %r564;
mov.b32 %f270, %r594;
add.ftz.f32 %f50, %f269, %f270;
@%p22 bra $L__BB69_22;

add.s32 %r595, %r30, 1;
cvt.u64.u32 %rd101, %r595;
add.s64 %rd102, %rd101, %rd12;
shl.b64 %rd104, %rd102, 2;
add.s64 %rd105, %rd80, %rd104;
st.global.f32 [%rd105], %f50;

$L__BB69_22:
shl.b32 %r596, %r34, 2;
mov.u32 %r597, 2;
ld.shared.f32 %f271, [%r29+1024];
add.ftz.f32 %f272, %f271, %f356;
mov.b32 %r598, %f272;
mov.u32 %r599, 31;
mov.u32 %r600, 16;
mov.u32 %r601, -1;
shfl.sync.bfly.b32 %r602|%p59, %r598, %r600, %r599, %r601;
mov.b32 %f273, %r602;
add.ftz.f32 %f274, %f272, %f273;
mov.b32 %r603, %f274;
mov.u32 %r604, 8;
shfl.sync.bfly.b32 %r605|%p60, %r603, %r604, %r599, %r601;
mov.b32 %f275, %r605;
add.ftz.f32 %f276, %f274, %f275;
mov.b32 %r606, %f276;
mov.u32 %r607, 4;
shfl.sync.bfly.b32 %r608|%p61, %r606, %r607, %r599, %r601;
mov.b32 %f277, %r608;
add.ftz.f32 %f278, %f276, %f277;
mov.b32 %r609, %f278;
shfl.sync.bfly.b32 %r610|%p62, %r609, %r597, %r599, %r601;
mov.b32 %f279, %r610;
add.ftz.f32 %f280, %f278, %f279;
mov.b32 %r611, %f280;
mov.u32 %r612, 1;
shfl.sync.bfly.b32 %r613|%p63, %r611, %r612, %r599, %r601;
mov.b32 %f281, %r613;
add.ftz.f32 %f51, %f280, %f281;
cvt.u64.u32 %rd106, %r596;
add.s64 %rd107, %rd106, %rd12;
shl.b64 %rd109, %rd107, 2;
add.s64 %rd15, %rd80, %rd109;
@%p14 bra $L__BB69_24;

st.global.f32 [%rd15], %f51;

$L__BB69_24:
ld.shared.f32 %f282, [%r29+1152];
add.ftz.f32 %f283, %f282, %f355;
mov.b32 %r614, %f283;
shfl.sync.bfly.b32 %r618|%p65, %r614, %r600, %r599, %r601;
mov.b32 %f284, %r618;
add.ftz.f32 %f285, %f283, %f284;
mov.b32 %r619, %f285;
shfl.sync.bfly.b32 %r621|%p66, %r619, %r604, %r599, %r601;
mov.b32 %f286, %r621;
add.ftz.f32 %f287, %f285, %f286;
mov.b32 %r622, %f287;
shfl.sync.bfly.b32 %r624|%p67, %r622, %r607, %r599, %r601;
mov.b32 %f288, %r624;
add.ftz.f32 %f289, %f287, %f288;
mov.b32 %r625, %f289;
shfl.sync.bfly.b32 %r627|%p68, %r625, %r597, %r599, %r601;
mov.b32 %f290, %r627;
add.ftz.f32 %f291, %f289, %f290;
mov.b32 %r628, %f291;
shfl.sync.bfly.b32 %r630|%p69, %r628, %r612, %r599, %r601;
mov.b32 %f292, %r630;
add.ftz.f32 %f52, %f291, %f292;
@%p22 bra $L__BB69_26;

st.global.f32 [%rd15+4], %f52;

$L__BB69_26:
mul.lo.s32 %r31, %r34, 5;
ld.shared.f32 %f293, [%r29+1280];
add.ftz.f32 %f294, %f293, %f354;
mov.b32 %r631, %f294;
mov.u32 %r632, 31;
mov.u32 %r633, 16;
mov.u32 %r634, -1;
shfl.sync.bfly.b32 %r635|%p71, %r631, %r633, %r632, %r634;
mov.b32 %f295, %r635;
add.ftz.f32 %f296, %f294, %f295;
mov.b32 %r636, %f296;
mov.u32 %r637, 8;
shfl.sync.bfly.b32 %r638|%p72, %r636, %r637, %r632, %r634;
mov.b32 %f297, %r638;
add.ftz.f32 %f298, %f296, %f297;
mov.b32 %r639, %f298;
mov.u32 %r640, 4;
shfl.sync.bfly.b32 %r641|%p73, %r639, %r640, %r632, %r634;
mov.b32 %f299, %r641;
add.ftz.f32 %f300, %f298, %f299;
mov.b32 %r642, %f300;
mov.u32 %r643, 2;
shfl.sync.bfly.b32 %r644|%p74, %r642, %r643, %r632, %r634;
mov.b32 %f301, %r644;
add.ftz.f32 %f302, %f300, %f301;
mov.b32 %r645, %f302;
mov.u32 %r646, 1;
shfl.sync.bfly.b32 %r647|%p75, %r645, %r646, %r632, %r634;
mov.b32 %f303, %r647;
add.ftz.f32 %f53, %f302, %f303;
@%p14 bra $L__BB69_28;

cvt.u64.u32 %rd110, %r31;
add.s64 %rd111, %rd110, %rd12;
shl.b64 %rd113, %rd111, 2;
add.s64 %rd114, %rd80, %rd113;
st.global.f32 [%rd114], %f53;

$L__BB69_28:
ld.shared.f32 %f304, [%r29+1408];
add.ftz.f32 %f305, %f304, %f353;
mov.b32 %r648, %f305;
shfl.sync.bfly.b32 %r652|%p77, %r648, %r633, %r632, %r634;
mov.b32 %f306, %r652;
add.ftz.f32 %f307, %f305, %f306;
mov.b32 %r653, %f307;
shfl.sync.bfly.b32 %r655|%p78, %r653, %r637, %r632, %r634;
mov.b32 %f308, %r655;
add.ftz.f32 %f309, %f307, %f308;
mov.b32 %r656, %f309;
shfl.sync.bfly.b32 %r658|%p79, %r656, %r640, %r632, %r634;
mov.b32 %f310, %r658;
add.ftz.f32 %f311, %f309, %f310;
mov.b32 %r659, %f311;
shfl.sync.bfly.b32 %r661|%p80, %r659, %r643, %r632, %r634;
mov.b32 %f312, %r661;
add.ftz.f32 %f313, %f311, %f312;
mov.b32 %r662, %f313;
shfl.sync.bfly.b32 %r664|%p81, %r662, %r646, %r632, %r634;
mov.b32 %f314, %r664;
add.ftz.f32 %f54, %f313, %f314;
@%p22 bra $L__BB69_30;

add.s32 %r665, %r31, 1;
cvt.u64.u32 %rd115, %r665;
add.s64 %rd116, %rd115, %rd12;
shl.b64 %rd118, %rd116, 2;
add.s64 %rd119, %rd80, %rd118;
st.global.f32 [%rd119], %f54;

$L__BB69_30:
mul.lo.s32 %r666, %r34, 6;
ld.shared.f32 %f315, [%r29+1536];
add.ftz.f32 %f316, %f315, %f352;
mov.b32 %r667, %f316;
mov.u32 %r668, 31;
mov.u32 %r669, 16;
mov.u32 %r670, -1;
shfl.sync.bfly.b32 %r671|%p83, %r667, %r669, %r668, %r670;
mov.b32 %f317, %r671;
add.ftz.f32 %f318, %f316, %f317;
mov.b32 %r672, %f318;
mov.u32 %r673, 8;
shfl.sync.bfly.b32 %r674|%p84, %r672, %r673, %r668, %r670;
mov.b32 %f319, %r674;
add.ftz.f32 %f320, %f318, %f319;
mov.b32 %r675, %f320;
mov.u32 %r676, 4;
shfl.sync.bfly.b32 %r677|%p85, %r675, %r676, %r668, %r670;
mov.b32 %f321, %r677;
add.ftz.f32 %f322, %f320, %f321;
mov.b32 %r678, %f322;
mov.u32 %r679, 2;
shfl.sync.bfly.b32 %r680|%p86, %r678, %r679, %r668, %r670;
mov.b32 %f323, %r680;
add.ftz.f32 %f324, %f322, %f323;
mov.b32 %r681, %f324;
mov.u32 %r682, 1;
shfl.sync.bfly.b32 %r683|%p87, %r681, %r682, %r668, %r670;
mov.b32 %f325, %r683;
add.ftz.f32 %f55, %f324, %f325;
cvt.u64.u32 %rd120, %r666;
add.s64 %rd121, %rd120, %rd12;
shl.b64 %rd123, %rd121, 2;
add.s64 %rd16, %rd80, %rd123;
@%p14 bra $L__BB69_32;

st.global.f32 [%rd16], %f55;

$L__BB69_32:
ld.shared.f32 %f326, [%r29+1664];
add.ftz.f32 %f327, %f326, %f351;
mov.b32 %r684, %f327;
shfl.sync.bfly.b32 %r688|%p89, %r684, %r669, %r668, %r670;
mov.b32 %f328, %r688;
add.ftz.f32 %f329, %f327, %f328;
mov.b32 %r689, %f329;
shfl.sync.bfly.b32 %r691|%p90, %r689, %r673, %r668, %r670;
mov.b32 %f330, %r691;
add.ftz.f32 %f331, %f329, %f330;
mov.b32 %r692, %f331;
shfl.sync.bfly.b32 %r694|%p91, %r692, %r676, %r668, %r670;
mov.b32 %f332, %r694;
add.ftz.f32 %f333, %f331, %f332;
mov.b32 %r695, %f333;
shfl.sync.bfly.b32 %r697|%p92, %r695, %r679, %r668, %r670;
mov.b32 %f334, %r697;
add.ftz.f32 %f335, %f333, %f334;
mov.b32 %r698, %f335;
shfl.sync.bfly.b32 %r700|%p93, %r698, %r682, %r668, %r670;
mov.b32 %f336, %r700;
add.ftz.f32 %f56, %f335, %f336;
@%p22 bra $L__BB69_34;

st.global.f32 [%rd16+4], %f56;

$L__BB69_34:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type6ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type6ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type6ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type6ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type6ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type6ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type6ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type6ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type6ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type6ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<107>;
.reg .b16 %rs<29>;
.reg .f32 %f<417>;
.reg .b32 %r<785>;
.reg .b64 %rd<142>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type6ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[2048];

ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type6ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type6ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type6ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r51, [_Z13mul_mat_vec_qIL9ggml_type6ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type6ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type6ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type6ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type6ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type6ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type6ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r48, [_Z13mul_mat_vec_qIL9ggml_type6ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r49, [_Z13mul_mat_vec_qIL9ggml_type6ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r50, [_Z13mul_mat_vec_qIL9ggml_type6ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r46, [_Z13mul_mat_vec_qIL9ggml_type6ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r45, [_Z13mul_mat_vec_qIL9ggml_type6ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type6ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type6ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r52, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r52, %r2;
mov.u32 %r53, %ctaid.x;
shl.b32 %r4, %r53, 1;
shr.u32 %r54, %r3, 31;
add.s32 %r55, %r3, %r54;
shr.s32 %r784, %r55, 1;
shr.u32 %r6, %r51, 5;
setp.ge.s32 %p3, %r784, %r6;
mov.f32 %f401, 0f00000000;
mov.f32 %f402, %f401;
mov.f32 %f403, %f401;
mov.f32 %f404, %f401;
mov.f32 %f405, %f401;
mov.f32 %f406, %f401;
mov.f32 %f407, %f401;
mov.f32 %f408, %f401;
mov.f32 %f409, %f401;
mov.f32 %f410, %f401;
mov.f32 %f411, %f401;
mov.f32 %f412, %f401;
mov.f32 %f413, %f401;
mov.f32 %f414, %f401;
mov.f32 %f415, %f401;
mov.f32 %f416, %f401;
@%p3 bra $L__BB70_3;

mov.u32 %r56, %ctaid.y;
mov.u32 %r57, %ctaid.z;
mul.lo.s32 %r58, %r56, %r43;
cvt.u64.u32 %rd20, %r58;
mul.lo.s32 %r59, %r57, %r49;
cvt.u64.u32 %rd21, %r59;
add.s64 %rd1, %rd21, %rd20;
mul.hi.u32 %r60, %r57, %r45;
add.s32 %r61, %r60, %r57;
shr.u32 %r62, %r61, %r46;
mul.lo.s32 %r63, %r62, %r48;
mul.hi.u32 %r64, %r56, %r39;
add.s32 %r65, %r64, %r56;
shr.u32 %r66, %r65, %r40;
mul.lo.s32 %r67, %r66, %r42;
cvta.to.global.u64 %rd141, %rd17;
and.b32 %r70, %r55, -2;
sub.s32 %r71, %r3, %r70;
shl.b32 %r72, %r71, 1;
shl.b32 %r73, %r71, 2;
or.b32 %r74, %r73, 1;
shl.b32 %r7, %r71, 3;
cvt.s64.s32 %rd3, %r72;
sub.s32 %r75, %r74, %r72;
shl.b32 %r8, %r75, 2;
cvt.s64.s32 %rd4, %r75;
shl.b32 %r76, %r37, 1;
add.s32 %r783, %r784, %r76;
mad.lo.s32 %r782, %r37, 3, %r784;
shl.b32 %r77, %r37, 2;
add.s32 %r781, %r784, %r77;
mad.lo.s32 %r780, %r37, 5, %r784;
mad.lo.s32 %r779, %r37, 6, %r784;
mad.lo.s32 %r778, %r37, 7, %r784;
add.s32 %r777, %r784, %r37;
mad.lo.s32 %r78, %r4, %r36, %r784;
add.s32 %r79, %r78, %r63;
add.s32 %r80, %r79, %r67;
mul.wide.s32 %rd5, %r80, 22;
mul.wide.s32 %rd22, %r73, 2;
add.s64 %rd23, %rd5, %rd22;
add.s64 %rd6, %rd23, 8;
add.s32 %r81, %r784, %r63;
add.s32 %r82, %r81, %r67;
add.s32 %r83, %r4, 1;
mad.lo.s32 %r84, %r36, %r83, %r82;
mul.wide.s32 %rd7, %r84, 22;
add.s64 %rd24, %rd7, %rd22;
add.s64 %rd8, %rd24, 6;
cvta.to.global.u64 %rd9, %rd18;
mov.f32 %f401, 0f00000000;
shl.b64 %rd32, %rd3, 2;
shl.b64 %rd34, %rd4, 2;

$L__BB70_2:
cvt.u64.u32 %rd25, %r784;
add.s64 %rd26, %rd1, %rd25;
mul.lo.s64 %rd27, %rd26, 36;
add.s64 %rd28, %rd9, %rd27;
add.s64 %rd29, %rd28, 4;
add.s64 %rd30, %rd141, %rd5;
ld.global.nc.u16 %rs17, [%rd30+2];
ld.global.nc.u16 %rs18, [%rd30+4];
mov.b32 %r373, {%rs17, %rs18};
add.s64 %rd31, %rd141, %rd6;
ld.global.nc.u16 %rs19, [%rd31+-2];
ld.global.nc.u16 %rs20, [%rd31];
mov.b32 %r374, {%rs19, %rs20};
shr.s32 %r375, %r373, %r7;
add.s64 %rd33, %rd29, %rd32;
ld.global.nc.u32 %r87, [%rd33];
ld.global.nc.u32 %r91, [%rd33+16];
ld.global.nc.u16 %rs21, [%rd31+2];
ld.global.nc.u16 %rs22, [%rd31+4];
mov.b32 %r376, {%rs21, %rs22};
shr.s32 %r377, %r373, %r8;
add.s64 %rd35, %rd29, %rd34;
ld.global.nc.u32 %r95, [%rd35];
ld.global.nc.u32 %r99, [%rd33+20];
ld.global.nc.u16 %rs1, [%rd30];

{ cvt.f32.f16 %f97, %rs1;}


and.b32 %r378, %r374, 252645135;
shl.b32 %r379, %r375, 4;
and.b32 %r380, %r379, 16;
or.b32 %r381, %r380, %r378;
shl.b32 %r382, %r375, 11;
and.b32 %r383, %r382, 4096;
or.b32 %r384, %r381, %r383;
shl.b32 %r385, %r375, 18;
and.b32 %r386, %r385, 1048576;
or.b32 %r387, %r384, %r386;
shl.b32 %r388, %r375, 25;
and.b32 %r389, %r388, 268435456;
or.b32 %r338, %r387, %r389;
mov.u32 %r358, 0;

dp4a.s32.s32 %r85, %r338, %r87, %r358;

shr.u32 %r390, %r374, 4;
and.b32 %r391, %r390, 252645135;
shr.u32 %r392, %r375, 12;
and.b32 %r393, %r392, 16;
or.b32 %r394, %r393, %r391;
shr.u32 %r395, %r375, 5;
and.b32 %r396, %r395, 4096;
or.b32 %r397, %r394, %r396;
shl.b32 %r398, %r375, 2;
and.b32 %r399, %r398, 1048576;
or.b32 %r400, %r397, %r399;
shl.b32 %r401, %r375, 9;
and.b32 %r402, %r401, 268435456;
or.b32 %r342, %r400, %r402;

dp4a.s32.s32 %r89, %r342, %r91, %r85;

and.b32 %r403, %r376, 252645135;
shl.b32 %r404, %r377, 4;
and.b32 %r405, %r404, 16;
or.b32 %r406, %r405, %r403;
shl.b32 %r407, %r377, 11;
and.b32 %r408, %r407, 4096;
or.b32 %r409, %r406, %r408;
shl.b32 %r410, %r377, 18;
and.b32 %r411, %r410, 1048576;
or.b32 %r412, %r409, %r411;
shl.b32 %r413, %r377, 25;
and.b32 %r414, %r413, 268435456;
or.b32 %r346, %r412, %r414;

dp4a.s32.s32 %r93, %r346, %r95, %r89;

shr.u32 %r415, %r376, 4;
and.b32 %r416, %r415, 252645135;
shr.u32 %r417, %r377, 12;
and.b32 %r418, %r417, 16;
or.b32 %r419, %r418, %r416;
shr.u32 %r420, %r377, 5;
and.b32 %r421, %r420, 4096;
or.b32 %r422, %r419, %r421;
shl.b32 %r423, %r377, 2;
and.b32 %r424, %r423, 1048576;
or.b32 %r425, %r422, %r424;
shl.b32 %r426, %r377, 9;
and.b32 %r427, %r426, 268435456;
or.b32 %r350, %r425, %r427;

dp4a.s32.s32 %r97, %r350, %r99, %r93;

ld.global.nc.u32 %r101, [%rd28];

{.reg .f16 low,high;
mov.b32 {low,high},%r101;
cvt.f32.f16 %f98, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r101;
cvt.f32.f16 %f99, high;}


cvt.rn.f32.s32 %f145, %r97;
mul.ftz.f32 %f146, %f98, %f145;
mul.ftz.f32 %f147, %f99, 0f41000000;
sub.ftz.f32 %f148, %f146, %f147;
fma.rn.ftz.f32 %f416, %f97, %f148, %f416;
add.s64 %rd36, %rd141, %rd7;
ld.global.nc.u16 %rs23, [%rd36+2];
ld.global.nc.u16 %rs24, [%rd36+4];
mov.b32 %r428, {%rs23, %rs24};
add.s64 %rd37, %rd141, %rd8;
ld.global.nc.u16 %rs25, [%rd37];
ld.global.nc.u16 %rs26, [%rd37+2];
mov.b32 %r429, {%rs25, %rs26};
shr.s32 %r430, %r428, %r7;
ld.global.nc.u16 %rs27, [%rd37+4];
ld.global.nc.u16 %rs28, [%rd37+6];
mov.b32 %r431, {%rs27, %rs28};
shr.s32 %r432, %r428, %r8;
ld.global.nc.u16 %rs2, [%rd36];

{ cvt.f32.f16 %f100, %rs2;}


and.b32 %r433, %r429, 252645135;
shl.b32 %r434, %r430, 4;
and.b32 %r435, %r434, 16;
or.b32 %r436, %r435, %r433;
shl.b32 %r437, %r430, 11;
and.b32 %r438, %r437, 4096;
or.b32 %r439, %r436, %r438;
shl.b32 %r440, %r430, 18;
and.b32 %r441, %r440, 1048576;
or.b32 %r442, %r439, %r441;
shl.b32 %r443, %r430, 25;
and.b32 %r444, %r443, 268435456;
or.b32 %r356, %r442, %r444;

dp4a.s32.s32 %r103, %r356, %r87, %r358;

shr.u32 %r445, %r429, 4;
and.b32 %r446, %r445, 252645135;
shr.u32 %r447, %r430, 12;
and.b32 %r448, %r447, 16;
or.b32 %r449, %r448, %r446;
shr.u32 %r450, %r430, 5;
and.b32 %r451, %r450, 4096;
or.b32 %r452, %r449, %r451;
shl.b32 %r453, %r430, 2;
and.b32 %r454, %r453, 1048576;
or.b32 %r455, %r452, %r454;
shl.b32 %r456, %r430, 9;
and.b32 %r457, %r456, 268435456;
or.b32 %r360, %r455, %r457;

dp4a.s32.s32 %r107, %r360, %r91, %r103;

and.b32 %r458, %r431, 252645135;
shl.b32 %r459, %r432, 4;
and.b32 %r460, %r459, 16;
or.b32 %r461, %r460, %r458;
shl.b32 %r462, %r432, 11;
and.b32 %r463, %r462, 4096;
or.b32 %r464, %r461, %r463;
shl.b32 %r465, %r432, 18;
and.b32 %r466, %r465, 1048576;
or.b32 %r467, %r464, %r466;
shl.b32 %r468, %r432, 25;
and.b32 %r469, %r468, 268435456;
or.b32 %r364, %r467, %r469;

dp4a.s32.s32 %r111, %r364, %r95, %r107;

shr.u32 %r470, %r431, 4;
and.b32 %r471, %r470, 252645135;
shr.u32 %r472, %r432, 12;
and.b32 %r473, %r472, 16;
or.b32 %r474, %r473, %r471;
shr.u32 %r475, %r432, 5;
and.b32 %r476, %r475, 4096;
or.b32 %r477, %r474, %r476;
shl.b32 %r478, %r432, 2;
and.b32 %r479, %r478, 1048576;
or.b32 %r480, %r477, %r479;
shl.b32 %r481, %r432, 9;
and.b32 %r482, %r481, 268435456;
or.b32 %r368, %r480, %r482;

dp4a.s32.s32 %r115, %r368, %r99, %r111;


{.reg .f16 low,high;
mov.b32 {low,high},%r101;
cvt.f32.f16 %f101, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r101;
cvt.f32.f16 %f102, high;}


cvt.rn.f32.s32 %f149, %r115;
mul.ftz.f32 %f150, %f101, %f149;
mul.ftz.f32 %f151, %f102, 0f41000000;
sub.ftz.f32 %f152, %f150, %f151;
fma.rn.ftz.f32 %f415, %f100, %f152, %f415;
cvt.u64.u32 %rd38, %r777;
add.s64 %rd39, %rd1, %rd38;
mul.lo.s64 %rd40, %rd39, 36;
add.s64 %rd41, %rd9, %rd40;
add.s64 %rd42, %rd41, 4;
add.s64 %rd43, %rd42, %rd32;
ld.global.nc.u32 %r123, [%rd43];
ld.global.nc.u32 %r127, [%rd43+16];
add.s64 %rd44, %rd42, %rd34;
ld.global.nc.u32 %r131, [%rd44];
ld.global.nc.u32 %r135, [%rd43+20];

{ cvt.f32.f16 %f103, %rs1;}



dp4a.s32.s32 %r121, %r338, %r123, %r358;


dp4a.s32.s32 %r125, %r342, %r127, %r121;


dp4a.s32.s32 %r129, %r346, %r131, %r125;


dp4a.s32.s32 %r133, %r350, %r135, %r129;

ld.global.nc.u32 %r137, [%rd41];

{.reg .f16 low,high;
mov.b32 {low,high},%r137;
cvt.f32.f16 %f104, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r137;
cvt.f32.f16 %f105, high;}


cvt.rn.f32.s32 %f153, %r133;
mul.ftz.f32 %f154, %f104, %f153;
mul.ftz.f32 %f155, %f105, 0f41000000;
sub.ftz.f32 %f156, %f154, %f155;
fma.rn.ftz.f32 %f414, %f103, %f156, %f414;

{ cvt.f32.f16 %f106, %rs2;}



dp4a.s32.s32 %r139, %r356, %r123, %r358;


dp4a.s32.s32 %r143, %r360, %r127, %r139;


dp4a.s32.s32 %r147, %r364, %r131, %r143;


dp4a.s32.s32 %r151, %r368, %r135, %r147;


{.reg .f16 low,high;
mov.b32 {low,high},%r137;
cvt.f32.f16 %f107, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r137;
cvt.f32.f16 %f108, high;}


cvt.rn.f32.s32 %f157, %r151;
mul.ftz.f32 %f158, %f107, %f157;
mul.ftz.f32 %f159, %f108, 0f41000000;
sub.ftz.f32 %f160, %f158, %f159;
fma.rn.ftz.f32 %f413, %f106, %f160, %f413;
cvt.u64.u32 %rd45, %r783;
add.s64 %rd46, %rd1, %rd45;
mul.lo.s64 %rd47, %rd46, 36;
add.s64 %rd48, %rd9, %rd47;
add.s64 %rd49, %rd48, 4;
add.s64 %rd50, %rd49, %rd32;
ld.global.nc.u32 %r159, [%rd50];
ld.global.nc.u32 %r163, [%rd50+16];
add.s64 %rd51, %rd49, %rd34;
ld.global.nc.u32 %r167, [%rd51];
ld.global.nc.u32 %r171, [%rd50+20];

{ cvt.f32.f16 %f109, %rs1;}



dp4a.s32.s32 %r157, %r338, %r159, %r358;


dp4a.s32.s32 %r161, %r342, %r163, %r157;


dp4a.s32.s32 %r165, %r346, %r167, %r161;


dp4a.s32.s32 %r169, %r350, %r171, %r165;

ld.global.nc.u32 %r173, [%rd48];

{.reg .f16 low,high;
mov.b32 {low,high},%r173;
cvt.f32.f16 %f110, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r173;
cvt.f32.f16 %f111, high;}


cvt.rn.f32.s32 %f161, %r169;
mul.ftz.f32 %f162, %f110, %f161;
mul.ftz.f32 %f163, %f111, 0f41000000;
sub.ftz.f32 %f164, %f162, %f163;
fma.rn.ftz.f32 %f412, %f109, %f164, %f412;

{ cvt.f32.f16 %f112, %rs2;}



dp4a.s32.s32 %r175, %r356, %r159, %r358;


dp4a.s32.s32 %r179, %r360, %r163, %r175;


dp4a.s32.s32 %r183, %r364, %r167, %r179;


dp4a.s32.s32 %r187, %r368, %r171, %r183;


{.reg .f16 low,high;
mov.b32 {low,high},%r173;
cvt.f32.f16 %f113, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r173;
cvt.f32.f16 %f114, high;}


cvt.rn.f32.s32 %f165, %r187;
mul.ftz.f32 %f166, %f113, %f165;
mul.ftz.f32 %f167, %f114, 0f41000000;
sub.ftz.f32 %f168, %f166, %f167;
fma.rn.ftz.f32 %f411, %f112, %f168, %f411;
cvt.u64.u32 %rd52, %r782;
add.s64 %rd53, %rd1, %rd52;
mul.lo.s64 %rd54, %rd53, 36;
add.s64 %rd55, %rd9, %rd54;
add.s64 %rd56, %rd55, 4;
add.s64 %rd57, %rd56, %rd32;
ld.global.nc.u32 %r195, [%rd57];
ld.global.nc.u32 %r199, [%rd57+16];
add.s64 %rd58, %rd56, %rd34;
ld.global.nc.u32 %r203, [%rd58];
ld.global.nc.u32 %r207, [%rd57+20];

{ cvt.f32.f16 %f115, %rs1;}



dp4a.s32.s32 %r193, %r338, %r195, %r358;


dp4a.s32.s32 %r197, %r342, %r199, %r193;


dp4a.s32.s32 %r201, %r346, %r203, %r197;


dp4a.s32.s32 %r205, %r350, %r207, %r201;

ld.global.nc.u32 %r209, [%rd55];

{.reg .f16 low,high;
mov.b32 {low,high},%r209;
cvt.f32.f16 %f116, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r209;
cvt.f32.f16 %f117, high;}


cvt.rn.f32.s32 %f169, %r205;
mul.ftz.f32 %f170, %f116, %f169;
mul.ftz.f32 %f171, %f117, 0f41000000;
sub.ftz.f32 %f172, %f170, %f171;
fma.rn.ftz.f32 %f410, %f115, %f172, %f410;

{ cvt.f32.f16 %f118, %rs2;}



dp4a.s32.s32 %r211, %r356, %r195, %r358;


dp4a.s32.s32 %r215, %r360, %r199, %r211;


dp4a.s32.s32 %r219, %r364, %r203, %r215;


dp4a.s32.s32 %r223, %r368, %r207, %r219;


{.reg .f16 low,high;
mov.b32 {low,high},%r209;
cvt.f32.f16 %f119, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r209;
cvt.f32.f16 %f120, high;}


cvt.rn.f32.s32 %f173, %r223;
mul.ftz.f32 %f174, %f119, %f173;
mul.ftz.f32 %f175, %f120, 0f41000000;
sub.ftz.f32 %f176, %f174, %f175;
fma.rn.ftz.f32 %f409, %f118, %f176, %f409;
cvt.u64.u32 %rd59, %r781;
add.s64 %rd60, %rd1, %rd59;
mul.lo.s64 %rd61, %rd60, 36;
add.s64 %rd62, %rd9, %rd61;
add.s64 %rd63, %rd62, 4;
add.s64 %rd64, %rd63, %rd32;
ld.global.nc.u32 %r231, [%rd64];
ld.global.nc.u32 %r235, [%rd64+16];
add.s64 %rd65, %rd63, %rd34;
ld.global.nc.u32 %r239, [%rd65];
ld.global.nc.u32 %r243, [%rd64+20];

{ cvt.f32.f16 %f121, %rs1;}



dp4a.s32.s32 %r229, %r338, %r231, %r358;


dp4a.s32.s32 %r233, %r342, %r235, %r229;


dp4a.s32.s32 %r237, %r346, %r239, %r233;


dp4a.s32.s32 %r241, %r350, %r243, %r237;

ld.global.nc.u32 %r245, [%rd62];

{.reg .f16 low,high;
mov.b32 {low,high},%r245;
cvt.f32.f16 %f122, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r245;
cvt.f32.f16 %f123, high;}


cvt.rn.f32.s32 %f177, %r241;
mul.ftz.f32 %f178, %f122, %f177;
mul.ftz.f32 %f179, %f123, 0f41000000;
sub.ftz.f32 %f180, %f178, %f179;
fma.rn.ftz.f32 %f408, %f121, %f180, %f408;

{ cvt.f32.f16 %f124, %rs2;}



dp4a.s32.s32 %r247, %r356, %r231, %r358;


dp4a.s32.s32 %r251, %r360, %r235, %r247;


dp4a.s32.s32 %r255, %r364, %r239, %r251;


dp4a.s32.s32 %r259, %r368, %r243, %r255;


{.reg .f16 low,high;
mov.b32 {low,high},%r245;
cvt.f32.f16 %f125, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r245;
cvt.f32.f16 %f126, high;}


cvt.rn.f32.s32 %f181, %r259;
mul.ftz.f32 %f182, %f125, %f181;
mul.ftz.f32 %f183, %f126, 0f41000000;
sub.ftz.f32 %f184, %f182, %f183;
fma.rn.ftz.f32 %f407, %f124, %f184, %f407;
cvt.u64.u32 %rd66, %r780;
add.s64 %rd67, %rd1, %rd66;
mul.lo.s64 %rd68, %rd67, 36;
add.s64 %rd69, %rd9, %rd68;
add.s64 %rd70, %rd69, 4;
add.s64 %rd71, %rd70, %rd32;
ld.global.nc.u32 %r267, [%rd71];
ld.global.nc.u32 %r271, [%rd71+16];
add.s64 %rd72, %rd70, %rd34;
ld.global.nc.u32 %r275, [%rd72];
ld.global.nc.u32 %r279, [%rd71+20];

{ cvt.f32.f16 %f127, %rs1;}



dp4a.s32.s32 %r265, %r338, %r267, %r358;


dp4a.s32.s32 %r269, %r342, %r271, %r265;


dp4a.s32.s32 %r273, %r346, %r275, %r269;


dp4a.s32.s32 %r277, %r350, %r279, %r273;

ld.global.nc.u32 %r281, [%rd69];

{.reg .f16 low,high;
mov.b32 {low,high},%r281;
cvt.f32.f16 %f128, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r281;
cvt.f32.f16 %f129, high;}


cvt.rn.f32.s32 %f185, %r277;
mul.ftz.f32 %f186, %f128, %f185;
mul.ftz.f32 %f187, %f129, 0f41000000;
sub.ftz.f32 %f188, %f186, %f187;
fma.rn.ftz.f32 %f406, %f127, %f188, %f406;

{ cvt.f32.f16 %f130, %rs2;}



dp4a.s32.s32 %r283, %r356, %r267, %r358;


dp4a.s32.s32 %r287, %r360, %r271, %r283;


dp4a.s32.s32 %r291, %r364, %r275, %r287;


dp4a.s32.s32 %r295, %r368, %r279, %r291;


{.reg .f16 low,high;
mov.b32 {low,high},%r281;
cvt.f32.f16 %f131, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r281;
cvt.f32.f16 %f132, high;}


cvt.rn.f32.s32 %f189, %r295;
mul.ftz.f32 %f190, %f131, %f189;
mul.ftz.f32 %f191, %f132, 0f41000000;
sub.ftz.f32 %f192, %f190, %f191;
fma.rn.ftz.f32 %f405, %f130, %f192, %f405;
cvt.u64.u32 %rd73, %r779;
add.s64 %rd74, %rd1, %rd73;
mul.lo.s64 %rd75, %rd74, 36;
add.s64 %rd76, %rd9, %rd75;
add.s64 %rd77, %rd76, 4;
add.s64 %rd78, %rd77, %rd32;
ld.global.nc.u32 %r303, [%rd78];
ld.global.nc.u32 %r307, [%rd78+16];
add.s64 %rd79, %rd77, %rd34;
ld.global.nc.u32 %r311, [%rd79];
ld.global.nc.u32 %r315, [%rd78+20];

{ cvt.f32.f16 %f133, %rs1;}



dp4a.s32.s32 %r301, %r338, %r303, %r358;


dp4a.s32.s32 %r305, %r342, %r307, %r301;


dp4a.s32.s32 %r309, %r346, %r311, %r305;


dp4a.s32.s32 %r313, %r350, %r315, %r309;

ld.global.nc.u32 %r317, [%rd76];

{.reg .f16 low,high;
mov.b32 {low,high},%r317;
cvt.f32.f16 %f134, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r317;
cvt.f32.f16 %f135, high;}


cvt.rn.f32.s32 %f193, %r313;
mul.ftz.f32 %f194, %f134, %f193;
mul.ftz.f32 %f195, %f135, 0f41000000;
sub.ftz.f32 %f196, %f194, %f195;
fma.rn.ftz.f32 %f404, %f133, %f196, %f404;

{ cvt.f32.f16 %f136, %rs2;}



dp4a.s32.s32 %r319, %r356, %r303, %r358;


dp4a.s32.s32 %r323, %r360, %r307, %r319;


dp4a.s32.s32 %r327, %r364, %r311, %r323;


dp4a.s32.s32 %r331, %r368, %r315, %r327;


{.reg .f16 low,high;
mov.b32 {low,high},%r317;
cvt.f32.f16 %f137, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r317;
cvt.f32.f16 %f138, high;}


cvt.rn.f32.s32 %f197, %r331;
mul.ftz.f32 %f198, %f137, %f197;
mul.ftz.f32 %f199, %f138, 0f41000000;
sub.ftz.f32 %f200, %f198, %f199;
fma.rn.ftz.f32 %f403, %f136, %f200, %f403;
cvt.u64.u32 %rd80, %r778;
add.s64 %rd81, %rd1, %rd80;
mul.lo.s64 %rd82, %rd81, 36;
add.s64 %rd83, %rd9, %rd82;
add.s64 %rd84, %rd83, 4;
add.s64 %rd85, %rd84, %rd32;
ld.global.nc.u32 %r339, [%rd85];
ld.global.nc.u32 %r343, [%rd85+16];
add.s64 %rd86, %rd84, %rd34;
ld.global.nc.u32 %r347, [%rd86];
ld.global.nc.u32 %r351, [%rd85+20];

{ cvt.f32.f16 %f139, %rs1;}



dp4a.s32.s32 %r337, %r338, %r339, %r358;


dp4a.s32.s32 %r341, %r342, %r343, %r337;


dp4a.s32.s32 %r345, %r346, %r347, %r341;


dp4a.s32.s32 %r349, %r350, %r351, %r345;

ld.global.nc.u32 %r353, [%rd83];

{.reg .f16 low,high;
mov.b32 {low,high},%r353;
cvt.f32.f16 %f140, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r353;
cvt.f32.f16 %f141, high;}


cvt.rn.f32.s32 %f201, %r349;
mul.ftz.f32 %f202, %f140, %f201;
mul.ftz.f32 %f203, %f141, 0f41000000;
sub.ftz.f32 %f204, %f202, %f203;
fma.rn.ftz.f32 %f402, %f139, %f204, %f402;

{ cvt.f32.f16 %f142, %rs2;}



dp4a.s32.s32 %r355, %r356, %r339, %r358;


dp4a.s32.s32 %r359, %r360, %r343, %r355;


dp4a.s32.s32 %r363, %r364, %r347, %r359;


dp4a.s32.s32 %r367, %r368, %r351, %r363;


{.reg .f16 low,high;
mov.b32 {low,high},%r353;
cvt.f32.f16 %f143, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r353;
cvt.f32.f16 %f144, high;}


cvt.rn.f32.s32 %f205, %r367;
mul.ftz.f32 %f206, %f143, %f205;
mul.ftz.f32 %f207, %f144, 0f41000000;
sub.ftz.f32 %f208, %f206, %f207;
fma.rn.ftz.f32 %f401, %f142, %f208, %f401;
add.s32 %r783, %r783, 32;
add.s32 %r782, %r782, 32;
add.s32 %r781, %r781, 32;
add.s32 %r780, %r780, 32;
add.s32 %r779, %r779, 32;
add.s32 %r778, %r778, 32;
add.s32 %r777, %r777, 32;
add.s64 %rd141, %rd141, 704;
add.s32 %r784, %r784, 32;
setp.lt.s32 %p4, %r784, %r6;
@%p4 bra $L__BB70_2;

$L__BB70_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB70_5;

shl.b32 %r483, %r1, 11;
mov.u32 %r484, _ZZ13mul_mat_vec_qIL9ggml_type6ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r485, %r484, %r483;
shl.b32 %r486, %r2, 2;
add.s32 %r487, %r485, %r486;
st.shared.f32 [%r487+-2048], %f416;
st.shared.f32 [%r487+-1920], %f415;
st.shared.f32 [%r487+-1792], %f414;
st.shared.f32 [%r487+-1664], %f413;
st.shared.f32 [%r487+-1536], %f412;
st.shared.f32 [%r487+-1408], %f411;
st.shared.f32 [%r487+-1280], %f410;
st.shared.f32 [%r487+-1152], %f409;
st.shared.f32 [%r487+-1024], %f408;
st.shared.f32 [%r487+-896], %f407;
st.shared.f32 [%r487+-768], %f406;
st.shared.f32 [%r487+-640], %f405;
st.shared.f32 [%r487+-512], %f404;
st.shared.f32 [%r487+-384], %f403;
st.shared.f32 [%r487+-256], %f402;
st.shared.f32 [%r487+-128], %f401;

$L__BB70_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB70_38;

mov.u32 %r488, %ctaid.z;
mov.u32 %r489, %ctaid.y;
mad.lo.s32 %r490, %r489, %r44, %r4;
mad.lo.s32 %r491, %r488, %r50, %r490;
cvt.u64.u32 %rd12, %r491;
shl.b32 %r492, %r2, 2;
mov.u32 %r493, _ZZ13mul_mat_vec_qIL9ggml_type6ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r32, %r493, %r492;
mov.u32 %r494, 2;
ld.shared.f32 %f209, [%r32];
add.ftz.f32 %f210, %f209, %f416;
mov.b32 %r495, %f210;
mov.u32 %r496, 31;
mov.u32 %r497, 16;
mov.u32 %r498, -1;
shfl.sync.bfly.b32 %r499|%p7, %r495, %r497, %r496, %r498;
mov.b32 %f211, %r499;
add.ftz.f32 %f212, %f210, %f211;
mov.b32 %r500, %f212;
mov.u32 %r501, 8;
shfl.sync.bfly.b32 %r502|%p8, %r500, %r501, %r496, %r498;
mov.b32 %f213, %r502;
add.ftz.f32 %f214, %f212, %f213;
mov.b32 %r503, %f214;
mov.u32 %r504, 4;
shfl.sync.bfly.b32 %r505|%p9, %r503, %r504, %r496, %r498;
mov.b32 %f215, %r505;
add.ftz.f32 %f216, %f214, %f215;
mov.b32 %r506, %f216;
shfl.sync.bfly.b32 %r507|%p10, %r506, %r494, %r496, %r498;
mov.b32 %f217, %r507;
add.ftz.f32 %f218, %f216, %f217;
mov.b32 %r508, %f218;
mov.u32 %r509, 1;
shfl.sync.bfly.b32 %r510|%p11, %r508, %r509, %r496, %r498;
mov.b32 %f219, %r510;
add.ftz.f32 %f49, %f218, %f219;
setp.lt.u32 %p12, %r4, %r38;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
cvta.to.global.u64 %rd87, %rd19;
mul.wide.u32 %rd88, %r491, 4;
add.s64 %rd13, %rd87, %rd88;
not.pred %p14, %p1;
@%p14 bra $L__BB70_8;

st.global.f32 [%rd13], %f49;

$L__BB70_8:
ld.shared.f32 %f220, [%r32+128];
add.ftz.f32 %f221, %f220, %f415;
mov.b32 %r511, %f221;
shfl.sync.bfly.b32 %r515|%p15, %r511, %r497, %r496, %r498;
mov.b32 %f222, %r515;
add.ftz.f32 %f223, %f221, %f222;
mov.b32 %r516, %f223;
shfl.sync.bfly.b32 %r518|%p16, %r516, %r501, %r496, %r498;
mov.b32 %f224, %r518;
add.ftz.f32 %f225, %f223, %f224;
mov.b32 %r519, %f225;
shfl.sync.bfly.b32 %r521|%p17, %r519, %r504, %r496, %r498;
mov.b32 %f226, %r521;
add.ftz.f32 %f227, %f225, %f226;
mov.b32 %r522, %f227;
shfl.sync.bfly.b32 %r524|%p18, %r522, %r494, %r496, %r498;
mov.b32 %f228, %r524;
add.ftz.f32 %f229, %f227, %f228;
mov.b32 %r525, %f229;
shfl.sync.bfly.b32 %r527|%p19, %r525, %r509, %r496, %r498;
mov.b32 %f230, %r527;
add.ftz.f32 %f50, %f229, %f230;
setp.eq.s32 %p20, %r2, 1;
or.b32 %r531, %r4, 1;
setp.lt.u32 %p21, %r531, %r38;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB70_10;

st.global.f32 [%rd13+4], %f50;

$L__BB70_10:
ld.shared.f32 %f231, [%r32+256];
add.ftz.f32 %f232, %f231, %f414;
mov.b32 %r532, %f232;
mov.u32 %r533, 31;
mov.u32 %r534, 16;
mov.u32 %r535, -1;
shfl.sync.bfly.b32 %r536|%p23, %r532, %r534, %r533, %r535;
mov.b32 %f233, %r536;
add.ftz.f32 %f234, %f232, %f233;
mov.b32 %r537, %f234;
mov.u32 %r538, 8;
shfl.sync.bfly.b32 %r539|%p24, %r537, %r538, %r533, %r535;
mov.b32 %f235, %r539;
add.ftz.f32 %f236, %f234, %f235;
mov.b32 %r540, %f236;
mov.u32 %r541, 4;
shfl.sync.bfly.b32 %r542|%p25, %r540, %r541, %r533, %r535;
mov.b32 %f237, %r542;
add.ftz.f32 %f238, %f236, %f237;
mov.b32 %r543, %f238;
mov.u32 %r544, 2;
shfl.sync.bfly.b32 %r545|%p26, %r543, %r544, %r533, %r535;
mov.b32 %f239, %r545;
add.ftz.f32 %f240, %f238, %f239;
mov.b32 %r546, %f240;
mov.u32 %r547, 1;
shfl.sync.bfly.b32 %r548|%p27, %r546, %r547, %r533, %r535;
mov.b32 %f241, %r548;
add.ftz.f32 %f51, %f240, %f241;
@%p14 bra $L__BB70_12;

cvt.u64.u32 %rd89, %r38;
add.s64 %rd90, %rd89, %rd12;
shl.b64 %rd92, %rd90, 2;
add.s64 %rd93, %rd87, %rd92;
st.global.f32 [%rd93], %f51;

$L__BB70_12:
ld.shared.f32 %f242, [%r32+384];
add.ftz.f32 %f243, %f242, %f413;
mov.b32 %r549, %f243;
shfl.sync.bfly.b32 %r553|%p29, %r549, %r534, %r533, %r535;
mov.b32 %f244, %r553;
add.ftz.f32 %f245, %f243, %f244;
mov.b32 %r554, %f245;
shfl.sync.bfly.b32 %r556|%p30, %r554, %r538, %r533, %r535;
mov.b32 %f246, %r556;
add.ftz.f32 %f247, %f245, %f246;
mov.b32 %r557, %f247;
shfl.sync.bfly.b32 %r559|%p31, %r557, %r541, %r533, %r535;
mov.b32 %f248, %r559;
add.ftz.f32 %f249, %f247, %f248;
mov.b32 %r560, %f249;
shfl.sync.bfly.b32 %r562|%p32, %r560, %r544, %r533, %r535;
mov.b32 %f250, %r562;
add.ftz.f32 %f251, %f249, %f250;
mov.b32 %r563, %f251;
shfl.sync.bfly.b32 %r565|%p33, %r563, %r547, %r533, %r535;
mov.b32 %f252, %r565;
add.ftz.f32 %f52, %f251, %f252;
@%p22 bra $L__BB70_14;

add.s32 %r566, %r38, 1;
cvt.u64.u32 %rd94, %r566;
add.s64 %rd95, %rd94, %rd12;
shl.b64 %rd97, %rd95, 2;
add.s64 %rd98, %rd87, %rd97;
st.global.f32 [%rd98], %f52;

$L__BB70_14:
shl.b32 %r567, %r38, 1;
mov.u32 %r568, 1;
ld.shared.f32 %f253, [%r32+512];
add.ftz.f32 %f254, %f253, %f412;
mov.b32 %r569, %f254;
mov.u32 %r570, 31;
mov.u32 %r571, 16;
mov.u32 %r572, -1;
shfl.sync.bfly.b32 %r573|%p35, %r569, %r571, %r570, %r572;
mov.b32 %f255, %r573;
add.ftz.f32 %f256, %f254, %f255;
mov.b32 %r574, %f256;
mov.u32 %r575, 8;
shfl.sync.bfly.b32 %r576|%p36, %r574, %r575, %r570, %r572;
mov.b32 %f257, %r576;
add.ftz.f32 %f258, %f256, %f257;
mov.b32 %r577, %f258;
mov.u32 %r578, 4;
shfl.sync.bfly.b32 %r579|%p37, %r577, %r578, %r570, %r572;
mov.b32 %f259, %r579;
add.ftz.f32 %f260, %f258, %f259;
mov.b32 %r580, %f260;
mov.u32 %r581, 2;
shfl.sync.bfly.b32 %r582|%p38, %r580, %r581, %r570, %r572;
mov.b32 %f261, %r582;
add.ftz.f32 %f262, %f260, %f261;
mov.b32 %r583, %f262;
shfl.sync.bfly.b32 %r584|%p39, %r583, %r568, %r570, %r572;
mov.b32 %f263, %r584;
add.ftz.f32 %f53, %f262, %f263;
cvt.u64.u32 %rd99, %r567;
add.s64 %rd100, %rd99, %rd12;
shl.b64 %rd102, %rd100, 2;
add.s64 %rd14, %rd87, %rd102;
@%p14 bra $L__BB70_16;

st.global.f32 [%rd14], %f53;

$L__BB70_16:
ld.shared.f32 %f264, [%r32+640];
add.ftz.f32 %f265, %f264, %f411;
mov.b32 %r585, %f265;
shfl.sync.bfly.b32 %r589|%p41, %r585, %r571, %r570, %r572;
mov.b32 %f266, %r589;
add.ftz.f32 %f267, %f265, %f266;
mov.b32 %r590, %f267;
shfl.sync.bfly.b32 %r592|%p42, %r590, %r575, %r570, %r572;
mov.b32 %f268, %r592;
add.ftz.f32 %f269, %f267, %f268;
mov.b32 %r593, %f269;
shfl.sync.bfly.b32 %r595|%p43, %r593, %r578, %r570, %r572;
mov.b32 %f270, %r595;
add.ftz.f32 %f271, %f269, %f270;
mov.b32 %r596, %f271;
shfl.sync.bfly.b32 %r598|%p44, %r596, %r581, %r570, %r572;
mov.b32 %f272, %r598;
add.ftz.f32 %f273, %f271, %f272;
mov.b32 %r599, %f273;
shfl.sync.bfly.b32 %r601|%p45, %r599, %r568, %r570, %r572;
mov.b32 %f274, %r601;
add.ftz.f32 %f54, %f273, %f274;
@%p22 bra $L__BB70_18;

st.global.f32 [%rd14+4], %f54;

$L__BB70_18:
mul.lo.s32 %r33, %r38, 3;
ld.shared.f32 %f275, [%r32+768];
add.ftz.f32 %f276, %f275, %f410;
mov.b32 %r602, %f276;
mov.u32 %r603, 31;
mov.u32 %r604, 16;
mov.u32 %r605, -1;
shfl.sync.bfly.b32 %r606|%p47, %r602, %r604, %r603, %r605;
mov.b32 %f277, %r606;
add.ftz.f32 %f278, %f276, %f277;
mov.b32 %r607, %f278;
mov.u32 %r608, 8;
shfl.sync.bfly.b32 %r609|%p48, %r607, %r608, %r603, %r605;
mov.b32 %f279, %r609;
add.ftz.f32 %f280, %f278, %f279;
mov.b32 %r610, %f280;
mov.u32 %r611, 4;
shfl.sync.bfly.b32 %r612|%p49, %r610, %r611, %r603, %r605;
mov.b32 %f281, %r612;
add.ftz.f32 %f282, %f280, %f281;
mov.b32 %r613, %f282;
mov.u32 %r614, 2;
shfl.sync.bfly.b32 %r615|%p50, %r613, %r614, %r603, %r605;
mov.b32 %f283, %r615;
add.ftz.f32 %f284, %f282, %f283;
mov.b32 %r616, %f284;
mov.u32 %r617, 1;
shfl.sync.bfly.b32 %r618|%p51, %r616, %r617, %r603, %r605;
mov.b32 %f285, %r618;
add.ftz.f32 %f55, %f284, %f285;
@%p14 bra $L__BB70_20;

cvt.u64.u32 %rd103, %r33;
add.s64 %rd104, %rd103, %rd12;
shl.b64 %rd106, %rd104, 2;
add.s64 %rd107, %rd87, %rd106;
st.global.f32 [%rd107], %f55;

$L__BB70_20:
ld.shared.f32 %f286, [%r32+896];
add.ftz.f32 %f287, %f286, %f409;
mov.b32 %r619, %f287;
shfl.sync.bfly.b32 %r623|%p53, %r619, %r604, %r603, %r605;
mov.b32 %f288, %r623;
add.ftz.f32 %f289, %f287, %f288;
mov.b32 %r624, %f289;
shfl.sync.bfly.b32 %r626|%p54, %r624, %r608, %r603, %r605;
mov.b32 %f290, %r626;
add.ftz.f32 %f291, %f289, %f290;
mov.b32 %r627, %f291;
shfl.sync.bfly.b32 %r629|%p55, %r627, %r611, %r603, %r605;
mov.b32 %f292, %r629;
add.ftz.f32 %f293, %f291, %f292;
mov.b32 %r630, %f293;
shfl.sync.bfly.b32 %r632|%p56, %r630, %r614, %r603, %r605;
mov.b32 %f294, %r632;
add.ftz.f32 %f295, %f293, %f294;
mov.b32 %r633, %f295;
shfl.sync.bfly.b32 %r635|%p57, %r633, %r617, %r603, %r605;
mov.b32 %f296, %r635;
add.ftz.f32 %f56, %f295, %f296;
@%p22 bra $L__BB70_22;

add.s32 %r636, %r33, 1;
cvt.u64.u32 %rd108, %r636;
add.s64 %rd109, %rd108, %rd12;
shl.b64 %rd111, %rd109, 2;
add.s64 %rd112, %rd87, %rd111;
st.global.f32 [%rd112], %f56;

$L__BB70_22:
shl.b32 %r637, %r38, 2;
mov.u32 %r638, 2;
ld.shared.f32 %f297, [%r32+1024];
add.ftz.f32 %f298, %f297, %f408;
mov.b32 %r639, %f298;
mov.u32 %r640, 31;
mov.u32 %r641, 16;
mov.u32 %r642, -1;
shfl.sync.bfly.b32 %r643|%p59, %r639, %r641, %r640, %r642;
mov.b32 %f299, %r643;
add.ftz.f32 %f300, %f298, %f299;
mov.b32 %r644, %f300;
mov.u32 %r645, 8;
shfl.sync.bfly.b32 %r646|%p60, %r644, %r645, %r640, %r642;
mov.b32 %f301, %r646;
add.ftz.f32 %f302, %f300, %f301;
mov.b32 %r647, %f302;
mov.u32 %r648, 4;
shfl.sync.bfly.b32 %r649|%p61, %r647, %r648, %r640, %r642;
mov.b32 %f303, %r649;
add.ftz.f32 %f304, %f302, %f303;
mov.b32 %r650, %f304;
shfl.sync.bfly.b32 %r651|%p62, %r650, %r638, %r640, %r642;
mov.b32 %f305, %r651;
add.ftz.f32 %f306, %f304, %f305;
mov.b32 %r652, %f306;
mov.u32 %r653, 1;
shfl.sync.bfly.b32 %r654|%p63, %r652, %r653, %r640, %r642;
mov.b32 %f307, %r654;
add.ftz.f32 %f57, %f306, %f307;
cvt.u64.u32 %rd113, %r637;
add.s64 %rd114, %rd113, %rd12;
shl.b64 %rd116, %rd114, 2;
add.s64 %rd15, %rd87, %rd116;
@%p14 bra $L__BB70_24;

st.global.f32 [%rd15], %f57;

$L__BB70_24:
ld.shared.f32 %f308, [%r32+1152];
add.ftz.f32 %f309, %f308, %f407;
mov.b32 %r655, %f309;
shfl.sync.bfly.b32 %r659|%p65, %r655, %r641, %r640, %r642;
mov.b32 %f310, %r659;
add.ftz.f32 %f311, %f309, %f310;
mov.b32 %r660, %f311;
shfl.sync.bfly.b32 %r662|%p66, %r660, %r645, %r640, %r642;
mov.b32 %f312, %r662;
add.ftz.f32 %f313, %f311, %f312;
mov.b32 %r663, %f313;
shfl.sync.bfly.b32 %r665|%p67, %r663, %r648, %r640, %r642;
mov.b32 %f314, %r665;
add.ftz.f32 %f315, %f313, %f314;
mov.b32 %r666, %f315;
shfl.sync.bfly.b32 %r668|%p68, %r666, %r638, %r640, %r642;
mov.b32 %f316, %r668;
add.ftz.f32 %f317, %f315, %f316;
mov.b32 %r669, %f317;
shfl.sync.bfly.b32 %r671|%p69, %r669, %r653, %r640, %r642;
mov.b32 %f318, %r671;
add.ftz.f32 %f58, %f317, %f318;
@%p22 bra $L__BB70_26;

st.global.f32 [%rd15+4], %f58;

$L__BB70_26:
mul.lo.s32 %r34, %r38, 5;
ld.shared.f32 %f319, [%r32+1280];
add.ftz.f32 %f320, %f319, %f406;
mov.b32 %r672, %f320;
mov.u32 %r673, 31;
mov.u32 %r674, 16;
mov.u32 %r675, -1;
shfl.sync.bfly.b32 %r676|%p71, %r672, %r674, %r673, %r675;
mov.b32 %f321, %r676;
add.ftz.f32 %f322, %f320, %f321;
mov.b32 %r677, %f322;
mov.u32 %r678, 8;
shfl.sync.bfly.b32 %r679|%p72, %r677, %r678, %r673, %r675;
mov.b32 %f323, %r679;
add.ftz.f32 %f324, %f322, %f323;
mov.b32 %r680, %f324;
mov.u32 %r681, 4;
shfl.sync.bfly.b32 %r682|%p73, %r680, %r681, %r673, %r675;
mov.b32 %f325, %r682;
add.ftz.f32 %f326, %f324, %f325;
mov.b32 %r683, %f326;
mov.u32 %r684, 2;
shfl.sync.bfly.b32 %r685|%p74, %r683, %r684, %r673, %r675;
mov.b32 %f327, %r685;
add.ftz.f32 %f328, %f326, %f327;
mov.b32 %r686, %f328;
mov.u32 %r687, 1;
shfl.sync.bfly.b32 %r688|%p75, %r686, %r687, %r673, %r675;
mov.b32 %f329, %r688;
add.ftz.f32 %f59, %f328, %f329;
@%p14 bra $L__BB70_28;

cvt.u64.u32 %rd117, %r34;
add.s64 %rd118, %rd117, %rd12;
shl.b64 %rd120, %rd118, 2;
add.s64 %rd121, %rd87, %rd120;
st.global.f32 [%rd121], %f59;

$L__BB70_28:
ld.shared.f32 %f330, [%r32+1408];
add.ftz.f32 %f331, %f330, %f405;
mov.b32 %r689, %f331;
shfl.sync.bfly.b32 %r693|%p77, %r689, %r674, %r673, %r675;
mov.b32 %f332, %r693;
add.ftz.f32 %f333, %f331, %f332;
mov.b32 %r694, %f333;
shfl.sync.bfly.b32 %r696|%p78, %r694, %r678, %r673, %r675;
mov.b32 %f334, %r696;
add.ftz.f32 %f335, %f333, %f334;
mov.b32 %r697, %f335;
shfl.sync.bfly.b32 %r699|%p79, %r697, %r681, %r673, %r675;
mov.b32 %f336, %r699;
add.ftz.f32 %f337, %f335, %f336;
mov.b32 %r700, %f337;
shfl.sync.bfly.b32 %r702|%p80, %r700, %r684, %r673, %r675;
mov.b32 %f338, %r702;
add.ftz.f32 %f339, %f337, %f338;
mov.b32 %r703, %f339;
shfl.sync.bfly.b32 %r705|%p81, %r703, %r687, %r673, %r675;
mov.b32 %f340, %r705;
add.ftz.f32 %f60, %f339, %f340;
@%p22 bra $L__BB70_30;

add.s32 %r706, %r34, 1;
cvt.u64.u32 %rd122, %r706;
add.s64 %rd123, %rd122, %rd12;
shl.b64 %rd125, %rd123, 2;
add.s64 %rd126, %rd87, %rd125;
st.global.f32 [%rd126], %f60;

$L__BB70_30:
mul.lo.s32 %r707, %r38, 6;
ld.shared.f32 %f341, [%r32+1536];
add.ftz.f32 %f342, %f341, %f404;
mov.b32 %r708, %f342;
mov.u32 %r709, 31;
mov.u32 %r710, 16;
mov.u32 %r711, -1;
shfl.sync.bfly.b32 %r712|%p83, %r708, %r710, %r709, %r711;
mov.b32 %f343, %r712;
add.ftz.f32 %f344, %f342, %f343;
mov.b32 %r713, %f344;
mov.u32 %r714, 8;
shfl.sync.bfly.b32 %r715|%p84, %r713, %r714, %r709, %r711;
mov.b32 %f345, %r715;
add.ftz.f32 %f346, %f344, %f345;
mov.b32 %r716, %f346;
mov.u32 %r717, 4;
shfl.sync.bfly.b32 %r718|%p85, %r716, %r717, %r709, %r711;
mov.b32 %f347, %r718;
add.ftz.f32 %f348, %f346, %f347;
mov.b32 %r719, %f348;
mov.u32 %r720, 2;
shfl.sync.bfly.b32 %r721|%p86, %r719, %r720, %r709, %r711;
mov.b32 %f349, %r721;
add.ftz.f32 %f350, %f348, %f349;
mov.b32 %r722, %f350;
mov.u32 %r723, 1;
shfl.sync.bfly.b32 %r724|%p87, %r722, %r723, %r709, %r711;
mov.b32 %f351, %r724;
add.ftz.f32 %f61, %f350, %f351;
cvt.u64.u32 %rd127, %r707;
add.s64 %rd128, %rd127, %rd12;
shl.b64 %rd130, %rd128, 2;
add.s64 %rd16, %rd87, %rd130;
@%p14 bra $L__BB70_32;

st.global.f32 [%rd16], %f61;

$L__BB70_32:
ld.shared.f32 %f352, [%r32+1664];
add.ftz.f32 %f353, %f352, %f403;
mov.b32 %r725, %f353;
shfl.sync.bfly.b32 %r729|%p89, %r725, %r710, %r709, %r711;
mov.b32 %f354, %r729;
add.ftz.f32 %f355, %f353, %f354;
mov.b32 %r730, %f355;
shfl.sync.bfly.b32 %r732|%p90, %r730, %r714, %r709, %r711;
mov.b32 %f356, %r732;
add.ftz.f32 %f357, %f355, %f356;
mov.b32 %r733, %f357;
shfl.sync.bfly.b32 %r735|%p91, %r733, %r717, %r709, %r711;
mov.b32 %f358, %r735;
add.ftz.f32 %f359, %f357, %f358;
mov.b32 %r736, %f359;
shfl.sync.bfly.b32 %r738|%p92, %r736, %r720, %r709, %r711;
mov.b32 %f360, %r738;
add.ftz.f32 %f361, %f359, %f360;
mov.b32 %r739, %f361;
shfl.sync.bfly.b32 %r741|%p93, %r739, %r723, %r709, %r711;
mov.b32 %f362, %r741;
add.ftz.f32 %f62, %f361, %f362;
@%p22 bra $L__BB70_34;

st.global.f32 [%rd16+4], %f62;

$L__BB70_34:
mul.lo.s32 %r35, %r38, 7;
ld.shared.f32 %f363, [%r32+1792];
add.ftz.f32 %f364, %f363, %f402;
mov.b32 %r742, %f364;
mov.u32 %r743, 31;
mov.u32 %r744, 16;
mov.u32 %r745, -1;
shfl.sync.bfly.b32 %r746|%p95, %r742, %r744, %r743, %r745;
mov.b32 %f365, %r746;
add.ftz.f32 %f366, %f364, %f365;
mov.b32 %r747, %f366;
mov.u32 %r748, 8;
shfl.sync.bfly.b32 %r749|%p96, %r747, %r748, %r743, %r745;
mov.b32 %f367, %r749;
add.ftz.f32 %f368, %f366, %f367;
mov.b32 %r750, %f368;
mov.u32 %r751, 4;
shfl.sync.bfly.b32 %r752|%p97, %r750, %r751, %r743, %r745;
mov.b32 %f369, %r752;
add.ftz.f32 %f370, %f368, %f369;
mov.b32 %r753, %f370;
mov.u32 %r754, 2;
shfl.sync.bfly.b32 %r755|%p98, %r753, %r754, %r743, %r745;
mov.b32 %f371, %r755;
add.ftz.f32 %f372, %f370, %f371;
mov.b32 %r756, %f372;
mov.u32 %r757, 1;
shfl.sync.bfly.b32 %r758|%p99, %r756, %r757, %r743, %r745;
mov.b32 %f373, %r758;
add.ftz.f32 %f63, %f372, %f373;
@%p14 bra $L__BB70_36;

cvt.u64.u32 %rd131, %r35;
add.s64 %rd132, %rd131, %rd12;
shl.b64 %rd134, %rd132, 2;
add.s64 %rd135, %rd87, %rd134;
st.global.f32 [%rd135], %f63;

$L__BB70_36:
ld.shared.f32 %f374, [%r32+1920];
add.ftz.f32 %f375, %f374, %f401;
mov.b32 %r759, %f375;
shfl.sync.bfly.b32 %r763|%p101, %r759, %r744, %r743, %r745;
mov.b32 %f376, %r763;
add.ftz.f32 %f377, %f375, %f376;
mov.b32 %r764, %f377;
shfl.sync.bfly.b32 %r766|%p102, %r764, %r748, %r743, %r745;
mov.b32 %f378, %r766;
add.ftz.f32 %f379, %f377, %f378;
mov.b32 %r767, %f379;
shfl.sync.bfly.b32 %r769|%p103, %r767, %r751, %r743, %r745;
mov.b32 %f380, %r769;
add.ftz.f32 %f381, %f379, %f380;
mov.b32 %r770, %f381;
shfl.sync.bfly.b32 %r772|%p104, %r770, %r754, %r743, %r745;
mov.b32 %f382, %r772;
add.ftz.f32 %f383, %f381, %f382;
mov.b32 %r773, %f383;
shfl.sync.bfly.b32 %r775|%p105, %r773, %r757, %r743, %r745;
mov.b32 %f384, %r775;
add.ftz.f32 %f64, %f383, %f384;
@%p22 bra $L__BB70_38;

add.s32 %r776, %r35, 1;
cvt.u64.u32 %rd136, %r776;
add.s64 %rd137, %rd136, %rd12;
shl.b64 %rd139, %rd137, 2;
add.s64 %rd140, %rd87, %rd139;
st.global.f32 [%rd140], %f64;

$L__BB70_38:
ret;

}
.entry _Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj(
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2,
.param .align 8 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3[48],
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5,
.param .align 4 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6[12],
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15
)
.maxntid 256, 1, 1
.minnctapersm 1
{
.local .align 8 .b8 __local_depot71[16];
.reg .b64 %SP;
.reg .b64 %SPL;
.reg .pred %p<48>;
.reg .f32 %f<177>;
.reg .b32 %r<477>;
.reg .b64 %rd<101>;


mov.u64 %SPL, __local_depot71;
ld.param.u64 %rd22, [_Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0];
ld.param.u64 %rd23, [_Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1];
ld.param.u64 %rd24, [_Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2];
ld.param.u32 %r17, [_Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5];
ld.param.u32 %r22, [_Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8];
ld.param.u32 %r23, [_Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9];
ld.param.u32 %r25, [_Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11];
ld.param.u32 %r26, [_Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12];
ld.param.u32 %r29, [_Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14];
ld.param.u32 %r28, [_Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15];
ld.param.u32 %r20, [_Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+8];
ld.param.u32 %r19, [_Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+4];
ld.param.u32 %r18, [_Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6];
ld.param.u64 %rd26, [_Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+8];
add.u64 %rd1, %SPL, 0;
add.u64 %rd2, %SPL, 8;
mov.u32 %r1, %tid.y;
setp.ge.u32 %p1, %r1, %r29;
@%p1 bra $L__BB71_29;

mov.u32 %r2, %ctaid.y;
cvta.to.global.u64 %rd33, %rd24;
mad.lo.s32 %r30, %r1, %r28, %r2;
mul.wide.u32 %rd34, %r30, 4;
add.s64 %rd35, %rd33, %rd34;
ld.global.nc.u32 %r3, [%rd35];
mov.f32 %f170, 0f00000000;
st.local.v2.f32 [%rd1], {%f170, %f170};
st.local.v2.f32 [%rd2], {%f170, %f170};
mov.u32 %r4, %tid.x;
shr.u32 %r5, %r4, 1;
shr.u32 %r31, %r17, 5;
setp.ge.u32 %p2, %r5, %r31;
mov.f32 %f171, %f170;
mov.f32 %f172, %f170;
mov.f32 %f173, %f170;
@%p2 bra $L__BB71_9;

cvta.to.global.u64 %rd36, %rd23;
mul.hi.u32 %r32, %r2, %r18;
add.s32 %r33, %r32, %r2;
shr.u32 %r34, %r33, %r19;
mul.lo.s32 %r35, %r34, %r20;
sub.s32 %r36, %r2, %r35;
mul.lo.s32 %r37, %r36, %r26;
cvt.u64.u32 %rd37, %r37;
mul.lo.s32 %r38, %r1, %r23;
cvt.u64.u32 %rd38, %r38;
cvt.u64.u32 %rd39, %r4;
shr.u64 %rd40, %rd39, 1;
mul.lo.s64 %rd41, %rd40, 36;
mul.wide.u32 %rd42, %r38, 36;
add.s64 %rd43, %rd41, %rd42;
add.s64 %rd4, %rd36, %rd43;
mul.wide.u32 %rd100, %r37, 36;
and.b32 %r39, %r4, 1;
mul.wide.u32 %rd44, %r39, 8;
add.s64 %rd45, %rd40, %rd38;
add.s64 %rd46, %rd45, %rd37;
mul.lo.s64 %rd47, %rd46, 36;
add.s64 %rd48, %rd47, %rd44;
add.s64 %rd49, %rd36, %rd48;
add.s64 %rd98, %rd49, 20;
mov.u32 %r40, %ctaid.x;
shl.b32 %r41, %r40, 1;
or.b32 %r42, %r41, 1;
mad.lo.s32 %r43, %r3, %r25, %r5;
mad.lo.s32 %r44, %r22, %r42, %r43;
mul.wide.s32 %rd7, %r44, 24;
add.s64 %rd8, %rd44, %rd7;
cvta.to.global.u64 %rd99, %rd22;
cvta.to.global.u64 %rd97, %rd26;
mov.f32 %f170, 0f00000000;
mov.u32 %r476, %r5;

$L__BB71_3:
add.s64 %rd50, %rd4, %rd100;
mad.lo.s32 %r70, %r41, %r22, %r5;
mad.lo.s32 %r71, %r3, %r25, %r70;
mul.wide.s32 %rd51, %r71, 24;
add.s64 %rd52, %rd99, %rd51;
add.s64 %rd54, %rd44, %rd51;
add.s64 %rd55, %rd99, %rd54;
ld.global.nc.u32 %r73, [%rd52+4];
shl.b32 %r74, %r4, 3;
and.b32 %r75, %r74, 8;
shr.s32 %r76, %r73, %r75;
ld.global.nc.u32 %r7, [%rd98+-16];
ld.global.nc.u32 %r8, [%rd98];
ld.global.nc.u32 %r77, [%rd55+12];
or.b32 %r78, %r75, 4;
shr.s32 %r79, %r73, %r78;
ld.global.nc.u32 %r9, [%rd98+-12];
ld.global.nc.u32 %r10, [%rd98+4];
ld.global.nc.u32 %r80, [%rd55+8];
and.b32 %r81, %r80, 252645135;
shl.b32 %r82, %r76, 4;
and.b32 %r83, %r82, 16;
or.b32 %r84, %r83, %r81;
shl.b32 %r85, %r76, 11;
and.b32 %r86, %r85, 4096;
or.b32 %r87, %r84, %r86;
shl.b32 %r88, %r76, 18;
and.b32 %r89, %r88, 1048576;
or.b32 %r90, %r87, %r89;
shl.b32 %r91, %r76, 25;
and.b32 %r92, %r91, 268435456;
or.b32 %r46, %r90, %r92;
mov.u32 %r48, 0;

dp4a.s32.s32 %r45, %r46, %r7, %r48;

shr.u32 %r93, %r80, 4;
and.b32 %r94, %r93, 252645135;
shr.u32 %r95, %r76, 12;
and.b32 %r96, %r95, 16;
or.b32 %r97, %r96, %r94;
shr.u32 %r98, %r76, 5;
and.b32 %r99, %r98, 4096;
or.b32 %r100, %r97, %r99;
shl.b32 %r101, %r76, 2;
and.b32 %r102, %r101, 1048576;
or.b32 %r103, %r100, %r102;
shl.b32 %r104, %r76, 9;
and.b32 %r105, %r104, 268435456;
or.b32 %r50, %r103, %r105;

dp4a.s32.s32 %r49, %r50, %r8, %r45;

and.b32 %r106, %r77, 252645135;
shl.b32 %r107, %r79, 4;
and.b32 %r108, %r107, 16;
or.b32 %r109, %r108, %r106;
shl.b32 %r110, %r79, 11;
and.b32 %r111, %r110, 4096;
or.b32 %r112, %r109, %r111;
shl.b32 %r113, %r79, 18;
and.b32 %r114, %r113, 1048576;
or.b32 %r115, %r112, %r114;
shl.b32 %r116, %r79, 25;
and.b32 %r117, %r116, 268435456;
or.b32 %r54, %r115, %r117;

dp4a.s32.s32 %r53, %r54, %r9, %r49;

shr.u32 %r118, %r77, 4;
and.b32 %r119, %r118, 252645135;
shr.u32 %r120, %r79, 12;
and.b32 %r121, %r120, 16;
or.b32 %r122, %r121, %r119;
shr.u32 %r123, %r79, 5;
and.b32 %r124, %r123, 4096;
or.b32 %r125, %r122, %r124;
shl.b32 %r126, %r79, 2;
and.b32 %r127, %r126, 1048576;
or.b32 %r128, %r125, %r127;
shl.b32 %r129, %r79, 9;
and.b32 %r130, %r129, 268435456;
or.b32 %r58, %r128, %r130;

dp4a.s32.s32 %r57, %r58, %r10, %r53;

ld.global.nc.u32 %r62, [%rd52];
ld.global.nc.u32 %r11, [%rd50];

{mul.f16x2 %r61,%r62,%r11;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r61;
cvt.f32.f16 %f42, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r61;
cvt.f32.f16 %f43, high;}


cvt.rn.f32.s32 %f44, %r57;
mov.f32 %f45, 0f40000000;
div.approx.ftz.f32 %f46, %f43, %f45;
fma.rn.ftz.f32 %f47, %f42, %f44, %f46;
add.ftz.f32 %f173, %f47, %f173;
setp.eq.s64 %p3, %rd26, 0;
@%p3 bra $L__BB71_5;

shl.b32 %r473, %r4, 3;
and.b32 %r472, %r473, 8;
shr.u32 %r153, %r4, 1;
mad.lo.s32 %r156, %r41, %r22, %r153;
mad.lo.s32 %r157, %r3, %r25, %r156;
mul.wide.s32 %rd56, %r157, 24;
add.s64 %rd57, %rd97, %rd56;
add.s64 %rd59, %rd44, %rd56;
add.s64 %rd60, %rd97, %rd59;
ld.global.u32 %r159, [%rd57+4];
shr.s32 %r162, %r159, %r472;
ld.global.u32 %r163, [%rd60+12];
shr.s32 %r165, %r159, %r78;
ld.global.u32 %r166, [%rd60+8];
and.b32 %r167, %r166, 252645135;
shl.b32 %r168, %r162, 4;
and.b32 %r169, %r168, 16;
or.b32 %r170, %r169, %r167;
shl.b32 %r171, %r162, 11;
and.b32 %r172, %r171, 4096;
or.b32 %r173, %r170, %r172;
shl.b32 %r174, %r162, 18;
and.b32 %r175, %r174, 1048576;
or.b32 %r176, %r173, %r175;
shl.b32 %r177, %r162, 25;
and.b32 %r178, %r177, 268435456;
or.b32 %r132, %r176, %r178;

dp4a.s32.s32 %r131, %r132, %r7, %r48;

shr.u32 %r179, %r166, 4;
and.b32 %r180, %r179, 252645135;
shr.u32 %r181, %r162, 12;
and.b32 %r182, %r181, 16;
or.b32 %r183, %r182, %r180;
shr.u32 %r184, %r162, 5;
and.b32 %r185, %r184, 4096;
or.b32 %r186, %r183, %r185;
shl.b32 %r187, %r162, 2;
and.b32 %r188, %r187, 1048576;
or.b32 %r189, %r186, %r188;
shl.b32 %r190, %r162, 9;
and.b32 %r191, %r190, 268435456;
or.b32 %r136, %r189, %r191;

dp4a.s32.s32 %r135, %r136, %r8, %r131;

and.b32 %r192, %r163, 252645135;
shl.b32 %r193, %r165, 4;
and.b32 %r194, %r193, 16;
or.b32 %r195, %r194, %r192;
shl.b32 %r196, %r165, 11;
and.b32 %r197, %r196, 4096;
or.b32 %r198, %r195, %r197;
shl.b32 %r199, %r165, 18;
and.b32 %r200, %r199, 1048576;
or.b32 %r201, %r198, %r200;
shl.b32 %r202, %r165, 25;
and.b32 %r203, %r202, 268435456;
or.b32 %r140, %r201, %r203;

dp4a.s32.s32 %r139, %r140, %r9, %r135;

shr.u32 %r204, %r163, 4;
and.b32 %r205, %r204, 252645135;
shr.u32 %r206, %r165, 12;
and.b32 %r207, %r206, 16;
or.b32 %r208, %r207, %r205;
shr.u32 %r209, %r165, 5;
and.b32 %r210, %r209, 4096;
or.b32 %r211, %r208, %r210;
shl.b32 %r212, %r165, 2;
and.b32 %r213, %r212, 1048576;
or.b32 %r214, %r211, %r213;
shl.b32 %r215, %r165, 9;
and.b32 %r216, %r215, 268435456;
or.b32 %r144, %r214, %r216;

dp4a.s32.s32 %r143, %r144, %r10, %r139;

ld.global.u32 %r148, [%rd57];

{mul.f16x2 %r147,%r148,%r11;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r147;
cvt.f32.f16 %f48, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r147;
cvt.f32.f16 %f49, high;}


cvt.rn.f32.s32 %f50, %r143;
div.approx.ftz.f32 %f52, %f49, %f45;
fma.rn.ftz.f32 %f53, %f48, %f50, %f52;
add.ftz.f32 %f171, %f53, %f171;

$L__BB71_5:
shl.b32 %r471, %r4, 3;
and.b32 %r470, %r471, 8;
add.s64 %rd61, %rd99, %rd7;
add.s64 %rd62, %rd99, %rd8;
ld.global.nc.u32 %r239, [%rd61+4];
shr.s32 %r242, %r239, %r470;
ld.global.nc.u32 %r243, [%rd62+12];
shr.s32 %r245, %r239, %r78;
ld.global.nc.u32 %r246, [%rd62+8];
and.b32 %r247, %r246, 252645135;
shl.b32 %r248, %r242, 4;
and.b32 %r249, %r248, 16;
or.b32 %r250, %r249, %r247;
shl.b32 %r251, %r242, 11;
and.b32 %r252, %r251, 4096;
or.b32 %r253, %r250, %r252;
shl.b32 %r254, %r242, 18;
and.b32 %r255, %r254, 1048576;
or.b32 %r256, %r253, %r255;
shl.b32 %r257, %r242, 25;
and.b32 %r258, %r257, 268435456;
or.b32 %r218, %r256, %r258;

dp4a.s32.s32 %r217, %r218, %r7, %r48;

shr.u32 %r259, %r246, 4;
and.b32 %r260, %r259, 252645135;
shr.u32 %r261, %r242, 12;
and.b32 %r262, %r261, 16;
or.b32 %r263, %r262, %r260;
shr.u32 %r264, %r242, 5;
and.b32 %r265, %r264, 4096;
or.b32 %r266, %r263, %r265;
shl.b32 %r267, %r242, 2;
and.b32 %r268, %r267, 1048576;
or.b32 %r269, %r266, %r268;
shl.b32 %r270, %r242, 9;
and.b32 %r271, %r270, 268435456;
or.b32 %r222, %r269, %r271;

dp4a.s32.s32 %r221, %r222, %r8, %r217;

and.b32 %r272, %r243, 252645135;
shl.b32 %r273, %r245, 4;
and.b32 %r274, %r273, 16;
or.b32 %r275, %r274, %r272;
shl.b32 %r276, %r245, 11;
and.b32 %r277, %r276, 4096;
or.b32 %r278, %r275, %r277;
shl.b32 %r279, %r245, 18;
and.b32 %r280, %r279, 1048576;
or.b32 %r281, %r278, %r280;
shl.b32 %r282, %r245, 25;
and.b32 %r283, %r282, 268435456;
or.b32 %r226, %r281, %r283;

dp4a.s32.s32 %r225, %r226, %r9, %r221;

shr.u32 %r284, %r243, 4;
and.b32 %r285, %r284, 252645135;
shr.u32 %r286, %r245, 12;
and.b32 %r287, %r286, 16;
or.b32 %r288, %r287, %r285;
shr.u32 %r289, %r245, 5;
and.b32 %r290, %r289, 4096;
or.b32 %r291, %r288, %r290;
shl.b32 %r292, %r245, 2;
and.b32 %r293, %r292, 1048576;
or.b32 %r294, %r291, %r293;
shl.b32 %r295, %r245, 9;
and.b32 %r296, %r295, 268435456;
or.b32 %r230, %r294, %r296;

dp4a.s32.s32 %r229, %r230, %r10, %r225;

ld.global.nc.u32 %r234, [%rd61];

{mul.f16x2 %r233,%r234,%r11;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r233;
cvt.f32.f16 %f54, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r233;
cvt.f32.f16 %f55, high;}


cvt.rn.f32.s32 %f56, %r229;
div.approx.ftz.f32 %f58, %f55, %f45;
fma.rn.ftz.f32 %f59, %f54, %f56, %f58;
add.ftz.f32 %f172, %f59, %f172;
@%p3 bra $L__BB71_7;

shl.b32 %r475, %r4, 3;
and.b32 %r474, %r475, 8;
add.s64 %rd63, %rd97, %rd7;
add.s64 %rd64, %rd97, %rd8;
ld.global.u32 %r319, [%rd63+4];
shr.s32 %r322, %r319, %r474;
ld.global.u32 %r323, [%rd64+12];
shr.s32 %r325, %r319, %r78;
ld.global.u32 %r326, [%rd64+8];
and.b32 %r327, %r326, 252645135;
shl.b32 %r328, %r322, 4;
and.b32 %r329, %r328, 16;
or.b32 %r330, %r329, %r327;
shl.b32 %r331, %r322, 11;
and.b32 %r332, %r331, 4096;
or.b32 %r333, %r330, %r332;
shl.b32 %r334, %r322, 18;
and.b32 %r335, %r334, 1048576;
or.b32 %r336, %r333, %r335;
shl.b32 %r337, %r322, 25;
and.b32 %r338, %r337, 268435456;
or.b32 %r298, %r336, %r338;
mov.u32 %r300, 0;

dp4a.s32.s32 %r297, %r298, %r7, %r300;

shr.u32 %r339, %r326, 4;
and.b32 %r340, %r339, 252645135;
shr.u32 %r341, %r322, 12;
and.b32 %r342, %r341, 16;
or.b32 %r343, %r342, %r340;
shr.u32 %r344, %r322, 5;
and.b32 %r345, %r344, 4096;
or.b32 %r346, %r343, %r345;
shl.b32 %r347, %r322, 2;
and.b32 %r348, %r347, 1048576;
or.b32 %r349, %r346, %r348;
shl.b32 %r350, %r322, 9;
and.b32 %r351, %r350, 268435456;
or.b32 %r302, %r349, %r351;

dp4a.s32.s32 %r301, %r302, %r8, %r297;

and.b32 %r352, %r323, 252645135;
shl.b32 %r353, %r325, 4;
and.b32 %r354, %r353, 16;
or.b32 %r355, %r354, %r352;
shl.b32 %r356, %r325, 11;
and.b32 %r357, %r356, 4096;
or.b32 %r358, %r355, %r357;
shl.b32 %r359, %r325, 18;
and.b32 %r360, %r359, 1048576;
or.b32 %r361, %r358, %r360;
shl.b32 %r362, %r325, 25;
and.b32 %r363, %r362, 268435456;
or.b32 %r306, %r361, %r363;

dp4a.s32.s32 %r305, %r306, %r9, %r301;

shr.u32 %r364, %r323, 4;
and.b32 %r365, %r364, 252645135;
shr.u32 %r366, %r325, 12;
and.b32 %r367, %r366, 16;
or.b32 %r368, %r367, %r365;
shr.u32 %r369, %r325, 5;
and.b32 %r370, %r369, 4096;
or.b32 %r371, %r368, %r370;
shl.b32 %r372, %r325, 2;
and.b32 %r373, %r372, 1048576;
or.b32 %r374, %r371, %r373;
shl.b32 %r375, %r325, 9;
and.b32 %r376, %r375, 268435456;
or.b32 %r310, %r374, %r376;

dp4a.s32.s32 %r309, %r310, %r10, %r305;

ld.global.u32 %r314, [%rd63];

{mul.f16x2 %r313,%r314,%r11;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r313;
cvt.f32.f16 %f60, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r313;
cvt.f32.f16 %f61, high;}


cvt.rn.f32.s32 %f62, %r309;
mov.f32 %f63, 0f40000000;
div.approx.ftz.f32 %f64, %f61, %f63;
fma.rn.ftz.f32 %f65, %f60, %f62, %f64;
add.ftz.f32 %f170, %f65, %f170;

$L__BB71_7:
add.s64 %rd100, %rd100, 576;
add.s64 %rd99, %rd99, 384;
add.s64 %rd98, %rd98, 576;
add.s64 %rd97, %rd97, 384;
add.s32 %r476, %r476, 16;
setp.lt.u32 %p5, %r476, %r31;
@%p5 bra $L__BB71_3;

add.u64 %rd89, %SPL, 0;
add.u64 %rd78, %SPL, 8;
st.local.v2.f32 [%rd89], {%f173, %f172};
st.local.v2.f32 [%rd78], {%f171, %f170};

$L__BB71_9:
mov.b32 %r13, %f173;
setp.ne.s64 %p6, %rd26, 0;
@%p6 bra $L__BB71_11;
bra.uni $L__BB71_10;

$L__BB71_11:
add.u64 %rd95, %SPL, 0;
add.u64 %rd87, %SPL, 8;
mov.u32 %r404, 31;
mov.u32 %r405, 16;
mov.u32 %r406, -1;
shfl.sync.bfly.b32 %r407|%p17, %r13, %r405, %r404, %r406;
mov.b32 %f86, %r407;
add.ftz.f32 %f87, %f173, %f86;
mov.b32 %r408, %f87;
mov.u32 %r409, 8;
shfl.sync.bfly.b32 %r410|%p18, %r408, %r409, %r404, %r406;
mov.b32 %f88, %r410;
add.ftz.f32 %f89, %f87, %f88;
mov.b32 %r411, %f89;
mov.u32 %r412, 4;
shfl.sync.bfly.b32 %r413|%p19, %r411, %r412, %r404, %r406;
mov.b32 %f90, %r413;
add.ftz.f32 %f91, %f89, %f90;
mov.b32 %r414, %f91;
mov.u32 %r415, 2;
shfl.sync.bfly.b32 %r416|%p20, %r414, %r415, %r404, %r406;
mov.b32 %f92, %r416;
add.ftz.f32 %f93, %f91, %f92;
mov.b32 %r417, %f93;
mov.u32 %r418, 1;
shfl.sync.bfly.b32 %r419|%p21, %r417, %r418, %r404, %r406;
mov.b32 %f94, %r419;
add.ftz.f32 %f95, %f93, %f94;
st.local.f32 [%rd95], %f95;
mov.b32 %r420, %f171;
shfl.sync.bfly.b32 %r421|%p22, %r420, %r405, %r404, %r406;
mov.b32 %f96, %r421;
add.ftz.f32 %f97, %f171, %f96;
mov.b32 %r422, %f97;
shfl.sync.bfly.b32 %r423|%p23, %r422, %r409, %r404, %r406;
mov.b32 %f98, %r423;
add.ftz.f32 %f99, %f97, %f98;
mov.b32 %r424, %f99;
shfl.sync.bfly.b32 %r425|%p24, %r424, %r412, %r404, %r406;
mov.b32 %f100, %r425;
add.ftz.f32 %f101, %f99, %f100;
mov.b32 %r426, %f101;
shfl.sync.bfly.b32 %r427|%p25, %r426, %r415, %r404, %r406;
mov.b32 %f102, %r427;
add.ftz.f32 %f103, %f101, %f102;
mov.b32 %r428, %f103;
shfl.sync.bfly.b32 %r429|%p26, %r428, %r418, %r404, %r406;
mov.b32 %f104, %r429;
add.ftz.f32 %f105, %f103, %f104;
st.local.f32 [%rd87], %f105;
mov.b32 %r430, %f172;
shfl.sync.bfly.b32 %r431|%p27, %r430, %r405, %r404, %r406;
mov.b32 %f106, %r431;
add.ftz.f32 %f107, %f172, %f106;
mov.b32 %r432, %f107;
shfl.sync.bfly.b32 %r433|%p28, %r432, %r409, %r404, %r406;
mov.b32 %f108, %r433;
add.ftz.f32 %f109, %f107, %f108;
mov.b32 %r434, %f109;
shfl.sync.bfly.b32 %r435|%p29, %r434, %r412, %r404, %r406;
mov.b32 %f110, %r435;
add.ftz.f32 %f111, %f109, %f110;
mov.b32 %r436, %f111;
shfl.sync.bfly.b32 %r437|%p30, %r436, %r415, %r404, %r406;
mov.b32 %f112, %r437;
add.ftz.f32 %f113, %f111, %f112;
mov.b32 %r438, %f113;
shfl.sync.bfly.b32 %r439|%p31, %r438, %r418, %r404, %r406;
mov.b32 %f114, %r439;
add.ftz.f32 %f115, %f113, %f114;
st.local.f32 [%rd95+4], %f115;
mov.b32 %r440, %f170;
shfl.sync.bfly.b32 %r441|%p32, %r440, %r405, %r404, %r406;
mov.b32 %f116, %r441;
add.ftz.f32 %f117, %f170, %f116;
mov.b32 %r442, %f117;
shfl.sync.bfly.b32 %r443|%p33, %r442, %r409, %r404, %r406;
mov.b32 %f118, %r443;
add.ftz.f32 %f119, %f117, %f118;
mov.b32 %r444, %f119;
shfl.sync.bfly.b32 %r445|%p34, %r444, %r412, %r404, %r406;
mov.b32 %f120, %r445;
add.ftz.f32 %f121, %f119, %f120;
mov.b32 %r446, %f121;
shfl.sync.bfly.b32 %r447|%p35, %r446, %r415, %r404, %r406;
mov.b32 %f122, %r447;
add.ftz.f32 %f123, %f121, %f122;
mov.b32 %r448, %f123;
shfl.sync.bfly.b32 %r449|%p36, %r448, %r418, %r404, %r406;
mov.b32 %f124, %r449;
add.ftz.f32 %f125, %f123, %f124;
st.local.f32 [%rd87+4], %f125;
bra.uni $L__BB71_12;

$L__BB71_10:
add.u64 %rd91, %SPL, 0;
mov.u32 %r378, 31;
mov.u32 %r379, 16;
mov.u32 %r380, -1;
shfl.sync.bfly.b32 %r381|%p7, %r13, %r379, %r378, %r380;
mov.b32 %f66, %r381;
add.ftz.f32 %f67, %f173, %f66;
mov.b32 %r382, %f67;
mov.u32 %r383, 8;
shfl.sync.bfly.b32 %r384|%p8, %r382, %r383, %r378, %r380;
mov.b32 %f68, %r384;
add.ftz.f32 %f69, %f67, %f68;
mov.b32 %r385, %f69;
mov.u32 %r386, 4;
shfl.sync.bfly.b32 %r387|%p9, %r385, %r386, %r378, %r380;
mov.b32 %f70, %r387;
add.ftz.f32 %f71, %f69, %f70;
mov.b32 %r388, %f71;
mov.u32 %r389, 2;
shfl.sync.bfly.b32 %r390|%p10, %r388, %r389, %r378, %r380;
mov.b32 %f72, %r390;
add.ftz.f32 %f73, %f71, %f72;
mov.b32 %r391, %f73;
mov.u32 %r392, 1;
shfl.sync.bfly.b32 %r393|%p11, %r391, %r392, %r378, %r380;
mov.b32 %f74, %r393;
add.ftz.f32 %f75, %f73, %f74;
st.local.f32 [%rd91], %f75;
mov.b32 %r394, %f172;
shfl.sync.bfly.b32 %r395|%p12, %r394, %r379, %r378, %r380;
mov.b32 %f76, %r395;
add.ftz.f32 %f77, %f172, %f76;
mov.b32 %r396, %f77;
shfl.sync.bfly.b32 %r397|%p13, %r396, %r383, %r378, %r380;
mov.b32 %f78, %r397;
add.ftz.f32 %f79, %f77, %f78;
mov.b32 %r398, %f79;
shfl.sync.bfly.b32 %r399|%p14, %r398, %r386, %r378, %r380;
mov.b32 %f80, %r399;
add.ftz.f32 %f81, %f79, %f80;
mov.b32 %r400, %f81;
shfl.sync.bfly.b32 %r401|%p15, %r400, %r389, %r378, %r380;
mov.b32 %f82, %r401;
add.ftz.f32 %f83, %f81, %f82;
mov.b32 %r402, %f83;
shfl.sync.bfly.b32 %r403|%p16, %r402, %r392, %r378, %r380;
mov.b32 %f84, %r403;
add.ftz.f32 %f85, %f83, %f84;
st.local.f32 [%rd91+4], %f85;

$L__BB71_12:
ld.param.u32 %r459, [_Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7];
setp.gt.u32 %p37, %r4, 1;
mov.u32 %r451, %ctaid.x;
shl.b32 %r452, %r451, 1;
add.s32 %r14, %r452, %r4;
setp.ge.u32 %p38, %r14, %r459;
or.pred %p39, %p37, %p38;
@%p39 bra $L__BB71_29;

add.u64 %rd93, %SPL, 0;
ld.param.u64 %rd80, [_Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3];
ld.param.u32 %r460, [_Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13];
cvt.u64.u32 %rd20, %r4;
mul.wide.u32 %rd65, %r4, 4;
add.s64 %rd66, %rd93, %rd65;
ld.local.f32 %f176, [%rd66];
mad.lo.s32 %r15, %r3, %r460, %r14;
setp.eq.s64 %p40, %rd80, 0;
@%p40 bra $L__BB71_15;

ld.param.u64 %rd81, [_Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3];
cvta.to.global.u64 %rd67, %rd81;
mul.wide.u32 %rd68, %r15, 4;
add.s64 %rd69, %rd67, %rd68;
ld.global.f32 %f126, [%rd69];
add.ftz.f32 %f176, %f176, %f126;

$L__BB71_15:
setp.eq.s64 %p41, %rd26, 0;
@%p41 bra $L__BB71_28;

ld.param.u64 %rd85, [_Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+16];
add.u64 %rd83, %SPL, 8;
shl.b64 %rd70, %rd20, 2;
add.s64 %rd71, %rd83, %rd70;
ld.local.f32 %f24, [%rd71];
setp.eq.s64 %p42, %rd85, 0;
@%p42 bra $L__BB71_18;

ld.param.u64 %rd86, [_Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+16];
cvta.to.global.u64 %rd72, %rd86;
mul.wide.u32 %rd73, %r15, 4;
add.s64 %rd74, %rd72, %rd73;
ld.global.f32 %f127, [%rd74];
add.ftz.f32 %f24, %f24, %f127;

$L__BB71_18:
ld.param.u32 %r465, [_Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.gt.s32 %p43, %r465, 2;
@%p43 bra $L__BB71_22;

ld.param.u32 %r468, [_Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.eq.s32 %p46, %r468, 1;
@%p46 bra $L__BB71_26;

ld.param.u32 %r469, [_Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.eq.s32 %p47, %r469, 2;
@%p47 bra $L__BB71_21;
bra.uni $L__BB71_27;

$L__BB71_21:
mul.ftz.f32 %f147, %f24, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f148, %f147;
add.ftz.f32 %f149, %f148, 0f3F800000;
div.approx.ftz.f32 %f150, %f24, %f149;
mul.ftz.f32 %f176, %f176, %f150;
bra.uni $L__BB71_28;

$L__BB71_22:
ld.param.u32 %r466, [_Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.eq.s32 %p44, %r466, 3;
@%p44 bra $L__BB71_25;

ld.param.u32 %r467, [_Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.ne.s32 %p45, %r467, 6;
@%p45 bra $L__BB71_27;

ld.param.f32 %f159, [_Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+44];
min.ftz.f32 %f128, %f24, %f159;
neg.ftz.f32 %f129, %f159;
min.ftz.f32 %f130, %f176, %f159;
max.ftz.f32 %f131, %f130, %f129;
mul.ftz.f32 %f132, %f128, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f133, %f132;
add.ftz.f32 %f134, %f133, 0f3F800000;
div.approx.ftz.f32 %f135, %f128, %f134;
mul.ftz.f32 %f176, %f131, %f135;
bra.uni $L__BB71_28;

$L__BB71_27:
mul.ftz.f32 %f176, %f176, %f24;
bra.uni $L__BB71_28;

$L__BB71_26:
mul.ftz.f32 %f151, %f24, 0f3F000000;
mul.ftz.f32 %f152, %f24, 0f3D372713;
fma.rn.ftz.f32 %f153, %f24, %f152, 0f3F800000;
mul.ftz.f32 %f154, %f24, 0f3F4C422A;
mul.ftz.f32 %f155, %f154, %f153;
tanh.approx.f32 %f156, %f155;
add.ftz.f32 %f157, %f156, 0f3F800000;
mul.ftz.f32 %f158, %f151, %f157;
mul.ftz.f32 %f176, %f176, %f158;
bra.uni $L__BB71_28;

$L__BB71_25:
mov.f32 %f136, 0f40E00000;
min.ftz.f32 %f137, %f24, %f136;
min.ftz.f32 %f138, %f176, %f136;
mov.f32 %f139, 0fC0E00000;
max.ftz.f32 %f140, %f138, %f139;
mul.ftz.f32 %f141, %f137, 0f3FD9DB23;
mul.ftz.f32 %f142, %f141, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f143, %f142;
add.ftz.f32 %f144, %f143, 0f3F800000;
div.approx.ftz.f32 %f145, %f137, %f144;
add.ftz.f32 %f146, %f140, 0f3F800000;
mul.ftz.f32 %f176, %f146, %f145;

$L__BB71_28:
ld.param.u64 %rd82, [_Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4];
ld.param.u32 %r464, [_Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10];
ld.param.u32 %r463, [_Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13];
mov.u32 %r462, %ctaid.y;
mov.u32 %r461, %tid.y;
mad.lo.s32 %r457, %r461, %r464, %r14;
mad.lo.s32 %r458, %r462, %r463, %r457;
cvta.to.global.u64 %rd75, %rd82;
mul.wide.u32 %rd76, %r458, 4;
add.s64 %rd77, %rd75, %rd76;
st.global.f32 [%rd77], %f176;

$L__BB71_29:
ret;

}
.entry _Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj(
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2,
.param .align 8 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3[48],
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5,
.param .align 4 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6[12],
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15
)
.maxntid 256, 1, 1
.minnctapersm 1
{
.local .align 8 .b8 __local_depot72[8];
.reg .b64 %SP;
.reg .b64 %SPL;
.reg .pred %p<17>;
.reg .f32 %f<47>;
.reg .b32 %r<227>;
.reg .b64 %rd<54>;


mov.u64 %SPL, __local_depot72;
ld.param.u64 %rd17, [_Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0];
ld.param.u64 %rd18, [_Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1];
ld.param.u64 %rd19, [_Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2];
ld.param.u64 %rd20, [_Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4];
ld.param.u32 %r22, [_Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5];
ld.param.u32 %r14, [_Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7];
ld.param.u32 %r15, [_Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8];
ld.param.u32 %r16, [_Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9];
ld.param.u32 %r17, [_Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10];
ld.param.u32 %r18, [_Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11];
ld.param.u32 %r19, [_Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12];
ld.param.u32 %r20, [_Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13];
ld.param.u32 %r23, [_Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14];
ld.param.u32 %r21, [_Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15];
ld.param.u32 %r13, [_Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+8];
ld.param.u32 %r12, [_Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+4];
ld.param.u32 %r11, [_Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6];
add.u64 %rd1, %SPL, 0;
mov.u32 %r24, %ctaid.x;
shl.b32 %r1, %r24, 1;
shr.u32 %r2, %r22, 5;
mov.u32 %r3, %tid.y;
setp.ge.u32 %p1, %r3, %r23;
@%p1 bra $L__BB72_7;

mov.f32 %f45, 0f00000000;
st.local.v2.f32 [%rd1], {%f45, %f45};
mov.u32 %r4, %tid.x;
shr.u32 %r226, %r4, 1;
setp.ge.u32 %p2, %r226, %r2;
mov.f32 %f46, %f45;
@%p2 bra $L__BB72_5;

mov.u32 %r25, %ctaid.y;
mul.hi.u32 %r26, %r25, %r11;
add.s32 %r27, %r26, %r25;
shr.u32 %r28, %r27, %r12;
mul.lo.s32 %r29, %r28, %r13;
sub.s32 %r30, %r25, %r29;
mul.lo.s32 %r31, %r30, %r19;
cvt.u64.u32 %rd22, %r31;
mul.lo.s32 %r32, %r3, %r16;
cvt.u64.u32 %rd23, %r32;
mad.lo.s32 %r33, %r3, %r21, %r25;
cvta.to.global.u64 %rd24, %rd19;
mul.wide.u32 %rd25, %r33, 4;
add.s64 %rd26, %rd24, %rd25;
ld.global.nc.u32 %r34, [%rd26];
mul.lo.s32 %r35, %r34, %r18;
shl.b32 %r36, %r4, 3;
and.b32 %r6, %r36, 8;
or.b32 %r7, %r6, 4;
cvt.u64.u32 %rd27, %r4;
shr.u64 %rd28, %rd27, 1;
mul.lo.s64 %rd29, %rd28, 36;
mul.wide.u32 %rd30, %r32, 36;
add.s64 %rd31, %rd29, %rd30;
cvta.to.global.u64 %rd32, %rd18;
add.s64 %rd3, %rd32, %rd31;
mul.wide.u32 %rd53, %r31, 36;
and.b32 %r37, %r4, 1;
mul.wide.u32 %rd33, %r37, 8;
mad.lo.s32 %r38, %r1, %r15, %r226;
add.s32 %r39, %r38, %r35;
mul.wide.s32 %rd5, %r39, 24;
add.s64 %rd34, %rd33, %rd5;
add.s64 %rd6, %rd34, 12;
add.s64 %rd35, %rd28, %rd23;
add.s64 %rd36, %rd35, %rd22;
mul.lo.s64 %rd37, %rd36, 36;
add.s64 %rd38, %rd37, %rd33;
add.s64 %rd39, %rd32, %rd38;
add.s64 %rd51, %rd39, 20;
add.s32 %r40, %r1, 1;
add.s32 %r41, %r226, %r35;
mad.lo.s32 %r42, %r15, %r40, %r41;
mul.wide.s32 %rd8, %r42, 24;
add.s64 %rd40, %rd33, %rd8;
add.s64 %rd9, %rd40, 12;
cvta.to.global.u64 %rd52, %rd17;
mov.f32 %f45, 0f00000000;

$L__BB72_3:
add.s64 %rd41, %rd3, %rd53;
add.s64 %rd42, %rd52, %rd5;
ld.global.nc.u32 %r85, [%rd42+4];
shr.s32 %r86, %r85, %r6;
ld.global.nc.u32 %r45, [%rd51+-16];
ld.global.nc.u32 %r49, [%rd51];
add.s64 %rd43, %rd52, %rd6;
ld.global.nc.u32 %r87, [%rd43];
shr.s32 %r88, %r85, %r7;
ld.global.nc.u32 %r53, [%rd51+-12];
ld.global.nc.u32 %r57, [%rd51+4];
ld.global.nc.u32 %r89, [%rd43+-4];
and.b32 %r90, %r89, 252645135;
shl.b32 %r91, %r86, 4;
and.b32 %r92, %r91, 16;
or.b32 %r93, %r92, %r90;
shl.b32 %r94, %r86, 11;
and.b32 %r95, %r94, 4096;
or.b32 %r96, %r93, %r95;
shl.b32 %r97, %r86, 18;
and.b32 %r98, %r97, 1048576;
or.b32 %r99, %r96, %r98;
shl.b32 %r100, %r86, 25;
and.b32 %r101, %r100, 268435456;
or.b32 %r44, %r99, %r101;
mov.u32 %r67, 0;

dp4a.s32.s32 %r43, %r44, %r45, %r67;

shr.u32 %r102, %r89, 4;
and.b32 %r103, %r102, 252645135;
shr.u32 %r104, %r86, 12;
and.b32 %r105, %r104, 16;
or.b32 %r106, %r105, %r103;
shr.u32 %r107, %r86, 5;
and.b32 %r108, %r107, 4096;
or.b32 %r109, %r106, %r108;
shl.b32 %r110, %r86, 2;
and.b32 %r111, %r110, 1048576;
or.b32 %r112, %r109, %r111;
shl.b32 %r113, %r86, 9;
and.b32 %r114, %r113, 268435456;
or.b32 %r48, %r112, %r114;

dp4a.s32.s32 %r47, %r48, %r49, %r43;

and.b32 %r115, %r87, 252645135;
shl.b32 %r116, %r88, 4;
and.b32 %r117, %r116, 16;
or.b32 %r118, %r117, %r115;
shl.b32 %r119, %r88, 11;
and.b32 %r120, %r119, 4096;
or.b32 %r121, %r118, %r120;
shl.b32 %r122, %r88, 18;
and.b32 %r123, %r122, 1048576;
or.b32 %r124, %r121, %r123;
shl.b32 %r125, %r88, 25;
and.b32 %r126, %r125, 268435456;
or.b32 %r52, %r124, %r126;

dp4a.s32.s32 %r51, %r52, %r53, %r47;

shr.u32 %r127, %r87, 4;
and.b32 %r128, %r127, 252645135;
shr.u32 %r129, %r88, 12;
and.b32 %r130, %r129, 16;
or.b32 %r131, %r130, %r128;
shr.u32 %r132, %r88, 5;
and.b32 %r133, %r132, 4096;
or.b32 %r134, %r131, %r133;
shl.b32 %r135, %r88, 2;
and.b32 %r136, %r135, 1048576;
or.b32 %r137, %r134, %r136;
shl.b32 %r138, %r88, 9;
and.b32 %r139, %r138, 268435456;
or.b32 %r56, %r137, %r139;

dp4a.s32.s32 %r55, %r56, %r57, %r51;

ld.global.nc.u32 %r60, [%rd42];
ld.global.nc.u32 %r61, [%rd41];

{mul.f16x2 %r59,%r60,%r61;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r59;
cvt.f32.f16 %f11, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r59;
cvt.f32.f16 %f12, high;}


cvt.rn.f32.s32 %f15, %r55;
mov.f32 %f16, 0f40000000;
div.approx.ftz.f32 %f17, %f12, %f16;
fma.rn.ftz.f32 %f18, %f11, %f15, %f17;
add.ftz.f32 %f46, %f18, %f46;
add.s64 %rd44, %rd52, %rd8;
ld.global.nc.u32 %r140, [%rd44+4];
shr.s32 %r141, %r140, %r6;
add.s64 %rd45, %rd52, %rd9;
ld.global.nc.u32 %r142, [%rd45];
shr.s32 %r143, %r140, %r7;
ld.global.nc.u32 %r144, [%rd45+-4];
and.b32 %r145, %r144, 252645135;
shl.b32 %r146, %r141, 4;
and.b32 %r147, %r146, 16;
or.b32 %r148, %r147, %r145;
shl.b32 %r149, %r141, 11;
and.b32 %r150, %r149, 4096;
or.b32 %r151, %r148, %r150;
shl.b32 %r152, %r141, 18;
and.b32 %r153, %r152, 1048576;
or.b32 %r154, %r151, %r153;
shl.b32 %r155, %r141, 25;
and.b32 %r156, %r155, 268435456;
or.b32 %r65, %r154, %r156;

dp4a.s32.s32 %r64, %r65, %r45, %r67;

shr.u32 %r157, %r144, 4;
and.b32 %r158, %r157, 252645135;
shr.u32 %r159, %r141, 12;
and.b32 %r160, %r159, 16;
or.b32 %r161, %r160, %r158;
shr.u32 %r162, %r141, 5;
and.b32 %r163, %r162, 4096;
or.b32 %r164, %r161, %r163;
shl.b32 %r165, %r141, 2;
and.b32 %r166, %r165, 1048576;
or.b32 %r167, %r164, %r166;
shl.b32 %r168, %r141, 9;
and.b32 %r169, %r168, 268435456;
or.b32 %r69, %r167, %r169;

dp4a.s32.s32 %r68, %r69, %r49, %r64;

and.b32 %r170, %r142, 252645135;
shl.b32 %r171, %r143, 4;
and.b32 %r172, %r171, 16;
or.b32 %r173, %r172, %r170;
shl.b32 %r174, %r143, 11;
and.b32 %r175, %r174, 4096;
or.b32 %r176, %r173, %r175;
shl.b32 %r177, %r143, 18;
and.b32 %r178, %r177, 1048576;
or.b32 %r179, %r176, %r178;
shl.b32 %r180, %r143, 25;
and.b32 %r181, %r180, 268435456;
or.b32 %r73, %r179, %r181;

dp4a.s32.s32 %r72, %r73, %r53, %r68;

shr.u32 %r182, %r142, 4;
and.b32 %r183, %r182, 252645135;
shr.u32 %r184, %r143, 12;
and.b32 %r185, %r184, 16;
or.b32 %r186, %r185, %r183;
shr.u32 %r187, %r143, 5;
and.b32 %r188, %r187, 4096;
or.b32 %r189, %r186, %r188;
shl.b32 %r190, %r143, 2;
and.b32 %r191, %r190, 1048576;
or.b32 %r192, %r189, %r191;
shl.b32 %r193, %r143, 9;
and.b32 %r194, %r193, 268435456;
or.b32 %r77, %r192, %r194;

dp4a.s32.s32 %r76, %r77, %r57, %r72;

ld.global.nc.u32 %r81, [%rd44];

{mul.f16x2 %r80,%r81,%r61;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r80;
cvt.f32.f16 %f13, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r80;
cvt.f32.f16 %f14, high;}


cvt.rn.f32.s32 %f19, %r76;
div.approx.ftz.f32 %f20, %f14, %f16;
fma.rn.ftz.f32 %f21, %f13, %f19, %f20;
add.ftz.f32 %f45, %f21, %f45;
add.s64 %rd53, %rd53, 576;
add.s64 %rd52, %rd52, 384;
add.s64 %rd51, %rd51, 576;
add.s32 %r226, %r226, 16;
setp.lt.u32 %p3, %r226, %r2;
@%p3 bra $L__BB72_3;

st.local.v2.f32 [%rd1], {%f46, %f45};

$L__BB72_5:
mov.b32 %r195, %f46;
mov.u32 %r196, 31;
mov.u32 %r197, 16;
mov.u32 %r198, -1;
shfl.sync.bfly.b32 %r199|%p4, %r195, %r197, %r196, %r198;
mov.b32 %f22, %r199;
add.ftz.f32 %f23, %f46, %f22;
mov.b32 %r200, %f23;
mov.u32 %r201, 8;
shfl.sync.bfly.b32 %r202|%p5, %r200, %r201, %r196, %r198;
mov.b32 %f24, %r202;
add.ftz.f32 %f25, %f23, %f24;
mov.b32 %r203, %f25;
mov.u32 %r204, 4;
shfl.sync.bfly.b32 %r205|%p6, %r203, %r204, %r196, %r198;
mov.b32 %f26, %r205;
add.ftz.f32 %f27, %f25, %f26;
mov.b32 %r206, %f27;
mov.u32 %r207, 2;
shfl.sync.bfly.b32 %r208|%p7, %r206, %r207, %r196, %r198;
mov.b32 %f28, %r208;
add.ftz.f32 %f29, %f27, %f28;
mov.b32 %r209, %f29;
mov.u32 %r210, 1;
shfl.sync.bfly.b32 %r211|%p8, %r209, %r210, %r196, %r198;
mov.b32 %f30, %r211;
add.ftz.f32 %f31, %f29, %f30;
st.local.f32 [%rd1], %f31;
mov.b32 %r212, %f45;
shfl.sync.bfly.b32 %r213|%p9, %r212, %r197, %r196, %r198;
mov.b32 %f32, %r213;
add.ftz.f32 %f33, %f45, %f32;
mov.b32 %r214, %f33;
shfl.sync.bfly.b32 %r215|%p10, %r214, %r201, %r196, %r198;
mov.b32 %f34, %r215;
add.ftz.f32 %f35, %f33, %f34;
mov.b32 %r216, %f35;
shfl.sync.bfly.b32 %r217|%p11, %r216, %r204, %r196, %r198;
mov.b32 %f36, %r217;
add.ftz.f32 %f37, %f35, %f36;
mov.b32 %r218, %f37;
shfl.sync.bfly.b32 %r219|%p12, %r218, %r207, %r196, %r198;
mov.b32 %f38, %r219;
add.ftz.f32 %f39, %f37, %f38;
mov.b32 %r220, %f39;
shfl.sync.bfly.b32 %r221|%p13, %r220, %r210, %r196, %r198;
mov.b32 %f40, %r221;
add.ftz.f32 %f41, %f39, %f40;
st.local.f32 [%rd1+4], %f41;
add.s32 %r10, %r1, %r4;
setp.ge.u32 %p14, %r10, %r14;
setp.gt.u32 %p15, %r4, 1;
or.pred %p16, %p15, %p14;
@%p16 bra $L__BB72_7;

mul.wide.u32 %rd46, %r4, 4;
add.s64 %rd47, %rd1, %rd46;
ld.local.f32 %f42, [%rd47];
mov.u32 %r222, %ctaid.y;
mad.lo.s32 %r224, %r3, %r17, %r10;
mad.lo.s32 %r225, %r222, %r20, %r224;
cvta.to.global.u64 %rd48, %rd20;
mul.wide.u32 %rd49, %r225, 4;
add.s64 %rd50, %rd48, %rd49;
st.global.f32 [%rd50], %f42;

$L__BB72_7:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<320>;
.reg .f32 %f<1021>;
.reg .b32 %r<1475>;
.reg .b64 %rd<92>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate[1536];

ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd20, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r45, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r49, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r50, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r51, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r55, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r56, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r57, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r53, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r52, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r47, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r46, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
ld.param.u64 %rd23, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+16];
ld.param.u64 %rd22, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+8];
ld.param.u64 %rd21, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3];
mov.u32 %r1, %tid.y;
shl.b32 %r58, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r58, %r2;
mov.u32 %r59, %ctaid.x;
shl.b32 %r4, %r59, 2;
mov.u32 %r5, %ctaid.y;
setp.eq.s64 %p1, %rd20, 0;
@%p1 bra $L__BB73_2;

cvta.to.global.u64 %rd27, %rd20;
mul.wide.u32 %rd28, %r5, 4;
add.s64 %rd29, %rd27, %rd28;
ld.global.nc.u32 %r1472, [%rd29];
mul.hi.u32 %r60, %r5, %r41;
add.s32 %r61, %r60, %r5;
shr.u32 %r62, %r61, %r42;
mul.lo.s32 %r63, %r62, %r43;
sub.s32 %r1473, %r5, %r63;
bra.uni $L__BB73_3;

$L__BB73_2:
mul.hi.u32 %r64, %r5, %r46;
add.s32 %r65, %r64, %r5;
shr.u32 %r1472, %r65, %r47;
mov.u32 %r1473, %r5;

$L__BB73_3:
mov.u32 %r11, %ctaid.z;
selp.b32 %r12, %r5, %r1472, %p1;
setp.ne.s32 %p3, %r1, 0;
setp.gt.u32 %p4, %r2, 3;
or.pred %p5, %p4, %p3;
add.s32 %r66, %r4, %r2;
setp.ge.u32 %p6, %r66, %r45;
mov.f32 %f26, 0f00000000;
or.pred %p7, %p5, %p6;
mov.f32 %f4, %f26;
mov.f32 %f5, %f26;
@%p7 bra $L__BB73_8;

setp.eq.s64 %p8, %rd21, 0;
mov.f32 %f4, 0f00000000;
mov.f32 %f5, %f4;
@%p8 bra $L__BB73_6;

cvta.to.global.u64 %rd30, %rd21;
mul.lo.s32 %r67, %r11, %r57;
cvt.u64.u32 %rd31, %r67;
mul.lo.s32 %r68, %r12, %r51;
cvt.u64.u32 %rd32, %r68;
cvt.u64.u32 %rd33, %r2;
cvt.s64.s32 %rd34, %r4;
add.s64 %rd35, %rd34, %rd33;
add.s64 %rd36, %rd35, %rd31;
add.s64 %rd37, %rd36, %rd32;
shl.b64 %rd38, %rd37, 2;
add.s64 %rd39, %rd30, %rd38;
ld.global.f32 %f5, [%rd39];

$L__BB73_6:
setp.eq.s64 %p9, %rd23, 0;
setp.eq.s64 %p10, %rd22, 0;
or.pred %p11, %p9, %p10;
@%p11 bra $L__BB73_8;

cvta.to.global.u64 %rd40, %rd23;
mul.lo.s32 %r69, %r11, %r57;
cvt.u64.u32 %rd41, %r69;
mul.lo.s32 %r70, %r12, %r51;
cvt.u64.u32 %rd42, %r70;
cvt.u64.u32 %rd43, %r2;
cvt.s64.s32 %rd44, %r4;
add.s64 %rd45, %rd44, %rd43;
add.s64 %rd46, %rd45, %rd41;
add.s64 %rd47, %rd46, %rd42;
shl.b64 %rd48, %rd47, 2;
add.s64 %rd49, %rd40, %rd48;
ld.global.f32 %f4, [%rd49];

$L__BB73_8:
shr.u32 %r71, %r3, 31;
add.s32 %r72, %r3, %r71;
shr.s32 %r1474, %r72, 1;
shr.u32 %r73, %r40, 5;
setp.ge.s32 %p12, %r1474, %r73;
mov.f32 %f27, %f26;
mov.f32 %f28, %f26;
mov.f32 %f29, %f26;
mov.f32 %f30, %f26;
mov.f32 %f31, %f26;
mov.f32 %f32, %f26;
mov.f32 %f33, %f26;
@%p12 bra $L__BB73_19;

mul.hi.u32 %r74, %r11, %r52;
add.s32 %r75, %r74, %r11;
shr.u32 %r76, %r75, %r53;
mul.lo.s32 %r77, %r76, %r55;
mul.lo.s32 %r78, %r1472, %r49;
and.b32 %r81, %r72, -2;
sub.s32 %r14, %r3, %r81;
add.s32 %r82, %r1474, %r77;
add.s32 %r15, %r82, %r78;
mad.lo.s32 %r83, %r4, %r44, %r1474;
add.s32 %r84, %r83, %r77;
add.s32 %r16, %r84, %r78;
cvta.to.global.u64 %rd91, %rd22;
cvta.to.global.u64 %rd90, %rd18;
mul.lo.s32 %r85, %r1473, %r50;
cvt.u64.u32 %rd50, %r85;
mul.lo.s32 %r86, %r11, %r56;
cvt.u64.u32 %rd51, %r86;
add.s64 %rd6, %rd50, %rd51;
mov.f32 %f26, 0f00000000;
cvta.to.global.u64 %rd52, %rd19;
mov.f32 %f27, %f26;
mov.f32 %f28, %f26;
mov.f32 %f29, %f26;
mov.f32 %f30, %f26;
mov.f32 %f31, %f26;
mov.f32 %f32, %f26;
mov.f32 %f33, %f26;

$L__BB73_10:
cvt.u64.u32 %rd53, %r1474;
add.s64 %rd54, %rd6, %rd53;
mul.wide.s32 %rd55, %r16, 24;
add.s64 %rd9, %rd90, %rd55;
mul.lo.s64 %rd56, %rd54, 36;
add.s64 %rd57, %rd52, %rd56;
shl.b32 %r108, %r14, 1;
mul.wide.s32 %rd58, %r108, 4;
add.s64 %rd59, %rd57, %rd58;
add.s64 %rd60, %rd55, %rd58;
add.s64 %rd61, %rd90, %rd60;
add.s64 %rd10, %rd61, 8;
ld.global.nc.u32 %r109, [%rd9+4];
shl.b32 %r110, %r14, 3;
shr.s32 %r111, %r109, %r110;
ld.global.nc.u32 %r18, [%rd59+4];
ld.global.nc.u32 %r19, [%rd59+20];
ld.global.nc.u32 %r112, [%rd61+12];
or.b32 %r113, %r110, 4;
shr.s32 %r114, %r109, %r113;
ld.global.nc.u32 %r20, [%rd59+8];
ld.global.nc.u32 %r21, [%rd59+24];
ld.global.nc.u32 %r115, [%rd61+8];
and.b32 %r116, %r115, 252645135;
shl.b32 %r117, %r111, 4;
and.b32 %r118, %r117, 16;
or.b32 %r119, %r118, %r116;
shl.b32 %r120, %r111, 11;
and.b32 %r121, %r120, 4096;
or.b32 %r122, %r119, %r121;
shl.b32 %r123, %r111, 18;
and.b32 %r124, %r123, 1048576;
or.b32 %r125, %r122, %r124;
shl.b32 %r126, %r111, 25;
and.b32 %r127, %r126, 268435456;
or.b32 %r88, %r125, %r127;
mov.u32 %r90, 0;

dp4a.s32.s32 %r87, %r88, %r18, %r90;

shr.u32 %r128, %r115, 4;
and.b32 %r129, %r128, 252645135;
shr.u32 %r130, %r111, 12;
and.b32 %r131, %r130, 16;
or.b32 %r132, %r131, %r129;
shr.u32 %r133, %r111, 5;
and.b32 %r134, %r133, 4096;
or.b32 %r135, %r132, %r134;
shl.b32 %r136, %r111, 2;
and.b32 %r137, %r136, 1048576;
or.b32 %r138, %r135, %r137;
shl.b32 %r139, %r111, 9;
and.b32 %r140, %r139, 268435456;
or.b32 %r92, %r138, %r140;

dp4a.s32.s32 %r91, %r92, %r19, %r87;

and.b32 %r141, %r112, 252645135;
shl.b32 %r142, %r114, 4;
and.b32 %r143, %r142, 16;
or.b32 %r144, %r143, %r141;
shl.b32 %r145, %r114, 11;
and.b32 %r146, %r145, 4096;
or.b32 %r147, %r144, %r146;
shl.b32 %r148, %r114, 18;
and.b32 %r149, %r148, 1048576;
or.b32 %r150, %r147, %r149;
shl.b32 %r151, %r114, 25;
and.b32 %r152, %r151, 268435456;
or.b32 %r96, %r150, %r152;

dp4a.s32.s32 %r95, %r96, %r20, %r91;

shr.u32 %r153, %r112, 4;
and.b32 %r154, %r153, 252645135;
shr.u32 %r155, %r114, 12;
and.b32 %r156, %r155, 16;
or.b32 %r157, %r156, %r154;
shr.u32 %r158, %r114, 5;
and.b32 %r159, %r158, 4096;
or.b32 %r160, %r157, %r159;
shl.b32 %r161, %r114, 2;
and.b32 %r162, %r161, 1048576;
or.b32 %r163, %r160, %r162;
shl.b32 %r164, %r114, 9;
and.b32 %r165, %r164, 268435456;
or.b32 %r100, %r163, %r165;

dp4a.s32.s32 %r99, %r100, %r21, %r95;

ld.global.nc.u32 %r104, [%rd9];
ld.global.nc.u32 %r22, [%rd57];

{mul.f16x2 %r103,%r104,%r22;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r103;
cvt.f32.f16 %f101, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r103;
cvt.f32.f16 %f102, high;}


cvt.rn.f32.s32 %f103, %r99;
mov.f32 %f104, 0f40000000;
div.approx.ftz.f32 %f105, %f102, %f104;
fma.rn.ftz.f32 %f106, %f101, %f103, %f105;
add.ftz.f32 %f29, %f106, %f29;
setp.eq.s64 %p13, %rd22, 0;
@%p13 bra $L__BB73_12;

add.s64 %rd63, %rd91, %rd55;
add.s64 %rd66, %rd91, %rd60;
ld.global.u32 %r188, [%rd63+4];
shr.s32 %r190, %r188, %r110;
ld.global.u32 %r191, [%rd66+12];
shr.s32 %r193, %r188, %r113;
ld.global.u32 %r194, [%rd66+8];
and.b32 %r195, %r194, 252645135;
shl.b32 %r196, %r190, 4;
and.b32 %r197, %r196, 16;
or.b32 %r198, %r197, %r195;
shl.b32 %r199, %r190, 11;
and.b32 %r200, %r199, 4096;
or.b32 %r201, %r198, %r200;
shl.b32 %r202, %r190, 18;
and.b32 %r203, %r202, 1048576;
or.b32 %r204, %r201, %r203;
shl.b32 %r205, %r190, 25;
and.b32 %r206, %r205, 268435456;
or.b32 %r167, %r204, %r206;

dp4a.s32.s32 %r166, %r167, %r18, %r90;

shr.u32 %r207, %r194, 4;
and.b32 %r208, %r207, 252645135;
shr.u32 %r209, %r190, 12;
and.b32 %r210, %r209, 16;
or.b32 %r211, %r210, %r208;
shr.u32 %r212, %r190, 5;
and.b32 %r213, %r212, 4096;
or.b32 %r214, %r211, %r213;
shl.b32 %r215, %r190, 2;
and.b32 %r216, %r215, 1048576;
or.b32 %r217, %r214, %r216;
shl.b32 %r218, %r190, 9;
and.b32 %r219, %r218, 268435456;
or.b32 %r171, %r217, %r219;

dp4a.s32.s32 %r170, %r171, %r19, %r166;

and.b32 %r220, %r191, 252645135;
shl.b32 %r221, %r193, 4;
and.b32 %r222, %r221, 16;
or.b32 %r223, %r222, %r220;
shl.b32 %r224, %r193, 11;
and.b32 %r225, %r224, 4096;
or.b32 %r226, %r223, %r225;
shl.b32 %r227, %r193, 18;
and.b32 %r228, %r227, 1048576;
or.b32 %r229, %r226, %r228;
shl.b32 %r230, %r193, 25;
and.b32 %r231, %r230, 268435456;
or.b32 %r175, %r229, %r231;

dp4a.s32.s32 %r174, %r175, %r20, %r170;

shr.u32 %r232, %r191, 4;
and.b32 %r233, %r232, 252645135;
shr.u32 %r234, %r193, 12;
and.b32 %r235, %r234, 16;
or.b32 %r236, %r235, %r233;
shr.u32 %r237, %r193, 5;
and.b32 %r238, %r237, 4096;
or.b32 %r239, %r236, %r238;
shl.b32 %r240, %r193, 2;
and.b32 %r241, %r240, 1048576;
or.b32 %r242, %r239, %r241;
shl.b32 %r243, %r193, 9;
and.b32 %r244, %r243, 268435456;
or.b32 %r179, %r242, %r244;

dp4a.s32.s32 %r178, %r179, %r21, %r174;

ld.global.u32 %r183, [%rd63];

{mul.f16x2 %r182,%r183,%r22;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r182;
cvt.f32.f16 %f107, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r182;
cvt.f32.f16 %f108, high;}


cvt.rn.f32.s32 %f109, %r178;
div.approx.ftz.f32 %f111, %f108, %f104;
fma.rn.ftz.f32 %f112, %f107, %f109, %f111;
add.ftz.f32 %f33, %f112, %f33;

$L__BB73_12:
mul.wide.s32 %rd67, %r44, 24;
add.s64 %rd11, %rd9, %rd67;
add.s64 %rd12, %rd10, %rd67;
ld.global.nc.u32 %r266, [%rd11+4];
shr.s32 %r268, %r266, %r110;
ld.global.nc.u32 %r269, [%rd12+4];
shr.s32 %r271, %r266, %r113;
ld.global.nc.u32 %r272, [%rd12];
and.b32 %r273, %r272, 252645135;
shl.b32 %r274, %r268, 4;
and.b32 %r275, %r274, 16;
or.b32 %r276, %r275, %r273;
shl.b32 %r277, %r268, 11;
and.b32 %r278, %r277, 4096;
or.b32 %r279, %r276, %r278;
shl.b32 %r280, %r268, 18;
and.b32 %r281, %r280, 1048576;
or.b32 %r282, %r279, %r281;
shl.b32 %r283, %r268, 25;
and.b32 %r284, %r283, 268435456;
or.b32 %r246, %r282, %r284;

dp4a.s32.s32 %r245, %r246, %r18, %r90;

shr.u32 %r285, %r272, 4;
and.b32 %r286, %r285, 252645135;
shr.u32 %r287, %r268, 12;
and.b32 %r288, %r287, 16;
or.b32 %r289, %r288, %r286;
shr.u32 %r290, %r268, 5;
and.b32 %r291, %r290, 4096;
or.b32 %r292, %r289, %r291;
shl.b32 %r293, %r268, 2;
and.b32 %r294, %r293, 1048576;
or.b32 %r295, %r292, %r294;
shl.b32 %r296, %r268, 9;
and.b32 %r297, %r296, 268435456;
or.b32 %r250, %r295, %r297;

dp4a.s32.s32 %r249, %r250, %r19, %r245;

and.b32 %r298, %r269, 252645135;
shl.b32 %r299, %r271, 4;
and.b32 %r300, %r299, 16;
or.b32 %r301, %r300, %r298;
shl.b32 %r302, %r271, 11;
and.b32 %r303, %r302, 4096;
or.b32 %r304, %r301, %r303;
shl.b32 %r305, %r271, 18;
and.b32 %r306, %r305, 1048576;
or.b32 %r307, %r304, %r306;
shl.b32 %r308, %r271, 25;
and.b32 %r309, %r308, 268435456;
or.b32 %r254, %r307, %r309;

dp4a.s32.s32 %r253, %r254, %r20, %r249;

shr.u32 %r310, %r269, 4;
and.b32 %r311, %r310, 252645135;
shr.u32 %r312, %r271, 12;
and.b32 %r313, %r312, 16;
or.b32 %r314, %r313, %r311;
shr.u32 %r315, %r271, 5;
and.b32 %r316, %r315, 4096;
or.b32 %r317, %r314, %r316;
shl.b32 %r318, %r271, 2;
and.b32 %r319, %r318, 1048576;
or.b32 %r320, %r317, %r319;
shl.b32 %r321, %r271, 9;
and.b32 %r322, %r321, 268435456;
or.b32 %r258, %r320, %r322;

dp4a.s32.s32 %r257, %r258, %r21, %r253;

ld.global.nc.u32 %r262, [%rd11];

{mul.f16x2 %r261,%r262,%r22;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r261;
cvt.f32.f16 %f113, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r261;
cvt.f32.f16 %f114, high;}


cvt.rn.f32.s32 %f115, %r257;
div.approx.ftz.f32 %f117, %f114, %f104;
fma.rn.ftz.f32 %f118, %f113, %f115, %f117;
add.ftz.f32 %f28, %f118, %f28;
@%p13 bra $L__BB73_14;

or.b32 %r346, %r4, 1;
mad.lo.s32 %r347, %r44, %r346, %r15;
mul.wide.s32 %rd68, %r347, 24;
add.s64 %rd69, %rd91, %rd68;
add.s64 %rd71, %rd68, %rd58;
add.s64 %rd72, %rd91, %rd71;
ld.global.u32 %r349, [%rd69+4];
shr.s32 %r351, %r349, %r110;
ld.global.u32 %r352, [%rd72+12];
shr.s32 %r354, %r349, %r113;
ld.global.u32 %r355, [%rd72+8];
and.b32 %r356, %r355, 252645135;
shl.b32 %r357, %r351, 4;
and.b32 %r358, %r357, 16;
or.b32 %r359, %r358, %r356;
shl.b32 %r360, %r351, 11;
and.b32 %r361, %r360, 4096;
or.b32 %r362, %r359, %r361;
shl.b32 %r363, %r351, 18;
and.b32 %r364, %r363, 1048576;
or.b32 %r365, %r362, %r364;
shl.b32 %r366, %r351, 25;
and.b32 %r367, %r366, 268435456;
or.b32 %r324, %r365, %r367;
mov.u32 %r326, 0;

dp4a.s32.s32 %r323, %r324, %r18, %r326;

shr.u32 %r368, %r355, 4;
and.b32 %r369, %r368, 252645135;
shr.u32 %r370, %r351, 12;
and.b32 %r371, %r370, 16;
or.b32 %r372, %r371, %r369;
shr.u32 %r373, %r351, 5;
and.b32 %r374, %r373, 4096;
or.b32 %r375, %r372, %r374;
shl.b32 %r376, %r351, 2;
and.b32 %r377, %r376, 1048576;
or.b32 %r378, %r375, %r377;
shl.b32 %r379, %r351, 9;
and.b32 %r380, %r379, 268435456;
or.b32 %r328, %r378, %r380;

dp4a.s32.s32 %r327, %r328, %r19, %r323;

and.b32 %r381, %r352, 252645135;
shl.b32 %r382, %r354, 4;
and.b32 %r383, %r382, 16;
or.b32 %r384, %r383, %r381;
shl.b32 %r385, %r354, 11;
and.b32 %r386, %r385, 4096;
or.b32 %r387, %r384, %r386;
shl.b32 %r388, %r354, 18;
and.b32 %r389, %r388, 1048576;
or.b32 %r390, %r387, %r389;
shl.b32 %r391, %r354, 25;
and.b32 %r392, %r391, 268435456;
or.b32 %r332, %r390, %r392;

dp4a.s32.s32 %r331, %r332, %r20, %r327;

shr.u32 %r393, %r352, 4;
and.b32 %r394, %r393, 252645135;
shr.u32 %r395, %r354, 12;
and.b32 %r396, %r395, 16;
or.b32 %r397, %r396, %r394;
shr.u32 %r398, %r354, 5;
and.b32 %r399, %r398, 4096;
or.b32 %r400, %r397, %r399;
shl.b32 %r401, %r354, 2;
and.b32 %r402, %r401, 1048576;
or.b32 %r403, %r400, %r402;
shl.b32 %r404, %r354, 9;
and.b32 %r405, %r404, 268435456;
or.b32 %r336, %r403, %r405;

dp4a.s32.s32 %r335, %r336, %r21, %r331;

ld.global.u32 %r340, [%rd69];

{mul.f16x2 %r339,%r340,%r22;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r339;
cvt.f32.f16 %f119, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r339;
cvt.f32.f16 %f120, high;}


cvt.rn.f32.s32 %f121, %r335;
mov.f32 %f122, 0f40000000;
div.approx.ftz.f32 %f123, %f120, %f122;
fma.rn.ftz.f32 %f124, %f119, %f121, %f123;
add.ftz.f32 %f32, %f124, %f32;

$L__BB73_14:
add.s64 %rd13, %rd11, %rd67;
add.s64 %rd14, %rd12, %rd67;
ld.global.nc.u32 %r427, [%rd13+4];
shr.s32 %r429, %r427, %r110;
ld.global.nc.u32 %r430, [%rd14+4];
shr.s32 %r432, %r427, %r113;
ld.global.nc.u32 %r433, [%rd14];
and.b32 %r434, %r433, 252645135;
shl.b32 %r435, %r429, 4;
and.b32 %r436, %r435, 16;
or.b32 %r437, %r436, %r434;
shl.b32 %r438, %r429, 11;
and.b32 %r439, %r438, 4096;
or.b32 %r440, %r437, %r439;
shl.b32 %r441, %r429, 18;
and.b32 %r442, %r441, 1048576;
or.b32 %r443, %r440, %r442;
shl.b32 %r444, %r429, 25;
and.b32 %r445, %r444, 268435456;
or.b32 %r407, %r443, %r445;
mov.u32 %r409, 0;

dp4a.s32.s32 %r406, %r407, %r18, %r409;

shr.u32 %r446, %r433, 4;
and.b32 %r447, %r446, 252645135;
shr.u32 %r448, %r429, 12;
and.b32 %r449, %r448, 16;
or.b32 %r450, %r449, %r447;
shr.u32 %r451, %r429, 5;
and.b32 %r452, %r451, 4096;
or.b32 %r453, %r450, %r452;
shl.b32 %r454, %r429, 2;
and.b32 %r455, %r454, 1048576;
or.b32 %r456, %r453, %r455;
shl.b32 %r457, %r429, 9;
and.b32 %r458, %r457, 268435456;
or.b32 %r411, %r456, %r458;

dp4a.s32.s32 %r410, %r411, %r19, %r406;

and.b32 %r459, %r430, 252645135;
shl.b32 %r460, %r432, 4;
and.b32 %r461, %r460, 16;
or.b32 %r462, %r461, %r459;
shl.b32 %r463, %r432, 11;
and.b32 %r464, %r463, 4096;
or.b32 %r465, %r462, %r464;
shl.b32 %r466, %r432, 18;
and.b32 %r467, %r466, 1048576;
or.b32 %r468, %r465, %r467;
shl.b32 %r469, %r432, 25;
and.b32 %r470, %r469, 268435456;
or.b32 %r415, %r468, %r470;

dp4a.s32.s32 %r414, %r415, %r20, %r410;

shr.u32 %r471, %r430, 4;
and.b32 %r472, %r471, 252645135;
shr.u32 %r473, %r432, 12;
and.b32 %r474, %r473, 16;
or.b32 %r475, %r474, %r472;
shr.u32 %r476, %r432, 5;
and.b32 %r477, %r476, 4096;
or.b32 %r478, %r475, %r477;
shl.b32 %r479, %r432, 2;
and.b32 %r480, %r479, 1048576;
or.b32 %r481, %r478, %r480;
shl.b32 %r482, %r432, 9;
and.b32 %r483, %r482, 268435456;
or.b32 %r419, %r481, %r483;

dp4a.s32.s32 %r418, %r419, %r21, %r414;

ld.global.nc.u32 %r423, [%rd13];

{mul.f16x2 %r422,%r423,%r22;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r422;
cvt.f32.f16 %f125, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r422;
cvt.f32.f16 %f126, high;}


cvt.rn.f32.s32 %f127, %r418;
mov.f32 %f128, 0f40000000;
div.approx.ftz.f32 %f129, %f126, %f128;
fma.rn.ftz.f32 %f130, %f125, %f127, %f129;
add.ftz.f32 %f27, %f130, %f27;
@%p13 bra $L__BB73_16;

or.b32 %r507, %r4, 2;
mad.lo.s32 %r508, %r44, %r507, %r15;
mul.wide.s32 %rd74, %r508, 24;
add.s64 %rd75, %rd91, %rd74;
add.s64 %rd77, %rd74, %rd58;
add.s64 %rd78, %rd91, %rd77;
ld.global.u32 %r510, [%rd75+4];
shr.s32 %r512, %r510, %r110;
ld.global.u32 %r513, [%rd78+12];
shr.s32 %r515, %r510, %r113;
ld.global.u32 %r516, [%rd78+8];
and.b32 %r517, %r516, 252645135;
shl.b32 %r518, %r512, 4;
and.b32 %r519, %r518, 16;
or.b32 %r520, %r519, %r517;
shl.b32 %r521, %r512, 11;
and.b32 %r522, %r521, 4096;
or.b32 %r523, %r520, %r522;
shl.b32 %r524, %r512, 18;
and.b32 %r525, %r524, 1048576;
or.b32 %r526, %r523, %r525;
shl.b32 %r527, %r512, 25;
and.b32 %r528, %r527, 268435456;
or.b32 %r485, %r526, %r528;

dp4a.s32.s32 %r484, %r485, %r18, %r409;

shr.u32 %r529, %r516, 4;
and.b32 %r530, %r529, 252645135;
shr.u32 %r531, %r512, 12;
and.b32 %r532, %r531, 16;
or.b32 %r533, %r532, %r530;
shr.u32 %r534, %r512, 5;
and.b32 %r535, %r534, 4096;
or.b32 %r536, %r533, %r535;
shl.b32 %r537, %r512, 2;
and.b32 %r538, %r537, 1048576;
or.b32 %r539, %r536, %r538;
shl.b32 %r540, %r512, 9;
and.b32 %r541, %r540, 268435456;
or.b32 %r489, %r539, %r541;

dp4a.s32.s32 %r488, %r489, %r19, %r484;

and.b32 %r542, %r513, 252645135;
shl.b32 %r543, %r515, 4;
and.b32 %r544, %r543, 16;
or.b32 %r545, %r544, %r542;
shl.b32 %r546, %r515, 11;
and.b32 %r547, %r546, 4096;
or.b32 %r548, %r545, %r547;
shl.b32 %r549, %r515, 18;
and.b32 %r550, %r549, 1048576;
or.b32 %r551, %r548, %r550;
shl.b32 %r552, %r515, 25;
and.b32 %r553, %r552, 268435456;
or.b32 %r493, %r551, %r553;

dp4a.s32.s32 %r492, %r493, %r20, %r488;

shr.u32 %r554, %r513, 4;
and.b32 %r555, %r554, 252645135;
shr.u32 %r556, %r515, 12;
and.b32 %r557, %r556, 16;
or.b32 %r558, %r557, %r555;
shr.u32 %r559, %r515, 5;
and.b32 %r560, %r559, 4096;
or.b32 %r561, %r558, %r560;
shl.b32 %r562, %r515, 2;
and.b32 %r563, %r562, 1048576;
or.b32 %r564, %r561, %r563;
shl.b32 %r565, %r515, 9;
and.b32 %r566, %r565, 268435456;
or.b32 %r497, %r564, %r566;

dp4a.s32.s32 %r496, %r497, %r21, %r492;

ld.global.u32 %r501, [%rd75];

{mul.f16x2 %r500,%r501,%r22;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r500;
cvt.f32.f16 %f131, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r500;
cvt.f32.f16 %f132, high;}


cvt.rn.f32.s32 %f133, %r496;
div.approx.ftz.f32 %f135, %f132, %f128;
fma.rn.ftz.f32 %f136, %f131, %f133, %f135;
add.ftz.f32 %f31, %f136, %f31;

$L__BB73_16:
add.s64 %rd80, %rd14, %rd67;
add.s64 %rd81, %rd13, %rd67;
ld.global.nc.u32 %r588, [%rd81+4];
shr.s32 %r590, %r588, %r110;
ld.global.nc.u32 %r591, [%rd80+4];
shr.s32 %r593, %r588, %r113;
ld.global.nc.u32 %r594, [%rd80];
and.b32 %r595, %r594, 252645135;
shl.b32 %r596, %r590, 4;
and.b32 %r597, %r596, 16;
or.b32 %r598, %r597, %r595;
shl.b32 %r599, %r590, 11;
and.b32 %r600, %r599, 4096;
or.b32 %r601, %r598, %r600;
shl.b32 %r602, %r590, 18;
and.b32 %r603, %r602, 1048576;
or.b32 %r604, %r601, %r603;
shl.b32 %r605, %r590, 25;
and.b32 %r606, %r605, 268435456;
or.b32 %r568, %r604, %r606;

dp4a.s32.s32 %r567, %r568, %r18, %r409;

shr.u32 %r607, %r594, 4;
and.b32 %r608, %r607, 252645135;
shr.u32 %r609, %r590, 12;
and.b32 %r610, %r609, 16;
or.b32 %r611, %r610, %r608;
shr.u32 %r612, %r590, 5;
and.b32 %r613, %r612, 4096;
or.b32 %r614, %r611, %r613;
shl.b32 %r615, %r590, 2;
and.b32 %r616, %r615, 1048576;
or.b32 %r617, %r614, %r616;
shl.b32 %r618, %r590, 9;
and.b32 %r619, %r618, 268435456;
or.b32 %r572, %r617, %r619;

dp4a.s32.s32 %r571, %r572, %r19, %r567;

and.b32 %r620, %r591, 252645135;
shl.b32 %r621, %r593, 4;
and.b32 %r622, %r621, 16;
or.b32 %r623, %r622, %r620;
shl.b32 %r624, %r593, 11;
and.b32 %r625, %r624, 4096;
or.b32 %r626, %r623, %r625;
shl.b32 %r627, %r593, 18;
and.b32 %r628, %r627, 1048576;
or.b32 %r629, %r626, %r628;
shl.b32 %r630, %r593, 25;
and.b32 %r631, %r630, 268435456;
or.b32 %r576, %r629, %r631;

dp4a.s32.s32 %r575, %r576, %r20, %r571;

shr.u32 %r632, %r591, 4;
and.b32 %r633, %r632, 252645135;
shr.u32 %r634, %r593, 12;
and.b32 %r635, %r634, 16;
or.b32 %r636, %r635, %r633;
shr.u32 %r637, %r593, 5;
and.b32 %r638, %r637, 4096;
or.b32 %r639, %r636, %r638;
shl.b32 %r640, %r593, 2;
and.b32 %r641, %r640, 1048576;
or.b32 %r642, %r639, %r641;
shl.b32 %r643, %r593, 9;
and.b32 %r644, %r643, 268435456;
or.b32 %r580, %r642, %r644;

dp4a.s32.s32 %r579, %r580, %r21, %r575;

ld.global.nc.u32 %r584, [%rd81];

{mul.f16x2 %r583,%r584,%r22;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r583;
cvt.f32.f16 %f137, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r583;
cvt.f32.f16 %f138, high;}


cvt.rn.f32.s32 %f139, %r579;
div.approx.ftz.f32 %f141, %f138, %f128;
fma.rn.ftz.f32 %f142, %f137, %f139, %f141;
add.ftz.f32 %f26, %f142, %f26;
@%p13 bra $L__BB73_18;

or.b32 %r668, %r4, 3;
mad.lo.s32 %r669, %r44, %r668, %r15;
mul.wide.s32 %rd82, %r669, 24;
add.s64 %rd83, %rd91, %rd82;
add.s64 %rd85, %rd82, %rd58;
add.s64 %rd86, %rd91, %rd85;
ld.global.u32 %r671, [%rd83+4];
shr.s32 %r673, %r671, %r110;
ld.global.u32 %r674, [%rd86+12];
shr.s32 %r676, %r671, %r113;
ld.global.u32 %r677, [%rd86+8];
and.b32 %r678, %r677, 252645135;
shl.b32 %r679, %r673, 4;
and.b32 %r680, %r679, 16;
or.b32 %r681, %r680, %r678;
shl.b32 %r682, %r673, 11;
and.b32 %r683, %r682, 4096;
or.b32 %r684, %r681, %r683;
shl.b32 %r685, %r673, 18;
and.b32 %r686, %r685, 1048576;
or.b32 %r687, %r684, %r686;
shl.b32 %r688, %r673, 25;
and.b32 %r689, %r688, 268435456;
or.b32 %r646, %r687, %r689;
mov.u32 %r648, 0;

dp4a.s32.s32 %r645, %r646, %r18, %r648;

shr.u32 %r690, %r677, 4;
and.b32 %r691, %r690, 252645135;
shr.u32 %r692, %r673, 12;
and.b32 %r693, %r692, 16;
or.b32 %r694, %r693, %r691;
shr.u32 %r695, %r673, 5;
and.b32 %r696, %r695, 4096;
or.b32 %r697, %r694, %r696;
shl.b32 %r698, %r673, 2;
and.b32 %r699, %r698, 1048576;
or.b32 %r700, %r697, %r699;
shl.b32 %r701, %r673, 9;
and.b32 %r702, %r701, 268435456;
or.b32 %r650, %r700, %r702;

dp4a.s32.s32 %r649, %r650, %r19, %r645;

and.b32 %r703, %r674, 252645135;
shl.b32 %r704, %r676, 4;
and.b32 %r705, %r704, 16;
or.b32 %r706, %r705, %r703;
shl.b32 %r707, %r676, 11;
and.b32 %r708, %r707, 4096;
or.b32 %r709, %r706, %r708;
shl.b32 %r710, %r676, 18;
and.b32 %r711, %r710, 1048576;
or.b32 %r712, %r709, %r711;
shl.b32 %r713, %r676, 25;
and.b32 %r714, %r713, 268435456;
or.b32 %r654, %r712, %r714;

dp4a.s32.s32 %r653, %r654, %r20, %r649;

shr.u32 %r715, %r674, 4;
and.b32 %r716, %r715, 252645135;
shr.u32 %r717, %r676, 12;
and.b32 %r718, %r717, 16;
or.b32 %r719, %r718, %r716;
shr.u32 %r720, %r676, 5;
and.b32 %r721, %r720, 4096;
or.b32 %r722, %r719, %r721;
shl.b32 %r723, %r676, 2;
and.b32 %r724, %r723, 1048576;
or.b32 %r725, %r722, %r724;
shl.b32 %r726, %r676, 9;
and.b32 %r727, %r726, 268435456;
or.b32 %r658, %r725, %r727;

dp4a.s32.s32 %r657, %r658, %r21, %r653;

ld.global.u32 %r662, [%rd83];

{mul.f16x2 %r661,%r662,%r22;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r661;
cvt.f32.f16 %f143, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r661;
cvt.f32.f16 %f144, high;}


cvt.rn.f32.s32 %f145, %r657;
mov.f32 %f146, 0f40000000;
div.approx.ftz.f32 %f147, %f144, %f146;
fma.rn.ftz.f32 %f148, %f143, %f145, %f147;
add.ftz.f32 %f30, %f148, %f30;

$L__BB73_18:
add.s64 %rd91, %rd91, 1536;
add.s64 %rd90, %rd90, 1536;
add.s32 %r1474, %r1474, 64;
setp.lt.s32 %p17, %r1474, %r73;
@%p17 bra $L__BB73_10;

$L__BB73_19:
mov.u32 %r1461, %tid.y;
setp.eq.s32 %p18, %r1461, 0;
@%p18 bra $L__BB73_23;

mov.u32 %r1462, %tid.y;
setp.ne.s64 %p19, %rd22, 0;
add.s32 %r24, %r1462, -1;
shl.b32 %r732, %r1462, 9;
mov.u32 %r733, _ZZ13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r734, %r733, %r732;
shl.b32 %r735, %r2, 2;
add.s32 %r736, %r734, %r735;
st.shared.f32 [%r736+-512], %f29;
@%p19 bra $L__BB73_22;
bra.uni $L__BB73_21;

$L__BB73_22:
shl.b32 %r738, %r24, 9;
mov.u32 %r739, _ZZ13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r740, %r739, %r738;
add.s32 %r742, %r740, %r735;
st.shared.f32 [%r742], %f33;
add.s32 %r1458, %r736, -512;
st.shared.f32 [%r1458+128], %f28;
st.shared.f32 [%r742+128], %f32;
add.s32 %r1459, %r736, -512;
st.shared.f32 [%r1459+256], %f27;
st.shared.f32 [%r742+256], %f31;
add.s32 %r1460, %r736, -512;
st.shared.f32 [%r1460+384], %f26;
st.shared.f32 [%r742+384], %f30;
bra.uni $L__BB73_23;

$L__BB73_21:
add.s32 %r1455, %r736, -512;
st.shared.f32 [%r1455+128], %f28;
add.s32 %r1456, %r736, -512;
st.shared.f32 [%r1456+256], %f27;
add.s32 %r1457, %r736, -512;
st.shared.f32 [%r1457+384], %f26;

$L__BB73_23:
mov.u32 %r1463, %tid.y;
setp.ne.s32 %p319, %r1463, 0;
bar.sync 0;
@%p319 bra $L__BB73_78;
bra.uni $L__BB73_24;

$L__BB73_78:
ret;

$L__BB73_24:
ld.param.u64 %rd89, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.f32 %f993, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
ld.param.u32 %r1466, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r1465, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
mov.u32 %r1464, %ctaid.z;
setp.ne.s64 %p21, %rd22, 0;
mov.u32 %r745, %ctaid.y;
mad.lo.s32 %r748, %r745, %r1466, %r4;
mad.lo.s32 %r749, %r1464, %r1465, %r748;
neg.ftz.f32 %f35, %f993;
shl.b32 %r751, %r2, 2;
mov.u32 %r752, _ZZ13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r26, %r752, %r751;
cvta.to.global.u64 %rd87, %rd89;
mul.wide.u32 %rd88, %r749, 4;
add.s64 %rd17, %rd87, %rd88;
@%p21 bra $L__BB73_33;
bra.uni $L__BB73_25;

$L__BB73_33:
ld.param.u32 %r1467, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
mov.u32 %r839, _ZZ13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r28, %r839, %r751;
setp.gt.s32 %p54, %r1467, 2;
@%p54 bra $L__BB73_44;

ld.param.u32 %r1470, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p57, %r1470, 1;
@%p57 bra $L__BB73_62;

ld.param.u32 %r1471, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p58, %r1471, 2;
@%p58 bra $L__BB73_36;
bra.uni $L__BB73_70;

$L__BB73_36:
ld.shared.f32 %f549, [%r26];
add.ftz.f32 %f550, %f549, %f29;
ld.shared.f32 %f551, [%r28];
add.ftz.f32 %f552, %f551, %f33;
ld.shared.f32 %f553, [%r26+512];
add.ftz.f32 %f554, %f553, %f550;
ld.shared.f32 %f555, [%r28+512];
add.ftz.f32 %f556, %f555, %f552;
ld.shared.f32 %f557, [%r26+1024];
add.ftz.f32 %f558, %f557, %f554;
ld.shared.f32 %f559, [%r28+1024];
add.ftz.f32 %f560, %f559, %f556;
mov.b32 %r1086, %f558;
mov.u32 %r1087, 31;
mov.u32 %r1088, 16;
mov.u32 %r1089, -1;
shfl.sync.bfly.b32 %r1090|%p163, %r1086, %r1088, %r1087, %r1089;
mov.b32 %f561, %r1090;
add.ftz.f32 %f562, %f558, %f561;
mov.b32 %r1091, %f562;
mov.u32 %r1092, 8;
shfl.sync.bfly.b32 %r1093|%p164, %r1091, %r1092, %r1087, %r1089;
mov.b32 %f563, %r1093;
add.ftz.f32 %f564, %f562, %f563;
mov.b32 %r1094, %f564;
mov.u32 %r1095, 4;
shfl.sync.bfly.b32 %r1096|%p165, %r1094, %r1095, %r1087, %r1089;
mov.b32 %f565, %r1096;
add.ftz.f32 %f566, %f564, %f565;
mov.b32 %r1097, %f566;
mov.u32 %r1098, 2;
shfl.sync.bfly.b32 %r1099|%p166, %r1097, %r1098, %r1087, %r1089;
mov.b32 %f567, %r1099;
add.ftz.f32 %f568, %f566, %f567;
mov.b32 %r1100, %f568;
mov.u32 %r1101, 1;
shfl.sync.bfly.b32 %r1102|%p167, %r1100, %r1101, %r1087, %r1089;
mov.b32 %f569, %r1102;
add.ftz.f32 %f56, %f568, %f569;
mov.b32 %r1103, %f560;
shfl.sync.bfly.b32 %r1104|%p168, %r1103, %r1088, %r1087, %r1089;
mov.b32 %f570, %r1104;
add.ftz.f32 %f571, %f560, %f570;
mov.b32 %r1105, %f571;
shfl.sync.bfly.b32 %r1106|%p169, %r1105, %r1092, %r1087, %r1089;
mov.b32 %f572, %r1106;
add.ftz.f32 %f573, %f571, %f572;
mov.b32 %r1107, %f573;
shfl.sync.bfly.b32 %r1108|%p170, %r1107, %r1095, %r1087, %r1089;
mov.b32 %f574, %r1108;
add.ftz.f32 %f575, %f573, %f574;
mov.b32 %r1109, %f575;
shfl.sync.bfly.b32 %r1110|%p171, %r1109, %r1098, %r1087, %r1089;
mov.b32 %f576, %r1110;
add.ftz.f32 %f577, %f575, %f576;
mov.b32 %r1111, %f577;
shfl.sync.bfly.b32 %r1112|%p172, %r1111, %r1101, %r1087, %r1089;
mov.b32 %f578, %r1112;
add.ftz.f32 %f57, %f577, %f578;
setp.ne.s32 %p173, %r2, 0;
setp.ge.u32 %p174, %r4, %r45;
or.pred %p175, %p173, %p174;
@%p175 bra $L__BB73_38;

add.ftz.f32 %f579, %f5, %f56;
add.ftz.f32 %f580, %f4, %f57;
mul.ftz.f32 %f581, %f580, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f582, %f581;
add.ftz.f32 %f583, %f582, 0f3F800000;
div.approx.ftz.f32 %f584, %f580, %f583;
mul.ftz.f32 %f585, %f579, %f584;
st.global.f32 [%rd17], %f585;

$L__BB73_38:
ld.shared.f32 %f586, [%r26+128];
add.ftz.f32 %f587, %f586, %f28;
ld.shared.f32 %f588, [%r28+128];
add.ftz.f32 %f589, %f588, %f32;
ld.shared.f32 %f590, [%r26+640];
add.ftz.f32 %f591, %f590, %f587;
ld.shared.f32 %f592, [%r28+640];
add.ftz.f32 %f593, %f592, %f589;
ld.shared.f32 %f594, [%r26+1152];
add.ftz.f32 %f595, %f594, %f591;
ld.shared.f32 %f596, [%r28+1152];
add.ftz.f32 %f597, %f596, %f593;
mov.b32 %r1116, %f595;
shfl.sync.bfly.b32 %r1120|%p176, %r1116, %r1088, %r1087, %r1089;
mov.b32 %f598, %r1120;
add.ftz.f32 %f599, %f595, %f598;
mov.b32 %r1121, %f599;
shfl.sync.bfly.b32 %r1123|%p177, %r1121, %r1092, %r1087, %r1089;
mov.b32 %f600, %r1123;
add.ftz.f32 %f601, %f599, %f600;
mov.b32 %r1124, %f601;
shfl.sync.bfly.b32 %r1126|%p178, %r1124, %r1095, %r1087, %r1089;
mov.b32 %f602, %r1126;
add.ftz.f32 %f603, %f601, %f602;
mov.b32 %r1127, %f603;
shfl.sync.bfly.b32 %r1129|%p179, %r1127, %r1098, %r1087, %r1089;
mov.b32 %f604, %r1129;
add.ftz.f32 %f605, %f603, %f604;
mov.b32 %r1130, %f605;
shfl.sync.bfly.b32 %r1132|%p180, %r1130, %r1101, %r1087, %r1089;
mov.b32 %f606, %r1132;
add.ftz.f32 %f58, %f605, %f606;
mov.b32 %r1133, %f597;
shfl.sync.bfly.b32 %r1134|%p181, %r1133, %r1088, %r1087, %r1089;
mov.b32 %f607, %r1134;
add.ftz.f32 %f608, %f597, %f607;
mov.b32 %r1135, %f608;
shfl.sync.bfly.b32 %r1136|%p182, %r1135, %r1092, %r1087, %r1089;
mov.b32 %f609, %r1136;
add.ftz.f32 %f610, %f608, %f609;
mov.b32 %r1137, %f610;
shfl.sync.bfly.b32 %r1138|%p183, %r1137, %r1095, %r1087, %r1089;
mov.b32 %f611, %r1138;
add.ftz.f32 %f612, %f610, %f611;
mov.b32 %r1139, %f612;
shfl.sync.bfly.b32 %r1140|%p184, %r1139, %r1098, %r1087, %r1089;
mov.b32 %f613, %r1140;
add.ftz.f32 %f614, %f612, %f613;
mov.b32 %r1141, %f614;
shfl.sync.bfly.b32 %r1142|%p185, %r1141, %r1101, %r1087, %r1089;
mov.b32 %f615, %r1142;
add.ftz.f32 %f59, %f614, %f615;
setp.ne.s32 %p186, %r2, 1;
or.b32 %r1146, %r4, 1;
setp.ge.u32 %p187, %r1146, %r45;
or.pred %p188, %p186, %p187;
@%p188 bra $L__BB73_40;

add.ftz.f32 %f616, %f5, %f58;
add.ftz.f32 %f617, %f4, %f59;
mul.ftz.f32 %f618, %f617, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f619, %f618;
add.ftz.f32 %f620, %f619, 0f3F800000;
div.approx.ftz.f32 %f621, %f617, %f620;
mul.ftz.f32 %f622, %f616, %f621;
st.global.f32 [%rd17+4], %f622;

$L__BB73_40:
ld.shared.f32 %f623, [%r26+256];
add.ftz.f32 %f624, %f623, %f27;
ld.shared.f32 %f625, [%r28+256];
add.ftz.f32 %f626, %f625, %f31;
ld.shared.f32 %f627, [%r26+768];
add.ftz.f32 %f628, %f627, %f624;
ld.shared.f32 %f629, [%r28+768];
add.ftz.f32 %f630, %f629, %f626;
ld.shared.f32 %f631, [%r26+1280];
add.ftz.f32 %f632, %f631, %f628;
ld.shared.f32 %f633, [%r28+1280];
add.ftz.f32 %f634, %f633, %f630;
mov.b32 %r1147, %f632;
mov.u32 %r1148, 31;
mov.u32 %r1149, 16;
mov.u32 %r1150, -1;
shfl.sync.bfly.b32 %r1151|%p189, %r1147, %r1149, %r1148, %r1150;
mov.b32 %f635, %r1151;
add.ftz.f32 %f636, %f632, %f635;
mov.b32 %r1152, %f636;
mov.u32 %r1153, 8;
shfl.sync.bfly.b32 %r1154|%p190, %r1152, %r1153, %r1148, %r1150;
mov.b32 %f637, %r1154;
add.ftz.f32 %f638, %f636, %f637;
mov.b32 %r1155, %f638;
mov.u32 %r1156, 4;
shfl.sync.bfly.b32 %r1157|%p191, %r1155, %r1156, %r1148, %r1150;
mov.b32 %f639, %r1157;
add.ftz.f32 %f640, %f638, %f639;
mov.b32 %r1158, %f640;
mov.u32 %r1159, 2;
shfl.sync.bfly.b32 %r1160|%p192, %r1158, %r1159, %r1148, %r1150;
mov.b32 %f641, %r1160;
add.ftz.f32 %f642, %f640, %f641;
mov.b32 %r1161, %f642;
mov.u32 %r1162, 1;
shfl.sync.bfly.b32 %r1163|%p193, %r1161, %r1162, %r1148, %r1150;
mov.b32 %f643, %r1163;
add.ftz.f32 %f60, %f642, %f643;
mov.b32 %r1164, %f634;
shfl.sync.bfly.b32 %r1165|%p194, %r1164, %r1149, %r1148, %r1150;
mov.b32 %f644, %r1165;
add.ftz.f32 %f645, %f634, %f644;
mov.b32 %r1166, %f645;
shfl.sync.bfly.b32 %r1167|%p195, %r1166, %r1153, %r1148, %r1150;
mov.b32 %f646, %r1167;
add.ftz.f32 %f647, %f645, %f646;
mov.b32 %r1168, %f647;
shfl.sync.bfly.b32 %r1169|%p196, %r1168, %r1156, %r1148, %r1150;
mov.b32 %f648, %r1169;
add.ftz.f32 %f649, %f647, %f648;
mov.b32 %r1170, %f649;
shfl.sync.bfly.b32 %r1171|%p197, %r1170, %r1159, %r1148, %r1150;
mov.b32 %f650, %r1171;
add.ftz.f32 %f651, %f649, %f650;
mov.b32 %r1172, %f651;
shfl.sync.bfly.b32 %r1173|%p198, %r1172, %r1162, %r1148, %r1150;
mov.b32 %f652, %r1173;
add.ftz.f32 %f61, %f651, %f652;
setp.ne.s32 %p199, %r2, 2;
or.b32 %r1177, %r4, 2;
setp.ge.u32 %p200, %r1177, %r45;
or.pred %p201, %p199, %p200;
@%p201 bra $L__BB73_42;

add.ftz.f32 %f653, %f5, %f60;
add.ftz.f32 %f654, %f4, %f61;
mul.ftz.f32 %f655, %f654, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f656, %f655;
add.ftz.f32 %f657, %f656, 0f3F800000;
div.approx.ftz.f32 %f658, %f654, %f657;
mul.ftz.f32 %f659, %f653, %f658;
st.global.f32 [%rd17+8], %f659;

$L__BB73_42:
ld.shared.f32 %f660, [%r26+384];
add.ftz.f32 %f661, %f660, %f26;
ld.shared.f32 %f662, [%r28+384];
add.ftz.f32 %f663, %f662, %f30;
ld.shared.f32 %f664, [%r26+896];
add.ftz.f32 %f665, %f664, %f661;
ld.shared.f32 %f666, [%r28+896];
add.ftz.f32 %f667, %f666, %f663;
ld.shared.f32 %f668, [%r26+1408];
add.ftz.f32 %f669, %f668, %f665;
ld.shared.f32 %f670, [%r28+1408];
add.ftz.f32 %f671, %f670, %f667;
mov.b32 %r1178, %f669;
shfl.sync.bfly.b32 %r1182|%p202, %r1178, %r1149, %r1148, %r1150;
mov.b32 %f672, %r1182;
add.ftz.f32 %f673, %f669, %f672;
mov.b32 %r1183, %f673;
shfl.sync.bfly.b32 %r1185|%p203, %r1183, %r1153, %r1148, %r1150;
mov.b32 %f674, %r1185;
add.ftz.f32 %f675, %f673, %f674;
mov.b32 %r1186, %f675;
shfl.sync.bfly.b32 %r1188|%p204, %r1186, %r1156, %r1148, %r1150;
mov.b32 %f676, %r1188;
add.ftz.f32 %f677, %f675, %f676;
mov.b32 %r1189, %f677;
shfl.sync.bfly.b32 %r1191|%p205, %r1189, %r1159, %r1148, %r1150;
mov.b32 %f678, %r1191;
add.ftz.f32 %f679, %f677, %f678;
mov.b32 %r1192, %f679;
shfl.sync.bfly.b32 %r1194|%p206, %r1192, %r1162, %r1148, %r1150;
mov.b32 %f680, %r1194;
add.ftz.f32 %f62, %f679, %f680;
mov.b32 %r1195, %f671;
shfl.sync.bfly.b32 %r1196|%p207, %r1195, %r1149, %r1148, %r1150;
mov.b32 %f681, %r1196;
add.ftz.f32 %f682, %f671, %f681;
mov.b32 %r1197, %f682;
shfl.sync.bfly.b32 %r1198|%p208, %r1197, %r1153, %r1148, %r1150;
mov.b32 %f683, %r1198;
add.ftz.f32 %f684, %f682, %f683;
mov.b32 %r1199, %f684;
shfl.sync.bfly.b32 %r1200|%p209, %r1199, %r1156, %r1148, %r1150;
mov.b32 %f685, %r1200;
add.ftz.f32 %f686, %f684, %f685;
mov.b32 %r1201, %f686;
shfl.sync.bfly.b32 %r1202|%p210, %r1201, %r1159, %r1148, %r1150;
mov.b32 %f687, %r1202;
add.ftz.f32 %f688, %f686, %f687;
mov.b32 %r1203, %f688;
shfl.sync.bfly.b32 %r1204|%p211, %r1203, %r1162, %r1148, %r1150;
mov.b32 %f689, %r1204;
add.ftz.f32 %f63, %f688, %f689;
setp.ne.s32 %p212, %r2, 3;
or.b32 %r1208, %r4, 3;
setp.ge.u32 %p213, %r1208, %r45;
or.pred %p214, %p212, %p213;
@%p214 bra $L__BB73_78;

add.ftz.f32 %f690, %f5, %f62;
add.ftz.f32 %f691, %f4, %f63;
mul.ftz.f32 %f692, %f691, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f693, %f692;
add.ftz.f32 %f694, %f693, 0f3F800000;
div.approx.ftz.f32 %f695, %f691, %f694;
mul.ftz.f32 %f696, %f690, %f695;
st.global.f32 [%rd17+12], %f696;
bra.uni $L__BB73_78;

$L__BB73_25:
ld.shared.f32 %f149, [%r26];
add.ftz.f32 %f150, %f149, %f29;
ld.shared.f32 %f151, [%r26+512];
add.ftz.f32 %f152, %f151, %f150;
ld.shared.f32 %f153, [%r26+1024];
add.ftz.f32 %f154, %f153, %f152;
mov.b32 %r753, %f154;
mov.u32 %r754, 31;
mov.u32 %r755, 16;
mov.u32 %r756, -1;
shfl.sync.bfly.b32 %r757|%p22, %r753, %r755, %r754, %r756;
mov.b32 %f155, %r757;
add.ftz.f32 %f156, %f154, %f155;
mov.b32 %r758, %f156;
mov.u32 %r759, 8;
shfl.sync.bfly.b32 %r760|%p23, %r758, %r759, %r754, %r756;
mov.b32 %f157, %r760;
add.ftz.f32 %f158, %f156, %f157;
mov.b32 %r761, %f158;
mov.u32 %r762, 4;
shfl.sync.bfly.b32 %r763|%p24, %r761, %r762, %r754, %r756;
mov.b32 %f159, %r763;
add.ftz.f32 %f160, %f158, %f159;
mov.b32 %r764, %f160;
mov.u32 %r765, 2;
shfl.sync.bfly.b32 %r766|%p25, %r764, %r765, %r754, %r756;
mov.b32 %f161, %r766;
add.ftz.f32 %f162, %f160, %f161;
mov.b32 %r767, %f162;
mov.u32 %r768, 1;
shfl.sync.bfly.b32 %r769|%p26, %r767, %r768, %r754, %r756;
mov.b32 %f163, %r769;
add.ftz.f32 %f36, %f162, %f163;
setp.ne.s32 %p27, %r2, 0;
setp.ge.u32 %p28, %r4, %r45;
or.pred %p29, %p27, %p28;
@%p29 bra $L__BB73_27;

add.ftz.f32 %f164, %f5, %f36;
st.global.f32 [%rd17], %f164;

$L__BB73_27:
ld.shared.f32 %f165, [%r26+128];
add.ftz.f32 %f166, %f165, %f28;
ld.shared.f32 %f167, [%r26+640];
add.ftz.f32 %f168, %f167, %f166;
ld.shared.f32 %f169, [%r26+1152];
add.ftz.f32 %f170, %f169, %f168;
mov.b32 %r773, %f170;
shfl.sync.bfly.b32 %r777|%p30, %r773, %r755, %r754, %r756;
mov.b32 %f171, %r777;
add.ftz.f32 %f172, %f170, %f171;
mov.b32 %r778, %f172;
shfl.sync.bfly.b32 %r780|%p31, %r778, %r759, %r754, %r756;
mov.b32 %f173, %r780;
add.ftz.f32 %f174, %f172, %f173;
mov.b32 %r781, %f174;
shfl.sync.bfly.b32 %r783|%p32, %r781, %r762, %r754, %r756;
mov.b32 %f175, %r783;
add.ftz.f32 %f176, %f174, %f175;
mov.b32 %r784, %f176;
shfl.sync.bfly.b32 %r786|%p33, %r784, %r765, %r754, %r756;
mov.b32 %f177, %r786;
add.ftz.f32 %f178, %f176, %f177;
mov.b32 %r787, %f178;
shfl.sync.bfly.b32 %r789|%p34, %r787, %r768, %r754, %r756;
mov.b32 %f179, %r789;
add.ftz.f32 %f37, %f178, %f179;
setp.ne.s32 %p35, %r2, 1;
or.b32 %r793, %r4, 1;
setp.ge.u32 %p36, %r793, %r45;
or.pred %p37, %p35, %p36;
@%p37 bra $L__BB73_29;

add.ftz.f32 %f180, %f5, %f37;
st.global.f32 [%rd17+4], %f180;

$L__BB73_29:
ld.shared.f32 %f181, [%r26+256];
add.ftz.f32 %f182, %f181, %f27;
ld.shared.f32 %f183, [%r26+768];
add.ftz.f32 %f184, %f183, %f182;
ld.shared.f32 %f185, [%r26+1280];
add.ftz.f32 %f186, %f185, %f184;
mov.b32 %r794, %f186;
mov.u32 %r795, 31;
mov.u32 %r796, 16;
mov.u32 %r797, -1;
shfl.sync.bfly.b32 %r798|%p38, %r794, %r796, %r795, %r797;
mov.b32 %f187, %r798;
add.ftz.f32 %f188, %f186, %f187;
mov.b32 %r799, %f188;
mov.u32 %r800, 8;
shfl.sync.bfly.b32 %r801|%p39, %r799, %r800, %r795, %r797;
mov.b32 %f189, %r801;
add.ftz.f32 %f190, %f188, %f189;
mov.b32 %r802, %f190;
mov.u32 %r803, 4;
shfl.sync.bfly.b32 %r804|%p40, %r802, %r803, %r795, %r797;
mov.b32 %f191, %r804;
add.ftz.f32 %f192, %f190, %f191;
mov.b32 %r805, %f192;
mov.u32 %r806, 2;
shfl.sync.bfly.b32 %r807|%p41, %r805, %r806, %r795, %r797;
mov.b32 %f193, %r807;
add.ftz.f32 %f194, %f192, %f193;
mov.b32 %r808, %f194;
mov.u32 %r809, 1;
shfl.sync.bfly.b32 %r810|%p42, %r808, %r809, %r795, %r797;
mov.b32 %f195, %r810;
add.ftz.f32 %f38, %f194, %f195;
setp.ne.s32 %p43, %r2, 2;
or.b32 %r814, %r4, 2;
setp.ge.u32 %p44, %r814, %r45;
or.pred %p45, %p43, %p44;
@%p45 bra $L__BB73_31;

add.ftz.f32 %f196, %f5, %f38;
st.global.f32 [%rd17+8], %f196;

$L__BB73_31:
ld.shared.f32 %f197, [%r26+384];
add.ftz.f32 %f198, %f197, %f26;
ld.shared.f32 %f199, [%r26+896];
add.ftz.f32 %f200, %f199, %f198;
ld.shared.f32 %f201, [%r26+1408];
add.ftz.f32 %f202, %f201, %f200;
mov.b32 %r815, %f202;
shfl.sync.bfly.b32 %r819|%p46, %r815, %r796, %r795, %r797;
mov.b32 %f203, %r819;
add.ftz.f32 %f204, %f202, %f203;
mov.b32 %r820, %f204;
shfl.sync.bfly.b32 %r822|%p47, %r820, %r800, %r795, %r797;
mov.b32 %f205, %r822;
add.ftz.f32 %f206, %f204, %f205;
mov.b32 %r823, %f206;
shfl.sync.bfly.b32 %r825|%p48, %r823, %r803, %r795, %r797;
mov.b32 %f207, %r825;
add.ftz.f32 %f208, %f206, %f207;
mov.b32 %r826, %f208;
shfl.sync.bfly.b32 %r828|%p49, %r826, %r806, %r795, %r797;
mov.b32 %f209, %r828;
add.ftz.f32 %f210, %f208, %f209;
mov.b32 %r829, %f210;
shfl.sync.bfly.b32 %r831|%p50, %r829, %r809, %r795, %r797;
mov.b32 %f211, %r831;
add.ftz.f32 %f39, %f210, %f211;
setp.ne.s32 %p51, %r2, 3;
or.b32 %r835, %r4, 3;
setp.ge.u32 %p52, %r835, %r45;
or.pred %p53, %p51, %p52;
@%p53 bra $L__BB73_78;

add.ftz.f32 %f212, %f5, %f39;
st.global.f32 [%rd17+12], %f212;
bra.uni $L__BB73_78;

$L__BB73_44:
ld.param.u32 %r1468, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p55, %r1468, 3;
@%p55 bra $L__BB73_54;

ld.param.u32 %r1469, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.ne.s32 %p56, %r1469, 6;
@%p56 bra $L__BB73_70;

ld.shared.f32 %f213, [%r26];
add.ftz.f32 %f214, %f213, %f29;
ld.shared.f32 %f215, [%r28];
add.ftz.f32 %f216, %f215, %f33;
ld.shared.f32 %f217, [%r26+512];
add.ftz.f32 %f218, %f217, %f214;
ld.shared.f32 %f219, [%r28+512];
add.ftz.f32 %f220, %f219, %f216;
ld.shared.f32 %f221, [%r26+1024];
add.ftz.f32 %f222, %f221, %f218;
ld.shared.f32 %f223, [%r28+1024];
add.ftz.f32 %f224, %f223, %f220;
mov.b32 %r840, %f222;
mov.u32 %r841, 31;
mov.u32 %r842, 16;
mov.u32 %r843, -1;
shfl.sync.bfly.b32 %r844|%p59, %r840, %r842, %r841, %r843;
mov.b32 %f225, %r844;
add.ftz.f32 %f226, %f222, %f225;
mov.b32 %r845, %f226;
mov.u32 %r846, 8;
shfl.sync.bfly.b32 %r847|%p60, %r845, %r846, %r841, %r843;
mov.b32 %f227, %r847;
add.ftz.f32 %f228, %f226, %f227;
mov.b32 %r848, %f228;
mov.u32 %r849, 4;
shfl.sync.bfly.b32 %r850|%p61, %r848, %r849, %r841, %r843;
mov.b32 %f229, %r850;
add.ftz.f32 %f230, %f228, %f229;
mov.b32 %r851, %f230;
mov.u32 %r852, 2;
shfl.sync.bfly.b32 %r853|%p62, %r851, %r852, %r841, %r843;
mov.b32 %f231, %r853;
add.ftz.f32 %f232, %f230, %f231;
mov.b32 %r854, %f232;
mov.u32 %r855, 1;
shfl.sync.bfly.b32 %r856|%p63, %r854, %r855, %r841, %r843;
mov.b32 %f233, %r856;
add.ftz.f32 %f40, %f232, %f233;
mov.b32 %r857, %f224;
shfl.sync.bfly.b32 %r858|%p64, %r857, %r842, %r841, %r843;
mov.b32 %f234, %r858;
add.ftz.f32 %f235, %f224, %f234;
mov.b32 %r859, %f235;
shfl.sync.bfly.b32 %r860|%p65, %r859, %r846, %r841, %r843;
mov.b32 %f236, %r860;
add.ftz.f32 %f237, %f235, %f236;
mov.b32 %r861, %f237;
shfl.sync.bfly.b32 %r862|%p66, %r861, %r849, %r841, %r843;
mov.b32 %f238, %r862;
add.ftz.f32 %f239, %f237, %f238;
mov.b32 %r863, %f239;
shfl.sync.bfly.b32 %r864|%p67, %r863, %r852, %r841, %r843;
mov.b32 %f240, %r864;
add.ftz.f32 %f241, %f239, %f240;
mov.b32 %r865, %f241;
shfl.sync.bfly.b32 %r866|%p68, %r865, %r855, %r841, %r843;
mov.b32 %f242, %r866;
add.ftz.f32 %f41, %f241, %f242;
setp.ne.s32 %p69, %r2, 0;
setp.ge.u32 %p70, %r4, %r45;
or.pred %p71, %p69, %p70;
@%p71 bra $L__BB73_48;

ld.param.f32 %f997, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f243, %f5, %f40;
add.ftz.f32 %f244, %f4, %f41;
min.ftz.f32 %f245, %f244, %f997;
min.ftz.f32 %f246, %f243, %f997;
max.ftz.f32 %f247, %f246, %f35;
mul.ftz.f32 %f248, %f245, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f249, %f248;
add.ftz.f32 %f250, %f249, 0f3F800000;
div.approx.ftz.f32 %f251, %f245, %f250;
mul.ftz.f32 %f252, %f247, %f251;
st.global.f32 [%rd17], %f252;

$L__BB73_48:
ld.shared.f32 %f253, [%r26+128];
add.ftz.f32 %f254, %f253, %f28;
ld.shared.f32 %f255, [%r28+128];
add.ftz.f32 %f256, %f255, %f32;
ld.shared.f32 %f257, [%r26+640];
add.ftz.f32 %f258, %f257, %f254;
ld.shared.f32 %f259, [%r28+640];
add.ftz.f32 %f260, %f259, %f256;
ld.shared.f32 %f261, [%r26+1152];
add.ftz.f32 %f262, %f261, %f258;
ld.shared.f32 %f263, [%r28+1152];
add.ftz.f32 %f264, %f263, %f260;
mov.b32 %r870, %f262;
shfl.sync.bfly.b32 %r874|%p72, %r870, %r842, %r841, %r843;
mov.b32 %f265, %r874;
add.ftz.f32 %f266, %f262, %f265;
mov.b32 %r875, %f266;
shfl.sync.bfly.b32 %r877|%p73, %r875, %r846, %r841, %r843;
mov.b32 %f267, %r877;
add.ftz.f32 %f268, %f266, %f267;
mov.b32 %r878, %f268;
shfl.sync.bfly.b32 %r880|%p74, %r878, %r849, %r841, %r843;
mov.b32 %f269, %r880;
add.ftz.f32 %f270, %f268, %f269;
mov.b32 %r881, %f270;
shfl.sync.bfly.b32 %r883|%p75, %r881, %r852, %r841, %r843;
mov.b32 %f271, %r883;
add.ftz.f32 %f272, %f270, %f271;
mov.b32 %r884, %f272;
shfl.sync.bfly.b32 %r886|%p76, %r884, %r855, %r841, %r843;
mov.b32 %f273, %r886;
add.ftz.f32 %f42, %f272, %f273;
mov.b32 %r887, %f264;
shfl.sync.bfly.b32 %r888|%p77, %r887, %r842, %r841, %r843;
mov.b32 %f274, %r888;
add.ftz.f32 %f275, %f264, %f274;
mov.b32 %r889, %f275;
shfl.sync.bfly.b32 %r890|%p78, %r889, %r846, %r841, %r843;
mov.b32 %f276, %r890;
add.ftz.f32 %f277, %f275, %f276;
mov.b32 %r891, %f277;
shfl.sync.bfly.b32 %r892|%p79, %r891, %r849, %r841, %r843;
mov.b32 %f278, %r892;
add.ftz.f32 %f279, %f277, %f278;
mov.b32 %r893, %f279;
shfl.sync.bfly.b32 %r894|%p80, %r893, %r852, %r841, %r843;
mov.b32 %f280, %r894;
add.ftz.f32 %f281, %f279, %f280;
mov.b32 %r895, %f281;
shfl.sync.bfly.b32 %r896|%p81, %r895, %r855, %r841, %r843;
mov.b32 %f282, %r896;
add.ftz.f32 %f43, %f281, %f282;
setp.ne.s32 %p82, %r2, 1;
or.b32 %r900, %r4, 1;
setp.ge.u32 %p83, %r900, %r45;
or.pred %p84, %p82, %p83;
@%p84 bra $L__BB73_50;

ld.param.f32 %f996, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f283, %f5, %f42;
add.ftz.f32 %f284, %f4, %f43;
min.ftz.f32 %f285, %f284, %f996;
min.ftz.f32 %f286, %f283, %f996;
max.ftz.f32 %f287, %f286, %f35;
mul.ftz.f32 %f288, %f285, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f289, %f288;
add.ftz.f32 %f290, %f289, 0f3F800000;
div.approx.ftz.f32 %f291, %f285, %f290;
mul.ftz.f32 %f292, %f287, %f291;
st.global.f32 [%rd17+4], %f292;

$L__BB73_50:
ld.shared.f32 %f293, [%r26+256];
add.ftz.f32 %f294, %f293, %f27;
ld.shared.f32 %f295, [%r28+256];
add.ftz.f32 %f296, %f295, %f31;
ld.shared.f32 %f297, [%r26+768];
add.ftz.f32 %f298, %f297, %f294;
ld.shared.f32 %f299, [%r28+768];
add.ftz.f32 %f300, %f299, %f296;
ld.shared.f32 %f301, [%r26+1280];
add.ftz.f32 %f302, %f301, %f298;
ld.shared.f32 %f303, [%r28+1280];
add.ftz.f32 %f304, %f303, %f300;
mov.b32 %r901, %f302;
mov.u32 %r902, 31;
mov.u32 %r903, 16;
mov.u32 %r904, -1;
shfl.sync.bfly.b32 %r905|%p85, %r901, %r903, %r902, %r904;
mov.b32 %f305, %r905;
add.ftz.f32 %f306, %f302, %f305;
mov.b32 %r906, %f306;
mov.u32 %r907, 8;
shfl.sync.bfly.b32 %r908|%p86, %r906, %r907, %r902, %r904;
mov.b32 %f307, %r908;
add.ftz.f32 %f308, %f306, %f307;
mov.b32 %r909, %f308;
mov.u32 %r910, 4;
shfl.sync.bfly.b32 %r911|%p87, %r909, %r910, %r902, %r904;
mov.b32 %f309, %r911;
add.ftz.f32 %f310, %f308, %f309;
mov.b32 %r912, %f310;
mov.u32 %r913, 2;
shfl.sync.bfly.b32 %r914|%p88, %r912, %r913, %r902, %r904;
mov.b32 %f311, %r914;
add.ftz.f32 %f312, %f310, %f311;
mov.b32 %r915, %f312;
mov.u32 %r916, 1;
shfl.sync.bfly.b32 %r917|%p89, %r915, %r916, %r902, %r904;
mov.b32 %f313, %r917;
add.ftz.f32 %f44, %f312, %f313;
mov.b32 %r918, %f304;
shfl.sync.bfly.b32 %r919|%p90, %r918, %r903, %r902, %r904;
mov.b32 %f314, %r919;
add.ftz.f32 %f315, %f304, %f314;
mov.b32 %r920, %f315;
shfl.sync.bfly.b32 %r921|%p91, %r920, %r907, %r902, %r904;
mov.b32 %f316, %r921;
add.ftz.f32 %f317, %f315, %f316;
mov.b32 %r922, %f317;
shfl.sync.bfly.b32 %r923|%p92, %r922, %r910, %r902, %r904;
mov.b32 %f318, %r923;
add.ftz.f32 %f319, %f317, %f318;
mov.b32 %r924, %f319;
shfl.sync.bfly.b32 %r925|%p93, %r924, %r913, %r902, %r904;
mov.b32 %f320, %r925;
add.ftz.f32 %f321, %f319, %f320;
mov.b32 %r926, %f321;
shfl.sync.bfly.b32 %r927|%p94, %r926, %r916, %r902, %r904;
mov.b32 %f322, %r927;
add.ftz.f32 %f45, %f321, %f322;
setp.ne.s32 %p95, %r2, 2;
or.b32 %r931, %r4, 2;
setp.ge.u32 %p96, %r931, %r45;
or.pred %p97, %p95, %p96;
@%p97 bra $L__BB73_52;

ld.param.f32 %f995, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f323, %f5, %f44;
add.ftz.f32 %f324, %f4, %f45;
min.ftz.f32 %f325, %f324, %f995;
min.ftz.f32 %f326, %f323, %f995;
max.ftz.f32 %f327, %f326, %f35;
mul.ftz.f32 %f328, %f325, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f329, %f328;
add.ftz.f32 %f330, %f329, 0f3F800000;
div.approx.ftz.f32 %f331, %f325, %f330;
mul.ftz.f32 %f332, %f327, %f331;
st.global.f32 [%rd17+8], %f332;

$L__BB73_52:
ld.shared.f32 %f333, [%r26+384];
add.ftz.f32 %f334, %f333, %f26;
ld.shared.f32 %f335, [%r28+384];
add.ftz.f32 %f336, %f335, %f30;
ld.shared.f32 %f337, [%r26+896];
add.ftz.f32 %f338, %f337, %f334;
ld.shared.f32 %f339, [%r28+896];
add.ftz.f32 %f340, %f339, %f336;
ld.shared.f32 %f341, [%r26+1408];
add.ftz.f32 %f342, %f341, %f338;
ld.shared.f32 %f343, [%r28+1408];
add.ftz.f32 %f344, %f343, %f340;
mov.b32 %r932, %f342;
shfl.sync.bfly.b32 %r936|%p98, %r932, %r903, %r902, %r904;
mov.b32 %f345, %r936;
add.ftz.f32 %f346, %f342, %f345;
mov.b32 %r937, %f346;
shfl.sync.bfly.b32 %r939|%p99, %r937, %r907, %r902, %r904;
mov.b32 %f347, %r939;
add.ftz.f32 %f348, %f346, %f347;
mov.b32 %r940, %f348;
shfl.sync.bfly.b32 %r942|%p100, %r940, %r910, %r902, %r904;
mov.b32 %f349, %r942;
add.ftz.f32 %f350, %f348, %f349;
mov.b32 %r943, %f350;
shfl.sync.bfly.b32 %r945|%p101, %r943, %r913, %r902, %r904;
mov.b32 %f351, %r945;
add.ftz.f32 %f352, %f350, %f351;
mov.b32 %r946, %f352;
shfl.sync.bfly.b32 %r948|%p102, %r946, %r916, %r902, %r904;
mov.b32 %f353, %r948;
add.ftz.f32 %f46, %f352, %f353;
mov.b32 %r949, %f344;
shfl.sync.bfly.b32 %r950|%p103, %r949, %r903, %r902, %r904;
mov.b32 %f354, %r950;
add.ftz.f32 %f355, %f344, %f354;
mov.b32 %r951, %f355;
shfl.sync.bfly.b32 %r952|%p104, %r951, %r907, %r902, %r904;
mov.b32 %f356, %r952;
add.ftz.f32 %f357, %f355, %f356;
mov.b32 %r953, %f357;
shfl.sync.bfly.b32 %r954|%p105, %r953, %r910, %r902, %r904;
mov.b32 %f358, %r954;
add.ftz.f32 %f359, %f357, %f358;
mov.b32 %r955, %f359;
shfl.sync.bfly.b32 %r956|%p106, %r955, %r913, %r902, %r904;
mov.b32 %f360, %r956;
add.ftz.f32 %f361, %f359, %f360;
mov.b32 %r957, %f361;
shfl.sync.bfly.b32 %r958|%p107, %r957, %r916, %r902, %r904;
mov.b32 %f362, %r958;
add.ftz.f32 %f47, %f361, %f362;
setp.ne.s32 %p108, %r2, 3;
or.b32 %r962, %r4, 3;
setp.ge.u32 %p109, %r962, %r45;
or.pred %p110, %p108, %p109;
@%p110 bra $L__BB73_78;

ld.param.f32 %f994, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f363, %f5, %f46;
add.ftz.f32 %f364, %f4, %f47;
min.ftz.f32 %f365, %f364, %f994;
min.ftz.f32 %f366, %f363, %f994;
max.ftz.f32 %f367, %f366, %f35;
mul.ftz.f32 %f368, %f365, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f369, %f368;
add.ftz.f32 %f370, %f369, 0f3F800000;
div.approx.ftz.f32 %f371, %f365, %f370;
mul.ftz.f32 %f372, %f367, %f371;
st.global.f32 [%rd17+12], %f372;
bra.uni $L__BB73_78;

$L__BB73_70:
ld.shared.f32 %f861, [%r26];
add.ftz.f32 %f862, %f861, %f29;
ld.shared.f32 %f863, [%r28];
add.ftz.f32 %f864, %f863, %f33;
ld.shared.f32 %f865, [%r26+512];
add.ftz.f32 %f866, %f865, %f862;
ld.shared.f32 %f867, [%r28+512];
add.ftz.f32 %f868, %f867, %f864;
ld.shared.f32 %f869, [%r26+1024];
add.ftz.f32 %f870, %f869, %f866;
ld.shared.f32 %f871, [%r28+1024];
add.ftz.f32 %f872, %f871, %f868;
mov.b32 %r1332, %f870;
mov.u32 %r1333, 31;
mov.u32 %r1334, 16;
mov.u32 %r1335, -1;
shfl.sync.bfly.b32 %r1336|%p267, %r1332, %r1334, %r1333, %r1335;
mov.b32 %f873, %r1336;
add.ftz.f32 %f874, %f870, %f873;
mov.b32 %r1337, %f874;
mov.u32 %r1338, 8;
shfl.sync.bfly.b32 %r1339|%p268, %r1337, %r1338, %r1333, %r1335;
mov.b32 %f875, %r1339;
add.ftz.f32 %f876, %f874, %f875;
mov.b32 %r1340, %f876;
mov.u32 %r1341, 4;
shfl.sync.bfly.b32 %r1342|%p269, %r1340, %r1341, %r1333, %r1335;
mov.b32 %f877, %r1342;
add.ftz.f32 %f878, %f876, %f877;
mov.b32 %r1343, %f878;
mov.u32 %r1344, 2;
shfl.sync.bfly.b32 %r1345|%p270, %r1343, %r1344, %r1333, %r1335;
mov.b32 %f879, %r1345;
add.ftz.f32 %f880, %f878, %f879;
mov.b32 %r1346, %f880;
mov.u32 %r1347, 1;
shfl.sync.bfly.b32 %r1348|%p271, %r1346, %r1347, %r1333, %r1335;
mov.b32 %f881, %r1348;
add.ftz.f32 %f72, %f880, %f881;
mov.b32 %r1349, %f872;
shfl.sync.bfly.b32 %r1350|%p272, %r1349, %r1334, %r1333, %r1335;
mov.b32 %f882, %r1350;
add.ftz.f32 %f883, %f872, %f882;
mov.b32 %r1351, %f883;
shfl.sync.bfly.b32 %r1352|%p273, %r1351, %r1338, %r1333, %r1335;
mov.b32 %f884, %r1352;
add.ftz.f32 %f885, %f883, %f884;
mov.b32 %r1353, %f885;
shfl.sync.bfly.b32 %r1354|%p274, %r1353, %r1341, %r1333, %r1335;
mov.b32 %f886, %r1354;
add.ftz.f32 %f887, %f885, %f886;
mov.b32 %r1355, %f887;
shfl.sync.bfly.b32 %r1356|%p275, %r1355, %r1344, %r1333, %r1335;
mov.b32 %f888, %r1356;
add.ftz.f32 %f889, %f887, %f888;
mov.b32 %r1357, %f889;
shfl.sync.bfly.b32 %r1358|%p276, %r1357, %r1347, %r1333, %r1335;
mov.b32 %f890, %r1358;
add.ftz.f32 %f73, %f889, %f890;
setp.ne.s32 %p277, %r2, 0;
setp.ge.u32 %p278, %r4, %r45;
or.pred %p279, %p277, %p278;
@%p279 bra $L__BB73_72;

add.ftz.f32 %f891, %f5, %f72;
add.ftz.f32 %f892, %f4, %f73;
mul.ftz.f32 %f893, %f891, %f892;
st.global.f32 [%rd17], %f893;

$L__BB73_72:
ld.shared.f32 %f894, [%r26+128];
add.ftz.f32 %f895, %f894, %f28;
ld.shared.f32 %f896, [%r28+128];
add.ftz.f32 %f897, %f896, %f32;
ld.shared.f32 %f898, [%r26+640];
add.ftz.f32 %f899, %f898, %f895;
ld.shared.f32 %f900, [%r28+640];
add.ftz.f32 %f901, %f900, %f897;
ld.shared.f32 %f902, [%r26+1152];
add.ftz.f32 %f903, %f902, %f899;
ld.shared.f32 %f904, [%r28+1152];
add.ftz.f32 %f905, %f904, %f901;
mov.b32 %r1362, %f903;
shfl.sync.bfly.b32 %r1366|%p280, %r1362, %r1334, %r1333, %r1335;
mov.b32 %f906, %r1366;
add.ftz.f32 %f907, %f903, %f906;
mov.b32 %r1367, %f907;
shfl.sync.bfly.b32 %r1369|%p281, %r1367, %r1338, %r1333, %r1335;
mov.b32 %f908, %r1369;
add.ftz.f32 %f909, %f907, %f908;
mov.b32 %r1370, %f909;
shfl.sync.bfly.b32 %r1372|%p282, %r1370, %r1341, %r1333, %r1335;
mov.b32 %f910, %r1372;
add.ftz.f32 %f911, %f909, %f910;
mov.b32 %r1373, %f911;
shfl.sync.bfly.b32 %r1375|%p283, %r1373, %r1344, %r1333, %r1335;
mov.b32 %f912, %r1375;
add.ftz.f32 %f913, %f911, %f912;
mov.b32 %r1376, %f913;
shfl.sync.bfly.b32 %r1378|%p284, %r1376, %r1347, %r1333, %r1335;
mov.b32 %f914, %r1378;
add.ftz.f32 %f74, %f913, %f914;
mov.b32 %r1379, %f905;
shfl.sync.bfly.b32 %r1380|%p285, %r1379, %r1334, %r1333, %r1335;
mov.b32 %f915, %r1380;
add.ftz.f32 %f916, %f905, %f915;
mov.b32 %r1381, %f916;
shfl.sync.bfly.b32 %r1382|%p286, %r1381, %r1338, %r1333, %r1335;
mov.b32 %f917, %r1382;
add.ftz.f32 %f918, %f916, %f917;
mov.b32 %r1383, %f918;
shfl.sync.bfly.b32 %r1384|%p287, %r1383, %r1341, %r1333, %r1335;
mov.b32 %f919, %r1384;
add.ftz.f32 %f920, %f918, %f919;
mov.b32 %r1385, %f920;
shfl.sync.bfly.b32 %r1386|%p288, %r1385, %r1344, %r1333, %r1335;
mov.b32 %f921, %r1386;
add.ftz.f32 %f922, %f920, %f921;
mov.b32 %r1387, %f922;
shfl.sync.bfly.b32 %r1388|%p289, %r1387, %r1347, %r1333, %r1335;
mov.b32 %f923, %r1388;
add.ftz.f32 %f75, %f922, %f923;
setp.ne.s32 %p290, %r2, 1;
or.b32 %r1392, %r4, 1;
setp.ge.u32 %p291, %r1392, %r45;
or.pred %p292, %p290, %p291;
@%p292 bra $L__BB73_74;

add.ftz.f32 %f924, %f5, %f74;
add.ftz.f32 %f925, %f4, %f75;
mul.ftz.f32 %f926, %f924, %f925;
st.global.f32 [%rd17+4], %f926;

$L__BB73_74:
ld.shared.f32 %f927, [%r26+256];
add.ftz.f32 %f928, %f927, %f27;
ld.shared.f32 %f929, [%r28+256];
add.ftz.f32 %f930, %f929, %f31;
ld.shared.f32 %f931, [%r26+768];
add.ftz.f32 %f932, %f931, %f928;
ld.shared.f32 %f933, [%r28+768];
add.ftz.f32 %f934, %f933, %f930;
ld.shared.f32 %f935, [%r26+1280];
add.ftz.f32 %f936, %f935, %f932;
ld.shared.f32 %f937, [%r28+1280];
add.ftz.f32 %f938, %f937, %f934;
mov.b32 %r1393, %f936;
mov.u32 %r1394, 31;
mov.u32 %r1395, 16;
mov.u32 %r1396, -1;
shfl.sync.bfly.b32 %r1397|%p293, %r1393, %r1395, %r1394, %r1396;
mov.b32 %f939, %r1397;
add.ftz.f32 %f940, %f936, %f939;
mov.b32 %r1398, %f940;
mov.u32 %r1399, 8;
shfl.sync.bfly.b32 %r1400|%p294, %r1398, %r1399, %r1394, %r1396;
mov.b32 %f941, %r1400;
add.ftz.f32 %f942, %f940, %f941;
mov.b32 %r1401, %f942;
mov.u32 %r1402, 4;
shfl.sync.bfly.b32 %r1403|%p295, %r1401, %r1402, %r1394, %r1396;
mov.b32 %f943, %r1403;
add.ftz.f32 %f944, %f942, %f943;
mov.b32 %r1404, %f944;
mov.u32 %r1405, 2;
shfl.sync.bfly.b32 %r1406|%p296, %r1404, %r1405, %r1394, %r1396;
mov.b32 %f945, %r1406;
add.ftz.f32 %f946, %f944, %f945;
mov.b32 %r1407, %f946;
mov.u32 %r1408, 1;
shfl.sync.bfly.b32 %r1409|%p297, %r1407, %r1408, %r1394, %r1396;
mov.b32 %f947, %r1409;
add.ftz.f32 %f76, %f946, %f947;
mov.b32 %r1410, %f938;
shfl.sync.bfly.b32 %r1411|%p298, %r1410, %r1395, %r1394, %r1396;
mov.b32 %f948, %r1411;
add.ftz.f32 %f949, %f938, %f948;
mov.b32 %r1412, %f949;
shfl.sync.bfly.b32 %r1413|%p299, %r1412, %r1399, %r1394, %r1396;
mov.b32 %f950, %r1413;
add.ftz.f32 %f951, %f949, %f950;
mov.b32 %r1414, %f951;
shfl.sync.bfly.b32 %r1415|%p300, %r1414, %r1402, %r1394, %r1396;
mov.b32 %f952, %r1415;
add.ftz.f32 %f953, %f951, %f952;
mov.b32 %r1416, %f953;
shfl.sync.bfly.b32 %r1417|%p301, %r1416, %r1405, %r1394, %r1396;
mov.b32 %f954, %r1417;
add.ftz.f32 %f955, %f953, %f954;
mov.b32 %r1418, %f955;
shfl.sync.bfly.b32 %r1419|%p302, %r1418, %r1408, %r1394, %r1396;
mov.b32 %f956, %r1419;
add.ftz.f32 %f77, %f955, %f956;
setp.ne.s32 %p303, %r2, 2;
or.b32 %r1423, %r4, 2;
setp.ge.u32 %p304, %r1423, %r45;
or.pred %p305, %p303, %p304;
@%p305 bra $L__BB73_76;

add.ftz.f32 %f957, %f5, %f76;
add.ftz.f32 %f958, %f4, %f77;
mul.ftz.f32 %f959, %f957, %f958;
st.global.f32 [%rd17+8], %f959;

$L__BB73_76:
ld.shared.f32 %f960, [%r26+384];
add.ftz.f32 %f961, %f960, %f26;
ld.shared.f32 %f962, [%r28+384];
add.ftz.f32 %f963, %f962, %f30;
ld.shared.f32 %f964, [%r26+896];
add.ftz.f32 %f965, %f964, %f961;
ld.shared.f32 %f966, [%r28+896];
add.ftz.f32 %f967, %f966, %f963;
ld.shared.f32 %f968, [%r26+1408];
add.ftz.f32 %f969, %f968, %f965;
ld.shared.f32 %f970, [%r28+1408];
add.ftz.f32 %f971, %f970, %f967;
mov.b32 %r1424, %f969;
shfl.sync.bfly.b32 %r1428|%p306, %r1424, %r1395, %r1394, %r1396;
mov.b32 %f972, %r1428;
add.ftz.f32 %f973, %f969, %f972;
mov.b32 %r1429, %f973;
shfl.sync.bfly.b32 %r1431|%p307, %r1429, %r1399, %r1394, %r1396;
mov.b32 %f974, %r1431;
add.ftz.f32 %f975, %f973, %f974;
mov.b32 %r1432, %f975;
shfl.sync.bfly.b32 %r1434|%p308, %r1432, %r1402, %r1394, %r1396;
mov.b32 %f976, %r1434;
add.ftz.f32 %f977, %f975, %f976;
mov.b32 %r1435, %f977;
shfl.sync.bfly.b32 %r1437|%p309, %r1435, %r1405, %r1394, %r1396;
mov.b32 %f978, %r1437;
add.ftz.f32 %f979, %f977, %f978;
mov.b32 %r1438, %f979;
shfl.sync.bfly.b32 %r1440|%p310, %r1438, %r1408, %r1394, %r1396;
mov.b32 %f980, %r1440;
add.ftz.f32 %f78, %f979, %f980;
mov.b32 %r1441, %f971;
shfl.sync.bfly.b32 %r1442|%p311, %r1441, %r1395, %r1394, %r1396;
mov.b32 %f981, %r1442;
add.ftz.f32 %f982, %f971, %f981;
mov.b32 %r1443, %f982;
shfl.sync.bfly.b32 %r1444|%p312, %r1443, %r1399, %r1394, %r1396;
mov.b32 %f983, %r1444;
add.ftz.f32 %f984, %f982, %f983;
mov.b32 %r1445, %f984;
shfl.sync.bfly.b32 %r1446|%p313, %r1445, %r1402, %r1394, %r1396;
mov.b32 %f985, %r1446;
add.ftz.f32 %f986, %f984, %f985;
mov.b32 %r1447, %f986;
shfl.sync.bfly.b32 %r1448|%p314, %r1447, %r1405, %r1394, %r1396;
mov.b32 %f987, %r1448;
add.ftz.f32 %f988, %f986, %f987;
mov.b32 %r1449, %f988;
shfl.sync.bfly.b32 %r1450|%p315, %r1449, %r1408, %r1394, %r1396;
mov.b32 %f989, %r1450;
add.ftz.f32 %f79, %f988, %f989;
setp.ne.s32 %p316, %r2, 3;
or.b32 %r1454, %r4, 3;
setp.ge.u32 %p317, %r1454, %r45;
or.pred %p318, %p316, %p317;
@%p318 bra $L__BB73_78;

add.ftz.f32 %f990, %f5, %f78;
add.ftz.f32 %f991, %f4, %f79;
mul.ftz.f32 %f992, %f990, %f991;
st.global.f32 [%rd17+12], %f992;
bra.uni $L__BB73_78;

$L__BB73_62:
ld.shared.f32 %f697, [%r26];
add.ftz.f32 %f698, %f697, %f29;
ld.shared.f32 %f699, [%r28];
add.ftz.f32 %f700, %f699, %f33;
ld.shared.f32 %f701, [%r26+512];
add.ftz.f32 %f702, %f701, %f698;
ld.shared.f32 %f703, [%r28+512];
add.ftz.f32 %f704, %f703, %f700;
ld.shared.f32 %f705, [%r26+1024];
add.ftz.f32 %f706, %f705, %f702;
ld.shared.f32 %f707, [%r28+1024];
add.ftz.f32 %f708, %f707, %f704;
mov.b32 %r1209, %f706;
mov.u32 %r1210, 31;
mov.u32 %r1211, 16;
mov.u32 %r1212, -1;
shfl.sync.bfly.b32 %r1213|%p215, %r1209, %r1211, %r1210, %r1212;
mov.b32 %f709, %r1213;
add.ftz.f32 %f710, %f706, %f709;
mov.b32 %r1214, %f710;
mov.u32 %r1215, 8;
shfl.sync.bfly.b32 %r1216|%p216, %r1214, %r1215, %r1210, %r1212;
mov.b32 %f711, %r1216;
add.ftz.f32 %f712, %f710, %f711;
mov.b32 %r1217, %f712;
mov.u32 %r1218, 4;
shfl.sync.bfly.b32 %r1219|%p217, %r1217, %r1218, %r1210, %r1212;
mov.b32 %f713, %r1219;
add.ftz.f32 %f714, %f712, %f713;
mov.b32 %r1220, %f714;
mov.u32 %r1221, 2;
shfl.sync.bfly.b32 %r1222|%p218, %r1220, %r1221, %r1210, %r1212;
mov.b32 %f715, %r1222;
add.ftz.f32 %f716, %f714, %f715;
mov.b32 %r1223, %f716;
mov.u32 %r1224, 1;
shfl.sync.bfly.b32 %r1225|%p219, %r1223, %r1224, %r1210, %r1212;
mov.b32 %f717, %r1225;
add.ftz.f32 %f64, %f716, %f717;
mov.b32 %r1226, %f708;
shfl.sync.bfly.b32 %r1227|%p220, %r1226, %r1211, %r1210, %r1212;
mov.b32 %f718, %r1227;
add.ftz.f32 %f719, %f708, %f718;
mov.b32 %r1228, %f719;
shfl.sync.bfly.b32 %r1229|%p221, %r1228, %r1215, %r1210, %r1212;
mov.b32 %f720, %r1229;
add.ftz.f32 %f721, %f719, %f720;
mov.b32 %r1230, %f721;
shfl.sync.bfly.b32 %r1231|%p222, %r1230, %r1218, %r1210, %r1212;
mov.b32 %f722, %r1231;
add.ftz.f32 %f723, %f721, %f722;
mov.b32 %r1232, %f723;
shfl.sync.bfly.b32 %r1233|%p223, %r1232, %r1221, %r1210, %r1212;
mov.b32 %f724, %r1233;
add.ftz.f32 %f725, %f723, %f724;
mov.b32 %r1234, %f725;
shfl.sync.bfly.b32 %r1235|%p224, %r1234, %r1224, %r1210, %r1212;
mov.b32 %f726, %r1235;
add.ftz.f32 %f65, %f725, %f726;
setp.ne.s32 %p225, %r2, 0;
setp.ge.u32 %p226, %r4, %r45;
or.pred %p227, %p225, %p226;
@%p227 bra $L__BB73_64;

add.ftz.f32 %f727, %f5, %f64;
add.ftz.f32 %f728, %f4, %f65;
mul.ftz.f32 %f729, %f728, 0f3F000000;
mul.ftz.f32 %f730, %f728, 0f3F4C422A;
mul.ftz.f32 %f731, %f728, 0f3D372713;
fma.rn.ftz.f32 %f732, %f728, %f731, 0f3F800000;
mul.ftz.f32 %f733, %f730, %f732;
tanh.approx.f32 %f734, %f733;
add.ftz.f32 %f735, %f734, 0f3F800000;
mul.ftz.f32 %f736, %f729, %f735;
mul.ftz.f32 %f737, %f727, %f736;
st.global.f32 [%rd17], %f737;

$L__BB73_64:
ld.shared.f32 %f738, [%r26+128];
add.ftz.f32 %f739, %f738, %f28;
ld.shared.f32 %f740, [%r28+128];
add.ftz.f32 %f741, %f740, %f32;
ld.shared.f32 %f742, [%r26+640];
add.ftz.f32 %f743, %f742, %f739;
ld.shared.f32 %f744, [%r28+640];
add.ftz.f32 %f745, %f744, %f741;
ld.shared.f32 %f746, [%r26+1152];
add.ftz.f32 %f747, %f746, %f743;
ld.shared.f32 %f748, [%r28+1152];
add.ftz.f32 %f749, %f748, %f745;
mov.b32 %r1239, %f747;
shfl.sync.bfly.b32 %r1243|%p228, %r1239, %r1211, %r1210, %r1212;
mov.b32 %f750, %r1243;
add.ftz.f32 %f751, %f747, %f750;
mov.b32 %r1244, %f751;
shfl.sync.bfly.b32 %r1246|%p229, %r1244, %r1215, %r1210, %r1212;
mov.b32 %f752, %r1246;
add.ftz.f32 %f753, %f751, %f752;
mov.b32 %r1247, %f753;
shfl.sync.bfly.b32 %r1249|%p230, %r1247, %r1218, %r1210, %r1212;
mov.b32 %f754, %r1249;
add.ftz.f32 %f755, %f753, %f754;
mov.b32 %r1250, %f755;
shfl.sync.bfly.b32 %r1252|%p231, %r1250, %r1221, %r1210, %r1212;
mov.b32 %f756, %r1252;
add.ftz.f32 %f757, %f755, %f756;
mov.b32 %r1253, %f757;
shfl.sync.bfly.b32 %r1255|%p232, %r1253, %r1224, %r1210, %r1212;
mov.b32 %f758, %r1255;
add.ftz.f32 %f66, %f757, %f758;
mov.b32 %r1256, %f749;
shfl.sync.bfly.b32 %r1257|%p233, %r1256, %r1211, %r1210, %r1212;
mov.b32 %f759, %r1257;
add.ftz.f32 %f760, %f749, %f759;
mov.b32 %r1258, %f760;
shfl.sync.bfly.b32 %r1259|%p234, %r1258, %r1215, %r1210, %r1212;
mov.b32 %f761, %r1259;
add.ftz.f32 %f762, %f760, %f761;
mov.b32 %r1260, %f762;
shfl.sync.bfly.b32 %r1261|%p235, %r1260, %r1218, %r1210, %r1212;
mov.b32 %f763, %r1261;
add.ftz.f32 %f764, %f762, %f763;
mov.b32 %r1262, %f764;
shfl.sync.bfly.b32 %r1263|%p236, %r1262, %r1221, %r1210, %r1212;
mov.b32 %f765, %r1263;
add.ftz.f32 %f766, %f764, %f765;
mov.b32 %r1264, %f766;
shfl.sync.bfly.b32 %r1265|%p237, %r1264, %r1224, %r1210, %r1212;
mov.b32 %f767, %r1265;
add.ftz.f32 %f67, %f766, %f767;
setp.ne.s32 %p238, %r2, 1;
or.b32 %r1269, %r4, 1;
setp.ge.u32 %p239, %r1269, %r45;
or.pred %p240, %p238, %p239;
@%p240 bra $L__BB73_66;

add.ftz.f32 %f768, %f5, %f66;
add.ftz.f32 %f769, %f4, %f67;
mul.ftz.f32 %f770, %f769, 0f3F000000;
mul.ftz.f32 %f771, %f769, 0f3F4C422A;
mul.ftz.f32 %f772, %f769, 0f3D372713;
fma.rn.ftz.f32 %f773, %f769, %f772, 0f3F800000;
mul.ftz.f32 %f774, %f771, %f773;
tanh.approx.f32 %f775, %f774;
add.ftz.f32 %f776, %f775, 0f3F800000;
mul.ftz.f32 %f777, %f770, %f776;
mul.ftz.f32 %f778, %f768, %f777;
st.global.f32 [%rd17+4], %f778;

$L__BB73_66:
ld.shared.f32 %f779, [%r26+256];
add.ftz.f32 %f780, %f779, %f27;
ld.shared.f32 %f781, [%r28+256];
add.ftz.f32 %f782, %f781, %f31;
ld.shared.f32 %f783, [%r26+768];
add.ftz.f32 %f784, %f783, %f780;
ld.shared.f32 %f785, [%r28+768];
add.ftz.f32 %f786, %f785, %f782;
ld.shared.f32 %f787, [%r26+1280];
add.ftz.f32 %f788, %f787, %f784;
ld.shared.f32 %f789, [%r28+1280];
add.ftz.f32 %f790, %f789, %f786;
mov.b32 %r1270, %f788;
mov.u32 %r1271, 31;
mov.u32 %r1272, 16;
mov.u32 %r1273, -1;
shfl.sync.bfly.b32 %r1274|%p241, %r1270, %r1272, %r1271, %r1273;
mov.b32 %f791, %r1274;
add.ftz.f32 %f792, %f788, %f791;
mov.b32 %r1275, %f792;
mov.u32 %r1276, 8;
shfl.sync.bfly.b32 %r1277|%p242, %r1275, %r1276, %r1271, %r1273;
mov.b32 %f793, %r1277;
add.ftz.f32 %f794, %f792, %f793;
mov.b32 %r1278, %f794;
mov.u32 %r1279, 4;
shfl.sync.bfly.b32 %r1280|%p243, %r1278, %r1279, %r1271, %r1273;
mov.b32 %f795, %r1280;
add.ftz.f32 %f796, %f794, %f795;
mov.b32 %r1281, %f796;
mov.u32 %r1282, 2;
shfl.sync.bfly.b32 %r1283|%p244, %r1281, %r1282, %r1271, %r1273;
mov.b32 %f797, %r1283;
add.ftz.f32 %f798, %f796, %f797;
mov.b32 %r1284, %f798;
mov.u32 %r1285, 1;
shfl.sync.bfly.b32 %r1286|%p245, %r1284, %r1285, %r1271, %r1273;
mov.b32 %f799, %r1286;
add.ftz.f32 %f68, %f798, %f799;
mov.b32 %r1287, %f790;
shfl.sync.bfly.b32 %r1288|%p246, %r1287, %r1272, %r1271, %r1273;
mov.b32 %f800, %r1288;
add.ftz.f32 %f801, %f790, %f800;
mov.b32 %r1289, %f801;
shfl.sync.bfly.b32 %r1290|%p247, %r1289, %r1276, %r1271, %r1273;
mov.b32 %f802, %r1290;
add.ftz.f32 %f803, %f801, %f802;
mov.b32 %r1291, %f803;
shfl.sync.bfly.b32 %r1292|%p248, %r1291, %r1279, %r1271, %r1273;
mov.b32 %f804, %r1292;
add.ftz.f32 %f805, %f803, %f804;
mov.b32 %r1293, %f805;
shfl.sync.bfly.b32 %r1294|%p249, %r1293, %r1282, %r1271, %r1273;
mov.b32 %f806, %r1294;
add.ftz.f32 %f807, %f805, %f806;
mov.b32 %r1295, %f807;
shfl.sync.bfly.b32 %r1296|%p250, %r1295, %r1285, %r1271, %r1273;
mov.b32 %f808, %r1296;
add.ftz.f32 %f69, %f807, %f808;
setp.ne.s32 %p251, %r2, 2;
or.b32 %r1300, %r4, 2;
setp.ge.u32 %p252, %r1300, %r45;
or.pred %p253, %p251, %p252;
@%p253 bra $L__BB73_68;

add.ftz.f32 %f809, %f5, %f68;
add.ftz.f32 %f810, %f4, %f69;
mul.ftz.f32 %f811, %f810, 0f3F000000;
mul.ftz.f32 %f812, %f810, 0f3F4C422A;
mul.ftz.f32 %f813, %f810, 0f3D372713;
fma.rn.ftz.f32 %f814, %f810, %f813, 0f3F800000;
mul.ftz.f32 %f815, %f812, %f814;
tanh.approx.f32 %f816, %f815;
add.ftz.f32 %f817, %f816, 0f3F800000;
mul.ftz.f32 %f818, %f811, %f817;
mul.ftz.f32 %f819, %f809, %f818;
st.global.f32 [%rd17+8], %f819;

$L__BB73_68:
ld.shared.f32 %f820, [%r26+384];
add.ftz.f32 %f821, %f820, %f26;
ld.shared.f32 %f822, [%r28+384];
add.ftz.f32 %f823, %f822, %f30;
ld.shared.f32 %f824, [%r26+896];
add.ftz.f32 %f825, %f824, %f821;
ld.shared.f32 %f826, [%r28+896];
add.ftz.f32 %f827, %f826, %f823;
ld.shared.f32 %f828, [%r26+1408];
add.ftz.f32 %f829, %f828, %f825;
ld.shared.f32 %f830, [%r28+1408];
add.ftz.f32 %f831, %f830, %f827;
mov.b32 %r1301, %f829;
shfl.sync.bfly.b32 %r1305|%p254, %r1301, %r1272, %r1271, %r1273;
mov.b32 %f832, %r1305;
add.ftz.f32 %f833, %f829, %f832;
mov.b32 %r1306, %f833;
shfl.sync.bfly.b32 %r1308|%p255, %r1306, %r1276, %r1271, %r1273;
mov.b32 %f834, %r1308;
add.ftz.f32 %f835, %f833, %f834;
mov.b32 %r1309, %f835;
shfl.sync.bfly.b32 %r1311|%p256, %r1309, %r1279, %r1271, %r1273;
mov.b32 %f836, %r1311;
add.ftz.f32 %f837, %f835, %f836;
mov.b32 %r1312, %f837;
shfl.sync.bfly.b32 %r1314|%p257, %r1312, %r1282, %r1271, %r1273;
mov.b32 %f838, %r1314;
add.ftz.f32 %f839, %f837, %f838;
mov.b32 %r1315, %f839;
shfl.sync.bfly.b32 %r1317|%p258, %r1315, %r1285, %r1271, %r1273;
mov.b32 %f840, %r1317;
add.ftz.f32 %f70, %f839, %f840;
mov.b32 %r1318, %f831;
shfl.sync.bfly.b32 %r1319|%p259, %r1318, %r1272, %r1271, %r1273;
mov.b32 %f841, %r1319;
add.ftz.f32 %f842, %f831, %f841;
mov.b32 %r1320, %f842;
shfl.sync.bfly.b32 %r1321|%p260, %r1320, %r1276, %r1271, %r1273;
mov.b32 %f843, %r1321;
add.ftz.f32 %f844, %f842, %f843;
mov.b32 %r1322, %f844;
shfl.sync.bfly.b32 %r1323|%p261, %r1322, %r1279, %r1271, %r1273;
mov.b32 %f845, %r1323;
add.ftz.f32 %f846, %f844, %f845;
mov.b32 %r1324, %f846;
shfl.sync.bfly.b32 %r1325|%p262, %r1324, %r1282, %r1271, %r1273;
mov.b32 %f847, %r1325;
add.ftz.f32 %f848, %f846, %f847;
mov.b32 %r1326, %f848;
shfl.sync.bfly.b32 %r1327|%p263, %r1326, %r1285, %r1271, %r1273;
mov.b32 %f849, %r1327;
add.ftz.f32 %f71, %f848, %f849;
setp.ne.s32 %p264, %r2, 3;
or.b32 %r1331, %r4, 3;
setp.ge.u32 %p265, %r1331, %r45;
or.pred %p266, %p264, %p265;
@%p266 bra $L__BB73_78;

add.ftz.f32 %f850, %f5, %f70;
add.ftz.f32 %f851, %f4, %f71;
mul.ftz.f32 %f852, %f851, 0f3F000000;
mul.ftz.f32 %f853, %f851, 0f3F4C422A;
mul.ftz.f32 %f854, %f851, 0f3D372713;
fma.rn.ftz.f32 %f855, %f851, %f854, 0f3F800000;
mul.ftz.f32 %f856, %f853, %f855;
tanh.approx.f32 %f857, %f856;
add.ftz.f32 %f858, %f857, 0f3F800000;
mul.ftz.f32 %f859, %f852, %f858;
mul.ftz.f32 %f860, %f850, %f859;
st.global.f32 [%rd17+12], %f860;
bra.uni $L__BB73_78;

$L__BB73_54:
ld.shared.f32 %f373, [%r26];
add.ftz.f32 %f374, %f373, %f29;
ld.shared.f32 %f375, [%r28];
add.ftz.f32 %f376, %f375, %f33;
ld.shared.f32 %f377, [%r26+512];
add.ftz.f32 %f378, %f377, %f374;
ld.shared.f32 %f379, [%r28+512];
add.ftz.f32 %f380, %f379, %f376;
ld.shared.f32 %f381, [%r26+1024];
add.ftz.f32 %f382, %f381, %f378;
ld.shared.f32 %f383, [%r28+1024];
add.ftz.f32 %f384, %f383, %f380;
mov.b32 %r963, %f382;
mov.u32 %r964, 31;
mov.u32 %r965, 16;
mov.u32 %r966, -1;
shfl.sync.bfly.b32 %r967|%p111, %r963, %r965, %r964, %r966;
mov.b32 %f385, %r967;
add.ftz.f32 %f386, %f382, %f385;
mov.b32 %r968, %f386;
mov.u32 %r969, 8;
shfl.sync.bfly.b32 %r970|%p112, %r968, %r969, %r964, %r966;
mov.b32 %f387, %r970;
add.ftz.f32 %f388, %f386, %f387;
mov.b32 %r971, %f388;
mov.u32 %r972, 4;
shfl.sync.bfly.b32 %r973|%p113, %r971, %r972, %r964, %r966;
mov.b32 %f389, %r973;
add.ftz.f32 %f390, %f388, %f389;
mov.b32 %r974, %f390;
mov.u32 %r975, 2;
shfl.sync.bfly.b32 %r976|%p114, %r974, %r975, %r964, %r966;
mov.b32 %f391, %r976;
add.ftz.f32 %f392, %f390, %f391;
mov.b32 %r977, %f392;
mov.u32 %r978, 1;
shfl.sync.bfly.b32 %r979|%p115, %r977, %r978, %r964, %r966;
mov.b32 %f393, %r979;
add.ftz.f32 %f48, %f392, %f393;
mov.b32 %r980, %f384;
shfl.sync.bfly.b32 %r981|%p116, %r980, %r965, %r964, %r966;
mov.b32 %f394, %r981;
add.ftz.f32 %f395, %f384, %f394;
mov.b32 %r982, %f395;
shfl.sync.bfly.b32 %r983|%p117, %r982, %r969, %r964, %r966;
mov.b32 %f396, %r983;
add.ftz.f32 %f397, %f395, %f396;
mov.b32 %r984, %f397;
shfl.sync.bfly.b32 %r985|%p118, %r984, %r972, %r964, %r966;
mov.b32 %f398, %r985;
add.ftz.f32 %f399, %f397, %f398;
mov.b32 %r986, %f399;
shfl.sync.bfly.b32 %r987|%p119, %r986, %r975, %r964, %r966;
mov.b32 %f400, %r987;
add.ftz.f32 %f401, %f399, %f400;
mov.b32 %r988, %f401;
shfl.sync.bfly.b32 %r989|%p120, %r988, %r978, %r964, %r966;
mov.b32 %f402, %r989;
add.ftz.f32 %f49, %f401, %f402;
setp.ne.s32 %p121, %r2, 0;
setp.ge.u32 %p122, %r4, %r45;
or.pred %p123, %p121, %p122;
@%p123 bra $L__BB73_56;

add.ftz.f32 %f403, %f5, %f48;
add.ftz.f32 %f404, %f4, %f49;
mov.f32 %f405, 0f40E00000;
min.ftz.f32 %f406, %f404, %f405;
min.ftz.f32 %f407, %f403, %f405;
mov.f32 %f408, 0fC0E00000;
max.ftz.f32 %f409, %f407, %f408;
mul.ftz.f32 %f410, %f406, 0f3FD9DB23;
mul.ftz.f32 %f411, %f410, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f412, %f411;
add.ftz.f32 %f413, %f412, 0f3F800000;
div.approx.ftz.f32 %f414, %f406, %f413;
add.ftz.f32 %f415, %f409, 0f3F800000;
mul.ftz.f32 %f416, %f415, %f414;
st.global.f32 [%rd17], %f416;

$L__BB73_56:
ld.shared.f32 %f417, [%r26+128];
add.ftz.f32 %f418, %f417, %f28;
ld.shared.f32 %f419, [%r28+128];
add.ftz.f32 %f420, %f419, %f32;
ld.shared.f32 %f421, [%r26+640];
add.ftz.f32 %f422, %f421, %f418;
ld.shared.f32 %f423, [%r28+640];
add.ftz.f32 %f424, %f423, %f420;
ld.shared.f32 %f425, [%r26+1152];
add.ftz.f32 %f426, %f425, %f422;
ld.shared.f32 %f427, [%r28+1152];
add.ftz.f32 %f428, %f427, %f424;
mov.b32 %r993, %f426;
shfl.sync.bfly.b32 %r997|%p124, %r993, %r965, %r964, %r966;
mov.b32 %f429, %r997;
add.ftz.f32 %f430, %f426, %f429;
mov.b32 %r998, %f430;
shfl.sync.bfly.b32 %r1000|%p125, %r998, %r969, %r964, %r966;
mov.b32 %f431, %r1000;
add.ftz.f32 %f432, %f430, %f431;
mov.b32 %r1001, %f432;
shfl.sync.bfly.b32 %r1003|%p126, %r1001, %r972, %r964, %r966;
mov.b32 %f433, %r1003;
add.ftz.f32 %f434, %f432, %f433;
mov.b32 %r1004, %f434;
shfl.sync.bfly.b32 %r1006|%p127, %r1004, %r975, %r964, %r966;
mov.b32 %f435, %r1006;
add.ftz.f32 %f436, %f434, %f435;
mov.b32 %r1007, %f436;
shfl.sync.bfly.b32 %r1009|%p128, %r1007, %r978, %r964, %r966;
mov.b32 %f437, %r1009;
add.ftz.f32 %f50, %f436, %f437;
mov.b32 %r1010, %f428;
shfl.sync.bfly.b32 %r1011|%p129, %r1010, %r965, %r964, %r966;
mov.b32 %f438, %r1011;
add.ftz.f32 %f439, %f428, %f438;
mov.b32 %r1012, %f439;
shfl.sync.bfly.b32 %r1013|%p130, %r1012, %r969, %r964, %r966;
mov.b32 %f440, %r1013;
add.ftz.f32 %f441, %f439, %f440;
mov.b32 %r1014, %f441;
shfl.sync.bfly.b32 %r1015|%p131, %r1014, %r972, %r964, %r966;
mov.b32 %f442, %r1015;
add.ftz.f32 %f443, %f441, %f442;
mov.b32 %r1016, %f443;
shfl.sync.bfly.b32 %r1017|%p132, %r1016, %r975, %r964, %r966;
mov.b32 %f444, %r1017;
add.ftz.f32 %f445, %f443, %f444;
mov.b32 %r1018, %f445;
shfl.sync.bfly.b32 %r1019|%p133, %r1018, %r978, %r964, %r966;
mov.b32 %f446, %r1019;
add.ftz.f32 %f51, %f445, %f446;
setp.ne.s32 %p134, %r2, 1;
or.b32 %r1023, %r4, 1;
setp.ge.u32 %p135, %r1023, %r45;
or.pred %p136, %p134, %p135;
@%p136 bra $L__BB73_58;

add.ftz.f32 %f447, %f5, %f50;
add.ftz.f32 %f448, %f4, %f51;
mov.f32 %f449, 0f40E00000;
min.ftz.f32 %f450, %f448, %f449;
min.ftz.f32 %f451, %f447, %f449;
mov.f32 %f452, 0fC0E00000;
max.ftz.f32 %f453, %f451, %f452;
mul.ftz.f32 %f454, %f450, 0f3FD9DB23;
mul.ftz.f32 %f455, %f454, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f456, %f455;
add.ftz.f32 %f457, %f456, 0f3F800000;
div.approx.ftz.f32 %f458, %f450, %f457;
add.ftz.f32 %f459, %f453, 0f3F800000;
mul.ftz.f32 %f460, %f459, %f458;
st.global.f32 [%rd17+4], %f460;

$L__BB73_58:
ld.shared.f32 %f461, [%r26+256];
add.ftz.f32 %f462, %f461, %f27;
ld.shared.f32 %f463, [%r28+256];
add.ftz.f32 %f464, %f463, %f31;
ld.shared.f32 %f465, [%r26+768];
add.ftz.f32 %f466, %f465, %f462;
ld.shared.f32 %f467, [%r28+768];
add.ftz.f32 %f468, %f467, %f464;
ld.shared.f32 %f469, [%r26+1280];
add.ftz.f32 %f470, %f469, %f466;
ld.shared.f32 %f471, [%r28+1280];
add.ftz.f32 %f472, %f471, %f468;
mov.b32 %r1024, %f470;
mov.u32 %r1025, 31;
mov.u32 %r1026, 16;
mov.u32 %r1027, -1;
shfl.sync.bfly.b32 %r1028|%p137, %r1024, %r1026, %r1025, %r1027;
mov.b32 %f473, %r1028;
add.ftz.f32 %f474, %f470, %f473;
mov.b32 %r1029, %f474;
mov.u32 %r1030, 8;
shfl.sync.bfly.b32 %r1031|%p138, %r1029, %r1030, %r1025, %r1027;
mov.b32 %f475, %r1031;
add.ftz.f32 %f476, %f474, %f475;
mov.b32 %r1032, %f476;
mov.u32 %r1033, 4;
shfl.sync.bfly.b32 %r1034|%p139, %r1032, %r1033, %r1025, %r1027;
mov.b32 %f477, %r1034;
add.ftz.f32 %f478, %f476, %f477;
mov.b32 %r1035, %f478;
mov.u32 %r1036, 2;
shfl.sync.bfly.b32 %r1037|%p140, %r1035, %r1036, %r1025, %r1027;
mov.b32 %f479, %r1037;
add.ftz.f32 %f480, %f478, %f479;
mov.b32 %r1038, %f480;
mov.u32 %r1039, 1;
shfl.sync.bfly.b32 %r1040|%p141, %r1038, %r1039, %r1025, %r1027;
mov.b32 %f481, %r1040;
add.ftz.f32 %f52, %f480, %f481;
mov.b32 %r1041, %f472;
shfl.sync.bfly.b32 %r1042|%p142, %r1041, %r1026, %r1025, %r1027;
mov.b32 %f482, %r1042;
add.ftz.f32 %f483, %f472, %f482;
mov.b32 %r1043, %f483;
shfl.sync.bfly.b32 %r1044|%p143, %r1043, %r1030, %r1025, %r1027;
mov.b32 %f484, %r1044;
add.ftz.f32 %f485, %f483, %f484;
mov.b32 %r1045, %f485;
shfl.sync.bfly.b32 %r1046|%p144, %r1045, %r1033, %r1025, %r1027;
mov.b32 %f486, %r1046;
add.ftz.f32 %f487, %f485, %f486;
mov.b32 %r1047, %f487;
shfl.sync.bfly.b32 %r1048|%p145, %r1047, %r1036, %r1025, %r1027;
mov.b32 %f488, %r1048;
add.ftz.f32 %f489, %f487, %f488;
mov.b32 %r1049, %f489;
shfl.sync.bfly.b32 %r1050|%p146, %r1049, %r1039, %r1025, %r1027;
mov.b32 %f490, %r1050;
add.ftz.f32 %f53, %f489, %f490;
setp.ne.s32 %p147, %r2, 2;
or.b32 %r1054, %r4, 2;
setp.ge.u32 %p148, %r1054, %r45;
or.pred %p149, %p147, %p148;
@%p149 bra $L__BB73_60;

add.ftz.f32 %f491, %f5, %f52;
add.ftz.f32 %f492, %f4, %f53;
mov.f32 %f493, 0f40E00000;
min.ftz.f32 %f494, %f492, %f493;
min.ftz.f32 %f495, %f491, %f493;
mov.f32 %f496, 0fC0E00000;
max.ftz.f32 %f497, %f495, %f496;
mul.ftz.f32 %f498, %f494, 0f3FD9DB23;
mul.ftz.f32 %f499, %f498, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f500, %f499;
add.ftz.f32 %f501, %f500, 0f3F800000;
div.approx.ftz.f32 %f502, %f494, %f501;
add.ftz.f32 %f503, %f497, 0f3F800000;
mul.ftz.f32 %f504, %f503, %f502;
st.global.f32 [%rd17+8], %f504;

$L__BB73_60:
ld.shared.f32 %f505, [%r26+384];
add.ftz.f32 %f506, %f505, %f26;
ld.shared.f32 %f507, [%r28+384];
add.ftz.f32 %f508, %f507, %f30;
ld.shared.f32 %f509, [%r26+896];
add.ftz.f32 %f510, %f509, %f506;
ld.shared.f32 %f511, [%r28+896];
add.ftz.f32 %f512, %f511, %f508;
ld.shared.f32 %f513, [%r26+1408];
add.ftz.f32 %f514, %f513, %f510;
ld.shared.f32 %f515, [%r28+1408];
add.ftz.f32 %f516, %f515, %f512;
mov.b32 %r1055, %f514;
shfl.sync.bfly.b32 %r1059|%p150, %r1055, %r1026, %r1025, %r1027;
mov.b32 %f517, %r1059;
add.ftz.f32 %f518, %f514, %f517;
mov.b32 %r1060, %f518;
shfl.sync.bfly.b32 %r1062|%p151, %r1060, %r1030, %r1025, %r1027;
mov.b32 %f519, %r1062;
add.ftz.f32 %f520, %f518, %f519;
mov.b32 %r1063, %f520;
shfl.sync.bfly.b32 %r1065|%p152, %r1063, %r1033, %r1025, %r1027;
mov.b32 %f521, %r1065;
add.ftz.f32 %f522, %f520, %f521;
mov.b32 %r1066, %f522;
shfl.sync.bfly.b32 %r1068|%p153, %r1066, %r1036, %r1025, %r1027;
mov.b32 %f523, %r1068;
add.ftz.f32 %f524, %f522, %f523;
mov.b32 %r1069, %f524;
shfl.sync.bfly.b32 %r1071|%p154, %r1069, %r1039, %r1025, %r1027;
mov.b32 %f525, %r1071;
add.ftz.f32 %f54, %f524, %f525;
mov.b32 %r1072, %f516;
shfl.sync.bfly.b32 %r1073|%p155, %r1072, %r1026, %r1025, %r1027;
mov.b32 %f526, %r1073;
add.ftz.f32 %f527, %f516, %f526;
mov.b32 %r1074, %f527;
shfl.sync.bfly.b32 %r1075|%p156, %r1074, %r1030, %r1025, %r1027;
mov.b32 %f528, %r1075;
add.ftz.f32 %f529, %f527, %f528;
mov.b32 %r1076, %f529;
shfl.sync.bfly.b32 %r1077|%p157, %r1076, %r1033, %r1025, %r1027;
mov.b32 %f530, %r1077;
add.ftz.f32 %f531, %f529, %f530;
mov.b32 %r1078, %f531;
shfl.sync.bfly.b32 %r1079|%p158, %r1078, %r1036, %r1025, %r1027;
mov.b32 %f532, %r1079;
add.ftz.f32 %f533, %f531, %f532;
mov.b32 %r1080, %f533;
shfl.sync.bfly.b32 %r1081|%p159, %r1080, %r1039, %r1025, %r1027;
mov.b32 %f534, %r1081;
add.ftz.f32 %f55, %f533, %f534;
setp.ne.s32 %p160, %r2, 3;
or.b32 %r1085, %r4, 3;
setp.ge.u32 %p161, %r1085, %r45;
or.pred %p162, %p160, %p161;
@%p162 bra $L__BB73_78;

add.ftz.f32 %f535, %f5, %f54;
add.ftz.f32 %f536, %f4, %f55;
mov.f32 %f537, 0f40E00000;
min.ftz.f32 %f538, %f536, %f537;
min.ftz.f32 %f539, %f535, %f537;
mov.f32 %f540, 0fC0E00000;
max.ftz.f32 %f541, %f539, %f540;
mul.ftz.f32 %f542, %f538, 0f3FD9DB23;
mul.ftz.f32 %f543, %f542, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f544, %f543;
add.ftz.f32 %f545, %f544, 0f3F800000;
div.approx.ftz.f32 %f546, %f538, %f545;
add.ftz.f32 %f547, %f541, 0f3F800000;
mul.ftz.f32 %f548, %f547, %f546;
st.global.f32 [%rd17+12], %f548;
bra.uni $L__BB73_78;

}
.entry _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<38>;
.reg .f32 %f<114>;
.reg .b32 %r<450>;
.reg .b64 %rd<45>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

ld.param.u64 %rd12, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd13, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd14, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r19, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
mov.u32 %r1, %tid.y;
shl.b32 %r37, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r37, %r2;
mov.u32 %r38, %ctaid.x;
shl.b32 %r4, %r38, 2;
shr.u32 %r5, %r36, 5;
mov.u32 %r448, %ctaid.y;
setp.eq.s64 %p1, %rd14, 0;
@%p1 bra $L__BB74_2;

cvta.to.global.u64 %rd16, %rd14;
mul.wide.u32 %rd17, %r448, 4;
add.s64 %rd18, %rd16, %rd17;
ld.global.nc.u32 %r447, [%rd18];
mul.hi.u32 %r39, %r448, %r19;
add.s32 %r40, %r39, %r448;
shr.u32 %r41, %r40, %r20;
mul.lo.s32 %r42, %r41, %r21;
sub.s32 %r448, %r448, %r42;
bra.uni $L__BB74_3;

$L__BB74_2:
mul.hi.u32 %r43, %r448, %r24;
add.s32 %r44, %r43, %r448;
shr.u32 %r447, %r44, %r25;

$L__BB74_3:
shr.u32 %r45, %r3, 31;
add.s32 %r46, %r3, %r45;
shr.s32 %r449, %r46, 1;
setp.ge.s32 %p2, %r449, %r5;
mov.f32 %f110, 0f00000000;
mov.f32 %f111, %f110;
mov.f32 %f112, %f110;
mov.f32 %f113, %f110;
@%p2 bra $L__BB74_6;

mov.u32 %r47, %ctaid.z;
and.b32 %r50, %r46, -2;
sub.s32 %r51, %r3, %r50;
shl.b32 %r52, %r51, 1;
cvt.s64.s32 %rd1, %r52;
shl.b32 %r13, %r51, 3;
or.b32 %r14, %r13, 4;
mad.lo.s32 %r53, %r4, %r22, %r449;
mul.hi.u32 %r54, %r47, %r30;
add.s32 %r55, %r54, %r47;
shr.u32 %r56, %r55, %r31;
mad.lo.s32 %r57, %r56, %r33, %r53;
mad.lo.s32 %r58, %r447, %r27, %r57;
mul.wide.s32 %rd2, %r58, 24;
mul.wide.s32 %rd19, %r52, 4;
add.s64 %rd20, %rd2, %rd19;
add.s64 %rd3, %rd20, 12;
mul.wide.s32 %rd4, %r22, 24;
or.b64 %rd5, %rd4, 4;
cvta.to.global.u64 %rd6, %rd13;
cvta.to.global.u64 %rd44, %rd12;
mul.lo.s32 %r59, %r448, %r28;
cvt.u64.u32 %rd21, %r59;
mul.lo.s32 %r60, %r47, %r34;
cvt.u64.u32 %rd22, %r60;
add.s64 %rd8, %rd21, %rd22;
mov.f32 %f110, 0f00000000;
shl.b64 %rd27, %rd1, 2;

$L__BB74_5:
cvt.u64.u32 %rd23, %r449;
add.s64 %rd24, %rd8, %rd23;
mul.lo.s64 %rd25, %rd24, 36;
add.s64 %rd26, %rd6, %rd25;
add.s64 %rd28, %rd26, %rd27;
add.s64 %rd29, %rd44, %rd2;
add.s64 %rd30, %rd29, 4;
add.s64 %rd31, %rd44, %rd3;
add.s64 %rd32, %rd31, -4;
ld.global.nc.u32 %r145, [%rd29+4];
shr.s32 %r146, %r145, %r13;
ld.global.nc.u32 %r63, [%rd28+4];
ld.global.nc.u32 %r67, [%rd28+20];
ld.global.nc.u32 %r147, [%rd31];
shr.s32 %r148, %r145, %r14;
ld.global.nc.u32 %r71, [%rd28+8];
ld.global.nc.u32 %r75, [%rd28+24];
ld.global.nc.u32 %r149, [%rd31+-4];
and.b32 %r150, %r149, 252645135;
shl.b32 %r151, %r146, 4;
and.b32 %r152, %r151, 16;
or.b32 %r153, %r152, %r150;
shl.b32 %r154, %r146, 11;
and.b32 %r155, %r154, 4096;
or.b32 %r156, %r153, %r155;
shl.b32 %r157, %r146, 18;
and.b32 %r158, %r157, 1048576;
or.b32 %r159, %r156, %r158;
shl.b32 %r160, %r146, 25;
and.b32 %r161, %r160, 268435456;
or.b32 %r62, %r159, %r161;
mov.u32 %r127, 0;

dp4a.s32.s32 %r61, %r62, %r63, %r127;

shr.u32 %r162, %r149, 4;
and.b32 %r163, %r162, 252645135;
shr.u32 %r164, %r146, 12;
and.b32 %r165, %r164, 16;
or.b32 %r166, %r165, %r163;
shr.u32 %r167, %r146, 5;
and.b32 %r168, %r167, 4096;
or.b32 %r169, %r166, %r168;
shl.b32 %r170, %r146, 2;
and.b32 %r171, %r170, 1048576;
or.b32 %r172, %r169, %r171;
shl.b32 %r173, %r146, 9;
and.b32 %r174, %r173, 268435456;
or.b32 %r66, %r172, %r174;

dp4a.s32.s32 %r65, %r66, %r67, %r61;

and.b32 %r175, %r147, 252645135;
shl.b32 %r176, %r148, 4;
and.b32 %r177, %r176, 16;
or.b32 %r178, %r177, %r175;
shl.b32 %r179, %r148, 11;
and.b32 %r180, %r179, 4096;
or.b32 %r181, %r178, %r180;
shl.b32 %r182, %r148, 18;
and.b32 %r183, %r182, 1048576;
or.b32 %r184, %r181, %r183;
shl.b32 %r185, %r148, 25;
and.b32 %r186, %r185, 268435456;
or.b32 %r70, %r184, %r186;

dp4a.s32.s32 %r69, %r70, %r71, %r65;

shr.u32 %r187, %r147, 4;
and.b32 %r188, %r187, 252645135;
shr.u32 %r189, %r148, 12;
and.b32 %r190, %r189, 16;
or.b32 %r191, %r190, %r188;
shr.u32 %r192, %r148, 5;
and.b32 %r193, %r192, 4096;
or.b32 %r194, %r191, %r193;
shl.b32 %r195, %r148, 2;
and.b32 %r196, %r195, 1048576;
or.b32 %r197, %r194, %r196;
shl.b32 %r198, %r148, 9;
and.b32 %r199, %r198, 268435456;
or.b32 %r74, %r197, %r199;

dp4a.s32.s32 %r73, %r74, %r75, %r69;

ld.global.nc.u32 %r78, [%rd29];
ld.global.nc.u32 %r79, [%rd26];

{mul.f16x2 %r77,%r78,%r79;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r77;
cvt.f32.f16 %f25, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r77;
cvt.f32.f16 %f26, high;}


cvt.rn.f32.s32 %f33, %r73;
mov.f32 %f34, 0f40000000;
div.approx.ftz.f32 %f35, %f26, %f34;
fma.rn.ftz.f32 %f36, %f25, %f33, %f35;
add.ftz.f32 %f113, %f36, %f113;
add.s64 %rd33, %rd30, %rd4;
add.s64 %rd34, %rd33, -4;
add.s64 %rd35, %rd29, %rd4;
add.s64 %rd36, %rd32, %rd4;
ld.global.nc.u32 %r200, [%rd35+4];
shr.s32 %r201, %r200, %r13;
ld.global.nc.u32 %r202, [%rd36+4];
shr.s32 %r203, %r200, %r14;
ld.global.nc.u32 %r204, [%rd36];
and.b32 %r205, %r204, 252645135;
shl.b32 %r206, %r201, 4;
and.b32 %r207, %r206, 16;
or.b32 %r208, %r207, %r205;
shl.b32 %r209, %r201, 11;
and.b32 %r210, %r209, 4096;
or.b32 %r211, %r208, %r210;
shl.b32 %r212, %r201, 18;
and.b32 %r213, %r212, 1048576;
or.b32 %r214, %r211, %r213;
shl.b32 %r215, %r201, 25;
and.b32 %r216, %r215, 268435456;
or.b32 %r83, %r214, %r216;

dp4a.s32.s32 %r82, %r83, %r63, %r127;

shr.u32 %r217, %r204, 4;
and.b32 %r218, %r217, 252645135;
shr.u32 %r219, %r201, 12;
and.b32 %r220, %r219, 16;
or.b32 %r221, %r220, %r218;
shr.u32 %r222, %r201, 5;
and.b32 %r223, %r222, 4096;
or.b32 %r224, %r221, %r223;
shl.b32 %r225, %r201, 2;
and.b32 %r226, %r225, 1048576;
or.b32 %r227, %r224, %r226;
shl.b32 %r228, %r201, 9;
and.b32 %r229, %r228, 268435456;
or.b32 %r87, %r227, %r229;

dp4a.s32.s32 %r86, %r87, %r67, %r82;

and.b32 %r230, %r202, 252645135;
shl.b32 %r231, %r203, 4;
and.b32 %r232, %r231, 16;
or.b32 %r233, %r232, %r230;
shl.b32 %r234, %r203, 11;
and.b32 %r235, %r234, 4096;
or.b32 %r236, %r233, %r235;
shl.b32 %r237, %r203, 18;
and.b32 %r238, %r237, 1048576;
or.b32 %r239, %r236, %r238;
shl.b32 %r240, %r203, 25;
and.b32 %r241, %r240, 268435456;
or.b32 %r91, %r239, %r241;

dp4a.s32.s32 %r90, %r91, %r71, %r86;

shr.u32 %r242, %r202, 4;
and.b32 %r243, %r242, 252645135;
shr.u32 %r244, %r203, 12;
and.b32 %r245, %r244, 16;
or.b32 %r246, %r245, %r243;
shr.u32 %r247, %r203, 5;
and.b32 %r248, %r247, 4096;
or.b32 %r249, %r246, %r248;
shl.b32 %r250, %r203, 2;
and.b32 %r251, %r250, 1048576;
or.b32 %r252, %r249, %r251;
shl.b32 %r253, %r203, 9;
and.b32 %r254, %r253, 268435456;
or.b32 %r95, %r252, %r254;

dp4a.s32.s32 %r94, %r95, %r75, %r90;

ld.global.nc.u32 %r99, [%rd35];

{mul.f16x2 %r98,%r99,%r79;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r98;
cvt.f32.f16 %f27, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r98;
cvt.f32.f16 %f28, high;}


cvt.rn.f32.s32 %f37, %r94;
div.approx.ftz.f32 %f38, %f28, %f34;
fma.rn.ftz.f32 %f39, %f27, %f37, %f38;
add.ftz.f32 %f112, %f39, %f112;
add.s64 %rd37, %rd34, %rd4;
add.s64 %rd38, %rd34, %rd5;
add.s64 %rd39, %rd36, %rd4;
ld.global.nc.u32 %r255, [%rd38];
shr.s32 %r256, %r255, %r13;
ld.global.nc.u32 %r257, [%rd39+4];
shr.s32 %r258, %r255, %r14;
ld.global.nc.u32 %r259, [%rd39];
and.b32 %r260, %r259, 252645135;
shl.b32 %r261, %r256, 4;
and.b32 %r262, %r261, 16;
or.b32 %r263, %r262, %r260;
shl.b32 %r264, %r256, 11;
and.b32 %r265, %r264, 4096;
or.b32 %r266, %r263, %r265;
shl.b32 %r267, %r256, 18;
and.b32 %r268, %r267, 1048576;
or.b32 %r269, %r266, %r268;
shl.b32 %r270, %r256, 25;
and.b32 %r271, %r270, 268435456;
or.b32 %r104, %r269, %r271;

dp4a.s32.s32 %r103, %r104, %r63, %r127;

shr.u32 %r272, %r259, 4;
and.b32 %r273, %r272, 252645135;
shr.u32 %r274, %r256, 12;
and.b32 %r275, %r274, 16;
or.b32 %r276, %r275, %r273;
shr.u32 %r277, %r256, 5;
and.b32 %r278, %r277, 4096;
or.b32 %r279, %r276, %r278;
shl.b32 %r280, %r256, 2;
and.b32 %r281, %r280, 1048576;
or.b32 %r282, %r279, %r281;
shl.b32 %r283, %r256, 9;
and.b32 %r284, %r283, 268435456;
or.b32 %r108, %r282, %r284;

dp4a.s32.s32 %r107, %r108, %r67, %r103;

and.b32 %r285, %r257, 252645135;
shl.b32 %r286, %r258, 4;
and.b32 %r287, %r286, 16;
or.b32 %r288, %r287, %r285;
shl.b32 %r289, %r258, 11;
and.b32 %r290, %r289, 4096;
or.b32 %r291, %r288, %r290;
shl.b32 %r292, %r258, 18;
and.b32 %r293, %r292, 1048576;
or.b32 %r294, %r291, %r293;
shl.b32 %r295, %r258, 25;
and.b32 %r296, %r295, 268435456;
or.b32 %r112, %r294, %r296;

dp4a.s32.s32 %r111, %r112, %r71, %r107;

shr.u32 %r297, %r257, 4;
and.b32 %r298, %r297, 252645135;
shr.u32 %r299, %r258, 12;
and.b32 %r300, %r299, 16;
or.b32 %r301, %r300, %r298;
shr.u32 %r302, %r258, 5;
and.b32 %r303, %r302, 4096;
or.b32 %r304, %r301, %r303;
shl.b32 %r305, %r258, 2;
and.b32 %r306, %r305, 1048576;
or.b32 %r307, %r304, %r306;
shl.b32 %r308, %r258, 9;
and.b32 %r309, %r308, 268435456;
or.b32 %r116, %r307, %r309;

dp4a.s32.s32 %r115, %r116, %r75, %r111;

ld.global.nc.u32 %r120, [%rd38+-4];

{mul.f16x2 %r119,%r120,%r79;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r119;
cvt.f32.f16 %f29, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r119;
cvt.f32.f16 %f30, high;}


cvt.rn.f32.s32 %f40, %r115;
div.approx.ftz.f32 %f41, %f30, %f34;
fma.rn.ftz.f32 %f42, %f29, %f40, %f41;
add.ftz.f32 %f111, %f42, %f111;
add.s64 %rd40, %rd37, %rd5;
add.s64 %rd41, %rd39, %rd4;
ld.global.nc.u32 %r310, [%rd40];
shr.s32 %r311, %r310, %r13;
ld.global.nc.u32 %r312, [%rd41+4];
shr.s32 %r313, %r310, %r14;
ld.global.nc.u32 %r314, [%rd41];
and.b32 %r315, %r314, 252645135;
shl.b32 %r316, %r311, 4;
and.b32 %r317, %r316, 16;
or.b32 %r318, %r317, %r315;
shl.b32 %r319, %r311, 11;
and.b32 %r320, %r319, 4096;
or.b32 %r321, %r318, %r320;
shl.b32 %r322, %r311, 18;
and.b32 %r323, %r322, 1048576;
or.b32 %r324, %r321, %r323;
shl.b32 %r325, %r311, 25;
and.b32 %r326, %r325, 268435456;
or.b32 %r125, %r324, %r326;

dp4a.s32.s32 %r124, %r125, %r63, %r127;

shr.u32 %r327, %r314, 4;
and.b32 %r328, %r327, 252645135;
shr.u32 %r329, %r311, 12;
and.b32 %r330, %r329, 16;
or.b32 %r331, %r330, %r328;
shr.u32 %r332, %r311, 5;
and.b32 %r333, %r332, 4096;
or.b32 %r334, %r331, %r333;
shl.b32 %r335, %r311, 2;
and.b32 %r336, %r335, 1048576;
or.b32 %r337, %r334, %r336;
shl.b32 %r338, %r311, 9;
and.b32 %r339, %r338, 268435456;
or.b32 %r129, %r337, %r339;

dp4a.s32.s32 %r128, %r129, %r67, %r124;

and.b32 %r340, %r312, 252645135;
shl.b32 %r341, %r313, 4;
and.b32 %r342, %r341, 16;
or.b32 %r343, %r342, %r340;
shl.b32 %r344, %r313, 11;
and.b32 %r345, %r344, 4096;
or.b32 %r346, %r343, %r345;
shl.b32 %r347, %r313, 18;
and.b32 %r348, %r347, 1048576;
or.b32 %r349, %r346, %r348;
shl.b32 %r350, %r313, 25;
and.b32 %r351, %r350, 268435456;
or.b32 %r133, %r349, %r351;

dp4a.s32.s32 %r132, %r133, %r71, %r128;

shr.u32 %r352, %r312, 4;
and.b32 %r353, %r352, 252645135;
shr.u32 %r354, %r313, 12;
and.b32 %r355, %r354, 16;
or.b32 %r356, %r355, %r353;
shr.u32 %r357, %r313, 5;
and.b32 %r358, %r357, 4096;
or.b32 %r359, %r356, %r358;
shl.b32 %r360, %r313, 2;
and.b32 %r361, %r360, 1048576;
or.b32 %r362, %r359, %r361;
shl.b32 %r363, %r313, 9;
and.b32 %r364, %r363, 268435456;
or.b32 %r137, %r362, %r364;

dp4a.s32.s32 %r136, %r137, %r75, %r132;

ld.global.nc.u32 %r141, [%rd40+-4];

{mul.f16x2 %r140,%r141,%r79;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r140;
cvt.f32.f16 %f31, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r140;
cvt.f32.f16 %f32, high;}


cvt.rn.f32.s32 %f43, %r136;
div.approx.ftz.f32 %f44, %f32, %f34;
fma.rn.ftz.f32 %f45, %f31, %f43, %f44;
add.ftz.f32 %f110, %f45, %f110;
add.s64 %rd44, %rd44, 1536;
add.s32 %r449, %r449, 64;
setp.lt.s32 %p3, %r449, %r5;
@%p3 bra $L__BB74_5;

$L__BB74_6:
setp.eq.s32 %p4, %r1, 0;
@%p4 bra $L__BB74_8;

shl.b32 %r365, %r1, 9;
mov.u32 %r366, _ZZ13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r367, %r366, %r365;
shl.b32 %r368, %r2, 2;
add.s32 %r369, %r367, %r368;
st.shared.f32 [%r369+-512], %f113;
st.shared.f32 [%r369+-384], %f112;
st.shared.f32 [%r369+-256], %f111;
st.shared.f32 [%r369+-128], %f110;

$L__BB74_8:
setp.ne.s32 %p5, %r1, 0;
bar.sync 0;
@%p5 bra $L__BB74_17;

mov.u32 %r370, %ctaid.z;
mov.u32 %r371, %ctaid.y;
mad.lo.s32 %r372, %r371, %r29, %r4;
mad.lo.s32 %r373, %r370, %r35, %r372;
shl.b32 %r374, %r2, 2;
mov.u32 %r375, _ZZ13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r17, %r375, %r374;
mov.u32 %r376, 2;
ld.shared.f32 %f46, [%r17];
add.ftz.f32 %f47, %f46, %f113;
ld.shared.f32 %f48, [%r17+512];
add.ftz.f32 %f49, %f48, %f47;
ld.shared.f32 %f50, [%r17+1024];
add.ftz.f32 %f51, %f50, %f49;
mov.b32 %r377, %f51;
mov.u32 %r378, 31;
mov.u32 %r379, 16;
mov.u32 %r380, -1;
shfl.sync.bfly.b32 %r381|%p6, %r377, %r379, %r378, %r380;
mov.b32 %f52, %r381;
add.ftz.f32 %f53, %f51, %f52;
mov.b32 %r382, %f53;
mov.u32 %r383, 8;
shfl.sync.bfly.b32 %r384|%p7, %r382, %r383, %r378, %r380;
mov.b32 %f54, %r384;
add.ftz.f32 %f55, %f53, %f54;
mov.b32 %r385, %f55;
mov.u32 %r386, 4;
shfl.sync.bfly.b32 %r387|%p8, %r385, %r386, %r378, %r380;
mov.b32 %f56, %r387;
add.ftz.f32 %f57, %f55, %f56;
mov.b32 %r388, %f57;
shfl.sync.bfly.b32 %r389|%p9, %r388, %r376, %r378, %r380;
mov.b32 %f58, %r389;
add.ftz.f32 %f59, %f57, %f58;
mov.b32 %r390, %f59;
mov.u32 %r391, 1;
shfl.sync.bfly.b32 %r392|%p10, %r390, %r391, %r378, %r380;
mov.b32 %f60, %r392;
add.ftz.f32 %f13, %f59, %f60;
setp.ge.u32 %p11, %r4, %r23;
setp.ne.s32 %p12, %r2, 0;
cvta.to.global.u64 %rd42, %rd15;
mul.wide.u32 %rd43, %r373, 4;
add.s64 %rd11, %rd42, %rd43;
or.pred %p13, %p12, %p11;
@%p13 bra $L__BB74_11;

st.global.f32 [%rd11], %f13;

$L__BB74_11:
ld.shared.f32 %f61, [%r17+128];
add.ftz.f32 %f62, %f61, %f112;
ld.shared.f32 %f63, [%r17+640];
add.ftz.f32 %f64, %f63, %f62;
ld.shared.f32 %f65, [%r17+1152];
add.ftz.f32 %f66, %f65, %f64;
mov.b32 %r393, %f66;
shfl.sync.bfly.b32 %r397|%p14, %r393, %r379, %r378, %r380;
mov.b32 %f67, %r397;
add.ftz.f32 %f68, %f66, %f67;
mov.b32 %r398, %f68;
shfl.sync.bfly.b32 %r400|%p15, %r398, %r383, %r378, %r380;
mov.b32 %f69, %r400;
add.ftz.f32 %f70, %f68, %f69;
mov.b32 %r401, %f70;
shfl.sync.bfly.b32 %r403|%p16, %r401, %r386, %r378, %r380;
mov.b32 %f71, %r403;
add.ftz.f32 %f72, %f70, %f71;
mov.b32 %r404, %f72;
shfl.sync.bfly.b32 %r406|%p17, %r404, %r376, %r378, %r380;
mov.b32 %f73, %r406;
add.ftz.f32 %f74, %f72, %f73;
mov.b32 %r407, %f74;
shfl.sync.bfly.b32 %r409|%p18, %r407, %r391, %r378, %r380;
mov.b32 %f75, %r409;
add.ftz.f32 %f14, %f74, %f75;
add.s32 %r410, %r4, 1;
setp.ge.u32 %p19, %r410, %r23;
setp.ne.s32 %p20, %r2, 1;
or.pred %p21, %p20, %p19;
@%p21 bra $L__BB74_13;

st.global.f32 [%rd11+4], %f14;

$L__BB74_13:
ld.shared.f32 %f76, [%r17+256];
add.ftz.f32 %f77, %f76, %f111;
ld.shared.f32 %f78, [%r17+768];
add.ftz.f32 %f79, %f78, %f77;
ld.shared.f32 %f80, [%r17+1280];
add.ftz.f32 %f81, %f80, %f79;
mov.b32 %r411, %f81;
mov.u32 %r412, 31;
mov.u32 %r413, 16;
mov.u32 %r414, -1;
shfl.sync.bfly.b32 %r415|%p22, %r411, %r413, %r412, %r414;
mov.b32 %f82, %r415;
add.ftz.f32 %f83, %f81, %f82;
mov.b32 %r416, %f83;
mov.u32 %r417, 8;
shfl.sync.bfly.b32 %r418|%p23, %r416, %r417, %r412, %r414;
mov.b32 %f84, %r418;
add.ftz.f32 %f85, %f83, %f84;
mov.b32 %r419, %f85;
mov.u32 %r420, 4;
shfl.sync.bfly.b32 %r421|%p24, %r419, %r420, %r412, %r414;
mov.b32 %f86, %r421;
add.ftz.f32 %f87, %f85, %f86;
mov.b32 %r422, %f87;
mov.u32 %r423, 2;
shfl.sync.bfly.b32 %r424|%p25, %r422, %r423, %r412, %r414;
mov.b32 %f88, %r424;
add.ftz.f32 %f89, %f87, %f88;
mov.b32 %r425, %f89;
mov.u32 %r426, 1;
shfl.sync.bfly.b32 %r427|%p26, %r425, %r426, %r412, %r414;
mov.b32 %f90, %r427;
add.ftz.f32 %f15, %f89, %f90;
add.s32 %r428, %r4, 2;
setp.ge.u32 %p27, %r428, %r23;
setp.ne.s32 %p28, %r2, 2;
or.pred %p29, %p28, %p27;
@%p29 bra $L__BB74_15;

st.global.f32 [%rd11+8], %f15;

$L__BB74_15:
ld.shared.f32 %f91, [%r17+384];
add.ftz.f32 %f92, %f91, %f110;
ld.shared.f32 %f93, [%r17+896];
add.ftz.f32 %f94, %f93, %f92;
ld.shared.f32 %f95, [%r17+1408];
add.ftz.f32 %f96, %f95, %f94;
mov.b32 %r429, %f96;
shfl.sync.bfly.b32 %r433|%p30, %r429, %r413, %r412, %r414;
mov.b32 %f97, %r433;
add.ftz.f32 %f98, %f96, %f97;
mov.b32 %r434, %f98;
shfl.sync.bfly.b32 %r436|%p31, %r434, %r417, %r412, %r414;
mov.b32 %f99, %r436;
add.ftz.f32 %f100, %f98, %f99;
mov.b32 %r437, %f100;
shfl.sync.bfly.b32 %r439|%p32, %r437, %r420, %r412, %r414;
mov.b32 %f101, %r439;
add.ftz.f32 %f102, %f100, %f101;
mov.b32 %r440, %f102;
shfl.sync.bfly.b32 %r442|%p33, %r440, %r423, %r412, %r414;
mov.b32 %f103, %r442;
add.ftz.f32 %f104, %f102, %f103;
mov.b32 %r443, %f104;
shfl.sync.bfly.b32 %r445|%p34, %r443, %r426, %r412, %r414;
mov.b32 %f105, %r445;
add.ftz.f32 %f16, %f104, %f105;
add.s32 %r446, %r4, 3;
setp.ge.u32 %p35, %r446, %r23;
setp.ne.s32 %p36, %r2, 3;
or.pred %p37, %p36, %p35;
@%p37 bra $L__BB74_17;

st.global.f32 [%rd11+12], %f16;

$L__BB74_17:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<34>;
.reg .f32 %f<124>;
.reg .b32 %r<302>;
.reg .b64 %rd<71>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[384];

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate[384];

ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd25, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
ld.param.f32 %f31, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
ld.param.u64 %rd22, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+16];
ld.param.u64 %rd21, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+8];
ld.param.u64 %rd20, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3];
mov.u32 %r1, %tid.y;
shl.b32 %r42, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r42, %r2;
mov.u32 %r4, %ctaid.y;
setp.eq.s64 %p2, %rd19, 0;
@%p2 bra $L__BB75_2;

cvta.to.global.u64 %rd26, %rd19;
mul.wide.u32 %rd27, %r4, 4;
add.s64 %rd28, %rd26, %rd27;
ld.global.nc.u32 %r299, [%rd28];
mul.hi.u32 %r43, %r4, %r26;
add.s32 %r44, %r43, %r4;
shr.u32 %r45, %r44, %r27;
mul.lo.s32 %r46, %r45, %r28;
sub.s32 %r300, %r4, %r46;
bra.uni $L__BB75_3;

$L__BB75_2:
mul.hi.u32 %r47, %r4, %r30;
add.s32 %r48, %r47, %r4;
shr.u32 %r299, %r48, %r31;
mov.u32 %r300, %r4;

$L__BB75_3:
or.b32 %r49, %r2, %r1;
setp.ne.s32 %p3, %r49, 0;
mov.f32 %f117, 0f00000000;
mov.f32 %f113, %f117;
mov.f32 %f114, %f117;
@%p3 bra $L__BB75_8;

setp.eq.s64 %p4, %rd20, 0;
mov.f32 %f113, 0f00000000;
mov.f32 %f114, %f113;
@%p4 bra $L__BB75_6;

cvta.to.global.u64 %rd29, %rd20;
mov.u32 %r50, %ctaid.z;
mul.lo.s32 %r51, %r50, %r41;
cvt.u64.u32 %rd30, %r51;
selp.b32 %r52, %r4, %r299, %p2;
mul.lo.s32 %r53, %r52, %r35;
cvt.u64.u32 %rd31, %r53;
mov.u32 %r54, %ctaid.x;
cvt.s64.s32 %rd32, %r54;
cvt.u64.u32 %rd33, %r2;
add.s64 %rd34, %rd33, %rd32;
add.s64 %rd35, %rd34, %rd30;
add.s64 %rd36, %rd35, %rd31;
shl.b64 %rd37, %rd36, 2;
add.s64 %rd38, %rd29, %rd37;
ld.global.f32 %f114, [%rd38];

$L__BB75_6:
setp.eq.s64 %p6, %rd22, 0;
setp.eq.s64 %p7, %rd21, 0;
or.pred %p8, %p6, %p7;
@%p8 bra $L__BB75_8;

cvta.to.global.u64 %rd39, %rd22;
mov.u32 %r56, %ctaid.z;
mul.lo.s32 %r57, %r56, %r41;
cvt.u64.u32 %rd40, %r57;
selp.b32 %r58, %r4, %r299, %p2;
mul.lo.s32 %r59, %r58, %r35;
cvt.u64.u32 %rd41, %r59;
mov.u32 %r60, %ctaid.x;
cvt.s64.s32 %rd42, %r60;
cvt.u64.u32 %rd43, %r2;
add.s64 %rd44, %rd43, %rd42;
add.s64 %rd45, %rd44, %rd40;
add.s64 %rd46, %rd45, %rd41;
shl.b64 %rd47, %rd46, 2;
add.s64 %rd48, %rd39, %rd47;
ld.global.f32 %f113, [%rd48];

$L__BB75_8:
shr.u32 %r62, %r25, 5;
shr.u32 %r63, %r3, 31;
add.s32 %r64, %r3, %r63;
shr.s32 %r301, %r64, 1;
setp.ge.s32 %p10, %r301, %r62;
mov.f32 %f119, %f117;
@%p10 bra $L__BB75_13;

mov.u32 %r65, %ctaid.z;
and.b32 %r68, %r64, -2;
sub.s32 %r13, %r3, %r68;
shl.b32 %r69, %r13, 1;
mul.hi.u32 %r70, %r65, %r36;
add.s32 %r71, %r70, %r65;
shr.u32 %r72, %r71, %r37;
mad.lo.s32 %r73, %r72, %r39, %r301;
mad.lo.s32 %r74, %r299, %r33, %r73;
mov.u32 %r75, %ctaid.x;
mad.lo.s32 %r76, %r75, %r29, %r74;
mul.wide.s32 %rd49, %r76, 24;
mul.wide.s32 %rd50, %r69, 4;
add.s64 %rd51, %rd49, %rd50;
add.s64 %rd52, %rd51, 12;
cvta.to.global.u64 %rd53, %rd21;
add.s64 %rd70, %rd53, %rd52;
cvta.to.global.u64 %rd54, %rd17;
add.s64 %rd69, %rd54, %rd52;
add.s64 %rd68, %rd53, %rd49;
add.s64 %rd67, %rd54, %rd49;
mul.lo.s32 %r77, %r300, %r34;
cvt.u64.u32 %rd55, %r77;
mul.lo.s32 %r78, %r65, %r40;
cvt.u64.u32 %rd56, %r78;
add.s64 %rd8, %rd55, %rd56;
mov.f32 %f119, 0f00000000;
cvta.to.global.u64 %rd57, %rd18;
mov.f32 %f117, %f119;

$L__BB75_10:
cvt.u64.u32 %rd58, %r301;
add.s64 %rd59, %rd8, %rd58;
mul.lo.s64 %rd60, %rd59, 36;
add.s64 %rd61, %rd57, %rd60;
add.s64 %rd63, %rd61, %rd50;
ld.global.nc.u32 %r101, [%rd67+4];
shl.b32 %r102, %r13, 3;
shr.s32 %r103, %r101, %r102;
ld.global.nc.u32 %r15, [%rd63+4];
ld.global.nc.u32 %r16, [%rd63+20];
ld.global.nc.u32 %r104, [%rd69];
or.b32 %r105, %r102, 4;
shr.s32 %r106, %r101, %r105;
ld.global.nc.u32 %r17, [%rd63+8];
ld.global.nc.u32 %r18, [%rd63+24];
ld.global.nc.u32 %r107, [%rd69+-4];
and.b32 %r108, %r107, 252645135;
shl.b32 %r109, %r103, 4;
and.b32 %r110, %r109, 16;
or.b32 %r111, %r110, %r108;
shl.b32 %r112, %r103, 11;
and.b32 %r113, %r112, 4096;
or.b32 %r114, %r111, %r113;
shl.b32 %r115, %r103, 18;
and.b32 %r116, %r115, 1048576;
or.b32 %r117, %r114, %r116;
shl.b32 %r118, %r103, 25;
and.b32 %r119, %r118, 268435456;
or.b32 %r80, %r117, %r119;
mov.u32 %r82, 0;

dp4a.s32.s32 %r79, %r80, %r15, %r82;

shr.u32 %r120, %r107, 4;
and.b32 %r121, %r120, 252645135;
shr.u32 %r122, %r103, 12;
and.b32 %r123, %r122, 16;
or.b32 %r124, %r123, %r121;
shr.u32 %r125, %r103, 5;
and.b32 %r126, %r125, 4096;
or.b32 %r127, %r124, %r126;
shl.b32 %r128, %r103, 2;
and.b32 %r129, %r128, 1048576;
or.b32 %r130, %r127, %r129;
shl.b32 %r131, %r103, 9;
and.b32 %r132, %r131, 268435456;
or.b32 %r84, %r130, %r132;

dp4a.s32.s32 %r83, %r84, %r16, %r79;

and.b32 %r133, %r104, 252645135;
shl.b32 %r134, %r106, 4;
and.b32 %r135, %r134, 16;
or.b32 %r136, %r135, %r133;
shl.b32 %r137, %r106, 11;
and.b32 %r138, %r137, 4096;
or.b32 %r139, %r136, %r138;
shl.b32 %r140, %r106, 18;
and.b32 %r141, %r140, 1048576;
or.b32 %r142, %r139, %r141;
shl.b32 %r143, %r106, 25;
and.b32 %r144, %r143, 268435456;
or.b32 %r88, %r142, %r144;

dp4a.s32.s32 %r87, %r88, %r17, %r83;

shr.u32 %r145, %r104, 4;
and.b32 %r146, %r145, 252645135;
shr.u32 %r147, %r106, 12;
and.b32 %r148, %r147, 16;
or.b32 %r149, %r148, %r146;
shr.u32 %r150, %r106, 5;
and.b32 %r151, %r150, 4096;
or.b32 %r152, %r149, %r151;
shl.b32 %r153, %r106, 2;
and.b32 %r154, %r153, 1048576;
or.b32 %r155, %r152, %r154;
shl.b32 %r156, %r106, 9;
and.b32 %r157, %r156, 268435456;
or.b32 %r92, %r155, %r157;

dp4a.s32.s32 %r91, %r92, %r18, %r87;

ld.global.nc.u32 %r96, [%rd67];
ld.global.nc.u32 %r19, [%rd61];

{mul.f16x2 %r95,%r96,%r19;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r95;
cvt.f32.f16 %f40, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r95;
cvt.f32.f16 %f41, high;}


cvt.rn.f32.s32 %f42, %r91;
mov.f32 %f43, 0f40000000;
div.approx.ftz.f32 %f44, %f41, %f43;
fma.rn.ftz.f32 %f45, %f40, %f42, %f44;
add.ftz.f32 %f119, %f119, %f45;
setp.eq.s64 %p11, %rd21, 0;
@%p11 bra $L__BB75_12;

ld.global.u32 %r179, [%rd68+4];
shr.s32 %r181, %r179, %r102;
ld.global.u32 %r182, [%rd70];
shr.s32 %r184, %r179, %r105;
ld.global.u32 %r185, [%rd70+-4];
and.b32 %r186, %r185, 252645135;
shl.b32 %r187, %r181, 4;
and.b32 %r188, %r187, 16;
or.b32 %r189, %r188, %r186;
shl.b32 %r190, %r181, 11;
and.b32 %r191, %r190, 4096;
or.b32 %r192, %r189, %r191;
shl.b32 %r193, %r181, 18;
and.b32 %r194, %r193, 1048576;
or.b32 %r195, %r192, %r194;
shl.b32 %r196, %r181, 25;
and.b32 %r197, %r196, 268435456;
or.b32 %r159, %r195, %r197;

dp4a.s32.s32 %r158, %r159, %r15, %r82;

shr.u32 %r198, %r185, 4;
and.b32 %r199, %r198, 252645135;
shr.u32 %r200, %r181, 12;
and.b32 %r201, %r200, 16;
or.b32 %r202, %r201, %r199;
shr.u32 %r203, %r181, 5;
and.b32 %r204, %r203, 4096;
or.b32 %r205, %r202, %r204;
shl.b32 %r206, %r181, 2;
and.b32 %r207, %r206, 1048576;
or.b32 %r208, %r205, %r207;
shl.b32 %r209, %r181, 9;
and.b32 %r210, %r209, 268435456;
or.b32 %r163, %r208, %r210;

dp4a.s32.s32 %r162, %r163, %r16, %r158;

and.b32 %r211, %r182, 252645135;
shl.b32 %r212, %r184, 4;
and.b32 %r213, %r212, 16;
or.b32 %r214, %r213, %r211;
shl.b32 %r215, %r184, 11;
and.b32 %r216, %r215, 4096;
or.b32 %r217, %r214, %r216;
shl.b32 %r218, %r184, 18;
and.b32 %r219, %r218, 1048576;
or.b32 %r220, %r217, %r219;
shl.b32 %r221, %r184, 25;
and.b32 %r222, %r221, 268435456;
or.b32 %r167, %r220, %r222;

dp4a.s32.s32 %r166, %r167, %r17, %r162;

shr.u32 %r223, %r182, 4;
and.b32 %r224, %r223, 252645135;
shr.u32 %r225, %r184, 12;
and.b32 %r226, %r225, 16;
or.b32 %r227, %r226, %r224;
shr.u32 %r228, %r184, 5;
and.b32 %r229, %r228, 4096;
or.b32 %r230, %r227, %r229;
shl.b32 %r231, %r184, 2;
and.b32 %r232, %r231, 1048576;
or.b32 %r233, %r230, %r232;
shl.b32 %r234, %r184, 9;
and.b32 %r235, %r234, 268435456;
or.b32 %r171, %r233, %r235;

dp4a.s32.s32 %r170, %r171, %r18, %r166;

ld.global.u32 %r175, [%rd68];

{mul.f16x2 %r174,%r175,%r19;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r174;
cvt.f32.f16 %f46, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r174;
cvt.f32.f16 %f47, high;}


cvt.rn.f32.s32 %f48, %r170;
div.approx.ftz.f32 %f50, %f47, %f43;
fma.rn.ftz.f32 %f51, %f46, %f48, %f50;
add.ftz.f32 %f117, %f117, %f51;

$L__BB75_12:
add.s64 %rd70, %rd70, 1536;
add.s64 %rd69, %rd69, 1536;
add.s64 %rd68, %rd68, 1536;
add.s64 %rd67, %rd67, 1536;
add.s32 %r301, %r301, 64;
setp.lt.s32 %p12, %r301, %r62;
@%p12 bra $L__BB75_10;

$L__BB75_13:
setp.eq.s32 %p13, %r1, 0;
@%p13 bra $L__BB75_16;

setp.eq.s64 %p14, %rd21, 0;
add.s32 %r21, %r1, -1;
shl.b32 %r240, %r1, 7;
mov.u32 %r241, _ZZ13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r242, %r241, %r240;
shl.b32 %r243, %r2, 2;
add.s32 %r244, %r242, %r243;
st.shared.f32 [%r244+-128], %f119;
@%p14 bra $L__BB75_16;

shl.b32 %r246, %r21, 7;
mov.u32 %r247, _ZZ13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r248, %r247, %r246;
add.s32 %r250, %r248, %r243;
st.shared.f32 [%r250], %f117;

$L__BB75_16:
setp.ne.s32 %p15, %r1, 0;
bar.sync 0;
@%p15 bra $L__BB75_35;

setp.ne.s64 %p16, %rd21, 0;
shl.b32 %r253, %r2, 2;
mov.u32 %r254, _ZZ13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r22, %r254, %r253;
ld.shared.f32 %f52, [%r22];
add.ftz.f32 %f13, %f119, %f52;
@%p16 bra $L__BB75_19;
bra.uni $L__BB75_18;

$L__BB75_19:
mov.u32 %r257, _ZZ13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r258, %r257, %r253;
ld.shared.f32 %f56, [%r258];
add.ftz.f32 %f57, %f117, %f56;
ld.shared.f32 %f58, [%r22+128];
add.ftz.f32 %f59, %f13, %f58;
ld.shared.f32 %f60, [%r258+128];
add.ftz.f32 %f61, %f57, %f60;
ld.shared.f32 %f62, [%r22+256];
add.ftz.f32 %f121, %f59, %f62;
ld.shared.f32 %f63, [%r258+256];
add.ftz.f32 %f117, %f61, %f63;
bra.uni $L__BB75_20;

$L__BB75_18:
ld.shared.f32 %f53, [%r22+128];
add.ftz.f32 %f54, %f13, %f53;
ld.shared.f32 %f55, [%r22+256];
add.ftz.f32 %f121, %f54, %f55;

$L__BB75_20:
mov.b32 %r259, %f121;
mov.u32 %r260, 31;
mov.u32 %r261, 16;
mov.u32 %r262, -1;
shfl.sync.bfly.b32 %r263|%p17, %r259, %r261, %r260, %r262;
mov.b32 %f64, %r263;
add.ftz.f32 %f65, %f121, %f64;
mov.b32 %r264, %f65;
mov.u32 %r265, 8;
shfl.sync.bfly.b32 %r266|%p18, %r264, %r265, %r260, %r262;
mov.b32 %f66, %r266;
add.ftz.f32 %f67, %f65, %f66;
mov.b32 %r267, %f67;
mov.u32 %r268, 4;
shfl.sync.bfly.b32 %r269|%p19, %r267, %r268, %r260, %r262;
mov.b32 %f68, %r269;
add.ftz.f32 %f69, %f67, %f68;
mov.b32 %r270, %f69;
mov.u32 %r271, 2;
shfl.sync.bfly.b32 %r272|%p20, %r270, %r271, %r260, %r262;
mov.b32 %f70, %r272;
add.ftz.f32 %f19, %f69, %f70;
mov.b32 %r273, %f19;
mov.u32 %r274, 1;
shfl.sync.bfly.b32 %r23|%p1, %r273, %r274, %r260, %r262;
setp.eq.s64 %p21, %rd21, 0;
@%p21 bra $L__BB75_22;

mov.b32 %r275, %f117;
shfl.sync.bfly.b32 %r279|%p22, %r275, %r261, %r260, %r262;
mov.b32 %f71, %r279;
add.ftz.f32 %f72, %f117, %f71;
mov.b32 %r280, %f72;
shfl.sync.bfly.b32 %r282|%p23, %r280, %r265, %r260, %r262;
mov.b32 %f73, %r282;
add.ftz.f32 %f74, %f72, %f73;
mov.b32 %r283, %f74;
shfl.sync.bfly.b32 %r285|%p24, %r283, %r268, %r260, %r262;
mov.b32 %f75, %r285;
add.ftz.f32 %f76, %f74, %f75;
mov.b32 %r286, %f76;
shfl.sync.bfly.b32 %r288|%p25, %r286, %r271, %r260, %r262;
mov.b32 %f77, %r288;
add.ftz.f32 %f78, %f76, %f77;
mov.b32 %r289, %f78;
shfl.sync.bfly.b32 %r291|%p26, %r289, %r274, %r260, %r262;
mov.b32 %f79, %r291;
add.ftz.f32 %f117, %f78, %f79;

$L__BB75_22:
setp.ne.s32 %p27, %r2, 0;
mov.b32 %f80, %r23;
add.ftz.f32 %f22, %f19, %f80;
@%p27 bra $L__BB75_35;

add.ftz.f32 %f123, %f114, %f22;
@%p21 bra $L__BB75_34;

add.ftz.f32 %f24, %f113, %f117;
setp.gt.s32 %p29, %r24, 2;
@%p29 bra $L__BB75_28;

setp.eq.s32 %p32, %r24, 1;
@%p32 bra $L__BB75_32;

setp.eq.s32 %p33, %r24, 2;
@%p33 bra $L__BB75_27;
bra.uni $L__BB75_33;

$L__BB75_27:
mul.ftz.f32 %f100, %f24, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f101, %f100;
add.ftz.f32 %f102, %f101, 0f3F800000;
div.approx.ftz.f32 %f103, %f24, %f102;
mul.ftz.f32 %f123, %f123, %f103;
bra.uni $L__BB75_34;

$L__BB75_28:
setp.eq.s32 %p30, %r24, 3;
@%p30 bra $L__BB75_31;

setp.ne.s32 %p31, %r24, 6;
@%p31 bra $L__BB75_33;

min.ftz.f32 %f81, %f24, %f31;
neg.ftz.f32 %f82, %f31;
min.ftz.f32 %f83, %f123, %f31;
max.ftz.f32 %f84, %f83, %f82;
mul.ftz.f32 %f85, %f81, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f86, %f85;
add.ftz.f32 %f87, %f86, 0f3F800000;
div.approx.ftz.f32 %f88, %f81, %f87;
mul.ftz.f32 %f123, %f84, %f88;
bra.uni $L__BB75_34;

$L__BB75_32:
mul.ftz.f32 %f104, %f24, 0f3F000000;
mul.ftz.f32 %f105, %f24, 0f3D372713;
fma.rn.ftz.f32 %f106, %f24, %f105, 0f3F800000;
mul.ftz.f32 %f107, %f24, 0f3F4C422A;
mul.ftz.f32 %f108, %f107, %f106;
tanh.approx.f32 %f109, %f108;
add.ftz.f32 %f110, %f109, 0f3F800000;
mul.ftz.f32 %f111, %f104, %f110;
mul.ftz.f32 %f123, %f123, %f111;
bra.uni $L__BB75_34;

$L__BB75_31:
mov.f32 %f89, 0f40E00000;
min.ftz.f32 %f90, %f24, %f89;
min.ftz.f32 %f91, %f123, %f89;
mov.f32 %f92, 0fC0E00000;
max.ftz.f32 %f93, %f91, %f92;
mul.ftz.f32 %f94, %f90, 0f3FD9DB23;
mul.ftz.f32 %f95, %f94, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f96, %f95;
add.ftz.f32 %f97, %f96, 0f3F800000;
div.approx.ftz.f32 %f98, %f90, %f97;
add.ftz.f32 %f99, %f93, 0f3F800000;
mul.ftz.f32 %f123, %f99, %f98;
bra.uni $L__BB75_34;

$L__BB75_33:
mul.ftz.f32 %f123, %f123, %f24;

$L__BB75_34:
cvta.to.global.u64 %rd64, %rd25;
mov.u32 %r294, %ctaid.z;
mov.u32 %r295, %ctaid.y;
mov.u32 %r296, %ctaid.x;
mad.lo.s32 %r297, %r295, %r35, %r296;
mad.lo.s32 %r298, %r294, %r41, %r297;
mul.wide.u32 %rd65, %r298, 4;
add.s64 %rd66, %rd64, %rd65;
st.global.f32 [%rd66], %f123;

$L__BB75_35:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<12>;
.reg .f32 %f<30>;
.reg .b32 %r<165>;
.reg .b64 %rd<31>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[384];

ld.param.u64 %rd8, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd9, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd10, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd11, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r19, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r18, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
mov.u32 %r1, %tid.y;
shl.b32 %r35, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r35, %r2;
shr.u32 %r4, %r34, 5;
mov.u32 %r163, %ctaid.y;
setp.eq.s64 %p1, %rd10, 0;
@%p1 bra $L__BB76_2;

cvta.to.global.u64 %rd12, %rd10;
mul.wide.u32 %rd13, %r163, 4;
add.s64 %rd14, %rd12, %rd13;
ld.global.nc.u32 %r162, [%rd14];
mul.hi.u32 %r36, %r163, %r18;
add.s32 %r37, %r36, %r163;
shr.u32 %r38, %r37, %r19;
mul.lo.s32 %r39, %r38, %r20;
sub.s32 %r163, %r163, %r39;
bra.uni $L__BB76_3;

$L__BB76_2:
mul.hi.u32 %r40, %r163, %r22;
add.s32 %r41, %r40, %r163;
shr.u32 %r162, %r41, %r23;

$L__BB76_3:
mov.u32 %r11, %ctaid.z;
shr.u32 %r42, %r3, 31;
add.s32 %r43, %r3, %r42;
shr.s32 %r164, %r43, 1;
setp.ge.s32 %p2, %r164, %r4;
mov.f32 %f29, 0f00000000;
@%p2 bra $L__BB76_6;

and.b32 %r46, %r43, -2;
sub.s32 %r47, %r3, %r46;
shl.b32 %r48, %r47, 1;
cvt.s64.s32 %rd1, %r48;
shl.b32 %r13, %r47, 3;
or.b32 %r14, %r13, 4;
mul.hi.u32 %r49, %r11, %r28;
add.s32 %r50, %r49, %r11;
shr.u32 %r51, %r50, %r29;
mad.lo.s32 %r52, %r51, %r31, %r164;
mad.lo.s32 %r53, %r162, %r25, %r52;
mov.u32 %r54, %ctaid.x;
mad.lo.s32 %r55, %r54, %r21, %r53;
mul.wide.s32 %rd15, %r55, 24;
cvta.to.global.u64 %rd16, %rd8;
add.s64 %rd30, %rd16, %rd15;
mul.wide.s32 %rd17, %r48, 4;
add.s64 %rd3, %rd17, 8;
cvta.to.global.u64 %rd4, %rd9;
mul.lo.s32 %r56, %r163, %r26;
cvt.u64.u32 %rd18, %r56;
mul.lo.s32 %r57, %r11, %r32;
cvt.u64.u32 %rd19, %r57;
add.s64 %rd5, %rd18, %rd19;
mov.f32 %f29, 0f00000000;
shl.b64 %rd24, %rd1, 2;

$L__BB76_5:
cvt.u64.u32 %rd20, %r164;
add.s64 %rd21, %rd5, %rd20;
mul.lo.s64 %rd22, %rd21, 36;
add.s64 %rd23, %rd4, %rd22;
add.s64 %rd25, %rd23, %rd24;
add.s64 %rd26, %rd30, %rd3;
ld.global.nc.u32 %r79, [%rd30+4];
shr.s32 %r80, %r79, %r13;
ld.global.nc.u32 %r60, [%rd25+4];
ld.global.nc.u32 %r64, [%rd25+20];
ld.global.nc.u32 %r81, [%rd26+4];
shr.s32 %r82, %r79, %r14;
ld.global.nc.u32 %r68, [%rd25+8];
ld.global.nc.u32 %r72, [%rd25+24];
ld.global.nc.u32 %r83, [%rd26];
and.b32 %r84, %r83, 252645135;
shl.b32 %r85, %r80, 4;
and.b32 %r86, %r85, 16;
or.b32 %r87, %r86, %r84;
shl.b32 %r88, %r80, 11;
and.b32 %r89, %r88, 4096;
or.b32 %r90, %r87, %r89;
shl.b32 %r91, %r80, 18;
and.b32 %r92, %r91, 1048576;
or.b32 %r93, %r90, %r92;
shl.b32 %r94, %r80, 25;
and.b32 %r95, %r94, 268435456;
or.b32 %r59, %r93, %r95;
mov.u32 %r61, 0;

dp4a.s32.s32 %r58, %r59, %r60, %r61;

shr.u32 %r96, %r83, 4;
and.b32 %r97, %r96, 252645135;
shr.u32 %r98, %r80, 12;
and.b32 %r99, %r98, 16;
or.b32 %r100, %r99, %r97;
shr.u32 %r101, %r80, 5;
and.b32 %r102, %r101, 4096;
or.b32 %r103, %r100, %r102;
shl.b32 %r104, %r80, 2;
and.b32 %r105, %r104, 1048576;
or.b32 %r106, %r103, %r105;
shl.b32 %r107, %r80, 9;
and.b32 %r108, %r107, 268435456;
or.b32 %r63, %r106, %r108;

dp4a.s32.s32 %r62, %r63, %r64, %r58;

and.b32 %r109, %r81, 252645135;
shl.b32 %r110, %r82, 4;
and.b32 %r111, %r110, 16;
or.b32 %r112, %r111, %r109;
shl.b32 %r113, %r82, 11;
and.b32 %r114, %r113, 4096;
or.b32 %r115, %r112, %r114;
shl.b32 %r116, %r82, 18;
and.b32 %r117, %r116, 1048576;
or.b32 %r118, %r115, %r117;
shl.b32 %r119, %r82, 25;
and.b32 %r120, %r119, 268435456;
or.b32 %r67, %r118, %r120;

dp4a.s32.s32 %r66, %r67, %r68, %r62;

shr.u32 %r121, %r81, 4;
and.b32 %r122, %r121, 252645135;
shr.u32 %r123, %r82, 12;
and.b32 %r124, %r123, 16;
or.b32 %r125, %r124, %r122;
shr.u32 %r126, %r82, 5;
and.b32 %r127, %r126, 4096;
or.b32 %r128, %r125, %r127;
shl.b32 %r129, %r82, 2;
and.b32 %r130, %r129, 1048576;
or.b32 %r131, %r128, %r130;
shl.b32 %r132, %r82, 9;
and.b32 %r133, %r132, 268435456;
or.b32 %r71, %r131, %r133;

dp4a.s32.s32 %r70, %r71, %r72, %r66;

ld.global.nc.u32 %r75, [%rd30];
ld.global.nc.u32 %r76, [%rd23];

{mul.f16x2 %r74,%r75,%r76;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r74;
cvt.f32.f16 %f7, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r74;
cvt.f32.f16 %f8, high;}


cvt.rn.f32.s32 %f9, %r70;
mov.f32 %f10, 0f40000000;
div.approx.ftz.f32 %f11, %f8, %f10;
fma.rn.ftz.f32 %f12, %f7, %f9, %f11;
add.ftz.f32 %f29, %f29, %f12;
add.s64 %rd30, %rd30, 1536;
add.s32 %r164, %r164, 64;
setp.lt.s32 %p3, %r164, %r4;
@%p3 bra $L__BB76_5;

$L__BB76_6:
setp.eq.s32 %p4, %r1, 0;
@%p4 bra $L__BB76_8;

shl.b32 %r134, %r1, 7;
mov.u32 %r135, _ZZ13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r136, %r135, %r134;
shl.b32 %r137, %r2, 2;
add.s32 %r138, %r136, %r137;
st.shared.f32 [%r138+-128], %f29;

$L__BB76_8:
setp.ne.s32 %p5, %r1, 0;
bar.sync 0;
@%p5 bra $L__BB76_11;

mov.u32 %r139, %ctaid.y;
mov.u32 %r140, %ctaid.x;
mad.lo.s32 %r141, %r139, %r27, %r140;
mad.lo.s32 %r17, %r11, %r33, %r141;
shl.b32 %r142, %r2, 2;
mov.u32 %r143, _ZZ13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r144, %r143, %r142;
mov.u32 %r145, 2;
ld.shared.f32 %f13, [%r144];
add.ftz.f32 %f14, %f29, %f13;
ld.shared.f32 %f15, [%r144+128];
add.ftz.f32 %f16, %f14, %f15;
ld.shared.f32 %f17, [%r144+256];
add.ftz.f32 %f18, %f16, %f17;
mov.b32 %r146, %f18;
mov.u32 %r147, 31;
mov.u32 %r148, 16;
mov.u32 %r149, -1;
shfl.sync.bfly.b32 %r150|%p6, %r146, %r148, %r147, %r149;
mov.b32 %f19, %r150;
add.ftz.f32 %f20, %f18, %f19;
mov.b32 %r151, %f20;
mov.u32 %r152, 8;
shfl.sync.bfly.b32 %r153|%p7, %r151, %r152, %r147, %r149;
mov.b32 %f21, %r153;
add.ftz.f32 %f22, %f20, %f21;
mov.b32 %r154, %f22;
mov.u32 %r155, 4;
shfl.sync.bfly.b32 %r156|%p8, %r154, %r155, %r147, %r149;
mov.b32 %f23, %r156;
add.ftz.f32 %f24, %f22, %f23;
mov.b32 %r157, %f24;
shfl.sync.bfly.b32 %r158|%p9, %r157, %r145, %r147, %r149;
mov.b32 %f25, %r158;
add.ftz.f32 %f26, %f24, %f25;
mov.b32 %r159, %f26;
mov.u32 %r160, 1;
shfl.sync.bfly.b32 %r161|%p10, %r159, %r160, %r147, %r149;
mov.b32 %f27, %r161;
add.ftz.f32 %f4, %f26, %f27;
setp.ne.s32 %p11, %r2, 0;
@%p11 bra $L__BB76_11;

cvta.to.global.u64 %rd27, %rd11;
mul.wide.u32 %rd28, %r17, 4;
add.s64 %rd29, %rd27, %rd28;
st.global.f32 [%rd29], %f4;

$L__BB76_11:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<34>;
.reg .f32 %f<124>;
.reg .b32 %r<302>;
.reg .b64 %rd<71>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[384];

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate[384];

ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd25, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
ld.param.f32 %f31, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
ld.param.u64 %rd22, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+16];
ld.param.u64 %rd21, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+8];
ld.param.u64 %rd20, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3];
mov.u32 %r1, %tid.y;
shl.b32 %r42, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r42, %r2;
mov.u32 %r4, %ctaid.y;
setp.eq.s64 %p2, %rd19, 0;
@%p2 bra $L__BB77_2;

cvta.to.global.u64 %rd26, %rd19;
mul.wide.u32 %rd27, %r4, 4;
add.s64 %rd28, %rd26, %rd27;
ld.global.nc.u32 %r299, [%rd28];
mul.hi.u32 %r43, %r4, %r26;
add.s32 %r44, %r43, %r4;
shr.u32 %r45, %r44, %r27;
mul.lo.s32 %r46, %r45, %r28;
sub.s32 %r300, %r4, %r46;
bra.uni $L__BB77_3;

$L__BB77_2:
mul.hi.u32 %r47, %r4, %r30;
add.s32 %r48, %r47, %r4;
shr.u32 %r299, %r48, %r31;
mov.u32 %r300, %r4;

$L__BB77_3:
or.b32 %r49, %r2, %r1;
setp.ne.s32 %p3, %r49, 0;
mov.f32 %f117, 0f00000000;
mov.f32 %f113, %f117;
mov.f32 %f114, %f117;
@%p3 bra $L__BB77_8;

setp.eq.s64 %p4, %rd20, 0;
mov.f32 %f113, 0f00000000;
mov.f32 %f114, %f113;
@%p4 bra $L__BB77_6;

cvta.to.global.u64 %rd29, %rd20;
mov.u32 %r50, %ctaid.z;
mul.lo.s32 %r51, %r50, %r41;
cvt.u64.u32 %rd30, %r51;
selp.b32 %r52, %r4, %r299, %p2;
mul.lo.s32 %r53, %r52, %r35;
cvt.u64.u32 %rd31, %r53;
mov.u32 %r54, %ctaid.x;
cvt.s64.s32 %rd32, %r54;
cvt.u64.u32 %rd33, %r2;
add.s64 %rd34, %rd33, %rd32;
add.s64 %rd35, %rd34, %rd30;
add.s64 %rd36, %rd35, %rd31;
shl.b64 %rd37, %rd36, 2;
add.s64 %rd38, %rd29, %rd37;
ld.global.f32 %f114, [%rd38];

$L__BB77_6:
setp.eq.s64 %p6, %rd22, 0;
setp.eq.s64 %p7, %rd21, 0;
or.pred %p8, %p6, %p7;
@%p8 bra $L__BB77_8;

cvta.to.global.u64 %rd39, %rd22;
mov.u32 %r56, %ctaid.z;
mul.lo.s32 %r57, %r56, %r41;
cvt.u64.u32 %rd40, %r57;
selp.b32 %r58, %r4, %r299, %p2;
mul.lo.s32 %r59, %r58, %r35;
cvt.u64.u32 %rd41, %r59;
mov.u32 %r60, %ctaid.x;
cvt.s64.s32 %rd42, %r60;
cvt.u64.u32 %rd43, %r2;
add.s64 %rd44, %rd43, %rd42;
add.s64 %rd45, %rd44, %rd40;
add.s64 %rd46, %rd45, %rd41;
shl.b64 %rd47, %rd46, 2;
add.s64 %rd48, %rd39, %rd47;
ld.global.f32 %f113, [%rd48];

$L__BB77_8:
shr.u32 %r62, %r25, 5;
shr.u32 %r63, %r3, 31;
add.s32 %r64, %r3, %r63;
shr.s32 %r301, %r64, 1;
setp.ge.s32 %p10, %r301, %r62;
mov.f32 %f119, %f117;
@%p10 bra $L__BB77_13;

mov.u32 %r65, %ctaid.z;
and.b32 %r68, %r64, -2;
sub.s32 %r13, %r3, %r68;
shl.b32 %r69, %r13, 1;
mul.hi.u32 %r70, %r65, %r36;
add.s32 %r71, %r70, %r65;
shr.u32 %r72, %r71, %r37;
mad.lo.s32 %r73, %r72, %r39, %r301;
mad.lo.s32 %r74, %r299, %r33, %r73;
mov.u32 %r75, %ctaid.x;
mad.lo.s32 %r76, %r75, %r29, %r74;
mul.wide.s32 %rd49, %r76, 24;
mul.wide.s32 %rd50, %r69, 4;
add.s64 %rd51, %rd49, %rd50;
add.s64 %rd52, %rd51, 12;
cvta.to.global.u64 %rd53, %rd21;
add.s64 %rd70, %rd53, %rd52;
cvta.to.global.u64 %rd54, %rd17;
add.s64 %rd69, %rd54, %rd52;
add.s64 %rd68, %rd53, %rd49;
add.s64 %rd67, %rd54, %rd49;
mul.lo.s32 %r77, %r300, %r34;
cvt.u64.u32 %rd55, %r77;
mul.lo.s32 %r78, %r65, %r40;
cvt.u64.u32 %rd56, %r78;
add.s64 %rd8, %rd55, %rd56;
mov.f32 %f119, 0f00000000;
cvta.to.global.u64 %rd57, %rd18;
mov.f32 %f117, %f119;

$L__BB77_10:
cvt.u64.u32 %rd58, %r301;
add.s64 %rd59, %rd8, %rd58;
mul.lo.s64 %rd60, %rd59, 36;
add.s64 %rd61, %rd57, %rd60;
add.s64 %rd63, %rd61, %rd50;
ld.global.nc.u32 %r101, [%rd67+4];
shl.b32 %r102, %r13, 3;
shr.s32 %r103, %r101, %r102;
ld.global.nc.u32 %r15, [%rd63+4];
ld.global.nc.u32 %r16, [%rd63+20];
ld.global.nc.u32 %r104, [%rd69];
or.b32 %r105, %r102, 4;
shr.s32 %r106, %r101, %r105;
ld.global.nc.u32 %r17, [%rd63+8];
ld.global.nc.u32 %r18, [%rd63+24];
ld.global.nc.u32 %r107, [%rd69+-4];
and.b32 %r108, %r107, 252645135;
shl.b32 %r109, %r103, 4;
and.b32 %r110, %r109, 16;
or.b32 %r111, %r110, %r108;
shl.b32 %r112, %r103, 11;
and.b32 %r113, %r112, 4096;
or.b32 %r114, %r111, %r113;
shl.b32 %r115, %r103, 18;
and.b32 %r116, %r115, 1048576;
or.b32 %r117, %r114, %r116;
shl.b32 %r118, %r103, 25;
and.b32 %r119, %r118, 268435456;
or.b32 %r80, %r117, %r119;
mov.u32 %r82, 0;

dp4a.s32.s32 %r79, %r80, %r15, %r82;

shr.u32 %r120, %r107, 4;
and.b32 %r121, %r120, 252645135;
shr.u32 %r122, %r103, 12;
and.b32 %r123, %r122, 16;
or.b32 %r124, %r123, %r121;
shr.u32 %r125, %r103, 5;
and.b32 %r126, %r125, 4096;
or.b32 %r127, %r124, %r126;
shl.b32 %r128, %r103, 2;
and.b32 %r129, %r128, 1048576;
or.b32 %r130, %r127, %r129;
shl.b32 %r131, %r103, 9;
and.b32 %r132, %r131, 268435456;
or.b32 %r84, %r130, %r132;

dp4a.s32.s32 %r83, %r84, %r16, %r79;

and.b32 %r133, %r104, 252645135;
shl.b32 %r134, %r106, 4;
and.b32 %r135, %r134, 16;
or.b32 %r136, %r135, %r133;
shl.b32 %r137, %r106, 11;
and.b32 %r138, %r137, 4096;
or.b32 %r139, %r136, %r138;
shl.b32 %r140, %r106, 18;
and.b32 %r141, %r140, 1048576;
or.b32 %r142, %r139, %r141;
shl.b32 %r143, %r106, 25;
and.b32 %r144, %r143, 268435456;
or.b32 %r88, %r142, %r144;

dp4a.s32.s32 %r87, %r88, %r17, %r83;

shr.u32 %r145, %r104, 4;
and.b32 %r146, %r145, 252645135;
shr.u32 %r147, %r106, 12;
and.b32 %r148, %r147, 16;
or.b32 %r149, %r148, %r146;
shr.u32 %r150, %r106, 5;
and.b32 %r151, %r150, 4096;
or.b32 %r152, %r149, %r151;
shl.b32 %r153, %r106, 2;
and.b32 %r154, %r153, 1048576;
or.b32 %r155, %r152, %r154;
shl.b32 %r156, %r106, 9;
and.b32 %r157, %r156, 268435456;
or.b32 %r92, %r155, %r157;

dp4a.s32.s32 %r91, %r92, %r18, %r87;

ld.global.nc.u32 %r96, [%rd67];
ld.global.nc.u32 %r19, [%rd61];

{mul.f16x2 %r95,%r96,%r19;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r95;
cvt.f32.f16 %f40, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r95;
cvt.f32.f16 %f41, high;}


cvt.rn.f32.s32 %f42, %r91;
mov.f32 %f43, 0f40000000;
div.approx.ftz.f32 %f44, %f41, %f43;
fma.rn.ftz.f32 %f45, %f40, %f42, %f44;
add.ftz.f32 %f119, %f119, %f45;
setp.eq.s64 %p11, %rd21, 0;
@%p11 bra $L__BB77_12;

ld.global.u32 %r179, [%rd68+4];
shr.s32 %r181, %r179, %r102;
ld.global.u32 %r182, [%rd70];
shr.s32 %r184, %r179, %r105;
ld.global.u32 %r185, [%rd70+-4];
and.b32 %r186, %r185, 252645135;
shl.b32 %r187, %r181, 4;
and.b32 %r188, %r187, 16;
or.b32 %r189, %r188, %r186;
shl.b32 %r190, %r181, 11;
and.b32 %r191, %r190, 4096;
or.b32 %r192, %r189, %r191;
shl.b32 %r193, %r181, 18;
and.b32 %r194, %r193, 1048576;
or.b32 %r195, %r192, %r194;
shl.b32 %r196, %r181, 25;
and.b32 %r197, %r196, 268435456;
or.b32 %r159, %r195, %r197;

dp4a.s32.s32 %r158, %r159, %r15, %r82;

shr.u32 %r198, %r185, 4;
and.b32 %r199, %r198, 252645135;
shr.u32 %r200, %r181, 12;
and.b32 %r201, %r200, 16;
or.b32 %r202, %r201, %r199;
shr.u32 %r203, %r181, 5;
and.b32 %r204, %r203, 4096;
or.b32 %r205, %r202, %r204;
shl.b32 %r206, %r181, 2;
and.b32 %r207, %r206, 1048576;
or.b32 %r208, %r205, %r207;
shl.b32 %r209, %r181, 9;
and.b32 %r210, %r209, 268435456;
or.b32 %r163, %r208, %r210;

dp4a.s32.s32 %r162, %r163, %r16, %r158;

and.b32 %r211, %r182, 252645135;
shl.b32 %r212, %r184, 4;
and.b32 %r213, %r212, 16;
or.b32 %r214, %r213, %r211;
shl.b32 %r215, %r184, 11;
and.b32 %r216, %r215, 4096;
or.b32 %r217, %r214, %r216;
shl.b32 %r218, %r184, 18;
and.b32 %r219, %r218, 1048576;
or.b32 %r220, %r217, %r219;
shl.b32 %r221, %r184, 25;
and.b32 %r222, %r221, 268435456;
or.b32 %r167, %r220, %r222;

dp4a.s32.s32 %r166, %r167, %r17, %r162;

shr.u32 %r223, %r182, 4;
and.b32 %r224, %r223, 252645135;
shr.u32 %r225, %r184, 12;
and.b32 %r226, %r225, 16;
or.b32 %r227, %r226, %r224;
shr.u32 %r228, %r184, 5;
and.b32 %r229, %r228, 4096;
or.b32 %r230, %r227, %r229;
shl.b32 %r231, %r184, 2;
and.b32 %r232, %r231, 1048576;
or.b32 %r233, %r230, %r232;
shl.b32 %r234, %r184, 9;
and.b32 %r235, %r234, 268435456;
or.b32 %r171, %r233, %r235;

dp4a.s32.s32 %r170, %r171, %r18, %r166;

ld.global.u32 %r175, [%rd68];

{mul.f16x2 %r174,%r175,%r19;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r174;
cvt.f32.f16 %f46, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r174;
cvt.f32.f16 %f47, high;}


cvt.rn.f32.s32 %f48, %r170;
div.approx.ftz.f32 %f50, %f47, %f43;
fma.rn.ftz.f32 %f51, %f46, %f48, %f50;
add.ftz.f32 %f117, %f117, %f51;

$L__BB77_12:
add.s64 %rd70, %rd70, 1536;
add.s64 %rd69, %rd69, 1536;
add.s64 %rd68, %rd68, 1536;
add.s64 %rd67, %rd67, 1536;
add.s32 %r301, %r301, 64;
setp.lt.s32 %p12, %r301, %r62;
@%p12 bra $L__BB77_10;

$L__BB77_13:
setp.eq.s32 %p13, %r1, 0;
@%p13 bra $L__BB77_16;

setp.eq.s64 %p14, %rd21, 0;
add.s32 %r21, %r1, -1;
shl.b32 %r240, %r1, 7;
mov.u32 %r241, _ZZ13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r242, %r241, %r240;
shl.b32 %r243, %r2, 2;
add.s32 %r244, %r242, %r243;
st.shared.f32 [%r244+-128], %f119;
@%p14 bra $L__BB77_16;

shl.b32 %r246, %r21, 7;
mov.u32 %r247, _ZZ13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r248, %r247, %r246;
add.s32 %r250, %r248, %r243;
st.shared.f32 [%r250], %f117;

$L__BB77_16:
setp.ne.s32 %p15, %r1, 0;
bar.sync 0;
@%p15 bra $L__BB77_35;

setp.ne.s64 %p16, %rd21, 0;
shl.b32 %r253, %r2, 2;
mov.u32 %r254, _ZZ13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r22, %r254, %r253;
ld.shared.f32 %f52, [%r22];
add.ftz.f32 %f13, %f119, %f52;
@%p16 bra $L__BB77_19;
bra.uni $L__BB77_18;

$L__BB77_19:
mov.u32 %r257, _ZZ13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r258, %r257, %r253;
ld.shared.f32 %f56, [%r258];
add.ftz.f32 %f57, %f117, %f56;
ld.shared.f32 %f58, [%r22+128];
add.ftz.f32 %f59, %f13, %f58;
ld.shared.f32 %f60, [%r258+128];
add.ftz.f32 %f61, %f57, %f60;
ld.shared.f32 %f62, [%r22+256];
add.ftz.f32 %f121, %f59, %f62;
ld.shared.f32 %f63, [%r258+256];
add.ftz.f32 %f117, %f61, %f63;
bra.uni $L__BB77_20;

$L__BB77_18:
ld.shared.f32 %f53, [%r22+128];
add.ftz.f32 %f54, %f13, %f53;
ld.shared.f32 %f55, [%r22+256];
add.ftz.f32 %f121, %f54, %f55;

$L__BB77_20:
mov.b32 %r259, %f121;
mov.u32 %r260, 31;
mov.u32 %r261, 16;
mov.u32 %r262, -1;
shfl.sync.bfly.b32 %r263|%p17, %r259, %r261, %r260, %r262;
mov.b32 %f64, %r263;
add.ftz.f32 %f65, %f121, %f64;
mov.b32 %r264, %f65;
mov.u32 %r265, 8;
shfl.sync.bfly.b32 %r266|%p18, %r264, %r265, %r260, %r262;
mov.b32 %f66, %r266;
add.ftz.f32 %f67, %f65, %f66;
mov.b32 %r267, %f67;
mov.u32 %r268, 4;
shfl.sync.bfly.b32 %r269|%p19, %r267, %r268, %r260, %r262;
mov.b32 %f68, %r269;
add.ftz.f32 %f69, %f67, %f68;
mov.b32 %r270, %f69;
mov.u32 %r271, 2;
shfl.sync.bfly.b32 %r272|%p20, %r270, %r271, %r260, %r262;
mov.b32 %f70, %r272;
add.ftz.f32 %f19, %f69, %f70;
mov.b32 %r273, %f19;
mov.u32 %r274, 1;
shfl.sync.bfly.b32 %r23|%p1, %r273, %r274, %r260, %r262;
setp.eq.s64 %p21, %rd21, 0;
@%p21 bra $L__BB77_22;

mov.b32 %r275, %f117;
shfl.sync.bfly.b32 %r279|%p22, %r275, %r261, %r260, %r262;
mov.b32 %f71, %r279;
add.ftz.f32 %f72, %f117, %f71;
mov.b32 %r280, %f72;
shfl.sync.bfly.b32 %r282|%p23, %r280, %r265, %r260, %r262;
mov.b32 %f73, %r282;
add.ftz.f32 %f74, %f72, %f73;
mov.b32 %r283, %f74;
shfl.sync.bfly.b32 %r285|%p24, %r283, %r268, %r260, %r262;
mov.b32 %f75, %r285;
add.ftz.f32 %f76, %f74, %f75;
mov.b32 %r286, %f76;
shfl.sync.bfly.b32 %r288|%p25, %r286, %r271, %r260, %r262;
mov.b32 %f77, %r288;
add.ftz.f32 %f78, %f76, %f77;
mov.b32 %r289, %f78;
shfl.sync.bfly.b32 %r291|%p26, %r289, %r274, %r260, %r262;
mov.b32 %f79, %r291;
add.ftz.f32 %f117, %f78, %f79;

$L__BB77_22:
setp.ne.s32 %p27, %r2, 0;
mov.b32 %f80, %r23;
add.ftz.f32 %f22, %f19, %f80;
@%p27 bra $L__BB77_35;

add.ftz.f32 %f123, %f114, %f22;
@%p21 bra $L__BB77_34;

add.ftz.f32 %f24, %f113, %f117;
setp.gt.s32 %p29, %r24, 2;
@%p29 bra $L__BB77_28;

setp.eq.s32 %p32, %r24, 1;
@%p32 bra $L__BB77_32;

setp.eq.s32 %p33, %r24, 2;
@%p33 bra $L__BB77_27;
bra.uni $L__BB77_33;

$L__BB77_27:
mul.ftz.f32 %f100, %f24, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f101, %f100;
add.ftz.f32 %f102, %f101, 0f3F800000;
div.approx.ftz.f32 %f103, %f24, %f102;
mul.ftz.f32 %f123, %f123, %f103;
bra.uni $L__BB77_34;

$L__BB77_28:
setp.eq.s32 %p30, %r24, 3;
@%p30 bra $L__BB77_31;

setp.ne.s32 %p31, %r24, 6;
@%p31 bra $L__BB77_33;

min.ftz.f32 %f81, %f24, %f31;
neg.ftz.f32 %f82, %f31;
min.ftz.f32 %f83, %f123, %f31;
max.ftz.f32 %f84, %f83, %f82;
mul.ftz.f32 %f85, %f81, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f86, %f85;
add.ftz.f32 %f87, %f86, 0f3F800000;
div.approx.ftz.f32 %f88, %f81, %f87;
mul.ftz.f32 %f123, %f84, %f88;
bra.uni $L__BB77_34;

$L__BB77_32:
mul.ftz.f32 %f104, %f24, 0f3F000000;
mul.ftz.f32 %f105, %f24, 0f3D372713;
fma.rn.ftz.f32 %f106, %f24, %f105, 0f3F800000;
mul.ftz.f32 %f107, %f24, 0f3F4C422A;
mul.ftz.f32 %f108, %f107, %f106;
tanh.approx.f32 %f109, %f108;
add.ftz.f32 %f110, %f109, 0f3F800000;
mul.ftz.f32 %f111, %f104, %f110;
mul.ftz.f32 %f123, %f123, %f111;
bra.uni $L__BB77_34;

$L__BB77_31:
mov.f32 %f89, 0f40E00000;
min.ftz.f32 %f90, %f24, %f89;
min.ftz.f32 %f91, %f123, %f89;
mov.f32 %f92, 0fC0E00000;
max.ftz.f32 %f93, %f91, %f92;
mul.ftz.f32 %f94, %f90, 0f3FD9DB23;
mul.ftz.f32 %f95, %f94, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f96, %f95;
add.ftz.f32 %f97, %f96, 0f3F800000;
div.approx.ftz.f32 %f98, %f90, %f97;
add.ftz.f32 %f99, %f93, 0f3F800000;
mul.ftz.f32 %f123, %f99, %f98;
bra.uni $L__BB77_34;

$L__BB77_33:
mul.ftz.f32 %f123, %f123, %f24;

$L__BB77_34:
cvta.to.global.u64 %rd64, %rd25;
mov.u32 %r294, %ctaid.z;
mov.u32 %r295, %ctaid.y;
mov.u32 %r296, %ctaid.x;
mad.lo.s32 %r297, %r295, %r35, %r296;
mad.lo.s32 %r298, %r294, %r41, %r297;
mul.wide.u32 %rd65, %r298, 4;
add.s64 %rd66, %rd64, %rd65;
st.global.f32 [%rd66], %f123;

$L__BB77_35:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<12>;
.reg .f32 %f<30>;
.reg .b32 %r<165>;
.reg .b64 %rd<31>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[384];

ld.param.u64 %rd8, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd9, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd10, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd11, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r19, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r18, [_Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
mov.u32 %r1, %tid.y;
shl.b32 %r35, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r35, %r2;
shr.u32 %r4, %r34, 5;
mov.u32 %r163, %ctaid.y;
setp.eq.s64 %p1, %rd10, 0;
@%p1 bra $L__BB78_2;

cvta.to.global.u64 %rd12, %rd10;
mul.wide.u32 %rd13, %r163, 4;
add.s64 %rd14, %rd12, %rd13;
ld.global.nc.u32 %r162, [%rd14];
mul.hi.u32 %r36, %r163, %r18;
add.s32 %r37, %r36, %r163;
shr.u32 %r38, %r37, %r19;
mul.lo.s32 %r39, %r38, %r20;
sub.s32 %r163, %r163, %r39;
bra.uni $L__BB78_3;

$L__BB78_2:
mul.hi.u32 %r40, %r163, %r22;
add.s32 %r41, %r40, %r163;
shr.u32 %r162, %r41, %r23;

$L__BB78_3:
mov.u32 %r11, %ctaid.z;
shr.u32 %r42, %r3, 31;
add.s32 %r43, %r3, %r42;
shr.s32 %r164, %r43, 1;
setp.ge.s32 %p2, %r164, %r4;
mov.f32 %f29, 0f00000000;
@%p2 bra $L__BB78_6;

and.b32 %r46, %r43, -2;
sub.s32 %r47, %r3, %r46;
shl.b32 %r48, %r47, 1;
cvt.s64.s32 %rd1, %r48;
shl.b32 %r13, %r47, 3;
or.b32 %r14, %r13, 4;
mul.hi.u32 %r49, %r11, %r28;
add.s32 %r50, %r49, %r11;
shr.u32 %r51, %r50, %r29;
mad.lo.s32 %r52, %r51, %r31, %r164;
mad.lo.s32 %r53, %r162, %r25, %r52;
mov.u32 %r54, %ctaid.x;
mad.lo.s32 %r55, %r54, %r21, %r53;
mul.wide.s32 %rd15, %r55, 24;
cvta.to.global.u64 %rd16, %rd8;
add.s64 %rd30, %rd16, %rd15;
mul.wide.s32 %rd17, %r48, 4;
add.s64 %rd3, %rd17, 8;
cvta.to.global.u64 %rd4, %rd9;
mul.lo.s32 %r56, %r163, %r26;
cvt.u64.u32 %rd18, %r56;
mul.lo.s32 %r57, %r11, %r32;
cvt.u64.u32 %rd19, %r57;
add.s64 %rd5, %rd18, %rd19;
mov.f32 %f29, 0f00000000;
shl.b64 %rd24, %rd1, 2;

$L__BB78_5:
cvt.u64.u32 %rd20, %r164;
add.s64 %rd21, %rd5, %rd20;
mul.lo.s64 %rd22, %rd21, 36;
add.s64 %rd23, %rd4, %rd22;
add.s64 %rd25, %rd23, %rd24;
add.s64 %rd26, %rd30, %rd3;
ld.global.nc.u32 %r79, [%rd30+4];
shr.s32 %r80, %r79, %r13;
ld.global.nc.u32 %r60, [%rd25+4];
ld.global.nc.u32 %r64, [%rd25+20];
ld.global.nc.u32 %r81, [%rd26+4];
shr.s32 %r82, %r79, %r14;
ld.global.nc.u32 %r68, [%rd25+8];
ld.global.nc.u32 %r72, [%rd25+24];
ld.global.nc.u32 %r83, [%rd26];
and.b32 %r84, %r83, 252645135;
shl.b32 %r85, %r80, 4;
and.b32 %r86, %r85, 16;
or.b32 %r87, %r86, %r84;
shl.b32 %r88, %r80, 11;
and.b32 %r89, %r88, 4096;
or.b32 %r90, %r87, %r89;
shl.b32 %r91, %r80, 18;
and.b32 %r92, %r91, 1048576;
or.b32 %r93, %r90, %r92;
shl.b32 %r94, %r80, 25;
and.b32 %r95, %r94, 268435456;
or.b32 %r59, %r93, %r95;
mov.u32 %r61, 0;

dp4a.s32.s32 %r58, %r59, %r60, %r61;

shr.u32 %r96, %r83, 4;
and.b32 %r97, %r96, 252645135;
shr.u32 %r98, %r80, 12;
and.b32 %r99, %r98, 16;
or.b32 %r100, %r99, %r97;
shr.u32 %r101, %r80, 5;
and.b32 %r102, %r101, 4096;
or.b32 %r103, %r100, %r102;
shl.b32 %r104, %r80, 2;
and.b32 %r105, %r104, 1048576;
or.b32 %r106, %r103, %r105;
shl.b32 %r107, %r80, 9;
and.b32 %r108, %r107, 268435456;
or.b32 %r63, %r106, %r108;

dp4a.s32.s32 %r62, %r63, %r64, %r58;

and.b32 %r109, %r81, 252645135;
shl.b32 %r110, %r82, 4;
and.b32 %r111, %r110, 16;
or.b32 %r112, %r111, %r109;
shl.b32 %r113, %r82, 11;
and.b32 %r114, %r113, 4096;
or.b32 %r115, %r112, %r114;
shl.b32 %r116, %r82, 18;
and.b32 %r117, %r116, 1048576;
or.b32 %r118, %r115, %r117;
shl.b32 %r119, %r82, 25;
and.b32 %r120, %r119, 268435456;
or.b32 %r67, %r118, %r120;

dp4a.s32.s32 %r66, %r67, %r68, %r62;

shr.u32 %r121, %r81, 4;
and.b32 %r122, %r121, 252645135;
shr.u32 %r123, %r82, 12;
and.b32 %r124, %r123, 16;
or.b32 %r125, %r124, %r122;
shr.u32 %r126, %r82, 5;
and.b32 %r127, %r126, 4096;
or.b32 %r128, %r125, %r127;
shl.b32 %r129, %r82, 2;
and.b32 %r130, %r129, 1048576;
or.b32 %r131, %r128, %r130;
shl.b32 %r132, %r82, 9;
and.b32 %r133, %r132, 268435456;
or.b32 %r71, %r131, %r133;

dp4a.s32.s32 %r70, %r71, %r72, %r66;

ld.global.nc.u32 %r75, [%rd30];
ld.global.nc.u32 %r76, [%rd23];

{mul.f16x2 %r74,%r75,%r76;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r74;
cvt.f32.f16 %f7, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r74;
cvt.f32.f16 %f8, high;}


cvt.rn.f32.s32 %f9, %r70;
mov.f32 %f10, 0f40000000;
div.approx.ftz.f32 %f11, %f8, %f10;
fma.rn.ftz.f32 %f12, %f7, %f9, %f11;
add.ftz.f32 %f29, %f29, %f12;
add.s64 %rd30, %rd30, 1536;
add.s32 %r164, %r164, 64;
setp.lt.s32 %p3, %r164, %r4;
@%p3 bra $L__BB78_5;

$L__BB78_6:
setp.eq.s32 %p4, %r1, 0;
@%p4 bra $L__BB78_8;

shl.b32 %r134, %r1, 7;
mov.u32 %r135, _ZZ13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r136, %r135, %r134;
shl.b32 %r137, %r2, 2;
add.s32 %r138, %r136, %r137;
st.shared.f32 [%r138+-128], %f29;

$L__BB78_8:
setp.ne.s32 %p5, %r1, 0;
bar.sync 0;
@%p5 bra $L__BB78_11;

mov.u32 %r139, %ctaid.y;
mov.u32 %r140, %ctaid.x;
mad.lo.s32 %r141, %r139, %r27, %r140;
mad.lo.s32 %r17, %r11, %r33, %r141;
shl.b32 %r142, %r2, 2;
mov.u32 %r143, _ZZ13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r144, %r143, %r142;
mov.u32 %r145, 2;
ld.shared.f32 %f13, [%r144];
add.ftz.f32 %f14, %f29, %f13;
ld.shared.f32 %f15, [%r144+128];
add.ftz.f32 %f16, %f14, %f15;
ld.shared.f32 %f17, [%r144+256];
add.ftz.f32 %f18, %f16, %f17;
mov.b32 %r146, %f18;
mov.u32 %r147, 31;
mov.u32 %r148, 16;
mov.u32 %r149, -1;
shfl.sync.bfly.b32 %r150|%p6, %r146, %r148, %r147, %r149;
mov.b32 %f19, %r150;
add.ftz.f32 %f20, %f18, %f19;
mov.b32 %r151, %f20;
mov.u32 %r152, 8;
shfl.sync.bfly.b32 %r153|%p7, %r151, %r152, %r147, %r149;
mov.b32 %f21, %r153;
add.ftz.f32 %f22, %f20, %f21;
mov.b32 %r154, %f22;
mov.u32 %r155, 4;
shfl.sync.bfly.b32 %r156|%p8, %r154, %r155, %r147, %r149;
mov.b32 %f23, %r156;
add.ftz.f32 %f24, %f22, %f23;
mov.b32 %r157, %f24;
shfl.sync.bfly.b32 %r158|%p9, %r157, %r145, %r147, %r149;
mov.b32 %f25, %r158;
add.ftz.f32 %f26, %f24, %f25;
mov.b32 %r159, %f26;
mov.u32 %r160, 1;
shfl.sync.bfly.b32 %r161|%p10, %r159, %r160, %r147, %r149;
mov.b32 %f27, %r161;
add.ftz.f32 %f4, %f26, %f27;
setp.ne.s32 %p11, %r2, 0;
@%p11 bra $L__BB78_11;

cvta.to.global.u64 %rd27, %rd11;
mul.wide.u32 %rd28, %r17, 4;
add.s64 %rd29, %rd27, %rd28;
st.global.f32 [%rd29], %f4;

$L__BB78_11:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type7ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type7ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type7ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type7ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type7ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type7ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type7ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type7ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type7ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<35>;
.reg .f32 %f<114>;
.reg .b32 %r<336>;
.reg .b64 %rd<48>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type7ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

ld.param.u64 %rd14, [_Z13mul_mat_vec_qIL9ggml_type7ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type7ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type7ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type7ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r17, [_Z13mul_mat_vec_qIL9ggml_type7ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r18, [_Z13mul_mat_vec_qIL9ggml_type7ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r19, [_Z13mul_mat_vec_qIL9ggml_type7ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type7ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type7ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type7ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type7ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type7ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type7ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type7ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type7ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type7ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type7ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
cvta.to.global.u64 %rd1, %rd16;
mov.u32 %r1, %tid.y;
shl.b32 %r33, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r33, %r2;
mov.u32 %r34, %ctaid.x;
shl.b32 %r4, %r34, 1;
mov.u32 %r5, %ctaid.y;
shr.u32 %r35, %r3, 31;
add.s32 %r36, %r3, %r35;
shr.s32 %r335, %r36, 1;
shr.u32 %r7, %r32, 5;
setp.ge.s32 %p3, %r335, %r7;
mov.f32 %f110, 0f00000000;
mov.f32 %f111, %f110;
mov.f32 %f112, %f110;
mov.f32 %f113, %f110;
@%p3 bra $L__BB79_3;

mov.u32 %r37, %ctaid.z;
mul.lo.s32 %r38, %r5, %r24;
cvt.u64.u32 %rd17, %r38;
mul.lo.s32 %r39, %r37, %r30;
cvt.u64.u32 %rd18, %r39;
add.s64 %rd2, %rd18, %rd17;
mul.hi.u32 %r40, %r37, %r26;
add.s32 %r41, %r40, %r37;
shr.u32 %r42, %r41, %r27;
mul.lo.s32 %r43, %r42, %r29;
mul.hi.u32 %r44, %r5, %r20;
add.s32 %r45, %r44, %r5;
shr.u32 %r46, %r45, %r21;
mul.lo.s32 %r47, %r46, %r23;
cvta.to.global.u64 %rd19, %rd14;
and.b32 %r50, %r36, -2;
sub.s32 %r51, %r3, %r50;
shl.b32 %r52, %r51, 1;
cvt.s64.s32 %rd3, %r52;
shl.b32 %r8, %r51, 3;
or.b32 %r9, %r8, 4;
add.s32 %r334, %r335, %r18;
add.s32 %r53, %r335, %r43;
add.s32 %r54, %r53, %r47;
add.s32 %r55, %r4, 1;
mad.lo.s32 %r56, %r17, %r55, %r54;
mul.wide.s32 %rd20, %r56, 24;
add.s64 %rd21, %rd19, %rd20;
add.s64 %rd47, %rd21, 8;
mul.wide.s32 %rd5, %r52, 4;
mad.lo.s32 %r57, %r4, %r17, %r335;
add.s32 %r58, %r57, %r43;
add.s32 %r59, %r58, %r47;
mul.wide.s32 %rd22, %r59, 24;
add.s64 %rd23, %rd19, %rd22;
add.s64 %rd46, %rd23, 8;
cvta.to.global.u64 %rd7, %rd15;
mov.f32 %f110, 0f00000000;
shl.b64 %rd28, %rd3, 2;

$L__BB79_2:
cvt.u64.u32 %rd24, %r335;
add.s64 %rd25, %rd2, %rd24;
mul.lo.s64 %rd26, %rd25, 36;
add.s64 %rd27, %rd7, %rd26;
add.s64 %rd29, %rd27, %rd28;
add.s64 %rd30, %rd46, %rd5;
ld.global.nc.u32 %r144, [%rd46+-4];
shr.s32 %r145, %r144, %r8;
ld.global.nc.u32 %r62, [%rd29+4];
ld.global.nc.u32 %r66, [%rd29+20];
ld.global.nc.u32 %r146, [%rd30+4];
shr.s32 %r147, %r144, %r9;
ld.global.nc.u32 %r70, [%rd29+8];
ld.global.nc.u32 %r74, [%rd29+24];
ld.global.nc.u32 %r148, [%rd30];
and.b32 %r149, %r148, 252645135;
shl.b32 %r150, %r145, 4;
and.b32 %r151, %r150, 16;
or.b32 %r152, %r151, %r149;
shl.b32 %r153, %r145, 11;
and.b32 %r154, %r153, 4096;
or.b32 %r155, %r152, %r154;
shl.b32 %r156, %r145, 18;
and.b32 %r157, %r156, 1048576;
or.b32 %r158, %r155, %r157;
shl.b32 %r159, %r145, 25;
and.b32 %r160, %r159, 268435456;
or.b32 %r103, %r158, %r160;
mov.u32 %r126, 0;

dp4a.s32.s32 %r60, %r103, %r62, %r126;

shr.u32 %r161, %r148, 4;
and.b32 %r162, %r161, 252645135;
shr.u32 %r163, %r145, 12;
and.b32 %r164, %r163, 16;
or.b32 %r165, %r164, %r162;
shr.u32 %r166, %r145, 5;
and.b32 %r167, %r166, 4096;
or.b32 %r168, %r165, %r167;
shl.b32 %r169, %r145, 2;
and.b32 %r170, %r169, 1048576;
or.b32 %r171, %r168, %r170;
shl.b32 %r172, %r145, 9;
and.b32 %r173, %r172, 268435456;
or.b32 %r107, %r171, %r173;

dp4a.s32.s32 %r64, %r107, %r66, %r60;

and.b32 %r174, %r146, 252645135;
shl.b32 %r175, %r147, 4;
and.b32 %r176, %r175, 16;
or.b32 %r177, %r176, %r174;
shl.b32 %r178, %r147, 11;
and.b32 %r179, %r178, 4096;
or.b32 %r180, %r177, %r179;
shl.b32 %r181, %r147, 18;
and.b32 %r182, %r181, 1048576;
or.b32 %r183, %r180, %r182;
shl.b32 %r184, %r147, 25;
and.b32 %r185, %r184, 268435456;
or.b32 %r111, %r183, %r185;

dp4a.s32.s32 %r68, %r111, %r70, %r64;

shr.u32 %r186, %r146, 4;
and.b32 %r187, %r186, 252645135;
shr.u32 %r188, %r147, 12;
and.b32 %r189, %r188, 16;
or.b32 %r190, %r189, %r187;
shr.u32 %r191, %r147, 5;
and.b32 %r192, %r191, 4096;
or.b32 %r193, %r190, %r192;
shl.b32 %r194, %r147, 2;
and.b32 %r195, %r194, 1048576;
or.b32 %r196, %r193, %r195;
shl.b32 %r197, %r147, 9;
and.b32 %r198, %r197, 268435456;
or.b32 %r115, %r196, %r198;

dp4a.s32.s32 %r72, %r115, %r74, %r68;

ld.global.nc.u32 %r77, [%rd46+-8];
ld.global.nc.u32 %r78, [%rd27];

{mul.f16x2 %r76,%r77,%r78;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r76;
cvt.f32.f16 %f25, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r76;
cvt.f32.f16 %f26, high;}


cvt.rn.f32.s32 %f33, %r72;
mov.f32 %f34, 0f40000000;
div.approx.ftz.f32 %f35, %f26, %f34;
fma.rn.ftz.f32 %f36, %f25, %f33, %f35;
add.ftz.f32 %f113, %f36, %f113;
add.s64 %rd31, %rd47, %rd5;
ld.global.nc.u32 %r199, [%rd47+-4];
shr.s32 %r200, %r199, %r8;
ld.global.nc.u32 %r201, [%rd31+4];
shr.s32 %r202, %r199, %r9;
ld.global.nc.u32 %r203, [%rd31];
and.b32 %r204, %r203, 252645135;
shl.b32 %r205, %r200, 4;
and.b32 %r206, %r205, 16;
or.b32 %r207, %r206, %r204;
shl.b32 %r208, %r200, 11;
and.b32 %r209, %r208, 4096;
or.b32 %r210, %r207, %r209;
shl.b32 %r211, %r200, 18;
and.b32 %r212, %r211, 1048576;
or.b32 %r213, %r210, %r212;
shl.b32 %r214, %r200, 25;
and.b32 %r215, %r214, 268435456;
or.b32 %r124, %r213, %r215;

dp4a.s32.s32 %r81, %r124, %r62, %r126;

shr.u32 %r216, %r203, 4;
and.b32 %r217, %r216, 252645135;
shr.u32 %r218, %r200, 12;
and.b32 %r219, %r218, 16;
or.b32 %r220, %r219, %r217;
shr.u32 %r221, %r200, 5;
and.b32 %r222, %r221, 4096;
or.b32 %r223, %r220, %r222;
shl.b32 %r224, %r200, 2;
and.b32 %r225, %r224, 1048576;
or.b32 %r226, %r223, %r225;
shl.b32 %r227, %r200, 9;
and.b32 %r228, %r227, 268435456;
or.b32 %r128, %r226, %r228;

dp4a.s32.s32 %r85, %r128, %r66, %r81;

and.b32 %r229, %r201, 252645135;
shl.b32 %r230, %r202, 4;
and.b32 %r231, %r230, 16;
or.b32 %r232, %r231, %r229;
shl.b32 %r233, %r202, 11;
and.b32 %r234, %r233, 4096;
or.b32 %r235, %r232, %r234;
shl.b32 %r236, %r202, 18;
and.b32 %r237, %r236, 1048576;
or.b32 %r238, %r235, %r237;
shl.b32 %r239, %r202, 25;
and.b32 %r240, %r239, 268435456;
or.b32 %r132, %r238, %r240;

dp4a.s32.s32 %r89, %r132, %r70, %r85;

shr.u32 %r241, %r201, 4;
and.b32 %r242, %r241, 252645135;
shr.u32 %r243, %r202, 12;
and.b32 %r244, %r243, 16;
or.b32 %r245, %r244, %r242;
shr.u32 %r246, %r202, 5;
and.b32 %r247, %r246, 4096;
or.b32 %r248, %r245, %r247;
shl.b32 %r249, %r202, 2;
and.b32 %r250, %r249, 1048576;
or.b32 %r251, %r248, %r250;
shl.b32 %r252, %r202, 9;
and.b32 %r253, %r252, 268435456;
or.b32 %r136, %r251, %r253;

dp4a.s32.s32 %r93, %r136, %r74, %r89;

ld.global.nc.u32 %r98, [%rd47+-8];

{mul.f16x2 %r97,%r98,%r78;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r97;
cvt.f32.f16 %f27, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r97;
cvt.f32.f16 %f28, high;}


cvt.rn.f32.s32 %f37, %r93;
div.approx.ftz.f32 %f38, %f28, %f34;
fma.rn.ftz.f32 %f39, %f27, %f37, %f38;
add.ftz.f32 %f112, %f39, %f112;
cvt.u64.u32 %rd32, %r334;
add.s64 %rd33, %rd2, %rd32;
mul.lo.s64 %rd34, %rd33, 36;
add.s64 %rd35, %rd7, %rd34;
add.s64 %rd36, %rd35, %rd28;
ld.global.nc.u32 %r104, [%rd36+4];
ld.global.nc.u32 %r108, [%rd36+20];
ld.global.nc.u32 %r112, [%rd36+8];
ld.global.nc.u32 %r116, [%rd36+24];

dp4a.s32.s32 %r102, %r103, %r104, %r126;


dp4a.s32.s32 %r106, %r107, %r108, %r102;


dp4a.s32.s32 %r110, %r111, %r112, %r106;


dp4a.s32.s32 %r114, %r115, %r116, %r110;

ld.global.nc.u32 %r120, [%rd35];

{mul.f16x2 %r118,%r77,%r120;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r118;
cvt.f32.f16 %f29, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r118;
cvt.f32.f16 %f30, high;}


cvt.rn.f32.s32 %f40, %r114;
div.approx.ftz.f32 %f41, %f30, %f34;
fma.rn.ftz.f32 %f42, %f29, %f40, %f41;
add.ftz.f32 %f111, %f42, %f111;

dp4a.s32.s32 %r123, %r124, %r104, %r126;


dp4a.s32.s32 %r127, %r128, %r108, %r123;


dp4a.s32.s32 %r131, %r132, %r112, %r127;


dp4a.s32.s32 %r135, %r136, %r116, %r131;


{mul.f16x2 %r139,%r98,%r120;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r139;
cvt.f32.f16 %f31, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r139;
cvt.f32.f16 %f32, high;}


cvt.rn.f32.s32 %f43, %r135;
div.approx.ftz.f32 %f44, %f32, %f34;
fma.rn.ftz.f32 %f45, %f31, %f43, %f44;
add.ftz.f32 %f110, %f45, %f110;
add.s32 %r334, %r334, 64;
add.s64 %rd47, %rd47, 1536;
add.s64 %rd46, %rd46, 1536;
add.s32 %r335, %r335, 64;
setp.lt.s32 %p4, %r335, %r7;
@%p4 bra $L__BB79_2;

$L__BB79_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB79_5;

shl.b32 %r254, %r1, 9;
mov.u32 %r255, _ZZ13mul_mat_vec_qIL9ggml_type7ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r256, %r255, %r254;
shl.b32 %r257, %r2, 2;
add.s32 %r258, %r256, %r257;
st.shared.f32 [%r258+-512], %f113;
st.shared.f32 [%r258+-384], %f112;
st.shared.f32 [%r258+-256], %f111;
st.shared.f32 [%r258+-128], %f110;

$L__BB79_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB79_14;

mov.u32 %r259, %ctaid.z;
mad.lo.s32 %r260, %r5, %r25, %r4;
mad.lo.s32 %r261, %r259, %r31, %r260;
cvt.u64.u32 %rd12, %r261;
shl.b32 %r262, %r2, 2;
mov.u32 %r263, _ZZ13mul_mat_vec_qIL9ggml_type7ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r15, %r263, %r262;
mov.u32 %r264, 2;
ld.shared.f32 %f46, [%r15];
add.ftz.f32 %f47, %f46, %f113;
ld.shared.f32 %f48, [%r15+512];
add.ftz.f32 %f49, %f48, %f47;
ld.shared.f32 %f50, [%r15+1024];
add.ftz.f32 %f51, %f50, %f49;
mov.b32 %r265, %f51;
mov.u32 %r266, 31;
mov.u32 %r267, 16;
mov.u32 %r268, -1;
shfl.sync.bfly.b32 %r269|%p7, %r265, %r267, %r266, %r268;
mov.b32 %f52, %r269;
add.ftz.f32 %f53, %f51, %f52;
mov.b32 %r270, %f53;
mov.u32 %r271, 8;
shfl.sync.bfly.b32 %r272|%p8, %r270, %r271, %r266, %r268;
mov.b32 %f54, %r272;
add.ftz.f32 %f55, %f53, %f54;
mov.b32 %r273, %f55;
mov.u32 %r274, 4;
shfl.sync.bfly.b32 %r275|%p9, %r273, %r274, %r266, %r268;
mov.b32 %f56, %r275;
add.ftz.f32 %f57, %f55, %f56;
mov.b32 %r276, %f57;
shfl.sync.bfly.b32 %r277|%p10, %r276, %r264, %r266, %r268;
mov.b32 %f58, %r277;
add.ftz.f32 %f59, %f57, %f58;
mov.b32 %r278, %f59;
mov.u32 %r279, 1;
shfl.sync.bfly.b32 %r280|%p11, %r278, %r279, %r266, %r268;
mov.b32 %f60, %r280;
add.ftz.f32 %f13, %f59, %f60;
setp.lt.u32 %p12, %r4, %r19;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
mul.wide.u32 %rd37, %r261, 4;
add.s64 %rd13, %rd1, %rd37;
not.pred %p14, %p1;
@%p14 bra $L__BB79_8;

st.global.f32 [%rd13], %f13;

$L__BB79_8:
ld.shared.f32 %f61, [%r15+128];
add.ftz.f32 %f62, %f61, %f112;
ld.shared.f32 %f63, [%r15+640];
add.ftz.f32 %f64, %f63, %f62;
ld.shared.f32 %f65, [%r15+1152];
add.ftz.f32 %f66, %f65, %f64;
mov.b32 %r281, %f66;
shfl.sync.bfly.b32 %r285|%p15, %r281, %r267, %r266, %r268;
mov.b32 %f67, %r285;
add.ftz.f32 %f68, %f66, %f67;
mov.b32 %r286, %f68;
shfl.sync.bfly.b32 %r288|%p16, %r286, %r271, %r266, %r268;
mov.b32 %f69, %r288;
add.ftz.f32 %f70, %f68, %f69;
mov.b32 %r289, %f70;
shfl.sync.bfly.b32 %r291|%p17, %r289, %r274, %r266, %r268;
mov.b32 %f71, %r291;
add.ftz.f32 %f72, %f70, %f71;
mov.b32 %r292, %f72;
shfl.sync.bfly.b32 %r294|%p18, %r292, %r264, %r266, %r268;
mov.b32 %f73, %r294;
add.ftz.f32 %f74, %f72, %f73;
mov.b32 %r295, %f74;
shfl.sync.bfly.b32 %r297|%p19, %r295, %r279, %r266, %r268;
mov.b32 %f75, %r297;
add.ftz.f32 %f14, %f74, %f75;
add.s32 %r298, %r4, 1;
setp.lt.u32 %p20, %r298, %r19;
setp.eq.s32 %p21, %r2, 1;
and.pred %p2, %p21, %p20;
not.pred %p22, %p2;
@%p22 bra $L__BB79_10;

st.global.f32 [%rd13+4], %f14;

$L__BB79_10:
ld.shared.f32 %f76, [%r15+256];
add.ftz.f32 %f77, %f76, %f111;
ld.shared.f32 %f78, [%r15+768];
add.ftz.f32 %f79, %f78, %f77;
ld.shared.f32 %f80, [%r15+1280];
add.ftz.f32 %f81, %f80, %f79;
mov.b32 %r299, %f81;
mov.u32 %r300, 31;
mov.u32 %r301, 16;
mov.u32 %r302, -1;
shfl.sync.bfly.b32 %r303|%p23, %r299, %r301, %r300, %r302;
mov.b32 %f82, %r303;
add.ftz.f32 %f83, %f81, %f82;
mov.b32 %r304, %f83;
mov.u32 %r305, 8;
shfl.sync.bfly.b32 %r306|%p24, %r304, %r305, %r300, %r302;
mov.b32 %f84, %r306;
add.ftz.f32 %f85, %f83, %f84;
mov.b32 %r307, %f85;
mov.u32 %r308, 4;
shfl.sync.bfly.b32 %r309|%p25, %r307, %r308, %r300, %r302;
mov.b32 %f86, %r309;
add.ftz.f32 %f87, %f85, %f86;
mov.b32 %r310, %f87;
mov.u32 %r311, 2;
shfl.sync.bfly.b32 %r312|%p26, %r310, %r311, %r300, %r302;
mov.b32 %f88, %r312;
add.ftz.f32 %f89, %f87, %f88;
mov.b32 %r313, %f89;
mov.u32 %r314, 1;
shfl.sync.bfly.b32 %r315|%p27, %r313, %r314, %r300, %r302;
mov.b32 %f90, %r315;
add.ftz.f32 %f15, %f89, %f90;
@%p14 bra $L__BB79_12;

cvt.u64.u32 %rd38, %r19;
add.s64 %rd39, %rd38, %rd12;
shl.b64 %rd40, %rd39, 2;
add.s64 %rd41, %rd1, %rd40;
st.global.f32 [%rd41], %f15;

$L__BB79_12:
ld.shared.f32 %f91, [%r15+384];
add.ftz.f32 %f92, %f91, %f110;
ld.shared.f32 %f93, [%r15+896];
add.ftz.f32 %f94, %f93, %f92;
ld.shared.f32 %f95, [%r15+1408];
add.ftz.f32 %f96, %f95, %f94;
mov.b32 %r316, %f96;
shfl.sync.bfly.b32 %r320|%p29, %r316, %r301, %r300, %r302;
mov.b32 %f97, %r320;
add.ftz.f32 %f98, %f96, %f97;
mov.b32 %r321, %f98;
shfl.sync.bfly.b32 %r323|%p30, %r321, %r305, %r300, %r302;
mov.b32 %f99, %r323;
add.ftz.f32 %f100, %f98, %f99;
mov.b32 %r324, %f100;
shfl.sync.bfly.b32 %r326|%p31, %r324, %r308, %r300, %r302;
mov.b32 %f101, %r326;
add.ftz.f32 %f102, %f100, %f101;
mov.b32 %r327, %f102;
shfl.sync.bfly.b32 %r329|%p32, %r327, %r311, %r300, %r302;
mov.b32 %f103, %r329;
add.ftz.f32 %f104, %f102, %f103;
mov.b32 %r330, %f104;
shfl.sync.bfly.b32 %r332|%p33, %r330, %r314, %r300, %r302;
mov.b32 %f105, %r332;
add.ftz.f32 %f16, %f104, %f105;
@%p22 bra $L__BB79_14;

add.s32 %r333, %r19, 1;
cvt.u64.u32 %rd42, %r333;
add.s64 %rd43, %rd42, %rd12;
shl.b64 %rd44, %rd43, 2;
add.s64 %rd45, %rd1, %rd44;
st.global.f32 [%rd45], %f16;

$L__BB79_14:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type7ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type7ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type7ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type7ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type7ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type7ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type7ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type7ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type7ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<47>;
.reg .f32 %f<170>;
.reg .b32 %r<419>;
.reg .b64 %rd<57>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type7ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[2304];

ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type7ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type7ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type7ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type7ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type7ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type7ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type7ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type7ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type7ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type7ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type7ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type7ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type7ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type7ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type7ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type7ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type7ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
cvta.to.global.u64 %rd1, %rd17;
mov.u32 %r1, %tid.y;
shl.b32 %r36, %r1, 5;
mov.u32 %r37, %tid.x;
add.s32 %r2, %r36, %r37;
mov.u32 %r38, %ctaid.x;
shl.b32 %r3, %r38, 1;
mov.u32 %r4, %ctaid.y;
mov.u32 %r5, %ctaid.z;
shr.u32 %r39, %r2, 31;
add.s32 %r40, %r2, %r39;
shr.s32 %r418, %r40, 1;
shr.u32 %r7, %r35, 5;
setp.ge.s32 %p3, %r418, %r7;
mov.f32 %f164, 0f00000000;
mov.f32 %f165, %f164;
mov.f32 %f166, %f164;
mov.f32 %f167, %f164;
mov.f32 %f168, %f164;
mov.f32 %f169, %f164;
@%p3 bra $L__BB80_3;

mul.lo.s32 %r41, %r4, %r27;
cvt.u64.u32 %rd18, %r41;
mul.lo.s32 %r42, %r5, %r33;
cvt.u64.u32 %rd19, %r42;
add.s64 %rd2, %rd19, %rd18;
mul.hi.u32 %r43, %r5, %r29;
add.s32 %r44, %r43, %r5;
shr.u32 %r45, %r44, %r30;
mul.lo.s32 %r46, %r45, %r32;
mul.hi.u32 %r47, %r4, %r23;
add.s32 %r48, %r47, %r4;
shr.u32 %r49, %r48, %r24;
mul.lo.s32 %r50, %r49, %r26;
cvta.to.global.u64 %rd20, %rd15;
and.b32 %r53, %r40, -2;
sub.s32 %r54, %r2, %r53;
shl.b32 %r55, %r54, 1;
cvt.s64.s32 %rd3, %r55;
shl.b32 %r8, %r54, 3;
or.b32 %r9, %r8, 4;
shl.b32 %r56, %r21, 1;
add.s32 %r417, %r418, %r56;
add.s32 %r416, %r418, %r21;
add.s32 %r57, %r418, %r46;
add.s32 %r58, %r57, %r50;
add.s32 %r59, %r3, 1;
mad.lo.s32 %r60, %r20, %r59, %r58;
mul.wide.s32 %rd21, %r60, 24;
add.s64 %rd22, %rd20, %rd21;
add.s64 %rd56, %rd22, 8;
mul.wide.s32 %rd5, %r55, 4;
mad.lo.s32 %r61, %r3, %r20, %r418;
add.s32 %r62, %r61, %r46;
add.s32 %r63, %r62, %r50;
mul.wide.s32 %rd23, %r63, 24;
add.s64 %rd24, %rd20, %rd23;
add.s64 %rd55, %rd24, 8;
cvta.to.global.u64 %rd7, %rd16;
mov.f32 %f164, 0f00000000;
shl.b64 %rd29, %rd3, 2;

$L__BB80_2:
cvt.u64.u32 %rd25, %r418;
add.s64 %rd26, %rd2, %rd25;
mul.lo.s64 %rd27, %rd26, 36;
add.s64 %rd28, %rd7, %rd27;
add.s64 %rd30, %rd28, %rd29;
add.s64 %rd31, %rd55, %rd5;
ld.global.nc.u32 %r190, [%rd55+-4];
shr.s32 %r191, %r190, %r8;
ld.global.nc.u32 %r66, [%rd30+4];
ld.global.nc.u32 %r70, [%rd30+20];
ld.global.nc.u32 %r192, [%rd31+4];
shr.s32 %r193, %r190, %r9;
ld.global.nc.u32 %r74, [%rd30+8];
ld.global.nc.u32 %r78, [%rd30+24];
ld.global.nc.u32 %r194, [%rd31];
and.b32 %r195, %r194, 252645135;
shl.b32 %r196, %r191, 4;
and.b32 %r197, %r196, 16;
or.b32 %r198, %r197, %r195;
shl.b32 %r199, %r191, 11;
and.b32 %r200, %r199, 4096;
or.b32 %r201, %r198, %r200;
shl.b32 %r202, %r191, 18;
and.b32 %r203, %r202, 1048576;
or.b32 %r204, %r201, %r203;
shl.b32 %r205, %r191, 25;
and.b32 %r206, %r205, 268435456;
or.b32 %r149, %r204, %r206;
mov.u32 %r172, 0;

dp4a.s32.s32 %r64, %r149, %r66, %r172;

shr.u32 %r207, %r194, 4;
and.b32 %r208, %r207, 252645135;
shr.u32 %r209, %r191, 12;
and.b32 %r210, %r209, 16;
or.b32 %r211, %r210, %r208;
shr.u32 %r212, %r191, 5;
and.b32 %r213, %r212, 4096;
or.b32 %r214, %r211, %r213;
shl.b32 %r215, %r191, 2;
and.b32 %r216, %r215, 1048576;
or.b32 %r217, %r214, %r216;
shl.b32 %r218, %r191, 9;
and.b32 %r219, %r218, 268435456;
or.b32 %r153, %r217, %r219;

dp4a.s32.s32 %r68, %r153, %r70, %r64;

and.b32 %r220, %r192, 252645135;
shl.b32 %r221, %r193, 4;
and.b32 %r222, %r221, 16;
or.b32 %r223, %r222, %r220;
shl.b32 %r224, %r193, 11;
and.b32 %r225, %r224, 4096;
or.b32 %r226, %r223, %r225;
shl.b32 %r227, %r193, 18;
and.b32 %r228, %r227, 1048576;
or.b32 %r229, %r226, %r228;
shl.b32 %r230, %r193, 25;
and.b32 %r231, %r230, 268435456;
or.b32 %r157, %r229, %r231;

dp4a.s32.s32 %r72, %r157, %r74, %r68;

shr.u32 %r232, %r192, 4;
and.b32 %r233, %r232, 252645135;
shr.u32 %r234, %r193, 12;
and.b32 %r235, %r234, 16;
or.b32 %r236, %r235, %r233;
shr.u32 %r237, %r193, 5;
and.b32 %r238, %r237, 4096;
or.b32 %r239, %r236, %r238;
shl.b32 %r240, %r193, 2;
and.b32 %r241, %r240, 1048576;
or.b32 %r242, %r239, %r241;
shl.b32 %r243, %r193, 9;
and.b32 %r244, %r243, 268435456;
or.b32 %r161, %r242, %r244;

dp4a.s32.s32 %r76, %r161, %r78, %r72;

ld.global.nc.u32 %r81, [%rd55+-8];
ld.global.nc.u32 %r82, [%rd28];

{mul.f16x2 %r80,%r81,%r82;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r80;
cvt.f32.f16 %f37, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r80;
cvt.f32.f16 %f38, high;}


cvt.rn.f32.s32 %f49, %r76;
mov.f32 %f50, 0f40000000;
div.approx.ftz.f32 %f51, %f38, %f50;
fma.rn.ftz.f32 %f52, %f37, %f49, %f51;
add.ftz.f32 %f169, %f52, %f169;
add.s64 %rd32, %rd56, %rd5;
ld.global.nc.u32 %r245, [%rd56+-4];
shr.s32 %r246, %r245, %r8;
ld.global.nc.u32 %r247, [%rd32+4];
shr.s32 %r248, %r245, %r9;
ld.global.nc.u32 %r249, [%rd32];
and.b32 %r250, %r249, 252645135;
shl.b32 %r251, %r246, 4;
and.b32 %r252, %r251, 16;
or.b32 %r253, %r252, %r250;
shl.b32 %r254, %r246, 11;
and.b32 %r255, %r254, 4096;
or.b32 %r256, %r253, %r255;
shl.b32 %r257, %r246, 18;
and.b32 %r258, %r257, 1048576;
or.b32 %r259, %r256, %r258;
shl.b32 %r260, %r246, 25;
and.b32 %r261, %r260, 268435456;
or.b32 %r170, %r259, %r261;

dp4a.s32.s32 %r85, %r170, %r66, %r172;

shr.u32 %r262, %r249, 4;
and.b32 %r263, %r262, 252645135;
shr.u32 %r264, %r246, 12;
and.b32 %r265, %r264, 16;
or.b32 %r266, %r265, %r263;
shr.u32 %r267, %r246, 5;
and.b32 %r268, %r267, 4096;
or.b32 %r269, %r266, %r268;
shl.b32 %r270, %r246, 2;
and.b32 %r271, %r270, 1048576;
or.b32 %r272, %r269, %r271;
shl.b32 %r273, %r246, 9;
and.b32 %r274, %r273, 268435456;
or.b32 %r174, %r272, %r274;

dp4a.s32.s32 %r89, %r174, %r70, %r85;

and.b32 %r275, %r247, 252645135;
shl.b32 %r276, %r248, 4;
and.b32 %r277, %r276, 16;
or.b32 %r278, %r277, %r275;
shl.b32 %r279, %r248, 11;
and.b32 %r280, %r279, 4096;
or.b32 %r281, %r278, %r280;
shl.b32 %r282, %r248, 18;
and.b32 %r283, %r282, 1048576;
or.b32 %r284, %r281, %r283;
shl.b32 %r285, %r248, 25;
and.b32 %r286, %r285, 268435456;
or.b32 %r178, %r284, %r286;

dp4a.s32.s32 %r93, %r178, %r74, %r89;

shr.u32 %r287, %r247, 4;
and.b32 %r288, %r287, 252645135;
shr.u32 %r289, %r248, 12;
and.b32 %r290, %r289, 16;
or.b32 %r291, %r290, %r288;
shr.u32 %r292, %r248, 5;
and.b32 %r293, %r292, 4096;
or.b32 %r294, %r291, %r293;
shl.b32 %r295, %r248, 2;
and.b32 %r296, %r295, 1048576;
or.b32 %r297, %r294, %r296;
shl.b32 %r298, %r248, 9;
and.b32 %r299, %r298, 268435456;
or.b32 %r182, %r297, %r299;

dp4a.s32.s32 %r97, %r182, %r78, %r93;

ld.global.nc.u32 %r102, [%rd56+-8];

{mul.f16x2 %r101,%r102,%r82;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r101;
cvt.f32.f16 %f39, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r101;
cvt.f32.f16 %f40, high;}


cvt.rn.f32.s32 %f53, %r97;
div.approx.ftz.f32 %f54, %f40, %f50;
fma.rn.ftz.f32 %f55, %f39, %f53, %f54;
add.ftz.f32 %f168, %f55, %f168;
cvt.u64.u32 %rd33, %r416;
add.s64 %rd34, %rd2, %rd33;
mul.lo.s64 %rd35, %rd34, 36;
add.s64 %rd36, %rd7, %rd35;
add.s64 %rd37, %rd36, %rd29;
ld.global.nc.u32 %r108, [%rd37+4];
ld.global.nc.u32 %r112, [%rd37+20];
ld.global.nc.u32 %r116, [%rd37+8];
ld.global.nc.u32 %r120, [%rd37+24];

dp4a.s32.s32 %r106, %r149, %r108, %r172;


dp4a.s32.s32 %r110, %r153, %r112, %r106;


dp4a.s32.s32 %r114, %r157, %r116, %r110;


dp4a.s32.s32 %r118, %r161, %r120, %r114;

ld.global.nc.u32 %r124, [%rd36];

{mul.f16x2 %r122,%r81,%r124;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r122;
cvt.f32.f16 %f41, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r122;
cvt.f32.f16 %f42, high;}


cvt.rn.f32.s32 %f56, %r118;
div.approx.ftz.f32 %f57, %f42, %f50;
fma.rn.ftz.f32 %f58, %f41, %f56, %f57;
add.ftz.f32 %f167, %f58, %f167;

dp4a.s32.s32 %r127, %r170, %r108, %r172;


dp4a.s32.s32 %r131, %r174, %r112, %r127;


dp4a.s32.s32 %r135, %r178, %r116, %r131;


dp4a.s32.s32 %r139, %r182, %r120, %r135;


{mul.f16x2 %r143,%r102,%r124;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r143;
cvt.f32.f16 %f43, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r143;
cvt.f32.f16 %f44, high;}


cvt.rn.f32.s32 %f59, %r139;
div.approx.ftz.f32 %f60, %f44, %f50;
fma.rn.ftz.f32 %f61, %f43, %f59, %f60;
add.ftz.f32 %f166, %f61, %f166;
cvt.u64.u32 %rd38, %r417;
add.s64 %rd39, %rd2, %rd38;
mul.lo.s64 %rd40, %rd39, 36;
add.s64 %rd41, %rd7, %rd40;
add.s64 %rd42, %rd41, %rd29;
ld.global.nc.u32 %r150, [%rd42+4];
ld.global.nc.u32 %r154, [%rd42+20];
ld.global.nc.u32 %r158, [%rd42+8];
ld.global.nc.u32 %r162, [%rd42+24];

dp4a.s32.s32 %r148, %r149, %r150, %r172;


dp4a.s32.s32 %r152, %r153, %r154, %r148;


dp4a.s32.s32 %r156, %r157, %r158, %r152;


dp4a.s32.s32 %r160, %r161, %r162, %r156;

ld.global.nc.u32 %r166, [%rd41];

{mul.f16x2 %r164,%r81,%r166;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r164;
cvt.f32.f16 %f45, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r164;
cvt.f32.f16 %f46, high;}


cvt.rn.f32.s32 %f62, %r160;
div.approx.ftz.f32 %f63, %f46, %f50;
fma.rn.ftz.f32 %f64, %f45, %f62, %f63;
add.ftz.f32 %f165, %f64, %f165;

dp4a.s32.s32 %r169, %r170, %r150, %r172;


dp4a.s32.s32 %r173, %r174, %r154, %r169;


dp4a.s32.s32 %r177, %r178, %r158, %r173;


dp4a.s32.s32 %r181, %r182, %r162, %r177;


{mul.f16x2 %r185,%r102,%r166;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r185;
cvt.f32.f16 %f47, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r185;
cvt.f32.f16 %f48, high;}


cvt.rn.f32.s32 %f65, %r181;
div.approx.ftz.f32 %f66, %f48, %f50;
fma.rn.ftz.f32 %f67, %f47, %f65, %f66;
add.ftz.f32 %f164, %f67, %f164;
add.s32 %r417, %r417, 64;
add.s32 %r416, %r416, 64;
add.s64 %rd56, %rd56, 1536;
add.s64 %rd55, %rd55, 1536;
add.s32 %r418, %r418, 64;
setp.lt.s32 %p4, %r418, %r7;
@%p4 bra $L__BB80_2;

$L__BB80_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB80_5;

mov.u32 %r301, _ZZ13mul_mat_vec_qIL9ggml_type7ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r302, %r1, 768, %r301;
shl.b32 %r303, %r37, 2;
add.s32 %r304, %r302, %r303;
st.shared.f32 [%r304+-768], %f169;
st.shared.f32 [%r304+-640], %f168;
st.shared.f32 [%r304+-512], %f167;
st.shared.f32 [%r304+-384], %f166;
st.shared.f32 [%r304+-256], %f165;
st.shared.f32 [%r304+-128], %f164;

$L__BB80_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB80_18;

mad.lo.s32 %r305, %r4, %r28, %r3;
mad.lo.s32 %r306, %r5, %r34, %r305;
cvt.u64.u32 %rd12, %r306;
shl.b32 %r308, %r37, 2;
mov.u32 %r309, _ZZ13mul_mat_vec_qIL9ggml_type7ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r18, %r309, %r308;
mov.u32 %r310, 2;
ld.shared.f32 %f68, [%r18];
add.ftz.f32 %f69, %f68, %f169;
ld.shared.f32 %f70, [%r18+768];
add.ftz.f32 %f71, %f70, %f69;
ld.shared.f32 %f72, [%r18+1536];
add.ftz.f32 %f73, %f72, %f71;
mov.b32 %r311, %f73;
mov.u32 %r312, 31;
mov.u32 %r313, 16;
mov.u32 %r314, -1;
shfl.sync.bfly.b32 %r315|%p7, %r311, %r313, %r312, %r314;
mov.b32 %f74, %r315;
add.ftz.f32 %f75, %f73, %f74;
mov.b32 %r316, %f75;
mov.u32 %r317, 8;
shfl.sync.bfly.b32 %r318|%p8, %r316, %r317, %r312, %r314;
mov.b32 %f76, %r318;
add.ftz.f32 %f77, %f75, %f76;
mov.b32 %r319, %f77;
mov.u32 %r320, 4;
shfl.sync.bfly.b32 %r321|%p9, %r319, %r320, %r312, %r314;
mov.b32 %f78, %r321;
add.ftz.f32 %f79, %f77, %f78;
mov.b32 %r322, %f79;
shfl.sync.bfly.b32 %r323|%p10, %r322, %r310, %r312, %r314;
mov.b32 %f80, %r323;
add.ftz.f32 %f81, %f79, %f80;
mov.b32 %r324, %f81;
mov.u32 %r325, 1;
shfl.sync.bfly.b32 %r326|%p11, %r324, %r325, %r312, %r314;
mov.b32 %f82, %r326;
add.ftz.f32 %f19, %f81, %f82;
setp.eq.s32 %p12, %r37, 0;
setp.lt.u32 %p13, %r3, %r22;
and.pred %p1, %p12, %p13;
mul.wide.u32 %rd43, %r306, 4;
add.s64 %rd13, %rd1, %rd43;
not.pred %p14, %p1;
@%p14 bra $L__BB80_8;

st.global.f32 [%rd13], %f19;

$L__BB80_8:
ld.shared.f32 %f83, [%r18+128];
add.ftz.f32 %f84, %f83, %f168;
ld.shared.f32 %f85, [%r18+896];
add.ftz.f32 %f86, %f85, %f84;
ld.shared.f32 %f87, [%r18+1664];
add.ftz.f32 %f88, %f87, %f86;
mov.b32 %r327, %f88;
shfl.sync.bfly.b32 %r331|%p15, %r327, %r313, %r312, %r314;
mov.b32 %f89, %r331;
add.ftz.f32 %f90, %f88, %f89;
mov.b32 %r332, %f90;
shfl.sync.bfly.b32 %r334|%p16, %r332, %r317, %r312, %r314;
mov.b32 %f91, %r334;
add.ftz.f32 %f92, %f90, %f91;
mov.b32 %r335, %f92;
shfl.sync.bfly.b32 %r337|%p17, %r335, %r320, %r312, %r314;
mov.b32 %f93, %r337;
add.ftz.f32 %f94, %f92, %f93;
mov.b32 %r338, %f94;
shfl.sync.bfly.b32 %r340|%p18, %r338, %r310, %r312, %r314;
mov.b32 %f95, %r340;
add.ftz.f32 %f96, %f94, %f95;
mov.b32 %r341, %f96;
shfl.sync.bfly.b32 %r343|%p19, %r341, %r325, %r312, %r314;
mov.b32 %f97, %r343;
add.ftz.f32 %f20, %f96, %f97;
setp.eq.s32 %p20, %r37, 1;
add.s32 %r345, %r3, 1;
setp.lt.u32 %p21, %r345, %r22;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB80_10;

st.global.f32 [%rd13+4], %f20;

$L__BB80_10:
ld.shared.f32 %f98, [%r18+256];
add.ftz.f32 %f99, %f98, %f167;
ld.shared.f32 %f100, [%r18+1024];
add.ftz.f32 %f101, %f100, %f99;
ld.shared.f32 %f102, [%r18+1792];
add.ftz.f32 %f103, %f102, %f101;
mov.b32 %r346, %f103;
mov.u32 %r347, 31;
mov.u32 %r348, 16;
mov.u32 %r349, -1;
shfl.sync.bfly.b32 %r350|%p23, %r346, %r348, %r347, %r349;
mov.b32 %f104, %r350;
add.ftz.f32 %f105, %f103, %f104;
mov.b32 %r351, %f105;
mov.u32 %r352, 8;
shfl.sync.bfly.b32 %r353|%p24, %r351, %r352, %r347, %r349;
mov.b32 %f106, %r353;
add.ftz.f32 %f107, %f105, %f106;
mov.b32 %r354, %f107;
mov.u32 %r355, 4;
shfl.sync.bfly.b32 %r356|%p25, %r354, %r355, %r347, %r349;
mov.b32 %f108, %r356;
add.ftz.f32 %f109, %f107, %f108;
mov.b32 %r357, %f109;
mov.u32 %r358, 2;
shfl.sync.bfly.b32 %r359|%p26, %r357, %r358, %r347, %r349;
mov.b32 %f110, %r359;
add.ftz.f32 %f111, %f109, %f110;
mov.b32 %r360, %f111;
mov.u32 %r361, 1;
shfl.sync.bfly.b32 %r362|%p27, %r360, %r361, %r347, %r349;
mov.b32 %f112, %r362;
add.ftz.f32 %f21, %f111, %f112;
@%p14 bra $L__BB80_12;

cvt.u64.u32 %rd44, %r22;
add.s64 %rd45, %rd44, %rd12;
shl.b64 %rd46, %rd45, 2;
add.s64 %rd47, %rd1, %rd46;
st.global.f32 [%rd47], %f21;

$L__BB80_12:
ld.shared.f32 %f113, [%r18+384];
add.ftz.f32 %f114, %f113, %f166;
ld.shared.f32 %f115, [%r18+1152];
add.ftz.f32 %f116, %f115, %f114;
ld.shared.f32 %f117, [%r18+1920];
add.ftz.f32 %f118, %f117, %f116;
mov.b32 %r363, %f118;
shfl.sync.bfly.b32 %r367|%p29, %r363, %r348, %r347, %r349;
mov.b32 %f119, %r367;
add.ftz.f32 %f120, %f118, %f119;
mov.b32 %r368, %f120;
shfl.sync.bfly.b32 %r370|%p30, %r368, %r352, %r347, %r349;
mov.b32 %f121, %r370;
add.ftz.f32 %f122, %f120, %f121;
mov.b32 %r371, %f122;
shfl.sync.bfly.b32 %r373|%p31, %r371, %r355, %r347, %r349;
mov.b32 %f123, %r373;
add.ftz.f32 %f124, %f122, %f123;
mov.b32 %r374, %f124;
shfl.sync.bfly.b32 %r376|%p32, %r374, %r358, %r347, %r349;
mov.b32 %f125, %r376;
add.ftz.f32 %f126, %f124, %f125;
mov.b32 %r377, %f126;
shfl.sync.bfly.b32 %r379|%p33, %r377, %r361, %r347, %r349;
mov.b32 %f127, %r379;
add.ftz.f32 %f22, %f126, %f127;
@%p22 bra $L__BB80_14;

add.s32 %r380, %r22, 1;
cvt.u64.u32 %rd48, %r380;
add.s64 %rd49, %rd48, %rd12;
shl.b64 %rd50, %rd49, 2;
add.s64 %rd51, %rd1, %rd50;
st.global.f32 [%rd51], %f22;

$L__BB80_14:
shl.b32 %r381, %r22, 1;
mov.u32 %r382, 1;
ld.shared.f32 %f128, [%r18+512];
add.ftz.f32 %f129, %f128, %f165;
ld.shared.f32 %f130, [%r18+1280];
add.ftz.f32 %f131, %f130, %f129;
ld.shared.f32 %f132, [%r18+2048];
add.ftz.f32 %f133, %f132, %f131;
mov.b32 %r383, %f133;
mov.u32 %r384, 31;
mov.u32 %r385, 16;
mov.u32 %r386, -1;
shfl.sync.bfly.b32 %r387|%p35, %r383, %r385, %r384, %r386;
mov.b32 %f134, %r387;
add.ftz.f32 %f135, %f133, %f134;
mov.b32 %r388, %f135;
mov.u32 %r389, 8;
shfl.sync.bfly.b32 %r390|%p36, %r388, %r389, %r384, %r386;
mov.b32 %f136, %r390;
add.ftz.f32 %f137, %f135, %f136;
mov.b32 %r391, %f137;
mov.u32 %r392, 4;
shfl.sync.bfly.b32 %r393|%p37, %r391, %r392, %r384, %r386;
mov.b32 %f138, %r393;
add.ftz.f32 %f139, %f137, %f138;
mov.b32 %r394, %f139;
mov.u32 %r395, 2;
shfl.sync.bfly.b32 %r396|%p38, %r394, %r395, %r384, %r386;
mov.b32 %f140, %r396;
add.ftz.f32 %f141, %f139, %f140;
mov.b32 %r397, %f141;
shfl.sync.bfly.b32 %r398|%p39, %r397, %r382, %r384, %r386;
mov.b32 %f142, %r398;
add.ftz.f32 %f23, %f141, %f142;
cvt.u64.u32 %rd52, %r381;
add.s64 %rd53, %rd52, %rd12;
shl.b64 %rd54, %rd53, 2;
add.s64 %rd14, %rd1, %rd54;
@%p14 bra $L__BB80_16;

st.global.f32 [%rd14], %f23;

$L__BB80_16:
ld.shared.f32 %f143, [%r18+640];
add.ftz.f32 %f144, %f143, %f164;
ld.shared.f32 %f145, [%r18+1408];
add.ftz.f32 %f146, %f145, %f144;
ld.shared.f32 %f147, [%r18+2176];
add.ftz.f32 %f148, %f147, %f146;
mov.b32 %r399, %f148;
shfl.sync.bfly.b32 %r403|%p41, %r399, %r385, %r384, %r386;
mov.b32 %f149, %r403;
add.ftz.f32 %f150, %f148, %f149;
mov.b32 %r404, %f150;
shfl.sync.bfly.b32 %r406|%p42, %r404, %r389, %r384, %r386;
mov.b32 %f151, %r406;
add.ftz.f32 %f152, %f150, %f151;
mov.b32 %r407, %f152;
shfl.sync.bfly.b32 %r409|%p43, %r407, %r392, %r384, %r386;
mov.b32 %f153, %r409;
add.ftz.f32 %f154, %f152, %f153;
mov.b32 %r410, %f154;
shfl.sync.bfly.b32 %r412|%p44, %r410, %r395, %r384, %r386;
mov.b32 %f155, %r412;
add.ftz.f32 %f156, %f154, %f155;
mov.b32 %r413, %f156;
shfl.sync.bfly.b32 %r415|%p45, %r413, %r382, %r384, %r386;
mov.b32 %f157, %r415;
add.ftz.f32 %f24, %f156, %f157;
@%p22 bra $L__BB80_18;

st.global.f32 [%rd14+4], %f24;

$L__BB80_18:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type7ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type7ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type7ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type7ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type7ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type7ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type7ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type7ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type7ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<59>;
.reg .f32 %f<226>;
.reg .b32 %r<506>;
.reg .b64 %rd<75>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type7ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[3072];

ld.param.u64 %rd14, [_Z13mul_mat_vec_qIL9ggml_type7ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type7ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type7ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type7ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type7ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type7ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type7ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type7ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type7ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type7ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type7ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type7ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type7ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type7ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type7ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type7ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type7ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r38, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r38, %r2;
mov.u32 %r39, %ctaid.x;
shl.b32 %r4, %r39, 1;
shr.u32 %r40, %r3, 31;
add.s32 %r41, %r3, %r40;
shr.s32 %r505, %r41, 1;
shr.u32 %r6, %r37, 5;
setp.ge.s32 %p3, %r505, %r6;
mov.f32 %f218, 0f00000000;
mov.f32 %f219, %f218;
mov.f32 %f220, %f218;
mov.f32 %f221, %f218;
mov.f32 %f222, %f218;
mov.f32 %f223, %f218;
mov.f32 %f224, %f218;
mov.f32 %f225, %f218;
@%p3 bra $L__BB81_3;

mov.u32 %r42, %ctaid.y;
mov.u32 %r43, %ctaid.z;
mul.lo.s32 %r44, %r42, %r29;
cvt.u64.u32 %rd17, %r44;
mul.lo.s32 %r45, %r43, %r35;
cvt.u64.u32 %rd18, %r45;
add.s64 %rd1, %rd18, %rd17;
mul.hi.u32 %r46, %r43, %r31;
add.s32 %r47, %r46, %r43;
shr.u32 %r48, %r47, %r32;
mul.lo.s32 %r49, %r48, %r34;
mul.hi.u32 %r50, %r42, %r25;
add.s32 %r51, %r50, %r42;
shr.u32 %r52, %r51, %r26;
mul.lo.s32 %r53, %r52, %r28;
cvta.to.global.u64 %rd19, %rd14;
and.b32 %r56, %r41, -2;
sub.s32 %r57, %r3, %r56;
shl.b32 %r58, %r57, 1;
cvt.s64.s32 %rd2, %r58;
shl.b32 %r7, %r57, 3;
or.b32 %r8, %r7, 4;
shl.b32 %r59, %r23, 1;
add.s32 %r504, %r505, %r59;
mad.lo.s32 %r503, %r23, 3, %r505;
add.s32 %r502, %r505, %r23;
add.s32 %r60, %r505, %r49;
add.s32 %r61, %r60, %r53;
add.s32 %r62, %r4, 1;
mad.lo.s32 %r63, %r22, %r62, %r61;
mul.wide.s32 %rd20, %r63, 24;
add.s64 %rd21, %rd19, %rd20;
add.s64 %rd74, %rd21, 8;
mul.wide.s32 %rd4, %r58, 4;
mad.lo.s32 %r64, %r4, %r22, %r505;
add.s32 %r65, %r64, %r49;
add.s32 %r66, %r65, %r53;
mul.wide.s32 %rd22, %r66, 24;
add.s64 %rd23, %rd19, %rd22;
add.s64 %rd73, %rd23, 8;
cvta.to.global.u64 %rd6, %rd15;
mov.f32 %f218, 0f00000000;
shl.b64 %rd28, %rd2, 2;

$L__BB81_2:
cvt.u64.u32 %rd24, %r505;
add.s64 %rd25, %rd1, %rd24;
mul.lo.s64 %rd26, %rd25, 36;
add.s64 %rd27, %rd6, %rd26;
add.s64 %rd29, %rd27, %rd28;
add.s64 %rd30, %rd73, %rd4;
ld.global.nc.u32 %r235, [%rd73+-4];
shr.s32 %r236, %r235, %r7;
ld.global.nc.u32 %r69, [%rd29+4];
ld.global.nc.u32 %r73, [%rd29+20];
ld.global.nc.u32 %r237, [%rd30+4];
shr.s32 %r238, %r235, %r8;
ld.global.nc.u32 %r77, [%rd29+8];
ld.global.nc.u32 %r81, [%rd29+24];
ld.global.nc.u32 %r239, [%rd30];
and.b32 %r240, %r239, 252645135;
shl.b32 %r241, %r236, 4;
and.b32 %r242, %r241, 16;
or.b32 %r243, %r242, %r240;
shl.b32 %r244, %r236, 11;
and.b32 %r245, %r244, 4096;
or.b32 %r246, %r243, %r245;
shl.b32 %r247, %r236, 18;
and.b32 %r248, %r247, 1048576;
or.b32 %r249, %r246, %r248;
shl.b32 %r250, %r236, 25;
and.b32 %r251, %r250, 268435456;
or.b32 %r194, %r249, %r251;
mov.u32 %r217, 0;

dp4a.s32.s32 %r67, %r194, %r69, %r217;

shr.u32 %r252, %r239, 4;
and.b32 %r253, %r252, 252645135;
shr.u32 %r254, %r236, 12;
and.b32 %r255, %r254, 16;
or.b32 %r256, %r255, %r253;
shr.u32 %r257, %r236, 5;
and.b32 %r258, %r257, 4096;
or.b32 %r259, %r256, %r258;
shl.b32 %r260, %r236, 2;
and.b32 %r261, %r260, 1048576;
or.b32 %r262, %r259, %r261;
shl.b32 %r263, %r236, 9;
and.b32 %r264, %r263, 268435456;
or.b32 %r198, %r262, %r264;

dp4a.s32.s32 %r71, %r198, %r73, %r67;

and.b32 %r265, %r237, 252645135;
shl.b32 %r266, %r238, 4;
and.b32 %r267, %r266, 16;
or.b32 %r268, %r267, %r265;
shl.b32 %r269, %r238, 11;
and.b32 %r270, %r269, 4096;
or.b32 %r271, %r268, %r270;
shl.b32 %r272, %r238, 18;
and.b32 %r273, %r272, 1048576;
or.b32 %r274, %r271, %r273;
shl.b32 %r275, %r238, 25;
and.b32 %r276, %r275, 268435456;
or.b32 %r202, %r274, %r276;

dp4a.s32.s32 %r75, %r202, %r77, %r71;

shr.u32 %r277, %r237, 4;
and.b32 %r278, %r277, 252645135;
shr.u32 %r279, %r238, 12;
and.b32 %r280, %r279, 16;
or.b32 %r281, %r280, %r278;
shr.u32 %r282, %r238, 5;
and.b32 %r283, %r282, 4096;
or.b32 %r284, %r281, %r283;
shl.b32 %r285, %r238, 2;
and.b32 %r286, %r285, 1048576;
or.b32 %r287, %r284, %r286;
shl.b32 %r288, %r238, 9;
and.b32 %r289, %r288, 268435456;
or.b32 %r206, %r287, %r289;

dp4a.s32.s32 %r79, %r206, %r81, %r75;

ld.global.nc.u32 %r84, [%rd73+-8];
ld.global.nc.u32 %r85, [%rd27];

{mul.f16x2 %r83,%r84,%r85;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r83;
cvt.f32.f16 %f49, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r83;
cvt.f32.f16 %f50, high;}


cvt.rn.f32.s32 %f65, %r79;
mov.f32 %f66, 0f40000000;
div.approx.ftz.f32 %f67, %f50, %f66;
fma.rn.ftz.f32 %f68, %f49, %f65, %f67;
add.ftz.f32 %f225, %f68, %f225;
add.s64 %rd31, %rd74, %rd4;
ld.global.nc.u32 %r290, [%rd74+-4];
shr.s32 %r291, %r290, %r7;
ld.global.nc.u32 %r292, [%rd31+4];
shr.s32 %r293, %r290, %r8;
ld.global.nc.u32 %r294, [%rd31];
and.b32 %r295, %r294, 252645135;
shl.b32 %r296, %r291, 4;
and.b32 %r297, %r296, 16;
or.b32 %r298, %r297, %r295;
shl.b32 %r299, %r291, 11;
and.b32 %r300, %r299, 4096;
or.b32 %r301, %r298, %r300;
shl.b32 %r302, %r291, 18;
and.b32 %r303, %r302, 1048576;
or.b32 %r304, %r301, %r303;
shl.b32 %r305, %r291, 25;
and.b32 %r306, %r305, 268435456;
or.b32 %r215, %r304, %r306;

dp4a.s32.s32 %r88, %r215, %r69, %r217;

shr.u32 %r307, %r294, 4;
and.b32 %r308, %r307, 252645135;
shr.u32 %r309, %r291, 12;
and.b32 %r310, %r309, 16;
or.b32 %r311, %r310, %r308;
shr.u32 %r312, %r291, 5;
and.b32 %r313, %r312, 4096;
or.b32 %r314, %r311, %r313;
shl.b32 %r315, %r291, 2;
and.b32 %r316, %r315, 1048576;
or.b32 %r317, %r314, %r316;
shl.b32 %r318, %r291, 9;
and.b32 %r319, %r318, 268435456;
or.b32 %r219, %r317, %r319;

dp4a.s32.s32 %r92, %r219, %r73, %r88;

and.b32 %r320, %r292, 252645135;
shl.b32 %r321, %r293, 4;
and.b32 %r322, %r321, 16;
or.b32 %r323, %r322, %r320;
shl.b32 %r324, %r293, 11;
and.b32 %r325, %r324, 4096;
or.b32 %r326, %r323, %r325;
shl.b32 %r327, %r293, 18;
and.b32 %r328, %r327, 1048576;
or.b32 %r329, %r326, %r328;
shl.b32 %r330, %r293, 25;
and.b32 %r331, %r330, 268435456;
or.b32 %r223, %r329, %r331;

dp4a.s32.s32 %r96, %r223, %r77, %r92;

shr.u32 %r332, %r292, 4;
and.b32 %r333, %r332, 252645135;
shr.u32 %r334, %r293, 12;
and.b32 %r335, %r334, 16;
or.b32 %r336, %r335, %r333;
shr.u32 %r337, %r293, 5;
and.b32 %r338, %r337, 4096;
or.b32 %r339, %r336, %r338;
shl.b32 %r340, %r293, 2;
and.b32 %r341, %r340, 1048576;
or.b32 %r342, %r339, %r341;
shl.b32 %r343, %r293, 9;
and.b32 %r344, %r343, 268435456;
or.b32 %r227, %r342, %r344;

dp4a.s32.s32 %r100, %r227, %r81, %r96;

ld.global.nc.u32 %r105, [%rd74+-8];

{mul.f16x2 %r104,%r105,%r85;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r104;
cvt.f32.f16 %f51, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r104;
cvt.f32.f16 %f52, high;}


cvt.rn.f32.s32 %f69, %r100;
div.approx.ftz.f32 %f70, %f52, %f66;
fma.rn.ftz.f32 %f71, %f51, %f69, %f70;
add.ftz.f32 %f224, %f71, %f224;
cvt.u64.u32 %rd32, %r502;
add.s64 %rd33, %rd1, %rd32;
mul.lo.s64 %rd34, %rd33, 36;
add.s64 %rd35, %rd6, %rd34;
add.s64 %rd36, %rd35, %rd28;
ld.global.nc.u32 %r111, [%rd36+4];
ld.global.nc.u32 %r115, [%rd36+20];
ld.global.nc.u32 %r119, [%rd36+8];
ld.global.nc.u32 %r123, [%rd36+24];

dp4a.s32.s32 %r109, %r194, %r111, %r217;


dp4a.s32.s32 %r113, %r198, %r115, %r109;


dp4a.s32.s32 %r117, %r202, %r119, %r113;


dp4a.s32.s32 %r121, %r206, %r123, %r117;

ld.global.nc.u32 %r127, [%rd35];

{mul.f16x2 %r125,%r84,%r127;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r125;
cvt.f32.f16 %f53, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r125;
cvt.f32.f16 %f54, high;}


cvt.rn.f32.s32 %f72, %r121;
div.approx.ftz.f32 %f73, %f54, %f66;
fma.rn.ftz.f32 %f74, %f53, %f72, %f73;
add.ftz.f32 %f223, %f74, %f223;

dp4a.s32.s32 %r130, %r215, %r111, %r217;


dp4a.s32.s32 %r134, %r219, %r115, %r130;


dp4a.s32.s32 %r138, %r223, %r119, %r134;


dp4a.s32.s32 %r142, %r227, %r123, %r138;


{mul.f16x2 %r146,%r105,%r127;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r146;
cvt.f32.f16 %f55, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r146;
cvt.f32.f16 %f56, high;}


cvt.rn.f32.s32 %f75, %r142;
div.approx.ftz.f32 %f76, %f56, %f66;
fma.rn.ftz.f32 %f77, %f55, %f75, %f76;
add.ftz.f32 %f222, %f77, %f222;
cvt.u64.u32 %rd37, %r504;
add.s64 %rd38, %rd1, %rd37;
mul.lo.s64 %rd39, %rd38, 36;
add.s64 %rd40, %rd6, %rd39;
add.s64 %rd41, %rd40, %rd28;
ld.global.nc.u32 %r153, [%rd41+4];
ld.global.nc.u32 %r157, [%rd41+20];
ld.global.nc.u32 %r161, [%rd41+8];
ld.global.nc.u32 %r165, [%rd41+24];

dp4a.s32.s32 %r151, %r194, %r153, %r217;


dp4a.s32.s32 %r155, %r198, %r157, %r151;


dp4a.s32.s32 %r159, %r202, %r161, %r155;


dp4a.s32.s32 %r163, %r206, %r165, %r159;

ld.global.nc.u32 %r169, [%rd40];

{mul.f16x2 %r167,%r84,%r169;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r167;
cvt.f32.f16 %f57, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r167;
cvt.f32.f16 %f58, high;}


cvt.rn.f32.s32 %f78, %r163;
div.approx.ftz.f32 %f79, %f58, %f66;
fma.rn.ftz.f32 %f80, %f57, %f78, %f79;
add.ftz.f32 %f221, %f80, %f221;

dp4a.s32.s32 %r172, %r215, %r153, %r217;


dp4a.s32.s32 %r176, %r219, %r157, %r172;


dp4a.s32.s32 %r180, %r223, %r161, %r176;


dp4a.s32.s32 %r184, %r227, %r165, %r180;


{mul.f16x2 %r188,%r105,%r169;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r188;
cvt.f32.f16 %f59, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r188;
cvt.f32.f16 %f60, high;}


cvt.rn.f32.s32 %f81, %r184;
div.approx.ftz.f32 %f82, %f60, %f66;
fma.rn.ftz.f32 %f83, %f59, %f81, %f82;
add.ftz.f32 %f220, %f83, %f220;
cvt.u64.u32 %rd42, %r503;
add.s64 %rd43, %rd1, %rd42;
mul.lo.s64 %rd44, %rd43, 36;
add.s64 %rd45, %rd6, %rd44;
add.s64 %rd46, %rd45, %rd28;
ld.global.nc.u32 %r195, [%rd46+4];
ld.global.nc.u32 %r199, [%rd46+20];
ld.global.nc.u32 %r203, [%rd46+8];
ld.global.nc.u32 %r207, [%rd46+24];

dp4a.s32.s32 %r193, %r194, %r195, %r217;


dp4a.s32.s32 %r197, %r198, %r199, %r193;


dp4a.s32.s32 %r201, %r202, %r203, %r197;


dp4a.s32.s32 %r205, %r206, %r207, %r201;

ld.global.nc.u32 %r211, [%rd45];

{mul.f16x2 %r209,%r84,%r211;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r209;
cvt.f32.f16 %f61, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r209;
cvt.f32.f16 %f62, high;}


cvt.rn.f32.s32 %f84, %r205;
div.approx.ftz.f32 %f85, %f62, %f66;
fma.rn.ftz.f32 %f86, %f61, %f84, %f85;
add.ftz.f32 %f219, %f86, %f219;

dp4a.s32.s32 %r214, %r215, %r195, %r217;


dp4a.s32.s32 %r218, %r219, %r199, %r214;


dp4a.s32.s32 %r222, %r223, %r203, %r218;


dp4a.s32.s32 %r226, %r227, %r207, %r222;


{mul.f16x2 %r230,%r105,%r211;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r230;
cvt.f32.f16 %f63, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r230;
cvt.f32.f16 %f64, high;}


cvt.rn.f32.s32 %f87, %r226;
div.approx.ftz.f32 %f88, %f64, %f66;
fma.rn.ftz.f32 %f89, %f63, %f87, %f88;
add.ftz.f32 %f218, %f89, %f218;
add.s32 %r504, %r504, 64;
add.s32 %r503, %r503, 64;
add.s32 %r502, %r502, 64;
add.s64 %rd74, %rd74, 1536;
add.s64 %rd73, %rd73, 1536;
add.s32 %r505, %r505, 64;
setp.lt.s32 %p4, %r505, %r6;
@%p4 bra $L__BB81_2;

$L__BB81_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB81_5;

shl.b32 %r345, %r1, 10;
mov.u32 %r346, _ZZ13mul_mat_vec_qIL9ggml_type7ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r347, %r346, %r345;
shl.b32 %r348, %r2, 2;
add.s32 %r349, %r347, %r348;
st.shared.f32 [%r349+-1024], %f225;
st.shared.f32 [%r349+-896], %f224;
st.shared.f32 [%r349+-768], %f223;
st.shared.f32 [%r349+-640], %f222;
st.shared.f32 [%r349+-512], %f221;
st.shared.f32 [%r349+-384], %f220;
st.shared.f32 [%r349+-256], %f219;
st.shared.f32 [%r349+-128], %f218;

$L__BB81_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB81_22;

mov.u32 %r350, %ctaid.z;
mov.u32 %r351, %ctaid.y;
mad.lo.s32 %r352, %r351, %r30, %r4;
mad.lo.s32 %r353, %r350, %r36, %r352;
cvt.u64.u32 %rd11, %r353;
shl.b32 %r354, %r2, 2;
mov.u32 %r355, _ZZ13mul_mat_vec_qIL9ggml_type7ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r356, %r355, %r354;
mov.u32 %r357, 2;
ld.shared.f32 %f90, [%r356];
add.ftz.f32 %f91, %f90, %f225;
ld.shared.f32 %f92, [%r356+1024];
add.ftz.f32 %f93, %f92, %f91;
ld.shared.f32 %f94, [%r356+2048];
add.ftz.f32 %f95, %f94, %f93;
mov.b32 %r358, %f95;
mov.u32 %r359, 31;
mov.u32 %r360, 16;
mov.u32 %r361, -1;
shfl.sync.bfly.b32 %r362|%p7, %r358, %r360, %r359, %r361;
mov.b32 %f96, %r362;
add.ftz.f32 %f97, %f95, %f96;
mov.b32 %r363, %f97;
mov.u32 %r364, 8;
shfl.sync.bfly.b32 %r365|%p8, %r363, %r364, %r359, %r361;
mov.b32 %f98, %r365;
add.ftz.f32 %f99, %f97, %f98;
mov.b32 %r366, %f99;
mov.u32 %r367, 4;
shfl.sync.bfly.b32 %r368|%p9, %r366, %r367, %r359, %r361;
mov.b32 %f100, %r368;
add.ftz.f32 %f101, %f99, %f100;
mov.b32 %r369, %f101;
shfl.sync.bfly.b32 %r370|%p10, %r369, %r357, %r359, %r361;
mov.b32 %f102, %r370;
add.ftz.f32 %f103, %f101, %f102;
mov.b32 %r371, %f103;
mov.u32 %r372, 1;
shfl.sync.bfly.b32 %r373|%p11, %r371, %r372, %r359, %r361;
mov.b32 %f104, %r373;
add.ftz.f32 %f25, %f103, %f104;
setp.lt.u32 %p12, %r4, %r24;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
cvta.to.global.u64 %rd47, %rd16;
mul.wide.u32 %rd48, %r353, 4;
add.s64 %rd12, %rd47, %rd48;
not.pred %p14, %p1;
@%p14 bra $L__BB81_8;

st.global.f32 [%rd12], %f25;

$L__BB81_8:
ld.shared.f32 %f105, [%r356+128];
add.ftz.f32 %f106, %f105, %f224;
ld.shared.f32 %f107, [%r356+1152];
add.ftz.f32 %f108, %f107, %f106;
ld.shared.f32 %f109, [%r356+2176];
add.ftz.f32 %f110, %f109, %f108;
mov.b32 %r378, %f110;
shfl.sync.bfly.b32 %r382|%p15, %r378, %r360, %r359, %r361;
mov.b32 %f111, %r382;
add.ftz.f32 %f112, %f110, %f111;
mov.b32 %r383, %f112;
shfl.sync.bfly.b32 %r385|%p16, %r383, %r364, %r359, %r361;
mov.b32 %f113, %r385;
add.ftz.f32 %f114, %f112, %f113;
mov.b32 %r386, %f114;
shfl.sync.bfly.b32 %r388|%p17, %r386, %r367, %r359, %r361;
mov.b32 %f115, %r388;
add.ftz.f32 %f116, %f114, %f115;
mov.b32 %r389, %f116;
shfl.sync.bfly.b32 %r390|%p18, %r389, %r357, %r359, %r361;
mov.b32 %f117, %r390;
add.ftz.f32 %f118, %f116, %f117;
mov.b32 %r391, %f118;
shfl.sync.bfly.b32 %r393|%p19, %r391, %r372, %r359, %r361;
mov.b32 %f119, %r393;
add.ftz.f32 %f26, %f118, %f119;
setp.eq.s32 %p20, %r2, 1;
or.b32 %r396, %r4, 1;
setp.lt.u32 %p21, %r396, %r24;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB81_10;

st.global.f32 [%rd12+4], %f26;

$L__BB81_10:
ld.shared.f32 %f120, [%r356+256];
add.ftz.f32 %f121, %f120, %f223;
ld.shared.f32 %f122, [%r356+1280];
add.ftz.f32 %f123, %f122, %f121;
ld.shared.f32 %f124, [%r356+2304];
add.ftz.f32 %f125, %f124, %f123;
mov.b32 %r397, %f125;
mov.u32 %r398, 31;
mov.u32 %r399, 16;
mov.u32 %r400, -1;
shfl.sync.bfly.b32 %r401|%p23, %r397, %r399, %r398, %r400;
mov.b32 %f126, %r401;
add.ftz.f32 %f127, %f125, %f126;
mov.b32 %r402, %f127;
mov.u32 %r403, 8;
shfl.sync.bfly.b32 %r404|%p24, %r402, %r403, %r398, %r400;
mov.b32 %f128, %r404;
add.ftz.f32 %f129, %f127, %f128;
mov.b32 %r405, %f129;
mov.u32 %r406, 4;
shfl.sync.bfly.b32 %r407|%p25, %r405, %r406, %r398, %r400;
mov.b32 %f130, %r407;
add.ftz.f32 %f131, %f129, %f130;
mov.b32 %r408, %f131;
mov.u32 %r409, 2;
shfl.sync.bfly.b32 %r410|%p26, %r408, %r409, %r398, %r400;
mov.b32 %f132, %r410;
add.ftz.f32 %f133, %f131, %f132;
mov.b32 %r411, %f133;
mov.u32 %r412, 1;
shfl.sync.bfly.b32 %r413|%p27, %r411, %r412, %r398, %r400;
mov.b32 %f134, %r413;
add.ftz.f32 %f27, %f133, %f134;
@%p14 bra $L__BB81_12;

cvt.u64.u32 %rd49, %r24;
add.s64 %rd50, %rd49, %rd11;
shl.b64 %rd52, %rd50, 2;
add.s64 %rd53, %rd47, %rd52;
st.global.f32 [%rd53], %f27;

$L__BB81_12:
ld.shared.f32 %f135, [%r356+384];
add.ftz.f32 %f136, %f135, %f222;
ld.shared.f32 %f137, [%r356+1408];
add.ftz.f32 %f138, %f137, %f136;
ld.shared.f32 %f139, [%r356+2432];
add.ftz.f32 %f140, %f139, %f138;
mov.b32 %r414, %f140;
shfl.sync.bfly.b32 %r418|%p29, %r414, %r399, %r398, %r400;
mov.b32 %f141, %r418;
add.ftz.f32 %f142, %f140, %f141;
mov.b32 %r419, %f142;
shfl.sync.bfly.b32 %r421|%p30, %r419, %r403, %r398, %r400;
mov.b32 %f143, %r421;
add.ftz.f32 %f144, %f142, %f143;
mov.b32 %r422, %f144;
shfl.sync.bfly.b32 %r424|%p31, %r422, %r406, %r398, %r400;
mov.b32 %f145, %r424;
add.ftz.f32 %f146, %f144, %f145;
mov.b32 %r425, %f146;
shfl.sync.bfly.b32 %r427|%p32, %r425, %r409, %r398, %r400;
mov.b32 %f147, %r427;
add.ftz.f32 %f148, %f146, %f147;
mov.b32 %r428, %f148;
shfl.sync.bfly.b32 %r430|%p33, %r428, %r412, %r398, %r400;
mov.b32 %f149, %r430;
add.ftz.f32 %f28, %f148, %f149;
@%p22 bra $L__BB81_14;

add.s32 %r431, %r24, 1;
cvt.u64.u32 %rd54, %r431;
add.s64 %rd55, %rd54, %rd11;
shl.b64 %rd57, %rd55, 2;
add.s64 %rd58, %rd47, %rd57;
st.global.f32 [%rd58], %f28;

$L__BB81_14:
shl.b32 %r432, %r24, 1;
mov.u32 %r433, 1;
ld.shared.f32 %f150, [%r356+512];
add.ftz.f32 %f151, %f150, %f221;
ld.shared.f32 %f152, [%r356+1536];
add.ftz.f32 %f153, %f152, %f151;
ld.shared.f32 %f154, [%r356+2560];
add.ftz.f32 %f155, %f154, %f153;
mov.b32 %r434, %f155;
mov.u32 %r435, 31;
mov.u32 %r436, 16;
mov.u32 %r437, -1;
shfl.sync.bfly.b32 %r438|%p35, %r434, %r436, %r435, %r437;
mov.b32 %f156, %r438;
add.ftz.f32 %f157, %f155, %f156;
mov.b32 %r439, %f157;
mov.u32 %r440, 8;
shfl.sync.bfly.b32 %r441|%p36, %r439, %r440, %r435, %r437;
mov.b32 %f158, %r441;
add.ftz.f32 %f159, %f157, %f158;
mov.b32 %r442, %f159;
mov.u32 %r443, 4;
shfl.sync.bfly.b32 %r444|%p37, %r442, %r443, %r435, %r437;
mov.b32 %f160, %r444;
add.ftz.f32 %f161, %f159, %f160;
mov.b32 %r445, %f161;
mov.u32 %r446, 2;
shfl.sync.bfly.b32 %r447|%p38, %r445, %r446, %r435, %r437;
mov.b32 %f162, %r447;
add.ftz.f32 %f163, %f161, %f162;
mov.b32 %r448, %f163;
shfl.sync.bfly.b32 %r449|%p39, %r448, %r433, %r435, %r437;
mov.b32 %f164, %r449;
add.ftz.f32 %f29, %f163, %f164;
cvt.u64.u32 %rd59, %r432;
add.s64 %rd60, %rd59, %rd11;
shl.b64 %rd62, %rd60, 2;
add.s64 %rd13, %rd47, %rd62;
@%p14 bra $L__BB81_16;

st.global.f32 [%rd13], %f29;

$L__BB81_16:
ld.shared.f32 %f165, [%r356+640];
add.ftz.f32 %f166, %f165, %f220;
ld.shared.f32 %f167, [%r356+1664];
add.ftz.f32 %f168, %f167, %f166;
ld.shared.f32 %f169, [%r356+2688];
add.ftz.f32 %f170, %f169, %f168;
mov.b32 %r450, %f170;
shfl.sync.bfly.b32 %r454|%p41, %r450, %r436, %r435, %r437;
mov.b32 %f171, %r454;
add.ftz.f32 %f172, %f170, %f171;
mov.b32 %r455, %f172;
shfl.sync.bfly.b32 %r457|%p42, %r455, %r440, %r435, %r437;
mov.b32 %f173, %r457;
add.ftz.f32 %f174, %f172, %f173;
mov.b32 %r458, %f174;
shfl.sync.bfly.b32 %r460|%p43, %r458, %r443, %r435, %r437;
mov.b32 %f175, %r460;
add.ftz.f32 %f176, %f174, %f175;
mov.b32 %r461, %f176;
shfl.sync.bfly.b32 %r463|%p44, %r461, %r446, %r435, %r437;
mov.b32 %f177, %r463;
add.ftz.f32 %f178, %f176, %f177;
mov.b32 %r464, %f178;
shfl.sync.bfly.b32 %r466|%p45, %r464, %r433, %r435, %r437;
mov.b32 %f179, %r466;
add.ftz.f32 %f30, %f178, %f179;
@%p22 bra $L__BB81_18;

st.global.f32 [%rd13+4], %f30;

$L__BB81_18:
mul.lo.s32 %r21, %r24, 3;
ld.shared.f32 %f180, [%r356+768];
add.ftz.f32 %f181, %f180, %f219;
ld.shared.f32 %f182, [%r356+1792];
add.ftz.f32 %f183, %f182, %f181;
ld.shared.f32 %f184, [%r356+2816];
add.ftz.f32 %f185, %f184, %f183;
mov.b32 %r467, %f185;
mov.u32 %r468, 31;
mov.u32 %r469, 16;
mov.u32 %r470, -1;
shfl.sync.bfly.b32 %r471|%p47, %r467, %r469, %r468, %r470;
mov.b32 %f186, %r471;
add.ftz.f32 %f187, %f185, %f186;
mov.b32 %r472, %f187;
mov.u32 %r473, 8;
shfl.sync.bfly.b32 %r474|%p48, %r472, %r473, %r468, %r470;
mov.b32 %f188, %r474;
add.ftz.f32 %f189, %f187, %f188;
mov.b32 %r475, %f189;
mov.u32 %r476, 4;
shfl.sync.bfly.b32 %r477|%p49, %r475, %r476, %r468, %r470;
mov.b32 %f190, %r477;
add.ftz.f32 %f191, %f189, %f190;
mov.b32 %r478, %f191;
mov.u32 %r479, 2;
shfl.sync.bfly.b32 %r480|%p50, %r478, %r479, %r468, %r470;
mov.b32 %f192, %r480;
add.ftz.f32 %f193, %f191, %f192;
mov.b32 %r481, %f193;
mov.u32 %r482, 1;
shfl.sync.bfly.b32 %r483|%p51, %r481, %r482, %r468, %r470;
mov.b32 %f194, %r483;
add.ftz.f32 %f31, %f193, %f194;
@%p14 bra $L__BB81_20;

cvt.u64.u32 %rd63, %r21;
add.s64 %rd64, %rd63, %rd11;
shl.b64 %rd66, %rd64, 2;
add.s64 %rd67, %rd47, %rd66;
st.global.f32 [%rd67], %f31;

$L__BB81_20:
ld.shared.f32 %f195, [%r356+896];
add.ftz.f32 %f196, %f195, %f218;
ld.shared.f32 %f197, [%r356+1920];
add.ftz.f32 %f198, %f197, %f196;
ld.shared.f32 %f199, [%r356+2944];
add.ftz.f32 %f200, %f199, %f198;
mov.b32 %r484, %f200;
shfl.sync.bfly.b32 %r488|%p53, %r484, %r469, %r468, %r470;
mov.b32 %f201, %r488;
add.ftz.f32 %f202, %f200, %f201;
mov.b32 %r489, %f202;
shfl.sync.bfly.b32 %r491|%p54, %r489, %r473, %r468, %r470;
mov.b32 %f203, %r491;
add.ftz.f32 %f204, %f202, %f203;
mov.b32 %r492, %f204;
shfl.sync.bfly.b32 %r494|%p55, %r492, %r476, %r468, %r470;
mov.b32 %f205, %r494;
add.ftz.f32 %f206, %f204, %f205;
mov.b32 %r495, %f206;
shfl.sync.bfly.b32 %r497|%p56, %r495, %r479, %r468, %r470;
mov.b32 %f207, %r497;
add.ftz.f32 %f208, %f206, %f207;
mov.b32 %r498, %f208;
shfl.sync.bfly.b32 %r500|%p57, %r498, %r482, %r468, %r470;
mov.b32 %f209, %r500;
add.ftz.f32 %f32, %f208, %f209;
@%p22 bra $L__BB81_22;

add.s32 %r501, %r21, 1;
cvt.u64.u32 %rd68, %r501;
add.s64 %rd69, %rd68, %rd11;
shl.b64 %rd71, %rd69, 2;
add.s64 %rd72, %rd47, %rd71;
st.global.f32 [%rd72], %f32;

$L__BB81_22:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type7ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type7ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type7ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type7ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type7ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type7ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type7ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type7ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type7ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<71>;
.reg .f32 %f<242>;
.reg .b32 %r<584>;
.reg .b64 %rd<85>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type7ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1280];

ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type7ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type7ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type7ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type7ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type7ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type7ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type7ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type7ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type7ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type7ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type7ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type7ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type7ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type7ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type7ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type7ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type7ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r41, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r41, %r2;
mov.u32 %r42, %ctaid.x;
shl.b32 %r4, %r42, 1;
shr.u32 %r43, %r3, 31;
add.s32 %r44, %r3, %r43;
shr.s32 %r583, %r44, 1;
shr.u32 %r6, %r40, 5;
setp.ge.s32 %p3, %r583, %r6;
mov.f32 %f232, 0f00000000;
mov.f32 %f233, %f232;
mov.f32 %f234, %f232;
mov.f32 %f235, %f232;
mov.f32 %f236, %f232;
mov.f32 %f237, %f232;
mov.f32 %f238, %f232;
mov.f32 %f239, %f232;
mov.f32 %f240, %f232;
mov.f32 %f241, %f232;
@%p3 bra $L__BB82_3;

mov.u32 %r45, %ctaid.y;
mov.u32 %r46, %ctaid.z;
mul.lo.s32 %r47, %r45, %r32;
cvt.u64.u32 %rd18, %r47;
mul.lo.s32 %r48, %r46, %r38;
cvt.u64.u32 %rd19, %r48;
add.s64 %rd1, %rd19, %rd18;
mul.hi.u32 %r49, %r46, %r34;
add.s32 %r50, %r49, %r46;
shr.u32 %r51, %r50, %r35;
mul.lo.s32 %r52, %r51, %r37;
mul.hi.u32 %r53, %r45, %r28;
add.s32 %r54, %r53, %r45;
shr.u32 %r55, %r54, %r29;
mul.lo.s32 %r56, %r55, %r31;
cvta.to.global.u64 %rd20, %rd15;
and.b32 %r59, %r44, -2;
sub.s32 %r60, %r3, %r59;
shl.b32 %r61, %r60, 1;
cvt.s64.s32 %rd2, %r61;
shl.b32 %r7, %r60, 3;
or.b32 %r8, %r7, 4;
shl.b32 %r62, %r26, 1;
add.s32 %r582, %r583, %r62;
mad.lo.s32 %r581, %r26, 3, %r583;
shl.b32 %r63, %r26, 2;
add.s32 %r580, %r583, %r63;
add.s32 %r579, %r583, %r26;
add.s32 %r64, %r583, %r52;
add.s32 %r65, %r64, %r56;
add.s32 %r66, %r4, 1;
mad.lo.s32 %r67, %r25, %r66, %r65;
mul.wide.s32 %rd21, %r67, 24;
add.s64 %rd22, %rd20, %rd21;
add.s64 %rd84, %rd22, 8;
mul.wide.s32 %rd4, %r61, 4;
mad.lo.s32 %r68, %r4, %r25, %r583;
add.s32 %r69, %r68, %r52;
add.s32 %r70, %r69, %r56;
mul.wide.s32 %rd23, %r70, 24;
add.s64 %rd24, %rd20, %rd23;
add.s64 %rd83, %rd24, 8;
cvta.to.global.u64 %rd6, %rd16;
mov.f32 %f232, 0f00000000;
shl.b64 %rd29, %rd2, 2;

$L__BB82_2:
cvt.u64.u32 %rd25, %r583;
add.s64 %rd26, %rd1, %rd25;
mul.lo.s64 %rd27, %rd26, 36;
add.s64 %rd28, %rd6, %rd27;
add.s64 %rd30, %rd28, %rd29;
add.s64 %rd31, %rd83, %rd4;
ld.global.nc.u32 %r281, [%rd83+-4];
shr.s32 %r282, %r281, %r7;
ld.global.nc.u32 %r73, [%rd30+4];
ld.global.nc.u32 %r77, [%rd30+20];
ld.global.nc.u32 %r283, [%rd31+4];
shr.s32 %r284, %r281, %r8;
ld.global.nc.u32 %r81, [%rd30+8];
ld.global.nc.u32 %r85, [%rd30+24];
ld.global.nc.u32 %r285, [%rd31];
and.b32 %r286, %r285, 252645135;
shl.b32 %r287, %r282, 4;
and.b32 %r288, %r287, 16;
or.b32 %r289, %r288, %r286;
shl.b32 %r290, %r282, 11;
and.b32 %r291, %r290, 4096;
or.b32 %r292, %r289, %r291;
shl.b32 %r293, %r282, 18;
and.b32 %r294, %r293, 1048576;
or.b32 %r295, %r292, %r294;
shl.b32 %r296, %r282, 25;
and.b32 %r297, %r296, 268435456;
or.b32 %r240, %r295, %r297;
mov.u32 %r263, 0;

dp4a.s32.s32 %r71, %r240, %r73, %r263;

shr.u32 %r298, %r285, 4;
and.b32 %r299, %r298, 252645135;
shr.u32 %r300, %r282, 12;
and.b32 %r301, %r300, 16;
or.b32 %r302, %r301, %r299;
shr.u32 %r303, %r282, 5;
and.b32 %r304, %r303, 4096;
or.b32 %r305, %r302, %r304;
shl.b32 %r306, %r282, 2;
and.b32 %r307, %r306, 1048576;
or.b32 %r308, %r305, %r307;
shl.b32 %r309, %r282, 9;
and.b32 %r310, %r309, 268435456;
or.b32 %r244, %r308, %r310;

dp4a.s32.s32 %r75, %r244, %r77, %r71;

and.b32 %r311, %r283, 252645135;
shl.b32 %r312, %r284, 4;
and.b32 %r313, %r312, 16;
or.b32 %r314, %r313, %r311;
shl.b32 %r315, %r284, 11;
and.b32 %r316, %r315, 4096;
or.b32 %r317, %r314, %r316;
shl.b32 %r318, %r284, 18;
and.b32 %r319, %r318, 1048576;
or.b32 %r320, %r317, %r319;
shl.b32 %r321, %r284, 25;
and.b32 %r322, %r321, 268435456;
or.b32 %r248, %r320, %r322;

dp4a.s32.s32 %r79, %r248, %r81, %r75;

shr.u32 %r323, %r283, 4;
and.b32 %r324, %r323, 252645135;
shr.u32 %r325, %r284, 12;
and.b32 %r326, %r325, 16;
or.b32 %r327, %r326, %r324;
shr.u32 %r328, %r284, 5;
and.b32 %r329, %r328, 4096;
or.b32 %r330, %r327, %r329;
shl.b32 %r331, %r284, 2;
and.b32 %r332, %r331, 1048576;
or.b32 %r333, %r330, %r332;
shl.b32 %r334, %r284, 9;
and.b32 %r335, %r334, 268435456;
or.b32 %r252, %r333, %r335;

dp4a.s32.s32 %r83, %r252, %r85, %r79;

ld.global.nc.u32 %r88, [%rd83+-8];
ld.global.nc.u32 %r89, [%rd28];

{mul.f16x2 %r87,%r88,%r89;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r87;
cvt.f32.f16 %f61, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r87;
cvt.f32.f16 %f62, high;}


cvt.rn.f32.s32 %f81, %r83;
mov.f32 %f82, 0f40000000;
div.approx.ftz.f32 %f83, %f62, %f82;
fma.rn.ftz.f32 %f84, %f61, %f81, %f83;
add.ftz.f32 %f241, %f84, %f241;
add.s64 %rd32, %rd84, %rd4;
ld.global.nc.u32 %r336, [%rd84+-4];
shr.s32 %r337, %r336, %r7;
ld.global.nc.u32 %r338, [%rd32+4];
shr.s32 %r339, %r336, %r8;
ld.global.nc.u32 %r340, [%rd32];
and.b32 %r341, %r340, 252645135;
shl.b32 %r342, %r337, 4;
and.b32 %r343, %r342, 16;
or.b32 %r344, %r343, %r341;
shl.b32 %r345, %r337, 11;
and.b32 %r346, %r345, 4096;
or.b32 %r347, %r344, %r346;
shl.b32 %r348, %r337, 18;
and.b32 %r349, %r348, 1048576;
or.b32 %r350, %r347, %r349;
shl.b32 %r351, %r337, 25;
and.b32 %r352, %r351, 268435456;
or.b32 %r261, %r350, %r352;

dp4a.s32.s32 %r92, %r261, %r73, %r263;

shr.u32 %r353, %r340, 4;
and.b32 %r354, %r353, 252645135;
shr.u32 %r355, %r337, 12;
and.b32 %r356, %r355, 16;
or.b32 %r357, %r356, %r354;
shr.u32 %r358, %r337, 5;
and.b32 %r359, %r358, 4096;
or.b32 %r360, %r357, %r359;
shl.b32 %r361, %r337, 2;
and.b32 %r362, %r361, 1048576;
or.b32 %r363, %r360, %r362;
shl.b32 %r364, %r337, 9;
and.b32 %r365, %r364, 268435456;
or.b32 %r265, %r363, %r365;

dp4a.s32.s32 %r96, %r265, %r77, %r92;

and.b32 %r366, %r338, 252645135;
shl.b32 %r367, %r339, 4;
and.b32 %r368, %r367, 16;
or.b32 %r369, %r368, %r366;
shl.b32 %r370, %r339, 11;
and.b32 %r371, %r370, 4096;
or.b32 %r372, %r369, %r371;
shl.b32 %r373, %r339, 18;
and.b32 %r374, %r373, 1048576;
or.b32 %r375, %r372, %r374;
shl.b32 %r376, %r339, 25;
and.b32 %r377, %r376, 268435456;
or.b32 %r269, %r375, %r377;

dp4a.s32.s32 %r100, %r269, %r81, %r96;

shr.u32 %r378, %r338, 4;
and.b32 %r379, %r378, 252645135;
shr.u32 %r380, %r339, 12;
and.b32 %r381, %r380, 16;
or.b32 %r382, %r381, %r379;
shr.u32 %r383, %r339, 5;
and.b32 %r384, %r383, 4096;
or.b32 %r385, %r382, %r384;
shl.b32 %r386, %r339, 2;
and.b32 %r387, %r386, 1048576;
or.b32 %r388, %r385, %r387;
shl.b32 %r389, %r339, 9;
and.b32 %r390, %r389, 268435456;
or.b32 %r273, %r388, %r390;

dp4a.s32.s32 %r104, %r273, %r85, %r100;

ld.global.nc.u32 %r109, [%rd84+-8];

{mul.f16x2 %r108,%r109,%r89;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r108;
cvt.f32.f16 %f63, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r108;
cvt.f32.f16 %f64, high;}


cvt.rn.f32.s32 %f85, %r104;
div.approx.ftz.f32 %f86, %f64, %f82;
fma.rn.ftz.f32 %f87, %f63, %f85, %f86;
add.ftz.f32 %f240, %f87, %f240;
cvt.u64.u32 %rd33, %r579;
add.s64 %rd34, %rd1, %rd33;
mul.lo.s64 %rd35, %rd34, 36;
add.s64 %rd36, %rd6, %rd35;
add.s64 %rd37, %rd36, %rd29;
ld.global.nc.u32 %r115, [%rd37+4];
ld.global.nc.u32 %r119, [%rd37+20];
ld.global.nc.u32 %r123, [%rd37+8];
ld.global.nc.u32 %r127, [%rd37+24];

dp4a.s32.s32 %r113, %r240, %r115, %r263;


dp4a.s32.s32 %r117, %r244, %r119, %r113;


dp4a.s32.s32 %r121, %r248, %r123, %r117;


dp4a.s32.s32 %r125, %r252, %r127, %r121;

ld.global.nc.u32 %r131, [%rd36];

{mul.f16x2 %r129,%r88,%r131;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r129;
cvt.f32.f16 %f65, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r129;
cvt.f32.f16 %f66, high;}


cvt.rn.f32.s32 %f88, %r125;
div.approx.ftz.f32 %f89, %f66, %f82;
fma.rn.ftz.f32 %f90, %f65, %f88, %f89;
add.ftz.f32 %f239, %f90, %f239;

dp4a.s32.s32 %r134, %r261, %r115, %r263;


dp4a.s32.s32 %r138, %r265, %r119, %r134;


dp4a.s32.s32 %r142, %r269, %r123, %r138;


dp4a.s32.s32 %r146, %r273, %r127, %r142;


{mul.f16x2 %r150,%r109,%r131;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r150;
cvt.f32.f16 %f67, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r150;
cvt.f32.f16 %f68, high;}


cvt.rn.f32.s32 %f91, %r146;
div.approx.ftz.f32 %f92, %f68, %f82;
fma.rn.ftz.f32 %f93, %f67, %f91, %f92;
add.ftz.f32 %f238, %f93, %f238;
cvt.u64.u32 %rd38, %r582;
add.s64 %rd39, %rd1, %rd38;
mul.lo.s64 %rd40, %rd39, 36;
add.s64 %rd41, %rd6, %rd40;
add.s64 %rd42, %rd41, %rd29;
ld.global.nc.u32 %r157, [%rd42+4];
ld.global.nc.u32 %r161, [%rd42+20];
ld.global.nc.u32 %r165, [%rd42+8];
ld.global.nc.u32 %r169, [%rd42+24];

dp4a.s32.s32 %r155, %r240, %r157, %r263;


dp4a.s32.s32 %r159, %r244, %r161, %r155;


dp4a.s32.s32 %r163, %r248, %r165, %r159;


dp4a.s32.s32 %r167, %r252, %r169, %r163;

ld.global.nc.u32 %r173, [%rd41];

{mul.f16x2 %r171,%r88,%r173;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r171;
cvt.f32.f16 %f69, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r171;
cvt.f32.f16 %f70, high;}


cvt.rn.f32.s32 %f94, %r167;
div.approx.ftz.f32 %f95, %f70, %f82;
fma.rn.ftz.f32 %f96, %f69, %f94, %f95;
add.ftz.f32 %f237, %f96, %f237;

dp4a.s32.s32 %r176, %r261, %r157, %r263;


dp4a.s32.s32 %r180, %r265, %r161, %r176;


dp4a.s32.s32 %r184, %r269, %r165, %r180;


dp4a.s32.s32 %r188, %r273, %r169, %r184;


{mul.f16x2 %r192,%r109,%r173;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r192;
cvt.f32.f16 %f71, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r192;
cvt.f32.f16 %f72, high;}


cvt.rn.f32.s32 %f97, %r188;
div.approx.ftz.f32 %f98, %f72, %f82;
fma.rn.ftz.f32 %f99, %f71, %f97, %f98;
add.ftz.f32 %f236, %f99, %f236;
cvt.u64.u32 %rd43, %r581;
add.s64 %rd44, %rd1, %rd43;
mul.lo.s64 %rd45, %rd44, 36;
add.s64 %rd46, %rd6, %rd45;
add.s64 %rd47, %rd46, %rd29;
ld.global.nc.u32 %r199, [%rd47+4];
ld.global.nc.u32 %r203, [%rd47+20];
ld.global.nc.u32 %r207, [%rd47+8];
ld.global.nc.u32 %r211, [%rd47+24];

dp4a.s32.s32 %r197, %r240, %r199, %r263;


dp4a.s32.s32 %r201, %r244, %r203, %r197;


dp4a.s32.s32 %r205, %r248, %r207, %r201;


dp4a.s32.s32 %r209, %r252, %r211, %r205;

ld.global.nc.u32 %r215, [%rd46];

{mul.f16x2 %r213,%r88,%r215;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r213;
cvt.f32.f16 %f73, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r213;
cvt.f32.f16 %f74, high;}


cvt.rn.f32.s32 %f100, %r209;
div.approx.ftz.f32 %f101, %f74, %f82;
fma.rn.ftz.f32 %f102, %f73, %f100, %f101;
add.ftz.f32 %f235, %f102, %f235;

dp4a.s32.s32 %r218, %r261, %r199, %r263;


dp4a.s32.s32 %r222, %r265, %r203, %r218;


dp4a.s32.s32 %r226, %r269, %r207, %r222;


dp4a.s32.s32 %r230, %r273, %r211, %r226;


{mul.f16x2 %r234,%r109,%r215;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r234;
cvt.f32.f16 %f75, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r234;
cvt.f32.f16 %f76, high;}


cvt.rn.f32.s32 %f103, %r230;
div.approx.ftz.f32 %f104, %f76, %f82;
fma.rn.ftz.f32 %f105, %f75, %f103, %f104;
add.ftz.f32 %f234, %f105, %f234;
cvt.u64.u32 %rd48, %r580;
add.s64 %rd49, %rd1, %rd48;
mul.lo.s64 %rd50, %rd49, 36;
add.s64 %rd51, %rd6, %rd50;
add.s64 %rd52, %rd51, %rd29;
ld.global.nc.u32 %r241, [%rd52+4];
ld.global.nc.u32 %r245, [%rd52+20];
ld.global.nc.u32 %r249, [%rd52+8];
ld.global.nc.u32 %r253, [%rd52+24];

dp4a.s32.s32 %r239, %r240, %r241, %r263;


dp4a.s32.s32 %r243, %r244, %r245, %r239;


dp4a.s32.s32 %r247, %r248, %r249, %r243;


dp4a.s32.s32 %r251, %r252, %r253, %r247;

ld.global.nc.u32 %r257, [%rd51];

{mul.f16x2 %r255,%r88,%r257;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r255;
cvt.f32.f16 %f77, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r255;
cvt.f32.f16 %f78, high;}


cvt.rn.f32.s32 %f106, %r251;
div.approx.ftz.f32 %f107, %f78, %f82;
fma.rn.ftz.f32 %f108, %f77, %f106, %f107;
add.ftz.f32 %f233, %f108, %f233;

dp4a.s32.s32 %r260, %r261, %r241, %r263;


dp4a.s32.s32 %r264, %r265, %r245, %r260;


dp4a.s32.s32 %r268, %r269, %r249, %r264;


dp4a.s32.s32 %r272, %r273, %r253, %r268;


{mul.f16x2 %r276,%r109,%r257;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r276;
cvt.f32.f16 %f79, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r276;
cvt.f32.f16 %f80, high;}


cvt.rn.f32.s32 %f109, %r272;
div.approx.ftz.f32 %f110, %f80, %f82;
fma.rn.ftz.f32 %f111, %f79, %f109, %f110;
add.ftz.f32 %f232, %f111, %f232;
add.s32 %r582, %r582, 32;
add.s32 %r581, %r581, 32;
add.s32 %r580, %r580, 32;
add.s32 %r579, %r579, 32;
add.s64 %rd84, %rd84, 768;
add.s64 %rd83, %rd83, 768;
add.s32 %r583, %r583, 32;
setp.lt.s32 %p4, %r583, %r6;
@%p4 bra $L__BB82_2;

$L__BB82_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB82_5;

mov.u32 %r391, _ZZ13mul_mat_vec_qIL9ggml_type7ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r392, %r1, 1280, %r391;
shl.b32 %r393, %r2, 2;
add.s32 %r394, %r392, %r393;
st.shared.f32 [%r394+-1280], %f241;
st.shared.f32 [%r394+-1152], %f240;
st.shared.f32 [%r394+-1024], %f239;
st.shared.f32 [%r394+-896], %f238;
st.shared.f32 [%r394+-768], %f237;
st.shared.f32 [%r394+-640], %f236;
st.shared.f32 [%r394+-512], %f235;
st.shared.f32 [%r394+-384], %f234;
st.shared.f32 [%r394+-256], %f233;
st.shared.f32 [%r394+-128], %f232;

$L__BB82_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB82_26;

mov.u32 %r395, %ctaid.z;
mov.u32 %r396, %ctaid.y;
mad.lo.s32 %r397, %r396, %r33, %r4;
mad.lo.s32 %r398, %r395, %r39, %r397;
cvt.u64.u32 %rd11, %r398;
shl.b32 %r399, %r2, 2;
mov.u32 %r400, _ZZ13mul_mat_vec_qIL9ggml_type7ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r23, %r400, %r399;
mov.u32 %r401, 2;
ld.shared.f32 %f112, [%r23];
add.ftz.f32 %f113, %f112, %f241;
mov.b32 %r402, %f113;
mov.u32 %r403, 31;
mov.u32 %r404, 16;
mov.u32 %r405, -1;
shfl.sync.bfly.b32 %r406|%p7, %r402, %r404, %r403, %r405;
mov.b32 %f114, %r406;
add.ftz.f32 %f115, %f113, %f114;
mov.b32 %r407, %f115;
mov.u32 %r408, 8;
shfl.sync.bfly.b32 %r409|%p8, %r407, %r408, %r403, %r405;
mov.b32 %f116, %r409;
add.ftz.f32 %f117, %f115, %f116;
mov.b32 %r410, %f117;
mov.u32 %r411, 4;
shfl.sync.bfly.b32 %r412|%p9, %r410, %r411, %r403, %r405;
mov.b32 %f118, %r412;
add.ftz.f32 %f119, %f117, %f118;
mov.b32 %r413, %f119;
shfl.sync.bfly.b32 %r414|%p10, %r413, %r401, %r403, %r405;
mov.b32 %f120, %r414;
add.ftz.f32 %f121, %f119, %f120;
mov.b32 %r415, %f121;
mov.u32 %r416, 1;
shfl.sync.bfly.b32 %r417|%p11, %r415, %r416, %r403, %r405;
mov.b32 %f122, %r417;
add.ftz.f32 %f31, %f121, %f122;
setp.lt.u32 %p12, %r4, %r27;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
cvta.to.global.u64 %rd53, %rd17;
mul.wide.u32 %rd54, %r398, 4;
add.s64 %rd12, %rd53, %rd54;
not.pred %p14, %p1;
@%p14 bra $L__BB82_8;

st.global.f32 [%rd12], %f31;

$L__BB82_8:
ld.shared.f32 %f123, [%r23+128];
add.ftz.f32 %f124, %f123, %f240;
mov.b32 %r418, %f124;
shfl.sync.bfly.b32 %r422|%p15, %r418, %r404, %r403, %r405;
mov.b32 %f125, %r422;
add.ftz.f32 %f126, %f124, %f125;
mov.b32 %r423, %f126;
shfl.sync.bfly.b32 %r425|%p16, %r423, %r408, %r403, %r405;
mov.b32 %f127, %r425;
add.ftz.f32 %f128, %f126, %f127;
mov.b32 %r426, %f128;
shfl.sync.bfly.b32 %r428|%p17, %r426, %r411, %r403, %r405;
mov.b32 %f129, %r428;
add.ftz.f32 %f130, %f128, %f129;
mov.b32 %r429, %f130;
shfl.sync.bfly.b32 %r431|%p18, %r429, %r401, %r403, %r405;
mov.b32 %f131, %r431;
add.ftz.f32 %f132, %f130, %f131;
mov.b32 %r432, %f132;
shfl.sync.bfly.b32 %r434|%p19, %r432, %r416, %r403, %r405;
mov.b32 %f133, %r434;
add.ftz.f32 %f32, %f132, %f133;
setp.eq.s32 %p20, %r2, 1;
or.b32 %r438, %r4, 1;
setp.lt.u32 %p21, %r438, %r27;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB82_10;

st.global.f32 [%rd12+4], %f32;

$L__BB82_10:
ld.shared.f32 %f134, [%r23+256];
add.ftz.f32 %f135, %f134, %f239;
mov.b32 %r439, %f135;
mov.u32 %r440, 31;
mov.u32 %r441, 16;
mov.u32 %r442, -1;
shfl.sync.bfly.b32 %r443|%p23, %r439, %r441, %r440, %r442;
mov.b32 %f136, %r443;
add.ftz.f32 %f137, %f135, %f136;
mov.b32 %r444, %f137;
mov.u32 %r445, 8;
shfl.sync.bfly.b32 %r446|%p24, %r444, %r445, %r440, %r442;
mov.b32 %f138, %r446;
add.ftz.f32 %f139, %f137, %f138;
mov.b32 %r447, %f139;
mov.u32 %r448, 4;
shfl.sync.bfly.b32 %r449|%p25, %r447, %r448, %r440, %r442;
mov.b32 %f140, %r449;
add.ftz.f32 %f141, %f139, %f140;
mov.b32 %r450, %f141;
mov.u32 %r451, 2;
shfl.sync.bfly.b32 %r452|%p26, %r450, %r451, %r440, %r442;
mov.b32 %f142, %r452;
add.ftz.f32 %f143, %f141, %f142;
mov.b32 %r453, %f143;
mov.u32 %r454, 1;
shfl.sync.bfly.b32 %r455|%p27, %r453, %r454, %r440, %r442;
mov.b32 %f144, %r455;
add.ftz.f32 %f33, %f143, %f144;
@%p14 bra $L__BB82_12;

cvt.u64.u32 %rd55, %r27;
add.s64 %rd56, %rd55, %rd11;
shl.b64 %rd58, %rd56, 2;
add.s64 %rd59, %rd53, %rd58;
st.global.f32 [%rd59], %f33;

$L__BB82_12:
ld.shared.f32 %f145, [%r23+384];
add.ftz.f32 %f146, %f145, %f238;
mov.b32 %r456, %f146;
shfl.sync.bfly.b32 %r460|%p29, %r456, %r441, %r440, %r442;
mov.b32 %f147, %r460;
add.ftz.f32 %f148, %f146, %f147;
mov.b32 %r461, %f148;
shfl.sync.bfly.b32 %r463|%p30, %r461, %r445, %r440, %r442;
mov.b32 %f149, %r463;
add.ftz.f32 %f150, %f148, %f149;
mov.b32 %r464, %f150;
shfl.sync.bfly.b32 %r466|%p31, %r464, %r448, %r440, %r442;
mov.b32 %f151, %r466;
add.ftz.f32 %f152, %f150, %f151;
mov.b32 %r467, %f152;
shfl.sync.bfly.b32 %r469|%p32, %r467, %r451, %r440, %r442;
mov.b32 %f153, %r469;
add.ftz.f32 %f154, %f152, %f153;
mov.b32 %r470, %f154;
shfl.sync.bfly.b32 %r472|%p33, %r470, %r454, %r440, %r442;
mov.b32 %f155, %r472;
add.ftz.f32 %f34, %f154, %f155;
@%p22 bra $L__BB82_14;

add.s32 %r473, %r27, 1;
cvt.u64.u32 %rd60, %r473;
add.s64 %rd61, %rd60, %rd11;
shl.b64 %rd63, %rd61, 2;
add.s64 %rd64, %rd53, %rd63;
st.global.f32 [%rd64], %f34;

$L__BB82_14:
shl.b32 %r474, %r27, 1;
mov.u32 %r475, 1;
ld.shared.f32 %f156, [%r23+512];
add.ftz.f32 %f157, %f156, %f237;
mov.b32 %r476, %f157;
mov.u32 %r477, 31;
mov.u32 %r478, 16;
mov.u32 %r479, -1;
shfl.sync.bfly.b32 %r480|%p35, %r476, %r478, %r477, %r479;
mov.b32 %f158, %r480;
add.ftz.f32 %f159, %f157, %f158;
mov.b32 %r481, %f159;
mov.u32 %r482, 8;
shfl.sync.bfly.b32 %r483|%p36, %r481, %r482, %r477, %r479;
mov.b32 %f160, %r483;
add.ftz.f32 %f161, %f159, %f160;
mov.b32 %r484, %f161;
mov.u32 %r485, 4;
shfl.sync.bfly.b32 %r486|%p37, %r484, %r485, %r477, %r479;
mov.b32 %f162, %r486;
add.ftz.f32 %f163, %f161, %f162;
mov.b32 %r487, %f163;
mov.u32 %r488, 2;
shfl.sync.bfly.b32 %r489|%p38, %r487, %r488, %r477, %r479;
mov.b32 %f164, %r489;
add.ftz.f32 %f165, %f163, %f164;
mov.b32 %r490, %f165;
shfl.sync.bfly.b32 %r491|%p39, %r490, %r475, %r477, %r479;
mov.b32 %f166, %r491;
add.ftz.f32 %f35, %f165, %f166;
cvt.u64.u32 %rd65, %r474;
add.s64 %rd66, %rd65, %rd11;
shl.b64 %rd68, %rd66, 2;
add.s64 %rd13, %rd53, %rd68;
@%p14 bra $L__BB82_16;

st.global.f32 [%rd13], %f35;

$L__BB82_16:
ld.shared.f32 %f167, [%r23+640];
add.ftz.f32 %f168, %f167, %f236;
mov.b32 %r492, %f168;
shfl.sync.bfly.b32 %r496|%p41, %r492, %r478, %r477, %r479;
mov.b32 %f169, %r496;
add.ftz.f32 %f170, %f168, %f169;
mov.b32 %r497, %f170;
shfl.sync.bfly.b32 %r499|%p42, %r497, %r482, %r477, %r479;
mov.b32 %f171, %r499;
add.ftz.f32 %f172, %f170, %f171;
mov.b32 %r500, %f172;
shfl.sync.bfly.b32 %r502|%p43, %r500, %r485, %r477, %r479;
mov.b32 %f173, %r502;
add.ftz.f32 %f174, %f172, %f173;
mov.b32 %r503, %f174;
shfl.sync.bfly.b32 %r505|%p44, %r503, %r488, %r477, %r479;
mov.b32 %f175, %r505;
add.ftz.f32 %f176, %f174, %f175;
mov.b32 %r506, %f176;
shfl.sync.bfly.b32 %r508|%p45, %r506, %r475, %r477, %r479;
mov.b32 %f177, %r508;
add.ftz.f32 %f36, %f176, %f177;
@%p22 bra $L__BB82_18;

st.global.f32 [%rd13+4], %f36;

$L__BB82_18:
mul.lo.s32 %r24, %r27, 3;
ld.shared.f32 %f178, [%r23+768];
add.ftz.f32 %f179, %f178, %f235;
mov.b32 %r509, %f179;
mov.u32 %r510, 31;
mov.u32 %r511, 16;
mov.u32 %r512, -1;
shfl.sync.bfly.b32 %r513|%p47, %r509, %r511, %r510, %r512;
mov.b32 %f180, %r513;
add.ftz.f32 %f181, %f179, %f180;
mov.b32 %r514, %f181;
mov.u32 %r515, 8;
shfl.sync.bfly.b32 %r516|%p48, %r514, %r515, %r510, %r512;
mov.b32 %f182, %r516;
add.ftz.f32 %f183, %f181, %f182;
mov.b32 %r517, %f183;
mov.u32 %r518, 4;
shfl.sync.bfly.b32 %r519|%p49, %r517, %r518, %r510, %r512;
mov.b32 %f184, %r519;
add.ftz.f32 %f185, %f183, %f184;
mov.b32 %r520, %f185;
mov.u32 %r521, 2;
shfl.sync.bfly.b32 %r522|%p50, %r520, %r521, %r510, %r512;
mov.b32 %f186, %r522;
add.ftz.f32 %f187, %f185, %f186;
mov.b32 %r523, %f187;
mov.u32 %r524, 1;
shfl.sync.bfly.b32 %r525|%p51, %r523, %r524, %r510, %r512;
mov.b32 %f188, %r525;
add.ftz.f32 %f37, %f187, %f188;
@%p14 bra $L__BB82_20;

cvt.u64.u32 %rd69, %r24;
add.s64 %rd70, %rd69, %rd11;
shl.b64 %rd72, %rd70, 2;
add.s64 %rd73, %rd53, %rd72;
st.global.f32 [%rd73], %f37;

$L__BB82_20:
ld.shared.f32 %f189, [%r23+896];
add.ftz.f32 %f190, %f189, %f234;
mov.b32 %r526, %f190;
shfl.sync.bfly.b32 %r530|%p53, %r526, %r511, %r510, %r512;
mov.b32 %f191, %r530;
add.ftz.f32 %f192, %f190, %f191;
mov.b32 %r531, %f192;
shfl.sync.bfly.b32 %r533|%p54, %r531, %r515, %r510, %r512;
mov.b32 %f193, %r533;
add.ftz.f32 %f194, %f192, %f193;
mov.b32 %r534, %f194;
shfl.sync.bfly.b32 %r536|%p55, %r534, %r518, %r510, %r512;
mov.b32 %f195, %r536;
add.ftz.f32 %f196, %f194, %f195;
mov.b32 %r537, %f196;
shfl.sync.bfly.b32 %r539|%p56, %r537, %r521, %r510, %r512;
mov.b32 %f197, %r539;
add.ftz.f32 %f198, %f196, %f197;
mov.b32 %r540, %f198;
shfl.sync.bfly.b32 %r542|%p57, %r540, %r524, %r510, %r512;
mov.b32 %f199, %r542;
add.ftz.f32 %f38, %f198, %f199;
@%p22 bra $L__BB82_22;

add.s32 %r543, %r24, 1;
cvt.u64.u32 %rd74, %r543;
add.s64 %rd75, %rd74, %rd11;
shl.b64 %rd77, %rd75, 2;
add.s64 %rd78, %rd53, %rd77;
st.global.f32 [%rd78], %f38;

$L__BB82_22:
shl.b32 %r544, %r27, 2;
mov.u32 %r545, 2;
ld.shared.f32 %f200, [%r23+1024];
add.ftz.f32 %f201, %f200, %f233;
mov.b32 %r546, %f201;
mov.u32 %r547, 31;
mov.u32 %r548, 16;
mov.u32 %r549, -1;
shfl.sync.bfly.b32 %r550|%p59, %r546, %r548, %r547, %r549;
mov.b32 %f202, %r550;
add.ftz.f32 %f203, %f201, %f202;
mov.b32 %r551, %f203;
mov.u32 %r552, 8;
shfl.sync.bfly.b32 %r553|%p60, %r551, %r552, %r547, %r549;
mov.b32 %f204, %r553;
add.ftz.f32 %f205, %f203, %f204;
mov.b32 %r554, %f205;
mov.u32 %r555, 4;
shfl.sync.bfly.b32 %r556|%p61, %r554, %r555, %r547, %r549;
mov.b32 %f206, %r556;
add.ftz.f32 %f207, %f205, %f206;
mov.b32 %r557, %f207;
shfl.sync.bfly.b32 %r558|%p62, %r557, %r545, %r547, %r549;
mov.b32 %f208, %r558;
add.ftz.f32 %f209, %f207, %f208;
mov.b32 %r559, %f209;
mov.u32 %r560, 1;
shfl.sync.bfly.b32 %r561|%p63, %r559, %r560, %r547, %r549;
mov.b32 %f210, %r561;
add.ftz.f32 %f39, %f209, %f210;
cvt.u64.u32 %rd79, %r544;
add.s64 %rd80, %rd79, %rd11;
shl.b64 %rd82, %rd80, 2;
add.s64 %rd14, %rd53, %rd82;
@%p14 bra $L__BB82_24;

st.global.f32 [%rd14], %f39;

$L__BB82_24:
ld.shared.f32 %f211, [%r23+1152];
add.ftz.f32 %f212, %f211, %f232;
mov.b32 %r562, %f212;
shfl.sync.bfly.b32 %r566|%p65, %r562, %r548, %r547, %r549;
mov.b32 %f213, %r566;
add.ftz.f32 %f214, %f212, %f213;
mov.b32 %r567, %f214;
shfl.sync.bfly.b32 %r569|%p66, %r567, %r552, %r547, %r549;
mov.b32 %f215, %r569;
add.ftz.f32 %f216, %f214, %f215;
mov.b32 %r570, %f216;
shfl.sync.bfly.b32 %r572|%p67, %r570, %r555, %r547, %r549;
mov.b32 %f217, %r572;
add.ftz.f32 %f218, %f216, %f217;
mov.b32 %r573, %f218;
shfl.sync.bfly.b32 %r575|%p68, %r573, %r545, %r547, %r549;
mov.b32 %f219, %r575;
add.ftz.f32 %f220, %f218, %f219;
mov.b32 %r576, %f220;
shfl.sync.bfly.b32 %r578|%p69, %r576, %r560, %r547, %r549;
mov.b32 %f221, %r578;
add.ftz.f32 %f40, %f220, %f221;
@%p22 bra $L__BB82_26;

st.global.f32 [%rd14+4], %f40;

$L__BB82_26:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type7ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type7ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type7ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type7ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type7ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type7ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type7ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type7ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type7ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<83>;
.reg .f32 %f<290>;
.reg .b32 %r<678>;
.reg .b64 %rd<103>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type7ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

ld.param.u64 %rd14, [_Z13mul_mat_vec_qIL9ggml_type7ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type7ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type7ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type7ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type7ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type7ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type7ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type7ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type7ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type7ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type7ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type7ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type7ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type7ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type7ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type7ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type7ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r45, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r45, %r2;
mov.u32 %r46, %ctaid.x;
shl.b32 %r4, %r46, 1;
shr.u32 %r47, %r3, 31;
add.s32 %r48, %r3, %r47;
shr.s32 %r677, %r48, 1;
shr.u32 %r6, %r44, 5;
setp.ge.s32 %p3, %r677, %r6;
mov.f32 %f278, 0f00000000;
mov.f32 %f279, %f278;
mov.f32 %f280, %f278;
mov.f32 %f281, %f278;
mov.f32 %f282, %f278;
mov.f32 %f283, %f278;
mov.f32 %f284, %f278;
mov.f32 %f285, %f278;
mov.f32 %f286, %f278;
mov.f32 %f287, %f278;
mov.f32 %f288, %f278;
mov.f32 %f289, %f278;
@%p3 bra $L__BB83_3;

mov.u32 %r49, %ctaid.y;
mov.u32 %r50, %ctaid.z;
mul.lo.s32 %r51, %r49, %r36;
cvt.u64.u32 %rd17, %r51;
mul.lo.s32 %r52, %r50, %r42;
cvt.u64.u32 %rd18, %r52;
add.s64 %rd1, %rd18, %rd17;
mul.hi.u32 %r53, %r50, %r38;
add.s32 %r54, %r53, %r50;
shr.u32 %r55, %r54, %r39;
mul.lo.s32 %r56, %r55, %r41;
mul.hi.u32 %r57, %r49, %r32;
add.s32 %r58, %r57, %r49;
shr.u32 %r59, %r58, %r33;
mul.lo.s32 %r60, %r59, %r35;
cvta.to.global.u64 %rd19, %rd14;
and.b32 %r63, %r48, -2;
sub.s32 %r64, %r3, %r63;
shl.b32 %r65, %r64, 1;
cvt.s64.s32 %rd2, %r65;
shl.b32 %r7, %r64, 3;
or.b32 %r8, %r7, 4;
shl.b32 %r66, %r30, 1;
add.s32 %r676, %r677, %r66;
mad.lo.s32 %r675, %r30, 3, %r677;
shl.b32 %r67, %r30, 2;
add.s32 %r674, %r677, %r67;
mad.lo.s32 %r673, %r30, 5, %r677;
add.s32 %r672, %r677, %r30;
add.s32 %r68, %r677, %r56;
add.s32 %r69, %r68, %r60;
add.s32 %r70, %r4, 1;
mad.lo.s32 %r71, %r29, %r70, %r69;
mul.wide.s32 %rd20, %r71, 24;
add.s64 %rd21, %rd19, %rd20;
add.s64 %rd102, %rd21, 8;
mul.wide.s32 %rd4, %r65, 4;
mad.lo.s32 %r72, %r4, %r29, %r677;
add.s32 %r73, %r72, %r56;
add.s32 %r74, %r73, %r60;
mul.wide.s32 %rd22, %r74, 24;
add.s64 %rd23, %rd19, %rd22;
add.s64 %rd101, %rd23, 8;
cvta.to.global.u64 %rd6, %rd15;
mov.f32 %f278, 0f00000000;
shl.b64 %rd28, %rd2, 2;

$L__BB83_2:
cvt.u64.u32 %rd24, %r677;
add.s64 %rd25, %rd1, %rd24;
mul.lo.s64 %rd26, %rd25, 36;
add.s64 %rd27, %rd6, %rd26;
add.s64 %rd29, %rd27, %rd28;
add.s64 %rd30, %rd101, %rd4;
ld.global.nc.u32 %r327, [%rd101+-4];
shr.s32 %r328, %r327, %r7;
ld.global.nc.u32 %r77, [%rd29+4];
ld.global.nc.u32 %r81, [%rd29+20];
ld.global.nc.u32 %r329, [%rd30+4];
shr.s32 %r330, %r327, %r8;
ld.global.nc.u32 %r85, [%rd29+8];
ld.global.nc.u32 %r89, [%rd29+24];
ld.global.nc.u32 %r331, [%rd30];
and.b32 %r332, %r331, 252645135;
shl.b32 %r333, %r328, 4;
and.b32 %r334, %r333, 16;
or.b32 %r335, %r334, %r332;
shl.b32 %r336, %r328, 11;
and.b32 %r337, %r336, 4096;
or.b32 %r338, %r335, %r337;
shl.b32 %r339, %r328, 18;
and.b32 %r340, %r339, 1048576;
or.b32 %r341, %r338, %r340;
shl.b32 %r342, %r328, 25;
and.b32 %r343, %r342, 268435456;
or.b32 %r286, %r341, %r343;
mov.u32 %r309, 0;

dp4a.s32.s32 %r75, %r286, %r77, %r309;

shr.u32 %r344, %r331, 4;
and.b32 %r345, %r344, 252645135;
shr.u32 %r346, %r328, 12;
and.b32 %r347, %r346, 16;
or.b32 %r348, %r347, %r345;
shr.u32 %r349, %r328, 5;
and.b32 %r350, %r349, 4096;
or.b32 %r351, %r348, %r350;
shl.b32 %r352, %r328, 2;
and.b32 %r353, %r352, 1048576;
or.b32 %r354, %r351, %r353;
shl.b32 %r355, %r328, 9;
and.b32 %r356, %r355, 268435456;
or.b32 %r290, %r354, %r356;

dp4a.s32.s32 %r79, %r290, %r81, %r75;

and.b32 %r357, %r329, 252645135;
shl.b32 %r358, %r330, 4;
and.b32 %r359, %r358, 16;
or.b32 %r360, %r359, %r357;
shl.b32 %r361, %r330, 11;
and.b32 %r362, %r361, 4096;
or.b32 %r363, %r360, %r362;
shl.b32 %r364, %r330, 18;
and.b32 %r365, %r364, 1048576;
or.b32 %r366, %r363, %r365;
shl.b32 %r367, %r330, 25;
and.b32 %r368, %r367, 268435456;
or.b32 %r294, %r366, %r368;

dp4a.s32.s32 %r83, %r294, %r85, %r79;

shr.u32 %r369, %r329, 4;
and.b32 %r370, %r369, 252645135;
shr.u32 %r371, %r330, 12;
and.b32 %r372, %r371, 16;
or.b32 %r373, %r372, %r370;
shr.u32 %r374, %r330, 5;
and.b32 %r375, %r374, 4096;
or.b32 %r376, %r373, %r375;
shl.b32 %r377, %r330, 2;
and.b32 %r378, %r377, 1048576;
or.b32 %r379, %r376, %r378;
shl.b32 %r380, %r330, 9;
and.b32 %r381, %r380, 268435456;
or.b32 %r298, %r379, %r381;

dp4a.s32.s32 %r87, %r298, %r89, %r83;

ld.global.nc.u32 %r92, [%rd101+-8];
ld.global.nc.u32 %r93, [%rd27];

{mul.f16x2 %r91,%r92,%r93;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r91;
cvt.f32.f16 %f73, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r91;
cvt.f32.f16 %f74, high;}


cvt.rn.f32.s32 %f97, %r87;
mov.f32 %f98, 0f40000000;
div.approx.ftz.f32 %f99, %f74, %f98;
fma.rn.ftz.f32 %f100, %f73, %f97, %f99;
add.ftz.f32 %f289, %f100, %f289;
add.s64 %rd31, %rd102, %rd4;
ld.global.nc.u32 %r382, [%rd102+-4];
shr.s32 %r383, %r382, %r7;
ld.global.nc.u32 %r384, [%rd31+4];
shr.s32 %r385, %r382, %r8;
ld.global.nc.u32 %r386, [%rd31];
and.b32 %r387, %r386, 252645135;
shl.b32 %r388, %r383, 4;
and.b32 %r389, %r388, 16;
or.b32 %r390, %r389, %r387;
shl.b32 %r391, %r383, 11;
and.b32 %r392, %r391, 4096;
or.b32 %r393, %r390, %r392;
shl.b32 %r394, %r383, 18;
and.b32 %r395, %r394, 1048576;
or.b32 %r396, %r393, %r395;
shl.b32 %r397, %r383, 25;
and.b32 %r398, %r397, 268435456;
or.b32 %r307, %r396, %r398;

dp4a.s32.s32 %r96, %r307, %r77, %r309;

shr.u32 %r399, %r386, 4;
and.b32 %r400, %r399, 252645135;
shr.u32 %r401, %r383, 12;
and.b32 %r402, %r401, 16;
or.b32 %r403, %r402, %r400;
shr.u32 %r404, %r383, 5;
and.b32 %r405, %r404, 4096;
or.b32 %r406, %r403, %r405;
shl.b32 %r407, %r383, 2;
and.b32 %r408, %r407, 1048576;
or.b32 %r409, %r406, %r408;
shl.b32 %r410, %r383, 9;
and.b32 %r411, %r410, 268435456;
or.b32 %r311, %r409, %r411;

dp4a.s32.s32 %r100, %r311, %r81, %r96;

and.b32 %r412, %r384, 252645135;
shl.b32 %r413, %r385, 4;
and.b32 %r414, %r413, 16;
or.b32 %r415, %r414, %r412;
shl.b32 %r416, %r385, 11;
and.b32 %r417, %r416, 4096;
or.b32 %r418, %r415, %r417;
shl.b32 %r419, %r385, 18;
and.b32 %r420, %r419, 1048576;
or.b32 %r421, %r418, %r420;
shl.b32 %r422, %r385, 25;
and.b32 %r423, %r422, 268435456;
or.b32 %r315, %r421, %r423;

dp4a.s32.s32 %r104, %r315, %r85, %r100;

shr.u32 %r424, %r384, 4;
and.b32 %r425, %r424, 252645135;
shr.u32 %r426, %r385, 12;
and.b32 %r427, %r426, 16;
or.b32 %r428, %r427, %r425;
shr.u32 %r429, %r385, 5;
and.b32 %r430, %r429, 4096;
or.b32 %r431, %r428, %r430;
shl.b32 %r432, %r385, 2;
and.b32 %r433, %r432, 1048576;
or.b32 %r434, %r431, %r433;
shl.b32 %r435, %r385, 9;
and.b32 %r436, %r435, 268435456;
or.b32 %r319, %r434, %r436;

dp4a.s32.s32 %r108, %r319, %r89, %r104;

ld.global.nc.u32 %r113, [%rd102+-8];

{mul.f16x2 %r112,%r113,%r93;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r112;
cvt.f32.f16 %f75, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r112;
cvt.f32.f16 %f76, high;}


cvt.rn.f32.s32 %f101, %r108;
div.approx.ftz.f32 %f102, %f76, %f98;
fma.rn.ftz.f32 %f103, %f75, %f101, %f102;
add.ftz.f32 %f288, %f103, %f288;
cvt.u64.u32 %rd32, %r672;
add.s64 %rd33, %rd1, %rd32;
mul.lo.s64 %rd34, %rd33, 36;
add.s64 %rd35, %rd6, %rd34;
add.s64 %rd36, %rd35, %rd28;
ld.global.nc.u32 %r119, [%rd36+4];
ld.global.nc.u32 %r123, [%rd36+20];
ld.global.nc.u32 %r127, [%rd36+8];
ld.global.nc.u32 %r131, [%rd36+24];

dp4a.s32.s32 %r117, %r286, %r119, %r309;


dp4a.s32.s32 %r121, %r290, %r123, %r117;


dp4a.s32.s32 %r125, %r294, %r127, %r121;


dp4a.s32.s32 %r129, %r298, %r131, %r125;

ld.global.nc.u32 %r135, [%rd35];

{mul.f16x2 %r133,%r92,%r135;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r133;
cvt.f32.f16 %f77, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r133;
cvt.f32.f16 %f78, high;}


cvt.rn.f32.s32 %f104, %r129;
div.approx.ftz.f32 %f105, %f78, %f98;
fma.rn.ftz.f32 %f106, %f77, %f104, %f105;
add.ftz.f32 %f287, %f106, %f287;

dp4a.s32.s32 %r138, %r307, %r119, %r309;


dp4a.s32.s32 %r142, %r311, %r123, %r138;


dp4a.s32.s32 %r146, %r315, %r127, %r142;


dp4a.s32.s32 %r150, %r319, %r131, %r146;


{mul.f16x2 %r154,%r113,%r135;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r154;
cvt.f32.f16 %f79, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r154;
cvt.f32.f16 %f80, high;}


cvt.rn.f32.s32 %f107, %r150;
div.approx.ftz.f32 %f108, %f80, %f98;
fma.rn.ftz.f32 %f109, %f79, %f107, %f108;
add.ftz.f32 %f286, %f109, %f286;
cvt.u64.u32 %rd37, %r676;
add.s64 %rd38, %rd1, %rd37;
mul.lo.s64 %rd39, %rd38, 36;
add.s64 %rd40, %rd6, %rd39;
add.s64 %rd41, %rd40, %rd28;
ld.global.nc.u32 %r161, [%rd41+4];
ld.global.nc.u32 %r165, [%rd41+20];
ld.global.nc.u32 %r169, [%rd41+8];
ld.global.nc.u32 %r173, [%rd41+24];

dp4a.s32.s32 %r159, %r286, %r161, %r309;


dp4a.s32.s32 %r163, %r290, %r165, %r159;


dp4a.s32.s32 %r167, %r294, %r169, %r163;


dp4a.s32.s32 %r171, %r298, %r173, %r167;

ld.global.nc.u32 %r177, [%rd40];

{mul.f16x2 %r175,%r92,%r177;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r175;
cvt.f32.f16 %f81, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r175;
cvt.f32.f16 %f82, high;}


cvt.rn.f32.s32 %f110, %r171;
div.approx.ftz.f32 %f111, %f82, %f98;
fma.rn.ftz.f32 %f112, %f81, %f110, %f111;
add.ftz.f32 %f285, %f112, %f285;

dp4a.s32.s32 %r180, %r307, %r161, %r309;


dp4a.s32.s32 %r184, %r311, %r165, %r180;


dp4a.s32.s32 %r188, %r315, %r169, %r184;


dp4a.s32.s32 %r192, %r319, %r173, %r188;


{mul.f16x2 %r196,%r113,%r177;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r196;
cvt.f32.f16 %f83, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r196;
cvt.f32.f16 %f84, high;}


cvt.rn.f32.s32 %f113, %r192;
div.approx.ftz.f32 %f114, %f84, %f98;
fma.rn.ftz.f32 %f115, %f83, %f113, %f114;
add.ftz.f32 %f284, %f115, %f284;
cvt.u64.u32 %rd42, %r675;
add.s64 %rd43, %rd1, %rd42;
mul.lo.s64 %rd44, %rd43, 36;
add.s64 %rd45, %rd6, %rd44;
add.s64 %rd46, %rd45, %rd28;
ld.global.nc.u32 %r203, [%rd46+4];
ld.global.nc.u32 %r207, [%rd46+20];
ld.global.nc.u32 %r211, [%rd46+8];
ld.global.nc.u32 %r215, [%rd46+24];

dp4a.s32.s32 %r201, %r286, %r203, %r309;


dp4a.s32.s32 %r205, %r290, %r207, %r201;


dp4a.s32.s32 %r209, %r294, %r211, %r205;


dp4a.s32.s32 %r213, %r298, %r215, %r209;

ld.global.nc.u32 %r219, [%rd45];

{mul.f16x2 %r217,%r92,%r219;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r217;
cvt.f32.f16 %f85, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r217;
cvt.f32.f16 %f86, high;}


cvt.rn.f32.s32 %f116, %r213;
div.approx.ftz.f32 %f117, %f86, %f98;
fma.rn.ftz.f32 %f118, %f85, %f116, %f117;
add.ftz.f32 %f283, %f118, %f283;

dp4a.s32.s32 %r222, %r307, %r203, %r309;


dp4a.s32.s32 %r226, %r311, %r207, %r222;


dp4a.s32.s32 %r230, %r315, %r211, %r226;


dp4a.s32.s32 %r234, %r319, %r215, %r230;


{mul.f16x2 %r238,%r113,%r219;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r238;
cvt.f32.f16 %f87, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r238;
cvt.f32.f16 %f88, high;}


cvt.rn.f32.s32 %f119, %r234;
div.approx.ftz.f32 %f120, %f88, %f98;
fma.rn.ftz.f32 %f121, %f87, %f119, %f120;
add.ftz.f32 %f282, %f121, %f282;
cvt.u64.u32 %rd47, %r674;
add.s64 %rd48, %rd1, %rd47;
mul.lo.s64 %rd49, %rd48, 36;
add.s64 %rd50, %rd6, %rd49;
add.s64 %rd51, %rd50, %rd28;
ld.global.nc.u32 %r245, [%rd51+4];
ld.global.nc.u32 %r249, [%rd51+20];
ld.global.nc.u32 %r253, [%rd51+8];
ld.global.nc.u32 %r257, [%rd51+24];

dp4a.s32.s32 %r243, %r286, %r245, %r309;


dp4a.s32.s32 %r247, %r290, %r249, %r243;


dp4a.s32.s32 %r251, %r294, %r253, %r247;


dp4a.s32.s32 %r255, %r298, %r257, %r251;

ld.global.nc.u32 %r261, [%rd50];

{mul.f16x2 %r259,%r92,%r261;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r259;
cvt.f32.f16 %f89, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r259;
cvt.f32.f16 %f90, high;}


cvt.rn.f32.s32 %f122, %r255;
div.approx.ftz.f32 %f123, %f90, %f98;
fma.rn.ftz.f32 %f124, %f89, %f122, %f123;
add.ftz.f32 %f281, %f124, %f281;

dp4a.s32.s32 %r264, %r307, %r245, %r309;


dp4a.s32.s32 %r268, %r311, %r249, %r264;


dp4a.s32.s32 %r272, %r315, %r253, %r268;


dp4a.s32.s32 %r276, %r319, %r257, %r272;


{mul.f16x2 %r280,%r113,%r261;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r280;
cvt.f32.f16 %f91, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r280;
cvt.f32.f16 %f92, high;}


cvt.rn.f32.s32 %f125, %r276;
div.approx.ftz.f32 %f126, %f92, %f98;
fma.rn.ftz.f32 %f127, %f91, %f125, %f126;
add.ftz.f32 %f280, %f127, %f280;
cvt.u64.u32 %rd52, %r673;
add.s64 %rd53, %rd1, %rd52;
mul.lo.s64 %rd54, %rd53, 36;
add.s64 %rd55, %rd6, %rd54;
add.s64 %rd56, %rd55, %rd28;
ld.global.nc.u32 %r287, [%rd56+4];
ld.global.nc.u32 %r291, [%rd56+20];
ld.global.nc.u32 %r295, [%rd56+8];
ld.global.nc.u32 %r299, [%rd56+24];

dp4a.s32.s32 %r285, %r286, %r287, %r309;


dp4a.s32.s32 %r289, %r290, %r291, %r285;


dp4a.s32.s32 %r293, %r294, %r295, %r289;


dp4a.s32.s32 %r297, %r298, %r299, %r293;

ld.global.nc.u32 %r303, [%rd55];

{mul.f16x2 %r301,%r92,%r303;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r301;
cvt.f32.f16 %f93, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r301;
cvt.f32.f16 %f94, high;}


cvt.rn.f32.s32 %f128, %r297;
div.approx.ftz.f32 %f129, %f94, %f98;
fma.rn.ftz.f32 %f130, %f93, %f128, %f129;
add.ftz.f32 %f279, %f130, %f279;

dp4a.s32.s32 %r306, %r307, %r287, %r309;


dp4a.s32.s32 %r310, %r311, %r291, %r306;


dp4a.s32.s32 %r314, %r315, %r295, %r310;


dp4a.s32.s32 %r318, %r319, %r299, %r314;


{mul.f16x2 %r322,%r113,%r303;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r322;
cvt.f32.f16 %f95, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r322;
cvt.f32.f16 %f96, high;}


cvt.rn.f32.s32 %f131, %r318;
div.approx.ftz.f32 %f132, %f96, %f98;
fma.rn.ftz.f32 %f133, %f95, %f131, %f132;
add.ftz.f32 %f278, %f133, %f278;
add.s32 %r676, %r676, 32;
add.s32 %r675, %r675, 32;
add.s32 %r674, %r674, 32;
add.s32 %r673, %r673, 32;
add.s32 %r672, %r672, 32;
add.s64 %rd102, %rd102, 768;
add.s64 %rd101, %rd101, 768;
add.s32 %r677, %r677, 32;
setp.lt.s32 %p4, %r677, %r6;
@%p4 bra $L__BB83_2;

$L__BB83_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB83_5;

mov.u32 %r437, _ZZ13mul_mat_vec_qIL9ggml_type7ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r438, %r1, 1536, %r437;
shl.b32 %r439, %r2, 2;
add.s32 %r440, %r438, %r439;
st.shared.f32 [%r440+-1536], %f289;
st.shared.f32 [%r440+-1408], %f288;
st.shared.f32 [%r440+-1280], %f287;
st.shared.f32 [%r440+-1152], %f286;
st.shared.f32 [%r440+-1024], %f285;
st.shared.f32 [%r440+-896], %f284;
st.shared.f32 [%r440+-768], %f283;
st.shared.f32 [%r440+-640], %f282;
st.shared.f32 [%r440+-512], %f281;
st.shared.f32 [%r440+-384], %f280;
st.shared.f32 [%r440+-256], %f279;
st.shared.f32 [%r440+-128], %f278;

$L__BB83_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB83_30;

mov.u32 %r441, %ctaid.z;
mov.u32 %r442, %ctaid.y;
mad.lo.s32 %r443, %r442, %r37, %r4;
mad.lo.s32 %r444, %r441, %r43, %r443;
cvt.u64.u32 %rd11, %r444;
shl.b32 %r445, %r2, 2;
mov.u32 %r446, _ZZ13mul_mat_vec_qIL9ggml_type7ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r26, %r446, %r445;
mov.u32 %r447, 2;
ld.shared.f32 %f134, [%r26];
add.ftz.f32 %f135, %f134, %f289;
mov.b32 %r448, %f135;
mov.u32 %r449, 31;
mov.u32 %r450, 16;
mov.u32 %r451, -1;
shfl.sync.bfly.b32 %r452|%p7, %r448, %r450, %r449, %r451;
mov.b32 %f136, %r452;
add.ftz.f32 %f137, %f135, %f136;
mov.b32 %r453, %f137;
mov.u32 %r454, 8;
shfl.sync.bfly.b32 %r455|%p8, %r453, %r454, %r449, %r451;
mov.b32 %f138, %r455;
add.ftz.f32 %f139, %f137, %f138;
mov.b32 %r456, %f139;
mov.u32 %r457, 4;
shfl.sync.bfly.b32 %r458|%p9, %r456, %r457, %r449, %r451;
mov.b32 %f140, %r458;
add.ftz.f32 %f141, %f139, %f140;
mov.b32 %r459, %f141;
shfl.sync.bfly.b32 %r460|%p10, %r459, %r447, %r449, %r451;
mov.b32 %f142, %r460;
add.ftz.f32 %f143, %f141, %f142;
mov.b32 %r461, %f143;
mov.u32 %r462, 1;
shfl.sync.bfly.b32 %r463|%p11, %r461, %r462, %r449, %r451;
mov.b32 %f144, %r463;
add.ftz.f32 %f37, %f143, %f144;
setp.lt.u32 %p12, %r4, %r31;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
not.pred %p14, %p1;
@%p14 bra $L__BB83_8;

cvta.to.global.u64 %rd57, %rd16;
mul.wide.u32 %rd58, %r444, 4;
add.s64 %rd59, %rd57, %rd58;
st.global.f32 [%rd59], %f37;

$L__BB83_8:
ld.shared.f32 %f145, [%r26+128];
add.ftz.f32 %f146, %f145, %f288;
mov.b32 %r470, %f146;
shfl.sync.bfly.b32 %r474|%p15, %r470, %r450, %r449, %r451;
mov.b32 %f147, %r474;
add.ftz.f32 %f148, %f146, %f147;
mov.b32 %r475, %f148;
shfl.sync.bfly.b32 %r477|%p16, %r475, %r454, %r449, %r451;
mov.b32 %f149, %r477;
add.ftz.f32 %f150, %f148, %f149;
mov.b32 %r478, %f150;
shfl.sync.bfly.b32 %r480|%p17, %r478, %r457, %r449, %r451;
mov.b32 %f151, %r480;
add.ftz.f32 %f152, %f150, %f151;
mov.b32 %r481, %f152;
shfl.sync.bfly.b32 %r483|%p18, %r481, %r447, %r449, %r451;
mov.b32 %f153, %r483;
add.ftz.f32 %f154, %f152, %f153;
mov.b32 %r484, %f154;
shfl.sync.bfly.b32 %r486|%p19, %r484, %r462, %r449, %r451;
mov.b32 %f155, %r486;
add.ftz.f32 %f38, %f154, %f155;
setp.eq.s32 %p20, %r2, 1;
or.b32 %r490, %r4, 1;
setp.lt.u32 %p21, %r490, %r31;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB83_10;

cvta.to.global.u64 %rd60, %rd16;
mul.wide.u32 %rd61, %r444, 4;
add.s64 %rd62, %rd60, %rd61;
st.global.f32 [%rd62+4], %f38;

$L__BB83_10:
ld.shared.f32 %f156, [%r26+256];
add.ftz.f32 %f157, %f156, %f287;
mov.b32 %r497, %f157;
mov.u32 %r498, 31;
mov.u32 %r499, 16;
mov.u32 %r500, -1;
shfl.sync.bfly.b32 %r501|%p23, %r497, %r499, %r498, %r500;
mov.b32 %f158, %r501;
add.ftz.f32 %f159, %f157, %f158;
mov.b32 %r502, %f159;
mov.u32 %r503, 8;
shfl.sync.bfly.b32 %r504|%p24, %r502, %r503, %r498, %r500;
mov.b32 %f160, %r504;
add.ftz.f32 %f161, %f159, %f160;
mov.b32 %r505, %f161;
mov.u32 %r506, 4;
shfl.sync.bfly.b32 %r507|%p25, %r505, %r506, %r498, %r500;
mov.b32 %f162, %r507;
add.ftz.f32 %f163, %f161, %f162;
mov.b32 %r508, %f163;
mov.u32 %r509, 2;
shfl.sync.bfly.b32 %r510|%p26, %r508, %r509, %r498, %r500;
mov.b32 %f164, %r510;
add.ftz.f32 %f165, %f163, %f164;
mov.b32 %r511, %f165;
mov.u32 %r512, 1;
shfl.sync.bfly.b32 %r513|%p27, %r511, %r512, %r498, %r500;
mov.b32 %f166, %r513;
add.ftz.f32 %f39, %f165, %f166;
@%p14 bra $L__BB83_12;

cvt.u64.u32 %rd63, %r31;
add.s64 %rd64, %rd63, %rd11;
cvta.to.global.u64 %rd65, %rd16;
shl.b64 %rd66, %rd64, 2;
add.s64 %rd67, %rd65, %rd66;
st.global.f32 [%rd67], %f39;

$L__BB83_12:
ld.shared.f32 %f167, [%r26+384];
add.ftz.f32 %f168, %f167, %f286;
mov.b32 %r514, %f168;
shfl.sync.bfly.b32 %r518|%p29, %r514, %r499, %r498, %r500;
mov.b32 %f169, %r518;
add.ftz.f32 %f170, %f168, %f169;
mov.b32 %r519, %f170;
shfl.sync.bfly.b32 %r521|%p30, %r519, %r503, %r498, %r500;
mov.b32 %f171, %r521;
add.ftz.f32 %f172, %f170, %f171;
mov.b32 %r522, %f172;
shfl.sync.bfly.b32 %r524|%p31, %r522, %r506, %r498, %r500;
mov.b32 %f173, %r524;
add.ftz.f32 %f174, %f172, %f173;
mov.b32 %r525, %f174;
shfl.sync.bfly.b32 %r527|%p32, %r525, %r509, %r498, %r500;
mov.b32 %f175, %r527;
add.ftz.f32 %f176, %f174, %f175;
mov.b32 %r528, %f176;
shfl.sync.bfly.b32 %r530|%p33, %r528, %r512, %r498, %r500;
mov.b32 %f177, %r530;
add.ftz.f32 %f40, %f176, %f177;
@%p22 bra $L__BB83_14;

add.s32 %r531, %r31, 1;
cvt.u64.u32 %rd68, %r531;
add.s64 %rd69, %rd68, %rd11;
cvta.to.global.u64 %rd70, %rd16;
shl.b64 %rd71, %rd69, 2;
add.s64 %rd72, %rd70, %rd71;
st.global.f32 [%rd72], %f40;

$L__BB83_14:
shl.b32 %r532, %r31, 1;
mov.u32 %r533, 1;
ld.shared.f32 %f178, [%r26+512];
add.ftz.f32 %f179, %f178, %f285;
mov.b32 %r534, %f179;
mov.u32 %r535, 31;
mov.u32 %r536, 16;
mov.u32 %r537, -1;
shfl.sync.bfly.b32 %r538|%p35, %r534, %r536, %r535, %r537;
mov.b32 %f180, %r538;
add.ftz.f32 %f181, %f179, %f180;
mov.b32 %r539, %f181;
mov.u32 %r540, 8;
shfl.sync.bfly.b32 %r541|%p36, %r539, %r540, %r535, %r537;
mov.b32 %f182, %r541;
add.ftz.f32 %f183, %f181, %f182;
mov.b32 %r542, %f183;
mov.u32 %r543, 4;
shfl.sync.bfly.b32 %r544|%p37, %r542, %r543, %r535, %r537;
mov.b32 %f184, %r544;
add.ftz.f32 %f185, %f183, %f184;
mov.b32 %r545, %f185;
mov.u32 %r546, 2;
shfl.sync.bfly.b32 %r547|%p38, %r545, %r546, %r535, %r537;
mov.b32 %f186, %r547;
add.ftz.f32 %f187, %f185, %f186;
mov.b32 %r548, %f187;
shfl.sync.bfly.b32 %r549|%p39, %r548, %r533, %r535, %r537;
mov.b32 %f188, %r549;
add.ftz.f32 %f41, %f187, %f188;
cvt.u64.u32 %rd73, %r532;
add.s64 %rd74, %rd73, %rd11;
cvta.to.global.u64 %rd75, %rd16;
shl.b64 %rd76, %rd74, 2;
add.s64 %rd12, %rd75, %rd76;
@%p14 bra $L__BB83_16;

st.global.f32 [%rd12], %f41;

$L__BB83_16:
ld.shared.f32 %f189, [%r26+640];
add.ftz.f32 %f190, %f189, %f284;
mov.b32 %r550, %f190;
shfl.sync.bfly.b32 %r554|%p41, %r550, %r536, %r535, %r537;
mov.b32 %f191, %r554;
add.ftz.f32 %f192, %f190, %f191;
mov.b32 %r555, %f192;
shfl.sync.bfly.b32 %r557|%p42, %r555, %r540, %r535, %r537;
mov.b32 %f193, %r557;
add.ftz.f32 %f194, %f192, %f193;
mov.b32 %r558, %f194;
shfl.sync.bfly.b32 %r560|%p43, %r558, %r543, %r535, %r537;
mov.b32 %f195, %r560;
add.ftz.f32 %f196, %f194, %f195;
mov.b32 %r561, %f196;
shfl.sync.bfly.b32 %r563|%p44, %r561, %r546, %r535, %r537;
mov.b32 %f197, %r563;
add.ftz.f32 %f198, %f196, %f197;
mov.b32 %r564, %f198;
shfl.sync.bfly.b32 %r566|%p45, %r564, %r533, %r535, %r537;
mov.b32 %f199, %r566;
add.ftz.f32 %f42, %f198, %f199;
@%p22 bra $L__BB83_18;

st.global.f32 [%rd12+4], %f42;

$L__BB83_18:
mul.lo.s32 %r27, %r31, 3;
ld.shared.f32 %f200, [%r26+768];
add.ftz.f32 %f201, %f200, %f283;
mov.b32 %r567, %f201;
mov.u32 %r568, 31;
mov.u32 %r569, 16;
mov.u32 %r570, -1;
shfl.sync.bfly.b32 %r571|%p47, %r567, %r569, %r568, %r570;
mov.b32 %f202, %r571;
add.ftz.f32 %f203, %f201, %f202;
mov.b32 %r572, %f203;
mov.u32 %r573, 8;
shfl.sync.bfly.b32 %r574|%p48, %r572, %r573, %r568, %r570;
mov.b32 %f204, %r574;
add.ftz.f32 %f205, %f203, %f204;
mov.b32 %r575, %f205;
mov.u32 %r576, 4;
shfl.sync.bfly.b32 %r577|%p49, %r575, %r576, %r568, %r570;
mov.b32 %f206, %r577;
add.ftz.f32 %f207, %f205, %f206;
mov.b32 %r578, %f207;
mov.u32 %r579, 2;
shfl.sync.bfly.b32 %r580|%p50, %r578, %r579, %r568, %r570;
mov.b32 %f208, %r580;
add.ftz.f32 %f209, %f207, %f208;
mov.b32 %r581, %f209;
mov.u32 %r582, 1;
shfl.sync.bfly.b32 %r583|%p51, %r581, %r582, %r568, %r570;
mov.b32 %f210, %r583;
add.ftz.f32 %f43, %f209, %f210;
@%p14 bra $L__BB83_20;

cvt.u64.u32 %rd77, %r27;
add.s64 %rd78, %rd77, %rd11;
shl.b64 %rd80, %rd78, 2;
add.s64 %rd81, %rd75, %rd80;
st.global.f32 [%rd81], %f43;

$L__BB83_20:
ld.shared.f32 %f211, [%r26+896];
add.ftz.f32 %f212, %f211, %f282;
mov.b32 %r584, %f212;
shfl.sync.bfly.b32 %r588|%p53, %r584, %r569, %r568, %r570;
mov.b32 %f213, %r588;
add.ftz.f32 %f214, %f212, %f213;
mov.b32 %r589, %f214;
shfl.sync.bfly.b32 %r591|%p54, %r589, %r573, %r568, %r570;
mov.b32 %f215, %r591;
add.ftz.f32 %f216, %f214, %f215;
mov.b32 %r592, %f216;
shfl.sync.bfly.b32 %r594|%p55, %r592, %r576, %r568, %r570;
mov.b32 %f217, %r594;
add.ftz.f32 %f218, %f216, %f217;
mov.b32 %r595, %f218;
shfl.sync.bfly.b32 %r597|%p56, %r595, %r579, %r568, %r570;
mov.b32 %f219, %r597;
add.ftz.f32 %f220, %f218, %f219;
mov.b32 %r598, %f220;
shfl.sync.bfly.b32 %r600|%p57, %r598, %r582, %r568, %r570;
mov.b32 %f221, %r600;
add.ftz.f32 %f44, %f220, %f221;
@%p22 bra $L__BB83_22;

add.s32 %r601, %r27, 1;
cvt.u64.u32 %rd82, %r601;
add.s64 %rd83, %rd82, %rd11;
shl.b64 %rd85, %rd83, 2;
add.s64 %rd86, %rd75, %rd85;
st.global.f32 [%rd86], %f44;

$L__BB83_22:
shl.b32 %r602, %r31, 2;
mov.u32 %r603, 2;
ld.shared.f32 %f222, [%r26+1024];
add.ftz.f32 %f223, %f222, %f281;
mov.b32 %r604, %f223;
mov.u32 %r605, 31;
mov.u32 %r606, 16;
mov.u32 %r607, -1;
shfl.sync.bfly.b32 %r608|%p59, %r604, %r606, %r605, %r607;
mov.b32 %f224, %r608;
add.ftz.f32 %f225, %f223, %f224;
mov.b32 %r609, %f225;
mov.u32 %r610, 8;
shfl.sync.bfly.b32 %r611|%p60, %r609, %r610, %r605, %r607;
mov.b32 %f226, %r611;
add.ftz.f32 %f227, %f225, %f226;
mov.b32 %r612, %f227;
mov.u32 %r613, 4;
shfl.sync.bfly.b32 %r614|%p61, %r612, %r613, %r605, %r607;
mov.b32 %f228, %r614;
add.ftz.f32 %f229, %f227, %f228;
mov.b32 %r615, %f229;
shfl.sync.bfly.b32 %r616|%p62, %r615, %r603, %r605, %r607;
mov.b32 %f230, %r616;
add.ftz.f32 %f231, %f229, %f230;
mov.b32 %r617, %f231;
mov.u32 %r618, 1;
shfl.sync.bfly.b32 %r619|%p63, %r617, %r618, %r605, %r607;
mov.b32 %f232, %r619;
add.ftz.f32 %f45, %f231, %f232;
cvt.u64.u32 %rd87, %r602;
add.s64 %rd88, %rd87, %rd11;
shl.b64 %rd90, %rd88, 2;
add.s64 %rd13, %rd75, %rd90;
@%p14 bra $L__BB83_24;

st.global.f32 [%rd13], %f45;

$L__BB83_24:
ld.shared.f32 %f233, [%r26+1152];
add.ftz.f32 %f234, %f233, %f280;
mov.b32 %r620, %f234;
shfl.sync.bfly.b32 %r624|%p65, %r620, %r606, %r605, %r607;
mov.b32 %f235, %r624;
add.ftz.f32 %f236, %f234, %f235;
mov.b32 %r625, %f236;
shfl.sync.bfly.b32 %r627|%p66, %r625, %r610, %r605, %r607;
mov.b32 %f237, %r627;
add.ftz.f32 %f238, %f236, %f237;
mov.b32 %r628, %f238;
shfl.sync.bfly.b32 %r630|%p67, %r628, %r613, %r605, %r607;
mov.b32 %f239, %r630;
add.ftz.f32 %f240, %f238, %f239;
mov.b32 %r631, %f240;
shfl.sync.bfly.b32 %r633|%p68, %r631, %r603, %r605, %r607;
mov.b32 %f241, %r633;
add.ftz.f32 %f242, %f240, %f241;
mov.b32 %r634, %f242;
shfl.sync.bfly.b32 %r636|%p69, %r634, %r618, %r605, %r607;
mov.b32 %f243, %r636;
add.ftz.f32 %f46, %f242, %f243;
@%p22 bra $L__BB83_26;

st.global.f32 [%rd13+4], %f46;

$L__BB83_26:
mul.lo.s32 %r28, %r31, 5;
ld.shared.f32 %f244, [%r26+1280];
add.ftz.f32 %f245, %f244, %f279;
mov.b32 %r637, %f245;
mov.u32 %r638, 31;
mov.u32 %r639, 16;
mov.u32 %r640, -1;
shfl.sync.bfly.b32 %r641|%p71, %r637, %r639, %r638, %r640;
mov.b32 %f246, %r641;
add.ftz.f32 %f247, %f245, %f246;
mov.b32 %r642, %f247;
mov.u32 %r643, 8;
shfl.sync.bfly.b32 %r644|%p72, %r642, %r643, %r638, %r640;
mov.b32 %f248, %r644;
add.ftz.f32 %f249, %f247, %f248;
mov.b32 %r645, %f249;
mov.u32 %r646, 4;
shfl.sync.bfly.b32 %r647|%p73, %r645, %r646, %r638, %r640;
mov.b32 %f250, %r647;
add.ftz.f32 %f251, %f249, %f250;
mov.b32 %r648, %f251;
mov.u32 %r649, 2;
shfl.sync.bfly.b32 %r650|%p74, %r648, %r649, %r638, %r640;
mov.b32 %f252, %r650;
add.ftz.f32 %f253, %f251, %f252;
mov.b32 %r651, %f253;
mov.u32 %r652, 1;
shfl.sync.bfly.b32 %r653|%p75, %r651, %r652, %r638, %r640;
mov.b32 %f254, %r653;
add.ftz.f32 %f47, %f253, %f254;
@%p14 bra $L__BB83_28;

cvt.u64.u32 %rd91, %r28;
add.s64 %rd92, %rd91, %rd11;
shl.b64 %rd94, %rd92, 2;
add.s64 %rd95, %rd75, %rd94;
st.global.f32 [%rd95], %f47;

$L__BB83_28:
ld.shared.f32 %f255, [%r26+1408];
add.ftz.f32 %f256, %f255, %f278;
mov.b32 %r654, %f256;
shfl.sync.bfly.b32 %r658|%p77, %r654, %r639, %r638, %r640;
mov.b32 %f257, %r658;
add.ftz.f32 %f258, %f256, %f257;
mov.b32 %r659, %f258;
shfl.sync.bfly.b32 %r661|%p78, %r659, %r643, %r638, %r640;
mov.b32 %f259, %r661;
add.ftz.f32 %f260, %f258, %f259;
mov.b32 %r662, %f260;
shfl.sync.bfly.b32 %r664|%p79, %r662, %r646, %r638, %r640;
mov.b32 %f261, %r664;
add.ftz.f32 %f262, %f260, %f261;
mov.b32 %r665, %f262;
shfl.sync.bfly.b32 %r667|%p80, %r665, %r649, %r638, %r640;
mov.b32 %f263, %r667;
add.ftz.f32 %f264, %f262, %f263;
mov.b32 %r668, %f264;
shfl.sync.bfly.b32 %r670|%p81, %r668, %r652, %r638, %r640;
mov.b32 %f265, %r670;
add.ftz.f32 %f48, %f264, %f265;
@%p22 bra $L__BB83_30;

add.s32 %r671, %r28, 1;
cvt.u64.u32 %rd96, %r671;
add.s64 %rd97, %rd96, %rd11;
shl.b64 %rd99, %rd97, 2;
add.s64 %rd100, %rd75, %rd99;
st.global.f32 [%rd100], %f48;

$L__BB83_30:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type7ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type7ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type7ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type7ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type7ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type7ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type7ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type7ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type7ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<95>;
.reg .f32 %f<338>;
.reg .b32 %r<747>;
.reg .b64 %rd<110>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type7ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1792];

ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type7ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type7ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type7ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r47, [_Z13mul_mat_vec_qIL9ggml_type7ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type7ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type7ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type7ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type7ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type7ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type7ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type7ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r45, [_Z13mul_mat_vec_qIL9ggml_type7ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r46, [_Z13mul_mat_vec_qIL9ggml_type7ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type7ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type7ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type7ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type7ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r48, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r48, %r2;
mov.u32 %r49, %ctaid.x;
shl.b32 %r4, %r49, 1;
shr.u32 %r50, %r3, 31;
add.s32 %r51, %r3, %r50;
shr.s32 %r746, %r51, 1;
shr.u32 %r6, %r47, 5;
setp.ge.s32 %p3, %r746, %r6;
mov.f32 %f324, 0f00000000;
mov.f32 %f325, %f324;
mov.f32 %f326, %f324;
mov.f32 %f327, %f324;
mov.f32 %f328, %f324;
mov.f32 %f329, %f324;
mov.f32 %f330, %f324;
mov.f32 %f331, %f324;
mov.f32 %f332, %f324;
mov.f32 %f333, %f324;
mov.f32 %f334, %f324;
mov.f32 %f335, %f324;
mov.f32 %f336, %f324;
mov.f32 %f337, %f324;
@%p3 bra $L__BB84_3;

mov.u32 %r52, %ctaid.y;
mov.u32 %r53, %ctaid.z;
mul.lo.s32 %r54, %r52, %r39;
cvt.u64.u32 %rd19, %r54;
mul.lo.s32 %r55, %r53, %r45;
cvt.u64.u32 %rd20, %r55;
add.s64 %rd1, %rd20, %rd19;
mul.hi.u32 %r56, %r53, %r41;
add.s32 %r57, %r56, %r53;
shr.u32 %r58, %r57, %r42;
mul.lo.s32 %r59, %r58, %r44;
mul.hi.u32 %r60, %r52, %r35;
add.s32 %r61, %r60, %r52;
shr.u32 %r62, %r61, %r36;
mul.lo.s32 %r63, %r62, %r38;
cvta.to.global.u64 %rd21, %rd16;
and.b32 %r66, %r51, -2;
sub.s32 %r67, %r3, %r66;
shl.b32 %r68, %r67, 1;
cvt.s64.s32 %rd2, %r68;
shl.b32 %r7, %r67, 3;
or.b32 %r8, %r7, 4;
shl.b32 %r69, %r33, 1;
add.s32 %r745, %r746, %r69;
mad.lo.s32 %r744, %r33, 3, %r746;
shl.b32 %r70, %r33, 2;
add.s32 %r743, %r746, %r70;
mad.lo.s32 %r742, %r33, 5, %r746;
mad.lo.s32 %r741, %r33, 6, %r746;
add.s32 %r740, %r746, %r33;
add.s32 %r71, %r746, %r59;
add.s32 %r72, %r71, %r63;
add.s32 %r73, %r4, 1;
mad.lo.s32 %r74, %r32, %r73, %r72;
mul.wide.s32 %rd22, %r74, 24;
add.s64 %rd23, %rd21, %rd22;
add.s64 %rd109, %rd23, 8;
mul.wide.s32 %rd4, %r68, 4;
mad.lo.s32 %r75, %r4, %r32, %r746;
add.s32 %r76, %r75, %r59;
add.s32 %r77, %r76, %r63;
mul.wide.s32 %rd24, %r77, 24;
add.s64 %rd25, %rd21, %rd24;
add.s64 %rd108, %rd25, 8;
cvta.to.global.u64 %rd6, %rd17;
mov.f32 %f324, 0f00000000;
shl.b64 %rd30, %rd2, 2;

$L__BB84_2:
cvt.u64.u32 %rd26, %r746;
add.s64 %rd27, %rd1, %rd26;
mul.lo.s64 %rd28, %rd27, 36;
add.s64 %rd29, %rd6, %rd28;
add.s64 %rd31, %rd29, %rd30;
add.s64 %rd32, %rd108, %rd4;
ld.global.nc.u32 %r372, [%rd108+-4];
shr.s32 %r373, %r372, %r7;
ld.global.nc.u32 %r80, [%rd31+4];
ld.global.nc.u32 %r84, [%rd31+20];
ld.global.nc.u32 %r374, [%rd32+4];
shr.s32 %r375, %r372, %r8;
ld.global.nc.u32 %r88, [%rd31+8];
ld.global.nc.u32 %r92, [%rd31+24];
ld.global.nc.u32 %r376, [%rd32];
and.b32 %r377, %r376, 252645135;
shl.b32 %r378, %r373, 4;
and.b32 %r379, %r378, 16;
or.b32 %r380, %r379, %r377;
shl.b32 %r381, %r373, 11;
and.b32 %r382, %r381, 4096;
or.b32 %r383, %r380, %r382;
shl.b32 %r384, %r373, 18;
and.b32 %r385, %r384, 1048576;
or.b32 %r386, %r383, %r385;
shl.b32 %r387, %r373, 25;
and.b32 %r388, %r387, 268435456;
or.b32 %r331, %r386, %r388;
mov.u32 %r354, 0;

dp4a.s32.s32 %r78, %r331, %r80, %r354;

shr.u32 %r389, %r376, 4;
and.b32 %r390, %r389, 252645135;
shr.u32 %r391, %r373, 12;
and.b32 %r392, %r391, 16;
or.b32 %r393, %r392, %r390;
shr.u32 %r394, %r373, 5;
and.b32 %r395, %r394, 4096;
or.b32 %r396, %r393, %r395;
shl.b32 %r397, %r373, 2;
and.b32 %r398, %r397, 1048576;
or.b32 %r399, %r396, %r398;
shl.b32 %r400, %r373, 9;
and.b32 %r401, %r400, 268435456;
or.b32 %r335, %r399, %r401;

dp4a.s32.s32 %r82, %r335, %r84, %r78;

and.b32 %r402, %r374, 252645135;
shl.b32 %r403, %r375, 4;
and.b32 %r404, %r403, 16;
or.b32 %r405, %r404, %r402;
shl.b32 %r406, %r375, 11;
and.b32 %r407, %r406, 4096;
or.b32 %r408, %r405, %r407;
shl.b32 %r409, %r375, 18;
and.b32 %r410, %r409, 1048576;
or.b32 %r411, %r408, %r410;
shl.b32 %r412, %r375, 25;
and.b32 %r413, %r412, 268435456;
or.b32 %r339, %r411, %r413;

dp4a.s32.s32 %r86, %r339, %r88, %r82;

shr.u32 %r414, %r374, 4;
and.b32 %r415, %r414, 252645135;
shr.u32 %r416, %r375, 12;
and.b32 %r417, %r416, 16;
or.b32 %r418, %r417, %r415;
shr.u32 %r419, %r375, 5;
and.b32 %r420, %r419, 4096;
or.b32 %r421, %r418, %r420;
shl.b32 %r422, %r375, 2;
and.b32 %r423, %r422, 1048576;
or.b32 %r424, %r421, %r423;
shl.b32 %r425, %r375, 9;
and.b32 %r426, %r425, 268435456;
or.b32 %r343, %r424, %r426;

dp4a.s32.s32 %r90, %r343, %r92, %r86;

ld.global.nc.u32 %r95, [%rd108+-8];
ld.global.nc.u32 %r96, [%rd29];

{mul.f16x2 %r94,%r95,%r96;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r94;
cvt.f32.f16 %f85, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r94;
cvt.f32.f16 %f86, high;}


cvt.rn.f32.s32 %f113, %r90;
mov.f32 %f114, 0f40000000;
div.approx.ftz.f32 %f115, %f86, %f114;
fma.rn.ftz.f32 %f116, %f85, %f113, %f115;
add.ftz.f32 %f337, %f116, %f337;
add.s64 %rd33, %rd109, %rd4;
ld.global.nc.u32 %r427, [%rd109+-4];
shr.s32 %r428, %r427, %r7;
ld.global.nc.u32 %r429, [%rd33+4];
shr.s32 %r430, %r427, %r8;
ld.global.nc.u32 %r431, [%rd33];
and.b32 %r432, %r431, 252645135;
shl.b32 %r433, %r428, 4;
and.b32 %r434, %r433, 16;
or.b32 %r435, %r434, %r432;
shl.b32 %r436, %r428, 11;
and.b32 %r437, %r436, 4096;
or.b32 %r438, %r435, %r437;
shl.b32 %r439, %r428, 18;
and.b32 %r440, %r439, 1048576;
or.b32 %r441, %r438, %r440;
shl.b32 %r442, %r428, 25;
and.b32 %r443, %r442, 268435456;
or.b32 %r352, %r441, %r443;

dp4a.s32.s32 %r99, %r352, %r80, %r354;

shr.u32 %r444, %r431, 4;
and.b32 %r445, %r444, 252645135;
shr.u32 %r446, %r428, 12;
and.b32 %r447, %r446, 16;
or.b32 %r448, %r447, %r445;
shr.u32 %r449, %r428, 5;
and.b32 %r450, %r449, 4096;
or.b32 %r451, %r448, %r450;
shl.b32 %r452, %r428, 2;
and.b32 %r453, %r452, 1048576;
or.b32 %r454, %r451, %r453;
shl.b32 %r455, %r428, 9;
and.b32 %r456, %r455, 268435456;
or.b32 %r356, %r454, %r456;

dp4a.s32.s32 %r103, %r356, %r84, %r99;

and.b32 %r457, %r429, 252645135;
shl.b32 %r458, %r430, 4;
and.b32 %r459, %r458, 16;
or.b32 %r460, %r459, %r457;
shl.b32 %r461, %r430, 11;
and.b32 %r462, %r461, 4096;
or.b32 %r463, %r460, %r462;
shl.b32 %r464, %r430, 18;
and.b32 %r465, %r464, 1048576;
or.b32 %r466, %r463, %r465;
shl.b32 %r467, %r430, 25;
and.b32 %r468, %r467, 268435456;
or.b32 %r360, %r466, %r468;

dp4a.s32.s32 %r107, %r360, %r88, %r103;

shr.u32 %r469, %r429, 4;
and.b32 %r470, %r469, 252645135;
shr.u32 %r471, %r430, 12;
and.b32 %r472, %r471, 16;
or.b32 %r473, %r472, %r470;
shr.u32 %r474, %r430, 5;
and.b32 %r475, %r474, 4096;
or.b32 %r476, %r473, %r475;
shl.b32 %r477, %r430, 2;
and.b32 %r478, %r477, 1048576;
or.b32 %r479, %r476, %r478;
shl.b32 %r480, %r430, 9;
and.b32 %r481, %r480, 268435456;
or.b32 %r364, %r479, %r481;

dp4a.s32.s32 %r111, %r364, %r92, %r107;

ld.global.nc.u32 %r116, [%rd109+-8];

{mul.f16x2 %r115,%r116,%r96;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r115;
cvt.f32.f16 %f87, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r115;
cvt.f32.f16 %f88, high;}


cvt.rn.f32.s32 %f117, %r111;
div.approx.ftz.f32 %f118, %f88, %f114;
fma.rn.ftz.f32 %f119, %f87, %f117, %f118;
add.ftz.f32 %f336, %f119, %f336;
cvt.u64.u32 %rd34, %r740;
add.s64 %rd35, %rd1, %rd34;
mul.lo.s64 %rd36, %rd35, 36;
add.s64 %rd37, %rd6, %rd36;
add.s64 %rd38, %rd37, %rd30;
ld.global.nc.u32 %r122, [%rd38+4];
ld.global.nc.u32 %r126, [%rd38+20];
ld.global.nc.u32 %r130, [%rd38+8];
ld.global.nc.u32 %r134, [%rd38+24];

dp4a.s32.s32 %r120, %r331, %r122, %r354;


dp4a.s32.s32 %r124, %r335, %r126, %r120;


dp4a.s32.s32 %r128, %r339, %r130, %r124;


dp4a.s32.s32 %r132, %r343, %r134, %r128;

ld.global.nc.u32 %r138, [%rd37];

{mul.f16x2 %r136,%r95,%r138;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r136;
cvt.f32.f16 %f89, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r136;
cvt.f32.f16 %f90, high;}


cvt.rn.f32.s32 %f120, %r132;
div.approx.ftz.f32 %f121, %f90, %f114;
fma.rn.ftz.f32 %f122, %f89, %f120, %f121;
add.ftz.f32 %f335, %f122, %f335;

dp4a.s32.s32 %r141, %r352, %r122, %r354;


dp4a.s32.s32 %r145, %r356, %r126, %r141;


dp4a.s32.s32 %r149, %r360, %r130, %r145;


dp4a.s32.s32 %r153, %r364, %r134, %r149;


{mul.f16x2 %r157,%r116,%r138;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r157;
cvt.f32.f16 %f91, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r157;
cvt.f32.f16 %f92, high;}


cvt.rn.f32.s32 %f123, %r153;
div.approx.ftz.f32 %f124, %f92, %f114;
fma.rn.ftz.f32 %f125, %f91, %f123, %f124;
add.ftz.f32 %f334, %f125, %f334;
cvt.u64.u32 %rd39, %r745;
add.s64 %rd40, %rd1, %rd39;
mul.lo.s64 %rd41, %rd40, 36;
add.s64 %rd42, %rd6, %rd41;
add.s64 %rd43, %rd42, %rd30;
ld.global.nc.u32 %r164, [%rd43+4];
ld.global.nc.u32 %r168, [%rd43+20];
ld.global.nc.u32 %r172, [%rd43+8];
ld.global.nc.u32 %r176, [%rd43+24];

dp4a.s32.s32 %r162, %r331, %r164, %r354;


dp4a.s32.s32 %r166, %r335, %r168, %r162;


dp4a.s32.s32 %r170, %r339, %r172, %r166;


dp4a.s32.s32 %r174, %r343, %r176, %r170;

ld.global.nc.u32 %r180, [%rd42];

{mul.f16x2 %r178,%r95,%r180;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r178;
cvt.f32.f16 %f93, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r178;
cvt.f32.f16 %f94, high;}


cvt.rn.f32.s32 %f126, %r174;
div.approx.ftz.f32 %f127, %f94, %f114;
fma.rn.ftz.f32 %f128, %f93, %f126, %f127;
add.ftz.f32 %f333, %f128, %f333;

dp4a.s32.s32 %r183, %r352, %r164, %r354;


dp4a.s32.s32 %r187, %r356, %r168, %r183;


dp4a.s32.s32 %r191, %r360, %r172, %r187;


dp4a.s32.s32 %r195, %r364, %r176, %r191;


{mul.f16x2 %r199,%r116,%r180;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r199;
cvt.f32.f16 %f95, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r199;
cvt.f32.f16 %f96, high;}


cvt.rn.f32.s32 %f129, %r195;
div.approx.ftz.f32 %f130, %f96, %f114;
fma.rn.ftz.f32 %f131, %f95, %f129, %f130;
add.ftz.f32 %f332, %f131, %f332;
cvt.u64.u32 %rd44, %r744;
add.s64 %rd45, %rd1, %rd44;
mul.lo.s64 %rd46, %rd45, 36;
add.s64 %rd47, %rd6, %rd46;
add.s64 %rd48, %rd47, %rd30;
ld.global.nc.u32 %r206, [%rd48+4];
ld.global.nc.u32 %r210, [%rd48+20];
ld.global.nc.u32 %r214, [%rd48+8];
ld.global.nc.u32 %r218, [%rd48+24];

dp4a.s32.s32 %r204, %r331, %r206, %r354;


dp4a.s32.s32 %r208, %r335, %r210, %r204;


dp4a.s32.s32 %r212, %r339, %r214, %r208;


dp4a.s32.s32 %r216, %r343, %r218, %r212;

ld.global.nc.u32 %r222, [%rd47];

{mul.f16x2 %r220,%r95,%r222;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r220;
cvt.f32.f16 %f97, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r220;
cvt.f32.f16 %f98, high;}


cvt.rn.f32.s32 %f132, %r216;
div.approx.ftz.f32 %f133, %f98, %f114;
fma.rn.ftz.f32 %f134, %f97, %f132, %f133;
add.ftz.f32 %f331, %f134, %f331;

dp4a.s32.s32 %r225, %r352, %r206, %r354;


dp4a.s32.s32 %r229, %r356, %r210, %r225;


dp4a.s32.s32 %r233, %r360, %r214, %r229;


dp4a.s32.s32 %r237, %r364, %r218, %r233;


{mul.f16x2 %r241,%r116,%r222;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r241;
cvt.f32.f16 %f99, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r241;
cvt.f32.f16 %f100, high;}


cvt.rn.f32.s32 %f135, %r237;
div.approx.ftz.f32 %f136, %f100, %f114;
fma.rn.ftz.f32 %f137, %f99, %f135, %f136;
add.ftz.f32 %f330, %f137, %f330;
cvt.u64.u32 %rd49, %r743;
add.s64 %rd50, %rd1, %rd49;
mul.lo.s64 %rd51, %rd50, 36;
add.s64 %rd52, %rd6, %rd51;
add.s64 %rd53, %rd52, %rd30;
ld.global.nc.u32 %r248, [%rd53+4];
ld.global.nc.u32 %r252, [%rd53+20];
ld.global.nc.u32 %r256, [%rd53+8];
ld.global.nc.u32 %r260, [%rd53+24];

dp4a.s32.s32 %r246, %r331, %r248, %r354;


dp4a.s32.s32 %r250, %r335, %r252, %r246;


dp4a.s32.s32 %r254, %r339, %r256, %r250;


dp4a.s32.s32 %r258, %r343, %r260, %r254;

ld.global.nc.u32 %r264, [%rd52];

{mul.f16x2 %r262,%r95,%r264;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r262;
cvt.f32.f16 %f101, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r262;
cvt.f32.f16 %f102, high;}


cvt.rn.f32.s32 %f138, %r258;
div.approx.ftz.f32 %f139, %f102, %f114;
fma.rn.ftz.f32 %f140, %f101, %f138, %f139;
add.ftz.f32 %f329, %f140, %f329;

dp4a.s32.s32 %r267, %r352, %r248, %r354;


dp4a.s32.s32 %r271, %r356, %r252, %r267;


dp4a.s32.s32 %r275, %r360, %r256, %r271;


dp4a.s32.s32 %r279, %r364, %r260, %r275;


{mul.f16x2 %r283,%r116,%r264;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r283;
cvt.f32.f16 %f103, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r283;
cvt.f32.f16 %f104, high;}


cvt.rn.f32.s32 %f141, %r279;
div.approx.ftz.f32 %f142, %f104, %f114;
fma.rn.ftz.f32 %f143, %f103, %f141, %f142;
add.ftz.f32 %f328, %f143, %f328;
cvt.u64.u32 %rd54, %r742;
add.s64 %rd55, %rd1, %rd54;
mul.lo.s64 %rd56, %rd55, 36;
add.s64 %rd57, %rd6, %rd56;
add.s64 %rd58, %rd57, %rd30;
ld.global.nc.u32 %r290, [%rd58+4];
ld.global.nc.u32 %r294, [%rd58+20];
ld.global.nc.u32 %r298, [%rd58+8];
ld.global.nc.u32 %r302, [%rd58+24];

dp4a.s32.s32 %r288, %r331, %r290, %r354;


dp4a.s32.s32 %r292, %r335, %r294, %r288;


dp4a.s32.s32 %r296, %r339, %r298, %r292;


dp4a.s32.s32 %r300, %r343, %r302, %r296;

ld.global.nc.u32 %r306, [%rd57];

{mul.f16x2 %r304,%r95,%r306;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r304;
cvt.f32.f16 %f105, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r304;
cvt.f32.f16 %f106, high;}


cvt.rn.f32.s32 %f144, %r300;
div.approx.ftz.f32 %f145, %f106, %f114;
fma.rn.ftz.f32 %f146, %f105, %f144, %f145;
add.ftz.f32 %f327, %f146, %f327;

dp4a.s32.s32 %r309, %r352, %r290, %r354;


dp4a.s32.s32 %r313, %r356, %r294, %r309;


dp4a.s32.s32 %r317, %r360, %r298, %r313;


dp4a.s32.s32 %r321, %r364, %r302, %r317;


{mul.f16x2 %r325,%r116,%r306;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r325;
cvt.f32.f16 %f107, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r325;
cvt.f32.f16 %f108, high;}


cvt.rn.f32.s32 %f147, %r321;
div.approx.ftz.f32 %f148, %f108, %f114;
fma.rn.ftz.f32 %f149, %f107, %f147, %f148;
add.ftz.f32 %f326, %f149, %f326;
cvt.u64.u32 %rd59, %r741;
add.s64 %rd60, %rd1, %rd59;
mul.lo.s64 %rd61, %rd60, 36;
add.s64 %rd62, %rd6, %rd61;
add.s64 %rd63, %rd62, %rd30;
ld.global.nc.u32 %r332, [%rd63+4];
ld.global.nc.u32 %r336, [%rd63+20];
ld.global.nc.u32 %r340, [%rd63+8];
ld.global.nc.u32 %r344, [%rd63+24];

dp4a.s32.s32 %r330, %r331, %r332, %r354;


dp4a.s32.s32 %r334, %r335, %r336, %r330;


dp4a.s32.s32 %r338, %r339, %r340, %r334;


dp4a.s32.s32 %r342, %r343, %r344, %r338;

ld.global.nc.u32 %r348, [%rd62];

{mul.f16x2 %r346,%r95,%r348;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r346;
cvt.f32.f16 %f109, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r346;
cvt.f32.f16 %f110, high;}


cvt.rn.f32.s32 %f150, %r342;
div.approx.ftz.f32 %f151, %f110, %f114;
fma.rn.ftz.f32 %f152, %f109, %f150, %f151;
add.ftz.f32 %f325, %f152, %f325;

dp4a.s32.s32 %r351, %r352, %r332, %r354;


dp4a.s32.s32 %r355, %r356, %r336, %r351;


dp4a.s32.s32 %r359, %r360, %r340, %r355;


dp4a.s32.s32 %r363, %r364, %r344, %r359;


{mul.f16x2 %r367,%r116,%r348;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r367;
cvt.f32.f16 %f111, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r367;
cvt.f32.f16 %f112, high;}


cvt.rn.f32.s32 %f153, %r363;
div.approx.ftz.f32 %f154, %f112, %f114;
fma.rn.ftz.f32 %f155, %f111, %f153, %f154;
add.ftz.f32 %f324, %f155, %f324;
add.s32 %r745, %r745, 32;
add.s32 %r744, %r744, 32;
add.s32 %r743, %r743, 32;
add.s32 %r742, %r742, 32;
add.s32 %r741, %r741, 32;
add.s32 %r740, %r740, 32;
add.s64 %rd109, %rd109, 768;
add.s64 %rd108, %rd108, 768;
add.s32 %r746, %r746, 32;
setp.lt.s32 %p4, %r746, %r6;
@%p4 bra $L__BB84_2;

$L__BB84_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB84_5;

mov.u32 %r482, _ZZ13mul_mat_vec_qIL9ggml_type7ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r483, %r1, 1792, %r482;
shl.b32 %r484, %r2, 2;
add.s32 %r485, %r483, %r484;
st.shared.f32 [%r485+-1792], %f337;
st.shared.f32 [%r485+-1664], %f336;
st.shared.f32 [%r485+-1536], %f335;
st.shared.f32 [%r485+-1408], %f334;
st.shared.f32 [%r485+-1280], %f333;
st.shared.f32 [%r485+-1152], %f332;
st.shared.f32 [%r485+-1024], %f331;
st.shared.f32 [%r485+-896], %f330;
st.shared.f32 [%r485+-768], %f329;
st.shared.f32 [%r485+-640], %f328;
st.shared.f32 [%r485+-512], %f327;
st.shared.f32 [%r485+-384], %f326;
st.shared.f32 [%r485+-256], %f325;
st.shared.f32 [%r485+-128], %f324;

$L__BB84_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB84_34;

mov.u32 %r486, %ctaid.z;
mov.u32 %r487, %ctaid.y;
mad.lo.s32 %r488, %r487, %r40, %r4;
mad.lo.s32 %r489, %r486, %r46, %r488;
cvt.u64.u32 %rd11, %r489;
shl.b32 %r490, %r2, 2;
mov.u32 %r491, _ZZ13mul_mat_vec_qIL9ggml_type7ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r29, %r491, %r490;
mov.u32 %r492, 2;
ld.shared.f32 %f156, [%r29];
add.ftz.f32 %f157, %f156, %f337;
mov.b32 %r493, %f157;
mov.u32 %r494, 31;
mov.u32 %r495, 16;
mov.u32 %r496, -1;
shfl.sync.bfly.b32 %r497|%p7, %r493, %r495, %r494, %r496;
mov.b32 %f158, %r497;
add.ftz.f32 %f159, %f157, %f158;
mov.b32 %r498, %f159;
mov.u32 %r499, 8;
shfl.sync.bfly.b32 %r500|%p8, %r498, %r499, %r494, %r496;
mov.b32 %f160, %r500;
add.ftz.f32 %f161, %f159, %f160;
mov.b32 %r501, %f161;
mov.u32 %r502, 4;
shfl.sync.bfly.b32 %r503|%p9, %r501, %r502, %r494, %r496;
mov.b32 %f162, %r503;
add.ftz.f32 %f163, %f161, %f162;
mov.b32 %r504, %f163;
shfl.sync.bfly.b32 %r505|%p10, %r504, %r492, %r494, %r496;
mov.b32 %f164, %r505;
add.ftz.f32 %f165, %f163, %f164;
mov.b32 %r506, %f165;
mov.u32 %r507, 1;
shfl.sync.bfly.b32 %r508|%p11, %r506, %r507, %r494, %r496;
mov.b32 %f166, %r508;
add.ftz.f32 %f43, %f165, %f166;
setp.lt.u32 %p12, %r4, %r34;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
cvta.to.global.u64 %rd64, %rd18;
mul.wide.u32 %rd65, %r489, 4;
add.s64 %rd12, %rd64, %rd65;
not.pred %p14, %p1;
@%p14 bra $L__BB84_8;

st.global.f32 [%rd12], %f43;

$L__BB84_8:
ld.shared.f32 %f167, [%r29+128];
add.ftz.f32 %f168, %f167, %f336;
mov.b32 %r509, %f168;
shfl.sync.bfly.b32 %r513|%p15, %r509, %r495, %r494, %r496;
mov.b32 %f169, %r513;
add.ftz.f32 %f170, %f168, %f169;
mov.b32 %r514, %f170;
shfl.sync.bfly.b32 %r516|%p16, %r514, %r499, %r494, %r496;
mov.b32 %f171, %r516;
add.ftz.f32 %f172, %f170, %f171;
mov.b32 %r517, %f172;
shfl.sync.bfly.b32 %r519|%p17, %r517, %r502, %r494, %r496;
mov.b32 %f173, %r519;
add.ftz.f32 %f174, %f172, %f173;
mov.b32 %r520, %f174;
shfl.sync.bfly.b32 %r522|%p18, %r520, %r492, %r494, %r496;
mov.b32 %f175, %r522;
add.ftz.f32 %f176, %f174, %f175;
mov.b32 %r523, %f176;
shfl.sync.bfly.b32 %r525|%p19, %r523, %r507, %r494, %r496;
mov.b32 %f177, %r525;
add.ftz.f32 %f44, %f176, %f177;
setp.eq.s32 %p20, %r2, 1;
or.b32 %r529, %r4, 1;
setp.lt.u32 %p21, %r529, %r34;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB84_10;

st.global.f32 [%rd12+4], %f44;

$L__BB84_10:
ld.shared.f32 %f178, [%r29+256];
add.ftz.f32 %f179, %f178, %f335;
mov.b32 %r530, %f179;
mov.u32 %r531, 31;
mov.u32 %r532, 16;
mov.u32 %r533, -1;
shfl.sync.bfly.b32 %r534|%p23, %r530, %r532, %r531, %r533;
mov.b32 %f180, %r534;
add.ftz.f32 %f181, %f179, %f180;
mov.b32 %r535, %f181;
mov.u32 %r536, 8;
shfl.sync.bfly.b32 %r537|%p24, %r535, %r536, %r531, %r533;
mov.b32 %f182, %r537;
add.ftz.f32 %f183, %f181, %f182;
mov.b32 %r538, %f183;
mov.u32 %r539, 4;
shfl.sync.bfly.b32 %r540|%p25, %r538, %r539, %r531, %r533;
mov.b32 %f184, %r540;
add.ftz.f32 %f185, %f183, %f184;
mov.b32 %r541, %f185;
mov.u32 %r542, 2;
shfl.sync.bfly.b32 %r543|%p26, %r541, %r542, %r531, %r533;
mov.b32 %f186, %r543;
add.ftz.f32 %f187, %f185, %f186;
mov.b32 %r544, %f187;
mov.u32 %r545, 1;
shfl.sync.bfly.b32 %r546|%p27, %r544, %r545, %r531, %r533;
mov.b32 %f188, %r546;
add.ftz.f32 %f45, %f187, %f188;
@%p14 bra $L__BB84_12;

cvt.u64.u32 %rd66, %r34;
add.s64 %rd67, %rd66, %rd11;
shl.b64 %rd69, %rd67, 2;
add.s64 %rd70, %rd64, %rd69;
st.global.f32 [%rd70], %f45;

$L__BB84_12:
ld.shared.f32 %f189, [%r29+384];
add.ftz.f32 %f190, %f189, %f334;
mov.b32 %r547, %f190;
shfl.sync.bfly.b32 %r551|%p29, %r547, %r532, %r531, %r533;
mov.b32 %f191, %r551;
add.ftz.f32 %f192, %f190, %f191;
mov.b32 %r552, %f192;
shfl.sync.bfly.b32 %r554|%p30, %r552, %r536, %r531, %r533;
mov.b32 %f193, %r554;
add.ftz.f32 %f194, %f192, %f193;
mov.b32 %r555, %f194;
shfl.sync.bfly.b32 %r557|%p31, %r555, %r539, %r531, %r533;
mov.b32 %f195, %r557;
add.ftz.f32 %f196, %f194, %f195;
mov.b32 %r558, %f196;
shfl.sync.bfly.b32 %r560|%p32, %r558, %r542, %r531, %r533;
mov.b32 %f197, %r560;
add.ftz.f32 %f198, %f196, %f197;
mov.b32 %r561, %f198;
shfl.sync.bfly.b32 %r563|%p33, %r561, %r545, %r531, %r533;
mov.b32 %f199, %r563;
add.ftz.f32 %f46, %f198, %f199;
@%p22 bra $L__BB84_14;

add.s32 %r564, %r34, 1;
cvt.u64.u32 %rd71, %r564;
add.s64 %rd72, %rd71, %rd11;
shl.b64 %rd74, %rd72, 2;
add.s64 %rd75, %rd64, %rd74;
st.global.f32 [%rd75], %f46;

$L__BB84_14:
shl.b32 %r565, %r34, 1;
mov.u32 %r566, 1;
ld.shared.f32 %f200, [%r29+512];
add.ftz.f32 %f201, %f200, %f333;
mov.b32 %r567, %f201;
mov.u32 %r568, 31;
mov.u32 %r569, 16;
mov.u32 %r570, -1;
shfl.sync.bfly.b32 %r571|%p35, %r567, %r569, %r568, %r570;
mov.b32 %f202, %r571;
add.ftz.f32 %f203, %f201, %f202;
mov.b32 %r572, %f203;
mov.u32 %r573, 8;
shfl.sync.bfly.b32 %r574|%p36, %r572, %r573, %r568, %r570;
mov.b32 %f204, %r574;
add.ftz.f32 %f205, %f203, %f204;
mov.b32 %r575, %f205;
mov.u32 %r576, 4;
shfl.sync.bfly.b32 %r577|%p37, %r575, %r576, %r568, %r570;
mov.b32 %f206, %r577;
add.ftz.f32 %f207, %f205, %f206;
mov.b32 %r578, %f207;
mov.u32 %r579, 2;
shfl.sync.bfly.b32 %r580|%p38, %r578, %r579, %r568, %r570;
mov.b32 %f208, %r580;
add.ftz.f32 %f209, %f207, %f208;
mov.b32 %r581, %f209;
shfl.sync.bfly.b32 %r582|%p39, %r581, %r566, %r568, %r570;
mov.b32 %f210, %r582;
add.ftz.f32 %f47, %f209, %f210;
cvt.u64.u32 %rd76, %r565;
add.s64 %rd77, %rd76, %rd11;
shl.b64 %rd79, %rd77, 2;
add.s64 %rd13, %rd64, %rd79;
@%p14 bra $L__BB84_16;

st.global.f32 [%rd13], %f47;

$L__BB84_16:
ld.shared.f32 %f211, [%r29+640];
add.ftz.f32 %f212, %f211, %f332;
mov.b32 %r583, %f212;
shfl.sync.bfly.b32 %r587|%p41, %r583, %r569, %r568, %r570;
mov.b32 %f213, %r587;
add.ftz.f32 %f214, %f212, %f213;
mov.b32 %r588, %f214;
shfl.sync.bfly.b32 %r590|%p42, %r588, %r573, %r568, %r570;
mov.b32 %f215, %r590;
add.ftz.f32 %f216, %f214, %f215;
mov.b32 %r591, %f216;
shfl.sync.bfly.b32 %r593|%p43, %r591, %r576, %r568, %r570;
mov.b32 %f217, %r593;
add.ftz.f32 %f218, %f216, %f217;
mov.b32 %r594, %f218;
shfl.sync.bfly.b32 %r596|%p44, %r594, %r579, %r568, %r570;
mov.b32 %f219, %r596;
add.ftz.f32 %f220, %f218, %f219;
mov.b32 %r597, %f220;
shfl.sync.bfly.b32 %r599|%p45, %r597, %r566, %r568, %r570;
mov.b32 %f221, %r599;
add.ftz.f32 %f48, %f220, %f221;
@%p22 bra $L__BB84_18;

st.global.f32 [%rd13+4], %f48;

$L__BB84_18:
mul.lo.s32 %r30, %r34, 3;
ld.shared.f32 %f222, [%r29+768];
add.ftz.f32 %f223, %f222, %f331;
mov.b32 %r600, %f223;
mov.u32 %r601, 31;
mov.u32 %r602, 16;
mov.u32 %r603, -1;
shfl.sync.bfly.b32 %r604|%p47, %r600, %r602, %r601, %r603;
mov.b32 %f224, %r604;
add.ftz.f32 %f225, %f223, %f224;
mov.b32 %r605, %f225;
mov.u32 %r606, 8;
shfl.sync.bfly.b32 %r607|%p48, %r605, %r606, %r601, %r603;
mov.b32 %f226, %r607;
add.ftz.f32 %f227, %f225, %f226;
mov.b32 %r608, %f227;
mov.u32 %r609, 4;
shfl.sync.bfly.b32 %r610|%p49, %r608, %r609, %r601, %r603;
mov.b32 %f228, %r610;
add.ftz.f32 %f229, %f227, %f228;
mov.b32 %r611, %f229;
mov.u32 %r612, 2;
shfl.sync.bfly.b32 %r613|%p50, %r611, %r612, %r601, %r603;
mov.b32 %f230, %r613;
add.ftz.f32 %f231, %f229, %f230;
mov.b32 %r614, %f231;
mov.u32 %r615, 1;
shfl.sync.bfly.b32 %r616|%p51, %r614, %r615, %r601, %r603;
mov.b32 %f232, %r616;
add.ftz.f32 %f49, %f231, %f232;
@%p14 bra $L__BB84_20;

cvt.u64.u32 %rd80, %r30;
add.s64 %rd81, %rd80, %rd11;
shl.b64 %rd83, %rd81, 2;
add.s64 %rd84, %rd64, %rd83;
st.global.f32 [%rd84], %f49;

$L__BB84_20:
ld.shared.f32 %f233, [%r29+896];
add.ftz.f32 %f234, %f233, %f330;
mov.b32 %r617, %f234;
shfl.sync.bfly.b32 %r621|%p53, %r617, %r602, %r601, %r603;
mov.b32 %f235, %r621;
add.ftz.f32 %f236, %f234, %f235;
mov.b32 %r622, %f236;
shfl.sync.bfly.b32 %r624|%p54, %r622, %r606, %r601, %r603;
mov.b32 %f237, %r624;
add.ftz.f32 %f238, %f236, %f237;
mov.b32 %r625, %f238;
shfl.sync.bfly.b32 %r627|%p55, %r625, %r609, %r601, %r603;
mov.b32 %f239, %r627;
add.ftz.f32 %f240, %f238, %f239;
mov.b32 %r628, %f240;
shfl.sync.bfly.b32 %r630|%p56, %r628, %r612, %r601, %r603;
mov.b32 %f241, %r630;
add.ftz.f32 %f242, %f240, %f241;
mov.b32 %r631, %f242;
shfl.sync.bfly.b32 %r633|%p57, %r631, %r615, %r601, %r603;
mov.b32 %f243, %r633;
add.ftz.f32 %f50, %f242, %f243;
@%p22 bra $L__BB84_22;

add.s32 %r634, %r30, 1;
cvt.u64.u32 %rd85, %r634;
add.s64 %rd86, %rd85, %rd11;
shl.b64 %rd88, %rd86, 2;
add.s64 %rd89, %rd64, %rd88;
st.global.f32 [%rd89], %f50;

$L__BB84_22:
shl.b32 %r635, %r34, 2;
mov.u32 %r636, 2;
ld.shared.f32 %f244, [%r29+1024];
add.ftz.f32 %f245, %f244, %f329;
mov.b32 %r637, %f245;
mov.u32 %r638, 31;
mov.u32 %r639, 16;
mov.u32 %r640, -1;
shfl.sync.bfly.b32 %r641|%p59, %r637, %r639, %r638, %r640;
mov.b32 %f246, %r641;
add.ftz.f32 %f247, %f245, %f246;
mov.b32 %r642, %f247;
mov.u32 %r643, 8;
shfl.sync.bfly.b32 %r644|%p60, %r642, %r643, %r638, %r640;
mov.b32 %f248, %r644;
add.ftz.f32 %f249, %f247, %f248;
mov.b32 %r645, %f249;
mov.u32 %r646, 4;
shfl.sync.bfly.b32 %r647|%p61, %r645, %r646, %r638, %r640;
mov.b32 %f250, %r647;
add.ftz.f32 %f251, %f249, %f250;
mov.b32 %r648, %f251;
shfl.sync.bfly.b32 %r649|%p62, %r648, %r636, %r638, %r640;
mov.b32 %f252, %r649;
add.ftz.f32 %f253, %f251, %f252;
mov.b32 %r650, %f253;
mov.u32 %r651, 1;
shfl.sync.bfly.b32 %r652|%p63, %r650, %r651, %r638, %r640;
mov.b32 %f254, %r652;
add.ftz.f32 %f51, %f253, %f254;
cvt.u64.u32 %rd90, %r635;
add.s64 %rd91, %rd90, %rd11;
shl.b64 %rd93, %rd91, 2;
add.s64 %rd14, %rd64, %rd93;
@%p14 bra $L__BB84_24;

st.global.f32 [%rd14], %f51;

$L__BB84_24:
ld.shared.f32 %f255, [%r29+1152];
add.ftz.f32 %f256, %f255, %f328;
mov.b32 %r653, %f256;
shfl.sync.bfly.b32 %r657|%p65, %r653, %r639, %r638, %r640;
mov.b32 %f257, %r657;
add.ftz.f32 %f258, %f256, %f257;
mov.b32 %r658, %f258;
shfl.sync.bfly.b32 %r660|%p66, %r658, %r643, %r638, %r640;
mov.b32 %f259, %r660;
add.ftz.f32 %f260, %f258, %f259;
mov.b32 %r661, %f260;
shfl.sync.bfly.b32 %r663|%p67, %r661, %r646, %r638, %r640;
mov.b32 %f261, %r663;
add.ftz.f32 %f262, %f260, %f261;
mov.b32 %r664, %f262;
shfl.sync.bfly.b32 %r666|%p68, %r664, %r636, %r638, %r640;
mov.b32 %f263, %r666;
add.ftz.f32 %f264, %f262, %f263;
mov.b32 %r667, %f264;
shfl.sync.bfly.b32 %r669|%p69, %r667, %r651, %r638, %r640;
mov.b32 %f265, %r669;
add.ftz.f32 %f52, %f264, %f265;
@%p22 bra $L__BB84_26;

st.global.f32 [%rd14+4], %f52;

$L__BB84_26:
mul.lo.s32 %r31, %r34, 5;
ld.shared.f32 %f266, [%r29+1280];
add.ftz.f32 %f267, %f266, %f327;
mov.b32 %r670, %f267;
mov.u32 %r671, 31;
mov.u32 %r672, 16;
mov.u32 %r673, -1;
shfl.sync.bfly.b32 %r674|%p71, %r670, %r672, %r671, %r673;
mov.b32 %f268, %r674;
add.ftz.f32 %f269, %f267, %f268;
mov.b32 %r675, %f269;
mov.u32 %r676, 8;
shfl.sync.bfly.b32 %r677|%p72, %r675, %r676, %r671, %r673;
mov.b32 %f270, %r677;
add.ftz.f32 %f271, %f269, %f270;
mov.b32 %r678, %f271;
mov.u32 %r679, 4;
shfl.sync.bfly.b32 %r680|%p73, %r678, %r679, %r671, %r673;
mov.b32 %f272, %r680;
add.ftz.f32 %f273, %f271, %f272;
mov.b32 %r681, %f273;
mov.u32 %r682, 2;
shfl.sync.bfly.b32 %r683|%p74, %r681, %r682, %r671, %r673;
mov.b32 %f274, %r683;
add.ftz.f32 %f275, %f273, %f274;
mov.b32 %r684, %f275;
mov.u32 %r685, 1;
shfl.sync.bfly.b32 %r686|%p75, %r684, %r685, %r671, %r673;
mov.b32 %f276, %r686;
add.ftz.f32 %f53, %f275, %f276;
@%p14 bra $L__BB84_28;

cvt.u64.u32 %rd94, %r31;
add.s64 %rd95, %rd94, %rd11;
shl.b64 %rd97, %rd95, 2;
add.s64 %rd98, %rd64, %rd97;
st.global.f32 [%rd98], %f53;

$L__BB84_28:
ld.shared.f32 %f277, [%r29+1408];
add.ftz.f32 %f278, %f277, %f326;
mov.b32 %r687, %f278;
shfl.sync.bfly.b32 %r691|%p77, %r687, %r672, %r671, %r673;
mov.b32 %f279, %r691;
add.ftz.f32 %f280, %f278, %f279;
mov.b32 %r692, %f280;
shfl.sync.bfly.b32 %r694|%p78, %r692, %r676, %r671, %r673;
mov.b32 %f281, %r694;
add.ftz.f32 %f282, %f280, %f281;
mov.b32 %r695, %f282;
shfl.sync.bfly.b32 %r697|%p79, %r695, %r679, %r671, %r673;
mov.b32 %f283, %r697;
add.ftz.f32 %f284, %f282, %f283;
mov.b32 %r698, %f284;
shfl.sync.bfly.b32 %r700|%p80, %r698, %r682, %r671, %r673;
mov.b32 %f285, %r700;
add.ftz.f32 %f286, %f284, %f285;
mov.b32 %r701, %f286;
shfl.sync.bfly.b32 %r703|%p81, %r701, %r685, %r671, %r673;
mov.b32 %f287, %r703;
add.ftz.f32 %f54, %f286, %f287;
@%p22 bra $L__BB84_30;

add.s32 %r704, %r31, 1;
cvt.u64.u32 %rd99, %r704;
add.s64 %rd100, %rd99, %rd11;
shl.b64 %rd102, %rd100, 2;
add.s64 %rd103, %rd64, %rd102;
st.global.f32 [%rd103], %f54;

$L__BB84_30:
mul.lo.s32 %r705, %r34, 6;
ld.shared.f32 %f288, [%r29+1536];
add.ftz.f32 %f289, %f288, %f325;
mov.b32 %r706, %f289;
mov.u32 %r707, 31;
mov.u32 %r708, 16;
mov.u32 %r709, -1;
shfl.sync.bfly.b32 %r710|%p83, %r706, %r708, %r707, %r709;
mov.b32 %f290, %r710;
add.ftz.f32 %f291, %f289, %f290;
mov.b32 %r711, %f291;
mov.u32 %r712, 8;
shfl.sync.bfly.b32 %r713|%p84, %r711, %r712, %r707, %r709;
mov.b32 %f292, %r713;
add.ftz.f32 %f293, %f291, %f292;
mov.b32 %r714, %f293;
mov.u32 %r715, 4;
shfl.sync.bfly.b32 %r716|%p85, %r714, %r715, %r707, %r709;
mov.b32 %f294, %r716;
add.ftz.f32 %f295, %f293, %f294;
mov.b32 %r717, %f295;
mov.u32 %r718, 2;
shfl.sync.bfly.b32 %r719|%p86, %r717, %r718, %r707, %r709;
mov.b32 %f296, %r719;
add.ftz.f32 %f297, %f295, %f296;
mov.b32 %r720, %f297;
mov.u32 %r721, 1;
shfl.sync.bfly.b32 %r722|%p87, %r720, %r721, %r707, %r709;
mov.b32 %f298, %r722;
add.ftz.f32 %f55, %f297, %f298;
cvt.u64.u32 %rd104, %r705;
add.s64 %rd105, %rd104, %rd11;
shl.b64 %rd107, %rd105, 2;
add.s64 %rd15, %rd64, %rd107;
@%p14 bra $L__BB84_32;

st.global.f32 [%rd15], %f55;

$L__BB84_32:
ld.shared.f32 %f299, [%r29+1664];
add.ftz.f32 %f300, %f299, %f324;
mov.b32 %r723, %f300;
shfl.sync.bfly.b32 %r727|%p89, %r723, %r708, %r707, %r709;
mov.b32 %f301, %r727;
add.ftz.f32 %f302, %f300, %f301;
mov.b32 %r728, %f302;
shfl.sync.bfly.b32 %r730|%p90, %r728, %r712, %r707, %r709;
mov.b32 %f303, %r730;
add.ftz.f32 %f304, %f302, %f303;
mov.b32 %r731, %f304;
shfl.sync.bfly.b32 %r733|%p91, %r731, %r715, %r707, %r709;
mov.b32 %f305, %r733;
add.ftz.f32 %f306, %f304, %f305;
mov.b32 %r734, %f306;
shfl.sync.bfly.b32 %r736|%p92, %r734, %r718, %r707, %r709;
mov.b32 %f307, %r736;
add.ftz.f32 %f308, %f306, %f307;
mov.b32 %r737, %f308;
shfl.sync.bfly.b32 %r739|%p93, %r737, %r721, %r707, %r709;
mov.b32 %f309, %r739;
add.ftz.f32 %f56, %f308, %f309;
@%p22 bra $L__BB84_34;

st.global.f32 [%rd15+4], %f56;

$L__BB84_34:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type7ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type7ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type7ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type7ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type7ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type7ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type7ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type7ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type7ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type7ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<107>;
.reg .f32 %f<386>;
.reg .b32 %r<830>;
.reg .b64 %rd<125>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type7ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[2048];

ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type7ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type7ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type7ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r51, [_Z13mul_mat_vec_qIL9ggml_type7ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type7ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type7ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type7ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type7ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type7ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type7ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r48, [_Z13mul_mat_vec_qIL9ggml_type7ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r49, [_Z13mul_mat_vec_qIL9ggml_type7ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r50, [_Z13mul_mat_vec_qIL9ggml_type7ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r46, [_Z13mul_mat_vec_qIL9ggml_type7ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r45, [_Z13mul_mat_vec_qIL9ggml_type7ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type7ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type7ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r52, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r52, %r2;
mov.u32 %r53, %ctaid.x;
shl.b32 %r4, %r53, 1;
shr.u32 %r54, %r3, 31;
add.s32 %r55, %r3, %r54;
shr.s32 %r829, %r55, 1;
shr.u32 %r6, %r51, 5;
setp.ge.s32 %p3, %r829, %r6;
mov.f32 %f370, 0f00000000;
mov.f32 %f371, %f370;
mov.f32 %f372, %f370;
mov.f32 %f373, %f370;
mov.f32 %f374, %f370;
mov.f32 %f375, %f370;
mov.f32 %f376, %f370;
mov.f32 %f377, %f370;
mov.f32 %f378, %f370;
mov.f32 %f379, %f370;
mov.f32 %f380, %f370;
mov.f32 %f381, %f370;
mov.f32 %f382, %f370;
mov.f32 %f383, %f370;
mov.f32 %f384, %f370;
mov.f32 %f385, %f370;
@%p3 bra $L__BB85_3;

mov.u32 %r56, %ctaid.y;
mov.u32 %r57, %ctaid.z;
mul.lo.s32 %r58, %r56, %r43;
cvt.u64.u32 %rd19, %r58;
mul.lo.s32 %r59, %r57, %r49;
cvt.u64.u32 %rd20, %r59;
add.s64 %rd1, %rd20, %rd19;
mul.hi.u32 %r60, %r57, %r45;
add.s32 %r61, %r60, %r57;
shr.u32 %r62, %r61, %r46;
mul.lo.s32 %r63, %r62, %r48;
mul.hi.u32 %r64, %r56, %r39;
add.s32 %r65, %r64, %r56;
shr.u32 %r66, %r65, %r40;
mul.lo.s32 %r67, %r66, %r42;
cvta.to.global.u64 %rd21, %rd16;
and.b32 %r70, %r55, -2;
sub.s32 %r71, %r3, %r70;
shl.b32 %r72, %r71, 1;
cvt.s64.s32 %rd2, %r72;
shl.b32 %r7, %r71, 3;
or.b32 %r8, %r7, 4;
shl.b32 %r73, %r37, 1;
add.s32 %r828, %r829, %r73;
mad.lo.s32 %r827, %r37, 3, %r829;
shl.b32 %r74, %r37, 2;
add.s32 %r826, %r829, %r74;
mad.lo.s32 %r825, %r37, 5, %r829;
mad.lo.s32 %r824, %r37, 6, %r829;
mad.lo.s32 %r823, %r37, 7, %r829;
add.s32 %r822, %r829, %r37;
add.s32 %r75, %r829, %r63;
add.s32 %r76, %r75, %r67;
add.s32 %r77, %r4, 1;
mad.lo.s32 %r78, %r36, %r77, %r76;
mul.wide.s32 %rd22, %r78, 24;
add.s64 %rd23, %rd21, %rd22;
add.s64 %rd124, %rd23, 8;
mul.wide.s32 %rd4, %r72, 4;
mad.lo.s32 %r79, %r4, %r36, %r829;
add.s32 %r80, %r79, %r63;
add.s32 %r81, %r80, %r67;
mul.wide.s32 %rd24, %r81, 24;
add.s64 %rd25, %rd21, %rd24;
add.s64 %rd123, %rd25, 8;
cvta.to.global.u64 %rd6, %rd17;
mov.f32 %f370, 0f00000000;
shl.b64 %rd30, %rd2, 2;

$L__BB85_2:
cvt.u64.u32 %rd26, %r829;
add.s64 %rd27, %rd1, %rd26;
mul.lo.s64 %rd28, %rd27, 36;
add.s64 %rd29, %rd6, %rd28;
add.s64 %rd31, %rd29, %rd30;
add.s64 %rd32, %rd123, %rd4;
ld.global.nc.u32 %r418, [%rd123+-4];
shr.s32 %r419, %r418, %r7;
ld.global.nc.u32 %r84, [%rd31+4];
ld.global.nc.u32 %r88, [%rd31+20];
ld.global.nc.u32 %r420, [%rd32+4];
shr.s32 %r421, %r418, %r8;
ld.global.nc.u32 %r92, [%rd31+8];
ld.global.nc.u32 %r96, [%rd31+24];
ld.global.nc.u32 %r422, [%rd32];
and.b32 %r423, %r422, 252645135;
shl.b32 %r424, %r419, 4;
and.b32 %r425, %r424, 16;
or.b32 %r426, %r425, %r423;
shl.b32 %r427, %r419, 11;
and.b32 %r428, %r427, 4096;
or.b32 %r429, %r426, %r428;
shl.b32 %r430, %r419, 18;
and.b32 %r431, %r430, 1048576;
or.b32 %r432, %r429, %r431;
shl.b32 %r433, %r419, 25;
and.b32 %r434, %r433, 268435456;
or.b32 %r377, %r432, %r434;
mov.u32 %r400, 0;

dp4a.s32.s32 %r82, %r377, %r84, %r400;

shr.u32 %r435, %r422, 4;
and.b32 %r436, %r435, 252645135;
shr.u32 %r437, %r419, 12;
and.b32 %r438, %r437, 16;
or.b32 %r439, %r438, %r436;
shr.u32 %r440, %r419, 5;
and.b32 %r441, %r440, 4096;
or.b32 %r442, %r439, %r441;
shl.b32 %r443, %r419, 2;
and.b32 %r444, %r443, 1048576;
or.b32 %r445, %r442, %r444;
shl.b32 %r446, %r419, 9;
and.b32 %r447, %r446, 268435456;
or.b32 %r381, %r445, %r447;

dp4a.s32.s32 %r86, %r381, %r88, %r82;

and.b32 %r448, %r420, 252645135;
shl.b32 %r449, %r421, 4;
and.b32 %r450, %r449, 16;
or.b32 %r451, %r450, %r448;
shl.b32 %r452, %r421, 11;
and.b32 %r453, %r452, 4096;
or.b32 %r454, %r451, %r453;
shl.b32 %r455, %r421, 18;
and.b32 %r456, %r455, 1048576;
or.b32 %r457, %r454, %r456;
shl.b32 %r458, %r421, 25;
and.b32 %r459, %r458, 268435456;
or.b32 %r385, %r457, %r459;

dp4a.s32.s32 %r90, %r385, %r92, %r86;

shr.u32 %r460, %r420, 4;
and.b32 %r461, %r460, 252645135;
shr.u32 %r462, %r421, 12;
and.b32 %r463, %r462, 16;
or.b32 %r464, %r463, %r461;
shr.u32 %r465, %r421, 5;
and.b32 %r466, %r465, 4096;
or.b32 %r467, %r464, %r466;
shl.b32 %r468, %r421, 2;
and.b32 %r469, %r468, 1048576;
or.b32 %r470, %r467, %r469;
shl.b32 %r471, %r421, 9;
and.b32 %r472, %r471, 268435456;
or.b32 %r389, %r470, %r472;

dp4a.s32.s32 %r94, %r389, %r96, %r90;

ld.global.nc.u32 %r99, [%rd123+-8];
ld.global.nc.u32 %r100, [%rd29];

{mul.f16x2 %r98,%r99,%r100;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r98;
cvt.f32.f16 %f97, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r98;
cvt.f32.f16 %f98, high;}


cvt.rn.f32.s32 %f129, %r94;
mov.f32 %f130, 0f40000000;
div.approx.ftz.f32 %f131, %f98, %f130;
fma.rn.ftz.f32 %f132, %f97, %f129, %f131;
add.ftz.f32 %f385, %f132, %f385;
add.s64 %rd33, %rd124, %rd4;
ld.global.nc.u32 %r473, [%rd124+-4];
shr.s32 %r474, %r473, %r7;
ld.global.nc.u32 %r475, [%rd33+4];
shr.s32 %r476, %r473, %r8;
ld.global.nc.u32 %r477, [%rd33];
and.b32 %r478, %r477, 252645135;
shl.b32 %r479, %r474, 4;
and.b32 %r480, %r479, 16;
or.b32 %r481, %r480, %r478;
shl.b32 %r482, %r474, 11;
and.b32 %r483, %r482, 4096;
or.b32 %r484, %r481, %r483;
shl.b32 %r485, %r474, 18;
and.b32 %r486, %r485, 1048576;
or.b32 %r487, %r484, %r486;
shl.b32 %r488, %r474, 25;
and.b32 %r489, %r488, 268435456;
or.b32 %r398, %r487, %r489;

dp4a.s32.s32 %r103, %r398, %r84, %r400;

shr.u32 %r490, %r477, 4;
and.b32 %r491, %r490, 252645135;
shr.u32 %r492, %r474, 12;
and.b32 %r493, %r492, 16;
or.b32 %r494, %r493, %r491;
shr.u32 %r495, %r474, 5;
and.b32 %r496, %r495, 4096;
or.b32 %r497, %r494, %r496;
shl.b32 %r498, %r474, 2;
and.b32 %r499, %r498, 1048576;
or.b32 %r500, %r497, %r499;
shl.b32 %r501, %r474, 9;
and.b32 %r502, %r501, 268435456;
or.b32 %r402, %r500, %r502;

dp4a.s32.s32 %r107, %r402, %r88, %r103;

and.b32 %r503, %r475, 252645135;
shl.b32 %r504, %r476, 4;
and.b32 %r505, %r504, 16;
or.b32 %r506, %r505, %r503;
shl.b32 %r507, %r476, 11;
and.b32 %r508, %r507, 4096;
or.b32 %r509, %r506, %r508;
shl.b32 %r510, %r476, 18;
and.b32 %r511, %r510, 1048576;
or.b32 %r512, %r509, %r511;
shl.b32 %r513, %r476, 25;
and.b32 %r514, %r513, 268435456;
or.b32 %r406, %r512, %r514;

dp4a.s32.s32 %r111, %r406, %r92, %r107;

shr.u32 %r515, %r475, 4;
and.b32 %r516, %r515, 252645135;
shr.u32 %r517, %r476, 12;
and.b32 %r518, %r517, 16;
or.b32 %r519, %r518, %r516;
shr.u32 %r520, %r476, 5;
and.b32 %r521, %r520, 4096;
or.b32 %r522, %r519, %r521;
shl.b32 %r523, %r476, 2;
and.b32 %r524, %r523, 1048576;
or.b32 %r525, %r522, %r524;
shl.b32 %r526, %r476, 9;
and.b32 %r527, %r526, 268435456;
or.b32 %r410, %r525, %r527;

dp4a.s32.s32 %r115, %r410, %r96, %r111;

ld.global.nc.u32 %r120, [%rd124+-8];

{mul.f16x2 %r119,%r120,%r100;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r119;
cvt.f32.f16 %f99, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r119;
cvt.f32.f16 %f100, high;}


cvt.rn.f32.s32 %f133, %r115;
div.approx.ftz.f32 %f134, %f100, %f130;
fma.rn.ftz.f32 %f135, %f99, %f133, %f134;
add.ftz.f32 %f384, %f135, %f384;
cvt.u64.u32 %rd34, %r822;
add.s64 %rd35, %rd1, %rd34;
mul.lo.s64 %rd36, %rd35, 36;
add.s64 %rd37, %rd6, %rd36;
add.s64 %rd38, %rd37, %rd30;
ld.global.nc.u32 %r126, [%rd38+4];
ld.global.nc.u32 %r130, [%rd38+20];
ld.global.nc.u32 %r134, [%rd38+8];
ld.global.nc.u32 %r138, [%rd38+24];

dp4a.s32.s32 %r124, %r377, %r126, %r400;


dp4a.s32.s32 %r128, %r381, %r130, %r124;


dp4a.s32.s32 %r132, %r385, %r134, %r128;


dp4a.s32.s32 %r136, %r389, %r138, %r132;

ld.global.nc.u32 %r142, [%rd37];

{mul.f16x2 %r140,%r99,%r142;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r140;
cvt.f32.f16 %f101, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r140;
cvt.f32.f16 %f102, high;}


cvt.rn.f32.s32 %f136, %r136;
div.approx.ftz.f32 %f137, %f102, %f130;
fma.rn.ftz.f32 %f138, %f101, %f136, %f137;
add.ftz.f32 %f383, %f138, %f383;

dp4a.s32.s32 %r145, %r398, %r126, %r400;


dp4a.s32.s32 %r149, %r402, %r130, %r145;


dp4a.s32.s32 %r153, %r406, %r134, %r149;


dp4a.s32.s32 %r157, %r410, %r138, %r153;


{mul.f16x2 %r161,%r120,%r142;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r161;
cvt.f32.f16 %f103, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r161;
cvt.f32.f16 %f104, high;}


cvt.rn.f32.s32 %f139, %r157;
div.approx.ftz.f32 %f140, %f104, %f130;
fma.rn.ftz.f32 %f141, %f103, %f139, %f140;
add.ftz.f32 %f382, %f141, %f382;
cvt.u64.u32 %rd39, %r828;
add.s64 %rd40, %rd1, %rd39;
mul.lo.s64 %rd41, %rd40, 36;
add.s64 %rd42, %rd6, %rd41;
add.s64 %rd43, %rd42, %rd30;
ld.global.nc.u32 %r168, [%rd43+4];
ld.global.nc.u32 %r172, [%rd43+20];
ld.global.nc.u32 %r176, [%rd43+8];
ld.global.nc.u32 %r180, [%rd43+24];

dp4a.s32.s32 %r166, %r377, %r168, %r400;


dp4a.s32.s32 %r170, %r381, %r172, %r166;


dp4a.s32.s32 %r174, %r385, %r176, %r170;


dp4a.s32.s32 %r178, %r389, %r180, %r174;

ld.global.nc.u32 %r184, [%rd42];

{mul.f16x2 %r182,%r99,%r184;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r182;
cvt.f32.f16 %f105, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r182;
cvt.f32.f16 %f106, high;}


cvt.rn.f32.s32 %f142, %r178;
div.approx.ftz.f32 %f143, %f106, %f130;
fma.rn.ftz.f32 %f144, %f105, %f142, %f143;
add.ftz.f32 %f381, %f144, %f381;

dp4a.s32.s32 %r187, %r398, %r168, %r400;


dp4a.s32.s32 %r191, %r402, %r172, %r187;


dp4a.s32.s32 %r195, %r406, %r176, %r191;


dp4a.s32.s32 %r199, %r410, %r180, %r195;


{mul.f16x2 %r203,%r120,%r184;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r203;
cvt.f32.f16 %f107, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r203;
cvt.f32.f16 %f108, high;}


cvt.rn.f32.s32 %f145, %r199;
div.approx.ftz.f32 %f146, %f108, %f130;
fma.rn.ftz.f32 %f147, %f107, %f145, %f146;
add.ftz.f32 %f380, %f147, %f380;
cvt.u64.u32 %rd44, %r827;
add.s64 %rd45, %rd1, %rd44;
mul.lo.s64 %rd46, %rd45, 36;
add.s64 %rd47, %rd6, %rd46;
add.s64 %rd48, %rd47, %rd30;
ld.global.nc.u32 %r210, [%rd48+4];
ld.global.nc.u32 %r214, [%rd48+20];
ld.global.nc.u32 %r218, [%rd48+8];
ld.global.nc.u32 %r222, [%rd48+24];

dp4a.s32.s32 %r208, %r377, %r210, %r400;


dp4a.s32.s32 %r212, %r381, %r214, %r208;


dp4a.s32.s32 %r216, %r385, %r218, %r212;


dp4a.s32.s32 %r220, %r389, %r222, %r216;

ld.global.nc.u32 %r226, [%rd47];

{mul.f16x2 %r224,%r99,%r226;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r224;
cvt.f32.f16 %f109, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r224;
cvt.f32.f16 %f110, high;}


cvt.rn.f32.s32 %f148, %r220;
div.approx.ftz.f32 %f149, %f110, %f130;
fma.rn.ftz.f32 %f150, %f109, %f148, %f149;
add.ftz.f32 %f379, %f150, %f379;

dp4a.s32.s32 %r229, %r398, %r210, %r400;


dp4a.s32.s32 %r233, %r402, %r214, %r229;


dp4a.s32.s32 %r237, %r406, %r218, %r233;


dp4a.s32.s32 %r241, %r410, %r222, %r237;


{mul.f16x2 %r245,%r120,%r226;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r245;
cvt.f32.f16 %f111, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r245;
cvt.f32.f16 %f112, high;}


cvt.rn.f32.s32 %f151, %r241;
div.approx.ftz.f32 %f152, %f112, %f130;
fma.rn.ftz.f32 %f153, %f111, %f151, %f152;
add.ftz.f32 %f378, %f153, %f378;
cvt.u64.u32 %rd49, %r826;
add.s64 %rd50, %rd1, %rd49;
mul.lo.s64 %rd51, %rd50, 36;
add.s64 %rd52, %rd6, %rd51;
add.s64 %rd53, %rd52, %rd30;
ld.global.nc.u32 %r252, [%rd53+4];
ld.global.nc.u32 %r256, [%rd53+20];
ld.global.nc.u32 %r260, [%rd53+8];
ld.global.nc.u32 %r264, [%rd53+24];

dp4a.s32.s32 %r250, %r377, %r252, %r400;


dp4a.s32.s32 %r254, %r381, %r256, %r250;


dp4a.s32.s32 %r258, %r385, %r260, %r254;


dp4a.s32.s32 %r262, %r389, %r264, %r258;

ld.global.nc.u32 %r268, [%rd52];

{mul.f16x2 %r266,%r99,%r268;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r266;
cvt.f32.f16 %f113, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r266;
cvt.f32.f16 %f114, high;}


cvt.rn.f32.s32 %f154, %r262;
div.approx.ftz.f32 %f155, %f114, %f130;
fma.rn.ftz.f32 %f156, %f113, %f154, %f155;
add.ftz.f32 %f377, %f156, %f377;

dp4a.s32.s32 %r271, %r398, %r252, %r400;


dp4a.s32.s32 %r275, %r402, %r256, %r271;


dp4a.s32.s32 %r279, %r406, %r260, %r275;


dp4a.s32.s32 %r283, %r410, %r264, %r279;


{mul.f16x2 %r287,%r120,%r268;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r287;
cvt.f32.f16 %f115, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r287;
cvt.f32.f16 %f116, high;}


cvt.rn.f32.s32 %f157, %r283;
div.approx.ftz.f32 %f158, %f116, %f130;
fma.rn.ftz.f32 %f159, %f115, %f157, %f158;
add.ftz.f32 %f376, %f159, %f376;
cvt.u64.u32 %rd54, %r825;
add.s64 %rd55, %rd1, %rd54;
mul.lo.s64 %rd56, %rd55, 36;
add.s64 %rd57, %rd6, %rd56;
add.s64 %rd58, %rd57, %rd30;
ld.global.nc.u32 %r294, [%rd58+4];
ld.global.nc.u32 %r298, [%rd58+20];
ld.global.nc.u32 %r302, [%rd58+8];
ld.global.nc.u32 %r306, [%rd58+24];

dp4a.s32.s32 %r292, %r377, %r294, %r400;


dp4a.s32.s32 %r296, %r381, %r298, %r292;


dp4a.s32.s32 %r300, %r385, %r302, %r296;


dp4a.s32.s32 %r304, %r389, %r306, %r300;

ld.global.nc.u32 %r310, [%rd57];

{mul.f16x2 %r308,%r99,%r310;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r308;
cvt.f32.f16 %f117, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r308;
cvt.f32.f16 %f118, high;}


cvt.rn.f32.s32 %f160, %r304;
div.approx.ftz.f32 %f161, %f118, %f130;
fma.rn.ftz.f32 %f162, %f117, %f160, %f161;
add.ftz.f32 %f375, %f162, %f375;

dp4a.s32.s32 %r313, %r398, %r294, %r400;


dp4a.s32.s32 %r317, %r402, %r298, %r313;


dp4a.s32.s32 %r321, %r406, %r302, %r317;


dp4a.s32.s32 %r325, %r410, %r306, %r321;


{mul.f16x2 %r329,%r120,%r310;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r329;
cvt.f32.f16 %f119, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r329;
cvt.f32.f16 %f120, high;}


cvt.rn.f32.s32 %f163, %r325;
div.approx.ftz.f32 %f164, %f120, %f130;
fma.rn.ftz.f32 %f165, %f119, %f163, %f164;
add.ftz.f32 %f374, %f165, %f374;
cvt.u64.u32 %rd59, %r824;
add.s64 %rd60, %rd1, %rd59;
mul.lo.s64 %rd61, %rd60, 36;
add.s64 %rd62, %rd6, %rd61;
add.s64 %rd63, %rd62, %rd30;
ld.global.nc.u32 %r336, [%rd63+4];
ld.global.nc.u32 %r340, [%rd63+20];
ld.global.nc.u32 %r344, [%rd63+8];
ld.global.nc.u32 %r348, [%rd63+24];

dp4a.s32.s32 %r334, %r377, %r336, %r400;


dp4a.s32.s32 %r338, %r381, %r340, %r334;


dp4a.s32.s32 %r342, %r385, %r344, %r338;


dp4a.s32.s32 %r346, %r389, %r348, %r342;

ld.global.nc.u32 %r352, [%rd62];

{mul.f16x2 %r350,%r99,%r352;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r350;
cvt.f32.f16 %f121, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r350;
cvt.f32.f16 %f122, high;}


cvt.rn.f32.s32 %f166, %r346;
div.approx.ftz.f32 %f167, %f122, %f130;
fma.rn.ftz.f32 %f168, %f121, %f166, %f167;
add.ftz.f32 %f373, %f168, %f373;

dp4a.s32.s32 %r355, %r398, %r336, %r400;


dp4a.s32.s32 %r359, %r402, %r340, %r355;


dp4a.s32.s32 %r363, %r406, %r344, %r359;


dp4a.s32.s32 %r367, %r410, %r348, %r363;


{mul.f16x2 %r371,%r120,%r352;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r371;
cvt.f32.f16 %f123, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r371;
cvt.f32.f16 %f124, high;}


cvt.rn.f32.s32 %f169, %r367;
div.approx.ftz.f32 %f170, %f124, %f130;
fma.rn.ftz.f32 %f171, %f123, %f169, %f170;
add.ftz.f32 %f372, %f171, %f372;
cvt.u64.u32 %rd64, %r823;
add.s64 %rd65, %rd1, %rd64;
mul.lo.s64 %rd66, %rd65, 36;
add.s64 %rd67, %rd6, %rd66;
add.s64 %rd68, %rd67, %rd30;
ld.global.nc.u32 %r378, [%rd68+4];
ld.global.nc.u32 %r382, [%rd68+20];
ld.global.nc.u32 %r386, [%rd68+8];
ld.global.nc.u32 %r390, [%rd68+24];

dp4a.s32.s32 %r376, %r377, %r378, %r400;


dp4a.s32.s32 %r380, %r381, %r382, %r376;


dp4a.s32.s32 %r384, %r385, %r386, %r380;


dp4a.s32.s32 %r388, %r389, %r390, %r384;

ld.global.nc.u32 %r394, [%rd67];

{mul.f16x2 %r392,%r99,%r394;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r392;
cvt.f32.f16 %f125, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r392;
cvt.f32.f16 %f126, high;}


cvt.rn.f32.s32 %f172, %r388;
div.approx.ftz.f32 %f173, %f126, %f130;
fma.rn.ftz.f32 %f174, %f125, %f172, %f173;
add.ftz.f32 %f371, %f174, %f371;

dp4a.s32.s32 %r397, %r398, %r378, %r400;


dp4a.s32.s32 %r401, %r402, %r382, %r397;


dp4a.s32.s32 %r405, %r406, %r386, %r401;


dp4a.s32.s32 %r409, %r410, %r390, %r405;


{mul.f16x2 %r413,%r120,%r394;
}


{.reg .f16 low,high;
mov.b32 {low,high},%r413;
cvt.f32.f16 %f127, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r413;
cvt.f32.f16 %f128, high;}


cvt.rn.f32.s32 %f175, %r409;
div.approx.ftz.f32 %f176, %f128, %f130;
fma.rn.ftz.f32 %f177, %f127, %f175, %f176;
add.ftz.f32 %f370, %f177, %f370;
add.s32 %r828, %r828, 32;
add.s32 %r827, %r827, 32;
add.s32 %r826, %r826, 32;
add.s32 %r825, %r825, 32;
add.s32 %r824, %r824, 32;
add.s32 %r823, %r823, 32;
add.s32 %r822, %r822, 32;
add.s64 %rd124, %rd124, 768;
add.s64 %rd123, %rd123, 768;
add.s32 %r829, %r829, 32;
setp.lt.s32 %p4, %r829, %r6;
@%p4 bra $L__BB85_2;

$L__BB85_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB85_5;

shl.b32 %r528, %r1, 11;
mov.u32 %r529, _ZZ13mul_mat_vec_qIL9ggml_type7ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r530, %r529, %r528;
shl.b32 %r531, %r2, 2;
add.s32 %r532, %r530, %r531;
st.shared.f32 [%r532+-2048], %f385;
st.shared.f32 [%r532+-1920], %f384;
st.shared.f32 [%r532+-1792], %f383;
st.shared.f32 [%r532+-1664], %f382;
st.shared.f32 [%r532+-1536], %f381;
st.shared.f32 [%r532+-1408], %f380;
st.shared.f32 [%r532+-1280], %f379;
st.shared.f32 [%r532+-1152], %f378;
st.shared.f32 [%r532+-1024], %f377;
st.shared.f32 [%r532+-896], %f376;
st.shared.f32 [%r532+-768], %f375;
st.shared.f32 [%r532+-640], %f374;
st.shared.f32 [%r532+-512], %f373;
st.shared.f32 [%r532+-384], %f372;
st.shared.f32 [%r532+-256], %f371;
st.shared.f32 [%r532+-128], %f370;

$L__BB85_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB85_38;

mov.u32 %r533, %ctaid.z;
mov.u32 %r534, %ctaid.y;
mad.lo.s32 %r535, %r534, %r44, %r4;
mad.lo.s32 %r536, %r533, %r50, %r535;
cvt.u64.u32 %rd11, %r536;
shl.b32 %r537, %r2, 2;
mov.u32 %r538, _ZZ13mul_mat_vec_qIL9ggml_type7ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r32, %r538, %r537;
mov.u32 %r539, 2;
ld.shared.f32 %f178, [%r32];
add.ftz.f32 %f179, %f178, %f385;
mov.b32 %r540, %f179;
mov.u32 %r541, 31;
mov.u32 %r542, 16;
mov.u32 %r543, -1;
shfl.sync.bfly.b32 %r544|%p7, %r540, %r542, %r541, %r543;
mov.b32 %f180, %r544;
add.ftz.f32 %f181, %f179, %f180;
mov.b32 %r545, %f181;
mov.u32 %r546, 8;
shfl.sync.bfly.b32 %r547|%p8, %r545, %r546, %r541, %r543;
mov.b32 %f182, %r547;
add.ftz.f32 %f183, %f181, %f182;
mov.b32 %r548, %f183;
mov.u32 %r549, 4;
shfl.sync.bfly.b32 %r550|%p9, %r548, %r549, %r541, %r543;
mov.b32 %f184, %r550;
add.ftz.f32 %f185, %f183, %f184;
mov.b32 %r551, %f185;
shfl.sync.bfly.b32 %r552|%p10, %r551, %r539, %r541, %r543;
mov.b32 %f186, %r552;
add.ftz.f32 %f187, %f185, %f186;
mov.b32 %r553, %f187;
mov.u32 %r554, 1;
shfl.sync.bfly.b32 %r555|%p11, %r553, %r554, %r541, %r543;
mov.b32 %f188, %r555;
add.ftz.f32 %f49, %f187, %f188;
setp.lt.u32 %p12, %r4, %r38;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
cvta.to.global.u64 %rd69, %rd18;
mul.wide.u32 %rd70, %r536, 4;
add.s64 %rd12, %rd69, %rd70;
not.pred %p14, %p1;
@%p14 bra $L__BB85_8;

st.global.f32 [%rd12], %f49;

$L__BB85_8:
ld.shared.f32 %f189, [%r32+128];
add.ftz.f32 %f190, %f189, %f384;
mov.b32 %r556, %f190;
shfl.sync.bfly.b32 %r560|%p15, %r556, %r542, %r541, %r543;
mov.b32 %f191, %r560;
add.ftz.f32 %f192, %f190, %f191;
mov.b32 %r561, %f192;
shfl.sync.bfly.b32 %r563|%p16, %r561, %r546, %r541, %r543;
mov.b32 %f193, %r563;
add.ftz.f32 %f194, %f192, %f193;
mov.b32 %r564, %f194;
shfl.sync.bfly.b32 %r566|%p17, %r564, %r549, %r541, %r543;
mov.b32 %f195, %r566;
add.ftz.f32 %f196, %f194, %f195;
mov.b32 %r567, %f196;
shfl.sync.bfly.b32 %r569|%p18, %r567, %r539, %r541, %r543;
mov.b32 %f197, %r569;
add.ftz.f32 %f198, %f196, %f197;
mov.b32 %r570, %f198;
shfl.sync.bfly.b32 %r572|%p19, %r570, %r554, %r541, %r543;
mov.b32 %f199, %r572;
add.ftz.f32 %f50, %f198, %f199;
setp.eq.s32 %p20, %r2, 1;
or.b32 %r576, %r4, 1;
setp.lt.u32 %p21, %r576, %r38;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB85_10;

st.global.f32 [%rd12+4], %f50;

$L__BB85_10:
ld.shared.f32 %f200, [%r32+256];
add.ftz.f32 %f201, %f200, %f383;
mov.b32 %r577, %f201;
mov.u32 %r578, 31;
mov.u32 %r579, 16;
mov.u32 %r580, -1;
shfl.sync.bfly.b32 %r581|%p23, %r577, %r579, %r578, %r580;
mov.b32 %f202, %r581;
add.ftz.f32 %f203, %f201, %f202;
mov.b32 %r582, %f203;
mov.u32 %r583, 8;
shfl.sync.bfly.b32 %r584|%p24, %r582, %r583, %r578, %r580;
mov.b32 %f204, %r584;
add.ftz.f32 %f205, %f203, %f204;
mov.b32 %r585, %f205;
mov.u32 %r586, 4;
shfl.sync.bfly.b32 %r587|%p25, %r585, %r586, %r578, %r580;
mov.b32 %f206, %r587;
add.ftz.f32 %f207, %f205, %f206;
mov.b32 %r588, %f207;
mov.u32 %r589, 2;
shfl.sync.bfly.b32 %r590|%p26, %r588, %r589, %r578, %r580;
mov.b32 %f208, %r590;
add.ftz.f32 %f209, %f207, %f208;
mov.b32 %r591, %f209;
mov.u32 %r592, 1;
shfl.sync.bfly.b32 %r593|%p27, %r591, %r592, %r578, %r580;
mov.b32 %f210, %r593;
add.ftz.f32 %f51, %f209, %f210;
@%p14 bra $L__BB85_12;

cvt.u64.u32 %rd71, %r38;
add.s64 %rd72, %rd71, %rd11;
shl.b64 %rd74, %rd72, 2;
add.s64 %rd75, %rd69, %rd74;
st.global.f32 [%rd75], %f51;

$L__BB85_12:
ld.shared.f32 %f211, [%r32+384];
add.ftz.f32 %f212, %f211, %f382;
mov.b32 %r594, %f212;
shfl.sync.bfly.b32 %r598|%p29, %r594, %r579, %r578, %r580;
mov.b32 %f213, %r598;
add.ftz.f32 %f214, %f212, %f213;
mov.b32 %r599, %f214;
shfl.sync.bfly.b32 %r601|%p30, %r599, %r583, %r578, %r580;
mov.b32 %f215, %r601;
add.ftz.f32 %f216, %f214, %f215;
mov.b32 %r602, %f216;
shfl.sync.bfly.b32 %r604|%p31, %r602, %r586, %r578, %r580;
mov.b32 %f217, %r604;
add.ftz.f32 %f218, %f216, %f217;
mov.b32 %r605, %f218;
shfl.sync.bfly.b32 %r607|%p32, %r605, %r589, %r578, %r580;
mov.b32 %f219, %r607;
add.ftz.f32 %f220, %f218, %f219;
mov.b32 %r608, %f220;
shfl.sync.bfly.b32 %r610|%p33, %r608, %r592, %r578, %r580;
mov.b32 %f221, %r610;
add.ftz.f32 %f52, %f220, %f221;
@%p22 bra $L__BB85_14;

add.s32 %r611, %r38, 1;
cvt.u64.u32 %rd76, %r611;
add.s64 %rd77, %rd76, %rd11;
shl.b64 %rd79, %rd77, 2;
add.s64 %rd80, %rd69, %rd79;
st.global.f32 [%rd80], %f52;

$L__BB85_14:
shl.b32 %r612, %r38, 1;
mov.u32 %r613, 1;
ld.shared.f32 %f222, [%r32+512];
add.ftz.f32 %f223, %f222, %f381;
mov.b32 %r614, %f223;
mov.u32 %r615, 31;
mov.u32 %r616, 16;
mov.u32 %r617, -1;
shfl.sync.bfly.b32 %r618|%p35, %r614, %r616, %r615, %r617;
mov.b32 %f224, %r618;
add.ftz.f32 %f225, %f223, %f224;
mov.b32 %r619, %f225;
mov.u32 %r620, 8;
shfl.sync.bfly.b32 %r621|%p36, %r619, %r620, %r615, %r617;
mov.b32 %f226, %r621;
add.ftz.f32 %f227, %f225, %f226;
mov.b32 %r622, %f227;
mov.u32 %r623, 4;
shfl.sync.bfly.b32 %r624|%p37, %r622, %r623, %r615, %r617;
mov.b32 %f228, %r624;
add.ftz.f32 %f229, %f227, %f228;
mov.b32 %r625, %f229;
mov.u32 %r626, 2;
shfl.sync.bfly.b32 %r627|%p38, %r625, %r626, %r615, %r617;
mov.b32 %f230, %r627;
add.ftz.f32 %f231, %f229, %f230;
mov.b32 %r628, %f231;
shfl.sync.bfly.b32 %r629|%p39, %r628, %r613, %r615, %r617;
mov.b32 %f232, %r629;
add.ftz.f32 %f53, %f231, %f232;
cvt.u64.u32 %rd81, %r612;
add.s64 %rd82, %rd81, %rd11;
shl.b64 %rd84, %rd82, 2;
add.s64 %rd13, %rd69, %rd84;
@%p14 bra $L__BB85_16;

st.global.f32 [%rd13], %f53;

$L__BB85_16:
ld.shared.f32 %f233, [%r32+640];
add.ftz.f32 %f234, %f233, %f380;
mov.b32 %r630, %f234;
shfl.sync.bfly.b32 %r634|%p41, %r630, %r616, %r615, %r617;
mov.b32 %f235, %r634;
add.ftz.f32 %f236, %f234, %f235;
mov.b32 %r635, %f236;
shfl.sync.bfly.b32 %r637|%p42, %r635, %r620, %r615, %r617;
mov.b32 %f237, %r637;
add.ftz.f32 %f238, %f236, %f237;
mov.b32 %r638, %f238;
shfl.sync.bfly.b32 %r640|%p43, %r638, %r623, %r615, %r617;
mov.b32 %f239, %r640;
add.ftz.f32 %f240, %f238, %f239;
mov.b32 %r641, %f240;
shfl.sync.bfly.b32 %r643|%p44, %r641, %r626, %r615, %r617;
mov.b32 %f241, %r643;
add.ftz.f32 %f242, %f240, %f241;
mov.b32 %r644, %f242;
shfl.sync.bfly.b32 %r646|%p45, %r644, %r613, %r615, %r617;
mov.b32 %f243, %r646;
add.ftz.f32 %f54, %f242, %f243;
@%p22 bra $L__BB85_18;

st.global.f32 [%rd13+4], %f54;

$L__BB85_18:
mul.lo.s32 %r33, %r38, 3;
ld.shared.f32 %f244, [%r32+768];
add.ftz.f32 %f245, %f244, %f379;
mov.b32 %r647, %f245;
mov.u32 %r648, 31;
mov.u32 %r649, 16;
mov.u32 %r650, -1;
shfl.sync.bfly.b32 %r651|%p47, %r647, %r649, %r648, %r650;
mov.b32 %f246, %r651;
add.ftz.f32 %f247, %f245, %f246;
mov.b32 %r652, %f247;
mov.u32 %r653, 8;
shfl.sync.bfly.b32 %r654|%p48, %r652, %r653, %r648, %r650;
mov.b32 %f248, %r654;
add.ftz.f32 %f249, %f247, %f248;
mov.b32 %r655, %f249;
mov.u32 %r656, 4;
shfl.sync.bfly.b32 %r657|%p49, %r655, %r656, %r648, %r650;
mov.b32 %f250, %r657;
add.ftz.f32 %f251, %f249, %f250;
mov.b32 %r658, %f251;
mov.u32 %r659, 2;
shfl.sync.bfly.b32 %r660|%p50, %r658, %r659, %r648, %r650;
mov.b32 %f252, %r660;
add.ftz.f32 %f253, %f251, %f252;
mov.b32 %r661, %f253;
mov.u32 %r662, 1;
shfl.sync.bfly.b32 %r663|%p51, %r661, %r662, %r648, %r650;
mov.b32 %f254, %r663;
add.ftz.f32 %f55, %f253, %f254;
@%p14 bra $L__BB85_20;

cvt.u64.u32 %rd85, %r33;
add.s64 %rd86, %rd85, %rd11;
shl.b64 %rd88, %rd86, 2;
add.s64 %rd89, %rd69, %rd88;
st.global.f32 [%rd89], %f55;

$L__BB85_20:
ld.shared.f32 %f255, [%r32+896];
add.ftz.f32 %f256, %f255, %f378;
mov.b32 %r664, %f256;
shfl.sync.bfly.b32 %r668|%p53, %r664, %r649, %r648, %r650;
mov.b32 %f257, %r668;
add.ftz.f32 %f258, %f256, %f257;
mov.b32 %r669, %f258;
shfl.sync.bfly.b32 %r671|%p54, %r669, %r653, %r648, %r650;
mov.b32 %f259, %r671;
add.ftz.f32 %f260, %f258, %f259;
mov.b32 %r672, %f260;
shfl.sync.bfly.b32 %r674|%p55, %r672, %r656, %r648, %r650;
mov.b32 %f261, %r674;
add.ftz.f32 %f262, %f260, %f261;
mov.b32 %r675, %f262;
shfl.sync.bfly.b32 %r677|%p56, %r675, %r659, %r648, %r650;
mov.b32 %f263, %r677;
add.ftz.f32 %f264, %f262, %f263;
mov.b32 %r678, %f264;
shfl.sync.bfly.b32 %r680|%p57, %r678, %r662, %r648, %r650;
mov.b32 %f265, %r680;
add.ftz.f32 %f56, %f264, %f265;
@%p22 bra $L__BB85_22;

add.s32 %r681, %r33, 1;
cvt.u64.u32 %rd90, %r681;
add.s64 %rd91, %rd90, %rd11;
shl.b64 %rd93, %rd91, 2;
add.s64 %rd94, %rd69, %rd93;
st.global.f32 [%rd94], %f56;

$L__BB85_22:
shl.b32 %r682, %r38, 2;
mov.u32 %r683, 2;
ld.shared.f32 %f266, [%r32+1024];
add.ftz.f32 %f267, %f266, %f377;
mov.b32 %r684, %f267;
mov.u32 %r685, 31;
mov.u32 %r686, 16;
mov.u32 %r687, -1;
shfl.sync.bfly.b32 %r688|%p59, %r684, %r686, %r685, %r687;
mov.b32 %f268, %r688;
add.ftz.f32 %f269, %f267, %f268;
mov.b32 %r689, %f269;
mov.u32 %r690, 8;
shfl.sync.bfly.b32 %r691|%p60, %r689, %r690, %r685, %r687;
mov.b32 %f270, %r691;
add.ftz.f32 %f271, %f269, %f270;
mov.b32 %r692, %f271;
mov.u32 %r693, 4;
shfl.sync.bfly.b32 %r694|%p61, %r692, %r693, %r685, %r687;
mov.b32 %f272, %r694;
add.ftz.f32 %f273, %f271, %f272;
mov.b32 %r695, %f273;
shfl.sync.bfly.b32 %r696|%p62, %r695, %r683, %r685, %r687;
mov.b32 %f274, %r696;
add.ftz.f32 %f275, %f273, %f274;
mov.b32 %r697, %f275;
mov.u32 %r698, 1;
shfl.sync.bfly.b32 %r699|%p63, %r697, %r698, %r685, %r687;
mov.b32 %f276, %r699;
add.ftz.f32 %f57, %f275, %f276;
cvt.u64.u32 %rd95, %r682;
add.s64 %rd96, %rd95, %rd11;
shl.b64 %rd98, %rd96, 2;
add.s64 %rd14, %rd69, %rd98;
@%p14 bra $L__BB85_24;

st.global.f32 [%rd14], %f57;

$L__BB85_24:
ld.shared.f32 %f277, [%r32+1152];
add.ftz.f32 %f278, %f277, %f376;
mov.b32 %r700, %f278;
shfl.sync.bfly.b32 %r704|%p65, %r700, %r686, %r685, %r687;
mov.b32 %f279, %r704;
add.ftz.f32 %f280, %f278, %f279;
mov.b32 %r705, %f280;
shfl.sync.bfly.b32 %r707|%p66, %r705, %r690, %r685, %r687;
mov.b32 %f281, %r707;
add.ftz.f32 %f282, %f280, %f281;
mov.b32 %r708, %f282;
shfl.sync.bfly.b32 %r710|%p67, %r708, %r693, %r685, %r687;
mov.b32 %f283, %r710;
add.ftz.f32 %f284, %f282, %f283;
mov.b32 %r711, %f284;
shfl.sync.bfly.b32 %r713|%p68, %r711, %r683, %r685, %r687;
mov.b32 %f285, %r713;
add.ftz.f32 %f286, %f284, %f285;
mov.b32 %r714, %f286;
shfl.sync.bfly.b32 %r716|%p69, %r714, %r698, %r685, %r687;
mov.b32 %f287, %r716;
add.ftz.f32 %f58, %f286, %f287;
@%p22 bra $L__BB85_26;

st.global.f32 [%rd14+4], %f58;

$L__BB85_26:
mul.lo.s32 %r34, %r38, 5;
ld.shared.f32 %f288, [%r32+1280];
add.ftz.f32 %f289, %f288, %f375;
mov.b32 %r717, %f289;
mov.u32 %r718, 31;
mov.u32 %r719, 16;
mov.u32 %r720, -1;
shfl.sync.bfly.b32 %r721|%p71, %r717, %r719, %r718, %r720;
mov.b32 %f290, %r721;
add.ftz.f32 %f291, %f289, %f290;
mov.b32 %r722, %f291;
mov.u32 %r723, 8;
shfl.sync.bfly.b32 %r724|%p72, %r722, %r723, %r718, %r720;
mov.b32 %f292, %r724;
add.ftz.f32 %f293, %f291, %f292;
mov.b32 %r725, %f293;
mov.u32 %r726, 4;
shfl.sync.bfly.b32 %r727|%p73, %r725, %r726, %r718, %r720;
mov.b32 %f294, %r727;
add.ftz.f32 %f295, %f293, %f294;
mov.b32 %r728, %f295;
mov.u32 %r729, 2;
shfl.sync.bfly.b32 %r730|%p74, %r728, %r729, %r718, %r720;
mov.b32 %f296, %r730;
add.ftz.f32 %f297, %f295, %f296;
mov.b32 %r731, %f297;
mov.u32 %r732, 1;
shfl.sync.bfly.b32 %r733|%p75, %r731, %r732, %r718, %r720;
mov.b32 %f298, %r733;
add.ftz.f32 %f59, %f297, %f298;
@%p14 bra $L__BB85_28;

cvt.u64.u32 %rd99, %r34;
add.s64 %rd100, %rd99, %rd11;
shl.b64 %rd102, %rd100, 2;
add.s64 %rd103, %rd69, %rd102;
st.global.f32 [%rd103], %f59;

$L__BB85_28:
ld.shared.f32 %f299, [%r32+1408];
add.ftz.f32 %f300, %f299, %f374;
mov.b32 %r734, %f300;
shfl.sync.bfly.b32 %r738|%p77, %r734, %r719, %r718, %r720;
mov.b32 %f301, %r738;
add.ftz.f32 %f302, %f300, %f301;
mov.b32 %r739, %f302;
shfl.sync.bfly.b32 %r741|%p78, %r739, %r723, %r718, %r720;
mov.b32 %f303, %r741;
add.ftz.f32 %f304, %f302, %f303;
mov.b32 %r742, %f304;
shfl.sync.bfly.b32 %r744|%p79, %r742, %r726, %r718, %r720;
mov.b32 %f305, %r744;
add.ftz.f32 %f306, %f304, %f305;
mov.b32 %r745, %f306;
shfl.sync.bfly.b32 %r747|%p80, %r745, %r729, %r718, %r720;
mov.b32 %f307, %r747;
add.ftz.f32 %f308, %f306, %f307;
mov.b32 %r748, %f308;
shfl.sync.bfly.b32 %r750|%p81, %r748, %r732, %r718, %r720;
mov.b32 %f309, %r750;
add.ftz.f32 %f60, %f308, %f309;
@%p22 bra $L__BB85_30;

add.s32 %r751, %r34, 1;
cvt.u64.u32 %rd104, %r751;
add.s64 %rd105, %rd104, %rd11;
shl.b64 %rd107, %rd105, 2;
add.s64 %rd108, %rd69, %rd107;
st.global.f32 [%rd108], %f60;

$L__BB85_30:
mul.lo.s32 %r752, %r38, 6;
ld.shared.f32 %f310, [%r32+1536];
add.ftz.f32 %f311, %f310, %f373;
mov.b32 %r753, %f311;
mov.u32 %r754, 31;
mov.u32 %r755, 16;
mov.u32 %r756, -1;
shfl.sync.bfly.b32 %r757|%p83, %r753, %r755, %r754, %r756;
mov.b32 %f312, %r757;
add.ftz.f32 %f313, %f311, %f312;
mov.b32 %r758, %f313;
mov.u32 %r759, 8;
shfl.sync.bfly.b32 %r760|%p84, %r758, %r759, %r754, %r756;
mov.b32 %f314, %r760;
add.ftz.f32 %f315, %f313, %f314;
mov.b32 %r761, %f315;
mov.u32 %r762, 4;
shfl.sync.bfly.b32 %r763|%p85, %r761, %r762, %r754, %r756;
mov.b32 %f316, %r763;
add.ftz.f32 %f317, %f315, %f316;
mov.b32 %r764, %f317;
mov.u32 %r765, 2;
shfl.sync.bfly.b32 %r766|%p86, %r764, %r765, %r754, %r756;
mov.b32 %f318, %r766;
add.ftz.f32 %f319, %f317, %f318;
mov.b32 %r767, %f319;
mov.u32 %r768, 1;
shfl.sync.bfly.b32 %r769|%p87, %r767, %r768, %r754, %r756;
mov.b32 %f320, %r769;
add.ftz.f32 %f61, %f319, %f320;
cvt.u64.u32 %rd109, %r752;
add.s64 %rd110, %rd109, %rd11;
shl.b64 %rd112, %rd110, 2;
add.s64 %rd15, %rd69, %rd112;
@%p14 bra $L__BB85_32;

st.global.f32 [%rd15], %f61;

$L__BB85_32:
ld.shared.f32 %f321, [%r32+1664];
add.ftz.f32 %f322, %f321, %f372;
mov.b32 %r770, %f322;
shfl.sync.bfly.b32 %r774|%p89, %r770, %r755, %r754, %r756;
mov.b32 %f323, %r774;
add.ftz.f32 %f324, %f322, %f323;
mov.b32 %r775, %f324;
shfl.sync.bfly.b32 %r777|%p90, %r775, %r759, %r754, %r756;
mov.b32 %f325, %r777;
add.ftz.f32 %f326, %f324, %f325;
mov.b32 %r778, %f326;
shfl.sync.bfly.b32 %r780|%p91, %r778, %r762, %r754, %r756;
mov.b32 %f327, %r780;
add.ftz.f32 %f328, %f326, %f327;
mov.b32 %r781, %f328;
shfl.sync.bfly.b32 %r783|%p92, %r781, %r765, %r754, %r756;
mov.b32 %f329, %r783;
add.ftz.f32 %f330, %f328, %f329;
mov.b32 %r784, %f330;
shfl.sync.bfly.b32 %r786|%p93, %r784, %r768, %r754, %r756;
mov.b32 %f331, %r786;
add.ftz.f32 %f62, %f330, %f331;
@%p22 bra $L__BB85_34;

st.global.f32 [%rd15+4], %f62;

$L__BB85_34:
mul.lo.s32 %r35, %r38, 7;
ld.shared.f32 %f332, [%r32+1792];
add.ftz.f32 %f333, %f332, %f371;
mov.b32 %r787, %f333;
mov.u32 %r788, 31;
mov.u32 %r789, 16;
mov.u32 %r790, -1;
shfl.sync.bfly.b32 %r791|%p95, %r787, %r789, %r788, %r790;
mov.b32 %f334, %r791;
add.ftz.f32 %f335, %f333, %f334;
mov.b32 %r792, %f335;
mov.u32 %r793, 8;
shfl.sync.bfly.b32 %r794|%p96, %r792, %r793, %r788, %r790;
mov.b32 %f336, %r794;
add.ftz.f32 %f337, %f335, %f336;
mov.b32 %r795, %f337;
mov.u32 %r796, 4;
shfl.sync.bfly.b32 %r797|%p97, %r795, %r796, %r788, %r790;
mov.b32 %f338, %r797;
add.ftz.f32 %f339, %f337, %f338;
mov.b32 %r798, %f339;
mov.u32 %r799, 2;
shfl.sync.bfly.b32 %r800|%p98, %r798, %r799, %r788, %r790;
mov.b32 %f340, %r800;
add.ftz.f32 %f341, %f339, %f340;
mov.b32 %r801, %f341;
mov.u32 %r802, 1;
shfl.sync.bfly.b32 %r803|%p99, %r801, %r802, %r788, %r790;
mov.b32 %f342, %r803;
add.ftz.f32 %f63, %f341, %f342;
@%p14 bra $L__BB85_36;

cvt.u64.u32 %rd113, %r35;
add.s64 %rd114, %rd113, %rd11;
shl.b64 %rd116, %rd114, 2;
add.s64 %rd117, %rd69, %rd116;
st.global.f32 [%rd117], %f63;

$L__BB85_36:
ld.shared.f32 %f343, [%r32+1920];
add.ftz.f32 %f344, %f343, %f370;
mov.b32 %r804, %f344;
shfl.sync.bfly.b32 %r808|%p101, %r804, %r789, %r788, %r790;
mov.b32 %f345, %r808;
add.ftz.f32 %f346, %f344, %f345;
mov.b32 %r809, %f346;
shfl.sync.bfly.b32 %r811|%p102, %r809, %r793, %r788, %r790;
mov.b32 %f347, %r811;
add.ftz.f32 %f348, %f346, %f347;
mov.b32 %r812, %f348;
shfl.sync.bfly.b32 %r814|%p103, %r812, %r796, %r788, %r790;
mov.b32 %f349, %r814;
add.ftz.f32 %f350, %f348, %f349;
mov.b32 %r815, %f350;
shfl.sync.bfly.b32 %r817|%p104, %r815, %r799, %r788, %r790;
mov.b32 %f351, %r817;
add.ftz.f32 %f352, %f350, %f351;
mov.b32 %r818, %f352;
shfl.sync.bfly.b32 %r820|%p105, %r818, %r802, %r788, %r790;
mov.b32 %f353, %r820;
add.ftz.f32 %f64, %f352, %f353;
@%p22 bra $L__BB85_38;

add.s32 %r821, %r35, 1;
cvt.u64.u32 %rd118, %r821;
add.s64 %rd119, %rd118, %rd11;
shl.b64 %rd121, %rd119, 2;
add.s64 %rd122, %rd69, %rd121;
st.global.f32 [%rd122], %f64;

$L__BB85_38:
ret;

}
.entry _Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj(
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2,
.param .align 8 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3[48],
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5,
.param .align 4 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6[12],
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15
)
.maxntid 256, 1, 1
.minnctapersm 1
{
.local .align 8 .b8 __local_depot86[16];
.reg .b64 %SP;
.reg .b64 %SPL;
.reg .pred %p<48>;
.reg .b16 %rs<29>;
.reg .f32 %f<168>;
.reg .b32 %r<162>;
.reg .b64 %rd<74>;


mov.u64 %SPL, __local_depot86;
ld.param.u64 %rd22, [_Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0];
ld.param.u64 %rd23, [_Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1];
ld.param.u64 %rd24, [_Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2];
ld.param.u64 %rd30, [_Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4];
ld.param.u32 %r28, [_Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5];
ld.param.u32 %r20, [_Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7];
ld.param.u32 %r21, [_Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8];
ld.param.u32 %r22, [_Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9];
ld.param.u32 %r23, [_Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10];
ld.param.u32 %r24, [_Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11];
ld.param.u32 %r25, [_Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12];
ld.param.u32 %r26, [_Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13];
ld.param.u32 %r29, [_Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14];
ld.param.u32 %r27, [_Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15];
ld.param.u32 %r19, [_Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+8];
ld.param.u32 %r18, [_Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+4];
ld.param.u32 %r17, [_Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6];
ld.param.f32 %f31, [_Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+44];
ld.param.u32 %r16, [_Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
ld.param.u64 %rd27, [_Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+16];
ld.param.u64 %rd26, [_Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+8];
ld.param.u64 %rd25, [_Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3];
add.u64 %rd1, %SPL, 0;
add.u64 %rd2, %SPL, 8;
mov.u32 %r30, %ctaid.x;
shl.b32 %r1, %r30, 1;
shr.u32 %r2, %r28, 5;
mov.u32 %r3, %tid.y;
setp.ge.u32 %p1, %r3, %r29;
@%p1 bra $L__BB86_29;

mov.u32 %r4, %ctaid.y;
cvta.to.global.u64 %rd33, %rd24;
mad.lo.s32 %r31, %r3, %r27, %r4;
mul.wide.u32 %rd34, %r31, 4;
add.s64 %rd35, %rd33, %rd34;
ld.global.nc.u32 %r5, [%rd35];
mov.f32 %f161, 0f00000000;
st.local.v2.f32 [%rd1], {%f161, %f161};
st.local.v2.f32 [%rd2], {%f161, %f161};
mov.u32 %r6, %tid.x;
shr.u32 %r161, %r6, 2;
setp.ge.u32 %p2, %r161, %r2;
mov.f32 %f162, %f161;
mov.f32 %f163, %f161;
mov.f32 %f164, %f161;
@%p2 bra $L__BB86_9;

mul.lo.s32 %r32, %r5, %r24;
cvt.u64.u32 %rd36, %r6;
shr.u64 %rd37, %rd36, 2;
mul.lo.s32 %r33, %r3, %r22;
cvt.u64.u32 %rd38, %r33;
add.s64 %rd39, %rd37, %rd38;
mul.hi.u32 %r34, %r4, %r17;
add.s32 %r35, %r34, %r4;
shr.u32 %r36, %r35, %r18;
mul.lo.s32 %r37, %r36, %r19;
sub.s32 %r38, %r4, %r37;
mul.lo.s32 %r39, %r38, %r25;
cvt.u64.u32 %rd40, %r39;
add.s64 %rd41, %rd39, %rd40;
mul.lo.s64 %rd42, %rd41, 36;
cvta.to.global.u64 %rd43, %rd23;
add.s64 %rd44, %rd43, %rd42;
add.s64 %rd73, %rd44, 8;
and.b32 %r40, %r6, 3;
mul.wide.u32 %rd45, %r40, 8;
mad.lo.s32 %r41, %r1, %r21, %r161;
add.s32 %r42, %r41, %r32;
mul.wide.s32 %rd7, %r42, 34;
add.s64 %rd46, %rd45, %rd7;
add.s64 %rd8, %rd46, 8;
add.s32 %r43, %r1, 1;
add.s32 %r44, %r161, %r32;
mad.lo.s32 %r45, %r21, %r43, %r44;
mul.wide.s32 %rd9, %r45, 34;
add.s64 %rd47, %rd45, %rd9;
add.s64 %rd10, %rd47, 4;
cvta.to.global.u64 %rd72, %rd22;
cvta.to.global.u64 %rd71, %rd26;
mov.f32 %f161, 0f00000000;

$L__BB86_3:
add.s64 %rd48, %rd72, %rd8;
ld.global.nc.u16 %rs4, [%rd48+-6];
ld.global.nc.u16 %rs5, [%rd48+-4];
mov.b32 %r48, {%rs4, %rs5};
add.s64 %rd50, %rd73, %rd45;
ld.global.nc.u32 %r9, [%rd50+-4];
ld.global.nc.u16 %rs6, [%rd48+-2];
ld.global.nc.u16 %rs7, [%rd48];
mov.b32 %r52, {%rs6, %rs7};
ld.global.nc.u32 %r10, [%rd50];
add.s64 %rd51, %rd72, %rd7;
ld.global.nc.u16 %rs1, [%rd51];

{ cvt.f32.f16 %f42, %rs1;}


ld.global.nc.u32 %r11, [%rd73+-8];

{.reg .f16 low,high;
mov.b32 {low,high}, %r11;
mov.b16 %rs2, low;}


{ cvt.f32.f16 %f43, %rs2;}


mov.u32 %r50, 0;

dp4a.s32.s32 %r47, %r48, %r9, %r50;


dp4a.s32.s32 %r51, %r52, %r10, %r47;

mul.ftz.f32 %f44, %f42, %f43;
cvt.rn.f32.s32 %f45, %r51;
fma.rn.ftz.f32 %f164, %f44, %f45, %f164;
setp.eq.s64 %p3, %rd26, 0;
@%p3 bra $L__BB86_5;

add.s64 %rd52, %rd71, %rd8;
ld.global.u16 %rs11, [%rd52+-6];
ld.global.u16 %rs12, [%rd52+-4];
mov.b32 %r59, {%rs11, %rs12};
ld.global.u16 %rs13, [%rd52+-2];
ld.global.u16 %rs14, [%rd52];
mov.b32 %r63, {%rs13, %rs14};
add.s64 %rd53, %rd71, %rd7;
ld.global.u16 %rs8, [%rd53];

{ cvt.f32.f16 %f46, %rs8;}



{.reg .f16 low,high;
mov.b32 {low,high}, %r11;
mov.b16 %rs9, low;}


{ cvt.f32.f16 %f47, %rs9;}



dp4a.s32.s32 %r58, %r59, %r9, %r50;


dp4a.s32.s32 %r62, %r63, %r10, %r58;

mul.ftz.f32 %f48, %f46, %f47;
cvt.rn.f32.s32 %f49, %r62;
fma.rn.ftz.f32 %f162, %f48, %f49, %f162;

$L__BB86_5:
add.s64 %rd54, %rd72, %rd10;
ld.global.nc.u16 %rs18, [%rd54+-2];
ld.global.nc.u16 %rs19, [%rd54];
mov.b32 %r68, {%rs18, %rs19};
ld.global.nc.u16 %rs20, [%rd54+2];
ld.global.nc.u16 %rs21, [%rd54+4];
mov.b32 %r72, {%rs20, %rs21};
add.s64 %rd55, %rd72, %rd9;
ld.global.nc.u16 %rs15, [%rd55];

{ cvt.f32.f16 %f50, %rs15;}



{.reg .f16 low,high;
mov.b32 {low,high}, %r11;
mov.b16 %rs16, low;}


{ cvt.f32.f16 %f51, %rs16;}



dp4a.s32.s32 %r67, %r68, %r9, %r50;


dp4a.s32.s32 %r71, %r72, %r10, %r67;

mul.ftz.f32 %f52, %f50, %f51;
cvt.rn.f32.s32 %f53, %r71;
fma.rn.ftz.f32 %f163, %f52, %f53, %f163;
@%p3 bra $L__BB86_7;

add.s64 %rd56, %rd71, %rd10;
ld.global.u16 %rs25, [%rd56+-2];
ld.global.u16 %rs26, [%rd56];
mov.b32 %r77, {%rs25, %rs26};
ld.global.u16 %rs27, [%rd56+2];
ld.global.u16 %rs28, [%rd56+4];
mov.b32 %r81, {%rs27, %rs28};
add.s64 %rd57, %rd71, %rd9;
ld.global.u16 %rs22, [%rd57];

{ cvt.f32.f16 %f54, %rs22;}



{.reg .f16 low,high;
mov.b32 {low,high}, %r11;
mov.b16 %rs23, low;}


{ cvt.f32.f16 %f55, %rs23;}


mov.u32 %r79, 0;

dp4a.s32.s32 %r76, %r77, %r9, %r79;


dp4a.s32.s32 %r80, %r81, %r10, %r76;

mul.ftz.f32 %f56, %f54, %f55;
cvt.rn.f32.s32 %f57, %r80;
fma.rn.ftz.f32 %f161, %f56, %f57, %f161;

$L__BB86_7:
add.s64 %rd73, %rd73, 288;
add.s64 %rd72, %rd72, 272;
add.s64 %rd71, %rd71, 272;
add.s32 %r161, %r161, 8;
setp.lt.u32 %p5, %r161, %r2;
@%p5 bra $L__BB86_3;

st.local.v2.f32 [%rd1], {%f164, %f163};
st.local.v2.f32 [%rd2], {%f162, %f161};

$L__BB86_9:
mov.b32 %r13, %f164;
setp.ne.s64 %p6, %rd26, 0;
@%p6 bra $L__BB86_11;
bra.uni $L__BB86_10;

$L__BB86_11:
mov.u32 %r110, 31;
mov.u32 %r111, 16;
mov.u32 %r112, -1;
shfl.sync.bfly.b32 %r113|%p17, %r13, %r111, %r110, %r112;
mov.b32 %f78, %r113;
add.ftz.f32 %f79, %f164, %f78;
mov.b32 %r114, %f79;
mov.u32 %r115, 8;
shfl.sync.bfly.b32 %r116|%p18, %r114, %r115, %r110, %r112;
mov.b32 %f80, %r116;
add.ftz.f32 %f81, %f79, %f80;
mov.b32 %r117, %f81;
mov.u32 %r118, 4;
shfl.sync.bfly.b32 %r119|%p19, %r117, %r118, %r110, %r112;
mov.b32 %f82, %r119;
add.ftz.f32 %f83, %f81, %f82;
mov.b32 %r120, %f83;
mov.u32 %r121, 2;
shfl.sync.bfly.b32 %r122|%p20, %r120, %r121, %r110, %r112;
mov.b32 %f84, %r122;
add.ftz.f32 %f85, %f83, %f84;
mov.b32 %r123, %f85;
mov.u32 %r124, 1;
shfl.sync.bfly.b32 %r125|%p21, %r123, %r124, %r110, %r112;
mov.b32 %f86, %r125;
add.ftz.f32 %f87, %f85, %f86;
st.local.f32 [%rd1], %f87;
mov.b32 %r126, %f162;
shfl.sync.bfly.b32 %r127|%p22, %r126, %r111, %r110, %r112;
mov.b32 %f88, %r127;
add.ftz.f32 %f89, %f162, %f88;
mov.b32 %r128, %f89;
shfl.sync.bfly.b32 %r129|%p23, %r128, %r115, %r110, %r112;
mov.b32 %f90, %r129;
add.ftz.f32 %f91, %f89, %f90;
mov.b32 %r130, %f91;
shfl.sync.bfly.b32 %r131|%p24, %r130, %r118, %r110, %r112;
mov.b32 %f92, %r131;
add.ftz.f32 %f93, %f91, %f92;
mov.b32 %r132, %f93;
shfl.sync.bfly.b32 %r133|%p25, %r132, %r121, %r110, %r112;
mov.b32 %f94, %r133;
add.ftz.f32 %f95, %f93, %f94;
mov.b32 %r134, %f95;
shfl.sync.bfly.b32 %r135|%p26, %r134, %r124, %r110, %r112;
mov.b32 %f96, %r135;
add.ftz.f32 %f97, %f95, %f96;
st.local.f32 [%rd2], %f97;
mov.b32 %r136, %f163;
shfl.sync.bfly.b32 %r137|%p27, %r136, %r111, %r110, %r112;
mov.b32 %f98, %r137;
add.ftz.f32 %f99, %f163, %f98;
mov.b32 %r138, %f99;
shfl.sync.bfly.b32 %r139|%p28, %r138, %r115, %r110, %r112;
mov.b32 %f100, %r139;
add.ftz.f32 %f101, %f99, %f100;
mov.b32 %r140, %f101;
shfl.sync.bfly.b32 %r141|%p29, %r140, %r118, %r110, %r112;
mov.b32 %f102, %r141;
add.ftz.f32 %f103, %f101, %f102;
mov.b32 %r142, %f103;
shfl.sync.bfly.b32 %r143|%p30, %r142, %r121, %r110, %r112;
mov.b32 %f104, %r143;
add.ftz.f32 %f105, %f103, %f104;
mov.b32 %r144, %f105;
shfl.sync.bfly.b32 %r145|%p31, %r144, %r124, %r110, %r112;
mov.b32 %f106, %r145;
add.ftz.f32 %f107, %f105, %f106;
st.local.f32 [%rd1+4], %f107;
mov.b32 %r146, %f161;
shfl.sync.bfly.b32 %r147|%p32, %r146, %r111, %r110, %r112;
mov.b32 %f108, %r147;
add.ftz.f32 %f109, %f161, %f108;
mov.b32 %r148, %f109;
shfl.sync.bfly.b32 %r149|%p33, %r148, %r115, %r110, %r112;
mov.b32 %f110, %r149;
add.ftz.f32 %f111, %f109, %f110;
mov.b32 %r150, %f111;
shfl.sync.bfly.b32 %r151|%p34, %r150, %r118, %r110, %r112;
mov.b32 %f112, %r151;
add.ftz.f32 %f113, %f111, %f112;
mov.b32 %r152, %f113;
shfl.sync.bfly.b32 %r153|%p35, %r152, %r121, %r110, %r112;
mov.b32 %f114, %r153;
add.ftz.f32 %f115, %f113, %f114;
mov.b32 %r154, %f115;
shfl.sync.bfly.b32 %r155|%p36, %r154, %r124, %r110, %r112;
mov.b32 %f116, %r155;
add.ftz.f32 %f117, %f115, %f116;
st.local.f32 [%rd2+4], %f117;
bra.uni $L__BB86_12;

$L__BB86_10:
mov.u32 %r84, 31;
mov.u32 %r85, 16;
mov.u32 %r86, -1;
shfl.sync.bfly.b32 %r87|%p7, %r13, %r85, %r84, %r86;
mov.b32 %f58, %r87;
add.ftz.f32 %f59, %f164, %f58;
mov.b32 %r88, %f59;
mov.u32 %r89, 8;
shfl.sync.bfly.b32 %r90|%p8, %r88, %r89, %r84, %r86;
mov.b32 %f60, %r90;
add.ftz.f32 %f61, %f59, %f60;
mov.b32 %r91, %f61;
mov.u32 %r92, 4;
shfl.sync.bfly.b32 %r93|%p9, %r91, %r92, %r84, %r86;
mov.b32 %f62, %r93;
add.ftz.f32 %f63, %f61, %f62;
mov.b32 %r94, %f63;
mov.u32 %r95, 2;
shfl.sync.bfly.b32 %r96|%p10, %r94, %r95, %r84, %r86;
mov.b32 %f64, %r96;
add.ftz.f32 %f65, %f63, %f64;
mov.b32 %r97, %f65;
mov.u32 %r98, 1;
shfl.sync.bfly.b32 %r99|%p11, %r97, %r98, %r84, %r86;
mov.b32 %f66, %r99;
add.ftz.f32 %f67, %f65, %f66;
st.local.f32 [%rd1], %f67;
mov.b32 %r100, %f163;
shfl.sync.bfly.b32 %r101|%p12, %r100, %r85, %r84, %r86;
mov.b32 %f68, %r101;
add.ftz.f32 %f69, %f163, %f68;
mov.b32 %r102, %f69;
shfl.sync.bfly.b32 %r103|%p13, %r102, %r89, %r84, %r86;
mov.b32 %f70, %r103;
add.ftz.f32 %f71, %f69, %f70;
mov.b32 %r104, %f71;
shfl.sync.bfly.b32 %r105|%p14, %r104, %r92, %r84, %r86;
mov.b32 %f72, %r105;
add.ftz.f32 %f73, %f71, %f72;
mov.b32 %r106, %f73;
shfl.sync.bfly.b32 %r107|%p15, %r106, %r95, %r84, %r86;
mov.b32 %f74, %r107;
add.ftz.f32 %f75, %f73, %f74;
mov.b32 %r108, %f75;
shfl.sync.bfly.b32 %r109|%p16, %r108, %r98, %r84, %r86;
mov.b32 %f76, %r109;
add.ftz.f32 %f77, %f75, %f76;
st.local.f32 [%rd1+4], %f77;

$L__BB86_12:
setp.gt.u32 %p37, %r6, 1;
add.s32 %r14, %r1, %r6;
setp.ge.u32 %p38, %r14, %r20;
or.pred %p39, %p37, %p38;
@%p39 bra $L__BB86_29;

cvt.u64.u32 %rd20, %r6;
mul.wide.u32 %rd58, %r6, 4;
add.s64 %rd59, %rd1, %rd58;
ld.local.f32 %f167, [%rd59];
mad.lo.s32 %r15, %r5, %r26, %r14;
setp.eq.s64 %p40, %rd25, 0;
@%p40 bra $L__BB86_15;

cvta.to.global.u64 %rd60, %rd25;
mul.wide.u32 %rd61, %r15, 4;
add.s64 %rd62, %rd60, %rd61;
ld.global.f32 %f118, [%rd62];
add.ftz.f32 %f167, %f167, %f118;

$L__BB86_15:
setp.eq.s64 %p41, %rd26, 0;
@%p41 bra $L__BB86_28;

shl.b64 %rd63, %rd20, 2;
add.s64 %rd64, %rd2, %rd63;
ld.local.f32 %f24, [%rd64];
setp.eq.s64 %p42, %rd27, 0;
@%p42 bra $L__BB86_18;

cvta.to.global.u64 %rd65, %rd27;
mul.wide.u32 %rd66, %r15, 4;
add.s64 %rd67, %rd65, %rd66;
ld.global.f32 %f119, [%rd67];
add.ftz.f32 %f24, %f24, %f119;

$L__BB86_18:
setp.gt.s32 %p43, %r16, 2;
@%p43 bra $L__BB86_22;

setp.eq.s32 %p46, %r16, 1;
@%p46 bra $L__BB86_26;

setp.eq.s32 %p47, %r16, 2;
@%p47 bra $L__BB86_21;
bra.uni $L__BB86_27;

$L__BB86_21:
mul.ftz.f32 %f139, %f24, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f140, %f139;
add.ftz.f32 %f141, %f140, 0f3F800000;
div.approx.ftz.f32 %f142, %f24, %f141;
mul.ftz.f32 %f167, %f167, %f142;
bra.uni $L__BB86_28;

$L__BB86_22:
setp.eq.s32 %p44, %r16, 3;
@%p44 bra $L__BB86_25;

setp.ne.s32 %p45, %r16, 6;
@%p45 bra $L__BB86_27;

min.ftz.f32 %f120, %f24, %f31;
neg.ftz.f32 %f121, %f31;
min.ftz.f32 %f122, %f167, %f31;
max.ftz.f32 %f123, %f122, %f121;
mul.ftz.f32 %f124, %f120, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f125, %f124;
add.ftz.f32 %f126, %f125, 0f3F800000;
div.approx.ftz.f32 %f127, %f120, %f126;
mul.ftz.f32 %f167, %f123, %f127;
bra.uni $L__BB86_28;

$L__BB86_27:
mul.ftz.f32 %f167, %f167, %f24;
bra.uni $L__BB86_28;

$L__BB86_26:
mul.ftz.f32 %f143, %f24, 0f3F000000;
mul.ftz.f32 %f144, %f24, 0f3D372713;
fma.rn.ftz.f32 %f145, %f24, %f144, 0f3F800000;
mul.ftz.f32 %f146, %f24, 0f3F4C422A;
mul.ftz.f32 %f147, %f146, %f145;
tanh.approx.f32 %f148, %f147;
add.ftz.f32 %f149, %f148, 0f3F800000;
mul.ftz.f32 %f150, %f143, %f149;
mul.ftz.f32 %f167, %f167, %f150;
bra.uni $L__BB86_28;

$L__BB86_25:
mov.f32 %f128, 0f40E00000;
min.ftz.f32 %f129, %f24, %f128;
min.ftz.f32 %f130, %f167, %f128;
mov.f32 %f131, 0fC0E00000;
max.ftz.f32 %f132, %f130, %f131;
mul.ftz.f32 %f133, %f129, 0f3FD9DB23;
mul.ftz.f32 %f134, %f133, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f135, %f134;
add.ftz.f32 %f136, %f135, 0f3F800000;
div.approx.ftz.f32 %f137, %f129, %f136;
add.ftz.f32 %f138, %f132, 0f3F800000;
mul.ftz.f32 %f167, %f138, %f137;

$L__BB86_28:
mad.lo.s32 %r159, %r3, %r23, %r14;
mad.lo.s32 %r160, %r4, %r26, %r159;
cvta.to.global.u64 %rd68, %rd30;
mul.wide.u32 %rd69, %r160, 4;
add.s64 %rd70, %rd68, %rd69;
st.global.f32 [%rd70], %f167;

$L__BB86_29:
ret;

}
.entry _Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj(
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2,
.param .align 8 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3[48],
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5,
.param .align 4 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6[12],
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15
)
.maxntid 256, 1, 1
.minnctapersm 1
{
.local .align 8 .b8 __local_depot87[8];
.reg .b64 %SP;
.reg .b64 %SPL;
.reg .pred %p<17>;
.reg .b16 %rs<15>;
.reg .f32 %f<44>;
.reg .b32 %r<90>;
.reg .b64 %rd<45>;


mov.u64 %SPL, __local_depot87;
ld.param.u64 %rd14, [_Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0];
ld.param.u64 %rd15, [_Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1];
ld.param.u64 %rd16, [_Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2];
ld.param.u64 %rd17, [_Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4];
ld.param.u32 %r20, [_Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5];
ld.param.u32 %r12, [_Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7];
ld.param.u32 %r13, [_Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8];
ld.param.u32 %r14, [_Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9];
ld.param.u32 %r15, [_Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10];
ld.param.u32 %r16, [_Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11];
ld.param.u32 %r17, [_Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12];
ld.param.u32 %r18, [_Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13];
ld.param.u32 %r21, [_Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14];
ld.param.u32 %r19, [_Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15];
ld.param.u32 %r11, [_Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+8];
ld.param.u32 %r10, [_Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+4];
ld.param.u32 %r9, [_Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6];
add.u64 %rd1, %SPL, 0;
mov.u32 %r22, %ctaid.x;
shl.b32 %r1, %r22, 1;
shr.u32 %r2, %r20, 5;
mov.u32 %r3, %tid.y;
setp.ge.u32 %p1, %r3, %r21;
@%p1 bra $L__BB87_7;

mov.f32 %f42, 0f00000000;
st.local.v2.f32 [%rd1], {%f42, %f42};
mov.u32 %r4, %tid.x;
shr.u32 %r89, %r4, 2;
setp.ge.u32 %p2, %r89, %r2;
mov.f32 %f43, %f42;
@%p2 bra $L__BB87_5;

mov.u32 %r23, %ctaid.y;
mad.lo.s32 %r24, %r3, %r19, %r23;
cvta.to.global.u64 %rd19, %rd16;
mul.wide.u32 %rd20, %r24, 4;
add.s64 %rd21, %rd19, %rd20;
ld.global.nc.u32 %r25, [%rd21];
mul.lo.s32 %r26, %r25, %r16;
cvt.u64.u32 %rd22, %r4;
shr.u64 %rd23, %rd22, 2;
mul.lo.s32 %r27, %r3, %r14;
cvt.u64.u32 %rd24, %r27;
add.s64 %rd25, %rd23, %rd24;
mul.hi.u32 %r28, %r23, %r9;
add.s32 %r29, %r28, %r23;
shr.u32 %r30, %r29, %r10;
mul.lo.s32 %r31, %r30, %r11;
sub.s32 %r32, %r23, %r31;
mul.lo.s32 %r33, %r32, %r17;
cvt.u64.u32 %rd26, %r33;
add.s64 %rd27, %rd25, %rd26;
mul.lo.s64 %rd28, %rd27, 36;
cvta.to.global.u64 %rd29, %rd15;
add.s64 %rd30, %rd29, %rd28;
add.s64 %rd44, %rd30, 8;
and.b32 %r34, %r4, 3;
mul.wide.u32 %rd4, %r34, 8;
mad.lo.s32 %r35, %r1, %r13, %r89;
add.s32 %r36, %r35, %r26;
mul.wide.s32 %rd5, %r36, 34;
add.s64 %rd31, %rd4, %rd5;
add.s64 %rd6, %rd31, 8;
add.s32 %r37, %r1, 1;
add.s32 %r38, %r89, %r26;
mad.lo.s32 %r39, %r13, %r37, %r38;
mul.wide.s32 %rd7, %r39, 34;
add.s64 %rd32, %rd4, %rd7;
add.s64 %rd8, %rd32, 4;
cvta.to.global.u64 %rd43, %rd14;
mov.f32 %f42, 0f00000000;

$L__BB87_3:
add.s64 %rd33, %rd43, %rd6;
ld.global.nc.u16 %rs7, [%rd33+-6];
ld.global.nc.u16 %rs8, [%rd33+-4];
mov.b32 %r42, {%rs7, %rs8};
add.s64 %rd34, %rd44, %rd4;
ld.global.nc.u32 %r43, [%rd34+-4];
ld.global.nc.u16 %rs9, [%rd33+-2];
ld.global.nc.u16 %rs10, [%rd33];
mov.b32 %r46, {%rs9, %rs10};
ld.global.nc.u32 %r47, [%rd34];
add.s64 %rd35, %rd43, %rd5;
ld.global.nc.u16 %rs1, [%rd35];

{ cvt.f32.f16 %f11, %rs1;}


ld.global.nc.u32 %r40, [%rd44+-8];

{.reg .f16 low,high;
mov.b32 {low,high}, %r40;
mov.b16 %rs2, low;}


{ cvt.f32.f16 %f12, %rs2;}


mov.u32 %r53, 0;

dp4a.s32.s32 %r41, %r42, %r43, %r53;


dp4a.s32.s32 %r45, %r46, %r47, %r41;

mul.ftz.f32 %f15, %f11, %f12;
cvt.rn.f32.s32 %f16, %r45;
fma.rn.ftz.f32 %f43, %f15, %f16, %f43;
add.s64 %rd36, %rd43, %rd8;
ld.global.nc.u16 %rs11, [%rd36+-2];
ld.global.nc.u16 %rs12, [%rd36];
mov.b32 %r51, {%rs11, %rs12};
ld.global.nc.u16 %rs13, [%rd36+2];
ld.global.nc.u16 %rs14, [%rd36+4];
mov.b32 %r55, {%rs13, %rs14};
add.s64 %rd37, %rd43, %rd7;
ld.global.nc.u16 %rs4, [%rd37];

{ cvt.f32.f16 %f13, %rs4;}



{.reg .f16 low,high;
mov.b32 {low,high}, %r40;
mov.b16 %rs5, low;}


{ cvt.f32.f16 %f14, %rs5;}



dp4a.s32.s32 %r50, %r51, %r43, %r53;


dp4a.s32.s32 %r54, %r55, %r47, %r50;

mul.ftz.f32 %f17, %f13, %f14;
cvt.rn.f32.s32 %f18, %r54;
fma.rn.ftz.f32 %f42, %f17, %f18, %f42;
add.s64 %rd44, %rd44, 288;
add.s64 %rd43, %rd43, 272;
add.s32 %r89, %r89, 8;
setp.lt.u32 %p3, %r89, %r2;
@%p3 bra $L__BB87_3;

st.local.v2.f32 [%rd1], {%f43, %f42};

$L__BB87_5:
mov.b32 %r58, %f43;
mov.u32 %r59, 31;
mov.u32 %r60, 16;
mov.u32 %r61, -1;
shfl.sync.bfly.b32 %r62|%p4, %r58, %r60, %r59, %r61;
mov.b32 %f19, %r62;
add.ftz.f32 %f20, %f43, %f19;
mov.b32 %r63, %f20;
mov.u32 %r64, 8;
shfl.sync.bfly.b32 %r65|%p5, %r63, %r64, %r59, %r61;
mov.b32 %f21, %r65;
add.ftz.f32 %f22, %f20, %f21;
mov.b32 %r66, %f22;
mov.u32 %r67, 4;
shfl.sync.bfly.b32 %r68|%p6, %r66, %r67, %r59, %r61;
mov.b32 %f23, %r68;
add.ftz.f32 %f24, %f22, %f23;
mov.b32 %r69, %f24;
mov.u32 %r70, 2;
shfl.sync.bfly.b32 %r71|%p7, %r69, %r70, %r59, %r61;
mov.b32 %f25, %r71;
add.ftz.f32 %f26, %f24, %f25;
mov.b32 %r72, %f26;
mov.u32 %r73, 1;
shfl.sync.bfly.b32 %r74|%p8, %r72, %r73, %r59, %r61;
mov.b32 %f27, %r74;
add.ftz.f32 %f28, %f26, %f27;
st.local.f32 [%rd1], %f28;
mov.b32 %r75, %f42;
shfl.sync.bfly.b32 %r76|%p9, %r75, %r60, %r59, %r61;
mov.b32 %f29, %r76;
add.ftz.f32 %f30, %f42, %f29;
mov.b32 %r77, %f30;
shfl.sync.bfly.b32 %r78|%p10, %r77, %r64, %r59, %r61;
mov.b32 %f31, %r78;
add.ftz.f32 %f32, %f30, %f31;
mov.b32 %r79, %f32;
shfl.sync.bfly.b32 %r80|%p11, %r79, %r67, %r59, %r61;
mov.b32 %f33, %r80;
add.ftz.f32 %f34, %f32, %f33;
mov.b32 %r81, %f34;
shfl.sync.bfly.b32 %r82|%p12, %r81, %r70, %r59, %r61;
mov.b32 %f35, %r82;
add.ftz.f32 %f36, %f34, %f35;
mov.b32 %r83, %f36;
shfl.sync.bfly.b32 %r84|%p13, %r83, %r73, %r59, %r61;
mov.b32 %f37, %r84;
add.ftz.f32 %f38, %f36, %f37;
st.local.f32 [%rd1+4], %f38;
add.s32 %r8, %r1, %r4;
setp.ge.u32 %p14, %r8, %r12;
setp.gt.u32 %p15, %r4, 1;
or.pred %p16, %p15, %p14;
@%p16 bra $L__BB87_7;

mul.wide.u32 %rd38, %r4, 4;
add.s64 %rd39, %rd1, %rd38;
ld.local.f32 %f39, [%rd39];
mov.u32 %r85, %ctaid.y;
mad.lo.s32 %r87, %r3, %r15, %r8;
mad.lo.s32 %r88, %r85, %r18, %r87;
cvta.to.global.u64 %rd40, %rd17;
mul.wide.u32 %rd41, %r88, 4;
add.s64 %rd42, %rd40, %rd41;
st.global.f32 [%rd42], %f39;

$L__BB87_7:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<319>;
.reg .b16 %rs<57>;
.reg .f32 %f<1005>;
.reg .b32 %r<922>;
.reg .b64 %rd<98>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate[1536];

ld.param.u64 %rd25, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd26, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd27, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r46, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r47, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r48, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r52, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r53, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r54, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r50, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r49, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
ld.param.u64 %rd30, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+16];
ld.param.u64 %rd29, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+8];
ld.param.u64 %rd28, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3];
mov.u32 %r1, %tid.y;
shl.b32 %r55, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r55, %r2;
mov.u32 %r56, %ctaid.x;
shl.b32 %r4, %r56, 2;
mov.u32 %r5, %ctaid.y;
setp.eq.s64 %p1, %rd27, 0;
@%p1 bra $L__BB88_2;

cvta.to.global.u64 %rd34, %rd27;
mul.wide.u32 %rd35, %r5, 4;
add.s64 %rd36, %rd34, %rd35;
ld.global.nc.u32 %r919, [%rd36];
mul.hi.u32 %r57, %r5, %r38;
add.s32 %r58, %r57, %r5;
shr.u32 %r59, %r58, %r39;
mul.lo.s32 %r60, %r59, %r40;
sub.s32 %r920, %r5, %r60;
bra.uni $L__BB88_3;

$L__BB88_2:
mul.hi.u32 %r61, %r5, %r43;
add.s32 %r62, %r61, %r5;
shr.u32 %r919, %r62, %r44;
mov.u32 %r920, %r5;

$L__BB88_3:
mov.u32 %r11, %ctaid.z;
selp.b32 %r12, %r5, %r919, %p1;
setp.ne.s32 %p3, %r1, 0;
setp.gt.u32 %p4, %r2, 3;
or.pred %p5, %p4, %p3;
add.s32 %r63, %r4, %r2;
setp.ge.u32 %p6, %r63, %r42;
mov.f32 %f27, 0f00000000;
or.pred %p7, %p5, %p6;
mov.f32 %f4, %f27;
mov.f32 %f5, %f27;
@%p7 bra $L__BB88_8;

setp.eq.s64 %p8, %rd28, 0;
mov.f32 %f4, 0f00000000;
mov.f32 %f5, %f4;
@%p8 bra $L__BB88_6;

cvta.to.global.u64 %rd37, %rd28;
mul.lo.s32 %r64, %r11, %r54;
cvt.u64.u32 %rd38, %r64;
mul.lo.s32 %r65, %r12, %r48;
cvt.u64.u32 %rd39, %r65;
cvt.u64.u32 %rd40, %r2;
cvt.s64.s32 %rd41, %r4;
add.s64 %rd42, %rd41, %rd40;
add.s64 %rd43, %rd42, %rd38;
add.s64 %rd44, %rd43, %rd39;
shl.b64 %rd45, %rd44, 2;
add.s64 %rd46, %rd37, %rd45;
ld.global.f32 %f5, [%rd46];

$L__BB88_6:
setp.eq.s64 %p9, %rd30, 0;
setp.eq.s64 %p10, %rd29, 0;
or.pred %p11, %p9, %p10;
@%p11 bra $L__BB88_8;

cvta.to.global.u64 %rd47, %rd30;
mul.lo.s32 %r66, %r11, %r54;
cvt.u64.u32 %rd48, %r66;
mul.lo.s32 %r67, %r12, %r48;
cvt.u64.u32 %rd49, %r67;
cvt.u64.u32 %rd50, %r2;
cvt.s64.s32 %rd51, %r4;
add.s64 %rd52, %rd51, %rd50;
add.s64 %rd53, %rd52, %rd48;
add.s64 %rd54, %rd53, %rd49;
shl.b64 %rd55, %rd54, 2;
add.s64 %rd56, %rd47, %rd55;
ld.global.f32 %f4, [%rd56];

$L__BB88_8:
mul.lo.s32 %r68, %r11, %r53;
cvt.u64.u32 %rd57, %r68;
mul.lo.s32 %r69, %r920, %r47;
cvt.u64.u32 %rd58, %r69;
add.s64 %rd4, %rd58, %rd57;
mul.hi.u32 %r70, %r11, %r49;
add.s32 %r71, %r70, %r11;
shr.u32 %r13, %r71, %r50;
shr.s32 %r72, %r3, 31;
shr.u32 %r73, %r72, 30;
add.s32 %r74, %r3, %r73;
shr.s32 %r921, %r74, 2;
shr.u32 %r75, %r37, 5;
setp.ge.s32 %p12, %r921, %r75;
mov.f32 %f28, %f27;
mov.f32 %f29, %f27;
mov.f32 %f30, %f27;
mov.f32 %f31, %f27;
mov.f32 %f32, %f27;
mov.f32 %f33, %f27;
mov.f32 %f34, %f27;
@%p12 bra $L__BB88_19;

mul.lo.s32 %r76, %r13, %r52;
mul.lo.s32 %r77, %r919, %r46;
and.b32 %r81, %r74, -4;
sub.s32 %r82, %r3, %r81;
shl.b32 %r83, %r82, 1;
shl.b32 %r84, %r82, 2;
or.b32 %r85, %r84, 1;
cvt.s64.s32 %rd5, %r83;
sub.s32 %r86, %r85, %r83;
cvt.s64.s32 %rd6, %r86;
add.s32 %r87, %r921, %r76;
add.s32 %r15, %r87, %r77;
add.s32 %r88, %r4, 1;
mad.lo.s32 %r89, %r41, %r88, %r15;
mul.wide.s32 %rd59, %r89, 34;
mad.lo.s32 %r90, %r4, %r41, %r921;
add.s32 %r91, %r90, %r76;
add.s32 %r92, %r91, %r77;
mul.wide.s32 %rd7, %r92, 34;
mul.wide.s32 %rd60, %r84, 2;
add.s64 %rd61, %rd7, %rd60;
add.s64 %rd8, %rd61, 4;
add.s64 %rd62, %rd59, %rd60;
add.s64 %rd9, %rd62, 2;
add.s32 %r93, %r4, 2;
mad.lo.s32 %r94, %r41, %r93, %r15;
mul.wide.s32 %rd63, %r94, 34;
add.s64 %rd64, %rd63, %rd60;
add.s64 %rd10, %rd64, 2;
add.s32 %r95, %r4, 3;
mad.lo.s32 %r96, %r41, %r95, %r15;
mul.wide.s32 %rd65, %r96, 34;
add.s64 %rd11, %rd65, %rd60;
cvta.to.global.u64 %rd97, %rd29;
cvta.to.global.u64 %rd96, %rd25;
mov.f32 %f27, 0f00000000;
cvta.to.global.u64 %rd66, %rd26;
shl.b64 %rd73, %rd5, 2;
shl.b64 %rd75, %rd6, 2;
mov.f32 %f28, %f27;
mov.f32 %f29, %f27;
mov.f32 %f30, %f27;
mov.f32 %f31, %f27;
mov.f32 %f32, %f27;
mov.f32 %f33, %f27;
mov.f32 %f34, %f27;

$L__BB88_10:
cvt.u64.u32 %rd67, %r921;
add.s64 %rd68, %rd4, %rd67;
mul.lo.s64 %rd69, %rd68, 36;
add.s64 %rd70, %rd66, %rd69;
add.s64 %rd71, %rd70, 4;
add.s64 %rd72, %rd96, %rd8;
add.s64 %rd16, %rd72, -2;
ld.global.nc.u16 %rs4, [%rd72+-2];
ld.global.nc.u16 %rs5, [%rd72];
mov.b32 %r99, {%rs4, %rs5};
add.s64 %rd74, %rd71, %rd73;
ld.global.nc.u32 %r17, [%rd74];
ld.global.nc.u16 %rs6, [%rd72+2];
ld.global.nc.u16 %rs7, [%rd72+4];
mov.b32 %r103, {%rs6, %rs7};
add.s64 %rd76, %rd71, %rd75;
ld.global.nc.u32 %r18, [%rd76];
add.s64 %rd17, %rd96, %rd7;
ld.global.nc.u16 %rs1, [%rd17];

{ cvt.f32.f16 %f101, %rs1;}


ld.global.nc.u32 %r19, [%rd70];

{.reg .f16 low,high;
mov.b32 {low,high}, %r19;
mov.b16 %rs2, low;}


{ cvt.f32.f16 %f102, %rs2;}


mov.u32 %r101, 0;

dp4a.s32.s32 %r98, %r99, %r17, %r101;


dp4a.s32.s32 %r102, %r103, %r18, %r98;

mul.ftz.f32 %f103, %f101, %f102;
cvt.rn.f32.s32 %f104, %r102;
fma.rn.ftz.f32 %f30, %f103, %f104, %f30;
setp.eq.s64 %p13, %rd29, 0;
@%p13 bra $L__BB88_12;

add.s64 %rd77, %rd97, %rd8;
ld.global.u16 %rs11, [%rd77+-2];
ld.global.u16 %rs12, [%rd77];
mov.b32 %r108, {%rs11, %rs12};
ld.global.u16 %rs13, [%rd77+2];
ld.global.u16 %rs14, [%rd77+4];
mov.b32 %r112, {%rs13, %rs14};
add.s64 %rd78, %rd97, %rd7;
ld.global.u16 %rs8, [%rd78];

{ cvt.f32.f16 %f105, %rs8;}



{.reg .f16 low,high;
mov.b32 {low,high}, %r19;
mov.b16 %rs9, low;}


{ cvt.f32.f16 %f106, %rs9;}



dp4a.s32.s32 %r107, %r108, %r17, %r101;


dp4a.s32.s32 %r111, %r112, %r18, %r107;

mul.ftz.f32 %f107, %f105, %f106;
cvt.rn.f32.s32 %f108, %r111;
fma.rn.ftz.f32 %f34, %f107, %f108, %f34;

$L__BB88_12:
mul.wide.s32 %rd79, %r41, 34;
add.s64 %rd18, %rd16, %rd79;
ld.global.nc.u16 %rs18, [%rd18];
ld.global.nc.u16 %rs19, [%rd18+2];
mov.b32 %r117, {%rs18, %rs19};
ld.global.nc.u16 %rs20, [%rd18+4];
ld.global.nc.u16 %rs21, [%rd18+6];
mov.b32 %r121, {%rs20, %rs21};
add.s64 %rd19, %rd17, %rd79;
ld.global.nc.u16 %rs15, [%rd19];

{ cvt.f32.f16 %f109, %rs15;}



{.reg .f16 low,high;
mov.b32 {low,high}, %r19;
mov.b16 %rs16, low;}


{ cvt.f32.f16 %f110, %rs16;}



dp4a.s32.s32 %r116, %r117, %r17, %r101;


dp4a.s32.s32 %r120, %r121, %r18, %r116;

mul.ftz.f32 %f111, %f109, %f110;
cvt.rn.f32.s32 %f112, %r120;
fma.rn.ftz.f32 %f29, %f111, %f112, %f29;
@%p13 bra $L__BB88_14;

add.s64 %rd80, %rd97, %rd9;
ld.global.u16 %rs25, [%rd80];
ld.global.u16 %rs26, [%rd80+2];
mov.b32 %r126, {%rs25, %rs26};
ld.global.u16 %rs27, [%rd80+4];
ld.global.u16 %rs28, [%rd80+6];
mov.b32 %r130, {%rs27, %rs28};
or.b32 %r135, %r4, 1;
mad.lo.s32 %r136, %r41, %r135, %r15;
mul.wide.s32 %rd81, %r136, 34;
add.s64 %rd82, %rd97, %rd81;
ld.global.u16 %rs22, [%rd82];

{ cvt.f32.f16 %f113, %rs22;}



{.reg .f16 low,high;
mov.b32 {low,high}, %r19;
mov.b16 %rs23, low;}


{ cvt.f32.f16 %f114, %rs23;}


mov.u32 %r128, 0;

dp4a.s32.s32 %r125, %r126, %r17, %r128;


dp4a.s32.s32 %r129, %r130, %r18, %r125;

mul.ftz.f32 %f115, %f113, %f114;
cvt.rn.f32.s32 %f116, %r129;
fma.rn.ftz.f32 %f33, %f115, %f116, %f33;

$L__BB88_14:
add.s64 %rd20, %rd18, %rd79;
ld.global.nc.u16 %rs32, [%rd20];
ld.global.nc.u16 %rs33, [%rd20+2];
mov.b32 %r139, {%rs32, %rs33};
ld.global.nc.u16 %rs34, [%rd20+4];
ld.global.nc.u16 %rs35, [%rd20+6];
mov.b32 %r143, {%rs34, %rs35};
add.s64 %rd21, %rd19, %rd79;
ld.global.nc.u16 %rs29, [%rd21];

{ cvt.f32.f16 %f117, %rs29;}



{.reg .f16 low,high;
mov.b32 {low,high}, %r19;
mov.b16 %rs30, low;}


{ cvt.f32.f16 %f118, %rs30;}


mov.u32 %r141, 0;

dp4a.s32.s32 %r138, %r139, %r17, %r141;


dp4a.s32.s32 %r142, %r143, %r18, %r138;

mul.ftz.f32 %f119, %f117, %f118;
cvt.rn.f32.s32 %f120, %r142;
fma.rn.ftz.f32 %f28, %f119, %f120, %f28;
@%p13 bra $L__BB88_16;

add.s64 %rd84, %rd97, %rd10;
ld.global.u16 %rs39, [%rd84];
ld.global.u16 %rs40, [%rd84+2];
mov.b32 %r148, {%rs39, %rs40};
ld.global.u16 %rs41, [%rd84+4];
ld.global.u16 %rs42, [%rd84+6];
mov.b32 %r152, {%rs41, %rs42};
or.b32 %r157, %r4, 2;
mad.lo.s32 %r158, %r41, %r157, %r15;
mul.wide.s32 %rd85, %r158, 34;
add.s64 %rd86, %rd97, %rd85;
ld.global.u16 %rs36, [%rd86];

{ cvt.f32.f16 %f121, %rs36;}



{.reg .f16 low,high;
mov.b32 {low,high}, %r19;
mov.b16 %rs37, low;}


{ cvt.f32.f16 %f122, %rs37;}



dp4a.s32.s32 %r147, %r148, %r17, %r141;


dp4a.s32.s32 %r151, %r152, %r18, %r147;

mul.ftz.f32 %f123, %f121, %f122;
cvt.rn.f32.s32 %f124, %r151;
fma.rn.ftz.f32 %f32, %f123, %f124, %f32;

$L__BB88_16:
add.s64 %rd88, %rd20, %rd79;
ld.global.nc.u16 %rs46, [%rd88];
ld.global.nc.u16 %rs47, [%rd88+2];
mov.b32 %r161, {%rs46, %rs47};
ld.global.nc.u16 %rs48, [%rd88+4];
ld.global.nc.u16 %rs49, [%rd88+6];
mov.b32 %r165, {%rs48, %rs49};
add.s64 %rd89, %rd21, %rd79;
ld.global.nc.u16 %rs43, [%rd89];

{ cvt.f32.f16 %f125, %rs43;}



{.reg .f16 low,high;
mov.b32 {low,high}, %r19;
mov.b16 %rs44, low;}


{ cvt.f32.f16 %f126, %rs44;}



dp4a.s32.s32 %r160, %r161, %r17, %r141;


dp4a.s32.s32 %r164, %r165, %r18, %r160;

mul.ftz.f32 %f127, %f125, %f126;
cvt.rn.f32.s32 %f128, %r164;
fma.rn.ftz.f32 %f27, %f127, %f128, %f27;
@%p13 bra $L__BB88_18;

add.s64 %rd90, %rd97, %rd11;
ld.global.u16 %rs53, [%rd90+2];
ld.global.u16 %rs54, [%rd90+4];
mov.b32 %r170, {%rs53, %rs54};
ld.global.u16 %rs55, [%rd90+6];
ld.global.u16 %rs56, [%rd90+8];
mov.b32 %r174, {%rs55, %rs56};
or.b32 %r179, %r4, 3;
mad.lo.s32 %r180, %r41, %r179, %r15;
mul.wide.s32 %rd91, %r180, 34;
add.s64 %rd92, %rd97, %rd91;
ld.global.u16 %rs50, [%rd92];

{ cvt.f32.f16 %f129, %rs50;}



{.reg .f16 low,high;
mov.b32 {low,high}, %r19;
mov.b16 %rs51, low;}


{ cvt.f32.f16 %f130, %rs51;}


mov.u32 %r172, 0;

dp4a.s32.s32 %r169, %r170, %r17, %r172;


dp4a.s32.s32 %r173, %r174, %r18, %r169;

mul.ftz.f32 %f131, %f129, %f130;
cvt.rn.f32.s32 %f132, %r173;
fma.rn.ftz.f32 %f31, %f131, %f132, %f31;

$L__BB88_18:
add.s64 %rd97, %rd97, 1088;
add.s64 %rd96, %rd96, 1088;
add.s32 %r921, %r921, 32;
setp.lt.s32 %p17, %r921, %r75;
@%p17 bra $L__BB88_10;

$L__BB88_19:
mov.u32 %r914, %tid.y;
setp.eq.s32 %p18, %r914, 0;
@%p18 bra $L__BB88_23;

mov.u32 %r915, %tid.y;
setp.ne.s64 %p19, %rd29, 0;
add.s32 %r21, %r915, -1;
shl.b32 %r185, %r915, 9;
mov.u32 %r186, _ZZ13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r187, %r186, %r185;
shl.b32 %r188, %r2, 2;
add.s32 %r189, %r187, %r188;
st.shared.f32 [%r189+-512], %f30;
@%p19 bra $L__BB88_22;
bra.uni $L__BB88_21;

$L__BB88_22:
shl.b32 %r191, %r21, 9;
mov.u32 %r192, _ZZ13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r193, %r192, %r191;
add.s32 %r195, %r193, %r188;
st.shared.f32 [%r195], %f34;
add.s32 %r911, %r189, -512;
st.shared.f32 [%r911+128], %f29;
st.shared.f32 [%r195+128], %f33;
add.s32 %r912, %r189, -512;
st.shared.f32 [%r912+256], %f28;
st.shared.f32 [%r195+256], %f32;
add.s32 %r913, %r189, -512;
st.shared.f32 [%r913+384], %f27;
st.shared.f32 [%r195+384], %f31;
bra.uni $L__BB88_23;

$L__BB88_21:
add.s32 %r908, %r189, -512;
st.shared.f32 [%r908+128], %f29;
add.s32 %r909, %r189, -512;
st.shared.f32 [%r909+256], %f28;
add.s32 %r910, %r189, -512;
st.shared.f32 [%r910+384], %f27;

$L__BB88_23:
bar.sync 0;
@%p3 bra $L__BB88_78;
bra.uni $L__BB88_24;

$L__BB88_78:
ret;

$L__BB88_24:
ld.param.u64 %rd95, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.f32 %f977, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
ld.param.u32 %r918, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r917, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
mov.u32 %r916, %ctaid.z;
setp.ne.s64 %p21, %rd29, 0;
mov.u32 %r198, %ctaid.y;
mad.lo.s32 %r201, %r198, %r918, %r4;
mad.lo.s32 %r202, %r916, %r917, %r201;
neg.ftz.f32 %f35, %f977;
shl.b32 %r204, %r2, 2;
mov.u32 %r205, _ZZ13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r23, %r205, %r204;
cvta.to.global.u64 %rd93, %rd95;
mul.wide.u32 %rd94, %r202, 4;
add.s64 %rd24, %rd93, %rd94;
@%p21 bra $L__BB88_33;
bra.uni $L__BB88_25;

$L__BB88_33:
mov.u32 %r292, _ZZ13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r25, %r292, %r204;
setp.gt.s32 %p54, %r36, 2;
@%p54 bra $L__BB88_44;

setp.eq.s32 %p57, %r36, 1;
@%p57 bra $L__BB88_62;

setp.eq.s32 %p58, %r36, 2;
@%p58 bra $L__BB88_36;
bra.uni $L__BB88_70;

$L__BB88_36:
ld.shared.f32 %f533, [%r23];
add.ftz.f32 %f534, %f533, %f30;
ld.shared.f32 %f535, [%r25];
add.ftz.f32 %f536, %f535, %f34;
ld.shared.f32 %f537, [%r23+512];
add.ftz.f32 %f538, %f537, %f534;
ld.shared.f32 %f539, [%r25+512];
add.ftz.f32 %f540, %f539, %f536;
ld.shared.f32 %f541, [%r23+1024];
add.ftz.f32 %f542, %f541, %f538;
ld.shared.f32 %f543, [%r25+1024];
add.ftz.f32 %f544, %f543, %f540;
mov.b32 %r539, %f542;
mov.u32 %r540, 31;
mov.u32 %r541, 16;
mov.u32 %r542, -1;
shfl.sync.bfly.b32 %r543|%p163, %r539, %r541, %r540, %r542;
mov.b32 %f545, %r543;
add.ftz.f32 %f546, %f542, %f545;
mov.b32 %r544, %f546;
mov.u32 %r545, 8;
shfl.sync.bfly.b32 %r546|%p164, %r544, %r545, %r540, %r542;
mov.b32 %f547, %r546;
add.ftz.f32 %f548, %f546, %f547;
mov.b32 %r547, %f548;
mov.u32 %r548, 4;
shfl.sync.bfly.b32 %r549|%p165, %r547, %r548, %r540, %r542;
mov.b32 %f549, %r549;
add.ftz.f32 %f550, %f548, %f549;
mov.b32 %r550, %f550;
mov.u32 %r551, 2;
shfl.sync.bfly.b32 %r552|%p166, %r550, %r551, %r540, %r542;
mov.b32 %f551, %r552;
add.ftz.f32 %f552, %f550, %f551;
mov.b32 %r553, %f552;
mov.u32 %r554, 1;
shfl.sync.bfly.b32 %r555|%p167, %r553, %r554, %r540, %r542;
mov.b32 %f553, %r555;
add.ftz.f32 %f56, %f552, %f553;
mov.b32 %r556, %f544;
shfl.sync.bfly.b32 %r557|%p168, %r556, %r541, %r540, %r542;
mov.b32 %f554, %r557;
add.ftz.f32 %f555, %f544, %f554;
mov.b32 %r558, %f555;
shfl.sync.bfly.b32 %r559|%p169, %r558, %r545, %r540, %r542;
mov.b32 %f556, %r559;
add.ftz.f32 %f557, %f555, %f556;
mov.b32 %r560, %f557;
shfl.sync.bfly.b32 %r561|%p170, %r560, %r548, %r540, %r542;
mov.b32 %f558, %r561;
add.ftz.f32 %f559, %f557, %f558;
mov.b32 %r562, %f559;
shfl.sync.bfly.b32 %r563|%p171, %r562, %r551, %r540, %r542;
mov.b32 %f560, %r563;
add.ftz.f32 %f561, %f559, %f560;
mov.b32 %r564, %f561;
shfl.sync.bfly.b32 %r565|%p172, %r564, %r554, %r540, %r542;
mov.b32 %f562, %r565;
add.ftz.f32 %f57, %f561, %f562;
setp.ne.s32 %p173, %r2, 0;
setp.ge.u32 %p174, %r4, %r42;
or.pred %p175, %p173, %p174;
@%p175 bra $L__BB88_38;

add.ftz.f32 %f563, %f5, %f56;
add.ftz.f32 %f564, %f4, %f57;
mul.ftz.f32 %f565, %f564, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f566, %f565;
add.ftz.f32 %f567, %f566, 0f3F800000;
div.approx.ftz.f32 %f568, %f564, %f567;
mul.ftz.f32 %f569, %f563, %f568;
st.global.f32 [%rd24], %f569;

$L__BB88_38:
ld.shared.f32 %f570, [%r23+128];
add.ftz.f32 %f571, %f570, %f29;
ld.shared.f32 %f572, [%r25+128];
add.ftz.f32 %f573, %f572, %f33;
ld.shared.f32 %f574, [%r23+640];
add.ftz.f32 %f575, %f574, %f571;
ld.shared.f32 %f576, [%r25+640];
add.ftz.f32 %f577, %f576, %f573;
ld.shared.f32 %f578, [%r23+1152];
add.ftz.f32 %f579, %f578, %f575;
ld.shared.f32 %f580, [%r25+1152];
add.ftz.f32 %f581, %f580, %f577;
mov.b32 %r569, %f579;
shfl.sync.bfly.b32 %r573|%p176, %r569, %r541, %r540, %r542;
mov.b32 %f582, %r573;
add.ftz.f32 %f583, %f579, %f582;
mov.b32 %r574, %f583;
shfl.sync.bfly.b32 %r576|%p177, %r574, %r545, %r540, %r542;
mov.b32 %f584, %r576;
add.ftz.f32 %f585, %f583, %f584;
mov.b32 %r577, %f585;
shfl.sync.bfly.b32 %r579|%p178, %r577, %r548, %r540, %r542;
mov.b32 %f586, %r579;
add.ftz.f32 %f587, %f585, %f586;
mov.b32 %r580, %f587;
shfl.sync.bfly.b32 %r582|%p179, %r580, %r551, %r540, %r542;
mov.b32 %f588, %r582;
add.ftz.f32 %f589, %f587, %f588;
mov.b32 %r583, %f589;
shfl.sync.bfly.b32 %r585|%p180, %r583, %r554, %r540, %r542;
mov.b32 %f590, %r585;
add.ftz.f32 %f58, %f589, %f590;
mov.b32 %r586, %f581;
shfl.sync.bfly.b32 %r587|%p181, %r586, %r541, %r540, %r542;
mov.b32 %f591, %r587;
add.ftz.f32 %f592, %f581, %f591;
mov.b32 %r588, %f592;
shfl.sync.bfly.b32 %r589|%p182, %r588, %r545, %r540, %r542;
mov.b32 %f593, %r589;
add.ftz.f32 %f594, %f592, %f593;
mov.b32 %r590, %f594;
shfl.sync.bfly.b32 %r591|%p183, %r590, %r548, %r540, %r542;
mov.b32 %f595, %r591;
add.ftz.f32 %f596, %f594, %f595;
mov.b32 %r592, %f596;
shfl.sync.bfly.b32 %r593|%p184, %r592, %r551, %r540, %r542;
mov.b32 %f597, %r593;
add.ftz.f32 %f598, %f596, %f597;
mov.b32 %r594, %f598;
shfl.sync.bfly.b32 %r595|%p185, %r594, %r554, %r540, %r542;
mov.b32 %f599, %r595;
add.ftz.f32 %f59, %f598, %f599;
setp.ne.s32 %p186, %r2, 1;
or.b32 %r599, %r4, 1;
setp.ge.u32 %p187, %r599, %r42;
or.pred %p188, %p186, %p187;
@%p188 bra $L__BB88_40;

add.ftz.f32 %f600, %f5, %f58;
add.ftz.f32 %f601, %f4, %f59;
mul.ftz.f32 %f602, %f601, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f603, %f602;
add.ftz.f32 %f604, %f603, 0f3F800000;
div.approx.ftz.f32 %f605, %f601, %f604;
mul.ftz.f32 %f606, %f600, %f605;
st.global.f32 [%rd24+4], %f606;

$L__BB88_40:
ld.shared.f32 %f607, [%r23+256];
add.ftz.f32 %f608, %f607, %f28;
ld.shared.f32 %f609, [%r25+256];
add.ftz.f32 %f610, %f609, %f32;
ld.shared.f32 %f611, [%r23+768];
add.ftz.f32 %f612, %f611, %f608;
ld.shared.f32 %f613, [%r25+768];
add.ftz.f32 %f614, %f613, %f610;
ld.shared.f32 %f615, [%r23+1280];
add.ftz.f32 %f616, %f615, %f612;
ld.shared.f32 %f617, [%r25+1280];
add.ftz.f32 %f618, %f617, %f614;
mov.b32 %r600, %f616;
mov.u32 %r601, 31;
mov.u32 %r602, 16;
mov.u32 %r603, -1;
shfl.sync.bfly.b32 %r604|%p189, %r600, %r602, %r601, %r603;
mov.b32 %f619, %r604;
add.ftz.f32 %f620, %f616, %f619;
mov.b32 %r605, %f620;
mov.u32 %r606, 8;
shfl.sync.bfly.b32 %r607|%p190, %r605, %r606, %r601, %r603;
mov.b32 %f621, %r607;
add.ftz.f32 %f622, %f620, %f621;
mov.b32 %r608, %f622;
mov.u32 %r609, 4;
shfl.sync.bfly.b32 %r610|%p191, %r608, %r609, %r601, %r603;
mov.b32 %f623, %r610;
add.ftz.f32 %f624, %f622, %f623;
mov.b32 %r611, %f624;
mov.u32 %r612, 2;
shfl.sync.bfly.b32 %r613|%p192, %r611, %r612, %r601, %r603;
mov.b32 %f625, %r613;
add.ftz.f32 %f626, %f624, %f625;
mov.b32 %r614, %f626;
mov.u32 %r615, 1;
shfl.sync.bfly.b32 %r616|%p193, %r614, %r615, %r601, %r603;
mov.b32 %f627, %r616;
add.ftz.f32 %f60, %f626, %f627;
mov.b32 %r617, %f618;
shfl.sync.bfly.b32 %r618|%p194, %r617, %r602, %r601, %r603;
mov.b32 %f628, %r618;
add.ftz.f32 %f629, %f618, %f628;
mov.b32 %r619, %f629;
shfl.sync.bfly.b32 %r620|%p195, %r619, %r606, %r601, %r603;
mov.b32 %f630, %r620;
add.ftz.f32 %f631, %f629, %f630;
mov.b32 %r621, %f631;
shfl.sync.bfly.b32 %r622|%p196, %r621, %r609, %r601, %r603;
mov.b32 %f632, %r622;
add.ftz.f32 %f633, %f631, %f632;
mov.b32 %r623, %f633;
shfl.sync.bfly.b32 %r624|%p197, %r623, %r612, %r601, %r603;
mov.b32 %f634, %r624;
add.ftz.f32 %f635, %f633, %f634;
mov.b32 %r625, %f635;
shfl.sync.bfly.b32 %r626|%p198, %r625, %r615, %r601, %r603;
mov.b32 %f636, %r626;
add.ftz.f32 %f61, %f635, %f636;
setp.ne.s32 %p199, %r2, 2;
or.b32 %r630, %r4, 2;
setp.ge.u32 %p200, %r630, %r42;
or.pred %p201, %p199, %p200;
@%p201 bra $L__BB88_42;

add.ftz.f32 %f637, %f5, %f60;
add.ftz.f32 %f638, %f4, %f61;
mul.ftz.f32 %f639, %f638, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f640, %f639;
add.ftz.f32 %f641, %f640, 0f3F800000;
div.approx.ftz.f32 %f642, %f638, %f641;
mul.ftz.f32 %f643, %f637, %f642;
st.global.f32 [%rd24+8], %f643;

$L__BB88_42:
ld.shared.f32 %f644, [%r23+384];
add.ftz.f32 %f645, %f644, %f27;
ld.shared.f32 %f646, [%r25+384];
add.ftz.f32 %f647, %f646, %f31;
ld.shared.f32 %f648, [%r23+896];
add.ftz.f32 %f649, %f648, %f645;
ld.shared.f32 %f650, [%r25+896];
add.ftz.f32 %f651, %f650, %f647;
ld.shared.f32 %f652, [%r23+1408];
add.ftz.f32 %f653, %f652, %f649;
ld.shared.f32 %f654, [%r25+1408];
add.ftz.f32 %f655, %f654, %f651;
mov.b32 %r631, %f653;
shfl.sync.bfly.b32 %r635|%p202, %r631, %r602, %r601, %r603;
mov.b32 %f656, %r635;
add.ftz.f32 %f657, %f653, %f656;
mov.b32 %r636, %f657;
shfl.sync.bfly.b32 %r638|%p203, %r636, %r606, %r601, %r603;
mov.b32 %f658, %r638;
add.ftz.f32 %f659, %f657, %f658;
mov.b32 %r639, %f659;
shfl.sync.bfly.b32 %r641|%p204, %r639, %r609, %r601, %r603;
mov.b32 %f660, %r641;
add.ftz.f32 %f661, %f659, %f660;
mov.b32 %r642, %f661;
shfl.sync.bfly.b32 %r644|%p205, %r642, %r612, %r601, %r603;
mov.b32 %f662, %r644;
add.ftz.f32 %f663, %f661, %f662;
mov.b32 %r645, %f663;
shfl.sync.bfly.b32 %r647|%p206, %r645, %r615, %r601, %r603;
mov.b32 %f664, %r647;
add.ftz.f32 %f62, %f663, %f664;
mov.b32 %r648, %f655;
shfl.sync.bfly.b32 %r649|%p207, %r648, %r602, %r601, %r603;
mov.b32 %f665, %r649;
add.ftz.f32 %f666, %f655, %f665;
mov.b32 %r650, %f666;
shfl.sync.bfly.b32 %r651|%p208, %r650, %r606, %r601, %r603;
mov.b32 %f667, %r651;
add.ftz.f32 %f668, %f666, %f667;
mov.b32 %r652, %f668;
shfl.sync.bfly.b32 %r653|%p209, %r652, %r609, %r601, %r603;
mov.b32 %f669, %r653;
add.ftz.f32 %f670, %f668, %f669;
mov.b32 %r654, %f670;
shfl.sync.bfly.b32 %r655|%p210, %r654, %r612, %r601, %r603;
mov.b32 %f671, %r655;
add.ftz.f32 %f672, %f670, %f671;
mov.b32 %r656, %f672;
shfl.sync.bfly.b32 %r657|%p211, %r656, %r615, %r601, %r603;
mov.b32 %f673, %r657;
add.ftz.f32 %f63, %f672, %f673;
setp.ne.s32 %p212, %r2, 3;
or.b32 %r661, %r4, 3;
setp.ge.u32 %p213, %r661, %r42;
or.pred %p214, %p212, %p213;
@%p214 bra $L__BB88_78;

add.ftz.f32 %f674, %f5, %f62;
add.ftz.f32 %f675, %f4, %f63;
mul.ftz.f32 %f676, %f675, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f677, %f676;
add.ftz.f32 %f678, %f677, 0f3F800000;
div.approx.ftz.f32 %f679, %f675, %f678;
mul.ftz.f32 %f680, %f674, %f679;
st.global.f32 [%rd24+12], %f680;
bra.uni $L__BB88_78;

$L__BB88_25:
ld.shared.f32 %f133, [%r23];
add.ftz.f32 %f134, %f133, %f30;
ld.shared.f32 %f135, [%r23+512];
add.ftz.f32 %f136, %f135, %f134;
ld.shared.f32 %f137, [%r23+1024];
add.ftz.f32 %f138, %f137, %f136;
mov.b32 %r206, %f138;
mov.u32 %r207, 31;
mov.u32 %r208, 16;
mov.u32 %r209, -1;
shfl.sync.bfly.b32 %r210|%p22, %r206, %r208, %r207, %r209;
mov.b32 %f139, %r210;
add.ftz.f32 %f140, %f138, %f139;
mov.b32 %r211, %f140;
mov.u32 %r212, 8;
shfl.sync.bfly.b32 %r213|%p23, %r211, %r212, %r207, %r209;
mov.b32 %f141, %r213;
add.ftz.f32 %f142, %f140, %f141;
mov.b32 %r214, %f142;
mov.u32 %r215, 4;
shfl.sync.bfly.b32 %r216|%p24, %r214, %r215, %r207, %r209;
mov.b32 %f143, %r216;
add.ftz.f32 %f144, %f142, %f143;
mov.b32 %r217, %f144;
mov.u32 %r218, 2;
shfl.sync.bfly.b32 %r219|%p25, %r217, %r218, %r207, %r209;
mov.b32 %f145, %r219;
add.ftz.f32 %f146, %f144, %f145;
mov.b32 %r220, %f146;
mov.u32 %r221, 1;
shfl.sync.bfly.b32 %r222|%p26, %r220, %r221, %r207, %r209;
mov.b32 %f147, %r222;
add.ftz.f32 %f36, %f146, %f147;
setp.ne.s32 %p27, %r2, 0;
setp.ge.u32 %p28, %r4, %r42;
or.pred %p29, %p27, %p28;
@%p29 bra $L__BB88_27;

add.ftz.f32 %f148, %f5, %f36;
st.global.f32 [%rd24], %f148;

$L__BB88_27:
ld.shared.f32 %f149, [%r23+128];
add.ftz.f32 %f150, %f149, %f29;
ld.shared.f32 %f151, [%r23+640];
add.ftz.f32 %f152, %f151, %f150;
ld.shared.f32 %f153, [%r23+1152];
add.ftz.f32 %f154, %f153, %f152;
mov.b32 %r226, %f154;
shfl.sync.bfly.b32 %r230|%p30, %r226, %r208, %r207, %r209;
mov.b32 %f155, %r230;
add.ftz.f32 %f156, %f154, %f155;
mov.b32 %r231, %f156;
shfl.sync.bfly.b32 %r233|%p31, %r231, %r212, %r207, %r209;
mov.b32 %f157, %r233;
add.ftz.f32 %f158, %f156, %f157;
mov.b32 %r234, %f158;
shfl.sync.bfly.b32 %r236|%p32, %r234, %r215, %r207, %r209;
mov.b32 %f159, %r236;
add.ftz.f32 %f160, %f158, %f159;
mov.b32 %r237, %f160;
shfl.sync.bfly.b32 %r239|%p33, %r237, %r218, %r207, %r209;
mov.b32 %f161, %r239;
add.ftz.f32 %f162, %f160, %f161;
mov.b32 %r240, %f162;
shfl.sync.bfly.b32 %r242|%p34, %r240, %r221, %r207, %r209;
mov.b32 %f163, %r242;
add.ftz.f32 %f37, %f162, %f163;
setp.ne.s32 %p35, %r2, 1;
or.b32 %r246, %r4, 1;
setp.ge.u32 %p36, %r246, %r42;
or.pred %p37, %p35, %p36;
@%p37 bra $L__BB88_29;

add.ftz.f32 %f164, %f5, %f37;
st.global.f32 [%rd24+4], %f164;

$L__BB88_29:
ld.shared.f32 %f165, [%r23+256];
add.ftz.f32 %f166, %f165, %f28;
ld.shared.f32 %f167, [%r23+768];
add.ftz.f32 %f168, %f167, %f166;
ld.shared.f32 %f169, [%r23+1280];
add.ftz.f32 %f170, %f169, %f168;
mov.b32 %r247, %f170;
mov.u32 %r248, 31;
mov.u32 %r249, 16;
mov.u32 %r250, -1;
shfl.sync.bfly.b32 %r251|%p38, %r247, %r249, %r248, %r250;
mov.b32 %f171, %r251;
add.ftz.f32 %f172, %f170, %f171;
mov.b32 %r252, %f172;
mov.u32 %r253, 8;
shfl.sync.bfly.b32 %r254|%p39, %r252, %r253, %r248, %r250;
mov.b32 %f173, %r254;
add.ftz.f32 %f174, %f172, %f173;
mov.b32 %r255, %f174;
mov.u32 %r256, 4;
shfl.sync.bfly.b32 %r257|%p40, %r255, %r256, %r248, %r250;
mov.b32 %f175, %r257;
add.ftz.f32 %f176, %f174, %f175;
mov.b32 %r258, %f176;
mov.u32 %r259, 2;
shfl.sync.bfly.b32 %r260|%p41, %r258, %r259, %r248, %r250;
mov.b32 %f177, %r260;
add.ftz.f32 %f178, %f176, %f177;
mov.b32 %r261, %f178;
mov.u32 %r262, 1;
shfl.sync.bfly.b32 %r263|%p42, %r261, %r262, %r248, %r250;
mov.b32 %f179, %r263;
add.ftz.f32 %f38, %f178, %f179;
setp.ne.s32 %p43, %r2, 2;
or.b32 %r267, %r4, 2;
setp.ge.u32 %p44, %r267, %r42;
or.pred %p45, %p43, %p44;
@%p45 bra $L__BB88_31;

add.ftz.f32 %f180, %f5, %f38;
st.global.f32 [%rd24+8], %f180;

$L__BB88_31:
ld.shared.f32 %f181, [%r23+384];
add.ftz.f32 %f182, %f181, %f27;
ld.shared.f32 %f183, [%r23+896];
add.ftz.f32 %f184, %f183, %f182;
ld.shared.f32 %f185, [%r23+1408];
add.ftz.f32 %f186, %f185, %f184;
mov.b32 %r268, %f186;
shfl.sync.bfly.b32 %r272|%p46, %r268, %r249, %r248, %r250;
mov.b32 %f187, %r272;
add.ftz.f32 %f188, %f186, %f187;
mov.b32 %r273, %f188;
shfl.sync.bfly.b32 %r275|%p47, %r273, %r253, %r248, %r250;
mov.b32 %f189, %r275;
add.ftz.f32 %f190, %f188, %f189;
mov.b32 %r276, %f190;
shfl.sync.bfly.b32 %r278|%p48, %r276, %r256, %r248, %r250;
mov.b32 %f191, %r278;
add.ftz.f32 %f192, %f190, %f191;
mov.b32 %r279, %f192;
shfl.sync.bfly.b32 %r281|%p49, %r279, %r259, %r248, %r250;
mov.b32 %f193, %r281;
add.ftz.f32 %f194, %f192, %f193;
mov.b32 %r282, %f194;
shfl.sync.bfly.b32 %r284|%p50, %r282, %r262, %r248, %r250;
mov.b32 %f195, %r284;
add.ftz.f32 %f39, %f194, %f195;
setp.ne.s32 %p51, %r2, 3;
or.b32 %r288, %r4, 3;
setp.ge.u32 %p52, %r288, %r42;
or.pred %p53, %p51, %p52;
@%p53 bra $L__BB88_78;

add.ftz.f32 %f196, %f5, %f39;
st.global.f32 [%rd24+12], %f196;
bra.uni $L__BB88_78;

$L__BB88_44:
setp.eq.s32 %p55, %r36, 3;
@%p55 bra $L__BB88_54;

setp.ne.s32 %p56, %r36, 6;
@%p56 bra $L__BB88_70;

ld.shared.f32 %f197, [%r23];
add.ftz.f32 %f198, %f197, %f30;
ld.shared.f32 %f199, [%r25];
add.ftz.f32 %f200, %f199, %f34;
ld.shared.f32 %f201, [%r23+512];
add.ftz.f32 %f202, %f201, %f198;
ld.shared.f32 %f203, [%r25+512];
add.ftz.f32 %f204, %f203, %f200;
ld.shared.f32 %f205, [%r23+1024];
add.ftz.f32 %f206, %f205, %f202;
ld.shared.f32 %f207, [%r25+1024];
add.ftz.f32 %f208, %f207, %f204;
mov.b32 %r293, %f206;
mov.u32 %r294, 31;
mov.u32 %r295, 16;
mov.u32 %r296, -1;
shfl.sync.bfly.b32 %r297|%p59, %r293, %r295, %r294, %r296;
mov.b32 %f209, %r297;
add.ftz.f32 %f210, %f206, %f209;
mov.b32 %r298, %f210;
mov.u32 %r299, 8;
shfl.sync.bfly.b32 %r300|%p60, %r298, %r299, %r294, %r296;
mov.b32 %f211, %r300;
add.ftz.f32 %f212, %f210, %f211;
mov.b32 %r301, %f212;
mov.u32 %r302, 4;
shfl.sync.bfly.b32 %r303|%p61, %r301, %r302, %r294, %r296;
mov.b32 %f213, %r303;
add.ftz.f32 %f214, %f212, %f213;
mov.b32 %r304, %f214;
mov.u32 %r305, 2;
shfl.sync.bfly.b32 %r306|%p62, %r304, %r305, %r294, %r296;
mov.b32 %f215, %r306;
add.ftz.f32 %f216, %f214, %f215;
mov.b32 %r307, %f216;
mov.u32 %r308, 1;
shfl.sync.bfly.b32 %r309|%p63, %r307, %r308, %r294, %r296;
mov.b32 %f217, %r309;
add.ftz.f32 %f40, %f216, %f217;
mov.b32 %r310, %f208;
shfl.sync.bfly.b32 %r311|%p64, %r310, %r295, %r294, %r296;
mov.b32 %f218, %r311;
add.ftz.f32 %f219, %f208, %f218;
mov.b32 %r312, %f219;
shfl.sync.bfly.b32 %r313|%p65, %r312, %r299, %r294, %r296;
mov.b32 %f220, %r313;
add.ftz.f32 %f221, %f219, %f220;
mov.b32 %r314, %f221;
shfl.sync.bfly.b32 %r315|%p66, %r314, %r302, %r294, %r296;
mov.b32 %f222, %r315;
add.ftz.f32 %f223, %f221, %f222;
mov.b32 %r316, %f223;
shfl.sync.bfly.b32 %r317|%p67, %r316, %r305, %r294, %r296;
mov.b32 %f224, %r317;
add.ftz.f32 %f225, %f223, %f224;
mov.b32 %r318, %f225;
shfl.sync.bfly.b32 %r319|%p68, %r318, %r308, %r294, %r296;
mov.b32 %f226, %r319;
add.ftz.f32 %f41, %f225, %f226;
setp.ne.s32 %p69, %r2, 0;
setp.ge.u32 %p70, %r4, %r42;
or.pred %p71, %p69, %p70;
@%p71 bra $L__BB88_48;

ld.param.f32 %f981, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f227, %f5, %f40;
add.ftz.f32 %f228, %f4, %f41;
min.ftz.f32 %f229, %f228, %f981;
min.ftz.f32 %f230, %f227, %f981;
max.ftz.f32 %f231, %f230, %f35;
mul.ftz.f32 %f232, %f229, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f233, %f232;
add.ftz.f32 %f234, %f233, 0f3F800000;
div.approx.ftz.f32 %f235, %f229, %f234;
mul.ftz.f32 %f236, %f231, %f235;
st.global.f32 [%rd24], %f236;

$L__BB88_48:
ld.shared.f32 %f237, [%r23+128];
add.ftz.f32 %f238, %f237, %f29;
ld.shared.f32 %f239, [%r25+128];
add.ftz.f32 %f240, %f239, %f33;
ld.shared.f32 %f241, [%r23+640];
add.ftz.f32 %f242, %f241, %f238;
ld.shared.f32 %f243, [%r25+640];
add.ftz.f32 %f244, %f243, %f240;
ld.shared.f32 %f245, [%r23+1152];
add.ftz.f32 %f246, %f245, %f242;
ld.shared.f32 %f247, [%r25+1152];
add.ftz.f32 %f248, %f247, %f244;
mov.b32 %r323, %f246;
shfl.sync.bfly.b32 %r327|%p72, %r323, %r295, %r294, %r296;
mov.b32 %f249, %r327;
add.ftz.f32 %f250, %f246, %f249;
mov.b32 %r328, %f250;
shfl.sync.bfly.b32 %r330|%p73, %r328, %r299, %r294, %r296;
mov.b32 %f251, %r330;
add.ftz.f32 %f252, %f250, %f251;
mov.b32 %r331, %f252;
shfl.sync.bfly.b32 %r333|%p74, %r331, %r302, %r294, %r296;
mov.b32 %f253, %r333;
add.ftz.f32 %f254, %f252, %f253;
mov.b32 %r334, %f254;
shfl.sync.bfly.b32 %r336|%p75, %r334, %r305, %r294, %r296;
mov.b32 %f255, %r336;
add.ftz.f32 %f256, %f254, %f255;
mov.b32 %r337, %f256;
shfl.sync.bfly.b32 %r339|%p76, %r337, %r308, %r294, %r296;
mov.b32 %f257, %r339;
add.ftz.f32 %f42, %f256, %f257;
mov.b32 %r340, %f248;
shfl.sync.bfly.b32 %r341|%p77, %r340, %r295, %r294, %r296;
mov.b32 %f258, %r341;
add.ftz.f32 %f259, %f248, %f258;
mov.b32 %r342, %f259;
shfl.sync.bfly.b32 %r343|%p78, %r342, %r299, %r294, %r296;
mov.b32 %f260, %r343;
add.ftz.f32 %f261, %f259, %f260;
mov.b32 %r344, %f261;
shfl.sync.bfly.b32 %r345|%p79, %r344, %r302, %r294, %r296;
mov.b32 %f262, %r345;
add.ftz.f32 %f263, %f261, %f262;
mov.b32 %r346, %f263;
shfl.sync.bfly.b32 %r347|%p80, %r346, %r305, %r294, %r296;
mov.b32 %f264, %r347;
add.ftz.f32 %f265, %f263, %f264;
mov.b32 %r348, %f265;
shfl.sync.bfly.b32 %r349|%p81, %r348, %r308, %r294, %r296;
mov.b32 %f266, %r349;
add.ftz.f32 %f43, %f265, %f266;
setp.ne.s32 %p82, %r2, 1;
or.b32 %r353, %r4, 1;
setp.ge.u32 %p83, %r353, %r42;
or.pred %p84, %p82, %p83;
@%p84 bra $L__BB88_50;

ld.param.f32 %f980, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f267, %f5, %f42;
add.ftz.f32 %f268, %f4, %f43;
min.ftz.f32 %f269, %f268, %f980;
min.ftz.f32 %f270, %f267, %f980;
max.ftz.f32 %f271, %f270, %f35;
mul.ftz.f32 %f272, %f269, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f273, %f272;
add.ftz.f32 %f274, %f273, 0f3F800000;
div.approx.ftz.f32 %f275, %f269, %f274;
mul.ftz.f32 %f276, %f271, %f275;
st.global.f32 [%rd24+4], %f276;

$L__BB88_50:
ld.shared.f32 %f277, [%r23+256];
add.ftz.f32 %f278, %f277, %f28;
ld.shared.f32 %f279, [%r25+256];
add.ftz.f32 %f280, %f279, %f32;
ld.shared.f32 %f281, [%r23+768];
add.ftz.f32 %f282, %f281, %f278;
ld.shared.f32 %f283, [%r25+768];
add.ftz.f32 %f284, %f283, %f280;
ld.shared.f32 %f285, [%r23+1280];
add.ftz.f32 %f286, %f285, %f282;
ld.shared.f32 %f287, [%r25+1280];
add.ftz.f32 %f288, %f287, %f284;
mov.b32 %r354, %f286;
mov.u32 %r355, 31;
mov.u32 %r356, 16;
mov.u32 %r357, -1;
shfl.sync.bfly.b32 %r358|%p85, %r354, %r356, %r355, %r357;
mov.b32 %f289, %r358;
add.ftz.f32 %f290, %f286, %f289;
mov.b32 %r359, %f290;
mov.u32 %r360, 8;
shfl.sync.bfly.b32 %r361|%p86, %r359, %r360, %r355, %r357;
mov.b32 %f291, %r361;
add.ftz.f32 %f292, %f290, %f291;
mov.b32 %r362, %f292;
mov.u32 %r363, 4;
shfl.sync.bfly.b32 %r364|%p87, %r362, %r363, %r355, %r357;
mov.b32 %f293, %r364;
add.ftz.f32 %f294, %f292, %f293;
mov.b32 %r365, %f294;
mov.u32 %r366, 2;
shfl.sync.bfly.b32 %r367|%p88, %r365, %r366, %r355, %r357;
mov.b32 %f295, %r367;
add.ftz.f32 %f296, %f294, %f295;
mov.b32 %r368, %f296;
mov.u32 %r369, 1;
shfl.sync.bfly.b32 %r370|%p89, %r368, %r369, %r355, %r357;
mov.b32 %f297, %r370;
add.ftz.f32 %f44, %f296, %f297;
mov.b32 %r371, %f288;
shfl.sync.bfly.b32 %r372|%p90, %r371, %r356, %r355, %r357;
mov.b32 %f298, %r372;
add.ftz.f32 %f299, %f288, %f298;
mov.b32 %r373, %f299;
shfl.sync.bfly.b32 %r374|%p91, %r373, %r360, %r355, %r357;
mov.b32 %f300, %r374;
add.ftz.f32 %f301, %f299, %f300;
mov.b32 %r375, %f301;
shfl.sync.bfly.b32 %r376|%p92, %r375, %r363, %r355, %r357;
mov.b32 %f302, %r376;
add.ftz.f32 %f303, %f301, %f302;
mov.b32 %r377, %f303;
shfl.sync.bfly.b32 %r378|%p93, %r377, %r366, %r355, %r357;
mov.b32 %f304, %r378;
add.ftz.f32 %f305, %f303, %f304;
mov.b32 %r379, %f305;
shfl.sync.bfly.b32 %r380|%p94, %r379, %r369, %r355, %r357;
mov.b32 %f306, %r380;
add.ftz.f32 %f45, %f305, %f306;
setp.ne.s32 %p95, %r2, 2;
or.b32 %r384, %r4, 2;
setp.ge.u32 %p96, %r384, %r42;
or.pred %p97, %p95, %p96;
@%p97 bra $L__BB88_52;

ld.param.f32 %f979, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f307, %f5, %f44;
add.ftz.f32 %f308, %f4, %f45;
min.ftz.f32 %f309, %f308, %f979;
min.ftz.f32 %f310, %f307, %f979;
max.ftz.f32 %f311, %f310, %f35;
mul.ftz.f32 %f312, %f309, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f313, %f312;
add.ftz.f32 %f314, %f313, 0f3F800000;
div.approx.ftz.f32 %f315, %f309, %f314;
mul.ftz.f32 %f316, %f311, %f315;
st.global.f32 [%rd24+8], %f316;

$L__BB88_52:
ld.shared.f32 %f317, [%r23+384];
add.ftz.f32 %f318, %f317, %f27;
ld.shared.f32 %f319, [%r25+384];
add.ftz.f32 %f320, %f319, %f31;
ld.shared.f32 %f321, [%r23+896];
add.ftz.f32 %f322, %f321, %f318;
ld.shared.f32 %f323, [%r25+896];
add.ftz.f32 %f324, %f323, %f320;
ld.shared.f32 %f325, [%r23+1408];
add.ftz.f32 %f326, %f325, %f322;
ld.shared.f32 %f327, [%r25+1408];
add.ftz.f32 %f328, %f327, %f324;
mov.b32 %r385, %f326;
shfl.sync.bfly.b32 %r389|%p98, %r385, %r356, %r355, %r357;
mov.b32 %f329, %r389;
add.ftz.f32 %f330, %f326, %f329;
mov.b32 %r390, %f330;
shfl.sync.bfly.b32 %r392|%p99, %r390, %r360, %r355, %r357;
mov.b32 %f331, %r392;
add.ftz.f32 %f332, %f330, %f331;
mov.b32 %r393, %f332;
shfl.sync.bfly.b32 %r395|%p100, %r393, %r363, %r355, %r357;
mov.b32 %f333, %r395;
add.ftz.f32 %f334, %f332, %f333;
mov.b32 %r396, %f334;
shfl.sync.bfly.b32 %r398|%p101, %r396, %r366, %r355, %r357;
mov.b32 %f335, %r398;
add.ftz.f32 %f336, %f334, %f335;
mov.b32 %r399, %f336;
shfl.sync.bfly.b32 %r401|%p102, %r399, %r369, %r355, %r357;
mov.b32 %f337, %r401;
add.ftz.f32 %f46, %f336, %f337;
mov.b32 %r402, %f328;
shfl.sync.bfly.b32 %r403|%p103, %r402, %r356, %r355, %r357;
mov.b32 %f338, %r403;
add.ftz.f32 %f339, %f328, %f338;
mov.b32 %r404, %f339;
shfl.sync.bfly.b32 %r405|%p104, %r404, %r360, %r355, %r357;
mov.b32 %f340, %r405;
add.ftz.f32 %f341, %f339, %f340;
mov.b32 %r406, %f341;
shfl.sync.bfly.b32 %r407|%p105, %r406, %r363, %r355, %r357;
mov.b32 %f342, %r407;
add.ftz.f32 %f343, %f341, %f342;
mov.b32 %r408, %f343;
shfl.sync.bfly.b32 %r409|%p106, %r408, %r366, %r355, %r357;
mov.b32 %f344, %r409;
add.ftz.f32 %f345, %f343, %f344;
mov.b32 %r410, %f345;
shfl.sync.bfly.b32 %r411|%p107, %r410, %r369, %r355, %r357;
mov.b32 %f346, %r411;
add.ftz.f32 %f47, %f345, %f346;
setp.ne.s32 %p108, %r2, 3;
or.b32 %r415, %r4, 3;
setp.ge.u32 %p109, %r415, %r42;
or.pred %p110, %p108, %p109;
@%p110 bra $L__BB88_78;

ld.param.f32 %f978, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f347, %f5, %f46;
add.ftz.f32 %f348, %f4, %f47;
min.ftz.f32 %f349, %f348, %f978;
min.ftz.f32 %f350, %f347, %f978;
max.ftz.f32 %f351, %f350, %f35;
mul.ftz.f32 %f352, %f349, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f353, %f352;
add.ftz.f32 %f354, %f353, 0f3F800000;
div.approx.ftz.f32 %f355, %f349, %f354;
mul.ftz.f32 %f356, %f351, %f355;
st.global.f32 [%rd24+12], %f356;
bra.uni $L__BB88_78;

$L__BB88_70:
ld.shared.f32 %f845, [%r23];
add.ftz.f32 %f846, %f845, %f30;
ld.shared.f32 %f847, [%r25];
add.ftz.f32 %f848, %f847, %f34;
ld.shared.f32 %f849, [%r23+512];
add.ftz.f32 %f850, %f849, %f846;
ld.shared.f32 %f851, [%r25+512];
add.ftz.f32 %f852, %f851, %f848;
ld.shared.f32 %f853, [%r23+1024];
add.ftz.f32 %f854, %f853, %f850;
ld.shared.f32 %f855, [%r25+1024];
add.ftz.f32 %f856, %f855, %f852;
mov.b32 %r785, %f854;
mov.u32 %r786, 31;
mov.u32 %r787, 16;
mov.u32 %r788, -1;
shfl.sync.bfly.b32 %r789|%p267, %r785, %r787, %r786, %r788;
mov.b32 %f857, %r789;
add.ftz.f32 %f858, %f854, %f857;
mov.b32 %r790, %f858;
mov.u32 %r791, 8;
shfl.sync.bfly.b32 %r792|%p268, %r790, %r791, %r786, %r788;
mov.b32 %f859, %r792;
add.ftz.f32 %f860, %f858, %f859;
mov.b32 %r793, %f860;
mov.u32 %r794, 4;
shfl.sync.bfly.b32 %r795|%p269, %r793, %r794, %r786, %r788;
mov.b32 %f861, %r795;
add.ftz.f32 %f862, %f860, %f861;
mov.b32 %r796, %f862;
mov.u32 %r797, 2;
shfl.sync.bfly.b32 %r798|%p270, %r796, %r797, %r786, %r788;
mov.b32 %f863, %r798;
add.ftz.f32 %f864, %f862, %f863;
mov.b32 %r799, %f864;
mov.u32 %r800, 1;
shfl.sync.bfly.b32 %r801|%p271, %r799, %r800, %r786, %r788;
mov.b32 %f865, %r801;
add.ftz.f32 %f72, %f864, %f865;
mov.b32 %r802, %f856;
shfl.sync.bfly.b32 %r803|%p272, %r802, %r787, %r786, %r788;
mov.b32 %f866, %r803;
add.ftz.f32 %f867, %f856, %f866;
mov.b32 %r804, %f867;
shfl.sync.bfly.b32 %r805|%p273, %r804, %r791, %r786, %r788;
mov.b32 %f868, %r805;
add.ftz.f32 %f869, %f867, %f868;
mov.b32 %r806, %f869;
shfl.sync.bfly.b32 %r807|%p274, %r806, %r794, %r786, %r788;
mov.b32 %f870, %r807;
add.ftz.f32 %f871, %f869, %f870;
mov.b32 %r808, %f871;
shfl.sync.bfly.b32 %r809|%p275, %r808, %r797, %r786, %r788;
mov.b32 %f872, %r809;
add.ftz.f32 %f873, %f871, %f872;
mov.b32 %r810, %f873;
shfl.sync.bfly.b32 %r811|%p276, %r810, %r800, %r786, %r788;
mov.b32 %f874, %r811;
add.ftz.f32 %f73, %f873, %f874;
setp.ne.s32 %p277, %r2, 0;
setp.ge.u32 %p278, %r4, %r42;
or.pred %p279, %p277, %p278;
@%p279 bra $L__BB88_72;

add.ftz.f32 %f875, %f5, %f72;
add.ftz.f32 %f876, %f4, %f73;
mul.ftz.f32 %f877, %f875, %f876;
st.global.f32 [%rd24], %f877;

$L__BB88_72:
ld.shared.f32 %f878, [%r23+128];
add.ftz.f32 %f879, %f878, %f29;
ld.shared.f32 %f880, [%r25+128];
add.ftz.f32 %f881, %f880, %f33;
ld.shared.f32 %f882, [%r23+640];
add.ftz.f32 %f883, %f882, %f879;
ld.shared.f32 %f884, [%r25+640];
add.ftz.f32 %f885, %f884, %f881;
ld.shared.f32 %f886, [%r23+1152];
add.ftz.f32 %f887, %f886, %f883;
ld.shared.f32 %f888, [%r25+1152];
add.ftz.f32 %f889, %f888, %f885;
mov.b32 %r815, %f887;
shfl.sync.bfly.b32 %r819|%p280, %r815, %r787, %r786, %r788;
mov.b32 %f890, %r819;
add.ftz.f32 %f891, %f887, %f890;
mov.b32 %r820, %f891;
shfl.sync.bfly.b32 %r822|%p281, %r820, %r791, %r786, %r788;
mov.b32 %f892, %r822;
add.ftz.f32 %f893, %f891, %f892;
mov.b32 %r823, %f893;
shfl.sync.bfly.b32 %r825|%p282, %r823, %r794, %r786, %r788;
mov.b32 %f894, %r825;
add.ftz.f32 %f895, %f893, %f894;
mov.b32 %r826, %f895;
shfl.sync.bfly.b32 %r828|%p283, %r826, %r797, %r786, %r788;
mov.b32 %f896, %r828;
add.ftz.f32 %f897, %f895, %f896;
mov.b32 %r829, %f897;
shfl.sync.bfly.b32 %r831|%p284, %r829, %r800, %r786, %r788;
mov.b32 %f898, %r831;
add.ftz.f32 %f74, %f897, %f898;
mov.b32 %r832, %f889;
shfl.sync.bfly.b32 %r833|%p285, %r832, %r787, %r786, %r788;
mov.b32 %f899, %r833;
add.ftz.f32 %f900, %f889, %f899;
mov.b32 %r834, %f900;
shfl.sync.bfly.b32 %r835|%p286, %r834, %r791, %r786, %r788;
mov.b32 %f901, %r835;
add.ftz.f32 %f902, %f900, %f901;
mov.b32 %r836, %f902;
shfl.sync.bfly.b32 %r837|%p287, %r836, %r794, %r786, %r788;
mov.b32 %f903, %r837;
add.ftz.f32 %f904, %f902, %f903;
mov.b32 %r838, %f904;
shfl.sync.bfly.b32 %r839|%p288, %r838, %r797, %r786, %r788;
mov.b32 %f905, %r839;
add.ftz.f32 %f906, %f904, %f905;
mov.b32 %r840, %f906;
shfl.sync.bfly.b32 %r841|%p289, %r840, %r800, %r786, %r788;
mov.b32 %f907, %r841;
add.ftz.f32 %f75, %f906, %f907;
setp.ne.s32 %p290, %r2, 1;
or.b32 %r845, %r4, 1;
setp.ge.u32 %p291, %r845, %r42;
or.pred %p292, %p290, %p291;
@%p292 bra $L__BB88_74;

add.ftz.f32 %f908, %f5, %f74;
add.ftz.f32 %f909, %f4, %f75;
mul.ftz.f32 %f910, %f908, %f909;
st.global.f32 [%rd24+4], %f910;

$L__BB88_74:
ld.shared.f32 %f911, [%r23+256];
add.ftz.f32 %f912, %f911, %f28;
ld.shared.f32 %f913, [%r25+256];
add.ftz.f32 %f914, %f913, %f32;
ld.shared.f32 %f915, [%r23+768];
add.ftz.f32 %f916, %f915, %f912;
ld.shared.f32 %f917, [%r25+768];
add.ftz.f32 %f918, %f917, %f914;
ld.shared.f32 %f919, [%r23+1280];
add.ftz.f32 %f920, %f919, %f916;
ld.shared.f32 %f921, [%r25+1280];
add.ftz.f32 %f922, %f921, %f918;
mov.b32 %r846, %f920;
mov.u32 %r847, 31;
mov.u32 %r848, 16;
mov.u32 %r849, -1;
shfl.sync.bfly.b32 %r850|%p293, %r846, %r848, %r847, %r849;
mov.b32 %f923, %r850;
add.ftz.f32 %f924, %f920, %f923;
mov.b32 %r851, %f924;
mov.u32 %r852, 8;
shfl.sync.bfly.b32 %r853|%p294, %r851, %r852, %r847, %r849;
mov.b32 %f925, %r853;
add.ftz.f32 %f926, %f924, %f925;
mov.b32 %r854, %f926;
mov.u32 %r855, 4;
shfl.sync.bfly.b32 %r856|%p295, %r854, %r855, %r847, %r849;
mov.b32 %f927, %r856;
add.ftz.f32 %f928, %f926, %f927;
mov.b32 %r857, %f928;
mov.u32 %r858, 2;
shfl.sync.bfly.b32 %r859|%p296, %r857, %r858, %r847, %r849;
mov.b32 %f929, %r859;
add.ftz.f32 %f930, %f928, %f929;
mov.b32 %r860, %f930;
mov.u32 %r861, 1;
shfl.sync.bfly.b32 %r862|%p297, %r860, %r861, %r847, %r849;
mov.b32 %f931, %r862;
add.ftz.f32 %f76, %f930, %f931;
mov.b32 %r863, %f922;
shfl.sync.bfly.b32 %r864|%p298, %r863, %r848, %r847, %r849;
mov.b32 %f932, %r864;
add.ftz.f32 %f933, %f922, %f932;
mov.b32 %r865, %f933;
shfl.sync.bfly.b32 %r866|%p299, %r865, %r852, %r847, %r849;
mov.b32 %f934, %r866;
add.ftz.f32 %f935, %f933, %f934;
mov.b32 %r867, %f935;
shfl.sync.bfly.b32 %r868|%p300, %r867, %r855, %r847, %r849;
mov.b32 %f936, %r868;
add.ftz.f32 %f937, %f935, %f936;
mov.b32 %r869, %f937;
shfl.sync.bfly.b32 %r870|%p301, %r869, %r858, %r847, %r849;
mov.b32 %f938, %r870;
add.ftz.f32 %f939, %f937, %f938;
mov.b32 %r871, %f939;
shfl.sync.bfly.b32 %r872|%p302, %r871, %r861, %r847, %r849;
mov.b32 %f940, %r872;
add.ftz.f32 %f77, %f939, %f940;
setp.ne.s32 %p303, %r2, 2;
or.b32 %r876, %r4, 2;
setp.ge.u32 %p304, %r876, %r42;
or.pred %p305, %p303, %p304;
@%p305 bra $L__BB88_76;

add.ftz.f32 %f941, %f5, %f76;
add.ftz.f32 %f942, %f4, %f77;
mul.ftz.f32 %f943, %f941, %f942;
st.global.f32 [%rd24+8], %f943;

$L__BB88_76:
ld.shared.f32 %f944, [%r23+384];
add.ftz.f32 %f945, %f944, %f27;
ld.shared.f32 %f946, [%r25+384];
add.ftz.f32 %f947, %f946, %f31;
ld.shared.f32 %f948, [%r23+896];
add.ftz.f32 %f949, %f948, %f945;
ld.shared.f32 %f950, [%r25+896];
add.ftz.f32 %f951, %f950, %f947;
ld.shared.f32 %f952, [%r23+1408];
add.ftz.f32 %f953, %f952, %f949;
ld.shared.f32 %f954, [%r25+1408];
add.ftz.f32 %f955, %f954, %f951;
mov.b32 %r877, %f953;
shfl.sync.bfly.b32 %r881|%p306, %r877, %r848, %r847, %r849;
mov.b32 %f956, %r881;
add.ftz.f32 %f957, %f953, %f956;
mov.b32 %r882, %f957;
shfl.sync.bfly.b32 %r884|%p307, %r882, %r852, %r847, %r849;
mov.b32 %f958, %r884;
add.ftz.f32 %f959, %f957, %f958;
mov.b32 %r885, %f959;
shfl.sync.bfly.b32 %r887|%p308, %r885, %r855, %r847, %r849;
mov.b32 %f960, %r887;
add.ftz.f32 %f961, %f959, %f960;
mov.b32 %r888, %f961;
shfl.sync.bfly.b32 %r890|%p309, %r888, %r858, %r847, %r849;
mov.b32 %f962, %r890;
add.ftz.f32 %f963, %f961, %f962;
mov.b32 %r891, %f963;
shfl.sync.bfly.b32 %r893|%p310, %r891, %r861, %r847, %r849;
mov.b32 %f964, %r893;
add.ftz.f32 %f78, %f963, %f964;
mov.b32 %r894, %f955;
shfl.sync.bfly.b32 %r895|%p311, %r894, %r848, %r847, %r849;
mov.b32 %f965, %r895;
add.ftz.f32 %f966, %f955, %f965;
mov.b32 %r896, %f966;
shfl.sync.bfly.b32 %r897|%p312, %r896, %r852, %r847, %r849;
mov.b32 %f967, %r897;
add.ftz.f32 %f968, %f966, %f967;
mov.b32 %r898, %f968;
shfl.sync.bfly.b32 %r899|%p313, %r898, %r855, %r847, %r849;
mov.b32 %f969, %r899;
add.ftz.f32 %f970, %f968, %f969;
mov.b32 %r900, %f970;
shfl.sync.bfly.b32 %r901|%p314, %r900, %r858, %r847, %r849;
mov.b32 %f971, %r901;
add.ftz.f32 %f972, %f970, %f971;
mov.b32 %r902, %f972;
shfl.sync.bfly.b32 %r903|%p315, %r902, %r861, %r847, %r849;
mov.b32 %f973, %r903;
add.ftz.f32 %f79, %f972, %f973;
setp.ne.s32 %p316, %r2, 3;
or.b32 %r907, %r4, 3;
setp.ge.u32 %p317, %r907, %r42;
or.pred %p318, %p316, %p317;
@%p318 bra $L__BB88_78;

add.ftz.f32 %f974, %f5, %f78;
add.ftz.f32 %f975, %f4, %f79;
mul.ftz.f32 %f976, %f974, %f975;
st.global.f32 [%rd24+12], %f976;
bra.uni $L__BB88_78;

$L__BB88_62:
ld.shared.f32 %f681, [%r23];
add.ftz.f32 %f682, %f681, %f30;
ld.shared.f32 %f683, [%r25];
add.ftz.f32 %f684, %f683, %f34;
ld.shared.f32 %f685, [%r23+512];
add.ftz.f32 %f686, %f685, %f682;
ld.shared.f32 %f687, [%r25+512];
add.ftz.f32 %f688, %f687, %f684;
ld.shared.f32 %f689, [%r23+1024];
add.ftz.f32 %f690, %f689, %f686;
ld.shared.f32 %f691, [%r25+1024];
add.ftz.f32 %f692, %f691, %f688;
mov.b32 %r662, %f690;
mov.u32 %r663, 31;
mov.u32 %r664, 16;
mov.u32 %r665, -1;
shfl.sync.bfly.b32 %r666|%p215, %r662, %r664, %r663, %r665;
mov.b32 %f693, %r666;
add.ftz.f32 %f694, %f690, %f693;
mov.b32 %r667, %f694;
mov.u32 %r668, 8;
shfl.sync.bfly.b32 %r669|%p216, %r667, %r668, %r663, %r665;
mov.b32 %f695, %r669;
add.ftz.f32 %f696, %f694, %f695;
mov.b32 %r670, %f696;
mov.u32 %r671, 4;
shfl.sync.bfly.b32 %r672|%p217, %r670, %r671, %r663, %r665;
mov.b32 %f697, %r672;
add.ftz.f32 %f698, %f696, %f697;
mov.b32 %r673, %f698;
mov.u32 %r674, 2;
shfl.sync.bfly.b32 %r675|%p218, %r673, %r674, %r663, %r665;
mov.b32 %f699, %r675;
add.ftz.f32 %f700, %f698, %f699;
mov.b32 %r676, %f700;
mov.u32 %r677, 1;
shfl.sync.bfly.b32 %r678|%p219, %r676, %r677, %r663, %r665;
mov.b32 %f701, %r678;
add.ftz.f32 %f64, %f700, %f701;
mov.b32 %r679, %f692;
shfl.sync.bfly.b32 %r680|%p220, %r679, %r664, %r663, %r665;
mov.b32 %f702, %r680;
add.ftz.f32 %f703, %f692, %f702;
mov.b32 %r681, %f703;
shfl.sync.bfly.b32 %r682|%p221, %r681, %r668, %r663, %r665;
mov.b32 %f704, %r682;
add.ftz.f32 %f705, %f703, %f704;
mov.b32 %r683, %f705;
shfl.sync.bfly.b32 %r684|%p222, %r683, %r671, %r663, %r665;
mov.b32 %f706, %r684;
add.ftz.f32 %f707, %f705, %f706;
mov.b32 %r685, %f707;
shfl.sync.bfly.b32 %r686|%p223, %r685, %r674, %r663, %r665;
mov.b32 %f708, %r686;
add.ftz.f32 %f709, %f707, %f708;
mov.b32 %r687, %f709;
shfl.sync.bfly.b32 %r688|%p224, %r687, %r677, %r663, %r665;
mov.b32 %f710, %r688;
add.ftz.f32 %f65, %f709, %f710;
setp.ne.s32 %p225, %r2, 0;
setp.ge.u32 %p226, %r4, %r42;
or.pred %p227, %p225, %p226;
@%p227 bra $L__BB88_64;

add.ftz.f32 %f711, %f5, %f64;
add.ftz.f32 %f712, %f4, %f65;
mul.ftz.f32 %f713, %f712, 0f3F000000;
mul.ftz.f32 %f714, %f712, 0f3F4C422A;
mul.ftz.f32 %f715, %f712, 0f3D372713;
fma.rn.ftz.f32 %f716, %f712, %f715, 0f3F800000;
mul.ftz.f32 %f717, %f714, %f716;
tanh.approx.f32 %f718, %f717;
add.ftz.f32 %f719, %f718, 0f3F800000;
mul.ftz.f32 %f720, %f713, %f719;
mul.ftz.f32 %f721, %f711, %f720;
st.global.f32 [%rd24], %f721;

$L__BB88_64:
ld.shared.f32 %f722, [%r23+128];
add.ftz.f32 %f723, %f722, %f29;
ld.shared.f32 %f724, [%r25+128];
add.ftz.f32 %f725, %f724, %f33;
ld.shared.f32 %f726, [%r23+640];
add.ftz.f32 %f727, %f726, %f723;
ld.shared.f32 %f728, [%r25+640];
add.ftz.f32 %f729, %f728, %f725;
ld.shared.f32 %f730, [%r23+1152];
add.ftz.f32 %f731, %f730, %f727;
ld.shared.f32 %f732, [%r25+1152];
add.ftz.f32 %f733, %f732, %f729;
mov.b32 %r692, %f731;
shfl.sync.bfly.b32 %r696|%p228, %r692, %r664, %r663, %r665;
mov.b32 %f734, %r696;
add.ftz.f32 %f735, %f731, %f734;
mov.b32 %r697, %f735;
shfl.sync.bfly.b32 %r699|%p229, %r697, %r668, %r663, %r665;
mov.b32 %f736, %r699;
add.ftz.f32 %f737, %f735, %f736;
mov.b32 %r700, %f737;
shfl.sync.bfly.b32 %r702|%p230, %r700, %r671, %r663, %r665;
mov.b32 %f738, %r702;
add.ftz.f32 %f739, %f737, %f738;
mov.b32 %r703, %f739;
shfl.sync.bfly.b32 %r705|%p231, %r703, %r674, %r663, %r665;
mov.b32 %f740, %r705;
add.ftz.f32 %f741, %f739, %f740;
mov.b32 %r706, %f741;
shfl.sync.bfly.b32 %r708|%p232, %r706, %r677, %r663, %r665;
mov.b32 %f742, %r708;
add.ftz.f32 %f66, %f741, %f742;
mov.b32 %r709, %f733;
shfl.sync.bfly.b32 %r710|%p233, %r709, %r664, %r663, %r665;
mov.b32 %f743, %r710;
add.ftz.f32 %f744, %f733, %f743;
mov.b32 %r711, %f744;
shfl.sync.bfly.b32 %r712|%p234, %r711, %r668, %r663, %r665;
mov.b32 %f745, %r712;
add.ftz.f32 %f746, %f744, %f745;
mov.b32 %r713, %f746;
shfl.sync.bfly.b32 %r714|%p235, %r713, %r671, %r663, %r665;
mov.b32 %f747, %r714;
add.ftz.f32 %f748, %f746, %f747;
mov.b32 %r715, %f748;
shfl.sync.bfly.b32 %r716|%p236, %r715, %r674, %r663, %r665;
mov.b32 %f749, %r716;
add.ftz.f32 %f750, %f748, %f749;
mov.b32 %r717, %f750;
shfl.sync.bfly.b32 %r718|%p237, %r717, %r677, %r663, %r665;
mov.b32 %f751, %r718;
add.ftz.f32 %f67, %f750, %f751;
setp.ne.s32 %p238, %r2, 1;
or.b32 %r722, %r4, 1;
setp.ge.u32 %p239, %r722, %r42;
or.pred %p240, %p238, %p239;
@%p240 bra $L__BB88_66;

add.ftz.f32 %f752, %f5, %f66;
add.ftz.f32 %f753, %f4, %f67;
mul.ftz.f32 %f754, %f753, 0f3F000000;
mul.ftz.f32 %f755, %f753, 0f3F4C422A;
mul.ftz.f32 %f756, %f753, 0f3D372713;
fma.rn.ftz.f32 %f757, %f753, %f756, 0f3F800000;
mul.ftz.f32 %f758, %f755, %f757;
tanh.approx.f32 %f759, %f758;
add.ftz.f32 %f760, %f759, 0f3F800000;
mul.ftz.f32 %f761, %f754, %f760;
mul.ftz.f32 %f762, %f752, %f761;
st.global.f32 [%rd24+4], %f762;

$L__BB88_66:
ld.shared.f32 %f763, [%r23+256];
add.ftz.f32 %f764, %f763, %f28;
ld.shared.f32 %f765, [%r25+256];
add.ftz.f32 %f766, %f765, %f32;
ld.shared.f32 %f767, [%r23+768];
add.ftz.f32 %f768, %f767, %f764;
ld.shared.f32 %f769, [%r25+768];
add.ftz.f32 %f770, %f769, %f766;
ld.shared.f32 %f771, [%r23+1280];
add.ftz.f32 %f772, %f771, %f768;
ld.shared.f32 %f773, [%r25+1280];
add.ftz.f32 %f774, %f773, %f770;
mov.b32 %r723, %f772;
mov.u32 %r724, 31;
mov.u32 %r725, 16;
mov.u32 %r726, -1;
shfl.sync.bfly.b32 %r727|%p241, %r723, %r725, %r724, %r726;
mov.b32 %f775, %r727;
add.ftz.f32 %f776, %f772, %f775;
mov.b32 %r728, %f776;
mov.u32 %r729, 8;
shfl.sync.bfly.b32 %r730|%p242, %r728, %r729, %r724, %r726;
mov.b32 %f777, %r730;
add.ftz.f32 %f778, %f776, %f777;
mov.b32 %r731, %f778;
mov.u32 %r732, 4;
shfl.sync.bfly.b32 %r733|%p243, %r731, %r732, %r724, %r726;
mov.b32 %f779, %r733;
add.ftz.f32 %f780, %f778, %f779;
mov.b32 %r734, %f780;
mov.u32 %r735, 2;
shfl.sync.bfly.b32 %r736|%p244, %r734, %r735, %r724, %r726;
mov.b32 %f781, %r736;
add.ftz.f32 %f782, %f780, %f781;
mov.b32 %r737, %f782;
mov.u32 %r738, 1;
shfl.sync.bfly.b32 %r739|%p245, %r737, %r738, %r724, %r726;
mov.b32 %f783, %r739;
add.ftz.f32 %f68, %f782, %f783;
mov.b32 %r740, %f774;
shfl.sync.bfly.b32 %r741|%p246, %r740, %r725, %r724, %r726;
mov.b32 %f784, %r741;
add.ftz.f32 %f785, %f774, %f784;
mov.b32 %r742, %f785;
shfl.sync.bfly.b32 %r743|%p247, %r742, %r729, %r724, %r726;
mov.b32 %f786, %r743;
add.ftz.f32 %f787, %f785, %f786;
mov.b32 %r744, %f787;
shfl.sync.bfly.b32 %r745|%p248, %r744, %r732, %r724, %r726;
mov.b32 %f788, %r745;
add.ftz.f32 %f789, %f787, %f788;
mov.b32 %r746, %f789;
shfl.sync.bfly.b32 %r747|%p249, %r746, %r735, %r724, %r726;
mov.b32 %f790, %r747;
add.ftz.f32 %f791, %f789, %f790;
mov.b32 %r748, %f791;
shfl.sync.bfly.b32 %r749|%p250, %r748, %r738, %r724, %r726;
mov.b32 %f792, %r749;
add.ftz.f32 %f69, %f791, %f792;
setp.ne.s32 %p251, %r2, 2;
or.b32 %r753, %r4, 2;
setp.ge.u32 %p252, %r753, %r42;
or.pred %p253, %p251, %p252;
@%p253 bra $L__BB88_68;

add.ftz.f32 %f793, %f5, %f68;
add.ftz.f32 %f794, %f4, %f69;
mul.ftz.f32 %f795, %f794, 0f3F000000;
mul.ftz.f32 %f796, %f794, 0f3F4C422A;
mul.ftz.f32 %f797, %f794, 0f3D372713;
fma.rn.ftz.f32 %f798, %f794, %f797, 0f3F800000;
mul.ftz.f32 %f799, %f796, %f798;
tanh.approx.f32 %f800, %f799;
add.ftz.f32 %f801, %f800, 0f3F800000;
mul.ftz.f32 %f802, %f795, %f801;
mul.ftz.f32 %f803, %f793, %f802;
st.global.f32 [%rd24+8], %f803;

$L__BB88_68:
ld.shared.f32 %f804, [%r23+384];
add.ftz.f32 %f805, %f804, %f27;
ld.shared.f32 %f806, [%r25+384];
add.ftz.f32 %f807, %f806, %f31;
ld.shared.f32 %f808, [%r23+896];
add.ftz.f32 %f809, %f808, %f805;
ld.shared.f32 %f810, [%r25+896];
add.ftz.f32 %f811, %f810, %f807;
ld.shared.f32 %f812, [%r23+1408];
add.ftz.f32 %f813, %f812, %f809;
ld.shared.f32 %f814, [%r25+1408];
add.ftz.f32 %f815, %f814, %f811;
mov.b32 %r754, %f813;
shfl.sync.bfly.b32 %r758|%p254, %r754, %r725, %r724, %r726;
mov.b32 %f816, %r758;
add.ftz.f32 %f817, %f813, %f816;
mov.b32 %r759, %f817;
shfl.sync.bfly.b32 %r761|%p255, %r759, %r729, %r724, %r726;
mov.b32 %f818, %r761;
add.ftz.f32 %f819, %f817, %f818;
mov.b32 %r762, %f819;
shfl.sync.bfly.b32 %r764|%p256, %r762, %r732, %r724, %r726;
mov.b32 %f820, %r764;
add.ftz.f32 %f821, %f819, %f820;
mov.b32 %r765, %f821;
shfl.sync.bfly.b32 %r767|%p257, %r765, %r735, %r724, %r726;
mov.b32 %f822, %r767;
add.ftz.f32 %f823, %f821, %f822;
mov.b32 %r768, %f823;
shfl.sync.bfly.b32 %r770|%p258, %r768, %r738, %r724, %r726;
mov.b32 %f824, %r770;
add.ftz.f32 %f70, %f823, %f824;
mov.b32 %r771, %f815;
shfl.sync.bfly.b32 %r772|%p259, %r771, %r725, %r724, %r726;
mov.b32 %f825, %r772;
add.ftz.f32 %f826, %f815, %f825;
mov.b32 %r773, %f826;
shfl.sync.bfly.b32 %r774|%p260, %r773, %r729, %r724, %r726;
mov.b32 %f827, %r774;
add.ftz.f32 %f828, %f826, %f827;
mov.b32 %r775, %f828;
shfl.sync.bfly.b32 %r776|%p261, %r775, %r732, %r724, %r726;
mov.b32 %f829, %r776;
add.ftz.f32 %f830, %f828, %f829;
mov.b32 %r777, %f830;
shfl.sync.bfly.b32 %r778|%p262, %r777, %r735, %r724, %r726;
mov.b32 %f831, %r778;
add.ftz.f32 %f832, %f830, %f831;
mov.b32 %r779, %f832;
shfl.sync.bfly.b32 %r780|%p263, %r779, %r738, %r724, %r726;
mov.b32 %f833, %r780;
add.ftz.f32 %f71, %f832, %f833;
setp.ne.s32 %p264, %r2, 3;
or.b32 %r784, %r4, 3;
setp.ge.u32 %p265, %r784, %r42;
or.pred %p266, %p264, %p265;
@%p266 bra $L__BB88_78;

add.ftz.f32 %f834, %f5, %f70;
add.ftz.f32 %f835, %f4, %f71;
mul.ftz.f32 %f836, %f835, 0f3F000000;
mul.ftz.f32 %f837, %f835, 0f3F4C422A;
mul.ftz.f32 %f838, %f835, 0f3D372713;
fma.rn.ftz.f32 %f839, %f835, %f838, 0f3F800000;
mul.ftz.f32 %f840, %f837, %f839;
tanh.approx.f32 %f841, %f840;
add.ftz.f32 %f842, %f841, 0f3F800000;
mul.ftz.f32 %f843, %f836, %f842;
mul.ftz.f32 %f844, %f834, %f843;
st.global.f32 [%rd24+12], %f844;
bra.uni $L__BB88_78;

$L__BB88_54:
ld.shared.f32 %f357, [%r23];
add.ftz.f32 %f358, %f357, %f30;
ld.shared.f32 %f359, [%r25];
add.ftz.f32 %f360, %f359, %f34;
ld.shared.f32 %f361, [%r23+512];
add.ftz.f32 %f362, %f361, %f358;
ld.shared.f32 %f363, [%r25+512];
add.ftz.f32 %f364, %f363, %f360;
ld.shared.f32 %f365, [%r23+1024];
add.ftz.f32 %f366, %f365, %f362;
ld.shared.f32 %f367, [%r25+1024];
add.ftz.f32 %f368, %f367, %f364;
mov.b32 %r416, %f366;
mov.u32 %r417, 31;
mov.u32 %r418, 16;
mov.u32 %r419, -1;
shfl.sync.bfly.b32 %r420|%p111, %r416, %r418, %r417, %r419;
mov.b32 %f369, %r420;
add.ftz.f32 %f370, %f366, %f369;
mov.b32 %r421, %f370;
mov.u32 %r422, 8;
shfl.sync.bfly.b32 %r423|%p112, %r421, %r422, %r417, %r419;
mov.b32 %f371, %r423;
add.ftz.f32 %f372, %f370, %f371;
mov.b32 %r424, %f372;
mov.u32 %r425, 4;
shfl.sync.bfly.b32 %r426|%p113, %r424, %r425, %r417, %r419;
mov.b32 %f373, %r426;
add.ftz.f32 %f374, %f372, %f373;
mov.b32 %r427, %f374;
mov.u32 %r428, 2;
shfl.sync.bfly.b32 %r429|%p114, %r427, %r428, %r417, %r419;
mov.b32 %f375, %r429;
add.ftz.f32 %f376, %f374, %f375;
mov.b32 %r430, %f376;
mov.u32 %r431, 1;
shfl.sync.bfly.b32 %r432|%p115, %r430, %r431, %r417, %r419;
mov.b32 %f377, %r432;
add.ftz.f32 %f48, %f376, %f377;
mov.b32 %r433, %f368;
shfl.sync.bfly.b32 %r434|%p116, %r433, %r418, %r417, %r419;
mov.b32 %f378, %r434;
add.ftz.f32 %f379, %f368, %f378;
mov.b32 %r435, %f379;
shfl.sync.bfly.b32 %r436|%p117, %r435, %r422, %r417, %r419;
mov.b32 %f380, %r436;
add.ftz.f32 %f381, %f379, %f380;
mov.b32 %r437, %f381;
shfl.sync.bfly.b32 %r438|%p118, %r437, %r425, %r417, %r419;
mov.b32 %f382, %r438;
add.ftz.f32 %f383, %f381, %f382;
mov.b32 %r439, %f383;
shfl.sync.bfly.b32 %r440|%p119, %r439, %r428, %r417, %r419;
mov.b32 %f384, %r440;
add.ftz.f32 %f385, %f383, %f384;
mov.b32 %r441, %f385;
shfl.sync.bfly.b32 %r442|%p120, %r441, %r431, %r417, %r419;
mov.b32 %f386, %r442;
add.ftz.f32 %f49, %f385, %f386;
setp.ne.s32 %p121, %r2, 0;
setp.ge.u32 %p122, %r4, %r42;
or.pred %p123, %p121, %p122;
@%p123 bra $L__BB88_56;

add.ftz.f32 %f387, %f5, %f48;
add.ftz.f32 %f388, %f4, %f49;
mov.f32 %f389, 0f40E00000;
min.ftz.f32 %f390, %f388, %f389;
min.ftz.f32 %f391, %f387, %f389;
mov.f32 %f392, 0fC0E00000;
max.ftz.f32 %f393, %f391, %f392;
mul.ftz.f32 %f394, %f390, 0f3FD9DB23;
mul.ftz.f32 %f395, %f394, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f396, %f395;
add.ftz.f32 %f397, %f396, 0f3F800000;
div.approx.ftz.f32 %f398, %f390, %f397;
add.ftz.f32 %f399, %f393, 0f3F800000;
mul.ftz.f32 %f400, %f399, %f398;
st.global.f32 [%rd24], %f400;

$L__BB88_56:
ld.shared.f32 %f401, [%r23+128];
add.ftz.f32 %f402, %f401, %f29;
ld.shared.f32 %f403, [%r25+128];
add.ftz.f32 %f404, %f403, %f33;
ld.shared.f32 %f405, [%r23+640];
add.ftz.f32 %f406, %f405, %f402;
ld.shared.f32 %f407, [%r25+640];
add.ftz.f32 %f408, %f407, %f404;
ld.shared.f32 %f409, [%r23+1152];
add.ftz.f32 %f410, %f409, %f406;
ld.shared.f32 %f411, [%r25+1152];
add.ftz.f32 %f412, %f411, %f408;
mov.b32 %r446, %f410;
shfl.sync.bfly.b32 %r450|%p124, %r446, %r418, %r417, %r419;
mov.b32 %f413, %r450;
add.ftz.f32 %f414, %f410, %f413;
mov.b32 %r451, %f414;
shfl.sync.bfly.b32 %r453|%p125, %r451, %r422, %r417, %r419;
mov.b32 %f415, %r453;
add.ftz.f32 %f416, %f414, %f415;
mov.b32 %r454, %f416;
shfl.sync.bfly.b32 %r456|%p126, %r454, %r425, %r417, %r419;
mov.b32 %f417, %r456;
add.ftz.f32 %f418, %f416, %f417;
mov.b32 %r457, %f418;
shfl.sync.bfly.b32 %r459|%p127, %r457, %r428, %r417, %r419;
mov.b32 %f419, %r459;
add.ftz.f32 %f420, %f418, %f419;
mov.b32 %r460, %f420;
shfl.sync.bfly.b32 %r462|%p128, %r460, %r431, %r417, %r419;
mov.b32 %f421, %r462;
add.ftz.f32 %f50, %f420, %f421;
mov.b32 %r463, %f412;
shfl.sync.bfly.b32 %r464|%p129, %r463, %r418, %r417, %r419;
mov.b32 %f422, %r464;
add.ftz.f32 %f423, %f412, %f422;
mov.b32 %r465, %f423;
shfl.sync.bfly.b32 %r466|%p130, %r465, %r422, %r417, %r419;
mov.b32 %f424, %r466;
add.ftz.f32 %f425, %f423, %f424;
mov.b32 %r467, %f425;
shfl.sync.bfly.b32 %r468|%p131, %r467, %r425, %r417, %r419;
mov.b32 %f426, %r468;
add.ftz.f32 %f427, %f425, %f426;
mov.b32 %r469, %f427;
shfl.sync.bfly.b32 %r470|%p132, %r469, %r428, %r417, %r419;
mov.b32 %f428, %r470;
add.ftz.f32 %f429, %f427, %f428;
mov.b32 %r471, %f429;
shfl.sync.bfly.b32 %r472|%p133, %r471, %r431, %r417, %r419;
mov.b32 %f430, %r472;
add.ftz.f32 %f51, %f429, %f430;
setp.ne.s32 %p134, %r2, 1;
or.b32 %r476, %r4, 1;
setp.ge.u32 %p135, %r476, %r42;
or.pred %p136, %p134, %p135;
@%p136 bra $L__BB88_58;

add.ftz.f32 %f431, %f5, %f50;
add.ftz.f32 %f432, %f4, %f51;
mov.f32 %f433, 0f40E00000;
min.ftz.f32 %f434, %f432, %f433;
min.ftz.f32 %f435, %f431, %f433;
mov.f32 %f436, 0fC0E00000;
max.ftz.f32 %f437, %f435, %f436;
mul.ftz.f32 %f438, %f434, 0f3FD9DB23;
mul.ftz.f32 %f439, %f438, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f440, %f439;
add.ftz.f32 %f441, %f440, 0f3F800000;
div.approx.ftz.f32 %f442, %f434, %f441;
add.ftz.f32 %f443, %f437, 0f3F800000;
mul.ftz.f32 %f444, %f443, %f442;
st.global.f32 [%rd24+4], %f444;

$L__BB88_58:
ld.shared.f32 %f445, [%r23+256];
add.ftz.f32 %f446, %f445, %f28;
ld.shared.f32 %f447, [%r25+256];
add.ftz.f32 %f448, %f447, %f32;
ld.shared.f32 %f449, [%r23+768];
add.ftz.f32 %f450, %f449, %f446;
ld.shared.f32 %f451, [%r25+768];
add.ftz.f32 %f452, %f451, %f448;
ld.shared.f32 %f453, [%r23+1280];
add.ftz.f32 %f454, %f453, %f450;
ld.shared.f32 %f455, [%r25+1280];
add.ftz.f32 %f456, %f455, %f452;
mov.b32 %r477, %f454;
mov.u32 %r478, 31;
mov.u32 %r479, 16;
mov.u32 %r480, -1;
shfl.sync.bfly.b32 %r481|%p137, %r477, %r479, %r478, %r480;
mov.b32 %f457, %r481;
add.ftz.f32 %f458, %f454, %f457;
mov.b32 %r482, %f458;
mov.u32 %r483, 8;
shfl.sync.bfly.b32 %r484|%p138, %r482, %r483, %r478, %r480;
mov.b32 %f459, %r484;
add.ftz.f32 %f460, %f458, %f459;
mov.b32 %r485, %f460;
mov.u32 %r486, 4;
shfl.sync.bfly.b32 %r487|%p139, %r485, %r486, %r478, %r480;
mov.b32 %f461, %r487;
add.ftz.f32 %f462, %f460, %f461;
mov.b32 %r488, %f462;
mov.u32 %r489, 2;
shfl.sync.bfly.b32 %r490|%p140, %r488, %r489, %r478, %r480;
mov.b32 %f463, %r490;
add.ftz.f32 %f464, %f462, %f463;
mov.b32 %r491, %f464;
mov.u32 %r492, 1;
shfl.sync.bfly.b32 %r493|%p141, %r491, %r492, %r478, %r480;
mov.b32 %f465, %r493;
add.ftz.f32 %f52, %f464, %f465;
mov.b32 %r494, %f456;
shfl.sync.bfly.b32 %r495|%p142, %r494, %r479, %r478, %r480;
mov.b32 %f466, %r495;
add.ftz.f32 %f467, %f456, %f466;
mov.b32 %r496, %f467;
shfl.sync.bfly.b32 %r497|%p143, %r496, %r483, %r478, %r480;
mov.b32 %f468, %r497;
add.ftz.f32 %f469, %f467, %f468;
mov.b32 %r498, %f469;
shfl.sync.bfly.b32 %r499|%p144, %r498, %r486, %r478, %r480;
mov.b32 %f470, %r499;
add.ftz.f32 %f471, %f469, %f470;
mov.b32 %r500, %f471;
shfl.sync.bfly.b32 %r501|%p145, %r500, %r489, %r478, %r480;
mov.b32 %f472, %r501;
add.ftz.f32 %f473, %f471, %f472;
mov.b32 %r502, %f473;
shfl.sync.bfly.b32 %r503|%p146, %r502, %r492, %r478, %r480;
mov.b32 %f474, %r503;
add.ftz.f32 %f53, %f473, %f474;
setp.ne.s32 %p147, %r2, 2;
or.b32 %r507, %r4, 2;
setp.ge.u32 %p148, %r507, %r42;
or.pred %p149, %p147, %p148;
@%p149 bra $L__BB88_60;

add.ftz.f32 %f475, %f5, %f52;
add.ftz.f32 %f476, %f4, %f53;
mov.f32 %f477, 0f40E00000;
min.ftz.f32 %f478, %f476, %f477;
min.ftz.f32 %f479, %f475, %f477;
mov.f32 %f480, 0fC0E00000;
max.ftz.f32 %f481, %f479, %f480;
mul.ftz.f32 %f482, %f478, 0f3FD9DB23;
mul.ftz.f32 %f483, %f482, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f484, %f483;
add.ftz.f32 %f485, %f484, 0f3F800000;
div.approx.ftz.f32 %f486, %f478, %f485;
add.ftz.f32 %f487, %f481, 0f3F800000;
mul.ftz.f32 %f488, %f487, %f486;
st.global.f32 [%rd24+8], %f488;

$L__BB88_60:
ld.shared.f32 %f489, [%r23+384];
add.ftz.f32 %f490, %f489, %f27;
ld.shared.f32 %f491, [%r25+384];
add.ftz.f32 %f492, %f491, %f31;
ld.shared.f32 %f493, [%r23+896];
add.ftz.f32 %f494, %f493, %f490;
ld.shared.f32 %f495, [%r25+896];
add.ftz.f32 %f496, %f495, %f492;
ld.shared.f32 %f497, [%r23+1408];
add.ftz.f32 %f498, %f497, %f494;
ld.shared.f32 %f499, [%r25+1408];
add.ftz.f32 %f500, %f499, %f496;
mov.b32 %r508, %f498;
shfl.sync.bfly.b32 %r512|%p150, %r508, %r479, %r478, %r480;
mov.b32 %f501, %r512;
add.ftz.f32 %f502, %f498, %f501;
mov.b32 %r513, %f502;
shfl.sync.bfly.b32 %r515|%p151, %r513, %r483, %r478, %r480;
mov.b32 %f503, %r515;
add.ftz.f32 %f504, %f502, %f503;
mov.b32 %r516, %f504;
shfl.sync.bfly.b32 %r518|%p152, %r516, %r486, %r478, %r480;
mov.b32 %f505, %r518;
add.ftz.f32 %f506, %f504, %f505;
mov.b32 %r519, %f506;
shfl.sync.bfly.b32 %r521|%p153, %r519, %r489, %r478, %r480;
mov.b32 %f507, %r521;
add.ftz.f32 %f508, %f506, %f507;
mov.b32 %r522, %f508;
shfl.sync.bfly.b32 %r524|%p154, %r522, %r492, %r478, %r480;
mov.b32 %f509, %r524;
add.ftz.f32 %f54, %f508, %f509;
mov.b32 %r525, %f500;
shfl.sync.bfly.b32 %r526|%p155, %r525, %r479, %r478, %r480;
mov.b32 %f510, %r526;
add.ftz.f32 %f511, %f500, %f510;
mov.b32 %r527, %f511;
shfl.sync.bfly.b32 %r528|%p156, %r527, %r483, %r478, %r480;
mov.b32 %f512, %r528;
add.ftz.f32 %f513, %f511, %f512;
mov.b32 %r529, %f513;
shfl.sync.bfly.b32 %r530|%p157, %r529, %r486, %r478, %r480;
mov.b32 %f514, %r530;
add.ftz.f32 %f515, %f513, %f514;
mov.b32 %r531, %f515;
shfl.sync.bfly.b32 %r532|%p158, %r531, %r489, %r478, %r480;
mov.b32 %f516, %r532;
add.ftz.f32 %f517, %f515, %f516;
mov.b32 %r533, %f517;
shfl.sync.bfly.b32 %r534|%p159, %r533, %r492, %r478, %r480;
mov.b32 %f518, %r534;
add.ftz.f32 %f55, %f517, %f518;
setp.ne.s32 %p160, %r2, 3;
or.b32 %r538, %r4, 3;
setp.ge.u32 %p161, %r538, %r42;
or.pred %p162, %p160, %p161;
@%p162 bra $L__BB88_78;

add.ftz.f32 %f519, %f5, %f54;
add.ftz.f32 %f520, %f4, %f55;
mov.f32 %f521, 0f40E00000;
min.ftz.f32 %f522, %f520, %f521;
min.ftz.f32 %f523, %f519, %f521;
mov.f32 %f524, 0fC0E00000;
max.ftz.f32 %f525, %f523, %f524;
mul.ftz.f32 %f526, %f522, 0f3FD9DB23;
mul.ftz.f32 %f527, %f526, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f528, %f527;
add.ftz.f32 %f529, %f528, 0f3F800000;
div.approx.ftz.f32 %f530, %f522, %f529;
add.ftz.f32 %f531, %f525, 0f3F800000;
mul.ftz.f32 %f532, %f531, %f530;
st.global.f32 [%rd24+12], %f532;
bra.uni $L__BB88_78;

}
.entry _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<38>;
.reg .b16 %rs<29>;
.reg .f32 %f<109>;
.reg .b32 %r<185>;
.reg .b64 %rd<44>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

ld.param.u64 %rd12, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd13, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd14, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r17, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r19, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r18, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
mov.u32 %r1, %tid.y;
shl.b32 %r35, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r35, %r2;
mov.u32 %r36, %ctaid.x;
shl.b32 %r4, %r36, 2;
mov.u32 %r183, %ctaid.y;
setp.eq.s64 %p1, %rd14, 0;
@%p1 bra $L__BB89_2;

cvta.to.global.u64 %rd16, %rd14;
mul.wide.u32 %rd17, %r183, 4;
add.s64 %rd18, %rd16, %rd17;
ld.global.nc.u32 %r182, [%rd18];
mul.hi.u32 %r37, %r183, %r18;
add.s32 %r38, %r37, %r183;
shr.u32 %r39, %r38, %r19;
mul.lo.s32 %r40, %r39, %r20;
sub.s32 %r183, %r183, %r40;
bra.uni $L__BB89_3;

$L__BB89_2:
mul.hi.u32 %r41, %r183, %r23;
add.s32 %r42, %r41, %r183;
shr.u32 %r182, %r42, %r24;

$L__BB89_3:
mov.u32 %r11, %ctaid.z;
shr.s32 %r43, %r3, 31;
shr.u32 %r44, %r43, 30;
add.s32 %r45, %r3, %r44;
shr.s32 %r184, %r45, 2;
shr.u32 %r46, %r17, 5;
setp.ge.s32 %p2, %r184, %r46;
mov.f32 %f105, 0f00000000;
mov.f32 %f106, %f105;
mov.f32 %f107, %f105;
mov.f32 %f108, %f105;
@%p2 bra $L__BB89_6;

and.b32 %r50, %r45, -4;
sub.s32 %r51, %r3, %r50;
shl.b32 %r52, %r51, 1;
shl.b32 %r53, %r51, 2;
or.b32 %r54, %r53, 1;
cvt.s64.s32 %rd1, %r52;
sub.s32 %r55, %r54, %r52;
cvt.s64.s32 %rd2, %r55;
mad.lo.s32 %r56, %r4, %r21, %r184;
mul.hi.u32 %r57, %r11, %r29;
add.s32 %r58, %r57, %r11;
shr.u32 %r59, %r58, %r30;
mad.lo.s32 %r60, %r59, %r32, %r56;
mad.lo.s32 %r61, %r182, %r26, %r60;
mul.wide.s32 %rd3, %r61, 34;
mul.wide.s32 %rd19, %r53, 2;
add.s64 %rd20, %rd3, %rd19;
add.s64 %rd4, %rd20, 4;
mul.wide.s32 %rd5, %r21, 34;
cvta.to.global.u64 %rd6, %rd13;
cvta.to.global.u64 %rd43, %rd12;
mul.lo.s32 %r62, %r183, %r27;
cvt.u64.u32 %rd21, %r62;
mul.lo.s32 %r63, %r11, %r33;
cvt.u64.u32 %rd22, %r63;
add.s64 %rd8, %rd21, %rd22;
mov.f32 %f105, 0f00000000;
shl.b64 %rd30, %rd1, 2;
shl.b64 %rd32, %rd2, 2;

$L__BB89_5:
cvt.u64.u32 %rd23, %r184;
add.s64 %rd24, %rd8, %rd23;
mul.lo.s64 %rd25, %rd24, 36;
add.s64 %rd26, %rd6, %rd25;
add.s64 %rd27, %rd26, 4;
add.s64 %rd28, %rd43, %rd4;
add.s64 %rd29, %rd28, -2;
ld.global.nc.u16 %rs13, [%rd28+-2];
ld.global.nc.u16 %rs14, [%rd28];
mov.b32 %r66, {%rs13, %rs14};
add.s64 %rd31, %rd27, %rd30;
ld.global.nc.u32 %r67, [%rd31];
ld.global.nc.u16 %rs15, [%rd28+2];
ld.global.nc.u16 %rs16, [%rd28+4];
mov.b32 %r70, {%rs15, %rs16};
add.s64 %rd33, %rd27, %rd32;
ld.global.nc.u32 %r71, [%rd33];
add.s64 %rd34, %rd43, %rd3;
ld.global.nc.u16 %rs1, [%rd34];

{ cvt.f32.f16 %f25, %rs1;}


ld.global.nc.u32 %r64, [%rd26];

{.reg .f16 low,high;
mov.b32 {low,high}, %r64;
mov.b16 %rs2, low;}


{ cvt.f32.f16 %f26, %rs2;}


mov.u32 %r95, 0;

dp4a.s32.s32 %r65, %r66, %r67, %r95;


dp4a.s32.s32 %r69, %r70, %r71, %r65;

mul.ftz.f32 %f33, %f25, %f26;
cvt.rn.f32.s32 %f34, %r69;
fma.rn.ftz.f32 %f108, %f33, %f34, %f108;
add.s64 %rd35, %rd29, %rd5;
ld.global.nc.u16 %rs17, [%rd35];
ld.global.nc.u16 %rs18, [%rd35+2];
mov.b32 %r75, {%rs17, %rs18};
ld.global.nc.u16 %rs19, [%rd35+4];
ld.global.nc.u16 %rs20, [%rd35+6];
mov.b32 %r79, {%rs19, %rs20};
add.s64 %rd36, %rd34, %rd5;
ld.global.nc.u16 %rs4, [%rd36];

{ cvt.f32.f16 %f27, %rs4;}



{.reg .f16 low,high;
mov.b32 {low,high}, %r64;
mov.b16 %rs5, low;}


{ cvt.f32.f16 %f28, %rs5;}



dp4a.s32.s32 %r74, %r75, %r67, %r95;


dp4a.s32.s32 %r78, %r79, %r71, %r74;

mul.ftz.f32 %f35, %f27, %f28;
cvt.rn.f32.s32 %f36, %r78;
fma.rn.ftz.f32 %f107, %f35, %f36, %f107;
add.s64 %rd37, %rd35, %rd5;
ld.global.nc.u16 %rs21, [%rd37];
ld.global.nc.u16 %rs22, [%rd37+2];
mov.b32 %r84, {%rs21, %rs22};
ld.global.nc.u16 %rs23, [%rd37+4];
ld.global.nc.u16 %rs24, [%rd37+6];
mov.b32 %r88, {%rs23, %rs24};
add.s64 %rd38, %rd36, %rd5;
ld.global.nc.u16 %rs7, [%rd38];

{ cvt.f32.f16 %f29, %rs7;}



{.reg .f16 low,high;
mov.b32 {low,high}, %r64;
mov.b16 %rs8, low;}


{ cvt.f32.f16 %f30, %rs8;}



dp4a.s32.s32 %r83, %r84, %r67, %r95;


dp4a.s32.s32 %r87, %r88, %r71, %r83;

mul.ftz.f32 %f37, %f29, %f30;
cvt.rn.f32.s32 %f38, %r87;
fma.rn.ftz.f32 %f106, %f37, %f38, %f106;
add.s64 %rd39, %rd37, %rd5;
ld.global.nc.u16 %rs25, [%rd39];
ld.global.nc.u16 %rs26, [%rd39+2];
mov.b32 %r93, {%rs25, %rs26};
ld.global.nc.u16 %rs27, [%rd39+4];
ld.global.nc.u16 %rs28, [%rd39+6];
mov.b32 %r97, {%rs27, %rs28};
add.s64 %rd40, %rd38, %rd5;
ld.global.nc.u16 %rs10, [%rd40];

{ cvt.f32.f16 %f31, %rs10;}



{.reg .f16 low,high;
mov.b32 {low,high}, %r64;
mov.b16 %rs11, low;}


{ cvt.f32.f16 %f32, %rs11;}



dp4a.s32.s32 %r92, %r93, %r67, %r95;


dp4a.s32.s32 %r96, %r97, %r71, %r92;

mul.ftz.f32 %f39, %f31, %f32;
cvt.rn.f32.s32 %f40, %r96;
fma.rn.ftz.f32 %f105, %f39, %f40, %f105;
add.s64 %rd43, %rd43, 1088;
add.s32 %r184, %r184, 32;
setp.lt.s32 %p3, %r184, %r46;
@%p3 bra $L__BB89_5;

$L__BB89_6:
setp.eq.s32 %p4, %r1, 0;
@%p4 bra $L__BB89_8;

shl.b32 %r101, %r1, 9;
mov.u32 %r102, _ZZ13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r103, %r102, %r101;
shl.b32 %r104, %r2, 2;
add.s32 %r105, %r103, %r104;
st.shared.f32 [%r105+-512], %f108;
st.shared.f32 [%r105+-384], %f107;
st.shared.f32 [%r105+-256], %f106;
st.shared.f32 [%r105+-128], %f105;

$L__BB89_8:
setp.ne.s32 %p5, %r1, 0;
bar.sync 0;
@%p5 bra $L__BB89_17;

mov.u32 %r106, %ctaid.y;
mad.lo.s32 %r107, %r106, %r28, %r4;
mad.lo.s32 %r108, %r11, %r34, %r107;
shl.b32 %r109, %r2, 2;
mov.u32 %r110, _ZZ13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r15, %r110, %r109;
mov.u32 %r111, 2;
ld.shared.f32 %f41, [%r15];
add.ftz.f32 %f42, %f41, %f108;
ld.shared.f32 %f43, [%r15+512];
add.ftz.f32 %f44, %f43, %f42;
ld.shared.f32 %f45, [%r15+1024];
add.ftz.f32 %f46, %f45, %f44;
mov.b32 %r112, %f46;
mov.u32 %r113, 31;
mov.u32 %r114, 16;
mov.u32 %r115, -1;
shfl.sync.bfly.b32 %r116|%p6, %r112, %r114, %r113, %r115;
mov.b32 %f47, %r116;
add.ftz.f32 %f48, %f46, %f47;
mov.b32 %r117, %f48;
mov.u32 %r118, 8;
shfl.sync.bfly.b32 %r119|%p7, %r117, %r118, %r113, %r115;
mov.b32 %f49, %r119;
add.ftz.f32 %f50, %f48, %f49;
mov.b32 %r120, %f50;
mov.u32 %r121, 4;
shfl.sync.bfly.b32 %r122|%p8, %r120, %r121, %r113, %r115;
mov.b32 %f51, %r122;
add.ftz.f32 %f52, %f50, %f51;
mov.b32 %r123, %f52;
shfl.sync.bfly.b32 %r124|%p9, %r123, %r111, %r113, %r115;
mov.b32 %f53, %r124;
add.ftz.f32 %f54, %f52, %f53;
mov.b32 %r125, %f54;
mov.u32 %r126, 1;
shfl.sync.bfly.b32 %r127|%p10, %r125, %r126, %r113, %r115;
mov.b32 %f55, %r127;
add.ftz.f32 %f13, %f54, %f55;
setp.ge.u32 %p11, %r4, %r22;
setp.ne.s32 %p12, %r2, 0;
cvta.to.global.u64 %rd41, %rd15;
mul.wide.u32 %rd42, %r108, 4;
add.s64 %rd11, %rd41, %rd42;
or.pred %p13, %p12, %p11;
@%p13 bra $L__BB89_11;

st.global.f32 [%rd11], %f13;

$L__BB89_11:
ld.shared.f32 %f56, [%r15+128];
add.ftz.f32 %f57, %f56, %f107;
ld.shared.f32 %f58, [%r15+640];
add.ftz.f32 %f59, %f58, %f57;
ld.shared.f32 %f60, [%r15+1152];
add.ftz.f32 %f61, %f60, %f59;
mov.b32 %r128, %f61;
shfl.sync.bfly.b32 %r132|%p14, %r128, %r114, %r113, %r115;
mov.b32 %f62, %r132;
add.ftz.f32 %f63, %f61, %f62;
mov.b32 %r133, %f63;
shfl.sync.bfly.b32 %r135|%p15, %r133, %r118, %r113, %r115;
mov.b32 %f64, %r135;
add.ftz.f32 %f65, %f63, %f64;
mov.b32 %r136, %f65;
shfl.sync.bfly.b32 %r138|%p16, %r136, %r121, %r113, %r115;
mov.b32 %f66, %r138;
add.ftz.f32 %f67, %f65, %f66;
mov.b32 %r139, %f67;
shfl.sync.bfly.b32 %r141|%p17, %r139, %r111, %r113, %r115;
mov.b32 %f68, %r141;
add.ftz.f32 %f69, %f67, %f68;
mov.b32 %r142, %f69;
shfl.sync.bfly.b32 %r144|%p18, %r142, %r126, %r113, %r115;
mov.b32 %f70, %r144;
add.ftz.f32 %f14, %f69, %f70;
add.s32 %r145, %r4, 1;
setp.ge.u32 %p19, %r145, %r22;
setp.ne.s32 %p20, %r2, 1;
or.pred %p21, %p20, %p19;
@%p21 bra $L__BB89_13;

st.global.f32 [%rd11+4], %f14;

$L__BB89_13:
ld.shared.f32 %f71, [%r15+256];
add.ftz.f32 %f72, %f71, %f106;
ld.shared.f32 %f73, [%r15+768];
add.ftz.f32 %f74, %f73, %f72;
ld.shared.f32 %f75, [%r15+1280];
add.ftz.f32 %f76, %f75, %f74;
mov.b32 %r146, %f76;
mov.u32 %r147, 31;
mov.u32 %r148, 16;
mov.u32 %r149, -1;
shfl.sync.bfly.b32 %r150|%p22, %r146, %r148, %r147, %r149;
mov.b32 %f77, %r150;
add.ftz.f32 %f78, %f76, %f77;
mov.b32 %r151, %f78;
mov.u32 %r152, 8;
shfl.sync.bfly.b32 %r153|%p23, %r151, %r152, %r147, %r149;
mov.b32 %f79, %r153;
add.ftz.f32 %f80, %f78, %f79;
mov.b32 %r154, %f80;
mov.u32 %r155, 4;
shfl.sync.bfly.b32 %r156|%p24, %r154, %r155, %r147, %r149;
mov.b32 %f81, %r156;
add.ftz.f32 %f82, %f80, %f81;
mov.b32 %r157, %f82;
mov.u32 %r158, 2;
shfl.sync.bfly.b32 %r159|%p25, %r157, %r158, %r147, %r149;
mov.b32 %f83, %r159;
add.ftz.f32 %f84, %f82, %f83;
mov.b32 %r160, %f84;
mov.u32 %r161, 1;
shfl.sync.bfly.b32 %r162|%p26, %r160, %r161, %r147, %r149;
mov.b32 %f85, %r162;
add.ftz.f32 %f15, %f84, %f85;
add.s32 %r163, %r4, 2;
setp.ge.u32 %p27, %r163, %r22;
setp.ne.s32 %p28, %r2, 2;
or.pred %p29, %p28, %p27;
@%p29 bra $L__BB89_15;

st.global.f32 [%rd11+8], %f15;

$L__BB89_15:
ld.shared.f32 %f86, [%r15+384];
add.ftz.f32 %f87, %f86, %f105;
ld.shared.f32 %f88, [%r15+896];
add.ftz.f32 %f89, %f88, %f87;
ld.shared.f32 %f90, [%r15+1408];
add.ftz.f32 %f91, %f90, %f89;
mov.b32 %r164, %f91;
shfl.sync.bfly.b32 %r168|%p30, %r164, %r148, %r147, %r149;
mov.b32 %f92, %r168;
add.ftz.f32 %f93, %f91, %f92;
mov.b32 %r169, %f93;
shfl.sync.bfly.b32 %r171|%p31, %r169, %r152, %r147, %r149;
mov.b32 %f94, %r171;
add.ftz.f32 %f95, %f93, %f94;
mov.b32 %r172, %f95;
shfl.sync.bfly.b32 %r174|%p32, %r172, %r155, %r147, %r149;
mov.b32 %f96, %r174;
add.ftz.f32 %f97, %f95, %f96;
mov.b32 %r175, %f97;
shfl.sync.bfly.b32 %r177|%p33, %r175, %r158, %r147, %r149;
mov.b32 %f98, %r177;
add.ftz.f32 %f99, %f97, %f98;
mov.b32 %r178, %f99;
shfl.sync.bfly.b32 %r180|%p34, %r178, %r161, %r147, %r149;
mov.b32 %f100, %r180;
add.ftz.f32 %f16, %f99, %f100;
add.s32 %r181, %r4, 3;
setp.ge.u32 %p35, %r181, %r22;
setp.ne.s32 %p36, %r2, 3;
or.pred %p37, %p36, %p35;
@%p37 bra $L__BB89_17;

st.global.f32 [%rd11+12], %f16;

$L__BB89_17:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<38>;
.reg .b16 %rs<43>;
.reg .f32 %f<154>;
.reg .b32 %r<226>;
.reg .b64 %rd<104>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[384];

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate[384];

ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd20, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd21, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd27, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r47, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r48, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r49, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r45, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
ld.param.f32 %f41, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
ld.param.u64 %rd24, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+16];
ld.param.u64 %rd23, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+8];
ld.param.u64 %rd22, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3];
mov.u32 %r1, %tid.y;
shl.b32 %r50, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r50, %r2;
mov.u32 %r4, %ctaid.y;
setp.eq.s64 %p2, %rd21, 0;
@%p2 bra $L__BB90_2;

cvta.to.global.u64 %rd28, %rd21;
mul.wide.u32 %rd29, %r4, 4;
add.s64 %rd30, %rd28, %rd29;
ld.global.nc.u32 %r222, [%rd30];
mul.hi.u32 %r51, %r4, %r34;
add.s32 %r52, %r51, %r4;
shr.u32 %r53, %r52, %r35;
mul.lo.s32 %r54, %r53, %r36;
sub.s32 %r223, %r4, %r54;
bra.uni $L__BB90_3;

$L__BB90_2:
mul.hi.u32 %r55, %r4, %r38;
add.s32 %r56, %r55, %r4;
shr.u32 %r222, %r56, %r39;
mov.u32 %r223, %r4;

$L__BB90_3:
or.b32 %r57, %r2, %r1;
setp.ne.s32 %p3, %r57, 0;
mov.f32 %f146, 0f00000000;
mov.f32 %f137, %f146;
mov.f32 %f138, %f146;
@%p3 bra $L__BB90_8;

setp.eq.s64 %p4, %rd22, 0;
mov.f32 %f137, 0f00000000;
mov.f32 %f138, %f137;
@%p4 bra $L__BB90_6;

cvta.to.global.u64 %rd31, %rd22;
mov.u32 %r58, %ctaid.z;
mul.lo.s32 %r59, %r58, %r49;
cvt.u64.u32 %rd32, %r59;
selp.b32 %r60, %r4, %r222, %p2;
mul.lo.s32 %r61, %r60, %r43;
cvt.u64.u32 %rd33, %r61;
mov.u32 %r62, %ctaid.x;
cvt.s64.s32 %rd34, %r62;
cvt.u64.u32 %rd35, %r2;
add.s64 %rd36, %rd35, %rd34;
add.s64 %rd37, %rd36, %rd32;
add.s64 %rd38, %rd37, %rd33;
shl.b64 %rd39, %rd38, 2;
add.s64 %rd40, %rd31, %rd39;
ld.global.f32 %f138, [%rd40];

$L__BB90_6:
setp.eq.s64 %p6, %rd24, 0;
setp.eq.s64 %p7, %rd23, 0;
or.pred %p8, %p6, %p7;
@%p8 bra $L__BB90_8;

cvta.to.global.u64 %rd41, %rd24;
mov.u32 %r64, %ctaid.z;
mul.lo.s32 %r65, %r64, %r49;
cvt.u64.u32 %rd42, %r65;
selp.b32 %r66, %r4, %r222, %p2;
mul.lo.s32 %r67, %r66, %r43;
cvt.u64.u32 %rd43, %r67;
mov.u32 %r68, %ctaid.x;
cvt.s64.s32 %rd44, %r68;
cvt.u64.u32 %rd45, %r2;
add.s64 %rd46, %rd45, %rd44;
add.s64 %rd47, %rd46, %rd42;
add.s64 %rd48, %rd47, %rd43;
shl.b64 %rd49, %rd48, 2;
add.s64 %rd50, %rd41, %rd49;
ld.global.f32 %f137, [%rd50];

$L__BB90_8:
mov.u32 %r70, %ctaid.z;
mul.lo.s32 %r71, %r70, %r48;
cvt.u64.u32 %rd51, %r71;
mul.lo.s32 %r72, %r223, %r42;
cvt.u64.u32 %rd52, %r72;
add.s64 %rd4, %rd52, %rd51;
mul.hi.u32 %r73, %r70, %r44;
add.s32 %r74, %r73, %r70;
shr.u32 %r75, %r74, %r45;
mul.lo.s32 %r10, %r75, %r47;
mul.lo.s32 %r11, %r222, %r41;
shr.u32 %r76, %r33, 5;
shr.s32 %r77, %r3, 31;
shr.u32 %r78, %r77, 30;
add.s32 %r79, %r3, %r78;
shr.s32 %r225, %r79, 2;
setp.ge.s32 %p10, %r225, %r76;
mov.f32 %f149, %f146;
@%p10 bra $L__BB90_20;

and.b32 %r83, %r79, -4;
sub.s32 %r13, %r3, %r83;
not.b32 %r85, %r225;
add.s32 %r14, %r76, %r85;
and.b32 %r86, %r14, 32;
setp.ne.s32 %p11, %r86, 0;
mov.f32 %f149, 0f00000000;
mov.f32 %f146, %f149;
@%p11 bra $L__BB90_13;

cvta.to.global.u64 %rd53, %rd19;
cvta.to.global.u64 %rd54, %rd20;
cvt.u64.u32 %rd55, %r225;
add.s64 %rd56, %rd4, %rd55;
mov.u32 %r96, %ctaid.x;
mad.lo.s32 %r97, %r96, %r37, %r11;
add.s32 %r98, %r97, %r10;
add.s32 %r99, %r98, %r225;
cvt.s64.s32 %rd5, %r99;
mul.wide.s32 %rd57, %r99, 34;
add.s64 %rd58, %rd53, %rd57;
shl.b32 %r100, %r13, 2;
mul.wide.s32 %rd59, %r100, 2;
add.s64 %rd60, %rd58, %rd59;
mul.lo.s64 %rd61, %rd56, 36;
add.s64 %rd62, %rd54, %rd61;
add.s64 %rd63, %rd62, 4;
ld.global.nc.u16 %rs4, [%rd60+2];
or.b32 %r101, %r100, 1;
ld.global.nc.u16 %rs5, [%rd60+4];
mov.b32 %r89, {%rs4, %rs5};
shl.b32 %r102, %r13, 1;
mul.wide.s32 %rd64, %r102, 4;
add.s64 %rd65, %rd63, %rd64;
ld.global.nc.u32 %r15, [%rd65];
sub.s32 %r103, %r101, %r102;
ld.global.nc.u16 %rs6, [%rd60+6];
ld.global.nc.u16 %rs7, [%rd60+8];
mov.b32 %r93, {%rs6, %rs7};
mul.wide.s32 %rd66, %r103, 4;
add.s64 %rd67, %rd63, %rd66;
ld.global.nc.u32 %r16, [%rd67];
ld.global.nc.u16 %rs1, [%rd58];

{ cvt.f32.f16 %f51, %rs1;}


ld.global.nc.u32 %r17, [%rd62];

{.reg .f16 low,high;
mov.b32 {low,high}, %r17;
mov.b16 %rs2, low;}


{ cvt.f32.f16 %f52, %rs2;}


mov.u32 %r91, 0;

dp4a.s32.s32 %r88, %r89, %r15, %r91;


dp4a.s32.s32 %r92, %r93, %r16, %r88;

mul.ftz.f32 %f54, %f51, %f52;
cvt.rn.f32.s32 %f55, %r92;
fma.rn.ftz.f32 %f149, %f54, %f55, 0f00000000;
mov.f32 %f146, 0f00000000;
setp.eq.s64 %p12, %rd23, 0;
@%p12 bra $L__BB90_12;

cvta.to.global.u64 %rd68, %rd23;
mul.lo.s64 %rd69, %rd5, 34;
add.s64 %rd70, %rd68, %rd69;
add.s64 %rd72, %rd70, %rd59;
ld.global.u16 %rs11, [%rd72+2];
ld.global.u16 %rs12, [%rd72+4];
mov.b32 %r106, {%rs11, %rs12};
ld.global.u16 %rs13, [%rd72+6];
ld.global.u16 %rs14, [%rd72+8];
mov.b32 %r110, {%rs13, %rs14};
ld.global.u16 %rs8, [%rd70];

{ cvt.f32.f16 %f56, %rs8;}



{.reg .f16 low,high;
mov.b32 {low,high}, %r17;
mov.b16 %rs9, low;}


{ cvt.f32.f16 %f57, %rs9;}



dp4a.s32.s32 %r105, %r106, %r15, %r91;


dp4a.s32.s32 %r109, %r110, %r16, %r105;

mul.ftz.f32 %f58, %f56, %f57;
cvt.rn.f32.s32 %f59, %r109;
fma.rn.ftz.f32 %f146, %f58, %f59, 0f00000000;

$L__BB90_12:
add.s32 %r225, %r225, 32;

$L__BB90_13:
and.b32 %r114, %r14, -32;
setp.eq.s32 %p13, %r114, 0;
@%p13 bra $L__BB90_20;

shl.b32 %r115, %r13, 2;
or.b32 %r116, %r115, 1;
shl.b32 %r117, %r13, 1;
cvt.s64.s32 %rd6, %r117;
sub.s32 %r118, %r116, %r117;
cvt.s64.s32 %rd7, %r118;
add.s32 %r119, %r225, %r10;
add.s32 %r120, %r119, %r11;
mov.u32 %r121, %ctaid.x;
mad.lo.s32 %r20, %r121, %r37, %r120;
mul.wide.s32 %rd73, %r20, 34;
mul.wide.s32 %rd74, %r115, 2;
add.s64 %rd75, %rd73, %rd74;
add.s64 %rd8, %rd75, 2;
cvta.to.global.u64 %rd102, %rd19;
cvta.to.global.u64 %rd103, %rd23;
cvta.to.global.u64 %rd11, %rd20;

$L__BB90_15:
cvt.u64.u32 %rd76, %r225;
add.s64 %rd77, %rd4, %rd76;
mul.lo.s64 %rd78, %rd77, 36;
add.s64 %rd79, %rd11, %rd78;
add.s64 %rd80, %rd79, 4;
add.s64 %rd14, %rd102, %rd8;
ld.global.nc.u16 %rs18, [%rd14];
ld.global.nc.u16 %rs19, [%rd14+2];
mov.b32 %r124, {%rs18, %rs19};
shl.b64 %rd81, %rd6, 2;
add.s64 %rd82, %rd80, %rd81;
ld.global.nc.u32 %r22, [%rd82];
ld.global.nc.u16 %rs20, [%rd14+4];
ld.global.nc.u16 %rs21, [%rd14+6];
mov.b32 %r128, {%rs20, %rs21};
shl.b64 %rd83, %rd7, 2;
add.s64 %rd84, %rd80, %rd83;
ld.global.nc.u32 %r23, [%rd84];
add.s64 %rd15, %rd102, %rd73;
ld.global.nc.u16 %rs15, [%rd15];

{ cvt.f32.f16 %f60, %rs15;}


ld.global.nc.u32 %r24, [%rd79];

{.reg .f16 low,high;
mov.b32 {low,high}, %r24;
mov.b16 %rs16, low;}


{ cvt.f32.f16 %f61, %rs16;}


mov.u32 %r126, 0;

dp4a.s32.s32 %r123, %r124, %r22, %r126;


dp4a.s32.s32 %r127, %r128, %r23, %r123;

mul.ftz.f32 %f62, %f60, %f61;
cvt.rn.f32.s32 %f63, %r127;
fma.rn.ftz.f32 %f15, %f62, %f63, %f149;
add.s64 %rd16, %rd103, %rd8;
setp.eq.s64 %p14, %rd23, 0;
@%p14 bra $L__BB90_17;

ld.global.u16 %rs25, [%rd16];
ld.global.u16 %rs26, [%rd16+2];
mov.b32 %r133, {%rs25, %rs26};
ld.global.u16 %rs27, [%rd16+4];
ld.global.u16 %rs28, [%rd16+6];
mov.b32 %r137, {%rs27, %rs28};
add.s64 %rd87, %rd103, %rd73;
ld.global.u16 %rs22, [%rd87];

{ cvt.f32.f16 %f64, %rs22;}



{.reg .f16 low,high;
mov.b32 {low,high}, %r24;
mov.b16 %rs23, low;}


{ cvt.f32.f16 %f65, %rs23;}



dp4a.s32.s32 %r132, %r133, %r22, %r126;


dp4a.s32.s32 %r136, %r137, %r23, %r132;

mul.ftz.f32 %f66, %f64, %f65;
cvt.rn.f32.s32 %f67, %r136;
fma.rn.ftz.f32 %f146, %f66, %f67, %f146;

$L__BB90_17:
add.s32 %r149, %r225, 32;
cvt.u64.u32 %rd88, %r149;
add.s64 %rd89, %rd4, %rd88;
mul.lo.s64 %rd90, %rd89, 36;
add.s64 %rd91, %rd11, %rd90;
add.s64 %rd92, %rd91, 4;
ld.global.nc.u16 %rs32, [%rd14+1088];
ld.global.nc.u16 %rs33, [%rd14+1090];
mov.b32 %r142, {%rs32, %rs33};
add.s64 %rd94, %rd92, %rd81;
ld.global.nc.u32 %r25, [%rd94];
ld.global.nc.u16 %rs34, [%rd14+1092];
ld.global.nc.u16 %rs35, [%rd14+1094];
mov.b32 %r146, {%rs34, %rs35};
add.s64 %rd96, %rd92, %rd83;
ld.global.nc.u32 %r26, [%rd96];
ld.global.nc.u16 %rs29, [%rd15+1088];

{ cvt.f32.f16 %f68, %rs29;}


ld.global.nc.u32 %r27, [%rd91];

{.reg .f16 low,high;
mov.b32 {low,high}, %r27;
mov.b16 %rs30, low;}


{ cvt.f32.f16 %f69, %rs30;}



dp4a.s32.s32 %r141, %r142, %r25, %r126;


dp4a.s32.s32 %r145, %r146, %r26, %r141;

mul.ftz.f32 %f70, %f68, %f69;
cvt.rn.f32.s32 %f71, %r145;
fma.rn.ftz.f32 %f149, %f70, %f71, %f15;
@%p14 bra $L__BB90_19;

ld.global.u16 %rs39, [%rd16+1088];
ld.global.u16 %rs40, [%rd16+1090];
mov.b32 %r152, {%rs39, %rs40};
ld.global.u16 %rs41, [%rd16+1092];
ld.global.u16 %rs42, [%rd16+1094];
mov.b32 %r156, {%rs41, %rs42};
add.s64 %rd98, %rd103, %rd73;
ld.global.u16 %rs36, [%rd98+1088];

{ cvt.f32.f16 %f72, %rs36;}



{.reg .f16 low,high;
mov.b32 {low,high}, %r27;
mov.b16 %rs37, low;}


{ cvt.f32.f16 %f73, %rs37;}


mov.u32 %r154, 0;

dp4a.s32.s32 %r151, %r152, %r25, %r154;


dp4a.s32.s32 %r155, %r156, %r26, %r151;

mul.ftz.f32 %f74, %f72, %f73;
cvt.rn.f32.s32 %f75, %r155;
fma.rn.ftz.f32 %f146, %f74, %f75, %f146;

$L__BB90_19:
add.s64 %rd103, %rd103, 2176;
add.s64 %rd102, %rd102, 2176;
add.s32 %r225, %r225, 64;
setp.lt.s32 %p16, %r225, %r76;
@%p16 bra $L__BB90_15;

$L__BB90_20:
setp.eq.s32 %p17, %r1, 0;
@%p17 bra $L__BB90_23;

setp.eq.s64 %p18, %rd23, 0;
add.s32 %r29, %r1, -1;
shl.b32 %r163, %r1, 7;
mov.u32 %r164, _ZZ13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r165, %r164, %r163;
shl.b32 %r166, %r2, 2;
add.s32 %r167, %r165, %r166;
st.shared.f32 [%r167+-128], %f149;
@%p18 bra $L__BB90_23;

shl.b32 %r169, %r29, 7;
mov.u32 %r170, _ZZ13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r171, %r170, %r169;
add.s32 %r173, %r171, %r166;
st.shared.f32 [%r173], %f146;

$L__BB90_23:
setp.ne.s32 %p19, %r1, 0;
bar.sync 0;
@%p19 bra $L__BB90_42;

setp.ne.s64 %p20, %rd23, 0;
shl.b32 %r176, %r2, 2;
mov.u32 %r177, _ZZ13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r30, %r177, %r176;
ld.shared.f32 %f76, [%r30];
add.ftz.f32 %f23, %f149, %f76;
@%p20 bra $L__BB90_26;
bra.uni $L__BB90_25;

$L__BB90_26:
mov.u32 %r180, _ZZ13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r181, %r180, %r176;
ld.shared.f32 %f80, [%r181];
add.ftz.f32 %f81, %f146, %f80;
ld.shared.f32 %f82, [%r30+128];
add.ftz.f32 %f83, %f23, %f82;
ld.shared.f32 %f84, [%r181+128];
add.ftz.f32 %f85, %f81, %f84;
ld.shared.f32 %f86, [%r30+256];
add.ftz.f32 %f151, %f83, %f86;
ld.shared.f32 %f87, [%r181+256];
add.ftz.f32 %f146, %f85, %f87;
bra.uni $L__BB90_27;

$L__BB90_25:
ld.shared.f32 %f77, [%r30+128];
add.ftz.f32 %f78, %f23, %f77;
ld.shared.f32 %f79, [%r30+256];
add.ftz.f32 %f151, %f78, %f79;

$L__BB90_27:
mov.b32 %r182, %f151;
mov.u32 %r183, 31;
mov.u32 %r184, 16;
mov.u32 %r185, -1;
shfl.sync.bfly.b32 %r186|%p21, %r182, %r184, %r183, %r185;
mov.b32 %f88, %r186;
add.ftz.f32 %f89, %f151, %f88;
mov.b32 %r187, %f89;
mov.u32 %r188, 8;
shfl.sync.bfly.b32 %r189|%p22, %r187, %r188, %r183, %r185;
mov.b32 %f90, %r189;
add.ftz.f32 %f91, %f89, %f90;
mov.b32 %r190, %f91;
mov.u32 %r191, 4;
shfl.sync.bfly.b32 %r192|%p23, %r190, %r191, %r183, %r185;
mov.b32 %f92, %r192;
add.ftz.f32 %f93, %f91, %f92;
mov.b32 %r193, %f93;
mov.u32 %r194, 2;
shfl.sync.bfly.b32 %r195|%p24, %r193, %r194, %r183, %r185;
mov.b32 %f94, %r195;
add.ftz.f32 %f29, %f93, %f94;
mov.b32 %r196, %f29;
mov.u32 %r197, 1;
shfl.sync.bfly.b32 %r31|%p1, %r196, %r197, %r183, %r185;
setp.eq.s64 %p25, %rd23, 0;
@%p25 bra $L__BB90_29;

mov.b32 %r198, %f146;
shfl.sync.bfly.b32 %r202|%p26, %r198, %r184, %r183, %r185;
mov.b32 %f95, %r202;
add.ftz.f32 %f96, %f146, %f95;
mov.b32 %r203, %f96;
shfl.sync.bfly.b32 %r205|%p27, %r203, %r188, %r183, %r185;
mov.b32 %f97, %r205;
add.ftz.f32 %f98, %f96, %f97;
mov.b32 %r206, %f98;
shfl.sync.bfly.b32 %r208|%p28, %r206, %r191, %r183, %r185;
mov.b32 %f99, %r208;
add.ftz.f32 %f100, %f98, %f99;
mov.b32 %r209, %f100;
shfl.sync.bfly.b32 %r211|%p29, %r209, %r194, %r183, %r185;
mov.b32 %f101, %r211;
add.ftz.f32 %f102, %f100, %f101;
mov.b32 %r212, %f102;
shfl.sync.bfly.b32 %r214|%p30, %r212, %r197, %r183, %r185;
mov.b32 %f103, %r214;
add.ftz.f32 %f146, %f102, %f103;

$L__BB90_29:
setp.ne.s32 %p31, %r2, 0;
mov.b32 %f104, %r31;
add.ftz.f32 %f32, %f29, %f104;
@%p31 bra $L__BB90_42;

add.ftz.f32 %f153, %f138, %f32;
@%p25 bra $L__BB90_41;

add.ftz.f32 %f34, %f137, %f146;
setp.gt.s32 %p33, %r32, 2;
@%p33 bra $L__BB90_35;

setp.eq.s32 %p36, %r32, 1;
@%p36 bra $L__BB90_39;

setp.eq.s32 %p37, %r32, 2;
@%p37 bra $L__BB90_34;
bra.uni $L__BB90_40;

$L__BB90_34:
mul.ftz.f32 %f124, %f34, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f125, %f124;
add.ftz.f32 %f126, %f125, 0f3F800000;
div.approx.ftz.f32 %f127, %f34, %f126;
mul.ftz.f32 %f153, %f153, %f127;
bra.uni $L__BB90_41;

$L__BB90_35:
setp.eq.s32 %p34, %r32, 3;
@%p34 bra $L__BB90_38;

setp.ne.s32 %p35, %r32, 6;
@%p35 bra $L__BB90_40;

min.ftz.f32 %f105, %f34, %f41;
neg.ftz.f32 %f106, %f41;
min.ftz.f32 %f107, %f153, %f41;
max.ftz.f32 %f108, %f107, %f106;
mul.ftz.f32 %f109, %f105, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f110, %f109;
add.ftz.f32 %f111, %f110, 0f3F800000;
div.approx.ftz.f32 %f112, %f105, %f111;
mul.ftz.f32 %f153, %f108, %f112;
bra.uni $L__BB90_41;

$L__BB90_39:
mul.ftz.f32 %f128, %f34, 0f3F000000;
mul.ftz.f32 %f129, %f34, 0f3D372713;
fma.rn.ftz.f32 %f130, %f34, %f129, 0f3F800000;
mul.ftz.f32 %f131, %f34, 0f3F4C422A;
mul.ftz.f32 %f132, %f131, %f130;
tanh.approx.f32 %f133, %f132;
add.ftz.f32 %f134, %f133, 0f3F800000;
mul.ftz.f32 %f135, %f128, %f134;
mul.ftz.f32 %f153, %f153, %f135;
bra.uni $L__BB90_41;

$L__BB90_38:
mov.f32 %f113, 0f40E00000;
min.ftz.f32 %f114, %f34, %f113;
min.ftz.f32 %f115, %f153, %f113;
mov.f32 %f116, 0fC0E00000;
max.ftz.f32 %f117, %f115, %f116;
mul.ftz.f32 %f118, %f114, 0f3FD9DB23;
mul.ftz.f32 %f119, %f118, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f120, %f119;
add.ftz.f32 %f121, %f120, 0f3F800000;
div.approx.ftz.f32 %f122, %f114, %f121;
add.ftz.f32 %f123, %f117, 0f3F800000;
mul.ftz.f32 %f153, %f123, %f122;
bra.uni $L__BB90_41;

$L__BB90_40:
mul.ftz.f32 %f153, %f153, %f34;

$L__BB90_41:
cvta.to.global.u64 %rd99, %rd27;
mov.u32 %r218, %ctaid.y;
mov.u32 %r219, %ctaid.x;
mad.lo.s32 %r220, %r218, %r43, %r219;
mad.lo.s32 %r221, %r70, %r49, %r220;
mul.wide.u32 %rd100, %r221, 4;
add.s64 %rd101, %rd99, %rd100;
st.global.f32 [%rd101], %f153;

$L__BB90_42:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<15>;
.reg .b16 %rs<36>;
.reg .f32 %f<56>;
.reg .b32 %r<163>;
.reg .b64 %rd<75>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[384];

ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r45, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r46, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
cvta.to.global.u64 %rd1, %rd18;
cvta.to.global.u64 %rd74, %rd17;
mov.u32 %r1, %tid.y;
shl.b32 %r47, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r47, %r2;
mov.u32 %r4, %ctaid.x;
shr.u32 %r5, %r30, 5;
mov.u32 %r158, %ctaid.y;
setp.eq.s64 %p1, %rd15, 0;
@%p1 bra $L__BB91_2;

cvta.to.global.u64 %rd19, %rd15;
mul.wide.u32 %rd20, %r158, 4;
add.s64 %rd21, %rd19, %rd20;
ld.global.nc.u32 %r157, [%rd21];
mul.hi.u32 %r48, %r158, %r31;
add.s32 %r49, %r48, %r158;
shr.u32 %r50, %r49, %r32;
mul.lo.s32 %r51, %r50, %r33;
sub.s32 %r158, %r158, %r51;
bra.uni $L__BB91_3;

$L__BB91_2:
mul.hi.u32 %r52, %r158, %r35;
add.s32 %r53, %r52, %r158;
shr.u32 %r157, %r53, %r36;

$L__BB91_3:
mov.u32 %r12, %ctaid.z;
mul.hi.u32 %r54, %r12, %r41;
add.s32 %r55, %r54, %r12;
shr.u32 %r56, %r55, %r42;
mul.lo.s32 %r57, %r12, %r45;
cvt.u64.u32 %rd22, %r57;
mul.lo.s32 %r58, %r158, %r39;
cvt.u64.u32 %rd23, %r58;
add.s64 %rd3, %rd23, %rd22;
mul.lo.s32 %r13, %r56, %r44;
mul.lo.s32 %r14, %r157, %r38;
mul.lo.s32 %r15, %r4, %r34;
shr.s32 %r59, %r3, 31;
shr.u32 %r60, %r59, 30;
add.s32 %r61, %r3, %r60;
shr.s32 %r161, %r61, 2;
setp.ge.s32 %p2, %r161, %r5;
mov.f32 %f55, 0f00000000;
@%p2 bra $L__BB91_10;

and.b32 %r65, %r61, -4;
sub.s32 %r17, %r3, %r65;
shl.b32 %r18, %r17, 1;
not.b32 %r66, %r161;
add.s32 %r19, %r5, %r66;
shr.u32 %r67, %r19, 5;
add.s32 %r68, %r67, 1;
and.b32 %r160, %r68, 3;
setp.eq.s32 %p3, %r160, 0;
mov.f32 %f55, 0f00000000;
@%p3 bra $L__BB91_7;

shl.b32 %r69, %r17, 2;
or.b32 %r70, %r69, 1;
sub.s32 %r71, %r70, %r18;
cvt.s64.s32 %rd4, %r71;
add.s32 %r72, %r161, %r13;
add.s32 %r73, %r72, %r14;
add.s32 %r74, %r73, %r15;
mul.wide.s32 %rd24, %r74, 34;
add.s64 %rd73, %rd74, %rd24;
mul.wide.s32 %rd25, %r69, 2;
or.b64 %rd6, %rd25, 2;
mov.f32 %f55, 0f00000000;
shl.b64 %rd34, %rd4, 2;

$L__BB91_6:
.pragma "nounroll";
cvt.u64.u32 %rd26, %r161;
add.s64 %rd27, %rd3, %rd26;
mul.lo.s64 %rd28, %rd27, 36;
add.s64 %rd29, %rd1, %rd28;
add.s64 %rd30, %rd29, 4;
add.s64 %rd31, %rd73, %rd6;
ld.global.nc.u16 %rs4, [%rd31];
ld.global.nc.u16 %rs5, [%rd31+2];
mov.b32 %r77, {%rs4, %rs5};
mul.wide.s32 %rd32, %r18, 4;
add.s64 %rd33, %rd30, %rd32;
ld.global.nc.u32 %r78, [%rd33];
ld.global.nc.u16 %rs6, [%rd31+4];
ld.global.nc.u16 %rs7, [%rd31+6];
mov.b32 %r81, {%rs6, %rs7};
add.s64 %rd35, %rd30, %rd34;
ld.global.nc.u32 %r82, [%rd35];
ld.global.nc.u16 %rs1, [%rd73];

{ cvt.f32.f16 %f13, %rs1;}


ld.global.nc.u32 %r75, [%rd29];

{.reg .f16 low,high;
mov.b32 {low,high}, %r75;
mov.b16 %rs2, low;}


{ cvt.f32.f16 %f14, %rs2;}


mov.u32 %r79, 0;

dp4a.s32.s32 %r76, %r77, %r78, %r79;


dp4a.s32.s32 %r80, %r81, %r82, %r76;

mul.ftz.f32 %f15, %f13, %f14;
cvt.rn.f32.s32 %f16, %r80;
fma.rn.ftz.f32 %f55, %f15, %f16, %f55;
add.s32 %r161, %r161, 32;
add.s64 %rd73, %rd73, 1088;
add.s32 %r160, %r160, -1;
setp.ne.s32 %p4, %r160, 0;
@%p4 bra $L__BB91_6;

$L__BB91_7:
setp.lt.u32 %p5, %r19, 96;
@%p5 bra $L__BB91_10;

shl.b32 %r85, %r17, 2;
or.b32 %r86, %r85, 1;
cvt.s64.s32 %rd9, %r18;
sub.s32 %r87, %r86, %r18;
cvt.s64.s32 %rd10, %r87;
add.s32 %r88, %r161, %r13;
add.s32 %r89, %r88, %r14;
add.s32 %r90, %r89, %r15;
mul.wide.s32 %rd36, %r85, 2;
mul.wide.s32 %rd11, %r90, 34;
add.s64 %rd37, %rd11, %rd36;
add.s64 %rd12, %rd37, 4;
shl.b64 %rd44, %rd9, 2;
shl.b64 %rd46, %rd10, 2;

$L__BB91_9:
cvt.u64.u32 %rd38, %r161;
add.s64 %rd39, %rd3, %rd38;
mul.lo.s64 %rd40, %rd39, 36;
add.s64 %rd41, %rd1, %rd40;
add.s64 %rd42, %rd41, 4;
add.s64 %rd43, %rd74, %rd12;
ld.global.nc.u16 %rs20, [%rd43+-2];
ld.global.nc.u16 %rs21, [%rd43];
mov.b32 %r93, {%rs20, %rs21};
add.s64 %rd45, %rd42, %rd44;
ld.global.nc.u32 %r94, [%rd45];
ld.global.nc.u16 %rs22, [%rd43+2];
ld.global.nc.u16 %rs23, [%rd43+4];
mov.b32 %r97, {%rs22, %rs23};
add.s64 %rd47, %rd42, %rd46;
ld.global.nc.u32 %r98, [%rd47];
add.s64 %rd48, %rd74, %rd11;
ld.global.nc.u16 %rs8, [%rd48];

{ cvt.f32.f16 %f17, %rs8;}


ld.global.nc.u32 %r91, [%rd41];

{.reg .f16 low,high;
mov.b32 {low,high}, %r91;
mov.b16 %rs9, low;}


{ cvt.f32.f16 %f18, %rs9;}


mov.u32 %r122, 0;

dp4a.s32.s32 %r92, %r93, %r94, %r122;


dp4a.s32.s32 %r96, %r97, %r98, %r92;

mul.ftz.f32 %f25, %f17, %f18;
cvt.rn.f32.s32 %f26, %r96;
fma.rn.ftz.f32 %f27, %f25, %f26, %f55;
add.s32 %r127, %r161, 32;
cvt.u64.u32 %rd49, %r127;
add.s64 %rd50, %rd3, %rd49;
mul.lo.s64 %rd51, %rd50, 36;
add.s64 %rd52, %rd1, %rd51;
add.s64 %rd53, %rd52, 4;
ld.global.nc.u16 %rs24, [%rd43+1086];
ld.global.nc.u16 %rs25, [%rd43+1088];
mov.b32 %r102, {%rs24, %rs25};
add.s64 %rd54, %rd53, %rd44;
ld.global.nc.u32 %r103, [%rd54];
ld.global.nc.u16 %rs26, [%rd43+1090];
ld.global.nc.u16 %rs27, [%rd43+1092];
mov.b32 %r106, {%rs26, %rs27};
add.s64 %rd55, %rd53, %rd46;
ld.global.nc.u32 %r107, [%rd55];
ld.global.nc.u16 %rs11, [%rd48+1088];

{ cvt.f32.f16 %f19, %rs11;}


ld.global.nc.u32 %r100, [%rd52];

{.reg .f16 low,high;
mov.b32 {low,high}, %r100;
mov.b16 %rs12, low;}


{ cvt.f32.f16 %f20, %rs12;}



dp4a.s32.s32 %r101, %r102, %r103, %r122;


dp4a.s32.s32 %r105, %r106, %r107, %r101;

mul.ftz.f32 %f28, %f19, %f20;
cvt.rn.f32.s32 %f29, %r105;
fma.rn.ftz.f32 %f30, %f28, %f29, %f27;
add.s32 %r128, %r161, 64;
cvt.u64.u32 %rd56, %r128;
add.s64 %rd57, %rd3, %rd56;
mul.lo.s64 %rd58, %rd57, 36;
add.s64 %rd59, %rd1, %rd58;
add.s64 %rd60, %rd59, 4;
ld.global.nc.u16 %rs28, [%rd43+2174];
ld.global.nc.u16 %rs29, [%rd43+2176];
mov.b32 %r111, {%rs28, %rs29};
add.s64 %rd61, %rd60, %rd44;
ld.global.nc.u32 %r112, [%rd61];
ld.global.nc.u16 %rs30, [%rd43+2178];
ld.global.nc.u16 %rs31, [%rd43+2180];
mov.b32 %r115, {%rs30, %rs31};
add.s64 %rd62, %rd60, %rd46;
ld.global.nc.u32 %r116, [%rd62];
ld.global.nc.u16 %rs14, [%rd48+2176];

{ cvt.f32.f16 %f21, %rs14;}


ld.global.nc.u32 %r109, [%rd59];

{.reg .f16 low,high;
mov.b32 {low,high}, %r109;
mov.b16 %rs15, low;}


{ cvt.f32.f16 %f22, %rs15;}



dp4a.s32.s32 %r110, %r111, %r112, %r122;


dp4a.s32.s32 %r114, %r115, %r116, %r110;

mul.ftz.f32 %f31, %f21, %f22;
cvt.rn.f32.s32 %f32, %r114;
fma.rn.ftz.f32 %f33, %f31, %f32, %f30;
add.s32 %r129, %r161, 96;
cvt.u64.u32 %rd63, %r129;
add.s64 %rd64, %rd3, %rd63;
mul.lo.s64 %rd65, %rd64, 36;
add.s64 %rd66, %rd1, %rd65;
add.s64 %rd67, %rd66, 4;
ld.global.nc.u16 %rs32, [%rd43+3262];
ld.global.nc.u16 %rs33, [%rd43+3264];
mov.b32 %r120, {%rs32, %rs33};
add.s64 %rd68, %rd67, %rd44;
ld.global.nc.u32 %r121, [%rd68];
ld.global.nc.u16 %rs34, [%rd43+3266];
ld.global.nc.u16 %rs35, [%rd43+3268];
mov.b32 %r124, {%rs34, %rs35};
add.s64 %rd69, %rd67, %rd46;
ld.global.nc.u32 %r125, [%rd69];
ld.global.nc.u16 %rs17, [%rd48+3264];

{ cvt.f32.f16 %f23, %rs17;}


ld.global.nc.u32 %r118, [%rd66];

{.reg .f16 low,high;
mov.b32 {low,high}, %r118;
mov.b16 %rs18, low;}


{ cvt.f32.f16 %f24, %rs18;}



dp4a.s32.s32 %r119, %r120, %r121, %r122;


dp4a.s32.s32 %r123, %r124, %r125, %r119;

mul.ftz.f32 %f34, %f23, %f24;
cvt.rn.f32.s32 %f35, %r123;
fma.rn.ftz.f32 %f55, %f34, %f35, %f33;
add.s64 %rd74, %rd74, 4352;
add.s32 %r161, %r161, 128;
setp.lt.s32 %p6, %r161, %r5;
@%p6 bra $L__BB91_9;

$L__BB91_10:
setp.eq.s32 %p7, %r1, 0;
@%p7 bra $L__BB91_12;

shl.b32 %r130, %r1, 7;
mov.u32 %r131, _ZZ13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r132, %r131, %r130;
shl.b32 %r133, %r2, 2;
add.s32 %r134, %r132, %r133;
st.shared.f32 [%r134+-128], %f55;

$L__BB91_12:
setp.ne.s32 %p8, %r1, 0;
bar.sync 0;
@%p8 bra $L__BB91_15;

mov.u32 %r135, %ctaid.y;
mad.lo.s32 %r136, %r135, %r40, %r4;
mad.lo.s32 %r29, %r12, %r46, %r136;
shl.b32 %r137, %r2, 2;
mov.u32 %r138, _ZZ13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r139, %r138, %r137;
mov.u32 %r140, 2;
ld.shared.f32 %f36, [%r139];
add.ftz.f32 %f37, %f55, %f36;
ld.shared.f32 %f38, [%r139+128];
add.ftz.f32 %f39, %f37, %f38;
ld.shared.f32 %f40, [%r139+256];
add.ftz.f32 %f41, %f39, %f40;
mov.b32 %r141, %f41;
mov.u32 %r142, 31;
mov.u32 %r143, 16;
mov.u32 %r144, -1;
shfl.sync.bfly.b32 %r145|%p9, %r141, %r143, %r142, %r144;
mov.b32 %f42, %r145;
add.ftz.f32 %f43, %f41, %f42;
mov.b32 %r146, %f43;
mov.u32 %r147, 8;
shfl.sync.bfly.b32 %r148|%p10, %r146, %r147, %r142, %r144;
mov.b32 %f44, %r148;
add.ftz.f32 %f45, %f43, %f44;
mov.b32 %r149, %f45;
mov.u32 %r150, 4;
shfl.sync.bfly.b32 %r151|%p11, %r149, %r150, %r142, %r144;
mov.b32 %f46, %r151;
add.ftz.f32 %f47, %f45, %f46;
mov.b32 %r152, %f47;
shfl.sync.bfly.b32 %r153|%p12, %r152, %r140, %r142, %r144;
mov.b32 %f48, %r153;
add.ftz.f32 %f49, %f47, %f48;
mov.b32 %r154, %f49;
mov.u32 %r155, 1;
shfl.sync.bfly.b32 %r156|%p13, %r154, %r155, %r142, %r144;
mov.b32 %f50, %r156;
add.ftz.f32 %f8, %f49, %f50;
setp.ne.s32 %p14, %r2, 0;
@%p14 bra $L__BB91_15;

cvta.to.global.u64 %rd70, %rd16;
mul.wide.u32 %rd71, %r29, 4;
add.s64 %rd72, %rd70, %rd71;
st.global.f32 [%rd72], %f8;

$L__BB91_15:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<38>;
.reg .b16 %rs<43>;
.reg .f32 %f<154>;
.reg .b32 %r<226>;
.reg .b64 %rd<104>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[384];

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate[384];

ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd20, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd21, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd27, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r47, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r48, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r49, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r45, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
ld.param.f32 %f41, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
ld.param.u64 %rd24, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+16];
ld.param.u64 %rd23, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+8];
ld.param.u64 %rd22, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3];
mov.u32 %r1, %tid.y;
shl.b32 %r50, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r50, %r2;
mov.u32 %r4, %ctaid.y;
setp.eq.s64 %p2, %rd21, 0;
@%p2 bra $L__BB92_2;

cvta.to.global.u64 %rd28, %rd21;
mul.wide.u32 %rd29, %r4, 4;
add.s64 %rd30, %rd28, %rd29;
ld.global.nc.u32 %r222, [%rd30];
mul.hi.u32 %r51, %r4, %r34;
add.s32 %r52, %r51, %r4;
shr.u32 %r53, %r52, %r35;
mul.lo.s32 %r54, %r53, %r36;
sub.s32 %r223, %r4, %r54;
bra.uni $L__BB92_3;

$L__BB92_2:
mul.hi.u32 %r55, %r4, %r38;
add.s32 %r56, %r55, %r4;
shr.u32 %r222, %r56, %r39;
mov.u32 %r223, %r4;

$L__BB92_3:
or.b32 %r57, %r2, %r1;
setp.ne.s32 %p3, %r57, 0;
mov.f32 %f146, 0f00000000;
mov.f32 %f137, %f146;
mov.f32 %f138, %f146;
@%p3 bra $L__BB92_8;

setp.eq.s64 %p4, %rd22, 0;
mov.f32 %f137, 0f00000000;
mov.f32 %f138, %f137;
@%p4 bra $L__BB92_6;

cvta.to.global.u64 %rd31, %rd22;
mov.u32 %r58, %ctaid.z;
mul.lo.s32 %r59, %r58, %r49;
cvt.u64.u32 %rd32, %r59;
selp.b32 %r60, %r4, %r222, %p2;
mul.lo.s32 %r61, %r60, %r43;
cvt.u64.u32 %rd33, %r61;
mov.u32 %r62, %ctaid.x;
cvt.s64.s32 %rd34, %r62;
cvt.u64.u32 %rd35, %r2;
add.s64 %rd36, %rd35, %rd34;
add.s64 %rd37, %rd36, %rd32;
add.s64 %rd38, %rd37, %rd33;
shl.b64 %rd39, %rd38, 2;
add.s64 %rd40, %rd31, %rd39;
ld.global.f32 %f138, [%rd40];

$L__BB92_6:
setp.eq.s64 %p6, %rd24, 0;
setp.eq.s64 %p7, %rd23, 0;
or.pred %p8, %p6, %p7;
@%p8 bra $L__BB92_8;

cvta.to.global.u64 %rd41, %rd24;
mov.u32 %r64, %ctaid.z;
mul.lo.s32 %r65, %r64, %r49;
cvt.u64.u32 %rd42, %r65;
selp.b32 %r66, %r4, %r222, %p2;
mul.lo.s32 %r67, %r66, %r43;
cvt.u64.u32 %rd43, %r67;
mov.u32 %r68, %ctaid.x;
cvt.s64.s32 %rd44, %r68;
cvt.u64.u32 %rd45, %r2;
add.s64 %rd46, %rd45, %rd44;
add.s64 %rd47, %rd46, %rd42;
add.s64 %rd48, %rd47, %rd43;
shl.b64 %rd49, %rd48, 2;
add.s64 %rd50, %rd41, %rd49;
ld.global.f32 %f137, [%rd50];

$L__BB92_8:
mov.u32 %r70, %ctaid.z;
mul.lo.s32 %r71, %r70, %r48;
cvt.u64.u32 %rd51, %r71;
mul.lo.s32 %r72, %r223, %r42;
cvt.u64.u32 %rd52, %r72;
add.s64 %rd4, %rd52, %rd51;
mul.hi.u32 %r73, %r70, %r44;
add.s32 %r74, %r73, %r70;
shr.u32 %r75, %r74, %r45;
mul.lo.s32 %r10, %r75, %r47;
mul.lo.s32 %r11, %r222, %r41;
shr.u32 %r76, %r33, 5;
shr.s32 %r77, %r3, 31;
shr.u32 %r78, %r77, 30;
add.s32 %r79, %r3, %r78;
shr.s32 %r225, %r79, 2;
setp.ge.s32 %p10, %r225, %r76;
mov.f32 %f149, %f146;
@%p10 bra $L__BB92_20;

and.b32 %r83, %r79, -4;
sub.s32 %r13, %r3, %r83;
not.b32 %r85, %r225;
add.s32 %r14, %r76, %r85;
and.b32 %r86, %r14, 32;
setp.ne.s32 %p11, %r86, 0;
mov.f32 %f149, 0f00000000;
mov.f32 %f146, %f149;
@%p11 bra $L__BB92_13;

cvta.to.global.u64 %rd53, %rd19;
cvta.to.global.u64 %rd54, %rd20;
cvt.u64.u32 %rd55, %r225;
add.s64 %rd56, %rd4, %rd55;
mov.u32 %r96, %ctaid.x;
mad.lo.s32 %r97, %r96, %r37, %r11;
add.s32 %r98, %r97, %r10;
add.s32 %r99, %r98, %r225;
cvt.s64.s32 %rd5, %r99;
mul.wide.s32 %rd57, %r99, 34;
add.s64 %rd58, %rd53, %rd57;
shl.b32 %r100, %r13, 2;
mul.wide.s32 %rd59, %r100, 2;
add.s64 %rd60, %rd58, %rd59;
mul.lo.s64 %rd61, %rd56, 36;
add.s64 %rd62, %rd54, %rd61;
add.s64 %rd63, %rd62, 4;
ld.global.nc.u16 %rs4, [%rd60+2];
or.b32 %r101, %r100, 1;
ld.global.nc.u16 %rs5, [%rd60+4];
mov.b32 %r89, {%rs4, %rs5};
shl.b32 %r102, %r13, 1;
mul.wide.s32 %rd64, %r102, 4;
add.s64 %rd65, %rd63, %rd64;
ld.global.nc.u32 %r15, [%rd65];
sub.s32 %r103, %r101, %r102;
ld.global.nc.u16 %rs6, [%rd60+6];
ld.global.nc.u16 %rs7, [%rd60+8];
mov.b32 %r93, {%rs6, %rs7};
mul.wide.s32 %rd66, %r103, 4;
add.s64 %rd67, %rd63, %rd66;
ld.global.nc.u32 %r16, [%rd67];
ld.global.nc.u16 %rs1, [%rd58];

{ cvt.f32.f16 %f51, %rs1;}


ld.global.nc.u32 %r17, [%rd62];

{.reg .f16 low,high;
mov.b32 {low,high}, %r17;
mov.b16 %rs2, low;}


{ cvt.f32.f16 %f52, %rs2;}


mov.u32 %r91, 0;

dp4a.s32.s32 %r88, %r89, %r15, %r91;


dp4a.s32.s32 %r92, %r93, %r16, %r88;

mul.ftz.f32 %f54, %f51, %f52;
cvt.rn.f32.s32 %f55, %r92;
fma.rn.ftz.f32 %f149, %f54, %f55, 0f00000000;
mov.f32 %f146, 0f00000000;
setp.eq.s64 %p12, %rd23, 0;
@%p12 bra $L__BB92_12;

cvta.to.global.u64 %rd68, %rd23;
mul.lo.s64 %rd69, %rd5, 34;
add.s64 %rd70, %rd68, %rd69;
add.s64 %rd72, %rd70, %rd59;
ld.global.u16 %rs11, [%rd72+2];
ld.global.u16 %rs12, [%rd72+4];
mov.b32 %r106, {%rs11, %rs12};
ld.global.u16 %rs13, [%rd72+6];
ld.global.u16 %rs14, [%rd72+8];
mov.b32 %r110, {%rs13, %rs14};
ld.global.u16 %rs8, [%rd70];

{ cvt.f32.f16 %f56, %rs8;}



{.reg .f16 low,high;
mov.b32 {low,high}, %r17;
mov.b16 %rs9, low;}


{ cvt.f32.f16 %f57, %rs9;}



dp4a.s32.s32 %r105, %r106, %r15, %r91;


dp4a.s32.s32 %r109, %r110, %r16, %r105;

mul.ftz.f32 %f58, %f56, %f57;
cvt.rn.f32.s32 %f59, %r109;
fma.rn.ftz.f32 %f146, %f58, %f59, 0f00000000;

$L__BB92_12:
add.s32 %r225, %r225, 32;

$L__BB92_13:
and.b32 %r114, %r14, -32;
setp.eq.s32 %p13, %r114, 0;
@%p13 bra $L__BB92_20;

shl.b32 %r115, %r13, 2;
or.b32 %r116, %r115, 1;
shl.b32 %r117, %r13, 1;
cvt.s64.s32 %rd6, %r117;
sub.s32 %r118, %r116, %r117;
cvt.s64.s32 %rd7, %r118;
add.s32 %r119, %r225, %r10;
add.s32 %r120, %r119, %r11;
mov.u32 %r121, %ctaid.x;
mad.lo.s32 %r20, %r121, %r37, %r120;
mul.wide.s32 %rd73, %r20, 34;
mul.wide.s32 %rd74, %r115, 2;
add.s64 %rd75, %rd73, %rd74;
add.s64 %rd8, %rd75, 2;
cvta.to.global.u64 %rd102, %rd19;
cvta.to.global.u64 %rd103, %rd23;
cvta.to.global.u64 %rd11, %rd20;

$L__BB92_15:
cvt.u64.u32 %rd76, %r225;
add.s64 %rd77, %rd4, %rd76;
mul.lo.s64 %rd78, %rd77, 36;
add.s64 %rd79, %rd11, %rd78;
add.s64 %rd80, %rd79, 4;
add.s64 %rd14, %rd102, %rd8;
ld.global.nc.u16 %rs18, [%rd14];
ld.global.nc.u16 %rs19, [%rd14+2];
mov.b32 %r124, {%rs18, %rs19};
shl.b64 %rd81, %rd6, 2;
add.s64 %rd82, %rd80, %rd81;
ld.global.nc.u32 %r22, [%rd82];
ld.global.nc.u16 %rs20, [%rd14+4];
ld.global.nc.u16 %rs21, [%rd14+6];
mov.b32 %r128, {%rs20, %rs21};
shl.b64 %rd83, %rd7, 2;
add.s64 %rd84, %rd80, %rd83;
ld.global.nc.u32 %r23, [%rd84];
add.s64 %rd15, %rd102, %rd73;
ld.global.nc.u16 %rs15, [%rd15];

{ cvt.f32.f16 %f60, %rs15;}


ld.global.nc.u32 %r24, [%rd79];

{.reg .f16 low,high;
mov.b32 {low,high}, %r24;
mov.b16 %rs16, low;}


{ cvt.f32.f16 %f61, %rs16;}


mov.u32 %r126, 0;

dp4a.s32.s32 %r123, %r124, %r22, %r126;


dp4a.s32.s32 %r127, %r128, %r23, %r123;

mul.ftz.f32 %f62, %f60, %f61;
cvt.rn.f32.s32 %f63, %r127;
fma.rn.ftz.f32 %f15, %f62, %f63, %f149;
add.s64 %rd16, %rd103, %rd8;
setp.eq.s64 %p14, %rd23, 0;
@%p14 bra $L__BB92_17;

ld.global.u16 %rs25, [%rd16];
ld.global.u16 %rs26, [%rd16+2];
mov.b32 %r133, {%rs25, %rs26};
ld.global.u16 %rs27, [%rd16+4];
ld.global.u16 %rs28, [%rd16+6];
mov.b32 %r137, {%rs27, %rs28};
add.s64 %rd87, %rd103, %rd73;
ld.global.u16 %rs22, [%rd87];

{ cvt.f32.f16 %f64, %rs22;}



{.reg .f16 low,high;
mov.b32 {low,high}, %r24;
mov.b16 %rs23, low;}


{ cvt.f32.f16 %f65, %rs23;}



dp4a.s32.s32 %r132, %r133, %r22, %r126;


dp4a.s32.s32 %r136, %r137, %r23, %r132;

mul.ftz.f32 %f66, %f64, %f65;
cvt.rn.f32.s32 %f67, %r136;
fma.rn.ftz.f32 %f146, %f66, %f67, %f146;

$L__BB92_17:
add.s32 %r149, %r225, 32;
cvt.u64.u32 %rd88, %r149;
add.s64 %rd89, %rd4, %rd88;
mul.lo.s64 %rd90, %rd89, 36;
add.s64 %rd91, %rd11, %rd90;
add.s64 %rd92, %rd91, 4;
ld.global.nc.u16 %rs32, [%rd14+1088];
ld.global.nc.u16 %rs33, [%rd14+1090];
mov.b32 %r142, {%rs32, %rs33};
add.s64 %rd94, %rd92, %rd81;
ld.global.nc.u32 %r25, [%rd94];
ld.global.nc.u16 %rs34, [%rd14+1092];
ld.global.nc.u16 %rs35, [%rd14+1094];
mov.b32 %r146, {%rs34, %rs35};
add.s64 %rd96, %rd92, %rd83;
ld.global.nc.u32 %r26, [%rd96];
ld.global.nc.u16 %rs29, [%rd15+1088];

{ cvt.f32.f16 %f68, %rs29;}


ld.global.nc.u32 %r27, [%rd91];

{.reg .f16 low,high;
mov.b32 {low,high}, %r27;
mov.b16 %rs30, low;}


{ cvt.f32.f16 %f69, %rs30;}



dp4a.s32.s32 %r141, %r142, %r25, %r126;


dp4a.s32.s32 %r145, %r146, %r26, %r141;

mul.ftz.f32 %f70, %f68, %f69;
cvt.rn.f32.s32 %f71, %r145;
fma.rn.ftz.f32 %f149, %f70, %f71, %f15;
@%p14 bra $L__BB92_19;

ld.global.u16 %rs39, [%rd16+1088];
ld.global.u16 %rs40, [%rd16+1090];
mov.b32 %r152, {%rs39, %rs40};
ld.global.u16 %rs41, [%rd16+1092];
ld.global.u16 %rs42, [%rd16+1094];
mov.b32 %r156, {%rs41, %rs42};
add.s64 %rd98, %rd103, %rd73;
ld.global.u16 %rs36, [%rd98+1088];

{ cvt.f32.f16 %f72, %rs36;}



{.reg .f16 low,high;
mov.b32 {low,high}, %r27;
mov.b16 %rs37, low;}


{ cvt.f32.f16 %f73, %rs37;}


mov.u32 %r154, 0;

dp4a.s32.s32 %r151, %r152, %r25, %r154;


dp4a.s32.s32 %r155, %r156, %r26, %r151;

mul.ftz.f32 %f74, %f72, %f73;
cvt.rn.f32.s32 %f75, %r155;
fma.rn.ftz.f32 %f146, %f74, %f75, %f146;

$L__BB92_19:
add.s64 %rd103, %rd103, 2176;
add.s64 %rd102, %rd102, 2176;
add.s32 %r225, %r225, 64;
setp.lt.s32 %p16, %r225, %r76;
@%p16 bra $L__BB92_15;

$L__BB92_20:
setp.eq.s32 %p17, %r1, 0;
@%p17 bra $L__BB92_23;

setp.eq.s64 %p18, %rd23, 0;
add.s32 %r29, %r1, -1;
shl.b32 %r163, %r1, 7;
mov.u32 %r164, _ZZ13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r165, %r164, %r163;
shl.b32 %r166, %r2, 2;
add.s32 %r167, %r165, %r166;
st.shared.f32 [%r167+-128], %f149;
@%p18 bra $L__BB92_23;

shl.b32 %r169, %r29, 7;
mov.u32 %r170, _ZZ13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r171, %r170, %r169;
add.s32 %r173, %r171, %r166;
st.shared.f32 [%r173], %f146;

$L__BB92_23:
setp.ne.s32 %p19, %r1, 0;
bar.sync 0;
@%p19 bra $L__BB92_42;

setp.ne.s64 %p20, %rd23, 0;
shl.b32 %r176, %r2, 2;
mov.u32 %r177, _ZZ13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r30, %r177, %r176;
ld.shared.f32 %f76, [%r30];
add.ftz.f32 %f23, %f149, %f76;
@%p20 bra $L__BB92_26;
bra.uni $L__BB92_25;

$L__BB92_26:
mov.u32 %r180, _ZZ13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r181, %r180, %r176;
ld.shared.f32 %f80, [%r181];
add.ftz.f32 %f81, %f146, %f80;
ld.shared.f32 %f82, [%r30+128];
add.ftz.f32 %f83, %f23, %f82;
ld.shared.f32 %f84, [%r181+128];
add.ftz.f32 %f85, %f81, %f84;
ld.shared.f32 %f86, [%r30+256];
add.ftz.f32 %f151, %f83, %f86;
ld.shared.f32 %f87, [%r181+256];
add.ftz.f32 %f146, %f85, %f87;
bra.uni $L__BB92_27;

$L__BB92_25:
ld.shared.f32 %f77, [%r30+128];
add.ftz.f32 %f78, %f23, %f77;
ld.shared.f32 %f79, [%r30+256];
add.ftz.f32 %f151, %f78, %f79;

$L__BB92_27:
mov.b32 %r182, %f151;
mov.u32 %r183, 31;
mov.u32 %r184, 16;
mov.u32 %r185, -1;
shfl.sync.bfly.b32 %r186|%p21, %r182, %r184, %r183, %r185;
mov.b32 %f88, %r186;
add.ftz.f32 %f89, %f151, %f88;
mov.b32 %r187, %f89;
mov.u32 %r188, 8;
shfl.sync.bfly.b32 %r189|%p22, %r187, %r188, %r183, %r185;
mov.b32 %f90, %r189;
add.ftz.f32 %f91, %f89, %f90;
mov.b32 %r190, %f91;
mov.u32 %r191, 4;
shfl.sync.bfly.b32 %r192|%p23, %r190, %r191, %r183, %r185;
mov.b32 %f92, %r192;
add.ftz.f32 %f93, %f91, %f92;
mov.b32 %r193, %f93;
mov.u32 %r194, 2;
shfl.sync.bfly.b32 %r195|%p24, %r193, %r194, %r183, %r185;
mov.b32 %f94, %r195;
add.ftz.f32 %f29, %f93, %f94;
mov.b32 %r196, %f29;
mov.u32 %r197, 1;
shfl.sync.bfly.b32 %r31|%p1, %r196, %r197, %r183, %r185;
setp.eq.s64 %p25, %rd23, 0;
@%p25 bra $L__BB92_29;

mov.b32 %r198, %f146;
shfl.sync.bfly.b32 %r202|%p26, %r198, %r184, %r183, %r185;
mov.b32 %f95, %r202;
add.ftz.f32 %f96, %f146, %f95;
mov.b32 %r203, %f96;
shfl.sync.bfly.b32 %r205|%p27, %r203, %r188, %r183, %r185;
mov.b32 %f97, %r205;
add.ftz.f32 %f98, %f96, %f97;
mov.b32 %r206, %f98;
shfl.sync.bfly.b32 %r208|%p28, %r206, %r191, %r183, %r185;
mov.b32 %f99, %r208;
add.ftz.f32 %f100, %f98, %f99;
mov.b32 %r209, %f100;
shfl.sync.bfly.b32 %r211|%p29, %r209, %r194, %r183, %r185;
mov.b32 %f101, %r211;
add.ftz.f32 %f102, %f100, %f101;
mov.b32 %r212, %f102;
shfl.sync.bfly.b32 %r214|%p30, %r212, %r197, %r183, %r185;
mov.b32 %f103, %r214;
add.ftz.f32 %f146, %f102, %f103;

$L__BB92_29:
setp.ne.s32 %p31, %r2, 0;
mov.b32 %f104, %r31;
add.ftz.f32 %f32, %f29, %f104;
@%p31 bra $L__BB92_42;

add.ftz.f32 %f153, %f138, %f32;
@%p25 bra $L__BB92_41;

add.ftz.f32 %f34, %f137, %f146;
setp.gt.s32 %p33, %r32, 2;
@%p33 bra $L__BB92_35;

setp.eq.s32 %p36, %r32, 1;
@%p36 bra $L__BB92_39;

setp.eq.s32 %p37, %r32, 2;
@%p37 bra $L__BB92_34;
bra.uni $L__BB92_40;

$L__BB92_34:
mul.ftz.f32 %f124, %f34, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f125, %f124;
add.ftz.f32 %f126, %f125, 0f3F800000;
div.approx.ftz.f32 %f127, %f34, %f126;
mul.ftz.f32 %f153, %f153, %f127;
bra.uni $L__BB92_41;

$L__BB92_35:
setp.eq.s32 %p34, %r32, 3;
@%p34 bra $L__BB92_38;

setp.ne.s32 %p35, %r32, 6;
@%p35 bra $L__BB92_40;

min.ftz.f32 %f105, %f34, %f41;
neg.ftz.f32 %f106, %f41;
min.ftz.f32 %f107, %f153, %f41;
max.ftz.f32 %f108, %f107, %f106;
mul.ftz.f32 %f109, %f105, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f110, %f109;
add.ftz.f32 %f111, %f110, 0f3F800000;
div.approx.ftz.f32 %f112, %f105, %f111;
mul.ftz.f32 %f153, %f108, %f112;
bra.uni $L__BB92_41;

$L__BB92_39:
mul.ftz.f32 %f128, %f34, 0f3F000000;
mul.ftz.f32 %f129, %f34, 0f3D372713;
fma.rn.ftz.f32 %f130, %f34, %f129, 0f3F800000;
mul.ftz.f32 %f131, %f34, 0f3F4C422A;
mul.ftz.f32 %f132, %f131, %f130;
tanh.approx.f32 %f133, %f132;
add.ftz.f32 %f134, %f133, 0f3F800000;
mul.ftz.f32 %f135, %f128, %f134;
mul.ftz.f32 %f153, %f153, %f135;
bra.uni $L__BB92_41;

$L__BB92_38:
mov.f32 %f113, 0f40E00000;
min.ftz.f32 %f114, %f34, %f113;
min.ftz.f32 %f115, %f153, %f113;
mov.f32 %f116, 0fC0E00000;
max.ftz.f32 %f117, %f115, %f116;
mul.ftz.f32 %f118, %f114, 0f3FD9DB23;
mul.ftz.f32 %f119, %f118, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f120, %f119;
add.ftz.f32 %f121, %f120, 0f3F800000;
div.approx.ftz.f32 %f122, %f114, %f121;
add.ftz.f32 %f123, %f117, 0f3F800000;
mul.ftz.f32 %f153, %f123, %f122;
bra.uni $L__BB92_41;

$L__BB92_40:
mul.ftz.f32 %f153, %f153, %f34;

$L__BB92_41:
cvta.to.global.u64 %rd99, %rd27;
mov.u32 %r218, %ctaid.y;
mov.u32 %r219, %ctaid.x;
mad.lo.s32 %r220, %r218, %r43, %r219;
mad.lo.s32 %r221, %r70, %r49, %r220;
mul.wide.u32 %rd100, %r221, 4;
add.s64 %rd101, %rd99, %rd100;
st.global.f32 [%rd101], %f153;

$L__BB92_42:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<15>;
.reg .b16 %rs<36>;
.reg .f32 %f<56>;
.reg .b32 %r<163>;
.reg .b64 %rd<75>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[384];

ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r45, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r46, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
cvta.to.global.u64 %rd1, %rd18;
cvta.to.global.u64 %rd74, %rd17;
mov.u32 %r1, %tid.y;
shl.b32 %r47, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r47, %r2;
mov.u32 %r4, %ctaid.x;
shr.u32 %r5, %r30, 5;
mov.u32 %r158, %ctaid.y;
setp.eq.s64 %p1, %rd15, 0;
@%p1 bra $L__BB93_2;

cvta.to.global.u64 %rd19, %rd15;
mul.wide.u32 %rd20, %r158, 4;
add.s64 %rd21, %rd19, %rd20;
ld.global.nc.u32 %r157, [%rd21];
mul.hi.u32 %r48, %r158, %r31;
add.s32 %r49, %r48, %r158;
shr.u32 %r50, %r49, %r32;
mul.lo.s32 %r51, %r50, %r33;
sub.s32 %r158, %r158, %r51;
bra.uni $L__BB93_3;

$L__BB93_2:
mul.hi.u32 %r52, %r158, %r35;
add.s32 %r53, %r52, %r158;
shr.u32 %r157, %r53, %r36;

$L__BB93_3:
mov.u32 %r12, %ctaid.z;
mul.hi.u32 %r54, %r12, %r41;
add.s32 %r55, %r54, %r12;
shr.u32 %r56, %r55, %r42;
mul.lo.s32 %r57, %r12, %r45;
cvt.u64.u32 %rd22, %r57;
mul.lo.s32 %r58, %r158, %r39;
cvt.u64.u32 %rd23, %r58;
add.s64 %rd3, %rd23, %rd22;
mul.lo.s32 %r13, %r56, %r44;
mul.lo.s32 %r14, %r157, %r38;
mul.lo.s32 %r15, %r4, %r34;
shr.s32 %r59, %r3, 31;
shr.u32 %r60, %r59, 30;
add.s32 %r61, %r3, %r60;
shr.s32 %r161, %r61, 2;
setp.ge.s32 %p2, %r161, %r5;
mov.f32 %f55, 0f00000000;
@%p2 bra $L__BB93_10;

and.b32 %r65, %r61, -4;
sub.s32 %r17, %r3, %r65;
shl.b32 %r18, %r17, 1;
not.b32 %r66, %r161;
add.s32 %r19, %r5, %r66;
shr.u32 %r67, %r19, 5;
add.s32 %r68, %r67, 1;
and.b32 %r160, %r68, 3;
setp.eq.s32 %p3, %r160, 0;
mov.f32 %f55, 0f00000000;
@%p3 bra $L__BB93_7;

shl.b32 %r69, %r17, 2;
or.b32 %r70, %r69, 1;
sub.s32 %r71, %r70, %r18;
cvt.s64.s32 %rd4, %r71;
add.s32 %r72, %r161, %r13;
add.s32 %r73, %r72, %r14;
add.s32 %r74, %r73, %r15;
mul.wide.s32 %rd24, %r74, 34;
add.s64 %rd73, %rd74, %rd24;
mul.wide.s32 %rd25, %r69, 2;
or.b64 %rd6, %rd25, 2;
mov.f32 %f55, 0f00000000;
shl.b64 %rd34, %rd4, 2;

$L__BB93_6:
.pragma "nounroll";
cvt.u64.u32 %rd26, %r161;
add.s64 %rd27, %rd3, %rd26;
mul.lo.s64 %rd28, %rd27, 36;
add.s64 %rd29, %rd1, %rd28;
add.s64 %rd30, %rd29, 4;
add.s64 %rd31, %rd73, %rd6;
ld.global.nc.u16 %rs4, [%rd31];
ld.global.nc.u16 %rs5, [%rd31+2];
mov.b32 %r77, {%rs4, %rs5};
mul.wide.s32 %rd32, %r18, 4;
add.s64 %rd33, %rd30, %rd32;
ld.global.nc.u32 %r78, [%rd33];
ld.global.nc.u16 %rs6, [%rd31+4];
ld.global.nc.u16 %rs7, [%rd31+6];
mov.b32 %r81, {%rs6, %rs7};
add.s64 %rd35, %rd30, %rd34;
ld.global.nc.u32 %r82, [%rd35];
ld.global.nc.u16 %rs1, [%rd73];

{ cvt.f32.f16 %f13, %rs1;}


ld.global.nc.u32 %r75, [%rd29];

{.reg .f16 low,high;
mov.b32 {low,high}, %r75;
mov.b16 %rs2, low;}


{ cvt.f32.f16 %f14, %rs2;}


mov.u32 %r79, 0;

dp4a.s32.s32 %r76, %r77, %r78, %r79;


dp4a.s32.s32 %r80, %r81, %r82, %r76;

mul.ftz.f32 %f15, %f13, %f14;
cvt.rn.f32.s32 %f16, %r80;
fma.rn.ftz.f32 %f55, %f15, %f16, %f55;
add.s32 %r161, %r161, 32;
add.s64 %rd73, %rd73, 1088;
add.s32 %r160, %r160, -1;
setp.ne.s32 %p4, %r160, 0;
@%p4 bra $L__BB93_6;

$L__BB93_7:
setp.lt.u32 %p5, %r19, 96;
@%p5 bra $L__BB93_10;

shl.b32 %r85, %r17, 2;
or.b32 %r86, %r85, 1;
cvt.s64.s32 %rd9, %r18;
sub.s32 %r87, %r86, %r18;
cvt.s64.s32 %rd10, %r87;
add.s32 %r88, %r161, %r13;
add.s32 %r89, %r88, %r14;
add.s32 %r90, %r89, %r15;
mul.wide.s32 %rd36, %r85, 2;
mul.wide.s32 %rd11, %r90, 34;
add.s64 %rd37, %rd11, %rd36;
add.s64 %rd12, %rd37, 4;
shl.b64 %rd44, %rd9, 2;
shl.b64 %rd46, %rd10, 2;

$L__BB93_9:
cvt.u64.u32 %rd38, %r161;
add.s64 %rd39, %rd3, %rd38;
mul.lo.s64 %rd40, %rd39, 36;
add.s64 %rd41, %rd1, %rd40;
add.s64 %rd42, %rd41, 4;
add.s64 %rd43, %rd74, %rd12;
ld.global.nc.u16 %rs20, [%rd43+-2];
ld.global.nc.u16 %rs21, [%rd43];
mov.b32 %r93, {%rs20, %rs21};
add.s64 %rd45, %rd42, %rd44;
ld.global.nc.u32 %r94, [%rd45];
ld.global.nc.u16 %rs22, [%rd43+2];
ld.global.nc.u16 %rs23, [%rd43+4];
mov.b32 %r97, {%rs22, %rs23};
add.s64 %rd47, %rd42, %rd46;
ld.global.nc.u32 %r98, [%rd47];
add.s64 %rd48, %rd74, %rd11;
ld.global.nc.u16 %rs8, [%rd48];

{ cvt.f32.f16 %f17, %rs8;}


ld.global.nc.u32 %r91, [%rd41];

{.reg .f16 low,high;
mov.b32 {low,high}, %r91;
mov.b16 %rs9, low;}


{ cvt.f32.f16 %f18, %rs9;}


mov.u32 %r122, 0;

dp4a.s32.s32 %r92, %r93, %r94, %r122;


dp4a.s32.s32 %r96, %r97, %r98, %r92;

mul.ftz.f32 %f25, %f17, %f18;
cvt.rn.f32.s32 %f26, %r96;
fma.rn.ftz.f32 %f27, %f25, %f26, %f55;
add.s32 %r127, %r161, 32;
cvt.u64.u32 %rd49, %r127;
add.s64 %rd50, %rd3, %rd49;
mul.lo.s64 %rd51, %rd50, 36;
add.s64 %rd52, %rd1, %rd51;
add.s64 %rd53, %rd52, 4;
ld.global.nc.u16 %rs24, [%rd43+1086];
ld.global.nc.u16 %rs25, [%rd43+1088];
mov.b32 %r102, {%rs24, %rs25};
add.s64 %rd54, %rd53, %rd44;
ld.global.nc.u32 %r103, [%rd54];
ld.global.nc.u16 %rs26, [%rd43+1090];
ld.global.nc.u16 %rs27, [%rd43+1092];
mov.b32 %r106, {%rs26, %rs27};
add.s64 %rd55, %rd53, %rd46;
ld.global.nc.u32 %r107, [%rd55];
ld.global.nc.u16 %rs11, [%rd48+1088];

{ cvt.f32.f16 %f19, %rs11;}


ld.global.nc.u32 %r100, [%rd52];

{.reg .f16 low,high;
mov.b32 {low,high}, %r100;
mov.b16 %rs12, low;}


{ cvt.f32.f16 %f20, %rs12;}



dp4a.s32.s32 %r101, %r102, %r103, %r122;


dp4a.s32.s32 %r105, %r106, %r107, %r101;

mul.ftz.f32 %f28, %f19, %f20;
cvt.rn.f32.s32 %f29, %r105;
fma.rn.ftz.f32 %f30, %f28, %f29, %f27;
add.s32 %r128, %r161, 64;
cvt.u64.u32 %rd56, %r128;
add.s64 %rd57, %rd3, %rd56;
mul.lo.s64 %rd58, %rd57, 36;
add.s64 %rd59, %rd1, %rd58;
add.s64 %rd60, %rd59, 4;
ld.global.nc.u16 %rs28, [%rd43+2174];
ld.global.nc.u16 %rs29, [%rd43+2176];
mov.b32 %r111, {%rs28, %rs29};
add.s64 %rd61, %rd60, %rd44;
ld.global.nc.u32 %r112, [%rd61];
ld.global.nc.u16 %rs30, [%rd43+2178];
ld.global.nc.u16 %rs31, [%rd43+2180];
mov.b32 %r115, {%rs30, %rs31};
add.s64 %rd62, %rd60, %rd46;
ld.global.nc.u32 %r116, [%rd62];
ld.global.nc.u16 %rs14, [%rd48+2176];

{ cvt.f32.f16 %f21, %rs14;}


ld.global.nc.u32 %r109, [%rd59];

{.reg .f16 low,high;
mov.b32 {low,high}, %r109;
mov.b16 %rs15, low;}


{ cvt.f32.f16 %f22, %rs15;}



dp4a.s32.s32 %r110, %r111, %r112, %r122;


dp4a.s32.s32 %r114, %r115, %r116, %r110;

mul.ftz.f32 %f31, %f21, %f22;
cvt.rn.f32.s32 %f32, %r114;
fma.rn.ftz.f32 %f33, %f31, %f32, %f30;
add.s32 %r129, %r161, 96;
cvt.u64.u32 %rd63, %r129;
add.s64 %rd64, %rd3, %rd63;
mul.lo.s64 %rd65, %rd64, 36;
add.s64 %rd66, %rd1, %rd65;
add.s64 %rd67, %rd66, 4;
ld.global.nc.u16 %rs32, [%rd43+3262];
ld.global.nc.u16 %rs33, [%rd43+3264];
mov.b32 %r120, {%rs32, %rs33};
add.s64 %rd68, %rd67, %rd44;
ld.global.nc.u32 %r121, [%rd68];
ld.global.nc.u16 %rs34, [%rd43+3266];
ld.global.nc.u16 %rs35, [%rd43+3268];
mov.b32 %r124, {%rs34, %rs35};
add.s64 %rd69, %rd67, %rd46;
ld.global.nc.u32 %r125, [%rd69];
ld.global.nc.u16 %rs17, [%rd48+3264];

{ cvt.f32.f16 %f23, %rs17;}


ld.global.nc.u32 %r118, [%rd66];

{.reg .f16 low,high;
mov.b32 {low,high}, %r118;
mov.b16 %rs18, low;}


{ cvt.f32.f16 %f24, %rs18;}



dp4a.s32.s32 %r119, %r120, %r121, %r122;


dp4a.s32.s32 %r123, %r124, %r125, %r119;

mul.ftz.f32 %f34, %f23, %f24;
cvt.rn.f32.s32 %f35, %r123;
fma.rn.ftz.f32 %f55, %f34, %f35, %f33;
add.s64 %rd74, %rd74, 4352;
add.s32 %r161, %r161, 128;
setp.lt.s32 %p6, %r161, %r5;
@%p6 bra $L__BB93_9;

$L__BB93_10:
setp.eq.s32 %p7, %r1, 0;
@%p7 bra $L__BB93_12;

shl.b32 %r130, %r1, 7;
mov.u32 %r131, _ZZ13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r132, %r131, %r130;
shl.b32 %r133, %r2, 2;
add.s32 %r134, %r132, %r133;
st.shared.f32 [%r134+-128], %f55;

$L__BB93_12:
setp.ne.s32 %p8, %r1, 0;
bar.sync 0;
@%p8 bra $L__BB93_15;

mov.u32 %r135, %ctaid.y;
mad.lo.s32 %r136, %r135, %r40, %r4;
mad.lo.s32 %r29, %r12, %r46, %r136;
shl.b32 %r137, %r2, 2;
mov.u32 %r138, _ZZ13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r139, %r138, %r137;
mov.u32 %r140, 2;
ld.shared.f32 %f36, [%r139];
add.ftz.f32 %f37, %f55, %f36;
ld.shared.f32 %f38, [%r139+128];
add.ftz.f32 %f39, %f37, %f38;
ld.shared.f32 %f40, [%r139+256];
add.ftz.f32 %f41, %f39, %f40;
mov.b32 %r141, %f41;
mov.u32 %r142, 31;
mov.u32 %r143, 16;
mov.u32 %r144, -1;
shfl.sync.bfly.b32 %r145|%p9, %r141, %r143, %r142, %r144;
mov.b32 %f42, %r145;
add.ftz.f32 %f43, %f41, %f42;
mov.b32 %r146, %f43;
mov.u32 %r147, 8;
shfl.sync.bfly.b32 %r148|%p10, %r146, %r147, %r142, %r144;
mov.b32 %f44, %r148;
add.ftz.f32 %f45, %f43, %f44;
mov.b32 %r149, %f45;
mov.u32 %r150, 4;
shfl.sync.bfly.b32 %r151|%p11, %r149, %r150, %r142, %r144;
mov.b32 %f46, %r151;
add.ftz.f32 %f47, %f45, %f46;
mov.b32 %r152, %f47;
shfl.sync.bfly.b32 %r153|%p12, %r152, %r140, %r142, %r144;
mov.b32 %f48, %r153;
add.ftz.f32 %f49, %f47, %f48;
mov.b32 %r154, %f49;
mov.u32 %r155, 1;
shfl.sync.bfly.b32 %r156|%p13, %r154, %r155, %r142, %r144;
mov.b32 %f50, %r156;
add.ftz.f32 %f8, %f49, %f50;
setp.ne.s32 %p14, %r2, 0;
@%p14 bra $L__BB93_15;

cvta.to.global.u64 %rd70, %rd16;
mul.wide.u32 %rd71, %r29, 4;
add.s64 %rd72, %rd70, %rd71;
st.global.f32 [%rd72], %f8;

$L__BB93_15:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type8ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type8ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type8ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type8ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type8ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type8ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type8ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type8ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type8ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<35>;
.reg .b16 %rs<21>;
.reg .f32 %f<109>;
.reg .b32 %r<182>;
.reg .b64 %rd<53>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type8ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type8ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type8ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type8ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r14, [_Z13mul_mat_vec_qIL9ggml_type8ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r15, [_Z13mul_mat_vec_qIL9ggml_type8ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r16, [_Z13mul_mat_vec_qIL9ggml_type8ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r17, [_Z13mul_mat_vec_qIL9ggml_type8ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type8ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type8ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type8ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type8ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type8ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type8ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type8ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type8ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r19, [_Z13mul_mat_vec_qIL9ggml_type8ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r18, [_Z13mul_mat_vec_qIL9ggml_type8ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
cvta.to.global.u64 %rd1, %rd17;
mov.u32 %r1, %tid.y;
shl.b32 %r30, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r30, %r2;
mov.u32 %r31, %ctaid.x;
shl.b32 %r4, %r31, 1;
mov.u32 %r5, %ctaid.y;
shr.s32 %r32, %r3, 31;
shr.u32 %r33, %r32, 30;
add.s32 %r34, %r3, %r33;
shr.s32 %r181, %r34, 2;
shr.u32 %r35, %r14, 5;
setp.ge.s32 %p3, %r181, %r35;
mov.f32 %f105, 0f00000000;
mov.f32 %f106, %f105;
mov.f32 %f107, %f105;
mov.f32 %f108, %f105;
@%p3 bra $L__BB94_3;

mov.u32 %r36, %ctaid.z;
mul.lo.s32 %r37, %r5, %r22;
cvt.u64.u32 %rd18, %r37;
mul.lo.s32 %r38, %r36, %r28;
cvt.u64.u32 %rd19, %r38;
add.s64 %rd2, %rd19, %rd18;
mul.hi.u32 %r39, %r36, %r24;
add.s32 %r40, %r39, %r36;
shr.u32 %r41, %r40, %r25;
mul.lo.s32 %r42, %r41, %r27;
mul.hi.u32 %r43, %r5, %r18;
add.s32 %r44, %r43, %r5;
shr.u32 %r45, %r44, %r19;
mul.lo.s32 %r46, %r45, %r21;
cvta.to.global.u64 %rd52, %rd15;
and.b32 %r50, %r34, -4;
sub.s32 %r51, %r3, %r50;
shl.b32 %r52, %r51, 1;
shl.b32 %r53, %r51, 2;
or.b32 %r54, %r53, 1;
cvt.s64.s32 %rd4, %r52;
sub.s32 %r55, %r54, %r52;
cvt.s64.s32 %rd5, %r55;
add.s32 %r180, %r181, %r16;
add.s32 %r56, %r181, %r42;
add.s32 %r57, %r56, %r46;
add.s32 %r58, %r4, 1;
mad.lo.s32 %r59, %r15, %r58, %r57;
mul.wide.s32 %rd6, %r59, 34;
mad.lo.s32 %r60, %r4, %r15, %r181;
add.s32 %r61, %r60, %r42;
add.s32 %r62, %r61, %r46;
mul.wide.s32 %rd7, %r62, 34;
mul.wide.s32 %rd20, %r53, 2;
add.s64 %rd21, %rd7, %rd20;
add.s64 %rd8, %rd21, 4;
add.s64 %rd22, %rd6, %rd20;
add.s64 %rd9, %rd22, 2;
cvta.to.global.u64 %rd10, %rd16;
mov.f32 %f105, 0f00000000;
shl.b64 %rd29, %rd4, 2;
shl.b64 %rd31, %rd5, 2;

$L__BB94_2:
cvt.u64.u32 %rd23, %r181;
add.s64 %rd24, %rd2, %rd23;
mul.lo.s64 %rd25, %rd24, 36;
add.s64 %rd26, %rd10, %rd25;
add.s64 %rd27, %rd26, 4;
add.s64 %rd28, %rd52, %rd8;
ld.global.nc.u16 %rs13, [%rd28+-2];
ld.global.nc.u16 %rs14, [%rd28];
mov.b32 %r65, {%rs13, %rs14};
add.s64 %rd30, %rd27, %rd29;
ld.global.nc.u32 %r66, [%rd30];
ld.global.nc.u16 %rs15, [%rd28+2];
ld.global.nc.u16 %rs16, [%rd28+4];
mov.b32 %r69, {%rs15, %rs16};
add.s64 %rd32, %rd27, %rd31;
ld.global.nc.u32 %r70, [%rd32];
add.s64 %rd33, %rd52, %rd7;
ld.global.nc.u16 %rs1, [%rd33];

{ cvt.f32.f16 %f25, %rs1;}


ld.global.nc.u32 %r63, [%rd26];

{.reg .f16 low,high;
mov.b32 {low,high}, %r63;
mov.b16 %rs2, low;}


{ cvt.f32.f16 %f26, %rs2;}


mov.u32 %r94, 0;

dp4a.s32.s32 %r64, %r65, %r66, %r94;


dp4a.s32.s32 %r68, %r69, %r70, %r64;

mul.ftz.f32 %f33, %f25, %f26;
cvt.rn.f32.s32 %f34, %r68;
fma.rn.ftz.f32 %f108, %f33, %f34, %f108;
add.s64 %rd34, %rd52, %rd9;
ld.global.nc.u16 %rs17, [%rd34];
ld.global.nc.u16 %rs18, [%rd34+2];
mov.b32 %r74, {%rs17, %rs18};
ld.global.nc.u16 %rs19, [%rd34+4];
ld.global.nc.u16 %rs20, [%rd34+6];
mov.b32 %r78, {%rs19, %rs20};
add.s64 %rd35, %rd52, %rd6;
ld.global.nc.u16 %rs4, [%rd35];

{ cvt.f32.f16 %f27, %rs4;}



{.reg .f16 low,high;
mov.b32 {low,high}, %r63;
mov.b16 %rs5, low;}


{ cvt.f32.f16 %f28, %rs5;}



dp4a.s32.s32 %r73, %r74, %r66, %r94;


dp4a.s32.s32 %r77, %r78, %r70, %r73;

mul.ftz.f32 %f35, %f27, %f28;
cvt.rn.f32.s32 %f36, %r77;
fma.rn.ftz.f32 %f107, %f35, %f36, %f107;
cvt.u64.u32 %rd36, %r180;
add.s64 %rd37, %rd2, %rd36;
mul.lo.s64 %rd38, %rd37, 36;
add.s64 %rd39, %rd10, %rd38;
add.s64 %rd40, %rd39, 4;
add.s64 %rd41, %rd40, %rd29;
ld.global.nc.u32 %r84, [%rd41];
add.s64 %rd42, %rd40, %rd31;
ld.global.nc.u32 %r88, [%rd42];

{ cvt.f32.f16 %f29, %rs1;}


ld.global.nc.u32 %r81, [%rd39];

{.reg .f16 low,high;
mov.b32 {low,high}, %r81;
mov.b16 %rs8, low;}


{ cvt.f32.f16 %f30, %rs8;}



dp4a.s32.s32 %r82, %r65, %r84, %r94;


dp4a.s32.s32 %r86, %r69, %r88, %r82;

mul.ftz.f32 %f37, %f29, %f30;
cvt.rn.f32.s32 %f38, %r86;
fma.rn.ftz.f32 %f106, %f37, %f38, %f106;

{ cvt.f32.f16 %f31, %rs4;}



{.reg .f16 low,high;
mov.b32 {low,high}, %r81;
mov.b16 %rs11, low;}


{ cvt.f32.f16 %f32, %rs11;}



dp4a.s32.s32 %r91, %r74, %r84, %r94;


dp4a.s32.s32 %r95, %r78, %r88, %r91;

mul.ftz.f32 %f39, %f31, %f32;
cvt.rn.f32.s32 %f40, %r95;
fma.rn.ftz.f32 %f105, %f39, %f40, %f105;
add.s32 %r180, %r180, 32;
add.s64 %rd52, %rd52, 1088;
add.s32 %r181, %r181, 32;
setp.lt.s32 %p4, %r181, %r35;
@%p4 bra $L__BB94_2;

$L__BB94_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB94_5;

shl.b32 %r100, %r1, 9;
mov.u32 %r101, _ZZ13mul_mat_vec_qIL9ggml_type8ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r102, %r101, %r100;
shl.b32 %r103, %r2, 2;
add.s32 %r104, %r102, %r103;
st.shared.f32 [%r104+-512], %f108;
st.shared.f32 [%r104+-384], %f107;
st.shared.f32 [%r104+-256], %f106;
st.shared.f32 [%r104+-128], %f105;

$L__BB94_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB94_14;

mov.u32 %r105, %ctaid.z;
mad.lo.s32 %r106, %r5, %r23, %r4;
mad.lo.s32 %r107, %r105, %r29, %r106;
cvt.u64.u32 %rd13, %r107;
shl.b32 %r108, %r2, 2;
mov.u32 %r109, _ZZ13mul_mat_vec_qIL9ggml_type8ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r12, %r109, %r108;
mov.u32 %r110, 2;
ld.shared.f32 %f41, [%r12];
add.ftz.f32 %f42, %f41, %f108;
ld.shared.f32 %f43, [%r12+512];
add.ftz.f32 %f44, %f43, %f42;
ld.shared.f32 %f45, [%r12+1024];
add.ftz.f32 %f46, %f45, %f44;
mov.b32 %r111, %f46;
mov.u32 %r112, 31;
mov.u32 %r113, 16;
mov.u32 %r114, -1;
shfl.sync.bfly.b32 %r115|%p7, %r111, %r113, %r112, %r114;
mov.b32 %f47, %r115;
add.ftz.f32 %f48, %f46, %f47;
mov.b32 %r116, %f48;
mov.u32 %r117, 8;
shfl.sync.bfly.b32 %r118|%p8, %r116, %r117, %r112, %r114;
mov.b32 %f49, %r118;
add.ftz.f32 %f50, %f48, %f49;
mov.b32 %r119, %f50;
mov.u32 %r120, 4;
shfl.sync.bfly.b32 %r121|%p9, %r119, %r120, %r112, %r114;
mov.b32 %f51, %r121;
add.ftz.f32 %f52, %f50, %f51;
mov.b32 %r122, %f52;
shfl.sync.bfly.b32 %r123|%p10, %r122, %r110, %r112, %r114;
mov.b32 %f53, %r123;
add.ftz.f32 %f54, %f52, %f53;
mov.b32 %r124, %f54;
mov.u32 %r125, 1;
shfl.sync.bfly.b32 %r126|%p11, %r124, %r125, %r112, %r114;
mov.b32 %f55, %r126;
add.ftz.f32 %f13, %f54, %f55;
setp.lt.u32 %p12, %r4, %r17;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
mul.wide.u32 %rd43, %r107, 4;
add.s64 %rd14, %rd1, %rd43;
not.pred %p14, %p1;
@%p14 bra $L__BB94_8;

st.global.f32 [%rd14], %f13;

$L__BB94_8:
ld.shared.f32 %f56, [%r12+128];
add.ftz.f32 %f57, %f56, %f107;
ld.shared.f32 %f58, [%r12+640];
add.ftz.f32 %f59, %f58, %f57;
ld.shared.f32 %f60, [%r12+1152];
add.ftz.f32 %f61, %f60, %f59;
mov.b32 %r127, %f61;
shfl.sync.bfly.b32 %r131|%p15, %r127, %r113, %r112, %r114;
mov.b32 %f62, %r131;
add.ftz.f32 %f63, %f61, %f62;
mov.b32 %r132, %f63;
shfl.sync.bfly.b32 %r134|%p16, %r132, %r117, %r112, %r114;
mov.b32 %f64, %r134;
add.ftz.f32 %f65, %f63, %f64;
mov.b32 %r135, %f65;
shfl.sync.bfly.b32 %r137|%p17, %r135, %r120, %r112, %r114;
mov.b32 %f66, %r137;
add.ftz.f32 %f67, %f65, %f66;
mov.b32 %r138, %f67;
shfl.sync.bfly.b32 %r140|%p18, %r138, %r110, %r112, %r114;
mov.b32 %f68, %r140;
add.ftz.f32 %f69, %f67, %f68;
mov.b32 %r141, %f69;
shfl.sync.bfly.b32 %r143|%p19, %r141, %r125, %r112, %r114;
mov.b32 %f70, %r143;
add.ftz.f32 %f14, %f69, %f70;
add.s32 %r144, %r4, 1;
setp.lt.u32 %p20, %r144, %r17;
setp.eq.s32 %p21, %r2, 1;
and.pred %p2, %p21, %p20;
not.pred %p22, %p2;
@%p22 bra $L__BB94_10;

st.global.f32 [%rd14+4], %f14;

$L__BB94_10:
ld.shared.f32 %f71, [%r12+256];
add.ftz.f32 %f72, %f71, %f106;
ld.shared.f32 %f73, [%r12+768];
add.ftz.f32 %f74, %f73, %f72;
ld.shared.f32 %f75, [%r12+1280];
add.ftz.f32 %f76, %f75, %f74;
mov.b32 %r145, %f76;
mov.u32 %r146, 31;
mov.u32 %r147, 16;
mov.u32 %r148, -1;
shfl.sync.bfly.b32 %r149|%p23, %r145, %r147, %r146, %r148;
mov.b32 %f77, %r149;
add.ftz.f32 %f78, %f76, %f77;
mov.b32 %r150, %f78;
mov.u32 %r151, 8;
shfl.sync.bfly.b32 %r152|%p24, %r150, %r151, %r146, %r148;
mov.b32 %f79, %r152;
add.ftz.f32 %f80, %f78, %f79;
mov.b32 %r153, %f80;
mov.u32 %r154, 4;
shfl.sync.bfly.b32 %r155|%p25, %r153, %r154, %r146, %r148;
mov.b32 %f81, %r155;
add.ftz.f32 %f82, %f80, %f81;
mov.b32 %r156, %f82;
mov.u32 %r157, 2;
shfl.sync.bfly.b32 %r158|%p26, %r156, %r157, %r146, %r148;
mov.b32 %f83, %r158;
add.ftz.f32 %f84, %f82, %f83;
mov.b32 %r159, %f84;
mov.u32 %r160, 1;
shfl.sync.bfly.b32 %r161|%p27, %r159, %r160, %r146, %r148;
mov.b32 %f85, %r161;
add.ftz.f32 %f15, %f84, %f85;
@%p14 bra $L__BB94_12;

cvt.u64.u32 %rd44, %r17;
add.s64 %rd45, %rd44, %rd13;
shl.b64 %rd46, %rd45, 2;
add.s64 %rd47, %rd1, %rd46;
st.global.f32 [%rd47], %f15;

$L__BB94_12:
ld.shared.f32 %f86, [%r12+384];
add.ftz.f32 %f87, %f86, %f105;
ld.shared.f32 %f88, [%r12+896];
add.ftz.f32 %f89, %f88, %f87;
ld.shared.f32 %f90, [%r12+1408];
add.ftz.f32 %f91, %f90, %f89;
mov.b32 %r162, %f91;
shfl.sync.bfly.b32 %r166|%p29, %r162, %r147, %r146, %r148;
mov.b32 %f92, %r166;
add.ftz.f32 %f93, %f91, %f92;
mov.b32 %r167, %f93;
shfl.sync.bfly.b32 %r169|%p30, %r167, %r151, %r146, %r148;
mov.b32 %f94, %r169;
add.ftz.f32 %f95, %f93, %f94;
mov.b32 %r170, %f95;
shfl.sync.bfly.b32 %r172|%p31, %r170, %r154, %r146, %r148;
mov.b32 %f96, %r172;
add.ftz.f32 %f97, %f95, %f96;
mov.b32 %r173, %f97;
shfl.sync.bfly.b32 %r175|%p32, %r173, %r157, %r146, %r148;
mov.b32 %f98, %r175;
add.ftz.f32 %f99, %f97, %f98;
mov.b32 %r176, %f99;
shfl.sync.bfly.b32 %r178|%p33, %r176, %r160, %r146, %r148;
mov.b32 %f100, %r178;
add.ftz.f32 %f16, %f99, %f100;
@%p22 bra $L__BB94_14;

add.s32 %r179, %r17, 1;
cvt.u64.u32 %rd48, %r179;
add.s64 %rd49, %rd48, %rd13;
shl.b64 %rd50, %rd49, 2;
add.s64 %rd51, %rd1, %rd50;
st.global.f32 [%rd51], %f16;

$L__BB94_14:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type8ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type8ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type8ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type8ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type8ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type8ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type8ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type8ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type8ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<47>;
.reg .b16 %rs<27>;
.reg .f32 %f<163>;
.reg .b32 %r<254>;
.reg .b64 %rd<67>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type8ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[2304];

ld.param.u64 %rd14, [_Z13mul_mat_vec_qIL9ggml_type8ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type8ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type8ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r13, [_Z13mul_mat_vec_qIL9ggml_type8ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r14, [_Z13mul_mat_vec_qIL9ggml_type8ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r15, [_Z13mul_mat_vec_qIL9ggml_type8ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r16, [_Z13mul_mat_vec_qIL9ggml_type8ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type8ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type8ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type8ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type8ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type8ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type8ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type8ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type8ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r18, [_Z13mul_mat_vec_qIL9ggml_type8ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r17, [_Z13mul_mat_vec_qIL9ggml_type8ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r29, %tid.y;
shl.b32 %r30, %r29, 5;
mov.u32 %r31, %tid.x;
add.s32 %r32, %r30, %r31;
mov.u32 %r33, %ctaid.x;
shl.b32 %r1, %r33, 1;
shr.s32 %r34, %r32, 31;
shr.u32 %r35, %r34, 30;
add.s32 %r36, %r32, %r35;
shr.s32 %r253, %r36, 2;
shr.u32 %r37, %r13, 5;
setp.ge.s32 %p3, %r253, %r37;
mov.f32 %f157, 0f00000000;
mov.f32 %f158, %f157;
mov.f32 %f159, %f157;
mov.f32 %f160, %f157;
mov.f32 %f161, %f157;
mov.f32 %f162, %f157;
@%p3 bra $L__BB95_3;

mov.u32 %r38, %ctaid.y;
mov.u32 %r39, %ctaid.z;
mul.hi.u32 %r40, %r39, %r23;
add.s32 %r41, %r40, %r39;
shr.u32 %r42, %r41, %r24;
mul.lo.s32 %r43, %r42, %r26;
mul.hi.u32 %r44, %r38, %r17;
add.s32 %r45, %r44, %r38;
shr.u32 %r46, %r45, %r18;
mul.lo.s32 %r47, %r46, %r20;
cvta.to.global.u64 %rd66, %rd14;
and.b32 %r55, %r36, -4;
sub.s32 %r56, %r32, %r55;
shl.b32 %r57, %r56, 1;
shl.b32 %r58, %r56, 2;
or.b32 %r59, %r58, 1;
cvt.s64.s32 %rd2, %r57;
sub.s32 %r60, %r59, %r57;
cvt.s64.s32 %rd3, %r60;
shl.b32 %r61, %r15, 1;
add.s32 %r252, %r253, %r61;
add.s32 %r251, %r253, %r15;
add.s32 %r62, %r253, %r43;
add.s32 %r63, %r62, %r47;
add.s32 %r64, %r1, 1;
mad.lo.s32 %r65, %r14, %r64, %r63;
mul.wide.s32 %rd4, %r65, 34;
mad.lo.s32 %r66, %r1, %r14, %r253;
add.s32 %r67, %r66, %r43;
add.s32 %r68, %r67, %r47;
mul.wide.s32 %rd5, %r68, 34;
mul.wide.s32 %rd17, %r58, 2;
add.s64 %rd18, %rd5, %rd17;
add.s64 %rd6, %rd18, 4;
add.s64 %rd19, %rd4, %rd17;
add.s64 %rd7, %rd19, 2;
cvta.to.global.u64 %rd8, %rd15;
mov.f32 %f157, 0f00000000;
shl.b64 %rd29, %rd2, 2;
shl.b64 %rd31, %rd3, 2;

$L__BB95_2:
mul.lo.s32 %r124, %r39, %r27;
cvt.u64.u32 %rd20, %r124;
mul.lo.s32 %r126, %r38, %r21;
cvt.u64.u32 %rd21, %r126;
add.s64 %rd22, %rd20, %rd21;
cvt.u64.u32 %rd23, %r253;
add.s64 %rd24, %rd22, %rd23;
mul.lo.s64 %rd25, %rd24, 36;
add.s64 %rd26, %rd8, %rd25;
add.s64 %rd27, %rd26, 4;
add.s64 %rd28, %rd66, %rd6;
ld.global.nc.u16 %rs19, [%rd28+-2];
ld.global.nc.u16 %rs20, [%rd28];
mov.b32 %r71, {%rs19, %rs20};
add.s64 %rd30, %rd27, %rd29;
ld.global.nc.u32 %r72, [%rd30];
ld.global.nc.u16 %rs21, [%rd28+2];
ld.global.nc.u16 %rs22, [%rd28+4];
mov.b32 %r75, {%rs21, %rs22};
add.s64 %rd32, %rd27, %rd31;
ld.global.nc.u32 %r76, [%rd32];
add.s64 %rd33, %rd66, %rd5;
ld.global.nc.u16 %rs1, [%rd33];

{ cvt.f32.f16 %f37, %rs1;}


ld.global.nc.u32 %r69, [%rd26];

{.reg .f16 low,high;
mov.b32 {low,high}, %r69;
mov.b16 %rs2, low;}


{ cvt.f32.f16 %f38, %rs2;}


mov.u32 %r118, 0;

dp4a.s32.s32 %r70, %r71, %r72, %r118;


dp4a.s32.s32 %r74, %r75, %r76, %r70;

mul.ftz.f32 %f49, %f37, %f38;
cvt.rn.f32.s32 %f50, %r74;
fma.rn.ftz.f32 %f162, %f49, %f50, %f162;
add.s64 %rd34, %rd66, %rd7;
ld.global.nc.u16 %rs23, [%rd34];
ld.global.nc.u16 %rs24, [%rd34+2];
mov.b32 %r80, {%rs23, %rs24};
ld.global.nc.u16 %rs25, [%rd34+4];
ld.global.nc.u16 %rs26, [%rd34+6];
mov.b32 %r84, {%rs25, %rs26};
add.s64 %rd35, %rd66, %rd4;
ld.global.nc.u16 %rs4, [%rd35];

{ cvt.f32.f16 %f39, %rs4;}



{.reg .f16 low,high;
mov.b32 {low,high}, %r69;
mov.b16 %rs5, low;}


{ cvt.f32.f16 %f40, %rs5;}



dp4a.s32.s32 %r79, %r80, %r72, %r118;


dp4a.s32.s32 %r83, %r84, %r76, %r79;

mul.ftz.f32 %f51, %f39, %f40;
cvt.rn.f32.s32 %f52, %r83;
fma.rn.ftz.f32 %f161, %f51, %f52, %f161;
cvt.u64.u32 %rd36, %r251;
add.s64 %rd37, %rd22, %rd36;
mul.lo.s64 %rd38, %rd37, 36;
add.s64 %rd39, %rd8, %rd38;
add.s64 %rd40, %rd39, 4;
add.s64 %rd41, %rd40, %rd29;
ld.global.nc.u32 %r90, [%rd41];
add.s64 %rd42, %rd40, %rd31;
ld.global.nc.u32 %r94, [%rd42];

{ cvt.f32.f16 %f41, %rs1;}


ld.global.nc.u32 %r87, [%rd39];

{.reg .f16 low,high;
mov.b32 {low,high}, %r87;
mov.b16 %rs8, low;}


{ cvt.f32.f16 %f42, %rs8;}



dp4a.s32.s32 %r88, %r71, %r90, %r118;


dp4a.s32.s32 %r92, %r75, %r94, %r88;

mul.ftz.f32 %f53, %f41, %f42;
cvt.rn.f32.s32 %f54, %r92;
fma.rn.ftz.f32 %f160, %f53, %f54, %f160;

{ cvt.f32.f16 %f43, %rs4;}



{.reg .f16 low,high;
mov.b32 {low,high}, %r87;
mov.b16 %rs11, low;}


{ cvt.f32.f16 %f44, %rs11;}



dp4a.s32.s32 %r97, %r80, %r90, %r118;


dp4a.s32.s32 %r101, %r84, %r94, %r97;

mul.ftz.f32 %f55, %f43, %f44;
cvt.rn.f32.s32 %f56, %r101;
fma.rn.ftz.f32 %f159, %f55, %f56, %f159;
cvt.u64.u32 %rd43, %r252;
add.s64 %rd44, %rd22, %rd43;
mul.lo.s64 %rd45, %rd44, 36;
add.s64 %rd46, %rd8, %rd45;
add.s64 %rd47, %rd46, 4;
add.s64 %rd48, %rd47, %rd29;
ld.global.nc.u32 %r108, [%rd48];
add.s64 %rd49, %rd47, %rd31;
ld.global.nc.u32 %r112, [%rd49];

{ cvt.f32.f16 %f45, %rs1;}


ld.global.nc.u32 %r105, [%rd46];

{.reg .f16 low,high;
mov.b32 {low,high}, %r105;
mov.b16 %rs14, low;}


{ cvt.f32.f16 %f46, %rs14;}



dp4a.s32.s32 %r106, %r71, %r108, %r118;


dp4a.s32.s32 %r110, %r75, %r112, %r106;

mul.ftz.f32 %f57, %f45, %f46;
cvt.rn.f32.s32 %f58, %r110;
fma.rn.ftz.f32 %f158, %f57, %f58, %f158;

{ cvt.f32.f16 %f47, %rs4;}



{.reg .f16 low,high;
mov.b32 {low,high}, %r105;
mov.b16 %rs17, low;}


{ cvt.f32.f16 %f48, %rs17;}



dp4a.s32.s32 %r115, %r80, %r108, %r118;


dp4a.s32.s32 %r119, %r84, %r112, %r115;

mul.ftz.f32 %f59, %f47, %f48;
cvt.rn.f32.s32 %f60, %r119;
fma.rn.ftz.f32 %f157, %f59, %f60, %f157;
add.s32 %r252, %r252, 32;
add.s32 %r251, %r251, 32;
add.s64 %rd66, %rd66, 1088;
add.s32 %r253, %r253, 32;
setp.lt.s32 %p4, %r253, %r37;
@%p4 bra $L__BB95_2;

$L__BB95_3:
setp.eq.s32 %p5, %r29, 0;
@%p5 bra $L__BB95_5;

mov.u32 %r131, _ZZ13mul_mat_vec_qIL9ggml_type8ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r132, %r29, 768, %r131;
shl.b32 %r133, %r31, 2;
add.s32 %r134, %r132, %r133;
st.shared.f32 [%r134+-768], %f162;
st.shared.f32 [%r134+-640], %f161;
st.shared.f32 [%r134+-512], %f160;
st.shared.f32 [%r134+-384], %f159;
st.shared.f32 [%r134+-256], %f158;
st.shared.f32 [%r134+-128], %f157;

$L__BB95_5:
setp.ne.s32 %p6, %r29, 0;
bar.sync 0;
@%p6 bra $L__BB95_18;

mov.u32 %r136, %ctaid.z;
mov.u32 %r137, %ctaid.y;
mad.lo.s32 %r138, %r137, %r22, %r1;
mad.lo.s32 %r139, %r136, %r28, %r138;
cvt.u64.u32 %rd11, %r139;
shl.b32 %r141, %r31, 2;
mov.u32 %r142, _ZZ13mul_mat_vec_qIL9ggml_type8ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r11, %r142, %r141;
mov.u32 %r143, 2;
ld.shared.f32 %f61, [%r11];
add.ftz.f32 %f62, %f61, %f162;
ld.shared.f32 %f63, [%r11+768];
add.ftz.f32 %f64, %f63, %f62;
ld.shared.f32 %f65, [%r11+1536];
add.ftz.f32 %f66, %f65, %f64;
mov.b32 %r144, %f66;
mov.u32 %r145, 31;
mov.u32 %r146, 16;
mov.u32 %r147, -1;
shfl.sync.bfly.b32 %r148|%p7, %r144, %r146, %r145, %r147;
mov.b32 %f67, %r148;
add.ftz.f32 %f68, %f66, %f67;
mov.b32 %r149, %f68;
mov.u32 %r150, 8;
shfl.sync.bfly.b32 %r151|%p8, %r149, %r150, %r145, %r147;
mov.b32 %f69, %r151;
add.ftz.f32 %f70, %f68, %f69;
mov.b32 %r152, %f70;
mov.u32 %r153, 4;
shfl.sync.bfly.b32 %r154|%p9, %r152, %r153, %r145, %r147;
mov.b32 %f71, %r154;
add.ftz.f32 %f72, %f70, %f71;
mov.b32 %r155, %f72;
shfl.sync.bfly.b32 %r156|%p10, %r155, %r143, %r145, %r147;
mov.b32 %f73, %r156;
add.ftz.f32 %f74, %f72, %f73;
mov.b32 %r157, %f74;
mov.u32 %r158, 1;
shfl.sync.bfly.b32 %r159|%p11, %r157, %r158, %r145, %r147;
mov.b32 %f75, %r159;
add.ftz.f32 %f19, %f74, %f75;
setp.eq.s32 %p12, %r31, 0;
setp.lt.u32 %p13, %r1, %r16;
and.pred %p1, %p12, %p13;
cvta.to.global.u64 %rd50, %rd16;
mul.wide.u32 %rd51, %r139, 4;
add.s64 %rd12, %rd50, %rd51;
not.pred %p14, %p1;
@%p14 bra $L__BB95_8;

st.global.f32 [%rd12], %f19;

$L__BB95_8:
ld.shared.f32 %f76, [%r11+128];
add.ftz.f32 %f77, %f76, %f161;
ld.shared.f32 %f78, [%r11+896];
add.ftz.f32 %f79, %f78, %f77;
ld.shared.f32 %f80, [%r11+1664];
add.ftz.f32 %f81, %f80, %f79;
mov.b32 %r160, %f81;
shfl.sync.bfly.b32 %r164|%p15, %r160, %r146, %r145, %r147;
mov.b32 %f82, %r164;
add.ftz.f32 %f83, %f81, %f82;
mov.b32 %r165, %f83;
shfl.sync.bfly.b32 %r167|%p16, %r165, %r150, %r145, %r147;
mov.b32 %f84, %r167;
add.ftz.f32 %f85, %f83, %f84;
mov.b32 %r168, %f85;
shfl.sync.bfly.b32 %r170|%p17, %r168, %r153, %r145, %r147;
mov.b32 %f86, %r170;
add.ftz.f32 %f87, %f85, %f86;
mov.b32 %r171, %f87;
shfl.sync.bfly.b32 %r173|%p18, %r171, %r143, %r145, %r147;
mov.b32 %f88, %r173;
add.ftz.f32 %f89, %f87, %f88;
mov.b32 %r174, %f89;
shfl.sync.bfly.b32 %r176|%p19, %r174, %r158, %r145, %r147;
mov.b32 %f90, %r176;
add.ftz.f32 %f20, %f89, %f90;
setp.eq.s32 %p20, %r31, 1;
or.b32 %r180, %r1, 1;
setp.lt.u32 %p21, %r180, %r16;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB95_10;

st.global.f32 [%rd12+4], %f20;

$L__BB95_10:
ld.shared.f32 %f91, [%r11+256];
add.ftz.f32 %f92, %f91, %f160;
ld.shared.f32 %f93, [%r11+1024];
add.ftz.f32 %f94, %f93, %f92;
ld.shared.f32 %f95, [%r11+1792];
add.ftz.f32 %f96, %f95, %f94;
mov.b32 %r181, %f96;
mov.u32 %r182, 31;
mov.u32 %r183, 16;
mov.u32 %r184, -1;
shfl.sync.bfly.b32 %r185|%p23, %r181, %r183, %r182, %r184;
mov.b32 %f97, %r185;
add.ftz.f32 %f98, %f96, %f97;
mov.b32 %r186, %f98;
mov.u32 %r187, 8;
shfl.sync.bfly.b32 %r188|%p24, %r186, %r187, %r182, %r184;
mov.b32 %f99, %r188;
add.ftz.f32 %f100, %f98, %f99;
mov.b32 %r189, %f100;
mov.u32 %r190, 4;
shfl.sync.bfly.b32 %r191|%p25, %r189, %r190, %r182, %r184;
mov.b32 %f101, %r191;
add.ftz.f32 %f102, %f100, %f101;
mov.b32 %r192, %f102;
mov.u32 %r193, 2;
shfl.sync.bfly.b32 %r194|%p26, %r192, %r193, %r182, %r184;
mov.b32 %f103, %r194;
add.ftz.f32 %f104, %f102, %f103;
mov.b32 %r195, %f104;
mov.u32 %r196, 1;
shfl.sync.bfly.b32 %r197|%p27, %r195, %r196, %r182, %r184;
mov.b32 %f105, %r197;
add.ftz.f32 %f21, %f104, %f105;
@%p14 bra $L__BB95_12;

cvt.u64.u32 %rd52, %r16;
add.s64 %rd53, %rd52, %rd11;
shl.b64 %rd55, %rd53, 2;
add.s64 %rd56, %rd50, %rd55;
st.global.f32 [%rd56], %f21;

$L__BB95_12:
ld.shared.f32 %f106, [%r11+384];
add.ftz.f32 %f107, %f106, %f159;
ld.shared.f32 %f108, [%r11+1152];
add.ftz.f32 %f109, %f108, %f107;
ld.shared.f32 %f110, [%r11+1920];
add.ftz.f32 %f111, %f110, %f109;
mov.b32 %r198, %f111;
shfl.sync.bfly.b32 %r202|%p29, %r198, %r183, %r182, %r184;
mov.b32 %f112, %r202;
add.ftz.f32 %f113, %f111, %f112;
mov.b32 %r203, %f113;
shfl.sync.bfly.b32 %r205|%p30, %r203, %r187, %r182, %r184;
mov.b32 %f114, %r205;
add.ftz.f32 %f115, %f113, %f114;
mov.b32 %r206, %f115;
shfl.sync.bfly.b32 %r208|%p31, %r206, %r190, %r182, %r184;
mov.b32 %f116, %r208;
add.ftz.f32 %f117, %f115, %f116;
mov.b32 %r209, %f117;
shfl.sync.bfly.b32 %r211|%p32, %r209, %r193, %r182, %r184;
mov.b32 %f118, %r211;
add.ftz.f32 %f119, %f117, %f118;
mov.b32 %r212, %f119;
shfl.sync.bfly.b32 %r214|%p33, %r212, %r196, %r182, %r184;
mov.b32 %f120, %r214;
add.ftz.f32 %f22, %f119, %f120;
@%p22 bra $L__BB95_14;

add.s32 %r215, %r16, 1;
cvt.u64.u32 %rd57, %r215;
add.s64 %rd58, %rd57, %rd11;
shl.b64 %rd60, %rd58, 2;
add.s64 %rd61, %rd50, %rd60;
st.global.f32 [%rd61], %f22;

$L__BB95_14:
shl.b32 %r216, %r16, 1;
mov.u32 %r217, 1;
ld.shared.f32 %f121, [%r11+512];
add.ftz.f32 %f122, %f121, %f158;
ld.shared.f32 %f123, [%r11+1280];
add.ftz.f32 %f124, %f123, %f122;
ld.shared.f32 %f125, [%r11+2048];
add.ftz.f32 %f126, %f125, %f124;
mov.b32 %r218, %f126;
mov.u32 %r219, 31;
mov.u32 %r220, 16;
mov.u32 %r221, -1;
shfl.sync.bfly.b32 %r222|%p35, %r218, %r220, %r219, %r221;
mov.b32 %f127, %r222;
add.ftz.f32 %f128, %f126, %f127;
mov.b32 %r223, %f128;
mov.u32 %r224, 8;
shfl.sync.bfly.b32 %r225|%p36, %r223, %r224, %r219, %r221;
mov.b32 %f129, %r225;
add.ftz.f32 %f130, %f128, %f129;
mov.b32 %r226, %f130;
mov.u32 %r227, 4;
shfl.sync.bfly.b32 %r228|%p37, %r226, %r227, %r219, %r221;
mov.b32 %f131, %r228;
add.ftz.f32 %f132, %f130, %f131;
mov.b32 %r229, %f132;
mov.u32 %r230, 2;
shfl.sync.bfly.b32 %r231|%p38, %r229, %r230, %r219, %r221;
mov.b32 %f133, %r231;
add.ftz.f32 %f134, %f132, %f133;
mov.b32 %r232, %f134;
shfl.sync.bfly.b32 %r233|%p39, %r232, %r217, %r219, %r221;
mov.b32 %f135, %r233;
add.ftz.f32 %f23, %f134, %f135;
cvt.u64.u32 %rd62, %r216;
add.s64 %rd63, %rd62, %rd11;
shl.b64 %rd65, %rd63, 2;
add.s64 %rd13, %rd50, %rd65;
@%p14 bra $L__BB95_16;

st.global.f32 [%rd13], %f23;

$L__BB95_16:
ld.shared.f32 %f136, [%r11+640];
add.ftz.f32 %f137, %f136, %f157;
ld.shared.f32 %f138, [%r11+1408];
add.ftz.f32 %f139, %f138, %f137;
ld.shared.f32 %f140, [%r11+2176];
add.ftz.f32 %f141, %f140, %f139;
mov.b32 %r234, %f141;
shfl.sync.bfly.b32 %r238|%p41, %r234, %r220, %r219, %r221;
mov.b32 %f142, %r238;
add.ftz.f32 %f143, %f141, %f142;
mov.b32 %r239, %f143;
shfl.sync.bfly.b32 %r241|%p42, %r239, %r224, %r219, %r221;
mov.b32 %f144, %r241;
add.ftz.f32 %f145, %f143, %f144;
mov.b32 %r242, %f145;
shfl.sync.bfly.b32 %r244|%p43, %r242, %r227, %r219, %r221;
mov.b32 %f146, %r244;
add.ftz.f32 %f147, %f145, %f146;
mov.b32 %r245, %f147;
shfl.sync.bfly.b32 %r247|%p44, %r245, %r230, %r219, %r221;
mov.b32 %f148, %r247;
add.ftz.f32 %f149, %f147, %f148;
mov.b32 %r248, %f149;
shfl.sync.bfly.b32 %r250|%p45, %r248, %r217, %r219, %r221;
mov.b32 %f150, %r250;
add.ftz.f32 %f24, %f149, %f150;
@%p22 bra $L__BB95_18;

st.global.f32 [%rd13+4], %f24;

$L__BB95_18:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type8ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type8ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type8ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type8ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type8ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type8ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type8ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type8ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type8ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<59>;
.reg .b16 %rs<33>;
.reg .f32 %f<217>;
.reg .b32 %r<304>;
.reg .b64 %rd<84>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type8ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[3072];

ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type8ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type8ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type8ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type8ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type8ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type8ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type8ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type8ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type8ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type8ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type8ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type8ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type8ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type8ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type8ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type8ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type8ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r37, %r1, 5;
mov.u32 %r38, %tid.x;
add.s32 %r2, %r37, %r38;
mov.u32 %r3, %ctaid.y;
mov.u32 %r4, %ctaid.z;
shr.s32 %r39, %r2, 31;
shr.u32 %r40, %r39, 30;
add.s32 %r41, %r2, %r40;
shr.s32 %r303, %r41, 2;
shr.u32 %r42, %r21, 5;
setp.ge.s32 %p3, %r303, %r42;
mov.f32 %f209, 0f00000000;
mov.f32 %f210, %f209;
mov.f32 %f211, %f209;
mov.f32 %f212, %f209;
mov.f32 %f213, %f209;
mov.f32 %f214, %f209;
mov.f32 %f215, %f209;
mov.f32 %f216, %f209;
@%p3 bra $L__BB96_3;

mov.u32 %r43, %ctaid.x;
shl.b32 %r44, %r43, 1;
mul.lo.s32 %r45, %r3, %r29;
cvt.u64.u32 %rd18, %r45;
mul.lo.s32 %r46, %r4, %r35;
cvt.u64.u32 %rd19, %r46;
add.s64 %rd1, %rd19, %rd18;
mul.hi.u32 %r47, %r4, %r31;
add.s32 %r48, %r47, %r4;
shr.u32 %r49, %r48, %r32;
mul.lo.s32 %r50, %r49, %r34;
mul.hi.u32 %r51, %r3, %r25;
add.s32 %r52, %r51, %r3;
shr.u32 %r53, %r52, %r26;
mul.lo.s32 %r54, %r53, %r28;
cvta.to.global.u64 %rd83, %rd15;
and.b32 %r58, %r41, -4;
sub.s32 %r59, %r2, %r58;
shl.b32 %r60, %r59, 1;
shl.b32 %r61, %r59, 2;
or.b32 %r62, %r61, 1;
cvt.s64.s32 %rd3, %r60;
sub.s32 %r63, %r62, %r60;
cvt.s64.s32 %rd4, %r63;
shl.b32 %r64, %r23, 1;
add.s32 %r302, %r303, %r64;
mad.lo.s32 %r301, %r23, 3, %r303;
add.s32 %r300, %r303, %r23;
add.s32 %r65, %r303, %r50;
add.s32 %r66, %r65, %r54;
or.b32 %r67, %r44, 1;
mad.lo.s32 %r68, %r22, %r67, %r66;
mul.wide.s32 %rd5, %r68, 34;
mad.lo.s32 %r69, %r44, %r22, %r303;
add.s32 %r70, %r69, %r50;
add.s32 %r71, %r70, %r54;
mul.wide.s32 %rd6, %r71, 34;
mul.wide.s32 %rd20, %r61, 2;
add.s64 %rd21, %rd6, %rd20;
add.s64 %rd7, %rd21, 4;
add.s64 %rd22, %rd5, %rd20;
add.s64 %rd8, %rd22, 2;
cvta.to.global.u64 %rd9, %rd16;
mov.f32 %f209, 0f00000000;
shl.b64 %rd29, %rd3, 2;
shl.b64 %rd31, %rd4, 2;

$L__BB96_2:
cvt.u64.u32 %rd23, %r303;
add.s64 %rd24, %rd1, %rd23;
mul.lo.s64 %rd25, %rd24, 36;
add.s64 %rd26, %rd9, %rd25;
add.s64 %rd27, %rd26, 4;
add.s64 %rd28, %rd83, %rd7;
ld.global.nc.u16 %rs25, [%rd28+-2];
ld.global.nc.u16 %rs26, [%rd28];
mov.b32 %r74, {%rs25, %rs26};
add.s64 %rd30, %rd27, %rd29;
ld.global.nc.u32 %r75, [%rd30];
ld.global.nc.u16 %rs27, [%rd28+2];
ld.global.nc.u16 %rs28, [%rd28+4];
mov.b32 %r78, {%rs27, %rs28};
add.s64 %rd32, %rd27, %rd31;
ld.global.nc.u32 %r79, [%rd32];
add.s64 %rd33, %rd83, %rd6;
ld.global.nc.u16 %rs1, [%rd33];

{ cvt.f32.f16 %f49, %rs1;}


ld.global.nc.u32 %r72, [%rd26];

{.reg .f16 low,high;
mov.b32 {low,high}, %r72;
mov.b16 %rs2, low;}


{ cvt.f32.f16 %f50, %rs2;}


mov.u32 %r139, 0;

dp4a.s32.s32 %r73, %r74, %r75, %r139;


dp4a.s32.s32 %r77, %r78, %r79, %r73;

mul.ftz.f32 %f65, %f49, %f50;
cvt.rn.f32.s32 %f66, %r77;
fma.rn.ftz.f32 %f216, %f65, %f66, %f216;
add.s64 %rd34, %rd83, %rd8;
ld.global.nc.u16 %rs29, [%rd34];
ld.global.nc.u16 %rs30, [%rd34+2];
mov.b32 %r83, {%rs29, %rs30};
ld.global.nc.u16 %rs31, [%rd34+4];
ld.global.nc.u16 %rs32, [%rd34+6];
mov.b32 %r87, {%rs31, %rs32};
add.s64 %rd35, %rd83, %rd5;
ld.global.nc.u16 %rs4, [%rd35];

{ cvt.f32.f16 %f51, %rs4;}



{.reg .f16 low,high;
mov.b32 {low,high}, %r72;
mov.b16 %rs5, low;}


{ cvt.f32.f16 %f52, %rs5;}



dp4a.s32.s32 %r82, %r83, %r75, %r139;


dp4a.s32.s32 %r86, %r87, %r79, %r82;

mul.ftz.f32 %f67, %f51, %f52;
cvt.rn.f32.s32 %f68, %r86;
fma.rn.ftz.f32 %f215, %f67, %f68, %f215;
cvt.u64.u32 %rd36, %r300;
add.s64 %rd37, %rd1, %rd36;
mul.lo.s64 %rd38, %rd37, 36;
add.s64 %rd39, %rd9, %rd38;
add.s64 %rd40, %rd39, 4;
add.s64 %rd41, %rd40, %rd29;
ld.global.nc.u32 %r93, [%rd41];
add.s64 %rd42, %rd40, %rd31;
ld.global.nc.u32 %r97, [%rd42];

{ cvt.f32.f16 %f53, %rs1;}


ld.global.nc.u32 %r90, [%rd39];

{.reg .f16 low,high;
mov.b32 {low,high}, %r90;
mov.b16 %rs8, low;}


{ cvt.f32.f16 %f54, %rs8;}



dp4a.s32.s32 %r91, %r74, %r93, %r139;


dp4a.s32.s32 %r95, %r78, %r97, %r91;

mul.ftz.f32 %f69, %f53, %f54;
cvt.rn.f32.s32 %f70, %r95;
fma.rn.ftz.f32 %f214, %f69, %f70, %f214;

{ cvt.f32.f16 %f55, %rs4;}



{.reg .f16 low,high;
mov.b32 {low,high}, %r90;
mov.b16 %rs11, low;}


{ cvt.f32.f16 %f56, %rs11;}



dp4a.s32.s32 %r100, %r83, %r93, %r139;


dp4a.s32.s32 %r104, %r87, %r97, %r100;

mul.ftz.f32 %f71, %f55, %f56;
cvt.rn.f32.s32 %f72, %r104;
fma.rn.ftz.f32 %f213, %f71, %f72, %f213;
cvt.u64.u32 %rd43, %r302;
add.s64 %rd44, %rd1, %rd43;
mul.lo.s64 %rd45, %rd44, 36;
add.s64 %rd46, %rd9, %rd45;
add.s64 %rd47, %rd46, 4;
add.s64 %rd48, %rd47, %rd29;
ld.global.nc.u32 %r111, [%rd48];
add.s64 %rd49, %rd47, %rd31;
ld.global.nc.u32 %r115, [%rd49];

{ cvt.f32.f16 %f57, %rs1;}


ld.global.nc.u32 %r108, [%rd46];

{.reg .f16 low,high;
mov.b32 {low,high}, %r108;
mov.b16 %rs14, low;}


{ cvt.f32.f16 %f58, %rs14;}



dp4a.s32.s32 %r109, %r74, %r111, %r139;


dp4a.s32.s32 %r113, %r78, %r115, %r109;

mul.ftz.f32 %f73, %f57, %f58;
cvt.rn.f32.s32 %f74, %r113;
fma.rn.ftz.f32 %f212, %f73, %f74, %f212;

{ cvt.f32.f16 %f59, %rs4;}



{.reg .f16 low,high;
mov.b32 {low,high}, %r108;
mov.b16 %rs17, low;}


{ cvt.f32.f16 %f60, %rs17;}



dp4a.s32.s32 %r118, %r83, %r111, %r139;


dp4a.s32.s32 %r122, %r87, %r115, %r118;

mul.ftz.f32 %f75, %f59, %f60;
cvt.rn.f32.s32 %f76, %r122;
fma.rn.ftz.f32 %f211, %f75, %f76, %f211;
cvt.u64.u32 %rd50, %r301;
add.s64 %rd51, %rd1, %rd50;
mul.lo.s64 %rd52, %rd51, 36;
add.s64 %rd53, %rd9, %rd52;
add.s64 %rd54, %rd53, 4;
add.s64 %rd55, %rd54, %rd29;
ld.global.nc.u32 %r129, [%rd55];
add.s64 %rd56, %rd54, %rd31;
ld.global.nc.u32 %r133, [%rd56];

{ cvt.f32.f16 %f61, %rs1;}


ld.global.nc.u32 %r126, [%rd53];

{.reg .f16 low,high;
mov.b32 {low,high}, %r126;
mov.b16 %rs20, low;}


{ cvt.f32.f16 %f62, %rs20;}



dp4a.s32.s32 %r127, %r74, %r129, %r139;


dp4a.s32.s32 %r131, %r78, %r133, %r127;

mul.ftz.f32 %f77, %f61, %f62;
cvt.rn.f32.s32 %f78, %r131;
fma.rn.ftz.f32 %f210, %f77, %f78, %f210;

{ cvt.f32.f16 %f63, %rs4;}



{.reg .f16 low,high;
mov.b32 {low,high}, %r126;
mov.b16 %rs23, low;}


{ cvt.f32.f16 %f64, %rs23;}



dp4a.s32.s32 %r136, %r83, %r129, %r139;


dp4a.s32.s32 %r140, %r87, %r133, %r136;

mul.ftz.f32 %f79, %f63, %f64;
cvt.rn.f32.s32 %f80, %r140;
fma.rn.ftz.f32 %f209, %f79, %f80, %f209;
add.s32 %r302, %r302, 32;
add.s32 %r301, %r301, 32;
add.s32 %r300, %r300, 32;
add.s64 %rd83, %rd83, 1088;
add.s32 %r303, %r303, 32;
setp.lt.s32 %p4, %r303, %r42;
@%p4 bra $L__BB96_2;

$L__BB96_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB96_5;

shl.b32 %r145, %r1, 10;
mov.u32 %r146, _ZZ13mul_mat_vec_qIL9ggml_type8ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r147, %r146, %r145;
shl.b32 %r148, %r38, 2;
add.s32 %r149, %r147, %r148;
st.shared.f32 [%r149+-1024], %f216;
st.shared.f32 [%r149+-896], %f215;
st.shared.f32 [%r149+-768], %f214;
st.shared.f32 [%r149+-640], %f213;
st.shared.f32 [%r149+-512], %f212;
st.shared.f32 [%r149+-384], %f211;
st.shared.f32 [%r149+-256], %f210;
st.shared.f32 [%r149+-128], %f209;

$L__BB96_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB96_22;

mov.u32 %r150, %ctaid.x;
shl.b32 %r151, %r150, 1;
mov.u32 %r152, 1;
mad.lo.s32 %r153, %r3, %r30, %r151;
mad.lo.s32 %r154, %r4, %r36, %r153;
cvt.u64.u32 %rd12, %r154;
shl.b32 %r156, %r38, 2;
mov.u32 %r157, _ZZ13mul_mat_vec_qIL9ggml_type8ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r18, %r157, %r156;
mov.u32 %r158, 2;
ld.shared.f32 %f81, [%r18];
add.ftz.f32 %f82, %f81, %f216;
ld.shared.f32 %f83, [%r18+1024];
add.ftz.f32 %f84, %f83, %f82;
ld.shared.f32 %f85, [%r18+2048];
add.ftz.f32 %f86, %f85, %f84;
mov.b32 %r159, %f86;
mov.u32 %r160, 31;
mov.u32 %r161, 16;
mov.u32 %r162, -1;
shfl.sync.bfly.b32 %r163|%p7, %r159, %r161, %r160, %r162;
mov.b32 %f87, %r163;
add.ftz.f32 %f88, %f86, %f87;
mov.b32 %r164, %f88;
mov.u32 %r165, 8;
shfl.sync.bfly.b32 %r166|%p8, %r164, %r165, %r160, %r162;
mov.b32 %f89, %r166;
add.ftz.f32 %f90, %f88, %f89;
mov.b32 %r167, %f90;
mov.u32 %r168, 4;
shfl.sync.bfly.b32 %r169|%p9, %r167, %r168, %r160, %r162;
mov.b32 %f91, %r169;
add.ftz.f32 %f92, %f90, %f91;
mov.b32 %r170, %f92;
shfl.sync.bfly.b32 %r171|%p10, %r170, %r158, %r160, %r162;
mov.b32 %f93, %r171;
add.ftz.f32 %f94, %f92, %f93;
mov.b32 %r172, %f94;
shfl.sync.bfly.b32 %r173|%p11, %r172, %r152, %r160, %r162;
mov.b32 %f95, %r173;
add.ftz.f32 %f25, %f94, %f95;
setp.eq.s32 %p12, %r38, 0;
setp.lt.u32 %p13, %r151, %r24;
and.pred %p1, %p12, %p13;
cvta.to.global.u64 %rd57, %rd17;
mul.wide.u32 %rd58, %r154, 4;
add.s64 %rd13, %rd57, %rd58;
not.pred %p14, %p1;
@%p14 bra $L__BB96_8;

st.global.f32 [%rd13], %f25;

$L__BB96_8:
ld.shared.f32 %f96, [%r18+128];
add.ftz.f32 %f97, %f96, %f215;
ld.shared.f32 %f98, [%r18+1152];
add.ftz.f32 %f99, %f98, %f97;
ld.shared.f32 %f100, [%r18+2176];
add.ftz.f32 %f101, %f100, %f99;
mov.b32 %r174, %f101;
shfl.sync.bfly.b32 %r178|%p15, %r174, %r161, %r160, %r162;
mov.b32 %f102, %r178;
add.ftz.f32 %f103, %f101, %f102;
mov.b32 %r179, %f103;
shfl.sync.bfly.b32 %r181|%p16, %r179, %r165, %r160, %r162;
mov.b32 %f104, %r181;
add.ftz.f32 %f105, %f103, %f104;
mov.b32 %r182, %f105;
shfl.sync.bfly.b32 %r184|%p17, %r182, %r168, %r160, %r162;
mov.b32 %f106, %r184;
add.ftz.f32 %f107, %f105, %f106;
mov.b32 %r185, %f107;
shfl.sync.bfly.b32 %r187|%p18, %r185, %r158, %r160, %r162;
mov.b32 %f108, %r187;
add.ftz.f32 %f109, %f107, %f108;
mov.b32 %r188, %f109;
shfl.sync.bfly.b32 %r190|%p19, %r188, %r152, %r160, %r162;
mov.b32 %f110, %r190;
add.ftz.f32 %f26, %f109, %f110;
setp.eq.s32 %p20, %r38, 1;
or.b32 %r194, %r151, 1;
setp.lt.u32 %p21, %r194, %r24;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB96_10;

st.global.f32 [%rd13+4], %f26;

$L__BB96_10:
ld.shared.f32 %f111, [%r18+256];
add.ftz.f32 %f112, %f111, %f214;
ld.shared.f32 %f113, [%r18+1280];
add.ftz.f32 %f114, %f113, %f112;
ld.shared.f32 %f115, [%r18+2304];
add.ftz.f32 %f116, %f115, %f114;
mov.b32 %r195, %f116;
mov.u32 %r196, 31;
mov.u32 %r197, 16;
mov.u32 %r198, -1;
shfl.sync.bfly.b32 %r199|%p23, %r195, %r197, %r196, %r198;
mov.b32 %f117, %r199;
add.ftz.f32 %f118, %f116, %f117;
mov.b32 %r200, %f118;
mov.u32 %r201, 8;
shfl.sync.bfly.b32 %r202|%p24, %r200, %r201, %r196, %r198;
mov.b32 %f119, %r202;
add.ftz.f32 %f120, %f118, %f119;
mov.b32 %r203, %f120;
mov.u32 %r204, 4;
shfl.sync.bfly.b32 %r205|%p25, %r203, %r204, %r196, %r198;
mov.b32 %f121, %r205;
add.ftz.f32 %f122, %f120, %f121;
mov.b32 %r206, %f122;
mov.u32 %r207, 2;
shfl.sync.bfly.b32 %r208|%p26, %r206, %r207, %r196, %r198;
mov.b32 %f123, %r208;
add.ftz.f32 %f124, %f122, %f123;
mov.b32 %r209, %f124;
mov.u32 %r210, 1;
shfl.sync.bfly.b32 %r211|%p27, %r209, %r210, %r196, %r198;
mov.b32 %f125, %r211;
add.ftz.f32 %f27, %f124, %f125;
@%p14 bra $L__BB96_12;

cvt.u64.u32 %rd59, %r24;
add.s64 %rd60, %rd59, %rd12;
shl.b64 %rd62, %rd60, 2;
add.s64 %rd63, %rd57, %rd62;
st.global.f32 [%rd63], %f27;

$L__BB96_12:
ld.shared.f32 %f126, [%r18+384];
add.ftz.f32 %f127, %f126, %f213;
ld.shared.f32 %f128, [%r18+1408];
add.ftz.f32 %f129, %f128, %f127;
ld.shared.f32 %f130, [%r18+2432];
add.ftz.f32 %f131, %f130, %f129;
mov.b32 %r212, %f131;
shfl.sync.bfly.b32 %r216|%p29, %r212, %r197, %r196, %r198;
mov.b32 %f132, %r216;
add.ftz.f32 %f133, %f131, %f132;
mov.b32 %r217, %f133;
shfl.sync.bfly.b32 %r219|%p30, %r217, %r201, %r196, %r198;
mov.b32 %f134, %r219;
add.ftz.f32 %f135, %f133, %f134;
mov.b32 %r220, %f135;
shfl.sync.bfly.b32 %r222|%p31, %r220, %r204, %r196, %r198;
mov.b32 %f136, %r222;
add.ftz.f32 %f137, %f135, %f136;
mov.b32 %r223, %f137;
shfl.sync.bfly.b32 %r225|%p32, %r223, %r207, %r196, %r198;
mov.b32 %f138, %r225;
add.ftz.f32 %f139, %f137, %f138;
mov.b32 %r226, %f139;
shfl.sync.bfly.b32 %r228|%p33, %r226, %r210, %r196, %r198;
mov.b32 %f140, %r228;
add.ftz.f32 %f28, %f139, %f140;
@%p22 bra $L__BB96_14;

add.s32 %r229, %r24, 1;
cvt.u64.u32 %rd64, %r229;
add.s64 %rd65, %rd64, %rd12;
shl.b64 %rd67, %rd65, 2;
add.s64 %rd68, %rd57, %rd67;
st.global.f32 [%rd68], %f28;

$L__BB96_14:
shl.b32 %r230, %r24, 1;
mov.u32 %r231, 1;
ld.shared.f32 %f141, [%r18+512];
add.ftz.f32 %f142, %f141, %f212;
ld.shared.f32 %f143, [%r18+1536];
add.ftz.f32 %f144, %f143, %f142;
ld.shared.f32 %f145, [%r18+2560];
add.ftz.f32 %f146, %f145, %f144;
mov.b32 %r232, %f146;
mov.u32 %r233, 31;
mov.u32 %r234, 16;
mov.u32 %r235, -1;
shfl.sync.bfly.b32 %r236|%p35, %r232, %r234, %r233, %r235;
mov.b32 %f147, %r236;
add.ftz.f32 %f148, %f146, %f147;
mov.b32 %r237, %f148;
mov.u32 %r238, 8;
shfl.sync.bfly.b32 %r239|%p36, %r237, %r238, %r233, %r235;
mov.b32 %f149, %r239;
add.ftz.f32 %f150, %f148, %f149;
mov.b32 %r240, %f150;
mov.u32 %r241, 4;
shfl.sync.bfly.b32 %r242|%p37, %r240, %r241, %r233, %r235;
mov.b32 %f151, %r242;
add.ftz.f32 %f152, %f150, %f151;
mov.b32 %r243, %f152;
mov.u32 %r244, 2;
shfl.sync.bfly.b32 %r245|%p38, %r243, %r244, %r233, %r235;
mov.b32 %f153, %r245;
add.ftz.f32 %f154, %f152, %f153;
mov.b32 %r246, %f154;
shfl.sync.bfly.b32 %r247|%p39, %r246, %r231, %r233, %r235;
mov.b32 %f155, %r247;
add.ftz.f32 %f29, %f154, %f155;
cvt.u64.u32 %rd69, %r230;
add.s64 %rd70, %rd69, %rd12;
shl.b64 %rd72, %rd70, 2;
add.s64 %rd14, %rd57, %rd72;
@%p14 bra $L__BB96_16;

st.global.f32 [%rd14], %f29;

$L__BB96_16:
ld.shared.f32 %f156, [%r18+640];
add.ftz.f32 %f157, %f156, %f211;
ld.shared.f32 %f158, [%r18+1664];
add.ftz.f32 %f159, %f158, %f157;
ld.shared.f32 %f160, [%r18+2688];
add.ftz.f32 %f161, %f160, %f159;
mov.b32 %r248, %f161;
shfl.sync.bfly.b32 %r252|%p41, %r248, %r234, %r233, %r235;
mov.b32 %f162, %r252;
add.ftz.f32 %f163, %f161, %f162;
mov.b32 %r253, %f163;
shfl.sync.bfly.b32 %r255|%p42, %r253, %r238, %r233, %r235;
mov.b32 %f164, %r255;
add.ftz.f32 %f165, %f163, %f164;
mov.b32 %r256, %f165;
shfl.sync.bfly.b32 %r258|%p43, %r256, %r241, %r233, %r235;
mov.b32 %f166, %r258;
add.ftz.f32 %f167, %f165, %f166;
mov.b32 %r259, %f167;
shfl.sync.bfly.b32 %r261|%p44, %r259, %r244, %r233, %r235;
mov.b32 %f168, %r261;
add.ftz.f32 %f169, %f167, %f168;
mov.b32 %r262, %f169;
shfl.sync.bfly.b32 %r264|%p45, %r262, %r231, %r233, %r235;
mov.b32 %f170, %r264;
add.ftz.f32 %f30, %f169, %f170;
@%p22 bra $L__BB96_18;

st.global.f32 [%rd14+4], %f30;

$L__BB96_18:
mul.lo.s32 %r20, %r24, 3;
ld.shared.f32 %f171, [%r18+768];
add.ftz.f32 %f172, %f171, %f210;
ld.shared.f32 %f173, [%r18+1792];
add.ftz.f32 %f174, %f173, %f172;
ld.shared.f32 %f175, [%r18+2816];
add.ftz.f32 %f176, %f175, %f174;
mov.b32 %r265, %f176;
mov.u32 %r266, 31;
mov.u32 %r267, 16;
mov.u32 %r268, -1;
shfl.sync.bfly.b32 %r269|%p47, %r265, %r267, %r266, %r268;
mov.b32 %f177, %r269;
add.ftz.f32 %f178, %f176, %f177;
mov.b32 %r270, %f178;
mov.u32 %r271, 8;
shfl.sync.bfly.b32 %r272|%p48, %r270, %r271, %r266, %r268;
mov.b32 %f179, %r272;
add.ftz.f32 %f180, %f178, %f179;
mov.b32 %r273, %f180;
mov.u32 %r274, 4;
shfl.sync.bfly.b32 %r275|%p49, %r273, %r274, %r266, %r268;
mov.b32 %f181, %r275;
add.ftz.f32 %f182, %f180, %f181;
mov.b32 %r276, %f182;
mov.u32 %r277, 2;
shfl.sync.bfly.b32 %r278|%p50, %r276, %r277, %r266, %r268;
mov.b32 %f183, %r278;
add.ftz.f32 %f184, %f182, %f183;
mov.b32 %r279, %f184;
mov.u32 %r280, 1;
shfl.sync.bfly.b32 %r281|%p51, %r279, %r280, %r266, %r268;
mov.b32 %f185, %r281;
add.ftz.f32 %f31, %f184, %f185;
@%p14 bra $L__BB96_20;

cvt.u64.u32 %rd73, %r20;
add.s64 %rd74, %rd73, %rd12;
shl.b64 %rd76, %rd74, 2;
add.s64 %rd77, %rd57, %rd76;
st.global.f32 [%rd77], %f31;

$L__BB96_20:
ld.shared.f32 %f186, [%r18+896];
add.ftz.f32 %f187, %f186, %f209;
ld.shared.f32 %f188, [%r18+1920];
add.ftz.f32 %f189, %f188, %f187;
ld.shared.f32 %f190, [%r18+2944];
add.ftz.f32 %f191, %f190, %f189;
mov.b32 %r282, %f191;
shfl.sync.bfly.b32 %r286|%p53, %r282, %r267, %r266, %r268;
mov.b32 %f192, %r286;
add.ftz.f32 %f193, %f191, %f192;
mov.b32 %r287, %f193;
shfl.sync.bfly.b32 %r289|%p54, %r287, %r271, %r266, %r268;
mov.b32 %f194, %r289;
add.ftz.f32 %f195, %f193, %f194;
mov.b32 %r290, %f195;
shfl.sync.bfly.b32 %r292|%p55, %r290, %r274, %r266, %r268;
mov.b32 %f196, %r292;
add.ftz.f32 %f197, %f195, %f196;
mov.b32 %r293, %f197;
shfl.sync.bfly.b32 %r295|%p56, %r293, %r277, %r266, %r268;
mov.b32 %f198, %r295;
add.ftz.f32 %f199, %f197, %f198;
mov.b32 %r296, %f199;
shfl.sync.bfly.b32 %r298|%p57, %r296, %r280, %r266, %r268;
mov.b32 %f200, %r298;
add.ftz.f32 %f32, %f199, %f200;
@%p22 bra $L__BB96_22;

add.s32 %r299, %r20, 1;
cvt.u64.u32 %rd78, %r299;
add.s64 %rd79, %rd78, %rd12;
shl.b64 %rd81, %rd79, 2;
add.s64 %rd82, %rd57, %rd81;
st.global.f32 [%rd82], %f32;

$L__BB96_22:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type8ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type8ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type8ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type8ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type8ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type8ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type8ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type8ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type8ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<71>;
.reg .b16 %rs<39>;
.reg .f32 %f<231>;
.reg .b32 %r<358>;
.reg .b64 %rd<96>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type8ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1280];

ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type8ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type8ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type8ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type8ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type8ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type8ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type8ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type8ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type8ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type8ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type8ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type8ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type8ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type8ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type8ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type8ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type8ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r38, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r38, %r2;
mov.u32 %r39, %ctaid.x;
shl.b32 %r4, %r39, 1;
shr.s32 %r40, %r3, 31;
shr.u32 %r41, %r40, 30;
add.s32 %r42, %r3, %r41;
shr.s32 %r357, %r42, 2;
shr.u32 %r43, %r22, 5;
setp.ge.s32 %p3, %r357, %r43;
mov.f32 %f221, 0f00000000;
mov.f32 %f222, %f221;
mov.f32 %f223, %f221;
mov.f32 %f224, %f221;
mov.f32 %f225, %f221;
mov.f32 %f226, %f221;
mov.f32 %f227, %f221;
mov.f32 %f228, %f221;
mov.f32 %f229, %f221;
mov.f32 %f230, %f221;
@%p3 bra $L__BB97_3;

mov.u32 %r44, %ctaid.y;
mov.u32 %r45, %ctaid.z;
mul.lo.s32 %r46, %r44, %r30;
cvt.u64.u32 %rd19, %r46;
mul.lo.s32 %r47, %r45, %r36;
cvt.u64.u32 %rd20, %r47;
add.s64 %rd1, %rd20, %rd19;
mul.hi.u32 %r48, %r45, %r32;
add.s32 %r49, %r48, %r45;
shr.u32 %r50, %r49, %r33;
mul.lo.s32 %r51, %r50, %r35;
mul.hi.u32 %r52, %r44, %r26;
add.s32 %r53, %r52, %r44;
shr.u32 %r54, %r53, %r27;
mul.lo.s32 %r55, %r54, %r29;
cvta.to.global.u64 %rd95, %rd16;
and.b32 %r59, %r42, -4;
sub.s32 %r60, %r3, %r59;
shl.b32 %r61, %r60, 1;
shl.b32 %r62, %r60, 2;
or.b32 %r63, %r62, 1;
cvt.s64.s32 %rd3, %r61;
sub.s32 %r64, %r63, %r61;
cvt.s64.s32 %rd4, %r64;
shl.b32 %r65, %r24, 1;
add.s32 %r356, %r357, %r65;
mad.lo.s32 %r355, %r24, 3, %r357;
shl.b32 %r66, %r24, 2;
add.s32 %r354, %r357, %r66;
add.s32 %r353, %r357, %r24;
add.s32 %r67, %r357, %r51;
add.s32 %r68, %r67, %r55;
add.s32 %r69, %r4, 1;
mad.lo.s32 %r70, %r23, %r69, %r68;
mul.wide.s32 %rd5, %r70, 34;
mad.lo.s32 %r71, %r4, %r23, %r357;
add.s32 %r72, %r71, %r51;
add.s32 %r73, %r72, %r55;
mul.wide.s32 %rd6, %r73, 34;
mul.wide.s32 %rd21, %r62, 2;
add.s64 %rd22, %rd6, %rd21;
add.s64 %rd7, %rd22, 4;
add.s64 %rd23, %rd5, %rd21;
add.s64 %rd8, %rd23, 2;
cvta.to.global.u64 %rd9, %rd17;
mov.f32 %f221, 0f00000000;
shl.b64 %rd30, %rd3, 2;
shl.b64 %rd32, %rd4, 2;

$L__BB97_2:
cvt.u64.u32 %rd24, %r357;
add.s64 %rd25, %rd1, %rd24;
mul.lo.s64 %rd26, %rd25, 36;
add.s64 %rd27, %rd9, %rd26;
add.s64 %rd28, %rd27, 4;
add.s64 %rd29, %rd95, %rd7;
ld.global.nc.u16 %rs31, [%rd29+-2];
ld.global.nc.u16 %rs32, [%rd29];
mov.b32 %r76, {%rs31, %rs32};
add.s64 %rd31, %rd28, %rd30;
ld.global.nc.u32 %r77, [%rd31];
ld.global.nc.u16 %rs33, [%rd29+2];
ld.global.nc.u16 %rs34, [%rd29+4];
mov.b32 %r80, {%rs33, %rs34};
add.s64 %rd33, %rd28, %rd32;
ld.global.nc.u32 %r81, [%rd33];
add.s64 %rd34, %rd95, %rd6;
ld.global.nc.u16 %rs1, [%rd34];

{ cvt.f32.f16 %f61, %rs1;}


ld.global.nc.u32 %r74, [%rd27];

{.reg .f16 low,high;
mov.b32 {low,high}, %r74;
mov.b16 %rs2, low;}


{ cvt.f32.f16 %f62, %rs2;}


mov.u32 %r159, 0;

dp4a.s32.s32 %r75, %r76, %r77, %r159;


dp4a.s32.s32 %r79, %r80, %r81, %r75;

mul.ftz.f32 %f81, %f61, %f62;
cvt.rn.f32.s32 %f82, %r79;
fma.rn.ftz.f32 %f230, %f81, %f82, %f230;
add.s64 %rd35, %rd95, %rd8;
ld.global.nc.u16 %rs35, [%rd35];
ld.global.nc.u16 %rs36, [%rd35+2];
mov.b32 %r85, {%rs35, %rs36};
ld.global.nc.u16 %rs37, [%rd35+4];
ld.global.nc.u16 %rs38, [%rd35+6];
mov.b32 %r89, {%rs37, %rs38};
add.s64 %rd36, %rd95, %rd5;
ld.global.nc.u16 %rs4, [%rd36];

{ cvt.f32.f16 %f63, %rs4;}



{.reg .f16 low,high;
mov.b32 {low,high}, %r74;
mov.b16 %rs5, low;}


{ cvt.f32.f16 %f64, %rs5;}



dp4a.s32.s32 %r84, %r85, %r77, %r159;


dp4a.s32.s32 %r88, %r89, %r81, %r84;

mul.ftz.f32 %f83, %f63, %f64;
cvt.rn.f32.s32 %f84, %r88;
fma.rn.ftz.f32 %f229, %f83, %f84, %f229;
cvt.u64.u32 %rd37, %r353;
add.s64 %rd38, %rd1, %rd37;
mul.lo.s64 %rd39, %rd38, 36;
add.s64 %rd40, %rd9, %rd39;
add.s64 %rd41, %rd40, 4;
add.s64 %rd42, %rd41, %rd30;
ld.global.nc.u32 %r95, [%rd42];
add.s64 %rd43, %rd41, %rd32;
ld.global.nc.u32 %r99, [%rd43];

{ cvt.f32.f16 %f65, %rs1;}


ld.global.nc.u32 %r92, [%rd40];

{.reg .f16 low,high;
mov.b32 {low,high}, %r92;
mov.b16 %rs8, low;}


{ cvt.f32.f16 %f66, %rs8;}



dp4a.s32.s32 %r93, %r76, %r95, %r159;


dp4a.s32.s32 %r97, %r80, %r99, %r93;

mul.ftz.f32 %f85, %f65, %f66;
cvt.rn.f32.s32 %f86, %r97;
fma.rn.ftz.f32 %f228, %f85, %f86, %f228;

{ cvt.f32.f16 %f67, %rs4;}



{.reg .f16 low,high;
mov.b32 {low,high}, %r92;
mov.b16 %rs11, low;}


{ cvt.f32.f16 %f68, %rs11;}



dp4a.s32.s32 %r102, %r85, %r95, %r159;


dp4a.s32.s32 %r106, %r89, %r99, %r102;

mul.ftz.f32 %f87, %f67, %f68;
cvt.rn.f32.s32 %f88, %r106;
fma.rn.ftz.f32 %f227, %f87, %f88, %f227;
cvt.u64.u32 %rd44, %r356;
add.s64 %rd45, %rd1, %rd44;
mul.lo.s64 %rd46, %rd45, 36;
add.s64 %rd47, %rd9, %rd46;
add.s64 %rd48, %rd47, 4;
add.s64 %rd49, %rd48, %rd30;
ld.global.nc.u32 %r113, [%rd49];
add.s64 %rd50, %rd48, %rd32;
ld.global.nc.u32 %r117, [%rd50];

{ cvt.f32.f16 %f69, %rs1;}


ld.global.nc.u32 %r110, [%rd47];

{.reg .f16 low,high;
mov.b32 {low,high}, %r110;
mov.b16 %rs14, low;}


{ cvt.f32.f16 %f70, %rs14;}



dp4a.s32.s32 %r111, %r76, %r113, %r159;


dp4a.s32.s32 %r115, %r80, %r117, %r111;

mul.ftz.f32 %f89, %f69, %f70;
cvt.rn.f32.s32 %f90, %r115;
fma.rn.ftz.f32 %f226, %f89, %f90, %f226;

{ cvt.f32.f16 %f71, %rs4;}



{.reg .f16 low,high;
mov.b32 {low,high}, %r110;
mov.b16 %rs17, low;}


{ cvt.f32.f16 %f72, %rs17;}



dp4a.s32.s32 %r120, %r85, %r113, %r159;


dp4a.s32.s32 %r124, %r89, %r117, %r120;

mul.ftz.f32 %f91, %f71, %f72;
cvt.rn.f32.s32 %f92, %r124;
fma.rn.ftz.f32 %f225, %f91, %f92, %f225;
cvt.u64.u32 %rd51, %r355;
add.s64 %rd52, %rd1, %rd51;
mul.lo.s64 %rd53, %rd52, 36;
add.s64 %rd54, %rd9, %rd53;
add.s64 %rd55, %rd54, 4;
add.s64 %rd56, %rd55, %rd30;
ld.global.nc.u32 %r131, [%rd56];
add.s64 %rd57, %rd55, %rd32;
ld.global.nc.u32 %r135, [%rd57];

{ cvt.f32.f16 %f73, %rs1;}


ld.global.nc.u32 %r128, [%rd54];

{.reg .f16 low,high;
mov.b32 {low,high}, %r128;
mov.b16 %rs20, low;}


{ cvt.f32.f16 %f74, %rs20;}



dp4a.s32.s32 %r129, %r76, %r131, %r159;


dp4a.s32.s32 %r133, %r80, %r135, %r129;

mul.ftz.f32 %f93, %f73, %f74;
cvt.rn.f32.s32 %f94, %r133;
fma.rn.ftz.f32 %f224, %f93, %f94, %f224;

{ cvt.f32.f16 %f75, %rs4;}



{.reg .f16 low,high;
mov.b32 {low,high}, %r128;
mov.b16 %rs23, low;}


{ cvt.f32.f16 %f76, %rs23;}



dp4a.s32.s32 %r138, %r85, %r131, %r159;


dp4a.s32.s32 %r142, %r89, %r135, %r138;

mul.ftz.f32 %f95, %f75, %f76;
cvt.rn.f32.s32 %f96, %r142;
fma.rn.ftz.f32 %f223, %f95, %f96, %f223;
cvt.u64.u32 %rd58, %r354;
add.s64 %rd59, %rd1, %rd58;
mul.lo.s64 %rd60, %rd59, 36;
add.s64 %rd61, %rd9, %rd60;
add.s64 %rd62, %rd61, 4;
add.s64 %rd63, %rd62, %rd30;
ld.global.nc.u32 %r149, [%rd63];
add.s64 %rd64, %rd62, %rd32;
ld.global.nc.u32 %r153, [%rd64];

{ cvt.f32.f16 %f77, %rs1;}


ld.global.nc.u32 %r146, [%rd61];

{.reg .f16 low,high;
mov.b32 {low,high}, %r146;
mov.b16 %rs26, low;}


{ cvt.f32.f16 %f78, %rs26;}



dp4a.s32.s32 %r147, %r76, %r149, %r159;


dp4a.s32.s32 %r151, %r80, %r153, %r147;

mul.ftz.f32 %f97, %f77, %f78;
cvt.rn.f32.s32 %f98, %r151;
fma.rn.ftz.f32 %f222, %f97, %f98, %f222;

{ cvt.f32.f16 %f79, %rs4;}



{.reg .f16 low,high;
mov.b32 {low,high}, %r146;
mov.b16 %rs29, low;}


{ cvt.f32.f16 %f80, %rs29;}



dp4a.s32.s32 %r156, %r85, %r149, %r159;


dp4a.s32.s32 %r160, %r89, %r153, %r156;

mul.ftz.f32 %f99, %f79, %f80;
cvt.rn.f32.s32 %f100, %r160;
fma.rn.ftz.f32 %f221, %f99, %f100, %f221;
add.s32 %r356, %r356, 16;
add.s32 %r355, %r355, 16;
add.s32 %r354, %r354, 16;
add.s32 %r353, %r353, 16;
add.s64 %rd95, %rd95, 544;
add.s32 %r357, %r357, 16;
setp.lt.s32 %p4, %r357, %r43;
@%p4 bra $L__BB97_2;

$L__BB97_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB97_5;

mov.u32 %r165, _ZZ13mul_mat_vec_qIL9ggml_type8ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r166, %r1, 1280, %r165;
shl.b32 %r167, %r2, 2;
add.s32 %r168, %r166, %r167;
st.shared.f32 [%r168+-1280], %f230;
st.shared.f32 [%r168+-1152], %f229;
st.shared.f32 [%r168+-1024], %f228;
st.shared.f32 [%r168+-896], %f227;
st.shared.f32 [%r168+-768], %f226;
st.shared.f32 [%r168+-640], %f225;
st.shared.f32 [%r168+-512], %f224;
st.shared.f32 [%r168+-384], %f223;
st.shared.f32 [%r168+-256], %f222;
st.shared.f32 [%r168+-128], %f221;

$L__BB97_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB97_26;

mov.u32 %r169, %ctaid.z;
mov.u32 %r170, %ctaid.y;
mad.lo.s32 %r171, %r170, %r31, %r4;
mad.lo.s32 %r172, %r169, %r37, %r171;
cvt.u64.u32 %rd12, %r172;
shl.b32 %r173, %r2, 2;
mov.u32 %r174, _ZZ13mul_mat_vec_qIL9ggml_type8ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r20, %r174, %r173;
mov.u32 %r175, 2;
ld.shared.f32 %f101, [%r20];
add.ftz.f32 %f102, %f101, %f230;
mov.b32 %r176, %f102;
mov.u32 %r177, 31;
mov.u32 %r178, 16;
mov.u32 %r179, -1;
shfl.sync.bfly.b32 %r180|%p7, %r176, %r178, %r177, %r179;
mov.b32 %f103, %r180;
add.ftz.f32 %f104, %f102, %f103;
mov.b32 %r181, %f104;
mov.u32 %r182, 8;
shfl.sync.bfly.b32 %r183|%p8, %r181, %r182, %r177, %r179;
mov.b32 %f105, %r183;
add.ftz.f32 %f106, %f104, %f105;
mov.b32 %r184, %f106;
mov.u32 %r185, 4;
shfl.sync.bfly.b32 %r186|%p9, %r184, %r185, %r177, %r179;
mov.b32 %f107, %r186;
add.ftz.f32 %f108, %f106, %f107;
mov.b32 %r187, %f108;
shfl.sync.bfly.b32 %r188|%p10, %r187, %r175, %r177, %r179;
mov.b32 %f109, %r188;
add.ftz.f32 %f110, %f108, %f109;
mov.b32 %r189, %f110;
mov.u32 %r190, 1;
shfl.sync.bfly.b32 %r191|%p11, %r189, %r190, %r177, %r179;
mov.b32 %f111, %r191;
add.ftz.f32 %f31, %f110, %f111;
setp.lt.u32 %p12, %r4, %r25;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
cvta.to.global.u64 %rd65, %rd18;
mul.wide.u32 %rd66, %r172, 4;
add.s64 %rd13, %rd65, %rd66;
not.pred %p14, %p1;
@%p14 bra $L__BB97_8;

st.global.f32 [%rd13], %f31;

$L__BB97_8:
ld.shared.f32 %f112, [%r20+128];
add.ftz.f32 %f113, %f112, %f229;
mov.b32 %r192, %f113;
shfl.sync.bfly.b32 %r196|%p15, %r192, %r178, %r177, %r179;
mov.b32 %f114, %r196;
add.ftz.f32 %f115, %f113, %f114;
mov.b32 %r197, %f115;
shfl.sync.bfly.b32 %r199|%p16, %r197, %r182, %r177, %r179;
mov.b32 %f116, %r199;
add.ftz.f32 %f117, %f115, %f116;
mov.b32 %r200, %f117;
shfl.sync.bfly.b32 %r202|%p17, %r200, %r185, %r177, %r179;
mov.b32 %f118, %r202;
add.ftz.f32 %f119, %f117, %f118;
mov.b32 %r203, %f119;
shfl.sync.bfly.b32 %r205|%p18, %r203, %r175, %r177, %r179;
mov.b32 %f120, %r205;
add.ftz.f32 %f121, %f119, %f120;
mov.b32 %r206, %f121;
shfl.sync.bfly.b32 %r208|%p19, %r206, %r190, %r177, %r179;
mov.b32 %f122, %r208;
add.ftz.f32 %f32, %f121, %f122;
setp.eq.s32 %p20, %r2, 1;
or.b32 %r212, %r4, 1;
setp.lt.u32 %p21, %r212, %r25;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB97_10;

st.global.f32 [%rd13+4], %f32;

$L__BB97_10:
ld.shared.f32 %f123, [%r20+256];
add.ftz.f32 %f124, %f123, %f228;
mov.b32 %r213, %f124;
mov.u32 %r214, 31;
mov.u32 %r215, 16;
mov.u32 %r216, -1;
shfl.sync.bfly.b32 %r217|%p23, %r213, %r215, %r214, %r216;
mov.b32 %f125, %r217;
add.ftz.f32 %f126, %f124, %f125;
mov.b32 %r218, %f126;
mov.u32 %r219, 8;
shfl.sync.bfly.b32 %r220|%p24, %r218, %r219, %r214, %r216;
mov.b32 %f127, %r220;
add.ftz.f32 %f128, %f126, %f127;
mov.b32 %r221, %f128;
mov.u32 %r222, 4;
shfl.sync.bfly.b32 %r223|%p25, %r221, %r222, %r214, %r216;
mov.b32 %f129, %r223;
add.ftz.f32 %f130, %f128, %f129;
mov.b32 %r224, %f130;
mov.u32 %r225, 2;
shfl.sync.bfly.b32 %r226|%p26, %r224, %r225, %r214, %r216;
mov.b32 %f131, %r226;
add.ftz.f32 %f132, %f130, %f131;
mov.b32 %r227, %f132;
mov.u32 %r228, 1;
shfl.sync.bfly.b32 %r229|%p27, %r227, %r228, %r214, %r216;
mov.b32 %f133, %r229;
add.ftz.f32 %f33, %f132, %f133;
@%p14 bra $L__BB97_12;

cvt.u64.u32 %rd67, %r25;
add.s64 %rd68, %rd67, %rd12;
shl.b64 %rd70, %rd68, 2;
add.s64 %rd71, %rd65, %rd70;
st.global.f32 [%rd71], %f33;

$L__BB97_12:
ld.shared.f32 %f134, [%r20+384];
add.ftz.f32 %f135, %f134, %f227;
mov.b32 %r230, %f135;
shfl.sync.bfly.b32 %r234|%p29, %r230, %r215, %r214, %r216;
mov.b32 %f136, %r234;
add.ftz.f32 %f137, %f135, %f136;
mov.b32 %r235, %f137;
shfl.sync.bfly.b32 %r237|%p30, %r235, %r219, %r214, %r216;
mov.b32 %f138, %r237;
add.ftz.f32 %f139, %f137, %f138;
mov.b32 %r238, %f139;
shfl.sync.bfly.b32 %r240|%p31, %r238, %r222, %r214, %r216;
mov.b32 %f140, %r240;
add.ftz.f32 %f141, %f139, %f140;
mov.b32 %r241, %f141;
shfl.sync.bfly.b32 %r243|%p32, %r241, %r225, %r214, %r216;
mov.b32 %f142, %r243;
add.ftz.f32 %f143, %f141, %f142;
mov.b32 %r244, %f143;
shfl.sync.bfly.b32 %r246|%p33, %r244, %r228, %r214, %r216;
mov.b32 %f144, %r246;
add.ftz.f32 %f34, %f143, %f144;
@%p22 bra $L__BB97_14;

add.s32 %r247, %r25, 1;
cvt.u64.u32 %rd72, %r247;
add.s64 %rd73, %rd72, %rd12;
shl.b64 %rd75, %rd73, 2;
add.s64 %rd76, %rd65, %rd75;
st.global.f32 [%rd76], %f34;

$L__BB97_14:
shl.b32 %r248, %r25, 1;
mov.u32 %r249, 1;
ld.shared.f32 %f145, [%r20+512];
add.ftz.f32 %f146, %f145, %f226;
mov.b32 %r250, %f146;
mov.u32 %r251, 31;
mov.u32 %r252, 16;
mov.u32 %r253, -1;
shfl.sync.bfly.b32 %r254|%p35, %r250, %r252, %r251, %r253;
mov.b32 %f147, %r254;
add.ftz.f32 %f148, %f146, %f147;
mov.b32 %r255, %f148;
mov.u32 %r256, 8;
shfl.sync.bfly.b32 %r257|%p36, %r255, %r256, %r251, %r253;
mov.b32 %f149, %r257;
add.ftz.f32 %f150, %f148, %f149;
mov.b32 %r258, %f150;
mov.u32 %r259, 4;
shfl.sync.bfly.b32 %r260|%p37, %r258, %r259, %r251, %r253;
mov.b32 %f151, %r260;
add.ftz.f32 %f152, %f150, %f151;
mov.b32 %r261, %f152;
mov.u32 %r262, 2;
shfl.sync.bfly.b32 %r263|%p38, %r261, %r262, %r251, %r253;
mov.b32 %f153, %r263;
add.ftz.f32 %f154, %f152, %f153;
mov.b32 %r264, %f154;
shfl.sync.bfly.b32 %r265|%p39, %r264, %r249, %r251, %r253;
mov.b32 %f155, %r265;
add.ftz.f32 %f35, %f154, %f155;
cvt.u64.u32 %rd77, %r248;
add.s64 %rd78, %rd77, %rd12;
shl.b64 %rd80, %rd78, 2;
add.s64 %rd14, %rd65, %rd80;
@%p14 bra $L__BB97_16;

st.global.f32 [%rd14], %f35;

$L__BB97_16:
ld.shared.f32 %f156, [%r20+640];
add.ftz.f32 %f157, %f156, %f225;
mov.b32 %r266, %f157;
shfl.sync.bfly.b32 %r270|%p41, %r266, %r252, %r251, %r253;
mov.b32 %f158, %r270;
add.ftz.f32 %f159, %f157, %f158;
mov.b32 %r271, %f159;
shfl.sync.bfly.b32 %r273|%p42, %r271, %r256, %r251, %r253;
mov.b32 %f160, %r273;
add.ftz.f32 %f161, %f159, %f160;
mov.b32 %r274, %f161;
shfl.sync.bfly.b32 %r276|%p43, %r274, %r259, %r251, %r253;
mov.b32 %f162, %r276;
add.ftz.f32 %f163, %f161, %f162;
mov.b32 %r277, %f163;
shfl.sync.bfly.b32 %r279|%p44, %r277, %r262, %r251, %r253;
mov.b32 %f164, %r279;
add.ftz.f32 %f165, %f163, %f164;
mov.b32 %r280, %f165;
shfl.sync.bfly.b32 %r282|%p45, %r280, %r249, %r251, %r253;
mov.b32 %f166, %r282;
add.ftz.f32 %f36, %f165, %f166;
@%p22 bra $L__BB97_18;

st.global.f32 [%rd14+4], %f36;

$L__BB97_18:
mul.lo.s32 %r21, %r25, 3;
ld.shared.f32 %f167, [%r20+768];
add.ftz.f32 %f168, %f167, %f224;
mov.b32 %r283, %f168;
mov.u32 %r284, 31;
mov.u32 %r285, 16;
mov.u32 %r286, -1;
shfl.sync.bfly.b32 %r287|%p47, %r283, %r285, %r284, %r286;
mov.b32 %f169, %r287;
add.ftz.f32 %f170, %f168, %f169;
mov.b32 %r288, %f170;
mov.u32 %r289, 8;
shfl.sync.bfly.b32 %r290|%p48, %r288, %r289, %r284, %r286;
mov.b32 %f171, %r290;
add.ftz.f32 %f172, %f170, %f171;
mov.b32 %r291, %f172;
mov.u32 %r292, 4;
shfl.sync.bfly.b32 %r293|%p49, %r291, %r292, %r284, %r286;
mov.b32 %f173, %r293;
add.ftz.f32 %f174, %f172, %f173;
mov.b32 %r294, %f174;
mov.u32 %r295, 2;
shfl.sync.bfly.b32 %r296|%p50, %r294, %r295, %r284, %r286;
mov.b32 %f175, %r296;
add.ftz.f32 %f176, %f174, %f175;
mov.b32 %r297, %f176;
mov.u32 %r298, 1;
shfl.sync.bfly.b32 %r299|%p51, %r297, %r298, %r284, %r286;
mov.b32 %f177, %r299;
add.ftz.f32 %f37, %f176, %f177;
@%p14 bra $L__BB97_20;

cvt.u64.u32 %rd81, %r21;
add.s64 %rd82, %rd81, %rd12;
shl.b64 %rd84, %rd82, 2;
add.s64 %rd85, %rd65, %rd84;
st.global.f32 [%rd85], %f37;

$L__BB97_20:
ld.shared.f32 %f178, [%r20+896];
add.ftz.f32 %f179, %f178, %f223;
mov.b32 %r300, %f179;
shfl.sync.bfly.b32 %r304|%p53, %r300, %r285, %r284, %r286;
mov.b32 %f180, %r304;
add.ftz.f32 %f181, %f179, %f180;
mov.b32 %r305, %f181;
shfl.sync.bfly.b32 %r307|%p54, %r305, %r289, %r284, %r286;
mov.b32 %f182, %r307;
add.ftz.f32 %f183, %f181, %f182;
mov.b32 %r308, %f183;
shfl.sync.bfly.b32 %r310|%p55, %r308, %r292, %r284, %r286;
mov.b32 %f184, %r310;
add.ftz.f32 %f185, %f183, %f184;
mov.b32 %r311, %f185;
shfl.sync.bfly.b32 %r313|%p56, %r311, %r295, %r284, %r286;
mov.b32 %f186, %r313;
add.ftz.f32 %f187, %f185, %f186;
mov.b32 %r314, %f187;
shfl.sync.bfly.b32 %r316|%p57, %r314, %r298, %r284, %r286;
mov.b32 %f188, %r316;
add.ftz.f32 %f38, %f187, %f188;
@%p22 bra $L__BB97_22;

add.s32 %r317, %r21, 1;
cvt.u64.u32 %rd86, %r317;
add.s64 %rd87, %rd86, %rd12;
shl.b64 %rd89, %rd87, 2;
add.s64 %rd90, %rd65, %rd89;
st.global.f32 [%rd90], %f38;

$L__BB97_22:
shl.b32 %r318, %r25, 2;
mov.u32 %r319, 2;
ld.shared.f32 %f189, [%r20+1024];
add.ftz.f32 %f190, %f189, %f222;
mov.b32 %r320, %f190;
mov.u32 %r321, 31;
mov.u32 %r322, 16;
mov.u32 %r323, -1;
shfl.sync.bfly.b32 %r324|%p59, %r320, %r322, %r321, %r323;
mov.b32 %f191, %r324;
add.ftz.f32 %f192, %f190, %f191;
mov.b32 %r325, %f192;
mov.u32 %r326, 8;
shfl.sync.bfly.b32 %r327|%p60, %r325, %r326, %r321, %r323;
mov.b32 %f193, %r327;
add.ftz.f32 %f194, %f192, %f193;
mov.b32 %r328, %f194;
mov.u32 %r329, 4;
shfl.sync.bfly.b32 %r330|%p61, %r328, %r329, %r321, %r323;
mov.b32 %f195, %r330;
add.ftz.f32 %f196, %f194, %f195;
mov.b32 %r331, %f196;
shfl.sync.bfly.b32 %r332|%p62, %r331, %r319, %r321, %r323;
mov.b32 %f197, %r332;
add.ftz.f32 %f198, %f196, %f197;
mov.b32 %r333, %f198;
mov.u32 %r334, 1;
shfl.sync.bfly.b32 %r335|%p63, %r333, %r334, %r321, %r323;
mov.b32 %f199, %r335;
add.ftz.f32 %f39, %f198, %f199;
cvt.u64.u32 %rd91, %r318;
add.s64 %rd92, %rd91, %rd12;
shl.b64 %rd94, %rd92, 2;
add.s64 %rd15, %rd65, %rd94;
@%p14 bra $L__BB97_24;

st.global.f32 [%rd15], %f39;

$L__BB97_24:
ld.shared.f32 %f200, [%r20+1152];
add.ftz.f32 %f201, %f200, %f221;
mov.b32 %r336, %f201;
shfl.sync.bfly.b32 %r340|%p65, %r336, %r322, %r321, %r323;
mov.b32 %f202, %r340;
add.ftz.f32 %f203, %f201, %f202;
mov.b32 %r341, %f203;
shfl.sync.bfly.b32 %r343|%p66, %r341, %r326, %r321, %r323;
mov.b32 %f204, %r343;
add.ftz.f32 %f205, %f203, %f204;
mov.b32 %r344, %f205;
shfl.sync.bfly.b32 %r346|%p67, %r344, %r329, %r321, %r323;
mov.b32 %f206, %r346;
add.ftz.f32 %f207, %f205, %f206;
mov.b32 %r347, %f207;
shfl.sync.bfly.b32 %r349|%p68, %r347, %r319, %r321, %r323;
mov.b32 %f208, %r349;
add.ftz.f32 %f209, %f207, %f208;
mov.b32 %r350, %f209;
shfl.sync.bfly.b32 %r352|%p69, %r350, %r334, %r321, %r323;
mov.b32 %f210, %r352;
add.ftz.f32 %f40, %f209, %f210;
@%p22 bra $L__BB97_26;

st.global.f32 [%rd15+4], %f40;

$L__BB97_26:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type8ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type8ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type8ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type8ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type8ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type8ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type8ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type8ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type8ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<83>;
.reg .b16 %rs<45>;
.reg .f32 %f<277>;
.reg .b32 %r<427>;
.reg .b64 %rd<116>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type8ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type8ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type8ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type8ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type8ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type8ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type8ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type8ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type8ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type8ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type8ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type8ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type8ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type8ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type8ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type8ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type8ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type8ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r43, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r43, %r2;
mov.u32 %r44, %ctaid.x;
shl.b32 %r4, %r44, 1;
shr.s32 %r45, %r3, 31;
shr.u32 %r46, %r45, 30;
add.s32 %r47, %r3, %r46;
shr.s32 %r426, %r47, 2;
shr.u32 %r6, %r42, 5;
setp.ge.s32 %p3, %r426, %r6;
mov.f32 %f265, 0f00000000;
mov.f32 %f266, %f265;
mov.f32 %f267, %f265;
mov.f32 %f268, %f265;
mov.f32 %f269, %f265;
mov.f32 %f270, %f265;
mov.f32 %f271, %f265;
mov.f32 %f272, %f265;
mov.f32 %f273, %f265;
mov.f32 %f274, %f265;
mov.f32 %f275, %f265;
mov.f32 %f276, %f265;
@%p3 bra $L__BB98_3;

mov.u32 %r48, %ctaid.y;
mov.u32 %r49, %ctaid.z;
mul.lo.s32 %r50, %r48, %r34;
cvt.u64.u32 %rd18, %r50;
mul.lo.s32 %r51, %r49, %r40;
cvt.u64.u32 %rd19, %r51;
add.s64 %rd1, %rd19, %rd18;
mul.hi.u32 %r52, %r49, %r36;
add.s32 %r53, %r52, %r49;
shr.u32 %r54, %r53, %r37;
mul.lo.s32 %r55, %r54, %r39;
mul.hi.u32 %r56, %r48, %r30;
add.s32 %r57, %r56, %r48;
shr.u32 %r58, %r57, %r31;
mul.lo.s32 %r59, %r58, %r33;
cvta.to.global.u64 %rd115, %rd15;
and.b32 %r63, %r47, -4;
sub.s32 %r64, %r3, %r63;
shl.b32 %r65, %r64, 1;
shl.b32 %r66, %r64, 2;
or.b32 %r67, %r66, 1;
cvt.s64.s32 %rd3, %r65;
sub.s32 %r68, %r67, %r65;
cvt.s64.s32 %rd4, %r68;
shl.b32 %r69, %r28, 1;
add.s32 %r425, %r426, %r69;
mad.lo.s32 %r424, %r28, 3, %r426;
shl.b32 %r70, %r28, 2;
add.s32 %r423, %r426, %r70;
mad.lo.s32 %r422, %r28, 5, %r426;
add.s32 %r421, %r426, %r28;
add.s32 %r71, %r426, %r55;
add.s32 %r72, %r71, %r59;
add.s32 %r73, %r4, 1;
mad.lo.s32 %r74, %r27, %r73, %r72;
mul.wide.s32 %rd5, %r74, 34;
mad.lo.s32 %r75, %r4, %r27, %r426;
add.s32 %r76, %r75, %r55;
add.s32 %r77, %r76, %r59;
mul.wide.s32 %rd6, %r77, 34;
mul.wide.s32 %rd20, %r66, 2;
add.s64 %rd21, %rd6, %rd20;
add.s64 %rd7, %rd21, 4;
add.s64 %rd22, %rd5, %rd20;
add.s64 %rd8, %rd22, 2;
cvta.to.global.u64 %rd9, %rd16;
mov.f32 %f265, 0f00000000;
shl.b64 %rd29, %rd3, 2;
shl.b64 %rd31, %rd4, 2;

$L__BB98_2:
cvt.u64.u32 %rd23, %r426;
add.s64 %rd24, %rd1, %rd23;
mul.lo.s64 %rd25, %rd24, 36;
add.s64 %rd26, %rd9, %rd25;
add.s64 %rd27, %rd26, 4;
add.s64 %rd28, %rd115, %rd7;
ld.global.nc.u16 %rs37, [%rd28+-2];
ld.global.nc.u16 %rs38, [%rd28];
mov.b32 %r80, {%rs37, %rs38};
add.s64 %rd30, %rd27, %rd29;
ld.global.nc.u32 %r81, [%rd30];
ld.global.nc.u16 %rs39, [%rd28+2];
ld.global.nc.u16 %rs40, [%rd28+4];
mov.b32 %r84, {%rs39, %rs40};
add.s64 %rd32, %rd27, %rd31;
ld.global.nc.u32 %r85, [%rd32];
add.s64 %rd33, %rd115, %rd6;
ld.global.nc.u16 %rs1, [%rd33];

{ cvt.f32.f16 %f73, %rs1;}


ld.global.nc.u32 %r78, [%rd26];

{.reg .f16 low,high;
mov.b32 {low,high}, %r78;
mov.b16 %rs2, low;}


{ cvt.f32.f16 %f74, %rs2;}


mov.u32 %r181, 0;

dp4a.s32.s32 %r79, %r80, %r81, %r181;


dp4a.s32.s32 %r83, %r84, %r85, %r79;

mul.ftz.f32 %f97, %f73, %f74;
cvt.rn.f32.s32 %f98, %r83;
fma.rn.ftz.f32 %f276, %f97, %f98, %f276;
add.s64 %rd34, %rd115, %rd8;
ld.global.nc.u16 %rs41, [%rd34];
ld.global.nc.u16 %rs42, [%rd34+2];
mov.b32 %r89, {%rs41, %rs42};
ld.global.nc.u16 %rs43, [%rd34+4];
ld.global.nc.u16 %rs44, [%rd34+6];
mov.b32 %r93, {%rs43, %rs44};
add.s64 %rd35, %rd115, %rd5;
ld.global.nc.u16 %rs4, [%rd35];

{ cvt.f32.f16 %f75, %rs4;}



{.reg .f16 low,high;
mov.b32 {low,high}, %r78;
mov.b16 %rs5, low;}


{ cvt.f32.f16 %f76, %rs5;}



dp4a.s32.s32 %r88, %r89, %r81, %r181;


dp4a.s32.s32 %r92, %r93, %r85, %r88;

mul.ftz.f32 %f99, %f75, %f76;
cvt.rn.f32.s32 %f100, %r92;
fma.rn.ftz.f32 %f275, %f99, %f100, %f275;
cvt.u64.u32 %rd36, %r421;
add.s64 %rd37, %rd1, %rd36;
mul.lo.s64 %rd38, %rd37, 36;
add.s64 %rd39, %rd9, %rd38;
add.s64 %rd40, %rd39, 4;
add.s64 %rd41, %rd40, %rd29;
ld.global.nc.u32 %r99, [%rd41];
add.s64 %rd42, %rd40, %rd31;
ld.global.nc.u32 %r103, [%rd42];

{ cvt.f32.f16 %f77, %rs1;}


ld.global.nc.u32 %r96, [%rd39];

{.reg .f16 low,high;
mov.b32 {low,high}, %r96;
mov.b16 %rs8, low;}


{ cvt.f32.f16 %f78, %rs8;}



dp4a.s32.s32 %r97, %r80, %r99, %r181;


dp4a.s32.s32 %r101, %r84, %r103, %r97;

mul.ftz.f32 %f101, %f77, %f78;
cvt.rn.f32.s32 %f102, %r101;
fma.rn.ftz.f32 %f274, %f101, %f102, %f274;

{ cvt.f32.f16 %f79, %rs4;}



{.reg .f16 low,high;
mov.b32 {low,high}, %r96;
mov.b16 %rs11, low;}


{ cvt.f32.f16 %f80, %rs11;}



dp4a.s32.s32 %r106, %r89, %r99, %r181;


dp4a.s32.s32 %r110, %r93, %r103, %r106;

mul.ftz.f32 %f103, %f79, %f80;
cvt.rn.f32.s32 %f104, %r110;
fma.rn.ftz.f32 %f273, %f103, %f104, %f273;
cvt.u64.u32 %rd43, %r425;
add.s64 %rd44, %rd1, %rd43;
mul.lo.s64 %rd45, %rd44, 36;
add.s64 %rd46, %rd9, %rd45;
add.s64 %rd47, %rd46, 4;
add.s64 %rd48, %rd47, %rd29;
ld.global.nc.u32 %r117, [%rd48];
add.s64 %rd49, %rd47, %rd31;
ld.global.nc.u32 %r121, [%rd49];

{ cvt.f32.f16 %f81, %rs1;}


ld.global.nc.u32 %r114, [%rd46];

{.reg .f16 low,high;
mov.b32 {low,high}, %r114;
mov.b16 %rs14, low;}


{ cvt.f32.f16 %f82, %rs14;}



dp4a.s32.s32 %r115, %r80, %r117, %r181;


dp4a.s32.s32 %r119, %r84, %r121, %r115;

mul.ftz.f32 %f105, %f81, %f82;
cvt.rn.f32.s32 %f106, %r119;
fma.rn.ftz.f32 %f272, %f105, %f106, %f272;

{ cvt.f32.f16 %f83, %rs4;}



{.reg .f16 low,high;
mov.b32 {low,high}, %r114;
mov.b16 %rs17, low;}


{ cvt.f32.f16 %f84, %rs17;}



dp4a.s32.s32 %r124, %r89, %r117, %r181;


dp4a.s32.s32 %r128, %r93, %r121, %r124;

mul.ftz.f32 %f107, %f83, %f84;
cvt.rn.f32.s32 %f108, %r128;
fma.rn.ftz.f32 %f271, %f107, %f108, %f271;
cvt.u64.u32 %rd50, %r424;
add.s64 %rd51, %rd1, %rd50;
mul.lo.s64 %rd52, %rd51, 36;
add.s64 %rd53, %rd9, %rd52;
add.s64 %rd54, %rd53, 4;
add.s64 %rd55, %rd54, %rd29;
ld.global.nc.u32 %r135, [%rd55];
add.s64 %rd56, %rd54, %rd31;
ld.global.nc.u32 %r139, [%rd56];

{ cvt.f32.f16 %f85, %rs1;}


ld.global.nc.u32 %r132, [%rd53];

{.reg .f16 low,high;
mov.b32 {low,high}, %r132;
mov.b16 %rs20, low;}


{ cvt.f32.f16 %f86, %rs20;}



dp4a.s32.s32 %r133, %r80, %r135, %r181;


dp4a.s32.s32 %r137, %r84, %r139, %r133;

mul.ftz.f32 %f109, %f85, %f86;
cvt.rn.f32.s32 %f110, %r137;
fma.rn.ftz.f32 %f270, %f109, %f110, %f270;

{ cvt.f32.f16 %f87, %rs4;}



{.reg .f16 low,high;
mov.b32 {low,high}, %r132;
mov.b16 %rs23, low;}


{ cvt.f32.f16 %f88, %rs23;}



dp4a.s32.s32 %r142, %r89, %r135, %r181;


dp4a.s32.s32 %r146, %r93, %r139, %r142;

mul.ftz.f32 %f111, %f87, %f88;
cvt.rn.f32.s32 %f112, %r146;
fma.rn.ftz.f32 %f269, %f111, %f112, %f269;
cvt.u64.u32 %rd57, %r423;
add.s64 %rd58, %rd1, %rd57;
mul.lo.s64 %rd59, %rd58, 36;
add.s64 %rd60, %rd9, %rd59;
add.s64 %rd61, %rd60, 4;
add.s64 %rd62, %rd61, %rd29;
ld.global.nc.u32 %r153, [%rd62];
add.s64 %rd63, %rd61, %rd31;
ld.global.nc.u32 %r157, [%rd63];

{ cvt.f32.f16 %f89, %rs1;}


ld.global.nc.u32 %r150, [%rd60];

{.reg .f16 low,high;
mov.b32 {low,high}, %r150;
mov.b16 %rs26, low;}


{ cvt.f32.f16 %f90, %rs26;}



dp4a.s32.s32 %r151, %r80, %r153, %r181;


dp4a.s32.s32 %r155, %r84, %r157, %r151;

mul.ftz.f32 %f113, %f89, %f90;
cvt.rn.f32.s32 %f114, %r155;
fma.rn.ftz.f32 %f268, %f113, %f114, %f268;

{ cvt.f32.f16 %f91, %rs4;}



{.reg .f16 low,high;
mov.b32 {low,high}, %r150;
mov.b16 %rs29, low;}


{ cvt.f32.f16 %f92, %rs29;}



dp4a.s32.s32 %r160, %r89, %r153, %r181;


dp4a.s32.s32 %r164, %r93, %r157, %r160;

mul.ftz.f32 %f115, %f91, %f92;
cvt.rn.f32.s32 %f116, %r164;
fma.rn.ftz.f32 %f267, %f115, %f116, %f267;
cvt.u64.u32 %rd64, %r422;
add.s64 %rd65, %rd1, %rd64;
mul.lo.s64 %rd66, %rd65, 36;
add.s64 %rd67, %rd9, %rd66;
add.s64 %rd68, %rd67, 4;
add.s64 %rd69, %rd68, %rd29;
ld.global.nc.u32 %r171, [%rd69];
add.s64 %rd70, %rd68, %rd31;
ld.global.nc.u32 %r175, [%rd70];

{ cvt.f32.f16 %f93, %rs1;}


ld.global.nc.u32 %r168, [%rd67];

{.reg .f16 low,high;
mov.b32 {low,high}, %r168;
mov.b16 %rs32, low;}


{ cvt.f32.f16 %f94, %rs32;}



dp4a.s32.s32 %r169, %r80, %r171, %r181;


dp4a.s32.s32 %r173, %r84, %r175, %r169;

mul.ftz.f32 %f117, %f93, %f94;
cvt.rn.f32.s32 %f118, %r173;
fma.rn.ftz.f32 %f266, %f117, %f118, %f266;

{ cvt.f32.f16 %f95, %rs4;}



{.reg .f16 low,high;
mov.b32 {low,high}, %r168;
mov.b16 %rs35, low;}


{ cvt.f32.f16 %f96, %rs35;}



dp4a.s32.s32 %r178, %r89, %r171, %r181;


dp4a.s32.s32 %r182, %r93, %r175, %r178;

mul.ftz.f32 %f119, %f95, %f96;
cvt.rn.f32.s32 %f120, %r182;
fma.rn.ftz.f32 %f265, %f119, %f120, %f265;
add.s32 %r425, %r425, 16;
add.s32 %r424, %r424, 16;
add.s32 %r423, %r423, 16;
add.s32 %r422, %r422, 16;
add.s32 %r421, %r421, 16;
add.s64 %rd115, %rd115, 544;
add.s32 %r426, %r426, 16;
setp.lt.s32 %p4, %r426, %r6;
@%p4 bra $L__BB98_2;

$L__BB98_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB98_5;

mov.u32 %r186, _ZZ13mul_mat_vec_qIL9ggml_type8ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r187, %r1, 1536, %r186;
shl.b32 %r188, %r2, 2;
add.s32 %r189, %r187, %r188;
st.shared.f32 [%r189+-1536], %f276;
st.shared.f32 [%r189+-1408], %f275;
st.shared.f32 [%r189+-1280], %f274;
st.shared.f32 [%r189+-1152], %f273;
st.shared.f32 [%r189+-1024], %f272;
st.shared.f32 [%r189+-896], %f271;
st.shared.f32 [%r189+-768], %f270;
st.shared.f32 [%r189+-640], %f269;
st.shared.f32 [%r189+-512], %f268;
st.shared.f32 [%r189+-384], %f267;
st.shared.f32 [%r189+-256], %f266;
st.shared.f32 [%r189+-128], %f265;

$L__BB98_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB98_30;

mov.u32 %r190, %ctaid.z;
mov.u32 %r191, %ctaid.y;
mad.lo.s32 %r192, %r191, %r35, %r4;
mad.lo.s32 %r193, %r190, %r41, %r192;
cvt.u64.u32 %rd12, %r193;
shl.b32 %r194, %r2, 2;
mov.u32 %r195, _ZZ13mul_mat_vec_qIL9ggml_type8ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r24, %r195, %r194;
mov.u32 %r196, 2;
ld.shared.f32 %f121, [%r24];
add.ftz.f32 %f122, %f121, %f276;
mov.b32 %r197, %f122;
mov.u32 %r198, 31;
mov.u32 %r199, 16;
mov.u32 %r200, -1;
shfl.sync.bfly.b32 %r201|%p7, %r197, %r199, %r198, %r200;
mov.b32 %f123, %r201;
add.ftz.f32 %f124, %f122, %f123;
mov.b32 %r202, %f124;
mov.u32 %r203, 8;
shfl.sync.bfly.b32 %r204|%p8, %r202, %r203, %r198, %r200;
mov.b32 %f125, %r204;
add.ftz.f32 %f126, %f124, %f125;
mov.b32 %r205, %f126;
mov.u32 %r206, 4;
shfl.sync.bfly.b32 %r207|%p9, %r205, %r206, %r198, %r200;
mov.b32 %f127, %r207;
add.ftz.f32 %f128, %f126, %f127;
mov.b32 %r208, %f128;
shfl.sync.bfly.b32 %r209|%p10, %r208, %r196, %r198, %r200;
mov.b32 %f129, %r209;
add.ftz.f32 %f130, %f128, %f129;
mov.b32 %r210, %f130;
mov.u32 %r211, 1;
shfl.sync.bfly.b32 %r212|%p11, %r210, %r211, %r198, %r200;
mov.b32 %f131, %r212;
add.ftz.f32 %f37, %f130, %f131;
setp.lt.u32 %p12, %r4, %r29;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
not.pred %p14, %p1;
@%p14 bra $L__BB98_8;

cvta.to.global.u64 %rd71, %rd17;
mul.wide.u32 %rd72, %r193, 4;
add.s64 %rd73, %rd71, %rd72;
st.global.f32 [%rd73], %f37;

$L__BB98_8:
ld.shared.f32 %f132, [%r24+128];
add.ftz.f32 %f133, %f132, %f275;
mov.b32 %r219, %f133;
shfl.sync.bfly.b32 %r223|%p15, %r219, %r199, %r198, %r200;
mov.b32 %f134, %r223;
add.ftz.f32 %f135, %f133, %f134;
mov.b32 %r224, %f135;
shfl.sync.bfly.b32 %r226|%p16, %r224, %r203, %r198, %r200;
mov.b32 %f136, %r226;
add.ftz.f32 %f137, %f135, %f136;
mov.b32 %r227, %f137;
shfl.sync.bfly.b32 %r229|%p17, %r227, %r206, %r198, %r200;
mov.b32 %f138, %r229;
add.ftz.f32 %f139, %f137, %f138;
mov.b32 %r230, %f139;
shfl.sync.bfly.b32 %r232|%p18, %r230, %r196, %r198, %r200;
mov.b32 %f140, %r232;
add.ftz.f32 %f141, %f139, %f140;
mov.b32 %r233, %f141;
shfl.sync.bfly.b32 %r235|%p19, %r233, %r211, %r198, %r200;
mov.b32 %f142, %r235;
add.ftz.f32 %f38, %f141, %f142;
setp.eq.s32 %p20, %r2, 1;
or.b32 %r239, %r4, 1;
setp.lt.u32 %p21, %r239, %r29;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB98_10;

cvta.to.global.u64 %rd74, %rd17;
mul.wide.u32 %rd75, %r193, 4;
add.s64 %rd76, %rd74, %rd75;
st.global.f32 [%rd76+4], %f38;

$L__BB98_10:
ld.shared.f32 %f143, [%r24+256];
add.ftz.f32 %f144, %f143, %f274;
mov.b32 %r246, %f144;
mov.u32 %r247, 31;
mov.u32 %r248, 16;
mov.u32 %r249, -1;
shfl.sync.bfly.b32 %r250|%p23, %r246, %r248, %r247, %r249;
mov.b32 %f145, %r250;
add.ftz.f32 %f146, %f144, %f145;
mov.b32 %r251, %f146;
mov.u32 %r252, 8;
shfl.sync.bfly.b32 %r253|%p24, %r251, %r252, %r247, %r249;
mov.b32 %f147, %r253;
add.ftz.f32 %f148, %f146, %f147;
mov.b32 %r254, %f148;
mov.u32 %r255, 4;
shfl.sync.bfly.b32 %r256|%p25, %r254, %r255, %r247, %r249;
mov.b32 %f149, %r256;
add.ftz.f32 %f150, %f148, %f149;
mov.b32 %r257, %f150;
mov.u32 %r258, 2;
shfl.sync.bfly.b32 %r259|%p26, %r257, %r258, %r247, %r249;
mov.b32 %f151, %r259;
add.ftz.f32 %f152, %f150, %f151;
mov.b32 %r260, %f152;
mov.u32 %r261, 1;
shfl.sync.bfly.b32 %r262|%p27, %r260, %r261, %r247, %r249;
mov.b32 %f153, %r262;
add.ftz.f32 %f39, %f152, %f153;
@%p14 bra $L__BB98_12;

cvt.u64.u32 %rd77, %r29;
add.s64 %rd78, %rd77, %rd12;
cvta.to.global.u64 %rd79, %rd17;
shl.b64 %rd80, %rd78, 2;
add.s64 %rd81, %rd79, %rd80;
st.global.f32 [%rd81], %f39;

$L__BB98_12:
ld.shared.f32 %f154, [%r24+384];
add.ftz.f32 %f155, %f154, %f273;
mov.b32 %r263, %f155;
shfl.sync.bfly.b32 %r267|%p29, %r263, %r248, %r247, %r249;
mov.b32 %f156, %r267;
add.ftz.f32 %f157, %f155, %f156;
mov.b32 %r268, %f157;
shfl.sync.bfly.b32 %r270|%p30, %r268, %r252, %r247, %r249;
mov.b32 %f158, %r270;
add.ftz.f32 %f159, %f157, %f158;
mov.b32 %r271, %f159;
shfl.sync.bfly.b32 %r273|%p31, %r271, %r255, %r247, %r249;
mov.b32 %f160, %r273;
add.ftz.f32 %f161, %f159, %f160;
mov.b32 %r274, %f161;
shfl.sync.bfly.b32 %r276|%p32, %r274, %r258, %r247, %r249;
mov.b32 %f162, %r276;
add.ftz.f32 %f163, %f161, %f162;
mov.b32 %r277, %f163;
shfl.sync.bfly.b32 %r279|%p33, %r277, %r261, %r247, %r249;
mov.b32 %f164, %r279;
add.ftz.f32 %f40, %f163, %f164;
@%p22 bra $L__BB98_14;

add.s32 %r280, %r29, 1;
cvt.u64.u32 %rd82, %r280;
add.s64 %rd83, %rd82, %rd12;
cvta.to.global.u64 %rd84, %rd17;
shl.b64 %rd85, %rd83, 2;
add.s64 %rd86, %rd84, %rd85;
st.global.f32 [%rd86], %f40;

$L__BB98_14:
shl.b32 %r281, %r29, 1;
mov.u32 %r282, 1;
ld.shared.f32 %f165, [%r24+512];
add.ftz.f32 %f166, %f165, %f272;
mov.b32 %r283, %f166;
mov.u32 %r284, 31;
mov.u32 %r285, 16;
mov.u32 %r286, -1;
shfl.sync.bfly.b32 %r287|%p35, %r283, %r285, %r284, %r286;
mov.b32 %f167, %r287;
add.ftz.f32 %f168, %f166, %f167;
mov.b32 %r288, %f168;
mov.u32 %r289, 8;
shfl.sync.bfly.b32 %r290|%p36, %r288, %r289, %r284, %r286;
mov.b32 %f169, %r290;
add.ftz.f32 %f170, %f168, %f169;
mov.b32 %r291, %f170;
mov.u32 %r292, 4;
shfl.sync.bfly.b32 %r293|%p37, %r291, %r292, %r284, %r286;
mov.b32 %f171, %r293;
add.ftz.f32 %f172, %f170, %f171;
mov.b32 %r294, %f172;
mov.u32 %r295, 2;
shfl.sync.bfly.b32 %r296|%p38, %r294, %r295, %r284, %r286;
mov.b32 %f173, %r296;
add.ftz.f32 %f174, %f172, %f173;
mov.b32 %r297, %f174;
shfl.sync.bfly.b32 %r298|%p39, %r297, %r282, %r284, %r286;
mov.b32 %f175, %r298;
add.ftz.f32 %f41, %f174, %f175;
cvt.u64.u32 %rd87, %r281;
add.s64 %rd88, %rd87, %rd12;
cvta.to.global.u64 %rd89, %rd17;
shl.b64 %rd90, %rd88, 2;
add.s64 %rd13, %rd89, %rd90;
@%p14 bra $L__BB98_16;

st.global.f32 [%rd13], %f41;

$L__BB98_16:
ld.shared.f32 %f176, [%r24+640];
add.ftz.f32 %f177, %f176, %f271;
mov.b32 %r299, %f177;
shfl.sync.bfly.b32 %r303|%p41, %r299, %r285, %r284, %r286;
mov.b32 %f178, %r303;
add.ftz.f32 %f179, %f177, %f178;
mov.b32 %r304, %f179;
shfl.sync.bfly.b32 %r306|%p42, %r304, %r289, %r284, %r286;
mov.b32 %f180, %r306;
add.ftz.f32 %f181, %f179, %f180;
mov.b32 %r307, %f181;
shfl.sync.bfly.b32 %r309|%p43, %r307, %r292, %r284, %r286;
mov.b32 %f182, %r309;
add.ftz.f32 %f183, %f181, %f182;
mov.b32 %r310, %f183;
shfl.sync.bfly.b32 %r312|%p44, %r310, %r295, %r284, %r286;
mov.b32 %f184, %r312;
add.ftz.f32 %f185, %f183, %f184;
mov.b32 %r313, %f185;
shfl.sync.bfly.b32 %r315|%p45, %r313, %r282, %r284, %r286;
mov.b32 %f186, %r315;
add.ftz.f32 %f42, %f185, %f186;
@%p22 bra $L__BB98_18;

st.global.f32 [%rd13+4], %f42;

$L__BB98_18:
mul.lo.s32 %r25, %r29, 3;
ld.shared.f32 %f187, [%r24+768];
add.ftz.f32 %f188, %f187, %f270;
mov.b32 %r316, %f188;
mov.u32 %r317, 31;
mov.u32 %r318, 16;
mov.u32 %r319, -1;
shfl.sync.bfly.b32 %r320|%p47, %r316, %r318, %r317, %r319;
mov.b32 %f189, %r320;
add.ftz.f32 %f190, %f188, %f189;
mov.b32 %r321, %f190;
mov.u32 %r322, 8;
shfl.sync.bfly.b32 %r323|%p48, %r321, %r322, %r317, %r319;
mov.b32 %f191, %r323;
add.ftz.f32 %f192, %f190, %f191;
mov.b32 %r324, %f192;
mov.u32 %r325, 4;
shfl.sync.bfly.b32 %r326|%p49, %r324, %r325, %r317, %r319;
mov.b32 %f193, %r326;
add.ftz.f32 %f194, %f192, %f193;
mov.b32 %r327, %f194;
mov.u32 %r328, 2;
shfl.sync.bfly.b32 %r329|%p50, %r327, %r328, %r317, %r319;
mov.b32 %f195, %r329;
add.ftz.f32 %f196, %f194, %f195;
mov.b32 %r330, %f196;
mov.u32 %r331, 1;
shfl.sync.bfly.b32 %r332|%p51, %r330, %r331, %r317, %r319;
mov.b32 %f197, %r332;
add.ftz.f32 %f43, %f196, %f197;
@%p14 bra $L__BB98_20;

cvt.u64.u32 %rd91, %r25;
add.s64 %rd92, %rd91, %rd12;
shl.b64 %rd94, %rd92, 2;
add.s64 %rd95, %rd89, %rd94;
st.global.f32 [%rd95], %f43;

$L__BB98_20:
ld.shared.f32 %f198, [%r24+896];
add.ftz.f32 %f199, %f198, %f269;
mov.b32 %r333, %f199;
shfl.sync.bfly.b32 %r337|%p53, %r333, %r318, %r317, %r319;
mov.b32 %f200, %r337;
add.ftz.f32 %f201, %f199, %f200;
mov.b32 %r338, %f201;
shfl.sync.bfly.b32 %r340|%p54, %r338, %r322, %r317, %r319;
mov.b32 %f202, %r340;
add.ftz.f32 %f203, %f201, %f202;
mov.b32 %r341, %f203;
shfl.sync.bfly.b32 %r343|%p55, %r341, %r325, %r317, %r319;
mov.b32 %f204, %r343;
add.ftz.f32 %f205, %f203, %f204;
mov.b32 %r344, %f205;
shfl.sync.bfly.b32 %r346|%p56, %r344, %r328, %r317, %r319;
mov.b32 %f206, %r346;
add.ftz.f32 %f207, %f205, %f206;
mov.b32 %r347, %f207;
shfl.sync.bfly.b32 %r349|%p57, %r347, %r331, %r317, %r319;
mov.b32 %f208, %r349;
add.ftz.f32 %f44, %f207, %f208;
@%p22 bra $L__BB98_22;

add.s32 %r350, %r25, 1;
cvt.u64.u32 %rd96, %r350;
add.s64 %rd97, %rd96, %rd12;
shl.b64 %rd99, %rd97, 2;
add.s64 %rd100, %rd89, %rd99;
st.global.f32 [%rd100], %f44;

$L__BB98_22:
shl.b32 %r351, %r29, 2;
mov.u32 %r352, 2;
ld.shared.f32 %f209, [%r24+1024];
add.ftz.f32 %f210, %f209, %f268;
mov.b32 %r353, %f210;
mov.u32 %r354, 31;
mov.u32 %r355, 16;
mov.u32 %r356, -1;
shfl.sync.bfly.b32 %r357|%p59, %r353, %r355, %r354, %r356;
mov.b32 %f211, %r357;
add.ftz.f32 %f212, %f210, %f211;
mov.b32 %r358, %f212;
mov.u32 %r359, 8;
shfl.sync.bfly.b32 %r360|%p60, %r358, %r359, %r354, %r356;
mov.b32 %f213, %r360;
add.ftz.f32 %f214, %f212, %f213;
mov.b32 %r361, %f214;
mov.u32 %r362, 4;
shfl.sync.bfly.b32 %r363|%p61, %r361, %r362, %r354, %r356;
mov.b32 %f215, %r363;
add.ftz.f32 %f216, %f214, %f215;
mov.b32 %r364, %f216;
shfl.sync.bfly.b32 %r365|%p62, %r364, %r352, %r354, %r356;
mov.b32 %f217, %r365;
add.ftz.f32 %f218, %f216, %f217;
mov.b32 %r366, %f218;
mov.u32 %r367, 1;
shfl.sync.bfly.b32 %r368|%p63, %r366, %r367, %r354, %r356;
mov.b32 %f219, %r368;
add.ftz.f32 %f45, %f218, %f219;
cvt.u64.u32 %rd101, %r351;
add.s64 %rd102, %rd101, %rd12;
shl.b64 %rd104, %rd102, 2;
add.s64 %rd14, %rd89, %rd104;
@%p14 bra $L__BB98_24;

st.global.f32 [%rd14], %f45;

$L__BB98_24:
ld.shared.f32 %f220, [%r24+1152];
add.ftz.f32 %f221, %f220, %f267;
mov.b32 %r369, %f221;
shfl.sync.bfly.b32 %r373|%p65, %r369, %r355, %r354, %r356;
mov.b32 %f222, %r373;
add.ftz.f32 %f223, %f221, %f222;
mov.b32 %r374, %f223;
shfl.sync.bfly.b32 %r376|%p66, %r374, %r359, %r354, %r356;
mov.b32 %f224, %r376;
add.ftz.f32 %f225, %f223, %f224;
mov.b32 %r377, %f225;
shfl.sync.bfly.b32 %r379|%p67, %r377, %r362, %r354, %r356;
mov.b32 %f226, %r379;
add.ftz.f32 %f227, %f225, %f226;
mov.b32 %r380, %f227;
shfl.sync.bfly.b32 %r382|%p68, %r380, %r352, %r354, %r356;
mov.b32 %f228, %r382;
add.ftz.f32 %f229, %f227, %f228;
mov.b32 %r383, %f229;
shfl.sync.bfly.b32 %r385|%p69, %r383, %r367, %r354, %r356;
mov.b32 %f230, %r385;
add.ftz.f32 %f46, %f229, %f230;
@%p22 bra $L__BB98_26;

st.global.f32 [%rd14+4], %f46;

$L__BB98_26:
mul.lo.s32 %r26, %r29, 5;
ld.shared.f32 %f231, [%r24+1280];
add.ftz.f32 %f232, %f231, %f266;
mov.b32 %r386, %f232;
mov.u32 %r387, 31;
mov.u32 %r388, 16;
mov.u32 %r389, -1;
shfl.sync.bfly.b32 %r390|%p71, %r386, %r388, %r387, %r389;
mov.b32 %f233, %r390;
add.ftz.f32 %f234, %f232, %f233;
mov.b32 %r391, %f234;
mov.u32 %r392, 8;
shfl.sync.bfly.b32 %r393|%p72, %r391, %r392, %r387, %r389;
mov.b32 %f235, %r393;
add.ftz.f32 %f236, %f234, %f235;
mov.b32 %r394, %f236;
mov.u32 %r395, 4;
shfl.sync.bfly.b32 %r396|%p73, %r394, %r395, %r387, %r389;
mov.b32 %f237, %r396;
add.ftz.f32 %f238, %f236, %f237;
mov.b32 %r397, %f238;
mov.u32 %r398, 2;
shfl.sync.bfly.b32 %r399|%p74, %r397, %r398, %r387, %r389;
mov.b32 %f239, %r399;
add.ftz.f32 %f240, %f238, %f239;
mov.b32 %r400, %f240;
mov.u32 %r401, 1;
shfl.sync.bfly.b32 %r402|%p75, %r400, %r401, %r387, %r389;
mov.b32 %f241, %r402;
add.ftz.f32 %f47, %f240, %f241;
@%p14 bra $L__BB98_28;

cvt.u64.u32 %rd105, %r26;
add.s64 %rd106, %rd105, %rd12;
shl.b64 %rd108, %rd106, 2;
add.s64 %rd109, %rd89, %rd108;
st.global.f32 [%rd109], %f47;

$L__BB98_28:
ld.shared.f32 %f242, [%r24+1408];
add.ftz.f32 %f243, %f242, %f265;
mov.b32 %r403, %f243;
shfl.sync.bfly.b32 %r407|%p77, %r403, %r388, %r387, %r389;
mov.b32 %f244, %r407;
add.ftz.f32 %f245, %f243, %f244;
mov.b32 %r408, %f245;
shfl.sync.bfly.b32 %r410|%p78, %r408, %r392, %r387, %r389;
mov.b32 %f246, %r410;
add.ftz.f32 %f247, %f245, %f246;
mov.b32 %r411, %f247;
shfl.sync.bfly.b32 %r413|%p79, %r411, %r395, %r387, %r389;
mov.b32 %f248, %r413;
add.ftz.f32 %f249, %f247, %f248;
mov.b32 %r414, %f249;
shfl.sync.bfly.b32 %r416|%p80, %r414, %r398, %r387, %r389;
mov.b32 %f250, %r416;
add.ftz.f32 %f251, %f249, %f250;
mov.b32 %r417, %f251;
shfl.sync.bfly.b32 %r419|%p81, %r417, %r401, %r387, %r389;
mov.b32 %f252, %r419;
add.ftz.f32 %f48, %f251, %f252;
@%p22 bra $L__BB98_30;

add.s32 %r420, %r26, 1;
cvt.u64.u32 %rd110, %r420;
add.s64 %rd111, %rd110, %rd12;
shl.b64 %rd113, %rd111, 2;
add.s64 %rd114, %rd89, %rd113;
st.global.f32 [%rd114], %f48;

$L__BB98_30:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type8ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type8ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type8ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type8ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type8ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type8ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type8ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type8ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type8ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<95>;
.reg .b16 %rs<51>;
.reg .f32 %f<323>;
.reg .b32 %r<472>;
.reg .b64 %rd<125>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type8ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1792];

ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type8ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type8ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type8ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r45, [_Z13mul_mat_vec_qIL9ggml_type8ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type8ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type8ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type8ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type8ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type8ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type8ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type8ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type8ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type8ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type8ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type8ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type8ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type8ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r46, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r46, %r2;
mov.u32 %r47, %ctaid.x;
shl.b32 %r4, %r47, 1;
shr.s32 %r48, %r3, 31;
shr.u32 %r49, %r48, 30;
add.s32 %r50, %r3, %r49;
shr.s32 %r471, %r50, 2;
shr.u32 %r6, %r45, 5;
setp.ge.s32 %p3, %r471, %r6;
mov.f32 %f309, 0f00000000;
mov.f32 %f310, %f309;
mov.f32 %f311, %f309;
mov.f32 %f312, %f309;
mov.f32 %f313, %f309;
mov.f32 %f314, %f309;
mov.f32 %f315, %f309;
mov.f32 %f316, %f309;
mov.f32 %f317, %f309;
mov.f32 %f318, %f309;
mov.f32 %f319, %f309;
mov.f32 %f320, %f309;
mov.f32 %f321, %f309;
mov.f32 %f322, %f309;
@%p3 bra $L__BB99_3;

mov.u32 %r51, %ctaid.y;
mov.u32 %r52, %ctaid.z;
mul.lo.s32 %r53, %r51, %r37;
cvt.u64.u32 %rd20, %r53;
mul.lo.s32 %r54, %r52, %r43;
cvt.u64.u32 %rd21, %r54;
add.s64 %rd1, %rd21, %rd20;
mul.hi.u32 %r55, %r52, %r39;
add.s32 %r56, %r55, %r52;
shr.u32 %r57, %r56, %r40;
mul.lo.s32 %r58, %r57, %r42;
mul.hi.u32 %r59, %r51, %r33;
add.s32 %r60, %r59, %r51;
shr.u32 %r61, %r60, %r34;
mul.lo.s32 %r62, %r61, %r36;
cvta.to.global.u64 %rd124, %rd17;
and.b32 %r66, %r50, -4;
sub.s32 %r67, %r3, %r66;
shl.b32 %r68, %r67, 1;
shl.b32 %r69, %r67, 2;
or.b32 %r70, %r69, 1;
cvt.s64.s32 %rd3, %r68;
sub.s32 %r71, %r70, %r68;
cvt.s64.s32 %rd4, %r71;
shl.b32 %r72, %r31, 1;
add.s32 %r470, %r471, %r72;
mad.lo.s32 %r469, %r31, 3, %r471;
shl.b32 %r73, %r31, 2;
add.s32 %r468, %r471, %r73;
mad.lo.s32 %r467, %r31, 5, %r471;
mad.lo.s32 %r466, %r31, 6, %r471;
add.s32 %r465, %r471, %r31;
add.s32 %r74, %r471, %r58;
add.s32 %r75, %r74, %r62;
add.s32 %r76, %r4, 1;
mad.lo.s32 %r77, %r30, %r76, %r75;
mul.wide.s32 %rd5, %r77, 34;
mad.lo.s32 %r78, %r4, %r30, %r471;
add.s32 %r79, %r78, %r58;
add.s32 %r80, %r79, %r62;
mul.wide.s32 %rd6, %r80, 34;
mul.wide.s32 %rd22, %r69, 2;
add.s64 %rd23, %rd6, %rd22;
add.s64 %rd7, %rd23, 4;
add.s64 %rd24, %rd5, %rd22;
add.s64 %rd8, %rd24, 2;
cvta.to.global.u64 %rd9, %rd18;
mov.f32 %f309, 0f00000000;
shl.b64 %rd31, %rd3, 2;
shl.b64 %rd33, %rd4, 2;

$L__BB99_2:
cvt.u64.u32 %rd25, %r471;
add.s64 %rd26, %rd1, %rd25;
mul.lo.s64 %rd27, %rd26, 36;
add.s64 %rd28, %rd9, %rd27;
add.s64 %rd29, %rd28, 4;
add.s64 %rd30, %rd124, %rd7;
ld.global.nc.u16 %rs43, [%rd30+-2];
ld.global.nc.u16 %rs44, [%rd30];
mov.b32 %r83, {%rs43, %rs44};
add.s64 %rd32, %rd29, %rd31;
ld.global.nc.u32 %r84, [%rd32];
ld.global.nc.u16 %rs45, [%rd30+2];
ld.global.nc.u16 %rs46, [%rd30+4];
mov.b32 %r87, {%rs45, %rs46};
add.s64 %rd34, %rd29, %rd33;
ld.global.nc.u32 %r88, [%rd34];
add.s64 %rd35, %rd124, %rd6;
ld.global.nc.u16 %rs1, [%rd35];

{ cvt.f32.f16 %f85, %rs1;}


ld.global.nc.u32 %r81, [%rd28];

{.reg .f16 low,high;
mov.b32 {low,high}, %r81;
mov.b16 %rs2, low;}


{ cvt.f32.f16 %f86, %rs2;}


mov.u32 %r202, 0;

dp4a.s32.s32 %r82, %r83, %r84, %r202;


dp4a.s32.s32 %r86, %r87, %r88, %r82;

mul.ftz.f32 %f113, %f85, %f86;
cvt.rn.f32.s32 %f114, %r86;
fma.rn.ftz.f32 %f322, %f113, %f114, %f322;
add.s64 %rd36, %rd124, %rd8;
ld.global.nc.u16 %rs47, [%rd36];
ld.global.nc.u16 %rs48, [%rd36+2];
mov.b32 %r92, {%rs47, %rs48};
ld.global.nc.u16 %rs49, [%rd36+4];
ld.global.nc.u16 %rs50, [%rd36+6];
mov.b32 %r96, {%rs49, %rs50};
add.s64 %rd37, %rd124, %rd5;
ld.global.nc.u16 %rs4, [%rd37];

{ cvt.f32.f16 %f87, %rs4;}



{.reg .f16 low,high;
mov.b32 {low,high}, %r81;
mov.b16 %rs5, low;}


{ cvt.f32.f16 %f88, %rs5;}



dp4a.s32.s32 %r91, %r92, %r84, %r202;


dp4a.s32.s32 %r95, %r96, %r88, %r91;

mul.ftz.f32 %f115, %f87, %f88;
cvt.rn.f32.s32 %f116, %r95;
fma.rn.ftz.f32 %f321, %f115, %f116, %f321;
cvt.u64.u32 %rd38, %r465;
add.s64 %rd39, %rd1, %rd38;
mul.lo.s64 %rd40, %rd39, 36;
add.s64 %rd41, %rd9, %rd40;
add.s64 %rd42, %rd41, 4;
add.s64 %rd43, %rd42, %rd31;
ld.global.nc.u32 %r102, [%rd43];
add.s64 %rd44, %rd42, %rd33;
ld.global.nc.u32 %r106, [%rd44];

{ cvt.f32.f16 %f89, %rs1;}


ld.global.nc.u32 %r99, [%rd41];

{.reg .f16 low,high;
mov.b32 {low,high}, %r99;
mov.b16 %rs8, low;}


{ cvt.f32.f16 %f90, %rs8;}



dp4a.s32.s32 %r100, %r83, %r102, %r202;


dp4a.s32.s32 %r104, %r87, %r106, %r100;

mul.ftz.f32 %f117, %f89, %f90;
cvt.rn.f32.s32 %f118, %r104;
fma.rn.ftz.f32 %f320, %f117, %f118, %f320;

{ cvt.f32.f16 %f91, %rs4;}



{.reg .f16 low,high;
mov.b32 {low,high}, %r99;
mov.b16 %rs11, low;}


{ cvt.f32.f16 %f92, %rs11;}



dp4a.s32.s32 %r109, %r92, %r102, %r202;


dp4a.s32.s32 %r113, %r96, %r106, %r109;

mul.ftz.f32 %f119, %f91, %f92;
cvt.rn.f32.s32 %f120, %r113;
fma.rn.ftz.f32 %f319, %f119, %f120, %f319;
cvt.u64.u32 %rd45, %r470;
add.s64 %rd46, %rd1, %rd45;
mul.lo.s64 %rd47, %rd46, 36;
add.s64 %rd48, %rd9, %rd47;
add.s64 %rd49, %rd48, 4;
add.s64 %rd50, %rd49, %rd31;
ld.global.nc.u32 %r120, [%rd50];
add.s64 %rd51, %rd49, %rd33;
ld.global.nc.u32 %r124, [%rd51];

{ cvt.f32.f16 %f93, %rs1;}


ld.global.nc.u32 %r117, [%rd48];

{.reg .f16 low,high;
mov.b32 {low,high}, %r117;
mov.b16 %rs14, low;}


{ cvt.f32.f16 %f94, %rs14;}



dp4a.s32.s32 %r118, %r83, %r120, %r202;


dp4a.s32.s32 %r122, %r87, %r124, %r118;

mul.ftz.f32 %f121, %f93, %f94;
cvt.rn.f32.s32 %f122, %r122;
fma.rn.ftz.f32 %f318, %f121, %f122, %f318;

{ cvt.f32.f16 %f95, %rs4;}



{.reg .f16 low,high;
mov.b32 {low,high}, %r117;
mov.b16 %rs17, low;}


{ cvt.f32.f16 %f96, %rs17;}



dp4a.s32.s32 %r127, %r92, %r120, %r202;


dp4a.s32.s32 %r131, %r96, %r124, %r127;

mul.ftz.f32 %f123, %f95, %f96;
cvt.rn.f32.s32 %f124, %r131;
fma.rn.ftz.f32 %f317, %f123, %f124, %f317;
cvt.u64.u32 %rd52, %r469;
add.s64 %rd53, %rd1, %rd52;
mul.lo.s64 %rd54, %rd53, 36;
add.s64 %rd55, %rd9, %rd54;
add.s64 %rd56, %rd55, 4;
add.s64 %rd57, %rd56, %rd31;
ld.global.nc.u32 %r138, [%rd57];
add.s64 %rd58, %rd56, %rd33;
ld.global.nc.u32 %r142, [%rd58];

{ cvt.f32.f16 %f97, %rs1;}


ld.global.nc.u32 %r135, [%rd55];

{.reg .f16 low,high;
mov.b32 {low,high}, %r135;
mov.b16 %rs20, low;}


{ cvt.f32.f16 %f98, %rs20;}



dp4a.s32.s32 %r136, %r83, %r138, %r202;


dp4a.s32.s32 %r140, %r87, %r142, %r136;

mul.ftz.f32 %f125, %f97, %f98;
cvt.rn.f32.s32 %f126, %r140;
fma.rn.ftz.f32 %f316, %f125, %f126, %f316;

{ cvt.f32.f16 %f99, %rs4;}



{.reg .f16 low,high;
mov.b32 {low,high}, %r135;
mov.b16 %rs23, low;}


{ cvt.f32.f16 %f100, %rs23;}



dp4a.s32.s32 %r145, %r92, %r138, %r202;


dp4a.s32.s32 %r149, %r96, %r142, %r145;

mul.ftz.f32 %f127, %f99, %f100;
cvt.rn.f32.s32 %f128, %r149;
fma.rn.ftz.f32 %f315, %f127, %f128, %f315;
cvt.u64.u32 %rd59, %r468;
add.s64 %rd60, %rd1, %rd59;
mul.lo.s64 %rd61, %rd60, 36;
add.s64 %rd62, %rd9, %rd61;
add.s64 %rd63, %rd62, 4;
add.s64 %rd64, %rd63, %rd31;
ld.global.nc.u32 %r156, [%rd64];
add.s64 %rd65, %rd63, %rd33;
ld.global.nc.u32 %r160, [%rd65];

{ cvt.f32.f16 %f101, %rs1;}


ld.global.nc.u32 %r153, [%rd62];

{.reg .f16 low,high;
mov.b32 {low,high}, %r153;
mov.b16 %rs26, low;}


{ cvt.f32.f16 %f102, %rs26;}



dp4a.s32.s32 %r154, %r83, %r156, %r202;


dp4a.s32.s32 %r158, %r87, %r160, %r154;

mul.ftz.f32 %f129, %f101, %f102;
cvt.rn.f32.s32 %f130, %r158;
fma.rn.ftz.f32 %f314, %f129, %f130, %f314;

{ cvt.f32.f16 %f103, %rs4;}



{.reg .f16 low,high;
mov.b32 {low,high}, %r153;
mov.b16 %rs29, low;}


{ cvt.f32.f16 %f104, %rs29;}



dp4a.s32.s32 %r163, %r92, %r156, %r202;


dp4a.s32.s32 %r167, %r96, %r160, %r163;

mul.ftz.f32 %f131, %f103, %f104;
cvt.rn.f32.s32 %f132, %r167;
fma.rn.ftz.f32 %f313, %f131, %f132, %f313;
cvt.u64.u32 %rd66, %r467;
add.s64 %rd67, %rd1, %rd66;
mul.lo.s64 %rd68, %rd67, 36;
add.s64 %rd69, %rd9, %rd68;
add.s64 %rd70, %rd69, 4;
add.s64 %rd71, %rd70, %rd31;
ld.global.nc.u32 %r174, [%rd71];
add.s64 %rd72, %rd70, %rd33;
ld.global.nc.u32 %r178, [%rd72];

{ cvt.f32.f16 %f105, %rs1;}


ld.global.nc.u32 %r171, [%rd69];

{.reg .f16 low,high;
mov.b32 {low,high}, %r171;
mov.b16 %rs32, low;}


{ cvt.f32.f16 %f106, %rs32;}



dp4a.s32.s32 %r172, %r83, %r174, %r202;


dp4a.s32.s32 %r176, %r87, %r178, %r172;

mul.ftz.f32 %f133, %f105, %f106;
cvt.rn.f32.s32 %f134, %r176;
fma.rn.ftz.f32 %f312, %f133, %f134, %f312;

{ cvt.f32.f16 %f107, %rs4;}



{.reg .f16 low,high;
mov.b32 {low,high}, %r171;
mov.b16 %rs35, low;}


{ cvt.f32.f16 %f108, %rs35;}



dp4a.s32.s32 %r181, %r92, %r174, %r202;


dp4a.s32.s32 %r185, %r96, %r178, %r181;

mul.ftz.f32 %f135, %f107, %f108;
cvt.rn.f32.s32 %f136, %r185;
fma.rn.ftz.f32 %f311, %f135, %f136, %f311;
cvt.u64.u32 %rd73, %r466;
add.s64 %rd74, %rd1, %rd73;
mul.lo.s64 %rd75, %rd74, 36;
add.s64 %rd76, %rd9, %rd75;
add.s64 %rd77, %rd76, 4;
add.s64 %rd78, %rd77, %rd31;
ld.global.nc.u32 %r192, [%rd78];
add.s64 %rd79, %rd77, %rd33;
ld.global.nc.u32 %r196, [%rd79];

{ cvt.f32.f16 %f109, %rs1;}


ld.global.nc.u32 %r189, [%rd76];

{.reg .f16 low,high;
mov.b32 {low,high}, %r189;
mov.b16 %rs38, low;}


{ cvt.f32.f16 %f110, %rs38;}



dp4a.s32.s32 %r190, %r83, %r192, %r202;


dp4a.s32.s32 %r194, %r87, %r196, %r190;

mul.ftz.f32 %f137, %f109, %f110;
cvt.rn.f32.s32 %f138, %r194;
fma.rn.ftz.f32 %f310, %f137, %f138, %f310;

{ cvt.f32.f16 %f111, %rs4;}



{.reg .f16 low,high;
mov.b32 {low,high}, %r189;
mov.b16 %rs41, low;}


{ cvt.f32.f16 %f112, %rs41;}



dp4a.s32.s32 %r199, %r92, %r192, %r202;


dp4a.s32.s32 %r203, %r96, %r196, %r199;

mul.ftz.f32 %f139, %f111, %f112;
cvt.rn.f32.s32 %f140, %r203;
fma.rn.ftz.f32 %f309, %f139, %f140, %f309;
add.s32 %r470, %r470, 16;
add.s32 %r469, %r469, 16;
add.s32 %r468, %r468, 16;
add.s32 %r467, %r467, 16;
add.s32 %r466, %r466, 16;
add.s32 %r465, %r465, 16;
add.s64 %rd124, %rd124, 544;
add.s32 %r471, %r471, 16;
setp.lt.s32 %p4, %r471, %r6;
@%p4 bra $L__BB99_2;

$L__BB99_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB99_5;

mov.u32 %r207, _ZZ13mul_mat_vec_qIL9ggml_type8ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r208, %r1, 1792, %r207;
shl.b32 %r209, %r2, 2;
add.s32 %r210, %r208, %r209;
st.shared.f32 [%r210+-1792], %f322;
st.shared.f32 [%r210+-1664], %f321;
st.shared.f32 [%r210+-1536], %f320;
st.shared.f32 [%r210+-1408], %f319;
st.shared.f32 [%r210+-1280], %f318;
st.shared.f32 [%r210+-1152], %f317;
st.shared.f32 [%r210+-1024], %f316;
st.shared.f32 [%r210+-896], %f315;
st.shared.f32 [%r210+-768], %f314;
st.shared.f32 [%r210+-640], %f313;
st.shared.f32 [%r210+-512], %f312;
st.shared.f32 [%r210+-384], %f311;
st.shared.f32 [%r210+-256], %f310;
st.shared.f32 [%r210+-128], %f309;

$L__BB99_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB99_34;

mov.u32 %r211, %ctaid.z;
mov.u32 %r212, %ctaid.y;
mad.lo.s32 %r213, %r212, %r38, %r4;
mad.lo.s32 %r214, %r211, %r44, %r213;
cvt.u64.u32 %rd12, %r214;
shl.b32 %r215, %r2, 2;
mov.u32 %r216, _ZZ13mul_mat_vec_qIL9ggml_type8ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r27, %r216, %r215;
mov.u32 %r217, 2;
ld.shared.f32 %f141, [%r27];
add.ftz.f32 %f142, %f141, %f322;
mov.b32 %r218, %f142;
mov.u32 %r219, 31;
mov.u32 %r220, 16;
mov.u32 %r221, -1;
shfl.sync.bfly.b32 %r222|%p7, %r218, %r220, %r219, %r221;
mov.b32 %f143, %r222;
add.ftz.f32 %f144, %f142, %f143;
mov.b32 %r223, %f144;
mov.u32 %r224, 8;
shfl.sync.bfly.b32 %r225|%p8, %r223, %r224, %r219, %r221;
mov.b32 %f145, %r225;
add.ftz.f32 %f146, %f144, %f145;
mov.b32 %r226, %f146;
mov.u32 %r227, 4;
shfl.sync.bfly.b32 %r228|%p9, %r226, %r227, %r219, %r221;
mov.b32 %f147, %r228;
add.ftz.f32 %f148, %f146, %f147;
mov.b32 %r229, %f148;
shfl.sync.bfly.b32 %r230|%p10, %r229, %r217, %r219, %r221;
mov.b32 %f149, %r230;
add.ftz.f32 %f150, %f148, %f149;
mov.b32 %r231, %f150;
mov.u32 %r232, 1;
shfl.sync.bfly.b32 %r233|%p11, %r231, %r232, %r219, %r221;
mov.b32 %f151, %r233;
add.ftz.f32 %f43, %f150, %f151;
setp.lt.u32 %p12, %r4, %r32;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
cvta.to.global.u64 %rd80, %rd19;
mul.wide.u32 %rd81, %r214, 4;
add.s64 %rd13, %rd80, %rd81;
not.pred %p14, %p1;
@%p14 bra $L__BB99_8;

st.global.f32 [%rd13], %f43;

$L__BB99_8:
ld.shared.f32 %f152, [%r27+128];
add.ftz.f32 %f153, %f152, %f321;
mov.b32 %r234, %f153;
shfl.sync.bfly.b32 %r238|%p15, %r234, %r220, %r219, %r221;
mov.b32 %f154, %r238;
add.ftz.f32 %f155, %f153, %f154;
mov.b32 %r239, %f155;
shfl.sync.bfly.b32 %r241|%p16, %r239, %r224, %r219, %r221;
mov.b32 %f156, %r241;
add.ftz.f32 %f157, %f155, %f156;
mov.b32 %r242, %f157;
shfl.sync.bfly.b32 %r244|%p17, %r242, %r227, %r219, %r221;
mov.b32 %f158, %r244;
add.ftz.f32 %f159, %f157, %f158;
mov.b32 %r245, %f159;
shfl.sync.bfly.b32 %r247|%p18, %r245, %r217, %r219, %r221;
mov.b32 %f160, %r247;
add.ftz.f32 %f161, %f159, %f160;
mov.b32 %r248, %f161;
shfl.sync.bfly.b32 %r250|%p19, %r248, %r232, %r219, %r221;
mov.b32 %f162, %r250;
add.ftz.f32 %f44, %f161, %f162;
setp.eq.s32 %p20, %r2, 1;
or.b32 %r254, %r4, 1;
setp.lt.u32 %p21, %r254, %r32;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB99_10;

st.global.f32 [%rd13+4], %f44;

$L__BB99_10:
ld.shared.f32 %f163, [%r27+256];
add.ftz.f32 %f164, %f163, %f320;
mov.b32 %r255, %f164;
mov.u32 %r256, 31;
mov.u32 %r257, 16;
mov.u32 %r258, -1;
shfl.sync.bfly.b32 %r259|%p23, %r255, %r257, %r256, %r258;
mov.b32 %f165, %r259;
add.ftz.f32 %f166, %f164, %f165;
mov.b32 %r260, %f166;
mov.u32 %r261, 8;
shfl.sync.bfly.b32 %r262|%p24, %r260, %r261, %r256, %r258;
mov.b32 %f167, %r262;
add.ftz.f32 %f168, %f166, %f167;
mov.b32 %r263, %f168;
mov.u32 %r264, 4;
shfl.sync.bfly.b32 %r265|%p25, %r263, %r264, %r256, %r258;
mov.b32 %f169, %r265;
add.ftz.f32 %f170, %f168, %f169;
mov.b32 %r266, %f170;
mov.u32 %r267, 2;
shfl.sync.bfly.b32 %r268|%p26, %r266, %r267, %r256, %r258;
mov.b32 %f171, %r268;
add.ftz.f32 %f172, %f170, %f171;
mov.b32 %r269, %f172;
mov.u32 %r270, 1;
shfl.sync.bfly.b32 %r271|%p27, %r269, %r270, %r256, %r258;
mov.b32 %f173, %r271;
add.ftz.f32 %f45, %f172, %f173;
@%p14 bra $L__BB99_12;

cvt.u64.u32 %rd82, %r32;
add.s64 %rd83, %rd82, %rd12;
shl.b64 %rd85, %rd83, 2;
add.s64 %rd86, %rd80, %rd85;
st.global.f32 [%rd86], %f45;

$L__BB99_12:
ld.shared.f32 %f174, [%r27+384];
add.ftz.f32 %f175, %f174, %f319;
mov.b32 %r272, %f175;
shfl.sync.bfly.b32 %r276|%p29, %r272, %r257, %r256, %r258;
mov.b32 %f176, %r276;
add.ftz.f32 %f177, %f175, %f176;
mov.b32 %r277, %f177;
shfl.sync.bfly.b32 %r279|%p30, %r277, %r261, %r256, %r258;
mov.b32 %f178, %r279;
add.ftz.f32 %f179, %f177, %f178;
mov.b32 %r280, %f179;
shfl.sync.bfly.b32 %r282|%p31, %r280, %r264, %r256, %r258;
mov.b32 %f180, %r282;
add.ftz.f32 %f181, %f179, %f180;
mov.b32 %r283, %f181;
shfl.sync.bfly.b32 %r285|%p32, %r283, %r267, %r256, %r258;
mov.b32 %f182, %r285;
add.ftz.f32 %f183, %f181, %f182;
mov.b32 %r286, %f183;
shfl.sync.bfly.b32 %r288|%p33, %r286, %r270, %r256, %r258;
mov.b32 %f184, %r288;
add.ftz.f32 %f46, %f183, %f184;
@%p22 bra $L__BB99_14;

add.s32 %r289, %r32, 1;
cvt.u64.u32 %rd87, %r289;
add.s64 %rd88, %rd87, %rd12;
shl.b64 %rd90, %rd88, 2;
add.s64 %rd91, %rd80, %rd90;
st.global.f32 [%rd91], %f46;

$L__BB99_14:
shl.b32 %r290, %r32, 1;
mov.u32 %r291, 1;
ld.shared.f32 %f185, [%r27+512];
add.ftz.f32 %f186, %f185, %f318;
mov.b32 %r292, %f186;
mov.u32 %r293, 31;
mov.u32 %r294, 16;
mov.u32 %r295, -1;
shfl.sync.bfly.b32 %r296|%p35, %r292, %r294, %r293, %r295;
mov.b32 %f187, %r296;
add.ftz.f32 %f188, %f186, %f187;
mov.b32 %r297, %f188;
mov.u32 %r298, 8;
shfl.sync.bfly.b32 %r299|%p36, %r297, %r298, %r293, %r295;
mov.b32 %f189, %r299;
add.ftz.f32 %f190, %f188, %f189;
mov.b32 %r300, %f190;
mov.u32 %r301, 4;
shfl.sync.bfly.b32 %r302|%p37, %r300, %r301, %r293, %r295;
mov.b32 %f191, %r302;
add.ftz.f32 %f192, %f190, %f191;
mov.b32 %r303, %f192;
mov.u32 %r304, 2;
shfl.sync.bfly.b32 %r305|%p38, %r303, %r304, %r293, %r295;
mov.b32 %f193, %r305;
add.ftz.f32 %f194, %f192, %f193;
mov.b32 %r306, %f194;
shfl.sync.bfly.b32 %r307|%p39, %r306, %r291, %r293, %r295;
mov.b32 %f195, %r307;
add.ftz.f32 %f47, %f194, %f195;
cvt.u64.u32 %rd92, %r290;
add.s64 %rd93, %rd92, %rd12;
shl.b64 %rd95, %rd93, 2;
add.s64 %rd14, %rd80, %rd95;
@%p14 bra $L__BB99_16;

st.global.f32 [%rd14], %f47;

$L__BB99_16:
ld.shared.f32 %f196, [%r27+640];
add.ftz.f32 %f197, %f196, %f317;
mov.b32 %r308, %f197;
shfl.sync.bfly.b32 %r312|%p41, %r308, %r294, %r293, %r295;
mov.b32 %f198, %r312;
add.ftz.f32 %f199, %f197, %f198;
mov.b32 %r313, %f199;
shfl.sync.bfly.b32 %r315|%p42, %r313, %r298, %r293, %r295;
mov.b32 %f200, %r315;
add.ftz.f32 %f201, %f199, %f200;
mov.b32 %r316, %f201;
shfl.sync.bfly.b32 %r318|%p43, %r316, %r301, %r293, %r295;
mov.b32 %f202, %r318;
add.ftz.f32 %f203, %f201, %f202;
mov.b32 %r319, %f203;
shfl.sync.bfly.b32 %r321|%p44, %r319, %r304, %r293, %r295;
mov.b32 %f204, %r321;
add.ftz.f32 %f205, %f203, %f204;
mov.b32 %r322, %f205;
shfl.sync.bfly.b32 %r324|%p45, %r322, %r291, %r293, %r295;
mov.b32 %f206, %r324;
add.ftz.f32 %f48, %f205, %f206;
@%p22 bra $L__BB99_18;

st.global.f32 [%rd14+4], %f48;

$L__BB99_18:
mul.lo.s32 %r28, %r32, 3;
ld.shared.f32 %f207, [%r27+768];
add.ftz.f32 %f208, %f207, %f316;
mov.b32 %r325, %f208;
mov.u32 %r326, 31;
mov.u32 %r327, 16;
mov.u32 %r328, -1;
shfl.sync.bfly.b32 %r329|%p47, %r325, %r327, %r326, %r328;
mov.b32 %f209, %r329;
add.ftz.f32 %f210, %f208, %f209;
mov.b32 %r330, %f210;
mov.u32 %r331, 8;
shfl.sync.bfly.b32 %r332|%p48, %r330, %r331, %r326, %r328;
mov.b32 %f211, %r332;
add.ftz.f32 %f212, %f210, %f211;
mov.b32 %r333, %f212;
mov.u32 %r334, 4;
shfl.sync.bfly.b32 %r335|%p49, %r333, %r334, %r326, %r328;
mov.b32 %f213, %r335;
add.ftz.f32 %f214, %f212, %f213;
mov.b32 %r336, %f214;
mov.u32 %r337, 2;
shfl.sync.bfly.b32 %r338|%p50, %r336, %r337, %r326, %r328;
mov.b32 %f215, %r338;
add.ftz.f32 %f216, %f214, %f215;
mov.b32 %r339, %f216;
mov.u32 %r340, 1;
shfl.sync.bfly.b32 %r341|%p51, %r339, %r340, %r326, %r328;
mov.b32 %f217, %r341;
add.ftz.f32 %f49, %f216, %f217;
@%p14 bra $L__BB99_20;

cvt.u64.u32 %rd96, %r28;
add.s64 %rd97, %rd96, %rd12;
shl.b64 %rd99, %rd97, 2;
add.s64 %rd100, %rd80, %rd99;
st.global.f32 [%rd100], %f49;

$L__BB99_20:
ld.shared.f32 %f218, [%r27+896];
add.ftz.f32 %f219, %f218, %f315;
mov.b32 %r342, %f219;
shfl.sync.bfly.b32 %r346|%p53, %r342, %r327, %r326, %r328;
mov.b32 %f220, %r346;
add.ftz.f32 %f221, %f219, %f220;
mov.b32 %r347, %f221;
shfl.sync.bfly.b32 %r349|%p54, %r347, %r331, %r326, %r328;
mov.b32 %f222, %r349;
add.ftz.f32 %f223, %f221, %f222;
mov.b32 %r350, %f223;
shfl.sync.bfly.b32 %r352|%p55, %r350, %r334, %r326, %r328;
mov.b32 %f224, %r352;
add.ftz.f32 %f225, %f223, %f224;
mov.b32 %r353, %f225;
shfl.sync.bfly.b32 %r355|%p56, %r353, %r337, %r326, %r328;
mov.b32 %f226, %r355;
add.ftz.f32 %f227, %f225, %f226;
mov.b32 %r356, %f227;
shfl.sync.bfly.b32 %r358|%p57, %r356, %r340, %r326, %r328;
mov.b32 %f228, %r358;
add.ftz.f32 %f50, %f227, %f228;
@%p22 bra $L__BB99_22;

add.s32 %r359, %r28, 1;
cvt.u64.u32 %rd101, %r359;
add.s64 %rd102, %rd101, %rd12;
shl.b64 %rd104, %rd102, 2;
add.s64 %rd105, %rd80, %rd104;
st.global.f32 [%rd105], %f50;

$L__BB99_22:
shl.b32 %r360, %r32, 2;
mov.u32 %r361, 2;
ld.shared.f32 %f229, [%r27+1024];
add.ftz.f32 %f230, %f229, %f314;
mov.b32 %r362, %f230;
mov.u32 %r363, 31;
mov.u32 %r364, 16;
mov.u32 %r365, -1;
shfl.sync.bfly.b32 %r366|%p59, %r362, %r364, %r363, %r365;
mov.b32 %f231, %r366;
add.ftz.f32 %f232, %f230, %f231;
mov.b32 %r367, %f232;
mov.u32 %r368, 8;
shfl.sync.bfly.b32 %r369|%p60, %r367, %r368, %r363, %r365;
mov.b32 %f233, %r369;
add.ftz.f32 %f234, %f232, %f233;
mov.b32 %r370, %f234;
mov.u32 %r371, 4;
shfl.sync.bfly.b32 %r372|%p61, %r370, %r371, %r363, %r365;
mov.b32 %f235, %r372;
add.ftz.f32 %f236, %f234, %f235;
mov.b32 %r373, %f236;
shfl.sync.bfly.b32 %r374|%p62, %r373, %r361, %r363, %r365;
mov.b32 %f237, %r374;
add.ftz.f32 %f238, %f236, %f237;
mov.b32 %r375, %f238;
mov.u32 %r376, 1;
shfl.sync.bfly.b32 %r377|%p63, %r375, %r376, %r363, %r365;
mov.b32 %f239, %r377;
add.ftz.f32 %f51, %f238, %f239;
cvt.u64.u32 %rd106, %r360;
add.s64 %rd107, %rd106, %rd12;
shl.b64 %rd109, %rd107, 2;
add.s64 %rd15, %rd80, %rd109;
@%p14 bra $L__BB99_24;

st.global.f32 [%rd15], %f51;

$L__BB99_24:
ld.shared.f32 %f240, [%r27+1152];
add.ftz.f32 %f241, %f240, %f313;
mov.b32 %r378, %f241;
shfl.sync.bfly.b32 %r382|%p65, %r378, %r364, %r363, %r365;
mov.b32 %f242, %r382;
add.ftz.f32 %f243, %f241, %f242;
mov.b32 %r383, %f243;
shfl.sync.bfly.b32 %r385|%p66, %r383, %r368, %r363, %r365;
mov.b32 %f244, %r385;
add.ftz.f32 %f245, %f243, %f244;
mov.b32 %r386, %f245;
shfl.sync.bfly.b32 %r388|%p67, %r386, %r371, %r363, %r365;
mov.b32 %f246, %r388;
add.ftz.f32 %f247, %f245, %f246;
mov.b32 %r389, %f247;
shfl.sync.bfly.b32 %r391|%p68, %r389, %r361, %r363, %r365;
mov.b32 %f248, %r391;
add.ftz.f32 %f249, %f247, %f248;
mov.b32 %r392, %f249;
shfl.sync.bfly.b32 %r394|%p69, %r392, %r376, %r363, %r365;
mov.b32 %f250, %r394;
add.ftz.f32 %f52, %f249, %f250;
@%p22 bra $L__BB99_26;

st.global.f32 [%rd15+4], %f52;

$L__BB99_26:
mul.lo.s32 %r29, %r32, 5;
ld.shared.f32 %f251, [%r27+1280];
add.ftz.f32 %f252, %f251, %f312;
mov.b32 %r395, %f252;
mov.u32 %r396, 31;
mov.u32 %r397, 16;
mov.u32 %r398, -1;
shfl.sync.bfly.b32 %r399|%p71, %r395, %r397, %r396, %r398;
mov.b32 %f253, %r399;
add.ftz.f32 %f254, %f252, %f253;
mov.b32 %r400, %f254;
mov.u32 %r401, 8;
shfl.sync.bfly.b32 %r402|%p72, %r400, %r401, %r396, %r398;
mov.b32 %f255, %r402;
add.ftz.f32 %f256, %f254, %f255;
mov.b32 %r403, %f256;
mov.u32 %r404, 4;
shfl.sync.bfly.b32 %r405|%p73, %r403, %r404, %r396, %r398;
mov.b32 %f257, %r405;
add.ftz.f32 %f258, %f256, %f257;
mov.b32 %r406, %f258;
mov.u32 %r407, 2;
shfl.sync.bfly.b32 %r408|%p74, %r406, %r407, %r396, %r398;
mov.b32 %f259, %r408;
add.ftz.f32 %f260, %f258, %f259;
mov.b32 %r409, %f260;
mov.u32 %r410, 1;
shfl.sync.bfly.b32 %r411|%p75, %r409, %r410, %r396, %r398;
mov.b32 %f261, %r411;
add.ftz.f32 %f53, %f260, %f261;
@%p14 bra $L__BB99_28;

cvt.u64.u32 %rd110, %r29;
add.s64 %rd111, %rd110, %rd12;
shl.b64 %rd113, %rd111, 2;
add.s64 %rd114, %rd80, %rd113;
st.global.f32 [%rd114], %f53;

$L__BB99_28:
ld.shared.f32 %f262, [%r27+1408];
add.ftz.f32 %f263, %f262, %f311;
mov.b32 %r412, %f263;
shfl.sync.bfly.b32 %r416|%p77, %r412, %r397, %r396, %r398;
mov.b32 %f264, %r416;
add.ftz.f32 %f265, %f263, %f264;
mov.b32 %r417, %f265;
shfl.sync.bfly.b32 %r419|%p78, %r417, %r401, %r396, %r398;
mov.b32 %f266, %r419;
add.ftz.f32 %f267, %f265, %f266;
mov.b32 %r420, %f267;
shfl.sync.bfly.b32 %r422|%p79, %r420, %r404, %r396, %r398;
mov.b32 %f268, %r422;
add.ftz.f32 %f269, %f267, %f268;
mov.b32 %r423, %f269;
shfl.sync.bfly.b32 %r425|%p80, %r423, %r407, %r396, %r398;
mov.b32 %f270, %r425;
add.ftz.f32 %f271, %f269, %f270;
mov.b32 %r426, %f271;
shfl.sync.bfly.b32 %r428|%p81, %r426, %r410, %r396, %r398;
mov.b32 %f272, %r428;
add.ftz.f32 %f54, %f271, %f272;
@%p22 bra $L__BB99_30;

add.s32 %r429, %r29, 1;
cvt.u64.u32 %rd115, %r429;
add.s64 %rd116, %rd115, %rd12;
shl.b64 %rd118, %rd116, 2;
add.s64 %rd119, %rd80, %rd118;
st.global.f32 [%rd119], %f54;

$L__BB99_30:
mul.lo.s32 %r430, %r32, 6;
ld.shared.f32 %f273, [%r27+1536];
add.ftz.f32 %f274, %f273, %f310;
mov.b32 %r431, %f274;
mov.u32 %r432, 31;
mov.u32 %r433, 16;
mov.u32 %r434, -1;
shfl.sync.bfly.b32 %r435|%p83, %r431, %r433, %r432, %r434;
mov.b32 %f275, %r435;
add.ftz.f32 %f276, %f274, %f275;
mov.b32 %r436, %f276;
mov.u32 %r437, 8;
shfl.sync.bfly.b32 %r438|%p84, %r436, %r437, %r432, %r434;
mov.b32 %f277, %r438;
add.ftz.f32 %f278, %f276, %f277;
mov.b32 %r439, %f278;
mov.u32 %r440, 4;
shfl.sync.bfly.b32 %r441|%p85, %r439, %r440, %r432, %r434;
mov.b32 %f279, %r441;
add.ftz.f32 %f280, %f278, %f279;
mov.b32 %r442, %f280;
mov.u32 %r443, 2;
shfl.sync.bfly.b32 %r444|%p86, %r442, %r443, %r432, %r434;
mov.b32 %f281, %r444;
add.ftz.f32 %f282, %f280, %f281;
mov.b32 %r445, %f282;
mov.u32 %r446, 1;
shfl.sync.bfly.b32 %r447|%p87, %r445, %r446, %r432, %r434;
mov.b32 %f283, %r447;
add.ftz.f32 %f55, %f282, %f283;
cvt.u64.u32 %rd120, %r430;
add.s64 %rd121, %rd120, %rd12;
shl.b64 %rd123, %rd121, 2;
add.s64 %rd16, %rd80, %rd123;
@%p14 bra $L__BB99_32;

st.global.f32 [%rd16], %f55;

$L__BB99_32:
ld.shared.f32 %f284, [%r27+1664];
add.ftz.f32 %f285, %f284, %f309;
mov.b32 %r448, %f285;
shfl.sync.bfly.b32 %r452|%p89, %r448, %r433, %r432, %r434;
mov.b32 %f286, %r452;
add.ftz.f32 %f287, %f285, %f286;
mov.b32 %r453, %f287;
shfl.sync.bfly.b32 %r455|%p90, %r453, %r437, %r432, %r434;
mov.b32 %f288, %r455;
add.ftz.f32 %f289, %f287, %f288;
mov.b32 %r456, %f289;
shfl.sync.bfly.b32 %r458|%p91, %r456, %r440, %r432, %r434;
mov.b32 %f290, %r458;
add.ftz.f32 %f291, %f289, %f290;
mov.b32 %r459, %f291;
shfl.sync.bfly.b32 %r461|%p92, %r459, %r443, %r432, %r434;
mov.b32 %f292, %r461;
add.ftz.f32 %f293, %f291, %f292;
mov.b32 %r462, %f293;
shfl.sync.bfly.b32 %r464|%p93, %r462, %r446, %r432, %r434;
mov.b32 %f294, %r464;
add.ftz.f32 %f56, %f293, %f294;
@%p22 bra $L__BB99_34;

st.global.f32 [%rd16+4], %f56;

$L__BB99_34:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type8ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type8ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type8ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type8ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type8ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type8ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type8ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type8ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type8ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type8ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<107>;
.reg .b16 %rs<57>;
.reg .f32 %f<369>;
.reg .b32 %r<531>;
.reg .b64 %rd<142>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type8ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[2048];

ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type8ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type8ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type8ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r49, [_Z13mul_mat_vec_qIL9ggml_type8ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type8ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type8ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type8ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type8ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type8ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type8ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r46, [_Z13mul_mat_vec_qIL9ggml_type8ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r47, [_Z13mul_mat_vec_qIL9ggml_type8ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r48, [_Z13mul_mat_vec_qIL9ggml_type8ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type8ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type8ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type8ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type8ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r50, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r50, %r2;
mov.u32 %r51, %ctaid.x;
shl.b32 %r4, %r51, 1;
shr.s32 %r52, %r3, 31;
shr.u32 %r53, %r52, 30;
add.s32 %r54, %r3, %r53;
shr.s32 %r530, %r54, 2;
shr.u32 %r6, %r49, 5;
setp.ge.s32 %p3, %r530, %r6;
mov.f32 %f353, 0f00000000;
mov.f32 %f354, %f353;
mov.f32 %f355, %f353;
mov.f32 %f356, %f353;
mov.f32 %f357, %f353;
mov.f32 %f358, %f353;
mov.f32 %f359, %f353;
mov.f32 %f360, %f353;
mov.f32 %f361, %f353;
mov.f32 %f362, %f353;
mov.f32 %f363, %f353;
mov.f32 %f364, %f353;
mov.f32 %f365, %f353;
mov.f32 %f366, %f353;
mov.f32 %f367, %f353;
mov.f32 %f368, %f353;
@%p3 bra $L__BB100_3;

mov.u32 %r55, %ctaid.y;
mov.u32 %r56, %ctaid.z;
mul.lo.s32 %r57, %r55, %r41;
cvt.u64.u32 %rd20, %r57;
mul.lo.s32 %r58, %r56, %r47;
cvt.u64.u32 %rd21, %r58;
add.s64 %rd1, %rd21, %rd20;
mul.hi.u32 %r59, %r56, %r43;
add.s32 %r60, %r59, %r56;
shr.u32 %r61, %r60, %r44;
mul.lo.s32 %r62, %r61, %r46;
mul.hi.u32 %r63, %r55, %r37;
add.s32 %r64, %r63, %r55;
shr.u32 %r65, %r64, %r38;
mul.lo.s32 %r66, %r65, %r40;
cvta.to.global.u64 %rd141, %rd17;
and.b32 %r70, %r54, -4;
sub.s32 %r71, %r3, %r70;
shl.b32 %r72, %r71, 1;
shl.b32 %r73, %r71, 2;
or.b32 %r74, %r73, 1;
cvt.s64.s32 %rd3, %r72;
sub.s32 %r75, %r74, %r72;
cvt.s64.s32 %rd4, %r75;
shl.b32 %r76, %r35, 1;
add.s32 %r529, %r530, %r76;
mad.lo.s32 %r528, %r35, 3, %r530;
shl.b32 %r77, %r35, 2;
add.s32 %r527, %r530, %r77;
mad.lo.s32 %r526, %r35, 5, %r530;
mad.lo.s32 %r525, %r35, 6, %r530;
mad.lo.s32 %r524, %r35, 7, %r530;
add.s32 %r523, %r530, %r35;
add.s32 %r78, %r530, %r62;
add.s32 %r79, %r78, %r66;
add.s32 %r80, %r4, 1;
mad.lo.s32 %r81, %r34, %r80, %r79;
mul.wide.s32 %rd5, %r81, 34;
mad.lo.s32 %r82, %r4, %r34, %r530;
add.s32 %r83, %r82, %r62;
add.s32 %r84, %r83, %r66;
mul.wide.s32 %rd6, %r84, 34;
mul.wide.s32 %rd22, %r73, 2;
add.s64 %rd23, %rd6, %rd22;
add.s64 %rd7, %rd23, 4;
add.s64 %rd24, %rd5, %rd22;
add.s64 %rd8, %rd24, 2;
cvta.to.global.u64 %rd9, %rd18;
mov.f32 %f353, 0f00000000;
shl.b64 %rd31, %rd3, 2;
shl.b64 %rd33, %rd4, 2;

$L__BB100_2:
cvt.u64.u32 %rd25, %r530;
add.s64 %rd26, %rd1, %rd25;
mul.lo.s64 %rd27, %rd26, 36;
add.s64 %rd28, %rd9, %rd27;
add.s64 %rd29, %rd28, 4;
add.s64 %rd30, %rd141, %rd7;
ld.global.nc.u16 %rs49, [%rd30+-2];
ld.global.nc.u16 %rs50, [%rd30];
mov.b32 %r87, {%rs49, %rs50};
add.s64 %rd32, %rd29, %rd31;
ld.global.nc.u32 %r88, [%rd32];
ld.global.nc.u16 %rs51, [%rd30+2];
ld.global.nc.u16 %rs52, [%rd30+4];
mov.b32 %r91, {%rs51, %rs52};
add.s64 %rd34, %rd29, %rd33;
ld.global.nc.u32 %r92, [%rd34];
add.s64 %rd35, %rd141, %rd6;
ld.global.nc.u16 %rs1, [%rd35];

{ cvt.f32.f16 %f97, %rs1;}


ld.global.nc.u32 %r85, [%rd28];

{.reg .f16 low,high;
mov.b32 {low,high}, %r85;
mov.b16 %rs2, low;}


{ cvt.f32.f16 %f98, %rs2;}


mov.u32 %r224, 0;

dp4a.s32.s32 %r86, %r87, %r88, %r224;


dp4a.s32.s32 %r90, %r91, %r92, %r86;

mul.ftz.f32 %f129, %f97, %f98;
cvt.rn.f32.s32 %f130, %r90;
fma.rn.ftz.f32 %f368, %f129, %f130, %f368;
add.s64 %rd36, %rd141, %rd8;
ld.global.nc.u16 %rs53, [%rd36];
ld.global.nc.u16 %rs54, [%rd36+2];
mov.b32 %r96, {%rs53, %rs54};
ld.global.nc.u16 %rs55, [%rd36+4];
ld.global.nc.u16 %rs56, [%rd36+6];
mov.b32 %r100, {%rs55, %rs56};
add.s64 %rd37, %rd141, %rd5;
ld.global.nc.u16 %rs4, [%rd37];

{ cvt.f32.f16 %f99, %rs4;}



{.reg .f16 low,high;
mov.b32 {low,high}, %r85;
mov.b16 %rs5, low;}


{ cvt.f32.f16 %f100, %rs5;}



dp4a.s32.s32 %r95, %r96, %r88, %r224;


dp4a.s32.s32 %r99, %r100, %r92, %r95;

mul.ftz.f32 %f131, %f99, %f100;
cvt.rn.f32.s32 %f132, %r99;
fma.rn.ftz.f32 %f367, %f131, %f132, %f367;
cvt.u64.u32 %rd38, %r523;
add.s64 %rd39, %rd1, %rd38;
mul.lo.s64 %rd40, %rd39, 36;
add.s64 %rd41, %rd9, %rd40;
add.s64 %rd42, %rd41, 4;
add.s64 %rd43, %rd42, %rd31;
ld.global.nc.u32 %r106, [%rd43];
add.s64 %rd44, %rd42, %rd33;
ld.global.nc.u32 %r110, [%rd44];

{ cvt.f32.f16 %f101, %rs1;}


ld.global.nc.u32 %r103, [%rd41];

{.reg .f16 low,high;
mov.b32 {low,high}, %r103;
mov.b16 %rs8, low;}


{ cvt.f32.f16 %f102, %rs8;}



dp4a.s32.s32 %r104, %r87, %r106, %r224;


dp4a.s32.s32 %r108, %r91, %r110, %r104;

mul.ftz.f32 %f133, %f101, %f102;
cvt.rn.f32.s32 %f134, %r108;
fma.rn.ftz.f32 %f366, %f133, %f134, %f366;

{ cvt.f32.f16 %f103, %rs4;}



{.reg .f16 low,high;
mov.b32 {low,high}, %r103;
mov.b16 %rs11, low;}


{ cvt.f32.f16 %f104, %rs11;}



dp4a.s32.s32 %r113, %r96, %r106, %r224;


dp4a.s32.s32 %r117, %r100, %r110, %r113;

mul.ftz.f32 %f135, %f103, %f104;
cvt.rn.f32.s32 %f136, %r117;
fma.rn.ftz.f32 %f365, %f135, %f136, %f365;
cvt.u64.u32 %rd45, %r529;
add.s64 %rd46, %rd1, %rd45;
mul.lo.s64 %rd47, %rd46, 36;
add.s64 %rd48, %rd9, %rd47;
add.s64 %rd49, %rd48, 4;
add.s64 %rd50, %rd49, %rd31;
ld.global.nc.u32 %r124, [%rd50];
add.s64 %rd51, %rd49, %rd33;
ld.global.nc.u32 %r128, [%rd51];

{ cvt.f32.f16 %f105, %rs1;}


ld.global.nc.u32 %r121, [%rd48];

{.reg .f16 low,high;
mov.b32 {low,high}, %r121;
mov.b16 %rs14, low;}


{ cvt.f32.f16 %f106, %rs14;}



dp4a.s32.s32 %r122, %r87, %r124, %r224;


dp4a.s32.s32 %r126, %r91, %r128, %r122;

mul.ftz.f32 %f137, %f105, %f106;
cvt.rn.f32.s32 %f138, %r126;
fma.rn.ftz.f32 %f364, %f137, %f138, %f364;

{ cvt.f32.f16 %f107, %rs4;}



{.reg .f16 low,high;
mov.b32 {low,high}, %r121;
mov.b16 %rs17, low;}


{ cvt.f32.f16 %f108, %rs17;}



dp4a.s32.s32 %r131, %r96, %r124, %r224;


dp4a.s32.s32 %r135, %r100, %r128, %r131;

mul.ftz.f32 %f139, %f107, %f108;
cvt.rn.f32.s32 %f140, %r135;
fma.rn.ftz.f32 %f363, %f139, %f140, %f363;
cvt.u64.u32 %rd52, %r528;
add.s64 %rd53, %rd1, %rd52;
mul.lo.s64 %rd54, %rd53, 36;
add.s64 %rd55, %rd9, %rd54;
add.s64 %rd56, %rd55, 4;
add.s64 %rd57, %rd56, %rd31;
ld.global.nc.u32 %r142, [%rd57];
add.s64 %rd58, %rd56, %rd33;
ld.global.nc.u32 %r146, [%rd58];

{ cvt.f32.f16 %f109, %rs1;}


ld.global.nc.u32 %r139, [%rd55];

{.reg .f16 low,high;
mov.b32 {low,high}, %r139;
mov.b16 %rs20, low;}


{ cvt.f32.f16 %f110, %rs20;}



dp4a.s32.s32 %r140, %r87, %r142, %r224;


dp4a.s32.s32 %r144, %r91, %r146, %r140;

mul.ftz.f32 %f141, %f109, %f110;
cvt.rn.f32.s32 %f142, %r144;
fma.rn.ftz.f32 %f362, %f141, %f142, %f362;

{ cvt.f32.f16 %f111, %rs4;}



{.reg .f16 low,high;
mov.b32 {low,high}, %r139;
mov.b16 %rs23, low;}


{ cvt.f32.f16 %f112, %rs23;}



dp4a.s32.s32 %r149, %r96, %r142, %r224;


dp4a.s32.s32 %r153, %r100, %r146, %r149;

mul.ftz.f32 %f143, %f111, %f112;
cvt.rn.f32.s32 %f144, %r153;
fma.rn.ftz.f32 %f361, %f143, %f144, %f361;
cvt.u64.u32 %rd59, %r527;
add.s64 %rd60, %rd1, %rd59;
mul.lo.s64 %rd61, %rd60, 36;
add.s64 %rd62, %rd9, %rd61;
add.s64 %rd63, %rd62, 4;
add.s64 %rd64, %rd63, %rd31;
ld.global.nc.u32 %r160, [%rd64];
add.s64 %rd65, %rd63, %rd33;
ld.global.nc.u32 %r164, [%rd65];

{ cvt.f32.f16 %f113, %rs1;}


ld.global.nc.u32 %r157, [%rd62];

{.reg .f16 low,high;
mov.b32 {low,high}, %r157;
mov.b16 %rs26, low;}


{ cvt.f32.f16 %f114, %rs26;}



dp4a.s32.s32 %r158, %r87, %r160, %r224;


dp4a.s32.s32 %r162, %r91, %r164, %r158;

mul.ftz.f32 %f145, %f113, %f114;
cvt.rn.f32.s32 %f146, %r162;
fma.rn.ftz.f32 %f360, %f145, %f146, %f360;

{ cvt.f32.f16 %f115, %rs4;}



{.reg .f16 low,high;
mov.b32 {low,high}, %r157;
mov.b16 %rs29, low;}


{ cvt.f32.f16 %f116, %rs29;}



dp4a.s32.s32 %r167, %r96, %r160, %r224;


dp4a.s32.s32 %r171, %r100, %r164, %r167;

mul.ftz.f32 %f147, %f115, %f116;
cvt.rn.f32.s32 %f148, %r171;
fma.rn.ftz.f32 %f359, %f147, %f148, %f359;
cvt.u64.u32 %rd66, %r526;
add.s64 %rd67, %rd1, %rd66;
mul.lo.s64 %rd68, %rd67, 36;
add.s64 %rd69, %rd9, %rd68;
add.s64 %rd70, %rd69, 4;
add.s64 %rd71, %rd70, %rd31;
ld.global.nc.u32 %r178, [%rd71];
add.s64 %rd72, %rd70, %rd33;
ld.global.nc.u32 %r182, [%rd72];

{ cvt.f32.f16 %f117, %rs1;}


ld.global.nc.u32 %r175, [%rd69];

{.reg .f16 low,high;
mov.b32 {low,high}, %r175;
mov.b16 %rs32, low;}


{ cvt.f32.f16 %f118, %rs32;}



dp4a.s32.s32 %r176, %r87, %r178, %r224;


dp4a.s32.s32 %r180, %r91, %r182, %r176;

mul.ftz.f32 %f149, %f117, %f118;
cvt.rn.f32.s32 %f150, %r180;
fma.rn.ftz.f32 %f358, %f149, %f150, %f358;

{ cvt.f32.f16 %f119, %rs4;}



{.reg .f16 low,high;
mov.b32 {low,high}, %r175;
mov.b16 %rs35, low;}


{ cvt.f32.f16 %f120, %rs35;}



dp4a.s32.s32 %r185, %r96, %r178, %r224;


dp4a.s32.s32 %r189, %r100, %r182, %r185;

mul.ftz.f32 %f151, %f119, %f120;
cvt.rn.f32.s32 %f152, %r189;
fma.rn.ftz.f32 %f357, %f151, %f152, %f357;
cvt.u64.u32 %rd73, %r525;
add.s64 %rd74, %rd1, %rd73;
mul.lo.s64 %rd75, %rd74, 36;
add.s64 %rd76, %rd9, %rd75;
add.s64 %rd77, %rd76, 4;
add.s64 %rd78, %rd77, %rd31;
ld.global.nc.u32 %r196, [%rd78];
add.s64 %rd79, %rd77, %rd33;
ld.global.nc.u32 %r200, [%rd79];

{ cvt.f32.f16 %f121, %rs1;}


ld.global.nc.u32 %r193, [%rd76];

{.reg .f16 low,high;
mov.b32 {low,high}, %r193;
mov.b16 %rs38, low;}


{ cvt.f32.f16 %f122, %rs38;}



dp4a.s32.s32 %r194, %r87, %r196, %r224;


dp4a.s32.s32 %r198, %r91, %r200, %r194;

mul.ftz.f32 %f153, %f121, %f122;
cvt.rn.f32.s32 %f154, %r198;
fma.rn.ftz.f32 %f356, %f153, %f154, %f356;

{ cvt.f32.f16 %f123, %rs4;}



{.reg .f16 low,high;
mov.b32 {low,high}, %r193;
mov.b16 %rs41, low;}


{ cvt.f32.f16 %f124, %rs41;}



dp4a.s32.s32 %r203, %r96, %r196, %r224;


dp4a.s32.s32 %r207, %r100, %r200, %r203;

mul.ftz.f32 %f155, %f123, %f124;
cvt.rn.f32.s32 %f156, %r207;
fma.rn.ftz.f32 %f355, %f155, %f156, %f355;
cvt.u64.u32 %rd80, %r524;
add.s64 %rd81, %rd1, %rd80;
mul.lo.s64 %rd82, %rd81, 36;
add.s64 %rd83, %rd9, %rd82;
add.s64 %rd84, %rd83, 4;
add.s64 %rd85, %rd84, %rd31;
ld.global.nc.u32 %r214, [%rd85];
add.s64 %rd86, %rd84, %rd33;
ld.global.nc.u32 %r218, [%rd86];

{ cvt.f32.f16 %f125, %rs1;}


ld.global.nc.u32 %r211, [%rd83];

{.reg .f16 low,high;
mov.b32 {low,high}, %r211;
mov.b16 %rs44, low;}


{ cvt.f32.f16 %f126, %rs44;}



dp4a.s32.s32 %r212, %r87, %r214, %r224;


dp4a.s32.s32 %r216, %r91, %r218, %r212;

mul.ftz.f32 %f157, %f125, %f126;
cvt.rn.f32.s32 %f158, %r216;
fma.rn.ftz.f32 %f354, %f157, %f158, %f354;

{ cvt.f32.f16 %f127, %rs4;}



{.reg .f16 low,high;
mov.b32 {low,high}, %r211;
mov.b16 %rs47, low;}


{ cvt.f32.f16 %f128, %rs47;}



dp4a.s32.s32 %r221, %r96, %r214, %r224;


dp4a.s32.s32 %r225, %r100, %r218, %r221;

mul.ftz.f32 %f159, %f127, %f128;
cvt.rn.f32.s32 %f160, %r225;
fma.rn.ftz.f32 %f353, %f159, %f160, %f353;
add.s32 %r529, %r529, 16;
add.s32 %r528, %r528, 16;
add.s32 %r527, %r527, 16;
add.s32 %r526, %r526, 16;
add.s32 %r525, %r525, 16;
add.s32 %r524, %r524, 16;
add.s32 %r523, %r523, 16;
add.s64 %rd141, %rd141, 544;
add.s32 %r530, %r530, 16;
setp.lt.s32 %p4, %r530, %r6;
@%p4 bra $L__BB100_2;

$L__BB100_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB100_5;

shl.b32 %r229, %r1, 11;
mov.u32 %r230, _ZZ13mul_mat_vec_qIL9ggml_type8ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r231, %r230, %r229;
shl.b32 %r232, %r2, 2;
add.s32 %r233, %r231, %r232;
st.shared.f32 [%r233+-2048], %f368;
st.shared.f32 [%r233+-1920], %f367;
st.shared.f32 [%r233+-1792], %f366;
st.shared.f32 [%r233+-1664], %f365;
st.shared.f32 [%r233+-1536], %f364;
st.shared.f32 [%r233+-1408], %f363;
st.shared.f32 [%r233+-1280], %f362;
st.shared.f32 [%r233+-1152], %f361;
st.shared.f32 [%r233+-1024], %f360;
st.shared.f32 [%r233+-896], %f359;
st.shared.f32 [%r233+-768], %f358;
st.shared.f32 [%r233+-640], %f357;
st.shared.f32 [%r233+-512], %f356;
st.shared.f32 [%r233+-384], %f355;
st.shared.f32 [%r233+-256], %f354;
st.shared.f32 [%r233+-128], %f353;

$L__BB100_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB100_38;

mov.u32 %r234, %ctaid.z;
mov.u32 %r235, %ctaid.y;
mad.lo.s32 %r236, %r235, %r42, %r4;
mad.lo.s32 %r237, %r234, %r48, %r236;
cvt.u64.u32 %rd12, %r237;
shl.b32 %r238, %r2, 2;
mov.u32 %r239, _ZZ13mul_mat_vec_qIL9ggml_type8ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r30, %r239, %r238;
mov.u32 %r240, 2;
ld.shared.f32 %f161, [%r30];
add.ftz.f32 %f162, %f161, %f368;
mov.b32 %r241, %f162;
mov.u32 %r242, 31;
mov.u32 %r243, 16;
mov.u32 %r244, -1;
shfl.sync.bfly.b32 %r245|%p7, %r241, %r243, %r242, %r244;
mov.b32 %f163, %r245;
add.ftz.f32 %f164, %f162, %f163;
mov.b32 %r246, %f164;
mov.u32 %r247, 8;
shfl.sync.bfly.b32 %r248|%p8, %r246, %r247, %r242, %r244;
mov.b32 %f165, %r248;
add.ftz.f32 %f166, %f164, %f165;
mov.b32 %r249, %f166;
mov.u32 %r250, 4;
shfl.sync.bfly.b32 %r251|%p9, %r249, %r250, %r242, %r244;
mov.b32 %f167, %r251;
add.ftz.f32 %f168, %f166, %f167;
mov.b32 %r252, %f168;
shfl.sync.bfly.b32 %r253|%p10, %r252, %r240, %r242, %r244;
mov.b32 %f169, %r253;
add.ftz.f32 %f170, %f168, %f169;
mov.b32 %r254, %f170;
mov.u32 %r255, 1;
shfl.sync.bfly.b32 %r256|%p11, %r254, %r255, %r242, %r244;
mov.b32 %f171, %r256;
add.ftz.f32 %f49, %f170, %f171;
setp.lt.u32 %p12, %r4, %r36;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
cvta.to.global.u64 %rd87, %rd19;
mul.wide.u32 %rd88, %r237, 4;
add.s64 %rd13, %rd87, %rd88;
not.pred %p14, %p1;
@%p14 bra $L__BB100_8;

st.global.f32 [%rd13], %f49;

$L__BB100_8:
ld.shared.f32 %f172, [%r30+128];
add.ftz.f32 %f173, %f172, %f367;
mov.b32 %r257, %f173;
shfl.sync.bfly.b32 %r261|%p15, %r257, %r243, %r242, %r244;
mov.b32 %f174, %r261;
add.ftz.f32 %f175, %f173, %f174;
mov.b32 %r262, %f175;
shfl.sync.bfly.b32 %r264|%p16, %r262, %r247, %r242, %r244;
mov.b32 %f176, %r264;
add.ftz.f32 %f177, %f175, %f176;
mov.b32 %r265, %f177;
shfl.sync.bfly.b32 %r267|%p17, %r265, %r250, %r242, %r244;
mov.b32 %f178, %r267;
add.ftz.f32 %f179, %f177, %f178;
mov.b32 %r268, %f179;
shfl.sync.bfly.b32 %r270|%p18, %r268, %r240, %r242, %r244;
mov.b32 %f180, %r270;
add.ftz.f32 %f181, %f179, %f180;
mov.b32 %r271, %f181;
shfl.sync.bfly.b32 %r273|%p19, %r271, %r255, %r242, %r244;
mov.b32 %f182, %r273;
add.ftz.f32 %f50, %f181, %f182;
setp.eq.s32 %p20, %r2, 1;
or.b32 %r277, %r4, 1;
setp.lt.u32 %p21, %r277, %r36;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB100_10;

st.global.f32 [%rd13+4], %f50;

$L__BB100_10:
ld.shared.f32 %f183, [%r30+256];
add.ftz.f32 %f184, %f183, %f366;
mov.b32 %r278, %f184;
mov.u32 %r279, 31;
mov.u32 %r280, 16;
mov.u32 %r281, -1;
shfl.sync.bfly.b32 %r282|%p23, %r278, %r280, %r279, %r281;
mov.b32 %f185, %r282;
add.ftz.f32 %f186, %f184, %f185;
mov.b32 %r283, %f186;
mov.u32 %r284, 8;
shfl.sync.bfly.b32 %r285|%p24, %r283, %r284, %r279, %r281;
mov.b32 %f187, %r285;
add.ftz.f32 %f188, %f186, %f187;
mov.b32 %r286, %f188;
mov.u32 %r287, 4;
shfl.sync.bfly.b32 %r288|%p25, %r286, %r287, %r279, %r281;
mov.b32 %f189, %r288;
add.ftz.f32 %f190, %f188, %f189;
mov.b32 %r289, %f190;
mov.u32 %r290, 2;
shfl.sync.bfly.b32 %r291|%p26, %r289, %r290, %r279, %r281;
mov.b32 %f191, %r291;
add.ftz.f32 %f192, %f190, %f191;
mov.b32 %r292, %f192;
mov.u32 %r293, 1;
shfl.sync.bfly.b32 %r294|%p27, %r292, %r293, %r279, %r281;
mov.b32 %f193, %r294;
add.ftz.f32 %f51, %f192, %f193;
@%p14 bra $L__BB100_12;

cvt.u64.u32 %rd89, %r36;
add.s64 %rd90, %rd89, %rd12;
shl.b64 %rd92, %rd90, 2;
add.s64 %rd93, %rd87, %rd92;
st.global.f32 [%rd93], %f51;

$L__BB100_12:
ld.shared.f32 %f194, [%r30+384];
add.ftz.f32 %f195, %f194, %f365;
mov.b32 %r295, %f195;
shfl.sync.bfly.b32 %r299|%p29, %r295, %r280, %r279, %r281;
mov.b32 %f196, %r299;
add.ftz.f32 %f197, %f195, %f196;
mov.b32 %r300, %f197;
shfl.sync.bfly.b32 %r302|%p30, %r300, %r284, %r279, %r281;
mov.b32 %f198, %r302;
add.ftz.f32 %f199, %f197, %f198;
mov.b32 %r303, %f199;
shfl.sync.bfly.b32 %r305|%p31, %r303, %r287, %r279, %r281;
mov.b32 %f200, %r305;
add.ftz.f32 %f201, %f199, %f200;
mov.b32 %r306, %f201;
shfl.sync.bfly.b32 %r308|%p32, %r306, %r290, %r279, %r281;
mov.b32 %f202, %r308;
add.ftz.f32 %f203, %f201, %f202;
mov.b32 %r309, %f203;
shfl.sync.bfly.b32 %r311|%p33, %r309, %r293, %r279, %r281;
mov.b32 %f204, %r311;
add.ftz.f32 %f52, %f203, %f204;
@%p22 bra $L__BB100_14;

add.s32 %r312, %r36, 1;
cvt.u64.u32 %rd94, %r312;
add.s64 %rd95, %rd94, %rd12;
shl.b64 %rd97, %rd95, 2;
add.s64 %rd98, %rd87, %rd97;
st.global.f32 [%rd98], %f52;

$L__BB100_14:
shl.b32 %r313, %r36, 1;
mov.u32 %r314, 1;
ld.shared.f32 %f205, [%r30+512];
add.ftz.f32 %f206, %f205, %f364;
mov.b32 %r315, %f206;
mov.u32 %r316, 31;
mov.u32 %r317, 16;
mov.u32 %r318, -1;
shfl.sync.bfly.b32 %r319|%p35, %r315, %r317, %r316, %r318;
mov.b32 %f207, %r319;
add.ftz.f32 %f208, %f206, %f207;
mov.b32 %r320, %f208;
mov.u32 %r321, 8;
shfl.sync.bfly.b32 %r322|%p36, %r320, %r321, %r316, %r318;
mov.b32 %f209, %r322;
add.ftz.f32 %f210, %f208, %f209;
mov.b32 %r323, %f210;
mov.u32 %r324, 4;
shfl.sync.bfly.b32 %r325|%p37, %r323, %r324, %r316, %r318;
mov.b32 %f211, %r325;
add.ftz.f32 %f212, %f210, %f211;
mov.b32 %r326, %f212;
mov.u32 %r327, 2;
shfl.sync.bfly.b32 %r328|%p38, %r326, %r327, %r316, %r318;
mov.b32 %f213, %r328;
add.ftz.f32 %f214, %f212, %f213;
mov.b32 %r329, %f214;
shfl.sync.bfly.b32 %r330|%p39, %r329, %r314, %r316, %r318;
mov.b32 %f215, %r330;
add.ftz.f32 %f53, %f214, %f215;
cvt.u64.u32 %rd99, %r313;
add.s64 %rd100, %rd99, %rd12;
shl.b64 %rd102, %rd100, 2;
add.s64 %rd14, %rd87, %rd102;
@%p14 bra $L__BB100_16;

st.global.f32 [%rd14], %f53;

$L__BB100_16:
ld.shared.f32 %f216, [%r30+640];
add.ftz.f32 %f217, %f216, %f363;
mov.b32 %r331, %f217;
shfl.sync.bfly.b32 %r335|%p41, %r331, %r317, %r316, %r318;
mov.b32 %f218, %r335;
add.ftz.f32 %f219, %f217, %f218;
mov.b32 %r336, %f219;
shfl.sync.bfly.b32 %r338|%p42, %r336, %r321, %r316, %r318;
mov.b32 %f220, %r338;
add.ftz.f32 %f221, %f219, %f220;
mov.b32 %r339, %f221;
shfl.sync.bfly.b32 %r341|%p43, %r339, %r324, %r316, %r318;
mov.b32 %f222, %r341;
add.ftz.f32 %f223, %f221, %f222;
mov.b32 %r342, %f223;
shfl.sync.bfly.b32 %r344|%p44, %r342, %r327, %r316, %r318;
mov.b32 %f224, %r344;
add.ftz.f32 %f225, %f223, %f224;
mov.b32 %r345, %f225;
shfl.sync.bfly.b32 %r347|%p45, %r345, %r314, %r316, %r318;
mov.b32 %f226, %r347;
add.ftz.f32 %f54, %f225, %f226;
@%p22 bra $L__BB100_18;

st.global.f32 [%rd14+4], %f54;

$L__BB100_18:
mul.lo.s32 %r31, %r36, 3;
ld.shared.f32 %f227, [%r30+768];
add.ftz.f32 %f228, %f227, %f362;
mov.b32 %r348, %f228;
mov.u32 %r349, 31;
mov.u32 %r350, 16;
mov.u32 %r351, -1;
shfl.sync.bfly.b32 %r352|%p47, %r348, %r350, %r349, %r351;
mov.b32 %f229, %r352;
add.ftz.f32 %f230, %f228, %f229;
mov.b32 %r353, %f230;
mov.u32 %r354, 8;
shfl.sync.bfly.b32 %r355|%p48, %r353, %r354, %r349, %r351;
mov.b32 %f231, %r355;
add.ftz.f32 %f232, %f230, %f231;
mov.b32 %r356, %f232;
mov.u32 %r357, 4;
shfl.sync.bfly.b32 %r358|%p49, %r356, %r357, %r349, %r351;
mov.b32 %f233, %r358;
add.ftz.f32 %f234, %f232, %f233;
mov.b32 %r359, %f234;
mov.u32 %r360, 2;
shfl.sync.bfly.b32 %r361|%p50, %r359, %r360, %r349, %r351;
mov.b32 %f235, %r361;
add.ftz.f32 %f236, %f234, %f235;
mov.b32 %r362, %f236;
mov.u32 %r363, 1;
shfl.sync.bfly.b32 %r364|%p51, %r362, %r363, %r349, %r351;
mov.b32 %f237, %r364;
add.ftz.f32 %f55, %f236, %f237;
@%p14 bra $L__BB100_20;

cvt.u64.u32 %rd103, %r31;
add.s64 %rd104, %rd103, %rd12;
shl.b64 %rd106, %rd104, 2;
add.s64 %rd107, %rd87, %rd106;
st.global.f32 [%rd107], %f55;

$L__BB100_20:
ld.shared.f32 %f238, [%r30+896];
add.ftz.f32 %f239, %f238, %f361;
mov.b32 %r365, %f239;
shfl.sync.bfly.b32 %r369|%p53, %r365, %r350, %r349, %r351;
mov.b32 %f240, %r369;
add.ftz.f32 %f241, %f239, %f240;
mov.b32 %r370, %f241;
shfl.sync.bfly.b32 %r372|%p54, %r370, %r354, %r349, %r351;
mov.b32 %f242, %r372;
add.ftz.f32 %f243, %f241, %f242;
mov.b32 %r373, %f243;
shfl.sync.bfly.b32 %r375|%p55, %r373, %r357, %r349, %r351;
mov.b32 %f244, %r375;
add.ftz.f32 %f245, %f243, %f244;
mov.b32 %r376, %f245;
shfl.sync.bfly.b32 %r378|%p56, %r376, %r360, %r349, %r351;
mov.b32 %f246, %r378;
add.ftz.f32 %f247, %f245, %f246;
mov.b32 %r379, %f247;
shfl.sync.bfly.b32 %r381|%p57, %r379, %r363, %r349, %r351;
mov.b32 %f248, %r381;
add.ftz.f32 %f56, %f247, %f248;
@%p22 bra $L__BB100_22;

add.s32 %r382, %r31, 1;
cvt.u64.u32 %rd108, %r382;
add.s64 %rd109, %rd108, %rd12;
shl.b64 %rd111, %rd109, 2;
add.s64 %rd112, %rd87, %rd111;
st.global.f32 [%rd112], %f56;

$L__BB100_22:
shl.b32 %r383, %r36, 2;
mov.u32 %r384, 2;
ld.shared.f32 %f249, [%r30+1024];
add.ftz.f32 %f250, %f249, %f360;
mov.b32 %r385, %f250;
mov.u32 %r386, 31;
mov.u32 %r387, 16;
mov.u32 %r388, -1;
shfl.sync.bfly.b32 %r389|%p59, %r385, %r387, %r386, %r388;
mov.b32 %f251, %r389;
add.ftz.f32 %f252, %f250, %f251;
mov.b32 %r390, %f252;
mov.u32 %r391, 8;
shfl.sync.bfly.b32 %r392|%p60, %r390, %r391, %r386, %r388;
mov.b32 %f253, %r392;
add.ftz.f32 %f254, %f252, %f253;
mov.b32 %r393, %f254;
mov.u32 %r394, 4;
shfl.sync.bfly.b32 %r395|%p61, %r393, %r394, %r386, %r388;
mov.b32 %f255, %r395;
add.ftz.f32 %f256, %f254, %f255;
mov.b32 %r396, %f256;
shfl.sync.bfly.b32 %r397|%p62, %r396, %r384, %r386, %r388;
mov.b32 %f257, %r397;
add.ftz.f32 %f258, %f256, %f257;
mov.b32 %r398, %f258;
mov.u32 %r399, 1;
shfl.sync.bfly.b32 %r400|%p63, %r398, %r399, %r386, %r388;
mov.b32 %f259, %r400;
add.ftz.f32 %f57, %f258, %f259;
cvt.u64.u32 %rd113, %r383;
add.s64 %rd114, %rd113, %rd12;
shl.b64 %rd116, %rd114, 2;
add.s64 %rd15, %rd87, %rd116;
@%p14 bra $L__BB100_24;

st.global.f32 [%rd15], %f57;

$L__BB100_24:
ld.shared.f32 %f260, [%r30+1152];
add.ftz.f32 %f261, %f260, %f359;
mov.b32 %r401, %f261;
shfl.sync.bfly.b32 %r405|%p65, %r401, %r387, %r386, %r388;
mov.b32 %f262, %r405;
add.ftz.f32 %f263, %f261, %f262;
mov.b32 %r406, %f263;
shfl.sync.bfly.b32 %r408|%p66, %r406, %r391, %r386, %r388;
mov.b32 %f264, %r408;
add.ftz.f32 %f265, %f263, %f264;
mov.b32 %r409, %f265;
shfl.sync.bfly.b32 %r411|%p67, %r409, %r394, %r386, %r388;
mov.b32 %f266, %r411;
add.ftz.f32 %f267, %f265, %f266;
mov.b32 %r412, %f267;
shfl.sync.bfly.b32 %r414|%p68, %r412, %r384, %r386, %r388;
mov.b32 %f268, %r414;
add.ftz.f32 %f269, %f267, %f268;
mov.b32 %r415, %f269;
shfl.sync.bfly.b32 %r417|%p69, %r415, %r399, %r386, %r388;
mov.b32 %f270, %r417;
add.ftz.f32 %f58, %f269, %f270;
@%p22 bra $L__BB100_26;

st.global.f32 [%rd15+4], %f58;

$L__BB100_26:
mul.lo.s32 %r32, %r36, 5;
ld.shared.f32 %f271, [%r30+1280];
add.ftz.f32 %f272, %f271, %f358;
mov.b32 %r418, %f272;
mov.u32 %r419, 31;
mov.u32 %r420, 16;
mov.u32 %r421, -1;
shfl.sync.bfly.b32 %r422|%p71, %r418, %r420, %r419, %r421;
mov.b32 %f273, %r422;
add.ftz.f32 %f274, %f272, %f273;
mov.b32 %r423, %f274;
mov.u32 %r424, 8;
shfl.sync.bfly.b32 %r425|%p72, %r423, %r424, %r419, %r421;
mov.b32 %f275, %r425;
add.ftz.f32 %f276, %f274, %f275;
mov.b32 %r426, %f276;
mov.u32 %r427, 4;
shfl.sync.bfly.b32 %r428|%p73, %r426, %r427, %r419, %r421;
mov.b32 %f277, %r428;
add.ftz.f32 %f278, %f276, %f277;
mov.b32 %r429, %f278;
mov.u32 %r430, 2;
shfl.sync.bfly.b32 %r431|%p74, %r429, %r430, %r419, %r421;
mov.b32 %f279, %r431;
add.ftz.f32 %f280, %f278, %f279;
mov.b32 %r432, %f280;
mov.u32 %r433, 1;
shfl.sync.bfly.b32 %r434|%p75, %r432, %r433, %r419, %r421;
mov.b32 %f281, %r434;
add.ftz.f32 %f59, %f280, %f281;
@%p14 bra $L__BB100_28;

cvt.u64.u32 %rd117, %r32;
add.s64 %rd118, %rd117, %rd12;
shl.b64 %rd120, %rd118, 2;
add.s64 %rd121, %rd87, %rd120;
st.global.f32 [%rd121], %f59;

$L__BB100_28:
ld.shared.f32 %f282, [%r30+1408];
add.ftz.f32 %f283, %f282, %f357;
mov.b32 %r435, %f283;
shfl.sync.bfly.b32 %r439|%p77, %r435, %r420, %r419, %r421;
mov.b32 %f284, %r439;
add.ftz.f32 %f285, %f283, %f284;
mov.b32 %r440, %f285;
shfl.sync.bfly.b32 %r442|%p78, %r440, %r424, %r419, %r421;
mov.b32 %f286, %r442;
add.ftz.f32 %f287, %f285, %f286;
mov.b32 %r443, %f287;
shfl.sync.bfly.b32 %r445|%p79, %r443, %r427, %r419, %r421;
mov.b32 %f288, %r445;
add.ftz.f32 %f289, %f287, %f288;
mov.b32 %r446, %f289;
shfl.sync.bfly.b32 %r448|%p80, %r446, %r430, %r419, %r421;
mov.b32 %f290, %r448;
add.ftz.f32 %f291, %f289, %f290;
mov.b32 %r449, %f291;
shfl.sync.bfly.b32 %r451|%p81, %r449, %r433, %r419, %r421;
mov.b32 %f292, %r451;
add.ftz.f32 %f60, %f291, %f292;
@%p22 bra $L__BB100_30;

add.s32 %r452, %r32, 1;
cvt.u64.u32 %rd122, %r452;
add.s64 %rd123, %rd122, %rd12;
shl.b64 %rd125, %rd123, 2;
add.s64 %rd126, %rd87, %rd125;
st.global.f32 [%rd126], %f60;

$L__BB100_30:
mul.lo.s32 %r453, %r36, 6;
ld.shared.f32 %f293, [%r30+1536];
add.ftz.f32 %f294, %f293, %f356;
mov.b32 %r454, %f294;
mov.u32 %r455, 31;
mov.u32 %r456, 16;
mov.u32 %r457, -1;
shfl.sync.bfly.b32 %r458|%p83, %r454, %r456, %r455, %r457;
mov.b32 %f295, %r458;
add.ftz.f32 %f296, %f294, %f295;
mov.b32 %r459, %f296;
mov.u32 %r460, 8;
shfl.sync.bfly.b32 %r461|%p84, %r459, %r460, %r455, %r457;
mov.b32 %f297, %r461;
add.ftz.f32 %f298, %f296, %f297;
mov.b32 %r462, %f298;
mov.u32 %r463, 4;
shfl.sync.bfly.b32 %r464|%p85, %r462, %r463, %r455, %r457;
mov.b32 %f299, %r464;
add.ftz.f32 %f300, %f298, %f299;
mov.b32 %r465, %f300;
mov.u32 %r466, 2;
shfl.sync.bfly.b32 %r467|%p86, %r465, %r466, %r455, %r457;
mov.b32 %f301, %r467;
add.ftz.f32 %f302, %f300, %f301;
mov.b32 %r468, %f302;
mov.u32 %r469, 1;
shfl.sync.bfly.b32 %r470|%p87, %r468, %r469, %r455, %r457;
mov.b32 %f303, %r470;
add.ftz.f32 %f61, %f302, %f303;
cvt.u64.u32 %rd127, %r453;
add.s64 %rd128, %rd127, %rd12;
shl.b64 %rd130, %rd128, 2;
add.s64 %rd16, %rd87, %rd130;
@%p14 bra $L__BB100_32;

st.global.f32 [%rd16], %f61;

$L__BB100_32:
ld.shared.f32 %f304, [%r30+1664];
add.ftz.f32 %f305, %f304, %f355;
mov.b32 %r471, %f305;
shfl.sync.bfly.b32 %r475|%p89, %r471, %r456, %r455, %r457;
mov.b32 %f306, %r475;
add.ftz.f32 %f307, %f305, %f306;
mov.b32 %r476, %f307;
shfl.sync.bfly.b32 %r478|%p90, %r476, %r460, %r455, %r457;
mov.b32 %f308, %r478;
add.ftz.f32 %f309, %f307, %f308;
mov.b32 %r479, %f309;
shfl.sync.bfly.b32 %r481|%p91, %r479, %r463, %r455, %r457;
mov.b32 %f310, %r481;
add.ftz.f32 %f311, %f309, %f310;
mov.b32 %r482, %f311;
shfl.sync.bfly.b32 %r484|%p92, %r482, %r466, %r455, %r457;
mov.b32 %f312, %r484;
add.ftz.f32 %f313, %f311, %f312;
mov.b32 %r485, %f313;
shfl.sync.bfly.b32 %r487|%p93, %r485, %r469, %r455, %r457;
mov.b32 %f314, %r487;
add.ftz.f32 %f62, %f313, %f314;
@%p22 bra $L__BB100_34;

st.global.f32 [%rd16+4], %f62;

$L__BB100_34:
mul.lo.s32 %r33, %r36, 7;
ld.shared.f32 %f315, [%r30+1792];
add.ftz.f32 %f316, %f315, %f354;
mov.b32 %r488, %f316;
mov.u32 %r489, 31;
mov.u32 %r490, 16;
mov.u32 %r491, -1;
shfl.sync.bfly.b32 %r492|%p95, %r488, %r490, %r489, %r491;
mov.b32 %f317, %r492;
add.ftz.f32 %f318, %f316, %f317;
mov.b32 %r493, %f318;
mov.u32 %r494, 8;
shfl.sync.bfly.b32 %r495|%p96, %r493, %r494, %r489, %r491;
mov.b32 %f319, %r495;
add.ftz.f32 %f320, %f318, %f319;
mov.b32 %r496, %f320;
mov.u32 %r497, 4;
shfl.sync.bfly.b32 %r498|%p97, %r496, %r497, %r489, %r491;
mov.b32 %f321, %r498;
add.ftz.f32 %f322, %f320, %f321;
mov.b32 %r499, %f322;
mov.u32 %r500, 2;
shfl.sync.bfly.b32 %r501|%p98, %r499, %r500, %r489, %r491;
mov.b32 %f323, %r501;
add.ftz.f32 %f324, %f322, %f323;
mov.b32 %r502, %f324;
mov.u32 %r503, 1;
shfl.sync.bfly.b32 %r504|%p99, %r502, %r503, %r489, %r491;
mov.b32 %f325, %r504;
add.ftz.f32 %f63, %f324, %f325;
@%p14 bra $L__BB100_36;

cvt.u64.u32 %rd131, %r33;
add.s64 %rd132, %rd131, %rd12;
shl.b64 %rd134, %rd132, 2;
add.s64 %rd135, %rd87, %rd134;
st.global.f32 [%rd135], %f63;

$L__BB100_36:
ld.shared.f32 %f326, [%r30+1920];
add.ftz.f32 %f327, %f326, %f353;
mov.b32 %r505, %f327;
shfl.sync.bfly.b32 %r509|%p101, %r505, %r490, %r489, %r491;
mov.b32 %f328, %r509;
add.ftz.f32 %f329, %f327, %f328;
mov.b32 %r510, %f329;
shfl.sync.bfly.b32 %r512|%p102, %r510, %r494, %r489, %r491;
mov.b32 %f330, %r512;
add.ftz.f32 %f331, %f329, %f330;
mov.b32 %r513, %f331;
shfl.sync.bfly.b32 %r515|%p103, %r513, %r497, %r489, %r491;
mov.b32 %f332, %r515;
add.ftz.f32 %f333, %f331, %f332;
mov.b32 %r516, %f333;
shfl.sync.bfly.b32 %r518|%p104, %r516, %r500, %r489, %r491;
mov.b32 %f334, %r518;
add.ftz.f32 %f335, %f333, %f334;
mov.b32 %r519, %f335;
shfl.sync.bfly.b32 %r521|%p105, %r519, %r503, %r489, %r491;
mov.b32 %f336, %r521;
add.ftz.f32 %f64, %f335, %f336;
@%p22 bra $L__BB100_38;

add.s32 %r522, %r33, 1;
cvt.u64.u32 %rd136, %r522;
add.s64 %rd137, %rd136, %rd12;
shl.b64 %rd139, %rd137, 2;
add.s64 %rd140, %rd87, %rd139;
st.global.f32 [%rd140], %f64;

$L__BB100_38:
ret;

}
.entry _Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj(
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2,
.param .align 8 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3[48],
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5,
.param .align 4 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6[12],
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15
)
.maxntid 256, 1, 1
.minnctapersm 1
{
.local .align 8 .b8 __local_depot101[16];
.reg .b64 %SP;
.reg .b64 %SPL;
.reg .pred %p<56>;
.reg .b16 %rs<37>;
.reg .f32 %f<173>;
.reg .b32 %r<435>;
.reg .b64 %rd<108>;


mov.u64 %SPL, __local_depot101;
ld.param.u64 %rd19, [_Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0];
ld.param.u64 %rd20, [_Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1];
ld.param.u64 %rd21, [_Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2];
ld.param.u32 %r17, [_Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5];
ld.param.u32 %r22, [_Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8];
ld.param.u32 %r23, [_Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9];
ld.param.u32 %r25, [_Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11];
ld.param.u32 %r26, [_Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12];
ld.param.u32 %r29, [_Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14];
ld.param.u32 %r28, [_Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15];
ld.param.u32 %r20, [_Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+8];
ld.param.u32 %r19, [_Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+4];
ld.param.u32 %r18, [_Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6];
ld.param.u64 %rd23, [_Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+8];
add.u64 %rd1, %SPL, 0;
add.u64 %rd2, %SPL, 8;
mov.u32 %r1, %tid.y;
setp.ge.u32 %p1, %r1, %r29;
@%p1 bra $L__BB101_29;

mov.u32 %r2, %ctaid.y;
cvta.to.global.u64 %rd30, %rd21;
mad.lo.s32 %r30, %r1, %r28, %r2;
mul.wide.u32 %rd31, %r30, 4;
add.s64 %rd32, %rd30, %rd31;
ld.global.nc.u32 %r3, [%rd32];
mov.f32 %f166, 0f00000000;
st.local.v2.f32 [%rd1], {%f166, %f166};
st.local.v2.f32 [%rd2], {%f166, %f166};
mov.u32 %r4, %tid.x;
shr.u32 %r5, %r4, 1;
shr.u32 %r31, %r17, 5;
setp.ge.u32 %p2, %r5, %r31;
mov.f32 %f167, %f166;
mov.f32 %f168, %f166;
mov.f32 %f169, %f166;
@%p2 bra $L__BB101_9;

mul.hi.u32 %r32, %r2, %r18;
add.s32 %r33, %r32, %r2;
shr.u32 %r34, %r33, %r19;
mul.lo.s32 %r35, %r34, %r20;
sub.s32 %r36, %r2, %r35;
mul.lo.s32 %r37, %r36, %r26;
cvt.u64.u32 %rd33, %r37;
shl.b32 %r38, %r4, 3;
and.b32 %r39, %r38, 8;
cvt.u64.u32 %rd34, %r39;
cvt.u64.u32 %rd35, %r4;
shr.u64 %rd36, %rd35, 1;
mul.wide.u32 %rd107, %r37, 36;
mul.lo.s32 %r40, %r1, %r23;
cvt.u64.u32 %rd37, %r40;
add.s64 %rd38, %rd36, %rd37;
add.s64 %rd39, %rd38, %rd33;
mul.lo.s64 %rd40, %rd39, 36;
add.s64 %rd41, %rd40, %rd34;
cvta.to.global.u64 %rd42, %rd20;
add.s64 %rd43, %rd42, %rd41;
add.s64 %rd105, %rd43, 20;
cvta.to.global.u64 %rd106, %rd19;
cvta.to.global.u64 %rd104, %rd23;
mov.f32 %f166, 0f00000000;
mov.u32 %r434, %r5;

$L__BB101_3:
mul.lo.s64 %rd46, %rd36, 36;
mul.wide.u32 %rd47, %r40, 36;
add.s64 %rd48, %rd46, %rd47;
add.s64 %rd50, %rd42, %rd48;
add.s64 %rd51, %rd50, %rd107;
mov.u32 %r64, %ctaid.x;
shl.b32 %r65, %r64, 1;
mad.lo.s32 %r66, %r65, %r22, %r5;
mad.lo.s32 %r67, %r3, %r25, %r66;
mul.wide.s32 %rd53, %r67, 17;
add.s64 %rd54, %rd34, %rd53;
add.s64 %rd55, %rd106, %rd54;
ld.global.nc.u8 %rs2, [%rd55+1];
cvt.u32.u16 %r68, %rs2;
and.b32 %r69, %r68, 255;
ld.global.nc.u8 %rs3, [%rd55+2];
cvt.u32.u16 %r70, %rs3;
prmt.b32 %r71, %r70, %r69, 30212;
ld.global.nc.u8 %rs4, [%rd55+3];
cvt.u32.u16 %r72, %rs4;
prmt.b32 %r73, %r72, %r71, 28756;
ld.global.nc.u8 %rs5, [%rd55+4];
cvt.u32.u16 %r74, %rs5;
prmt.b32 %r75, %r74, %r73, 1620;
and.b32 %r76, %r75, -2004318072;
shr.u32 %r77, %r76, 1;
or.b32 %r78, %r77, 839922192;
and.b32 %r79, %r71, 30583;
mov.u32 %r80, 201852420;
mov.u32 %r81, 50462976;
prmt.b32 %r82, %r81, %r80, %r79;
mov.u32 %r83, -185009412;
mov.u32 %r84, -33620224;
prmt.b32 %r85, %r84, %r83, %r79;
and.b32 %r86, %r78, 30292;
prmt.b32 %r87, %r82, %r85, %r86;
shr.u32 %r88, %r75, 16;
and.b32 %r89, %r88, 30583;
prmt.b32 %r90, %r81, %r80, %r89;
prmt.b32 %r91, %r84, %r83, %r89;
shr.u32 %r92, %r78, 16;
prmt.b32 %r93, %r90, %r91, %r92;
mov.u32 %r94, 25632;
prmt.b32 %r42, %r87, %r93, %r94;
mov.u32 %r95, 30001;
prmt.b32 %r46, %r87, %r93, %r95;
ld.global.nc.u32 %r7, [%rd105+-16];
mov.u32 %r44, 0;

dp4a.s32.s32 %r41, %r42, %r7, %r44;

ld.global.nc.u32 %r8, [%rd105];

dp4a.s32.s32 %r45, %r46, %r8, %r41;

ld.global.nc.u8 %rs6, [%rd55+5];
cvt.u32.u16 %r96, %rs6;
and.b32 %r97, %r96, 255;
ld.global.nc.u8 %rs7, [%rd55+6];
cvt.u32.u16 %r98, %rs7;
prmt.b32 %r99, %r98, %r97, 30212;
ld.global.nc.u8 %rs8, [%rd55+7];
cvt.u32.u16 %r100, %rs8;
prmt.b32 %r101, %r100, %r99, 28756;
ld.global.nc.u8 %rs9, [%rd55+8];
cvt.u32.u16 %r102, %rs9;
prmt.b32 %r103, %r102, %r101, 1620;
and.b32 %r104, %r103, -2004318072;
shr.u32 %r105, %r104, 1;
or.b32 %r106, %r105, 839922192;
and.b32 %r107, %r99, 30583;
prmt.b32 %r108, %r81, %r80, %r107;
prmt.b32 %r109, %r84, %r83, %r107;
and.b32 %r110, %r106, 30292;
prmt.b32 %r111, %r108, %r109, %r110;
shr.u32 %r112, %r103, 16;
and.b32 %r113, %r112, 30583;
prmt.b32 %r114, %r81, %r80, %r113;
prmt.b32 %r115, %r84, %r83, %r113;
shr.u32 %r116, %r106, 16;
prmt.b32 %r117, %r114, %r115, %r116;
prmt.b32 %r50, %r111, %r117, %r94;
prmt.b32 %r54, %r111, %r117, %r95;
ld.global.nc.u32 %r9, [%rd105+-12];

dp4a.s32.s32 %r49, %r50, %r9, %r45;

ld.global.nc.u32 %r10, [%rd105+4];

dp4a.s32.s32 %r53, %r54, %r10, %r49;

add.s64 %rd56, %rd106, %rd53;
ld.global.nc.u8 %rs10, [%rd56];
shl.b16 %rs12, %rs10, 7;
and.b16 %rs13, %rs12, 32640;
setp.eq.s16 %p3, %rs10, 255;
setp.eq.s16 %p4, %rs10, 0;
selp.b16 %rs14, 64, %rs13, %p4;
selp.b16 %rs1, 32767, %rs14, %p3;

{ mov.b32 %f42, {0,%rs1};}


mul.ftz.f32 %f44, %f42, 0f3F000000;
ld.global.nc.u32 %r11, [%rd51];

{.reg .f16 low,high;
mov.b32 {low,high},%r11;
cvt.f32.f16 %f43, low;}


mul.ftz.f32 %f45, %f44, %f43;
cvt.rn.f32.s32 %f46, %r53;
fma.rn.ftz.f32 %f169, %f45, %f46, %f169;
setp.eq.s64 %p5, %rd23, 0;
@%p5 bra $L__BB101_5;

shr.u32 %r138, %r4, 1;
mad.lo.s32 %r141, %r65, %r22, %r138;
mad.lo.s32 %r142, %r3, %r25, %r141;
mul.wide.s32 %rd58, %r142, 17;
add.s64 %rd59, %rd34, %rd58;
add.s64 %rd60, %rd104, %rd59;
ld.global.u8 %r143, [%rd60+1];
ld.global.u8 %r144, [%rd60+2];
prmt.b32 %r145, %r144, %r143, 30212;
ld.global.u8 %r146, [%rd60+3];
prmt.b32 %r147, %r146, %r145, 28756;
ld.global.u8 %r148, [%rd60+4];
prmt.b32 %r149, %r148, %r147, 1620;
and.b32 %r150, %r149, -2004318072;
shr.u32 %r151, %r150, 1;
or.b32 %r152, %r151, 839922192;
and.b32 %r153, %r145, 30583;
prmt.b32 %r156, %r81, %r80, %r153;
prmt.b32 %r159, %r84, %r83, %r153;
and.b32 %r160, %r152, 30292;
prmt.b32 %r161, %r156, %r159, %r160;
shr.u32 %r162, %r149, 16;
and.b32 %r163, %r162, 30583;
prmt.b32 %r164, %r81, %r80, %r163;
prmt.b32 %r165, %r84, %r83, %r163;
shr.u32 %r166, %r152, 16;
prmt.b32 %r167, %r164, %r165, %r166;
prmt.b32 %r119, %r161, %r167, %r94;
prmt.b32 %r123, %r161, %r167, %r95;

dp4a.s32.s32 %r118, %r119, %r7, %r44;


dp4a.s32.s32 %r122, %r123, %r8, %r118;

ld.global.u8 %r170, [%rd60+5];
ld.global.u8 %r171, [%rd60+6];
prmt.b32 %r172, %r171, %r170, 30212;
ld.global.u8 %r173, [%rd60+7];
prmt.b32 %r174, %r173, %r172, 28756;
ld.global.u8 %r175, [%rd60+8];
prmt.b32 %r176, %r175, %r174, 1620;
and.b32 %r177, %r176, -2004318072;
shr.u32 %r178, %r177, 1;
or.b32 %r179, %r178, 839922192;
and.b32 %r180, %r172, 30583;
prmt.b32 %r181, %r81, %r80, %r180;
prmt.b32 %r182, %r84, %r83, %r180;
and.b32 %r183, %r179, 30292;
prmt.b32 %r184, %r181, %r182, %r183;
shr.u32 %r185, %r176, 16;
and.b32 %r186, %r185, 30583;
prmt.b32 %r187, %r81, %r80, %r186;
prmt.b32 %r188, %r84, %r83, %r186;
shr.u32 %r189, %r179, 16;
prmt.b32 %r190, %r187, %r188, %r189;
prmt.b32 %r127, %r184, %r190, %r94;
prmt.b32 %r131, %r184, %r190, %r95;

dp4a.s32.s32 %r126, %r127, %r9, %r122;


dp4a.s32.s32 %r130, %r131, %r10, %r126;

add.s64 %rd61, %rd104, %rd58;
ld.global.u8 %rs16, [%rd61];
shl.b16 %rs17, %rs16, 7;
setp.eq.s16 %p6, %rs16, 255;
setp.eq.s16 %p7, %rs16, 0;
selp.b16 %rs18, 64, %rs17, %p7;
selp.b16 %rs15, 32767, %rs18, %p6;

{ mov.b32 %f47, {0,%rs15};}


mul.ftz.f32 %f49, %f47, 0f3F000000;

{.reg .f16 low,high;
mov.b32 {low,high},%r11;
cvt.f32.f16 %f48, low;}


mul.ftz.f32 %f50, %f49, %f48;
cvt.rn.f32.s32 %f51, %r130;
fma.rn.ftz.f32 %f167, %f50, %f51, %f167;

$L__BB101_5:
shr.u32 %r211, %r4, 1;
mad.lo.s32 %r212, %r3, %r25, %r211;
or.b32 %r215, %r65, 1;
mad.lo.s32 %r216, %r22, %r215, %r212;
mul.wide.s32 %rd63, %r216, 17;
add.s64 %rd64, %rd34, %rd63;
add.s64 %rd65, %rd106, %rd64;
ld.global.nc.u8 %rs20, [%rd65+1];
cvt.u32.u16 %r217, %rs20;
and.b32 %r218, %r217, 255;
ld.global.nc.u8 %rs21, [%rd65+2];
cvt.u32.u16 %r219, %rs21;
prmt.b32 %r220, %r219, %r218, 30212;
ld.global.nc.u8 %rs22, [%rd65+3];
cvt.u32.u16 %r221, %rs22;
prmt.b32 %r222, %r221, %r220, 28756;
ld.global.nc.u8 %rs23, [%rd65+4];
cvt.u32.u16 %r223, %rs23;
prmt.b32 %r224, %r223, %r222, 1620;
and.b32 %r225, %r224, -2004318072;
shr.u32 %r226, %r225, 1;
or.b32 %r227, %r226, 839922192;
and.b32 %r228, %r220, 30583;
prmt.b32 %r231, %r81, %r80, %r228;
prmt.b32 %r234, %r84, %r83, %r228;
and.b32 %r235, %r227, 30292;
prmt.b32 %r236, %r231, %r234, %r235;
shr.u32 %r237, %r224, 16;
and.b32 %r238, %r237, 30583;
prmt.b32 %r239, %r81, %r80, %r238;
prmt.b32 %r240, %r84, %r83, %r238;
shr.u32 %r241, %r227, 16;
prmt.b32 %r242, %r239, %r240, %r241;
prmt.b32 %r192, %r236, %r242, %r94;
prmt.b32 %r196, %r236, %r242, %r95;

dp4a.s32.s32 %r191, %r192, %r7, %r44;


dp4a.s32.s32 %r195, %r196, %r8, %r191;

ld.global.nc.u8 %rs24, [%rd65+5];
cvt.u32.u16 %r245, %rs24;
and.b32 %r246, %r245, 255;
ld.global.nc.u8 %rs25, [%rd65+6];
cvt.u32.u16 %r247, %rs25;
prmt.b32 %r248, %r247, %r246, 30212;
ld.global.nc.u8 %rs26, [%rd65+7];
cvt.u32.u16 %r249, %rs26;
prmt.b32 %r250, %r249, %r248, 28756;
ld.global.nc.u8 %rs27, [%rd65+8];
cvt.u32.u16 %r251, %rs27;
prmt.b32 %r252, %r251, %r250, 1620;
and.b32 %r253, %r252, -2004318072;
shr.u32 %r254, %r253, 1;
or.b32 %r255, %r254, 839922192;
and.b32 %r256, %r248, 30583;
prmt.b32 %r257, %r81, %r80, %r256;
prmt.b32 %r258, %r84, %r83, %r256;
and.b32 %r259, %r255, 30292;
prmt.b32 %r260, %r257, %r258, %r259;
shr.u32 %r261, %r252, 16;
and.b32 %r262, %r261, 30583;
prmt.b32 %r263, %r81, %r80, %r262;
prmt.b32 %r264, %r84, %r83, %r262;
shr.u32 %r265, %r255, 16;
prmt.b32 %r266, %r263, %r264, %r265;
prmt.b32 %r200, %r260, %r266, %r94;
prmt.b32 %r204, %r260, %r266, %r95;

dp4a.s32.s32 %r199, %r200, %r9, %r195;


dp4a.s32.s32 %r203, %r204, %r10, %r199;

add.s64 %rd66, %rd106, %rd63;
ld.global.nc.u8 %rs28, [%rd66];
shl.b16 %rs30, %rs28, 7;
and.b16 %rs31, %rs30, 32640;
setp.eq.s16 %p8, %rs28, 255;
setp.eq.s16 %p9, %rs28, 0;
selp.b16 %rs32, 64, %rs31, %p9;
selp.b16 %rs19, 32767, %rs32, %p8;

{ mov.b32 %f52, {0,%rs19};}


mul.ftz.f32 %f54, %f52, 0f3F000000;

{.reg .f16 low,high;
mov.b32 {low,high},%r11;
cvt.f32.f16 %f53, low;}


mul.ftz.f32 %f55, %f54, %f53;
cvt.rn.f32.s32 %f56, %r203;
fma.rn.ftz.f32 %f168, %f55, %f56, %f168;
@%p5 bra $L__BB101_7;

add.s64 %rd70, %rd104, %rd64;
ld.global.u8 %r293, [%rd70+1];
ld.global.u8 %r294, [%rd70+2];
prmt.b32 %r295, %r294, %r293, 30212;
ld.global.u8 %r296, [%rd70+3];
prmt.b32 %r297, %r296, %r295, 28756;
ld.global.u8 %r298, [%rd70+4];
prmt.b32 %r299, %r298, %r297, 1620;
and.b32 %r300, %r299, -2004318072;
shr.u32 %r301, %r300, 1;
or.b32 %r302, %r301, 839922192;
and.b32 %r303, %r295, 30583;
mov.u32 %r304, 201852420;
mov.u32 %r305, 50462976;
prmt.b32 %r306, %r305, %r304, %r303;
mov.u32 %r307, -185009412;
mov.u32 %r308, -33620224;
prmt.b32 %r309, %r308, %r307, %r303;
and.b32 %r310, %r302, 30292;
prmt.b32 %r311, %r306, %r309, %r310;
shr.u32 %r312, %r299, 16;
and.b32 %r313, %r312, 30583;
prmt.b32 %r314, %r305, %r304, %r313;
prmt.b32 %r315, %r308, %r307, %r313;
shr.u32 %r316, %r302, 16;
prmt.b32 %r317, %r314, %r315, %r316;
mov.u32 %r318, 25632;
prmt.b32 %r268, %r311, %r317, %r318;
mov.u32 %r319, 30001;
prmt.b32 %r272, %r311, %r317, %r319;
mov.u32 %r270, 0;

dp4a.s32.s32 %r267, %r268, %r7, %r270;


dp4a.s32.s32 %r271, %r272, %r8, %r267;

ld.global.u8 %r320, [%rd70+5];
ld.global.u8 %r321, [%rd70+6];
prmt.b32 %r322, %r321, %r320, 30212;
ld.global.u8 %r323, [%rd70+7];
prmt.b32 %r324, %r323, %r322, 28756;
ld.global.u8 %r325, [%rd70+8];
prmt.b32 %r326, %r325, %r324, 1620;
and.b32 %r327, %r326, -2004318072;
shr.u32 %r328, %r327, 1;
or.b32 %r329, %r328, 839922192;
and.b32 %r330, %r322, 30583;
prmt.b32 %r331, %r305, %r304, %r330;
prmt.b32 %r332, %r308, %r307, %r330;
and.b32 %r333, %r329, 30292;
prmt.b32 %r334, %r331, %r332, %r333;
shr.u32 %r335, %r326, 16;
and.b32 %r336, %r335, 30583;
prmt.b32 %r337, %r305, %r304, %r336;
prmt.b32 %r338, %r308, %r307, %r336;
shr.u32 %r339, %r329, 16;
prmt.b32 %r340, %r337, %r338, %r339;
prmt.b32 %r276, %r334, %r340, %r318;
prmt.b32 %r280, %r334, %r340, %r319;

dp4a.s32.s32 %r275, %r276, %r9, %r271;


dp4a.s32.s32 %r279, %r280, %r10, %r275;

add.s64 %rd71, %rd104, %rd63;
ld.global.u8 %rs34, [%rd71];
shl.b16 %rs35, %rs34, 7;
setp.eq.s16 %p11, %rs34, 255;
setp.eq.s16 %p12, %rs34, 0;
selp.b16 %rs36, 64, %rs35, %p12;
selp.b16 %rs33, 32767, %rs36, %p11;

{ mov.b32 %f57, {0,%rs33};}


mul.ftz.f32 %f59, %f57, 0f3F000000;

{.reg .f16 low,high;
mov.b32 {low,high},%r11;
cvt.f32.f16 %f58, low;}


mul.ftz.f32 %f60, %f59, %f58;
cvt.rn.f32.s32 %f61, %r279;
fma.rn.ftz.f32 %f166, %f60, %f61, %f166;

$L__BB101_7:
add.s64 %rd107, %rd107, 576;
add.s64 %rd106, %rd106, 272;
add.s64 %rd105, %rd105, 576;
add.s64 %rd104, %rd104, 272;
add.s32 %r434, %r434, 16;
setp.lt.u32 %p13, %r434, %r31;
@%p13 bra $L__BB101_3;

add.u64 %rd96, %SPL, 0;
add.u64 %rd90, %SPL, 8;
st.local.v2.f32 [%rd96], {%f169, %f168};
st.local.v2.f32 [%rd90], {%f167, %f166};

$L__BB101_9:
mov.b32 %r13, %f169;
setp.ne.s64 %p14, %rd23, 0;
@%p14 bra $L__BB101_11;
bra.uni $L__BB101_10;

$L__BB101_11:
add.u64 %rd102, %SPL, 0;
add.u64 %rd94, %SPL, 8;
mov.u32 %r368, 31;
mov.u32 %r369, 16;
mov.u32 %r370, -1;
shfl.sync.bfly.b32 %r371|%p25, %r13, %r369, %r368, %r370;
mov.b32 %f82, %r371;
add.ftz.f32 %f83, %f169, %f82;
mov.b32 %r372, %f83;
mov.u32 %r373, 8;
shfl.sync.bfly.b32 %r374|%p26, %r372, %r373, %r368, %r370;
mov.b32 %f84, %r374;
add.ftz.f32 %f85, %f83, %f84;
mov.b32 %r375, %f85;
mov.u32 %r376, 4;
shfl.sync.bfly.b32 %r377|%p27, %r375, %r376, %r368, %r370;
mov.b32 %f86, %r377;
add.ftz.f32 %f87, %f85, %f86;
mov.b32 %r378, %f87;
mov.u32 %r379, 2;
shfl.sync.bfly.b32 %r380|%p28, %r378, %r379, %r368, %r370;
mov.b32 %f88, %r380;
add.ftz.f32 %f89, %f87, %f88;
mov.b32 %r381, %f89;
mov.u32 %r382, 1;
shfl.sync.bfly.b32 %r383|%p29, %r381, %r382, %r368, %r370;
mov.b32 %f90, %r383;
add.ftz.f32 %f91, %f89, %f90;
st.local.f32 [%rd102], %f91;
mov.b32 %r384, %f167;
shfl.sync.bfly.b32 %r385|%p30, %r384, %r369, %r368, %r370;
mov.b32 %f92, %r385;
add.ftz.f32 %f93, %f167, %f92;
mov.b32 %r386, %f93;
shfl.sync.bfly.b32 %r387|%p31, %r386, %r373, %r368, %r370;
mov.b32 %f94, %r387;
add.ftz.f32 %f95, %f93, %f94;
mov.b32 %r388, %f95;
shfl.sync.bfly.b32 %r389|%p32, %r388, %r376, %r368, %r370;
mov.b32 %f96, %r389;
add.ftz.f32 %f97, %f95, %f96;
mov.b32 %r390, %f97;
shfl.sync.bfly.b32 %r391|%p33, %r390, %r379, %r368, %r370;
mov.b32 %f98, %r391;
add.ftz.f32 %f99, %f97, %f98;
mov.b32 %r392, %f99;
shfl.sync.bfly.b32 %r393|%p34, %r392, %r382, %r368, %r370;
mov.b32 %f100, %r393;
add.ftz.f32 %f101, %f99, %f100;
st.local.f32 [%rd94], %f101;
mov.b32 %r394, %f168;
shfl.sync.bfly.b32 %r395|%p35, %r394, %r369, %r368, %r370;
mov.b32 %f102, %r395;
add.ftz.f32 %f103, %f168, %f102;
mov.b32 %r396, %f103;
shfl.sync.bfly.b32 %r397|%p36, %r396, %r373, %r368, %r370;
mov.b32 %f104, %r397;
add.ftz.f32 %f105, %f103, %f104;
mov.b32 %r398, %f105;
shfl.sync.bfly.b32 %r399|%p37, %r398, %r376, %r368, %r370;
mov.b32 %f106, %r399;
add.ftz.f32 %f107, %f105, %f106;
mov.b32 %r400, %f107;
shfl.sync.bfly.b32 %r401|%p38, %r400, %r379, %r368, %r370;
mov.b32 %f108, %r401;
add.ftz.f32 %f109, %f107, %f108;
mov.b32 %r402, %f109;
shfl.sync.bfly.b32 %r403|%p39, %r402, %r382, %r368, %r370;
mov.b32 %f110, %r403;
add.ftz.f32 %f111, %f109, %f110;
st.local.f32 [%rd102+4], %f111;
mov.b32 %r404, %f166;
shfl.sync.bfly.b32 %r405|%p40, %r404, %r369, %r368, %r370;
mov.b32 %f112, %r405;
add.ftz.f32 %f113, %f166, %f112;
mov.b32 %r406, %f113;
shfl.sync.bfly.b32 %r407|%p41, %r406, %r373, %r368, %r370;
mov.b32 %f114, %r407;
add.ftz.f32 %f115, %f113, %f114;
mov.b32 %r408, %f115;
shfl.sync.bfly.b32 %r409|%p42, %r408, %r376, %r368, %r370;
mov.b32 %f116, %r409;
add.ftz.f32 %f117, %f115, %f116;
mov.b32 %r410, %f117;
shfl.sync.bfly.b32 %r411|%p43, %r410, %r379, %r368, %r370;
mov.b32 %f118, %r411;
add.ftz.f32 %f119, %f117, %f118;
mov.b32 %r412, %f119;
shfl.sync.bfly.b32 %r413|%p44, %r412, %r382, %r368, %r370;
mov.b32 %f120, %r413;
add.ftz.f32 %f121, %f119, %f120;
st.local.f32 [%rd94+4], %f121;
bra.uni $L__BB101_12;

$L__BB101_10:
add.u64 %rd98, %SPL, 0;
mov.u32 %r342, 31;
mov.u32 %r343, 16;
mov.u32 %r344, -1;
shfl.sync.bfly.b32 %r345|%p15, %r13, %r343, %r342, %r344;
mov.b32 %f62, %r345;
add.ftz.f32 %f63, %f169, %f62;
mov.b32 %r346, %f63;
mov.u32 %r347, 8;
shfl.sync.bfly.b32 %r348|%p16, %r346, %r347, %r342, %r344;
mov.b32 %f64, %r348;
add.ftz.f32 %f65, %f63, %f64;
mov.b32 %r349, %f65;
mov.u32 %r350, 4;
shfl.sync.bfly.b32 %r351|%p17, %r349, %r350, %r342, %r344;
mov.b32 %f66, %r351;
add.ftz.f32 %f67, %f65, %f66;
mov.b32 %r352, %f67;
mov.u32 %r353, 2;
shfl.sync.bfly.b32 %r354|%p18, %r352, %r353, %r342, %r344;
mov.b32 %f68, %r354;
add.ftz.f32 %f69, %f67, %f68;
mov.b32 %r355, %f69;
mov.u32 %r356, 1;
shfl.sync.bfly.b32 %r357|%p19, %r355, %r356, %r342, %r344;
mov.b32 %f70, %r357;
add.ftz.f32 %f71, %f69, %f70;
st.local.f32 [%rd98], %f71;
mov.b32 %r358, %f168;
shfl.sync.bfly.b32 %r359|%p20, %r358, %r343, %r342, %r344;
mov.b32 %f72, %r359;
add.ftz.f32 %f73, %f168, %f72;
mov.b32 %r360, %f73;
shfl.sync.bfly.b32 %r361|%p21, %r360, %r347, %r342, %r344;
mov.b32 %f74, %r361;
add.ftz.f32 %f75, %f73, %f74;
mov.b32 %r362, %f75;
shfl.sync.bfly.b32 %r363|%p22, %r362, %r350, %r342, %r344;
mov.b32 %f76, %r363;
add.ftz.f32 %f77, %f75, %f76;
mov.b32 %r364, %f77;
shfl.sync.bfly.b32 %r365|%p23, %r364, %r353, %r342, %r344;
mov.b32 %f78, %r365;
add.ftz.f32 %f79, %f77, %f78;
mov.b32 %r366, %f79;
shfl.sync.bfly.b32 %r367|%p24, %r366, %r356, %r342, %r344;
mov.b32 %f80, %r367;
add.ftz.f32 %f81, %f79, %f80;
st.local.f32 [%rd98+4], %f81;

$L__BB101_12:
ld.param.u32 %r423, [_Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7];
setp.gt.u32 %p45, %r4, 1;
mov.u32 %r415, %ctaid.x;
shl.b32 %r416, %r415, 1;
add.s32 %r14, %r416, %r4;
setp.ge.u32 %p46, %r14, %r423;
or.pred %p47, %p45, %p46;
@%p47 bra $L__BB101_29;

add.u64 %rd100, %SPL, 0;
ld.param.u64 %rd85, [_Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3];
ld.param.u32 %r424, [_Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13];
cvt.u64.u32 %rd17, %r4;
mul.wide.u32 %rd72, %r4, 4;
add.s64 %rd73, %rd100, %rd72;
ld.local.f32 %f172, [%rd73];
mad.lo.s32 %r15, %r3, %r424, %r14;
setp.eq.s64 %p48, %rd85, 0;
@%p48 bra $L__BB101_15;

ld.param.u64 %rd86, [_Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3];
cvta.to.global.u64 %rd74, %rd86;
mul.wide.u32 %rd75, %r15, 4;
add.s64 %rd76, %rd74, %rd75;
ld.global.f32 %f122, [%rd76];
add.ftz.f32 %f172, %f172, %f122;

$L__BB101_15:
setp.eq.s64 %p49, %rd23, 0;
@%p49 bra $L__BB101_28;

add.u64 %rd92, %SPL, 8;
ld.param.u64 %rd88, [_Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+16];
shl.b64 %rd77, %rd17, 2;
add.s64 %rd78, %rd92, %rd77;
ld.local.f32 %f24, [%rd78];
setp.eq.s64 %p50, %rd88, 0;
@%p50 bra $L__BB101_18;

ld.param.u64 %rd89, [_Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+16];
cvta.to.global.u64 %rd79, %rd89;
mul.wide.u32 %rd80, %r15, 4;
add.s64 %rd81, %rd79, %rd80;
ld.global.f32 %f123, [%rd81];
add.ftz.f32 %f24, %f24, %f123;

$L__BB101_18:
ld.param.u32 %r429, [_Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.gt.s32 %p51, %r429, 2;
@%p51 bra $L__BB101_22;

ld.param.u32 %r432, [_Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.eq.s32 %p54, %r432, 1;
@%p54 bra $L__BB101_26;

ld.param.u32 %r433, [_Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.eq.s32 %p55, %r433, 2;
@%p55 bra $L__BB101_21;
bra.uni $L__BB101_27;

$L__BB101_21:
mul.ftz.f32 %f143, %f24, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f144, %f143;
add.ftz.f32 %f145, %f144, 0f3F800000;
div.approx.ftz.f32 %f146, %f24, %f145;
mul.ftz.f32 %f172, %f172, %f146;
bra.uni $L__BB101_28;

$L__BB101_22:
ld.param.u32 %r430, [_Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.eq.s32 %p52, %r430, 3;
@%p52 bra $L__BB101_25;

ld.param.u32 %r431, [_Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.ne.s32 %p53, %r431, 6;
@%p53 bra $L__BB101_27;

ld.param.f32 %f155, [_Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+44];
min.ftz.f32 %f124, %f24, %f155;
neg.ftz.f32 %f125, %f155;
min.ftz.f32 %f126, %f172, %f155;
max.ftz.f32 %f127, %f126, %f125;
mul.ftz.f32 %f128, %f124, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f129, %f128;
add.ftz.f32 %f130, %f129, 0f3F800000;
div.approx.ftz.f32 %f131, %f124, %f130;
mul.ftz.f32 %f172, %f127, %f131;
bra.uni $L__BB101_28;

$L__BB101_27:
mul.ftz.f32 %f172, %f172, %f24;
bra.uni $L__BB101_28;

$L__BB101_26:
mul.ftz.f32 %f147, %f24, 0f3F000000;
mul.ftz.f32 %f148, %f24, 0f3D372713;
fma.rn.ftz.f32 %f149, %f24, %f148, 0f3F800000;
mul.ftz.f32 %f150, %f24, 0f3F4C422A;
mul.ftz.f32 %f151, %f150, %f149;
tanh.approx.f32 %f152, %f151;
add.ftz.f32 %f153, %f152, 0f3F800000;
mul.ftz.f32 %f154, %f147, %f153;
mul.ftz.f32 %f172, %f172, %f154;
bra.uni $L__BB101_28;

$L__BB101_25:
mov.f32 %f132, 0f40E00000;
min.ftz.f32 %f133, %f24, %f132;
min.ftz.f32 %f134, %f172, %f132;
mov.f32 %f135, 0fC0E00000;
max.ftz.f32 %f136, %f134, %f135;
mul.ftz.f32 %f137, %f133, 0f3FD9DB23;
mul.ftz.f32 %f138, %f137, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f139, %f138;
add.ftz.f32 %f140, %f139, 0f3F800000;
div.approx.ftz.f32 %f141, %f133, %f140;
add.ftz.f32 %f142, %f136, 0f3F800000;
mul.ftz.f32 %f172, %f142, %f141;

$L__BB101_28:
ld.param.u64 %rd87, [_Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4];
ld.param.u32 %r428, [_Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10];
ld.param.u32 %r427, [_Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13];
mov.u32 %r426, %ctaid.y;
mov.u32 %r425, %tid.y;
mad.lo.s32 %r421, %r425, %r428, %r14;
mad.lo.s32 %r422, %r426, %r427, %r421;
cvta.to.global.u64 %rd82, %rd87;
mul.wide.u32 %rd83, %r422, 4;
add.s64 %rd84, %rd82, %rd83;
st.global.f32 [%rd84], %f172;

$L__BB101_29:
ret;

}
.entry _Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj(
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2,
.param .align 8 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3[48],
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5,
.param .align 4 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6[12],
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15
)
.maxntid 256, 1, 1
.minnctapersm 1
{
.local .align 8 .b8 __local_depot102[8];
.reg .b64 %SP;
.reg .b64 %SPL;
.reg .pred %p<21>;
.reg .b16 %rs<29>;
.reg .f32 %f<46>;
.reg .b32 %r<201>;
.reg .b64 %rd<54>;


mov.u64 %SPL, __local_depot102;
ld.param.u64 %rd17, [_Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0];
ld.param.u64 %rd18, [_Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1];
ld.param.u64 %rd19, [_Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2];
ld.param.u64 %rd20, [_Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4];
ld.param.u32 %r20, [_Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5];
ld.param.u32 %r12, [_Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7];
ld.param.u32 %r13, [_Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8];
ld.param.u32 %r14, [_Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9];
ld.param.u32 %r15, [_Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10];
ld.param.u32 %r16, [_Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11];
ld.param.u32 %r17, [_Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12];
ld.param.u32 %r18, [_Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13];
ld.param.u32 %r21, [_Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14];
ld.param.u32 %r19, [_Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15];
ld.param.u32 %r11, [_Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+8];
ld.param.u32 %r10, [_Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+4];
ld.param.u32 %r9, [_Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6];
add.u64 %rd1, %SPL, 0;
mov.u32 %r22, %ctaid.x;
shl.b32 %r1, %r22, 1;
shr.u32 %r2, %r20, 5;
mov.u32 %r3, %tid.y;
setp.ge.u32 %p1, %r3, %r21;
@%p1 bra $L__BB102_7;

mov.f32 %f44, 0f00000000;
st.local.v2.f32 [%rd1], {%f44, %f44};
mov.u32 %r4, %tid.x;
shr.u32 %r200, %r4, 1;
setp.ge.u32 %p2, %r200, %r2;
mov.f32 %f45, %f44;
@%p2 bra $L__BB102_5;

mov.u32 %r23, %ctaid.y;
cvta.to.global.u64 %rd22, %rd18;
mul.hi.u32 %r24, %r23, %r9;
add.s32 %r25, %r24, %r23;
shr.u32 %r26, %r25, %r10;
mul.lo.s32 %r27, %r26, %r11;
sub.s32 %r28, %r23, %r27;
mul.lo.s32 %r29, %r28, %r17;
cvt.u64.u32 %rd23, %r29;
mul.lo.s32 %r30, %r3, %r14;
cvt.u64.u32 %rd24, %r30;
mad.lo.s32 %r31, %r3, %r19, %r23;
cvta.to.global.u64 %rd25, %rd19;
mul.wide.u32 %rd26, %r31, 4;
add.s64 %rd27, %rd25, %rd26;
ld.global.nc.u32 %r32, [%rd27];
mul.lo.s32 %r33, %r32, %r16;
shl.b32 %r34, %r4, 3;
and.b32 %r35, %r34, 8;
cvt.u64.u32 %rd28, %r35;
cvt.u64.u32 %rd29, %r4;
shr.u64 %rd30, %rd29, 1;
mul.lo.s64 %rd31, %rd30, 36;
mul.wide.u32 %rd32, %r30, 36;
add.s64 %rd33, %rd31, %rd32;
add.s64 %rd3, %rd22, %rd33;
mul.wide.u32 %rd53, %r29, 36;
mad.lo.s32 %r36, %r1, %r13, %r200;
add.s32 %r37, %r36, %r33;
mul.wide.s32 %rd5, %r37, 17;
add.s64 %rd34, %rd28, %rd5;
add.s64 %rd6, %rd34, 8;
add.s64 %rd35, %rd30, %rd24;
add.s64 %rd36, %rd35, %rd23;
mul.lo.s64 %rd37, %rd36, 36;
add.s64 %rd38, %rd37, %rd28;
add.s64 %rd39, %rd22, %rd38;
add.s64 %rd51, %rd39, 20;
add.s32 %r38, %r1, 1;
add.s32 %r39, %r200, %r33;
mad.lo.s32 %r40, %r13, %r38, %r39;
mul.wide.s32 %rd8, %r40, 17;
add.s64 %rd40, %rd28, %rd8;
add.s64 %rd9, %rd40, 8;
cvta.to.global.u64 %rd52, %rd17;
mov.f32 %f44, 0f00000000;

$L__BB102_3:
add.s64 %rd41, %rd3, %rd53;
add.s64 %rd42, %rd52, %rd6;
ld.global.nc.u8 %rs3, [%rd42+-7];
cvt.u32.u16 %r75, %rs3;
and.b32 %r76, %r75, 255;
ld.global.nc.u8 %rs4, [%rd42+-6];
cvt.u32.u16 %r77, %rs4;
prmt.b32 %r78, %r77, %r76, 30212;
ld.global.nc.u8 %rs5, [%rd42+-5];
cvt.u32.u16 %r79, %rs5;
prmt.b32 %r80, %r79, %r78, 28756;
ld.global.nc.u8 %rs6, [%rd42+-4];
cvt.u32.u16 %r81, %rs6;
prmt.b32 %r82, %r81, %r80, 1620;
and.b32 %r83, %r82, -2004318072;
shr.u32 %r84, %r83, 1;
or.b32 %r85, %r84, 839922192;
and.b32 %r86, %r78, 30583;
mov.u32 %r87, 201852420;
mov.u32 %r88, 50462976;
prmt.b32 %r89, %r88, %r87, %r86;
mov.u32 %r90, -185009412;
mov.u32 %r91, -33620224;
prmt.b32 %r92, %r91, %r90, %r86;
and.b32 %r93, %r85, 30292;
prmt.b32 %r94, %r89, %r92, %r93;
shr.u32 %r95, %r82, 16;
and.b32 %r96, %r95, 30583;
prmt.b32 %r97, %r88, %r87, %r96;
prmt.b32 %r98, %r91, %r90, %r96;
shr.u32 %r99, %r85, 16;
prmt.b32 %r100, %r97, %r98, %r99;
mov.u32 %r101, 25632;
prmt.b32 %r42, %r94, %r100, %r101;
mov.u32 %r102, 30001;
prmt.b32 %r46, %r94, %r100, %r102;
ld.global.nc.u32 %r43, [%rd51+-16];
mov.u32 %r61, 0;

dp4a.s32.s32 %r41, %r42, %r43, %r61;

ld.global.nc.u32 %r47, [%rd51];

dp4a.s32.s32 %r45, %r46, %r47, %r41;

ld.global.nc.u8 %rs7, [%rd42+-3];
cvt.u32.u16 %r103, %rs7;
and.b32 %r104, %r103, 255;
ld.global.nc.u8 %rs8, [%rd42+-2];
cvt.u32.u16 %r105, %rs8;
prmt.b32 %r106, %r105, %r104, 30212;
ld.global.nc.u8 %rs9, [%rd42+-1];
cvt.u32.u16 %r107, %rs9;
prmt.b32 %r108, %r107, %r106, 28756;
ld.global.nc.u8 %rs10, [%rd42];
cvt.u32.u16 %r109, %rs10;
prmt.b32 %r110, %r109, %r108, 1620;
and.b32 %r111, %r110, -2004318072;
shr.u32 %r112, %r111, 1;
or.b32 %r113, %r112, 839922192;
and.b32 %r114, %r106, 30583;
prmt.b32 %r115, %r88, %r87, %r114;
prmt.b32 %r116, %r91, %r90, %r114;
and.b32 %r117, %r113, 30292;
prmt.b32 %r118, %r115, %r116, %r117;
shr.u32 %r119, %r110, 16;
and.b32 %r120, %r119, 30583;
prmt.b32 %r121, %r88, %r87, %r120;
prmt.b32 %r122, %r91, %r90, %r120;
shr.u32 %r123, %r113, 16;
prmt.b32 %r124, %r121, %r122, %r123;
prmt.b32 %r50, %r118, %r124, %r101;
prmt.b32 %r54, %r118, %r124, %r102;
ld.global.nc.u32 %r51, [%rd51+-12];

dp4a.s32.s32 %r49, %r50, %r51, %r45;

ld.global.nc.u32 %r55, [%rd51+4];

dp4a.s32.s32 %r53, %r54, %r55, %r49;

add.s64 %rd43, %rd52, %rd5;
ld.global.nc.u8 %rs11, [%rd43];
shl.b16 %rs13, %rs11, 7;
and.b16 %rs14, %rs13, 32640;
setp.eq.s16 %p3, %rs11, 255;
setp.eq.s16 %p4, %rs11, 0;
selp.b16 %rs15, 64, %rs14, %p4;
selp.b16 %rs1, 32767, %rs15, %p3;

{ mov.b32 %f11, {0,%rs1};}


mul.ftz.f32 %f15, %f11, 0f3F000000;
ld.global.nc.u32 %r57, [%rd41];

{.reg .f16 low,high;
mov.b32 {low,high},%r57;
cvt.f32.f16 %f12, low;}


mul.ftz.f32 %f16, %f15, %f12;
cvt.rn.f32.s32 %f17, %r53;
fma.rn.ftz.f32 %f45, %f16, %f17, %f45;
add.s64 %rd44, %rd52, %rd9;
ld.global.nc.u8 %rs16, [%rd44+-7];
cvt.u32.u16 %r125, %rs16;
and.b32 %r126, %r125, 255;
ld.global.nc.u8 %rs17, [%rd44+-6];
cvt.u32.u16 %r127, %rs17;
prmt.b32 %r128, %r127, %r126, 30212;
ld.global.nc.u8 %rs18, [%rd44+-5];
cvt.u32.u16 %r129, %rs18;
prmt.b32 %r130, %r129, %r128, 28756;
ld.global.nc.u8 %rs19, [%rd44+-4];
cvt.u32.u16 %r131, %rs19;
prmt.b32 %r132, %r131, %r130, 1620;
and.b32 %r133, %r132, -2004318072;
shr.u32 %r134, %r133, 1;
or.b32 %r135, %r134, 839922192;
and.b32 %r136, %r128, 30583;
prmt.b32 %r137, %r88, %r87, %r136;
prmt.b32 %r138, %r91, %r90, %r136;
and.b32 %r139, %r135, 30292;
prmt.b32 %r140, %r137, %r138, %r139;
shr.u32 %r141, %r132, 16;
and.b32 %r142, %r141, 30583;
prmt.b32 %r143, %r88, %r87, %r142;
prmt.b32 %r144, %r91, %r90, %r142;
shr.u32 %r145, %r135, 16;
prmt.b32 %r146, %r143, %r144, %r145;
prmt.b32 %r59, %r140, %r146, %r101;
prmt.b32 %r63, %r140, %r146, %r102;

dp4a.s32.s32 %r58, %r59, %r43, %r61;


dp4a.s32.s32 %r62, %r63, %r47, %r58;

ld.global.nc.u8 %rs20, [%rd44+-3];
cvt.u32.u16 %r147, %rs20;
and.b32 %r148, %r147, 255;
ld.global.nc.u8 %rs21, [%rd44+-2];
cvt.u32.u16 %r149, %rs21;
prmt.b32 %r150, %r149, %r148, 30212;
ld.global.nc.u8 %rs22, [%rd44+-1];
cvt.u32.u16 %r151, %rs22;
prmt.b32 %r152, %r151, %r150, 28756;
ld.global.nc.u8 %rs23, [%rd44];
cvt.u32.u16 %r153, %rs23;
prmt.b32 %r154, %r153, %r152, 1620;
and.b32 %r155, %r154, -2004318072;
shr.u32 %r156, %r155, 1;
or.b32 %r157, %r156, 839922192;
and.b32 %r158, %r150, 30583;
prmt.b32 %r159, %r88, %r87, %r158;
prmt.b32 %r160, %r91, %r90, %r158;
and.b32 %r161, %r157, 30292;
prmt.b32 %r162, %r159, %r160, %r161;
shr.u32 %r163, %r154, 16;
and.b32 %r164, %r163, 30583;
prmt.b32 %r165, %r88, %r87, %r164;
prmt.b32 %r166, %r91, %r90, %r164;
shr.u32 %r167, %r157, 16;
prmt.b32 %r168, %r165, %r166, %r167;
prmt.b32 %r67, %r162, %r168, %r101;
prmt.b32 %r71, %r162, %r168, %r102;

dp4a.s32.s32 %r66, %r67, %r51, %r62;


dp4a.s32.s32 %r70, %r71, %r55, %r66;

add.s64 %rd45, %rd52, %rd8;
ld.global.nc.u8 %rs24, [%rd45];
shl.b16 %rs26, %rs24, 7;
and.b16 %rs27, %rs26, 32640;
setp.eq.s16 %p5, %rs24, 255;
setp.eq.s16 %p6, %rs24, 0;
selp.b16 %rs28, 64, %rs27, %p6;
selp.b16 %rs2, 32767, %rs28, %p5;

{ mov.b32 %f13, {0,%rs2};}


mul.ftz.f32 %f18, %f13, 0f3F000000;

{.reg .f16 low,high;
mov.b32 {low,high},%r57;
cvt.f32.f16 %f14, low;}


mul.ftz.f32 %f19, %f18, %f14;
cvt.rn.f32.s32 %f20, %r70;
fma.rn.ftz.f32 %f44, %f19, %f20, %f44;
add.s64 %rd53, %rd53, 576;
add.s64 %rd52, %rd52, 272;
add.s64 %rd51, %rd51, 576;
add.s32 %r200, %r200, 16;
setp.lt.u32 %p7, %r200, %r2;
@%p7 bra $L__BB102_3;

st.local.v2.f32 [%rd1], {%f45, %f44};

$L__BB102_5:
mov.b32 %r169, %f45;
mov.u32 %r170, 31;
mov.u32 %r171, 16;
mov.u32 %r172, -1;
shfl.sync.bfly.b32 %r173|%p8, %r169, %r171, %r170, %r172;
mov.b32 %f21, %r173;
add.ftz.f32 %f22, %f45, %f21;
mov.b32 %r174, %f22;
mov.u32 %r175, 8;
shfl.sync.bfly.b32 %r176|%p9, %r174, %r175, %r170, %r172;
mov.b32 %f23, %r176;
add.ftz.f32 %f24, %f22, %f23;
mov.b32 %r177, %f24;
mov.u32 %r178, 4;
shfl.sync.bfly.b32 %r179|%p10, %r177, %r178, %r170, %r172;
mov.b32 %f25, %r179;
add.ftz.f32 %f26, %f24, %f25;
mov.b32 %r180, %f26;
mov.u32 %r181, 2;
shfl.sync.bfly.b32 %r182|%p11, %r180, %r181, %r170, %r172;
mov.b32 %f27, %r182;
add.ftz.f32 %f28, %f26, %f27;
mov.b32 %r183, %f28;
mov.u32 %r184, 1;
shfl.sync.bfly.b32 %r185|%p12, %r183, %r184, %r170, %r172;
mov.b32 %f29, %r185;
add.ftz.f32 %f30, %f28, %f29;
st.local.f32 [%rd1], %f30;
mov.b32 %r186, %f44;
shfl.sync.bfly.b32 %r187|%p13, %r186, %r171, %r170, %r172;
mov.b32 %f31, %r187;
add.ftz.f32 %f32, %f44, %f31;
mov.b32 %r188, %f32;
shfl.sync.bfly.b32 %r189|%p14, %r188, %r175, %r170, %r172;
mov.b32 %f33, %r189;
add.ftz.f32 %f34, %f32, %f33;
mov.b32 %r190, %f34;
shfl.sync.bfly.b32 %r191|%p15, %r190, %r178, %r170, %r172;
mov.b32 %f35, %r191;
add.ftz.f32 %f36, %f34, %f35;
mov.b32 %r192, %f36;
shfl.sync.bfly.b32 %r193|%p16, %r192, %r181, %r170, %r172;
mov.b32 %f37, %r193;
add.ftz.f32 %f38, %f36, %f37;
mov.b32 %r194, %f38;
shfl.sync.bfly.b32 %r195|%p17, %r194, %r184, %r170, %r172;
mov.b32 %f39, %r195;
add.ftz.f32 %f40, %f38, %f39;
st.local.f32 [%rd1+4], %f40;
add.s32 %r8, %r1, %r4;
setp.ge.u32 %p18, %r8, %r12;
setp.gt.u32 %p19, %r4, 1;
or.pred %p20, %p19, %p18;
@%p20 bra $L__BB102_7;

mul.wide.u32 %rd46, %r4, 4;
add.s64 %rd47, %rd1, %rd46;
ld.local.f32 %f41, [%rd47];
mov.u32 %r196, %ctaid.y;
mad.lo.s32 %r198, %r3, %r15, %r8;
mad.lo.s32 %r199, %r196, %r18, %r198;
cvta.to.global.u64 %rd48, %rd20;
mul.wide.u32 %rd49, %r199, 4;
add.s64 %rd50, %rd48, %rd49;
st.global.f32 [%rd50], %f41;

$L__BB102_7:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<336>;
.reg .b16 %rs<73>;
.reg .f32 %f<1013>;
.reg .b32 %r<1304>;
.reg .b64 %rd<97>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate[1536];

ld.param.u64 %rd25, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd26, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd27, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r48, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r49, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r50, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r54, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r55, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r56, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r52, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r51, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r46, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r45, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
ld.param.u64 %rd30, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+16];
ld.param.u64 %rd29, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+8];
ld.param.u64 %rd28, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3];
mov.u32 %r1, %tid.y;
shl.b32 %r57, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r57, %r2;
mov.u32 %r58, %ctaid.x;
shl.b32 %r4, %r58, 2;
mov.u32 %r5, %ctaid.y;
setp.eq.s64 %p1, %rd27, 0;
@%p1 bra $L__BB103_2;

cvta.to.global.u64 %rd34, %rd27;
mul.wide.u32 %rd35, %r5, 4;
add.s64 %rd36, %rd34, %rd35;
ld.global.nc.u32 %r1301, [%rd36];
mul.hi.u32 %r59, %r5, %r40;
add.s32 %r60, %r59, %r5;
shr.u32 %r61, %r60, %r41;
mul.lo.s32 %r62, %r61, %r42;
sub.s32 %r1302, %r5, %r62;
bra.uni $L__BB103_3;

$L__BB103_2:
mul.hi.u32 %r63, %r5, %r45;
add.s32 %r64, %r63, %r5;
shr.u32 %r1301, %r64, %r46;
mov.u32 %r1302, %r5;

$L__BB103_3:
mov.u32 %r11, %ctaid.z;
selp.b32 %r12, %r5, %r1301, %p1;
setp.ne.s32 %p3, %r1, 0;
setp.gt.u32 %p4, %r2, 3;
or.pred %p5, %p4, %p3;
add.s32 %r65, %r4, %r2;
setp.ge.u32 %p6, %r65, %r44;
mov.f32 %f27, 0f00000000;
or.pred %p7, %p5, %p6;
mov.f32 %f4, %f27;
mov.f32 %f5, %f27;
@%p7 bra $L__BB103_8;

setp.eq.s64 %p8, %rd28, 0;
mov.f32 %f4, 0f00000000;
mov.f32 %f5, %f4;
@%p8 bra $L__BB103_6;

cvta.to.global.u64 %rd37, %rd28;
mul.lo.s32 %r66, %r11, %r56;
cvt.u64.u32 %rd38, %r66;
mul.lo.s32 %r67, %r12, %r50;
cvt.u64.u32 %rd39, %r67;
cvt.u64.u32 %rd40, %r2;
cvt.s64.s32 %rd41, %r4;
add.s64 %rd42, %rd41, %rd40;
add.s64 %rd43, %rd42, %rd38;
add.s64 %rd44, %rd43, %rd39;
shl.b64 %rd45, %rd44, 2;
add.s64 %rd46, %rd37, %rd45;
ld.global.f32 %f5, [%rd46];

$L__BB103_6:
setp.eq.s64 %p9, %rd30, 0;
setp.eq.s64 %p10, %rd29, 0;
or.pred %p11, %p9, %p10;
@%p11 bra $L__BB103_8;

cvta.to.global.u64 %rd47, %rd30;
mul.lo.s32 %r68, %r11, %r56;
cvt.u64.u32 %rd48, %r68;
mul.lo.s32 %r69, %r12, %r50;
cvt.u64.u32 %rd49, %r69;
cvt.u64.u32 %rd50, %r2;
cvt.s64.s32 %rd51, %r4;
add.s64 %rd52, %rd51, %rd50;
add.s64 %rd53, %rd52, %rd48;
add.s64 %rd54, %rd53, %rd49;
shl.b64 %rd55, %rd54, 2;
add.s64 %rd56, %rd47, %rd55;
ld.global.f32 %f4, [%rd56];

$L__BB103_8:
mul.lo.s32 %r70, %r11, %r55;
cvt.u64.u32 %rd57, %r70;
mul.lo.s32 %r71, %r1302, %r49;
cvt.u64.u32 %rd58, %r71;
add.s64 %rd4, %rd58, %rd57;
mul.hi.u32 %r72, %r11, %r51;
add.s32 %r73, %r72, %r11;
shr.u32 %r13, %r73, %r52;
shr.u32 %r74, %r3, 31;
add.s32 %r75, %r3, %r74;
shr.s32 %r1303, %r75, 1;
shr.u32 %r76, %r39, 5;
setp.ge.s32 %p12, %r1303, %r76;
mov.f32 %f28, %f27;
mov.f32 %f29, %f27;
mov.f32 %f30, %f27;
mov.f32 %f31, %f27;
mov.f32 %f32, %f27;
mov.f32 %f33, %f27;
mov.f32 %f34, %f27;
@%p12 bra $L__BB103_19;

mul.lo.s32 %r77, %r13, %r54;
mul.lo.s32 %r78, %r1301, %r48;
and.b32 %r81, %r75, -2;
sub.s32 %r82, %r3, %r81;
shl.b32 %r83, %r82, 1;
cvt.s64.s32 %rd5, %r83;
shl.b32 %r84, %r82, 3;
cvt.s64.s32 %rd59, %r84;
add.s32 %r85, %r1303, %r77;
add.s32 %r15, %r85, %r78;
add.s32 %r86, %r4, 1;
mad.lo.s32 %r87, %r43, %r86, %r15;
mul.wide.s32 %rd60, %r87, 17;
mad.lo.s32 %r88, %r4, %r43, %r1303;
add.s32 %r89, %r88, %r77;
add.s32 %r90, %r89, %r78;
mul.wide.s32 %rd6, %r90, 17;
add.s64 %rd61, %rd6, %rd59;
add.s64 %rd7, %rd61, 8;
add.s64 %rd62, %rd60, %rd59;
add.s64 %rd8, %rd62, 4;
add.s32 %r91, %r4, 3;
mad.lo.s32 %r92, %r43, %r91, %r15;
mul.wide.s32 %rd63, %r92, 17;
add.s64 %rd64, %rd63, %rd59;
add.s64 %rd9, %rd64, 8;
add.s32 %r93, %r4, 2;
mad.lo.s32 %r94, %r43, %r93, %r15;
mul.wide.s32 %rd65, %r94, 17;
add.s64 %rd66, %rd65, %rd59;
add.s64 %rd10, %rd66, 8;
cvta.to.global.u64 %rd96, %rd29;
cvta.to.global.u64 %rd95, %rd25;
cvta.to.global.u64 %rd13, %rd26;
mov.f32 %f27, 0f00000000;
shl.b64 %rd71, %rd5, 2;
mov.f32 %f28, %f27;
mov.f32 %f29, %f27;
mov.f32 %f30, %f27;
mov.f32 %f31, %f27;
mov.f32 %f32, %f27;
mov.f32 %f33, %f27;
mov.f32 %f34, %f27;

$L__BB103_10:
cvt.u64.u32 %rd67, %r1303;
add.s64 %rd68, %rd4, %rd67;
mul.lo.s64 %rd69, %rd68, 36;
add.s64 %rd70, %rd13, %rd69;
add.s64 %rd72, %rd70, %rd71;
add.s64 %rd73, %rd95, %rd7;
add.s64 %rd16, %rd73, -7;
ld.global.nc.u8 %rs2, [%rd73+-7];
cvt.u32.u16 %r112, %rs2;
and.b32 %r113, %r112, 255;
ld.global.nc.u8 %rs3, [%rd73+-6];
cvt.u32.u16 %r114, %rs3;
prmt.b32 %r115, %r114, %r113, 30212;
ld.global.nc.u8 %rs4, [%rd73+-5];
cvt.u32.u16 %r116, %rs4;
prmt.b32 %r117, %r116, %r115, 28756;
ld.global.nc.u8 %rs5, [%rd73+-4];
cvt.u32.u16 %r118, %rs5;
prmt.b32 %r119, %r118, %r117, 1620;
and.b32 %r120, %r119, -2004318072;
shr.u32 %r121, %r120, 1;
or.b32 %r122, %r121, 839922192;
and.b32 %r123, %r115, 30583;
mov.u32 %r124, 201852420;
mov.u32 %r125, 50462976;
prmt.b32 %r126, %r125, %r124, %r123;
mov.u32 %r127, -185009412;
mov.u32 %r128, -33620224;
prmt.b32 %r129, %r128, %r127, %r123;
and.b32 %r130, %r122, 30292;
prmt.b32 %r131, %r126, %r129, %r130;
shr.u32 %r132, %r119, 16;
and.b32 %r133, %r132, 30583;
prmt.b32 %r134, %r125, %r124, %r133;
prmt.b32 %r135, %r128, %r127, %r133;
shr.u32 %r136, %r122, 16;
prmt.b32 %r137, %r134, %r135, %r136;
mov.u32 %r138, 25632;
prmt.b32 %r96, %r131, %r137, %r138;
mov.u32 %r139, 30001;
prmt.b32 %r100, %r131, %r137, %r139;
ld.global.nc.u32 %r17, [%rd72+4];
mov.u32 %r98, 0;

dp4a.s32.s32 %r95, %r96, %r17, %r98;

ld.global.nc.u32 %r18, [%rd72+20];

dp4a.s32.s32 %r99, %r100, %r18, %r95;

ld.global.nc.u8 %rs6, [%rd73+-3];
cvt.u32.u16 %r140, %rs6;
and.b32 %r141, %r140, 255;
ld.global.nc.u8 %rs7, [%rd73+-2];
cvt.u32.u16 %r142, %rs7;
prmt.b32 %r143, %r142, %r141, 30212;
ld.global.nc.u8 %rs8, [%rd73+-1];
cvt.u32.u16 %r144, %rs8;
prmt.b32 %r145, %r144, %r143, 28756;
ld.global.nc.u8 %rs9, [%rd73];
cvt.u32.u16 %r146, %rs9;
prmt.b32 %r147, %r146, %r145, 1620;
and.b32 %r148, %r147, -2004318072;
shr.u32 %r149, %r148, 1;
or.b32 %r150, %r149, 839922192;
and.b32 %r151, %r143, 30583;
prmt.b32 %r152, %r125, %r124, %r151;
prmt.b32 %r153, %r128, %r127, %r151;
and.b32 %r154, %r150, 30292;
prmt.b32 %r155, %r152, %r153, %r154;
shr.u32 %r156, %r147, 16;
and.b32 %r157, %r156, 30583;
prmt.b32 %r158, %r125, %r124, %r157;
prmt.b32 %r159, %r128, %r127, %r157;
shr.u32 %r160, %r150, 16;
prmt.b32 %r161, %r158, %r159, %r160;
prmt.b32 %r104, %r155, %r161, %r138;
prmt.b32 %r108, %r155, %r161, %r139;
ld.global.nc.u32 %r19, [%rd72+8];

dp4a.s32.s32 %r103, %r104, %r19, %r99;

ld.global.nc.u32 %r20, [%rd72+24];

dp4a.s32.s32 %r107, %r108, %r20, %r103;

add.s64 %rd17, %rd95, %rd6;
ld.global.nc.u8 %rs10, [%rd17];
shl.b16 %rs12, %rs10, 7;
and.b16 %rs13, %rs12, 32640;
setp.eq.s16 %p13, %rs10, 255;
setp.eq.s16 %p14, %rs10, 0;
selp.b16 %rs14, 64, %rs13, %p14;
selp.b16 %rs1, 32767, %rs14, %p13;

{ mov.b32 %f101, {0,%rs1};}


mul.ftz.f32 %f103, %f101, 0f3F000000;
ld.global.nc.u32 %r21, [%rd70];

{.reg .f16 low,high;
mov.b32 {low,high},%r21;
cvt.f32.f16 %f102, low;}


mul.ftz.f32 %f104, %f103, %f102;
cvt.rn.f32.s32 %f105, %r107;
fma.rn.ftz.f32 %f30, %f104, %f105, %f30;
setp.eq.s64 %p15, %rd29, 0;
@%p15 bra $L__BB103_12;

add.s64 %rd74, %rd96, %rd7;
ld.global.u8 %r179, [%rd74+-7];
ld.global.u8 %r180, [%rd74+-6];
prmt.b32 %r181, %r180, %r179, 30212;
ld.global.u8 %r182, [%rd74+-5];
prmt.b32 %r183, %r182, %r181, 28756;
ld.global.u8 %r184, [%rd74+-4];
prmt.b32 %r185, %r184, %r183, 1620;
and.b32 %r186, %r185, -2004318072;
shr.u32 %r187, %r186, 1;
or.b32 %r188, %r187, 839922192;
and.b32 %r189, %r181, 30583;
prmt.b32 %r192, %r125, %r124, %r189;
prmt.b32 %r195, %r128, %r127, %r189;
and.b32 %r196, %r188, 30292;
prmt.b32 %r197, %r192, %r195, %r196;
shr.u32 %r198, %r185, 16;
and.b32 %r199, %r198, 30583;
prmt.b32 %r200, %r125, %r124, %r199;
prmt.b32 %r201, %r128, %r127, %r199;
shr.u32 %r202, %r188, 16;
prmt.b32 %r203, %r200, %r201, %r202;
prmt.b32 %r163, %r197, %r203, %r138;
prmt.b32 %r167, %r197, %r203, %r139;

dp4a.s32.s32 %r162, %r163, %r17, %r98;


dp4a.s32.s32 %r166, %r167, %r18, %r162;

ld.global.u8 %r206, [%rd74+-3];
ld.global.u8 %r207, [%rd74+-2];
prmt.b32 %r208, %r207, %r206, 30212;
ld.global.u8 %r209, [%rd74+-1];
prmt.b32 %r210, %r209, %r208, 28756;
ld.global.u8 %r211, [%rd74];
prmt.b32 %r212, %r211, %r210, 1620;
and.b32 %r213, %r212, -2004318072;
shr.u32 %r214, %r213, 1;
or.b32 %r215, %r214, 839922192;
and.b32 %r216, %r208, 30583;
prmt.b32 %r217, %r125, %r124, %r216;
prmt.b32 %r218, %r128, %r127, %r216;
and.b32 %r219, %r215, 30292;
prmt.b32 %r220, %r217, %r218, %r219;
shr.u32 %r221, %r212, 16;
and.b32 %r222, %r221, 30583;
prmt.b32 %r223, %r125, %r124, %r222;
prmt.b32 %r224, %r128, %r127, %r222;
shr.u32 %r225, %r215, 16;
prmt.b32 %r226, %r223, %r224, %r225;
prmt.b32 %r171, %r220, %r226, %r138;
prmt.b32 %r175, %r220, %r226, %r139;

dp4a.s32.s32 %r170, %r171, %r19, %r166;


dp4a.s32.s32 %r174, %r175, %r20, %r170;

add.s64 %rd75, %rd96, %rd6;
ld.global.u8 %rs16, [%rd75];
shl.b16 %rs17, %rs16, 7;
setp.eq.s16 %p16, %rs16, 255;
setp.eq.s16 %p17, %rs16, 0;
selp.b16 %rs18, 64, %rs17, %p17;
selp.b16 %rs15, 32767, %rs18, %p16;

{ mov.b32 %f106, {0,%rs15};}


mul.ftz.f32 %f108, %f106, 0f3F000000;

{.reg .f16 low,high;
mov.b32 {low,high},%r21;
cvt.f32.f16 %f107, low;}


mul.ftz.f32 %f109, %f108, %f107;
cvt.rn.f32.s32 %f110, %r174;
fma.rn.ftz.f32 %f34, %f109, %f110, %f34;

$L__BB103_12:
mul.wide.s32 %rd76, %r43, 17;
add.s64 %rd18, %rd16, %rd76;
ld.global.nc.u8 %rs20, [%rd18];
cvt.u32.u16 %r244, %rs20;
and.b32 %r245, %r244, 255;
ld.global.nc.u8 %rs21, [%rd18+1];
cvt.u32.u16 %r246, %rs21;
prmt.b32 %r247, %r246, %r245, 30212;
ld.global.nc.u8 %rs22, [%rd18+2];
cvt.u32.u16 %r248, %rs22;
prmt.b32 %r249, %r248, %r247, 28756;
ld.global.nc.u8 %rs23, [%rd18+3];
cvt.u32.u16 %r250, %rs23;
prmt.b32 %r251, %r250, %r249, 1620;
and.b32 %r252, %r251, -2004318072;
shr.u32 %r253, %r252, 1;
or.b32 %r254, %r253, 839922192;
and.b32 %r255, %r247, 30583;
prmt.b32 %r258, %r125, %r124, %r255;
prmt.b32 %r261, %r128, %r127, %r255;
and.b32 %r262, %r254, 30292;
prmt.b32 %r263, %r258, %r261, %r262;
shr.u32 %r264, %r251, 16;
and.b32 %r265, %r264, 30583;
prmt.b32 %r266, %r125, %r124, %r265;
prmt.b32 %r267, %r128, %r127, %r265;
shr.u32 %r268, %r254, 16;
prmt.b32 %r269, %r266, %r267, %r268;
prmt.b32 %r228, %r263, %r269, %r138;
prmt.b32 %r232, %r263, %r269, %r139;

dp4a.s32.s32 %r227, %r228, %r17, %r98;


dp4a.s32.s32 %r231, %r232, %r18, %r227;

ld.global.nc.u8 %rs24, [%rd18+4];
cvt.u32.u16 %r272, %rs24;
and.b32 %r273, %r272, 255;
ld.global.nc.u8 %rs25, [%rd18+5];
cvt.u32.u16 %r274, %rs25;
prmt.b32 %r275, %r274, %r273, 30212;
ld.global.nc.u8 %rs26, [%rd18+6];
cvt.u32.u16 %r276, %rs26;
prmt.b32 %r277, %r276, %r275, 28756;
ld.global.nc.u8 %rs27, [%rd18+7];
cvt.u32.u16 %r278, %rs27;
prmt.b32 %r279, %r278, %r277, 1620;
and.b32 %r280, %r279, -2004318072;
shr.u32 %r281, %r280, 1;
or.b32 %r282, %r281, 839922192;
and.b32 %r283, %r275, 30583;
prmt.b32 %r284, %r125, %r124, %r283;
prmt.b32 %r285, %r128, %r127, %r283;
and.b32 %r286, %r282, 30292;
prmt.b32 %r287, %r284, %r285, %r286;
shr.u32 %r288, %r279, 16;
and.b32 %r289, %r288, 30583;
prmt.b32 %r290, %r125, %r124, %r289;
prmt.b32 %r291, %r128, %r127, %r289;
shr.u32 %r292, %r282, 16;
prmt.b32 %r293, %r290, %r291, %r292;
prmt.b32 %r236, %r287, %r293, %r138;
prmt.b32 %r240, %r287, %r293, %r139;

dp4a.s32.s32 %r235, %r236, %r19, %r231;


dp4a.s32.s32 %r239, %r240, %r20, %r235;

add.s64 %rd19, %rd17, %rd76;
ld.global.nc.u8 %rs28, [%rd19];
shl.b16 %rs30, %rs28, 7;
and.b16 %rs31, %rs30, 32640;
setp.eq.s16 %p18, %rs28, 255;
setp.eq.s16 %p19, %rs28, 0;
selp.b16 %rs32, 64, %rs31, %p19;
selp.b16 %rs19, 32767, %rs32, %p18;

{ mov.b32 %f111, {0,%rs19};}


mul.ftz.f32 %f113, %f111, 0f3F000000;

{.reg .f16 low,high;
mov.b32 {low,high},%r21;
cvt.f32.f16 %f112, low;}


mul.ftz.f32 %f114, %f113, %f112;
cvt.rn.f32.s32 %f115, %r239;
fma.rn.ftz.f32 %f29, %f114, %f115, %f29;
@%p15 bra $L__BB103_14;

add.s64 %rd77, %rd96, %rd8;
ld.global.u8 %r311, [%rd77+-3];
ld.global.u8 %r312, [%rd77+-2];
prmt.b32 %r313, %r312, %r311, 30212;
ld.global.u8 %r314, [%rd77+-1];
prmt.b32 %r315, %r314, %r313, 28756;
ld.global.u8 %r316, [%rd77];
prmt.b32 %r317, %r316, %r315, 1620;
and.b32 %r318, %r317, -2004318072;
shr.u32 %r319, %r318, 1;
or.b32 %r320, %r319, 839922192;
and.b32 %r321, %r313, 30583;
mov.u32 %r322, 201852420;
mov.u32 %r323, 50462976;
prmt.b32 %r324, %r323, %r322, %r321;
mov.u32 %r325, -185009412;
mov.u32 %r326, -33620224;
prmt.b32 %r327, %r326, %r325, %r321;
and.b32 %r328, %r320, 30292;
prmt.b32 %r329, %r324, %r327, %r328;
shr.u32 %r330, %r317, 16;
and.b32 %r331, %r330, 30583;
prmt.b32 %r332, %r323, %r322, %r331;
prmt.b32 %r333, %r326, %r325, %r331;
shr.u32 %r334, %r320, 16;
prmt.b32 %r335, %r332, %r333, %r334;
mov.u32 %r336, 25632;
prmt.b32 %r295, %r329, %r335, %r336;
mov.u32 %r337, 30001;
prmt.b32 %r299, %r329, %r335, %r337;
mov.u32 %r297, 0;

dp4a.s32.s32 %r294, %r295, %r17, %r297;


dp4a.s32.s32 %r298, %r299, %r18, %r294;

ld.global.u8 %r338, [%rd77+1];
ld.global.u8 %r339, [%rd77+2];
prmt.b32 %r340, %r339, %r338, 30212;
ld.global.u8 %r341, [%rd77+3];
prmt.b32 %r342, %r341, %r340, 28756;
ld.global.u8 %r343, [%rd77+4];
prmt.b32 %r344, %r343, %r342, 1620;
and.b32 %r345, %r344, -2004318072;
shr.u32 %r346, %r345, 1;
or.b32 %r347, %r346, 839922192;
and.b32 %r348, %r340, 30583;
prmt.b32 %r349, %r323, %r322, %r348;
prmt.b32 %r350, %r326, %r325, %r348;
and.b32 %r351, %r347, 30292;
prmt.b32 %r352, %r349, %r350, %r351;
shr.u32 %r353, %r344, 16;
and.b32 %r354, %r353, 30583;
prmt.b32 %r355, %r323, %r322, %r354;
prmt.b32 %r356, %r326, %r325, %r354;
shr.u32 %r357, %r347, 16;
prmt.b32 %r358, %r355, %r356, %r357;
prmt.b32 %r303, %r352, %r358, %r336;
prmt.b32 %r307, %r352, %r358, %r337;

dp4a.s32.s32 %r302, %r303, %r19, %r298;


dp4a.s32.s32 %r306, %r307, %r20, %r302;

or.b32 %r361, %r4, 1;
mad.lo.s32 %r362, %r43, %r361, %r15;
mul.wide.s32 %rd78, %r362, 17;
add.s64 %rd79, %rd96, %rd78;
ld.global.u8 %rs34, [%rd79];
shl.b16 %rs35, %rs34, 7;
setp.eq.s16 %p21, %rs34, 255;
setp.eq.s16 %p22, %rs34, 0;
selp.b16 %rs36, 64, %rs35, %p22;
selp.b16 %rs33, 32767, %rs36, %p21;

{ mov.b32 %f116, {0,%rs33};}


mul.ftz.f32 %f118, %f116, 0f3F000000;

{.reg .f16 low,high;
mov.b32 {low,high},%r21;
cvt.f32.f16 %f117, low;}


mul.ftz.f32 %f119, %f118, %f117;
cvt.rn.f32.s32 %f120, %r306;
fma.rn.ftz.f32 %f33, %f119, %f120, %f33;

$L__BB103_14:
mul.wide.s32 %rd94, %r43, 17;
add.s64 %rd20, %rd18, %rd94;
ld.global.nc.u8 %rs38, [%rd20];
cvt.u32.u16 %r380, %rs38;
and.b32 %r381, %r380, 255;
ld.global.nc.u8 %rs39, [%rd20+1];
cvt.u32.u16 %r382, %rs39;
prmt.b32 %r383, %r382, %r381, 30212;
ld.global.nc.u8 %rs40, [%rd20+2];
cvt.u32.u16 %r384, %rs40;
prmt.b32 %r385, %r384, %r383, 28756;
ld.global.nc.u8 %rs41, [%rd20+3];
cvt.u32.u16 %r386, %rs41;
prmt.b32 %r387, %r386, %r385, 1620;
and.b32 %r388, %r387, -2004318072;
shr.u32 %r389, %r388, 1;
or.b32 %r390, %r389, 839922192;
and.b32 %r391, %r383, 30583;
mov.u32 %r392, 201852420;
mov.u32 %r393, 50462976;
prmt.b32 %r394, %r393, %r392, %r391;
mov.u32 %r395, -185009412;
mov.u32 %r396, -33620224;
prmt.b32 %r397, %r396, %r395, %r391;
and.b32 %r398, %r390, 30292;
prmt.b32 %r399, %r394, %r397, %r398;
shr.u32 %r400, %r387, 16;
and.b32 %r401, %r400, 30583;
prmt.b32 %r402, %r393, %r392, %r401;
prmt.b32 %r403, %r396, %r395, %r401;
shr.u32 %r404, %r390, 16;
prmt.b32 %r405, %r402, %r403, %r404;
mov.u32 %r406, 25632;
prmt.b32 %r364, %r399, %r405, %r406;
mov.u32 %r407, 30001;
prmt.b32 %r368, %r399, %r405, %r407;
mov.u32 %r366, 0;

dp4a.s32.s32 %r363, %r364, %r17, %r366;


dp4a.s32.s32 %r367, %r368, %r18, %r363;

ld.global.nc.u8 %rs42, [%rd20+4];
cvt.u32.u16 %r408, %rs42;
and.b32 %r409, %r408, 255;
ld.global.nc.u8 %rs43, [%rd20+5];
cvt.u32.u16 %r410, %rs43;
prmt.b32 %r411, %r410, %r409, 30212;
ld.global.nc.u8 %rs44, [%rd20+6];
cvt.u32.u16 %r412, %rs44;
prmt.b32 %r413, %r412, %r411, 28756;
ld.global.nc.u8 %rs45, [%rd20+7];
cvt.u32.u16 %r414, %rs45;
prmt.b32 %r415, %r414, %r413, 1620;
and.b32 %r416, %r415, -2004318072;
shr.u32 %r417, %r416, 1;
or.b32 %r418, %r417, 839922192;
and.b32 %r419, %r411, 30583;
prmt.b32 %r420, %r393, %r392, %r419;
prmt.b32 %r421, %r396, %r395, %r419;
and.b32 %r422, %r418, 30292;
prmt.b32 %r423, %r420, %r421, %r422;
shr.u32 %r424, %r415, 16;
and.b32 %r425, %r424, 30583;
prmt.b32 %r426, %r393, %r392, %r425;
prmt.b32 %r427, %r396, %r395, %r425;
shr.u32 %r428, %r418, 16;
prmt.b32 %r429, %r426, %r427, %r428;
prmt.b32 %r372, %r423, %r429, %r406;
prmt.b32 %r376, %r423, %r429, %r407;

dp4a.s32.s32 %r371, %r372, %r19, %r367;


dp4a.s32.s32 %r375, %r376, %r20, %r371;

add.s64 %rd21, %rd19, %rd94;
ld.global.nc.u8 %rs46, [%rd21];
shl.b16 %rs48, %rs46, 7;
and.b16 %rs49, %rs48, 32640;
setp.eq.s16 %p23, %rs46, 255;
setp.eq.s16 %p24, %rs46, 0;
selp.b16 %rs50, 64, %rs49, %p24;
selp.b16 %rs37, 32767, %rs50, %p23;

{ mov.b32 %f121, {0,%rs37};}


mul.ftz.f32 %f123, %f121, 0f3F000000;

{.reg .f16 low,high;
mov.b32 {low,high},%r21;
cvt.f32.f16 %f122, low;}


mul.ftz.f32 %f124, %f123, %f122;
cvt.rn.f32.s32 %f125, %r375;
fma.rn.ftz.f32 %f28, %f124, %f125, %f28;
@%p15 bra $L__BB103_16;

add.s64 %rd81, %rd96, %rd10;
ld.global.u8 %r447, [%rd81+-7];
ld.global.u8 %r448, [%rd81+-6];
prmt.b32 %r449, %r448, %r447, 30212;
ld.global.u8 %r450, [%rd81+-5];
prmt.b32 %r451, %r450, %r449, 28756;
ld.global.u8 %r452, [%rd81+-4];
prmt.b32 %r453, %r452, %r451, 1620;
and.b32 %r454, %r453, -2004318072;
shr.u32 %r455, %r454, 1;
or.b32 %r456, %r455, 839922192;
and.b32 %r457, %r449, 30583;
prmt.b32 %r460, %r393, %r392, %r457;
prmt.b32 %r463, %r396, %r395, %r457;
and.b32 %r464, %r456, 30292;
prmt.b32 %r465, %r460, %r463, %r464;
shr.u32 %r466, %r453, 16;
and.b32 %r467, %r466, 30583;
prmt.b32 %r468, %r393, %r392, %r467;
prmt.b32 %r469, %r396, %r395, %r467;
shr.u32 %r470, %r456, 16;
prmt.b32 %r471, %r468, %r469, %r470;
prmt.b32 %r431, %r465, %r471, %r406;
prmt.b32 %r435, %r465, %r471, %r407;

dp4a.s32.s32 %r430, %r431, %r17, %r366;


dp4a.s32.s32 %r434, %r435, %r18, %r430;

ld.global.u8 %r474, [%rd81+-3];
ld.global.u8 %r475, [%rd81+-2];
prmt.b32 %r476, %r475, %r474, 30212;
ld.global.u8 %r477, [%rd81+-1];
prmt.b32 %r478, %r477, %r476, 28756;
ld.global.u8 %r479, [%rd81];
prmt.b32 %r480, %r479, %r478, 1620;
and.b32 %r481, %r480, -2004318072;
shr.u32 %r482, %r481, 1;
or.b32 %r483, %r482, 839922192;
and.b32 %r484, %r476, 30583;
prmt.b32 %r485, %r393, %r392, %r484;
prmt.b32 %r486, %r396, %r395, %r484;
and.b32 %r487, %r483, 30292;
prmt.b32 %r488, %r485, %r486, %r487;
shr.u32 %r489, %r480, 16;
and.b32 %r490, %r489, 30583;
prmt.b32 %r491, %r393, %r392, %r490;
prmt.b32 %r492, %r396, %r395, %r490;
shr.u32 %r493, %r483, 16;
prmt.b32 %r494, %r491, %r492, %r493;
prmt.b32 %r439, %r488, %r494, %r406;
prmt.b32 %r443, %r488, %r494, %r407;

dp4a.s32.s32 %r438, %r439, %r19, %r434;


dp4a.s32.s32 %r442, %r443, %r20, %r438;

or.b32 %r497, %r4, 2;
mad.lo.s32 %r498, %r43, %r497, %r15;
mul.wide.s32 %rd82, %r498, 17;
add.s64 %rd83, %rd96, %rd82;
ld.global.u8 %rs52, [%rd83];
shl.b16 %rs53, %rs52, 7;
setp.eq.s16 %p26, %rs52, 255;
setp.eq.s16 %p27, %rs52, 0;
selp.b16 %rs54, 64, %rs53, %p27;
selp.b16 %rs51, 32767, %rs54, %p26;

{ mov.b32 %f126, {0,%rs51};}


mul.ftz.f32 %f128, %f126, 0f3F000000;

{.reg .f16 low,high;
mov.b32 {low,high},%r21;
cvt.f32.f16 %f127, low;}


mul.ftz.f32 %f129, %f128, %f127;
cvt.rn.f32.s32 %f130, %r442;
fma.rn.ftz.f32 %f32, %f129, %f130, %f32;

$L__BB103_16:
mul.wide.s32 %rd92, %r43, 17;
add.s64 %rd85, %rd20, %rd92;
ld.global.nc.u8 %rs56, [%rd85];
cvt.u32.u16 %r516, %rs56;
and.b32 %r517, %r516, 255;
ld.global.nc.u8 %rs57, [%rd85+1];
cvt.u32.u16 %r518, %rs57;
prmt.b32 %r519, %r518, %r517, 30212;
ld.global.nc.u8 %rs58, [%rd85+2];
cvt.u32.u16 %r520, %rs58;
prmt.b32 %r521, %r520, %r519, 28756;
ld.global.nc.u8 %rs59, [%rd85+3];
cvt.u32.u16 %r522, %rs59;
prmt.b32 %r523, %r522, %r521, 1620;
and.b32 %r524, %r523, -2004318072;
shr.u32 %r525, %r524, 1;
or.b32 %r526, %r525, 839922192;
and.b32 %r527, %r519, 30583;
prmt.b32 %r530, %r393, %r392, %r527;
prmt.b32 %r533, %r396, %r395, %r527;
and.b32 %r534, %r526, 30292;
prmt.b32 %r535, %r530, %r533, %r534;
shr.u32 %r536, %r523, 16;
and.b32 %r537, %r536, 30583;
prmt.b32 %r538, %r393, %r392, %r537;
prmt.b32 %r539, %r396, %r395, %r537;
shr.u32 %r540, %r526, 16;
prmt.b32 %r541, %r538, %r539, %r540;
prmt.b32 %r500, %r535, %r541, %r406;
prmt.b32 %r504, %r535, %r541, %r407;

dp4a.s32.s32 %r499, %r500, %r17, %r366;


dp4a.s32.s32 %r503, %r504, %r18, %r499;

ld.global.nc.u8 %rs60, [%rd85+4];
cvt.u32.u16 %r544, %rs60;
and.b32 %r545, %r544, 255;
ld.global.nc.u8 %rs61, [%rd85+5];
cvt.u32.u16 %r546, %rs61;
prmt.b32 %r547, %r546, %r545, 30212;
ld.global.nc.u8 %rs62, [%rd85+6];
cvt.u32.u16 %r548, %rs62;
prmt.b32 %r549, %r548, %r547, 28756;
ld.global.nc.u8 %rs63, [%rd85+7];
cvt.u32.u16 %r550, %rs63;
prmt.b32 %r551, %r550, %r549, 1620;
and.b32 %r552, %r551, -2004318072;
shr.u32 %r553, %r552, 1;
or.b32 %r554, %r553, 839922192;
and.b32 %r555, %r547, 30583;
prmt.b32 %r556, %r393, %r392, %r555;
prmt.b32 %r557, %r396, %r395, %r555;
and.b32 %r558, %r554, 30292;
prmt.b32 %r559, %r556, %r557, %r558;
shr.u32 %r560, %r551, 16;
and.b32 %r561, %r560, 30583;
prmt.b32 %r562, %r393, %r392, %r561;
prmt.b32 %r563, %r396, %r395, %r561;
shr.u32 %r564, %r554, 16;
prmt.b32 %r565, %r562, %r563, %r564;
prmt.b32 %r508, %r559, %r565, %r406;
prmt.b32 %r512, %r559, %r565, %r407;

dp4a.s32.s32 %r507, %r508, %r19, %r503;


dp4a.s32.s32 %r511, %r512, %r20, %r507;

add.s64 %rd86, %rd21, %rd92;
ld.global.nc.u8 %rs64, [%rd86];
shl.b16 %rs66, %rs64, 7;
and.b16 %rs67, %rs66, 32640;
setp.eq.s16 %p28, %rs64, 255;
setp.eq.s16 %p29, %rs64, 0;
selp.b16 %rs68, 64, %rs67, %p29;
selp.b16 %rs55, 32767, %rs68, %p28;

{ mov.b32 %f131, {0,%rs55};}


mul.ftz.f32 %f133, %f131, 0f3F000000;

{.reg .f16 low,high;
mov.b32 {low,high},%r21;
cvt.f32.f16 %f132, low;}


mul.ftz.f32 %f134, %f133, %f132;
cvt.rn.f32.s32 %f135, %r511;
fma.rn.ftz.f32 %f27, %f134, %f135, %f27;
@%p15 bra $L__BB103_18;

add.s64 %rd87, %rd96, %rd9;
ld.global.u8 %r583, [%rd87+-7];
ld.global.u8 %r584, [%rd87+-6];
prmt.b32 %r585, %r584, %r583, 30212;
ld.global.u8 %r586, [%rd87+-5];
prmt.b32 %r587, %r586, %r585, 28756;
ld.global.u8 %r588, [%rd87+-4];
prmt.b32 %r589, %r588, %r587, 1620;
and.b32 %r590, %r589, -2004318072;
shr.u32 %r591, %r590, 1;
or.b32 %r592, %r591, 839922192;
and.b32 %r593, %r585, 30583;
mov.u32 %r594, 201852420;
mov.u32 %r595, 50462976;
prmt.b32 %r596, %r595, %r594, %r593;
mov.u32 %r597, -185009412;
mov.u32 %r598, -33620224;
prmt.b32 %r599, %r598, %r597, %r593;
and.b32 %r600, %r592, 30292;
prmt.b32 %r601, %r596, %r599, %r600;
shr.u32 %r602, %r589, 16;
and.b32 %r603, %r602, 30583;
prmt.b32 %r604, %r595, %r594, %r603;
prmt.b32 %r605, %r598, %r597, %r603;
shr.u32 %r606, %r592, 16;
prmt.b32 %r607, %r604, %r605, %r606;
mov.u32 %r608, 25632;
prmt.b32 %r567, %r601, %r607, %r608;
mov.u32 %r609, 30001;
prmt.b32 %r571, %r601, %r607, %r609;
mov.u32 %r569, 0;

dp4a.s32.s32 %r566, %r567, %r17, %r569;


dp4a.s32.s32 %r570, %r571, %r18, %r566;

ld.global.u8 %r610, [%rd87+-3];
ld.global.u8 %r611, [%rd87+-2];
prmt.b32 %r612, %r611, %r610, 30212;
ld.global.u8 %r613, [%rd87+-1];
prmt.b32 %r614, %r613, %r612, 28756;
ld.global.u8 %r615, [%rd87];
prmt.b32 %r616, %r615, %r614, 1620;
and.b32 %r617, %r616, -2004318072;
shr.u32 %r618, %r617, 1;
or.b32 %r619, %r618, 839922192;
and.b32 %r620, %r612, 30583;
prmt.b32 %r621, %r595, %r594, %r620;
prmt.b32 %r622, %r598, %r597, %r620;
and.b32 %r623, %r619, 30292;
prmt.b32 %r624, %r621, %r622, %r623;
shr.u32 %r625, %r616, 16;
and.b32 %r626, %r625, 30583;
prmt.b32 %r627, %r595, %r594, %r626;
prmt.b32 %r628, %r598, %r597, %r626;
shr.u32 %r629, %r619, 16;
prmt.b32 %r630, %r627, %r628, %r629;
prmt.b32 %r575, %r624, %r630, %r608;
prmt.b32 %r579, %r624, %r630, %r609;

dp4a.s32.s32 %r574, %r575, %r19, %r570;


dp4a.s32.s32 %r578, %r579, %r20, %r574;

or.b32 %r633, %r4, 3;
mad.lo.s32 %r634, %r43, %r633, %r15;
mul.wide.s32 %rd88, %r634, 17;
add.s64 %rd89, %rd96, %rd88;
ld.global.u8 %rs70, [%rd89];
shl.b16 %rs71, %rs70, 7;
setp.eq.s16 %p31, %rs70, 255;
setp.eq.s16 %p32, %rs70, 0;
selp.b16 %rs72, 64, %rs71, %p32;
selp.b16 %rs69, 32767, %rs72, %p31;

{ mov.b32 %f136, {0,%rs69};}


mul.ftz.f32 %f138, %f136, 0f3F000000;

{.reg .f16 low,high;
mov.b32 {low,high},%r21;
cvt.f32.f16 %f137, low;}


mul.ftz.f32 %f139, %f138, %f137;
cvt.rn.f32.s32 %f140, %r578;
fma.rn.ftz.f32 %f31, %f139, %f140, %f31;

$L__BB103_18:
add.s64 %rd96, %rd96, 1088;
add.s64 %rd95, %rd95, 1088;
add.s32 %r1303, %r1303, 64;
setp.lt.s32 %p33, %r1303, %r76;
@%p33 bra $L__BB103_10;

$L__BB103_19:
mov.u32 %r1290, %tid.y;
setp.eq.s32 %p34, %r1290, 0;
@%p34 bra $L__BB103_23;

mov.u32 %r1291, %tid.y;
setp.ne.s64 %p35, %rd29, 0;
add.s32 %r23, %r1291, -1;
shl.b32 %r638, %r1291, 9;
mov.u32 %r639, _ZZ13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r640, %r639, %r638;
shl.b32 %r641, %r2, 2;
add.s32 %r642, %r640, %r641;
st.shared.f32 [%r642+-512], %f30;
@%p35 bra $L__BB103_22;
bra.uni $L__BB103_21;

$L__BB103_22:
shl.b32 %r643, %r23, 9;
mov.u32 %r644, _ZZ13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r645, %r644, %r643;
add.s32 %r647, %r645, %r641;
st.shared.f32 [%r647], %f34;
add.s32 %r1287, %r642, -512;
st.shared.f32 [%r1287+128], %f29;
st.shared.f32 [%r647+128], %f33;
add.s32 %r1288, %r642, -512;
st.shared.f32 [%r1288+256], %f28;
st.shared.f32 [%r647+256], %f32;
add.s32 %r1289, %r642, -512;
st.shared.f32 [%r1289+384], %f27;
st.shared.f32 [%r647+384], %f31;
bra.uni $L__BB103_23;

$L__BB103_21:
add.s32 %r1284, %r642, -512;
st.shared.f32 [%r1284+128], %f29;
add.s32 %r1285, %r642, -512;
st.shared.f32 [%r1285+256], %f28;
add.s32 %r1286, %r642, -512;
st.shared.f32 [%r1286+384], %f27;

$L__BB103_23:
mov.u32 %r1292, %tid.y;
setp.ne.s32 %p335, %r1292, 0;
bar.sync 0;
@%p335 bra $L__BB103_78;
bra.uni $L__BB103_24;

$L__BB103_78:
ret;

$L__BB103_24:
ld.param.u64 %rd93, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.f32 %f985, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
ld.param.u32 %r1295, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r1294, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
mov.u32 %r1293, %ctaid.z;
setp.ne.s64 %p37, %rd29, 0;
mov.u32 %r650, %ctaid.y;
mad.lo.s32 %r651, %r650, %r1295, %r4;
mad.lo.s32 %r652, %r1293, %r1294, %r651;
neg.ftz.f32 %f35, %f985;
shl.b32 %r653, %r2, 2;
mov.u32 %r654, _ZZ13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r25, %r654, %r653;
cvta.to.global.u64 %rd90, %rd93;
mul.wide.u32 %rd91, %r652, 4;
add.s64 %rd24, %rd90, %rd91;
@%p37 bra $L__BB103_33;
bra.uni $L__BB103_25;

$L__BB103_33:
ld.param.u32 %r1296, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
mov.u32 %r728, _ZZ13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r27, %r728, %r653;
setp.gt.s32 %p70, %r1296, 2;
@%p70 bra $L__BB103_44;

ld.param.u32 %r1299, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p73, %r1299, 1;
@%p73 bra $L__BB103_62;

ld.param.u32 %r1300, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p74, %r1300, 2;
@%p74 bra $L__BB103_36;
bra.uni $L__BB103_70;

$L__BB103_36:
ld.shared.f32 %f541, [%r25];
add.ftz.f32 %f542, %f541, %f30;
ld.shared.f32 %f543, [%r27];
add.ftz.f32 %f544, %f543, %f34;
ld.shared.f32 %f545, [%r25+512];
add.ftz.f32 %f546, %f545, %f542;
ld.shared.f32 %f547, [%r27+512];
add.ftz.f32 %f548, %f547, %f544;
ld.shared.f32 %f549, [%r25+1024];
add.ftz.f32 %f550, %f549, %f546;
ld.shared.f32 %f551, [%r27+1024];
add.ftz.f32 %f552, %f551, %f548;
mov.b32 %r951, %f550;
mov.u32 %r952, 31;
mov.u32 %r953, 16;
mov.u32 %r954, -1;
shfl.sync.bfly.b32 %r955|%p179, %r951, %r953, %r952, %r954;
mov.b32 %f553, %r955;
add.ftz.f32 %f554, %f550, %f553;
mov.b32 %r956, %f554;
mov.u32 %r957, 8;
shfl.sync.bfly.b32 %r958|%p180, %r956, %r957, %r952, %r954;
mov.b32 %f555, %r958;
add.ftz.f32 %f556, %f554, %f555;
mov.b32 %r959, %f556;
mov.u32 %r960, 4;
shfl.sync.bfly.b32 %r961|%p181, %r959, %r960, %r952, %r954;
mov.b32 %f557, %r961;
add.ftz.f32 %f558, %f556, %f557;
mov.b32 %r962, %f558;
mov.u32 %r963, 2;
shfl.sync.bfly.b32 %r964|%p182, %r962, %r963, %r952, %r954;
mov.b32 %f559, %r964;
add.ftz.f32 %f560, %f558, %f559;
mov.b32 %r965, %f560;
mov.u32 %r966, 1;
shfl.sync.bfly.b32 %r967|%p183, %r965, %r966, %r952, %r954;
mov.b32 %f561, %r967;
add.ftz.f32 %f56, %f560, %f561;
mov.b32 %r968, %f552;
shfl.sync.bfly.b32 %r969|%p184, %r968, %r953, %r952, %r954;
mov.b32 %f562, %r969;
add.ftz.f32 %f563, %f552, %f562;
mov.b32 %r970, %f563;
shfl.sync.bfly.b32 %r971|%p185, %r970, %r957, %r952, %r954;
mov.b32 %f564, %r971;
add.ftz.f32 %f565, %f563, %f564;
mov.b32 %r972, %f565;
shfl.sync.bfly.b32 %r973|%p186, %r972, %r960, %r952, %r954;
mov.b32 %f566, %r973;
add.ftz.f32 %f567, %f565, %f566;
mov.b32 %r974, %f567;
shfl.sync.bfly.b32 %r975|%p187, %r974, %r963, %r952, %r954;
mov.b32 %f568, %r975;
add.ftz.f32 %f569, %f567, %f568;
mov.b32 %r976, %f569;
shfl.sync.bfly.b32 %r977|%p188, %r976, %r966, %r952, %r954;
mov.b32 %f570, %r977;
add.ftz.f32 %f57, %f569, %f570;
setp.ge.u32 %p189, %r4, %r44;
setp.ne.s32 %p190, %r2, 0;
or.pred %p191, %p190, %p189;
@%p191 bra $L__BB103_38;

add.ftz.f32 %f571, %f5, %f56;
add.ftz.f32 %f572, %f4, %f57;
mul.ftz.f32 %f573, %f572, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f574, %f573;
add.ftz.f32 %f575, %f574, 0f3F800000;
div.approx.ftz.f32 %f576, %f572, %f575;
mul.ftz.f32 %f577, %f571, %f576;
st.global.f32 [%rd24], %f577;

$L__BB103_38:
ld.shared.f32 %f578, [%r25+128];
add.ftz.f32 %f579, %f578, %f29;
ld.shared.f32 %f580, [%r27+128];
add.ftz.f32 %f581, %f580, %f33;
ld.shared.f32 %f582, [%r25+640];
add.ftz.f32 %f583, %f582, %f579;
ld.shared.f32 %f584, [%r27+640];
add.ftz.f32 %f585, %f584, %f581;
ld.shared.f32 %f586, [%r25+1152];
add.ftz.f32 %f587, %f586, %f583;
ld.shared.f32 %f588, [%r27+1152];
add.ftz.f32 %f589, %f588, %f585;
mov.b32 %r978, %f587;
shfl.sync.bfly.b32 %r982|%p192, %r978, %r953, %r952, %r954;
mov.b32 %f590, %r982;
add.ftz.f32 %f591, %f587, %f590;
mov.b32 %r983, %f591;
shfl.sync.bfly.b32 %r985|%p193, %r983, %r957, %r952, %r954;
mov.b32 %f592, %r985;
add.ftz.f32 %f593, %f591, %f592;
mov.b32 %r986, %f593;
shfl.sync.bfly.b32 %r988|%p194, %r986, %r960, %r952, %r954;
mov.b32 %f594, %r988;
add.ftz.f32 %f595, %f593, %f594;
mov.b32 %r989, %f595;
shfl.sync.bfly.b32 %r991|%p195, %r989, %r963, %r952, %r954;
mov.b32 %f596, %r991;
add.ftz.f32 %f597, %f595, %f596;
mov.b32 %r992, %f597;
shfl.sync.bfly.b32 %r994|%p196, %r992, %r966, %r952, %r954;
mov.b32 %f598, %r994;
add.ftz.f32 %f58, %f597, %f598;
mov.b32 %r995, %f589;
shfl.sync.bfly.b32 %r996|%p197, %r995, %r953, %r952, %r954;
mov.b32 %f599, %r996;
add.ftz.f32 %f600, %f589, %f599;
mov.b32 %r997, %f600;
shfl.sync.bfly.b32 %r998|%p198, %r997, %r957, %r952, %r954;
mov.b32 %f601, %r998;
add.ftz.f32 %f602, %f600, %f601;
mov.b32 %r999, %f602;
shfl.sync.bfly.b32 %r1000|%p199, %r999, %r960, %r952, %r954;
mov.b32 %f603, %r1000;
add.ftz.f32 %f604, %f602, %f603;
mov.b32 %r1001, %f604;
shfl.sync.bfly.b32 %r1002|%p200, %r1001, %r963, %r952, %r954;
mov.b32 %f605, %r1002;
add.ftz.f32 %f606, %f604, %f605;
mov.b32 %r1003, %f606;
shfl.sync.bfly.b32 %r1004|%p201, %r1003, %r966, %r952, %r954;
mov.b32 %f607, %r1004;
add.ftz.f32 %f59, %f606, %f607;
add.s32 %r1005, %r4, 1;
setp.ge.u32 %p202, %r1005, %r44;
setp.ne.s32 %p203, %r2, 1;
or.pred %p204, %p203, %p202;
@%p204 bra $L__BB103_40;

add.ftz.f32 %f608, %f5, %f58;
add.ftz.f32 %f609, %f4, %f59;
mul.ftz.f32 %f610, %f609, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f611, %f610;
add.ftz.f32 %f612, %f611, 0f3F800000;
div.approx.ftz.f32 %f613, %f609, %f612;
mul.ftz.f32 %f614, %f608, %f613;
st.global.f32 [%rd24+4], %f614;

$L__BB103_40:
ld.shared.f32 %f615, [%r25+256];
add.ftz.f32 %f616, %f615, %f28;
ld.shared.f32 %f617, [%r27+256];
add.ftz.f32 %f618, %f617, %f32;
ld.shared.f32 %f619, [%r25+768];
add.ftz.f32 %f620, %f619, %f616;
ld.shared.f32 %f621, [%r27+768];
add.ftz.f32 %f622, %f621, %f618;
ld.shared.f32 %f623, [%r25+1280];
add.ftz.f32 %f624, %f623, %f620;
ld.shared.f32 %f625, [%r27+1280];
add.ftz.f32 %f626, %f625, %f622;
mov.b32 %r1006, %f624;
mov.u32 %r1007, 31;
mov.u32 %r1008, 16;
mov.u32 %r1009, -1;
shfl.sync.bfly.b32 %r1010|%p205, %r1006, %r1008, %r1007, %r1009;
mov.b32 %f627, %r1010;
add.ftz.f32 %f628, %f624, %f627;
mov.b32 %r1011, %f628;
mov.u32 %r1012, 8;
shfl.sync.bfly.b32 %r1013|%p206, %r1011, %r1012, %r1007, %r1009;
mov.b32 %f629, %r1013;
add.ftz.f32 %f630, %f628, %f629;
mov.b32 %r1014, %f630;
mov.u32 %r1015, 4;
shfl.sync.bfly.b32 %r1016|%p207, %r1014, %r1015, %r1007, %r1009;
mov.b32 %f631, %r1016;
add.ftz.f32 %f632, %f630, %f631;
mov.b32 %r1017, %f632;
mov.u32 %r1018, 2;
shfl.sync.bfly.b32 %r1019|%p208, %r1017, %r1018, %r1007, %r1009;
mov.b32 %f633, %r1019;
add.ftz.f32 %f634, %f632, %f633;
mov.b32 %r1020, %f634;
mov.u32 %r1021, 1;
shfl.sync.bfly.b32 %r1022|%p209, %r1020, %r1021, %r1007, %r1009;
mov.b32 %f635, %r1022;
add.ftz.f32 %f60, %f634, %f635;
mov.b32 %r1023, %f626;
shfl.sync.bfly.b32 %r1024|%p210, %r1023, %r1008, %r1007, %r1009;
mov.b32 %f636, %r1024;
add.ftz.f32 %f637, %f626, %f636;
mov.b32 %r1025, %f637;
shfl.sync.bfly.b32 %r1026|%p211, %r1025, %r1012, %r1007, %r1009;
mov.b32 %f638, %r1026;
add.ftz.f32 %f639, %f637, %f638;
mov.b32 %r1027, %f639;
shfl.sync.bfly.b32 %r1028|%p212, %r1027, %r1015, %r1007, %r1009;
mov.b32 %f640, %r1028;
add.ftz.f32 %f641, %f639, %f640;
mov.b32 %r1029, %f641;
shfl.sync.bfly.b32 %r1030|%p213, %r1029, %r1018, %r1007, %r1009;
mov.b32 %f642, %r1030;
add.ftz.f32 %f643, %f641, %f642;
mov.b32 %r1031, %f643;
shfl.sync.bfly.b32 %r1032|%p214, %r1031, %r1021, %r1007, %r1009;
mov.b32 %f644, %r1032;
add.ftz.f32 %f61, %f643, %f644;
add.s32 %r1033, %r4, 2;
setp.ge.u32 %p215, %r1033, %r44;
setp.ne.s32 %p216, %r2, 2;
or.pred %p217, %p216, %p215;
@%p217 bra $L__BB103_42;

add.ftz.f32 %f645, %f5, %f60;
add.ftz.f32 %f646, %f4, %f61;
mul.ftz.f32 %f647, %f646, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f648, %f647;
add.ftz.f32 %f649, %f648, 0f3F800000;
div.approx.ftz.f32 %f650, %f646, %f649;
mul.ftz.f32 %f651, %f645, %f650;
st.global.f32 [%rd24+8], %f651;

$L__BB103_42:
ld.shared.f32 %f652, [%r25+384];
add.ftz.f32 %f653, %f652, %f27;
ld.shared.f32 %f654, [%r27+384];
add.ftz.f32 %f655, %f654, %f31;
ld.shared.f32 %f656, [%r25+896];
add.ftz.f32 %f657, %f656, %f653;
ld.shared.f32 %f658, [%r27+896];
add.ftz.f32 %f659, %f658, %f655;
ld.shared.f32 %f660, [%r25+1408];
add.ftz.f32 %f661, %f660, %f657;
ld.shared.f32 %f662, [%r27+1408];
add.ftz.f32 %f663, %f662, %f659;
mov.b32 %r1034, %f661;
shfl.sync.bfly.b32 %r1038|%p218, %r1034, %r1008, %r1007, %r1009;
mov.b32 %f664, %r1038;
add.ftz.f32 %f665, %f661, %f664;
mov.b32 %r1039, %f665;
shfl.sync.bfly.b32 %r1041|%p219, %r1039, %r1012, %r1007, %r1009;
mov.b32 %f666, %r1041;
add.ftz.f32 %f667, %f665, %f666;
mov.b32 %r1042, %f667;
shfl.sync.bfly.b32 %r1044|%p220, %r1042, %r1015, %r1007, %r1009;
mov.b32 %f668, %r1044;
add.ftz.f32 %f669, %f667, %f668;
mov.b32 %r1045, %f669;
shfl.sync.bfly.b32 %r1047|%p221, %r1045, %r1018, %r1007, %r1009;
mov.b32 %f670, %r1047;
add.ftz.f32 %f671, %f669, %f670;
mov.b32 %r1048, %f671;
shfl.sync.bfly.b32 %r1050|%p222, %r1048, %r1021, %r1007, %r1009;
mov.b32 %f672, %r1050;
add.ftz.f32 %f62, %f671, %f672;
mov.b32 %r1051, %f663;
shfl.sync.bfly.b32 %r1052|%p223, %r1051, %r1008, %r1007, %r1009;
mov.b32 %f673, %r1052;
add.ftz.f32 %f674, %f663, %f673;
mov.b32 %r1053, %f674;
shfl.sync.bfly.b32 %r1054|%p224, %r1053, %r1012, %r1007, %r1009;
mov.b32 %f675, %r1054;
add.ftz.f32 %f676, %f674, %f675;
mov.b32 %r1055, %f676;
shfl.sync.bfly.b32 %r1056|%p225, %r1055, %r1015, %r1007, %r1009;
mov.b32 %f677, %r1056;
add.ftz.f32 %f678, %f676, %f677;
mov.b32 %r1057, %f678;
shfl.sync.bfly.b32 %r1058|%p226, %r1057, %r1018, %r1007, %r1009;
mov.b32 %f679, %r1058;
add.ftz.f32 %f680, %f678, %f679;
mov.b32 %r1059, %f680;
shfl.sync.bfly.b32 %r1060|%p227, %r1059, %r1021, %r1007, %r1009;
mov.b32 %f681, %r1060;
add.ftz.f32 %f63, %f680, %f681;
add.s32 %r1061, %r4, 3;
setp.ge.u32 %p228, %r1061, %r44;
setp.ne.s32 %p229, %r2, 3;
or.pred %p230, %p229, %p228;
@%p230 bra $L__BB103_78;

add.ftz.f32 %f682, %f5, %f62;
add.ftz.f32 %f683, %f4, %f63;
mul.ftz.f32 %f684, %f683, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f685, %f684;
add.ftz.f32 %f686, %f685, 0f3F800000;
div.approx.ftz.f32 %f687, %f683, %f686;
mul.ftz.f32 %f688, %f682, %f687;
st.global.f32 [%rd24+12], %f688;
bra.uni $L__BB103_78;

$L__BB103_25:
ld.shared.f32 %f141, [%r25];
add.ftz.f32 %f142, %f141, %f30;
ld.shared.f32 %f143, [%r25+512];
add.ftz.f32 %f144, %f143, %f142;
ld.shared.f32 %f145, [%r25+1024];
add.ftz.f32 %f146, %f145, %f144;
mov.b32 %r655, %f146;
mov.u32 %r656, 31;
mov.u32 %r657, 16;
mov.u32 %r658, -1;
shfl.sync.bfly.b32 %r659|%p38, %r655, %r657, %r656, %r658;
mov.b32 %f147, %r659;
add.ftz.f32 %f148, %f146, %f147;
mov.b32 %r660, %f148;
mov.u32 %r661, 8;
shfl.sync.bfly.b32 %r662|%p39, %r660, %r661, %r656, %r658;
mov.b32 %f149, %r662;
add.ftz.f32 %f150, %f148, %f149;
mov.b32 %r663, %f150;
mov.u32 %r664, 4;
shfl.sync.bfly.b32 %r665|%p40, %r663, %r664, %r656, %r658;
mov.b32 %f151, %r665;
add.ftz.f32 %f152, %f150, %f151;
mov.b32 %r666, %f152;
mov.u32 %r667, 2;
shfl.sync.bfly.b32 %r668|%p41, %r666, %r667, %r656, %r658;
mov.b32 %f153, %r668;
add.ftz.f32 %f154, %f152, %f153;
mov.b32 %r669, %f154;
mov.u32 %r670, 1;
shfl.sync.bfly.b32 %r671|%p42, %r669, %r670, %r656, %r658;
mov.b32 %f155, %r671;
add.ftz.f32 %f36, %f154, %f155;
setp.ge.u32 %p43, %r4, %r44;
setp.ne.s32 %p44, %r2, 0;
or.pred %p45, %p44, %p43;
@%p45 bra $L__BB103_27;

add.ftz.f32 %f156, %f5, %f36;
st.global.f32 [%rd24], %f156;

$L__BB103_27:
ld.shared.f32 %f157, [%r25+128];
add.ftz.f32 %f158, %f157, %f29;
ld.shared.f32 %f159, [%r25+640];
add.ftz.f32 %f160, %f159, %f158;
ld.shared.f32 %f161, [%r25+1152];
add.ftz.f32 %f162, %f161, %f160;
mov.b32 %r672, %f162;
shfl.sync.bfly.b32 %r676|%p46, %r672, %r657, %r656, %r658;
mov.b32 %f163, %r676;
add.ftz.f32 %f164, %f162, %f163;
mov.b32 %r677, %f164;
shfl.sync.bfly.b32 %r679|%p47, %r677, %r661, %r656, %r658;
mov.b32 %f165, %r679;
add.ftz.f32 %f166, %f164, %f165;
mov.b32 %r680, %f166;
shfl.sync.bfly.b32 %r682|%p48, %r680, %r664, %r656, %r658;
mov.b32 %f167, %r682;
add.ftz.f32 %f168, %f166, %f167;
mov.b32 %r683, %f168;
shfl.sync.bfly.b32 %r685|%p49, %r683, %r667, %r656, %r658;
mov.b32 %f169, %r685;
add.ftz.f32 %f170, %f168, %f169;
mov.b32 %r686, %f170;
shfl.sync.bfly.b32 %r688|%p50, %r686, %r670, %r656, %r658;
mov.b32 %f171, %r688;
add.ftz.f32 %f37, %f170, %f171;
add.s32 %r689, %r4, 1;
setp.ge.u32 %p51, %r689, %r44;
setp.ne.s32 %p52, %r2, 1;
or.pred %p53, %p52, %p51;
@%p53 bra $L__BB103_29;

add.ftz.f32 %f172, %f5, %f37;
st.global.f32 [%rd24+4], %f172;

$L__BB103_29:
ld.shared.f32 %f173, [%r25+256];
add.ftz.f32 %f174, %f173, %f28;
ld.shared.f32 %f175, [%r25+768];
add.ftz.f32 %f176, %f175, %f174;
ld.shared.f32 %f177, [%r25+1280];
add.ftz.f32 %f178, %f177, %f176;
mov.b32 %r690, %f178;
mov.u32 %r691, 31;
mov.u32 %r692, 16;
mov.u32 %r693, -1;
shfl.sync.bfly.b32 %r694|%p54, %r690, %r692, %r691, %r693;
mov.b32 %f179, %r694;
add.ftz.f32 %f180, %f178, %f179;
mov.b32 %r695, %f180;
mov.u32 %r696, 8;
shfl.sync.bfly.b32 %r697|%p55, %r695, %r696, %r691, %r693;
mov.b32 %f181, %r697;
add.ftz.f32 %f182, %f180, %f181;
mov.b32 %r698, %f182;
mov.u32 %r699, 4;
shfl.sync.bfly.b32 %r700|%p56, %r698, %r699, %r691, %r693;
mov.b32 %f183, %r700;
add.ftz.f32 %f184, %f182, %f183;
mov.b32 %r701, %f184;
mov.u32 %r702, 2;
shfl.sync.bfly.b32 %r703|%p57, %r701, %r702, %r691, %r693;
mov.b32 %f185, %r703;
add.ftz.f32 %f186, %f184, %f185;
mov.b32 %r704, %f186;
mov.u32 %r705, 1;
shfl.sync.bfly.b32 %r706|%p58, %r704, %r705, %r691, %r693;
mov.b32 %f187, %r706;
add.ftz.f32 %f38, %f186, %f187;
add.s32 %r707, %r4, 2;
setp.ge.u32 %p59, %r707, %r44;
setp.ne.s32 %p60, %r2, 2;
or.pred %p61, %p60, %p59;
@%p61 bra $L__BB103_31;

add.ftz.f32 %f188, %f5, %f38;
st.global.f32 [%rd24+8], %f188;

$L__BB103_31:
ld.shared.f32 %f189, [%r25+384];
add.ftz.f32 %f190, %f189, %f27;
ld.shared.f32 %f191, [%r25+896];
add.ftz.f32 %f192, %f191, %f190;
ld.shared.f32 %f193, [%r25+1408];
add.ftz.f32 %f194, %f193, %f192;
mov.b32 %r708, %f194;
shfl.sync.bfly.b32 %r712|%p62, %r708, %r692, %r691, %r693;
mov.b32 %f195, %r712;
add.ftz.f32 %f196, %f194, %f195;
mov.b32 %r713, %f196;
shfl.sync.bfly.b32 %r715|%p63, %r713, %r696, %r691, %r693;
mov.b32 %f197, %r715;
add.ftz.f32 %f198, %f196, %f197;
mov.b32 %r716, %f198;
shfl.sync.bfly.b32 %r718|%p64, %r716, %r699, %r691, %r693;
mov.b32 %f199, %r718;
add.ftz.f32 %f200, %f198, %f199;
mov.b32 %r719, %f200;
shfl.sync.bfly.b32 %r721|%p65, %r719, %r702, %r691, %r693;
mov.b32 %f201, %r721;
add.ftz.f32 %f202, %f200, %f201;
mov.b32 %r722, %f202;
shfl.sync.bfly.b32 %r724|%p66, %r722, %r705, %r691, %r693;
mov.b32 %f203, %r724;
add.ftz.f32 %f39, %f202, %f203;
add.s32 %r725, %r4, 3;
setp.ge.u32 %p67, %r725, %r44;
setp.ne.s32 %p68, %r2, 3;
or.pred %p69, %p68, %p67;
@%p69 bra $L__BB103_78;

add.ftz.f32 %f204, %f5, %f39;
st.global.f32 [%rd24+12], %f204;
bra.uni $L__BB103_78;

$L__BB103_44:
ld.param.u32 %r1297, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p71, %r1297, 3;
@%p71 bra $L__BB103_54;

ld.param.u32 %r1298, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.ne.s32 %p72, %r1298, 6;
@%p72 bra $L__BB103_70;

ld.shared.f32 %f205, [%r25];
add.ftz.f32 %f206, %f205, %f30;
ld.shared.f32 %f207, [%r27];
add.ftz.f32 %f208, %f207, %f34;
ld.shared.f32 %f209, [%r25+512];
add.ftz.f32 %f210, %f209, %f206;
ld.shared.f32 %f211, [%r27+512];
add.ftz.f32 %f212, %f211, %f208;
ld.shared.f32 %f213, [%r25+1024];
add.ftz.f32 %f214, %f213, %f210;
ld.shared.f32 %f215, [%r27+1024];
add.ftz.f32 %f216, %f215, %f212;
mov.b32 %r729, %f214;
mov.u32 %r730, 31;
mov.u32 %r731, 16;
mov.u32 %r732, -1;
shfl.sync.bfly.b32 %r733|%p75, %r729, %r731, %r730, %r732;
mov.b32 %f217, %r733;
add.ftz.f32 %f218, %f214, %f217;
mov.b32 %r734, %f218;
mov.u32 %r735, 8;
shfl.sync.bfly.b32 %r736|%p76, %r734, %r735, %r730, %r732;
mov.b32 %f219, %r736;
add.ftz.f32 %f220, %f218, %f219;
mov.b32 %r737, %f220;
mov.u32 %r738, 4;
shfl.sync.bfly.b32 %r739|%p77, %r737, %r738, %r730, %r732;
mov.b32 %f221, %r739;
add.ftz.f32 %f222, %f220, %f221;
mov.b32 %r740, %f222;
mov.u32 %r741, 2;
shfl.sync.bfly.b32 %r742|%p78, %r740, %r741, %r730, %r732;
mov.b32 %f223, %r742;
add.ftz.f32 %f224, %f222, %f223;
mov.b32 %r743, %f224;
mov.u32 %r744, 1;
shfl.sync.bfly.b32 %r745|%p79, %r743, %r744, %r730, %r732;
mov.b32 %f225, %r745;
add.ftz.f32 %f40, %f224, %f225;
mov.b32 %r746, %f216;
shfl.sync.bfly.b32 %r747|%p80, %r746, %r731, %r730, %r732;
mov.b32 %f226, %r747;
add.ftz.f32 %f227, %f216, %f226;
mov.b32 %r748, %f227;
shfl.sync.bfly.b32 %r749|%p81, %r748, %r735, %r730, %r732;
mov.b32 %f228, %r749;
add.ftz.f32 %f229, %f227, %f228;
mov.b32 %r750, %f229;
shfl.sync.bfly.b32 %r751|%p82, %r750, %r738, %r730, %r732;
mov.b32 %f230, %r751;
add.ftz.f32 %f231, %f229, %f230;
mov.b32 %r752, %f231;
shfl.sync.bfly.b32 %r753|%p83, %r752, %r741, %r730, %r732;
mov.b32 %f232, %r753;
add.ftz.f32 %f233, %f231, %f232;
mov.b32 %r754, %f233;
shfl.sync.bfly.b32 %r755|%p84, %r754, %r744, %r730, %r732;
mov.b32 %f234, %r755;
add.ftz.f32 %f41, %f233, %f234;
setp.ge.u32 %p85, %r4, %r44;
setp.ne.s32 %p86, %r2, 0;
or.pred %p87, %p86, %p85;
@%p87 bra $L__BB103_48;

ld.param.f32 %f989, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f235, %f5, %f40;
add.ftz.f32 %f236, %f4, %f41;
min.ftz.f32 %f237, %f236, %f989;
min.ftz.f32 %f238, %f235, %f989;
max.ftz.f32 %f239, %f238, %f35;
mul.ftz.f32 %f240, %f237, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f241, %f240;
add.ftz.f32 %f242, %f241, 0f3F800000;
div.approx.ftz.f32 %f243, %f237, %f242;
mul.ftz.f32 %f244, %f239, %f243;
st.global.f32 [%rd24], %f244;

$L__BB103_48:
ld.shared.f32 %f245, [%r25+128];
add.ftz.f32 %f246, %f245, %f29;
ld.shared.f32 %f247, [%r27+128];
add.ftz.f32 %f248, %f247, %f33;
ld.shared.f32 %f249, [%r25+640];
add.ftz.f32 %f250, %f249, %f246;
ld.shared.f32 %f251, [%r27+640];
add.ftz.f32 %f252, %f251, %f248;
ld.shared.f32 %f253, [%r25+1152];
add.ftz.f32 %f254, %f253, %f250;
ld.shared.f32 %f255, [%r27+1152];
add.ftz.f32 %f256, %f255, %f252;
mov.b32 %r756, %f254;
shfl.sync.bfly.b32 %r760|%p88, %r756, %r731, %r730, %r732;
mov.b32 %f257, %r760;
add.ftz.f32 %f258, %f254, %f257;
mov.b32 %r761, %f258;
shfl.sync.bfly.b32 %r763|%p89, %r761, %r735, %r730, %r732;
mov.b32 %f259, %r763;
add.ftz.f32 %f260, %f258, %f259;
mov.b32 %r764, %f260;
shfl.sync.bfly.b32 %r766|%p90, %r764, %r738, %r730, %r732;
mov.b32 %f261, %r766;
add.ftz.f32 %f262, %f260, %f261;
mov.b32 %r767, %f262;
shfl.sync.bfly.b32 %r769|%p91, %r767, %r741, %r730, %r732;
mov.b32 %f263, %r769;
add.ftz.f32 %f264, %f262, %f263;
mov.b32 %r770, %f264;
shfl.sync.bfly.b32 %r772|%p92, %r770, %r744, %r730, %r732;
mov.b32 %f265, %r772;
add.ftz.f32 %f42, %f264, %f265;
mov.b32 %r773, %f256;
shfl.sync.bfly.b32 %r774|%p93, %r773, %r731, %r730, %r732;
mov.b32 %f266, %r774;
add.ftz.f32 %f267, %f256, %f266;
mov.b32 %r775, %f267;
shfl.sync.bfly.b32 %r776|%p94, %r775, %r735, %r730, %r732;
mov.b32 %f268, %r776;
add.ftz.f32 %f269, %f267, %f268;
mov.b32 %r777, %f269;
shfl.sync.bfly.b32 %r778|%p95, %r777, %r738, %r730, %r732;
mov.b32 %f270, %r778;
add.ftz.f32 %f271, %f269, %f270;
mov.b32 %r779, %f271;
shfl.sync.bfly.b32 %r780|%p96, %r779, %r741, %r730, %r732;
mov.b32 %f272, %r780;
add.ftz.f32 %f273, %f271, %f272;
mov.b32 %r781, %f273;
shfl.sync.bfly.b32 %r782|%p97, %r781, %r744, %r730, %r732;
mov.b32 %f274, %r782;
add.ftz.f32 %f43, %f273, %f274;
add.s32 %r783, %r4, 1;
setp.ge.u32 %p98, %r783, %r44;
setp.ne.s32 %p99, %r2, 1;
or.pred %p100, %p99, %p98;
@%p100 bra $L__BB103_50;

ld.param.f32 %f988, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f275, %f5, %f42;
add.ftz.f32 %f276, %f4, %f43;
min.ftz.f32 %f277, %f276, %f988;
min.ftz.f32 %f278, %f275, %f988;
max.ftz.f32 %f279, %f278, %f35;
mul.ftz.f32 %f280, %f277, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f281, %f280;
add.ftz.f32 %f282, %f281, 0f3F800000;
div.approx.ftz.f32 %f283, %f277, %f282;
mul.ftz.f32 %f284, %f279, %f283;
st.global.f32 [%rd24+4], %f284;

$L__BB103_50:
ld.shared.f32 %f285, [%r25+256];
add.ftz.f32 %f286, %f285, %f28;
ld.shared.f32 %f287, [%r27+256];
add.ftz.f32 %f288, %f287, %f32;
ld.shared.f32 %f289, [%r25+768];
add.ftz.f32 %f290, %f289, %f286;
ld.shared.f32 %f291, [%r27+768];
add.ftz.f32 %f292, %f291, %f288;
ld.shared.f32 %f293, [%r25+1280];
add.ftz.f32 %f294, %f293, %f290;
ld.shared.f32 %f295, [%r27+1280];
add.ftz.f32 %f296, %f295, %f292;
mov.b32 %r784, %f294;
mov.u32 %r785, 31;
mov.u32 %r786, 16;
mov.u32 %r787, -1;
shfl.sync.bfly.b32 %r788|%p101, %r784, %r786, %r785, %r787;
mov.b32 %f297, %r788;
add.ftz.f32 %f298, %f294, %f297;
mov.b32 %r789, %f298;
mov.u32 %r790, 8;
shfl.sync.bfly.b32 %r791|%p102, %r789, %r790, %r785, %r787;
mov.b32 %f299, %r791;
add.ftz.f32 %f300, %f298, %f299;
mov.b32 %r792, %f300;
mov.u32 %r793, 4;
shfl.sync.bfly.b32 %r794|%p103, %r792, %r793, %r785, %r787;
mov.b32 %f301, %r794;
add.ftz.f32 %f302, %f300, %f301;
mov.b32 %r795, %f302;
mov.u32 %r796, 2;
shfl.sync.bfly.b32 %r797|%p104, %r795, %r796, %r785, %r787;
mov.b32 %f303, %r797;
add.ftz.f32 %f304, %f302, %f303;
mov.b32 %r798, %f304;
mov.u32 %r799, 1;
shfl.sync.bfly.b32 %r800|%p105, %r798, %r799, %r785, %r787;
mov.b32 %f305, %r800;
add.ftz.f32 %f44, %f304, %f305;
mov.b32 %r801, %f296;
shfl.sync.bfly.b32 %r802|%p106, %r801, %r786, %r785, %r787;
mov.b32 %f306, %r802;
add.ftz.f32 %f307, %f296, %f306;
mov.b32 %r803, %f307;
shfl.sync.bfly.b32 %r804|%p107, %r803, %r790, %r785, %r787;
mov.b32 %f308, %r804;
add.ftz.f32 %f309, %f307, %f308;
mov.b32 %r805, %f309;
shfl.sync.bfly.b32 %r806|%p108, %r805, %r793, %r785, %r787;
mov.b32 %f310, %r806;
add.ftz.f32 %f311, %f309, %f310;
mov.b32 %r807, %f311;
shfl.sync.bfly.b32 %r808|%p109, %r807, %r796, %r785, %r787;
mov.b32 %f312, %r808;
add.ftz.f32 %f313, %f311, %f312;
mov.b32 %r809, %f313;
shfl.sync.bfly.b32 %r810|%p110, %r809, %r799, %r785, %r787;
mov.b32 %f314, %r810;
add.ftz.f32 %f45, %f313, %f314;
add.s32 %r811, %r4, 2;
setp.ge.u32 %p111, %r811, %r44;
setp.ne.s32 %p112, %r2, 2;
or.pred %p113, %p112, %p111;
@%p113 bra $L__BB103_52;

ld.param.f32 %f987, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f315, %f5, %f44;
add.ftz.f32 %f316, %f4, %f45;
min.ftz.f32 %f317, %f316, %f987;
min.ftz.f32 %f318, %f315, %f987;
max.ftz.f32 %f319, %f318, %f35;
mul.ftz.f32 %f320, %f317, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f321, %f320;
add.ftz.f32 %f322, %f321, 0f3F800000;
div.approx.ftz.f32 %f323, %f317, %f322;
mul.ftz.f32 %f324, %f319, %f323;
st.global.f32 [%rd24+8], %f324;

$L__BB103_52:
ld.shared.f32 %f325, [%r25+384];
add.ftz.f32 %f326, %f325, %f27;
ld.shared.f32 %f327, [%r27+384];
add.ftz.f32 %f328, %f327, %f31;
ld.shared.f32 %f329, [%r25+896];
add.ftz.f32 %f330, %f329, %f326;
ld.shared.f32 %f331, [%r27+896];
add.ftz.f32 %f332, %f331, %f328;
ld.shared.f32 %f333, [%r25+1408];
add.ftz.f32 %f334, %f333, %f330;
ld.shared.f32 %f335, [%r27+1408];
add.ftz.f32 %f336, %f335, %f332;
mov.b32 %r812, %f334;
shfl.sync.bfly.b32 %r816|%p114, %r812, %r786, %r785, %r787;
mov.b32 %f337, %r816;
add.ftz.f32 %f338, %f334, %f337;
mov.b32 %r817, %f338;
shfl.sync.bfly.b32 %r819|%p115, %r817, %r790, %r785, %r787;
mov.b32 %f339, %r819;
add.ftz.f32 %f340, %f338, %f339;
mov.b32 %r820, %f340;
shfl.sync.bfly.b32 %r822|%p116, %r820, %r793, %r785, %r787;
mov.b32 %f341, %r822;
add.ftz.f32 %f342, %f340, %f341;
mov.b32 %r823, %f342;
shfl.sync.bfly.b32 %r825|%p117, %r823, %r796, %r785, %r787;
mov.b32 %f343, %r825;
add.ftz.f32 %f344, %f342, %f343;
mov.b32 %r826, %f344;
shfl.sync.bfly.b32 %r828|%p118, %r826, %r799, %r785, %r787;
mov.b32 %f345, %r828;
add.ftz.f32 %f46, %f344, %f345;
mov.b32 %r829, %f336;
shfl.sync.bfly.b32 %r830|%p119, %r829, %r786, %r785, %r787;
mov.b32 %f346, %r830;
add.ftz.f32 %f347, %f336, %f346;
mov.b32 %r831, %f347;
shfl.sync.bfly.b32 %r832|%p120, %r831, %r790, %r785, %r787;
mov.b32 %f348, %r832;
add.ftz.f32 %f349, %f347, %f348;
mov.b32 %r833, %f349;
shfl.sync.bfly.b32 %r834|%p121, %r833, %r793, %r785, %r787;
mov.b32 %f350, %r834;
add.ftz.f32 %f351, %f349, %f350;
mov.b32 %r835, %f351;
shfl.sync.bfly.b32 %r836|%p122, %r835, %r796, %r785, %r787;
mov.b32 %f352, %r836;
add.ftz.f32 %f353, %f351, %f352;
mov.b32 %r837, %f353;
shfl.sync.bfly.b32 %r838|%p123, %r837, %r799, %r785, %r787;
mov.b32 %f354, %r838;
add.ftz.f32 %f47, %f353, %f354;
add.s32 %r839, %r4, 3;
setp.ge.u32 %p124, %r839, %r44;
setp.ne.s32 %p125, %r2, 3;
or.pred %p126, %p125, %p124;
@%p126 bra $L__BB103_78;

ld.param.f32 %f986, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f355, %f5, %f46;
add.ftz.f32 %f356, %f4, %f47;
min.ftz.f32 %f357, %f356, %f986;
min.ftz.f32 %f358, %f355, %f986;
max.ftz.f32 %f359, %f358, %f35;
mul.ftz.f32 %f360, %f357, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f361, %f360;
add.ftz.f32 %f362, %f361, 0f3F800000;
div.approx.ftz.f32 %f363, %f357, %f362;
mul.ftz.f32 %f364, %f359, %f363;
st.global.f32 [%rd24+12], %f364;
bra.uni $L__BB103_78;

$L__BB103_70:
ld.shared.f32 %f853, [%r25];
add.ftz.f32 %f854, %f853, %f30;
ld.shared.f32 %f855, [%r27];
add.ftz.f32 %f856, %f855, %f34;
ld.shared.f32 %f857, [%r25+512];
add.ftz.f32 %f858, %f857, %f854;
ld.shared.f32 %f859, [%r27+512];
add.ftz.f32 %f860, %f859, %f856;
ld.shared.f32 %f861, [%r25+1024];
add.ftz.f32 %f862, %f861, %f858;
ld.shared.f32 %f863, [%r27+1024];
add.ftz.f32 %f864, %f863, %f860;
mov.b32 %r1173, %f862;
mov.u32 %r1174, 31;
mov.u32 %r1175, 16;
mov.u32 %r1176, -1;
shfl.sync.bfly.b32 %r1177|%p283, %r1173, %r1175, %r1174, %r1176;
mov.b32 %f865, %r1177;
add.ftz.f32 %f866, %f862, %f865;
mov.b32 %r1178, %f866;
mov.u32 %r1179, 8;
shfl.sync.bfly.b32 %r1180|%p284, %r1178, %r1179, %r1174, %r1176;
mov.b32 %f867, %r1180;
add.ftz.f32 %f868, %f866, %f867;
mov.b32 %r1181, %f868;
mov.u32 %r1182, 4;
shfl.sync.bfly.b32 %r1183|%p285, %r1181, %r1182, %r1174, %r1176;
mov.b32 %f869, %r1183;
add.ftz.f32 %f870, %f868, %f869;
mov.b32 %r1184, %f870;
mov.u32 %r1185, 2;
shfl.sync.bfly.b32 %r1186|%p286, %r1184, %r1185, %r1174, %r1176;
mov.b32 %f871, %r1186;
add.ftz.f32 %f872, %f870, %f871;
mov.b32 %r1187, %f872;
mov.u32 %r1188, 1;
shfl.sync.bfly.b32 %r1189|%p287, %r1187, %r1188, %r1174, %r1176;
mov.b32 %f873, %r1189;
add.ftz.f32 %f72, %f872, %f873;
mov.b32 %r1190, %f864;
shfl.sync.bfly.b32 %r1191|%p288, %r1190, %r1175, %r1174, %r1176;
mov.b32 %f874, %r1191;
add.ftz.f32 %f875, %f864, %f874;
mov.b32 %r1192, %f875;
shfl.sync.bfly.b32 %r1193|%p289, %r1192, %r1179, %r1174, %r1176;
mov.b32 %f876, %r1193;
add.ftz.f32 %f877, %f875, %f876;
mov.b32 %r1194, %f877;
shfl.sync.bfly.b32 %r1195|%p290, %r1194, %r1182, %r1174, %r1176;
mov.b32 %f878, %r1195;
add.ftz.f32 %f879, %f877, %f878;
mov.b32 %r1196, %f879;
shfl.sync.bfly.b32 %r1197|%p291, %r1196, %r1185, %r1174, %r1176;
mov.b32 %f880, %r1197;
add.ftz.f32 %f881, %f879, %f880;
mov.b32 %r1198, %f881;
shfl.sync.bfly.b32 %r1199|%p292, %r1198, %r1188, %r1174, %r1176;
mov.b32 %f882, %r1199;
add.ftz.f32 %f73, %f881, %f882;
setp.ge.u32 %p293, %r4, %r44;
setp.ne.s32 %p294, %r2, 0;
or.pred %p295, %p294, %p293;
@%p295 bra $L__BB103_72;

add.ftz.f32 %f883, %f5, %f72;
add.ftz.f32 %f884, %f4, %f73;
mul.ftz.f32 %f885, %f883, %f884;
st.global.f32 [%rd24], %f885;

$L__BB103_72:
ld.shared.f32 %f886, [%r25+128];
add.ftz.f32 %f887, %f886, %f29;
ld.shared.f32 %f888, [%r27+128];
add.ftz.f32 %f889, %f888, %f33;
ld.shared.f32 %f890, [%r25+640];
add.ftz.f32 %f891, %f890, %f887;
ld.shared.f32 %f892, [%r27+640];
add.ftz.f32 %f893, %f892, %f889;
ld.shared.f32 %f894, [%r25+1152];
add.ftz.f32 %f895, %f894, %f891;
ld.shared.f32 %f896, [%r27+1152];
add.ftz.f32 %f897, %f896, %f893;
mov.b32 %r1200, %f895;
shfl.sync.bfly.b32 %r1204|%p296, %r1200, %r1175, %r1174, %r1176;
mov.b32 %f898, %r1204;
add.ftz.f32 %f899, %f895, %f898;
mov.b32 %r1205, %f899;
shfl.sync.bfly.b32 %r1207|%p297, %r1205, %r1179, %r1174, %r1176;
mov.b32 %f900, %r1207;
add.ftz.f32 %f901, %f899, %f900;
mov.b32 %r1208, %f901;
shfl.sync.bfly.b32 %r1210|%p298, %r1208, %r1182, %r1174, %r1176;
mov.b32 %f902, %r1210;
add.ftz.f32 %f903, %f901, %f902;
mov.b32 %r1211, %f903;
shfl.sync.bfly.b32 %r1213|%p299, %r1211, %r1185, %r1174, %r1176;
mov.b32 %f904, %r1213;
add.ftz.f32 %f905, %f903, %f904;
mov.b32 %r1214, %f905;
shfl.sync.bfly.b32 %r1216|%p300, %r1214, %r1188, %r1174, %r1176;
mov.b32 %f906, %r1216;
add.ftz.f32 %f74, %f905, %f906;
mov.b32 %r1217, %f897;
shfl.sync.bfly.b32 %r1218|%p301, %r1217, %r1175, %r1174, %r1176;
mov.b32 %f907, %r1218;
add.ftz.f32 %f908, %f897, %f907;
mov.b32 %r1219, %f908;
shfl.sync.bfly.b32 %r1220|%p302, %r1219, %r1179, %r1174, %r1176;
mov.b32 %f909, %r1220;
add.ftz.f32 %f910, %f908, %f909;
mov.b32 %r1221, %f910;
shfl.sync.bfly.b32 %r1222|%p303, %r1221, %r1182, %r1174, %r1176;
mov.b32 %f911, %r1222;
add.ftz.f32 %f912, %f910, %f911;
mov.b32 %r1223, %f912;
shfl.sync.bfly.b32 %r1224|%p304, %r1223, %r1185, %r1174, %r1176;
mov.b32 %f913, %r1224;
add.ftz.f32 %f914, %f912, %f913;
mov.b32 %r1225, %f914;
shfl.sync.bfly.b32 %r1226|%p305, %r1225, %r1188, %r1174, %r1176;
mov.b32 %f915, %r1226;
add.ftz.f32 %f75, %f914, %f915;
add.s32 %r1227, %r4, 1;
setp.ge.u32 %p306, %r1227, %r44;
setp.ne.s32 %p307, %r2, 1;
or.pred %p308, %p307, %p306;
@%p308 bra $L__BB103_74;

add.ftz.f32 %f916, %f5, %f74;
add.ftz.f32 %f917, %f4, %f75;
mul.ftz.f32 %f918, %f916, %f917;
st.global.f32 [%rd24+4], %f918;

$L__BB103_74:
ld.shared.f32 %f919, [%r25+256];
add.ftz.f32 %f920, %f919, %f28;
ld.shared.f32 %f921, [%r27+256];
add.ftz.f32 %f922, %f921, %f32;
ld.shared.f32 %f923, [%r25+768];
add.ftz.f32 %f924, %f923, %f920;
ld.shared.f32 %f925, [%r27+768];
add.ftz.f32 %f926, %f925, %f922;
ld.shared.f32 %f927, [%r25+1280];
add.ftz.f32 %f928, %f927, %f924;
ld.shared.f32 %f929, [%r27+1280];
add.ftz.f32 %f930, %f929, %f926;
mov.b32 %r1228, %f928;
mov.u32 %r1229, 31;
mov.u32 %r1230, 16;
mov.u32 %r1231, -1;
shfl.sync.bfly.b32 %r1232|%p309, %r1228, %r1230, %r1229, %r1231;
mov.b32 %f931, %r1232;
add.ftz.f32 %f932, %f928, %f931;
mov.b32 %r1233, %f932;
mov.u32 %r1234, 8;
shfl.sync.bfly.b32 %r1235|%p310, %r1233, %r1234, %r1229, %r1231;
mov.b32 %f933, %r1235;
add.ftz.f32 %f934, %f932, %f933;
mov.b32 %r1236, %f934;
mov.u32 %r1237, 4;
shfl.sync.bfly.b32 %r1238|%p311, %r1236, %r1237, %r1229, %r1231;
mov.b32 %f935, %r1238;
add.ftz.f32 %f936, %f934, %f935;
mov.b32 %r1239, %f936;
mov.u32 %r1240, 2;
shfl.sync.bfly.b32 %r1241|%p312, %r1239, %r1240, %r1229, %r1231;
mov.b32 %f937, %r1241;
add.ftz.f32 %f938, %f936, %f937;
mov.b32 %r1242, %f938;
mov.u32 %r1243, 1;
shfl.sync.bfly.b32 %r1244|%p313, %r1242, %r1243, %r1229, %r1231;
mov.b32 %f939, %r1244;
add.ftz.f32 %f76, %f938, %f939;
mov.b32 %r1245, %f930;
shfl.sync.bfly.b32 %r1246|%p314, %r1245, %r1230, %r1229, %r1231;
mov.b32 %f940, %r1246;
add.ftz.f32 %f941, %f930, %f940;
mov.b32 %r1247, %f941;
shfl.sync.bfly.b32 %r1248|%p315, %r1247, %r1234, %r1229, %r1231;
mov.b32 %f942, %r1248;
add.ftz.f32 %f943, %f941, %f942;
mov.b32 %r1249, %f943;
shfl.sync.bfly.b32 %r1250|%p316, %r1249, %r1237, %r1229, %r1231;
mov.b32 %f944, %r1250;
add.ftz.f32 %f945, %f943, %f944;
mov.b32 %r1251, %f945;
shfl.sync.bfly.b32 %r1252|%p317, %r1251, %r1240, %r1229, %r1231;
mov.b32 %f946, %r1252;
add.ftz.f32 %f947, %f945, %f946;
mov.b32 %r1253, %f947;
shfl.sync.bfly.b32 %r1254|%p318, %r1253, %r1243, %r1229, %r1231;
mov.b32 %f948, %r1254;
add.ftz.f32 %f77, %f947, %f948;
add.s32 %r1255, %r4, 2;
setp.ge.u32 %p319, %r1255, %r44;
setp.ne.s32 %p320, %r2, 2;
or.pred %p321, %p320, %p319;
@%p321 bra $L__BB103_76;

add.ftz.f32 %f949, %f5, %f76;
add.ftz.f32 %f950, %f4, %f77;
mul.ftz.f32 %f951, %f949, %f950;
st.global.f32 [%rd24+8], %f951;

$L__BB103_76:
ld.shared.f32 %f952, [%r25+384];
add.ftz.f32 %f953, %f952, %f27;
ld.shared.f32 %f954, [%r27+384];
add.ftz.f32 %f955, %f954, %f31;
ld.shared.f32 %f956, [%r25+896];
add.ftz.f32 %f957, %f956, %f953;
ld.shared.f32 %f958, [%r27+896];
add.ftz.f32 %f959, %f958, %f955;
ld.shared.f32 %f960, [%r25+1408];
add.ftz.f32 %f961, %f960, %f957;
ld.shared.f32 %f962, [%r27+1408];
add.ftz.f32 %f963, %f962, %f959;
mov.b32 %r1256, %f961;
shfl.sync.bfly.b32 %r1260|%p322, %r1256, %r1230, %r1229, %r1231;
mov.b32 %f964, %r1260;
add.ftz.f32 %f965, %f961, %f964;
mov.b32 %r1261, %f965;
shfl.sync.bfly.b32 %r1263|%p323, %r1261, %r1234, %r1229, %r1231;
mov.b32 %f966, %r1263;
add.ftz.f32 %f967, %f965, %f966;
mov.b32 %r1264, %f967;
shfl.sync.bfly.b32 %r1266|%p324, %r1264, %r1237, %r1229, %r1231;
mov.b32 %f968, %r1266;
add.ftz.f32 %f969, %f967, %f968;
mov.b32 %r1267, %f969;
shfl.sync.bfly.b32 %r1269|%p325, %r1267, %r1240, %r1229, %r1231;
mov.b32 %f970, %r1269;
add.ftz.f32 %f971, %f969, %f970;
mov.b32 %r1270, %f971;
shfl.sync.bfly.b32 %r1272|%p326, %r1270, %r1243, %r1229, %r1231;
mov.b32 %f972, %r1272;
add.ftz.f32 %f78, %f971, %f972;
mov.b32 %r1273, %f963;
shfl.sync.bfly.b32 %r1274|%p327, %r1273, %r1230, %r1229, %r1231;
mov.b32 %f973, %r1274;
add.ftz.f32 %f974, %f963, %f973;
mov.b32 %r1275, %f974;
shfl.sync.bfly.b32 %r1276|%p328, %r1275, %r1234, %r1229, %r1231;
mov.b32 %f975, %r1276;
add.ftz.f32 %f976, %f974, %f975;
mov.b32 %r1277, %f976;
shfl.sync.bfly.b32 %r1278|%p329, %r1277, %r1237, %r1229, %r1231;
mov.b32 %f977, %r1278;
add.ftz.f32 %f978, %f976, %f977;
mov.b32 %r1279, %f978;
shfl.sync.bfly.b32 %r1280|%p330, %r1279, %r1240, %r1229, %r1231;
mov.b32 %f979, %r1280;
add.ftz.f32 %f980, %f978, %f979;
mov.b32 %r1281, %f980;
shfl.sync.bfly.b32 %r1282|%p331, %r1281, %r1243, %r1229, %r1231;
mov.b32 %f981, %r1282;
add.ftz.f32 %f79, %f980, %f981;
add.s32 %r1283, %r4, 3;
setp.ge.u32 %p332, %r1283, %r44;
setp.ne.s32 %p333, %r2, 3;
or.pred %p334, %p333, %p332;
@%p334 bra $L__BB103_78;

add.ftz.f32 %f982, %f5, %f78;
add.ftz.f32 %f983, %f4, %f79;
mul.ftz.f32 %f984, %f982, %f983;
st.global.f32 [%rd24+12], %f984;
bra.uni $L__BB103_78;

$L__BB103_62:
ld.shared.f32 %f689, [%r25];
add.ftz.f32 %f690, %f689, %f30;
ld.shared.f32 %f691, [%r27];
add.ftz.f32 %f692, %f691, %f34;
ld.shared.f32 %f693, [%r25+512];
add.ftz.f32 %f694, %f693, %f690;
ld.shared.f32 %f695, [%r27+512];
add.ftz.f32 %f696, %f695, %f692;
ld.shared.f32 %f697, [%r25+1024];
add.ftz.f32 %f698, %f697, %f694;
ld.shared.f32 %f699, [%r27+1024];
add.ftz.f32 %f700, %f699, %f696;
mov.b32 %r1062, %f698;
mov.u32 %r1063, 31;
mov.u32 %r1064, 16;
mov.u32 %r1065, -1;
shfl.sync.bfly.b32 %r1066|%p231, %r1062, %r1064, %r1063, %r1065;
mov.b32 %f701, %r1066;
add.ftz.f32 %f702, %f698, %f701;
mov.b32 %r1067, %f702;
mov.u32 %r1068, 8;
shfl.sync.bfly.b32 %r1069|%p232, %r1067, %r1068, %r1063, %r1065;
mov.b32 %f703, %r1069;
add.ftz.f32 %f704, %f702, %f703;
mov.b32 %r1070, %f704;
mov.u32 %r1071, 4;
shfl.sync.bfly.b32 %r1072|%p233, %r1070, %r1071, %r1063, %r1065;
mov.b32 %f705, %r1072;
add.ftz.f32 %f706, %f704, %f705;
mov.b32 %r1073, %f706;
mov.u32 %r1074, 2;
shfl.sync.bfly.b32 %r1075|%p234, %r1073, %r1074, %r1063, %r1065;
mov.b32 %f707, %r1075;
add.ftz.f32 %f708, %f706, %f707;
mov.b32 %r1076, %f708;
mov.u32 %r1077, 1;
shfl.sync.bfly.b32 %r1078|%p235, %r1076, %r1077, %r1063, %r1065;
mov.b32 %f709, %r1078;
add.ftz.f32 %f64, %f708, %f709;
mov.b32 %r1079, %f700;
shfl.sync.bfly.b32 %r1080|%p236, %r1079, %r1064, %r1063, %r1065;
mov.b32 %f710, %r1080;
add.ftz.f32 %f711, %f700, %f710;
mov.b32 %r1081, %f711;
shfl.sync.bfly.b32 %r1082|%p237, %r1081, %r1068, %r1063, %r1065;
mov.b32 %f712, %r1082;
add.ftz.f32 %f713, %f711, %f712;
mov.b32 %r1083, %f713;
shfl.sync.bfly.b32 %r1084|%p238, %r1083, %r1071, %r1063, %r1065;
mov.b32 %f714, %r1084;
add.ftz.f32 %f715, %f713, %f714;
mov.b32 %r1085, %f715;
shfl.sync.bfly.b32 %r1086|%p239, %r1085, %r1074, %r1063, %r1065;
mov.b32 %f716, %r1086;
add.ftz.f32 %f717, %f715, %f716;
mov.b32 %r1087, %f717;
shfl.sync.bfly.b32 %r1088|%p240, %r1087, %r1077, %r1063, %r1065;
mov.b32 %f718, %r1088;
add.ftz.f32 %f65, %f717, %f718;
setp.ge.u32 %p241, %r4, %r44;
setp.ne.s32 %p242, %r2, 0;
or.pred %p243, %p242, %p241;
@%p243 bra $L__BB103_64;

add.ftz.f32 %f719, %f5, %f64;
add.ftz.f32 %f720, %f4, %f65;
mul.ftz.f32 %f721, %f720, 0f3F000000;
mul.ftz.f32 %f722, %f720, 0f3F4C422A;
mul.ftz.f32 %f723, %f720, 0f3D372713;
fma.rn.ftz.f32 %f724, %f720, %f723, 0f3F800000;
mul.ftz.f32 %f725, %f722, %f724;
tanh.approx.f32 %f726, %f725;
add.ftz.f32 %f727, %f726, 0f3F800000;
mul.ftz.f32 %f728, %f721, %f727;
mul.ftz.f32 %f729, %f719, %f728;
st.global.f32 [%rd24], %f729;

$L__BB103_64:
ld.shared.f32 %f730, [%r25+128];
add.ftz.f32 %f731, %f730, %f29;
ld.shared.f32 %f732, [%r27+128];
add.ftz.f32 %f733, %f732, %f33;
ld.shared.f32 %f734, [%r25+640];
add.ftz.f32 %f735, %f734, %f731;
ld.shared.f32 %f736, [%r27+640];
add.ftz.f32 %f737, %f736, %f733;
ld.shared.f32 %f738, [%r25+1152];
add.ftz.f32 %f739, %f738, %f735;
ld.shared.f32 %f740, [%r27+1152];
add.ftz.f32 %f741, %f740, %f737;
mov.b32 %r1089, %f739;
shfl.sync.bfly.b32 %r1093|%p244, %r1089, %r1064, %r1063, %r1065;
mov.b32 %f742, %r1093;
add.ftz.f32 %f743, %f739, %f742;
mov.b32 %r1094, %f743;
shfl.sync.bfly.b32 %r1096|%p245, %r1094, %r1068, %r1063, %r1065;
mov.b32 %f744, %r1096;
add.ftz.f32 %f745, %f743, %f744;
mov.b32 %r1097, %f745;
shfl.sync.bfly.b32 %r1099|%p246, %r1097, %r1071, %r1063, %r1065;
mov.b32 %f746, %r1099;
add.ftz.f32 %f747, %f745, %f746;
mov.b32 %r1100, %f747;
shfl.sync.bfly.b32 %r1102|%p247, %r1100, %r1074, %r1063, %r1065;
mov.b32 %f748, %r1102;
add.ftz.f32 %f749, %f747, %f748;
mov.b32 %r1103, %f749;
shfl.sync.bfly.b32 %r1105|%p248, %r1103, %r1077, %r1063, %r1065;
mov.b32 %f750, %r1105;
add.ftz.f32 %f66, %f749, %f750;
mov.b32 %r1106, %f741;
shfl.sync.bfly.b32 %r1107|%p249, %r1106, %r1064, %r1063, %r1065;
mov.b32 %f751, %r1107;
add.ftz.f32 %f752, %f741, %f751;
mov.b32 %r1108, %f752;
shfl.sync.bfly.b32 %r1109|%p250, %r1108, %r1068, %r1063, %r1065;
mov.b32 %f753, %r1109;
add.ftz.f32 %f754, %f752, %f753;
mov.b32 %r1110, %f754;
shfl.sync.bfly.b32 %r1111|%p251, %r1110, %r1071, %r1063, %r1065;
mov.b32 %f755, %r1111;
add.ftz.f32 %f756, %f754, %f755;
mov.b32 %r1112, %f756;
shfl.sync.bfly.b32 %r1113|%p252, %r1112, %r1074, %r1063, %r1065;
mov.b32 %f757, %r1113;
add.ftz.f32 %f758, %f756, %f757;
mov.b32 %r1114, %f758;
shfl.sync.bfly.b32 %r1115|%p253, %r1114, %r1077, %r1063, %r1065;
mov.b32 %f759, %r1115;
add.ftz.f32 %f67, %f758, %f759;
add.s32 %r1116, %r4, 1;
setp.ge.u32 %p254, %r1116, %r44;
setp.ne.s32 %p255, %r2, 1;
or.pred %p256, %p255, %p254;
@%p256 bra $L__BB103_66;

add.ftz.f32 %f760, %f5, %f66;
add.ftz.f32 %f761, %f4, %f67;
mul.ftz.f32 %f762, %f761, 0f3F000000;
mul.ftz.f32 %f763, %f761, 0f3F4C422A;
mul.ftz.f32 %f764, %f761, 0f3D372713;
fma.rn.ftz.f32 %f765, %f761, %f764, 0f3F800000;
mul.ftz.f32 %f766, %f763, %f765;
tanh.approx.f32 %f767, %f766;
add.ftz.f32 %f768, %f767, 0f3F800000;
mul.ftz.f32 %f769, %f762, %f768;
mul.ftz.f32 %f770, %f760, %f769;
st.global.f32 [%rd24+4], %f770;

$L__BB103_66:
ld.shared.f32 %f771, [%r25+256];
add.ftz.f32 %f772, %f771, %f28;
ld.shared.f32 %f773, [%r27+256];
add.ftz.f32 %f774, %f773, %f32;
ld.shared.f32 %f775, [%r25+768];
add.ftz.f32 %f776, %f775, %f772;
ld.shared.f32 %f777, [%r27+768];
add.ftz.f32 %f778, %f777, %f774;
ld.shared.f32 %f779, [%r25+1280];
add.ftz.f32 %f780, %f779, %f776;
ld.shared.f32 %f781, [%r27+1280];
add.ftz.f32 %f782, %f781, %f778;
mov.b32 %r1117, %f780;
mov.u32 %r1118, 31;
mov.u32 %r1119, 16;
mov.u32 %r1120, -1;
shfl.sync.bfly.b32 %r1121|%p257, %r1117, %r1119, %r1118, %r1120;
mov.b32 %f783, %r1121;
add.ftz.f32 %f784, %f780, %f783;
mov.b32 %r1122, %f784;
mov.u32 %r1123, 8;
shfl.sync.bfly.b32 %r1124|%p258, %r1122, %r1123, %r1118, %r1120;
mov.b32 %f785, %r1124;
add.ftz.f32 %f786, %f784, %f785;
mov.b32 %r1125, %f786;
mov.u32 %r1126, 4;
shfl.sync.bfly.b32 %r1127|%p259, %r1125, %r1126, %r1118, %r1120;
mov.b32 %f787, %r1127;
add.ftz.f32 %f788, %f786, %f787;
mov.b32 %r1128, %f788;
mov.u32 %r1129, 2;
shfl.sync.bfly.b32 %r1130|%p260, %r1128, %r1129, %r1118, %r1120;
mov.b32 %f789, %r1130;
add.ftz.f32 %f790, %f788, %f789;
mov.b32 %r1131, %f790;
mov.u32 %r1132, 1;
shfl.sync.bfly.b32 %r1133|%p261, %r1131, %r1132, %r1118, %r1120;
mov.b32 %f791, %r1133;
add.ftz.f32 %f68, %f790, %f791;
mov.b32 %r1134, %f782;
shfl.sync.bfly.b32 %r1135|%p262, %r1134, %r1119, %r1118, %r1120;
mov.b32 %f792, %r1135;
add.ftz.f32 %f793, %f782, %f792;
mov.b32 %r1136, %f793;
shfl.sync.bfly.b32 %r1137|%p263, %r1136, %r1123, %r1118, %r1120;
mov.b32 %f794, %r1137;
add.ftz.f32 %f795, %f793, %f794;
mov.b32 %r1138, %f795;
shfl.sync.bfly.b32 %r1139|%p264, %r1138, %r1126, %r1118, %r1120;
mov.b32 %f796, %r1139;
add.ftz.f32 %f797, %f795, %f796;
mov.b32 %r1140, %f797;
shfl.sync.bfly.b32 %r1141|%p265, %r1140, %r1129, %r1118, %r1120;
mov.b32 %f798, %r1141;
add.ftz.f32 %f799, %f797, %f798;
mov.b32 %r1142, %f799;
shfl.sync.bfly.b32 %r1143|%p266, %r1142, %r1132, %r1118, %r1120;
mov.b32 %f800, %r1143;
add.ftz.f32 %f69, %f799, %f800;
add.s32 %r1144, %r4, 2;
setp.ge.u32 %p267, %r1144, %r44;
setp.ne.s32 %p268, %r2, 2;
or.pred %p269, %p268, %p267;
@%p269 bra $L__BB103_68;

add.ftz.f32 %f801, %f5, %f68;
add.ftz.f32 %f802, %f4, %f69;
mul.ftz.f32 %f803, %f802, 0f3F000000;
mul.ftz.f32 %f804, %f802, 0f3F4C422A;
mul.ftz.f32 %f805, %f802, 0f3D372713;
fma.rn.ftz.f32 %f806, %f802, %f805, 0f3F800000;
mul.ftz.f32 %f807, %f804, %f806;
tanh.approx.f32 %f808, %f807;
add.ftz.f32 %f809, %f808, 0f3F800000;
mul.ftz.f32 %f810, %f803, %f809;
mul.ftz.f32 %f811, %f801, %f810;
st.global.f32 [%rd24+8], %f811;

$L__BB103_68:
ld.shared.f32 %f812, [%r25+384];
add.ftz.f32 %f813, %f812, %f27;
ld.shared.f32 %f814, [%r27+384];
add.ftz.f32 %f815, %f814, %f31;
ld.shared.f32 %f816, [%r25+896];
add.ftz.f32 %f817, %f816, %f813;
ld.shared.f32 %f818, [%r27+896];
add.ftz.f32 %f819, %f818, %f815;
ld.shared.f32 %f820, [%r25+1408];
add.ftz.f32 %f821, %f820, %f817;
ld.shared.f32 %f822, [%r27+1408];
add.ftz.f32 %f823, %f822, %f819;
mov.b32 %r1145, %f821;
shfl.sync.bfly.b32 %r1149|%p270, %r1145, %r1119, %r1118, %r1120;
mov.b32 %f824, %r1149;
add.ftz.f32 %f825, %f821, %f824;
mov.b32 %r1150, %f825;
shfl.sync.bfly.b32 %r1152|%p271, %r1150, %r1123, %r1118, %r1120;
mov.b32 %f826, %r1152;
add.ftz.f32 %f827, %f825, %f826;
mov.b32 %r1153, %f827;
shfl.sync.bfly.b32 %r1155|%p272, %r1153, %r1126, %r1118, %r1120;
mov.b32 %f828, %r1155;
add.ftz.f32 %f829, %f827, %f828;
mov.b32 %r1156, %f829;
shfl.sync.bfly.b32 %r1158|%p273, %r1156, %r1129, %r1118, %r1120;
mov.b32 %f830, %r1158;
add.ftz.f32 %f831, %f829, %f830;
mov.b32 %r1159, %f831;
shfl.sync.bfly.b32 %r1161|%p274, %r1159, %r1132, %r1118, %r1120;
mov.b32 %f832, %r1161;
add.ftz.f32 %f70, %f831, %f832;
mov.b32 %r1162, %f823;
shfl.sync.bfly.b32 %r1163|%p275, %r1162, %r1119, %r1118, %r1120;
mov.b32 %f833, %r1163;
add.ftz.f32 %f834, %f823, %f833;
mov.b32 %r1164, %f834;
shfl.sync.bfly.b32 %r1165|%p276, %r1164, %r1123, %r1118, %r1120;
mov.b32 %f835, %r1165;
add.ftz.f32 %f836, %f834, %f835;
mov.b32 %r1166, %f836;
shfl.sync.bfly.b32 %r1167|%p277, %r1166, %r1126, %r1118, %r1120;
mov.b32 %f837, %r1167;
add.ftz.f32 %f838, %f836, %f837;
mov.b32 %r1168, %f838;
shfl.sync.bfly.b32 %r1169|%p278, %r1168, %r1129, %r1118, %r1120;
mov.b32 %f839, %r1169;
add.ftz.f32 %f840, %f838, %f839;
mov.b32 %r1170, %f840;
shfl.sync.bfly.b32 %r1171|%p279, %r1170, %r1132, %r1118, %r1120;
mov.b32 %f841, %r1171;
add.ftz.f32 %f71, %f840, %f841;
add.s32 %r1172, %r4, 3;
setp.ge.u32 %p280, %r1172, %r44;
setp.ne.s32 %p281, %r2, 3;
or.pred %p282, %p281, %p280;
@%p282 bra $L__BB103_78;

add.ftz.f32 %f842, %f5, %f70;
add.ftz.f32 %f843, %f4, %f71;
mul.ftz.f32 %f844, %f843, 0f3F000000;
mul.ftz.f32 %f845, %f843, 0f3F4C422A;
mul.ftz.f32 %f846, %f843, 0f3D372713;
fma.rn.ftz.f32 %f847, %f843, %f846, 0f3F800000;
mul.ftz.f32 %f848, %f845, %f847;
tanh.approx.f32 %f849, %f848;
add.ftz.f32 %f850, %f849, 0f3F800000;
mul.ftz.f32 %f851, %f844, %f850;
mul.ftz.f32 %f852, %f842, %f851;
st.global.f32 [%rd24+12], %f852;
bra.uni $L__BB103_78;

$L__BB103_54:
ld.shared.f32 %f365, [%r25];
add.ftz.f32 %f366, %f365, %f30;
ld.shared.f32 %f367, [%r27];
add.ftz.f32 %f368, %f367, %f34;
ld.shared.f32 %f369, [%r25+512];
add.ftz.f32 %f370, %f369, %f366;
ld.shared.f32 %f371, [%r27+512];
add.ftz.f32 %f372, %f371, %f368;
ld.shared.f32 %f373, [%r25+1024];
add.ftz.f32 %f374, %f373, %f370;
ld.shared.f32 %f375, [%r27+1024];
add.ftz.f32 %f376, %f375, %f372;
mov.b32 %r840, %f374;
mov.u32 %r841, 31;
mov.u32 %r842, 16;
mov.u32 %r843, -1;
shfl.sync.bfly.b32 %r844|%p127, %r840, %r842, %r841, %r843;
mov.b32 %f377, %r844;
add.ftz.f32 %f378, %f374, %f377;
mov.b32 %r845, %f378;
mov.u32 %r846, 8;
shfl.sync.bfly.b32 %r847|%p128, %r845, %r846, %r841, %r843;
mov.b32 %f379, %r847;
add.ftz.f32 %f380, %f378, %f379;
mov.b32 %r848, %f380;
mov.u32 %r849, 4;
shfl.sync.bfly.b32 %r850|%p129, %r848, %r849, %r841, %r843;
mov.b32 %f381, %r850;
add.ftz.f32 %f382, %f380, %f381;
mov.b32 %r851, %f382;
mov.u32 %r852, 2;
shfl.sync.bfly.b32 %r853|%p130, %r851, %r852, %r841, %r843;
mov.b32 %f383, %r853;
add.ftz.f32 %f384, %f382, %f383;
mov.b32 %r854, %f384;
mov.u32 %r855, 1;
shfl.sync.bfly.b32 %r856|%p131, %r854, %r855, %r841, %r843;
mov.b32 %f385, %r856;
add.ftz.f32 %f48, %f384, %f385;
mov.b32 %r857, %f376;
shfl.sync.bfly.b32 %r858|%p132, %r857, %r842, %r841, %r843;
mov.b32 %f386, %r858;
add.ftz.f32 %f387, %f376, %f386;
mov.b32 %r859, %f387;
shfl.sync.bfly.b32 %r860|%p133, %r859, %r846, %r841, %r843;
mov.b32 %f388, %r860;
add.ftz.f32 %f389, %f387, %f388;
mov.b32 %r861, %f389;
shfl.sync.bfly.b32 %r862|%p134, %r861, %r849, %r841, %r843;
mov.b32 %f390, %r862;
add.ftz.f32 %f391, %f389, %f390;
mov.b32 %r863, %f391;
shfl.sync.bfly.b32 %r864|%p135, %r863, %r852, %r841, %r843;
mov.b32 %f392, %r864;
add.ftz.f32 %f393, %f391, %f392;
mov.b32 %r865, %f393;
shfl.sync.bfly.b32 %r866|%p136, %r865, %r855, %r841, %r843;
mov.b32 %f394, %r866;
add.ftz.f32 %f49, %f393, %f394;
setp.ge.u32 %p137, %r4, %r44;
setp.ne.s32 %p138, %r2, 0;
or.pred %p139, %p138, %p137;
@%p139 bra $L__BB103_56;

add.ftz.f32 %f395, %f5, %f48;
add.ftz.f32 %f396, %f4, %f49;
mov.f32 %f397, 0f40E00000;
min.ftz.f32 %f398, %f396, %f397;
min.ftz.f32 %f399, %f395, %f397;
mov.f32 %f400, 0fC0E00000;
max.ftz.f32 %f401, %f399, %f400;
mul.ftz.f32 %f402, %f398, 0f3FD9DB23;
mul.ftz.f32 %f403, %f402, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f404, %f403;
add.ftz.f32 %f405, %f404, 0f3F800000;
div.approx.ftz.f32 %f406, %f398, %f405;
add.ftz.f32 %f407, %f401, 0f3F800000;
mul.ftz.f32 %f408, %f407, %f406;
st.global.f32 [%rd24], %f408;

$L__BB103_56:
ld.shared.f32 %f409, [%r25+128];
add.ftz.f32 %f410, %f409, %f29;
ld.shared.f32 %f411, [%r27+128];
add.ftz.f32 %f412, %f411, %f33;
ld.shared.f32 %f413, [%r25+640];
add.ftz.f32 %f414, %f413, %f410;
ld.shared.f32 %f415, [%r27+640];
add.ftz.f32 %f416, %f415, %f412;
ld.shared.f32 %f417, [%r25+1152];
add.ftz.f32 %f418, %f417, %f414;
ld.shared.f32 %f419, [%r27+1152];
add.ftz.f32 %f420, %f419, %f416;
mov.b32 %r867, %f418;
shfl.sync.bfly.b32 %r871|%p140, %r867, %r842, %r841, %r843;
mov.b32 %f421, %r871;
add.ftz.f32 %f422, %f418, %f421;
mov.b32 %r872, %f422;
shfl.sync.bfly.b32 %r874|%p141, %r872, %r846, %r841, %r843;
mov.b32 %f423, %r874;
add.ftz.f32 %f424, %f422, %f423;
mov.b32 %r875, %f424;
shfl.sync.bfly.b32 %r877|%p142, %r875, %r849, %r841, %r843;
mov.b32 %f425, %r877;
add.ftz.f32 %f426, %f424, %f425;
mov.b32 %r878, %f426;
shfl.sync.bfly.b32 %r880|%p143, %r878, %r852, %r841, %r843;
mov.b32 %f427, %r880;
add.ftz.f32 %f428, %f426, %f427;
mov.b32 %r881, %f428;
shfl.sync.bfly.b32 %r883|%p144, %r881, %r855, %r841, %r843;
mov.b32 %f429, %r883;
add.ftz.f32 %f50, %f428, %f429;
mov.b32 %r884, %f420;
shfl.sync.bfly.b32 %r885|%p145, %r884, %r842, %r841, %r843;
mov.b32 %f430, %r885;
add.ftz.f32 %f431, %f420, %f430;
mov.b32 %r886, %f431;
shfl.sync.bfly.b32 %r887|%p146, %r886, %r846, %r841, %r843;
mov.b32 %f432, %r887;
add.ftz.f32 %f433, %f431, %f432;
mov.b32 %r888, %f433;
shfl.sync.bfly.b32 %r889|%p147, %r888, %r849, %r841, %r843;
mov.b32 %f434, %r889;
add.ftz.f32 %f435, %f433, %f434;
mov.b32 %r890, %f435;
shfl.sync.bfly.b32 %r891|%p148, %r890, %r852, %r841, %r843;
mov.b32 %f436, %r891;
add.ftz.f32 %f437, %f435, %f436;
mov.b32 %r892, %f437;
shfl.sync.bfly.b32 %r893|%p149, %r892, %r855, %r841, %r843;
mov.b32 %f438, %r893;
add.ftz.f32 %f51, %f437, %f438;
add.s32 %r894, %r4, 1;
setp.ge.u32 %p150, %r894, %r44;
setp.ne.s32 %p151, %r2, 1;
or.pred %p152, %p151, %p150;
@%p152 bra $L__BB103_58;

add.ftz.f32 %f439, %f5, %f50;
add.ftz.f32 %f440, %f4, %f51;
mov.f32 %f441, 0f40E00000;
min.ftz.f32 %f442, %f440, %f441;
min.ftz.f32 %f443, %f439, %f441;
mov.f32 %f444, 0fC0E00000;
max.ftz.f32 %f445, %f443, %f444;
mul.ftz.f32 %f446, %f442, 0f3FD9DB23;
mul.ftz.f32 %f447, %f446, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f448, %f447;
add.ftz.f32 %f449, %f448, 0f3F800000;
div.approx.ftz.f32 %f450, %f442, %f449;
add.ftz.f32 %f451, %f445, 0f3F800000;
mul.ftz.f32 %f452, %f451, %f450;
st.global.f32 [%rd24+4], %f452;

$L__BB103_58:
ld.shared.f32 %f453, [%r25+256];
add.ftz.f32 %f454, %f453, %f28;
ld.shared.f32 %f455, [%r27+256];
add.ftz.f32 %f456, %f455, %f32;
ld.shared.f32 %f457, [%r25+768];
add.ftz.f32 %f458, %f457, %f454;
ld.shared.f32 %f459, [%r27+768];
add.ftz.f32 %f460, %f459, %f456;
ld.shared.f32 %f461, [%r25+1280];
add.ftz.f32 %f462, %f461, %f458;
ld.shared.f32 %f463, [%r27+1280];
add.ftz.f32 %f464, %f463, %f460;
mov.b32 %r895, %f462;
mov.u32 %r896, 31;
mov.u32 %r897, 16;
mov.u32 %r898, -1;
shfl.sync.bfly.b32 %r899|%p153, %r895, %r897, %r896, %r898;
mov.b32 %f465, %r899;
add.ftz.f32 %f466, %f462, %f465;
mov.b32 %r900, %f466;
mov.u32 %r901, 8;
shfl.sync.bfly.b32 %r902|%p154, %r900, %r901, %r896, %r898;
mov.b32 %f467, %r902;
add.ftz.f32 %f468, %f466, %f467;
mov.b32 %r903, %f468;
mov.u32 %r904, 4;
shfl.sync.bfly.b32 %r905|%p155, %r903, %r904, %r896, %r898;
mov.b32 %f469, %r905;
add.ftz.f32 %f470, %f468, %f469;
mov.b32 %r906, %f470;
mov.u32 %r907, 2;
shfl.sync.bfly.b32 %r908|%p156, %r906, %r907, %r896, %r898;
mov.b32 %f471, %r908;
add.ftz.f32 %f472, %f470, %f471;
mov.b32 %r909, %f472;
mov.u32 %r910, 1;
shfl.sync.bfly.b32 %r911|%p157, %r909, %r910, %r896, %r898;
mov.b32 %f473, %r911;
add.ftz.f32 %f52, %f472, %f473;
mov.b32 %r912, %f464;
shfl.sync.bfly.b32 %r913|%p158, %r912, %r897, %r896, %r898;
mov.b32 %f474, %r913;
add.ftz.f32 %f475, %f464, %f474;
mov.b32 %r914, %f475;
shfl.sync.bfly.b32 %r915|%p159, %r914, %r901, %r896, %r898;
mov.b32 %f476, %r915;
add.ftz.f32 %f477, %f475, %f476;
mov.b32 %r916, %f477;
shfl.sync.bfly.b32 %r917|%p160, %r916, %r904, %r896, %r898;
mov.b32 %f478, %r917;
add.ftz.f32 %f479, %f477, %f478;
mov.b32 %r918, %f479;
shfl.sync.bfly.b32 %r919|%p161, %r918, %r907, %r896, %r898;
mov.b32 %f480, %r919;
add.ftz.f32 %f481, %f479, %f480;
mov.b32 %r920, %f481;
shfl.sync.bfly.b32 %r921|%p162, %r920, %r910, %r896, %r898;
mov.b32 %f482, %r921;
add.ftz.f32 %f53, %f481, %f482;
add.s32 %r922, %r4, 2;
setp.ge.u32 %p163, %r922, %r44;
setp.ne.s32 %p164, %r2, 2;
or.pred %p165, %p164, %p163;
@%p165 bra $L__BB103_60;

add.ftz.f32 %f483, %f5, %f52;
add.ftz.f32 %f484, %f4, %f53;
mov.f32 %f485, 0f40E00000;
min.ftz.f32 %f486, %f484, %f485;
min.ftz.f32 %f487, %f483, %f485;
mov.f32 %f488, 0fC0E00000;
max.ftz.f32 %f489, %f487, %f488;
mul.ftz.f32 %f490, %f486, 0f3FD9DB23;
mul.ftz.f32 %f491, %f490, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f492, %f491;
add.ftz.f32 %f493, %f492, 0f3F800000;
div.approx.ftz.f32 %f494, %f486, %f493;
add.ftz.f32 %f495, %f489, 0f3F800000;
mul.ftz.f32 %f496, %f495, %f494;
st.global.f32 [%rd24+8], %f496;

$L__BB103_60:
ld.shared.f32 %f497, [%r25+384];
add.ftz.f32 %f498, %f497, %f27;
ld.shared.f32 %f499, [%r27+384];
add.ftz.f32 %f500, %f499, %f31;
ld.shared.f32 %f501, [%r25+896];
add.ftz.f32 %f502, %f501, %f498;
ld.shared.f32 %f503, [%r27+896];
add.ftz.f32 %f504, %f503, %f500;
ld.shared.f32 %f505, [%r25+1408];
add.ftz.f32 %f506, %f505, %f502;
ld.shared.f32 %f507, [%r27+1408];
add.ftz.f32 %f508, %f507, %f504;
mov.b32 %r923, %f506;
shfl.sync.bfly.b32 %r927|%p166, %r923, %r897, %r896, %r898;
mov.b32 %f509, %r927;
add.ftz.f32 %f510, %f506, %f509;
mov.b32 %r928, %f510;
shfl.sync.bfly.b32 %r930|%p167, %r928, %r901, %r896, %r898;
mov.b32 %f511, %r930;
add.ftz.f32 %f512, %f510, %f511;
mov.b32 %r931, %f512;
shfl.sync.bfly.b32 %r933|%p168, %r931, %r904, %r896, %r898;
mov.b32 %f513, %r933;
add.ftz.f32 %f514, %f512, %f513;
mov.b32 %r934, %f514;
shfl.sync.bfly.b32 %r936|%p169, %r934, %r907, %r896, %r898;
mov.b32 %f515, %r936;
add.ftz.f32 %f516, %f514, %f515;
mov.b32 %r937, %f516;
shfl.sync.bfly.b32 %r939|%p170, %r937, %r910, %r896, %r898;
mov.b32 %f517, %r939;
add.ftz.f32 %f54, %f516, %f517;
mov.b32 %r940, %f508;
shfl.sync.bfly.b32 %r941|%p171, %r940, %r897, %r896, %r898;
mov.b32 %f518, %r941;
add.ftz.f32 %f519, %f508, %f518;
mov.b32 %r942, %f519;
shfl.sync.bfly.b32 %r943|%p172, %r942, %r901, %r896, %r898;
mov.b32 %f520, %r943;
add.ftz.f32 %f521, %f519, %f520;
mov.b32 %r944, %f521;
shfl.sync.bfly.b32 %r945|%p173, %r944, %r904, %r896, %r898;
mov.b32 %f522, %r945;
add.ftz.f32 %f523, %f521, %f522;
mov.b32 %r946, %f523;
shfl.sync.bfly.b32 %r947|%p174, %r946, %r907, %r896, %r898;
mov.b32 %f524, %r947;
add.ftz.f32 %f525, %f523, %f524;
mov.b32 %r948, %f525;
shfl.sync.bfly.b32 %r949|%p175, %r948, %r910, %r896, %r898;
mov.b32 %f526, %r949;
add.ftz.f32 %f55, %f525, %f526;
add.s32 %r950, %r4, 3;
setp.ge.u32 %p176, %r950, %r44;
setp.ne.s32 %p177, %r2, 3;
or.pred %p178, %p177, %p176;
@%p178 bra $L__BB103_78;

add.ftz.f32 %f527, %f5, %f54;
add.ftz.f32 %f528, %f4, %f55;
mov.f32 %f529, 0f40E00000;
min.ftz.f32 %f530, %f528, %f529;
min.ftz.f32 %f531, %f527, %f529;
mov.f32 %f532, 0fC0E00000;
max.ftz.f32 %f533, %f531, %f532;
mul.ftz.f32 %f534, %f530, 0f3FD9DB23;
mul.ftz.f32 %f535, %f534, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f536, %f535;
add.ftz.f32 %f537, %f536, 0f3F800000;
div.approx.ftz.f32 %f538, %f530, %f537;
add.ftz.f32 %f539, %f533, 0f3F800000;
mul.ftz.f32 %f540, %f539, %f538;
st.global.f32 [%rd24+12], %f540;
bra.uni $L__BB103_78;

}
.entry _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<46>;
.reg .b16 %rs<57>;
.reg .f32 %f<113>;
.reg .b32 %r<396>;
.reg .b64 %rd<40>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

ld.param.u64 %rd11, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd12, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd13, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd14, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r16, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r19, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r18, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r17, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
mov.u32 %r1, %tid.y;
shl.b32 %r34, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r34, %r2;
mov.u32 %r35, %ctaid.x;
shl.b32 %r4, %r35, 2;
mov.u32 %r394, %ctaid.y;
setp.eq.s64 %p1, %rd13, 0;
@%p1 bra $L__BB104_2;

cvta.to.global.u64 %rd15, %rd13;
mul.wide.u32 %rd16, %r394, 4;
add.s64 %rd17, %rd15, %rd16;
ld.global.nc.u32 %r393, [%rd17];
mul.hi.u32 %r36, %r394, %r17;
add.s32 %r37, %r36, %r394;
shr.u32 %r38, %r37, %r18;
mul.lo.s32 %r39, %r38, %r19;
sub.s32 %r394, %r394, %r39;
bra.uni $L__BB104_3;

$L__BB104_2:
mul.hi.u32 %r40, %r394, %r22;
add.s32 %r41, %r40, %r394;
shr.u32 %r393, %r41, %r23;

$L__BB104_3:
shr.u32 %r42, %r3, 31;
add.s32 %r43, %r3, %r42;
shr.s32 %r395, %r43, 1;
shr.u32 %r44, %r16, 5;
setp.ge.s32 %p2, %r395, %r44;
mov.f32 %f109, 0f00000000;
mov.f32 %f110, %f109;
mov.f32 %f111, %f109;
mov.f32 %f112, %f109;
@%p2 bra $L__BB104_6;

mov.u32 %r45, %ctaid.z;
and.b32 %r48, %r43, -2;
sub.s32 %r49, %r3, %r48;
shl.b32 %r50, %r49, 1;
cvt.s64.s32 %rd1, %r50;
shl.b32 %r51, %r49, 3;
cvt.s64.s32 %rd18, %r51;
mad.lo.s32 %r52, %r4, %r20, %r395;
mul.hi.u32 %r53, %r45, %r28;
add.s32 %r54, %r53, %r45;
shr.u32 %r55, %r54, %r29;
mad.lo.s32 %r56, %r55, %r31, %r52;
mad.lo.s32 %r57, %r393, %r25, %r56;
mul.wide.s32 %rd2, %r57, 17;
add.s64 %rd19, %rd2, %rd18;
add.s64 %rd3, %rd19, 8;
mul.wide.s32 %rd4, %r20, 17;
cvta.to.global.u64 %rd39, %rd11;
cvta.to.global.u64 %rd6, %rd12;
mul.lo.s32 %r58, %r394, %r26;
cvt.u64.u32 %rd20, %r58;
mul.lo.s32 %r59, %r45, %r32;
cvt.u64.u32 %rd21, %r59;
add.s64 %rd7, %rd20, %rd21;
mov.f32 %f109, 0f00000000;
shl.b64 %rd26, %rd1, 2;

$L__BB104_5:
cvt.u64.u32 %rd22, %r395;
add.s64 %rd23, %rd7, %rd22;
mul.lo.s64 %rd24, %rd23, 36;
add.s64 %rd25, %rd6, %rd24;
add.s64 %rd27, %rd25, %rd26;
add.s64 %rd28, %rd39, %rd3;
add.s64 %rd29, %rd28, -7;
ld.global.nc.u8 %rs5, [%rd28+-7];
cvt.u32.u16 %r128, %rs5;
and.b32 %r129, %r128, 255;
ld.global.nc.u8 %rs6, [%rd28+-6];
cvt.u32.u16 %r130, %rs6;
prmt.b32 %r131, %r130, %r129, 30212;
ld.global.nc.u8 %rs7, [%rd28+-5];
cvt.u32.u16 %r132, %rs7;
prmt.b32 %r133, %r132, %r131, 28756;
ld.global.nc.u8 %rs8, [%rd28+-4];
cvt.u32.u16 %r134, %rs8;
prmt.b32 %r135, %r134, %r133, 1620;
and.b32 %r136, %r135, -2004318072;
shr.u32 %r137, %r136, 1;
or.b32 %r138, %r137, 839922192;
and.b32 %r139, %r131, 30583;
mov.u32 %r140, 201852420;
mov.u32 %r141, 50462976;
prmt.b32 %r142, %r141, %r140, %r139;
mov.u32 %r143, -185009412;
mov.u32 %r144, -33620224;
prmt.b32 %r145, %r144, %r143, %r139;
and.b32 %r146, %r138, 30292;
prmt.b32 %r147, %r142, %r145, %r146;
shr.u32 %r148, %r135, 16;
and.b32 %r149, %r148, 30583;
prmt.b32 %r150, %r141, %r140, %r149;
prmt.b32 %r151, %r144, %r143, %r149;
shr.u32 %r152, %r138, 16;
prmt.b32 %r153, %r150, %r151, %r152;
mov.u32 %r154, 25632;
prmt.b32 %r61, %r147, %r153, %r154;
mov.u32 %r155, 30001;
prmt.b32 %r65, %r147, %r153, %r155;
ld.global.nc.u32 %r62, [%rd27+4];
mov.u32 %r114, 0;

dp4a.s32.s32 %r60, %r61, %r62, %r114;

ld.global.nc.u32 %r66, [%rd27+20];

dp4a.s32.s32 %r64, %r65, %r66, %r60;

ld.global.nc.u8 %rs9, [%rd28+-3];
cvt.u32.u16 %r156, %rs9;
and.b32 %r157, %r156, 255;
ld.global.nc.u8 %rs10, [%rd28+-2];
cvt.u32.u16 %r158, %rs10;
prmt.b32 %r159, %r158, %r157, 30212;
ld.global.nc.u8 %rs11, [%rd28+-1];
cvt.u32.u16 %r160, %rs11;
prmt.b32 %r161, %r160, %r159, 28756;
ld.global.nc.u8 %rs12, [%rd28];
cvt.u32.u16 %r162, %rs12;
prmt.b32 %r163, %r162, %r161, 1620;
and.b32 %r164, %r163, -2004318072;
shr.u32 %r165, %r164, 1;
or.b32 %r166, %r165, 839922192;
and.b32 %r167, %r159, 30583;
prmt.b32 %r168, %r141, %r140, %r167;
prmt.b32 %r169, %r144, %r143, %r167;
and.b32 %r170, %r166, 30292;
prmt.b32 %r171, %r168, %r169, %r170;
shr.u32 %r172, %r163, 16;
and.b32 %r173, %r172, 30583;
prmt.b32 %r174, %r141, %r140, %r173;
prmt.b32 %r175, %r144, %r143, %r173;
shr.u32 %r176, %r166, 16;
prmt.b32 %r177, %r174, %r175, %r176;
prmt.b32 %r69, %r171, %r177, %r154;
prmt.b32 %r73, %r171, %r177, %r155;
ld.global.nc.u32 %r70, [%rd27+8];

dp4a.s32.s32 %r68, %r69, %r70, %r64;

ld.global.nc.u32 %r74, [%rd27+24];

dp4a.s32.s32 %r72, %r73, %r74, %r68;

add.s64 %rd30, %rd39, %rd2;
ld.global.nc.u8 %rs13, [%rd30];
shl.b16 %rs15, %rs13, 7;
and.b16 %rs16, %rs15, 32640;
setp.eq.s16 %p3, %rs13, 255;
setp.eq.s16 %p4, %rs13, 0;
selp.b16 %rs17, 64, %rs16, %p4;
selp.b16 %rs1, 32767, %rs17, %p3;

{ mov.b32 %f25, {0,%rs1};}


mul.ftz.f32 %f33, %f25, 0f3F000000;
ld.global.nc.u32 %r76, [%rd25];

{.reg .f16 low,high;
mov.b32 {low,high},%r76;
cvt.f32.f16 %f26, low;}


mul.ftz.f32 %f34, %f33, %f26;
cvt.rn.f32.s32 %f35, %r72;
fma.rn.ftz.f32 %f112, %f34, %f35, %f112;
add.s64 %rd31, %rd29, %rd4;
ld.global.nc.u8 %rs18, [%rd31];
cvt.u32.u16 %r178, %rs18;
and.b32 %r179, %r178, 255;
ld.global.nc.u8 %rs19, [%rd31+1];
cvt.u32.u16 %r180, %rs19;
prmt.b32 %r181, %r180, %r179, 30212;
ld.global.nc.u8 %rs20, [%rd31+2];
cvt.u32.u16 %r182, %rs20;
prmt.b32 %r183, %r182, %r181, 28756;
ld.global.nc.u8 %rs21, [%rd31+3];
cvt.u32.u16 %r184, %rs21;
prmt.b32 %r185, %r184, %r183, 1620;
and.b32 %r186, %r185, -2004318072;
shr.u32 %r187, %r186, 1;
or.b32 %r188, %r187, 839922192;
and.b32 %r189, %r181, 30583;
prmt.b32 %r190, %r141, %r140, %r189;
prmt.b32 %r191, %r144, %r143, %r189;
and.b32 %r192, %r188, 30292;
prmt.b32 %r193, %r190, %r191, %r192;
shr.u32 %r194, %r185, 16;
and.b32 %r195, %r194, 30583;
prmt.b32 %r196, %r141, %r140, %r195;
prmt.b32 %r197, %r144, %r143, %r195;
shr.u32 %r198, %r188, 16;
prmt.b32 %r199, %r196, %r197, %r198;
prmt.b32 %r78, %r193, %r199, %r154;
prmt.b32 %r82, %r193, %r199, %r155;

dp4a.s32.s32 %r77, %r78, %r62, %r114;


dp4a.s32.s32 %r81, %r82, %r66, %r77;

ld.global.nc.u8 %rs22, [%rd31+4];
cvt.u32.u16 %r200, %rs22;
and.b32 %r201, %r200, 255;
ld.global.nc.u8 %rs23, [%rd31+5];
cvt.u32.u16 %r202, %rs23;
prmt.b32 %r203, %r202, %r201, 30212;
ld.global.nc.u8 %rs24, [%rd31+6];
cvt.u32.u16 %r204, %rs24;
prmt.b32 %r205, %r204, %r203, 28756;
ld.global.nc.u8 %rs25, [%rd31+7];
cvt.u32.u16 %r206, %rs25;
prmt.b32 %r207, %r206, %r205, 1620;
and.b32 %r208, %r207, -2004318072;
shr.u32 %r209, %r208, 1;
or.b32 %r210, %r209, 839922192;
and.b32 %r211, %r203, 30583;
prmt.b32 %r212, %r141, %r140, %r211;
prmt.b32 %r213, %r144, %r143, %r211;
and.b32 %r214, %r210, 30292;
prmt.b32 %r215, %r212, %r213, %r214;
shr.u32 %r216, %r207, 16;
and.b32 %r217, %r216, 30583;
prmt.b32 %r218, %r141, %r140, %r217;
prmt.b32 %r219, %r144, %r143, %r217;
shr.u32 %r220, %r210, 16;
prmt.b32 %r221, %r218, %r219, %r220;
prmt.b32 %r86, %r215, %r221, %r154;
prmt.b32 %r90, %r215, %r221, %r155;

dp4a.s32.s32 %r85, %r86, %r70, %r81;


dp4a.s32.s32 %r89, %r90, %r74, %r85;

add.s64 %rd32, %rd30, %rd4;
ld.global.nc.u8 %rs26, [%rd32];
shl.b16 %rs28, %rs26, 7;
and.b16 %rs29, %rs28, 32640;
setp.eq.s16 %p5, %rs26, 255;
setp.eq.s16 %p6, %rs26, 0;
selp.b16 %rs30, 64, %rs29, %p6;
selp.b16 %rs2, 32767, %rs30, %p5;

{ mov.b32 %f27, {0,%rs2};}


mul.ftz.f32 %f36, %f27, 0f3F000000;

{.reg .f16 low,high;
mov.b32 {low,high},%r76;
cvt.f32.f16 %f28, low;}


mul.ftz.f32 %f37, %f36, %f28;
cvt.rn.f32.s32 %f38, %r89;
fma.rn.ftz.f32 %f111, %f37, %f38, %f111;
add.s64 %rd33, %rd31, %rd4;
ld.global.nc.u8 %rs31, [%rd33];
cvt.u32.u16 %r222, %rs31;
and.b32 %r223, %r222, 255;
ld.global.nc.u8 %rs32, [%rd33+1];
cvt.u32.u16 %r224, %rs32;
prmt.b32 %r225, %r224, %r223, 30212;
ld.global.nc.u8 %rs33, [%rd33+2];
cvt.u32.u16 %r226, %rs33;
prmt.b32 %r227, %r226, %r225, 28756;
ld.global.nc.u8 %rs34, [%rd33+3];
cvt.u32.u16 %r228, %rs34;
prmt.b32 %r229, %r228, %r227, 1620;
and.b32 %r230, %r229, -2004318072;
shr.u32 %r231, %r230, 1;
or.b32 %r232, %r231, 839922192;
and.b32 %r233, %r225, 30583;
prmt.b32 %r234, %r141, %r140, %r233;
prmt.b32 %r235, %r144, %r143, %r233;
and.b32 %r236, %r232, 30292;
prmt.b32 %r237, %r234, %r235, %r236;
shr.u32 %r238, %r229, 16;
and.b32 %r239, %r238, 30583;
prmt.b32 %r240, %r141, %r140, %r239;
prmt.b32 %r241, %r144, %r143, %r239;
shr.u32 %r242, %r232, 16;
prmt.b32 %r243, %r240, %r241, %r242;
prmt.b32 %r95, %r237, %r243, %r154;
prmt.b32 %r99, %r237, %r243, %r155;

dp4a.s32.s32 %r94, %r95, %r62, %r114;


dp4a.s32.s32 %r98, %r99, %r66, %r94;

ld.global.nc.u8 %rs35, [%rd33+4];
cvt.u32.u16 %r244, %rs35;
and.b32 %r245, %r244, 255;
ld.global.nc.u8 %rs36, [%rd33+5];
cvt.u32.u16 %r246, %rs36;
prmt.b32 %r247, %r246, %r245, 30212;
ld.global.nc.u8 %rs37, [%rd33+6];
cvt.u32.u16 %r248, %rs37;
prmt.b32 %r249, %r248, %r247, 28756;
ld.global.nc.u8 %rs38, [%rd33+7];
cvt.u32.u16 %r250, %rs38;
prmt.b32 %r251, %r250, %r249, 1620;
and.b32 %r252, %r251, -2004318072;
shr.u32 %r253, %r252, 1;
or.b32 %r254, %r253, 839922192;
and.b32 %r255, %r247, 30583;
prmt.b32 %r256, %r141, %r140, %r255;
prmt.b32 %r257, %r144, %r143, %r255;
and.b32 %r258, %r254, 30292;
prmt.b32 %r259, %r256, %r257, %r258;
shr.u32 %r260, %r251, 16;
and.b32 %r261, %r260, 30583;
prmt.b32 %r262, %r141, %r140, %r261;
prmt.b32 %r263, %r144, %r143, %r261;
shr.u32 %r264, %r254, 16;
prmt.b32 %r265, %r262, %r263, %r264;
prmt.b32 %r103, %r259, %r265, %r154;
prmt.b32 %r107, %r259, %r265, %r155;

dp4a.s32.s32 %r102, %r103, %r70, %r98;


dp4a.s32.s32 %r106, %r107, %r74, %r102;

add.s64 %rd34, %rd32, %rd4;
ld.global.nc.u8 %rs39, [%rd34];
shl.b16 %rs41, %rs39, 7;
and.b16 %rs42, %rs41, 32640;
setp.eq.s16 %p7, %rs39, 255;
setp.eq.s16 %p8, %rs39, 0;
selp.b16 %rs43, 64, %rs42, %p8;
selp.b16 %rs3, 32767, %rs43, %p7;

{ mov.b32 %f29, {0,%rs3};}


mul.ftz.f32 %f39, %f29, 0f3F000000;

{.reg .f16 low,high;
mov.b32 {low,high},%r76;
cvt.f32.f16 %f30, low;}


mul.ftz.f32 %f40, %f39, %f30;
cvt.rn.f32.s32 %f41, %r106;
fma.rn.ftz.f32 %f110, %f40, %f41, %f110;
add.s64 %rd35, %rd33, %rd4;
ld.global.nc.u8 %rs44, [%rd35];
cvt.u32.u16 %r266, %rs44;
and.b32 %r267, %r266, 255;
ld.global.nc.u8 %rs45, [%rd35+1];
cvt.u32.u16 %r268, %rs45;
prmt.b32 %r269, %r268, %r267, 30212;
ld.global.nc.u8 %rs46, [%rd35+2];
cvt.u32.u16 %r270, %rs46;
prmt.b32 %r271, %r270, %r269, 28756;
ld.global.nc.u8 %rs47, [%rd35+3];
cvt.u32.u16 %r272, %rs47;
prmt.b32 %r273, %r272, %r271, 1620;
and.b32 %r274, %r273, -2004318072;
shr.u32 %r275, %r274, 1;
or.b32 %r276, %r275, 839922192;
and.b32 %r277, %r269, 30583;
prmt.b32 %r278, %r141, %r140, %r277;
prmt.b32 %r279, %r144, %r143, %r277;
and.b32 %r280, %r276, 30292;
prmt.b32 %r281, %r278, %r279, %r280;
shr.u32 %r282, %r273, 16;
and.b32 %r283, %r282, 30583;
prmt.b32 %r284, %r141, %r140, %r283;
prmt.b32 %r285, %r144, %r143, %r283;
shr.u32 %r286, %r276, 16;
prmt.b32 %r287, %r284, %r285, %r286;
prmt.b32 %r112, %r281, %r287, %r154;
prmt.b32 %r116, %r281, %r287, %r155;

dp4a.s32.s32 %r111, %r112, %r62, %r114;


dp4a.s32.s32 %r115, %r116, %r66, %r111;

ld.global.nc.u8 %rs48, [%rd35+4];
cvt.u32.u16 %r288, %rs48;
and.b32 %r289, %r288, 255;
ld.global.nc.u8 %rs49, [%rd35+5];
cvt.u32.u16 %r290, %rs49;
prmt.b32 %r291, %r290, %r289, 30212;
ld.global.nc.u8 %rs50, [%rd35+6];
cvt.u32.u16 %r292, %rs50;
prmt.b32 %r293, %r292, %r291, 28756;
ld.global.nc.u8 %rs51, [%rd35+7];
cvt.u32.u16 %r294, %rs51;
prmt.b32 %r295, %r294, %r293, 1620;
and.b32 %r296, %r295, -2004318072;
shr.u32 %r297, %r296, 1;
or.b32 %r298, %r297, 839922192;
and.b32 %r299, %r291, 30583;
prmt.b32 %r300, %r141, %r140, %r299;
prmt.b32 %r301, %r144, %r143, %r299;
and.b32 %r302, %r298, 30292;
prmt.b32 %r303, %r300, %r301, %r302;
shr.u32 %r304, %r295, 16;
and.b32 %r305, %r304, 30583;
prmt.b32 %r306, %r141, %r140, %r305;
prmt.b32 %r307, %r144, %r143, %r305;
shr.u32 %r308, %r298, 16;
prmt.b32 %r309, %r306, %r307, %r308;
prmt.b32 %r120, %r303, %r309, %r154;
prmt.b32 %r124, %r303, %r309, %r155;

dp4a.s32.s32 %r119, %r120, %r70, %r115;


dp4a.s32.s32 %r123, %r124, %r74, %r119;

add.s64 %rd36, %rd34, %rd4;
ld.global.nc.u8 %rs52, [%rd36];
shl.b16 %rs54, %rs52, 7;
and.b16 %rs55, %rs54, 32640;
setp.eq.s16 %p9, %rs52, 255;
setp.eq.s16 %p10, %rs52, 0;
selp.b16 %rs56, 64, %rs55, %p10;
selp.b16 %rs4, 32767, %rs56, %p9;

{ mov.b32 %f31, {0,%rs4};}


mul.ftz.f32 %f42, %f31, 0f3F000000;

{.reg .f16 low,high;
mov.b32 {low,high},%r76;
cvt.f32.f16 %f32, low;}


mul.ftz.f32 %f43, %f42, %f32;
cvt.rn.f32.s32 %f44, %r123;
fma.rn.ftz.f32 %f109, %f43, %f44, %f109;
add.s64 %rd39, %rd39, 1088;
add.s32 %r395, %r395, 64;
setp.lt.s32 %p11, %r395, %r44;
@%p11 bra $L__BB104_5;

$L__BB104_6:
setp.eq.s32 %p12, %r1, 0;
@%p12 bra $L__BB104_8;

shl.b32 %r311, %r1, 9;
mov.u32 %r312, _ZZ13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r313, %r312, %r311;
shl.b32 %r314, %r2, 2;
add.s32 %r315, %r313, %r314;
st.shared.f32 [%r315+-512], %f112;
st.shared.f32 [%r315+-384], %f111;
st.shared.f32 [%r315+-256], %f110;
st.shared.f32 [%r315+-128], %f109;

$L__BB104_8:
setp.ne.s32 %p13, %r1, 0;
bar.sync 0;
@%p13 bra $L__BB104_17;

mov.u32 %r316, %ctaid.z;
mov.u32 %r317, %ctaid.y;
mad.lo.s32 %r318, %r317, %r27, %r4;
mad.lo.s32 %r319, %r316, %r33, %r318;
shl.b32 %r320, %r2, 2;
mov.u32 %r321, _ZZ13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r14, %r321, %r320;
mov.u32 %r322, 2;
ld.shared.f32 %f45, [%r14];
add.ftz.f32 %f46, %f45, %f112;
ld.shared.f32 %f47, [%r14+512];
add.ftz.f32 %f48, %f47, %f46;
ld.shared.f32 %f49, [%r14+1024];
add.ftz.f32 %f50, %f49, %f48;
mov.b32 %r323, %f50;
mov.u32 %r324, 31;
mov.u32 %r325, 16;
mov.u32 %r326, -1;
shfl.sync.bfly.b32 %r327|%p14, %r323, %r325, %r324, %r326;
mov.b32 %f51, %r327;
add.ftz.f32 %f52, %f50, %f51;
mov.b32 %r328, %f52;
mov.u32 %r329, 8;
shfl.sync.bfly.b32 %r330|%p15, %r328, %r329, %r324, %r326;
mov.b32 %f53, %r330;
add.ftz.f32 %f54, %f52, %f53;
mov.b32 %r331, %f54;
mov.u32 %r332, 4;
shfl.sync.bfly.b32 %r333|%p16, %r331, %r332, %r324, %r326;
mov.b32 %f55, %r333;
add.ftz.f32 %f56, %f54, %f55;
mov.b32 %r334, %f56;
shfl.sync.bfly.b32 %r335|%p17, %r334, %r322, %r324, %r326;
mov.b32 %f57, %r335;
add.ftz.f32 %f58, %f56, %f57;
mov.b32 %r336, %f58;
mov.u32 %r337, 1;
shfl.sync.bfly.b32 %r338|%p18, %r336, %r337, %r324, %r326;
mov.b32 %f59, %r338;
add.ftz.f32 %f13, %f58, %f59;
setp.ge.u32 %p19, %r4, %r21;
setp.ne.s32 %p20, %r2, 0;
cvta.to.global.u64 %rd37, %rd14;
mul.wide.u32 %rd38, %r319, 4;
add.s64 %rd10, %rd37, %rd38;
or.pred %p21, %p20, %p19;
@%p21 bra $L__BB104_11;

st.global.f32 [%rd10], %f13;

$L__BB104_11:
ld.shared.f32 %f60, [%r14+128];
add.ftz.f32 %f61, %f60, %f111;
ld.shared.f32 %f62, [%r14+640];
add.ftz.f32 %f63, %f62, %f61;
ld.shared.f32 %f64, [%r14+1152];
add.ftz.f32 %f65, %f64, %f63;
mov.b32 %r339, %f65;
shfl.sync.bfly.b32 %r343|%p22, %r339, %r325, %r324, %r326;
mov.b32 %f66, %r343;
add.ftz.f32 %f67, %f65, %f66;
mov.b32 %r344, %f67;
shfl.sync.bfly.b32 %r346|%p23, %r344, %r329, %r324, %r326;
mov.b32 %f68, %r346;
add.ftz.f32 %f69, %f67, %f68;
mov.b32 %r347, %f69;
shfl.sync.bfly.b32 %r349|%p24, %r347, %r332, %r324, %r326;
mov.b32 %f70, %r349;
add.ftz.f32 %f71, %f69, %f70;
mov.b32 %r350, %f71;
shfl.sync.bfly.b32 %r352|%p25, %r350, %r322, %r324, %r326;
mov.b32 %f72, %r352;
add.ftz.f32 %f73, %f71, %f72;
mov.b32 %r353, %f73;
shfl.sync.bfly.b32 %r355|%p26, %r353, %r337, %r324, %r326;
mov.b32 %f74, %r355;
add.ftz.f32 %f14, %f73, %f74;
add.s32 %r356, %r4, 1;
setp.ge.u32 %p27, %r356, %r21;
setp.ne.s32 %p28, %r2, 1;
or.pred %p29, %p28, %p27;
@%p29 bra $L__BB104_13;

st.global.f32 [%rd10+4], %f14;

$L__BB104_13:
ld.shared.f32 %f75, [%r14+256];
add.ftz.f32 %f76, %f75, %f110;
ld.shared.f32 %f77, [%r14+768];
add.ftz.f32 %f78, %f77, %f76;
ld.shared.f32 %f79, [%r14+1280];
add.ftz.f32 %f80, %f79, %f78;
mov.b32 %r357, %f80;
mov.u32 %r358, 31;
mov.u32 %r359, 16;
mov.u32 %r360, -1;
shfl.sync.bfly.b32 %r361|%p30, %r357, %r359, %r358, %r360;
mov.b32 %f81, %r361;
add.ftz.f32 %f82, %f80, %f81;
mov.b32 %r362, %f82;
mov.u32 %r363, 8;
shfl.sync.bfly.b32 %r364|%p31, %r362, %r363, %r358, %r360;
mov.b32 %f83, %r364;
add.ftz.f32 %f84, %f82, %f83;
mov.b32 %r365, %f84;
mov.u32 %r366, 4;
shfl.sync.bfly.b32 %r367|%p32, %r365, %r366, %r358, %r360;
mov.b32 %f85, %r367;
add.ftz.f32 %f86, %f84, %f85;
mov.b32 %r368, %f86;
mov.u32 %r369, 2;
shfl.sync.bfly.b32 %r370|%p33, %r368, %r369, %r358, %r360;
mov.b32 %f87, %r370;
add.ftz.f32 %f88, %f86, %f87;
mov.b32 %r371, %f88;
mov.u32 %r372, 1;
shfl.sync.bfly.b32 %r373|%p34, %r371, %r372, %r358, %r360;
mov.b32 %f89, %r373;
add.ftz.f32 %f15, %f88, %f89;
add.s32 %r374, %r4, 2;
setp.ge.u32 %p35, %r374, %r21;
setp.ne.s32 %p36, %r2, 2;
or.pred %p37, %p36, %p35;
@%p37 bra $L__BB104_15;

st.global.f32 [%rd10+8], %f15;

$L__BB104_15:
ld.shared.f32 %f90, [%r14+384];
add.ftz.f32 %f91, %f90, %f109;
ld.shared.f32 %f92, [%r14+896];
add.ftz.f32 %f93, %f92, %f91;
ld.shared.f32 %f94, [%r14+1408];
add.ftz.f32 %f95, %f94, %f93;
mov.b32 %r375, %f95;
shfl.sync.bfly.b32 %r379|%p38, %r375, %r359, %r358, %r360;
mov.b32 %f96, %r379;
add.ftz.f32 %f97, %f95, %f96;
mov.b32 %r380, %f97;
shfl.sync.bfly.b32 %r382|%p39, %r380, %r363, %r358, %r360;
mov.b32 %f98, %r382;
add.ftz.f32 %f99, %f97, %f98;
mov.b32 %r383, %f99;
shfl.sync.bfly.b32 %r385|%p40, %r383, %r366, %r358, %r360;
mov.b32 %f100, %r385;
add.ftz.f32 %f101, %f99, %f100;
mov.b32 %r386, %f101;
shfl.sync.bfly.b32 %r388|%p41, %r386, %r369, %r358, %r360;
mov.b32 %f102, %r388;
add.ftz.f32 %f103, %f101, %f102;
mov.b32 %r389, %f103;
shfl.sync.bfly.b32 %r391|%p42, %r389, %r372, %r358, %r360;
mov.b32 %f104, %r391;
add.ftz.f32 %f16, %f103, %f104;
add.s32 %r392, %r4, 3;
setp.ge.u32 %p43, %r392, %r21;
setp.ne.s32 %p44, %r2, 3;
or.pred %p45, %p44, %p43;
@%p45 bra $L__BB104_17;

st.global.f32 [%rd10+12], %f16;

$L__BB104_17:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<38>;
.reg .b16 %rs<19>;
.reg .f32 %f<122>;
.reg .b32 %r<288>;
.reg .b64 %rd<70>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[384];

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate[384];

ld.param.u64 %rd14, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd22, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
ld.param.f32 %f31, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+16];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+8];
ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3];
setp.eq.s64 %p2, %rd16, 0;
@%p2 bra $L__BB105_2;

cvta.to.global.u64 %rd23, %rd16;
mov.u32 %r39, %ctaid.y;
mul.wide.u32 %rd24, %r39, 4;
add.s64 %rd25, %rd23, %rd24;
ld.global.nc.u32 %r285, [%rd25];
mul.hi.u32 %r40, %r39, %r23;
add.s32 %r41, %r40, %r39;
shr.u32 %r42, %r41, %r24;
mul.lo.s32 %r43, %r42, %r25;
sub.s32 %r286, %r39, %r43;
bra.uni $L__BB105_3;

$L__BB105_2:
mov.u32 %r286, %ctaid.y;
mul.hi.u32 %r44, %r286, %r27;
add.s32 %r45, %r44, %r286;
shr.u32 %r285, %r45, %r28;

$L__BB105_3:
mov.u32 %r46, %tid.y;
mov.u32 %r47, %tid.x;
or.b32 %r48, %r47, %r46;
setp.ne.s32 %p3, %r48, 0;
mov.f32 %f115, 0f00000000;
mov.f32 %f111, %f115;
mov.f32 %f112, %f115;
@%p3 bra $L__BB105_8;

setp.eq.s64 %p4, %rd17, 0;
mov.f32 %f111, 0f00000000;
mov.f32 %f112, %f111;
@%p4 bra $L__BB105_6;

cvta.to.global.u64 %rd26, %rd17;
mov.u32 %r49, %ctaid.z;
mul.lo.s32 %r50, %r49, %r38;
cvt.u64.u32 %rd27, %r50;
mov.u32 %r51, %ctaid.y;
selp.b32 %r52, %r51, %r285, %p2;
mul.lo.s32 %r53, %r52, %r32;
cvt.u64.u32 %rd28, %r53;
mov.u32 %r54, %ctaid.x;
cvt.s64.s32 %rd29, %r54;
cvt.u64.u32 %rd30, %r47;
add.s64 %rd31, %rd30, %rd29;
add.s64 %rd32, %rd31, %rd27;
add.s64 %rd33, %rd32, %rd28;
shl.b64 %rd34, %rd33, 2;
add.s64 %rd35, %rd26, %rd34;
ld.global.f32 %f112, [%rd35];

$L__BB105_6:
setp.eq.s64 %p6, %rd19, 0;
setp.eq.s64 %p7, %rd18, 0;
or.pred %p8, %p6, %p7;
@%p8 bra $L__BB105_8;

cvta.to.global.u64 %rd36, %rd19;
mov.u32 %r56, %ctaid.z;
mul.lo.s32 %r57, %r56, %r38;
cvt.u64.u32 %rd37, %r57;
mov.u32 %r58, %ctaid.y;
selp.b32 %r59, %r58, %r285, %p2;
mul.lo.s32 %r60, %r59, %r32;
cvt.u64.u32 %rd38, %r60;
mov.u32 %r61, %ctaid.x;
cvt.s64.s32 %rd39, %r61;
cvt.u64.u32 %rd40, %r47;
add.s64 %rd41, %rd40, %rd39;
add.s64 %rd42, %rd41, %rd37;
add.s64 %rd43, %rd42, %rd38;
shl.b64 %rd44, %rd43, 2;
add.s64 %rd45, %rd36, %rd44;
ld.global.f32 %f111, [%rd45];

$L__BB105_8:
shl.b32 %r64, %r46, 5;
add.s32 %r66, %r64, %r47;
shr.u32 %r67, %r66, 31;
add.s32 %r68, %r66, %r67;
shr.s32 %r287, %r68, 1;
shr.u32 %r69, %r22, 5;
setp.ge.s32 %p10, %r287, %r69;
mov.f32 %f117, %f115;
@%p10 bra $L__BB105_13;

mov.u32 %r70, %ctaid.z;
and.b32 %r77, %r68, -2;
sub.s32 %r10, %r66, %r77;
shl.b32 %r78, %r10, 3;
cvt.s64.s32 %rd46, %r78;
mul.hi.u32 %r79, %r70, %r33;
add.s32 %r80, %r79, %r70;
shr.u32 %r81, %r80, %r34;
mad.lo.s32 %r82, %r81, %r36, %r287;
mad.lo.s32 %r83, %r285, %r30, %r82;
mov.u32 %r84, %ctaid.x;
mad.lo.s32 %r85, %r84, %r26, %r83;
mul.wide.s32 %rd47, %r85, 17;
cvta.to.global.u64 %rd48, %rd18;
add.s64 %rd69, %rd48, %rd47;
cvta.to.global.u64 %rd49, %rd14;
add.s64 %rd50, %rd49, %rd47;
add.s64 %rd68, %rd50, 4;
add.s64 %rd51, %rd47, %rd46;
add.s64 %rd52, %rd49, %rd51;
add.s64 %rd67, %rd52, 4;
mul.lo.s32 %r86, %r286, %r31;
cvt.u64.u32 %rd53, %r86;
mul.lo.s32 %r87, %r70, %r37;
cvt.u64.u32 %rd54, %r87;
add.s64 %rd7, %rd53, %rd54;
mov.f32 %f117, 0f00000000;
cvta.to.global.u64 %rd55, %rd15;
mov.f32 %f115, %f117;

$L__BB105_10:
cvt.u64.u32 %rd56, %r287;
add.s64 %rd57, %rd7, %rd56;
mul.lo.s64 %rd58, %rd57, 36;
add.s64 %rd59, %rd55, %rd58;
shl.b32 %r105, %r10, 1;
mul.wide.s32 %rd60, %r105, 4;
add.s64 %rd61, %rd59, %rd60;
ld.global.nc.u8 %rs2, [%rd67+-3];
cvt.u32.u16 %r106, %rs2;
and.b32 %r107, %r106, 255;
ld.global.nc.u8 %rs3, [%rd67+-2];
cvt.u32.u16 %r108, %rs3;
prmt.b32 %r109, %r108, %r107, 30212;
ld.global.nc.u8 %rs4, [%rd67+-1];
cvt.u32.u16 %r110, %rs4;
prmt.b32 %r111, %r110, %r109, 28756;
ld.global.nc.u8 %rs5, [%rd67];
cvt.u32.u16 %r112, %rs5;
prmt.b32 %r113, %r112, %r111, 1620;
and.b32 %r114, %r113, -2004318072;
shr.u32 %r115, %r114, 1;
or.b32 %r116, %r115, 839922192;
and.b32 %r117, %r109, 30583;
mov.u32 %r118, 201852420;
mov.u32 %r119, 50462976;
prmt.b32 %r120, %r119, %r118, %r117;
mov.u32 %r121, -185009412;
mov.u32 %r122, -33620224;
prmt.b32 %r123, %r122, %r121, %r117;
and.b32 %r124, %r116, 30292;
prmt.b32 %r125, %r120, %r123, %r124;
shr.u32 %r126, %r113, 16;
and.b32 %r127, %r126, 30583;
prmt.b32 %r128, %r119, %r118, %r127;
prmt.b32 %r129, %r122, %r121, %r127;
shr.u32 %r130, %r116, 16;
prmt.b32 %r131, %r128, %r129, %r130;
mov.u32 %r132, 25632;
prmt.b32 %r89, %r125, %r131, %r132;
mov.u32 %r133, 30001;
prmt.b32 %r93, %r125, %r131, %r133;
ld.global.nc.u32 %r12, [%rd61+4];
mov.u32 %r91, 0;

dp4a.s32.s32 %r88, %r89, %r12, %r91;

ld.global.nc.u32 %r13, [%rd61+20];

dp4a.s32.s32 %r92, %r93, %r13, %r88;

ld.global.nc.u8 %rs6, [%rd67+1];
cvt.u32.u16 %r134, %rs6;
and.b32 %r135, %r134, 255;
ld.global.nc.u8 %rs7, [%rd67+2];
cvt.u32.u16 %r136, %rs7;
prmt.b32 %r137, %r136, %r135, 30212;
ld.global.nc.u8 %rs8, [%rd67+3];
cvt.u32.u16 %r138, %rs8;
prmt.b32 %r139, %r138, %r137, 28756;
ld.global.nc.u8 %rs9, [%rd67+4];
cvt.u32.u16 %r140, %rs9;
prmt.b32 %r141, %r140, %r139, 1620;
and.b32 %r142, %r141, -2004318072;
shr.u32 %r143, %r142, 1;
or.b32 %r144, %r143, 839922192;
and.b32 %r145, %r137, 30583;
prmt.b32 %r146, %r119, %r118, %r145;
prmt.b32 %r147, %r122, %r121, %r145;
and.b32 %r148, %r144, 30292;
prmt.b32 %r149, %r146, %r147, %r148;
shr.u32 %r150, %r141, 16;
and.b32 %r151, %r150, 30583;
prmt.b32 %r152, %r119, %r118, %r151;
prmt.b32 %r153, %r122, %r121, %r151;
shr.u32 %r154, %r144, 16;
prmt.b32 %r155, %r152, %r153, %r154;
prmt.b32 %r97, %r149, %r155, %r132;
prmt.b32 %r101, %r149, %r155, %r133;
ld.global.nc.u32 %r14, [%rd61+8];

dp4a.s32.s32 %r96, %r97, %r14, %r92;

ld.global.nc.u32 %r15, [%rd61+24];

dp4a.s32.s32 %r100, %r101, %r15, %r96;

ld.global.nc.u8 %rs10, [%rd68+-4];
shl.b16 %rs12, %rs10, 7;
and.b16 %rs13, %rs12, 32640;
setp.eq.s16 %p11, %rs10, 255;
setp.eq.s16 %p12, %rs10, 0;
selp.b16 %rs14, 64, %rs13, %p12;
selp.b16 %rs1, 32767, %rs14, %p11;

{ mov.b32 %f40, {0,%rs1};}


mul.ftz.f32 %f42, %f40, 0f3F000000;
ld.global.nc.u32 %r16, [%rd59];

{.reg .f16 low,high;
mov.b32 {low,high},%r16;
cvt.f32.f16 %f41, low;}


mul.ftz.f32 %f43, %f42, %f41;
cvt.rn.f32.s32 %f44, %r100;
fma.rn.ftz.f32 %f117, %f43, %f44, %f117;
setp.eq.s64 %p13, %rd18, 0;
@%p13 bra $L__BB105_12;

add.s64 %rd63, %rd69, %rd46;
ld.global.u8 %r174, [%rd63+1];
ld.global.u8 %r175, [%rd63+2];
prmt.b32 %r176, %r175, %r174, 30212;
ld.global.u8 %r177, [%rd63+3];
prmt.b32 %r178, %r177, %r176, 28756;
ld.global.u8 %r179, [%rd63+4];
prmt.b32 %r180, %r179, %r178, 1620;
and.b32 %r181, %r180, -2004318072;
shr.u32 %r182, %r181, 1;
or.b32 %r183, %r182, 839922192;
and.b32 %r184, %r176, 30583;
prmt.b32 %r187, %r119, %r118, %r184;
prmt.b32 %r190, %r122, %r121, %r184;
and.b32 %r191, %r183, 30292;
prmt.b32 %r192, %r187, %r190, %r191;
shr.u32 %r193, %r180, 16;
and.b32 %r194, %r193, 30583;
prmt.b32 %r195, %r119, %r118, %r194;
prmt.b32 %r196, %r122, %r121, %r194;
shr.u32 %r197, %r183, 16;
prmt.b32 %r198, %r195, %r196, %r197;
prmt.b32 %r157, %r192, %r198, %r132;
prmt.b32 %r161, %r192, %r198, %r133;

dp4a.s32.s32 %r156, %r157, %r12, %r91;


dp4a.s32.s32 %r160, %r161, %r13, %r156;

ld.global.u8 %r201, [%rd63+5];
ld.global.u8 %r202, [%rd63+6];
prmt.b32 %r203, %r202, %r201, 30212;
ld.global.u8 %r204, [%rd63+7];
prmt.b32 %r205, %r204, %r203, 28756;
ld.global.u8 %r206, [%rd63+8];
prmt.b32 %r207, %r206, %r205, 1620;
and.b32 %r208, %r207, -2004318072;
shr.u32 %r209, %r208, 1;
or.b32 %r210, %r209, 839922192;
and.b32 %r211, %r203, 30583;
prmt.b32 %r212, %r119, %r118, %r211;
prmt.b32 %r213, %r122, %r121, %r211;
and.b32 %r214, %r210, 30292;
prmt.b32 %r215, %r212, %r213, %r214;
shr.u32 %r216, %r207, 16;
and.b32 %r217, %r216, 30583;
prmt.b32 %r218, %r119, %r118, %r217;
prmt.b32 %r219, %r122, %r121, %r217;
shr.u32 %r220, %r210, 16;
prmt.b32 %r221, %r218, %r219, %r220;
prmt.b32 %r165, %r215, %r221, %r132;
prmt.b32 %r169, %r215, %r221, %r133;

dp4a.s32.s32 %r164, %r165, %r14, %r160;


dp4a.s32.s32 %r168, %r169, %r15, %r164;

ld.global.u8 %rs16, [%rd69];
shl.b16 %rs17, %rs16, 7;
setp.eq.s16 %p14, %rs16, 255;
setp.eq.s16 %p15, %rs16, 0;
selp.b16 %rs18, 64, %rs17, %p15;
selp.b16 %rs15, 32767, %rs18, %p14;

{ mov.b32 %f45, {0,%rs15};}


mul.ftz.f32 %f47, %f45, 0f3F000000;

{.reg .f16 low,high;
mov.b32 {low,high},%r16;
cvt.f32.f16 %f46, low;}


mul.ftz.f32 %f48, %f47, %f46;
cvt.rn.f32.s32 %f49, %r168;
fma.rn.ftz.f32 %f115, %f48, %f49, %f115;

$L__BB105_12:
add.s64 %rd69, %rd69, 1088;
add.s64 %rd68, %rd68, 1088;
add.s64 %rd67, %rd67, 1088;
add.s32 %r287, %r287, 64;
setp.lt.s32 %p16, %r287, %r69;
@%p16 bra $L__BB105_10;

$L__BB105_13:
setp.eq.s32 %p17, %r46, 0;
@%p17 bra $L__BB105_16;

setp.eq.s64 %p18, %rd18, 0;
add.s32 %r18, %r46, -1;
shl.b32 %r226, %r46, 7;
mov.u32 %r227, _ZZ13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r228, %r227, %r226;
shl.b32 %r229, %r47, 2;
add.s32 %r230, %r228, %r229;
st.shared.f32 [%r230+-128], %f117;
@%p18 bra $L__BB105_16;

shl.b32 %r232, %r18, 7;
mov.u32 %r233, _ZZ13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r234, %r233, %r232;
add.s32 %r236, %r234, %r229;
st.shared.f32 [%r236], %f115;

$L__BB105_16:
setp.ne.s32 %p19, %r46, 0;
bar.sync 0;
@%p19 bra $L__BB105_35;

setp.ne.s64 %p20, %rd18, 0;
shl.b32 %r239, %r47, 2;
mov.u32 %r240, _ZZ13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r19, %r240, %r239;
ld.shared.f32 %f50, [%r19];
add.ftz.f32 %f13, %f117, %f50;
@%p20 bra $L__BB105_19;
bra.uni $L__BB105_18;

$L__BB105_19:
mov.u32 %r243, _ZZ13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r244, %r243, %r239;
ld.shared.f32 %f54, [%r244];
add.ftz.f32 %f55, %f115, %f54;
ld.shared.f32 %f56, [%r19+128];
add.ftz.f32 %f57, %f13, %f56;
ld.shared.f32 %f58, [%r244+128];
add.ftz.f32 %f59, %f55, %f58;
ld.shared.f32 %f60, [%r19+256];
add.ftz.f32 %f119, %f57, %f60;
ld.shared.f32 %f61, [%r244+256];
add.ftz.f32 %f115, %f59, %f61;
bra.uni $L__BB105_20;

$L__BB105_18:
ld.shared.f32 %f51, [%r19+128];
add.ftz.f32 %f52, %f13, %f51;
ld.shared.f32 %f53, [%r19+256];
add.ftz.f32 %f119, %f52, %f53;

$L__BB105_20:
mov.b32 %r245, %f119;
mov.u32 %r246, 31;
mov.u32 %r247, 16;
mov.u32 %r248, -1;
shfl.sync.bfly.b32 %r249|%p21, %r245, %r247, %r246, %r248;
mov.b32 %f62, %r249;
add.ftz.f32 %f63, %f119, %f62;
mov.b32 %r250, %f63;
mov.u32 %r251, 8;
shfl.sync.bfly.b32 %r252|%p22, %r250, %r251, %r246, %r248;
mov.b32 %f64, %r252;
add.ftz.f32 %f65, %f63, %f64;
mov.b32 %r253, %f65;
mov.u32 %r254, 4;
shfl.sync.bfly.b32 %r255|%p23, %r253, %r254, %r246, %r248;
mov.b32 %f66, %r255;
add.ftz.f32 %f67, %f65, %f66;
mov.b32 %r256, %f67;
mov.u32 %r257, 2;
shfl.sync.bfly.b32 %r258|%p24, %r256, %r257, %r246, %r248;
mov.b32 %f68, %r258;
add.ftz.f32 %f19, %f67, %f68;
mov.b32 %r259, %f19;
mov.u32 %r260, 1;
shfl.sync.bfly.b32 %r20|%p1, %r259, %r260, %r246, %r248;
setp.eq.s64 %p25, %rd18, 0;
@%p25 bra $L__BB105_22;

mov.b32 %r261, %f115;
shfl.sync.bfly.b32 %r265|%p26, %r261, %r247, %r246, %r248;
mov.b32 %f69, %r265;
add.ftz.f32 %f70, %f115, %f69;
mov.b32 %r266, %f70;
shfl.sync.bfly.b32 %r268|%p27, %r266, %r251, %r246, %r248;
mov.b32 %f71, %r268;
add.ftz.f32 %f72, %f70, %f71;
mov.b32 %r269, %f72;
shfl.sync.bfly.b32 %r271|%p28, %r269, %r254, %r246, %r248;
mov.b32 %f73, %r271;
add.ftz.f32 %f74, %f72, %f73;
mov.b32 %r272, %f74;
shfl.sync.bfly.b32 %r274|%p29, %r272, %r257, %r246, %r248;
mov.b32 %f75, %r274;
add.ftz.f32 %f76, %f74, %f75;
mov.b32 %r275, %f76;
shfl.sync.bfly.b32 %r277|%p30, %r275, %r260, %r246, %r248;
mov.b32 %f77, %r277;
add.ftz.f32 %f115, %f76, %f77;

$L__BB105_22:
setp.ne.s32 %p31, %r47, 0;
mov.b32 %f78, %r20;
add.ftz.f32 %f22, %f19, %f78;
@%p31 bra $L__BB105_35;

add.ftz.f32 %f121, %f112, %f22;
@%p25 bra $L__BB105_34;

add.ftz.f32 %f24, %f111, %f115;
setp.gt.s32 %p33, %r21, 2;
@%p33 bra $L__BB105_28;

setp.eq.s32 %p36, %r21, 1;
@%p36 bra $L__BB105_32;

setp.eq.s32 %p37, %r21, 2;
@%p37 bra $L__BB105_27;
bra.uni $L__BB105_33;

$L__BB105_27:
mul.ftz.f32 %f98, %f24, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f99, %f98;
add.ftz.f32 %f100, %f99, 0f3F800000;
div.approx.ftz.f32 %f101, %f24, %f100;
mul.ftz.f32 %f121, %f121, %f101;
bra.uni $L__BB105_34;

$L__BB105_28:
setp.eq.s32 %p34, %r21, 3;
@%p34 bra $L__BB105_31;

setp.ne.s32 %p35, %r21, 6;
@%p35 bra $L__BB105_33;

min.ftz.f32 %f79, %f24, %f31;
neg.ftz.f32 %f80, %f31;
min.ftz.f32 %f81, %f121, %f31;
max.ftz.f32 %f82, %f81, %f80;
mul.ftz.f32 %f83, %f79, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f84, %f83;
add.ftz.f32 %f85, %f84, 0f3F800000;
div.approx.ftz.f32 %f86, %f79, %f85;
mul.ftz.f32 %f121, %f82, %f86;
bra.uni $L__BB105_34;

$L__BB105_32:
mul.ftz.f32 %f102, %f24, 0f3F000000;
mul.ftz.f32 %f103, %f24, 0f3D372713;
fma.rn.ftz.f32 %f104, %f24, %f103, 0f3F800000;
mul.ftz.f32 %f105, %f24, 0f3F4C422A;
mul.ftz.f32 %f106, %f105, %f104;
tanh.approx.f32 %f107, %f106;
add.ftz.f32 %f108, %f107, 0f3F800000;
mul.ftz.f32 %f109, %f102, %f108;
mul.ftz.f32 %f121, %f121, %f109;
bra.uni $L__BB105_34;

$L__BB105_31:
mov.f32 %f87, 0f40E00000;
min.ftz.f32 %f88, %f24, %f87;
min.ftz.f32 %f89, %f121, %f87;
mov.f32 %f90, 0fC0E00000;
max.ftz.f32 %f91, %f89, %f90;
mul.ftz.f32 %f92, %f88, 0f3FD9DB23;
mul.ftz.f32 %f93, %f92, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f94, %f93;
add.ftz.f32 %f95, %f94, 0f3F800000;
div.approx.ftz.f32 %f96, %f88, %f95;
add.ftz.f32 %f97, %f91, 0f3F800000;
mul.ftz.f32 %f121, %f97, %f96;
bra.uni $L__BB105_34;

$L__BB105_33:
mul.ftz.f32 %f121, %f121, %f24;

$L__BB105_34:
cvta.to.global.u64 %rd64, %rd22;
mov.u32 %r280, %ctaid.z;
mov.u32 %r281, %ctaid.y;
mov.u32 %r282, %ctaid.x;
mad.lo.s32 %r283, %r281, %r32, %r282;
mad.lo.s32 %r284, %r280, %r38, %r283;
mul.wide.u32 %rd65, %r284, 4;
add.s64 %rd66, %rd64, %rd65;
st.global.f32 [%rd66], %f121;

$L__BB105_35:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<14>;
.reg .b16 %rs<15>;
.reg .f32 %f<29>;
.reg .b32 %r<157>;
.reg .b64 %rd<36>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[384];

ld.param.u64 %rd10, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd11, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd12, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd13, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r15, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r19, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r18, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r17, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r16, [_Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
mov.u32 %r1, %tid.y;
shl.b32 %r32, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r32, %r2;
mov.u32 %r155, %ctaid.y;
setp.eq.s64 %p2, %rd12, 0;
@%p2 bra $L__BB106_2;

cvta.to.global.u64 %rd14, %rd12;
mul.wide.u32 %rd15, %r155, 4;
add.s64 %rd16, %rd14, %rd15;
ld.global.nc.u32 %r154, [%rd16];
mul.hi.u32 %r33, %r155, %r16;
add.s32 %r34, %r33, %r155;
shr.u32 %r35, %r34, %r17;
mul.lo.s32 %r36, %r35, %r18;
sub.s32 %r155, %r155, %r36;
bra.uni $L__BB106_3;

$L__BB106_2:
mul.hi.u32 %r37, %r155, %r20;
add.s32 %r38, %r37, %r155;
shr.u32 %r154, %r38, %r21;

$L__BB106_3:
shr.u32 %r39, %r3, 31;
add.s32 %r40, %r3, %r39;
shr.s32 %r156, %r40, 1;
shr.u32 %r41, %r15, 5;
setp.ge.s32 %p3, %r156, %r41;
mov.f32 %f28, 0f00000000;
@%p3 bra $L__BB106_6;

mov.u32 %r42, %ctaid.z;
and.b32 %r45, %r40, -2;
sub.s32 %r46, %r3, %r45;
shl.b32 %r47, %r46, 1;
cvt.s64.s32 %rd1, %r47;
shl.b32 %r48, %r46, 3;
cvt.s64.s32 %rd17, %r48;
mul.hi.u32 %r49, %r42, %r26;
add.s32 %r50, %r49, %r42;
shr.u32 %r51, %r50, %r27;
mad.lo.s32 %r52, %r51, %r29, %r156;
mad.lo.s32 %r53, %r154, %r23, %r52;
mov.u32 %r54, %ctaid.x;
mad.lo.s32 %r55, %r54, %r19, %r53;
mul.wide.s32 %rd18, %r55, 17;
cvta.to.global.u64 %rd19, %rd10;
add.s64 %rd20, %rd19, %rd18;
add.s64 %rd35, %rd20, 8;
add.s64 %rd21, %rd18, %rd17;
add.s64 %rd22, %rd19, %rd21;
add.s64 %rd34, %rd22, 4;
cvta.to.global.u64 %rd4, %rd11;
mul.lo.s32 %r56, %r155, %r24;
cvt.u64.u32 %rd23, %r56;
mul.lo.s32 %r57, %r42, %r30;
cvt.u64.u32 %rd24, %r57;
add.s64 %rd5, %rd23, %rd24;
mov.f32 %f28, 0f00000000;
shl.b64 %rd29, %rd1, 2;

$L__BB106_5:
cvt.u64.u32 %rd25, %r156;
add.s64 %rd26, %rd5, %rd25;
mul.lo.s64 %rd27, %rd26, 36;
add.s64 %rd28, %rd4, %rd27;
add.s64 %rd30, %rd28, %rd29;
ld.global.nc.u8 %rs2, [%rd34+-3];
cvt.u32.u16 %r75, %rs2;
and.b32 %r76, %r75, 255;
ld.global.nc.u8 %rs3, [%rd34+-2];
cvt.u32.u16 %r77, %rs3;
prmt.b32 %r78, %r77, %r76, 30212;
ld.global.nc.u8 %rs4, [%rd34+-1];
cvt.u32.u16 %r79, %rs4;
prmt.b32 %r80, %r79, %r78, 28756;
ld.global.nc.u8 %rs5, [%rd34];
cvt.u32.u16 %r81, %rs5;
prmt.b32 %r82, %r81, %r80, 1620;
and.b32 %r83, %r82, -2004318072;
shr.u32 %r84, %r83, 1;
or.b32 %r85, %r84, 839922192;
and.b32 %r86, %r78, 30583;
mov.u32 %r87, 201852420;
mov.u32 %r88, 50462976;
prmt.b32 %r89, %r88, %r87, %r86;
mov.u32 %r90, -185009412;
mov.u32 %r91, -33620224;
prmt.b32 %r92, %r91, %r90, %r86;
and.b32 %r93, %r85, 30292;
prmt.b32 %r94, %r89, %r92, %r93;
shr.u32 %r95, %r82, 16;
and.b32 %r96, %r95, 30583;
prmt.b32 %r97, %r88, %r87, %r96;
prmt.b32 %r98, %r91, %r90, %r96;
shr.u32 %r99, %r85, 16;
prmt.b32 %r100, %r97, %r98, %r99;
mov.u32 %r101, 25632;
prmt.b32 %r59, %r94, %r100, %r101;
mov.u32 %r102, 30001;
prmt.b32 %r63, %r94, %r100, %r102;
ld.global.nc.u32 %r60, [%rd30+4];
mov.u32 %r61, 0;

dp4a.s32.s32 %r58, %r59, %r60, %r61;

ld.global.nc.u32 %r64, [%rd30+20];

dp4a.s32.s32 %r62, %r63, %r64, %r58;

ld.global.nc.u8 %rs6, [%rd34+1];
cvt.u32.u16 %r103, %rs6;
and.b32 %r104, %r103, 255;
ld.global.nc.u8 %rs7, [%rd34+2];
cvt.u32.u16 %r105, %rs7;
prmt.b32 %r106, %r105, %r104, 30212;
ld.global.nc.u8 %rs8, [%rd34+3];
cvt.u32.u16 %r107, %rs8;
prmt.b32 %r108, %r107, %r106, 28756;
ld.global.nc.u8 %rs9, [%rd34+4];
cvt.u32.u16 %r109, %rs9;
prmt.b32 %r110, %r109, %r108, 1620;
and.b32 %r111, %r110, -2004318072;
shr.u32 %r112, %r111, 1;
or.b32 %r113, %r112, 839922192;
and.b32 %r114, %r106, 30583;
prmt.b32 %r115, %r88, %r87, %r114;
prmt.b32 %r116, %r91, %r90, %r114;
and.b32 %r117, %r113, 30292;
prmt.b32 %r118, %r115, %r116, %r117;
shr.u32 %r119, %r110, 16;
and.b32 %r120, %r119, 30583;
prmt.b32 %r121, %r88, %r87, %r120;
prmt.b32 %r122, %r91, %r90, %r120;
shr.u32 %r123, %r113, 16;
prmt.b32 %r124, %r121, %r122, %r123;
prmt.b32 %r67, %r118, %r124, %r101;
prmt.b32 %r71, %r118, %r124, %r102;
ld.global.nc.u32 %r68, [%rd30+8];

dp4a.s32.s32 %r66, %r67, %r68, %r62;

ld.global.nc.u32 %r72, [%rd30+24];

dp4a.s32.s32 %r70, %r71, %r72, %r66;

ld.global.nc.u8 %rs10, [%rd35+-8];
shl.b16 %rs12, %rs10, 7;
and.b16 %rs13, %rs12, 32640;
setp.eq.s16 %p4, %rs10, 255;
setp.eq.s16 %p5, %rs10, 0;
selp.b16 %rs14, 64, %rs13, %p5;
selp.b16 %rs1, 32767, %rs14, %p4;

{ mov.b32 %f7, {0,%rs1};}


mul.ftz.f32 %f9, %f7, 0f3F000000;
ld.global.nc.u32 %r74, [%rd28];

{.reg .f16 low,high;
mov.b32 {low,high},%r74;
cvt.f32.f16 %f8, low;}


mul.ftz.f32 %f10, %f9, %f8;
cvt.rn.f32.s32 %f11, %r70;
fma.rn.ftz.f32 %f28, %f10, %f11, %f28;
add.s64 %rd35, %rd35, 1088;
add.s64 %rd34, %rd34, 1088;
add.s32 %r156, %r156, 64;
setp.lt.s32 %p6, %r156, %r41;
@%p6 bra $L__BB106_5;

$L__BB106_6:
setp.eq.s32 %p7, %r1, 0;
@%p7 bra $L__BB106_8;

shl.b32 %r125, %r1, 7;
mov.u32 %r126, _ZZ13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r127, %r126, %r125;
shl.b32 %r128, %r2, 2;
add.s32 %r129, %r127, %r128;
st.shared.f32 [%r129+-128], %f28;

$L__BB106_8:
setp.ne.s32 %p8, %r1, 0;
bar.sync 0;
@%p8 bra $L__BB106_11;

shl.b32 %r130, %r2, 2;
mov.u32 %r131, _ZZ13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r132, %r131, %r130;
mov.u32 %r133, 2;
ld.shared.f32 %f12, [%r132];
add.ftz.f32 %f13, %f28, %f12;
ld.shared.f32 %f14, [%r132+128];
add.ftz.f32 %f15, %f13, %f14;
ld.shared.f32 %f16, [%r132+256];
add.ftz.f32 %f17, %f15, %f16;
mov.b32 %r134, %f17;
mov.u32 %r135, 31;
mov.u32 %r136, 16;
mov.u32 %r137, -1;
shfl.sync.bfly.b32 %r138|%p9, %r134, %r136, %r135, %r137;
mov.b32 %f18, %r138;
add.ftz.f32 %f19, %f17, %f18;
mov.b32 %r139, %f19;
mov.u32 %r140, 8;
shfl.sync.bfly.b32 %r141|%p10, %r139, %r140, %r135, %r137;
mov.b32 %f20, %r141;
add.ftz.f32 %f21, %f19, %f20;
mov.b32 %r142, %f21;
mov.u32 %r143, 4;
shfl.sync.bfly.b32 %r144|%p11, %r142, %r143, %r135, %r137;
mov.b32 %f22, %r144;
add.ftz.f32 %f23, %f21, %f22;
mov.b32 %r145, %f23;
shfl.sync.bfly.b32 %r146|%p12, %r145, %r133, %r135, %r137;
mov.b32 %f24, %r146;
add.ftz.f32 %f4, %f23, %f24;
mov.b32 %r147, %f4;
mov.u32 %r148, 1;
shfl.sync.bfly.b32 %r14|%p1, %r147, %r148, %r135, %r137;
setp.ne.s32 %p13, %r2, 0;
@%p13 bra $L__BB106_11;

mov.u32 %r149, %ctaid.x;
mov.u32 %r150, %ctaid.y;
mad.lo.s32 %r151, %r150, %r25, %r149;
mov.u32 %r152, %ctaid.z;
mad.lo.s32 %r153, %r152, %r31, %r151;
cvta.to.global.u64 %rd31, %rd13;
mul.wide.u32 %rd32, %r153, 4;
add.s64 %rd33, %rd31, %rd32;
mov.b32 %f25, %r14;
add.ftz.f32 %f26, %f4, %f25;
st.global.f32 [%rd33], %f26;

$L__BB106_11:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type39ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type39ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type39ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type39ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type39ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type39ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type39ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type39ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type39ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<39>;
.reg .b16 %rs<31>;
.reg .f32 %f<113>;
.reg .b32 %r<310>;
.reg .b64 %rd<53>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type39ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type39ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type39ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type39ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r13, [_Z13mul_mat_vec_qIL9ggml_type39ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r14, [_Z13mul_mat_vec_qIL9ggml_type39ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r15, [_Z13mul_mat_vec_qIL9ggml_type39ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r16, [_Z13mul_mat_vec_qIL9ggml_type39ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type39ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type39ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type39ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type39ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type39ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type39ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type39ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type39ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r18, [_Z13mul_mat_vec_qIL9ggml_type39ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r17, [_Z13mul_mat_vec_qIL9ggml_type39ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
cvta.to.global.u64 %rd1, %rd19;
mov.u32 %r1, %tid.y;
shl.b32 %r29, %r1, 5;
mov.u32 %r30, %tid.x;
add.s32 %r2, %r29, %r30;
mov.u32 %r31, %ctaid.x;
shl.b32 %r3, %r31, 1;
shr.u32 %r32, %r2, 31;
add.s32 %r33, %r2, %r32;
shr.s32 %r309, %r33, 1;
shr.u32 %r34, %r13, 5;
setp.ge.s32 %p3, %r309, %r34;
mov.f32 %f109, 0f00000000;
mov.f32 %f110, %f109;
mov.f32 %f111, %f109;
mov.f32 %f112, %f109;
@%p3 bra $L__BB107_3;

mov.u32 %r35, %ctaid.y;
mov.u32 %r36, %ctaid.z;
mul.hi.u32 %r37, %r36, %r23;
add.s32 %r38, %r37, %r36;
shr.u32 %r39, %r38, %r24;
mul.lo.s32 %r40, %r39, %r26;
mul.hi.u32 %r41, %r35, %r17;
add.s32 %r42, %r41, %r35;
shr.u32 %r43, %r42, %r18;
mul.lo.s32 %r44, %r43, %r20;
cvta.to.global.u64 %rd20, %rd17;
and.b32 %r47, %r33, -2;
sub.s32 %r48, %r2, %r47;
shl.b32 %r49, %r48, 1;
cvt.s64.s32 %rd2, %r49;
shl.b32 %r50, %r48, 3;
cvt.s64.s32 %rd21, %r50;
add.s32 %r308, %r309, %r15;
add.s32 %r51, %r309, %r40;
add.s32 %r52, %r51, %r44;
add.s32 %r53, %r3, 1;
mad.lo.s32 %r54, %r14, %r53, %r52;
mul.wide.s32 %rd22, %r54, 17;
add.s64 %rd52, %rd20, %rd22;
mad.lo.s32 %r55, %r3, %r14, %r309;
add.s32 %r56, %r55, %r40;
add.s32 %r57, %r56, %r44;
mul.wide.s32 %rd23, %r57, 17;
add.s64 %rd24, %rd20, %rd23;
add.s64 %rd51, %rd24, 4;
add.s64 %rd25, %rd23, %rd21;
add.s64 %rd26, %rd20, %rd25;
add.s64 %rd50, %rd26, 4;
or.b64 %rd6, %rd21, 4;
mul.lo.s32 %r58, %r35, %r21;
cvt.u64.u32 %rd27, %r58;
mul.lo.s32 %r59, %r36, %r27;
cvt.u64.u32 %rd28, %r59;
add.s64 %rd7, %rd28, %rd27;
cvta.to.global.u64 %rd8, %rd18;
mov.f32 %f109, 0f00000000;
shl.b64 %rd33, %rd2, 2;

$L__BB107_2:
cvt.u64.u32 %rd29, %r309;
add.s64 %rd30, %rd7, %rd29;
mul.lo.s64 %rd31, %rd30, 36;
add.s64 %rd32, %rd8, %rd31;
add.s64 %rd34, %rd32, %rd33;
ld.global.nc.u8 %rs5, [%rd50+-3];
cvt.u32.u16 %r128, %rs5;
and.b32 %r129, %r128, 255;
ld.global.nc.u8 %rs6, [%rd50+-2];
cvt.u32.u16 %r130, %rs6;
prmt.b32 %r131, %r130, %r129, 30212;
ld.global.nc.u8 %rs7, [%rd50+-1];
cvt.u32.u16 %r132, %rs7;
prmt.b32 %r133, %r132, %r131, 28756;
ld.global.nc.u8 %rs8, [%rd50];
cvt.u32.u16 %r134, %rs8;
prmt.b32 %r135, %r134, %r133, 1620;
and.b32 %r136, %r135, -2004318072;
shr.u32 %r137, %r136, 1;
or.b32 %r138, %r137, 839922192;
and.b32 %r139, %r131, 30583;
mov.u32 %r140, 201852420;
mov.u32 %r141, 50462976;
prmt.b32 %r142, %r141, %r140, %r139;
mov.u32 %r143, -185009412;
mov.u32 %r144, -33620224;
prmt.b32 %r145, %r144, %r143, %r139;
and.b32 %r146, %r138, 30292;
prmt.b32 %r147, %r142, %r145, %r146;
shr.u32 %r148, %r135, 16;
and.b32 %r149, %r148, 30583;
prmt.b32 %r150, %r141, %r140, %r149;
prmt.b32 %r151, %r144, %r143, %r149;
shr.u32 %r152, %r138, 16;
prmt.b32 %r153, %r150, %r151, %r152;
mov.u32 %r154, 25632;
prmt.b32 %r95, %r147, %r153, %r154;
mov.u32 %r155, 30001;
prmt.b32 %r99, %r147, %r153, %r155;
ld.global.nc.u32 %r62, [%rd34+4];
mov.u32 %r114, 0;

dp4a.s32.s32 %r60, %r95, %r62, %r114;

ld.global.nc.u32 %r66, [%rd34+20];

dp4a.s32.s32 %r64, %r99, %r66, %r60;

ld.global.nc.u8 %rs9, [%rd50+1];
cvt.u32.u16 %r156, %rs9;
and.b32 %r157, %r156, 255;
ld.global.nc.u8 %rs10, [%rd50+2];
cvt.u32.u16 %r158, %rs10;
prmt.b32 %r159, %r158, %r157, 30212;
ld.global.nc.u8 %rs11, [%rd50+3];
cvt.u32.u16 %r160, %rs11;
prmt.b32 %r161, %r160, %r159, 28756;
ld.global.nc.u8 %rs12, [%rd50+4];
cvt.u32.u16 %r162, %rs12;
prmt.b32 %r163, %r162, %r161, 1620;
and.b32 %r164, %r163, -2004318072;
shr.u32 %r165, %r164, 1;
or.b32 %r166, %r165, 839922192;
and.b32 %r167, %r159, 30583;
prmt.b32 %r168, %r141, %r140, %r167;
prmt.b32 %r169, %r144, %r143, %r167;
and.b32 %r170, %r166, 30292;
prmt.b32 %r171, %r168, %r169, %r170;
shr.u32 %r172, %r163, 16;
and.b32 %r173, %r172, 30583;
prmt.b32 %r174, %r141, %r140, %r173;
prmt.b32 %r175, %r144, %r143, %r173;
shr.u32 %r176, %r166, 16;
prmt.b32 %r177, %r174, %r175, %r176;
prmt.b32 %r103, %r171, %r177, %r154;
prmt.b32 %r107, %r171, %r177, %r155;
ld.global.nc.u32 %r70, [%rd34+8];

dp4a.s32.s32 %r68, %r103, %r70, %r64;

ld.global.nc.u32 %r74, [%rd34+24];

dp4a.s32.s32 %r72, %r107, %r74, %r68;

ld.global.nc.u8 %rs13, [%rd51+-4];
shl.b16 %rs15, %rs13, 7;
and.b16 %rs16, %rs15, 32640;
setp.eq.s16 %p4, %rs13, 255;
setp.eq.s16 %p5, %rs13, 0;
selp.b16 %rs17, 64, %rs16, %p5;
selp.b16 %rs3, 32767, %rs17, %p4;

{ mov.b32 %f25, {0,%rs3};}


mul.ftz.f32 %f33, %f25, 0f3F000000;
ld.global.nc.u32 %r76, [%rd32];

{.reg .f16 low,high;
mov.b32 {low,high},%r76;
cvt.f32.f16 %f26, low;}


mul.ftz.f32 %f34, %f33, %f26;
cvt.rn.f32.s32 %f35, %r72;
fma.rn.ftz.f32 %f112, %f34, %f35, %f112;
add.s64 %rd35, %rd52, %rd6;
ld.global.nc.u8 %rs18, [%rd35+-3];
cvt.u32.u16 %r178, %rs18;
and.b32 %r179, %r178, 255;
ld.global.nc.u8 %rs19, [%rd35+-2];
cvt.u32.u16 %r180, %rs19;
prmt.b32 %r181, %r180, %r179, 30212;
ld.global.nc.u8 %rs20, [%rd35+-1];
cvt.u32.u16 %r182, %rs20;
prmt.b32 %r183, %r182, %r181, 28756;
ld.global.nc.u8 %rs21, [%rd35];
cvt.u32.u16 %r184, %rs21;
prmt.b32 %r185, %r184, %r183, 1620;
and.b32 %r186, %r185, -2004318072;
shr.u32 %r187, %r186, 1;
or.b32 %r188, %r187, 839922192;
and.b32 %r189, %r181, 30583;
prmt.b32 %r190, %r141, %r140, %r189;
prmt.b32 %r191, %r144, %r143, %r189;
and.b32 %r192, %r188, 30292;
prmt.b32 %r193, %r190, %r191, %r192;
shr.u32 %r194, %r185, 16;
and.b32 %r195, %r194, 30583;
prmt.b32 %r196, %r141, %r140, %r195;
prmt.b32 %r197, %r144, %r143, %r195;
shr.u32 %r198, %r188, 16;
prmt.b32 %r199, %r196, %r197, %r198;
prmt.b32 %r112, %r193, %r199, %r154;
prmt.b32 %r116, %r193, %r199, %r155;

dp4a.s32.s32 %r77, %r112, %r62, %r114;


dp4a.s32.s32 %r81, %r116, %r66, %r77;

ld.global.nc.u8 %rs22, [%rd35+1];
cvt.u32.u16 %r200, %rs22;
and.b32 %r201, %r200, 255;
ld.global.nc.u8 %rs23, [%rd35+2];
cvt.u32.u16 %r202, %rs23;
prmt.b32 %r203, %r202, %r201, 30212;
ld.global.nc.u8 %rs24, [%rd35+3];
cvt.u32.u16 %r204, %rs24;
prmt.b32 %r205, %r204, %r203, 28756;
ld.global.nc.u8 %rs25, [%rd35+4];
cvt.u32.u16 %r206, %rs25;
prmt.b32 %r207, %r206, %r205, 1620;
and.b32 %r208, %r207, -2004318072;
shr.u32 %r209, %r208, 1;
or.b32 %r210, %r209, 839922192;
and.b32 %r211, %r203, 30583;
prmt.b32 %r212, %r141, %r140, %r211;
prmt.b32 %r213, %r144, %r143, %r211;
and.b32 %r214, %r210, 30292;
prmt.b32 %r215, %r212, %r213, %r214;
shr.u32 %r216, %r207, 16;
and.b32 %r217, %r216, 30583;
prmt.b32 %r218, %r141, %r140, %r217;
prmt.b32 %r219, %r144, %r143, %r217;
shr.u32 %r220, %r210, 16;
prmt.b32 %r221, %r218, %r219, %r220;
prmt.b32 %r120, %r215, %r221, %r154;
prmt.b32 %r124, %r215, %r221, %r155;

dp4a.s32.s32 %r85, %r120, %r70, %r81;


dp4a.s32.s32 %r89, %r124, %r74, %r85;

ld.global.nc.u8 %rs26, [%rd52];
shl.b16 %rs28, %rs26, 7;
and.b16 %rs29, %rs28, 32640;
setp.eq.s16 %p6, %rs26, 255;
setp.eq.s16 %p7, %rs26, 0;
selp.b16 %rs30, 64, %rs29, %p7;
selp.b16 %rs4, 32767, %rs30, %p6;

{ mov.b32 %f27, {0,%rs4};}


mul.ftz.f32 %f36, %f27, 0f3F000000;

{.reg .f16 low,high;
mov.b32 {low,high},%r76;
cvt.f32.f16 %f28, low;}


mul.ftz.f32 %f37, %f36, %f28;
cvt.rn.f32.s32 %f38, %r89;
fma.rn.ftz.f32 %f111, %f37, %f38, %f111;
cvt.u64.u32 %rd36, %r308;
add.s64 %rd37, %rd7, %rd36;
mul.lo.s64 %rd38, %rd37, 36;
add.s64 %rd39, %rd8, %rd38;
add.s64 %rd40, %rd39, %rd33;
ld.global.nc.u32 %r96, [%rd40+4];

dp4a.s32.s32 %r94, %r95, %r96, %r114;

ld.global.nc.u32 %r100, [%rd40+20];

dp4a.s32.s32 %r98, %r99, %r100, %r94;

ld.global.nc.u32 %r104, [%rd40+8];

dp4a.s32.s32 %r102, %r103, %r104, %r98;

ld.global.nc.u32 %r108, [%rd40+24];

dp4a.s32.s32 %r106, %r107, %r108, %r102;


{ mov.b32 %f29, {0,%rs3};}


mul.ftz.f32 %f39, %f29, 0f3F000000;
ld.global.nc.u32 %r110, [%rd39];

{.reg .f16 low,high;
mov.b32 {low,high},%r110;
cvt.f32.f16 %f30, low;}


mul.ftz.f32 %f40, %f39, %f30;
cvt.rn.f32.s32 %f41, %r106;
fma.rn.ftz.f32 %f110, %f40, %f41, %f110;

dp4a.s32.s32 %r111, %r112, %r96, %r114;


dp4a.s32.s32 %r115, %r116, %r100, %r111;


dp4a.s32.s32 %r119, %r120, %r104, %r115;


dp4a.s32.s32 %r123, %r124, %r108, %r119;


{ mov.b32 %f31, {0,%rs4};}


mul.ftz.f32 %f42, %f31, 0f3F000000;

{.reg .f16 low,high;
mov.b32 {low,high},%r110;
cvt.f32.f16 %f32, low;}


mul.ftz.f32 %f43, %f42, %f32;
cvt.rn.f32.s32 %f44, %r123;
fma.rn.ftz.f32 %f109, %f43, %f44, %f109;
add.s32 %r308, %r308, 64;
add.s64 %rd52, %rd52, 1088;
add.s64 %rd51, %rd51, 1088;
add.s64 %rd50, %rd50, 1088;
add.s32 %r309, %r309, 64;
setp.lt.s32 %p8, %r309, %r34;
@%p8 bra $L__BB107_2;

$L__BB107_3:
setp.eq.s32 %p9, %r1, 0;
@%p9 bra $L__BB107_5;

shl.b32 %r223, %r1, 9;
mov.u32 %r224, _ZZ13mul_mat_vec_qIL9ggml_type39ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r225, %r224, %r223;
shl.b32 %r226, %r30, 2;
add.s32 %r227, %r225, %r226;
st.shared.f32 [%r227+-512], %f112;
st.shared.f32 [%r227+-384], %f111;
st.shared.f32 [%r227+-256], %f110;
st.shared.f32 [%r227+-128], %f109;

$L__BB107_5:
setp.ne.s32 %p10, %r1, 0;
bar.sync 0;
@%p10 bra $L__BB107_14;

mov.u32 %r228, %ctaid.z;
mov.u32 %r229, %ctaid.y;
mad.lo.s32 %r230, %r229, %r22, %r3;
mad.lo.s32 %r231, %r228, %r28, %r230;
cvt.u64.u32 %rd15, %r231;
shl.b32 %r233, %r30, 2;
mov.u32 %r234, _ZZ13mul_mat_vec_qIL9ggml_type39ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r11, %r234, %r233;
mov.u32 %r235, 2;
ld.shared.f32 %f45, [%r11];
add.ftz.f32 %f46, %f45, %f112;
ld.shared.f32 %f47, [%r11+512];
add.ftz.f32 %f48, %f47, %f46;
ld.shared.f32 %f49, [%r11+1024];
add.ftz.f32 %f50, %f49, %f48;
mov.b32 %r236, %f50;
mov.u32 %r237, 31;
mov.u32 %r238, 16;
mov.u32 %r239, -1;
shfl.sync.bfly.b32 %r240|%p11, %r236, %r238, %r237, %r239;
mov.b32 %f51, %r240;
add.ftz.f32 %f52, %f50, %f51;
mov.b32 %r241, %f52;
mov.u32 %r242, 8;
shfl.sync.bfly.b32 %r243|%p12, %r241, %r242, %r237, %r239;
mov.b32 %f53, %r243;
add.ftz.f32 %f54, %f52, %f53;
mov.b32 %r244, %f54;
mov.u32 %r245, 4;
shfl.sync.bfly.b32 %r246|%p13, %r244, %r245, %r237, %r239;
mov.b32 %f55, %r246;
add.ftz.f32 %f56, %f54, %f55;
mov.b32 %r247, %f56;
shfl.sync.bfly.b32 %r248|%p14, %r247, %r235, %r237, %r239;
mov.b32 %f57, %r248;
add.ftz.f32 %f58, %f56, %f57;
mov.b32 %r249, %f58;
mov.u32 %r250, 1;
shfl.sync.bfly.b32 %r251|%p15, %r249, %r250, %r237, %r239;
mov.b32 %f59, %r251;
add.ftz.f32 %f13, %f58, %f59;
setp.eq.s32 %p16, %r30, 0;
setp.lt.u32 %p17, %r3, %r16;
and.pred %p1, %p16, %p17;
mul.wide.u32 %rd41, %r231, 4;
add.s64 %rd16, %rd1, %rd41;
not.pred %p18, %p1;
@%p18 bra $L__BB107_8;

st.global.f32 [%rd16], %f13;

$L__BB107_8:
ld.shared.f32 %f60, [%r11+128];
add.ftz.f32 %f61, %f60, %f111;
ld.shared.f32 %f62, [%r11+640];
add.ftz.f32 %f63, %f62, %f61;
ld.shared.f32 %f64, [%r11+1152];
add.ftz.f32 %f65, %f64, %f63;
mov.b32 %r252, %f65;
shfl.sync.bfly.b32 %r256|%p19, %r252, %r238, %r237, %r239;
mov.b32 %f66, %r256;
add.ftz.f32 %f67, %f65, %f66;
mov.b32 %r257, %f67;
shfl.sync.bfly.b32 %r259|%p20, %r257, %r242, %r237, %r239;
mov.b32 %f68, %r259;
add.ftz.f32 %f69, %f67, %f68;
mov.b32 %r260, %f69;
shfl.sync.bfly.b32 %r262|%p21, %r260, %r245, %r237, %r239;
mov.b32 %f70, %r262;
add.ftz.f32 %f71, %f69, %f70;
mov.b32 %r263, %f71;
shfl.sync.bfly.b32 %r265|%p22, %r263, %r235, %r237, %r239;
mov.b32 %f72, %r265;
add.ftz.f32 %f73, %f71, %f72;
mov.b32 %r266, %f73;
shfl.sync.bfly.b32 %r268|%p23, %r266, %r250, %r237, %r239;
mov.b32 %f74, %r268;
add.ftz.f32 %f14, %f73, %f74;
setp.eq.s32 %p24, %r30, 1;
or.b32 %r272, %r3, 1;
setp.lt.u32 %p25, %r272, %r16;
and.pred %p2, %p24, %p25;
not.pred %p26, %p2;
@%p26 bra $L__BB107_10;

st.global.f32 [%rd16+4], %f14;

$L__BB107_10:
ld.shared.f32 %f75, [%r11+256];
add.ftz.f32 %f76, %f75, %f110;
ld.shared.f32 %f77, [%r11+768];
add.ftz.f32 %f78, %f77, %f76;
ld.shared.f32 %f79, [%r11+1280];
add.ftz.f32 %f80, %f79, %f78;
mov.b32 %r273, %f80;
mov.u32 %r274, 31;
mov.u32 %r275, 16;
mov.u32 %r276, -1;
shfl.sync.bfly.b32 %r277|%p27, %r273, %r275, %r274, %r276;
mov.b32 %f81, %r277;
add.ftz.f32 %f82, %f80, %f81;
mov.b32 %r278, %f82;
mov.u32 %r279, 8;
shfl.sync.bfly.b32 %r280|%p28, %r278, %r279, %r274, %r276;
mov.b32 %f83, %r280;
add.ftz.f32 %f84, %f82, %f83;
mov.b32 %r281, %f84;
mov.u32 %r282, 4;
shfl.sync.bfly.b32 %r283|%p29, %r281, %r282, %r274, %r276;
mov.b32 %f85, %r283;
add.ftz.f32 %f86, %f84, %f85;
mov.b32 %r284, %f86;
mov.u32 %r285, 2;
shfl.sync.bfly.b32 %r286|%p30, %r284, %r285, %r274, %r276;
mov.b32 %f87, %r286;
add.ftz.f32 %f88, %f86, %f87;
mov.b32 %r287, %f88;
mov.u32 %r288, 1;
shfl.sync.bfly.b32 %r289|%p31, %r287, %r288, %r274, %r276;
mov.b32 %f89, %r289;
add.ftz.f32 %f15, %f88, %f89;
@%p18 bra $L__BB107_12;

cvt.u64.u32 %rd42, %r16;
add.s64 %rd43, %rd42, %rd15;
shl.b64 %rd44, %rd43, 2;
add.s64 %rd45, %rd1, %rd44;
st.global.f32 [%rd45], %f15;

$L__BB107_12:
ld.shared.f32 %f90, [%r11+384];
add.ftz.f32 %f91, %f90, %f109;
ld.shared.f32 %f92, [%r11+896];
add.ftz.f32 %f93, %f92, %f91;
ld.shared.f32 %f94, [%r11+1408];
add.ftz.f32 %f95, %f94, %f93;
mov.b32 %r290, %f95;
shfl.sync.bfly.b32 %r294|%p33, %r290, %r275, %r274, %r276;
mov.b32 %f96, %r294;
add.ftz.f32 %f97, %f95, %f96;
mov.b32 %r295, %f97;
shfl.sync.bfly.b32 %r297|%p34, %r295, %r279, %r274, %r276;
mov.b32 %f98, %r297;
add.ftz.f32 %f99, %f97, %f98;
mov.b32 %r298, %f99;
shfl.sync.bfly.b32 %r300|%p35, %r298, %r282, %r274, %r276;
mov.b32 %f100, %r300;
add.ftz.f32 %f101, %f99, %f100;
mov.b32 %r301, %f101;
shfl.sync.bfly.b32 %r303|%p36, %r301, %r285, %r274, %r276;
mov.b32 %f102, %r303;
add.ftz.f32 %f103, %f101, %f102;
mov.b32 %r304, %f103;
shfl.sync.bfly.b32 %r306|%p37, %r304, %r288, %r274, %r276;
mov.b32 %f104, %r306;
add.ftz.f32 %f16, %f103, %f104;
@%p26 bra $L__BB107_14;

add.s32 %r307, %r16, 1;
cvt.u64.u32 %rd46, %r307;
add.s64 %rd47, %rd46, %rd15;
shl.b64 %rd48, %rd47, 2;
add.s64 %rd49, %rd1, %rd48;
st.global.f32 [%rd49], %f16;

$L__BB107_14:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type39ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type39ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type39ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type39ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type39ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type39ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type39ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type39ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type39ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<54>;
.reg .b16 %rs<33>;
.reg .f32 %f<169>;
.reg .b32 %r<389>;
.reg .b64 %rd<65>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type39ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[2304];

ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type39ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type39ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type39ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r17, [_Z13mul_mat_vec_qIL9ggml_type39ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r18, [_Z13mul_mat_vec_qIL9ggml_type39ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r19, [_Z13mul_mat_vec_qIL9ggml_type39ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type39ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type39ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type39ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type39ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type39ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type39ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type39ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type39ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type39ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type39ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type39ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r33, %r1, 5;
mov.u32 %r34, %tid.x;
add.s32 %r2, %r33, %r34;
mov.u32 %r3, %ctaid.y;
mov.u32 %r4, %ctaid.z;
shr.u32 %r35, %r2, 31;
add.s32 %r36, %r2, %r35;
shr.s32 %r388, %r36, 1;
shr.u32 %r37, %r17, 5;
setp.ge.s32 %p2, %r388, %r37;
mov.f32 %f163, 0f00000000;
mov.f32 %f164, %f163;
mov.f32 %f165, %f163;
mov.f32 %f166, %f163;
mov.f32 %f167, %f163;
mov.f32 %f168, %f163;
@%p2 bra $L__BB108_3;

mul.lo.s32 %r38, %r3, %r25;
cvt.u64.u32 %rd20, %r38;
mul.lo.s32 %r39, %r4, %r31;
cvt.u64.u32 %rd21, %r39;
add.s64 %rd1, %rd21, %rd20;
mul.hi.u32 %r40, %r4, %r27;
add.s32 %r41, %r40, %r4;
shr.u32 %r42, %r41, %r28;
mul.lo.s32 %r43, %r42, %r30;
mul.hi.u32 %r44, %r3, %r21;
add.s32 %r45, %r44, %r3;
shr.u32 %r46, %r45, %r22;
mul.lo.s32 %r47, %r46, %r24;
cvta.to.global.u64 %rd22, %rd17;
and.b32 %r50, %r36, -2;
sub.s32 %r51, %r2, %r50;
shl.b32 %r52, %r51, 1;
cvt.s64.s32 %rd2, %r52;
mov.u32 %r53, %ctaid.x;
shl.b32 %r54, %r53, 1;
shl.b32 %r55, %r51, 3;
cvt.s64.s32 %rd23, %r55;
shl.b32 %r56, %r19, 1;
add.s32 %r387, %r388, %r56;
add.s32 %r386, %r388, %r19;
add.s32 %r57, %r388, %r43;
add.s32 %r58, %r57, %r47;
or.b32 %r59, %r54, 1;
mad.lo.s32 %r60, %r18, %r59, %r58;
mul.wide.s32 %rd24, %r60, 17;
add.s64 %rd64, %rd22, %rd24;
mad.lo.s32 %r61, %r54, %r18, %r388;
add.s32 %r62, %r61, %r43;
add.s32 %r63, %r62, %r47;
mul.wide.s32 %rd25, %r63, 17;
add.s64 %rd26, %rd22, %rd25;
add.s64 %rd63, %rd26, 4;
add.s64 %rd27, %rd25, %rd23;
add.s64 %rd28, %rd22, %rd27;
add.s64 %rd62, %rd28, 4;
or.b64 %rd6, %rd23, 4;
cvta.to.global.u64 %rd7, %rd18;
mov.f32 %f163, 0f00000000;
shl.b64 %rd33, %rd2, 2;

$L__BB108_2:
cvt.u64.u32 %rd29, %r388;
add.s64 %rd30, %rd1, %rd29;
mul.lo.s64 %rd31, %rd30, 36;
add.s64 %rd32, %rd7, %rd31;
add.s64 %rd34, %rd32, %rd33;
ld.global.nc.u8 %rs7, [%rd62+-3];
cvt.u32.u16 %r166, %rs7;
and.b32 %r167, %r166, 255;
ld.global.nc.u8 %rs8, [%rd62+-2];
cvt.u32.u16 %r168, %rs8;
prmt.b32 %r169, %r168, %r167, 30212;
ld.global.nc.u8 %rs9, [%rd62+-1];
cvt.u32.u16 %r170, %rs9;
prmt.b32 %r171, %r170, %r169, 28756;
ld.global.nc.u8 %rs10, [%rd62];
cvt.u32.u16 %r172, %rs10;
prmt.b32 %r173, %r172, %r171, 1620;
and.b32 %r174, %r173, -2004318072;
shr.u32 %r175, %r174, 1;
or.b32 %r176, %r175, 839922192;
and.b32 %r177, %r169, 30583;
mov.u32 %r178, 201852420;
mov.u32 %r179, 50462976;
prmt.b32 %r180, %r179, %r178, %r177;
mov.u32 %r181, -185009412;
mov.u32 %r182, -33620224;
prmt.b32 %r183, %r182, %r181, %r177;
and.b32 %r184, %r176, 30292;
prmt.b32 %r185, %r180, %r183, %r184;
shr.u32 %r186, %r173, 16;
and.b32 %r187, %r186, 30583;
prmt.b32 %r188, %r179, %r178, %r187;
prmt.b32 %r189, %r182, %r181, %r187;
shr.u32 %r190, %r176, 16;
prmt.b32 %r191, %r188, %r189, %r190;
mov.u32 %r192, 25632;
prmt.b32 %r133, %r185, %r191, %r192;
mov.u32 %r193, 30001;
prmt.b32 %r137, %r185, %r191, %r193;
ld.global.nc.u32 %r66, [%rd34+4];
mov.u32 %r152, 0;

dp4a.s32.s32 %r64, %r133, %r66, %r152;

ld.global.nc.u32 %r70, [%rd34+20];

dp4a.s32.s32 %r68, %r137, %r70, %r64;

ld.global.nc.u8 %rs11, [%rd62+1];
cvt.u32.u16 %r194, %rs11;
and.b32 %r195, %r194, 255;
ld.global.nc.u8 %rs12, [%rd62+2];
cvt.u32.u16 %r196, %rs12;
prmt.b32 %r197, %r196, %r195, 30212;
ld.global.nc.u8 %rs13, [%rd62+3];
cvt.u32.u16 %r198, %rs13;
prmt.b32 %r199, %r198, %r197, 28756;
ld.global.nc.u8 %rs14, [%rd62+4];
cvt.u32.u16 %r200, %rs14;
prmt.b32 %r201, %r200, %r199, 1620;
and.b32 %r202, %r201, -2004318072;
shr.u32 %r203, %r202, 1;
or.b32 %r204, %r203, 839922192;
and.b32 %r205, %r197, 30583;
prmt.b32 %r206, %r179, %r178, %r205;
prmt.b32 %r207, %r182, %r181, %r205;
and.b32 %r208, %r204, 30292;
prmt.b32 %r209, %r206, %r207, %r208;
shr.u32 %r210, %r201, 16;
and.b32 %r211, %r210, 30583;
prmt.b32 %r212, %r179, %r178, %r211;
prmt.b32 %r213, %r182, %r181, %r211;
shr.u32 %r214, %r204, 16;
prmt.b32 %r215, %r212, %r213, %r214;
prmt.b32 %r141, %r209, %r215, %r192;
prmt.b32 %r145, %r209, %r215, %r193;
ld.global.nc.u32 %r74, [%rd34+8];

dp4a.s32.s32 %r72, %r141, %r74, %r68;

ld.global.nc.u32 %r78, [%rd34+24];

dp4a.s32.s32 %r76, %r145, %r78, %r72;

ld.global.nc.u8 %rs15, [%rd63+-4];
shl.b16 %rs17, %rs15, 7;
and.b16 %rs18, %rs17, 32640;
setp.eq.s16 %p3, %rs15, 255;
setp.eq.s16 %p4, %rs15, 0;
selp.b16 %rs19, 64, %rs18, %p4;
selp.b16 %rs5, 32767, %rs19, %p3;

{ mov.b32 %f37, {0,%rs5};}


mul.ftz.f32 %f49, %f37, 0f3F000000;
ld.global.nc.u32 %r80, [%rd32];

{.reg .f16 low,high;
mov.b32 {low,high},%r80;
cvt.f32.f16 %f38, low;}


mul.ftz.f32 %f50, %f49, %f38;
cvt.rn.f32.s32 %f51, %r76;
fma.rn.ftz.f32 %f168, %f50, %f51, %f168;
add.s64 %rd35, %rd64, %rd6;
ld.global.nc.u8 %rs20, [%rd35+-3];
cvt.u32.u16 %r216, %rs20;
and.b32 %r217, %r216, 255;
ld.global.nc.u8 %rs21, [%rd35+-2];
cvt.u32.u16 %r218, %rs21;
prmt.b32 %r219, %r218, %r217, 30212;
ld.global.nc.u8 %rs22, [%rd35+-1];
cvt.u32.u16 %r220, %rs22;
prmt.b32 %r221, %r220, %r219, 28756;
ld.global.nc.u8 %rs23, [%rd35];
cvt.u32.u16 %r222, %rs23;
prmt.b32 %r223, %r222, %r221, 1620;
and.b32 %r224, %r223, -2004318072;
shr.u32 %r225, %r224, 1;
or.b32 %r226, %r225, 839922192;
and.b32 %r227, %r219, 30583;
prmt.b32 %r228, %r179, %r178, %r227;
prmt.b32 %r229, %r182, %r181, %r227;
and.b32 %r230, %r226, 30292;
prmt.b32 %r231, %r228, %r229, %r230;
shr.u32 %r232, %r223, 16;
and.b32 %r233, %r232, 30583;
prmt.b32 %r234, %r179, %r178, %r233;
prmt.b32 %r235, %r182, %r181, %r233;
shr.u32 %r236, %r226, 16;
prmt.b32 %r237, %r234, %r235, %r236;
prmt.b32 %r150, %r231, %r237, %r192;
prmt.b32 %r154, %r231, %r237, %r193;

dp4a.s32.s32 %r81, %r150, %r66, %r152;


dp4a.s32.s32 %r85, %r154, %r70, %r81;

ld.global.nc.u8 %rs24, [%rd35+1];
cvt.u32.u16 %r238, %rs24;
and.b32 %r239, %r238, 255;
ld.global.nc.u8 %rs25, [%rd35+2];
cvt.u32.u16 %r240, %rs25;
prmt.b32 %r241, %r240, %r239, 30212;
ld.global.nc.u8 %rs26, [%rd35+3];
cvt.u32.u16 %r242, %rs26;
prmt.b32 %r243, %r242, %r241, 28756;
ld.global.nc.u8 %rs27, [%rd35+4];
cvt.u32.u16 %r244, %rs27;
prmt.b32 %r245, %r244, %r243, 1620;
and.b32 %r246, %r245, -2004318072;
shr.u32 %r247, %r246, 1;
or.b32 %r248, %r247, 839922192;
and.b32 %r249, %r241, 30583;
prmt.b32 %r250, %r179, %r178, %r249;
prmt.b32 %r251, %r182, %r181, %r249;
and.b32 %r252, %r248, 30292;
prmt.b32 %r253, %r250, %r251, %r252;
shr.u32 %r254, %r245, 16;
and.b32 %r255, %r254, 30583;
prmt.b32 %r256, %r179, %r178, %r255;
prmt.b32 %r257, %r182, %r181, %r255;
shr.u32 %r258, %r248, 16;
prmt.b32 %r259, %r256, %r257, %r258;
prmt.b32 %r158, %r253, %r259, %r192;
prmt.b32 %r162, %r253, %r259, %r193;

dp4a.s32.s32 %r89, %r158, %r74, %r85;


dp4a.s32.s32 %r93, %r162, %r78, %r89;

ld.global.nc.u8 %rs28, [%rd64];
shl.b16 %rs30, %rs28, 7;
and.b16 %rs31, %rs30, 32640;
setp.eq.s16 %p5, %rs28, 255;
setp.eq.s16 %p6, %rs28, 0;
selp.b16 %rs32, 64, %rs31, %p6;
selp.b16 %rs6, 32767, %rs32, %p5;

{ mov.b32 %f39, {0,%rs6};}


mul.ftz.f32 %f52, %f39, 0f3F000000;

{.reg .f16 low,high;
mov.b32 {low,high},%r80;
cvt.f32.f16 %f40, low;}


mul.ftz.f32 %f53, %f52, %f40;
cvt.rn.f32.s32 %f54, %r93;
fma.rn.ftz.f32 %f167, %f53, %f54, %f167;
cvt.u64.u32 %rd36, %r386;
add.s64 %rd37, %rd1, %rd36;
mul.lo.s64 %rd38, %rd37, 36;
add.s64 %rd39, %rd7, %rd38;
add.s64 %rd40, %rd39, %rd33;
ld.global.nc.u32 %r100, [%rd40+4];

dp4a.s32.s32 %r98, %r133, %r100, %r152;

ld.global.nc.u32 %r104, [%rd40+20];

dp4a.s32.s32 %r102, %r137, %r104, %r98;

ld.global.nc.u32 %r108, [%rd40+8];

dp4a.s32.s32 %r106, %r141, %r108, %r102;

ld.global.nc.u32 %r112, [%rd40+24];

dp4a.s32.s32 %r110, %r145, %r112, %r106;


{ mov.b32 %f41, {0,%rs5};}


mul.ftz.f32 %f55, %f41, 0f3F000000;
ld.global.nc.u32 %r114, [%rd39];

{.reg .f16 low,high;
mov.b32 {low,high},%r114;
cvt.f32.f16 %f42, low;}


mul.ftz.f32 %f56, %f55, %f42;
cvt.rn.f32.s32 %f57, %r110;
fma.rn.ftz.f32 %f166, %f56, %f57, %f166;

dp4a.s32.s32 %r115, %r150, %r100, %r152;


dp4a.s32.s32 %r119, %r154, %r104, %r115;


dp4a.s32.s32 %r123, %r158, %r108, %r119;


dp4a.s32.s32 %r127, %r162, %r112, %r123;


{ mov.b32 %f43, {0,%rs6};}


mul.ftz.f32 %f58, %f43, 0f3F000000;

{.reg .f16 low,high;
mov.b32 {low,high},%r114;
cvt.f32.f16 %f44, low;}


mul.ftz.f32 %f59, %f58, %f44;
cvt.rn.f32.s32 %f60, %r127;
fma.rn.ftz.f32 %f165, %f59, %f60, %f165;
cvt.u64.u32 %rd41, %r387;
add.s64 %rd42, %rd1, %rd41;
mul.lo.s64 %rd43, %rd42, 36;
add.s64 %rd44, %rd7, %rd43;
add.s64 %rd45, %rd44, %rd33;
ld.global.nc.u32 %r134, [%rd45+4];

dp4a.s32.s32 %r132, %r133, %r134, %r152;

ld.global.nc.u32 %r138, [%rd45+20];

dp4a.s32.s32 %r136, %r137, %r138, %r132;

ld.global.nc.u32 %r142, [%rd45+8];

dp4a.s32.s32 %r140, %r141, %r142, %r136;

ld.global.nc.u32 %r146, [%rd45+24];

dp4a.s32.s32 %r144, %r145, %r146, %r140;


{ mov.b32 %f45, {0,%rs5};}


mul.ftz.f32 %f61, %f45, 0f3F000000;
ld.global.nc.u32 %r148, [%rd44];

{.reg .f16 low,high;
mov.b32 {low,high},%r148;
cvt.f32.f16 %f46, low;}


mul.ftz.f32 %f62, %f61, %f46;
cvt.rn.f32.s32 %f63, %r144;
fma.rn.ftz.f32 %f164, %f62, %f63, %f164;

dp4a.s32.s32 %r149, %r150, %r134, %r152;


dp4a.s32.s32 %r153, %r154, %r138, %r149;


dp4a.s32.s32 %r157, %r158, %r142, %r153;


dp4a.s32.s32 %r161, %r162, %r146, %r157;


{ mov.b32 %f47, {0,%rs6};}


mul.ftz.f32 %f64, %f47, 0f3F000000;

{.reg .f16 low,high;
mov.b32 {low,high},%r148;
cvt.f32.f16 %f48, low;}


mul.ftz.f32 %f65, %f64, %f48;
cvt.rn.f32.s32 %f66, %r161;
fma.rn.ftz.f32 %f163, %f65, %f66, %f163;
add.s32 %r387, %r387, 64;
add.s32 %r386, %r386, 64;
add.s64 %rd64, %rd64, 1088;
add.s64 %rd63, %rd63, 1088;
add.s64 %rd62, %rd62, 1088;
add.s32 %r388, %r388, 64;
setp.lt.s32 %p7, %r388, %r37;
@%p7 bra $L__BB108_2;

$L__BB108_3:
setp.eq.s32 %p8, %r1, 0;
@%p8 bra $L__BB108_5;

mov.u32 %r261, _ZZ13mul_mat_vec_qIL9ggml_type39ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r262, %r1, 768, %r261;
shl.b32 %r263, %r34, 2;
add.s32 %r264, %r262, %r263;
st.shared.f32 [%r264+-768], %f168;
st.shared.f32 [%r264+-640], %f167;
st.shared.f32 [%r264+-512], %f166;
st.shared.f32 [%r264+-384], %f165;
st.shared.f32 [%r264+-256], %f164;
st.shared.f32 [%r264+-128], %f163;

$L__BB108_5:
setp.ne.s32 %p9, %r1, 0;
bar.sync 0;
@%p9 bra $L__BB108_18;

mov.u32 %r265, %ctaid.x;
shl.b32 %r266, %r265, 1;
mov.u32 %r267, 1;
mad.lo.s32 %r268, %r3, %r26, %r266;
mad.lo.s32 %r269, %r4, %r32, %r268;
cvt.u64.u32 %rd14, %r269;
shl.b32 %r271, %r34, 2;
mov.u32 %r272, _ZZ13mul_mat_vec_qIL9ggml_type39ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r15, %r272, %r271;
mov.u32 %r273, 2;
ld.shared.f32 %f67, [%r15];
add.ftz.f32 %f68, %f67, %f168;
ld.shared.f32 %f69, [%r15+768];
add.ftz.f32 %f70, %f69, %f68;
ld.shared.f32 %f71, [%r15+1536];
add.ftz.f32 %f72, %f71, %f70;
mov.b32 %r274, %f72;
mov.u32 %r275, 31;
mov.u32 %r276, 16;
mov.u32 %r277, -1;
shfl.sync.bfly.b32 %r278|%p10, %r274, %r276, %r275, %r277;
mov.b32 %f73, %r278;
add.ftz.f32 %f74, %f72, %f73;
mov.b32 %r279, %f74;
mov.u32 %r280, 8;
shfl.sync.bfly.b32 %r281|%p11, %r279, %r280, %r275, %r277;
mov.b32 %f75, %r281;
add.ftz.f32 %f76, %f74, %f75;
mov.b32 %r282, %f76;
mov.u32 %r283, 4;
shfl.sync.bfly.b32 %r284|%p12, %r282, %r283, %r275, %r277;
mov.b32 %f77, %r284;
add.ftz.f32 %f78, %f76, %f77;
mov.b32 %r285, %f78;
shfl.sync.bfly.b32 %r286|%p13, %r285, %r273, %r275, %r277;
mov.b32 %f79, %r286;
add.ftz.f32 %f80, %f78, %f79;
mov.b32 %r287, %f80;
shfl.sync.bfly.b32 %r288|%p14, %r287, %r267, %r275, %r277;
mov.b32 %f81, %r288;
add.ftz.f32 %f19, %f80, %f81;
setp.ne.s32 %p15, %r34, 0;
setp.ge.u32 %p16, %r266, %r20;
cvta.to.global.u64 %rd46, %rd19;
mul.wide.u32 %rd47, %r269, 4;
add.s64 %rd15, %rd46, %rd47;
or.pred %p17, %p15, %p16;
@%p17 bra $L__BB108_8;

st.global.f32 [%rd15], %f19;

$L__BB108_8:
ld.shared.f32 %f82, [%r15+128];
add.ftz.f32 %f83, %f82, %f167;
ld.shared.f32 %f84, [%r15+896];
add.ftz.f32 %f85, %f84, %f83;
ld.shared.f32 %f86, [%r15+1664];
add.ftz.f32 %f87, %f86, %f85;
mov.b32 %r289, %f87;
shfl.sync.bfly.b32 %r293|%p18, %r289, %r276, %r275, %r277;
mov.b32 %f88, %r293;
add.ftz.f32 %f89, %f87, %f88;
mov.b32 %r294, %f89;
shfl.sync.bfly.b32 %r296|%p19, %r294, %r280, %r275, %r277;
mov.b32 %f90, %r296;
add.ftz.f32 %f91, %f89, %f90;
mov.b32 %r297, %f91;
shfl.sync.bfly.b32 %r299|%p20, %r297, %r283, %r275, %r277;
mov.b32 %f92, %r299;
add.ftz.f32 %f93, %f91, %f92;
mov.b32 %r300, %f93;
shfl.sync.bfly.b32 %r302|%p21, %r300, %r273, %r275, %r277;
mov.b32 %f94, %r302;
add.ftz.f32 %f95, %f93, %f94;
mov.b32 %r303, %f95;
shfl.sync.bfly.b32 %r305|%p22, %r303, %r267, %r275, %r277;
mov.b32 %f96, %r305;
add.ftz.f32 %f20, %f95, %f96;
setp.eq.s32 %p23, %r34, 1;
or.b32 %r309, %r266, 1;
setp.lt.u32 %p24, %r309, %r20;
and.pred %p1, %p23, %p24;
not.pred %p25, %p1;
@%p25 bra $L__BB108_10;

st.global.f32 [%rd15+4], %f20;

$L__BB108_10:
ld.shared.f32 %f97, [%r15+256];
add.ftz.f32 %f98, %f97, %f166;
ld.shared.f32 %f99, [%r15+1024];
add.ftz.f32 %f100, %f99, %f98;
ld.shared.f32 %f101, [%r15+1792];
add.ftz.f32 %f102, %f101, %f100;
mov.b32 %r310, %f102;
mov.u32 %r311, 31;
mov.u32 %r312, 16;
mov.u32 %r313, -1;
shfl.sync.bfly.b32 %r314|%p26, %r310, %r312, %r311, %r313;
mov.b32 %f103, %r314;
add.ftz.f32 %f104, %f102, %f103;
mov.b32 %r315, %f104;
mov.u32 %r316, 8;
shfl.sync.bfly.b32 %r317|%p27, %r315, %r316, %r311, %r313;
mov.b32 %f105, %r317;
add.ftz.f32 %f106, %f104, %f105;
mov.b32 %r318, %f106;
mov.u32 %r319, 4;
shfl.sync.bfly.b32 %r320|%p28, %r318, %r319, %r311, %r313;
mov.b32 %f107, %r320;
add.ftz.f32 %f108, %f106, %f107;
mov.b32 %r321, %f108;
mov.u32 %r322, 2;
shfl.sync.bfly.b32 %r323|%p29, %r321, %r322, %r311, %r313;
mov.b32 %f109, %r323;
add.ftz.f32 %f110, %f108, %f109;
mov.b32 %r324, %f110;
mov.u32 %r325, 1;
shfl.sync.bfly.b32 %r326|%p30, %r324, %r325, %r311, %r313;
mov.b32 %f111, %r326;
add.ftz.f32 %f21, %f110, %f111;
@%p17 bra $L__BB108_12;

cvt.u64.u32 %rd48, %r20;
add.s64 %rd49, %rd48, %rd14;
shl.b64 %rd51, %rd49, 2;
add.s64 %rd52, %rd46, %rd51;
st.global.f32 [%rd52], %f21;

$L__BB108_12:
ld.shared.f32 %f112, [%r15+384];
add.ftz.f32 %f113, %f112, %f165;
ld.shared.f32 %f114, [%r15+1152];
add.ftz.f32 %f115, %f114, %f113;
ld.shared.f32 %f116, [%r15+1920];
add.ftz.f32 %f117, %f116, %f115;
mov.b32 %r330, %f117;
shfl.sync.bfly.b32 %r334|%p34, %r330, %r312, %r311, %r313;
mov.b32 %f118, %r334;
add.ftz.f32 %f119, %f117, %f118;
mov.b32 %r335, %f119;
shfl.sync.bfly.b32 %r337|%p35, %r335, %r316, %r311, %r313;
mov.b32 %f120, %r337;
add.ftz.f32 %f121, %f119, %f120;
mov.b32 %r338, %f121;
shfl.sync.bfly.b32 %r340|%p36, %r338, %r319, %r311, %r313;
mov.b32 %f122, %r340;
add.ftz.f32 %f123, %f121, %f122;
mov.b32 %r341, %f123;
shfl.sync.bfly.b32 %r343|%p37, %r341, %r322, %r311, %r313;
mov.b32 %f124, %r343;
add.ftz.f32 %f125, %f123, %f124;
mov.b32 %r344, %f125;
shfl.sync.bfly.b32 %r346|%p38, %r344, %r325, %r311, %r313;
mov.b32 %f126, %r346;
add.ftz.f32 %f22, %f125, %f126;
@%p25 bra $L__BB108_14;

add.s32 %r347, %r20, 1;
cvt.u64.u32 %rd53, %r347;
add.s64 %rd54, %rd53, %rd14;
shl.b64 %rd56, %rd54, 2;
add.s64 %rd57, %rd46, %rd56;
st.global.f32 [%rd57], %f22;

$L__BB108_14:
shl.b32 %r348, %r20, 1;
mov.u32 %r349, 1;
ld.shared.f32 %f127, [%r15+512];
add.ftz.f32 %f128, %f127, %f164;
ld.shared.f32 %f129, [%r15+1280];
add.ftz.f32 %f130, %f129, %f128;
ld.shared.f32 %f131, [%r15+2048];
add.ftz.f32 %f132, %f131, %f130;
mov.b32 %r350, %f132;
mov.u32 %r351, 31;
mov.u32 %r352, 16;
mov.u32 %r353, -1;
shfl.sync.bfly.b32 %r354|%p40, %r350, %r352, %r351, %r353;
mov.b32 %f133, %r354;
add.ftz.f32 %f134, %f132, %f133;
mov.b32 %r355, %f134;
mov.u32 %r356, 8;
shfl.sync.bfly.b32 %r357|%p41, %r355, %r356, %r351, %r353;
mov.b32 %f135, %r357;
add.ftz.f32 %f136, %f134, %f135;
mov.b32 %r358, %f136;
mov.u32 %r359, 4;
shfl.sync.bfly.b32 %r360|%p42, %r358, %r359, %r351, %r353;
mov.b32 %f137, %r360;
add.ftz.f32 %f138, %f136, %f137;
mov.b32 %r361, %f138;
mov.u32 %r362, 2;
shfl.sync.bfly.b32 %r363|%p43, %r361, %r362, %r351, %r353;
mov.b32 %f139, %r363;
add.ftz.f32 %f140, %f138, %f139;
mov.b32 %r364, %f140;
shfl.sync.bfly.b32 %r365|%p44, %r364, %r349, %r351, %r353;
mov.b32 %f141, %r365;
add.ftz.f32 %f23, %f140, %f141;
cvt.u64.u32 %rd58, %r348;
add.s64 %rd59, %rd58, %rd14;
shl.b64 %rd61, %rd59, 2;
add.s64 %rd16, %rd46, %rd61;
@%p17 bra $L__BB108_16;

st.global.f32 [%rd16], %f23;

$L__BB108_16:
ld.shared.f32 %f142, [%r15+640];
add.ftz.f32 %f143, %f142, %f163;
ld.shared.f32 %f144, [%r15+1408];
add.ftz.f32 %f145, %f144, %f143;
ld.shared.f32 %f146, [%r15+2176];
add.ftz.f32 %f147, %f146, %f145;
mov.b32 %r369, %f147;
shfl.sync.bfly.b32 %r373|%p48, %r369, %r352, %r351, %r353;
mov.b32 %f148, %r373;
add.ftz.f32 %f149, %f147, %f148;
mov.b32 %r374, %f149;
shfl.sync.bfly.b32 %r376|%p49, %r374, %r356, %r351, %r353;
mov.b32 %f150, %r376;
add.ftz.f32 %f151, %f149, %f150;
mov.b32 %r377, %f151;
shfl.sync.bfly.b32 %r379|%p50, %r377, %r359, %r351, %r353;
mov.b32 %f152, %r379;
add.ftz.f32 %f153, %f151, %f152;
mov.b32 %r380, %f153;
shfl.sync.bfly.b32 %r382|%p51, %r380, %r362, %r351, %r353;
mov.b32 %f154, %r382;
add.ftz.f32 %f155, %f153, %f154;
mov.b32 %r383, %f155;
shfl.sync.bfly.b32 %r385|%p52, %r383, %r349, %r351, %r353;
mov.b32 %f156, %r385;
add.ftz.f32 %f24, %f155, %f156;
@%p25 bra $L__BB108_18;

st.global.f32 [%rd16+4], %f24;

$L__BB108_18:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type39ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type39ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type39ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type39ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type39ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type39ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type39ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type39ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type39ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<68>;
.reg .b16 %rs<35>;
.reg .f32 %f<225>;
.reg .b32 %r<508>;
.reg .b64 %rd<83>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type39ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[3072];

ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type39ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type39ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type39ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r16, [_Z13mul_mat_vec_qIL9ggml_type39ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r17, [_Z13mul_mat_vec_qIL9ggml_type39ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r18, [_Z13mul_mat_vec_qIL9ggml_type39ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r19, [_Z13mul_mat_vec_qIL9ggml_type39ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type39ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type39ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type39ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type39ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type39ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type39ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type39ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type39ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type39ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type39ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r32, %r1, 5;
mov.u32 %r33, %tid.x;
add.s32 %r34, %r32, %r33;
shr.u32 %r35, %r34, 31;
add.s32 %r36, %r34, %r35;
shr.s32 %r507, %r36, 1;
shr.u32 %r37, %r16, 5;
setp.ge.s32 %p2, %r507, %r37;
mov.f32 %f217, 0f00000000;
mov.f32 %f218, %f217;
mov.f32 %f219, %f217;
mov.f32 %f220, %f217;
mov.f32 %f221, %f217;
mov.f32 %f222, %f217;
mov.f32 %f223, %f217;
mov.f32 %f224, %f217;
@%p2 bra $L__BB109_3;

mov.u32 %r38, %ctaid.y;
mov.u32 %r39, %ctaid.z;
mul.hi.u32 %r40, %r39, %r26;
add.s32 %r41, %r40, %r39;
shr.u32 %r42, %r41, %r27;
mul.lo.s32 %r43, %r42, %r29;
mul.hi.u32 %r44, %r38, %r20;
add.s32 %r45, %r44, %r38;
shr.u32 %r46, %r45, %r21;
mul.lo.s32 %r47, %r46, %r23;
cvta.to.global.u64 %rd19, %rd16;
and.b32 %r54, %r36, -2;
sub.s32 %r55, %r34, %r54;
shl.b32 %r56, %r55, 1;
cvt.s64.s32 %rd1, %r56;
mov.u32 %r57, %ctaid.x;
shl.b32 %r58, %r57, 1;
shl.b32 %r59, %r55, 3;
cvt.s64.s32 %rd20, %r59;
shl.b32 %r60, %r18, 1;
add.s32 %r506, %r507, %r60;
mad.lo.s32 %r505, %r18, 3, %r507;
add.s32 %r504, %r507, %r18;
add.s32 %r61, %r507, %r43;
add.s32 %r62, %r61, %r47;
or.b32 %r63, %r58, 1;
mad.lo.s32 %r64, %r17, %r63, %r62;
mul.wide.s32 %rd21, %r64, 17;
add.s64 %rd82, %rd19, %rd21;
mad.lo.s32 %r65, %r58, %r17, %r507;
add.s32 %r66, %r65, %r43;
add.s32 %r67, %r66, %r47;
mul.wide.s32 %rd22, %r67, 17;
add.s64 %rd23, %rd19, %rd22;
add.s64 %rd81, %rd23, 4;
add.s64 %rd24, %rd22, %rd20;
add.s64 %rd25, %rd19, %rd24;
add.s64 %rd80, %rd25, 4;
or.b64 %rd5, %rd20, 4;
mul.lo.s32 %r68, %r38, %r24;
cvt.u64.u32 %rd26, %r68;
mul.lo.s32 %r69, %r39, %r30;
cvt.u64.u32 %rd27, %r69;
add.s64 %rd6, %rd27, %rd26;
cvta.to.global.u64 %rd7, %rd17;
mov.f32 %f217, 0f00000000;
shl.b64 %rd32, %rd1, 2;

$L__BB109_2:
cvt.u64.u32 %rd28, %r507;
add.s64 %rd29, %rd6, %rd28;
mul.lo.s64 %rd30, %rd29, 36;
add.s64 %rd31, %rd7, %rd30;
add.s64 %rd33, %rd31, %rd32;
ld.global.nc.u8 %rs9, [%rd80+-3];
cvt.u32.u16 %r206, %rs9;
and.b32 %r207, %r206, 255;
ld.global.nc.u8 %rs10, [%rd80+-2];
cvt.u32.u16 %r208, %rs10;
prmt.b32 %r209, %r208, %r207, 30212;
ld.global.nc.u8 %rs11, [%rd80+-1];
cvt.u32.u16 %r210, %rs11;
prmt.b32 %r211, %r210, %r209, 28756;
ld.global.nc.u8 %rs12, [%rd80];
cvt.u32.u16 %r212, %rs12;
prmt.b32 %r213, %r212, %r211, 1620;
and.b32 %r214, %r213, -2004318072;
shr.u32 %r215, %r214, 1;
or.b32 %r216, %r215, 839922192;
and.b32 %r217, %r209, 30583;
mov.u32 %r218, 201852420;
mov.u32 %r219, 50462976;
prmt.b32 %r220, %r219, %r218, %r217;
mov.u32 %r221, -185009412;
mov.u32 %r222, -33620224;
prmt.b32 %r223, %r222, %r221, %r217;
and.b32 %r224, %r216, 30292;
prmt.b32 %r225, %r220, %r223, %r224;
shr.u32 %r226, %r213, 16;
and.b32 %r227, %r226, 30583;
prmt.b32 %r228, %r219, %r218, %r227;
prmt.b32 %r229, %r222, %r221, %r227;
shr.u32 %r230, %r216, 16;
prmt.b32 %r231, %r228, %r229, %r230;
mov.u32 %r232, 25632;
prmt.b32 %r173, %r225, %r231, %r232;
mov.u32 %r233, 30001;
prmt.b32 %r177, %r225, %r231, %r233;
ld.global.nc.u32 %r72, [%rd33+4];
mov.u32 %r192, 0;

dp4a.s32.s32 %r70, %r173, %r72, %r192;

ld.global.nc.u32 %r76, [%rd33+20];

dp4a.s32.s32 %r74, %r177, %r76, %r70;

ld.global.nc.u8 %rs13, [%rd80+1];
cvt.u32.u16 %r234, %rs13;
and.b32 %r235, %r234, 255;
ld.global.nc.u8 %rs14, [%rd80+2];
cvt.u32.u16 %r236, %rs14;
prmt.b32 %r237, %r236, %r235, 30212;
ld.global.nc.u8 %rs15, [%rd80+3];
cvt.u32.u16 %r238, %rs15;
prmt.b32 %r239, %r238, %r237, 28756;
ld.global.nc.u8 %rs16, [%rd80+4];
cvt.u32.u16 %r240, %rs16;
prmt.b32 %r241, %r240, %r239, 1620;
and.b32 %r242, %r241, -2004318072;
shr.u32 %r243, %r242, 1;
or.b32 %r244, %r243, 839922192;
and.b32 %r245, %r237, 30583;
prmt.b32 %r246, %r219, %r218, %r245;
prmt.b32 %r247, %r222, %r221, %r245;
and.b32 %r248, %r244, 30292;
prmt.b32 %r249, %r246, %r247, %r248;
shr.u32 %r250, %r241, 16;
and.b32 %r251, %r250, 30583;
prmt.b32 %r252, %r219, %r218, %r251;
prmt.b32 %r253, %r222, %r221, %r251;
shr.u32 %r254, %r244, 16;
prmt.b32 %r255, %r252, %r253, %r254;
prmt.b32 %r181, %r249, %r255, %r232;
prmt.b32 %r185, %r249, %r255, %r233;
ld.global.nc.u32 %r80, [%rd33+8];

dp4a.s32.s32 %r78, %r181, %r80, %r74;

ld.global.nc.u32 %r84, [%rd33+24];

dp4a.s32.s32 %r82, %r185, %r84, %r78;

ld.global.nc.u8 %rs17, [%rd81+-4];
shl.b16 %rs19, %rs17, 7;
and.b16 %rs20, %rs19, 32640;
setp.eq.s16 %p3, %rs17, 255;
setp.eq.s16 %p4, %rs17, 0;
selp.b16 %rs21, 64, %rs20, %p4;
selp.b16 %rs7, 32767, %rs21, %p3;

{ mov.b32 %f49, {0,%rs7};}


mul.ftz.f32 %f65, %f49, 0f3F000000;
ld.global.nc.u32 %r86, [%rd31];

{.reg .f16 low,high;
mov.b32 {low,high},%r86;
cvt.f32.f16 %f50, low;}


mul.ftz.f32 %f66, %f65, %f50;
cvt.rn.f32.s32 %f67, %r82;
fma.rn.ftz.f32 %f224, %f66, %f67, %f224;
add.s64 %rd34, %rd82, %rd5;
ld.global.nc.u8 %rs22, [%rd34+-3];
cvt.u32.u16 %r256, %rs22;
and.b32 %r257, %r256, 255;
ld.global.nc.u8 %rs23, [%rd34+-2];
cvt.u32.u16 %r258, %rs23;
prmt.b32 %r259, %r258, %r257, 30212;
ld.global.nc.u8 %rs24, [%rd34+-1];
cvt.u32.u16 %r260, %rs24;
prmt.b32 %r261, %r260, %r259, 28756;
ld.global.nc.u8 %rs25, [%rd34];
cvt.u32.u16 %r262, %rs25;
prmt.b32 %r263, %r262, %r261, 1620;
and.b32 %r264, %r263, -2004318072;
shr.u32 %r265, %r264, 1;
or.b32 %r266, %r265, 839922192;
and.b32 %r267, %r259, 30583;
prmt.b32 %r268, %r219, %r218, %r267;
prmt.b32 %r269, %r222, %r221, %r267;
and.b32 %r270, %r266, 30292;
prmt.b32 %r271, %r268, %r269, %r270;
shr.u32 %r272, %r263, 16;
and.b32 %r273, %r272, 30583;
prmt.b32 %r274, %r219, %r218, %r273;
prmt.b32 %r275, %r222, %r221, %r273;
shr.u32 %r276, %r266, 16;
prmt.b32 %r277, %r274, %r275, %r276;
prmt.b32 %r190, %r271, %r277, %r232;
prmt.b32 %r194, %r271, %r277, %r233;

dp4a.s32.s32 %r87, %r190, %r72, %r192;


dp4a.s32.s32 %r91, %r194, %r76, %r87;

ld.global.nc.u8 %rs26, [%rd34+1];
cvt.u32.u16 %r278, %rs26;
and.b32 %r279, %r278, 255;
ld.global.nc.u8 %rs27, [%rd34+2];
cvt.u32.u16 %r280, %rs27;
prmt.b32 %r281, %r280, %r279, 30212;
ld.global.nc.u8 %rs28, [%rd34+3];
cvt.u32.u16 %r282, %rs28;
prmt.b32 %r283, %r282, %r281, 28756;
ld.global.nc.u8 %rs29, [%rd34+4];
cvt.u32.u16 %r284, %rs29;
prmt.b32 %r285, %r284, %r283, 1620;
and.b32 %r286, %r285, -2004318072;
shr.u32 %r287, %r286, 1;
or.b32 %r288, %r287, 839922192;
and.b32 %r289, %r281, 30583;
prmt.b32 %r290, %r219, %r218, %r289;
prmt.b32 %r291, %r222, %r221, %r289;
and.b32 %r292, %r288, 30292;
prmt.b32 %r293, %r290, %r291, %r292;
shr.u32 %r294, %r285, 16;
and.b32 %r295, %r294, 30583;
prmt.b32 %r296, %r219, %r218, %r295;
prmt.b32 %r297, %r222, %r221, %r295;
shr.u32 %r298, %r288, 16;
prmt.b32 %r299, %r296, %r297, %r298;
prmt.b32 %r198, %r293, %r299, %r232;
prmt.b32 %r202, %r293, %r299, %r233;

dp4a.s32.s32 %r95, %r198, %r80, %r91;


dp4a.s32.s32 %r99, %r202, %r84, %r95;

ld.global.nc.u8 %rs30, [%rd82];
shl.b16 %rs32, %rs30, 7;
and.b16 %rs33, %rs32, 32640;
setp.eq.s16 %p5, %rs30, 255;
setp.eq.s16 %p6, %rs30, 0;
selp.b16 %rs34, 64, %rs33, %p6;
selp.b16 %rs8, 32767, %rs34, %p5;

{ mov.b32 %f51, {0,%rs8};}


mul.ftz.f32 %f68, %f51, 0f3F000000;

{.reg .f16 low,high;
mov.b32 {low,high},%r86;
cvt.f32.f16 %f52, low;}


mul.ftz.f32 %f69, %f68, %f52;
cvt.rn.f32.s32 %f70, %r99;
fma.rn.ftz.f32 %f223, %f69, %f70, %f223;
cvt.u64.u32 %rd35, %r504;
add.s64 %rd36, %rd6, %rd35;
mul.lo.s64 %rd37, %rd36, 36;
add.s64 %rd38, %rd7, %rd37;
add.s64 %rd39, %rd38, %rd32;
ld.global.nc.u32 %r106, [%rd39+4];

dp4a.s32.s32 %r104, %r173, %r106, %r192;

ld.global.nc.u32 %r110, [%rd39+20];

dp4a.s32.s32 %r108, %r177, %r110, %r104;

ld.global.nc.u32 %r114, [%rd39+8];

dp4a.s32.s32 %r112, %r181, %r114, %r108;

ld.global.nc.u32 %r118, [%rd39+24];

dp4a.s32.s32 %r116, %r185, %r118, %r112;


{ mov.b32 %f53, {0,%rs7};}


mul.ftz.f32 %f71, %f53, 0f3F000000;
ld.global.nc.u32 %r120, [%rd38];

{.reg .f16 low,high;
mov.b32 {low,high},%r120;
cvt.f32.f16 %f54, low;}


mul.ftz.f32 %f72, %f71, %f54;
cvt.rn.f32.s32 %f73, %r116;
fma.rn.ftz.f32 %f222, %f72, %f73, %f222;

dp4a.s32.s32 %r121, %r190, %r106, %r192;


dp4a.s32.s32 %r125, %r194, %r110, %r121;


dp4a.s32.s32 %r129, %r198, %r114, %r125;


dp4a.s32.s32 %r133, %r202, %r118, %r129;


{ mov.b32 %f55, {0,%rs8};}


mul.ftz.f32 %f74, %f55, 0f3F000000;

{.reg .f16 low,high;
mov.b32 {low,high},%r120;
cvt.f32.f16 %f56, low;}


mul.ftz.f32 %f75, %f74, %f56;
cvt.rn.f32.s32 %f76, %r133;
fma.rn.ftz.f32 %f221, %f75, %f76, %f221;
cvt.u64.u32 %rd40, %r506;
add.s64 %rd41, %rd6, %rd40;
mul.lo.s64 %rd42, %rd41, 36;
add.s64 %rd43, %rd7, %rd42;
add.s64 %rd44, %rd43, %rd32;
ld.global.nc.u32 %r140, [%rd44+4];

dp4a.s32.s32 %r138, %r173, %r140, %r192;

ld.global.nc.u32 %r144, [%rd44+20];

dp4a.s32.s32 %r142, %r177, %r144, %r138;

ld.global.nc.u32 %r148, [%rd44+8];

dp4a.s32.s32 %r146, %r181, %r148, %r142;

ld.global.nc.u32 %r152, [%rd44+24];

dp4a.s32.s32 %r150, %r185, %r152, %r146;


{ mov.b32 %f57, {0,%rs7};}


mul.ftz.f32 %f77, %f57, 0f3F000000;
ld.global.nc.u32 %r154, [%rd43];

{.reg .f16 low,high;
mov.b32 {low,high},%r154;
cvt.f32.f16 %f58, low;}


mul.ftz.f32 %f78, %f77, %f58;
cvt.rn.f32.s32 %f79, %r150;
fma.rn.ftz.f32 %f220, %f78, %f79, %f220;

dp4a.s32.s32 %r155, %r190, %r140, %r192;


dp4a.s32.s32 %r159, %r194, %r144, %r155;


dp4a.s32.s32 %r163, %r198, %r148, %r159;


dp4a.s32.s32 %r167, %r202, %r152, %r163;


{ mov.b32 %f59, {0,%rs8};}


mul.ftz.f32 %f80, %f59, 0f3F000000;

{.reg .f16 low,high;
mov.b32 {low,high},%r154;
cvt.f32.f16 %f60, low;}


mul.ftz.f32 %f81, %f80, %f60;
cvt.rn.f32.s32 %f82, %r167;
fma.rn.ftz.f32 %f219, %f81, %f82, %f219;
cvt.u64.u32 %rd45, %r505;
add.s64 %rd46, %rd6, %rd45;
mul.lo.s64 %rd47, %rd46, 36;
add.s64 %rd48, %rd7, %rd47;
add.s64 %rd49, %rd48, %rd32;
ld.global.nc.u32 %r174, [%rd49+4];

dp4a.s32.s32 %r172, %r173, %r174, %r192;

ld.global.nc.u32 %r178, [%rd49+20];

dp4a.s32.s32 %r176, %r177, %r178, %r172;

ld.global.nc.u32 %r182, [%rd49+8];

dp4a.s32.s32 %r180, %r181, %r182, %r176;

ld.global.nc.u32 %r186, [%rd49+24];

dp4a.s32.s32 %r184, %r185, %r186, %r180;


{ mov.b32 %f61, {0,%rs7};}


mul.ftz.f32 %f83, %f61, 0f3F000000;
ld.global.nc.u32 %r188, [%rd48];

{.reg .f16 low,high;
mov.b32 {low,high},%r188;
cvt.f32.f16 %f62, low;}


mul.ftz.f32 %f84, %f83, %f62;
cvt.rn.f32.s32 %f85, %r184;
fma.rn.ftz.f32 %f218, %f84, %f85, %f218;

dp4a.s32.s32 %r189, %r190, %r174, %r192;


dp4a.s32.s32 %r193, %r194, %r178, %r189;


dp4a.s32.s32 %r197, %r198, %r182, %r193;


dp4a.s32.s32 %r201, %r202, %r186, %r197;


{ mov.b32 %f63, {0,%rs8};}


mul.ftz.f32 %f86, %f63, 0f3F000000;

{.reg .f16 low,high;
mov.b32 {low,high},%r188;
cvt.f32.f16 %f64, low;}


mul.ftz.f32 %f87, %f86, %f64;
cvt.rn.f32.s32 %f88, %r201;
fma.rn.ftz.f32 %f217, %f87, %f88, %f217;
add.s32 %r506, %r506, 64;
add.s32 %r505, %r505, 64;
add.s32 %r504, %r504, 64;
add.s64 %rd82, %rd82, 1088;
add.s64 %rd81, %rd81, 1088;
add.s64 %rd80, %rd80, 1088;
add.s32 %r507, %r507, 64;
setp.lt.s32 %p7, %r507, %r37;
@%p7 bra $L__BB109_2;

$L__BB109_3:
setp.eq.s32 %p8, %r1, 0;
@%p8 bra $L__BB109_5;

shl.b32 %r301, %r1, 10;
mov.u32 %r302, _ZZ13mul_mat_vec_qIL9ggml_type39ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r303, %r302, %r301;
shl.b32 %r304, %r33, 2;
add.s32 %r305, %r303, %r304;
st.shared.f32 [%r305+-1024], %f224;
st.shared.f32 [%r305+-896], %f223;
st.shared.f32 [%r305+-768], %f222;
st.shared.f32 [%r305+-640], %f221;
st.shared.f32 [%r305+-512], %f220;
st.shared.f32 [%r305+-384], %f219;
st.shared.f32 [%r305+-256], %f218;
st.shared.f32 [%r305+-128], %f217;

$L__BB109_5:
setp.ne.s32 %p9, %r1, 0;
bar.sync 0;
@%p9 bra $L__BB109_22;

mov.u32 %r306, %ctaid.z;
mov.u32 %r307, %ctaid.y;
mov.u32 %r308, %ctaid.x;
shl.b32 %r309, %r308, 1;
mov.u32 %r310, 1;
mad.lo.s32 %r311, %r307, %r25, %r309;
mad.lo.s32 %r312, %r306, %r31, %r311;
cvt.u64.u32 %rd14, %r312;
shl.b32 %r314, %r33, 2;
mov.u32 %r315, _ZZ13mul_mat_vec_qIL9ggml_type39ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r316, %r315, %r314;
mov.u32 %r317, 2;
ld.shared.f32 %f89, [%r316];
add.ftz.f32 %f90, %f89, %f224;
ld.shared.f32 %f91, [%r316+1024];
add.ftz.f32 %f92, %f91, %f90;
ld.shared.f32 %f93, [%r316+2048];
add.ftz.f32 %f94, %f93, %f92;
mov.b32 %r318, %f94;
mov.u32 %r319, 31;
mov.u32 %r320, 16;
mov.u32 %r321, -1;
shfl.sync.bfly.b32 %r322|%p10, %r318, %r320, %r319, %r321;
mov.b32 %f95, %r322;
add.ftz.f32 %f96, %f94, %f95;
mov.b32 %r323, %f96;
mov.u32 %r324, 8;
shfl.sync.bfly.b32 %r325|%p11, %r323, %r324, %r319, %r321;
mov.b32 %f97, %r325;
add.ftz.f32 %f98, %f96, %f97;
mov.b32 %r326, %f98;
mov.u32 %r327, 4;
shfl.sync.bfly.b32 %r328|%p12, %r326, %r327, %r319, %r321;
mov.b32 %f99, %r328;
add.ftz.f32 %f100, %f98, %f99;
mov.b32 %r329, %f100;
shfl.sync.bfly.b32 %r330|%p13, %r329, %r317, %r319, %r321;
mov.b32 %f101, %r330;
add.ftz.f32 %f102, %f100, %f101;
mov.b32 %r331, %f102;
shfl.sync.bfly.b32 %r332|%p14, %r331, %r310, %r319, %r321;
mov.b32 %f103, %r332;
add.ftz.f32 %f25, %f102, %f103;
setp.ne.s32 %p15, %r33, 0;
setp.ge.u32 %p16, %r309, %r19;
or.pred %p17, %p15, %p16;
@%p17 bra $L__BB109_8;

cvta.to.global.u64 %rd50, %rd18;
mul.wide.u32 %rd51, %r312, 4;
add.s64 %rd52, %rd50, %rd51;
st.global.f32 [%rd52], %f25;

$L__BB109_8:
ld.shared.f32 %f104, [%r316+128];
add.ftz.f32 %f105, %f104, %f223;
ld.shared.f32 %f106, [%r316+1152];
add.ftz.f32 %f107, %f106, %f105;
ld.shared.f32 %f108, [%r316+2176];
add.ftz.f32 %f109, %f108, %f107;
mov.b32 %r344, %f109;
shfl.sync.bfly.b32 %r348|%p18, %r344, %r320, %r319, %r321;
mov.b32 %f110, %r348;
add.ftz.f32 %f111, %f109, %f110;
mov.b32 %r349, %f111;
shfl.sync.bfly.b32 %r351|%p19, %r349, %r324, %r319, %r321;
mov.b32 %f112, %r351;
add.ftz.f32 %f113, %f111, %f112;
mov.b32 %r352, %f113;
shfl.sync.bfly.b32 %r354|%p20, %r352, %r327, %r319, %r321;
mov.b32 %f114, %r354;
add.ftz.f32 %f115, %f113, %f114;
mov.b32 %r355, %f115;
shfl.sync.bfly.b32 %r356|%p21, %r355, %r317, %r319, %r321;
mov.b32 %f116, %r356;
add.ftz.f32 %f117, %f115, %f116;
mov.b32 %r357, %f117;
shfl.sync.bfly.b32 %r359|%p22, %r357, %r310, %r319, %r321;
mov.b32 %f118, %r359;
add.ftz.f32 %f26, %f117, %f118;
setp.eq.s32 %p23, %r33, 1;
or.b32 %r362, %r309, 1;
setp.lt.u32 %p24, %r362, %r19;
and.pred %p1, %p23, %p24;
not.pred %p25, %p1;
@%p25 bra $L__BB109_10;

cvta.to.global.u64 %rd53, %rd18;
mul.wide.u32 %rd54, %r312, 4;
add.s64 %rd55, %rd53, %rd54;
st.global.f32 [%rd55+4], %f26;

$L__BB109_10:
mov.u32 %r373, 2;
ld.shared.f32 %f119, [%r316+256];
add.ftz.f32 %f120, %f119, %f222;
ld.shared.f32 %f121, [%r316+1280];
add.ftz.f32 %f122, %f121, %f120;
ld.shared.f32 %f123, [%r316+2304];
add.ftz.f32 %f124, %f123, %f122;
mov.b32 %r374, %f124;
mov.u32 %r375, 31;
mov.u32 %r376, 16;
mov.u32 %r377, -1;
shfl.sync.bfly.b32 %r378|%p26, %r374, %r376, %r375, %r377;
mov.b32 %f125, %r378;
add.ftz.f32 %f126, %f124, %f125;
mov.b32 %r379, %f126;
mov.u32 %r380, 8;
shfl.sync.bfly.b32 %r381|%p27, %r379, %r380, %r375, %r377;
mov.b32 %f127, %r381;
add.ftz.f32 %f128, %f126, %f127;
mov.b32 %r382, %f128;
mov.u32 %r383, 4;
shfl.sync.bfly.b32 %r384|%p28, %r382, %r383, %r375, %r377;
mov.b32 %f129, %r384;
add.ftz.f32 %f130, %f128, %f129;
mov.b32 %r385, %f130;
shfl.sync.bfly.b32 %r386|%p29, %r385, %r373, %r375, %r377;
mov.b32 %f131, %r386;
add.ftz.f32 %f132, %f130, %f131;
mov.b32 %r387, %f132;
mov.u32 %r388, 1;
shfl.sync.bfly.b32 %r389|%p30, %r387, %r388, %r375, %r377;
mov.b32 %f133, %r389;
add.ftz.f32 %f27, %f132, %f133;
@%p17 bra $L__BB109_12;

cvt.u64.u32 %rd56, %r19;
add.s64 %rd57, %rd56, %rd14;
cvta.to.global.u64 %rd58, %rd18;
shl.b64 %rd59, %rd57, 2;
add.s64 %rd60, %rd58, %rd59;
st.global.f32 [%rd60], %f27;

$L__BB109_12:
ld.shared.f32 %f134, [%r316+384];
add.ftz.f32 %f135, %f134, %f221;
ld.shared.f32 %f136, [%r316+1408];
add.ftz.f32 %f137, %f136, %f135;
ld.shared.f32 %f138, [%r316+2432];
add.ftz.f32 %f139, %f138, %f137;
mov.b32 %r397, %f139;
shfl.sync.bfly.b32 %r401|%p34, %r397, %r376, %r375, %r377;
mov.b32 %f140, %r401;
add.ftz.f32 %f141, %f139, %f140;
mov.b32 %r402, %f141;
shfl.sync.bfly.b32 %r404|%p35, %r402, %r380, %r375, %r377;
mov.b32 %f142, %r404;
add.ftz.f32 %f143, %f141, %f142;
mov.b32 %r405, %f143;
shfl.sync.bfly.b32 %r407|%p36, %r405, %r383, %r375, %r377;
mov.b32 %f144, %r407;
add.ftz.f32 %f145, %f143, %f144;
mov.b32 %r408, %f145;
shfl.sync.bfly.b32 %r409|%p37, %r408, %r373, %r375, %r377;
mov.b32 %f146, %r409;
add.ftz.f32 %f147, %f145, %f146;
mov.b32 %r410, %f147;
shfl.sync.bfly.b32 %r412|%p38, %r410, %r388, %r375, %r377;
mov.b32 %f148, %r412;
add.ftz.f32 %f28, %f147, %f148;
@%p25 bra $L__BB109_14;

add.s32 %r413, %r19, 1;
cvt.u64.u32 %rd61, %r413;
add.s64 %rd62, %rd61, %rd14;
cvta.to.global.u64 %rd63, %rd18;
shl.b64 %rd64, %rd62, 2;
add.s64 %rd65, %rd63, %rd64;
st.global.f32 [%rd65], %f28;

$L__BB109_14:
shl.b32 %r414, %r19, 1;
mov.u32 %r415, 1;
mov.u32 %r420, 2;
ld.shared.f32 %f149, [%r316+512];
add.ftz.f32 %f150, %f149, %f220;
ld.shared.f32 %f151, [%r316+1536];
add.ftz.f32 %f152, %f151, %f150;
ld.shared.f32 %f153, [%r316+2560];
add.ftz.f32 %f154, %f153, %f152;
mov.b32 %r421, %f154;
mov.u32 %r422, 31;
mov.u32 %r423, 16;
mov.u32 %r424, -1;
shfl.sync.bfly.b32 %r425|%p40, %r421, %r423, %r422, %r424;
mov.b32 %f155, %r425;
add.ftz.f32 %f156, %f154, %f155;
mov.b32 %r426, %f156;
mov.u32 %r427, 8;
shfl.sync.bfly.b32 %r428|%p41, %r426, %r427, %r422, %r424;
mov.b32 %f157, %r428;
add.ftz.f32 %f158, %f156, %f157;
mov.b32 %r429, %f158;
mov.u32 %r430, 4;
shfl.sync.bfly.b32 %r431|%p42, %r429, %r430, %r422, %r424;
mov.b32 %f159, %r431;
add.ftz.f32 %f160, %f158, %f159;
mov.b32 %r432, %f160;
shfl.sync.bfly.b32 %r433|%p43, %r432, %r420, %r422, %r424;
mov.b32 %f161, %r433;
add.ftz.f32 %f162, %f160, %f161;
mov.b32 %r434, %f162;
shfl.sync.bfly.b32 %r435|%p44, %r434, %r415, %r422, %r424;
mov.b32 %f163, %r435;
add.ftz.f32 %f29, %f162, %f163;
cvt.u64.u32 %rd66, %r414;
add.s64 %rd67, %rd66, %rd14;
cvta.to.global.u64 %rd68, %rd18;
shl.b64 %rd69, %rd67, 2;
add.s64 %rd15, %rd68, %rd69;
@%p17 bra $L__BB109_16;

st.global.f32 [%rd15], %f29;

$L__BB109_16:
ld.shared.f32 %f164, [%r316+640];
add.ftz.f32 %f165, %f164, %f219;
ld.shared.f32 %f166, [%r316+1664];
add.ftz.f32 %f167, %f166, %f165;
ld.shared.f32 %f168, [%r316+2688];
add.ftz.f32 %f169, %f168, %f167;
mov.b32 %r443, %f169;
shfl.sync.bfly.b32 %r447|%p48, %r443, %r423, %r422, %r424;
mov.b32 %f170, %r447;
add.ftz.f32 %f171, %f169, %f170;
mov.b32 %r448, %f171;
shfl.sync.bfly.b32 %r450|%p49, %r448, %r427, %r422, %r424;
mov.b32 %f172, %r450;
add.ftz.f32 %f173, %f171, %f172;
mov.b32 %r451, %f173;
shfl.sync.bfly.b32 %r453|%p50, %r451, %r430, %r422, %r424;
mov.b32 %f174, %r453;
add.ftz.f32 %f175, %f173, %f174;
mov.b32 %r454, %f175;
shfl.sync.bfly.b32 %r455|%p51, %r454, %r420, %r422, %r424;
mov.b32 %f176, %r455;
add.ftz.f32 %f177, %f175, %f176;
mov.b32 %r456, %f177;
shfl.sync.bfly.b32 %r458|%p52, %r456, %r415, %r422, %r424;
mov.b32 %f178, %r458;
add.ftz.f32 %f30, %f177, %f178;
@%p25 bra $L__BB109_18;

st.global.f32 [%rd15+4], %f30;

$L__BB109_18:
mul.lo.s32 %r15, %r19, 3;
mov.u32 %r463, 2;
ld.shared.f32 %f179, [%r316+768];
add.ftz.f32 %f180, %f179, %f218;
ld.shared.f32 %f181, [%r316+1792];
add.ftz.f32 %f182, %f181, %f180;
ld.shared.f32 %f183, [%r316+2816];
add.ftz.f32 %f184, %f183, %f182;
mov.b32 %r464, %f184;
mov.u32 %r465, 31;
mov.u32 %r466, 16;
mov.u32 %r467, -1;
shfl.sync.bfly.b32 %r468|%p54, %r464, %r466, %r465, %r467;
mov.b32 %f185, %r468;
add.ftz.f32 %f186, %f184, %f185;
mov.b32 %r469, %f186;
mov.u32 %r470, 8;
shfl.sync.bfly.b32 %r471|%p55, %r469, %r470, %r465, %r467;
mov.b32 %f187, %r471;
add.ftz.f32 %f188, %f186, %f187;
mov.b32 %r472, %f188;
mov.u32 %r473, 4;
shfl.sync.bfly.b32 %r474|%p56, %r472, %r473, %r465, %r467;
mov.b32 %f189, %r474;
add.ftz.f32 %f190, %f188, %f189;
mov.b32 %r475, %f190;
shfl.sync.bfly.b32 %r476|%p57, %r475, %r463, %r465, %r467;
mov.b32 %f191, %r476;
add.ftz.f32 %f192, %f190, %f191;
mov.b32 %r477, %f192;
mov.u32 %r478, 1;
shfl.sync.bfly.b32 %r479|%p58, %r477, %r478, %r465, %r467;
mov.b32 %f193, %r479;
add.ftz.f32 %f31, %f192, %f193;
@%p17 bra $L__BB109_20;

cvt.u64.u32 %rd70, %r15;
add.s64 %rd71, %rd70, %rd14;
shl.b64 %rd73, %rd71, 2;
add.s64 %rd74, %rd68, %rd73;
st.global.f32 [%rd74], %f31;

$L__BB109_20:
ld.shared.f32 %f194, [%r316+896];
add.ftz.f32 %f195, %f194, %f217;
ld.shared.f32 %f196, [%r316+1920];
add.ftz.f32 %f197, %f196, %f195;
ld.shared.f32 %f198, [%r316+2944];
add.ftz.f32 %f199, %f198, %f197;
mov.b32 %r487, %f199;
shfl.sync.bfly.b32 %r491|%p62, %r487, %r466, %r465, %r467;
mov.b32 %f200, %r491;
add.ftz.f32 %f201, %f199, %f200;
mov.b32 %r492, %f201;
shfl.sync.bfly.b32 %r494|%p63, %r492, %r470, %r465, %r467;
mov.b32 %f202, %r494;
add.ftz.f32 %f203, %f201, %f202;
mov.b32 %r495, %f203;
shfl.sync.bfly.b32 %r497|%p64, %r495, %r473, %r465, %r467;
mov.b32 %f204, %r497;
add.ftz.f32 %f205, %f203, %f204;
mov.b32 %r498, %f205;
shfl.sync.bfly.b32 %r499|%p65, %r498, %r463, %r465, %r467;
mov.b32 %f206, %r499;
add.ftz.f32 %f207, %f205, %f206;
mov.b32 %r500, %f207;
shfl.sync.bfly.b32 %r502|%p66, %r500, %r478, %r465, %r467;
mov.b32 %f208, %r502;
add.ftz.f32 %f32, %f207, %f208;
@%p25 bra $L__BB109_22;

add.s32 %r503, %r15, 1;
cvt.u64.u32 %rd75, %r503;
add.s64 %rd76, %rd75, %rd14;
shl.b64 %rd78, %rd76, 2;
add.s64 %rd79, %rd68, %rd78;
st.global.f32 [%rd79], %f32;

$L__BB109_22:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type39ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type39ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type39ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type39ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type39ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type39ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type39ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type39ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type39ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<82>;
.reg .b16 %rs<37>;
.reg .f32 %f<241>;
.reg .b32 %r<617>;
.reg .b64 %rd<98>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type39ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1280];

ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type39ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type39ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type39ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type39ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type39ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type39ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type39ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type39ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type39ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type39ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type39ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type39ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type39ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type39ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type39ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type39ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type39ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r36, %r1, 5;
mov.u32 %r37, %tid.x;
add.s32 %r2, %r36, %r37;
shr.u32 %r38, %r2, 31;
add.s32 %r39, %r2, %r38;
shr.s32 %r616, %r39, 1;
shr.u32 %r40, %r20, 5;
setp.ge.s32 %p2, %r616, %r40;
mov.f32 %f231, 0f00000000;
mov.f32 %f232, %f231;
mov.f32 %f233, %f231;
mov.f32 %f234, %f231;
mov.f32 %f235, %f231;
mov.f32 %f236, %f231;
mov.f32 %f237, %f231;
mov.f32 %f238, %f231;
mov.f32 %f239, %f231;
mov.f32 %f240, %f231;
@%p2 bra $L__BB110_3;

mov.u32 %r41, %ctaid.y;
mov.u32 %r42, %ctaid.z;
mul.hi.u32 %r43, %r42, %r30;
add.s32 %r44, %r43, %r42;
shr.u32 %r45, %r44, %r31;
mul.lo.s32 %r46, %r45, %r33;
mul.hi.u32 %r47, %r41, %r24;
add.s32 %r48, %r47, %r41;
shr.u32 %r49, %r48, %r25;
mul.lo.s32 %r50, %r49, %r27;
cvta.to.global.u64 %rd19, %rd16;
and.b32 %r53, %r39, -2;
sub.s32 %r54, %r2, %r53;
shl.b32 %r55, %r54, 1;
cvt.s64.s32 %rd1, %r55;
mov.u32 %r56, %ctaid.x;
shl.b32 %r57, %r56, 1;
shl.b32 %r58, %r54, 3;
cvt.s64.s32 %rd20, %r58;
shl.b32 %r59, %r22, 1;
add.s32 %r615, %r616, %r59;
mad.lo.s32 %r614, %r22, 3, %r616;
shl.b32 %r60, %r22, 2;
add.s32 %r613, %r616, %r60;
add.s32 %r612, %r616, %r22;
add.s32 %r61, %r616, %r46;
add.s32 %r62, %r61, %r50;
or.b32 %r63, %r57, 1;
mad.lo.s32 %r64, %r21, %r63, %r62;
mul.wide.s32 %rd21, %r64, 17;
add.s64 %rd97, %rd19, %rd21;
mad.lo.s32 %r65, %r57, %r21, %r616;
add.s32 %r66, %r65, %r46;
add.s32 %r67, %r66, %r50;
mul.wide.s32 %rd22, %r67, 17;
add.s64 %rd23, %rd19, %rd22;
add.s64 %rd96, %rd23, 4;
add.s64 %rd24, %rd22, %rd20;
add.s64 %rd25, %rd19, %rd24;
add.s64 %rd95, %rd25, 4;
or.b64 %rd5, %rd20, 4;
mul.lo.s32 %r68, %r41, %r28;
cvt.u64.u32 %rd26, %r68;
mul.lo.s32 %r69, %r42, %r34;
cvt.u64.u32 %rd27, %r69;
add.s64 %rd6, %rd27, %rd26;
cvta.to.global.u64 %rd7, %rd17;
mov.f32 %f231, 0f00000000;
shl.b64 %rd32, %rd1, 2;

$L__BB110_2:
cvt.u64.u32 %rd28, %r616;
add.s64 %rd29, %rd6, %rd28;
mul.lo.s64 %rd30, %rd29, 36;
add.s64 %rd31, %rd7, %rd30;
add.s64 %rd33, %rd31, %rd32;
ld.global.nc.u8 %rs11, [%rd95+-3];
cvt.u32.u16 %r240, %rs11;
and.b32 %r241, %r240, 255;
ld.global.nc.u8 %rs12, [%rd95+-2];
cvt.u32.u16 %r242, %rs12;
prmt.b32 %r243, %r242, %r241, 30212;
ld.global.nc.u8 %rs13, [%rd95+-1];
cvt.u32.u16 %r244, %rs13;
prmt.b32 %r245, %r244, %r243, 28756;
ld.global.nc.u8 %rs14, [%rd95];
cvt.u32.u16 %r246, %rs14;
prmt.b32 %r247, %r246, %r245, 1620;
and.b32 %r248, %r247, -2004318072;
shr.u32 %r249, %r248, 1;
or.b32 %r250, %r249, 839922192;
and.b32 %r251, %r243, 30583;
mov.u32 %r252, 201852420;
mov.u32 %r253, 50462976;
prmt.b32 %r254, %r253, %r252, %r251;
mov.u32 %r255, -185009412;
mov.u32 %r256, -33620224;
prmt.b32 %r257, %r256, %r255, %r251;
and.b32 %r258, %r250, 30292;
prmt.b32 %r259, %r254, %r257, %r258;
shr.u32 %r260, %r247, 16;
and.b32 %r261, %r260, 30583;
prmt.b32 %r262, %r253, %r252, %r261;
prmt.b32 %r263, %r256, %r255, %r261;
shr.u32 %r264, %r250, 16;
prmt.b32 %r265, %r262, %r263, %r264;
mov.u32 %r266, 25632;
prmt.b32 %r207, %r259, %r265, %r266;
mov.u32 %r267, 30001;
prmt.b32 %r211, %r259, %r265, %r267;
ld.global.nc.u32 %r72, [%rd33+4];
mov.u32 %r226, 0;

dp4a.s32.s32 %r70, %r207, %r72, %r226;

ld.global.nc.u32 %r76, [%rd33+20];

dp4a.s32.s32 %r74, %r211, %r76, %r70;

ld.global.nc.u8 %rs15, [%rd95+1];
cvt.u32.u16 %r268, %rs15;
and.b32 %r269, %r268, 255;
ld.global.nc.u8 %rs16, [%rd95+2];
cvt.u32.u16 %r270, %rs16;
prmt.b32 %r271, %r270, %r269, 30212;
ld.global.nc.u8 %rs17, [%rd95+3];
cvt.u32.u16 %r272, %rs17;
prmt.b32 %r273, %r272, %r271, 28756;
ld.global.nc.u8 %rs18, [%rd95+4];
cvt.u32.u16 %r274, %rs18;
prmt.b32 %r275, %r274, %r273, 1620;
and.b32 %r276, %r275, -2004318072;
shr.u32 %r277, %r276, 1;
or.b32 %r278, %r277, 839922192;
and.b32 %r279, %r271, 30583;
prmt.b32 %r280, %r253, %r252, %r279;
prmt.b32 %r281, %r256, %r255, %r279;
and.b32 %r282, %r278, 30292;
prmt.b32 %r283, %r280, %r281, %r282;
shr.u32 %r284, %r275, 16;
and.b32 %r285, %r284, 30583;
prmt.b32 %r286, %r253, %r252, %r285;
prmt.b32 %r287, %r256, %r255, %r285;
shr.u32 %r288, %r278, 16;
prmt.b32 %r289, %r286, %r287, %r288;
prmt.b32 %r215, %r283, %r289, %r266;
prmt.b32 %r219, %r283, %r289, %r267;
ld.global.nc.u32 %r80, [%rd33+8];

dp4a.s32.s32 %r78, %r215, %r80, %r74;

ld.global.nc.u32 %r84, [%rd33+24];

dp4a.s32.s32 %r82, %r219, %r84, %r78;

ld.global.nc.u8 %rs19, [%rd96+-4];
shl.b16 %rs21, %rs19, 7;
and.b16 %rs22, %rs21, 32640;
setp.eq.s16 %p3, %rs19, 255;
setp.eq.s16 %p4, %rs19, 0;
selp.b16 %rs23, 64, %rs22, %p4;
selp.b16 %rs9, 32767, %rs23, %p3;

{ mov.b32 %f61, {0,%rs9};}


mul.ftz.f32 %f81, %f61, 0f3F000000;
ld.global.nc.u32 %r86, [%rd31];

{.reg .f16 low,high;
mov.b32 {low,high},%r86;
cvt.f32.f16 %f62, low;}


mul.ftz.f32 %f82, %f81, %f62;
cvt.rn.f32.s32 %f83, %r82;
fma.rn.ftz.f32 %f240, %f82, %f83, %f240;
add.s64 %rd34, %rd97, %rd5;
ld.global.nc.u8 %rs24, [%rd34+-3];
cvt.u32.u16 %r290, %rs24;
and.b32 %r291, %r290, 255;
ld.global.nc.u8 %rs25, [%rd34+-2];
cvt.u32.u16 %r292, %rs25;
prmt.b32 %r293, %r292, %r291, 30212;
ld.global.nc.u8 %rs26, [%rd34+-1];
cvt.u32.u16 %r294, %rs26;
prmt.b32 %r295, %r294, %r293, 28756;
ld.global.nc.u8 %rs27, [%rd34];
cvt.u32.u16 %r296, %rs27;
prmt.b32 %r297, %r296, %r295, 1620;
and.b32 %r298, %r297, -2004318072;
shr.u32 %r299, %r298, 1;
or.b32 %r300, %r299, 839922192;
and.b32 %r301, %r293, 30583;
prmt.b32 %r302, %r253, %r252, %r301;
prmt.b32 %r303, %r256, %r255, %r301;
and.b32 %r304, %r300, 30292;
prmt.b32 %r305, %r302, %r303, %r304;
shr.u32 %r306, %r297, 16;
and.b32 %r307, %r306, 30583;
prmt.b32 %r308, %r253, %r252, %r307;
prmt.b32 %r309, %r256, %r255, %r307;
shr.u32 %r310, %r300, 16;
prmt.b32 %r311, %r308, %r309, %r310;
prmt.b32 %r224, %r305, %r311, %r266;
prmt.b32 %r228, %r305, %r311, %r267;

dp4a.s32.s32 %r87, %r224, %r72, %r226;


dp4a.s32.s32 %r91, %r228, %r76, %r87;

ld.global.nc.u8 %rs28, [%rd34+1];
cvt.u32.u16 %r312, %rs28;
and.b32 %r313, %r312, 255;
ld.global.nc.u8 %rs29, [%rd34+2];
cvt.u32.u16 %r314, %rs29;
prmt.b32 %r315, %r314, %r313, 30212;
ld.global.nc.u8 %rs30, [%rd34+3];
cvt.u32.u16 %r316, %rs30;
prmt.b32 %r317, %r316, %r315, 28756;
ld.global.nc.u8 %rs31, [%rd34+4];
cvt.u32.u16 %r318, %rs31;
prmt.b32 %r319, %r318, %r317, 1620;
and.b32 %r320, %r319, -2004318072;
shr.u32 %r321, %r320, 1;
or.b32 %r322, %r321, 839922192;
and.b32 %r323, %r315, 30583;
prmt.b32 %r324, %r253, %r252, %r323;
prmt.b32 %r325, %r256, %r255, %r323;
and.b32 %r326, %r322, 30292;
prmt.b32 %r327, %r324, %r325, %r326;
shr.u32 %r328, %r319, 16;
and.b32 %r329, %r328, 30583;
prmt.b32 %r330, %r253, %r252, %r329;
prmt.b32 %r331, %r256, %r255, %r329;
shr.u32 %r332, %r322, 16;
prmt.b32 %r333, %r330, %r331, %r332;
prmt.b32 %r232, %r327, %r333, %r266;
prmt.b32 %r236, %r327, %r333, %r267;

dp4a.s32.s32 %r95, %r232, %r80, %r91;


dp4a.s32.s32 %r99, %r236, %r84, %r95;

ld.global.nc.u8 %rs32, [%rd97];
shl.b16 %rs34, %rs32, 7;
and.b16 %rs35, %rs34, 32640;
setp.eq.s16 %p5, %rs32, 255;
setp.eq.s16 %p6, %rs32, 0;
selp.b16 %rs36, 64, %rs35, %p6;
selp.b16 %rs10, 32767, %rs36, %p5;

{ mov.b32 %f63, {0,%rs10};}


mul.ftz.f32 %f84, %f63, 0f3F000000;

{.reg .f16 low,high;
mov.b32 {low,high},%r86;
cvt.f32.f16 %f64, low;}


mul.ftz.f32 %f85, %f84, %f64;
cvt.rn.f32.s32 %f86, %r99;
fma.rn.ftz.f32 %f239, %f85, %f86, %f239;
cvt.u64.u32 %rd35, %r612;
add.s64 %rd36, %rd6, %rd35;
mul.lo.s64 %rd37, %rd36, 36;
add.s64 %rd38, %rd7, %rd37;
add.s64 %rd39, %rd38, %rd32;
ld.global.nc.u32 %r106, [%rd39+4];

dp4a.s32.s32 %r104, %r207, %r106, %r226;

ld.global.nc.u32 %r110, [%rd39+20];

dp4a.s32.s32 %r108, %r211, %r110, %r104;

ld.global.nc.u32 %r114, [%rd39+8];

dp4a.s32.s32 %r112, %r215, %r114, %r108;

ld.global.nc.u32 %r118, [%rd39+24];

dp4a.s32.s32 %r116, %r219, %r118, %r112;


{ mov.b32 %f65, {0,%rs9};}


mul.ftz.f32 %f87, %f65, 0f3F000000;
ld.global.nc.u32 %r120, [%rd38];

{.reg .f16 low,high;
mov.b32 {low,high},%r120;
cvt.f32.f16 %f66, low;}


mul.ftz.f32 %f88, %f87, %f66;
cvt.rn.f32.s32 %f89, %r116;
fma.rn.ftz.f32 %f238, %f88, %f89, %f238;

dp4a.s32.s32 %r121, %r224, %r106, %r226;


dp4a.s32.s32 %r125, %r228, %r110, %r121;


dp4a.s32.s32 %r129, %r232, %r114, %r125;


dp4a.s32.s32 %r133, %r236, %r118, %r129;


{ mov.b32 %f67, {0,%rs10};}


mul.ftz.f32 %f90, %f67, 0f3F000000;

{.reg .f16 low,high;
mov.b32 {low,high},%r120;
cvt.f32.f16 %f68, low;}


mul.ftz.f32 %f91, %f90, %f68;
cvt.rn.f32.s32 %f92, %r133;
fma.rn.ftz.f32 %f237, %f91, %f92, %f237;
cvt.u64.u32 %rd40, %r615;
add.s64 %rd41, %rd6, %rd40;
mul.lo.s64 %rd42, %rd41, 36;
add.s64 %rd43, %rd7, %rd42;
add.s64 %rd44, %rd43, %rd32;
ld.global.nc.u32 %r140, [%rd44+4];

dp4a.s32.s32 %r138, %r207, %r140, %r226;

ld.global.nc.u32 %r144, [%rd44+20];

dp4a.s32.s32 %r142, %r211, %r144, %r138;

ld.global.nc.u32 %r148, [%rd44+8];

dp4a.s32.s32 %r146, %r215, %r148, %r142;

ld.global.nc.u32 %r152, [%rd44+24];

dp4a.s32.s32 %r150, %r219, %r152, %r146;


{ mov.b32 %f69, {0,%rs9};}


mul.ftz.f32 %f93, %f69, 0f3F000000;
ld.global.nc.u32 %r154, [%rd43];

{.reg .f16 low,high;
mov.b32 {low,high},%r154;
cvt.f32.f16 %f70, low;}


mul.ftz.f32 %f94, %f93, %f70;
cvt.rn.f32.s32 %f95, %r150;
fma.rn.ftz.f32 %f236, %f94, %f95, %f236;

dp4a.s32.s32 %r155, %r224, %r140, %r226;


dp4a.s32.s32 %r159, %r228, %r144, %r155;


dp4a.s32.s32 %r163, %r232, %r148, %r159;


dp4a.s32.s32 %r167, %r236, %r152, %r163;


{ mov.b32 %f71, {0,%rs10};}


mul.ftz.f32 %f96, %f71, 0f3F000000;

{.reg .f16 low,high;
mov.b32 {low,high},%r154;
cvt.f32.f16 %f72, low;}


mul.ftz.f32 %f97, %f96, %f72;
cvt.rn.f32.s32 %f98, %r167;
fma.rn.ftz.f32 %f235, %f97, %f98, %f235;
cvt.u64.u32 %rd45, %r614;
add.s64 %rd46, %rd6, %rd45;
mul.lo.s64 %rd47, %rd46, 36;
add.s64 %rd48, %rd7, %rd47;
add.s64 %rd49, %rd48, %rd32;
ld.global.nc.u32 %r174, [%rd49+4];

dp4a.s32.s32 %r172, %r207, %r174, %r226;

ld.global.nc.u32 %r178, [%rd49+20];

dp4a.s32.s32 %r176, %r211, %r178, %r172;

ld.global.nc.u32 %r182, [%rd49+8];

dp4a.s32.s32 %r180, %r215, %r182, %r176;

ld.global.nc.u32 %r186, [%rd49+24];

dp4a.s32.s32 %r184, %r219, %r186, %r180;


{ mov.b32 %f73, {0,%rs9};}


mul.ftz.f32 %f99, %f73, 0f3F000000;
ld.global.nc.u32 %r188, [%rd48];

{.reg .f16 low,high;
mov.b32 {low,high},%r188;
cvt.f32.f16 %f74, low;}


mul.ftz.f32 %f100, %f99, %f74;
cvt.rn.f32.s32 %f101, %r184;
fma.rn.ftz.f32 %f234, %f100, %f101, %f234;

dp4a.s32.s32 %r189, %r224, %r174, %r226;


dp4a.s32.s32 %r193, %r228, %r178, %r189;


dp4a.s32.s32 %r197, %r232, %r182, %r193;


dp4a.s32.s32 %r201, %r236, %r186, %r197;


{ mov.b32 %f75, {0,%rs10};}


mul.ftz.f32 %f102, %f75, 0f3F000000;

{.reg .f16 low,high;
mov.b32 {low,high},%r188;
cvt.f32.f16 %f76, low;}


mul.ftz.f32 %f103, %f102, %f76;
cvt.rn.f32.s32 %f104, %r201;
fma.rn.ftz.f32 %f233, %f103, %f104, %f233;
cvt.u64.u32 %rd50, %r613;
add.s64 %rd51, %rd6, %rd50;
mul.lo.s64 %rd52, %rd51, 36;
add.s64 %rd53, %rd7, %rd52;
add.s64 %rd54, %rd53, %rd32;
ld.global.nc.u32 %r208, [%rd54+4];

dp4a.s32.s32 %r206, %r207, %r208, %r226;

ld.global.nc.u32 %r212, [%rd54+20];

dp4a.s32.s32 %r210, %r211, %r212, %r206;

ld.global.nc.u32 %r216, [%rd54+8];

dp4a.s32.s32 %r214, %r215, %r216, %r210;

ld.global.nc.u32 %r220, [%rd54+24];

dp4a.s32.s32 %r218, %r219, %r220, %r214;


{ mov.b32 %f77, {0,%rs9};}


mul.ftz.f32 %f105, %f77, 0f3F000000;
ld.global.nc.u32 %r222, [%rd53];

{.reg .f16 low,high;
mov.b32 {low,high},%r222;
cvt.f32.f16 %f78, low;}


mul.ftz.f32 %f106, %f105, %f78;
cvt.rn.f32.s32 %f107, %r218;
fma.rn.ftz.f32 %f232, %f106, %f107, %f232;

dp4a.s32.s32 %r223, %r224, %r208, %r226;


dp4a.s32.s32 %r227, %r228, %r212, %r223;


dp4a.s32.s32 %r231, %r232, %r216, %r227;


dp4a.s32.s32 %r235, %r236, %r220, %r231;


{ mov.b32 %f79, {0,%rs10};}


mul.ftz.f32 %f108, %f79, 0f3F000000;

{.reg .f16 low,high;
mov.b32 {low,high},%r222;
cvt.f32.f16 %f80, low;}


mul.ftz.f32 %f109, %f108, %f80;
cvt.rn.f32.s32 %f110, %r235;
fma.rn.ftz.f32 %f231, %f109, %f110, %f231;
add.s32 %r615, %r615, 32;
add.s32 %r614, %r614, 32;
add.s32 %r613, %r613, 32;
add.s32 %r612, %r612, 32;
add.s64 %rd97, %rd97, 544;
add.s64 %rd96, %rd96, 544;
add.s64 %rd95, %rd95, 544;
add.s32 %r616, %r616, 32;
setp.lt.s32 %p7, %r616, %r40;
@%p7 bra $L__BB110_2;

$L__BB110_3:
setp.eq.s32 %p8, %r1, 0;
@%p8 bra $L__BB110_5;

mov.u32 %r336, _ZZ13mul_mat_vec_qIL9ggml_type39ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r337, %r1, 1280, %r336;
shl.b32 %r338, %r37, 2;
add.s32 %r339, %r337, %r338;
st.shared.f32 [%r339+-1280], %f240;
st.shared.f32 [%r339+-1152], %f239;
st.shared.f32 [%r339+-1024], %f238;
st.shared.f32 [%r339+-896], %f237;
st.shared.f32 [%r339+-768], %f236;
st.shared.f32 [%r339+-640], %f235;
st.shared.f32 [%r339+-512], %f234;
st.shared.f32 [%r339+-384], %f233;
st.shared.f32 [%r339+-256], %f232;
st.shared.f32 [%r339+-128], %f231;

$L__BB110_5:
setp.ne.s32 %p9, %r1, 0;
bar.sync 0;
@%p9 bra $L__BB110_26;

mov.u32 %r341, %ctaid.x;
shl.b32 %r342, %r341, 1;
mov.u32 %r343, 1;
shl.b32 %r345, %r37, 2;
mov.u32 %r346, _ZZ13mul_mat_vec_qIL9ggml_type39ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r347, %r346, %r345;
mov.u32 %r348, 2;
ld.shared.f32 %f111, [%r347];
add.ftz.f32 %f112, %f111, %f240;
mov.b32 %r349, %f112;
mov.u32 %r350, 31;
mov.u32 %r351, 16;
mov.u32 %r352, -1;
shfl.sync.bfly.b32 %r353|%p10, %r349, %r351, %r350, %r352;
mov.b32 %f113, %r353;
add.ftz.f32 %f114, %f112, %f113;
mov.b32 %r354, %f114;
mov.u32 %r355, 8;
shfl.sync.bfly.b32 %r356|%p11, %r354, %r355, %r350, %r352;
mov.b32 %f115, %r356;
add.ftz.f32 %f116, %f114, %f115;
mov.b32 %r357, %f116;
mov.u32 %r358, 4;
shfl.sync.bfly.b32 %r359|%p12, %r357, %r358, %r350, %r352;
mov.b32 %f117, %r359;
add.ftz.f32 %f118, %f116, %f117;
mov.b32 %r360, %f118;
shfl.sync.bfly.b32 %r361|%p13, %r360, %r348, %r350, %r352;
mov.b32 %f119, %r361;
add.ftz.f32 %f120, %f118, %f119;
mov.b32 %r362, %f120;
shfl.sync.bfly.b32 %r363|%p14, %r362, %r343, %r350, %r352;
mov.b32 %f121, %r363;
add.ftz.f32 %f31, %f120, %f121;
setp.ne.s32 %p15, %r37, 0;
setp.ge.u32 %p16, %r342, %r23;
or.pred %p17, %p15, %p16;
@%p17 bra $L__BB110_8;

cvta.to.global.u64 %rd55, %rd18;
mov.u32 %r364, %ctaid.y;
mad.lo.s32 %r367, %r364, %r29, %r342;
mov.u32 %r368, %ctaid.z;
mad.lo.s32 %r369, %r368, %r35, %r367;
mul.wide.u32 %rd56, %r369, 4;
add.s64 %rd57, %rd55, %rd56;
st.global.f32 [%rd57], %f31;

$L__BB110_8:
ld.shared.f32 %f122, [%r347+128];
add.ftz.f32 %f123, %f122, %f239;
mov.b32 %r375, %f123;
shfl.sync.bfly.b32 %r379|%p18, %r375, %r351, %r350, %r352;
mov.b32 %f124, %r379;
add.ftz.f32 %f125, %f123, %f124;
mov.b32 %r380, %f125;
shfl.sync.bfly.b32 %r382|%p19, %r380, %r355, %r350, %r352;
mov.b32 %f126, %r382;
add.ftz.f32 %f127, %f125, %f126;
mov.b32 %r383, %f127;
shfl.sync.bfly.b32 %r385|%p20, %r383, %r358, %r350, %r352;
mov.b32 %f128, %r385;
add.ftz.f32 %f129, %f127, %f128;
mov.b32 %r386, %f129;
shfl.sync.bfly.b32 %r387|%p21, %r386, %r348, %r350, %r352;
mov.b32 %f130, %r387;
add.ftz.f32 %f131, %f129, %f130;
mov.b32 %r388, %f131;
shfl.sync.bfly.b32 %r390|%p22, %r388, %r343, %r350, %r352;
mov.b32 %f132, %r390;
add.ftz.f32 %f32, %f131, %f132;
setp.eq.s32 %p23, %r37, 1;
or.b32 %r393, %r342, 1;
setp.lt.u32 %p24, %r393, %r23;
and.pred %p1, %p23, %p24;
not.pred %p25, %p1;
@%p25 bra $L__BB110_10;

cvta.to.global.u64 %rd58, %rd18;
mov.u32 %r394, %ctaid.y;
mad.lo.s32 %r397, %r394, %r29, %r342;
mov.u32 %r398, %ctaid.z;
mad.lo.s32 %r399, %r398, %r35, %r397;
mul.wide.u32 %rd59, %r399, 4;
add.s64 %rd60, %rd58, %rd59;
st.global.f32 [%rd60+4], %f32;

$L__BB110_10:
mov.u32 %r404, 2;
ld.shared.f32 %f133, [%r347+256];
add.ftz.f32 %f134, %f133, %f238;
mov.b32 %r405, %f134;
mov.u32 %r406, 31;
mov.u32 %r407, 16;
mov.u32 %r408, -1;
shfl.sync.bfly.b32 %r409|%p26, %r405, %r407, %r406, %r408;
mov.b32 %f135, %r409;
add.ftz.f32 %f136, %f134, %f135;
mov.b32 %r410, %f136;
mov.u32 %r411, 8;
shfl.sync.bfly.b32 %r412|%p27, %r410, %r411, %r406, %r408;
mov.b32 %f137, %r412;
add.ftz.f32 %f138, %f136, %f137;
mov.b32 %r413, %f138;
mov.u32 %r414, 4;
shfl.sync.bfly.b32 %r415|%p28, %r413, %r414, %r406, %r408;
mov.b32 %f139, %r415;
add.ftz.f32 %f140, %f138, %f139;
mov.b32 %r416, %f140;
shfl.sync.bfly.b32 %r417|%p29, %r416, %r404, %r406, %r408;
mov.b32 %f141, %r417;
add.ftz.f32 %f142, %f140, %f141;
mov.b32 %r418, %f142;
mov.u32 %r419, 1;
shfl.sync.bfly.b32 %r420|%p30, %r418, %r419, %r406, %r408;
mov.b32 %f143, %r420;
add.ftz.f32 %f33, %f142, %f143;
@%p17 bra $L__BB110_12;

cvt.u64.u32 %rd61, %r23;
mov.u32 %r423, %ctaid.y;
mad.lo.s32 %r426, %r423, %r29, %r342;
mov.u32 %r427, %ctaid.z;
mad.lo.s32 %r428, %r427, %r35, %r426;
cvt.u64.u32 %rd62, %r428;
add.s64 %rd63, %rd61, %rd62;
cvta.to.global.u64 %rd64, %rd18;
shl.b64 %rd65, %rd63, 2;
add.s64 %rd66, %rd64, %rd65;
st.global.f32 [%rd66], %f33;

$L__BB110_12:
ld.shared.f32 %f144, [%r347+384];
add.ftz.f32 %f145, %f144, %f237;
mov.b32 %r434, %f145;
shfl.sync.bfly.b32 %r438|%p34, %r434, %r407, %r406, %r408;
mov.b32 %f146, %r438;
add.ftz.f32 %f147, %f145, %f146;
mov.b32 %r439, %f147;
shfl.sync.bfly.b32 %r441|%p35, %r439, %r411, %r406, %r408;
mov.b32 %f148, %r441;
add.ftz.f32 %f149, %f147, %f148;
mov.b32 %r442, %f149;
shfl.sync.bfly.b32 %r444|%p36, %r442, %r414, %r406, %r408;
mov.b32 %f150, %r444;
add.ftz.f32 %f151, %f149, %f150;
mov.b32 %r445, %f151;
shfl.sync.bfly.b32 %r446|%p37, %r445, %r404, %r406, %r408;
mov.b32 %f152, %r446;
add.ftz.f32 %f153, %f151, %f152;
mov.b32 %r447, %f153;
shfl.sync.bfly.b32 %r449|%p38, %r447, %r419, %r406, %r408;
mov.b32 %f154, %r449;
add.ftz.f32 %f34, %f153, %f154;
@%p25 bra $L__BB110_14;

add.s32 %r450, %r23, 1;
cvt.u64.u32 %rd67, %r450;
mov.u32 %r451, %ctaid.y;
mad.lo.s32 %r454, %r451, %r29, %r342;
mov.u32 %r455, %ctaid.z;
mad.lo.s32 %r456, %r455, %r35, %r454;
cvt.u64.u32 %rd68, %r456;
add.s64 %rd69, %rd67, %rd68;
cvta.to.global.u64 %rd70, %rd18;
shl.b64 %rd71, %rd69, 2;
add.s64 %rd72, %rd70, %rd71;
st.global.f32 [%rd72], %f34;

$L__BB110_14:
shl.b32 %r457, %r23, 1;
mov.u32 %r458, 1;
mov.u32 %r463, 2;
ld.shared.f32 %f155, [%r347+512];
add.ftz.f32 %f156, %f155, %f236;
mov.b32 %r464, %f156;
mov.u32 %r465, 31;
mov.u32 %r466, 16;
mov.u32 %r467, -1;
shfl.sync.bfly.b32 %r468|%p40, %r464, %r466, %r465, %r467;
mov.b32 %f157, %r468;
add.ftz.f32 %f158, %f156, %f157;
mov.b32 %r469, %f158;
mov.u32 %r470, 8;
shfl.sync.bfly.b32 %r471|%p41, %r469, %r470, %r465, %r467;
mov.b32 %f159, %r471;
add.ftz.f32 %f160, %f158, %f159;
mov.b32 %r472, %f160;
mov.u32 %r473, 4;
shfl.sync.bfly.b32 %r474|%p42, %r472, %r473, %r465, %r467;
mov.b32 %f161, %r474;
add.ftz.f32 %f162, %f160, %f161;
mov.b32 %r475, %f162;
shfl.sync.bfly.b32 %r476|%p43, %r475, %r463, %r465, %r467;
mov.b32 %f163, %r476;
add.ftz.f32 %f164, %f162, %f163;
mov.b32 %r477, %f164;
shfl.sync.bfly.b32 %r478|%p44, %r477, %r458, %r465, %r467;
mov.b32 %f165, %r478;
add.ftz.f32 %f35, %f164, %f165;
cvt.u64.u32 %rd73, %r457;
mov.u32 %r479, %ctaid.y;
mad.lo.s32 %r482, %r479, %r29, %r342;
mov.u32 %r483, %ctaid.z;
mad.lo.s32 %r484, %r483, %r35, %r482;
cvt.u64.u32 %rd74, %r484;
add.s64 %rd75, %rd73, %rd74;
cvta.to.global.u64 %rd76, %rd18;
shl.b64 %rd77, %rd75, 2;
add.s64 %rd14, %rd76, %rd77;
@%p17 bra $L__BB110_16;

st.global.f32 [%rd14], %f35;

$L__BB110_16:
ld.shared.f32 %f166, [%r347+640];
add.ftz.f32 %f167, %f166, %f235;
mov.b32 %r490, %f167;
shfl.sync.bfly.b32 %r494|%p48, %r490, %r466, %r465, %r467;
mov.b32 %f168, %r494;
add.ftz.f32 %f169, %f167, %f168;
mov.b32 %r495, %f169;
shfl.sync.bfly.b32 %r497|%p49, %r495, %r470, %r465, %r467;
mov.b32 %f170, %r497;
add.ftz.f32 %f171, %f169, %f170;
mov.b32 %r498, %f171;
shfl.sync.bfly.b32 %r500|%p50, %r498, %r473, %r465, %r467;
mov.b32 %f172, %r500;
add.ftz.f32 %f173, %f171, %f172;
mov.b32 %r501, %f173;
shfl.sync.bfly.b32 %r502|%p51, %r501, %r463, %r465, %r467;
mov.b32 %f174, %r502;
add.ftz.f32 %f175, %f173, %f174;
mov.b32 %r503, %f175;
shfl.sync.bfly.b32 %r505|%p52, %r503, %r458, %r465, %r467;
mov.b32 %f176, %r505;
add.ftz.f32 %f36, %f175, %f176;
@%p25 bra $L__BB110_18;

st.global.f32 [%rd14+4], %f36;

$L__BB110_18:
mul.lo.s32 %r19, %r23, 3;
mov.u32 %r510, 2;
ld.shared.f32 %f177, [%r347+768];
add.ftz.f32 %f178, %f177, %f234;
mov.b32 %r511, %f178;
mov.u32 %r512, 31;
mov.u32 %r513, 16;
mov.u32 %r514, -1;
shfl.sync.bfly.b32 %r515|%p54, %r511, %r513, %r512, %r514;
mov.b32 %f179, %r515;
add.ftz.f32 %f180, %f178, %f179;
mov.b32 %r516, %f180;
mov.u32 %r517, 8;
shfl.sync.bfly.b32 %r518|%p55, %r516, %r517, %r512, %r514;
mov.b32 %f181, %r518;
add.ftz.f32 %f182, %f180, %f181;
mov.b32 %r519, %f182;
mov.u32 %r520, 4;
shfl.sync.bfly.b32 %r521|%p56, %r519, %r520, %r512, %r514;
mov.b32 %f183, %r521;
add.ftz.f32 %f184, %f182, %f183;
mov.b32 %r522, %f184;
shfl.sync.bfly.b32 %r523|%p57, %r522, %r510, %r512, %r514;
mov.b32 %f185, %r523;
add.ftz.f32 %f186, %f184, %f185;
mov.b32 %r524, %f186;
mov.u32 %r525, 1;
shfl.sync.bfly.b32 %r526|%p58, %r524, %r525, %r512, %r514;
mov.b32 %f187, %r526;
add.ftz.f32 %f37, %f186, %f187;
@%p17 bra $L__BB110_20;

cvt.u64.u32 %rd78, %r19;
add.s64 %rd80, %rd78, %rd74;
shl.b64 %rd82, %rd80, 2;
add.s64 %rd83, %rd76, %rd82;
st.global.f32 [%rd83], %f37;

$L__BB110_20:
ld.shared.f32 %f188, [%r347+896];
add.ftz.f32 %f189, %f188, %f233;
mov.b32 %r540, %f189;
shfl.sync.bfly.b32 %r544|%p62, %r540, %r513, %r512, %r514;
mov.b32 %f190, %r544;
add.ftz.f32 %f191, %f189, %f190;
mov.b32 %r545, %f191;
shfl.sync.bfly.b32 %r547|%p63, %r545, %r517, %r512, %r514;
mov.b32 %f192, %r547;
add.ftz.f32 %f193, %f191, %f192;
mov.b32 %r548, %f193;
shfl.sync.bfly.b32 %r550|%p64, %r548, %r520, %r512, %r514;
mov.b32 %f194, %r550;
add.ftz.f32 %f195, %f193, %f194;
mov.b32 %r551, %f195;
shfl.sync.bfly.b32 %r552|%p65, %r551, %r510, %r512, %r514;
mov.b32 %f196, %r552;
add.ftz.f32 %f197, %f195, %f196;
mov.b32 %r553, %f197;
shfl.sync.bfly.b32 %r555|%p66, %r553, %r525, %r512, %r514;
mov.b32 %f198, %r555;
add.ftz.f32 %f38, %f197, %f198;
@%p25 bra $L__BB110_22;

add.s32 %r556, %r19, 1;
cvt.u64.u32 %rd84, %r556;
add.s64 %rd86, %rd84, %rd74;
shl.b64 %rd88, %rd86, 2;
add.s64 %rd89, %rd76, %rd88;
st.global.f32 [%rd89], %f38;

$L__BB110_22:
shl.b32 %r563, %r23, 2;
mov.u32 %r564, 2;
ld.shared.f32 %f199, [%r347+1024];
add.ftz.f32 %f200, %f199, %f232;
mov.b32 %r569, %f200;
mov.u32 %r570, 31;
mov.u32 %r571, 16;
mov.u32 %r572, -1;
shfl.sync.bfly.b32 %r573|%p68, %r569, %r571, %r570, %r572;
mov.b32 %f201, %r573;
add.ftz.f32 %f202, %f200, %f201;
mov.b32 %r574, %f202;
mov.u32 %r575, 8;
shfl.sync.bfly.b32 %r576|%p69, %r574, %r575, %r570, %r572;
mov.b32 %f203, %r576;
add.ftz.f32 %f204, %f202, %f203;
mov.b32 %r577, %f204;
mov.u32 %r578, 4;
shfl.sync.bfly.b32 %r579|%p70, %r577, %r578, %r570, %r572;
mov.b32 %f205, %r579;
add.ftz.f32 %f206, %f204, %f205;
mov.b32 %r580, %f206;
shfl.sync.bfly.b32 %r581|%p71, %r580, %r564, %r570, %r572;
mov.b32 %f207, %r581;
add.ftz.f32 %f208, %f206, %f207;
mov.b32 %r582, %f208;
mov.u32 %r583, 1;
shfl.sync.bfly.b32 %r584|%p72, %r582, %r583, %r570, %r572;
mov.b32 %f209, %r584;
add.ftz.f32 %f39, %f208, %f209;
cvt.u64.u32 %rd90, %r563;
add.s64 %rd92, %rd90, %rd74;
shl.b64 %rd94, %rd92, 2;
add.s64 %rd15, %rd76, %rd94;
@%p17 bra $L__BB110_24;

st.global.f32 [%rd15], %f39;

$L__BB110_24:
ld.shared.f32 %f210, [%r347+1152];
add.ftz.f32 %f211, %f210, %f231;
mov.b32 %r596, %f211;
shfl.sync.bfly.b32 %r600|%p76, %r596, %r571, %r570, %r572;
mov.b32 %f212, %r600;
add.ftz.f32 %f213, %f211, %f212;
mov.b32 %r601, %f213;
shfl.sync.bfly.b32 %r603|%p77, %r601, %r575, %r570, %r572;
mov.b32 %f214, %r603;
add.ftz.f32 %f215, %f213, %f214;
mov.b32 %r604, %f215;
shfl.sync.bfly.b32 %r606|%p78, %r604, %r578, %r570, %r572;
mov.b32 %f216, %r606;
add.ftz.f32 %f217, %f215, %f216;
mov.b32 %r607, %f217;
shfl.sync.bfly.b32 %r608|%p79, %r607, %r564, %r570, %r572;
mov.b32 %f218, %r608;
add.ftz.f32 %f219, %f217, %f218;
mov.b32 %r609, %f219;
shfl.sync.bfly.b32 %r611|%p80, %r609, %r583, %r570, %r572;
mov.b32 %f220, %r611;
add.ftz.f32 %f40, %f219, %f220;
@%p25 bra $L__BB110_26;

st.global.f32 [%rd15+4], %f40;

$L__BB110_26:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type39ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type39ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type39ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type39ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type39ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type39ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type39ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type39ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type39ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<87>;
.reg .b16 %rs<39>;
.reg .f32 %f<289>;
.reg .b32 %r<613>;
.reg .b64 %rd<108>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type39ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type39ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type39ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type39ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type39ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type39ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type39ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type39ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type39ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type39ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type39ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type39ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type39ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type39ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type39ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type39ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type39ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type39ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r43, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r43, %r2;
mov.u32 %r44, %ctaid.x;
shl.b32 %r4, %r44, 1;
shr.u32 %r45, %r3, 31;
add.s32 %r46, %r3, %r45;
shr.s32 %r612, %r46, 1;
shr.u32 %r6, %r42, 5;
setp.ge.s32 %p3, %r612, %r6;
mov.f32 %f277, 0f00000000;
mov.f32 %f278, %f277;
mov.f32 %f279, %f277;
mov.f32 %f280, %f277;
mov.f32 %f281, %f277;
mov.f32 %f282, %f277;
mov.f32 %f283, %f277;
mov.f32 %f284, %f277;
mov.f32 %f285, %f277;
mov.f32 %f286, %f277;
mov.f32 %f287, %f277;
mov.f32 %f288, %f277;
@%p3 bra $L__BB111_3;

mov.u32 %r47, %ctaid.y;
mov.u32 %r48, %ctaid.z;
mul.lo.s32 %r49, %r47, %r34;
cvt.u64.u32 %rd20, %r49;
mul.lo.s32 %r50, %r48, %r40;
cvt.u64.u32 %rd21, %r50;
add.s64 %rd1, %rd21, %rd20;
mul.hi.u32 %r51, %r48, %r36;
add.s32 %r52, %r51, %r48;
shr.u32 %r53, %r52, %r37;
mul.lo.s32 %r54, %r53, %r39;
mul.hi.u32 %r55, %r47, %r30;
add.s32 %r56, %r55, %r47;
shr.u32 %r57, %r56, %r31;
mul.lo.s32 %r58, %r57, %r33;
cvta.to.global.u64 %rd22, %rd17;
and.b32 %r61, %r46, -2;
sub.s32 %r62, %r3, %r61;
shl.b32 %r63, %r62, 1;
cvt.s64.s32 %rd2, %r63;
shl.b32 %r64, %r62, 3;
cvt.s64.s32 %rd23, %r64;
shl.b32 %r65, %r28, 1;
add.s32 %r611, %r612, %r65;
mad.lo.s32 %r610, %r28, 3, %r612;
shl.b32 %r66, %r28, 2;
add.s32 %r609, %r612, %r66;
mad.lo.s32 %r608, %r28, 5, %r612;
add.s32 %r607, %r612, %r28;
add.s32 %r67, %r612, %r54;
add.s32 %r68, %r67, %r58;
add.s32 %r69, %r4, 1;
mad.lo.s32 %r70, %r27, %r69, %r68;
mul.wide.s32 %rd24, %r70, 17;
add.s64 %rd107, %rd22, %rd24;
mad.lo.s32 %r71, %r4, %r27, %r612;
add.s32 %r72, %r71, %r54;
add.s32 %r73, %r72, %r58;
mul.wide.s32 %rd25, %r73, 17;
add.s64 %rd26, %rd22, %rd25;
add.s64 %rd106, %rd26, 4;
add.s64 %rd27, %rd25, %rd23;
add.s64 %rd28, %rd22, %rd27;
add.s64 %rd105, %rd28, 4;
or.b64 %rd6, %rd23, 4;
cvta.to.global.u64 %rd7, %rd18;
mov.f32 %f277, 0f00000000;
shl.b64 %rd33, %rd2, 2;

$L__BB111_2:
cvt.u64.u32 %rd29, %r612;
add.s64 %rd30, %rd1, %rd29;
mul.lo.s64 %rd31, %rd30, 36;
add.s64 %rd32, %rd7, %rd31;
add.s64 %rd34, %rd32, %rd33;
ld.global.nc.u8 %rs13, [%rd105+-3];
cvt.u32.u16 %r278, %rs13;
and.b32 %r279, %r278, 255;
ld.global.nc.u8 %rs14, [%rd105+-2];
cvt.u32.u16 %r280, %rs14;
prmt.b32 %r281, %r280, %r279, 30212;
ld.global.nc.u8 %rs15, [%rd105+-1];
cvt.u32.u16 %r282, %rs15;
prmt.b32 %r283, %r282, %r281, 28756;
ld.global.nc.u8 %rs16, [%rd105];
cvt.u32.u16 %r284, %rs16;
prmt.b32 %r285, %r284, %r283, 1620;
and.b32 %r286, %r285, -2004318072;
shr.u32 %r287, %r286, 1;
or.b32 %r288, %r287, 839922192;
and.b32 %r289, %r281, 30583;
mov.u32 %r290, 201852420;
mov.u32 %r291, 50462976;
prmt.b32 %r292, %r291, %r290, %r289;
mov.u32 %r293, -185009412;
mov.u32 %r294, -33620224;
prmt.b32 %r295, %r294, %r293, %r289;
and.b32 %r296, %r288, 30292;
prmt.b32 %r297, %r292, %r295, %r296;
shr.u32 %r298, %r285, 16;
and.b32 %r299, %r298, 30583;
prmt.b32 %r300, %r291, %r290, %r299;
prmt.b32 %r301, %r294, %r293, %r299;
shr.u32 %r302, %r288, 16;
prmt.b32 %r303, %r300, %r301, %r302;
mov.u32 %r304, 25632;
prmt.b32 %r245, %r297, %r303, %r304;
mov.u32 %r305, 30001;
prmt.b32 %r249, %r297, %r303, %r305;
ld.global.nc.u32 %r76, [%rd34+4];
mov.u32 %r264, 0;

dp4a.s32.s32 %r74, %r245, %r76, %r264;

ld.global.nc.u32 %r80, [%rd34+20];

dp4a.s32.s32 %r78, %r249, %r80, %r74;

ld.global.nc.u8 %rs17, [%rd105+1];
cvt.u32.u16 %r306, %rs17;
and.b32 %r307, %r306, 255;
ld.global.nc.u8 %rs18, [%rd105+2];
cvt.u32.u16 %r308, %rs18;
prmt.b32 %r309, %r308, %r307, 30212;
ld.global.nc.u8 %rs19, [%rd105+3];
cvt.u32.u16 %r310, %rs19;
prmt.b32 %r311, %r310, %r309, 28756;
ld.global.nc.u8 %rs20, [%rd105+4];
cvt.u32.u16 %r312, %rs20;
prmt.b32 %r313, %r312, %r311, 1620;
and.b32 %r314, %r313, -2004318072;
shr.u32 %r315, %r314, 1;
or.b32 %r316, %r315, 839922192;
and.b32 %r317, %r309, 30583;
prmt.b32 %r318, %r291, %r290, %r317;
prmt.b32 %r319, %r294, %r293, %r317;
and.b32 %r320, %r316, 30292;
prmt.b32 %r321, %r318, %r319, %r320;
shr.u32 %r322, %r313, 16;
and.b32 %r323, %r322, 30583;
prmt.b32 %r324, %r291, %r290, %r323;
prmt.b32 %r325, %r294, %r293, %r323;
shr.u32 %r326, %r316, 16;
prmt.b32 %r327, %r324, %r325, %r326;
prmt.b32 %r253, %r321, %r327, %r304;
prmt.b32 %r257, %r321, %r327, %r305;
ld.global.nc.u32 %r84, [%rd34+8];

dp4a.s32.s32 %r82, %r253, %r84, %r78;

ld.global.nc.u32 %r88, [%rd34+24];

dp4a.s32.s32 %r86, %r257, %r88, %r82;

ld.global.nc.u8 %rs21, [%rd106+-4];
shl.b16 %rs23, %rs21, 7;
and.b16 %rs24, %rs23, 32640;
setp.eq.s16 %p4, %rs21, 255;
setp.eq.s16 %p5, %rs21, 0;
selp.b16 %rs25, 64, %rs24, %p5;
selp.b16 %rs11, 32767, %rs25, %p4;

{ mov.b32 %f73, {0,%rs11};}


mul.ftz.f32 %f97, %f73, 0f3F000000;
ld.global.nc.u32 %r90, [%rd32];

{.reg .f16 low,high;
mov.b32 {low,high},%r90;
cvt.f32.f16 %f74, low;}


mul.ftz.f32 %f98, %f97, %f74;
cvt.rn.f32.s32 %f99, %r86;
fma.rn.ftz.f32 %f288, %f98, %f99, %f288;
add.s64 %rd35, %rd107, %rd6;
ld.global.nc.u8 %rs26, [%rd35+-3];
cvt.u32.u16 %r328, %rs26;
and.b32 %r329, %r328, 255;
ld.global.nc.u8 %rs27, [%rd35+-2];
cvt.u32.u16 %r330, %rs27;
prmt.b32 %r331, %r330, %r329, 30212;
ld.global.nc.u8 %rs28, [%rd35+-1];
cvt.u32.u16 %r332, %rs28;
prmt.b32 %r333, %r332, %r331, 28756;
ld.global.nc.u8 %rs29, [%rd35];
cvt.u32.u16 %r334, %rs29;
prmt.b32 %r335, %r334, %r333, 1620;
and.b32 %r336, %r335, -2004318072;
shr.u32 %r337, %r336, 1;
or.b32 %r338, %r337, 839922192;
and.b32 %r339, %r331, 30583;
prmt.b32 %r340, %r291, %r290, %r339;
prmt.b32 %r341, %r294, %r293, %r339;
and.b32 %r342, %r338, 30292;
prmt.b32 %r343, %r340, %r341, %r342;
shr.u32 %r344, %r335, 16;
and.b32 %r345, %r344, 30583;
prmt.b32 %r346, %r291, %r290, %r345;
prmt.b32 %r347, %r294, %r293, %r345;
shr.u32 %r348, %r338, 16;
prmt.b32 %r349, %r346, %r347, %r348;
prmt.b32 %r262, %r343, %r349, %r304;
prmt.b32 %r266, %r343, %r349, %r305;

dp4a.s32.s32 %r91, %r262, %r76, %r264;


dp4a.s32.s32 %r95, %r266, %r80, %r91;

ld.global.nc.u8 %rs30, [%rd35+1];
cvt.u32.u16 %r350, %rs30;
and.b32 %r351, %r350, 255;
ld.global.nc.u8 %rs31, [%rd35+2];
cvt.u32.u16 %r352, %rs31;
prmt.b32 %r353, %r352, %r351, 30212;
ld.global.nc.u8 %rs32, [%rd35+3];
cvt.u32.u16 %r354, %rs32;
prmt.b32 %r355, %r354, %r353, 28756;
ld.global.nc.u8 %rs33, [%rd35+4];
cvt.u32.u16 %r356, %rs33;
prmt.b32 %r357, %r356, %r355, 1620;
and.b32 %r358, %r357, -2004318072;
shr.u32 %r359, %r358, 1;
or.b32 %r360, %r359, 839922192;
and.b32 %r361, %r353, 30583;
prmt.b32 %r362, %r291, %r290, %r361;
prmt.b32 %r363, %r294, %r293, %r361;
and.b32 %r364, %r360, 30292;
prmt.b32 %r365, %r362, %r363, %r364;
shr.u32 %r366, %r357, 16;
and.b32 %r367, %r366, 30583;
prmt.b32 %r368, %r291, %r290, %r367;
prmt.b32 %r369, %r294, %r293, %r367;
shr.u32 %r370, %r360, 16;
prmt.b32 %r371, %r368, %r369, %r370;
prmt.b32 %r270, %r365, %r371, %r304;
prmt.b32 %r274, %r365, %r371, %r305;

dp4a.s32.s32 %r99, %r270, %r84, %r95;


dp4a.s32.s32 %r103, %r274, %r88, %r99;

ld.global.nc.u8 %rs34, [%rd107];
shl.b16 %rs36, %rs34, 7;
and.b16 %rs37, %rs36, 32640;
setp.eq.s16 %p6, %rs34, 255;
setp.eq.s16 %p7, %rs34, 0;
selp.b16 %rs38, 64, %rs37, %p7;
selp.b16 %rs12, 32767, %rs38, %p6;

{ mov.b32 %f75, {0,%rs12};}


mul.ftz.f32 %f100, %f75, 0f3F000000;

{.reg .f16 low,high;
mov.b32 {low,high},%r90;
cvt.f32.f16 %f76, low;}


mul.ftz.f32 %f101, %f100, %f76;
cvt.rn.f32.s32 %f102, %r103;
fma.rn.ftz.f32 %f287, %f101, %f102, %f287;
cvt.u64.u32 %rd36, %r607;
add.s64 %rd37, %rd1, %rd36;
mul.lo.s64 %rd38, %rd37, 36;
add.s64 %rd39, %rd7, %rd38;
add.s64 %rd40, %rd39, %rd33;
ld.global.nc.u32 %r110, [%rd40+4];

dp4a.s32.s32 %r108, %r245, %r110, %r264;

ld.global.nc.u32 %r114, [%rd40+20];

dp4a.s32.s32 %r112, %r249, %r114, %r108;

ld.global.nc.u32 %r118, [%rd40+8];

dp4a.s32.s32 %r116, %r253, %r118, %r112;

ld.global.nc.u32 %r122, [%rd40+24];

dp4a.s32.s32 %r120, %r257, %r122, %r116;


{ mov.b32 %f77, {0,%rs11};}


mul.ftz.f32 %f103, %f77, 0f3F000000;
ld.global.nc.u32 %r124, [%rd39];

{.reg .f16 low,high;
mov.b32 {low,high},%r124;
cvt.f32.f16 %f78, low;}


mul.ftz.f32 %f104, %f103, %f78;
cvt.rn.f32.s32 %f105, %r120;
fma.rn.ftz.f32 %f286, %f104, %f105, %f286;

dp4a.s32.s32 %r125, %r262, %r110, %r264;


dp4a.s32.s32 %r129, %r266, %r114, %r125;


dp4a.s32.s32 %r133, %r270, %r118, %r129;


dp4a.s32.s32 %r137, %r274, %r122, %r133;


{ mov.b32 %f79, {0,%rs12};}


mul.ftz.f32 %f106, %f79, 0f3F000000;

{.reg .f16 low,high;
mov.b32 {low,high},%r124;
cvt.f32.f16 %f80, low;}


mul.ftz.f32 %f107, %f106, %f80;
cvt.rn.f32.s32 %f108, %r137;
fma.rn.ftz.f32 %f285, %f107, %f108, %f285;
cvt.u64.u32 %rd41, %r611;
add.s64 %rd42, %rd1, %rd41;
mul.lo.s64 %rd43, %rd42, 36;
add.s64 %rd44, %rd7, %rd43;
add.s64 %rd45, %rd44, %rd33;
ld.global.nc.u32 %r144, [%rd45+4];

dp4a.s32.s32 %r142, %r245, %r144, %r264;

ld.global.nc.u32 %r148, [%rd45+20];

dp4a.s32.s32 %r146, %r249, %r148, %r142;

ld.global.nc.u32 %r152, [%rd45+8];

dp4a.s32.s32 %r150, %r253, %r152, %r146;

ld.global.nc.u32 %r156, [%rd45+24];

dp4a.s32.s32 %r154, %r257, %r156, %r150;


{ mov.b32 %f81, {0,%rs11};}


mul.ftz.f32 %f109, %f81, 0f3F000000;
ld.global.nc.u32 %r158, [%rd44];

{.reg .f16 low,high;
mov.b32 {low,high},%r158;
cvt.f32.f16 %f82, low;}


mul.ftz.f32 %f110, %f109, %f82;
cvt.rn.f32.s32 %f111, %r154;
fma.rn.ftz.f32 %f284, %f110, %f111, %f284;

dp4a.s32.s32 %r159, %r262, %r144, %r264;


dp4a.s32.s32 %r163, %r266, %r148, %r159;


dp4a.s32.s32 %r167, %r270, %r152, %r163;


dp4a.s32.s32 %r171, %r274, %r156, %r167;


{ mov.b32 %f83, {0,%rs12};}


mul.ftz.f32 %f112, %f83, 0f3F000000;

{.reg .f16 low,high;
mov.b32 {low,high},%r158;
cvt.f32.f16 %f84, low;}


mul.ftz.f32 %f113, %f112, %f84;
cvt.rn.f32.s32 %f114, %r171;
fma.rn.ftz.f32 %f283, %f113, %f114, %f283;
cvt.u64.u32 %rd46, %r610;
add.s64 %rd47, %rd1, %rd46;
mul.lo.s64 %rd48, %rd47, 36;
add.s64 %rd49, %rd7, %rd48;
add.s64 %rd50, %rd49, %rd33;
ld.global.nc.u32 %r178, [%rd50+4];

dp4a.s32.s32 %r176, %r245, %r178, %r264;

ld.global.nc.u32 %r182, [%rd50+20];

dp4a.s32.s32 %r180, %r249, %r182, %r176;

ld.global.nc.u32 %r186, [%rd50+8];

dp4a.s32.s32 %r184, %r253, %r186, %r180;

ld.global.nc.u32 %r190, [%rd50+24];

dp4a.s32.s32 %r188, %r257, %r190, %r184;


{ mov.b32 %f85, {0,%rs11};}


mul.ftz.f32 %f115, %f85, 0f3F000000;
ld.global.nc.u32 %r192, [%rd49];

{.reg .f16 low,high;
mov.b32 {low,high},%r192;
cvt.f32.f16 %f86, low;}


mul.ftz.f32 %f116, %f115, %f86;
cvt.rn.f32.s32 %f117, %r188;
fma.rn.ftz.f32 %f282, %f116, %f117, %f282;

dp4a.s32.s32 %r193, %r262, %r178, %r264;


dp4a.s32.s32 %r197, %r266, %r182, %r193;


dp4a.s32.s32 %r201, %r270, %r186, %r197;


dp4a.s32.s32 %r205, %r274, %r190, %r201;


{ mov.b32 %f87, {0,%rs12};}


mul.ftz.f32 %f118, %f87, 0f3F000000;

{.reg .f16 low,high;
mov.b32 {low,high},%r192;
cvt.f32.f16 %f88, low;}


mul.ftz.f32 %f119, %f118, %f88;
cvt.rn.f32.s32 %f120, %r205;
fma.rn.ftz.f32 %f281, %f119, %f120, %f281;
cvt.u64.u32 %rd51, %r609;
add.s64 %rd52, %rd1, %rd51;
mul.lo.s64 %rd53, %rd52, 36;
add.s64 %rd54, %rd7, %rd53;
add.s64 %rd55, %rd54, %rd33;
ld.global.nc.u32 %r212, [%rd55+4];

dp4a.s32.s32 %r210, %r245, %r212, %r264;

ld.global.nc.u32 %r216, [%rd55+20];

dp4a.s32.s32 %r214, %r249, %r216, %r210;

ld.global.nc.u32 %r220, [%rd55+8];

dp4a.s32.s32 %r218, %r253, %r220, %r214;

ld.global.nc.u32 %r224, [%rd55+24];

dp4a.s32.s32 %r222, %r257, %r224, %r218;


{ mov.b32 %f89, {0,%rs11};}


mul.ftz.f32 %f121, %f89, 0f3F000000;
ld.global.nc.u32 %r226, [%rd54];

{.reg .f16 low,high;
mov.b32 {low,high},%r226;
cvt.f32.f16 %f90, low;}


mul.ftz.f32 %f122, %f121, %f90;
cvt.rn.f32.s32 %f123, %r222;
fma.rn.ftz.f32 %f280, %f122, %f123, %f280;

dp4a.s32.s32 %r227, %r262, %r212, %r264;


dp4a.s32.s32 %r231, %r266, %r216, %r227;


dp4a.s32.s32 %r235, %r270, %r220, %r231;


dp4a.s32.s32 %r239, %r274, %r224, %r235;


{ mov.b32 %f91, {0,%rs12};}


mul.ftz.f32 %f124, %f91, 0f3F000000;

{.reg .f16 low,high;
mov.b32 {low,high},%r226;
cvt.f32.f16 %f92, low;}


mul.ftz.f32 %f125, %f124, %f92;
cvt.rn.f32.s32 %f126, %r239;
fma.rn.ftz.f32 %f279, %f125, %f126, %f279;
cvt.u64.u32 %rd56, %r608;
add.s64 %rd57, %rd1, %rd56;
mul.lo.s64 %rd58, %rd57, 36;
add.s64 %rd59, %rd7, %rd58;
add.s64 %rd60, %rd59, %rd33;
ld.global.nc.u32 %r246, [%rd60+4];

dp4a.s32.s32 %r244, %r245, %r246, %r264;

ld.global.nc.u32 %r250, [%rd60+20];

dp4a.s32.s32 %r248, %r249, %r250, %r244;

ld.global.nc.u32 %r254, [%rd60+8];

dp4a.s32.s32 %r252, %r253, %r254, %r248;

ld.global.nc.u32 %r258, [%rd60+24];

dp4a.s32.s32 %r256, %r257, %r258, %r252;


{ mov.b32 %f93, {0,%rs11};}


mul.ftz.f32 %f127, %f93, 0f3F000000;
ld.global.nc.u32 %r260, [%rd59];

{.reg .f16 low,high;
mov.b32 {low,high},%r260;
cvt.f32.f16 %f94, low;}


mul.ftz.f32 %f128, %f127, %f94;
cvt.rn.f32.s32 %f129, %r256;
fma.rn.ftz.f32 %f278, %f128, %f129, %f278;

dp4a.s32.s32 %r261, %r262, %r246, %r264;


dp4a.s32.s32 %r265, %r266, %r250, %r261;


dp4a.s32.s32 %r269, %r270, %r254, %r265;


dp4a.s32.s32 %r273, %r274, %r258, %r269;


{ mov.b32 %f95, {0,%rs12};}


mul.ftz.f32 %f130, %f95, 0f3F000000;

{.reg .f16 low,high;
mov.b32 {low,high},%r260;
cvt.f32.f16 %f96, low;}


mul.ftz.f32 %f131, %f130, %f96;
cvt.rn.f32.s32 %f132, %r273;
fma.rn.ftz.f32 %f277, %f131, %f132, %f277;
add.s32 %r611, %r611, 32;
add.s32 %r610, %r610, 32;
add.s32 %r609, %r609, 32;
add.s32 %r608, %r608, 32;
add.s32 %r607, %r607, 32;
add.s64 %rd107, %rd107, 544;
add.s64 %rd106, %rd106, 544;
add.s64 %rd105, %rd105, 544;
add.s32 %r612, %r612, 32;
setp.lt.s32 %p8, %r612, %r6;
@%p8 bra $L__BB111_2;

$L__BB111_3:
setp.eq.s32 %p9, %r1, 0;
@%p9 bra $L__BB111_5;

mov.u32 %r372, _ZZ13mul_mat_vec_qIL9ggml_type39ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r373, %r1, 1536, %r372;
shl.b32 %r374, %r2, 2;
add.s32 %r375, %r373, %r374;
st.shared.f32 [%r375+-1536], %f288;
st.shared.f32 [%r375+-1408], %f287;
st.shared.f32 [%r375+-1280], %f286;
st.shared.f32 [%r375+-1152], %f285;
st.shared.f32 [%r375+-1024], %f284;
st.shared.f32 [%r375+-896], %f283;
st.shared.f32 [%r375+-768], %f282;
st.shared.f32 [%r375+-640], %f281;
st.shared.f32 [%r375+-512], %f280;
st.shared.f32 [%r375+-384], %f279;
st.shared.f32 [%r375+-256], %f278;
st.shared.f32 [%r375+-128], %f277;

$L__BB111_5:
setp.ne.s32 %p10, %r1, 0;
bar.sync 0;
@%p10 bra $L__BB111_30;

mov.u32 %r376, %ctaid.z;
mov.u32 %r377, %ctaid.y;
mad.lo.s32 %r378, %r377, %r35, %r4;
mad.lo.s32 %r379, %r376, %r41, %r378;
cvt.u64.u32 %rd14, %r379;
shl.b32 %r380, %r2, 2;
mov.u32 %r381, _ZZ13mul_mat_vec_qIL9ggml_type39ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r24, %r381, %r380;
mov.u32 %r382, 2;
ld.shared.f32 %f133, [%r24];
add.ftz.f32 %f134, %f133, %f288;
mov.b32 %r383, %f134;
mov.u32 %r384, 31;
mov.u32 %r385, 16;
mov.u32 %r386, -1;
shfl.sync.bfly.b32 %r387|%p11, %r383, %r385, %r384, %r386;
mov.b32 %f135, %r387;
add.ftz.f32 %f136, %f134, %f135;
mov.b32 %r388, %f136;
mov.u32 %r389, 8;
shfl.sync.bfly.b32 %r390|%p12, %r388, %r389, %r384, %r386;
mov.b32 %f137, %r390;
add.ftz.f32 %f138, %f136, %f137;
mov.b32 %r391, %f138;
mov.u32 %r392, 4;
shfl.sync.bfly.b32 %r393|%p13, %r391, %r392, %r384, %r386;
mov.b32 %f139, %r393;
add.ftz.f32 %f140, %f138, %f139;
mov.b32 %r394, %f140;
shfl.sync.bfly.b32 %r395|%p14, %r394, %r382, %r384, %r386;
mov.b32 %f141, %r395;
add.ftz.f32 %f142, %f140, %f141;
mov.b32 %r396, %f142;
mov.u32 %r397, 1;
shfl.sync.bfly.b32 %r398|%p15, %r396, %r397, %r384, %r386;
mov.b32 %f143, %r398;
add.ftz.f32 %f37, %f142, %f143;
setp.lt.u32 %p16, %r4, %r29;
setp.eq.s32 %p17, %r2, 0;
and.pred %p1, %p17, %p16;
not.pred %p18, %p1;
@%p18 bra $L__BB111_8;

cvta.to.global.u64 %rd61, %rd19;
mul.wide.u32 %rd62, %r379, 4;
add.s64 %rd63, %rd61, %rd62;
st.global.f32 [%rd63], %f37;

$L__BB111_8:
ld.shared.f32 %f144, [%r24+128];
add.ftz.f32 %f145, %f144, %f287;
mov.b32 %r405, %f145;
shfl.sync.bfly.b32 %r409|%p19, %r405, %r385, %r384, %r386;
mov.b32 %f146, %r409;
add.ftz.f32 %f147, %f145, %f146;
mov.b32 %r410, %f147;
shfl.sync.bfly.b32 %r412|%p20, %r410, %r389, %r384, %r386;
mov.b32 %f148, %r412;
add.ftz.f32 %f149, %f147, %f148;
mov.b32 %r413, %f149;
shfl.sync.bfly.b32 %r415|%p21, %r413, %r392, %r384, %r386;
mov.b32 %f150, %r415;
add.ftz.f32 %f151, %f149, %f150;
mov.b32 %r416, %f151;
shfl.sync.bfly.b32 %r418|%p22, %r416, %r382, %r384, %r386;
mov.b32 %f152, %r418;
add.ftz.f32 %f153, %f151, %f152;
mov.b32 %r419, %f153;
shfl.sync.bfly.b32 %r421|%p23, %r419, %r397, %r384, %r386;
mov.b32 %f154, %r421;
add.ftz.f32 %f38, %f153, %f154;
setp.eq.s32 %p24, %r2, 1;
or.b32 %r425, %r4, 1;
setp.lt.u32 %p25, %r425, %r29;
and.pred %p2, %p24, %p25;
not.pred %p26, %p2;
@%p26 bra $L__BB111_10;

cvta.to.global.u64 %rd64, %rd19;
mul.wide.u32 %rd65, %r379, 4;
add.s64 %rd66, %rd64, %rd65;
st.global.f32 [%rd66+4], %f38;

$L__BB111_10:
ld.shared.f32 %f155, [%r24+256];
add.ftz.f32 %f156, %f155, %f286;
mov.b32 %r432, %f156;
mov.u32 %r433, 31;
mov.u32 %r434, 16;
mov.u32 %r435, -1;
shfl.sync.bfly.b32 %r436|%p27, %r432, %r434, %r433, %r435;
mov.b32 %f157, %r436;
add.ftz.f32 %f158, %f156, %f157;
mov.b32 %r437, %f158;
mov.u32 %r438, 8;
shfl.sync.bfly.b32 %r439|%p28, %r437, %r438, %r433, %r435;
mov.b32 %f159, %r439;
add.ftz.f32 %f160, %f158, %f159;
mov.b32 %r440, %f160;
mov.u32 %r441, 4;
shfl.sync.bfly.b32 %r442|%p29, %r440, %r441, %r433, %r435;
mov.b32 %f161, %r442;
add.ftz.f32 %f162, %f160, %f161;
mov.b32 %r443, %f162;
mov.u32 %r444, 2;
shfl.sync.bfly.b32 %r445|%p30, %r443, %r444, %r433, %r435;
mov.b32 %f163, %r445;
add.ftz.f32 %f164, %f162, %f163;
mov.b32 %r446, %f164;
mov.u32 %r447, 1;
shfl.sync.bfly.b32 %r448|%p31, %r446, %r447, %r433, %r435;
mov.b32 %f165, %r448;
add.ftz.f32 %f39, %f164, %f165;
@%p18 bra $L__BB111_12;

cvt.u64.u32 %rd67, %r29;
add.s64 %rd68, %rd67, %rd14;
cvta.to.global.u64 %rd69, %rd19;
shl.b64 %rd70, %rd68, 2;
add.s64 %rd71, %rd69, %rd70;
st.global.f32 [%rd71], %f39;

$L__BB111_12:
ld.shared.f32 %f166, [%r24+384];
add.ftz.f32 %f167, %f166, %f285;
mov.b32 %r449, %f167;
shfl.sync.bfly.b32 %r453|%p33, %r449, %r434, %r433, %r435;
mov.b32 %f168, %r453;
add.ftz.f32 %f169, %f167, %f168;
mov.b32 %r454, %f169;
shfl.sync.bfly.b32 %r456|%p34, %r454, %r438, %r433, %r435;
mov.b32 %f170, %r456;
add.ftz.f32 %f171, %f169, %f170;
mov.b32 %r457, %f171;
shfl.sync.bfly.b32 %r459|%p35, %r457, %r441, %r433, %r435;
mov.b32 %f172, %r459;
add.ftz.f32 %f173, %f171, %f172;
mov.b32 %r460, %f173;
shfl.sync.bfly.b32 %r462|%p36, %r460, %r444, %r433, %r435;
mov.b32 %f174, %r462;
add.ftz.f32 %f175, %f173, %f174;
mov.b32 %r463, %f175;
shfl.sync.bfly.b32 %r465|%p37, %r463, %r447, %r433, %r435;
mov.b32 %f176, %r465;
add.ftz.f32 %f40, %f175, %f176;
@%p26 bra $L__BB111_14;

add.s32 %r466, %r29, 1;
cvt.u64.u32 %rd72, %r466;
add.s64 %rd73, %rd72, %rd14;
cvta.to.global.u64 %rd74, %rd19;
shl.b64 %rd75, %rd73, 2;
add.s64 %rd76, %rd74, %rd75;
st.global.f32 [%rd76], %f40;

$L__BB111_14:
shl.b32 %r467, %r29, 1;
mov.u32 %r468, 1;
ld.shared.f32 %f177, [%r24+512];
add.ftz.f32 %f178, %f177, %f284;
mov.b32 %r469, %f178;
mov.u32 %r470, 31;
mov.u32 %r471, 16;
mov.u32 %r472, -1;
shfl.sync.bfly.b32 %r473|%p39, %r469, %r471, %r470, %r472;
mov.b32 %f179, %r473;
add.ftz.f32 %f180, %f178, %f179;
mov.b32 %r474, %f180;
mov.u32 %r475, 8;
shfl.sync.bfly.b32 %r476|%p40, %r474, %r475, %r470, %r472;
mov.b32 %f181, %r476;
add.ftz.f32 %f182, %f180, %f181;
mov.b32 %r477, %f182;
mov.u32 %r478, 4;
shfl.sync.bfly.b32 %r479|%p41, %r477, %r478, %r470, %r472;
mov.b32 %f183, %r479;
add.ftz.f32 %f184, %f182, %f183;
mov.b32 %r480, %f184;
mov.u32 %r481, 2;
shfl.sync.bfly.b32 %r482|%p42, %r480, %r481, %r470, %r472;
mov.b32 %f185, %r482;
add.ftz.f32 %f186, %f184, %f185;
mov.b32 %r483, %f186;
shfl.sync.bfly.b32 %r484|%p43, %r483, %r468, %r470, %r472;
mov.b32 %f187, %r484;
add.ftz.f32 %f41, %f186, %f187;
cvt.u64.u32 %rd77, %r467;
add.s64 %rd78, %rd77, %rd14;
cvta.to.global.u64 %rd79, %rd19;
shl.b64 %rd80, %rd78, 2;
add.s64 %rd15, %rd79, %rd80;
@%p18 bra $L__BB111_16;

st.global.f32 [%rd15], %f41;

$L__BB111_16:
ld.shared.f32 %f188, [%r24+640];
add.ftz.f32 %f189, %f188, %f283;
mov.b32 %r485, %f189;
shfl.sync.bfly.b32 %r489|%p45, %r485, %r471, %r470, %r472;
mov.b32 %f190, %r489;
add.ftz.f32 %f191, %f189, %f190;
mov.b32 %r490, %f191;
shfl.sync.bfly.b32 %r492|%p46, %r490, %r475, %r470, %r472;
mov.b32 %f192, %r492;
add.ftz.f32 %f193, %f191, %f192;
mov.b32 %r493, %f193;
shfl.sync.bfly.b32 %r495|%p47, %r493, %r478, %r470, %r472;
mov.b32 %f194, %r495;
add.ftz.f32 %f195, %f193, %f194;
mov.b32 %r496, %f195;
shfl.sync.bfly.b32 %r498|%p48, %r496, %r481, %r470, %r472;
mov.b32 %f196, %r498;
add.ftz.f32 %f197, %f195, %f196;
mov.b32 %r499, %f197;
shfl.sync.bfly.b32 %r501|%p49, %r499, %r468, %r470, %r472;
mov.b32 %f198, %r501;
add.ftz.f32 %f42, %f197, %f198;
@%p26 bra $L__BB111_18;

st.global.f32 [%rd15+4], %f42;

$L__BB111_18:
mul.lo.s32 %r25, %r29, 3;
ld.shared.f32 %f199, [%r24+768];
add.ftz.f32 %f200, %f199, %f282;
mov.b32 %r502, %f200;
mov.u32 %r503, 31;
mov.u32 %r504, 16;
mov.u32 %r505, -1;
shfl.sync.bfly.b32 %r506|%p51, %r502, %r504, %r503, %r505;
mov.b32 %f201, %r506;
add.ftz.f32 %f202, %f200, %f201;
mov.b32 %r507, %f202;
mov.u32 %r508, 8;
shfl.sync.bfly.b32 %r509|%p52, %r507, %r508, %r503, %r505;
mov.b32 %f203, %r509;
add.ftz.f32 %f204, %f202, %f203;
mov.b32 %r510, %f204;
mov.u32 %r511, 4;
shfl.sync.bfly.b32 %r512|%p53, %r510, %r511, %r503, %r505;
mov.b32 %f205, %r512;
add.ftz.f32 %f206, %f204, %f205;
mov.b32 %r513, %f206;
mov.u32 %r514, 2;
shfl.sync.bfly.b32 %r515|%p54, %r513, %r514, %r503, %r505;
mov.b32 %f207, %r515;
add.ftz.f32 %f208, %f206, %f207;
mov.b32 %r516, %f208;
mov.u32 %r517, 1;
shfl.sync.bfly.b32 %r518|%p55, %r516, %r517, %r503, %r505;
mov.b32 %f209, %r518;
add.ftz.f32 %f43, %f208, %f209;
@%p18 bra $L__BB111_20;

cvt.u64.u32 %rd81, %r25;
add.s64 %rd82, %rd81, %rd14;
shl.b64 %rd84, %rd82, 2;
add.s64 %rd85, %rd79, %rd84;
st.global.f32 [%rd85], %f43;

$L__BB111_20:
ld.shared.f32 %f210, [%r24+896];
add.ftz.f32 %f211, %f210, %f281;
mov.b32 %r519, %f211;
shfl.sync.bfly.b32 %r523|%p57, %r519, %r504, %r503, %r505;
mov.b32 %f212, %r523;
add.ftz.f32 %f213, %f211, %f212;
mov.b32 %r524, %f213;
shfl.sync.bfly.b32 %r526|%p58, %r524, %r508, %r503, %r505;
mov.b32 %f214, %r526;
add.ftz.f32 %f215, %f213, %f214;
mov.b32 %r527, %f215;
shfl.sync.bfly.b32 %r529|%p59, %r527, %r511, %r503, %r505;
mov.b32 %f216, %r529;
add.ftz.f32 %f217, %f215, %f216;
mov.b32 %r530, %f217;
shfl.sync.bfly.b32 %r532|%p60, %r530, %r514, %r503, %r505;
mov.b32 %f218, %r532;
add.ftz.f32 %f219, %f217, %f218;
mov.b32 %r533, %f219;
shfl.sync.bfly.b32 %r535|%p61, %r533, %r517, %r503, %r505;
mov.b32 %f220, %r535;
add.ftz.f32 %f44, %f219, %f220;
@%p26 bra $L__BB111_22;

add.s32 %r536, %r25, 1;
cvt.u64.u32 %rd86, %r536;
add.s64 %rd87, %rd86, %rd14;
shl.b64 %rd89, %rd87, 2;
add.s64 %rd90, %rd79, %rd89;
st.global.f32 [%rd90], %f44;

$L__BB111_22:
shl.b32 %r537, %r29, 2;
mov.u32 %r538, 2;
ld.shared.f32 %f221, [%r24+1024];
add.ftz.f32 %f222, %f221, %f280;
mov.b32 %r539, %f222;
mov.u32 %r540, 31;
mov.u32 %r541, 16;
mov.u32 %r542, -1;
shfl.sync.bfly.b32 %r543|%p63, %r539, %r541, %r540, %r542;
mov.b32 %f223, %r543;
add.ftz.f32 %f224, %f222, %f223;
mov.b32 %r544, %f224;
mov.u32 %r545, 8;
shfl.sync.bfly.b32 %r546|%p64, %r544, %r545, %r540, %r542;
mov.b32 %f225, %r546;
add.ftz.f32 %f226, %f224, %f225;
mov.b32 %r547, %f226;
mov.u32 %r548, 4;
shfl.sync.bfly.b32 %r549|%p65, %r547, %r548, %r540, %r542;
mov.b32 %f227, %r549;
add.ftz.f32 %f228, %f226, %f227;
mov.b32 %r550, %f228;
shfl.sync.bfly.b32 %r551|%p66, %r550, %r538, %r540, %r542;
mov.b32 %f229, %r551;
add.ftz.f32 %f230, %f228, %f229;
mov.b32 %r552, %f230;
mov.u32 %r553, 1;
shfl.sync.bfly.b32 %r554|%p67, %r552, %r553, %r540, %r542;
mov.b32 %f231, %r554;
add.ftz.f32 %f45, %f230, %f231;
cvt.u64.u32 %rd91, %r537;
add.s64 %rd92, %rd91, %rd14;
shl.b64 %rd94, %rd92, 2;
add.s64 %rd16, %rd79, %rd94;
@%p18 bra $L__BB111_24;

st.global.f32 [%rd16], %f45;

$L__BB111_24:
ld.shared.f32 %f232, [%r24+1152];
add.ftz.f32 %f233, %f232, %f279;
mov.b32 %r555, %f233;
shfl.sync.bfly.b32 %r559|%p69, %r555, %r541, %r540, %r542;
mov.b32 %f234, %r559;
add.ftz.f32 %f235, %f233, %f234;
mov.b32 %r560, %f235;
shfl.sync.bfly.b32 %r562|%p70, %r560, %r545, %r540, %r542;
mov.b32 %f236, %r562;
add.ftz.f32 %f237, %f235, %f236;
mov.b32 %r563, %f237;
shfl.sync.bfly.b32 %r565|%p71, %r563, %r548, %r540, %r542;
mov.b32 %f238, %r565;
add.ftz.f32 %f239, %f237, %f238;
mov.b32 %r566, %f239;
shfl.sync.bfly.b32 %r568|%p72, %r566, %r538, %r540, %r542;
mov.b32 %f240, %r568;
add.ftz.f32 %f241, %f239, %f240;
mov.b32 %r569, %f241;
shfl.sync.bfly.b32 %r571|%p73, %r569, %r553, %r540, %r542;
mov.b32 %f242, %r571;
add.ftz.f32 %f46, %f241, %f242;
@%p26 bra $L__BB111_26;

st.global.f32 [%rd16+4], %f46;

$L__BB111_26:
mul.lo.s32 %r26, %r29, 5;
ld.shared.f32 %f243, [%r24+1280];
add.ftz.f32 %f244, %f243, %f278;
mov.b32 %r572, %f244;
mov.u32 %r573, 31;
mov.u32 %r574, 16;
mov.u32 %r575, -1;
shfl.sync.bfly.b32 %r576|%p75, %r572, %r574, %r573, %r575;
mov.b32 %f245, %r576;
add.ftz.f32 %f246, %f244, %f245;
mov.b32 %r577, %f246;
mov.u32 %r578, 8;
shfl.sync.bfly.b32 %r579|%p76, %r577, %r578, %r573, %r575;
mov.b32 %f247, %r579;
add.ftz.f32 %f248, %f246, %f247;
mov.b32 %r580, %f248;
mov.u32 %r581, 4;
shfl.sync.bfly.b32 %r582|%p77, %r580, %r581, %r573, %r575;
mov.b32 %f249, %r582;
add.ftz.f32 %f250, %f248, %f249;
mov.b32 %r583, %f250;
mov.u32 %r584, 2;
shfl.sync.bfly.b32 %r585|%p78, %r583, %r584, %r573, %r575;
mov.b32 %f251, %r585;
add.ftz.f32 %f252, %f250, %f251;
mov.b32 %r586, %f252;
mov.u32 %r587, 1;
shfl.sync.bfly.b32 %r588|%p79, %r586, %r587, %r573, %r575;
mov.b32 %f253, %r588;
add.ftz.f32 %f47, %f252, %f253;
@%p18 bra $L__BB111_28;

cvt.u64.u32 %rd95, %r26;
add.s64 %rd96, %rd95, %rd14;
shl.b64 %rd98, %rd96, 2;
add.s64 %rd99, %rd79, %rd98;
st.global.f32 [%rd99], %f47;

$L__BB111_28:
ld.shared.f32 %f254, [%r24+1408];
add.ftz.f32 %f255, %f254, %f277;
mov.b32 %r589, %f255;
shfl.sync.bfly.b32 %r593|%p81, %r589, %r574, %r573, %r575;
mov.b32 %f256, %r593;
add.ftz.f32 %f257, %f255, %f256;
mov.b32 %r594, %f257;
shfl.sync.bfly.b32 %r596|%p82, %r594, %r578, %r573, %r575;
mov.b32 %f258, %r596;
add.ftz.f32 %f259, %f257, %f258;
mov.b32 %r597, %f259;
shfl.sync.bfly.b32 %r599|%p83, %r597, %r581, %r573, %r575;
mov.b32 %f260, %r599;
add.ftz.f32 %f261, %f259, %f260;
mov.b32 %r600, %f261;
shfl.sync.bfly.b32 %r602|%p84, %r600, %r584, %r573, %r575;
mov.b32 %f262, %r602;
add.ftz.f32 %f263, %f261, %f262;
mov.b32 %r603, %f263;
shfl.sync.bfly.b32 %r605|%p85, %r603, %r587, %r573, %r575;
mov.b32 %f264, %r605;
add.ftz.f32 %f48, %f263, %f264;
@%p26 bra $L__BB111_30;

add.s32 %r606, %r26, 1;
cvt.u64.u32 %rd100, %r606;
add.s64 %rd101, %rd100, %rd14;
shl.b64 %rd103, %rd101, 2;
add.s64 %rd104, %rd79, %rd103;
st.global.f32 [%rd104], %f48;

$L__BB111_30:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type39ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type39ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type39ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type39ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type39ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type39ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type39ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type39ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type39ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<99>;
.reg .b16 %rs<41>;
.reg .f32 %f<337>;
.reg .b32 %r<674>;
.reg .b64 %rd<115>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type39ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1792];

ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type39ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd20, [_Z13mul_mat_vec_qIL9ggml_type39ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd21, [_Z13mul_mat_vec_qIL9ggml_type39ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r45, [_Z13mul_mat_vec_qIL9ggml_type39ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type39ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type39ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type39ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type39ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type39ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type39ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type39ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type39ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type39ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type39ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type39ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type39ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type39ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r46, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r46, %r2;
mov.u32 %r47, %ctaid.x;
shl.b32 %r4, %r47, 1;
shr.u32 %r48, %r3, 31;
add.s32 %r49, %r3, %r48;
shr.s32 %r673, %r49, 1;
shr.u32 %r6, %r45, 5;
setp.ge.s32 %p3, %r673, %r6;
mov.f32 %f323, 0f00000000;
mov.f32 %f324, %f323;
mov.f32 %f325, %f323;
mov.f32 %f326, %f323;
mov.f32 %f327, %f323;
mov.f32 %f328, %f323;
mov.f32 %f329, %f323;
mov.f32 %f330, %f323;
mov.f32 %f331, %f323;
mov.f32 %f332, %f323;
mov.f32 %f333, %f323;
mov.f32 %f334, %f323;
mov.f32 %f335, %f323;
mov.f32 %f336, %f323;
@%p3 bra $L__BB112_3;

mov.u32 %r50, %ctaid.y;
mov.u32 %r51, %ctaid.z;
mul.lo.s32 %r52, %r50, %r37;
cvt.u64.u32 %rd22, %r52;
mul.lo.s32 %r53, %r51, %r43;
cvt.u64.u32 %rd23, %r53;
add.s64 %rd1, %rd23, %rd22;
mul.hi.u32 %r54, %r51, %r39;
add.s32 %r55, %r54, %r51;
shr.u32 %r56, %r55, %r40;
mul.lo.s32 %r57, %r56, %r42;
mul.hi.u32 %r58, %r50, %r33;
add.s32 %r59, %r58, %r50;
shr.u32 %r60, %r59, %r34;
mul.lo.s32 %r61, %r60, %r36;
cvta.to.global.u64 %rd24, %rd19;
and.b32 %r64, %r49, -2;
sub.s32 %r65, %r3, %r64;
shl.b32 %r66, %r65, 1;
cvt.s64.s32 %rd2, %r66;
shl.b32 %r67, %r65, 3;
cvt.s64.s32 %rd25, %r67;
shl.b32 %r68, %r31, 1;
add.s32 %r672, %r673, %r68;
mad.lo.s32 %r671, %r31, 3, %r673;
shl.b32 %r69, %r31, 2;
add.s32 %r670, %r673, %r69;
mad.lo.s32 %r669, %r31, 5, %r673;
mad.lo.s32 %r668, %r31, 6, %r673;
add.s32 %r667, %r673, %r31;
add.s32 %r70, %r673, %r57;
add.s32 %r71, %r70, %r61;
add.s32 %r72, %r4, 1;
mad.lo.s32 %r73, %r30, %r72, %r71;
mul.wide.s32 %rd26, %r73, 17;
add.s64 %rd114, %rd24, %rd26;
mad.lo.s32 %r74, %r4, %r30, %r673;
add.s32 %r75, %r74, %r57;
add.s32 %r76, %r75, %r61;
mul.wide.s32 %rd27, %r76, 17;
add.s64 %rd28, %rd24, %rd27;
add.s64 %rd113, %rd28, 4;
add.s64 %rd29, %rd27, %rd25;
add.s64 %rd30, %rd24, %rd29;
add.s64 %rd112, %rd30, 4;
or.b64 %rd6, %rd25, 4;
cvta.to.global.u64 %rd7, %rd20;
mov.f32 %f323, 0f00000000;
shl.b64 %rd35, %rd2, 2;

$L__BB112_2:
cvt.u64.u32 %rd31, %r673;
add.s64 %rd32, %rd1, %rd31;
mul.lo.s64 %rd33, %rd32, 36;
add.s64 %rd34, %rd7, %rd33;
add.s64 %rd36, %rd34, %rd35;
ld.global.nc.u8 %rs15, [%rd112+-3];
cvt.u32.u16 %r315, %rs15;
and.b32 %r316, %r315, 255;
ld.global.nc.u8 %rs16, [%rd112+-2];
cvt.u32.u16 %r317, %rs16;
prmt.b32 %r318, %r317, %r316, 30212;
ld.global.nc.u8 %rs17, [%rd112+-1];
cvt.u32.u16 %r319, %rs17;
prmt.b32 %r320, %r319, %r318, 28756;
ld.global.nc.u8 %rs18, [%rd112];
cvt.u32.u16 %r321, %rs18;
prmt.b32 %r322, %r321, %r320, 1620;
and.b32 %r323, %r322, -2004318072;
shr.u32 %r324, %r323, 1;
or.b32 %r325, %r324, 839922192;
and.b32 %r326, %r318, 30583;
mov.u32 %r327, 201852420;
mov.u32 %r328, 50462976;
prmt.b32 %r329, %r328, %r327, %r326;
mov.u32 %r330, -185009412;
mov.u32 %r331, -33620224;
prmt.b32 %r332, %r331, %r330, %r326;
and.b32 %r333, %r325, 30292;
prmt.b32 %r334, %r329, %r332, %r333;
shr.u32 %r335, %r322, 16;
and.b32 %r336, %r335, 30583;
prmt.b32 %r337, %r328, %r327, %r336;
prmt.b32 %r338, %r331, %r330, %r336;
shr.u32 %r339, %r325, 16;
prmt.b32 %r340, %r337, %r338, %r339;
mov.u32 %r341, 25632;
prmt.b32 %r282, %r334, %r340, %r341;
mov.u32 %r342, 30001;
prmt.b32 %r286, %r334, %r340, %r342;
ld.global.nc.u32 %r79, [%rd36+4];
mov.u32 %r301, 0;

dp4a.s32.s32 %r77, %r282, %r79, %r301;

ld.global.nc.u32 %r83, [%rd36+20];

dp4a.s32.s32 %r81, %r286, %r83, %r77;

ld.global.nc.u8 %rs19, [%rd112+1];
cvt.u32.u16 %r343, %rs19;
and.b32 %r344, %r343, 255;
ld.global.nc.u8 %rs20, [%rd112+2];
cvt.u32.u16 %r345, %rs20;
prmt.b32 %r346, %r345, %r344, 30212;
ld.global.nc.u8 %rs21, [%rd112+3];
cvt.u32.u16 %r347, %rs21;
prmt.b32 %r348, %r347, %r346, 28756;
ld.global.nc.u8 %rs22, [%rd112+4];
cvt.u32.u16 %r349, %rs22;
prmt.b32 %r350, %r349, %r348, 1620;
and.b32 %r351, %r350, -2004318072;
shr.u32 %r352, %r351, 1;
or.b32 %r353, %r352, 839922192;
and.b32 %r354, %r346, 30583;
prmt.b32 %r355, %r328, %r327, %r354;
prmt.b32 %r356, %r331, %r330, %r354;
and.b32 %r357, %r353, 30292;
prmt.b32 %r358, %r355, %r356, %r357;
shr.u32 %r359, %r350, 16;
and.b32 %r360, %r359, 30583;
prmt.b32 %r361, %r328, %r327, %r360;
prmt.b32 %r362, %r331, %r330, %r360;
shr.u32 %r363, %r353, 16;
prmt.b32 %r364, %r361, %r362, %r363;
prmt.b32 %r290, %r358, %r364, %r341;
prmt.b32 %r294, %r358, %r364, %r342;
ld.global.nc.u32 %r87, [%rd36+8];

dp4a.s32.s32 %r85, %r290, %r87, %r81;

ld.global.nc.u32 %r91, [%rd36+24];

dp4a.s32.s32 %r89, %r294, %r91, %r85;

ld.global.nc.u8 %rs23, [%rd113+-4];
shl.b16 %rs25, %rs23, 7;
and.b16 %rs26, %rs25, 32640;
setp.eq.s16 %p4, %rs23, 255;
setp.eq.s16 %p5, %rs23, 0;
selp.b16 %rs27, 64, %rs26, %p5;
selp.b16 %rs13, 32767, %rs27, %p4;

{ mov.b32 %f85, {0,%rs13};}


mul.ftz.f32 %f113, %f85, 0f3F000000;
ld.global.nc.u32 %r93, [%rd34];

{.reg .f16 low,high;
mov.b32 {low,high},%r93;
cvt.f32.f16 %f86, low;}


mul.ftz.f32 %f114, %f113, %f86;
cvt.rn.f32.s32 %f115, %r89;
fma.rn.ftz.f32 %f336, %f114, %f115, %f336;
add.s64 %rd37, %rd114, %rd6;
ld.global.nc.u8 %rs28, [%rd37+-3];
cvt.u32.u16 %r365, %rs28;
and.b32 %r366, %r365, 255;
ld.global.nc.u8 %rs29, [%rd37+-2];
cvt.u32.u16 %r367, %rs29;
prmt.b32 %r368, %r367, %r366, 30212;
ld.global.nc.u8 %rs30, [%rd37+-1];
cvt.u32.u16 %r369, %rs30;
prmt.b32 %r370, %r369, %r368, 28756;
ld.global.nc.u8 %rs31, [%rd37];
cvt.u32.u16 %r371, %rs31;
prmt.b32 %r372, %r371, %r370, 1620;
and.b32 %r373, %r372, -2004318072;
shr.u32 %r374, %r373, 1;
or.b32 %r375, %r374, 839922192;
and.b32 %r376, %r368, 30583;
prmt.b32 %r377, %r328, %r327, %r376;
prmt.b32 %r378, %r331, %r330, %r376;
and.b32 %r379, %r375, 30292;
prmt.b32 %r380, %r377, %r378, %r379;
shr.u32 %r381, %r372, 16;
and.b32 %r382, %r381, 30583;
prmt.b32 %r383, %r328, %r327, %r382;
prmt.b32 %r384, %r331, %r330, %r382;
shr.u32 %r385, %r375, 16;
prmt.b32 %r386, %r383, %r384, %r385;
prmt.b32 %r299, %r380, %r386, %r341;
prmt.b32 %r303, %r380, %r386, %r342;

dp4a.s32.s32 %r94, %r299, %r79, %r301;


dp4a.s32.s32 %r98, %r303, %r83, %r94;

ld.global.nc.u8 %rs32, [%rd37+1];
cvt.u32.u16 %r387, %rs32;
and.b32 %r388, %r387, 255;
ld.global.nc.u8 %rs33, [%rd37+2];
cvt.u32.u16 %r389, %rs33;
prmt.b32 %r390, %r389, %r388, 30212;
ld.global.nc.u8 %rs34, [%rd37+3];
cvt.u32.u16 %r391, %rs34;
prmt.b32 %r392, %r391, %r390, 28756;
ld.global.nc.u8 %rs35, [%rd37+4];
cvt.u32.u16 %r393, %rs35;
prmt.b32 %r394, %r393, %r392, 1620;
and.b32 %r395, %r394, -2004318072;
shr.u32 %r396, %r395, 1;
or.b32 %r397, %r396, 839922192;
and.b32 %r398, %r390, 30583;
prmt.b32 %r399, %r328, %r327, %r398;
prmt.b32 %r400, %r331, %r330, %r398;
and.b32 %r401, %r397, 30292;
prmt.b32 %r402, %r399, %r400, %r401;
shr.u32 %r403, %r394, 16;
and.b32 %r404, %r403, 30583;
prmt.b32 %r405, %r328, %r327, %r404;
prmt.b32 %r406, %r331, %r330, %r404;
shr.u32 %r407, %r397, 16;
prmt.b32 %r408, %r405, %r406, %r407;
prmt.b32 %r307, %r402, %r408, %r341;
prmt.b32 %r311, %r402, %r408, %r342;

dp4a.s32.s32 %r102, %r307, %r87, %r98;


dp4a.s32.s32 %r106, %r311, %r91, %r102;

ld.global.nc.u8 %rs36, [%rd114];
shl.b16 %rs38, %rs36, 7;
and.b16 %rs39, %rs38, 32640;
setp.eq.s16 %p6, %rs36, 255;
setp.eq.s16 %p7, %rs36, 0;
selp.b16 %rs40, 64, %rs39, %p7;
selp.b16 %rs14, 32767, %rs40, %p6;

{ mov.b32 %f87, {0,%rs14};}


mul.ftz.f32 %f116, %f87, 0f3F000000;

{.reg .f16 low,high;
mov.b32 {low,high},%r93;
cvt.f32.f16 %f88, low;}


mul.ftz.f32 %f117, %f116, %f88;
cvt.rn.f32.s32 %f118, %r106;
fma.rn.ftz.f32 %f335, %f117, %f118, %f335;
cvt.u64.u32 %rd38, %r667;
add.s64 %rd39, %rd1, %rd38;
mul.lo.s64 %rd40, %rd39, 36;
add.s64 %rd41, %rd7, %rd40;
add.s64 %rd42, %rd41, %rd35;
ld.global.nc.u32 %r113, [%rd42+4];

dp4a.s32.s32 %r111, %r282, %r113, %r301;

ld.global.nc.u32 %r117, [%rd42+20];

dp4a.s32.s32 %r115, %r286, %r117, %r111;

ld.global.nc.u32 %r121, [%rd42+8];

dp4a.s32.s32 %r119, %r290, %r121, %r115;

ld.global.nc.u32 %r125, [%rd42+24];

dp4a.s32.s32 %r123, %r294, %r125, %r119;


{ mov.b32 %f89, {0,%rs13};}


mul.ftz.f32 %f119, %f89, 0f3F000000;
ld.global.nc.u32 %r127, [%rd41];

{.reg .f16 low,high;
mov.b32 {low,high},%r127;
cvt.f32.f16 %f90, low;}


mul.ftz.f32 %f120, %f119, %f90;
cvt.rn.f32.s32 %f121, %r123;
fma.rn.ftz.f32 %f334, %f120, %f121, %f334;

dp4a.s32.s32 %r128, %r299, %r113, %r301;


dp4a.s32.s32 %r132, %r303, %r117, %r128;


dp4a.s32.s32 %r136, %r307, %r121, %r132;


dp4a.s32.s32 %r140, %r311, %r125, %r136;


{ mov.b32 %f91, {0,%rs14};}


mul.ftz.f32 %f122, %f91, 0f3F000000;

{.reg .f16 low,high;
mov.b32 {low,high},%r127;
cvt.f32.f16 %f92, low;}


mul.ftz.f32 %f123, %f122, %f92;
cvt.rn.f32.s32 %f124, %r140;
fma.rn.ftz.f32 %f333, %f123, %f124, %f333;
cvt.u64.u32 %rd43, %r672;
add.s64 %rd44, %rd1, %rd43;
mul.lo.s64 %rd45, %rd44, 36;
add.s64 %rd46, %rd7, %rd45;
add.s64 %rd47, %rd46, %rd35;
ld.global.nc.u32 %r147, [%rd47+4];

dp4a.s32.s32 %r145, %r282, %r147, %r301;

ld.global.nc.u32 %r151, [%rd47+20];

dp4a.s32.s32 %r149, %r286, %r151, %r145;

ld.global.nc.u32 %r155, [%rd47+8];

dp4a.s32.s32 %r153, %r290, %r155, %r149;

ld.global.nc.u32 %r159, [%rd47+24];

dp4a.s32.s32 %r157, %r294, %r159, %r153;


{ mov.b32 %f93, {0,%rs13};}


mul.ftz.f32 %f125, %f93, 0f3F000000;
ld.global.nc.u32 %r161, [%rd46];

{.reg .f16 low,high;
mov.b32 {low,high},%r161;
cvt.f32.f16 %f94, low;}


mul.ftz.f32 %f126, %f125, %f94;
cvt.rn.f32.s32 %f127, %r157;
fma.rn.ftz.f32 %f332, %f126, %f127, %f332;

dp4a.s32.s32 %r162, %r299, %r147, %r301;


dp4a.s32.s32 %r166, %r303, %r151, %r162;


dp4a.s32.s32 %r170, %r307, %r155, %r166;


dp4a.s32.s32 %r174, %r311, %r159, %r170;


{ mov.b32 %f95, {0,%rs14};}


mul.ftz.f32 %f128, %f95, 0f3F000000;

{.reg .f16 low,high;
mov.b32 {low,high},%r161;
cvt.f32.f16 %f96, low;}


mul.ftz.f32 %f129, %f128, %f96;
cvt.rn.f32.s32 %f130, %r174;
fma.rn.ftz.f32 %f331, %f129, %f130, %f331;
cvt.u64.u32 %rd48, %r671;
add.s64 %rd49, %rd1, %rd48;
mul.lo.s64 %rd50, %rd49, 36;
add.s64 %rd51, %rd7, %rd50;
add.s64 %rd52, %rd51, %rd35;
ld.global.nc.u32 %r181, [%rd52+4];

dp4a.s32.s32 %r179, %r282, %r181, %r301;

ld.global.nc.u32 %r185, [%rd52+20];

dp4a.s32.s32 %r183, %r286, %r185, %r179;

ld.global.nc.u32 %r189, [%rd52+8];

dp4a.s32.s32 %r187, %r290, %r189, %r183;

ld.global.nc.u32 %r193, [%rd52+24];

dp4a.s32.s32 %r191, %r294, %r193, %r187;


{ mov.b32 %f97, {0,%rs13};}


mul.ftz.f32 %f131, %f97, 0f3F000000;
ld.global.nc.u32 %r195, [%rd51];

{.reg .f16 low,high;
mov.b32 {low,high},%r195;
cvt.f32.f16 %f98, low;}


mul.ftz.f32 %f132, %f131, %f98;
cvt.rn.f32.s32 %f133, %r191;
fma.rn.ftz.f32 %f330, %f132, %f133, %f330;

dp4a.s32.s32 %r196, %r299, %r181, %r301;


dp4a.s32.s32 %r200, %r303, %r185, %r196;


dp4a.s32.s32 %r204, %r307, %r189, %r200;


dp4a.s32.s32 %r208, %r311, %r193, %r204;


{ mov.b32 %f99, {0,%rs14};}


mul.ftz.f32 %f134, %f99, 0f3F000000;

{.reg .f16 low,high;
mov.b32 {low,high},%r195;
cvt.f32.f16 %f100, low;}


mul.ftz.f32 %f135, %f134, %f100;
cvt.rn.f32.s32 %f136, %r208;
fma.rn.ftz.f32 %f329, %f135, %f136, %f329;
cvt.u64.u32 %rd53, %r670;
add.s64 %rd54, %rd1, %rd53;
mul.lo.s64 %rd55, %rd54, 36;
add.s64 %rd56, %rd7, %rd55;
add.s64 %rd57, %rd56, %rd35;
ld.global.nc.u32 %r215, [%rd57+4];

dp4a.s32.s32 %r213, %r282, %r215, %r301;

ld.global.nc.u32 %r219, [%rd57+20];

dp4a.s32.s32 %r217, %r286, %r219, %r213;

ld.global.nc.u32 %r223, [%rd57+8];

dp4a.s32.s32 %r221, %r290, %r223, %r217;

ld.global.nc.u32 %r227, [%rd57+24];

dp4a.s32.s32 %r225, %r294, %r227, %r221;


{ mov.b32 %f101, {0,%rs13};}


mul.ftz.f32 %f137, %f101, 0f3F000000;
ld.global.nc.u32 %r229, [%rd56];

{.reg .f16 low,high;
mov.b32 {low,high},%r229;
cvt.f32.f16 %f102, low;}


mul.ftz.f32 %f138, %f137, %f102;
cvt.rn.f32.s32 %f139, %r225;
fma.rn.ftz.f32 %f328, %f138, %f139, %f328;

dp4a.s32.s32 %r230, %r299, %r215, %r301;


dp4a.s32.s32 %r234, %r303, %r219, %r230;


dp4a.s32.s32 %r238, %r307, %r223, %r234;


dp4a.s32.s32 %r242, %r311, %r227, %r238;


{ mov.b32 %f103, {0,%rs14};}


mul.ftz.f32 %f140, %f103, 0f3F000000;

{.reg .f16 low,high;
mov.b32 {low,high},%r229;
cvt.f32.f16 %f104, low;}


mul.ftz.f32 %f141, %f140, %f104;
cvt.rn.f32.s32 %f142, %r242;
fma.rn.ftz.f32 %f327, %f141, %f142, %f327;
cvt.u64.u32 %rd58, %r669;
add.s64 %rd59, %rd1, %rd58;
mul.lo.s64 %rd60, %rd59, 36;
add.s64 %rd61, %rd7, %rd60;
add.s64 %rd62, %rd61, %rd35;
ld.global.nc.u32 %r249, [%rd62+4];

dp4a.s32.s32 %r247, %r282, %r249, %r301;

ld.global.nc.u32 %r253, [%rd62+20];

dp4a.s32.s32 %r251, %r286, %r253, %r247;

ld.global.nc.u32 %r257, [%rd62+8];

dp4a.s32.s32 %r255, %r290, %r257, %r251;

ld.global.nc.u32 %r261, [%rd62+24];

dp4a.s32.s32 %r259, %r294, %r261, %r255;


{ mov.b32 %f105, {0,%rs13};}


mul.ftz.f32 %f143, %f105, 0f3F000000;
ld.global.nc.u32 %r263, [%rd61];

{.reg .f16 low,high;
mov.b32 {low,high},%r263;
cvt.f32.f16 %f106, low;}


mul.ftz.f32 %f144, %f143, %f106;
cvt.rn.f32.s32 %f145, %r259;
fma.rn.ftz.f32 %f326, %f144, %f145, %f326;

dp4a.s32.s32 %r264, %r299, %r249, %r301;


dp4a.s32.s32 %r268, %r303, %r253, %r264;


dp4a.s32.s32 %r272, %r307, %r257, %r268;


dp4a.s32.s32 %r276, %r311, %r261, %r272;


{ mov.b32 %f107, {0,%rs14};}


mul.ftz.f32 %f146, %f107, 0f3F000000;

{.reg .f16 low,high;
mov.b32 {low,high},%r263;
cvt.f32.f16 %f108, low;}


mul.ftz.f32 %f147, %f146, %f108;
cvt.rn.f32.s32 %f148, %r276;
fma.rn.ftz.f32 %f325, %f147, %f148, %f325;
cvt.u64.u32 %rd63, %r668;
add.s64 %rd64, %rd1, %rd63;
mul.lo.s64 %rd65, %rd64, 36;
add.s64 %rd66, %rd7, %rd65;
add.s64 %rd67, %rd66, %rd35;
ld.global.nc.u32 %r283, [%rd67+4];

dp4a.s32.s32 %r281, %r282, %r283, %r301;

ld.global.nc.u32 %r287, [%rd67+20];

dp4a.s32.s32 %r285, %r286, %r287, %r281;

ld.global.nc.u32 %r291, [%rd67+8];

dp4a.s32.s32 %r289, %r290, %r291, %r285;

ld.global.nc.u32 %r295, [%rd67+24];

dp4a.s32.s32 %r293, %r294, %r295, %r289;


{ mov.b32 %f109, {0,%rs13};}


mul.ftz.f32 %f149, %f109, 0f3F000000;
ld.global.nc.u32 %r297, [%rd66];

{.reg .f16 low,high;
mov.b32 {low,high},%r297;
cvt.f32.f16 %f110, low;}


mul.ftz.f32 %f150, %f149, %f110;
cvt.rn.f32.s32 %f151, %r293;
fma.rn.ftz.f32 %f324, %f150, %f151, %f324;

dp4a.s32.s32 %r298, %r299, %r283, %r301;


dp4a.s32.s32 %r302, %r303, %r287, %r298;


dp4a.s32.s32 %r306, %r307, %r291, %r302;


dp4a.s32.s32 %r310, %r311, %r295, %r306;


{ mov.b32 %f111, {0,%rs14};}


mul.ftz.f32 %f152, %f111, 0f3F000000;

{.reg .f16 low,high;
mov.b32 {low,high},%r297;
cvt.f32.f16 %f112, low;}


mul.ftz.f32 %f153, %f152, %f112;
cvt.rn.f32.s32 %f154, %r310;
fma.rn.ftz.f32 %f323, %f153, %f154, %f323;
add.s32 %r672, %r672, 32;
add.s32 %r671, %r671, 32;
add.s32 %r670, %r670, 32;
add.s32 %r669, %r669, 32;
add.s32 %r668, %r668, 32;
add.s32 %r667, %r667, 32;
add.s64 %rd114, %rd114, 544;
add.s64 %rd113, %rd113, 544;
add.s64 %rd112, %rd112, 544;
add.s32 %r673, %r673, 32;
setp.lt.s32 %p8, %r673, %r6;
@%p8 bra $L__BB112_2;

$L__BB112_3:
setp.eq.s32 %p9, %r1, 0;
@%p9 bra $L__BB112_5;

mov.u32 %r409, _ZZ13mul_mat_vec_qIL9ggml_type39ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r410, %r1, 1792, %r409;
shl.b32 %r411, %r2, 2;
add.s32 %r412, %r410, %r411;
st.shared.f32 [%r412+-1792], %f336;
st.shared.f32 [%r412+-1664], %f335;
st.shared.f32 [%r412+-1536], %f334;
st.shared.f32 [%r412+-1408], %f333;
st.shared.f32 [%r412+-1280], %f332;
st.shared.f32 [%r412+-1152], %f331;
st.shared.f32 [%r412+-1024], %f330;
st.shared.f32 [%r412+-896], %f329;
st.shared.f32 [%r412+-768], %f328;
st.shared.f32 [%r412+-640], %f327;
st.shared.f32 [%r412+-512], %f326;
st.shared.f32 [%r412+-384], %f325;
st.shared.f32 [%r412+-256], %f324;
st.shared.f32 [%r412+-128], %f323;

$L__BB112_5:
setp.ne.s32 %p10, %r1, 0;
bar.sync 0;
@%p10 bra $L__BB112_34;

mov.u32 %r413, %ctaid.z;
mov.u32 %r414, %ctaid.y;
mad.lo.s32 %r415, %r414, %r38, %r4;
mad.lo.s32 %r416, %r413, %r44, %r415;
cvt.u64.u32 %rd14, %r416;
shl.b32 %r417, %r2, 2;
mov.u32 %r418, _ZZ13mul_mat_vec_qIL9ggml_type39ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r27, %r418, %r417;
mov.u32 %r419, 2;
ld.shared.f32 %f155, [%r27];
add.ftz.f32 %f156, %f155, %f336;
mov.b32 %r420, %f156;
mov.u32 %r421, 31;
mov.u32 %r422, 16;
mov.u32 %r423, -1;
shfl.sync.bfly.b32 %r424|%p11, %r420, %r422, %r421, %r423;
mov.b32 %f157, %r424;
add.ftz.f32 %f158, %f156, %f157;
mov.b32 %r425, %f158;
mov.u32 %r426, 8;
shfl.sync.bfly.b32 %r427|%p12, %r425, %r426, %r421, %r423;
mov.b32 %f159, %r427;
add.ftz.f32 %f160, %f158, %f159;
mov.b32 %r428, %f160;
mov.u32 %r429, 4;
shfl.sync.bfly.b32 %r430|%p13, %r428, %r429, %r421, %r423;
mov.b32 %f161, %r430;
add.ftz.f32 %f162, %f160, %f161;
mov.b32 %r431, %f162;
shfl.sync.bfly.b32 %r432|%p14, %r431, %r419, %r421, %r423;
mov.b32 %f163, %r432;
add.ftz.f32 %f164, %f162, %f163;
mov.b32 %r433, %f164;
mov.u32 %r434, 1;
shfl.sync.bfly.b32 %r435|%p15, %r433, %r434, %r421, %r423;
mov.b32 %f165, %r435;
add.ftz.f32 %f43, %f164, %f165;
setp.lt.u32 %p16, %r4, %r32;
setp.eq.s32 %p17, %r2, 0;
and.pred %p1, %p17, %p16;
cvta.to.global.u64 %rd68, %rd21;
mul.wide.u32 %rd69, %r416, 4;
add.s64 %rd15, %rd68, %rd69;
not.pred %p18, %p1;
@%p18 bra $L__BB112_8;

st.global.f32 [%rd15], %f43;

$L__BB112_8:
ld.shared.f32 %f166, [%r27+128];
add.ftz.f32 %f167, %f166, %f335;
mov.b32 %r436, %f167;
shfl.sync.bfly.b32 %r440|%p19, %r436, %r422, %r421, %r423;
mov.b32 %f168, %r440;
add.ftz.f32 %f169, %f167, %f168;
mov.b32 %r441, %f169;
shfl.sync.bfly.b32 %r443|%p20, %r441, %r426, %r421, %r423;
mov.b32 %f170, %r443;
add.ftz.f32 %f171, %f169, %f170;
mov.b32 %r444, %f171;
shfl.sync.bfly.b32 %r446|%p21, %r444, %r429, %r421, %r423;
mov.b32 %f172, %r446;
add.ftz.f32 %f173, %f171, %f172;
mov.b32 %r447, %f173;
shfl.sync.bfly.b32 %r449|%p22, %r447, %r419, %r421, %r423;
mov.b32 %f174, %r449;
add.ftz.f32 %f175, %f173, %f174;
mov.b32 %r450, %f175;
shfl.sync.bfly.b32 %r452|%p23, %r450, %r434, %r421, %r423;
mov.b32 %f176, %r452;
add.ftz.f32 %f44, %f175, %f176;
setp.eq.s32 %p24, %r2, 1;
or.b32 %r456, %r4, 1;
setp.lt.u32 %p25, %r456, %r32;
and.pred %p2, %p24, %p25;
not.pred %p26, %p2;
@%p26 bra $L__BB112_10;

st.global.f32 [%rd15+4], %f44;

$L__BB112_10:
ld.shared.f32 %f177, [%r27+256];
add.ftz.f32 %f178, %f177, %f334;
mov.b32 %r457, %f178;
mov.u32 %r458, 31;
mov.u32 %r459, 16;
mov.u32 %r460, -1;
shfl.sync.bfly.b32 %r461|%p27, %r457, %r459, %r458, %r460;
mov.b32 %f179, %r461;
add.ftz.f32 %f180, %f178, %f179;
mov.b32 %r462, %f180;
mov.u32 %r463, 8;
shfl.sync.bfly.b32 %r464|%p28, %r462, %r463, %r458, %r460;
mov.b32 %f181, %r464;
add.ftz.f32 %f182, %f180, %f181;
mov.b32 %r465, %f182;
mov.u32 %r466, 4;
shfl.sync.bfly.b32 %r467|%p29, %r465, %r466, %r458, %r460;
mov.b32 %f183, %r467;
add.ftz.f32 %f184, %f182, %f183;
mov.b32 %r468, %f184;
mov.u32 %r469, 2;
shfl.sync.bfly.b32 %r470|%p30, %r468, %r469, %r458, %r460;
mov.b32 %f185, %r470;
add.ftz.f32 %f186, %f184, %f185;
mov.b32 %r471, %f186;
mov.u32 %r472, 1;
shfl.sync.bfly.b32 %r473|%p31, %r471, %r472, %r458, %r460;
mov.b32 %f187, %r473;
add.ftz.f32 %f45, %f186, %f187;
@%p18 bra $L__BB112_12;

cvt.u64.u32 %rd70, %r32;
add.s64 %rd71, %rd70, %rd14;
shl.b64 %rd73, %rd71, 2;
add.s64 %rd74, %rd68, %rd73;
st.global.f32 [%rd74], %f45;

$L__BB112_12:
ld.shared.f32 %f188, [%r27+384];
add.ftz.f32 %f189, %f188, %f333;
mov.b32 %r474, %f189;
shfl.sync.bfly.b32 %r478|%p33, %r474, %r459, %r458, %r460;
mov.b32 %f190, %r478;
add.ftz.f32 %f191, %f189, %f190;
mov.b32 %r479, %f191;
shfl.sync.bfly.b32 %r481|%p34, %r479, %r463, %r458, %r460;
mov.b32 %f192, %r481;
add.ftz.f32 %f193, %f191, %f192;
mov.b32 %r482, %f193;
shfl.sync.bfly.b32 %r484|%p35, %r482, %r466, %r458, %r460;
mov.b32 %f194, %r484;
add.ftz.f32 %f195, %f193, %f194;
mov.b32 %r485, %f195;
shfl.sync.bfly.b32 %r487|%p36, %r485, %r469, %r458, %r460;
mov.b32 %f196, %r487;
add.ftz.f32 %f197, %f195, %f196;
mov.b32 %r488, %f197;
shfl.sync.bfly.b32 %r490|%p37, %r488, %r472, %r458, %r460;
mov.b32 %f198, %r490;
add.ftz.f32 %f46, %f197, %f198;
@%p26 bra $L__BB112_14;

add.s32 %r491, %r32, 1;
cvt.u64.u32 %rd75, %r491;
add.s64 %rd76, %rd75, %rd14;
shl.b64 %rd78, %rd76, 2;
add.s64 %rd79, %rd68, %rd78;
st.global.f32 [%rd79], %f46;

$L__BB112_14:
shl.b32 %r492, %r32, 1;
mov.u32 %r493, 1;
ld.shared.f32 %f199, [%r27+512];
add.ftz.f32 %f200, %f199, %f332;
mov.b32 %r494, %f200;
mov.u32 %r495, 31;
mov.u32 %r496, 16;
mov.u32 %r497, -1;
shfl.sync.bfly.b32 %r498|%p39, %r494, %r496, %r495, %r497;
mov.b32 %f201, %r498;
add.ftz.f32 %f202, %f200, %f201;
mov.b32 %r499, %f202;
mov.u32 %r500, 8;
shfl.sync.bfly.b32 %r501|%p40, %r499, %r500, %r495, %r497;
mov.b32 %f203, %r501;
add.ftz.f32 %f204, %f202, %f203;
mov.b32 %r502, %f204;
mov.u32 %r503, 4;
shfl.sync.bfly.b32 %r504|%p41, %r502, %r503, %r495, %r497;
mov.b32 %f205, %r504;
add.ftz.f32 %f206, %f204, %f205;
mov.b32 %r505, %f206;
mov.u32 %r506, 2;
shfl.sync.bfly.b32 %r507|%p42, %r505, %r506, %r495, %r497;
mov.b32 %f207, %r507;
add.ftz.f32 %f208, %f206, %f207;
mov.b32 %r508, %f208;
shfl.sync.bfly.b32 %r509|%p43, %r508, %r493, %r495, %r497;
mov.b32 %f209, %r509;
add.ftz.f32 %f47, %f208, %f209;
cvt.u64.u32 %rd80, %r492;
add.s64 %rd81, %rd80, %rd14;
shl.b64 %rd83, %rd81, 2;
add.s64 %rd16, %rd68, %rd83;
@%p18 bra $L__BB112_16;

st.global.f32 [%rd16], %f47;

$L__BB112_16:
ld.shared.f32 %f210, [%r27+640];
add.ftz.f32 %f211, %f210, %f331;
mov.b32 %r510, %f211;
shfl.sync.bfly.b32 %r514|%p45, %r510, %r496, %r495, %r497;
mov.b32 %f212, %r514;
add.ftz.f32 %f213, %f211, %f212;
mov.b32 %r515, %f213;
shfl.sync.bfly.b32 %r517|%p46, %r515, %r500, %r495, %r497;
mov.b32 %f214, %r517;
add.ftz.f32 %f215, %f213, %f214;
mov.b32 %r518, %f215;
shfl.sync.bfly.b32 %r520|%p47, %r518, %r503, %r495, %r497;
mov.b32 %f216, %r520;
add.ftz.f32 %f217, %f215, %f216;
mov.b32 %r521, %f217;
shfl.sync.bfly.b32 %r523|%p48, %r521, %r506, %r495, %r497;
mov.b32 %f218, %r523;
add.ftz.f32 %f219, %f217, %f218;
mov.b32 %r524, %f219;
shfl.sync.bfly.b32 %r526|%p49, %r524, %r493, %r495, %r497;
mov.b32 %f220, %r526;
add.ftz.f32 %f48, %f219, %f220;
@%p26 bra $L__BB112_18;

st.global.f32 [%rd16+4], %f48;

$L__BB112_18:
mul.lo.s32 %r28, %r32, 3;
ld.shared.f32 %f221, [%r27+768];
add.ftz.f32 %f222, %f221, %f330;
mov.b32 %r527, %f222;
mov.u32 %r528, 31;
mov.u32 %r529, 16;
mov.u32 %r530, -1;
shfl.sync.bfly.b32 %r531|%p51, %r527, %r529, %r528, %r530;
mov.b32 %f223, %r531;
add.ftz.f32 %f224, %f222, %f223;
mov.b32 %r532, %f224;
mov.u32 %r533, 8;
shfl.sync.bfly.b32 %r534|%p52, %r532, %r533, %r528, %r530;
mov.b32 %f225, %r534;
add.ftz.f32 %f226, %f224, %f225;
mov.b32 %r535, %f226;
mov.u32 %r536, 4;
shfl.sync.bfly.b32 %r537|%p53, %r535, %r536, %r528, %r530;
mov.b32 %f227, %r537;
add.ftz.f32 %f228, %f226, %f227;
mov.b32 %r538, %f228;
mov.u32 %r539, 2;
shfl.sync.bfly.b32 %r540|%p54, %r538, %r539, %r528, %r530;
mov.b32 %f229, %r540;
add.ftz.f32 %f230, %f228, %f229;
mov.b32 %r541, %f230;
mov.u32 %r542, 1;
shfl.sync.bfly.b32 %r543|%p55, %r541, %r542, %r528, %r530;
mov.b32 %f231, %r543;
add.ftz.f32 %f49, %f230, %f231;
@%p18 bra $L__BB112_20;

cvt.u64.u32 %rd84, %r28;
add.s64 %rd85, %rd84, %rd14;
shl.b64 %rd87, %rd85, 2;
add.s64 %rd88, %rd68, %rd87;
st.global.f32 [%rd88], %f49;

$L__BB112_20:
ld.shared.f32 %f232, [%r27+896];
add.ftz.f32 %f233, %f232, %f329;
mov.b32 %r544, %f233;
shfl.sync.bfly.b32 %r548|%p57, %r544, %r529, %r528, %r530;
mov.b32 %f234, %r548;
add.ftz.f32 %f235, %f233, %f234;
mov.b32 %r549, %f235;
shfl.sync.bfly.b32 %r551|%p58, %r549, %r533, %r528, %r530;
mov.b32 %f236, %r551;
add.ftz.f32 %f237, %f235, %f236;
mov.b32 %r552, %f237;
shfl.sync.bfly.b32 %r554|%p59, %r552, %r536, %r528, %r530;
mov.b32 %f238, %r554;
add.ftz.f32 %f239, %f237, %f238;
mov.b32 %r555, %f239;
shfl.sync.bfly.b32 %r557|%p60, %r555, %r539, %r528, %r530;
mov.b32 %f240, %r557;
add.ftz.f32 %f241, %f239, %f240;
mov.b32 %r558, %f241;
shfl.sync.bfly.b32 %r560|%p61, %r558, %r542, %r528, %r530;
mov.b32 %f242, %r560;
add.ftz.f32 %f50, %f241, %f242;
@%p26 bra $L__BB112_22;

add.s32 %r561, %r28, 1;
cvt.u64.u32 %rd89, %r561;
add.s64 %rd90, %rd89, %rd14;
shl.b64 %rd92, %rd90, 2;
add.s64 %rd93, %rd68, %rd92;
st.global.f32 [%rd93], %f50;

$L__BB112_22:
shl.b32 %r562, %r32, 2;
mov.u32 %r563, 2;
ld.shared.f32 %f243, [%r27+1024];
add.ftz.f32 %f244, %f243, %f328;
mov.b32 %r564, %f244;
mov.u32 %r565, 31;
mov.u32 %r566, 16;
mov.u32 %r567, -1;
shfl.sync.bfly.b32 %r568|%p63, %r564, %r566, %r565, %r567;
mov.b32 %f245, %r568;
add.ftz.f32 %f246, %f244, %f245;
mov.b32 %r569, %f246;
mov.u32 %r570, 8;
shfl.sync.bfly.b32 %r571|%p64, %r569, %r570, %r565, %r567;
mov.b32 %f247, %r571;
add.ftz.f32 %f248, %f246, %f247;
mov.b32 %r572, %f248;
mov.u32 %r573, 4;
shfl.sync.bfly.b32 %r574|%p65, %r572, %r573, %r565, %r567;
mov.b32 %f249, %r574;
add.ftz.f32 %f250, %f248, %f249;
mov.b32 %r575, %f250;
shfl.sync.bfly.b32 %r576|%p66, %r575, %r563, %r565, %r567;
mov.b32 %f251, %r576;
add.ftz.f32 %f252, %f250, %f251;
mov.b32 %r577, %f252;
mov.u32 %r578, 1;
shfl.sync.bfly.b32 %r579|%p67, %r577, %r578, %r565, %r567;
mov.b32 %f253, %r579;
add.ftz.f32 %f51, %f252, %f253;
cvt.u64.u32 %rd94, %r562;
add.s64 %rd95, %rd94, %rd14;
shl.b64 %rd97, %rd95, 2;
add.s64 %rd17, %rd68, %rd97;
@%p18 bra $L__BB112_24;

st.global.f32 [%rd17], %f51;

$L__BB112_24:
ld.shared.f32 %f254, [%r27+1152];
add.ftz.f32 %f255, %f254, %f327;
mov.b32 %r580, %f255;
shfl.sync.bfly.b32 %r584|%p69, %r580, %r566, %r565, %r567;
mov.b32 %f256, %r584;
add.ftz.f32 %f257, %f255, %f256;
mov.b32 %r585, %f257;
shfl.sync.bfly.b32 %r587|%p70, %r585, %r570, %r565, %r567;
mov.b32 %f258, %r587;
add.ftz.f32 %f259, %f257, %f258;
mov.b32 %r588, %f259;
shfl.sync.bfly.b32 %r590|%p71, %r588, %r573, %r565, %r567;
mov.b32 %f260, %r590;
add.ftz.f32 %f261, %f259, %f260;
mov.b32 %r591, %f261;
shfl.sync.bfly.b32 %r593|%p72, %r591, %r563, %r565, %r567;
mov.b32 %f262, %r593;
add.ftz.f32 %f263, %f261, %f262;
mov.b32 %r594, %f263;
shfl.sync.bfly.b32 %r596|%p73, %r594, %r578, %r565, %r567;
mov.b32 %f264, %r596;
add.ftz.f32 %f52, %f263, %f264;
@%p26 bra $L__BB112_26;

st.global.f32 [%rd17+4], %f52;

$L__BB112_26:
mul.lo.s32 %r29, %r32, 5;
ld.shared.f32 %f265, [%r27+1280];
add.ftz.f32 %f266, %f265, %f326;
mov.b32 %r597, %f266;
mov.u32 %r598, 31;
mov.u32 %r599, 16;
mov.u32 %r600, -1;
shfl.sync.bfly.b32 %r601|%p75, %r597, %r599, %r598, %r600;
mov.b32 %f267, %r601;
add.ftz.f32 %f268, %f266, %f267;
mov.b32 %r602, %f268;
mov.u32 %r603, 8;
shfl.sync.bfly.b32 %r604|%p76, %r602, %r603, %r598, %r600;
mov.b32 %f269, %r604;
add.ftz.f32 %f270, %f268, %f269;
mov.b32 %r605, %f270;
mov.u32 %r606, 4;
shfl.sync.bfly.b32 %r607|%p77, %r605, %r606, %r598, %r600;
mov.b32 %f271, %r607;
add.ftz.f32 %f272, %f270, %f271;
mov.b32 %r608, %f272;
mov.u32 %r609, 2;
shfl.sync.bfly.b32 %r610|%p78, %r608, %r609, %r598, %r600;
mov.b32 %f273, %r610;
add.ftz.f32 %f274, %f272, %f273;
mov.b32 %r611, %f274;
mov.u32 %r612, 1;
shfl.sync.bfly.b32 %r613|%p79, %r611, %r612, %r598, %r600;
mov.b32 %f275, %r613;
add.ftz.f32 %f53, %f274, %f275;
@%p18 bra $L__BB112_28;

cvt.u64.u32 %rd98, %r29;
add.s64 %rd99, %rd98, %rd14;
shl.b64 %rd101, %rd99, 2;
add.s64 %rd102, %rd68, %rd101;
st.global.f32 [%rd102], %f53;

$L__BB112_28:
ld.shared.f32 %f276, [%r27+1408];
add.ftz.f32 %f277, %f276, %f325;
mov.b32 %r614, %f277;
shfl.sync.bfly.b32 %r618|%p81, %r614, %r599, %r598, %r600;
mov.b32 %f278, %r618;
add.ftz.f32 %f279, %f277, %f278;
mov.b32 %r619, %f279;
shfl.sync.bfly.b32 %r621|%p82, %r619, %r603, %r598, %r600;
mov.b32 %f280, %r621;
add.ftz.f32 %f281, %f279, %f280;
mov.b32 %r622, %f281;
shfl.sync.bfly.b32 %r624|%p83, %r622, %r606, %r598, %r600;
mov.b32 %f282, %r624;
add.ftz.f32 %f283, %f281, %f282;
mov.b32 %r625, %f283;
shfl.sync.bfly.b32 %r627|%p84, %r625, %r609, %r598, %r600;
mov.b32 %f284, %r627;
add.ftz.f32 %f285, %f283, %f284;
mov.b32 %r628, %f285;
shfl.sync.bfly.b32 %r630|%p85, %r628, %r612, %r598, %r600;
mov.b32 %f286, %r630;
add.ftz.f32 %f54, %f285, %f286;
@%p26 bra $L__BB112_30;

add.s32 %r631, %r29, 1;
cvt.u64.u32 %rd103, %r631;
add.s64 %rd104, %rd103, %rd14;
shl.b64 %rd106, %rd104, 2;
add.s64 %rd107, %rd68, %rd106;
st.global.f32 [%rd107], %f54;

$L__BB112_30:
mul.lo.s32 %r632, %r32, 6;
ld.shared.f32 %f287, [%r27+1536];
add.ftz.f32 %f288, %f287, %f324;
mov.b32 %r633, %f288;
mov.u32 %r634, 31;
mov.u32 %r635, 16;
mov.u32 %r636, -1;
shfl.sync.bfly.b32 %r637|%p87, %r633, %r635, %r634, %r636;
mov.b32 %f289, %r637;
add.ftz.f32 %f290, %f288, %f289;
mov.b32 %r638, %f290;
mov.u32 %r639, 8;
shfl.sync.bfly.b32 %r640|%p88, %r638, %r639, %r634, %r636;
mov.b32 %f291, %r640;
add.ftz.f32 %f292, %f290, %f291;
mov.b32 %r641, %f292;
mov.u32 %r642, 4;
shfl.sync.bfly.b32 %r643|%p89, %r641, %r642, %r634, %r636;
mov.b32 %f293, %r643;
add.ftz.f32 %f294, %f292, %f293;
mov.b32 %r644, %f294;
mov.u32 %r645, 2;
shfl.sync.bfly.b32 %r646|%p90, %r644, %r645, %r634, %r636;
mov.b32 %f295, %r646;
add.ftz.f32 %f296, %f294, %f295;
mov.b32 %r647, %f296;
mov.u32 %r648, 1;
shfl.sync.bfly.b32 %r649|%p91, %r647, %r648, %r634, %r636;
mov.b32 %f297, %r649;
add.ftz.f32 %f55, %f296, %f297;
cvt.u64.u32 %rd108, %r632;
add.s64 %rd109, %rd108, %rd14;
shl.b64 %rd111, %rd109, 2;
add.s64 %rd18, %rd68, %rd111;
@%p18 bra $L__BB112_32;

st.global.f32 [%rd18], %f55;

$L__BB112_32:
ld.shared.f32 %f298, [%r27+1664];
add.ftz.f32 %f299, %f298, %f323;
mov.b32 %r650, %f299;
shfl.sync.bfly.b32 %r654|%p93, %r650, %r635, %r634, %r636;
mov.b32 %f300, %r654;
add.ftz.f32 %f301, %f299, %f300;
mov.b32 %r655, %f301;
shfl.sync.bfly.b32 %r657|%p94, %r655, %r639, %r634, %r636;
mov.b32 %f302, %r657;
add.ftz.f32 %f303, %f301, %f302;
mov.b32 %r658, %f303;
shfl.sync.bfly.b32 %r660|%p95, %r658, %r642, %r634, %r636;
mov.b32 %f304, %r660;
add.ftz.f32 %f305, %f303, %f304;
mov.b32 %r661, %f305;
shfl.sync.bfly.b32 %r663|%p96, %r661, %r645, %r634, %r636;
mov.b32 %f306, %r663;
add.ftz.f32 %f307, %f305, %f306;
mov.b32 %r664, %f307;
shfl.sync.bfly.b32 %r666|%p97, %r664, %r648, %r634, %r636;
mov.b32 %f308, %r666;
add.ftz.f32 %f56, %f307, %f308;
@%p26 bra $L__BB112_34;

st.global.f32 [%rd18+4], %f56;

$L__BB112_34:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type39ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type39ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type39ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type39ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type39ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type39ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type39ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type39ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type39ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type39ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<111>;
.reg .b16 %rs<43>;
.reg .f32 %f<385>;
.reg .b32 %r<749>;
.reg .b64 %rd<130>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type39ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[2048];

ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type39ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd20, [_Z13mul_mat_vec_qIL9ggml_type39ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd21, [_Z13mul_mat_vec_qIL9ggml_type39ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r49, [_Z13mul_mat_vec_qIL9ggml_type39ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type39ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type39ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type39ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type39ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type39ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type39ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r46, [_Z13mul_mat_vec_qIL9ggml_type39ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r47, [_Z13mul_mat_vec_qIL9ggml_type39ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r48, [_Z13mul_mat_vec_qIL9ggml_type39ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type39ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type39ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type39ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type39ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r50, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r50, %r2;
mov.u32 %r51, %ctaid.x;
shl.b32 %r4, %r51, 1;
shr.u32 %r52, %r3, 31;
add.s32 %r53, %r3, %r52;
shr.s32 %r748, %r53, 1;
shr.u32 %r6, %r49, 5;
setp.ge.s32 %p3, %r748, %r6;
mov.f32 %f369, 0f00000000;
mov.f32 %f370, %f369;
mov.f32 %f371, %f369;
mov.f32 %f372, %f369;
mov.f32 %f373, %f369;
mov.f32 %f374, %f369;
mov.f32 %f375, %f369;
mov.f32 %f376, %f369;
mov.f32 %f377, %f369;
mov.f32 %f378, %f369;
mov.f32 %f379, %f369;
mov.f32 %f380, %f369;
mov.f32 %f381, %f369;
mov.f32 %f382, %f369;
mov.f32 %f383, %f369;
mov.f32 %f384, %f369;
@%p3 bra $L__BB113_3;

mov.u32 %r54, %ctaid.y;
mov.u32 %r55, %ctaid.z;
mul.lo.s32 %r56, %r54, %r41;
cvt.u64.u32 %rd22, %r56;
mul.lo.s32 %r57, %r55, %r47;
cvt.u64.u32 %rd23, %r57;
add.s64 %rd1, %rd23, %rd22;
mul.hi.u32 %r58, %r55, %r43;
add.s32 %r59, %r58, %r55;
shr.u32 %r60, %r59, %r44;
mul.lo.s32 %r61, %r60, %r46;
mul.hi.u32 %r62, %r54, %r37;
add.s32 %r63, %r62, %r54;
shr.u32 %r64, %r63, %r38;
mul.lo.s32 %r65, %r64, %r40;
cvta.to.global.u64 %rd24, %rd19;
and.b32 %r68, %r53, -2;
sub.s32 %r69, %r3, %r68;
shl.b32 %r70, %r69, 1;
cvt.s64.s32 %rd2, %r70;
shl.b32 %r71, %r69, 3;
cvt.s64.s32 %rd25, %r71;
shl.b32 %r72, %r35, 1;
add.s32 %r747, %r748, %r72;
mad.lo.s32 %r746, %r35, 3, %r748;
shl.b32 %r73, %r35, 2;
add.s32 %r745, %r748, %r73;
mad.lo.s32 %r744, %r35, 5, %r748;
mad.lo.s32 %r743, %r35, 6, %r748;
mad.lo.s32 %r742, %r35, 7, %r748;
add.s32 %r741, %r748, %r35;
add.s32 %r74, %r748, %r61;
add.s32 %r75, %r74, %r65;
add.s32 %r76, %r4, 1;
mad.lo.s32 %r77, %r34, %r76, %r75;
mul.wide.s32 %rd26, %r77, 17;
add.s64 %rd129, %rd24, %rd26;
mad.lo.s32 %r78, %r4, %r34, %r748;
add.s32 %r79, %r78, %r61;
add.s32 %r80, %r79, %r65;
mul.wide.s32 %rd27, %r80, 17;
add.s64 %rd28, %rd24, %rd27;
add.s64 %rd128, %rd28, 4;
add.s64 %rd29, %rd27, %rd25;
add.s64 %rd30, %rd24, %rd29;
add.s64 %rd127, %rd30, 4;
or.b64 %rd6, %rd25, 4;
cvta.to.global.u64 %rd7, %rd20;
mov.f32 %f369, 0f00000000;
shl.b64 %rd35, %rd2, 2;

$L__BB113_2:
cvt.u64.u32 %rd31, %r748;
add.s64 %rd32, %rd1, %rd31;
mul.lo.s64 %rd33, %rd32, 36;
add.s64 %rd34, %rd7, %rd33;
add.s64 %rd36, %rd34, %rd35;
ld.global.nc.u8 %rs17, [%rd127+-3];
cvt.u32.u16 %r353, %rs17;
and.b32 %r354, %r353, 255;
ld.global.nc.u8 %rs18, [%rd127+-2];
cvt.u32.u16 %r355, %rs18;
prmt.b32 %r356, %r355, %r354, 30212;
ld.global.nc.u8 %rs19, [%rd127+-1];
cvt.u32.u16 %r357, %rs19;
prmt.b32 %r358, %r357, %r356, 28756;
ld.global.nc.u8 %rs20, [%rd127];
cvt.u32.u16 %r359, %rs20;
prmt.b32 %r360, %r359, %r358, 1620;
and.b32 %r361, %r360, -2004318072;
shr.u32 %r362, %r361, 1;
or.b32 %r363, %r362, 839922192;
and.b32 %r364, %r356, 30583;
mov.u32 %r365, 201852420;
mov.u32 %r366, 50462976;
prmt.b32 %r367, %r366, %r365, %r364;
mov.u32 %r368, -185009412;
mov.u32 %r369, -33620224;
prmt.b32 %r370, %r369, %r368, %r364;
and.b32 %r371, %r363, 30292;
prmt.b32 %r372, %r367, %r370, %r371;
shr.u32 %r373, %r360, 16;
and.b32 %r374, %r373, 30583;
prmt.b32 %r375, %r366, %r365, %r374;
prmt.b32 %r376, %r369, %r368, %r374;
shr.u32 %r377, %r363, 16;
prmt.b32 %r378, %r375, %r376, %r377;
mov.u32 %r379, 25632;
prmt.b32 %r320, %r372, %r378, %r379;
mov.u32 %r380, 30001;
prmt.b32 %r324, %r372, %r378, %r380;
ld.global.nc.u32 %r83, [%rd36+4];
mov.u32 %r339, 0;

dp4a.s32.s32 %r81, %r320, %r83, %r339;

ld.global.nc.u32 %r87, [%rd36+20];

dp4a.s32.s32 %r85, %r324, %r87, %r81;

ld.global.nc.u8 %rs21, [%rd127+1];
cvt.u32.u16 %r381, %rs21;
and.b32 %r382, %r381, 255;
ld.global.nc.u8 %rs22, [%rd127+2];
cvt.u32.u16 %r383, %rs22;
prmt.b32 %r384, %r383, %r382, 30212;
ld.global.nc.u8 %rs23, [%rd127+3];
cvt.u32.u16 %r385, %rs23;
prmt.b32 %r386, %r385, %r384, 28756;
ld.global.nc.u8 %rs24, [%rd127+4];
cvt.u32.u16 %r387, %rs24;
prmt.b32 %r388, %r387, %r386, 1620;
and.b32 %r389, %r388, -2004318072;
shr.u32 %r390, %r389, 1;
or.b32 %r391, %r390, 839922192;
and.b32 %r392, %r384, 30583;
prmt.b32 %r393, %r366, %r365, %r392;
prmt.b32 %r394, %r369, %r368, %r392;
and.b32 %r395, %r391, 30292;
prmt.b32 %r396, %r393, %r394, %r395;
shr.u32 %r397, %r388, 16;
and.b32 %r398, %r397, 30583;
prmt.b32 %r399, %r366, %r365, %r398;
prmt.b32 %r400, %r369, %r368, %r398;
shr.u32 %r401, %r391, 16;
prmt.b32 %r402, %r399, %r400, %r401;
prmt.b32 %r328, %r396, %r402, %r379;
prmt.b32 %r332, %r396, %r402, %r380;
ld.global.nc.u32 %r91, [%rd36+8];

dp4a.s32.s32 %r89, %r328, %r91, %r85;

ld.global.nc.u32 %r95, [%rd36+24];

dp4a.s32.s32 %r93, %r332, %r95, %r89;

ld.global.nc.u8 %rs25, [%rd128+-4];
shl.b16 %rs27, %rs25, 7;
and.b16 %rs28, %rs27, 32640;
setp.eq.s16 %p4, %rs25, 255;
setp.eq.s16 %p5, %rs25, 0;
selp.b16 %rs29, 64, %rs28, %p5;
selp.b16 %rs15, 32767, %rs29, %p4;

{ mov.b32 %f97, {0,%rs15};}


mul.ftz.f32 %f129, %f97, 0f3F000000;
ld.global.nc.u32 %r97, [%rd34];

{.reg .f16 low,high;
mov.b32 {low,high},%r97;
cvt.f32.f16 %f98, low;}


mul.ftz.f32 %f130, %f129, %f98;
cvt.rn.f32.s32 %f131, %r93;
fma.rn.ftz.f32 %f384, %f130, %f131, %f384;
add.s64 %rd37, %rd129, %rd6;
ld.global.nc.u8 %rs30, [%rd37+-3];
cvt.u32.u16 %r403, %rs30;
and.b32 %r404, %r403, 255;
ld.global.nc.u8 %rs31, [%rd37+-2];
cvt.u32.u16 %r405, %rs31;
prmt.b32 %r406, %r405, %r404, 30212;
ld.global.nc.u8 %rs32, [%rd37+-1];
cvt.u32.u16 %r407, %rs32;
prmt.b32 %r408, %r407, %r406, 28756;
ld.global.nc.u8 %rs33, [%rd37];
cvt.u32.u16 %r409, %rs33;
prmt.b32 %r410, %r409, %r408, 1620;
and.b32 %r411, %r410, -2004318072;
shr.u32 %r412, %r411, 1;
or.b32 %r413, %r412, 839922192;
and.b32 %r414, %r406, 30583;
prmt.b32 %r415, %r366, %r365, %r414;
prmt.b32 %r416, %r369, %r368, %r414;
and.b32 %r417, %r413, 30292;
prmt.b32 %r418, %r415, %r416, %r417;
shr.u32 %r419, %r410, 16;
and.b32 %r420, %r419, 30583;
prmt.b32 %r421, %r366, %r365, %r420;
prmt.b32 %r422, %r369, %r368, %r420;
shr.u32 %r423, %r413, 16;
prmt.b32 %r424, %r421, %r422, %r423;
prmt.b32 %r337, %r418, %r424, %r379;
prmt.b32 %r341, %r418, %r424, %r380;

dp4a.s32.s32 %r98, %r337, %r83, %r339;


dp4a.s32.s32 %r102, %r341, %r87, %r98;

ld.global.nc.u8 %rs34, [%rd37+1];
cvt.u32.u16 %r425, %rs34;
and.b32 %r426, %r425, 255;
ld.global.nc.u8 %rs35, [%rd37+2];
cvt.u32.u16 %r427, %rs35;
prmt.b32 %r428, %r427, %r426, 30212;
ld.global.nc.u8 %rs36, [%rd37+3];
cvt.u32.u16 %r429, %rs36;
prmt.b32 %r430, %r429, %r428, 28756;
ld.global.nc.u8 %rs37, [%rd37+4];
cvt.u32.u16 %r431, %rs37;
prmt.b32 %r432, %r431, %r430, 1620;
and.b32 %r433, %r432, -2004318072;
shr.u32 %r434, %r433, 1;
or.b32 %r435, %r434, 839922192;
and.b32 %r436, %r428, 30583;
prmt.b32 %r437, %r366, %r365, %r436;
prmt.b32 %r438, %r369, %r368, %r436;
and.b32 %r439, %r435, 30292;
prmt.b32 %r440, %r437, %r438, %r439;
shr.u32 %r441, %r432, 16;
and.b32 %r442, %r441, 30583;
prmt.b32 %r443, %r366, %r365, %r442;
prmt.b32 %r444, %r369, %r368, %r442;
shr.u32 %r445, %r435, 16;
prmt.b32 %r446, %r443, %r444, %r445;
prmt.b32 %r345, %r440, %r446, %r379;
prmt.b32 %r349, %r440, %r446, %r380;

dp4a.s32.s32 %r106, %r345, %r91, %r102;


dp4a.s32.s32 %r110, %r349, %r95, %r106;

ld.global.nc.u8 %rs38, [%rd129];
shl.b16 %rs40, %rs38, 7;
and.b16 %rs41, %rs40, 32640;
setp.eq.s16 %p6, %rs38, 255;
setp.eq.s16 %p7, %rs38, 0;
selp.b16 %rs42, 64, %rs41, %p7;
selp.b16 %rs16, 32767, %rs42, %p6;

{ mov.b32 %f99, {0,%rs16};}


mul.ftz.f32 %f132, %f99, 0f3F000000;

{.reg .f16 low,high;
mov.b32 {low,high},%r97;
cvt.f32.f16 %f100, low;}


mul.ftz.f32 %f133, %f132, %f100;
cvt.rn.f32.s32 %f134, %r110;
fma.rn.ftz.f32 %f383, %f133, %f134, %f383;
cvt.u64.u32 %rd38, %r741;
add.s64 %rd39, %rd1, %rd38;
mul.lo.s64 %rd40, %rd39, 36;
add.s64 %rd41, %rd7, %rd40;
add.s64 %rd42, %rd41, %rd35;
ld.global.nc.u32 %r117, [%rd42+4];

dp4a.s32.s32 %r115, %r320, %r117, %r339;

ld.global.nc.u32 %r121, [%rd42+20];

dp4a.s32.s32 %r119, %r324, %r121, %r115;

ld.global.nc.u32 %r125, [%rd42+8];

dp4a.s32.s32 %r123, %r328, %r125, %r119;

ld.global.nc.u32 %r129, [%rd42+24];

dp4a.s32.s32 %r127, %r332, %r129, %r123;


{ mov.b32 %f101, {0,%rs15};}


mul.ftz.f32 %f135, %f101, 0f3F000000;
ld.global.nc.u32 %r131, [%rd41];

{.reg .f16 low,high;
mov.b32 {low,high},%r131;
cvt.f32.f16 %f102, low;}


mul.ftz.f32 %f136, %f135, %f102;
cvt.rn.f32.s32 %f137, %r127;
fma.rn.ftz.f32 %f382, %f136, %f137, %f382;

dp4a.s32.s32 %r132, %r337, %r117, %r339;


dp4a.s32.s32 %r136, %r341, %r121, %r132;


dp4a.s32.s32 %r140, %r345, %r125, %r136;


dp4a.s32.s32 %r144, %r349, %r129, %r140;


{ mov.b32 %f103, {0,%rs16};}


mul.ftz.f32 %f138, %f103, 0f3F000000;

{.reg .f16 low,high;
mov.b32 {low,high},%r131;
cvt.f32.f16 %f104, low;}


mul.ftz.f32 %f139, %f138, %f104;
cvt.rn.f32.s32 %f140, %r144;
fma.rn.ftz.f32 %f381, %f139, %f140, %f381;
cvt.u64.u32 %rd43, %r747;
add.s64 %rd44, %rd1, %rd43;
mul.lo.s64 %rd45, %rd44, 36;
add.s64 %rd46, %rd7, %rd45;
add.s64 %rd47, %rd46, %rd35;
ld.global.nc.u32 %r151, [%rd47+4];

dp4a.s32.s32 %r149, %r320, %r151, %r339;

ld.global.nc.u32 %r155, [%rd47+20];

dp4a.s32.s32 %r153, %r324, %r155, %r149;

ld.global.nc.u32 %r159, [%rd47+8];

dp4a.s32.s32 %r157, %r328, %r159, %r153;

ld.global.nc.u32 %r163, [%rd47+24];

dp4a.s32.s32 %r161, %r332, %r163, %r157;


{ mov.b32 %f105, {0,%rs15};}


mul.ftz.f32 %f141, %f105, 0f3F000000;
ld.global.nc.u32 %r165, [%rd46];

{.reg .f16 low,high;
mov.b32 {low,high},%r165;
cvt.f32.f16 %f106, low;}


mul.ftz.f32 %f142, %f141, %f106;
cvt.rn.f32.s32 %f143, %r161;
fma.rn.ftz.f32 %f380, %f142, %f143, %f380;

dp4a.s32.s32 %r166, %r337, %r151, %r339;


dp4a.s32.s32 %r170, %r341, %r155, %r166;


dp4a.s32.s32 %r174, %r345, %r159, %r170;


dp4a.s32.s32 %r178, %r349, %r163, %r174;


{ mov.b32 %f107, {0,%rs16};}


mul.ftz.f32 %f144, %f107, 0f3F000000;

{.reg .f16 low,high;
mov.b32 {low,high},%r165;
cvt.f32.f16 %f108, low;}


mul.ftz.f32 %f145, %f144, %f108;
cvt.rn.f32.s32 %f146, %r178;
fma.rn.ftz.f32 %f379, %f145, %f146, %f379;
cvt.u64.u32 %rd48, %r746;
add.s64 %rd49, %rd1, %rd48;
mul.lo.s64 %rd50, %rd49, 36;
add.s64 %rd51, %rd7, %rd50;
add.s64 %rd52, %rd51, %rd35;
ld.global.nc.u32 %r185, [%rd52+4];

dp4a.s32.s32 %r183, %r320, %r185, %r339;

ld.global.nc.u32 %r189, [%rd52+20];

dp4a.s32.s32 %r187, %r324, %r189, %r183;

ld.global.nc.u32 %r193, [%rd52+8];

dp4a.s32.s32 %r191, %r328, %r193, %r187;

ld.global.nc.u32 %r197, [%rd52+24];

dp4a.s32.s32 %r195, %r332, %r197, %r191;


{ mov.b32 %f109, {0,%rs15};}


mul.ftz.f32 %f147, %f109, 0f3F000000;
ld.global.nc.u32 %r199, [%rd51];

{.reg .f16 low,high;
mov.b32 {low,high},%r199;
cvt.f32.f16 %f110, low;}


mul.ftz.f32 %f148, %f147, %f110;
cvt.rn.f32.s32 %f149, %r195;
fma.rn.ftz.f32 %f378, %f148, %f149, %f378;

dp4a.s32.s32 %r200, %r337, %r185, %r339;


dp4a.s32.s32 %r204, %r341, %r189, %r200;


dp4a.s32.s32 %r208, %r345, %r193, %r204;


dp4a.s32.s32 %r212, %r349, %r197, %r208;


{ mov.b32 %f111, {0,%rs16};}


mul.ftz.f32 %f150, %f111, 0f3F000000;

{.reg .f16 low,high;
mov.b32 {low,high},%r199;
cvt.f32.f16 %f112, low;}


mul.ftz.f32 %f151, %f150, %f112;
cvt.rn.f32.s32 %f152, %r212;
fma.rn.ftz.f32 %f377, %f151, %f152, %f377;
cvt.u64.u32 %rd53, %r745;
add.s64 %rd54, %rd1, %rd53;
mul.lo.s64 %rd55, %rd54, 36;
add.s64 %rd56, %rd7, %rd55;
add.s64 %rd57, %rd56, %rd35;
ld.global.nc.u32 %r219, [%rd57+4];

dp4a.s32.s32 %r217, %r320, %r219, %r339;

ld.global.nc.u32 %r223, [%rd57+20];

dp4a.s32.s32 %r221, %r324, %r223, %r217;

ld.global.nc.u32 %r227, [%rd57+8];

dp4a.s32.s32 %r225, %r328, %r227, %r221;

ld.global.nc.u32 %r231, [%rd57+24];

dp4a.s32.s32 %r229, %r332, %r231, %r225;


{ mov.b32 %f113, {0,%rs15};}


mul.ftz.f32 %f153, %f113, 0f3F000000;
ld.global.nc.u32 %r233, [%rd56];

{.reg .f16 low,high;
mov.b32 {low,high},%r233;
cvt.f32.f16 %f114, low;}


mul.ftz.f32 %f154, %f153, %f114;
cvt.rn.f32.s32 %f155, %r229;
fma.rn.ftz.f32 %f376, %f154, %f155, %f376;

dp4a.s32.s32 %r234, %r337, %r219, %r339;


dp4a.s32.s32 %r238, %r341, %r223, %r234;


dp4a.s32.s32 %r242, %r345, %r227, %r238;


dp4a.s32.s32 %r246, %r349, %r231, %r242;


{ mov.b32 %f115, {0,%rs16};}


mul.ftz.f32 %f156, %f115, 0f3F000000;

{.reg .f16 low,high;
mov.b32 {low,high},%r233;
cvt.f32.f16 %f116, low;}


mul.ftz.f32 %f157, %f156, %f116;
cvt.rn.f32.s32 %f158, %r246;
fma.rn.ftz.f32 %f375, %f157, %f158, %f375;
cvt.u64.u32 %rd58, %r744;
add.s64 %rd59, %rd1, %rd58;
mul.lo.s64 %rd60, %rd59, 36;
add.s64 %rd61, %rd7, %rd60;
add.s64 %rd62, %rd61, %rd35;
ld.global.nc.u32 %r253, [%rd62+4];

dp4a.s32.s32 %r251, %r320, %r253, %r339;

ld.global.nc.u32 %r257, [%rd62+20];

dp4a.s32.s32 %r255, %r324, %r257, %r251;

ld.global.nc.u32 %r261, [%rd62+8];

dp4a.s32.s32 %r259, %r328, %r261, %r255;

ld.global.nc.u32 %r265, [%rd62+24];

dp4a.s32.s32 %r263, %r332, %r265, %r259;


{ mov.b32 %f117, {0,%rs15};}


mul.ftz.f32 %f159, %f117, 0f3F000000;
ld.global.nc.u32 %r267, [%rd61];

{.reg .f16 low,high;
mov.b32 {low,high},%r267;
cvt.f32.f16 %f118, low;}


mul.ftz.f32 %f160, %f159, %f118;
cvt.rn.f32.s32 %f161, %r263;
fma.rn.ftz.f32 %f374, %f160, %f161, %f374;

dp4a.s32.s32 %r268, %r337, %r253, %r339;


dp4a.s32.s32 %r272, %r341, %r257, %r268;


dp4a.s32.s32 %r276, %r345, %r261, %r272;


dp4a.s32.s32 %r280, %r349, %r265, %r276;


{ mov.b32 %f119, {0,%rs16};}


mul.ftz.f32 %f162, %f119, 0f3F000000;

{.reg .f16 low,high;
mov.b32 {low,high},%r267;
cvt.f32.f16 %f120, low;}


mul.ftz.f32 %f163, %f162, %f120;
cvt.rn.f32.s32 %f164, %r280;
fma.rn.ftz.f32 %f373, %f163, %f164, %f373;
cvt.u64.u32 %rd63, %r743;
add.s64 %rd64, %rd1, %rd63;
mul.lo.s64 %rd65, %rd64, 36;
add.s64 %rd66, %rd7, %rd65;
add.s64 %rd67, %rd66, %rd35;
ld.global.nc.u32 %r287, [%rd67+4];

dp4a.s32.s32 %r285, %r320, %r287, %r339;

ld.global.nc.u32 %r291, [%rd67+20];

dp4a.s32.s32 %r289, %r324, %r291, %r285;

ld.global.nc.u32 %r295, [%rd67+8];

dp4a.s32.s32 %r293, %r328, %r295, %r289;

ld.global.nc.u32 %r299, [%rd67+24];

dp4a.s32.s32 %r297, %r332, %r299, %r293;


{ mov.b32 %f121, {0,%rs15};}


mul.ftz.f32 %f165, %f121, 0f3F000000;
ld.global.nc.u32 %r301, [%rd66];

{.reg .f16 low,high;
mov.b32 {low,high},%r301;
cvt.f32.f16 %f122, low;}


mul.ftz.f32 %f166, %f165, %f122;
cvt.rn.f32.s32 %f167, %r297;
fma.rn.ftz.f32 %f372, %f166, %f167, %f372;

dp4a.s32.s32 %r302, %r337, %r287, %r339;


dp4a.s32.s32 %r306, %r341, %r291, %r302;


dp4a.s32.s32 %r310, %r345, %r295, %r306;


dp4a.s32.s32 %r314, %r349, %r299, %r310;


{ mov.b32 %f123, {0,%rs16};}


mul.ftz.f32 %f168, %f123, 0f3F000000;

{.reg .f16 low,high;
mov.b32 {low,high},%r301;
cvt.f32.f16 %f124, low;}


mul.ftz.f32 %f169, %f168, %f124;
cvt.rn.f32.s32 %f170, %r314;
fma.rn.ftz.f32 %f371, %f169, %f170, %f371;
cvt.u64.u32 %rd68, %r742;
add.s64 %rd69, %rd1, %rd68;
mul.lo.s64 %rd70, %rd69, 36;
add.s64 %rd71, %rd7, %rd70;
add.s64 %rd72, %rd71, %rd35;
ld.global.nc.u32 %r321, [%rd72+4];

dp4a.s32.s32 %r319, %r320, %r321, %r339;

ld.global.nc.u32 %r325, [%rd72+20];

dp4a.s32.s32 %r323, %r324, %r325, %r319;

ld.global.nc.u32 %r329, [%rd72+8];

dp4a.s32.s32 %r327, %r328, %r329, %r323;

ld.global.nc.u32 %r333, [%rd72+24];

dp4a.s32.s32 %r331, %r332, %r333, %r327;


{ mov.b32 %f125, {0,%rs15};}


mul.ftz.f32 %f171, %f125, 0f3F000000;
ld.global.nc.u32 %r335, [%rd71];

{.reg .f16 low,high;
mov.b32 {low,high},%r335;
cvt.f32.f16 %f126, low;}


mul.ftz.f32 %f172, %f171, %f126;
cvt.rn.f32.s32 %f173, %r331;
fma.rn.ftz.f32 %f370, %f172, %f173, %f370;

dp4a.s32.s32 %r336, %r337, %r321, %r339;


dp4a.s32.s32 %r340, %r341, %r325, %r336;


dp4a.s32.s32 %r344, %r345, %r329, %r340;


dp4a.s32.s32 %r348, %r349, %r333, %r344;


{ mov.b32 %f127, {0,%rs16};}


mul.ftz.f32 %f174, %f127, 0f3F000000;

{.reg .f16 low,high;
mov.b32 {low,high},%r335;
cvt.f32.f16 %f128, low;}


mul.ftz.f32 %f175, %f174, %f128;
cvt.rn.f32.s32 %f176, %r348;
fma.rn.ftz.f32 %f369, %f175, %f176, %f369;
add.s32 %r747, %r747, 32;
add.s32 %r746, %r746, 32;
add.s32 %r745, %r745, 32;
add.s32 %r744, %r744, 32;
add.s32 %r743, %r743, 32;
add.s32 %r742, %r742, 32;
add.s32 %r741, %r741, 32;
add.s64 %rd129, %rd129, 544;
add.s64 %rd128, %rd128, 544;
add.s64 %rd127, %rd127, 544;
add.s32 %r748, %r748, 32;
setp.lt.s32 %p8, %r748, %r6;
@%p8 bra $L__BB113_2;

$L__BB113_3:
setp.eq.s32 %p9, %r1, 0;
@%p9 bra $L__BB113_5;

shl.b32 %r447, %r1, 11;
mov.u32 %r448, _ZZ13mul_mat_vec_qIL9ggml_type39ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r449, %r448, %r447;
shl.b32 %r450, %r2, 2;
add.s32 %r451, %r449, %r450;
st.shared.f32 [%r451+-2048], %f384;
st.shared.f32 [%r451+-1920], %f383;
st.shared.f32 [%r451+-1792], %f382;
st.shared.f32 [%r451+-1664], %f381;
st.shared.f32 [%r451+-1536], %f380;
st.shared.f32 [%r451+-1408], %f379;
st.shared.f32 [%r451+-1280], %f378;
st.shared.f32 [%r451+-1152], %f377;
st.shared.f32 [%r451+-1024], %f376;
st.shared.f32 [%r451+-896], %f375;
st.shared.f32 [%r451+-768], %f374;
st.shared.f32 [%r451+-640], %f373;
st.shared.f32 [%r451+-512], %f372;
st.shared.f32 [%r451+-384], %f371;
st.shared.f32 [%r451+-256], %f370;
st.shared.f32 [%r451+-128], %f369;

$L__BB113_5:
setp.ne.s32 %p10, %r1, 0;
bar.sync 0;
@%p10 bra $L__BB113_38;

mov.u32 %r452, %ctaid.z;
mov.u32 %r453, %ctaid.y;
mad.lo.s32 %r454, %r453, %r42, %r4;
mad.lo.s32 %r455, %r452, %r48, %r454;
cvt.u64.u32 %rd14, %r455;
shl.b32 %r456, %r2, 2;
mov.u32 %r457, _ZZ13mul_mat_vec_qIL9ggml_type39ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r30, %r457, %r456;
mov.u32 %r458, 2;
ld.shared.f32 %f177, [%r30];
add.ftz.f32 %f178, %f177, %f384;
mov.b32 %r459, %f178;
mov.u32 %r460, 31;
mov.u32 %r461, 16;
mov.u32 %r462, -1;
shfl.sync.bfly.b32 %r463|%p11, %r459, %r461, %r460, %r462;
mov.b32 %f179, %r463;
add.ftz.f32 %f180, %f178, %f179;
mov.b32 %r464, %f180;
mov.u32 %r465, 8;
shfl.sync.bfly.b32 %r466|%p12, %r464, %r465, %r460, %r462;
mov.b32 %f181, %r466;
add.ftz.f32 %f182, %f180, %f181;
mov.b32 %r467, %f182;
mov.u32 %r468, 4;
shfl.sync.bfly.b32 %r469|%p13, %r467, %r468, %r460, %r462;
mov.b32 %f183, %r469;
add.ftz.f32 %f184, %f182, %f183;
mov.b32 %r470, %f184;
shfl.sync.bfly.b32 %r471|%p14, %r470, %r458, %r460, %r462;
mov.b32 %f185, %r471;
add.ftz.f32 %f186, %f184, %f185;
mov.b32 %r472, %f186;
mov.u32 %r473, 1;
shfl.sync.bfly.b32 %r474|%p15, %r472, %r473, %r460, %r462;
mov.b32 %f187, %r474;
add.ftz.f32 %f49, %f186, %f187;
setp.lt.u32 %p16, %r4, %r36;
setp.eq.s32 %p17, %r2, 0;
and.pred %p1, %p17, %p16;
cvta.to.global.u64 %rd73, %rd21;
mul.wide.u32 %rd74, %r455, 4;
add.s64 %rd15, %rd73, %rd74;
not.pred %p18, %p1;
@%p18 bra $L__BB113_8;

st.global.f32 [%rd15], %f49;

$L__BB113_8:
ld.shared.f32 %f188, [%r30+128];
add.ftz.f32 %f189, %f188, %f383;
mov.b32 %r475, %f189;
shfl.sync.bfly.b32 %r479|%p19, %r475, %r461, %r460, %r462;
mov.b32 %f190, %r479;
add.ftz.f32 %f191, %f189, %f190;
mov.b32 %r480, %f191;
shfl.sync.bfly.b32 %r482|%p20, %r480, %r465, %r460, %r462;
mov.b32 %f192, %r482;
add.ftz.f32 %f193, %f191, %f192;
mov.b32 %r483, %f193;
shfl.sync.bfly.b32 %r485|%p21, %r483, %r468, %r460, %r462;
mov.b32 %f194, %r485;
add.ftz.f32 %f195, %f193, %f194;
mov.b32 %r486, %f195;
shfl.sync.bfly.b32 %r488|%p22, %r486, %r458, %r460, %r462;
mov.b32 %f196, %r488;
add.ftz.f32 %f197, %f195, %f196;
mov.b32 %r489, %f197;
shfl.sync.bfly.b32 %r491|%p23, %r489, %r473, %r460, %r462;
mov.b32 %f198, %r491;
add.ftz.f32 %f50, %f197, %f198;
setp.eq.s32 %p24, %r2, 1;
or.b32 %r495, %r4, 1;
setp.lt.u32 %p25, %r495, %r36;
and.pred %p2, %p24, %p25;
not.pred %p26, %p2;
@%p26 bra $L__BB113_10;

st.global.f32 [%rd15+4], %f50;

$L__BB113_10:
ld.shared.f32 %f199, [%r30+256];
add.ftz.f32 %f200, %f199, %f382;
mov.b32 %r496, %f200;
mov.u32 %r497, 31;
mov.u32 %r498, 16;
mov.u32 %r499, -1;
shfl.sync.bfly.b32 %r500|%p27, %r496, %r498, %r497, %r499;
mov.b32 %f201, %r500;
add.ftz.f32 %f202, %f200, %f201;
mov.b32 %r501, %f202;
mov.u32 %r502, 8;
shfl.sync.bfly.b32 %r503|%p28, %r501, %r502, %r497, %r499;
mov.b32 %f203, %r503;
add.ftz.f32 %f204, %f202, %f203;
mov.b32 %r504, %f204;
mov.u32 %r505, 4;
shfl.sync.bfly.b32 %r506|%p29, %r504, %r505, %r497, %r499;
mov.b32 %f205, %r506;
add.ftz.f32 %f206, %f204, %f205;
mov.b32 %r507, %f206;
mov.u32 %r508, 2;
shfl.sync.bfly.b32 %r509|%p30, %r507, %r508, %r497, %r499;
mov.b32 %f207, %r509;
add.ftz.f32 %f208, %f206, %f207;
mov.b32 %r510, %f208;
mov.u32 %r511, 1;
shfl.sync.bfly.b32 %r512|%p31, %r510, %r511, %r497, %r499;
mov.b32 %f209, %r512;
add.ftz.f32 %f51, %f208, %f209;
@%p18 bra $L__BB113_12;

cvt.u64.u32 %rd75, %r36;
add.s64 %rd76, %rd75, %rd14;
shl.b64 %rd78, %rd76, 2;
add.s64 %rd79, %rd73, %rd78;
st.global.f32 [%rd79], %f51;

$L__BB113_12:
ld.shared.f32 %f210, [%r30+384];
add.ftz.f32 %f211, %f210, %f381;
mov.b32 %r513, %f211;
shfl.sync.bfly.b32 %r517|%p33, %r513, %r498, %r497, %r499;
mov.b32 %f212, %r517;
add.ftz.f32 %f213, %f211, %f212;
mov.b32 %r518, %f213;
shfl.sync.bfly.b32 %r520|%p34, %r518, %r502, %r497, %r499;
mov.b32 %f214, %r520;
add.ftz.f32 %f215, %f213, %f214;
mov.b32 %r521, %f215;
shfl.sync.bfly.b32 %r523|%p35, %r521, %r505, %r497, %r499;
mov.b32 %f216, %r523;
add.ftz.f32 %f217, %f215, %f216;
mov.b32 %r524, %f217;
shfl.sync.bfly.b32 %r526|%p36, %r524, %r508, %r497, %r499;
mov.b32 %f218, %r526;
add.ftz.f32 %f219, %f217, %f218;
mov.b32 %r527, %f219;
shfl.sync.bfly.b32 %r529|%p37, %r527, %r511, %r497, %r499;
mov.b32 %f220, %r529;
add.ftz.f32 %f52, %f219, %f220;
@%p26 bra $L__BB113_14;

add.s32 %r530, %r36, 1;
cvt.u64.u32 %rd80, %r530;
add.s64 %rd81, %rd80, %rd14;
shl.b64 %rd83, %rd81, 2;
add.s64 %rd84, %rd73, %rd83;
st.global.f32 [%rd84], %f52;

$L__BB113_14:
shl.b32 %r531, %r36, 1;
mov.u32 %r532, 1;
ld.shared.f32 %f221, [%r30+512];
add.ftz.f32 %f222, %f221, %f380;
mov.b32 %r533, %f222;
mov.u32 %r534, 31;
mov.u32 %r535, 16;
mov.u32 %r536, -1;
shfl.sync.bfly.b32 %r537|%p39, %r533, %r535, %r534, %r536;
mov.b32 %f223, %r537;
add.ftz.f32 %f224, %f222, %f223;
mov.b32 %r538, %f224;
mov.u32 %r539, 8;
shfl.sync.bfly.b32 %r540|%p40, %r538, %r539, %r534, %r536;
mov.b32 %f225, %r540;
add.ftz.f32 %f226, %f224, %f225;
mov.b32 %r541, %f226;
mov.u32 %r542, 4;
shfl.sync.bfly.b32 %r543|%p41, %r541, %r542, %r534, %r536;
mov.b32 %f227, %r543;
add.ftz.f32 %f228, %f226, %f227;
mov.b32 %r544, %f228;
mov.u32 %r545, 2;
shfl.sync.bfly.b32 %r546|%p42, %r544, %r545, %r534, %r536;
mov.b32 %f229, %r546;
add.ftz.f32 %f230, %f228, %f229;
mov.b32 %r547, %f230;
shfl.sync.bfly.b32 %r548|%p43, %r547, %r532, %r534, %r536;
mov.b32 %f231, %r548;
add.ftz.f32 %f53, %f230, %f231;
cvt.u64.u32 %rd85, %r531;
add.s64 %rd86, %rd85, %rd14;
shl.b64 %rd88, %rd86, 2;
add.s64 %rd16, %rd73, %rd88;
@%p18 bra $L__BB113_16;

st.global.f32 [%rd16], %f53;

$L__BB113_16:
ld.shared.f32 %f232, [%r30+640];
add.ftz.f32 %f233, %f232, %f379;
mov.b32 %r549, %f233;
shfl.sync.bfly.b32 %r553|%p45, %r549, %r535, %r534, %r536;
mov.b32 %f234, %r553;
add.ftz.f32 %f235, %f233, %f234;
mov.b32 %r554, %f235;
shfl.sync.bfly.b32 %r556|%p46, %r554, %r539, %r534, %r536;
mov.b32 %f236, %r556;
add.ftz.f32 %f237, %f235, %f236;
mov.b32 %r557, %f237;
shfl.sync.bfly.b32 %r559|%p47, %r557, %r542, %r534, %r536;
mov.b32 %f238, %r559;
add.ftz.f32 %f239, %f237, %f238;
mov.b32 %r560, %f239;
shfl.sync.bfly.b32 %r562|%p48, %r560, %r545, %r534, %r536;
mov.b32 %f240, %r562;
add.ftz.f32 %f241, %f239, %f240;
mov.b32 %r563, %f241;
shfl.sync.bfly.b32 %r565|%p49, %r563, %r532, %r534, %r536;
mov.b32 %f242, %r565;
add.ftz.f32 %f54, %f241, %f242;
@%p26 bra $L__BB113_18;

st.global.f32 [%rd16+4], %f54;

$L__BB113_18:
mul.lo.s32 %r31, %r36, 3;
ld.shared.f32 %f243, [%r30+768];
add.ftz.f32 %f244, %f243, %f378;
mov.b32 %r566, %f244;
mov.u32 %r567, 31;
mov.u32 %r568, 16;
mov.u32 %r569, -1;
shfl.sync.bfly.b32 %r570|%p51, %r566, %r568, %r567, %r569;
mov.b32 %f245, %r570;
add.ftz.f32 %f246, %f244, %f245;
mov.b32 %r571, %f246;
mov.u32 %r572, 8;
shfl.sync.bfly.b32 %r573|%p52, %r571, %r572, %r567, %r569;
mov.b32 %f247, %r573;
add.ftz.f32 %f248, %f246, %f247;
mov.b32 %r574, %f248;
mov.u32 %r575, 4;
shfl.sync.bfly.b32 %r576|%p53, %r574, %r575, %r567, %r569;
mov.b32 %f249, %r576;
add.ftz.f32 %f250, %f248, %f249;
mov.b32 %r577, %f250;
mov.u32 %r578, 2;
shfl.sync.bfly.b32 %r579|%p54, %r577, %r578, %r567, %r569;
mov.b32 %f251, %r579;
add.ftz.f32 %f252, %f250, %f251;
mov.b32 %r580, %f252;
mov.u32 %r581, 1;
shfl.sync.bfly.b32 %r582|%p55, %r580, %r581, %r567, %r569;
mov.b32 %f253, %r582;
add.ftz.f32 %f55, %f252, %f253;
@%p18 bra $L__BB113_20;

cvt.u64.u32 %rd89, %r31;
add.s64 %rd90, %rd89, %rd14;
shl.b64 %rd92, %rd90, 2;
add.s64 %rd93, %rd73, %rd92;
st.global.f32 [%rd93], %f55;

$L__BB113_20:
ld.shared.f32 %f254, [%r30+896];
add.ftz.f32 %f255, %f254, %f377;
mov.b32 %r583, %f255;
shfl.sync.bfly.b32 %r587|%p57, %r583, %r568, %r567, %r569;
mov.b32 %f256, %r587;
add.ftz.f32 %f257, %f255, %f256;
mov.b32 %r588, %f257;
shfl.sync.bfly.b32 %r590|%p58, %r588, %r572, %r567, %r569;
mov.b32 %f258, %r590;
add.ftz.f32 %f259, %f257, %f258;
mov.b32 %r591, %f259;
shfl.sync.bfly.b32 %r593|%p59, %r591, %r575, %r567, %r569;
mov.b32 %f260, %r593;
add.ftz.f32 %f261, %f259, %f260;
mov.b32 %r594, %f261;
shfl.sync.bfly.b32 %r596|%p60, %r594, %r578, %r567, %r569;
mov.b32 %f262, %r596;
add.ftz.f32 %f263, %f261, %f262;
mov.b32 %r597, %f263;
shfl.sync.bfly.b32 %r599|%p61, %r597, %r581, %r567, %r569;
mov.b32 %f264, %r599;
add.ftz.f32 %f56, %f263, %f264;
@%p26 bra $L__BB113_22;

add.s32 %r600, %r31, 1;
cvt.u64.u32 %rd94, %r600;
add.s64 %rd95, %rd94, %rd14;
shl.b64 %rd97, %rd95, 2;
add.s64 %rd98, %rd73, %rd97;
st.global.f32 [%rd98], %f56;

$L__BB113_22:
shl.b32 %r601, %r36, 2;
mov.u32 %r602, 2;
ld.shared.f32 %f265, [%r30+1024];
add.ftz.f32 %f266, %f265, %f376;
mov.b32 %r603, %f266;
mov.u32 %r604, 31;
mov.u32 %r605, 16;
mov.u32 %r606, -1;
shfl.sync.bfly.b32 %r607|%p63, %r603, %r605, %r604, %r606;
mov.b32 %f267, %r607;
add.ftz.f32 %f268, %f266, %f267;
mov.b32 %r608, %f268;
mov.u32 %r609, 8;
shfl.sync.bfly.b32 %r610|%p64, %r608, %r609, %r604, %r606;
mov.b32 %f269, %r610;
add.ftz.f32 %f270, %f268, %f269;
mov.b32 %r611, %f270;
mov.u32 %r612, 4;
shfl.sync.bfly.b32 %r613|%p65, %r611, %r612, %r604, %r606;
mov.b32 %f271, %r613;
add.ftz.f32 %f272, %f270, %f271;
mov.b32 %r614, %f272;
shfl.sync.bfly.b32 %r615|%p66, %r614, %r602, %r604, %r606;
mov.b32 %f273, %r615;
add.ftz.f32 %f274, %f272, %f273;
mov.b32 %r616, %f274;
mov.u32 %r617, 1;
shfl.sync.bfly.b32 %r618|%p67, %r616, %r617, %r604, %r606;
mov.b32 %f275, %r618;
add.ftz.f32 %f57, %f274, %f275;
cvt.u64.u32 %rd99, %r601;
add.s64 %rd100, %rd99, %rd14;
shl.b64 %rd102, %rd100, 2;
add.s64 %rd17, %rd73, %rd102;
@%p18 bra $L__BB113_24;

st.global.f32 [%rd17], %f57;

$L__BB113_24:
ld.shared.f32 %f276, [%r30+1152];
add.ftz.f32 %f277, %f276, %f375;
mov.b32 %r619, %f277;
shfl.sync.bfly.b32 %r623|%p69, %r619, %r605, %r604, %r606;
mov.b32 %f278, %r623;
add.ftz.f32 %f279, %f277, %f278;
mov.b32 %r624, %f279;
shfl.sync.bfly.b32 %r626|%p70, %r624, %r609, %r604, %r606;
mov.b32 %f280, %r626;
add.ftz.f32 %f281, %f279, %f280;
mov.b32 %r627, %f281;
shfl.sync.bfly.b32 %r629|%p71, %r627, %r612, %r604, %r606;
mov.b32 %f282, %r629;
add.ftz.f32 %f283, %f281, %f282;
mov.b32 %r630, %f283;
shfl.sync.bfly.b32 %r632|%p72, %r630, %r602, %r604, %r606;
mov.b32 %f284, %r632;
add.ftz.f32 %f285, %f283, %f284;
mov.b32 %r633, %f285;
shfl.sync.bfly.b32 %r635|%p73, %r633, %r617, %r604, %r606;
mov.b32 %f286, %r635;
add.ftz.f32 %f58, %f285, %f286;
@%p26 bra $L__BB113_26;

st.global.f32 [%rd17+4], %f58;

$L__BB113_26:
mul.lo.s32 %r32, %r36, 5;
ld.shared.f32 %f287, [%r30+1280];
add.ftz.f32 %f288, %f287, %f374;
mov.b32 %r636, %f288;
mov.u32 %r637, 31;
mov.u32 %r638, 16;
mov.u32 %r639, -1;
shfl.sync.bfly.b32 %r640|%p75, %r636, %r638, %r637, %r639;
mov.b32 %f289, %r640;
add.ftz.f32 %f290, %f288, %f289;
mov.b32 %r641, %f290;
mov.u32 %r642, 8;
shfl.sync.bfly.b32 %r643|%p76, %r641, %r642, %r637, %r639;
mov.b32 %f291, %r643;
add.ftz.f32 %f292, %f290, %f291;
mov.b32 %r644, %f292;
mov.u32 %r645, 4;
shfl.sync.bfly.b32 %r646|%p77, %r644, %r645, %r637, %r639;
mov.b32 %f293, %r646;
add.ftz.f32 %f294, %f292, %f293;
mov.b32 %r647, %f294;
mov.u32 %r648, 2;
shfl.sync.bfly.b32 %r649|%p78, %r647, %r648, %r637, %r639;
mov.b32 %f295, %r649;
add.ftz.f32 %f296, %f294, %f295;
mov.b32 %r650, %f296;
mov.u32 %r651, 1;
shfl.sync.bfly.b32 %r652|%p79, %r650, %r651, %r637, %r639;
mov.b32 %f297, %r652;
add.ftz.f32 %f59, %f296, %f297;
@%p18 bra $L__BB113_28;

cvt.u64.u32 %rd103, %r32;
add.s64 %rd104, %rd103, %rd14;
shl.b64 %rd106, %rd104, 2;
add.s64 %rd107, %rd73, %rd106;
st.global.f32 [%rd107], %f59;

$L__BB113_28:
ld.shared.f32 %f298, [%r30+1408];
add.ftz.f32 %f299, %f298, %f373;
mov.b32 %r653, %f299;
shfl.sync.bfly.b32 %r657|%p81, %r653, %r638, %r637, %r639;
mov.b32 %f300, %r657;
add.ftz.f32 %f301, %f299, %f300;
mov.b32 %r658, %f301;
shfl.sync.bfly.b32 %r660|%p82, %r658, %r642, %r637, %r639;
mov.b32 %f302, %r660;
add.ftz.f32 %f303, %f301, %f302;
mov.b32 %r661, %f303;
shfl.sync.bfly.b32 %r663|%p83, %r661, %r645, %r637, %r639;
mov.b32 %f304, %r663;
add.ftz.f32 %f305, %f303, %f304;
mov.b32 %r664, %f305;
shfl.sync.bfly.b32 %r666|%p84, %r664, %r648, %r637, %r639;
mov.b32 %f306, %r666;
add.ftz.f32 %f307, %f305, %f306;
mov.b32 %r667, %f307;
shfl.sync.bfly.b32 %r669|%p85, %r667, %r651, %r637, %r639;
mov.b32 %f308, %r669;
add.ftz.f32 %f60, %f307, %f308;
@%p26 bra $L__BB113_30;

add.s32 %r670, %r32, 1;
cvt.u64.u32 %rd108, %r670;
add.s64 %rd109, %rd108, %rd14;
shl.b64 %rd111, %rd109, 2;
add.s64 %rd112, %rd73, %rd111;
st.global.f32 [%rd112], %f60;

$L__BB113_30:
mul.lo.s32 %r671, %r36, 6;
ld.shared.f32 %f309, [%r30+1536];
add.ftz.f32 %f310, %f309, %f372;
mov.b32 %r672, %f310;
mov.u32 %r673, 31;
mov.u32 %r674, 16;
mov.u32 %r675, -1;
shfl.sync.bfly.b32 %r676|%p87, %r672, %r674, %r673, %r675;
mov.b32 %f311, %r676;
add.ftz.f32 %f312, %f310, %f311;
mov.b32 %r677, %f312;
mov.u32 %r678, 8;
shfl.sync.bfly.b32 %r679|%p88, %r677, %r678, %r673, %r675;
mov.b32 %f313, %r679;
add.ftz.f32 %f314, %f312, %f313;
mov.b32 %r680, %f314;
mov.u32 %r681, 4;
shfl.sync.bfly.b32 %r682|%p89, %r680, %r681, %r673, %r675;
mov.b32 %f315, %r682;
add.ftz.f32 %f316, %f314, %f315;
mov.b32 %r683, %f316;
mov.u32 %r684, 2;
shfl.sync.bfly.b32 %r685|%p90, %r683, %r684, %r673, %r675;
mov.b32 %f317, %r685;
add.ftz.f32 %f318, %f316, %f317;
mov.b32 %r686, %f318;
mov.u32 %r687, 1;
shfl.sync.bfly.b32 %r688|%p91, %r686, %r687, %r673, %r675;
mov.b32 %f319, %r688;
add.ftz.f32 %f61, %f318, %f319;
cvt.u64.u32 %rd113, %r671;
add.s64 %rd114, %rd113, %rd14;
shl.b64 %rd116, %rd114, 2;
add.s64 %rd18, %rd73, %rd116;
@%p18 bra $L__BB113_32;

st.global.f32 [%rd18], %f61;

$L__BB113_32:
ld.shared.f32 %f320, [%r30+1664];
add.ftz.f32 %f321, %f320, %f371;
mov.b32 %r689, %f321;
shfl.sync.bfly.b32 %r693|%p93, %r689, %r674, %r673, %r675;
mov.b32 %f322, %r693;
add.ftz.f32 %f323, %f321, %f322;
mov.b32 %r694, %f323;
shfl.sync.bfly.b32 %r696|%p94, %r694, %r678, %r673, %r675;
mov.b32 %f324, %r696;
add.ftz.f32 %f325, %f323, %f324;
mov.b32 %r697, %f325;
shfl.sync.bfly.b32 %r699|%p95, %r697, %r681, %r673, %r675;
mov.b32 %f326, %r699;
add.ftz.f32 %f327, %f325, %f326;
mov.b32 %r700, %f327;
shfl.sync.bfly.b32 %r702|%p96, %r700, %r684, %r673, %r675;
mov.b32 %f328, %r702;
add.ftz.f32 %f329, %f327, %f328;
mov.b32 %r703, %f329;
shfl.sync.bfly.b32 %r705|%p97, %r703, %r687, %r673, %r675;
mov.b32 %f330, %r705;
add.ftz.f32 %f62, %f329, %f330;
@%p26 bra $L__BB113_34;

st.global.f32 [%rd18+4], %f62;

$L__BB113_34:
mul.lo.s32 %r33, %r36, 7;
ld.shared.f32 %f331, [%r30+1792];
add.ftz.f32 %f332, %f331, %f370;
mov.b32 %r706, %f332;
mov.u32 %r707, 31;
mov.u32 %r708, 16;
mov.u32 %r709, -1;
shfl.sync.bfly.b32 %r710|%p99, %r706, %r708, %r707, %r709;
mov.b32 %f333, %r710;
add.ftz.f32 %f334, %f332, %f333;
mov.b32 %r711, %f334;
mov.u32 %r712, 8;
shfl.sync.bfly.b32 %r713|%p100, %r711, %r712, %r707, %r709;
mov.b32 %f335, %r713;
add.ftz.f32 %f336, %f334, %f335;
mov.b32 %r714, %f336;
mov.u32 %r715, 4;
shfl.sync.bfly.b32 %r716|%p101, %r714, %r715, %r707, %r709;
mov.b32 %f337, %r716;
add.ftz.f32 %f338, %f336, %f337;
mov.b32 %r717, %f338;
mov.u32 %r718, 2;
shfl.sync.bfly.b32 %r719|%p102, %r717, %r718, %r707, %r709;
mov.b32 %f339, %r719;
add.ftz.f32 %f340, %f338, %f339;
mov.b32 %r720, %f340;
mov.u32 %r721, 1;
shfl.sync.bfly.b32 %r722|%p103, %r720, %r721, %r707, %r709;
mov.b32 %f341, %r722;
add.ftz.f32 %f63, %f340, %f341;
@%p18 bra $L__BB113_36;

cvt.u64.u32 %rd117, %r33;
add.s64 %rd118, %rd117, %rd14;
shl.b64 %rd120, %rd118, 2;
add.s64 %rd121, %rd73, %rd120;
st.global.f32 [%rd121], %f63;

$L__BB113_36:
ld.shared.f32 %f342, [%r30+1920];
add.ftz.f32 %f343, %f342, %f369;
mov.b32 %r723, %f343;
shfl.sync.bfly.b32 %r727|%p105, %r723, %r708, %r707, %r709;
mov.b32 %f344, %r727;
add.ftz.f32 %f345, %f343, %f344;
mov.b32 %r728, %f345;
shfl.sync.bfly.b32 %r730|%p106, %r728, %r712, %r707, %r709;
mov.b32 %f346, %r730;
add.ftz.f32 %f347, %f345, %f346;
mov.b32 %r731, %f347;
shfl.sync.bfly.b32 %r733|%p107, %r731, %r715, %r707, %r709;
mov.b32 %f348, %r733;
add.ftz.f32 %f349, %f347, %f348;
mov.b32 %r734, %f349;
shfl.sync.bfly.b32 %r736|%p108, %r734, %r718, %r707, %r709;
mov.b32 %f350, %r736;
add.ftz.f32 %f351, %f349, %f350;
mov.b32 %r737, %f351;
shfl.sync.bfly.b32 %r739|%p109, %r737, %r721, %r707, %r709;
mov.b32 %f352, %r739;
add.ftz.f32 %f64, %f351, %f352;
@%p26 bra $L__BB113_38;

add.s32 %r740, %r33, 1;
cvt.u64.u32 %rd122, %r740;
add.s64 %rd123, %rd122, %rd14;
shl.b64 %rd125, %rd123, 2;
add.s64 %rd126, %rd73, %rd125;
st.global.f32 [%rd126], %f64;

$L__BB113_38:
ret;

}
.entry _Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj(
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2,
.param .align 8 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3[48],
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5,
.param .align 4 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6[12],
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15
)
.maxntid 256, 1, 1
.minnctapersm 1
{
.local .align 8 .b8 __local_depot114[16];
.reg .b64 %SP;
.reg .b64 %SPL;
.reg .pred %p<58>;
.reg .b16 %rs<37>;
.reg .f32 %f<213>;
.reg .b32 %r<579>;
.reg .b64 %rd<122>;


mov.u64 %SPL, __local_depot114;
ld.param.u64 %rd23, [_Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0];
ld.param.u64 %rd24, [_Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1];
ld.param.u64 %rd25, [_Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2];
ld.param.u32 %r22, [_Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5];
ld.param.u32 %r27, [_Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8];
ld.param.u32 %r28, [_Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9];
ld.param.u32 %r30, [_Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11];
ld.param.u32 %r31, [_Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12];
ld.param.u32 %r34, [_Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14];
ld.param.u32 %r33, [_Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15];
ld.param.u32 %r25, [_Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+8];
ld.param.u32 %r24, [_Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+4];
ld.param.u32 %r23, [_Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6];
ld.param.u64 %rd27, [_Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+8];
add.u64 %rd1, %SPL, 0;
add.u64 %rd2, %SPL, 8;
mov.u32 %r1, %tid.y;
setp.ge.u32 %p1, %r1, %r34;
@%p1 bra $L__BB114_33;

mov.u32 %r2, %ctaid.y;
cvta.to.global.u64 %rd34, %rd25;
mad.lo.s32 %r35, %r1, %r33, %r2;
mul.wide.u32 %rd35, %r35, 4;
add.s64 %rd36, %rd34, %rd35;
ld.global.nc.u32 %r3, [%rd36];
mov.f32 %f204, 0f00000000;
st.local.v2.f32 [%rd1], {%f204, %f204};
st.local.v2.f32 [%rd2], {%f204, %f204};
mov.u32 %r4, %tid.x;
shr.u32 %r5, %r4, 1;
shr.u32 %r36, %r22, 6;
setp.ge.u32 %p2, %r5, %r36;
mov.f32 %f205, %f204;
mov.f32 %f206, %f204;
mov.f32 %f207, %f204;
@%p2 bra $L__BB114_9;

cvta.to.global.u64 %rd37, %rd24;
and.b32 %r37, %r4, 1;
shl.b32 %r38, %r4, 2;
and.b32 %r39, %r38, 4;
or.b32 %r40, %r39, 2;
shr.u32 %r41, %r40, 1;
shr.u32 %r42, %r40, 2;
cvt.u64.u32 %rd38, %r41;
shl.b32 %r43, %r4, 4;
cvt.u64.u32 %rd39, %r43;
and.b64 %rd40, %rd39, 16;
mul.wide.u32 %rd41, %r42, 36;
mul.lo.s32 %r44, %r1, %r28;
mul.wide.u32 %rd42, %r44, 36;
add.s64 %rd43, %rd41, %rd42;
mul.hi.u32 %r45, %r2, %r23;
add.s32 %r46, %r45, %r2;
shr.u32 %r47, %r46, %r24;
mul.lo.s32 %r48, %r47, %r25;
sub.s32 %r49, %r2, %r48;
mul.lo.s32 %r50, %r49, %r31;
mul.wide.u32 %rd44, %r50, 36;
add.s64 %rd45, %rd43, %rd44;
shl.b32 %r51, %r5, 1;
mul.wide.s32 %rd46, %r51, 36;
add.s64 %rd47, %rd45, %rd46;
add.s64 %rd48, %rd37, %rd47;
add.s64 %rd119, %rd48, 32;
mov.u32 %r52, %ctaid.x;
shl.b32 %r53, %r52, 1;
or.b32 %r54, %r53, 1;
mad.lo.s32 %r55, %r3, %r30, %r5;
mad.lo.s32 %r56, %r27, %r54, %r55;
mul.wide.s32 %rd49, %r56, 36;
add.s64 %rd50, %rd40, %rd49;
add.s64 %rd5, %rd50, 16;
add.s64 %rd6, %rd38, %rd49;
add.s64 %rd51, %rd42, %rd44;
mul.wide.s32 %rd52, %r37, 36;
add.s64 %rd53, %rd51, %rd52;
add.s64 %rd54, %rd53, %rd46;
add.s64 %rd55, %rd37, %rd54;
add.s64 %rd118, %rd55, 8;
cvta.to.global.u64 %rd121, %rd23;
cvta.to.global.u64 %rd120, %rd27;
mov.f32 %f204, 0f00000000;
mov.u32 %r578, %r5;

$L__BB114_3:
mad.lo.s32 %r98, %r53, %r27, %r5;
mad.lo.s32 %r99, %r3, %r30, %r98;
mul.wide.s32 %rd58, %r99, 36;
add.s64 %rd59, %rd40, %rd58;
add.s64 %rd60, %rd121, %rd59;
ld.global.nc.u32 %r100, [%rd60+4];
and.b32 %r101, %r100, -2004318072;
shr.u32 %r102, %r101, 1;
or.b32 %r103, %r102, 839922192;
and.b32 %r104, %r100, 30583;
mov.u32 %r105, 201852420;
mov.u32 %r106, 50462976;
prmt.b32 %r107, %r106, %r105, %r104;
mov.u32 %r108, -185009412;
mov.u32 %r109, -33620224;
prmt.b32 %r110, %r109, %r108, %r104;
and.b32 %r111, %r103, 30292;
prmt.b32 %r112, %r107, %r110, %r111;
shr.u32 %r113, %r100, 16;
and.b32 %r114, %r113, 30583;
prmt.b32 %r115, %r106, %r105, %r114;
prmt.b32 %r116, %r109, %r108, %r114;
shr.u32 %r117, %r103, 16;
prmt.b32 %r118, %r115, %r116, %r117;
mov.u32 %r119, 25632;
prmt.b32 %r58, %r112, %r118, %r119;
mov.u32 %r120, 30001;
prmt.b32 %r62, %r112, %r118, %r120;
ld.global.nc.u32 %r121, [%rd60+8];
and.b32 %r122, %r121, -2004318072;
shr.u32 %r123, %r122, 1;
or.b32 %r124, %r123, 839922192;
and.b32 %r125, %r121, 30583;
prmt.b32 %r126, %r106, %r105, %r125;
prmt.b32 %r127, %r109, %r108, %r125;
and.b32 %r128, %r124, 30292;
prmt.b32 %r129, %r126, %r127, %r128;
shr.u32 %r130, %r121, 16;
and.b32 %r131, %r130, 30583;
prmt.b32 %r132, %r106, %r105, %r131;
prmt.b32 %r133, %r109, %r108, %r131;
shr.u32 %r134, %r124, 16;
prmt.b32 %r135, %r132, %r133, %r134;
prmt.b32 %r66, %r129, %r135, %r119;
prmt.b32 %r70, %r129, %r135, %r120;
ld.global.nc.u32 %r7, [%rd118+-4];
mov.u32 %r78, 0;

dp4a.s32.s32 %r57, %r58, %r7, %r78;

ld.global.nc.u32 %r8, [%rd118+4];

dp4a.s32.s32 %r61, %r62, %r8, %r57;

ld.global.nc.u32 %r9, [%rd118];

dp4a.s32.s32 %r65, %r66, %r9, %r61;

ld.global.nc.u32 %r10, [%rd118+8];

dp4a.s32.s32 %r69, %r70, %r10, %r65;

shl.b32 %r136, %r4, 1;
and.b32 %r137, %r136, 2;
or.b32 %r138, %r137, 1;
cvt.u64.u32 %rd61, %r138;
add.s64 %rd62, %rd61, %rd58;
add.s64 %rd63, %rd121, %rd62;
ld.global.nc.u8 %rs5, [%rd63+-1];
and.b16 %rs6, %rs5, 127;
setp.eq.s16 %p3, %rs6, 127;
selp.b16 %rs7, 0, %rs5, %p3;
and.b16 %rs1, %rs7, 255;

{cvt.rn.f16x2.e4m3x2 %r73, %rs1;}


cvt.u16.u32 %rs2, %r73;

{cvt.f32.f16 %f46, %rs2;}


mov.f32 %f50, 0f40000000;
div.approx.ftz.f32 %f51, %f46, %f50;
ld.global.nc.u32 %r11, [%rd118+-8];

{.reg .f16 low,high;
mov.b32 {low,high},%r11;
cvt.f32.f16 %f47, low;}


mul.ftz.f32 %f52, %f51, %f47;
cvt.rn.f32.s32 %f53, %r69;
fma.rn.ftz.f32 %f54, %f52, %f53, 0f00000000;
ld.global.nc.u32 %r139, [%rd60+12];
and.b32 %r140, %r139, -2004318072;
shr.u32 %r141, %r140, 1;
or.b32 %r142, %r141, 839922192;
and.b32 %r143, %r139, 30583;
prmt.b32 %r144, %r106, %r105, %r143;
prmt.b32 %r145, %r109, %r108, %r143;
and.b32 %r146, %r142, 30292;
prmt.b32 %r147, %r144, %r145, %r146;
shr.u32 %r148, %r139, 16;
and.b32 %r149, %r148, 30583;
prmt.b32 %r150, %r106, %r105, %r149;
prmt.b32 %r151, %r109, %r108, %r149;
shr.u32 %r152, %r142, 16;
prmt.b32 %r153, %r150, %r151, %r152;
prmt.b32 %r76, %r147, %r153, %r119;
prmt.b32 %r80, %r147, %r153, %r120;
ld.global.nc.u32 %r154, [%rd60+16];
and.b32 %r155, %r154, -2004318072;
shr.u32 %r156, %r155, 1;
or.b32 %r157, %r156, 839922192;
and.b32 %r158, %r154, 30583;
prmt.b32 %r159, %r106, %r105, %r158;
prmt.b32 %r160, %r109, %r108, %r158;
and.b32 %r161, %r157, 30292;
prmt.b32 %r162, %r159, %r160, %r161;
shr.u32 %r163, %r154, 16;
and.b32 %r164, %r163, 30583;
prmt.b32 %r165, %r106, %r105, %r164;
prmt.b32 %r166, %r109, %r108, %r164;
shr.u32 %r167, %r157, 16;
prmt.b32 %r168, %r165, %r166, %r167;
prmt.b32 %r84, %r162, %r168, %r119;
prmt.b32 %r88, %r162, %r168, %r120;
ld.global.nc.u32 %r12, [%rd119+-12];

dp4a.s32.s32 %r75, %r76, %r12, %r78;

ld.global.nc.u32 %r13, [%rd119+-4];

dp4a.s32.s32 %r79, %r80, %r13, %r75;

ld.global.nc.u32 %r14, [%rd119+-8];

dp4a.s32.s32 %r83, %r84, %r14, %r79;

ld.global.nc.u32 %r15, [%rd119];

dp4a.s32.s32 %r87, %r88, %r15, %r83;

ld.global.nc.u8 %rs8, [%rd63];
and.b16 %rs9, %rs8, 127;
setp.eq.s16 %p4, %rs9, 127;
selp.b16 %rs10, 0, %rs8, %p4;
and.b16 %rs3, %rs10, 255;

{cvt.rn.f16x2.e4m3x2 %r91, %rs3;}


cvt.u16.u32 %rs4, %r91;

{cvt.f32.f16 %f48, %rs4;}


div.approx.ftz.f32 %f55, %f48, %f50;
ld.global.nc.u32 %r16, [%rd119+-32];

{.reg .f16 low,high;
mov.b32 {low,high},%r16;
cvt.f32.f16 %f49, low;}


mul.ftz.f32 %f56, %f55, %f49;
cvt.rn.f32.s32 %f57, %r87;
fma.rn.ftz.f32 %f58, %f56, %f57, %f54;
add.ftz.f32 %f207, %f58, %f207;
setp.eq.s64 %p5, %rd27, 0;
@%p5 bra $L__BB114_5;

shr.u32 %r207, %r4, 1;
mad.lo.s32 %r210, %r53, %r27, %r207;
mad.lo.s32 %r211, %r3, %r30, %r210;
mul.wide.s32 %rd66, %r211, 36;
add.s64 %rd67, %rd40, %rd66;
add.s64 %rd68, %rd120, %rd67;
ld.global.u32 %r212, [%rd68+4];
and.b32 %r213, %r212, -2004318072;
shr.u32 %r214, %r213, 1;
or.b32 %r215, %r214, 839922192;
and.b32 %r216, %r212, 30583;
prmt.b32 %r219, %r106, %r105, %r216;
prmt.b32 %r222, %r109, %r108, %r216;
and.b32 %r223, %r215, 30292;
prmt.b32 %r224, %r219, %r222, %r223;
shr.u32 %r225, %r212, 16;
and.b32 %r226, %r225, 30583;
prmt.b32 %r227, %r106, %r105, %r226;
prmt.b32 %r228, %r109, %r108, %r226;
shr.u32 %r229, %r215, 16;
prmt.b32 %r230, %r227, %r228, %r229;
prmt.b32 %r170, %r224, %r230, %r119;
prmt.b32 %r174, %r224, %r230, %r120;
ld.global.u32 %r233, [%rd68+8];
and.b32 %r234, %r233, -2004318072;
shr.u32 %r235, %r234, 1;
or.b32 %r236, %r235, 839922192;
and.b32 %r237, %r233, 30583;
prmt.b32 %r238, %r106, %r105, %r237;
prmt.b32 %r239, %r109, %r108, %r237;
and.b32 %r240, %r236, 30292;
prmt.b32 %r241, %r238, %r239, %r240;
shr.u32 %r242, %r233, 16;
and.b32 %r243, %r242, 30583;
prmt.b32 %r244, %r106, %r105, %r243;
prmt.b32 %r245, %r109, %r108, %r243;
shr.u32 %r246, %r236, 16;
prmt.b32 %r247, %r244, %r245, %r246;
prmt.b32 %r178, %r241, %r247, %r119;
prmt.b32 %r182, %r241, %r247, %r120;

dp4a.s32.s32 %r169, %r170, %r7, %r78;


dp4a.s32.s32 %r173, %r174, %r8, %r169;


dp4a.s32.s32 %r177, %r178, %r9, %r173;


dp4a.s32.s32 %r181, %r182, %r10, %r177;

add.s64 %rd70, %rd61, %rd66;
add.s64 %rd71, %rd120, %rd70;
ld.global.u8 %rs15, [%rd71+-1];
and.b16 %rs16, %rs15, 127;
setp.eq.s16 %p6, %rs16, 127;
selp.b16 %rs11, 0, %rs15, %p6;

{cvt.rn.f16x2.e4m3x2 %r185, %rs11;}


cvt.u16.u32 %rs12, %r185;

{cvt.f32.f16 %f59, %rs12;}


div.approx.ftz.f32 %f64, %f59, %f50;

{.reg .f16 low,high;
mov.b32 {low,high},%r11;
cvt.f32.f16 %f60, low;}


mul.ftz.f32 %f65, %f64, %f60;
cvt.rn.f32.s32 %f66, %r181;
fma.rn.ftz.f32 %f67, %f65, %f66, 0f00000000;
ld.global.u32 %r251, [%rd68+12];
and.b32 %r252, %r251, -2004318072;
shr.u32 %r253, %r252, 1;
or.b32 %r254, %r253, 839922192;
and.b32 %r255, %r251, 30583;
prmt.b32 %r256, %r106, %r105, %r255;
prmt.b32 %r257, %r109, %r108, %r255;
and.b32 %r258, %r254, 30292;
prmt.b32 %r259, %r256, %r257, %r258;
shr.u32 %r260, %r251, 16;
and.b32 %r261, %r260, 30583;
prmt.b32 %r262, %r106, %r105, %r261;
prmt.b32 %r263, %r109, %r108, %r261;
shr.u32 %r264, %r254, 16;
prmt.b32 %r265, %r262, %r263, %r264;
prmt.b32 %r188, %r259, %r265, %r119;
prmt.b32 %r192, %r259, %r265, %r120;
ld.global.u32 %r266, [%rd68+16];
and.b32 %r267, %r266, -2004318072;
shr.u32 %r268, %r267, 1;
or.b32 %r269, %r268, 839922192;
and.b32 %r270, %r266, 30583;
prmt.b32 %r271, %r106, %r105, %r270;
prmt.b32 %r272, %r109, %r108, %r270;
and.b32 %r273, %r269, 30292;
prmt.b32 %r274, %r271, %r272, %r273;
shr.u32 %r275, %r266, 16;
and.b32 %r276, %r275, 30583;
prmt.b32 %r277, %r106, %r105, %r276;
prmt.b32 %r278, %r109, %r108, %r276;
shr.u32 %r279, %r269, 16;
prmt.b32 %r280, %r277, %r278, %r279;
prmt.b32 %r196, %r274, %r280, %r119;
prmt.b32 %r200, %r274, %r280, %r120;

dp4a.s32.s32 %r187, %r188, %r12, %r78;


dp4a.s32.s32 %r191, %r192, %r13, %r187;


dp4a.s32.s32 %r195, %r196, %r14, %r191;


dp4a.s32.s32 %r199, %r200, %r15, %r195;

ld.global.u8 %rs17, [%rd71];
and.b16 %rs18, %rs17, 127;
setp.eq.s16 %p7, %rs18, 127;
selp.b16 %rs13, 0, %rs17, %p7;

{cvt.rn.f16x2.e4m3x2 %r203, %rs13;}


cvt.u16.u32 %rs14, %r203;

{cvt.f32.f16 %f61, %rs14;}


div.approx.ftz.f32 %f68, %f61, %f50;

{.reg .f16 low,high;
mov.b32 {low,high},%r16;
cvt.f32.f16 %f62, low;}


mul.ftz.f32 %f69, %f68, %f62;
cvt.rn.f32.s32 %f70, %r199;
fma.rn.ftz.f32 %f71, %f69, %f70, %f67;
add.ftz.f32 %f205, %f71, %f205;

$L__BB114_5:
add.s64 %rd72, %rd121, %rd5;
ld.global.nc.u32 %r317, [%rd72+-12];
and.b32 %r318, %r317, -2004318072;
shr.u32 %r319, %r318, 1;
or.b32 %r320, %r319, 839922192;
and.b32 %r321, %r317, 30583;
prmt.b32 %r324, %r106, %r105, %r321;
prmt.b32 %r327, %r109, %r108, %r321;
and.b32 %r328, %r320, 30292;
prmt.b32 %r329, %r324, %r327, %r328;
shr.u32 %r330, %r317, 16;
and.b32 %r331, %r330, 30583;
prmt.b32 %r332, %r106, %r105, %r331;
prmt.b32 %r333, %r109, %r108, %r331;
shr.u32 %r334, %r320, 16;
prmt.b32 %r335, %r332, %r333, %r334;
prmt.b32 %r282, %r329, %r335, %r119;
prmt.b32 %r286, %r329, %r335, %r120;
ld.global.nc.u32 %r338, [%rd72+-8];
and.b32 %r339, %r338, -2004318072;
shr.u32 %r340, %r339, 1;
or.b32 %r341, %r340, 839922192;
and.b32 %r342, %r338, 30583;
prmt.b32 %r343, %r106, %r105, %r342;
prmt.b32 %r344, %r109, %r108, %r342;
and.b32 %r345, %r341, 30292;
prmt.b32 %r346, %r343, %r344, %r345;
shr.u32 %r347, %r338, 16;
and.b32 %r348, %r347, 30583;
prmt.b32 %r349, %r106, %r105, %r348;
prmt.b32 %r350, %r109, %r108, %r348;
shr.u32 %r351, %r341, 16;
prmt.b32 %r352, %r349, %r350, %r351;
prmt.b32 %r290, %r346, %r352, %r119;
prmt.b32 %r294, %r346, %r352, %r120;

dp4a.s32.s32 %r281, %r282, %r7, %r78;


dp4a.s32.s32 %r285, %r286, %r8, %r281;


dp4a.s32.s32 %r289, %r290, %r9, %r285;


dp4a.s32.s32 %r293, %r294, %r10, %r289;

add.s64 %rd73, %rd121, %rd6;
ld.global.nc.u8 %rs23, [%rd73+-1];
and.b16 %rs24, %rs23, 127;
setp.eq.s16 %p8, %rs24, 127;
selp.b16 %rs25, 0, %rs23, %p8;
and.b16 %rs19, %rs25, 255;

{cvt.rn.f16x2.e4m3x2 %r297, %rs19;}


cvt.u16.u32 %rs20, %r297;

{cvt.f32.f16 %f72, %rs20;}


div.approx.ftz.f32 %f77, %f72, %f50;

{.reg .f16 low,high;
mov.b32 {low,high},%r11;
cvt.f32.f16 %f73, low;}


mul.ftz.f32 %f78, %f77, %f73;
cvt.rn.f32.s32 %f79, %r293;
fma.rn.ftz.f32 %f80, %f78, %f79, 0f00000000;
ld.global.nc.u32 %r353, [%rd72+-4];
and.b32 %r354, %r353, -2004318072;
shr.u32 %r355, %r354, 1;
or.b32 %r356, %r355, 839922192;
and.b32 %r357, %r353, 30583;
prmt.b32 %r358, %r106, %r105, %r357;
prmt.b32 %r359, %r109, %r108, %r357;
and.b32 %r360, %r356, 30292;
prmt.b32 %r361, %r358, %r359, %r360;
shr.u32 %r362, %r353, 16;
and.b32 %r363, %r362, 30583;
prmt.b32 %r364, %r106, %r105, %r363;
prmt.b32 %r365, %r109, %r108, %r363;
shr.u32 %r366, %r356, 16;
prmt.b32 %r367, %r364, %r365, %r366;
prmt.b32 %r300, %r361, %r367, %r119;
prmt.b32 %r304, %r361, %r367, %r120;
ld.global.nc.u32 %r368, [%rd72];
and.b32 %r369, %r368, -2004318072;
shr.u32 %r370, %r369, 1;
or.b32 %r371, %r370, 839922192;
and.b32 %r372, %r368, 30583;
prmt.b32 %r373, %r106, %r105, %r372;
prmt.b32 %r374, %r109, %r108, %r372;
and.b32 %r375, %r371, 30292;
prmt.b32 %r376, %r373, %r374, %r375;
shr.u32 %r377, %r368, 16;
and.b32 %r378, %r377, 30583;
prmt.b32 %r379, %r106, %r105, %r378;
prmt.b32 %r380, %r109, %r108, %r378;
shr.u32 %r381, %r371, 16;
prmt.b32 %r382, %r379, %r380, %r381;
prmt.b32 %r308, %r376, %r382, %r119;
prmt.b32 %r312, %r376, %r382, %r120;

dp4a.s32.s32 %r299, %r300, %r12, %r78;


dp4a.s32.s32 %r303, %r304, %r13, %r299;


dp4a.s32.s32 %r307, %r308, %r14, %r303;


dp4a.s32.s32 %r311, %r312, %r15, %r307;

ld.global.nc.u8 %rs26, [%rd73];
and.b16 %rs27, %rs26, 127;
setp.eq.s16 %p9, %rs27, 127;
selp.b16 %rs28, 0, %rs26, %p9;
and.b16 %rs21, %rs28, 255;

{cvt.rn.f16x2.e4m3x2 %r315, %rs21;}


cvt.u16.u32 %rs22, %r315;

{cvt.f32.f16 %f74, %rs22;}


div.approx.ftz.f32 %f81, %f74, %f50;

{.reg .f16 low,high;
mov.b32 {low,high},%r16;
cvt.f32.f16 %f75, low;}


mul.ftz.f32 %f82, %f81, %f75;
cvt.rn.f32.s32 %f83, %r311;
fma.rn.ftz.f32 %f84, %f82, %f83, %f80;
add.ftz.f32 %f206, %f84, %f206;
@%p5 bra $L__BB114_7;

add.s64 %rd74, %rd120, %rd5;
ld.global.u32 %r419, [%rd74+-12];
and.b32 %r420, %r419, -2004318072;
shr.u32 %r421, %r420, 1;
or.b32 %r422, %r421, 839922192;
and.b32 %r423, %r419, 30583;
mov.u32 %r424, 201852420;
mov.u32 %r425, 50462976;
prmt.b32 %r426, %r425, %r424, %r423;
mov.u32 %r427, -185009412;
mov.u32 %r428, -33620224;
prmt.b32 %r429, %r428, %r427, %r423;
and.b32 %r430, %r422, 30292;
prmt.b32 %r431, %r426, %r429, %r430;
shr.u32 %r432, %r419, 16;
and.b32 %r433, %r432, 30583;
prmt.b32 %r434, %r425, %r424, %r433;
prmt.b32 %r435, %r428, %r427, %r433;
shr.u32 %r436, %r422, 16;
prmt.b32 %r437, %r434, %r435, %r436;
mov.u32 %r438, 25632;
prmt.b32 %r384, %r431, %r437, %r438;
mov.u32 %r439, 30001;
prmt.b32 %r388, %r431, %r437, %r439;
ld.global.u32 %r440, [%rd74+-8];
and.b32 %r441, %r440, -2004318072;
shr.u32 %r442, %r441, 1;
or.b32 %r443, %r442, 839922192;
and.b32 %r444, %r440, 30583;
prmt.b32 %r445, %r425, %r424, %r444;
prmt.b32 %r446, %r428, %r427, %r444;
and.b32 %r447, %r443, 30292;
prmt.b32 %r448, %r445, %r446, %r447;
shr.u32 %r449, %r440, 16;
and.b32 %r450, %r449, 30583;
prmt.b32 %r451, %r425, %r424, %r450;
prmt.b32 %r452, %r428, %r427, %r450;
shr.u32 %r453, %r443, 16;
prmt.b32 %r454, %r451, %r452, %r453;
prmt.b32 %r392, %r448, %r454, %r438;
prmt.b32 %r396, %r448, %r454, %r439;
mov.u32 %r404, 0;

dp4a.s32.s32 %r383, %r384, %r7, %r404;


dp4a.s32.s32 %r387, %r388, %r8, %r383;


dp4a.s32.s32 %r391, %r392, %r9, %r387;


dp4a.s32.s32 %r395, %r396, %r10, %r391;

add.s64 %rd75, %rd120, %rd6;
ld.global.u8 %rs33, [%rd75+-1];
and.b16 %rs34, %rs33, 127;
setp.eq.s16 %p11, %rs34, 127;
selp.b16 %rs29, 0, %rs33, %p11;

{cvt.rn.f16x2.e4m3x2 %r399, %rs29;}


cvt.u16.u32 %rs30, %r399;

{cvt.f32.f16 %f85, %rs30;}


mov.f32 %f89, 0f40000000;
div.approx.ftz.f32 %f90, %f85, %f89;

{.reg .f16 low,high;
mov.b32 {low,high},%r11;
cvt.f32.f16 %f86, low;}


mul.ftz.f32 %f91, %f90, %f86;
cvt.rn.f32.s32 %f92, %r395;
fma.rn.ftz.f32 %f93, %f91, %f92, 0f00000000;
ld.global.u32 %r455, [%rd74+-4];
and.b32 %r456, %r455, -2004318072;
shr.u32 %r457, %r456, 1;
or.b32 %r458, %r457, 839922192;
and.b32 %r459, %r455, 30583;
prmt.b32 %r460, %r425, %r424, %r459;
prmt.b32 %r461, %r428, %r427, %r459;
and.b32 %r462, %r458, 30292;
prmt.b32 %r463, %r460, %r461, %r462;
shr.u32 %r464, %r455, 16;
and.b32 %r465, %r464, 30583;
prmt.b32 %r466, %r425, %r424, %r465;
prmt.b32 %r467, %r428, %r427, %r465;
shr.u32 %r468, %r458, 16;
prmt.b32 %r469, %r466, %r467, %r468;
prmt.b32 %r402, %r463, %r469, %r438;
prmt.b32 %r406, %r463, %r469, %r439;
ld.global.u32 %r470, [%rd74];
and.b32 %r471, %r470, -2004318072;
shr.u32 %r472, %r471, 1;
or.b32 %r473, %r472, 839922192;
and.b32 %r474, %r470, 30583;
prmt.b32 %r475, %r425, %r424, %r474;
prmt.b32 %r476, %r428, %r427, %r474;
and.b32 %r477, %r473, 30292;
prmt.b32 %r478, %r475, %r476, %r477;
shr.u32 %r479, %r470, 16;
and.b32 %r480, %r479, 30583;
prmt.b32 %r481, %r425, %r424, %r480;
prmt.b32 %r482, %r428, %r427, %r480;
shr.u32 %r483, %r473, 16;
prmt.b32 %r484, %r481, %r482, %r483;
prmt.b32 %r410, %r478, %r484, %r438;
prmt.b32 %r414, %r478, %r484, %r439;

dp4a.s32.s32 %r401, %r402, %r12, %r404;


dp4a.s32.s32 %r405, %r406, %r13, %r401;


dp4a.s32.s32 %r409, %r410, %r14, %r405;


dp4a.s32.s32 %r413, %r414, %r15, %r409;

ld.global.u8 %rs35, [%rd75];
and.b16 %rs36, %rs35, 127;
setp.eq.s16 %p12, %rs36, 127;
selp.b16 %rs31, 0, %rs35, %p12;

{cvt.rn.f16x2.e4m3x2 %r417, %rs31;}


cvt.u16.u32 %rs32, %r417;

{cvt.f32.f16 %f87, %rs32;}


div.approx.ftz.f32 %f94, %f87, %f89;

{.reg .f16 low,high;
mov.b32 {low,high},%r16;
cvt.f32.f16 %f88, low;}


mul.ftz.f32 %f95, %f94, %f88;
cvt.rn.f32.s32 %f96, %r413;
fma.rn.ftz.f32 %f97, %f95, %f96, %f93;
add.ftz.f32 %f204, %f97, %f204;

$L__BB114_7:
add.s64 %rd121, %rd121, 576;
add.s64 %rd120, %rd120, 576;
add.s64 %rd119, %rd119, 1152;
add.s64 %rd118, %rd118, 1152;
add.s32 %r578, %r578, 16;
setp.lt.u32 %p13, %r578, %r36;
@%p13 bra $L__BB114_3;

add.u64 %rd110, %SPL, 0;
add.u64 %rd102, %SPL, 8;
st.local.v2.f32 [%rd110], {%f207, %f206};
st.local.v2.f32 [%rd102], {%f205, %f204};

$L__BB114_9:
mov.b32 %r18, %f207;
setp.ne.s64 %p14, %rd27, 0;
@%p14 bra $L__BB114_11;
bra.uni $L__BB114_10;

$L__BB114_11:
add.u64 %rd116, %SPL, 0;
add.u64 %rd108, %SPL, 8;
mov.u32 %r512, 31;
mov.u32 %r513, 16;
mov.u32 %r514, -1;
shfl.sync.bfly.b32 %r515|%p25, %r18, %r513, %r512, %r514;
mov.b32 %f118, %r515;
add.ftz.f32 %f119, %f207, %f118;
mov.b32 %r516, %f119;
mov.u32 %r517, 8;
shfl.sync.bfly.b32 %r518|%p26, %r516, %r517, %r512, %r514;
mov.b32 %f120, %r518;
add.ftz.f32 %f121, %f119, %f120;
mov.b32 %r519, %f121;
mov.u32 %r520, 4;
shfl.sync.bfly.b32 %r521|%p27, %r519, %r520, %r512, %r514;
mov.b32 %f122, %r521;
add.ftz.f32 %f123, %f121, %f122;
mov.b32 %r522, %f123;
mov.u32 %r523, 2;
shfl.sync.bfly.b32 %r524|%p28, %r522, %r523, %r512, %r514;
mov.b32 %f124, %r524;
add.ftz.f32 %f125, %f123, %f124;
mov.b32 %r525, %f125;
mov.u32 %r526, 1;
shfl.sync.bfly.b32 %r527|%p29, %r525, %r526, %r512, %r514;
mov.b32 %f126, %r527;
add.ftz.f32 %f127, %f125, %f126;
st.local.f32 [%rd116], %f127;
mov.b32 %r528, %f205;
shfl.sync.bfly.b32 %r529|%p30, %r528, %r513, %r512, %r514;
mov.b32 %f128, %r529;
add.ftz.f32 %f129, %f205, %f128;
mov.b32 %r530, %f129;
shfl.sync.bfly.b32 %r531|%p31, %r530, %r517, %r512, %r514;
mov.b32 %f130, %r531;
add.ftz.f32 %f131, %f129, %f130;
mov.b32 %r532, %f131;
shfl.sync.bfly.b32 %r533|%p32, %r532, %r520, %r512, %r514;
mov.b32 %f132, %r533;
add.ftz.f32 %f133, %f131, %f132;
mov.b32 %r534, %f133;
shfl.sync.bfly.b32 %r535|%p33, %r534, %r523, %r512, %r514;
mov.b32 %f134, %r535;
add.ftz.f32 %f135, %f133, %f134;
mov.b32 %r536, %f135;
shfl.sync.bfly.b32 %r537|%p34, %r536, %r526, %r512, %r514;
mov.b32 %f136, %r537;
add.ftz.f32 %f137, %f135, %f136;
st.local.f32 [%rd108], %f137;
mov.b32 %r538, %f206;
shfl.sync.bfly.b32 %r539|%p35, %r538, %r513, %r512, %r514;
mov.b32 %f138, %r539;
add.ftz.f32 %f139, %f206, %f138;
mov.b32 %r540, %f139;
shfl.sync.bfly.b32 %r541|%p36, %r540, %r517, %r512, %r514;
mov.b32 %f140, %r541;
add.ftz.f32 %f141, %f139, %f140;
mov.b32 %r542, %f141;
shfl.sync.bfly.b32 %r543|%p37, %r542, %r520, %r512, %r514;
mov.b32 %f142, %r543;
add.ftz.f32 %f143, %f141, %f142;
mov.b32 %r544, %f143;
shfl.sync.bfly.b32 %r545|%p38, %r544, %r523, %r512, %r514;
mov.b32 %f144, %r545;
add.ftz.f32 %f145, %f143, %f144;
mov.b32 %r546, %f145;
shfl.sync.bfly.b32 %r547|%p39, %r546, %r526, %r512, %r514;
mov.b32 %f146, %r547;
add.ftz.f32 %f147, %f145, %f146;
st.local.f32 [%rd116+4], %f147;
mov.b32 %r548, %f204;
shfl.sync.bfly.b32 %r549|%p40, %r548, %r513, %r512, %r514;
mov.b32 %f148, %r549;
add.ftz.f32 %f149, %f204, %f148;
mov.b32 %r550, %f149;
shfl.sync.bfly.b32 %r551|%p41, %r550, %r517, %r512, %r514;
mov.b32 %f150, %r551;
add.ftz.f32 %f151, %f149, %f150;
mov.b32 %r552, %f151;
shfl.sync.bfly.b32 %r553|%p42, %r552, %r520, %r512, %r514;
mov.b32 %f152, %r553;
add.ftz.f32 %f153, %f151, %f152;
mov.b32 %r554, %f153;
shfl.sync.bfly.b32 %r555|%p43, %r554, %r523, %r512, %r514;
mov.b32 %f154, %r555;
add.ftz.f32 %f155, %f153, %f154;
mov.b32 %r556, %f155;
shfl.sync.bfly.b32 %r557|%p44, %r556, %r526, %r512, %r514;
mov.b32 %f156, %r557;
add.ftz.f32 %f157, %f155, %f156;
st.local.f32 [%rd108+4], %f157;
bra.uni $L__BB114_12;

$L__BB114_10:
add.u64 %rd112, %SPL, 0;
mov.u32 %r486, 31;
mov.u32 %r487, 16;
mov.u32 %r488, -1;
shfl.sync.bfly.b32 %r489|%p15, %r18, %r487, %r486, %r488;
mov.b32 %f98, %r489;
add.ftz.f32 %f99, %f207, %f98;
mov.b32 %r490, %f99;
mov.u32 %r491, 8;
shfl.sync.bfly.b32 %r492|%p16, %r490, %r491, %r486, %r488;
mov.b32 %f100, %r492;
add.ftz.f32 %f101, %f99, %f100;
mov.b32 %r493, %f101;
mov.u32 %r494, 4;
shfl.sync.bfly.b32 %r495|%p17, %r493, %r494, %r486, %r488;
mov.b32 %f102, %r495;
add.ftz.f32 %f103, %f101, %f102;
mov.b32 %r496, %f103;
mov.u32 %r497, 2;
shfl.sync.bfly.b32 %r498|%p18, %r496, %r497, %r486, %r488;
mov.b32 %f104, %r498;
add.ftz.f32 %f105, %f103, %f104;
mov.b32 %r499, %f105;
mov.u32 %r500, 1;
shfl.sync.bfly.b32 %r501|%p19, %r499, %r500, %r486, %r488;
mov.b32 %f106, %r501;
add.ftz.f32 %f107, %f105, %f106;
st.local.f32 [%rd112], %f107;
mov.b32 %r502, %f206;
shfl.sync.bfly.b32 %r503|%p20, %r502, %r487, %r486, %r488;
mov.b32 %f108, %r503;
add.ftz.f32 %f109, %f206, %f108;
mov.b32 %r504, %f109;
shfl.sync.bfly.b32 %r505|%p21, %r504, %r491, %r486, %r488;
mov.b32 %f110, %r505;
add.ftz.f32 %f111, %f109, %f110;
mov.b32 %r506, %f111;
shfl.sync.bfly.b32 %r507|%p22, %r506, %r494, %r486, %r488;
mov.b32 %f112, %r507;
add.ftz.f32 %f113, %f111, %f112;
mov.b32 %r508, %f113;
shfl.sync.bfly.b32 %r509|%p23, %r508, %r497, %r486, %r488;
mov.b32 %f114, %r509;
add.ftz.f32 %f115, %f113, %f114;
mov.b32 %r510, %f115;
shfl.sync.bfly.b32 %r511|%p24, %r510, %r500, %r486, %r488;
mov.b32 %f116, %r511;
add.ftz.f32 %f117, %f115, %f116;
st.local.f32 [%rd112+4], %f117;

$L__BB114_12:
ld.param.u32 %r567, [_Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7];
setp.gt.u32 %p45, %r4, 1;
mov.u32 %r559, %ctaid.x;
shl.b32 %r560, %r559, 1;
add.s32 %r19, %r560, %r4;
setp.ge.u32 %p46, %r19, %r567;
or.pred %p47, %p45, %p46;
@%p47 bra $L__BB114_33;

add.u64 %rd114, %SPL, 0;
ld.param.u64 %rd95, [_Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+24];
ld.param.u32 %r568, [_Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13];
cvt.u64.u32 %rd19, %r4;
mul.wide.u32 %rd76, %r4, 4;
add.s64 %rd77, %rd114, %rd76;
ld.local.f32 %f23, [%rd77];
mad.lo.s32 %r20, %r3, %r568, %r19;
setp.eq.s64 %p48, %rd95, 0;
@%p48 bra $L__BB114_15;

ld.param.u64 %rd96, [_Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+24];
cvta.to.global.u64 %rd78, %rd96;
mul.wide.u32 %rd79, %r3, 4;
add.s64 %rd80, %rd78, %rd79;
ld.global.f32 %f158, [%rd80];
mul.ftz.f32 %f23, %f23, %f158;

$L__BB114_15:
ld.param.u64 %rd97, [_Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3];
setp.eq.s64 %p49, %rd97, 0;
@%p49 bra $L__BB114_17;

ld.param.u64 %rd98, [_Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3];
cvta.to.global.u64 %rd81, %rd98;
mul.wide.u32 %rd82, %r20, 4;
add.s64 %rd83, %rd81, %rd82;
ld.global.f32 %f159, [%rd83];
add.ftz.f32 %f23, %f23, %f159;

$L__BB114_17:
setp.eq.s64 %p50, %rd27, 0;
@%p50 bra $L__BB114_32;

ld.param.u64 %rd106, [_Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+32];
add.u64 %rd104, %SPL, 8;
shl.b64 %rd84, %rd19, 2;
add.s64 %rd85, %rd104, %rd84;
ld.local.f32 %f28, [%rd85];
setp.eq.s64 %p51, %rd106, 0;
@%p51 bra $L__BB114_20;

ld.param.u64 %rd107, [_Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+32];
cvta.to.global.u64 %rd86, %rd107;
mul.wide.u32 %rd87, %r3, 4;
add.s64 %rd88, %rd86, %rd87;
ld.global.f32 %f160, [%rd88];
mul.ftz.f32 %f28, %f28, %f160;

$L__BB114_20:
ld.param.u64 %rd100, [_Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+16];
setp.eq.s64 %p52, %rd100, 0;
@%p52 bra $L__BB114_22;

ld.param.u64 %rd101, [_Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+16];
cvta.to.global.u64 %rd89, %rd101;
mul.wide.u32 %rd90, %r20, 4;
add.s64 %rd91, %rd89, %rd90;
ld.global.f32 %f161, [%rd91];
add.ftz.f32 %f28, %f28, %f161;

$L__BB114_22:
ld.param.u32 %r573, [_Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.gt.s32 %p53, %r573, 2;
@%p53 bra $L__BB114_26;

ld.param.u32 %r576, [_Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.eq.s32 %p56, %r576, 1;
@%p56 bra $L__BB114_30;

ld.param.u32 %r577, [_Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.eq.s32 %p57, %r577, 2;
@%p57 bra $L__BB114_25;
bra.uni $L__BB114_31;

$L__BB114_25:
mul.ftz.f32 %f181, %f28, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f182, %f181;
add.ftz.f32 %f183, %f182, 0f3F800000;
div.approx.ftz.f32 %f184, %f28, %f183;
mul.ftz.f32 %f23, %f23, %f184;
bra.uni $L__BB114_32;

$L__BB114_26:
ld.param.u32 %r574, [_Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.eq.s32 %p54, %r574, 3;
@%p54 bra $L__BB114_29;

ld.param.u32 %r575, [_Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.ne.s32 %p55, %r575, 6;
@%p55 bra $L__BB114_31;

ld.param.f32 %f193, [_Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+44];
min.ftz.f32 %f162, %f28, %f193;
neg.ftz.f32 %f163, %f193;
min.ftz.f32 %f164, %f23, %f193;
max.ftz.f32 %f165, %f164, %f163;
mul.ftz.f32 %f166, %f162, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f167, %f166;
add.ftz.f32 %f168, %f167, 0f3F800000;
div.approx.ftz.f32 %f169, %f162, %f168;
mul.ftz.f32 %f23, %f165, %f169;
bra.uni $L__BB114_32;

$L__BB114_31:
mul.ftz.f32 %f23, %f23, %f28;
bra.uni $L__BB114_32;

$L__BB114_30:
mul.ftz.f32 %f185, %f28, 0f3F000000;
mul.ftz.f32 %f186, %f28, 0f3D372713;
fma.rn.ftz.f32 %f187, %f28, %f186, 0f3F800000;
mul.ftz.f32 %f188, %f28, 0f3F4C422A;
mul.ftz.f32 %f189, %f188, %f187;
tanh.approx.f32 %f190, %f189;
add.ftz.f32 %f191, %f190, 0f3F800000;
mul.ftz.f32 %f192, %f185, %f191;
mul.ftz.f32 %f23, %f23, %f192;
bra.uni $L__BB114_32;

$L__BB114_29:
mov.f32 %f170, 0f40E00000;
min.ftz.f32 %f171, %f28, %f170;
min.ftz.f32 %f172, %f23, %f170;
mov.f32 %f173, 0fC0E00000;
max.ftz.f32 %f174, %f172, %f173;
mul.ftz.f32 %f175, %f171, 0f3FD9DB23;
mul.ftz.f32 %f176, %f175, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f177, %f176;
add.ftz.f32 %f178, %f177, 0f3F800000;
div.approx.ftz.f32 %f179, %f171, %f178;
add.ftz.f32 %f180, %f174, 0f3F800000;
mul.ftz.f32 %f23, %f180, %f179;

$L__BB114_32:
ld.param.u64 %rd99, [_Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4];
ld.param.u32 %r572, [_Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10];
ld.param.u32 %r571, [_Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13];
mov.u32 %r570, %ctaid.y;
mov.u32 %r569, %tid.y;
mad.lo.s32 %r565, %r569, %r572, %r19;
mad.lo.s32 %r566, %r570, %r571, %r565;
cvta.to.global.u64 %rd92, %rd99;
mul.wide.u32 %rd93, %r566, 4;
add.s64 %rd94, %rd92, %rd93;
st.global.f32 [%rd94], %f23;

$L__BB114_33:
ret;

}
.entry _Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj(
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2,
.param .align 8 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3[48],
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5,
.param .align 4 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6[12],
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15
)
.maxntid 256, 1, 1
.minnctapersm 1
{
.local .align 8 .b8 __local_depot115[8];
.reg .b64 %SP;
.reg .b64 %SPL;
.reg .pred %p<21>;
.reg .b16 %rs<21>;
.reg .f32 %f<61>;
.reg .b32 %r<277>;
.reg .b64 %rd<60>;


mov.u64 %SPL, __local_depot115;
ld.param.u64 %rd16, [_Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0];
ld.param.u64 %rd17, [_Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1];
ld.param.u64 %rd18, [_Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2];
ld.param.u64 %rd19, [_Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4];
ld.param.u32 %r20, [_Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5];
ld.param.u32 %r12, [_Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7];
ld.param.u32 %r13, [_Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8];
ld.param.u32 %r14, [_Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9];
ld.param.u32 %r15, [_Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10];
ld.param.u32 %r16, [_Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11];
ld.param.u32 %r17, [_Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12];
ld.param.u32 %r18, [_Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13];
ld.param.u32 %r21, [_Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14];
ld.param.u32 %r19, [_Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15];
ld.param.u32 %r11, [_Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+8];
ld.param.u32 %r10, [_Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+4];
ld.param.u32 %r9, [_Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6];
add.u64 %rd1, %SPL, 0;
mov.u32 %r22, %ctaid.x;
shl.b32 %r1, %r22, 1;
shr.u32 %r2, %r20, 6;
mov.u32 %r3, %tid.y;
setp.ge.u32 %p1, %r3, %r21;
@%p1 bra $L__BB115_7;

mov.f32 %f59, 0f00000000;
st.local.v2.f32 [%rd1], {%f59, %f59};
mov.u32 %r4, %tid.x;
shr.u32 %r276, %r4, 1;
setp.ge.u32 %p2, %r276, %r2;
mov.f32 %f60, %f59;
@%p2 bra $L__BB115_5;

mov.u32 %r23, %ctaid.y;
cvta.to.global.u64 %rd21, %rd17;
mad.lo.s32 %r24, %r3, %r19, %r23;
cvta.to.global.u64 %rd22, %rd18;
mul.wide.u32 %rd23, %r24, 4;
add.s64 %rd24, %rd22, %rd23;
ld.global.nc.u32 %r25, [%rd24];
mul.lo.s32 %r26, %r25, %r16;
and.b32 %r27, %r4, 1;
shl.b32 %r28, %r4, 2;
and.b32 %r29, %r28, 4;
or.b32 %r30, %r29, 2;
shr.u32 %r31, %r30, 1;
shr.u32 %r32, %r30, 2;
cvt.u64.u32 %rd25, %r31;
shl.b32 %r33, %r4, 4;
cvt.u64.u32 %rd26, %r33;
and.b64 %rd27, %rd26, 16;
mad.lo.s32 %r34, %r1, %r13, %r276;
add.s32 %r35, %r34, %r26;
mul.wide.s32 %rd28, %r35, 36;
add.s64 %rd29, %rd27, %rd28;
add.s64 %rd3, %rd29, 16;
add.s64 %rd30, %rd25, -1;
add.s64 %rd31, %rd30, %rd28;
add.s64 %rd4, %rd31, 1;
add.s32 %r36, %r1, 1;
add.s32 %r37, %r276, %r26;
mad.lo.s32 %r38, %r13, %r36, %r37;
mul.wide.s32 %rd32, %r38, 36;
add.s64 %rd33, %rd27, %rd32;
add.s64 %rd5, %rd33, 16;
add.s64 %rd34, %rd30, %rd32;
add.s64 %rd6, %rd34, 1;
mul.wide.u32 %rd35, %r32, 36;
mul.lo.s32 %r39, %r3, %r14;
mul.wide.u32 %rd36, %r39, 36;
add.s64 %rd37, %rd35, %rd36;
mul.hi.u32 %r40, %r23, %r9;
add.s32 %r41, %r40, %r23;
shr.u32 %r42, %r41, %r10;
mul.lo.s32 %r43, %r42, %r11;
sub.s32 %r44, %r23, %r43;
mul.lo.s32 %r45, %r44, %r17;
mul.wide.u32 %rd38, %r45, 36;
add.s64 %rd39, %rd37, %rd38;
shl.b32 %r46, %r276, 1;
mul.wide.s32 %rd40, %r46, 36;
add.s64 %rd41, %rd39, %rd40;
add.s64 %rd42, %rd21, %rd41;
add.s64 %rd58, %rd42, 32;
add.s64 %rd43, %rd36, %rd38;
mul.wide.s32 %rd44, %r27, 36;
add.s64 %rd45, %rd43, %rd44;
add.s64 %rd46, %rd45, %rd40;
add.s64 %rd47, %rd21, %rd46;
add.s64 %rd57, %rd47, 8;
cvta.to.global.u64 %rd59, %rd16;
mov.f32 %f59, 0f00000000;

$L__BB115_3:
add.s64 %rd48, %rd59, %rd3;
ld.global.nc.u32 %r119, [%rd48+-12];
and.b32 %r120, %r119, -2004318072;
shr.u32 %r121, %r120, 1;
or.b32 %r122, %r121, 839922192;
and.b32 %r123, %r119, 30583;
mov.u32 %r124, 201852420;
mov.u32 %r125, 50462976;
prmt.b32 %r126, %r125, %r124, %r123;
mov.u32 %r127, -185009412;
mov.u32 %r128, -33620224;
prmt.b32 %r129, %r128, %r127, %r123;
and.b32 %r130, %r122, 30292;
prmt.b32 %r131, %r126, %r129, %r130;
shr.u32 %r132, %r119, 16;
and.b32 %r133, %r132, 30583;
prmt.b32 %r134, %r125, %r124, %r133;
prmt.b32 %r135, %r128, %r127, %r133;
shr.u32 %r136, %r122, 16;
prmt.b32 %r137, %r134, %r135, %r136;
mov.u32 %r138, 25632;
prmt.b32 %r48, %r131, %r137, %r138;
mov.u32 %r139, 30001;
prmt.b32 %r52, %r131, %r137, %r139;
ld.global.nc.u32 %r140, [%rd48+-8];
and.b32 %r141, %r140, -2004318072;
shr.u32 %r142, %r141, 1;
or.b32 %r143, %r142, 839922192;
and.b32 %r144, %r140, 30583;
prmt.b32 %r145, %r125, %r124, %r144;
prmt.b32 %r146, %r128, %r127, %r144;
and.b32 %r147, %r143, 30292;
prmt.b32 %r148, %r145, %r146, %r147;
shr.u32 %r149, %r140, 16;
and.b32 %r150, %r149, 30583;
prmt.b32 %r151, %r125, %r124, %r150;
prmt.b32 %r152, %r128, %r127, %r150;
shr.u32 %r153, %r143, 16;
prmt.b32 %r154, %r151, %r152, %r153;
prmt.b32 %r56, %r148, %r154, %r138;
prmt.b32 %r60, %r148, %r154, %r139;
ld.global.nc.u32 %r49, [%rd57+-4];
mov.u32 %r104, 0;

dp4a.s32.s32 %r47, %r48, %r49, %r104;

ld.global.nc.u32 %r53, [%rd57+4];

dp4a.s32.s32 %r51, %r52, %r53, %r47;

ld.global.nc.u32 %r57, [%rd57];

dp4a.s32.s32 %r55, %r56, %r57, %r51;

ld.global.nc.u32 %r61, [%rd57+8];

dp4a.s32.s32 %r59, %r60, %r61, %r55;

add.s64 %rd49, %rd59, %rd4;
ld.global.nc.u8 %rs9, [%rd49+-1];
and.b16 %rs10, %rs9, 127;
setp.eq.s16 %p3, %rs10, 127;
selp.b16 %rs11, 0, %rs9, %p3;
and.b16 %rs1, %rs11, 255;

{cvt.rn.f16x2.e4m3x2 %r63, %rs1;}


cvt.u16.u32 %rs2, %r63;

{cvt.f32.f16 %f11, %rs2;}


mov.f32 %f19, 0f40000000;
div.approx.ftz.f32 %f20, %f11, %f19;
ld.global.nc.u32 %r64, [%rd57+-8];

{.reg .f16 low,high;
mov.b32 {low,high},%r64;
cvt.f32.f16 %f12, low;}


mul.ftz.f32 %f21, %f20, %f12;
cvt.rn.f32.s32 %f22, %r59;
fma.rn.ftz.f32 %f23, %f21, %f22, 0f00000000;
ld.global.nc.u32 %r155, [%rd48+-4];
and.b32 %r156, %r155, -2004318072;
shr.u32 %r157, %r156, 1;
or.b32 %r158, %r157, 839922192;
and.b32 %r159, %r155, 30583;
prmt.b32 %r160, %r125, %r124, %r159;
prmt.b32 %r161, %r128, %r127, %r159;
and.b32 %r162, %r158, 30292;
prmt.b32 %r163, %r160, %r161, %r162;
shr.u32 %r164, %r155, 16;
and.b32 %r165, %r164, 30583;
prmt.b32 %r166, %r125, %r124, %r165;
prmt.b32 %r167, %r128, %r127, %r165;
shr.u32 %r168, %r158, 16;
prmt.b32 %r169, %r166, %r167, %r168;
prmt.b32 %r66, %r163, %r169, %r138;
prmt.b32 %r70, %r163, %r169, %r139;
ld.global.nc.u32 %r170, [%rd48];
and.b32 %r171, %r170, -2004318072;
shr.u32 %r172, %r171, 1;
or.b32 %r173, %r172, 839922192;
and.b32 %r174, %r170, 30583;
prmt.b32 %r175, %r125, %r124, %r174;
prmt.b32 %r176, %r128, %r127, %r174;
and.b32 %r177, %r173, 30292;
prmt.b32 %r178, %r175, %r176, %r177;
shr.u32 %r179, %r170, 16;
and.b32 %r180, %r179, 30583;
prmt.b32 %r181, %r125, %r124, %r180;
prmt.b32 %r182, %r128, %r127, %r180;
shr.u32 %r183, %r173, 16;
prmt.b32 %r184, %r181, %r182, %r183;
prmt.b32 %r74, %r178, %r184, %r138;
prmt.b32 %r78, %r178, %r184, %r139;
ld.global.nc.u32 %r67, [%rd58+-12];

dp4a.s32.s32 %r65, %r66, %r67, %r104;

ld.global.nc.u32 %r71, [%rd58+-4];

dp4a.s32.s32 %r69, %r70, %r71, %r65;

ld.global.nc.u32 %r75, [%rd58+-8];

dp4a.s32.s32 %r73, %r74, %r75, %r69;

ld.global.nc.u32 %r79, [%rd58];

dp4a.s32.s32 %r77, %r78, %r79, %r73;

ld.global.nc.u8 %rs12, [%rd49];
and.b16 %rs13, %rs12, 127;
setp.eq.s16 %p4, %rs13, 127;
selp.b16 %rs14, 0, %rs12, %p4;
and.b16 %rs3, %rs14, 255;

{cvt.rn.f16x2.e4m3x2 %r81, %rs3;}


cvt.u16.u32 %rs4, %r81;

{cvt.f32.f16 %f13, %rs4;}


div.approx.ftz.f32 %f24, %f13, %f19;
ld.global.nc.u32 %r82, [%rd58+-32];

{.reg .f16 low,high;
mov.b32 {low,high},%r82;
cvt.f32.f16 %f14, low;}


mul.ftz.f32 %f25, %f24, %f14;
cvt.rn.f32.s32 %f26, %r77;
fma.rn.ftz.f32 %f27, %f25, %f26, %f23;
add.ftz.f32 %f60, %f27, %f60;
add.s64 %rd50, %rd59, %rd5;
ld.global.nc.u32 %r185, [%rd50+-12];
and.b32 %r186, %r185, -2004318072;
shr.u32 %r187, %r186, 1;
or.b32 %r188, %r187, 839922192;
and.b32 %r189, %r185, 30583;
prmt.b32 %r190, %r125, %r124, %r189;
prmt.b32 %r191, %r128, %r127, %r189;
and.b32 %r192, %r188, 30292;
prmt.b32 %r193, %r190, %r191, %r192;
shr.u32 %r194, %r185, 16;
and.b32 %r195, %r194, 30583;
prmt.b32 %r196, %r125, %r124, %r195;
prmt.b32 %r197, %r128, %r127, %r195;
shr.u32 %r198, %r188, 16;
prmt.b32 %r199, %r196, %r197, %r198;
prmt.b32 %r84, %r193, %r199, %r138;
prmt.b32 %r88, %r193, %r199, %r139;
ld.global.nc.u32 %r200, [%rd50+-8];
and.b32 %r201, %r200, -2004318072;
shr.u32 %r202, %r201, 1;
or.b32 %r203, %r202, 839922192;
and.b32 %r204, %r200, 30583;
prmt.b32 %r205, %r125, %r124, %r204;
prmt.b32 %r206, %r128, %r127, %r204;
and.b32 %r207, %r203, 30292;
prmt.b32 %r208, %r205, %r206, %r207;
shr.u32 %r209, %r200, 16;
and.b32 %r210, %r209, 30583;
prmt.b32 %r211, %r125, %r124, %r210;
prmt.b32 %r212, %r128, %r127, %r210;
shr.u32 %r213, %r203, 16;
prmt.b32 %r214, %r211, %r212, %r213;
prmt.b32 %r92, %r208, %r214, %r138;
prmt.b32 %r96, %r208, %r214, %r139;

dp4a.s32.s32 %r83, %r84, %r49, %r104;


dp4a.s32.s32 %r87, %r88, %r53, %r83;


dp4a.s32.s32 %r91, %r92, %r57, %r87;


dp4a.s32.s32 %r95, %r96, %r61, %r91;

add.s64 %rd51, %rd59, %rd6;
ld.global.nc.u8 %rs15, [%rd51+-1];
and.b16 %rs16, %rs15, 127;
setp.eq.s16 %p5, %rs16, 127;
selp.b16 %rs17, 0, %rs15, %p5;
and.b16 %rs5, %rs17, 255;

{cvt.rn.f16x2.e4m3x2 %r99, %rs5;}


cvt.u16.u32 %rs6, %r99;

{cvt.f32.f16 %f15, %rs6;}


div.approx.ftz.f32 %f28, %f15, %f19;

{.reg .f16 low,high;
mov.b32 {low,high},%r64;
cvt.f32.f16 %f16, low;}


mul.ftz.f32 %f29, %f28, %f16;
cvt.rn.f32.s32 %f30, %r95;
fma.rn.ftz.f32 %f31, %f29, %f30, 0f00000000;
ld.global.nc.u32 %r215, [%rd50+-4];
and.b32 %r216, %r215, -2004318072;
shr.u32 %r217, %r216, 1;
or.b32 %r218, %r217, 839922192;
and.b32 %r219, %r215, 30583;
prmt.b32 %r220, %r125, %r124, %r219;
prmt.b32 %r221, %r128, %r127, %r219;
and.b32 %r222, %r218, 30292;
prmt.b32 %r223, %r220, %r221, %r222;
shr.u32 %r224, %r215, 16;
and.b32 %r225, %r224, 30583;
prmt.b32 %r226, %r125, %r124, %r225;
prmt.b32 %r227, %r128, %r127, %r225;
shr.u32 %r228, %r218, 16;
prmt.b32 %r229, %r226, %r227, %r228;
prmt.b32 %r102, %r223, %r229, %r138;
prmt.b32 %r106, %r223, %r229, %r139;
ld.global.nc.u32 %r230, [%rd50];
and.b32 %r231, %r230, -2004318072;
shr.u32 %r232, %r231, 1;
or.b32 %r233, %r232, 839922192;
and.b32 %r234, %r230, 30583;
prmt.b32 %r235, %r125, %r124, %r234;
prmt.b32 %r236, %r128, %r127, %r234;
and.b32 %r237, %r233, 30292;
prmt.b32 %r238, %r235, %r236, %r237;
shr.u32 %r239, %r230, 16;
and.b32 %r240, %r239, 30583;
prmt.b32 %r241, %r125, %r124, %r240;
prmt.b32 %r242, %r128, %r127, %r240;
shr.u32 %r243, %r233, 16;
prmt.b32 %r244, %r241, %r242, %r243;
prmt.b32 %r110, %r238, %r244, %r138;
prmt.b32 %r114, %r238, %r244, %r139;

dp4a.s32.s32 %r101, %r102, %r67, %r104;


dp4a.s32.s32 %r105, %r106, %r71, %r101;


dp4a.s32.s32 %r109, %r110, %r75, %r105;


dp4a.s32.s32 %r113, %r114, %r79, %r109;

ld.global.nc.u8 %rs18, [%rd51];
and.b16 %rs19, %rs18, 127;
setp.eq.s16 %p6, %rs19, 127;
selp.b16 %rs20, 0, %rs18, %p6;
and.b16 %rs7, %rs20, 255;

{cvt.rn.f16x2.e4m3x2 %r117, %rs7;}


cvt.u16.u32 %rs8, %r117;

{cvt.f32.f16 %f17, %rs8;}


div.approx.ftz.f32 %f32, %f17, %f19;

{.reg .f16 low,high;
mov.b32 {low,high},%r82;
cvt.f32.f16 %f18, low;}


mul.ftz.f32 %f33, %f32, %f18;
cvt.rn.f32.s32 %f34, %r113;
fma.rn.ftz.f32 %f35, %f33, %f34, %f31;
add.ftz.f32 %f59, %f35, %f59;
add.s64 %rd59, %rd59, 576;
add.s64 %rd58, %rd58, 1152;
add.s64 %rd57, %rd57, 1152;
add.s32 %r276, %r276, 16;
setp.lt.u32 %p7, %r276, %r2;
@%p7 bra $L__BB115_3;

st.local.v2.f32 [%rd1], {%f60, %f59};

$L__BB115_5:
mov.b32 %r245, %f60;
mov.u32 %r246, 31;
mov.u32 %r247, 16;
mov.u32 %r248, -1;
shfl.sync.bfly.b32 %r249|%p8, %r245, %r247, %r246, %r248;
mov.b32 %f36, %r249;
add.ftz.f32 %f37, %f60, %f36;
mov.b32 %r250, %f37;
mov.u32 %r251, 8;
shfl.sync.bfly.b32 %r252|%p9, %r250, %r251, %r246, %r248;
mov.b32 %f38, %r252;
add.ftz.f32 %f39, %f37, %f38;
mov.b32 %r253, %f39;
mov.u32 %r254, 4;
shfl.sync.bfly.b32 %r255|%p10, %r253, %r254, %r246, %r248;
mov.b32 %f40, %r255;
add.ftz.f32 %f41, %f39, %f40;
mov.b32 %r256, %f41;
mov.u32 %r257, 2;
shfl.sync.bfly.b32 %r258|%p11, %r256, %r257, %r246, %r248;
mov.b32 %f42, %r258;
add.ftz.f32 %f43, %f41, %f42;
mov.b32 %r259, %f43;
mov.u32 %r260, 1;
shfl.sync.bfly.b32 %r261|%p12, %r259, %r260, %r246, %r248;
mov.b32 %f44, %r261;
add.ftz.f32 %f45, %f43, %f44;
st.local.f32 [%rd1], %f45;
mov.b32 %r262, %f59;
shfl.sync.bfly.b32 %r263|%p13, %r262, %r247, %r246, %r248;
mov.b32 %f46, %r263;
add.ftz.f32 %f47, %f59, %f46;
mov.b32 %r264, %f47;
shfl.sync.bfly.b32 %r265|%p14, %r264, %r251, %r246, %r248;
mov.b32 %f48, %r265;
add.ftz.f32 %f49, %f47, %f48;
mov.b32 %r266, %f49;
shfl.sync.bfly.b32 %r267|%p15, %r266, %r254, %r246, %r248;
mov.b32 %f50, %r267;
add.ftz.f32 %f51, %f49, %f50;
mov.b32 %r268, %f51;
shfl.sync.bfly.b32 %r269|%p16, %r268, %r257, %r246, %r248;
mov.b32 %f52, %r269;
add.ftz.f32 %f53, %f51, %f52;
mov.b32 %r270, %f53;
shfl.sync.bfly.b32 %r271|%p17, %r270, %r260, %r246, %r248;
mov.b32 %f54, %r271;
add.ftz.f32 %f55, %f53, %f54;
st.local.f32 [%rd1+4], %f55;
add.s32 %r8, %r1, %r4;
setp.ge.u32 %p18, %r8, %r12;
setp.gt.u32 %p19, %r4, 1;
or.pred %p20, %p19, %p18;
@%p20 bra $L__BB115_7;

mul.wide.u32 %rd52, %r4, 4;
add.s64 %rd53, %rd1, %rd52;
ld.local.f32 %f56, [%rd53];
mov.u32 %r272, %ctaid.y;
mad.lo.s32 %r274, %r3, %r15, %r8;
mad.lo.s32 %r275, %r272, %r18, %r274;
cvta.to.global.u64 %rd54, %rd19;
mul.wide.u32 %rd55, %r275, 4;
add.s64 %rd56, %rd54, %rd55;
st.global.f32 [%rd56], %f56;

$L__BB115_7:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<343>;
.reg .b16 %rs<73>;
.reg .f32 %f<1091>;
.reg .b32 %r<1589>;
.reg .b64 %rd<125>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate[1536];

ld.param.u64 %rd38, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd39, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd40, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u32 %r65, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r51, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r52, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r56, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r57, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r58, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r62, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r63, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r64, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r60, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r59, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r54, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r53, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r50, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r49, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r48, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
ld.param.u64 %rd45, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+32];
ld.param.u64 %rd44, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+24];
ld.param.u64 %rd43, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+16];
ld.param.u64 %rd42, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+8];
ld.param.u64 %rd41, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3];
mov.u32 %r1, %tid.y;
shl.b32 %r66, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r66, %r2;
mov.u32 %r67, %ctaid.x;
shl.b32 %r4, %r67, 2;
shr.u32 %r5, %r65, 6;
mov.u32 %r6, %ctaid.y;
setp.eq.s64 %p2, %rd40, 0;
@%p2 bra $L__BB116_2;

cvta.to.global.u64 %rd47, %rd40;
mul.wide.u32 %rd48, %r6, 4;
add.s64 %rd49, %rd47, %rd48;
ld.global.nc.u32 %r1585, [%rd49];
mul.hi.u32 %r68, %r6, %r48;
add.s32 %r69, %r68, %r6;
shr.u32 %r70, %r69, %r49;
mul.lo.s32 %r71, %r70, %r50;
sub.s32 %r1586, %r6, %r71;
bra.uni $L__BB116_3;

$L__BB116_2:
mul.hi.u32 %r72, %r6, %r53;
add.s32 %r73, %r72, %r6;
shr.u32 %r1585, %r73, %r54;
mov.u32 %r1586, %r6;

$L__BB116_3:
mov.u32 %r12, %ctaid.z;
selp.b32 %r13, %r6, %r1585, %p2;
setp.ne.s32 %p4, %r1, 0;
setp.gt.u32 %p5, %r2, 3;
or.pred %p6, %p5, %p4;
add.s32 %r74, %r4, %r2;
setp.ge.u32 %p7, %r74, %r52;
mov.f32 %f10, 0f3F800000;
mov.f32 %f33, 0f00000000;
or.pred %p8, %p6, %p7;
mov.f32 %f8, %f33;
mov.f32 %f9, %f33;
mov.f32 %f11, %f10;
@%p8 bra $L__BB116_12;

setp.eq.s64 %p9, %rd41, 0;
mov.f32 %f8, 0f00000000;
mov.f32 %f9, %f8;
@%p9 bra $L__BB116_6;

cvta.to.global.u64 %rd50, %rd41;
mul.lo.s32 %r75, %r12, %r64;
cvt.u64.u32 %rd51, %r75;
mul.lo.s32 %r76, %r13, %r58;
cvt.u64.u32 %rd52, %r76;
cvt.u64.u32 %rd53, %r2;
cvt.s64.s32 %rd54, %r4;
add.s64 %rd55, %rd54, %rd53;
add.s64 %rd56, %rd55, %rd51;
add.s64 %rd57, %rd56, %rd52;
shl.b64 %rd58, %rd57, 2;
add.s64 %rd59, %rd50, %rd58;
ld.global.f32 %f9, [%rd59];

$L__BB116_6:
setp.ne.s64 %p1, %rd42, 0;
setp.eq.s64 %p10, %rd42, 0;
setp.eq.s64 %p11, %rd43, 0;
or.pred %p12, %p11, %p10;
@%p12 bra $L__BB116_8;

cvta.to.global.u64 %rd60, %rd43;
mul.lo.s32 %r77, %r12, %r64;
cvt.u64.u32 %rd61, %r77;
mul.lo.s32 %r78, %r13, %r58;
cvt.u64.u32 %rd62, %r78;
cvt.u64.u32 %rd63, %r2;
cvt.s64.s32 %rd64, %r4;
add.s64 %rd65, %rd64, %rd63;
add.s64 %rd66, %rd65, %rd61;
add.s64 %rd67, %rd66, %rd62;
shl.b64 %rd68, %rd67, 2;
add.s64 %rd69, %rd60, %rd68;
ld.global.f32 %f8, [%rd69];

$L__BB116_8:
setp.eq.s64 %p13, %rd44, 0;
mov.f32 %f11, 0f3F800000;
mov.f32 %f10, %f11;
@%p13 bra $L__BB116_10;

cvta.to.global.u64 %rd70, %rd44;
cvt.u64.u32 %rd71, %r1585;
selp.b64 %rd72, 0, %rd71, %p2;
shl.b64 %rd73, %rd72, 2;
add.s64 %rd74, %rd70, %rd73;
ld.global.f32 %f10, [%rd74];

$L__BB116_10:
setp.eq.s64 %p15, %rd45, 0;
not.pred %p16, %p1;
or.pred %p17, %p15, %p16;
@%p17 bra $L__BB116_12;

cvta.to.global.u64 %rd75, %rd45;
cvt.u64.u32 %rd76, %r1585;
selp.b64 %rd77, 0, %rd76, %p2;
shl.b64 %rd78, %rd77, 2;
add.s64 %rd79, %rd75, %rd78;
ld.global.f32 %f11, [%rd79];

$L__BB116_12:
mul.lo.s32 %r79, %r12, %r63;
cvt.u64.u32 %rd80, %r79;
mul.lo.s32 %r80, %r1586, %r57;
cvt.u64.u32 %rd81, %r80;
add.s64 %rd6, %rd81, %rd80;
mul.hi.u32 %r81, %r12, %r59;
add.s32 %r82, %r81, %r12;
shr.u32 %r14, %r82, %r60;
shr.u32 %r83, %r3, 31;
add.s32 %r84, %r3, %r83;
shr.s32 %r1588, %r84, 1;
setp.ge.s32 %p19, %r1588, %r5;
mov.f32 %f34, %f33;
mov.f32 %f35, %f33;
mov.f32 %f36, %f33;
mov.f32 %f37, %f33;
mov.f32 %f38, %f33;
mov.f32 %f39, %f33;
mov.f32 %f40, %f33;
@%p19 bra $L__BB116_23;

mul.lo.s32 %r85, %r14, %r62;
mul.lo.s32 %r86, %r1585, %r56;
and.b32 %r89, %r84, -2;
sub.s32 %r16, %r3, %r89;
shl.b32 %r90, %r16, 2;
shr.s32 %r91, %r90, 1;
shl.b32 %r92, %r16, 1;
cvt.s64.s32 %rd7, %r16;
cvt.s64.s32 %rd82, %r92;
or.b32 %r93, %r91, 1;
cvt.s64.s32 %rd83, %r93;
shl.b32 %r1587, %r1588, 1;
add.s32 %r94, %r1588, %r85;
add.s32 %r95, %r94, %r86;
add.s32 %r96, %r4, 1;
mad.lo.s32 %r97, %r51, %r96, %r95;
mul.wide.s32 %rd84, %r97, 36;
add.s64 %rd8, %rd84, %rd83;
add.s64 %rd9, %rd84, %rd82;
mad.lo.s32 %r98, %r4, %r51, %r1588;
add.s32 %r99, %r98, %r85;
add.s32 %r100, %r99, %r86;
mul.wide.s32 %rd85, %r100, 36;
mul.wide.s32 %rd86, %r90, 4;
add.s64 %rd87, %rd85, %rd86;
add.s64 %rd10, %rd87, 16;
add.s64 %rd11, %rd85, %rd82;
add.s64 %rd88, %rd84, %rd86;
add.s64 %rd12, %rd88, 8;
add.s64 %rd13, %rd85, %rd83;
add.s32 %r101, %r4, 2;
mad.lo.s32 %r102, %r51, %r101, %r95;
mul.wide.s32 %rd89, %r102, 36;
add.s64 %rd14, %rd89, %rd83;
add.s64 %rd15, %rd89, %rd82;
add.s32 %r103, %r4, 3;
mad.lo.s32 %r104, %r51, %r103, %r95;
mul.wide.s32 %rd90, %r104, 36;
add.s64 %rd91, %rd90, %rd86;
add.s64 %rd16, %rd91, 16;
add.s64 %rd17, %rd90, %rd82;
add.s64 %rd92, %rd89, %rd86;
add.s64 %rd18, %rd92, 8;
add.s64 %rd19, %rd90, %rd83;
mul.wide.s32 %rd20, %r51, 36;
cvta.to.global.u64 %rd124, %rd42;
cvta.to.global.u64 %rd22, %rd39;
cvta.to.global.u64 %rd123, %rd38;
mov.f32 %f33, 0f00000000;
mov.f32 %f34, %f33;
mov.f32 %f35, %f33;
mov.f32 %f36, %f33;
mov.f32 %f37, %f33;
mov.f32 %f38, %f33;
mov.f32 %f39, %f33;
mov.f32 %f40, %f33;

$L__BB116_14:
cvt.u64.u32 %rd93, %r1587;
add.s64 %rd94, %rd6, %rd93;
add.s64 %rd95, %rd123, %rd10;
ld.global.nc.u32 %r141, [%rd95+-12];
and.b32 %r142, %r141, -2004318072;
shr.u32 %r143, %r142, 1;
or.b32 %r144, %r143, 839922192;
and.b32 %r145, %r141, 30583;
mov.u32 %r146, 201852420;
mov.u32 %r147, 50462976;
prmt.b32 %r148, %r147, %r146, %r145;
mov.u32 %r149, -185009412;
mov.u32 %r150, -33620224;
prmt.b32 %r151, %r150, %r149, %r145;
and.b32 %r152, %r144, 30292;
prmt.b32 %r153, %r148, %r151, %r152;
shr.u32 %r154, %r141, 16;
and.b32 %r155, %r154, 30583;
prmt.b32 %r156, %r147, %r146, %r155;
prmt.b32 %r157, %r150, %r149, %r155;
shr.u32 %r158, %r144, 16;
prmt.b32 %r159, %r156, %r157, %r158;
mov.u32 %r160, 25632;
prmt.b32 %r106, %r153, %r159, %r160;
mov.u32 %r161, 30001;
prmt.b32 %r110, %r153, %r159, %r161;
ld.global.nc.u32 %r162, [%rd95+-8];
and.b32 %r163, %r162, -2004318072;
shr.u32 %r164, %r163, 1;
or.b32 %r165, %r164, 839922192;
and.b32 %r166, %r162, 30583;
prmt.b32 %r167, %r147, %r146, %r166;
prmt.b32 %r168, %r150, %r149, %r166;
and.b32 %r169, %r165, 30292;
prmt.b32 %r170, %r167, %r168, %r169;
shr.u32 %r171, %r162, 16;
and.b32 %r172, %r171, 30583;
prmt.b32 %r173, %r147, %r146, %r172;
prmt.b32 %r174, %r150, %r149, %r172;
shr.u32 %r175, %r165, 16;
prmt.b32 %r176, %r173, %r174, %r175;
prmt.b32 %r114, %r170, %r176, %r160;
prmt.b32 %r118, %r170, %r176, %r161;
add.s64 %rd96, %rd94, %rd7;
mul.lo.s64 %rd97, %rd96, 36;
add.s64 %rd98, %rd22, %rd97;
ld.global.nc.u32 %r20, [%rd98+4];
mov.u32 %r126, 0;

dp4a.s32.s32 %r105, %r106, %r20, %r126;

ld.global.nc.u32 %r21, [%rd98+12];

dp4a.s32.s32 %r109, %r110, %r21, %r105;

ld.global.nc.u32 %r22, [%rd98+8];

dp4a.s32.s32 %r113, %r114, %r22, %r109;

ld.global.nc.u32 %r23, [%rd98+16];

dp4a.s32.s32 %r117, %r118, %r23, %r113;

add.s64 %rd27, %rd123, %rd11;
ld.global.nc.u8 %rs5, [%rd27];
and.b16 %rs6, %rs5, 127;
setp.eq.s16 %p20, %rs6, 127;
selp.b16 %rs7, 0, %rs5, %p20;
and.b16 %rs1, %rs7, 255;

{cvt.rn.f16x2.e4m3x2 %r121, %rs1;}


cvt.u16.u32 %rs2, %r121;

{cvt.f32.f16 %f111, %rs2;}


mov.f32 %f115, 0f40000000;
div.approx.ftz.f32 %f116, %f111, %f115;
ld.global.nc.u32 %r24, [%rd98];

{.reg .f16 low,high;
mov.b32 {low,high},%r24;
cvt.f32.f16 %f112, low;}


mul.ftz.f32 %f117, %f116, %f112;
cvt.rn.f32.s32 %f118, %r117;
fma.rn.ftz.f32 %f119, %f117, %f118, 0f00000000;
ld.global.nc.u32 %r177, [%rd95+-4];
and.b32 %r178, %r177, -2004318072;
shr.u32 %r179, %r178, 1;
or.b32 %r180, %r179, 839922192;
and.b32 %r181, %r177, 30583;
prmt.b32 %r182, %r147, %r146, %r181;
prmt.b32 %r183, %r150, %r149, %r181;
and.b32 %r184, %r180, 30292;
prmt.b32 %r185, %r182, %r183, %r184;
shr.u32 %r186, %r177, 16;
and.b32 %r187, %r186, 30583;
prmt.b32 %r188, %r147, %r146, %r187;
prmt.b32 %r189, %r150, %r149, %r187;
shr.u32 %r190, %r180, 16;
prmt.b32 %r191, %r188, %r189, %r190;
prmt.b32 %r124, %r185, %r191, %r160;
prmt.b32 %r128, %r185, %r191, %r161;
ld.global.nc.u32 %r192, [%rd95];
and.b32 %r193, %r192, -2004318072;
shr.u32 %r194, %r193, 1;
or.b32 %r195, %r194, 839922192;
and.b32 %r196, %r192, 30583;
prmt.b32 %r197, %r147, %r146, %r196;
prmt.b32 %r198, %r150, %r149, %r196;
and.b32 %r199, %r195, 30292;
prmt.b32 %r200, %r197, %r198, %r199;
shr.u32 %r201, %r192, 16;
and.b32 %r202, %r201, 30583;
prmt.b32 %r203, %r147, %r146, %r202;
prmt.b32 %r204, %r150, %r149, %r202;
shr.u32 %r205, %r195, 16;
prmt.b32 %r206, %r203, %r204, %r205;
prmt.b32 %r132, %r200, %r206, %r160;
prmt.b32 %r136, %r200, %r206, %r161;
shr.s32 %r208, %r90, 2;
cvt.s64.s32 %rd99, %r208;
add.s64 %rd100, %rd94, %rd99;
mul.lo.s64 %rd101, %rd100, 36;
add.s64 %rd102, %rd22, %rd101;
ld.global.nc.u32 %r25, [%rd102+20];

dp4a.s32.s32 %r123, %r124, %r25, %r126;

ld.global.nc.u32 %r26, [%rd102+28];

dp4a.s32.s32 %r127, %r128, %r26, %r123;

ld.global.nc.u32 %r27, [%rd102+24];

dp4a.s32.s32 %r131, %r132, %r27, %r127;

ld.global.nc.u32 %r28, [%rd102+32];

dp4a.s32.s32 %r135, %r136, %r28, %r131;

add.s64 %rd28, %rd123, %rd13;
ld.global.nc.u8 %rs8, [%rd28];
and.b16 %rs9, %rs8, 127;
setp.eq.s16 %p21, %rs9, 127;
selp.b16 %rs10, 0, %rs8, %p21;
and.b16 %rs3, %rs10, 255;

{cvt.rn.f16x2.e4m3x2 %r139, %rs3;}


cvt.u16.u32 %rs4, %r139;

{cvt.f32.f16 %f113, %rs4;}


div.approx.ftz.f32 %f120, %f113, %f115;
ld.global.nc.u32 %r29, [%rd102];

{.reg .f16 low,high;
mov.b32 {low,high},%r29;
cvt.f32.f16 %f114, low;}


mul.ftz.f32 %f121, %f120, %f114;
cvt.rn.f32.s32 %f122, %r135;
fma.rn.ftz.f32 %f123, %f121, %f122, %f119;
add.ftz.f32 %f36, %f123, %f36;
setp.eq.s64 %p22, %rd42, 0;
@%p22 bra $L__BB116_16;

add.s64 %rd103, %rd124, %rd10;
ld.global.u32 %r245, [%rd103+-12];
and.b32 %r246, %r245, -2004318072;
shr.u32 %r247, %r246, 1;
or.b32 %r248, %r247, 839922192;
and.b32 %r249, %r245, 30583;
prmt.b32 %r252, %r147, %r146, %r249;
prmt.b32 %r255, %r150, %r149, %r249;
and.b32 %r256, %r248, 30292;
prmt.b32 %r257, %r252, %r255, %r256;
shr.u32 %r258, %r245, 16;
and.b32 %r259, %r258, 30583;
prmt.b32 %r260, %r147, %r146, %r259;
prmt.b32 %r261, %r150, %r149, %r259;
shr.u32 %r262, %r248, 16;
prmt.b32 %r263, %r260, %r261, %r262;
prmt.b32 %r210, %r257, %r263, %r160;
prmt.b32 %r214, %r257, %r263, %r161;
ld.global.u32 %r266, [%rd103+-8];
and.b32 %r267, %r266, -2004318072;
shr.u32 %r268, %r267, 1;
or.b32 %r269, %r268, 839922192;
and.b32 %r270, %r266, 30583;
prmt.b32 %r271, %r147, %r146, %r270;
prmt.b32 %r272, %r150, %r149, %r270;
and.b32 %r273, %r269, 30292;
prmt.b32 %r274, %r271, %r272, %r273;
shr.u32 %r275, %r266, 16;
and.b32 %r276, %r275, 30583;
prmt.b32 %r277, %r147, %r146, %r276;
prmt.b32 %r278, %r150, %r149, %r276;
shr.u32 %r279, %r269, 16;
prmt.b32 %r280, %r277, %r278, %r279;
prmt.b32 %r218, %r274, %r280, %r160;
prmt.b32 %r222, %r274, %r280, %r161;

dp4a.s32.s32 %r209, %r210, %r20, %r126;


dp4a.s32.s32 %r213, %r214, %r21, %r209;


dp4a.s32.s32 %r217, %r218, %r22, %r213;


dp4a.s32.s32 %r221, %r222, %r23, %r217;

add.s64 %rd104, %rd124, %rd11;
ld.global.u8 %rs15, [%rd104];
and.b16 %rs16, %rs15, 127;
setp.eq.s16 %p23, %rs16, 127;
selp.b16 %rs11, 0, %rs15, %p23;

{cvt.rn.f16x2.e4m3x2 %r225, %rs11;}


cvt.u16.u32 %rs12, %r225;

{cvt.f32.f16 %f124, %rs12;}


div.approx.ftz.f32 %f129, %f124, %f115;

{.reg .f16 low,high;
mov.b32 {low,high},%r24;
cvt.f32.f16 %f125, low;}


mul.ftz.f32 %f130, %f129, %f125;
cvt.rn.f32.s32 %f131, %r221;
fma.rn.ftz.f32 %f132, %f130, %f131, 0f00000000;
ld.global.u32 %r281, [%rd103+-4];
and.b32 %r282, %r281, -2004318072;
shr.u32 %r283, %r282, 1;
or.b32 %r284, %r283, 839922192;
and.b32 %r285, %r281, 30583;
prmt.b32 %r286, %r147, %r146, %r285;
prmt.b32 %r287, %r150, %r149, %r285;
and.b32 %r288, %r284, 30292;
prmt.b32 %r289, %r286, %r287, %r288;
shr.u32 %r290, %r281, 16;
and.b32 %r291, %r290, 30583;
prmt.b32 %r292, %r147, %r146, %r291;
prmt.b32 %r293, %r150, %r149, %r291;
shr.u32 %r294, %r284, 16;
prmt.b32 %r295, %r292, %r293, %r294;
prmt.b32 %r228, %r289, %r295, %r160;
prmt.b32 %r232, %r289, %r295, %r161;
ld.global.u32 %r296, [%rd103];
and.b32 %r297, %r296, -2004318072;
shr.u32 %r298, %r297, 1;
or.b32 %r299, %r298, 839922192;
and.b32 %r300, %r296, 30583;
prmt.b32 %r301, %r147, %r146, %r300;
prmt.b32 %r302, %r150, %r149, %r300;
and.b32 %r303, %r299, 30292;
prmt.b32 %r304, %r301, %r302, %r303;
shr.u32 %r305, %r296, 16;
and.b32 %r306, %r305, 30583;
prmt.b32 %r307, %r147, %r146, %r306;
prmt.b32 %r308, %r150, %r149, %r306;
shr.u32 %r309, %r299, 16;
prmt.b32 %r310, %r307, %r308, %r309;
prmt.b32 %r236, %r304, %r310, %r160;
prmt.b32 %r240, %r304, %r310, %r161;

dp4a.s32.s32 %r227, %r228, %r25, %r126;


dp4a.s32.s32 %r231, %r232, %r26, %r227;


dp4a.s32.s32 %r235, %r236, %r27, %r231;


dp4a.s32.s32 %r239, %r240, %r28, %r235;

add.s64 %rd105, %rd124, %rd13;
ld.global.u8 %rs17, [%rd105];
and.b16 %rs18, %rs17, 127;
setp.eq.s16 %p24, %rs18, 127;
selp.b16 %rs13, 0, %rs17, %p24;

{cvt.rn.f16x2.e4m3x2 %r243, %rs13;}


cvt.u16.u32 %rs14, %r243;

{cvt.f32.f16 %f126, %rs14;}


div.approx.ftz.f32 %f133, %f126, %f115;

{.reg .f16 low,high;
mov.b32 {low,high},%r29;
cvt.f32.f16 %f127, low;}


mul.ftz.f32 %f134, %f133, %f127;
cvt.rn.f32.s32 %f135, %r239;
fma.rn.ftz.f32 %f136, %f134, %f135, %f132;
add.ftz.f32 %f40, %f136, %f40;

$L__BB116_16:
add.s64 %rd121, %rd123, %rd10;
add.s64 %rd120, %rd121, -12;
add.s64 %rd29, %rd120, %rd20;
ld.global.nc.u32 %r347, [%rd29];
and.b32 %r348, %r347, -2004318072;
shr.u32 %r349, %r348, 1;
or.b32 %r350, %r349, 839922192;
and.b32 %r351, %r347, 30583;
prmt.b32 %r354, %r147, %r146, %r351;
prmt.b32 %r357, %r150, %r149, %r351;
and.b32 %r358, %r350, 30292;
prmt.b32 %r359, %r354, %r357, %r358;
shr.u32 %r360, %r347, 16;
and.b32 %r361, %r360, 30583;
prmt.b32 %r362, %r147, %r146, %r361;
prmt.b32 %r363, %r150, %r149, %r361;
shr.u32 %r364, %r350, 16;
prmt.b32 %r365, %r362, %r363, %r364;
prmt.b32 %r312, %r359, %r365, %r160;
prmt.b32 %r316, %r359, %r365, %r161;
ld.global.nc.u32 %r368, [%rd29+4];
and.b32 %r369, %r368, -2004318072;
shr.u32 %r370, %r369, 1;
or.b32 %r371, %r370, 839922192;
and.b32 %r372, %r368, 30583;
prmt.b32 %r373, %r147, %r146, %r372;
prmt.b32 %r374, %r150, %r149, %r372;
and.b32 %r375, %r371, 30292;
prmt.b32 %r376, %r373, %r374, %r375;
shr.u32 %r377, %r368, 16;
and.b32 %r378, %r377, 30583;
prmt.b32 %r379, %r147, %r146, %r378;
prmt.b32 %r380, %r150, %r149, %r378;
shr.u32 %r381, %r371, 16;
prmt.b32 %r382, %r379, %r380, %r381;
prmt.b32 %r320, %r376, %r382, %r160;
prmt.b32 %r324, %r376, %r382, %r161;

dp4a.s32.s32 %r311, %r312, %r20, %r126;


dp4a.s32.s32 %r315, %r316, %r21, %r311;


dp4a.s32.s32 %r319, %r320, %r22, %r315;


dp4a.s32.s32 %r323, %r324, %r23, %r319;

add.s64 %rd30, %rd27, %rd20;
ld.global.nc.u8 %rs23, [%rd30];
and.b16 %rs24, %rs23, 127;
setp.eq.s16 %p25, %rs24, 127;
selp.b16 %rs25, 0, %rs23, %p25;
and.b16 %rs19, %rs25, 255;

{cvt.rn.f16x2.e4m3x2 %r327, %rs19;}


cvt.u16.u32 %rs20, %r327;

{cvt.f32.f16 %f137, %rs20;}


div.approx.ftz.f32 %f142, %f137, %f115;

{.reg .f16 low,high;
mov.b32 {low,high},%r24;
cvt.f32.f16 %f138, low;}


mul.ftz.f32 %f143, %f142, %f138;
cvt.rn.f32.s32 %f144, %r323;
fma.rn.ftz.f32 %f145, %f143, %f144, 0f00000000;
ld.global.nc.u32 %r383, [%rd29+8];
and.b32 %r384, %r383, -2004318072;
shr.u32 %r385, %r384, 1;
or.b32 %r386, %r385, 839922192;
and.b32 %r387, %r383, 30583;
prmt.b32 %r388, %r147, %r146, %r387;
prmt.b32 %r389, %r150, %r149, %r387;
and.b32 %r390, %r386, 30292;
prmt.b32 %r391, %r388, %r389, %r390;
shr.u32 %r392, %r383, 16;
and.b32 %r393, %r392, 30583;
prmt.b32 %r394, %r147, %r146, %r393;
prmt.b32 %r395, %r150, %r149, %r393;
shr.u32 %r396, %r386, 16;
prmt.b32 %r397, %r394, %r395, %r396;
prmt.b32 %r330, %r391, %r397, %r160;
prmt.b32 %r334, %r391, %r397, %r161;
ld.global.nc.u32 %r398, [%rd29+12];
and.b32 %r399, %r398, -2004318072;
shr.u32 %r400, %r399, 1;
or.b32 %r401, %r400, 839922192;
and.b32 %r402, %r398, 30583;
prmt.b32 %r403, %r147, %r146, %r402;
prmt.b32 %r404, %r150, %r149, %r402;
and.b32 %r405, %r401, 30292;
prmt.b32 %r406, %r403, %r404, %r405;
shr.u32 %r407, %r398, 16;
and.b32 %r408, %r407, 30583;
prmt.b32 %r409, %r147, %r146, %r408;
prmt.b32 %r410, %r150, %r149, %r408;
shr.u32 %r411, %r401, 16;
prmt.b32 %r412, %r409, %r410, %r411;
prmt.b32 %r338, %r406, %r412, %r160;
prmt.b32 %r342, %r406, %r412, %r161;

dp4a.s32.s32 %r329, %r330, %r25, %r126;


dp4a.s32.s32 %r333, %r334, %r26, %r329;


dp4a.s32.s32 %r337, %r338, %r27, %r333;


dp4a.s32.s32 %r341, %r342, %r28, %r337;

add.s64 %rd31, %rd28, %rd20;
ld.global.nc.u8 %rs26, [%rd31];
and.b16 %rs27, %rs26, 127;
setp.eq.s16 %p26, %rs27, 127;
selp.b16 %rs28, 0, %rs26, %p26;
and.b16 %rs21, %rs28, 255;

{cvt.rn.f16x2.e4m3x2 %r345, %rs21;}


cvt.u16.u32 %rs22, %r345;

{cvt.f32.f16 %f139, %rs22;}


div.approx.ftz.f32 %f146, %f139, %f115;

{.reg .f16 low,high;
mov.b32 {low,high},%r29;
cvt.f32.f16 %f140, low;}


mul.ftz.f32 %f147, %f146, %f140;
cvt.rn.f32.s32 %f148, %r341;
fma.rn.ftz.f32 %f149, %f147, %f148, %f145;
add.ftz.f32 %f35, %f149, %f35;
@%p22 bra $L__BB116_18;

add.s64 %rd106, %rd124, %rd12;
ld.global.u32 %r449, [%rd106+-4];
and.b32 %r450, %r449, -2004318072;
shr.u32 %r451, %r450, 1;
or.b32 %r452, %r451, 839922192;
and.b32 %r453, %r449, 30583;
mov.u32 %r454, 201852420;
mov.u32 %r455, 50462976;
prmt.b32 %r456, %r455, %r454, %r453;
mov.u32 %r457, -185009412;
mov.u32 %r458, -33620224;
prmt.b32 %r459, %r458, %r457, %r453;
and.b32 %r460, %r452, 30292;
prmt.b32 %r461, %r456, %r459, %r460;
shr.u32 %r462, %r449, 16;
and.b32 %r463, %r462, 30583;
prmt.b32 %r464, %r455, %r454, %r463;
prmt.b32 %r465, %r458, %r457, %r463;
shr.u32 %r466, %r452, 16;
prmt.b32 %r467, %r464, %r465, %r466;
mov.u32 %r468, 25632;
prmt.b32 %r414, %r461, %r467, %r468;
mov.u32 %r469, 30001;
prmt.b32 %r418, %r461, %r467, %r469;
ld.global.u32 %r470, [%rd106];
and.b32 %r471, %r470, -2004318072;
shr.u32 %r472, %r471, 1;
or.b32 %r473, %r472, 839922192;
and.b32 %r474, %r470, 30583;
prmt.b32 %r475, %r455, %r454, %r474;
prmt.b32 %r476, %r458, %r457, %r474;
and.b32 %r477, %r473, 30292;
prmt.b32 %r478, %r475, %r476, %r477;
shr.u32 %r479, %r470, 16;
and.b32 %r480, %r479, 30583;
prmt.b32 %r481, %r455, %r454, %r480;
prmt.b32 %r482, %r458, %r457, %r480;
shr.u32 %r483, %r473, 16;
prmt.b32 %r484, %r481, %r482, %r483;
prmt.b32 %r422, %r478, %r484, %r468;
prmt.b32 %r426, %r478, %r484, %r469;
mov.u32 %r434, 0;

dp4a.s32.s32 %r413, %r414, %r20, %r434;


dp4a.s32.s32 %r417, %r418, %r21, %r413;


dp4a.s32.s32 %r421, %r422, %r22, %r417;


dp4a.s32.s32 %r425, %r426, %r23, %r421;

add.s64 %rd107, %rd124, %rd9;
ld.global.u8 %rs33, [%rd107];
and.b16 %rs34, %rs33, 127;
setp.eq.s16 %p28, %rs34, 127;
selp.b16 %rs29, 0, %rs33, %p28;

{cvt.rn.f16x2.e4m3x2 %r429, %rs29;}


cvt.u16.u32 %rs30, %r429;

{cvt.f32.f16 %f150, %rs30;}


mov.f32 %f154, 0f40000000;
div.approx.ftz.f32 %f155, %f150, %f154;

{.reg .f16 low,high;
mov.b32 {low,high},%r24;
cvt.f32.f16 %f151, low;}


mul.ftz.f32 %f156, %f155, %f151;
cvt.rn.f32.s32 %f157, %r425;
fma.rn.ftz.f32 %f158, %f156, %f157, 0f00000000;
ld.global.u32 %r485, [%rd106+4];
and.b32 %r486, %r485, -2004318072;
shr.u32 %r487, %r486, 1;
or.b32 %r488, %r487, 839922192;
and.b32 %r489, %r485, 30583;
prmt.b32 %r490, %r455, %r454, %r489;
prmt.b32 %r491, %r458, %r457, %r489;
and.b32 %r492, %r488, 30292;
prmt.b32 %r493, %r490, %r491, %r492;
shr.u32 %r494, %r485, 16;
and.b32 %r495, %r494, 30583;
prmt.b32 %r496, %r455, %r454, %r495;
prmt.b32 %r497, %r458, %r457, %r495;
shr.u32 %r498, %r488, 16;
prmt.b32 %r499, %r496, %r497, %r498;
prmt.b32 %r432, %r493, %r499, %r468;
prmt.b32 %r436, %r493, %r499, %r469;
ld.global.u32 %r500, [%rd106+8];
and.b32 %r501, %r500, -2004318072;
shr.u32 %r502, %r501, 1;
or.b32 %r503, %r502, 839922192;
and.b32 %r504, %r500, 30583;
prmt.b32 %r505, %r455, %r454, %r504;
prmt.b32 %r506, %r458, %r457, %r504;
and.b32 %r507, %r503, 30292;
prmt.b32 %r508, %r505, %r506, %r507;
shr.u32 %r509, %r500, 16;
and.b32 %r510, %r509, 30583;
prmt.b32 %r511, %r455, %r454, %r510;
prmt.b32 %r512, %r458, %r457, %r510;
shr.u32 %r513, %r503, 16;
prmt.b32 %r514, %r511, %r512, %r513;
prmt.b32 %r440, %r508, %r514, %r468;
prmt.b32 %r444, %r508, %r514, %r469;

dp4a.s32.s32 %r431, %r432, %r25, %r434;


dp4a.s32.s32 %r435, %r436, %r26, %r431;


dp4a.s32.s32 %r439, %r440, %r27, %r435;


dp4a.s32.s32 %r443, %r444, %r28, %r439;

add.s64 %rd108, %rd124, %rd8;
ld.global.u8 %rs35, [%rd108];
and.b16 %rs36, %rs35, 127;
setp.eq.s16 %p29, %rs36, 127;
selp.b16 %rs31, 0, %rs35, %p29;

{cvt.rn.f16x2.e4m3x2 %r447, %rs31;}


cvt.u16.u32 %rs32, %r447;

{cvt.f32.f16 %f152, %rs32;}


div.approx.ftz.f32 %f159, %f152, %f154;

{.reg .f16 low,high;
mov.b32 {low,high},%r29;
cvt.f32.f16 %f153, low;}


mul.ftz.f32 %f160, %f159, %f153;
cvt.rn.f32.s32 %f161, %r443;
fma.rn.ftz.f32 %f162, %f160, %f161, %f158;
add.ftz.f32 %f39, %f162, %f39;

$L__BB116_18:
add.s64 %rd32, %rd29, %rd20;
ld.global.nc.u32 %r551, [%rd32];
and.b32 %r552, %r551, -2004318072;
shr.u32 %r553, %r552, 1;
or.b32 %r554, %r553, 839922192;
and.b32 %r555, %r551, 30583;
mov.u32 %r556, 201852420;
mov.u32 %r557, 50462976;
prmt.b32 %r558, %r557, %r556, %r555;
mov.u32 %r559, -185009412;
mov.u32 %r560, -33620224;
prmt.b32 %r561, %r560, %r559, %r555;
and.b32 %r562, %r554, 30292;
prmt.b32 %r563, %r558, %r561, %r562;
shr.u32 %r564, %r551, 16;
and.b32 %r565, %r564, 30583;
prmt.b32 %r566, %r557, %r556, %r565;
prmt.b32 %r567, %r560, %r559, %r565;
shr.u32 %r568, %r554, 16;
prmt.b32 %r569, %r566, %r567, %r568;
mov.u32 %r570, 25632;
prmt.b32 %r516, %r563, %r569, %r570;
mov.u32 %r571, 30001;
prmt.b32 %r520, %r563, %r569, %r571;
ld.global.nc.u32 %r572, [%rd32+4];
and.b32 %r573, %r572, -2004318072;
shr.u32 %r574, %r573, 1;
or.b32 %r575, %r574, 839922192;
and.b32 %r576, %r572, 30583;
prmt.b32 %r577, %r557, %r556, %r576;
prmt.b32 %r578, %r560, %r559, %r576;
and.b32 %r579, %r575, 30292;
prmt.b32 %r580, %r577, %r578, %r579;
shr.u32 %r581, %r572, 16;
and.b32 %r582, %r581, 30583;
prmt.b32 %r583, %r557, %r556, %r582;
prmt.b32 %r584, %r560, %r559, %r582;
shr.u32 %r585, %r575, 16;
prmt.b32 %r586, %r583, %r584, %r585;
prmt.b32 %r524, %r580, %r586, %r570;
prmt.b32 %r528, %r580, %r586, %r571;
mov.u32 %r536, 0;

dp4a.s32.s32 %r515, %r516, %r20, %r536;


dp4a.s32.s32 %r519, %r520, %r21, %r515;


dp4a.s32.s32 %r523, %r524, %r22, %r519;


dp4a.s32.s32 %r527, %r528, %r23, %r523;

add.s64 %rd33, %rd30, %rd20;
ld.global.nc.u8 %rs41, [%rd33];
and.b16 %rs42, %rs41, 127;
setp.eq.s16 %p30, %rs42, 127;
selp.b16 %rs43, 0, %rs41, %p30;
and.b16 %rs37, %rs43, 255;

{cvt.rn.f16x2.e4m3x2 %r531, %rs37;}


cvt.u16.u32 %rs38, %r531;

{cvt.f32.f16 %f163, %rs38;}


mov.f32 %f167, 0f40000000;
div.approx.ftz.f32 %f168, %f163, %f167;

{.reg .f16 low,high;
mov.b32 {low,high},%r24;
cvt.f32.f16 %f164, low;}


mul.ftz.f32 %f169, %f168, %f164;
cvt.rn.f32.s32 %f170, %r527;
fma.rn.ftz.f32 %f171, %f169, %f170, 0f00000000;
ld.global.nc.u32 %r587, [%rd32+8];
and.b32 %r588, %r587, -2004318072;
shr.u32 %r589, %r588, 1;
or.b32 %r590, %r589, 839922192;
and.b32 %r591, %r587, 30583;
prmt.b32 %r592, %r557, %r556, %r591;
prmt.b32 %r593, %r560, %r559, %r591;
and.b32 %r594, %r590, 30292;
prmt.b32 %r595, %r592, %r593, %r594;
shr.u32 %r596, %r587, 16;
and.b32 %r597, %r596, 30583;
prmt.b32 %r598, %r557, %r556, %r597;
prmt.b32 %r599, %r560, %r559, %r597;
shr.u32 %r600, %r590, 16;
prmt.b32 %r601, %r598, %r599, %r600;
prmt.b32 %r534, %r595, %r601, %r570;
prmt.b32 %r538, %r595, %r601, %r571;
ld.global.nc.u32 %r602, [%rd32+12];
and.b32 %r603, %r602, -2004318072;
shr.u32 %r604, %r603, 1;
or.b32 %r605, %r604, 839922192;
and.b32 %r606, %r602, 30583;
prmt.b32 %r607, %r557, %r556, %r606;
prmt.b32 %r608, %r560, %r559, %r606;
and.b32 %r609, %r605, 30292;
prmt.b32 %r610, %r607, %r608, %r609;
shr.u32 %r611, %r602, 16;
and.b32 %r612, %r611, 30583;
prmt.b32 %r613, %r557, %r556, %r612;
prmt.b32 %r614, %r560, %r559, %r612;
shr.u32 %r615, %r605, 16;
prmt.b32 %r616, %r613, %r614, %r615;
prmt.b32 %r542, %r610, %r616, %r570;
prmt.b32 %r546, %r610, %r616, %r571;

dp4a.s32.s32 %r533, %r534, %r25, %r536;


dp4a.s32.s32 %r537, %r538, %r26, %r533;


dp4a.s32.s32 %r541, %r542, %r27, %r537;


dp4a.s32.s32 %r545, %r546, %r28, %r541;

add.s64 %rd34, %rd31, %rd20;
ld.global.nc.u8 %rs44, [%rd34];
and.b16 %rs45, %rs44, 127;
setp.eq.s16 %p31, %rs45, 127;
selp.b16 %rs46, 0, %rs44, %p31;
and.b16 %rs39, %rs46, 255;

{cvt.rn.f16x2.e4m3x2 %r549, %rs39;}


cvt.u16.u32 %rs40, %r549;

{cvt.f32.f16 %f165, %rs40;}


div.approx.ftz.f32 %f172, %f165, %f167;

{.reg .f16 low,high;
mov.b32 {low,high},%r29;
cvt.f32.f16 %f166, low;}


mul.ftz.f32 %f173, %f172, %f166;
cvt.rn.f32.s32 %f174, %r545;
fma.rn.ftz.f32 %f175, %f173, %f174, %f171;
add.ftz.f32 %f34, %f175, %f34;
@%p22 bra $L__BB116_20;

add.s64 %rd109, %rd124, %rd18;
ld.global.u32 %r653, [%rd109+-4];
and.b32 %r654, %r653, -2004318072;
shr.u32 %r655, %r654, 1;
or.b32 %r656, %r655, 839922192;
and.b32 %r657, %r653, 30583;
prmt.b32 %r660, %r557, %r556, %r657;
prmt.b32 %r663, %r560, %r559, %r657;
and.b32 %r664, %r656, 30292;
prmt.b32 %r665, %r660, %r663, %r664;
shr.u32 %r666, %r653, 16;
and.b32 %r667, %r666, 30583;
prmt.b32 %r668, %r557, %r556, %r667;
prmt.b32 %r669, %r560, %r559, %r667;
shr.u32 %r670, %r656, 16;
prmt.b32 %r671, %r668, %r669, %r670;
prmt.b32 %r618, %r665, %r671, %r570;
prmt.b32 %r622, %r665, %r671, %r571;
ld.global.u32 %r674, [%rd109];
and.b32 %r675, %r674, -2004318072;
shr.u32 %r676, %r675, 1;
or.b32 %r677, %r676, 839922192;
and.b32 %r678, %r674, 30583;
prmt.b32 %r679, %r557, %r556, %r678;
prmt.b32 %r680, %r560, %r559, %r678;
and.b32 %r681, %r677, 30292;
prmt.b32 %r682, %r679, %r680, %r681;
shr.u32 %r683, %r674, 16;
and.b32 %r684, %r683, 30583;
prmt.b32 %r685, %r557, %r556, %r684;
prmt.b32 %r686, %r560, %r559, %r684;
shr.u32 %r687, %r677, 16;
prmt.b32 %r688, %r685, %r686, %r687;
prmt.b32 %r626, %r682, %r688, %r570;
prmt.b32 %r630, %r682, %r688, %r571;

dp4a.s32.s32 %r617, %r618, %r20, %r536;


dp4a.s32.s32 %r621, %r622, %r21, %r617;


dp4a.s32.s32 %r625, %r626, %r22, %r621;


dp4a.s32.s32 %r629, %r630, %r23, %r625;

add.s64 %rd110, %rd124, %rd15;
ld.global.u8 %rs51, [%rd110];
and.b16 %rs52, %rs51, 127;
setp.eq.s16 %p33, %rs52, 127;
selp.b16 %rs47, 0, %rs51, %p33;

{cvt.rn.f16x2.e4m3x2 %r633, %rs47;}


cvt.u16.u32 %rs48, %r633;

{cvt.f32.f16 %f176, %rs48;}


div.approx.ftz.f32 %f181, %f176, %f167;

{.reg .f16 low,high;
mov.b32 {low,high},%r24;
cvt.f32.f16 %f177, low;}


mul.ftz.f32 %f182, %f181, %f177;
cvt.rn.f32.s32 %f183, %r629;
fma.rn.ftz.f32 %f184, %f182, %f183, 0f00000000;
ld.global.u32 %r689, [%rd109+4];
and.b32 %r690, %r689, -2004318072;
shr.u32 %r691, %r690, 1;
or.b32 %r692, %r691, 839922192;
and.b32 %r693, %r689, 30583;
prmt.b32 %r694, %r557, %r556, %r693;
prmt.b32 %r695, %r560, %r559, %r693;
and.b32 %r696, %r692, 30292;
prmt.b32 %r697, %r694, %r695, %r696;
shr.u32 %r698, %r689, 16;
and.b32 %r699, %r698, 30583;
prmt.b32 %r700, %r557, %r556, %r699;
prmt.b32 %r701, %r560, %r559, %r699;
shr.u32 %r702, %r692, 16;
prmt.b32 %r703, %r700, %r701, %r702;
prmt.b32 %r636, %r697, %r703, %r570;
prmt.b32 %r640, %r697, %r703, %r571;
ld.global.u32 %r704, [%rd109+8];
and.b32 %r705, %r704, -2004318072;
shr.u32 %r706, %r705, 1;
or.b32 %r707, %r706, 839922192;
and.b32 %r708, %r704, 30583;
prmt.b32 %r709, %r557, %r556, %r708;
prmt.b32 %r710, %r560, %r559, %r708;
and.b32 %r711, %r707, 30292;
prmt.b32 %r712, %r709, %r710, %r711;
shr.u32 %r713, %r704, 16;
and.b32 %r714, %r713, 30583;
prmt.b32 %r715, %r557, %r556, %r714;
prmt.b32 %r716, %r560, %r559, %r714;
shr.u32 %r717, %r707, 16;
prmt.b32 %r718, %r715, %r716, %r717;
prmt.b32 %r644, %r712, %r718, %r570;
prmt.b32 %r648, %r712, %r718, %r571;

dp4a.s32.s32 %r635, %r636, %r25, %r536;


dp4a.s32.s32 %r639, %r640, %r26, %r635;


dp4a.s32.s32 %r643, %r644, %r27, %r639;


dp4a.s32.s32 %r647, %r648, %r28, %r643;

add.s64 %rd111, %rd124, %rd14;
ld.global.u8 %rs53, [%rd111];
and.b16 %rs54, %rs53, 127;
setp.eq.s16 %p34, %rs54, 127;
selp.b16 %rs49, 0, %rs53, %p34;

{cvt.rn.f16x2.e4m3x2 %r651, %rs49;}


cvt.u16.u32 %rs50, %r651;

{cvt.f32.f16 %f178, %rs50;}


div.approx.ftz.f32 %f185, %f178, %f167;

{.reg .f16 low,high;
mov.b32 {low,high},%r29;
cvt.f32.f16 %f179, low;}


mul.ftz.f32 %f186, %f185, %f179;
cvt.rn.f32.s32 %f187, %r647;
fma.rn.ftz.f32 %f188, %f186, %f187, %f184;
add.ftz.f32 %f38, %f188, %f38;

$L__BB116_20:
add.s64 %rd112, %rd32, %rd20;
ld.global.nc.u32 %r755, [%rd112];
and.b32 %r756, %r755, -2004318072;
shr.u32 %r757, %r756, 1;
or.b32 %r758, %r757, 839922192;
and.b32 %r759, %r755, 30583;
prmt.b32 %r762, %r557, %r556, %r759;
prmt.b32 %r765, %r560, %r559, %r759;
and.b32 %r766, %r758, 30292;
prmt.b32 %r767, %r762, %r765, %r766;
shr.u32 %r768, %r755, 16;
and.b32 %r769, %r768, 30583;
prmt.b32 %r770, %r557, %r556, %r769;
prmt.b32 %r771, %r560, %r559, %r769;
shr.u32 %r772, %r758, 16;
prmt.b32 %r773, %r770, %r771, %r772;
prmt.b32 %r720, %r767, %r773, %r570;
prmt.b32 %r724, %r767, %r773, %r571;
ld.global.nc.u32 %r776, [%rd112+4];
and.b32 %r777, %r776, -2004318072;
shr.u32 %r778, %r777, 1;
or.b32 %r779, %r778, 839922192;
and.b32 %r780, %r776, 30583;
prmt.b32 %r781, %r557, %r556, %r780;
prmt.b32 %r782, %r560, %r559, %r780;
and.b32 %r783, %r779, 30292;
prmt.b32 %r784, %r781, %r782, %r783;
shr.u32 %r785, %r776, 16;
and.b32 %r786, %r785, 30583;
prmt.b32 %r787, %r557, %r556, %r786;
prmt.b32 %r788, %r560, %r559, %r786;
shr.u32 %r789, %r779, 16;
prmt.b32 %r790, %r787, %r788, %r789;
prmt.b32 %r728, %r784, %r790, %r570;
prmt.b32 %r732, %r784, %r790, %r571;

dp4a.s32.s32 %r719, %r720, %r20, %r536;


dp4a.s32.s32 %r723, %r724, %r21, %r719;


dp4a.s32.s32 %r727, %r728, %r22, %r723;


dp4a.s32.s32 %r731, %r732, %r23, %r727;

add.s64 %rd113, %rd33, %rd20;
ld.global.nc.u8 %rs59, [%rd113];
and.b16 %rs60, %rs59, 127;
setp.eq.s16 %p35, %rs60, 127;
selp.b16 %rs61, 0, %rs59, %p35;
and.b16 %rs55, %rs61, 255;

{cvt.rn.f16x2.e4m3x2 %r735, %rs55;}


cvt.u16.u32 %rs56, %r735;

{cvt.f32.f16 %f189, %rs56;}


div.approx.ftz.f32 %f194, %f189, %f167;

{.reg .f16 low,high;
mov.b32 {low,high},%r24;
cvt.f32.f16 %f190, low;}


mul.ftz.f32 %f195, %f194, %f190;
cvt.rn.f32.s32 %f196, %r731;
fma.rn.ftz.f32 %f197, %f195, %f196, 0f00000000;
ld.global.nc.u32 %r791, [%rd112+8];
and.b32 %r792, %r791, -2004318072;
shr.u32 %r793, %r792, 1;
or.b32 %r794, %r793, 839922192;
and.b32 %r795, %r791, 30583;
prmt.b32 %r796, %r557, %r556, %r795;
prmt.b32 %r797, %r560, %r559, %r795;
and.b32 %r798, %r794, 30292;
prmt.b32 %r799, %r796, %r797, %r798;
shr.u32 %r800, %r791, 16;
and.b32 %r801, %r800, 30583;
prmt.b32 %r802, %r557, %r556, %r801;
prmt.b32 %r803, %r560, %r559, %r801;
shr.u32 %r804, %r794, 16;
prmt.b32 %r805, %r802, %r803, %r804;
prmt.b32 %r738, %r799, %r805, %r570;
prmt.b32 %r742, %r799, %r805, %r571;
ld.global.nc.u32 %r806, [%rd112+12];
and.b32 %r807, %r806, -2004318072;
shr.u32 %r808, %r807, 1;
or.b32 %r809, %r808, 839922192;
and.b32 %r810, %r806, 30583;
prmt.b32 %r811, %r557, %r556, %r810;
prmt.b32 %r812, %r560, %r559, %r810;
and.b32 %r813, %r809, 30292;
prmt.b32 %r814, %r811, %r812, %r813;
shr.u32 %r815, %r806, 16;
and.b32 %r816, %r815, 30583;
prmt.b32 %r817, %r557, %r556, %r816;
prmt.b32 %r818, %r560, %r559, %r816;
shr.u32 %r819, %r809, 16;
prmt.b32 %r820, %r817, %r818, %r819;
prmt.b32 %r746, %r814, %r820, %r570;
prmt.b32 %r750, %r814, %r820, %r571;

dp4a.s32.s32 %r737, %r738, %r25, %r536;


dp4a.s32.s32 %r741, %r742, %r26, %r737;


dp4a.s32.s32 %r745, %r746, %r27, %r741;


dp4a.s32.s32 %r749, %r750, %r28, %r745;

add.s64 %rd114, %rd34, %rd20;
ld.global.nc.u8 %rs62, [%rd114];
and.b16 %rs63, %rs62, 127;
setp.eq.s16 %p36, %rs63, 127;
selp.b16 %rs64, 0, %rs62, %p36;
and.b16 %rs57, %rs64, 255;

{cvt.rn.f16x2.e4m3x2 %r753, %rs57;}


cvt.u16.u32 %rs58, %r753;

{cvt.f32.f16 %f191, %rs58;}


div.approx.ftz.f32 %f198, %f191, %f167;

{.reg .f16 low,high;
mov.b32 {low,high},%r29;
cvt.f32.f16 %f192, low;}


mul.ftz.f32 %f199, %f198, %f192;
cvt.rn.f32.s32 %f200, %r749;
fma.rn.ftz.f32 %f201, %f199, %f200, %f197;
add.ftz.f32 %f33, %f201, %f33;
@%p22 bra $L__BB116_22;

add.s64 %rd115, %rd124, %rd16;
ld.global.u32 %r857, [%rd115+-12];
and.b32 %r858, %r857, -2004318072;
shr.u32 %r859, %r858, 1;
or.b32 %r860, %r859, 839922192;
and.b32 %r861, %r857, 30583;
mov.u32 %r862, 201852420;
mov.u32 %r863, 50462976;
prmt.b32 %r864, %r863, %r862, %r861;
mov.u32 %r865, -185009412;
mov.u32 %r866, -33620224;
prmt.b32 %r867, %r866, %r865, %r861;
and.b32 %r868, %r860, 30292;
prmt.b32 %r869, %r864, %r867, %r868;
shr.u32 %r870, %r857, 16;
and.b32 %r871, %r870, 30583;
prmt.b32 %r872, %r863, %r862, %r871;
prmt.b32 %r873, %r866, %r865, %r871;
shr.u32 %r874, %r860, 16;
prmt.b32 %r875, %r872, %r873, %r874;
mov.u32 %r876, 25632;
prmt.b32 %r822, %r869, %r875, %r876;
mov.u32 %r877, 30001;
prmt.b32 %r826, %r869, %r875, %r877;
ld.global.u32 %r878, [%rd115+-8];
and.b32 %r879, %r878, -2004318072;
shr.u32 %r880, %r879, 1;
or.b32 %r881, %r880, 839922192;
and.b32 %r882, %r878, 30583;
prmt.b32 %r883, %r863, %r862, %r882;
prmt.b32 %r884, %r866, %r865, %r882;
and.b32 %r885, %r881, 30292;
prmt.b32 %r886, %r883, %r884, %r885;
shr.u32 %r887, %r878, 16;
and.b32 %r888, %r887, 30583;
prmt.b32 %r889, %r863, %r862, %r888;
prmt.b32 %r890, %r866, %r865, %r888;
shr.u32 %r891, %r881, 16;
prmt.b32 %r892, %r889, %r890, %r891;
prmt.b32 %r830, %r886, %r892, %r876;
prmt.b32 %r834, %r886, %r892, %r877;
mov.u32 %r842, 0;

dp4a.s32.s32 %r821, %r822, %r20, %r842;


dp4a.s32.s32 %r825, %r826, %r21, %r821;


dp4a.s32.s32 %r829, %r830, %r22, %r825;


dp4a.s32.s32 %r833, %r834, %r23, %r829;

add.s64 %rd116, %rd124, %rd17;
ld.global.u8 %rs69, [%rd116];
and.b16 %rs70, %rs69, 127;
setp.eq.s16 %p38, %rs70, 127;
selp.b16 %rs65, 0, %rs69, %p38;

{cvt.rn.f16x2.e4m3x2 %r837, %rs65;}


cvt.u16.u32 %rs66, %r837;

{cvt.f32.f16 %f202, %rs66;}


mov.f32 %f206, 0f40000000;
div.approx.ftz.f32 %f207, %f202, %f206;

{.reg .f16 low,high;
mov.b32 {low,high},%r24;
cvt.f32.f16 %f203, low;}


mul.ftz.f32 %f208, %f207, %f203;
cvt.rn.f32.s32 %f209, %r833;
fma.rn.ftz.f32 %f210, %f208, %f209, 0f00000000;
ld.global.u32 %r893, [%rd115+-4];
and.b32 %r894, %r893, -2004318072;
shr.u32 %r895, %r894, 1;
or.b32 %r896, %r895, 839922192;
and.b32 %r897, %r893, 30583;
prmt.b32 %r898, %r863, %r862, %r897;
prmt.b32 %r899, %r866, %r865, %r897;
and.b32 %r900, %r896, 30292;
prmt.b32 %r901, %r898, %r899, %r900;
shr.u32 %r902, %r893, 16;
and.b32 %r903, %r902, 30583;
prmt.b32 %r904, %r863, %r862, %r903;
prmt.b32 %r905, %r866, %r865, %r903;
shr.u32 %r906, %r896, 16;
prmt.b32 %r907, %r904, %r905, %r906;
prmt.b32 %r840, %r901, %r907, %r876;
prmt.b32 %r844, %r901, %r907, %r877;
ld.global.u32 %r908, [%rd115];
and.b32 %r909, %r908, -2004318072;
shr.u32 %r910, %r909, 1;
or.b32 %r911, %r910, 839922192;
and.b32 %r912, %r908, 30583;
prmt.b32 %r913, %r863, %r862, %r912;
prmt.b32 %r914, %r866, %r865, %r912;
and.b32 %r915, %r911, 30292;
prmt.b32 %r916, %r913, %r914, %r915;
shr.u32 %r917, %r908, 16;
and.b32 %r918, %r917, 30583;
prmt.b32 %r919, %r863, %r862, %r918;
prmt.b32 %r920, %r866, %r865, %r918;
shr.u32 %r921, %r911, 16;
prmt.b32 %r922, %r919, %r920, %r921;
prmt.b32 %r848, %r916, %r922, %r876;
prmt.b32 %r852, %r916, %r922, %r877;

dp4a.s32.s32 %r839, %r840, %r25, %r842;


dp4a.s32.s32 %r843, %r844, %r26, %r839;


dp4a.s32.s32 %r847, %r848, %r27, %r843;


dp4a.s32.s32 %r851, %r852, %r28, %r847;

add.s64 %rd117, %rd124, %rd19;
ld.global.u8 %rs71, [%rd117];
and.b16 %rs72, %rs71, 127;
setp.eq.s16 %p39, %rs72, 127;
selp.b16 %rs67, 0, %rs71, %p39;

{cvt.rn.f16x2.e4m3x2 %r855, %rs67;}


cvt.u16.u32 %rs68, %r855;

{cvt.f32.f16 %f204, %rs68;}


div.approx.ftz.f32 %f211, %f204, %f206;

{.reg .f16 low,high;
mov.b32 {low,high},%r29;
cvt.f32.f16 %f205, low;}


mul.ftz.f32 %f212, %f211, %f205;
cvt.rn.f32.s32 %f213, %r851;
fma.rn.ftz.f32 %f214, %f212, %f213, %f210;
add.ftz.f32 %f37, %f214, %f37;

$L__BB116_22:
add.s32 %r1587, %r1587, 128;
add.s64 %rd124, %rd124, 2304;
add.s64 %rd123, %rd123, 2304;
add.s32 %r1588, %r1588, 64;
setp.lt.s32 %p40, %r1588, %r5;
@%p40 bra $L__BB116_14;

$L__BB116_23:
mov.u32 %r1574, %tid.y;
setp.eq.s32 %p41, %r1574, 0;
@%p41 bra $L__BB116_27;

mov.u32 %r1575, %tid.y;
setp.ne.s64 %p42, %rd42, 0;
add.s32 %r32, %r1575, -1;
shl.b32 %r923, %r1575, 9;
mov.u32 %r924, _ZZ13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r925, %r924, %r923;
shl.b32 %r926, %r2, 2;
add.s32 %r927, %r925, %r926;
st.shared.f32 [%r927+-512], %f36;
@%p42 bra $L__BB116_26;
bra.uni $L__BB116_25;

$L__BB116_26:
shl.b32 %r928, %r32, 9;
mov.u32 %r929, _ZZ13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r930, %r929, %r928;
add.s32 %r932, %r930, %r926;
st.shared.f32 [%r932], %f40;
add.s32 %r1571, %r927, -512;
st.shared.f32 [%r1571+128], %f35;
st.shared.f32 [%r932+128], %f39;
add.s32 %r1572, %r927, -512;
st.shared.f32 [%r1572+256], %f34;
st.shared.f32 [%r932+256], %f38;
add.s32 %r1573, %r927, -512;
st.shared.f32 [%r1573+384], %f33;
st.shared.f32 [%r932+384], %f37;
bra.uni $L__BB116_27;

$L__BB116_25:
add.s32 %r1568, %r927, -512;
st.shared.f32 [%r1568+128], %f35;
add.s32 %r1569, %r927, -512;
st.shared.f32 [%r1569+256], %f34;
add.s32 %r1570, %r927, -512;
st.shared.f32 [%r1570+384], %f33;

$L__BB116_27:
mov.u32 %r1576, %tid.y;
setp.ne.s32 %p342, %r1576, 0;
bar.sync 0;
@%p342 bra $L__BB116_82;
bra.uni $L__BB116_28;

$L__BB116_82:
ret;

$L__BB116_28:
ld.param.f32 %f1059, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
ld.param.u64 %rd122, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r1579, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r1578, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
mov.u32 %r1577, %ctaid.z;
setp.ne.s64 %p44, %rd42, 0;
mov.u32 %r934, %ctaid.y;
mad.lo.s32 %r935, %r934, %r1579, %r4;
mad.lo.s32 %r936, %r1577, %r1578, %r935;
neg.ftz.f32 %f41, %f1059;
shl.b32 %r937, %r2, 2;
mov.u32 %r938, _ZZ13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r34, %r938, %r937;
cvta.to.global.u64 %rd118, %rd122;
mul.wide.u32 %rd119, %r936, 4;
add.s64 %rd37, %rd118, %rd119;
@%p44 bra $L__BB116_37;
bra.uni $L__BB116_29;

$L__BB116_37:
ld.param.u32 %r1580, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
mov.u32 %r1012, _ZZ13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r36, %r1012, %r937;
setp.gt.s32 %p77, %r1580, 2;
@%p77 bra $L__BB116_48;

ld.param.u32 %r1583, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p80, %r1583, 1;
@%p80 bra $L__BB116_66;

ld.param.u32 %r1584, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p81, %r1584, 2;
@%p81 bra $L__BB116_40;
bra.uni $L__BB116_74;

$L__BB116_40:
ld.shared.f32 %f615, [%r34];
add.ftz.f32 %f616, %f615, %f36;
ld.shared.f32 %f617, [%r36];
add.ftz.f32 %f618, %f617, %f40;
ld.shared.f32 %f619, [%r34+512];
add.ftz.f32 %f620, %f619, %f616;
ld.shared.f32 %f621, [%r36+512];
add.ftz.f32 %f622, %f621, %f618;
ld.shared.f32 %f623, [%r34+1024];
add.ftz.f32 %f624, %f623, %f620;
ld.shared.f32 %f625, [%r36+1024];
add.ftz.f32 %f626, %f625, %f622;
mov.b32 %r1235, %f624;
mov.u32 %r1236, 31;
mov.u32 %r1237, 16;
mov.u32 %r1238, -1;
shfl.sync.bfly.b32 %r1239|%p186, %r1235, %r1237, %r1236, %r1238;
mov.b32 %f627, %r1239;
add.ftz.f32 %f628, %f624, %f627;
mov.b32 %r1240, %f628;
mov.u32 %r1241, 8;
shfl.sync.bfly.b32 %r1242|%p187, %r1240, %r1241, %r1236, %r1238;
mov.b32 %f629, %r1242;
add.ftz.f32 %f630, %f628, %f629;
mov.b32 %r1243, %f630;
mov.u32 %r1244, 4;
shfl.sync.bfly.b32 %r1245|%p188, %r1243, %r1244, %r1236, %r1238;
mov.b32 %f631, %r1245;
add.ftz.f32 %f632, %f630, %f631;
mov.b32 %r1246, %f632;
mov.u32 %r1247, 2;
shfl.sync.bfly.b32 %r1248|%p189, %r1246, %r1247, %r1236, %r1238;
mov.b32 %f633, %r1248;
add.ftz.f32 %f634, %f632, %f633;
mov.b32 %r1249, %f634;
mov.u32 %r1250, 1;
shfl.sync.bfly.b32 %r1251|%p190, %r1249, %r1250, %r1236, %r1238;
mov.b32 %f635, %r1251;
add.ftz.f32 %f62, %f634, %f635;
mov.b32 %r1252, %f626;
shfl.sync.bfly.b32 %r1253|%p191, %r1252, %r1237, %r1236, %r1238;
mov.b32 %f636, %r1253;
add.ftz.f32 %f637, %f626, %f636;
mov.b32 %r1254, %f637;
shfl.sync.bfly.b32 %r1255|%p192, %r1254, %r1241, %r1236, %r1238;
mov.b32 %f638, %r1255;
add.ftz.f32 %f639, %f637, %f638;
mov.b32 %r1256, %f639;
shfl.sync.bfly.b32 %r1257|%p193, %r1256, %r1244, %r1236, %r1238;
mov.b32 %f640, %r1257;
add.ftz.f32 %f641, %f639, %f640;
mov.b32 %r1258, %f641;
shfl.sync.bfly.b32 %r1259|%p194, %r1258, %r1247, %r1236, %r1238;
mov.b32 %f642, %r1259;
add.ftz.f32 %f643, %f641, %f642;
mov.b32 %r1260, %f643;
shfl.sync.bfly.b32 %r1261|%p195, %r1260, %r1250, %r1236, %r1238;
mov.b32 %f644, %r1261;
add.ftz.f32 %f63, %f643, %f644;
setp.ge.u32 %p196, %r4, %r52;
setp.ne.s32 %p197, %r2, 0;
or.pred %p198, %p197, %p196;
@%p198 bra $L__BB116_42;

fma.rn.ftz.f32 %f645, %f10, %f62, %f9;
fma.rn.ftz.f32 %f646, %f11, %f63, %f8;
mul.ftz.f32 %f647, %f646, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f648, %f647;
add.ftz.f32 %f649, %f648, 0f3F800000;
div.approx.ftz.f32 %f650, %f646, %f649;
mul.ftz.f32 %f651, %f645, %f650;
st.global.f32 [%rd37], %f651;

$L__BB116_42:
ld.shared.f32 %f652, [%r34+128];
add.ftz.f32 %f653, %f652, %f35;
ld.shared.f32 %f654, [%r36+128];
add.ftz.f32 %f655, %f654, %f39;
ld.shared.f32 %f656, [%r34+640];
add.ftz.f32 %f657, %f656, %f653;
ld.shared.f32 %f658, [%r36+640];
add.ftz.f32 %f659, %f658, %f655;
ld.shared.f32 %f660, [%r34+1152];
add.ftz.f32 %f661, %f660, %f657;
ld.shared.f32 %f662, [%r36+1152];
add.ftz.f32 %f663, %f662, %f659;
mov.b32 %r1262, %f661;
shfl.sync.bfly.b32 %r1266|%p199, %r1262, %r1237, %r1236, %r1238;
mov.b32 %f664, %r1266;
add.ftz.f32 %f665, %f661, %f664;
mov.b32 %r1267, %f665;
shfl.sync.bfly.b32 %r1269|%p200, %r1267, %r1241, %r1236, %r1238;
mov.b32 %f666, %r1269;
add.ftz.f32 %f667, %f665, %f666;
mov.b32 %r1270, %f667;
shfl.sync.bfly.b32 %r1272|%p201, %r1270, %r1244, %r1236, %r1238;
mov.b32 %f668, %r1272;
add.ftz.f32 %f669, %f667, %f668;
mov.b32 %r1273, %f669;
shfl.sync.bfly.b32 %r1275|%p202, %r1273, %r1247, %r1236, %r1238;
mov.b32 %f670, %r1275;
add.ftz.f32 %f671, %f669, %f670;
mov.b32 %r1276, %f671;
shfl.sync.bfly.b32 %r1278|%p203, %r1276, %r1250, %r1236, %r1238;
mov.b32 %f672, %r1278;
add.ftz.f32 %f64, %f671, %f672;
mov.b32 %r1279, %f663;
shfl.sync.bfly.b32 %r1280|%p204, %r1279, %r1237, %r1236, %r1238;
mov.b32 %f673, %r1280;
add.ftz.f32 %f674, %f663, %f673;
mov.b32 %r1281, %f674;
shfl.sync.bfly.b32 %r1282|%p205, %r1281, %r1241, %r1236, %r1238;
mov.b32 %f675, %r1282;
add.ftz.f32 %f676, %f674, %f675;
mov.b32 %r1283, %f676;
shfl.sync.bfly.b32 %r1284|%p206, %r1283, %r1244, %r1236, %r1238;
mov.b32 %f677, %r1284;
add.ftz.f32 %f678, %f676, %f677;
mov.b32 %r1285, %f678;
shfl.sync.bfly.b32 %r1286|%p207, %r1285, %r1247, %r1236, %r1238;
mov.b32 %f679, %r1286;
add.ftz.f32 %f680, %f678, %f679;
mov.b32 %r1287, %f680;
shfl.sync.bfly.b32 %r1288|%p208, %r1287, %r1250, %r1236, %r1238;
mov.b32 %f681, %r1288;
add.ftz.f32 %f65, %f680, %f681;
add.s32 %r1289, %r4, 1;
setp.ge.u32 %p209, %r1289, %r52;
setp.ne.s32 %p210, %r2, 1;
or.pred %p211, %p210, %p209;
@%p211 bra $L__BB116_44;

fma.rn.ftz.f32 %f682, %f10, %f64, %f9;
fma.rn.ftz.f32 %f683, %f11, %f65, %f8;
mul.ftz.f32 %f684, %f683, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f685, %f684;
add.ftz.f32 %f686, %f685, 0f3F800000;
div.approx.ftz.f32 %f687, %f683, %f686;
mul.ftz.f32 %f688, %f682, %f687;
st.global.f32 [%rd37+4], %f688;

$L__BB116_44:
ld.shared.f32 %f689, [%r34+256];
add.ftz.f32 %f690, %f689, %f34;
ld.shared.f32 %f691, [%r36+256];
add.ftz.f32 %f692, %f691, %f38;
ld.shared.f32 %f693, [%r34+768];
add.ftz.f32 %f694, %f693, %f690;
ld.shared.f32 %f695, [%r36+768];
add.ftz.f32 %f696, %f695, %f692;
ld.shared.f32 %f697, [%r34+1280];
add.ftz.f32 %f698, %f697, %f694;
ld.shared.f32 %f699, [%r36+1280];
add.ftz.f32 %f700, %f699, %f696;
mov.b32 %r1290, %f698;
mov.u32 %r1291, 31;
mov.u32 %r1292, 16;
mov.u32 %r1293, -1;
shfl.sync.bfly.b32 %r1294|%p212, %r1290, %r1292, %r1291, %r1293;
mov.b32 %f701, %r1294;
add.ftz.f32 %f702, %f698, %f701;
mov.b32 %r1295, %f702;
mov.u32 %r1296, 8;
shfl.sync.bfly.b32 %r1297|%p213, %r1295, %r1296, %r1291, %r1293;
mov.b32 %f703, %r1297;
add.ftz.f32 %f704, %f702, %f703;
mov.b32 %r1298, %f704;
mov.u32 %r1299, 4;
shfl.sync.bfly.b32 %r1300|%p214, %r1298, %r1299, %r1291, %r1293;
mov.b32 %f705, %r1300;
add.ftz.f32 %f706, %f704, %f705;
mov.b32 %r1301, %f706;
mov.u32 %r1302, 2;
shfl.sync.bfly.b32 %r1303|%p215, %r1301, %r1302, %r1291, %r1293;
mov.b32 %f707, %r1303;
add.ftz.f32 %f708, %f706, %f707;
mov.b32 %r1304, %f708;
mov.u32 %r1305, 1;
shfl.sync.bfly.b32 %r1306|%p216, %r1304, %r1305, %r1291, %r1293;
mov.b32 %f709, %r1306;
add.ftz.f32 %f66, %f708, %f709;
mov.b32 %r1307, %f700;
shfl.sync.bfly.b32 %r1308|%p217, %r1307, %r1292, %r1291, %r1293;
mov.b32 %f710, %r1308;
add.ftz.f32 %f711, %f700, %f710;
mov.b32 %r1309, %f711;
shfl.sync.bfly.b32 %r1310|%p218, %r1309, %r1296, %r1291, %r1293;
mov.b32 %f712, %r1310;
add.ftz.f32 %f713, %f711, %f712;
mov.b32 %r1311, %f713;
shfl.sync.bfly.b32 %r1312|%p219, %r1311, %r1299, %r1291, %r1293;
mov.b32 %f714, %r1312;
add.ftz.f32 %f715, %f713, %f714;
mov.b32 %r1313, %f715;
shfl.sync.bfly.b32 %r1314|%p220, %r1313, %r1302, %r1291, %r1293;
mov.b32 %f716, %r1314;
add.ftz.f32 %f717, %f715, %f716;
mov.b32 %r1315, %f717;
shfl.sync.bfly.b32 %r1316|%p221, %r1315, %r1305, %r1291, %r1293;
mov.b32 %f718, %r1316;
add.ftz.f32 %f67, %f717, %f718;
add.s32 %r1317, %r4, 2;
setp.ge.u32 %p222, %r1317, %r52;
setp.ne.s32 %p223, %r2, 2;
or.pred %p224, %p223, %p222;
@%p224 bra $L__BB116_46;

fma.rn.ftz.f32 %f719, %f10, %f66, %f9;
fma.rn.ftz.f32 %f720, %f11, %f67, %f8;
mul.ftz.f32 %f721, %f720, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f722, %f721;
add.ftz.f32 %f723, %f722, 0f3F800000;
div.approx.ftz.f32 %f724, %f720, %f723;
mul.ftz.f32 %f725, %f719, %f724;
st.global.f32 [%rd37+8], %f725;

$L__BB116_46:
ld.shared.f32 %f726, [%r34+384];
add.ftz.f32 %f727, %f726, %f33;
ld.shared.f32 %f728, [%r36+384];
add.ftz.f32 %f729, %f728, %f37;
ld.shared.f32 %f730, [%r34+896];
add.ftz.f32 %f731, %f730, %f727;
ld.shared.f32 %f732, [%r36+896];
add.ftz.f32 %f733, %f732, %f729;
ld.shared.f32 %f734, [%r34+1408];
add.ftz.f32 %f735, %f734, %f731;
ld.shared.f32 %f736, [%r36+1408];
add.ftz.f32 %f737, %f736, %f733;
mov.b32 %r1318, %f735;
shfl.sync.bfly.b32 %r1322|%p225, %r1318, %r1292, %r1291, %r1293;
mov.b32 %f738, %r1322;
add.ftz.f32 %f739, %f735, %f738;
mov.b32 %r1323, %f739;
shfl.sync.bfly.b32 %r1325|%p226, %r1323, %r1296, %r1291, %r1293;
mov.b32 %f740, %r1325;
add.ftz.f32 %f741, %f739, %f740;
mov.b32 %r1326, %f741;
shfl.sync.bfly.b32 %r1328|%p227, %r1326, %r1299, %r1291, %r1293;
mov.b32 %f742, %r1328;
add.ftz.f32 %f743, %f741, %f742;
mov.b32 %r1329, %f743;
shfl.sync.bfly.b32 %r1331|%p228, %r1329, %r1302, %r1291, %r1293;
mov.b32 %f744, %r1331;
add.ftz.f32 %f745, %f743, %f744;
mov.b32 %r1332, %f745;
shfl.sync.bfly.b32 %r1334|%p229, %r1332, %r1305, %r1291, %r1293;
mov.b32 %f746, %r1334;
add.ftz.f32 %f68, %f745, %f746;
mov.b32 %r1335, %f737;
shfl.sync.bfly.b32 %r1336|%p230, %r1335, %r1292, %r1291, %r1293;
mov.b32 %f747, %r1336;
add.ftz.f32 %f748, %f737, %f747;
mov.b32 %r1337, %f748;
shfl.sync.bfly.b32 %r1338|%p231, %r1337, %r1296, %r1291, %r1293;
mov.b32 %f749, %r1338;
add.ftz.f32 %f750, %f748, %f749;
mov.b32 %r1339, %f750;
shfl.sync.bfly.b32 %r1340|%p232, %r1339, %r1299, %r1291, %r1293;
mov.b32 %f751, %r1340;
add.ftz.f32 %f752, %f750, %f751;
mov.b32 %r1341, %f752;
shfl.sync.bfly.b32 %r1342|%p233, %r1341, %r1302, %r1291, %r1293;
mov.b32 %f753, %r1342;
add.ftz.f32 %f754, %f752, %f753;
mov.b32 %r1343, %f754;
shfl.sync.bfly.b32 %r1344|%p234, %r1343, %r1305, %r1291, %r1293;
mov.b32 %f755, %r1344;
add.ftz.f32 %f69, %f754, %f755;
add.s32 %r1345, %r4, 3;
setp.ge.u32 %p235, %r1345, %r52;
setp.ne.s32 %p236, %r2, 3;
or.pred %p237, %p236, %p235;
@%p237 bra $L__BB116_82;

fma.rn.ftz.f32 %f756, %f10, %f68, %f9;
fma.rn.ftz.f32 %f757, %f11, %f69, %f8;
mul.ftz.f32 %f758, %f757, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f759, %f758;
add.ftz.f32 %f760, %f759, 0f3F800000;
div.approx.ftz.f32 %f761, %f757, %f760;
mul.ftz.f32 %f762, %f756, %f761;
st.global.f32 [%rd37+12], %f762;
bra.uni $L__BB116_82;

$L__BB116_29:
ld.shared.f32 %f215, [%r34];
add.ftz.f32 %f216, %f215, %f36;
ld.shared.f32 %f217, [%r34+512];
add.ftz.f32 %f218, %f217, %f216;
ld.shared.f32 %f219, [%r34+1024];
add.ftz.f32 %f220, %f219, %f218;
mov.b32 %r939, %f220;
mov.u32 %r940, 31;
mov.u32 %r941, 16;
mov.u32 %r942, -1;
shfl.sync.bfly.b32 %r943|%p45, %r939, %r941, %r940, %r942;
mov.b32 %f221, %r943;
add.ftz.f32 %f222, %f220, %f221;
mov.b32 %r944, %f222;
mov.u32 %r945, 8;
shfl.sync.bfly.b32 %r946|%p46, %r944, %r945, %r940, %r942;
mov.b32 %f223, %r946;
add.ftz.f32 %f224, %f222, %f223;
mov.b32 %r947, %f224;
mov.u32 %r948, 4;
shfl.sync.bfly.b32 %r949|%p47, %r947, %r948, %r940, %r942;
mov.b32 %f225, %r949;
add.ftz.f32 %f226, %f224, %f225;
mov.b32 %r950, %f226;
mov.u32 %r951, 2;
shfl.sync.bfly.b32 %r952|%p48, %r950, %r951, %r940, %r942;
mov.b32 %f227, %r952;
add.ftz.f32 %f228, %f226, %f227;
mov.b32 %r953, %f228;
mov.u32 %r954, 1;
shfl.sync.bfly.b32 %r955|%p49, %r953, %r954, %r940, %r942;
mov.b32 %f229, %r955;
add.ftz.f32 %f42, %f228, %f229;
setp.ge.u32 %p50, %r4, %r52;
setp.ne.s32 %p51, %r2, 0;
or.pred %p52, %p51, %p50;
@%p52 bra $L__BB116_31;

fma.rn.ftz.f32 %f230, %f10, %f42, %f9;
st.global.f32 [%rd37], %f230;

$L__BB116_31:
ld.shared.f32 %f231, [%r34+128];
add.ftz.f32 %f232, %f231, %f35;
ld.shared.f32 %f233, [%r34+640];
add.ftz.f32 %f234, %f233, %f232;
ld.shared.f32 %f235, [%r34+1152];
add.ftz.f32 %f236, %f235, %f234;
mov.b32 %r956, %f236;
shfl.sync.bfly.b32 %r960|%p53, %r956, %r941, %r940, %r942;
mov.b32 %f237, %r960;
add.ftz.f32 %f238, %f236, %f237;
mov.b32 %r961, %f238;
shfl.sync.bfly.b32 %r963|%p54, %r961, %r945, %r940, %r942;
mov.b32 %f239, %r963;
add.ftz.f32 %f240, %f238, %f239;
mov.b32 %r964, %f240;
shfl.sync.bfly.b32 %r966|%p55, %r964, %r948, %r940, %r942;
mov.b32 %f241, %r966;
add.ftz.f32 %f242, %f240, %f241;
mov.b32 %r967, %f242;
shfl.sync.bfly.b32 %r969|%p56, %r967, %r951, %r940, %r942;
mov.b32 %f243, %r969;
add.ftz.f32 %f244, %f242, %f243;
mov.b32 %r970, %f244;
shfl.sync.bfly.b32 %r972|%p57, %r970, %r954, %r940, %r942;
mov.b32 %f245, %r972;
add.ftz.f32 %f43, %f244, %f245;
add.s32 %r973, %r4, 1;
setp.ge.u32 %p58, %r973, %r52;
setp.ne.s32 %p59, %r2, 1;
or.pred %p60, %p59, %p58;
@%p60 bra $L__BB116_33;

fma.rn.ftz.f32 %f246, %f10, %f43, %f9;
st.global.f32 [%rd37+4], %f246;

$L__BB116_33:
ld.shared.f32 %f247, [%r34+256];
add.ftz.f32 %f248, %f247, %f34;
ld.shared.f32 %f249, [%r34+768];
add.ftz.f32 %f250, %f249, %f248;
ld.shared.f32 %f251, [%r34+1280];
add.ftz.f32 %f252, %f251, %f250;
mov.b32 %r974, %f252;
mov.u32 %r975, 31;
mov.u32 %r976, 16;
mov.u32 %r977, -1;
shfl.sync.bfly.b32 %r978|%p61, %r974, %r976, %r975, %r977;
mov.b32 %f253, %r978;
add.ftz.f32 %f254, %f252, %f253;
mov.b32 %r979, %f254;
mov.u32 %r980, 8;
shfl.sync.bfly.b32 %r981|%p62, %r979, %r980, %r975, %r977;
mov.b32 %f255, %r981;
add.ftz.f32 %f256, %f254, %f255;
mov.b32 %r982, %f256;
mov.u32 %r983, 4;
shfl.sync.bfly.b32 %r984|%p63, %r982, %r983, %r975, %r977;
mov.b32 %f257, %r984;
add.ftz.f32 %f258, %f256, %f257;
mov.b32 %r985, %f258;
mov.u32 %r986, 2;
shfl.sync.bfly.b32 %r987|%p64, %r985, %r986, %r975, %r977;
mov.b32 %f259, %r987;
add.ftz.f32 %f260, %f258, %f259;
mov.b32 %r988, %f260;
mov.u32 %r989, 1;
shfl.sync.bfly.b32 %r990|%p65, %r988, %r989, %r975, %r977;
mov.b32 %f261, %r990;
add.ftz.f32 %f44, %f260, %f261;
add.s32 %r991, %r4, 2;
setp.ge.u32 %p66, %r991, %r52;
setp.ne.s32 %p67, %r2, 2;
or.pred %p68, %p67, %p66;
@%p68 bra $L__BB116_35;

fma.rn.ftz.f32 %f262, %f10, %f44, %f9;
st.global.f32 [%rd37+8], %f262;

$L__BB116_35:
ld.shared.f32 %f263, [%r34+384];
add.ftz.f32 %f264, %f263, %f33;
ld.shared.f32 %f265, [%r34+896];
add.ftz.f32 %f266, %f265, %f264;
ld.shared.f32 %f267, [%r34+1408];
add.ftz.f32 %f268, %f267, %f266;
mov.b32 %r992, %f268;
shfl.sync.bfly.b32 %r996|%p69, %r992, %r976, %r975, %r977;
mov.b32 %f269, %r996;
add.ftz.f32 %f270, %f268, %f269;
mov.b32 %r997, %f270;
shfl.sync.bfly.b32 %r999|%p70, %r997, %r980, %r975, %r977;
mov.b32 %f271, %r999;
add.ftz.f32 %f272, %f270, %f271;
mov.b32 %r1000, %f272;
shfl.sync.bfly.b32 %r1002|%p71, %r1000, %r983, %r975, %r977;
mov.b32 %f273, %r1002;
add.ftz.f32 %f274, %f272, %f273;
mov.b32 %r1003, %f274;
shfl.sync.bfly.b32 %r1005|%p72, %r1003, %r986, %r975, %r977;
mov.b32 %f275, %r1005;
add.ftz.f32 %f276, %f274, %f275;
mov.b32 %r1006, %f276;
shfl.sync.bfly.b32 %r1008|%p73, %r1006, %r989, %r975, %r977;
mov.b32 %f277, %r1008;
add.ftz.f32 %f45, %f276, %f277;
add.s32 %r1009, %r4, 3;
setp.ge.u32 %p74, %r1009, %r52;
setp.ne.s32 %p75, %r2, 3;
or.pred %p76, %p75, %p74;
@%p76 bra $L__BB116_82;

fma.rn.ftz.f32 %f278, %f10, %f45, %f9;
st.global.f32 [%rd37+12], %f278;
bra.uni $L__BB116_82;

$L__BB116_48:
ld.param.u32 %r1581, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p78, %r1581, 3;
@%p78 bra $L__BB116_58;

ld.param.u32 %r1582, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.ne.s32 %p79, %r1582, 6;
@%p79 bra $L__BB116_74;

ld.shared.f32 %f279, [%r34];
add.ftz.f32 %f280, %f279, %f36;
ld.shared.f32 %f281, [%r36];
add.ftz.f32 %f282, %f281, %f40;
ld.shared.f32 %f283, [%r34+512];
add.ftz.f32 %f284, %f283, %f280;
ld.shared.f32 %f285, [%r36+512];
add.ftz.f32 %f286, %f285, %f282;
ld.shared.f32 %f287, [%r34+1024];
add.ftz.f32 %f288, %f287, %f284;
ld.shared.f32 %f289, [%r36+1024];
add.ftz.f32 %f290, %f289, %f286;
mov.b32 %r1013, %f288;
mov.u32 %r1014, 31;
mov.u32 %r1015, 16;
mov.u32 %r1016, -1;
shfl.sync.bfly.b32 %r1017|%p82, %r1013, %r1015, %r1014, %r1016;
mov.b32 %f291, %r1017;
add.ftz.f32 %f292, %f288, %f291;
mov.b32 %r1018, %f292;
mov.u32 %r1019, 8;
shfl.sync.bfly.b32 %r1020|%p83, %r1018, %r1019, %r1014, %r1016;
mov.b32 %f293, %r1020;
add.ftz.f32 %f294, %f292, %f293;
mov.b32 %r1021, %f294;
mov.u32 %r1022, 4;
shfl.sync.bfly.b32 %r1023|%p84, %r1021, %r1022, %r1014, %r1016;
mov.b32 %f295, %r1023;
add.ftz.f32 %f296, %f294, %f295;
mov.b32 %r1024, %f296;
mov.u32 %r1025, 2;
shfl.sync.bfly.b32 %r1026|%p85, %r1024, %r1025, %r1014, %r1016;
mov.b32 %f297, %r1026;
add.ftz.f32 %f298, %f296, %f297;
mov.b32 %r1027, %f298;
mov.u32 %r1028, 1;
shfl.sync.bfly.b32 %r1029|%p86, %r1027, %r1028, %r1014, %r1016;
mov.b32 %f299, %r1029;
add.ftz.f32 %f46, %f298, %f299;
mov.b32 %r1030, %f290;
shfl.sync.bfly.b32 %r1031|%p87, %r1030, %r1015, %r1014, %r1016;
mov.b32 %f300, %r1031;
add.ftz.f32 %f301, %f290, %f300;
mov.b32 %r1032, %f301;
shfl.sync.bfly.b32 %r1033|%p88, %r1032, %r1019, %r1014, %r1016;
mov.b32 %f302, %r1033;
add.ftz.f32 %f303, %f301, %f302;
mov.b32 %r1034, %f303;
shfl.sync.bfly.b32 %r1035|%p89, %r1034, %r1022, %r1014, %r1016;
mov.b32 %f304, %r1035;
add.ftz.f32 %f305, %f303, %f304;
mov.b32 %r1036, %f305;
shfl.sync.bfly.b32 %r1037|%p90, %r1036, %r1025, %r1014, %r1016;
mov.b32 %f306, %r1037;
add.ftz.f32 %f307, %f305, %f306;
mov.b32 %r1038, %f307;
shfl.sync.bfly.b32 %r1039|%p91, %r1038, %r1028, %r1014, %r1016;
mov.b32 %f308, %r1039;
add.ftz.f32 %f47, %f307, %f308;
setp.ge.u32 %p92, %r4, %r52;
setp.ne.s32 %p93, %r2, 0;
or.pred %p94, %p93, %p92;
@%p94 bra $L__BB116_52;

ld.param.f32 %f1063, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
fma.rn.ftz.f32 %f309, %f10, %f46, %f9;
fma.rn.ftz.f32 %f310, %f11, %f47, %f8;
min.ftz.f32 %f311, %f310, %f1063;
min.ftz.f32 %f312, %f309, %f1063;
max.ftz.f32 %f313, %f312, %f41;
mul.ftz.f32 %f314, %f311, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f315, %f314;
add.ftz.f32 %f316, %f315, 0f3F800000;
div.approx.ftz.f32 %f317, %f311, %f316;
mul.ftz.f32 %f318, %f313, %f317;
st.global.f32 [%rd37], %f318;

$L__BB116_52:
ld.shared.f32 %f319, [%r34+128];
add.ftz.f32 %f320, %f319, %f35;
ld.shared.f32 %f321, [%r36+128];
add.ftz.f32 %f322, %f321, %f39;
ld.shared.f32 %f323, [%r34+640];
add.ftz.f32 %f324, %f323, %f320;
ld.shared.f32 %f325, [%r36+640];
add.ftz.f32 %f326, %f325, %f322;
ld.shared.f32 %f327, [%r34+1152];
add.ftz.f32 %f328, %f327, %f324;
ld.shared.f32 %f329, [%r36+1152];
add.ftz.f32 %f330, %f329, %f326;
mov.b32 %r1040, %f328;
shfl.sync.bfly.b32 %r1044|%p95, %r1040, %r1015, %r1014, %r1016;
mov.b32 %f331, %r1044;
add.ftz.f32 %f332, %f328, %f331;
mov.b32 %r1045, %f332;
shfl.sync.bfly.b32 %r1047|%p96, %r1045, %r1019, %r1014, %r1016;
mov.b32 %f333, %r1047;
add.ftz.f32 %f334, %f332, %f333;
mov.b32 %r1048, %f334;
shfl.sync.bfly.b32 %r1050|%p97, %r1048, %r1022, %r1014, %r1016;
mov.b32 %f335, %r1050;
add.ftz.f32 %f336, %f334, %f335;
mov.b32 %r1051, %f336;
shfl.sync.bfly.b32 %r1053|%p98, %r1051, %r1025, %r1014, %r1016;
mov.b32 %f337, %r1053;
add.ftz.f32 %f338, %f336, %f337;
mov.b32 %r1054, %f338;
shfl.sync.bfly.b32 %r1056|%p99, %r1054, %r1028, %r1014, %r1016;
mov.b32 %f339, %r1056;
add.ftz.f32 %f48, %f338, %f339;
mov.b32 %r1057, %f330;
shfl.sync.bfly.b32 %r1058|%p100, %r1057, %r1015, %r1014, %r1016;
mov.b32 %f340, %r1058;
add.ftz.f32 %f341, %f330, %f340;
mov.b32 %r1059, %f341;
shfl.sync.bfly.b32 %r1060|%p101, %r1059, %r1019, %r1014, %r1016;
mov.b32 %f342, %r1060;
add.ftz.f32 %f343, %f341, %f342;
mov.b32 %r1061, %f343;
shfl.sync.bfly.b32 %r1062|%p102, %r1061, %r1022, %r1014, %r1016;
mov.b32 %f344, %r1062;
add.ftz.f32 %f345, %f343, %f344;
mov.b32 %r1063, %f345;
shfl.sync.bfly.b32 %r1064|%p103, %r1063, %r1025, %r1014, %r1016;
mov.b32 %f346, %r1064;
add.ftz.f32 %f347, %f345, %f346;
mov.b32 %r1065, %f347;
shfl.sync.bfly.b32 %r1066|%p104, %r1065, %r1028, %r1014, %r1016;
mov.b32 %f348, %r1066;
add.ftz.f32 %f49, %f347, %f348;
add.s32 %r1067, %r4, 1;
setp.ge.u32 %p105, %r1067, %r52;
setp.ne.s32 %p106, %r2, 1;
or.pred %p107, %p106, %p105;
@%p107 bra $L__BB116_54;

ld.param.f32 %f1062, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
fma.rn.ftz.f32 %f349, %f10, %f48, %f9;
fma.rn.ftz.f32 %f350, %f11, %f49, %f8;
min.ftz.f32 %f351, %f350, %f1062;
min.ftz.f32 %f352, %f349, %f1062;
max.ftz.f32 %f353, %f352, %f41;
mul.ftz.f32 %f354, %f351, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f355, %f354;
add.ftz.f32 %f356, %f355, 0f3F800000;
div.approx.ftz.f32 %f357, %f351, %f356;
mul.ftz.f32 %f358, %f353, %f357;
st.global.f32 [%rd37+4], %f358;

$L__BB116_54:
ld.shared.f32 %f359, [%r34+256];
add.ftz.f32 %f360, %f359, %f34;
ld.shared.f32 %f361, [%r36+256];
add.ftz.f32 %f362, %f361, %f38;
ld.shared.f32 %f363, [%r34+768];
add.ftz.f32 %f364, %f363, %f360;
ld.shared.f32 %f365, [%r36+768];
add.ftz.f32 %f366, %f365, %f362;
ld.shared.f32 %f367, [%r34+1280];
add.ftz.f32 %f368, %f367, %f364;
ld.shared.f32 %f369, [%r36+1280];
add.ftz.f32 %f370, %f369, %f366;
mov.b32 %r1068, %f368;
mov.u32 %r1069, 31;
mov.u32 %r1070, 16;
mov.u32 %r1071, -1;
shfl.sync.bfly.b32 %r1072|%p108, %r1068, %r1070, %r1069, %r1071;
mov.b32 %f371, %r1072;
add.ftz.f32 %f372, %f368, %f371;
mov.b32 %r1073, %f372;
mov.u32 %r1074, 8;
shfl.sync.bfly.b32 %r1075|%p109, %r1073, %r1074, %r1069, %r1071;
mov.b32 %f373, %r1075;
add.ftz.f32 %f374, %f372, %f373;
mov.b32 %r1076, %f374;
mov.u32 %r1077, 4;
shfl.sync.bfly.b32 %r1078|%p110, %r1076, %r1077, %r1069, %r1071;
mov.b32 %f375, %r1078;
add.ftz.f32 %f376, %f374, %f375;
mov.b32 %r1079, %f376;
mov.u32 %r1080, 2;
shfl.sync.bfly.b32 %r1081|%p111, %r1079, %r1080, %r1069, %r1071;
mov.b32 %f377, %r1081;
add.ftz.f32 %f378, %f376, %f377;
mov.b32 %r1082, %f378;
mov.u32 %r1083, 1;
shfl.sync.bfly.b32 %r1084|%p112, %r1082, %r1083, %r1069, %r1071;
mov.b32 %f379, %r1084;
add.ftz.f32 %f50, %f378, %f379;
mov.b32 %r1085, %f370;
shfl.sync.bfly.b32 %r1086|%p113, %r1085, %r1070, %r1069, %r1071;
mov.b32 %f380, %r1086;
add.ftz.f32 %f381, %f370, %f380;
mov.b32 %r1087, %f381;
shfl.sync.bfly.b32 %r1088|%p114, %r1087, %r1074, %r1069, %r1071;
mov.b32 %f382, %r1088;
add.ftz.f32 %f383, %f381, %f382;
mov.b32 %r1089, %f383;
shfl.sync.bfly.b32 %r1090|%p115, %r1089, %r1077, %r1069, %r1071;
mov.b32 %f384, %r1090;
add.ftz.f32 %f385, %f383, %f384;
mov.b32 %r1091, %f385;
shfl.sync.bfly.b32 %r1092|%p116, %r1091, %r1080, %r1069, %r1071;
mov.b32 %f386, %r1092;
add.ftz.f32 %f387, %f385, %f386;
mov.b32 %r1093, %f387;
shfl.sync.bfly.b32 %r1094|%p117, %r1093, %r1083, %r1069, %r1071;
mov.b32 %f388, %r1094;
add.ftz.f32 %f51, %f387, %f388;
add.s32 %r1095, %r4, 2;
setp.ge.u32 %p118, %r1095, %r52;
setp.ne.s32 %p119, %r2, 2;
or.pred %p120, %p119, %p118;
@%p120 bra $L__BB116_56;

ld.param.f32 %f1061, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
fma.rn.ftz.f32 %f389, %f10, %f50, %f9;
fma.rn.ftz.f32 %f390, %f11, %f51, %f8;
min.ftz.f32 %f391, %f390, %f1061;
min.ftz.f32 %f392, %f389, %f1061;
max.ftz.f32 %f393, %f392, %f41;
mul.ftz.f32 %f394, %f391, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f395, %f394;
add.ftz.f32 %f396, %f395, 0f3F800000;
div.approx.ftz.f32 %f397, %f391, %f396;
mul.ftz.f32 %f398, %f393, %f397;
st.global.f32 [%rd37+8], %f398;

$L__BB116_56:
ld.shared.f32 %f399, [%r34+384];
add.ftz.f32 %f400, %f399, %f33;
ld.shared.f32 %f401, [%r36+384];
add.ftz.f32 %f402, %f401, %f37;
ld.shared.f32 %f403, [%r34+896];
add.ftz.f32 %f404, %f403, %f400;
ld.shared.f32 %f405, [%r36+896];
add.ftz.f32 %f406, %f405, %f402;
ld.shared.f32 %f407, [%r34+1408];
add.ftz.f32 %f408, %f407, %f404;
ld.shared.f32 %f409, [%r36+1408];
add.ftz.f32 %f410, %f409, %f406;
mov.b32 %r1096, %f408;
shfl.sync.bfly.b32 %r1100|%p121, %r1096, %r1070, %r1069, %r1071;
mov.b32 %f411, %r1100;
add.ftz.f32 %f412, %f408, %f411;
mov.b32 %r1101, %f412;
shfl.sync.bfly.b32 %r1103|%p122, %r1101, %r1074, %r1069, %r1071;
mov.b32 %f413, %r1103;
add.ftz.f32 %f414, %f412, %f413;
mov.b32 %r1104, %f414;
shfl.sync.bfly.b32 %r1106|%p123, %r1104, %r1077, %r1069, %r1071;
mov.b32 %f415, %r1106;
add.ftz.f32 %f416, %f414, %f415;
mov.b32 %r1107, %f416;
shfl.sync.bfly.b32 %r1109|%p124, %r1107, %r1080, %r1069, %r1071;
mov.b32 %f417, %r1109;
add.ftz.f32 %f418, %f416, %f417;
mov.b32 %r1110, %f418;
shfl.sync.bfly.b32 %r1112|%p125, %r1110, %r1083, %r1069, %r1071;
mov.b32 %f419, %r1112;
add.ftz.f32 %f52, %f418, %f419;
mov.b32 %r1113, %f410;
shfl.sync.bfly.b32 %r1114|%p126, %r1113, %r1070, %r1069, %r1071;
mov.b32 %f420, %r1114;
add.ftz.f32 %f421, %f410, %f420;
mov.b32 %r1115, %f421;
shfl.sync.bfly.b32 %r1116|%p127, %r1115, %r1074, %r1069, %r1071;
mov.b32 %f422, %r1116;
add.ftz.f32 %f423, %f421, %f422;
mov.b32 %r1117, %f423;
shfl.sync.bfly.b32 %r1118|%p128, %r1117, %r1077, %r1069, %r1071;
mov.b32 %f424, %r1118;
add.ftz.f32 %f425, %f423, %f424;
mov.b32 %r1119, %f425;
shfl.sync.bfly.b32 %r1120|%p129, %r1119, %r1080, %r1069, %r1071;
mov.b32 %f426, %r1120;
add.ftz.f32 %f427, %f425, %f426;
mov.b32 %r1121, %f427;
shfl.sync.bfly.b32 %r1122|%p130, %r1121, %r1083, %r1069, %r1071;
mov.b32 %f428, %r1122;
add.ftz.f32 %f53, %f427, %f428;
add.s32 %r1123, %r4, 3;
setp.ge.u32 %p131, %r1123, %r52;
setp.ne.s32 %p132, %r2, 3;
or.pred %p133, %p132, %p131;
@%p133 bra $L__BB116_82;

ld.param.f32 %f1060, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
fma.rn.ftz.f32 %f429, %f10, %f52, %f9;
fma.rn.ftz.f32 %f430, %f11, %f53, %f8;
min.ftz.f32 %f431, %f430, %f1060;
min.ftz.f32 %f432, %f429, %f1060;
max.ftz.f32 %f433, %f432, %f41;
mul.ftz.f32 %f434, %f431, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f435, %f434;
add.ftz.f32 %f436, %f435, 0f3F800000;
div.approx.ftz.f32 %f437, %f431, %f436;
mul.ftz.f32 %f438, %f433, %f437;
st.global.f32 [%rd37+12], %f438;
bra.uni $L__BB116_82;

$L__BB116_74:
ld.shared.f32 %f927, [%r34];
add.ftz.f32 %f928, %f927, %f36;
ld.shared.f32 %f929, [%r36];
add.ftz.f32 %f930, %f929, %f40;
ld.shared.f32 %f931, [%r34+512];
add.ftz.f32 %f932, %f931, %f928;
ld.shared.f32 %f933, [%r36+512];
add.ftz.f32 %f934, %f933, %f930;
ld.shared.f32 %f935, [%r34+1024];
add.ftz.f32 %f936, %f935, %f932;
ld.shared.f32 %f937, [%r36+1024];
add.ftz.f32 %f938, %f937, %f934;
mov.b32 %r1457, %f936;
mov.u32 %r1458, 31;
mov.u32 %r1459, 16;
mov.u32 %r1460, -1;
shfl.sync.bfly.b32 %r1461|%p290, %r1457, %r1459, %r1458, %r1460;
mov.b32 %f939, %r1461;
add.ftz.f32 %f940, %f936, %f939;
mov.b32 %r1462, %f940;
mov.u32 %r1463, 8;
shfl.sync.bfly.b32 %r1464|%p291, %r1462, %r1463, %r1458, %r1460;
mov.b32 %f941, %r1464;
add.ftz.f32 %f942, %f940, %f941;
mov.b32 %r1465, %f942;
mov.u32 %r1466, 4;
shfl.sync.bfly.b32 %r1467|%p292, %r1465, %r1466, %r1458, %r1460;
mov.b32 %f943, %r1467;
add.ftz.f32 %f944, %f942, %f943;
mov.b32 %r1468, %f944;
mov.u32 %r1469, 2;
shfl.sync.bfly.b32 %r1470|%p293, %r1468, %r1469, %r1458, %r1460;
mov.b32 %f945, %r1470;
add.ftz.f32 %f946, %f944, %f945;
mov.b32 %r1471, %f946;
mov.u32 %r1472, 1;
shfl.sync.bfly.b32 %r1473|%p294, %r1471, %r1472, %r1458, %r1460;
mov.b32 %f947, %r1473;
add.ftz.f32 %f78, %f946, %f947;
mov.b32 %r1474, %f938;
shfl.sync.bfly.b32 %r1475|%p295, %r1474, %r1459, %r1458, %r1460;
mov.b32 %f948, %r1475;
add.ftz.f32 %f949, %f938, %f948;
mov.b32 %r1476, %f949;
shfl.sync.bfly.b32 %r1477|%p296, %r1476, %r1463, %r1458, %r1460;
mov.b32 %f950, %r1477;
add.ftz.f32 %f951, %f949, %f950;
mov.b32 %r1478, %f951;
shfl.sync.bfly.b32 %r1479|%p297, %r1478, %r1466, %r1458, %r1460;
mov.b32 %f952, %r1479;
add.ftz.f32 %f953, %f951, %f952;
mov.b32 %r1480, %f953;
shfl.sync.bfly.b32 %r1481|%p298, %r1480, %r1469, %r1458, %r1460;
mov.b32 %f954, %r1481;
add.ftz.f32 %f955, %f953, %f954;
mov.b32 %r1482, %f955;
shfl.sync.bfly.b32 %r1483|%p299, %r1482, %r1472, %r1458, %r1460;
mov.b32 %f956, %r1483;
add.ftz.f32 %f79, %f955, %f956;
setp.ge.u32 %p300, %r4, %r52;
setp.ne.s32 %p301, %r2, 0;
or.pred %p302, %p301, %p300;
@%p302 bra $L__BB116_76;

fma.rn.ftz.f32 %f957, %f10, %f78, %f9;
fma.rn.ftz.f32 %f958, %f11, %f79, %f8;
mul.ftz.f32 %f959, %f957, %f958;
st.global.f32 [%rd37], %f959;

$L__BB116_76:
ld.shared.f32 %f960, [%r34+128];
add.ftz.f32 %f961, %f960, %f35;
ld.shared.f32 %f962, [%r36+128];
add.ftz.f32 %f963, %f962, %f39;
ld.shared.f32 %f964, [%r34+640];
add.ftz.f32 %f965, %f964, %f961;
ld.shared.f32 %f966, [%r36+640];
add.ftz.f32 %f967, %f966, %f963;
ld.shared.f32 %f968, [%r34+1152];
add.ftz.f32 %f969, %f968, %f965;
ld.shared.f32 %f970, [%r36+1152];
add.ftz.f32 %f971, %f970, %f967;
mov.b32 %r1484, %f969;
shfl.sync.bfly.b32 %r1488|%p303, %r1484, %r1459, %r1458, %r1460;
mov.b32 %f972, %r1488;
add.ftz.f32 %f973, %f969, %f972;
mov.b32 %r1489, %f973;
shfl.sync.bfly.b32 %r1491|%p304, %r1489, %r1463, %r1458, %r1460;
mov.b32 %f974, %r1491;
add.ftz.f32 %f975, %f973, %f974;
mov.b32 %r1492, %f975;
shfl.sync.bfly.b32 %r1494|%p305, %r1492, %r1466, %r1458, %r1460;
mov.b32 %f976, %r1494;
add.ftz.f32 %f977, %f975, %f976;
mov.b32 %r1495, %f977;
shfl.sync.bfly.b32 %r1497|%p306, %r1495, %r1469, %r1458, %r1460;
mov.b32 %f978, %r1497;
add.ftz.f32 %f979, %f977, %f978;
mov.b32 %r1498, %f979;
shfl.sync.bfly.b32 %r1500|%p307, %r1498, %r1472, %r1458, %r1460;
mov.b32 %f980, %r1500;
add.ftz.f32 %f80, %f979, %f980;
mov.b32 %r1501, %f971;
shfl.sync.bfly.b32 %r1502|%p308, %r1501, %r1459, %r1458, %r1460;
mov.b32 %f981, %r1502;
add.ftz.f32 %f982, %f971, %f981;
mov.b32 %r1503, %f982;
shfl.sync.bfly.b32 %r1504|%p309, %r1503, %r1463, %r1458, %r1460;
mov.b32 %f983, %r1504;
add.ftz.f32 %f984, %f982, %f983;
mov.b32 %r1505, %f984;
shfl.sync.bfly.b32 %r1506|%p310, %r1505, %r1466, %r1458, %r1460;
mov.b32 %f985, %r1506;
add.ftz.f32 %f986, %f984, %f985;
mov.b32 %r1507, %f986;
shfl.sync.bfly.b32 %r1508|%p311, %r1507, %r1469, %r1458, %r1460;
mov.b32 %f987, %r1508;
add.ftz.f32 %f988, %f986, %f987;
mov.b32 %r1509, %f988;
shfl.sync.bfly.b32 %r1510|%p312, %r1509, %r1472, %r1458, %r1460;
mov.b32 %f989, %r1510;
add.ftz.f32 %f81, %f988, %f989;
add.s32 %r1511, %r4, 1;
setp.ge.u32 %p313, %r1511, %r52;
setp.ne.s32 %p314, %r2, 1;
or.pred %p315, %p314, %p313;
@%p315 bra $L__BB116_78;

fma.rn.ftz.f32 %f990, %f10, %f80, %f9;
fma.rn.ftz.f32 %f991, %f11, %f81, %f8;
mul.ftz.f32 %f992, %f990, %f991;
st.global.f32 [%rd37+4], %f992;

$L__BB116_78:
ld.shared.f32 %f993, [%r34+256];
add.ftz.f32 %f994, %f993, %f34;
ld.shared.f32 %f995, [%r36+256];
add.ftz.f32 %f996, %f995, %f38;
ld.shared.f32 %f997, [%r34+768];
add.ftz.f32 %f998, %f997, %f994;
ld.shared.f32 %f999, [%r36+768];
add.ftz.f32 %f1000, %f999, %f996;
ld.shared.f32 %f1001, [%r34+1280];
add.ftz.f32 %f1002, %f1001, %f998;
ld.shared.f32 %f1003, [%r36+1280];
add.ftz.f32 %f1004, %f1003, %f1000;
mov.b32 %r1512, %f1002;
mov.u32 %r1513, 31;
mov.u32 %r1514, 16;
mov.u32 %r1515, -1;
shfl.sync.bfly.b32 %r1516|%p316, %r1512, %r1514, %r1513, %r1515;
mov.b32 %f1005, %r1516;
add.ftz.f32 %f1006, %f1002, %f1005;
mov.b32 %r1517, %f1006;
mov.u32 %r1518, 8;
shfl.sync.bfly.b32 %r1519|%p317, %r1517, %r1518, %r1513, %r1515;
mov.b32 %f1007, %r1519;
add.ftz.f32 %f1008, %f1006, %f1007;
mov.b32 %r1520, %f1008;
mov.u32 %r1521, 4;
shfl.sync.bfly.b32 %r1522|%p318, %r1520, %r1521, %r1513, %r1515;
mov.b32 %f1009, %r1522;
add.ftz.f32 %f1010, %f1008, %f1009;
mov.b32 %r1523, %f1010;
mov.u32 %r1524, 2;
shfl.sync.bfly.b32 %r1525|%p319, %r1523, %r1524, %r1513, %r1515;
mov.b32 %f1011, %r1525;
add.ftz.f32 %f1012, %f1010, %f1011;
mov.b32 %r1526, %f1012;
mov.u32 %r1527, 1;
shfl.sync.bfly.b32 %r1528|%p320, %r1526, %r1527, %r1513, %r1515;
mov.b32 %f1013, %r1528;
add.ftz.f32 %f82, %f1012, %f1013;
mov.b32 %r1529, %f1004;
shfl.sync.bfly.b32 %r1530|%p321, %r1529, %r1514, %r1513, %r1515;
mov.b32 %f1014, %r1530;
add.ftz.f32 %f1015, %f1004, %f1014;
mov.b32 %r1531, %f1015;
shfl.sync.bfly.b32 %r1532|%p322, %r1531, %r1518, %r1513, %r1515;
mov.b32 %f1016, %r1532;
add.ftz.f32 %f1017, %f1015, %f1016;
mov.b32 %r1533, %f1017;
shfl.sync.bfly.b32 %r1534|%p323, %r1533, %r1521, %r1513, %r1515;
mov.b32 %f1018, %r1534;
add.ftz.f32 %f1019, %f1017, %f1018;
mov.b32 %r1535, %f1019;
shfl.sync.bfly.b32 %r1536|%p324, %r1535, %r1524, %r1513, %r1515;
mov.b32 %f1020, %r1536;
add.ftz.f32 %f1021, %f1019, %f1020;
mov.b32 %r1537, %f1021;
shfl.sync.bfly.b32 %r1538|%p325, %r1537, %r1527, %r1513, %r1515;
mov.b32 %f1022, %r1538;
add.ftz.f32 %f83, %f1021, %f1022;
add.s32 %r1539, %r4, 2;
setp.ge.u32 %p326, %r1539, %r52;
setp.ne.s32 %p327, %r2, 2;
or.pred %p328, %p327, %p326;
@%p328 bra $L__BB116_80;

fma.rn.ftz.f32 %f1023, %f10, %f82, %f9;
fma.rn.ftz.f32 %f1024, %f11, %f83, %f8;
mul.ftz.f32 %f1025, %f1023, %f1024;
st.global.f32 [%rd37+8], %f1025;

$L__BB116_80:
ld.shared.f32 %f1026, [%r34+384];
add.ftz.f32 %f1027, %f1026, %f33;
ld.shared.f32 %f1028, [%r36+384];
add.ftz.f32 %f1029, %f1028, %f37;
ld.shared.f32 %f1030, [%r34+896];
add.ftz.f32 %f1031, %f1030, %f1027;
ld.shared.f32 %f1032, [%r36+896];
add.ftz.f32 %f1033, %f1032, %f1029;
ld.shared.f32 %f1034, [%r34+1408];
add.ftz.f32 %f1035, %f1034, %f1031;
ld.shared.f32 %f1036, [%r36+1408];
add.ftz.f32 %f1037, %f1036, %f1033;
mov.b32 %r1540, %f1035;
shfl.sync.bfly.b32 %r1544|%p329, %r1540, %r1514, %r1513, %r1515;
mov.b32 %f1038, %r1544;
add.ftz.f32 %f1039, %f1035, %f1038;
mov.b32 %r1545, %f1039;
shfl.sync.bfly.b32 %r1547|%p330, %r1545, %r1518, %r1513, %r1515;
mov.b32 %f1040, %r1547;
add.ftz.f32 %f1041, %f1039, %f1040;
mov.b32 %r1548, %f1041;
shfl.sync.bfly.b32 %r1550|%p331, %r1548, %r1521, %r1513, %r1515;
mov.b32 %f1042, %r1550;
add.ftz.f32 %f1043, %f1041, %f1042;
mov.b32 %r1551, %f1043;
shfl.sync.bfly.b32 %r1553|%p332, %r1551, %r1524, %r1513, %r1515;
mov.b32 %f1044, %r1553;
add.ftz.f32 %f1045, %f1043, %f1044;
mov.b32 %r1554, %f1045;
shfl.sync.bfly.b32 %r1556|%p333, %r1554, %r1527, %r1513, %r1515;
mov.b32 %f1046, %r1556;
add.ftz.f32 %f84, %f1045, %f1046;
mov.b32 %r1557, %f1037;
shfl.sync.bfly.b32 %r1558|%p334, %r1557, %r1514, %r1513, %r1515;
mov.b32 %f1047, %r1558;
add.ftz.f32 %f1048, %f1037, %f1047;
mov.b32 %r1559, %f1048;
shfl.sync.bfly.b32 %r1560|%p335, %r1559, %r1518, %r1513, %r1515;
mov.b32 %f1049, %r1560;
add.ftz.f32 %f1050, %f1048, %f1049;
mov.b32 %r1561, %f1050;
shfl.sync.bfly.b32 %r1562|%p336, %r1561, %r1521, %r1513, %r1515;
mov.b32 %f1051, %r1562;
add.ftz.f32 %f1052, %f1050, %f1051;
mov.b32 %r1563, %f1052;
shfl.sync.bfly.b32 %r1564|%p337, %r1563, %r1524, %r1513, %r1515;
mov.b32 %f1053, %r1564;
add.ftz.f32 %f1054, %f1052, %f1053;
mov.b32 %r1565, %f1054;
shfl.sync.bfly.b32 %r1566|%p338, %r1565, %r1527, %r1513, %r1515;
mov.b32 %f1055, %r1566;
add.ftz.f32 %f85, %f1054, %f1055;
add.s32 %r1567, %r4, 3;
setp.ge.u32 %p339, %r1567, %r52;
setp.ne.s32 %p340, %r2, 3;
or.pred %p341, %p340, %p339;
@%p341 bra $L__BB116_82;

fma.rn.ftz.f32 %f1056, %f10, %f84, %f9;
fma.rn.ftz.f32 %f1057, %f11, %f85, %f8;
mul.ftz.f32 %f1058, %f1056, %f1057;
st.global.f32 [%rd37+12], %f1058;
bra.uni $L__BB116_82;

$L__BB116_66:
ld.shared.f32 %f763, [%r34];
add.ftz.f32 %f764, %f763, %f36;
ld.shared.f32 %f765, [%r36];
add.ftz.f32 %f766, %f765, %f40;
ld.shared.f32 %f767, [%r34+512];
add.ftz.f32 %f768, %f767, %f764;
ld.shared.f32 %f769, [%r36+512];
add.ftz.f32 %f770, %f769, %f766;
ld.shared.f32 %f771, [%r34+1024];
add.ftz.f32 %f772, %f771, %f768;
ld.shared.f32 %f773, [%r36+1024];
add.ftz.f32 %f774, %f773, %f770;
mov.b32 %r1346, %f772;
mov.u32 %r1347, 31;
mov.u32 %r1348, 16;
mov.u32 %r1349, -1;
shfl.sync.bfly.b32 %r1350|%p238, %r1346, %r1348, %r1347, %r1349;
mov.b32 %f775, %r1350;
add.ftz.f32 %f776, %f772, %f775;
mov.b32 %r1351, %f776;
mov.u32 %r1352, 8;
shfl.sync.bfly.b32 %r1353|%p239, %r1351, %r1352, %r1347, %r1349;
mov.b32 %f777, %r1353;
add.ftz.f32 %f778, %f776, %f777;
mov.b32 %r1354, %f778;
mov.u32 %r1355, 4;
shfl.sync.bfly.b32 %r1356|%p240, %r1354, %r1355, %r1347, %r1349;
mov.b32 %f779, %r1356;
add.ftz.f32 %f780, %f778, %f779;
mov.b32 %r1357, %f780;
mov.u32 %r1358, 2;
shfl.sync.bfly.b32 %r1359|%p241, %r1357, %r1358, %r1347, %r1349;
mov.b32 %f781, %r1359;
add.ftz.f32 %f782, %f780, %f781;
mov.b32 %r1360, %f782;
mov.u32 %r1361, 1;
shfl.sync.bfly.b32 %r1362|%p242, %r1360, %r1361, %r1347, %r1349;
mov.b32 %f783, %r1362;
add.ftz.f32 %f70, %f782, %f783;
mov.b32 %r1363, %f774;
shfl.sync.bfly.b32 %r1364|%p243, %r1363, %r1348, %r1347, %r1349;
mov.b32 %f784, %r1364;
add.ftz.f32 %f785, %f774, %f784;
mov.b32 %r1365, %f785;
shfl.sync.bfly.b32 %r1366|%p244, %r1365, %r1352, %r1347, %r1349;
mov.b32 %f786, %r1366;
add.ftz.f32 %f787, %f785, %f786;
mov.b32 %r1367, %f787;
shfl.sync.bfly.b32 %r1368|%p245, %r1367, %r1355, %r1347, %r1349;
mov.b32 %f788, %r1368;
add.ftz.f32 %f789, %f787, %f788;
mov.b32 %r1369, %f789;
shfl.sync.bfly.b32 %r1370|%p246, %r1369, %r1358, %r1347, %r1349;
mov.b32 %f790, %r1370;
add.ftz.f32 %f791, %f789, %f790;
mov.b32 %r1371, %f791;
shfl.sync.bfly.b32 %r1372|%p247, %r1371, %r1361, %r1347, %r1349;
mov.b32 %f792, %r1372;
add.ftz.f32 %f71, %f791, %f792;
setp.ge.u32 %p248, %r4, %r52;
setp.ne.s32 %p249, %r2, 0;
or.pred %p250, %p249, %p248;
@%p250 bra $L__BB116_68;

fma.rn.ftz.f32 %f793, %f10, %f70, %f9;
fma.rn.ftz.f32 %f794, %f11, %f71, %f8;
mul.ftz.f32 %f795, %f794, 0f3F000000;
mul.ftz.f32 %f796, %f794, 0f3F4C422A;
mul.ftz.f32 %f797, %f794, 0f3D372713;
fma.rn.ftz.f32 %f798, %f794, %f797, 0f3F800000;
mul.ftz.f32 %f799, %f796, %f798;
tanh.approx.f32 %f800, %f799;
add.ftz.f32 %f801, %f800, 0f3F800000;
mul.ftz.f32 %f802, %f795, %f801;
mul.ftz.f32 %f803, %f793, %f802;
st.global.f32 [%rd37], %f803;

$L__BB116_68:
ld.shared.f32 %f804, [%r34+128];
add.ftz.f32 %f805, %f804, %f35;
ld.shared.f32 %f806, [%r36+128];
add.ftz.f32 %f807, %f806, %f39;
ld.shared.f32 %f808, [%r34+640];
add.ftz.f32 %f809, %f808, %f805;
ld.shared.f32 %f810, [%r36+640];
add.ftz.f32 %f811, %f810, %f807;
ld.shared.f32 %f812, [%r34+1152];
add.ftz.f32 %f813, %f812, %f809;
ld.shared.f32 %f814, [%r36+1152];
add.ftz.f32 %f815, %f814, %f811;
mov.b32 %r1373, %f813;
shfl.sync.bfly.b32 %r1377|%p251, %r1373, %r1348, %r1347, %r1349;
mov.b32 %f816, %r1377;
add.ftz.f32 %f817, %f813, %f816;
mov.b32 %r1378, %f817;
shfl.sync.bfly.b32 %r1380|%p252, %r1378, %r1352, %r1347, %r1349;
mov.b32 %f818, %r1380;
add.ftz.f32 %f819, %f817, %f818;
mov.b32 %r1381, %f819;
shfl.sync.bfly.b32 %r1383|%p253, %r1381, %r1355, %r1347, %r1349;
mov.b32 %f820, %r1383;
add.ftz.f32 %f821, %f819, %f820;
mov.b32 %r1384, %f821;
shfl.sync.bfly.b32 %r1386|%p254, %r1384, %r1358, %r1347, %r1349;
mov.b32 %f822, %r1386;
add.ftz.f32 %f823, %f821, %f822;
mov.b32 %r1387, %f823;
shfl.sync.bfly.b32 %r1389|%p255, %r1387, %r1361, %r1347, %r1349;
mov.b32 %f824, %r1389;
add.ftz.f32 %f72, %f823, %f824;
mov.b32 %r1390, %f815;
shfl.sync.bfly.b32 %r1391|%p256, %r1390, %r1348, %r1347, %r1349;
mov.b32 %f825, %r1391;
add.ftz.f32 %f826, %f815, %f825;
mov.b32 %r1392, %f826;
shfl.sync.bfly.b32 %r1393|%p257, %r1392, %r1352, %r1347, %r1349;
mov.b32 %f827, %r1393;
add.ftz.f32 %f828, %f826, %f827;
mov.b32 %r1394, %f828;
shfl.sync.bfly.b32 %r1395|%p258, %r1394, %r1355, %r1347, %r1349;
mov.b32 %f829, %r1395;
add.ftz.f32 %f830, %f828, %f829;
mov.b32 %r1396, %f830;
shfl.sync.bfly.b32 %r1397|%p259, %r1396, %r1358, %r1347, %r1349;
mov.b32 %f831, %r1397;
add.ftz.f32 %f832, %f830, %f831;
mov.b32 %r1398, %f832;
shfl.sync.bfly.b32 %r1399|%p260, %r1398, %r1361, %r1347, %r1349;
mov.b32 %f833, %r1399;
add.ftz.f32 %f73, %f832, %f833;
add.s32 %r1400, %r4, 1;
setp.ge.u32 %p261, %r1400, %r52;
setp.ne.s32 %p262, %r2, 1;
or.pred %p263, %p262, %p261;
@%p263 bra $L__BB116_70;

fma.rn.ftz.f32 %f834, %f10, %f72, %f9;
fma.rn.ftz.f32 %f835, %f11, %f73, %f8;
mul.ftz.f32 %f836, %f835, 0f3F000000;
mul.ftz.f32 %f837, %f835, 0f3F4C422A;
mul.ftz.f32 %f838, %f835, 0f3D372713;
fma.rn.ftz.f32 %f839, %f835, %f838, 0f3F800000;
mul.ftz.f32 %f840, %f837, %f839;
tanh.approx.f32 %f841, %f840;
add.ftz.f32 %f842, %f841, 0f3F800000;
mul.ftz.f32 %f843, %f836, %f842;
mul.ftz.f32 %f844, %f834, %f843;
st.global.f32 [%rd37+4], %f844;

$L__BB116_70:
ld.shared.f32 %f845, [%r34+256];
add.ftz.f32 %f846, %f845, %f34;
ld.shared.f32 %f847, [%r36+256];
add.ftz.f32 %f848, %f847, %f38;
ld.shared.f32 %f849, [%r34+768];
add.ftz.f32 %f850, %f849, %f846;
ld.shared.f32 %f851, [%r36+768];
add.ftz.f32 %f852, %f851, %f848;
ld.shared.f32 %f853, [%r34+1280];
add.ftz.f32 %f854, %f853, %f850;
ld.shared.f32 %f855, [%r36+1280];
add.ftz.f32 %f856, %f855, %f852;
mov.b32 %r1401, %f854;
mov.u32 %r1402, 31;
mov.u32 %r1403, 16;
mov.u32 %r1404, -1;
shfl.sync.bfly.b32 %r1405|%p264, %r1401, %r1403, %r1402, %r1404;
mov.b32 %f857, %r1405;
add.ftz.f32 %f858, %f854, %f857;
mov.b32 %r1406, %f858;
mov.u32 %r1407, 8;
shfl.sync.bfly.b32 %r1408|%p265, %r1406, %r1407, %r1402, %r1404;
mov.b32 %f859, %r1408;
add.ftz.f32 %f860, %f858, %f859;
mov.b32 %r1409, %f860;
mov.u32 %r1410, 4;
shfl.sync.bfly.b32 %r1411|%p266, %r1409, %r1410, %r1402, %r1404;
mov.b32 %f861, %r1411;
add.ftz.f32 %f862, %f860, %f861;
mov.b32 %r1412, %f862;
mov.u32 %r1413, 2;
shfl.sync.bfly.b32 %r1414|%p267, %r1412, %r1413, %r1402, %r1404;
mov.b32 %f863, %r1414;
add.ftz.f32 %f864, %f862, %f863;
mov.b32 %r1415, %f864;
mov.u32 %r1416, 1;
shfl.sync.bfly.b32 %r1417|%p268, %r1415, %r1416, %r1402, %r1404;
mov.b32 %f865, %r1417;
add.ftz.f32 %f74, %f864, %f865;
mov.b32 %r1418, %f856;
shfl.sync.bfly.b32 %r1419|%p269, %r1418, %r1403, %r1402, %r1404;
mov.b32 %f866, %r1419;
add.ftz.f32 %f867, %f856, %f866;
mov.b32 %r1420, %f867;
shfl.sync.bfly.b32 %r1421|%p270, %r1420, %r1407, %r1402, %r1404;
mov.b32 %f868, %r1421;
add.ftz.f32 %f869, %f867, %f868;
mov.b32 %r1422, %f869;
shfl.sync.bfly.b32 %r1423|%p271, %r1422, %r1410, %r1402, %r1404;
mov.b32 %f870, %r1423;
add.ftz.f32 %f871, %f869, %f870;
mov.b32 %r1424, %f871;
shfl.sync.bfly.b32 %r1425|%p272, %r1424, %r1413, %r1402, %r1404;
mov.b32 %f872, %r1425;
add.ftz.f32 %f873, %f871, %f872;
mov.b32 %r1426, %f873;
shfl.sync.bfly.b32 %r1427|%p273, %r1426, %r1416, %r1402, %r1404;
mov.b32 %f874, %r1427;
add.ftz.f32 %f75, %f873, %f874;
add.s32 %r1428, %r4, 2;
setp.ge.u32 %p274, %r1428, %r52;
setp.ne.s32 %p275, %r2, 2;
or.pred %p276, %p275, %p274;
@%p276 bra $L__BB116_72;

fma.rn.ftz.f32 %f875, %f10, %f74, %f9;
fma.rn.ftz.f32 %f876, %f11, %f75, %f8;
mul.ftz.f32 %f877, %f876, 0f3F000000;
mul.ftz.f32 %f878, %f876, 0f3F4C422A;
mul.ftz.f32 %f879, %f876, 0f3D372713;
fma.rn.ftz.f32 %f880, %f876, %f879, 0f3F800000;
mul.ftz.f32 %f881, %f878, %f880;
tanh.approx.f32 %f882, %f881;
add.ftz.f32 %f883, %f882, 0f3F800000;
mul.ftz.f32 %f884, %f877, %f883;
mul.ftz.f32 %f885, %f875, %f884;
st.global.f32 [%rd37+8], %f885;

$L__BB116_72:
ld.shared.f32 %f886, [%r34+384];
add.ftz.f32 %f887, %f886, %f33;
ld.shared.f32 %f888, [%r36+384];
add.ftz.f32 %f889, %f888, %f37;
ld.shared.f32 %f890, [%r34+896];
add.ftz.f32 %f891, %f890, %f887;
ld.shared.f32 %f892, [%r36+896];
add.ftz.f32 %f893, %f892, %f889;
ld.shared.f32 %f894, [%r34+1408];
add.ftz.f32 %f895, %f894, %f891;
ld.shared.f32 %f896, [%r36+1408];
add.ftz.f32 %f897, %f896, %f893;
mov.b32 %r1429, %f895;
shfl.sync.bfly.b32 %r1433|%p277, %r1429, %r1403, %r1402, %r1404;
mov.b32 %f898, %r1433;
add.ftz.f32 %f899, %f895, %f898;
mov.b32 %r1434, %f899;
shfl.sync.bfly.b32 %r1436|%p278, %r1434, %r1407, %r1402, %r1404;
mov.b32 %f900, %r1436;
add.ftz.f32 %f901, %f899, %f900;
mov.b32 %r1437, %f901;
shfl.sync.bfly.b32 %r1439|%p279, %r1437, %r1410, %r1402, %r1404;
mov.b32 %f902, %r1439;
add.ftz.f32 %f903, %f901, %f902;
mov.b32 %r1440, %f903;
shfl.sync.bfly.b32 %r1442|%p280, %r1440, %r1413, %r1402, %r1404;
mov.b32 %f904, %r1442;
add.ftz.f32 %f905, %f903, %f904;
mov.b32 %r1443, %f905;
shfl.sync.bfly.b32 %r1445|%p281, %r1443, %r1416, %r1402, %r1404;
mov.b32 %f906, %r1445;
add.ftz.f32 %f76, %f905, %f906;
mov.b32 %r1446, %f897;
shfl.sync.bfly.b32 %r1447|%p282, %r1446, %r1403, %r1402, %r1404;
mov.b32 %f907, %r1447;
add.ftz.f32 %f908, %f897, %f907;
mov.b32 %r1448, %f908;
shfl.sync.bfly.b32 %r1449|%p283, %r1448, %r1407, %r1402, %r1404;
mov.b32 %f909, %r1449;
add.ftz.f32 %f910, %f908, %f909;
mov.b32 %r1450, %f910;
shfl.sync.bfly.b32 %r1451|%p284, %r1450, %r1410, %r1402, %r1404;
mov.b32 %f911, %r1451;
add.ftz.f32 %f912, %f910, %f911;
mov.b32 %r1452, %f912;
shfl.sync.bfly.b32 %r1453|%p285, %r1452, %r1413, %r1402, %r1404;
mov.b32 %f913, %r1453;
add.ftz.f32 %f914, %f912, %f913;
mov.b32 %r1454, %f914;
shfl.sync.bfly.b32 %r1455|%p286, %r1454, %r1416, %r1402, %r1404;
mov.b32 %f915, %r1455;
add.ftz.f32 %f77, %f914, %f915;
add.s32 %r1456, %r4, 3;
setp.ge.u32 %p287, %r1456, %r52;
setp.ne.s32 %p288, %r2, 3;
or.pred %p289, %p288, %p287;
@%p289 bra $L__BB116_82;

fma.rn.ftz.f32 %f916, %f10, %f76, %f9;
fma.rn.ftz.f32 %f917, %f11, %f77, %f8;
mul.ftz.f32 %f918, %f917, 0f3F000000;
mul.ftz.f32 %f919, %f917, 0f3F4C422A;
mul.ftz.f32 %f920, %f917, 0f3D372713;
fma.rn.ftz.f32 %f921, %f917, %f920, 0f3F800000;
mul.ftz.f32 %f922, %f919, %f921;
tanh.approx.f32 %f923, %f922;
add.ftz.f32 %f924, %f923, 0f3F800000;
mul.ftz.f32 %f925, %f918, %f924;
mul.ftz.f32 %f926, %f916, %f925;
st.global.f32 [%rd37+12], %f926;
bra.uni $L__BB116_82;

$L__BB116_58:
ld.shared.f32 %f439, [%r34];
add.ftz.f32 %f440, %f439, %f36;
ld.shared.f32 %f441, [%r36];
add.ftz.f32 %f442, %f441, %f40;
ld.shared.f32 %f443, [%r34+512];
add.ftz.f32 %f444, %f443, %f440;
ld.shared.f32 %f445, [%r36+512];
add.ftz.f32 %f446, %f445, %f442;
ld.shared.f32 %f447, [%r34+1024];
add.ftz.f32 %f448, %f447, %f444;
ld.shared.f32 %f449, [%r36+1024];
add.ftz.f32 %f450, %f449, %f446;
mov.b32 %r1124, %f448;
mov.u32 %r1125, 31;
mov.u32 %r1126, 16;
mov.u32 %r1127, -1;
shfl.sync.bfly.b32 %r1128|%p134, %r1124, %r1126, %r1125, %r1127;
mov.b32 %f451, %r1128;
add.ftz.f32 %f452, %f448, %f451;
mov.b32 %r1129, %f452;
mov.u32 %r1130, 8;
shfl.sync.bfly.b32 %r1131|%p135, %r1129, %r1130, %r1125, %r1127;
mov.b32 %f453, %r1131;
add.ftz.f32 %f454, %f452, %f453;
mov.b32 %r1132, %f454;
mov.u32 %r1133, 4;
shfl.sync.bfly.b32 %r1134|%p136, %r1132, %r1133, %r1125, %r1127;
mov.b32 %f455, %r1134;
add.ftz.f32 %f456, %f454, %f455;
mov.b32 %r1135, %f456;
mov.u32 %r1136, 2;
shfl.sync.bfly.b32 %r1137|%p137, %r1135, %r1136, %r1125, %r1127;
mov.b32 %f457, %r1137;
add.ftz.f32 %f458, %f456, %f457;
mov.b32 %r1138, %f458;
mov.u32 %r1139, 1;
shfl.sync.bfly.b32 %r1140|%p138, %r1138, %r1139, %r1125, %r1127;
mov.b32 %f459, %r1140;
add.ftz.f32 %f54, %f458, %f459;
mov.b32 %r1141, %f450;
shfl.sync.bfly.b32 %r1142|%p139, %r1141, %r1126, %r1125, %r1127;
mov.b32 %f460, %r1142;
add.ftz.f32 %f461, %f450, %f460;
mov.b32 %r1143, %f461;
shfl.sync.bfly.b32 %r1144|%p140, %r1143, %r1130, %r1125, %r1127;
mov.b32 %f462, %r1144;
add.ftz.f32 %f463, %f461, %f462;
mov.b32 %r1145, %f463;
shfl.sync.bfly.b32 %r1146|%p141, %r1145, %r1133, %r1125, %r1127;
mov.b32 %f464, %r1146;
add.ftz.f32 %f465, %f463, %f464;
mov.b32 %r1147, %f465;
shfl.sync.bfly.b32 %r1148|%p142, %r1147, %r1136, %r1125, %r1127;
mov.b32 %f466, %r1148;
add.ftz.f32 %f467, %f465, %f466;
mov.b32 %r1149, %f467;
shfl.sync.bfly.b32 %r1150|%p143, %r1149, %r1139, %r1125, %r1127;
mov.b32 %f468, %r1150;
add.ftz.f32 %f55, %f467, %f468;
setp.ge.u32 %p144, %r4, %r52;
setp.ne.s32 %p145, %r2, 0;
or.pred %p146, %p145, %p144;
@%p146 bra $L__BB116_60;

fma.rn.ftz.f32 %f469, %f10, %f54, %f9;
fma.rn.ftz.f32 %f470, %f11, %f55, %f8;
mov.f32 %f471, 0f40E00000;
min.ftz.f32 %f472, %f470, %f471;
min.ftz.f32 %f473, %f469, %f471;
mov.f32 %f474, 0fC0E00000;
max.ftz.f32 %f475, %f473, %f474;
mul.ftz.f32 %f476, %f472, 0f3FD9DB23;
mul.ftz.f32 %f477, %f476, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f478, %f477;
add.ftz.f32 %f479, %f478, 0f3F800000;
div.approx.ftz.f32 %f480, %f472, %f479;
add.ftz.f32 %f481, %f475, 0f3F800000;
mul.ftz.f32 %f482, %f481, %f480;
st.global.f32 [%rd37], %f482;

$L__BB116_60:
ld.shared.f32 %f483, [%r34+128];
add.ftz.f32 %f484, %f483, %f35;
ld.shared.f32 %f485, [%r36+128];
add.ftz.f32 %f486, %f485, %f39;
ld.shared.f32 %f487, [%r34+640];
add.ftz.f32 %f488, %f487, %f484;
ld.shared.f32 %f489, [%r36+640];
add.ftz.f32 %f490, %f489, %f486;
ld.shared.f32 %f491, [%r34+1152];
add.ftz.f32 %f492, %f491, %f488;
ld.shared.f32 %f493, [%r36+1152];
add.ftz.f32 %f494, %f493, %f490;
mov.b32 %r1151, %f492;
shfl.sync.bfly.b32 %r1155|%p147, %r1151, %r1126, %r1125, %r1127;
mov.b32 %f495, %r1155;
add.ftz.f32 %f496, %f492, %f495;
mov.b32 %r1156, %f496;
shfl.sync.bfly.b32 %r1158|%p148, %r1156, %r1130, %r1125, %r1127;
mov.b32 %f497, %r1158;
add.ftz.f32 %f498, %f496, %f497;
mov.b32 %r1159, %f498;
shfl.sync.bfly.b32 %r1161|%p149, %r1159, %r1133, %r1125, %r1127;
mov.b32 %f499, %r1161;
add.ftz.f32 %f500, %f498, %f499;
mov.b32 %r1162, %f500;
shfl.sync.bfly.b32 %r1164|%p150, %r1162, %r1136, %r1125, %r1127;
mov.b32 %f501, %r1164;
add.ftz.f32 %f502, %f500, %f501;
mov.b32 %r1165, %f502;
shfl.sync.bfly.b32 %r1167|%p151, %r1165, %r1139, %r1125, %r1127;
mov.b32 %f503, %r1167;
add.ftz.f32 %f56, %f502, %f503;
mov.b32 %r1168, %f494;
shfl.sync.bfly.b32 %r1169|%p152, %r1168, %r1126, %r1125, %r1127;
mov.b32 %f504, %r1169;
add.ftz.f32 %f505, %f494, %f504;
mov.b32 %r1170, %f505;
shfl.sync.bfly.b32 %r1171|%p153, %r1170, %r1130, %r1125, %r1127;
mov.b32 %f506, %r1171;
add.ftz.f32 %f507, %f505, %f506;
mov.b32 %r1172, %f507;
shfl.sync.bfly.b32 %r1173|%p154, %r1172, %r1133, %r1125, %r1127;
mov.b32 %f508, %r1173;
add.ftz.f32 %f509, %f507, %f508;
mov.b32 %r1174, %f509;
shfl.sync.bfly.b32 %r1175|%p155, %r1174, %r1136, %r1125, %r1127;
mov.b32 %f510, %r1175;
add.ftz.f32 %f511, %f509, %f510;
mov.b32 %r1176, %f511;
shfl.sync.bfly.b32 %r1177|%p156, %r1176, %r1139, %r1125, %r1127;
mov.b32 %f512, %r1177;
add.ftz.f32 %f57, %f511, %f512;
add.s32 %r1178, %r4, 1;
setp.ge.u32 %p157, %r1178, %r52;
setp.ne.s32 %p158, %r2, 1;
or.pred %p159, %p158, %p157;
@%p159 bra $L__BB116_62;

fma.rn.ftz.f32 %f513, %f10, %f56, %f9;
fma.rn.ftz.f32 %f514, %f11, %f57, %f8;
mov.f32 %f515, 0f40E00000;
min.ftz.f32 %f516, %f514, %f515;
min.ftz.f32 %f517, %f513, %f515;
mov.f32 %f518, 0fC0E00000;
max.ftz.f32 %f519, %f517, %f518;
mul.ftz.f32 %f520, %f516, 0f3FD9DB23;
mul.ftz.f32 %f521, %f520, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f522, %f521;
add.ftz.f32 %f523, %f522, 0f3F800000;
div.approx.ftz.f32 %f524, %f516, %f523;
add.ftz.f32 %f525, %f519, 0f3F800000;
mul.ftz.f32 %f526, %f525, %f524;
st.global.f32 [%rd37+4], %f526;

$L__BB116_62:
ld.shared.f32 %f527, [%r34+256];
add.ftz.f32 %f528, %f527, %f34;
ld.shared.f32 %f529, [%r36+256];
add.ftz.f32 %f530, %f529, %f38;
ld.shared.f32 %f531, [%r34+768];
add.ftz.f32 %f532, %f531, %f528;
ld.shared.f32 %f533, [%r36+768];
add.ftz.f32 %f534, %f533, %f530;
ld.shared.f32 %f535, [%r34+1280];
add.ftz.f32 %f536, %f535, %f532;
ld.shared.f32 %f537, [%r36+1280];
add.ftz.f32 %f538, %f537, %f534;
mov.b32 %r1179, %f536;
mov.u32 %r1180, 31;
mov.u32 %r1181, 16;
mov.u32 %r1182, -1;
shfl.sync.bfly.b32 %r1183|%p160, %r1179, %r1181, %r1180, %r1182;
mov.b32 %f539, %r1183;
add.ftz.f32 %f540, %f536, %f539;
mov.b32 %r1184, %f540;
mov.u32 %r1185, 8;
shfl.sync.bfly.b32 %r1186|%p161, %r1184, %r1185, %r1180, %r1182;
mov.b32 %f541, %r1186;
add.ftz.f32 %f542, %f540, %f541;
mov.b32 %r1187, %f542;
mov.u32 %r1188, 4;
shfl.sync.bfly.b32 %r1189|%p162, %r1187, %r1188, %r1180, %r1182;
mov.b32 %f543, %r1189;
add.ftz.f32 %f544, %f542, %f543;
mov.b32 %r1190, %f544;
mov.u32 %r1191, 2;
shfl.sync.bfly.b32 %r1192|%p163, %r1190, %r1191, %r1180, %r1182;
mov.b32 %f545, %r1192;
add.ftz.f32 %f546, %f544, %f545;
mov.b32 %r1193, %f546;
mov.u32 %r1194, 1;
shfl.sync.bfly.b32 %r1195|%p164, %r1193, %r1194, %r1180, %r1182;
mov.b32 %f547, %r1195;
add.ftz.f32 %f58, %f546, %f547;
mov.b32 %r1196, %f538;
shfl.sync.bfly.b32 %r1197|%p165, %r1196, %r1181, %r1180, %r1182;
mov.b32 %f548, %r1197;
add.ftz.f32 %f549, %f538, %f548;
mov.b32 %r1198, %f549;
shfl.sync.bfly.b32 %r1199|%p166, %r1198, %r1185, %r1180, %r1182;
mov.b32 %f550, %r1199;
add.ftz.f32 %f551, %f549, %f550;
mov.b32 %r1200, %f551;
shfl.sync.bfly.b32 %r1201|%p167, %r1200, %r1188, %r1180, %r1182;
mov.b32 %f552, %r1201;
add.ftz.f32 %f553, %f551, %f552;
mov.b32 %r1202, %f553;
shfl.sync.bfly.b32 %r1203|%p168, %r1202, %r1191, %r1180, %r1182;
mov.b32 %f554, %r1203;
add.ftz.f32 %f555, %f553, %f554;
mov.b32 %r1204, %f555;
shfl.sync.bfly.b32 %r1205|%p169, %r1204, %r1194, %r1180, %r1182;
mov.b32 %f556, %r1205;
add.ftz.f32 %f59, %f555, %f556;
add.s32 %r1206, %r4, 2;
setp.ge.u32 %p170, %r1206, %r52;
setp.ne.s32 %p171, %r2, 2;
or.pred %p172, %p171, %p170;
@%p172 bra $L__BB116_64;

fma.rn.ftz.f32 %f557, %f10, %f58, %f9;
fma.rn.ftz.f32 %f558, %f11, %f59, %f8;
mov.f32 %f559, 0f40E00000;
min.ftz.f32 %f560, %f558, %f559;
min.ftz.f32 %f561, %f557, %f559;
mov.f32 %f562, 0fC0E00000;
max.ftz.f32 %f563, %f561, %f562;
mul.ftz.f32 %f564, %f560, 0f3FD9DB23;
mul.ftz.f32 %f565, %f564, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f566, %f565;
add.ftz.f32 %f567, %f566, 0f3F800000;
div.approx.ftz.f32 %f568, %f560, %f567;
add.ftz.f32 %f569, %f563, 0f3F800000;
mul.ftz.f32 %f570, %f569, %f568;
st.global.f32 [%rd37+8], %f570;

$L__BB116_64:
ld.shared.f32 %f571, [%r34+384];
add.ftz.f32 %f572, %f571, %f33;
ld.shared.f32 %f573, [%r36+384];
add.ftz.f32 %f574, %f573, %f37;
ld.shared.f32 %f575, [%r34+896];
add.ftz.f32 %f576, %f575, %f572;
ld.shared.f32 %f577, [%r36+896];
add.ftz.f32 %f578, %f577, %f574;
ld.shared.f32 %f579, [%r34+1408];
add.ftz.f32 %f580, %f579, %f576;
ld.shared.f32 %f581, [%r36+1408];
add.ftz.f32 %f582, %f581, %f578;
mov.b32 %r1207, %f580;
shfl.sync.bfly.b32 %r1211|%p173, %r1207, %r1181, %r1180, %r1182;
mov.b32 %f583, %r1211;
add.ftz.f32 %f584, %f580, %f583;
mov.b32 %r1212, %f584;
shfl.sync.bfly.b32 %r1214|%p174, %r1212, %r1185, %r1180, %r1182;
mov.b32 %f585, %r1214;
add.ftz.f32 %f586, %f584, %f585;
mov.b32 %r1215, %f586;
shfl.sync.bfly.b32 %r1217|%p175, %r1215, %r1188, %r1180, %r1182;
mov.b32 %f587, %r1217;
add.ftz.f32 %f588, %f586, %f587;
mov.b32 %r1218, %f588;
shfl.sync.bfly.b32 %r1220|%p176, %r1218, %r1191, %r1180, %r1182;
mov.b32 %f589, %r1220;
add.ftz.f32 %f590, %f588, %f589;
mov.b32 %r1221, %f590;
shfl.sync.bfly.b32 %r1223|%p177, %r1221, %r1194, %r1180, %r1182;
mov.b32 %f591, %r1223;
add.ftz.f32 %f60, %f590, %f591;
mov.b32 %r1224, %f582;
shfl.sync.bfly.b32 %r1225|%p178, %r1224, %r1181, %r1180, %r1182;
mov.b32 %f592, %r1225;
add.ftz.f32 %f593, %f582, %f592;
mov.b32 %r1226, %f593;
shfl.sync.bfly.b32 %r1227|%p179, %r1226, %r1185, %r1180, %r1182;
mov.b32 %f594, %r1227;
add.ftz.f32 %f595, %f593, %f594;
mov.b32 %r1228, %f595;
shfl.sync.bfly.b32 %r1229|%p180, %r1228, %r1188, %r1180, %r1182;
mov.b32 %f596, %r1229;
add.ftz.f32 %f597, %f595, %f596;
mov.b32 %r1230, %f597;
shfl.sync.bfly.b32 %r1231|%p181, %r1230, %r1191, %r1180, %r1182;
mov.b32 %f598, %r1231;
add.ftz.f32 %f599, %f597, %f598;
mov.b32 %r1232, %f599;
shfl.sync.bfly.b32 %r1233|%p182, %r1232, %r1194, %r1180, %r1182;
mov.b32 %f600, %r1233;
add.ftz.f32 %f61, %f599, %f600;
add.s32 %r1234, %r4, 3;
setp.ge.u32 %p183, %r1234, %r52;
setp.ne.s32 %p184, %r2, 3;
or.pred %p185, %p184, %p183;
@%p185 bra $L__BB116_82;

fma.rn.ftz.f32 %f601, %f10, %f60, %f9;
fma.rn.ftz.f32 %f602, %f11, %f61, %f8;
mov.f32 %f603, 0f40E00000;
min.ftz.f32 %f604, %f602, %f603;
min.ftz.f32 %f605, %f601, %f603;
mov.f32 %f606, 0fC0E00000;
max.ftz.f32 %f607, %f605, %f606;
mul.ftz.f32 %f608, %f604, 0f3FD9DB23;
mul.ftz.f32 %f609, %f608, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f610, %f609;
add.ftz.f32 %f611, %f610, 0f3F800000;
div.approx.ftz.f32 %f612, %f604, %f611;
add.ftz.f32 %f613, %f607, 0f3F800000;
mul.ftz.f32 %f614, %f613, %f612;
st.global.f32 [%rd37+12], %f614;
bra.uni $L__BB116_82;

}
.entry _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<46>;
.reg .b16 %rs<41>;
.reg .f32 %f<142>;
.reg .b32 %r<546>;
.reg .b64 %rd<51>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

ld.param.u64 %rd13, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd14, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r19, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
mov.u32 %r1, %tid.y;
mov.u32 %r2, %tid.x;
mov.u32 %r37, %ctaid.x;
shl.b32 %r3, %r37, 2;
shr.u32 %r4, %r36, 6;
mov.u32 %r543, %ctaid.y;
setp.eq.s64 %p1, %rd15, 0;
@%p1 bra $L__BB117_2;

cvta.to.global.u64 %rd17, %rd15;
mul.wide.u32 %rd18, %r543, 4;
add.s64 %rd19, %rd17, %rd18;
ld.global.nc.u32 %r542, [%rd19];
mul.hi.u32 %r38, %r543, %r19;
add.s32 %r39, %r38, %r543;
shr.u32 %r40, %r39, %r20;
mul.lo.s32 %r41, %r40, %r21;
sub.s32 %r543, %r543, %r41;
bra.uni $L__BB117_3;

$L__BB117_2:
mul.hi.u32 %r42, %r543, %r24;
add.s32 %r43, %r42, %r543;
shr.u32 %r542, %r43, %r25;

$L__BB117_3:
shl.b32 %r44, %r1, 5;
add.s32 %r45, %r44, %r2;
shr.u32 %r46, %r45, 31;
add.s32 %r47, %r45, %r46;
shr.s32 %r545, %r47, 1;
setp.ge.s32 %p2, %r545, %r4;
mov.f32 %f138, 0f00000000;
mov.f32 %f139, %f138;
mov.f32 %f140, %f138;
mov.f32 %f141, %f138;
@%p2 bra $L__BB117_6;

mov.u32 %r48, %ctaid.z;
and.b32 %r55, %r47, -2;
sub.s32 %r56, %r45, %r55;
shl.b32 %r57, %r56, 2;
shl.b32 %r58, %r56, 1;
cvt.s64.s32 %rd1, %r56;
cvt.s64.s32 %rd20, %r58;
or.b32 %r59, %r57, 2;
shr.s32 %r60, %r59, 1;
shr.s32 %r61, %r59, 2;
cvt.s64.s32 %rd2, %r61;
cvt.s64.s32 %rd21, %r60;
shl.b32 %r544, %r545, 1;
mad.lo.s32 %r62, %r3, %r22, %r545;
mul.hi.u32 %r63, %r48, %r30;
add.s32 %r64, %r63, %r48;
shr.u32 %r65, %r64, %r31;
mad.lo.s32 %r66, %r65, %r33, %r62;
mad.lo.s32 %r67, %r542, %r27, %r66;
mul.wide.s32 %rd22, %r67, 36;
mul.wide.s32 %rd23, %r57, 4;
add.s64 %rd24, %rd22, %rd23;
add.s64 %rd3, %rd24, 16;
add.s64 %rd4, %rd22, %rd20;
add.s64 %rd5, %rd22, %rd21;
mul.wide.s32 %rd6, %r22, 36;
cvta.to.global.u64 %rd50, %rd13;
cvta.to.global.u64 %rd8, %rd14;
mul.lo.s32 %r68, %r543, %r28;
cvt.u64.u32 %rd25, %r68;
mul.lo.s32 %r69, %r48, %r34;
cvt.u64.u32 %rd26, %r69;
add.s64 %rd9, %rd25, %rd26;
mov.f32 %f138, 0f00000000;

$L__BB117_5:
cvt.u64.u32 %rd27, %r544;
add.s64 %rd28, %rd9, %rd27;
add.s64 %rd29, %rd50, %rd3;
add.s64 %rd30, %rd29, -12;
ld.global.nc.u32 %r214, [%rd29+-12];
and.b32 %r215, %r214, -2004318072;
shr.u32 %r216, %r215, 1;
or.b32 %r217, %r216, 839922192;
and.b32 %r218, %r214, 30583;
mov.u32 %r219, 201852420;
mov.u32 %r220, 50462976;
prmt.b32 %r221, %r220, %r219, %r218;
mov.u32 %r222, -185009412;
mov.u32 %r223, -33620224;
prmt.b32 %r224, %r223, %r222, %r218;
and.b32 %r225, %r217, 30292;
prmt.b32 %r226, %r221, %r224, %r225;
shr.u32 %r227, %r214, 16;
and.b32 %r228, %r227, 30583;
prmt.b32 %r229, %r220, %r219, %r228;
prmt.b32 %r230, %r223, %r222, %r228;
shr.u32 %r231, %r217, 16;
prmt.b32 %r232, %r229, %r230, %r231;
mov.u32 %r233, 25632;
prmt.b32 %r71, %r226, %r232, %r233;
mov.u32 %r234, 30001;
prmt.b32 %r75, %r226, %r232, %r234;
ld.global.nc.u32 %r235, [%rd29+-8];
and.b32 %r236, %r235, -2004318072;
shr.u32 %r237, %r236, 1;
or.b32 %r238, %r237, 839922192;
and.b32 %r239, %r235, 30583;
prmt.b32 %r240, %r220, %r219, %r239;
prmt.b32 %r241, %r223, %r222, %r239;
and.b32 %r242, %r238, 30292;
prmt.b32 %r243, %r240, %r241, %r242;
shr.u32 %r244, %r235, 16;
and.b32 %r245, %r244, 30583;
prmt.b32 %r246, %r220, %r219, %r245;
prmt.b32 %r247, %r223, %r222, %r245;
shr.u32 %r248, %r238, 16;
prmt.b32 %r249, %r246, %r247, %r248;
prmt.b32 %r79, %r243, %r249, %r233;
prmt.b32 %r83, %r243, %r249, %r234;
add.s64 %rd31, %rd28, %rd1;
mul.lo.s64 %rd32, %rd31, 36;
add.s64 %rd33, %rd8, %rd32;
ld.global.nc.u32 %r72, [%rd33+4];
mov.u32 %r199, 0;

dp4a.s32.s32 %r70, %r71, %r72, %r199;

ld.global.nc.u32 %r76, [%rd33+12];

dp4a.s32.s32 %r74, %r75, %r76, %r70;

ld.global.nc.u32 %r80, [%rd33+8];

dp4a.s32.s32 %r78, %r79, %r80, %r74;

ld.global.nc.u32 %r84, [%rd33+16];

dp4a.s32.s32 %r82, %r83, %r84, %r78;

add.s64 %rd34, %rd50, %rd4;
ld.global.nc.u8 %rs17, [%rd34];
and.b16 %rs18, %rs17, 127;
setp.eq.s16 %p3, %rs18, 127;
selp.b16 %rs19, 0, %rs17, %p3;
and.b16 %rs1, %rs19, 255;

{cvt.rn.f16x2.e4m3x2 %r86, %rs1;}


cvt.u16.u32 %rs2, %r86;

{cvt.f32.f16 %f25, %rs2;}


mov.f32 %f41, 0f40000000;
div.approx.ftz.f32 %f42, %f25, %f41;
ld.global.nc.u32 %r87, [%rd33];

{.reg .f16 low,high;
mov.b32 {low,high},%r87;
cvt.f32.f16 %f26, low;}


mul.ftz.f32 %f43, %f42, %f26;
cvt.rn.f32.s32 %f44, %r82;
fma.rn.ftz.f32 %f45, %f43, %f44, 0f00000000;
ld.global.nc.u32 %r250, [%rd29+-4];
and.b32 %r251, %r250, -2004318072;
shr.u32 %r252, %r251, 1;
or.b32 %r253, %r252, 839922192;
and.b32 %r254, %r250, 30583;
prmt.b32 %r255, %r220, %r219, %r254;
prmt.b32 %r256, %r223, %r222, %r254;
and.b32 %r257, %r253, 30292;
prmt.b32 %r258, %r255, %r256, %r257;
shr.u32 %r259, %r250, 16;
and.b32 %r260, %r259, 30583;
prmt.b32 %r261, %r220, %r219, %r260;
prmt.b32 %r262, %r223, %r222, %r260;
shr.u32 %r263, %r253, 16;
prmt.b32 %r264, %r261, %r262, %r263;
prmt.b32 %r89, %r258, %r264, %r233;
prmt.b32 %r93, %r258, %r264, %r234;
ld.global.nc.u32 %r265, [%rd29];
and.b32 %r266, %r265, -2004318072;
shr.u32 %r267, %r266, 1;
or.b32 %r268, %r267, 839922192;
and.b32 %r269, %r265, 30583;
prmt.b32 %r270, %r220, %r219, %r269;
prmt.b32 %r271, %r223, %r222, %r269;
and.b32 %r272, %r268, 30292;
prmt.b32 %r273, %r270, %r271, %r272;
shr.u32 %r274, %r265, 16;
and.b32 %r275, %r274, 30583;
prmt.b32 %r276, %r220, %r219, %r275;
prmt.b32 %r277, %r223, %r222, %r275;
shr.u32 %r278, %r268, 16;
prmt.b32 %r279, %r276, %r277, %r278;
prmt.b32 %r97, %r273, %r279, %r233;
prmt.b32 %r101, %r273, %r279, %r234;
add.s64 %rd35, %rd28, %rd2;
mul.lo.s64 %rd36, %rd35, 36;
add.s64 %rd37, %rd8, %rd36;
ld.global.nc.u32 %r90, [%rd37+20];

dp4a.s32.s32 %r88, %r89, %r90, %r199;

ld.global.nc.u32 %r94, [%rd37+28];

dp4a.s32.s32 %r92, %r93, %r94, %r88;

ld.global.nc.u32 %r98, [%rd37+24];

dp4a.s32.s32 %r96, %r97, %r98, %r92;

ld.global.nc.u32 %r102, [%rd37+32];

dp4a.s32.s32 %r100, %r101, %r102, %r96;

add.s64 %rd38, %rd50, %rd5;
ld.global.nc.u8 %rs20, [%rd38];
and.b16 %rs21, %rs20, 127;
setp.eq.s16 %p4, %rs21, 127;
selp.b16 %rs22, 0, %rs20, %p4;
and.b16 %rs3, %rs22, 255;

{cvt.rn.f16x2.e4m3x2 %r104, %rs3;}


cvt.u16.u32 %rs4, %r104;

{cvt.f32.f16 %f27, %rs4;}


div.approx.ftz.f32 %f46, %f27, %f41;
ld.global.nc.u32 %r105, [%rd37];

{.reg .f16 low,high;
mov.b32 {low,high},%r105;
cvt.f32.f16 %f28, low;}


mul.ftz.f32 %f47, %f46, %f28;
cvt.rn.f32.s32 %f48, %r100;
fma.rn.ftz.f32 %f49, %f47, %f48, %f45;
add.ftz.f32 %f141, %f49, %f141;
add.s64 %rd39, %rd30, %rd6;
ld.global.nc.u32 %r280, [%rd39];
and.b32 %r281, %r280, -2004318072;
shr.u32 %r282, %r281, 1;
or.b32 %r283, %r282, 839922192;
and.b32 %r284, %r280, 30583;
prmt.b32 %r285, %r220, %r219, %r284;
prmt.b32 %r286, %r223, %r222, %r284;
and.b32 %r287, %r283, 30292;
prmt.b32 %r288, %r285, %r286, %r287;
shr.u32 %r289, %r280, 16;
and.b32 %r290, %r289, 30583;
prmt.b32 %r291, %r220, %r219, %r290;
prmt.b32 %r292, %r223, %r222, %r290;
shr.u32 %r293, %r283, 16;
prmt.b32 %r294, %r291, %r292, %r293;
prmt.b32 %r107, %r288, %r294, %r233;
prmt.b32 %r111, %r288, %r294, %r234;
ld.global.nc.u32 %r295, [%rd39+4];
and.b32 %r296, %r295, -2004318072;
shr.u32 %r297, %r296, 1;
or.b32 %r298, %r297, 839922192;
and.b32 %r299, %r295, 30583;
prmt.b32 %r300, %r220, %r219, %r299;
prmt.b32 %r301, %r223, %r222, %r299;
and.b32 %r302, %r298, 30292;
prmt.b32 %r303, %r300, %r301, %r302;
shr.u32 %r304, %r295, 16;
and.b32 %r305, %r304, 30583;
prmt.b32 %r306, %r220, %r219, %r305;
prmt.b32 %r307, %r223, %r222, %r305;
shr.u32 %r308, %r298, 16;
prmt.b32 %r309, %r306, %r307, %r308;
prmt.b32 %r115, %r303, %r309, %r233;
prmt.b32 %r119, %r303, %r309, %r234;

dp4a.s32.s32 %r106, %r107, %r72, %r199;


dp4a.s32.s32 %r110, %r111, %r76, %r106;


dp4a.s32.s32 %r114, %r115, %r80, %r110;


dp4a.s32.s32 %r118, %r119, %r84, %r114;

add.s64 %rd40, %rd34, %rd6;
ld.global.nc.u8 %rs23, [%rd40];
and.b16 %rs24, %rs23, 127;
setp.eq.s16 %p5, %rs24, 127;
selp.b16 %rs25, 0, %rs23, %p5;
and.b16 %rs5, %rs25, 255;

{cvt.rn.f16x2.e4m3x2 %r122, %rs5;}


cvt.u16.u32 %rs6, %r122;

{cvt.f32.f16 %f29, %rs6;}


div.approx.ftz.f32 %f50, %f29, %f41;

{.reg .f16 low,high;
mov.b32 {low,high},%r87;
cvt.f32.f16 %f30, low;}


mul.ftz.f32 %f51, %f50, %f30;
cvt.rn.f32.s32 %f52, %r118;
fma.rn.ftz.f32 %f53, %f51, %f52, 0f00000000;
ld.global.nc.u32 %r310, [%rd39+8];
and.b32 %r311, %r310, -2004318072;
shr.u32 %r312, %r311, 1;
or.b32 %r313, %r312, 839922192;
and.b32 %r314, %r310, 30583;
prmt.b32 %r315, %r220, %r219, %r314;
prmt.b32 %r316, %r223, %r222, %r314;
and.b32 %r317, %r313, 30292;
prmt.b32 %r318, %r315, %r316, %r317;
shr.u32 %r319, %r310, 16;
and.b32 %r320, %r319, 30583;
prmt.b32 %r321, %r220, %r219, %r320;
prmt.b32 %r322, %r223, %r222, %r320;
shr.u32 %r323, %r313, 16;
prmt.b32 %r324, %r321, %r322, %r323;
prmt.b32 %r125, %r318, %r324, %r233;
prmt.b32 %r129, %r318, %r324, %r234;
ld.global.nc.u32 %r325, [%rd39+12];
and.b32 %r326, %r325, -2004318072;
shr.u32 %r327, %r326, 1;
or.b32 %r328, %r327, 839922192;
and.b32 %r329, %r325, 30583;
prmt.b32 %r330, %r220, %r219, %r329;
prmt.b32 %r331, %r223, %r222, %r329;
and.b32 %r332, %r328, 30292;
prmt.b32 %r333, %r330, %r331, %r332;
shr.u32 %r334, %r325, 16;
and.b32 %r335, %r334, 30583;
prmt.b32 %r336, %r220, %r219, %r335;
prmt.b32 %r337, %r223, %r222, %r335;
shr.u32 %r338, %r328, 16;
prmt.b32 %r339, %r336, %r337, %r338;
prmt.b32 %r133, %r333, %r339, %r233;
prmt.b32 %r137, %r333, %r339, %r234;

dp4a.s32.s32 %r124, %r125, %r90, %r199;


dp4a.s32.s32 %r128, %r129, %r94, %r124;


dp4a.s32.s32 %r132, %r133, %r98, %r128;


dp4a.s32.s32 %r136, %r137, %r102, %r132;

add.s64 %rd41, %rd38, %rd6;
ld.global.nc.u8 %rs26, [%rd41];
and.b16 %rs27, %rs26, 127;
setp.eq.s16 %p6, %rs27, 127;
selp.b16 %rs28, 0, %rs26, %p6;
and.b16 %rs7, %rs28, 255;

{cvt.rn.f16x2.e4m3x2 %r140, %rs7;}


cvt.u16.u32 %rs8, %r140;

{cvt.f32.f16 %f31, %rs8;}


div.approx.ftz.f32 %f54, %f31, %f41;

{.reg .f16 low,high;
mov.b32 {low,high},%r105;
cvt.f32.f16 %f32, low;}


mul.ftz.f32 %f55, %f54, %f32;
cvt.rn.f32.s32 %f56, %r136;
fma.rn.ftz.f32 %f57, %f55, %f56, %f53;
add.ftz.f32 %f140, %f57, %f140;
add.s64 %rd42, %rd39, %rd6;
ld.global.nc.u32 %r340, [%rd42];
and.b32 %r341, %r340, -2004318072;
shr.u32 %r342, %r341, 1;
or.b32 %r343, %r342, 839922192;
and.b32 %r344, %r340, 30583;
prmt.b32 %r345, %r220, %r219, %r344;
prmt.b32 %r346, %r223, %r222, %r344;
and.b32 %r347, %r343, 30292;
prmt.b32 %r348, %r345, %r346, %r347;
shr.u32 %r349, %r340, 16;
and.b32 %r350, %r349, 30583;
prmt.b32 %r351, %r220, %r219, %r350;
prmt.b32 %r352, %r223, %r222, %r350;
shr.u32 %r353, %r343, 16;
prmt.b32 %r354, %r351, %r352, %r353;
prmt.b32 %r143, %r348, %r354, %r233;
prmt.b32 %r147, %r348, %r354, %r234;
ld.global.nc.u32 %r355, [%rd42+4];
and.b32 %r356, %r355, -2004318072;
shr.u32 %r357, %r356, 1;
or.b32 %r358, %r357, 839922192;
and.b32 %r359, %r355, 30583;
prmt.b32 %r360, %r220, %r219, %r359;
prmt.b32 %r361, %r223, %r222, %r359;
and.b32 %r362, %r358, 30292;
prmt.b32 %r363, %r360, %r361, %r362;
shr.u32 %r364, %r355, 16;
and.b32 %r365, %r364, 30583;
prmt.b32 %r366, %r220, %r219, %r365;
prmt.b32 %r367, %r223, %r222, %r365;
shr.u32 %r368, %r358, 16;
prmt.b32 %r369, %r366, %r367, %r368;
prmt.b32 %r151, %r363, %r369, %r233;
prmt.b32 %r155, %r363, %r369, %r234;

dp4a.s32.s32 %r142, %r143, %r72, %r199;


dp4a.s32.s32 %r146, %r147, %r76, %r142;


dp4a.s32.s32 %r150, %r151, %r80, %r146;


dp4a.s32.s32 %r154, %r155, %r84, %r150;

add.s64 %rd43, %rd40, %rd6;
ld.global.nc.u8 %rs29, [%rd43];
and.b16 %rs30, %rs29, 127;
setp.eq.s16 %p7, %rs30, 127;
selp.b16 %rs31, 0, %rs29, %p7;
and.b16 %rs9, %rs31, 255;

{cvt.rn.f16x2.e4m3x2 %r158, %rs9;}


cvt.u16.u32 %rs10, %r158;

{cvt.f32.f16 %f33, %rs10;}


div.approx.ftz.f32 %f58, %f33, %f41;

{.reg .f16 low,high;
mov.b32 {low,high},%r87;
cvt.f32.f16 %f34, low;}


mul.ftz.f32 %f59, %f58, %f34;
cvt.rn.f32.s32 %f60, %r154;
fma.rn.ftz.f32 %f61, %f59, %f60, 0f00000000;
ld.global.nc.u32 %r370, [%rd42+8];
and.b32 %r371, %r370, -2004318072;
shr.u32 %r372, %r371, 1;
or.b32 %r373, %r372, 839922192;
and.b32 %r374, %r370, 30583;
prmt.b32 %r375, %r220, %r219, %r374;
prmt.b32 %r376, %r223, %r222, %r374;
and.b32 %r377, %r373, 30292;
prmt.b32 %r378, %r375, %r376, %r377;
shr.u32 %r379, %r370, 16;
and.b32 %r380, %r379, 30583;
prmt.b32 %r381, %r220, %r219, %r380;
prmt.b32 %r382, %r223, %r222, %r380;
shr.u32 %r383, %r373, 16;
prmt.b32 %r384, %r381, %r382, %r383;
prmt.b32 %r161, %r378, %r384, %r233;
prmt.b32 %r165, %r378, %r384, %r234;
ld.global.nc.u32 %r385, [%rd42+12];
and.b32 %r386, %r385, -2004318072;
shr.u32 %r387, %r386, 1;
or.b32 %r388, %r387, 839922192;
and.b32 %r389, %r385, 30583;
prmt.b32 %r390, %r220, %r219, %r389;
prmt.b32 %r391, %r223, %r222, %r389;
and.b32 %r392, %r388, 30292;
prmt.b32 %r393, %r390, %r391, %r392;
shr.u32 %r394, %r385, 16;
and.b32 %r395, %r394, 30583;
prmt.b32 %r396, %r220, %r219, %r395;
prmt.b32 %r397, %r223, %r222, %r395;
shr.u32 %r398, %r388, 16;
prmt.b32 %r399, %r396, %r397, %r398;
prmt.b32 %r169, %r393, %r399, %r233;
prmt.b32 %r173, %r393, %r399, %r234;

dp4a.s32.s32 %r160, %r161, %r90, %r199;


dp4a.s32.s32 %r164, %r165, %r94, %r160;


dp4a.s32.s32 %r168, %r169, %r98, %r164;


dp4a.s32.s32 %r172, %r173, %r102, %r168;

add.s64 %rd44, %rd41, %rd6;
ld.global.nc.u8 %rs32, [%rd44];
and.b16 %rs33, %rs32, 127;
setp.eq.s16 %p8, %rs33, 127;
selp.b16 %rs34, 0, %rs32, %p8;
and.b16 %rs11, %rs34, 255;

{cvt.rn.f16x2.e4m3x2 %r176, %rs11;}


cvt.u16.u32 %rs12, %r176;

{cvt.f32.f16 %f35, %rs12;}


div.approx.ftz.f32 %f62, %f35, %f41;

{.reg .f16 low,high;
mov.b32 {low,high},%r105;
cvt.f32.f16 %f36, low;}


mul.ftz.f32 %f63, %f62, %f36;
cvt.rn.f32.s32 %f64, %r172;
fma.rn.ftz.f32 %f65, %f63, %f64, %f61;
add.ftz.f32 %f139, %f65, %f139;
add.s64 %rd45, %rd42, %rd6;
ld.global.nc.u32 %r400, [%rd45];
and.b32 %r401, %r400, -2004318072;
shr.u32 %r402, %r401, 1;
or.b32 %r403, %r402, 839922192;
and.b32 %r404, %r400, 30583;
prmt.b32 %r405, %r220, %r219, %r404;
prmt.b32 %r406, %r223, %r222, %r404;
and.b32 %r407, %r403, 30292;
prmt.b32 %r408, %r405, %r406, %r407;
shr.u32 %r409, %r400, 16;
and.b32 %r410, %r409, 30583;
prmt.b32 %r411, %r220, %r219, %r410;
prmt.b32 %r412, %r223, %r222, %r410;
shr.u32 %r413, %r403, 16;
prmt.b32 %r414, %r411, %r412, %r413;
prmt.b32 %r179, %r408, %r414, %r233;
prmt.b32 %r183, %r408, %r414, %r234;
ld.global.nc.u32 %r415, [%rd45+4];
and.b32 %r416, %r415, -2004318072;
shr.u32 %r417, %r416, 1;
or.b32 %r418, %r417, 839922192;
and.b32 %r419, %r415, 30583;
prmt.b32 %r420, %r220, %r219, %r419;
prmt.b32 %r421, %r223, %r222, %r419;
and.b32 %r422, %r418, 30292;
prmt.b32 %r423, %r420, %r421, %r422;
shr.u32 %r424, %r415, 16;
and.b32 %r425, %r424, 30583;
prmt.b32 %r426, %r220, %r219, %r425;
prmt.b32 %r427, %r223, %r222, %r425;
shr.u32 %r428, %r418, 16;
prmt.b32 %r429, %r426, %r427, %r428;
prmt.b32 %r187, %r423, %r429, %r233;
prmt.b32 %r191, %r423, %r429, %r234;

dp4a.s32.s32 %r178, %r179, %r72, %r199;


dp4a.s32.s32 %r182, %r183, %r76, %r178;


dp4a.s32.s32 %r186, %r187, %r80, %r182;


dp4a.s32.s32 %r190, %r191, %r84, %r186;

add.s64 %rd46, %rd43, %rd6;
ld.global.nc.u8 %rs35, [%rd46];
and.b16 %rs36, %rs35, 127;
setp.eq.s16 %p9, %rs36, 127;
selp.b16 %rs37, 0, %rs35, %p9;
and.b16 %rs13, %rs37, 255;

{cvt.rn.f16x2.e4m3x2 %r194, %rs13;}


cvt.u16.u32 %rs14, %r194;

{cvt.f32.f16 %f37, %rs14;}


div.approx.ftz.f32 %f66, %f37, %f41;

{.reg .f16 low,high;
mov.b32 {low,high},%r87;
cvt.f32.f16 %f38, low;}


mul.ftz.f32 %f67, %f66, %f38;
cvt.rn.f32.s32 %f68, %r190;
fma.rn.ftz.f32 %f69, %f67, %f68, 0f00000000;
ld.global.nc.u32 %r430, [%rd45+8];
and.b32 %r431, %r430, -2004318072;
shr.u32 %r432, %r431, 1;
or.b32 %r433, %r432, 839922192;
and.b32 %r434, %r430, 30583;
prmt.b32 %r435, %r220, %r219, %r434;
prmt.b32 %r436, %r223, %r222, %r434;
and.b32 %r437, %r433, 30292;
prmt.b32 %r438, %r435, %r436, %r437;
shr.u32 %r439, %r430, 16;
and.b32 %r440, %r439, 30583;
prmt.b32 %r441, %r220, %r219, %r440;
prmt.b32 %r442, %r223, %r222, %r440;
shr.u32 %r443, %r433, 16;
prmt.b32 %r444, %r441, %r442, %r443;
prmt.b32 %r197, %r438, %r444, %r233;
prmt.b32 %r201, %r438, %r444, %r234;
ld.global.nc.u32 %r445, [%rd45+12];
and.b32 %r446, %r445, -2004318072;
shr.u32 %r447, %r446, 1;
or.b32 %r448, %r447, 839922192;
and.b32 %r449, %r445, 30583;
prmt.b32 %r450, %r220, %r219, %r449;
prmt.b32 %r451, %r223, %r222, %r449;
and.b32 %r452, %r448, 30292;
prmt.b32 %r453, %r450, %r451, %r452;
shr.u32 %r454, %r445, 16;
and.b32 %r455, %r454, 30583;
prmt.b32 %r456, %r220, %r219, %r455;
prmt.b32 %r457, %r223, %r222, %r455;
shr.u32 %r458, %r448, 16;
prmt.b32 %r459, %r456, %r457, %r458;
prmt.b32 %r205, %r453, %r459, %r233;
prmt.b32 %r209, %r453, %r459, %r234;

dp4a.s32.s32 %r196, %r197, %r90, %r199;


dp4a.s32.s32 %r200, %r201, %r94, %r196;


dp4a.s32.s32 %r204, %r205, %r98, %r200;


dp4a.s32.s32 %r208, %r209, %r102, %r204;

add.s64 %rd47, %rd44, %rd6;
ld.global.nc.u8 %rs38, [%rd47];
and.b16 %rs39, %rs38, 127;
setp.eq.s16 %p10, %rs39, 127;
selp.b16 %rs40, 0, %rs38, %p10;
and.b16 %rs15, %rs40, 255;

{cvt.rn.f16x2.e4m3x2 %r212, %rs15;}


cvt.u16.u32 %rs16, %r212;

{cvt.f32.f16 %f39, %rs16;}


div.approx.ftz.f32 %f70, %f39, %f41;

{.reg .f16 low,high;
mov.b32 {low,high},%r105;
cvt.f32.f16 %f40, low;}


mul.ftz.f32 %f71, %f70, %f40;
cvt.rn.f32.s32 %f72, %r208;
fma.rn.ftz.f32 %f73, %f71, %f72, %f69;
add.ftz.f32 %f138, %f73, %f138;
add.s32 %r544, %r544, 128;
add.s64 %rd50, %rd50, 2304;
add.s32 %r545, %r545, 64;
setp.lt.s32 %p11, %r545, %r4;
@%p11 bra $L__BB117_5;

$L__BB117_6:
setp.eq.s32 %p12, %r1, 0;
@%p12 bra $L__BB117_8;

shl.b32 %r460, %r1, 9;
mov.u32 %r461, _ZZ13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r462, %r461, %r460;
shl.b32 %r463, %r2, 2;
add.s32 %r464, %r462, %r463;
st.shared.f32 [%r464+-512], %f141;
st.shared.f32 [%r464+-384], %f140;
st.shared.f32 [%r464+-256], %f139;
st.shared.f32 [%r464+-128], %f138;

$L__BB117_8:
setp.ne.s32 %p13, %r1, 0;
bar.sync 0;
@%p13 bra $L__BB117_17;

mov.u32 %r465, %ctaid.z;
mov.u32 %r466, %ctaid.y;
mad.lo.s32 %r467, %r466, %r29, %r3;
mad.lo.s32 %r468, %r465, %r35, %r467;
shl.b32 %r469, %r2, 2;
mov.u32 %r470, _ZZ13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r17, %r470, %r469;
mov.u32 %r471, 2;
ld.shared.f32 %f74, [%r17];
add.ftz.f32 %f75, %f74, %f141;
ld.shared.f32 %f76, [%r17+512];
add.ftz.f32 %f77, %f76, %f75;
ld.shared.f32 %f78, [%r17+1024];
add.ftz.f32 %f79, %f78, %f77;
mov.b32 %r472, %f79;
mov.u32 %r473, 31;
mov.u32 %r474, 16;
mov.u32 %r475, -1;
shfl.sync.bfly.b32 %r476|%p14, %r472, %r474, %r473, %r475;
mov.b32 %f80, %r476;
add.ftz.f32 %f81, %f79, %f80;
mov.b32 %r477, %f81;
mov.u32 %r478, 8;
shfl.sync.bfly.b32 %r479|%p15, %r477, %r478, %r473, %r475;
mov.b32 %f82, %r479;
add.ftz.f32 %f83, %f81, %f82;
mov.b32 %r480, %f83;
mov.u32 %r481, 4;
shfl.sync.bfly.b32 %r482|%p16, %r480, %r481, %r473, %r475;
mov.b32 %f84, %r482;
add.ftz.f32 %f85, %f83, %f84;
mov.b32 %r483, %f85;
shfl.sync.bfly.b32 %r484|%p17, %r483, %r471, %r473, %r475;
mov.b32 %f86, %r484;
add.ftz.f32 %f87, %f85, %f86;
mov.b32 %r485, %f87;
mov.u32 %r486, 1;
shfl.sync.bfly.b32 %r487|%p18, %r485, %r486, %r473, %r475;
mov.b32 %f88, %r487;
add.ftz.f32 %f13, %f87, %f88;
setp.ge.u32 %p19, %r3, %r23;
setp.ne.s32 %p20, %r2, 0;
cvta.to.global.u64 %rd48, %rd16;
mul.wide.u32 %rd49, %r468, 4;
add.s64 %rd12, %rd48, %rd49;
or.pred %p21, %p20, %p19;
@%p21 bra $L__BB117_11;

st.global.f32 [%rd12], %f13;

$L__BB117_11:
ld.shared.f32 %f89, [%r17+128];
add.ftz.f32 %f90, %f89, %f140;
ld.shared.f32 %f91, [%r17+640];
add.ftz.f32 %f92, %f91, %f90;
ld.shared.f32 %f93, [%r17+1152];
add.ftz.f32 %f94, %f93, %f92;
mov.b32 %r488, %f94;
shfl.sync.bfly.b32 %r492|%p22, %r488, %r474, %r473, %r475;
mov.b32 %f95, %r492;
add.ftz.f32 %f96, %f94, %f95;
mov.b32 %r493, %f96;
shfl.sync.bfly.b32 %r495|%p23, %r493, %r478, %r473, %r475;
mov.b32 %f97, %r495;
add.ftz.f32 %f98, %f96, %f97;
mov.b32 %r496, %f98;
shfl.sync.bfly.b32 %r498|%p24, %r496, %r481, %r473, %r475;
mov.b32 %f99, %r498;
add.ftz.f32 %f100, %f98, %f99;
mov.b32 %r499, %f100;
shfl.sync.bfly.b32 %r501|%p25, %r499, %r471, %r473, %r475;
mov.b32 %f101, %r501;
add.ftz.f32 %f102, %f100, %f101;
mov.b32 %r502, %f102;
shfl.sync.bfly.b32 %r504|%p26, %r502, %r486, %r473, %r475;
mov.b32 %f103, %r504;
add.ftz.f32 %f14, %f102, %f103;
add.s32 %r505, %r3, 1;
setp.ge.u32 %p27, %r505, %r23;
setp.ne.s32 %p28, %r2, 1;
or.pred %p29, %p28, %p27;
@%p29 bra $L__BB117_13;

st.global.f32 [%rd12+4], %f14;

$L__BB117_13:
ld.shared.f32 %f104, [%r17+256];
add.ftz.f32 %f105, %f104, %f139;
ld.shared.f32 %f106, [%r17+768];
add.ftz.f32 %f107, %f106, %f105;
ld.shared.f32 %f108, [%r17+1280];
add.ftz.f32 %f109, %f108, %f107;
mov.b32 %r506, %f109;
mov.u32 %r507, 31;
mov.u32 %r508, 16;
mov.u32 %r509, -1;
shfl.sync.bfly.b32 %r510|%p30, %r506, %r508, %r507, %r509;
mov.b32 %f110, %r510;
add.ftz.f32 %f111, %f109, %f110;
mov.b32 %r511, %f111;
mov.u32 %r512, 8;
shfl.sync.bfly.b32 %r513|%p31, %r511, %r512, %r507, %r509;
mov.b32 %f112, %r513;
add.ftz.f32 %f113, %f111, %f112;
mov.b32 %r514, %f113;
mov.u32 %r515, 4;
shfl.sync.bfly.b32 %r516|%p32, %r514, %r515, %r507, %r509;
mov.b32 %f114, %r516;
add.ftz.f32 %f115, %f113, %f114;
mov.b32 %r517, %f115;
mov.u32 %r518, 2;
shfl.sync.bfly.b32 %r519|%p33, %r517, %r518, %r507, %r509;
mov.b32 %f116, %r519;
add.ftz.f32 %f117, %f115, %f116;
mov.b32 %r520, %f117;
mov.u32 %r521, 1;
shfl.sync.bfly.b32 %r522|%p34, %r520, %r521, %r507, %r509;
mov.b32 %f118, %r522;
add.ftz.f32 %f15, %f117, %f118;
add.s32 %r523, %r3, 2;
setp.ge.u32 %p35, %r523, %r23;
setp.ne.s32 %p36, %r2, 2;
or.pred %p37, %p36, %p35;
@%p37 bra $L__BB117_15;

st.global.f32 [%rd12+8], %f15;

$L__BB117_15:
ld.shared.f32 %f119, [%r17+384];
add.ftz.f32 %f120, %f119, %f138;
ld.shared.f32 %f121, [%r17+896];
add.ftz.f32 %f122, %f121, %f120;
ld.shared.f32 %f123, [%r17+1408];
add.ftz.f32 %f124, %f123, %f122;
mov.b32 %r524, %f124;
shfl.sync.bfly.b32 %r528|%p38, %r524, %r508, %r507, %r509;
mov.b32 %f125, %r528;
add.ftz.f32 %f126, %f124, %f125;
mov.b32 %r529, %f126;
shfl.sync.bfly.b32 %r531|%p39, %r529, %r512, %r507, %r509;
mov.b32 %f127, %r531;
add.ftz.f32 %f128, %f126, %f127;
mov.b32 %r532, %f128;
shfl.sync.bfly.b32 %r534|%p40, %r532, %r515, %r507, %r509;
mov.b32 %f129, %r534;
add.ftz.f32 %f130, %f128, %f129;
mov.b32 %r535, %f130;
shfl.sync.bfly.b32 %r537|%p41, %r535, %r518, %r507, %r509;
mov.b32 %f131, %r537;
add.ftz.f32 %f132, %f130, %f131;
mov.b32 %r538, %f132;
shfl.sync.bfly.b32 %r540|%p42, %r538, %r521, %r507, %r509;
mov.b32 %f133, %r540;
add.ftz.f32 %f16, %f132, %f133;
add.s32 %r541, %r3, 3;
setp.ge.u32 %p43, %r541, %r23;
setp.ne.s32 %p44, %r2, 3;
or.pred %p45, %p44, %p43;
@%p45 bra $L__BB117_17;

st.global.f32 [%rd12+12], %f16;

$L__BB117_17:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<45>;
.reg .b16 %rs<19>;
.reg .f32 %f<153>;
.reg .b32 %r<381>;
.reg .b64 %rd<90>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[384];

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate[384];

ld.param.u64 %rd10, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd11, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd12, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r45, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+32];
ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+24];
ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+16];
ld.param.u64 %rd14, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+8];
ld.param.u64 %rd13, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3];
mov.u32 %r1, %tid.y;
mov.u32 %r2, %ctaid.y;
setp.eq.s64 %p3, %rd12, 0;
@%p3 bra $L__BB118_2;

cvta.to.global.u64 %rd19, %rd12;
mul.wide.u32 %rd20, %r2, 4;
add.s64 %rd21, %rd19, %rd20;
ld.global.nc.u32 %r377, [%rd21];
mul.hi.u32 %r46, %r2, %r30;
add.s32 %r47, %r46, %r2;
shr.u32 %r48, %r47, %r31;
mul.lo.s32 %r49, %r48, %r32;
sub.s32 %r378, %r2, %r49;
bra.uni $L__BB118_3;

$L__BB118_2:
mul.hi.u32 %r50, %r2, %r34;
add.s32 %r51, %r50, %r2;
shr.u32 %r377, %r51, %r35;
mov.u32 %r378, %r2;

$L__BB118_3:
mov.u32 %r52, %tid.x;
or.b32 %r53, %r52, %r1;
setp.ne.s32 %p4, %r53, 0;
mov.f32 %f142, 0f3F800000;
mov.f32 %f146, 0f00000000;
mov.f32 %f140, %f146;
mov.f32 %f141, %f146;
mov.f32 %f143, %f142;
@%p4 bra $L__BB118_12;

setp.eq.s64 %p5, %rd13, 0;
mov.f32 %f140, 0f00000000;
mov.f32 %f141, %f140;
@%p5 bra $L__BB118_6;

cvta.to.global.u64 %rd22, %rd13;
mov.u32 %r54, %ctaid.z;
mul.lo.s32 %r55, %r54, %r45;
cvt.u64.u32 %rd23, %r55;
selp.b32 %r56, %r2, %r377, %p3;
mul.lo.s32 %r57, %r56, %r39;
cvt.u64.u32 %rd24, %r57;
mov.u32 %r58, %ctaid.x;
cvt.s64.s32 %rd25, %r58;
cvt.u64.u32 %rd26, %r52;
add.s64 %rd27, %rd26, %rd25;
add.s64 %rd28, %rd27, %rd23;
add.s64 %rd29, %rd28, %rd24;
shl.b64 %rd30, %rd29, 2;
add.s64 %rd31, %rd22, %rd30;
ld.global.f32 %f141, [%rd31];

$L__BB118_6:
setp.ne.s64 %p1, %rd14, 0;
setp.eq.s64 %p7, %rd14, 0;
setp.eq.s64 %p8, %rd15, 0;
or.pred %p9, %p8, %p7;
@%p9 bra $L__BB118_8;

cvta.to.global.u64 %rd32, %rd15;
mov.u32 %r60, %ctaid.z;
mul.lo.s32 %r61, %r60, %r45;
cvt.u64.u32 %rd33, %r61;
selp.b32 %r62, %r2, %r377, %p3;
mul.lo.s32 %r63, %r62, %r39;
cvt.u64.u32 %rd34, %r63;
mov.u32 %r64, %ctaid.x;
cvt.s64.s32 %rd35, %r64;
cvt.u64.u32 %rd36, %r52;
add.s64 %rd37, %rd36, %rd35;
add.s64 %rd38, %rd37, %rd33;
add.s64 %rd39, %rd38, %rd34;
shl.b64 %rd40, %rd39, 2;
add.s64 %rd41, %rd32, %rd40;
ld.global.f32 %f140, [%rd41];

$L__BB118_8:
setp.eq.s64 %p11, %rd16, 0;
mov.f32 %f143, 0f3F800000;
mov.f32 %f142, %f143;
@%p11 bra $L__BB118_10;

cvta.to.global.u64 %rd42, %rd16;
cvt.u64.u32 %rd43, %r377;
selp.b64 %rd44, 0, %rd43, %p3;
shl.b64 %rd45, %rd44, 2;
add.s64 %rd46, %rd42, %rd45;
ld.global.f32 %f142, [%rd46];

$L__BB118_10:
setp.eq.s64 %p13, %rd17, 0;
not.pred %p14, %p1;
or.pred %p15, %p13, %p14;
@%p15 bra $L__BB118_12;

cvta.to.global.u64 %rd47, %rd17;
cvt.u64.u32 %rd48, %r377;
selp.b64 %rd49, 0, %rd48, %p3;
shl.b64 %rd50, %rd49, 2;
add.s64 %rd51, %rd47, %rd50;
ld.global.f32 %f143, [%rd51];

$L__BB118_12:
shl.b32 %r67, %r1, 5;
add.s32 %r68, %r67, %r52;
shr.u32 %r69, %r68, 31;
add.s32 %r70, %r68, %r69;
shr.s32 %r380, %r70, 1;
shr.u32 %r71, %r29, 6;
setp.ge.s32 %p17, %r380, %r71;
mov.f32 %f148, %f146;
@%p17 bra $L__BB118_17;

mov.u32 %r72, %ctaid.z;
and.b32 %r78, %r70, -2;
sub.s32 %r9, %r68, %r78;
mul.hi.u32 %r79, %r72, %r40;
add.s32 %r80, %r79, %r72;
shr.u32 %r81, %r80, %r41;
mad.lo.s32 %r82, %r81, %r43, %r380;
mad.lo.s32 %r83, %r377, %r37, %r82;
mov.u32 %r84, %ctaid.x;
mad.lo.s32 %r85, %r84, %r33, %r83;
mul.wide.s32 %rd89, %r85, 36;
shl.b32 %r379, %r380, 1;
mul.lo.s32 %r86, %r378, %r38;
cvt.u64.u32 %rd52, %r86;
mul.lo.s32 %r87, %r72, %r44;
cvt.u64.u32 %rd53, %r87;
add.s64 %rd7, %rd52, %rd53;
mov.f32 %f148, 0f00000000;
cvta.to.global.u64 %rd54, %rd11;
cvta.to.global.u64 %rd58, %rd10;
cvt.s64.s32 %rd61, %r9;
mov.f32 %f146, %f148;

$L__BB118_14:
cvt.u64.u32 %rd55, %r379;
add.s64 %rd56, %rd7, %rd55;
shl.b32 %r124, %r9, 2;
mul.wide.s32 %rd57, %r124, 4;
add.s64 %rd59, %rd58, %rd57;
add.s64 %rd60, %rd59, %rd89;
ld.global.nc.u32 %r125, [%rd60+4];
and.b32 %r126, %r125, -2004318072;
shr.u32 %r127, %r126, 1;
or.b32 %r128, %r127, 839922192;
and.b32 %r129, %r125, 30583;
mov.u32 %r130, 201852420;
mov.u32 %r131, 50462976;
prmt.b32 %r132, %r131, %r130, %r129;
mov.u32 %r133, -185009412;
mov.u32 %r134, -33620224;
prmt.b32 %r135, %r134, %r133, %r129;
and.b32 %r136, %r128, 30292;
prmt.b32 %r137, %r132, %r135, %r136;
shr.u32 %r138, %r125, 16;
and.b32 %r139, %r138, 30583;
prmt.b32 %r140, %r131, %r130, %r139;
prmt.b32 %r141, %r134, %r133, %r139;
shr.u32 %r142, %r128, 16;
prmt.b32 %r143, %r140, %r141, %r142;
mov.u32 %r144, 25632;
prmt.b32 %r89, %r137, %r143, %r144;
mov.u32 %r145, 30001;
prmt.b32 %r93, %r137, %r143, %r145;
ld.global.nc.u32 %r146, [%rd60+8];
and.b32 %r147, %r146, -2004318072;
shr.u32 %r148, %r147, 1;
or.b32 %r149, %r148, 839922192;
and.b32 %r150, %r146, 30583;
prmt.b32 %r151, %r131, %r130, %r150;
prmt.b32 %r152, %r134, %r133, %r150;
and.b32 %r153, %r149, 30292;
prmt.b32 %r154, %r151, %r152, %r153;
shr.u32 %r155, %r146, 16;
and.b32 %r156, %r155, 30583;
prmt.b32 %r157, %r131, %r130, %r156;
prmt.b32 %r158, %r134, %r133, %r156;
shr.u32 %r159, %r149, 16;
prmt.b32 %r160, %r157, %r158, %r159;
prmt.b32 %r97, %r154, %r160, %r144;
prmt.b32 %r101, %r154, %r160, %r145;
add.s64 %rd62, %rd56, %rd61;
mul.lo.s64 %rd63, %rd62, 36;
add.s64 %rd64, %rd54, %rd63;
ld.global.nc.u32 %r13, [%rd64+4];
mov.u32 %r109, 0;

dp4a.s32.s32 %r88, %r89, %r13, %r109;

ld.global.nc.u32 %r14, [%rd64+12];

dp4a.s32.s32 %r92, %r93, %r14, %r88;

ld.global.nc.u32 %r15, [%rd64+8];

dp4a.s32.s32 %r96, %r97, %r15, %r92;

ld.global.nc.u32 %r16, [%rd64+16];

dp4a.s32.s32 %r100, %r101, %r16, %r96;

shl.b32 %r161, %r9, 1;
cvt.s64.s32 %rd65, %r161;
add.s64 %rd66, %rd58, %rd65;
add.s64 %rd67, %rd66, %rd89;
ld.global.nc.u8 %rs5, [%rd67];
and.b16 %rs6, %rs5, 127;
setp.eq.s16 %p18, %rs6, 127;
selp.b16 %rs7, 0, %rs5, %p18;
and.b16 %rs1, %rs7, 255;

{cvt.rn.f16x2.e4m3x2 %r104, %rs1;}


cvt.u16.u32 %rs2, %r104;

{cvt.f32.f16 %f50, %rs2;}


mov.f32 %f54, 0f40000000;
div.approx.ftz.f32 %f55, %f50, %f54;
ld.global.nc.u32 %r17, [%rd64];

{.reg .f16 low,high;
mov.b32 {low,high},%r17;
cvt.f32.f16 %f51, low;}


mul.ftz.f32 %f56, %f55, %f51;
cvt.rn.f32.s32 %f57, %r100;
fma.rn.ftz.f32 %f58, %f56, %f57, 0f00000000;
ld.global.nc.u32 %r162, [%rd60+12];
and.b32 %r163, %r162, -2004318072;
shr.u32 %r164, %r163, 1;
or.b32 %r165, %r164, 839922192;
and.b32 %r166, %r162, 30583;
prmt.b32 %r167, %r131, %r130, %r166;
prmt.b32 %r168, %r134, %r133, %r166;
and.b32 %r169, %r165, 30292;
prmt.b32 %r170, %r167, %r168, %r169;
shr.u32 %r171, %r162, 16;
and.b32 %r172, %r171, 30583;
prmt.b32 %r173, %r131, %r130, %r172;
prmt.b32 %r174, %r134, %r133, %r172;
shr.u32 %r175, %r165, 16;
prmt.b32 %r176, %r173, %r174, %r175;
prmt.b32 %r107, %r170, %r176, %r144;
prmt.b32 %r111, %r170, %r176, %r145;
ld.global.nc.u32 %r177, [%rd60+16];
and.b32 %r178, %r177, -2004318072;
shr.u32 %r179, %r178, 1;
or.b32 %r180, %r179, 839922192;
and.b32 %r181, %r177, 30583;
prmt.b32 %r182, %r131, %r130, %r181;
prmt.b32 %r183, %r134, %r133, %r181;
and.b32 %r184, %r180, 30292;
prmt.b32 %r185, %r182, %r183, %r184;
shr.u32 %r186, %r177, 16;
and.b32 %r187, %r186, 30583;
prmt.b32 %r188, %r131, %r130, %r187;
prmt.b32 %r189, %r134, %r133, %r187;
shr.u32 %r190, %r180, 16;
prmt.b32 %r191, %r188, %r189, %r190;
prmt.b32 %r115, %r185, %r191, %r144;
prmt.b32 %r119, %r185, %r191, %r145;
or.b32 %r192, %r124, 2;
shr.s32 %r193, %r192, 2;
cvt.s64.s32 %rd68, %r193;
add.s64 %rd69, %rd56, %rd68;
mul.lo.s64 %rd70, %rd69, 36;
add.s64 %rd71, %rd54, %rd70;
ld.global.nc.u32 %r18, [%rd71+20];

dp4a.s32.s32 %r106, %r107, %r18, %r109;

ld.global.nc.u32 %r19, [%rd71+28];

dp4a.s32.s32 %r110, %r111, %r19, %r106;

ld.global.nc.u32 %r20, [%rd71+24];

dp4a.s32.s32 %r114, %r115, %r20, %r110;

ld.global.nc.u32 %r21, [%rd71+32];

dp4a.s32.s32 %r118, %r119, %r21, %r114;

shr.s32 %r194, %r192, 1;
cvt.s64.s32 %rd72, %r194;
add.s64 %rd73, %rd58, %rd72;
add.s64 %rd74, %rd73, %rd89;
ld.global.nc.u8 %rs8, [%rd74];
and.b16 %rs9, %rs8, 127;
setp.eq.s16 %p19, %rs9, 127;
selp.b16 %rs10, 0, %rs8, %p19;
and.b16 %rs3, %rs10, 255;

{cvt.rn.f16x2.e4m3x2 %r122, %rs3;}


cvt.u16.u32 %rs4, %r122;

{cvt.f32.f16 %f52, %rs4;}


div.approx.ftz.f32 %f59, %f52, %f54;
ld.global.nc.u32 %r22, [%rd71];

{.reg .f16 low,high;
mov.b32 {low,high},%r22;
cvt.f32.f16 %f53, low;}


mul.ftz.f32 %f60, %f59, %f53;
cvt.rn.f32.s32 %f61, %r118;
fma.rn.ftz.f32 %f62, %f60, %f61, %f58;
add.ftz.f32 %f148, %f148, %f62;
setp.eq.s64 %p20, %rd14, 0;
@%p20 bra $L__BB118_16;

shr.s32 %r232, %r124, 1;
cvta.to.global.u64 %rd76, %rd14;
add.s64 %rd77, %rd76, %rd57;
add.s64 %rd78, %rd77, %rd89;
ld.global.u32 %r233, [%rd78+4];
and.b32 %r234, %r233, -2004318072;
shr.u32 %r235, %r234, 1;
or.b32 %r236, %r235, 839922192;
and.b32 %r237, %r233, 30583;
prmt.b32 %r240, %r131, %r130, %r237;
prmt.b32 %r243, %r134, %r133, %r237;
and.b32 %r244, %r236, 30292;
prmt.b32 %r245, %r240, %r243, %r244;
shr.u32 %r246, %r233, 16;
and.b32 %r247, %r246, 30583;
prmt.b32 %r248, %r131, %r130, %r247;
prmt.b32 %r249, %r134, %r133, %r247;
shr.u32 %r250, %r236, 16;
prmt.b32 %r251, %r248, %r249, %r250;
prmt.b32 %r196, %r245, %r251, %r144;
prmt.b32 %r200, %r245, %r251, %r145;
ld.global.u32 %r254, [%rd78+8];
and.b32 %r255, %r254, -2004318072;
shr.u32 %r256, %r255, 1;
or.b32 %r257, %r256, 839922192;
and.b32 %r258, %r254, 30583;
prmt.b32 %r259, %r131, %r130, %r258;
prmt.b32 %r260, %r134, %r133, %r258;
and.b32 %r261, %r257, 30292;
prmt.b32 %r262, %r259, %r260, %r261;
shr.u32 %r263, %r254, 16;
and.b32 %r264, %r263, 30583;
prmt.b32 %r265, %r131, %r130, %r264;
prmt.b32 %r266, %r134, %r133, %r264;
shr.u32 %r267, %r257, 16;
prmt.b32 %r268, %r265, %r266, %r267;
prmt.b32 %r204, %r262, %r268, %r144;
prmt.b32 %r208, %r262, %r268, %r145;

dp4a.s32.s32 %r195, %r196, %r13, %r109;


dp4a.s32.s32 %r199, %r200, %r14, %r195;


dp4a.s32.s32 %r203, %r204, %r15, %r199;


dp4a.s32.s32 %r207, %r208, %r16, %r203;

add.s64 %rd80, %rd76, %rd65;
add.s64 %rd81, %rd80, %rd89;
ld.global.u8 %rs15, [%rd81];
and.b16 %rs16, %rs15, 127;
setp.eq.s16 %p21, %rs16, 127;
selp.b16 %rs11, 0, %rs15, %p21;

{cvt.rn.f16x2.e4m3x2 %r211, %rs11;}


cvt.u16.u32 %rs12, %r211;

{cvt.f32.f16 %f63, %rs12;}


div.approx.ftz.f32 %f68, %f63, %f54;

{.reg .f16 low,high;
mov.b32 {low,high},%r17;
cvt.f32.f16 %f64, low;}


mul.ftz.f32 %f69, %f68, %f64;
cvt.rn.f32.s32 %f70, %r207;
fma.rn.ftz.f32 %f71, %f69, %f70, 0f00000000;
ld.global.u32 %r270, [%rd78+12];
and.b32 %r271, %r270, -2004318072;
shr.u32 %r272, %r271, 1;
or.b32 %r273, %r272, 839922192;
and.b32 %r274, %r270, 30583;
prmt.b32 %r275, %r131, %r130, %r274;
prmt.b32 %r276, %r134, %r133, %r274;
and.b32 %r277, %r273, 30292;
prmt.b32 %r278, %r275, %r276, %r277;
shr.u32 %r279, %r270, 16;
and.b32 %r280, %r279, 30583;
prmt.b32 %r281, %r131, %r130, %r280;
prmt.b32 %r282, %r134, %r133, %r280;
shr.u32 %r283, %r273, 16;
prmt.b32 %r284, %r281, %r282, %r283;
prmt.b32 %r214, %r278, %r284, %r144;
prmt.b32 %r218, %r278, %r284, %r145;
ld.global.u32 %r285, [%rd78+16];
and.b32 %r286, %r285, -2004318072;
shr.u32 %r287, %r286, 1;
or.b32 %r288, %r287, 839922192;
and.b32 %r289, %r285, 30583;
prmt.b32 %r290, %r131, %r130, %r289;
prmt.b32 %r291, %r134, %r133, %r289;
and.b32 %r292, %r288, 30292;
prmt.b32 %r293, %r290, %r291, %r292;
shr.u32 %r294, %r285, 16;
and.b32 %r295, %r294, 30583;
prmt.b32 %r296, %r131, %r130, %r295;
prmt.b32 %r297, %r134, %r133, %r295;
shr.u32 %r298, %r288, 16;
prmt.b32 %r299, %r296, %r297, %r298;
prmt.b32 %r222, %r293, %r299, %r144;
prmt.b32 %r226, %r293, %r299, %r145;

dp4a.s32.s32 %r213, %r214, %r18, %r109;


dp4a.s32.s32 %r217, %r218, %r19, %r213;


dp4a.s32.s32 %r221, %r222, %r20, %r217;


dp4a.s32.s32 %r225, %r226, %r21, %r221;

or.b32 %r300, %r232, 1;
cvt.s64.s32 %rd82, %r300;
add.s64 %rd83, %rd76, %rd82;
add.s64 %rd84, %rd83, %rd89;
ld.global.u8 %rs17, [%rd84];
and.b16 %rs18, %rs17, 127;
setp.eq.s16 %p22, %rs18, 127;
selp.b16 %rs13, 0, %rs17, %p22;

{cvt.rn.f16x2.e4m3x2 %r229, %rs13;}


cvt.u16.u32 %rs14, %r229;

{cvt.f32.f16 %f65, %rs14;}


div.approx.ftz.f32 %f72, %f65, %f54;

{.reg .f16 low,high;
mov.b32 {low,high},%r22;
cvt.f32.f16 %f66, low;}


mul.ftz.f32 %f73, %f72, %f66;
cvt.rn.f32.s32 %f74, %r225;
fma.rn.ftz.f32 %f75, %f73, %f74, %f71;
add.ftz.f32 %f146, %f146, %f75;

$L__BB118_16:
add.s64 %rd89, %rd89, 2304;
add.s32 %r379, %r379, 128;
add.s32 %r380, %r380, 64;
setp.lt.s32 %p23, %r380, %r71;
@%p23 bra $L__BB118_14;

$L__BB118_17:
mov.u32 %r364, %tid.y;
setp.eq.s32 %p24, %r364, 0;
@%p24 bra $L__BB118_20;

mov.u32 %r366, %tid.x;
mov.u32 %r365, %tid.y;
setp.eq.s64 %p25, %rd14, 0;
add.s32 %r25, %r365, -1;
shl.b32 %r305, %r365, 7;
mov.u32 %r306, _ZZ13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r307, %r306, %r305;
shl.b32 %r308, %r366, 2;
add.s32 %r309, %r307, %r308;
st.shared.f32 [%r309+-128], %f148;
@%p25 bra $L__BB118_20;

shl.b32 %r311, %r25, 7;
mov.u32 %r312, _ZZ13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r313, %r312, %r311;
add.s32 %r315, %r313, %r308;
st.shared.f32 [%r315], %f146;

$L__BB118_20:
mov.u32 %r367, %tid.y;
setp.ne.s32 %p26, %r367, 0;
bar.sync 0;
@%p26 bra $L__BB118_39;

mov.u32 %r368, %tid.x;
setp.ne.s64 %p27, %rd14, 0;
shl.b32 %r318, %r368, 2;
mov.u32 %r319, _ZZ13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r26, %r319, %r318;
ld.shared.f32 %f76, [%r26];
add.ftz.f32 %f19, %f148, %f76;
@%p27 bra $L__BB118_23;
bra.uni $L__BB118_22;

$L__BB118_23:
mov.u32 %r322, _ZZ13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r323, %r322, %r318;
ld.shared.f32 %f80, [%r323];
add.ftz.f32 %f81, %f146, %f80;
ld.shared.f32 %f82, [%r26+128];
add.ftz.f32 %f83, %f19, %f82;
ld.shared.f32 %f84, [%r323+128];
add.ftz.f32 %f85, %f81, %f84;
ld.shared.f32 %f86, [%r26+256];
add.ftz.f32 %f150, %f83, %f86;
ld.shared.f32 %f87, [%r323+256];
add.ftz.f32 %f146, %f85, %f87;
bra.uni $L__BB118_24;

$L__BB118_22:
ld.shared.f32 %f77, [%r26+128];
add.ftz.f32 %f78, %f19, %f77;
ld.shared.f32 %f79, [%r26+256];
add.ftz.f32 %f150, %f78, %f79;

$L__BB118_24:
mov.b32 %r324, %f150;
mov.u32 %r325, 31;
mov.u32 %r326, 16;
mov.u32 %r327, -1;
shfl.sync.bfly.b32 %r328|%p28, %r324, %r326, %r325, %r327;
mov.b32 %f88, %r328;
add.ftz.f32 %f89, %f150, %f88;
mov.b32 %r329, %f89;
mov.u32 %r330, 8;
shfl.sync.bfly.b32 %r331|%p29, %r329, %r330, %r325, %r327;
mov.b32 %f90, %r331;
add.ftz.f32 %f91, %f89, %f90;
mov.b32 %r332, %f91;
mov.u32 %r333, 4;
shfl.sync.bfly.b32 %r334|%p30, %r332, %r333, %r325, %r327;
mov.b32 %f92, %r334;
add.ftz.f32 %f93, %f91, %f92;
mov.b32 %r335, %f93;
mov.u32 %r336, 2;
shfl.sync.bfly.b32 %r337|%p31, %r335, %r336, %r325, %r327;
mov.b32 %f94, %r337;
add.ftz.f32 %f25, %f93, %f94;
mov.b32 %r338, %f25;
mov.u32 %r339, 1;
shfl.sync.bfly.b32 %r27|%p2, %r338, %r339, %r325, %r327;
setp.eq.s64 %p32, %rd14, 0;
@%p32 bra $L__BB118_26;

mov.b32 %r340, %f146;
shfl.sync.bfly.b32 %r344|%p33, %r340, %r326, %r325, %r327;
mov.b32 %f95, %r344;
add.ftz.f32 %f96, %f146, %f95;
mov.b32 %r345, %f96;
shfl.sync.bfly.b32 %r347|%p34, %r345, %r330, %r325, %r327;
mov.b32 %f97, %r347;
add.ftz.f32 %f98, %f96, %f97;
mov.b32 %r348, %f98;
shfl.sync.bfly.b32 %r350|%p35, %r348, %r333, %r325, %r327;
mov.b32 %f99, %r350;
add.ftz.f32 %f100, %f98, %f99;
mov.b32 %r351, %f100;
shfl.sync.bfly.b32 %r353|%p36, %r351, %r336, %r325, %r327;
mov.b32 %f101, %r353;
add.ftz.f32 %f102, %f100, %f101;
mov.b32 %r354, %f102;
shfl.sync.bfly.b32 %r356|%p37, %r354, %r339, %r325, %r327;
mov.b32 %f103, %r356;
add.ftz.f32 %f146, %f102, %f103;

$L__BB118_26:
mov.u32 %r369, %tid.x;
setp.ne.s32 %p38, %r369, 0;
mov.b32 %f104, %r27;
add.ftz.f32 %f28, %f25, %f104;
@%p38 bra $L__BB118_39;

fma.rn.ftz.f32 %f152, %f142, %f28, %f141;
@%p32 bra $L__BB118_38;

ld.param.u32 %r372, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
fma.rn.ftz.f32 %f30, %f143, %f146, %f140;
setp.gt.s32 %p40, %r372, 2;
@%p40 bra $L__BB118_32;

ld.param.u32 %r375, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p43, %r375, 1;
@%p43 bra $L__BB118_36;

ld.param.u32 %r376, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p44, %r376, 2;
@%p44 bra $L__BB118_31;
bra.uni $L__BB118_37;

$L__BB118_31:
mul.ftz.f32 %f124, %f30, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f125, %f124;
add.ftz.f32 %f126, %f125, 0f3F800000;
div.approx.ftz.f32 %f127, %f30, %f126;
mul.ftz.f32 %f152, %f152, %f127;
bra.uni $L__BB118_38;

$L__BB118_32:
ld.param.u32 %r373, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p41, %r373, 3;
@%p41 bra $L__BB118_35;

ld.param.u32 %r374, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.ne.s32 %p42, %r374, 6;
@%p42 bra $L__BB118_37;

ld.param.f32 %f136, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
min.ftz.f32 %f105, %f30, %f136;
neg.ftz.f32 %f106, %f136;
min.ftz.f32 %f107, %f152, %f136;
max.ftz.f32 %f108, %f107, %f106;
mul.ftz.f32 %f109, %f105, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f110, %f109;
add.ftz.f32 %f111, %f110, 0f3F800000;
div.approx.ftz.f32 %f112, %f105, %f111;
mul.ftz.f32 %f152, %f108, %f112;
bra.uni $L__BB118_38;

$L__BB118_36:
mul.ftz.f32 %f128, %f30, 0f3F000000;
mul.ftz.f32 %f129, %f30, 0f3D372713;
fma.rn.ftz.f32 %f130, %f30, %f129, 0f3F800000;
mul.ftz.f32 %f131, %f30, 0f3F4C422A;
mul.ftz.f32 %f132, %f131, %f130;
tanh.approx.f32 %f133, %f132;
add.ftz.f32 %f134, %f133, 0f3F800000;
mul.ftz.f32 %f135, %f128, %f134;
mul.ftz.f32 %f152, %f152, %f135;
bra.uni $L__BB118_38;

$L__BB118_35:
mov.f32 %f113, 0f40E00000;
min.ftz.f32 %f114, %f30, %f113;
min.ftz.f32 %f115, %f152, %f113;
mov.f32 %f116, 0fC0E00000;
max.ftz.f32 %f117, %f115, %f116;
mul.ftz.f32 %f118, %f114, 0f3FD9DB23;
mul.ftz.f32 %f119, %f118, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f120, %f119;
add.ftz.f32 %f121, %f120, 0f3F800000;
div.approx.ftz.f32 %f122, %f114, %f121;
add.ftz.f32 %f123, %f117, 0f3F800000;
mul.ftz.f32 %f152, %f123, %f122;
bra.uni $L__BB118_38;

$L__BB118_37:
mul.ftz.f32 %f152, %f152, %f30;

$L__BB118_38:
ld.param.u64 %rd88, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r371, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r370, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
cvta.to.global.u64 %rd85, %rd88;
mov.u32 %r359, %ctaid.z;
mov.u32 %r360, %ctaid.y;
mov.u32 %r361, %ctaid.x;
mad.lo.s32 %r362, %r360, %r371, %r361;
mad.lo.s32 %r363, %r359, %r370, %r362;
mul.wide.u32 %rd86, %r363, 4;
add.s64 %rd87, %rd85, %rd86;
st.global.f32 [%rd87], %f152;

$L__BB118_39:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<14>;
.reg .b16 %rs<11>;
.reg .f32 %f<37>;
.reg .b32 %r<199>;
.reg .b64 %rd<46>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[384];

ld.param.u64 %rd14, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r18, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r19, [_Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
mov.u32 %r1, %tid.y;
shl.b32 %r35, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r35, %r2;
mov.u32 %r196, %ctaid.y;
setp.eq.s64 %p2, %rd16, 0;
@%p2 bra $L__BB119_2;

cvta.to.global.u64 %rd18, %rd16;
mul.wide.u32 %rd19, %r196, 4;
add.s64 %rd20, %rd18, %rd19;
ld.global.nc.u32 %r195, [%rd20];
mul.hi.u32 %r36, %r196, %r19;
add.s32 %r37, %r36, %r196;
shr.u32 %r38, %r37, %r20;
mul.lo.s32 %r39, %r38, %r21;
sub.s32 %r196, %r196, %r39;
bra.uni $L__BB119_3;

$L__BB119_2:
mul.hi.u32 %r40, %r196, %r23;
add.s32 %r41, %r40, %r196;
shr.u32 %r195, %r41, %r24;

$L__BB119_3:
shr.u32 %r42, %r3, 31;
add.s32 %r43, %r3, %r42;
shr.s32 %r198, %r43, 1;
shr.u32 %r44, %r18, 6;
setp.ge.s32 %p3, %r198, %r44;
mov.f32 %f36, 0f00000000;
@%p3 bra $L__BB119_6;

mov.u32 %r45, %ctaid.z;
and.b32 %r48, %r43, -2;
sub.s32 %r49, %r3, %r48;
shl.b32 %r50, %r49, 2;
shl.b32 %r51, %r49, 1;
cvt.s64.s32 %rd1, %r49;
cvt.s64.s32 %rd21, %r51;
or.b32 %r52, %r50, 2;
shr.s32 %r53, %r52, 1;
shr.s32 %r54, %r52, 2;
cvt.s64.s32 %rd2, %r54;
cvt.s64.s32 %rd22, %r53;
mul.hi.u32 %r55, %r45, %r29;
add.s32 %r56, %r55, %r45;
shr.u32 %r57, %r56, %r30;
mad.lo.s32 %r58, %r57, %r32, %r198;
mad.lo.s32 %r59, %r195, %r26, %r58;
mov.u32 %r60, %ctaid.x;
mad.lo.s32 %r61, %r60, %r22, %r59;
mul.wide.s32 %rd23, %r61, 36;
add.s64 %rd24, %rd23, %rd22;
cvta.to.global.u64 %rd25, %rd14;
add.s64 %rd45, %rd25, %rd24;
add.s64 %rd26, %rd23, %rd21;
add.s64 %rd44, %rd25, %rd26;
shl.b32 %r197, %r198, 1;
mul.wide.s32 %rd27, %r50, 4;
add.s64 %rd28, %rd23, %rd27;
add.s64 %rd29, %rd25, %rd28;
add.s64 %rd43, %rd29, 8;
cvta.to.global.u64 %rd6, %rd15;
mul.lo.s32 %r62, %r196, %r27;
cvt.u64.u32 %rd30, %r62;
mul.lo.s32 %r63, %r45, %r33;
cvt.u64.u32 %rd31, %r63;
add.s64 %rd7, %rd30, %rd31;
mov.f32 %f36, 0f00000000;

$L__BB119_5:
cvt.u64.u32 %rd32, %r197;
add.s64 %rd33, %rd7, %rd32;
ld.global.nc.u32 %r100, [%rd43+-4];
and.b32 %r101, %r100, -2004318072;
shr.u32 %r102, %r101, 1;
or.b32 %r103, %r102, 839922192;
and.b32 %r104, %r100, 30583;
mov.u32 %r105, 201852420;
mov.u32 %r106, 50462976;
prmt.b32 %r107, %r106, %r105, %r104;
mov.u32 %r108, -185009412;
mov.u32 %r109, -33620224;
prmt.b32 %r110, %r109, %r108, %r104;
and.b32 %r111, %r103, 30292;
prmt.b32 %r112, %r107, %r110, %r111;
shr.u32 %r113, %r100, 16;
and.b32 %r114, %r113, 30583;
prmt.b32 %r115, %r106, %r105, %r114;
prmt.b32 %r116, %r109, %r108, %r114;
shr.u32 %r117, %r103, 16;
prmt.b32 %r118, %r115, %r116, %r117;
mov.u32 %r119, 25632;
prmt.b32 %r65, %r112, %r118, %r119;
mov.u32 %r120, 30001;
prmt.b32 %r69, %r112, %r118, %r120;
ld.global.nc.u32 %r121, [%rd43];
and.b32 %r122, %r121, -2004318072;
shr.u32 %r123, %r122, 1;
or.b32 %r124, %r123, 839922192;
and.b32 %r125, %r121, 30583;
prmt.b32 %r126, %r106, %r105, %r125;
prmt.b32 %r127, %r109, %r108, %r125;
and.b32 %r128, %r124, 30292;
prmt.b32 %r129, %r126, %r127, %r128;
shr.u32 %r130, %r121, 16;
and.b32 %r131, %r130, 30583;
prmt.b32 %r132, %r106, %r105, %r131;
prmt.b32 %r133, %r109, %r108, %r131;
shr.u32 %r134, %r124, 16;
prmt.b32 %r135, %r132, %r133, %r134;
prmt.b32 %r73, %r129, %r135, %r119;
prmt.b32 %r77, %r129, %r135, %r120;
add.s64 %rd34, %rd33, %rd1;
mul.lo.s64 %rd35, %rd34, 36;
add.s64 %rd36, %rd6, %rd35;
ld.global.nc.u32 %r66, [%rd36+4];
mov.u32 %r85, 0;

dp4a.s32.s32 %r64, %r65, %r66, %r85;

ld.global.nc.u32 %r70, [%rd36+12];

dp4a.s32.s32 %r68, %r69, %r70, %r64;

ld.global.nc.u32 %r74, [%rd36+8];

dp4a.s32.s32 %r72, %r73, %r74, %r68;

ld.global.nc.u32 %r78, [%rd36+16];

dp4a.s32.s32 %r76, %r77, %r78, %r72;

ld.global.nc.u8 %rs5, [%rd44];
and.b16 %rs6, %rs5, 127;
setp.eq.s16 %p4, %rs6, 127;
selp.b16 %rs7, 0, %rs5, %p4;
and.b16 %rs1, %rs7, 255;

{cvt.rn.f16x2.e4m3x2 %r80, %rs1;}


cvt.u16.u32 %rs2, %r80;

{cvt.f32.f16 %f7, %rs2;}


mov.f32 %f11, 0f40000000;
div.approx.ftz.f32 %f12, %f7, %f11;
ld.global.nc.u32 %r81, [%rd36];

{.reg .f16 low,high;
mov.b32 {low,high},%r81;
cvt.f32.f16 %f8, low;}


mul.ftz.f32 %f13, %f12, %f8;
cvt.rn.f32.s32 %f14, %r76;
fma.rn.ftz.f32 %f15, %f13, %f14, 0f00000000;
ld.global.nc.u32 %r136, [%rd43+4];
and.b32 %r137, %r136, -2004318072;
shr.u32 %r138, %r137, 1;
or.b32 %r139, %r138, 839922192;
and.b32 %r140, %r136, 30583;
prmt.b32 %r141, %r106, %r105, %r140;
prmt.b32 %r142, %r109, %r108, %r140;
and.b32 %r143, %r139, 30292;
prmt.b32 %r144, %r141, %r142, %r143;
shr.u32 %r145, %r136, 16;
and.b32 %r146, %r145, 30583;
prmt.b32 %r147, %r106, %r105, %r146;
prmt.b32 %r148, %r109, %r108, %r146;
shr.u32 %r149, %r139, 16;
prmt.b32 %r150, %r147, %r148, %r149;
prmt.b32 %r83, %r144, %r150, %r119;
prmt.b32 %r87, %r144, %r150, %r120;
ld.global.nc.u32 %r151, [%rd43+8];
and.b32 %r152, %r151, -2004318072;
shr.u32 %r153, %r152, 1;
or.b32 %r154, %r153, 839922192;
and.b32 %r155, %r151, 30583;
prmt.b32 %r156, %r106, %r105, %r155;
prmt.b32 %r157, %r109, %r108, %r155;
and.b32 %r158, %r154, 30292;
prmt.b32 %r159, %r156, %r157, %r158;
shr.u32 %r160, %r151, 16;
and.b32 %r161, %r160, 30583;
prmt.b32 %r162, %r106, %r105, %r161;
prmt.b32 %r163, %r109, %r108, %r161;
shr.u32 %r164, %r154, 16;
prmt.b32 %r165, %r162, %r163, %r164;
prmt.b32 %r91, %r159, %r165, %r119;
prmt.b32 %r95, %r159, %r165, %r120;
add.s64 %rd37, %rd33, %rd2;
mul.lo.s64 %rd38, %rd37, 36;
add.s64 %rd39, %rd6, %rd38;
ld.global.nc.u32 %r84, [%rd39+20];

dp4a.s32.s32 %r82, %r83, %r84, %r85;

ld.global.nc.u32 %r88, [%rd39+28];

dp4a.s32.s32 %r86, %r87, %r88, %r82;

ld.global.nc.u32 %r92, [%rd39+24];

dp4a.s32.s32 %r90, %r91, %r92, %r86;

ld.global.nc.u32 %r96, [%rd39+32];

dp4a.s32.s32 %r94, %r95, %r96, %r90;

ld.global.nc.u8 %rs8, [%rd45];
and.b16 %rs9, %rs8, 127;
setp.eq.s16 %p5, %rs9, 127;
selp.b16 %rs10, 0, %rs8, %p5;
and.b16 %rs3, %rs10, 255;

{cvt.rn.f16x2.e4m3x2 %r98, %rs3;}


cvt.u16.u32 %rs4, %r98;

{cvt.f32.f16 %f9, %rs4;}


div.approx.ftz.f32 %f16, %f9, %f11;
ld.global.nc.u32 %r99, [%rd39];

{.reg .f16 low,high;
mov.b32 {low,high},%r99;
cvt.f32.f16 %f10, low;}


mul.ftz.f32 %f17, %f16, %f10;
cvt.rn.f32.s32 %f18, %r94;
fma.rn.ftz.f32 %f19, %f17, %f18, %f15;
add.ftz.f32 %f36, %f36, %f19;
add.s64 %rd45, %rd45, 2304;
add.s64 %rd44, %rd44, 2304;
add.s32 %r197, %r197, 128;
add.s64 %rd43, %rd43, 2304;
add.s32 %r198, %r198, 64;
setp.lt.s32 %p6, %r198, %r44;
@%p6 bra $L__BB119_5;

$L__BB119_6:
setp.eq.s32 %p7, %r1, 0;
@%p7 bra $L__BB119_8;

shl.b32 %r166, %r1, 7;
mov.u32 %r167, _ZZ13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r168, %r167, %r166;
shl.b32 %r169, %r2, 2;
add.s32 %r170, %r168, %r169;
st.shared.f32 [%r170+-128], %f36;

$L__BB119_8:
setp.ne.s32 %p8, %r1, 0;
bar.sync 0;
@%p8 bra $L__BB119_11;

shl.b32 %r171, %r2, 2;
mov.u32 %r172, _ZZ13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r173, %r172, %r171;
mov.u32 %r174, 2;
ld.shared.f32 %f20, [%r173];
add.ftz.f32 %f21, %f36, %f20;
ld.shared.f32 %f22, [%r173+128];
add.ftz.f32 %f23, %f21, %f22;
ld.shared.f32 %f24, [%r173+256];
add.ftz.f32 %f25, %f23, %f24;
mov.b32 %r175, %f25;
mov.u32 %r176, 31;
mov.u32 %r177, 16;
mov.u32 %r178, -1;
shfl.sync.bfly.b32 %r179|%p9, %r175, %r177, %r176, %r178;
mov.b32 %f26, %r179;
add.ftz.f32 %f27, %f25, %f26;
mov.b32 %r180, %f27;
mov.u32 %r181, 8;
shfl.sync.bfly.b32 %r182|%p10, %r180, %r181, %r176, %r178;
mov.b32 %f28, %r182;
add.ftz.f32 %f29, %f27, %f28;
mov.b32 %r183, %f29;
mov.u32 %r184, 4;
shfl.sync.bfly.b32 %r185|%p11, %r183, %r184, %r176, %r178;
mov.b32 %f30, %r185;
add.ftz.f32 %f31, %f29, %f30;
mov.b32 %r186, %f31;
shfl.sync.bfly.b32 %r187|%p12, %r186, %r174, %r176, %r178;
mov.b32 %f32, %r187;
add.ftz.f32 %f4, %f31, %f32;
mov.b32 %r188, %f4;
mov.u32 %r189, 1;
shfl.sync.bfly.b32 %r17|%p1, %r188, %r189, %r176, %r178;
setp.ne.s32 %p13, %r2, 0;
@%p13 bra $L__BB119_11;

mov.u32 %r190, %ctaid.x;
mov.u32 %r191, %ctaid.y;
mad.lo.s32 %r192, %r191, %r28, %r190;
mov.u32 %r193, %ctaid.z;
mad.lo.s32 %r194, %r193, %r34, %r192;
cvta.to.global.u64 %rd40, %rd17;
mul.wide.u32 %rd41, %r194, 4;
add.s64 %rd42, %rd40, %rd41;
mov.b32 %f33, %r17;
add.ftz.f32 %f34, %f4, %f33;
st.global.f32 [%rd42], %f34;

$L__BB119_11:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type40ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type40ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type40ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type40ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type40ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type40ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type40ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type40ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type40ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<39>;
.reg .b16 %rs<29>;
.reg .f32 %f<142>;
.reg .b32 %r<419>;
.reg .b64 %rd<61>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type40ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type40ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type40ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type40ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type40ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r17, [_Z13mul_mat_vec_qIL9ggml_type40ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r18, [_Z13mul_mat_vec_qIL9ggml_type40ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r19, [_Z13mul_mat_vec_qIL9ggml_type40ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type40ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type40ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type40ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type40ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type40ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type40ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type40ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type40ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type40ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type40ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
cvta.to.global.u64 %rd1, %rd19;
mov.u32 %r1, %tid.y;
shl.b32 %r33, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r33, %r2;
mov.u32 %r34, %ctaid.x;
shl.b32 %r4, %r34, 1;
shr.u32 %r35, %r3, 31;
add.s32 %r36, %r3, %r35;
shr.s32 %r418, %r36, 1;
shr.u32 %r6, %r32, 6;
setp.ge.s32 %p3, %r418, %r6;
mov.f32 %f138, 0f00000000;
mov.f32 %f139, %f138;
mov.f32 %f140, %f138;
mov.f32 %f141, %f138;
@%p3 bra $L__BB120_3;

mov.u32 %r37, %ctaid.y;
mov.u32 %r38, %ctaid.z;
mul.lo.s32 %r39, %r37, %r24;
cvt.u64.u32 %rd20, %r39;
mul.lo.s32 %r40, %r38, %r30;
cvt.u64.u32 %rd21, %r40;
add.s64 %rd2, %rd21, %rd20;
mul.hi.u32 %r41, %r38, %r26;
add.s32 %r42, %r41, %r38;
shr.u32 %r43, %r42, %r27;
mul.lo.s32 %r44, %r43, %r29;
mul.hi.u32 %r45, %r37, %r20;
add.s32 %r46, %r45, %r37;
shr.u32 %r47, %r46, %r21;
mul.lo.s32 %r48, %r47, %r23;
cvta.to.global.u64 %rd60, %rd17;
and.b32 %r51, %r36, -2;
sub.s32 %r52, %r3, %r51;
shl.b32 %r53, %r52, 2;
shl.b32 %r54, %r52, 1;
cvt.s64.s32 %rd4, %r52;
cvt.s64.s32 %rd22, %r54;
or.b32 %r55, %r53, 2;
shr.s32 %r56, %r55, 1;
shr.s32 %r57, %r55, 2;
cvt.s64.s32 %rd5, %r57;
cvt.s64.s32 %rd23, %r56;
shl.b32 %r416, %r418, 1;
add.s32 %r417, %r18, %r416;
add.s32 %r58, %r418, %r44;
add.s32 %r59, %r58, %r48;
add.s32 %r60, %r4, 1;
mad.lo.s32 %r61, %r17, %r60, %r59;
mul.wide.s32 %rd24, %r61, 36;
add.s64 %rd6, %rd24, %rd23;
add.s64 %rd7, %rd24, %rd22;
mad.lo.s32 %r62, %r4, %r17, %r418;
add.s32 %r63, %r62, %r44;
add.s32 %r64, %r63, %r48;
mul.wide.s32 %rd25, %r64, 36;
mul.wide.s32 %rd26, %r53, 4;
add.s64 %rd27, %rd25, %rd26;
add.s64 %rd8, %rd27, 16;
add.s64 %rd9, %rd25, %rd22;
add.s64 %rd28, %rd24, %rd26;
add.s64 %rd10, %rd28, 8;
add.s64 %rd11, %rd25, %rd23;
cvta.to.global.u64 %rd12, %rd18;
mov.f32 %f138, 0f00000000;

$L__BB120_2:
cvt.u64.u32 %rd29, %r416;
add.s64 %rd30, %rd2, %rd29;
add.s64 %rd31, %rd60, %rd8;
ld.global.nc.u32 %r209, [%rd31+-12];
and.b32 %r210, %r209, -2004318072;
shr.u32 %r211, %r210, 1;
or.b32 %r212, %r211, 839922192;
and.b32 %r213, %r209, 30583;
mov.u32 %r214, 201852420;
mov.u32 %r215, 50462976;
prmt.b32 %r216, %r215, %r214, %r213;
mov.u32 %r217, -185009412;
mov.u32 %r218, -33620224;
prmt.b32 %r219, %r218, %r217, %r213;
and.b32 %r220, %r212, 30292;
prmt.b32 %r221, %r216, %r219, %r220;
shr.u32 %r222, %r209, 16;
and.b32 %r223, %r222, 30583;
prmt.b32 %r224, %r215, %r214, %r223;
prmt.b32 %r225, %r218, %r217, %r223;
shr.u32 %r226, %r212, 16;
prmt.b32 %r227, %r224, %r225, %r226;
mov.u32 %r228, 25632;
prmt.b32 %r138, %r221, %r227, %r228;
mov.u32 %r229, 30001;
prmt.b32 %r142, %r221, %r227, %r229;
ld.global.nc.u32 %r230, [%rd31+-8];
and.b32 %r231, %r230, -2004318072;
shr.u32 %r232, %r231, 1;
or.b32 %r233, %r232, 839922192;
and.b32 %r234, %r230, 30583;
prmt.b32 %r235, %r215, %r214, %r234;
prmt.b32 %r236, %r218, %r217, %r234;
and.b32 %r237, %r233, 30292;
prmt.b32 %r238, %r235, %r236, %r237;
shr.u32 %r239, %r230, 16;
and.b32 %r240, %r239, 30583;
prmt.b32 %r241, %r215, %r214, %r240;
prmt.b32 %r242, %r218, %r217, %r240;
shr.u32 %r243, %r233, 16;
prmt.b32 %r244, %r241, %r242, %r243;
prmt.b32 %r146, %r238, %r244, %r228;
prmt.b32 %r150, %r238, %r244, %r229;
add.s64 %rd32, %rd30, %rd4;
mul.lo.s64 %rd33, %rd32, 36;
add.s64 %rd34, %rd12, %rd33;
ld.global.nc.u32 %r67, [%rd34+4];
mov.u32 %r194, 0;

dp4a.s32.s32 %r65, %r138, %r67, %r194;

ld.global.nc.u32 %r71, [%rd34+12];

dp4a.s32.s32 %r69, %r142, %r71, %r65;

ld.global.nc.u32 %r75, [%rd34+8];

dp4a.s32.s32 %r73, %r146, %r75, %r69;

ld.global.nc.u32 %r79, [%rd34+16];

dp4a.s32.s32 %r77, %r150, %r79, %r73;

add.s64 %rd35, %rd60, %rd9;
ld.global.nc.u8 %rs17, [%rd35];
and.b16 %rs18, %rs17, 127;
setp.eq.s16 %p4, %rs18, 127;
selp.b16 %rs19, 0, %rs17, %p4;
and.b16 %rs1, %rs19, 255;

{cvt.rn.f16x2.e4m3x2 %r81, %rs1;}


cvt.u16.u32 %rs2, %r81;

{cvt.f32.f16 %f25, %rs2;}


mov.f32 %f41, 0f40000000;
div.approx.ftz.f32 %f42, %f25, %f41;
ld.global.nc.u32 %r82, [%rd34];

{.reg .f16 low,high;
mov.b32 {low,high},%r82;
cvt.f32.f16 %f26, low;}


mul.ftz.f32 %f43, %f42, %f26;
cvt.rn.f32.s32 %f44, %r77;
fma.rn.ftz.f32 %f45, %f43, %f44, 0f00000000;
ld.global.nc.u32 %r245, [%rd31+-4];
and.b32 %r246, %r245, -2004318072;
shr.u32 %r247, %r246, 1;
or.b32 %r248, %r247, 839922192;
and.b32 %r249, %r245, 30583;
prmt.b32 %r250, %r215, %r214, %r249;
prmt.b32 %r251, %r218, %r217, %r249;
and.b32 %r252, %r248, 30292;
prmt.b32 %r253, %r250, %r251, %r252;
shr.u32 %r254, %r245, 16;
and.b32 %r255, %r254, 30583;
prmt.b32 %r256, %r215, %r214, %r255;
prmt.b32 %r257, %r218, %r217, %r255;
shr.u32 %r258, %r248, 16;
prmt.b32 %r259, %r256, %r257, %r258;
prmt.b32 %r156, %r253, %r259, %r228;
prmt.b32 %r160, %r253, %r259, %r229;
ld.global.nc.u32 %r260, [%rd31];
and.b32 %r261, %r260, -2004318072;
shr.u32 %r262, %r261, 1;
or.b32 %r263, %r262, 839922192;
and.b32 %r264, %r260, 30583;
prmt.b32 %r265, %r215, %r214, %r264;
prmt.b32 %r266, %r218, %r217, %r264;
and.b32 %r267, %r263, 30292;
prmt.b32 %r268, %r265, %r266, %r267;
shr.u32 %r269, %r260, 16;
and.b32 %r270, %r269, 30583;
prmt.b32 %r271, %r215, %r214, %r270;
prmt.b32 %r272, %r218, %r217, %r270;
shr.u32 %r273, %r263, 16;
prmt.b32 %r274, %r271, %r272, %r273;
prmt.b32 %r164, %r268, %r274, %r228;
prmt.b32 %r168, %r268, %r274, %r229;
add.s64 %rd36, %rd30, %rd5;
mul.lo.s64 %rd37, %rd36, 36;
add.s64 %rd38, %rd12, %rd37;
ld.global.nc.u32 %r85, [%rd38+20];

dp4a.s32.s32 %r83, %r156, %r85, %r194;

ld.global.nc.u32 %r89, [%rd38+28];

dp4a.s32.s32 %r87, %r160, %r89, %r83;

ld.global.nc.u32 %r93, [%rd38+24];

dp4a.s32.s32 %r91, %r164, %r93, %r87;

ld.global.nc.u32 %r97, [%rd38+32];

dp4a.s32.s32 %r95, %r168, %r97, %r91;

add.s64 %rd39, %rd60, %rd11;
ld.global.nc.u8 %rs20, [%rd39];
and.b16 %rs21, %rs20, 127;
setp.eq.s16 %p5, %rs21, 127;
selp.b16 %rs22, 0, %rs20, %p5;
and.b16 %rs3, %rs22, 255;

{cvt.rn.f16x2.e4m3x2 %r99, %rs3;}


cvt.u16.u32 %rs4, %r99;

{cvt.f32.f16 %f27, %rs4;}


div.approx.ftz.f32 %f46, %f27, %f41;
ld.global.nc.u32 %r100, [%rd38];

{.reg .f16 low,high;
mov.b32 {low,high},%r100;
cvt.f32.f16 %f28, low;}


mul.ftz.f32 %f47, %f46, %f28;
cvt.rn.f32.s32 %f48, %r95;
fma.rn.ftz.f32 %f49, %f47, %f48, %f45;
add.ftz.f32 %f141, %f49, %f141;
add.s64 %rd40, %rd60, %rd10;
ld.global.nc.u32 %r275, [%rd40+-4];
and.b32 %r276, %r275, -2004318072;
shr.u32 %r277, %r276, 1;
or.b32 %r278, %r277, 839922192;
and.b32 %r279, %r275, 30583;
prmt.b32 %r280, %r215, %r214, %r279;
prmt.b32 %r281, %r218, %r217, %r279;
and.b32 %r282, %r278, 30292;
prmt.b32 %r283, %r280, %r281, %r282;
shr.u32 %r284, %r275, 16;
and.b32 %r285, %r284, 30583;
prmt.b32 %r286, %r215, %r214, %r285;
prmt.b32 %r287, %r218, %r217, %r285;
shr.u32 %r288, %r278, 16;
prmt.b32 %r289, %r286, %r287, %r288;
prmt.b32 %r174, %r283, %r289, %r228;
prmt.b32 %r178, %r283, %r289, %r229;
ld.global.nc.u32 %r290, [%rd40];
and.b32 %r291, %r290, -2004318072;
shr.u32 %r292, %r291, 1;
or.b32 %r293, %r292, 839922192;
and.b32 %r294, %r290, 30583;
prmt.b32 %r295, %r215, %r214, %r294;
prmt.b32 %r296, %r218, %r217, %r294;
and.b32 %r297, %r293, 30292;
prmt.b32 %r298, %r295, %r296, %r297;
shr.u32 %r299, %r290, 16;
and.b32 %r300, %r299, 30583;
prmt.b32 %r301, %r215, %r214, %r300;
prmt.b32 %r302, %r218, %r217, %r300;
shr.u32 %r303, %r293, 16;
prmt.b32 %r304, %r301, %r302, %r303;
prmt.b32 %r182, %r298, %r304, %r228;
prmt.b32 %r186, %r298, %r304, %r229;

dp4a.s32.s32 %r101, %r174, %r67, %r194;


dp4a.s32.s32 %r105, %r178, %r71, %r101;


dp4a.s32.s32 %r109, %r182, %r75, %r105;


dp4a.s32.s32 %r113, %r186, %r79, %r109;

add.s64 %rd41, %rd60, %rd7;
ld.global.nc.u8 %rs23, [%rd41];
and.b16 %rs24, %rs23, 127;
setp.eq.s16 %p6, %rs24, 127;
selp.b16 %rs25, 0, %rs23, %p6;
and.b16 %rs5, %rs25, 255;

{cvt.rn.f16x2.e4m3x2 %r117, %rs5;}


cvt.u16.u32 %rs6, %r117;

{cvt.f32.f16 %f29, %rs6;}


div.approx.ftz.f32 %f50, %f29, %f41;

{.reg .f16 low,high;
mov.b32 {low,high},%r82;
cvt.f32.f16 %f30, low;}


mul.ftz.f32 %f51, %f50, %f30;
cvt.rn.f32.s32 %f52, %r113;
fma.rn.ftz.f32 %f53, %f51, %f52, 0f00000000;
ld.global.nc.u32 %r305, [%rd40+4];
and.b32 %r306, %r305, -2004318072;
shr.u32 %r307, %r306, 1;
or.b32 %r308, %r307, 839922192;
and.b32 %r309, %r305, 30583;
prmt.b32 %r310, %r215, %r214, %r309;
prmt.b32 %r311, %r218, %r217, %r309;
and.b32 %r312, %r308, 30292;
prmt.b32 %r313, %r310, %r311, %r312;
shr.u32 %r314, %r305, 16;
and.b32 %r315, %r314, 30583;
prmt.b32 %r316, %r215, %r214, %r315;
prmt.b32 %r317, %r218, %r217, %r315;
shr.u32 %r318, %r308, 16;
prmt.b32 %r319, %r316, %r317, %r318;
prmt.b32 %r192, %r313, %r319, %r228;
prmt.b32 %r196, %r313, %r319, %r229;
ld.global.nc.u32 %r320, [%rd40+8];
and.b32 %r321, %r320, -2004318072;
shr.u32 %r322, %r321, 1;
or.b32 %r323, %r322, 839922192;
and.b32 %r324, %r320, 30583;
prmt.b32 %r325, %r215, %r214, %r324;
prmt.b32 %r326, %r218, %r217, %r324;
and.b32 %r327, %r323, 30292;
prmt.b32 %r328, %r325, %r326, %r327;
shr.u32 %r329, %r320, 16;
and.b32 %r330, %r329, 30583;
prmt.b32 %r331, %r215, %r214, %r330;
prmt.b32 %r332, %r218, %r217, %r330;
shr.u32 %r333, %r323, 16;
prmt.b32 %r334, %r331, %r332, %r333;
prmt.b32 %r200, %r328, %r334, %r228;
prmt.b32 %r204, %r328, %r334, %r229;

dp4a.s32.s32 %r119, %r192, %r85, %r194;


dp4a.s32.s32 %r123, %r196, %r89, %r119;


dp4a.s32.s32 %r127, %r200, %r93, %r123;


dp4a.s32.s32 %r131, %r204, %r97, %r127;

add.s64 %rd42, %rd60, %rd6;
ld.global.nc.u8 %rs26, [%rd42];
and.b16 %rs27, %rs26, 127;
setp.eq.s16 %p7, %rs27, 127;
selp.b16 %rs28, 0, %rs26, %p7;
and.b16 %rs7, %rs28, 255;

{cvt.rn.f16x2.e4m3x2 %r135, %rs7;}


cvt.u16.u32 %rs8, %r135;

{cvt.f32.f16 %f31, %rs8;}


div.approx.ftz.f32 %f54, %f31, %f41;

{.reg .f16 low,high;
mov.b32 {low,high},%r100;
cvt.f32.f16 %f32, low;}


mul.ftz.f32 %f55, %f54, %f32;
cvt.rn.f32.s32 %f56, %r131;
fma.rn.ftz.f32 %f57, %f55, %f56, %f53;
add.ftz.f32 %f140, %f57, %f140;
cvt.u64.u32 %rd43, %r417;
add.s64 %rd44, %rd2, %rd43;
add.s64 %rd45, %rd44, %rd4;
mul.lo.s64 %rd46, %rd45, 36;
add.s64 %rd47, %rd12, %rd46;
ld.global.nc.u32 %r139, [%rd47+4];

dp4a.s32.s32 %r137, %r138, %r139, %r194;

ld.global.nc.u32 %r143, [%rd47+12];

dp4a.s32.s32 %r141, %r142, %r143, %r137;

ld.global.nc.u32 %r147, [%rd47+8];

dp4a.s32.s32 %r145, %r146, %r147, %r141;

ld.global.nc.u32 %r151, [%rd47+16];

dp4a.s32.s32 %r149, %r150, %r151, %r145;


{cvt.rn.f16x2.e4m3x2 %r153, %rs1;}


cvt.u16.u32 %rs10, %r153;

{cvt.f32.f16 %f33, %rs10;}


div.approx.ftz.f32 %f58, %f33, %f41;
ld.global.nc.u32 %r154, [%rd47];

{.reg .f16 low,high;
mov.b32 {low,high},%r154;
cvt.f32.f16 %f34, low;}


mul.ftz.f32 %f59, %f58, %f34;
cvt.rn.f32.s32 %f60, %r149;
fma.rn.ftz.f32 %f61, %f59, %f60, 0f00000000;
add.s64 %rd48, %rd44, %rd5;
mul.lo.s64 %rd49, %rd48, 36;
add.s64 %rd50, %rd12, %rd49;
ld.global.nc.u32 %r157, [%rd50+20];

dp4a.s32.s32 %r155, %r156, %r157, %r194;

ld.global.nc.u32 %r161, [%rd50+28];

dp4a.s32.s32 %r159, %r160, %r161, %r155;

ld.global.nc.u32 %r165, [%rd50+24];

dp4a.s32.s32 %r163, %r164, %r165, %r159;

ld.global.nc.u32 %r169, [%rd50+32];

dp4a.s32.s32 %r167, %r168, %r169, %r163;


{cvt.rn.f16x2.e4m3x2 %r171, %rs3;}


cvt.u16.u32 %rs12, %r171;

{cvt.f32.f16 %f35, %rs12;}


div.approx.ftz.f32 %f62, %f35, %f41;
ld.global.nc.u32 %r172, [%rd50];

{.reg .f16 low,high;
mov.b32 {low,high},%r172;
cvt.f32.f16 %f36, low;}


mul.ftz.f32 %f63, %f62, %f36;
cvt.rn.f32.s32 %f64, %r167;
fma.rn.ftz.f32 %f65, %f63, %f64, %f61;
add.ftz.f32 %f139, %f65, %f139;

dp4a.s32.s32 %r173, %r174, %r139, %r194;


dp4a.s32.s32 %r177, %r178, %r143, %r173;


dp4a.s32.s32 %r181, %r182, %r147, %r177;


dp4a.s32.s32 %r185, %r186, %r151, %r181;


{cvt.rn.f16x2.e4m3x2 %r189, %rs5;}


cvt.u16.u32 %rs14, %r189;

{cvt.f32.f16 %f37, %rs14;}


div.approx.ftz.f32 %f66, %f37, %f41;

{.reg .f16 low,high;
mov.b32 {low,high},%r154;
cvt.f32.f16 %f38, low;}


mul.ftz.f32 %f67, %f66, %f38;
cvt.rn.f32.s32 %f68, %r185;
fma.rn.ftz.f32 %f69, %f67, %f68, 0f00000000;

dp4a.s32.s32 %r191, %r192, %r157, %r194;


dp4a.s32.s32 %r195, %r196, %r161, %r191;


dp4a.s32.s32 %r199, %r200, %r165, %r195;


dp4a.s32.s32 %r203, %r204, %r169, %r199;


{cvt.rn.f16x2.e4m3x2 %r207, %rs7;}


cvt.u16.u32 %rs16, %r207;

{cvt.f32.f16 %f39, %rs16;}


div.approx.ftz.f32 %f70, %f39, %f41;

{.reg .f16 low,high;
mov.b32 {low,high},%r172;
cvt.f32.f16 %f40, low;}


mul.ftz.f32 %f71, %f70, %f40;
cvt.rn.f32.s32 %f72, %r203;
fma.rn.ftz.f32 %f73, %f71, %f72, %f69;
add.ftz.f32 %f138, %f73, %f138;
add.s32 %r417, %r417, 128;
add.s32 %r416, %r416, 128;
add.s64 %rd60, %rd60, 2304;
add.s32 %r418, %r418, 64;
setp.lt.s32 %p8, %r418, %r6;
@%p8 bra $L__BB120_2;

$L__BB120_3:
setp.eq.s32 %p9, %r1, 0;
@%p9 bra $L__BB120_5;

shl.b32 %r335, %r1, 9;
mov.u32 %r336, _ZZ13mul_mat_vec_qIL9ggml_type40ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r337, %r336, %r335;
shl.b32 %r338, %r2, 2;
add.s32 %r339, %r337, %r338;
st.shared.f32 [%r339+-512], %f141;
st.shared.f32 [%r339+-384], %f140;
st.shared.f32 [%r339+-256], %f139;
st.shared.f32 [%r339+-128], %f138;

$L__BB120_5:
setp.ne.s32 %p10, %r1, 0;
bar.sync 0;
@%p10 bra $L__BB120_14;

mov.u32 %r340, %ctaid.z;
mov.u32 %r341, %ctaid.y;
mad.lo.s32 %r342, %r341, %r25, %r4;
mad.lo.s32 %r343, %r340, %r31, %r342;
cvt.u64.u32 %rd15, %r343;
shl.b32 %r344, %r2, 2;
mov.u32 %r345, _ZZ13mul_mat_vec_qIL9ggml_type40ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r15, %r345, %r344;
mov.u32 %r346, 2;
ld.shared.f32 %f74, [%r15];
add.ftz.f32 %f75, %f74, %f141;
ld.shared.f32 %f76, [%r15+512];
add.ftz.f32 %f77, %f76, %f75;
ld.shared.f32 %f78, [%r15+1024];
add.ftz.f32 %f79, %f78, %f77;
mov.b32 %r347, %f79;
mov.u32 %r348, 31;
mov.u32 %r349, 16;
mov.u32 %r350, -1;
shfl.sync.bfly.b32 %r351|%p11, %r347, %r349, %r348, %r350;
mov.b32 %f80, %r351;
add.ftz.f32 %f81, %f79, %f80;
mov.b32 %r352, %f81;
mov.u32 %r353, 8;
shfl.sync.bfly.b32 %r354|%p12, %r352, %r353, %r348, %r350;
mov.b32 %f82, %r354;
add.ftz.f32 %f83, %f81, %f82;
mov.b32 %r355, %f83;
mov.u32 %r356, 4;
shfl.sync.bfly.b32 %r357|%p13, %r355, %r356, %r348, %r350;
mov.b32 %f84, %r357;
add.ftz.f32 %f85, %f83, %f84;
mov.b32 %r358, %f85;
shfl.sync.bfly.b32 %r359|%p14, %r358, %r346, %r348, %r350;
mov.b32 %f86, %r359;
add.ftz.f32 %f87, %f85, %f86;
mov.b32 %r360, %f87;
mov.u32 %r361, 1;
shfl.sync.bfly.b32 %r362|%p15, %r360, %r361, %r348, %r350;
mov.b32 %f88, %r362;
add.ftz.f32 %f13, %f87, %f88;
setp.lt.u32 %p16, %r4, %r19;
setp.eq.s32 %p17, %r2, 0;
and.pred %p1, %p17, %p16;
mul.wide.u32 %rd51, %r343, 4;
add.s64 %rd16, %rd1, %rd51;
not.pred %p18, %p1;
@%p18 bra $L__BB120_8;

st.global.f32 [%rd16], %f13;

$L__BB120_8:
ld.shared.f32 %f89, [%r15+128];
add.ftz.f32 %f90, %f89, %f140;
ld.shared.f32 %f91, [%r15+640];
add.ftz.f32 %f92, %f91, %f90;
ld.shared.f32 %f93, [%r15+1152];
add.ftz.f32 %f94, %f93, %f92;
mov.b32 %r363, %f94;
shfl.sync.bfly.b32 %r367|%p19, %r363, %r349, %r348, %r350;
mov.b32 %f95, %r367;
add.ftz.f32 %f96, %f94, %f95;
mov.b32 %r368, %f96;
shfl.sync.bfly.b32 %r370|%p20, %r368, %r353, %r348, %r350;
mov.b32 %f97, %r370;
add.ftz.f32 %f98, %f96, %f97;
mov.b32 %r371, %f98;
shfl.sync.bfly.b32 %r373|%p21, %r371, %r356, %r348, %r350;
mov.b32 %f99, %r373;
add.ftz.f32 %f100, %f98, %f99;
mov.b32 %r374, %f100;
shfl.sync.bfly.b32 %r376|%p22, %r374, %r346, %r348, %r350;
mov.b32 %f101, %r376;
add.ftz.f32 %f102, %f100, %f101;
mov.b32 %r377, %f102;
shfl.sync.bfly.b32 %r379|%p23, %r377, %r361, %r348, %r350;
mov.b32 %f103, %r379;
add.ftz.f32 %f14, %f102, %f103;
add.s32 %r380, %r4, 1;
setp.lt.u32 %p24, %r380, %r19;
setp.eq.s32 %p25, %r2, 1;
and.pred %p2, %p25, %p24;
not.pred %p26, %p2;
@%p26 bra $L__BB120_10;

st.global.f32 [%rd16+4], %f14;

$L__BB120_10:
ld.shared.f32 %f104, [%r15+256];
add.ftz.f32 %f105, %f104, %f139;
ld.shared.f32 %f106, [%r15+768];
add.ftz.f32 %f107, %f106, %f105;
ld.shared.f32 %f108, [%r15+1280];
add.ftz.f32 %f109, %f108, %f107;
mov.b32 %r381, %f109;
mov.u32 %r382, 31;
mov.u32 %r383, 16;
mov.u32 %r384, -1;
shfl.sync.bfly.b32 %r385|%p27, %r381, %r383, %r382, %r384;
mov.b32 %f110, %r385;
add.ftz.f32 %f111, %f109, %f110;
mov.b32 %r386, %f111;
mov.u32 %r387, 8;
shfl.sync.bfly.b32 %r388|%p28, %r386, %r387, %r382, %r384;
mov.b32 %f112, %r388;
add.ftz.f32 %f113, %f111, %f112;
mov.b32 %r389, %f113;
mov.u32 %r390, 4;
shfl.sync.bfly.b32 %r391|%p29, %r389, %r390, %r382, %r384;
mov.b32 %f114, %r391;
add.ftz.f32 %f115, %f113, %f114;
mov.b32 %r392, %f115;
mov.u32 %r393, 2;
shfl.sync.bfly.b32 %r394|%p30, %r392, %r393, %r382, %r384;
mov.b32 %f116, %r394;
add.ftz.f32 %f117, %f115, %f116;
mov.b32 %r395, %f117;
mov.u32 %r396, 1;
shfl.sync.bfly.b32 %r397|%p31, %r395, %r396, %r382, %r384;
mov.b32 %f118, %r397;
add.ftz.f32 %f15, %f117, %f118;
@%p18 bra $L__BB120_12;

cvt.u64.u32 %rd52, %r19;
add.s64 %rd53, %rd52, %rd15;
shl.b64 %rd54, %rd53, 2;
add.s64 %rd55, %rd1, %rd54;
st.global.f32 [%rd55], %f15;

$L__BB120_12:
ld.shared.f32 %f119, [%r15+384];
add.ftz.f32 %f120, %f119, %f138;
ld.shared.f32 %f121, [%r15+896];
add.ftz.f32 %f122, %f121, %f120;
ld.shared.f32 %f123, [%r15+1408];
add.ftz.f32 %f124, %f123, %f122;
mov.b32 %r398, %f124;
shfl.sync.bfly.b32 %r402|%p33, %r398, %r383, %r382, %r384;
mov.b32 %f125, %r402;
add.ftz.f32 %f126, %f124, %f125;
mov.b32 %r403, %f126;
shfl.sync.bfly.b32 %r405|%p34, %r403, %r387, %r382, %r384;
mov.b32 %f127, %r405;
add.ftz.f32 %f128, %f126, %f127;
mov.b32 %r406, %f128;
shfl.sync.bfly.b32 %r408|%p35, %r406, %r390, %r382, %r384;
mov.b32 %f129, %r408;
add.ftz.f32 %f130, %f128, %f129;
mov.b32 %r409, %f130;
shfl.sync.bfly.b32 %r411|%p36, %r409, %r393, %r382, %r384;
mov.b32 %f131, %r411;
add.ftz.f32 %f132, %f130, %f131;
mov.b32 %r412, %f132;
shfl.sync.bfly.b32 %r414|%p37, %r412, %r396, %r382, %r384;
mov.b32 %f133, %r414;
add.ftz.f32 %f16, %f132, %f133;
@%p26 bra $L__BB120_14;

add.s32 %r415, %r19, 1;
cvt.u64.u32 %rd56, %r415;
add.s64 %rd57, %rd56, %rd15;
shl.b64 %rd58, %rd57, 2;
add.s64 %rd59, %rd1, %rd58;
st.global.f32 [%rd59], %f16;

$L__BB120_14:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type40ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type40ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type40ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type40ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type40ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type40ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type40ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type40ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type40ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<51>;
.reg .b16 %rs<37>;
.reg .f32 %f<212>;
.reg .b32 %r<530>;
.reg .b64 %rd<73>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type40ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[2304];

ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type40ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type40ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd20, [_Z13mul_mat_vec_qIL9ggml_type40ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type40ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type40ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type40ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type40ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type40ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type40ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type40ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type40ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type40ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type40ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type40ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type40ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type40ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type40ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
cvta.to.global.u64 %rd1, %rd20;
mov.u32 %r1, %tid.y;
shl.b32 %r36, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r36, %r2;
mov.u32 %r37, %ctaid.x;
shl.b32 %r4, %r37, 1;
shr.u32 %r38, %r3, 31;
add.s32 %r39, %r3, %r38;
shr.s32 %r529, %r39, 1;
shr.u32 %r6, %r35, 6;
setp.ge.s32 %p3, %r529, %r6;
mov.f32 %f206, 0f00000000;
mov.f32 %f207, %f206;
mov.f32 %f208, %f206;
mov.f32 %f209, %f206;
mov.f32 %f210, %f206;
mov.f32 %f211, %f206;
@%p3 bra $L__BB121_3;

mov.u32 %r40, %ctaid.y;
mov.u32 %r41, %ctaid.z;
mul.lo.s32 %r42, %r40, %r27;
cvt.u64.u32 %rd21, %r42;
mul.lo.s32 %r43, %r41, %r33;
cvt.u64.u32 %rd22, %r43;
add.s64 %rd2, %rd22, %rd21;
mul.hi.u32 %r44, %r41, %r29;
add.s32 %r45, %r44, %r41;
shr.u32 %r46, %r45, %r30;
mul.lo.s32 %r47, %r46, %r32;
mul.hi.u32 %r48, %r40, %r23;
add.s32 %r49, %r48, %r40;
shr.u32 %r50, %r49, %r24;
mul.lo.s32 %r51, %r50, %r26;
cvta.to.global.u64 %rd72, %rd18;
and.b32 %r54, %r39, -2;
sub.s32 %r55, %r3, %r54;
shl.b32 %r56, %r55, 2;
shl.b32 %r57, %r55, 1;
cvt.s64.s32 %rd4, %r55;
cvt.s64.s32 %rd23, %r57;
or.b32 %r58, %r56, 2;
shr.s32 %r59, %r58, 1;
shr.s32 %r60, %r58, 2;
cvt.s64.s32 %rd5, %r60;
cvt.s64.s32 %rd24, %r59;
shl.b32 %r61, %r21, 1;
shl.b32 %r526, %r529, 1;
add.s32 %r528, %r526, %r61;
add.s32 %r527, %r21, %r526;
add.s32 %r62, %r529, %r47;
add.s32 %r63, %r62, %r51;
add.s32 %r64, %r4, 1;
mad.lo.s32 %r65, %r20, %r64, %r63;
mul.wide.s32 %rd25, %r65, 36;
add.s64 %rd6, %rd25, %rd24;
add.s64 %rd7, %rd25, %rd23;
mad.lo.s32 %r66, %r4, %r20, %r529;
add.s32 %r67, %r66, %r47;
add.s32 %r68, %r67, %r51;
mul.wide.s32 %rd26, %r68, 36;
mul.wide.s32 %rd27, %r56, 4;
add.s64 %rd28, %rd26, %rd27;
add.s64 %rd8, %rd28, 16;
add.s64 %rd9, %rd26, %rd23;
add.s64 %rd29, %rd25, %rd27;
add.s64 %rd10, %rd29, 8;
add.s64 %rd11, %rd26, %rd24;
cvta.to.global.u64 %rd12, %rd19;
mov.f32 %f206, 0f00000000;

$L__BB121_2:
cvt.u64.u32 %rd30, %r526;
add.s64 %rd31, %rd2, %rd30;
add.s64 %rd32, %rd72, %rd8;
ld.global.nc.u32 %r285, [%rd32+-12];
and.b32 %r286, %r285, -2004318072;
shr.u32 %r287, %r286, 1;
or.b32 %r288, %r287, 839922192;
and.b32 %r289, %r285, 30583;
mov.u32 %r290, 201852420;
mov.u32 %r291, 50462976;
prmt.b32 %r292, %r291, %r290, %r289;
mov.u32 %r293, -185009412;
mov.u32 %r294, -33620224;
prmt.b32 %r295, %r294, %r293, %r289;
and.b32 %r296, %r288, 30292;
prmt.b32 %r297, %r292, %r295, %r296;
shr.u32 %r298, %r285, 16;
and.b32 %r299, %r298, 30583;
prmt.b32 %r300, %r291, %r290, %r299;
prmt.b32 %r301, %r294, %r293, %r299;
shr.u32 %r302, %r288, 16;
prmt.b32 %r303, %r300, %r301, %r302;
mov.u32 %r304, 25632;
prmt.b32 %r214, %r297, %r303, %r304;
mov.u32 %r305, 30001;
prmt.b32 %r218, %r297, %r303, %r305;
ld.global.nc.u32 %r306, [%rd32+-8];
and.b32 %r307, %r306, -2004318072;
shr.u32 %r308, %r307, 1;
or.b32 %r309, %r308, 839922192;
and.b32 %r310, %r306, 30583;
prmt.b32 %r311, %r291, %r290, %r310;
prmt.b32 %r312, %r294, %r293, %r310;
and.b32 %r313, %r309, 30292;
prmt.b32 %r314, %r311, %r312, %r313;
shr.u32 %r315, %r306, 16;
and.b32 %r316, %r315, 30583;
prmt.b32 %r317, %r291, %r290, %r316;
prmt.b32 %r318, %r294, %r293, %r316;
shr.u32 %r319, %r309, 16;
prmt.b32 %r320, %r317, %r318, %r319;
prmt.b32 %r222, %r314, %r320, %r304;
prmt.b32 %r226, %r314, %r320, %r305;
add.s64 %rd33, %rd31, %rd4;
mul.lo.s64 %rd34, %rd33, 36;
add.s64 %rd35, %rd12, %rd34;
ld.global.nc.u32 %r71, [%rd35+4];
mov.u32 %r270, 0;

dp4a.s32.s32 %r69, %r214, %r71, %r270;

ld.global.nc.u32 %r75, [%rd35+12];

dp4a.s32.s32 %r73, %r218, %r75, %r69;

ld.global.nc.u32 %r79, [%rd35+8];

dp4a.s32.s32 %r77, %r222, %r79, %r73;

ld.global.nc.u32 %r83, [%rd35+16];

dp4a.s32.s32 %r81, %r226, %r83, %r77;

add.s64 %rd36, %rd72, %rd9;
ld.global.nc.u8 %rs25, [%rd36];
and.b16 %rs26, %rs25, 127;
setp.eq.s16 %p4, %rs26, 127;
selp.b16 %rs27, 0, %rs25, %p4;
and.b16 %rs1, %rs27, 255;

{cvt.rn.f16x2.e4m3x2 %r85, %rs1;}


cvt.u16.u32 %rs2, %r85;

{cvt.f32.f16 %f37, %rs2;}


mov.f32 %f61, 0f40000000;
div.approx.ftz.f32 %f62, %f37, %f61;
ld.global.nc.u32 %r86, [%rd35];

{.reg .f16 low,high;
mov.b32 {low,high},%r86;
cvt.f32.f16 %f38, low;}


mul.ftz.f32 %f63, %f62, %f38;
cvt.rn.f32.s32 %f64, %r81;
fma.rn.ftz.f32 %f65, %f63, %f64, 0f00000000;
ld.global.nc.u32 %r321, [%rd32+-4];
and.b32 %r322, %r321, -2004318072;
shr.u32 %r323, %r322, 1;
or.b32 %r324, %r323, 839922192;
and.b32 %r325, %r321, 30583;
prmt.b32 %r326, %r291, %r290, %r325;
prmt.b32 %r327, %r294, %r293, %r325;
and.b32 %r328, %r324, 30292;
prmt.b32 %r329, %r326, %r327, %r328;
shr.u32 %r330, %r321, 16;
and.b32 %r331, %r330, 30583;
prmt.b32 %r332, %r291, %r290, %r331;
prmt.b32 %r333, %r294, %r293, %r331;
shr.u32 %r334, %r324, 16;
prmt.b32 %r335, %r332, %r333, %r334;
prmt.b32 %r232, %r329, %r335, %r304;
prmt.b32 %r236, %r329, %r335, %r305;
ld.global.nc.u32 %r336, [%rd32];
and.b32 %r337, %r336, -2004318072;
shr.u32 %r338, %r337, 1;
or.b32 %r339, %r338, 839922192;
and.b32 %r340, %r336, 30583;
prmt.b32 %r341, %r291, %r290, %r340;
prmt.b32 %r342, %r294, %r293, %r340;
and.b32 %r343, %r339, 30292;
prmt.b32 %r344, %r341, %r342, %r343;
shr.u32 %r345, %r336, 16;
and.b32 %r346, %r345, 30583;
prmt.b32 %r347, %r291, %r290, %r346;
prmt.b32 %r348, %r294, %r293, %r346;
shr.u32 %r349, %r339, 16;
prmt.b32 %r350, %r347, %r348, %r349;
prmt.b32 %r240, %r344, %r350, %r304;
prmt.b32 %r244, %r344, %r350, %r305;
add.s64 %rd37, %rd31, %rd5;
mul.lo.s64 %rd38, %rd37, 36;
add.s64 %rd39, %rd12, %rd38;
ld.global.nc.u32 %r89, [%rd39+20];

dp4a.s32.s32 %r87, %r232, %r89, %r270;

ld.global.nc.u32 %r93, [%rd39+28];

dp4a.s32.s32 %r91, %r236, %r93, %r87;

ld.global.nc.u32 %r97, [%rd39+24];

dp4a.s32.s32 %r95, %r240, %r97, %r91;

ld.global.nc.u32 %r101, [%rd39+32];

dp4a.s32.s32 %r99, %r244, %r101, %r95;

add.s64 %rd40, %rd72, %rd11;
ld.global.nc.u8 %rs28, [%rd40];
and.b16 %rs29, %rs28, 127;
setp.eq.s16 %p5, %rs29, 127;
selp.b16 %rs30, 0, %rs28, %p5;
and.b16 %rs3, %rs30, 255;

{cvt.rn.f16x2.e4m3x2 %r103, %rs3;}


cvt.u16.u32 %rs4, %r103;

{cvt.f32.f16 %f39, %rs4;}


div.approx.ftz.f32 %f66, %f39, %f61;
ld.global.nc.u32 %r104, [%rd39];

{.reg .f16 low,high;
mov.b32 {low,high},%r104;
cvt.f32.f16 %f40, low;}


mul.ftz.f32 %f67, %f66, %f40;
cvt.rn.f32.s32 %f68, %r99;
fma.rn.ftz.f32 %f69, %f67, %f68, %f65;
add.ftz.f32 %f211, %f69, %f211;
add.s64 %rd41, %rd72, %rd10;
ld.global.nc.u32 %r351, [%rd41+-4];
and.b32 %r352, %r351, -2004318072;
shr.u32 %r353, %r352, 1;
or.b32 %r354, %r353, 839922192;
and.b32 %r355, %r351, 30583;
prmt.b32 %r356, %r291, %r290, %r355;
prmt.b32 %r357, %r294, %r293, %r355;
and.b32 %r358, %r354, 30292;
prmt.b32 %r359, %r356, %r357, %r358;
shr.u32 %r360, %r351, 16;
and.b32 %r361, %r360, 30583;
prmt.b32 %r362, %r291, %r290, %r361;
prmt.b32 %r363, %r294, %r293, %r361;
shr.u32 %r364, %r354, 16;
prmt.b32 %r365, %r362, %r363, %r364;
prmt.b32 %r250, %r359, %r365, %r304;
prmt.b32 %r254, %r359, %r365, %r305;
ld.global.nc.u32 %r366, [%rd41];
and.b32 %r367, %r366, -2004318072;
shr.u32 %r368, %r367, 1;
or.b32 %r369, %r368, 839922192;
and.b32 %r370, %r366, 30583;
prmt.b32 %r371, %r291, %r290, %r370;
prmt.b32 %r372, %r294, %r293, %r370;
and.b32 %r373, %r369, 30292;
prmt.b32 %r374, %r371, %r372, %r373;
shr.u32 %r375, %r366, 16;
and.b32 %r376, %r375, 30583;
prmt.b32 %r377, %r291, %r290, %r376;
prmt.b32 %r378, %r294, %r293, %r376;
shr.u32 %r379, %r369, 16;
prmt.b32 %r380, %r377, %r378, %r379;
prmt.b32 %r258, %r374, %r380, %r304;
prmt.b32 %r262, %r374, %r380, %r305;

dp4a.s32.s32 %r105, %r250, %r71, %r270;


dp4a.s32.s32 %r109, %r254, %r75, %r105;


dp4a.s32.s32 %r113, %r258, %r79, %r109;


dp4a.s32.s32 %r117, %r262, %r83, %r113;

add.s64 %rd42, %rd72, %rd7;
ld.global.nc.u8 %rs31, [%rd42];
and.b16 %rs32, %rs31, 127;
setp.eq.s16 %p6, %rs32, 127;
selp.b16 %rs33, 0, %rs31, %p6;
and.b16 %rs5, %rs33, 255;

{cvt.rn.f16x2.e4m3x2 %r121, %rs5;}


cvt.u16.u32 %rs6, %r121;

{cvt.f32.f16 %f41, %rs6;}


div.approx.ftz.f32 %f70, %f41, %f61;

{.reg .f16 low,high;
mov.b32 {low,high},%r86;
cvt.f32.f16 %f42, low;}


mul.ftz.f32 %f71, %f70, %f42;
cvt.rn.f32.s32 %f72, %r117;
fma.rn.ftz.f32 %f73, %f71, %f72, 0f00000000;
ld.global.nc.u32 %r381, [%rd41+4];
and.b32 %r382, %r381, -2004318072;
shr.u32 %r383, %r382, 1;
or.b32 %r384, %r383, 839922192;
and.b32 %r385, %r381, 30583;
prmt.b32 %r386, %r291, %r290, %r385;
prmt.b32 %r387, %r294, %r293, %r385;
and.b32 %r388, %r384, 30292;
prmt.b32 %r389, %r386, %r387, %r388;
shr.u32 %r390, %r381, 16;
and.b32 %r391, %r390, 30583;
prmt.b32 %r392, %r291, %r290, %r391;
prmt.b32 %r393, %r294, %r293, %r391;
shr.u32 %r394, %r384, 16;
prmt.b32 %r395, %r392, %r393, %r394;
prmt.b32 %r268, %r389, %r395, %r304;
prmt.b32 %r272, %r389, %r395, %r305;
ld.global.nc.u32 %r396, [%rd41+8];
and.b32 %r397, %r396, -2004318072;
shr.u32 %r398, %r397, 1;
or.b32 %r399, %r398, 839922192;
and.b32 %r400, %r396, 30583;
prmt.b32 %r401, %r291, %r290, %r400;
prmt.b32 %r402, %r294, %r293, %r400;
and.b32 %r403, %r399, 30292;
prmt.b32 %r404, %r401, %r402, %r403;
shr.u32 %r405, %r396, 16;
and.b32 %r406, %r405, 30583;
prmt.b32 %r407, %r291, %r290, %r406;
prmt.b32 %r408, %r294, %r293, %r406;
shr.u32 %r409, %r399, 16;
prmt.b32 %r410, %r407, %r408, %r409;
prmt.b32 %r276, %r404, %r410, %r304;
prmt.b32 %r280, %r404, %r410, %r305;

dp4a.s32.s32 %r123, %r268, %r89, %r270;


dp4a.s32.s32 %r127, %r272, %r93, %r123;


dp4a.s32.s32 %r131, %r276, %r97, %r127;


dp4a.s32.s32 %r135, %r280, %r101, %r131;

add.s64 %rd43, %rd72, %rd6;
ld.global.nc.u8 %rs34, [%rd43];
and.b16 %rs35, %rs34, 127;
setp.eq.s16 %p7, %rs35, 127;
selp.b16 %rs36, 0, %rs34, %p7;
and.b16 %rs7, %rs36, 255;

{cvt.rn.f16x2.e4m3x2 %r139, %rs7;}


cvt.u16.u32 %rs8, %r139;

{cvt.f32.f16 %f43, %rs8;}


div.approx.ftz.f32 %f74, %f43, %f61;

{.reg .f16 low,high;
mov.b32 {low,high},%r104;
cvt.f32.f16 %f44, low;}


mul.ftz.f32 %f75, %f74, %f44;
cvt.rn.f32.s32 %f76, %r135;
fma.rn.ftz.f32 %f77, %f75, %f76, %f73;
add.ftz.f32 %f210, %f77, %f210;
cvt.u64.u32 %rd44, %r527;
add.s64 %rd45, %rd2, %rd44;
add.s64 %rd46, %rd45, %rd4;
mul.lo.s64 %rd47, %rd46, 36;
add.s64 %rd48, %rd12, %rd47;
ld.global.nc.u32 %r143, [%rd48+4];

dp4a.s32.s32 %r141, %r214, %r143, %r270;

ld.global.nc.u32 %r147, [%rd48+12];

dp4a.s32.s32 %r145, %r218, %r147, %r141;

ld.global.nc.u32 %r151, [%rd48+8];

dp4a.s32.s32 %r149, %r222, %r151, %r145;

ld.global.nc.u32 %r155, [%rd48+16];

dp4a.s32.s32 %r153, %r226, %r155, %r149;


{cvt.rn.f16x2.e4m3x2 %r157, %rs1;}


cvt.u16.u32 %rs10, %r157;

{cvt.f32.f16 %f45, %rs10;}


div.approx.ftz.f32 %f78, %f45, %f61;
ld.global.nc.u32 %r158, [%rd48];

{.reg .f16 low,high;
mov.b32 {low,high},%r158;
cvt.f32.f16 %f46, low;}


mul.ftz.f32 %f79, %f78, %f46;
cvt.rn.f32.s32 %f80, %r153;
fma.rn.ftz.f32 %f81, %f79, %f80, 0f00000000;
add.s64 %rd49, %rd45, %rd5;
mul.lo.s64 %rd50, %rd49, 36;
add.s64 %rd51, %rd12, %rd50;
ld.global.nc.u32 %r161, [%rd51+20];

dp4a.s32.s32 %r159, %r232, %r161, %r270;

ld.global.nc.u32 %r165, [%rd51+28];

dp4a.s32.s32 %r163, %r236, %r165, %r159;

ld.global.nc.u32 %r169, [%rd51+24];

dp4a.s32.s32 %r167, %r240, %r169, %r163;

ld.global.nc.u32 %r173, [%rd51+32];

dp4a.s32.s32 %r171, %r244, %r173, %r167;


{cvt.rn.f16x2.e4m3x2 %r175, %rs3;}


cvt.u16.u32 %rs12, %r175;

{cvt.f32.f16 %f47, %rs12;}


div.approx.ftz.f32 %f82, %f47, %f61;
ld.global.nc.u32 %r176, [%rd51];

{.reg .f16 low,high;
mov.b32 {low,high},%r176;
cvt.f32.f16 %f48, low;}


mul.ftz.f32 %f83, %f82, %f48;
cvt.rn.f32.s32 %f84, %r171;
fma.rn.ftz.f32 %f85, %f83, %f84, %f81;
add.ftz.f32 %f209, %f85, %f209;

dp4a.s32.s32 %r177, %r250, %r143, %r270;


dp4a.s32.s32 %r181, %r254, %r147, %r177;


dp4a.s32.s32 %r185, %r258, %r151, %r181;


dp4a.s32.s32 %r189, %r262, %r155, %r185;


{cvt.rn.f16x2.e4m3x2 %r193, %rs5;}


cvt.u16.u32 %rs14, %r193;

{cvt.f32.f16 %f49, %rs14;}


div.approx.ftz.f32 %f86, %f49, %f61;

{.reg .f16 low,high;
mov.b32 {low,high},%r158;
cvt.f32.f16 %f50, low;}


mul.ftz.f32 %f87, %f86, %f50;
cvt.rn.f32.s32 %f88, %r189;
fma.rn.ftz.f32 %f89, %f87, %f88, 0f00000000;

dp4a.s32.s32 %r195, %r268, %r161, %r270;


dp4a.s32.s32 %r199, %r272, %r165, %r195;


dp4a.s32.s32 %r203, %r276, %r169, %r199;


dp4a.s32.s32 %r207, %r280, %r173, %r203;


{cvt.rn.f16x2.e4m3x2 %r211, %rs7;}


cvt.u16.u32 %rs16, %r211;

{cvt.f32.f16 %f51, %rs16;}


div.approx.ftz.f32 %f90, %f51, %f61;

{.reg .f16 low,high;
mov.b32 {low,high},%r176;
cvt.f32.f16 %f52, low;}


mul.ftz.f32 %f91, %f90, %f52;
cvt.rn.f32.s32 %f92, %r207;
fma.rn.ftz.f32 %f93, %f91, %f92, %f89;
add.ftz.f32 %f208, %f93, %f208;
cvt.u64.u32 %rd52, %r528;
add.s64 %rd53, %rd2, %rd52;
add.s64 %rd54, %rd53, %rd4;
mul.lo.s64 %rd55, %rd54, 36;
add.s64 %rd56, %rd12, %rd55;
ld.global.nc.u32 %r215, [%rd56+4];

dp4a.s32.s32 %r213, %r214, %r215, %r270;

ld.global.nc.u32 %r219, [%rd56+12];

dp4a.s32.s32 %r217, %r218, %r219, %r213;

ld.global.nc.u32 %r223, [%rd56+8];

dp4a.s32.s32 %r221, %r222, %r223, %r217;

ld.global.nc.u32 %r227, [%rd56+16];

dp4a.s32.s32 %r225, %r226, %r227, %r221;


{cvt.rn.f16x2.e4m3x2 %r229, %rs1;}


cvt.u16.u32 %rs18, %r229;

{cvt.f32.f16 %f53, %rs18;}


div.approx.ftz.f32 %f94, %f53, %f61;
ld.global.nc.u32 %r230, [%rd56];

{.reg .f16 low,high;
mov.b32 {low,high},%r230;
cvt.f32.f16 %f54, low;}


mul.ftz.f32 %f95, %f94, %f54;
cvt.rn.f32.s32 %f96, %r225;
fma.rn.ftz.f32 %f97, %f95, %f96, 0f00000000;
add.s64 %rd57, %rd53, %rd5;
mul.lo.s64 %rd58, %rd57, 36;
add.s64 %rd59, %rd12, %rd58;
ld.global.nc.u32 %r233, [%rd59+20];

dp4a.s32.s32 %r231, %r232, %r233, %r270;

ld.global.nc.u32 %r237, [%rd59+28];

dp4a.s32.s32 %r235, %r236, %r237, %r231;

ld.global.nc.u32 %r241, [%rd59+24];

dp4a.s32.s32 %r239, %r240, %r241, %r235;

ld.global.nc.u32 %r245, [%rd59+32];

dp4a.s32.s32 %r243, %r244, %r245, %r239;


{cvt.rn.f16x2.e4m3x2 %r247, %rs3;}


cvt.u16.u32 %rs20, %r247;

{cvt.f32.f16 %f55, %rs20;}


div.approx.ftz.f32 %f98, %f55, %f61;
ld.global.nc.u32 %r248, [%rd59];

{.reg .f16 low,high;
mov.b32 {low,high},%r248;
cvt.f32.f16 %f56, low;}


mul.ftz.f32 %f99, %f98, %f56;
cvt.rn.f32.s32 %f100, %r243;
fma.rn.ftz.f32 %f101, %f99, %f100, %f97;
add.ftz.f32 %f207, %f101, %f207;

dp4a.s32.s32 %r249, %r250, %r215, %r270;


dp4a.s32.s32 %r253, %r254, %r219, %r249;


dp4a.s32.s32 %r257, %r258, %r223, %r253;


dp4a.s32.s32 %r261, %r262, %r227, %r257;


{cvt.rn.f16x2.e4m3x2 %r265, %rs5;}


cvt.u16.u32 %rs22, %r265;

{cvt.f32.f16 %f57, %rs22;}


div.approx.ftz.f32 %f102, %f57, %f61;

{.reg .f16 low,high;
mov.b32 {low,high},%r230;
cvt.f32.f16 %f58, low;}


mul.ftz.f32 %f103, %f102, %f58;
cvt.rn.f32.s32 %f104, %r261;
fma.rn.ftz.f32 %f105, %f103, %f104, 0f00000000;

dp4a.s32.s32 %r267, %r268, %r233, %r270;


dp4a.s32.s32 %r271, %r272, %r237, %r267;


dp4a.s32.s32 %r275, %r276, %r241, %r271;


dp4a.s32.s32 %r279, %r280, %r245, %r275;


{cvt.rn.f16x2.e4m3x2 %r283, %rs7;}


cvt.u16.u32 %rs24, %r283;

{cvt.f32.f16 %f59, %rs24;}


div.approx.ftz.f32 %f106, %f59, %f61;

{.reg .f16 low,high;
mov.b32 {low,high},%r248;
cvt.f32.f16 %f60, low;}


mul.ftz.f32 %f107, %f106, %f60;
cvt.rn.f32.s32 %f108, %r279;
fma.rn.ftz.f32 %f109, %f107, %f108, %f105;
add.ftz.f32 %f206, %f109, %f206;
add.s32 %r528, %r528, 128;
add.s32 %r527, %r527, 128;
add.s32 %r526, %r526, 128;
add.s64 %rd72, %rd72, 2304;
add.s32 %r529, %r529, 64;
setp.lt.s32 %p8, %r529, %r6;
@%p8 bra $L__BB121_2;

$L__BB121_3:
setp.eq.s32 %p9, %r1, 0;
@%p9 bra $L__BB121_5;

mov.u32 %r411, _ZZ13mul_mat_vec_qIL9ggml_type40ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r412, %r1, 768, %r411;
shl.b32 %r413, %r2, 2;
add.s32 %r414, %r412, %r413;
st.shared.f32 [%r414+-768], %f211;
st.shared.f32 [%r414+-640], %f210;
st.shared.f32 [%r414+-512], %f209;
st.shared.f32 [%r414+-384], %f208;
st.shared.f32 [%r414+-256], %f207;
st.shared.f32 [%r414+-128], %f206;

$L__BB121_5:
setp.ne.s32 %p10, %r1, 0;
bar.sync 0;
@%p10 bra $L__BB121_18;

mov.u32 %r415, %ctaid.z;
mov.u32 %r416, %ctaid.y;
mad.lo.s32 %r417, %r416, %r28, %r4;
mad.lo.s32 %r418, %r415, %r34, %r417;
cvt.u64.u32 %rd15, %r418;
shl.b32 %r419, %r2, 2;
mov.u32 %r420, _ZZ13mul_mat_vec_qIL9ggml_type40ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r18, %r420, %r419;
mov.u32 %r421, 2;
ld.shared.f32 %f110, [%r18];
add.ftz.f32 %f111, %f110, %f211;
ld.shared.f32 %f112, [%r18+768];
add.ftz.f32 %f113, %f112, %f111;
ld.shared.f32 %f114, [%r18+1536];
add.ftz.f32 %f115, %f114, %f113;
mov.b32 %r422, %f115;
mov.u32 %r423, 31;
mov.u32 %r424, 16;
mov.u32 %r425, -1;
shfl.sync.bfly.b32 %r426|%p11, %r422, %r424, %r423, %r425;
mov.b32 %f116, %r426;
add.ftz.f32 %f117, %f115, %f116;
mov.b32 %r427, %f117;
mov.u32 %r428, 8;
shfl.sync.bfly.b32 %r429|%p12, %r427, %r428, %r423, %r425;
mov.b32 %f118, %r429;
add.ftz.f32 %f119, %f117, %f118;
mov.b32 %r430, %f119;
mov.u32 %r431, 4;
shfl.sync.bfly.b32 %r432|%p13, %r430, %r431, %r423, %r425;
mov.b32 %f120, %r432;
add.ftz.f32 %f121, %f119, %f120;
mov.b32 %r433, %f121;
shfl.sync.bfly.b32 %r434|%p14, %r433, %r421, %r423, %r425;
mov.b32 %f122, %r434;
add.ftz.f32 %f123, %f121, %f122;
mov.b32 %r435, %f123;
mov.u32 %r436, 1;
shfl.sync.bfly.b32 %r437|%p15, %r435, %r436, %r423, %r425;
mov.b32 %f124, %r437;
add.ftz.f32 %f19, %f123, %f124;
setp.lt.u32 %p16, %r4, %r22;
setp.eq.s32 %p17, %r2, 0;
and.pred %p1, %p17, %p16;
mul.wide.u32 %rd60, %r418, 4;
add.s64 %rd16, %rd1, %rd60;
not.pred %p18, %p1;
@%p18 bra $L__BB121_8;

st.global.f32 [%rd16], %f19;

$L__BB121_8:
ld.shared.f32 %f125, [%r18+128];
add.ftz.f32 %f126, %f125, %f210;
ld.shared.f32 %f127, [%r18+896];
add.ftz.f32 %f128, %f127, %f126;
ld.shared.f32 %f129, [%r18+1664];
add.ftz.f32 %f130, %f129, %f128;
mov.b32 %r438, %f130;
shfl.sync.bfly.b32 %r442|%p19, %r438, %r424, %r423, %r425;
mov.b32 %f131, %r442;
add.ftz.f32 %f132, %f130, %f131;
mov.b32 %r443, %f132;
shfl.sync.bfly.b32 %r445|%p20, %r443, %r428, %r423, %r425;
mov.b32 %f133, %r445;
add.ftz.f32 %f134, %f132, %f133;
mov.b32 %r446, %f134;
shfl.sync.bfly.b32 %r448|%p21, %r446, %r431, %r423, %r425;
mov.b32 %f135, %r448;
add.ftz.f32 %f136, %f134, %f135;
mov.b32 %r449, %f136;
shfl.sync.bfly.b32 %r451|%p22, %r449, %r421, %r423, %r425;
mov.b32 %f137, %r451;
add.ftz.f32 %f138, %f136, %f137;
mov.b32 %r452, %f138;
shfl.sync.bfly.b32 %r454|%p23, %r452, %r436, %r423, %r425;
mov.b32 %f139, %r454;
add.ftz.f32 %f20, %f138, %f139;
add.s32 %r455, %r4, 1;
setp.lt.u32 %p24, %r455, %r22;
setp.eq.s32 %p25, %r2, 1;
and.pred %p2, %p25, %p24;
not.pred %p26, %p2;
@%p26 bra $L__BB121_10;

st.global.f32 [%rd16+4], %f20;

$L__BB121_10:
ld.shared.f32 %f140, [%r18+256];
add.ftz.f32 %f141, %f140, %f209;
ld.shared.f32 %f142, [%r18+1024];
add.ftz.f32 %f143, %f142, %f141;
ld.shared.f32 %f144, [%r18+1792];
add.ftz.f32 %f145, %f144, %f143;
mov.b32 %r456, %f145;
mov.u32 %r457, 31;
mov.u32 %r458, 16;
mov.u32 %r459, -1;
shfl.sync.bfly.b32 %r460|%p27, %r456, %r458, %r457, %r459;
mov.b32 %f146, %r460;
add.ftz.f32 %f147, %f145, %f146;
mov.b32 %r461, %f147;
mov.u32 %r462, 8;
shfl.sync.bfly.b32 %r463|%p28, %r461, %r462, %r457, %r459;
mov.b32 %f148, %r463;
add.ftz.f32 %f149, %f147, %f148;
mov.b32 %r464, %f149;
mov.u32 %r465, 4;
shfl.sync.bfly.b32 %r466|%p29, %r464, %r465, %r457, %r459;
mov.b32 %f150, %r466;
add.ftz.f32 %f151, %f149, %f150;
mov.b32 %r467, %f151;
mov.u32 %r468, 2;
shfl.sync.bfly.b32 %r469|%p30, %r467, %r468, %r457, %r459;
mov.b32 %f152, %r469;
add.ftz.f32 %f153, %f151, %f152;
mov.b32 %r470, %f153;
mov.u32 %r471, 1;
shfl.sync.bfly.b32 %r472|%p31, %r470, %r471, %r457, %r459;
mov.b32 %f154, %r472;
add.ftz.f32 %f21, %f153, %f154;
@%p18 bra $L__BB121_12;

cvt.u64.u32 %rd61, %r22;
add.s64 %rd62, %rd61, %rd15;
shl.b64 %rd63, %rd62, 2;
add.s64 %rd64, %rd1, %rd63;
st.global.f32 [%rd64], %f21;

$L__BB121_12:
ld.shared.f32 %f155, [%r18+384];
add.ftz.f32 %f156, %f155, %f208;
ld.shared.f32 %f157, [%r18+1152];
add.ftz.f32 %f158, %f157, %f156;
ld.shared.f32 %f159, [%r18+1920];
add.ftz.f32 %f160, %f159, %f158;
mov.b32 %r473, %f160;
shfl.sync.bfly.b32 %r477|%p33, %r473, %r458, %r457, %r459;
mov.b32 %f161, %r477;
add.ftz.f32 %f162, %f160, %f161;
mov.b32 %r478, %f162;
shfl.sync.bfly.b32 %r480|%p34, %r478, %r462, %r457, %r459;
mov.b32 %f163, %r480;
add.ftz.f32 %f164, %f162, %f163;
mov.b32 %r481, %f164;
shfl.sync.bfly.b32 %r483|%p35, %r481, %r465, %r457, %r459;
mov.b32 %f165, %r483;
add.ftz.f32 %f166, %f164, %f165;
mov.b32 %r484, %f166;
shfl.sync.bfly.b32 %r486|%p36, %r484, %r468, %r457, %r459;
mov.b32 %f167, %r486;
add.ftz.f32 %f168, %f166, %f167;
mov.b32 %r487, %f168;
shfl.sync.bfly.b32 %r489|%p37, %r487, %r471, %r457, %r459;
mov.b32 %f169, %r489;
add.ftz.f32 %f22, %f168, %f169;
@%p26 bra $L__BB121_14;

add.s32 %r490, %r22, 1;
cvt.u64.u32 %rd65, %r490;
add.s64 %rd66, %rd65, %rd15;
shl.b64 %rd67, %rd66, 2;
add.s64 %rd68, %rd1, %rd67;
st.global.f32 [%rd68], %f22;

$L__BB121_14:
shl.b32 %r491, %r22, 1;
mov.u32 %r492, 1;
ld.shared.f32 %f170, [%r18+512];
add.ftz.f32 %f171, %f170, %f207;
ld.shared.f32 %f172, [%r18+1280];
add.ftz.f32 %f173, %f172, %f171;
ld.shared.f32 %f174, [%r18+2048];
add.ftz.f32 %f175, %f174, %f173;
mov.b32 %r493, %f175;
mov.u32 %r494, 31;
mov.u32 %r495, 16;
mov.u32 %r496, -1;
shfl.sync.bfly.b32 %r497|%p39, %r493, %r495, %r494, %r496;
mov.b32 %f176, %r497;
add.ftz.f32 %f177, %f175, %f176;
mov.b32 %r498, %f177;
mov.u32 %r499, 8;
shfl.sync.bfly.b32 %r500|%p40, %r498, %r499, %r494, %r496;
mov.b32 %f178, %r500;
add.ftz.f32 %f179, %f177, %f178;
mov.b32 %r501, %f179;
mov.u32 %r502, 4;
shfl.sync.bfly.b32 %r503|%p41, %r501, %r502, %r494, %r496;
mov.b32 %f180, %r503;
add.ftz.f32 %f181, %f179, %f180;
mov.b32 %r504, %f181;
mov.u32 %r505, 2;
shfl.sync.bfly.b32 %r506|%p42, %r504, %r505, %r494, %r496;
mov.b32 %f182, %r506;
add.ftz.f32 %f183, %f181, %f182;
mov.b32 %r507, %f183;
shfl.sync.bfly.b32 %r508|%p43, %r507, %r492, %r494, %r496;
mov.b32 %f184, %r508;
add.ftz.f32 %f23, %f183, %f184;
cvt.u64.u32 %rd69, %r491;
add.s64 %rd70, %rd69, %rd15;
shl.b64 %rd71, %rd70, 2;
add.s64 %rd17, %rd1, %rd71;
@%p18 bra $L__BB121_16;

st.global.f32 [%rd17], %f23;

$L__BB121_16:
ld.shared.f32 %f185, [%r18+640];
add.ftz.f32 %f186, %f185, %f206;
ld.shared.f32 %f187, [%r18+1408];
add.ftz.f32 %f188, %f187, %f186;
ld.shared.f32 %f189, [%r18+2176];
add.ftz.f32 %f190, %f189, %f188;
mov.b32 %r509, %f190;
shfl.sync.bfly.b32 %r513|%p45, %r509, %r495, %r494, %r496;
mov.b32 %f191, %r513;
add.ftz.f32 %f192, %f190, %f191;
mov.b32 %r514, %f192;
shfl.sync.bfly.b32 %r516|%p46, %r514, %r499, %r494, %r496;
mov.b32 %f193, %r516;
add.ftz.f32 %f194, %f192, %f193;
mov.b32 %r517, %f194;
shfl.sync.bfly.b32 %r519|%p47, %r517, %r502, %r494, %r496;
mov.b32 %f195, %r519;
add.ftz.f32 %f196, %f194, %f195;
mov.b32 %r520, %f196;
shfl.sync.bfly.b32 %r522|%p48, %r520, %r505, %r494, %r496;
mov.b32 %f197, %r522;
add.ftz.f32 %f198, %f196, %f197;
mov.b32 %r523, %f198;
shfl.sync.bfly.b32 %r525|%p49, %r523, %r492, %r494, %r496;
mov.b32 %f199, %r525;
add.ftz.f32 %f24, %f198, %f199;
@%p26 bra $L__BB121_18;

st.global.f32 [%rd17+4], %f24;

$L__BB121_18:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type40ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type40ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type40ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type40ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type40ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type40ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type40ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type40ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type40ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<63>;
.reg .b16 %rs<45>;
.reg .f32 %f<282>;
.reg .b32 %r<643>;
.reg .b64 %rd<89>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type40ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[3072];

ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type40ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type40ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type40ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type40ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type40ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type40ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type40ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type40ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type40ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type40ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type40ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type40ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type40ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type40ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type40ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type40ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type40ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
cvta.to.global.u64 %rd1, %rd18;
mov.u32 %r1, %tid.y;
shl.b32 %r42, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r42, %r2;
mov.u32 %r43, %ctaid.x;
shl.b32 %r4, %r43, 1;
mov.u32 %r5, %ctaid.y;
mov.u32 %r6, %ctaid.z;
shr.u32 %r44, %r3, 31;
add.s32 %r45, %r3, %r44;
shr.s32 %r642, %r45, 1;
shr.u32 %r46, %r26, 6;
setp.ge.s32 %p3, %r642, %r46;
mov.f32 %f274, 0f00000000;
mov.f32 %f275, %f274;
mov.f32 %f276, %f274;
mov.f32 %f277, %f274;
mov.f32 %f278, %f274;
mov.f32 %f279, %f274;
mov.f32 %f280, %f274;
mov.f32 %f281, %f274;
@%p3 bra $L__BB122_3;

mul.lo.s32 %r47, %r5, %r34;
cvt.u64.u32 %rd19, %r47;
mul.lo.s32 %r48, %r6, %r40;
cvt.u64.u32 %rd20, %r48;
add.s64 %rd2, %rd20, %rd19;
mul.hi.u32 %r49, %r6, %r36;
add.s32 %r50, %r49, %r6;
shr.u32 %r51, %r50, %r37;
mul.lo.s32 %r52, %r51, %r39;
mul.hi.u32 %r53, %r5, %r30;
add.s32 %r54, %r53, %r5;
shr.u32 %r55, %r54, %r31;
mul.lo.s32 %r56, %r55, %r33;
cvta.to.global.u64 %rd88, %rd16;
and.b32 %r59, %r45, -2;
sub.s32 %r8, %r3, %r59;
shl.b32 %r60, %r8, 2;
shr.s32 %r61, %r60, 1;
shl.b32 %r62, %r8, 1;
cvt.s64.s32 %rd21, %r62;
or.b32 %r63, %r61, 1;
cvt.s64.s32 %rd22, %r63;
shl.b32 %r64, %r28, 1;
shl.b32 %r638, %r642, 1;
add.s32 %r641, %r638, %r64;
mad.lo.s32 %r640, %r28, 3, %r638;
add.s32 %r639, %r28, %r638;
add.s32 %r65, %r642, %r52;
add.s32 %r66, %r65, %r56;
add.s32 %r67, %r4, 1;
mad.lo.s32 %r68, %r27, %r67, %r66;
mul.wide.s32 %rd23, %r68, 36;
add.s64 %rd4, %rd23, %rd22;
add.s64 %rd5, %rd23, %rd21;
mad.lo.s32 %r69, %r4, %r27, %r642;
add.s32 %r70, %r69, %r52;
add.s32 %r71, %r70, %r56;
mul.wide.s32 %rd24, %r71, 36;
mul.wide.s32 %rd25, %r60, 4;
add.s64 %rd26, %rd24, %rd25;
add.s64 %rd6, %rd26, 16;
add.s64 %rd7, %rd24, %rd21;
add.s64 %rd27, %rd23, %rd25;
add.s64 %rd8, %rd27, 8;
add.s64 %rd9, %rd24, %rd22;
cvta.to.global.u64 %rd10, %rd17;
mov.f32 %f274, 0f00000000;

$L__BB122_2:
cvt.u64.u32 %rd28, %r638;
add.s64 %rd29, %rd2, %rd28;
add.s64 %rd30, %rd88, %rd6;
ld.global.nc.u32 %r360, [%rd30+-12];
and.b32 %r361, %r360, -2004318072;
shr.u32 %r362, %r361, 1;
or.b32 %r363, %r362, 839922192;
and.b32 %r364, %r360, 30583;
mov.u32 %r365, 201852420;
mov.u32 %r366, 50462976;
prmt.b32 %r367, %r366, %r365, %r364;
mov.u32 %r368, -185009412;
mov.u32 %r369, -33620224;
prmt.b32 %r370, %r369, %r368, %r364;
and.b32 %r371, %r363, 30292;
prmt.b32 %r372, %r367, %r370, %r371;
shr.u32 %r373, %r360, 16;
and.b32 %r374, %r373, 30583;
prmt.b32 %r375, %r366, %r365, %r374;
prmt.b32 %r376, %r369, %r368, %r374;
shr.u32 %r377, %r363, 16;
prmt.b32 %r378, %r375, %r376, %r377;
mov.u32 %r379, 25632;
prmt.b32 %r289, %r372, %r378, %r379;
mov.u32 %r380, 30001;
prmt.b32 %r293, %r372, %r378, %r380;
ld.global.nc.u32 %r381, [%rd30+-8];
and.b32 %r382, %r381, -2004318072;
shr.u32 %r383, %r382, 1;
or.b32 %r384, %r383, 839922192;
and.b32 %r385, %r381, 30583;
prmt.b32 %r386, %r366, %r365, %r385;
prmt.b32 %r387, %r369, %r368, %r385;
and.b32 %r388, %r384, 30292;
prmt.b32 %r389, %r386, %r387, %r388;
shr.u32 %r390, %r381, 16;
and.b32 %r391, %r390, 30583;
prmt.b32 %r392, %r366, %r365, %r391;
prmt.b32 %r393, %r369, %r368, %r391;
shr.u32 %r394, %r384, 16;
prmt.b32 %r395, %r392, %r393, %r394;
prmt.b32 %r297, %r389, %r395, %r379;
prmt.b32 %r301, %r389, %r395, %r380;
cvt.s64.s32 %rd31, %r8;
add.s64 %rd32, %rd29, %rd31;
mul.lo.s64 %rd33, %rd32, 36;
add.s64 %rd34, %rd10, %rd33;
ld.global.nc.u32 %r74, [%rd34+4];
mov.u32 %r345, 0;

dp4a.s32.s32 %r72, %r289, %r74, %r345;

ld.global.nc.u32 %r78, [%rd34+12];

dp4a.s32.s32 %r76, %r293, %r78, %r72;

ld.global.nc.u32 %r82, [%rd34+8];

dp4a.s32.s32 %r80, %r297, %r82, %r76;

ld.global.nc.u32 %r86, [%rd34+16];

dp4a.s32.s32 %r84, %r301, %r86, %r80;

add.s64 %rd35, %rd88, %rd7;
ld.global.nc.u8 %rs33, [%rd35];
and.b16 %rs34, %rs33, 127;
setp.eq.s16 %p4, %rs34, 127;
selp.b16 %rs35, 0, %rs33, %p4;
and.b16 %rs1, %rs35, 255;

{cvt.rn.f16x2.e4m3x2 %r88, %rs1;}


cvt.u16.u32 %rs2, %r88;

{cvt.f32.f16 %f49, %rs2;}


mov.f32 %f81, 0f40000000;
div.approx.ftz.f32 %f82, %f49, %f81;
ld.global.nc.u32 %r89, [%rd34];

{.reg .f16 low,high;
mov.b32 {low,high},%r89;
cvt.f32.f16 %f50, low;}


mul.ftz.f32 %f83, %f82, %f50;
cvt.rn.f32.s32 %f84, %r84;
fma.rn.ftz.f32 %f85, %f83, %f84, 0f00000000;
ld.global.nc.u32 %r396, [%rd30+-4];
and.b32 %r397, %r396, -2004318072;
shr.u32 %r398, %r397, 1;
or.b32 %r399, %r398, 839922192;
and.b32 %r400, %r396, 30583;
prmt.b32 %r401, %r366, %r365, %r400;
prmt.b32 %r402, %r369, %r368, %r400;
and.b32 %r403, %r399, 30292;
prmt.b32 %r404, %r401, %r402, %r403;
shr.u32 %r405, %r396, 16;
and.b32 %r406, %r405, 30583;
prmt.b32 %r407, %r366, %r365, %r406;
prmt.b32 %r408, %r369, %r368, %r406;
shr.u32 %r409, %r399, 16;
prmt.b32 %r410, %r407, %r408, %r409;
prmt.b32 %r307, %r404, %r410, %r379;
prmt.b32 %r311, %r404, %r410, %r380;
ld.global.nc.u32 %r411, [%rd30];
and.b32 %r412, %r411, -2004318072;
shr.u32 %r413, %r412, 1;
or.b32 %r414, %r413, 839922192;
and.b32 %r415, %r411, 30583;
prmt.b32 %r416, %r366, %r365, %r415;
prmt.b32 %r417, %r369, %r368, %r415;
and.b32 %r418, %r414, 30292;
prmt.b32 %r419, %r416, %r417, %r418;
shr.u32 %r420, %r411, 16;
and.b32 %r421, %r420, 30583;
prmt.b32 %r422, %r366, %r365, %r421;
prmt.b32 %r423, %r369, %r368, %r421;
shr.u32 %r424, %r414, 16;
prmt.b32 %r425, %r422, %r423, %r424;
prmt.b32 %r315, %r419, %r425, %r379;
prmt.b32 %r319, %r419, %r425, %r380;
shr.s32 %r427, %r60, 2;
cvt.s64.s32 %rd36, %r427;
add.s64 %rd37, %rd29, %rd36;
mul.lo.s64 %rd38, %rd37, 36;
add.s64 %rd39, %rd10, %rd38;
ld.global.nc.u32 %r92, [%rd39+20];

dp4a.s32.s32 %r90, %r307, %r92, %r345;

ld.global.nc.u32 %r96, [%rd39+28];

dp4a.s32.s32 %r94, %r311, %r96, %r90;

ld.global.nc.u32 %r100, [%rd39+24];

dp4a.s32.s32 %r98, %r315, %r100, %r94;

ld.global.nc.u32 %r104, [%rd39+32];

dp4a.s32.s32 %r102, %r319, %r104, %r98;

add.s64 %rd40, %rd88, %rd9;
ld.global.nc.u8 %rs36, [%rd40];
and.b16 %rs37, %rs36, 127;
setp.eq.s16 %p5, %rs37, 127;
selp.b16 %rs38, 0, %rs36, %p5;
and.b16 %rs3, %rs38, 255;

{cvt.rn.f16x2.e4m3x2 %r106, %rs3;}


cvt.u16.u32 %rs4, %r106;

{cvt.f32.f16 %f51, %rs4;}


div.approx.ftz.f32 %f86, %f51, %f81;
ld.global.nc.u32 %r107, [%rd39];

{.reg .f16 low,high;
mov.b32 {low,high},%r107;
cvt.f32.f16 %f52, low;}


mul.ftz.f32 %f87, %f86, %f52;
cvt.rn.f32.s32 %f88, %r102;
fma.rn.ftz.f32 %f89, %f87, %f88, %f85;
add.ftz.f32 %f281, %f89, %f281;
add.s64 %rd41, %rd88, %rd8;
ld.global.nc.u32 %r428, [%rd41+-4];
and.b32 %r429, %r428, -2004318072;
shr.u32 %r430, %r429, 1;
or.b32 %r431, %r430, 839922192;
and.b32 %r432, %r428, 30583;
prmt.b32 %r433, %r366, %r365, %r432;
prmt.b32 %r434, %r369, %r368, %r432;
and.b32 %r435, %r431, 30292;
prmt.b32 %r436, %r433, %r434, %r435;
shr.u32 %r437, %r428, 16;
and.b32 %r438, %r437, 30583;
prmt.b32 %r439, %r366, %r365, %r438;
prmt.b32 %r440, %r369, %r368, %r438;
shr.u32 %r441, %r431, 16;
prmt.b32 %r442, %r439, %r440, %r441;
prmt.b32 %r325, %r436, %r442, %r379;
prmt.b32 %r329, %r436, %r442, %r380;
ld.global.nc.u32 %r443, [%rd41];
and.b32 %r444, %r443, -2004318072;
shr.u32 %r445, %r444, 1;
or.b32 %r446, %r445, 839922192;
and.b32 %r447, %r443, 30583;
prmt.b32 %r448, %r366, %r365, %r447;
prmt.b32 %r449, %r369, %r368, %r447;
and.b32 %r450, %r446, 30292;
prmt.b32 %r451, %r448, %r449, %r450;
shr.u32 %r452, %r443, 16;
and.b32 %r453, %r452, 30583;
prmt.b32 %r454, %r366, %r365, %r453;
prmt.b32 %r455, %r369, %r368, %r453;
shr.u32 %r456, %r446, 16;
prmt.b32 %r457, %r454, %r455, %r456;
prmt.b32 %r333, %r451, %r457, %r379;
prmt.b32 %r337, %r451, %r457, %r380;

dp4a.s32.s32 %r108, %r325, %r74, %r345;


dp4a.s32.s32 %r112, %r329, %r78, %r108;


dp4a.s32.s32 %r116, %r333, %r82, %r112;


dp4a.s32.s32 %r120, %r337, %r86, %r116;

add.s64 %rd42, %rd88, %rd5;
ld.global.nc.u8 %rs39, [%rd42];
and.b16 %rs40, %rs39, 127;
setp.eq.s16 %p6, %rs40, 127;
selp.b16 %rs41, 0, %rs39, %p6;
and.b16 %rs5, %rs41, 255;

{cvt.rn.f16x2.e4m3x2 %r124, %rs5;}


cvt.u16.u32 %rs6, %r124;

{cvt.f32.f16 %f53, %rs6;}


div.approx.ftz.f32 %f90, %f53, %f81;

{.reg .f16 low,high;
mov.b32 {low,high},%r89;
cvt.f32.f16 %f54, low;}


mul.ftz.f32 %f91, %f90, %f54;
cvt.rn.f32.s32 %f92, %r120;
fma.rn.ftz.f32 %f93, %f91, %f92, 0f00000000;
ld.global.nc.u32 %r458, [%rd41+4];
and.b32 %r459, %r458, -2004318072;
shr.u32 %r460, %r459, 1;
or.b32 %r461, %r460, 839922192;
and.b32 %r462, %r458, 30583;
prmt.b32 %r463, %r366, %r365, %r462;
prmt.b32 %r464, %r369, %r368, %r462;
and.b32 %r465, %r461, 30292;
prmt.b32 %r466, %r463, %r464, %r465;
shr.u32 %r467, %r458, 16;
and.b32 %r468, %r467, 30583;
prmt.b32 %r469, %r366, %r365, %r468;
prmt.b32 %r470, %r369, %r368, %r468;
shr.u32 %r471, %r461, 16;
prmt.b32 %r472, %r469, %r470, %r471;
prmt.b32 %r343, %r466, %r472, %r379;
prmt.b32 %r347, %r466, %r472, %r380;
ld.global.nc.u32 %r473, [%rd41+8];
and.b32 %r474, %r473, -2004318072;
shr.u32 %r475, %r474, 1;
or.b32 %r476, %r475, 839922192;
and.b32 %r477, %r473, 30583;
prmt.b32 %r478, %r366, %r365, %r477;
prmt.b32 %r479, %r369, %r368, %r477;
and.b32 %r480, %r476, 30292;
prmt.b32 %r481, %r478, %r479, %r480;
shr.u32 %r482, %r473, 16;
and.b32 %r483, %r482, 30583;
prmt.b32 %r484, %r366, %r365, %r483;
prmt.b32 %r485, %r369, %r368, %r483;
shr.u32 %r486, %r476, 16;
prmt.b32 %r487, %r484, %r485, %r486;
prmt.b32 %r351, %r481, %r487, %r379;
prmt.b32 %r355, %r481, %r487, %r380;

dp4a.s32.s32 %r126, %r343, %r92, %r345;


dp4a.s32.s32 %r130, %r347, %r96, %r126;


dp4a.s32.s32 %r134, %r351, %r100, %r130;


dp4a.s32.s32 %r138, %r355, %r104, %r134;

add.s64 %rd43, %rd88, %rd4;
ld.global.nc.u8 %rs42, [%rd43];
and.b16 %rs43, %rs42, 127;
setp.eq.s16 %p7, %rs43, 127;
selp.b16 %rs44, 0, %rs42, %p7;
and.b16 %rs7, %rs44, 255;

{cvt.rn.f16x2.e4m3x2 %r142, %rs7;}


cvt.u16.u32 %rs8, %r142;

{cvt.f32.f16 %f55, %rs8;}


div.approx.ftz.f32 %f94, %f55, %f81;

{.reg .f16 low,high;
mov.b32 {low,high},%r107;
cvt.f32.f16 %f56, low;}


mul.ftz.f32 %f95, %f94, %f56;
cvt.rn.f32.s32 %f96, %r138;
fma.rn.ftz.f32 %f97, %f95, %f96, %f93;
add.ftz.f32 %f280, %f97, %f280;
cvt.u64.u32 %rd44, %r639;
add.s64 %rd45, %rd2, %rd44;
add.s64 %rd46, %rd45, %rd31;
mul.lo.s64 %rd47, %rd46, 36;
add.s64 %rd48, %rd10, %rd47;
ld.global.nc.u32 %r146, [%rd48+4];

dp4a.s32.s32 %r144, %r289, %r146, %r345;

ld.global.nc.u32 %r150, [%rd48+12];

dp4a.s32.s32 %r148, %r293, %r150, %r144;

ld.global.nc.u32 %r154, [%rd48+8];

dp4a.s32.s32 %r152, %r297, %r154, %r148;

ld.global.nc.u32 %r158, [%rd48+16];

dp4a.s32.s32 %r156, %r301, %r158, %r152;


{cvt.rn.f16x2.e4m3x2 %r160, %rs1;}


cvt.u16.u32 %rs10, %r160;

{cvt.f32.f16 %f57, %rs10;}


div.approx.ftz.f32 %f98, %f57, %f81;
ld.global.nc.u32 %r161, [%rd48];

{.reg .f16 low,high;
mov.b32 {low,high},%r161;
cvt.f32.f16 %f58, low;}


mul.ftz.f32 %f99, %f98, %f58;
cvt.rn.f32.s32 %f100, %r156;
fma.rn.ftz.f32 %f101, %f99, %f100, 0f00000000;
add.s64 %rd49, %rd45, %rd36;
mul.lo.s64 %rd50, %rd49, 36;
add.s64 %rd51, %rd10, %rd50;
ld.global.nc.u32 %r164, [%rd51+20];

dp4a.s32.s32 %r162, %r307, %r164, %r345;

ld.global.nc.u32 %r168, [%rd51+28];

dp4a.s32.s32 %r166, %r311, %r168, %r162;

ld.global.nc.u32 %r172, [%rd51+24];

dp4a.s32.s32 %r170, %r315, %r172, %r166;

ld.global.nc.u32 %r176, [%rd51+32];

dp4a.s32.s32 %r174, %r319, %r176, %r170;


{cvt.rn.f16x2.e4m3x2 %r178, %rs3;}


cvt.u16.u32 %rs12, %r178;

{cvt.f32.f16 %f59, %rs12;}


div.approx.ftz.f32 %f102, %f59, %f81;
ld.global.nc.u32 %r179, [%rd51];

{.reg .f16 low,high;
mov.b32 {low,high},%r179;
cvt.f32.f16 %f60, low;}


mul.ftz.f32 %f103, %f102, %f60;
cvt.rn.f32.s32 %f104, %r174;
fma.rn.ftz.f32 %f105, %f103, %f104, %f101;
add.ftz.f32 %f279, %f105, %f279;

dp4a.s32.s32 %r180, %r325, %r146, %r345;


dp4a.s32.s32 %r184, %r329, %r150, %r180;


dp4a.s32.s32 %r188, %r333, %r154, %r184;


dp4a.s32.s32 %r192, %r337, %r158, %r188;


{cvt.rn.f16x2.e4m3x2 %r196, %rs5;}


cvt.u16.u32 %rs14, %r196;

{cvt.f32.f16 %f61, %rs14;}


div.approx.ftz.f32 %f106, %f61, %f81;

{.reg .f16 low,high;
mov.b32 {low,high},%r161;
cvt.f32.f16 %f62, low;}


mul.ftz.f32 %f107, %f106, %f62;
cvt.rn.f32.s32 %f108, %r192;
fma.rn.ftz.f32 %f109, %f107, %f108, 0f00000000;

dp4a.s32.s32 %r198, %r343, %r164, %r345;


dp4a.s32.s32 %r202, %r347, %r168, %r198;


dp4a.s32.s32 %r206, %r351, %r172, %r202;


dp4a.s32.s32 %r210, %r355, %r176, %r206;


{cvt.rn.f16x2.e4m3x2 %r214, %rs7;}


cvt.u16.u32 %rs16, %r214;

{cvt.f32.f16 %f63, %rs16;}


div.approx.ftz.f32 %f110, %f63, %f81;

{.reg .f16 low,high;
mov.b32 {low,high},%r179;
cvt.f32.f16 %f64, low;}


mul.ftz.f32 %f111, %f110, %f64;
cvt.rn.f32.s32 %f112, %r210;
fma.rn.ftz.f32 %f113, %f111, %f112, %f109;
add.ftz.f32 %f278, %f113, %f278;
cvt.u64.u32 %rd52, %r641;
add.s64 %rd53, %rd2, %rd52;
add.s64 %rd54, %rd53, %rd31;
mul.lo.s64 %rd55, %rd54, 36;
add.s64 %rd56, %rd10, %rd55;
ld.global.nc.u32 %r218, [%rd56+4];

dp4a.s32.s32 %r216, %r289, %r218, %r345;

ld.global.nc.u32 %r222, [%rd56+12];

dp4a.s32.s32 %r220, %r293, %r222, %r216;

ld.global.nc.u32 %r226, [%rd56+8];

dp4a.s32.s32 %r224, %r297, %r226, %r220;

ld.global.nc.u32 %r230, [%rd56+16];

dp4a.s32.s32 %r228, %r301, %r230, %r224;


{cvt.rn.f16x2.e4m3x2 %r232, %rs1;}


cvt.u16.u32 %rs18, %r232;

{cvt.f32.f16 %f65, %rs18;}


div.approx.ftz.f32 %f114, %f65, %f81;
ld.global.nc.u32 %r233, [%rd56];

{.reg .f16 low,high;
mov.b32 {low,high},%r233;
cvt.f32.f16 %f66, low;}


mul.ftz.f32 %f115, %f114, %f66;
cvt.rn.f32.s32 %f116, %r228;
fma.rn.ftz.f32 %f117, %f115, %f116, 0f00000000;
add.s64 %rd57, %rd53, %rd36;
mul.lo.s64 %rd58, %rd57, 36;
add.s64 %rd59, %rd10, %rd58;
ld.global.nc.u32 %r236, [%rd59+20];

dp4a.s32.s32 %r234, %r307, %r236, %r345;

ld.global.nc.u32 %r240, [%rd59+28];

dp4a.s32.s32 %r238, %r311, %r240, %r234;

ld.global.nc.u32 %r244, [%rd59+24];

dp4a.s32.s32 %r242, %r315, %r244, %r238;

ld.global.nc.u32 %r248, [%rd59+32];

dp4a.s32.s32 %r246, %r319, %r248, %r242;


{cvt.rn.f16x2.e4m3x2 %r250, %rs3;}


cvt.u16.u32 %rs20, %r250;

{cvt.f32.f16 %f67, %rs20;}


div.approx.ftz.f32 %f118, %f67, %f81;
ld.global.nc.u32 %r251, [%rd59];

{.reg .f16 low,high;
mov.b32 {low,high},%r251;
cvt.f32.f16 %f68, low;}


mul.ftz.f32 %f119, %f118, %f68;
cvt.rn.f32.s32 %f120, %r246;
fma.rn.ftz.f32 %f121, %f119, %f120, %f117;
add.ftz.f32 %f277, %f121, %f277;

dp4a.s32.s32 %r252, %r325, %r218, %r345;


dp4a.s32.s32 %r256, %r329, %r222, %r252;


dp4a.s32.s32 %r260, %r333, %r226, %r256;


dp4a.s32.s32 %r264, %r337, %r230, %r260;


{cvt.rn.f16x2.e4m3x2 %r268, %rs5;}


cvt.u16.u32 %rs22, %r268;

{cvt.f32.f16 %f69, %rs22;}


div.approx.ftz.f32 %f122, %f69, %f81;

{.reg .f16 low,high;
mov.b32 {low,high},%r233;
cvt.f32.f16 %f70, low;}


mul.ftz.f32 %f123, %f122, %f70;
cvt.rn.f32.s32 %f124, %r264;
fma.rn.ftz.f32 %f125, %f123, %f124, 0f00000000;

dp4a.s32.s32 %r270, %r343, %r236, %r345;


dp4a.s32.s32 %r274, %r347, %r240, %r270;


dp4a.s32.s32 %r278, %r351, %r244, %r274;


dp4a.s32.s32 %r282, %r355, %r248, %r278;


{cvt.rn.f16x2.e4m3x2 %r286, %rs7;}


cvt.u16.u32 %rs24, %r286;

{cvt.f32.f16 %f71, %rs24;}


div.approx.ftz.f32 %f126, %f71, %f81;

{.reg .f16 low,high;
mov.b32 {low,high},%r251;
cvt.f32.f16 %f72, low;}


mul.ftz.f32 %f127, %f126, %f72;
cvt.rn.f32.s32 %f128, %r282;
fma.rn.ftz.f32 %f129, %f127, %f128, %f125;
add.ftz.f32 %f276, %f129, %f276;
cvt.u64.u32 %rd60, %r640;
add.s64 %rd61, %rd2, %rd60;
add.s64 %rd62, %rd61, %rd31;
mul.lo.s64 %rd63, %rd62, 36;
add.s64 %rd64, %rd10, %rd63;
ld.global.nc.u32 %r290, [%rd64+4];

dp4a.s32.s32 %r288, %r289, %r290, %r345;

ld.global.nc.u32 %r294, [%rd64+12];

dp4a.s32.s32 %r292, %r293, %r294, %r288;

ld.global.nc.u32 %r298, [%rd64+8];

dp4a.s32.s32 %r296, %r297, %r298, %r292;

ld.global.nc.u32 %r302, [%rd64+16];

dp4a.s32.s32 %r300, %r301, %r302, %r296;


{cvt.rn.f16x2.e4m3x2 %r304, %rs1;}


cvt.u16.u32 %rs26, %r304;

{cvt.f32.f16 %f73, %rs26;}


div.approx.ftz.f32 %f130, %f73, %f81;
ld.global.nc.u32 %r305, [%rd64];

{.reg .f16 low,high;
mov.b32 {low,high},%r305;
cvt.f32.f16 %f74, low;}


mul.ftz.f32 %f131, %f130, %f74;
cvt.rn.f32.s32 %f132, %r300;
fma.rn.ftz.f32 %f133, %f131, %f132, 0f00000000;
add.s64 %rd65, %rd61, %rd36;
mul.lo.s64 %rd66, %rd65, 36;
add.s64 %rd67, %rd10, %rd66;
ld.global.nc.u32 %r308, [%rd67+20];

dp4a.s32.s32 %r306, %r307, %r308, %r345;

ld.global.nc.u32 %r312, [%rd67+28];

dp4a.s32.s32 %r310, %r311, %r312, %r306;

ld.global.nc.u32 %r316, [%rd67+24];

dp4a.s32.s32 %r314, %r315, %r316, %r310;

ld.global.nc.u32 %r320, [%rd67+32];

dp4a.s32.s32 %r318, %r319, %r320, %r314;


{cvt.rn.f16x2.e4m3x2 %r322, %rs3;}


cvt.u16.u32 %rs28, %r322;

{cvt.f32.f16 %f75, %rs28;}


div.approx.ftz.f32 %f134, %f75, %f81;
ld.global.nc.u32 %r323, [%rd67];

{.reg .f16 low,high;
mov.b32 {low,high},%r323;
cvt.f32.f16 %f76, low;}


mul.ftz.f32 %f135, %f134, %f76;
cvt.rn.f32.s32 %f136, %r318;
fma.rn.ftz.f32 %f137, %f135, %f136, %f133;
add.ftz.f32 %f275, %f137, %f275;

dp4a.s32.s32 %r324, %r325, %r290, %r345;


dp4a.s32.s32 %r328, %r329, %r294, %r324;


dp4a.s32.s32 %r332, %r333, %r298, %r328;


dp4a.s32.s32 %r336, %r337, %r302, %r332;


{cvt.rn.f16x2.e4m3x2 %r340, %rs5;}


cvt.u16.u32 %rs30, %r340;

{cvt.f32.f16 %f77, %rs30;}


div.approx.ftz.f32 %f138, %f77, %f81;

{.reg .f16 low,high;
mov.b32 {low,high},%r305;
cvt.f32.f16 %f78, low;}


mul.ftz.f32 %f139, %f138, %f78;
cvt.rn.f32.s32 %f140, %r336;
fma.rn.ftz.f32 %f141, %f139, %f140, 0f00000000;

dp4a.s32.s32 %r342, %r343, %r308, %r345;


dp4a.s32.s32 %r346, %r347, %r312, %r342;


dp4a.s32.s32 %r350, %r351, %r316, %r346;


dp4a.s32.s32 %r354, %r355, %r320, %r350;


{cvt.rn.f16x2.e4m3x2 %r358, %rs7;}


cvt.u16.u32 %rs32, %r358;

{cvt.f32.f16 %f79, %rs32;}


div.approx.ftz.f32 %f142, %f79, %f81;

{.reg .f16 low,high;
mov.b32 {low,high},%r323;
cvt.f32.f16 %f80, low;}


mul.ftz.f32 %f143, %f142, %f80;
cvt.rn.f32.s32 %f144, %r354;
fma.rn.ftz.f32 %f145, %f143, %f144, %f141;
add.ftz.f32 %f274, %f145, %f274;
add.s32 %r641, %r641, 128;
add.s32 %r640, %r640, 128;
add.s32 %r639, %r639, 128;
add.s32 %r638, %r638, 128;
add.s64 %rd88, %rd88, 2304;
add.s32 %r642, %r642, 64;
setp.lt.s32 %p8, %r642, %r46;
@%p8 bra $L__BB122_2;

$L__BB122_3:
setp.eq.s32 %p9, %r1, 0;
@%p9 bra $L__BB122_5;

shl.b32 %r489, %r1, 10;
mov.u32 %r490, _ZZ13mul_mat_vec_qIL9ggml_type40ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r491, %r490, %r489;
shl.b32 %r492, %r2, 2;
add.s32 %r493, %r491, %r492;
st.shared.f32 [%r493+-1024], %f281;
st.shared.f32 [%r493+-896], %f280;
st.shared.f32 [%r493+-768], %f279;
st.shared.f32 [%r493+-640], %f278;
st.shared.f32 [%r493+-512], %f277;
st.shared.f32 [%r493+-384], %f276;
st.shared.f32 [%r493+-256], %f275;
st.shared.f32 [%r493+-128], %f274;

$L__BB122_5:
setp.ne.s32 %p10, %r1, 0;
bar.sync 0;
@%p10 bra $L__BB122_22;

mad.lo.s32 %r494, %r5, %r35, %r4;
mad.lo.s32 %r495, %r6, %r41, %r494;
cvt.u64.u32 %rd13, %r495;
shl.b32 %r496, %r2, 2;
mov.u32 %r497, _ZZ13mul_mat_vec_qIL9ggml_type40ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r23, %r497, %r496;
mov.u32 %r498, 2;
ld.shared.f32 %f146, [%r23];
add.ftz.f32 %f147, %f146, %f281;
ld.shared.f32 %f148, [%r23+1024];
add.ftz.f32 %f149, %f148, %f147;
ld.shared.f32 %f150, [%r23+2048];
add.ftz.f32 %f151, %f150, %f149;
mov.b32 %r499, %f151;
mov.u32 %r500, 31;
mov.u32 %r501, 16;
mov.u32 %r502, -1;
shfl.sync.bfly.b32 %r503|%p11, %r499, %r501, %r500, %r502;
mov.b32 %f152, %r503;
add.ftz.f32 %f153, %f151, %f152;
mov.b32 %r504, %f153;
mov.u32 %r505, 8;
shfl.sync.bfly.b32 %r506|%p12, %r504, %r505, %r500, %r502;
mov.b32 %f154, %r506;
add.ftz.f32 %f155, %f153, %f154;
mov.b32 %r507, %f155;
mov.u32 %r508, 4;
shfl.sync.bfly.b32 %r509|%p13, %r507, %r508, %r500, %r502;
mov.b32 %f156, %r509;
add.ftz.f32 %f157, %f155, %f156;
mov.b32 %r510, %f157;
shfl.sync.bfly.b32 %r511|%p14, %r510, %r498, %r500, %r502;
mov.b32 %f158, %r511;
add.ftz.f32 %f159, %f157, %f158;
mov.b32 %r512, %f159;
mov.u32 %r513, 1;
shfl.sync.bfly.b32 %r514|%p15, %r512, %r513, %r500, %r502;
mov.b32 %f160, %r514;
add.ftz.f32 %f25, %f159, %f160;
setp.lt.u32 %p16, %r4, %r29;
setp.eq.s32 %p17, %r2, 0;
and.pred %p1, %p17, %p16;
mul.wide.u32 %rd68, %r495, 4;
add.s64 %rd14, %rd1, %rd68;
not.pred %p18, %p1;
@%p18 bra $L__BB122_8;

st.global.f32 [%rd14], %f25;

$L__BB122_8:
ld.shared.f32 %f161, [%r23+128];
add.ftz.f32 %f162, %f161, %f280;
ld.shared.f32 %f163, [%r23+1152];
add.ftz.f32 %f164, %f163, %f162;
ld.shared.f32 %f165, [%r23+2176];
add.ftz.f32 %f166, %f165, %f164;
mov.b32 %r515, %f166;
shfl.sync.bfly.b32 %r519|%p19, %r515, %r501, %r500, %r502;
mov.b32 %f167, %r519;
add.ftz.f32 %f168, %f166, %f167;
mov.b32 %r520, %f168;
shfl.sync.bfly.b32 %r522|%p20, %r520, %r505, %r500, %r502;
mov.b32 %f169, %r522;
add.ftz.f32 %f170, %f168, %f169;
mov.b32 %r523, %f170;
shfl.sync.bfly.b32 %r525|%p21, %r523, %r508, %r500, %r502;
mov.b32 %f171, %r525;
add.ftz.f32 %f172, %f170, %f171;
mov.b32 %r526, %f172;
shfl.sync.bfly.b32 %r528|%p22, %r526, %r498, %r500, %r502;
mov.b32 %f173, %r528;
add.ftz.f32 %f174, %f172, %f173;
mov.b32 %r529, %f174;
shfl.sync.bfly.b32 %r531|%p23, %r529, %r513, %r500, %r502;
mov.b32 %f175, %r531;
add.ftz.f32 %f26, %f174, %f175;
add.s32 %r532, %r4, 1;
setp.lt.u32 %p24, %r532, %r29;
setp.eq.s32 %p25, %r2, 1;
and.pred %p2, %p25, %p24;
not.pred %p26, %p2;
@%p26 bra $L__BB122_10;

st.global.f32 [%rd14+4], %f26;

$L__BB122_10:
ld.shared.f32 %f176, [%r23+256];
add.ftz.f32 %f177, %f176, %f279;
ld.shared.f32 %f178, [%r23+1280];
add.ftz.f32 %f179, %f178, %f177;
ld.shared.f32 %f180, [%r23+2304];
add.ftz.f32 %f181, %f180, %f179;
mov.b32 %r533, %f181;
mov.u32 %r534, 31;
mov.u32 %r535, 16;
mov.u32 %r536, -1;
shfl.sync.bfly.b32 %r537|%p27, %r533, %r535, %r534, %r536;
mov.b32 %f182, %r537;
add.ftz.f32 %f183, %f181, %f182;
mov.b32 %r538, %f183;
mov.u32 %r539, 8;
shfl.sync.bfly.b32 %r540|%p28, %r538, %r539, %r534, %r536;
mov.b32 %f184, %r540;
add.ftz.f32 %f185, %f183, %f184;
mov.b32 %r541, %f185;
mov.u32 %r542, 4;
shfl.sync.bfly.b32 %r543|%p29, %r541, %r542, %r534, %r536;
mov.b32 %f186, %r543;
add.ftz.f32 %f187, %f185, %f186;
mov.b32 %r544, %f187;
mov.u32 %r545, 2;
shfl.sync.bfly.b32 %r546|%p30, %r544, %r545, %r534, %r536;
mov.b32 %f188, %r546;
add.ftz.f32 %f189, %f187, %f188;
mov.b32 %r547, %f189;
mov.u32 %r548, 1;
shfl.sync.bfly.b32 %r549|%p31, %r547, %r548, %r534, %r536;
mov.b32 %f190, %r549;
add.ftz.f32 %f27, %f189, %f190;
@%p18 bra $L__BB122_12;

cvt.u64.u32 %rd69, %r29;
add.s64 %rd70, %rd69, %rd13;
shl.b64 %rd71, %rd70, 2;
add.s64 %rd72, %rd1, %rd71;
st.global.f32 [%rd72], %f27;

$L__BB122_12:
ld.shared.f32 %f191, [%r23+384];
add.ftz.f32 %f192, %f191, %f278;
ld.shared.f32 %f193, [%r23+1408];
add.ftz.f32 %f194, %f193, %f192;
ld.shared.f32 %f195, [%r23+2432];
add.ftz.f32 %f196, %f195, %f194;
mov.b32 %r550, %f196;
shfl.sync.bfly.b32 %r554|%p33, %r550, %r535, %r534, %r536;
mov.b32 %f197, %r554;
add.ftz.f32 %f198, %f196, %f197;
mov.b32 %r555, %f198;
shfl.sync.bfly.b32 %r557|%p34, %r555, %r539, %r534, %r536;
mov.b32 %f199, %r557;
add.ftz.f32 %f200, %f198, %f199;
mov.b32 %r558, %f200;
shfl.sync.bfly.b32 %r560|%p35, %r558, %r542, %r534, %r536;
mov.b32 %f201, %r560;
add.ftz.f32 %f202, %f200, %f201;
mov.b32 %r561, %f202;
shfl.sync.bfly.b32 %r563|%p36, %r561, %r545, %r534, %r536;
mov.b32 %f203, %r563;
add.ftz.f32 %f204, %f202, %f203;
mov.b32 %r564, %f204;
shfl.sync.bfly.b32 %r566|%p37, %r564, %r548, %r534, %r536;
mov.b32 %f205, %r566;
add.ftz.f32 %f28, %f204, %f205;
@%p26 bra $L__BB122_14;

add.s32 %r567, %r29, 1;
cvt.u64.u32 %rd73, %r567;
add.s64 %rd74, %rd73, %rd13;
shl.b64 %rd75, %rd74, 2;
add.s64 %rd76, %rd1, %rd75;
st.global.f32 [%rd76], %f28;

$L__BB122_14:
shl.b32 %r568, %r29, 1;
mov.u32 %r569, 1;
ld.shared.f32 %f206, [%r23+512];
add.ftz.f32 %f207, %f206, %f277;
ld.shared.f32 %f208, [%r23+1536];
add.ftz.f32 %f209, %f208, %f207;
ld.shared.f32 %f210, [%r23+2560];
add.ftz.f32 %f211, %f210, %f209;
mov.b32 %r570, %f211;
mov.u32 %r571, 31;
mov.u32 %r572, 16;
mov.u32 %r573, -1;
shfl.sync.bfly.b32 %r574|%p39, %r570, %r572, %r571, %r573;
mov.b32 %f212, %r574;
add.ftz.f32 %f213, %f211, %f212;
mov.b32 %r575, %f213;
mov.u32 %r576, 8;
shfl.sync.bfly.b32 %r577|%p40, %r575, %r576, %r571, %r573;
mov.b32 %f214, %r577;
add.ftz.f32 %f215, %f213, %f214;
mov.b32 %r578, %f215;
mov.u32 %r579, 4;
shfl.sync.bfly.b32 %r580|%p41, %r578, %r579, %r571, %r573;
mov.b32 %f216, %r580;
add.ftz.f32 %f217, %f215, %f216;
mov.b32 %r581, %f217;
mov.u32 %r582, 2;
shfl.sync.bfly.b32 %r583|%p42, %r581, %r582, %r571, %r573;
mov.b32 %f218, %r583;
add.ftz.f32 %f219, %f217, %f218;
mov.b32 %r584, %f219;
shfl.sync.bfly.b32 %r585|%p43, %r584, %r569, %r571, %r573;
mov.b32 %f220, %r585;
add.ftz.f32 %f29, %f219, %f220;
cvt.u64.u32 %rd77, %r568;
add.s64 %rd78, %rd77, %rd13;
shl.b64 %rd79, %rd78, 2;
add.s64 %rd15, %rd1, %rd79;
@%p18 bra $L__BB122_16;

st.global.f32 [%rd15], %f29;

$L__BB122_16:
ld.shared.f32 %f221, [%r23+640];
add.ftz.f32 %f222, %f221, %f276;
ld.shared.f32 %f223, [%r23+1664];
add.ftz.f32 %f224, %f223, %f222;
ld.shared.f32 %f225, [%r23+2688];
add.ftz.f32 %f226, %f225, %f224;
mov.b32 %r586, %f226;
shfl.sync.bfly.b32 %r590|%p45, %r586, %r572, %r571, %r573;
mov.b32 %f227, %r590;
add.ftz.f32 %f228, %f226, %f227;
mov.b32 %r591, %f228;
shfl.sync.bfly.b32 %r593|%p46, %r591, %r576, %r571, %r573;
mov.b32 %f229, %r593;
add.ftz.f32 %f230, %f228, %f229;
mov.b32 %r594, %f230;
shfl.sync.bfly.b32 %r596|%p47, %r594, %r579, %r571, %r573;
mov.b32 %f231, %r596;
add.ftz.f32 %f232, %f230, %f231;
mov.b32 %r597, %f232;
shfl.sync.bfly.b32 %r599|%p48, %r597, %r582, %r571, %r573;
mov.b32 %f233, %r599;
add.ftz.f32 %f234, %f232, %f233;
mov.b32 %r600, %f234;
shfl.sync.bfly.b32 %r602|%p49, %r600, %r569, %r571, %r573;
mov.b32 %f235, %r602;
add.ftz.f32 %f30, %f234, %f235;
@%p26 bra $L__BB122_18;

st.global.f32 [%rd15+4], %f30;

$L__BB122_18:
mul.lo.s32 %r25, %r29, 3;
ld.shared.f32 %f236, [%r23+768];
add.ftz.f32 %f237, %f236, %f275;
ld.shared.f32 %f238, [%r23+1792];
add.ftz.f32 %f239, %f238, %f237;
ld.shared.f32 %f240, [%r23+2816];
add.ftz.f32 %f241, %f240, %f239;
mov.b32 %r603, %f241;
mov.u32 %r604, 31;
mov.u32 %r605, 16;
mov.u32 %r606, -1;
shfl.sync.bfly.b32 %r607|%p51, %r603, %r605, %r604, %r606;
mov.b32 %f242, %r607;
add.ftz.f32 %f243, %f241, %f242;
mov.b32 %r608, %f243;
mov.u32 %r609, 8;
shfl.sync.bfly.b32 %r610|%p52, %r608, %r609, %r604, %r606;
mov.b32 %f244, %r610;
add.ftz.f32 %f245, %f243, %f244;
mov.b32 %r611, %f245;
mov.u32 %r612, 4;
shfl.sync.bfly.b32 %r613|%p53, %r611, %r612, %r604, %r606;
mov.b32 %f246, %r613;
add.ftz.f32 %f247, %f245, %f246;
mov.b32 %r614, %f247;
mov.u32 %r615, 2;
shfl.sync.bfly.b32 %r616|%p54, %r614, %r615, %r604, %r606;
mov.b32 %f248, %r616;
add.ftz.f32 %f249, %f247, %f248;
mov.b32 %r617, %f249;
mov.u32 %r618, 1;
shfl.sync.bfly.b32 %r619|%p55, %r617, %r618, %r604, %r606;
mov.b32 %f250, %r619;
add.ftz.f32 %f31, %f249, %f250;
@%p18 bra $L__BB122_20;

cvt.u64.u32 %rd80, %r25;
add.s64 %rd81, %rd80, %rd13;
shl.b64 %rd82, %rd81, 2;
add.s64 %rd83, %rd1, %rd82;
st.global.f32 [%rd83], %f31;

$L__BB122_20:
ld.shared.f32 %f251, [%r23+896];
add.ftz.f32 %f252, %f251, %f274;
ld.shared.f32 %f253, [%r23+1920];
add.ftz.f32 %f254, %f253, %f252;
ld.shared.f32 %f255, [%r23+2944];
add.ftz.f32 %f256, %f255, %f254;
mov.b32 %r620, %f256;
shfl.sync.bfly.b32 %r624|%p57, %r620, %r605, %r604, %r606;
mov.b32 %f257, %r624;
add.ftz.f32 %f258, %f256, %f257;
mov.b32 %r625, %f258;
shfl.sync.bfly.b32 %r627|%p58, %r625, %r609, %r604, %r606;
mov.b32 %f259, %r627;
add.ftz.f32 %f260, %f258, %f259;
mov.b32 %r628, %f260;
shfl.sync.bfly.b32 %r630|%p59, %r628, %r612, %r604, %r606;
mov.b32 %f261, %r630;
add.ftz.f32 %f262, %f260, %f261;
mov.b32 %r631, %f262;
shfl.sync.bfly.b32 %r633|%p60, %r631, %r615, %r604, %r606;
mov.b32 %f263, %r633;
add.ftz.f32 %f264, %f262, %f263;
mov.b32 %r634, %f264;
shfl.sync.bfly.b32 %r636|%p61, %r634, %r618, %r604, %r606;
mov.b32 %f265, %r636;
add.ftz.f32 %f32, %f264, %f265;
@%p26 bra $L__BB122_22;

add.s32 %r637, %r25, 1;
cvt.u64.u32 %rd84, %r637;
add.s64 %rd85, %rd84, %rd13;
shl.b64 %rd86, %rd85, 2;
add.s64 %rd87, %rd1, %rd86;
st.global.f32 [%rd87], %f32;

$L__BB122_22:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type40ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type40ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type40ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type40ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type40ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type40ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type40ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type40ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type40ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<75>;
.reg .b16 %rs<53>;
.reg .f32 %f<312>;
.reg .b32 %r<754>;
.reg .b64 %rd<101>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type40ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1280];

ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type40ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd20, [_Z13mul_mat_vec_qIL9ggml_type40ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd21, [_Z13mul_mat_vec_qIL9ggml_type40ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type40ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type40ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type40ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type40ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type40ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type40ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type40ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type40ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type40ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type40ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type40ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type40ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type40ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type40ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
cvta.to.global.u64 %rd1, %rd21;
mov.u32 %r1, %tid.y;
shl.b32 %r43, %r1, 5;
mov.u32 %r44, %tid.x;
add.s32 %r2, %r43, %r44;
mov.u32 %r45, %ctaid.x;
shl.b32 %r3, %r45, 1;
mov.u32 %r4, %ctaid.y;
mov.u32 %r5, %ctaid.z;
shr.u32 %r46, %r2, 31;
add.s32 %r47, %r2, %r46;
shr.s32 %r753, %r47, 1;
shr.u32 %r7, %r42, 6;
setp.ge.s32 %p3, %r753, %r7;
mov.f32 %f302, 0f00000000;
mov.f32 %f303, %f302;
mov.f32 %f304, %f302;
mov.f32 %f305, %f302;
mov.f32 %f306, %f302;
mov.f32 %f307, %f302;
mov.f32 %f308, %f302;
mov.f32 %f309, %f302;
mov.f32 %f310, %f302;
mov.f32 %f311, %f302;
@%p3 bra $L__BB123_3;

mul.lo.s32 %r48, %r4, %r34;
cvt.u64.u32 %rd22, %r48;
mul.lo.s32 %r49, %r5, %r40;
cvt.u64.u32 %rd23, %r49;
add.s64 %rd2, %rd23, %rd22;
mul.hi.u32 %r50, %r5, %r36;
add.s32 %r51, %r50, %r5;
shr.u32 %r52, %r51, %r37;
mul.lo.s32 %r53, %r52, %r39;
mul.hi.u32 %r54, %r4, %r30;
add.s32 %r55, %r54, %r4;
shr.u32 %r56, %r55, %r31;
mul.lo.s32 %r57, %r56, %r33;
cvta.to.global.u64 %rd100, %rd19;
and.b32 %r60, %r47, -2;
sub.s32 %r61, %r2, %r60;
shl.b32 %r62, %r61, 2;
shl.b32 %r63, %r61, 1;
cvt.s64.s32 %rd4, %r61;
cvt.s64.s32 %rd24, %r63;
or.b32 %r64, %r62, 2;
shr.s32 %r65, %r64, 1;
shr.s32 %r66, %r64, 2;
cvt.s64.s32 %rd5, %r66;
cvt.s64.s32 %rd25, %r65;
shl.b32 %r67, %r28, 1;
shl.b32 %r748, %r753, 1;
add.s32 %r752, %r748, %r67;
mad.lo.s32 %r751, %r28, 3, %r748;
shl.b32 %r68, %r28, 2;
add.s32 %r750, %r748, %r68;
add.s32 %r749, %r28, %r748;
add.s32 %r69, %r753, %r53;
add.s32 %r70, %r69, %r57;
add.s32 %r71, %r3, 1;
mad.lo.s32 %r72, %r27, %r71, %r70;
mul.wide.s32 %rd26, %r72, 36;
add.s64 %rd6, %rd26, %rd25;
add.s64 %rd7, %rd26, %rd24;
mad.lo.s32 %r73, %r3, %r27, %r753;
add.s32 %r74, %r73, %r53;
add.s32 %r75, %r74, %r57;
mul.wide.s32 %rd27, %r75, 36;
mul.wide.s32 %rd28, %r62, 4;
add.s64 %rd29, %rd27, %rd28;
add.s64 %rd8, %rd29, 16;
add.s64 %rd9, %rd27, %rd24;
add.s64 %rd30, %rd26, %rd28;
add.s64 %rd10, %rd30, 8;
add.s64 %rd11, %rd27, %rd25;
cvta.to.global.u64 %rd12, %rd20;
mov.f32 %f302, 0f00000000;

$L__BB123_2:
cvt.u64.u32 %rd31, %r748;
add.s64 %rd32, %rd2, %rd31;
add.s64 %rd33, %rd100, %rd8;
ld.global.nc.u32 %r436, [%rd33+-12];
and.b32 %r437, %r436, -2004318072;
shr.u32 %r438, %r437, 1;
or.b32 %r439, %r438, 839922192;
and.b32 %r440, %r436, 30583;
mov.u32 %r441, 201852420;
mov.u32 %r442, 50462976;
prmt.b32 %r443, %r442, %r441, %r440;
mov.u32 %r444, -185009412;
mov.u32 %r445, -33620224;
prmt.b32 %r446, %r445, %r444, %r440;
and.b32 %r447, %r439, 30292;
prmt.b32 %r448, %r443, %r446, %r447;
shr.u32 %r449, %r436, 16;
and.b32 %r450, %r449, 30583;
prmt.b32 %r451, %r442, %r441, %r450;
prmt.b32 %r452, %r445, %r444, %r450;
shr.u32 %r453, %r439, 16;
prmt.b32 %r454, %r451, %r452, %r453;
mov.u32 %r455, 25632;
prmt.b32 %r365, %r448, %r454, %r455;
mov.u32 %r456, 30001;
prmt.b32 %r369, %r448, %r454, %r456;
ld.global.nc.u32 %r457, [%rd33+-8];
and.b32 %r458, %r457, -2004318072;
shr.u32 %r459, %r458, 1;
or.b32 %r460, %r459, 839922192;
and.b32 %r461, %r457, 30583;
prmt.b32 %r462, %r442, %r441, %r461;
prmt.b32 %r463, %r445, %r444, %r461;
and.b32 %r464, %r460, 30292;
prmt.b32 %r465, %r462, %r463, %r464;
shr.u32 %r466, %r457, 16;
and.b32 %r467, %r466, 30583;
prmt.b32 %r468, %r442, %r441, %r467;
prmt.b32 %r469, %r445, %r444, %r467;
shr.u32 %r470, %r460, 16;
prmt.b32 %r471, %r468, %r469, %r470;
prmt.b32 %r373, %r465, %r471, %r455;
prmt.b32 %r377, %r465, %r471, %r456;
add.s64 %rd34, %rd32, %rd4;
mul.lo.s64 %rd35, %rd34, 36;
add.s64 %rd36, %rd12, %rd35;
ld.global.nc.u32 %r78, [%rd36+4];
mov.u32 %r421, 0;

dp4a.s32.s32 %r76, %r365, %r78, %r421;

ld.global.nc.u32 %r82, [%rd36+12];

dp4a.s32.s32 %r80, %r369, %r82, %r76;

ld.global.nc.u32 %r86, [%rd36+8];

dp4a.s32.s32 %r84, %r373, %r86, %r80;

ld.global.nc.u32 %r90, [%rd36+16];

dp4a.s32.s32 %r88, %r377, %r90, %r84;

add.s64 %rd37, %rd100, %rd9;
ld.global.nc.u8 %rs41, [%rd37];
and.b16 %rs42, %rs41, 127;
setp.eq.s16 %p4, %rs42, 127;
selp.b16 %rs43, 0, %rs41, %p4;
and.b16 %rs1, %rs43, 255;

{cvt.rn.f16x2.e4m3x2 %r92, %rs1;}


cvt.u16.u32 %rs2, %r92;

{cvt.f32.f16 %f61, %rs2;}


mov.f32 %f101, 0f40000000;
div.approx.ftz.f32 %f102, %f61, %f101;
ld.global.nc.u32 %r93, [%rd36];

{.reg .f16 low,high;
mov.b32 {low,high},%r93;
cvt.f32.f16 %f62, low;}


mul.ftz.f32 %f103, %f102, %f62;
cvt.rn.f32.s32 %f104, %r88;
fma.rn.ftz.f32 %f105, %f103, %f104, 0f00000000;
ld.global.nc.u32 %r472, [%rd33+-4];
and.b32 %r473, %r472, -2004318072;
shr.u32 %r474, %r473, 1;
or.b32 %r475, %r474, 839922192;
and.b32 %r476, %r472, 30583;
prmt.b32 %r477, %r442, %r441, %r476;
prmt.b32 %r478, %r445, %r444, %r476;
and.b32 %r479, %r475, 30292;
prmt.b32 %r480, %r477, %r478, %r479;
shr.u32 %r481, %r472, 16;
and.b32 %r482, %r481, 30583;
prmt.b32 %r483, %r442, %r441, %r482;
prmt.b32 %r484, %r445, %r444, %r482;
shr.u32 %r485, %r475, 16;
prmt.b32 %r486, %r483, %r484, %r485;
prmt.b32 %r383, %r480, %r486, %r455;
prmt.b32 %r387, %r480, %r486, %r456;
ld.global.nc.u32 %r487, [%rd33];
and.b32 %r488, %r487, -2004318072;
shr.u32 %r489, %r488, 1;
or.b32 %r490, %r489, 839922192;
and.b32 %r491, %r487, 30583;
prmt.b32 %r492, %r442, %r441, %r491;
prmt.b32 %r493, %r445, %r444, %r491;
and.b32 %r494, %r490, 30292;
prmt.b32 %r495, %r492, %r493, %r494;
shr.u32 %r496, %r487, 16;
and.b32 %r497, %r496, 30583;
prmt.b32 %r498, %r442, %r441, %r497;
prmt.b32 %r499, %r445, %r444, %r497;
shr.u32 %r500, %r490, 16;
prmt.b32 %r501, %r498, %r499, %r500;
prmt.b32 %r391, %r495, %r501, %r455;
prmt.b32 %r395, %r495, %r501, %r456;
add.s64 %rd38, %rd32, %rd5;
mul.lo.s64 %rd39, %rd38, 36;
add.s64 %rd40, %rd12, %rd39;
ld.global.nc.u32 %r96, [%rd40+20];

dp4a.s32.s32 %r94, %r383, %r96, %r421;

ld.global.nc.u32 %r100, [%rd40+28];

dp4a.s32.s32 %r98, %r387, %r100, %r94;

ld.global.nc.u32 %r104, [%rd40+24];

dp4a.s32.s32 %r102, %r391, %r104, %r98;

ld.global.nc.u32 %r108, [%rd40+32];

dp4a.s32.s32 %r106, %r395, %r108, %r102;

add.s64 %rd41, %rd100, %rd11;
ld.global.nc.u8 %rs44, [%rd41];
and.b16 %rs45, %rs44, 127;
setp.eq.s16 %p5, %rs45, 127;
selp.b16 %rs46, 0, %rs44, %p5;
and.b16 %rs3, %rs46, 255;

{cvt.rn.f16x2.e4m3x2 %r110, %rs3;}


cvt.u16.u32 %rs4, %r110;

{cvt.f32.f16 %f63, %rs4;}


div.approx.ftz.f32 %f106, %f63, %f101;
ld.global.nc.u32 %r111, [%rd40];

{.reg .f16 low,high;
mov.b32 {low,high},%r111;
cvt.f32.f16 %f64, low;}


mul.ftz.f32 %f107, %f106, %f64;
cvt.rn.f32.s32 %f108, %r106;
fma.rn.ftz.f32 %f109, %f107, %f108, %f105;
add.ftz.f32 %f311, %f109, %f311;
add.s64 %rd42, %rd100, %rd10;
ld.global.nc.u32 %r502, [%rd42+-4];
and.b32 %r503, %r502, -2004318072;
shr.u32 %r504, %r503, 1;
or.b32 %r505, %r504, 839922192;
and.b32 %r506, %r502, 30583;
prmt.b32 %r507, %r442, %r441, %r506;
prmt.b32 %r508, %r445, %r444, %r506;
and.b32 %r509, %r505, 30292;
prmt.b32 %r510, %r507, %r508, %r509;
shr.u32 %r511, %r502, 16;
and.b32 %r512, %r511, 30583;
prmt.b32 %r513, %r442, %r441, %r512;
prmt.b32 %r514, %r445, %r444, %r512;
shr.u32 %r515, %r505, 16;
prmt.b32 %r516, %r513, %r514, %r515;
prmt.b32 %r401, %r510, %r516, %r455;
prmt.b32 %r405, %r510, %r516, %r456;
ld.global.nc.u32 %r517, [%rd42];
and.b32 %r518, %r517, -2004318072;
shr.u32 %r519, %r518, 1;
or.b32 %r520, %r519, 839922192;
and.b32 %r521, %r517, 30583;
prmt.b32 %r522, %r442, %r441, %r521;
prmt.b32 %r523, %r445, %r444, %r521;
and.b32 %r524, %r520, 30292;
prmt.b32 %r525, %r522, %r523, %r524;
shr.u32 %r526, %r517, 16;
and.b32 %r527, %r526, 30583;
prmt.b32 %r528, %r442, %r441, %r527;
prmt.b32 %r529, %r445, %r444, %r527;
shr.u32 %r530, %r520, 16;
prmt.b32 %r531, %r528, %r529, %r530;
prmt.b32 %r409, %r525, %r531, %r455;
prmt.b32 %r413, %r525, %r531, %r456;

dp4a.s32.s32 %r112, %r401, %r78, %r421;


dp4a.s32.s32 %r116, %r405, %r82, %r112;


dp4a.s32.s32 %r120, %r409, %r86, %r116;


dp4a.s32.s32 %r124, %r413, %r90, %r120;

add.s64 %rd43, %rd100, %rd7;
ld.global.nc.u8 %rs47, [%rd43];
and.b16 %rs48, %rs47, 127;
setp.eq.s16 %p6, %rs48, 127;
selp.b16 %rs49, 0, %rs47, %p6;
and.b16 %rs5, %rs49, 255;

{cvt.rn.f16x2.e4m3x2 %r128, %rs5;}


cvt.u16.u32 %rs6, %r128;

{cvt.f32.f16 %f65, %rs6;}


div.approx.ftz.f32 %f110, %f65, %f101;

{.reg .f16 low,high;
mov.b32 {low,high},%r93;
cvt.f32.f16 %f66, low;}


mul.ftz.f32 %f111, %f110, %f66;
cvt.rn.f32.s32 %f112, %r124;
fma.rn.ftz.f32 %f113, %f111, %f112, 0f00000000;
ld.global.nc.u32 %r532, [%rd42+4];
and.b32 %r533, %r532, -2004318072;
shr.u32 %r534, %r533, 1;
or.b32 %r535, %r534, 839922192;
and.b32 %r536, %r532, 30583;
prmt.b32 %r537, %r442, %r441, %r536;
prmt.b32 %r538, %r445, %r444, %r536;
and.b32 %r539, %r535, 30292;
prmt.b32 %r540, %r537, %r538, %r539;
shr.u32 %r541, %r532, 16;
and.b32 %r542, %r541, 30583;
prmt.b32 %r543, %r442, %r441, %r542;
prmt.b32 %r544, %r445, %r444, %r542;
shr.u32 %r545, %r535, 16;
prmt.b32 %r546, %r543, %r544, %r545;
prmt.b32 %r419, %r540, %r546, %r455;
prmt.b32 %r423, %r540, %r546, %r456;
ld.global.nc.u32 %r547, [%rd42+8];
and.b32 %r548, %r547, -2004318072;
shr.u32 %r549, %r548, 1;
or.b32 %r550, %r549, 839922192;
and.b32 %r551, %r547, 30583;
prmt.b32 %r552, %r442, %r441, %r551;
prmt.b32 %r553, %r445, %r444, %r551;
and.b32 %r554, %r550, 30292;
prmt.b32 %r555, %r552, %r553, %r554;
shr.u32 %r556, %r547, 16;
and.b32 %r557, %r556, 30583;
prmt.b32 %r558, %r442, %r441, %r557;
prmt.b32 %r559, %r445, %r444, %r557;
shr.u32 %r560, %r550, 16;
prmt.b32 %r561, %r558, %r559, %r560;
prmt.b32 %r427, %r555, %r561, %r455;
prmt.b32 %r431, %r555, %r561, %r456;

dp4a.s32.s32 %r130, %r419, %r96, %r421;


dp4a.s32.s32 %r134, %r423, %r100, %r130;


dp4a.s32.s32 %r138, %r427, %r104, %r134;


dp4a.s32.s32 %r142, %r431, %r108, %r138;

add.s64 %rd44, %rd100, %rd6;
ld.global.nc.u8 %rs50, [%rd44];
and.b16 %rs51, %rs50, 127;
setp.eq.s16 %p7, %rs51, 127;
selp.b16 %rs52, 0, %rs50, %p7;
and.b16 %rs7, %rs52, 255;

{cvt.rn.f16x2.e4m3x2 %r146, %rs7;}


cvt.u16.u32 %rs8, %r146;

{cvt.f32.f16 %f67, %rs8;}


div.approx.ftz.f32 %f114, %f67, %f101;

{.reg .f16 low,high;
mov.b32 {low,high},%r111;
cvt.f32.f16 %f68, low;}


mul.ftz.f32 %f115, %f114, %f68;
cvt.rn.f32.s32 %f116, %r142;
fma.rn.ftz.f32 %f117, %f115, %f116, %f113;
add.ftz.f32 %f310, %f117, %f310;
cvt.u64.u32 %rd45, %r749;
add.s64 %rd46, %rd2, %rd45;
add.s64 %rd47, %rd46, %rd4;
mul.lo.s64 %rd48, %rd47, 36;
add.s64 %rd49, %rd12, %rd48;
ld.global.nc.u32 %r150, [%rd49+4];

dp4a.s32.s32 %r148, %r365, %r150, %r421;

ld.global.nc.u32 %r154, [%rd49+12];

dp4a.s32.s32 %r152, %r369, %r154, %r148;

ld.global.nc.u32 %r158, [%rd49+8];

dp4a.s32.s32 %r156, %r373, %r158, %r152;

ld.global.nc.u32 %r162, [%rd49+16];

dp4a.s32.s32 %r160, %r377, %r162, %r156;


{cvt.rn.f16x2.e4m3x2 %r164, %rs1;}


cvt.u16.u32 %rs10, %r164;

{cvt.f32.f16 %f69, %rs10;}


div.approx.ftz.f32 %f118, %f69, %f101;
ld.global.nc.u32 %r165, [%rd49];

{.reg .f16 low,high;
mov.b32 {low,high},%r165;
cvt.f32.f16 %f70, low;}


mul.ftz.f32 %f119, %f118, %f70;
cvt.rn.f32.s32 %f120, %r160;
fma.rn.ftz.f32 %f121, %f119, %f120, 0f00000000;
add.s64 %rd50, %rd46, %rd5;
mul.lo.s64 %rd51, %rd50, 36;
add.s64 %rd52, %rd12, %rd51;
ld.global.nc.u32 %r168, [%rd52+20];

dp4a.s32.s32 %r166, %r383, %r168, %r421;

ld.global.nc.u32 %r172, [%rd52+28];

dp4a.s32.s32 %r170, %r387, %r172, %r166;

ld.global.nc.u32 %r176, [%rd52+24];

dp4a.s32.s32 %r174, %r391, %r176, %r170;

ld.global.nc.u32 %r180, [%rd52+32];

dp4a.s32.s32 %r178, %r395, %r180, %r174;


{cvt.rn.f16x2.e4m3x2 %r182, %rs3;}


cvt.u16.u32 %rs12, %r182;

{cvt.f32.f16 %f71, %rs12;}


div.approx.ftz.f32 %f122, %f71, %f101;
ld.global.nc.u32 %r183, [%rd52];

{.reg .f16 low,high;
mov.b32 {low,high},%r183;
cvt.f32.f16 %f72, low;}


mul.ftz.f32 %f123, %f122, %f72;
cvt.rn.f32.s32 %f124, %r178;
fma.rn.ftz.f32 %f125, %f123, %f124, %f121;
add.ftz.f32 %f309, %f125, %f309;

dp4a.s32.s32 %r184, %r401, %r150, %r421;


dp4a.s32.s32 %r188, %r405, %r154, %r184;


dp4a.s32.s32 %r192, %r409, %r158, %r188;


dp4a.s32.s32 %r196, %r413, %r162, %r192;


{cvt.rn.f16x2.e4m3x2 %r200, %rs5;}


cvt.u16.u32 %rs14, %r200;

{cvt.f32.f16 %f73, %rs14;}


div.approx.ftz.f32 %f126, %f73, %f101;

{.reg .f16 low,high;
mov.b32 {low,high},%r165;
cvt.f32.f16 %f74, low;}


mul.ftz.f32 %f127, %f126, %f74;
cvt.rn.f32.s32 %f128, %r196;
fma.rn.ftz.f32 %f129, %f127, %f128, 0f00000000;

dp4a.s32.s32 %r202, %r419, %r168, %r421;


dp4a.s32.s32 %r206, %r423, %r172, %r202;


dp4a.s32.s32 %r210, %r427, %r176, %r206;


dp4a.s32.s32 %r214, %r431, %r180, %r210;


{cvt.rn.f16x2.e4m3x2 %r218, %rs7;}


cvt.u16.u32 %rs16, %r218;

{cvt.f32.f16 %f75, %rs16;}


div.approx.ftz.f32 %f130, %f75, %f101;

{.reg .f16 low,high;
mov.b32 {low,high},%r183;
cvt.f32.f16 %f76, low;}


mul.ftz.f32 %f131, %f130, %f76;
cvt.rn.f32.s32 %f132, %r214;
fma.rn.ftz.f32 %f133, %f131, %f132, %f129;
add.ftz.f32 %f308, %f133, %f308;
cvt.u64.u32 %rd53, %r752;
add.s64 %rd54, %rd2, %rd53;
add.s64 %rd55, %rd54, %rd4;
mul.lo.s64 %rd56, %rd55, 36;
add.s64 %rd57, %rd12, %rd56;
ld.global.nc.u32 %r222, [%rd57+4];

dp4a.s32.s32 %r220, %r365, %r222, %r421;

ld.global.nc.u32 %r226, [%rd57+12];

dp4a.s32.s32 %r224, %r369, %r226, %r220;

ld.global.nc.u32 %r230, [%rd57+8];

dp4a.s32.s32 %r228, %r373, %r230, %r224;

ld.global.nc.u32 %r234, [%rd57+16];

dp4a.s32.s32 %r232, %r377, %r234, %r228;


{cvt.rn.f16x2.e4m3x2 %r236, %rs1;}


cvt.u16.u32 %rs18, %r236;

{cvt.f32.f16 %f77, %rs18;}


div.approx.ftz.f32 %f134, %f77, %f101;
ld.global.nc.u32 %r237, [%rd57];

{.reg .f16 low,high;
mov.b32 {low,high},%r237;
cvt.f32.f16 %f78, low;}


mul.ftz.f32 %f135, %f134, %f78;
cvt.rn.f32.s32 %f136, %r232;
fma.rn.ftz.f32 %f137, %f135, %f136, 0f00000000;
add.s64 %rd58, %rd54, %rd5;
mul.lo.s64 %rd59, %rd58, 36;
add.s64 %rd60, %rd12, %rd59;
ld.global.nc.u32 %r240, [%rd60+20];

dp4a.s32.s32 %r238, %r383, %r240, %r421;

ld.global.nc.u32 %r244, [%rd60+28];

dp4a.s32.s32 %r242, %r387, %r244, %r238;

ld.global.nc.u32 %r248, [%rd60+24];

dp4a.s32.s32 %r246, %r391, %r248, %r242;

ld.global.nc.u32 %r252, [%rd60+32];

dp4a.s32.s32 %r250, %r395, %r252, %r246;


{cvt.rn.f16x2.e4m3x2 %r254, %rs3;}


cvt.u16.u32 %rs20, %r254;

{cvt.f32.f16 %f79, %rs20;}


div.approx.ftz.f32 %f138, %f79, %f101;
ld.global.nc.u32 %r255, [%rd60];

{.reg .f16 low,high;
mov.b32 {low,high},%r255;
cvt.f32.f16 %f80, low;}


mul.ftz.f32 %f139, %f138, %f80;
cvt.rn.f32.s32 %f140, %r250;
fma.rn.ftz.f32 %f141, %f139, %f140, %f137;
add.ftz.f32 %f307, %f141, %f307;

dp4a.s32.s32 %r256, %r401, %r222, %r421;


dp4a.s32.s32 %r260, %r405, %r226, %r256;


dp4a.s32.s32 %r264, %r409, %r230, %r260;


dp4a.s32.s32 %r268, %r413, %r234, %r264;


{cvt.rn.f16x2.e4m3x2 %r272, %rs5;}


cvt.u16.u32 %rs22, %r272;

{cvt.f32.f16 %f81, %rs22;}


div.approx.ftz.f32 %f142, %f81, %f101;

{.reg .f16 low,high;
mov.b32 {low,high},%r237;
cvt.f32.f16 %f82, low;}


mul.ftz.f32 %f143, %f142, %f82;
cvt.rn.f32.s32 %f144, %r268;
fma.rn.ftz.f32 %f145, %f143, %f144, 0f00000000;

dp4a.s32.s32 %r274, %r419, %r240, %r421;


dp4a.s32.s32 %r278, %r423, %r244, %r274;


dp4a.s32.s32 %r282, %r427, %r248, %r278;


dp4a.s32.s32 %r286, %r431, %r252, %r282;


{cvt.rn.f16x2.e4m3x2 %r290, %rs7;}


cvt.u16.u32 %rs24, %r290;

{cvt.f32.f16 %f83, %rs24;}


div.approx.ftz.f32 %f146, %f83, %f101;

{.reg .f16 low,high;
mov.b32 {low,high},%r255;
cvt.f32.f16 %f84, low;}


mul.ftz.f32 %f147, %f146, %f84;
cvt.rn.f32.s32 %f148, %r286;
fma.rn.ftz.f32 %f149, %f147, %f148, %f145;
add.ftz.f32 %f306, %f149, %f306;
cvt.u64.u32 %rd61, %r751;
add.s64 %rd62, %rd2, %rd61;
add.s64 %rd63, %rd62, %rd4;
mul.lo.s64 %rd64, %rd63, 36;
add.s64 %rd65, %rd12, %rd64;
ld.global.nc.u32 %r294, [%rd65+4];

dp4a.s32.s32 %r292, %r365, %r294, %r421;

ld.global.nc.u32 %r298, [%rd65+12];

dp4a.s32.s32 %r296, %r369, %r298, %r292;

ld.global.nc.u32 %r302, [%rd65+8];

dp4a.s32.s32 %r300, %r373, %r302, %r296;

ld.global.nc.u32 %r306, [%rd65+16];

dp4a.s32.s32 %r304, %r377, %r306, %r300;


{cvt.rn.f16x2.e4m3x2 %r308, %rs1;}


cvt.u16.u32 %rs26, %r308;

{cvt.f32.f16 %f85, %rs26;}


div.approx.ftz.f32 %f150, %f85, %f101;
ld.global.nc.u32 %r309, [%rd65];

{.reg .f16 low,high;
mov.b32 {low,high},%r309;
cvt.f32.f16 %f86, low;}


mul.ftz.f32 %f151, %f150, %f86;
cvt.rn.f32.s32 %f152, %r304;
fma.rn.ftz.f32 %f153, %f151, %f152, 0f00000000;
add.s64 %rd66, %rd62, %rd5;
mul.lo.s64 %rd67, %rd66, 36;
add.s64 %rd68, %rd12, %rd67;
ld.global.nc.u32 %r312, [%rd68+20];

dp4a.s32.s32 %r310, %r383, %r312, %r421;

ld.global.nc.u32 %r316, [%rd68+28];

dp4a.s32.s32 %r314, %r387, %r316, %r310;

ld.global.nc.u32 %r320, [%rd68+24];

dp4a.s32.s32 %r318, %r391, %r320, %r314;

ld.global.nc.u32 %r324, [%rd68+32];

dp4a.s32.s32 %r322, %r395, %r324, %r318;


{cvt.rn.f16x2.e4m3x2 %r326, %rs3;}


cvt.u16.u32 %rs28, %r326;

{cvt.f32.f16 %f87, %rs28;}


div.approx.ftz.f32 %f154, %f87, %f101;
ld.global.nc.u32 %r327, [%rd68];

{.reg .f16 low,high;
mov.b32 {low,high},%r327;
cvt.f32.f16 %f88, low;}


mul.ftz.f32 %f155, %f154, %f88;
cvt.rn.f32.s32 %f156, %r322;
fma.rn.ftz.f32 %f157, %f155, %f156, %f153;
add.ftz.f32 %f305, %f157, %f305;

dp4a.s32.s32 %r328, %r401, %r294, %r421;


dp4a.s32.s32 %r332, %r405, %r298, %r328;


dp4a.s32.s32 %r336, %r409, %r302, %r332;


dp4a.s32.s32 %r340, %r413, %r306, %r336;


{cvt.rn.f16x2.e4m3x2 %r344, %rs5;}


cvt.u16.u32 %rs30, %r344;

{cvt.f32.f16 %f89, %rs30;}


div.approx.ftz.f32 %f158, %f89, %f101;

{.reg .f16 low,high;
mov.b32 {low,high},%r309;
cvt.f32.f16 %f90, low;}


mul.ftz.f32 %f159, %f158, %f90;
cvt.rn.f32.s32 %f160, %r340;
fma.rn.ftz.f32 %f161, %f159, %f160, 0f00000000;

dp4a.s32.s32 %r346, %r419, %r312, %r421;


dp4a.s32.s32 %r350, %r423, %r316, %r346;


dp4a.s32.s32 %r354, %r427, %r320, %r350;


dp4a.s32.s32 %r358, %r431, %r324, %r354;


{cvt.rn.f16x2.e4m3x2 %r362, %rs7;}


cvt.u16.u32 %rs32, %r362;

{cvt.f32.f16 %f91, %rs32;}


div.approx.ftz.f32 %f162, %f91, %f101;

{.reg .f16 low,high;
mov.b32 {low,high},%r327;
cvt.f32.f16 %f92, low;}


mul.ftz.f32 %f163, %f162, %f92;
cvt.rn.f32.s32 %f164, %r358;
fma.rn.ftz.f32 %f165, %f163, %f164, %f161;
add.ftz.f32 %f304, %f165, %f304;
cvt.u64.u32 %rd69, %r750;
add.s64 %rd70, %rd2, %rd69;
add.s64 %rd71, %rd70, %rd4;
mul.lo.s64 %rd72, %rd71, 36;
add.s64 %rd73, %rd12, %rd72;
ld.global.nc.u32 %r366, [%rd73+4];

dp4a.s32.s32 %r364, %r365, %r366, %r421;

ld.global.nc.u32 %r370, [%rd73+12];

dp4a.s32.s32 %r368, %r369, %r370, %r364;

ld.global.nc.u32 %r374, [%rd73+8];

dp4a.s32.s32 %r372, %r373, %r374, %r368;

ld.global.nc.u32 %r378, [%rd73+16];

dp4a.s32.s32 %r376, %r377, %r378, %r372;


{cvt.rn.f16x2.e4m3x2 %r380, %rs1;}


cvt.u16.u32 %rs34, %r380;

{cvt.f32.f16 %f93, %rs34;}


div.approx.ftz.f32 %f166, %f93, %f101;
ld.global.nc.u32 %r381, [%rd73];

{.reg .f16 low,high;
mov.b32 {low,high},%r381;
cvt.f32.f16 %f94, low;}


mul.ftz.f32 %f167, %f166, %f94;
cvt.rn.f32.s32 %f168, %r376;
fma.rn.ftz.f32 %f169, %f167, %f168, 0f00000000;
add.s64 %rd74, %rd70, %rd5;
mul.lo.s64 %rd75, %rd74, 36;
add.s64 %rd76, %rd12, %rd75;
ld.global.nc.u32 %r384, [%rd76+20];

dp4a.s32.s32 %r382, %r383, %r384, %r421;

ld.global.nc.u32 %r388, [%rd76+28];

dp4a.s32.s32 %r386, %r387, %r388, %r382;

ld.global.nc.u32 %r392, [%rd76+24];

dp4a.s32.s32 %r390, %r391, %r392, %r386;

ld.global.nc.u32 %r396, [%rd76+32];

dp4a.s32.s32 %r394, %r395, %r396, %r390;


{cvt.rn.f16x2.e4m3x2 %r398, %rs3;}


cvt.u16.u32 %rs36, %r398;

{cvt.f32.f16 %f95, %rs36;}


div.approx.ftz.f32 %f170, %f95, %f101;
ld.global.nc.u32 %r399, [%rd76];

{.reg .f16 low,high;
mov.b32 {low,high},%r399;
cvt.f32.f16 %f96, low;}


mul.ftz.f32 %f171, %f170, %f96;
cvt.rn.f32.s32 %f172, %r394;
fma.rn.ftz.f32 %f173, %f171, %f172, %f169;
add.ftz.f32 %f303, %f173, %f303;

dp4a.s32.s32 %r400, %r401, %r366, %r421;


dp4a.s32.s32 %r404, %r405, %r370, %r400;


dp4a.s32.s32 %r408, %r409, %r374, %r404;


dp4a.s32.s32 %r412, %r413, %r378, %r408;


{cvt.rn.f16x2.e4m3x2 %r416, %rs5;}


cvt.u16.u32 %rs38, %r416;

{cvt.f32.f16 %f97, %rs38;}


div.approx.ftz.f32 %f174, %f97, %f101;

{.reg .f16 low,high;
mov.b32 {low,high},%r381;
cvt.f32.f16 %f98, low;}


mul.ftz.f32 %f175, %f174, %f98;
cvt.rn.f32.s32 %f176, %r412;
fma.rn.ftz.f32 %f177, %f175, %f176, 0f00000000;

dp4a.s32.s32 %r418, %r419, %r384, %r421;


dp4a.s32.s32 %r422, %r423, %r388, %r418;


dp4a.s32.s32 %r426, %r427, %r392, %r422;


dp4a.s32.s32 %r430, %r431, %r396, %r426;


{cvt.rn.f16x2.e4m3x2 %r434, %rs7;}


cvt.u16.u32 %rs40, %r434;

{cvt.f32.f16 %f99, %rs40;}


div.approx.ftz.f32 %f178, %f99, %f101;

{.reg .f16 low,high;
mov.b32 {low,high},%r399;
cvt.f32.f16 %f100, low;}


mul.ftz.f32 %f179, %f178, %f100;
cvt.rn.f32.s32 %f180, %r430;
fma.rn.ftz.f32 %f181, %f179, %f180, %f177;
add.ftz.f32 %f302, %f181, %f302;
add.s32 %r752, %r752, 64;
add.s32 %r751, %r751, 64;
add.s32 %r750, %r750, 64;
add.s32 %r749, %r749, 64;
add.s32 %r748, %r748, 64;
add.s64 %rd100, %rd100, 1152;
add.s32 %r753, %r753, 32;
setp.lt.s32 %p8, %r753, %r7;
@%p8 bra $L__BB123_2;

$L__BB123_3:
setp.eq.s32 %p9, %r1, 0;
@%p9 bra $L__BB123_5;

mov.u32 %r563, _ZZ13mul_mat_vec_qIL9ggml_type40ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r564, %r1, 1280, %r563;
shl.b32 %r565, %r44, 2;
add.s32 %r566, %r564, %r565;
st.shared.f32 [%r566+-1280], %f311;
st.shared.f32 [%r566+-1152], %f310;
st.shared.f32 [%r566+-1024], %f309;
st.shared.f32 [%r566+-896], %f308;
st.shared.f32 [%r566+-768], %f307;
st.shared.f32 [%r566+-640], %f306;
st.shared.f32 [%r566+-512], %f305;
st.shared.f32 [%r566+-384], %f304;
st.shared.f32 [%r566+-256], %f303;
st.shared.f32 [%r566+-128], %f302;

$L__BB123_5:
setp.ne.s32 %p10, %r1, 0;
bar.sync 0;
@%p10 bra $L__BB123_26;

mad.lo.s32 %r567, %r4, %r35, %r3;
mad.lo.s32 %r568, %r5, %r41, %r567;
cvt.u64.u32 %rd15, %r568;
shl.b32 %r570, %r44, 2;
mov.u32 %r571, _ZZ13mul_mat_vec_qIL9ggml_type40ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r25, %r571, %r570;
mov.u32 %r572, 2;
ld.shared.f32 %f182, [%r25];
add.ftz.f32 %f183, %f182, %f311;
mov.b32 %r573, %f183;
mov.u32 %r574, 31;
mov.u32 %r575, 16;
mov.u32 %r576, -1;
shfl.sync.bfly.b32 %r577|%p11, %r573, %r575, %r574, %r576;
mov.b32 %f184, %r577;
add.ftz.f32 %f185, %f183, %f184;
mov.b32 %r578, %f185;
mov.u32 %r579, 8;
shfl.sync.bfly.b32 %r580|%p12, %r578, %r579, %r574, %r576;
mov.b32 %f186, %r580;
add.ftz.f32 %f187, %f185, %f186;
mov.b32 %r581, %f187;
mov.u32 %r582, 4;
shfl.sync.bfly.b32 %r583|%p13, %r581, %r582, %r574, %r576;
mov.b32 %f188, %r583;
add.ftz.f32 %f189, %f187, %f188;
mov.b32 %r584, %f189;
shfl.sync.bfly.b32 %r585|%p14, %r584, %r572, %r574, %r576;
mov.b32 %f190, %r585;
add.ftz.f32 %f191, %f189, %f190;
mov.b32 %r586, %f191;
mov.u32 %r587, 1;
shfl.sync.bfly.b32 %r588|%p15, %r586, %r587, %r574, %r576;
mov.b32 %f192, %r588;
add.ftz.f32 %f31, %f191, %f192;
setp.eq.s32 %p16, %r44, 0;
setp.lt.u32 %p17, %r3, %r29;
and.pred %p1, %p16, %p17;
mul.wide.u32 %rd77, %r568, 4;
add.s64 %rd16, %rd1, %rd77;
not.pred %p18, %p1;
@%p18 bra $L__BB123_8;

st.global.f32 [%rd16], %f31;

$L__BB123_8:
ld.shared.f32 %f193, [%r25+128];
add.ftz.f32 %f194, %f193, %f310;
mov.b32 %r589, %f194;
shfl.sync.bfly.b32 %r593|%p19, %r589, %r575, %r574, %r576;
mov.b32 %f195, %r593;
add.ftz.f32 %f196, %f194, %f195;
mov.b32 %r594, %f196;
shfl.sync.bfly.b32 %r596|%p20, %r594, %r579, %r574, %r576;
mov.b32 %f197, %r596;
add.ftz.f32 %f198, %f196, %f197;
mov.b32 %r597, %f198;
shfl.sync.bfly.b32 %r599|%p21, %r597, %r582, %r574, %r576;
mov.b32 %f199, %r599;
add.ftz.f32 %f200, %f198, %f199;
mov.b32 %r600, %f200;
shfl.sync.bfly.b32 %r602|%p22, %r600, %r572, %r574, %r576;
mov.b32 %f201, %r602;
add.ftz.f32 %f202, %f200, %f201;
mov.b32 %r603, %f202;
shfl.sync.bfly.b32 %r605|%p23, %r603, %r587, %r574, %r576;
mov.b32 %f203, %r605;
add.ftz.f32 %f32, %f202, %f203;
setp.eq.s32 %p24, %r44, 1;
add.s32 %r607, %r3, 1;
setp.lt.u32 %p25, %r607, %r29;
and.pred %p2, %p24, %p25;
not.pred %p26, %p2;
@%p26 bra $L__BB123_10;

st.global.f32 [%rd16+4], %f32;

$L__BB123_10:
ld.shared.f32 %f204, [%r25+256];
add.ftz.f32 %f205, %f204, %f309;
mov.b32 %r608, %f205;
mov.u32 %r609, 31;
mov.u32 %r610, 16;
mov.u32 %r611, -1;
shfl.sync.bfly.b32 %r612|%p27, %r608, %r610, %r609, %r611;
mov.b32 %f206, %r612;
add.ftz.f32 %f207, %f205, %f206;
mov.b32 %r613, %f207;
mov.u32 %r614, 8;
shfl.sync.bfly.b32 %r615|%p28, %r613, %r614, %r609, %r611;
mov.b32 %f208, %r615;
add.ftz.f32 %f209, %f207, %f208;
mov.b32 %r616, %f209;
mov.u32 %r617, 4;
shfl.sync.bfly.b32 %r618|%p29, %r616, %r617, %r609, %r611;
mov.b32 %f210, %r618;
add.ftz.f32 %f211, %f209, %f210;
mov.b32 %r619, %f211;
mov.u32 %r620, 2;
shfl.sync.bfly.b32 %r621|%p30, %r619, %r620, %r609, %r611;
mov.b32 %f212, %r621;
add.ftz.f32 %f213, %f211, %f212;
mov.b32 %r622, %f213;
mov.u32 %r623, 1;
shfl.sync.bfly.b32 %r624|%p31, %r622, %r623, %r609, %r611;
mov.b32 %f214, %r624;
add.ftz.f32 %f33, %f213, %f214;
@%p18 bra $L__BB123_12;

cvt.u64.u32 %rd78, %r29;
add.s64 %rd79, %rd78, %rd15;
shl.b64 %rd80, %rd79, 2;
add.s64 %rd81, %rd1, %rd80;
st.global.f32 [%rd81], %f33;

$L__BB123_12:
ld.shared.f32 %f215, [%r25+384];
add.ftz.f32 %f216, %f215, %f308;
mov.b32 %r625, %f216;
shfl.sync.bfly.b32 %r629|%p33, %r625, %r610, %r609, %r611;
mov.b32 %f217, %r629;
add.ftz.f32 %f218, %f216, %f217;
mov.b32 %r630, %f218;
shfl.sync.bfly.b32 %r632|%p34, %r630, %r614, %r609, %r611;
mov.b32 %f219, %r632;
add.ftz.f32 %f220, %f218, %f219;
mov.b32 %r633, %f220;
shfl.sync.bfly.b32 %r635|%p35, %r633, %r617, %r609, %r611;
mov.b32 %f221, %r635;
add.ftz.f32 %f222, %f220, %f221;
mov.b32 %r636, %f222;
shfl.sync.bfly.b32 %r638|%p36, %r636, %r620, %r609, %r611;
mov.b32 %f223, %r638;
add.ftz.f32 %f224, %f222, %f223;
mov.b32 %r639, %f224;
shfl.sync.bfly.b32 %r641|%p37, %r639, %r623, %r609, %r611;
mov.b32 %f225, %r641;
add.ftz.f32 %f34, %f224, %f225;
@%p26 bra $L__BB123_14;

add.s32 %r642, %r29, 1;
cvt.u64.u32 %rd82, %r642;
add.s64 %rd83, %rd82, %rd15;
shl.b64 %rd84, %rd83, 2;
add.s64 %rd85, %rd1, %rd84;
st.global.f32 [%rd85], %f34;

$L__BB123_14:
shl.b32 %r643, %r29, 1;
mov.u32 %r644, 1;
ld.shared.f32 %f226, [%r25+512];
add.ftz.f32 %f227, %f226, %f307;
mov.b32 %r645, %f227;
mov.u32 %r646, 31;
mov.u32 %r647, 16;
mov.u32 %r648, -1;
shfl.sync.bfly.b32 %r649|%p39, %r645, %r647, %r646, %r648;
mov.b32 %f228, %r649;
add.ftz.f32 %f229, %f227, %f228;
mov.b32 %r650, %f229;
mov.u32 %r651, 8;
shfl.sync.bfly.b32 %r652|%p40, %r650, %r651, %r646, %r648;
mov.b32 %f230, %r652;
add.ftz.f32 %f231, %f229, %f230;
mov.b32 %r653, %f231;
mov.u32 %r654, 4;
shfl.sync.bfly.b32 %r655|%p41, %r653, %r654, %r646, %r648;
mov.b32 %f232, %r655;
add.ftz.f32 %f233, %f231, %f232;
mov.b32 %r656, %f233;
mov.u32 %r657, 2;
shfl.sync.bfly.b32 %r658|%p42, %r656, %r657, %r646, %r648;
mov.b32 %f234, %r658;
add.ftz.f32 %f235, %f233, %f234;
mov.b32 %r659, %f235;
shfl.sync.bfly.b32 %r660|%p43, %r659, %r644, %r646, %r648;
mov.b32 %f236, %r660;
add.ftz.f32 %f35, %f235, %f236;
cvt.u64.u32 %rd86, %r643;
add.s64 %rd87, %rd86, %rd15;
shl.b64 %rd88, %rd87, 2;
add.s64 %rd17, %rd1, %rd88;
@%p18 bra $L__BB123_16;

st.global.f32 [%rd17], %f35;

$L__BB123_16:
ld.shared.f32 %f237, [%r25+640];
add.ftz.f32 %f238, %f237, %f306;
mov.b32 %r661, %f238;
shfl.sync.bfly.b32 %r665|%p45, %r661, %r647, %r646, %r648;
mov.b32 %f239, %r665;
add.ftz.f32 %f240, %f238, %f239;
mov.b32 %r666, %f240;
shfl.sync.bfly.b32 %r668|%p46, %r666, %r651, %r646, %r648;
mov.b32 %f241, %r668;
add.ftz.f32 %f242, %f240, %f241;
mov.b32 %r669, %f242;
shfl.sync.bfly.b32 %r671|%p47, %r669, %r654, %r646, %r648;
mov.b32 %f243, %r671;
add.ftz.f32 %f244, %f242, %f243;
mov.b32 %r672, %f244;
shfl.sync.bfly.b32 %r674|%p48, %r672, %r657, %r646, %r648;
mov.b32 %f245, %r674;
add.ftz.f32 %f246, %f244, %f245;
mov.b32 %r675, %f246;
shfl.sync.bfly.b32 %r677|%p49, %r675, %r644, %r646, %r648;
mov.b32 %f247, %r677;
add.ftz.f32 %f36, %f246, %f247;
@%p26 bra $L__BB123_18;

st.global.f32 [%rd17+4], %f36;

$L__BB123_18:
mul.lo.s32 %r26, %r29, 3;
ld.shared.f32 %f248, [%r25+768];
add.ftz.f32 %f249, %f248, %f305;
mov.b32 %r678, %f249;
mov.u32 %r679, 31;
mov.u32 %r680, 16;
mov.u32 %r681, -1;
shfl.sync.bfly.b32 %r682|%p51, %r678, %r680, %r679, %r681;
mov.b32 %f250, %r682;
add.ftz.f32 %f251, %f249, %f250;
mov.b32 %r683, %f251;
mov.u32 %r684, 8;
shfl.sync.bfly.b32 %r685|%p52, %r683, %r684, %r679, %r681;
mov.b32 %f252, %r685;
add.ftz.f32 %f253, %f251, %f252;
mov.b32 %r686, %f253;
mov.u32 %r687, 4;
shfl.sync.bfly.b32 %r688|%p53, %r686, %r687, %r679, %r681;
mov.b32 %f254, %r688;
add.ftz.f32 %f255, %f253, %f254;
mov.b32 %r689, %f255;
mov.u32 %r690, 2;
shfl.sync.bfly.b32 %r691|%p54, %r689, %r690, %r679, %r681;
mov.b32 %f256, %r691;
add.ftz.f32 %f257, %f255, %f256;
mov.b32 %r692, %f257;
mov.u32 %r693, 1;
shfl.sync.bfly.b32 %r694|%p55, %r692, %r693, %r679, %r681;
mov.b32 %f258, %r694;
add.ftz.f32 %f37, %f257, %f258;
@%p18 bra $L__BB123_20;

cvt.u64.u32 %rd89, %r26;
add.s64 %rd90, %rd89, %rd15;
shl.b64 %rd91, %rd90, 2;
add.s64 %rd92, %rd1, %rd91;
st.global.f32 [%rd92], %f37;

$L__BB123_20:
ld.shared.f32 %f259, [%r25+896];
add.ftz.f32 %f260, %f259, %f304;
mov.b32 %r695, %f260;
shfl.sync.bfly.b32 %r699|%p57, %r695, %r680, %r679, %r681;
mov.b32 %f261, %r699;
add.ftz.f32 %f262, %f260, %f261;
mov.b32 %r700, %f262;
shfl.sync.bfly.b32 %r702|%p58, %r700, %r684, %r679, %r681;
mov.b32 %f263, %r702;
add.ftz.f32 %f264, %f262, %f263;
mov.b32 %r703, %f264;
shfl.sync.bfly.b32 %r705|%p59, %r703, %r687, %r679, %r681;
mov.b32 %f265, %r705;
add.ftz.f32 %f266, %f264, %f265;
mov.b32 %r706, %f266;
shfl.sync.bfly.b32 %r708|%p60, %r706, %r690, %r679, %r681;
mov.b32 %f267, %r708;
add.ftz.f32 %f268, %f266, %f267;
mov.b32 %r709, %f268;
shfl.sync.bfly.b32 %r711|%p61, %r709, %r693, %r679, %r681;
mov.b32 %f269, %r711;
add.ftz.f32 %f38, %f268, %f269;
@%p26 bra $L__BB123_22;

add.s32 %r712, %r26, 1;
cvt.u64.u32 %rd93, %r712;
add.s64 %rd94, %rd93, %rd15;
shl.b64 %rd95, %rd94, 2;
add.s64 %rd96, %rd1, %rd95;
st.global.f32 [%rd96], %f38;

$L__BB123_22:
shl.b32 %r713, %r29, 2;
mov.u32 %r714, 2;
ld.shared.f32 %f270, [%r25+1024];
add.ftz.f32 %f271, %f270, %f303;
mov.b32 %r715, %f271;
mov.u32 %r716, 31;
mov.u32 %r717, 16;
mov.u32 %r718, -1;
shfl.sync.bfly.b32 %r719|%p63, %r715, %r717, %r716, %r718;
mov.b32 %f272, %r719;
add.ftz.f32 %f273, %f271, %f272;
mov.b32 %r720, %f273;
mov.u32 %r721, 8;
shfl.sync.bfly.b32 %r722|%p64, %r720, %r721, %r716, %r718;
mov.b32 %f274, %r722;
add.ftz.f32 %f275, %f273, %f274;
mov.b32 %r723, %f275;
mov.u32 %r724, 4;
shfl.sync.bfly.b32 %r725|%p65, %r723, %r724, %r716, %r718;
mov.b32 %f276, %r725;
add.ftz.f32 %f277, %f275, %f276;
mov.b32 %r726, %f277;
shfl.sync.bfly.b32 %r727|%p66, %r726, %r714, %r716, %r718;
mov.b32 %f278, %r727;
add.ftz.f32 %f279, %f277, %f278;
mov.b32 %r728, %f279;
mov.u32 %r729, 1;
shfl.sync.bfly.b32 %r730|%p67, %r728, %r729, %r716, %r718;
mov.b32 %f280, %r730;
add.ftz.f32 %f39, %f279, %f280;
cvt.u64.u32 %rd97, %r713;
add.s64 %rd98, %rd97, %rd15;
shl.b64 %rd99, %rd98, 2;
add.s64 %rd18, %rd1, %rd99;
@%p18 bra $L__BB123_24;

st.global.f32 [%rd18], %f39;

$L__BB123_24:
ld.shared.f32 %f281, [%r25+1152];
add.ftz.f32 %f282, %f281, %f302;
mov.b32 %r731, %f282;
shfl.sync.bfly.b32 %r735|%p69, %r731, %r717, %r716, %r718;
mov.b32 %f283, %r735;
add.ftz.f32 %f284, %f282, %f283;
mov.b32 %r736, %f284;
shfl.sync.bfly.b32 %r738|%p70, %r736, %r721, %r716, %r718;
mov.b32 %f285, %r738;
add.ftz.f32 %f286, %f284, %f285;
mov.b32 %r739, %f286;
shfl.sync.bfly.b32 %r741|%p71, %r739, %r724, %r716, %r718;
mov.b32 %f287, %r741;
add.ftz.f32 %f288, %f286, %f287;
mov.b32 %r742, %f288;
shfl.sync.bfly.b32 %r744|%p72, %r742, %r714, %r716, %r718;
mov.b32 %f289, %r744;
add.ftz.f32 %f290, %f288, %f289;
mov.b32 %r745, %f290;
shfl.sync.bfly.b32 %r747|%p73, %r745, %r729, %r716, %r718;
mov.b32 %f291, %r747;
add.ftz.f32 %f40, %f290, %f291;
@%p26 bra $L__BB123_26;

st.global.f32 [%rd18+4], %f40;

$L__BB123_26:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type40ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type40ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type40ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type40ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type40ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type40ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type40ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type40ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type40ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<87>;
.reg .b16 %rs<61>;
.reg .f32 %f<374>;
.reg .b32 %r<869>;
.reg .b64 %rd<125>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type40ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type40ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type40ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd20, [_Z13mul_mat_vec_qIL9ggml_type40ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type40ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type40ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type40ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type40ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type40ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type40ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type40ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type40ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type40ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type40ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type40ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type40ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type40ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type40ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r45, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r45, %r2;
mov.u32 %r46, %ctaid.x;
shl.b32 %r4, %r46, 1;
shr.u32 %r47, %r3, 31;
add.s32 %r48, %r3, %r47;
shr.s32 %r868, %r48, 1;
shr.u32 %r49, %r29, 6;
setp.ge.s32 %p3, %r868, %r49;
mov.f32 %f362, 0f00000000;
mov.f32 %f363, %f362;
mov.f32 %f364, %f362;
mov.f32 %f365, %f362;
mov.f32 %f366, %f362;
mov.f32 %f367, %f362;
mov.f32 %f368, %f362;
mov.f32 %f369, %f362;
mov.f32 %f370, %f362;
mov.f32 %f371, %f362;
mov.f32 %f372, %f362;
mov.f32 %f373, %f362;
@%p3 bra $L__BB124_3;

mov.u32 %r50, %ctaid.y;
mov.u32 %r51, %ctaid.z;
mul.lo.s32 %r52, %r50, %r37;
cvt.u64.u32 %rd21, %r52;
mul.lo.s32 %r53, %r51, %r43;
cvt.u64.u32 %rd22, %r53;
add.s64 %rd1, %rd22, %rd21;
mul.hi.u32 %r54, %r51, %r39;
add.s32 %r55, %r54, %r51;
shr.u32 %r56, %r55, %r40;
mul.lo.s32 %r57, %r56, %r42;
mul.hi.u32 %r58, %r50, %r33;
add.s32 %r59, %r58, %r50;
shr.u32 %r60, %r59, %r34;
mul.lo.s32 %r61, %r60, %r36;
cvta.to.global.u64 %rd124, %rd18;
and.b32 %r64, %r48, -2;
sub.s32 %r65, %r3, %r64;
shl.b32 %r66, %r65, 2;
shl.b32 %r67, %r65, 1;
cvt.s64.s32 %rd3, %r65;
cvt.s64.s32 %rd23, %r67;
or.b32 %r68, %r66, 2;
shr.s32 %r69, %r68, 1;
shr.s32 %r70, %r68, 2;
cvt.s64.s32 %rd4, %r70;
cvt.s64.s32 %rd24, %r69;
shl.b32 %r71, %r31, 1;
shl.b32 %r862, %r868, 1;
add.s32 %r867, %r862, %r71;
mad.lo.s32 %r866, %r31, 3, %r862;
shl.b32 %r72, %r31, 2;
add.s32 %r865, %r862, %r72;
mad.lo.s32 %r864, %r31, 5, %r862;
add.s32 %r863, %r31, %r862;
add.s32 %r73, %r868, %r57;
add.s32 %r74, %r73, %r61;
add.s32 %r75, %r4, 1;
mad.lo.s32 %r76, %r30, %r75, %r74;
mul.wide.s32 %rd25, %r76, 36;
add.s64 %rd5, %rd25, %rd24;
add.s64 %rd6, %rd25, %rd23;
mad.lo.s32 %r77, %r4, %r30, %r868;
add.s32 %r78, %r77, %r57;
add.s32 %r79, %r78, %r61;
mul.wide.s32 %rd26, %r79, 36;
mul.wide.s32 %rd27, %r66, 4;
add.s64 %rd28, %rd26, %rd27;
add.s64 %rd7, %rd28, 16;
add.s64 %rd8, %rd26, %rd23;
add.s64 %rd29, %rd25, %rd27;
add.s64 %rd9, %rd29, 8;
add.s64 %rd10, %rd26, %rd24;
cvta.to.global.u64 %rd11, %rd19;
mov.f32 %f362, 0f00000000;

$L__BB124_2:
cvt.u64.u32 %rd30, %r862;
add.s64 %rd31, %rd1, %rd30;
add.s64 %rd32, %rd124, %rd7;
ld.global.nc.u32 %r512, [%rd32+-12];
and.b32 %r513, %r512, -2004318072;
shr.u32 %r514, %r513, 1;
or.b32 %r515, %r514, 839922192;
and.b32 %r516, %r512, 30583;
mov.u32 %r517, 201852420;
mov.u32 %r518, 50462976;
prmt.b32 %r519, %r518, %r517, %r516;
mov.u32 %r520, -185009412;
mov.u32 %r521, -33620224;
prmt.b32 %r522, %r521, %r520, %r516;
and.b32 %r523, %r515, 30292;
prmt.b32 %r524, %r519, %r522, %r523;
shr.u32 %r525, %r512, 16;
and.b32 %r526, %r525, 30583;
prmt.b32 %r527, %r518, %r517, %r526;
prmt.b32 %r528, %r521, %r520, %r526;
shr.u32 %r529, %r515, 16;
prmt.b32 %r530, %r527, %r528, %r529;
mov.u32 %r531, 25632;
prmt.b32 %r441, %r524, %r530, %r531;
mov.u32 %r532, 30001;
prmt.b32 %r445, %r524, %r530, %r532;
ld.global.nc.u32 %r533, [%rd32+-8];
and.b32 %r534, %r533, -2004318072;
shr.u32 %r535, %r534, 1;
or.b32 %r536, %r535, 839922192;
and.b32 %r537, %r533, 30583;
prmt.b32 %r538, %r518, %r517, %r537;
prmt.b32 %r539, %r521, %r520, %r537;
and.b32 %r540, %r536, 30292;
prmt.b32 %r541, %r538, %r539, %r540;
shr.u32 %r542, %r533, 16;
and.b32 %r543, %r542, 30583;
prmt.b32 %r544, %r518, %r517, %r543;
prmt.b32 %r545, %r521, %r520, %r543;
shr.u32 %r546, %r536, 16;
prmt.b32 %r547, %r544, %r545, %r546;
prmt.b32 %r449, %r541, %r547, %r531;
prmt.b32 %r453, %r541, %r547, %r532;
add.s64 %rd33, %rd31, %rd3;
mul.lo.s64 %rd34, %rd33, 36;
add.s64 %rd35, %rd11, %rd34;
ld.global.nc.u32 %r82, [%rd35+4];
mov.u32 %r497, 0;

dp4a.s32.s32 %r80, %r441, %r82, %r497;

ld.global.nc.u32 %r86, [%rd35+12];

dp4a.s32.s32 %r84, %r445, %r86, %r80;

ld.global.nc.u32 %r90, [%rd35+8];

dp4a.s32.s32 %r88, %r449, %r90, %r84;

ld.global.nc.u32 %r94, [%rd35+16];

dp4a.s32.s32 %r92, %r453, %r94, %r88;

add.s64 %rd36, %rd124, %rd8;
ld.global.nc.u8 %rs49, [%rd36];
and.b16 %rs50, %rs49, 127;
setp.eq.s16 %p4, %rs50, 127;
selp.b16 %rs51, 0, %rs49, %p4;
and.b16 %rs1, %rs51, 255;

{cvt.rn.f16x2.e4m3x2 %r96, %rs1;}


cvt.u16.u32 %rs2, %r96;

{cvt.f32.f16 %f73, %rs2;}


mov.f32 %f121, 0f40000000;
div.approx.ftz.f32 %f122, %f73, %f121;
ld.global.nc.u32 %r97, [%rd35];

{.reg .f16 low,high;
mov.b32 {low,high},%r97;
cvt.f32.f16 %f74, low;}


mul.ftz.f32 %f123, %f122, %f74;
cvt.rn.f32.s32 %f124, %r92;
fma.rn.ftz.f32 %f125, %f123, %f124, 0f00000000;
ld.global.nc.u32 %r548, [%rd32+-4];
and.b32 %r549, %r548, -2004318072;
shr.u32 %r550, %r549, 1;
or.b32 %r551, %r550, 839922192;
and.b32 %r552, %r548, 30583;
prmt.b32 %r553, %r518, %r517, %r552;
prmt.b32 %r554, %r521, %r520, %r552;
and.b32 %r555, %r551, 30292;
prmt.b32 %r556, %r553, %r554, %r555;
shr.u32 %r557, %r548, 16;
and.b32 %r558, %r557, 30583;
prmt.b32 %r559, %r518, %r517, %r558;
prmt.b32 %r560, %r521, %r520, %r558;
shr.u32 %r561, %r551, 16;
prmt.b32 %r562, %r559, %r560, %r561;
prmt.b32 %r459, %r556, %r562, %r531;
prmt.b32 %r463, %r556, %r562, %r532;
ld.global.nc.u32 %r563, [%rd32];
and.b32 %r564, %r563, -2004318072;
shr.u32 %r565, %r564, 1;
or.b32 %r566, %r565, 839922192;
and.b32 %r567, %r563, 30583;
prmt.b32 %r568, %r518, %r517, %r567;
prmt.b32 %r569, %r521, %r520, %r567;
and.b32 %r570, %r566, 30292;
prmt.b32 %r571, %r568, %r569, %r570;
shr.u32 %r572, %r563, 16;
and.b32 %r573, %r572, 30583;
prmt.b32 %r574, %r518, %r517, %r573;
prmt.b32 %r575, %r521, %r520, %r573;
shr.u32 %r576, %r566, 16;
prmt.b32 %r577, %r574, %r575, %r576;
prmt.b32 %r467, %r571, %r577, %r531;
prmt.b32 %r471, %r571, %r577, %r532;
add.s64 %rd37, %rd31, %rd4;
mul.lo.s64 %rd38, %rd37, 36;
add.s64 %rd39, %rd11, %rd38;
ld.global.nc.u32 %r100, [%rd39+20];

dp4a.s32.s32 %r98, %r459, %r100, %r497;

ld.global.nc.u32 %r104, [%rd39+28];

dp4a.s32.s32 %r102, %r463, %r104, %r98;

ld.global.nc.u32 %r108, [%rd39+24];

dp4a.s32.s32 %r106, %r467, %r108, %r102;

ld.global.nc.u32 %r112, [%rd39+32];

dp4a.s32.s32 %r110, %r471, %r112, %r106;

add.s64 %rd40, %rd124, %rd10;
ld.global.nc.u8 %rs52, [%rd40];
and.b16 %rs53, %rs52, 127;
setp.eq.s16 %p5, %rs53, 127;
selp.b16 %rs54, 0, %rs52, %p5;
and.b16 %rs3, %rs54, 255;

{cvt.rn.f16x2.e4m3x2 %r114, %rs3;}


cvt.u16.u32 %rs4, %r114;

{cvt.f32.f16 %f75, %rs4;}


div.approx.ftz.f32 %f126, %f75, %f121;
ld.global.nc.u32 %r115, [%rd39];

{.reg .f16 low,high;
mov.b32 {low,high},%r115;
cvt.f32.f16 %f76, low;}


mul.ftz.f32 %f127, %f126, %f76;
cvt.rn.f32.s32 %f128, %r110;
fma.rn.ftz.f32 %f129, %f127, %f128, %f125;
add.ftz.f32 %f373, %f129, %f373;
add.s64 %rd41, %rd124, %rd9;
ld.global.nc.u32 %r578, [%rd41+-4];
and.b32 %r579, %r578, -2004318072;
shr.u32 %r580, %r579, 1;
or.b32 %r581, %r580, 839922192;
and.b32 %r582, %r578, 30583;
prmt.b32 %r583, %r518, %r517, %r582;
prmt.b32 %r584, %r521, %r520, %r582;
and.b32 %r585, %r581, 30292;
prmt.b32 %r586, %r583, %r584, %r585;
shr.u32 %r587, %r578, 16;
and.b32 %r588, %r587, 30583;
prmt.b32 %r589, %r518, %r517, %r588;
prmt.b32 %r590, %r521, %r520, %r588;
shr.u32 %r591, %r581, 16;
prmt.b32 %r592, %r589, %r590, %r591;
prmt.b32 %r477, %r586, %r592, %r531;
prmt.b32 %r481, %r586, %r592, %r532;
ld.global.nc.u32 %r593, [%rd41];
and.b32 %r594, %r593, -2004318072;
shr.u32 %r595, %r594, 1;
or.b32 %r596, %r595, 839922192;
and.b32 %r597, %r593, 30583;
prmt.b32 %r598, %r518, %r517, %r597;
prmt.b32 %r599, %r521, %r520, %r597;
and.b32 %r600, %r596, 30292;
prmt.b32 %r601, %r598, %r599, %r600;
shr.u32 %r602, %r593, 16;
and.b32 %r603, %r602, 30583;
prmt.b32 %r604, %r518, %r517, %r603;
prmt.b32 %r605, %r521, %r520, %r603;
shr.u32 %r606, %r596, 16;
prmt.b32 %r607, %r604, %r605, %r606;
prmt.b32 %r485, %r601, %r607, %r531;
prmt.b32 %r489, %r601, %r607, %r532;

dp4a.s32.s32 %r116, %r477, %r82, %r497;


dp4a.s32.s32 %r120, %r481, %r86, %r116;


dp4a.s32.s32 %r124, %r485, %r90, %r120;


dp4a.s32.s32 %r128, %r489, %r94, %r124;

add.s64 %rd42, %rd124, %rd6;
ld.global.nc.u8 %rs55, [%rd42];
and.b16 %rs56, %rs55, 127;
setp.eq.s16 %p6, %rs56, 127;
selp.b16 %rs57, 0, %rs55, %p6;
and.b16 %rs5, %rs57, 255;

{cvt.rn.f16x2.e4m3x2 %r132, %rs5;}


cvt.u16.u32 %rs6, %r132;

{cvt.f32.f16 %f77, %rs6;}


div.approx.ftz.f32 %f130, %f77, %f121;

{.reg .f16 low,high;
mov.b32 {low,high},%r97;
cvt.f32.f16 %f78, low;}


mul.ftz.f32 %f131, %f130, %f78;
cvt.rn.f32.s32 %f132, %r128;
fma.rn.ftz.f32 %f133, %f131, %f132, 0f00000000;
ld.global.nc.u32 %r608, [%rd41+4];
and.b32 %r609, %r608, -2004318072;
shr.u32 %r610, %r609, 1;
or.b32 %r611, %r610, 839922192;
and.b32 %r612, %r608, 30583;
prmt.b32 %r613, %r518, %r517, %r612;
prmt.b32 %r614, %r521, %r520, %r612;
and.b32 %r615, %r611, 30292;
prmt.b32 %r616, %r613, %r614, %r615;
shr.u32 %r617, %r608, 16;
and.b32 %r618, %r617, 30583;
prmt.b32 %r619, %r518, %r517, %r618;
prmt.b32 %r620, %r521, %r520, %r618;
shr.u32 %r621, %r611, 16;
prmt.b32 %r622, %r619, %r620, %r621;
prmt.b32 %r495, %r616, %r622, %r531;
prmt.b32 %r499, %r616, %r622, %r532;
ld.global.nc.u32 %r623, [%rd41+8];
and.b32 %r624, %r623, -2004318072;
shr.u32 %r625, %r624, 1;
or.b32 %r626, %r625, 839922192;
and.b32 %r627, %r623, 30583;
prmt.b32 %r628, %r518, %r517, %r627;
prmt.b32 %r629, %r521, %r520, %r627;
and.b32 %r630, %r626, 30292;
prmt.b32 %r631, %r628, %r629, %r630;
shr.u32 %r632, %r623, 16;
and.b32 %r633, %r632, 30583;
prmt.b32 %r634, %r518, %r517, %r633;
prmt.b32 %r635, %r521, %r520, %r633;
shr.u32 %r636, %r626, 16;
prmt.b32 %r637, %r634, %r635, %r636;
prmt.b32 %r503, %r631, %r637, %r531;
prmt.b32 %r507, %r631, %r637, %r532;

dp4a.s32.s32 %r134, %r495, %r100, %r497;


dp4a.s32.s32 %r138, %r499, %r104, %r134;


dp4a.s32.s32 %r142, %r503, %r108, %r138;


dp4a.s32.s32 %r146, %r507, %r112, %r142;

add.s64 %rd43, %rd124, %rd5;
ld.global.nc.u8 %rs58, [%rd43];
and.b16 %rs59, %rs58, 127;
setp.eq.s16 %p7, %rs59, 127;
selp.b16 %rs60, 0, %rs58, %p7;
and.b16 %rs7, %rs60, 255;

{cvt.rn.f16x2.e4m3x2 %r150, %rs7;}


cvt.u16.u32 %rs8, %r150;

{cvt.f32.f16 %f79, %rs8;}


div.approx.ftz.f32 %f134, %f79, %f121;

{.reg .f16 low,high;
mov.b32 {low,high},%r115;
cvt.f32.f16 %f80, low;}


mul.ftz.f32 %f135, %f134, %f80;
cvt.rn.f32.s32 %f136, %r146;
fma.rn.ftz.f32 %f137, %f135, %f136, %f133;
add.ftz.f32 %f372, %f137, %f372;
cvt.u64.u32 %rd44, %r863;
add.s64 %rd45, %rd1, %rd44;
add.s64 %rd46, %rd45, %rd3;
mul.lo.s64 %rd47, %rd46, 36;
add.s64 %rd48, %rd11, %rd47;
ld.global.nc.u32 %r154, [%rd48+4];

dp4a.s32.s32 %r152, %r441, %r154, %r497;

ld.global.nc.u32 %r158, [%rd48+12];

dp4a.s32.s32 %r156, %r445, %r158, %r152;

ld.global.nc.u32 %r162, [%rd48+8];

dp4a.s32.s32 %r160, %r449, %r162, %r156;

ld.global.nc.u32 %r166, [%rd48+16];

dp4a.s32.s32 %r164, %r453, %r166, %r160;


{cvt.rn.f16x2.e4m3x2 %r168, %rs1;}


cvt.u16.u32 %rs10, %r168;

{cvt.f32.f16 %f81, %rs10;}


div.approx.ftz.f32 %f138, %f81, %f121;
ld.global.nc.u32 %r169, [%rd48];

{.reg .f16 low,high;
mov.b32 {low,high},%r169;
cvt.f32.f16 %f82, low;}


mul.ftz.f32 %f139, %f138, %f82;
cvt.rn.f32.s32 %f140, %r164;
fma.rn.ftz.f32 %f141, %f139, %f140, 0f00000000;
add.s64 %rd49, %rd45, %rd4;
mul.lo.s64 %rd50, %rd49, 36;
add.s64 %rd51, %rd11, %rd50;
ld.global.nc.u32 %r172, [%rd51+20];

dp4a.s32.s32 %r170, %r459, %r172, %r497;

ld.global.nc.u32 %r176, [%rd51+28];

dp4a.s32.s32 %r174, %r463, %r176, %r170;

ld.global.nc.u32 %r180, [%rd51+24];

dp4a.s32.s32 %r178, %r467, %r180, %r174;

ld.global.nc.u32 %r184, [%rd51+32];

dp4a.s32.s32 %r182, %r471, %r184, %r178;


{cvt.rn.f16x2.e4m3x2 %r186, %rs3;}


cvt.u16.u32 %rs12, %r186;

{cvt.f32.f16 %f83, %rs12;}


div.approx.ftz.f32 %f142, %f83, %f121;
ld.global.nc.u32 %r187, [%rd51];

{.reg .f16 low,high;
mov.b32 {low,high},%r187;
cvt.f32.f16 %f84, low;}


mul.ftz.f32 %f143, %f142, %f84;
cvt.rn.f32.s32 %f144, %r182;
fma.rn.ftz.f32 %f145, %f143, %f144, %f141;
add.ftz.f32 %f371, %f145, %f371;

dp4a.s32.s32 %r188, %r477, %r154, %r497;


dp4a.s32.s32 %r192, %r481, %r158, %r188;


dp4a.s32.s32 %r196, %r485, %r162, %r192;


dp4a.s32.s32 %r200, %r489, %r166, %r196;


{cvt.rn.f16x2.e4m3x2 %r204, %rs5;}


cvt.u16.u32 %rs14, %r204;

{cvt.f32.f16 %f85, %rs14;}


div.approx.ftz.f32 %f146, %f85, %f121;

{.reg .f16 low,high;
mov.b32 {low,high},%r169;
cvt.f32.f16 %f86, low;}


mul.ftz.f32 %f147, %f146, %f86;
cvt.rn.f32.s32 %f148, %r200;
fma.rn.ftz.f32 %f149, %f147, %f148, 0f00000000;

dp4a.s32.s32 %r206, %r495, %r172, %r497;


dp4a.s32.s32 %r210, %r499, %r176, %r206;


dp4a.s32.s32 %r214, %r503, %r180, %r210;


dp4a.s32.s32 %r218, %r507, %r184, %r214;


{cvt.rn.f16x2.e4m3x2 %r222, %rs7;}


cvt.u16.u32 %rs16, %r222;

{cvt.f32.f16 %f87, %rs16;}


div.approx.ftz.f32 %f150, %f87, %f121;

{.reg .f16 low,high;
mov.b32 {low,high},%r187;
cvt.f32.f16 %f88, low;}


mul.ftz.f32 %f151, %f150, %f88;
cvt.rn.f32.s32 %f152, %r218;
fma.rn.ftz.f32 %f153, %f151, %f152, %f149;
add.ftz.f32 %f370, %f153, %f370;
cvt.u64.u32 %rd52, %r867;
add.s64 %rd53, %rd1, %rd52;
add.s64 %rd54, %rd53, %rd3;
mul.lo.s64 %rd55, %rd54, 36;
add.s64 %rd56, %rd11, %rd55;
ld.global.nc.u32 %r226, [%rd56+4];

dp4a.s32.s32 %r224, %r441, %r226, %r497;

ld.global.nc.u32 %r230, [%rd56+12];

dp4a.s32.s32 %r228, %r445, %r230, %r224;

ld.global.nc.u32 %r234, [%rd56+8];

dp4a.s32.s32 %r232, %r449, %r234, %r228;

ld.global.nc.u32 %r238, [%rd56+16];

dp4a.s32.s32 %r236, %r453, %r238, %r232;


{cvt.rn.f16x2.e4m3x2 %r240, %rs1;}


cvt.u16.u32 %rs18, %r240;

{cvt.f32.f16 %f89, %rs18;}


div.approx.ftz.f32 %f154, %f89, %f121;
ld.global.nc.u32 %r241, [%rd56];

{.reg .f16 low,high;
mov.b32 {low,high},%r241;
cvt.f32.f16 %f90, low;}


mul.ftz.f32 %f155, %f154, %f90;
cvt.rn.f32.s32 %f156, %r236;
fma.rn.ftz.f32 %f157, %f155, %f156, 0f00000000;
add.s64 %rd57, %rd53, %rd4;
mul.lo.s64 %rd58, %rd57, 36;
add.s64 %rd59, %rd11, %rd58;
ld.global.nc.u32 %r244, [%rd59+20];

dp4a.s32.s32 %r242, %r459, %r244, %r497;

ld.global.nc.u32 %r248, [%rd59+28];

dp4a.s32.s32 %r246, %r463, %r248, %r242;

ld.global.nc.u32 %r252, [%rd59+24];

dp4a.s32.s32 %r250, %r467, %r252, %r246;

ld.global.nc.u32 %r256, [%rd59+32];

dp4a.s32.s32 %r254, %r471, %r256, %r250;


{cvt.rn.f16x2.e4m3x2 %r258, %rs3;}


cvt.u16.u32 %rs20, %r258;

{cvt.f32.f16 %f91, %rs20;}


div.approx.ftz.f32 %f158, %f91, %f121;
ld.global.nc.u32 %r259, [%rd59];

{.reg .f16 low,high;
mov.b32 {low,high},%r259;
cvt.f32.f16 %f92, low;}


mul.ftz.f32 %f159, %f158, %f92;
cvt.rn.f32.s32 %f160, %r254;
fma.rn.ftz.f32 %f161, %f159, %f160, %f157;
add.ftz.f32 %f369, %f161, %f369;

dp4a.s32.s32 %r260, %r477, %r226, %r497;


dp4a.s32.s32 %r264, %r481, %r230, %r260;


dp4a.s32.s32 %r268, %r485, %r234, %r264;


dp4a.s32.s32 %r272, %r489, %r238, %r268;


{cvt.rn.f16x2.e4m3x2 %r276, %rs5;}


cvt.u16.u32 %rs22, %r276;

{cvt.f32.f16 %f93, %rs22;}


div.approx.ftz.f32 %f162, %f93, %f121;

{.reg .f16 low,high;
mov.b32 {low,high},%r241;
cvt.f32.f16 %f94, low;}


mul.ftz.f32 %f163, %f162, %f94;
cvt.rn.f32.s32 %f164, %r272;
fma.rn.ftz.f32 %f165, %f163, %f164, 0f00000000;

dp4a.s32.s32 %r278, %r495, %r244, %r497;


dp4a.s32.s32 %r282, %r499, %r248, %r278;


dp4a.s32.s32 %r286, %r503, %r252, %r282;


dp4a.s32.s32 %r290, %r507, %r256, %r286;


{cvt.rn.f16x2.e4m3x2 %r294, %rs7;}


cvt.u16.u32 %rs24, %r294;

{cvt.f32.f16 %f95, %rs24;}


div.approx.ftz.f32 %f166, %f95, %f121;

{.reg .f16 low,high;
mov.b32 {low,high},%r259;
cvt.f32.f16 %f96, low;}


mul.ftz.f32 %f167, %f166, %f96;
cvt.rn.f32.s32 %f168, %r290;
fma.rn.ftz.f32 %f169, %f167, %f168, %f165;
add.ftz.f32 %f368, %f169, %f368;
cvt.u64.u32 %rd60, %r866;
add.s64 %rd61, %rd1, %rd60;
add.s64 %rd62, %rd61, %rd3;
mul.lo.s64 %rd63, %rd62, 36;
add.s64 %rd64, %rd11, %rd63;
ld.global.nc.u32 %r298, [%rd64+4];

dp4a.s32.s32 %r296, %r441, %r298, %r497;

ld.global.nc.u32 %r302, [%rd64+12];

dp4a.s32.s32 %r300, %r445, %r302, %r296;

ld.global.nc.u32 %r306, [%rd64+8];

dp4a.s32.s32 %r304, %r449, %r306, %r300;

ld.global.nc.u32 %r310, [%rd64+16];

dp4a.s32.s32 %r308, %r453, %r310, %r304;


{cvt.rn.f16x2.e4m3x2 %r312, %rs1;}


cvt.u16.u32 %rs26, %r312;

{cvt.f32.f16 %f97, %rs26;}


div.approx.ftz.f32 %f170, %f97, %f121;
ld.global.nc.u32 %r313, [%rd64];

{.reg .f16 low,high;
mov.b32 {low,high},%r313;
cvt.f32.f16 %f98, low;}


mul.ftz.f32 %f171, %f170, %f98;
cvt.rn.f32.s32 %f172, %r308;
fma.rn.ftz.f32 %f173, %f171, %f172, 0f00000000;
add.s64 %rd65, %rd61, %rd4;
mul.lo.s64 %rd66, %rd65, 36;
add.s64 %rd67, %rd11, %rd66;
ld.global.nc.u32 %r316, [%rd67+20];

dp4a.s32.s32 %r314, %r459, %r316, %r497;

ld.global.nc.u32 %r320, [%rd67+28];

dp4a.s32.s32 %r318, %r463, %r320, %r314;

ld.global.nc.u32 %r324, [%rd67+24];

dp4a.s32.s32 %r322, %r467, %r324, %r318;

ld.global.nc.u32 %r328, [%rd67+32];

dp4a.s32.s32 %r326, %r471, %r328, %r322;


{cvt.rn.f16x2.e4m3x2 %r330, %rs3;}


cvt.u16.u32 %rs28, %r330;

{cvt.f32.f16 %f99, %rs28;}


div.approx.ftz.f32 %f174, %f99, %f121;
ld.global.nc.u32 %r331, [%rd67];

{.reg .f16 low,high;
mov.b32 {low,high},%r331;
cvt.f32.f16 %f100, low;}


mul.ftz.f32 %f175, %f174, %f100;
cvt.rn.f32.s32 %f176, %r326;
fma.rn.ftz.f32 %f177, %f175, %f176, %f173;
add.ftz.f32 %f367, %f177, %f367;

dp4a.s32.s32 %r332, %r477, %r298, %r497;


dp4a.s32.s32 %r336, %r481, %r302, %r332;


dp4a.s32.s32 %r340, %r485, %r306, %r336;


dp4a.s32.s32 %r344, %r489, %r310, %r340;


{cvt.rn.f16x2.e4m3x2 %r348, %rs5;}


cvt.u16.u32 %rs30, %r348;

{cvt.f32.f16 %f101, %rs30;}


div.approx.ftz.f32 %f178, %f101, %f121;

{.reg .f16 low,high;
mov.b32 {low,high},%r313;
cvt.f32.f16 %f102, low;}


mul.ftz.f32 %f179, %f178, %f102;
cvt.rn.f32.s32 %f180, %r344;
fma.rn.ftz.f32 %f181, %f179, %f180, 0f00000000;

dp4a.s32.s32 %r350, %r495, %r316, %r497;


dp4a.s32.s32 %r354, %r499, %r320, %r350;


dp4a.s32.s32 %r358, %r503, %r324, %r354;


dp4a.s32.s32 %r362, %r507, %r328, %r358;


{cvt.rn.f16x2.e4m3x2 %r366, %rs7;}


cvt.u16.u32 %rs32, %r366;

{cvt.f32.f16 %f103, %rs32;}


div.approx.ftz.f32 %f182, %f103, %f121;

{.reg .f16 low,high;
mov.b32 {low,high},%r331;
cvt.f32.f16 %f104, low;}


mul.ftz.f32 %f183, %f182, %f104;
cvt.rn.f32.s32 %f184, %r362;
fma.rn.ftz.f32 %f185, %f183, %f184, %f181;
add.ftz.f32 %f366, %f185, %f366;
cvt.u64.u32 %rd68, %r865;
add.s64 %rd69, %rd1, %rd68;
add.s64 %rd70, %rd69, %rd3;
mul.lo.s64 %rd71, %rd70, 36;
add.s64 %rd72, %rd11, %rd71;
ld.global.nc.u32 %r370, [%rd72+4];

dp4a.s32.s32 %r368, %r441, %r370, %r497;

ld.global.nc.u32 %r374, [%rd72+12];

dp4a.s32.s32 %r372, %r445, %r374, %r368;

ld.global.nc.u32 %r378, [%rd72+8];

dp4a.s32.s32 %r376, %r449, %r378, %r372;

ld.global.nc.u32 %r382, [%rd72+16];

dp4a.s32.s32 %r380, %r453, %r382, %r376;


{cvt.rn.f16x2.e4m3x2 %r384, %rs1;}


cvt.u16.u32 %rs34, %r384;

{cvt.f32.f16 %f105, %rs34;}


div.approx.ftz.f32 %f186, %f105, %f121;
ld.global.nc.u32 %r385, [%rd72];

{.reg .f16 low,high;
mov.b32 {low,high},%r385;
cvt.f32.f16 %f106, low;}


mul.ftz.f32 %f187, %f186, %f106;
cvt.rn.f32.s32 %f188, %r380;
fma.rn.ftz.f32 %f189, %f187, %f188, 0f00000000;
add.s64 %rd73, %rd69, %rd4;
mul.lo.s64 %rd74, %rd73, 36;
add.s64 %rd75, %rd11, %rd74;
ld.global.nc.u32 %r388, [%rd75+20];

dp4a.s32.s32 %r386, %r459, %r388, %r497;

ld.global.nc.u32 %r392, [%rd75+28];

dp4a.s32.s32 %r390, %r463, %r392, %r386;

ld.global.nc.u32 %r396, [%rd75+24];

dp4a.s32.s32 %r394, %r467, %r396, %r390;

ld.global.nc.u32 %r400, [%rd75+32];

dp4a.s32.s32 %r398, %r471, %r400, %r394;


{cvt.rn.f16x2.e4m3x2 %r402, %rs3;}


cvt.u16.u32 %rs36, %r402;

{cvt.f32.f16 %f107, %rs36;}


div.approx.ftz.f32 %f190, %f107, %f121;
ld.global.nc.u32 %r403, [%rd75];

{.reg .f16 low,high;
mov.b32 {low,high},%r403;
cvt.f32.f16 %f108, low;}


mul.ftz.f32 %f191, %f190, %f108;
cvt.rn.f32.s32 %f192, %r398;
fma.rn.ftz.f32 %f193, %f191, %f192, %f189;
add.ftz.f32 %f365, %f193, %f365;

dp4a.s32.s32 %r404, %r477, %r370, %r497;


dp4a.s32.s32 %r408, %r481, %r374, %r404;


dp4a.s32.s32 %r412, %r485, %r378, %r408;


dp4a.s32.s32 %r416, %r489, %r382, %r412;


{cvt.rn.f16x2.e4m3x2 %r420, %rs5;}


cvt.u16.u32 %rs38, %r420;

{cvt.f32.f16 %f109, %rs38;}


div.approx.ftz.f32 %f194, %f109, %f121;

{.reg .f16 low,high;
mov.b32 {low,high},%r385;
cvt.f32.f16 %f110, low;}


mul.ftz.f32 %f195, %f194, %f110;
cvt.rn.f32.s32 %f196, %r416;
fma.rn.ftz.f32 %f197, %f195, %f196, 0f00000000;

dp4a.s32.s32 %r422, %r495, %r388, %r497;


dp4a.s32.s32 %r426, %r499, %r392, %r422;


dp4a.s32.s32 %r430, %r503, %r396, %r426;


dp4a.s32.s32 %r434, %r507, %r400, %r430;


{cvt.rn.f16x2.e4m3x2 %r438, %rs7;}


cvt.u16.u32 %rs40, %r438;

{cvt.f32.f16 %f111, %rs40;}


div.approx.ftz.f32 %f198, %f111, %f121;

{.reg .f16 low,high;
mov.b32 {low,high},%r403;
cvt.f32.f16 %f112, low;}


mul.ftz.f32 %f199, %f198, %f112;
cvt.rn.f32.s32 %f200, %r434;
fma.rn.ftz.f32 %f201, %f199, %f200, %f197;
add.ftz.f32 %f364, %f201, %f364;
cvt.u64.u32 %rd76, %r864;
add.s64 %rd77, %rd1, %rd76;
add.s64 %rd78, %rd77, %rd3;
mul.lo.s64 %rd79, %rd78, 36;
add.s64 %rd80, %rd11, %rd79;
ld.global.nc.u32 %r442, [%rd80+4];

dp4a.s32.s32 %r440, %r441, %r442, %r497;

ld.global.nc.u32 %r446, [%rd80+12];

dp4a.s32.s32 %r444, %r445, %r446, %r440;

ld.global.nc.u32 %r450, [%rd80+8];

dp4a.s32.s32 %r448, %r449, %r450, %r444;

ld.global.nc.u32 %r454, [%rd80+16];

dp4a.s32.s32 %r452, %r453, %r454, %r448;


{cvt.rn.f16x2.e4m3x2 %r456, %rs1;}


cvt.u16.u32 %rs42, %r456;

{cvt.f32.f16 %f113, %rs42;}


div.approx.ftz.f32 %f202, %f113, %f121;
ld.global.nc.u32 %r457, [%rd80];

{.reg .f16 low,high;
mov.b32 {low,high},%r457;
cvt.f32.f16 %f114, low;}


mul.ftz.f32 %f203, %f202, %f114;
cvt.rn.f32.s32 %f204, %r452;
fma.rn.ftz.f32 %f205, %f203, %f204, 0f00000000;
add.s64 %rd81, %rd77, %rd4;
mul.lo.s64 %rd82, %rd81, 36;
add.s64 %rd83, %rd11, %rd82;
ld.global.nc.u32 %r460, [%rd83+20];

dp4a.s32.s32 %r458, %r459, %r460, %r497;

ld.global.nc.u32 %r464, [%rd83+28];

dp4a.s32.s32 %r462, %r463, %r464, %r458;

ld.global.nc.u32 %r468, [%rd83+24];

dp4a.s32.s32 %r466, %r467, %r468, %r462;

ld.global.nc.u32 %r472, [%rd83+32];

dp4a.s32.s32 %r470, %r471, %r472, %r466;


{cvt.rn.f16x2.e4m3x2 %r474, %rs3;}


cvt.u16.u32 %rs44, %r474;

{cvt.f32.f16 %f115, %rs44;}


div.approx.ftz.f32 %f206, %f115, %f121;
ld.global.nc.u32 %r475, [%rd83];

{.reg .f16 low,high;
mov.b32 {low,high},%r475;
cvt.f32.f16 %f116, low;}


mul.ftz.f32 %f207, %f206, %f116;
cvt.rn.f32.s32 %f208, %r470;
fma.rn.ftz.f32 %f209, %f207, %f208, %f205;
add.ftz.f32 %f363, %f209, %f363;

dp4a.s32.s32 %r476, %r477, %r442, %r497;


dp4a.s32.s32 %r480, %r481, %r446, %r476;


dp4a.s32.s32 %r484, %r485, %r450, %r480;


dp4a.s32.s32 %r488, %r489, %r454, %r484;


{cvt.rn.f16x2.e4m3x2 %r492, %rs5;}


cvt.u16.u32 %rs46, %r492;

{cvt.f32.f16 %f117, %rs46;}


div.approx.ftz.f32 %f210, %f117, %f121;

{.reg .f16 low,high;
mov.b32 {low,high},%r457;
cvt.f32.f16 %f118, low;}


mul.ftz.f32 %f211, %f210, %f118;
cvt.rn.f32.s32 %f212, %r488;
fma.rn.ftz.f32 %f213, %f211, %f212, 0f00000000;

dp4a.s32.s32 %r494, %r495, %r460, %r497;


dp4a.s32.s32 %r498, %r499, %r464, %r494;


dp4a.s32.s32 %r502, %r503, %r468, %r498;


dp4a.s32.s32 %r506, %r507, %r472, %r502;


{cvt.rn.f16x2.e4m3x2 %r510, %rs7;}


cvt.u16.u32 %rs48, %r510;

{cvt.f32.f16 %f119, %rs48;}


div.approx.ftz.f32 %f214, %f119, %f121;

{.reg .f16 low,high;
mov.b32 {low,high},%r475;
cvt.f32.f16 %f120, low;}


mul.ftz.f32 %f215, %f214, %f120;
cvt.rn.f32.s32 %f216, %r506;
fma.rn.ftz.f32 %f217, %f215, %f216, %f213;
add.ftz.f32 %f362, %f217, %f362;
add.s32 %r867, %r867, 64;
add.s32 %r866, %r866, 64;
add.s32 %r865, %r865, 64;
add.s32 %r864, %r864, 64;
add.s32 %r863, %r863, 64;
add.s32 %r862, %r862, 64;
add.s64 %rd124, %rd124, 1152;
add.s32 %r868, %r868, 32;
setp.lt.s32 %p8, %r868, %r49;
@%p8 bra $L__BB124_2;

$L__BB124_3:
setp.eq.s32 %p9, %r1, 0;
@%p9 bra $L__BB124_5;

mov.u32 %r639, _ZZ13mul_mat_vec_qIL9ggml_type40ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r640, %r1, 1536, %r639;
shl.b32 %r641, %r2, 2;
add.s32 %r642, %r640, %r641;
st.shared.f32 [%r642+-1536], %f373;
st.shared.f32 [%r642+-1408], %f372;
st.shared.f32 [%r642+-1280], %f371;
st.shared.f32 [%r642+-1152], %f370;
st.shared.f32 [%r642+-1024], %f369;
st.shared.f32 [%r642+-896], %f368;
st.shared.f32 [%r642+-768], %f367;
st.shared.f32 [%r642+-640], %f366;
st.shared.f32 [%r642+-512], %f365;
st.shared.f32 [%r642+-384], %f364;
st.shared.f32 [%r642+-256], %f363;
st.shared.f32 [%r642+-128], %f362;

$L__BB124_5:
setp.ne.s32 %p10, %r1, 0;
bar.sync 0;
@%p10 bra $L__BB124_30;

mov.u32 %r643, %ctaid.z;
mov.u32 %r644, %ctaid.y;
mad.lo.s32 %r645, %r644, %r38, %r4;
mad.lo.s32 %r646, %r643, %r44, %r645;
cvt.u64.u32 %rd14, %r646;
shl.b32 %r647, %r2, 2;
mov.u32 %r648, _ZZ13mul_mat_vec_qIL9ggml_type40ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r26, %r648, %r647;
mov.u32 %r649, 2;
ld.shared.f32 %f218, [%r26];
add.ftz.f32 %f219, %f218, %f373;
mov.b32 %r650, %f219;
mov.u32 %r651, 31;
mov.u32 %r652, 16;
mov.u32 %r653, -1;
shfl.sync.bfly.b32 %r654|%p11, %r650, %r652, %r651, %r653;
mov.b32 %f220, %r654;
add.ftz.f32 %f221, %f219, %f220;
mov.b32 %r655, %f221;
mov.u32 %r656, 8;
shfl.sync.bfly.b32 %r657|%p12, %r655, %r656, %r651, %r653;
mov.b32 %f222, %r657;
add.ftz.f32 %f223, %f221, %f222;
mov.b32 %r658, %f223;
mov.u32 %r659, 4;
shfl.sync.bfly.b32 %r660|%p13, %r658, %r659, %r651, %r653;
mov.b32 %f224, %r660;
add.ftz.f32 %f225, %f223, %f224;
mov.b32 %r661, %f225;
shfl.sync.bfly.b32 %r662|%p14, %r661, %r649, %r651, %r653;
mov.b32 %f226, %r662;
add.ftz.f32 %f227, %f225, %f226;
mov.b32 %r663, %f227;
mov.u32 %r664, 1;
shfl.sync.bfly.b32 %r665|%p15, %r663, %r664, %r651, %r653;
mov.b32 %f228, %r665;
add.ftz.f32 %f37, %f227, %f228;
setp.lt.u32 %p16, %r4, %r32;
setp.eq.s32 %p17, %r2, 0;
and.pred %p1, %p17, %p16;
cvta.to.global.u64 %rd84, %rd20;
mul.wide.u32 %rd85, %r646, 4;
add.s64 %rd15, %rd84, %rd85;
not.pred %p18, %p1;
@%p18 bra $L__BB124_8;

st.global.f32 [%rd15], %f37;

$L__BB124_8:
ld.shared.f32 %f229, [%r26+128];
add.ftz.f32 %f230, %f229, %f372;
mov.b32 %r666, %f230;
shfl.sync.bfly.b32 %r670|%p19, %r666, %r652, %r651, %r653;
mov.b32 %f231, %r670;
add.ftz.f32 %f232, %f230, %f231;
mov.b32 %r671, %f232;
shfl.sync.bfly.b32 %r673|%p20, %r671, %r656, %r651, %r653;
mov.b32 %f233, %r673;
add.ftz.f32 %f234, %f232, %f233;
mov.b32 %r674, %f234;
shfl.sync.bfly.b32 %r676|%p21, %r674, %r659, %r651, %r653;
mov.b32 %f235, %r676;
add.ftz.f32 %f236, %f234, %f235;
mov.b32 %r677, %f236;
shfl.sync.bfly.b32 %r679|%p22, %r677, %r649, %r651, %r653;
mov.b32 %f237, %r679;
add.ftz.f32 %f238, %f236, %f237;
mov.b32 %r680, %f238;
shfl.sync.bfly.b32 %r682|%p23, %r680, %r664, %r651, %r653;
mov.b32 %f239, %r682;
add.ftz.f32 %f38, %f238, %f239;
setp.eq.s32 %p24, %r2, 1;
or.b32 %r686, %r4, 1;
setp.lt.u32 %p25, %r686, %r32;
and.pred %p2, %p24, %p25;
not.pred %p26, %p2;
@%p26 bra $L__BB124_10;

st.global.f32 [%rd15+4], %f38;

$L__BB124_10:
ld.shared.f32 %f240, [%r26+256];
add.ftz.f32 %f241, %f240, %f371;
mov.b32 %r687, %f241;
mov.u32 %r688, 31;
mov.u32 %r689, 16;
mov.u32 %r690, -1;
shfl.sync.bfly.b32 %r691|%p27, %r687, %r689, %r688, %r690;
mov.b32 %f242, %r691;
add.ftz.f32 %f243, %f241, %f242;
mov.b32 %r692, %f243;
mov.u32 %r693, 8;
shfl.sync.bfly.b32 %r694|%p28, %r692, %r693, %r688, %r690;
mov.b32 %f244, %r694;
add.ftz.f32 %f245, %f243, %f244;
mov.b32 %r695, %f245;
mov.u32 %r696, 4;
shfl.sync.bfly.b32 %r697|%p29, %r695, %r696, %r688, %r690;
mov.b32 %f246, %r697;
add.ftz.f32 %f247, %f245, %f246;
mov.b32 %r698, %f247;
mov.u32 %r699, 2;
shfl.sync.bfly.b32 %r700|%p30, %r698, %r699, %r688, %r690;
mov.b32 %f248, %r700;
add.ftz.f32 %f249, %f247, %f248;
mov.b32 %r701, %f249;
mov.u32 %r702, 1;
shfl.sync.bfly.b32 %r703|%p31, %r701, %r702, %r688, %r690;
mov.b32 %f250, %r703;
add.ftz.f32 %f39, %f249, %f250;
@%p18 bra $L__BB124_12;

cvt.u64.u32 %rd86, %r32;
add.s64 %rd87, %rd86, %rd14;
shl.b64 %rd89, %rd87, 2;
add.s64 %rd90, %rd84, %rd89;
st.global.f32 [%rd90], %f39;

$L__BB124_12:
ld.shared.f32 %f251, [%r26+384];
add.ftz.f32 %f252, %f251, %f370;
mov.b32 %r704, %f252;
shfl.sync.bfly.b32 %r708|%p33, %r704, %r689, %r688, %r690;
mov.b32 %f253, %r708;
add.ftz.f32 %f254, %f252, %f253;
mov.b32 %r709, %f254;
shfl.sync.bfly.b32 %r711|%p34, %r709, %r693, %r688, %r690;
mov.b32 %f255, %r711;
add.ftz.f32 %f256, %f254, %f255;
mov.b32 %r712, %f256;
shfl.sync.bfly.b32 %r714|%p35, %r712, %r696, %r688, %r690;
mov.b32 %f257, %r714;
add.ftz.f32 %f258, %f256, %f257;
mov.b32 %r715, %f258;
shfl.sync.bfly.b32 %r717|%p36, %r715, %r699, %r688, %r690;
mov.b32 %f259, %r717;
add.ftz.f32 %f260, %f258, %f259;
mov.b32 %r718, %f260;
shfl.sync.bfly.b32 %r720|%p37, %r718, %r702, %r688, %r690;
mov.b32 %f261, %r720;
add.ftz.f32 %f40, %f260, %f261;
@%p26 bra $L__BB124_14;

add.s32 %r721, %r32, 1;
cvt.u64.u32 %rd91, %r721;
add.s64 %rd92, %rd91, %rd14;
shl.b64 %rd94, %rd92, 2;
add.s64 %rd95, %rd84, %rd94;
st.global.f32 [%rd95], %f40;

$L__BB124_14:
shl.b32 %r722, %r32, 1;
mov.u32 %r723, 1;
ld.shared.f32 %f262, [%r26+512];
add.ftz.f32 %f263, %f262, %f369;
mov.b32 %r724, %f263;
mov.u32 %r725, 31;
mov.u32 %r726, 16;
mov.u32 %r727, -1;
shfl.sync.bfly.b32 %r728|%p39, %r724, %r726, %r725, %r727;
mov.b32 %f264, %r728;
add.ftz.f32 %f265, %f263, %f264;
mov.b32 %r729, %f265;
mov.u32 %r730, 8;
shfl.sync.bfly.b32 %r731|%p40, %r729, %r730, %r725, %r727;
mov.b32 %f266, %r731;
add.ftz.f32 %f267, %f265, %f266;
mov.b32 %r732, %f267;
mov.u32 %r733, 4;
shfl.sync.bfly.b32 %r734|%p41, %r732, %r733, %r725, %r727;
mov.b32 %f268, %r734;
add.ftz.f32 %f269, %f267, %f268;
mov.b32 %r735, %f269;
mov.u32 %r736, 2;
shfl.sync.bfly.b32 %r737|%p42, %r735, %r736, %r725, %r727;
mov.b32 %f270, %r737;
add.ftz.f32 %f271, %f269, %f270;
mov.b32 %r738, %f271;
shfl.sync.bfly.b32 %r739|%p43, %r738, %r723, %r725, %r727;
mov.b32 %f272, %r739;
add.ftz.f32 %f41, %f271, %f272;
cvt.u64.u32 %rd96, %r722;
add.s64 %rd97, %rd96, %rd14;
shl.b64 %rd99, %rd97, 2;
add.s64 %rd16, %rd84, %rd99;
@%p18 bra $L__BB124_16;

st.global.f32 [%rd16], %f41;

$L__BB124_16:
ld.shared.f32 %f273, [%r26+640];
add.ftz.f32 %f274, %f273, %f368;
mov.b32 %r740, %f274;
shfl.sync.bfly.b32 %r744|%p45, %r740, %r726, %r725, %r727;
mov.b32 %f275, %r744;
add.ftz.f32 %f276, %f274, %f275;
mov.b32 %r745, %f276;
shfl.sync.bfly.b32 %r747|%p46, %r745, %r730, %r725, %r727;
mov.b32 %f277, %r747;
add.ftz.f32 %f278, %f276, %f277;
mov.b32 %r748, %f278;
shfl.sync.bfly.b32 %r750|%p47, %r748, %r733, %r725, %r727;
mov.b32 %f279, %r750;
add.ftz.f32 %f280, %f278, %f279;
mov.b32 %r751, %f280;
shfl.sync.bfly.b32 %r753|%p48, %r751, %r736, %r725, %r727;
mov.b32 %f281, %r753;
add.ftz.f32 %f282, %f280, %f281;
mov.b32 %r754, %f282;
shfl.sync.bfly.b32 %r756|%p49, %r754, %r723, %r725, %r727;
mov.b32 %f283, %r756;
add.ftz.f32 %f42, %f282, %f283;
@%p26 bra $L__BB124_18;

st.global.f32 [%rd16+4], %f42;

$L__BB124_18:
mul.lo.s32 %r27, %r32, 3;
ld.shared.f32 %f284, [%r26+768];
add.ftz.f32 %f285, %f284, %f367;
mov.b32 %r757, %f285;
mov.u32 %r758, 31;
mov.u32 %r759, 16;
mov.u32 %r760, -1;
shfl.sync.bfly.b32 %r761|%p51, %r757, %r759, %r758, %r760;
mov.b32 %f286, %r761;
add.ftz.f32 %f287, %f285, %f286;
mov.b32 %r762, %f287;
mov.u32 %r763, 8;
shfl.sync.bfly.b32 %r764|%p52, %r762, %r763, %r758, %r760;
mov.b32 %f288, %r764;
add.ftz.f32 %f289, %f287, %f288;
mov.b32 %r765, %f289;
mov.u32 %r766, 4;
shfl.sync.bfly.b32 %r767|%p53, %r765, %r766, %r758, %r760;
mov.b32 %f290, %r767;
add.ftz.f32 %f291, %f289, %f290;
mov.b32 %r768, %f291;
mov.u32 %r769, 2;
shfl.sync.bfly.b32 %r770|%p54, %r768, %r769, %r758, %r760;
mov.b32 %f292, %r770;
add.ftz.f32 %f293, %f291, %f292;
mov.b32 %r771, %f293;
mov.u32 %r772, 1;
shfl.sync.bfly.b32 %r773|%p55, %r771, %r772, %r758, %r760;
mov.b32 %f294, %r773;
add.ftz.f32 %f43, %f293, %f294;
@%p18 bra $L__BB124_20;

cvt.u64.u32 %rd100, %r27;
add.s64 %rd101, %rd100, %rd14;
shl.b64 %rd103, %rd101, 2;
add.s64 %rd104, %rd84, %rd103;
st.global.f32 [%rd104], %f43;

$L__BB124_20:
ld.shared.f32 %f295, [%r26+896];
add.ftz.f32 %f296, %f295, %f366;
mov.b32 %r774, %f296;
shfl.sync.bfly.b32 %r778|%p57, %r774, %r759, %r758, %r760;
mov.b32 %f297, %r778;
add.ftz.f32 %f298, %f296, %f297;
mov.b32 %r779, %f298;
shfl.sync.bfly.b32 %r781|%p58, %r779, %r763, %r758, %r760;
mov.b32 %f299, %r781;
add.ftz.f32 %f300, %f298, %f299;
mov.b32 %r782, %f300;
shfl.sync.bfly.b32 %r784|%p59, %r782, %r766, %r758, %r760;
mov.b32 %f301, %r784;
add.ftz.f32 %f302, %f300, %f301;
mov.b32 %r785, %f302;
shfl.sync.bfly.b32 %r787|%p60, %r785, %r769, %r758, %r760;
mov.b32 %f303, %r787;
add.ftz.f32 %f304, %f302, %f303;
mov.b32 %r788, %f304;
shfl.sync.bfly.b32 %r790|%p61, %r788, %r772, %r758, %r760;
mov.b32 %f305, %r790;
add.ftz.f32 %f44, %f304, %f305;
@%p26 bra $L__BB124_22;

add.s32 %r791, %r27, 1;
cvt.u64.u32 %rd105, %r791;
add.s64 %rd106, %rd105, %rd14;
shl.b64 %rd108, %rd106, 2;
add.s64 %rd109, %rd84, %rd108;
st.global.f32 [%rd109], %f44;

$L__BB124_22:
shl.b32 %r792, %r32, 2;
mov.u32 %r793, 2;
ld.shared.f32 %f306, [%r26+1024];
add.ftz.f32 %f307, %f306, %f365;
mov.b32 %r794, %f307;
mov.u32 %r795, 31;
mov.u32 %r796, 16;
mov.u32 %r797, -1;
shfl.sync.bfly.b32 %r798|%p63, %r794, %r796, %r795, %r797;
mov.b32 %f308, %r798;
add.ftz.f32 %f309, %f307, %f308;
mov.b32 %r799, %f309;
mov.u32 %r800, 8;
shfl.sync.bfly.b32 %r801|%p64, %r799, %r800, %r795, %r797;
mov.b32 %f310, %r801;
add.ftz.f32 %f311, %f309, %f310;
mov.b32 %r802, %f311;
mov.u32 %r803, 4;
shfl.sync.bfly.b32 %r804|%p65, %r802, %r803, %r795, %r797;
mov.b32 %f312, %r804;
add.ftz.f32 %f313, %f311, %f312;
mov.b32 %r805, %f313;
shfl.sync.bfly.b32 %r806|%p66, %r805, %r793, %r795, %r797;
mov.b32 %f314, %r806;
add.ftz.f32 %f315, %f313, %f314;
mov.b32 %r807, %f315;
mov.u32 %r808, 1;
shfl.sync.bfly.b32 %r809|%p67, %r807, %r808, %r795, %r797;
mov.b32 %f316, %r809;
add.ftz.f32 %f45, %f315, %f316;
cvt.u64.u32 %rd110, %r792;
add.s64 %rd111, %rd110, %rd14;
shl.b64 %rd113, %rd111, 2;
add.s64 %rd17, %rd84, %rd113;
@%p18 bra $L__BB124_24;

st.global.f32 [%rd17], %f45;

$L__BB124_24:
ld.shared.f32 %f317, [%r26+1152];
add.ftz.f32 %f318, %f317, %f364;
mov.b32 %r810, %f318;
shfl.sync.bfly.b32 %r814|%p69, %r810, %r796, %r795, %r797;
mov.b32 %f319, %r814;
add.ftz.f32 %f320, %f318, %f319;
mov.b32 %r815, %f320;
shfl.sync.bfly.b32 %r817|%p70, %r815, %r800, %r795, %r797;
mov.b32 %f321, %r817;
add.ftz.f32 %f322, %f320, %f321;
mov.b32 %r818, %f322;
shfl.sync.bfly.b32 %r820|%p71, %r818, %r803, %r795, %r797;
mov.b32 %f323, %r820;
add.ftz.f32 %f324, %f322, %f323;
mov.b32 %r821, %f324;
shfl.sync.bfly.b32 %r823|%p72, %r821, %r793, %r795, %r797;
mov.b32 %f325, %r823;
add.ftz.f32 %f326, %f324, %f325;
mov.b32 %r824, %f326;
shfl.sync.bfly.b32 %r826|%p73, %r824, %r808, %r795, %r797;
mov.b32 %f327, %r826;
add.ftz.f32 %f46, %f326, %f327;
@%p26 bra $L__BB124_26;

st.global.f32 [%rd17+4], %f46;

$L__BB124_26:
mul.lo.s32 %r28, %r32, 5;
ld.shared.f32 %f328, [%r26+1280];
add.ftz.f32 %f329, %f328, %f363;
mov.b32 %r827, %f329;
mov.u32 %r828, 31;
mov.u32 %r829, 16;
mov.u32 %r830, -1;
shfl.sync.bfly.b32 %r831|%p75, %r827, %r829, %r828, %r830;
mov.b32 %f330, %r831;
add.ftz.f32 %f331, %f329, %f330;
mov.b32 %r832, %f331;
mov.u32 %r833, 8;
shfl.sync.bfly.b32 %r834|%p76, %r832, %r833, %r828, %r830;
mov.b32 %f332, %r834;
add.ftz.f32 %f333, %f331, %f332;
mov.b32 %r835, %f333;
mov.u32 %r836, 4;
shfl.sync.bfly.b32 %r837|%p77, %r835, %r836, %r828, %r830;
mov.b32 %f334, %r837;
add.ftz.f32 %f335, %f333, %f334;
mov.b32 %r838, %f335;
mov.u32 %r839, 2;
shfl.sync.bfly.b32 %r840|%p78, %r838, %r839, %r828, %r830;
mov.b32 %f336, %r840;
add.ftz.f32 %f337, %f335, %f336;
mov.b32 %r841, %f337;
mov.u32 %r842, 1;
shfl.sync.bfly.b32 %r843|%p79, %r841, %r842, %r828, %r830;
mov.b32 %f338, %r843;
add.ftz.f32 %f47, %f337, %f338;
@%p18 bra $L__BB124_28;

cvt.u64.u32 %rd114, %r28;
add.s64 %rd115, %rd114, %rd14;
shl.b64 %rd117, %rd115, 2;
add.s64 %rd118, %rd84, %rd117;
st.global.f32 [%rd118], %f47;

$L__BB124_28:
ld.shared.f32 %f339, [%r26+1408];
add.ftz.f32 %f340, %f339, %f362;
mov.b32 %r844, %f340;
shfl.sync.bfly.b32 %r848|%p81, %r844, %r829, %r828, %r830;
mov.b32 %f341, %r848;
add.ftz.f32 %f342, %f340, %f341;
mov.b32 %r849, %f342;
shfl.sync.bfly.b32 %r851|%p82, %r849, %r833, %r828, %r830;
mov.b32 %f343, %r851;
add.ftz.f32 %f344, %f342, %f343;
mov.b32 %r852, %f344;
shfl.sync.bfly.b32 %r854|%p83, %r852, %r836, %r828, %r830;
mov.b32 %f345, %r854;
add.ftz.f32 %f346, %f344, %f345;
mov.b32 %r855, %f346;
shfl.sync.bfly.b32 %r857|%p84, %r855, %r839, %r828, %r830;
mov.b32 %f347, %r857;
add.ftz.f32 %f348, %f346, %f347;
mov.b32 %r858, %f348;
shfl.sync.bfly.b32 %r860|%p85, %r858, %r842, %r828, %r830;
mov.b32 %f349, %r860;
add.ftz.f32 %f48, %f348, %f349;
@%p26 bra $L__BB124_30;

add.s32 %r861, %r28, 1;
cvt.u64.u32 %rd119, %r861;
add.s64 %rd120, %rd119, %rd14;
shl.b64 %rd122, %rd120, 2;
add.s64 %rd123, %rd84, %rd122;
st.global.f32 [%rd123], %f48;

$L__BB124_30:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type40ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type40ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type40ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type40ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type40ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type40ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type40ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type40ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type40ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<99>;
.reg .b16 %rs<69>;
.reg .f32 %f<436>;
.reg .b32 %r<1045>;
.reg .b64 %rd<141>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type40ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1792];

ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type40ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type40ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd20, [_Z13mul_mat_vec_qIL9ggml_type40ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type40ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type40ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type40ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type40ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type40ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type40ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type40ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type40ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type40ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r45, [_Z13mul_mat_vec_qIL9ggml_type40ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type40ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type40ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type40ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type40ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r46, %r1, 5;
mov.u32 %r47, %tid.x;
add.s32 %r2, %r46, %r47;
mov.u32 %r48, %ctaid.x;
shl.b32 %r3, %r48, 1;
shr.u32 %r49, %r2, 31;
add.s32 %r50, %r2, %r49;
shr.s32 %r1044, %r50, 1;
shr.u32 %r51, %r30, 6;
setp.ge.s32 %p3, %r1044, %r51;
mov.f32 %f422, 0f00000000;
mov.f32 %f423, %f422;
mov.f32 %f424, %f422;
mov.f32 %f425, %f422;
mov.f32 %f426, %f422;
mov.f32 %f427, %f422;
mov.f32 %f428, %f422;
mov.f32 %f429, %f422;
mov.f32 %f430, %f422;
mov.f32 %f431, %f422;
mov.f32 %f432, %f422;
mov.f32 %f433, %f422;
mov.f32 %f434, %f422;
mov.f32 %f435, %f422;
@%p3 bra $L__BB125_3;

mov.u32 %r52, %ctaid.y;
mov.u32 %r53, %ctaid.z;
mul.lo.s32 %r54, %r52, %r38;
cvt.u64.u32 %rd21, %r54;
mul.lo.s32 %r55, %r53, %r44;
cvt.u64.u32 %rd22, %r55;
add.s64 %rd1, %rd22, %rd21;
mul.hi.u32 %r56, %r53, %r40;
add.s32 %r57, %r56, %r53;
shr.u32 %r58, %r57, %r41;
mul.lo.s32 %r59, %r58, %r43;
mul.hi.u32 %r60, %r52, %r34;
add.s32 %r61, %r60, %r52;
shr.u32 %r62, %r61, %r35;
mul.lo.s32 %r63, %r62, %r37;
cvta.to.global.u64 %rd140, %rd18;
and.b32 %r66, %r50, -2;
sub.s32 %r67, %r2, %r66;
shl.b32 %r68, %r67, 2;
shl.b32 %r69, %r67, 1;
cvt.s64.s32 %rd3, %r67;
cvt.s64.s32 %rd23, %r69;
or.b32 %r70, %r68, 2;
shr.s32 %r71, %r70, 1;
shr.s32 %r72, %r70, 2;
cvt.s64.s32 %rd4, %r72;
cvt.s64.s32 %rd24, %r71;
shl.b32 %r73, %r32, 1;
shl.b32 %r1037, %r1044, 1;
add.s32 %r1043, %r1037, %r73;
mad.lo.s32 %r1042, %r32, 3, %r1037;
shl.b32 %r74, %r32, 2;
add.s32 %r1041, %r1037, %r74;
mad.lo.s32 %r1040, %r32, 5, %r1037;
mad.lo.s32 %r1039, %r32, 6, %r1037;
add.s32 %r1038, %r32, %r1037;
add.s32 %r75, %r1044, %r59;
add.s32 %r76, %r75, %r63;
add.s32 %r77, %r3, 1;
mad.lo.s32 %r78, %r31, %r77, %r76;
mul.wide.s32 %rd25, %r78, 36;
add.s64 %rd5, %rd25, %rd24;
add.s64 %rd6, %rd25, %rd23;
mad.lo.s32 %r79, %r3, %r31, %r1044;
add.s32 %r80, %r79, %r59;
add.s32 %r81, %r80, %r63;
mul.wide.s32 %rd26, %r81, 36;
mul.wide.s32 %rd27, %r68, 4;
add.s64 %rd28, %rd26, %rd27;
add.s64 %rd7, %rd28, 16;
add.s64 %rd8, %rd26, %rd23;
add.s64 %rd29, %rd25, %rd27;
add.s64 %rd9, %rd29, 8;
add.s64 %rd10, %rd26, %rd24;
cvta.to.global.u64 %rd11, %rd19;
mov.f32 %f422, 0f00000000;

$L__BB125_2:
cvt.u64.u32 %rd30, %r1037;
add.s64 %rd31, %rd1, %rd30;
add.s64 %rd32, %rd140, %rd7;
ld.global.nc.u32 %r586, [%rd32+-12];
and.b32 %r587, %r586, -2004318072;
shr.u32 %r588, %r587, 1;
or.b32 %r589, %r588, 839922192;
and.b32 %r590, %r586, 30583;
mov.u32 %r591, 201852420;
mov.u32 %r592, 50462976;
prmt.b32 %r593, %r592, %r591, %r590;
mov.u32 %r594, -185009412;
mov.u32 %r595, -33620224;
prmt.b32 %r596, %r595, %r594, %r590;
and.b32 %r597, %r589, 30292;
prmt.b32 %r598, %r593, %r596, %r597;
shr.u32 %r599, %r586, 16;
and.b32 %r600, %r599, 30583;
prmt.b32 %r601, %r592, %r591, %r600;
prmt.b32 %r602, %r595, %r594, %r600;
shr.u32 %r603, %r589, 16;
prmt.b32 %r604, %r601, %r602, %r603;
mov.u32 %r605, 25632;
prmt.b32 %r515, %r598, %r604, %r605;
mov.u32 %r606, 30001;
prmt.b32 %r519, %r598, %r604, %r606;
ld.global.nc.u32 %r607, [%rd32+-8];
and.b32 %r608, %r607, -2004318072;
shr.u32 %r609, %r608, 1;
or.b32 %r610, %r609, 839922192;
and.b32 %r611, %r607, 30583;
prmt.b32 %r612, %r592, %r591, %r611;
prmt.b32 %r613, %r595, %r594, %r611;
and.b32 %r614, %r610, 30292;
prmt.b32 %r615, %r612, %r613, %r614;
shr.u32 %r616, %r607, 16;
and.b32 %r617, %r616, 30583;
prmt.b32 %r618, %r592, %r591, %r617;
prmt.b32 %r619, %r595, %r594, %r617;
shr.u32 %r620, %r610, 16;
prmt.b32 %r621, %r618, %r619, %r620;
prmt.b32 %r523, %r615, %r621, %r605;
prmt.b32 %r527, %r615, %r621, %r606;
add.s64 %rd33, %rd31, %rd3;
mul.lo.s64 %rd34, %rd33, 36;
add.s64 %rd35, %rd11, %rd34;
ld.global.nc.u32 %r84, [%rd35+4];
mov.u32 %r571, 0;

dp4a.s32.s32 %r82, %r515, %r84, %r571;

ld.global.nc.u32 %r88, [%rd35+12];

dp4a.s32.s32 %r86, %r519, %r88, %r82;

ld.global.nc.u32 %r92, [%rd35+8];

dp4a.s32.s32 %r90, %r523, %r92, %r86;

ld.global.nc.u32 %r96, [%rd35+16];

dp4a.s32.s32 %r94, %r527, %r96, %r90;

add.s64 %rd36, %rd140, %rd8;
ld.global.nc.u8 %rs57, [%rd36];
and.b16 %rs58, %rs57, 127;
setp.eq.s16 %p4, %rs58, 127;
selp.b16 %rs59, 0, %rs57, %p4;
and.b16 %rs1, %rs59, 255;

{cvt.rn.f16x2.e4m3x2 %r98, %rs1;}


cvt.u16.u32 %rs2, %r98;

{cvt.f32.f16 %f85, %rs2;}


mov.f32 %f141, 0f40000000;
div.approx.ftz.f32 %f142, %f85, %f141;
ld.global.nc.u32 %r99, [%rd35];

{.reg .f16 low,high;
mov.b32 {low,high},%r99;
cvt.f32.f16 %f86, low;}


mul.ftz.f32 %f143, %f142, %f86;
cvt.rn.f32.s32 %f144, %r94;
fma.rn.ftz.f32 %f145, %f143, %f144, 0f00000000;
ld.global.nc.u32 %r622, [%rd32+-4];
and.b32 %r623, %r622, -2004318072;
shr.u32 %r624, %r623, 1;
or.b32 %r625, %r624, 839922192;
and.b32 %r626, %r622, 30583;
prmt.b32 %r627, %r592, %r591, %r626;
prmt.b32 %r628, %r595, %r594, %r626;
and.b32 %r629, %r625, 30292;
prmt.b32 %r630, %r627, %r628, %r629;
shr.u32 %r631, %r622, 16;
and.b32 %r632, %r631, 30583;
prmt.b32 %r633, %r592, %r591, %r632;
prmt.b32 %r634, %r595, %r594, %r632;
shr.u32 %r635, %r625, 16;
prmt.b32 %r636, %r633, %r634, %r635;
prmt.b32 %r533, %r630, %r636, %r605;
prmt.b32 %r537, %r630, %r636, %r606;
ld.global.nc.u32 %r637, [%rd32];
and.b32 %r638, %r637, -2004318072;
shr.u32 %r639, %r638, 1;
or.b32 %r640, %r639, 839922192;
and.b32 %r641, %r637, 30583;
prmt.b32 %r642, %r592, %r591, %r641;
prmt.b32 %r643, %r595, %r594, %r641;
and.b32 %r644, %r640, 30292;
prmt.b32 %r645, %r642, %r643, %r644;
shr.u32 %r646, %r637, 16;
and.b32 %r647, %r646, 30583;
prmt.b32 %r648, %r592, %r591, %r647;
prmt.b32 %r649, %r595, %r594, %r647;
shr.u32 %r650, %r640, 16;
prmt.b32 %r651, %r648, %r649, %r650;
prmt.b32 %r541, %r645, %r651, %r605;
prmt.b32 %r545, %r645, %r651, %r606;
add.s64 %rd37, %rd31, %rd4;
mul.lo.s64 %rd38, %rd37, 36;
add.s64 %rd39, %rd11, %rd38;
ld.global.nc.u32 %r102, [%rd39+20];

dp4a.s32.s32 %r100, %r533, %r102, %r571;

ld.global.nc.u32 %r106, [%rd39+28];

dp4a.s32.s32 %r104, %r537, %r106, %r100;

ld.global.nc.u32 %r110, [%rd39+24];

dp4a.s32.s32 %r108, %r541, %r110, %r104;

ld.global.nc.u32 %r114, [%rd39+32];

dp4a.s32.s32 %r112, %r545, %r114, %r108;

add.s64 %rd40, %rd140, %rd10;
ld.global.nc.u8 %rs60, [%rd40];
and.b16 %rs61, %rs60, 127;
setp.eq.s16 %p5, %rs61, 127;
selp.b16 %rs62, 0, %rs60, %p5;
and.b16 %rs3, %rs62, 255;

{cvt.rn.f16x2.e4m3x2 %r116, %rs3;}


cvt.u16.u32 %rs4, %r116;

{cvt.f32.f16 %f87, %rs4;}


div.approx.ftz.f32 %f146, %f87, %f141;
ld.global.nc.u32 %r117, [%rd39];

{.reg .f16 low,high;
mov.b32 {low,high},%r117;
cvt.f32.f16 %f88, low;}


mul.ftz.f32 %f147, %f146, %f88;
cvt.rn.f32.s32 %f148, %r112;
fma.rn.ftz.f32 %f149, %f147, %f148, %f145;
add.ftz.f32 %f435, %f149, %f435;
add.s64 %rd41, %rd140, %rd9;
ld.global.nc.u32 %r652, [%rd41+-4];
and.b32 %r653, %r652, -2004318072;
shr.u32 %r654, %r653, 1;
or.b32 %r655, %r654, 839922192;
and.b32 %r656, %r652, 30583;
prmt.b32 %r657, %r592, %r591, %r656;
prmt.b32 %r658, %r595, %r594, %r656;
and.b32 %r659, %r655, 30292;
prmt.b32 %r660, %r657, %r658, %r659;
shr.u32 %r661, %r652, 16;
and.b32 %r662, %r661, 30583;
prmt.b32 %r663, %r592, %r591, %r662;
prmt.b32 %r664, %r595, %r594, %r662;
shr.u32 %r665, %r655, 16;
prmt.b32 %r666, %r663, %r664, %r665;
prmt.b32 %r551, %r660, %r666, %r605;
prmt.b32 %r555, %r660, %r666, %r606;
ld.global.nc.u32 %r667, [%rd41];
and.b32 %r668, %r667, -2004318072;
shr.u32 %r669, %r668, 1;
or.b32 %r670, %r669, 839922192;
and.b32 %r671, %r667, 30583;
prmt.b32 %r672, %r592, %r591, %r671;
prmt.b32 %r673, %r595, %r594, %r671;
and.b32 %r674, %r670, 30292;
prmt.b32 %r675, %r672, %r673, %r674;
shr.u32 %r676, %r667, 16;
and.b32 %r677, %r676, 30583;
prmt.b32 %r678, %r592, %r591, %r677;
prmt.b32 %r679, %r595, %r594, %r677;
shr.u32 %r680, %r670, 16;
prmt.b32 %r681, %r678, %r679, %r680;
prmt.b32 %r559, %r675, %r681, %r605;
prmt.b32 %r563, %r675, %r681, %r606;

dp4a.s32.s32 %r118, %r551, %r84, %r571;


dp4a.s32.s32 %r122, %r555, %r88, %r118;


dp4a.s32.s32 %r126, %r559, %r92, %r122;


dp4a.s32.s32 %r130, %r563, %r96, %r126;

add.s64 %rd42, %rd140, %rd6;
ld.global.nc.u8 %rs63, [%rd42];
and.b16 %rs64, %rs63, 127;
setp.eq.s16 %p6, %rs64, 127;
selp.b16 %rs65, 0, %rs63, %p6;
and.b16 %rs5, %rs65, 255;

{cvt.rn.f16x2.e4m3x2 %r134, %rs5;}


cvt.u16.u32 %rs6, %r134;

{cvt.f32.f16 %f89, %rs6;}


div.approx.ftz.f32 %f150, %f89, %f141;

{.reg .f16 low,high;
mov.b32 {low,high},%r99;
cvt.f32.f16 %f90, low;}


mul.ftz.f32 %f151, %f150, %f90;
cvt.rn.f32.s32 %f152, %r130;
fma.rn.ftz.f32 %f153, %f151, %f152, 0f00000000;
ld.global.nc.u32 %r682, [%rd41+4];
and.b32 %r683, %r682, -2004318072;
shr.u32 %r684, %r683, 1;
or.b32 %r685, %r684, 839922192;
and.b32 %r686, %r682, 30583;
prmt.b32 %r687, %r592, %r591, %r686;
prmt.b32 %r688, %r595, %r594, %r686;
and.b32 %r689, %r685, 30292;
prmt.b32 %r690, %r687, %r688, %r689;
shr.u32 %r691, %r682, 16;
and.b32 %r692, %r691, 30583;
prmt.b32 %r693, %r592, %r591, %r692;
prmt.b32 %r694, %r595, %r594, %r692;
shr.u32 %r695, %r685, 16;
prmt.b32 %r696, %r693, %r694, %r695;
prmt.b32 %r569, %r690, %r696, %r605;
prmt.b32 %r573, %r690, %r696, %r606;
ld.global.nc.u32 %r697, [%rd41+8];
and.b32 %r698, %r697, -2004318072;
shr.u32 %r699, %r698, 1;
or.b32 %r700, %r699, 839922192;
and.b32 %r701, %r697, 30583;
prmt.b32 %r702, %r592, %r591, %r701;
prmt.b32 %r703, %r595, %r594, %r701;
and.b32 %r704, %r700, 30292;
prmt.b32 %r705, %r702, %r703, %r704;
shr.u32 %r706, %r697, 16;
and.b32 %r707, %r706, 30583;
prmt.b32 %r708, %r592, %r591, %r707;
prmt.b32 %r709, %r595, %r594, %r707;
shr.u32 %r710, %r700, 16;
prmt.b32 %r711, %r708, %r709, %r710;
prmt.b32 %r577, %r705, %r711, %r605;
prmt.b32 %r581, %r705, %r711, %r606;

dp4a.s32.s32 %r136, %r569, %r102, %r571;


dp4a.s32.s32 %r140, %r573, %r106, %r136;


dp4a.s32.s32 %r144, %r577, %r110, %r140;


dp4a.s32.s32 %r148, %r581, %r114, %r144;

add.s64 %rd43, %rd140, %rd5;
ld.global.nc.u8 %rs66, [%rd43];
and.b16 %rs67, %rs66, 127;
setp.eq.s16 %p7, %rs67, 127;
selp.b16 %rs68, 0, %rs66, %p7;
and.b16 %rs7, %rs68, 255;

{cvt.rn.f16x2.e4m3x2 %r152, %rs7;}


cvt.u16.u32 %rs8, %r152;

{cvt.f32.f16 %f91, %rs8;}


div.approx.ftz.f32 %f154, %f91, %f141;

{.reg .f16 low,high;
mov.b32 {low,high},%r117;
cvt.f32.f16 %f92, low;}


mul.ftz.f32 %f155, %f154, %f92;
cvt.rn.f32.s32 %f156, %r148;
fma.rn.ftz.f32 %f157, %f155, %f156, %f153;
add.ftz.f32 %f434, %f157, %f434;
cvt.u64.u32 %rd44, %r1038;
add.s64 %rd45, %rd1, %rd44;
add.s64 %rd46, %rd45, %rd3;
mul.lo.s64 %rd47, %rd46, 36;
add.s64 %rd48, %rd11, %rd47;
ld.global.nc.u32 %r156, [%rd48+4];

dp4a.s32.s32 %r154, %r515, %r156, %r571;

ld.global.nc.u32 %r160, [%rd48+12];

dp4a.s32.s32 %r158, %r519, %r160, %r154;

ld.global.nc.u32 %r164, [%rd48+8];

dp4a.s32.s32 %r162, %r523, %r164, %r158;

ld.global.nc.u32 %r168, [%rd48+16];

dp4a.s32.s32 %r166, %r527, %r168, %r162;


{cvt.rn.f16x2.e4m3x2 %r170, %rs1;}


cvt.u16.u32 %rs10, %r170;

{cvt.f32.f16 %f93, %rs10;}


div.approx.ftz.f32 %f158, %f93, %f141;
ld.global.nc.u32 %r171, [%rd48];

{.reg .f16 low,high;
mov.b32 {low,high},%r171;
cvt.f32.f16 %f94, low;}


mul.ftz.f32 %f159, %f158, %f94;
cvt.rn.f32.s32 %f160, %r166;
fma.rn.ftz.f32 %f161, %f159, %f160, 0f00000000;
add.s64 %rd49, %rd45, %rd4;
mul.lo.s64 %rd50, %rd49, 36;
add.s64 %rd51, %rd11, %rd50;
ld.global.nc.u32 %r174, [%rd51+20];

dp4a.s32.s32 %r172, %r533, %r174, %r571;

ld.global.nc.u32 %r178, [%rd51+28];

dp4a.s32.s32 %r176, %r537, %r178, %r172;

ld.global.nc.u32 %r182, [%rd51+24];

dp4a.s32.s32 %r180, %r541, %r182, %r176;

ld.global.nc.u32 %r186, [%rd51+32];

dp4a.s32.s32 %r184, %r545, %r186, %r180;


{cvt.rn.f16x2.e4m3x2 %r188, %rs3;}


cvt.u16.u32 %rs12, %r188;

{cvt.f32.f16 %f95, %rs12;}


div.approx.ftz.f32 %f162, %f95, %f141;
ld.global.nc.u32 %r189, [%rd51];

{.reg .f16 low,high;
mov.b32 {low,high},%r189;
cvt.f32.f16 %f96, low;}


mul.ftz.f32 %f163, %f162, %f96;
cvt.rn.f32.s32 %f164, %r184;
fma.rn.ftz.f32 %f165, %f163, %f164, %f161;
add.ftz.f32 %f433, %f165, %f433;

dp4a.s32.s32 %r190, %r551, %r156, %r571;


dp4a.s32.s32 %r194, %r555, %r160, %r190;


dp4a.s32.s32 %r198, %r559, %r164, %r194;


dp4a.s32.s32 %r202, %r563, %r168, %r198;


{cvt.rn.f16x2.e4m3x2 %r206, %rs5;}


cvt.u16.u32 %rs14, %r206;

{cvt.f32.f16 %f97, %rs14;}


div.approx.ftz.f32 %f166, %f97, %f141;

{.reg .f16 low,high;
mov.b32 {low,high},%r171;
cvt.f32.f16 %f98, low;}


mul.ftz.f32 %f167, %f166, %f98;
cvt.rn.f32.s32 %f168, %r202;
fma.rn.ftz.f32 %f169, %f167, %f168, 0f00000000;

dp4a.s32.s32 %r208, %r569, %r174, %r571;


dp4a.s32.s32 %r212, %r573, %r178, %r208;


dp4a.s32.s32 %r216, %r577, %r182, %r212;


dp4a.s32.s32 %r220, %r581, %r186, %r216;


{cvt.rn.f16x2.e4m3x2 %r224, %rs7;}


cvt.u16.u32 %rs16, %r224;

{cvt.f32.f16 %f99, %rs16;}


div.approx.ftz.f32 %f170, %f99, %f141;

{.reg .f16 low,high;
mov.b32 {low,high},%r189;
cvt.f32.f16 %f100, low;}


mul.ftz.f32 %f171, %f170, %f100;
cvt.rn.f32.s32 %f172, %r220;
fma.rn.ftz.f32 %f173, %f171, %f172, %f169;
add.ftz.f32 %f432, %f173, %f432;
cvt.u64.u32 %rd52, %r1043;
add.s64 %rd53, %rd1, %rd52;
add.s64 %rd54, %rd53, %rd3;
mul.lo.s64 %rd55, %rd54, 36;
add.s64 %rd56, %rd11, %rd55;
ld.global.nc.u32 %r228, [%rd56+4];

dp4a.s32.s32 %r226, %r515, %r228, %r571;

ld.global.nc.u32 %r232, [%rd56+12];

dp4a.s32.s32 %r230, %r519, %r232, %r226;

ld.global.nc.u32 %r236, [%rd56+8];

dp4a.s32.s32 %r234, %r523, %r236, %r230;

ld.global.nc.u32 %r240, [%rd56+16];

dp4a.s32.s32 %r238, %r527, %r240, %r234;


{cvt.rn.f16x2.e4m3x2 %r242, %rs1;}


cvt.u16.u32 %rs18, %r242;

{cvt.f32.f16 %f101, %rs18;}


div.approx.ftz.f32 %f174, %f101, %f141;
ld.global.nc.u32 %r243, [%rd56];

{.reg .f16 low,high;
mov.b32 {low,high},%r243;
cvt.f32.f16 %f102, low;}


mul.ftz.f32 %f175, %f174, %f102;
cvt.rn.f32.s32 %f176, %r238;
fma.rn.ftz.f32 %f177, %f175, %f176, 0f00000000;
add.s64 %rd57, %rd53, %rd4;
mul.lo.s64 %rd58, %rd57, 36;
add.s64 %rd59, %rd11, %rd58;
ld.global.nc.u32 %r246, [%rd59+20];

dp4a.s32.s32 %r244, %r533, %r246, %r571;

ld.global.nc.u32 %r250, [%rd59+28];

dp4a.s32.s32 %r248, %r537, %r250, %r244;

ld.global.nc.u32 %r254, [%rd59+24];

dp4a.s32.s32 %r252, %r541, %r254, %r248;

ld.global.nc.u32 %r258, [%rd59+32];

dp4a.s32.s32 %r256, %r545, %r258, %r252;


{cvt.rn.f16x2.e4m3x2 %r260, %rs3;}


cvt.u16.u32 %rs20, %r260;

{cvt.f32.f16 %f103, %rs20;}


div.approx.ftz.f32 %f178, %f103, %f141;
ld.global.nc.u32 %r261, [%rd59];

{.reg .f16 low,high;
mov.b32 {low,high},%r261;
cvt.f32.f16 %f104, low;}


mul.ftz.f32 %f179, %f178, %f104;
cvt.rn.f32.s32 %f180, %r256;
fma.rn.ftz.f32 %f181, %f179, %f180, %f177;
add.ftz.f32 %f431, %f181, %f431;

dp4a.s32.s32 %r262, %r551, %r228, %r571;


dp4a.s32.s32 %r266, %r555, %r232, %r262;


dp4a.s32.s32 %r270, %r559, %r236, %r266;


dp4a.s32.s32 %r274, %r563, %r240, %r270;


{cvt.rn.f16x2.e4m3x2 %r278, %rs5;}


cvt.u16.u32 %rs22, %r278;

{cvt.f32.f16 %f105, %rs22;}


div.approx.ftz.f32 %f182, %f105, %f141;

{.reg .f16 low,high;
mov.b32 {low,high},%r243;
cvt.f32.f16 %f106, low;}


mul.ftz.f32 %f183, %f182, %f106;
cvt.rn.f32.s32 %f184, %r274;
fma.rn.ftz.f32 %f185, %f183, %f184, 0f00000000;

dp4a.s32.s32 %r280, %r569, %r246, %r571;


dp4a.s32.s32 %r284, %r573, %r250, %r280;


dp4a.s32.s32 %r288, %r577, %r254, %r284;


dp4a.s32.s32 %r292, %r581, %r258, %r288;


{cvt.rn.f16x2.e4m3x2 %r296, %rs7;}


cvt.u16.u32 %rs24, %r296;

{cvt.f32.f16 %f107, %rs24;}


div.approx.ftz.f32 %f186, %f107, %f141;

{.reg .f16 low,high;
mov.b32 {low,high},%r261;
cvt.f32.f16 %f108, low;}


mul.ftz.f32 %f187, %f186, %f108;
cvt.rn.f32.s32 %f188, %r292;
fma.rn.ftz.f32 %f189, %f187, %f188, %f185;
add.ftz.f32 %f430, %f189, %f430;
cvt.u64.u32 %rd60, %r1042;
add.s64 %rd61, %rd1, %rd60;
add.s64 %rd62, %rd61, %rd3;
mul.lo.s64 %rd63, %rd62, 36;
add.s64 %rd64, %rd11, %rd63;
ld.global.nc.u32 %r300, [%rd64+4];

dp4a.s32.s32 %r298, %r515, %r300, %r571;

ld.global.nc.u32 %r304, [%rd64+12];

dp4a.s32.s32 %r302, %r519, %r304, %r298;

ld.global.nc.u32 %r308, [%rd64+8];

dp4a.s32.s32 %r306, %r523, %r308, %r302;

ld.global.nc.u32 %r312, [%rd64+16];

dp4a.s32.s32 %r310, %r527, %r312, %r306;


{cvt.rn.f16x2.e4m3x2 %r314, %rs1;}


cvt.u16.u32 %rs26, %r314;

{cvt.f32.f16 %f109, %rs26;}


div.approx.ftz.f32 %f190, %f109, %f141;
ld.global.nc.u32 %r315, [%rd64];

{.reg .f16 low,high;
mov.b32 {low,high},%r315;
cvt.f32.f16 %f110, low;}


mul.ftz.f32 %f191, %f190, %f110;
cvt.rn.f32.s32 %f192, %r310;
fma.rn.ftz.f32 %f193, %f191, %f192, 0f00000000;
add.s64 %rd65, %rd61, %rd4;
mul.lo.s64 %rd66, %rd65, 36;
add.s64 %rd67, %rd11, %rd66;
ld.global.nc.u32 %r318, [%rd67+20];

dp4a.s32.s32 %r316, %r533, %r318, %r571;

ld.global.nc.u32 %r322, [%rd67+28];

dp4a.s32.s32 %r320, %r537, %r322, %r316;

ld.global.nc.u32 %r326, [%rd67+24];

dp4a.s32.s32 %r324, %r541, %r326, %r320;

ld.global.nc.u32 %r330, [%rd67+32];

dp4a.s32.s32 %r328, %r545, %r330, %r324;


{cvt.rn.f16x2.e4m3x2 %r332, %rs3;}


cvt.u16.u32 %rs28, %r332;

{cvt.f32.f16 %f111, %rs28;}


div.approx.ftz.f32 %f194, %f111, %f141;
ld.global.nc.u32 %r333, [%rd67];

{.reg .f16 low,high;
mov.b32 {low,high},%r333;
cvt.f32.f16 %f112, low;}


mul.ftz.f32 %f195, %f194, %f112;
cvt.rn.f32.s32 %f196, %r328;
fma.rn.ftz.f32 %f197, %f195, %f196, %f193;
add.ftz.f32 %f429, %f197, %f429;

dp4a.s32.s32 %r334, %r551, %r300, %r571;


dp4a.s32.s32 %r338, %r555, %r304, %r334;


dp4a.s32.s32 %r342, %r559, %r308, %r338;


dp4a.s32.s32 %r346, %r563, %r312, %r342;


{cvt.rn.f16x2.e4m3x2 %r350, %rs5;}


cvt.u16.u32 %rs30, %r350;

{cvt.f32.f16 %f113, %rs30;}


div.approx.ftz.f32 %f198, %f113, %f141;

{.reg .f16 low,high;
mov.b32 {low,high},%r315;
cvt.f32.f16 %f114, low;}


mul.ftz.f32 %f199, %f198, %f114;
cvt.rn.f32.s32 %f200, %r346;
fma.rn.ftz.f32 %f201, %f199, %f200, 0f00000000;

dp4a.s32.s32 %r352, %r569, %r318, %r571;


dp4a.s32.s32 %r356, %r573, %r322, %r352;


dp4a.s32.s32 %r360, %r577, %r326, %r356;


dp4a.s32.s32 %r364, %r581, %r330, %r360;


{cvt.rn.f16x2.e4m3x2 %r368, %rs7;}


cvt.u16.u32 %rs32, %r368;

{cvt.f32.f16 %f115, %rs32;}


div.approx.ftz.f32 %f202, %f115, %f141;

{.reg .f16 low,high;
mov.b32 {low,high},%r333;
cvt.f32.f16 %f116, low;}


mul.ftz.f32 %f203, %f202, %f116;
cvt.rn.f32.s32 %f204, %r364;
fma.rn.ftz.f32 %f205, %f203, %f204, %f201;
add.ftz.f32 %f428, %f205, %f428;
cvt.u64.u32 %rd68, %r1041;
add.s64 %rd69, %rd1, %rd68;
add.s64 %rd70, %rd69, %rd3;
mul.lo.s64 %rd71, %rd70, 36;
add.s64 %rd72, %rd11, %rd71;
ld.global.nc.u32 %r372, [%rd72+4];

dp4a.s32.s32 %r370, %r515, %r372, %r571;

ld.global.nc.u32 %r376, [%rd72+12];

dp4a.s32.s32 %r374, %r519, %r376, %r370;

ld.global.nc.u32 %r380, [%rd72+8];

dp4a.s32.s32 %r378, %r523, %r380, %r374;

ld.global.nc.u32 %r384, [%rd72+16];

dp4a.s32.s32 %r382, %r527, %r384, %r378;


{cvt.rn.f16x2.e4m3x2 %r386, %rs1;}


cvt.u16.u32 %rs34, %r386;

{cvt.f32.f16 %f117, %rs34;}


div.approx.ftz.f32 %f206, %f117, %f141;
ld.global.nc.u32 %r387, [%rd72];

{.reg .f16 low,high;
mov.b32 {low,high},%r387;
cvt.f32.f16 %f118, low;}


mul.ftz.f32 %f207, %f206, %f118;
cvt.rn.f32.s32 %f208, %r382;
fma.rn.ftz.f32 %f209, %f207, %f208, 0f00000000;
add.s64 %rd73, %rd69, %rd4;
mul.lo.s64 %rd74, %rd73, 36;
add.s64 %rd75, %rd11, %rd74;
ld.global.nc.u32 %r390, [%rd75+20];

dp4a.s32.s32 %r388, %r533, %r390, %r571;

ld.global.nc.u32 %r394, [%rd75+28];

dp4a.s32.s32 %r392, %r537, %r394, %r388;

ld.global.nc.u32 %r398, [%rd75+24];

dp4a.s32.s32 %r396, %r541, %r398, %r392;

ld.global.nc.u32 %r402, [%rd75+32];

dp4a.s32.s32 %r400, %r545, %r402, %r396;


{cvt.rn.f16x2.e4m3x2 %r404, %rs3;}


cvt.u16.u32 %rs36, %r404;

{cvt.f32.f16 %f119, %rs36;}


div.approx.ftz.f32 %f210, %f119, %f141;
ld.global.nc.u32 %r405, [%rd75];

{.reg .f16 low,high;
mov.b32 {low,high},%r405;
cvt.f32.f16 %f120, low;}


mul.ftz.f32 %f211, %f210, %f120;
cvt.rn.f32.s32 %f212, %r400;
fma.rn.ftz.f32 %f213, %f211, %f212, %f209;
add.ftz.f32 %f427, %f213, %f427;

dp4a.s32.s32 %r406, %r551, %r372, %r571;


dp4a.s32.s32 %r410, %r555, %r376, %r406;


dp4a.s32.s32 %r414, %r559, %r380, %r410;


dp4a.s32.s32 %r418, %r563, %r384, %r414;


{cvt.rn.f16x2.e4m3x2 %r422, %rs5;}


cvt.u16.u32 %rs38, %r422;

{cvt.f32.f16 %f121, %rs38;}


div.approx.ftz.f32 %f214, %f121, %f141;

{.reg .f16 low,high;
mov.b32 {low,high},%r387;
cvt.f32.f16 %f122, low;}


mul.ftz.f32 %f215, %f214, %f122;
cvt.rn.f32.s32 %f216, %r418;
fma.rn.ftz.f32 %f217, %f215, %f216, 0f00000000;

dp4a.s32.s32 %r424, %r569, %r390, %r571;


dp4a.s32.s32 %r428, %r573, %r394, %r424;


dp4a.s32.s32 %r432, %r577, %r398, %r428;


dp4a.s32.s32 %r436, %r581, %r402, %r432;


{cvt.rn.f16x2.e4m3x2 %r440, %rs7;}


cvt.u16.u32 %rs40, %r440;

{cvt.f32.f16 %f123, %rs40;}


div.approx.ftz.f32 %f218, %f123, %f141;

{.reg .f16 low,high;
mov.b32 {low,high},%r405;
cvt.f32.f16 %f124, low;}


mul.ftz.f32 %f219, %f218, %f124;
cvt.rn.f32.s32 %f220, %r436;
fma.rn.ftz.f32 %f221, %f219, %f220, %f217;
add.ftz.f32 %f426, %f221, %f426;
cvt.u64.u32 %rd76, %r1040;
add.s64 %rd77, %rd1, %rd76;
add.s64 %rd78, %rd77, %rd3;
mul.lo.s64 %rd79, %rd78, 36;
add.s64 %rd80, %rd11, %rd79;
ld.global.nc.u32 %r444, [%rd80+4];

dp4a.s32.s32 %r442, %r515, %r444, %r571;

ld.global.nc.u32 %r448, [%rd80+12];

dp4a.s32.s32 %r446, %r519, %r448, %r442;

ld.global.nc.u32 %r452, [%rd80+8];

dp4a.s32.s32 %r450, %r523, %r452, %r446;

ld.global.nc.u32 %r456, [%rd80+16];

dp4a.s32.s32 %r454, %r527, %r456, %r450;


{cvt.rn.f16x2.e4m3x2 %r458, %rs1;}


cvt.u16.u32 %rs42, %r458;

{cvt.f32.f16 %f125, %rs42;}


div.approx.ftz.f32 %f222, %f125, %f141;
ld.global.nc.u32 %r459, [%rd80];

{.reg .f16 low,high;
mov.b32 {low,high},%r459;
cvt.f32.f16 %f126, low;}


mul.ftz.f32 %f223, %f222, %f126;
cvt.rn.f32.s32 %f224, %r454;
fma.rn.ftz.f32 %f225, %f223, %f224, 0f00000000;
add.s64 %rd81, %rd77, %rd4;
mul.lo.s64 %rd82, %rd81, 36;
add.s64 %rd83, %rd11, %rd82;
ld.global.nc.u32 %r462, [%rd83+20];

dp4a.s32.s32 %r460, %r533, %r462, %r571;

ld.global.nc.u32 %r466, [%rd83+28];

dp4a.s32.s32 %r464, %r537, %r466, %r460;

ld.global.nc.u32 %r470, [%rd83+24];

dp4a.s32.s32 %r468, %r541, %r470, %r464;

ld.global.nc.u32 %r474, [%rd83+32];

dp4a.s32.s32 %r472, %r545, %r474, %r468;


{cvt.rn.f16x2.e4m3x2 %r476, %rs3;}


cvt.u16.u32 %rs44, %r476;

{cvt.f32.f16 %f127, %rs44;}


div.approx.ftz.f32 %f226, %f127, %f141;
ld.global.nc.u32 %r477, [%rd83];

{.reg .f16 low,high;
mov.b32 {low,high},%r477;
cvt.f32.f16 %f128, low;}


mul.ftz.f32 %f227, %f226, %f128;
cvt.rn.f32.s32 %f228, %r472;
fma.rn.ftz.f32 %f229, %f227, %f228, %f225;
add.ftz.f32 %f425, %f229, %f425;

dp4a.s32.s32 %r478, %r551, %r444, %r571;


dp4a.s32.s32 %r482, %r555, %r448, %r478;


dp4a.s32.s32 %r486, %r559, %r452, %r482;


dp4a.s32.s32 %r490, %r563, %r456, %r486;


{cvt.rn.f16x2.e4m3x2 %r494, %rs5;}


cvt.u16.u32 %rs46, %r494;

{cvt.f32.f16 %f129, %rs46;}


div.approx.ftz.f32 %f230, %f129, %f141;

{.reg .f16 low,high;
mov.b32 {low,high},%r459;
cvt.f32.f16 %f130, low;}


mul.ftz.f32 %f231, %f230, %f130;
cvt.rn.f32.s32 %f232, %r490;
fma.rn.ftz.f32 %f233, %f231, %f232, 0f00000000;

dp4a.s32.s32 %r496, %r569, %r462, %r571;


dp4a.s32.s32 %r500, %r573, %r466, %r496;


dp4a.s32.s32 %r504, %r577, %r470, %r500;


dp4a.s32.s32 %r508, %r581, %r474, %r504;


{cvt.rn.f16x2.e4m3x2 %r512, %rs7;}


cvt.u16.u32 %rs48, %r512;

{cvt.f32.f16 %f131, %rs48;}


div.approx.ftz.f32 %f234, %f131, %f141;

{.reg .f16 low,high;
mov.b32 {low,high},%r477;
cvt.f32.f16 %f132, low;}


mul.ftz.f32 %f235, %f234, %f132;
cvt.rn.f32.s32 %f236, %r508;
fma.rn.ftz.f32 %f237, %f235, %f236, %f233;
add.ftz.f32 %f424, %f237, %f424;
cvt.u64.u32 %rd84, %r1039;
add.s64 %rd85, %rd1, %rd84;
add.s64 %rd86, %rd85, %rd3;
mul.lo.s64 %rd87, %rd86, 36;
add.s64 %rd88, %rd11, %rd87;
ld.global.nc.u32 %r516, [%rd88+4];

dp4a.s32.s32 %r514, %r515, %r516, %r571;

ld.global.nc.u32 %r520, [%rd88+12];

dp4a.s32.s32 %r518, %r519, %r520, %r514;

ld.global.nc.u32 %r524, [%rd88+8];

dp4a.s32.s32 %r522, %r523, %r524, %r518;

ld.global.nc.u32 %r528, [%rd88+16];

dp4a.s32.s32 %r526, %r527, %r528, %r522;


{cvt.rn.f16x2.e4m3x2 %r530, %rs1;}


cvt.u16.u32 %rs50, %r530;

{cvt.f32.f16 %f133, %rs50;}


div.approx.ftz.f32 %f238, %f133, %f141;
ld.global.nc.u32 %r531, [%rd88];

{.reg .f16 low,high;
mov.b32 {low,high},%r531;
cvt.f32.f16 %f134, low;}


mul.ftz.f32 %f239, %f238, %f134;
cvt.rn.f32.s32 %f240, %r526;
fma.rn.ftz.f32 %f241, %f239, %f240, 0f00000000;
add.s64 %rd89, %rd85, %rd4;
mul.lo.s64 %rd90, %rd89, 36;
add.s64 %rd91, %rd11, %rd90;
ld.global.nc.u32 %r534, [%rd91+20];

dp4a.s32.s32 %r532, %r533, %r534, %r571;

ld.global.nc.u32 %r538, [%rd91+28];

dp4a.s32.s32 %r536, %r537, %r538, %r532;

ld.global.nc.u32 %r542, [%rd91+24];

dp4a.s32.s32 %r540, %r541, %r542, %r536;

ld.global.nc.u32 %r546, [%rd91+32];

dp4a.s32.s32 %r544, %r545, %r546, %r540;


{cvt.rn.f16x2.e4m3x2 %r548, %rs3;}


cvt.u16.u32 %rs52, %r548;

{cvt.f32.f16 %f135, %rs52;}


div.approx.ftz.f32 %f242, %f135, %f141;
ld.global.nc.u32 %r549, [%rd91];

{.reg .f16 low,high;
mov.b32 {low,high},%r549;
cvt.f32.f16 %f136, low;}


mul.ftz.f32 %f243, %f242, %f136;
cvt.rn.f32.s32 %f244, %r544;
fma.rn.ftz.f32 %f245, %f243, %f244, %f241;
add.ftz.f32 %f423, %f245, %f423;

dp4a.s32.s32 %r550, %r551, %r516, %r571;


dp4a.s32.s32 %r554, %r555, %r520, %r550;


dp4a.s32.s32 %r558, %r559, %r524, %r554;


dp4a.s32.s32 %r562, %r563, %r528, %r558;


{cvt.rn.f16x2.e4m3x2 %r566, %rs5;}


cvt.u16.u32 %rs54, %r566;

{cvt.f32.f16 %f137, %rs54;}


div.approx.ftz.f32 %f246, %f137, %f141;

{.reg .f16 low,high;
mov.b32 {low,high},%r531;
cvt.f32.f16 %f138, low;}


mul.ftz.f32 %f247, %f246, %f138;
cvt.rn.f32.s32 %f248, %r562;
fma.rn.ftz.f32 %f249, %f247, %f248, 0f00000000;

dp4a.s32.s32 %r568, %r569, %r534, %r571;


dp4a.s32.s32 %r572, %r573, %r538, %r568;


dp4a.s32.s32 %r576, %r577, %r542, %r572;


dp4a.s32.s32 %r580, %r581, %r546, %r576;


{cvt.rn.f16x2.e4m3x2 %r584, %rs7;}


cvt.u16.u32 %rs56, %r584;

{cvt.f32.f16 %f139, %rs56;}


div.approx.ftz.f32 %f250, %f139, %f141;

{.reg .f16 low,high;
mov.b32 {low,high},%r549;
cvt.f32.f16 %f140, low;}


mul.ftz.f32 %f251, %f250, %f140;
cvt.rn.f32.s32 %f252, %r580;
fma.rn.ftz.f32 %f253, %f251, %f252, %f249;
add.ftz.f32 %f422, %f253, %f422;
add.s32 %r1043, %r1043, 64;
add.s32 %r1042, %r1042, 64;
add.s32 %r1041, %r1041, 64;
add.s32 %r1040, %r1040, 64;
add.s32 %r1039, %r1039, 64;
add.s32 %r1038, %r1038, 64;
add.s32 %r1037, %r1037, 64;
add.s64 %rd140, %rd140, 1152;
add.s32 %r1044, %r1044, 32;
setp.lt.s32 %p8, %r1044, %r51;
@%p8 bra $L__BB125_2;

$L__BB125_3:
setp.eq.s32 %p9, %r1, 0;
@%p9 bra $L__BB125_5;

mov.u32 %r714, _ZZ13mul_mat_vec_qIL9ggml_type40ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r715, %r1, 1792, %r714;
shl.b32 %r716, %r47, 2;
add.s32 %r717, %r715, %r716;
st.shared.f32 [%r717+-1792], %f435;
st.shared.f32 [%r717+-1664], %f434;
st.shared.f32 [%r717+-1536], %f433;
st.shared.f32 [%r717+-1408], %f432;
st.shared.f32 [%r717+-1280], %f431;
st.shared.f32 [%r717+-1152], %f430;
st.shared.f32 [%r717+-1024], %f429;
st.shared.f32 [%r717+-896], %f428;
st.shared.f32 [%r717+-768], %f427;
st.shared.f32 [%r717+-640], %f426;
st.shared.f32 [%r717+-512], %f425;
st.shared.f32 [%r717+-384], %f424;
st.shared.f32 [%r717+-256], %f423;
st.shared.f32 [%r717+-128], %f422;

$L__BB125_5:
setp.ne.s32 %p10, %r1, 0;
bar.sync 0;
@%p10 bra $L__BB125_34;

mov.u32 %r718, %ctaid.z;
mov.u32 %r719, %ctaid.y;
mad.lo.s32 %r720, %r719, %r39, %r3;
mad.lo.s32 %r721, %r718, %r45, %r720;
cvt.u64.u32 %rd14, %r721;
shl.b32 %r723, %r47, 2;
mov.u32 %r724, _ZZ13mul_mat_vec_qIL9ggml_type40ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r725, %r724, %r723;
mov.u32 %r726, 2;
ld.shared.f32 %f254, [%r725];
add.ftz.f32 %f255, %f254, %f435;
mov.b32 %r727, %f255;
mov.u32 %r728, 31;
mov.u32 %r729, 16;
mov.u32 %r730, -1;
shfl.sync.bfly.b32 %r731|%p11, %r727, %r729, %r728, %r730;
mov.b32 %f256, %r731;
add.ftz.f32 %f257, %f255, %f256;
mov.b32 %r732, %f257;
mov.u32 %r733, 8;
shfl.sync.bfly.b32 %r734|%p12, %r732, %r733, %r728, %r730;
mov.b32 %f258, %r734;
add.ftz.f32 %f259, %f257, %f258;
mov.b32 %r735, %f259;
mov.u32 %r736, 4;
shfl.sync.bfly.b32 %r737|%p13, %r735, %r736, %r728, %r730;
mov.b32 %f260, %r737;
add.ftz.f32 %f261, %f259, %f260;
mov.b32 %r738, %f261;
shfl.sync.bfly.b32 %r739|%p14, %r738, %r726, %r728, %r730;
mov.b32 %f262, %r739;
add.ftz.f32 %f263, %f261, %f262;
mov.b32 %r740, %f263;
mov.u32 %r741, 1;
shfl.sync.bfly.b32 %r742|%p15, %r740, %r741, %r728, %r730;
mov.b32 %f264, %r742;
add.ftz.f32 %f43, %f263, %f264;
setp.eq.s32 %p16, %r47, 0;
setp.lt.u32 %p17, %r3, %r33;
and.pred %p1, %p16, %p17;
not.pred %p18, %p1;
@%p18 bra $L__BB125_8;

cvta.to.global.u64 %rd92, %rd20;
mul.wide.u32 %rd93, %r721, 4;
add.s64 %rd94, %rd92, %rd93;
st.global.f32 [%rd94], %f43;

$L__BB125_8:
ld.shared.f32 %f265, [%r725+128];
add.ftz.f32 %f266, %f265, %f434;
mov.b32 %r754, %f266;
shfl.sync.bfly.b32 %r758|%p19, %r754, %r729, %r728, %r730;
mov.b32 %f267, %r758;
add.ftz.f32 %f268, %f266, %f267;
mov.b32 %r759, %f268;
shfl.sync.bfly.b32 %r761|%p20, %r759, %r733, %r728, %r730;
mov.b32 %f269, %r761;
add.ftz.f32 %f270, %f268, %f269;
mov.b32 %r762, %f270;
shfl.sync.bfly.b32 %r764|%p21, %r762, %r736, %r728, %r730;
mov.b32 %f271, %r764;
add.ftz.f32 %f272, %f270, %f271;
mov.b32 %r765, %f272;
shfl.sync.bfly.b32 %r766|%p22, %r765, %r726, %r728, %r730;
mov.b32 %f273, %r766;
add.ftz.f32 %f274, %f272, %f273;
mov.b32 %r767, %f274;
shfl.sync.bfly.b32 %r769|%p23, %r767, %r741, %r728, %r730;
mov.b32 %f275, %r769;
add.ftz.f32 %f44, %f274, %f275;
setp.eq.s32 %p24, %r47, 1;
or.b32 %r772, %r3, 1;
setp.lt.u32 %p25, %r772, %r33;
and.pred %p2, %p24, %p25;
not.pred %p26, %p2;
@%p26 bra $L__BB125_10;

cvta.to.global.u64 %rd95, %rd20;
mul.wide.u32 %rd96, %r721, 4;
add.s64 %rd97, %rd95, %rd96;
st.global.f32 [%rd97+4], %f44;

$L__BB125_10:
mov.u32 %r783, 2;
ld.shared.f32 %f276, [%r725+256];
add.ftz.f32 %f277, %f276, %f433;
mov.b32 %r784, %f277;
mov.u32 %r785, 31;
mov.u32 %r786, 16;
mov.u32 %r787, -1;
shfl.sync.bfly.b32 %r788|%p27, %r784, %r786, %r785, %r787;
mov.b32 %f278, %r788;
add.ftz.f32 %f279, %f277, %f278;
mov.b32 %r789, %f279;
mov.u32 %r790, 8;
shfl.sync.bfly.b32 %r791|%p28, %r789, %r790, %r785, %r787;
mov.b32 %f280, %r791;
add.ftz.f32 %f281, %f279, %f280;
mov.b32 %r792, %f281;
mov.u32 %r793, 4;
shfl.sync.bfly.b32 %r794|%p29, %r792, %r793, %r785, %r787;
mov.b32 %f282, %r794;
add.ftz.f32 %f283, %f281, %f282;
mov.b32 %r795, %f283;
shfl.sync.bfly.b32 %r796|%p30, %r795, %r783, %r785, %r787;
mov.b32 %f284, %r796;
add.ftz.f32 %f285, %f283, %f284;
mov.b32 %r797, %f285;
mov.u32 %r798, 1;
shfl.sync.bfly.b32 %r799|%p31, %r797, %r798, %r785, %r787;
mov.b32 %f286, %r799;
add.ftz.f32 %f45, %f285, %f286;
@%p18 bra $L__BB125_12;

cvt.u64.u32 %rd98, %r33;
add.s64 %rd99, %rd98, %rd14;
cvta.to.global.u64 %rd100, %rd20;
shl.b64 %rd101, %rd99, 2;
add.s64 %rd102, %rd100, %rd101;
st.global.f32 [%rd102], %f45;

$L__BB125_12:
ld.shared.f32 %f287, [%r725+384];
add.ftz.f32 %f288, %f287, %f432;
mov.b32 %r805, %f288;
shfl.sync.bfly.b32 %r809|%p33, %r805, %r786, %r785, %r787;
mov.b32 %f289, %r809;
add.ftz.f32 %f290, %f288, %f289;
mov.b32 %r810, %f290;
shfl.sync.bfly.b32 %r812|%p34, %r810, %r790, %r785, %r787;
mov.b32 %f291, %r812;
add.ftz.f32 %f292, %f290, %f291;
mov.b32 %r813, %f292;
shfl.sync.bfly.b32 %r815|%p35, %r813, %r793, %r785, %r787;
mov.b32 %f293, %r815;
add.ftz.f32 %f294, %f292, %f293;
mov.b32 %r816, %f294;
shfl.sync.bfly.b32 %r817|%p36, %r816, %r783, %r785, %r787;
mov.b32 %f295, %r817;
add.ftz.f32 %f296, %f294, %f295;
mov.b32 %r818, %f296;
shfl.sync.bfly.b32 %r820|%p37, %r818, %r798, %r785, %r787;
mov.b32 %f297, %r820;
add.ftz.f32 %f46, %f296, %f297;
@%p26 bra $L__BB125_14;

add.s32 %r821, %r33, 1;
cvt.u64.u32 %rd103, %r821;
add.s64 %rd104, %rd103, %rd14;
cvta.to.global.u64 %rd105, %rd20;
shl.b64 %rd106, %rd104, 2;
add.s64 %rd107, %rd105, %rd106;
st.global.f32 [%rd107], %f46;

$L__BB125_14:
shl.b32 %r822, %r33, 1;
mov.u32 %r823, 1;
mov.u32 %r828, 2;
ld.shared.f32 %f298, [%r725+512];
add.ftz.f32 %f299, %f298, %f431;
mov.b32 %r829, %f299;
mov.u32 %r830, 31;
mov.u32 %r831, 16;
mov.u32 %r832, -1;
shfl.sync.bfly.b32 %r833|%p39, %r829, %r831, %r830, %r832;
mov.b32 %f300, %r833;
add.ftz.f32 %f301, %f299, %f300;
mov.b32 %r834, %f301;
mov.u32 %r835, 8;
shfl.sync.bfly.b32 %r836|%p40, %r834, %r835, %r830, %r832;
mov.b32 %f302, %r836;
add.ftz.f32 %f303, %f301, %f302;
mov.b32 %r837, %f303;
mov.u32 %r838, 4;
shfl.sync.bfly.b32 %r839|%p41, %r837, %r838, %r830, %r832;
mov.b32 %f304, %r839;
add.ftz.f32 %f305, %f303, %f304;
mov.b32 %r840, %f305;
shfl.sync.bfly.b32 %r841|%p42, %r840, %r828, %r830, %r832;
mov.b32 %f306, %r841;
add.ftz.f32 %f307, %f305, %f306;
mov.b32 %r842, %f307;
shfl.sync.bfly.b32 %r843|%p43, %r842, %r823, %r830, %r832;
mov.b32 %f308, %r843;
add.ftz.f32 %f47, %f307, %f308;
cvt.u64.u32 %rd108, %r822;
add.s64 %rd109, %rd108, %rd14;
cvta.to.global.u64 %rd110, %rd20;
shl.b64 %rd111, %rd109, 2;
add.s64 %rd15, %rd110, %rd111;
@%p18 bra $L__BB125_16;

st.global.f32 [%rd15], %f47;

$L__BB125_16:
ld.shared.f32 %f309, [%r725+640];
add.ftz.f32 %f310, %f309, %f430;
mov.b32 %r849, %f310;
shfl.sync.bfly.b32 %r853|%p45, %r849, %r831, %r830, %r832;
mov.b32 %f311, %r853;
add.ftz.f32 %f312, %f310, %f311;
mov.b32 %r854, %f312;
shfl.sync.bfly.b32 %r856|%p46, %r854, %r835, %r830, %r832;
mov.b32 %f313, %r856;
add.ftz.f32 %f314, %f312, %f313;
mov.b32 %r857, %f314;
shfl.sync.bfly.b32 %r859|%p47, %r857, %r838, %r830, %r832;
mov.b32 %f315, %r859;
add.ftz.f32 %f316, %f314, %f315;
mov.b32 %r860, %f316;
shfl.sync.bfly.b32 %r861|%p48, %r860, %r828, %r830, %r832;
mov.b32 %f317, %r861;
add.ftz.f32 %f318, %f316, %f317;
mov.b32 %r862, %f318;
shfl.sync.bfly.b32 %r864|%p49, %r862, %r823, %r830, %r832;
mov.b32 %f319, %r864;
add.ftz.f32 %f48, %f318, %f319;
@%p26 bra $L__BB125_18;

st.global.f32 [%rd15+4], %f48;

$L__BB125_18:
mul.lo.s32 %r28, %r33, 3;
mov.u32 %r869, 2;
ld.shared.f32 %f320, [%r725+768];
add.ftz.f32 %f321, %f320, %f429;
mov.b32 %r870, %f321;
mov.u32 %r871, 31;
mov.u32 %r872, 16;
mov.u32 %r873, -1;
shfl.sync.bfly.b32 %r874|%p51, %r870, %r872, %r871, %r873;
mov.b32 %f322, %r874;
add.ftz.f32 %f323, %f321, %f322;
mov.b32 %r875, %f323;
mov.u32 %r876, 8;
shfl.sync.bfly.b32 %r877|%p52, %r875, %r876, %r871, %r873;
mov.b32 %f324, %r877;
add.ftz.f32 %f325, %f323, %f324;
mov.b32 %r878, %f325;
mov.u32 %r879, 4;
shfl.sync.bfly.b32 %r880|%p53, %r878, %r879, %r871, %r873;
mov.b32 %f326, %r880;
add.ftz.f32 %f327, %f325, %f326;
mov.b32 %r881, %f327;
shfl.sync.bfly.b32 %r882|%p54, %r881, %r869, %r871, %r873;
mov.b32 %f328, %r882;
add.ftz.f32 %f329, %f327, %f328;
mov.b32 %r883, %f329;
mov.u32 %r884, 1;
shfl.sync.bfly.b32 %r885|%p55, %r883, %r884, %r871, %r873;
mov.b32 %f330, %r885;
add.ftz.f32 %f49, %f329, %f330;
@%p18 bra $L__BB125_20;

cvt.u64.u32 %rd112, %r28;
add.s64 %rd113, %rd112, %rd14;
shl.b64 %rd115, %rd113, 2;
add.s64 %rd116, %rd110, %rd115;
st.global.f32 [%rd116], %f49;

$L__BB125_20:
ld.shared.f32 %f331, [%r725+896];
add.ftz.f32 %f332, %f331, %f428;
mov.b32 %r891, %f332;
shfl.sync.bfly.b32 %r895|%p57, %r891, %r872, %r871, %r873;
mov.b32 %f333, %r895;
add.ftz.f32 %f334, %f332, %f333;
mov.b32 %r896, %f334;
shfl.sync.bfly.b32 %r898|%p58, %r896, %r876, %r871, %r873;
mov.b32 %f335, %r898;
add.ftz.f32 %f336, %f334, %f335;
mov.b32 %r899, %f336;
shfl.sync.bfly.b32 %r901|%p59, %r899, %r879, %r871, %r873;
mov.b32 %f337, %r901;
add.ftz.f32 %f338, %f336, %f337;
mov.b32 %r902, %f338;
shfl.sync.bfly.b32 %r903|%p60, %r902, %r869, %r871, %r873;
mov.b32 %f339, %r903;
add.ftz.f32 %f340, %f338, %f339;
mov.b32 %r904, %f340;
shfl.sync.bfly.b32 %r906|%p61, %r904, %r884, %r871, %r873;
mov.b32 %f341, %r906;
add.ftz.f32 %f50, %f340, %f341;
@%p26 bra $L__BB125_22;

add.s32 %r907, %r28, 1;
cvt.u64.u32 %rd117, %r907;
add.s64 %rd118, %rd117, %rd14;
shl.b64 %rd120, %rd118, 2;
add.s64 %rd121, %rd110, %rd120;
st.global.f32 [%rd121], %f50;

$L__BB125_22:
shl.b32 %r908, %r33, 2;
mov.u32 %r909, 2;
ld.shared.f32 %f342, [%r725+1024];
add.ftz.f32 %f343, %f342, %f427;
mov.b32 %r914, %f343;
mov.u32 %r915, 31;
mov.u32 %r916, 16;
mov.u32 %r917, -1;
shfl.sync.bfly.b32 %r918|%p63, %r914, %r916, %r915, %r917;
mov.b32 %f344, %r918;
add.ftz.f32 %f345, %f343, %f344;
mov.b32 %r919, %f345;
mov.u32 %r920, 8;
shfl.sync.bfly.b32 %r921|%p64, %r919, %r920, %r915, %r917;
mov.b32 %f346, %r921;
add.ftz.f32 %f347, %f345, %f346;
mov.b32 %r922, %f347;
mov.u32 %r923, 4;
shfl.sync.bfly.b32 %r924|%p65, %r922, %r923, %r915, %r917;
mov.b32 %f348, %r924;
add.ftz.f32 %f349, %f347, %f348;
mov.b32 %r925, %f349;
shfl.sync.bfly.b32 %r926|%p66, %r925, %r909, %r915, %r917;
mov.b32 %f350, %r926;
add.ftz.f32 %f351, %f349, %f350;
mov.b32 %r927, %f351;
mov.u32 %r928, 1;
shfl.sync.bfly.b32 %r929|%p67, %r927, %r928, %r915, %r917;
mov.b32 %f352, %r929;
add.ftz.f32 %f51, %f351, %f352;
cvt.u64.u32 %rd122, %r908;
add.s64 %rd123, %rd122, %rd14;
shl.b64 %rd125, %rd123, 2;
add.s64 %rd16, %rd110, %rd125;
@%p18 bra $L__BB125_24;

st.global.f32 [%rd16], %f51;

$L__BB125_24:
ld.shared.f32 %f353, [%r725+1152];
add.ftz.f32 %f354, %f353, %f426;
mov.b32 %r935, %f354;
shfl.sync.bfly.b32 %r939|%p69, %r935, %r916, %r915, %r917;
mov.b32 %f355, %r939;
add.ftz.f32 %f356, %f354, %f355;
mov.b32 %r940, %f356;
shfl.sync.bfly.b32 %r942|%p70, %r940, %r920, %r915, %r917;
mov.b32 %f357, %r942;
add.ftz.f32 %f358, %f356, %f357;
mov.b32 %r943, %f358;
shfl.sync.bfly.b32 %r945|%p71, %r943, %r923, %r915, %r917;
mov.b32 %f359, %r945;
add.ftz.f32 %f360, %f358, %f359;
mov.b32 %r946, %f360;
shfl.sync.bfly.b32 %r947|%p72, %r946, %r909, %r915, %r917;
mov.b32 %f361, %r947;
add.ftz.f32 %f362, %f360, %f361;
mov.b32 %r948, %f362;
shfl.sync.bfly.b32 %r950|%p73, %r948, %r928, %r915, %r917;
mov.b32 %f363, %r950;
add.ftz.f32 %f52, %f362, %f363;
@%p26 bra $L__BB125_26;

st.global.f32 [%rd16+4], %f52;

$L__BB125_26:
mul.lo.s32 %r29, %r33, 5;
mov.u32 %r955, 2;
ld.shared.f32 %f364, [%r725+1280];
add.ftz.f32 %f365, %f364, %f425;
mov.b32 %r956, %f365;
mov.u32 %r957, 31;
mov.u32 %r958, 16;
mov.u32 %r959, -1;
shfl.sync.bfly.b32 %r960|%p75, %r956, %r958, %r957, %r959;
mov.b32 %f366, %r960;
add.ftz.f32 %f367, %f365, %f366;
mov.b32 %r961, %f367;
mov.u32 %r962, 8;
shfl.sync.bfly.b32 %r963|%p76, %r961, %r962, %r957, %r959;
mov.b32 %f368, %r963;
add.ftz.f32 %f369, %f367, %f368;
mov.b32 %r964, %f369;
mov.u32 %r965, 4;
shfl.sync.bfly.b32 %r966|%p77, %r964, %r965, %r957, %r959;
mov.b32 %f370, %r966;
add.ftz.f32 %f371, %f369, %f370;
mov.b32 %r967, %f371;
shfl.sync.bfly.b32 %r968|%p78, %r967, %r955, %r957, %r959;
mov.b32 %f372, %r968;
add.ftz.f32 %f373, %f371, %f372;
mov.b32 %r969, %f373;
mov.u32 %r970, 1;
shfl.sync.bfly.b32 %r971|%p79, %r969, %r970, %r957, %r959;
mov.b32 %f374, %r971;
add.ftz.f32 %f53, %f373, %f374;
@%p18 bra $L__BB125_28;

cvt.u64.u32 %rd126, %r29;
add.s64 %rd127, %rd126, %rd14;
shl.b64 %rd129, %rd127, 2;
add.s64 %rd130, %rd110, %rd129;
st.global.f32 [%rd130], %f53;

$L__BB125_28:
ld.shared.f32 %f375, [%r725+1408];
add.ftz.f32 %f376, %f375, %f424;
mov.b32 %r977, %f376;
shfl.sync.bfly.b32 %r981|%p81, %r977, %r958, %r957, %r959;
mov.b32 %f377, %r981;
add.ftz.f32 %f378, %f376, %f377;
mov.b32 %r982, %f378;
shfl.sync.bfly.b32 %r984|%p82, %r982, %r962, %r957, %r959;
mov.b32 %f379, %r984;
add.ftz.f32 %f380, %f378, %f379;
mov.b32 %r985, %f380;
shfl.sync.bfly.b32 %r987|%p83, %r985, %r965, %r957, %r959;
mov.b32 %f381, %r987;
add.ftz.f32 %f382, %f380, %f381;
mov.b32 %r988, %f382;
shfl.sync.bfly.b32 %r989|%p84, %r988, %r955, %r957, %r959;
mov.b32 %f383, %r989;
add.ftz.f32 %f384, %f382, %f383;
mov.b32 %r990, %f384;
shfl.sync.bfly.b32 %r992|%p85, %r990, %r970, %r957, %r959;
mov.b32 %f385, %r992;
add.ftz.f32 %f54, %f384, %f385;
@%p26 bra $L__BB125_30;

add.s32 %r993, %r29, 1;
cvt.u64.u32 %rd131, %r993;
add.s64 %rd132, %rd131, %rd14;
shl.b64 %rd134, %rd132, 2;
add.s64 %rd135, %rd110, %rd134;
st.global.f32 [%rd135], %f54;

$L__BB125_30:
mul.lo.s32 %r994, %r33, 6;
mov.u32 %r999, 2;
ld.shared.f32 %f386, [%r725+1536];
add.ftz.f32 %f387, %f386, %f423;
mov.b32 %r1000, %f387;
mov.u32 %r1001, 31;
mov.u32 %r1002, 16;
mov.u32 %r1003, -1;
shfl.sync.bfly.b32 %r1004|%p87, %r1000, %r1002, %r1001, %r1003;
mov.b32 %f388, %r1004;
add.ftz.f32 %f389, %f387, %f388;
mov.b32 %r1005, %f389;
mov.u32 %r1006, 8;
shfl.sync.bfly.b32 %r1007|%p88, %r1005, %r1006, %r1001, %r1003;
mov.b32 %f390, %r1007;
add.ftz.f32 %f391, %f389, %f390;
mov.b32 %r1008, %f391;
mov.u32 %r1009, 4;
shfl.sync.bfly.b32 %r1010|%p89, %r1008, %r1009, %r1001, %r1003;
mov.b32 %f392, %r1010;
add.ftz.f32 %f393, %f391, %f392;
mov.b32 %r1011, %f393;
shfl.sync.bfly.b32 %r1012|%p90, %r1011, %r999, %r1001, %r1003;
mov.b32 %f394, %r1012;
add.ftz.f32 %f395, %f393, %f394;
mov.b32 %r1013, %f395;
mov.u32 %r1014, 1;
shfl.sync.bfly.b32 %r1015|%p91, %r1013, %r1014, %r1001, %r1003;
mov.b32 %f396, %r1015;
add.ftz.f32 %f55, %f395, %f396;
cvt.u64.u32 %rd136, %r994;
add.s64 %rd137, %rd136, %rd14;
shl.b64 %rd139, %rd137, 2;
add.s64 %rd17, %rd110, %rd139;
@%p18 bra $L__BB125_32;

st.global.f32 [%rd17], %f55;

$L__BB125_32:
ld.shared.f32 %f397, [%r725+1664];
add.ftz.f32 %f398, %f397, %f422;
mov.b32 %r1021, %f398;
shfl.sync.bfly.b32 %r1025|%p93, %r1021, %r1002, %r1001, %r1003;
mov.b32 %f399, %r1025;
add.ftz.f32 %f400, %f398, %f399;
mov.b32 %r1026, %f400;
shfl.sync.bfly.b32 %r1028|%p94, %r1026, %r1006, %r1001, %r1003;
mov.b32 %f401, %r1028;
add.ftz.f32 %f402, %f400, %f401;
mov.b32 %r1029, %f402;
shfl.sync.bfly.b32 %r1031|%p95, %r1029, %r1009, %r1001, %r1003;
mov.b32 %f403, %r1031;
add.ftz.f32 %f404, %f402, %f403;
mov.b32 %r1032, %f404;
shfl.sync.bfly.b32 %r1033|%p96, %r1032, %r999, %r1001, %r1003;
mov.b32 %f405, %r1033;
add.ftz.f32 %f406, %f404, %f405;
mov.b32 %r1034, %f406;
shfl.sync.bfly.b32 %r1036|%p97, %r1034, %r1014, %r1001, %r1003;
mov.b32 %f407, %r1036;
add.ftz.f32 %f56, %f406, %f407;
@%p26 bra $L__BB125_34;

st.global.f32 [%rd17+4], %f56;

$L__BB125_34:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type40ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type40ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type40ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type40ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type40ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type40ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type40ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type40ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type40ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type40ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<124>;
.reg .b16 %rs<77>;
.reg .f32 %f<498>;
.reg .b32 %r<1240>;
.reg .b64 %rd<169>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type40ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[2048];

ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type40ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type40ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type40ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type40ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type40ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type40ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type40ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type40ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type40ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type40ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r46, [_Z13mul_mat_vec_qIL9ggml_type40ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r47, [_Z13mul_mat_vec_qIL9ggml_type40ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r48, [_Z13mul_mat_vec_qIL9ggml_type40ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type40ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type40ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type40ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type40ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r49, %r1, 5;
mov.u32 %r50, %tid.x;
add.s32 %r2, %r49, %r50;
shr.u32 %r51, %r2, 31;
add.s32 %r52, %r2, %r51;
shr.s32 %r1239, %r52, 1;
shr.u32 %r53, %r33, 6;
setp.ge.s32 %p2, %r1239, %r53;
mov.f32 %f482, 0f00000000;
mov.f32 %f483, %f482;
mov.f32 %f484, %f482;
mov.f32 %f485, %f482;
mov.f32 %f486, %f482;
mov.f32 %f487, %f482;
mov.f32 %f488, %f482;
mov.f32 %f489, %f482;
mov.f32 %f490, %f482;
mov.f32 %f491, %f482;
mov.f32 %f492, %f482;
mov.f32 %f493, %f482;
mov.f32 %f494, %f482;
mov.f32 %f495, %f482;
mov.f32 %f496, %f482;
mov.f32 %f497, %f482;
@%p2 bra $L__BB126_3;

mov.u32 %r54, %ctaid.x;
shl.b32 %r55, %r54, 1;
mov.u32 %r56, %ctaid.y;
mov.u32 %r57, %ctaid.z;
mul.lo.s32 %r58, %r56, %r41;
cvt.u64.u32 %rd20, %r58;
mul.lo.s32 %r59, %r57, %r47;
cvt.u64.u32 %rd21, %r59;
add.s64 %rd1, %rd21, %rd20;
mul.hi.u32 %r60, %r57, %r43;
add.s32 %r61, %r60, %r57;
shr.u32 %r62, %r61, %r44;
mul.lo.s32 %r63, %r62, %r46;
mul.hi.u32 %r64, %r56, %r37;
add.s32 %r65, %r64, %r56;
shr.u32 %r66, %r65, %r38;
mul.lo.s32 %r67, %r66, %r40;
cvta.to.global.u64 %rd168, %rd17;
and.b32 %r70, %r52, -2;
sub.s32 %r71, %r2, %r70;
shl.b32 %r72, %r71, 2;
shl.b32 %r73, %r71, 1;
cvt.s64.s32 %rd3, %r71;
cvt.s64.s32 %rd22, %r73;
or.b32 %r74, %r72, 2;
shr.s32 %r75, %r74, 1;
shr.s32 %r76, %r74, 2;
cvt.s64.s32 %rd4, %r76;
cvt.s64.s32 %rd23, %r75;
shl.b32 %r77, %r35, 1;
shl.b32 %r1231, %r1239, 1;
add.s32 %r1238, %r1231, %r77;
mad.lo.s32 %r1237, %r35, 3, %r1231;
shl.b32 %r78, %r35, 2;
add.s32 %r1236, %r1231, %r78;
mad.lo.s32 %r1235, %r35, 5, %r1231;
mad.lo.s32 %r1234, %r35, 6, %r1231;
mad.lo.s32 %r1233, %r35, 7, %r1231;
add.s32 %r1232, %r35, %r1231;
add.s32 %r79, %r1239, %r63;
add.s32 %r80, %r79, %r67;
or.b32 %r81, %r55, 1;
mad.lo.s32 %r82, %r34, %r81, %r80;
mul.wide.s32 %rd24, %r82, 36;
add.s64 %rd5, %rd24, %rd23;
add.s64 %rd6, %rd24, %rd22;
mad.lo.s32 %r83, %r55, %r34, %r1239;
add.s32 %r84, %r83, %r63;
add.s32 %r85, %r84, %r67;
mul.wide.s32 %rd25, %r85, 36;
mul.wide.s32 %rd26, %r72, 4;
add.s64 %rd27, %rd25, %rd26;
add.s64 %rd7, %rd27, 16;
add.s64 %rd8, %rd25, %rd22;
add.s64 %rd28, %rd24, %rd26;
add.s64 %rd9, %rd28, 8;
add.s64 %rd10, %rd25, %rd23;
cvta.to.global.u64 %rd11, %rd18;
mov.f32 %f482, 0f00000000;

$L__BB126_2:
cvt.u64.u32 %rd29, %r1231;
add.s64 %rd30, %rd1, %rd29;
add.s64 %rd31, %rd168, %rd7;
ld.global.nc.u32 %r662, [%rd31+-12];
and.b32 %r663, %r662, -2004318072;
shr.u32 %r664, %r663, 1;
or.b32 %r665, %r664, 839922192;
and.b32 %r666, %r662, 30583;
mov.u32 %r667, 201852420;
mov.u32 %r668, 50462976;
prmt.b32 %r669, %r668, %r667, %r666;
mov.u32 %r670, -185009412;
mov.u32 %r671, -33620224;
prmt.b32 %r672, %r671, %r670, %r666;
and.b32 %r673, %r665, 30292;
prmt.b32 %r674, %r669, %r672, %r673;
shr.u32 %r675, %r662, 16;
and.b32 %r676, %r675, 30583;
prmt.b32 %r677, %r668, %r667, %r676;
prmt.b32 %r678, %r671, %r670, %r676;
shr.u32 %r679, %r665, 16;
prmt.b32 %r680, %r677, %r678, %r679;
mov.u32 %r681, 25632;
prmt.b32 %r591, %r674, %r680, %r681;
mov.u32 %r682, 30001;
prmt.b32 %r595, %r674, %r680, %r682;
ld.global.nc.u32 %r683, [%rd31+-8];
and.b32 %r684, %r683, -2004318072;
shr.u32 %r685, %r684, 1;
or.b32 %r686, %r685, 839922192;
and.b32 %r687, %r683, 30583;
prmt.b32 %r688, %r668, %r667, %r687;
prmt.b32 %r689, %r671, %r670, %r687;
and.b32 %r690, %r686, 30292;
prmt.b32 %r691, %r688, %r689, %r690;
shr.u32 %r692, %r683, 16;
and.b32 %r693, %r692, 30583;
prmt.b32 %r694, %r668, %r667, %r693;
prmt.b32 %r695, %r671, %r670, %r693;
shr.u32 %r696, %r686, 16;
prmt.b32 %r697, %r694, %r695, %r696;
prmt.b32 %r599, %r691, %r697, %r681;
prmt.b32 %r603, %r691, %r697, %r682;
add.s64 %rd32, %rd30, %rd3;
mul.lo.s64 %rd33, %rd32, 36;
add.s64 %rd34, %rd11, %rd33;
ld.global.nc.u32 %r88, [%rd34+4];
mov.u32 %r647, 0;

dp4a.s32.s32 %r86, %r591, %r88, %r647;

ld.global.nc.u32 %r92, [%rd34+12];

dp4a.s32.s32 %r90, %r595, %r92, %r86;

ld.global.nc.u32 %r96, [%rd34+8];

dp4a.s32.s32 %r94, %r599, %r96, %r90;

ld.global.nc.u32 %r100, [%rd34+16];

dp4a.s32.s32 %r98, %r603, %r100, %r94;

add.s64 %rd35, %rd168, %rd8;
ld.global.nc.u8 %rs65, [%rd35];
and.b16 %rs66, %rs65, 127;
setp.eq.s16 %p3, %rs66, 127;
selp.b16 %rs67, 0, %rs65, %p3;
and.b16 %rs1, %rs67, 255;

{cvt.rn.f16x2.e4m3x2 %r102, %rs1;}


cvt.u16.u32 %rs2, %r102;

{cvt.f32.f16 %f97, %rs2;}


mov.f32 %f161, 0f40000000;
div.approx.ftz.f32 %f162, %f97, %f161;
ld.global.nc.u32 %r103, [%rd34];

{.reg .f16 low,high;
mov.b32 {low,high},%r103;
cvt.f32.f16 %f98, low;}


mul.ftz.f32 %f163, %f162, %f98;
cvt.rn.f32.s32 %f164, %r98;
fma.rn.ftz.f32 %f165, %f163, %f164, 0f00000000;
ld.global.nc.u32 %r698, [%rd31+-4];
and.b32 %r699, %r698, -2004318072;
shr.u32 %r700, %r699, 1;
or.b32 %r701, %r700, 839922192;
and.b32 %r702, %r698, 30583;
prmt.b32 %r703, %r668, %r667, %r702;
prmt.b32 %r704, %r671, %r670, %r702;
and.b32 %r705, %r701, 30292;
prmt.b32 %r706, %r703, %r704, %r705;
shr.u32 %r707, %r698, 16;
and.b32 %r708, %r707, 30583;
prmt.b32 %r709, %r668, %r667, %r708;
prmt.b32 %r710, %r671, %r670, %r708;
shr.u32 %r711, %r701, 16;
prmt.b32 %r712, %r709, %r710, %r711;
prmt.b32 %r609, %r706, %r712, %r681;
prmt.b32 %r613, %r706, %r712, %r682;
ld.global.nc.u32 %r713, [%rd31];
and.b32 %r714, %r713, -2004318072;
shr.u32 %r715, %r714, 1;
or.b32 %r716, %r715, 839922192;
and.b32 %r717, %r713, 30583;
prmt.b32 %r718, %r668, %r667, %r717;
prmt.b32 %r719, %r671, %r670, %r717;
and.b32 %r720, %r716, 30292;
prmt.b32 %r721, %r718, %r719, %r720;
shr.u32 %r722, %r713, 16;
and.b32 %r723, %r722, 30583;
prmt.b32 %r724, %r668, %r667, %r723;
prmt.b32 %r725, %r671, %r670, %r723;
shr.u32 %r726, %r716, 16;
prmt.b32 %r727, %r724, %r725, %r726;
prmt.b32 %r617, %r721, %r727, %r681;
prmt.b32 %r621, %r721, %r727, %r682;
add.s64 %rd36, %rd30, %rd4;
mul.lo.s64 %rd37, %rd36, 36;
add.s64 %rd38, %rd11, %rd37;
ld.global.nc.u32 %r106, [%rd38+20];

dp4a.s32.s32 %r104, %r609, %r106, %r647;

ld.global.nc.u32 %r110, [%rd38+28];

dp4a.s32.s32 %r108, %r613, %r110, %r104;

ld.global.nc.u32 %r114, [%rd38+24];

dp4a.s32.s32 %r112, %r617, %r114, %r108;

ld.global.nc.u32 %r118, [%rd38+32];

dp4a.s32.s32 %r116, %r621, %r118, %r112;

add.s64 %rd39, %rd168, %rd10;
ld.global.nc.u8 %rs68, [%rd39];
and.b16 %rs69, %rs68, 127;
setp.eq.s16 %p4, %rs69, 127;
selp.b16 %rs70, 0, %rs68, %p4;
and.b16 %rs3, %rs70, 255;

{cvt.rn.f16x2.e4m3x2 %r120, %rs3;}


cvt.u16.u32 %rs4, %r120;

{cvt.f32.f16 %f99, %rs4;}


div.approx.ftz.f32 %f166, %f99, %f161;
ld.global.nc.u32 %r121, [%rd38];

{.reg .f16 low,high;
mov.b32 {low,high},%r121;
cvt.f32.f16 %f100, low;}


mul.ftz.f32 %f167, %f166, %f100;
cvt.rn.f32.s32 %f168, %r116;
fma.rn.ftz.f32 %f169, %f167, %f168, %f165;
add.ftz.f32 %f497, %f169, %f497;
add.s64 %rd40, %rd168, %rd9;
ld.global.nc.u32 %r728, [%rd40+-4];
and.b32 %r729, %r728, -2004318072;
shr.u32 %r730, %r729, 1;
or.b32 %r731, %r730, 839922192;
and.b32 %r732, %r728, 30583;
prmt.b32 %r733, %r668, %r667, %r732;
prmt.b32 %r734, %r671, %r670, %r732;
and.b32 %r735, %r731, 30292;
prmt.b32 %r736, %r733, %r734, %r735;
shr.u32 %r737, %r728, 16;
and.b32 %r738, %r737, 30583;
prmt.b32 %r739, %r668, %r667, %r738;
prmt.b32 %r740, %r671, %r670, %r738;
shr.u32 %r741, %r731, 16;
prmt.b32 %r742, %r739, %r740, %r741;
prmt.b32 %r627, %r736, %r742, %r681;
prmt.b32 %r631, %r736, %r742, %r682;
ld.global.nc.u32 %r743, [%rd40];
and.b32 %r744, %r743, -2004318072;
shr.u32 %r745, %r744, 1;
or.b32 %r746, %r745, 839922192;
and.b32 %r747, %r743, 30583;
prmt.b32 %r748, %r668, %r667, %r747;
prmt.b32 %r749, %r671, %r670, %r747;
and.b32 %r750, %r746, 30292;
prmt.b32 %r751, %r748, %r749, %r750;
shr.u32 %r752, %r743, 16;
and.b32 %r753, %r752, 30583;
prmt.b32 %r754, %r668, %r667, %r753;
prmt.b32 %r755, %r671, %r670, %r753;
shr.u32 %r756, %r746, 16;
prmt.b32 %r757, %r754, %r755, %r756;
prmt.b32 %r635, %r751, %r757, %r681;
prmt.b32 %r639, %r751, %r757, %r682;

dp4a.s32.s32 %r122, %r627, %r88, %r647;


dp4a.s32.s32 %r126, %r631, %r92, %r122;


dp4a.s32.s32 %r130, %r635, %r96, %r126;


dp4a.s32.s32 %r134, %r639, %r100, %r130;

add.s64 %rd41, %rd168, %rd6;
ld.global.nc.u8 %rs71, [%rd41];
and.b16 %rs72, %rs71, 127;
setp.eq.s16 %p5, %rs72, 127;
selp.b16 %rs73, 0, %rs71, %p5;
and.b16 %rs5, %rs73, 255;

{cvt.rn.f16x2.e4m3x2 %r138, %rs5;}


cvt.u16.u32 %rs6, %r138;

{cvt.f32.f16 %f101, %rs6;}


div.approx.ftz.f32 %f170, %f101, %f161;

{.reg .f16 low,high;
mov.b32 {low,high},%r103;
cvt.f32.f16 %f102, low;}


mul.ftz.f32 %f171, %f170, %f102;
cvt.rn.f32.s32 %f172, %r134;
fma.rn.ftz.f32 %f173, %f171, %f172, 0f00000000;
ld.global.nc.u32 %r758, [%rd40+4];
and.b32 %r759, %r758, -2004318072;
shr.u32 %r760, %r759, 1;
or.b32 %r761, %r760, 839922192;
and.b32 %r762, %r758, 30583;
prmt.b32 %r763, %r668, %r667, %r762;
prmt.b32 %r764, %r671, %r670, %r762;
and.b32 %r765, %r761, 30292;
prmt.b32 %r766, %r763, %r764, %r765;
shr.u32 %r767, %r758, 16;
and.b32 %r768, %r767, 30583;
prmt.b32 %r769, %r668, %r667, %r768;
prmt.b32 %r770, %r671, %r670, %r768;
shr.u32 %r771, %r761, 16;
prmt.b32 %r772, %r769, %r770, %r771;
prmt.b32 %r645, %r766, %r772, %r681;
prmt.b32 %r649, %r766, %r772, %r682;
ld.global.nc.u32 %r773, [%rd40+8];
and.b32 %r774, %r773, -2004318072;
shr.u32 %r775, %r774, 1;
or.b32 %r776, %r775, 839922192;
and.b32 %r777, %r773, 30583;
prmt.b32 %r778, %r668, %r667, %r777;
prmt.b32 %r779, %r671, %r670, %r777;
and.b32 %r780, %r776, 30292;
prmt.b32 %r781, %r778, %r779, %r780;
shr.u32 %r782, %r773, 16;
and.b32 %r783, %r782, 30583;
prmt.b32 %r784, %r668, %r667, %r783;
prmt.b32 %r785, %r671, %r670, %r783;
shr.u32 %r786, %r776, 16;
prmt.b32 %r787, %r784, %r785, %r786;
prmt.b32 %r653, %r781, %r787, %r681;
prmt.b32 %r657, %r781, %r787, %r682;

dp4a.s32.s32 %r140, %r645, %r106, %r647;


dp4a.s32.s32 %r144, %r649, %r110, %r140;


dp4a.s32.s32 %r148, %r653, %r114, %r144;


dp4a.s32.s32 %r152, %r657, %r118, %r148;

add.s64 %rd42, %rd168, %rd5;
ld.global.nc.u8 %rs74, [%rd42];
and.b16 %rs75, %rs74, 127;
setp.eq.s16 %p6, %rs75, 127;
selp.b16 %rs76, 0, %rs74, %p6;
and.b16 %rs7, %rs76, 255;

{cvt.rn.f16x2.e4m3x2 %r156, %rs7;}


cvt.u16.u32 %rs8, %r156;

{cvt.f32.f16 %f103, %rs8;}


div.approx.ftz.f32 %f174, %f103, %f161;

{.reg .f16 low,high;
mov.b32 {low,high},%r121;
cvt.f32.f16 %f104, low;}


mul.ftz.f32 %f175, %f174, %f104;
cvt.rn.f32.s32 %f176, %r152;
fma.rn.ftz.f32 %f177, %f175, %f176, %f173;
add.ftz.f32 %f496, %f177, %f496;
cvt.u64.u32 %rd43, %r1232;
add.s64 %rd44, %rd1, %rd43;
add.s64 %rd45, %rd44, %rd3;
mul.lo.s64 %rd46, %rd45, 36;
add.s64 %rd47, %rd11, %rd46;
ld.global.nc.u32 %r160, [%rd47+4];

dp4a.s32.s32 %r158, %r591, %r160, %r647;

ld.global.nc.u32 %r164, [%rd47+12];

dp4a.s32.s32 %r162, %r595, %r164, %r158;

ld.global.nc.u32 %r168, [%rd47+8];

dp4a.s32.s32 %r166, %r599, %r168, %r162;

ld.global.nc.u32 %r172, [%rd47+16];

dp4a.s32.s32 %r170, %r603, %r172, %r166;


{cvt.rn.f16x2.e4m3x2 %r174, %rs1;}


cvt.u16.u32 %rs10, %r174;

{cvt.f32.f16 %f105, %rs10;}


div.approx.ftz.f32 %f178, %f105, %f161;
ld.global.nc.u32 %r175, [%rd47];

{.reg .f16 low,high;
mov.b32 {low,high},%r175;
cvt.f32.f16 %f106, low;}


mul.ftz.f32 %f179, %f178, %f106;
cvt.rn.f32.s32 %f180, %r170;
fma.rn.ftz.f32 %f181, %f179, %f180, 0f00000000;
add.s64 %rd48, %rd44, %rd4;
mul.lo.s64 %rd49, %rd48, 36;
add.s64 %rd50, %rd11, %rd49;
ld.global.nc.u32 %r178, [%rd50+20];

dp4a.s32.s32 %r176, %r609, %r178, %r647;

ld.global.nc.u32 %r182, [%rd50+28];

dp4a.s32.s32 %r180, %r613, %r182, %r176;

ld.global.nc.u32 %r186, [%rd50+24];

dp4a.s32.s32 %r184, %r617, %r186, %r180;

ld.global.nc.u32 %r190, [%rd50+32];

dp4a.s32.s32 %r188, %r621, %r190, %r184;


{cvt.rn.f16x2.e4m3x2 %r192, %rs3;}


cvt.u16.u32 %rs12, %r192;

{cvt.f32.f16 %f107, %rs12;}


div.approx.ftz.f32 %f182, %f107, %f161;
ld.global.nc.u32 %r193, [%rd50];

{.reg .f16 low,high;
mov.b32 {low,high},%r193;
cvt.f32.f16 %f108, low;}


mul.ftz.f32 %f183, %f182, %f108;
cvt.rn.f32.s32 %f184, %r188;
fma.rn.ftz.f32 %f185, %f183, %f184, %f181;
add.ftz.f32 %f495, %f185, %f495;

dp4a.s32.s32 %r194, %r627, %r160, %r647;


dp4a.s32.s32 %r198, %r631, %r164, %r194;


dp4a.s32.s32 %r202, %r635, %r168, %r198;


dp4a.s32.s32 %r206, %r639, %r172, %r202;


{cvt.rn.f16x2.e4m3x2 %r210, %rs5;}


cvt.u16.u32 %rs14, %r210;

{cvt.f32.f16 %f109, %rs14;}


div.approx.ftz.f32 %f186, %f109, %f161;

{.reg .f16 low,high;
mov.b32 {low,high},%r175;
cvt.f32.f16 %f110, low;}


mul.ftz.f32 %f187, %f186, %f110;
cvt.rn.f32.s32 %f188, %r206;
fma.rn.ftz.f32 %f189, %f187, %f188, 0f00000000;

dp4a.s32.s32 %r212, %r645, %r178, %r647;


dp4a.s32.s32 %r216, %r649, %r182, %r212;


dp4a.s32.s32 %r220, %r653, %r186, %r216;


dp4a.s32.s32 %r224, %r657, %r190, %r220;


{cvt.rn.f16x2.e4m3x2 %r228, %rs7;}


cvt.u16.u32 %rs16, %r228;

{cvt.f32.f16 %f111, %rs16;}


div.approx.ftz.f32 %f190, %f111, %f161;

{.reg .f16 low,high;
mov.b32 {low,high},%r193;
cvt.f32.f16 %f112, low;}


mul.ftz.f32 %f191, %f190, %f112;
cvt.rn.f32.s32 %f192, %r224;
fma.rn.ftz.f32 %f193, %f191, %f192, %f189;
add.ftz.f32 %f494, %f193, %f494;
cvt.u64.u32 %rd51, %r1238;
add.s64 %rd52, %rd1, %rd51;
add.s64 %rd53, %rd52, %rd3;
mul.lo.s64 %rd54, %rd53, 36;
add.s64 %rd55, %rd11, %rd54;
ld.global.nc.u32 %r232, [%rd55+4];

dp4a.s32.s32 %r230, %r591, %r232, %r647;

ld.global.nc.u32 %r236, [%rd55+12];

dp4a.s32.s32 %r234, %r595, %r236, %r230;

ld.global.nc.u32 %r240, [%rd55+8];

dp4a.s32.s32 %r238, %r599, %r240, %r234;

ld.global.nc.u32 %r244, [%rd55+16];

dp4a.s32.s32 %r242, %r603, %r244, %r238;


{cvt.rn.f16x2.e4m3x2 %r246, %rs1;}


cvt.u16.u32 %rs18, %r246;

{cvt.f32.f16 %f113, %rs18;}


div.approx.ftz.f32 %f194, %f113, %f161;
ld.global.nc.u32 %r247, [%rd55];

{.reg .f16 low,high;
mov.b32 {low,high},%r247;
cvt.f32.f16 %f114, low;}


mul.ftz.f32 %f195, %f194, %f114;
cvt.rn.f32.s32 %f196, %r242;
fma.rn.ftz.f32 %f197, %f195, %f196, 0f00000000;
add.s64 %rd56, %rd52, %rd4;
mul.lo.s64 %rd57, %rd56, 36;
add.s64 %rd58, %rd11, %rd57;
ld.global.nc.u32 %r250, [%rd58+20];

dp4a.s32.s32 %r248, %r609, %r250, %r647;

ld.global.nc.u32 %r254, [%rd58+28];

dp4a.s32.s32 %r252, %r613, %r254, %r248;

ld.global.nc.u32 %r258, [%rd58+24];

dp4a.s32.s32 %r256, %r617, %r258, %r252;

ld.global.nc.u32 %r262, [%rd58+32];

dp4a.s32.s32 %r260, %r621, %r262, %r256;


{cvt.rn.f16x2.e4m3x2 %r264, %rs3;}


cvt.u16.u32 %rs20, %r264;

{cvt.f32.f16 %f115, %rs20;}


div.approx.ftz.f32 %f198, %f115, %f161;
ld.global.nc.u32 %r265, [%rd58];

{.reg .f16 low,high;
mov.b32 {low,high},%r265;
cvt.f32.f16 %f116, low;}


mul.ftz.f32 %f199, %f198, %f116;
cvt.rn.f32.s32 %f200, %r260;
fma.rn.ftz.f32 %f201, %f199, %f200, %f197;
add.ftz.f32 %f493, %f201, %f493;

dp4a.s32.s32 %r266, %r627, %r232, %r647;


dp4a.s32.s32 %r270, %r631, %r236, %r266;


dp4a.s32.s32 %r274, %r635, %r240, %r270;


dp4a.s32.s32 %r278, %r639, %r244, %r274;


{cvt.rn.f16x2.e4m3x2 %r282, %rs5;}


cvt.u16.u32 %rs22, %r282;

{cvt.f32.f16 %f117, %rs22;}


div.approx.ftz.f32 %f202, %f117, %f161;

{.reg .f16 low,high;
mov.b32 {low,high},%r247;
cvt.f32.f16 %f118, low;}


mul.ftz.f32 %f203, %f202, %f118;
cvt.rn.f32.s32 %f204, %r278;
fma.rn.ftz.f32 %f205, %f203, %f204, 0f00000000;

dp4a.s32.s32 %r284, %r645, %r250, %r647;


dp4a.s32.s32 %r288, %r649, %r254, %r284;


dp4a.s32.s32 %r292, %r653, %r258, %r288;


dp4a.s32.s32 %r296, %r657, %r262, %r292;


{cvt.rn.f16x2.e4m3x2 %r300, %rs7;}


cvt.u16.u32 %rs24, %r300;

{cvt.f32.f16 %f119, %rs24;}


div.approx.ftz.f32 %f206, %f119, %f161;

{.reg .f16 low,high;
mov.b32 {low,high},%r265;
cvt.f32.f16 %f120, low;}


mul.ftz.f32 %f207, %f206, %f120;
cvt.rn.f32.s32 %f208, %r296;
fma.rn.ftz.f32 %f209, %f207, %f208, %f205;
add.ftz.f32 %f492, %f209, %f492;
cvt.u64.u32 %rd59, %r1237;
add.s64 %rd60, %rd1, %rd59;
add.s64 %rd61, %rd60, %rd3;
mul.lo.s64 %rd62, %rd61, 36;
add.s64 %rd63, %rd11, %rd62;
ld.global.nc.u32 %r304, [%rd63+4];

dp4a.s32.s32 %r302, %r591, %r304, %r647;

ld.global.nc.u32 %r308, [%rd63+12];

dp4a.s32.s32 %r306, %r595, %r308, %r302;

ld.global.nc.u32 %r312, [%rd63+8];

dp4a.s32.s32 %r310, %r599, %r312, %r306;

ld.global.nc.u32 %r316, [%rd63+16];

dp4a.s32.s32 %r314, %r603, %r316, %r310;


{cvt.rn.f16x2.e4m3x2 %r318, %rs1;}


cvt.u16.u32 %rs26, %r318;

{cvt.f32.f16 %f121, %rs26;}


div.approx.ftz.f32 %f210, %f121, %f161;
ld.global.nc.u32 %r319, [%rd63];

{.reg .f16 low,high;
mov.b32 {low,high},%r319;
cvt.f32.f16 %f122, low;}


mul.ftz.f32 %f211, %f210, %f122;
cvt.rn.f32.s32 %f212, %r314;
fma.rn.ftz.f32 %f213, %f211, %f212, 0f00000000;
add.s64 %rd64, %rd60, %rd4;
mul.lo.s64 %rd65, %rd64, 36;
add.s64 %rd66, %rd11, %rd65;
ld.global.nc.u32 %r322, [%rd66+20];

dp4a.s32.s32 %r320, %r609, %r322, %r647;

ld.global.nc.u32 %r326, [%rd66+28];

dp4a.s32.s32 %r324, %r613, %r326, %r320;

ld.global.nc.u32 %r330, [%rd66+24];

dp4a.s32.s32 %r328, %r617, %r330, %r324;

ld.global.nc.u32 %r334, [%rd66+32];

dp4a.s32.s32 %r332, %r621, %r334, %r328;


{cvt.rn.f16x2.e4m3x2 %r336, %rs3;}


cvt.u16.u32 %rs28, %r336;

{cvt.f32.f16 %f123, %rs28;}


div.approx.ftz.f32 %f214, %f123, %f161;
ld.global.nc.u32 %r337, [%rd66];

{.reg .f16 low,high;
mov.b32 {low,high},%r337;
cvt.f32.f16 %f124, low;}


mul.ftz.f32 %f215, %f214, %f124;
cvt.rn.f32.s32 %f216, %r332;
fma.rn.ftz.f32 %f217, %f215, %f216, %f213;
add.ftz.f32 %f491, %f217, %f491;

dp4a.s32.s32 %r338, %r627, %r304, %r647;


dp4a.s32.s32 %r342, %r631, %r308, %r338;


dp4a.s32.s32 %r346, %r635, %r312, %r342;


dp4a.s32.s32 %r350, %r639, %r316, %r346;


{cvt.rn.f16x2.e4m3x2 %r354, %rs5;}


cvt.u16.u32 %rs30, %r354;

{cvt.f32.f16 %f125, %rs30;}


div.approx.ftz.f32 %f218, %f125, %f161;

{.reg .f16 low,high;
mov.b32 {low,high},%r319;
cvt.f32.f16 %f126, low;}


mul.ftz.f32 %f219, %f218, %f126;
cvt.rn.f32.s32 %f220, %r350;
fma.rn.ftz.f32 %f221, %f219, %f220, 0f00000000;

dp4a.s32.s32 %r356, %r645, %r322, %r647;


dp4a.s32.s32 %r360, %r649, %r326, %r356;


dp4a.s32.s32 %r364, %r653, %r330, %r360;


dp4a.s32.s32 %r368, %r657, %r334, %r364;


{cvt.rn.f16x2.e4m3x2 %r372, %rs7;}


cvt.u16.u32 %rs32, %r372;

{cvt.f32.f16 %f127, %rs32;}


div.approx.ftz.f32 %f222, %f127, %f161;

{.reg .f16 low,high;
mov.b32 {low,high},%r337;
cvt.f32.f16 %f128, low;}


mul.ftz.f32 %f223, %f222, %f128;
cvt.rn.f32.s32 %f224, %r368;
fma.rn.ftz.f32 %f225, %f223, %f224, %f221;
add.ftz.f32 %f490, %f225, %f490;
cvt.u64.u32 %rd67, %r1236;
add.s64 %rd68, %rd1, %rd67;
add.s64 %rd69, %rd68, %rd3;
mul.lo.s64 %rd70, %rd69, 36;
add.s64 %rd71, %rd11, %rd70;
ld.global.nc.u32 %r376, [%rd71+4];

dp4a.s32.s32 %r374, %r591, %r376, %r647;

ld.global.nc.u32 %r380, [%rd71+12];

dp4a.s32.s32 %r378, %r595, %r380, %r374;

ld.global.nc.u32 %r384, [%rd71+8];

dp4a.s32.s32 %r382, %r599, %r384, %r378;

ld.global.nc.u32 %r388, [%rd71+16];

dp4a.s32.s32 %r386, %r603, %r388, %r382;


{cvt.rn.f16x2.e4m3x2 %r390, %rs1;}


cvt.u16.u32 %rs34, %r390;

{cvt.f32.f16 %f129, %rs34;}


div.approx.ftz.f32 %f226, %f129, %f161;
ld.global.nc.u32 %r391, [%rd71];

{.reg .f16 low,high;
mov.b32 {low,high},%r391;
cvt.f32.f16 %f130, low;}


mul.ftz.f32 %f227, %f226, %f130;
cvt.rn.f32.s32 %f228, %r386;
fma.rn.ftz.f32 %f229, %f227, %f228, 0f00000000;
add.s64 %rd72, %rd68, %rd4;
mul.lo.s64 %rd73, %rd72, 36;
add.s64 %rd74, %rd11, %rd73;
ld.global.nc.u32 %r394, [%rd74+20];

dp4a.s32.s32 %r392, %r609, %r394, %r647;

ld.global.nc.u32 %r398, [%rd74+28];

dp4a.s32.s32 %r396, %r613, %r398, %r392;

ld.global.nc.u32 %r402, [%rd74+24];

dp4a.s32.s32 %r400, %r617, %r402, %r396;

ld.global.nc.u32 %r406, [%rd74+32];

dp4a.s32.s32 %r404, %r621, %r406, %r400;


{cvt.rn.f16x2.e4m3x2 %r408, %rs3;}


cvt.u16.u32 %rs36, %r408;

{cvt.f32.f16 %f131, %rs36;}


div.approx.ftz.f32 %f230, %f131, %f161;
ld.global.nc.u32 %r409, [%rd74];

{.reg .f16 low,high;
mov.b32 {low,high},%r409;
cvt.f32.f16 %f132, low;}


mul.ftz.f32 %f231, %f230, %f132;
cvt.rn.f32.s32 %f232, %r404;
fma.rn.ftz.f32 %f233, %f231, %f232, %f229;
add.ftz.f32 %f489, %f233, %f489;

dp4a.s32.s32 %r410, %r627, %r376, %r647;


dp4a.s32.s32 %r414, %r631, %r380, %r410;


dp4a.s32.s32 %r418, %r635, %r384, %r414;


dp4a.s32.s32 %r422, %r639, %r388, %r418;


{cvt.rn.f16x2.e4m3x2 %r426, %rs5;}


cvt.u16.u32 %rs38, %r426;

{cvt.f32.f16 %f133, %rs38;}


div.approx.ftz.f32 %f234, %f133, %f161;

{.reg .f16 low,high;
mov.b32 {low,high},%r391;
cvt.f32.f16 %f134, low;}


mul.ftz.f32 %f235, %f234, %f134;
cvt.rn.f32.s32 %f236, %r422;
fma.rn.ftz.f32 %f237, %f235, %f236, 0f00000000;

dp4a.s32.s32 %r428, %r645, %r394, %r647;


dp4a.s32.s32 %r432, %r649, %r398, %r428;


dp4a.s32.s32 %r436, %r653, %r402, %r432;


dp4a.s32.s32 %r440, %r657, %r406, %r436;


{cvt.rn.f16x2.e4m3x2 %r444, %rs7;}


cvt.u16.u32 %rs40, %r444;

{cvt.f32.f16 %f135, %rs40;}


div.approx.ftz.f32 %f238, %f135, %f161;

{.reg .f16 low,high;
mov.b32 {low,high},%r409;
cvt.f32.f16 %f136, low;}


mul.ftz.f32 %f239, %f238, %f136;
cvt.rn.f32.s32 %f240, %r440;
fma.rn.ftz.f32 %f241, %f239, %f240, %f237;
add.ftz.f32 %f488, %f241, %f488;
cvt.u64.u32 %rd75, %r1235;
add.s64 %rd76, %rd1, %rd75;
add.s64 %rd77, %rd76, %rd3;
mul.lo.s64 %rd78, %rd77, 36;
add.s64 %rd79, %rd11, %rd78;
ld.global.nc.u32 %r448, [%rd79+4];

dp4a.s32.s32 %r446, %r591, %r448, %r647;

ld.global.nc.u32 %r452, [%rd79+12];

dp4a.s32.s32 %r450, %r595, %r452, %r446;

ld.global.nc.u32 %r456, [%rd79+8];

dp4a.s32.s32 %r454, %r599, %r456, %r450;

ld.global.nc.u32 %r460, [%rd79+16];

dp4a.s32.s32 %r458, %r603, %r460, %r454;


{cvt.rn.f16x2.e4m3x2 %r462, %rs1;}


cvt.u16.u32 %rs42, %r462;

{cvt.f32.f16 %f137, %rs42;}


div.approx.ftz.f32 %f242, %f137, %f161;
ld.global.nc.u32 %r463, [%rd79];

{.reg .f16 low,high;
mov.b32 {low,high},%r463;
cvt.f32.f16 %f138, low;}


mul.ftz.f32 %f243, %f242, %f138;
cvt.rn.f32.s32 %f244, %r458;
fma.rn.ftz.f32 %f245, %f243, %f244, 0f00000000;
add.s64 %rd80, %rd76, %rd4;
mul.lo.s64 %rd81, %rd80, 36;
add.s64 %rd82, %rd11, %rd81;
ld.global.nc.u32 %r466, [%rd82+20];

dp4a.s32.s32 %r464, %r609, %r466, %r647;

ld.global.nc.u32 %r470, [%rd82+28];

dp4a.s32.s32 %r468, %r613, %r470, %r464;

ld.global.nc.u32 %r474, [%rd82+24];

dp4a.s32.s32 %r472, %r617, %r474, %r468;

ld.global.nc.u32 %r478, [%rd82+32];

dp4a.s32.s32 %r476, %r621, %r478, %r472;


{cvt.rn.f16x2.e4m3x2 %r480, %rs3;}


cvt.u16.u32 %rs44, %r480;

{cvt.f32.f16 %f139, %rs44;}


div.approx.ftz.f32 %f246, %f139, %f161;
ld.global.nc.u32 %r481, [%rd82];

{.reg .f16 low,high;
mov.b32 {low,high},%r481;
cvt.f32.f16 %f140, low;}


mul.ftz.f32 %f247, %f246, %f140;
cvt.rn.f32.s32 %f248, %r476;
fma.rn.ftz.f32 %f249, %f247, %f248, %f245;
add.ftz.f32 %f487, %f249, %f487;

dp4a.s32.s32 %r482, %r627, %r448, %r647;


dp4a.s32.s32 %r486, %r631, %r452, %r482;


dp4a.s32.s32 %r490, %r635, %r456, %r486;


dp4a.s32.s32 %r494, %r639, %r460, %r490;


{cvt.rn.f16x2.e4m3x2 %r498, %rs5;}


cvt.u16.u32 %rs46, %r498;

{cvt.f32.f16 %f141, %rs46;}


div.approx.ftz.f32 %f250, %f141, %f161;

{.reg .f16 low,high;
mov.b32 {low,high},%r463;
cvt.f32.f16 %f142, low;}


mul.ftz.f32 %f251, %f250, %f142;
cvt.rn.f32.s32 %f252, %r494;
fma.rn.ftz.f32 %f253, %f251, %f252, 0f00000000;

dp4a.s32.s32 %r500, %r645, %r466, %r647;


dp4a.s32.s32 %r504, %r649, %r470, %r500;


dp4a.s32.s32 %r508, %r653, %r474, %r504;


dp4a.s32.s32 %r512, %r657, %r478, %r508;


{cvt.rn.f16x2.e4m3x2 %r516, %rs7;}


cvt.u16.u32 %rs48, %r516;

{cvt.f32.f16 %f143, %rs48;}


div.approx.ftz.f32 %f254, %f143, %f161;

{.reg .f16 low,high;
mov.b32 {low,high},%r481;
cvt.f32.f16 %f144, low;}


mul.ftz.f32 %f255, %f254, %f144;
cvt.rn.f32.s32 %f256, %r512;
fma.rn.ftz.f32 %f257, %f255, %f256, %f253;
add.ftz.f32 %f486, %f257, %f486;
cvt.u64.u32 %rd83, %r1234;
add.s64 %rd84, %rd1, %rd83;
add.s64 %rd85, %rd84, %rd3;
mul.lo.s64 %rd86, %rd85, 36;
add.s64 %rd87, %rd11, %rd86;
ld.global.nc.u32 %r520, [%rd87+4];

dp4a.s32.s32 %r518, %r591, %r520, %r647;

ld.global.nc.u32 %r524, [%rd87+12];

dp4a.s32.s32 %r522, %r595, %r524, %r518;

ld.global.nc.u32 %r528, [%rd87+8];

dp4a.s32.s32 %r526, %r599, %r528, %r522;

ld.global.nc.u32 %r532, [%rd87+16];

dp4a.s32.s32 %r530, %r603, %r532, %r526;


{cvt.rn.f16x2.e4m3x2 %r534, %rs1;}


cvt.u16.u32 %rs50, %r534;

{cvt.f32.f16 %f145, %rs50;}


div.approx.ftz.f32 %f258, %f145, %f161;
ld.global.nc.u32 %r535, [%rd87];

{.reg .f16 low,high;
mov.b32 {low,high},%r535;
cvt.f32.f16 %f146, low;}


mul.ftz.f32 %f259, %f258, %f146;
cvt.rn.f32.s32 %f260, %r530;
fma.rn.ftz.f32 %f261, %f259, %f260, 0f00000000;
add.s64 %rd88, %rd84, %rd4;
mul.lo.s64 %rd89, %rd88, 36;
add.s64 %rd90, %rd11, %rd89;
ld.global.nc.u32 %r538, [%rd90+20];

dp4a.s32.s32 %r536, %r609, %r538, %r647;

ld.global.nc.u32 %r542, [%rd90+28];

dp4a.s32.s32 %r540, %r613, %r542, %r536;

ld.global.nc.u32 %r546, [%rd90+24];

dp4a.s32.s32 %r544, %r617, %r546, %r540;

ld.global.nc.u32 %r550, [%rd90+32];

dp4a.s32.s32 %r548, %r621, %r550, %r544;


{cvt.rn.f16x2.e4m3x2 %r552, %rs3;}


cvt.u16.u32 %rs52, %r552;

{cvt.f32.f16 %f147, %rs52;}


div.approx.ftz.f32 %f262, %f147, %f161;
ld.global.nc.u32 %r553, [%rd90];

{.reg .f16 low,high;
mov.b32 {low,high},%r553;
cvt.f32.f16 %f148, low;}


mul.ftz.f32 %f263, %f262, %f148;
cvt.rn.f32.s32 %f264, %r548;
fma.rn.ftz.f32 %f265, %f263, %f264, %f261;
add.ftz.f32 %f485, %f265, %f485;

dp4a.s32.s32 %r554, %r627, %r520, %r647;


dp4a.s32.s32 %r558, %r631, %r524, %r554;


dp4a.s32.s32 %r562, %r635, %r528, %r558;


dp4a.s32.s32 %r566, %r639, %r532, %r562;


{cvt.rn.f16x2.e4m3x2 %r570, %rs5;}


cvt.u16.u32 %rs54, %r570;

{cvt.f32.f16 %f149, %rs54;}


div.approx.ftz.f32 %f266, %f149, %f161;

{.reg .f16 low,high;
mov.b32 {low,high},%r535;
cvt.f32.f16 %f150, low;}


mul.ftz.f32 %f267, %f266, %f150;
cvt.rn.f32.s32 %f268, %r566;
fma.rn.ftz.f32 %f269, %f267, %f268, 0f00000000;

dp4a.s32.s32 %r572, %r645, %r538, %r647;


dp4a.s32.s32 %r576, %r649, %r542, %r572;


dp4a.s32.s32 %r580, %r653, %r546, %r576;


dp4a.s32.s32 %r584, %r657, %r550, %r580;


{cvt.rn.f16x2.e4m3x2 %r588, %rs7;}


cvt.u16.u32 %rs56, %r588;

{cvt.f32.f16 %f151, %rs56;}


div.approx.ftz.f32 %f270, %f151, %f161;

{.reg .f16 low,high;
mov.b32 {low,high},%r553;
cvt.f32.f16 %f152, low;}


mul.ftz.f32 %f271, %f270, %f152;
cvt.rn.f32.s32 %f272, %r584;
fma.rn.ftz.f32 %f273, %f271, %f272, %f269;
add.ftz.f32 %f484, %f273, %f484;
cvt.u64.u32 %rd91, %r1233;
add.s64 %rd92, %rd1, %rd91;
add.s64 %rd93, %rd92, %rd3;
mul.lo.s64 %rd94, %rd93, 36;
add.s64 %rd95, %rd11, %rd94;
ld.global.nc.u32 %r592, [%rd95+4];

dp4a.s32.s32 %r590, %r591, %r592, %r647;

ld.global.nc.u32 %r596, [%rd95+12];

dp4a.s32.s32 %r594, %r595, %r596, %r590;

ld.global.nc.u32 %r600, [%rd95+8];

dp4a.s32.s32 %r598, %r599, %r600, %r594;

ld.global.nc.u32 %r604, [%rd95+16];

dp4a.s32.s32 %r602, %r603, %r604, %r598;


{cvt.rn.f16x2.e4m3x2 %r606, %rs1;}


cvt.u16.u32 %rs58, %r606;

{cvt.f32.f16 %f153, %rs58;}


div.approx.ftz.f32 %f274, %f153, %f161;
ld.global.nc.u32 %r607, [%rd95];

{.reg .f16 low,high;
mov.b32 {low,high},%r607;
cvt.f32.f16 %f154, low;}


mul.ftz.f32 %f275, %f274, %f154;
cvt.rn.f32.s32 %f276, %r602;
fma.rn.ftz.f32 %f277, %f275, %f276, 0f00000000;
add.s64 %rd96, %rd92, %rd4;
mul.lo.s64 %rd97, %rd96, 36;
add.s64 %rd98, %rd11, %rd97;
ld.global.nc.u32 %r610, [%rd98+20];

dp4a.s32.s32 %r608, %r609, %r610, %r647;

ld.global.nc.u32 %r614, [%rd98+28];

dp4a.s32.s32 %r612, %r613, %r614, %r608;

ld.global.nc.u32 %r618, [%rd98+24];

dp4a.s32.s32 %r616, %r617, %r618, %r612;

ld.global.nc.u32 %r622, [%rd98+32];

dp4a.s32.s32 %r620, %r621, %r622, %r616;


{cvt.rn.f16x2.e4m3x2 %r624, %rs3;}


cvt.u16.u32 %rs60, %r624;

{cvt.f32.f16 %f155, %rs60;}


div.approx.ftz.f32 %f278, %f155, %f161;
ld.global.nc.u32 %r625, [%rd98];

{.reg .f16 low,high;
mov.b32 {low,high},%r625;
cvt.f32.f16 %f156, low;}


mul.ftz.f32 %f279, %f278, %f156;
cvt.rn.f32.s32 %f280, %r620;
fma.rn.ftz.f32 %f281, %f279, %f280, %f277;
add.ftz.f32 %f483, %f281, %f483;

dp4a.s32.s32 %r626, %r627, %r592, %r647;


dp4a.s32.s32 %r630, %r631, %r596, %r626;


dp4a.s32.s32 %r634, %r635, %r600, %r630;


dp4a.s32.s32 %r638, %r639, %r604, %r634;


{cvt.rn.f16x2.e4m3x2 %r642, %rs5;}


cvt.u16.u32 %rs62, %r642;

{cvt.f32.f16 %f157, %rs62;}


div.approx.ftz.f32 %f282, %f157, %f161;

{.reg .f16 low,high;
mov.b32 {low,high},%r607;
cvt.f32.f16 %f158, low;}


mul.ftz.f32 %f283, %f282, %f158;
cvt.rn.f32.s32 %f284, %r638;
fma.rn.ftz.f32 %f285, %f283, %f284, 0f00000000;

dp4a.s32.s32 %r644, %r645, %r610, %r647;


dp4a.s32.s32 %r648, %r649, %r614, %r644;


dp4a.s32.s32 %r652, %r653, %r618, %r648;


dp4a.s32.s32 %r656, %r657, %r622, %r652;


{cvt.rn.f16x2.e4m3x2 %r660, %rs7;}


cvt.u16.u32 %rs64, %r660;

{cvt.f32.f16 %f159, %rs64;}


div.approx.ftz.f32 %f286, %f159, %f161;

{.reg .f16 low,high;
mov.b32 {low,high},%r625;
cvt.f32.f16 %f160, low;}


mul.ftz.f32 %f287, %f286, %f160;
cvt.rn.f32.s32 %f288, %r656;
fma.rn.ftz.f32 %f289, %f287, %f288, %f285;
add.ftz.f32 %f482, %f289, %f482;
add.s32 %r1238, %r1238, 64;
add.s32 %r1237, %r1237, 64;
add.s32 %r1236, %r1236, 64;
add.s32 %r1235, %r1235, 64;
add.s32 %r1234, %r1234, 64;
add.s32 %r1233, %r1233, 64;
add.s32 %r1232, %r1232, 64;
add.s32 %r1231, %r1231, 64;
add.s64 %rd168, %rd168, 1152;
add.s32 %r1239, %r1239, 32;
setp.lt.s32 %p7, %r1239, %r53;
@%p7 bra $L__BB126_2;

$L__BB126_3:
setp.eq.s32 %p8, %r1, 0;
@%p8 bra $L__BB126_5;

shl.b32 %r791, %r1, 11;
mov.u32 %r792, _ZZ13mul_mat_vec_qIL9ggml_type40ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r793, %r792, %r791;
shl.b32 %r794, %r50, 2;
add.s32 %r795, %r793, %r794;
st.shared.f32 [%r795+-2048], %f497;
st.shared.f32 [%r795+-1920], %f496;
st.shared.f32 [%r795+-1792], %f495;
st.shared.f32 [%r795+-1664], %f494;
st.shared.f32 [%r795+-1536], %f493;
st.shared.f32 [%r795+-1408], %f492;
st.shared.f32 [%r795+-1280], %f491;
st.shared.f32 [%r795+-1152], %f490;
st.shared.f32 [%r795+-1024], %f489;
st.shared.f32 [%r795+-896], %f488;
st.shared.f32 [%r795+-768], %f487;
st.shared.f32 [%r795+-640], %f486;
st.shared.f32 [%r795+-512], %f485;
st.shared.f32 [%r795+-384], %f484;
st.shared.f32 [%r795+-256], %f483;
st.shared.f32 [%r795+-128], %f482;

$L__BB126_5:
setp.ne.s32 %p9, %r1, 0;
bar.sync 0;
@%p9 bra $L__BB126_38;

mov.u32 %r797, %ctaid.x;
shl.b32 %r798, %r797, 1;
mov.u32 %r799, 1;
shl.b32 %r801, %r50, 2;
mov.u32 %r802, _ZZ13mul_mat_vec_qIL9ggml_type40ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r803, %r802, %r801;
mov.u32 %r804, 2;
ld.shared.f32 %f290, [%r803];
add.ftz.f32 %f291, %f290, %f497;
mov.b32 %r805, %f291;
mov.u32 %r806, 31;
mov.u32 %r807, 16;
mov.u32 %r808, -1;
shfl.sync.bfly.b32 %r809|%p10, %r805, %r807, %r806, %r808;
mov.b32 %f292, %r809;
add.ftz.f32 %f293, %f291, %f292;
mov.b32 %r810, %f293;
mov.u32 %r811, 8;
shfl.sync.bfly.b32 %r812|%p11, %r810, %r811, %r806, %r808;
mov.b32 %f294, %r812;
add.ftz.f32 %f295, %f293, %f294;
mov.b32 %r813, %f295;
mov.u32 %r814, 4;
shfl.sync.bfly.b32 %r815|%p12, %r813, %r814, %r806, %r808;
mov.b32 %f296, %r815;
add.ftz.f32 %f297, %f295, %f296;
mov.b32 %r816, %f297;
shfl.sync.bfly.b32 %r817|%p13, %r816, %r804, %r806, %r808;
mov.b32 %f298, %r817;
add.ftz.f32 %f299, %f297, %f298;
mov.b32 %r818, %f299;
shfl.sync.bfly.b32 %r819|%p14, %r818, %r799, %r806, %r808;
mov.b32 %f300, %r819;
add.ftz.f32 %f49, %f299, %f300;
setp.ne.s32 %p15, %r50, 0;
setp.ge.u32 %p16, %r798, %r36;
or.pred %p17, %p15, %p16;
@%p17 bra $L__BB126_8;

cvta.to.global.u64 %rd99, %rd19;
mov.u32 %r820, %ctaid.y;
mad.lo.s32 %r823, %r820, %r42, %r798;
mov.u32 %r824, %ctaid.z;
mad.lo.s32 %r825, %r824, %r48, %r823;
mul.wide.u32 %rd100, %r825, 4;
add.s64 %rd101, %rd99, %rd100;
st.global.f32 [%rd101], %f49;

$L__BB126_8:
ld.shared.f32 %f301, [%r803+128];
add.ftz.f32 %f302, %f301, %f496;
mov.b32 %r831, %f302;
shfl.sync.bfly.b32 %r835|%p18, %r831, %r807, %r806, %r808;
mov.b32 %f303, %r835;
add.ftz.f32 %f304, %f302, %f303;
mov.b32 %r836, %f304;
shfl.sync.bfly.b32 %r838|%p19, %r836, %r811, %r806, %r808;
mov.b32 %f305, %r838;
add.ftz.f32 %f306, %f304, %f305;
mov.b32 %r839, %f306;
shfl.sync.bfly.b32 %r841|%p20, %r839, %r814, %r806, %r808;
mov.b32 %f307, %r841;
add.ftz.f32 %f308, %f306, %f307;
mov.b32 %r842, %f308;
shfl.sync.bfly.b32 %r843|%p21, %r842, %r804, %r806, %r808;
mov.b32 %f309, %r843;
add.ftz.f32 %f310, %f308, %f309;
mov.b32 %r844, %f310;
shfl.sync.bfly.b32 %r846|%p22, %r844, %r799, %r806, %r808;
mov.b32 %f311, %r846;
add.ftz.f32 %f50, %f310, %f311;
setp.eq.s32 %p23, %r50, 1;
or.b32 %r849, %r798, 1;
setp.lt.u32 %p24, %r849, %r36;
and.pred %p1, %p23, %p24;
not.pred %p25, %p1;
@%p25 bra $L__BB126_10;

cvta.to.global.u64 %rd102, %rd19;
mov.u32 %r850, %ctaid.y;
mad.lo.s32 %r853, %r850, %r42, %r798;
mov.u32 %r854, %ctaid.z;
mad.lo.s32 %r855, %r854, %r48, %r853;
mul.wide.u32 %rd103, %r855, 4;
add.s64 %rd104, %rd102, %rd103;
st.global.f32 [%rd104+4], %f50;

$L__BB126_10:
mov.u32 %r860, 2;
ld.shared.f32 %f312, [%r803+256];
add.ftz.f32 %f313, %f312, %f495;
mov.b32 %r861, %f313;
mov.u32 %r862, 31;
mov.u32 %r863, 16;
mov.u32 %r864, -1;
shfl.sync.bfly.b32 %r865|%p26, %r861, %r863, %r862, %r864;
mov.b32 %f314, %r865;
add.ftz.f32 %f315, %f313, %f314;
mov.b32 %r866, %f315;
mov.u32 %r867, 8;
shfl.sync.bfly.b32 %r868|%p27, %r866, %r867, %r862, %r864;
mov.b32 %f316, %r868;
add.ftz.f32 %f317, %f315, %f316;
mov.b32 %r869, %f317;
mov.u32 %r870, 4;
shfl.sync.bfly.b32 %r871|%p28, %r869, %r870, %r862, %r864;
mov.b32 %f318, %r871;
add.ftz.f32 %f319, %f317, %f318;
mov.b32 %r872, %f319;
shfl.sync.bfly.b32 %r873|%p29, %r872, %r860, %r862, %r864;
mov.b32 %f320, %r873;
add.ftz.f32 %f321, %f319, %f320;
mov.b32 %r874, %f321;
mov.u32 %r875, 1;
shfl.sync.bfly.b32 %r876|%p30, %r874, %r875, %r862, %r864;
mov.b32 %f322, %r876;
add.ftz.f32 %f51, %f321, %f322;
@%p17 bra $L__BB126_12;

cvt.u64.u32 %rd105, %r36;
mov.u32 %r879, %ctaid.y;
mad.lo.s32 %r882, %r879, %r42, %r798;
mov.u32 %r883, %ctaid.z;
mad.lo.s32 %r884, %r883, %r48, %r882;
cvt.u64.u32 %rd106, %r884;
add.s64 %rd107, %rd105, %rd106;
cvta.to.global.u64 %rd108, %rd19;
shl.b64 %rd109, %rd107, 2;
add.s64 %rd110, %rd108, %rd109;
st.global.f32 [%rd110], %f51;

$L__BB126_12:
ld.shared.f32 %f323, [%r803+384];
add.ftz.f32 %f324, %f323, %f494;
mov.b32 %r890, %f324;
shfl.sync.bfly.b32 %r894|%p34, %r890, %r863, %r862, %r864;
mov.b32 %f325, %r894;
add.ftz.f32 %f326, %f324, %f325;
mov.b32 %r895, %f326;
shfl.sync.bfly.b32 %r897|%p35, %r895, %r867, %r862, %r864;
mov.b32 %f327, %r897;
add.ftz.f32 %f328, %f326, %f327;
mov.b32 %r898, %f328;
shfl.sync.bfly.b32 %r900|%p36, %r898, %r870, %r862, %r864;
mov.b32 %f329, %r900;
add.ftz.f32 %f330, %f328, %f329;
mov.b32 %r901, %f330;
shfl.sync.bfly.b32 %r902|%p37, %r901, %r860, %r862, %r864;
mov.b32 %f331, %r902;
add.ftz.f32 %f332, %f330, %f331;
mov.b32 %r903, %f332;
shfl.sync.bfly.b32 %r905|%p38, %r903, %r875, %r862, %r864;
mov.b32 %f333, %r905;
add.ftz.f32 %f52, %f332, %f333;
@%p25 bra $L__BB126_14;

add.s32 %r906, %r36, 1;
cvt.u64.u32 %rd111, %r906;
mov.u32 %r907, %ctaid.y;
mad.lo.s32 %r910, %r907, %r42, %r798;
mov.u32 %r911, %ctaid.z;
mad.lo.s32 %r912, %r911, %r48, %r910;
cvt.u64.u32 %rd112, %r912;
add.s64 %rd113, %rd111, %rd112;
cvta.to.global.u64 %rd114, %rd19;
shl.b64 %rd115, %rd113, 2;
add.s64 %rd116, %rd114, %rd115;
st.global.f32 [%rd116], %f52;

$L__BB126_14:
shl.b32 %r913, %r36, 1;
mov.u32 %r914, 1;
mov.u32 %r919, 2;
ld.shared.f32 %f334, [%r803+512];
add.ftz.f32 %f335, %f334, %f493;
mov.b32 %r920, %f335;
mov.u32 %r921, 31;
mov.u32 %r922, 16;
mov.u32 %r923, -1;
shfl.sync.bfly.b32 %r924|%p40, %r920, %r922, %r921, %r923;
mov.b32 %f336, %r924;
add.ftz.f32 %f337, %f335, %f336;
mov.b32 %r925, %f337;
mov.u32 %r926, 8;
shfl.sync.bfly.b32 %r927|%p41, %r925, %r926, %r921, %r923;
mov.b32 %f338, %r927;
add.ftz.f32 %f339, %f337, %f338;
mov.b32 %r928, %f339;
mov.u32 %r929, 4;
shfl.sync.bfly.b32 %r930|%p42, %r928, %r929, %r921, %r923;
mov.b32 %f340, %r930;
add.ftz.f32 %f341, %f339, %f340;
mov.b32 %r931, %f341;
shfl.sync.bfly.b32 %r932|%p43, %r931, %r919, %r921, %r923;
mov.b32 %f342, %r932;
add.ftz.f32 %f343, %f341, %f342;
mov.b32 %r933, %f343;
shfl.sync.bfly.b32 %r934|%p44, %r933, %r914, %r921, %r923;
mov.b32 %f344, %r934;
add.ftz.f32 %f53, %f343, %f344;
cvt.u64.u32 %rd117, %r913;
mov.u32 %r935, %ctaid.y;
mad.lo.s32 %r938, %r935, %r42, %r798;
mov.u32 %r939, %ctaid.z;
mad.lo.s32 %r940, %r939, %r48, %r938;
cvt.u64.u32 %rd118, %r940;
add.s64 %rd119, %rd117, %rd118;
cvta.to.global.u64 %rd120, %rd19;
shl.b64 %rd121, %rd119, 2;
add.s64 %rd14, %rd120, %rd121;
@%p17 bra $L__BB126_16;

st.global.f32 [%rd14], %f53;

$L__BB126_16:
ld.shared.f32 %f345, [%r803+640];
add.ftz.f32 %f346, %f345, %f492;
mov.b32 %r946, %f346;
shfl.sync.bfly.b32 %r950|%p48, %r946, %r922, %r921, %r923;
mov.b32 %f347, %r950;
add.ftz.f32 %f348, %f346, %f347;
mov.b32 %r951, %f348;
shfl.sync.bfly.b32 %r953|%p49, %r951, %r926, %r921, %r923;
mov.b32 %f349, %r953;
add.ftz.f32 %f350, %f348, %f349;
mov.b32 %r954, %f350;
shfl.sync.bfly.b32 %r956|%p50, %r954, %r929, %r921, %r923;
mov.b32 %f351, %r956;
add.ftz.f32 %f352, %f350, %f351;
mov.b32 %r957, %f352;
shfl.sync.bfly.b32 %r958|%p51, %r957, %r919, %r921, %r923;
mov.b32 %f353, %r958;
add.ftz.f32 %f354, %f352, %f353;
mov.b32 %r959, %f354;
shfl.sync.bfly.b32 %r961|%p52, %r959, %r914, %r921, %r923;
mov.b32 %f355, %r961;
add.ftz.f32 %f54, %f354, %f355;
@%p25 bra $L__BB126_18;

st.global.f32 [%rd14+4], %f54;

$L__BB126_18:
mul.lo.s32 %r30, %r36, 3;
mov.u32 %r966, 2;
ld.shared.f32 %f356, [%r803+768];
add.ftz.f32 %f357, %f356, %f491;
mov.b32 %r967, %f357;
mov.u32 %r968, 31;
mov.u32 %r969, 16;
mov.u32 %r970, -1;
shfl.sync.bfly.b32 %r971|%p54, %r967, %r969, %r968, %r970;
mov.b32 %f358, %r971;
add.ftz.f32 %f359, %f357, %f358;
mov.b32 %r972, %f359;
mov.u32 %r973, 8;
shfl.sync.bfly.b32 %r974|%p55, %r972, %r973, %r968, %r970;
mov.b32 %f360, %r974;
add.ftz.f32 %f361, %f359, %f360;
mov.b32 %r975, %f361;
mov.u32 %r976, 4;
shfl.sync.bfly.b32 %r977|%p56, %r975, %r976, %r968, %r970;
mov.b32 %f362, %r977;
add.ftz.f32 %f363, %f361, %f362;
mov.b32 %r978, %f363;
shfl.sync.bfly.b32 %r979|%p57, %r978, %r966, %r968, %r970;
mov.b32 %f364, %r979;
add.ftz.f32 %f365, %f363, %f364;
mov.b32 %r980, %f365;
mov.u32 %r981, 1;
shfl.sync.bfly.b32 %r982|%p58, %r980, %r981, %r968, %r970;
mov.b32 %f366, %r982;
add.ftz.f32 %f55, %f365, %f366;
@%p17 bra $L__BB126_20;

cvt.u64.u32 %rd122, %r30;
add.s64 %rd124, %rd122, %rd118;
shl.b64 %rd126, %rd124, 2;
add.s64 %rd127, %rd120, %rd126;
st.global.f32 [%rd127], %f55;

$L__BB126_20:
ld.shared.f32 %f367, [%r803+896];
add.ftz.f32 %f368, %f367, %f490;
mov.b32 %r996, %f368;
shfl.sync.bfly.b32 %r1000|%p62, %r996, %r969, %r968, %r970;
mov.b32 %f369, %r1000;
add.ftz.f32 %f370, %f368, %f369;
mov.b32 %r1001, %f370;
shfl.sync.bfly.b32 %r1003|%p63, %r1001, %r973, %r968, %r970;
mov.b32 %f371, %r1003;
add.ftz.f32 %f372, %f370, %f371;
mov.b32 %r1004, %f372;
shfl.sync.bfly.b32 %r1006|%p64, %r1004, %r976, %r968, %r970;
mov.b32 %f373, %r1006;
add.ftz.f32 %f374, %f372, %f373;
mov.b32 %r1007, %f374;
shfl.sync.bfly.b32 %r1008|%p65, %r1007, %r966, %r968, %r970;
mov.b32 %f375, %r1008;
add.ftz.f32 %f376, %f374, %f375;
mov.b32 %r1009, %f376;
shfl.sync.bfly.b32 %r1011|%p66, %r1009, %r981, %r968, %r970;
mov.b32 %f377, %r1011;
add.ftz.f32 %f56, %f376, %f377;
@%p25 bra $L__BB126_22;

add.s32 %r1012, %r30, 1;
cvt.u64.u32 %rd128, %r1012;
add.s64 %rd130, %rd128, %rd118;
shl.b64 %rd132, %rd130, 2;
add.s64 %rd133, %rd120, %rd132;
st.global.f32 [%rd133], %f56;

$L__BB126_22:
shl.b32 %r1019, %r36, 2;
mov.u32 %r1020, 2;
ld.shared.f32 %f378, [%r803+1024];
add.ftz.f32 %f379, %f378, %f489;
mov.b32 %r1025, %f379;
mov.u32 %r1026, 31;
mov.u32 %r1027, 16;
mov.u32 %r1028, -1;
shfl.sync.bfly.b32 %r1029|%p68, %r1025, %r1027, %r1026, %r1028;
mov.b32 %f380, %r1029;
add.ftz.f32 %f381, %f379, %f380;
mov.b32 %r1030, %f381;
mov.u32 %r1031, 8;
shfl.sync.bfly.b32 %r1032|%p69, %r1030, %r1031, %r1026, %r1028;
mov.b32 %f382, %r1032;
add.ftz.f32 %f383, %f381, %f382;
mov.b32 %r1033, %f383;
mov.u32 %r1034, 4;
shfl.sync.bfly.b32 %r1035|%p70, %r1033, %r1034, %r1026, %r1028;
mov.b32 %f384, %r1035;
add.ftz.f32 %f385, %f383, %f384;
mov.b32 %r1036, %f385;
shfl.sync.bfly.b32 %r1037|%p71, %r1036, %r1020, %r1026, %r1028;
mov.b32 %f386, %r1037;
add.ftz.f32 %f387, %f385, %f386;
mov.b32 %r1038, %f387;
mov.u32 %r1039, 1;
shfl.sync.bfly.b32 %r1040|%p72, %r1038, %r1039, %r1026, %r1028;
mov.b32 %f388, %r1040;
add.ftz.f32 %f57, %f387, %f388;
cvt.u64.u32 %rd134, %r1019;
add.s64 %rd136, %rd134, %rd118;
shl.b64 %rd138, %rd136, 2;
add.s64 %rd15, %rd120, %rd138;
@%p17 bra $L__BB126_24;

st.global.f32 [%rd15], %f57;

$L__BB126_24:
ld.shared.f32 %f389, [%r803+1152];
add.ftz.f32 %f390, %f389, %f488;
mov.b32 %r1052, %f390;
shfl.sync.bfly.b32 %r1056|%p76, %r1052, %r1027, %r1026, %r1028;
mov.b32 %f391, %r1056;
add.ftz.f32 %f392, %f390, %f391;
mov.b32 %r1057, %f392;
shfl.sync.bfly.b32 %r1059|%p77, %r1057, %r1031, %r1026, %r1028;
mov.b32 %f393, %r1059;
add.ftz.f32 %f394, %f392, %f393;
mov.b32 %r1060, %f394;
shfl.sync.bfly.b32 %r1062|%p78, %r1060, %r1034, %r1026, %r1028;
mov.b32 %f395, %r1062;
add.ftz.f32 %f396, %f394, %f395;
mov.b32 %r1063, %f396;
shfl.sync.bfly.b32 %r1064|%p79, %r1063, %r1020, %r1026, %r1028;
mov.b32 %f397, %r1064;
add.ftz.f32 %f398, %f396, %f397;
mov.b32 %r1065, %f398;
shfl.sync.bfly.b32 %r1067|%p80, %r1065, %r1039, %r1026, %r1028;
mov.b32 %f399, %r1067;
add.ftz.f32 %f58, %f398, %f399;
@%p25 bra $L__BB126_26;

st.global.f32 [%rd15+4], %f58;

$L__BB126_26:
mul.lo.s32 %r31, %r36, 5;
mov.u32 %r1072, 2;
ld.shared.f32 %f400, [%r803+1280];
add.ftz.f32 %f401, %f400, %f487;
mov.b32 %r1073, %f401;
mov.u32 %r1074, 31;
mov.u32 %r1075, 16;
mov.u32 %r1076, -1;
shfl.sync.bfly.b32 %r1077|%p82, %r1073, %r1075, %r1074, %r1076;
mov.b32 %f402, %r1077;
add.ftz.f32 %f403, %f401, %f402;
mov.b32 %r1078, %f403;
mov.u32 %r1079, 8;
shfl.sync.bfly.b32 %r1080|%p83, %r1078, %r1079, %r1074, %r1076;
mov.b32 %f404, %r1080;
add.ftz.f32 %f405, %f403, %f404;
mov.b32 %r1081, %f405;
mov.u32 %r1082, 4;
shfl.sync.bfly.b32 %r1083|%p84, %r1081, %r1082, %r1074, %r1076;
mov.b32 %f406, %r1083;
add.ftz.f32 %f407, %f405, %f406;
mov.b32 %r1084, %f407;
shfl.sync.bfly.b32 %r1085|%p85, %r1084, %r1072, %r1074, %r1076;
mov.b32 %f408, %r1085;
add.ftz.f32 %f409, %f407, %f408;
mov.b32 %r1086, %f409;
mov.u32 %r1087, 1;
shfl.sync.bfly.b32 %r1088|%p86, %r1086, %r1087, %r1074, %r1076;
mov.b32 %f410, %r1088;
add.ftz.f32 %f59, %f409, %f410;
@%p17 bra $L__BB126_28;

cvt.u64.u32 %rd139, %r31;
add.s64 %rd141, %rd139, %rd118;
shl.b64 %rd143, %rd141, 2;
add.s64 %rd144, %rd120, %rd143;
st.global.f32 [%rd144], %f59;

$L__BB126_28:
ld.shared.f32 %f411, [%r803+1408];
add.ftz.f32 %f412, %f411, %f486;
mov.b32 %r1102, %f412;
shfl.sync.bfly.b32 %r1106|%p90, %r1102, %r1075, %r1074, %r1076;
mov.b32 %f413, %r1106;
add.ftz.f32 %f414, %f412, %f413;
mov.b32 %r1107, %f414;
shfl.sync.bfly.b32 %r1109|%p91, %r1107, %r1079, %r1074, %r1076;
mov.b32 %f415, %r1109;
add.ftz.f32 %f416, %f414, %f415;
mov.b32 %r1110, %f416;
shfl.sync.bfly.b32 %r1112|%p92, %r1110, %r1082, %r1074, %r1076;
mov.b32 %f417, %r1112;
add.ftz.f32 %f418, %f416, %f417;
mov.b32 %r1113, %f418;
shfl.sync.bfly.b32 %r1114|%p93, %r1113, %r1072, %r1074, %r1076;
mov.b32 %f419, %r1114;
add.ftz.f32 %f420, %f418, %f419;
mov.b32 %r1115, %f420;
shfl.sync.bfly.b32 %r1117|%p94, %r1115, %r1087, %r1074, %r1076;
mov.b32 %f421, %r1117;
add.ftz.f32 %f60, %f420, %f421;
@%p25 bra $L__BB126_30;

add.s32 %r1118, %r31, 1;
cvt.u64.u32 %rd145, %r1118;
add.s64 %rd147, %rd145, %rd118;
shl.b64 %rd149, %rd147, 2;
add.s64 %rd150, %rd120, %rd149;
st.global.f32 [%rd150], %f60;

$L__BB126_30:
mul.lo.s32 %r1125, %r36, 6;
mov.u32 %r1130, 2;
ld.shared.f32 %f422, [%r803+1536];
add.ftz.f32 %f423, %f422, %f485;
mov.b32 %r1131, %f423;
mov.u32 %r1132, 31;
mov.u32 %r1133, 16;
mov.u32 %r1134, -1;
shfl.sync.bfly.b32 %r1135|%p96, %r1131, %r1133, %r1132, %r1134;
mov.b32 %f424, %r1135;
add.ftz.f32 %f425, %f423, %f424;
mov.b32 %r1136, %f425;
mov.u32 %r1137, 8;
shfl.sync.bfly.b32 %r1138|%p97, %r1136, %r1137, %r1132, %r1134;
mov.b32 %f426, %r1138;
add.ftz.f32 %f427, %f425, %f426;
mov.b32 %r1139, %f427;
mov.u32 %r1140, 4;
shfl.sync.bfly.b32 %r1141|%p98, %r1139, %r1140, %r1132, %r1134;
mov.b32 %f428, %r1141;
add.ftz.f32 %f429, %f427, %f428;
mov.b32 %r1142, %f429;
shfl.sync.bfly.b32 %r1143|%p99, %r1142, %r1130, %r1132, %r1134;
mov.b32 %f430, %r1143;
add.ftz.f32 %f431, %f429, %f430;
mov.b32 %r1144, %f431;
mov.u32 %r1145, 1;
shfl.sync.bfly.b32 %r1146|%p100, %r1144, %r1145, %r1132, %r1134;
mov.b32 %f432, %r1146;
add.ftz.f32 %f61, %f431, %f432;
cvt.u64.u32 %rd151, %r1125;
add.s64 %rd153, %rd151, %rd118;
shl.b64 %rd155, %rd153, 2;
add.s64 %rd16, %rd120, %rd155;
@%p17 bra $L__BB126_32;

st.global.f32 [%rd16], %f61;

$L__BB126_32:
ld.shared.f32 %f433, [%r803+1664];
add.ftz.f32 %f434, %f433, %f484;
mov.b32 %r1158, %f434;
shfl.sync.bfly.b32 %r1162|%p104, %r1158, %r1133, %r1132, %r1134;
mov.b32 %f435, %r1162;
add.ftz.f32 %f436, %f434, %f435;
mov.b32 %r1163, %f436;
shfl.sync.bfly.b32 %r1165|%p105, %r1163, %r1137, %r1132, %r1134;
mov.b32 %f437, %r1165;
add.ftz.f32 %f438, %f436, %f437;
mov.b32 %r1166, %f438;
shfl.sync.bfly.b32 %r1168|%p106, %r1166, %r1140, %r1132, %r1134;
mov.b32 %f439, %r1168;
add.ftz.f32 %f440, %f438, %f439;
mov.b32 %r1169, %f440;
shfl.sync.bfly.b32 %r1170|%p107, %r1169, %r1130, %r1132, %r1134;
mov.b32 %f441, %r1170;
add.ftz.f32 %f442, %f440, %f441;
mov.b32 %r1171, %f442;
shfl.sync.bfly.b32 %r1173|%p108, %r1171, %r1145, %r1132, %r1134;
mov.b32 %f443, %r1173;
add.ftz.f32 %f62, %f442, %f443;
@%p25 bra $L__BB126_34;

st.global.f32 [%rd16+4], %f62;

$L__BB126_34:
mul.lo.s32 %r32, %r36, 7;
mov.u32 %r1178, 2;
ld.shared.f32 %f444, [%r803+1792];
add.ftz.f32 %f445, %f444, %f483;
mov.b32 %r1179, %f445;
mov.u32 %r1180, 31;
mov.u32 %r1181, 16;
mov.u32 %r1182, -1;
shfl.sync.bfly.b32 %r1183|%p110, %r1179, %r1181, %r1180, %r1182;
mov.b32 %f446, %r1183;
add.ftz.f32 %f447, %f445, %f446;
mov.b32 %r1184, %f447;
mov.u32 %r1185, 8;
shfl.sync.bfly.b32 %r1186|%p111, %r1184, %r1185, %r1180, %r1182;
mov.b32 %f448, %r1186;
add.ftz.f32 %f449, %f447, %f448;
mov.b32 %r1187, %f449;
mov.u32 %r1188, 4;
shfl.sync.bfly.b32 %r1189|%p112, %r1187, %r1188, %r1180, %r1182;
mov.b32 %f450, %r1189;
add.ftz.f32 %f451, %f449, %f450;
mov.b32 %r1190, %f451;
shfl.sync.bfly.b32 %r1191|%p113, %r1190, %r1178, %r1180, %r1182;
mov.b32 %f452, %r1191;
add.ftz.f32 %f453, %f451, %f452;
mov.b32 %r1192, %f453;
mov.u32 %r1193, 1;
shfl.sync.bfly.b32 %r1194|%p114, %r1192, %r1193, %r1180, %r1182;
mov.b32 %f454, %r1194;
add.ftz.f32 %f63, %f453, %f454;
@%p17 bra $L__BB126_36;

cvt.u64.u32 %rd156, %r32;
add.s64 %rd158, %rd156, %rd118;
shl.b64 %rd160, %rd158, 2;
add.s64 %rd161, %rd120, %rd160;
st.global.f32 [%rd161], %f63;

$L__BB126_36:
ld.shared.f32 %f455, [%r803+1920];
add.ftz.f32 %f456, %f455, %f482;
mov.b32 %r1208, %f456;
shfl.sync.bfly.b32 %r1212|%p118, %r1208, %r1181, %r1180, %r1182;
mov.b32 %f457, %r1212;
add.ftz.f32 %f458, %f456, %f457;
mov.b32 %r1213, %f458;
shfl.sync.bfly.b32 %r1215|%p119, %r1213, %r1185, %r1180, %r1182;
mov.b32 %f459, %r1215;
add.ftz.f32 %f460, %f458, %f459;
mov.b32 %r1216, %f460;
shfl.sync.bfly.b32 %r1218|%p120, %r1216, %r1188, %r1180, %r1182;
mov.b32 %f461, %r1218;
add.ftz.f32 %f462, %f460, %f461;
mov.b32 %r1219, %f462;
shfl.sync.bfly.b32 %r1220|%p121, %r1219, %r1178, %r1180, %r1182;
mov.b32 %f463, %r1220;
add.ftz.f32 %f464, %f462, %f463;
mov.b32 %r1221, %f464;
shfl.sync.bfly.b32 %r1223|%p122, %r1221, %r1193, %r1180, %r1182;
mov.b32 %f465, %r1223;
add.ftz.f32 %f64, %f464, %f465;
@%p25 bra $L__BB126_38;

add.s32 %r1224, %r32, 1;
cvt.u64.u32 %rd162, %r1224;
add.s64 %rd164, %rd162, %rd118;
shl.b64 %rd166, %rd164, 2;
add.s64 %rd167, %rd120, %rd166;
st.global.f32 [%rd167], %f64;

$L__BB126_38:
ret;

}
.entry _Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj(
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2,
.param .align 8 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3[48],
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5,
.param .align 4 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6[12],
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15
)
.maxntid 256, 1, 1
.minnctapersm 1
{
.local .align 8 .b8 __local_depot127[16];
.reg .b64 %SP;
.reg .b64 %SPL;
.reg .pred %p<48>;
.reg .b16 %rs<9>;
.reg .f32 %f<253>;
.reg .b32 %r<482>;
.reg .b64 %rd<114>;


mov.u64 %SPL, __local_depot127;
ld.param.u64 %rd16, [_Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0];
ld.param.u64 %rd17, [_Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1];
ld.param.u64 %rd18, [_Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2];
ld.param.u32 %r20, [_Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5];
ld.param.u32 %r25, [_Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8];
ld.param.u32 %r26, [_Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9];
ld.param.u32 %r28, [_Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11];
ld.param.u32 %r29, [_Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12];
ld.param.u32 %r32, [_Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14];
ld.param.u32 %r31, [_Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15];
ld.param.u32 %r23, [_Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+8];
ld.param.u32 %r22, [_Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+4];
ld.param.u32 %r21, [_Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6];
ld.param.u64 %rd20, [_Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+8];
add.u64 %rd1, %SPL, 0;
add.u64 %rd2, %SPL, 8;
mov.u32 %r1, %tid.y;
setp.ge.u32 %p1, %r1, %r32;
@%p1 bra $L__BB127_29;

mov.u32 %r2, %ctaid.y;
cvta.to.global.u64 %rd27, %rd18;
mad.lo.s32 %r33, %r1, %r31, %r2;
mul.wide.u32 %rd28, %r33, 4;
add.s64 %rd29, %rd27, %rd28;
ld.global.nc.u32 %r3, [%rd29];
mov.f32 %f246, 0f00000000;
st.local.v2.f32 [%rd1], {%f246, %f246};
st.local.v2.f32 [%rd2], {%f246, %f246};
mov.u32 %r4, %tid.x;
shr.u32 %r5, %r4, 4;
shr.u32 %r34, %r20, 8;
setp.ge.u32 %p2, %r5, %r34;
mov.f32 %f247, %f246;
mov.f32 %f248, %f246;
mov.f32 %f249, %f246;
@%p2 bra $L__BB127_9;

and.b32 %r35, %r4, 8;
shr.u32 %r36, %r35, 1;
or.b32 %r37, %r36, 3;
cvt.u64.u32 %rd30, %r37;
shr.u32 %r38, %r4, 1;
cvt.u64.u32 %rd31, %r38;
and.b64 %rd32, %rd31, 2147483640;
or.b64 %rd33, %rd32, %rd30;
mul.lo.s32 %r39, %r1, %r26;
cvt.u64.u32 %rd34, %r39;
add.s64 %rd35, %rd33, %rd34;
mul.hi.u32 %r40, %r2, %r21;
add.s32 %r41, %r40, %r2;
shr.u32 %r42, %r41, %r22;
mul.lo.s32 %r43, %r42, %r23;
sub.s32 %r44, %r2, %r43;
mul.lo.s32 %r45, %r44, %r29;
cvt.u64.u32 %rd36, %r45;
add.s64 %rd37, %rd35, %rd36;
mul.lo.s64 %rd38, %rd37, 36;
cvta.to.global.u64 %rd39, %rd17;
add.s64 %rd40, %rd39, %rd38;
add.s64 %rd111, %rd40, -104;
cvta.to.global.u64 %rd113, %rd16;
cvta.to.global.u64 %rd112, %rd20;
mov.f32 %f246, 0f00000000;
mov.u32 %r481, %r5;

$L__BB127_3:
and.b32 %r85, %r4, 15;
shl.b32 %r86, %r4, 2;
cvt.u64.u32 %rd41, %r86;
and.b64 %rd42, %rd41, 60;
mov.u32 %r88, %ctaid.x;
shl.b32 %r89, %r88, 1;
mad.lo.s32 %r90, %r89, %r25, %r5;
mad.lo.s32 %r91, %r3, %r28, %r90;
mul.wide.s32 %rd43, %r91, 84;
add.s64 %rd44, %rd42, %rd43;
add.s64 %rd45, %rd113, %rd44;
ld.global.nc.u32 %r92, [%rd45+16];
and.b32 %r93, %r4, 7;
and.b64 %rd46, %rd41, 28;
add.s64 %rd47, %rd111, %rd46;
ld.global.nc.u32 %r7, [%rd47];
ld.global.nc.u32 %r8, [%rd111+-4];

{.reg .f16 low,high;
mov.b32 {low,high},%r8;
cvt.f32.f16 %f42, low;}


ld.global.nc.u32 %r9, [%rd47+36];
ld.global.nc.u32 %r10, [%rd111+32];

{.reg .f16 low,high;
mov.b32 {low,high},%r10;
cvt.f32.f16 %f43, low;}


ld.global.nc.u32 %r11, [%rd47+72];
ld.global.nc.u32 %r12, [%rd111+68];

{.reg .f16 low,high;
mov.b32 {low,high},%r12;
cvt.f32.f16 %f44, low;}


ld.global.nc.u32 %r13, [%rd47+108];
ld.global.nc.u32 %r14, [%rd111+104];

{.reg .f16 low,high;
mov.b32 {low,high},%r14;
cvt.f32.f16 %f45, low;}


sub.s32 %r94, %r85, %r93;
and.b32 %r95, %r4, 4;
shr.u32 %r96, %r95, 2;
add.s32 %r97, %r94, %r96;
cvt.s64.s32 %rd48, %r97;
add.s64 %rd49, %rd43, %rd48;
add.s64 %rd50, %rd113, %rd49;
ld.global.nc.u8 %rs1, [%rd50];
cvt.u32.u16 %r98, %rs1;
and.b32 %r99, %r98, 240;
and.b32 %r51, %r92, 50529027;
mov.u32 %r81, 0;

dp4a.s32.s32 %r50, %r51, %r7, %r81;

and.b32 %r100, %r98, 15;
mul.lo.s32 %r101, %r50, %r100;
cvt.rn.f32.s32 %f48, %r101;
fma.rn.ftz.f32 %f49, %f42, %f48, 0f00000000;
shr.u32 %r102, %r99, 4;
shl.b32 %r103, %r102, 8;
or.b32 %r104, %r103, %r102;
shl.b32 %r105, %r104, 16;
or.b32 %r55, %r105, %r104;

dp4a.s32.s32 %r54, %r55, %r7, %r81;

cvt.rn.f32.s32 %f50, %r54;
fma.rn.ftz.f32 %f51, %f42, %f50, 0f00000000;
ld.global.nc.u8 %rs2, [%rd50+2];
cvt.u32.u16 %r106, %rs2;
and.b32 %r107, %r106, 240;
shr.u32 %r108, %r92, 2;
and.b32 %r59, %r108, 50529027;

dp4a.s32.s32 %r58, %r59, %r9, %r81;

and.b32 %r109, %r106, 15;
mul.lo.s32 %r110, %r58, %r109;
cvt.rn.f32.s32 %f52, %r110;
fma.rn.ftz.f32 %f53, %f43, %f52, %f49;
shr.u32 %r111, %r107, 4;
shl.b32 %r112, %r111, 8;
or.b32 %r113, %r112, %r111;
shl.b32 %r114, %r113, 16;
or.b32 %r63, %r114, %r113;

dp4a.s32.s32 %r62, %r63, %r9, %r81;

cvt.rn.f32.s32 %f54, %r62;
fma.rn.ftz.f32 %f55, %f43, %f54, %f51;
ld.global.nc.u8 %rs3, [%rd50+4];
cvt.u32.u16 %r115, %rs3;
and.b32 %r116, %r115, 240;
shr.u32 %r117, %r92, 4;
and.b32 %r67, %r117, 50529027;

dp4a.s32.s32 %r66, %r67, %r11, %r81;

and.b32 %r118, %r115, 15;
mul.lo.s32 %r119, %r66, %r118;
cvt.rn.f32.s32 %f56, %r119;
fma.rn.ftz.f32 %f57, %f44, %f56, %f53;
shr.u32 %r120, %r116, 4;
shl.b32 %r121, %r120, 8;
or.b32 %r122, %r121, %r120;
shl.b32 %r123, %r122, 16;
or.b32 %r71, %r123, %r122;

dp4a.s32.s32 %r70, %r71, %r11, %r81;

cvt.rn.f32.s32 %f58, %r70;
fma.rn.ftz.f32 %f59, %f44, %f58, %f55;
ld.global.nc.u8 %rs4, [%rd50+6];
cvt.u32.u16 %r124, %rs4;
and.b32 %r125, %r124, 240;
shr.u32 %r126, %r92, 6;
and.b32 %r75, %r126, 50529027;

dp4a.s32.s32 %r74, %r75, %r13, %r81;

and.b32 %r127, %r124, 15;
mul.lo.s32 %r128, %r74, %r127;
cvt.rn.f32.s32 %f60, %r128;
fma.rn.ftz.f32 %f61, %f45, %f60, %f57;
shr.u32 %r129, %r125, 4;
shl.b32 %r130, %r129, 8;
or.b32 %r131, %r130, %r129;
shl.b32 %r132, %r131, 16;
or.b32 %r79, %r132, %r131;

dp4a.s32.s32 %r78, %r79, %r13, %r81;

cvt.rn.f32.s32 %f62, %r78;
fma.rn.ftz.f32 %f63, %f45, %f62, %f59;
add.s64 %rd51, %rd113, %rd43;
ld.global.nc.u32 %r82, [%rd51+80];

{.reg .f16 low,high;
mov.b32 {low,high},%r82;
cvt.f32.f16 %f46, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r82;
cvt.f32.f16 %f47, high;}


mul.ftz.f32 %f64, %f61, %f46;
mul.ftz.f32 %f65, %f63, %f47;
sub.ftz.f32 %f66, %f64, %f65;
add.ftz.f32 %f249, %f66, %f249;
setp.eq.s64 %p3, %rd20, 0;
@%p3 bra $L__BB127_5;

shr.u32 %r174, %r4, 4;
mad.lo.s32 %r177, %r89, %r25, %r174;
mad.lo.s32 %r178, %r3, %r28, %r177;
mul.wide.s32 %rd54, %r178, 84;
add.s64 %rd55, %rd42, %rd54;
add.s64 %rd56, %rd112, %rd55;
ld.global.u32 %r179, [%rd56+16];

{.reg .f16 low,high;
mov.b32 {low,high},%r8;
cvt.f32.f16 %f67, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r10;
cvt.f32.f16 %f68, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r12;
cvt.f32.f16 %f69, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r14;
cvt.f32.f16 %f70, low;}


add.s64 %rd58, %rd54, %rd48;
add.s64 %rd59, %rd112, %rd58;
ld.global.u8 %r185, [%rd59];
and.b32 %r138, %r179, 50529027;

dp4a.s32.s32 %r137, %r138, %r7, %r81;

and.b32 %r186, %r185, 15;
mul.lo.s32 %r187, %r137, %r186;
cvt.rn.f32.s32 %f73, %r187;
fma.rn.ftz.f32 %f74, %f67, %f73, 0f00000000;
shr.u32 %r188, %r185, 4;
shl.b32 %r189, %r188, 8;
or.b32 %r190, %r189, %r188;
shl.b32 %r191, %r190, 16;
or.b32 %r142, %r191, %r190;

dp4a.s32.s32 %r141, %r142, %r7, %r81;

cvt.rn.f32.s32 %f75, %r141;
fma.rn.ftz.f32 %f76, %f67, %f75, 0f00000000;
ld.global.u8 %r192, [%rd59+2];
shr.u32 %r193, %r179, 2;
and.b32 %r146, %r193, 50529027;

dp4a.s32.s32 %r145, %r146, %r9, %r81;

and.b32 %r194, %r192, 15;
mul.lo.s32 %r195, %r145, %r194;
cvt.rn.f32.s32 %f77, %r195;
fma.rn.ftz.f32 %f78, %f68, %f77, %f74;
shr.u32 %r196, %r192, 4;
shl.b32 %r197, %r196, 8;
or.b32 %r198, %r197, %r196;
shl.b32 %r199, %r198, 16;
or.b32 %r150, %r199, %r198;

dp4a.s32.s32 %r149, %r150, %r9, %r81;

cvt.rn.f32.s32 %f79, %r149;
fma.rn.ftz.f32 %f80, %f68, %f79, %f76;
ld.global.u8 %r200, [%rd59+4];
shr.u32 %r201, %r179, 4;
and.b32 %r154, %r201, 50529027;

dp4a.s32.s32 %r153, %r154, %r11, %r81;

and.b32 %r202, %r200, 15;
mul.lo.s32 %r203, %r153, %r202;
cvt.rn.f32.s32 %f81, %r203;
fma.rn.ftz.f32 %f82, %f69, %f81, %f78;
shr.u32 %r204, %r200, 4;
shl.b32 %r205, %r204, 8;
or.b32 %r206, %r205, %r204;
shl.b32 %r207, %r206, 16;
or.b32 %r158, %r207, %r206;

dp4a.s32.s32 %r157, %r158, %r11, %r81;

cvt.rn.f32.s32 %f83, %r157;
fma.rn.ftz.f32 %f84, %f69, %f83, %f80;
ld.global.u8 %r208, [%rd59+6];
shr.u32 %r209, %r179, 6;
and.b32 %r162, %r209, 50529027;

dp4a.s32.s32 %r161, %r162, %r13, %r81;

and.b32 %r210, %r208, 15;
mul.lo.s32 %r211, %r161, %r210;
cvt.rn.f32.s32 %f85, %r211;
fma.rn.ftz.f32 %f86, %f70, %f85, %f82;
shr.u32 %r212, %r208, 4;
shl.b32 %r213, %r212, 8;
or.b32 %r214, %r213, %r212;
shl.b32 %r215, %r214, 16;
or.b32 %r166, %r215, %r214;

dp4a.s32.s32 %r165, %r166, %r13, %r81;

cvt.rn.f32.s32 %f87, %r165;
fma.rn.ftz.f32 %f88, %f70, %f87, %f84;
add.s64 %rd60, %rd112, %rd54;
ld.global.u32 %r170, [%rd60+80];

{.reg .f16 low,high;
mov.b32 {low,high},%r170;
cvt.f32.f16 %f71, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r170;
cvt.f32.f16 %f72, high;}


mul.ftz.f32 %f89, %f86, %f71;
mul.ftz.f32 %f90, %f88, %f72;
sub.ftz.f32 %f91, %f89, %f90;
add.ftz.f32 %f247, %f91, %f247;

$L__BB127_5:
shr.u32 %r257, %r4, 4;
mad.lo.s32 %r258, %r3, %r28, %r257;
or.b32 %r261, %r89, 1;
mad.lo.s32 %r262, %r25, %r261, %r258;
mul.wide.s32 %rd63, %r262, 84;
add.s64 %rd64, %rd42, %rd63;
add.s64 %rd65, %rd113, %rd64;
ld.global.nc.u32 %r263, [%rd65+16];

{.reg .f16 low,high;
mov.b32 {low,high},%r8;
cvt.f32.f16 %f92, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r10;
cvt.f32.f16 %f93, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r12;
cvt.f32.f16 %f94, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r14;
cvt.f32.f16 %f95, low;}


add.s64 %rd67, %rd63, %rd48;
add.s64 %rd68, %rd113, %rd67;
ld.global.nc.u8 %rs5, [%rd68];
cvt.u32.u16 %r269, %rs5;
and.b32 %r270, %r269, 240;
and.b32 %r221, %r263, 50529027;

dp4a.s32.s32 %r220, %r221, %r7, %r81;

and.b32 %r271, %r269, 15;
mul.lo.s32 %r272, %r220, %r271;
cvt.rn.f32.s32 %f98, %r272;
fma.rn.ftz.f32 %f99, %f92, %f98, 0f00000000;
shr.u32 %r273, %r270, 4;
shl.b32 %r274, %r273, 8;
or.b32 %r275, %r274, %r273;
shl.b32 %r276, %r275, 16;
or.b32 %r225, %r276, %r275;

dp4a.s32.s32 %r224, %r225, %r7, %r81;

cvt.rn.f32.s32 %f100, %r224;
fma.rn.ftz.f32 %f101, %f92, %f100, 0f00000000;
ld.global.nc.u8 %rs6, [%rd68+2];
cvt.u32.u16 %r277, %rs6;
and.b32 %r278, %r277, 240;
shr.u32 %r279, %r263, 2;
and.b32 %r229, %r279, 50529027;

dp4a.s32.s32 %r228, %r229, %r9, %r81;

and.b32 %r280, %r277, 15;
mul.lo.s32 %r281, %r228, %r280;
cvt.rn.f32.s32 %f102, %r281;
fma.rn.ftz.f32 %f103, %f93, %f102, %f99;
shr.u32 %r282, %r278, 4;
shl.b32 %r283, %r282, 8;
or.b32 %r284, %r283, %r282;
shl.b32 %r285, %r284, 16;
or.b32 %r233, %r285, %r284;

dp4a.s32.s32 %r232, %r233, %r9, %r81;

cvt.rn.f32.s32 %f104, %r232;
fma.rn.ftz.f32 %f105, %f93, %f104, %f101;
ld.global.nc.u8 %rs7, [%rd68+4];
cvt.u32.u16 %r286, %rs7;
and.b32 %r287, %r286, 240;
shr.u32 %r288, %r263, 4;
and.b32 %r237, %r288, 50529027;

dp4a.s32.s32 %r236, %r237, %r11, %r81;

and.b32 %r289, %r286, 15;
mul.lo.s32 %r290, %r236, %r289;
cvt.rn.f32.s32 %f106, %r290;
fma.rn.ftz.f32 %f107, %f94, %f106, %f103;
shr.u32 %r291, %r287, 4;
shl.b32 %r292, %r291, 8;
or.b32 %r293, %r292, %r291;
shl.b32 %r294, %r293, 16;
or.b32 %r241, %r294, %r293;

dp4a.s32.s32 %r240, %r241, %r11, %r81;

cvt.rn.f32.s32 %f108, %r240;
fma.rn.ftz.f32 %f109, %f94, %f108, %f105;
ld.global.nc.u8 %rs8, [%rd68+6];
cvt.u32.u16 %r295, %rs8;
and.b32 %r296, %r295, 240;
shr.u32 %r297, %r263, 6;
and.b32 %r245, %r297, 50529027;

dp4a.s32.s32 %r244, %r245, %r13, %r81;

and.b32 %r298, %r295, 15;
mul.lo.s32 %r299, %r244, %r298;
cvt.rn.f32.s32 %f110, %r299;
fma.rn.ftz.f32 %f111, %f95, %f110, %f107;
shr.u32 %r300, %r296, 4;
shl.b32 %r301, %r300, 8;
or.b32 %r302, %r301, %r300;
shl.b32 %r303, %r302, 16;
or.b32 %r249, %r303, %r302;

dp4a.s32.s32 %r248, %r249, %r13, %r81;

cvt.rn.f32.s32 %f112, %r248;
fma.rn.ftz.f32 %f113, %f95, %f112, %f109;
add.s64 %rd69, %rd113, %rd63;
ld.global.nc.u32 %r252, [%rd69+80];

{.reg .f16 low,high;
mov.b32 {low,high},%r252;
cvt.f32.f16 %f96, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r252;
cvt.f32.f16 %f97, high;}


mul.ftz.f32 %f114, %f111, %f96;
mul.ftz.f32 %f115, %f113, %f97;
sub.ftz.f32 %f116, %f114, %f115;
add.ftz.f32 %f248, %f116, %f248;
@%p3 bra $L__BB127_7;

add.s64 %rd74, %rd112, %rd64;
ld.global.u32 %r351, [%rd74+16];

{.reg .f16 low,high;
mov.b32 {low,high},%r8;
cvt.f32.f16 %f117, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r10;
cvt.f32.f16 %f118, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r12;
cvt.f32.f16 %f119, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r14;
cvt.f32.f16 %f120, low;}


add.s64 %rd77, %rd112, %rd67;
ld.global.u8 %r357, [%rd77];
and.b32 %r309, %r351, 50529027;
mov.u32 %r339, 0;

dp4a.s32.s32 %r308, %r309, %r7, %r339;

and.b32 %r358, %r357, 15;
mul.lo.s32 %r359, %r308, %r358;
cvt.rn.f32.s32 %f123, %r359;
fma.rn.ftz.f32 %f124, %f117, %f123, 0f00000000;
shr.u32 %r360, %r357, 4;
shl.b32 %r361, %r360, 8;
or.b32 %r362, %r361, %r360;
shl.b32 %r363, %r362, 16;
or.b32 %r313, %r363, %r362;

dp4a.s32.s32 %r312, %r313, %r7, %r339;

cvt.rn.f32.s32 %f125, %r312;
fma.rn.ftz.f32 %f126, %f117, %f125, 0f00000000;
ld.global.u8 %r364, [%rd77+2];
shr.u32 %r365, %r351, 2;
and.b32 %r317, %r365, 50529027;

dp4a.s32.s32 %r316, %r317, %r9, %r339;

and.b32 %r366, %r364, 15;
mul.lo.s32 %r367, %r316, %r366;
cvt.rn.f32.s32 %f127, %r367;
fma.rn.ftz.f32 %f128, %f118, %f127, %f124;
shr.u32 %r368, %r364, 4;
shl.b32 %r369, %r368, 8;
or.b32 %r370, %r369, %r368;
shl.b32 %r371, %r370, 16;
or.b32 %r321, %r371, %r370;

dp4a.s32.s32 %r320, %r321, %r9, %r339;

cvt.rn.f32.s32 %f129, %r320;
fma.rn.ftz.f32 %f130, %f118, %f129, %f126;
ld.global.u8 %r372, [%rd77+4];
shr.u32 %r373, %r351, 4;
and.b32 %r325, %r373, 50529027;

dp4a.s32.s32 %r324, %r325, %r11, %r339;

and.b32 %r374, %r372, 15;
mul.lo.s32 %r375, %r324, %r374;
cvt.rn.f32.s32 %f131, %r375;
fma.rn.ftz.f32 %f132, %f119, %f131, %f128;
shr.u32 %r376, %r372, 4;
shl.b32 %r377, %r376, 8;
or.b32 %r378, %r377, %r376;
shl.b32 %r379, %r378, 16;
or.b32 %r329, %r379, %r378;

dp4a.s32.s32 %r328, %r329, %r11, %r339;

cvt.rn.f32.s32 %f133, %r328;
fma.rn.ftz.f32 %f134, %f119, %f133, %f130;
ld.global.u8 %r380, [%rd77+6];
shr.u32 %r381, %r351, 6;
and.b32 %r333, %r381, 50529027;

dp4a.s32.s32 %r332, %r333, %r13, %r339;

and.b32 %r382, %r380, 15;
mul.lo.s32 %r383, %r332, %r382;
cvt.rn.f32.s32 %f135, %r383;
fma.rn.ftz.f32 %f136, %f120, %f135, %f132;
shr.u32 %r384, %r380, 4;
shl.b32 %r385, %r384, 8;
or.b32 %r386, %r385, %r384;
shl.b32 %r387, %r386, 16;
or.b32 %r337, %r387, %r386;

dp4a.s32.s32 %r336, %r337, %r13, %r339;

cvt.rn.f32.s32 %f137, %r336;
fma.rn.ftz.f32 %f138, %f120, %f137, %f134;
add.s64 %rd78, %rd112, %rd63;
ld.global.u32 %r341, [%rd78+80];

{.reg .f16 low,high;
mov.b32 {low,high},%r341;
cvt.f32.f16 %f121, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r341;
cvt.f32.f16 %f122, high;}


mul.ftz.f32 %f139, %f136, %f121;
mul.ftz.f32 %f140, %f138, %f122;
sub.ftz.f32 %f141, %f139, %f140;
add.ftz.f32 %f246, %f141, %f246;

$L__BB127_7:
add.s64 %rd113, %rd113, 168;
add.s64 %rd112, %rd112, 168;
add.s64 %rd111, %rd111, 576;
add.s32 %r481, %r481, 2;
setp.lt.u32 %p5, %r481, %r34;
@%p5 bra $L__BB127_3;

add.u64 %rd103, %SPL, 0;
add.u64 %rd97, %SPL, 8;
st.local.v2.f32 [%rd103], {%f249, %f248};
st.local.v2.f32 [%rd97], {%f247, %f246};

$L__BB127_9:
mov.b32 %r16, %f249;
setp.ne.s64 %p6, %rd20, 0;
@%p6 bra $L__BB127_11;
bra.uni $L__BB127_10;

$L__BB127_11:
add.u64 %rd109, %SPL, 0;
add.u64 %rd101, %SPL, 8;
mov.u32 %r415, 31;
mov.u32 %r416, 16;
mov.u32 %r417, -1;
shfl.sync.bfly.b32 %r418|%p17, %r16, %r416, %r415, %r417;
mov.b32 %f162, %r418;
add.ftz.f32 %f163, %f249, %f162;
mov.b32 %r419, %f163;
mov.u32 %r420, 8;
shfl.sync.bfly.b32 %r421|%p18, %r419, %r420, %r415, %r417;
mov.b32 %f164, %r421;
add.ftz.f32 %f165, %f163, %f164;
mov.b32 %r422, %f165;
mov.u32 %r423, 4;
shfl.sync.bfly.b32 %r424|%p19, %r422, %r423, %r415, %r417;
mov.b32 %f166, %r424;
add.ftz.f32 %f167, %f165, %f166;
mov.b32 %r425, %f167;
mov.u32 %r426, 2;
shfl.sync.bfly.b32 %r427|%p20, %r425, %r426, %r415, %r417;
mov.b32 %f168, %r427;
add.ftz.f32 %f169, %f167, %f168;
mov.b32 %r428, %f169;
mov.u32 %r429, 1;
shfl.sync.bfly.b32 %r430|%p21, %r428, %r429, %r415, %r417;
mov.b32 %f170, %r430;
add.ftz.f32 %f171, %f169, %f170;
st.local.f32 [%rd109], %f171;
mov.b32 %r431, %f247;
shfl.sync.bfly.b32 %r432|%p22, %r431, %r416, %r415, %r417;
mov.b32 %f172, %r432;
add.ftz.f32 %f173, %f247, %f172;
mov.b32 %r433, %f173;
shfl.sync.bfly.b32 %r434|%p23, %r433, %r420, %r415, %r417;
mov.b32 %f174, %r434;
add.ftz.f32 %f175, %f173, %f174;
mov.b32 %r435, %f175;
shfl.sync.bfly.b32 %r436|%p24, %r435, %r423, %r415, %r417;
mov.b32 %f176, %r436;
add.ftz.f32 %f177, %f175, %f176;
mov.b32 %r437, %f177;
shfl.sync.bfly.b32 %r438|%p25, %r437, %r426, %r415, %r417;
mov.b32 %f178, %r438;
add.ftz.f32 %f179, %f177, %f178;
mov.b32 %r439, %f179;
shfl.sync.bfly.b32 %r440|%p26, %r439, %r429, %r415, %r417;
mov.b32 %f180, %r440;
add.ftz.f32 %f181, %f179, %f180;
st.local.f32 [%rd101], %f181;
mov.b32 %r441, %f248;
shfl.sync.bfly.b32 %r442|%p27, %r441, %r416, %r415, %r417;
mov.b32 %f182, %r442;
add.ftz.f32 %f183, %f248, %f182;
mov.b32 %r443, %f183;
shfl.sync.bfly.b32 %r444|%p28, %r443, %r420, %r415, %r417;
mov.b32 %f184, %r444;
add.ftz.f32 %f185, %f183, %f184;
mov.b32 %r445, %f185;
shfl.sync.bfly.b32 %r446|%p29, %r445, %r423, %r415, %r417;
mov.b32 %f186, %r446;
add.ftz.f32 %f187, %f185, %f186;
mov.b32 %r447, %f187;
shfl.sync.bfly.b32 %r448|%p30, %r447, %r426, %r415, %r417;
mov.b32 %f188, %r448;
add.ftz.f32 %f189, %f187, %f188;
mov.b32 %r449, %f189;
shfl.sync.bfly.b32 %r450|%p31, %r449, %r429, %r415, %r417;
mov.b32 %f190, %r450;
add.ftz.f32 %f191, %f189, %f190;
st.local.f32 [%rd109+4], %f191;
mov.b32 %r451, %f246;
shfl.sync.bfly.b32 %r452|%p32, %r451, %r416, %r415, %r417;
mov.b32 %f192, %r452;
add.ftz.f32 %f193, %f246, %f192;
mov.b32 %r453, %f193;
shfl.sync.bfly.b32 %r454|%p33, %r453, %r420, %r415, %r417;
mov.b32 %f194, %r454;
add.ftz.f32 %f195, %f193, %f194;
mov.b32 %r455, %f195;
shfl.sync.bfly.b32 %r456|%p34, %r455, %r423, %r415, %r417;
mov.b32 %f196, %r456;
add.ftz.f32 %f197, %f195, %f196;
mov.b32 %r457, %f197;
shfl.sync.bfly.b32 %r458|%p35, %r457, %r426, %r415, %r417;
mov.b32 %f198, %r458;
add.ftz.f32 %f199, %f197, %f198;
mov.b32 %r459, %f199;
shfl.sync.bfly.b32 %r460|%p36, %r459, %r429, %r415, %r417;
mov.b32 %f200, %r460;
add.ftz.f32 %f201, %f199, %f200;
st.local.f32 [%rd101+4], %f201;
bra.uni $L__BB127_12;

$L__BB127_10:
add.u64 %rd105, %SPL, 0;
mov.u32 %r389, 31;
mov.u32 %r390, 16;
mov.u32 %r391, -1;
shfl.sync.bfly.b32 %r392|%p7, %r16, %r390, %r389, %r391;
mov.b32 %f142, %r392;
add.ftz.f32 %f143, %f249, %f142;
mov.b32 %r393, %f143;
mov.u32 %r394, 8;
shfl.sync.bfly.b32 %r395|%p8, %r393, %r394, %r389, %r391;
mov.b32 %f144, %r395;
add.ftz.f32 %f145, %f143, %f144;
mov.b32 %r396, %f145;
mov.u32 %r397, 4;
shfl.sync.bfly.b32 %r398|%p9, %r396, %r397, %r389, %r391;
mov.b32 %f146, %r398;
add.ftz.f32 %f147, %f145, %f146;
mov.b32 %r399, %f147;
mov.u32 %r400, 2;
shfl.sync.bfly.b32 %r401|%p10, %r399, %r400, %r389, %r391;
mov.b32 %f148, %r401;
add.ftz.f32 %f149, %f147, %f148;
mov.b32 %r402, %f149;
mov.u32 %r403, 1;
shfl.sync.bfly.b32 %r404|%p11, %r402, %r403, %r389, %r391;
mov.b32 %f150, %r404;
add.ftz.f32 %f151, %f149, %f150;
st.local.f32 [%rd105], %f151;
mov.b32 %r405, %f248;
shfl.sync.bfly.b32 %r406|%p12, %r405, %r390, %r389, %r391;
mov.b32 %f152, %r406;
add.ftz.f32 %f153, %f248, %f152;
mov.b32 %r407, %f153;
shfl.sync.bfly.b32 %r408|%p13, %r407, %r394, %r389, %r391;
mov.b32 %f154, %r408;
add.ftz.f32 %f155, %f153, %f154;
mov.b32 %r409, %f155;
shfl.sync.bfly.b32 %r410|%p14, %r409, %r397, %r389, %r391;
mov.b32 %f156, %r410;
add.ftz.f32 %f157, %f155, %f156;
mov.b32 %r411, %f157;
shfl.sync.bfly.b32 %r412|%p15, %r411, %r400, %r389, %r391;
mov.b32 %f158, %r412;
add.ftz.f32 %f159, %f157, %f158;
mov.b32 %r413, %f159;
shfl.sync.bfly.b32 %r414|%p16, %r413, %r403, %r389, %r391;
mov.b32 %f160, %r414;
add.ftz.f32 %f161, %f159, %f160;
st.local.f32 [%rd105+4], %f161;

$L__BB127_12:
ld.param.u32 %r470, [_Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7];
setp.gt.u32 %p37, %r4, 1;
mov.u32 %r462, %ctaid.x;
shl.b32 %r463, %r462, 1;
add.s32 %r17, %r463, %r4;
setp.ge.u32 %p38, %r17, %r470;
or.pred %p39, %p37, %p38;
@%p39 bra $L__BB127_29;

add.u64 %rd107, %SPL, 0;
ld.param.u64 %rd92, [_Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3];
ld.param.u32 %r471, [_Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13];
cvt.u64.u32 %rd14, %r4;
mul.wide.u32 %rd79, %r4, 4;
add.s64 %rd80, %rd107, %rd79;
ld.local.f32 %f252, [%rd80];
mad.lo.s32 %r18, %r3, %r471, %r17;
setp.eq.s64 %p40, %rd92, 0;
@%p40 bra $L__BB127_15;

ld.param.u64 %rd93, [_Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3];
cvta.to.global.u64 %rd81, %rd93;
mul.wide.u32 %rd82, %r18, 4;
add.s64 %rd83, %rd81, %rd82;
ld.global.f32 %f202, [%rd83];
add.ftz.f32 %f252, %f252, %f202;

$L__BB127_15:
setp.eq.s64 %p41, %rd20, 0;
@%p41 bra $L__BB127_28;

add.u64 %rd99, %SPL, 8;
ld.param.u64 %rd95, [_Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+16];
shl.b64 %rd84, %rd14, 2;
add.s64 %rd85, %rd99, %rd84;
ld.local.f32 %f24, [%rd85];
setp.eq.s64 %p42, %rd95, 0;
@%p42 bra $L__BB127_18;

ld.param.u64 %rd96, [_Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+16];
cvta.to.global.u64 %rd86, %rd96;
mul.wide.u32 %rd87, %r18, 4;
add.s64 %rd88, %rd86, %rd87;
ld.global.f32 %f203, [%rd88];
add.ftz.f32 %f24, %f24, %f203;

$L__BB127_18:
ld.param.u32 %r476, [_Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.gt.s32 %p43, %r476, 2;
@%p43 bra $L__BB127_22;

ld.param.u32 %r479, [_Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.eq.s32 %p46, %r479, 1;
@%p46 bra $L__BB127_26;

ld.param.u32 %r480, [_Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.eq.s32 %p47, %r480, 2;
@%p47 bra $L__BB127_21;
bra.uni $L__BB127_27;

$L__BB127_21:
mul.ftz.f32 %f223, %f24, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f224, %f223;
add.ftz.f32 %f225, %f224, 0f3F800000;
div.approx.ftz.f32 %f226, %f24, %f225;
mul.ftz.f32 %f252, %f252, %f226;
bra.uni $L__BB127_28;

$L__BB127_22:
ld.param.u32 %r477, [_Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.eq.s32 %p44, %r477, 3;
@%p44 bra $L__BB127_25;

ld.param.u32 %r478, [_Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.ne.s32 %p45, %r478, 6;
@%p45 bra $L__BB127_27;

ld.param.f32 %f235, [_Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+44];
min.ftz.f32 %f204, %f24, %f235;
neg.ftz.f32 %f205, %f235;
min.ftz.f32 %f206, %f252, %f235;
max.ftz.f32 %f207, %f206, %f205;
mul.ftz.f32 %f208, %f204, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f209, %f208;
add.ftz.f32 %f210, %f209, 0f3F800000;
div.approx.ftz.f32 %f211, %f204, %f210;
mul.ftz.f32 %f252, %f207, %f211;
bra.uni $L__BB127_28;

$L__BB127_27:
mul.ftz.f32 %f252, %f252, %f24;
bra.uni $L__BB127_28;

$L__BB127_26:
mul.ftz.f32 %f227, %f24, 0f3F000000;
mul.ftz.f32 %f228, %f24, 0f3D372713;
fma.rn.ftz.f32 %f229, %f24, %f228, 0f3F800000;
mul.ftz.f32 %f230, %f24, 0f3F4C422A;
mul.ftz.f32 %f231, %f230, %f229;
tanh.approx.f32 %f232, %f231;
add.ftz.f32 %f233, %f232, 0f3F800000;
mul.ftz.f32 %f234, %f227, %f233;
mul.ftz.f32 %f252, %f252, %f234;
bra.uni $L__BB127_28;

$L__BB127_25:
mov.f32 %f212, 0f40E00000;
min.ftz.f32 %f213, %f24, %f212;
min.ftz.f32 %f214, %f252, %f212;
mov.f32 %f215, 0fC0E00000;
max.ftz.f32 %f216, %f214, %f215;
mul.ftz.f32 %f217, %f213, 0f3FD9DB23;
mul.ftz.f32 %f218, %f217, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f219, %f218;
add.ftz.f32 %f220, %f219, 0f3F800000;
div.approx.ftz.f32 %f221, %f213, %f220;
add.ftz.f32 %f222, %f216, 0f3F800000;
mul.ftz.f32 %f252, %f222, %f221;

$L__BB127_28:
ld.param.u64 %rd94, [_Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4];
ld.param.u32 %r475, [_Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10];
ld.param.u32 %r474, [_Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13];
mov.u32 %r473, %ctaid.y;
mov.u32 %r472, %tid.y;
mad.lo.s32 %r468, %r472, %r475, %r17;
mad.lo.s32 %r469, %r473, %r474, %r468;
cvta.to.global.u64 %rd89, %rd94;
mul.wide.u32 %rd90, %r469, 4;
add.s64 %rd91, %rd89, %rd90;
st.global.f32 [%rd91], %f252;

$L__BB127_29:
ret;

}
.entry _Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj(
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2,
.param .align 8 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3[48],
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5,
.param .align 4 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6[12],
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15
)
.maxntid 256, 1, 1
.minnctapersm 1
{
.local .align 8 .b8 __local_depot128[8];
.reg .b64 %SP;
.reg .b64 %SPL;
.reg .pred %p<17>;
.reg .b16 %rs<9>;
.reg .f32 %f<86>;
.reg .b32 %r<231>;
.reg .b64 %rd<56>;


mov.u64 %SPL, __local_depot128;
ld.param.u64 %rd16, [_Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0];
ld.param.u64 %rd17, [_Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1];
ld.param.u64 %rd18, [_Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2];
ld.param.u64 %rd19, [_Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4];
ld.param.u32 %r19, [_Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5];
ld.param.u32 %r11, [_Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7];
ld.param.u32 %r12, [_Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8];
ld.param.u32 %r13, [_Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9];
ld.param.u32 %r14, [_Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10];
ld.param.u32 %r15, [_Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11];
ld.param.u32 %r16, [_Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12];
ld.param.u32 %r17, [_Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13];
ld.param.u32 %r20, [_Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14];
ld.param.u32 %r18, [_Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15];
ld.param.u32 %r10, [_Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+8];
ld.param.u32 %r9, [_Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+4];
ld.param.u32 %r8, [_Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6];
add.u64 %rd1, %SPL, 0;
shr.u32 %r1, %r19, 8;
mov.u32 %r2, %tid.y;
setp.ge.u32 %p1, %r2, %r20;
@%p1 bra $L__BB128_7;

mov.f32 %f84, 0f00000000;
st.local.v2.f32 [%rd1], {%f84, %f84};
mov.u32 %r3, %tid.x;
shr.u32 %r230, %r3, 4;
setp.ge.u32 %p2, %r230, %r1;
mov.f32 %f85, %f84;
@%p2 bra $L__BB128_5;

mov.u32 %r21, %ctaid.x;
shl.b32 %r22, %r21, 1;
mov.u32 %r23, %ctaid.y;
mad.lo.s32 %r24, %r2, %r18, %r23;
cvta.to.global.u64 %rd21, %rd18;
mul.wide.u32 %rd22, %r24, 4;
add.s64 %rd23, %rd21, %rd22;
ld.global.nc.u32 %r25, [%rd23];
mul.lo.s32 %r26, %r25, %r15;
and.b32 %r27, %r3, 15;
shr.u32 %r28, %r27, 1;
and.b32 %r29, %r3, 7;
sub.s32 %r30, %r27, %r29;
and.b32 %r31, %r3, 4;
shr.u32 %r32, %r31, 2;
add.s32 %r33, %r30, %r32;
cvt.s64.s32 %rd24, %r33;
or.b32 %r34, %r28, 3;
cvt.u64.u32 %rd25, %r34;
shl.b32 %r35, %r3, 2;
cvt.u64.u32 %rd26, %r35;
and.b64 %rd27, %rd26, 60;
mad.lo.s32 %r36, %r22, %r12, %r230;
add.s32 %r37, %r36, %r26;
mul.wide.s32 %rd3, %r37, 84;
add.s64 %rd28, %rd27, %rd3;
add.s64 %rd4, %rd28, 16;
add.s64 %rd29, %rd3, %rd24;
add.s64 %rd5, %rd29, 6;
shr.u32 %r38, %r3, 1;
and.b64 %rd6, %rd26, 28;
or.b32 %r39, %r22, 1;
add.s32 %r40, %r230, %r26;
mad.lo.s32 %r41, %r12, %r39, %r40;
mul.wide.s32 %rd7, %r41, 84;
add.s64 %rd30, %rd27, %rd7;
add.s64 %rd8, %rd30, 16;
add.s64 %rd31, %rd7, %rd24;
add.s64 %rd9, %rd31, 6;
cvt.u64.u32 %rd32, %r38;
and.b64 %rd33, %rd32, 2147483640;
or.b64 %rd34, %rd33, %rd25;
mul.lo.s32 %r42, %r2, %r13;
cvt.u64.u32 %rd35, %r42;
add.s64 %rd36, %rd34, %rd35;
mul.hi.u32 %r43, %r23, %r8;
add.s32 %r44, %r43, %r23;
shr.u32 %r45, %r44, %r9;
mul.lo.s32 %r46, %r45, %r10;
sub.s32 %r47, %r23, %r46;
mul.lo.s32 %r48, %r47, %r16;
cvt.u64.u32 %rd37, %r48;
add.s64 %rd38, %rd36, %rd37;
mul.lo.s64 %rd39, %rd38, 36;
cvta.to.global.u64 %rd40, %rd17;
add.s64 %rd41, %rd40, %rd39;
add.s64 %rd54, %rd41, -104;
cvta.to.global.u64 %rd55, %rd16;
mov.f32 %f84, 0f00000000;

$L__BB128_3:
add.s64 %rd42, %rd55, %rd4;
ld.global.nc.u32 %r125, [%rd42];
add.s64 %rd43, %rd54, %rd6;
ld.global.nc.u32 %r55, [%rd43];
ld.global.nc.u32 %r49, [%rd54+-4];

{.reg .f16 low,high;
mov.b32 {low,high},%r49;
cvt.f32.f16 %f11, low;}


ld.global.nc.u32 %r63, [%rd43+36];
ld.global.nc.u32 %r50, [%rd54+32];

{.reg .f16 low,high;
mov.b32 {low,high},%r50;
cvt.f32.f16 %f12, low;}


ld.global.nc.u32 %r71, [%rd43+72];
ld.global.nc.u32 %r51, [%rd54+68];

{.reg .f16 low,high;
mov.b32 {low,high},%r51;
cvt.f32.f16 %f13, low;}


ld.global.nc.u32 %r79, [%rd43+108];
ld.global.nc.u32 %r52, [%rd54+104];

{.reg .f16 low,high;
mov.b32 {low,high},%r52;
cvt.f32.f16 %f14, low;}


add.s64 %rd44, %rd55, %rd5;
ld.global.nc.u8 %rs1, [%rd44+-6];
cvt.u32.u16 %r126, %rs1;
and.b32 %r127, %r126, 240;
and.b32 %r54, %r125, 50529027;
mov.u32 %r122, 0;

dp4a.s32.s32 %r53, %r54, %r55, %r122;

and.b32 %r128, %r126, 15;
mul.lo.s32 %r129, %r53, %r128;
cvt.rn.f32.s32 %f23, %r129;
fma.rn.ftz.f32 %f24, %f11, %f23, 0f00000000;
shr.u32 %r130, %r127, 4;
shl.b32 %r131, %r130, 8;
or.b32 %r132, %r131, %r130;
shl.b32 %r133, %r132, 16;
or.b32 %r58, %r133, %r132;

dp4a.s32.s32 %r57, %r58, %r55, %r122;

cvt.rn.f32.s32 %f25, %r57;
fma.rn.ftz.f32 %f26, %f11, %f25, 0f00000000;
ld.global.nc.u8 %rs2, [%rd44+-4];
cvt.u32.u16 %r134, %rs2;
and.b32 %r135, %r134, 240;
shr.u32 %r136, %r125, 2;
and.b32 %r62, %r136, 50529027;

dp4a.s32.s32 %r61, %r62, %r63, %r122;

and.b32 %r137, %r134, 15;
mul.lo.s32 %r138, %r61, %r137;
cvt.rn.f32.s32 %f27, %r138;
fma.rn.ftz.f32 %f28, %f12, %f27, %f24;
shr.u32 %r139, %r135, 4;
shl.b32 %r140, %r139, 8;
or.b32 %r141, %r140, %r139;
shl.b32 %r142, %r141, 16;
or.b32 %r66, %r142, %r141;

dp4a.s32.s32 %r65, %r66, %r63, %r122;

cvt.rn.f32.s32 %f29, %r65;
fma.rn.ftz.f32 %f30, %f12, %f29, %f26;
ld.global.nc.u8 %rs3, [%rd44+-2];
cvt.u32.u16 %r143, %rs3;
and.b32 %r144, %r143, 240;
shr.u32 %r145, %r125, 4;
and.b32 %r70, %r145, 50529027;

dp4a.s32.s32 %r69, %r70, %r71, %r122;

and.b32 %r146, %r143, 15;
mul.lo.s32 %r147, %r69, %r146;
cvt.rn.f32.s32 %f31, %r147;
fma.rn.ftz.f32 %f32, %f13, %f31, %f28;
shr.u32 %r148, %r144, 4;
shl.b32 %r149, %r148, 8;
or.b32 %r150, %r149, %r148;
shl.b32 %r151, %r150, 16;
or.b32 %r74, %r151, %r150;

dp4a.s32.s32 %r73, %r74, %r71, %r122;

cvt.rn.f32.s32 %f33, %r73;
fma.rn.ftz.f32 %f34, %f13, %f33, %f30;
ld.global.nc.u8 %rs4, [%rd44];
cvt.u32.u16 %r152, %rs4;
and.b32 %r153, %r152, 240;
shr.u32 %r154, %r125, 6;
and.b32 %r78, %r154, 50529027;

dp4a.s32.s32 %r77, %r78, %r79, %r122;

and.b32 %r155, %r152, 15;
mul.lo.s32 %r156, %r77, %r155;
cvt.rn.f32.s32 %f35, %r156;
fma.rn.ftz.f32 %f36, %f14, %f35, %f32;
shr.u32 %r157, %r153, 4;
shl.b32 %r158, %r157, 8;
or.b32 %r159, %r158, %r157;
shl.b32 %r160, %r159, 16;
or.b32 %r82, %r160, %r159;

dp4a.s32.s32 %r81, %r82, %r79, %r122;

cvt.rn.f32.s32 %f37, %r81;
fma.rn.ftz.f32 %f38, %f14, %f37, %f34;
add.s64 %rd45, %rd55, %rd3;
ld.global.nc.u32 %r85, [%rd45+80];

{.reg .f16 low,high;
mov.b32 {low,high},%r85;
cvt.f32.f16 %f15, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r85;
cvt.f32.f16 %f16, high;}


mul.ftz.f32 %f39, %f36, %f15;
mul.ftz.f32 %f40, %f38, %f16;
sub.ftz.f32 %f41, %f39, %f40;
add.ftz.f32 %f85, %f41, %f85;
add.s64 %rd46, %rd55, %rd8;
ld.global.nc.u32 %r161, [%rd46];

{.reg .f16 low,high;
mov.b32 {low,high},%r49;
cvt.f32.f16 %f17, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r50;
cvt.f32.f16 %f18, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r51;
cvt.f32.f16 %f19, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r52;
cvt.f32.f16 %f20, low;}


add.s64 %rd47, %rd55, %rd9;
ld.global.nc.u8 %rs5, [%rd47+-6];
cvt.u32.u16 %r162, %rs5;
and.b32 %r163, %r162, 240;
and.b32 %r92, %r161, 50529027;

dp4a.s32.s32 %r91, %r92, %r55, %r122;

and.b32 %r164, %r162, 15;
mul.lo.s32 %r165, %r91, %r164;
cvt.rn.f32.s32 %f42, %r165;
fma.rn.ftz.f32 %f43, %f17, %f42, 0f00000000;
shr.u32 %r166, %r163, 4;
shl.b32 %r167, %r166, 8;
or.b32 %r168, %r167, %r166;
shl.b32 %r169, %r168, 16;
or.b32 %r96, %r169, %r168;

dp4a.s32.s32 %r95, %r96, %r55, %r122;

cvt.rn.f32.s32 %f44, %r95;
fma.rn.ftz.f32 %f45, %f17, %f44, 0f00000000;
ld.global.nc.u8 %rs6, [%rd47+-4];
cvt.u32.u16 %r170, %rs6;
and.b32 %r171, %r170, 240;
shr.u32 %r172, %r161, 2;
and.b32 %r100, %r172, 50529027;

dp4a.s32.s32 %r99, %r100, %r63, %r122;

and.b32 %r173, %r170, 15;
mul.lo.s32 %r174, %r99, %r173;
cvt.rn.f32.s32 %f46, %r174;
fma.rn.ftz.f32 %f47, %f18, %f46, %f43;
shr.u32 %r175, %r171, 4;
shl.b32 %r176, %r175, 8;
or.b32 %r177, %r176, %r175;
shl.b32 %r178, %r177, 16;
or.b32 %r104, %r178, %r177;

dp4a.s32.s32 %r103, %r104, %r63, %r122;

cvt.rn.f32.s32 %f48, %r103;
fma.rn.ftz.f32 %f49, %f18, %f48, %f45;
ld.global.nc.u8 %rs7, [%rd47+-2];
cvt.u32.u16 %r179, %rs7;
and.b32 %r180, %r179, 240;
shr.u32 %r181, %r161, 4;
and.b32 %r108, %r181, 50529027;

dp4a.s32.s32 %r107, %r108, %r71, %r122;

and.b32 %r182, %r179, 15;
mul.lo.s32 %r183, %r107, %r182;
cvt.rn.f32.s32 %f50, %r183;
fma.rn.ftz.f32 %f51, %f19, %f50, %f47;
shr.u32 %r184, %r180, 4;
shl.b32 %r185, %r184, 8;
or.b32 %r186, %r185, %r184;
shl.b32 %r187, %r186, 16;
or.b32 %r112, %r187, %r186;

dp4a.s32.s32 %r111, %r112, %r71, %r122;

cvt.rn.f32.s32 %f52, %r111;
fma.rn.ftz.f32 %f53, %f19, %f52, %f49;
ld.global.nc.u8 %rs8, [%rd47];
cvt.u32.u16 %r188, %rs8;
and.b32 %r189, %r188, 240;
shr.u32 %r190, %r161, 6;
and.b32 %r116, %r190, 50529027;

dp4a.s32.s32 %r115, %r116, %r79, %r122;

and.b32 %r191, %r188, 15;
mul.lo.s32 %r192, %r115, %r191;
cvt.rn.f32.s32 %f54, %r192;
fma.rn.ftz.f32 %f55, %f20, %f54, %f51;
shr.u32 %r193, %r189, 4;
shl.b32 %r194, %r193, 8;
or.b32 %r195, %r194, %r193;
shl.b32 %r196, %r195, 16;
or.b32 %r120, %r196, %r195;

dp4a.s32.s32 %r119, %r120, %r79, %r122;

cvt.rn.f32.s32 %f56, %r119;
fma.rn.ftz.f32 %f57, %f20, %f56, %f53;
add.s64 %rd48, %rd55, %rd7;
ld.global.nc.u32 %r123, [%rd48+80];

{.reg .f16 low,high;
mov.b32 {low,high},%r123;
cvt.f32.f16 %f21, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r123;
cvt.f32.f16 %f22, high;}


mul.ftz.f32 %f58, %f55, %f21;
mul.ftz.f32 %f59, %f57, %f22;
sub.ftz.f32 %f60, %f58, %f59;
add.ftz.f32 %f84, %f60, %f84;
add.s64 %rd55, %rd55, 168;
add.s64 %rd54, %rd54, 576;
add.s32 %r230, %r230, 2;
setp.lt.u32 %p3, %r230, %r1;
@%p3 bra $L__BB128_3;

st.local.v2.f32 [%rd1], {%f85, %f84};

$L__BB128_5:
mov.b32 %r197, %f85;
mov.u32 %r198, 31;
mov.u32 %r199, 16;
mov.u32 %r200, -1;
shfl.sync.bfly.b32 %r201|%p4, %r197, %r199, %r198, %r200;
mov.b32 %f61, %r201;
add.ftz.f32 %f62, %f85, %f61;
mov.b32 %r202, %f62;
mov.u32 %r203, 8;
shfl.sync.bfly.b32 %r204|%p5, %r202, %r203, %r198, %r200;
mov.b32 %f63, %r204;
add.ftz.f32 %f64, %f62, %f63;
mov.b32 %r205, %f64;
mov.u32 %r206, 4;
shfl.sync.bfly.b32 %r207|%p6, %r205, %r206, %r198, %r200;
mov.b32 %f65, %r207;
add.ftz.f32 %f66, %f64, %f65;
mov.b32 %r208, %f66;
mov.u32 %r209, 2;
shfl.sync.bfly.b32 %r210|%p7, %r208, %r209, %r198, %r200;
mov.b32 %f67, %r210;
add.ftz.f32 %f68, %f66, %f67;
mov.b32 %r211, %f68;
mov.u32 %r212, 1;
shfl.sync.bfly.b32 %r213|%p8, %r211, %r212, %r198, %r200;
mov.b32 %f69, %r213;
add.ftz.f32 %f70, %f68, %f69;
st.local.f32 [%rd1], %f70;
mov.b32 %r214, %f84;
shfl.sync.bfly.b32 %r215|%p9, %r214, %r199, %r198, %r200;
mov.b32 %f71, %r215;
add.ftz.f32 %f72, %f84, %f71;
mov.b32 %r216, %f72;
shfl.sync.bfly.b32 %r217|%p10, %r216, %r203, %r198, %r200;
mov.b32 %f73, %r217;
add.ftz.f32 %f74, %f72, %f73;
mov.b32 %r218, %f74;
shfl.sync.bfly.b32 %r219|%p11, %r218, %r206, %r198, %r200;
mov.b32 %f75, %r219;
add.ftz.f32 %f76, %f74, %f75;
mov.b32 %r220, %f76;
shfl.sync.bfly.b32 %r221|%p12, %r220, %r209, %r198, %r200;
mov.b32 %f77, %r221;
add.ftz.f32 %f78, %f76, %f77;
mov.b32 %r222, %f78;
shfl.sync.bfly.b32 %r223|%p13, %r222, %r212, %r198, %r200;
mov.b32 %f79, %r223;
add.ftz.f32 %f80, %f78, %f79;
st.local.f32 [%rd1+4], %f80;
mov.u32 %r224, %ctaid.x;
shl.b32 %r225, %r224, 1;
add.s32 %r7, %r225, %r3;
setp.ge.u32 %p14, %r7, %r11;
setp.gt.u32 %p15, %r3, 1;
or.pred %p16, %p15, %p14;
@%p16 bra $L__BB128_7;

mul.wide.u32 %rd49, %r3, 4;
add.s64 %rd50, %rd1, %rd49;
ld.local.f32 %f81, [%rd50];
mov.u32 %r226, %ctaid.y;
mad.lo.s32 %r228, %r2, %r14, %r7;
mad.lo.s32 %r229, %r226, %r17, %r228;
cvta.to.global.u64 %rd51, %rd19;
mul.wide.u32 %rd52, %r229, 4;
add.s64 %rd53, %rd51, %rd52;
st.global.f32 [%rd53], %f81;

$L__BB128_7:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<320>;
.reg .b16 %rs<17>;
.reg .f32 %f<1173>;
.reg .b32 %r<1437>;
.reg .b64 %rd<112>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate[1536];

ld.param.u64 %rd36, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd37, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd38, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r48, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r49, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r53, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r54, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r55, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r59, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r60, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r61, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r57, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r56, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r51, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r50, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r47, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r46, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r45, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
ld.param.u64 %rd41, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+16];
ld.param.u64 %rd40, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+8];
ld.param.u64 %rd39, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3];
mov.u32 %r1, %tid.y;
shl.b32 %r62, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r62, %r2;
mov.u32 %r63, %ctaid.x;
shl.b32 %r4, %r63, 2;
mov.u32 %r5, %ctaid.y;
setp.eq.s64 %p1, %rd38, 0;
@%p1 bra $L__BB129_2;

cvta.to.global.u64 %rd45, %rd38;
mul.wide.u32 %rd46, %r5, 4;
add.s64 %rd47, %rd45, %rd46;
ld.global.nc.u32 %r1433, [%rd47];
mul.hi.u32 %r64, %r5, %r45;
add.s32 %r65, %r64, %r5;
shr.u32 %r66, %r65, %r46;
mul.lo.s32 %r67, %r66, %r47;
sub.s32 %r1434, %r5, %r67;
bra.uni $L__BB129_3;

$L__BB129_2:
mul.hi.u32 %r68, %r5, %r50;
add.s32 %r69, %r68, %r5;
shr.u32 %r1433, %r69, %r51;
mov.u32 %r1434, %r5;

$L__BB129_3:
mov.u32 %r11, %ctaid.z;
selp.b32 %r12, %r5, %r1433, %p1;
setp.ne.s32 %p3, %r1, 0;
setp.gt.u32 %p4, %r2, 3;
or.pred %p5, %p4, %p3;
add.s32 %r70, %r4, %r2;
setp.ge.u32 %p6, %r70, %r49;
mov.f32 %f27, 0f00000000;
or.pred %p7, %p5, %p6;
mov.f32 %f4, %f27;
mov.f32 %f5, %f27;
@%p7 bra $L__BB129_8;

setp.eq.s64 %p8, %rd39, 0;
mov.f32 %f4, 0f00000000;
mov.f32 %f5, %f4;
@%p8 bra $L__BB129_6;

cvta.to.global.u64 %rd48, %rd39;
mul.lo.s32 %r71, %r11, %r61;
cvt.u64.u32 %rd49, %r71;
mul.lo.s32 %r72, %r12, %r55;
cvt.u64.u32 %rd50, %r72;
cvt.u64.u32 %rd51, %r2;
cvt.s64.s32 %rd52, %r4;
add.s64 %rd53, %rd52, %rd51;
add.s64 %rd54, %rd53, %rd49;
add.s64 %rd55, %rd54, %rd50;
shl.b64 %rd56, %rd55, 2;
add.s64 %rd57, %rd48, %rd56;
ld.global.f32 %f5, [%rd57];

$L__BB129_6:
setp.eq.s64 %p9, %rd41, 0;
setp.eq.s64 %p10, %rd40, 0;
or.pred %p11, %p9, %p10;
@%p11 bra $L__BB129_8;

cvta.to.global.u64 %rd58, %rd41;
mul.lo.s32 %r73, %r11, %r61;
cvt.u64.u32 %rd59, %r73;
mul.lo.s32 %r74, %r12, %r55;
cvt.u64.u32 %rd60, %r74;
cvt.u64.u32 %rd61, %r2;
cvt.s64.s32 %rd62, %r4;
add.s64 %rd63, %rd62, %rd61;
add.s64 %rd64, %rd63, %rd59;
add.s64 %rd65, %rd64, %rd60;
shl.b64 %rd66, %rd65, 2;
add.s64 %rd67, %rd58, %rd66;
ld.global.f32 %f4, [%rd67];

$L__BB129_8:
mul.lo.s32 %r75, %r11, %r60;
cvt.u64.u32 %rd68, %r75;
mul.lo.s32 %r76, %r1434, %r54;
cvt.u64.u32 %rd69, %r76;
add.s64 %rd4, %rd69, %rd68;
mul.hi.u32 %r77, %r11, %r56;
add.s32 %r78, %r77, %r11;
shr.u32 %r13, %r78, %r57;
shr.s32 %r79, %r3, 31;
shr.u32 %r80, %r79, 28;
add.s32 %r81, %r3, %r80;
shr.s32 %r1436, %r81, 4;
shr.u32 %r82, %r44, 8;
setp.ge.s32 %p12, %r1436, %r82;
mov.f32 %f28, %f27;
mov.f32 %f29, %f27;
mov.f32 %f30, %f27;
mov.f32 %f31, %f27;
mov.f32 %f32, %f27;
mov.f32 %f33, %f27;
mov.f32 %f34, %f27;
@%p12 bra $L__BB129_19;

mul.lo.s32 %r83, %r13, %r59;
mul.lo.s32 %r84, %r1433, %r53;
and.b32 %r88, %r81, -16;
sub.s32 %r89, %r3, %r88;
shr.s32 %r90, %r89, 31;
shr.u32 %r91, %r90, 29;
add.s32 %r92, %r89, %r91;
shr.s32 %r93, %r92, 3;
shl.b32 %r94, %r93, 2;
and.b32 %r95, %r92, -8;
sub.s32 %r96, %r89, %r95;
shr.s32 %r97, %r96, 31;
shr.u32 %r98, %r97, 30;
add.s32 %r99, %r96, %r98;
shr.s32 %r100, %r99, 2;
add.s32 %r101, %r95, %r100;
cvt.s64.s32 %rd5, %r96;
cvt.s64.s32 %rd6, %r94;
cvt.s64.s32 %rd70, %r101;
shl.b32 %r1435, %r1436, 3;
add.s32 %r102, %r1436, %r83;
add.s32 %r103, %r102, %r84;
add.s32 %r104, %r4, 1;
mad.lo.s32 %r105, %r48, %r104, %r103;
mul.wide.s32 %rd7, %r105, 84;
mad.lo.s32 %r106, %r4, %r48, %r1436;
add.s32 %r107, %r106, %r83;
add.s32 %r108, %r107, %r84;
mul.wide.s32 %rd8, %r108, 84;
mul.wide.s32 %rd71, %r89, 4;
add.s64 %rd72, %rd8, %rd71;
add.s64 %rd9, %rd72, 16;
add.s64 %rd73, %rd8, %rd70;
add.s64 %rd10, %rd73, 6;
add.s64 %rd11, %rd7, %rd70;
add.s64 %rd74, %rd7, %rd71;
add.s64 %rd12, %rd74, 16;
add.s32 %r109, %r4, 2;
mad.lo.s32 %r110, %r48, %r109, %r103;
mul.wide.s32 %rd13, %r110, 84;
add.s32 %r111, %r4, 3;
mad.lo.s32 %r112, %r48, %r111, %r103;
mul.wide.s32 %rd14, %r112, 84;
add.s64 %rd75, %rd14, %rd71;
add.s64 %rd15, %rd75, 16;
add.s64 %rd76, %rd14, %rd70;
add.s64 %rd16, %rd76, 6;
add.s64 %rd17, %rd13, %rd70;
add.s64 %rd77, %rd13, %rd71;
add.s64 %rd18, %rd77, 16;
mul.wide.s32 %rd19, %r48, 84;
cvta.to.global.u64 %rd111, %rd40;
cvta.to.global.u64 %rd110, %rd36;
mov.f32 %f27, 0f00000000;
cvta.to.global.u64 %rd78, %rd37;
shl.b64 %rd84, %rd5, 2;
mov.f32 %f28, %f27;
mov.f32 %f29, %f27;
mov.f32 %f30, %f27;
mov.f32 %f31, %f27;
mov.f32 %f32, %f27;
mov.f32 %f33, %f27;
mov.f32 %f34, %f27;

$L__BB129_10:
cvt.u64.u32 %rd79, %r1435;
add.s64 %rd80, %rd4, %rd79;
add.s64 %rd24, %rd110, %rd9;
ld.global.nc.u32 %r151, [%rd24];
add.s64 %rd81, %rd80, %rd6;
mul.lo.s64 %rd82, %rd81, 36;
add.s64 %rd83, %rd78, %rd82;
add.s64 %rd85, %rd83, %rd84;
ld.global.nc.u32 %r18, [%rd85+4];
ld.global.nc.u32 %r19, [%rd83];

{.reg .f16 low,high;
mov.b32 {low,high},%r19;
cvt.f32.f16 %f101, low;}


ld.global.nc.u32 %r20, [%rd85+40];
ld.global.nc.u32 %r21, [%rd83+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r21;
cvt.f32.f16 %f102, low;}


ld.global.nc.u32 %r22, [%rd85+76];
ld.global.nc.u32 %r23, [%rd83+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r23;
cvt.f32.f16 %f103, low;}


ld.global.nc.u32 %r24, [%rd85+112];
ld.global.nc.u32 %r25, [%rd83+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r25;
cvt.f32.f16 %f104, low;}


add.s64 %rd86, %rd110, %rd10;
ld.global.nc.u8 %rs1, [%rd86+-6];
cvt.u32.u16 %r152, %rs1;
and.b32 %r153, %r152, 240;
and.b32 %r118, %r151, 50529027;
mov.u32 %r148, 0;

dp4a.s32.s32 %r117, %r118, %r18, %r148;

and.b32 %r154, %r152, 15;
mul.lo.s32 %r155, %r117, %r154;
cvt.rn.f32.s32 %f107, %r155;
fma.rn.ftz.f32 %f108, %f101, %f107, 0f00000000;
shr.u32 %r156, %r153, 4;
shl.b32 %r157, %r156, 8;
or.b32 %r158, %r157, %r156;
shl.b32 %r159, %r158, 16;
or.b32 %r122, %r159, %r158;

dp4a.s32.s32 %r121, %r122, %r18, %r148;

cvt.rn.f32.s32 %f109, %r121;
fma.rn.ftz.f32 %f110, %f101, %f109, 0f00000000;
ld.global.nc.u8 %rs2, [%rd86+-4];
cvt.u32.u16 %r160, %rs2;
and.b32 %r161, %r160, 240;
shr.u32 %r162, %r151, 2;
and.b32 %r126, %r162, 50529027;

dp4a.s32.s32 %r125, %r126, %r20, %r148;

and.b32 %r163, %r160, 15;
mul.lo.s32 %r164, %r125, %r163;
cvt.rn.f32.s32 %f111, %r164;
fma.rn.ftz.f32 %f112, %f102, %f111, %f108;
shr.u32 %r165, %r161, 4;
shl.b32 %r166, %r165, 8;
or.b32 %r167, %r166, %r165;
shl.b32 %r168, %r167, 16;
or.b32 %r130, %r168, %r167;

dp4a.s32.s32 %r129, %r130, %r20, %r148;

cvt.rn.f32.s32 %f113, %r129;
fma.rn.ftz.f32 %f114, %f102, %f113, %f110;
ld.global.nc.u8 %rs3, [%rd86+-2];
cvt.u32.u16 %r169, %rs3;
and.b32 %r170, %r169, 240;
shr.u32 %r171, %r151, 4;
and.b32 %r134, %r171, 50529027;

dp4a.s32.s32 %r133, %r134, %r22, %r148;

and.b32 %r172, %r169, 15;
mul.lo.s32 %r173, %r133, %r172;
cvt.rn.f32.s32 %f115, %r173;
fma.rn.ftz.f32 %f116, %f103, %f115, %f112;
shr.u32 %r174, %r170, 4;
shl.b32 %r175, %r174, 8;
or.b32 %r176, %r175, %r174;
shl.b32 %r177, %r176, 16;
or.b32 %r138, %r177, %r176;

dp4a.s32.s32 %r137, %r138, %r22, %r148;

cvt.rn.f32.s32 %f117, %r137;
fma.rn.ftz.f32 %f118, %f103, %f117, %f114;
ld.global.nc.u8 %rs4, [%rd86];
cvt.u32.u16 %r178, %rs4;
and.b32 %r179, %r178, 240;
shr.u32 %r180, %r151, 6;
and.b32 %r142, %r180, 50529027;

dp4a.s32.s32 %r141, %r142, %r24, %r148;

and.b32 %r181, %r178, 15;
mul.lo.s32 %r182, %r141, %r181;
cvt.rn.f32.s32 %f119, %r182;
fma.rn.ftz.f32 %f120, %f104, %f119, %f116;
shr.u32 %r183, %r179, 4;
shl.b32 %r184, %r183, 8;
or.b32 %r185, %r184, %r183;
shl.b32 %r186, %r185, 16;
or.b32 %r146, %r186, %r185;

dp4a.s32.s32 %r145, %r146, %r24, %r148;

cvt.rn.f32.s32 %f121, %r145;
fma.rn.ftz.f32 %f122, %f104, %f121, %f118;
add.s64 %rd87, %rd110, %rd8;
ld.global.nc.u32 %r149, [%rd87+80];

{.reg .f16 low,high;
mov.b32 {low,high},%r149;
cvt.f32.f16 %f105, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r149;
cvt.f32.f16 %f106, high;}


mul.ftz.f32 %f123, %f120, %f105;
mul.ftz.f32 %f124, %f122, %f106;
sub.ftz.f32 %f125, %f123, %f124;
add.ftz.f32 %f30, %f125, %f30;
setp.eq.s64 %p13, %rd40, 0;
@%p13 bra $L__BB129_12;

add.s64 %rd88, %rd111, %rd9;
ld.global.u32 %r225, [%rd88];

{.reg .f16 low,high;
mov.b32 {low,high},%r19;
cvt.f32.f16 %f126, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r21;
cvt.f32.f16 %f127, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r23;
cvt.f32.f16 %f128, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r25;
cvt.f32.f16 %f129, low;}


add.s64 %rd89, %rd111, %rd10;
ld.global.u8 %r226, [%rd89+-6];
and.b32 %r192, %r225, 50529027;

dp4a.s32.s32 %r191, %r192, %r18, %r148;

and.b32 %r227, %r226, 15;
mul.lo.s32 %r228, %r191, %r227;
cvt.rn.f32.s32 %f132, %r228;
fma.rn.ftz.f32 %f133, %f126, %f132, 0f00000000;
shr.u32 %r229, %r226, 4;
shl.b32 %r230, %r229, 8;
or.b32 %r231, %r230, %r229;
shl.b32 %r232, %r231, 16;
or.b32 %r196, %r232, %r231;

dp4a.s32.s32 %r195, %r196, %r18, %r148;

cvt.rn.f32.s32 %f134, %r195;
fma.rn.ftz.f32 %f135, %f126, %f134, 0f00000000;
ld.global.u8 %r233, [%rd89+-4];
shr.u32 %r234, %r225, 2;
and.b32 %r200, %r234, 50529027;

dp4a.s32.s32 %r199, %r200, %r20, %r148;

and.b32 %r235, %r233, 15;
mul.lo.s32 %r236, %r199, %r235;
cvt.rn.f32.s32 %f136, %r236;
fma.rn.ftz.f32 %f137, %f127, %f136, %f133;
shr.u32 %r237, %r233, 4;
shl.b32 %r238, %r237, 8;
or.b32 %r239, %r238, %r237;
shl.b32 %r240, %r239, 16;
or.b32 %r204, %r240, %r239;

dp4a.s32.s32 %r203, %r204, %r20, %r148;

cvt.rn.f32.s32 %f138, %r203;
fma.rn.ftz.f32 %f139, %f127, %f138, %f135;
ld.global.u8 %r241, [%rd89+-2];
shr.u32 %r242, %r225, 4;
and.b32 %r208, %r242, 50529027;

dp4a.s32.s32 %r207, %r208, %r22, %r148;

and.b32 %r243, %r241, 15;
mul.lo.s32 %r244, %r207, %r243;
cvt.rn.f32.s32 %f140, %r244;
fma.rn.ftz.f32 %f141, %f128, %f140, %f137;
shr.u32 %r245, %r241, 4;
shl.b32 %r246, %r245, 8;
or.b32 %r247, %r246, %r245;
shl.b32 %r248, %r247, 16;
or.b32 %r212, %r248, %r247;

dp4a.s32.s32 %r211, %r212, %r22, %r148;

cvt.rn.f32.s32 %f142, %r211;
fma.rn.ftz.f32 %f143, %f128, %f142, %f139;
ld.global.u8 %r249, [%rd89];
shr.u32 %r250, %r225, 6;
and.b32 %r216, %r250, 50529027;

dp4a.s32.s32 %r215, %r216, %r24, %r148;

and.b32 %r251, %r249, 15;
mul.lo.s32 %r252, %r215, %r251;
cvt.rn.f32.s32 %f144, %r252;
fma.rn.ftz.f32 %f145, %f129, %f144, %f141;
shr.u32 %r253, %r249, 4;
shl.b32 %r254, %r253, 8;
or.b32 %r255, %r254, %r253;
shl.b32 %r256, %r255, 16;
or.b32 %r220, %r256, %r255;

dp4a.s32.s32 %r219, %r220, %r24, %r148;

cvt.rn.f32.s32 %f146, %r219;
fma.rn.ftz.f32 %f147, %f129, %f146, %f143;
add.s64 %rd90, %rd111, %rd8;
ld.global.u32 %r224, [%rd90+80];

{.reg .f16 low,high;
mov.b32 {low,high},%r224;
cvt.f32.f16 %f130, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r224;
cvt.f32.f16 %f131, high;}


mul.ftz.f32 %f148, %f145, %f130;
mul.ftz.f32 %f149, %f147, %f131;
sub.ftz.f32 %f150, %f148, %f149;
add.ftz.f32 %f34, %f150, %f34;

$L__BB129_12:
add.s64 %rd108, %rd110, %rd8;
add.s64 %rd107, %rd108, 80;
add.s64 %rd106, %rd110, %rd10;
add.s64 %rd105, %rd106, -6;
add.s64 %rd27, %rd24, %rd19;
ld.global.nc.u32 %r295, [%rd27];

{.reg .f16 low,high;
mov.b32 {low,high},%r19;
cvt.f32.f16 %f151, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r21;
cvt.f32.f16 %f152, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r23;
cvt.f32.f16 %f153, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r25;
cvt.f32.f16 %f154, low;}


add.s64 %rd28, %rd105, %rd19;
ld.global.nc.u8 %rs5, [%rd28];
cvt.u32.u16 %r296, %rs5;
and.b32 %r297, %r296, 240;
and.b32 %r262, %r295, 50529027;

dp4a.s32.s32 %r261, %r262, %r18, %r148;

and.b32 %r298, %r296, 15;
mul.lo.s32 %r299, %r261, %r298;
cvt.rn.f32.s32 %f157, %r299;
fma.rn.ftz.f32 %f158, %f151, %f157, 0f00000000;
shr.u32 %r300, %r297, 4;
shl.b32 %r301, %r300, 8;
or.b32 %r302, %r301, %r300;
shl.b32 %r303, %r302, 16;
or.b32 %r266, %r303, %r302;

dp4a.s32.s32 %r265, %r266, %r18, %r148;

cvt.rn.f32.s32 %f159, %r265;
fma.rn.ftz.f32 %f160, %f151, %f159, 0f00000000;
ld.global.nc.u8 %rs6, [%rd28+2];
cvt.u32.u16 %r304, %rs6;
and.b32 %r305, %r304, 240;
shr.u32 %r306, %r295, 2;
and.b32 %r270, %r306, 50529027;

dp4a.s32.s32 %r269, %r270, %r20, %r148;

and.b32 %r307, %r304, 15;
mul.lo.s32 %r308, %r269, %r307;
cvt.rn.f32.s32 %f161, %r308;
fma.rn.ftz.f32 %f162, %f152, %f161, %f158;
shr.u32 %r309, %r305, 4;
shl.b32 %r310, %r309, 8;
or.b32 %r311, %r310, %r309;
shl.b32 %r312, %r311, 16;
or.b32 %r274, %r312, %r311;

dp4a.s32.s32 %r273, %r274, %r20, %r148;

cvt.rn.f32.s32 %f163, %r273;
fma.rn.ftz.f32 %f164, %f152, %f163, %f160;
ld.global.nc.u8 %rs7, [%rd28+4];
cvt.u32.u16 %r313, %rs7;
and.b32 %r314, %r313, 240;
shr.u32 %r315, %r295, 4;
and.b32 %r278, %r315, 50529027;

dp4a.s32.s32 %r277, %r278, %r22, %r148;

and.b32 %r316, %r313, 15;
mul.lo.s32 %r317, %r277, %r316;
cvt.rn.f32.s32 %f165, %r317;
fma.rn.ftz.f32 %f166, %f153, %f165, %f162;
shr.u32 %r318, %r314, 4;
shl.b32 %r319, %r318, 8;
or.b32 %r320, %r319, %r318;
shl.b32 %r321, %r320, 16;
or.b32 %r282, %r321, %r320;

dp4a.s32.s32 %r281, %r282, %r22, %r148;

cvt.rn.f32.s32 %f167, %r281;
fma.rn.ftz.f32 %f168, %f153, %f167, %f164;
ld.global.nc.u8 %rs8, [%rd28+6];
cvt.u32.u16 %r322, %rs8;
and.b32 %r323, %r322, 240;
shr.u32 %r324, %r295, 6;
and.b32 %r286, %r324, 50529027;

dp4a.s32.s32 %r285, %r286, %r24, %r148;

and.b32 %r325, %r322, 15;
mul.lo.s32 %r326, %r285, %r325;
cvt.rn.f32.s32 %f169, %r326;
fma.rn.ftz.f32 %f170, %f154, %f169, %f166;
shr.u32 %r327, %r323, 4;
shl.b32 %r328, %r327, 8;
or.b32 %r329, %r328, %r327;
shl.b32 %r330, %r329, 16;
or.b32 %r290, %r330, %r329;

dp4a.s32.s32 %r289, %r290, %r24, %r148;

cvt.rn.f32.s32 %f171, %r289;
fma.rn.ftz.f32 %f172, %f154, %f171, %f168;
add.s64 %rd29, %rd107, %rd19;
ld.global.nc.u32 %r293, [%rd29];

{.reg .f16 low,high;
mov.b32 {low,high},%r293;
cvt.f32.f16 %f155, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r293;
cvt.f32.f16 %f156, high;}


mul.ftz.f32 %f173, %f170, %f155;
mul.ftz.f32 %f174, %f172, %f156;
sub.ftz.f32 %f175, %f173, %f174;
add.ftz.f32 %f29, %f175, %f29;
@%p13 bra $L__BB129_14;

add.s64 %rd91, %rd111, %rd12;
ld.global.u32 %r369, [%rd91];

{.reg .f16 low,high;
mov.b32 {low,high},%r19;
cvt.f32.f16 %f176, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r21;
cvt.f32.f16 %f177, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r23;
cvt.f32.f16 %f178, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r25;
cvt.f32.f16 %f179, low;}


add.s64 %rd92, %rd111, %rd11;
ld.global.u8 %r370, [%rd92];
and.b32 %r336, %r369, 50529027;
mov.u32 %r366, 0;

dp4a.s32.s32 %r335, %r336, %r18, %r366;

and.b32 %r371, %r370, 15;
mul.lo.s32 %r372, %r335, %r371;
cvt.rn.f32.s32 %f182, %r372;
fma.rn.ftz.f32 %f183, %f176, %f182, 0f00000000;
shr.u32 %r373, %r370, 4;
shl.b32 %r374, %r373, 8;
or.b32 %r375, %r374, %r373;
shl.b32 %r376, %r375, 16;
or.b32 %r340, %r376, %r375;

dp4a.s32.s32 %r339, %r340, %r18, %r366;

cvt.rn.f32.s32 %f184, %r339;
fma.rn.ftz.f32 %f185, %f176, %f184, 0f00000000;
ld.global.u8 %r377, [%rd92+2];
shr.u32 %r378, %r369, 2;
and.b32 %r344, %r378, 50529027;

dp4a.s32.s32 %r343, %r344, %r20, %r366;

and.b32 %r379, %r377, 15;
mul.lo.s32 %r380, %r343, %r379;
cvt.rn.f32.s32 %f186, %r380;
fma.rn.ftz.f32 %f187, %f177, %f186, %f183;
shr.u32 %r381, %r377, 4;
shl.b32 %r382, %r381, 8;
or.b32 %r383, %r382, %r381;
shl.b32 %r384, %r383, 16;
or.b32 %r348, %r384, %r383;

dp4a.s32.s32 %r347, %r348, %r20, %r366;

cvt.rn.f32.s32 %f188, %r347;
fma.rn.ftz.f32 %f189, %f177, %f188, %f185;
ld.global.u8 %r385, [%rd92+4];
shr.u32 %r386, %r369, 4;
and.b32 %r352, %r386, 50529027;

dp4a.s32.s32 %r351, %r352, %r22, %r366;

and.b32 %r387, %r385, 15;
mul.lo.s32 %r388, %r351, %r387;
cvt.rn.f32.s32 %f190, %r388;
fma.rn.ftz.f32 %f191, %f178, %f190, %f187;
shr.u32 %r389, %r385, 4;
shl.b32 %r390, %r389, 8;
or.b32 %r391, %r390, %r389;
shl.b32 %r392, %r391, 16;
or.b32 %r356, %r392, %r391;

dp4a.s32.s32 %r355, %r356, %r22, %r366;

cvt.rn.f32.s32 %f192, %r355;
fma.rn.ftz.f32 %f193, %f178, %f192, %f189;
ld.global.u8 %r393, [%rd92+6];
shr.u32 %r394, %r369, 6;
and.b32 %r360, %r394, 50529027;

dp4a.s32.s32 %r359, %r360, %r24, %r366;

and.b32 %r395, %r393, 15;
mul.lo.s32 %r396, %r359, %r395;
cvt.rn.f32.s32 %f194, %r396;
fma.rn.ftz.f32 %f195, %f179, %f194, %f191;
shr.u32 %r397, %r393, 4;
shl.b32 %r398, %r397, 8;
or.b32 %r399, %r398, %r397;
shl.b32 %r400, %r399, 16;
or.b32 %r364, %r400, %r399;

dp4a.s32.s32 %r363, %r364, %r24, %r366;

cvt.rn.f32.s32 %f196, %r363;
fma.rn.ftz.f32 %f197, %f179, %f196, %f193;
add.s64 %rd93, %rd111, %rd7;
ld.global.u32 %r368, [%rd93+80];

{.reg .f16 low,high;
mov.b32 {low,high},%r368;
cvt.f32.f16 %f180, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r368;
cvt.f32.f16 %f181, high;}


mul.ftz.f32 %f198, %f195, %f180;
mul.ftz.f32 %f199, %f197, %f181;
sub.ftz.f32 %f200, %f198, %f199;
add.ftz.f32 %f33, %f200, %f33;

$L__BB129_14:
add.s64 %rd30, %rd27, %rd19;
ld.global.nc.u32 %r439, [%rd30];

{.reg .f16 low,high;
mov.b32 {low,high},%r19;
cvt.f32.f16 %f201, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r21;
cvt.f32.f16 %f202, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r23;
cvt.f32.f16 %f203, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r25;
cvt.f32.f16 %f204, low;}


add.s64 %rd31, %rd28, %rd19;
ld.global.nc.u8 %rs9, [%rd31];
cvt.u32.u16 %r440, %rs9;
and.b32 %r441, %r440, 240;
and.b32 %r406, %r439, 50529027;
mov.u32 %r436, 0;

dp4a.s32.s32 %r405, %r406, %r18, %r436;

and.b32 %r442, %r440, 15;
mul.lo.s32 %r443, %r405, %r442;
cvt.rn.f32.s32 %f207, %r443;
fma.rn.ftz.f32 %f208, %f201, %f207, 0f00000000;
shr.u32 %r444, %r441, 4;
shl.b32 %r445, %r444, 8;
or.b32 %r446, %r445, %r444;
shl.b32 %r447, %r446, 16;
or.b32 %r410, %r447, %r446;

dp4a.s32.s32 %r409, %r410, %r18, %r436;

cvt.rn.f32.s32 %f209, %r409;
fma.rn.ftz.f32 %f210, %f201, %f209, 0f00000000;
ld.global.nc.u8 %rs10, [%rd31+2];
cvt.u32.u16 %r448, %rs10;
and.b32 %r449, %r448, 240;
shr.u32 %r450, %r439, 2;
and.b32 %r414, %r450, 50529027;

dp4a.s32.s32 %r413, %r414, %r20, %r436;

and.b32 %r451, %r448, 15;
mul.lo.s32 %r452, %r413, %r451;
cvt.rn.f32.s32 %f211, %r452;
fma.rn.ftz.f32 %f212, %f202, %f211, %f208;
shr.u32 %r453, %r449, 4;
shl.b32 %r454, %r453, 8;
or.b32 %r455, %r454, %r453;
shl.b32 %r456, %r455, 16;
or.b32 %r418, %r456, %r455;

dp4a.s32.s32 %r417, %r418, %r20, %r436;

cvt.rn.f32.s32 %f213, %r417;
fma.rn.ftz.f32 %f214, %f202, %f213, %f210;
ld.global.nc.u8 %rs11, [%rd31+4];
cvt.u32.u16 %r457, %rs11;
and.b32 %r458, %r457, 240;
shr.u32 %r459, %r439, 4;
and.b32 %r422, %r459, 50529027;

dp4a.s32.s32 %r421, %r422, %r22, %r436;

and.b32 %r460, %r457, 15;
mul.lo.s32 %r461, %r421, %r460;
cvt.rn.f32.s32 %f215, %r461;
fma.rn.ftz.f32 %f216, %f203, %f215, %f212;
shr.u32 %r462, %r458, 4;
shl.b32 %r463, %r462, 8;
or.b32 %r464, %r463, %r462;
shl.b32 %r465, %r464, 16;
or.b32 %r426, %r465, %r464;

dp4a.s32.s32 %r425, %r426, %r22, %r436;

cvt.rn.f32.s32 %f217, %r425;
fma.rn.ftz.f32 %f218, %f203, %f217, %f214;
ld.global.nc.u8 %rs12, [%rd31+6];
cvt.u32.u16 %r466, %rs12;
and.b32 %r467, %r466, 240;
shr.u32 %r468, %r439, 6;
and.b32 %r430, %r468, 50529027;

dp4a.s32.s32 %r429, %r430, %r24, %r436;

and.b32 %r469, %r466, 15;
mul.lo.s32 %r470, %r429, %r469;
cvt.rn.f32.s32 %f219, %r470;
fma.rn.ftz.f32 %f220, %f204, %f219, %f216;
shr.u32 %r471, %r467, 4;
shl.b32 %r472, %r471, 8;
or.b32 %r473, %r472, %r471;
shl.b32 %r474, %r473, 16;
or.b32 %r434, %r474, %r473;

dp4a.s32.s32 %r433, %r434, %r24, %r436;

cvt.rn.f32.s32 %f221, %r433;
fma.rn.ftz.f32 %f222, %f204, %f221, %f218;
add.s64 %rd32, %rd29, %rd19;
ld.global.nc.u32 %r437, [%rd32];

{.reg .f16 low,high;
mov.b32 {low,high},%r437;
cvt.f32.f16 %f205, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r437;
cvt.f32.f16 %f206, high;}


mul.ftz.f32 %f223, %f220, %f205;
mul.ftz.f32 %f224, %f222, %f206;
sub.ftz.f32 %f225, %f223, %f224;
add.ftz.f32 %f28, %f225, %f28;
@%p13 bra $L__BB129_16;

add.s64 %rd94, %rd111, %rd18;
ld.global.u32 %r513, [%rd94];

{.reg .f16 low,high;
mov.b32 {low,high},%r19;
cvt.f32.f16 %f226, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r21;
cvt.f32.f16 %f227, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r23;
cvt.f32.f16 %f228, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r25;
cvt.f32.f16 %f229, low;}


add.s64 %rd95, %rd111, %rd17;
ld.global.u8 %r514, [%rd95];
and.b32 %r480, %r513, 50529027;

dp4a.s32.s32 %r479, %r480, %r18, %r436;

and.b32 %r515, %r514, 15;
mul.lo.s32 %r516, %r479, %r515;
cvt.rn.f32.s32 %f232, %r516;
fma.rn.ftz.f32 %f233, %f226, %f232, 0f00000000;
shr.u32 %r517, %r514, 4;
shl.b32 %r518, %r517, 8;
or.b32 %r519, %r518, %r517;
shl.b32 %r520, %r519, 16;
or.b32 %r484, %r520, %r519;

dp4a.s32.s32 %r483, %r484, %r18, %r436;

cvt.rn.f32.s32 %f234, %r483;
fma.rn.ftz.f32 %f235, %f226, %f234, 0f00000000;
ld.global.u8 %r521, [%rd95+2];
shr.u32 %r522, %r513, 2;
and.b32 %r488, %r522, 50529027;

dp4a.s32.s32 %r487, %r488, %r20, %r436;

and.b32 %r523, %r521, 15;
mul.lo.s32 %r524, %r487, %r523;
cvt.rn.f32.s32 %f236, %r524;
fma.rn.ftz.f32 %f237, %f227, %f236, %f233;
shr.u32 %r525, %r521, 4;
shl.b32 %r526, %r525, 8;
or.b32 %r527, %r526, %r525;
shl.b32 %r528, %r527, 16;
or.b32 %r492, %r528, %r527;

dp4a.s32.s32 %r491, %r492, %r20, %r436;

cvt.rn.f32.s32 %f238, %r491;
fma.rn.ftz.f32 %f239, %f227, %f238, %f235;
ld.global.u8 %r529, [%rd95+4];
shr.u32 %r530, %r513, 4;
and.b32 %r496, %r530, 50529027;

dp4a.s32.s32 %r495, %r496, %r22, %r436;

and.b32 %r531, %r529, 15;
mul.lo.s32 %r532, %r495, %r531;
cvt.rn.f32.s32 %f240, %r532;
fma.rn.ftz.f32 %f241, %f228, %f240, %f237;
shr.u32 %r533, %r529, 4;
shl.b32 %r534, %r533, 8;
or.b32 %r535, %r534, %r533;
shl.b32 %r536, %r535, 16;
or.b32 %r500, %r536, %r535;

dp4a.s32.s32 %r499, %r500, %r22, %r436;

cvt.rn.f32.s32 %f242, %r499;
fma.rn.ftz.f32 %f243, %f228, %f242, %f239;
ld.global.u8 %r537, [%rd95+6];
shr.u32 %r538, %r513, 6;
and.b32 %r504, %r538, 50529027;

dp4a.s32.s32 %r503, %r504, %r24, %r436;

and.b32 %r539, %r537, 15;
mul.lo.s32 %r540, %r503, %r539;
cvt.rn.f32.s32 %f244, %r540;
fma.rn.ftz.f32 %f245, %f229, %f244, %f241;
shr.u32 %r541, %r537, 4;
shl.b32 %r542, %r541, 8;
or.b32 %r543, %r542, %r541;
shl.b32 %r544, %r543, 16;
or.b32 %r508, %r544, %r543;

dp4a.s32.s32 %r507, %r508, %r24, %r436;

cvt.rn.f32.s32 %f246, %r507;
fma.rn.ftz.f32 %f247, %f229, %f246, %f243;
add.s64 %rd96, %rd111, %rd13;
ld.global.u32 %r512, [%rd96+80];

{.reg .f16 low,high;
mov.b32 {low,high},%r512;
cvt.f32.f16 %f230, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r512;
cvt.f32.f16 %f231, high;}


mul.ftz.f32 %f248, %f245, %f230;
mul.ftz.f32 %f249, %f247, %f231;
sub.ftz.f32 %f250, %f248, %f249;
add.ftz.f32 %f32, %f250, %f32;

$L__BB129_16:
add.s64 %rd97, %rd30, %rd19;
ld.global.nc.u32 %r583, [%rd97];

{.reg .f16 low,high;
mov.b32 {low,high},%r19;
cvt.f32.f16 %f251, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r21;
cvt.f32.f16 %f252, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r23;
cvt.f32.f16 %f253, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r25;
cvt.f32.f16 %f254, low;}


add.s64 %rd98, %rd31, %rd19;
ld.global.nc.u8 %rs13, [%rd98];
cvt.u32.u16 %r584, %rs13;
and.b32 %r585, %r584, 240;
and.b32 %r550, %r583, 50529027;

dp4a.s32.s32 %r549, %r550, %r18, %r436;

and.b32 %r586, %r584, 15;
mul.lo.s32 %r587, %r549, %r586;
cvt.rn.f32.s32 %f257, %r587;
fma.rn.ftz.f32 %f258, %f251, %f257, 0f00000000;
shr.u32 %r588, %r585, 4;
shl.b32 %r589, %r588, 8;
or.b32 %r590, %r589, %r588;
shl.b32 %r591, %r590, 16;
or.b32 %r554, %r591, %r590;

dp4a.s32.s32 %r553, %r554, %r18, %r436;

cvt.rn.f32.s32 %f259, %r553;
fma.rn.ftz.f32 %f260, %f251, %f259, 0f00000000;
ld.global.nc.u8 %rs14, [%rd98+2];
cvt.u32.u16 %r592, %rs14;
and.b32 %r593, %r592, 240;
shr.u32 %r594, %r583, 2;
and.b32 %r558, %r594, 50529027;

dp4a.s32.s32 %r557, %r558, %r20, %r436;

and.b32 %r595, %r592, 15;
mul.lo.s32 %r596, %r557, %r595;
cvt.rn.f32.s32 %f261, %r596;
fma.rn.ftz.f32 %f262, %f252, %f261, %f258;
shr.u32 %r597, %r593, 4;
shl.b32 %r598, %r597, 8;
or.b32 %r599, %r598, %r597;
shl.b32 %r600, %r599, 16;
or.b32 %r562, %r600, %r599;

dp4a.s32.s32 %r561, %r562, %r20, %r436;

cvt.rn.f32.s32 %f263, %r561;
fma.rn.ftz.f32 %f264, %f252, %f263, %f260;
ld.global.nc.u8 %rs15, [%rd98+4];
cvt.u32.u16 %r601, %rs15;
and.b32 %r602, %r601, 240;
shr.u32 %r603, %r583, 4;
and.b32 %r566, %r603, 50529027;

dp4a.s32.s32 %r565, %r566, %r22, %r436;

and.b32 %r604, %r601, 15;
mul.lo.s32 %r605, %r565, %r604;
cvt.rn.f32.s32 %f265, %r605;
fma.rn.ftz.f32 %f266, %f253, %f265, %f262;
shr.u32 %r606, %r602, 4;
shl.b32 %r607, %r606, 8;
or.b32 %r608, %r607, %r606;
shl.b32 %r609, %r608, 16;
or.b32 %r570, %r609, %r608;

dp4a.s32.s32 %r569, %r570, %r22, %r436;

cvt.rn.f32.s32 %f267, %r569;
fma.rn.ftz.f32 %f268, %f253, %f267, %f264;
ld.global.nc.u8 %rs16, [%rd98+6];
cvt.u32.u16 %r610, %rs16;
and.b32 %r611, %r610, 240;
shr.u32 %r612, %r583, 6;
and.b32 %r574, %r612, 50529027;

dp4a.s32.s32 %r573, %r574, %r24, %r436;

and.b32 %r613, %r610, 15;
mul.lo.s32 %r614, %r573, %r613;
cvt.rn.f32.s32 %f269, %r614;
fma.rn.ftz.f32 %f270, %f254, %f269, %f266;
shr.u32 %r615, %r611, 4;
shl.b32 %r616, %r615, 8;
or.b32 %r617, %r616, %r615;
shl.b32 %r618, %r617, 16;
or.b32 %r578, %r618, %r617;

dp4a.s32.s32 %r577, %r578, %r24, %r436;

cvt.rn.f32.s32 %f271, %r577;
fma.rn.ftz.f32 %f272, %f254, %f271, %f268;
add.s64 %rd99, %rd32, %rd19;
ld.global.nc.u32 %r581, [%rd99];

{.reg .f16 low,high;
mov.b32 {low,high},%r581;
cvt.f32.f16 %f255, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r581;
cvt.f32.f16 %f256, high;}


mul.ftz.f32 %f273, %f270, %f255;
mul.ftz.f32 %f274, %f272, %f256;
sub.ftz.f32 %f275, %f273, %f274;
add.ftz.f32 %f27, %f275, %f27;
@%p13 bra $L__BB129_18;

add.s64 %rd100, %rd111, %rd15;
ld.global.u32 %r657, [%rd100];

{.reg .f16 low,high;
mov.b32 {low,high},%r19;
cvt.f32.f16 %f276, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r21;
cvt.f32.f16 %f277, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r23;
cvt.f32.f16 %f278, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r25;
cvt.f32.f16 %f279, low;}


add.s64 %rd101, %rd111, %rd16;
ld.global.u8 %r658, [%rd101+-6];
and.b32 %r624, %r657, 50529027;
mov.u32 %r654, 0;

dp4a.s32.s32 %r623, %r624, %r18, %r654;

and.b32 %r659, %r658, 15;
mul.lo.s32 %r660, %r623, %r659;
cvt.rn.f32.s32 %f282, %r660;
fma.rn.ftz.f32 %f283, %f276, %f282, 0f00000000;
shr.u32 %r661, %r658, 4;
shl.b32 %r662, %r661, 8;
or.b32 %r663, %r662, %r661;
shl.b32 %r664, %r663, 16;
or.b32 %r628, %r664, %r663;

dp4a.s32.s32 %r627, %r628, %r18, %r654;

cvt.rn.f32.s32 %f284, %r627;
fma.rn.ftz.f32 %f285, %f276, %f284, 0f00000000;
ld.global.u8 %r665, [%rd101+-4];
shr.u32 %r666, %r657, 2;
and.b32 %r632, %r666, 50529027;

dp4a.s32.s32 %r631, %r632, %r20, %r654;

and.b32 %r667, %r665, 15;
mul.lo.s32 %r668, %r631, %r667;
cvt.rn.f32.s32 %f286, %r668;
fma.rn.ftz.f32 %f287, %f277, %f286, %f283;
shr.u32 %r669, %r665, 4;
shl.b32 %r670, %r669, 8;
or.b32 %r671, %r670, %r669;
shl.b32 %r672, %r671, 16;
or.b32 %r636, %r672, %r671;

dp4a.s32.s32 %r635, %r636, %r20, %r654;

cvt.rn.f32.s32 %f288, %r635;
fma.rn.ftz.f32 %f289, %f277, %f288, %f285;
ld.global.u8 %r673, [%rd101+-2];
shr.u32 %r674, %r657, 4;
and.b32 %r640, %r674, 50529027;

dp4a.s32.s32 %r639, %r640, %r22, %r654;

and.b32 %r675, %r673, 15;
mul.lo.s32 %r676, %r639, %r675;
cvt.rn.f32.s32 %f290, %r676;
fma.rn.ftz.f32 %f291, %f278, %f290, %f287;
shr.u32 %r677, %r673, 4;
shl.b32 %r678, %r677, 8;
or.b32 %r679, %r678, %r677;
shl.b32 %r680, %r679, 16;
or.b32 %r644, %r680, %r679;

dp4a.s32.s32 %r643, %r644, %r22, %r654;

cvt.rn.f32.s32 %f292, %r643;
fma.rn.ftz.f32 %f293, %f278, %f292, %f289;
ld.global.u8 %r681, [%rd101];
shr.u32 %r682, %r657, 6;
and.b32 %r648, %r682, 50529027;

dp4a.s32.s32 %r647, %r648, %r24, %r654;

and.b32 %r683, %r681, 15;
mul.lo.s32 %r684, %r647, %r683;
cvt.rn.f32.s32 %f294, %r684;
fma.rn.ftz.f32 %f295, %f279, %f294, %f291;
shr.u32 %r685, %r681, 4;
shl.b32 %r686, %r685, 8;
or.b32 %r687, %r686, %r685;
shl.b32 %r688, %r687, 16;
or.b32 %r652, %r688, %r687;

dp4a.s32.s32 %r651, %r652, %r24, %r654;

cvt.rn.f32.s32 %f296, %r651;
fma.rn.ftz.f32 %f297, %f279, %f296, %f293;
add.s64 %rd102, %rd111, %rd14;
ld.global.u32 %r656, [%rd102+80];

{.reg .f16 low,high;
mov.b32 {low,high},%r656;
cvt.f32.f16 %f280, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r656;
cvt.f32.f16 %f281, high;}


mul.ftz.f32 %f298, %f295, %f280;
mul.ftz.f32 %f299, %f297, %f281;
sub.ftz.f32 %f300, %f298, %f299;
add.ftz.f32 %f31, %f300, %f31;

$L__BB129_18:
add.s32 %r1435, %r1435, 64;
add.s64 %rd111, %rd111, 672;
add.s64 %rd110, %rd110, 672;
add.s32 %r1436, %r1436, 8;
setp.lt.s32 %p17, %r1436, %r82;
@%p17 bra $L__BB129_10;

$L__BB129_19:
mov.u32 %r1422, %tid.y;
setp.eq.s32 %p18, %r1422, 0;
@%p18 bra $L__BB129_23;

mov.u32 %r1423, %tid.y;
setp.ne.s64 %p19, %rd40, 0;
add.s32 %r28, %r1423, -1;
shl.b32 %r693, %r1423, 9;
mov.u32 %r694, _ZZ13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r695, %r694, %r693;
shl.b32 %r696, %r2, 2;
add.s32 %r697, %r695, %r696;
st.shared.f32 [%r697+-512], %f30;
@%p19 bra $L__BB129_22;
bra.uni $L__BB129_21;

$L__BB129_22:
shl.b32 %r699, %r28, 9;
mov.u32 %r700, _ZZ13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r701, %r700, %r699;
add.s32 %r703, %r701, %r696;
st.shared.f32 [%r703], %f34;
add.s32 %r1419, %r697, -512;
st.shared.f32 [%r1419+128], %f29;
st.shared.f32 [%r703+128], %f33;
add.s32 %r1420, %r697, -512;
st.shared.f32 [%r1420+256], %f28;
st.shared.f32 [%r703+256], %f32;
add.s32 %r1421, %r697, -512;
st.shared.f32 [%r1421+384], %f27;
st.shared.f32 [%r703+384], %f31;
bra.uni $L__BB129_23;

$L__BB129_21:
add.s32 %r1416, %r697, -512;
st.shared.f32 [%r1416+128], %f29;
add.s32 %r1417, %r697, -512;
st.shared.f32 [%r1417+256], %f28;
add.s32 %r1418, %r697, -512;
st.shared.f32 [%r1418+384], %f27;

$L__BB129_23:
mov.u32 %r1432, %tid.y;
setp.ne.s32 %p319, %r1432, 0;
bar.sync 0;
@%p319 bra $L__BB129_78;
bra.uni $L__BB129_24;

$L__BB129_78:
ret;

$L__BB129_24:
ld.param.f32 %f1145, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
ld.param.u64 %rd109, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r1426, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r1425, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
mov.u32 %r1424, %ctaid.z;
setp.ne.s64 %p21, %rd40, 0;
mov.u32 %r706, %ctaid.y;
mad.lo.s32 %r709, %r706, %r1426, %r4;
mad.lo.s32 %r710, %r1424, %r1425, %r709;
neg.ftz.f32 %f35, %f1145;
shl.b32 %r712, %r2, 2;
mov.u32 %r713, _ZZ13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r30, %r713, %r712;
cvta.to.global.u64 %rd103, %rd109;
mul.wide.u32 %rd104, %r710, 4;
add.s64 %rd35, %rd103, %rd104;
@%p21 bra $L__BB129_33;
bra.uni $L__BB129_25;

$L__BB129_33:
ld.param.u32 %r1427, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
mov.u32 %r800, _ZZ13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r32, %r800, %r712;
setp.gt.s32 %p54, %r1427, 2;
@%p54 bra $L__BB129_44;

ld.param.u32 %r1430, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p57, %r1430, 1;
@%p57 bra $L__BB129_62;

ld.param.u32 %r1431, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p58, %r1431, 2;
@%p58 bra $L__BB129_36;
bra.uni $L__BB129_70;

$L__BB129_36:
ld.shared.f32 %f701, [%r30];
add.ftz.f32 %f702, %f701, %f30;
ld.shared.f32 %f703, [%r32];
add.ftz.f32 %f704, %f703, %f34;
ld.shared.f32 %f705, [%r30+512];
add.ftz.f32 %f706, %f705, %f702;
ld.shared.f32 %f707, [%r32+512];
add.ftz.f32 %f708, %f707, %f704;
ld.shared.f32 %f709, [%r30+1024];
add.ftz.f32 %f710, %f709, %f706;
ld.shared.f32 %f711, [%r32+1024];
add.ftz.f32 %f712, %f711, %f708;
mov.b32 %r1047, %f710;
mov.u32 %r1048, 31;
mov.u32 %r1049, 16;
mov.u32 %r1050, -1;
shfl.sync.bfly.b32 %r1051|%p163, %r1047, %r1049, %r1048, %r1050;
mov.b32 %f713, %r1051;
add.ftz.f32 %f714, %f710, %f713;
mov.b32 %r1052, %f714;
mov.u32 %r1053, 8;
shfl.sync.bfly.b32 %r1054|%p164, %r1052, %r1053, %r1048, %r1050;
mov.b32 %f715, %r1054;
add.ftz.f32 %f716, %f714, %f715;
mov.b32 %r1055, %f716;
mov.u32 %r1056, 4;
shfl.sync.bfly.b32 %r1057|%p165, %r1055, %r1056, %r1048, %r1050;
mov.b32 %f717, %r1057;
add.ftz.f32 %f718, %f716, %f717;
mov.b32 %r1058, %f718;
mov.u32 %r1059, 2;
shfl.sync.bfly.b32 %r1060|%p166, %r1058, %r1059, %r1048, %r1050;
mov.b32 %f719, %r1060;
add.ftz.f32 %f720, %f718, %f719;
mov.b32 %r1061, %f720;
mov.u32 %r1062, 1;
shfl.sync.bfly.b32 %r1063|%p167, %r1061, %r1062, %r1048, %r1050;
mov.b32 %f721, %r1063;
add.ftz.f32 %f56, %f720, %f721;
mov.b32 %r1064, %f712;
shfl.sync.bfly.b32 %r1065|%p168, %r1064, %r1049, %r1048, %r1050;
mov.b32 %f722, %r1065;
add.ftz.f32 %f723, %f712, %f722;
mov.b32 %r1066, %f723;
shfl.sync.bfly.b32 %r1067|%p169, %r1066, %r1053, %r1048, %r1050;
mov.b32 %f724, %r1067;
add.ftz.f32 %f725, %f723, %f724;
mov.b32 %r1068, %f725;
shfl.sync.bfly.b32 %r1069|%p170, %r1068, %r1056, %r1048, %r1050;
mov.b32 %f726, %r1069;
add.ftz.f32 %f727, %f725, %f726;
mov.b32 %r1070, %f727;
shfl.sync.bfly.b32 %r1071|%p171, %r1070, %r1059, %r1048, %r1050;
mov.b32 %f728, %r1071;
add.ftz.f32 %f729, %f727, %f728;
mov.b32 %r1072, %f729;
shfl.sync.bfly.b32 %r1073|%p172, %r1072, %r1062, %r1048, %r1050;
mov.b32 %f730, %r1073;
add.ftz.f32 %f57, %f729, %f730;
setp.ne.s32 %p173, %r2, 0;
setp.ge.u32 %p174, %r4, %r49;
or.pred %p175, %p173, %p174;
@%p175 bra $L__BB129_38;

add.ftz.f32 %f731, %f5, %f56;
add.ftz.f32 %f732, %f4, %f57;
mul.ftz.f32 %f733, %f732, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f734, %f733;
add.ftz.f32 %f735, %f734, 0f3F800000;
div.approx.ftz.f32 %f736, %f732, %f735;
mul.ftz.f32 %f737, %f731, %f736;
st.global.f32 [%rd35], %f737;

$L__BB129_38:
ld.shared.f32 %f738, [%r30+128];
add.ftz.f32 %f739, %f738, %f29;
ld.shared.f32 %f740, [%r32+128];
add.ftz.f32 %f741, %f740, %f33;
ld.shared.f32 %f742, [%r30+640];
add.ftz.f32 %f743, %f742, %f739;
ld.shared.f32 %f744, [%r32+640];
add.ftz.f32 %f745, %f744, %f741;
ld.shared.f32 %f746, [%r30+1152];
add.ftz.f32 %f747, %f746, %f743;
ld.shared.f32 %f748, [%r32+1152];
add.ftz.f32 %f749, %f748, %f745;
mov.b32 %r1077, %f747;
shfl.sync.bfly.b32 %r1081|%p176, %r1077, %r1049, %r1048, %r1050;
mov.b32 %f750, %r1081;
add.ftz.f32 %f751, %f747, %f750;
mov.b32 %r1082, %f751;
shfl.sync.bfly.b32 %r1084|%p177, %r1082, %r1053, %r1048, %r1050;
mov.b32 %f752, %r1084;
add.ftz.f32 %f753, %f751, %f752;
mov.b32 %r1085, %f753;
shfl.sync.bfly.b32 %r1087|%p178, %r1085, %r1056, %r1048, %r1050;
mov.b32 %f754, %r1087;
add.ftz.f32 %f755, %f753, %f754;
mov.b32 %r1088, %f755;
shfl.sync.bfly.b32 %r1090|%p179, %r1088, %r1059, %r1048, %r1050;
mov.b32 %f756, %r1090;
add.ftz.f32 %f757, %f755, %f756;
mov.b32 %r1091, %f757;
shfl.sync.bfly.b32 %r1093|%p180, %r1091, %r1062, %r1048, %r1050;
mov.b32 %f758, %r1093;
add.ftz.f32 %f58, %f757, %f758;
mov.b32 %r1094, %f749;
shfl.sync.bfly.b32 %r1095|%p181, %r1094, %r1049, %r1048, %r1050;
mov.b32 %f759, %r1095;
add.ftz.f32 %f760, %f749, %f759;
mov.b32 %r1096, %f760;
shfl.sync.bfly.b32 %r1097|%p182, %r1096, %r1053, %r1048, %r1050;
mov.b32 %f761, %r1097;
add.ftz.f32 %f762, %f760, %f761;
mov.b32 %r1098, %f762;
shfl.sync.bfly.b32 %r1099|%p183, %r1098, %r1056, %r1048, %r1050;
mov.b32 %f763, %r1099;
add.ftz.f32 %f764, %f762, %f763;
mov.b32 %r1100, %f764;
shfl.sync.bfly.b32 %r1101|%p184, %r1100, %r1059, %r1048, %r1050;
mov.b32 %f765, %r1101;
add.ftz.f32 %f766, %f764, %f765;
mov.b32 %r1102, %f766;
shfl.sync.bfly.b32 %r1103|%p185, %r1102, %r1062, %r1048, %r1050;
mov.b32 %f767, %r1103;
add.ftz.f32 %f59, %f766, %f767;
setp.ne.s32 %p186, %r2, 1;
or.b32 %r1107, %r4, 1;
setp.ge.u32 %p187, %r1107, %r49;
or.pred %p188, %p186, %p187;
@%p188 bra $L__BB129_40;

add.ftz.f32 %f768, %f5, %f58;
add.ftz.f32 %f769, %f4, %f59;
mul.ftz.f32 %f770, %f769, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f771, %f770;
add.ftz.f32 %f772, %f771, 0f3F800000;
div.approx.ftz.f32 %f773, %f769, %f772;
mul.ftz.f32 %f774, %f768, %f773;
st.global.f32 [%rd35+4], %f774;

$L__BB129_40:
ld.shared.f32 %f775, [%r30+256];
add.ftz.f32 %f776, %f775, %f28;
ld.shared.f32 %f777, [%r32+256];
add.ftz.f32 %f778, %f777, %f32;
ld.shared.f32 %f779, [%r30+768];
add.ftz.f32 %f780, %f779, %f776;
ld.shared.f32 %f781, [%r32+768];
add.ftz.f32 %f782, %f781, %f778;
ld.shared.f32 %f783, [%r30+1280];
add.ftz.f32 %f784, %f783, %f780;
ld.shared.f32 %f785, [%r32+1280];
add.ftz.f32 %f786, %f785, %f782;
mov.b32 %r1108, %f784;
mov.u32 %r1109, 31;
mov.u32 %r1110, 16;
mov.u32 %r1111, -1;
shfl.sync.bfly.b32 %r1112|%p189, %r1108, %r1110, %r1109, %r1111;
mov.b32 %f787, %r1112;
add.ftz.f32 %f788, %f784, %f787;
mov.b32 %r1113, %f788;
mov.u32 %r1114, 8;
shfl.sync.bfly.b32 %r1115|%p190, %r1113, %r1114, %r1109, %r1111;
mov.b32 %f789, %r1115;
add.ftz.f32 %f790, %f788, %f789;
mov.b32 %r1116, %f790;
mov.u32 %r1117, 4;
shfl.sync.bfly.b32 %r1118|%p191, %r1116, %r1117, %r1109, %r1111;
mov.b32 %f791, %r1118;
add.ftz.f32 %f792, %f790, %f791;
mov.b32 %r1119, %f792;
mov.u32 %r1120, 2;
shfl.sync.bfly.b32 %r1121|%p192, %r1119, %r1120, %r1109, %r1111;
mov.b32 %f793, %r1121;
add.ftz.f32 %f794, %f792, %f793;
mov.b32 %r1122, %f794;
mov.u32 %r1123, 1;
shfl.sync.bfly.b32 %r1124|%p193, %r1122, %r1123, %r1109, %r1111;
mov.b32 %f795, %r1124;
add.ftz.f32 %f60, %f794, %f795;
mov.b32 %r1125, %f786;
shfl.sync.bfly.b32 %r1126|%p194, %r1125, %r1110, %r1109, %r1111;
mov.b32 %f796, %r1126;
add.ftz.f32 %f797, %f786, %f796;
mov.b32 %r1127, %f797;
shfl.sync.bfly.b32 %r1128|%p195, %r1127, %r1114, %r1109, %r1111;
mov.b32 %f798, %r1128;
add.ftz.f32 %f799, %f797, %f798;
mov.b32 %r1129, %f799;
shfl.sync.bfly.b32 %r1130|%p196, %r1129, %r1117, %r1109, %r1111;
mov.b32 %f800, %r1130;
add.ftz.f32 %f801, %f799, %f800;
mov.b32 %r1131, %f801;
shfl.sync.bfly.b32 %r1132|%p197, %r1131, %r1120, %r1109, %r1111;
mov.b32 %f802, %r1132;
add.ftz.f32 %f803, %f801, %f802;
mov.b32 %r1133, %f803;
shfl.sync.bfly.b32 %r1134|%p198, %r1133, %r1123, %r1109, %r1111;
mov.b32 %f804, %r1134;
add.ftz.f32 %f61, %f803, %f804;
setp.ne.s32 %p199, %r2, 2;
or.b32 %r1138, %r4, 2;
setp.ge.u32 %p200, %r1138, %r49;
or.pred %p201, %p199, %p200;
@%p201 bra $L__BB129_42;

add.ftz.f32 %f805, %f5, %f60;
add.ftz.f32 %f806, %f4, %f61;
mul.ftz.f32 %f807, %f806, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f808, %f807;
add.ftz.f32 %f809, %f808, 0f3F800000;
div.approx.ftz.f32 %f810, %f806, %f809;
mul.ftz.f32 %f811, %f805, %f810;
st.global.f32 [%rd35+8], %f811;

$L__BB129_42:
ld.shared.f32 %f812, [%r30+384];
add.ftz.f32 %f813, %f812, %f27;
ld.shared.f32 %f814, [%r32+384];
add.ftz.f32 %f815, %f814, %f31;
ld.shared.f32 %f816, [%r30+896];
add.ftz.f32 %f817, %f816, %f813;
ld.shared.f32 %f818, [%r32+896];
add.ftz.f32 %f819, %f818, %f815;
ld.shared.f32 %f820, [%r30+1408];
add.ftz.f32 %f821, %f820, %f817;
ld.shared.f32 %f822, [%r32+1408];
add.ftz.f32 %f823, %f822, %f819;
mov.b32 %r1139, %f821;
shfl.sync.bfly.b32 %r1143|%p202, %r1139, %r1110, %r1109, %r1111;
mov.b32 %f824, %r1143;
add.ftz.f32 %f825, %f821, %f824;
mov.b32 %r1144, %f825;
shfl.sync.bfly.b32 %r1146|%p203, %r1144, %r1114, %r1109, %r1111;
mov.b32 %f826, %r1146;
add.ftz.f32 %f827, %f825, %f826;
mov.b32 %r1147, %f827;
shfl.sync.bfly.b32 %r1149|%p204, %r1147, %r1117, %r1109, %r1111;
mov.b32 %f828, %r1149;
add.ftz.f32 %f829, %f827, %f828;
mov.b32 %r1150, %f829;
shfl.sync.bfly.b32 %r1152|%p205, %r1150, %r1120, %r1109, %r1111;
mov.b32 %f830, %r1152;
add.ftz.f32 %f831, %f829, %f830;
mov.b32 %r1153, %f831;
shfl.sync.bfly.b32 %r1155|%p206, %r1153, %r1123, %r1109, %r1111;
mov.b32 %f832, %r1155;
add.ftz.f32 %f62, %f831, %f832;
mov.b32 %r1156, %f823;
shfl.sync.bfly.b32 %r1157|%p207, %r1156, %r1110, %r1109, %r1111;
mov.b32 %f833, %r1157;
add.ftz.f32 %f834, %f823, %f833;
mov.b32 %r1158, %f834;
shfl.sync.bfly.b32 %r1159|%p208, %r1158, %r1114, %r1109, %r1111;
mov.b32 %f835, %r1159;
add.ftz.f32 %f836, %f834, %f835;
mov.b32 %r1160, %f836;
shfl.sync.bfly.b32 %r1161|%p209, %r1160, %r1117, %r1109, %r1111;
mov.b32 %f837, %r1161;
add.ftz.f32 %f838, %f836, %f837;
mov.b32 %r1162, %f838;
shfl.sync.bfly.b32 %r1163|%p210, %r1162, %r1120, %r1109, %r1111;
mov.b32 %f839, %r1163;
add.ftz.f32 %f840, %f838, %f839;
mov.b32 %r1164, %f840;
shfl.sync.bfly.b32 %r1165|%p211, %r1164, %r1123, %r1109, %r1111;
mov.b32 %f841, %r1165;
add.ftz.f32 %f63, %f840, %f841;
setp.ne.s32 %p212, %r2, 3;
or.b32 %r1169, %r4, 3;
setp.ge.u32 %p213, %r1169, %r49;
or.pred %p214, %p212, %p213;
@%p214 bra $L__BB129_78;

add.ftz.f32 %f842, %f5, %f62;
add.ftz.f32 %f843, %f4, %f63;
mul.ftz.f32 %f844, %f843, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f845, %f844;
add.ftz.f32 %f846, %f845, 0f3F800000;
div.approx.ftz.f32 %f847, %f843, %f846;
mul.ftz.f32 %f848, %f842, %f847;
st.global.f32 [%rd35+12], %f848;
bra.uni $L__BB129_78;

$L__BB129_25:
ld.shared.f32 %f301, [%r30];
add.ftz.f32 %f302, %f301, %f30;
ld.shared.f32 %f303, [%r30+512];
add.ftz.f32 %f304, %f303, %f302;
ld.shared.f32 %f305, [%r30+1024];
add.ftz.f32 %f306, %f305, %f304;
mov.b32 %r714, %f306;
mov.u32 %r715, 31;
mov.u32 %r716, 16;
mov.u32 %r717, -1;
shfl.sync.bfly.b32 %r718|%p22, %r714, %r716, %r715, %r717;
mov.b32 %f307, %r718;
add.ftz.f32 %f308, %f306, %f307;
mov.b32 %r719, %f308;
mov.u32 %r720, 8;
shfl.sync.bfly.b32 %r721|%p23, %r719, %r720, %r715, %r717;
mov.b32 %f309, %r721;
add.ftz.f32 %f310, %f308, %f309;
mov.b32 %r722, %f310;
mov.u32 %r723, 4;
shfl.sync.bfly.b32 %r724|%p24, %r722, %r723, %r715, %r717;
mov.b32 %f311, %r724;
add.ftz.f32 %f312, %f310, %f311;
mov.b32 %r725, %f312;
mov.u32 %r726, 2;
shfl.sync.bfly.b32 %r727|%p25, %r725, %r726, %r715, %r717;
mov.b32 %f313, %r727;
add.ftz.f32 %f314, %f312, %f313;
mov.b32 %r728, %f314;
mov.u32 %r729, 1;
shfl.sync.bfly.b32 %r730|%p26, %r728, %r729, %r715, %r717;
mov.b32 %f315, %r730;
add.ftz.f32 %f36, %f314, %f315;
setp.ne.s32 %p27, %r2, 0;
setp.ge.u32 %p28, %r4, %r49;
or.pred %p29, %p27, %p28;
@%p29 bra $L__BB129_27;

add.ftz.f32 %f316, %f5, %f36;
st.global.f32 [%rd35], %f316;

$L__BB129_27:
ld.shared.f32 %f317, [%r30+128];
add.ftz.f32 %f318, %f317, %f29;
ld.shared.f32 %f319, [%r30+640];
add.ftz.f32 %f320, %f319, %f318;
ld.shared.f32 %f321, [%r30+1152];
add.ftz.f32 %f322, %f321, %f320;
mov.b32 %r734, %f322;
shfl.sync.bfly.b32 %r738|%p30, %r734, %r716, %r715, %r717;
mov.b32 %f323, %r738;
add.ftz.f32 %f324, %f322, %f323;
mov.b32 %r739, %f324;
shfl.sync.bfly.b32 %r741|%p31, %r739, %r720, %r715, %r717;
mov.b32 %f325, %r741;
add.ftz.f32 %f326, %f324, %f325;
mov.b32 %r742, %f326;
shfl.sync.bfly.b32 %r744|%p32, %r742, %r723, %r715, %r717;
mov.b32 %f327, %r744;
add.ftz.f32 %f328, %f326, %f327;
mov.b32 %r745, %f328;
shfl.sync.bfly.b32 %r747|%p33, %r745, %r726, %r715, %r717;
mov.b32 %f329, %r747;
add.ftz.f32 %f330, %f328, %f329;
mov.b32 %r748, %f330;
shfl.sync.bfly.b32 %r750|%p34, %r748, %r729, %r715, %r717;
mov.b32 %f331, %r750;
add.ftz.f32 %f37, %f330, %f331;
setp.ne.s32 %p35, %r2, 1;
or.b32 %r754, %r4, 1;
setp.ge.u32 %p36, %r754, %r49;
or.pred %p37, %p35, %p36;
@%p37 bra $L__BB129_29;

add.ftz.f32 %f332, %f5, %f37;
st.global.f32 [%rd35+4], %f332;

$L__BB129_29:
ld.shared.f32 %f333, [%r30+256];
add.ftz.f32 %f334, %f333, %f28;
ld.shared.f32 %f335, [%r30+768];
add.ftz.f32 %f336, %f335, %f334;
ld.shared.f32 %f337, [%r30+1280];
add.ftz.f32 %f338, %f337, %f336;
mov.b32 %r755, %f338;
mov.u32 %r756, 31;
mov.u32 %r757, 16;
mov.u32 %r758, -1;
shfl.sync.bfly.b32 %r759|%p38, %r755, %r757, %r756, %r758;
mov.b32 %f339, %r759;
add.ftz.f32 %f340, %f338, %f339;
mov.b32 %r760, %f340;
mov.u32 %r761, 8;
shfl.sync.bfly.b32 %r762|%p39, %r760, %r761, %r756, %r758;
mov.b32 %f341, %r762;
add.ftz.f32 %f342, %f340, %f341;
mov.b32 %r763, %f342;
mov.u32 %r764, 4;
shfl.sync.bfly.b32 %r765|%p40, %r763, %r764, %r756, %r758;
mov.b32 %f343, %r765;
add.ftz.f32 %f344, %f342, %f343;
mov.b32 %r766, %f344;
mov.u32 %r767, 2;
shfl.sync.bfly.b32 %r768|%p41, %r766, %r767, %r756, %r758;
mov.b32 %f345, %r768;
add.ftz.f32 %f346, %f344, %f345;
mov.b32 %r769, %f346;
mov.u32 %r770, 1;
shfl.sync.bfly.b32 %r771|%p42, %r769, %r770, %r756, %r758;
mov.b32 %f347, %r771;
add.ftz.f32 %f38, %f346, %f347;
setp.ne.s32 %p43, %r2, 2;
or.b32 %r775, %r4, 2;
setp.ge.u32 %p44, %r775, %r49;
or.pred %p45, %p43, %p44;
@%p45 bra $L__BB129_31;

add.ftz.f32 %f348, %f5, %f38;
st.global.f32 [%rd35+8], %f348;

$L__BB129_31:
ld.shared.f32 %f349, [%r30+384];
add.ftz.f32 %f350, %f349, %f27;
ld.shared.f32 %f351, [%r30+896];
add.ftz.f32 %f352, %f351, %f350;
ld.shared.f32 %f353, [%r30+1408];
add.ftz.f32 %f354, %f353, %f352;
mov.b32 %r776, %f354;
shfl.sync.bfly.b32 %r780|%p46, %r776, %r757, %r756, %r758;
mov.b32 %f355, %r780;
add.ftz.f32 %f356, %f354, %f355;
mov.b32 %r781, %f356;
shfl.sync.bfly.b32 %r783|%p47, %r781, %r761, %r756, %r758;
mov.b32 %f357, %r783;
add.ftz.f32 %f358, %f356, %f357;
mov.b32 %r784, %f358;
shfl.sync.bfly.b32 %r786|%p48, %r784, %r764, %r756, %r758;
mov.b32 %f359, %r786;
add.ftz.f32 %f360, %f358, %f359;
mov.b32 %r787, %f360;
shfl.sync.bfly.b32 %r789|%p49, %r787, %r767, %r756, %r758;
mov.b32 %f361, %r789;
add.ftz.f32 %f362, %f360, %f361;
mov.b32 %r790, %f362;
shfl.sync.bfly.b32 %r792|%p50, %r790, %r770, %r756, %r758;
mov.b32 %f363, %r792;
add.ftz.f32 %f39, %f362, %f363;
setp.ne.s32 %p51, %r2, 3;
or.b32 %r796, %r4, 3;
setp.ge.u32 %p52, %r796, %r49;
or.pred %p53, %p51, %p52;
@%p53 bra $L__BB129_78;

add.ftz.f32 %f364, %f5, %f39;
st.global.f32 [%rd35+12], %f364;
bra.uni $L__BB129_78;

$L__BB129_44:
ld.param.u32 %r1428, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p55, %r1428, 3;
@%p55 bra $L__BB129_54;

ld.param.u32 %r1429, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.ne.s32 %p56, %r1429, 6;
@%p56 bra $L__BB129_70;

ld.shared.f32 %f365, [%r30];
add.ftz.f32 %f366, %f365, %f30;
ld.shared.f32 %f367, [%r32];
add.ftz.f32 %f368, %f367, %f34;
ld.shared.f32 %f369, [%r30+512];
add.ftz.f32 %f370, %f369, %f366;
ld.shared.f32 %f371, [%r32+512];
add.ftz.f32 %f372, %f371, %f368;
ld.shared.f32 %f373, [%r30+1024];
add.ftz.f32 %f374, %f373, %f370;
ld.shared.f32 %f375, [%r32+1024];
add.ftz.f32 %f376, %f375, %f372;
mov.b32 %r801, %f374;
mov.u32 %r802, 31;
mov.u32 %r803, 16;
mov.u32 %r804, -1;
shfl.sync.bfly.b32 %r805|%p59, %r801, %r803, %r802, %r804;
mov.b32 %f377, %r805;
add.ftz.f32 %f378, %f374, %f377;
mov.b32 %r806, %f378;
mov.u32 %r807, 8;
shfl.sync.bfly.b32 %r808|%p60, %r806, %r807, %r802, %r804;
mov.b32 %f379, %r808;
add.ftz.f32 %f380, %f378, %f379;
mov.b32 %r809, %f380;
mov.u32 %r810, 4;
shfl.sync.bfly.b32 %r811|%p61, %r809, %r810, %r802, %r804;
mov.b32 %f381, %r811;
add.ftz.f32 %f382, %f380, %f381;
mov.b32 %r812, %f382;
mov.u32 %r813, 2;
shfl.sync.bfly.b32 %r814|%p62, %r812, %r813, %r802, %r804;
mov.b32 %f383, %r814;
add.ftz.f32 %f384, %f382, %f383;
mov.b32 %r815, %f384;
mov.u32 %r816, 1;
shfl.sync.bfly.b32 %r817|%p63, %r815, %r816, %r802, %r804;
mov.b32 %f385, %r817;
add.ftz.f32 %f40, %f384, %f385;
mov.b32 %r818, %f376;
shfl.sync.bfly.b32 %r819|%p64, %r818, %r803, %r802, %r804;
mov.b32 %f386, %r819;
add.ftz.f32 %f387, %f376, %f386;
mov.b32 %r820, %f387;
shfl.sync.bfly.b32 %r821|%p65, %r820, %r807, %r802, %r804;
mov.b32 %f388, %r821;
add.ftz.f32 %f389, %f387, %f388;
mov.b32 %r822, %f389;
shfl.sync.bfly.b32 %r823|%p66, %r822, %r810, %r802, %r804;
mov.b32 %f390, %r823;
add.ftz.f32 %f391, %f389, %f390;
mov.b32 %r824, %f391;
shfl.sync.bfly.b32 %r825|%p67, %r824, %r813, %r802, %r804;
mov.b32 %f392, %r825;
add.ftz.f32 %f393, %f391, %f392;
mov.b32 %r826, %f393;
shfl.sync.bfly.b32 %r827|%p68, %r826, %r816, %r802, %r804;
mov.b32 %f394, %r827;
add.ftz.f32 %f41, %f393, %f394;
setp.ne.s32 %p69, %r2, 0;
setp.ge.u32 %p70, %r4, %r49;
or.pred %p71, %p69, %p70;
@%p71 bra $L__BB129_48;

ld.param.f32 %f1149, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f395, %f5, %f40;
add.ftz.f32 %f396, %f4, %f41;
min.ftz.f32 %f397, %f396, %f1149;
min.ftz.f32 %f398, %f395, %f1149;
max.ftz.f32 %f399, %f398, %f35;
mul.ftz.f32 %f400, %f397, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f401, %f400;
add.ftz.f32 %f402, %f401, 0f3F800000;
div.approx.ftz.f32 %f403, %f397, %f402;
mul.ftz.f32 %f404, %f399, %f403;
st.global.f32 [%rd35], %f404;

$L__BB129_48:
ld.shared.f32 %f405, [%r30+128];
add.ftz.f32 %f406, %f405, %f29;
ld.shared.f32 %f407, [%r32+128];
add.ftz.f32 %f408, %f407, %f33;
ld.shared.f32 %f409, [%r30+640];
add.ftz.f32 %f410, %f409, %f406;
ld.shared.f32 %f411, [%r32+640];
add.ftz.f32 %f412, %f411, %f408;
ld.shared.f32 %f413, [%r30+1152];
add.ftz.f32 %f414, %f413, %f410;
ld.shared.f32 %f415, [%r32+1152];
add.ftz.f32 %f416, %f415, %f412;
mov.b32 %r831, %f414;
shfl.sync.bfly.b32 %r835|%p72, %r831, %r803, %r802, %r804;
mov.b32 %f417, %r835;
add.ftz.f32 %f418, %f414, %f417;
mov.b32 %r836, %f418;
shfl.sync.bfly.b32 %r838|%p73, %r836, %r807, %r802, %r804;
mov.b32 %f419, %r838;
add.ftz.f32 %f420, %f418, %f419;
mov.b32 %r839, %f420;
shfl.sync.bfly.b32 %r841|%p74, %r839, %r810, %r802, %r804;
mov.b32 %f421, %r841;
add.ftz.f32 %f422, %f420, %f421;
mov.b32 %r842, %f422;
shfl.sync.bfly.b32 %r844|%p75, %r842, %r813, %r802, %r804;
mov.b32 %f423, %r844;
add.ftz.f32 %f424, %f422, %f423;
mov.b32 %r845, %f424;
shfl.sync.bfly.b32 %r847|%p76, %r845, %r816, %r802, %r804;
mov.b32 %f425, %r847;
add.ftz.f32 %f42, %f424, %f425;
mov.b32 %r848, %f416;
shfl.sync.bfly.b32 %r849|%p77, %r848, %r803, %r802, %r804;
mov.b32 %f426, %r849;
add.ftz.f32 %f427, %f416, %f426;
mov.b32 %r850, %f427;
shfl.sync.bfly.b32 %r851|%p78, %r850, %r807, %r802, %r804;
mov.b32 %f428, %r851;
add.ftz.f32 %f429, %f427, %f428;
mov.b32 %r852, %f429;
shfl.sync.bfly.b32 %r853|%p79, %r852, %r810, %r802, %r804;
mov.b32 %f430, %r853;
add.ftz.f32 %f431, %f429, %f430;
mov.b32 %r854, %f431;
shfl.sync.bfly.b32 %r855|%p80, %r854, %r813, %r802, %r804;
mov.b32 %f432, %r855;
add.ftz.f32 %f433, %f431, %f432;
mov.b32 %r856, %f433;
shfl.sync.bfly.b32 %r857|%p81, %r856, %r816, %r802, %r804;
mov.b32 %f434, %r857;
add.ftz.f32 %f43, %f433, %f434;
setp.ne.s32 %p82, %r2, 1;
or.b32 %r861, %r4, 1;
setp.ge.u32 %p83, %r861, %r49;
or.pred %p84, %p82, %p83;
@%p84 bra $L__BB129_50;

ld.param.f32 %f1148, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f435, %f5, %f42;
add.ftz.f32 %f436, %f4, %f43;
min.ftz.f32 %f437, %f436, %f1148;
min.ftz.f32 %f438, %f435, %f1148;
max.ftz.f32 %f439, %f438, %f35;
mul.ftz.f32 %f440, %f437, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f441, %f440;
add.ftz.f32 %f442, %f441, 0f3F800000;
div.approx.ftz.f32 %f443, %f437, %f442;
mul.ftz.f32 %f444, %f439, %f443;
st.global.f32 [%rd35+4], %f444;

$L__BB129_50:
ld.shared.f32 %f445, [%r30+256];
add.ftz.f32 %f446, %f445, %f28;
ld.shared.f32 %f447, [%r32+256];
add.ftz.f32 %f448, %f447, %f32;
ld.shared.f32 %f449, [%r30+768];
add.ftz.f32 %f450, %f449, %f446;
ld.shared.f32 %f451, [%r32+768];
add.ftz.f32 %f452, %f451, %f448;
ld.shared.f32 %f453, [%r30+1280];
add.ftz.f32 %f454, %f453, %f450;
ld.shared.f32 %f455, [%r32+1280];
add.ftz.f32 %f456, %f455, %f452;
mov.b32 %r862, %f454;
mov.u32 %r863, 31;
mov.u32 %r864, 16;
mov.u32 %r865, -1;
shfl.sync.bfly.b32 %r866|%p85, %r862, %r864, %r863, %r865;
mov.b32 %f457, %r866;
add.ftz.f32 %f458, %f454, %f457;
mov.b32 %r867, %f458;
mov.u32 %r868, 8;
shfl.sync.bfly.b32 %r869|%p86, %r867, %r868, %r863, %r865;
mov.b32 %f459, %r869;
add.ftz.f32 %f460, %f458, %f459;
mov.b32 %r870, %f460;
mov.u32 %r871, 4;
shfl.sync.bfly.b32 %r872|%p87, %r870, %r871, %r863, %r865;
mov.b32 %f461, %r872;
add.ftz.f32 %f462, %f460, %f461;
mov.b32 %r873, %f462;
mov.u32 %r874, 2;
shfl.sync.bfly.b32 %r875|%p88, %r873, %r874, %r863, %r865;
mov.b32 %f463, %r875;
add.ftz.f32 %f464, %f462, %f463;
mov.b32 %r876, %f464;
mov.u32 %r877, 1;
shfl.sync.bfly.b32 %r878|%p89, %r876, %r877, %r863, %r865;
mov.b32 %f465, %r878;
add.ftz.f32 %f44, %f464, %f465;
mov.b32 %r879, %f456;
shfl.sync.bfly.b32 %r880|%p90, %r879, %r864, %r863, %r865;
mov.b32 %f466, %r880;
add.ftz.f32 %f467, %f456, %f466;
mov.b32 %r881, %f467;
shfl.sync.bfly.b32 %r882|%p91, %r881, %r868, %r863, %r865;
mov.b32 %f468, %r882;
add.ftz.f32 %f469, %f467, %f468;
mov.b32 %r883, %f469;
shfl.sync.bfly.b32 %r884|%p92, %r883, %r871, %r863, %r865;
mov.b32 %f470, %r884;
add.ftz.f32 %f471, %f469, %f470;
mov.b32 %r885, %f471;
shfl.sync.bfly.b32 %r886|%p93, %r885, %r874, %r863, %r865;
mov.b32 %f472, %r886;
add.ftz.f32 %f473, %f471, %f472;
mov.b32 %r887, %f473;
shfl.sync.bfly.b32 %r888|%p94, %r887, %r877, %r863, %r865;
mov.b32 %f474, %r888;
add.ftz.f32 %f45, %f473, %f474;
setp.ne.s32 %p95, %r2, 2;
or.b32 %r892, %r4, 2;
setp.ge.u32 %p96, %r892, %r49;
or.pred %p97, %p95, %p96;
@%p97 bra $L__BB129_52;

ld.param.f32 %f1147, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f475, %f5, %f44;
add.ftz.f32 %f476, %f4, %f45;
min.ftz.f32 %f477, %f476, %f1147;
min.ftz.f32 %f478, %f475, %f1147;
max.ftz.f32 %f479, %f478, %f35;
mul.ftz.f32 %f480, %f477, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f481, %f480;
add.ftz.f32 %f482, %f481, 0f3F800000;
div.approx.ftz.f32 %f483, %f477, %f482;
mul.ftz.f32 %f484, %f479, %f483;
st.global.f32 [%rd35+8], %f484;

$L__BB129_52:
ld.shared.f32 %f485, [%r30+384];
add.ftz.f32 %f486, %f485, %f27;
ld.shared.f32 %f487, [%r32+384];
add.ftz.f32 %f488, %f487, %f31;
ld.shared.f32 %f489, [%r30+896];
add.ftz.f32 %f490, %f489, %f486;
ld.shared.f32 %f491, [%r32+896];
add.ftz.f32 %f492, %f491, %f488;
ld.shared.f32 %f493, [%r30+1408];
add.ftz.f32 %f494, %f493, %f490;
ld.shared.f32 %f495, [%r32+1408];
add.ftz.f32 %f496, %f495, %f492;
mov.b32 %r893, %f494;
shfl.sync.bfly.b32 %r897|%p98, %r893, %r864, %r863, %r865;
mov.b32 %f497, %r897;
add.ftz.f32 %f498, %f494, %f497;
mov.b32 %r898, %f498;
shfl.sync.bfly.b32 %r900|%p99, %r898, %r868, %r863, %r865;
mov.b32 %f499, %r900;
add.ftz.f32 %f500, %f498, %f499;
mov.b32 %r901, %f500;
shfl.sync.bfly.b32 %r903|%p100, %r901, %r871, %r863, %r865;
mov.b32 %f501, %r903;
add.ftz.f32 %f502, %f500, %f501;
mov.b32 %r904, %f502;
shfl.sync.bfly.b32 %r906|%p101, %r904, %r874, %r863, %r865;
mov.b32 %f503, %r906;
add.ftz.f32 %f504, %f502, %f503;
mov.b32 %r907, %f504;
shfl.sync.bfly.b32 %r909|%p102, %r907, %r877, %r863, %r865;
mov.b32 %f505, %r909;
add.ftz.f32 %f46, %f504, %f505;
mov.b32 %r910, %f496;
shfl.sync.bfly.b32 %r911|%p103, %r910, %r864, %r863, %r865;
mov.b32 %f506, %r911;
add.ftz.f32 %f507, %f496, %f506;
mov.b32 %r912, %f507;
shfl.sync.bfly.b32 %r913|%p104, %r912, %r868, %r863, %r865;
mov.b32 %f508, %r913;
add.ftz.f32 %f509, %f507, %f508;
mov.b32 %r914, %f509;
shfl.sync.bfly.b32 %r915|%p105, %r914, %r871, %r863, %r865;
mov.b32 %f510, %r915;
add.ftz.f32 %f511, %f509, %f510;
mov.b32 %r916, %f511;
shfl.sync.bfly.b32 %r917|%p106, %r916, %r874, %r863, %r865;
mov.b32 %f512, %r917;
add.ftz.f32 %f513, %f511, %f512;
mov.b32 %r918, %f513;
shfl.sync.bfly.b32 %r919|%p107, %r918, %r877, %r863, %r865;
mov.b32 %f514, %r919;
add.ftz.f32 %f47, %f513, %f514;
setp.ne.s32 %p108, %r2, 3;
or.b32 %r923, %r4, 3;
setp.ge.u32 %p109, %r923, %r49;
or.pred %p110, %p108, %p109;
@%p110 bra $L__BB129_78;

ld.param.f32 %f1146, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f515, %f5, %f46;
add.ftz.f32 %f516, %f4, %f47;
min.ftz.f32 %f517, %f516, %f1146;
min.ftz.f32 %f518, %f515, %f1146;
max.ftz.f32 %f519, %f518, %f35;
mul.ftz.f32 %f520, %f517, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f521, %f520;
add.ftz.f32 %f522, %f521, 0f3F800000;
div.approx.ftz.f32 %f523, %f517, %f522;
mul.ftz.f32 %f524, %f519, %f523;
st.global.f32 [%rd35+12], %f524;
bra.uni $L__BB129_78;

$L__BB129_70:
ld.shared.f32 %f1013, [%r30];
add.ftz.f32 %f1014, %f1013, %f30;
ld.shared.f32 %f1015, [%r32];
add.ftz.f32 %f1016, %f1015, %f34;
ld.shared.f32 %f1017, [%r30+512];
add.ftz.f32 %f1018, %f1017, %f1014;
ld.shared.f32 %f1019, [%r32+512];
add.ftz.f32 %f1020, %f1019, %f1016;
ld.shared.f32 %f1021, [%r30+1024];
add.ftz.f32 %f1022, %f1021, %f1018;
ld.shared.f32 %f1023, [%r32+1024];
add.ftz.f32 %f1024, %f1023, %f1020;
mov.b32 %r1293, %f1022;
mov.u32 %r1294, 31;
mov.u32 %r1295, 16;
mov.u32 %r1296, -1;
shfl.sync.bfly.b32 %r1297|%p267, %r1293, %r1295, %r1294, %r1296;
mov.b32 %f1025, %r1297;
add.ftz.f32 %f1026, %f1022, %f1025;
mov.b32 %r1298, %f1026;
mov.u32 %r1299, 8;
shfl.sync.bfly.b32 %r1300|%p268, %r1298, %r1299, %r1294, %r1296;
mov.b32 %f1027, %r1300;
add.ftz.f32 %f1028, %f1026, %f1027;
mov.b32 %r1301, %f1028;
mov.u32 %r1302, 4;
shfl.sync.bfly.b32 %r1303|%p269, %r1301, %r1302, %r1294, %r1296;
mov.b32 %f1029, %r1303;
add.ftz.f32 %f1030, %f1028, %f1029;
mov.b32 %r1304, %f1030;
mov.u32 %r1305, 2;
shfl.sync.bfly.b32 %r1306|%p270, %r1304, %r1305, %r1294, %r1296;
mov.b32 %f1031, %r1306;
add.ftz.f32 %f1032, %f1030, %f1031;
mov.b32 %r1307, %f1032;
mov.u32 %r1308, 1;
shfl.sync.bfly.b32 %r1309|%p271, %r1307, %r1308, %r1294, %r1296;
mov.b32 %f1033, %r1309;
add.ftz.f32 %f72, %f1032, %f1033;
mov.b32 %r1310, %f1024;
shfl.sync.bfly.b32 %r1311|%p272, %r1310, %r1295, %r1294, %r1296;
mov.b32 %f1034, %r1311;
add.ftz.f32 %f1035, %f1024, %f1034;
mov.b32 %r1312, %f1035;
shfl.sync.bfly.b32 %r1313|%p273, %r1312, %r1299, %r1294, %r1296;
mov.b32 %f1036, %r1313;
add.ftz.f32 %f1037, %f1035, %f1036;
mov.b32 %r1314, %f1037;
shfl.sync.bfly.b32 %r1315|%p274, %r1314, %r1302, %r1294, %r1296;
mov.b32 %f1038, %r1315;
add.ftz.f32 %f1039, %f1037, %f1038;
mov.b32 %r1316, %f1039;
shfl.sync.bfly.b32 %r1317|%p275, %r1316, %r1305, %r1294, %r1296;
mov.b32 %f1040, %r1317;
add.ftz.f32 %f1041, %f1039, %f1040;
mov.b32 %r1318, %f1041;
shfl.sync.bfly.b32 %r1319|%p276, %r1318, %r1308, %r1294, %r1296;
mov.b32 %f1042, %r1319;
add.ftz.f32 %f73, %f1041, %f1042;
setp.ne.s32 %p277, %r2, 0;
setp.ge.u32 %p278, %r4, %r49;
or.pred %p279, %p277, %p278;
@%p279 bra $L__BB129_72;

add.ftz.f32 %f1043, %f5, %f72;
add.ftz.f32 %f1044, %f4, %f73;
mul.ftz.f32 %f1045, %f1043, %f1044;
st.global.f32 [%rd35], %f1045;

$L__BB129_72:
ld.shared.f32 %f1046, [%r30+128];
add.ftz.f32 %f1047, %f1046, %f29;
ld.shared.f32 %f1048, [%r32+128];
add.ftz.f32 %f1049, %f1048, %f33;
ld.shared.f32 %f1050, [%r30+640];
add.ftz.f32 %f1051, %f1050, %f1047;
ld.shared.f32 %f1052, [%r32+640];
add.ftz.f32 %f1053, %f1052, %f1049;
ld.shared.f32 %f1054, [%r30+1152];
add.ftz.f32 %f1055, %f1054, %f1051;
ld.shared.f32 %f1056, [%r32+1152];
add.ftz.f32 %f1057, %f1056, %f1053;
mov.b32 %r1323, %f1055;
shfl.sync.bfly.b32 %r1327|%p280, %r1323, %r1295, %r1294, %r1296;
mov.b32 %f1058, %r1327;
add.ftz.f32 %f1059, %f1055, %f1058;
mov.b32 %r1328, %f1059;
shfl.sync.bfly.b32 %r1330|%p281, %r1328, %r1299, %r1294, %r1296;
mov.b32 %f1060, %r1330;
add.ftz.f32 %f1061, %f1059, %f1060;
mov.b32 %r1331, %f1061;
shfl.sync.bfly.b32 %r1333|%p282, %r1331, %r1302, %r1294, %r1296;
mov.b32 %f1062, %r1333;
add.ftz.f32 %f1063, %f1061, %f1062;
mov.b32 %r1334, %f1063;
shfl.sync.bfly.b32 %r1336|%p283, %r1334, %r1305, %r1294, %r1296;
mov.b32 %f1064, %r1336;
add.ftz.f32 %f1065, %f1063, %f1064;
mov.b32 %r1337, %f1065;
shfl.sync.bfly.b32 %r1339|%p284, %r1337, %r1308, %r1294, %r1296;
mov.b32 %f1066, %r1339;
add.ftz.f32 %f74, %f1065, %f1066;
mov.b32 %r1340, %f1057;
shfl.sync.bfly.b32 %r1341|%p285, %r1340, %r1295, %r1294, %r1296;
mov.b32 %f1067, %r1341;
add.ftz.f32 %f1068, %f1057, %f1067;
mov.b32 %r1342, %f1068;
shfl.sync.bfly.b32 %r1343|%p286, %r1342, %r1299, %r1294, %r1296;
mov.b32 %f1069, %r1343;
add.ftz.f32 %f1070, %f1068, %f1069;
mov.b32 %r1344, %f1070;
shfl.sync.bfly.b32 %r1345|%p287, %r1344, %r1302, %r1294, %r1296;
mov.b32 %f1071, %r1345;
add.ftz.f32 %f1072, %f1070, %f1071;
mov.b32 %r1346, %f1072;
shfl.sync.bfly.b32 %r1347|%p288, %r1346, %r1305, %r1294, %r1296;
mov.b32 %f1073, %r1347;
add.ftz.f32 %f1074, %f1072, %f1073;
mov.b32 %r1348, %f1074;
shfl.sync.bfly.b32 %r1349|%p289, %r1348, %r1308, %r1294, %r1296;
mov.b32 %f1075, %r1349;
add.ftz.f32 %f75, %f1074, %f1075;
setp.ne.s32 %p290, %r2, 1;
or.b32 %r1353, %r4, 1;
setp.ge.u32 %p291, %r1353, %r49;
or.pred %p292, %p290, %p291;
@%p292 bra $L__BB129_74;

add.ftz.f32 %f1076, %f5, %f74;
add.ftz.f32 %f1077, %f4, %f75;
mul.ftz.f32 %f1078, %f1076, %f1077;
st.global.f32 [%rd35+4], %f1078;

$L__BB129_74:
ld.shared.f32 %f1079, [%r30+256];
add.ftz.f32 %f1080, %f1079, %f28;
ld.shared.f32 %f1081, [%r32+256];
add.ftz.f32 %f1082, %f1081, %f32;
ld.shared.f32 %f1083, [%r30+768];
add.ftz.f32 %f1084, %f1083, %f1080;
ld.shared.f32 %f1085, [%r32+768];
add.ftz.f32 %f1086, %f1085, %f1082;
ld.shared.f32 %f1087, [%r30+1280];
add.ftz.f32 %f1088, %f1087, %f1084;
ld.shared.f32 %f1089, [%r32+1280];
add.ftz.f32 %f1090, %f1089, %f1086;
mov.b32 %r1354, %f1088;
mov.u32 %r1355, 31;
mov.u32 %r1356, 16;
mov.u32 %r1357, -1;
shfl.sync.bfly.b32 %r1358|%p293, %r1354, %r1356, %r1355, %r1357;
mov.b32 %f1091, %r1358;
add.ftz.f32 %f1092, %f1088, %f1091;
mov.b32 %r1359, %f1092;
mov.u32 %r1360, 8;
shfl.sync.bfly.b32 %r1361|%p294, %r1359, %r1360, %r1355, %r1357;
mov.b32 %f1093, %r1361;
add.ftz.f32 %f1094, %f1092, %f1093;
mov.b32 %r1362, %f1094;
mov.u32 %r1363, 4;
shfl.sync.bfly.b32 %r1364|%p295, %r1362, %r1363, %r1355, %r1357;
mov.b32 %f1095, %r1364;
add.ftz.f32 %f1096, %f1094, %f1095;
mov.b32 %r1365, %f1096;
mov.u32 %r1366, 2;
shfl.sync.bfly.b32 %r1367|%p296, %r1365, %r1366, %r1355, %r1357;
mov.b32 %f1097, %r1367;
add.ftz.f32 %f1098, %f1096, %f1097;
mov.b32 %r1368, %f1098;
mov.u32 %r1369, 1;
shfl.sync.bfly.b32 %r1370|%p297, %r1368, %r1369, %r1355, %r1357;
mov.b32 %f1099, %r1370;
add.ftz.f32 %f76, %f1098, %f1099;
mov.b32 %r1371, %f1090;
shfl.sync.bfly.b32 %r1372|%p298, %r1371, %r1356, %r1355, %r1357;
mov.b32 %f1100, %r1372;
add.ftz.f32 %f1101, %f1090, %f1100;
mov.b32 %r1373, %f1101;
shfl.sync.bfly.b32 %r1374|%p299, %r1373, %r1360, %r1355, %r1357;
mov.b32 %f1102, %r1374;
add.ftz.f32 %f1103, %f1101, %f1102;
mov.b32 %r1375, %f1103;
shfl.sync.bfly.b32 %r1376|%p300, %r1375, %r1363, %r1355, %r1357;
mov.b32 %f1104, %r1376;
add.ftz.f32 %f1105, %f1103, %f1104;
mov.b32 %r1377, %f1105;
shfl.sync.bfly.b32 %r1378|%p301, %r1377, %r1366, %r1355, %r1357;
mov.b32 %f1106, %r1378;
add.ftz.f32 %f1107, %f1105, %f1106;
mov.b32 %r1379, %f1107;
shfl.sync.bfly.b32 %r1380|%p302, %r1379, %r1369, %r1355, %r1357;
mov.b32 %f1108, %r1380;
add.ftz.f32 %f77, %f1107, %f1108;
setp.ne.s32 %p303, %r2, 2;
or.b32 %r1384, %r4, 2;
setp.ge.u32 %p304, %r1384, %r49;
or.pred %p305, %p303, %p304;
@%p305 bra $L__BB129_76;

add.ftz.f32 %f1109, %f5, %f76;
add.ftz.f32 %f1110, %f4, %f77;
mul.ftz.f32 %f1111, %f1109, %f1110;
st.global.f32 [%rd35+8], %f1111;

$L__BB129_76:
ld.shared.f32 %f1112, [%r30+384];
add.ftz.f32 %f1113, %f1112, %f27;
ld.shared.f32 %f1114, [%r32+384];
add.ftz.f32 %f1115, %f1114, %f31;
ld.shared.f32 %f1116, [%r30+896];
add.ftz.f32 %f1117, %f1116, %f1113;
ld.shared.f32 %f1118, [%r32+896];
add.ftz.f32 %f1119, %f1118, %f1115;
ld.shared.f32 %f1120, [%r30+1408];
add.ftz.f32 %f1121, %f1120, %f1117;
ld.shared.f32 %f1122, [%r32+1408];
add.ftz.f32 %f1123, %f1122, %f1119;
mov.b32 %r1385, %f1121;
shfl.sync.bfly.b32 %r1389|%p306, %r1385, %r1356, %r1355, %r1357;
mov.b32 %f1124, %r1389;
add.ftz.f32 %f1125, %f1121, %f1124;
mov.b32 %r1390, %f1125;
shfl.sync.bfly.b32 %r1392|%p307, %r1390, %r1360, %r1355, %r1357;
mov.b32 %f1126, %r1392;
add.ftz.f32 %f1127, %f1125, %f1126;
mov.b32 %r1393, %f1127;
shfl.sync.bfly.b32 %r1395|%p308, %r1393, %r1363, %r1355, %r1357;
mov.b32 %f1128, %r1395;
add.ftz.f32 %f1129, %f1127, %f1128;
mov.b32 %r1396, %f1129;
shfl.sync.bfly.b32 %r1398|%p309, %r1396, %r1366, %r1355, %r1357;
mov.b32 %f1130, %r1398;
add.ftz.f32 %f1131, %f1129, %f1130;
mov.b32 %r1399, %f1131;
shfl.sync.bfly.b32 %r1401|%p310, %r1399, %r1369, %r1355, %r1357;
mov.b32 %f1132, %r1401;
add.ftz.f32 %f78, %f1131, %f1132;
mov.b32 %r1402, %f1123;
shfl.sync.bfly.b32 %r1403|%p311, %r1402, %r1356, %r1355, %r1357;
mov.b32 %f1133, %r1403;
add.ftz.f32 %f1134, %f1123, %f1133;
mov.b32 %r1404, %f1134;
shfl.sync.bfly.b32 %r1405|%p312, %r1404, %r1360, %r1355, %r1357;
mov.b32 %f1135, %r1405;
add.ftz.f32 %f1136, %f1134, %f1135;
mov.b32 %r1406, %f1136;
shfl.sync.bfly.b32 %r1407|%p313, %r1406, %r1363, %r1355, %r1357;
mov.b32 %f1137, %r1407;
add.ftz.f32 %f1138, %f1136, %f1137;
mov.b32 %r1408, %f1138;
shfl.sync.bfly.b32 %r1409|%p314, %r1408, %r1366, %r1355, %r1357;
mov.b32 %f1139, %r1409;
add.ftz.f32 %f1140, %f1138, %f1139;
mov.b32 %r1410, %f1140;
shfl.sync.bfly.b32 %r1411|%p315, %r1410, %r1369, %r1355, %r1357;
mov.b32 %f1141, %r1411;
add.ftz.f32 %f79, %f1140, %f1141;
setp.ne.s32 %p316, %r2, 3;
or.b32 %r1415, %r4, 3;
setp.ge.u32 %p317, %r1415, %r49;
or.pred %p318, %p316, %p317;
@%p318 bra $L__BB129_78;

add.ftz.f32 %f1142, %f5, %f78;
add.ftz.f32 %f1143, %f4, %f79;
mul.ftz.f32 %f1144, %f1142, %f1143;
st.global.f32 [%rd35+12], %f1144;
bra.uni $L__BB129_78;

$L__BB129_62:
ld.shared.f32 %f849, [%r30];
add.ftz.f32 %f850, %f849, %f30;
ld.shared.f32 %f851, [%r32];
add.ftz.f32 %f852, %f851, %f34;
ld.shared.f32 %f853, [%r30+512];
add.ftz.f32 %f854, %f853, %f850;
ld.shared.f32 %f855, [%r32+512];
add.ftz.f32 %f856, %f855, %f852;
ld.shared.f32 %f857, [%r30+1024];
add.ftz.f32 %f858, %f857, %f854;
ld.shared.f32 %f859, [%r32+1024];
add.ftz.f32 %f860, %f859, %f856;
mov.b32 %r1170, %f858;
mov.u32 %r1171, 31;
mov.u32 %r1172, 16;
mov.u32 %r1173, -1;
shfl.sync.bfly.b32 %r1174|%p215, %r1170, %r1172, %r1171, %r1173;
mov.b32 %f861, %r1174;
add.ftz.f32 %f862, %f858, %f861;
mov.b32 %r1175, %f862;
mov.u32 %r1176, 8;
shfl.sync.bfly.b32 %r1177|%p216, %r1175, %r1176, %r1171, %r1173;
mov.b32 %f863, %r1177;
add.ftz.f32 %f864, %f862, %f863;
mov.b32 %r1178, %f864;
mov.u32 %r1179, 4;
shfl.sync.bfly.b32 %r1180|%p217, %r1178, %r1179, %r1171, %r1173;
mov.b32 %f865, %r1180;
add.ftz.f32 %f866, %f864, %f865;
mov.b32 %r1181, %f866;
mov.u32 %r1182, 2;
shfl.sync.bfly.b32 %r1183|%p218, %r1181, %r1182, %r1171, %r1173;
mov.b32 %f867, %r1183;
add.ftz.f32 %f868, %f866, %f867;
mov.b32 %r1184, %f868;
mov.u32 %r1185, 1;
shfl.sync.bfly.b32 %r1186|%p219, %r1184, %r1185, %r1171, %r1173;
mov.b32 %f869, %r1186;
add.ftz.f32 %f64, %f868, %f869;
mov.b32 %r1187, %f860;
shfl.sync.bfly.b32 %r1188|%p220, %r1187, %r1172, %r1171, %r1173;
mov.b32 %f870, %r1188;
add.ftz.f32 %f871, %f860, %f870;
mov.b32 %r1189, %f871;
shfl.sync.bfly.b32 %r1190|%p221, %r1189, %r1176, %r1171, %r1173;
mov.b32 %f872, %r1190;
add.ftz.f32 %f873, %f871, %f872;
mov.b32 %r1191, %f873;
shfl.sync.bfly.b32 %r1192|%p222, %r1191, %r1179, %r1171, %r1173;
mov.b32 %f874, %r1192;
add.ftz.f32 %f875, %f873, %f874;
mov.b32 %r1193, %f875;
shfl.sync.bfly.b32 %r1194|%p223, %r1193, %r1182, %r1171, %r1173;
mov.b32 %f876, %r1194;
add.ftz.f32 %f877, %f875, %f876;
mov.b32 %r1195, %f877;
shfl.sync.bfly.b32 %r1196|%p224, %r1195, %r1185, %r1171, %r1173;
mov.b32 %f878, %r1196;
add.ftz.f32 %f65, %f877, %f878;
setp.ne.s32 %p225, %r2, 0;
setp.ge.u32 %p226, %r4, %r49;
or.pred %p227, %p225, %p226;
@%p227 bra $L__BB129_64;

add.ftz.f32 %f879, %f5, %f64;
add.ftz.f32 %f880, %f4, %f65;
mul.ftz.f32 %f881, %f880, 0f3F000000;
mul.ftz.f32 %f882, %f880, 0f3F4C422A;
mul.ftz.f32 %f883, %f880, 0f3D372713;
fma.rn.ftz.f32 %f884, %f880, %f883, 0f3F800000;
mul.ftz.f32 %f885, %f882, %f884;
tanh.approx.f32 %f886, %f885;
add.ftz.f32 %f887, %f886, 0f3F800000;
mul.ftz.f32 %f888, %f881, %f887;
mul.ftz.f32 %f889, %f879, %f888;
st.global.f32 [%rd35], %f889;

$L__BB129_64:
ld.shared.f32 %f890, [%r30+128];
add.ftz.f32 %f891, %f890, %f29;
ld.shared.f32 %f892, [%r32+128];
add.ftz.f32 %f893, %f892, %f33;
ld.shared.f32 %f894, [%r30+640];
add.ftz.f32 %f895, %f894, %f891;
ld.shared.f32 %f896, [%r32+640];
add.ftz.f32 %f897, %f896, %f893;
ld.shared.f32 %f898, [%r30+1152];
add.ftz.f32 %f899, %f898, %f895;
ld.shared.f32 %f900, [%r32+1152];
add.ftz.f32 %f901, %f900, %f897;
mov.b32 %r1200, %f899;
shfl.sync.bfly.b32 %r1204|%p228, %r1200, %r1172, %r1171, %r1173;
mov.b32 %f902, %r1204;
add.ftz.f32 %f903, %f899, %f902;
mov.b32 %r1205, %f903;
shfl.sync.bfly.b32 %r1207|%p229, %r1205, %r1176, %r1171, %r1173;
mov.b32 %f904, %r1207;
add.ftz.f32 %f905, %f903, %f904;
mov.b32 %r1208, %f905;
shfl.sync.bfly.b32 %r1210|%p230, %r1208, %r1179, %r1171, %r1173;
mov.b32 %f906, %r1210;
add.ftz.f32 %f907, %f905, %f906;
mov.b32 %r1211, %f907;
shfl.sync.bfly.b32 %r1213|%p231, %r1211, %r1182, %r1171, %r1173;
mov.b32 %f908, %r1213;
add.ftz.f32 %f909, %f907, %f908;
mov.b32 %r1214, %f909;
shfl.sync.bfly.b32 %r1216|%p232, %r1214, %r1185, %r1171, %r1173;
mov.b32 %f910, %r1216;
add.ftz.f32 %f66, %f909, %f910;
mov.b32 %r1217, %f901;
shfl.sync.bfly.b32 %r1218|%p233, %r1217, %r1172, %r1171, %r1173;
mov.b32 %f911, %r1218;
add.ftz.f32 %f912, %f901, %f911;
mov.b32 %r1219, %f912;
shfl.sync.bfly.b32 %r1220|%p234, %r1219, %r1176, %r1171, %r1173;
mov.b32 %f913, %r1220;
add.ftz.f32 %f914, %f912, %f913;
mov.b32 %r1221, %f914;
shfl.sync.bfly.b32 %r1222|%p235, %r1221, %r1179, %r1171, %r1173;
mov.b32 %f915, %r1222;
add.ftz.f32 %f916, %f914, %f915;
mov.b32 %r1223, %f916;
shfl.sync.bfly.b32 %r1224|%p236, %r1223, %r1182, %r1171, %r1173;
mov.b32 %f917, %r1224;
add.ftz.f32 %f918, %f916, %f917;
mov.b32 %r1225, %f918;
shfl.sync.bfly.b32 %r1226|%p237, %r1225, %r1185, %r1171, %r1173;
mov.b32 %f919, %r1226;
add.ftz.f32 %f67, %f918, %f919;
setp.ne.s32 %p238, %r2, 1;
or.b32 %r1230, %r4, 1;
setp.ge.u32 %p239, %r1230, %r49;
or.pred %p240, %p238, %p239;
@%p240 bra $L__BB129_66;

add.ftz.f32 %f920, %f5, %f66;
add.ftz.f32 %f921, %f4, %f67;
mul.ftz.f32 %f922, %f921, 0f3F000000;
mul.ftz.f32 %f923, %f921, 0f3F4C422A;
mul.ftz.f32 %f924, %f921, 0f3D372713;
fma.rn.ftz.f32 %f925, %f921, %f924, 0f3F800000;
mul.ftz.f32 %f926, %f923, %f925;
tanh.approx.f32 %f927, %f926;
add.ftz.f32 %f928, %f927, 0f3F800000;
mul.ftz.f32 %f929, %f922, %f928;
mul.ftz.f32 %f930, %f920, %f929;
st.global.f32 [%rd35+4], %f930;

$L__BB129_66:
ld.shared.f32 %f931, [%r30+256];
add.ftz.f32 %f932, %f931, %f28;
ld.shared.f32 %f933, [%r32+256];
add.ftz.f32 %f934, %f933, %f32;
ld.shared.f32 %f935, [%r30+768];
add.ftz.f32 %f936, %f935, %f932;
ld.shared.f32 %f937, [%r32+768];
add.ftz.f32 %f938, %f937, %f934;
ld.shared.f32 %f939, [%r30+1280];
add.ftz.f32 %f940, %f939, %f936;
ld.shared.f32 %f941, [%r32+1280];
add.ftz.f32 %f942, %f941, %f938;
mov.b32 %r1231, %f940;
mov.u32 %r1232, 31;
mov.u32 %r1233, 16;
mov.u32 %r1234, -1;
shfl.sync.bfly.b32 %r1235|%p241, %r1231, %r1233, %r1232, %r1234;
mov.b32 %f943, %r1235;
add.ftz.f32 %f944, %f940, %f943;
mov.b32 %r1236, %f944;
mov.u32 %r1237, 8;
shfl.sync.bfly.b32 %r1238|%p242, %r1236, %r1237, %r1232, %r1234;
mov.b32 %f945, %r1238;
add.ftz.f32 %f946, %f944, %f945;
mov.b32 %r1239, %f946;
mov.u32 %r1240, 4;
shfl.sync.bfly.b32 %r1241|%p243, %r1239, %r1240, %r1232, %r1234;
mov.b32 %f947, %r1241;
add.ftz.f32 %f948, %f946, %f947;
mov.b32 %r1242, %f948;
mov.u32 %r1243, 2;
shfl.sync.bfly.b32 %r1244|%p244, %r1242, %r1243, %r1232, %r1234;
mov.b32 %f949, %r1244;
add.ftz.f32 %f950, %f948, %f949;
mov.b32 %r1245, %f950;
mov.u32 %r1246, 1;
shfl.sync.bfly.b32 %r1247|%p245, %r1245, %r1246, %r1232, %r1234;
mov.b32 %f951, %r1247;
add.ftz.f32 %f68, %f950, %f951;
mov.b32 %r1248, %f942;
shfl.sync.bfly.b32 %r1249|%p246, %r1248, %r1233, %r1232, %r1234;
mov.b32 %f952, %r1249;
add.ftz.f32 %f953, %f942, %f952;
mov.b32 %r1250, %f953;
shfl.sync.bfly.b32 %r1251|%p247, %r1250, %r1237, %r1232, %r1234;
mov.b32 %f954, %r1251;
add.ftz.f32 %f955, %f953, %f954;
mov.b32 %r1252, %f955;
shfl.sync.bfly.b32 %r1253|%p248, %r1252, %r1240, %r1232, %r1234;
mov.b32 %f956, %r1253;
add.ftz.f32 %f957, %f955, %f956;
mov.b32 %r1254, %f957;
shfl.sync.bfly.b32 %r1255|%p249, %r1254, %r1243, %r1232, %r1234;
mov.b32 %f958, %r1255;
add.ftz.f32 %f959, %f957, %f958;
mov.b32 %r1256, %f959;
shfl.sync.bfly.b32 %r1257|%p250, %r1256, %r1246, %r1232, %r1234;
mov.b32 %f960, %r1257;
add.ftz.f32 %f69, %f959, %f960;
setp.ne.s32 %p251, %r2, 2;
or.b32 %r1261, %r4, 2;
setp.ge.u32 %p252, %r1261, %r49;
or.pred %p253, %p251, %p252;
@%p253 bra $L__BB129_68;

add.ftz.f32 %f961, %f5, %f68;
add.ftz.f32 %f962, %f4, %f69;
mul.ftz.f32 %f963, %f962, 0f3F000000;
mul.ftz.f32 %f964, %f962, 0f3F4C422A;
mul.ftz.f32 %f965, %f962, 0f3D372713;
fma.rn.ftz.f32 %f966, %f962, %f965, 0f3F800000;
mul.ftz.f32 %f967, %f964, %f966;
tanh.approx.f32 %f968, %f967;
add.ftz.f32 %f969, %f968, 0f3F800000;
mul.ftz.f32 %f970, %f963, %f969;
mul.ftz.f32 %f971, %f961, %f970;
st.global.f32 [%rd35+8], %f971;

$L__BB129_68:
ld.shared.f32 %f972, [%r30+384];
add.ftz.f32 %f973, %f972, %f27;
ld.shared.f32 %f974, [%r32+384];
add.ftz.f32 %f975, %f974, %f31;
ld.shared.f32 %f976, [%r30+896];
add.ftz.f32 %f977, %f976, %f973;
ld.shared.f32 %f978, [%r32+896];
add.ftz.f32 %f979, %f978, %f975;
ld.shared.f32 %f980, [%r30+1408];
add.ftz.f32 %f981, %f980, %f977;
ld.shared.f32 %f982, [%r32+1408];
add.ftz.f32 %f983, %f982, %f979;
mov.b32 %r1262, %f981;
shfl.sync.bfly.b32 %r1266|%p254, %r1262, %r1233, %r1232, %r1234;
mov.b32 %f984, %r1266;
add.ftz.f32 %f985, %f981, %f984;
mov.b32 %r1267, %f985;
shfl.sync.bfly.b32 %r1269|%p255, %r1267, %r1237, %r1232, %r1234;
mov.b32 %f986, %r1269;
add.ftz.f32 %f987, %f985, %f986;
mov.b32 %r1270, %f987;
shfl.sync.bfly.b32 %r1272|%p256, %r1270, %r1240, %r1232, %r1234;
mov.b32 %f988, %r1272;
add.ftz.f32 %f989, %f987, %f988;
mov.b32 %r1273, %f989;
shfl.sync.bfly.b32 %r1275|%p257, %r1273, %r1243, %r1232, %r1234;
mov.b32 %f990, %r1275;
add.ftz.f32 %f991, %f989, %f990;
mov.b32 %r1276, %f991;
shfl.sync.bfly.b32 %r1278|%p258, %r1276, %r1246, %r1232, %r1234;
mov.b32 %f992, %r1278;
add.ftz.f32 %f70, %f991, %f992;
mov.b32 %r1279, %f983;
shfl.sync.bfly.b32 %r1280|%p259, %r1279, %r1233, %r1232, %r1234;
mov.b32 %f993, %r1280;
add.ftz.f32 %f994, %f983, %f993;
mov.b32 %r1281, %f994;
shfl.sync.bfly.b32 %r1282|%p260, %r1281, %r1237, %r1232, %r1234;
mov.b32 %f995, %r1282;
add.ftz.f32 %f996, %f994, %f995;
mov.b32 %r1283, %f996;
shfl.sync.bfly.b32 %r1284|%p261, %r1283, %r1240, %r1232, %r1234;
mov.b32 %f997, %r1284;
add.ftz.f32 %f998, %f996, %f997;
mov.b32 %r1285, %f998;
shfl.sync.bfly.b32 %r1286|%p262, %r1285, %r1243, %r1232, %r1234;
mov.b32 %f999, %r1286;
add.ftz.f32 %f1000, %f998, %f999;
mov.b32 %r1287, %f1000;
shfl.sync.bfly.b32 %r1288|%p263, %r1287, %r1246, %r1232, %r1234;
mov.b32 %f1001, %r1288;
add.ftz.f32 %f71, %f1000, %f1001;
setp.ne.s32 %p264, %r2, 3;
or.b32 %r1292, %r4, 3;
setp.ge.u32 %p265, %r1292, %r49;
or.pred %p266, %p264, %p265;
@%p266 bra $L__BB129_78;

add.ftz.f32 %f1002, %f5, %f70;
add.ftz.f32 %f1003, %f4, %f71;
mul.ftz.f32 %f1004, %f1003, 0f3F000000;
mul.ftz.f32 %f1005, %f1003, 0f3F4C422A;
mul.ftz.f32 %f1006, %f1003, 0f3D372713;
fma.rn.ftz.f32 %f1007, %f1003, %f1006, 0f3F800000;
mul.ftz.f32 %f1008, %f1005, %f1007;
tanh.approx.f32 %f1009, %f1008;
add.ftz.f32 %f1010, %f1009, 0f3F800000;
mul.ftz.f32 %f1011, %f1004, %f1010;
mul.ftz.f32 %f1012, %f1002, %f1011;
st.global.f32 [%rd35+12], %f1012;
bra.uni $L__BB129_78;

$L__BB129_54:
ld.shared.f32 %f525, [%r30];
add.ftz.f32 %f526, %f525, %f30;
ld.shared.f32 %f527, [%r32];
add.ftz.f32 %f528, %f527, %f34;
ld.shared.f32 %f529, [%r30+512];
add.ftz.f32 %f530, %f529, %f526;
ld.shared.f32 %f531, [%r32+512];
add.ftz.f32 %f532, %f531, %f528;
ld.shared.f32 %f533, [%r30+1024];
add.ftz.f32 %f534, %f533, %f530;
ld.shared.f32 %f535, [%r32+1024];
add.ftz.f32 %f536, %f535, %f532;
mov.b32 %r924, %f534;
mov.u32 %r925, 31;
mov.u32 %r926, 16;
mov.u32 %r927, -1;
shfl.sync.bfly.b32 %r928|%p111, %r924, %r926, %r925, %r927;
mov.b32 %f537, %r928;
add.ftz.f32 %f538, %f534, %f537;
mov.b32 %r929, %f538;
mov.u32 %r930, 8;
shfl.sync.bfly.b32 %r931|%p112, %r929, %r930, %r925, %r927;
mov.b32 %f539, %r931;
add.ftz.f32 %f540, %f538, %f539;
mov.b32 %r932, %f540;
mov.u32 %r933, 4;
shfl.sync.bfly.b32 %r934|%p113, %r932, %r933, %r925, %r927;
mov.b32 %f541, %r934;
add.ftz.f32 %f542, %f540, %f541;
mov.b32 %r935, %f542;
mov.u32 %r936, 2;
shfl.sync.bfly.b32 %r937|%p114, %r935, %r936, %r925, %r927;
mov.b32 %f543, %r937;
add.ftz.f32 %f544, %f542, %f543;
mov.b32 %r938, %f544;
mov.u32 %r939, 1;
shfl.sync.bfly.b32 %r940|%p115, %r938, %r939, %r925, %r927;
mov.b32 %f545, %r940;
add.ftz.f32 %f48, %f544, %f545;
mov.b32 %r941, %f536;
shfl.sync.bfly.b32 %r942|%p116, %r941, %r926, %r925, %r927;
mov.b32 %f546, %r942;
add.ftz.f32 %f547, %f536, %f546;
mov.b32 %r943, %f547;
shfl.sync.bfly.b32 %r944|%p117, %r943, %r930, %r925, %r927;
mov.b32 %f548, %r944;
add.ftz.f32 %f549, %f547, %f548;
mov.b32 %r945, %f549;
shfl.sync.bfly.b32 %r946|%p118, %r945, %r933, %r925, %r927;
mov.b32 %f550, %r946;
add.ftz.f32 %f551, %f549, %f550;
mov.b32 %r947, %f551;
shfl.sync.bfly.b32 %r948|%p119, %r947, %r936, %r925, %r927;
mov.b32 %f552, %r948;
add.ftz.f32 %f553, %f551, %f552;
mov.b32 %r949, %f553;
shfl.sync.bfly.b32 %r950|%p120, %r949, %r939, %r925, %r927;
mov.b32 %f554, %r950;
add.ftz.f32 %f49, %f553, %f554;
setp.ne.s32 %p121, %r2, 0;
setp.ge.u32 %p122, %r4, %r49;
or.pred %p123, %p121, %p122;
@%p123 bra $L__BB129_56;

add.ftz.f32 %f555, %f5, %f48;
add.ftz.f32 %f556, %f4, %f49;
mov.f32 %f557, 0f40E00000;
min.ftz.f32 %f558, %f556, %f557;
min.ftz.f32 %f559, %f555, %f557;
mov.f32 %f560, 0fC0E00000;
max.ftz.f32 %f561, %f559, %f560;
mul.ftz.f32 %f562, %f558, 0f3FD9DB23;
mul.ftz.f32 %f563, %f562, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f564, %f563;
add.ftz.f32 %f565, %f564, 0f3F800000;
div.approx.ftz.f32 %f566, %f558, %f565;
add.ftz.f32 %f567, %f561, 0f3F800000;
mul.ftz.f32 %f568, %f567, %f566;
st.global.f32 [%rd35], %f568;

$L__BB129_56:
ld.shared.f32 %f569, [%r30+128];
add.ftz.f32 %f570, %f569, %f29;
ld.shared.f32 %f571, [%r32+128];
add.ftz.f32 %f572, %f571, %f33;
ld.shared.f32 %f573, [%r30+640];
add.ftz.f32 %f574, %f573, %f570;
ld.shared.f32 %f575, [%r32+640];
add.ftz.f32 %f576, %f575, %f572;
ld.shared.f32 %f577, [%r30+1152];
add.ftz.f32 %f578, %f577, %f574;
ld.shared.f32 %f579, [%r32+1152];
add.ftz.f32 %f580, %f579, %f576;
mov.b32 %r954, %f578;
shfl.sync.bfly.b32 %r958|%p124, %r954, %r926, %r925, %r927;
mov.b32 %f581, %r958;
add.ftz.f32 %f582, %f578, %f581;
mov.b32 %r959, %f582;
shfl.sync.bfly.b32 %r961|%p125, %r959, %r930, %r925, %r927;
mov.b32 %f583, %r961;
add.ftz.f32 %f584, %f582, %f583;
mov.b32 %r962, %f584;
shfl.sync.bfly.b32 %r964|%p126, %r962, %r933, %r925, %r927;
mov.b32 %f585, %r964;
add.ftz.f32 %f586, %f584, %f585;
mov.b32 %r965, %f586;
shfl.sync.bfly.b32 %r967|%p127, %r965, %r936, %r925, %r927;
mov.b32 %f587, %r967;
add.ftz.f32 %f588, %f586, %f587;
mov.b32 %r968, %f588;
shfl.sync.bfly.b32 %r970|%p128, %r968, %r939, %r925, %r927;
mov.b32 %f589, %r970;
add.ftz.f32 %f50, %f588, %f589;
mov.b32 %r971, %f580;
shfl.sync.bfly.b32 %r972|%p129, %r971, %r926, %r925, %r927;
mov.b32 %f590, %r972;
add.ftz.f32 %f591, %f580, %f590;
mov.b32 %r973, %f591;
shfl.sync.bfly.b32 %r974|%p130, %r973, %r930, %r925, %r927;
mov.b32 %f592, %r974;
add.ftz.f32 %f593, %f591, %f592;
mov.b32 %r975, %f593;
shfl.sync.bfly.b32 %r976|%p131, %r975, %r933, %r925, %r927;
mov.b32 %f594, %r976;
add.ftz.f32 %f595, %f593, %f594;
mov.b32 %r977, %f595;
shfl.sync.bfly.b32 %r978|%p132, %r977, %r936, %r925, %r927;
mov.b32 %f596, %r978;
add.ftz.f32 %f597, %f595, %f596;
mov.b32 %r979, %f597;
shfl.sync.bfly.b32 %r980|%p133, %r979, %r939, %r925, %r927;
mov.b32 %f598, %r980;
add.ftz.f32 %f51, %f597, %f598;
setp.ne.s32 %p134, %r2, 1;
or.b32 %r984, %r4, 1;
setp.ge.u32 %p135, %r984, %r49;
or.pred %p136, %p134, %p135;
@%p136 bra $L__BB129_58;

add.ftz.f32 %f599, %f5, %f50;
add.ftz.f32 %f600, %f4, %f51;
mov.f32 %f601, 0f40E00000;
min.ftz.f32 %f602, %f600, %f601;
min.ftz.f32 %f603, %f599, %f601;
mov.f32 %f604, 0fC0E00000;
max.ftz.f32 %f605, %f603, %f604;
mul.ftz.f32 %f606, %f602, 0f3FD9DB23;
mul.ftz.f32 %f607, %f606, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f608, %f607;
add.ftz.f32 %f609, %f608, 0f3F800000;
div.approx.ftz.f32 %f610, %f602, %f609;
add.ftz.f32 %f611, %f605, 0f3F800000;
mul.ftz.f32 %f612, %f611, %f610;
st.global.f32 [%rd35+4], %f612;

$L__BB129_58:
ld.shared.f32 %f613, [%r30+256];
add.ftz.f32 %f614, %f613, %f28;
ld.shared.f32 %f615, [%r32+256];
add.ftz.f32 %f616, %f615, %f32;
ld.shared.f32 %f617, [%r30+768];
add.ftz.f32 %f618, %f617, %f614;
ld.shared.f32 %f619, [%r32+768];
add.ftz.f32 %f620, %f619, %f616;
ld.shared.f32 %f621, [%r30+1280];
add.ftz.f32 %f622, %f621, %f618;
ld.shared.f32 %f623, [%r32+1280];
add.ftz.f32 %f624, %f623, %f620;
mov.b32 %r985, %f622;
mov.u32 %r986, 31;
mov.u32 %r987, 16;
mov.u32 %r988, -1;
shfl.sync.bfly.b32 %r989|%p137, %r985, %r987, %r986, %r988;
mov.b32 %f625, %r989;
add.ftz.f32 %f626, %f622, %f625;
mov.b32 %r990, %f626;
mov.u32 %r991, 8;
shfl.sync.bfly.b32 %r992|%p138, %r990, %r991, %r986, %r988;
mov.b32 %f627, %r992;
add.ftz.f32 %f628, %f626, %f627;
mov.b32 %r993, %f628;
mov.u32 %r994, 4;
shfl.sync.bfly.b32 %r995|%p139, %r993, %r994, %r986, %r988;
mov.b32 %f629, %r995;
add.ftz.f32 %f630, %f628, %f629;
mov.b32 %r996, %f630;
mov.u32 %r997, 2;
shfl.sync.bfly.b32 %r998|%p140, %r996, %r997, %r986, %r988;
mov.b32 %f631, %r998;
add.ftz.f32 %f632, %f630, %f631;
mov.b32 %r999, %f632;
mov.u32 %r1000, 1;
shfl.sync.bfly.b32 %r1001|%p141, %r999, %r1000, %r986, %r988;
mov.b32 %f633, %r1001;
add.ftz.f32 %f52, %f632, %f633;
mov.b32 %r1002, %f624;
shfl.sync.bfly.b32 %r1003|%p142, %r1002, %r987, %r986, %r988;
mov.b32 %f634, %r1003;
add.ftz.f32 %f635, %f624, %f634;
mov.b32 %r1004, %f635;
shfl.sync.bfly.b32 %r1005|%p143, %r1004, %r991, %r986, %r988;
mov.b32 %f636, %r1005;
add.ftz.f32 %f637, %f635, %f636;
mov.b32 %r1006, %f637;
shfl.sync.bfly.b32 %r1007|%p144, %r1006, %r994, %r986, %r988;
mov.b32 %f638, %r1007;
add.ftz.f32 %f639, %f637, %f638;
mov.b32 %r1008, %f639;
shfl.sync.bfly.b32 %r1009|%p145, %r1008, %r997, %r986, %r988;
mov.b32 %f640, %r1009;
add.ftz.f32 %f641, %f639, %f640;
mov.b32 %r1010, %f641;
shfl.sync.bfly.b32 %r1011|%p146, %r1010, %r1000, %r986, %r988;
mov.b32 %f642, %r1011;
add.ftz.f32 %f53, %f641, %f642;
setp.ne.s32 %p147, %r2, 2;
or.b32 %r1015, %r4, 2;
setp.ge.u32 %p148, %r1015, %r49;
or.pred %p149, %p147, %p148;
@%p149 bra $L__BB129_60;

add.ftz.f32 %f643, %f5, %f52;
add.ftz.f32 %f644, %f4, %f53;
mov.f32 %f645, 0f40E00000;
min.ftz.f32 %f646, %f644, %f645;
min.ftz.f32 %f647, %f643, %f645;
mov.f32 %f648, 0fC0E00000;
max.ftz.f32 %f649, %f647, %f648;
mul.ftz.f32 %f650, %f646, 0f3FD9DB23;
mul.ftz.f32 %f651, %f650, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f652, %f651;
add.ftz.f32 %f653, %f652, 0f3F800000;
div.approx.ftz.f32 %f654, %f646, %f653;
add.ftz.f32 %f655, %f649, 0f3F800000;
mul.ftz.f32 %f656, %f655, %f654;
st.global.f32 [%rd35+8], %f656;

$L__BB129_60:
ld.shared.f32 %f657, [%r30+384];
add.ftz.f32 %f658, %f657, %f27;
ld.shared.f32 %f659, [%r32+384];
add.ftz.f32 %f660, %f659, %f31;
ld.shared.f32 %f661, [%r30+896];
add.ftz.f32 %f662, %f661, %f658;
ld.shared.f32 %f663, [%r32+896];
add.ftz.f32 %f664, %f663, %f660;
ld.shared.f32 %f665, [%r30+1408];
add.ftz.f32 %f666, %f665, %f662;
ld.shared.f32 %f667, [%r32+1408];
add.ftz.f32 %f668, %f667, %f664;
mov.b32 %r1016, %f666;
shfl.sync.bfly.b32 %r1020|%p150, %r1016, %r987, %r986, %r988;
mov.b32 %f669, %r1020;
add.ftz.f32 %f670, %f666, %f669;
mov.b32 %r1021, %f670;
shfl.sync.bfly.b32 %r1023|%p151, %r1021, %r991, %r986, %r988;
mov.b32 %f671, %r1023;
add.ftz.f32 %f672, %f670, %f671;
mov.b32 %r1024, %f672;
shfl.sync.bfly.b32 %r1026|%p152, %r1024, %r994, %r986, %r988;
mov.b32 %f673, %r1026;
add.ftz.f32 %f674, %f672, %f673;
mov.b32 %r1027, %f674;
shfl.sync.bfly.b32 %r1029|%p153, %r1027, %r997, %r986, %r988;
mov.b32 %f675, %r1029;
add.ftz.f32 %f676, %f674, %f675;
mov.b32 %r1030, %f676;
shfl.sync.bfly.b32 %r1032|%p154, %r1030, %r1000, %r986, %r988;
mov.b32 %f677, %r1032;
add.ftz.f32 %f54, %f676, %f677;
mov.b32 %r1033, %f668;
shfl.sync.bfly.b32 %r1034|%p155, %r1033, %r987, %r986, %r988;
mov.b32 %f678, %r1034;
add.ftz.f32 %f679, %f668, %f678;
mov.b32 %r1035, %f679;
shfl.sync.bfly.b32 %r1036|%p156, %r1035, %r991, %r986, %r988;
mov.b32 %f680, %r1036;
add.ftz.f32 %f681, %f679, %f680;
mov.b32 %r1037, %f681;
shfl.sync.bfly.b32 %r1038|%p157, %r1037, %r994, %r986, %r988;
mov.b32 %f682, %r1038;
add.ftz.f32 %f683, %f681, %f682;
mov.b32 %r1039, %f683;
shfl.sync.bfly.b32 %r1040|%p158, %r1039, %r997, %r986, %r988;
mov.b32 %f684, %r1040;
add.ftz.f32 %f685, %f683, %f684;
mov.b32 %r1041, %f685;
shfl.sync.bfly.b32 %r1042|%p159, %r1041, %r1000, %r986, %r988;
mov.b32 %f686, %r1042;
add.ftz.f32 %f55, %f685, %f686;
setp.ne.s32 %p160, %r2, 3;
or.b32 %r1046, %r4, 3;
setp.ge.u32 %p161, %r1046, %r49;
or.pred %p162, %p160, %p161;
@%p162 bra $L__BB129_78;

add.ftz.f32 %f687, %f5, %f54;
add.ftz.f32 %f688, %f4, %f55;
mov.f32 %f689, 0f40E00000;
min.ftz.f32 %f690, %f688, %f689;
min.ftz.f32 %f691, %f687, %f689;
mov.f32 %f692, 0fC0E00000;
max.ftz.f32 %f693, %f691, %f692;
mul.ftz.f32 %f694, %f690, 0f3FD9DB23;
mul.ftz.f32 %f695, %f694, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f696, %f695;
add.ftz.f32 %f697, %f696, 0f3F800000;
div.approx.ftz.f32 %f698, %f690, %f697;
add.ftz.f32 %f699, %f693, 0f3F800000;
mul.ftz.f32 %f700, %f699, %f698;
st.global.f32 [%rd35+12], %f700;
bra.uni $L__BB129_78;

}
.entry _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<38>;
.reg .b16 %rs<17>;
.reg .f32 %f<193>;
.reg .b32 %r<456>;
.reg .b64 %rd<50>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

ld.param.u64 %rd13, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd14, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
mov.u32 %r1, %tid.y;
shl.b32 %r39, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r39, %r2;
mov.u32 %r40, %ctaid.x;
shl.b32 %r4, %r40, 2;
shr.u32 %r5, %r38, 8;
mov.u32 %r453, %ctaid.y;
setp.eq.s64 %p1, %rd15, 0;
@%p1 bra $L__BB130_2;

cvta.to.global.u64 %rd17, %rd15;
mul.wide.u32 %rd18, %r453, 4;
add.s64 %rd19, %rd17, %rd18;
ld.global.nc.u32 %r452, [%rd19];
mul.hi.u32 %r41, %r453, %r21;
add.s32 %r42, %r41, %r453;
shr.u32 %r43, %r42, %r22;
mul.lo.s32 %r44, %r43, %r23;
sub.s32 %r453, %r453, %r44;
bra.uni $L__BB130_3;

$L__BB130_2:
mul.hi.u32 %r45, %r453, %r26;
add.s32 %r46, %r45, %r453;
shr.u32 %r452, %r46, %r27;

$L__BB130_3:
mov.u32 %r12, %ctaid.z;
shr.s32 %r47, %r3, 31;
shr.u32 %r48, %r47, 28;
add.s32 %r49, %r3, %r48;
shr.s32 %r455, %r49, 4;
setp.ge.s32 %p2, %r455, %r5;
mov.f32 %f189, 0f00000000;
mov.f32 %f190, %f189;
mov.f32 %f191, %f189;
mov.f32 %f192, %f189;
@%p2 bra $L__BB130_6;

and.b32 %r53, %r49, -16;
sub.s32 %r54, %r3, %r53;
shr.s32 %r55, %r54, 31;
shr.u32 %r56, %r55, 29;
add.s32 %r57, %r54, %r56;
shr.s32 %r58, %r57, 3;
shl.b32 %r59, %r58, 2;
and.b32 %r60, %r57, -8;
sub.s32 %r61, %r54, %r60;
shr.s32 %r62, %r61, 31;
shr.u32 %r63, %r62, 30;
add.s32 %r64, %r61, %r63;
shr.s32 %r65, %r64, 2;
add.s32 %r66, %r60, %r65;
cvt.s64.s32 %rd20, %r66;
cvt.s64.s32 %rd1, %r61;
cvt.s64.s32 %rd2, %r59;
shl.b32 %r454, %r455, 3;
mad.lo.s32 %r67, %r4, %r24, %r455;
mul.hi.u32 %r68, %r12, %r32;
add.s32 %r69, %r68, %r12;
shr.u32 %r70, %r69, %r33;
mad.lo.s32 %r71, %r70, %r35, %r67;
mad.lo.s32 %r72, %r452, %r29, %r71;
mul.wide.s32 %rd3, %r72, 84;
mul.wide.s32 %rd21, %r54, 4;
add.s64 %rd22, %rd3, %rd21;
add.s64 %rd4, %rd22, 16;
add.s64 %rd23, %rd3, %rd20;
add.s64 %rd5, %rd23, 6;
mul.wide.s32 %rd6, %r24, 84;
cvta.to.global.u64 %rd7, %rd14;
cvta.to.global.u64 %rd49, %rd13;
mul.lo.s32 %r73, %r453, %r30;
cvt.u64.u32 %rd24, %r73;
mul.lo.s32 %r74, %r12, %r36;
cvt.u64.u32 %rd25, %r74;
add.s64 %rd9, %rd24, %rd25;
mov.f32 %f189, 0f00000000;
shl.b64 %rd32, %rd1, 2;

$L__BB130_5:
cvt.u64.u32 %rd26, %r454;
add.s64 %rd27, %rd9, %rd26;
add.s64 %rd28, %rd49, %rd4;
ld.global.nc.u32 %r227, [%rd28];
add.s64 %rd29, %rd27, %rd2;
mul.lo.s64 %rd30, %rd29, 36;
add.s64 %rd31, %rd7, %rd30;
add.s64 %rd33, %rd31, %rd32;
ld.global.nc.u32 %r81, [%rd33+4];
ld.global.nc.u32 %r75, [%rd31];

{.reg .f16 low,high;
mov.b32 {low,high},%r75;
cvt.f32.f16 %f25, low;}


ld.global.nc.u32 %r89, [%rd33+40];
ld.global.nc.u32 %r76, [%rd31+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r76;
cvt.f32.f16 %f26, low;}


ld.global.nc.u32 %r97, [%rd33+76];
ld.global.nc.u32 %r77, [%rd31+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r77;
cvt.f32.f16 %f27, low;}


ld.global.nc.u32 %r105, [%rd33+112];
ld.global.nc.u32 %r78, [%rd31+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r78;
cvt.f32.f16 %f28, low;}


add.s64 %rd34, %rd49, %rd5;
add.s64 %rd35, %rd34, -6;
ld.global.nc.u8 %rs1, [%rd34+-6];
cvt.u32.u16 %r228, %rs1;
and.b32 %r229, %r228, 240;
and.b32 %r80, %r227, 50529027;
mov.u32 %r224, 0;

dp4a.s32.s32 %r79, %r80, %r81, %r224;

and.b32 %r230, %r228, 15;
mul.lo.s32 %r231, %r79, %r230;
cvt.rn.f32.s32 %f49, %r231;
fma.rn.ftz.f32 %f50, %f25, %f49, 0f00000000;
shr.u32 %r232, %r229, 4;
shl.b32 %r233, %r232, 8;
or.b32 %r234, %r233, %r232;
shl.b32 %r235, %r234, 16;
or.b32 %r84, %r235, %r234;

dp4a.s32.s32 %r83, %r84, %r81, %r224;

cvt.rn.f32.s32 %f51, %r83;
fma.rn.ftz.f32 %f52, %f25, %f51, 0f00000000;
ld.global.nc.u8 %rs2, [%rd34+-4];
cvt.u32.u16 %r236, %rs2;
and.b32 %r237, %r236, 240;
shr.u32 %r238, %r227, 2;
and.b32 %r88, %r238, 50529027;

dp4a.s32.s32 %r87, %r88, %r89, %r224;

and.b32 %r239, %r236, 15;
mul.lo.s32 %r240, %r87, %r239;
cvt.rn.f32.s32 %f53, %r240;
fma.rn.ftz.f32 %f54, %f26, %f53, %f50;
shr.u32 %r241, %r237, 4;
shl.b32 %r242, %r241, 8;
or.b32 %r243, %r242, %r241;
shl.b32 %r244, %r243, 16;
or.b32 %r92, %r244, %r243;

dp4a.s32.s32 %r91, %r92, %r89, %r224;

cvt.rn.f32.s32 %f55, %r91;
fma.rn.ftz.f32 %f56, %f26, %f55, %f52;
ld.global.nc.u8 %rs3, [%rd34+-2];
cvt.u32.u16 %r245, %rs3;
and.b32 %r246, %r245, 240;
shr.u32 %r247, %r227, 4;
and.b32 %r96, %r247, 50529027;

dp4a.s32.s32 %r95, %r96, %r97, %r224;

and.b32 %r248, %r245, 15;
mul.lo.s32 %r249, %r95, %r248;
cvt.rn.f32.s32 %f57, %r249;
fma.rn.ftz.f32 %f58, %f27, %f57, %f54;
shr.u32 %r250, %r246, 4;
shl.b32 %r251, %r250, 8;
or.b32 %r252, %r251, %r250;
shl.b32 %r253, %r252, 16;
or.b32 %r100, %r253, %r252;

dp4a.s32.s32 %r99, %r100, %r97, %r224;

cvt.rn.f32.s32 %f59, %r99;
fma.rn.ftz.f32 %f60, %f27, %f59, %f56;
ld.global.nc.u8 %rs4, [%rd34];
cvt.u32.u16 %r254, %rs4;
and.b32 %r255, %r254, 240;
shr.u32 %r256, %r227, 6;
and.b32 %r104, %r256, 50529027;

dp4a.s32.s32 %r103, %r104, %r105, %r224;

and.b32 %r257, %r254, 15;
mul.lo.s32 %r258, %r103, %r257;
cvt.rn.f32.s32 %f61, %r258;
fma.rn.ftz.f32 %f62, %f28, %f61, %f58;
shr.u32 %r259, %r255, 4;
shl.b32 %r260, %r259, 8;
or.b32 %r261, %r260, %r259;
shl.b32 %r262, %r261, 16;
or.b32 %r108, %r262, %r261;

dp4a.s32.s32 %r107, %r108, %r105, %r224;

cvt.rn.f32.s32 %f63, %r107;
fma.rn.ftz.f32 %f64, %f28, %f63, %f60;
add.s64 %rd36, %rd49, %rd3;
add.s64 %rd37, %rd36, 80;
ld.global.nc.u32 %r111, [%rd36+80];

{.reg .f16 low,high;
mov.b32 {low,high},%r111;
cvt.f32.f16 %f29, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r111;
cvt.f32.f16 %f30, high;}


mul.ftz.f32 %f65, %f62, %f29;
mul.ftz.f32 %f66, %f64, %f30;
sub.ftz.f32 %f67, %f65, %f66;
add.ftz.f32 %f192, %f67, %f192;
add.s64 %rd38, %rd28, %rd6;
ld.global.nc.u32 %r263, [%rd38];

{.reg .f16 low,high;
mov.b32 {low,high},%r75;
cvt.f32.f16 %f31, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r76;
cvt.f32.f16 %f32, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r77;
cvt.f32.f16 %f33, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r78;
cvt.f32.f16 %f34, low;}


add.s64 %rd39, %rd35, %rd6;
ld.global.nc.u8 %rs5, [%rd39];
cvt.u32.u16 %r264, %rs5;
and.b32 %r265, %r264, 240;
and.b32 %r118, %r263, 50529027;

dp4a.s32.s32 %r117, %r118, %r81, %r224;

and.b32 %r266, %r264, 15;
mul.lo.s32 %r267, %r117, %r266;
cvt.rn.f32.s32 %f68, %r267;
fma.rn.ftz.f32 %f69, %f31, %f68, 0f00000000;
shr.u32 %r268, %r265, 4;
shl.b32 %r269, %r268, 8;
or.b32 %r270, %r269, %r268;
shl.b32 %r271, %r270, 16;
or.b32 %r122, %r271, %r270;

dp4a.s32.s32 %r121, %r122, %r81, %r224;

cvt.rn.f32.s32 %f70, %r121;
fma.rn.ftz.f32 %f71, %f31, %f70, 0f00000000;
ld.global.nc.u8 %rs6, [%rd39+2];
cvt.u32.u16 %r272, %rs6;
and.b32 %r273, %r272, 240;
shr.u32 %r274, %r263, 2;
and.b32 %r126, %r274, 50529027;

dp4a.s32.s32 %r125, %r126, %r89, %r224;

and.b32 %r275, %r272, 15;
mul.lo.s32 %r276, %r125, %r275;
cvt.rn.f32.s32 %f72, %r276;
fma.rn.ftz.f32 %f73, %f32, %f72, %f69;
shr.u32 %r277, %r273, 4;
shl.b32 %r278, %r277, 8;
or.b32 %r279, %r278, %r277;
shl.b32 %r280, %r279, 16;
or.b32 %r130, %r280, %r279;

dp4a.s32.s32 %r129, %r130, %r89, %r224;

cvt.rn.f32.s32 %f74, %r129;
fma.rn.ftz.f32 %f75, %f32, %f74, %f71;
ld.global.nc.u8 %rs7, [%rd39+4];
cvt.u32.u16 %r281, %rs7;
and.b32 %r282, %r281, 240;
shr.u32 %r283, %r263, 4;
and.b32 %r134, %r283, 50529027;

dp4a.s32.s32 %r133, %r134, %r97, %r224;

and.b32 %r284, %r281, 15;
mul.lo.s32 %r285, %r133, %r284;
cvt.rn.f32.s32 %f76, %r285;
fma.rn.ftz.f32 %f77, %f33, %f76, %f73;
shr.u32 %r286, %r282, 4;
shl.b32 %r287, %r286, 8;
or.b32 %r288, %r287, %r286;
shl.b32 %r289, %r288, 16;
or.b32 %r138, %r289, %r288;

dp4a.s32.s32 %r137, %r138, %r97, %r224;

cvt.rn.f32.s32 %f78, %r137;
fma.rn.ftz.f32 %f79, %f33, %f78, %f75;
ld.global.nc.u8 %rs8, [%rd39+6];
cvt.u32.u16 %r290, %rs8;
and.b32 %r291, %r290, 240;
shr.u32 %r292, %r263, 6;
and.b32 %r142, %r292, 50529027;

dp4a.s32.s32 %r141, %r142, %r105, %r224;

and.b32 %r293, %r290, 15;
mul.lo.s32 %r294, %r141, %r293;
cvt.rn.f32.s32 %f80, %r294;
fma.rn.ftz.f32 %f81, %f34, %f80, %f77;
shr.u32 %r295, %r291, 4;
shl.b32 %r296, %r295, 8;
or.b32 %r297, %r296, %r295;
shl.b32 %r298, %r297, 16;
or.b32 %r146, %r298, %r297;

dp4a.s32.s32 %r145, %r146, %r105, %r224;

cvt.rn.f32.s32 %f82, %r145;
fma.rn.ftz.f32 %f83, %f34, %f82, %f79;
add.s64 %rd40, %rd37, %rd6;
ld.global.nc.u32 %r149, [%rd40];

{.reg .f16 low,high;
mov.b32 {low,high},%r149;
cvt.f32.f16 %f35, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r149;
cvt.f32.f16 %f36, high;}


mul.ftz.f32 %f84, %f81, %f35;
mul.ftz.f32 %f85, %f83, %f36;
sub.ftz.f32 %f86, %f84, %f85;
add.ftz.f32 %f191, %f86, %f191;
add.s64 %rd41, %rd38, %rd6;
ld.global.nc.u32 %r299, [%rd41];

{.reg .f16 low,high;
mov.b32 {low,high},%r75;
cvt.f32.f16 %f37, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r76;
cvt.f32.f16 %f38, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r77;
cvt.f32.f16 %f39, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r78;
cvt.f32.f16 %f40, low;}


add.s64 %rd42, %rd39, %rd6;
ld.global.nc.u8 %rs9, [%rd42];
cvt.u32.u16 %r300, %rs9;
and.b32 %r301, %r300, 240;
and.b32 %r156, %r299, 50529027;

dp4a.s32.s32 %r155, %r156, %r81, %r224;

and.b32 %r302, %r300, 15;
mul.lo.s32 %r303, %r155, %r302;
cvt.rn.f32.s32 %f87, %r303;
fma.rn.ftz.f32 %f88, %f37, %f87, 0f00000000;
shr.u32 %r304, %r301, 4;
shl.b32 %r305, %r304, 8;
or.b32 %r306, %r305, %r304;
shl.b32 %r307, %r306, 16;
or.b32 %r160, %r307, %r306;

dp4a.s32.s32 %r159, %r160, %r81, %r224;

cvt.rn.f32.s32 %f89, %r159;
fma.rn.ftz.f32 %f90, %f37, %f89, 0f00000000;
ld.global.nc.u8 %rs10, [%rd42+2];
cvt.u32.u16 %r308, %rs10;
and.b32 %r309, %r308, 240;
shr.u32 %r310, %r299, 2;
and.b32 %r164, %r310, 50529027;

dp4a.s32.s32 %r163, %r164, %r89, %r224;

and.b32 %r311, %r308, 15;
mul.lo.s32 %r312, %r163, %r311;
cvt.rn.f32.s32 %f91, %r312;
fma.rn.ftz.f32 %f92, %f38, %f91, %f88;
shr.u32 %r313, %r309, 4;
shl.b32 %r314, %r313, 8;
or.b32 %r315, %r314, %r313;
shl.b32 %r316, %r315, 16;
or.b32 %r168, %r316, %r315;

dp4a.s32.s32 %r167, %r168, %r89, %r224;

cvt.rn.f32.s32 %f93, %r167;
fma.rn.ftz.f32 %f94, %f38, %f93, %f90;
ld.global.nc.u8 %rs11, [%rd42+4];
cvt.u32.u16 %r317, %rs11;
and.b32 %r318, %r317, 240;
shr.u32 %r319, %r299, 4;
and.b32 %r172, %r319, 50529027;

dp4a.s32.s32 %r171, %r172, %r97, %r224;

and.b32 %r320, %r317, 15;
mul.lo.s32 %r321, %r171, %r320;
cvt.rn.f32.s32 %f95, %r321;
fma.rn.ftz.f32 %f96, %f39, %f95, %f92;
shr.u32 %r322, %r318, 4;
shl.b32 %r323, %r322, 8;
or.b32 %r324, %r323, %r322;
shl.b32 %r325, %r324, 16;
or.b32 %r176, %r325, %r324;

dp4a.s32.s32 %r175, %r176, %r97, %r224;

cvt.rn.f32.s32 %f97, %r175;
fma.rn.ftz.f32 %f98, %f39, %f97, %f94;
ld.global.nc.u8 %rs12, [%rd42+6];
cvt.u32.u16 %r326, %rs12;
and.b32 %r327, %r326, 240;
shr.u32 %r328, %r299, 6;
and.b32 %r180, %r328, 50529027;

dp4a.s32.s32 %r179, %r180, %r105, %r224;

and.b32 %r329, %r326, 15;
mul.lo.s32 %r330, %r179, %r329;
cvt.rn.f32.s32 %f99, %r330;
fma.rn.ftz.f32 %f100, %f40, %f99, %f96;
shr.u32 %r331, %r327, 4;
shl.b32 %r332, %r331, 8;
or.b32 %r333, %r332, %r331;
shl.b32 %r334, %r333, 16;
or.b32 %r184, %r334, %r333;

dp4a.s32.s32 %r183, %r184, %r105, %r224;

cvt.rn.f32.s32 %f101, %r183;
fma.rn.ftz.f32 %f102, %f40, %f101, %f98;
add.s64 %rd43, %rd40, %rd6;
ld.global.nc.u32 %r187, [%rd43];

{.reg .f16 low,high;
mov.b32 {low,high},%r187;
cvt.f32.f16 %f41, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r187;
cvt.f32.f16 %f42, high;}


mul.ftz.f32 %f103, %f100, %f41;
mul.ftz.f32 %f104, %f102, %f42;
sub.ftz.f32 %f105, %f103, %f104;
add.ftz.f32 %f190, %f105, %f190;
add.s64 %rd44, %rd41, %rd6;
ld.global.nc.u32 %r335, [%rd44];

{.reg .f16 low,high;
mov.b32 {low,high},%r75;
cvt.f32.f16 %f43, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r76;
cvt.f32.f16 %f44, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r77;
cvt.f32.f16 %f45, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r78;
cvt.f32.f16 %f46, low;}


add.s64 %rd45, %rd42, %rd6;
ld.global.nc.u8 %rs13, [%rd45];
cvt.u32.u16 %r336, %rs13;
and.b32 %r337, %r336, 240;
and.b32 %r194, %r335, 50529027;

dp4a.s32.s32 %r193, %r194, %r81, %r224;

and.b32 %r338, %r336, 15;
mul.lo.s32 %r339, %r193, %r338;
cvt.rn.f32.s32 %f106, %r339;
fma.rn.ftz.f32 %f107, %f43, %f106, 0f00000000;
shr.u32 %r340, %r337, 4;
shl.b32 %r341, %r340, 8;
or.b32 %r342, %r341, %r340;
shl.b32 %r343, %r342, 16;
or.b32 %r198, %r343, %r342;

dp4a.s32.s32 %r197, %r198, %r81, %r224;

cvt.rn.f32.s32 %f108, %r197;
fma.rn.ftz.f32 %f109, %f43, %f108, 0f00000000;
ld.global.nc.u8 %rs14, [%rd45+2];
cvt.u32.u16 %r344, %rs14;
and.b32 %r345, %r344, 240;
shr.u32 %r346, %r335, 2;
and.b32 %r202, %r346, 50529027;

dp4a.s32.s32 %r201, %r202, %r89, %r224;

and.b32 %r347, %r344, 15;
mul.lo.s32 %r348, %r201, %r347;
cvt.rn.f32.s32 %f110, %r348;
fma.rn.ftz.f32 %f111, %f44, %f110, %f107;
shr.u32 %r349, %r345, 4;
shl.b32 %r350, %r349, 8;
or.b32 %r351, %r350, %r349;
shl.b32 %r352, %r351, 16;
or.b32 %r206, %r352, %r351;

dp4a.s32.s32 %r205, %r206, %r89, %r224;

cvt.rn.f32.s32 %f112, %r205;
fma.rn.ftz.f32 %f113, %f44, %f112, %f109;
ld.global.nc.u8 %rs15, [%rd45+4];
cvt.u32.u16 %r353, %rs15;
and.b32 %r354, %r353, 240;
shr.u32 %r355, %r335, 4;
and.b32 %r210, %r355, 50529027;

dp4a.s32.s32 %r209, %r210, %r97, %r224;

and.b32 %r356, %r353, 15;
mul.lo.s32 %r357, %r209, %r356;
cvt.rn.f32.s32 %f114, %r357;
fma.rn.ftz.f32 %f115, %f45, %f114, %f111;
shr.u32 %r358, %r354, 4;
shl.b32 %r359, %r358, 8;
or.b32 %r360, %r359, %r358;
shl.b32 %r361, %r360, 16;
or.b32 %r214, %r361, %r360;

dp4a.s32.s32 %r213, %r214, %r97, %r224;

cvt.rn.f32.s32 %f116, %r213;
fma.rn.ftz.f32 %f117, %f45, %f116, %f113;
ld.global.nc.u8 %rs16, [%rd45+6];
cvt.u32.u16 %r362, %rs16;
and.b32 %r363, %r362, 240;
shr.u32 %r364, %r335, 6;
and.b32 %r218, %r364, 50529027;

dp4a.s32.s32 %r217, %r218, %r105, %r224;

and.b32 %r365, %r362, 15;
mul.lo.s32 %r366, %r217, %r365;
cvt.rn.f32.s32 %f118, %r366;
fma.rn.ftz.f32 %f119, %f46, %f118, %f115;
shr.u32 %r367, %r363, 4;
shl.b32 %r368, %r367, 8;
or.b32 %r369, %r368, %r367;
shl.b32 %r370, %r369, 16;
or.b32 %r222, %r370, %r369;

dp4a.s32.s32 %r221, %r222, %r105, %r224;

cvt.rn.f32.s32 %f120, %r221;
fma.rn.ftz.f32 %f121, %f46, %f120, %f117;
add.s64 %rd46, %rd43, %rd6;
ld.global.nc.u32 %r225, [%rd46];

{.reg .f16 low,high;
mov.b32 {low,high},%r225;
cvt.f32.f16 %f47, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r225;
cvt.f32.f16 %f48, high;}


mul.ftz.f32 %f122, %f119, %f47;
mul.ftz.f32 %f123, %f121, %f48;
sub.ftz.f32 %f124, %f122, %f123;
add.ftz.f32 %f189, %f124, %f189;
add.s32 %r454, %r454, 64;
add.s64 %rd49, %rd49, 672;
add.s32 %r455, %r455, 8;
setp.lt.s32 %p3, %r455, %r5;
@%p3 bra $L__BB130_5;

$L__BB130_6:
setp.eq.s32 %p4, %r1, 0;
@%p4 bra $L__BB130_8;

shl.b32 %r371, %r1, 9;
mov.u32 %r372, _ZZ13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r373, %r372, %r371;
shl.b32 %r374, %r2, 2;
add.s32 %r375, %r373, %r374;
st.shared.f32 [%r375+-512], %f192;
st.shared.f32 [%r375+-384], %f191;
st.shared.f32 [%r375+-256], %f190;
st.shared.f32 [%r375+-128], %f189;

$L__BB130_8:
setp.ne.s32 %p5, %r1, 0;
bar.sync 0;
@%p5 bra $L__BB130_17;

mov.u32 %r376, %ctaid.y;
mad.lo.s32 %r377, %r376, %r31, %r4;
mad.lo.s32 %r378, %r12, %r37, %r377;
shl.b32 %r379, %r2, 2;
mov.u32 %r380, _ZZ13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r19, %r380, %r379;
mov.u32 %r381, 2;
ld.shared.f32 %f125, [%r19];
add.ftz.f32 %f126, %f125, %f192;
ld.shared.f32 %f127, [%r19+512];
add.ftz.f32 %f128, %f127, %f126;
ld.shared.f32 %f129, [%r19+1024];
add.ftz.f32 %f130, %f129, %f128;
mov.b32 %r382, %f130;
mov.u32 %r383, 31;
mov.u32 %r384, 16;
mov.u32 %r385, -1;
shfl.sync.bfly.b32 %r386|%p6, %r382, %r384, %r383, %r385;
mov.b32 %f131, %r386;
add.ftz.f32 %f132, %f130, %f131;
mov.b32 %r387, %f132;
mov.u32 %r388, 8;
shfl.sync.bfly.b32 %r389|%p7, %r387, %r388, %r383, %r385;
mov.b32 %f133, %r389;
add.ftz.f32 %f134, %f132, %f133;
mov.b32 %r390, %f134;
mov.u32 %r391, 4;
shfl.sync.bfly.b32 %r392|%p8, %r390, %r391, %r383, %r385;
mov.b32 %f135, %r392;
add.ftz.f32 %f136, %f134, %f135;
mov.b32 %r393, %f136;
shfl.sync.bfly.b32 %r394|%p9, %r393, %r381, %r383, %r385;
mov.b32 %f137, %r394;
add.ftz.f32 %f138, %f136, %f137;
mov.b32 %r395, %f138;
mov.u32 %r396, 1;
shfl.sync.bfly.b32 %r397|%p10, %r395, %r396, %r383, %r385;
mov.b32 %f139, %r397;
add.ftz.f32 %f13, %f138, %f139;
setp.ge.u32 %p11, %r4, %r25;
setp.ne.s32 %p12, %r2, 0;
cvta.to.global.u64 %rd47, %rd16;
mul.wide.u32 %rd48, %r378, 4;
add.s64 %rd12, %rd47, %rd48;
or.pred %p13, %p12, %p11;
@%p13 bra $L__BB130_11;

st.global.f32 [%rd12], %f13;

$L__BB130_11:
ld.shared.f32 %f140, [%r19+128];
add.ftz.f32 %f141, %f140, %f191;
ld.shared.f32 %f142, [%r19+640];
add.ftz.f32 %f143, %f142, %f141;
ld.shared.f32 %f144, [%r19+1152];
add.ftz.f32 %f145, %f144, %f143;
mov.b32 %r398, %f145;
shfl.sync.bfly.b32 %r402|%p14, %r398, %r384, %r383, %r385;
mov.b32 %f146, %r402;
add.ftz.f32 %f147, %f145, %f146;
mov.b32 %r403, %f147;
shfl.sync.bfly.b32 %r405|%p15, %r403, %r388, %r383, %r385;
mov.b32 %f148, %r405;
add.ftz.f32 %f149, %f147, %f148;
mov.b32 %r406, %f149;
shfl.sync.bfly.b32 %r408|%p16, %r406, %r391, %r383, %r385;
mov.b32 %f150, %r408;
add.ftz.f32 %f151, %f149, %f150;
mov.b32 %r409, %f151;
shfl.sync.bfly.b32 %r411|%p17, %r409, %r381, %r383, %r385;
mov.b32 %f152, %r411;
add.ftz.f32 %f153, %f151, %f152;
mov.b32 %r412, %f153;
shfl.sync.bfly.b32 %r414|%p18, %r412, %r396, %r383, %r385;
mov.b32 %f154, %r414;
add.ftz.f32 %f14, %f153, %f154;
add.s32 %r415, %r4, 1;
setp.ge.u32 %p19, %r415, %r25;
setp.ne.s32 %p20, %r2, 1;
or.pred %p21, %p20, %p19;
@%p21 bra $L__BB130_13;

st.global.f32 [%rd12+4], %f14;

$L__BB130_13:
ld.shared.f32 %f155, [%r19+256];
add.ftz.f32 %f156, %f155, %f190;
ld.shared.f32 %f157, [%r19+768];
add.ftz.f32 %f158, %f157, %f156;
ld.shared.f32 %f159, [%r19+1280];
add.ftz.f32 %f160, %f159, %f158;
mov.b32 %r416, %f160;
mov.u32 %r417, 31;
mov.u32 %r418, 16;
mov.u32 %r419, -1;
shfl.sync.bfly.b32 %r420|%p22, %r416, %r418, %r417, %r419;
mov.b32 %f161, %r420;
add.ftz.f32 %f162, %f160, %f161;
mov.b32 %r421, %f162;
mov.u32 %r422, 8;
shfl.sync.bfly.b32 %r423|%p23, %r421, %r422, %r417, %r419;
mov.b32 %f163, %r423;
add.ftz.f32 %f164, %f162, %f163;
mov.b32 %r424, %f164;
mov.u32 %r425, 4;
shfl.sync.bfly.b32 %r426|%p24, %r424, %r425, %r417, %r419;
mov.b32 %f165, %r426;
add.ftz.f32 %f166, %f164, %f165;
mov.b32 %r427, %f166;
mov.u32 %r428, 2;
shfl.sync.bfly.b32 %r429|%p25, %r427, %r428, %r417, %r419;
mov.b32 %f167, %r429;
add.ftz.f32 %f168, %f166, %f167;
mov.b32 %r430, %f168;
mov.u32 %r431, 1;
shfl.sync.bfly.b32 %r432|%p26, %r430, %r431, %r417, %r419;
mov.b32 %f169, %r432;
add.ftz.f32 %f15, %f168, %f169;
add.s32 %r433, %r4, 2;
setp.ge.u32 %p27, %r433, %r25;
setp.ne.s32 %p28, %r2, 2;
or.pred %p29, %p28, %p27;
@%p29 bra $L__BB130_15;

st.global.f32 [%rd12+8], %f15;

$L__BB130_15:
ld.shared.f32 %f170, [%r19+384];
add.ftz.f32 %f171, %f170, %f189;
ld.shared.f32 %f172, [%r19+896];
add.ftz.f32 %f173, %f172, %f171;
ld.shared.f32 %f174, [%r19+1408];
add.ftz.f32 %f175, %f174, %f173;
mov.b32 %r434, %f175;
shfl.sync.bfly.b32 %r438|%p30, %r434, %r418, %r417, %r419;
mov.b32 %f176, %r438;
add.ftz.f32 %f177, %f175, %f176;
mov.b32 %r439, %f177;
shfl.sync.bfly.b32 %r441|%p31, %r439, %r422, %r417, %r419;
mov.b32 %f178, %r441;
add.ftz.f32 %f179, %f177, %f178;
mov.b32 %r442, %f179;
shfl.sync.bfly.b32 %r444|%p32, %r442, %r425, %r417, %r419;
mov.b32 %f180, %r444;
add.ftz.f32 %f181, %f179, %f180;
mov.b32 %r445, %f181;
shfl.sync.bfly.b32 %r447|%p33, %r445, %r428, %r417, %r419;
mov.b32 %f182, %r447;
add.ftz.f32 %f183, %f181, %f182;
mov.b32 %r448, %f183;
shfl.sync.bfly.b32 %r450|%p34, %r448, %r431, %r417, %r419;
mov.b32 %f184, %r450;
add.ftz.f32 %f16, %f183, %f184;
add.s32 %r451, %r4, 3;
setp.ge.u32 %p35, %r451, %r25;
setp.ne.s32 %p36, %r2, 3;
or.pred %p37, %p36, %p35;
@%p37 bra $L__BB130_17;

st.global.f32 [%rd12+12], %f16;

$L__BB130_17:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<34>;
.reg .b16 %rs<5>;
.reg .f32 %f<162>;
.reg .b32 %r<305>;
.reg .b64 %rd<72>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[384];

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate[384];

ld.param.u64 %rd10, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd11, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd12, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r45, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r46, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r47, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
ld.param.f32 %f31, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+16];
ld.param.u64 %rd14, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+8];
ld.param.u64 %rd13, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3];
mov.u32 %r1, %tid.y;
shl.b32 %r48, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r48, %r2;
mov.u32 %r4, %ctaid.x;
mov.u32 %r5, %ctaid.y;
setp.eq.s64 %p2, %rd12, 0;
@%p2 bra $L__BB131_2;

cvta.to.global.u64 %rd19, %rd12;
mul.wide.u32 %rd20, %r5, 4;
add.s64 %rd21, %rd19, %rd20;
ld.global.nc.u32 %r301, [%rd21];
mul.hi.u32 %r49, %r5, %r32;
add.s32 %r50, %r49, %r5;
shr.u32 %r51, %r50, %r33;
mul.lo.s32 %r52, %r51, %r34;
sub.s32 %r302, %r5, %r52;
bra.uni $L__BB131_3;

$L__BB131_2:
mul.hi.u32 %r53, %r5, %r36;
add.s32 %r54, %r53, %r5;
shr.u32 %r301, %r54, %r37;
mov.u32 %r302, %r5;

$L__BB131_3:
or.b32 %r55, %r2, %r1;
setp.ne.s32 %p3, %r55, 0;
mov.f32 %f155, 0f00000000;
mov.f32 %f151, %f155;
mov.f32 %f152, %f155;
@%p3 bra $L__BB131_8;

setp.eq.s64 %p4, %rd13, 0;
mov.f32 %f151, 0f00000000;
mov.f32 %f152, %f151;
@%p4 bra $L__BB131_6;

cvta.to.global.u64 %rd22, %rd13;
mov.u32 %r56, %ctaid.z;
mul.lo.s32 %r57, %r56, %r47;
cvt.u64.u32 %rd23, %r57;
selp.b32 %r58, %r5, %r301, %p2;
mul.lo.s32 %r59, %r58, %r41;
cvt.u64.u32 %rd24, %r59;
cvt.u64.u32 %rd25, %r2;
cvt.s64.s32 %rd26, %r4;
add.s64 %rd27, %rd25, %rd26;
add.s64 %rd28, %rd27, %rd23;
add.s64 %rd29, %rd28, %rd24;
shl.b64 %rd30, %rd29, 2;
add.s64 %rd31, %rd22, %rd30;
ld.global.f32 %f152, [%rd31];

$L__BB131_6:
setp.eq.s64 %p6, %rd15, 0;
setp.eq.s64 %p7, %rd14, 0;
or.pred %p8, %p6, %p7;
@%p8 bra $L__BB131_8;

cvta.to.global.u64 %rd32, %rd15;
mov.u32 %r61, %ctaid.z;
mul.lo.s32 %r62, %r61, %r47;
cvt.u64.u32 %rd33, %r62;
selp.b32 %r63, %r5, %r301, %p2;
mul.lo.s32 %r64, %r63, %r41;
cvt.u64.u32 %rd34, %r64;
cvt.u64.u32 %rd35, %r2;
cvt.s64.s32 %rd36, %r4;
add.s64 %rd37, %rd35, %rd36;
add.s64 %rd38, %rd37, %rd33;
add.s64 %rd39, %rd38, %rd34;
shl.b64 %rd40, %rd39, 2;
add.s64 %rd41, %rd32, %rd40;
ld.global.f32 %f151, [%rd41];

$L__BB131_8:
shr.u32 %r66, %r31, 8;
shr.s32 %r67, %r3, 31;
shr.u32 %r68, %r67, 28;
add.s32 %r69, %r3, %r68;
shr.s32 %r304, %r69, 4;
setp.ge.s32 %p10, %r304, %r66;
mov.f32 %f157, %f155;
@%p10 bra $L__BB131_13;

mov.u32 %r70, %ctaid.z;
and.b32 %r74, %r69, -16;
sub.s32 %r12, %r3, %r74;
shr.s32 %r75, %r12, 31;
shr.u32 %r76, %r75, 29;
add.s32 %r77, %r12, %r76;
shr.s32 %r78, %r77, 3;
shl.b32 %r79, %r78, 2;
and.b32 %r80, %r77, -8;
sub.s32 %r81, %r12, %r80;
shr.s32 %r82, %r81, 31;
shr.u32 %r83, %r82, 30;
add.s32 %r84, %r81, %r83;
shr.s32 %r85, %r84, 2;
add.s32 %r13, %r80, %r85;
cvt.s64.s32 %rd4, %r81;
cvt.s64.s32 %rd5, %r79;
mul.hi.u32 %r86, %r70, %r42;
add.s32 %r87, %r86, %r70;
shr.u32 %r88, %r87, %r43;
mad.lo.s32 %r89, %r88, %r45, %r304;
mad.lo.s32 %r90, %r301, %r39, %r89;
mad.lo.s32 %r91, %r4, %r35, %r90;
mul.wide.s32 %rd71, %r91, 84;
shl.b32 %r303, %r304, 3;
mul.lo.s32 %r92, %r302, %r40;
cvt.u64.u32 %rd42, %r92;
mul.lo.s32 %r93, %r70, %r46;
cvt.u64.u32 %rd43, %r93;
add.s64 %rd7, %rd42, %rd43;
mov.f32 %f157, 0f00000000;
cvta.to.global.u64 %rd44, %rd11;
cvta.to.global.u64 %rd48, %rd10;
shl.b64 %rd54, %rd4, 2;
mov.f32 %f155, %f157;

$L__BB131_10:
cvt.u64.u32 %rd45, %r303;
add.s64 %rd46, %rd7, %rd45;
mul.wide.s32 %rd47, %r12, 4;
add.s64 %rd49, %rd48, %rd47;
add.s64 %rd50, %rd49, %rd71;
ld.global.nc.u32 %r132, [%rd50+16];
add.s64 %rd51, %rd46, %rd5;
mul.lo.s64 %rd52, %rd51, 36;
add.s64 %rd53, %rd44, %rd52;
add.s64 %rd55, %rd53, %rd54;
ld.global.nc.u32 %r17, [%rd55+4];
ld.global.nc.u32 %r18, [%rd53];

{.reg .f16 low,high;
mov.b32 {low,high},%r18;
cvt.f32.f16 %f40, low;}


ld.global.nc.u32 %r19, [%rd55+40];
ld.global.nc.u32 %r20, [%rd53+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r20;
cvt.f32.f16 %f41, low;}


ld.global.nc.u32 %r21, [%rd55+76];
ld.global.nc.u32 %r22, [%rd53+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r22;
cvt.f32.f16 %f42, low;}


ld.global.nc.u32 %r23, [%rd55+112];
ld.global.nc.u32 %r24, [%rd53+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r24;
cvt.f32.f16 %f43, low;}


cvt.s64.s32 %rd56, %r13;
add.s64 %rd57, %rd48, %rd56;
add.s64 %rd58, %rd57, %rd71;
ld.global.nc.u8 %rs1, [%rd58];
cvt.u32.u16 %r133, %rs1;
and.b32 %r134, %r133, 240;
and.b32 %r99, %r132, 50529027;
mov.u32 %r129, 0;

dp4a.s32.s32 %r98, %r99, %r17, %r129;

and.b32 %r135, %r133, 15;
mul.lo.s32 %r136, %r98, %r135;
cvt.rn.f32.s32 %f46, %r136;
fma.rn.ftz.f32 %f47, %f40, %f46, 0f00000000;
shr.u32 %r137, %r134, 4;
shl.b32 %r138, %r137, 8;
or.b32 %r139, %r138, %r137;
shl.b32 %r140, %r139, 16;
or.b32 %r103, %r140, %r139;

dp4a.s32.s32 %r102, %r103, %r17, %r129;

cvt.rn.f32.s32 %f48, %r102;
fma.rn.ftz.f32 %f49, %f40, %f48, 0f00000000;
ld.global.nc.u8 %rs2, [%rd58+2];
cvt.u32.u16 %r141, %rs2;
and.b32 %r142, %r141, 240;
shr.u32 %r143, %r132, 2;
and.b32 %r107, %r143, 50529027;

dp4a.s32.s32 %r106, %r107, %r19, %r129;

and.b32 %r144, %r141, 15;
mul.lo.s32 %r145, %r106, %r144;
cvt.rn.f32.s32 %f50, %r145;
fma.rn.ftz.f32 %f51, %f41, %f50, %f47;
shr.u32 %r146, %r142, 4;
shl.b32 %r147, %r146, 8;
or.b32 %r148, %r147, %r146;
shl.b32 %r149, %r148, 16;
or.b32 %r111, %r149, %r148;

dp4a.s32.s32 %r110, %r111, %r19, %r129;

cvt.rn.f32.s32 %f52, %r110;
fma.rn.ftz.f32 %f53, %f41, %f52, %f49;
ld.global.nc.u8 %rs3, [%rd58+4];
cvt.u32.u16 %r150, %rs3;
and.b32 %r151, %r150, 240;
shr.u32 %r152, %r132, 4;
and.b32 %r115, %r152, 50529027;

dp4a.s32.s32 %r114, %r115, %r21, %r129;

and.b32 %r153, %r150, 15;
mul.lo.s32 %r154, %r114, %r153;
cvt.rn.f32.s32 %f54, %r154;
fma.rn.ftz.f32 %f55, %f42, %f54, %f51;
shr.u32 %r155, %r151, 4;
shl.b32 %r156, %r155, 8;
or.b32 %r157, %r156, %r155;
shl.b32 %r158, %r157, 16;
or.b32 %r119, %r158, %r157;

dp4a.s32.s32 %r118, %r119, %r21, %r129;

cvt.rn.f32.s32 %f56, %r118;
fma.rn.ftz.f32 %f57, %f42, %f56, %f53;
ld.global.nc.u8 %rs4, [%rd58+6];
cvt.u32.u16 %r159, %rs4;
and.b32 %r160, %r159, 240;
shr.u32 %r161, %r132, 6;
and.b32 %r123, %r161, 50529027;

dp4a.s32.s32 %r122, %r123, %r23, %r129;

and.b32 %r162, %r159, 15;
mul.lo.s32 %r163, %r122, %r162;
cvt.rn.f32.s32 %f58, %r163;
fma.rn.ftz.f32 %f59, %f43, %f58, %f55;
shr.u32 %r164, %r160, 4;
shl.b32 %r165, %r164, 8;
or.b32 %r166, %r165, %r164;
shl.b32 %r167, %r166, 16;
or.b32 %r127, %r167, %r166;

dp4a.s32.s32 %r126, %r127, %r23, %r129;

cvt.rn.f32.s32 %f60, %r126;
fma.rn.ftz.f32 %f61, %f43, %f60, %f57;
add.s64 %rd59, %rd48, %rd71;
ld.global.nc.u32 %r130, [%rd59+80];

{.reg .f16 low,high;
mov.b32 {low,high},%r130;
cvt.f32.f16 %f44, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r130;
cvt.f32.f16 %f45, high;}


mul.ftz.f32 %f62, %f59, %f44;
mul.ftz.f32 %f63, %f61, %f45;
sub.ftz.f32 %f64, %f62, %f63;
add.ftz.f32 %f157, %f157, %f64;
setp.eq.s64 %p11, %rd14, 0;
@%p11 bra $L__BB131_12;

cvta.to.global.u64 %rd61, %rd14;
add.s64 %rd62, %rd61, %rd47;
add.s64 %rd63, %rd62, %rd71;
ld.global.u32 %r206, [%rd63+16];

{.reg .f16 low,high;
mov.b32 {low,high},%r18;
cvt.f32.f16 %f65, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r20;
cvt.f32.f16 %f66, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r22;
cvt.f32.f16 %f67, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r24;
cvt.f32.f16 %f68, low;}


add.s64 %rd65, %rd61, %rd56;
add.s64 %rd66, %rd65, %rd71;
ld.global.u8 %r207, [%rd66];
and.b32 %r173, %r206, 50529027;

dp4a.s32.s32 %r172, %r173, %r17, %r129;

and.b32 %r208, %r207, 15;
mul.lo.s32 %r209, %r172, %r208;
cvt.rn.f32.s32 %f71, %r209;
fma.rn.ftz.f32 %f72, %f65, %f71, 0f00000000;
shr.u32 %r210, %r207, 4;
shl.b32 %r211, %r210, 8;
or.b32 %r212, %r211, %r210;
shl.b32 %r213, %r212, 16;
or.b32 %r177, %r213, %r212;

dp4a.s32.s32 %r176, %r177, %r17, %r129;

cvt.rn.f32.s32 %f73, %r176;
fma.rn.ftz.f32 %f74, %f65, %f73, 0f00000000;
ld.global.u8 %r214, [%rd66+2];
shr.u32 %r215, %r206, 2;
and.b32 %r181, %r215, 50529027;

dp4a.s32.s32 %r180, %r181, %r19, %r129;

and.b32 %r216, %r214, 15;
mul.lo.s32 %r217, %r180, %r216;
cvt.rn.f32.s32 %f75, %r217;
fma.rn.ftz.f32 %f76, %f66, %f75, %f72;
shr.u32 %r218, %r214, 4;
shl.b32 %r219, %r218, 8;
or.b32 %r220, %r219, %r218;
shl.b32 %r221, %r220, 16;
or.b32 %r185, %r221, %r220;

dp4a.s32.s32 %r184, %r185, %r19, %r129;

cvt.rn.f32.s32 %f77, %r184;
fma.rn.ftz.f32 %f78, %f66, %f77, %f74;
ld.global.u8 %r222, [%rd66+4];
shr.u32 %r223, %r206, 4;
and.b32 %r189, %r223, 50529027;

dp4a.s32.s32 %r188, %r189, %r21, %r129;

and.b32 %r224, %r222, 15;
mul.lo.s32 %r225, %r188, %r224;
cvt.rn.f32.s32 %f79, %r225;
fma.rn.ftz.f32 %f80, %f67, %f79, %f76;
shr.u32 %r226, %r222, 4;
shl.b32 %r227, %r226, 8;
or.b32 %r228, %r227, %r226;
shl.b32 %r229, %r228, 16;
or.b32 %r193, %r229, %r228;

dp4a.s32.s32 %r192, %r193, %r21, %r129;

cvt.rn.f32.s32 %f81, %r192;
fma.rn.ftz.f32 %f82, %f67, %f81, %f78;
ld.global.u8 %r230, [%rd66+6];
shr.u32 %r231, %r206, 6;
and.b32 %r197, %r231, 50529027;

dp4a.s32.s32 %r196, %r197, %r23, %r129;

and.b32 %r232, %r230, 15;
mul.lo.s32 %r233, %r196, %r232;
cvt.rn.f32.s32 %f83, %r233;
fma.rn.ftz.f32 %f84, %f68, %f83, %f80;
shr.u32 %r234, %r230, 4;
shl.b32 %r235, %r234, 8;
or.b32 %r236, %r235, %r234;
shl.b32 %r237, %r236, 16;
or.b32 %r201, %r237, %r236;

dp4a.s32.s32 %r200, %r201, %r23, %r129;

cvt.rn.f32.s32 %f85, %r200;
fma.rn.ftz.f32 %f86, %f68, %f85, %f82;
add.s64 %rd67, %rd61, %rd71;
ld.global.u32 %r205, [%rd67+80];

{.reg .f16 low,high;
mov.b32 {low,high},%r205;
cvt.f32.f16 %f69, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r205;
cvt.f32.f16 %f70, high;}


mul.ftz.f32 %f87, %f84, %f69;
mul.ftz.f32 %f88, %f86, %f70;
sub.ftz.f32 %f89, %f87, %f88;
add.ftz.f32 %f155, %f155, %f89;

$L__BB131_12:
add.s64 %rd71, %rd71, 672;
add.s32 %r303, %r303, 64;
add.s32 %r304, %r304, 8;
setp.lt.s32 %p12, %r304, %r66;
@%p12 bra $L__BB131_10;

$L__BB131_13:
setp.eq.s32 %p13, %r1, 0;
@%p13 bra $L__BB131_16;

setp.eq.s64 %p14, %rd14, 0;
add.s32 %r27, %r1, -1;
shl.b32 %r242, %r1, 7;
mov.u32 %r243, _ZZ13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r244, %r243, %r242;
shl.b32 %r245, %r2, 2;
add.s32 %r246, %r244, %r245;
st.shared.f32 [%r246+-128], %f157;
@%p14 bra $L__BB131_16;

shl.b32 %r248, %r27, 7;
mov.u32 %r249, _ZZ13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r250, %r249, %r248;
add.s32 %r252, %r250, %r245;
st.shared.f32 [%r252], %f155;

$L__BB131_16:
setp.ne.s32 %p15, %r1, 0;
bar.sync 0;
@%p15 bra $L__BB131_35;

setp.ne.s64 %p16, %rd14, 0;
shl.b32 %r255, %r2, 2;
mov.u32 %r256, _ZZ13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r28, %r256, %r255;
ld.shared.f32 %f90, [%r28];
add.ftz.f32 %f13, %f157, %f90;
@%p16 bra $L__BB131_19;
bra.uni $L__BB131_18;

$L__BB131_19:
mov.u32 %r259, _ZZ13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r260, %r259, %r255;
ld.shared.f32 %f94, [%r260];
add.ftz.f32 %f95, %f155, %f94;
ld.shared.f32 %f96, [%r28+128];
add.ftz.f32 %f97, %f13, %f96;
ld.shared.f32 %f98, [%r260+128];
add.ftz.f32 %f99, %f95, %f98;
ld.shared.f32 %f100, [%r28+256];
add.ftz.f32 %f159, %f97, %f100;
ld.shared.f32 %f101, [%r260+256];
add.ftz.f32 %f155, %f99, %f101;
bra.uni $L__BB131_20;

$L__BB131_18:
ld.shared.f32 %f91, [%r28+128];
add.ftz.f32 %f92, %f13, %f91;
ld.shared.f32 %f93, [%r28+256];
add.ftz.f32 %f159, %f92, %f93;

$L__BB131_20:
mov.b32 %r261, %f159;
mov.u32 %r262, 31;
mov.u32 %r263, 16;
mov.u32 %r264, -1;
shfl.sync.bfly.b32 %r265|%p17, %r261, %r263, %r262, %r264;
mov.b32 %f102, %r265;
add.ftz.f32 %f103, %f159, %f102;
mov.b32 %r266, %f103;
mov.u32 %r267, 8;
shfl.sync.bfly.b32 %r268|%p18, %r266, %r267, %r262, %r264;
mov.b32 %f104, %r268;
add.ftz.f32 %f105, %f103, %f104;
mov.b32 %r269, %f105;
mov.u32 %r270, 4;
shfl.sync.bfly.b32 %r271|%p19, %r269, %r270, %r262, %r264;
mov.b32 %f106, %r271;
add.ftz.f32 %f107, %f105, %f106;
mov.b32 %r272, %f107;
mov.u32 %r273, 2;
shfl.sync.bfly.b32 %r274|%p20, %r272, %r273, %r262, %r264;
mov.b32 %f108, %r274;
add.ftz.f32 %f19, %f107, %f108;
mov.b32 %r275, %f19;
mov.u32 %r276, 1;
shfl.sync.bfly.b32 %r29|%p1, %r275, %r276, %r262, %r264;
setp.eq.s64 %p21, %rd14, 0;
@%p21 bra $L__BB131_22;

mov.b32 %r277, %f155;
shfl.sync.bfly.b32 %r281|%p22, %r277, %r263, %r262, %r264;
mov.b32 %f109, %r281;
add.ftz.f32 %f110, %f155, %f109;
mov.b32 %r282, %f110;
shfl.sync.bfly.b32 %r284|%p23, %r282, %r267, %r262, %r264;
mov.b32 %f111, %r284;
add.ftz.f32 %f112, %f110, %f111;
mov.b32 %r285, %f112;
shfl.sync.bfly.b32 %r287|%p24, %r285, %r270, %r262, %r264;
mov.b32 %f113, %r287;
add.ftz.f32 %f114, %f112, %f113;
mov.b32 %r288, %f114;
shfl.sync.bfly.b32 %r290|%p25, %r288, %r273, %r262, %r264;
mov.b32 %f115, %r290;
add.ftz.f32 %f116, %f114, %f115;
mov.b32 %r291, %f116;
shfl.sync.bfly.b32 %r293|%p26, %r291, %r276, %r262, %r264;
mov.b32 %f117, %r293;
add.ftz.f32 %f155, %f116, %f117;

$L__BB131_22:
setp.ne.s32 %p27, %r2, 0;
mov.b32 %f118, %r29;
add.ftz.f32 %f22, %f19, %f118;
@%p27 bra $L__BB131_35;

add.ftz.f32 %f161, %f152, %f22;
@%p21 bra $L__BB131_34;

add.ftz.f32 %f24, %f151, %f155;
setp.gt.s32 %p29, %r30, 2;
@%p29 bra $L__BB131_28;

setp.eq.s32 %p32, %r30, 1;
@%p32 bra $L__BB131_32;

setp.eq.s32 %p33, %r30, 2;
@%p33 bra $L__BB131_27;
bra.uni $L__BB131_33;

$L__BB131_27:
mul.ftz.f32 %f138, %f24, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f139, %f138;
add.ftz.f32 %f140, %f139, 0f3F800000;
div.approx.ftz.f32 %f141, %f24, %f140;
mul.ftz.f32 %f161, %f161, %f141;
bra.uni $L__BB131_34;

$L__BB131_28:
setp.eq.s32 %p30, %r30, 3;
@%p30 bra $L__BB131_31;

setp.ne.s32 %p31, %r30, 6;
@%p31 bra $L__BB131_33;

min.ftz.f32 %f119, %f24, %f31;
neg.ftz.f32 %f120, %f31;
min.ftz.f32 %f121, %f161, %f31;
max.ftz.f32 %f122, %f121, %f120;
mul.ftz.f32 %f123, %f119, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f124, %f123;
add.ftz.f32 %f125, %f124, 0f3F800000;
div.approx.ftz.f32 %f126, %f119, %f125;
mul.ftz.f32 %f161, %f122, %f126;
bra.uni $L__BB131_34;

$L__BB131_32:
mul.ftz.f32 %f142, %f24, 0f3F000000;
mul.ftz.f32 %f143, %f24, 0f3D372713;
fma.rn.ftz.f32 %f144, %f24, %f143, 0f3F800000;
mul.ftz.f32 %f145, %f24, 0f3F4C422A;
mul.ftz.f32 %f146, %f145, %f144;
tanh.approx.f32 %f147, %f146;
add.ftz.f32 %f148, %f147, 0f3F800000;
mul.ftz.f32 %f149, %f142, %f148;
mul.ftz.f32 %f161, %f161, %f149;
bra.uni $L__BB131_34;

$L__BB131_31:
mov.f32 %f127, 0f40E00000;
min.ftz.f32 %f128, %f24, %f127;
min.ftz.f32 %f129, %f161, %f127;
mov.f32 %f130, 0fC0E00000;
max.ftz.f32 %f131, %f129, %f130;
mul.ftz.f32 %f132, %f128, 0f3FD9DB23;
mul.ftz.f32 %f133, %f132, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f134, %f133;
add.ftz.f32 %f135, %f134, 0f3F800000;
div.approx.ftz.f32 %f136, %f128, %f135;
add.ftz.f32 %f137, %f131, 0f3F800000;
mul.ftz.f32 %f161, %f137, %f136;
bra.uni $L__BB131_34;

$L__BB131_33:
mul.ftz.f32 %f161, %f161, %f24;

$L__BB131_34:
cvta.to.global.u64 %rd68, %rd18;
mov.u32 %r296, %ctaid.z;
mov.u32 %r297, %ctaid.y;
mad.lo.s32 %r299, %r297, %r41, %r4;
mad.lo.s32 %r300, %r296, %r47, %r299;
mul.wide.u32 %rd69, %r300, 4;
add.s64 %rd70, %rd68, %rd69;
st.global.f32 [%rd70], %f161;

$L__BB131_35:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<12>;
.reg .b16 %rs<5>;
.reg .f32 %f<49>;
.reg .b32 %r<180>;
.reg .b64 %rd<45>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[384];

ld.param.u64 %rd14, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r18, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r19, [_Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
mov.u32 %r1, %tid.y;
shl.b32 %r35, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r35, %r2;
mov.u32 %r177, %ctaid.y;
setp.eq.s64 %p2, %rd16, 0;
@%p2 bra $L__BB132_2;

cvta.to.global.u64 %rd18, %rd16;
mul.wide.u32 %rd19, %r177, 4;
add.s64 %rd20, %rd18, %rd19;
ld.global.nc.u32 %r176, [%rd20];
mul.hi.u32 %r36, %r177, %r19;
add.s32 %r37, %r36, %r177;
shr.u32 %r38, %r37, %r20;
mul.lo.s32 %r39, %r38, %r21;
sub.s32 %r177, %r177, %r39;
bra.uni $L__BB132_3;

$L__BB132_2:
mul.hi.u32 %r40, %r177, %r23;
add.s32 %r41, %r40, %r177;
shr.u32 %r176, %r41, %r24;

$L__BB132_3:
shr.s32 %r42, %r3, 31;
shr.u32 %r43, %r42, 28;
add.s32 %r44, %r3, %r43;
shr.s32 %r179, %r44, 4;
shr.u32 %r45, %r18, 8;
setp.ge.s32 %p3, %r179, %r45;
mov.f32 %f48, 0f00000000;
@%p3 bra $L__BB132_6;

mov.u32 %r46, %ctaid.z;
and.b32 %r50, %r44, -16;
sub.s32 %r51, %r3, %r50;
shr.s32 %r52, %r51, 31;
shr.u32 %r53, %r52, 29;
add.s32 %r54, %r51, %r53;
shr.s32 %r55, %r54, 3;
shl.b32 %r56, %r55, 2;
and.b32 %r57, %r54, -8;
sub.s32 %r58, %r51, %r57;
shr.s32 %r59, %r58, 31;
shr.u32 %r60, %r59, 30;
add.s32 %r61, %r58, %r60;
shr.s32 %r62, %r61, 2;
add.s32 %r63, %r57, %r62;
cvt.s64.s32 %rd21, %r63;
cvt.s64.s32 %rd1, %r58;
cvt.s64.s32 %rd2, %r56;
shl.b32 %r178, %r179, 3;
mul.hi.u32 %r64, %r46, %r29;
add.s32 %r65, %r64, %r46;
shr.u32 %r66, %r65, %r30;
mad.lo.s32 %r67, %r66, %r32, %r179;
mad.lo.s32 %r68, %r176, %r26, %r67;
mov.u32 %r69, %ctaid.x;
mad.lo.s32 %r70, %r69, %r22, %r68;
mul.wide.s32 %rd22, %r70, 84;
mul.wide.s32 %rd23, %r51, 4;
add.s64 %rd24, %rd22, %rd23;
cvta.to.global.u64 %rd25, %rd14;
add.s64 %rd26, %rd25, %rd24;
add.s64 %rd44, %rd26, 16;
add.s64 %rd27, %rd22, %rd21;
add.s64 %rd28, %rd25, %rd27;
add.s64 %rd43, %rd28, 4;
add.s64 %rd29, %rd25, %rd22;
add.s64 %rd42, %rd29, 80;
cvta.to.global.u64 %rd6, %rd15;
mul.lo.s32 %r71, %r177, %r27;
cvt.u64.u32 %rd30, %r71;
mul.lo.s32 %r72, %r46, %r33;
cvt.u64.u32 %rd31, %r72;
add.s64 %rd7, %rd30, %rd31;
mov.f32 %f48, 0f00000000;
shl.b64 %rd37, %rd1, 2;

$L__BB132_5:
cvt.u64.u32 %rd32, %r178;
add.s64 %rd33, %rd7, %rd32;
ld.global.nc.u32 %r111, [%rd44];
add.s64 %rd34, %rd33, %rd2;
mul.lo.s64 %rd35, %rd34, 36;
add.s64 %rd36, %rd6, %rd35;
add.s64 %rd38, %rd36, %rd37;
ld.global.nc.u32 %r79, [%rd38+4];
ld.global.nc.u32 %r73, [%rd36];

{.reg .f16 low,high;
mov.b32 {low,high},%r73;
cvt.f32.f16 %f7, low;}


ld.global.nc.u32 %r87, [%rd38+40];
ld.global.nc.u32 %r74, [%rd36+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r74;
cvt.f32.f16 %f8, low;}


ld.global.nc.u32 %r95, [%rd38+76];
ld.global.nc.u32 %r75, [%rd36+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r75;
cvt.f32.f16 %f9, low;}


ld.global.nc.u32 %r103, [%rd38+112];
ld.global.nc.u32 %r76, [%rd36+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r76;
cvt.f32.f16 %f10, low;}


ld.global.nc.u8 %rs1, [%rd43+-4];
cvt.u32.u16 %r112, %rs1;
and.b32 %r113, %r112, 240;
and.b32 %r78, %r111, 50529027;
mov.u32 %r108, 0;

dp4a.s32.s32 %r77, %r78, %r79, %r108;

and.b32 %r114, %r112, 15;
mul.lo.s32 %r115, %r77, %r114;
cvt.rn.f32.s32 %f13, %r115;
fma.rn.ftz.f32 %f14, %f7, %f13, 0f00000000;
shr.u32 %r116, %r113, 4;
shl.b32 %r117, %r116, 8;
or.b32 %r118, %r117, %r116;
shl.b32 %r119, %r118, 16;
or.b32 %r82, %r119, %r118;

dp4a.s32.s32 %r81, %r82, %r79, %r108;

cvt.rn.f32.s32 %f15, %r81;
fma.rn.ftz.f32 %f16, %f7, %f15, 0f00000000;
ld.global.nc.u8 %rs2, [%rd43+-2];
cvt.u32.u16 %r120, %rs2;
and.b32 %r121, %r120, 240;
shr.u32 %r122, %r111, 2;
and.b32 %r86, %r122, 50529027;

dp4a.s32.s32 %r85, %r86, %r87, %r108;

and.b32 %r123, %r120, 15;
mul.lo.s32 %r124, %r85, %r123;
cvt.rn.f32.s32 %f17, %r124;
fma.rn.ftz.f32 %f18, %f8, %f17, %f14;
shr.u32 %r125, %r121, 4;
shl.b32 %r126, %r125, 8;
or.b32 %r127, %r126, %r125;
shl.b32 %r128, %r127, 16;
or.b32 %r90, %r128, %r127;

dp4a.s32.s32 %r89, %r90, %r87, %r108;

cvt.rn.f32.s32 %f19, %r89;
fma.rn.ftz.f32 %f20, %f8, %f19, %f16;
ld.global.nc.u8 %rs3, [%rd43];
cvt.u32.u16 %r129, %rs3;
and.b32 %r130, %r129, 240;
shr.u32 %r131, %r111, 4;
and.b32 %r94, %r131, 50529027;

dp4a.s32.s32 %r93, %r94, %r95, %r108;

and.b32 %r132, %r129, 15;
mul.lo.s32 %r133, %r93, %r132;
cvt.rn.f32.s32 %f21, %r133;
fma.rn.ftz.f32 %f22, %f9, %f21, %f18;
shr.u32 %r134, %r130, 4;
shl.b32 %r135, %r134, 8;
or.b32 %r136, %r135, %r134;
shl.b32 %r137, %r136, 16;
or.b32 %r98, %r137, %r136;

dp4a.s32.s32 %r97, %r98, %r95, %r108;

cvt.rn.f32.s32 %f23, %r97;
fma.rn.ftz.f32 %f24, %f9, %f23, %f20;
ld.global.nc.u8 %rs4, [%rd43+2];
cvt.u32.u16 %r138, %rs4;
and.b32 %r139, %r138, 240;
shr.u32 %r140, %r111, 6;
and.b32 %r102, %r140, 50529027;

dp4a.s32.s32 %r101, %r102, %r103, %r108;

and.b32 %r141, %r138, 15;
mul.lo.s32 %r142, %r101, %r141;
cvt.rn.f32.s32 %f25, %r142;
fma.rn.ftz.f32 %f26, %f10, %f25, %f22;
shr.u32 %r143, %r139, 4;
shl.b32 %r144, %r143, 8;
or.b32 %r145, %r144, %r143;
shl.b32 %r146, %r145, 16;
or.b32 %r106, %r146, %r145;

dp4a.s32.s32 %r105, %r106, %r103, %r108;

cvt.rn.f32.s32 %f27, %r105;
fma.rn.ftz.f32 %f28, %f10, %f27, %f24;
ld.global.nc.u32 %r109, [%rd42];

{.reg .f16 low,high;
mov.b32 {low,high},%r109;
cvt.f32.f16 %f11, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r109;
cvt.f32.f16 %f12, high;}


mul.ftz.f32 %f29, %f26, %f11;
mul.ftz.f32 %f30, %f28, %f12;
sub.ftz.f32 %f31, %f29, %f30;
add.ftz.f32 %f48, %f48, %f31;
add.s32 %r178, %r178, 64;
add.s64 %rd44, %rd44, 672;
add.s64 %rd43, %rd43, 672;
add.s64 %rd42, %rd42, 672;
add.s32 %r179, %r179, 8;
setp.lt.s32 %p4, %r179, %r45;
@%p4 bra $L__BB132_5;

$L__BB132_6:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB132_8;

shl.b32 %r147, %r1, 7;
mov.u32 %r148, _ZZ13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r149, %r148, %r147;
shl.b32 %r150, %r2, 2;
add.s32 %r151, %r149, %r150;
st.shared.f32 [%r151+-128], %f48;

$L__BB132_8:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB132_11;

shl.b32 %r152, %r2, 2;
mov.u32 %r153, _ZZ13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r154, %r153, %r152;
mov.u32 %r155, 2;
ld.shared.f32 %f32, [%r154];
add.ftz.f32 %f33, %f48, %f32;
ld.shared.f32 %f34, [%r154+128];
add.ftz.f32 %f35, %f33, %f34;
ld.shared.f32 %f36, [%r154+256];
add.ftz.f32 %f37, %f35, %f36;
mov.b32 %r156, %f37;
mov.u32 %r157, 31;
mov.u32 %r158, 16;
mov.u32 %r159, -1;
shfl.sync.bfly.b32 %r160|%p7, %r156, %r158, %r157, %r159;
mov.b32 %f38, %r160;
add.ftz.f32 %f39, %f37, %f38;
mov.b32 %r161, %f39;
mov.u32 %r162, 8;
shfl.sync.bfly.b32 %r163|%p8, %r161, %r162, %r157, %r159;
mov.b32 %f40, %r163;
add.ftz.f32 %f41, %f39, %f40;
mov.b32 %r164, %f41;
mov.u32 %r165, 4;
shfl.sync.bfly.b32 %r166|%p9, %r164, %r165, %r157, %r159;
mov.b32 %f42, %r166;
add.ftz.f32 %f43, %f41, %f42;
mov.b32 %r167, %f43;
shfl.sync.bfly.b32 %r168|%p10, %r167, %r155, %r157, %r159;
mov.b32 %f44, %r168;
add.ftz.f32 %f4, %f43, %f44;
mov.b32 %r169, %f4;
mov.u32 %r170, 1;
shfl.sync.bfly.b32 %r17|%p1, %r169, %r170, %r157, %r159;
setp.ne.s32 %p11, %r2, 0;
@%p11 bra $L__BB132_11;

mov.u32 %r171, %ctaid.x;
mov.u32 %r172, %ctaid.y;
mad.lo.s32 %r173, %r172, %r28, %r171;
mov.u32 %r174, %ctaid.z;
mad.lo.s32 %r175, %r174, %r34, %r173;
cvta.to.global.u64 %rd39, %rd17;
mul.wide.u32 %rd40, %r175, 4;
add.s64 %rd41, %rd39, %rd40;
mov.b32 %f45, %r17;
add.ftz.f32 %f46, %f4, %f45;
st.global.f32 [%rd41], %f46;

$L__BB132_11:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type10ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type10ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type10ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type10ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type10ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type10ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type10ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type10ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type10ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<35>;
.reg .b16 %rs<9>;
.reg .f32 %f<193>;
.reg .b32 %r<389>;
.reg .b64 %rd<56>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type10ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type10ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type10ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type10ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type10ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r18, [_Z13mul_mat_vec_qIL9ggml_type10ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r19, [_Z13mul_mat_vec_qIL9ggml_type10ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type10ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type10ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type10ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type10ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type10ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type10ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type10ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type10ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type10ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type10ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type10ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
cvta.to.global.u64 %rd1, %rd19;
mov.u32 %r1, %tid.y;
shl.b32 %r34, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r34, %r2;
mov.u32 %r35, %ctaid.x;
shl.b32 %r4, %r35, 1;
mov.u32 %r5, %ctaid.y;
shr.s32 %r36, %r3, 31;
shr.u32 %r37, %r36, 28;
add.s32 %r38, %r3, %r37;
shr.s32 %r388, %r38, 4;
shr.u32 %r7, %r33, 8;
setp.ge.s32 %p3, %r388, %r7;
mov.f32 %f189, 0f00000000;
mov.f32 %f190, %f189;
mov.f32 %f191, %f189;
mov.f32 %f192, %f189;
@%p3 bra $L__BB133_3;

mov.u32 %r39, %ctaid.z;
mul.lo.s32 %r40, %r5, %r25;
cvt.u64.u32 %rd20, %r40;
mul.lo.s32 %r41, %r39, %r31;
cvt.u64.u32 %rd21, %r41;
add.s64 %rd2, %rd21, %rd20;
mul.hi.u32 %r42, %r39, %r27;
add.s32 %r43, %r42, %r39;
shr.u32 %r44, %r43, %r28;
mul.lo.s32 %r45, %r44, %r30;
mul.hi.u32 %r46, %r5, %r21;
add.s32 %r47, %r46, %r5;
shr.u32 %r48, %r47, %r22;
mul.lo.s32 %r49, %r48, %r24;
cvta.to.global.u64 %rd55, %rd17;
and.b32 %r53, %r38, -16;
sub.s32 %r54, %r3, %r53;
shr.s32 %r55, %r54, 31;
shr.u32 %r56, %r55, 29;
add.s32 %r57, %r54, %r56;
shr.s32 %r58, %r57, 3;
shl.b32 %r59, %r58, 2;
and.b32 %r60, %r57, -8;
sub.s32 %r61, %r54, %r60;
shr.s32 %r62, %r61, 31;
shr.u32 %r63, %r62, 30;
add.s32 %r64, %r61, %r63;
shr.s32 %r65, %r64, 2;
add.s32 %r66, %r60, %r65;
cvt.s64.s32 %rd22, %r66;
cvt.s64.s32 %rd4, %r61;
add.s32 %r67, %r4, 1;
cvt.s64.s32 %rd5, %r59;
shl.b32 %r386, %r388, 3;
add.s32 %r387, %r19, %r386;
add.s32 %r68, %r388, %r45;
add.s32 %r69, %r68, %r49;
mad.lo.s32 %r70, %r67, %r18, %r69;
mul.wide.s32 %rd6, %r70, 84;
mad.lo.s32 %r71, %r4, %r18, %r388;
add.s32 %r72, %r71, %r45;
add.s32 %r73, %r72, %r49;
mul.wide.s32 %rd7, %r73, 84;
mul.wide.s32 %rd23, %r54, 4;
add.s64 %rd24, %rd7, %rd23;
add.s64 %rd8, %rd24, 16;
add.s64 %rd25, %rd7, %rd22;
add.s64 %rd9, %rd25, 6;
add.s64 %rd10, %rd6, %rd22;
add.s64 %rd26, %rd6, %rd23;
add.s64 %rd11, %rd26, 16;
cvta.to.global.u64 %rd12, %rd18;
mov.f32 %f189, 0f00000000;
shl.b64 %rd33, %rd4, 2;

$L__BB133_2:
cvt.u64.u32 %rd27, %r386;
add.s64 %rd28, %rd2, %rd27;
add.s64 %rd29, %rd55, %rd8;
ld.global.nc.u32 %r226, [%rd29];
add.s64 %rd30, %rd28, %rd5;
mul.lo.s64 %rd31, %rd30, 36;
add.s64 %rd32, %rd12, %rd31;
add.s64 %rd34, %rd32, %rd33;
ld.global.nc.u32 %r80, [%rd34+4];
ld.global.nc.u32 %r74, [%rd32];

{.reg .f16 low,high;
mov.b32 {low,high},%r74;
cvt.f32.f16 %f25, low;}


ld.global.nc.u32 %r88, [%rd34+40];
ld.global.nc.u32 %r75, [%rd32+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r75;
cvt.f32.f16 %f26, low;}


ld.global.nc.u32 %r96, [%rd34+76];
ld.global.nc.u32 %r76, [%rd32+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r76;
cvt.f32.f16 %f27, low;}


ld.global.nc.u32 %r104, [%rd34+112];
ld.global.nc.u32 %r77, [%rd32+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r77;
cvt.f32.f16 %f28, low;}


add.s64 %rd35, %rd55, %rd9;
ld.global.nc.u8 %rs1, [%rd35+-6];
cvt.u32.u16 %r227, %rs1;
and.b32 %r228, %r227, 240;
and.b32 %r155, %r226, 50529027;
mov.u32 %r223, 0;

dp4a.s32.s32 %r78, %r155, %r80, %r223;

and.b32 %r229, %r227, 15;
mul.lo.s32 %r230, %r78, %r229;
cvt.rn.f32.s32 %f49, %r230;
fma.rn.ftz.f32 %f50, %f25, %f49, 0f00000000;
shr.u32 %r231, %r228, 4;
shl.b32 %r232, %r231, 8;
or.b32 %r233, %r232, %r231;
shl.b32 %r234, %r233, 16;
or.b32 %r159, %r234, %r233;

dp4a.s32.s32 %r82, %r159, %r80, %r223;

cvt.rn.f32.s32 %f51, %r82;
fma.rn.ftz.f32 %f52, %f25, %f51, 0f00000000;
ld.global.nc.u8 %rs2, [%rd35+-4];
cvt.u32.u16 %r235, %rs2;
and.b32 %r236, %r235, 240;
shr.u32 %r237, %r226, 2;
and.b32 %r163, %r237, 50529027;

dp4a.s32.s32 %r86, %r163, %r88, %r223;

and.b32 %r238, %r235, 15;
mul.lo.s32 %r239, %r86, %r238;
cvt.rn.f32.s32 %f53, %r239;
fma.rn.ftz.f32 %f54, %f26, %f53, %f50;
shr.u32 %r240, %r236, 4;
shl.b32 %r241, %r240, 8;
or.b32 %r242, %r241, %r240;
shl.b32 %r243, %r242, 16;
or.b32 %r167, %r243, %r242;

dp4a.s32.s32 %r90, %r167, %r88, %r223;

cvt.rn.f32.s32 %f55, %r90;
fma.rn.ftz.f32 %f56, %f26, %f55, %f52;
ld.global.nc.u8 %rs3, [%rd35+-2];
cvt.u32.u16 %r244, %rs3;
and.b32 %r245, %r244, 240;
shr.u32 %r246, %r226, 4;
and.b32 %r171, %r246, 50529027;

dp4a.s32.s32 %r94, %r171, %r96, %r223;

and.b32 %r247, %r244, 15;
mul.lo.s32 %r248, %r94, %r247;
cvt.rn.f32.s32 %f57, %r248;
fma.rn.ftz.f32 %f58, %f27, %f57, %f54;
shr.u32 %r249, %r245, 4;
shl.b32 %r250, %r249, 8;
or.b32 %r251, %r250, %r249;
shl.b32 %r252, %r251, 16;
or.b32 %r175, %r252, %r251;

dp4a.s32.s32 %r98, %r175, %r96, %r223;

cvt.rn.f32.s32 %f59, %r98;
fma.rn.ftz.f32 %f60, %f27, %f59, %f56;
ld.global.nc.u8 %rs4, [%rd35];
cvt.u32.u16 %r253, %rs4;
and.b32 %r254, %r253, 240;
shr.u32 %r255, %r226, 6;
and.b32 %r179, %r255, 50529027;

dp4a.s32.s32 %r102, %r179, %r104, %r223;

and.b32 %r256, %r253, 15;
mul.lo.s32 %r257, %r102, %r256;
cvt.rn.f32.s32 %f61, %r257;
fma.rn.ftz.f32 %f62, %f28, %f61, %f58;
shr.u32 %r258, %r254, 4;
shl.b32 %r259, %r258, 8;
or.b32 %r260, %r259, %r258;
shl.b32 %r261, %r260, 16;
or.b32 %r183, %r261, %r260;

dp4a.s32.s32 %r106, %r183, %r104, %r223;

cvt.rn.f32.s32 %f63, %r106;
fma.rn.ftz.f32 %f64, %f28, %f63, %f60;
add.s64 %rd36, %rd55, %rd7;
ld.global.nc.u32 %r110, [%rd36+80];

{.reg .f16 low,high;
mov.b32 {low,high},%r110;
cvt.f32.f16 %f29, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r110;
cvt.f32.f16 %f30, high;}


mul.ftz.f32 %f65, %f62, %f29;
mul.ftz.f32 %f66, %f64, %f30;
sub.ftz.f32 %f67, %f65, %f66;
add.ftz.f32 %f192, %f67, %f192;
add.s64 %rd37, %rd55, %rd11;
ld.global.nc.u32 %r262, [%rd37];

{.reg .f16 low,high;
mov.b32 {low,high},%r74;
cvt.f32.f16 %f31, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r75;
cvt.f32.f16 %f32, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r76;
cvt.f32.f16 %f33, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r77;
cvt.f32.f16 %f34, low;}


add.s64 %rd38, %rd55, %rd10;
ld.global.nc.u8 %rs5, [%rd38];
cvt.u32.u16 %r263, %rs5;
and.b32 %r264, %r263, 240;
and.b32 %r193, %r262, 50529027;

dp4a.s32.s32 %r116, %r193, %r80, %r223;

and.b32 %r265, %r263, 15;
mul.lo.s32 %r266, %r116, %r265;
cvt.rn.f32.s32 %f68, %r266;
fma.rn.ftz.f32 %f69, %f31, %f68, 0f00000000;
shr.u32 %r267, %r264, 4;
shl.b32 %r268, %r267, 8;
or.b32 %r269, %r268, %r267;
shl.b32 %r270, %r269, 16;
or.b32 %r197, %r270, %r269;

dp4a.s32.s32 %r120, %r197, %r80, %r223;

cvt.rn.f32.s32 %f70, %r120;
fma.rn.ftz.f32 %f71, %f31, %f70, 0f00000000;
ld.global.nc.u8 %rs6, [%rd38+2];
cvt.u32.u16 %r271, %rs6;
and.b32 %r272, %r271, 240;
shr.u32 %r273, %r262, 2;
and.b32 %r201, %r273, 50529027;

dp4a.s32.s32 %r124, %r201, %r88, %r223;

and.b32 %r274, %r271, 15;
mul.lo.s32 %r275, %r124, %r274;
cvt.rn.f32.s32 %f72, %r275;
fma.rn.ftz.f32 %f73, %f32, %f72, %f69;
shr.u32 %r276, %r272, 4;
shl.b32 %r277, %r276, 8;
or.b32 %r278, %r277, %r276;
shl.b32 %r279, %r278, 16;
or.b32 %r205, %r279, %r278;

dp4a.s32.s32 %r128, %r205, %r88, %r223;

cvt.rn.f32.s32 %f74, %r128;
fma.rn.ftz.f32 %f75, %f32, %f74, %f71;
ld.global.nc.u8 %rs7, [%rd38+4];
cvt.u32.u16 %r280, %rs7;
and.b32 %r281, %r280, 240;
shr.u32 %r282, %r262, 4;
and.b32 %r209, %r282, 50529027;

dp4a.s32.s32 %r132, %r209, %r96, %r223;

and.b32 %r283, %r280, 15;
mul.lo.s32 %r284, %r132, %r283;
cvt.rn.f32.s32 %f76, %r284;
fma.rn.ftz.f32 %f77, %f33, %f76, %f73;
shr.u32 %r285, %r281, 4;
shl.b32 %r286, %r285, 8;
or.b32 %r287, %r286, %r285;
shl.b32 %r288, %r287, 16;
or.b32 %r213, %r288, %r287;

dp4a.s32.s32 %r136, %r213, %r96, %r223;

cvt.rn.f32.s32 %f78, %r136;
fma.rn.ftz.f32 %f79, %f33, %f78, %f75;
ld.global.nc.u8 %rs8, [%rd38+6];
cvt.u32.u16 %r289, %rs8;
and.b32 %r290, %r289, 240;
shr.u32 %r291, %r262, 6;
and.b32 %r217, %r291, 50529027;

dp4a.s32.s32 %r140, %r217, %r104, %r223;

and.b32 %r292, %r289, 15;
mul.lo.s32 %r293, %r140, %r292;
cvt.rn.f32.s32 %f80, %r293;
fma.rn.ftz.f32 %f81, %f34, %f80, %f77;
shr.u32 %r294, %r290, 4;
shl.b32 %r295, %r294, 8;
or.b32 %r296, %r295, %r294;
shl.b32 %r297, %r296, 16;
or.b32 %r221, %r297, %r296;

dp4a.s32.s32 %r144, %r221, %r104, %r223;

cvt.rn.f32.s32 %f82, %r144;
fma.rn.ftz.f32 %f83, %f34, %f82, %f79;
add.s64 %rd39, %rd55, %rd6;
ld.global.nc.u32 %r148, [%rd39+80];

{.reg .f16 low,high;
mov.b32 {low,high},%r148;
cvt.f32.f16 %f35, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r148;
cvt.f32.f16 %f36, high;}


mul.ftz.f32 %f84, %f81, %f35;
mul.ftz.f32 %f85, %f83, %f36;
sub.ftz.f32 %f86, %f84, %f85;
add.ftz.f32 %f191, %f86, %f191;
cvt.u64.u32 %rd40, %r387;
add.s64 %rd41, %rd2, %rd40;
add.s64 %rd42, %rd41, %rd5;
mul.lo.s64 %rd43, %rd42, 36;
add.s64 %rd44, %rd12, %rd43;
add.s64 %rd45, %rd44, %rd33;
ld.global.nc.u32 %r156, [%rd45+4];
ld.global.nc.u32 %r150, [%rd44];

{.reg .f16 low,high;
mov.b32 {low,high},%r150;
cvt.f32.f16 %f37, low;}


ld.global.nc.u32 %r164, [%rd45+40];
ld.global.nc.u32 %r151, [%rd44+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r151;
cvt.f32.f16 %f38, low;}


ld.global.nc.u32 %r172, [%rd45+76];
ld.global.nc.u32 %r152, [%rd44+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r152;
cvt.f32.f16 %f39, low;}


ld.global.nc.u32 %r180, [%rd45+112];
ld.global.nc.u32 %r153, [%rd44+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r153;
cvt.f32.f16 %f40, low;}



dp4a.s32.s32 %r154, %r155, %r156, %r223;

mul.lo.s32 %r298, %r154, %r229;
cvt.rn.f32.s32 %f87, %r298;
fma.rn.ftz.f32 %f88, %f37, %f87, 0f00000000;

dp4a.s32.s32 %r158, %r159, %r156, %r223;

cvt.rn.f32.s32 %f89, %r158;
fma.rn.ftz.f32 %f90, %f37, %f89, 0f00000000;

dp4a.s32.s32 %r162, %r163, %r164, %r223;

mul.lo.s32 %r299, %r162, %r238;
cvt.rn.f32.s32 %f91, %r299;
fma.rn.ftz.f32 %f92, %f38, %f91, %f88;

dp4a.s32.s32 %r166, %r167, %r164, %r223;

cvt.rn.f32.s32 %f93, %r166;
fma.rn.ftz.f32 %f94, %f38, %f93, %f90;

dp4a.s32.s32 %r170, %r171, %r172, %r223;

mul.lo.s32 %r300, %r170, %r247;
cvt.rn.f32.s32 %f95, %r300;
fma.rn.ftz.f32 %f96, %f39, %f95, %f92;

dp4a.s32.s32 %r174, %r175, %r172, %r223;

cvt.rn.f32.s32 %f97, %r174;
fma.rn.ftz.f32 %f98, %f39, %f97, %f94;

dp4a.s32.s32 %r178, %r179, %r180, %r223;

mul.lo.s32 %r301, %r178, %r256;
cvt.rn.f32.s32 %f99, %r301;
fma.rn.ftz.f32 %f100, %f40, %f99, %f96;

dp4a.s32.s32 %r182, %r183, %r180, %r223;

cvt.rn.f32.s32 %f101, %r182;
fma.rn.ftz.f32 %f102, %f40, %f101, %f98;

{.reg .f16 low,high;
mov.b32 {low,high},%r110;
cvt.f32.f16 %f41, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r110;
cvt.f32.f16 %f42, high;}


mul.ftz.f32 %f103, %f100, %f41;
mul.ftz.f32 %f104, %f102, %f42;
sub.ftz.f32 %f105, %f103, %f104;
add.ftz.f32 %f190, %f105, %f190;

{.reg .f16 low,high;
mov.b32 {low,high},%r150;
cvt.f32.f16 %f43, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r151;
cvt.f32.f16 %f44, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r152;
cvt.f32.f16 %f45, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r153;
cvt.f32.f16 %f46, low;}



dp4a.s32.s32 %r192, %r193, %r156, %r223;

mul.lo.s32 %r302, %r192, %r265;
cvt.rn.f32.s32 %f106, %r302;
fma.rn.ftz.f32 %f107, %f43, %f106, 0f00000000;

dp4a.s32.s32 %r196, %r197, %r156, %r223;

cvt.rn.f32.s32 %f108, %r196;
fma.rn.ftz.f32 %f109, %f43, %f108, 0f00000000;

dp4a.s32.s32 %r200, %r201, %r164, %r223;

mul.lo.s32 %r303, %r200, %r274;
cvt.rn.f32.s32 %f110, %r303;
fma.rn.ftz.f32 %f111, %f44, %f110, %f107;

dp4a.s32.s32 %r204, %r205, %r164, %r223;

cvt.rn.f32.s32 %f112, %r204;
fma.rn.ftz.f32 %f113, %f44, %f112, %f109;

dp4a.s32.s32 %r208, %r209, %r172, %r223;

mul.lo.s32 %r304, %r208, %r283;
cvt.rn.f32.s32 %f114, %r304;
fma.rn.ftz.f32 %f115, %f45, %f114, %f111;

dp4a.s32.s32 %r212, %r213, %r172, %r223;

cvt.rn.f32.s32 %f116, %r212;
fma.rn.ftz.f32 %f117, %f45, %f116, %f113;

dp4a.s32.s32 %r216, %r217, %r180, %r223;

mul.lo.s32 %r305, %r216, %r292;
cvt.rn.f32.s32 %f118, %r305;
fma.rn.ftz.f32 %f119, %f46, %f118, %f115;

dp4a.s32.s32 %r220, %r221, %r180, %r223;

cvt.rn.f32.s32 %f120, %r220;
fma.rn.ftz.f32 %f121, %f46, %f120, %f117;

{.reg .f16 low,high;
mov.b32 {low,high},%r148;
cvt.f32.f16 %f47, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r148;
cvt.f32.f16 %f48, high;}


mul.ftz.f32 %f122, %f119, %f47;
mul.ftz.f32 %f123, %f121, %f48;
sub.ftz.f32 %f124, %f122, %f123;
add.ftz.f32 %f189, %f124, %f189;
add.s32 %r387, %r387, 64;
add.s32 %r386, %r386, 64;
add.s64 %rd55, %rd55, 672;
add.s32 %r388, %r388, 8;
setp.lt.s32 %p4, %r388, %r7;
@%p4 bra $L__BB133_2;

$L__BB133_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB133_5;

shl.b32 %r306, %r1, 9;
mov.u32 %r307, _ZZ13mul_mat_vec_qIL9ggml_type10ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r308, %r307, %r306;
shl.b32 %r309, %r2, 2;
add.s32 %r310, %r308, %r309;
st.shared.f32 [%r310+-512], %f192;
st.shared.f32 [%r310+-384], %f191;
st.shared.f32 [%r310+-256], %f190;
st.shared.f32 [%r310+-128], %f189;

$L__BB133_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB133_14;

mov.u32 %r311, %ctaid.z;
mad.lo.s32 %r312, %r5, %r26, %r4;
mad.lo.s32 %r313, %r311, %r32, %r312;
cvt.u64.u32 %rd15, %r313;
shl.b32 %r314, %r2, 2;
mov.u32 %r315, _ZZ13mul_mat_vec_qIL9ggml_type10ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r16, %r315, %r314;
mov.u32 %r316, 2;
ld.shared.f32 %f125, [%r16];
add.ftz.f32 %f126, %f125, %f192;
ld.shared.f32 %f127, [%r16+512];
add.ftz.f32 %f128, %f127, %f126;
ld.shared.f32 %f129, [%r16+1024];
add.ftz.f32 %f130, %f129, %f128;
mov.b32 %r317, %f130;
mov.u32 %r318, 31;
mov.u32 %r319, 16;
mov.u32 %r320, -1;
shfl.sync.bfly.b32 %r321|%p7, %r317, %r319, %r318, %r320;
mov.b32 %f131, %r321;
add.ftz.f32 %f132, %f130, %f131;
mov.b32 %r322, %f132;
mov.u32 %r323, 8;
shfl.sync.bfly.b32 %r324|%p8, %r322, %r323, %r318, %r320;
mov.b32 %f133, %r324;
add.ftz.f32 %f134, %f132, %f133;
mov.b32 %r325, %f134;
mov.u32 %r326, 4;
shfl.sync.bfly.b32 %r327|%p9, %r325, %r326, %r318, %r320;
mov.b32 %f135, %r327;
add.ftz.f32 %f136, %f134, %f135;
mov.b32 %r328, %f136;
shfl.sync.bfly.b32 %r329|%p10, %r328, %r316, %r318, %r320;
mov.b32 %f137, %r329;
add.ftz.f32 %f138, %f136, %f137;
mov.b32 %r330, %f138;
mov.u32 %r331, 1;
shfl.sync.bfly.b32 %r332|%p11, %r330, %r331, %r318, %r320;
mov.b32 %f139, %r332;
add.ftz.f32 %f13, %f138, %f139;
setp.lt.u32 %p12, %r4, %r20;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
mul.wide.u32 %rd46, %r313, 4;
add.s64 %rd16, %rd1, %rd46;
not.pred %p14, %p1;
@%p14 bra $L__BB133_8;

st.global.f32 [%rd16], %f13;

$L__BB133_8:
ld.shared.f32 %f140, [%r16+128];
add.ftz.f32 %f141, %f140, %f191;
ld.shared.f32 %f142, [%r16+640];
add.ftz.f32 %f143, %f142, %f141;
ld.shared.f32 %f144, [%r16+1152];
add.ftz.f32 %f145, %f144, %f143;
mov.b32 %r333, %f145;
shfl.sync.bfly.b32 %r337|%p15, %r333, %r319, %r318, %r320;
mov.b32 %f146, %r337;
add.ftz.f32 %f147, %f145, %f146;
mov.b32 %r338, %f147;
shfl.sync.bfly.b32 %r340|%p16, %r338, %r323, %r318, %r320;
mov.b32 %f148, %r340;
add.ftz.f32 %f149, %f147, %f148;
mov.b32 %r341, %f149;
shfl.sync.bfly.b32 %r343|%p17, %r341, %r326, %r318, %r320;
mov.b32 %f150, %r343;
add.ftz.f32 %f151, %f149, %f150;
mov.b32 %r344, %f151;
shfl.sync.bfly.b32 %r346|%p18, %r344, %r316, %r318, %r320;
mov.b32 %f152, %r346;
add.ftz.f32 %f153, %f151, %f152;
mov.b32 %r347, %f153;
shfl.sync.bfly.b32 %r349|%p19, %r347, %r331, %r318, %r320;
mov.b32 %f154, %r349;
add.ftz.f32 %f14, %f153, %f154;
add.s32 %r350, %r4, 1;
setp.lt.u32 %p20, %r350, %r20;
setp.eq.s32 %p21, %r2, 1;
and.pred %p2, %p21, %p20;
not.pred %p22, %p2;
@%p22 bra $L__BB133_10;

st.global.f32 [%rd16+4], %f14;

$L__BB133_10:
ld.shared.f32 %f155, [%r16+256];
add.ftz.f32 %f156, %f155, %f190;
ld.shared.f32 %f157, [%r16+768];
add.ftz.f32 %f158, %f157, %f156;
ld.shared.f32 %f159, [%r16+1280];
add.ftz.f32 %f160, %f159, %f158;
mov.b32 %r351, %f160;
mov.u32 %r352, 31;
mov.u32 %r353, 16;
mov.u32 %r354, -1;
shfl.sync.bfly.b32 %r355|%p23, %r351, %r353, %r352, %r354;
mov.b32 %f161, %r355;
add.ftz.f32 %f162, %f160, %f161;
mov.b32 %r356, %f162;
mov.u32 %r357, 8;
shfl.sync.bfly.b32 %r358|%p24, %r356, %r357, %r352, %r354;
mov.b32 %f163, %r358;
add.ftz.f32 %f164, %f162, %f163;
mov.b32 %r359, %f164;
mov.u32 %r360, 4;
shfl.sync.bfly.b32 %r361|%p25, %r359, %r360, %r352, %r354;
mov.b32 %f165, %r361;
add.ftz.f32 %f166, %f164, %f165;
mov.b32 %r362, %f166;
mov.u32 %r363, 2;
shfl.sync.bfly.b32 %r364|%p26, %r362, %r363, %r352, %r354;
mov.b32 %f167, %r364;
add.ftz.f32 %f168, %f166, %f167;
mov.b32 %r365, %f168;
mov.u32 %r366, 1;
shfl.sync.bfly.b32 %r367|%p27, %r365, %r366, %r352, %r354;
mov.b32 %f169, %r367;
add.ftz.f32 %f15, %f168, %f169;
@%p14 bra $L__BB133_12;

cvt.u64.u32 %rd47, %r20;
add.s64 %rd48, %rd47, %rd15;
shl.b64 %rd49, %rd48, 2;
add.s64 %rd50, %rd1, %rd49;
st.global.f32 [%rd50], %f15;

$L__BB133_12:
ld.shared.f32 %f170, [%r16+384];
add.ftz.f32 %f171, %f170, %f189;
ld.shared.f32 %f172, [%r16+896];
add.ftz.f32 %f173, %f172, %f171;
ld.shared.f32 %f174, [%r16+1408];
add.ftz.f32 %f175, %f174, %f173;
mov.b32 %r368, %f175;
shfl.sync.bfly.b32 %r372|%p29, %r368, %r353, %r352, %r354;
mov.b32 %f176, %r372;
add.ftz.f32 %f177, %f175, %f176;
mov.b32 %r373, %f177;
shfl.sync.bfly.b32 %r375|%p30, %r373, %r357, %r352, %r354;
mov.b32 %f178, %r375;
add.ftz.f32 %f179, %f177, %f178;
mov.b32 %r376, %f179;
shfl.sync.bfly.b32 %r378|%p31, %r376, %r360, %r352, %r354;
mov.b32 %f180, %r378;
add.ftz.f32 %f181, %f179, %f180;
mov.b32 %r379, %f181;
shfl.sync.bfly.b32 %r381|%p32, %r379, %r363, %r352, %r354;
mov.b32 %f182, %r381;
add.ftz.f32 %f183, %f181, %f182;
mov.b32 %r382, %f183;
shfl.sync.bfly.b32 %r384|%p33, %r382, %r366, %r352, %r354;
mov.b32 %f184, %r384;
add.ftz.f32 %f16, %f183, %f184;
@%p22 bra $L__BB133_14;

add.s32 %r385, %r20, 1;
cvt.u64.u32 %rd51, %r385;
add.s64 %rd52, %rd51, %rd15;
shl.b64 %rd53, %rd52, 2;
add.s64 %rd54, %rd1, %rd53;
st.global.f32 [%rd54], %f16;

$L__BB133_14:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type10ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type10ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type10ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type10ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type10ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type10ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type10ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type10ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type10ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<47>;
.reg .b16 %rs<9>;
.reg .f32 %f<289>;
.reg .b32 %r<512>;
.reg .b64 %rd<66>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type10ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[2304];

ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type10ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type10ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd20, [_Z13mul_mat_vec_qIL9ggml_type10ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type10ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type10ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type10ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type10ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type10ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type10ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type10ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type10ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type10ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type10ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type10ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type10ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type10ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type10ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
cvta.to.global.u64 %rd1, %rd20;
mov.u32 %r1, %tid.y;
shl.b32 %r37, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r37, %r2;
mov.u32 %r38, %ctaid.x;
shl.b32 %r4, %r38, 1;
mov.u32 %r5, %ctaid.y;
shr.s32 %r39, %r3, 31;
shr.u32 %r40, %r39, 28;
add.s32 %r41, %r3, %r40;
shr.s32 %r511, %r41, 4;
shr.u32 %r7, %r36, 8;
setp.ge.s32 %p3, %r511, %r7;
mov.f32 %f283, 0f00000000;
mov.f32 %f284, %f283;
mov.f32 %f285, %f283;
mov.f32 %f286, %f283;
mov.f32 %f287, %f283;
mov.f32 %f288, %f283;
@%p3 bra $L__BB134_3;

mov.u32 %r42, %ctaid.z;
mul.lo.s32 %r43, %r5, %r28;
cvt.u64.u32 %rd21, %r43;
mul.lo.s32 %r44, %r42, %r34;
cvt.u64.u32 %rd22, %r44;
add.s64 %rd2, %rd22, %rd21;
mul.hi.u32 %r45, %r42, %r30;
add.s32 %r46, %r45, %r42;
shr.u32 %r47, %r46, %r31;
mul.lo.s32 %r48, %r47, %r33;
mul.hi.u32 %r49, %r5, %r24;
add.s32 %r50, %r49, %r5;
shr.u32 %r51, %r50, %r25;
mul.lo.s32 %r52, %r51, %r27;
cvta.to.global.u64 %rd65, %rd18;
and.b32 %r56, %r41, -16;
sub.s32 %r57, %r3, %r56;
shr.s32 %r58, %r57, 31;
shr.u32 %r59, %r58, 29;
add.s32 %r60, %r57, %r59;
shr.s32 %r61, %r60, 3;
shl.b32 %r62, %r61, 2;
and.b32 %r63, %r60, -8;
sub.s32 %r64, %r57, %r63;
shr.s32 %r65, %r64, 31;
shr.u32 %r66, %r65, 30;
add.s32 %r67, %r64, %r66;
shr.s32 %r68, %r67, 2;
add.s32 %r69, %r63, %r68;
cvt.s64.s32 %rd23, %r69;
cvt.s64.s32 %rd4, %r64;
add.s32 %r70, %r4, 1;
cvt.s64.s32 %rd5, %r62;
shl.b32 %r71, %r22, 1;
shl.b32 %r508, %r511, 3;
add.s32 %r510, %r508, %r71;
add.s32 %r509, %r22, %r508;
add.s32 %r72, %r511, %r48;
add.s32 %r73, %r72, %r52;
mad.lo.s32 %r74, %r70, %r21, %r73;
mul.wide.s32 %rd6, %r74, 84;
mad.lo.s32 %r75, %r4, %r21, %r511;
add.s32 %r76, %r75, %r48;
add.s32 %r77, %r76, %r52;
mul.wide.s32 %rd7, %r77, 84;
mul.wide.s32 %rd24, %r57, 4;
add.s64 %rd25, %rd7, %rd24;
add.s64 %rd8, %rd25, 16;
add.s64 %rd26, %rd7, %rd23;
add.s64 %rd9, %rd26, 6;
add.s64 %rd10, %rd6, %rd23;
add.s64 %rd27, %rd6, %rd24;
add.s64 %rd11, %rd27, 16;
cvta.to.global.u64 %rd12, %rd19;
mov.f32 %f283, 0f00000000;
shl.b64 %rd34, %rd4, 2;

$L__BB134_2:
cvt.u64.u32 %rd28, %r508;
add.s64 %rd29, %rd2, %rd28;
add.s64 %rd30, %rd65, %rd8;
ld.global.nc.u32 %r306, [%rd30];
add.s64 %rd31, %rd29, %rd5;
mul.lo.s64 %rd32, %rd31, 36;
add.s64 %rd33, %rd12, %rd32;
add.s64 %rd35, %rd33, %rd34;
ld.global.nc.u32 %r84, [%rd35+4];
ld.global.nc.u32 %r78, [%rd33];

{.reg .f16 low,high;
mov.b32 {low,high},%r78;
cvt.f32.f16 %f37, low;}


ld.global.nc.u32 %r92, [%rd35+40];
ld.global.nc.u32 %r79, [%rd33+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r79;
cvt.f32.f16 %f38, low;}


ld.global.nc.u32 %r100, [%rd35+76];
ld.global.nc.u32 %r80, [%rd33+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r80;
cvt.f32.f16 %f39, low;}


ld.global.nc.u32 %r108, [%rd35+112];
ld.global.nc.u32 %r81, [%rd33+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r81;
cvt.f32.f16 %f40, low;}


add.s64 %rd36, %rd65, %rd9;
ld.global.nc.u8 %rs1, [%rd36+-6];
cvt.u32.u16 %r307, %rs1;
and.b32 %r308, %r307, 240;
and.b32 %r235, %r306, 50529027;
mov.u32 %r303, 0;

dp4a.s32.s32 %r82, %r235, %r84, %r303;

and.b32 %r309, %r307, 15;
mul.lo.s32 %r310, %r82, %r309;
cvt.rn.f32.s32 %f73, %r310;
fma.rn.ftz.f32 %f74, %f37, %f73, 0f00000000;
shr.u32 %r311, %r308, 4;
shl.b32 %r312, %r311, 8;
or.b32 %r313, %r312, %r311;
shl.b32 %r314, %r313, 16;
or.b32 %r239, %r314, %r313;

dp4a.s32.s32 %r86, %r239, %r84, %r303;

cvt.rn.f32.s32 %f75, %r86;
fma.rn.ftz.f32 %f76, %f37, %f75, 0f00000000;
ld.global.nc.u8 %rs2, [%rd36+-4];
cvt.u32.u16 %r315, %rs2;
and.b32 %r316, %r315, 240;
shr.u32 %r317, %r306, 2;
and.b32 %r243, %r317, 50529027;

dp4a.s32.s32 %r90, %r243, %r92, %r303;

and.b32 %r318, %r315, 15;
mul.lo.s32 %r319, %r90, %r318;
cvt.rn.f32.s32 %f77, %r319;
fma.rn.ftz.f32 %f78, %f38, %f77, %f74;
shr.u32 %r320, %r316, 4;
shl.b32 %r321, %r320, 8;
or.b32 %r322, %r321, %r320;
shl.b32 %r323, %r322, 16;
or.b32 %r247, %r323, %r322;

dp4a.s32.s32 %r94, %r247, %r92, %r303;

cvt.rn.f32.s32 %f79, %r94;
fma.rn.ftz.f32 %f80, %f38, %f79, %f76;
ld.global.nc.u8 %rs3, [%rd36+-2];
cvt.u32.u16 %r324, %rs3;
and.b32 %r325, %r324, 240;
shr.u32 %r326, %r306, 4;
and.b32 %r251, %r326, 50529027;

dp4a.s32.s32 %r98, %r251, %r100, %r303;

and.b32 %r327, %r324, 15;
mul.lo.s32 %r328, %r98, %r327;
cvt.rn.f32.s32 %f81, %r328;
fma.rn.ftz.f32 %f82, %f39, %f81, %f78;
shr.u32 %r329, %r325, 4;
shl.b32 %r330, %r329, 8;
or.b32 %r331, %r330, %r329;
shl.b32 %r332, %r331, 16;
or.b32 %r255, %r332, %r331;

dp4a.s32.s32 %r102, %r255, %r100, %r303;

cvt.rn.f32.s32 %f83, %r102;
fma.rn.ftz.f32 %f84, %f39, %f83, %f80;
ld.global.nc.u8 %rs4, [%rd36];
cvt.u32.u16 %r333, %rs4;
and.b32 %r334, %r333, 240;
shr.u32 %r335, %r306, 6;
and.b32 %r259, %r335, 50529027;

dp4a.s32.s32 %r106, %r259, %r108, %r303;

and.b32 %r336, %r333, 15;
mul.lo.s32 %r337, %r106, %r336;
cvt.rn.f32.s32 %f85, %r337;
fma.rn.ftz.f32 %f86, %f40, %f85, %f82;
shr.u32 %r338, %r334, 4;
shl.b32 %r339, %r338, 8;
or.b32 %r340, %r339, %r338;
shl.b32 %r341, %r340, 16;
or.b32 %r263, %r341, %r340;

dp4a.s32.s32 %r110, %r263, %r108, %r303;

cvt.rn.f32.s32 %f87, %r110;
fma.rn.ftz.f32 %f88, %f40, %f87, %f84;
add.s64 %rd37, %rd65, %rd7;
ld.global.nc.u32 %r114, [%rd37+80];

{.reg .f16 low,high;
mov.b32 {low,high},%r114;
cvt.f32.f16 %f41, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r114;
cvt.f32.f16 %f42, high;}


mul.ftz.f32 %f89, %f86, %f41;
mul.ftz.f32 %f90, %f88, %f42;
sub.ftz.f32 %f91, %f89, %f90;
add.ftz.f32 %f288, %f91, %f288;
add.s64 %rd38, %rd65, %rd11;
ld.global.nc.u32 %r342, [%rd38];

{.reg .f16 low,high;
mov.b32 {low,high},%r78;
cvt.f32.f16 %f43, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r79;
cvt.f32.f16 %f44, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r80;
cvt.f32.f16 %f45, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r81;
cvt.f32.f16 %f46, low;}


add.s64 %rd39, %rd65, %rd10;
ld.global.nc.u8 %rs5, [%rd39];
cvt.u32.u16 %r343, %rs5;
and.b32 %r344, %r343, 240;
and.b32 %r273, %r342, 50529027;

dp4a.s32.s32 %r120, %r273, %r84, %r303;

and.b32 %r345, %r343, 15;
mul.lo.s32 %r346, %r120, %r345;
cvt.rn.f32.s32 %f92, %r346;
fma.rn.ftz.f32 %f93, %f43, %f92, 0f00000000;
shr.u32 %r347, %r344, 4;
shl.b32 %r348, %r347, 8;
or.b32 %r349, %r348, %r347;
shl.b32 %r350, %r349, 16;
or.b32 %r277, %r350, %r349;

dp4a.s32.s32 %r124, %r277, %r84, %r303;

cvt.rn.f32.s32 %f94, %r124;
fma.rn.ftz.f32 %f95, %f43, %f94, 0f00000000;
ld.global.nc.u8 %rs6, [%rd39+2];
cvt.u32.u16 %r351, %rs6;
and.b32 %r352, %r351, 240;
shr.u32 %r353, %r342, 2;
and.b32 %r281, %r353, 50529027;

dp4a.s32.s32 %r128, %r281, %r92, %r303;

and.b32 %r354, %r351, 15;
mul.lo.s32 %r355, %r128, %r354;
cvt.rn.f32.s32 %f96, %r355;
fma.rn.ftz.f32 %f97, %f44, %f96, %f93;
shr.u32 %r356, %r352, 4;
shl.b32 %r357, %r356, 8;
or.b32 %r358, %r357, %r356;
shl.b32 %r359, %r358, 16;
or.b32 %r285, %r359, %r358;

dp4a.s32.s32 %r132, %r285, %r92, %r303;

cvt.rn.f32.s32 %f98, %r132;
fma.rn.ftz.f32 %f99, %f44, %f98, %f95;
ld.global.nc.u8 %rs7, [%rd39+4];
cvt.u32.u16 %r360, %rs7;
and.b32 %r361, %r360, 240;
shr.u32 %r362, %r342, 4;
and.b32 %r289, %r362, 50529027;

dp4a.s32.s32 %r136, %r289, %r100, %r303;

and.b32 %r363, %r360, 15;
mul.lo.s32 %r364, %r136, %r363;
cvt.rn.f32.s32 %f100, %r364;
fma.rn.ftz.f32 %f101, %f45, %f100, %f97;
shr.u32 %r365, %r361, 4;
shl.b32 %r366, %r365, 8;
or.b32 %r367, %r366, %r365;
shl.b32 %r368, %r367, 16;
or.b32 %r293, %r368, %r367;

dp4a.s32.s32 %r140, %r293, %r100, %r303;

cvt.rn.f32.s32 %f102, %r140;
fma.rn.ftz.f32 %f103, %f45, %f102, %f99;
ld.global.nc.u8 %rs8, [%rd39+6];
cvt.u32.u16 %r369, %rs8;
and.b32 %r370, %r369, 240;
shr.u32 %r371, %r342, 6;
and.b32 %r297, %r371, 50529027;

dp4a.s32.s32 %r144, %r297, %r108, %r303;

and.b32 %r372, %r369, 15;
mul.lo.s32 %r373, %r144, %r372;
cvt.rn.f32.s32 %f104, %r373;
fma.rn.ftz.f32 %f105, %f46, %f104, %f101;
shr.u32 %r374, %r370, 4;
shl.b32 %r375, %r374, 8;
or.b32 %r376, %r375, %r374;
shl.b32 %r377, %r376, 16;
or.b32 %r301, %r377, %r376;

dp4a.s32.s32 %r148, %r301, %r108, %r303;

cvt.rn.f32.s32 %f106, %r148;
fma.rn.ftz.f32 %f107, %f46, %f106, %f103;
add.s64 %rd40, %rd65, %rd6;
ld.global.nc.u32 %r152, [%rd40+80];

{.reg .f16 low,high;
mov.b32 {low,high},%r152;
cvt.f32.f16 %f47, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r152;
cvt.f32.f16 %f48, high;}


mul.ftz.f32 %f108, %f105, %f47;
mul.ftz.f32 %f109, %f107, %f48;
sub.ftz.f32 %f110, %f108, %f109;
add.ftz.f32 %f287, %f110, %f287;
cvt.u64.u32 %rd41, %r509;
add.s64 %rd42, %rd2, %rd41;
add.s64 %rd43, %rd42, %rd5;
mul.lo.s64 %rd44, %rd43, 36;
add.s64 %rd45, %rd12, %rd44;
add.s64 %rd46, %rd45, %rd34;
ld.global.nc.u32 %r160, [%rd46+4];
ld.global.nc.u32 %r154, [%rd45];

{.reg .f16 low,high;
mov.b32 {low,high},%r154;
cvt.f32.f16 %f49, low;}


ld.global.nc.u32 %r168, [%rd46+40];
ld.global.nc.u32 %r155, [%rd45+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r155;
cvt.f32.f16 %f50, low;}


ld.global.nc.u32 %r176, [%rd46+76];
ld.global.nc.u32 %r156, [%rd45+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r156;
cvt.f32.f16 %f51, low;}


ld.global.nc.u32 %r184, [%rd46+112];
ld.global.nc.u32 %r157, [%rd45+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r157;
cvt.f32.f16 %f52, low;}



dp4a.s32.s32 %r158, %r235, %r160, %r303;

mul.lo.s32 %r378, %r158, %r309;
cvt.rn.f32.s32 %f111, %r378;
fma.rn.ftz.f32 %f112, %f49, %f111, 0f00000000;

dp4a.s32.s32 %r162, %r239, %r160, %r303;

cvt.rn.f32.s32 %f113, %r162;
fma.rn.ftz.f32 %f114, %f49, %f113, 0f00000000;

dp4a.s32.s32 %r166, %r243, %r168, %r303;

mul.lo.s32 %r379, %r166, %r318;
cvt.rn.f32.s32 %f115, %r379;
fma.rn.ftz.f32 %f116, %f50, %f115, %f112;

dp4a.s32.s32 %r170, %r247, %r168, %r303;

cvt.rn.f32.s32 %f117, %r170;
fma.rn.ftz.f32 %f118, %f50, %f117, %f114;

dp4a.s32.s32 %r174, %r251, %r176, %r303;

mul.lo.s32 %r380, %r174, %r327;
cvt.rn.f32.s32 %f119, %r380;
fma.rn.ftz.f32 %f120, %f51, %f119, %f116;

dp4a.s32.s32 %r178, %r255, %r176, %r303;

cvt.rn.f32.s32 %f121, %r178;
fma.rn.ftz.f32 %f122, %f51, %f121, %f118;

dp4a.s32.s32 %r182, %r259, %r184, %r303;

mul.lo.s32 %r381, %r182, %r336;
cvt.rn.f32.s32 %f123, %r381;
fma.rn.ftz.f32 %f124, %f52, %f123, %f120;

dp4a.s32.s32 %r186, %r263, %r184, %r303;

cvt.rn.f32.s32 %f125, %r186;
fma.rn.ftz.f32 %f126, %f52, %f125, %f122;

{.reg .f16 low,high;
mov.b32 {low,high},%r114;
cvt.f32.f16 %f53, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r114;
cvt.f32.f16 %f54, high;}


mul.ftz.f32 %f127, %f124, %f53;
mul.ftz.f32 %f128, %f126, %f54;
sub.ftz.f32 %f129, %f127, %f128;
add.ftz.f32 %f286, %f129, %f286;

{.reg .f16 low,high;
mov.b32 {low,high},%r154;
cvt.f32.f16 %f55, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r155;
cvt.f32.f16 %f56, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r156;
cvt.f32.f16 %f57, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r157;
cvt.f32.f16 %f58, low;}



dp4a.s32.s32 %r196, %r273, %r160, %r303;

mul.lo.s32 %r382, %r196, %r345;
cvt.rn.f32.s32 %f130, %r382;
fma.rn.ftz.f32 %f131, %f55, %f130, 0f00000000;

dp4a.s32.s32 %r200, %r277, %r160, %r303;

cvt.rn.f32.s32 %f132, %r200;
fma.rn.ftz.f32 %f133, %f55, %f132, 0f00000000;

dp4a.s32.s32 %r204, %r281, %r168, %r303;

mul.lo.s32 %r383, %r204, %r354;
cvt.rn.f32.s32 %f134, %r383;
fma.rn.ftz.f32 %f135, %f56, %f134, %f131;

dp4a.s32.s32 %r208, %r285, %r168, %r303;

cvt.rn.f32.s32 %f136, %r208;
fma.rn.ftz.f32 %f137, %f56, %f136, %f133;

dp4a.s32.s32 %r212, %r289, %r176, %r303;

mul.lo.s32 %r384, %r212, %r363;
cvt.rn.f32.s32 %f138, %r384;
fma.rn.ftz.f32 %f139, %f57, %f138, %f135;

dp4a.s32.s32 %r216, %r293, %r176, %r303;

cvt.rn.f32.s32 %f140, %r216;
fma.rn.ftz.f32 %f141, %f57, %f140, %f137;

dp4a.s32.s32 %r220, %r297, %r184, %r303;

mul.lo.s32 %r385, %r220, %r372;
cvt.rn.f32.s32 %f142, %r385;
fma.rn.ftz.f32 %f143, %f58, %f142, %f139;

dp4a.s32.s32 %r224, %r301, %r184, %r303;

cvt.rn.f32.s32 %f144, %r224;
fma.rn.ftz.f32 %f145, %f58, %f144, %f141;

{.reg .f16 low,high;
mov.b32 {low,high},%r152;
cvt.f32.f16 %f59, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r152;
cvt.f32.f16 %f60, high;}


mul.ftz.f32 %f146, %f143, %f59;
mul.ftz.f32 %f147, %f145, %f60;
sub.ftz.f32 %f148, %f146, %f147;
add.ftz.f32 %f285, %f148, %f285;
cvt.u64.u32 %rd47, %r510;
add.s64 %rd48, %rd2, %rd47;
add.s64 %rd49, %rd48, %rd5;
mul.lo.s64 %rd50, %rd49, 36;
add.s64 %rd51, %rd12, %rd50;
add.s64 %rd52, %rd51, %rd34;
ld.global.nc.u32 %r236, [%rd52+4];
ld.global.nc.u32 %r230, [%rd51];

{.reg .f16 low,high;
mov.b32 {low,high},%r230;
cvt.f32.f16 %f61, low;}


ld.global.nc.u32 %r244, [%rd52+40];
ld.global.nc.u32 %r231, [%rd51+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r231;
cvt.f32.f16 %f62, low;}


ld.global.nc.u32 %r252, [%rd52+76];
ld.global.nc.u32 %r232, [%rd51+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r232;
cvt.f32.f16 %f63, low;}


ld.global.nc.u32 %r260, [%rd52+112];
ld.global.nc.u32 %r233, [%rd51+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r233;
cvt.f32.f16 %f64, low;}



dp4a.s32.s32 %r234, %r235, %r236, %r303;

mul.lo.s32 %r386, %r234, %r309;
cvt.rn.f32.s32 %f149, %r386;
fma.rn.ftz.f32 %f150, %f61, %f149, 0f00000000;

dp4a.s32.s32 %r238, %r239, %r236, %r303;

cvt.rn.f32.s32 %f151, %r238;
fma.rn.ftz.f32 %f152, %f61, %f151, 0f00000000;

dp4a.s32.s32 %r242, %r243, %r244, %r303;

mul.lo.s32 %r387, %r242, %r318;
cvt.rn.f32.s32 %f153, %r387;
fma.rn.ftz.f32 %f154, %f62, %f153, %f150;

dp4a.s32.s32 %r246, %r247, %r244, %r303;

cvt.rn.f32.s32 %f155, %r246;
fma.rn.ftz.f32 %f156, %f62, %f155, %f152;

dp4a.s32.s32 %r250, %r251, %r252, %r303;

mul.lo.s32 %r388, %r250, %r327;
cvt.rn.f32.s32 %f157, %r388;
fma.rn.ftz.f32 %f158, %f63, %f157, %f154;

dp4a.s32.s32 %r254, %r255, %r252, %r303;

cvt.rn.f32.s32 %f159, %r254;
fma.rn.ftz.f32 %f160, %f63, %f159, %f156;

dp4a.s32.s32 %r258, %r259, %r260, %r303;

mul.lo.s32 %r389, %r258, %r336;
cvt.rn.f32.s32 %f161, %r389;
fma.rn.ftz.f32 %f162, %f64, %f161, %f158;

dp4a.s32.s32 %r262, %r263, %r260, %r303;

cvt.rn.f32.s32 %f163, %r262;
fma.rn.ftz.f32 %f164, %f64, %f163, %f160;

{.reg .f16 low,high;
mov.b32 {low,high},%r114;
cvt.f32.f16 %f65, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r114;
cvt.f32.f16 %f66, high;}


mul.ftz.f32 %f165, %f162, %f65;
mul.ftz.f32 %f166, %f164, %f66;
sub.ftz.f32 %f167, %f165, %f166;
add.ftz.f32 %f284, %f167, %f284;

{.reg .f16 low,high;
mov.b32 {low,high},%r230;
cvt.f32.f16 %f67, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r231;
cvt.f32.f16 %f68, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r232;
cvt.f32.f16 %f69, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r233;
cvt.f32.f16 %f70, low;}



dp4a.s32.s32 %r272, %r273, %r236, %r303;

mul.lo.s32 %r390, %r272, %r345;
cvt.rn.f32.s32 %f168, %r390;
fma.rn.ftz.f32 %f169, %f67, %f168, 0f00000000;

dp4a.s32.s32 %r276, %r277, %r236, %r303;

cvt.rn.f32.s32 %f170, %r276;
fma.rn.ftz.f32 %f171, %f67, %f170, 0f00000000;

dp4a.s32.s32 %r280, %r281, %r244, %r303;

mul.lo.s32 %r391, %r280, %r354;
cvt.rn.f32.s32 %f172, %r391;
fma.rn.ftz.f32 %f173, %f68, %f172, %f169;

dp4a.s32.s32 %r284, %r285, %r244, %r303;

cvt.rn.f32.s32 %f174, %r284;
fma.rn.ftz.f32 %f175, %f68, %f174, %f171;

dp4a.s32.s32 %r288, %r289, %r252, %r303;

mul.lo.s32 %r392, %r288, %r363;
cvt.rn.f32.s32 %f176, %r392;
fma.rn.ftz.f32 %f177, %f69, %f176, %f173;

dp4a.s32.s32 %r292, %r293, %r252, %r303;

cvt.rn.f32.s32 %f178, %r292;
fma.rn.ftz.f32 %f179, %f69, %f178, %f175;

dp4a.s32.s32 %r296, %r297, %r260, %r303;

mul.lo.s32 %r393, %r296, %r372;
cvt.rn.f32.s32 %f180, %r393;
fma.rn.ftz.f32 %f181, %f70, %f180, %f177;

dp4a.s32.s32 %r300, %r301, %r260, %r303;

cvt.rn.f32.s32 %f182, %r300;
fma.rn.ftz.f32 %f183, %f70, %f182, %f179;

{.reg .f16 low,high;
mov.b32 {low,high},%r152;
cvt.f32.f16 %f71, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r152;
cvt.f32.f16 %f72, high;}


mul.ftz.f32 %f184, %f181, %f71;
mul.ftz.f32 %f185, %f183, %f72;
sub.ftz.f32 %f186, %f184, %f185;
add.ftz.f32 %f283, %f186, %f283;
add.s32 %r510, %r510, 64;
add.s32 %r509, %r509, 64;
add.s32 %r508, %r508, 64;
add.s64 %rd65, %rd65, 672;
add.s32 %r511, %r511, 8;
setp.lt.s32 %p4, %r511, %r7;
@%p4 bra $L__BB134_2;

$L__BB134_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB134_5;

mov.u32 %r394, _ZZ13mul_mat_vec_qIL9ggml_type10ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r395, %r1, 768, %r394;
shl.b32 %r396, %r2, 2;
add.s32 %r397, %r395, %r396;
st.shared.f32 [%r397+-768], %f288;
st.shared.f32 [%r397+-640], %f287;
st.shared.f32 [%r397+-512], %f286;
st.shared.f32 [%r397+-384], %f285;
st.shared.f32 [%r397+-256], %f284;
st.shared.f32 [%r397+-128], %f283;

$L__BB134_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB134_18;

mov.u32 %r398, %ctaid.z;
mad.lo.s32 %r399, %r5, %r29, %r4;
mad.lo.s32 %r400, %r398, %r35, %r399;
cvt.u64.u32 %rd15, %r400;
shl.b32 %r401, %r2, 2;
mov.u32 %r402, _ZZ13mul_mat_vec_qIL9ggml_type10ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r19, %r402, %r401;
mov.u32 %r403, 2;
ld.shared.f32 %f187, [%r19];
add.ftz.f32 %f188, %f187, %f288;
ld.shared.f32 %f189, [%r19+768];
add.ftz.f32 %f190, %f189, %f188;
ld.shared.f32 %f191, [%r19+1536];
add.ftz.f32 %f192, %f191, %f190;
mov.b32 %r404, %f192;
mov.u32 %r405, 31;
mov.u32 %r406, 16;
mov.u32 %r407, -1;
shfl.sync.bfly.b32 %r408|%p7, %r404, %r406, %r405, %r407;
mov.b32 %f193, %r408;
add.ftz.f32 %f194, %f192, %f193;
mov.b32 %r409, %f194;
mov.u32 %r410, 8;
shfl.sync.bfly.b32 %r411|%p8, %r409, %r410, %r405, %r407;
mov.b32 %f195, %r411;
add.ftz.f32 %f196, %f194, %f195;
mov.b32 %r412, %f196;
mov.u32 %r413, 4;
shfl.sync.bfly.b32 %r414|%p9, %r412, %r413, %r405, %r407;
mov.b32 %f197, %r414;
add.ftz.f32 %f198, %f196, %f197;
mov.b32 %r415, %f198;
shfl.sync.bfly.b32 %r416|%p10, %r415, %r403, %r405, %r407;
mov.b32 %f199, %r416;
add.ftz.f32 %f200, %f198, %f199;
mov.b32 %r417, %f200;
mov.u32 %r418, 1;
shfl.sync.bfly.b32 %r419|%p11, %r417, %r418, %r405, %r407;
mov.b32 %f201, %r419;
add.ftz.f32 %f19, %f200, %f201;
setp.lt.u32 %p12, %r4, %r23;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
mul.wide.u32 %rd53, %r400, 4;
add.s64 %rd16, %rd1, %rd53;
not.pred %p14, %p1;
@%p14 bra $L__BB134_8;

st.global.f32 [%rd16], %f19;

$L__BB134_8:
ld.shared.f32 %f202, [%r19+128];
add.ftz.f32 %f203, %f202, %f287;
ld.shared.f32 %f204, [%r19+896];
add.ftz.f32 %f205, %f204, %f203;
ld.shared.f32 %f206, [%r19+1664];
add.ftz.f32 %f207, %f206, %f205;
mov.b32 %r420, %f207;
shfl.sync.bfly.b32 %r424|%p15, %r420, %r406, %r405, %r407;
mov.b32 %f208, %r424;
add.ftz.f32 %f209, %f207, %f208;
mov.b32 %r425, %f209;
shfl.sync.bfly.b32 %r427|%p16, %r425, %r410, %r405, %r407;
mov.b32 %f210, %r427;
add.ftz.f32 %f211, %f209, %f210;
mov.b32 %r428, %f211;
shfl.sync.bfly.b32 %r430|%p17, %r428, %r413, %r405, %r407;
mov.b32 %f212, %r430;
add.ftz.f32 %f213, %f211, %f212;
mov.b32 %r431, %f213;
shfl.sync.bfly.b32 %r433|%p18, %r431, %r403, %r405, %r407;
mov.b32 %f214, %r433;
add.ftz.f32 %f215, %f213, %f214;
mov.b32 %r434, %f215;
shfl.sync.bfly.b32 %r436|%p19, %r434, %r418, %r405, %r407;
mov.b32 %f216, %r436;
add.ftz.f32 %f20, %f215, %f216;
add.s32 %r437, %r4, 1;
setp.lt.u32 %p20, %r437, %r23;
setp.eq.s32 %p21, %r2, 1;
and.pred %p2, %p21, %p20;
not.pred %p22, %p2;
@%p22 bra $L__BB134_10;

st.global.f32 [%rd16+4], %f20;

$L__BB134_10:
ld.shared.f32 %f217, [%r19+256];
add.ftz.f32 %f218, %f217, %f286;
ld.shared.f32 %f219, [%r19+1024];
add.ftz.f32 %f220, %f219, %f218;
ld.shared.f32 %f221, [%r19+1792];
add.ftz.f32 %f222, %f221, %f220;
mov.b32 %r438, %f222;
mov.u32 %r439, 31;
mov.u32 %r440, 16;
mov.u32 %r441, -1;
shfl.sync.bfly.b32 %r442|%p23, %r438, %r440, %r439, %r441;
mov.b32 %f223, %r442;
add.ftz.f32 %f224, %f222, %f223;
mov.b32 %r443, %f224;
mov.u32 %r444, 8;
shfl.sync.bfly.b32 %r445|%p24, %r443, %r444, %r439, %r441;
mov.b32 %f225, %r445;
add.ftz.f32 %f226, %f224, %f225;
mov.b32 %r446, %f226;
mov.u32 %r447, 4;
shfl.sync.bfly.b32 %r448|%p25, %r446, %r447, %r439, %r441;
mov.b32 %f227, %r448;
add.ftz.f32 %f228, %f226, %f227;
mov.b32 %r449, %f228;
mov.u32 %r450, 2;
shfl.sync.bfly.b32 %r451|%p26, %r449, %r450, %r439, %r441;
mov.b32 %f229, %r451;
add.ftz.f32 %f230, %f228, %f229;
mov.b32 %r452, %f230;
mov.u32 %r453, 1;
shfl.sync.bfly.b32 %r454|%p27, %r452, %r453, %r439, %r441;
mov.b32 %f231, %r454;
add.ftz.f32 %f21, %f230, %f231;
@%p14 bra $L__BB134_12;

cvt.u64.u32 %rd54, %r23;
add.s64 %rd55, %rd54, %rd15;
shl.b64 %rd56, %rd55, 2;
add.s64 %rd57, %rd1, %rd56;
st.global.f32 [%rd57], %f21;

$L__BB134_12:
ld.shared.f32 %f232, [%r19+384];
add.ftz.f32 %f233, %f232, %f285;
ld.shared.f32 %f234, [%r19+1152];
add.ftz.f32 %f235, %f234, %f233;
ld.shared.f32 %f236, [%r19+1920];
add.ftz.f32 %f237, %f236, %f235;
mov.b32 %r455, %f237;
shfl.sync.bfly.b32 %r459|%p29, %r455, %r440, %r439, %r441;
mov.b32 %f238, %r459;
add.ftz.f32 %f239, %f237, %f238;
mov.b32 %r460, %f239;
shfl.sync.bfly.b32 %r462|%p30, %r460, %r444, %r439, %r441;
mov.b32 %f240, %r462;
add.ftz.f32 %f241, %f239, %f240;
mov.b32 %r463, %f241;
shfl.sync.bfly.b32 %r465|%p31, %r463, %r447, %r439, %r441;
mov.b32 %f242, %r465;
add.ftz.f32 %f243, %f241, %f242;
mov.b32 %r466, %f243;
shfl.sync.bfly.b32 %r468|%p32, %r466, %r450, %r439, %r441;
mov.b32 %f244, %r468;
add.ftz.f32 %f245, %f243, %f244;
mov.b32 %r469, %f245;
shfl.sync.bfly.b32 %r471|%p33, %r469, %r453, %r439, %r441;
mov.b32 %f246, %r471;
add.ftz.f32 %f22, %f245, %f246;
@%p22 bra $L__BB134_14;

add.s32 %r472, %r23, 1;
cvt.u64.u32 %rd58, %r472;
add.s64 %rd59, %rd58, %rd15;
shl.b64 %rd60, %rd59, 2;
add.s64 %rd61, %rd1, %rd60;
st.global.f32 [%rd61], %f22;

$L__BB134_14:
shl.b32 %r473, %r23, 1;
mov.u32 %r474, 1;
ld.shared.f32 %f247, [%r19+512];
add.ftz.f32 %f248, %f247, %f284;
ld.shared.f32 %f249, [%r19+1280];
add.ftz.f32 %f250, %f249, %f248;
ld.shared.f32 %f251, [%r19+2048];
add.ftz.f32 %f252, %f251, %f250;
mov.b32 %r475, %f252;
mov.u32 %r476, 31;
mov.u32 %r477, 16;
mov.u32 %r478, -1;
shfl.sync.bfly.b32 %r479|%p35, %r475, %r477, %r476, %r478;
mov.b32 %f253, %r479;
add.ftz.f32 %f254, %f252, %f253;
mov.b32 %r480, %f254;
mov.u32 %r481, 8;
shfl.sync.bfly.b32 %r482|%p36, %r480, %r481, %r476, %r478;
mov.b32 %f255, %r482;
add.ftz.f32 %f256, %f254, %f255;
mov.b32 %r483, %f256;
mov.u32 %r484, 4;
shfl.sync.bfly.b32 %r485|%p37, %r483, %r484, %r476, %r478;
mov.b32 %f257, %r485;
add.ftz.f32 %f258, %f256, %f257;
mov.b32 %r486, %f258;
mov.u32 %r487, 2;
shfl.sync.bfly.b32 %r488|%p38, %r486, %r487, %r476, %r478;
mov.b32 %f259, %r488;
add.ftz.f32 %f260, %f258, %f259;
mov.b32 %r489, %f260;
shfl.sync.bfly.b32 %r490|%p39, %r489, %r474, %r476, %r478;
mov.b32 %f261, %r490;
add.ftz.f32 %f23, %f260, %f261;
cvt.u64.u32 %rd62, %r473;
add.s64 %rd63, %rd62, %rd15;
shl.b64 %rd64, %rd63, 2;
add.s64 %rd17, %rd1, %rd64;
@%p14 bra $L__BB134_16;

st.global.f32 [%rd17], %f23;

$L__BB134_16:
ld.shared.f32 %f262, [%r19+640];
add.ftz.f32 %f263, %f262, %f283;
ld.shared.f32 %f264, [%r19+1408];
add.ftz.f32 %f265, %f264, %f263;
ld.shared.f32 %f266, [%r19+2176];
add.ftz.f32 %f267, %f266, %f265;
mov.b32 %r491, %f267;
shfl.sync.bfly.b32 %r495|%p41, %r491, %r477, %r476, %r478;
mov.b32 %f268, %r495;
add.ftz.f32 %f269, %f267, %f268;
mov.b32 %r496, %f269;
shfl.sync.bfly.b32 %r498|%p42, %r496, %r481, %r476, %r478;
mov.b32 %f270, %r498;
add.ftz.f32 %f271, %f269, %f270;
mov.b32 %r499, %f271;
shfl.sync.bfly.b32 %r501|%p43, %r499, %r484, %r476, %r478;
mov.b32 %f272, %r501;
add.ftz.f32 %f273, %f271, %f272;
mov.b32 %r502, %f273;
shfl.sync.bfly.b32 %r504|%p44, %r502, %r487, %r476, %r478;
mov.b32 %f274, %r504;
add.ftz.f32 %f275, %f273, %f274;
mov.b32 %r505, %f275;
shfl.sync.bfly.b32 %r507|%p45, %r505, %r474, %r476, %r478;
mov.b32 %f276, %r507;
add.ftz.f32 %f24, %f275, %f276;
@%p22 bra $L__BB134_18;

st.global.f32 [%rd17+4], %f24;

$L__BB134_18:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type10ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type10ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type10ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type10ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type10ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type10ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type10ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type10ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type10ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<59>;
.reg .b16 %rs<9>;
.reg .f32 %f<385>;
.reg .b32 %r<637>;
.reg .b64 %rd<80>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type10ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[3072];

ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type10ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type10ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd20, [_Z13mul_mat_vec_qIL9ggml_type10ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type10ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type10ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type10ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type10ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type10ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type10ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type10ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type10ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type10ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type10ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type10ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type10ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type10ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type10ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
cvta.to.global.u64 %rd1, %rd20;
mov.u32 %r1, %tid.y;
shl.b32 %r41, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r41, %r2;
mov.u32 %r42, %ctaid.x;
shl.b32 %r4, %r42, 1;
mov.u32 %r5, %ctaid.y;
shr.s32 %r43, %r3, 31;
shr.u32 %r44, %r43, 28;
add.s32 %r45, %r3, %r44;
shr.s32 %r636, %r45, 4;
shr.u32 %r7, %r40, 8;
setp.ge.s32 %p3, %r636, %r7;
mov.f32 %f377, 0f00000000;
mov.f32 %f378, %f377;
mov.f32 %f379, %f377;
mov.f32 %f380, %f377;
mov.f32 %f381, %f377;
mov.f32 %f382, %f377;
mov.f32 %f383, %f377;
mov.f32 %f384, %f377;
@%p3 bra $L__BB135_3;

mov.u32 %r46, %ctaid.z;
mul.lo.s32 %r47, %r5, %r32;
cvt.u64.u32 %rd21, %r47;
mul.lo.s32 %r48, %r46, %r38;
cvt.u64.u32 %rd22, %r48;
add.s64 %rd2, %rd22, %rd21;
mul.hi.u32 %r49, %r46, %r34;
add.s32 %r50, %r49, %r46;
shr.u32 %r51, %r50, %r35;
mul.lo.s32 %r52, %r51, %r37;
mul.hi.u32 %r53, %r5, %r28;
add.s32 %r54, %r53, %r5;
shr.u32 %r55, %r54, %r29;
mul.lo.s32 %r56, %r55, %r31;
cvta.to.global.u64 %rd79, %rd18;
and.b32 %r60, %r45, -16;
sub.s32 %r61, %r3, %r60;
shr.s32 %r62, %r61, 31;
shr.u32 %r63, %r62, 29;
add.s32 %r64, %r61, %r63;
shr.s32 %r65, %r64, 3;
shl.b32 %r66, %r65, 2;
and.b32 %r67, %r64, -8;
sub.s32 %r68, %r61, %r67;
shr.s32 %r69, %r68, 31;
shr.u32 %r70, %r69, 30;
add.s32 %r71, %r68, %r70;
shr.s32 %r72, %r71, 2;
add.s32 %r73, %r67, %r72;
cvt.s64.s32 %rd23, %r73;
cvt.s64.s32 %rd4, %r68;
add.s32 %r74, %r4, 1;
cvt.s64.s32 %rd5, %r66;
shl.b32 %r75, %r26, 1;
shl.b32 %r632, %r636, 3;
add.s32 %r635, %r632, %r75;
mad.lo.s32 %r634, %r26, 3, %r632;
add.s32 %r633, %r26, %r632;
add.s32 %r76, %r636, %r52;
add.s32 %r77, %r76, %r56;
mad.lo.s32 %r78, %r74, %r25, %r77;
mul.wide.s32 %rd6, %r78, 84;
mad.lo.s32 %r79, %r4, %r25, %r636;
add.s32 %r80, %r79, %r52;
add.s32 %r81, %r80, %r56;
mul.wide.s32 %rd7, %r81, 84;
mul.wide.s32 %rd24, %r61, 4;
add.s64 %rd25, %rd7, %rd24;
add.s64 %rd8, %rd25, 16;
add.s64 %rd26, %rd7, %rd23;
add.s64 %rd9, %rd26, 6;
add.s64 %rd10, %rd6, %rd23;
add.s64 %rd27, %rd6, %rd24;
add.s64 %rd11, %rd27, 16;
cvta.to.global.u64 %rd12, %rd19;
mov.f32 %f377, 0f00000000;
shl.b64 %rd34, %rd4, 2;

$L__BB135_2:
cvt.u64.u32 %rd28, %r632;
add.s64 %rd29, %rd2, %rd28;
add.s64 %rd30, %rd79, %rd8;
ld.global.nc.u32 %r386, [%rd30];
add.s64 %rd31, %rd29, %rd5;
mul.lo.s64 %rd32, %rd31, 36;
add.s64 %rd33, %rd12, %rd32;
add.s64 %rd35, %rd33, %rd34;
ld.global.nc.u32 %r88, [%rd35+4];
ld.global.nc.u32 %r82, [%rd33];

{.reg .f16 low,high;
mov.b32 {low,high},%r82;
cvt.f32.f16 %f49, low;}


ld.global.nc.u32 %r96, [%rd35+40];
ld.global.nc.u32 %r83, [%rd33+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r83;
cvt.f32.f16 %f50, low;}


ld.global.nc.u32 %r104, [%rd35+76];
ld.global.nc.u32 %r84, [%rd33+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r84;
cvt.f32.f16 %f51, low;}


ld.global.nc.u32 %r112, [%rd35+112];
ld.global.nc.u32 %r85, [%rd33+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r85;
cvt.f32.f16 %f52, low;}


add.s64 %rd36, %rd79, %rd9;
ld.global.nc.u8 %rs1, [%rd36+-6];
cvt.u32.u16 %r387, %rs1;
and.b32 %r388, %r387, 240;
and.b32 %r315, %r386, 50529027;
mov.u32 %r383, 0;

dp4a.s32.s32 %r86, %r315, %r88, %r383;

and.b32 %r389, %r387, 15;
mul.lo.s32 %r390, %r86, %r389;
cvt.rn.f32.s32 %f97, %r390;
fma.rn.ftz.f32 %f98, %f49, %f97, 0f00000000;
shr.u32 %r391, %r388, 4;
shl.b32 %r392, %r391, 8;
or.b32 %r393, %r392, %r391;
shl.b32 %r394, %r393, 16;
or.b32 %r319, %r394, %r393;

dp4a.s32.s32 %r90, %r319, %r88, %r383;

cvt.rn.f32.s32 %f99, %r90;
fma.rn.ftz.f32 %f100, %f49, %f99, 0f00000000;
ld.global.nc.u8 %rs2, [%rd36+-4];
cvt.u32.u16 %r395, %rs2;
and.b32 %r396, %r395, 240;
shr.u32 %r397, %r386, 2;
and.b32 %r323, %r397, 50529027;

dp4a.s32.s32 %r94, %r323, %r96, %r383;

and.b32 %r398, %r395, 15;
mul.lo.s32 %r399, %r94, %r398;
cvt.rn.f32.s32 %f101, %r399;
fma.rn.ftz.f32 %f102, %f50, %f101, %f98;
shr.u32 %r400, %r396, 4;
shl.b32 %r401, %r400, 8;
or.b32 %r402, %r401, %r400;
shl.b32 %r403, %r402, 16;
or.b32 %r327, %r403, %r402;

dp4a.s32.s32 %r98, %r327, %r96, %r383;

cvt.rn.f32.s32 %f103, %r98;
fma.rn.ftz.f32 %f104, %f50, %f103, %f100;
ld.global.nc.u8 %rs3, [%rd36+-2];
cvt.u32.u16 %r404, %rs3;
and.b32 %r405, %r404, 240;
shr.u32 %r406, %r386, 4;
and.b32 %r331, %r406, 50529027;

dp4a.s32.s32 %r102, %r331, %r104, %r383;

and.b32 %r407, %r404, 15;
mul.lo.s32 %r408, %r102, %r407;
cvt.rn.f32.s32 %f105, %r408;
fma.rn.ftz.f32 %f106, %f51, %f105, %f102;
shr.u32 %r409, %r405, 4;
shl.b32 %r410, %r409, 8;
or.b32 %r411, %r410, %r409;
shl.b32 %r412, %r411, 16;
or.b32 %r335, %r412, %r411;

dp4a.s32.s32 %r106, %r335, %r104, %r383;

cvt.rn.f32.s32 %f107, %r106;
fma.rn.ftz.f32 %f108, %f51, %f107, %f104;
ld.global.nc.u8 %rs4, [%rd36];
cvt.u32.u16 %r413, %rs4;
and.b32 %r414, %r413, 240;
shr.u32 %r415, %r386, 6;
and.b32 %r339, %r415, 50529027;

dp4a.s32.s32 %r110, %r339, %r112, %r383;

and.b32 %r416, %r413, 15;
mul.lo.s32 %r417, %r110, %r416;
cvt.rn.f32.s32 %f109, %r417;
fma.rn.ftz.f32 %f110, %f52, %f109, %f106;
shr.u32 %r418, %r414, 4;
shl.b32 %r419, %r418, 8;
or.b32 %r420, %r419, %r418;
shl.b32 %r421, %r420, 16;
or.b32 %r343, %r421, %r420;

dp4a.s32.s32 %r114, %r343, %r112, %r383;

cvt.rn.f32.s32 %f111, %r114;
fma.rn.ftz.f32 %f112, %f52, %f111, %f108;
add.s64 %rd37, %rd79, %rd7;
ld.global.nc.u32 %r118, [%rd37+80];

{.reg .f16 low,high;
mov.b32 {low,high},%r118;
cvt.f32.f16 %f53, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r118;
cvt.f32.f16 %f54, high;}


mul.ftz.f32 %f113, %f110, %f53;
mul.ftz.f32 %f114, %f112, %f54;
sub.ftz.f32 %f115, %f113, %f114;
add.ftz.f32 %f384, %f115, %f384;
add.s64 %rd38, %rd79, %rd11;
ld.global.nc.u32 %r422, [%rd38];

{.reg .f16 low,high;
mov.b32 {low,high},%r82;
cvt.f32.f16 %f55, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r83;
cvt.f32.f16 %f56, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r84;
cvt.f32.f16 %f57, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r85;
cvt.f32.f16 %f58, low;}


add.s64 %rd39, %rd79, %rd10;
ld.global.nc.u8 %rs5, [%rd39];
cvt.u32.u16 %r423, %rs5;
and.b32 %r424, %r423, 240;
and.b32 %r353, %r422, 50529027;

dp4a.s32.s32 %r124, %r353, %r88, %r383;

and.b32 %r425, %r423, 15;
mul.lo.s32 %r426, %r124, %r425;
cvt.rn.f32.s32 %f116, %r426;
fma.rn.ftz.f32 %f117, %f55, %f116, 0f00000000;
shr.u32 %r427, %r424, 4;
shl.b32 %r428, %r427, 8;
or.b32 %r429, %r428, %r427;
shl.b32 %r430, %r429, 16;
or.b32 %r357, %r430, %r429;

dp4a.s32.s32 %r128, %r357, %r88, %r383;

cvt.rn.f32.s32 %f118, %r128;
fma.rn.ftz.f32 %f119, %f55, %f118, 0f00000000;
ld.global.nc.u8 %rs6, [%rd39+2];
cvt.u32.u16 %r431, %rs6;
and.b32 %r432, %r431, 240;
shr.u32 %r433, %r422, 2;
and.b32 %r361, %r433, 50529027;

dp4a.s32.s32 %r132, %r361, %r96, %r383;

and.b32 %r434, %r431, 15;
mul.lo.s32 %r435, %r132, %r434;
cvt.rn.f32.s32 %f120, %r435;
fma.rn.ftz.f32 %f121, %f56, %f120, %f117;
shr.u32 %r436, %r432, 4;
shl.b32 %r437, %r436, 8;
or.b32 %r438, %r437, %r436;
shl.b32 %r439, %r438, 16;
or.b32 %r365, %r439, %r438;

dp4a.s32.s32 %r136, %r365, %r96, %r383;

cvt.rn.f32.s32 %f122, %r136;
fma.rn.ftz.f32 %f123, %f56, %f122, %f119;
ld.global.nc.u8 %rs7, [%rd39+4];
cvt.u32.u16 %r440, %rs7;
and.b32 %r441, %r440, 240;
shr.u32 %r442, %r422, 4;
and.b32 %r369, %r442, 50529027;

dp4a.s32.s32 %r140, %r369, %r104, %r383;

and.b32 %r443, %r440, 15;
mul.lo.s32 %r444, %r140, %r443;
cvt.rn.f32.s32 %f124, %r444;
fma.rn.ftz.f32 %f125, %f57, %f124, %f121;
shr.u32 %r445, %r441, 4;
shl.b32 %r446, %r445, 8;
or.b32 %r447, %r446, %r445;
shl.b32 %r448, %r447, 16;
or.b32 %r373, %r448, %r447;

dp4a.s32.s32 %r144, %r373, %r104, %r383;

cvt.rn.f32.s32 %f126, %r144;
fma.rn.ftz.f32 %f127, %f57, %f126, %f123;
ld.global.nc.u8 %rs8, [%rd39+6];
cvt.u32.u16 %r449, %rs8;
and.b32 %r450, %r449, 240;
shr.u32 %r451, %r422, 6;
and.b32 %r377, %r451, 50529027;

dp4a.s32.s32 %r148, %r377, %r112, %r383;

and.b32 %r452, %r449, 15;
mul.lo.s32 %r453, %r148, %r452;
cvt.rn.f32.s32 %f128, %r453;
fma.rn.ftz.f32 %f129, %f58, %f128, %f125;
shr.u32 %r454, %r450, 4;
shl.b32 %r455, %r454, 8;
or.b32 %r456, %r455, %r454;
shl.b32 %r457, %r456, 16;
or.b32 %r381, %r457, %r456;

dp4a.s32.s32 %r152, %r381, %r112, %r383;

cvt.rn.f32.s32 %f130, %r152;
fma.rn.ftz.f32 %f131, %f58, %f130, %f127;
add.s64 %rd40, %rd79, %rd6;
ld.global.nc.u32 %r156, [%rd40+80];

{.reg .f16 low,high;
mov.b32 {low,high},%r156;
cvt.f32.f16 %f59, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r156;
cvt.f32.f16 %f60, high;}


mul.ftz.f32 %f132, %f129, %f59;
mul.ftz.f32 %f133, %f131, %f60;
sub.ftz.f32 %f134, %f132, %f133;
add.ftz.f32 %f383, %f134, %f383;
cvt.u64.u32 %rd41, %r633;
add.s64 %rd42, %rd2, %rd41;
add.s64 %rd43, %rd42, %rd5;
mul.lo.s64 %rd44, %rd43, 36;
add.s64 %rd45, %rd12, %rd44;
add.s64 %rd46, %rd45, %rd34;
ld.global.nc.u32 %r164, [%rd46+4];
ld.global.nc.u32 %r158, [%rd45];

{.reg .f16 low,high;
mov.b32 {low,high},%r158;
cvt.f32.f16 %f61, low;}


ld.global.nc.u32 %r172, [%rd46+40];
ld.global.nc.u32 %r159, [%rd45+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r159;
cvt.f32.f16 %f62, low;}


ld.global.nc.u32 %r180, [%rd46+76];
ld.global.nc.u32 %r160, [%rd45+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r160;
cvt.f32.f16 %f63, low;}


ld.global.nc.u32 %r188, [%rd46+112];
ld.global.nc.u32 %r161, [%rd45+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r161;
cvt.f32.f16 %f64, low;}



dp4a.s32.s32 %r162, %r315, %r164, %r383;

mul.lo.s32 %r458, %r162, %r389;
cvt.rn.f32.s32 %f135, %r458;
fma.rn.ftz.f32 %f136, %f61, %f135, 0f00000000;

dp4a.s32.s32 %r166, %r319, %r164, %r383;

cvt.rn.f32.s32 %f137, %r166;
fma.rn.ftz.f32 %f138, %f61, %f137, 0f00000000;

dp4a.s32.s32 %r170, %r323, %r172, %r383;

mul.lo.s32 %r459, %r170, %r398;
cvt.rn.f32.s32 %f139, %r459;
fma.rn.ftz.f32 %f140, %f62, %f139, %f136;

dp4a.s32.s32 %r174, %r327, %r172, %r383;

cvt.rn.f32.s32 %f141, %r174;
fma.rn.ftz.f32 %f142, %f62, %f141, %f138;

dp4a.s32.s32 %r178, %r331, %r180, %r383;

mul.lo.s32 %r460, %r178, %r407;
cvt.rn.f32.s32 %f143, %r460;
fma.rn.ftz.f32 %f144, %f63, %f143, %f140;

dp4a.s32.s32 %r182, %r335, %r180, %r383;

cvt.rn.f32.s32 %f145, %r182;
fma.rn.ftz.f32 %f146, %f63, %f145, %f142;

dp4a.s32.s32 %r186, %r339, %r188, %r383;

mul.lo.s32 %r461, %r186, %r416;
cvt.rn.f32.s32 %f147, %r461;
fma.rn.ftz.f32 %f148, %f64, %f147, %f144;

dp4a.s32.s32 %r190, %r343, %r188, %r383;

cvt.rn.f32.s32 %f149, %r190;
fma.rn.ftz.f32 %f150, %f64, %f149, %f146;

{.reg .f16 low,high;
mov.b32 {low,high},%r118;
cvt.f32.f16 %f65, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r118;
cvt.f32.f16 %f66, high;}


mul.ftz.f32 %f151, %f148, %f65;
mul.ftz.f32 %f152, %f150, %f66;
sub.ftz.f32 %f153, %f151, %f152;
add.ftz.f32 %f382, %f153, %f382;

{.reg .f16 low,high;
mov.b32 {low,high},%r158;
cvt.f32.f16 %f67, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r159;
cvt.f32.f16 %f68, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r160;
cvt.f32.f16 %f69, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r161;
cvt.f32.f16 %f70, low;}



dp4a.s32.s32 %r200, %r353, %r164, %r383;

mul.lo.s32 %r462, %r200, %r425;
cvt.rn.f32.s32 %f154, %r462;
fma.rn.ftz.f32 %f155, %f67, %f154, 0f00000000;

dp4a.s32.s32 %r204, %r357, %r164, %r383;

cvt.rn.f32.s32 %f156, %r204;
fma.rn.ftz.f32 %f157, %f67, %f156, 0f00000000;

dp4a.s32.s32 %r208, %r361, %r172, %r383;

mul.lo.s32 %r463, %r208, %r434;
cvt.rn.f32.s32 %f158, %r463;
fma.rn.ftz.f32 %f159, %f68, %f158, %f155;

dp4a.s32.s32 %r212, %r365, %r172, %r383;

cvt.rn.f32.s32 %f160, %r212;
fma.rn.ftz.f32 %f161, %f68, %f160, %f157;

dp4a.s32.s32 %r216, %r369, %r180, %r383;

mul.lo.s32 %r464, %r216, %r443;
cvt.rn.f32.s32 %f162, %r464;
fma.rn.ftz.f32 %f163, %f69, %f162, %f159;

dp4a.s32.s32 %r220, %r373, %r180, %r383;

cvt.rn.f32.s32 %f164, %r220;
fma.rn.ftz.f32 %f165, %f69, %f164, %f161;

dp4a.s32.s32 %r224, %r377, %r188, %r383;

mul.lo.s32 %r465, %r224, %r452;
cvt.rn.f32.s32 %f166, %r465;
fma.rn.ftz.f32 %f167, %f70, %f166, %f163;

dp4a.s32.s32 %r228, %r381, %r188, %r383;

cvt.rn.f32.s32 %f168, %r228;
fma.rn.ftz.f32 %f169, %f70, %f168, %f165;

{.reg .f16 low,high;
mov.b32 {low,high},%r156;
cvt.f32.f16 %f71, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r156;
cvt.f32.f16 %f72, high;}


mul.ftz.f32 %f170, %f167, %f71;
mul.ftz.f32 %f171, %f169, %f72;
sub.ftz.f32 %f172, %f170, %f171;
add.ftz.f32 %f381, %f172, %f381;
cvt.u64.u32 %rd47, %r635;
add.s64 %rd48, %rd2, %rd47;
add.s64 %rd49, %rd48, %rd5;
mul.lo.s64 %rd50, %rd49, 36;
add.s64 %rd51, %rd12, %rd50;
add.s64 %rd52, %rd51, %rd34;
ld.global.nc.u32 %r240, [%rd52+4];
ld.global.nc.u32 %r234, [%rd51];

{.reg .f16 low,high;
mov.b32 {low,high},%r234;
cvt.f32.f16 %f73, low;}


ld.global.nc.u32 %r248, [%rd52+40];
ld.global.nc.u32 %r235, [%rd51+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r235;
cvt.f32.f16 %f74, low;}


ld.global.nc.u32 %r256, [%rd52+76];
ld.global.nc.u32 %r236, [%rd51+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r236;
cvt.f32.f16 %f75, low;}


ld.global.nc.u32 %r264, [%rd52+112];
ld.global.nc.u32 %r237, [%rd51+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r237;
cvt.f32.f16 %f76, low;}



dp4a.s32.s32 %r238, %r315, %r240, %r383;

mul.lo.s32 %r466, %r238, %r389;
cvt.rn.f32.s32 %f173, %r466;
fma.rn.ftz.f32 %f174, %f73, %f173, 0f00000000;

dp4a.s32.s32 %r242, %r319, %r240, %r383;

cvt.rn.f32.s32 %f175, %r242;
fma.rn.ftz.f32 %f176, %f73, %f175, 0f00000000;

dp4a.s32.s32 %r246, %r323, %r248, %r383;

mul.lo.s32 %r467, %r246, %r398;
cvt.rn.f32.s32 %f177, %r467;
fma.rn.ftz.f32 %f178, %f74, %f177, %f174;

dp4a.s32.s32 %r250, %r327, %r248, %r383;

cvt.rn.f32.s32 %f179, %r250;
fma.rn.ftz.f32 %f180, %f74, %f179, %f176;

dp4a.s32.s32 %r254, %r331, %r256, %r383;

mul.lo.s32 %r468, %r254, %r407;
cvt.rn.f32.s32 %f181, %r468;
fma.rn.ftz.f32 %f182, %f75, %f181, %f178;

dp4a.s32.s32 %r258, %r335, %r256, %r383;

cvt.rn.f32.s32 %f183, %r258;
fma.rn.ftz.f32 %f184, %f75, %f183, %f180;

dp4a.s32.s32 %r262, %r339, %r264, %r383;

mul.lo.s32 %r469, %r262, %r416;
cvt.rn.f32.s32 %f185, %r469;
fma.rn.ftz.f32 %f186, %f76, %f185, %f182;

dp4a.s32.s32 %r266, %r343, %r264, %r383;

cvt.rn.f32.s32 %f187, %r266;
fma.rn.ftz.f32 %f188, %f76, %f187, %f184;

{.reg .f16 low,high;
mov.b32 {low,high},%r118;
cvt.f32.f16 %f77, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r118;
cvt.f32.f16 %f78, high;}


mul.ftz.f32 %f189, %f186, %f77;
mul.ftz.f32 %f190, %f188, %f78;
sub.ftz.f32 %f191, %f189, %f190;
add.ftz.f32 %f380, %f191, %f380;

{.reg .f16 low,high;
mov.b32 {low,high},%r234;
cvt.f32.f16 %f79, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r235;
cvt.f32.f16 %f80, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r236;
cvt.f32.f16 %f81, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r237;
cvt.f32.f16 %f82, low;}



dp4a.s32.s32 %r276, %r353, %r240, %r383;

mul.lo.s32 %r470, %r276, %r425;
cvt.rn.f32.s32 %f192, %r470;
fma.rn.ftz.f32 %f193, %f79, %f192, 0f00000000;

dp4a.s32.s32 %r280, %r357, %r240, %r383;

cvt.rn.f32.s32 %f194, %r280;
fma.rn.ftz.f32 %f195, %f79, %f194, 0f00000000;

dp4a.s32.s32 %r284, %r361, %r248, %r383;

mul.lo.s32 %r471, %r284, %r434;
cvt.rn.f32.s32 %f196, %r471;
fma.rn.ftz.f32 %f197, %f80, %f196, %f193;

dp4a.s32.s32 %r288, %r365, %r248, %r383;

cvt.rn.f32.s32 %f198, %r288;
fma.rn.ftz.f32 %f199, %f80, %f198, %f195;

dp4a.s32.s32 %r292, %r369, %r256, %r383;

mul.lo.s32 %r472, %r292, %r443;
cvt.rn.f32.s32 %f200, %r472;
fma.rn.ftz.f32 %f201, %f81, %f200, %f197;

dp4a.s32.s32 %r296, %r373, %r256, %r383;

cvt.rn.f32.s32 %f202, %r296;
fma.rn.ftz.f32 %f203, %f81, %f202, %f199;

dp4a.s32.s32 %r300, %r377, %r264, %r383;

mul.lo.s32 %r473, %r300, %r452;
cvt.rn.f32.s32 %f204, %r473;
fma.rn.ftz.f32 %f205, %f82, %f204, %f201;

dp4a.s32.s32 %r304, %r381, %r264, %r383;

cvt.rn.f32.s32 %f206, %r304;
fma.rn.ftz.f32 %f207, %f82, %f206, %f203;

{.reg .f16 low,high;
mov.b32 {low,high},%r156;
cvt.f32.f16 %f83, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r156;
cvt.f32.f16 %f84, high;}


mul.ftz.f32 %f208, %f205, %f83;
mul.ftz.f32 %f209, %f207, %f84;
sub.ftz.f32 %f210, %f208, %f209;
add.ftz.f32 %f379, %f210, %f379;
cvt.u64.u32 %rd53, %r634;
add.s64 %rd54, %rd2, %rd53;
add.s64 %rd55, %rd54, %rd5;
mul.lo.s64 %rd56, %rd55, 36;
add.s64 %rd57, %rd12, %rd56;
add.s64 %rd58, %rd57, %rd34;
ld.global.nc.u32 %r316, [%rd58+4];
ld.global.nc.u32 %r310, [%rd57];

{.reg .f16 low,high;
mov.b32 {low,high},%r310;
cvt.f32.f16 %f85, low;}


ld.global.nc.u32 %r324, [%rd58+40];
ld.global.nc.u32 %r311, [%rd57+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r311;
cvt.f32.f16 %f86, low;}


ld.global.nc.u32 %r332, [%rd58+76];
ld.global.nc.u32 %r312, [%rd57+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r312;
cvt.f32.f16 %f87, low;}


ld.global.nc.u32 %r340, [%rd58+112];
ld.global.nc.u32 %r313, [%rd57+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r313;
cvt.f32.f16 %f88, low;}



dp4a.s32.s32 %r314, %r315, %r316, %r383;

mul.lo.s32 %r474, %r314, %r389;
cvt.rn.f32.s32 %f211, %r474;
fma.rn.ftz.f32 %f212, %f85, %f211, 0f00000000;

dp4a.s32.s32 %r318, %r319, %r316, %r383;

cvt.rn.f32.s32 %f213, %r318;
fma.rn.ftz.f32 %f214, %f85, %f213, 0f00000000;

dp4a.s32.s32 %r322, %r323, %r324, %r383;

mul.lo.s32 %r475, %r322, %r398;
cvt.rn.f32.s32 %f215, %r475;
fma.rn.ftz.f32 %f216, %f86, %f215, %f212;

dp4a.s32.s32 %r326, %r327, %r324, %r383;

cvt.rn.f32.s32 %f217, %r326;
fma.rn.ftz.f32 %f218, %f86, %f217, %f214;

dp4a.s32.s32 %r330, %r331, %r332, %r383;

mul.lo.s32 %r476, %r330, %r407;
cvt.rn.f32.s32 %f219, %r476;
fma.rn.ftz.f32 %f220, %f87, %f219, %f216;

dp4a.s32.s32 %r334, %r335, %r332, %r383;

cvt.rn.f32.s32 %f221, %r334;
fma.rn.ftz.f32 %f222, %f87, %f221, %f218;

dp4a.s32.s32 %r338, %r339, %r340, %r383;

mul.lo.s32 %r477, %r338, %r416;
cvt.rn.f32.s32 %f223, %r477;
fma.rn.ftz.f32 %f224, %f88, %f223, %f220;

dp4a.s32.s32 %r342, %r343, %r340, %r383;

cvt.rn.f32.s32 %f225, %r342;
fma.rn.ftz.f32 %f226, %f88, %f225, %f222;

{.reg .f16 low,high;
mov.b32 {low,high},%r118;
cvt.f32.f16 %f89, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r118;
cvt.f32.f16 %f90, high;}


mul.ftz.f32 %f227, %f224, %f89;
mul.ftz.f32 %f228, %f226, %f90;
sub.ftz.f32 %f229, %f227, %f228;
add.ftz.f32 %f378, %f229, %f378;

{.reg .f16 low,high;
mov.b32 {low,high},%r310;
cvt.f32.f16 %f91, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r311;
cvt.f32.f16 %f92, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r312;
cvt.f32.f16 %f93, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r313;
cvt.f32.f16 %f94, low;}



dp4a.s32.s32 %r352, %r353, %r316, %r383;

mul.lo.s32 %r478, %r352, %r425;
cvt.rn.f32.s32 %f230, %r478;
fma.rn.ftz.f32 %f231, %f91, %f230, 0f00000000;

dp4a.s32.s32 %r356, %r357, %r316, %r383;

cvt.rn.f32.s32 %f232, %r356;
fma.rn.ftz.f32 %f233, %f91, %f232, 0f00000000;

dp4a.s32.s32 %r360, %r361, %r324, %r383;

mul.lo.s32 %r479, %r360, %r434;
cvt.rn.f32.s32 %f234, %r479;
fma.rn.ftz.f32 %f235, %f92, %f234, %f231;

dp4a.s32.s32 %r364, %r365, %r324, %r383;

cvt.rn.f32.s32 %f236, %r364;
fma.rn.ftz.f32 %f237, %f92, %f236, %f233;

dp4a.s32.s32 %r368, %r369, %r332, %r383;

mul.lo.s32 %r480, %r368, %r443;
cvt.rn.f32.s32 %f238, %r480;
fma.rn.ftz.f32 %f239, %f93, %f238, %f235;

dp4a.s32.s32 %r372, %r373, %r332, %r383;

cvt.rn.f32.s32 %f240, %r372;
fma.rn.ftz.f32 %f241, %f93, %f240, %f237;

dp4a.s32.s32 %r376, %r377, %r340, %r383;

mul.lo.s32 %r481, %r376, %r452;
cvt.rn.f32.s32 %f242, %r481;
fma.rn.ftz.f32 %f243, %f94, %f242, %f239;

dp4a.s32.s32 %r380, %r381, %r340, %r383;

cvt.rn.f32.s32 %f244, %r380;
fma.rn.ftz.f32 %f245, %f94, %f244, %f241;

{.reg .f16 low,high;
mov.b32 {low,high},%r156;
cvt.f32.f16 %f95, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r156;
cvt.f32.f16 %f96, high;}


mul.ftz.f32 %f246, %f243, %f95;
mul.ftz.f32 %f247, %f245, %f96;
sub.ftz.f32 %f248, %f246, %f247;
add.ftz.f32 %f377, %f248, %f377;
add.s32 %r635, %r635, 64;
add.s32 %r634, %r634, 64;
add.s32 %r633, %r633, 64;
add.s32 %r632, %r632, 64;
add.s64 %rd79, %rd79, 672;
add.s32 %r636, %r636, 8;
setp.lt.s32 %p4, %r636, %r7;
@%p4 bra $L__BB135_2;

$L__BB135_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB135_5;

shl.b32 %r482, %r1, 10;
mov.u32 %r483, _ZZ13mul_mat_vec_qIL9ggml_type10ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r484, %r483, %r482;
shl.b32 %r485, %r2, 2;
add.s32 %r486, %r484, %r485;
st.shared.f32 [%r486+-1024], %f384;
st.shared.f32 [%r486+-896], %f383;
st.shared.f32 [%r486+-768], %f382;
st.shared.f32 [%r486+-640], %f381;
st.shared.f32 [%r486+-512], %f380;
st.shared.f32 [%r486+-384], %f379;
st.shared.f32 [%r486+-256], %f378;
st.shared.f32 [%r486+-128], %f377;

$L__BB135_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB135_22;

mov.u32 %r487, %ctaid.z;
mad.lo.s32 %r488, %r5, %r33, %r4;
mad.lo.s32 %r489, %r487, %r39, %r488;
cvt.u64.u32 %rd15, %r489;
shl.b32 %r490, %r2, 2;
mov.u32 %r491, _ZZ13mul_mat_vec_qIL9ggml_type10ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r22, %r491, %r490;
mov.u32 %r492, 2;
ld.shared.f32 %f249, [%r22];
add.ftz.f32 %f250, %f249, %f384;
ld.shared.f32 %f251, [%r22+1024];
add.ftz.f32 %f252, %f251, %f250;
ld.shared.f32 %f253, [%r22+2048];
add.ftz.f32 %f254, %f253, %f252;
mov.b32 %r493, %f254;
mov.u32 %r494, 31;
mov.u32 %r495, 16;
mov.u32 %r496, -1;
shfl.sync.bfly.b32 %r497|%p7, %r493, %r495, %r494, %r496;
mov.b32 %f255, %r497;
add.ftz.f32 %f256, %f254, %f255;
mov.b32 %r498, %f256;
mov.u32 %r499, 8;
shfl.sync.bfly.b32 %r500|%p8, %r498, %r499, %r494, %r496;
mov.b32 %f257, %r500;
add.ftz.f32 %f258, %f256, %f257;
mov.b32 %r501, %f258;
mov.u32 %r502, 4;
shfl.sync.bfly.b32 %r503|%p9, %r501, %r502, %r494, %r496;
mov.b32 %f259, %r503;
add.ftz.f32 %f260, %f258, %f259;
mov.b32 %r504, %f260;
shfl.sync.bfly.b32 %r505|%p10, %r504, %r492, %r494, %r496;
mov.b32 %f261, %r505;
add.ftz.f32 %f262, %f260, %f261;
mov.b32 %r506, %f262;
mov.u32 %r507, 1;
shfl.sync.bfly.b32 %r508|%p11, %r506, %r507, %r494, %r496;
mov.b32 %f263, %r508;
add.ftz.f32 %f25, %f262, %f263;
setp.lt.u32 %p12, %r4, %r27;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
mul.wide.u32 %rd59, %r489, 4;
add.s64 %rd16, %rd1, %rd59;
not.pred %p14, %p1;
@%p14 bra $L__BB135_8;

st.global.f32 [%rd16], %f25;

$L__BB135_8:
ld.shared.f32 %f264, [%r22+128];
add.ftz.f32 %f265, %f264, %f383;
ld.shared.f32 %f266, [%r22+1152];
add.ftz.f32 %f267, %f266, %f265;
ld.shared.f32 %f268, [%r22+2176];
add.ftz.f32 %f269, %f268, %f267;
mov.b32 %r509, %f269;
shfl.sync.bfly.b32 %r513|%p15, %r509, %r495, %r494, %r496;
mov.b32 %f270, %r513;
add.ftz.f32 %f271, %f269, %f270;
mov.b32 %r514, %f271;
shfl.sync.bfly.b32 %r516|%p16, %r514, %r499, %r494, %r496;
mov.b32 %f272, %r516;
add.ftz.f32 %f273, %f271, %f272;
mov.b32 %r517, %f273;
shfl.sync.bfly.b32 %r519|%p17, %r517, %r502, %r494, %r496;
mov.b32 %f274, %r519;
add.ftz.f32 %f275, %f273, %f274;
mov.b32 %r520, %f275;
shfl.sync.bfly.b32 %r522|%p18, %r520, %r492, %r494, %r496;
mov.b32 %f276, %r522;
add.ftz.f32 %f277, %f275, %f276;
mov.b32 %r523, %f277;
shfl.sync.bfly.b32 %r525|%p19, %r523, %r507, %r494, %r496;
mov.b32 %f278, %r525;
add.ftz.f32 %f26, %f277, %f278;
add.s32 %r526, %r4, 1;
setp.lt.u32 %p20, %r526, %r27;
setp.eq.s32 %p21, %r2, 1;
and.pred %p2, %p21, %p20;
not.pred %p22, %p2;
@%p22 bra $L__BB135_10;

st.global.f32 [%rd16+4], %f26;

$L__BB135_10:
ld.shared.f32 %f279, [%r22+256];
add.ftz.f32 %f280, %f279, %f382;
ld.shared.f32 %f281, [%r22+1280];
add.ftz.f32 %f282, %f281, %f280;
ld.shared.f32 %f283, [%r22+2304];
add.ftz.f32 %f284, %f283, %f282;
mov.b32 %r527, %f284;
mov.u32 %r528, 31;
mov.u32 %r529, 16;
mov.u32 %r530, -1;
shfl.sync.bfly.b32 %r531|%p23, %r527, %r529, %r528, %r530;
mov.b32 %f285, %r531;
add.ftz.f32 %f286, %f284, %f285;
mov.b32 %r532, %f286;
mov.u32 %r533, 8;
shfl.sync.bfly.b32 %r534|%p24, %r532, %r533, %r528, %r530;
mov.b32 %f287, %r534;
add.ftz.f32 %f288, %f286, %f287;
mov.b32 %r535, %f288;
mov.u32 %r536, 4;
shfl.sync.bfly.b32 %r537|%p25, %r535, %r536, %r528, %r530;
mov.b32 %f289, %r537;
add.ftz.f32 %f290, %f288, %f289;
mov.b32 %r538, %f290;
mov.u32 %r539, 2;
shfl.sync.bfly.b32 %r540|%p26, %r538, %r539, %r528, %r530;
mov.b32 %f291, %r540;
add.ftz.f32 %f292, %f290, %f291;
mov.b32 %r541, %f292;
mov.u32 %r542, 1;
shfl.sync.bfly.b32 %r543|%p27, %r541, %r542, %r528, %r530;
mov.b32 %f293, %r543;
add.ftz.f32 %f27, %f292, %f293;
@%p14 bra $L__BB135_12;

cvt.u64.u32 %rd60, %r27;
add.s64 %rd61, %rd60, %rd15;
shl.b64 %rd62, %rd61, 2;
add.s64 %rd63, %rd1, %rd62;
st.global.f32 [%rd63], %f27;

$L__BB135_12:
ld.shared.f32 %f294, [%r22+384];
add.ftz.f32 %f295, %f294, %f381;
ld.shared.f32 %f296, [%r22+1408];
add.ftz.f32 %f297, %f296, %f295;
ld.shared.f32 %f298, [%r22+2432];
add.ftz.f32 %f299, %f298, %f297;
mov.b32 %r544, %f299;
shfl.sync.bfly.b32 %r548|%p29, %r544, %r529, %r528, %r530;
mov.b32 %f300, %r548;
add.ftz.f32 %f301, %f299, %f300;
mov.b32 %r549, %f301;
shfl.sync.bfly.b32 %r551|%p30, %r549, %r533, %r528, %r530;
mov.b32 %f302, %r551;
add.ftz.f32 %f303, %f301, %f302;
mov.b32 %r552, %f303;
shfl.sync.bfly.b32 %r554|%p31, %r552, %r536, %r528, %r530;
mov.b32 %f304, %r554;
add.ftz.f32 %f305, %f303, %f304;
mov.b32 %r555, %f305;
shfl.sync.bfly.b32 %r557|%p32, %r555, %r539, %r528, %r530;
mov.b32 %f306, %r557;
add.ftz.f32 %f307, %f305, %f306;
mov.b32 %r558, %f307;
shfl.sync.bfly.b32 %r560|%p33, %r558, %r542, %r528, %r530;
mov.b32 %f308, %r560;
add.ftz.f32 %f28, %f307, %f308;
@%p22 bra $L__BB135_14;

add.s32 %r561, %r27, 1;
cvt.u64.u32 %rd64, %r561;
add.s64 %rd65, %rd64, %rd15;
shl.b64 %rd66, %rd65, 2;
add.s64 %rd67, %rd1, %rd66;
st.global.f32 [%rd67], %f28;

$L__BB135_14:
shl.b32 %r562, %r27, 1;
mov.u32 %r563, 1;
ld.shared.f32 %f309, [%r22+512];
add.ftz.f32 %f310, %f309, %f380;
ld.shared.f32 %f311, [%r22+1536];
add.ftz.f32 %f312, %f311, %f310;
ld.shared.f32 %f313, [%r22+2560];
add.ftz.f32 %f314, %f313, %f312;
mov.b32 %r564, %f314;
mov.u32 %r565, 31;
mov.u32 %r566, 16;
mov.u32 %r567, -1;
shfl.sync.bfly.b32 %r568|%p35, %r564, %r566, %r565, %r567;
mov.b32 %f315, %r568;
add.ftz.f32 %f316, %f314, %f315;
mov.b32 %r569, %f316;
mov.u32 %r570, 8;
shfl.sync.bfly.b32 %r571|%p36, %r569, %r570, %r565, %r567;
mov.b32 %f317, %r571;
add.ftz.f32 %f318, %f316, %f317;
mov.b32 %r572, %f318;
mov.u32 %r573, 4;
shfl.sync.bfly.b32 %r574|%p37, %r572, %r573, %r565, %r567;
mov.b32 %f319, %r574;
add.ftz.f32 %f320, %f318, %f319;
mov.b32 %r575, %f320;
mov.u32 %r576, 2;
shfl.sync.bfly.b32 %r577|%p38, %r575, %r576, %r565, %r567;
mov.b32 %f321, %r577;
add.ftz.f32 %f322, %f320, %f321;
mov.b32 %r578, %f322;
shfl.sync.bfly.b32 %r579|%p39, %r578, %r563, %r565, %r567;
mov.b32 %f323, %r579;
add.ftz.f32 %f29, %f322, %f323;
cvt.u64.u32 %rd68, %r562;
add.s64 %rd69, %rd68, %rd15;
shl.b64 %rd70, %rd69, 2;
add.s64 %rd17, %rd1, %rd70;
@%p14 bra $L__BB135_16;

st.global.f32 [%rd17], %f29;

$L__BB135_16:
ld.shared.f32 %f324, [%r22+640];
add.ftz.f32 %f325, %f324, %f379;
ld.shared.f32 %f326, [%r22+1664];
add.ftz.f32 %f327, %f326, %f325;
ld.shared.f32 %f328, [%r22+2688];
add.ftz.f32 %f329, %f328, %f327;
mov.b32 %r580, %f329;
shfl.sync.bfly.b32 %r584|%p41, %r580, %r566, %r565, %r567;
mov.b32 %f330, %r584;
add.ftz.f32 %f331, %f329, %f330;
mov.b32 %r585, %f331;
shfl.sync.bfly.b32 %r587|%p42, %r585, %r570, %r565, %r567;
mov.b32 %f332, %r587;
add.ftz.f32 %f333, %f331, %f332;
mov.b32 %r588, %f333;
shfl.sync.bfly.b32 %r590|%p43, %r588, %r573, %r565, %r567;
mov.b32 %f334, %r590;
add.ftz.f32 %f335, %f333, %f334;
mov.b32 %r591, %f335;
shfl.sync.bfly.b32 %r593|%p44, %r591, %r576, %r565, %r567;
mov.b32 %f336, %r593;
add.ftz.f32 %f337, %f335, %f336;
mov.b32 %r594, %f337;
shfl.sync.bfly.b32 %r596|%p45, %r594, %r563, %r565, %r567;
mov.b32 %f338, %r596;
add.ftz.f32 %f30, %f337, %f338;
@%p22 bra $L__BB135_18;

st.global.f32 [%rd17+4], %f30;

$L__BB135_18:
mul.lo.s32 %r24, %r27, 3;
ld.shared.f32 %f339, [%r22+768];
add.ftz.f32 %f340, %f339, %f378;
ld.shared.f32 %f341, [%r22+1792];
add.ftz.f32 %f342, %f341, %f340;
ld.shared.f32 %f343, [%r22+2816];
add.ftz.f32 %f344, %f343, %f342;
mov.b32 %r597, %f344;
mov.u32 %r598, 31;
mov.u32 %r599, 16;
mov.u32 %r600, -1;
shfl.sync.bfly.b32 %r601|%p47, %r597, %r599, %r598, %r600;
mov.b32 %f345, %r601;
add.ftz.f32 %f346, %f344, %f345;
mov.b32 %r602, %f346;
mov.u32 %r603, 8;
shfl.sync.bfly.b32 %r604|%p48, %r602, %r603, %r598, %r600;
mov.b32 %f347, %r604;
add.ftz.f32 %f348, %f346, %f347;
mov.b32 %r605, %f348;
mov.u32 %r606, 4;
shfl.sync.bfly.b32 %r607|%p49, %r605, %r606, %r598, %r600;
mov.b32 %f349, %r607;
add.ftz.f32 %f350, %f348, %f349;
mov.b32 %r608, %f350;
mov.u32 %r609, 2;
shfl.sync.bfly.b32 %r610|%p50, %r608, %r609, %r598, %r600;
mov.b32 %f351, %r610;
add.ftz.f32 %f352, %f350, %f351;
mov.b32 %r611, %f352;
mov.u32 %r612, 1;
shfl.sync.bfly.b32 %r613|%p51, %r611, %r612, %r598, %r600;
mov.b32 %f353, %r613;
add.ftz.f32 %f31, %f352, %f353;
@%p14 bra $L__BB135_20;

cvt.u64.u32 %rd71, %r24;
add.s64 %rd72, %rd71, %rd15;
shl.b64 %rd73, %rd72, 2;
add.s64 %rd74, %rd1, %rd73;
st.global.f32 [%rd74], %f31;

$L__BB135_20:
ld.shared.f32 %f354, [%r22+896];
add.ftz.f32 %f355, %f354, %f377;
ld.shared.f32 %f356, [%r22+1920];
add.ftz.f32 %f357, %f356, %f355;
ld.shared.f32 %f358, [%r22+2944];
add.ftz.f32 %f359, %f358, %f357;
mov.b32 %r614, %f359;
shfl.sync.bfly.b32 %r618|%p53, %r614, %r599, %r598, %r600;
mov.b32 %f360, %r618;
add.ftz.f32 %f361, %f359, %f360;
mov.b32 %r619, %f361;
shfl.sync.bfly.b32 %r621|%p54, %r619, %r603, %r598, %r600;
mov.b32 %f362, %r621;
add.ftz.f32 %f363, %f361, %f362;
mov.b32 %r622, %f363;
shfl.sync.bfly.b32 %r624|%p55, %r622, %r606, %r598, %r600;
mov.b32 %f364, %r624;
add.ftz.f32 %f365, %f363, %f364;
mov.b32 %r625, %f365;
shfl.sync.bfly.b32 %r627|%p56, %r625, %r609, %r598, %r600;
mov.b32 %f366, %r627;
add.ftz.f32 %f367, %f365, %f366;
mov.b32 %r628, %f367;
shfl.sync.bfly.b32 %r630|%p57, %r628, %r612, %r598, %r600;
mov.b32 %f368, %r630;
add.ftz.f32 %f32, %f367, %f368;
@%p22 bra $L__BB135_22;

add.s32 %r631, %r24, 1;
cvt.u64.u32 %rd75, %r631;
add.s64 %rd76, %rd75, %rd15;
shl.b64 %rd77, %rd76, 2;
add.s64 %rd78, %rd1, %rd77;
st.global.f32 [%rd78], %f32;

$L__BB135_22:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type10ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type10ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type10ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type10ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type10ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type10ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type10ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type10ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type10ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<71>;
.reg .b16 %rs<9>;
.reg .f32 %f<441>;
.reg .b32 %r<765>;
.reg .b64 %rd<96>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type10ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1280];

ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type10ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type10ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd20, [_Z13mul_mat_vec_qIL9ggml_type10ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type10ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type10ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type10ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type10ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type10ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type10ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type10ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type10ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type10ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type10ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type10ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type10ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type10ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type10ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r42, %r1, 5;
mov.u32 %r43, %tid.x;
add.s32 %r2, %r42, %r43;
mov.u32 %r3, %ctaid.y;
mov.u32 %r4, %ctaid.z;
shr.s32 %r44, %r2, 31;
shr.u32 %r45, %r44, 28;
add.s32 %r46, %r2, %r45;
shr.s32 %r764, %r46, 4;
shr.u32 %r6, %r41, 8;
setp.ge.s32 %p3, %r764, %r6;
mov.f32 %f431, 0f00000000;
mov.f32 %f432, %f431;
mov.f32 %f433, %f431;
mov.f32 %f434, %f431;
mov.f32 %f435, %f431;
mov.f32 %f436, %f431;
mov.f32 %f437, %f431;
mov.f32 %f438, %f431;
mov.f32 %f439, %f431;
mov.f32 %f440, %f431;
@%p3 bra $L__BB136_3;

mov.u32 %r47, %ctaid.x;
shl.b32 %r48, %r47, 1;
mul.lo.s32 %r49, %r3, %r33;
cvt.u64.u32 %rd21, %r49;
mul.lo.s32 %r50, %r4, %r39;
cvt.u64.u32 %rd22, %r50;
add.s64 %rd1, %rd22, %rd21;
mul.hi.u32 %r51, %r4, %r35;
add.s32 %r52, %r51, %r4;
shr.u32 %r53, %r52, %r36;
mul.lo.s32 %r54, %r53, %r38;
mul.hi.u32 %r55, %r3, %r29;
add.s32 %r56, %r55, %r3;
shr.u32 %r57, %r56, %r30;
mul.lo.s32 %r58, %r57, %r32;
cvta.to.global.u64 %rd95, %rd18;
or.b32 %r59, %r48, 1;
and.b32 %r63, %r46, -16;
sub.s32 %r64, %r2, %r63;
shr.s32 %r65, %r64, 31;
shr.u32 %r66, %r65, 29;
add.s32 %r67, %r64, %r66;
and.b32 %r68, %r67, -8;
sub.s32 %r69, %r64, %r68;
shr.s32 %r70, %r69, 31;
shr.u32 %r71, %r70, 30;
add.s32 %r72, %r69, %r71;
shr.s32 %r73, %r72, 2;
add.s32 %r74, %r68, %r73;
cvt.s64.s32 %rd23, %r74;
shr.s32 %r75, %r67, 3;
shl.b32 %r76, %r75, 2;
cvt.s64.s32 %rd3, %r69;
cvt.s64.s32 %rd4, %r76;
shl.b32 %r77, %r27, 1;
shl.b32 %r759, %r764, 3;
add.s32 %r763, %r759, %r77;
mad.lo.s32 %r762, %r27, 3, %r759;
shl.b32 %r78, %r27, 2;
add.s32 %r761, %r759, %r78;
add.s32 %r760, %r27, %r759;
add.s32 %r79, %r764, %r54;
add.s32 %r80, %r79, %r58;
mad.lo.s32 %r81, %r59, %r26, %r80;
mul.wide.s32 %rd5, %r81, 84;
mad.lo.s32 %r82, %r48, %r26, %r764;
add.s32 %r83, %r82, %r54;
add.s32 %r84, %r83, %r58;
mul.wide.s32 %rd6, %r84, 84;
mul.wide.s32 %rd24, %r64, 4;
add.s64 %rd25, %rd6, %rd24;
add.s64 %rd7, %rd25, 16;
add.s64 %rd26, %rd6, %rd23;
add.s64 %rd8, %rd26, 6;
add.s64 %rd9, %rd5, %rd23;
add.s64 %rd27, %rd5, %rd24;
add.s64 %rd10, %rd27, 16;
cvta.to.global.u64 %rd11, %rd19;
mov.f32 %f431, 0f00000000;
shl.b64 %rd34, %rd3, 2;

$L__BB136_2:
cvt.u64.u32 %rd28, %r759;
add.s64 %rd29, %rd1, %rd28;
add.s64 %rd30, %rd95, %rd7;
ld.global.nc.u32 %r465, [%rd30];
add.s64 %rd31, %rd29, %rd4;
mul.lo.s64 %rd32, %rd31, 36;
add.s64 %rd33, %rd11, %rd32;
add.s64 %rd35, %rd33, %rd34;
ld.global.nc.u32 %r91, [%rd35+4];
ld.global.nc.u32 %r85, [%rd33];

{.reg .f16 low,high;
mov.b32 {low,high},%r85;
cvt.f32.f16 %f61, low;}


ld.global.nc.u32 %r99, [%rd35+40];
ld.global.nc.u32 %r86, [%rd33+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r86;
cvt.f32.f16 %f62, low;}


ld.global.nc.u32 %r107, [%rd35+76];
ld.global.nc.u32 %r87, [%rd33+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r87;
cvt.f32.f16 %f63, low;}


ld.global.nc.u32 %r115, [%rd35+112];
ld.global.nc.u32 %r88, [%rd33+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r88;
cvt.f32.f16 %f64, low;}


add.s64 %rd36, %rd95, %rd8;
ld.global.nc.u8 %rs1, [%rd36+-6];
cvt.u32.u16 %r466, %rs1;
and.b32 %r467, %r466, 240;
and.b32 %r394, %r465, 50529027;
mov.u32 %r462, 0;

dp4a.s32.s32 %r89, %r394, %r91, %r462;

and.b32 %r468, %r466, 15;
mul.lo.s32 %r469, %r89, %r468;
cvt.rn.f32.s32 %f121, %r469;
fma.rn.ftz.f32 %f122, %f61, %f121, 0f00000000;
shr.u32 %r470, %r467, 4;
shl.b32 %r471, %r470, 8;
or.b32 %r472, %r471, %r470;
shl.b32 %r473, %r472, 16;
or.b32 %r398, %r473, %r472;

dp4a.s32.s32 %r93, %r398, %r91, %r462;

cvt.rn.f32.s32 %f123, %r93;
fma.rn.ftz.f32 %f124, %f61, %f123, 0f00000000;
ld.global.nc.u8 %rs2, [%rd36+-4];
cvt.u32.u16 %r474, %rs2;
and.b32 %r475, %r474, 240;
shr.u32 %r476, %r465, 2;
and.b32 %r402, %r476, 50529027;

dp4a.s32.s32 %r97, %r402, %r99, %r462;

and.b32 %r477, %r474, 15;
mul.lo.s32 %r478, %r97, %r477;
cvt.rn.f32.s32 %f125, %r478;
fma.rn.ftz.f32 %f126, %f62, %f125, %f122;
shr.u32 %r479, %r475, 4;
shl.b32 %r480, %r479, 8;
or.b32 %r481, %r480, %r479;
shl.b32 %r482, %r481, 16;
or.b32 %r406, %r482, %r481;

dp4a.s32.s32 %r101, %r406, %r99, %r462;

cvt.rn.f32.s32 %f127, %r101;
fma.rn.ftz.f32 %f128, %f62, %f127, %f124;
ld.global.nc.u8 %rs3, [%rd36+-2];
cvt.u32.u16 %r483, %rs3;
and.b32 %r484, %r483, 240;
shr.u32 %r485, %r465, 4;
and.b32 %r410, %r485, 50529027;

dp4a.s32.s32 %r105, %r410, %r107, %r462;

and.b32 %r486, %r483, 15;
mul.lo.s32 %r487, %r105, %r486;
cvt.rn.f32.s32 %f129, %r487;
fma.rn.ftz.f32 %f130, %f63, %f129, %f126;
shr.u32 %r488, %r484, 4;
shl.b32 %r489, %r488, 8;
or.b32 %r490, %r489, %r488;
shl.b32 %r491, %r490, 16;
or.b32 %r414, %r491, %r490;

dp4a.s32.s32 %r109, %r414, %r107, %r462;

cvt.rn.f32.s32 %f131, %r109;
fma.rn.ftz.f32 %f132, %f63, %f131, %f128;
ld.global.nc.u8 %rs4, [%rd36];
cvt.u32.u16 %r492, %rs4;
and.b32 %r493, %r492, 240;
shr.u32 %r494, %r465, 6;
and.b32 %r418, %r494, 50529027;

dp4a.s32.s32 %r113, %r418, %r115, %r462;

and.b32 %r495, %r492, 15;
mul.lo.s32 %r496, %r113, %r495;
cvt.rn.f32.s32 %f133, %r496;
fma.rn.ftz.f32 %f134, %f64, %f133, %f130;
shr.u32 %r497, %r493, 4;
shl.b32 %r498, %r497, 8;
or.b32 %r499, %r498, %r497;
shl.b32 %r500, %r499, 16;
or.b32 %r422, %r500, %r499;

dp4a.s32.s32 %r117, %r422, %r115, %r462;

cvt.rn.f32.s32 %f135, %r117;
fma.rn.ftz.f32 %f136, %f64, %f135, %f132;
add.s64 %rd37, %rd95, %rd6;
ld.global.nc.u32 %r121, [%rd37+80];

{.reg .f16 low,high;
mov.b32 {low,high},%r121;
cvt.f32.f16 %f65, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r121;
cvt.f32.f16 %f66, high;}


mul.ftz.f32 %f137, %f134, %f65;
mul.ftz.f32 %f138, %f136, %f66;
sub.ftz.f32 %f139, %f137, %f138;
add.ftz.f32 %f440, %f139, %f440;
add.s64 %rd38, %rd95, %rd10;
ld.global.nc.u32 %r501, [%rd38];

{.reg .f16 low,high;
mov.b32 {low,high},%r85;
cvt.f32.f16 %f67, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r86;
cvt.f32.f16 %f68, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r87;
cvt.f32.f16 %f69, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r88;
cvt.f32.f16 %f70, low;}


add.s64 %rd39, %rd95, %rd9;
ld.global.nc.u8 %rs5, [%rd39];
cvt.u32.u16 %r502, %rs5;
and.b32 %r503, %r502, 240;
and.b32 %r432, %r501, 50529027;

dp4a.s32.s32 %r127, %r432, %r91, %r462;

and.b32 %r504, %r502, 15;
mul.lo.s32 %r505, %r127, %r504;
cvt.rn.f32.s32 %f140, %r505;
fma.rn.ftz.f32 %f141, %f67, %f140, 0f00000000;
shr.u32 %r506, %r503, 4;
shl.b32 %r507, %r506, 8;
or.b32 %r508, %r507, %r506;
shl.b32 %r509, %r508, 16;
or.b32 %r436, %r509, %r508;

dp4a.s32.s32 %r131, %r436, %r91, %r462;

cvt.rn.f32.s32 %f142, %r131;
fma.rn.ftz.f32 %f143, %f67, %f142, 0f00000000;
ld.global.nc.u8 %rs6, [%rd39+2];
cvt.u32.u16 %r510, %rs6;
and.b32 %r511, %r510, 240;
shr.u32 %r512, %r501, 2;
and.b32 %r440, %r512, 50529027;

dp4a.s32.s32 %r135, %r440, %r99, %r462;

and.b32 %r513, %r510, 15;
mul.lo.s32 %r514, %r135, %r513;
cvt.rn.f32.s32 %f144, %r514;
fma.rn.ftz.f32 %f145, %f68, %f144, %f141;
shr.u32 %r515, %r511, 4;
shl.b32 %r516, %r515, 8;
or.b32 %r517, %r516, %r515;
shl.b32 %r518, %r517, 16;
or.b32 %r444, %r518, %r517;

dp4a.s32.s32 %r139, %r444, %r99, %r462;

cvt.rn.f32.s32 %f146, %r139;
fma.rn.ftz.f32 %f147, %f68, %f146, %f143;
ld.global.nc.u8 %rs7, [%rd39+4];
cvt.u32.u16 %r519, %rs7;
and.b32 %r520, %r519, 240;
shr.u32 %r521, %r501, 4;
and.b32 %r448, %r521, 50529027;

dp4a.s32.s32 %r143, %r448, %r107, %r462;

and.b32 %r522, %r519, 15;
mul.lo.s32 %r523, %r143, %r522;
cvt.rn.f32.s32 %f148, %r523;
fma.rn.ftz.f32 %f149, %f69, %f148, %f145;
shr.u32 %r524, %r520, 4;
shl.b32 %r525, %r524, 8;
or.b32 %r526, %r525, %r524;
shl.b32 %r527, %r526, 16;
or.b32 %r452, %r527, %r526;

dp4a.s32.s32 %r147, %r452, %r107, %r462;

cvt.rn.f32.s32 %f150, %r147;
fma.rn.ftz.f32 %f151, %f69, %f150, %f147;
ld.global.nc.u8 %rs8, [%rd39+6];
cvt.u32.u16 %r528, %rs8;
and.b32 %r529, %r528, 240;
shr.u32 %r530, %r501, 6;
and.b32 %r456, %r530, 50529027;

dp4a.s32.s32 %r151, %r456, %r115, %r462;

and.b32 %r531, %r528, 15;
mul.lo.s32 %r532, %r151, %r531;
cvt.rn.f32.s32 %f152, %r532;
fma.rn.ftz.f32 %f153, %f70, %f152, %f149;
shr.u32 %r533, %r529, 4;
shl.b32 %r534, %r533, 8;
or.b32 %r535, %r534, %r533;
shl.b32 %r536, %r535, 16;
or.b32 %r460, %r536, %r535;

dp4a.s32.s32 %r155, %r460, %r115, %r462;

cvt.rn.f32.s32 %f154, %r155;
fma.rn.ftz.f32 %f155, %f70, %f154, %f151;
add.s64 %rd40, %rd95, %rd5;
ld.global.nc.u32 %r159, [%rd40+80];

{.reg .f16 low,high;
mov.b32 {low,high},%r159;
cvt.f32.f16 %f71, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r159;
cvt.f32.f16 %f72, high;}


mul.ftz.f32 %f156, %f153, %f71;
mul.ftz.f32 %f157, %f155, %f72;
sub.ftz.f32 %f158, %f156, %f157;
add.ftz.f32 %f439, %f158, %f439;
cvt.u64.u32 %rd41, %r760;
add.s64 %rd42, %rd1, %rd41;
add.s64 %rd43, %rd42, %rd4;
mul.lo.s64 %rd44, %rd43, 36;
add.s64 %rd45, %rd11, %rd44;
add.s64 %rd46, %rd45, %rd34;
ld.global.nc.u32 %r167, [%rd46+4];
ld.global.nc.u32 %r161, [%rd45];

{.reg .f16 low,high;
mov.b32 {low,high},%r161;
cvt.f32.f16 %f73, low;}


ld.global.nc.u32 %r175, [%rd46+40];
ld.global.nc.u32 %r162, [%rd45+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r162;
cvt.f32.f16 %f74, low;}


ld.global.nc.u32 %r183, [%rd46+76];
ld.global.nc.u32 %r163, [%rd45+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r163;
cvt.f32.f16 %f75, low;}


ld.global.nc.u32 %r191, [%rd46+112];
ld.global.nc.u32 %r164, [%rd45+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r164;
cvt.f32.f16 %f76, low;}



dp4a.s32.s32 %r165, %r394, %r167, %r462;

mul.lo.s32 %r537, %r165, %r468;
cvt.rn.f32.s32 %f159, %r537;
fma.rn.ftz.f32 %f160, %f73, %f159, 0f00000000;

dp4a.s32.s32 %r169, %r398, %r167, %r462;

cvt.rn.f32.s32 %f161, %r169;
fma.rn.ftz.f32 %f162, %f73, %f161, 0f00000000;

dp4a.s32.s32 %r173, %r402, %r175, %r462;

mul.lo.s32 %r538, %r173, %r477;
cvt.rn.f32.s32 %f163, %r538;
fma.rn.ftz.f32 %f164, %f74, %f163, %f160;

dp4a.s32.s32 %r177, %r406, %r175, %r462;

cvt.rn.f32.s32 %f165, %r177;
fma.rn.ftz.f32 %f166, %f74, %f165, %f162;

dp4a.s32.s32 %r181, %r410, %r183, %r462;

mul.lo.s32 %r539, %r181, %r486;
cvt.rn.f32.s32 %f167, %r539;
fma.rn.ftz.f32 %f168, %f75, %f167, %f164;

dp4a.s32.s32 %r185, %r414, %r183, %r462;

cvt.rn.f32.s32 %f169, %r185;
fma.rn.ftz.f32 %f170, %f75, %f169, %f166;

dp4a.s32.s32 %r189, %r418, %r191, %r462;

mul.lo.s32 %r540, %r189, %r495;
cvt.rn.f32.s32 %f171, %r540;
fma.rn.ftz.f32 %f172, %f76, %f171, %f168;

dp4a.s32.s32 %r193, %r422, %r191, %r462;

cvt.rn.f32.s32 %f173, %r193;
fma.rn.ftz.f32 %f174, %f76, %f173, %f170;

{.reg .f16 low,high;
mov.b32 {low,high},%r121;
cvt.f32.f16 %f77, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r121;
cvt.f32.f16 %f78, high;}


mul.ftz.f32 %f175, %f172, %f77;
mul.ftz.f32 %f176, %f174, %f78;
sub.ftz.f32 %f177, %f175, %f176;
add.ftz.f32 %f438, %f177, %f438;

{.reg .f16 low,high;
mov.b32 {low,high},%r161;
cvt.f32.f16 %f79, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r162;
cvt.f32.f16 %f80, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r163;
cvt.f32.f16 %f81, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r164;
cvt.f32.f16 %f82, low;}



dp4a.s32.s32 %r203, %r432, %r167, %r462;

mul.lo.s32 %r541, %r203, %r504;
cvt.rn.f32.s32 %f178, %r541;
fma.rn.ftz.f32 %f179, %f79, %f178, 0f00000000;

dp4a.s32.s32 %r207, %r436, %r167, %r462;

cvt.rn.f32.s32 %f180, %r207;
fma.rn.ftz.f32 %f181, %f79, %f180, 0f00000000;

dp4a.s32.s32 %r211, %r440, %r175, %r462;

mul.lo.s32 %r542, %r211, %r513;
cvt.rn.f32.s32 %f182, %r542;
fma.rn.ftz.f32 %f183, %f80, %f182, %f179;

dp4a.s32.s32 %r215, %r444, %r175, %r462;

cvt.rn.f32.s32 %f184, %r215;
fma.rn.ftz.f32 %f185, %f80, %f184, %f181;

dp4a.s32.s32 %r219, %r448, %r183, %r462;

mul.lo.s32 %r543, %r219, %r522;
cvt.rn.f32.s32 %f186, %r543;
fma.rn.ftz.f32 %f187, %f81, %f186, %f183;

dp4a.s32.s32 %r223, %r452, %r183, %r462;

cvt.rn.f32.s32 %f188, %r223;
fma.rn.ftz.f32 %f189, %f81, %f188, %f185;

dp4a.s32.s32 %r227, %r456, %r191, %r462;

mul.lo.s32 %r544, %r227, %r531;
cvt.rn.f32.s32 %f190, %r544;
fma.rn.ftz.f32 %f191, %f82, %f190, %f187;

dp4a.s32.s32 %r231, %r460, %r191, %r462;

cvt.rn.f32.s32 %f192, %r231;
fma.rn.ftz.f32 %f193, %f82, %f192, %f189;

{.reg .f16 low,high;
mov.b32 {low,high},%r159;
cvt.f32.f16 %f83, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r159;
cvt.f32.f16 %f84, high;}


mul.ftz.f32 %f194, %f191, %f83;
mul.ftz.f32 %f195, %f193, %f84;
sub.ftz.f32 %f196, %f194, %f195;
add.ftz.f32 %f437, %f196, %f437;
cvt.u64.u32 %rd47, %r763;
add.s64 %rd48, %rd1, %rd47;
add.s64 %rd49, %rd48, %rd4;
mul.lo.s64 %rd50, %rd49, 36;
add.s64 %rd51, %rd11, %rd50;
add.s64 %rd52, %rd51, %rd34;
ld.global.nc.u32 %r243, [%rd52+4];
ld.global.nc.u32 %r237, [%rd51];

{.reg .f16 low,high;
mov.b32 {low,high},%r237;
cvt.f32.f16 %f85, low;}


ld.global.nc.u32 %r251, [%rd52+40];
ld.global.nc.u32 %r238, [%rd51+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r238;
cvt.f32.f16 %f86, low;}


ld.global.nc.u32 %r259, [%rd52+76];
ld.global.nc.u32 %r239, [%rd51+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r239;
cvt.f32.f16 %f87, low;}


ld.global.nc.u32 %r267, [%rd52+112];
ld.global.nc.u32 %r240, [%rd51+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r240;
cvt.f32.f16 %f88, low;}



dp4a.s32.s32 %r241, %r394, %r243, %r462;

mul.lo.s32 %r545, %r241, %r468;
cvt.rn.f32.s32 %f197, %r545;
fma.rn.ftz.f32 %f198, %f85, %f197, 0f00000000;

dp4a.s32.s32 %r245, %r398, %r243, %r462;

cvt.rn.f32.s32 %f199, %r245;
fma.rn.ftz.f32 %f200, %f85, %f199, 0f00000000;

dp4a.s32.s32 %r249, %r402, %r251, %r462;

mul.lo.s32 %r546, %r249, %r477;
cvt.rn.f32.s32 %f201, %r546;
fma.rn.ftz.f32 %f202, %f86, %f201, %f198;

dp4a.s32.s32 %r253, %r406, %r251, %r462;

cvt.rn.f32.s32 %f203, %r253;
fma.rn.ftz.f32 %f204, %f86, %f203, %f200;

dp4a.s32.s32 %r257, %r410, %r259, %r462;

mul.lo.s32 %r547, %r257, %r486;
cvt.rn.f32.s32 %f205, %r547;
fma.rn.ftz.f32 %f206, %f87, %f205, %f202;

dp4a.s32.s32 %r261, %r414, %r259, %r462;

cvt.rn.f32.s32 %f207, %r261;
fma.rn.ftz.f32 %f208, %f87, %f207, %f204;

dp4a.s32.s32 %r265, %r418, %r267, %r462;

mul.lo.s32 %r548, %r265, %r495;
cvt.rn.f32.s32 %f209, %r548;
fma.rn.ftz.f32 %f210, %f88, %f209, %f206;

dp4a.s32.s32 %r269, %r422, %r267, %r462;

cvt.rn.f32.s32 %f211, %r269;
fma.rn.ftz.f32 %f212, %f88, %f211, %f208;

{.reg .f16 low,high;
mov.b32 {low,high},%r121;
cvt.f32.f16 %f89, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r121;
cvt.f32.f16 %f90, high;}


mul.ftz.f32 %f213, %f210, %f89;
mul.ftz.f32 %f214, %f212, %f90;
sub.ftz.f32 %f215, %f213, %f214;
add.ftz.f32 %f436, %f215, %f436;

{.reg .f16 low,high;
mov.b32 {low,high},%r237;
cvt.f32.f16 %f91, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r238;
cvt.f32.f16 %f92, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r239;
cvt.f32.f16 %f93, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r240;
cvt.f32.f16 %f94, low;}



dp4a.s32.s32 %r279, %r432, %r243, %r462;

mul.lo.s32 %r549, %r279, %r504;
cvt.rn.f32.s32 %f216, %r549;
fma.rn.ftz.f32 %f217, %f91, %f216, 0f00000000;

dp4a.s32.s32 %r283, %r436, %r243, %r462;

cvt.rn.f32.s32 %f218, %r283;
fma.rn.ftz.f32 %f219, %f91, %f218, 0f00000000;

dp4a.s32.s32 %r287, %r440, %r251, %r462;

mul.lo.s32 %r550, %r287, %r513;
cvt.rn.f32.s32 %f220, %r550;
fma.rn.ftz.f32 %f221, %f92, %f220, %f217;

dp4a.s32.s32 %r291, %r444, %r251, %r462;

cvt.rn.f32.s32 %f222, %r291;
fma.rn.ftz.f32 %f223, %f92, %f222, %f219;

dp4a.s32.s32 %r295, %r448, %r259, %r462;

mul.lo.s32 %r551, %r295, %r522;
cvt.rn.f32.s32 %f224, %r551;
fma.rn.ftz.f32 %f225, %f93, %f224, %f221;

dp4a.s32.s32 %r299, %r452, %r259, %r462;

cvt.rn.f32.s32 %f226, %r299;
fma.rn.ftz.f32 %f227, %f93, %f226, %f223;

dp4a.s32.s32 %r303, %r456, %r267, %r462;

mul.lo.s32 %r552, %r303, %r531;
cvt.rn.f32.s32 %f228, %r552;
fma.rn.ftz.f32 %f229, %f94, %f228, %f225;

dp4a.s32.s32 %r307, %r460, %r267, %r462;

cvt.rn.f32.s32 %f230, %r307;
fma.rn.ftz.f32 %f231, %f94, %f230, %f227;

{.reg .f16 low,high;
mov.b32 {low,high},%r159;
cvt.f32.f16 %f95, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r159;
cvt.f32.f16 %f96, high;}


mul.ftz.f32 %f232, %f229, %f95;
mul.ftz.f32 %f233, %f231, %f96;
sub.ftz.f32 %f234, %f232, %f233;
add.ftz.f32 %f435, %f234, %f435;
cvt.u64.u32 %rd53, %r762;
add.s64 %rd54, %rd1, %rd53;
add.s64 %rd55, %rd54, %rd4;
mul.lo.s64 %rd56, %rd55, 36;
add.s64 %rd57, %rd11, %rd56;
add.s64 %rd58, %rd57, %rd34;
ld.global.nc.u32 %r319, [%rd58+4];
ld.global.nc.u32 %r313, [%rd57];

{.reg .f16 low,high;
mov.b32 {low,high},%r313;
cvt.f32.f16 %f97, low;}


ld.global.nc.u32 %r327, [%rd58+40];
ld.global.nc.u32 %r314, [%rd57+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r314;
cvt.f32.f16 %f98, low;}


ld.global.nc.u32 %r335, [%rd58+76];
ld.global.nc.u32 %r315, [%rd57+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r315;
cvt.f32.f16 %f99, low;}


ld.global.nc.u32 %r343, [%rd58+112];
ld.global.nc.u32 %r316, [%rd57+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r316;
cvt.f32.f16 %f100, low;}



dp4a.s32.s32 %r317, %r394, %r319, %r462;

mul.lo.s32 %r553, %r317, %r468;
cvt.rn.f32.s32 %f235, %r553;
fma.rn.ftz.f32 %f236, %f97, %f235, 0f00000000;

dp4a.s32.s32 %r321, %r398, %r319, %r462;

cvt.rn.f32.s32 %f237, %r321;
fma.rn.ftz.f32 %f238, %f97, %f237, 0f00000000;

dp4a.s32.s32 %r325, %r402, %r327, %r462;

mul.lo.s32 %r554, %r325, %r477;
cvt.rn.f32.s32 %f239, %r554;
fma.rn.ftz.f32 %f240, %f98, %f239, %f236;

dp4a.s32.s32 %r329, %r406, %r327, %r462;

cvt.rn.f32.s32 %f241, %r329;
fma.rn.ftz.f32 %f242, %f98, %f241, %f238;

dp4a.s32.s32 %r333, %r410, %r335, %r462;

mul.lo.s32 %r555, %r333, %r486;
cvt.rn.f32.s32 %f243, %r555;
fma.rn.ftz.f32 %f244, %f99, %f243, %f240;

dp4a.s32.s32 %r337, %r414, %r335, %r462;

cvt.rn.f32.s32 %f245, %r337;
fma.rn.ftz.f32 %f246, %f99, %f245, %f242;

dp4a.s32.s32 %r341, %r418, %r343, %r462;

mul.lo.s32 %r556, %r341, %r495;
cvt.rn.f32.s32 %f247, %r556;
fma.rn.ftz.f32 %f248, %f100, %f247, %f244;

dp4a.s32.s32 %r345, %r422, %r343, %r462;

cvt.rn.f32.s32 %f249, %r345;
fma.rn.ftz.f32 %f250, %f100, %f249, %f246;

{.reg .f16 low,high;
mov.b32 {low,high},%r121;
cvt.f32.f16 %f101, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r121;
cvt.f32.f16 %f102, high;}


mul.ftz.f32 %f251, %f248, %f101;
mul.ftz.f32 %f252, %f250, %f102;
sub.ftz.f32 %f253, %f251, %f252;
add.ftz.f32 %f434, %f253, %f434;

{.reg .f16 low,high;
mov.b32 {low,high},%r313;
cvt.f32.f16 %f103, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r314;
cvt.f32.f16 %f104, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r315;
cvt.f32.f16 %f105, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r316;
cvt.f32.f16 %f106, low;}



dp4a.s32.s32 %r355, %r432, %r319, %r462;

mul.lo.s32 %r557, %r355, %r504;
cvt.rn.f32.s32 %f254, %r557;
fma.rn.ftz.f32 %f255, %f103, %f254, 0f00000000;

dp4a.s32.s32 %r359, %r436, %r319, %r462;

cvt.rn.f32.s32 %f256, %r359;
fma.rn.ftz.f32 %f257, %f103, %f256, 0f00000000;

dp4a.s32.s32 %r363, %r440, %r327, %r462;

mul.lo.s32 %r558, %r363, %r513;
cvt.rn.f32.s32 %f258, %r558;
fma.rn.ftz.f32 %f259, %f104, %f258, %f255;

dp4a.s32.s32 %r367, %r444, %r327, %r462;

cvt.rn.f32.s32 %f260, %r367;
fma.rn.ftz.f32 %f261, %f104, %f260, %f257;

dp4a.s32.s32 %r371, %r448, %r335, %r462;

mul.lo.s32 %r559, %r371, %r522;
cvt.rn.f32.s32 %f262, %r559;
fma.rn.ftz.f32 %f263, %f105, %f262, %f259;

dp4a.s32.s32 %r375, %r452, %r335, %r462;

cvt.rn.f32.s32 %f264, %r375;
fma.rn.ftz.f32 %f265, %f105, %f264, %f261;

dp4a.s32.s32 %r379, %r456, %r343, %r462;

mul.lo.s32 %r560, %r379, %r531;
cvt.rn.f32.s32 %f266, %r560;
fma.rn.ftz.f32 %f267, %f106, %f266, %f263;

dp4a.s32.s32 %r383, %r460, %r343, %r462;

cvt.rn.f32.s32 %f268, %r383;
fma.rn.ftz.f32 %f269, %f106, %f268, %f265;

{.reg .f16 low,high;
mov.b32 {low,high},%r159;
cvt.f32.f16 %f107, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r159;
cvt.f32.f16 %f108, high;}


mul.ftz.f32 %f270, %f267, %f107;
mul.ftz.f32 %f271, %f269, %f108;
sub.ftz.f32 %f272, %f270, %f271;
add.ftz.f32 %f433, %f272, %f433;
cvt.u64.u32 %rd59, %r761;
add.s64 %rd60, %rd1, %rd59;
add.s64 %rd61, %rd60, %rd4;
mul.lo.s64 %rd62, %rd61, 36;
add.s64 %rd63, %rd11, %rd62;
add.s64 %rd64, %rd63, %rd34;
ld.global.nc.u32 %r395, [%rd64+4];
ld.global.nc.u32 %r389, [%rd63];

{.reg .f16 low,high;
mov.b32 {low,high},%r389;
cvt.f32.f16 %f109, low;}


ld.global.nc.u32 %r403, [%rd64+40];
ld.global.nc.u32 %r390, [%rd63+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r390;
cvt.f32.f16 %f110, low;}


ld.global.nc.u32 %r411, [%rd64+76];
ld.global.nc.u32 %r391, [%rd63+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r391;
cvt.f32.f16 %f111, low;}


ld.global.nc.u32 %r419, [%rd64+112];
ld.global.nc.u32 %r392, [%rd63+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r392;
cvt.f32.f16 %f112, low;}



dp4a.s32.s32 %r393, %r394, %r395, %r462;

mul.lo.s32 %r561, %r393, %r468;
cvt.rn.f32.s32 %f273, %r561;
fma.rn.ftz.f32 %f274, %f109, %f273, 0f00000000;

dp4a.s32.s32 %r397, %r398, %r395, %r462;

cvt.rn.f32.s32 %f275, %r397;
fma.rn.ftz.f32 %f276, %f109, %f275, 0f00000000;

dp4a.s32.s32 %r401, %r402, %r403, %r462;

mul.lo.s32 %r562, %r401, %r477;
cvt.rn.f32.s32 %f277, %r562;
fma.rn.ftz.f32 %f278, %f110, %f277, %f274;

dp4a.s32.s32 %r405, %r406, %r403, %r462;

cvt.rn.f32.s32 %f279, %r405;
fma.rn.ftz.f32 %f280, %f110, %f279, %f276;

dp4a.s32.s32 %r409, %r410, %r411, %r462;

mul.lo.s32 %r563, %r409, %r486;
cvt.rn.f32.s32 %f281, %r563;
fma.rn.ftz.f32 %f282, %f111, %f281, %f278;

dp4a.s32.s32 %r413, %r414, %r411, %r462;

cvt.rn.f32.s32 %f283, %r413;
fma.rn.ftz.f32 %f284, %f111, %f283, %f280;

dp4a.s32.s32 %r417, %r418, %r419, %r462;

mul.lo.s32 %r564, %r417, %r495;
cvt.rn.f32.s32 %f285, %r564;
fma.rn.ftz.f32 %f286, %f112, %f285, %f282;

dp4a.s32.s32 %r421, %r422, %r419, %r462;

cvt.rn.f32.s32 %f287, %r421;
fma.rn.ftz.f32 %f288, %f112, %f287, %f284;

{.reg .f16 low,high;
mov.b32 {low,high},%r121;
cvt.f32.f16 %f113, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r121;
cvt.f32.f16 %f114, high;}


mul.ftz.f32 %f289, %f286, %f113;
mul.ftz.f32 %f290, %f288, %f114;
sub.ftz.f32 %f291, %f289, %f290;
add.ftz.f32 %f432, %f291, %f432;

{.reg .f16 low,high;
mov.b32 {low,high},%r389;
cvt.f32.f16 %f115, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r390;
cvt.f32.f16 %f116, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r391;
cvt.f32.f16 %f117, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r392;
cvt.f32.f16 %f118, low;}



dp4a.s32.s32 %r431, %r432, %r395, %r462;

mul.lo.s32 %r565, %r431, %r504;
cvt.rn.f32.s32 %f292, %r565;
fma.rn.ftz.f32 %f293, %f115, %f292, 0f00000000;

dp4a.s32.s32 %r435, %r436, %r395, %r462;

cvt.rn.f32.s32 %f294, %r435;
fma.rn.ftz.f32 %f295, %f115, %f294, 0f00000000;

dp4a.s32.s32 %r439, %r440, %r403, %r462;

mul.lo.s32 %r566, %r439, %r513;
cvt.rn.f32.s32 %f296, %r566;
fma.rn.ftz.f32 %f297, %f116, %f296, %f293;

dp4a.s32.s32 %r443, %r444, %r403, %r462;

cvt.rn.f32.s32 %f298, %r443;
fma.rn.ftz.f32 %f299, %f116, %f298, %f295;

dp4a.s32.s32 %r447, %r448, %r411, %r462;

mul.lo.s32 %r567, %r447, %r522;
cvt.rn.f32.s32 %f300, %r567;
fma.rn.ftz.f32 %f301, %f117, %f300, %f297;

dp4a.s32.s32 %r451, %r452, %r411, %r462;

cvt.rn.f32.s32 %f302, %r451;
fma.rn.ftz.f32 %f303, %f117, %f302, %f299;

dp4a.s32.s32 %r455, %r456, %r419, %r462;

mul.lo.s32 %r568, %r455, %r531;
cvt.rn.f32.s32 %f304, %r568;
fma.rn.ftz.f32 %f305, %f118, %f304, %f301;

dp4a.s32.s32 %r459, %r460, %r419, %r462;

cvt.rn.f32.s32 %f306, %r459;
fma.rn.ftz.f32 %f307, %f118, %f306, %f303;

{.reg .f16 low,high;
mov.b32 {low,high},%r159;
cvt.f32.f16 %f119, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r159;
cvt.f32.f16 %f120, high;}


mul.ftz.f32 %f308, %f305, %f119;
mul.ftz.f32 %f309, %f307, %f120;
sub.ftz.f32 %f310, %f308, %f309;
add.ftz.f32 %f431, %f310, %f431;
add.s32 %r763, %r763, 32;
add.s32 %r762, %r762, 32;
add.s32 %r761, %r761, 32;
add.s32 %r760, %r760, 32;
add.s32 %r759, %r759, 32;
add.s64 %rd95, %rd95, 336;
add.s32 %r764, %r764, 4;
setp.lt.s32 %p4, %r764, %r6;
@%p4 bra $L__BB136_2;

$L__BB136_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB136_5;

mov.u32 %r570, _ZZ13mul_mat_vec_qIL9ggml_type10ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r571, %r1, 1280, %r570;
shl.b32 %r572, %r43, 2;
add.s32 %r573, %r571, %r572;
st.shared.f32 [%r573+-1280], %f440;
st.shared.f32 [%r573+-1152], %f439;
st.shared.f32 [%r573+-1024], %f438;
st.shared.f32 [%r573+-896], %f437;
st.shared.f32 [%r573+-768], %f436;
st.shared.f32 [%r573+-640], %f435;
st.shared.f32 [%r573+-512], %f434;
st.shared.f32 [%r573+-384], %f433;
st.shared.f32 [%r573+-256], %f432;
st.shared.f32 [%r573+-128], %f431;

$L__BB136_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB136_26;

mov.u32 %r574, %ctaid.x;
shl.b32 %r575, %r574, 1;
mov.u32 %r576, 1;
mad.lo.s32 %r577, %r3, %r34, %r575;
mad.lo.s32 %r578, %r4, %r40, %r577;
cvt.u64.u32 %rd14, %r578;
shl.b32 %r580, %r43, 2;
mov.u32 %r581, _ZZ13mul_mat_vec_qIL9ggml_type10ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r24, %r581, %r580;
mov.u32 %r582, 2;
ld.shared.f32 %f311, [%r24];
add.ftz.f32 %f312, %f311, %f440;
mov.b32 %r583, %f312;
mov.u32 %r584, 31;
mov.u32 %r585, 16;
mov.u32 %r586, -1;
shfl.sync.bfly.b32 %r587|%p7, %r583, %r585, %r584, %r586;
mov.b32 %f313, %r587;
add.ftz.f32 %f314, %f312, %f313;
mov.b32 %r588, %f314;
mov.u32 %r589, 8;
shfl.sync.bfly.b32 %r590|%p8, %r588, %r589, %r584, %r586;
mov.b32 %f315, %r590;
add.ftz.f32 %f316, %f314, %f315;
mov.b32 %r591, %f316;
mov.u32 %r592, 4;
shfl.sync.bfly.b32 %r593|%p9, %r591, %r592, %r584, %r586;
mov.b32 %f317, %r593;
add.ftz.f32 %f318, %f316, %f317;
mov.b32 %r594, %f318;
shfl.sync.bfly.b32 %r595|%p10, %r594, %r582, %r584, %r586;
mov.b32 %f319, %r595;
add.ftz.f32 %f320, %f318, %f319;
mov.b32 %r596, %f320;
shfl.sync.bfly.b32 %r597|%p11, %r596, %r576, %r584, %r586;
mov.b32 %f321, %r597;
add.ftz.f32 %f31, %f320, %f321;
setp.eq.s32 %p12, %r43, 0;
setp.lt.u32 %p13, %r575, %r28;
and.pred %p1, %p12, %p13;
cvta.to.global.u64 %rd65, %rd20;
mul.wide.u32 %rd66, %r578, 4;
add.s64 %rd15, %rd65, %rd66;
not.pred %p14, %p1;
@%p14 bra $L__BB136_8;

st.global.f32 [%rd15], %f31;

$L__BB136_8:
ld.shared.f32 %f322, [%r24+128];
add.ftz.f32 %f323, %f322, %f439;
mov.b32 %r598, %f323;
shfl.sync.bfly.b32 %r602|%p15, %r598, %r585, %r584, %r586;
mov.b32 %f324, %r602;
add.ftz.f32 %f325, %f323, %f324;
mov.b32 %r603, %f325;
shfl.sync.bfly.b32 %r605|%p16, %r603, %r589, %r584, %r586;
mov.b32 %f326, %r605;
add.ftz.f32 %f327, %f325, %f326;
mov.b32 %r606, %f327;
shfl.sync.bfly.b32 %r608|%p17, %r606, %r592, %r584, %r586;
mov.b32 %f328, %r608;
add.ftz.f32 %f329, %f327, %f328;
mov.b32 %r609, %f329;
shfl.sync.bfly.b32 %r611|%p18, %r609, %r582, %r584, %r586;
mov.b32 %f330, %r611;
add.ftz.f32 %f331, %f329, %f330;
mov.b32 %r612, %f331;
shfl.sync.bfly.b32 %r614|%p19, %r612, %r576, %r584, %r586;
mov.b32 %f332, %r614;
add.ftz.f32 %f32, %f331, %f332;
setp.eq.s32 %p20, %r43, 1;
or.b32 %r618, %r575, 1;
setp.lt.u32 %p21, %r618, %r28;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB136_10;

st.global.f32 [%rd15+4], %f32;

$L__BB136_10:
ld.shared.f32 %f333, [%r24+256];
add.ftz.f32 %f334, %f333, %f438;
mov.b32 %r619, %f334;
mov.u32 %r620, 31;
mov.u32 %r621, 16;
mov.u32 %r622, -1;
shfl.sync.bfly.b32 %r623|%p23, %r619, %r621, %r620, %r622;
mov.b32 %f335, %r623;
add.ftz.f32 %f336, %f334, %f335;
mov.b32 %r624, %f336;
mov.u32 %r625, 8;
shfl.sync.bfly.b32 %r626|%p24, %r624, %r625, %r620, %r622;
mov.b32 %f337, %r626;
add.ftz.f32 %f338, %f336, %f337;
mov.b32 %r627, %f338;
mov.u32 %r628, 4;
shfl.sync.bfly.b32 %r629|%p25, %r627, %r628, %r620, %r622;
mov.b32 %f339, %r629;
add.ftz.f32 %f340, %f338, %f339;
mov.b32 %r630, %f340;
mov.u32 %r631, 2;
shfl.sync.bfly.b32 %r632|%p26, %r630, %r631, %r620, %r622;
mov.b32 %f341, %r632;
add.ftz.f32 %f342, %f340, %f341;
mov.b32 %r633, %f342;
mov.u32 %r634, 1;
shfl.sync.bfly.b32 %r635|%p27, %r633, %r634, %r620, %r622;
mov.b32 %f343, %r635;
add.ftz.f32 %f33, %f342, %f343;
@%p14 bra $L__BB136_12;

cvt.u64.u32 %rd67, %r28;
add.s64 %rd68, %rd67, %rd14;
shl.b64 %rd70, %rd68, 2;
add.s64 %rd71, %rd65, %rd70;
st.global.f32 [%rd71], %f33;

$L__BB136_12:
ld.shared.f32 %f344, [%r24+384];
add.ftz.f32 %f345, %f344, %f437;
mov.b32 %r636, %f345;
shfl.sync.bfly.b32 %r640|%p29, %r636, %r621, %r620, %r622;
mov.b32 %f346, %r640;
add.ftz.f32 %f347, %f345, %f346;
mov.b32 %r641, %f347;
shfl.sync.bfly.b32 %r643|%p30, %r641, %r625, %r620, %r622;
mov.b32 %f348, %r643;
add.ftz.f32 %f349, %f347, %f348;
mov.b32 %r644, %f349;
shfl.sync.bfly.b32 %r646|%p31, %r644, %r628, %r620, %r622;
mov.b32 %f350, %r646;
add.ftz.f32 %f351, %f349, %f350;
mov.b32 %r647, %f351;
shfl.sync.bfly.b32 %r649|%p32, %r647, %r631, %r620, %r622;
mov.b32 %f352, %r649;
add.ftz.f32 %f353, %f351, %f352;
mov.b32 %r650, %f353;
shfl.sync.bfly.b32 %r652|%p33, %r650, %r634, %r620, %r622;
mov.b32 %f354, %r652;
add.ftz.f32 %f34, %f353, %f354;
@%p22 bra $L__BB136_14;

add.s32 %r653, %r28, 1;
cvt.u64.u32 %rd72, %r653;
add.s64 %rd73, %rd72, %rd14;
shl.b64 %rd75, %rd73, 2;
add.s64 %rd76, %rd65, %rd75;
st.global.f32 [%rd76], %f34;

$L__BB136_14:
shl.b32 %r654, %r28, 1;
mov.u32 %r655, 1;
ld.shared.f32 %f355, [%r24+512];
add.ftz.f32 %f356, %f355, %f436;
mov.b32 %r656, %f356;
mov.u32 %r657, 31;
mov.u32 %r658, 16;
mov.u32 %r659, -1;
shfl.sync.bfly.b32 %r660|%p35, %r656, %r658, %r657, %r659;
mov.b32 %f357, %r660;
add.ftz.f32 %f358, %f356, %f357;
mov.b32 %r661, %f358;
mov.u32 %r662, 8;
shfl.sync.bfly.b32 %r663|%p36, %r661, %r662, %r657, %r659;
mov.b32 %f359, %r663;
add.ftz.f32 %f360, %f358, %f359;
mov.b32 %r664, %f360;
mov.u32 %r665, 4;
shfl.sync.bfly.b32 %r666|%p37, %r664, %r665, %r657, %r659;
mov.b32 %f361, %r666;
add.ftz.f32 %f362, %f360, %f361;
mov.b32 %r667, %f362;
mov.u32 %r668, 2;
shfl.sync.bfly.b32 %r669|%p38, %r667, %r668, %r657, %r659;
mov.b32 %f363, %r669;
add.ftz.f32 %f364, %f362, %f363;
mov.b32 %r670, %f364;
shfl.sync.bfly.b32 %r671|%p39, %r670, %r655, %r657, %r659;
mov.b32 %f365, %r671;
add.ftz.f32 %f35, %f364, %f365;
cvt.u64.u32 %rd77, %r654;
add.s64 %rd78, %rd77, %rd14;
shl.b64 %rd80, %rd78, 2;
add.s64 %rd16, %rd65, %rd80;
@%p14 bra $L__BB136_16;

st.global.f32 [%rd16], %f35;

$L__BB136_16:
ld.shared.f32 %f366, [%r24+640];
add.ftz.f32 %f367, %f366, %f435;
mov.b32 %r672, %f367;
shfl.sync.bfly.b32 %r676|%p41, %r672, %r658, %r657, %r659;
mov.b32 %f368, %r676;
add.ftz.f32 %f369, %f367, %f368;
mov.b32 %r677, %f369;
shfl.sync.bfly.b32 %r679|%p42, %r677, %r662, %r657, %r659;
mov.b32 %f370, %r679;
add.ftz.f32 %f371, %f369, %f370;
mov.b32 %r680, %f371;
shfl.sync.bfly.b32 %r682|%p43, %r680, %r665, %r657, %r659;
mov.b32 %f372, %r682;
add.ftz.f32 %f373, %f371, %f372;
mov.b32 %r683, %f373;
shfl.sync.bfly.b32 %r685|%p44, %r683, %r668, %r657, %r659;
mov.b32 %f374, %r685;
add.ftz.f32 %f375, %f373, %f374;
mov.b32 %r686, %f375;
shfl.sync.bfly.b32 %r688|%p45, %r686, %r655, %r657, %r659;
mov.b32 %f376, %r688;
add.ftz.f32 %f36, %f375, %f376;
@%p22 bra $L__BB136_18;

st.global.f32 [%rd16+4], %f36;

$L__BB136_18:
mul.lo.s32 %r25, %r28, 3;
ld.shared.f32 %f377, [%r24+768];
add.ftz.f32 %f378, %f377, %f434;
mov.b32 %r689, %f378;
mov.u32 %r690, 31;
mov.u32 %r691, 16;
mov.u32 %r692, -1;
shfl.sync.bfly.b32 %r693|%p47, %r689, %r691, %r690, %r692;
mov.b32 %f379, %r693;
add.ftz.f32 %f380, %f378, %f379;
mov.b32 %r694, %f380;
mov.u32 %r695, 8;
shfl.sync.bfly.b32 %r696|%p48, %r694, %r695, %r690, %r692;
mov.b32 %f381, %r696;
add.ftz.f32 %f382, %f380, %f381;
mov.b32 %r697, %f382;
mov.u32 %r698, 4;
shfl.sync.bfly.b32 %r699|%p49, %r697, %r698, %r690, %r692;
mov.b32 %f383, %r699;
add.ftz.f32 %f384, %f382, %f383;
mov.b32 %r700, %f384;
mov.u32 %r701, 2;
shfl.sync.bfly.b32 %r702|%p50, %r700, %r701, %r690, %r692;
mov.b32 %f385, %r702;
add.ftz.f32 %f386, %f384, %f385;
mov.b32 %r703, %f386;
mov.u32 %r704, 1;
shfl.sync.bfly.b32 %r705|%p51, %r703, %r704, %r690, %r692;
mov.b32 %f387, %r705;
add.ftz.f32 %f37, %f386, %f387;
@%p14 bra $L__BB136_20;

cvt.u64.u32 %rd81, %r25;
add.s64 %rd82, %rd81, %rd14;
shl.b64 %rd84, %rd82, 2;
add.s64 %rd85, %rd65, %rd84;
st.global.f32 [%rd85], %f37;

$L__BB136_20:
ld.shared.f32 %f388, [%r24+896];
add.ftz.f32 %f389, %f388, %f433;
mov.b32 %r706, %f389;
shfl.sync.bfly.b32 %r710|%p53, %r706, %r691, %r690, %r692;
mov.b32 %f390, %r710;
add.ftz.f32 %f391, %f389, %f390;
mov.b32 %r711, %f391;
shfl.sync.bfly.b32 %r713|%p54, %r711, %r695, %r690, %r692;
mov.b32 %f392, %r713;
add.ftz.f32 %f393, %f391, %f392;
mov.b32 %r714, %f393;
shfl.sync.bfly.b32 %r716|%p55, %r714, %r698, %r690, %r692;
mov.b32 %f394, %r716;
add.ftz.f32 %f395, %f393, %f394;
mov.b32 %r717, %f395;
shfl.sync.bfly.b32 %r719|%p56, %r717, %r701, %r690, %r692;
mov.b32 %f396, %r719;
add.ftz.f32 %f397, %f395, %f396;
mov.b32 %r720, %f397;
shfl.sync.bfly.b32 %r722|%p57, %r720, %r704, %r690, %r692;
mov.b32 %f398, %r722;
add.ftz.f32 %f38, %f397, %f398;
@%p22 bra $L__BB136_22;

add.s32 %r723, %r25, 1;
cvt.u64.u32 %rd86, %r723;
add.s64 %rd87, %rd86, %rd14;
shl.b64 %rd89, %rd87, 2;
add.s64 %rd90, %rd65, %rd89;
st.global.f32 [%rd90], %f38;

$L__BB136_22:
shl.b32 %r724, %r28, 2;
mov.u32 %r725, 2;
ld.shared.f32 %f399, [%r24+1024];
add.ftz.f32 %f400, %f399, %f432;
mov.b32 %r726, %f400;
mov.u32 %r727, 31;
mov.u32 %r728, 16;
mov.u32 %r729, -1;
shfl.sync.bfly.b32 %r730|%p59, %r726, %r728, %r727, %r729;
mov.b32 %f401, %r730;
add.ftz.f32 %f402, %f400, %f401;
mov.b32 %r731, %f402;
mov.u32 %r732, 8;
shfl.sync.bfly.b32 %r733|%p60, %r731, %r732, %r727, %r729;
mov.b32 %f403, %r733;
add.ftz.f32 %f404, %f402, %f403;
mov.b32 %r734, %f404;
mov.u32 %r735, 4;
shfl.sync.bfly.b32 %r736|%p61, %r734, %r735, %r727, %r729;
mov.b32 %f405, %r736;
add.ftz.f32 %f406, %f404, %f405;
mov.b32 %r737, %f406;
shfl.sync.bfly.b32 %r738|%p62, %r737, %r725, %r727, %r729;
mov.b32 %f407, %r738;
add.ftz.f32 %f408, %f406, %f407;
mov.b32 %r739, %f408;
mov.u32 %r740, 1;
shfl.sync.bfly.b32 %r741|%p63, %r739, %r740, %r727, %r729;
mov.b32 %f409, %r741;
add.ftz.f32 %f39, %f408, %f409;
cvt.u64.u32 %rd91, %r724;
add.s64 %rd92, %rd91, %rd14;
shl.b64 %rd94, %rd92, 2;
add.s64 %rd17, %rd65, %rd94;
@%p14 bra $L__BB136_24;

st.global.f32 [%rd17], %f39;

$L__BB136_24:
ld.shared.f32 %f410, [%r24+1152];
add.ftz.f32 %f411, %f410, %f431;
mov.b32 %r742, %f411;
shfl.sync.bfly.b32 %r746|%p65, %r742, %r728, %r727, %r729;
mov.b32 %f412, %r746;
add.ftz.f32 %f413, %f411, %f412;
mov.b32 %r747, %f413;
shfl.sync.bfly.b32 %r749|%p66, %r747, %r732, %r727, %r729;
mov.b32 %f414, %r749;
add.ftz.f32 %f415, %f413, %f414;
mov.b32 %r750, %f415;
shfl.sync.bfly.b32 %r752|%p67, %r750, %r735, %r727, %r729;
mov.b32 %f416, %r752;
add.ftz.f32 %f417, %f415, %f416;
mov.b32 %r753, %f417;
shfl.sync.bfly.b32 %r755|%p68, %r753, %r725, %r727, %r729;
mov.b32 %f418, %r755;
add.ftz.f32 %f419, %f417, %f418;
mov.b32 %r756, %f419;
shfl.sync.bfly.b32 %r758|%p69, %r756, %r740, %r727, %r729;
mov.b32 %f420, %r758;
add.ftz.f32 %f40, %f419, %f420;
@%p22 bra $L__BB136_26;

st.global.f32 [%rd17+4], %f40;

$L__BB136_26:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type10ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type10ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type10ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type10ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type10ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type10ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type10ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type10ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type10ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<83>;
.reg .b16 %rs<9>;
.reg .f32 %f<529>;
.reg .b32 %r<903>;
.reg .b64 %rd<115>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type10ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type10ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type10ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type10ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type10ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type10ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type10ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type10ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type10ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type10ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type10ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type10ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type10ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type10ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type10ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type10ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type10ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type10ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r44, %r1, 5;
mov.u32 %r45, %tid.x;
add.s32 %r2, %r44, %r45;
shr.s32 %r46, %r2, 31;
shr.u32 %r47, %r46, 28;
add.s32 %r48, %r2, %r47;
shr.s32 %r902, %r48, 4;
shr.u32 %r4, %r43, 8;
setp.ge.s32 %p3, %r902, %r4;
mov.f32 %f517, 0f00000000;
mov.f32 %f518, %f517;
mov.f32 %f519, %f517;
mov.f32 %f520, %f517;
mov.f32 %f521, %f517;
mov.f32 %f522, %f517;
mov.f32 %f523, %f517;
mov.f32 %f524, %f517;
mov.f32 %f525, %f517;
mov.f32 %f526, %f517;
mov.f32 %f527, %f517;
mov.f32 %f528, %f517;
@%p3 bra $L__BB137_3;

mov.u32 %r49, %ctaid.x;
shl.b32 %r50, %r49, 1;
mov.u32 %r51, %ctaid.y;
mov.u32 %r52, %ctaid.z;
mul.lo.s32 %r53, %r51, %r35;
cvt.u64.u32 %rd20, %r53;
mul.lo.s32 %r54, %r52, %r41;
cvt.u64.u32 %rd21, %r54;
add.s64 %rd1, %rd21, %rd20;
mul.hi.u32 %r55, %r52, %r37;
add.s32 %r56, %r55, %r52;
shr.u32 %r57, %r56, %r38;
mul.lo.s32 %r58, %r57, %r40;
mul.hi.u32 %r59, %r51, %r31;
add.s32 %r60, %r59, %r51;
shr.u32 %r61, %r60, %r32;
mul.lo.s32 %r62, %r61, %r34;
cvta.to.global.u64 %rd114, %rd17;
or.b32 %r63, %r50, 1;
and.b32 %r67, %r48, -16;
sub.s32 %r68, %r2, %r67;
shr.s32 %r69, %r68, 31;
shr.u32 %r70, %r69, 29;
add.s32 %r71, %r68, %r70;
and.b32 %r72, %r71, -8;
sub.s32 %r73, %r68, %r72;
shr.s32 %r74, %r73, 31;
shr.u32 %r75, %r74, 30;
add.s32 %r76, %r73, %r75;
shr.s32 %r77, %r76, 2;
add.s32 %r78, %r72, %r77;
cvt.s64.s32 %rd22, %r78;
shr.s32 %r79, %r71, 3;
shl.b32 %r80, %r79, 2;
cvt.s64.s32 %rd3, %r73;
cvt.s64.s32 %rd4, %r80;
shl.b32 %r81, %r29, 1;
shl.b32 %r896, %r902, 3;
add.s32 %r901, %r896, %r81;
mad.lo.s32 %r900, %r29, 3, %r896;
shl.b32 %r82, %r29, 2;
add.s32 %r899, %r896, %r82;
mad.lo.s32 %r898, %r29, 5, %r896;
add.s32 %r897, %r29, %r896;
add.s32 %r83, %r902, %r58;
add.s32 %r84, %r83, %r62;
mad.lo.s32 %r85, %r63, %r28, %r84;
mul.wide.s32 %rd5, %r85, 84;
mad.lo.s32 %r86, %r50, %r28, %r902;
add.s32 %r87, %r86, %r58;
add.s32 %r88, %r87, %r62;
mul.wide.s32 %rd6, %r88, 84;
mul.wide.s32 %rd23, %r68, 4;
add.s64 %rd24, %rd6, %rd23;
add.s64 %rd7, %rd24, 16;
add.s64 %rd25, %rd6, %rd22;
add.s64 %rd8, %rd25, 6;
add.s64 %rd9, %rd5, %rd22;
add.s64 %rd26, %rd5, %rd23;
add.s64 %rd10, %rd26, 16;
cvta.to.global.u64 %rd11, %rd18;
mov.f32 %f517, 0f00000000;
shl.b64 %rd33, %rd3, 2;

$L__BB137_2:
cvt.u64.u32 %rd27, %r896;
add.s64 %rd28, %rd1, %rd27;
add.s64 %rd29, %rd114, %rd7;
ld.global.nc.u32 %r545, [%rd29];
add.s64 %rd30, %rd28, %rd4;
mul.lo.s64 %rd31, %rd30, 36;
add.s64 %rd32, %rd11, %rd31;
add.s64 %rd34, %rd32, %rd33;
ld.global.nc.u32 %r95, [%rd34+4];
ld.global.nc.u32 %r89, [%rd32];

{.reg .f16 low,high;
mov.b32 {low,high},%r89;
cvt.f32.f16 %f73, low;}


ld.global.nc.u32 %r103, [%rd34+40];
ld.global.nc.u32 %r90, [%rd32+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r90;
cvt.f32.f16 %f74, low;}


ld.global.nc.u32 %r111, [%rd34+76];
ld.global.nc.u32 %r91, [%rd32+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r91;
cvt.f32.f16 %f75, low;}


ld.global.nc.u32 %r119, [%rd34+112];
ld.global.nc.u32 %r92, [%rd32+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r92;
cvt.f32.f16 %f76, low;}


add.s64 %rd35, %rd114, %rd8;
ld.global.nc.u8 %rs1, [%rd35+-6];
cvt.u32.u16 %r546, %rs1;
and.b32 %r547, %r546, 240;
and.b32 %r474, %r545, 50529027;
mov.u32 %r542, 0;

dp4a.s32.s32 %r93, %r474, %r95, %r542;

and.b32 %r548, %r546, 15;
mul.lo.s32 %r549, %r93, %r548;
cvt.rn.f32.s32 %f145, %r549;
fma.rn.ftz.f32 %f146, %f73, %f145, 0f00000000;
shr.u32 %r550, %r547, 4;
shl.b32 %r551, %r550, 8;
or.b32 %r552, %r551, %r550;
shl.b32 %r553, %r552, 16;
or.b32 %r478, %r553, %r552;

dp4a.s32.s32 %r97, %r478, %r95, %r542;

cvt.rn.f32.s32 %f147, %r97;
fma.rn.ftz.f32 %f148, %f73, %f147, 0f00000000;
ld.global.nc.u8 %rs2, [%rd35+-4];
cvt.u32.u16 %r554, %rs2;
and.b32 %r555, %r554, 240;
shr.u32 %r556, %r545, 2;
and.b32 %r482, %r556, 50529027;

dp4a.s32.s32 %r101, %r482, %r103, %r542;

and.b32 %r557, %r554, 15;
mul.lo.s32 %r558, %r101, %r557;
cvt.rn.f32.s32 %f149, %r558;
fma.rn.ftz.f32 %f150, %f74, %f149, %f146;
shr.u32 %r559, %r555, 4;
shl.b32 %r560, %r559, 8;
or.b32 %r561, %r560, %r559;
shl.b32 %r562, %r561, 16;
or.b32 %r486, %r562, %r561;

dp4a.s32.s32 %r105, %r486, %r103, %r542;

cvt.rn.f32.s32 %f151, %r105;
fma.rn.ftz.f32 %f152, %f74, %f151, %f148;
ld.global.nc.u8 %rs3, [%rd35+-2];
cvt.u32.u16 %r563, %rs3;
and.b32 %r564, %r563, 240;
shr.u32 %r565, %r545, 4;
and.b32 %r490, %r565, 50529027;

dp4a.s32.s32 %r109, %r490, %r111, %r542;

and.b32 %r566, %r563, 15;
mul.lo.s32 %r567, %r109, %r566;
cvt.rn.f32.s32 %f153, %r567;
fma.rn.ftz.f32 %f154, %f75, %f153, %f150;
shr.u32 %r568, %r564, 4;
shl.b32 %r569, %r568, 8;
or.b32 %r570, %r569, %r568;
shl.b32 %r571, %r570, 16;
or.b32 %r494, %r571, %r570;

dp4a.s32.s32 %r113, %r494, %r111, %r542;

cvt.rn.f32.s32 %f155, %r113;
fma.rn.ftz.f32 %f156, %f75, %f155, %f152;
ld.global.nc.u8 %rs4, [%rd35];
cvt.u32.u16 %r572, %rs4;
and.b32 %r573, %r572, 240;
shr.u32 %r574, %r545, 6;
and.b32 %r498, %r574, 50529027;

dp4a.s32.s32 %r117, %r498, %r119, %r542;

and.b32 %r575, %r572, 15;
mul.lo.s32 %r576, %r117, %r575;
cvt.rn.f32.s32 %f157, %r576;
fma.rn.ftz.f32 %f158, %f76, %f157, %f154;
shr.u32 %r577, %r573, 4;
shl.b32 %r578, %r577, 8;
or.b32 %r579, %r578, %r577;
shl.b32 %r580, %r579, 16;
or.b32 %r502, %r580, %r579;

dp4a.s32.s32 %r121, %r502, %r119, %r542;

cvt.rn.f32.s32 %f159, %r121;
fma.rn.ftz.f32 %f160, %f76, %f159, %f156;
add.s64 %rd36, %rd114, %rd6;
ld.global.nc.u32 %r125, [%rd36+80];

{.reg .f16 low,high;
mov.b32 {low,high},%r125;
cvt.f32.f16 %f77, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r125;
cvt.f32.f16 %f78, high;}


mul.ftz.f32 %f161, %f158, %f77;
mul.ftz.f32 %f162, %f160, %f78;
sub.ftz.f32 %f163, %f161, %f162;
add.ftz.f32 %f528, %f163, %f528;
add.s64 %rd37, %rd114, %rd10;
ld.global.nc.u32 %r581, [%rd37];

{.reg .f16 low,high;
mov.b32 {low,high},%r89;
cvt.f32.f16 %f79, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r90;
cvt.f32.f16 %f80, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r91;
cvt.f32.f16 %f81, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r92;
cvt.f32.f16 %f82, low;}


add.s64 %rd38, %rd114, %rd9;
ld.global.nc.u8 %rs5, [%rd38];
cvt.u32.u16 %r582, %rs5;
and.b32 %r583, %r582, 240;
and.b32 %r512, %r581, 50529027;

dp4a.s32.s32 %r131, %r512, %r95, %r542;

and.b32 %r584, %r582, 15;
mul.lo.s32 %r585, %r131, %r584;
cvt.rn.f32.s32 %f164, %r585;
fma.rn.ftz.f32 %f165, %f79, %f164, 0f00000000;
shr.u32 %r586, %r583, 4;
shl.b32 %r587, %r586, 8;
or.b32 %r588, %r587, %r586;
shl.b32 %r589, %r588, 16;
or.b32 %r516, %r589, %r588;

dp4a.s32.s32 %r135, %r516, %r95, %r542;

cvt.rn.f32.s32 %f166, %r135;
fma.rn.ftz.f32 %f167, %f79, %f166, 0f00000000;
ld.global.nc.u8 %rs6, [%rd38+2];
cvt.u32.u16 %r590, %rs6;
and.b32 %r591, %r590, 240;
shr.u32 %r592, %r581, 2;
and.b32 %r520, %r592, 50529027;

dp4a.s32.s32 %r139, %r520, %r103, %r542;

and.b32 %r593, %r590, 15;
mul.lo.s32 %r594, %r139, %r593;
cvt.rn.f32.s32 %f168, %r594;
fma.rn.ftz.f32 %f169, %f80, %f168, %f165;
shr.u32 %r595, %r591, 4;
shl.b32 %r596, %r595, 8;
or.b32 %r597, %r596, %r595;
shl.b32 %r598, %r597, 16;
or.b32 %r524, %r598, %r597;

dp4a.s32.s32 %r143, %r524, %r103, %r542;

cvt.rn.f32.s32 %f170, %r143;
fma.rn.ftz.f32 %f171, %f80, %f170, %f167;
ld.global.nc.u8 %rs7, [%rd38+4];
cvt.u32.u16 %r599, %rs7;
and.b32 %r600, %r599, 240;
shr.u32 %r601, %r581, 4;
and.b32 %r528, %r601, 50529027;

dp4a.s32.s32 %r147, %r528, %r111, %r542;

and.b32 %r602, %r599, 15;
mul.lo.s32 %r603, %r147, %r602;
cvt.rn.f32.s32 %f172, %r603;
fma.rn.ftz.f32 %f173, %f81, %f172, %f169;
shr.u32 %r604, %r600, 4;
shl.b32 %r605, %r604, 8;
or.b32 %r606, %r605, %r604;
shl.b32 %r607, %r606, 16;
or.b32 %r532, %r607, %r606;

dp4a.s32.s32 %r151, %r532, %r111, %r542;

cvt.rn.f32.s32 %f174, %r151;
fma.rn.ftz.f32 %f175, %f81, %f174, %f171;
ld.global.nc.u8 %rs8, [%rd38+6];
cvt.u32.u16 %r608, %rs8;
and.b32 %r609, %r608, 240;
shr.u32 %r610, %r581, 6;
and.b32 %r536, %r610, 50529027;

dp4a.s32.s32 %r155, %r536, %r119, %r542;

and.b32 %r611, %r608, 15;
mul.lo.s32 %r612, %r155, %r611;
cvt.rn.f32.s32 %f176, %r612;
fma.rn.ftz.f32 %f177, %f82, %f176, %f173;
shr.u32 %r613, %r609, 4;
shl.b32 %r614, %r613, 8;
or.b32 %r615, %r614, %r613;
shl.b32 %r616, %r615, 16;
or.b32 %r540, %r616, %r615;

dp4a.s32.s32 %r159, %r540, %r119, %r542;

cvt.rn.f32.s32 %f178, %r159;
fma.rn.ftz.f32 %f179, %f82, %f178, %f175;
add.s64 %rd39, %rd114, %rd5;
ld.global.nc.u32 %r163, [%rd39+80];

{.reg .f16 low,high;
mov.b32 {low,high},%r163;
cvt.f32.f16 %f83, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r163;
cvt.f32.f16 %f84, high;}


mul.ftz.f32 %f180, %f177, %f83;
mul.ftz.f32 %f181, %f179, %f84;
sub.ftz.f32 %f182, %f180, %f181;
add.ftz.f32 %f527, %f182, %f527;
cvt.u64.u32 %rd40, %r897;
add.s64 %rd41, %rd1, %rd40;
add.s64 %rd42, %rd41, %rd4;
mul.lo.s64 %rd43, %rd42, 36;
add.s64 %rd44, %rd11, %rd43;
add.s64 %rd45, %rd44, %rd33;
ld.global.nc.u32 %r171, [%rd45+4];
ld.global.nc.u32 %r165, [%rd44];

{.reg .f16 low,high;
mov.b32 {low,high},%r165;
cvt.f32.f16 %f85, low;}


ld.global.nc.u32 %r179, [%rd45+40];
ld.global.nc.u32 %r166, [%rd44+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r166;
cvt.f32.f16 %f86, low;}


ld.global.nc.u32 %r187, [%rd45+76];
ld.global.nc.u32 %r167, [%rd44+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r167;
cvt.f32.f16 %f87, low;}


ld.global.nc.u32 %r195, [%rd45+112];
ld.global.nc.u32 %r168, [%rd44+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r168;
cvt.f32.f16 %f88, low;}



dp4a.s32.s32 %r169, %r474, %r171, %r542;

mul.lo.s32 %r617, %r169, %r548;
cvt.rn.f32.s32 %f183, %r617;
fma.rn.ftz.f32 %f184, %f85, %f183, 0f00000000;

dp4a.s32.s32 %r173, %r478, %r171, %r542;

cvt.rn.f32.s32 %f185, %r173;
fma.rn.ftz.f32 %f186, %f85, %f185, 0f00000000;

dp4a.s32.s32 %r177, %r482, %r179, %r542;

mul.lo.s32 %r618, %r177, %r557;
cvt.rn.f32.s32 %f187, %r618;
fma.rn.ftz.f32 %f188, %f86, %f187, %f184;

dp4a.s32.s32 %r181, %r486, %r179, %r542;

cvt.rn.f32.s32 %f189, %r181;
fma.rn.ftz.f32 %f190, %f86, %f189, %f186;

dp4a.s32.s32 %r185, %r490, %r187, %r542;

mul.lo.s32 %r619, %r185, %r566;
cvt.rn.f32.s32 %f191, %r619;
fma.rn.ftz.f32 %f192, %f87, %f191, %f188;

dp4a.s32.s32 %r189, %r494, %r187, %r542;

cvt.rn.f32.s32 %f193, %r189;
fma.rn.ftz.f32 %f194, %f87, %f193, %f190;

dp4a.s32.s32 %r193, %r498, %r195, %r542;

mul.lo.s32 %r620, %r193, %r575;
cvt.rn.f32.s32 %f195, %r620;
fma.rn.ftz.f32 %f196, %f88, %f195, %f192;

dp4a.s32.s32 %r197, %r502, %r195, %r542;

cvt.rn.f32.s32 %f197, %r197;
fma.rn.ftz.f32 %f198, %f88, %f197, %f194;

{.reg .f16 low,high;
mov.b32 {low,high},%r125;
cvt.f32.f16 %f89, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r125;
cvt.f32.f16 %f90, high;}


mul.ftz.f32 %f199, %f196, %f89;
mul.ftz.f32 %f200, %f198, %f90;
sub.ftz.f32 %f201, %f199, %f200;
add.ftz.f32 %f526, %f201, %f526;

{.reg .f16 low,high;
mov.b32 {low,high},%r165;
cvt.f32.f16 %f91, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r166;
cvt.f32.f16 %f92, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r167;
cvt.f32.f16 %f93, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r168;
cvt.f32.f16 %f94, low;}



dp4a.s32.s32 %r207, %r512, %r171, %r542;

mul.lo.s32 %r621, %r207, %r584;
cvt.rn.f32.s32 %f202, %r621;
fma.rn.ftz.f32 %f203, %f91, %f202, 0f00000000;

dp4a.s32.s32 %r211, %r516, %r171, %r542;

cvt.rn.f32.s32 %f204, %r211;
fma.rn.ftz.f32 %f205, %f91, %f204, 0f00000000;

dp4a.s32.s32 %r215, %r520, %r179, %r542;

mul.lo.s32 %r622, %r215, %r593;
cvt.rn.f32.s32 %f206, %r622;
fma.rn.ftz.f32 %f207, %f92, %f206, %f203;

dp4a.s32.s32 %r219, %r524, %r179, %r542;

cvt.rn.f32.s32 %f208, %r219;
fma.rn.ftz.f32 %f209, %f92, %f208, %f205;

dp4a.s32.s32 %r223, %r528, %r187, %r542;

mul.lo.s32 %r623, %r223, %r602;
cvt.rn.f32.s32 %f210, %r623;
fma.rn.ftz.f32 %f211, %f93, %f210, %f207;

dp4a.s32.s32 %r227, %r532, %r187, %r542;

cvt.rn.f32.s32 %f212, %r227;
fma.rn.ftz.f32 %f213, %f93, %f212, %f209;

dp4a.s32.s32 %r231, %r536, %r195, %r542;

mul.lo.s32 %r624, %r231, %r611;
cvt.rn.f32.s32 %f214, %r624;
fma.rn.ftz.f32 %f215, %f94, %f214, %f211;

dp4a.s32.s32 %r235, %r540, %r195, %r542;

cvt.rn.f32.s32 %f216, %r235;
fma.rn.ftz.f32 %f217, %f94, %f216, %f213;

{.reg .f16 low,high;
mov.b32 {low,high},%r163;
cvt.f32.f16 %f95, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r163;
cvt.f32.f16 %f96, high;}


mul.ftz.f32 %f218, %f215, %f95;
mul.ftz.f32 %f219, %f217, %f96;
sub.ftz.f32 %f220, %f218, %f219;
add.ftz.f32 %f525, %f220, %f525;
cvt.u64.u32 %rd46, %r901;
add.s64 %rd47, %rd1, %rd46;
add.s64 %rd48, %rd47, %rd4;
mul.lo.s64 %rd49, %rd48, 36;
add.s64 %rd50, %rd11, %rd49;
add.s64 %rd51, %rd50, %rd33;
ld.global.nc.u32 %r247, [%rd51+4];
ld.global.nc.u32 %r241, [%rd50];

{.reg .f16 low,high;
mov.b32 {low,high},%r241;
cvt.f32.f16 %f97, low;}


ld.global.nc.u32 %r255, [%rd51+40];
ld.global.nc.u32 %r242, [%rd50+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r242;
cvt.f32.f16 %f98, low;}


ld.global.nc.u32 %r263, [%rd51+76];
ld.global.nc.u32 %r243, [%rd50+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r243;
cvt.f32.f16 %f99, low;}


ld.global.nc.u32 %r271, [%rd51+112];
ld.global.nc.u32 %r244, [%rd50+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r244;
cvt.f32.f16 %f100, low;}



dp4a.s32.s32 %r245, %r474, %r247, %r542;

mul.lo.s32 %r625, %r245, %r548;
cvt.rn.f32.s32 %f221, %r625;
fma.rn.ftz.f32 %f222, %f97, %f221, 0f00000000;

dp4a.s32.s32 %r249, %r478, %r247, %r542;

cvt.rn.f32.s32 %f223, %r249;
fma.rn.ftz.f32 %f224, %f97, %f223, 0f00000000;

dp4a.s32.s32 %r253, %r482, %r255, %r542;

mul.lo.s32 %r626, %r253, %r557;
cvt.rn.f32.s32 %f225, %r626;
fma.rn.ftz.f32 %f226, %f98, %f225, %f222;

dp4a.s32.s32 %r257, %r486, %r255, %r542;

cvt.rn.f32.s32 %f227, %r257;
fma.rn.ftz.f32 %f228, %f98, %f227, %f224;

dp4a.s32.s32 %r261, %r490, %r263, %r542;

mul.lo.s32 %r627, %r261, %r566;
cvt.rn.f32.s32 %f229, %r627;
fma.rn.ftz.f32 %f230, %f99, %f229, %f226;

dp4a.s32.s32 %r265, %r494, %r263, %r542;

cvt.rn.f32.s32 %f231, %r265;
fma.rn.ftz.f32 %f232, %f99, %f231, %f228;

dp4a.s32.s32 %r269, %r498, %r271, %r542;

mul.lo.s32 %r628, %r269, %r575;
cvt.rn.f32.s32 %f233, %r628;
fma.rn.ftz.f32 %f234, %f100, %f233, %f230;

dp4a.s32.s32 %r273, %r502, %r271, %r542;

cvt.rn.f32.s32 %f235, %r273;
fma.rn.ftz.f32 %f236, %f100, %f235, %f232;

{.reg .f16 low,high;
mov.b32 {low,high},%r125;
cvt.f32.f16 %f101, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r125;
cvt.f32.f16 %f102, high;}


mul.ftz.f32 %f237, %f234, %f101;
mul.ftz.f32 %f238, %f236, %f102;
sub.ftz.f32 %f239, %f237, %f238;
add.ftz.f32 %f524, %f239, %f524;

{.reg .f16 low,high;
mov.b32 {low,high},%r241;
cvt.f32.f16 %f103, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r242;
cvt.f32.f16 %f104, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r243;
cvt.f32.f16 %f105, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r244;
cvt.f32.f16 %f106, low;}



dp4a.s32.s32 %r283, %r512, %r247, %r542;

mul.lo.s32 %r629, %r283, %r584;
cvt.rn.f32.s32 %f240, %r629;
fma.rn.ftz.f32 %f241, %f103, %f240, 0f00000000;

dp4a.s32.s32 %r287, %r516, %r247, %r542;

cvt.rn.f32.s32 %f242, %r287;
fma.rn.ftz.f32 %f243, %f103, %f242, 0f00000000;

dp4a.s32.s32 %r291, %r520, %r255, %r542;

mul.lo.s32 %r630, %r291, %r593;
cvt.rn.f32.s32 %f244, %r630;
fma.rn.ftz.f32 %f245, %f104, %f244, %f241;

dp4a.s32.s32 %r295, %r524, %r255, %r542;

cvt.rn.f32.s32 %f246, %r295;
fma.rn.ftz.f32 %f247, %f104, %f246, %f243;

dp4a.s32.s32 %r299, %r528, %r263, %r542;

mul.lo.s32 %r631, %r299, %r602;
cvt.rn.f32.s32 %f248, %r631;
fma.rn.ftz.f32 %f249, %f105, %f248, %f245;

dp4a.s32.s32 %r303, %r532, %r263, %r542;

cvt.rn.f32.s32 %f250, %r303;
fma.rn.ftz.f32 %f251, %f105, %f250, %f247;

dp4a.s32.s32 %r307, %r536, %r271, %r542;

mul.lo.s32 %r632, %r307, %r611;
cvt.rn.f32.s32 %f252, %r632;
fma.rn.ftz.f32 %f253, %f106, %f252, %f249;

dp4a.s32.s32 %r311, %r540, %r271, %r542;

cvt.rn.f32.s32 %f254, %r311;
fma.rn.ftz.f32 %f255, %f106, %f254, %f251;

{.reg .f16 low,high;
mov.b32 {low,high},%r163;
cvt.f32.f16 %f107, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r163;
cvt.f32.f16 %f108, high;}


mul.ftz.f32 %f256, %f253, %f107;
mul.ftz.f32 %f257, %f255, %f108;
sub.ftz.f32 %f258, %f256, %f257;
add.ftz.f32 %f523, %f258, %f523;
cvt.u64.u32 %rd52, %r900;
add.s64 %rd53, %rd1, %rd52;
add.s64 %rd54, %rd53, %rd4;
mul.lo.s64 %rd55, %rd54, 36;
add.s64 %rd56, %rd11, %rd55;
add.s64 %rd57, %rd56, %rd33;
ld.global.nc.u32 %r323, [%rd57+4];
ld.global.nc.u32 %r317, [%rd56];

{.reg .f16 low,high;
mov.b32 {low,high},%r317;
cvt.f32.f16 %f109, low;}


ld.global.nc.u32 %r331, [%rd57+40];
ld.global.nc.u32 %r318, [%rd56+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r318;
cvt.f32.f16 %f110, low;}


ld.global.nc.u32 %r339, [%rd57+76];
ld.global.nc.u32 %r319, [%rd56+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r319;
cvt.f32.f16 %f111, low;}


ld.global.nc.u32 %r347, [%rd57+112];
ld.global.nc.u32 %r320, [%rd56+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r320;
cvt.f32.f16 %f112, low;}



dp4a.s32.s32 %r321, %r474, %r323, %r542;

mul.lo.s32 %r633, %r321, %r548;
cvt.rn.f32.s32 %f259, %r633;
fma.rn.ftz.f32 %f260, %f109, %f259, 0f00000000;

dp4a.s32.s32 %r325, %r478, %r323, %r542;

cvt.rn.f32.s32 %f261, %r325;
fma.rn.ftz.f32 %f262, %f109, %f261, 0f00000000;

dp4a.s32.s32 %r329, %r482, %r331, %r542;

mul.lo.s32 %r634, %r329, %r557;
cvt.rn.f32.s32 %f263, %r634;
fma.rn.ftz.f32 %f264, %f110, %f263, %f260;

dp4a.s32.s32 %r333, %r486, %r331, %r542;

cvt.rn.f32.s32 %f265, %r333;
fma.rn.ftz.f32 %f266, %f110, %f265, %f262;

dp4a.s32.s32 %r337, %r490, %r339, %r542;

mul.lo.s32 %r635, %r337, %r566;
cvt.rn.f32.s32 %f267, %r635;
fma.rn.ftz.f32 %f268, %f111, %f267, %f264;

dp4a.s32.s32 %r341, %r494, %r339, %r542;

cvt.rn.f32.s32 %f269, %r341;
fma.rn.ftz.f32 %f270, %f111, %f269, %f266;

dp4a.s32.s32 %r345, %r498, %r347, %r542;

mul.lo.s32 %r636, %r345, %r575;
cvt.rn.f32.s32 %f271, %r636;
fma.rn.ftz.f32 %f272, %f112, %f271, %f268;

dp4a.s32.s32 %r349, %r502, %r347, %r542;

cvt.rn.f32.s32 %f273, %r349;
fma.rn.ftz.f32 %f274, %f112, %f273, %f270;

{.reg .f16 low,high;
mov.b32 {low,high},%r125;
cvt.f32.f16 %f113, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r125;
cvt.f32.f16 %f114, high;}


mul.ftz.f32 %f275, %f272, %f113;
mul.ftz.f32 %f276, %f274, %f114;
sub.ftz.f32 %f277, %f275, %f276;
add.ftz.f32 %f522, %f277, %f522;

{.reg .f16 low,high;
mov.b32 {low,high},%r317;
cvt.f32.f16 %f115, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r318;
cvt.f32.f16 %f116, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r319;
cvt.f32.f16 %f117, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r320;
cvt.f32.f16 %f118, low;}



dp4a.s32.s32 %r359, %r512, %r323, %r542;

mul.lo.s32 %r637, %r359, %r584;
cvt.rn.f32.s32 %f278, %r637;
fma.rn.ftz.f32 %f279, %f115, %f278, 0f00000000;

dp4a.s32.s32 %r363, %r516, %r323, %r542;

cvt.rn.f32.s32 %f280, %r363;
fma.rn.ftz.f32 %f281, %f115, %f280, 0f00000000;

dp4a.s32.s32 %r367, %r520, %r331, %r542;

mul.lo.s32 %r638, %r367, %r593;
cvt.rn.f32.s32 %f282, %r638;
fma.rn.ftz.f32 %f283, %f116, %f282, %f279;

dp4a.s32.s32 %r371, %r524, %r331, %r542;

cvt.rn.f32.s32 %f284, %r371;
fma.rn.ftz.f32 %f285, %f116, %f284, %f281;

dp4a.s32.s32 %r375, %r528, %r339, %r542;

mul.lo.s32 %r639, %r375, %r602;
cvt.rn.f32.s32 %f286, %r639;
fma.rn.ftz.f32 %f287, %f117, %f286, %f283;

dp4a.s32.s32 %r379, %r532, %r339, %r542;

cvt.rn.f32.s32 %f288, %r379;
fma.rn.ftz.f32 %f289, %f117, %f288, %f285;

dp4a.s32.s32 %r383, %r536, %r347, %r542;

mul.lo.s32 %r640, %r383, %r611;
cvt.rn.f32.s32 %f290, %r640;
fma.rn.ftz.f32 %f291, %f118, %f290, %f287;

dp4a.s32.s32 %r387, %r540, %r347, %r542;

cvt.rn.f32.s32 %f292, %r387;
fma.rn.ftz.f32 %f293, %f118, %f292, %f289;

{.reg .f16 low,high;
mov.b32 {low,high},%r163;
cvt.f32.f16 %f119, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r163;
cvt.f32.f16 %f120, high;}


mul.ftz.f32 %f294, %f291, %f119;
mul.ftz.f32 %f295, %f293, %f120;
sub.ftz.f32 %f296, %f294, %f295;
add.ftz.f32 %f521, %f296, %f521;
cvt.u64.u32 %rd58, %r899;
add.s64 %rd59, %rd1, %rd58;
add.s64 %rd60, %rd59, %rd4;
mul.lo.s64 %rd61, %rd60, 36;
add.s64 %rd62, %rd11, %rd61;
add.s64 %rd63, %rd62, %rd33;
ld.global.nc.u32 %r399, [%rd63+4];
ld.global.nc.u32 %r393, [%rd62];

{.reg .f16 low,high;
mov.b32 {low,high},%r393;
cvt.f32.f16 %f121, low;}


ld.global.nc.u32 %r407, [%rd63+40];
ld.global.nc.u32 %r394, [%rd62+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r394;
cvt.f32.f16 %f122, low;}


ld.global.nc.u32 %r415, [%rd63+76];
ld.global.nc.u32 %r395, [%rd62+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r395;
cvt.f32.f16 %f123, low;}


ld.global.nc.u32 %r423, [%rd63+112];
ld.global.nc.u32 %r396, [%rd62+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r396;
cvt.f32.f16 %f124, low;}



dp4a.s32.s32 %r397, %r474, %r399, %r542;

mul.lo.s32 %r641, %r397, %r548;
cvt.rn.f32.s32 %f297, %r641;
fma.rn.ftz.f32 %f298, %f121, %f297, 0f00000000;

dp4a.s32.s32 %r401, %r478, %r399, %r542;

cvt.rn.f32.s32 %f299, %r401;
fma.rn.ftz.f32 %f300, %f121, %f299, 0f00000000;

dp4a.s32.s32 %r405, %r482, %r407, %r542;

mul.lo.s32 %r642, %r405, %r557;
cvt.rn.f32.s32 %f301, %r642;
fma.rn.ftz.f32 %f302, %f122, %f301, %f298;

dp4a.s32.s32 %r409, %r486, %r407, %r542;

cvt.rn.f32.s32 %f303, %r409;
fma.rn.ftz.f32 %f304, %f122, %f303, %f300;

dp4a.s32.s32 %r413, %r490, %r415, %r542;

mul.lo.s32 %r643, %r413, %r566;
cvt.rn.f32.s32 %f305, %r643;
fma.rn.ftz.f32 %f306, %f123, %f305, %f302;

dp4a.s32.s32 %r417, %r494, %r415, %r542;

cvt.rn.f32.s32 %f307, %r417;
fma.rn.ftz.f32 %f308, %f123, %f307, %f304;

dp4a.s32.s32 %r421, %r498, %r423, %r542;

mul.lo.s32 %r644, %r421, %r575;
cvt.rn.f32.s32 %f309, %r644;
fma.rn.ftz.f32 %f310, %f124, %f309, %f306;

dp4a.s32.s32 %r425, %r502, %r423, %r542;

cvt.rn.f32.s32 %f311, %r425;
fma.rn.ftz.f32 %f312, %f124, %f311, %f308;

{.reg .f16 low,high;
mov.b32 {low,high},%r125;
cvt.f32.f16 %f125, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r125;
cvt.f32.f16 %f126, high;}


mul.ftz.f32 %f313, %f310, %f125;
mul.ftz.f32 %f314, %f312, %f126;
sub.ftz.f32 %f315, %f313, %f314;
add.ftz.f32 %f520, %f315, %f520;

{.reg .f16 low,high;
mov.b32 {low,high},%r393;
cvt.f32.f16 %f127, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r394;
cvt.f32.f16 %f128, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r395;
cvt.f32.f16 %f129, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r396;
cvt.f32.f16 %f130, low;}



dp4a.s32.s32 %r435, %r512, %r399, %r542;

mul.lo.s32 %r645, %r435, %r584;
cvt.rn.f32.s32 %f316, %r645;
fma.rn.ftz.f32 %f317, %f127, %f316, 0f00000000;

dp4a.s32.s32 %r439, %r516, %r399, %r542;

cvt.rn.f32.s32 %f318, %r439;
fma.rn.ftz.f32 %f319, %f127, %f318, 0f00000000;

dp4a.s32.s32 %r443, %r520, %r407, %r542;

mul.lo.s32 %r646, %r443, %r593;
cvt.rn.f32.s32 %f320, %r646;
fma.rn.ftz.f32 %f321, %f128, %f320, %f317;

dp4a.s32.s32 %r447, %r524, %r407, %r542;

cvt.rn.f32.s32 %f322, %r447;
fma.rn.ftz.f32 %f323, %f128, %f322, %f319;

dp4a.s32.s32 %r451, %r528, %r415, %r542;

mul.lo.s32 %r647, %r451, %r602;
cvt.rn.f32.s32 %f324, %r647;
fma.rn.ftz.f32 %f325, %f129, %f324, %f321;

dp4a.s32.s32 %r455, %r532, %r415, %r542;

cvt.rn.f32.s32 %f326, %r455;
fma.rn.ftz.f32 %f327, %f129, %f326, %f323;

dp4a.s32.s32 %r459, %r536, %r423, %r542;

mul.lo.s32 %r648, %r459, %r611;
cvt.rn.f32.s32 %f328, %r648;
fma.rn.ftz.f32 %f329, %f130, %f328, %f325;

dp4a.s32.s32 %r463, %r540, %r423, %r542;

cvt.rn.f32.s32 %f330, %r463;
fma.rn.ftz.f32 %f331, %f130, %f330, %f327;

{.reg .f16 low,high;
mov.b32 {low,high},%r163;
cvt.f32.f16 %f131, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r163;
cvt.f32.f16 %f132, high;}


mul.ftz.f32 %f332, %f329, %f131;
mul.ftz.f32 %f333, %f331, %f132;
sub.ftz.f32 %f334, %f332, %f333;
add.ftz.f32 %f519, %f334, %f519;
cvt.u64.u32 %rd64, %r898;
add.s64 %rd65, %rd1, %rd64;
add.s64 %rd66, %rd65, %rd4;
mul.lo.s64 %rd67, %rd66, 36;
add.s64 %rd68, %rd11, %rd67;
add.s64 %rd69, %rd68, %rd33;
ld.global.nc.u32 %r475, [%rd69+4];
ld.global.nc.u32 %r469, [%rd68];

{.reg .f16 low,high;
mov.b32 {low,high},%r469;
cvt.f32.f16 %f133, low;}


ld.global.nc.u32 %r483, [%rd69+40];
ld.global.nc.u32 %r470, [%rd68+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r470;
cvt.f32.f16 %f134, low;}


ld.global.nc.u32 %r491, [%rd69+76];
ld.global.nc.u32 %r471, [%rd68+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r471;
cvt.f32.f16 %f135, low;}


ld.global.nc.u32 %r499, [%rd69+112];
ld.global.nc.u32 %r472, [%rd68+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r472;
cvt.f32.f16 %f136, low;}



dp4a.s32.s32 %r473, %r474, %r475, %r542;

mul.lo.s32 %r649, %r473, %r548;
cvt.rn.f32.s32 %f335, %r649;
fma.rn.ftz.f32 %f336, %f133, %f335, 0f00000000;

dp4a.s32.s32 %r477, %r478, %r475, %r542;

cvt.rn.f32.s32 %f337, %r477;
fma.rn.ftz.f32 %f338, %f133, %f337, 0f00000000;

dp4a.s32.s32 %r481, %r482, %r483, %r542;

mul.lo.s32 %r650, %r481, %r557;
cvt.rn.f32.s32 %f339, %r650;
fma.rn.ftz.f32 %f340, %f134, %f339, %f336;

dp4a.s32.s32 %r485, %r486, %r483, %r542;

cvt.rn.f32.s32 %f341, %r485;
fma.rn.ftz.f32 %f342, %f134, %f341, %f338;

dp4a.s32.s32 %r489, %r490, %r491, %r542;

mul.lo.s32 %r651, %r489, %r566;
cvt.rn.f32.s32 %f343, %r651;
fma.rn.ftz.f32 %f344, %f135, %f343, %f340;

dp4a.s32.s32 %r493, %r494, %r491, %r542;

cvt.rn.f32.s32 %f345, %r493;
fma.rn.ftz.f32 %f346, %f135, %f345, %f342;

dp4a.s32.s32 %r497, %r498, %r499, %r542;

mul.lo.s32 %r652, %r497, %r575;
cvt.rn.f32.s32 %f347, %r652;
fma.rn.ftz.f32 %f348, %f136, %f347, %f344;

dp4a.s32.s32 %r501, %r502, %r499, %r542;

cvt.rn.f32.s32 %f349, %r501;
fma.rn.ftz.f32 %f350, %f136, %f349, %f346;

{.reg .f16 low,high;
mov.b32 {low,high},%r125;
cvt.f32.f16 %f137, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r125;
cvt.f32.f16 %f138, high;}


mul.ftz.f32 %f351, %f348, %f137;
mul.ftz.f32 %f352, %f350, %f138;
sub.ftz.f32 %f353, %f351, %f352;
add.ftz.f32 %f518, %f353, %f518;

{.reg .f16 low,high;
mov.b32 {low,high},%r469;
cvt.f32.f16 %f139, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r470;
cvt.f32.f16 %f140, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r471;
cvt.f32.f16 %f141, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r472;
cvt.f32.f16 %f142, low;}



dp4a.s32.s32 %r511, %r512, %r475, %r542;

mul.lo.s32 %r653, %r511, %r584;
cvt.rn.f32.s32 %f354, %r653;
fma.rn.ftz.f32 %f355, %f139, %f354, 0f00000000;

dp4a.s32.s32 %r515, %r516, %r475, %r542;

cvt.rn.f32.s32 %f356, %r515;
fma.rn.ftz.f32 %f357, %f139, %f356, 0f00000000;

dp4a.s32.s32 %r519, %r520, %r483, %r542;

mul.lo.s32 %r654, %r519, %r593;
cvt.rn.f32.s32 %f358, %r654;
fma.rn.ftz.f32 %f359, %f140, %f358, %f355;

dp4a.s32.s32 %r523, %r524, %r483, %r542;

cvt.rn.f32.s32 %f360, %r523;
fma.rn.ftz.f32 %f361, %f140, %f360, %f357;

dp4a.s32.s32 %r527, %r528, %r491, %r542;

mul.lo.s32 %r655, %r527, %r602;
cvt.rn.f32.s32 %f362, %r655;
fma.rn.ftz.f32 %f363, %f141, %f362, %f359;

dp4a.s32.s32 %r531, %r532, %r491, %r542;

cvt.rn.f32.s32 %f364, %r531;
fma.rn.ftz.f32 %f365, %f141, %f364, %f361;

dp4a.s32.s32 %r535, %r536, %r499, %r542;

mul.lo.s32 %r656, %r535, %r611;
cvt.rn.f32.s32 %f366, %r656;
fma.rn.ftz.f32 %f367, %f142, %f366, %f363;

dp4a.s32.s32 %r539, %r540, %r499, %r542;

cvt.rn.f32.s32 %f368, %r539;
fma.rn.ftz.f32 %f369, %f142, %f368, %f365;

{.reg .f16 low,high;
mov.b32 {low,high},%r163;
cvt.f32.f16 %f143, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r163;
cvt.f32.f16 %f144, high;}


mul.ftz.f32 %f370, %f367, %f143;
mul.ftz.f32 %f371, %f369, %f144;
sub.ftz.f32 %f372, %f370, %f371;
add.ftz.f32 %f517, %f372, %f517;
add.s32 %r901, %r901, 32;
add.s32 %r900, %r900, 32;
add.s32 %r899, %r899, 32;
add.s32 %r898, %r898, 32;
add.s32 %r897, %r897, 32;
add.s32 %r896, %r896, 32;
add.s64 %rd114, %rd114, 336;
add.s32 %r902, %r902, 4;
setp.lt.s32 %p4, %r902, %r4;
@%p4 bra $L__BB137_2;

$L__BB137_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB137_5;

mov.u32 %r658, _ZZ13mul_mat_vec_qIL9ggml_type10ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r659, %r1, 1536, %r658;
shl.b32 %r660, %r45, 2;
add.s32 %r661, %r659, %r660;
st.shared.f32 [%r661+-1536], %f528;
st.shared.f32 [%r661+-1408], %f527;
st.shared.f32 [%r661+-1280], %f526;
st.shared.f32 [%r661+-1152], %f525;
st.shared.f32 [%r661+-1024], %f524;
st.shared.f32 [%r661+-896], %f523;
st.shared.f32 [%r661+-768], %f522;
st.shared.f32 [%r661+-640], %f521;
st.shared.f32 [%r661+-512], %f520;
st.shared.f32 [%r661+-384], %f519;
st.shared.f32 [%r661+-256], %f518;
st.shared.f32 [%r661+-128], %f517;

$L__BB137_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB137_30;

mov.u32 %r662, %ctaid.z;
mov.u32 %r663, %ctaid.y;
mov.u32 %r664, %ctaid.x;
shl.b32 %r665, %r664, 1;
mov.u32 %r666, 1;
mad.lo.s32 %r667, %r663, %r36, %r665;
mad.lo.s32 %r668, %r662, %r42, %r667;
cvt.u64.u32 %rd14, %r668;
shl.b32 %r670, %r45, 2;
mov.u32 %r671, _ZZ13mul_mat_vec_qIL9ggml_type10ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r25, %r671, %r670;
mov.u32 %r672, 2;
ld.shared.f32 %f373, [%r25];
add.ftz.f32 %f374, %f373, %f528;
mov.b32 %r673, %f374;
mov.u32 %r674, 31;
mov.u32 %r675, 16;
mov.u32 %r676, -1;
shfl.sync.bfly.b32 %r677|%p7, %r673, %r675, %r674, %r676;
mov.b32 %f375, %r677;
add.ftz.f32 %f376, %f374, %f375;
mov.b32 %r678, %f376;
mov.u32 %r679, 8;
shfl.sync.bfly.b32 %r680|%p8, %r678, %r679, %r674, %r676;
mov.b32 %f377, %r680;
add.ftz.f32 %f378, %f376, %f377;
mov.b32 %r681, %f378;
mov.u32 %r682, 4;
shfl.sync.bfly.b32 %r683|%p9, %r681, %r682, %r674, %r676;
mov.b32 %f379, %r683;
add.ftz.f32 %f380, %f378, %f379;
mov.b32 %r684, %f380;
shfl.sync.bfly.b32 %r685|%p10, %r684, %r672, %r674, %r676;
mov.b32 %f381, %r685;
add.ftz.f32 %f382, %f380, %f381;
mov.b32 %r686, %f382;
shfl.sync.bfly.b32 %r687|%p11, %r686, %r666, %r674, %r676;
mov.b32 %f383, %r687;
add.ftz.f32 %f37, %f382, %f383;
setp.eq.s32 %p12, %r45, 0;
setp.lt.u32 %p13, %r665, %r30;
and.pred %p1, %p12, %p13;
not.pred %p14, %p1;
@%p14 bra $L__BB137_8;

cvta.to.global.u64 %rd70, %rd19;
mul.wide.u32 %rd71, %r668, 4;
add.s64 %rd72, %rd70, %rd71;
st.global.f32 [%rd72], %f37;

$L__BB137_8:
ld.shared.f32 %f384, [%r25+128];
add.ftz.f32 %f385, %f384, %f527;
mov.b32 %r694, %f385;
shfl.sync.bfly.b32 %r698|%p15, %r694, %r675, %r674, %r676;
mov.b32 %f386, %r698;
add.ftz.f32 %f387, %f385, %f386;
mov.b32 %r699, %f387;
shfl.sync.bfly.b32 %r701|%p16, %r699, %r679, %r674, %r676;
mov.b32 %f388, %r701;
add.ftz.f32 %f389, %f387, %f388;
mov.b32 %r702, %f389;
shfl.sync.bfly.b32 %r704|%p17, %r702, %r682, %r674, %r676;
mov.b32 %f390, %r704;
add.ftz.f32 %f391, %f389, %f390;
mov.b32 %r705, %f391;
shfl.sync.bfly.b32 %r707|%p18, %r705, %r672, %r674, %r676;
mov.b32 %f392, %r707;
add.ftz.f32 %f393, %f391, %f392;
mov.b32 %r708, %f393;
shfl.sync.bfly.b32 %r710|%p19, %r708, %r666, %r674, %r676;
mov.b32 %f394, %r710;
add.ftz.f32 %f38, %f393, %f394;
setp.eq.s32 %p20, %r45, 1;
or.b32 %r714, %r665, 1;
setp.lt.u32 %p21, %r714, %r30;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB137_10;

cvta.to.global.u64 %rd73, %rd19;
mul.wide.u32 %rd74, %r668, 4;
add.s64 %rd75, %rd73, %rd74;
st.global.f32 [%rd75+4], %f38;

$L__BB137_10:
ld.shared.f32 %f395, [%r25+256];
add.ftz.f32 %f396, %f395, %f526;
mov.b32 %r721, %f396;
mov.u32 %r722, 31;
mov.u32 %r723, 16;
mov.u32 %r724, -1;
shfl.sync.bfly.b32 %r725|%p23, %r721, %r723, %r722, %r724;
mov.b32 %f397, %r725;
add.ftz.f32 %f398, %f396, %f397;
mov.b32 %r726, %f398;
mov.u32 %r727, 8;
shfl.sync.bfly.b32 %r728|%p24, %r726, %r727, %r722, %r724;
mov.b32 %f399, %r728;
add.ftz.f32 %f400, %f398, %f399;
mov.b32 %r729, %f400;
mov.u32 %r730, 4;
shfl.sync.bfly.b32 %r731|%p25, %r729, %r730, %r722, %r724;
mov.b32 %f401, %r731;
add.ftz.f32 %f402, %f400, %f401;
mov.b32 %r732, %f402;
mov.u32 %r733, 2;
shfl.sync.bfly.b32 %r734|%p26, %r732, %r733, %r722, %r724;
mov.b32 %f403, %r734;
add.ftz.f32 %f404, %f402, %f403;
mov.b32 %r735, %f404;
mov.u32 %r736, 1;
shfl.sync.bfly.b32 %r737|%p27, %r735, %r736, %r722, %r724;
mov.b32 %f405, %r737;
add.ftz.f32 %f39, %f404, %f405;
@%p14 bra $L__BB137_12;

cvt.u64.u32 %rd76, %r30;
add.s64 %rd77, %rd76, %rd14;
cvta.to.global.u64 %rd78, %rd19;
shl.b64 %rd79, %rd77, 2;
add.s64 %rd80, %rd78, %rd79;
st.global.f32 [%rd80], %f39;

$L__BB137_12:
ld.shared.f32 %f406, [%r25+384];
add.ftz.f32 %f407, %f406, %f525;
mov.b32 %r738, %f407;
shfl.sync.bfly.b32 %r742|%p29, %r738, %r723, %r722, %r724;
mov.b32 %f408, %r742;
add.ftz.f32 %f409, %f407, %f408;
mov.b32 %r743, %f409;
shfl.sync.bfly.b32 %r745|%p30, %r743, %r727, %r722, %r724;
mov.b32 %f410, %r745;
add.ftz.f32 %f411, %f409, %f410;
mov.b32 %r746, %f411;
shfl.sync.bfly.b32 %r748|%p31, %r746, %r730, %r722, %r724;
mov.b32 %f412, %r748;
add.ftz.f32 %f413, %f411, %f412;
mov.b32 %r749, %f413;
shfl.sync.bfly.b32 %r751|%p32, %r749, %r733, %r722, %r724;
mov.b32 %f414, %r751;
add.ftz.f32 %f415, %f413, %f414;
mov.b32 %r752, %f415;
shfl.sync.bfly.b32 %r754|%p33, %r752, %r736, %r722, %r724;
mov.b32 %f416, %r754;
add.ftz.f32 %f40, %f415, %f416;
@%p22 bra $L__BB137_14;

add.s32 %r755, %r30, 1;
cvt.u64.u32 %rd81, %r755;
add.s64 %rd82, %rd81, %rd14;
cvta.to.global.u64 %rd83, %rd19;
shl.b64 %rd84, %rd82, 2;
add.s64 %rd85, %rd83, %rd84;
st.global.f32 [%rd85], %f40;

$L__BB137_14:
shl.b32 %r756, %r30, 1;
mov.u32 %r757, 1;
ld.shared.f32 %f417, [%r25+512];
add.ftz.f32 %f418, %f417, %f524;
mov.b32 %r758, %f418;
mov.u32 %r759, 31;
mov.u32 %r760, 16;
mov.u32 %r761, -1;
shfl.sync.bfly.b32 %r762|%p35, %r758, %r760, %r759, %r761;
mov.b32 %f419, %r762;
add.ftz.f32 %f420, %f418, %f419;
mov.b32 %r763, %f420;
mov.u32 %r764, 8;
shfl.sync.bfly.b32 %r765|%p36, %r763, %r764, %r759, %r761;
mov.b32 %f421, %r765;
add.ftz.f32 %f422, %f420, %f421;
mov.b32 %r766, %f422;
mov.u32 %r767, 4;
shfl.sync.bfly.b32 %r768|%p37, %r766, %r767, %r759, %r761;
mov.b32 %f423, %r768;
add.ftz.f32 %f424, %f422, %f423;
mov.b32 %r769, %f424;
mov.u32 %r770, 2;
shfl.sync.bfly.b32 %r771|%p38, %r769, %r770, %r759, %r761;
mov.b32 %f425, %r771;
add.ftz.f32 %f426, %f424, %f425;
mov.b32 %r772, %f426;
shfl.sync.bfly.b32 %r773|%p39, %r772, %r757, %r759, %r761;
mov.b32 %f427, %r773;
add.ftz.f32 %f41, %f426, %f427;
cvt.u64.u32 %rd86, %r756;
add.s64 %rd87, %rd86, %rd14;
cvta.to.global.u64 %rd88, %rd19;
shl.b64 %rd89, %rd87, 2;
add.s64 %rd15, %rd88, %rd89;
@%p14 bra $L__BB137_16;

st.global.f32 [%rd15], %f41;

$L__BB137_16:
ld.shared.f32 %f428, [%r25+640];
add.ftz.f32 %f429, %f428, %f523;
mov.b32 %r774, %f429;
shfl.sync.bfly.b32 %r778|%p41, %r774, %r760, %r759, %r761;
mov.b32 %f430, %r778;
add.ftz.f32 %f431, %f429, %f430;
mov.b32 %r779, %f431;
shfl.sync.bfly.b32 %r781|%p42, %r779, %r764, %r759, %r761;
mov.b32 %f432, %r781;
add.ftz.f32 %f433, %f431, %f432;
mov.b32 %r782, %f433;
shfl.sync.bfly.b32 %r784|%p43, %r782, %r767, %r759, %r761;
mov.b32 %f434, %r784;
add.ftz.f32 %f435, %f433, %f434;
mov.b32 %r785, %f435;
shfl.sync.bfly.b32 %r787|%p44, %r785, %r770, %r759, %r761;
mov.b32 %f436, %r787;
add.ftz.f32 %f437, %f435, %f436;
mov.b32 %r788, %f437;
shfl.sync.bfly.b32 %r790|%p45, %r788, %r757, %r759, %r761;
mov.b32 %f438, %r790;
add.ftz.f32 %f42, %f437, %f438;
@%p22 bra $L__BB137_18;

st.global.f32 [%rd15+4], %f42;

$L__BB137_18:
mul.lo.s32 %r26, %r30, 3;
ld.shared.f32 %f439, [%r25+768];
add.ftz.f32 %f440, %f439, %f522;
mov.b32 %r791, %f440;
mov.u32 %r792, 31;
mov.u32 %r793, 16;
mov.u32 %r794, -1;
shfl.sync.bfly.b32 %r795|%p47, %r791, %r793, %r792, %r794;
mov.b32 %f441, %r795;
add.ftz.f32 %f442, %f440, %f441;
mov.b32 %r796, %f442;
mov.u32 %r797, 8;
shfl.sync.bfly.b32 %r798|%p48, %r796, %r797, %r792, %r794;
mov.b32 %f443, %r798;
add.ftz.f32 %f444, %f442, %f443;
mov.b32 %r799, %f444;
mov.u32 %r800, 4;
shfl.sync.bfly.b32 %r801|%p49, %r799, %r800, %r792, %r794;
mov.b32 %f445, %r801;
add.ftz.f32 %f446, %f444, %f445;
mov.b32 %r802, %f446;
mov.u32 %r803, 2;
shfl.sync.bfly.b32 %r804|%p50, %r802, %r803, %r792, %r794;
mov.b32 %f447, %r804;
add.ftz.f32 %f448, %f446, %f447;
mov.b32 %r805, %f448;
mov.u32 %r806, 1;
shfl.sync.bfly.b32 %r807|%p51, %r805, %r806, %r792, %r794;
mov.b32 %f449, %r807;
add.ftz.f32 %f43, %f448, %f449;
@%p14 bra $L__BB137_20;

cvt.u64.u32 %rd90, %r26;
add.s64 %rd91, %rd90, %rd14;
shl.b64 %rd93, %rd91, 2;
add.s64 %rd94, %rd88, %rd93;
st.global.f32 [%rd94], %f43;

$L__BB137_20:
ld.shared.f32 %f450, [%r25+896];
add.ftz.f32 %f451, %f450, %f521;
mov.b32 %r808, %f451;
shfl.sync.bfly.b32 %r812|%p53, %r808, %r793, %r792, %r794;
mov.b32 %f452, %r812;
add.ftz.f32 %f453, %f451, %f452;
mov.b32 %r813, %f453;
shfl.sync.bfly.b32 %r815|%p54, %r813, %r797, %r792, %r794;
mov.b32 %f454, %r815;
add.ftz.f32 %f455, %f453, %f454;
mov.b32 %r816, %f455;
shfl.sync.bfly.b32 %r818|%p55, %r816, %r800, %r792, %r794;
mov.b32 %f456, %r818;
add.ftz.f32 %f457, %f455, %f456;
mov.b32 %r819, %f457;
shfl.sync.bfly.b32 %r821|%p56, %r819, %r803, %r792, %r794;
mov.b32 %f458, %r821;
add.ftz.f32 %f459, %f457, %f458;
mov.b32 %r822, %f459;
shfl.sync.bfly.b32 %r824|%p57, %r822, %r806, %r792, %r794;
mov.b32 %f460, %r824;
add.ftz.f32 %f44, %f459, %f460;
@%p22 bra $L__BB137_22;

add.s32 %r825, %r26, 1;
cvt.u64.u32 %rd95, %r825;
add.s64 %rd96, %rd95, %rd14;
shl.b64 %rd98, %rd96, 2;
add.s64 %rd99, %rd88, %rd98;
st.global.f32 [%rd99], %f44;

$L__BB137_22:
shl.b32 %r826, %r30, 2;
mov.u32 %r827, 2;
ld.shared.f32 %f461, [%r25+1024];
add.ftz.f32 %f462, %f461, %f520;
mov.b32 %r828, %f462;
mov.u32 %r829, 31;
mov.u32 %r830, 16;
mov.u32 %r831, -1;
shfl.sync.bfly.b32 %r832|%p59, %r828, %r830, %r829, %r831;
mov.b32 %f463, %r832;
add.ftz.f32 %f464, %f462, %f463;
mov.b32 %r833, %f464;
mov.u32 %r834, 8;
shfl.sync.bfly.b32 %r835|%p60, %r833, %r834, %r829, %r831;
mov.b32 %f465, %r835;
add.ftz.f32 %f466, %f464, %f465;
mov.b32 %r836, %f466;
mov.u32 %r837, 4;
shfl.sync.bfly.b32 %r838|%p61, %r836, %r837, %r829, %r831;
mov.b32 %f467, %r838;
add.ftz.f32 %f468, %f466, %f467;
mov.b32 %r839, %f468;
shfl.sync.bfly.b32 %r840|%p62, %r839, %r827, %r829, %r831;
mov.b32 %f469, %r840;
add.ftz.f32 %f470, %f468, %f469;
mov.b32 %r841, %f470;
mov.u32 %r842, 1;
shfl.sync.bfly.b32 %r843|%p63, %r841, %r842, %r829, %r831;
mov.b32 %f471, %r843;
add.ftz.f32 %f45, %f470, %f471;
cvt.u64.u32 %rd100, %r826;
add.s64 %rd101, %rd100, %rd14;
shl.b64 %rd103, %rd101, 2;
add.s64 %rd16, %rd88, %rd103;
@%p14 bra $L__BB137_24;

st.global.f32 [%rd16], %f45;

$L__BB137_24:
ld.shared.f32 %f472, [%r25+1152];
add.ftz.f32 %f473, %f472, %f519;
mov.b32 %r844, %f473;
shfl.sync.bfly.b32 %r848|%p65, %r844, %r830, %r829, %r831;
mov.b32 %f474, %r848;
add.ftz.f32 %f475, %f473, %f474;
mov.b32 %r849, %f475;
shfl.sync.bfly.b32 %r851|%p66, %r849, %r834, %r829, %r831;
mov.b32 %f476, %r851;
add.ftz.f32 %f477, %f475, %f476;
mov.b32 %r852, %f477;
shfl.sync.bfly.b32 %r854|%p67, %r852, %r837, %r829, %r831;
mov.b32 %f478, %r854;
add.ftz.f32 %f479, %f477, %f478;
mov.b32 %r855, %f479;
shfl.sync.bfly.b32 %r857|%p68, %r855, %r827, %r829, %r831;
mov.b32 %f480, %r857;
add.ftz.f32 %f481, %f479, %f480;
mov.b32 %r858, %f481;
shfl.sync.bfly.b32 %r860|%p69, %r858, %r842, %r829, %r831;
mov.b32 %f482, %r860;
add.ftz.f32 %f46, %f481, %f482;
@%p22 bra $L__BB137_26;

st.global.f32 [%rd16+4], %f46;

$L__BB137_26:
mul.lo.s32 %r27, %r30, 5;
ld.shared.f32 %f483, [%r25+1280];
add.ftz.f32 %f484, %f483, %f518;
mov.b32 %r861, %f484;
mov.u32 %r862, 31;
mov.u32 %r863, 16;
mov.u32 %r864, -1;
shfl.sync.bfly.b32 %r865|%p71, %r861, %r863, %r862, %r864;
mov.b32 %f485, %r865;
add.ftz.f32 %f486, %f484, %f485;
mov.b32 %r866, %f486;
mov.u32 %r867, 8;
shfl.sync.bfly.b32 %r868|%p72, %r866, %r867, %r862, %r864;
mov.b32 %f487, %r868;
add.ftz.f32 %f488, %f486, %f487;
mov.b32 %r869, %f488;
mov.u32 %r870, 4;
shfl.sync.bfly.b32 %r871|%p73, %r869, %r870, %r862, %r864;
mov.b32 %f489, %r871;
add.ftz.f32 %f490, %f488, %f489;
mov.b32 %r872, %f490;
mov.u32 %r873, 2;
shfl.sync.bfly.b32 %r874|%p74, %r872, %r873, %r862, %r864;
mov.b32 %f491, %r874;
add.ftz.f32 %f492, %f490, %f491;
mov.b32 %r875, %f492;
mov.u32 %r876, 1;
shfl.sync.bfly.b32 %r877|%p75, %r875, %r876, %r862, %r864;
mov.b32 %f493, %r877;
add.ftz.f32 %f47, %f492, %f493;
@%p14 bra $L__BB137_28;

cvt.u64.u32 %rd104, %r27;
add.s64 %rd105, %rd104, %rd14;
shl.b64 %rd107, %rd105, 2;
add.s64 %rd108, %rd88, %rd107;
st.global.f32 [%rd108], %f47;

$L__BB137_28:
ld.shared.f32 %f494, [%r25+1408];
add.ftz.f32 %f495, %f494, %f517;
mov.b32 %r878, %f495;
shfl.sync.bfly.b32 %r882|%p77, %r878, %r863, %r862, %r864;
mov.b32 %f496, %r882;
add.ftz.f32 %f497, %f495, %f496;
mov.b32 %r883, %f497;
shfl.sync.bfly.b32 %r885|%p78, %r883, %r867, %r862, %r864;
mov.b32 %f498, %r885;
add.ftz.f32 %f499, %f497, %f498;
mov.b32 %r886, %f499;
shfl.sync.bfly.b32 %r888|%p79, %r886, %r870, %r862, %r864;
mov.b32 %f500, %r888;
add.ftz.f32 %f501, %f499, %f500;
mov.b32 %r889, %f501;
shfl.sync.bfly.b32 %r891|%p80, %r889, %r873, %r862, %r864;
mov.b32 %f502, %r891;
add.ftz.f32 %f503, %f501, %f502;
mov.b32 %r892, %f503;
shfl.sync.bfly.b32 %r894|%p81, %r892, %r876, %r862, %r864;
mov.b32 %f504, %r894;
add.ftz.f32 %f48, %f503, %f504;
@%p22 bra $L__BB137_30;

add.s32 %r895, %r27, 1;
cvt.u64.u32 %rd109, %r895;
add.s64 %rd110, %rd109, %rd14;
shl.b64 %rd112, %rd110, 2;
add.s64 %rd113, %rd88, %rd112;
st.global.f32 [%rd113], %f48;

$L__BB137_30:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type10ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type10ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type10ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type10ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type10ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type10ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type10ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type10ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type10ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<95>;
.reg .b16 %rs<9>;
.reg .f32 %f<617>;
.reg .b32 %r<1010>;
.reg .b64 %rd<123>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type10ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1792];

ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type10ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd20, [_Z13mul_mat_vec_qIL9ggml_type10ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd21, [_Z13mul_mat_vec_qIL9ggml_type10ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r48, [_Z13mul_mat_vec_qIL9ggml_type10ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type10ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type10ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type10ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type10ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type10ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type10ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r45, [_Z13mul_mat_vec_qIL9ggml_type10ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r46, [_Z13mul_mat_vec_qIL9ggml_type10ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r47, [_Z13mul_mat_vec_qIL9ggml_type10ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type10ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type10ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type10ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type10ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r49, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r49, %r2;
mov.u32 %r50, %ctaid.x;
shl.b32 %r4, %r50, 1;
shr.s32 %r51, %r3, 31;
shr.u32 %r52, %r51, 28;
add.s32 %r53, %r3, %r52;
shr.s32 %r1009, %r53, 4;
shr.u32 %r6, %r48, 8;
setp.ge.s32 %p3, %r1009, %r6;
mov.f32 %f603, 0f00000000;
mov.f32 %f604, %f603;
mov.f32 %f605, %f603;
mov.f32 %f606, %f603;
mov.f32 %f607, %f603;
mov.f32 %f608, %f603;
mov.f32 %f609, %f603;
mov.f32 %f610, %f603;
mov.f32 %f611, %f603;
mov.f32 %f612, %f603;
mov.f32 %f613, %f603;
mov.f32 %f614, %f603;
mov.f32 %f615, %f603;
mov.f32 %f616, %f603;
@%p3 bra $L__BB138_3;

mov.u32 %r54, %ctaid.y;
mov.u32 %r55, %ctaid.z;
mul.lo.s32 %r56, %r54, %r40;
cvt.u64.u32 %rd22, %r56;
mul.lo.s32 %r57, %r55, %r46;
cvt.u64.u32 %rd23, %r57;
add.s64 %rd1, %rd23, %rd22;
mul.hi.u32 %r58, %r55, %r42;
add.s32 %r59, %r58, %r55;
shr.u32 %r60, %r59, %r43;
mul.lo.s32 %r61, %r60, %r45;
mul.hi.u32 %r62, %r54, %r36;
add.s32 %r63, %r62, %r54;
shr.u32 %r64, %r63, %r37;
mul.lo.s32 %r65, %r64, %r39;
cvta.to.global.u64 %rd122, %rd19;
add.s32 %r66, %r4, 1;
and.b32 %r70, %r53, -16;
sub.s32 %r71, %r3, %r70;
shr.s32 %r72, %r71, 31;
shr.u32 %r73, %r72, 29;
add.s32 %r74, %r71, %r73;
and.b32 %r75, %r74, -8;
sub.s32 %r76, %r71, %r75;
shr.s32 %r77, %r76, 31;
shr.u32 %r78, %r77, 30;
add.s32 %r79, %r76, %r78;
shr.s32 %r80, %r79, 2;
add.s32 %r81, %r75, %r80;
cvt.s64.s32 %rd24, %r81;
shr.s32 %r82, %r74, 3;
shl.b32 %r83, %r82, 2;
cvt.s64.s32 %rd3, %r76;
cvt.s64.s32 %rd4, %r83;
shl.b32 %r84, %r34, 1;
shl.b32 %r1002, %r1009, 3;
add.s32 %r1008, %r1002, %r84;
mad.lo.s32 %r1007, %r34, 3, %r1002;
shl.b32 %r85, %r34, 2;
add.s32 %r1006, %r1002, %r85;
mad.lo.s32 %r1005, %r34, 5, %r1002;
mad.lo.s32 %r1004, %r34, 6, %r1002;
add.s32 %r1003, %r34, %r1002;
add.s32 %r86, %r1009, %r61;
add.s32 %r87, %r86, %r65;
mad.lo.s32 %r88, %r66, %r33, %r87;
mul.wide.s32 %rd5, %r88, 84;
mad.lo.s32 %r89, %r4, %r33, %r1009;
add.s32 %r90, %r89, %r61;
add.s32 %r91, %r90, %r65;
mul.wide.s32 %rd6, %r91, 84;
mul.wide.s32 %rd25, %r71, 4;
add.s64 %rd26, %rd6, %rd25;
add.s64 %rd7, %rd26, 16;
add.s64 %rd27, %rd6, %rd24;
add.s64 %rd8, %rd27, 6;
add.s64 %rd9, %rd5, %rd24;
add.s64 %rd28, %rd5, %rd25;
add.s64 %rd10, %rd28, 16;
cvta.to.global.u64 %rd11, %rd20;
mov.f32 %f603, 0f00000000;
shl.b64 %rd35, %rd3, 2;

$L__BB138_2:
cvt.u64.u32 %rd29, %r1002;
add.s64 %rd30, %rd1, %rd29;
add.s64 %rd31, %rd122, %rd7;
ld.global.nc.u32 %r624, [%rd31];
add.s64 %rd32, %rd30, %rd4;
mul.lo.s64 %rd33, %rd32, 36;
add.s64 %rd34, %rd11, %rd33;
add.s64 %rd36, %rd34, %rd35;
ld.global.nc.u32 %r98, [%rd36+4];
ld.global.nc.u32 %r92, [%rd34];

{.reg .f16 low,high;
mov.b32 {low,high},%r92;
cvt.f32.f16 %f85, low;}


ld.global.nc.u32 %r106, [%rd36+40];
ld.global.nc.u32 %r93, [%rd34+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r93;
cvt.f32.f16 %f86, low;}


ld.global.nc.u32 %r114, [%rd36+76];
ld.global.nc.u32 %r94, [%rd34+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r94;
cvt.f32.f16 %f87, low;}


ld.global.nc.u32 %r122, [%rd36+112];
ld.global.nc.u32 %r95, [%rd34+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r95;
cvt.f32.f16 %f88, low;}


add.s64 %rd37, %rd122, %rd8;
ld.global.nc.u8 %rs1, [%rd37+-6];
cvt.u32.u16 %r625, %rs1;
and.b32 %r626, %r625, 240;
and.b32 %r553, %r624, 50529027;
mov.u32 %r621, 0;

dp4a.s32.s32 %r96, %r553, %r98, %r621;

and.b32 %r627, %r625, 15;
mul.lo.s32 %r628, %r96, %r627;
cvt.rn.f32.s32 %f169, %r628;
fma.rn.ftz.f32 %f170, %f85, %f169, 0f00000000;
shr.u32 %r629, %r626, 4;
shl.b32 %r630, %r629, 8;
or.b32 %r631, %r630, %r629;
shl.b32 %r632, %r631, 16;
or.b32 %r557, %r632, %r631;

dp4a.s32.s32 %r100, %r557, %r98, %r621;

cvt.rn.f32.s32 %f171, %r100;
fma.rn.ftz.f32 %f172, %f85, %f171, 0f00000000;
ld.global.nc.u8 %rs2, [%rd37+-4];
cvt.u32.u16 %r633, %rs2;
and.b32 %r634, %r633, 240;
shr.u32 %r635, %r624, 2;
and.b32 %r561, %r635, 50529027;

dp4a.s32.s32 %r104, %r561, %r106, %r621;

and.b32 %r636, %r633, 15;
mul.lo.s32 %r637, %r104, %r636;
cvt.rn.f32.s32 %f173, %r637;
fma.rn.ftz.f32 %f174, %f86, %f173, %f170;
shr.u32 %r638, %r634, 4;
shl.b32 %r639, %r638, 8;
or.b32 %r640, %r639, %r638;
shl.b32 %r641, %r640, 16;
or.b32 %r565, %r641, %r640;

dp4a.s32.s32 %r108, %r565, %r106, %r621;

cvt.rn.f32.s32 %f175, %r108;
fma.rn.ftz.f32 %f176, %f86, %f175, %f172;
ld.global.nc.u8 %rs3, [%rd37+-2];
cvt.u32.u16 %r642, %rs3;
and.b32 %r643, %r642, 240;
shr.u32 %r644, %r624, 4;
and.b32 %r569, %r644, 50529027;

dp4a.s32.s32 %r112, %r569, %r114, %r621;

and.b32 %r645, %r642, 15;
mul.lo.s32 %r646, %r112, %r645;
cvt.rn.f32.s32 %f177, %r646;
fma.rn.ftz.f32 %f178, %f87, %f177, %f174;
shr.u32 %r647, %r643, 4;
shl.b32 %r648, %r647, 8;
or.b32 %r649, %r648, %r647;
shl.b32 %r650, %r649, 16;
or.b32 %r573, %r650, %r649;

dp4a.s32.s32 %r116, %r573, %r114, %r621;

cvt.rn.f32.s32 %f179, %r116;
fma.rn.ftz.f32 %f180, %f87, %f179, %f176;
ld.global.nc.u8 %rs4, [%rd37];
cvt.u32.u16 %r651, %rs4;
and.b32 %r652, %r651, 240;
shr.u32 %r653, %r624, 6;
and.b32 %r577, %r653, 50529027;

dp4a.s32.s32 %r120, %r577, %r122, %r621;

and.b32 %r654, %r651, 15;
mul.lo.s32 %r655, %r120, %r654;
cvt.rn.f32.s32 %f181, %r655;
fma.rn.ftz.f32 %f182, %f88, %f181, %f178;
shr.u32 %r656, %r652, 4;
shl.b32 %r657, %r656, 8;
or.b32 %r658, %r657, %r656;
shl.b32 %r659, %r658, 16;
or.b32 %r581, %r659, %r658;

dp4a.s32.s32 %r124, %r581, %r122, %r621;

cvt.rn.f32.s32 %f183, %r124;
fma.rn.ftz.f32 %f184, %f88, %f183, %f180;
add.s64 %rd38, %rd122, %rd6;
ld.global.nc.u32 %r128, [%rd38+80];

{.reg .f16 low,high;
mov.b32 {low,high},%r128;
cvt.f32.f16 %f89, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r128;
cvt.f32.f16 %f90, high;}


mul.ftz.f32 %f185, %f182, %f89;
mul.ftz.f32 %f186, %f184, %f90;
sub.ftz.f32 %f187, %f185, %f186;
add.ftz.f32 %f616, %f187, %f616;
add.s64 %rd39, %rd122, %rd10;
ld.global.nc.u32 %r660, [%rd39];

{.reg .f16 low,high;
mov.b32 {low,high},%r92;
cvt.f32.f16 %f91, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r93;
cvt.f32.f16 %f92, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r94;
cvt.f32.f16 %f93, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r95;
cvt.f32.f16 %f94, low;}


add.s64 %rd40, %rd122, %rd9;
ld.global.nc.u8 %rs5, [%rd40];
cvt.u32.u16 %r661, %rs5;
and.b32 %r662, %r661, 240;
and.b32 %r591, %r660, 50529027;

dp4a.s32.s32 %r134, %r591, %r98, %r621;

and.b32 %r663, %r661, 15;
mul.lo.s32 %r664, %r134, %r663;
cvt.rn.f32.s32 %f188, %r664;
fma.rn.ftz.f32 %f189, %f91, %f188, 0f00000000;
shr.u32 %r665, %r662, 4;
shl.b32 %r666, %r665, 8;
or.b32 %r667, %r666, %r665;
shl.b32 %r668, %r667, 16;
or.b32 %r595, %r668, %r667;

dp4a.s32.s32 %r138, %r595, %r98, %r621;

cvt.rn.f32.s32 %f190, %r138;
fma.rn.ftz.f32 %f191, %f91, %f190, 0f00000000;
ld.global.nc.u8 %rs6, [%rd40+2];
cvt.u32.u16 %r669, %rs6;
and.b32 %r670, %r669, 240;
shr.u32 %r671, %r660, 2;
and.b32 %r599, %r671, 50529027;

dp4a.s32.s32 %r142, %r599, %r106, %r621;

and.b32 %r672, %r669, 15;
mul.lo.s32 %r673, %r142, %r672;
cvt.rn.f32.s32 %f192, %r673;
fma.rn.ftz.f32 %f193, %f92, %f192, %f189;
shr.u32 %r674, %r670, 4;
shl.b32 %r675, %r674, 8;
or.b32 %r676, %r675, %r674;
shl.b32 %r677, %r676, 16;
or.b32 %r603, %r677, %r676;

dp4a.s32.s32 %r146, %r603, %r106, %r621;

cvt.rn.f32.s32 %f194, %r146;
fma.rn.ftz.f32 %f195, %f92, %f194, %f191;
ld.global.nc.u8 %rs7, [%rd40+4];
cvt.u32.u16 %r678, %rs7;
and.b32 %r679, %r678, 240;
shr.u32 %r680, %r660, 4;
and.b32 %r607, %r680, 50529027;

dp4a.s32.s32 %r150, %r607, %r114, %r621;

and.b32 %r681, %r678, 15;
mul.lo.s32 %r682, %r150, %r681;
cvt.rn.f32.s32 %f196, %r682;
fma.rn.ftz.f32 %f197, %f93, %f196, %f193;
shr.u32 %r683, %r679, 4;
shl.b32 %r684, %r683, 8;
or.b32 %r685, %r684, %r683;
shl.b32 %r686, %r685, 16;
or.b32 %r611, %r686, %r685;

dp4a.s32.s32 %r154, %r611, %r114, %r621;

cvt.rn.f32.s32 %f198, %r154;
fma.rn.ftz.f32 %f199, %f93, %f198, %f195;
ld.global.nc.u8 %rs8, [%rd40+6];
cvt.u32.u16 %r687, %rs8;
and.b32 %r688, %r687, 240;
shr.u32 %r689, %r660, 6;
and.b32 %r615, %r689, 50529027;

dp4a.s32.s32 %r158, %r615, %r122, %r621;

and.b32 %r690, %r687, 15;
mul.lo.s32 %r691, %r158, %r690;
cvt.rn.f32.s32 %f200, %r691;
fma.rn.ftz.f32 %f201, %f94, %f200, %f197;
shr.u32 %r692, %r688, 4;
shl.b32 %r693, %r692, 8;
or.b32 %r694, %r693, %r692;
shl.b32 %r695, %r694, 16;
or.b32 %r619, %r695, %r694;

dp4a.s32.s32 %r162, %r619, %r122, %r621;

cvt.rn.f32.s32 %f202, %r162;
fma.rn.ftz.f32 %f203, %f94, %f202, %f199;
add.s64 %rd41, %rd122, %rd5;
ld.global.nc.u32 %r166, [%rd41+80];

{.reg .f16 low,high;
mov.b32 {low,high},%r166;
cvt.f32.f16 %f95, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r166;
cvt.f32.f16 %f96, high;}


mul.ftz.f32 %f204, %f201, %f95;
mul.ftz.f32 %f205, %f203, %f96;
sub.ftz.f32 %f206, %f204, %f205;
add.ftz.f32 %f615, %f206, %f615;
cvt.u64.u32 %rd42, %r1003;
add.s64 %rd43, %rd1, %rd42;
add.s64 %rd44, %rd43, %rd4;
mul.lo.s64 %rd45, %rd44, 36;
add.s64 %rd46, %rd11, %rd45;
add.s64 %rd47, %rd46, %rd35;
ld.global.nc.u32 %r174, [%rd47+4];
ld.global.nc.u32 %r168, [%rd46];

{.reg .f16 low,high;
mov.b32 {low,high},%r168;
cvt.f32.f16 %f97, low;}


ld.global.nc.u32 %r182, [%rd47+40];
ld.global.nc.u32 %r169, [%rd46+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r169;
cvt.f32.f16 %f98, low;}


ld.global.nc.u32 %r190, [%rd47+76];
ld.global.nc.u32 %r170, [%rd46+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r170;
cvt.f32.f16 %f99, low;}


ld.global.nc.u32 %r198, [%rd47+112];
ld.global.nc.u32 %r171, [%rd46+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r171;
cvt.f32.f16 %f100, low;}



dp4a.s32.s32 %r172, %r553, %r174, %r621;

mul.lo.s32 %r696, %r172, %r627;
cvt.rn.f32.s32 %f207, %r696;
fma.rn.ftz.f32 %f208, %f97, %f207, 0f00000000;

dp4a.s32.s32 %r176, %r557, %r174, %r621;

cvt.rn.f32.s32 %f209, %r176;
fma.rn.ftz.f32 %f210, %f97, %f209, 0f00000000;

dp4a.s32.s32 %r180, %r561, %r182, %r621;

mul.lo.s32 %r697, %r180, %r636;
cvt.rn.f32.s32 %f211, %r697;
fma.rn.ftz.f32 %f212, %f98, %f211, %f208;

dp4a.s32.s32 %r184, %r565, %r182, %r621;

cvt.rn.f32.s32 %f213, %r184;
fma.rn.ftz.f32 %f214, %f98, %f213, %f210;

dp4a.s32.s32 %r188, %r569, %r190, %r621;

mul.lo.s32 %r698, %r188, %r645;
cvt.rn.f32.s32 %f215, %r698;
fma.rn.ftz.f32 %f216, %f99, %f215, %f212;

dp4a.s32.s32 %r192, %r573, %r190, %r621;

cvt.rn.f32.s32 %f217, %r192;
fma.rn.ftz.f32 %f218, %f99, %f217, %f214;

dp4a.s32.s32 %r196, %r577, %r198, %r621;

mul.lo.s32 %r699, %r196, %r654;
cvt.rn.f32.s32 %f219, %r699;
fma.rn.ftz.f32 %f220, %f100, %f219, %f216;

dp4a.s32.s32 %r200, %r581, %r198, %r621;

cvt.rn.f32.s32 %f221, %r200;
fma.rn.ftz.f32 %f222, %f100, %f221, %f218;

{.reg .f16 low,high;
mov.b32 {low,high},%r128;
cvt.f32.f16 %f101, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r128;
cvt.f32.f16 %f102, high;}


mul.ftz.f32 %f223, %f220, %f101;
mul.ftz.f32 %f224, %f222, %f102;
sub.ftz.f32 %f225, %f223, %f224;
add.ftz.f32 %f614, %f225, %f614;

{.reg .f16 low,high;
mov.b32 {low,high},%r168;
cvt.f32.f16 %f103, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r169;
cvt.f32.f16 %f104, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r170;
cvt.f32.f16 %f105, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r171;
cvt.f32.f16 %f106, low;}



dp4a.s32.s32 %r210, %r591, %r174, %r621;

mul.lo.s32 %r700, %r210, %r663;
cvt.rn.f32.s32 %f226, %r700;
fma.rn.ftz.f32 %f227, %f103, %f226, 0f00000000;

dp4a.s32.s32 %r214, %r595, %r174, %r621;

cvt.rn.f32.s32 %f228, %r214;
fma.rn.ftz.f32 %f229, %f103, %f228, 0f00000000;

dp4a.s32.s32 %r218, %r599, %r182, %r621;

mul.lo.s32 %r701, %r218, %r672;
cvt.rn.f32.s32 %f230, %r701;
fma.rn.ftz.f32 %f231, %f104, %f230, %f227;

dp4a.s32.s32 %r222, %r603, %r182, %r621;

cvt.rn.f32.s32 %f232, %r222;
fma.rn.ftz.f32 %f233, %f104, %f232, %f229;

dp4a.s32.s32 %r226, %r607, %r190, %r621;

mul.lo.s32 %r702, %r226, %r681;
cvt.rn.f32.s32 %f234, %r702;
fma.rn.ftz.f32 %f235, %f105, %f234, %f231;

dp4a.s32.s32 %r230, %r611, %r190, %r621;

cvt.rn.f32.s32 %f236, %r230;
fma.rn.ftz.f32 %f237, %f105, %f236, %f233;

dp4a.s32.s32 %r234, %r615, %r198, %r621;

mul.lo.s32 %r703, %r234, %r690;
cvt.rn.f32.s32 %f238, %r703;
fma.rn.ftz.f32 %f239, %f106, %f238, %f235;

dp4a.s32.s32 %r238, %r619, %r198, %r621;

cvt.rn.f32.s32 %f240, %r238;
fma.rn.ftz.f32 %f241, %f106, %f240, %f237;

{.reg .f16 low,high;
mov.b32 {low,high},%r166;
cvt.f32.f16 %f107, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r166;
cvt.f32.f16 %f108, high;}


mul.ftz.f32 %f242, %f239, %f107;
mul.ftz.f32 %f243, %f241, %f108;
sub.ftz.f32 %f244, %f242, %f243;
add.ftz.f32 %f613, %f244, %f613;
cvt.u64.u32 %rd48, %r1008;
add.s64 %rd49, %rd1, %rd48;
add.s64 %rd50, %rd49, %rd4;
mul.lo.s64 %rd51, %rd50, 36;
add.s64 %rd52, %rd11, %rd51;
add.s64 %rd53, %rd52, %rd35;
ld.global.nc.u32 %r250, [%rd53+4];
ld.global.nc.u32 %r244, [%rd52];

{.reg .f16 low,high;
mov.b32 {low,high},%r244;
cvt.f32.f16 %f109, low;}


ld.global.nc.u32 %r258, [%rd53+40];
ld.global.nc.u32 %r245, [%rd52+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r245;
cvt.f32.f16 %f110, low;}


ld.global.nc.u32 %r266, [%rd53+76];
ld.global.nc.u32 %r246, [%rd52+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r246;
cvt.f32.f16 %f111, low;}


ld.global.nc.u32 %r274, [%rd53+112];
ld.global.nc.u32 %r247, [%rd52+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r247;
cvt.f32.f16 %f112, low;}



dp4a.s32.s32 %r248, %r553, %r250, %r621;

mul.lo.s32 %r704, %r248, %r627;
cvt.rn.f32.s32 %f245, %r704;
fma.rn.ftz.f32 %f246, %f109, %f245, 0f00000000;

dp4a.s32.s32 %r252, %r557, %r250, %r621;

cvt.rn.f32.s32 %f247, %r252;
fma.rn.ftz.f32 %f248, %f109, %f247, 0f00000000;

dp4a.s32.s32 %r256, %r561, %r258, %r621;

mul.lo.s32 %r705, %r256, %r636;
cvt.rn.f32.s32 %f249, %r705;
fma.rn.ftz.f32 %f250, %f110, %f249, %f246;

dp4a.s32.s32 %r260, %r565, %r258, %r621;

cvt.rn.f32.s32 %f251, %r260;
fma.rn.ftz.f32 %f252, %f110, %f251, %f248;

dp4a.s32.s32 %r264, %r569, %r266, %r621;

mul.lo.s32 %r706, %r264, %r645;
cvt.rn.f32.s32 %f253, %r706;
fma.rn.ftz.f32 %f254, %f111, %f253, %f250;

dp4a.s32.s32 %r268, %r573, %r266, %r621;

cvt.rn.f32.s32 %f255, %r268;
fma.rn.ftz.f32 %f256, %f111, %f255, %f252;

dp4a.s32.s32 %r272, %r577, %r274, %r621;

mul.lo.s32 %r707, %r272, %r654;
cvt.rn.f32.s32 %f257, %r707;
fma.rn.ftz.f32 %f258, %f112, %f257, %f254;

dp4a.s32.s32 %r276, %r581, %r274, %r621;

cvt.rn.f32.s32 %f259, %r276;
fma.rn.ftz.f32 %f260, %f112, %f259, %f256;

{.reg .f16 low,high;
mov.b32 {low,high},%r128;
cvt.f32.f16 %f113, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r128;
cvt.f32.f16 %f114, high;}


mul.ftz.f32 %f261, %f258, %f113;
mul.ftz.f32 %f262, %f260, %f114;
sub.ftz.f32 %f263, %f261, %f262;
add.ftz.f32 %f612, %f263, %f612;

{.reg .f16 low,high;
mov.b32 {low,high},%r244;
cvt.f32.f16 %f115, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r245;
cvt.f32.f16 %f116, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r246;
cvt.f32.f16 %f117, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r247;
cvt.f32.f16 %f118, low;}



dp4a.s32.s32 %r286, %r591, %r250, %r621;

mul.lo.s32 %r708, %r286, %r663;
cvt.rn.f32.s32 %f264, %r708;
fma.rn.ftz.f32 %f265, %f115, %f264, 0f00000000;

dp4a.s32.s32 %r290, %r595, %r250, %r621;

cvt.rn.f32.s32 %f266, %r290;
fma.rn.ftz.f32 %f267, %f115, %f266, 0f00000000;

dp4a.s32.s32 %r294, %r599, %r258, %r621;

mul.lo.s32 %r709, %r294, %r672;
cvt.rn.f32.s32 %f268, %r709;
fma.rn.ftz.f32 %f269, %f116, %f268, %f265;

dp4a.s32.s32 %r298, %r603, %r258, %r621;

cvt.rn.f32.s32 %f270, %r298;
fma.rn.ftz.f32 %f271, %f116, %f270, %f267;

dp4a.s32.s32 %r302, %r607, %r266, %r621;

mul.lo.s32 %r710, %r302, %r681;
cvt.rn.f32.s32 %f272, %r710;
fma.rn.ftz.f32 %f273, %f117, %f272, %f269;

dp4a.s32.s32 %r306, %r611, %r266, %r621;

cvt.rn.f32.s32 %f274, %r306;
fma.rn.ftz.f32 %f275, %f117, %f274, %f271;

dp4a.s32.s32 %r310, %r615, %r274, %r621;

mul.lo.s32 %r711, %r310, %r690;
cvt.rn.f32.s32 %f276, %r711;
fma.rn.ftz.f32 %f277, %f118, %f276, %f273;

dp4a.s32.s32 %r314, %r619, %r274, %r621;

cvt.rn.f32.s32 %f278, %r314;
fma.rn.ftz.f32 %f279, %f118, %f278, %f275;

{.reg .f16 low,high;
mov.b32 {low,high},%r166;
cvt.f32.f16 %f119, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r166;
cvt.f32.f16 %f120, high;}


mul.ftz.f32 %f280, %f277, %f119;
mul.ftz.f32 %f281, %f279, %f120;
sub.ftz.f32 %f282, %f280, %f281;
add.ftz.f32 %f611, %f282, %f611;
cvt.u64.u32 %rd54, %r1007;
add.s64 %rd55, %rd1, %rd54;
add.s64 %rd56, %rd55, %rd4;
mul.lo.s64 %rd57, %rd56, 36;
add.s64 %rd58, %rd11, %rd57;
add.s64 %rd59, %rd58, %rd35;
ld.global.nc.u32 %r326, [%rd59+4];
ld.global.nc.u32 %r320, [%rd58];

{.reg .f16 low,high;
mov.b32 {low,high},%r320;
cvt.f32.f16 %f121, low;}


ld.global.nc.u32 %r334, [%rd59+40];
ld.global.nc.u32 %r321, [%rd58+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r321;
cvt.f32.f16 %f122, low;}


ld.global.nc.u32 %r342, [%rd59+76];
ld.global.nc.u32 %r322, [%rd58+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r322;
cvt.f32.f16 %f123, low;}


ld.global.nc.u32 %r350, [%rd59+112];
ld.global.nc.u32 %r323, [%rd58+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r323;
cvt.f32.f16 %f124, low;}



dp4a.s32.s32 %r324, %r553, %r326, %r621;

mul.lo.s32 %r712, %r324, %r627;
cvt.rn.f32.s32 %f283, %r712;
fma.rn.ftz.f32 %f284, %f121, %f283, 0f00000000;

dp4a.s32.s32 %r328, %r557, %r326, %r621;

cvt.rn.f32.s32 %f285, %r328;
fma.rn.ftz.f32 %f286, %f121, %f285, 0f00000000;

dp4a.s32.s32 %r332, %r561, %r334, %r621;

mul.lo.s32 %r713, %r332, %r636;
cvt.rn.f32.s32 %f287, %r713;
fma.rn.ftz.f32 %f288, %f122, %f287, %f284;

dp4a.s32.s32 %r336, %r565, %r334, %r621;

cvt.rn.f32.s32 %f289, %r336;
fma.rn.ftz.f32 %f290, %f122, %f289, %f286;

dp4a.s32.s32 %r340, %r569, %r342, %r621;

mul.lo.s32 %r714, %r340, %r645;
cvt.rn.f32.s32 %f291, %r714;
fma.rn.ftz.f32 %f292, %f123, %f291, %f288;

dp4a.s32.s32 %r344, %r573, %r342, %r621;

cvt.rn.f32.s32 %f293, %r344;
fma.rn.ftz.f32 %f294, %f123, %f293, %f290;

dp4a.s32.s32 %r348, %r577, %r350, %r621;

mul.lo.s32 %r715, %r348, %r654;
cvt.rn.f32.s32 %f295, %r715;
fma.rn.ftz.f32 %f296, %f124, %f295, %f292;

dp4a.s32.s32 %r352, %r581, %r350, %r621;

cvt.rn.f32.s32 %f297, %r352;
fma.rn.ftz.f32 %f298, %f124, %f297, %f294;

{.reg .f16 low,high;
mov.b32 {low,high},%r128;
cvt.f32.f16 %f125, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r128;
cvt.f32.f16 %f126, high;}


mul.ftz.f32 %f299, %f296, %f125;
mul.ftz.f32 %f300, %f298, %f126;
sub.ftz.f32 %f301, %f299, %f300;
add.ftz.f32 %f610, %f301, %f610;

{.reg .f16 low,high;
mov.b32 {low,high},%r320;
cvt.f32.f16 %f127, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r321;
cvt.f32.f16 %f128, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r322;
cvt.f32.f16 %f129, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r323;
cvt.f32.f16 %f130, low;}



dp4a.s32.s32 %r362, %r591, %r326, %r621;

mul.lo.s32 %r716, %r362, %r663;
cvt.rn.f32.s32 %f302, %r716;
fma.rn.ftz.f32 %f303, %f127, %f302, 0f00000000;

dp4a.s32.s32 %r366, %r595, %r326, %r621;

cvt.rn.f32.s32 %f304, %r366;
fma.rn.ftz.f32 %f305, %f127, %f304, 0f00000000;

dp4a.s32.s32 %r370, %r599, %r334, %r621;

mul.lo.s32 %r717, %r370, %r672;
cvt.rn.f32.s32 %f306, %r717;
fma.rn.ftz.f32 %f307, %f128, %f306, %f303;

dp4a.s32.s32 %r374, %r603, %r334, %r621;

cvt.rn.f32.s32 %f308, %r374;
fma.rn.ftz.f32 %f309, %f128, %f308, %f305;

dp4a.s32.s32 %r378, %r607, %r342, %r621;

mul.lo.s32 %r718, %r378, %r681;
cvt.rn.f32.s32 %f310, %r718;
fma.rn.ftz.f32 %f311, %f129, %f310, %f307;

dp4a.s32.s32 %r382, %r611, %r342, %r621;

cvt.rn.f32.s32 %f312, %r382;
fma.rn.ftz.f32 %f313, %f129, %f312, %f309;

dp4a.s32.s32 %r386, %r615, %r350, %r621;

mul.lo.s32 %r719, %r386, %r690;
cvt.rn.f32.s32 %f314, %r719;
fma.rn.ftz.f32 %f315, %f130, %f314, %f311;

dp4a.s32.s32 %r390, %r619, %r350, %r621;

cvt.rn.f32.s32 %f316, %r390;
fma.rn.ftz.f32 %f317, %f130, %f316, %f313;

{.reg .f16 low,high;
mov.b32 {low,high},%r166;
cvt.f32.f16 %f131, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r166;
cvt.f32.f16 %f132, high;}


mul.ftz.f32 %f318, %f315, %f131;
mul.ftz.f32 %f319, %f317, %f132;
sub.ftz.f32 %f320, %f318, %f319;
add.ftz.f32 %f609, %f320, %f609;
cvt.u64.u32 %rd60, %r1006;
add.s64 %rd61, %rd1, %rd60;
add.s64 %rd62, %rd61, %rd4;
mul.lo.s64 %rd63, %rd62, 36;
add.s64 %rd64, %rd11, %rd63;
add.s64 %rd65, %rd64, %rd35;
ld.global.nc.u32 %r402, [%rd65+4];
ld.global.nc.u32 %r396, [%rd64];

{.reg .f16 low,high;
mov.b32 {low,high},%r396;
cvt.f32.f16 %f133, low;}


ld.global.nc.u32 %r410, [%rd65+40];
ld.global.nc.u32 %r397, [%rd64+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r397;
cvt.f32.f16 %f134, low;}


ld.global.nc.u32 %r418, [%rd65+76];
ld.global.nc.u32 %r398, [%rd64+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r398;
cvt.f32.f16 %f135, low;}


ld.global.nc.u32 %r426, [%rd65+112];
ld.global.nc.u32 %r399, [%rd64+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r399;
cvt.f32.f16 %f136, low;}



dp4a.s32.s32 %r400, %r553, %r402, %r621;

mul.lo.s32 %r720, %r400, %r627;
cvt.rn.f32.s32 %f321, %r720;
fma.rn.ftz.f32 %f322, %f133, %f321, 0f00000000;

dp4a.s32.s32 %r404, %r557, %r402, %r621;

cvt.rn.f32.s32 %f323, %r404;
fma.rn.ftz.f32 %f324, %f133, %f323, 0f00000000;

dp4a.s32.s32 %r408, %r561, %r410, %r621;

mul.lo.s32 %r721, %r408, %r636;
cvt.rn.f32.s32 %f325, %r721;
fma.rn.ftz.f32 %f326, %f134, %f325, %f322;

dp4a.s32.s32 %r412, %r565, %r410, %r621;

cvt.rn.f32.s32 %f327, %r412;
fma.rn.ftz.f32 %f328, %f134, %f327, %f324;

dp4a.s32.s32 %r416, %r569, %r418, %r621;

mul.lo.s32 %r722, %r416, %r645;
cvt.rn.f32.s32 %f329, %r722;
fma.rn.ftz.f32 %f330, %f135, %f329, %f326;

dp4a.s32.s32 %r420, %r573, %r418, %r621;

cvt.rn.f32.s32 %f331, %r420;
fma.rn.ftz.f32 %f332, %f135, %f331, %f328;

dp4a.s32.s32 %r424, %r577, %r426, %r621;

mul.lo.s32 %r723, %r424, %r654;
cvt.rn.f32.s32 %f333, %r723;
fma.rn.ftz.f32 %f334, %f136, %f333, %f330;

dp4a.s32.s32 %r428, %r581, %r426, %r621;

cvt.rn.f32.s32 %f335, %r428;
fma.rn.ftz.f32 %f336, %f136, %f335, %f332;

{.reg .f16 low,high;
mov.b32 {low,high},%r128;
cvt.f32.f16 %f137, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r128;
cvt.f32.f16 %f138, high;}


mul.ftz.f32 %f337, %f334, %f137;
mul.ftz.f32 %f338, %f336, %f138;
sub.ftz.f32 %f339, %f337, %f338;
add.ftz.f32 %f608, %f339, %f608;

{.reg .f16 low,high;
mov.b32 {low,high},%r396;
cvt.f32.f16 %f139, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r397;
cvt.f32.f16 %f140, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r398;
cvt.f32.f16 %f141, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r399;
cvt.f32.f16 %f142, low;}



dp4a.s32.s32 %r438, %r591, %r402, %r621;

mul.lo.s32 %r724, %r438, %r663;
cvt.rn.f32.s32 %f340, %r724;
fma.rn.ftz.f32 %f341, %f139, %f340, 0f00000000;

dp4a.s32.s32 %r442, %r595, %r402, %r621;

cvt.rn.f32.s32 %f342, %r442;
fma.rn.ftz.f32 %f343, %f139, %f342, 0f00000000;

dp4a.s32.s32 %r446, %r599, %r410, %r621;

mul.lo.s32 %r725, %r446, %r672;
cvt.rn.f32.s32 %f344, %r725;
fma.rn.ftz.f32 %f345, %f140, %f344, %f341;

dp4a.s32.s32 %r450, %r603, %r410, %r621;

cvt.rn.f32.s32 %f346, %r450;
fma.rn.ftz.f32 %f347, %f140, %f346, %f343;

dp4a.s32.s32 %r454, %r607, %r418, %r621;

mul.lo.s32 %r726, %r454, %r681;
cvt.rn.f32.s32 %f348, %r726;
fma.rn.ftz.f32 %f349, %f141, %f348, %f345;

dp4a.s32.s32 %r458, %r611, %r418, %r621;

cvt.rn.f32.s32 %f350, %r458;
fma.rn.ftz.f32 %f351, %f141, %f350, %f347;

dp4a.s32.s32 %r462, %r615, %r426, %r621;

mul.lo.s32 %r727, %r462, %r690;
cvt.rn.f32.s32 %f352, %r727;
fma.rn.ftz.f32 %f353, %f142, %f352, %f349;

dp4a.s32.s32 %r466, %r619, %r426, %r621;

cvt.rn.f32.s32 %f354, %r466;
fma.rn.ftz.f32 %f355, %f142, %f354, %f351;

{.reg .f16 low,high;
mov.b32 {low,high},%r166;
cvt.f32.f16 %f143, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r166;
cvt.f32.f16 %f144, high;}


mul.ftz.f32 %f356, %f353, %f143;
mul.ftz.f32 %f357, %f355, %f144;
sub.ftz.f32 %f358, %f356, %f357;
add.ftz.f32 %f607, %f358, %f607;
cvt.u64.u32 %rd66, %r1005;
add.s64 %rd67, %rd1, %rd66;
add.s64 %rd68, %rd67, %rd4;
mul.lo.s64 %rd69, %rd68, 36;
add.s64 %rd70, %rd11, %rd69;
add.s64 %rd71, %rd70, %rd35;
ld.global.nc.u32 %r478, [%rd71+4];
ld.global.nc.u32 %r472, [%rd70];

{.reg .f16 low,high;
mov.b32 {low,high},%r472;
cvt.f32.f16 %f145, low;}


ld.global.nc.u32 %r486, [%rd71+40];
ld.global.nc.u32 %r473, [%rd70+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r473;
cvt.f32.f16 %f146, low;}


ld.global.nc.u32 %r494, [%rd71+76];
ld.global.nc.u32 %r474, [%rd70+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r474;
cvt.f32.f16 %f147, low;}


ld.global.nc.u32 %r502, [%rd71+112];
ld.global.nc.u32 %r475, [%rd70+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r475;
cvt.f32.f16 %f148, low;}



dp4a.s32.s32 %r476, %r553, %r478, %r621;

mul.lo.s32 %r728, %r476, %r627;
cvt.rn.f32.s32 %f359, %r728;
fma.rn.ftz.f32 %f360, %f145, %f359, 0f00000000;

dp4a.s32.s32 %r480, %r557, %r478, %r621;

cvt.rn.f32.s32 %f361, %r480;
fma.rn.ftz.f32 %f362, %f145, %f361, 0f00000000;

dp4a.s32.s32 %r484, %r561, %r486, %r621;

mul.lo.s32 %r729, %r484, %r636;
cvt.rn.f32.s32 %f363, %r729;
fma.rn.ftz.f32 %f364, %f146, %f363, %f360;

dp4a.s32.s32 %r488, %r565, %r486, %r621;

cvt.rn.f32.s32 %f365, %r488;
fma.rn.ftz.f32 %f366, %f146, %f365, %f362;

dp4a.s32.s32 %r492, %r569, %r494, %r621;

mul.lo.s32 %r730, %r492, %r645;
cvt.rn.f32.s32 %f367, %r730;
fma.rn.ftz.f32 %f368, %f147, %f367, %f364;

dp4a.s32.s32 %r496, %r573, %r494, %r621;

cvt.rn.f32.s32 %f369, %r496;
fma.rn.ftz.f32 %f370, %f147, %f369, %f366;

dp4a.s32.s32 %r500, %r577, %r502, %r621;

mul.lo.s32 %r731, %r500, %r654;
cvt.rn.f32.s32 %f371, %r731;
fma.rn.ftz.f32 %f372, %f148, %f371, %f368;

dp4a.s32.s32 %r504, %r581, %r502, %r621;

cvt.rn.f32.s32 %f373, %r504;
fma.rn.ftz.f32 %f374, %f148, %f373, %f370;

{.reg .f16 low,high;
mov.b32 {low,high},%r128;
cvt.f32.f16 %f149, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r128;
cvt.f32.f16 %f150, high;}


mul.ftz.f32 %f375, %f372, %f149;
mul.ftz.f32 %f376, %f374, %f150;
sub.ftz.f32 %f377, %f375, %f376;
add.ftz.f32 %f606, %f377, %f606;

{.reg .f16 low,high;
mov.b32 {low,high},%r472;
cvt.f32.f16 %f151, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r473;
cvt.f32.f16 %f152, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r474;
cvt.f32.f16 %f153, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r475;
cvt.f32.f16 %f154, low;}



dp4a.s32.s32 %r514, %r591, %r478, %r621;

mul.lo.s32 %r732, %r514, %r663;
cvt.rn.f32.s32 %f378, %r732;
fma.rn.ftz.f32 %f379, %f151, %f378, 0f00000000;

dp4a.s32.s32 %r518, %r595, %r478, %r621;

cvt.rn.f32.s32 %f380, %r518;
fma.rn.ftz.f32 %f381, %f151, %f380, 0f00000000;

dp4a.s32.s32 %r522, %r599, %r486, %r621;

mul.lo.s32 %r733, %r522, %r672;
cvt.rn.f32.s32 %f382, %r733;
fma.rn.ftz.f32 %f383, %f152, %f382, %f379;

dp4a.s32.s32 %r526, %r603, %r486, %r621;

cvt.rn.f32.s32 %f384, %r526;
fma.rn.ftz.f32 %f385, %f152, %f384, %f381;

dp4a.s32.s32 %r530, %r607, %r494, %r621;

mul.lo.s32 %r734, %r530, %r681;
cvt.rn.f32.s32 %f386, %r734;
fma.rn.ftz.f32 %f387, %f153, %f386, %f383;

dp4a.s32.s32 %r534, %r611, %r494, %r621;

cvt.rn.f32.s32 %f388, %r534;
fma.rn.ftz.f32 %f389, %f153, %f388, %f385;

dp4a.s32.s32 %r538, %r615, %r502, %r621;

mul.lo.s32 %r735, %r538, %r690;
cvt.rn.f32.s32 %f390, %r735;
fma.rn.ftz.f32 %f391, %f154, %f390, %f387;

dp4a.s32.s32 %r542, %r619, %r502, %r621;

cvt.rn.f32.s32 %f392, %r542;
fma.rn.ftz.f32 %f393, %f154, %f392, %f389;

{.reg .f16 low,high;
mov.b32 {low,high},%r166;
cvt.f32.f16 %f155, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r166;
cvt.f32.f16 %f156, high;}


mul.ftz.f32 %f394, %f391, %f155;
mul.ftz.f32 %f395, %f393, %f156;
sub.ftz.f32 %f396, %f394, %f395;
add.ftz.f32 %f605, %f396, %f605;
cvt.u64.u32 %rd72, %r1004;
add.s64 %rd73, %rd1, %rd72;
add.s64 %rd74, %rd73, %rd4;
mul.lo.s64 %rd75, %rd74, 36;
add.s64 %rd76, %rd11, %rd75;
add.s64 %rd77, %rd76, %rd35;
ld.global.nc.u32 %r554, [%rd77+4];
ld.global.nc.u32 %r548, [%rd76];

{.reg .f16 low,high;
mov.b32 {low,high},%r548;
cvt.f32.f16 %f157, low;}


ld.global.nc.u32 %r562, [%rd77+40];
ld.global.nc.u32 %r549, [%rd76+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r549;
cvt.f32.f16 %f158, low;}


ld.global.nc.u32 %r570, [%rd77+76];
ld.global.nc.u32 %r550, [%rd76+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r550;
cvt.f32.f16 %f159, low;}


ld.global.nc.u32 %r578, [%rd77+112];
ld.global.nc.u32 %r551, [%rd76+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r551;
cvt.f32.f16 %f160, low;}



dp4a.s32.s32 %r552, %r553, %r554, %r621;

mul.lo.s32 %r736, %r552, %r627;
cvt.rn.f32.s32 %f397, %r736;
fma.rn.ftz.f32 %f398, %f157, %f397, 0f00000000;

dp4a.s32.s32 %r556, %r557, %r554, %r621;

cvt.rn.f32.s32 %f399, %r556;
fma.rn.ftz.f32 %f400, %f157, %f399, 0f00000000;

dp4a.s32.s32 %r560, %r561, %r562, %r621;

mul.lo.s32 %r737, %r560, %r636;
cvt.rn.f32.s32 %f401, %r737;
fma.rn.ftz.f32 %f402, %f158, %f401, %f398;

dp4a.s32.s32 %r564, %r565, %r562, %r621;

cvt.rn.f32.s32 %f403, %r564;
fma.rn.ftz.f32 %f404, %f158, %f403, %f400;

dp4a.s32.s32 %r568, %r569, %r570, %r621;

mul.lo.s32 %r738, %r568, %r645;
cvt.rn.f32.s32 %f405, %r738;
fma.rn.ftz.f32 %f406, %f159, %f405, %f402;

dp4a.s32.s32 %r572, %r573, %r570, %r621;

cvt.rn.f32.s32 %f407, %r572;
fma.rn.ftz.f32 %f408, %f159, %f407, %f404;

dp4a.s32.s32 %r576, %r577, %r578, %r621;

mul.lo.s32 %r739, %r576, %r654;
cvt.rn.f32.s32 %f409, %r739;
fma.rn.ftz.f32 %f410, %f160, %f409, %f406;

dp4a.s32.s32 %r580, %r581, %r578, %r621;

cvt.rn.f32.s32 %f411, %r580;
fma.rn.ftz.f32 %f412, %f160, %f411, %f408;

{.reg .f16 low,high;
mov.b32 {low,high},%r128;
cvt.f32.f16 %f161, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r128;
cvt.f32.f16 %f162, high;}


mul.ftz.f32 %f413, %f410, %f161;
mul.ftz.f32 %f414, %f412, %f162;
sub.ftz.f32 %f415, %f413, %f414;
add.ftz.f32 %f604, %f415, %f604;

{.reg .f16 low,high;
mov.b32 {low,high},%r548;
cvt.f32.f16 %f163, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r549;
cvt.f32.f16 %f164, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r550;
cvt.f32.f16 %f165, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r551;
cvt.f32.f16 %f166, low;}



dp4a.s32.s32 %r590, %r591, %r554, %r621;

mul.lo.s32 %r740, %r590, %r663;
cvt.rn.f32.s32 %f416, %r740;
fma.rn.ftz.f32 %f417, %f163, %f416, 0f00000000;

dp4a.s32.s32 %r594, %r595, %r554, %r621;

cvt.rn.f32.s32 %f418, %r594;
fma.rn.ftz.f32 %f419, %f163, %f418, 0f00000000;

dp4a.s32.s32 %r598, %r599, %r562, %r621;

mul.lo.s32 %r741, %r598, %r672;
cvt.rn.f32.s32 %f420, %r741;
fma.rn.ftz.f32 %f421, %f164, %f420, %f417;

dp4a.s32.s32 %r602, %r603, %r562, %r621;

cvt.rn.f32.s32 %f422, %r602;
fma.rn.ftz.f32 %f423, %f164, %f422, %f419;

dp4a.s32.s32 %r606, %r607, %r570, %r621;

mul.lo.s32 %r742, %r606, %r681;
cvt.rn.f32.s32 %f424, %r742;
fma.rn.ftz.f32 %f425, %f165, %f424, %f421;

dp4a.s32.s32 %r610, %r611, %r570, %r621;

cvt.rn.f32.s32 %f426, %r610;
fma.rn.ftz.f32 %f427, %f165, %f426, %f423;

dp4a.s32.s32 %r614, %r615, %r578, %r621;

mul.lo.s32 %r743, %r614, %r690;
cvt.rn.f32.s32 %f428, %r743;
fma.rn.ftz.f32 %f429, %f166, %f428, %f425;

dp4a.s32.s32 %r618, %r619, %r578, %r621;

cvt.rn.f32.s32 %f430, %r618;
fma.rn.ftz.f32 %f431, %f166, %f430, %f427;

{.reg .f16 low,high;
mov.b32 {low,high},%r166;
cvt.f32.f16 %f167, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r166;
cvt.f32.f16 %f168, high;}


mul.ftz.f32 %f432, %f429, %f167;
mul.ftz.f32 %f433, %f431, %f168;
sub.ftz.f32 %f434, %f432, %f433;
add.ftz.f32 %f603, %f434, %f603;
add.s32 %r1008, %r1008, 32;
add.s32 %r1007, %r1007, 32;
add.s32 %r1006, %r1006, 32;
add.s32 %r1005, %r1005, 32;
add.s32 %r1004, %r1004, 32;
add.s32 %r1003, %r1003, 32;
add.s32 %r1002, %r1002, 32;
add.s64 %rd122, %rd122, 336;
add.s32 %r1009, %r1009, 4;
setp.lt.s32 %p4, %r1009, %r6;
@%p4 bra $L__BB138_2;

$L__BB138_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB138_5;

mov.u32 %r744, _ZZ13mul_mat_vec_qIL9ggml_type10ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r745, %r1, 1792, %r744;
shl.b32 %r746, %r2, 2;
add.s32 %r747, %r745, %r746;
st.shared.f32 [%r747+-1792], %f616;
st.shared.f32 [%r747+-1664], %f615;
st.shared.f32 [%r747+-1536], %f614;
st.shared.f32 [%r747+-1408], %f613;
st.shared.f32 [%r747+-1280], %f612;
st.shared.f32 [%r747+-1152], %f611;
st.shared.f32 [%r747+-1024], %f610;
st.shared.f32 [%r747+-896], %f609;
st.shared.f32 [%r747+-768], %f608;
st.shared.f32 [%r747+-640], %f607;
st.shared.f32 [%r747+-512], %f606;
st.shared.f32 [%r747+-384], %f605;
st.shared.f32 [%r747+-256], %f604;
st.shared.f32 [%r747+-128], %f603;

$L__BB138_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB138_34;

mov.u32 %r748, %ctaid.z;
mov.u32 %r749, %ctaid.y;
mad.lo.s32 %r750, %r749, %r41, %r4;
mad.lo.s32 %r751, %r748, %r47, %r750;
cvt.u64.u32 %rd14, %r751;
shl.b32 %r752, %r2, 2;
mov.u32 %r753, _ZZ13mul_mat_vec_qIL9ggml_type10ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r30, %r753, %r752;
mov.u32 %r754, 2;
ld.shared.f32 %f435, [%r30];
add.ftz.f32 %f436, %f435, %f616;
mov.b32 %r755, %f436;
mov.u32 %r756, 31;
mov.u32 %r757, 16;
mov.u32 %r758, -1;
shfl.sync.bfly.b32 %r759|%p7, %r755, %r757, %r756, %r758;
mov.b32 %f437, %r759;
add.ftz.f32 %f438, %f436, %f437;
mov.b32 %r760, %f438;
mov.u32 %r761, 8;
shfl.sync.bfly.b32 %r762|%p8, %r760, %r761, %r756, %r758;
mov.b32 %f439, %r762;
add.ftz.f32 %f440, %f438, %f439;
mov.b32 %r763, %f440;
mov.u32 %r764, 4;
shfl.sync.bfly.b32 %r765|%p9, %r763, %r764, %r756, %r758;
mov.b32 %f441, %r765;
add.ftz.f32 %f442, %f440, %f441;
mov.b32 %r766, %f442;
shfl.sync.bfly.b32 %r767|%p10, %r766, %r754, %r756, %r758;
mov.b32 %f443, %r767;
add.ftz.f32 %f444, %f442, %f443;
mov.b32 %r768, %f444;
mov.u32 %r769, 1;
shfl.sync.bfly.b32 %r770|%p11, %r768, %r769, %r756, %r758;
mov.b32 %f445, %r770;
add.ftz.f32 %f43, %f444, %f445;
setp.lt.u32 %p12, %r4, %r35;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
cvta.to.global.u64 %rd78, %rd21;
mul.wide.u32 %rd79, %r751, 4;
add.s64 %rd15, %rd78, %rd79;
not.pred %p14, %p1;
@%p14 bra $L__BB138_8;

st.global.f32 [%rd15], %f43;

$L__BB138_8:
ld.shared.f32 %f446, [%r30+128];
add.ftz.f32 %f447, %f446, %f615;
mov.b32 %r771, %f447;
shfl.sync.bfly.b32 %r775|%p15, %r771, %r757, %r756, %r758;
mov.b32 %f448, %r775;
add.ftz.f32 %f449, %f447, %f448;
mov.b32 %r776, %f449;
shfl.sync.bfly.b32 %r778|%p16, %r776, %r761, %r756, %r758;
mov.b32 %f450, %r778;
add.ftz.f32 %f451, %f449, %f450;
mov.b32 %r779, %f451;
shfl.sync.bfly.b32 %r781|%p17, %r779, %r764, %r756, %r758;
mov.b32 %f452, %r781;
add.ftz.f32 %f453, %f451, %f452;
mov.b32 %r782, %f453;
shfl.sync.bfly.b32 %r784|%p18, %r782, %r754, %r756, %r758;
mov.b32 %f454, %r784;
add.ftz.f32 %f455, %f453, %f454;
mov.b32 %r785, %f455;
shfl.sync.bfly.b32 %r787|%p19, %r785, %r769, %r756, %r758;
mov.b32 %f456, %r787;
add.ftz.f32 %f44, %f455, %f456;
setp.eq.s32 %p20, %r2, 1;
or.b32 %r791, %r4, 1;
setp.lt.u32 %p21, %r791, %r35;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB138_10;

st.global.f32 [%rd15+4], %f44;

$L__BB138_10:
ld.shared.f32 %f457, [%r30+256];
add.ftz.f32 %f458, %f457, %f614;
mov.b32 %r792, %f458;
mov.u32 %r793, 31;
mov.u32 %r794, 16;
mov.u32 %r795, -1;
shfl.sync.bfly.b32 %r796|%p23, %r792, %r794, %r793, %r795;
mov.b32 %f459, %r796;
add.ftz.f32 %f460, %f458, %f459;
mov.b32 %r797, %f460;
mov.u32 %r798, 8;
shfl.sync.bfly.b32 %r799|%p24, %r797, %r798, %r793, %r795;
mov.b32 %f461, %r799;
add.ftz.f32 %f462, %f460, %f461;
mov.b32 %r800, %f462;
mov.u32 %r801, 4;
shfl.sync.bfly.b32 %r802|%p25, %r800, %r801, %r793, %r795;
mov.b32 %f463, %r802;
add.ftz.f32 %f464, %f462, %f463;
mov.b32 %r803, %f464;
mov.u32 %r804, 2;
shfl.sync.bfly.b32 %r805|%p26, %r803, %r804, %r793, %r795;
mov.b32 %f465, %r805;
add.ftz.f32 %f466, %f464, %f465;
mov.b32 %r806, %f466;
mov.u32 %r807, 1;
shfl.sync.bfly.b32 %r808|%p27, %r806, %r807, %r793, %r795;
mov.b32 %f467, %r808;
add.ftz.f32 %f45, %f466, %f467;
@%p14 bra $L__BB138_12;

cvt.u64.u32 %rd80, %r35;
add.s64 %rd81, %rd80, %rd14;
shl.b64 %rd83, %rd81, 2;
add.s64 %rd84, %rd78, %rd83;
st.global.f32 [%rd84], %f45;

$L__BB138_12:
ld.shared.f32 %f468, [%r30+384];
add.ftz.f32 %f469, %f468, %f613;
mov.b32 %r809, %f469;
shfl.sync.bfly.b32 %r813|%p29, %r809, %r794, %r793, %r795;
mov.b32 %f470, %r813;
add.ftz.f32 %f471, %f469, %f470;
mov.b32 %r814, %f471;
shfl.sync.bfly.b32 %r816|%p30, %r814, %r798, %r793, %r795;
mov.b32 %f472, %r816;
add.ftz.f32 %f473, %f471, %f472;
mov.b32 %r817, %f473;
shfl.sync.bfly.b32 %r819|%p31, %r817, %r801, %r793, %r795;
mov.b32 %f474, %r819;
add.ftz.f32 %f475, %f473, %f474;
mov.b32 %r820, %f475;
shfl.sync.bfly.b32 %r822|%p32, %r820, %r804, %r793, %r795;
mov.b32 %f476, %r822;
add.ftz.f32 %f477, %f475, %f476;
mov.b32 %r823, %f477;
shfl.sync.bfly.b32 %r825|%p33, %r823, %r807, %r793, %r795;
mov.b32 %f478, %r825;
add.ftz.f32 %f46, %f477, %f478;
@%p22 bra $L__BB138_14;

add.s32 %r826, %r35, 1;
cvt.u64.u32 %rd85, %r826;
add.s64 %rd86, %rd85, %rd14;
shl.b64 %rd88, %rd86, 2;
add.s64 %rd89, %rd78, %rd88;
st.global.f32 [%rd89], %f46;

$L__BB138_14:
shl.b32 %r827, %r35, 1;
mov.u32 %r828, 1;
ld.shared.f32 %f479, [%r30+512];
add.ftz.f32 %f480, %f479, %f612;
mov.b32 %r829, %f480;
mov.u32 %r830, 31;
mov.u32 %r831, 16;
mov.u32 %r832, -1;
shfl.sync.bfly.b32 %r833|%p35, %r829, %r831, %r830, %r832;
mov.b32 %f481, %r833;
add.ftz.f32 %f482, %f480, %f481;
mov.b32 %r834, %f482;
mov.u32 %r835, 8;
shfl.sync.bfly.b32 %r836|%p36, %r834, %r835, %r830, %r832;
mov.b32 %f483, %r836;
add.ftz.f32 %f484, %f482, %f483;
mov.b32 %r837, %f484;
mov.u32 %r838, 4;
shfl.sync.bfly.b32 %r839|%p37, %r837, %r838, %r830, %r832;
mov.b32 %f485, %r839;
add.ftz.f32 %f486, %f484, %f485;
mov.b32 %r840, %f486;
mov.u32 %r841, 2;
shfl.sync.bfly.b32 %r842|%p38, %r840, %r841, %r830, %r832;
mov.b32 %f487, %r842;
add.ftz.f32 %f488, %f486, %f487;
mov.b32 %r843, %f488;
shfl.sync.bfly.b32 %r844|%p39, %r843, %r828, %r830, %r832;
mov.b32 %f489, %r844;
add.ftz.f32 %f47, %f488, %f489;
cvt.u64.u32 %rd90, %r827;
add.s64 %rd91, %rd90, %rd14;
shl.b64 %rd93, %rd91, 2;
add.s64 %rd16, %rd78, %rd93;
@%p14 bra $L__BB138_16;

st.global.f32 [%rd16], %f47;

$L__BB138_16:
ld.shared.f32 %f490, [%r30+640];
add.ftz.f32 %f491, %f490, %f611;
mov.b32 %r845, %f491;
shfl.sync.bfly.b32 %r849|%p41, %r845, %r831, %r830, %r832;
mov.b32 %f492, %r849;
add.ftz.f32 %f493, %f491, %f492;
mov.b32 %r850, %f493;
shfl.sync.bfly.b32 %r852|%p42, %r850, %r835, %r830, %r832;
mov.b32 %f494, %r852;
add.ftz.f32 %f495, %f493, %f494;
mov.b32 %r853, %f495;
shfl.sync.bfly.b32 %r855|%p43, %r853, %r838, %r830, %r832;
mov.b32 %f496, %r855;
add.ftz.f32 %f497, %f495, %f496;
mov.b32 %r856, %f497;
shfl.sync.bfly.b32 %r858|%p44, %r856, %r841, %r830, %r832;
mov.b32 %f498, %r858;
add.ftz.f32 %f499, %f497, %f498;
mov.b32 %r859, %f499;
shfl.sync.bfly.b32 %r861|%p45, %r859, %r828, %r830, %r832;
mov.b32 %f500, %r861;
add.ftz.f32 %f48, %f499, %f500;
@%p22 bra $L__BB138_18;

st.global.f32 [%rd16+4], %f48;

$L__BB138_18:
mul.lo.s32 %r31, %r35, 3;
ld.shared.f32 %f501, [%r30+768];
add.ftz.f32 %f502, %f501, %f610;
mov.b32 %r862, %f502;
mov.u32 %r863, 31;
mov.u32 %r864, 16;
mov.u32 %r865, -1;
shfl.sync.bfly.b32 %r866|%p47, %r862, %r864, %r863, %r865;
mov.b32 %f503, %r866;
add.ftz.f32 %f504, %f502, %f503;
mov.b32 %r867, %f504;
mov.u32 %r868, 8;
shfl.sync.bfly.b32 %r869|%p48, %r867, %r868, %r863, %r865;
mov.b32 %f505, %r869;
add.ftz.f32 %f506, %f504, %f505;
mov.b32 %r870, %f506;
mov.u32 %r871, 4;
shfl.sync.bfly.b32 %r872|%p49, %r870, %r871, %r863, %r865;
mov.b32 %f507, %r872;
add.ftz.f32 %f508, %f506, %f507;
mov.b32 %r873, %f508;
mov.u32 %r874, 2;
shfl.sync.bfly.b32 %r875|%p50, %r873, %r874, %r863, %r865;
mov.b32 %f509, %r875;
add.ftz.f32 %f510, %f508, %f509;
mov.b32 %r876, %f510;
mov.u32 %r877, 1;
shfl.sync.bfly.b32 %r878|%p51, %r876, %r877, %r863, %r865;
mov.b32 %f511, %r878;
add.ftz.f32 %f49, %f510, %f511;
@%p14 bra $L__BB138_20;

cvt.u64.u32 %rd94, %r31;
add.s64 %rd95, %rd94, %rd14;
shl.b64 %rd97, %rd95, 2;
add.s64 %rd98, %rd78, %rd97;
st.global.f32 [%rd98], %f49;

$L__BB138_20:
ld.shared.f32 %f512, [%r30+896];
add.ftz.f32 %f513, %f512, %f609;
mov.b32 %r879, %f513;
shfl.sync.bfly.b32 %r883|%p53, %r879, %r864, %r863, %r865;
mov.b32 %f514, %r883;
add.ftz.f32 %f515, %f513, %f514;
mov.b32 %r884, %f515;
shfl.sync.bfly.b32 %r886|%p54, %r884, %r868, %r863, %r865;
mov.b32 %f516, %r886;
add.ftz.f32 %f517, %f515, %f516;
mov.b32 %r887, %f517;
shfl.sync.bfly.b32 %r889|%p55, %r887, %r871, %r863, %r865;
mov.b32 %f518, %r889;
add.ftz.f32 %f519, %f517, %f518;
mov.b32 %r890, %f519;
shfl.sync.bfly.b32 %r892|%p56, %r890, %r874, %r863, %r865;
mov.b32 %f520, %r892;
add.ftz.f32 %f521, %f519, %f520;
mov.b32 %r893, %f521;
shfl.sync.bfly.b32 %r895|%p57, %r893, %r877, %r863, %r865;
mov.b32 %f522, %r895;
add.ftz.f32 %f50, %f521, %f522;
@%p22 bra $L__BB138_22;

add.s32 %r896, %r31, 1;
cvt.u64.u32 %rd99, %r896;
add.s64 %rd100, %rd99, %rd14;
shl.b64 %rd102, %rd100, 2;
add.s64 %rd103, %rd78, %rd102;
st.global.f32 [%rd103], %f50;

$L__BB138_22:
shl.b32 %r897, %r35, 2;
mov.u32 %r898, 2;
ld.shared.f32 %f523, [%r30+1024];
add.ftz.f32 %f524, %f523, %f608;
mov.b32 %r899, %f524;
mov.u32 %r900, 31;
mov.u32 %r901, 16;
mov.u32 %r902, -1;
shfl.sync.bfly.b32 %r903|%p59, %r899, %r901, %r900, %r902;
mov.b32 %f525, %r903;
add.ftz.f32 %f526, %f524, %f525;
mov.b32 %r904, %f526;
mov.u32 %r905, 8;
shfl.sync.bfly.b32 %r906|%p60, %r904, %r905, %r900, %r902;
mov.b32 %f527, %r906;
add.ftz.f32 %f528, %f526, %f527;
mov.b32 %r907, %f528;
mov.u32 %r908, 4;
shfl.sync.bfly.b32 %r909|%p61, %r907, %r908, %r900, %r902;
mov.b32 %f529, %r909;
add.ftz.f32 %f530, %f528, %f529;
mov.b32 %r910, %f530;
shfl.sync.bfly.b32 %r911|%p62, %r910, %r898, %r900, %r902;
mov.b32 %f531, %r911;
add.ftz.f32 %f532, %f530, %f531;
mov.b32 %r912, %f532;
mov.u32 %r913, 1;
shfl.sync.bfly.b32 %r914|%p63, %r912, %r913, %r900, %r902;
mov.b32 %f533, %r914;
add.ftz.f32 %f51, %f532, %f533;
cvt.u64.u32 %rd104, %r897;
add.s64 %rd105, %rd104, %rd14;
shl.b64 %rd107, %rd105, 2;
add.s64 %rd17, %rd78, %rd107;
@%p14 bra $L__BB138_24;

st.global.f32 [%rd17], %f51;

$L__BB138_24:
ld.shared.f32 %f534, [%r30+1152];
add.ftz.f32 %f535, %f534, %f607;
mov.b32 %r915, %f535;
shfl.sync.bfly.b32 %r919|%p65, %r915, %r901, %r900, %r902;
mov.b32 %f536, %r919;
add.ftz.f32 %f537, %f535, %f536;
mov.b32 %r920, %f537;
shfl.sync.bfly.b32 %r922|%p66, %r920, %r905, %r900, %r902;
mov.b32 %f538, %r922;
add.ftz.f32 %f539, %f537, %f538;
mov.b32 %r923, %f539;
shfl.sync.bfly.b32 %r925|%p67, %r923, %r908, %r900, %r902;
mov.b32 %f540, %r925;
add.ftz.f32 %f541, %f539, %f540;
mov.b32 %r926, %f541;
shfl.sync.bfly.b32 %r928|%p68, %r926, %r898, %r900, %r902;
mov.b32 %f542, %r928;
add.ftz.f32 %f543, %f541, %f542;
mov.b32 %r929, %f543;
shfl.sync.bfly.b32 %r931|%p69, %r929, %r913, %r900, %r902;
mov.b32 %f544, %r931;
add.ftz.f32 %f52, %f543, %f544;
@%p22 bra $L__BB138_26;

st.global.f32 [%rd17+4], %f52;

$L__BB138_26:
mul.lo.s32 %r32, %r35, 5;
ld.shared.f32 %f545, [%r30+1280];
add.ftz.f32 %f546, %f545, %f606;
mov.b32 %r932, %f546;
mov.u32 %r933, 31;
mov.u32 %r934, 16;
mov.u32 %r935, -1;
shfl.sync.bfly.b32 %r936|%p71, %r932, %r934, %r933, %r935;
mov.b32 %f547, %r936;
add.ftz.f32 %f548, %f546, %f547;
mov.b32 %r937, %f548;
mov.u32 %r938, 8;
shfl.sync.bfly.b32 %r939|%p72, %r937, %r938, %r933, %r935;
mov.b32 %f549, %r939;
add.ftz.f32 %f550, %f548, %f549;
mov.b32 %r940, %f550;
mov.u32 %r941, 4;
shfl.sync.bfly.b32 %r942|%p73, %r940, %r941, %r933, %r935;
mov.b32 %f551, %r942;
add.ftz.f32 %f552, %f550, %f551;
mov.b32 %r943, %f552;
mov.u32 %r944, 2;
shfl.sync.bfly.b32 %r945|%p74, %r943, %r944, %r933, %r935;
mov.b32 %f553, %r945;
add.ftz.f32 %f554, %f552, %f553;
mov.b32 %r946, %f554;
mov.u32 %r947, 1;
shfl.sync.bfly.b32 %r948|%p75, %r946, %r947, %r933, %r935;
mov.b32 %f555, %r948;
add.ftz.f32 %f53, %f554, %f555;
@%p14 bra $L__BB138_28;

cvt.u64.u32 %rd108, %r32;
add.s64 %rd109, %rd108, %rd14;
shl.b64 %rd111, %rd109, 2;
add.s64 %rd112, %rd78, %rd111;
st.global.f32 [%rd112], %f53;

$L__BB138_28:
ld.shared.f32 %f556, [%r30+1408];
add.ftz.f32 %f557, %f556, %f605;
mov.b32 %r949, %f557;
shfl.sync.bfly.b32 %r953|%p77, %r949, %r934, %r933, %r935;
mov.b32 %f558, %r953;
add.ftz.f32 %f559, %f557, %f558;
mov.b32 %r954, %f559;
shfl.sync.bfly.b32 %r956|%p78, %r954, %r938, %r933, %r935;
mov.b32 %f560, %r956;
add.ftz.f32 %f561, %f559, %f560;
mov.b32 %r957, %f561;
shfl.sync.bfly.b32 %r959|%p79, %r957, %r941, %r933, %r935;
mov.b32 %f562, %r959;
add.ftz.f32 %f563, %f561, %f562;
mov.b32 %r960, %f563;
shfl.sync.bfly.b32 %r962|%p80, %r960, %r944, %r933, %r935;
mov.b32 %f564, %r962;
add.ftz.f32 %f565, %f563, %f564;
mov.b32 %r963, %f565;
shfl.sync.bfly.b32 %r965|%p81, %r963, %r947, %r933, %r935;
mov.b32 %f566, %r965;
add.ftz.f32 %f54, %f565, %f566;
@%p22 bra $L__BB138_30;

add.s32 %r966, %r32, 1;
cvt.u64.u32 %rd113, %r966;
add.s64 %rd114, %rd113, %rd14;
shl.b64 %rd116, %rd114, 2;
add.s64 %rd117, %rd78, %rd116;
st.global.f32 [%rd117], %f54;

$L__BB138_30:
mul.lo.s32 %r967, %r35, 6;
ld.shared.f32 %f567, [%r30+1536];
add.ftz.f32 %f568, %f567, %f604;
mov.b32 %r968, %f568;
mov.u32 %r969, 31;
mov.u32 %r970, 16;
mov.u32 %r971, -1;
shfl.sync.bfly.b32 %r972|%p83, %r968, %r970, %r969, %r971;
mov.b32 %f569, %r972;
add.ftz.f32 %f570, %f568, %f569;
mov.b32 %r973, %f570;
mov.u32 %r974, 8;
shfl.sync.bfly.b32 %r975|%p84, %r973, %r974, %r969, %r971;
mov.b32 %f571, %r975;
add.ftz.f32 %f572, %f570, %f571;
mov.b32 %r976, %f572;
mov.u32 %r977, 4;
shfl.sync.bfly.b32 %r978|%p85, %r976, %r977, %r969, %r971;
mov.b32 %f573, %r978;
add.ftz.f32 %f574, %f572, %f573;
mov.b32 %r979, %f574;
mov.u32 %r980, 2;
shfl.sync.bfly.b32 %r981|%p86, %r979, %r980, %r969, %r971;
mov.b32 %f575, %r981;
add.ftz.f32 %f576, %f574, %f575;
mov.b32 %r982, %f576;
mov.u32 %r983, 1;
shfl.sync.bfly.b32 %r984|%p87, %r982, %r983, %r969, %r971;
mov.b32 %f577, %r984;
add.ftz.f32 %f55, %f576, %f577;
cvt.u64.u32 %rd118, %r967;
add.s64 %rd119, %rd118, %rd14;
shl.b64 %rd121, %rd119, 2;
add.s64 %rd18, %rd78, %rd121;
@%p14 bra $L__BB138_32;

st.global.f32 [%rd18], %f55;

$L__BB138_32:
ld.shared.f32 %f578, [%r30+1664];
add.ftz.f32 %f579, %f578, %f603;
mov.b32 %r985, %f579;
shfl.sync.bfly.b32 %r989|%p89, %r985, %r970, %r969, %r971;
mov.b32 %f580, %r989;
add.ftz.f32 %f581, %f579, %f580;
mov.b32 %r990, %f581;
shfl.sync.bfly.b32 %r992|%p90, %r990, %r974, %r969, %r971;
mov.b32 %f582, %r992;
add.ftz.f32 %f583, %f581, %f582;
mov.b32 %r993, %f583;
shfl.sync.bfly.b32 %r995|%p91, %r993, %r977, %r969, %r971;
mov.b32 %f584, %r995;
add.ftz.f32 %f585, %f583, %f584;
mov.b32 %r996, %f585;
shfl.sync.bfly.b32 %r998|%p92, %r996, %r980, %r969, %r971;
mov.b32 %f586, %r998;
add.ftz.f32 %f587, %f585, %f586;
mov.b32 %r999, %f587;
shfl.sync.bfly.b32 %r1001|%p93, %r999, %r983, %r969, %r971;
mov.b32 %f588, %r1001;
add.ftz.f32 %f56, %f587, %f588;
@%p22 bra $L__BB138_34;

st.global.f32 [%rd18+4], %f56;

$L__BB138_34:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type10ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type10ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type10ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type10ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type10ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type10ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type10ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type10ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type10ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type10ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<107>;
.reg .b16 %rs<9>;
.reg .f32 %f<705>;
.reg .b32 %r<1135>;
.reg .b64 %rd<139>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type10ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[2048];

ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type10ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd20, [_Z13mul_mat_vec_qIL9ggml_type10ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd21, [_Z13mul_mat_vec_qIL9ggml_type10ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r52, [_Z13mul_mat_vec_qIL9ggml_type10ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type10ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type10ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type10ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type10ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type10ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r45, [_Z13mul_mat_vec_qIL9ggml_type10ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r49, [_Z13mul_mat_vec_qIL9ggml_type10ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r50, [_Z13mul_mat_vec_qIL9ggml_type10ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r51, [_Z13mul_mat_vec_qIL9ggml_type10ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r47, [_Z13mul_mat_vec_qIL9ggml_type10ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r46, [_Z13mul_mat_vec_qIL9ggml_type10ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type10ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type10ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r53, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r53, %r2;
mov.u32 %r54, %ctaid.x;
shl.b32 %r4, %r54, 1;
shr.s32 %r55, %r3, 31;
shr.u32 %r56, %r55, 28;
add.s32 %r57, %r3, %r56;
shr.s32 %r1134, %r57, 4;
shr.u32 %r6, %r52, 8;
setp.ge.s32 %p3, %r1134, %r6;
mov.f32 %f689, 0f00000000;
mov.f32 %f690, %f689;
mov.f32 %f691, %f689;
mov.f32 %f692, %f689;
mov.f32 %f693, %f689;
mov.f32 %f694, %f689;
mov.f32 %f695, %f689;
mov.f32 %f696, %f689;
mov.f32 %f697, %f689;
mov.f32 %f698, %f689;
mov.f32 %f699, %f689;
mov.f32 %f700, %f689;
mov.f32 %f701, %f689;
mov.f32 %f702, %f689;
mov.f32 %f703, %f689;
mov.f32 %f704, %f689;
@%p3 bra $L__BB139_3;

mov.u32 %r58, %ctaid.y;
mov.u32 %r59, %ctaid.z;
mul.lo.s32 %r60, %r58, %r44;
cvt.u64.u32 %rd22, %r60;
mul.lo.s32 %r61, %r59, %r50;
cvt.u64.u32 %rd23, %r61;
add.s64 %rd1, %rd23, %rd22;
mul.hi.u32 %r62, %r59, %r46;
add.s32 %r63, %r62, %r59;
shr.u32 %r64, %r63, %r47;
mul.lo.s32 %r65, %r64, %r49;
mul.hi.u32 %r66, %r58, %r40;
add.s32 %r67, %r66, %r58;
shr.u32 %r68, %r67, %r41;
mul.lo.s32 %r69, %r68, %r43;
cvta.to.global.u64 %rd138, %rd19;
add.s32 %r70, %r4, 1;
and.b32 %r74, %r57, -16;
sub.s32 %r75, %r3, %r74;
shr.s32 %r76, %r75, 31;
shr.u32 %r77, %r76, 29;
add.s32 %r78, %r75, %r77;
and.b32 %r79, %r78, -8;
sub.s32 %r80, %r75, %r79;
shr.s32 %r81, %r80, 31;
shr.u32 %r82, %r81, 30;
add.s32 %r83, %r80, %r82;
shr.s32 %r84, %r83, 2;
add.s32 %r85, %r79, %r84;
cvt.s64.s32 %rd24, %r85;
shr.s32 %r86, %r78, 3;
shl.b32 %r87, %r86, 2;
cvt.s64.s32 %rd3, %r80;
cvt.s64.s32 %rd4, %r87;
shl.b32 %r88, %r38, 1;
shl.b32 %r1126, %r1134, 3;
add.s32 %r1133, %r1126, %r88;
mad.lo.s32 %r1132, %r38, 3, %r1126;
shl.b32 %r89, %r38, 2;
add.s32 %r1131, %r1126, %r89;
mad.lo.s32 %r1130, %r38, 5, %r1126;
mad.lo.s32 %r1129, %r38, 6, %r1126;
mad.lo.s32 %r1128, %r38, 7, %r1126;
add.s32 %r1127, %r38, %r1126;
add.s32 %r90, %r1134, %r65;
add.s32 %r91, %r90, %r69;
mad.lo.s32 %r92, %r70, %r37, %r91;
mul.wide.s32 %rd5, %r92, 84;
mad.lo.s32 %r93, %r4, %r37, %r1134;
add.s32 %r94, %r93, %r65;
add.s32 %r95, %r94, %r69;
mul.wide.s32 %rd6, %r95, 84;
mul.wide.s32 %rd25, %r75, 4;
add.s64 %rd26, %rd6, %rd25;
add.s64 %rd7, %rd26, 16;
add.s64 %rd27, %rd6, %rd24;
add.s64 %rd8, %rd27, 6;
add.s64 %rd9, %rd5, %rd24;
add.s64 %rd28, %rd5, %rd25;
add.s64 %rd10, %rd28, 16;
cvta.to.global.u64 %rd11, %rd20;
mov.f32 %f689, 0f00000000;
shl.b64 %rd35, %rd3, 2;

$L__BB139_2:
cvt.u64.u32 %rd29, %r1126;
add.s64 %rd30, %rd1, %rd29;
add.s64 %rd31, %rd138, %rd7;
ld.global.nc.u32 %r704, [%rd31];
add.s64 %rd32, %rd30, %rd4;
mul.lo.s64 %rd33, %rd32, 36;
add.s64 %rd34, %rd11, %rd33;
add.s64 %rd36, %rd34, %rd35;
ld.global.nc.u32 %r102, [%rd36+4];
ld.global.nc.u32 %r96, [%rd34];

{.reg .f16 low,high;
mov.b32 {low,high},%r96;
cvt.f32.f16 %f97, low;}


ld.global.nc.u32 %r110, [%rd36+40];
ld.global.nc.u32 %r97, [%rd34+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r97;
cvt.f32.f16 %f98, low;}


ld.global.nc.u32 %r118, [%rd36+76];
ld.global.nc.u32 %r98, [%rd34+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r98;
cvt.f32.f16 %f99, low;}


ld.global.nc.u32 %r126, [%rd36+112];
ld.global.nc.u32 %r99, [%rd34+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r99;
cvt.f32.f16 %f100, low;}


add.s64 %rd37, %rd138, %rd8;
ld.global.nc.u8 %rs1, [%rd37+-6];
cvt.u32.u16 %r705, %rs1;
and.b32 %r706, %r705, 240;
and.b32 %r633, %r704, 50529027;
mov.u32 %r701, 0;

dp4a.s32.s32 %r100, %r633, %r102, %r701;

and.b32 %r707, %r705, 15;
mul.lo.s32 %r708, %r100, %r707;
cvt.rn.f32.s32 %f193, %r708;
fma.rn.ftz.f32 %f194, %f97, %f193, 0f00000000;
shr.u32 %r709, %r706, 4;
shl.b32 %r710, %r709, 8;
or.b32 %r711, %r710, %r709;
shl.b32 %r712, %r711, 16;
or.b32 %r637, %r712, %r711;

dp4a.s32.s32 %r104, %r637, %r102, %r701;

cvt.rn.f32.s32 %f195, %r104;
fma.rn.ftz.f32 %f196, %f97, %f195, 0f00000000;
ld.global.nc.u8 %rs2, [%rd37+-4];
cvt.u32.u16 %r713, %rs2;
and.b32 %r714, %r713, 240;
shr.u32 %r715, %r704, 2;
and.b32 %r641, %r715, 50529027;

dp4a.s32.s32 %r108, %r641, %r110, %r701;

and.b32 %r716, %r713, 15;
mul.lo.s32 %r717, %r108, %r716;
cvt.rn.f32.s32 %f197, %r717;
fma.rn.ftz.f32 %f198, %f98, %f197, %f194;
shr.u32 %r718, %r714, 4;
shl.b32 %r719, %r718, 8;
or.b32 %r720, %r719, %r718;
shl.b32 %r721, %r720, 16;
or.b32 %r645, %r721, %r720;

dp4a.s32.s32 %r112, %r645, %r110, %r701;

cvt.rn.f32.s32 %f199, %r112;
fma.rn.ftz.f32 %f200, %f98, %f199, %f196;
ld.global.nc.u8 %rs3, [%rd37+-2];
cvt.u32.u16 %r722, %rs3;
and.b32 %r723, %r722, 240;
shr.u32 %r724, %r704, 4;
and.b32 %r649, %r724, 50529027;

dp4a.s32.s32 %r116, %r649, %r118, %r701;

and.b32 %r725, %r722, 15;
mul.lo.s32 %r726, %r116, %r725;
cvt.rn.f32.s32 %f201, %r726;
fma.rn.ftz.f32 %f202, %f99, %f201, %f198;
shr.u32 %r727, %r723, 4;
shl.b32 %r728, %r727, 8;
or.b32 %r729, %r728, %r727;
shl.b32 %r730, %r729, 16;
or.b32 %r653, %r730, %r729;

dp4a.s32.s32 %r120, %r653, %r118, %r701;

cvt.rn.f32.s32 %f203, %r120;
fma.rn.ftz.f32 %f204, %f99, %f203, %f200;
ld.global.nc.u8 %rs4, [%rd37];
cvt.u32.u16 %r731, %rs4;
and.b32 %r732, %r731, 240;
shr.u32 %r733, %r704, 6;
and.b32 %r657, %r733, 50529027;

dp4a.s32.s32 %r124, %r657, %r126, %r701;

and.b32 %r734, %r731, 15;
mul.lo.s32 %r735, %r124, %r734;
cvt.rn.f32.s32 %f205, %r735;
fma.rn.ftz.f32 %f206, %f100, %f205, %f202;
shr.u32 %r736, %r732, 4;
shl.b32 %r737, %r736, 8;
or.b32 %r738, %r737, %r736;
shl.b32 %r739, %r738, 16;
or.b32 %r661, %r739, %r738;

dp4a.s32.s32 %r128, %r661, %r126, %r701;

cvt.rn.f32.s32 %f207, %r128;
fma.rn.ftz.f32 %f208, %f100, %f207, %f204;
add.s64 %rd38, %rd138, %rd6;
ld.global.nc.u32 %r132, [%rd38+80];

{.reg .f16 low,high;
mov.b32 {low,high},%r132;
cvt.f32.f16 %f101, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r132;
cvt.f32.f16 %f102, high;}


mul.ftz.f32 %f209, %f206, %f101;
mul.ftz.f32 %f210, %f208, %f102;
sub.ftz.f32 %f211, %f209, %f210;
add.ftz.f32 %f704, %f211, %f704;
add.s64 %rd39, %rd138, %rd10;
ld.global.nc.u32 %r740, [%rd39];

{.reg .f16 low,high;
mov.b32 {low,high},%r96;
cvt.f32.f16 %f103, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r97;
cvt.f32.f16 %f104, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r98;
cvt.f32.f16 %f105, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r99;
cvt.f32.f16 %f106, low;}


add.s64 %rd40, %rd138, %rd9;
ld.global.nc.u8 %rs5, [%rd40];
cvt.u32.u16 %r741, %rs5;
and.b32 %r742, %r741, 240;
and.b32 %r671, %r740, 50529027;

dp4a.s32.s32 %r138, %r671, %r102, %r701;

and.b32 %r743, %r741, 15;
mul.lo.s32 %r744, %r138, %r743;
cvt.rn.f32.s32 %f212, %r744;
fma.rn.ftz.f32 %f213, %f103, %f212, 0f00000000;
shr.u32 %r745, %r742, 4;
shl.b32 %r746, %r745, 8;
or.b32 %r747, %r746, %r745;
shl.b32 %r748, %r747, 16;
or.b32 %r675, %r748, %r747;

dp4a.s32.s32 %r142, %r675, %r102, %r701;

cvt.rn.f32.s32 %f214, %r142;
fma.rn.ftz.f32 %f215, %f103, %f214, 0f00000000;
ld.global.nc.u8 %rs6, [%rd40+2];
cvt.u32.u16 %r749, %rs6;
and.b32 %r750, %r749, 240;
shr.u32 %r751, %r740, 2;
and.b32 %r679, %r751, 50529027;

dp4a.s32.s32 %r146, %r679, %r110, %r701;

and.b32 %r752, %r749, 15;
mul.lo.s32 %r753, %r146, %r752;
cvt.rn.f32.s32 %f216, %r753;
fma.rn.ftz.f32 %f217, %f104, %f216, %f213;
shr.u32 %r754, %r750, 4;
shl.b32 %r755, %r754, 8;
or.b32 %r756, %r755, %r754;
shl.b32 %r757, %r756, 16;
or.b32 %r683, %r757, %r756;

dp4a.s32.s32 %r150, %r683, %r110, %r701;

cvt.rn.f32.s32 %f218, %r150;
fma.rn.ftz.f32 %f219, %f104, %f218, %f215;
ld.global.nc.u8 %rs7, [%rd40+4];
cvt.u32.u16 %r758, %rs7;
and.b32 %r759, %r758, 240;
shr.u32 %r760, %r740, 4;
and.b32 %r687, %r760, 50529027;

dp4a.s32.s32 %r154, %r687, %r118, %r701;

and.b32 %r761, %r758, 15;
mul.lo.s32 %r762, %r154, %r761;
cvt.rn.f32.s32 %f220, %r762;
fma.rn.ftz.f32 %f221, %f105, %f220, %f217;
shr.u32 %r763, %r759, 4;
shl.b32 %r764, %r763, 8;
or.b32 %r765, %r764, %r763;
shl.b32 %r766, %r765, 16;
or.b32 %r691, %r766, %r765;

dp4a.s32.s32 %r158, %r691, %r118, %r701;

cvt.rn.f32.s32 %f222, %r158;
fma.rn.ftz.f32 %f223, %f105, %f222, %f219;
ld.global.nc.u8 %rs8, [%rd40+6];
cvt.u32.u16 %r767, %rs8;
and.b32 %r768, %r767, 240;
shr.u32 %r769, %r740, 6;
and.b32 %r695, %r769, 50529027;

dp4a.s32.s32 %r162, %r695, %r126, %r701;

and.b32 %r770, %r767, 15;
mul.lo.s32 %r771, %r162, %r770;
cvt.rn.f32.s32 %f224, %r771;
fma.rn.ftz.f32 %f225, %f106, %f224, %f221;
shr.u32 %r772, %r768, 4;
shl.b32 %r773, %r772, 8;
or.b32 %r774, %r773, %r772;
shl.b32 %r775, %r774, 16;
or.b32 %r699, %r775, %r774;

dp4a.s32.s32 %r166, %r699, %r126, %r701;

cvt.rn.f32.s32 %f226, %r166;
fma.rn.ftz.f32 %f227, %f106, %f226, %f223;
add.s64 %rd41, %rd138, %rd5;
ld.global.nc.u32 %r170, [%rd41+80];

{.reg .f16 low,high;
mov.b32 {low,high},%r170;
cvt.f32.f16 %f107, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r170;
cvt.f32.f16 %f108, high;}


mul.ftz.f32 %f228, %f225, %f107;
mul.ftz.f32 %f229, %f227, %f108;
sub.ftz.f32 %f230, %f228, %f229;
add.ftz.f32 %f703, %f230, %f703;
cvt.u64.u32 %rd42, %r1127;
add.s64 %rd43, %rd1, %rd42;
add.s64 %rd44, %rd43, %rd4;
mul.lo.s64 %rd45, %rd44, 36;
add.s64 %rd46, %rd11, %rd45;
add.s64 %rd47, %rd46, %rd35;
ld.global.nc.u32 %r178, [%rd47+4];
ld.global.nc.u32 %r172, [%rd46];

{.reg .f16 low,high;
mov.b32 {low,high},%r172;
cvt.f32.f16 %f109, low;}


ld.global.nc.u32 %r186, [%rd47+40];
ld.global.nc.u32 %r173, [%rd46+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r173;
cvt.f32.f16 %f110, low;}


ld.global.nc.u32 %r194, [%rd47+76];
ld.global.nc.u32 %r174, [%rd46+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r174;
cvt.f32.f16 %f111, low;}


ld.global.nc.u32 %r202, [%rd47+112];
ld.global.nc.u32 %r175, [%rd46+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r175;
cvt.f32.f16 %f112, low;}



dp4a.s32.s32 %r176, %r633, %r178, %r701;

mul.lo.s32 %r776, %r176, %r707;
cvt.rn.f32.s32 %f231, %r776;
fma.rn.ftz.f32 %f232, %f109, %f231, 0f00000000;

dp4a.s32.s32 %r180, %r637, %r178, %r701;

cvt.rn.f32.s32 %f233, %r180;
fma.rn.ftz.f32 %f234, %f109, %f233, 0f00000000;

dp4a.s32.s32 %r184, %r641, %r186, %r701;

mul.lo.s32 %r777, %r184, %r716;
cvt.rn.f32.s32 %f235, %r777;
fma.rn.ftz.f32 %f236, %f110, %f235, %f232;

dp4a.s32.s32 %r188, %r645, %r186, %r701;

cvt.rn.f32.s32 %f237, %r188;
fma.rn.ftz.f32 %f238, %f110, %f237, %f234;

dp4a.s32.s32 %r192, %r649, %r194, %r701;

mul.lo.s32 %r778, %r192, %r725;
cvt.rn.f32.s32 %f239, %r778;
fma.rn.ftz.f32 %f240, %f111, %f239, %f236;

dp4a.s32.s32 %r196, %r653, %r194, %r701;

cvt.rn.f32.s32 %f241, %r196;
fma.rn.ftz.f32 %f242, %f111, %f241, %f238;

dp4a.s32.s32 %r200, %r657, %r202, %r701;

mul.lo.s32 %r779, %r200, %r734;
cvt.rn.f32.s32 %f243, %r779;
fma.rn.ftz.f32 %f244, %f112, %f243, %f240;

dp4a.s32.s32 %r204, %r661, %r202, %r701;

cvt.rn.f32.s32 %f245, %r204;
fma.rn.ftz.f32 %f246, %f112, %f245, %f242;

{.reg .f16 low,high;
mov.b32 {low,high},%r132;
cvt.f32.f16 %f113, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r132;
cvt.f32.f16 %f114, high;}


mul.ftz.f32 %f247, %f244, %f113;
mul.ftz.f32 %f248, %f246, %f114;
sub.ftz.f32 %f249, %f247, %f248;
add.ftz.f32 %f702, %f249, %f702;

{.reg .f16 low,high;
mov.b32 {low,high},%r172;
cvt.f32.f16 %f115, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r173;
cvt.f32.f16 %f116, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r174;
cvt.f32.f16 %f117, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r175;
cvt.f32.f16 %f118, low;}



dp4a.s32.s32 %r214, %r671, %r178, %r701;

mul.lo.s32 %r780, %r214, %r743;
cvt.rn.f32.s32 %f250, %r780;
fma.rn.ftz.f32 %f251, %f115, %f250, 0f00000000;

dp4a.s32.s32 %r218, %r675, %r178, %r701;

cvt.rn.f32.s32 %f252, %r218;
fma.rn.ftz.f32 %f253, %f115, %f252, 0f00000000;

dp4a.s32.s32 %r222, %r679, %r186, %r701;

mul.lo.s32 %r781, %r222, %r752;
cvt.rn.f32.s32 %f254, %r781;
fma.rn.ftz.f32 %f255, %f116, %f254, %f251;

dp4a.s32.s32 %r226, %r683, %r186, %r701;

cvt.rn.f32.s32 %f256, %r226;
fma.rn.ftz.f32 %f257, %f116, %f256, %f253;

dp4a.s32.s32 %r230, %r687, %r194, %r701;

mul.lo.s32 %r782, %r230, %r761;
cvt.rn.f32.s32 %f258, %r782;
fma.rn.ftz.f32 %f259, %f117, %f258, %f255;

dp4a.s32.s32 %r234, %r691, %r194, %r701;

cvt.rn.f32.s32 %f260, %r234;
fma.rn.ftz.f32 %f261, %f117, %f260, %f257;

dp4a.s32.s32 %r238, %r695, %r202, %r701;

mul.lo.s32 %r783, %r238, %r770;
cvt.rn.f32.s32 %f262, %r783;
fma.rn.ftz.f32 %f263, %f118, %f262, %f259;

dp4a.s32.s32 %r242, %r699, %r202, %r701;

cvt.rn.f32.s32 %f264, %r242;
fma.rn.ftz.f32 %f265, %f118, %f264, %f261;

{.reg .f16 low,high;
mov.b32 {low,high},%r170;
cvt.f32.f16 %f119, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r170;
cvt.f32.f16 %f120, high;}


mul.ftz.f32 %f266, %f263, %f119;
mul.ftz.f32 %f267, %f265, %f120;
sub.ftz.f32 %f268, %f266, %f267;
add.ftz.f32 %f701, %f268, %f701;
cvt.u64.u32 %rd48, %r1133;
add.s64 %rd49, %rd1, %rd48;
add.s64 %rd50, %rd49, %rd4;
mul.lo.s64 %rd51, %rd50, 36;
add.s64 %rd52, %rd11, %rd51;
add.s64 %rd53, %rd52, %rd35;
ld.global.nc.u32 %r254, [%rd53+4];
ld.global.nc.u32 %r248, [%rd52];

{.reg .f16 low,high;
mov.b32 {low,high},%r248;
cvt.f32.f16 %f121, low;}


ld.global.nc.u32 %r262, [%rd53+40];
ld.global.nc.u32 %r249, [%rd52+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r249;
cvt.f32.f16 %f122, low;}


ld.global.nc.u32 %r270, [%rd53+76];
ld.global.nc.u32 %r250, [%rd52+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r250;
cvt.f32.f16 %f123, low;}


ld.global.nc.u32 %r278, [%rd53+112];
ld.global.nc.u32 %r251, [%rd52+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r251;
cvt.f32.f16 %f124, low;}



dp4a.s32.s32 %r252, %r633, %r254, %r701;

mul.lo.s32 %r784, %r252, %r707;
cvt.rn.f32.s32 %f269, %r784;
fma.rn.ftz.f32 %f270, %f121, %f269, 0f00000000;

dp4a.s32.s32 %r256, %r637, %r254, %r701;

cvt.rn.f32.s32 %f271, %r256;
fma.rn.ftz.f32 %f272, %f121, %f271, 0f00000000;

dp4a.s32.s32 %r260, %r641, %r262, %r701;

mul.lo.s32 %r785, %r260, %r716;
cvt.rn.f32.s32 %f273, %r785;
fma.rn.ftz.f32 %f274, %f122, %f273, %f270;

dp4a.s32.s32 %r264, %r645, %r262, %r701;

cvt.rn.f32.s32 %f275, %r264;
fma.rn.ftz.f32 %f276, %f122, %f275, %f272;

dp4a.s32.s32 %r268, %r649, %r270, %r701;

mul.lo.s32 %r786, %r268, %r725;
cvt.rn.f32.s32 %f277, %r786;
fma.rn.ftz.f32 %f278, %f123, %f277, %f274;

dp4a.s32.s32 %r272, %r653, %r270, %r701;

cvt.rn.f32.s32 %f279, %r272;
fma.rn.ftz.f32 %f280, %f123, %f279, %f276;

dp4a.s32.s32 %r276, %r657, %r278, %r701;

mul.lo.s32 %r787, %r276, %r734;
cvt.rn.f32.s32 %f281, %r787;
fma.rn.ftz.f32 %f282, %f124, %f281, %f278;

dp4a.s32.s32 %r280, %r661, %r278, %r701;

cvt.rn.f32.s32 %f283, %r280;
fma.rn.ftz.f32 %f284, %f124, %f283, %f280;

{.reg .f16 low,high;
mov.b32 {low,high},%r132;
cvt.f32.f16 %f125, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r132;
cvt.f32.f16 %f126, high;}


mul.ftz.f32 %f285, %f282, %f125;
mul.ftz.f32 %f286, %f284, %f126;
sub.ftz.f32 %f287, %f285, %f286;
add.ftz.f32 %f700, %f287, %f700;

{.reg .f16 low,high;
mov.b32 {low,high},%r248;
cvt.f32.f16 %f127, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r249;
cvt.f32.f16 %f128, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r250;
cvt.f32.f16 %f129, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r251;
cvt.f32.f16 %f130, low;}



dp4a.s32.s32 %r290, %r671, %r254, %r701;

mul.lo.s32 %r788, %r290, %r743;
cvt.rn.f32.s32 %f288, %r788;
fma.rn.ftz.f32 %f289, %f127, %f288, 0f00000000;

dp4a.s32.s32 %r294, %r675, %r254, %r701;

cvt.rn.f32.s32 %f290, %r294;
fma.rn.ftz.f32 %f291, %f127, %f290, 0f00000000;

dp4a.s32.s32 %r298, %r679, %r262, %r701;

mul.lo.s32 %r789, %r298, %r752;
cvt.rn.f32.s32 %f292, %r789;
fma.rn.ftz.f32 %f293, %f128, %f292, %f289;

dp4a.s32.s32 %r302, %r683, %r262, %r701;

cvt.rn.f32.s32 %f294, %r302;
fma.rn.ftz.f32 %f295, %f128, %f294, %f291;

dp4a.s32.s32 %r306, %r687, %r270, %r701;

mul.lo.s32 %r790, %r306, %r761;
cvt.rn.f32.s32 %f296, %r790;
fma.rn.ftz.f32 %f297, %f129, %f296, %f293;

dp4a.s32.s32 %r310, %r691, %r270, %r701;

cvt.rn.f32.s32 %f298, %r310;
fma.rn.ftz.f32 %f299, %f129, %f298, %f295;

dp4a.s32.s32 %r314, %r695, %r278, %r701;

mul.lo.s32 %r791, %r314, %r770;
cvt.rn.f32.s32 %f300, %r791;
fma.rn.ftz.f32 %f301, %f130, %f300, %f297;

dp4a.s32.s32 %r318, %r699, %r278, %r701;

cvt.rn.f32.s32 %f302, %r318;
fma.rn.ftz.f32 %f303, %f130, %f302, %f299;

{.reg .f16 low,high;
mov.b32 {low,high},%r170;
cvt.f32.f16 %f131, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r170;
cvt.f32.f16 %f132, high;}


mul.ftz.f32 %f304, %f301, %f131;
mul.ftz.f32 %f305, %f303, %f132;
sub.ftz.f32 %f306, %f304, %f305;
add.ftz.f32 %f699, %f306, %f699;
cvt.u64.u32 %rd54, %r1132;
add.s64 %rd55, %rd1, %rd54;
add.s64 %rd56, %rd55, %rd4;
mul.lo.s64 %rd57, %rd56, 36;
add.s64 %rd58, %rd11, %rd57;
add.s64 %rd59, %rd58, %rd35;
ld.global.nc.u32 %r330, [%rd59+4];
ld.global.nc.u32 %r324, [%rd58];

{.reg .f16 low,high;
mov.b32 {low,high},%r324;
cvt.f32.f16 %f133, low;}


ld.global.nc.u32 %r338, [%rd59+40];
ld.global.nc.u32 %r325, [%rd58+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r325;
cvt.f32.f16 %f134, low;}


ld.global.nc.u32 %r346, [%rd59+76];
ld.global.nc.u32 %r326, [%rd58+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r326;
cvt.f32.f16 %f135, low;}


ld.global.nc.u32 %r354, [%rd59+112];
ld.global.nc.u32 %r327, [%rd58+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r327;
cvt.f32.f16 %f136, low;}



dp4a.s32.s32 %r328, %r633, %r330, %r701;

mul.lo.s32 %r792, %r328, %r707;
cvt.rn.f32.s32 %f307, %r792;
fma.rn.ftz.f32 %f308, %f133, %f307, 0f00000000;

dp4a.s32.s32 %r332, %r637, %r330, %r701;

cvt.rn.f32.s32 %f309, %r332;
fma.rn.ftz.f32 %f310, %f133, %f309, 0f00000000;

dp4a.s32.s32 %r336, %r641, %r338, %r701;

mul.lo.s32 %r793, %r336, %r716;
cvt.rn.f32.s32 %f311, %r793;
fma.rn.ftz.f32 %f312, %f134, %f311, %f308;

dp4a.s32.s32 %r340, %r645, %r338, %r701;

cvt.rn.f32.s32 %f313, %r340;
fma.rn.ftz.f32 %f314, %f134, %f313, %f310;

dp4a.s32.s32 %r344, %r649, %r346, %r701;

mul.lo.s32 %r794, %r344, %r725;
cvt.rn.f32.s32 %f315, %r794;
fma.rn.ftz.f32 %f316, %f135, %f315, %f312;

dp4a.s32.s32 %r348, %r653, %r346, %r701;

cvt.rn.f32.s32 %f317, %r348;
fma.rn.ftz.f32 %f318, %f135, %f317, %f314;

dp4a.s32.s32 %r352, %r657, %r354, %r701;

mul.lo.s32 %r795, %r352, %r734;
cvt.rn.f32.s32 %f319, %r795;
fma.rn.ftz.f32 %f320, %f136, %f319, %f316;

dp4a.s32.s32 %r356, %r661, %r354, %r701;

cvt.rn.f32.s32 %f321, %r356;
fma.rn.ftz.f32 %f322, %f136, %f321, %f318;

{.reg .f16 low,high;
mov.b32 {low,high},%r132;
cvt.f32.f16 %f137, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r132;
cvt.f32.f16 %f138, high;}


mul.ftz.f32 %f323, %f320, %f137;
mul.ftz.f32 %f324, %f322, %f138;
sub.ftz.f32 %f325, %f323, %f324;
add.ftz.f32 %f698, %f325, %f698;

{.reg .f16 low,high;
mov.b32 {low,high},%r324;
cvt.f32.f16 %f139, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r325;
cvt.f32.f16 %f140, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r326;
cvt.f32.f16 %f141, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r327;
cvt.f32.f16 %f142, low;}



dp4a.s32.s32 %r366, %r671, %r330, %r701;

mul.lo.s32 %r796, %r366, %r743;
cvt.rn.f32.s32 %f326, %r796;
fma.rn.ftz.f32 %f327, %f139, %f326, 0f00000000;

dp4a.s32.s32 %r370, %r675, %r330, %r701;

cvt.rn.f32.s32 %f328, %r370;
fma.rn.ftz.f32 %f329, %f139, %f328, 0f00000000;

dp4a.s32.s32 %r374, %r679, %r338, %r701;

mul.lo.s32 %r797, %r374, %r752;
cvt.rn.f32.s32 %f330, %r797;
fma.rn.ftz.f32 %f331, %f140, %f330, %f327;

dp4a.s32.s32 %r378, %r683, %r338, %r701;

cvt.rn.f32.s32 %f332, %r378;
fma.rn.ftz.f32 %f333, %f140, %f332, %f329;

dp4a.s32.s32 %r382, %r687, %r346, %r701;

mul.lo.s32 %r798, %r382, %r761;
cvt.rn.f32.s32 %f334, %r798;
fma.rn.ftz.f32 %f335, %f141, %f334, %f331;

dp4a.s32.s32 %r386, %r691, %r346, %r701;

cvt.rn.f32.s32 %f336, %r386;
fma.rn.ftz.f32 %f337, %f141, %f336, %f333;

dp4a.s32.s32 %r390, %r695, %r354, %r701;

mul.lo.s32 %r799, %r390, %r770;
cvt.rn.f32.s32 %f338, %r799;
fma.rn.ftz.f32 %f339, %f142, %f338, %f335;

dp4a.s32.s32 %r394, %r699, %r354, %r701;

cvt.rn.f32.s32 %f340, %r394;
fma.rn.ftz.f32 %f341, %f142, %f340, %f337;

{.reg .f16 low,high;
mov.b32 {low,high},%r170;
cvt.f32.f16 %f143, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r170;
cvt.f32.f16 %f144, high;}


mul.ftz.f32 %f342, %f339, %f143;
mul.ftz.f32 %f343, %f341, %f144;
sub.ftz.f32 %f344, %f342, %f343;
add.ftz.f32 %f697, %f344, %f697;
cvt.u64.u32 %rd60, %r1131;
add.s64 %rd61, %rd1, %rd60;
add.s64 %rd62, %rd61, %rd4;
mul.lo.s64 %rd63, %rd62, 36;
add.s64 %rd64, %rd11, %rd63;
add.s64 %rd65, %rd64, %rd35;
ld.global.nc.u32 %r406, [%rd65+4];
ld.global.nc.u32 %r400, [%rd64];

{.reg .f16 low,high;
mov.b32 {low,high},%r400;
cvt.f32.f16 %f145, low;}


ld.global.nc.u32 %r414, [%rd65+40];
ld.global.nc.u32 %r401, [%rd64+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r401;
cvt.f32.f16 %f146, low;}


ld.global.nc.u32 %r422, [%rd65+76];
ld.global.nc.u32 %r402, [%rd64+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r402;
cvt.f32.f16 %f147, low;}


ld.global.nc.u32 %r430, [%rd65+112];
ld.global.nc.u32 %r403, [%rd64+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r403;
cvt.f32.f16 %f148, low;}



dp4a.s32.s32 %r404, %r633, %r406, %r701;

mul.lo.s32 %r800, %r404, %r707;
cvt.rn.f32.s32 %f345, %r800;
fma.rn.ftz.f32 %f346, %f145, %f345, 0f00000000;

dp4a.s32.s32 %r408, %r637, %r406, %r701;

cvt.rn.f32.s32 %f347, %r408;
fma.rn.ftz.f32 %f348, %f145, %f347, 0f00000000;

dp4a.s32.s32 %r412, %r641, %r414, %r701;

mul.lo.s32 %r801, %r412, %r716;
cvt.rn.f32.s32 %f349, %r801;
fma.rn.ftz.f32 %f350, %f146, %f349, %f346;

dp4a.s32.s32 %r416, %r645, %r414, %r701;

cvt.rn.f32.s32 %f351, %r416;
fma.rn.ftz.f32 %f352, %f146, %f351, %f348;

dp4a.s32.s32 %r420, %r649, %r422, %r701;

mul.lo.s32 %r802, %r420, %r725;
cvt.rn.f32.s32 %f353, %r802;
fma.rn.ftz.f32 %f354, %f147, %f353, %f350;

dp4a.s32.s32 %r424, %r653, %r422, %r701;

cvt.rn.f32.s32 %f355, %r424;
fma.rn.ftz.f32 %f356, %f147, %f355, %f352;

dp4a.s32.s32 %r428, %r657, %r430, %r701;

mul.lo.s32 %r803, %r428, %r734;
cvt.rn.f32.s32 %f357, %r803;
fma.rn.ftz.f32 %f358, %f148, %f357, %f354;

dp4a.s32.s32 %r432, %r661, %r430, %r701;

cvt.rn.f32.s32 %f359, %r432;
fma.rn.ftz.f32 %f360, %f148, %f359, %f356;

{.reg .f16 low,high;
mov.b32 {low,high},%r132;
cvt.f32.f16 %f149, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r132;
cvt.f32.f16 %f150, high;}


mul.ftz.f32 %f361, %f358, %f149;
mul.ftz.f32 %f362, %f360, %f150;
sub.ftz.f32 %f363, %f361, %f362;
add.ftz.f32 %f696, %f363, %f696;

{.reg .f16 low,high;
mov.b32 {low,high},%r400;
cvt.f32.f16 %f151, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r401;
cvt.f32.f16 %f152, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r402;
cvt.f32.f16 %f153, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r403;
cvt.f32.f16 %f154, low;}



dp4a.s32.s32 %r442, %r671, %r406, %r701;

mul.lo.s32 %r804, %r442, %r743;
cvt.rn.f32.s32 %f364, %r804;
fma.rn.ftz.f32 %f365, %f151, %f364, 0f00000000;

dp4a.s32.s32 %r446, %r675, %r406, %r701;

cvt.rn.f32.s32 %f366, %r446;
fma.rn.ftz.f32 %f367, %f151, %f366, 0f00000000;

dp4a.s32.s32 %r450, %r679, %r414, %r701;

mul.lo.s32 %r805, %r450, %r752;
cvt.rn.f32.s32 %f368, %r805;
fma.rn.ftz.f32 %f369, %f152, %f368, %f365;

dp4a.s32.s32 %r454, %r683, %r414, %r701;

cvt.rn.f32.s32 %f370, %r454;
fma.rn.ftz.f32 %f371, %f152, %f370, %f367;

dp4a.s32.s32 %r458, %r687, %r422, %r701;

mul.lo.s32 %r806, %r458, %r761;
cvt.rn.f32.s32 %f372, %r806;
fma.rn.ftz.f32 %f373, %f153, %f372, %f369;

dp4a.s32.s32 %r462, %r691, %r422, %r701;

cvt.rn.f32.s32 %f374, %r462;
fma.rn.ftz.f32 %f375, %f153, %f374, %f371;

dp4a.s32.s32 %r466, %r695, %r430, %r701;

mul.lo.s32 %r807, %r466, %r770;
cvt.rn.f32.s32 %f376, %r807;
fma.rn.ftz.f32 %f377, %f154, %f376, %f373;

dp4a.s32.s32 %r470, %r699, %r430, %r701;

cvt.rn.f32.s32 %f378, %r470;
fma.rn.ftz.f32 %f379, %f154, %f378, %f375;

{.reg .f16 low,high;
mov.b32 {low,high},%r170;
cvt.f32.f16 %f155, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r170;
cvt.f32.f16 %f156, high;}


mul.ftz.f32 %f380, %f377, %f155;
mul.ftz.f32 %f381, %f379, %f156;
sub.ftz.f32 %f382, %f380, %f381;
add.ftz.f32 %f695, %f382, %f695;
cvt.u64.u32 %rd66, %r1130;
add.s64 %rd67, %rd1, %rd66;
add.s64 %rd68, %rd67, %rd4;
mul.lo.s64 %rd69, %rd68, 36;
add.s64 %rd70, %rd11, %rd69;
add.s64 %rd71, %rd70, %rd35;
ld.global.nc.u32 %r482, [%rd71+4];
ld.global.nc.u32 %r476, [%rd70];

{.reg .f16 low,high;
mov.b32 {low,high},%r476;
cvt.f32.f16 %f157, low;}


ld.global.nc.u32 %r490, [%rd71+40];
ld.global.nc.u32 %r477, [%rd70+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r477;
cvt.f32.f16 %f158, low;}


ld.global.nc.u32 %r498, [%rd71+76];
ld.global.nc.u32 %r478, [%rd70+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r478;
cvt.f32.f16 %f159, low;}


ld.global.nc.u32 %r506, [%rd71+112];
ld.global.nc.u32 %r479, [%rd70+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r479;
cvt.f32.f16 %f160, low;}



dp4a.s32.s32 %r480, %r633, %r482, %r701;

mul.lo.s32 %r808, %r480, %r707;
cvt.rn.f32.s32 %f383, %r808;
fma.rn.ftz.f32 %f384, %f157, %f383, 0f00000000;

dp4a.s32.s32 %r484, %r637, %r482, %r701;

cvt.rn.f32.s32 %f385, %r484;
fma.rn.ftz.f32 %f386, %f157, %f385, 0f00000000;

dp4a.s32.s32 %r488, %r641, %r490, %r701;

mul.lo.s32 %r809, %r488, %r716;
cvt.rn.f32.s32 %f387, %r809;
fma.rn.ftz.f32 %f388, %f158, %f387, %f384;

dp4a.s32.s32 %r492, %r645, %r490, %r701;

cvt.rn.f32.s32 %f389, %r492;
fma.rn.ftz.f32 %f390, %f158, %f389, %f386;

dp4a.s32.s32 %r496, %r649, %r498, %r701;

mul.lo.s32 %r810, %r496, %r725;
cvt.rn.f32.s32 %f391, %r810;
fma.rn.ftz.f32 %f392, %f159, %f391, %f388;

dp4a.s32.s32 %r500, %r653, %r498, %r701;

cvt.rn.f32.s32 %f393, %r500;
fma.rn.ftz.f32 %f394, %f159, %f393, %f390;

dp4a.s32.s32 %r504, %r657, %r506, %r701;

mul.lo.s32 %r811, %r504, %r734;
cvt.rn.f32.s32 %f395, %r811;
fma.rn.ftz.f32 %f396, %f160, %f395, %f392;

dp4a.s32.s32 %r508, %r661, %r506, %r701;

cvt.rn.f32.s32 %f397, %r508;
fma.rn.ftz.f32 %f398, %f160, %f397, %f394;

{.reg .f16 low,high;
mov.b32 {low,high},%r132;
cvt.f32.f16 %f161, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r132;
cvt.f32.f16 %f162, high;}


mul.ftz.f32 %f399, %f396, %f161;
mul.ftz.f32 %f400, %f398, %f162;
sub.ftz.f32 %f401, %f399, %f400;
add.ftz.f32 %f694, %f401, %f694;

{.reg .f16 low,high;
mov.b32 {low,high},%r476;
cvt.f32.f16 %f163, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r477;
cvt.f32.f16 %f164, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r478;
cvt.f32.f16 %f165, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r479;
cvt.f32.f16 %f166, low;}



dp4a.s32.s32 %r518, %r671, %r482, %r701;

mul.lo.s32 %r812, %r518, %r743;
cvt.rn.f32.s32 %f402, %r812;
fma.rn.ftz.f32 %f403, %f163, %f402, 0f00000000;

dp4a.s32.s32 %r522, %r675, %r482, %r701;

cvt.rn.f32.s32 %f404, %r522;
fma.rn.ftz.f32 %f405, %f163, %f404, 0f00000000;

dp4a.s32.s32 %r526, %r679, %r490, %r701;

mul.lo.s32 %r813, %r526, %r752;
cvt.rn.f32.s32 %f406, %r813;
fma.rn.ftz.f32 %f407, %f164, %f406, %f403;

dp4a.s32.s32 %r530, %r683, %r490, %r701;

cvt.rn.f32.s32 %f408, %r530;
fma.rn.ftz.f32 %f409, %f164, %f408, %f405;

dp4a.s32.s32 %r534, %r687, %r498, %r701;

mul.lo.s32 %r814, %r534, %r761;
cvt.rn.f32.s32 %f410, %r814;
fma.rn.ftz.f32 %f411, %f165, %f410, %f407;

dp4a.s32.s32 %r538, %r691, %r498, %r701;

cvt.rn.f32.s32 %f412, %r538;
fma.rn.ftz.f32 %f413, %f165, %f412, %f409;

dp4a.s32.s32 %r542, %r695, %r506, %r701;

mul.lo.s32 %r815, %r542, %r770;
cvt.rn.f32.s32 %f414, %r815;
fma.rn.ftz.f32 %f415, %f166, %f414, %f411;

dp4a.s32.s32 %r546, %r699, %r506, %r701;

cvt.rn.f32.s32 %f416, %r546;
fma.rn.ftz.f32 %f417, %f166, %f416, %f413;

{.reg .f16 low,high;
mov.b32 {low,high},%r170;
cvt.f32.f16 %f167, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r170;
cvt.f32.f16 %f168, high;}


mul.ftz.f32 %f418, %f415, %f167;
mul.ftz.f32 %f419, %f417, %f168;
sub.ftz.f32 %f420, %f418, %f419;
add.ftz.f32 %f693, %f420, %f693;
cvt.u64.u32 %rd72, %r1129;
add.s64 %rd73, %rd1, %rd72;
add.s64 %rd74, %rd73, %rd4;
mul.lo.s64 %rd75, %rd74, 36;
add.s64 %rd76, %rd11, %rd75;
add.s64 %rd77, %rd76, %rd35;
ld.global.nc.u32 %r558, [%rd77+4];
ld.global.nc.u32 %r552, [%rd76];

{.reg .f16 low,high;
mov.b32 {low,high},%r552;
cvt.f32.f16 %f169, low;}


ld.global.nc.u32 %r566, [%rd77+40];
ld.global.nc.u32 %r553, [%rd76+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r553;
cvt.f32.f16 %f170, low;}


ld.global.nc.u32 %r574, [%rd77+76];
ld.global.nc.u32 %r554, [%rd76+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r554;
cvt.f32.f16 %f171, low;}


ld.global.nc.u32 %r582, [%rd77+112];
ld.global.nc.u32 %r555, [%rd76+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r555;
cvt.f32.f16 %f172, low;}



dp4a.s32.s32 %r556, %r633, %r558, %r701;

mul.lo.s32 %r816, %r556, %r707;
cvt.rn.f32.s32 %f421, %r816;
fma.rn.ftz.f32 %f422, %f169, %f421, 0f00000000;

dp4a.s32.s32 %r560, %r637, %r558, %r701;

cvt.rn.f32.s32 %f423, %r560;
fma.rn.ftz.f32 %f424, %f169, %f423, 0f00000000;

dp4a.s32.s32 %r564, %r641, %r566, %r701;

mul.lo.s32 %r817, %r564, %r716;
cvt.rn.f32.s32 %f425, %r817;
fma.rn.ftz.f32 %f426, %f170, %f425, %f422;

dp4a.s32.s32 %r568, %r645, %r566, %r701;

cvt.rn.f32.s32 %f427, %r568;
fma.rn.ftz.f32 %f428, %f170, %f427, %f424;

dp4a.s32.s32 %r572, %r649, %r574, %r701;

mul.lo.s32 %r818, %r572, %r725;
cvt.rn.f32.s32 %f429, %r818;
fma.rn.ftz.f32 %f430, %f171, %f429, %f426;

dp4a.s32.s32 %r576, %r653, %r574, %r701;

cvt.rn.f32.s32 %f431, %r576;
fma.rn.ftz.f32 %f432, %f171, %f431, %f428;

dp4a.s32.s32 %r580, %r657, %r582, %r701;

mul.lo.s32 %r819, %r580, %r734;
cvt.rn.f32.s32 %f433, %r819;
fma.rn.ftz.f32 %f434, %f172, %f433, %f430;

dp4a.s32.s32 %r584, %r661, %r582, %r701;

cvt.rn.f32.s32 %f435, %r584;
fma.rn.ftz.f32 %f436, %f172, %f435, %f432;

{.reg .f16 low,high;
mov.b32 {low,high},%r132;
cvt.f32.f16 %f173, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r132;
cvt.f32.f16 %f174, high;}


mul.ftz.f32 %f437, %f434, %f173;
mul.ftz.f32 %f438, %f436, %f174;
sub.ftz.f32 %f439, %f437, %f438;
add.ftz.f32 %f692, %f439, %f692;

{.reg .f16 low,high;
mov.b32 {low,high},%r552;
cvt.f32.f16 %f175, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r553;
cvt.f32.f16 %f176, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r554;
cvt.f32.f16 %f177, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r555;
cvt.f32.f16 %f178, low;}



dp4a.s32.s32 %r594, %r671, %r558, %r701;

mul.lo.s32 %r820, %r594, %r743;
cvt.rn.f32.s32 %f440, %r820;
fma.rn.ftz.f32 %f441, %f175, %f440, 0f00000000;

dp4a.s32.s32 %r598, %r675, %r558, %r701;

cvt.rn.f32.s32 %f442, %r598;
fma.rn.ftz.f32 %f443, %f175, %f442, 0f00000000;

dp4a.s32.s32 %r602, %r679, %r566, %r701;

mul.lo.s32 %r821, %r602, %r752;
cvt.rn.f32.s32 %f444, %r821;
fma.rn.ftz.f32 %f445, %f176, %f444, %f441;

dp4a.s32.s32 %r606, %r683, %r566, %r701;

cvt.rn.f32.s32 %f446, %r606;
fma.rn.ftz.f32 %f447, %f176, %f446, %f443;

dp4a.s32.s32 %r610, %r687, %r574, %r701;

mul.lo.s32 %r822, %r610, %r761;
cvt.rn.f32.s32 %f448, %r822;
fma.rn.ftz.f32 %f449, %f177, %f448, %f445;

dp4a.s32.s32 %r614, %r691, %r574, %r701;

cvt.rn.f32.s32 %f450, %r614;
fma.rn.ftz.f32 %f451, %f177, %f450, %f447;

dp4a.s32.s32 %r618, %r695, %r582, %r701;

mul.lo.s32 %r823, %r618, %r770;
cvt.rn.f32.s32 %f452, %r823;
fma.rn.ftz.f32 %f453, %f178, %f452, %f449;

dp4a.s32.s32 %r622, %r699, %r582, %r701;

cvt.rn.f32.s32 %f454, %r622;
fma.rn.ftz.f32 %f455, %f178, %f454, %f451;

{.reg .f16 low,high;
mov.b32 {low,high},%r170;
cvt.f32.f16 %f179, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r170;
cvt.f32.f16 %f180, high;}


mul.ftz.f32 %f456, %f453, %f179;
mul.ftz.f32 %f457, %f455, %f180;
sub.ftz.f32 %f458, %f456, %f457;
add.ftz.f32 %f691, %f458, %f691;
cvt.u64.u32 %rd78, %r1128;
add.s64 %rd79, %rd1, %rd78;
add.s64 %rd80, %rd79, %rd4;
mul.lo.s64 %rd81, %rd80, 36;
add.s64 %rd82, %rd11, %rd81;
add.s64 %rd83, %rd82, %rd35;
ld.global.nc.u32 %r634, [%rd83+4];
ld.global.nc.u32 %r628, [%rd82];

{.reg .f16 low,high;
mov.b32 {low,high},%r628;
cvt.f32.f16 %f181, low;}


ld.global.nc.u32 %r642, [%rd83+40];
ld.global.nc.u32 %r629, [%rd82+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r629;
cvt.f32.f16 %f182, low;}


ld.global.nc.u32 %r650, [%rd83+76];
ld.global.nc.u32 %r630, [%rd82+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r630;
cvt.f32.f16 %f183, low;}


ld.global.nc.u32 %r658, [%rd83+112];
ld.global.nc.u32 %r631, [%rd82+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r631;
cvt.f32.f16 %f184, low;}



dp4a.s32.s32 %r632, %r633, %r634, %r701;

mul.lo.s32 %r824, %r632, %r707;
cvt.rn.f32.s32 %f459, %r824;
fma.rn.ftz.f32 %f460, %f181, %f459, 0f00000000;

dp4a.s32.s32 %r636, %r637, %r634, %r701;

cvt.rn.f32.s32 %f461, %r636;
fma.rn.ftz.f32 %f462, %f181, %f461, 0f00000000;

dp4a.s32.s32 %r640, %r641, %r642, %r701;

mul.lo.s32 %r825, %r640, %r716;
cvt.rn.f32.s32 %f463, %r825;
fma.rn.ftz.f32 %f464, %f182, %f463, %f460;

dp4a.s32.s32 %r644, %r645, %r642, %r701;

cvt.rn.f32.s32 %f465, %r644;
fma.rn.ftz.f32 %f466, %f182, %f465, %f462;

dp4a.s32.s32 %r648, %r649, %r650, %r701;

mul.lo.s32 %r826, %r648, %r725;
cvt.rn.f32.s32 %f467, %r826;
fma.rn.ftz.f32 %f468, %f183, %f467, %f464;

dp4a.s32.s32 %r652, %r653, %r650, %r701;

cvt.rn.f32.s32 %f469, %r652;
fma.rn.ftz.f32 %f470, %f183, %f469, %f466;

dp4a.s32.s32 %r656, %r657, %r658, %r701;

mul.lo.s32 %r827, %r656, %r734;
cvt.rn.f32.s32 %f471, %r827;
fma.rn.ftz.f32 %f472, %f184, %f471, %f468;

dp4a.s32.s32 %r660, %r661, %r658, %r701;

cvt.rn.f32.s32 %f473, %r660;
fma.rn.ftz.f32 %f474, %f184, %f473, %f470;

{.reg .f16 low,high;
mov.b32 {low,high},%r132;
cvt.f32.f16 %f185, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r132;
cvt.f32.f16 %f186, high;}


mul.ftz.f32 %f475, %f472, %f185;
mul.ftz.f32 %f476, %f474, %f186;
sub.ftz.f32 %f477, %f475, %f476;
add.ftz.f32 %f690, %f477, %f690;

{.reg .f16 low,high;
mov.b32 {low,high},%r628;
cvt.f32.f16 %f187, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r629;
cvt.f32.f16 %f188, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r630;
cvt.f32.f16 %f189, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r631;
cvt.f32.f16 %f190, low;}



dp4a.s32.s32 %r670, %r671, %r634, %r701;

mul.lo.s32 %r828, %r670, %r743;
cvt.rn.f32.s32 %f478, %r828;
fma.rn.ftz.f32 %f479, %f187, %f478, 0f00000000;

dp4a.s32.s32 %r674, %r675, %r634, %r701;

cvt.rn.f32.s32 %f480, %r674;
fma.rn.ftz.f32 %f481, %f187, %f480, 0f00000000;

dp4a.s32.s32 %r678, %r679, %r642, %r701;

mul.lo.s32 %r829, %r678, %r752;
cvt.rn.f32.s32 %f482, %r829;
fma.rn.ftz.f32 %f483, %f188, %f482, %f479;

dp4a.s32.s32 %r682, %r683, %r642, %r701;

cvt.rn.f32.s32 %f484, %r682;
fma.rn.ftz.f32 %f485, %f188, %f484, %f481;

dp4a.s32.s32 %r686, %r687, %r650, %r701;

mul.lo.s32 %r830, %r686, %r761;
cvt.rn.f32.s32 %f486, %r830;
fma.rn.ftz.f32 %f487, %f189, %f486, %f483;

dp4a.s32.s32 %r690, %r691, %r650, %r701;

cvt.rn.f32.s32 %f488, %r690;
fma.rn.ftz.f32 %f489, %f189, %f488, %f485;

dp4a.s32.s32 %r694, %r695, %r658, %r701;

mul.lo.s32 %r831, %r694, %r770;
cvt.rn.f32.s32 %f490, %r831;
fma.rn.ftz.f32 %f491, %f190, %f490, %f487;

dp4a.s32.s32 %r698, %r699, %r658, %r701;

cvt.rn.f32.s32 %f492, %r698;
fma.rn.ftz.f32 %f493, %f190, %f492, %f489;

{.reg .f16 low,high;
mov.b32 {low,high},%r170;
cvt.f32.f16 %f191, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r170;
cvt.f32.f16 %f192, high;}


mul.ftz.f32 %f494, %f491, %f191;
mul.ftz.f32 %f495, %f493, %f192;
sub.ftz.f32 %f496, %f494, %f495;
add.ftz.f32 %f689, %f496, %f689;
add.s32 %r1133, %r1133, 32;
add.s32 %r1132, %r1132, 32;
add.s32 %r1131, %r1131, 32;
add.s32 %r1130, %r1130, 32;
add.s32 %r1129, %r1129, 32;
add.s32 %r1128, %r1128, 32;
add.s32 %r1127, %r1127, 32;
add.s32 %r1126, %r1126, 32;
add.s64 %rd138, %rd138, 336;
add.s32 %r1134, %r1134, 4;
setp.lt.s32 %p4, %r1134, %r6;
@%p4 bra $L__BB139_2;

$L__BB139_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB139_5;

shl.b32 %r832, %r1, 11;
mov.u32 %r833, _ZZ13mul_mat_vec_qIL9ggml_type10ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r834, %r833, %r832;
shl.b32 %r835, %r2, 2;
add.s32 %r836, %r834, %r835;
st.shared.f32 [%r836+-2048], %f704;
st.shared.f32 [%r836+-1920], %f703;
st.shared.f32 [%r836+-1792], %f702;
st.shared.f32 [%r836+-1664], %f701;
st.shared.f32 [%r836+-1536], %f700;
st.shared.f32 [%r836+-1408], %f699;
st.shared.f32 [%r836+-1280], %f698;
st.shared.f32 [%r836+-1152], %f697;
st.shared.f32 [%r836+-1024], %f696;
st.shared.f32 [%r836+-896], %f695;
st.shared.f32 [%r836+-768], %f694;
st.shared.f32 [%r836+-640], %f693;
st.shared.f32 [%r836+-512], %f692;
st.shared.f32 [%r836+-384], %f691;
st.shared.f32 [%r836+-256], %f690;
st.shared.f32 [%r836+-128], %f689;

$L__BB139_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB139_38;

mov.u32 %r837, %ctaid.z;
mov.u32 %r838, %ctaid.y;
mad.lo.s32 %r839, %r838, %r45, %r4;
mad.lo.s32 %r840, %r837, %r51, %r839;
cvt.u64.u32 %rd14, %r840;
shl.b32 %r841, %r2, 2;
mov.u32 %r842, _ZZ13mul_mat_vec_qIL9ggml_type10ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r33, %r842, %r841;
mov.u32 %r843, 2;
ld.shared.f32 %f497, [%r33];
add.ftz.f32 %f498, %f497, %f704;
mov.b32 %r844, %f498;
mov.u32 %r845, 31;
mov.u32 %r846, 16;
mov.u32 %r847, -1;
shfl.sync.bfly.b32 %r848|%p7, %r844, %r846, %r845, %r847;
mov.b32 %f499, %r848;
add.ftz.f32 %f500, %f498, %f499;
mov.b32 %r849, %f500;
mov.u32 %r850, 8;
shfl.sync.bfly.b32 %r851|%p8, %r849, %r850, %r845, %r847;
mov.b32 %f501, %r851;
add.ftz.f32 %f502, %f500, %f501;
mov.b32 %r852, %f502;
mov.u32 %r853, 4;
shfl.sync.bfly.b32 %r854|%p9, %r852, %r853, %r845, %r847;
mov.b32 %f503, %r854;
add.ftz.f32 %f504, %f502, %f503;
mov.b32 %r855, %f504;
shfl.sync.bfly.b32 %r856|%p10, %r855, %r843, %r845, %r847;
mov.b32 %f505, %r856;
add.ftz.f32 %f506, %f504, %f505;
mov.b32 %r857, %f506;
mov.u32 %r858, 1;
shfl.sync.bfly.b32 %r859|%p11, %r857, %r858, %r845, %r847;
mov.b32 %f507, %r859;
add.ftz.f32 %f49, %f506, %f507;
setp.lt.u32 %p12, %r4, %r39;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
cvta.to.global.u64 %rd84, %rd21;
mul.wide.u32 %rd85, %r840, 4;
add.s64 %rd15, %rd84, %rd85;
not.pred %p14, %p1;
@%p14 bra $L__BB139_8;

st.global.f32 [%rd15], %f49;

$L__BB139_8:
ld.shared.f32 %f508, [%r33+128];
add.ftz.f32 %f509, %f508, %f703;
mov.b32 %r860, %f509;
shfl.sync.bfly.b32 %r864|%p15, %r860, %r846, %r845, %r847;
mov.b32 %f510, %r864;
add.ftz.f32 %f511, %f509, %f510;
mov.b32 %r865, %f511;
shfl.sync.bfly.b32 %r867|%p16, %r865, %r850, %r845, %r847;
mov.b32 %f512, %r867;
add.ftz.f32 %f513, %f511, %f512;
mov.b32 %r868, %f513;
shfl.sync.bfly.b32 %r870|%p17, %r868, %r853, %r845, %r847;
mov.b32 %f514, %r870;
add.ftz.f32 %f515, %f513, %f514;
mov.b32 %r871, %f515;
shfl.sync.bfly.b32 %r873|%p18, %r871, %r843, %r845, %r847;
mov.b32 %f516, %r873;
add.ftz.f32 %f517, %f515, %f516;
mov.b32 %r874, %f517;
shfl.sync.bfly.b32 %r876|%p19, %r874, %r858, %r845, %r847;
mov.b32 %f518, %r876;
add.ftz.f32 %f50, %f517, %f518;
setp.eq.s32 %p20, %r2, 1;
or.b32 %r880, %r4, 1;
setp.lt.u32 %p21, %r880, %r39;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB139_10;

st.global.f32 [%rd15+4], %f50;

$L__BB139_10:
ld.shared.f32 %f519, [%r33+256];
add.ftz.f32 %f520, %f519, %f702;
mov.b32 %r881, %f520;
mov.u32 %r882, 31;
mov.u32 %r883, 16;
mov.u32 %r884, -1;
shfl.sync.bfly.b32 %r885|%p23, %r881, %r883, %r882, %r884;
mov.b32 %f521, %r885;
add.ftz.f32 %f522, %f520, %f521;
mov.b32 %r886, %f522;
mov.u32 %r887, 8;
shfl.sync.bfly.b32 %r888|%p24, %r886, %r887, %r882, %r884;
mov.b32 %f523, %r888;
add.ftz.f32 %f524, %f522, %f523;
mov.b32 %r889, %f524;
mov.u32 %r890, 4;
shfl.sync.bfly.b32 %r891|%p25, %r889, %r890, %r882, %r884;
mov.b32 %f525, %r891;
add.ftz.f32 %f526, %f524, %f525;
mov.b32 %r892, %f526;
mov.u32 %r893, 2;
shfl.sync.bfly.b32 %r894|%p26, %r892, %r893, %r882, %r884;
mov.b32 %f527, %r894;
add.ftz.f32 %f528, %f526, %f527;
mov.b32 %r895, %f528;
mov.u32 %r896, 1;
shfl.sync.bfly.b32 %r897|%p27, %r895, %r896, %r882, %r884;
mov.b32 %f529, %r897;
add.ftz.f32 %f51, %f528, %f529;
@%p14 bra $L__BB139_12;

cvt.u64.u32 %rd86, %r39;
add.s64 %rd87, %rd86, %rd14;
shl.b64 %rd89, %rd87, 2;
add.s64 %rd90, %rd84, %rd89;
st.global.f32 [%rd90], %f51;

$L__BB139_12:
ld.shared.f32 %f530, [%r33+384];
add.ftz.f32 %f531, %f530, %f701;
mov.b32 %r898, %f531;
shfl.sync.bfly.b32 %r902|%p29, %r898, %r883, %r882, %r884;
mov.b32 %f532, %r902;
add.ftz.f32 %f533, %f531, %f532;
mov.b32 %r903, %f533;
shfl.sync.bfly.b32 %r905|%p30, %r903, %r887, %r882, %r884;
mov.b32 %f534, %r905;
add.ftz.f32 %f535, %f533, %f534;
mov.b32 %r906, %f535;
shfl.sync.bfly.b32 %r908|%p31, %r906, %r890, %r882, %r884;
mov.b32 %f536, %r908;
add.ftz.f32 %f537, %f535, %f536;
mov.b32 %r909, %f537;
shfl.sync.bfly.b32 %r911|%p32, %r909, %r893, %r882, %r884;
mov.b32 %f538, %r911;
add.ftz.f32 %f539, %f537, %f538;
mov.b32 %r912, %f539;
shfl.sync.bfly.b32 %r914|%p33, %r912, %r896, %r882, %r884;
mov.b32 %f540, %r914;
add.ftz.f32 %f52, %f539, %f540;
@%p22 bra $L__BB139_14;

add.s32 %r915, %r39, 1;
cvt.u64.u32 %rd91, %r915;
add.s64 %rd92, %rd91, %rd14;
shl.b64 %rd94, %rd92, 2;
add.s64 %rd95, %rd84, %rd94;
st.global.f32 [%rd95], %f52;

$L__BB139_14:
shl.b32 %r916, %r39, 1;
mov.u32 %r917, 1;
ld.shared.f32 %f541, [%r33+512];
add.ftz.f32 %f542, %f541, %f700;
mov.b32 %r918, %f542;
mov.u32 %r919, 31;
mov.u32 %r920, 16;
mov.u32 %r921, -1;
shfl.sync.bfly.b32 %r922|%p35, %r918, %r920, %r919, %r921;
mov.b32 %f543, %r922;
add.ftz.f32 %f544, %f542, %f543;
mov.b32 %r923, %f544;
mov.u32 %r924, 8;
shfl.sync.bfly.b32 %r925|%p36, %r923, %r924, %r919, %r921;
mov.b32 %f545, %r925;
add.ftz.f32 %f546, %f544, %f545;
mov.b32 %r926, %f546;
mov.u32 %r927, 4;
shfl.sync.bfly.b32 %r928|%p37, %r926, %r927, %r919, %r921;
mov.b32 %f547, %r928;
add.ftz.f32 %f548, %f546, %f547;
mov.b32 %r929, %f548;
mov.u32 %r930, 2;
shfl.sync.bfly.b32 %r931|%p38, %r929, %r930, %r919, %r921;
mov.b32 %f549, %r931;
add.ftz.f32 %f550, %f548, %f549;
mov.b32 %r932, %f550;
shfl.sync.bfly.b32 %r933|%p39, %r932, %r917, %r919, %r921;
mov.b32 %f551, %r933;
add.ftz.f32 %f53, %f550, %f551;
cvt.u64.u32 %rd96, %r916;
add.s64 %rd97, %rd96, %rd14;
shl.b64 %rd99, %rd97, 2;
add.s64 %rd16, %rd84, %rd99;
@%p14 bra $L__BB139_16;

st.global.f32 [%rd16], %f53;

$L__BB139_16:
ld.shared.f32 %f552, [%r33+640];
add.ftz.f32 %f553, %f552, %f699;
mov.b32 %r934, %f553;
shfl.sync.bfly.b32 %r938|%p41, %r934, %r920, %r919, %r921;
mov.b32 %f554, %r938;
add.ftz.f32 %f555, %f553, %f554;
mov.b32 %r939, %f555;
shfl.sync.bfly.b32 %r941|%p42, %r939, %r924, %r919, %r921;
mov.b32 %f556, %r941;
add.ftz.f32 %f557, %f555, %f556;
mov.b32 %r942, %f557;
shfl.sync.bfly.b32 %r944|%p43, %r942, %r927, %r919, %r921;
mov.b32 %f558, %r944;
add.ftz.f32 %f559, %f557, %f558;
mov.b32 %r945, %f559;
shfl.sync.bfly.b32 %r947|%p44, %r945, %r930, %r919, %r921;
mov.b32 %f560, %r947;
add.ftz.f32 %f561, %f559, %f560;
mov.b32 %r948, %f561;
shfl.sync.bfly.b32 %r950|%p45, %r948, %r917, %r919, %r921;
mov.b32 %f562, %r950;
add.ftz.f32 %f54, %f561, %f562;
@%p22 bra $L__BB139_18;

st.global.f32 [%rd16+4], %f54;

$L__BB139_18:
mul.lo.s32 %r34, %r39, 3;
ld.shared.f32 %f563, [%r33+768];
add.ftz.f32 %f564, %f563, %f698;
mov.b32 %r951, %f564;
mov.u32 %r952, 31;
mov.u32 %r953, 16;
mov.u32 %r954, -1;
shfl.sync.bfly.b32 %r955|%p47, %r951, %r953, %r952, %r954;
mov.b32 %f565, %r955;
add.ftz.f32 %f566, %f564, %f565;
mov.b32 %r956, %f566;
mov.u32 %r957, 8;
shfl.sync.bfly.b32 %r958|%p48, %r956, %r957, %r952, %r954;
mov.b32 %f567, %r958;
add.ftz.f32 %f568, %f566, %f567;
mov.b32 %r959, %f568;
mov.u32 %r960, 4;
shfl.sync.bfly.b32 %r961|%p49, %r959, %r960, %r952, %r954;
mov.b32 %f569, %r961;
add.ftz.f32 %f570, %f568, %f569;
mov.b32 %r962, %f570;
mov.u32 %r963, 2;
shfl.sync.bfly.b32 %r964|%p50, %r962, %r963, %r952, %r954;
mov.b32 %f571, %r964;
add.ftz.f32 %f572, %f570, %f571;
mov.b32 %r965, %f572;
mov.u32 %r966, 1;
shfl.sync.bfly.b32 %r967|%p51, %r965, %r966, %r952, %r954;
mov.b32 %f573, %r967;
add.ftz.f32 %f55, %f572, %f573;
@%p14 bra $L__BB139_20;

cvt.u64.u32 %rd100, %r34;
add.s64 %rd101, %rd100, %rd14;
shl.b64 %rd103, %rd101, 2;
add.s64 %rd104, %rd84, %rd103;
st.global.f32 [%rd104], %f55;

$L__BB139_20:
ld.shared.f32 %f574, [%r33+896];
add.ftz.f32 %f575, %f574, %f697;
mov.b32 %r968, %f575;
shfl.sync.bfly.b32 %r972|%p53, %r968, %r953, %r952, %r954;
mov.b32 %f576, %r972;
add.ftz.f32 %f577, %f575, %f576;
mov.b32 %r973, %f577;
shfl.sync.bfly.b32 %r975|%p54, %r973, %r957, %r952, %r954;
mov.b32 %f578, %r975;
add.ftz.f32 %f579, %f577, %f578;
mov.b32 %r976, %f579;
shfl.sync.bfly.b32 %r978|%p55, %r976, %r960, %r952, %r954;
mov.b32 %f580, %r978;
add.ftz.f32 %f581, %f579, %f580;
mov.b32 %r979, %f581;
shfl.sync.bfly.b32 %r981|%p56, %r979, %r963, %r952, %r954;
mov.b32 %f582, %r981;
add.ftz.f32 %f583, %f581, %f582;
mov.b32 %r982, %f583;
shfl.sync.bfly.b32 %r984|%p57, %r982, %r966, %r952, %r954;
mov.b32 %f584, %r984;
add.ftz.f32 %f56, %f583, %f584;
@%p22 bra $L__BB139_22;

add.s32 %r985, %r34, 1;
cvt.u64.u32 %rd105, %r985;
add.s64 %rd106, %rd105, %rd14;
shl.b64 %rd108, %rd106, 2;
add.s64 %rd109, %rd84, %rd108;
st.global.f32 [%rd109], %f56;

$L__BB139_22:
shl.b32 %r986, %r39, 2;
mov.u32 %r987, 2;
ld.shared.f32 %f585, [%r33+1024];
add.ftz.f32 %f586, %f585, %f696;
mov.b32 %r988, %f586;
mov.u32 %r989, 31;
mov.u32 %r990, 16;
mov.u32 %r991, -1;
shfl.sync.bfly.b32 %r992|%p59, %r988, %r990, %r989, %r991;
mov.b32 %f587, %r992;
add.ftz.f32 %f588, %f586, %f587;
mov.b32 %r993, %f588;
mov.u32 %r994, 8;
shfl.sync.bfly.b32 %r995|%p60, %r993, %r994, %r989, %r991;
mov.b32 %f589, %r995;
add.ftz.f32 %f590, %f588, %f589;
mov.b32 %r996, %f590;
mov.u32 %r997, 4;
shfl.sync.bfly.b32 %r998|%p61, %r996, %r997, %r989, %r991;
mov.b32 %f591, %r998;
add.ftz.f32 %f592, %f590, %f591;
mov.b32 %r999, %f592;
shfl.sync.bfly.b32 %r1000|%p62, %r999, %r987, %r989, %r991;
mov.b32 %f593, %r1000;
add.ftz.f32 %f594, %f592, %f593;
mov.b32 %r1001, %f594;
mov.u32 %r1002, 1;
shfl.sync.bfly.b32 %r1003|%p63, %r1001, %r1002, %r989, %r991;
mov.b32 %f595, %r1003;
add.ftz.f32 %f57, %f594, %f595;
cvt.u64.u32 %rd110, %r986;
add.s64 %rd111, %rd110, %rd14;
shl.b64 %rd113, %rd111, 2;
add.s64 %rd17, %rd84, %rd113;
@%p14 bra $L__BB139_24;

st.global.f32 [%rd17], %f57;

$L__BB139_24:
ld.shared.f32 %f596, [%r33+1152];
add.ftz.f32 %f597, %f596, %f695;
mov.b32 %r1004, %f597;
shfl.sync.bfly.b32 %r1008|%p65, %r1004, %r990, %r989, %r991;
mov.b32 %f598, %r1008;
add.ftz.f32 %f599, %f597, %f598;
mov.b32 %r1009, %f599;
shfl.sync.bfly.b32 %r1011|%p66, %r1009, %r994, %r989, %r991;
mov.b32 %f600, %r1011;
add.ftz.f32 %f601, %f599, %f600;
mov.b32 %r1012, %f601;
shfl.sync.bfly.b32 %r1014|%p67, %r1012, %r997, %r989, %r991;
mov.b32 %f602, %r1014;
add.ftz.f32 %f603, %f601, %f602;
mov.b32 %r1015, %f603;
shfl.sync.bfly.b32 %r1017|%p68, %r1015, %r987, %r989, %r991;
mov.b32 %f604, %r1017;
add.ftz.f32 %f605, %f603, %f604;
mov.b32 %r1018, %f605;
shfl.sync.bfly.b32 %r1020|%p69, %r1018, %r1002, %r989, %r991;
mov.b32 %f606, %r1020;
add.ftz.f32 %f58, %f605, %f606;
@%p22 bra $L__BB139_26;

st.global.f32 [%rd17+4], %f58;

$L__BB139_26:
mul.lo.s32 %r35, %r39, 5;
ld.shared.f32 %f607, [%r33+1280];
add.ftz.f32 %f608, %f607, %f694;
mov.b32 %r1021, %f608;
mov.u32 %r1022, 31;
mov.u32 %r1023, 16;
mov.u32 %r1024, -1;
shfl.sync.bfly.b32 %r1025|%p71, %r1021, %r1023, %r1022, %r1024;
mov.b32 %f609, %r1025;
add.ftz.f32 %f610, %f608, %f609;
mov.b32 %r1026, %f610;
mov.u32 %r1027, 8;
shfl.sync.bfly.b32 %r1028|%p72, %r1026, %r1027, %r1022, %r1024;
mov.b32 %f611, %r1028;
add.ftz.f32 %f612, %f610, %f611;
mov.b32 %r1029, %f612;
mov.u32 %r1030, 4;
shfl.sync.bfly.b32 %r1031|%p73, %r1029, %r1030, %r1022, %r1024;
mov.b32 %f613, %r1031;
add.ftz.f32 %f614, %f612, %f613;
mov.b32 %r1032, %f614;
mov.u32 %r1033, 2;
shfl.sync.bfly.b32 %r1034|%p74, %r1032, %r1033, %r1022, %r1024;
mov.b32 %f615, %r1034;
add.ftz.f32 %f616, %f614, %f615;
mov.b32 %r1035, %f616;
mov.u32 %r1036, 1;
shfl.sync.bfly.b32 %r1037|%p75, %r1035, %r1036, %r1022, %r1024;
mov.b32 %f617, %r1037;
add.ftz.f32 %f59, %f616, %f617;
@%p14 bra $L__BB139_28;

cvt.u64.u32 %rd114, %r35;
add.s64 %rd115, %rd114, %rd14;
shl.b64 %rd117, %rd115, 2;
add.s64 %rd118, %rd84, %rd117;
st.global.f32 [%rd118], %f59;

$L__BB139_28:
ld.shared.f32 %f618, [%r33+1408];
add.ftz.f32 %f619, %f618, %f693;
mov.b32 %r1038, %f619;
shfl.sync.bfly.b32 %r1042|%p77, %r1038, %r1023, %r1022, %r1024;
mov.b32 %f620, %r1042;
add.ftz.f32 %f621, %f619, %f620;
mov.b32 %r1043, %f621;
shfl.sync.bfly.b32 %r1045|%p78, %r1043, %r1027, %r1022, %r1024;
mov.b32 %f622, %r1045;
add.ftz.f32 %f623, %f621, %f622;
mov.b32 %r1046, %f623;
shfl.sync.bfly.b32 %r1048|%p79, %r1046, %r1030, %r1022, %r1024;
mov.b32 %f624, %r1048;
add.ftz.f32 %f625, %f623, %f624;
mov.b32 %r1049, %f625;
shfl.sync.bfly.b32 %r1051|%p80, %r1049, %r1033, %r1022, %r1024;
mov.b32 %f626, %r1051;
add.ftz.f32 %f627, %f625, %f626;
mov.b32 %r1052, %f627;
shfl.sync.bfly.b32 %r1054|%p81, %r1052, %r1036, %r1022, %r1024;
mov.b32 %f628, %r1054;
add.ftz.f32 %f60, %f627, %f628;
@%p22 bra $L__BB139_30;

add.s32 %r1055, %r35, 1;
cvt.u64.u32 %rd119, %r1055;
add.s64 %rd120, %rd119, %rd14;
shl.b64 %rd122, %rd120, 2;
add.s64 %rd123, %rd84, %rd122;
st.global.f32 [%rd123], %f60;

$L__BB139_30:
mul.lo.s32 %r1056, %r39, 6;
ld.shared.f32 %f629, [%r33+1536];
add.ftz.f32 %f630, %f629, %f692;
mov.b32 %r1057, %f630;
mov.u32 %r1058, 31;
mov.u32 %r1059, 16;
mov.u32 %r1060, -1;
shfl.sync.bfly.b32 %r1061|%p83, %r1057, %r1059, %r1058, %r1060;
mov.b32 %f631, %r1061;
add.ftz.f32 %f632, %f630, %f631;
mov.b32 %r1062, %f632;
mov.u32 %r1063, 8;
shfl.sync.bfly.b32 %r1064|%p84, %r1062, %r1063, %r1058, %r1060;
mov.b32 %f633, %r1064;
add.ftz.f32 %f634, %f632, %f633;
mov.b32 %r1065, %f634;
mov.u32 %r1066, 4;
shfl.sync.bfly.b32 %r1067|%p85, %r1065, %r1066, %r1058, %r1060;
mov.b32 %f635, %r1067;
add.ftz.f32 %f636, %f634, %f635;
mov.b32 %r1068, %f636;
mov.u32 %r1069, 2;
shfl.sync.bfly.b32 %r1070|%p86, %r1068, %r1069, %r1058, %r1060;
mov.b32 %f637, %r1070;
add.ftz.f32 %f638, %f636, %f637;
mov.b32 %r1071, %f638;
mov.u32 %r1072, 1;
shfl.sync.bfly.b32 %r1073|%p87, %r1071, %r1072, %r1058, %r1060;
mov.b32 %f639, %r1073;
add.ftz.f32 %f61, %f638, %f639;
cvt.u64.u32 %rd124, %r1056;
add.s64 %rd125, %rd124, %rd14;
shl.b64 %rd127, %rd125, 2;
add.s64 %rd18, %rd84, %rd127;
@%p14 bra $L__BB139_32;

st.global.f32 [%rd18], %f61;

$L__BB139_32:
ld.shared.f32 %f640, [%r33+1664];
add.ftz.f32 %f641, %f640, %f691;
mov.b32 %r1074, %f641;
shfl.sync.bfly.b32 %r1078|%p89, %r1074, %r1059, %r1058, %r1060;
mov.b32 %f642, %r1078;
add.ftz.f32 %f643, %f641, %f642;
mov.b32 %r1079, %f643;
shfl.sync.bfly.b32 %r1081|%p90, %r1079, %r1063, %r1058, %r1060;
mov.b32 %f644, %r1081;
add.ftz.f32 %f645, %f643, %f644;
mov.b32 %r1082, %f645;
shfl.sync.bfly.b32 %r1084|%p91, %r1082, %r1066, %r1058, %r1060;
mov.b32 %f646, %r1084;
add.ftz.f32 %f647, %f645, %f646;
mov.b32 %r1085, %f647;
shfl.sync.bfly.b32 %r1087|%p92, %r1085, %r1069, %r1058, %r1060;
mov.b32 %f648, %r1087;
add.ftz.f32 %f649, %f647, %f648;
mov.b32 %r1088, %f649;
shfl.sync.bfly.b32 %r1090|%p93, %r1088, %r1072, %r1058, %r1060;
mov.b32 %f650, %r1090;
add.ftz.f32 %f62, %f649, %f650;
@%p22 bra $L__BB139_34;

st.global.f32 [%rd18+4], %f62;

$L__BB139_34:
mul.lo.s32 %r36, %r39, 7;
ld.shared.f32 %f651, [%r33+1792];
add.ftz.f32 %f652, %f651, %f690;
mov.b32 %r1091, %f652;
mov.u32 %r1092, 31;
mov.u32 %r1093, 16;
mov.u32 %r1094, -1;
shfl.sync.bfly.b32 %r1095|%p95, %r1091, %r1093, %r1092, %r1094;
mov.b32 %f653, %r1095;
add.ftz.f32 %f654, %f652, %f653;
mov.b32 %r1096, %f654;
mov.u32 %r1097, 8;
shfl.sync.bfly.b32 %r1098|%p96, %r1096, %r1097, %r1092, %r1094;
mov.b32 %f655, %r1098;
add.ftz.f32 %f656, %f654, %f655;
mov.b32 %r1099, %f656;
mov.u32 %r1100, 4;
shfl.sync.bfly.b32 %r1101|%p97, %r1099, %r1100, %r1092, %r1094;
mov.b32 %f657, %r1101;
add.ftz.f32 %f658, %f656, %f657;
mov.b32 %r1102, %f658;
mov.u32 %r1103, 2;
shfl.sync.bfly.b32 %r1104|%p98, %r1102, %r1103, %r1092, %r1094;
mov.b32 %f659, %r1104;
add.ftz.f32 %f660, %f658, %f659;
mov.b32 %r1105, %f660;
mov.u32 %r1106, 1;
shfl.sync.bfly.b32 %r1107|%p99, %r1105, %r1106, %r1092, %r1094;
mov.b32 %f661, %r1107;
add.ftz.f32 %f63, %f660, %f661;
@%p14 bra $L__BB139_36;

cvt.u64.u32 %rd128, %r36;
add.s64 %rd129, %rd128, %rd14;
shl.b64 %rd131, %rd129, 2;
add.s64 %rd132, %rd84, %rd131;
st.global.f32 [%rd132], %f63;

$L__BB139_36:
ld.shared.f32 %f662, [%r33+1920];
add.ftz.f32 %f663, %f662, %f689;
mov.b32 %r1108, %f663;
shfl.sync.bfly.b32 %r1112|%p101, %r1108, %r1093, %r1092, %r1094;
mov.b32 %f664, %r1112;
add.ftz.f32 %f665, %f663, %f664;
mov.b32 %r1113, %f665;
shfl.sync.bfly.b32 %r1115|%p102, %r1113, %r1097, %r1092, %r1094;
mov.b32 %f666, %r1115;
add.ftz.f32 %f667, %f665, %f666;
mov.b32 %r1116, %f667;
shfl.sync.bfly.b32 %r1118|%p103, %r1116, %r1100, %r1092, %r1094;
mov.b32 %f668, %r1118;
add.ftz.f32 %f669, %f667, %f668;
mov.b32 %r1119, %f669;
shfl.sync.bfly.b32 %r1121|%p104, %r1119, %r1103, %r1092, %r1094;
mov.b32 %f670, %r1121;
add.ftz.f32 %f671, %f669, %f670;
mov.b32 %r1122, %f671;
shfl.sync.bfly.b32 %r1124|%p105, %r1122, %r1106, %r1092, %r1094;
mov.b32 %f672, %r1124;
add.ftz.f32 %f64, %f671, %f672;
@%p22 bra $L__BB139_38;

add.s32 %r1125, %r36, 1;
cvt.u64.u32 %rd133, %r1125;
add.s64 %rd134, %rd133, %rd14;
shl.b64 %rd136, %rd134, 2;
add.s64 %rd137, %rd84, %rd136;
st.global.f32 [%rd137], %f64;

$L__BB139_38:
ret;

}
.entry _Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj(
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2,
.param .align 8 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3[48],
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5,
.param .align 4 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6[12],
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15
)
.maxntid 256, 1, 1
.minnctapersm 1
{
.local .align 8 .b8 __local_depot140[16];
.reg .b64 %SP;
.reg .b64 %SPL;
.reg .pred %p<48>;
.reg .b16 %rs<37>;
.reg .f32 %f<205>;
.reg .b32 %r<753>;
.reg .b64 %rd<217>;


mov.u64 %SPL, __local_depot140;
ld.param.u64 %rd19, [_Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0];
ld.param.u64 %rd20, [_Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1];
ld.param.u64 %rd21, [_Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2];
ld.param.u32 %r36, [_Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5];
ld.param.u32 %r41, [_Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8];
ld.param.u32 %r42, [_Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9];
ld.param.u32 %r44, [_Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11];
ld.param.u32 %r45, [_Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12];
ld.param.u32 %r48, [_Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14];
ld.param.u32 %r47, [_Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15];
ld.param.u32 %r39, [_Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+8];
ld.param.u32 %r38, [_Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+4];
ld.param.u32 %r37, [_Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6];
ld.param.u64 %rd23, [_Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+8];
add.u64 %rd1, %SPL, 0;
add.u64 %rd2, %SPL, 8;
mov.u32 %r1, %tid.y;
setp.ge.u32 %p1, %r1, %r48;
@%p1 bra $L__BB140_29;

mov.u32 %r2, %ctaid.y;
cvta.to.global.u64 %rd30, %rd21;
mad.lo.s32 %r49, %r1, %r47, %r2;
mul.wide.u32 %rd31, %r49, 4;
add.s64 %rd32, %rd30, %rd31;
ld.global.nc.u32 %r3, [%rd32];
mov.f32 %f198, 0f00000000;
st.local.v2.f32 [%rd1], {%f198, %f198};
st.local.v2.f32 [%rd2], {%f198, %f198};
mov.u32 %r4, %tid.x;
shr.u32 %r5, %r4, 4;
shr.u32 %r50, %r36, 8;
setp.ge.u32 %p2, %r5, %r50;
mov.f32 %f199, %f198;
mov.f32 %f200, %f198;
mov.f32 %f201, %f198;
@%p2 bra $L__BB140_9;

cvta.to.global.u64 %rd33, %rd20;
and.b32 %r51, %r4, 15;
and.b32 %r52, %r4, 7;
sub.s32 %r53, %r51, %r52;
and.b32 %r54, %r4, 4;
shr.u32 %r55, %r54, 2;
add.s32 %r56, %r53, %r55;
shr.s32 %r57, %r56, 31;
shr.u32 %r58, %r57, 29;
add.s32 %r59, %r56, %r58;
and.b32 %r60, %r59, -8;
sub.s32 %r6, %r56, %r60;
shr.s32 %r61, %r59, 3;
shl.b32 %r7, %r61, 2;
shr.u32 %r62, %r57, 30;
add.s32 %r63, %r56, %r62;
and.b32 %r64, %r63, -4;
sub.s32 %r65, %r56, %r64;
shr.s32 %r66, %r63, 2;
shl.b32 %r8, %r66, 1;
add.s32 %r9, %r65, 8;
add.s32 %r67, %r56, 2;
shr.s32 %r68, %r67, 31;
shr.u32 %r69, %r68, 29;
add.s32 %r70, %r67, %r69;
and.b32 %r71, %r70, -8;
sub.s32 %r10, %r67, %r71;
shr.s32 %r72, %r70, 3;
shl.b32 %r11, %r72, 2;
shr.u32 %r73, %r68, 30;
add.s32 %r74, %r67, %r73;
and.b32 %r75, %r74, -4;
sub.s32 %r76, %r67, %r75;
shr.s32 %r77, %r74, 2;
shl.b32 %r12, %r77, 1;
add.s32 %r13, %r76, 8;
add.s32 %r78, %r56, 4;
shr.s32 %r79, %r78, 31;
shr.u32 %r80, %r79, 29;
add.s32 %r81, %r78, %r80;
and.b32 %r82, %r81, -8;
sub.s32 %r14, %r78, %r82;
shr.s32 %r83, %r81, 3;
shl.b32 %r15, %r83, 2;
shr.u32 %r84, %r79, 30;
add.s32 %r85, %r78, %r84;
and.b32 %r86, %r85, -4;
sub.s32 %r87, %r78, %r86;
shr.s32 %r88, %r85, 2;
shl.b32 %r16, %r88, 1;
add.s32 %r17, %r87, 8;
add.s32 %r89, %r56, 6;
shr.s32 %r90, %r89, 31;
shr.u32 %r91, %r90, 29;
add.s32 %r92, %r89, %r91;
and.b32 %r93, %r92, -8;
sub.s32 %r18, %r89, %r93;
shr.s32 %r94, %r92, 3;
shl.b32 %r19, %r94, 2;
shr.u32 %r95, %r90, 30;
add.s32 %r96, %r89, %r95;
and.b32 %r97, %r96, -4;
sub.s32 %r98, %r89, %r97;
shr.s32 %r99, %r96, 2;
shl.b32 %r20, %r99, 1;
add.s32 %r21, %r98, 8;
cvt.u64.u32 %rd34, %r4;
and.b64 %rd35, %rd34, 8;
mul.wide.u32 %rd36, %r4, 4;
and.b64 %rd37, %rd36, 28;
shr.u64 %rd38, %rd34, 4;
mul.lo.s64 %rd39, %rd38, 288;
shr.u64 %rd40, %rd35, 3;
mul.lo.s64 %rd41, %rd40, 144;
add.s64 %rd42, %rd39, %rd41;
mul.lo.s32 %r100, %r1, %r42;
mul.wide.u32 %rd43, %r100, 36;
add.s64 %rd44, %rd42, %rd43;
mul.hi.u32 %r101, %r2, %r37;
add.s32 %r102, %r101, %r2;
shr.u32 %r103, %r102, %r38;
mul.lo.s32 %r104, %r103, %r39;
sub.s32 %r105, %r2, %r104;
mul.lo.s32 %r106, %r105, %r45;
mul.wide.u32 %rd45, %r106, 36;
add.s64 %rd46, %rd44, %rd45;
add.s64 %rd47, %rd46, %rd37;
add.s64 %rd48, %rd33, %rd47;
add.s64 %rd215, %rd48, 76;
add.s64 %rd214, %rd33, %rd46;
cvta.to.global.u64 %rd216, %rd19;
cvta.to.global.u64 %rd213, %rd23;
mov.f32 %f198, 0f00000000;
mov.u32 %r752, %r5;

$L__BB140_3:
mov.u32 %r145, %ctaid.x;
shl.b32 %r146, %r145, 1;
mad.lo.s32 %r147, %r146, %r41, %r5;
mad.lo.s32 %r148, %r3, %r44, %r147;
mul.wide.s32 %rd49, %r148, 110;
add.s64 %rd50, %rd216, %rd49;
ld.global.nc.u16 %rs1, [%rd50+108];

{ cvt.f32.f16 %f42, %rs1;}


mul.wide.u32 %rd51, %r51, 4;
add.s64 %rd52, %rd51, %rd49;
add.s64 %rd53, %rd216, %rd52;
shl.b32 %r150, %r4, 2;
cvt.u64.u32 %rd54, %r150;
and.b64 %rd55, %rd54, 28;
add.s64 %rd56, %rd55, %rd49;
add.s64 %rd57, %rd216, %rd56;
ld.global.nc.u16 %rs2, [%rd57];
ld.global.nc.u16 %rs3, [%rd57+2];
mov.b32 %r151, {%rs2, %rs3};
not.b32 %r152, %r151;
ld.global.nc.u16 %rs4, [%rd53+34];
ld.global.nc.u16 %rs5, [%rd53+32];
ld.global.nc.u32 %r23, [%rd215+-72];
ld.global.nc.u32 %r24, [%rd214];

{.reg .f16 low,high;
mov.b32 {low,high},%r24;
cvt.f32.f16 %f43, low;}


ld.global.nc.u32 %r25, [%rd215+-36];
ld.global.nc.u32 %r26, [%rd214+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r26;
cvt.f32.f16 %f44, low;}


ld.global.nc.u32 %r27, [%rd215];
ld.global.nc.u32 %r28, [%rd214+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r28;
cvt.f32.f16 %f45, low;}


ld.global.nc.u32 %r29, [%rd215+36];
ld.global.nc.u32 %r30, [%rd214+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r30;
cvt.f32.f16 %f46, low;}


mov.b32 %r153, {%rs5, %rs4};
and.b32 %r154, %r4, 8;
shr.u32 %r155, %r154, 1;
shr.s32 %r156, %r152, %r155;
cvt.s64.s32 %rd58, %r6;
add.s64 %rd59, %rd49, %rd58;
add.s64 %rd60, %rd216, %rd59;
ld.global.nc.u8 %rs6, [%rd60+96];
cvt.u32.u16 %r157, %rs6;
and.b32 %r158, %r157, 255;
shr.u32 %r159, %r158, %r7;
and.b32 %r160, %r159, 15;
cvt.s64.s32 %rd61, %r9;
add.s64 %rd62, %rd49, %rd61;
add.s64 %rd63, %rd216, %rd62;
ld.global.nc.u8 %rs7, [%rd63+96];
cvt.u32.u16 %r161, %rs7;
and.b32 %r162, %r161, 255;
shr.u32 %r163, %r162, %r8;
and.b32 %r164, %r163, 3;
bfi.b32 %r165, %r164, %r160, 4, 2;
add.s32 %r166, %r165, -32;
and.b32 %r113, %r153, 50529027;
shl.b32 %r167, %r156, 2;
and.b32 %r168, %r167, 67372036;
or.b32 %r169, %r113, -2139062144;
sub.s32 %r170, %r169, %r168;
xor.b32 %r171, %r170, -2139062144;
or.b32 %r172, %r113, %r168;
xor.b32 %r173, %r170, %r113;
and.b32 %r111, %r173, %r172;

prmt.b32 %r111, %r111, 0, 0xba98;


prmt.b32 %r113, %r113, 0, 0xba98;

xor.b32 %r174, %r113, 2139062143;
not.b32 %r175, %r111;
and.b32 %r176, %r171, %r175;
and.b32 %r177, %r174, %r111;
or.b32 %r116, %r177, %r176;
mov.u32 %r142, 0;

dp4a.s32.s32 %r115, %r116, %r23, %r142;

mul.lo.s32 %r178, %r166, %r115;
cvt.rn.f32.s32 %f47, %r178;
fma.rn.ftz.f32 %f48, %f43, %f47, 0f00000000;
cvt.s64.s32 %rd64, %r10;
add.s64 %rd65, %rd49, %rd64;
add.s64 %rd66, %rd216, %rd65;
ld.global.nc.u8 %rs8, [%rd66+96];
cvt.u32.u16 %r179, %rs8;
and.b32 %r180, %r179, 255;
shr.u32 %r181, %r180, %r11;
and.b32 %r182, %r181, 15;
cvt.s64.s32 %rd67, %r13;
add.s64 %rd68, %rd49, %rd67;
add.s64 %rd69, %rd216, %rd68;
ld.global.nc.u8 %rs9, [%rd69+96];
cvt.u32.u16 %r183, %rs9;
and.b32 %r184, %r183, 255;
shr.u32 %r185, %r184, %r12;
and.b32 %r186, %r185, 3;
bfi.b32 %r187, %r186, %r182, 4, 2;
add.s32 %r188, %r187, -32;
shr.u32 %r189, %r153, 2;
and.b32 %r121, %r189, 50529027;
shl.b32 %r190, %r156, 1;
and.b32 %r191, %r190, 67372036;
or.b32 %r192, %r121, -2139062144;
sub.s32 %r193, %r192, %r191;
xor.b32 %r194, %r193, -2139062144;
or.b32 %r195, %r121, %r191;
xor.b32 %r196, %r193, %r121;
and.b32 %r119, %r196, %r195;

prmt.b32 %r119, %r119, 0, 0xba98;


prmt.b32 %r121, %r121, 0, 0xba98;

xor.b32 %r197, %r121, 2139062143;
not.b32 %r198, %r119;
and.b32 %r199, %r194, %r198;
and.b32 %r200, %r197, %r119;
or.b32 %r124, %r200, %r199;

dp4a.s32.s32 %r123, %r124, %r25, %r142;

mul.lo.s32 %r201, %r188, %r123;
cvt.rn.f32.s32 %f49, %r201;
fma.rn.ftz.f32 %f50, %f44, %f49, %f48;
cvt.s64.s32 %rd70, %r14;
add.s64 %rd71, %rd49, %rd70;
add.s64 %rd72, %rd216, %rd71;
ld.global.nc.u8 %rs10, [%rd72+96];
cvt.u32.u16 %r202, %rs10;
and.b32 %r203, %r202, 255;
shr.u32 %r204, %r203, %r15;
and.b32 %r205, %r204, 15;
cvt.s64.s32 %rd73, %r17;
add.s64 %rd74, %rd49, %rd73;
add.s64 %rd75, %rd216, %rd74;
ld.global.nc.u8 %rs11, [%rd75+96];
cvt.u32.u16 %r206, %rs11;
and.b32 %r207, %r206, 255;
shr.u32 %r208, %r207, %r16;
and.b32 %r209, %r208, 3;
bfi.b32 %r210, %r209, %r205, 4, 2;
add.s32 %r211, %r210, -32;
shr.u32 %r212, %r153, 4;
and.b32 %r129, %r212, 50529027;
and.b32 %r213, %r156, 67372036;
or.b32 %r214, %r129, -2139062144;
sub.s32 %r215, %r214, %r213;
xor.b32 %r216, %r215, -2139062144;
or.b32 %r217, %r129, %r213;
xor.b32 %r218, %r215, %r129;
and.b32 %r127, %r218, %r217;

prmt.b32 %r127, %r127, 0, 0xba98;


prmt.b32 %r129, %r129, 0, 0xba98;

xor.b32 %r219, %r129, 2139062143;
not.b32 %r220, %r127;
and.b32 %r221, %r216, %r220;
and.b32 %r222, %r219, %r127;
or.b32 %r132, %r222, %r221;

dp4a.s32.s32 %r131, %r132, %r27, %r142;

mul.lo.s32 %r223, %r211, %r131;
cvt.rn.f32.s32 %f51, %r223;
fma.rn.ftz.f32 %f52, %f45, %f51, %f50;
cvt.s64.s32 %rd76, %r18;
add.s64 %rd77, %rd49, %rd76;
add.s64 %rd78, %rd216, %rd77;
ld.global.nc.u8 %rs12, [%rd78+96];
cvt.u32.u16 %r224, %rs12;
and.b32 %r225, %r224, 255;
shr.u32 %r226, %r225, %r19;
and.b32 %r227, %r226, 15;
cvt.s64.s32 %rd79, %r21;
add.s64 %rd80, %rd49, %rd79;
add.s64 %rd81, %rd216, %rd80;
ld.global.nc.u8 %rs13, [%rd81+96];
cvt.u32.u16 %r228, %rs13;
and.b32 %r229, %r228, 255;
shr.u32 %r230, %r229, %r20;
and.b32 %r231, %r230, 3;
bfi.b32 %r232, %r231, %r227, 4, 2;
add.s32 %r233, %r232, -32;
shr.u32 %r234, %r153, 6;
and.b32 %r137, %r234, 50529027;
shr.u32 %r235, %r156, 1;
and.b32 %r236, %r235, 67372036;
or.b32 %r237, %r137, -2139062144;
sub.s32 %r238, %r237, %r236;
xor.b32 %r239, %r238, -2139062144;
or.b32 %r240, %r137, %r236;
xor.b32 %r241, %r238, %r137;
and.b32 %r135, %r241, %r240;

prmt.b32 %r135, %r135, 0, 0xba98;


prmt.b32 %r137, %r137, 0, 0xba98;

xor.b32 %r242, %r137, 2139062143;
not.b32 %r243, %r135;
and.b32 %r244, %r239, %r243;
and.b32 %r245, %r242, %r135;
or.b32 %r140, %r245, %r244;

dp4a.s32.s32 %r139, %r140, %r29, %r142;

mul.lo.s32 %r246, %r233, %r139;
cvt.rn.f32.s32 %f53, %r246;
fma.rn.ftz.f32 %f54, %f46, %f53, %f52;
fma.rn.ftz.f32 %f201, %f42, %f54, %f201;
setp.eq.s64 %p3, %rd23, 0;
@%p3 bra $L__BB140_5;

and.b32 %r751, %r4, 8;
shr.u32 %r750, %r751, 1;
shr.u32 %r284, %r4, 4;
mad.lo.s32 %r287, %r146, %r41, %r284;
mad.lo.s32 %r288, %r3, %r44, %r287;
mul.wide.s32 %rd82, %r288, 110;
add.s64 %rd83, %rd213, %rd82;
ld.global.u16 %rs14, [%rd83+108];

{ cvt.f32.f16 %f55, %rs14;}


add.s64 %rd85, %rd51, %rd82;
add.s64 %rd86, %rd213, %rd85;
add.s64 %rd89, %rd55, %rd82;
add.s64 %rd90, %rd213, %rd89;
ld.global.u16 %rs15, [%rd90];
ld.global.u16 %rs16, [%rd90+2];
mov.b32 %r291, {%rs15, %rs16};
not.b32 %r292, %r291;
ld.global.u16 %rs17, [%rd86+34];
ld.global.u16 %rs18, [%rd86+32];

{.reg .f16 low,high;
mov.b32 {low,high},%r24;
cvt.f32.f16 %f56, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r26;
cvt.f32.f16 %f57, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r28;
cvt.f32.f16 %f58, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r30;
cvt.f32.f16 %f59, low;}


mov.b32 %r293, {%rs18, %rs17};
shr.s32 %r296, %r292, %r750;
add.s64 %rd92, %rd82, %rd58;
add.s64 %rd93, %rd213, %rd92;
ld.global.u8 %r297, [%rd93+96];
shr.u32 %r298, %r297, %r7;
and.b32 %r299, %r298, 15;
add.s64 %rd95, %rd82, %rd61;
add.s64 %rd96, %rd213, %rd95;
ld.global.u8 %r300, [%rd96+96];
shr.u32 %r301, %r300, %r8;
and.b32 %r302, %r301, 3;
bfi.b32 %r303, %r302, %r299, 4, 2;
add.s32 %r304, %r303, -32;
and.b32 %r253, %r293, 50529027;
shl.b32 %r305, %r296, 2;
and.b32 %r306, %r305, 67372036;
or.b32 %r307, %r253, -2139062144;
sub.s32 %r308, %r307, %r306;
xor.b32 %r309, %r308, -2139062144;
or.b32 %r310, %r253, %r306;
xor.b32 %r311, %r308, %r253;
and.b32 %r251, %r311, %r310;

prmt.b32 %r251, %r251, 0, 0xba98;


prmt.b32 %r253, %r253, 0, 0xba98;

xor.b32 %r312, %r253, 2139062143;
not.b32 %r313, %r251;
and.b32 %r314, %r309, %r313;
and.b32 %r315, %r312, %r251;
or.b32 %r256, %r315, %r314;

dp4a.s32.s32 %r255, %r256, %r23, %r142;

mul.lo.s32 %r316, %r304, %r255;
cvt.rn.f32.s32 %f60, %r316;
fma.rn.ftz.f32 %f61, %f56, %f60, 0f00000000;
add.s64 %rd98, %rd82, %rd64;
add.s64 %rd99, %rd213, %rd98;
ld.global.u8 %r317, [%rd99+96];
shr.u32 %r318, %r317, %r11;
and.b32 %r319, %r318, 15;
add.s64 %rd101, %rd82, %rd67;
add.s64 %rd102, %rd213, %rd101;
ld.global.u8 %r320, [%rd102+96];
shr.u32 %r321, %r320, %r12;
and.b32 %r322, %r321, 3;
bfi.b32 %r323, %r322, %r319, 4, 2;
add.s32 %r324, %r323, -32;
shr.u32 %r325, %r293, 2;
and.b32 %r261, %r325, 50529027;
shl.b32 %r326, %r296, 1;
and.b32 %r327, %r326, 67372036;
or.b32 %r328, %r261, -2139062144;
sub.s32 %r329, %r328, %r327;
xor.b32 %r330, %r329, -2139062144;
or.b32 %r331, %r261, %r327;
xor.b32 %r332, %r329, %r261;
and.b32 %r259, %r332, %r331;

prmt.b32 %r259, %r259, 0, 0xba98;


prmt.b32 %r261, %r261, 0, 0xba98;

xor.b32 %r333, %r261, 2139062143;
not.b32 %r334, %r259;
and.b32 %r335, %r330, %r334;
and.b32 %r336, %r333, %r259;
or.b32 %r264, %r336, %r335;

dp4a.s32.s32 %r263, %r264, %r25, %r142;

mul.lo.s32 %r337, %r324, %r263;
cvt.rn.f32.s32 %f62, %r337;
fma.rn.ftz.f32 %f63, %f57, %f62, %f61;
add.s64 %rd104, %rd82, %rd70;
add.s64 %rd105, %rd213, %rd104;
ld.global.u8 %r338, [%rd105+96];
shr.u32 %r339, %r338, %r15;
and.b32 %r340, %r339, 15;
add.s64 %rd107, %rd82, %rd73;
add.s64 %rd108, %rd213, %rd107;
ld.global.u8 %r341, [%rd108+96];
shr.u32 %r342, %r341, %r16;
and.b32 %r343, %r342, 3;
bfi.b32 %r344, %r343, %r340, 4, 2;
add.s32 %r345, %r344, -32;
shr.u32 %r346, %r293, 4;
and.b32 %r269, %r346, 50529027;
and.b32 %r347, %r296, 67372036;
or.b32 %r348, %r269, -2139062144;
sub.s32 %r349, %r348, %r347;
xor.b32 %r350, %r349, -2139062144;
or.b32 %r351, %r269, %r347;
xor.b32 %r352, %r349, %r269;
and.b32 %r267, %r352, %r351;

prmt.b32 %r267, %r267, 0, 0xba98;


prmt.b32 %r269, %r269, 0, 0xba98;

xor.b32 %r353, %r269, 2139062143;
not.b32 %r354, %r267;
and.b32 %r355, %r350, %r354;
and.b32 %r356, %r353, %r267;
or.b32 %r272, %r356, %r355;

dp4a.s32.s32 %r271, %r272, %r27, %r142;

mul.lo.s32 %r357, %r345, %r271;
cvt.rn.f32.s32 %f64, %r357;
fma.rn.ftz.f32 %f65, %f58, %f64, %f63;
add.s64 %rd110, %rd82, %rd76;
add.s64 %rd111, %rd213, %rd110;
ld.global.u8 %r358, [%rd111+96];
shr.u32 %r359, %r358, %r19;
and.b32 %r360, %r359, 15;
add.s64 %rd113, %rd82, %rd79;
add.s64 %rd114, %rd213, %rd113;
ld.global.u8 %r361, [%rd114+96];
shr.u32 %r362, %r361, %r20;
and.b32 %r363, %r362, 3;
bfi.b32 %r364, %r363, %r360, 4, 2;
add.s32 %r365, %r364, -32;
shr.u32 %r366, %r293, 6;
and.b32 %r277, %r366, 50529027;
shr.u32 %r367, %r296, 1;
and.b32 %r368, %r367, 67372036;
or.b32 %r369, %r277, -2139062144;
sub.s32 %r370, %r369, %r368;
xor.b32 %r371, %r370, -2139062144;
or.b32 %r372, %r277, %r368;
xor.b32 %r373, %r370, %r277;
and.b32 %r275, %r373, %r372;

prmt.b32 %r275, %r275, 0, 0xba98;


prmt.b32 %r277, %r277, 0, 0xba98;

xor.b32 %r374, %r277, 2139062143;
not.b32 %r375, %r275;
and.b32 %r376, %r371, %r375;
and.b32 %r377, %r374, %r275;
or.b32 %r280, %r377, %r376;

dp4a.s32.s32 %r279, %r280, %r29, %r142;

mul.lo.s32 %r378, %r365, %r279;
cvt.rn.f32.s32 %f66, %r378;
fma.rn.ftz.f32 %f67, %f59, %f66, %f65;
fma.rn.ftz.f32 %f199, %f55, %f67, %f199;

$L__BB140_5:
and.b32 %r747, %r4, 8;
shr.u32 %r746, %r747, 1;
or.b32 %r417, %r146, 1;
shr.u32 %r419, %r4, 4;
mad.lo.s32 %r420, %r3, %r44, %r419;
mad.lo.s32 %r421, %r41, %r417, %r420;
mul.wide.s32 %rd115, %r421, 110;
add.s64 %rd116, %rd216, %rd115;
ld.global.nc.u16 %rs19, [%rd116+108];

{ cvt.f32.f16 %f68, %rs19;}


add.s64 %rd118, %rd51, %rd115;
add.s64 %rd119, %rd216, %rd118;
add.s64 %rd122, %rd55, %rd115;
add.s64 %rd123, %rd216, %rd122;
ld.global.nc.u16 %rs20, [%rd123];
ld.global.nc.u16 %rs21, [%rd123+2];
mov.b32 %r424, {%rs20, %rs21};
not.b32 %r425, %r424;
ld.global.nc.u16 %rs22, [%rd119+34];
ld.global.nc.u16 %rs23, [%rd119+32];

{.reg .f16 low,high;
mov.b32 {low,high},%r24;
cvt.f32.f16 %f69, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r26;
cvt.f32.f16 %f70, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r28;
cvt.f32.f16 %f71, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r30;
cvt.f32.f16 %f72, low;}


mov.b32 %r426, {%rs23, %rs22};
shr.s32 %r429, %r425, %r746;
add.s64 %rd125, %rd115, %rd58;
add.s64 %rd126, %rd216, %rd125;
ld.global.nc.u8 %rs24, [%rd126+96];
cvt.u32.u16 %r430, %rs24;
and.b32 %r431, %r430, 255;
shr.u32 %r432, %r431, %r7;
and.b32 %r433, %r432, 15;
add.s64 %rd128, %rd115, %rd61;
add.s64 %rd129, %rd216, %rd128;
ld.global.nc.u8 %rs25, [%rd129+96];
cvt.u32.u16 %r434, %rs25;
and.b32 %r435, %r434, 255;
shr.u32 %r436, %r435, %r8;
and.b32 %r437, %r436, 3;
bfi.b32 %r438, %r437, %r433, 4, 2;
add.s32 %r439, %r438, -32;
and.b32 %r385, %r426, 50529027;
shl.b32 %r440, %r429, 2;
and.b32 %r441, %r440, 67372036;
or.b32 %r442, %r385, -2139062144;
sub.s32 %r443, %r442, %r441;
xor.b32 %r444, %r443, -2139062144;
or.b32 %r445, %r385, %r441;
xor.b32 %r446, %r443, %r385;
and.b32 %r383, %r446, %r445;

prmt.b32 %r383, %r383, 0, 0xba98;


prmt.b32 %r385, %r385, 0, 0xba98;

xor.b32 %r447, %r385, 2139062143;
not.b32 %r448, %r383;
and.b32 %r449, %r444, %r448;
and.b32 %r450, %r447, %r383;
or.b32 %r388, %r450, %r449;

dp4a.s32.s32 %r387, %r388, %r23, %r142;

mul.lo.s32 %r451, %r439, %r387;
cvt.rn.f32.s32 %f73, %r451;
fma.rn.ftz.f32 %f74, %f69, %f73, 0f00000000;
add.s64 %rd131, %rd115, %rd64;
add.s64 %rd132, %rd216, %rd131;
ld.global.nc.u8 %rs26, [%rd132+96];
cvt.u32.u16 %r452, %rs26;
and.b32 %r453, %r452, 255;
shr.u32 %r454, %r453, %r11;
and.b32 %r455, %r454, 15;
add.s64 %rd134, %rd115, %rd67;
add.s64 %rd135, %rd216, %rd134;
ld.global.nc.u8 %rs27, [%rd135+96];
cvt.u32.u16 %r456, %rs27;
and.b32 %r457, %r456, 255;
shr.u32 %r458, %r457, %r12;
and.b32 %r459, %r458, 3;
bfi.b32 %r460, %r459, %r455, 4, 2;
add.s32 %r461, %r460, -32;
shr.u32 %r462, %r426, 2;
and.b32 %r393, %r462, 50529027;
shl.b32 %r463, %r429, 1;
and.b32 %r464, %r463, 67372036;
or.b32 %r465, %r393, -2139062144;
sub.s32 %r466, %r465, %r464;
xor.b32 %r467, %r466, -2139062144;
or.b32 %r468, %r393, %r464;
xor.b32 %r469, %r466, %r393;
and.b32 %r391, %r469, %r468;

prmt.b32 %r391, %r391, 0, 0xba98;


prmt.b32 %r393, %r393, 0, 0xba98;

xor.b32 %r470, %r393, 2139062143;
not.b32 %r471, %r391;
and.b32 %r472, %r467, %r471;
and.b32 %r473, %r470, %r391;
or.b32 %r396, %r473, %r472;

dp4a.s32.s32 %r395, %r396, %r25, %r142;

mul.lo.s32 %r474, %r461, %r395;
cvt.rn.f32.s32 %f75, %r474;
fma.rn.ftz.f32 %f76, %f70, %f75, %f74;
add.s64 %rd137, %rd115, %rd70;
add.s64 %rd138, %rd216, %rd137;
ld.global.nc.u8 %rs28, [%rd138+96];
cvt.u32.u16 %r475, %rs28;
and.b32 %r476, %r475, 255;
shr.u32 %r477, %r476, %r15;
and.b32 %r478, %r477, 15;
add.s64 %rd140, %rd115, %rd73;
add.s64 %rd141, %rd216, %rd140;
ld.global.nc.u8 %rs29, [%rd141+96];
cvt.u32.u16 %r479, %rs29;
and.b32 %r480, %r479, 255;
shr.u32 %r481, %r480, %r16;
and.b32 %r482, %r481, 3;
bfi.b32 %r483, %r482, %r478, 4, 2;
add.s32 %r484, %r483, -32;
shr.u32 %r485, %r426, 4;
and.b32 %r401, %r485, 50529027;
and.b32 %r486, %r429, 67372036;
or.b32 %r487, %r401, -2139062144;
sub.s32 %r488, %r487, %r486;
xor.b32 %r489, %r488, -2139062144;
or.b32 %r490, %r401, %r486;
xor.b32 %r491, %r488, %r401;
and.b32 %r399, %r491, %r490;

prmt.b32 %r399, %r399, 0, 0xba98;


prmt.b32 %r401, %r401, 0, 0xba98;

xor.b32 %r492, %r401, 2139062143;
not.b32 %r493, %r399;
and.b32 %r494, %r489, %r493;
and.b32 %r495, %r492, %r399;
or.b32 %r404, %r495, %r494;

dp4a.s32.s32 %r403, %r404, %r27, %r142;

mul.lo.s32 %r496, %r484, %r403;
cvt.rn.f32.s32 %f77, %r496;
fma.rn.ftz.f32 %f78, %f71, %f77, %f76;
add.s64 %rd143, %rd115, %rd76;
add.s64 %rd144, %rd216, %rd143;
ld.global.nc.u8 %rs30, [%rd144+96];
cvt.u32.u16 %r497, %rs30;
and.b32 %r498, %r497, 255;
shr.u32 %r499, %r498, %r19;
and.b32 %r500, %r499, 15;
add.s64 %rd146, %rd115, %rd79;
add.s64 %rd147, %rd216, %rd146;
ld.global.nc.u8 %rs31, [%rd147+96];
cvt.u32.u16 %r501, %rs31;
and.b32 %r502, %r501, 255;
shr.u32 %r503, %r502, %r20;
and.b32 %r504, %r503, 3;
bfi.b32 %r505, %r504, %r500, 4, 2;
add.s32 %r506, %r505, -32;
shr.u32 %r507, %r426, 6;
and.b32 %r409, %r507, 50529027;
shr.u32 %r508, %r429, 1;
and.b32 %r509, %r508, 67372036;
or.b32 %r510, %r409, -2139062144;
sub.s32 %r511, %r510, %r509;
xor.b32 %r512, %r511, -2139062144;
or.b32 %r513, %r409, %r509;
xor.b32 %r514, %r511, %r409;
and.b32 %r407, %r514, %r513;

prmt.b32 %r407, %r407, 0, 0xba98;


prmt.b32 %r409, %r409, 0, 0xba98;

xor.b32 %r515, %r409, 2139062143;
not.b32 %r516, %r407;
and.b32 %r517, %r512, %r516;
and.b32 %r518, %r515, %r407;
or.b32 %r412, %r518, %r517;

dp4a.s32.s32 %r411, %r412, %r29, %r142;

mul.lo.s32 %r519, %r506, %r411;
cvt.rn.f32.s32 %f79, %r519;
fma.rn.ftz.f32 %f80, %f72, %f79, %f78;
fma.rn.ftz.f32 %f200, %f68, %f80, %f200;
@%p3 bra $L__BB140_7;

and.b32 %r749, %r4, 8;
shr.u32 %r748, %r749, 1;
add.s64 %rd149, %rd213, %rd115;
ld.global.u16 %rs32, [%rd149+108];

{ cvt.f32.f16 %f81, %rs32;}


add.s64 %rd152, %rd213, %rd118;
add.s64 %rd156, %rd213, %rd122;
ld.global.u16 %rs33, [%rd156];
ld.global.u16 %rs34, [%rd156+2];
mov.b32 %r565, {%rs33, %rs34};
not.b32 %r566, %r565;
ld.global.u16 %rs35, [%rd152+34];
ld.global.u16 %rs36, [%rd152+32];

{.reg .f16 low,high;
mov.b32 {low,high},%r24;
cvt.f32.f16 %f82, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r26;
cvt.f32.f16 %f83, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r28;
cvt.f32.f16 %f84, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r30;
cvt.f32.f16 %f85, low;}


mov.b32 %r567, {%rs36, %rs35};
shr.s32 %r570, %r566, %r748;
add.s64 %rd159, %rd213, %rd125;
ld.global.u8 %r571, [%rd159+96];
shr.u32 %r572, %r571, %r7;
and.b32 %r573, %r572, 15;
add.s64 %rd162, %rd213, %rd128;
ld.global.u8 %r574, [%rd162+96];
shr.u32 %r575, %r574, %r8;
and.b32 %r576, %r575, 3;
bfi.b32 %r577, %r576, %r573, 4, 2;
add.s32 %r578, %r577, -32;
and.b32 %r526, %r567, 50529027;
shl.b32 %r579, %r570, 2;
and.b32 %r580, %r579, 67372036;
or.b32 %r581, %r526, -2139062144;
sub.s32 %r582, %r581, %r580;
xor.b32 %r583, %r582, -2139062144;
or.b32 %r584, %r526, %r580;
xor.b32 %r585, %r582, %r526;
and.b32 %r524, %r585, %r584;

prmt.b32 %r524, %r524, 0, 0xba98;


prmt.b32 %r526, %r526, 0, 0xba98;

xor.b32 %r586, %r526, 2139062143;
not.b32 %r587, %r524;
and.b32 %r588, %r583, %r587;
and.b32 %r589, %r586, %r524;
or.b32 %r529, %r589, %r588;
mov.u32 %r555, 0;

dp4a.s32.s32 %r528, %r529, %r23, %r555;

mul.lo.s32 %r590, %r578, %r528;
cvt.rn.f32.s32 %f86, %r590;
fma.rn.ftz.f32 %f87, %f82, %f86, 0f00000000;
add.s64 %rd165, %rd213, %rd131;
ld.global.u8 %r591, [%rd165+96];
shr.u32 %r592, %r591, %r11;
and.b32 %r593, %r592, 15;
add.s64 %rd168, %rd213, %rd134;
ld.global.u8 %r594, [%rd168+96];
shr.u32 %r595, %r594, %r12;
and.b32 %r596, %r595, 3;
bfi.b32 %r597, %r596, %r593, 4, 2;
add.s32 %r598, %r597, -32;
shr.u32 %r599, %r567, 2;
and.b32 %r534, %r599, 50529027;
shl.b32 %r600, %r570, 1;
and.b32 %r601, %r600, 67372036;
or.b32 %r602, %r534, -2139062144;
sub.s32 %r603, %r602, %r601;
xor.b32 %r604, %r603, -2139062144;
or.b32 %r605, %r534, %r601;
xor.b32 %r606, %r603, %r534;
and.b32 %r532, %r606, %r605;

prmt.b32 %r532, %r532, 0, 0xba98;


prmt.b32 %r534, %r534, 0, 0xba98;

xor.b32 %r607, %r534, 2139062143;
not.b32 %r608, %r532;
and.b32 %r609, %r604, %r608;
and.b32 %r610, %r607, %r532;
or.b32 %r537, %r610, %r609;

dp4a.s32.s32 %r536, %r537, %r25, %r555;

mul.lo.s32 %r611, %r598, %r536;
cvt.rn.f32.s32 %f88, %r611;
fma.rn.ftz.f32 %f89, %f83, %f88, %f87;
add.s64 %rd171, %rd213, %rd137;
ld.global.u8 %r612, [%rd171+96];
shr.u32 %r613, %r612, %r15;
and.b32 %r614, %r613, 15;
add.s64 %rd174, %rd213, %rd140;
ld.global.u8 %r615, [%rd174+96];
shr.u32 %r616, %r615, %r16;
and.b32 %r617, %r616, 3;
bfi.b32 %r618, %r617, %r614, 4, 2;
add.s32 %r619, %r618, -32;
shr.u32 %r620, %r567, 4;
and.b32 %r542, %r620, 50529027;
and.b32 %r621, %r570, 67372036;
or.b32 %r622, %r542, -2139062144;
sub.s32 %r623, %r622, %r621;
xor.b32 %r624, %r623, -2139062144;
or.b32 %r625, %r542, %r621;
xor.b32 %r626, %r623, %r542;
and.b32 %r540, %r626, %r625;

prmt.b32 %r540, %r540, 0, 0xba98;


prmt.b32 %r542, %r542, 0, 0xba98;

xor.b32 %r627, %r542, 2139062143;
not.b32 %r628, %r540;
and.b32 %r629, %r624, %r628;
and.b32 %r630, %r627, %r540;
or.b32 %r545, %r630, %r629;

dp4a.s32.s32 %r544, %r545, %r27, %r555;

mul.lo.s32 %r631, %r619, %r544;
cvt.rn.f32.s32 %f90, %r631;
fma.rn.ftz.f32 %f91, %f84, %f90, %f89;
add.s64 %rd177, %rd213, %rd143;
ld.global.u8 %r632, [%rd177+96];
shr.u32 %r633, %r632, %r19;
and.b32 %r634, %r633, 15;
add.s64 %rd180, %rd213, %rd146;
ld.global.u8 %r635, [%rd180+96];
shr.u32 %r636, %r635, %r20;
and.b32 %r637, %r636, 3;
bfi.b32 %r638, %r637, %r634, 4, 2;
add.s32 %r639, %r638, -32;
shr.u32 %r640, %r567, 6;
and.b32 %r550, %r640, 50529027;
shr.u32 %r641, %r570, 1;
and.b32 %r642, %r641, 67372036;
or.b32 %r643, %r550, -2139062144;
sub.s32 %r644, %r643, %r642;
xor.b32 %r645, %r644, -2139062144;
or.b32 %r646, %r550, %r642;
xor.b32 %r647, %r644, %r550;
and.b32 %r548, %r647, %r646;

prmt.b32 %r548, %r548, 0, 0xba98;


prmt.b32 %r550, %r550, 0, 0xba98;

xor.b32 %r648, %r550, 2139062143;
not.b32 %r649, %r548;
and.b32 %r650, %r645, %r649;
and.b32 %r651, %r648, %r548;
or.b32 %r553, %r651, %r650;

dp4a.s32.s32 %r552, %r553, %r29, %r555;

mul.lo.s32 %r652, %r639, %r552;
cvt.rn.f32.s32 %f92, %r652;
fma.rn.ftz.f32 %f93, %f85, %f92, %f91;
fma.rn.ftz.f32 %f198, %f81, %f93, %f198;

$L__BB140_7:
add.s64 %rd216, %rd216, 220;
add.s64 %rd215, %rd215, 576;
add.s64 %rd214, %rd214, 576;
add.s64 %rd213, %rd213, 220;
add.s32 %r752, %r752, 2;
setp.lt.u32 %p5, %r752, %r50;
@%p5 bra $L__BB140_3;

add.u64 %rd201, %SPL, 8;
add.u64 %rd199, %SPL, 0;
st.local.v2.f32 [%rd199], {%f201, %f200};
st.local.v2.f32 [%rd201], {%f199, %f198};

$L__BB140_9:
mov.b32 %r32, %f201;
setp.ne.s64 %p6, %rd23, 0;
@%p6 bra $L__BB140_11;
bra.uni $L__BB140_10;

$L__BB140_11:
add.u64 %rd211, %SPL, 8;
add.u64 %rd209, %SPL, 0;
mov.u32 %r680, 31;
mov.u32 %r681, 16;
mov.u32 %r682, -1;
shfl.sync.bfly.b32 %r683|%p17, %r32, %r681, %r680, %r682;
mov.b32 %f114, %r683;
add.ftz.f32 %f115, %f201, %f114;
mov.b32 %r684, %f115;
mov.u32 %r685, 8;
shfl.sync.bfly.b32 %r686|%p18, %r684, %r685, %r680, %r682;
mov.b32 %f116, %r686;
add.ftz.f32 %f117, %f115, %f116;
mov.b32 %r687, %f117;
mov.u32 %r688, 4;
shfl.sync.bfly.b32 %r689|%p19, %r687, %r688, %r680, %r682;
mov.b32 %f118, %r689;
add.ftz.f32 %f119, %f117, %f118;
mov.b32 %r690, %f119;
mov.u32 %r691, 2;
shfl.sync.bfly.b32 %r692|%p20, %r690, %r691, %r680, %r682;
mov.b32 %f120, %r692;
add.ftz.f32 %f121, %f119, %f120;
mov.b32 %r693, %f121;
mov.u32 %r694, 1;
shfl.sync.bfly.b32 %r695|%p21, %r693, %r694, %r680, %r682;
mov.b32 %f122, %r695;
add.ftz.f32 %f123, %f121, %f122;
st.local.f32 [%rd209], %f123;
mov.b32 %r696, %f199;
shfl.sync.bfly.b32 %r697|%p22, %r696, %r681, %r680, %r682;
mov.b32 %f124, %r697;
add.ftz.f32 %f125, %f199, %f124;
mov.b32 %r698, %f125;
shfl.sync.bfly.b32 %r699|%p23, %r698, %r685, %r680, %r682;
mov.b32 %f126, %r699;
add.ftz.f32 %f127, %f125, %f126;
mov.b32 %r700, %f127;
shfl.sync.bfly.b32 %r701|%p24, %r700, %r688, %r680, %r682;
mov.b32 %f128, %r701;
add.ftz.f32 %f129, %f127, %f128;
mov.b32 %r702, %f129;
shfl.sync.bfly.b32 %r703|%p25, %r702, %r691, %r680, %r682;
mov.b32 %f130, %r703;
add.ftz.f32 %f131, %f129, %f130;
mov.b32 %r704, %f131;
shfl.sync.bfly.b32 %r705|%p26, %r704, %r694, %r680, %r682;
mov.b32 %f132, %r705;
add.ftz.f32 %f133, %f131, %f132;
st.local.f32 [%rd211], %f133;
mov.b32 %r706, %f200;
shfl.sync.bfly.b32 %r707|%p27, %r706, %r681, %r680, %r682;
mov.b32 %f134, %r707;
add.ftz.f32 %f135, %f200, %f134;
mov.b32 %r708, %f135;
shfl.sync.bfly.b32 %r709|%p28, %r708, %r685, %r680, %r682;
mov.b32 %f136, %r709;
add.ftz.f32 %f137, %f135, %f136;
mov.b32 %r710, %f137;
shfl.sync.bfly.b32 %r711|%p29, %r710, %r688, %r680, %r682;
mov.b32 %f138, %r711;
add.ftz.f32 %f139, %f137, %f138;
mov.b32 %r712, %f139;
shfl.sync.bfly.b32 %r713|%p30, %r712, %r691, %r680, %r682;
mov.b32 %f140, %r713;
add.ftz.f32 %f141, %f139, %f140;
mov.b32 %r714, %f141;
shfl.sync.bfly.b32 %r715|%p31, %r714, %r694, %r680, %r682;
mov.b32 %f142, %r715;
add.ftz.f32 %f143, %f141, %f142;
st.local.f32 [%rd209+4], %f143;
mov.b32 %r716, %f198;
shfl.sync.bfly.b32 %r717|%p32, %r716, %r681, %r680, %r682;
mov.b32 %f144, %r717;
add.ftz.f32 %f145, %f198, %f144;
mov.b32 %r718, %f145;
shfl.sync.bfly.b32 %r719|%p33, %r718, %r685, %r680, %r682;
mov.b32 %f146, %r719;
add.ftz.f32 %f147, %f145, %f146;
mov.b32 %r720, %f147;
shfl.sync.bfly.b32 %r721|%p34, %r720, %r688, %r680, %r682;
mov.b32 %f148, %r721;
add.ftz.f32 %f149, %f147, %f148;
mov.b32 %r722, %f149;
shfl.sync.bfly.b32 %r723|%p35, %r722, %r691, %r680, %r682;
mov.b32 %f150, %r723;
add.ftz.f32 %f151, %f149, %f150;
mov.b32 %r724, %f151;
shfl.sync.bfly.b32 %r725|%p36, %r724, %r694, %r680, %r682;
mov.b32 %f152, %r725;
add.ftz.f32 %f153, %f151, %f152;
st.local.f32 [%rd211+4], %f153;
bra.uni $L__BB140_12;

$L__BB140_10:
add.u64 %rd203, %SPL, 0;
mov.u32 %r654, 31;
mov.u32 %r655, 16;
mov.u32 %r656, -1;
shfl.sync.bfly.b32 %r657|%p7, %r32, %r655, %r654, %r656;
mov.b32 %f94, %r657;
add.ftz.f32 %f95, %f201, %f94;
mov.b32 %r658, %f95;
mov.u32 %r659, 8;
shfl.sync.bfly.b32 %r660|%p8, %r658, %r659, %r654, %r656;
mov.b32 %f96, %r660;
add.ftz.f32 %f97, %f95, %f96;
mov.b32 %r661, %f97;
mov.u32 %r662, 4;
shfl.sync.bfly.b32 %r663|%p9, %r661, %r662, %r654, %r656;
mov.b32 %f98, %r663;
add.ftz.f32 %f99, %f97, %f98;
mov.b32 %r664, %f99;
mov.u32 %r665, 2;
shfl.sync.bfly.b32 %r666|%p10, %r664, %r665, %r654, %r656;
mov.b32 %f100, %r666;
add.ftz.f32 %f101, %f99, %f100;
mov.b32 %r667, %f101;
mov.u32 %r668, 1;
shfl.sync.bfly.b32 %r669|%p11, %r667, %r668, %r654, %r656;
mov.b32 %f102, %r669;
add.ftz.f32 %f103, %f101, %f102;
st.local.f32 [%rd203], %f103;
mov.b32 %r670, %f200;
shfl.sync.bfly.b32 %r671|%p12, %r670, %r655, %r654, %r656;
mov.b32 %f104, %r671;
add.ftz.f32 %f105, %f200, %f104;
mov.b32 %r672, %f105;
shfl.sync.bfly.b32 %r673|%p13, %r672, %r659, %r654, %r656;
mov.b32 %f106, %r673;
add.ftz.f32 %f107, %f105, %f106;
mov.b32 %r674, %f107;
shfl.sync.bfly.b32 %r675|%p14, %r674, %r662, %r654, %r656;
mov.b32 %f108, %r675;
add.ftz.f32 %f109, %f107, %f108;
mov.b32 %r676, %f109;
shfl.sync.bfly.b32 %r677|%p15, %r676, %r665, %r654, %r656;
mov.b32 %f110, %r677;
add.ftz.f32 %f111, %f109, %f110;
mov.b32 %r678, %f111;
shfl.sync.bfly.b32 %r679|%p16, %r678, %r668, %r654, %r656;
mov.b32 %f112, %r679;
add.ftz.f32 %f113, %f111, %f112;
st.local.f32 [%rd203+4], %f113;

$L__BB140_12:
ld.param.u32 %r735, [_Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7];
setp.gt.u32 %p37, %r4, 1;
mov.u32 %r727, %ctaid.x;
shl.b32 %r728, %r727, 1;
add.s32 %r33, %r728, %r4;
setp.ge.u32 %p38, %r33, %r735;
or.pred %p39, %p37, %p38;
@%p39 bra $L__BB140_29;

add.u64 %rd205, %SPL, 0;
ld.param.u64 %rd194, [_Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3];
ld.param.u32 %r736, [_Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13];
cvt.u64.u32 %rd17, %r4;
mul.wide.u32 %rd181, %r4, 4;
add.s64 %rd182, %rd205, %rd181;
ld.local.f32 %f204, [%rd182];
mad.lo.s32 %r34, %r3, %r736, %r33;
setp.eq.s64 %p40, %rd194, 0;
@%p40 bra $L__BB140_15;

ld.param.u64 %rd195, [_Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3];
cvta.to.global.u64 %rd183, %rd195;
mul.wide.u32 %rd184, %r34, 4;
add.s64 %rd185, %rd183, %rd184;
ld.global.f32 %f154, [%rd185];
add.ftz.f32 %f204, %f204, %f154;

$L__BB140_15:
setp.eq.s64 %p41, %rd23, 0;
@%p41 bra $L__BB140_28;

add.u64 %rd207, %SPL, 8;
ld.param.u64 %rd197, [_Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+16];
shl.b64 %rd186, %rd17, 2;
add.s64 %rd187, %rd207, %rd186;
ld.local.f32 %f24, [%rd187];
setp.eq.s64 %p42, %rd197, 0;
@%p42 bra $L__BB140_18;

ld.param.u64 %rd198, [_Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+16];
cvta.to.global.u64 %rd188, %rd198;
mul.wide.u32 %rd189, %r34, 4;
add.s64 %rd190, %rd188, %rd189;
ld.global.f32 %f155, [%rd190];
add.ftz.f32 %f24, %f24, %f155;

$L__BB140_18:
ld.param.u32 %r741, [_Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.gt.s32 %p43, %r741, 2;
@%p43 bra $L__BB140_22;

ld.param.u32 %r744, [_Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.eq.s32 %p46, %r744, 1;
@%p46 bra $L__BB140_26;

ld.param.u32 %r745, [_Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.eq.s32 %p47, %r745, 2;
@%p47 bra $L__BB140_21;
bra.uni $L__BB140_27;

$L__BB140_21:
mul.ftz.f32 %f175, %f24, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f176, %f175;
add.ftz.f32 %f177, %f176, 0f3F800000;
div.approx.ftz.f32 %f178, %f24, %f177;
mul.ftz.f32 %f204, %f204, %f178;
bra.uni $L__BB140_28;

$L__BB140_22:
ld.param.u32 %r742, [_Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.eq.s32 %p44, %r742, 3;
@%p44 bra $L__BB140_25;

ld.param.u32 %r743, [_Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.ne.s32 %p45, %r743, 6;
@%p45 bra $L__BB140_27;

ld.param.f32 %f187, [_Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+44];
min.ftz.f32 %f156, %f24, %f187;
neg.ftz.f32 %f157, %f187;
min.ftz.f32 %f158, %f204, %f187;
max.ftz.f32 %f159, %f158, %f157;
mul.ftz.f32 %f160, %f156, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f161, %f160;
add.ftz.f32 %f162, %f161, 0f3F800000;
div.approx.ftz.f32 %f163, %f156, %f162;
mul.ftz.f32 %f204, %f159, %f163;
bra.uni $L__BB140_28;

$L__BB140_27:
mul.ftz.f32 %f204, %f204, %f24;
bra.uni $L__BB140_28;

$L__BB140_26:
mul.ftz.f32 %f179, %f24, 0f3F000000;
mul.ftz.f32 %f180, %f24, 0f3D372713;
fma.rn.ftz.f32 %f181, %f24, %f180, 0f3F800000;
mul.ftz.f32 %f182, %f24, 0f3F4C422A;
mul.ftz.f32 %f183, %f182, %f181;
tanh.approx.f32 %f184, %f183;
add.ftz.f32 %f185, %f184, 0f3F800000;
mul.ftz.f32 %f186, %f179, %f185;
mul.ftz.f32 %f204, %f204, %f186;
bra.uni $L__BB140_28;

$L__BB140_25:
mov.f32 %f164, 0f40E00000;
min.ftz.f32 %f165, %f24, %f164;
min.ftz.f32 %f166, %f204, %f164;
mov.f32 %f167, 0fC0E00000;
max.ftz.f32 %f168, %f166, %f167;
mul.ftz.f32 %f169, %f165, 0f3FD9DB23;
mul.ftz.f32 %f170, %f169, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f171, %f170;
add.ftz.f32 %f172, %f171, 0f3F800000;
div.approx.ftz.f32 %f173, %f165, %f172;
add.ftz.f32 %f174, %f168, 0f3F800000;
mul.ftz.f32 %f204, %f174, %f173;

$L__BB140_28:
ld.param.u64 %rd196, [_Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4];
ld.param.u32 %r740, [_Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10];
ld.param.u32 %r739, [_Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13];
mov.u32 %r738, %ctaid.y;
mov.u32 %r737, %tid.y;
mad.lo.s32 %r733, %r737, %r740, %r33;
mad.lo.s32 %r734, %r738, %r739, %r733;
cvta.to.global.u64 %rd191, %rd196;
mul.wide.u32 %rd192, %r734, 4;
add.s64 %rd193, %rd191, %rd192;
st.global.f32 [%rd193], %f204;

$L__BB140_29:
ret;

}
.entry _Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj(
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2,
.param .align 8 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3[48],
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5,
.param .align 4 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6[12],
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15
)
.maxntid 256, 1, 1
.minnctapersm 1
{
.local .align 8 .b8 __local_depot141[8];
.reg .b64 %SP;
.reg .b64 %SPL;
.reg .pred %p<17>;
.reg .b16 %rs<27>;
.reg .f32 %f<62>;
.reg .b32 %r<410>;
.reg .b64 %rd<122>;


mov.u64 %SPL, __local_depot141;
ld.param.u64 %rd28, [_Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0];
ld.param.u64 %rd29, [_Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1];
ld.param.u64 %rd30, [_Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2];
ld.param.u32 %r16, [_Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5];
ld.param.u32 %r21, [_Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8];
ld.param.u32 %r22, [_Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9];
ld.param.u32 %r24, [_Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11];
ld.param.u32 %r25, [_Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12];
ld.param.u32 %r28, [_Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14];
ld.param.u32 %r27, [_Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15];
ld.param.u32 %r19, [_Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+8];
ld.param.u32 %r18, [_Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+4];
ld.param.u32 %r17, [_Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6];
add.u64 %rd1, %SPL, 0;
mov.u32 %r1, %tid.y;
setp.ge.u32 %p1, %r1, %r28;
@%p1 bra $L__BB141_7;

mov.f32 %f60, 0f00000000;
st.local.v2.f32 [%rd1], {%f60, %f60};
mov.u32 %r2, %tid.x;
shr.u32 %r409, %r2, 4;
shr.u32 %r29, %r16, 8;
setp.ge.u32 %p2, %r409, %r29;
mov.f32 %f61, %f60;
@%p2 bra $L__BB141_5;

mov.u32 %r30, %ctaid.x;
shl.b32 %r31, %r30, 1;
mov.u32 %r32, %ctaid.y;
cvta.to.global.u64 %rd33, %rd29;
mad.lo.s32 %r33, %r1, %r27, %r32;
cvta.to.global.u64 %rd34, %rd30;
mul.wide.u32 %rd35, %r33, 4;
add.s64 %rd36, %rd34, %rd35;
ld.global.nc.u32 %r34, [%rd36];
mul.lo.s32 %r35, %r34, %r24;
and.b32 %r36, %r2, 15;
and.b32 %r37, %r2, 7;
sub.s32 %r38, %r36, %r37;
and.b32 %r39, %r2, 4;
shr.u32 %r40, %r39, 2;
add.s32 %r41, %r38, %r40;
shr.s32 %r42, %r41, 31;
shr.u32 %r43, %r42, 29;
add.s32 %r44, %r41, %r43;
and.b32 %r45, %r44, -8;
sub.s32 %r46, %r41, %r45;
shr.s32 %r47, %r44, 3;
shl.b32 %r4, %r47, 2;
cvt.s64.s32 %rd37, %r46;
shr.u32 %r48, %r42, 30;
add.s32 %r49, %r41, %r48;
and.b32 %r50, %r49, -4;
sub.s32 %r51, %r41, %r50;
shr.s32 %r52, %r49, 2;
shl.b32 %r5, %r52, 1;
add.s32 %r53, %r51, 8;
cvt.s64.s32 %rd38, %r53;
add.s32 %r54, %r41, 2;
shr.s32 %r55, %r54, 31;
shr.u32 %r56, %r55, 29;
add.s32 %r57, %r54, %r56;
and.b32 %r58, %r57, -8;
sub.s32 %r59, %r54, %r58;
shr.s32 %r60, %r57, 3;
shl.b32 %r6, %r60, 2;
cvt.s64.s32 %rd39, %r59;
shr.u32 %r61, %r55, 30;
add.s32 %r62, %r54, %r61;
and.b32 %r63, %r62, -4;
sub.s32 %r64, %r54, %r63;
shr.s32 %r65, %r62, 2;
shl.b32 %r7, %r65, 1;
add.s32 %r66, %r64, 8;
cvt.s64.s32 %rd40, %r66;
add.s32 %r67, %r41, 4;
shr.s32 %r68, %r67, 31;
shr.u32 %r69, %r68, 29;
add.s32 %r70, %r67, %r69;
and.b32 %r71, %r70, -8;
sub.s32 %r72, %r67, %r71;
shr.s32 %r73, %r70, 3;
shl.b32 %r8, %r73, 2;
cvt.s64.s32 %rd41, %r72;
shr.u32 %r74, %r68, 30;
add.s32 %r75, %r67, %r74;
and.b32 %r76, %r75, -4;
sub.s32 %r77, %r67, %r76;
shr.s32 %r78, %r75, 2;
shl.b32 %r9, %r78, 1;
add.s32 %r79, %r77, 8;
cvt.s64.s32 %rd42, %r79;
add.s32 %r80, %r41, 6;
shr.s32 %r81, %r80, 31;
shr.u32 %r82, %r81, 29;
add.s32 %r83, %r80, %r82;
and.b32 %r84, %r83, -8;
sub.s32 %r85, %r80, %r84;
shr.s32 %r86, %r83, 3;
shl.b32 %r10, %r86, 2;
cvt.s64.s32 %rd43, %r85;
shr.u32 %r87, %r81, 30;
add.s32 %r88, %r80, %r87;
and.b32 %r89, %r88, -4;
sub.s32 %r90, %r80, %r89;
shr.s32 %r91, %r88, 2;
shl.b32 %r11, %r91, 1;
add.s32 %r92, %r90, 8;
cvt.s64.s32 %rd44, %r92;
cvt.u64.u32 %rd45, %r2;
and.b64 %rd46, %rd45, 8;
mad.lo.s32 %r93, %r31, %r21, %r409;
add.s32 %r94, %r93, %r35;
mul.wide.s32 %rd2, %r94, 110;
mul.wide.u32 %rd47, %r2, 4;
and.b64 %rd48, %rd47, 28;
shr.u64 %rd49, %rd45, 4;
mul.lo.s64 %rd50, %rd49, 288;
shr.u64 %rd51, %rd46, 3;
mul.lo.s64 %rd52, %rd51, 144;
add.s64 %rd53, %rd50, %rd52;
mul.lo.s32 %r95, %r1, %r22;
mul.wide.u32 %rd54, %r95, 36;
add.s64 %rd55, %rd53, %rd54;
mul.hi.u32 %r96, %r32, %r17;
add.s32 %r97, %r96, %r32;
shr.u32 %r98, %r97, %r18;
mul.lo.s32 %r99, %r98, %r19;
sub.s32 %r100, %r32, %r99;
mul.lo.s32 %r101, %r100, %r25;
mul.wide.u32 %rd56, %r101, 36;
add.s64 %rd57, %rd55, %rd56;
add.s64 %rd58, %rd57, %rd48;
add.s64 %rd59, %rd33, %rd58;
add.s64 %rd120, %rd59, 76;
add.s64 %rd60, %rd2, %rd37;
add.s64 %rd4, %rd60, 96;
add.s64 %rd61, %rd2, %rd38;
add.s64 %rd5, %rd61, 96;
add.s64 %rd119, %rd33, %rd57;
add.s64 %rd62, %rd2, %rd39;
add.s64 %rd7, %rd62, 96;
or.b32 %r102, %r31, 1;
add.s32 %r103, %r409, %r35;
mad.lo.s32 %r12, %r21, %r102, %r103;
mul.wide.s32 %rd63, %r12, 110;
add.s64 %rd64, %rd63, %rd44;
add.s64 %rd8, %rd64, 96;
add.s64 %rd65, %rd63, %rd43;
add.s64 %rd9, %rd65, 96;
add.s64 %rd66, %rd63, %rd42;
add.s64 %rd10, %rd66, 96;
add.s64 %rd67, %rd63, %rd41;
add.s64 %rd11, %rd67, 96;
add.s64 %rd68, %rd63, %rd40;
add.s64 %rd12, %rd68, 96;
add.s64 %rd69, %rd63, %rd39;
add.s64 %rd13, %rd69, 96;
add.s64 %rd70, %rd63, %rd38;
add.s64 %rd14, %rd70, 96;
add.s64 %rd71, %rd63, %rd37;
add.s64 %rd15, %rd71, 96;
add.s64 %rd72, %rd2, %rd40;
add.s64 %rd16, %rd72, 96;
add.s64 %rd73, %rd2, %rd41;
add.s64 %rd17, %rd73, 96;
add.s64 %rd74, %rd2, %rd44;
add.s64 %rd18, %rd74, 96;
add.s64 %rd75, %rd2, %rd43;
add.s64 %rd19, %rd75, 96;
add.s64 %rd76, %rd2, %rd42;
add.s64 %rd20, %rd76, 96;
cvta.to.global.u64 %rd121, %rd28;
mov.f32 %f60, 0f00000000;

$L__BB141_3:
add.s64 %rd77, %rd121, %rd2;
ld.global.nc.u16 %rs1, [%rd77+108];

{ cvt.f32.f16 %f11, %rs1;}


mul.wide.u32 %rd78, %r36, 4;
add.s64 %rd79, %rd78, %rd2;
add.s64 %rd80, %rd121, %rd79;
shl.b32 %r178, %r2, 2;
cvt.u64.u32 %rd81, %r178;
and.b64 %rd82, %rd81, 28;
add.s64 %rd83, %rd82, %rd2;
add.s64 %rd84, %rd121, %rd83;
ld.global.nc.u16 %rs3, [%rd84];
ld.global.nc.u16 %rs4, [%rd84+2];
mov.b32 %r179, {%rs3, %rs4};
not.b32 %r180, %r179;
ld.global.nc.u16 %rs5, [%rd80+34];
ld.global.nc.u16 %rs6, [%rd80+32];
ld.global.nc.u32 %r114, [%rd120+-72];
ld.global.nc.u32 %r104, [%rd119];

{.reg .f16 low,high;
mov.b32 {low,high},%r104;
cvt.f32.f16 %f12, low;}


ld.global.nc.u32 %r122, [%rd120+-36];
ld.global.nc.u32 %r105, [%rd119+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r105;
cvt.f32.f16 %f13, low;}


ld.global.nc.u32 %r130, [%rd120];
ld.global.nc.u32 %r106, [%rd119+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r106;
cvt.f32.f16 %f14, low;}


ld.global.nc.u32 %r138, [%rd120+36];
ld.global.nc.u32 %r107, [%rd119+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r107;
cvt.f32.f16 %f15, low;}


mov.b32 %r181, {%rs6, %rs5};
and.b32 %r182, %r2, 8;
shr.u32 %r183, %r182, 1;
shr.s32 %r184, %r180, %r183;
add.s64 %rd85, %rd121, %rd4;
ld.global.nc.u8 %rs7, [%rd85];
cvt.u32.u16 %r185, %rs7;
and.b32 %r186, %r185, 255;
shr.u32 %r187, %r186, %r4;
and.b32 %r188, %r187, 15;
add.s64 %rd86, %rd121, %rd5;
ld.global.nc.u8 %rs8, [%rd86];
cvt.u32.u16 %r189, %rs8;
and.b32 %r190, %r189, 255;
shr.u32 %r191, %r190, %r5;
and.b32 %r192, %r191, 3;
bfi.b32 %r193, %r192, %r188, 4, 2;
add.s32 %r194, %r193, -32;
and.b32 %r110, %r181, 50529027;
shl.b32 %r195, %r184, 2;
and.b32 %r196, %r195, 67372036;
or.b32 %r197, %r110, -2139062144;
sub.s32 %r198, %r197, %r196;
xor.b32 %r199, %r198, -2139062144;
or.b32 %r200, %r110, %r196;
xor.b32 %r201, %r198, %r110;
and.b32 %r108, %r201, %r200;

prmt.b32 %r108, %r108, 0, 0xba98;


prmt.b32 %r110, %r110, 0, 0xba98;

xor.b32 %r202, %r110, 2139062143;
not.b32 %r203, %r108;
and.b32 %r204, %r199, %r203;
and.b32 %r205, %r202, %r108;
or.b32 %r113, %r205, %r204;
mov.u32 %r175, 0;

dp4a.s32.s32 %r112, %r113, %r114, %r175;

mul.lo.s32 %r206, %r194, %r112;
cvt.rn.f32.s32 %f21, %r206;
fma.rn.ftz.f32 %f22, %f12, %f21, 0f00000000;
add.s64 %rd87, %rd121, %rd7;
ld.global.nc.u8 %rs9, [%rd87];
cvt.u32.u16 %r207, %rs9;
and.b32 %r208, %r207, 255;
shr.u32 %r209, %r208, %r6;
and.b32 %r210, %r209, 15;
add.s64 %rd88, %rd121, %rd16;
ld.global.nc.u8 %rs10, [%rd88];
cvt.u32.u16 %r211, %rs10;
and.b32 %r212, %r211, 255;
shr.u32 %r213, %r212, %r7;
and.b32 %r214, %r213, 3;
bfi.b32 %r215, %r214, %r210, 4, 2;
add.s32 %r216, %r215, -32;
shr.u32 %r217, %r181, 2;
and.b32 %r118, %r217, 50529027;
shl.b32 %r218, %r184, 1;
and.b32 %r219, %r218, 67372036;
or.b32 %r220, %r118, -2139062144;
sub.s32 %r221, %r220, %r219;
xor.b32 %r222, %r221, -2139062144;
or.b32 %r223, %r118, %r219;
xor.b32 %r224, %r221, %r118;
and.b32 %r116, %r224, %r223;

prmt.b32 %r116, %r116, 0, 0xba98;


prmt.b32 %r118, %r118, 0, 0xba98;

xor.b32 %r225, %r118, 2139062143;
not.b32 %r226, %r116;
and.b32 %r227, %r222, %r226;
and.b32 %r228, %r225, %r116;
or.b32 %r121, %r228, %r227;

dp4a.s32.s32 %r120, %r121, %r122, %r175;

mul.lo.s32 %r229, %r216, %r120;
cvt.rn.f32.s32 %f23, %r229;
fma.rn.ftz.f32 %f24, %f13, %f23, %f22;
add.s64 %rd89, %rd121, %rd17;
ld.global.nc.u8 %rs11, [%rd89];
cvt.u32.u16 %r230, %rs11;
and.b32 %r231, %r230, 255;
shr.u32 %r232, %r231, %r8;
and.b32 %r233, %r232, 15;
add.s64 %rd90, %rd121, %rd20;
ld.global.nc.u8 %rs12, [%rd90];
cvt.u32.u16 %r234, %rs12;
and.b32 %r235, %r234, 255;
shr.u32 %r236, %r235, %r9;
and.b32 %r237, %r236, 3;
bfi.b32 %r238, %r237, %r233, 4, 2;
add.s32 %r239, %r238, -32;
shr.u32 %r240, %r181, 4;
and.b32 %r126, %r240, 50529027;
and.b32 %r241, %r184, 67372036;
or.b32 %r242, %r126, -2139062144;
sub.s32 %r243, %r242, %r241;
xor.b32 %r244, %r243, -2139062144;
or.b32 %r245, %r126, %r241;
xor.b32 %r246, %r243, %r126;
and.b32 %r124, %r246, %r245;

prmt.b32 %r124, %r124, 0, 0xba98;


prmt.b32 %r126, %r126, 0, 0xba98;

xor.b32 %r247, %r126, 2139062143;
not.b32 %r248, %r124;
and.b32 %r249, %r244, %r248;
and.b32 %r250, %r247, %r124;
or.b32 %r129, %r250, %r249;

dp4a.s32.s32 %r128, %r129, %r130, %r175;

mul.lo.s32 %r251, %r239, %r128;
cvt.rn.f32.s32 %f25, %r251;
fma.rn.ftz.f32 %f26, %f14, %f25, %f24;
add.s64 %rd91, %rd121, %rd19;
ld.global.nc.u8 %rs13, [%rd91];
cvt.u32.u16 %r252, %rs13;
and.b32 %r253, %r252, 255;
shr.u32 %r254, %r253, %r10;
and.b32 %r255, %r254, 15;
add.s64 %rd92, %rd121, %rd18;
ld.global.nc.u8 %rs14, [%rd92];
cvt.u32.u16 %r256, %rs14;
and.b32 %r257, %r256, 255;
shr.u32 %r258, %r257, %r11;
and.b32 %r259, %r258, 3;
bfi.b32 %r260, %r259, %r255, 4, 2;
add.s32 %r261, %r260, -32;
shr.u32 %r262, %r181, 6;
and.b32 %r134, %r262, 50529027;
shr.u32 %r263, %r184, 1;
and.b32 %r264, %r263, 67372036;
or.b32 %r265, %r134, -2139062144;
sub.s32 %r266, %r265, %r264;
xor.b32 %r267, %r266, -2139062144;
or.b32 %r268, %r134, %r264;
xor.b32 %r269, %r266, %r134;
and.b32 %r132, %r269, %r268;

prmt.b32 %r132, %r132, 0, 0xba98;


prmt.b32 %r134, %r134, 0, 0xba98;

xor.b32 %r270, %r134, 2139062143;
not.b32 %r271, %r132;
and.b32 %r272, %r267, %r271;
and.b32 %r273, %r270, %r132;
or.b32 %r137, %r273, %r272;

dp4a.s32.s32 %r136, %r137, %r138, %r175;

mul.lo.s32 %r274, %r261, %r136;
cvt.rn.f32.s32 %f27, %r274;
fma.rn.ftz.f32 %f28, %f15, %f27, %f26;
fma.rn.ftz.f32 %f61, %f11, %f28, %f61;
add.s64 %rd94, %rd121, %rd63;
ld.global.nc.u16 %rs2, [%rd94+108];

{ cvt.f32.f16 %f16, %rs2;}


add.s64 %rd95, %rd78, %rd63;
add.s64 %rd96, %rd121, %rd95;
add.s64 %rd97, %rd82, %rd63;
add.s64 %rd98, %rd121, %rd97;
ld.global.nc.u16 %rs15, [%rd98];
ld.global.nc.u16 %rs16, [%rd98+2];
mov.b32 %r275, {%rs15, %rs16};
not.b32 %r276, %r275;
ld.global.nc.u16 %rs17, [%rd96+34];
ld.global.nc.u16 %rs18, [%rd96+32];

{.reg .f16 low,high;
mov.b32 {low,high},%r104;
cvt.f32.f16 %f17, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r105;
cvt.f32.f16 %f18, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r106;
cvt.f32.f16 %f19, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r107;
cvt.f32.f16 %f20, low;}


mov.b32 %r277, {%rs18, %rs17};
shr.s32 %r278, %r276, %r183;
add.s64 %rd99, %rd121, %rd15;
ld.global.nc.u8 %rs19, [%rd99];
cvt.u32.u16 %r279, %rs19;
and.b32 %r280, %r279, 255;
shr.u32 %r281, %r280, %r4;
and.b32 %r282, %r281, 15;
add.s64 %rd100, %rd121, %rd14;
ld.global.nc.u8 %rs20, [%rd100];
cvt.u32.u16 %r283, %rs20;
and.b32 %r284, %r283, 255;
shr.u32 %r285, %r284, %r5;
and.b32 %r286, %r285, 3;
bfi.b32 %r287, %r286, %r282, 4, 2;
add.s32 %r288, %r287, -32;
and.b32 %r146, %r277, 50529027;
shl.b32 %r289, %r278, 2;
and.b32 %r290, %r289, 67372036;
or.b32 %r291, %r146, -2139062144;
sub.s32 %r292, %r291, %r290;
xor.b32 %r293, %r292, -2139062144;
or.b32 %r294, %r146, %r290;
xor.b32 %r295, %r292, %r146;
and.b32 %r144, %r295, %r294;

prmt.b32 %r144, %r144, 0, 0xba98;


prmt.b32 %r146, %r146, 0, 0xba98;

xor.b32 %r296, %r146, 2139062143;
not.b32 %r297, %r144;
and.b32 %r298, %r293, %r297;
and.b32 %r299, %r296, %r144;
or.b32 %r149, %r299, %r298;

dp4a.s32.s32 %r148, %r149, %r114, %r175;

mul.lo.s32 %r300, %r288, %r148;
cvt.rn.f32.s32 %f29, %r300;
fma.rn.ftz.f32 %f30, %f17, %f29, 0f00000000;
add.s64 %rd101, %rd121, %rd13;
ld.global.nc.u8 %rs21, [%rd101];
cvt.u32.u16 %r301, %rs21;
and.b32 %r302, %r301, 255;
shr.u32 %r303, %r302, %r6;
and.b32 %r304, %r303, 15;
add.s64 %rd102, %rd121, %rd12;
ld.global.nc.u8 %rs22, [%rd102];
cvt.u32.u16 %r305, %rs22;
and.b32 %r306, %r305, 255;
shr.u32 %r307, %r306, %r7;
and.b32 %r308, %r307, 3;
bfi.b32 %r309, %r308, %r304, 4, 2;
add.s32 %r310, %r309, -32;
shr.u32 %r311, %r277, 2;
and.b32 %r154, %r311, 50529027;
shl.b32 %r312, %r278, 1;
and.b32 %r313, %r312, 67372036;
or.b32 %r314, %r154, -2139062144;
sub.s32 %r315, %r314, %r313;
xor.b32 %r316, %r315, -2139062144;
or.b32 %r317, %r154, %r313;
xor.b32 %r318, %r315, %r154;
and.b32 %r152, %r318, %r317;

prmt.b32 %r152, %r152, 0, 0xba98;


prmt.b32 %r154, %r154, 0, 0xba98;

xor.b32 %r319, %r154, 2139062143;
not.b32 %r320, %r152;
and.b32 %r321, %r316, %r320;
and.b32 %r322, %r319, %r152;
or.b32 %r157, %r322, %r321;

dp4a.s32.s32 %r156, %r157, %r122, %r175;

mul.lo.s32 %r323, %r310, %r156;
cvt.rn.f32.s32 %f31, %r323;
fma.rn.ftz.f32 %f32, %f18, %f31, %f30;
add.s64 %rd103, %rd121, %rd11;
ld.global.nc.u8 %rs23, [%rd103];
cvt.u32.u16 %r324, %rs23;
and.b32 %r325, %r324, 255;
shr.u32 %r326, %r325, %r8;
and.b32 %r327, %r326, 15;
add.s64 %rd104, %rd121, %rd10;
ld.global.nc.u8 %rs24, [%rd104];
cvt.u32.u16 %r328, %rs24;
and.b32 %r329, %r328, 255;
shr.u32 %r330, %r329, %r9;
and.b32 %r331, %r330, 3;
bfi.b32 %r332, %r331, %r327, 4, 2;
add.s32 %r333, %r332, -32;
shr.u32 %r334, %r277, 4;
and.b32 %r162, %r334, 50529027;
and.b32 %r335, %r278, 67372036;
or.b32 %r336, %r162, -2139062144;
sub.s32 %r337, %r336, %r335;
xor.b32 %r338, %r337, -2139062144;
or.b32 %r339, %r162, %r335;
xor.b32 %r340, %r337, %r162;
and.b32 %r160, %r340, %r339;

prmt.b32 %r160, %r160, 0, 0xba98;


prmt.b32 %r162, %r162, 0, 0xba98;

xor.b32 %r341, %r162, 2139062143;
not.b32 %r342, %r160;
and.b32 %r343, %r338, %r342;
and.b32 %r344, %r341, %r160;
or.b32 %r165, %r344, %r343;

dp4a.s32.s32 %r164, %r165, %r130, %r175;

mul.lo.s32 %r345, %r333, %r164;
cvt.rn.f32.s32 %f33, %r345;
fma.rn.ftz.f32 %f34, %f19, %f33, %f32;
add.s64 %rd105, %rd121, %rd9;
ld.global.nc.u8 %rs25, [%rd105];
cvt.u32.u16 %r346, %rs25;
and.b32 %r347, %r346, 255;
shr.u32 %r348, %r347, %r10;
and.b32 %r349, %r348, 15;
add.s64 %rd106, %rd121, %rd8;
ld.global.nc.u8 %rs26, [%rd106];
cvt.u32.u16 %r350, %rs26;
and.b32 %r351, %r350, 255;
shr.u32 %r352, %r351, %r11;
and.b32 %r353, %r352, 3;
bfi.b32 %r354, %r353, %r349, 4, 2;
add.s32 %r355, %r354, -32;
shr.u32 %r356, %r277, 6;
and.b32 %r170, %r356, 50529027;
shr.u32 %r357, %r278, 1;
and.b32 %r358, %r357, 67372036;
or.b32 %r359, %r170, -2139062144;
sub.s32 %r360, %r359, %r358;
xor.b32 %r361, %r360, -2139062144;
or.b32 %r362, %r170, %r358;
xor.b32 %r363, %r360, %r170;
and.b32 %r168, %r363, %r362;

prmt.b32 %r168, %r168, 0, 0xba98;


prmt.b32 %r170, %r170, 0, 0xba98;

xor.b32 %r364, %r170, 2139062143;
not.b32 %r365, %r168;
and.b32 %r366, %r361, %r365;
and.b32 %r367, %r364, %r168;
or.b32 %r173, %r367, %r366;

dp4a.s32.s32 %r172, %r173, %r138, %r175;

mul.lo.s32 %r368, %r355, %r172;
cvt.rn.f32.s32 %f35, %r368;
fma.rn.ftz.f32 %f36, %f20, %f35, %f34;
fma.rn.ftz.f32 %f60, %f16, %f36, %f60;
add.s64 %rd121, %rd121, 220;
add.s64 %rd120, %rd120, 576;
add.s64 %rd119, %rd119, 576;
add.s32 %r409, %r409, 2;
setp.lt.u32 %p3, %r409, %r29;
@%p3 bra $L__BB141_3;

add.u64 %rd113, %SPL, 0;
st.local.v2.f32 [%rd113], {%f61, %f60};

$L__BB141_5:
add.u64 %rd115, %SPL, 0;
ld.param.u32 %r405, [_Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7];
mov.b32 %r370, %f61;
mov.u32 %r371, 31;
mov.u32 %r372, 16;
mov.u32 %r373, -1;
shfl.sync.bfly.b32 %r374|%p4, %r370, %r372, %r371, %r373;
mov.b32 %f37, %r374;
add.ftz.f32 %f38, %f61, %f37;
mov.b32 %r375, %f38;
mov.u32 %r376, 8;
shfl.sync.bfly.b32 %r377|%p5, %r375, %r376, %r371, %r373;
mov.b32 %f39, %r377;
add.ftz.f32 %f40, %f38, %f39;
mov.b32 %r378, %f40;
mov.u32 %r379, 4;
shfl.sync.bfly.b32 %r380|%p6, %r378, %r379, %r371, %r373;
mov.b32 %f41, %r380;
add.ftz.f32 %f42, %f40, %f41;
mov.b32 %r381, %f42;
mov.u32 %r382, 2;
shfl.sync.bfly.b32 %r383|%p7, %r381, %r382, %r371, %r373;
mov.b32 %f43, %r383;
add.ftz.f32 %f44, %f42, %f43;
mov.b32 %r384, %f44;
mov.u32 %r385, 1;
shfl.sync.bfly.b32 %r386|%p8, %r384, %r385, %r371, %r373;
mov.b32 %f45, %r386;
add.ftz.f32 %f46, %f44, %f45;
st.local.f32 [%rd115], %f46;
mov.b32 %r387, %f60;
shfl.sync.bfly.b32 %r388|%p9, %r387, %r372, %r371, %r373;
mov.b32 %f47, %r388;
add.ftz.f32 %f48, %f60, %f47;
mov.b32 %r389, %f48;
shfl.sync.bfly.b32 %r390|%p10, %r389, %r376, %r371, %r373;
mov.b32 %f49, %r390;
add.ftz.f32 %f50, %f48, %f49;
mov.b32 %r391, %f50;
shfl.sync.bfly.b32 %r392|%p11, %r391, %r379, %r371, %r373;
mov.b32 %f51, %r392;
add.ftz.f32 %f52, %f50, %f51;
mov.b32 %r393, %f52;
shfl.sync.bfly.b32 %r394|%p12, %r393, %r382, %r371, %r373;
mov.b32 %f53, %r394;
add.ftz.f32 %f54, %f52, %f53;
mov.b32 %r395, %f54;
shfl.sync.bfly.b32 %r396|%p13, %r395, %r385, %r371, %r373;
mov.b32 %f55, %r396;
add.ftz.f32 %f56, %f54, %f55;
st.local.f32 [%rd115+4], %f56;
setp.gt.u32 %p14, %r2, 1;
mov.u32 %r398, %ctaid.x;
shl.b32 %r399, %r398, 1;
add.s32 %r15, %r399, %r2;
setp.ge.u32 %p15, %r15, %r405;
or.pred %p16, %p14, %p15;
@%p16 bra $L__BB141_7;

add.u64 %rd117, %SPL, 0;
ld.param.u64 %rd112, [_Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4];
ld.param.u32 %r408, [_Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13];
ld.param.u32 %r407, [_Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10];
mov.u32 %r406, %tid.y;
mul.wide.u32 %rd107, %r2, 4;
add.s64 %rd108, %rd117, %rd107;
ld.local.f32 %f57, [%rd108];
mov.u32 %r401, %ctaid.y;
mad.lo.s32 %r403, %r406, %r407, %r15;
mad.lo.s32 %r404, %r401, %r408, %r403;
cvta.to.global.u64 %rd109, %rd112;
mul.wide.u32 %rd110, %r404, 4;
add.s64 %rd111, %rd109, %rd110;
st.global.f32 [%rd111], %f57;

$L__BB141_7:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<320>;
.reg .b16 %rs<73>;
.reg .f32 %f<1077>;
.reg .b32 %r<1863>;
.reg .b64 %rd<251>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate[1536];

ld.param.u64 %rd93, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd94, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd95, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u32 %r70, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r56, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r57, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r61, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r62, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r63, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r67, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r68, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r69, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r65, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r64, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r59, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r58, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r55, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r54, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r53, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
ld.param.u64 %rd98, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+16];
ld.param.u64 %rd97, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+8];
ld.param.u64 %rd96, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3];
mov.u32 %r1, %tid.y;
shl.b32 %r71, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r71, %r2;
mov.u32 %r72, %ctaid.x;
shl.b32 %r4, %r72, 2;
shr.u32 %r5, %r70, 8;
mov.u32 %r6, %ctaid.y;
setp.eq.s64 %p1, %rd95, 0;
@%p1 bra $L__BB142_2;

cvta.to.global.u64 %rd102, %rd95;
mul.wide.u32 %rd103, %r6, 4;
add.s64 %rd104, %rd102, %rd103;
ld.global.nc.u32 %r1859, [%rd104];
mul.hi.u32 %r73, %r6, %r53;
add.s32 %r74, %r73, %r6;
shr.u32 %r75, %r74, %r54;
mul.lo.s32 %r76, %r75, %r55;
sub.s32 %r1860, %r6, %r76;
bra.uni $L__BB142_3;

$L__BB142_2:
mul.hi.u32 %r77, %r6, %r58;
add.s32 %r78, %r77, %r6;
shr.u32 %r1859, %r78, %r59;
mov.u32 %r1860, %r6;

$L__BB142_3:
mov.u32 %r12, %ctaid.z;
selp.b32 %r13, %r6, %r1859, %p1;
setp.ne.s32 %p3, %r1, 0;
setp.gt.u32 %p4, %r2, 3;
or.pred %p5, %p4, %p3;
add.s32 %r79, %r4, %r2;
setp.ge.u32 %p6, %r79, %r57;
mov.f32 %f27, 0f00000000;
or.pred %p7, %p5, %p6;
mov.f32 %f4, %f27;
mov.f32 %f5, %f27;
@%p7 bra $L__BB142_8;

setp.eq.s64 %p8, %rd96, 0;
mov.f32 %f4, 0f00000000;
mov.f32 %f5, %f4;
@%p8 bra $L__BB142_6;

cvta.to.global.u64 %rd105, %rd96;
mul.lo.s32 %r80, %r12, %r69;
cvt.u64.u32 %rd106, %r80;
mul.lo.s32 %r81, %r13, %r63;
cvt.u64.u32 %rd107, %r81;
cvt.u64.u32 %rd108, %r2;
cvt.s64.s32 %rd109, %r4;
add.s64 %rd110, %rd109, %rd108;
add.s64 %rd111, %rd110, %rd106;
add.s64 %rd112, %rd111, %rd107;
shl.b64 %rd113, %rd112, 2;
add.s64 %rd114, %rd105, %rd113;
ld.global.f32 %f5, [%rd114];

$L__BB142_6:
setp.eq.s64 %p9, %rd98, 0;
setp.eq.s64 %p10, %rd97, 0;
or.pred %p11, %p9, %p10;
@%p11 bra $L__BB142_8;

cvta.to.global.u64 %rd115, %rd98;
mul.lo.s32 %r82, %r12, %r69;
cvt.u64.u32 %rd116, %r82;
mul.lo.s32 %r83, %r13, %r63;
cvt.u64.u32 %rd117, %r83;
cvt.u64.u32 %rd118, %r2;
cvt.s64.s32 %rd119, %r4;
add.s64 %rd120, %rd119, %rd118;
add.s64 %rd121, %rd120, %rd116;
add.s64 %rd122, %rd121, %rd117;
shl.b64 %rd123, %rd122, 2;
add.s64 %rd124, %rd115, %rd123;
ld.global.f32 %f4, [%rd124];

$L__BB142_8:
mul.lo.s32 %r84, %r12, %r68;
cvt.u64.u32 %rd125, %r84;
mul.lo.s32 %r85, %r1860, %r62;
cvt.u64.u32 %rd126, %r85;
add.s64 %rd4, %rd126, %rd125;
mul.hi.u32 %r86, %r12, %r64;
add.s32 %r87, %r86, %r12;
shr.u32 %r14, %r87, %r65;
shr.s32 %r88, %r3, 31;
shr.u32 %r89, %r88, 28;
add.s32 %r90, %r3, %r89;
shr.s32 %r1862, %r90, 4;
setp.ge.s32 %p12, %r1862, %r5;
mov.f32 %f28, %f27;
mov.f32 %f29, %f27;
mov.f32 %f30, %f27;
mov.f32 %f31, %f27;
mov.f32 %f32, %f27;
mov.f32 %f33, %f27;
mov.f32 %f34, %f27;
@%p12 bra $L__BB142_19;

mul.lo.s32 %r91, %r14, %r67;
mul.lo.s32 %r92, %r1859, %r61;
and.b32 %r96, %r90, -16;
sub.s32 %r97, %r3, %r96;
shr.s32 %r98, %r97, 31;
shr.u32 %r99, %r98, 29;
add.s32 %r100, %r97, %r99;
shr.s32 %r101, %r100, 3;
shl.b32 %r102, %r101, 2;
and.b32 %r103, %r100, -8;
sub.s32 %r104, %r97, %r103;
shl.b32 %r105, %r97, 1;
shl.b32 %r106, %r104, 1;
cvt.s64.s32 %rd5, %r104;
shr.s32 %r107, %r104, 31;
shr.u32 %r108, %r107, 30;
add.s32 %r109, %r104, %r108;
shr.s32 %r110, %r109, 2;
cvt.s64.s32 %rd6, %r102;
add.s32 %r111, %r103, %r110;
shr.s32 %r112, %r111, 31;
shr.u32 %r113, %r112, 29;
add.s32 %r114, %r111, %r113;
and.b32 %r115, %r114, -8;
sub.s32 %r116, %r111, %r115;
shr.s32 %r117, %r114, 3;
shl.b32 %r16, %r117, 2;
cvt.s64.s32 %rd127, %r116;
shr.u32 %r118, %r112, 30;
add.s32 %r119, %r111, %r118;
and.b32 %r120, %r119, -4;
sub.s32 %r121, %r111, %r120;
shr.s32 %r122, %r119, 2;
shl.b32 %r17, %r122, 1;
add.s32 %r123, %r121, 8;
cvt.s64.s32 %rd128, %r123;
add.s32 %r124, %r111, 2;
shr.s32 %r125, %r124, 31;
shr.u32 %r126, %r125, 29;
add.s32 %r127, %r124, %r126;
and.b32 %r128, %r127, -8;
sub.s32 %r129, %r124, %r128;
shr.s32 %r130, %r127, 3;
shl.b32 %r18, %r130, 2;
cvt.s64.s32 %rd129, %r129;
shr.u32 %r131, %r125, 30;
add.s32 %r132, %r124, %r131;
and.b32 %r133, %r132, -4;
sub.s32 %r134, %r124, %r133;
shr.s32 %r135, %r132, 2;
shl.b32 %r19, %r135, 1;
add.s32 %r136, %r134, 8;
cvt.s64.s32 %rd130, %r136;
add.s32 %r137, %r111, 4;
shr.s32 %r138, %r137, 31;
shr.u32 %r139, %r138, 29;
add.s32 %r140, %r137, %r139;
and.b32 %r141, %r140, -8;
sub.s32 %r142, %r137, %r141;
shr.s32 %r143, %r140, 3;
shl.b32 %r20, %r143, 2;
cvt.s64.s32 %rd131, %r142;
shr.u32 %r144, %r138, 30;
add.s32 %r145, %r137, %r144;
and.b32 %r146, %r145, -4;
sub.s32 %r147, %r137, %r146;
shr.s32 %r148, %r145, 2;
shl.b32 %r21, %r148, 1;
add.s32 %r149, %r147, 8;
cvt.s64.s32 %rd132, %r149;
add.s32 %r150, %r111, 6;
shr.s32 %r151, %r150, 31;
shr.u32 %r152, %r151, 29;
add.s32 %r153, %r150, %r152;
and.b32 %r154, %r153, -8;
sub.s32 %r155, %r150, %r154;
shr.s32 %r156, %r153, 3;
shl.b32 %r22, %r156, 2;
cvt.s64.s32 %rd133, %r155;
shr.u32 %r157, %r151, 30;
add.s32 %r158, %r150, %r157;
and.b32 %r159, %r158, -4;
sub.s32 %r160, %r150, %r159;
shr.s32 %r161, %r158, 2;
shl.b32 %r23, %r161, 1;
add.s32 %r162, %r160, 8;
cvt.s64.s32 %rd134, %r162;
shl.b32 %r1861, %r1862, 3;
add.s32 %r163, %r1862, %r91;
add.s32 %r164, %r163, %r92;
add.s32 %r165, %r4, 1;
mad.lo.s32 %r166, %r56, %r165, %r164;
mul.wide.s32 %rd7, %r166, 110;
add.s64 %rd135, %rd7, %rd134;
add.s64 %rd8, %rd135, 96;
add.s64 %rd136, %rd7, %rd133;
add.s64 %rd9, %rd136, 96;
add.s64 %rd137, %rd7, %rd132;
add.s64 %rd10, %rd137, 96;
add.s64 %rd138, %rd7, %rd131;
add.s64 %rd11, %rd138, 96;
add.s64 %rd139, %rd7, %rd130;
add.s64 %rd12, %rd139, 96;
add.s64 %rd140, %rd7, %rd129;
add.s64 %rd13, %rd140, 96;
add.s64 %rd141, %rd7, %rd128;
add.s64 %rd14, %rd141, 96;
add.s64 %rd142, %rd7, %rd127;
add.s64 %rd15, %rd142, 96;
mad.lo.s32 %r167, %r4, %r56, %r1862;
add.s32 %r168, %r167, %r91;
add.s32 %r169, %r168, %r92;
mul.wide.s32 %rd16, %r169, 110;
mul.wide.s32 %rd143, %r105, 2;
add.s64 %rd144, %rd16, %rd143;
add.s64 %rd17, %rd144, 34;
mul.wide.s32 %rd145, %r106, 2;
add.s64 %rd18, %rd7, %rd145;
add.s64 %rd146, %rd16, %rd145;
add.s64 %rd19, %rd146, 2;
add.s64 %rd147, %rd7, %rd143;
add.s64 %rd20, %rd147, 32;
add.s64 %rd148, %rd16, %rd127;
add.s64 %rd21, %rd148, 96;
add.s32 %r170, %r4, 2;
mad.lo.s32 %r171, %r56, %r170, %r164;
mul.wide.s32 %rd22, %r171, 110;
add.s64 %rd149, %rd22, %rd134;
add.s64 %rd23, %rd149, 96;
add.s64 %rd150, %rd22, %rd133;
add.s64 %rd24, %rd150, 96;
add.s64 %rd151, %rd22, %rd132;
add.s64 %rd25, %rd151, 96;
add.s64 %rd152, %rd22, %rd131;
add.s64 %rd26, %rd152, 96;
add.s64 %rd153, %rd22, %rd130;
add.s64 %rd27, %rd153, 96;
add.s64 %rd154, %rd22, %rd129;
add.s64 %rd28, %rd154, 96;
add.s64 %rd155, %rd22, %rd128;
add.s64 %rd29, %rd155, 96;
add.s64 %rd156, %rd22, %rd127;
add.s64 %rd30, %rd156, 96;
add.s64 %rd157, %rd16, %rd128;
add.s64 %rd31, %rd157, 96;
add.s64 %rd32, %rd22, %rd145;
add.s64 %rd158, %rd22, %rd143;
add.s64 %rd33, %rd158, 32;
add.s32 %r172, %r4, 3;
mad.lo.s32 %r173, %r56, %r172, %r164;
mul.wide.s32 %rd34, %r173, 110;
add.s64 %rd159, %rd34, %rd134;
add.s64 %rd35, %rd159, 96;
add.s64 %rd160, %rd34, %rd133;
add.s64 %rd36, %rd160, 96;
add.s64 %rd161, %rd34, %rd132;
add.s64 %rd37, %rd161, 96;
add.s64 %rd162, %rd34, %rd131;
add.s64 %rd38, %rd162, 96;
add.s64 %rd163, %rd34, %rd130;
add.s64 %rd39, %rd163, 96;
add.s64 %rd164, %rd34, %rd129;
add.s64 %rd40, %rd164, 96;
add.s64 %rd165, %rd34, %rd128;
add.s64 %rd41, %rd165, 96;
add.s64 %rd166, %rd34, %rd127;
add.s64 %rd42, %rd166, 96;
add.s64 %rd43, %rd34, %rd145;
add.s64 %rd167, %rd34, %rd143;
add.s64 %rd44, %rd167, 32;
add.s64 %rd168, %rd16, %rd129;
add.s64 %rd45, %rd168, 96;
add.s64 %rd169, %rd16, %rd130;
add.s64 %rd46, %rd169, 96;
add.s64 %rd170, %rd16, %rd134;
add.s64 %rd47, %rd170, 96;
add.s64 %rd171, %rd16, %rd133;
add.s64 %rd48, %rd171, 96;
add.s64 %rd172, %rd16, %rd132;
add.s64 %rd49, %rd172, 96;
add.s64 %rd173, %rd16, %rd131;
add.s64 %rd50, %rd173, 96;
mul.wide.s32 %rd51, %r56, 110;
add.s64 %rd52, %rd51, 2;
cvta.to.global.u64 %rd250, %rd97;
cvta.to.global.u64 %rd249, %rd93;
cvta.to.global.u64 %rd55, %rd94;
mov.f32 %f27, 0f00000000;
shl.b64 %rd181, %rd5, 2;
mov.f32 %f28, %f27;
mov.f32 %f29, %f27;
mov.f32 %f30, %f27;
mov.f32 %f31, %f27;
mov.f32 %f32, %f27;
mov.f32 %f33, %f27;
mov.f32 %f34, %f27;

$L__BB142_10:
cvt.u64.u32 %rd174, %r1861;
add.s64 %rd175, %rd4, %rd174;
add.s64 %rd176, %rd249, %rd16;
ld.global.nc.u16 %rs1, [%rd176+108];

{ cvt.f32.f16 %f101, %rs1;}


add.s64 %rd177, %rd249, %rd19;
ld.global.nc.u16 %rs2, [%rd177+-2];
ld.global.nc.u16 %rs3, [%rd177];
mov.b32 %r210, {%rs2, %rs3};
not.b32 %r211, %r210;
add.s64 %rd60, %rd249, %rd17;
ld.global.nc.u16 %rs4, [%rd60];
ld.global.nc.u16 %rs5, [%rd60+-2];
add.s64 %rd178, %rd175, %rd6;
mul.lo.s64 %rd179, %rd178, 36;
add.s64 %rd180, %rd55, %rd179;
add.s64 %rd182, %rd180, %rd181;
ld.global.nc.u32 %r27, [%rd182+4];
ld.global.nc.u32 %r28, [%rd180];

{.reg .f16 low,high;
mov.b32 {low,high},%r28;
cvt.f32.f16 %f102, low;}


ld.global.nc.u32 %r29, [%rd182+40];
ld.global.nc.u32 %r30, [%rd180+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r30;
cvt.f32.f16 %f103, low;}


ld.global.nc.u32 %r31, [%rd182+76];
ld.global.nc.u32 %r32, [%rd180+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r32;
cvt.f32.f16 %f104, low;}


ld.global.nc.u32 %r33, [%rd182+112];
ld.global.nc.u32 %r34, [%rd180+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r34;
cvt.f32.f16 %f105, low;}


mov.b32 %r212, {%rs5, %rs4};
cvt.u32.u64 %r213, %rd6;
shr.s32 %r214, %r211, %r213;
add.s64 %rd61, %rd249, %rd21;
ld.global.nc.u8 %rs6, [%rd61];
cvt.u32.u16 %r215, %rs6;
and.b32 %r216, %r215, 255;
shr.u32 %r217, %r216, %r16;
and.b32 %r218, %r217, 15;
add.s64 %rd62, %rd249, %rd31;
ld.global.nc.u8 %rs7, [%rd62];
cvt.u32.u16 %r219, %rs7;
and.b32 %r220, %r219, 255;
shr.u32 %r221, %r220, %r17;
and.b32 %r222, %r221, 3;
bfi.b32 %r223, %r222, %r218, 4, 2;
add.s32 %r224, %r223, -32;
and.b32 %r180, %r212, 50529027;
shl.b32 %r225, %r214, 2;
and.b32 %r226, %r225, 67372036;
or.b32 %r227, %r180, -2139062144;
sub.s32 %r228, %r227, %r226;
xor.b32 %r229, %r228, -2139062144;
or.b32 %r230, %r180, %r226;
xor.b32 %r231, %r228, %r180;
and.b32 %r178, %r231, %r230;

prmt.b32 %r178, %r178, 0, 0xba98;


prmt.b32 %r180, %r180, 0, 0xba98;

xor.b32 %r232, %r180, 2139062143;
not.b32 %r233, %r178;
and.b32 %r234, %r229, %r233;
and.b32 %r235, %r232, %r178;
or.b32 %r183, %r235, %r234;
mov.u32 %r209, 0;

dp4a.s32.s32 %r182, %r183, %r27, %r209;

mul.lo.s32 %r236, %r224, %r182;
cvt.rn.f32.s32 %f106, %r236;
fma.rn.ftz.f32 %f107, %f102, %f106, 0f00000000;
add.s64 %rd63, %rd249, %rd45;
ld.global.nc.u8 %rs8, [%rd63];
cvt.u32.u16 %r237, %rs8;
and.b32 %r238, %r237, 255;
shr.u32 %r239, %r238, %r18;
and.b32 %r240, %r239, 15;
add.s64 %rd64, %rd249, %rd46;
ld.global.nc.u8 %rs9, [%rd64];
cvt.u32.u16 %r241, %rs9;
and.b32 %r242, %r241, 255;
shr.u32 %r243, %r242, %r19;
and.b32 %r244, %r243, 3;
bfi.b32 %r245, %r244, %r240, 4, 2;
add.s32 %r246, %r245, -32;
shr.u32 %r247, %r212, 2;
and.b32 %r188, %r247, 50529027;
shl.b32 %r248, %r214, 1;
and.b32 %r249, %r248, 67372036;
or.b32 %r250, %r188, -2139062144;
sub.s32 %r251, %r250, %r249;
xor.b32 %r252, %r251, -2139062144;
or.b32 %r253, %r188, %r249;
xor.b32 %r254, %r251, %r188;
and.b32 %r186, %r254, %r253;

prmt.b32 %r186, %r186, 0, 0xba98;


prmt.b32 %r188, %r188, 0, 0xba98;

xor.b32 %r255, %r188, 2139062143;
not.b32 %r256, %r186;
and.b32 %r257, %r252, %r256;
and.b32 %r258, %r255, %r186;
or.b32 %r191, %r258, %r257;

dp4a.s32.s32 %r190, %r191, %r29, %r209;

mul.lo.s32 %r259, %r246, %r190;
cvt.rn.f32.s32 %f108, %r259;
fma.rn.ftz.f32 %f109, %f103, %f108, %f107;
add.s64 %rd65, %rd249, %rd50;
ld.global.nc.u8 %rs10, [%rd65];
cvt.u32.u16 %r260, %rs10;
and.b32 %r261, %r260, 255;
shr.u32 %r262, %r261, %r20;
and.b32 %r263, %r262, 15;
add.s64 %rd66, %rd249, %rd49;
ld.global.nc.u8 %rs11, [%rd66];
cvt.u32.u16 %r264, %rs11;
and.b32 %r265, %r264, 255;
shr.u32 %r266, %r265, %r21;
and.b32 %r267, %r266, 3;
bfi.b32 %r268, %r267, %r263, 4, 2;
add.s32 %r269, %r268, -32;
shr.u32 %r270, %r212, 4;
and.b32 %r196, %r270, 50529027;
and.b32 %r271, %r214, 67372036;
or.b32 %r272, %r196, -2139062144;
sub.s32 %r273, %r272, %r271;
xor.b32 %r274, %r273, -2139062144;
or.b32 %r275, %r196, %r271;
xor.b32 %r276, %r273, %r196;
and.b32 %r194, %r276, %r275;

prmt.b32 %r194, %r194, 0, 0xba98;


prmt.b32 %r196, %r196, 0, 0xba98;

xor.b32 %r277, %r196, 2139062143;
not.b32 %r278, %r194;
and.b32 %r279, %r274, %r278;
and.b32 %r280, %r277, %r194;
or.b32 %r199, %r280, %r279;

dp4a.s32.s32 %r198, %r199, %r31, %r209;

mul.lo.s32 %r281, %r269, %r198;
cvt.rn.f32.s32 %f110, %r281;
fma.rn.ftz.f32 %f111, %f104, %f110, %f109;
add.s64 %rd67, %rd249, %rd48;
ld.global.nc.u8 %rs12, [%rd67];
cvt.u32.u16 %r282, %rs12;
and.b32 %r283, %r282, 255;
shr.u32 %r284, %r283, %r22;
and.b32 %r285, %r284, 15;
add.s64 %rd68, %rd249, %rd47;
ld.global.nc.u8 %rs13, [%rd68];
cvt.u32.u16 %r286, %rs13;
and.b32 %r287, %r286, 255;
shr.u32 %r288, %r287, %r23;
and.b32 %r289, %r288, 3;
bfi.b32 %r290, %r289, %r285, 4, 2;
add.s32 %r291, %r290, -32;
shr.u32 %r292, %r212, 6;
and.b32 %r204, %r292, 50529027;
shr.u32 %r293, %r214, 1;
and.b32 %r294, %r293, 67372036;
or.b32 %r295, %r204, -2139062144;
sub.s32 %r296, %r295, %r294;
xor.b32 %r297, %r296, -2139062144;
or.b32 %r298, %r204, %r294;
xor.b32 %r299, %r296, %r204;
and.b32 %r202, %r299, %r298;

prmt.b32 %r202, %r202, 0, 0xba98;


prmt.b32 %r204, %r204, 0, 0xba98;

xor.b32 %r300, %r204, 2139062143;
not.b32 %r301, %r202;
and.b32 %r302, %r297, %r301;
and.b32 %r303, %r300, %r202;
or.b32 %r207, %r303, %r302;

dp4a.s32.s32 %r206, %r207, %r33, %r209;

mul.lo.s32 %r304, %r291, %r206;
cvt.rn.f32.s32 %f112, %r304;
fma.rn.ftz.f32 %f113, %f105, %f112, %f111;
fma.rn.ftz.f32 %f30, %f101, %f113, %f30;
setp.eq.s64 %p13, %rd97, 0;
@%p13 bra $L__BB142_12;

cvt.u32.u64 %r1856, %rd6;
add.s64 %rd183, %rd250, %rd16;
ld.global.u16 %rs14, [%rd183+108];

{ cvt.f32.f16 %f114, %rs14;}


add.s64 %rd184, %rd250, %rd17;
add.s64 %rd185, %rd250, %rd19;
ld.global.u16 %rs15, [%rd185+-2];
ld.global.u16 %rs16, [%rd185];
mov.b32 %r342, {%rs15, %rs16};
not.b32 %r343, %r342;
ld.global.u16 %rs17, [%rd184];
ld.global.u16 %rs18, [%rd184+-2];

{.reg .f16 low,high;
mov.b32 {low,high},%r28;
cvt.f32.f16 %f115, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r30;
cvt.f32.f16 %f116, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r32;
cvt.f32.f16 %f117, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r34;
cvt.f32.f16 %f118, low;}


mov.b32 %r344, {%rs18, %rs17};
shr.s32 %r345, %r343, %r1856;
add.s64 %rd186, %rd250, %rd21;
ld.global.u8 %r346, [%rd186];
shr.u32 %r347, %r346, %r16;
and.b32 %r348, %r347, 15;
add.s64 %rd187, %rd250, %rd31;
ld.global.u8 %r349, [%rd187];
shr.u32 %r350, %r349, %r17;
and.b32 %r351, %r350, 3;
bfi.b32 %r352, %r351, %r348, 4, 2;
add.s32 %r353, %r352, -32;
and.b32 %r311, %r344, 50529027;
shl.b32 %r354, %r345, 2;
and.b32 %r355, %r354, 67372036;
or.b32 %r356, %r311, -2139062144;
sub.s32 %r357, %r356, %r355;
xor.b32 %r358, %r357, -2139062144;
or.b32 %r359, %r311, %r355;
xor.b32 %r360, %r357, %r311;
and.b32 %r309, %r360, %r359;

prmt.b32 %r309, %r309, 0, 0xba98;


prmt.b32 %r311, %r311, 0, 0xba98;

xor.b32 %r361, %r311, 2139062143;
not.b32 %r362, %r309;
and.b32 %r363, %r358, %r362;
and.b32 %r364, %r361, %r309;
or.b32 %r314, %r364, %r363;

dp4a.s32.s32 %r313, %r314, %r27, %r209;

mul.lo.s32 %r365, %r353, %r313;
cvt.rn.f32.s32 %f119, %r365;
fma.rn.ftz.f32 %f120, %f115, %f119, 0f00000000;
add.s64 %rd188, %rd250, %rd45;
ld.global.u8 %r366, [%rd188];
shr.u32 %r367, %r366, %r18;
and.b32 %r368, %r367, 15;
add.s64 %rd189, %rd250, %rd46;
ld.global.u8 %r369, [%rd189];
shr.u32 %r370, %r369, %r19;
and.b32 %r371, %r370, 3;
bfi.b32 %r372, %r371, %r368, 4, 2;
add.s32 %r373, %r372, -32;
shr.u32 %r374, %r344, 2;
and.b32 %r319, %r374, 50529027;
shl.b32 %r375, %r345, 1;
and.b32 %r376, %r375, 67372036;
or.b32 %r377, %r319, -2139062144;
sub.s32 %r378, %r377, %r376;
xor.b32 %r379, %r378, -2139062144;
or.b32 %r380, %r319, %r376;
xor.b32 %r381, %r378, %r319;
and.b32 %r317, %r381, %r380;

prmt.b32 %r317, %r317, 0, 0xba98;


prmt.b32 %r319, %r319, 0, 0xba98;

xor.b32 %r382, %r319, 2139062143;
not.b32 %r383, %r317;
and.b32 %r384, %r379, %r383;
and.b32 %r385, %r382, %r317;
or.b32 %r322, %r385, %r384;

dp4a.s32.s32 %r321, %r322, %r29, %r209;

mul.lo.s32 %r386, %r373, %r321;
cvt.rn.f32.s32 %f121, %r386;
fma.rn.ftz.f32 %f122, %f116, %f121, %f120;
add.s64 %rd190, %rd250, %rd50;
ld.global.u8 %r387, [%rd190];
shr.u32 %r388, %r387, %r20;
and.b32 %r389, %r388, 15;
add.s64 %rd191, %rd250, %rd49;
ld.global.u8 %r390, [%rd191];
shr.u32 %r391, %r390, %r21;
and.b32 %r392, %r391, 3;
bfi.b32 %r393, %r392, %r389, 4, 2;
add.s32 %r394, %r393, -32;
shr.u32 %r395, %r344, 4;
and.b32 %r327, %r395, 50529027;
and.b32 %r396, %r345, 67372036;
or.b32 %r397, %r327, -2139062144;
sub.s32 %r398, %r397, %r396;
xor.b32 %r399, %r398, -2139062144;
or.b32 %r400, %r327, %r396;
xor.b32 %r401, %r398, %r327;
and.b32 %r325, %r401, %r400;

prmt.b32 %r325, %r325, 0, 0xba98;


prmt.b32 %r327, %r327, 0, 0xba98;

xor.b32 %r402, %r327, 2139062143;
not.b32 %r403, %r325;
and.b32 %r404, %r399, %r403;
and.b32 %r405, %r402, %r325;
or.b32 %r330, %r405, %r404;

dp4a.s32.s32 %r329, %r330, %r31, %r209;

mul.lo.s32 %r406, %r394, %r329;
cvt.rn.f32.s32 %f123, %r406;
fma.rn.ftz.f32 %f124, %f117, %f123, %f122;
add.s64 %rd192, %rd250, %rd48;
ld.global.u8 %r407, [%rd192];
shr.u32 %r408, %r407, %r22;
and.b32 %r409, %r408, 15;
add.s64 %rd193, %rd250, %rd47;
ld.global.u8 %r410, [%rd193];
shr.u32 %r411, %r410, %r23;
and.b32 %r412, %r411, 3;
bfi.b32 %r413, %r412, %r409, 4, 2;
add.s32 %r414, %r413, -32;
shr.u32 %r415, %r344, 6;
and.b32 %r335, %r415, 50529027;
shr.u32 %r416, %r345, 1;
and.b32 %r417, %r416, 67372036;
or.b32 %r418, %r335, -2139062144;
sub.s32 %r419, %r418, %r417;
xor.b32 %r420, %r419, -2139062144;
or.b32 %r421, %r335, %r417;
xor.b32 %r422, %r419, %r335;
and.b32 %r333, %r422, %r421;

prmt.b32 %r333, %r333, 0, 0xba98;


prmt.b32 %r335, %r335, 0, 0xba98;

xor.b32 %r423, %r335, 2139062143;
not.b32 %r424, %r333;
and.b32 %r425, %r420, %r424;
and.b32 %r426, %r423, %r333;
or.b32 %r338, %r426, %r425;

dp4a.s32.s32 %r337, %r338, %r33, %r209;

mul.lo.s32 %r427, %r414, %r337;
cvt.rn.f32.s32 %f125, %r427;
fma.rn.ftz.f32 %f126, %f118, %f125, %f124;
fma.rn.ftz.f32 %f34, %f114, %f126, %f34;

$L__BB142_12:
cvt.u32.u64 %r1853, %rd6;
add.s64 %rd247, %rd249, %rd19;
add.s64 %rd246, %rd247, -2;
add.s64 %rd245, %rd249, %rd16;
add.s64 %rd244, %rd245, 108;
add.s64 %rd69, %rd244, %rd51;
ld.global.nc.u16 %rs19, [%rd69];

{ cvt.f32.f16 %f127, %rs19;}


add.s64 %rd70, %rd246, %rd51;
ld.global.nc.u16 %rs20, [%rd70];
ld.global.nc.u16 %rs21, [%rd70+2];
mov.b32 %r464, {%rs20, %rs21};
not.b32 %r465, %r464;
add.s64 %rd194, %rd60, %rd51;
ld.global.nc.u16 %rs22, [%rd194];
ld.global.nc.u16 %rs23, [%rd194+-2];

{.reg .f16 low,high;
mov.b32 {low,high},%r28;
cvt.f32.f16 %f128, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r30;
cvt.f32.f16 %f129, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r32;
cvt.f32.f16 %f130, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r34;
cvt.f32.f16 %f131, low;}


mov.b32 %r466, {%rs23, %rs22};
shr.s32 %r468, %r465, %r1853;
add.s64 %rd71, %rd61, %rd51;
ld.global.nc.u8 %rs24, [%rd71];
cvt.u32.u16 %r469, %rs24;
and.b32 %r470, %r469, 255;
shr.u32 %r471, %r470, %r16;
and.b32 %r472, %r471, 15;
add.s64 %rd72, %rd62, %rd51;
ld.global.nc.u8 %rs25, [%rd72];
cvt.u32.u16 %r473, %rs25;
and.b32 %r474, %r473, 255;
shr.u32 %r475, %r474, %r17;
and.b32 %r476, %r475, 3;
bfi.b32 %r477, %r476, %r472, 4, 2;
add.s32 %r478, %r477, -32;
and.b32 %r434, %r466, 50529027;
shl.b32 %r479, %r468, 2;
and.b32 %r480, %r479, 67372036;
or.b32 %r481, %r434, -2139062144;
sub.s32 %r482, %r481, %r480;
xor.b32 %r483, %r482, -2139062144;
or.b32 %r484, %r434, %r480;
xor.b32 %r485, %r482, %r434;
and.b32 %r432, %r485, %r484;

prmt.b32 %r432, %r432, 0, 0xba98;


prmt.b32 %r434, %r434, 0, 0xba98;

xor.b32 %r486, %r434, 2139062143;
not.b32 %r487, %r432;
and.b32 %r488, %r483, %r487;
and.b32 %r489, %r486, %r432;
or.b32 %r437, %r489, %r488;

dp4a.s32.s32 %r436, %r437, %r27, %r209;

mul.lo.s32 %r490, %r478, %r436;
cvt.rn.f32.s32 %f132, %r490;
fma.rn.ftz.f32 %f133, %f128, %f132, 0f00000000;
add.s64 %rd73, %rd63, %rd51;
ld.global.nc.u8 %rs26, [%rd73];
cvt.u32.u16 %r491, %rs26;
and.b32 %r492, %r491, 255;
shr.u32 %r493, %r492, %r18;
and.b32 %r494, %r493, 15;
add.s64 %rd74, %rd64, %rd51;
ld.global.nc.u8 %rs27, [%rd74];
cvt.u32.u16 %r495, %rs27;
and.b32 %r496, %r495, 255;
shr.u32 %r497, %r496, %r19;
and.b32 %r498, %r497, 3;
bfi.b32 %r499, %r498, %r494, 4, 2;
add.s32 %r500, %r499, -32;
shr.u32 %r501, %r466, 2;
and.b32 %r442, %r501, 50529027;
shl.b32 %r502, %r468, 1;
and.b32 %r503, %r502, 67372036;
or.b32 %r504, %r442, -2139062144;
sub.s32 %r505, %r504, %r503;
xor.b32 %r506, %r505, -2139062144;
or.b32 %r507, %r442, %r503;
xor.b32 %r508, %r505, %r442;
and.b32 %r440, %r508, %r507;

prmt.b32 %r440, %r440, 0, 0xba98;


prmt.b32 %r442, %r442, 0, 0xba98;

xor.b32 %r509, %r442, 2139062143;
not.b32 %r510, %r440;
and.b32 %r511, %r506, %r510;
and.b32 %r512, %r509, %r440;
or.b32 %r445, %r512, %r511;

dp4a.s32.s32 %r444, %r445, %r29, %r209;

mul.lo.s32 %r513, %r500, %r444;
cvt.rn.f32.s32 %f134, %r513;
fma.rn.ftz.f32 %f135, %f129, %f134, %f133;
add.s64 %rd75, %rd65, %rd51;
ld.global.nc.u8 %rs28, [%rd75];
cvt.u32.u16 %r514, %rs28;
and.b32 %r515, %r514, 255;
shr.u32 %r516, %r515, %r20;
and.b32 %r517, %r516, 15;
add.s64 %rd76, %rd66, %rd51;
ld.global.nc.u8 %rs29, [%rd76];
cvt.u32.u16 %r518, %rs29;
and.b32 %r519, %r518, 255;
shr.u32 %r520, %r519, %r21;
and.b32 %r521, %r520, 3;
bfi.b32 %r522, %r521, %r517, 4, 2;
add.s32 %r523, %r522, -32;
shr.u32 %r524, %r466, 4;
and.b32 %r450, %r524, 50529027;
and.b32 %r525, %r468, 67372036;
or.b32 %r526, %r450, -2139062144;
sub.s32 %r527, %r526, %r525;
xor.b32 %r528, %r527, -2139062144;
or.b32 %r529, %r450, %r525;
xor.b32 %r530, %r527, %r450;
and.b32 %r448, %r530, %r529;

prmt.b32 %r448, %r448, 0, 0xba98;


prmt.b32 %r450, %r450, 0, 0xba98;

xor.b32 %r531, %r450, 2139062143;
not.b32 %r532, %r448;
and.b32 %r533, %r528, %r532;
and.b32 %r534, %r531, %r448;
or.b32 %r453, %r534, %r533;

dp4a.s32.s32 %r452, %r453, %r31, %r209;

mul.lo.s32 %r535, %r523, %r452;
cvt.rn.f32.s32 %f136, %r535;
fma.rn.ftz.f32 %f137, %f130, %f136, %f135;
add.s64 %rd77, %rd67, %rd51;
ld.global.nc.u8 %rs30, [%rd77];
cvt.u32.u16 %r536, %rs30;
and.b32 %r537, %r536, 255;
shr.u32 %r538, %r537, %r22;
and.b32 %r539, %r538, 15;
add.s64 %rd78, %rd68, %rd51;
ld.global.nc.u8 %rs31, [%rd78];
cvt.u32.u16 %r540, %rs31;
and.b32 %r541, %r540, 255;
shr.u32 %r542, %r541, %r23;
and.b32 %r543, %r542, 3;
bfi.b32 %r544, %r543, %r539, 4, 2;
add.s32 %r545, %r544, -32;
shr.u32 %r546, %r466, 6;
and.b32 %r458, %r546, 50529027;
shr.u32 %r547, %r468, 1;
and.b32 %r548, %r547, 67372036;
or.b32 %r549, %r458, -2139062144;
sub.s32 %r550, %r549, %r548;
xor.b32 %r551, %r550, -2139062144;
or.b32 %r552, %r458, %r548;
xor.b32 %r553, %r550, %r458;
and.b32 %r456, %r553, %r552;

prmt.b32 %r456, %r456, 0, 0xba98;


prmt.b32 %r458, %r458, 0, 0xba98;

xor.b32 %r554, %r458, 2139062143;
not.b32 %r555, %r456;
and.b32 %r556, %r551, %r555;
and.b32 %r557, %r554, %r456;
or.b32 %r461, %r557, %r556;

dp4a.s32.s32 %r460, %r461, %r33, %r209;

mul.lo.s32 %r558, %r545, %r460;
cvt.rn.f32.s32 %f138, %r558;
fma.rn.ftz.f32 %f139, %f131, %f138, %f137;
fma.rn.ftz.f32 %f29, %f127, %f139, %f29;
@%p13 bra $L__BB142_14;

cvt.u32.u64 %r1855, %rd6;
add.s64 %rd195, %rd250, %rd7;
ld.global.u16 %rs32, [%rd195+108];

{ cvt.f32.f16 %f140, %rs32;}


add.s64 %rd196, %rd250, %rd20;
add.s64 %rd197, %rd250, %rd18;
ld.global.u16 %rs33, [%rd197];
ld.global.u16 %rs34, [%rd197+2];
mov.b32 %r596, {%rs33, %rs34};
not.b32 %r597, %r596;
ld.global.u16 %rs35, [%rd196+2];
ld.global.u16 %rs36, [%rd196];

{.reg .f16 low,high;
mov.b32 {low,high},%r28;
cvt.f32.f16 %f141, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r30;
cvt.f32.f16 %f142, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r32;
cvt.f32.f16 %f143, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r34;
cvt.f32.f16 %f144, low;}


mov.b32 %r598, {%rs36, %rs35};
shr.s32 %r599, %r597, %r1855;
add.s64 %rd198, %rd250, %rd15;
ld.global.u8 %r600, [%rd198];
shr.u32 %r601, %r600, %r16;
and.b32 %r602, %r601, 15;
add.s64 %rd199, %rd250, %rd14;
ld.global.u8 %r603, [%rd199];
shr.u32 %r604, %r603, %r17;
and.b32 %r605, %r604, 3;
bfi.b32 %r606, %r605, %r602, 4, 2;
add.s32 %r607, %r606, -32;
and.b32 %r565, %r598, 50529027;
shl.b32 %r608, %r599, 2;
and.b32 %r609, %r608, 67372036;
or.b32 %r610, %r565, -2139062144;
sub.s32 %r611, %r610, %r609;
xor.b32 %r612, %r611, -2139062144;
or.b32 %r613, %r565, %r609;
xor.b32 %r614, %r611, %r565;
and.b32 %r563, %r614, %r613;

prmt.b32 %r563, %r563, 0, 0xba98;


prmt.b32 %r565, %r565, 0, 0xba98;

xor.b32 %r615, %r565, 2139062143;
not.b32 %r616, %r563;
and.b32 %r617, %r612, %r616;
and.b32 %r618, %r615, %r563;
or.b32 %r568, %r618, %r617;
mov.u32 %r594, 0;

dp4a.s32.s32 %r567, %r568, %r27, %r594;

mul.lo.s32 %r619, %r607, %r567;
cvt.rn.f32.s32 %f145, %r619;
fma.rn.ftz.f32 %f146, %f141, %f145, 0f00000000;
add.s64 %rd200, %rd250, %rd13;
ld.global.u8 %r620, [%rd200];
shr.u32 %r621, %r620, %r18;
and.b32 %r622, %r621, 15;
add.s64 %rd201, %rd250, %rd12;
ld.global.u8 %r623, [%rd201];
shr.u32 %r624, %r623, %r19;
and.b32 %r625, %r624, 3;
bfi.b32 %r626, %r625, %r622, 4, 2;
add.s32 %r627, %r626, -32;
shr.u32 %r628, %r598, 2;
and.b32 %r573, %r628, 50529027;
shl.b32 %r629, %r599, 1;
and.b32 %r630, %r629, 67372036;
or.b32 %r631, %r573, -2139062144;
sub.s32 %r632, %r631, %r630;
xor.b32 %r633, %r632, -2139062144;
or.b32 %r634, %r573, %r630;
xor.b32 %r635, %r632, %r573;
and.b32 %r571, %r635, %r634;

prmt.b32 %r571, %r571, 0, 0xba98;


prmt.b32 %r573, %r573, 0, 0xba98;

xor.b32 %r636, %r573, 2139062143;
not.b32 %r637, %r571;
and.b32 %r638, %r633, %r637;
and.b32 %r639, %r636, %r571;
or.b32 %r576, %r639, %r638;

dp4a.s32.s32 %r575, %r576, %r29, %r594;

mul.lo.s32 %r640, %r627, %r575;
cvt.rn.f32.s32 %f147, %r640;
fma.rn.ftz.f32 %f148, %f142, %f147, %f146;
add.s64 %rd202, %rd250, %rd11;
ld.global.u8 %r641, [%rd202];
shr.u32 %r642, %r641, %r20;
and.b32 %r643, %r642, 15;
add.s64 %rd203, %rd250, %rd10;
ld.global.u8 %r644, [%rd203];
shr.u32 %r645, %r644, %r21;
and.b32 %r646, %r645, 3;
bfi.b32 %r647, %r646, %r643, 4, 2;
add.s32 %r648, %r647, -32;
shr.u32 %r649, %r598, 4;
and.b32 %r581, %r649, 50529027;
and.b32 %r650, %r599, 67372036;
or.b32 %r651, %r581, -2139062144;
sub.s32 %r652, %r651, %r650;
xor.b32 %r653, %r652, -2139062144;
or.b32 %r654, %r581, %r650;
xor.b32 %r655, %r652, %r581;
and.b32 %r579, %r655, %r654;

prmt.b32 %r579, %r579, 0, 0xba98;


prmt.b32 %r581, %r581, 0, 0xba98;

xor.b32 %r656, %r581, 2139062143;
not.b32 %r657, %r579;
and.b32 %r658, %r653, %r657;
and.b32 %r659, %r656, %r579;
or.b32 %r584, %r659, %r658;

dp4a.s32.s32 %r583, %r584, %r31, %r594;

mul.lo.s32 %r660, %r648, %r583;
cvt.rn.f32.s32 %f149, %r660;
fma.rn.ftz.f32 %f150, %f143, %f149, %f148;
add.s64 %rd204, %rd250, %rd9;
ld.global.u8 %r661, [%rd204];
shr.u32 %r662, %r661, %r22;
and.b32 %r663, %r662, 15;
add.s64 %rd205, %rd250, %rd8;
ld.global.u8 %r664, [%rd205];
shr.u32 %r665, %r664, %r23;
and.b32 %r666, %r665, 3;
bfi.b32 %r667, %r666, %r663, 4, 2;
add.s32 %r668, %r667, -32;
shr.u32 %r669, %r598, 6;
and.b32 %r589, %r669, 50529027;
shr.u32 %r670, %r599, 1;
and.b32 %r671, %r670, 67372036;
or.b32 %r672, %r589, -2139062144;
sub.s32 %r673, %r672, %r671;
xor.b32 %r674, %r673, -2139062144;
or.b32 %r675, %r589, %r671;
xor.b32 %r676, %r673, %r589;
and.b32 %r587, %r676, %r675;

prmt.b32 %r587, %r587, 0, 0xba98;


prmt.b32 %r589, %r589, 0, 0xba98;

xor.b32 %r677, %r589, 2139062143;
not.b32 %r678, %r587;
and.b32 %r679, %r674, %r678;
and.b32 %r680, %r677, %r587;
or.b32 %r592, %r680, %r679;

dp4a.s32.s32 %r591, %r592, %r33, %r594;

mul.lo.s32 %r681, %r668, %r591;
cvt.rn.f32.s32 %f151, %r681;
fma.rn.ftz.f32 %f152, %f144, %f151, %f150;
fma.rn.ftz.f32 %f33, %f140, %f152, %f33;

$L__BB142_14:
cvt.u32.u64 %r1854, %rd6;
add.s64 %rd79, %rd194, -2;
add.s64 %rd80, %rd69, %rd51;
ld.global.nc.u16 %rs37, [%rd80];

{ cvt.f32.f16 %f153, %rs37;}


add.s64 %rd81, %rd70, %rd51;
ld.global.nc.u16 %rs38, [%rd81];
ld.global.nc.u16 %rs39, [%rd81+2];
mov.b32 %r718, {%rs38, %rs39};
not.b32 %r719, %r718;
add.s64 %rd207, %rd79, %rd52;
ld.global.nc.u16 %rs40, [%rd207];
ld.global.nc.u16 %rs41, [%rd207+-2];

{.reg .f16 low,high;
mov.b32 {low,high},%r28;
cvt.f32.f16 %f154, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r30;
cvt.f32.f16 %f155, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r32;
cvt.f32.f16 %f156, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r34;
cvt.f32.f16 %f157, low;}


mov.b32 %r720, {%rs41, %rs40};
shr.s32 %r722, %r719, %r1854;
add.s64 %rd82, %rd71, %rd51;
ld.global.nc.u8 %rs42, [%rd82];
cvt.u32.u16 %r723, %rs42;
and.b32 %r724, %r723, 255;
shr.u32 %r725, %r724, %r16;
and.b32 %r726, %r725, 15;
add.s64 %rd83, %rd72, %rd51;
ld.global.nc.u8 %rs43, [%rd83];
cvt.u32.u16 %r727, %rs43;
and.b32 %r728, %r727, 255;
shr.u32 %r729, %r728, %r17;
and.b32 %r730, %r729, 3;
bfi.b32 %r731, %r730, %r726, 4, 2;
add.s32 %r732, %r731, -32;
and.b32 %r688, %r720, 50529027;
shl.b32 %r733, %r722, 2;
and.b32 %r734, %r733, 67372036;
or.b32 %r735, %r688, -2139062144;
sub.s32 %r736, %r735, %r734;
xor.b32 %r737, %r736, -2139062144;
or.b32 %r738, %r688, %r734;
xor.b32 %r739, %r736, %r688;
and.b32 %r686, %r739, %r738;

prmt.b32 %r686, %r686, 0, 0xba98;


prmt.b32 %r688, %r688, 0, 0xba98;

xor.b32 %r740, %r688, 2139062143;
not.b32 %r741, %r686;
and.b32 %r742, %r737, %r741;
and.b32 %r743, %r740, %r686;
or.b32 %r691, %r743, %r742;
mov.u32 %r717, 0;

dp4a.s32.s32 %r690, %r691, %r27, %r717;

mul.lo.s32 %r744, %r732, %r690;
cvt.rn.f32.s32 %f158, %r744;
fma.rn.ftz.f32 %f159, %f154, %f158, 0f00000000;
add.s64 %rd84, %rd73, %rd51;
ld.global.nc.u8 %rs44, [%rd84];
cvt.u32.u16 %r745, %rs44;
and.b32 %r746, %r745, 255;
shr.u32 %r747, %r746, %r18;
and.b32 %r748, %r747, 15;
add.s64 %rd85, %rd74, %rd51;
ld.global.nc.u8 %rs45, [%rd85];
cvt.u32.u16 %r749, %rs45;
and.b32 %r750, %r749, 255;
shr.u32 %r751, %r750, %r19;
and.b32 %r752, %r751, 3;
bfi.b32 %r753, %r752, %r748, 4, 2;
add.s32 %r754, %r753, -32;
shr.u32 %r755, %r720, 2;
and.b32 %r696, %r755, 50529027;
shl.b32 %r756, %r722, 1;
and.b32 %r757, %r756, 67372036;
or.b32 %r758, %r696, -2139062144;
sub.s32 %r759, %r758, %r757;
xor.b32 %r760, %r759, -2139062144;
or.b32 %r761, %r696, %r757;
xor.b32 %r762, %r759, %r696;
and.b32 %r694, %r762, %r761;

prmt.b32 %r694, %r694, 0, 0xba98;


prmt.b32 %r696, %r696, 0, 0xba98;

xor.b32 %r763, %r696, 2139062143;
not.b32 %r764, %r694;
and.b32 %r765, %r760, %r764;
and.b32 %r766, %r763, %r694;
or.b32 %r699, %r766, %r765;

dp4a.s32.s32 %r698, %r699, %r29, %r717;

mul.lo.s32 %r767, %r754, %r698;
cvt.rn.f32.s32 %f160, %r767;
fma.rn.ftz.f32 %f161, %f155, %f160, %f159;
add.s64 %rd86, %rd75, %rd51;
ld.global.nc.u8 %rs46, [%rd86];
cvt.u32.u16 %r768, %rs46;
and.b32 %r769, %r768, 255;
shr.u32 %r770, %r769, %r20;
and.b32 %r771, %r770, 15;
add.s64 %rd87, %rd76, %rd51;
ld.global.nc.u8 %rs47, [%rd87];
cvt.u32.u16 %r772, %rs47;
and.b32 %r773, %r772, 255;
shr.u32 %r774, %r773, %r21;
and.b32 %r775, %r774, 3;
bfi.b32 %r776, %r775, %r771, 4, 2;
add.s32 %r777, %r776, -32;
shr.u32 %r778, %r720, 4;
and.b32 %r704, %r778, 50529027;
and.b32 %r779, %r722, 67372036;
or.b32 %r780, %r704, -2139062144;
sub.s32 %r781, %r780, %r779;
xor.b32 %r782, %r781, -2139062144;
or.b32 %r783, %r704, %r779;
xor.b32 %r784, %r781, %r704;
and.b32 %r702, %r784, %r783;

prmt.b32 %r702, %r702, 0, 0xba98;


prmt.b32 %r704, %r704, 0, 0xba98;

xor.b32 %r785, %r704, 2139062143;
not.b32 %r786, %r702;
and.b32 %r787, %r782, %r786;
and.b32 %r788, %r785, %r702;
or.b32 %r707, %r788, %r787;

dp4a.s32.s32 %r706, %r707, %r31, %r717;

mul.lo.s32 %r789, %r777, %r706;
cvt.rn.f32.s32 %f162, %r789;
fma.rn.ftz.f32 %f163, %f156, %f162, %f161;
add.s64 %rd88, %rd77, %rd51;
ld.global.nc.u8 %rs48, [%rd88];
cvt.u32.u16 %r790, %rs48;
and.b32 %r791, %r790, 255;
shr.u32 %r792, %r791, %r22;
and.b32 %r793, %r792, 15;
add.s64 %rd89, %rd78, %rd51;
ld.global.nc.u8 %rs49, [%rd89];
cvt.u32.u16 %r794, %rs49;
and.b32 %r795, %r794, 255;
shr.u32 %r796, %r795, %r23;
and.b32 %r797, %r796, 3;
bfi.b32 %r798, %r797, %r793, 4, 2;
add.s32 %r799, %r798, -32;
shr.u32 %r800, %r720, 6;
and.b32 %r712, %r800, 50529027;
shr.u32 %r801, %r722, 1;
and.b32 %r802, %r801, 67372036;
or.b32 %r803, %r712, -2139062144;
sub.s32 %r804, %r803, %r802;
xor.b32 %r805, %r804, -2139062144;
or.b32 %r806, %r712, %r802;
xor.b32 %r807, %r804, %r712;
and.b32 %r710, %r807, %r806;

prmt.b32 %r710, %r710, 0, 0xba98;


prmt.b32 %r712, %r712, 0, 0xba98;

xor.b32 %r808, %r712, 2139062143;
not.b32 %r809, %r710;
and.b32 %r810, %r805, %r809;
and.b32 %r811, %r808, %r710;
or.b32 %r715, %r811, %r810;

dp4a.s32.s32 %r714, %r715, %r33, %r717;

mul.lo.s32 %r812, %r799, %r714;
cvt.rn.f32.s32 %f164, %r812;
fma.rn.ftz.f32 %f165, %f157, %f164, %f163;
fma.rn.ftz.f32 %f28, %f153, %f165, %f28;
@%p13 bra $L__BB142_16;

cvt.u32.u64 %r1858, %rd6;
add.s64 %rd208, %rd250, %rd22;
ld.global.u16 %rs50, [%rd208+108];

{ cvt.f32.f16 %f166, %rs50;}


add.s64 %rd209, %rd250, %rd33;
add.s64 %rd210, %rd250, %rd32;
ld.global.u16 %rs51, [%rd210];
ld.global.u16 %rs52, [%rd210+2];
mov.b32 %r850, {%rs51, %rs52};
not.b32 %r851, %r850;
ld.global.u16 %rs53, [%rd209+2];
ld.global.u16 %rs54, [%rd209];

{.reg .f16 low,high;
mov.b32 {low,high},%r28;
cvt.f32.f16 %f167, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r30;
cvt.f32.f16 %f168, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r32;
cvt.f32.f16 %f169, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r34;
cvt.f32.f16 %f170, low;}


mov.b32 %r852, {%rs54, %rs53};
shr.s32 %r853, %r851, %r1858;
add.s64 %rd211, %rd250, %rd30;
ld.global.u8 %r854, [%rd211];
shr.u32 %r855, %r854, %r16;
and.b32 %r856, %r855, 15;
add.s64 %rd212, %rd250, %rd29;
ld.global.u8 %r857, [%rd212];
shr.u32 %r858, %r857, %r17;
and.b32 %r859, %r858, 3;
bfi.b32 %r860, %r859, %r856, 4, 2;
add.s32 %r861, %r860, -32;
and.b32 %r819, %r852, 50529027;
shl.b32 %r862, %r853, 2;
and.b32 %r863, %r862, 67372036;
or.b32 %r864, %r819, -2139062144;
sub.s32 %r865, %r864, %r863;
xor.b32 %r866, %r865, -2139062144;
or.b32 %r867, %r819, %r863;
xor.b32 %r868, %r865, %r819;
and.b32 %r817, %r868, %r867;

prmt.b32 %r817, %r817, 0, 0xba98;


prmt.b32 %r819, %r819, 0, 0xba98;

xor.b32 %r869, %r819, 2139062143;
not.b32 %r870, %r817;
and.b32 %r871, %r866, %r870;
and.b32 %r872, %r869, %r817;
or.b32 %r822, %r872, %r871;

dp4a.s32.s32 %r821, %r822, %r27, %r717;

mul.lo.s32 %r873, %r861, %r821;
cvt.rn.f32.s32 %f171, %r873;
fma.rn.ftz.f32 %f172, %f167, %f171, 0f00000000;
add.s64 %rd213, %rd250, %rd28;
ld.global.u8 %r874, [%rd213];
shr.u32 %r875, %r874, %r18;
and.b32 %r876, %r875, 15;
add.s64 %rd214, %rd250, %rd27;
ld.global.u8 %r877, [%rd214];
shr.u32 %r878, %r877, %r19;
and.b32 %r879, %r878, 3;
bfi.b32 %r880, %r879, %r876, 4, 2;
add.s32 %r881, %r880, -32;
shr.u32 %r882, %r852, 2;
and.b32 %r827, %r882, 50529027;
shl.b32 %r883, %r853, 1;
and.b32 %r884, %r883, 67372036;
or.b32 %r885, %r827, -2139062144;
sub.s32 %r886, %r885, %r884;
xor.b32 %r887, %r886, -2139062144;
or.b32 %r888, %r827, %r884;
xor.b32 %r889, %r886, %r827;
and.b32 %r825, %r889, %r888;

prmt.b32 %r825, %r825, 0, 0xba98;


prmt.b32 %r827, %r827, 0, 0xba98;

xor.b32 %r890, %r827, 2139062143;
not.b32 %r891, %r825;
and.b32 %r892, %r887, %r891;
and.b32 %r893, %r890, %r825;
or.b32 %r830, %r893, %r892;

dp4a.s32.s32 %r829, %r830, %r29, %r717;

mul.lo.s32 %r894, %r881, %r829;
cvt.rn.f32.s32 %f173, %r894;
fma.rn.ftz.f32 %f174, %f168, %f173, %f172;
add.s64 %rd215, %rd250, %rd26;
ld.global.u8 %r895, [%rd215];
shr.u32 %r896, %r895, %r20;
and.b32 %r897, %r896, 15;
add.s64 %rd216, %rd250, %rd25;
ld.global.u8 %r898, [%rd216];
shr.u32 %r899, %r898, %r21;
and.b32 %r900, %r899, 3;
bfi.b32 %r901, %r900, %r897, 4, 2;
add.s32 %r902, %r901, -32;
shr.u32 %r903, %r852, 4;
and.b32 %r835, %r903, 50529027;
and.b32 %r904, %r853, 67372036;
or.b32 %r905, %r835, -2139062144;
sub.s32 %r906, %r905, %r904;
xor.b32 %r907, %r906, -2139062144;
or.b32 %r908, %r835, %r904;
xor.b32 %r909, %r906, %r835;
and.b32 %r833, %r909, %r908;

prmt.b32 %r833, %r833, 0, 0xba98;


prmt.b32 %r835, %r835, 0, 0xba98;

xor.b32 %r910, %r835, 2139062143;
not.b32 %r911, %r833;
and.b32 %r912, %r907, %r911;
and.b32 %r913, %r910, %r833;
or.b32 %r838, %r913, %r912;

dp4a.s32.s32 %r837, %r838, %r31, %r717;

mul.lo.s32 %r914, %r902, %r837;
cvt.rn.f32.s32 %f175, %r914;
fma.rn.ftz.f32 %f176, %f169, %f175, %f174;
add.s64 %rd217, %rd250, %rd24;
ld.global.u8 %r915, [%rd217];
shr.u32 %r916, %r915, %r22;
and.b32 %r917, %r916, 15;
add.s64 %rd218, %rd250, %rd23;
ld.global.u8 %r918, [%rd218];
shr.u32 %r919, %r918, %r23;
and.b32 %r920, %r919, 3;
bfi.b32 %r921, %r920, %r917, 4, 2;
add.s32 %r922, %r921, -32;
shr.u32 %r923, %r852, 6;
and.b32 %r843, %r923, 50529027;
shr.u32 %r924, %r853, 1;
and.b32 %r925, %r924, 67372036;
or.b32 %r926, %r843, -2139062144;
sub.s32 %r927, %r926, %r925;
xor.b32 %r928, %r927, -2139062144;
or.b32 %r929, %r843, %r925;
xor.b32 %r930, %r927, %r843;
and.b32 %r841, %r930, %r929;

prmt.b32 %r841, %r841, 0, 0xba98;


prmt.b32 %r843, %r843, 0, 0xba98;

xor.b32 %r931, %r843, 2139062143;
not.b32 %r932, %r841;
and.b32 %r933, %r928, %r932;
and.b32 %r934, %r931, %r841;
or.b32 %r846, %r934, %r933;

dp4a.s32.s32 %r845, %r846, %r33, %r717;

mul.lo.s32 %r935, %r922, %r845;
cvt.rn.f32.s32 %f177, %r935;
fma.rn.ftz.f32 %f178, %f170, %f177, %f176;
fma.rn.ftz.f32 %f32, %f166, %f178, %f32;

$L__BB142_16:
cvt.u32.u64 %r1851, %rd6;
add.s64 %rd219, %rd80, %rd51;
ld.global.nc.u16 %rs55, [%rd219];

{ cvt.f32.f16 %f179, %rs55;}


add.s64 %rd220, %rd81, %rd51;
ld.global.nc.u16 %rs56, [%rd220];
ld.global.nc.u16 %rs57, [%rd220+2];
mov.b32 %r972, {%rs56, %rs57};
not.b32 %r973, %r972;
add.s64 %rd221, %rd79, %rd51;
add.s64 %rd222, %rd221, %rd52;
ld.global.nc.u16 %rs58, [%rd222];
ld.global.nc.u16 %rs59, [%rd222+-2];

{.reg .f16 low,high;
mov.b32 {low,high},%r28;
cvt.f32.f16 %f180, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r30;
cvt.f32.f16 %f181, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r32;
cvt.f32.f16 %f182, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r34;
cvt.f32.f16 %f183, low;}


mov.b32 %r974, {%rs59, %rs58};
shr.s32 %r976, %r973, %r1851;
add.s64 %rd223, %rd82, %rd51;
ld.global.nc.u8 %rs60, [%rd223];
cvt.u32.u16 %r977, %rs60;
and.b32 %r978, %r977, 255;
shr.u32 %r979, %r978, %r16;
and.b32 %r980, %r979, 15;
add.s64 %rd224, %rd83, %rd51;
ld.global.nc.u8 %rs61, [%rd224];
cvt.u32.u16 %r981, %rs61;
and.b32 %r982, %r981, 255;
shr.u32 %r983, %r982, %r17;
and.b32 %r984, %r983, 3;
bfi.b32 %r985, %r984, %r980, 4, 2;
add.s32 %r986, %r985, -32;
and.b32 %r942, %r974, 50529027;
shl.b32 %r987, %r976, 2;
and.b32 %r988, %r987, 67372036;
or.b32 %r989, %r942, -2139062144;
sub.s32 %r990, %r989, %r988;
xor.b32 %r991, %r990, -2139062144;
or.b32 %r992, %r942, %r988;
xor.b32 %r993, %r990, %r942;
and.b32 %r940, %r993, %r992;

prmt.b32 %r940, %r940, 0, 0xba98;


prmt.b32 %r942, %r942, 0, 0xba98;

xor.b32 %r994, %r942, 2139062143;
not.b32 %r995, %r940;
and.b32 %r996, %r991, %r995;
and.b32 %r997, %r994, %r940;
or.b32 %r945, %r997, %r996;

dp4a.s32.s32 %r944, %r945, %r27, %r717;

mul.lo.s32 %r998, %r986, %r944;
cvt.rn.f32.s32 %f184, %r998;
fma.rn.ftz.f32 %f185, %f180, %f184, 0f00000000;
add.s64 %rd225, %rd84, %rd51;
ld.global.nc.u8 %rs62, [%rd225];
cvt.u32.u16 %r999, %rs62;
and.b32 %r1000, %r999, 255;
shr.u32 %r1001, %r1000, %r18;
and.b32 %r1002, %r1001, 15;
add.s64 %rd226, %rd85, %rd51;
ld.global.nc.u8 %rs63, [%rd226];
cvt.u32.u16 %r1003, %rs63;
and.b32 %r1004, %r1003, 255;
shr.u32 %r1005, %r1004, %r19;
and.b32 %r1006, %r1005, 3;
bfi.b32 %r1007, %r1006, %r1002, 4, 2;
add.s32 %r1008, %r1007, -32;
shr.u32 %r1009, %r974, 2;
and.b32 %r950, %r1009, 50529027;
shl.b32 %r1010, %r976, 1;
and.b32 %r1011, %r1010, 67372036;
or.b32 %r1012, %r950, -2139062144;
sub.s32 %r1013, %r1012, %r1011;
xor.b32 %r1014, %r1013, -2139062144;
or.b32 %r1015, %r950, %r1011;
xor.b32 %r1016, %r1013, %r950;
and.b32 %r948, %r1016, %r1015;

prmt.b32 %r948, %r948, 0, 0xba98;


prmt.b32 %r950, %r950, 0, 0xba98;

xor.b32 %r1017, %r950, 2139062143;
not.b32 %r1018, %r948;
and.b32 %r1019, %r1014, %r1018;
and.b32 %r1020, %r1017, %r948;
or.b32 %r953, %r1020, %r1019;

dp4a.s32.s32 %r952, %r953, %r29, %r717;

mul.lo.s32 %r1021, %r1008, %r952;
cvt.rn.f32.s32 %f186, %r1021;
fma.rn.ftz.f32 %f187, %f181, %f186, %f185;
add.s64 %rd227, %rd86, %rd51;
ld.global.nc.u8 %rs64, [%rd227];
cvt.u32.u16 %r1022, %rs64;
and.b32 %r1023, %r1022, 255;
shr.u32 %r1024, %r1023, %r20;
and.b32 %r1025, %r1024, 15;
add.s64 %rd228, %rd87, %rd51;
ld.global.nc.u8 %rs65, [%rd228];
cvt.u32.u16 %r1026, %rs65;
and.b32 %r1027, %r1026, 255;
shr.u32 %r1028, %r1027, %r21;
and.b32 %r1029, %r1028, 3;
bfi.b32 %r1030, %r1029, %r1025, 4, 2;
add.s32 %r1031, %r1030, -32;
shr.u32 %r1032, %r974, 4;
and.b32 %r958, %r1032, 50529027;
and.b32 %r1033, %r976, 67372036;
or.b32 %r1034, %r958, -2139062144;
sub.s32 %r1035, %r1034, %r1033;
xor.b32 %r1036, %r1035, -2139062144;
or.b32 %r1037, %r958, %r1033;
xor.b32 %r1038, %r1035, %r958;
and.b32 %r956, %r1038, %r1037;

prmt.b32 %r956, %r956, 0, 0xba98;


prmt.b32 %r958, %r958, 0, 0xba98;

xor.b32 %r1039, %r958, 2139062143;
not.b32 %r1040, %r956;
and.b32 %r1041, %r1036, %r1040;
and.b32 %r1042, %r1039, %r956;
or.b32 %r961, %r1042, %r1041;

dp4a.s32.s32 %r960, %r961, %r31, %r717;

mul.lo.s32 %r1043, %r1031, %r960;
cvt.rn.f32.s32 %f188, %r1043;
fma.rn.ftz.f32 %f189, %f182, %f188, %f187;
add.s64 %rd229, %rd88, %rd51;
ld.global.nc.u8 %rs66, [%rd229];
cvt.u32.u16 %r1044, %rs66;
and.b32 %r1045, %r1044, 255;
shr.u32 %r1046, %r1045, %r22;
and.b32 %r1047, %r1046, 15;
add.s64 %rd230, %rd89, %rd51;
ld.global.nc.u8 %rs67, [%rd230];
cvt.u32.u16 %r1048, %rs67;
and.b32 %r1049, %r1048, 255;
shr.u32 %r1050, %r1049, %r23;
and.b32 %r1051, %r1050, 3;
bfi.b32 %r1052, %r1051, %r1047, 4, 2;
add.s32 %r1053, %r1052, -32;
shr.u32 %r1054, %r974, 6;
and.b32 %r966, %r1054, 50529027;
shr.u32 %r1055, %r976, 1;
and.b32 %r1056, %r1055, 67372036;
or.b32 %r1057, %r966, -2139062144;
sub.s32 %r1058, %r1057, %r1056;
xor.b32 %r1059, %r1058, -2139062144;
or.b32 %r1060, %r966, %r1056;
xor.b32 %r1061, %r1058, %r966;
and.b32 %r964, %r1061, %r1060;

prmt.b32 %r964, %r964, 0, 0xba98;


prmt.b32 %r966, %r966, 0, 0xba98;

xor.b32 %r1062, %r966, 2139062143;
not.b32 %r1063, %r964;
and.b32 %r1064, %r1059, %r1063;
and.b32 %r1065, %r1062, %r964;
or.b32 %r969, %r1065, %r1064;

dp4a.s32.s32 %r968, %r969, %r33, %r717;

mul.lo.s32 %r1066, %r1053, %r968;
cvt.rn.f32.s32 %f190, %r1066;
fma.rn.ftz.f32 %f191, %f183, %f190, %f189;
fma.rn.ftz.f32 %f27, %f179, %f191, %f27;
@%p13 bra $L__BB142_18;

cvt.u32.u64 %r1857, %rd6;
add.s64 %rd231, %rd250, %rd34;
ld.global.u16 %rs68, [%rd231+108];

{ cvt.f32.f16 %f192, %rs68;}


add.s64 %rd232, %rd250, %rd44;
add.s64 %rd233, %rd250, %rd43;
ld.global.u16 %rs69, [%rd233];
ld.global.u16 %rs70, [%rd233+2];
mov.b32 %r1104, {%rs69, %rs70};
not.b32 %r1105, %r1104;
ld.global.u16 %rs71, [%rd232+2];
ld.global.u16 %rs72, [%rd232];

{.reg .f16 low,high;
mov.b32 {low,high},%r28;
cvt.f32.f16 %f193, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r30;
cvt.f32.f16 %f194, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r32;
cvt.f32.f16 %f195, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r34;
cvt.f32.f16 %f196, low;}


mov.b32 %r1106, {%rs72, %rs71};
shr.s32 %r1107, %r1105, %r1857;
add.s64 %rd234, %rd250, %rd42;
ld.global.u8 %r1108, [%rd234];
shr.u32 %r1109, %r1108, %r16;
and.b32 %r1110, %r1109, 15;
add.s64 %rd235, %rd250, %rd41;
ld.global.u8 %r1111, [%rd235];
shr.u32 %r1112, %r1111, %r17;
and.b32 %r1113, %r1112, 3;
bfi.b32 %r1114, %r1113, %r1110, 4, 2;
add.s32 %r1115, %r1114, -32;
and.b32 %r1073, %r1106, 50529027;
shl.b32 %r1116, %r1107, 2;
and.b32 %r1117, %r1116, 67372036;
or.b32 %r1118, %r1073, -2139062144;
sub.s32 %r1119, %r1118, %r1117;
xor.b32 %r1120, %r1119, -2139062144;
or.b32 %r1121, %r1073, %r1117;
xor.b32 %r1122, %r1119, %r1073;
and.b32 %r1071, %r1122, %r1121;

prmt.b32 %r1071, %r1071, 0, 0xba98;


prmt.b32 %r1073, %r1073, 0, 0xba98;

xor.b32 %r1123, %r1073, 2139062143;
not.b32 %r1124, %r1071;
and.b32 %r1125, %r1120, %r1124;
and.b32 %r1126, %r1123, %r1071;
or.b32 %r1076, %r1126, %r1125;
mov.u32 %r1102, 0;

dp4a.s32.s32 %r1075, %r1076, %r27, %r1102;

mul.lo.s32 %r1127, %r1115, %r1075;
cvt.rn.f32.s32 %f197, %r1127;
fma.rn.ftz.f32 %f198, %f193, %f197, 0f00000000;
add.s64 %rd236, %rd250, %rd40;
ld.global.u8 %r1128, [%rd236];
shr.u32 %r1129, %r1128, %r18;
and.b32 %r1130, %r1129, 15;
add.s64 %rd237, %rd250, %rd39;
ld.global.u8 %r1131, [%rd237];
shr.u32 %r1132, %r1131, %r19;
and.b32 %r1133, %r1132, 3;
bfi.b32 %r1134, %r1133, %r1130, 4, 2;
add.s32 %r1135, %r1134, -32;
shr.u32 %r1136, %r1106, 2;
and.b32 %r1081, %r1136, 50529027;
shl.b32 %r1137, %r1107, 1;
and.b32 %r1138, %r1137, 67372036;
or.b32 %r1139, %r1081, -2139062144;
sub.s32 %r1140, %r1139, %r1138;
xor.b32 %r1141, %r1140, -2139062144;
or.b32 %r1142, %r1081, %r1138;
xor.b32 %r1143, %r1140, %r1081;
and.b32 %r1079, %r1143, %r1142;

prmt.b32 %r1079, %r1079, 0, 0xba98;


prmt.b32 %r1081, %r1081, 0, 0xba98;

xor.b32 %r1144, %r1081, 2139062143;
not.b32 %r1145, %r1079;
and.b32 %r1146, %r1141, %r1145;
and.b32 %r1147, %r1144, %r1079;
or.b32 %r1084, %r1147, %r1146;

dp4a.s32.s32 %r1083, %r1084, %r29, %r1102;

mul.lo.s32 %r1148, %r1135, %r1083;
cvt.rn.f32.s32 %f199, %r1148;
fma.rn.ftz.f32 %f200, %f194, %f199, %f198;
add.s64 %rd238, %rd250, %rd38;
ld.global.u8 %r1149, [%rd238];
shr.u32 %r1150, %r1149, %r20;
and.b32 %r1151, %r1150, 15;
add.s64 %rd239, %rd250, %rd37;
ld.global.u8 %r1152, [%rd239];
shr.u32 %r1153, %r1152, %r21;
and.b32 %r1154, %r1153, 3;
bfi.b32 %r1155, %r1154, %r1151, 4, 2;
add.s32 %r1156, %r1155, -32;
shr.u32 %r1157, %r1106, 4;
and.b32 %r1089, %r1157, 50529027;
and.b32 %r1158, %r1107, 67372036;
or.b32 %r1159, %r1089, -2139062144;
sub.s32 %r1160, %r1159, %r1158;
xor.b32 %r1161, %r1160, -2139062144;
or.b32 %r1162, %r1089, %r1158;
xor.b32 %r1163, %r1160, %r1089;
and.b32 %r1087, %r1163, %r1162;

prmt.b32 %r1087, %r1087, 0, 0xba98;


prmt.b32 %r1089, %r1089, 0, 0xba98;

xor.b32 %r1164, %r1089, 2139062143;
not.b32 %r1165, %r1087;
and.b32 %r1166, %r1161, %r1165;
and.b32 %r1167, %r1164, %r1087;
or.b32 %r1092, %r1167, %r1166;

dp4a.s32.s32 %r1091, %r1092, %r31, %r1102;

mul.lo.s32 %r1168, %r1156, %r1091;
cvt.rn.f32.s32 %f201, %r1168;
fma.rn.ftz.f32 %f202, %f195, %f201, %f200;
add.s64 %rd240, %rd250, %rd36;
ld.global.u8 %r1169, [%rd240];
shr.u32 %r1170, %r1169, %r22;
and.b32 %r1171, %r1170, 15;
add.s64 %rd241, %rd250, %rd35;
ld.global.u8 %r1172, [%rd241];
shr.u32 %r1173, %r1172, %r23;
and.b32 %r1174, %r1173, 3;
bfi.b32 %r1175, %r1174, %r1171, 4, 2;
add.s32 %r1176, %r1175, -32;
shr.u32 %r1177, %r1106, 6;
and.b32 %r1097, %r1177, 50529027;
shr.u32 %r1178, %r1107, 1;
and.b32 %r1179, %r1178, 67372036;
or.b32 %r1180, %r1097, -2139062144;
sub.s32 %r1181, %r1180, %r1179;
xor.b32 %r1182, %r1181, -2139062144;
or.b32 %r1183, %r1097, %r1179;
xor.b32 %r1184, %r1181, %r1097;
and.b32 %r1095, %r1184, %r1183;

prmt.b32 %r1095, %r1095, 0, 0xba98;


prmt.b32 %r1097, %r1097, 0, 0xba98;

xor.b32 %r1185, %r1097, 2139062143;
not.b32 %r1186, %r1095;
and.b32 %r1187, %r1182, %r1186;
and.b32 %r1188, %r1185, %r1095;
or.b32 %r1100, %r1188, %r1187;

dp4a.s32.s32 %r1099, %r1100, %r33, %r1102;

mul.lo.s32 %r1189, %r1176, %r1099;
cvt.rn.f32.s32 %f203, %r1189;
fma.rn.ftz.f32 %f204, %f196, %f203, %f202;
fma.rn.ftz.f32 %f31, %f192, %f204, %f31;

$L__BB142_18:
add.s32 %r1861, %r1861, 64;
add.s64 %rd250, %rd250, 880;
add.s64 %rd249, %rd249, 880;
add.s32 %r1862, %r1862, 8;
setp.lt.s32 %p17, %r1862, %r5;
@%p17 bra $L__BB142_10;

$L__BB142_19:
mov.u32 %r1841, %tid.y;
setp.eq.s32 %p18, %r1841, 0;
@%p18 bra $L__BB142_23;

mov.u32 %r1842, %tid.y;
setp.ne.s64 %p19, %rd97, 0;
add.s32 %r37, %r1842, -1;
shl.b32 %r1190, %r1842, 9;
mov.u32 %r1191, _ZZ13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r1192, %r1191, %r1190;
shl.b32 %r1193, %r2, 2;
add.s32 %r1194, %r1192, %r1193;
st.shared.f32 [%r1194+-512], %f30;
@%p19 bra $L__BB142_22;
bra.uni $L__BB142_21;

$L__BB142_22:
shl.b32 %r1195, %r37, 9;
mov.u32 %r1196, _ZZ13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r1197, %r1196, %r1195;
add.s32 %r1199, %r1197, %r1193;
st.shared.f32 [%r1199], %f34;
add.s32 %r1838, %r1194, -512;
st.shared.f32 [%r1838+128], %f29;
st.shared.f32 [%r1199+128], %f33;
add.s32 %r1839, %r1194, -512;
st.shared.f32 [%r1839+256], %f28;
st.shared.f32 [%r1199+256], %f32;
add.s32 %r1840, %r1194, -512;
st.shared.f32 [%r1840+384], %f27;
st.shared.f32 [%r1199+384], %f31;
bra.uni $L__BB142_23;

$L__BB142_21:
add.s32 %r1835, %r1194, -512;
st.shared.f32 [%r1835+128], %f29;
add.s32 %r1836, %r1194, -512;
st.shared.f32 [%r1836+256], %f28;
add.s32 %r1837, %r1194, -512;
st.shared.f32 [%r1837+384], %f27;

$L__BB142_23:
mov.u32 %r1852, %tid.y;
setp.ne.s32 %p319, %r1852, 0;
bar.sync 0;
@%p319 bra $L__BB142_78;
bra.uni $L__BB142_24;

$L__BB142_78:
ret;

$L__BB142_24:
ld.param.u64 %rd248, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.f32 %f1049, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
ld.param.u32 %r1845, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r1844, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
mov.u32 %r1843, %ctaid.z;
setp.ne.s64 %p21, %rd97, 0;
mov.u32 %r1201, %ctaid.y;
mad.lo.s32 %r1202, %r1201, %r1845, %r4;
mad.lo.s32 %r1203, %r1843, %r1844, %r1202;
neg.ftz.f32 %f35, %f1049;
shl.b32 %r1204, %r2, 2;
mov.u32 %r1205, _ZZ13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r39, %r1205, %r1204;
cvta.to.global.u64 %rd242, %rd248;
mul.wide.u32 %rd243, %r1203, 4;
add.s64 %rd92, %rd242, %rd243;
@%p21 bra $L__BB142_33;
bra.uni $L__BB142_25;

$L__BB142_33:
ld.param.u32 %r1846, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
mov.u32 %r1279, _ZZ13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r41, %r1279, %r1204;
setp.gt.s32 %p54, %r1846, 2;
@%p54 bra $L__BB142_44;

ld.param.u32 %r1849, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p57, %r1849, 1;
@%p57 bra $L__BB142_62;

ld.param.u32 %r1850, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p58, %r1850, 2;
@%p58 bra $L__BB142_36;
bra.uni $L__BB142_70;

$L__BB142_36:
ld.shared.f32 %f605, [%r39];
add.ftz.f32 %f606, %f605, %f30;
ld.shared.f32 %f607, [%r41];
add.ftz.f32 %f608, %f607, %f34;
ld.shared.f32 %f609, [%r39+512];
add.ftz.f32 %f610, %f609, %f606;
ld.shared.f32 %f611, [%r41+512];
add.ftz.f32 %f612, %f611, %f608;
ld.shared.f32 %f613, [%r39+1024];
add.ftz.f32 %f614, %f613, %f610;
ld.shared.f32 %f615, [%r41+1024];
add.ftz.f32 %f616, %f615, %f612;
mov.b32 %r1502, %f614;
mov.u32 %r1503, 31;
mov.u32 %r1504, 16;
mov.u32 %r1505, -1;
shfl.sync.bfly.b32 %r1506|%p163, %r1502, %r1504, %r1503, %r1505;
mov.b32 %f617, %r1506;
add.ftz.f32 %f618, %f614, %f617;
mov.b32 %r1507, %f618;
mov.u32 %r1508, 8;
shfl.sync.bfly.b32 %r1509|%p164, %r1507, %r1508, %r1503, %r1505;
mov.b32 %f619, %r1509;
add.ftz.f32 %f620, %f618, %f619;
mov.b32 %r1510, %f620;
mov.u32 %r1511, 4;
shfl.sync.bfly.b32 %r1512|%p165, %r1510, %r1511, %r1503, %r1505;
mov.b32 %f621, %r1512;
add.ftz.f32 %f622, %f620, %f621;
mov.b32 %r1513, %f622;
mov.u32 %r1514, 2;
shfl.sync.bfly.b32 %r1515|%p166, %r1513, %r1514, %r1503, %r1505;
mov.b32 %f623, %r1515;
add.ftz.f32 %f624, %f622, %f623;
mov.b32 %r1516, %f624;
mov.u32 %r1517, 1;
shfl.sync.bfly.b32 %r1518|%p167, %r1516, %r1517, %r1503, %r1505;
mov.b32 %f625, %r1518;
add.ftz.f32 %f56, %f624, %f625;
mov.b32 %r1519, %f616;
shfl.sync.bfly.b32 %r1520|%p168, %r1519, %r1504, %r1503, %r1505;
mov.b32 %f626, %r1520;
add.ftz.f32 %f627, %f616, %f626;
mov.b32 %r1521, %f627;
shfl.sync.bfly.b32 %r1522|%p169, %r1521, %r1508, %r1503, %r1505;
mov.b32 %f628, %r1522;
add.ftz.f32 %f629, %f627, %f628;
mov.b32 %r1523, %f629;
shfl.sync.bfly.b32 %r1524|%p170, %r1523, %r1511, %r1503, %r1505;
mov.b32 %f630, %r1524;
add.ftz.f32 %f631, %f629, %f630;
mov.b32 %r1525, %f631;
shfl.sync.bfly.b32 %r1526|%p171, %r1525, %r1514, %r1503, %r1505;
mov.b32 %f632, %r1526;
add.ftz.f32 %f633, %f631, %f632;
mov.b32 %r1527, %f633;
shfl.sync.bfly.b32 %r1528|%p172, %r1527, %r1517, %r1503, %r1505;
mov.b32 %f634, %r1528;
add.ftz.f32 %f57, %f633, %f634;
setp.ge.u32 %p173, %r4, %r57;
setp.ne.s32 %p174, %r2, 0;
or.pred %p175, %p174, %p173;
@%p175 bra $L__BB142_38;

add.ftz.f32 %f635, %f5, %f56;
add.ftz.f32 %f636, %f4, %f57;
mul.ftz.f32 %f637, %f636, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f638, %f637;
add.ftz.f32 %f639, %f638, 0f3F800000;
div.approx.ftz.f32 %f640, %f636, %f639;
mul.ftz.f32 %f641, %f635, %f640;
st.global.f32 [%rd92], %f641;

$L__BB142_38:
ld.shared.f32 %f642, [%r39+128];
add.ftz.f32 %f643, %f642, %f29;
ld.shared.f32 %f644, [%r41+128];
add.ftz.f32 %f645, %f644, %f33;
ld.shared.f32 %f646, [%r39+640];
add.ftz.f32 %f647, %f646, %f643;
ld.shared.f32 %f648, [%r41+640];
add.ftz.f32 %f649, %f648, %f645;
ld.shared.f32 %f650, [%r39+1152];
add.ftz.f32 %f651, %f650, %f647;
ld.shared.f32 %f652, [%r41+1152];
add.ftz.f32 %f653, %f652, %f649;
mov.b32 %r1529, %f651;
shfl.sync.bfly.b32 %r1533|%p176, %r1529, %r1504, %r1503, %r1505;
mov.b32 %f654, %r1533;
add.ftz.f32 %f655, %f651, %f654;
mov.b32 %r1534, %f655;
shfl.sync.bfly.b32 %r1536|%p177, %r1534, %r1508, %r1503, %r1505;
mov.b32 %f656, %r1536;
add.ftz.f32 %f657, %f655, %f656;
mov.b32 %r1537, %f657;
shfl.sync.bfly.b32 %r1539|%p178, %r1537, %r1511, %r1503, %r1505;
mov.b32 %f658, %r1539;
add.ftz.f32 %f659, %f657, %f658;
mov.b32 %r1540, %f659;
shfl.sync.bfly.b32 %r1542|%p179, %r1540, %r1514, %r1503, %r1505;
mov.b32 %f660, %r1542;
add.ftz.f32 %f661, %f659, %f660;
mov.b32 %r1543, %f661;
shfl.sync.bfly.b32 %r1545|%p180, %r1543, %r1517, %r1503, %r1505;
mov.b32 %f662, %r1545;
add.ftz.f32 %f58, %f661, %f662;
mov.b32 %r1546, %f653;
shfl.sync.bfly.b32 %r1547|%p181, %r1546, %r1504, %r1503, %r1505;
mov.b32 %f663, %r1547;
add.ftz.f32 %f664, %f653, %f663;
mov.b32 %r1548, %f664;
shfl.sync.bfly.b32 %r1549|%p182, %r1548, %r1508, %r1503, %r1505;
mov.b32 %f665, %r1549;
add.ftz.f32 %f666, %f664, %f665;
mov.b32 %r1550, %f666;
shfl.sync.bfly.b32 %r1551|%p183, %r1550, %r1511, %r1503, %r1505;
mov.b32 %f667, %r1551;
add.ftz.f32 %f668, %f666, %f667;
mov.b32 %r1552, %f668;
shfl.sync.bfly.b32 %r1553|%p184, %r1552, %r1514, %r1503, %r1505;
mov.b32 %f669, %r1553;
add.ftz.f32 %f670, %f668, %f669;
mov.b32 %r1554, %f670;
shfl.sync.bfly.b32 %r1555|%p185, %r1554, %r1517, %r1503, %r1505;
mov.b32 %f671, %r1555;
add.ftz.f32 %f59, %f670, %f671;
add.s32 %r1556, %r4, 1;
setp.ge.u32 %p186, %r1556, %r57;
setp.ne.s32 %p187, %r2, 1;
or.pred %p188, %p187, %p186;
@%p188 bra $L__BB142_40;

add.ftz.f32 %f672, %f5, %f58;
add.ftz.f32 %f673, %f4, %f59;
mul.ftz.f32 %f674, %f673, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f675, %f674;
add.ftz.f32 %f676, %f675, 0f3F800000;
div.approx.ftz.f32 %f677, %f673, %f676;
mul.ftz.f32 %f678, %f672, %f677;
st.global.f32 [%rd92+4], %f678;

$L__BB142_40:
ld.shared.f32 %f679, [%r39+256];
add.ftz.f32 %f680, %f679, %f28;
ld.shared.f32 %f681, [%r41+256];
add.ftz.f32 %f682, %f681, %f32;
ld.shared.f32 %f683, [%r39+768];
add.ftz.f32 %f684, %f683, %f680;
ld.shared.f32 %f685, [%r41+768];
add.ftz.f32 %f686, %f685, %f682;
ld.shared.f32 %f687, [%r39+1280];
add.ftz.f32 %f688, %f687, %f684;
ld.shared.f32 %f689, [%r41+1280];
add.ftz.f32 %f690, %f689, %f686;
mov.b32 %r1557, %f688;
mov.u32 %r1558, 31;
mov.u32 %r1559, 16;
mov.u32 %r1560, -1;
shfl.sync.bfly.b32 %r1561|%p189, %r1557, %r1559, %r1558, %r1560;
mov.b32 %f691, %r1561;
add.ftz.f32 %f692, %f688, %f691;
mov.b32 %r1562, %f692;
mov.u32 %r1563, 8;
shfl.sync.bfly.b32 %r1564|%p190, %r1562, %r1563, %r1558, %r1560;
mov.b32 %f693, %r1564;
add.ftz.f32 %f694, %f692, %f693;
mov.b32 %r1565, %f694;
mov.u32 %r1566, 4;
shfl.sync.bfly.b32 %r1567|%p191, %r1565, %r1566, %r1558, %r1560;
mov.b32 %f695, %r1567;
add.ftz.f32 %f696, %f694, %f695;
mov.b32 %r1568, %f696;
mov.u32 %r1569, 2;
shfl.sync.bfly.b32 %r1570|%p192, %r1568, %r1569, %r1558, %r1560;
mov.b32 %f697, %r1570;
add.ftz.f32 %f698, %f696, %f697;
mov.b32 %r1571, %f698;
mov.u32 %r1572, 1;
shfl.sync.bfly.b32 %r1573|%p193, %r1571, %r1572, %r1558, %r1560;
mov.b32 %f699, %r1573;
add.ftz.f32 %f60, %f698, %f699;
mov.b32 %r1574, %f690;
shfl.sync.bfly.b32 %r1575|%p194, %r1574, %r1559, %r1558, %r1560;
mov.b32 %f700, %r1575;
add.ftz.f32 %f701, %f690, %f700;
mov.b32 %r1576, %f701;
shfl.sync.bfly.b32 %r1577|%p195, %r1576, %r1563, %r1558, %r1560;
mov.b32 %f702, %r1577;
add.ftz.f32 %f703, %f701, %f702;
mov.b32 %r1578, %f703;
shfl.sync.bfly.b32 %r1579|%p196, %r1578, %r1566, %r1558, %r1560;
mov.b32 %f704, %r1579;
add.ftz.f32 %f705, %f703, %f704;
mov.b32 %r1580, %f705;
shfl.sync.bfly.b32 %r1581|%p197, %r1580, %r1569, %r1558, %r1560;
mov.b32 %f706, %r1581;
add.ftz.f32 %f707, %f705, %f706;
mov.b32 %r1582, %f707;
shfl.sync.bfly.b32 %r1583|%p198, %r1582, %r1572, %r1558, %r1560;
mov.b32 %f708, %r1583;
add.ftz.f32 %f61, %f707, %f708;
add.s32 %r1584, %r4, 2;
setp.ge.u32 %p199, %r1584, %r57;
setp.ne.s32 %p200, %r2, 2;
or.pred %p201, %p200, %p199;
@%p201 bra $L__BB142_42;

add.ftz.f32 %f709, %f5, %f60;
add.ftz.f32 %f710, %f4, %f61;
mul.ftz.f32 %f711, %f710, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f712, %f711;
add.ftz.f32 %f713, %f712, 0f3F800000;
div.approx.ftz.f32 %f714, %f710, %f713;
mul.ftz.f32 %f715, %f709, %f714;
st.global.f32 [%rd92+8], %f715;

$L__BB142_42:
ld.shared.f32 %f716, [%r39+384];
add.ftz.f32 %f717, %f716, %f27;
ld.shared.f32 %f718, [%r41+384];
add.ftz.f32 %f719, %f718, %f31;
ld.shared.f32 %f720, [%r39+896];
add.ftz.f32 %f721, %f720, %f717;
ld.shared.f32 %f722, [%r41+896];
add.ftz.f32 %f723, %f722, %f719;
ld.shared.f32 %f724, [%r39+1408];
add.ftz.f32 %f725, %f724, %f721;
ld.shared.f32 %f726, [%r41+1408];
add.ftz.f32 %f727, %f726, %f723;
mov.b32 %r1585, %f725;
shfl.sync.bfly.b32 %r1589|%p202, %r1585, %r1559, %r1558, %r1560;
mov.b32 %f728, %r1589;
add.ftz.f32 %f729, %f725, %f728;
mov.b32 %r1590, %f729;
shfl.sync.bfly.b32 %r1592|%p203, %r1590, %r1563, %r1558, %r1560;
mov.b32 %f730, %r1592;
add.ftz.f32 %f731, %f729, %f730;
mov.b32 %r1593, %f731;
shfl.sync.bfly.b32 %r1595|%p204, %r1593, %r1566, %r1558, %r1560;
mov.b32 %f732, %r1595;
add.ftz.f32 %f733, %f731, %f732;
mov.b32 %r1596, %f733;
shfl.sync.bfly.b32 %r1598|%p205, %r1596, %r1569, %r1558, %r1560;
mov.b32 %f734, %r1598;
add.ftz.f32 %f735, %f733, %f734;
mov.b32 %r1599, %f735;
shfl.sync.bfly.b32 %r1601|%p206, %r1599, %r1572, %r1558, %r1560;
mov.b32 %f736, %r1601;
add.ftz.f32 %f62, %f735, %f736;
mov.b32 %r1602, %f727;
shfl.sync.bfly.b32 %r1603|%p207, %r1602, %r1559, %r1558, %r1560;
mov.b32 %f737, %r1603;
add.ftz.f32 %f738, %f727, %f737;
mov.b32 %r1604, %f738;
shfl.sync.bfly.b32 %r1605|%p208, %r1604, %r1563, %r1558, %r1560;
mov.b32 %f739, %r1605;
add.ftz.f32 %f740, %f738, %f739;
mov.b32 %r1606, %f740;
shfl.sync.bfly.b32 %r1607|%p209, %r1606, %r1566, %r1558, %r1560;
mov.b32 %f741, %r1607;
add.ftz.f32 %f742, %f740, %f741;
mov.b32 %r1608, %f742;
shfl.sync.bfly.b32 %r1609|%p210, %r1608, %r1569, %r1558, %r1560;
mov.b32 %f743, %r1609;
add.ftz.f32 %f744, %f742, %f743;
mov.b32 %r1610, %f744;
shfl.sync.bfly.b32 %r1611|%p211, %r1610, %r1572, %r1558, %r1560;
mov.b32 %f745, %r1611;
add.ftz.f32 %f63, %f744, %f745;
add.s32 %r1612, %r4, 3;
setp.ge.u32 %p212, %r1612, %r57;
setp.ne.s32 %p213, %r2, 3;
or.pred %p214, %p213, %p212;
@%p214 bra $L__BB142_78;

add.ftz.f32 %f746, %f5, %f62;
add.ftz.f32 %f747, %f4, %f63;
mul.ftz.f32 %f748, %f747, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f749, %f748;
add.ftz.f32 %f750, %f749, 0f3F800000;
div.approx.ftz.f32 %f751, %f747, %f750;
mul.ftz.f32 %f752, %f746, %f751;
st.global.f32 [%rd92+12], %f752;
bra.uni $L__BB142_78;

$L__BB142_25:
ld.shared.f32 %f205, [%r39];
add.ftz.f32 %f206, %f205, %f30;
ld.shared.f32 %f207, [%r39+512];
add.ftz.f32 %f208, %f207, %f206;
ld.shared.f32 %f209, [%r39+1024];
add.ftz.f32 %f210, %f209, %f208;
mov.b32 %r1206, %f210;
mov.u32 %r1207, 31;
mov.u32 %r1208, 16;
mov.u32 %r1209, -1;
shfl.sync.bfly.b32 %r1210|%p22, %r1206, %r1208, %r1207, %r1209;
mov.b32 %f211, %r1210;
add.ftz.f32 %f212, %f210, %f211;
mov.b32 %r1211, %f212;
mov.u32 %r1212, 8;
shfl.sync.bfly.b32 %r1213|%p23, %r1211, %r1212, %r1207, %r1209;
mov.b32 %f213, %r1213;
add.ftz.f32 %f214, %f212, %f213;
mov.b32 %r1214, %f214;
mov.u32 %r1215, 4;
shfl.sync.bfly.b32 %r1216|%p24, %r1214, %r1215, %r1207, %r1209;
mov.b32 %f215, %r1216;
add.ftz.f32 %f216, %f214, %f215;
mov.b32 %r1217, %f216;
mov.u32 %r1218, 2;
shfl.sync.bfly.b32 %r1219|%p25, %r1217, %r1218, %r1207, %r1209;
mov.b32 %f217, %r1219;
add.ftz.f32 %f218, %f216, %f217;
mov.b32 %r1220, %f218;
mov.u32 %r1221, 1;
shfl.sync.bfly.b32 %r1222|%p26, %r1220, %r1221, %r1207, %r1209;
mov.b32 %f219, %r1222;
add.ftz.f32 %f36, %f218, %f219;
setp.ge.u32 %p27, %r4, %r57;
setp.ne.s32 %p28, %r2, 0;
or.pred %p29, %p28, %p27;
@%p29 bra $L__BB142_27;

add.ftz.f32 %f220, %f5, %f36;
st.global.f32 [%rd92], %f220;

$L__BB142_27:
ld.shared.f32 %f221, [%r39+128];
add.ftz.f32 %f222, %f221, %f29;
ld.shared.f32 %f223, [%r39+640];
add.ftz.f32 %f224, %f223, %f222;
ld.shared.f32 %f225, [%r39+1152];
add.ftz.f32 %f226, %f225, %f224;
mov.b32 %r1223, %f226;
shfl.sync.bfly.b32 %r1227|%p30, %r1223, %r1208, %r1207, %r1209;
mov.b32 %f227, %r1227;
add.ftz.f32 %f228, %f226, %f227;
mov.b32 %r1228, %f228;
shfl.sync.bfly.b32 %r1230|%p31, %r1228, %r1212, %r1207, %r1209;
mov.b32 %f229, %r1230;
add.ftz.f32 %f230, %f228, %f229;
mov.b32 %r1231, %f230;
shfl.sync.bfly.b32 %r1233|%p32, %r1231, %r1215, %r1207, %r1209;
mov.b32 %f231, %r1233;
add.ftz.f32 %f232, %f230, %f231;
mov.b32 %r1234, %f232;
shfl.sync.bfly.b32 %r1236|%p33, %r1234, %r1218, %r1207, %r1209;
mov.b32 %f233, %r1236;
add.ftz.f32 %f234, %f232, %f233;
mov.b32 %r1237, %f234;
shfl.sync.bfly.b32 %r1239|%p34, %r1237, %r1221, %r1207, %r1209;
mov.b32 %f235, %r1239;
add.ftz.f32 %f37, %f234, %f235;
add.s32 %r1240, %r4, 1;
setp.ge.u32 %p35, %r1240, %r57;
setp.ne.s32 %p36, %r2, 1;
or.pred %p37, %p36, %p35;
@%p37 bra $L__BB142_29;

add.ftz.f32 %f236, %f5, %f37;
st.global.f32 [%rd92+4], %f236;

$L__BB142_29:
ld.shared.f32 %f237, [%r39+256];
add.ftz.f32 %f238, %f237, %f28;
ld.shared.f32 %f239, [%r39+768];
add.ftz.f32 %f240, %f239, %f238;
ld.shared.f32 %f241, [%r39+1280];
add.ftz.f32 %f242, %f241, %f240;
mov.b32 %r1241, %f242;
mov.u32 %r1242, 31;
mov.u32 %r1243, 16;
mov.u32 %r1244, -1;
shfl.sync.bfly.b32 %r1245|%p38, %r1241, %r1243, %r1242, %r1244;
mov.b32 %f243, %r1245;
add.ftz.f32 %f244, %f242, %f243;
mov.b32 %r1246, %f244;
mov.u32 %r1247, 8;
shfl.sync.bfly.b32 %r1248|%p39, %r1246, %r1247, %r1242, %r1244;
mov.b32 %f245, %r1248;
add.ftz.f32 %f246, %f244, %f245;
mov.b32 %r1249, %f246;
mov.u32 %r1250, 4;
shfl.sync.bfly.b32 %r1251|%p40, %r1249, %r1250, %r1242, %r1244;
mov.b32 %f247, %r1251;
add.ftz.f32 %f248, %f246, %f247;
mov.b32 %r1252, %f248;
mov.u32 %r1253, 2;
shfl.sync.bfly.b32 %r1254|%p41, %r1252, %r1253, %r1242, %r1244;
mov.b32 %f249, %r1254;
add.ftz.f32 %f250, %f248, %f249;
mov.b32 %r1255, %f250;
mov.u32 %r1256, 1;
shfl.sync.bfly.b32 %r1257|%p42, %r1255, %r1256, %r1242, %r1244;
mov.b32 %f251, %r1257;
add.ftz.f32 %f38, %f250, %f251;
add.s32 %r1258, %r4, 2;
setp.ge.u32 %p43, %r1258, %r57;
setp.ne.s32 %p44, %r2, 2;
or.pred %p45, %p44, %p43;
@%p45 bra $L__BB142_31;

add.ftz.f32 %f252, %f5, %f38;
st.global.f32 [%rd92+8], %f252;

$L__BB142_31:
ld.shared.f32 %f253, [%r39+384];
add.ftz.f32 %f254, %f253, %f27;
ld.shared.f32 %f255, [%r39+896];
add.ftz.f32 %f256, %f255, %f254;
ld.shared.f32 %f257, [%r39+1408];
add.ftz.f32 %f258, %f257, %f256;
mov.b32 %r1259, %f258;
shfl.sync.bfly.b32 %r1263|%p46, %r1259, %r1243, %r1242, %r1244;
mov.b32 %f259, %r1263;
add.ftz.f32 %f260, %f258, %f259;
mov.b32 %r1264, %f260;
shfl.sync.bfly.b32 %r1266|%p47, %r1264, %r1247, %r1242, %r1244;
mov.b32 %f261, %r1266;
add.ftz.f32 %f262, %f260, %f261;
mov.b32 %r1267, %f262;
shfl.sync.bfly.b32 %r1269|%p48, %r1267, %r1250, %r1242, %r1244;
mov.b32 %f263, %r1269;
add.ftz.f32 %f264, %f262, %f263;
mov.b32 %r1270, %f264;
shfl.sync.bfly.b32 %r1272|%p49, %r1270, %r1253, %r1242, %r1244;
mov.b32 %f265, %r1272;
add.ftz.f32 %f266, %f264, %f265;
mov.b32 %r1273, %f266;
shfl.sync.bfly.b32 %r1275|%p50, %r1273, %r1256, %r1242, %r1244;
mov.b32 %f267, %r1275;
add.ftz.f32 %f39, %f266, %f267;
add.s32 %r1276, %r4, 3;
setp.ge.u32 %p51, %r1276, %r57;
setp.ne.s32 %p52, %r2, 3;
or.pred %p53, %p52, %p51;
@%p53 bra $L__BB142_78;

add.ftz.f32 %f268, %f5, %f39;
st.global.f32 [%rd92+12], %f268;
bra.uni $L__BB142_78;

$L__BB142_44:
ld.param.u32 %r1847, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p55, %r1847, 3;
@%p55 bra $L__BB142_54;

ld.param.u32 %r1848, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.ne.s32 %p56, %r1848, 6;
@%p56 bra $L__BB142_70;

ld.shared.f32 %f269, [%r39];
add.ftz.f32 %f270, %f269, %f30;
ld.shared.f32 %f271, [%r41];
add.ftz.f32 %f272, %f271, %f34;
ld.shared.f32 %f273, [%r39+512];
add.ftz.f32 %f274, %f273, %f270;
ld.shared.f32 %f275, [%r41+512];
add.ftz.f32 %f276, %f275, %f272;
ld.shared.f32 %f277, [%r39+1024];
add.ftz.f32 %f278, %f277, %f274;
ld.shared.f32 %f279, [%r41+1024];
add.ftz.f32 %f280, %f279, %f276;
mov.b32 %r1280, %f278;
mov.u32 %r1281, 31;
mov.u32 %r1282, 16;
mov.u32 %r1283, -1;
shfl.sync.bfly.b32 %r1284|%p59, %r1280, %r1282, %r1281, %r1283;
mov.b32 %f281, %r1284;
add.ftz.f32 %f282, %f278, %f281;
mov.b32 %r1285, %f282;
mov.u32 %r1286, 8;
shfl.sync.bfly.b32 %r1287|%p60, %r1285, %r1286, %r1281, %r1283;
mov.b32 %f283, %r1287;
add.ftz.f32 %f284, %f282, %f283;
mov.b32 %r1288, %f284;
mov.u32 %r1289, 4;
shfl.sync.bfly.b32 %r1290|%p61, %r1288, %r1289, %r1281, %r1283;
mov.b32 %f285, %r1290;
add.ftz.f32 %f286, %f284, %f285;
mov.b32 %r1291, %f286;
mov.u32 %r1292, 2;
shfl.sync.bfly.b32 %r1293|%p62, %r1291, %r1292, %r1281, %r1283;
mov.b32 %f287, %r1293;
add.ftz.f32 %f288, %f286, %f287;
mov.b32 %r1294, %f288;
mov.u32 %r1295, 1;
shfl.sync.bfly.b32 %r1296|%p63, %r1294, %r1295, %r1281, %r1283;
mov.b32 %f289, %r1296;
add.ftz.f32 %f40, %f288, %f289;
mov.b32 %r1297, %f280;
shfl.sync.bfly.b32 %r1298|%p64, %r1297, %r1282, %r1281, %r1283;
mov.b32 %f290, %r1298;
add.ftz.f32 %f291, %f280, %f290;
mov.b32 %r1299, %f291;
shfl.sync.bfly.b32 %r1300|%p65, %r1299, %r1286, %r1281, %r1283;
mov.b32 %f292, %r1300;
add.ftz.f32 %f293, %f291, %f292;
mov.b32 %r1301, %f293;
shfl.sync.bfly.b32 %r1302|%p66, %r1301, %r1289, %r1281, %r1283;
mov.b32 %f294, %r1302;
add.ftz.f32 %f295, %f293, %f294;
mov.b32 %r1303, %f295;
shfl.sync.bfly.b32 %r1304|%p67, %r1303, %r1292, %r1281, %r1283;
mov.b32 %f296, %r1304;
add.ftz.f32 %f297, %f295, %f296;
mov.b32 %r1305, %f297;
shfl.sync.bfly.b32 %r1306|%p68, %r1305, %r1295, %r1281, %r1283;
mov.b32 %f298, %r1306;
add.ftz.f32 %f41, %f297, %f298;
setp.ge.u32 %p69, %r4, %r57;
setp.ne.s32 %p70, %r2, 0;
or.pred %p71, %p70, %p69;
@%p71 bra $L__BB142_48;

ld.param.f32 %f1053, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f299, %f5, %f40;
add.ftz.f32 %f300, %f4, %f41;
min.ftz.f32 %f301, %f300, %f1053;
min.ftz.f32 %f302, %f299, %f1053;
max.ftz.f32 %f303, %f302, %f35;
mul.ftz.f32 %f304, %f301, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f305, %f304;
add.ftz.f32 %f306, %f305, 0f3F800000;
div.approx.ftz.f32 %f307, %f301, %f306;
mul.ftz.f32 %f308, %f303, %f307;
st.global.f32 [%rd92], %f308;

$L__BB142_48:
ld.shared.f32 %f309, [%r39+128];
add.ftz.f32 %f310, %f309, %f29;
ld.shared.f32 %f311, [%r41+128];
add.ftz.f32 %f312, %f311, %f33;
ld.shared.f32 %f313, [%r39+640];
add.ftz.f32 %f314, %f313, %f310;
ld.shared.f32 %f315, [%r41+640];
add.ftz.f32 %f316, %f315, %f312;
ld.shared.f32 %f317, [%r39+1152];
add.ftz.f32 %f318, %f317, %f314;
ld.shared.f32 %f319, [%r41+1152];
add.ftz.f32 %f320, %f319, %f316;
mov.b32 %r1307, %f318;
shfl.sync.bfly.b32 %r1311|%p72, %r1307, %r1282, %r1281, %r1283;
mov.b32 %f321, %r1311;
add.ftz.f32 %f322, %f318, %f321;
mov.b32 %r1312, %f322;
shfl.sync.bfly.b32 %r1314|%p73, %r1312, %r1286, %r1281, %r1283;
mov.b32 %f323, %r1314;
add.ftz.f32 %f324, %f322, %f323;
mov.b32 %r1315, %f324;
shfl.sync.bfly.b32 %r1317|%p74, %r1315, %r1289, %r1281, %r1283;
mov.b32 %f325, %r1317;
add.ftz.f32 %f326, %f324, %f325;
mov.b32 %r1318, %f326;
shfl.sync.bfly.b32 %r1320|%p75, %r1318, %r1292, %r1281, %r1283;
mov.b32 %f327, %r1320;
add.ftz.f32 %f328, %f326, %f327;
mov.b32 %r1321, %f328;
shfl.sync.bfly.b32 %r1323|%p76, %r1321, %r1295, %r1281, %r1283;
mov.b32 %f329, %r1323;
add.ftz.f32 %f42, %f328, %f329;
mov.b32 %r1324, %f320;
shfl.sync.bfly.b32 %r1325|%p77, %r1324, %r1282, %r1281, %r1283;
mov.b32 %f330, %r1325;
add.ftz.f32 %f331, %f320, %f330;
mov.b32 %r1326, %f331;
shfl.sync.bfly.b32 %r1327|%p78, %r1326, %r1286, %r1281, %r1283;
mov.b32 %f332, %r1327;
add.ftz.f32 %f333, %f331, %f332;
mov.b32 %r1328, %f333;
shfl.sync.bfly.b32 %r1329|%p79, %r1328, %r1289, %r1281, %r1283;
mov.b32 %f334, %r1329;
add.ftz.f32 %f335, %f333, %f334;
mov.b32 %r1330, %f335;
shfl.sync.bfly.b32 %r1331|%p80, %r1330, %r1292, %r1281, %r1283;
mov.b32 %f336, %r1331;
add.ftz.f32 %f337, %f335, %f336;
mov.b32 %r1332, %f337;
shfl.sync.bfly.b32 %r1333|%p81, %r1332, %r1295, %r1281, %r1283;
mov.b32 %f338, %r1333;
add.ftz.f32 %f43, %f337, %f338;
add.s32 %r1334, %r4, 1;
setp.ge.u32 %p82, %r1334, %r57;
setp.ne.s32 %p83, %r2, 1;
or.pred %p84, %p83, %p82;
@%p84 bra $L__BB142_50;

ld.param.f32 %f1052, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f339, %f5, %f42;
add.ftz.f32 %f340, %f4, %f43;
min.ftz.f32 %f341, %f340, %f1052;
min.ftz.f32 %f342, %f339, %f1052;
max.ftz.f32 %f343, %f342, %f35;
mul.ftz.f32 %f344, %f341, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f345, %f344;
add.ftz.f32 %f346, %f345, 0f3F800000;
div.approx.ftz.f32 %f347, %f341, %f346;
mul.ftz.f32 %f348, %f343, %f347;
st.global.f32 [%rd92+4], %f348;

$L__BB142_50:
ld.shared.f32 %f349, [%r39+256];
add.ftz.f32 %f350, %f349, %f28;
ld.shared.f32 %f351, [%r41+256];
add.ftz.f32 %f352, %f351, %f32;
ld.shared.f32 %f353, [%r39+768];
add.ftz.f32 %f354, %f353, %f350;
ld.shared.f32 %f355, [%r41+768];
add.ftz.f32 %f356, %f355, %f352;
ld.shared.f32 %f357, [%r39+1280];
add.ftz.f32 %f358, %f357, %f354;
ld.shared.f32 %f359, [%r41+1280];
add.ftz.f32 %f360, %f359, %f356;
mov.b32 %r1335, %f358;
mov.u32 %r1336, 31;
mov.u32 %r1337, 16;
mov.u32 %r1338, -1;
shfl.sync.bfly.b32 %r1339|%p85, %r1335, %r1337, %r1336, %r1338;
mov.b32 %f361, %r1339;
add.ftz.f32 %f362, %f358, %f361;
mov.b32 %r1340, %f362;
mov.u32 %r1341, 8;
shfl.sync.bfly.b32 %r1342|%p86, %r1340, %r1341, %r1336, %r1338;
mov.b32 %f363, %r1342;
add.ftz.f32 %f364, %f362, %f363;
mov.b32 %r1343, %f364;
mov.u32 %r1344, 4;
shfl.sync.bfly.b32 %r1345|%p87, %r1343, %r1344, %r1336, %r1338;
mov.b32 %f365, %r1345;
add.ftz.f32 %f366, %f364, %f365;
mov.b32 %r1346, %f366;
mov.u32 %r1347, 2;
shfl.sync.bfly.b32 %r1348|%p88, %r1346, %r1347, %r1336, %r1338;
mov.b32 %f367, %r1348;
add.ftz.f32 %f368, %f366, %f367;
mov.b32 %r1349, %f368;
mov.u32 %r1350, 1;
shfl.sync.bfly.b32 %r1351|%p89, %r1349, %r1350, %r1336, %r1338;
mov.b32 %f369, %r1351;
add.ftz.f32 %f44, %f368, %f369;
mov.b32 %r1352, %f360;
shfl.sync.bfly.b32 %r1353|%p90, %r1352, %r1337, %r1336, %r1338;
mov.b32 %f370, %r1353;
add.ftz.f32 %f371, %f360, %f370;
mov.b32 %r1354, %f371;
shfl.sync.bfly.b32 %r1355|%p91, %r1354, %r1341, %r1336, %r1338;
mov.b32 %f372, %r1355;
add.ftz.f32 %f373, %f371, %f372;
mov.b32 %r1356, %f373;
shfl.sync.bfly.b32 %r1357|%p92, %r1356, %r1344, %r1336, %r1338;
mov.b32 %f374, %r1357;
add.ftz.f32 %f375, %f373, %f374;
mov.b32 %r1358, %f375;
shfl.sync.bfly.b32 %r1359|%p93, %r1358, %r1347, %r1336, %r1338;
mov.b32 %f376, %r1359;
add.ftz.f32 %f377, %f375, %f376;
mov.b32 %r1360, %f377;
shfl.sync.bfly.b32 %r1361|%p94, %r1360, %r1350, %r1336, %r1338;
mov.b32 %f378, %r1361;
add.ftz.f32 %f45, %f377, %f378;
add.s32 %r1362, %r4, 2;
setp.ge.u32 %p95, %r1362, %r57;
setp.ne.s32 %p96, %r2, 2;
or.pred %p97, %p96, %p95;
@%p97 bra $L__BB142_52;

ld.param.f32 %f1051, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f379, %f5, %f44;
add.ftz.f32 %f380, %f4, %f45;
min.ftz.f32 %f381, %f380, %f1051;
min.ftz.f32 %f382, %f379, %f1051;
max.ftz.f32 %f383, %f382, %f35;
mul.ftz.f32 %f384, %f381, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f385, %f384;
add.ftz.f32 %f386, %f385, 0f3F800000;
div.approx.ftz.f32 %f387, %f381, %f386;
mul.ftz.f32 %f388, %f383, %f387;
st.global.f32 [%rd92+8], %f388;

$L__BB142_52:
ld.shared.f32 %f389, [%r39+384];
add.ftz.f32 %f390, %f389, %f27;
ld.shared.f32 %f391, [%r41+384];
add.ftz.f32 %f392, %f391, %f31;
ld.shared.f32 %f393, [%r39+896];
add.ftz.f32 %f394, %f393, %f390;
ld.shared.f32 %f395, [%r41+896];
add.ftz.f32 %f396, %f395, %f392;
ld.shared.f32 %f397, [%r39+1408];
add.ftz.f32 %f398, %f397, %f394;
ld.shared.f32 %f399, [%r41+1408];
add.ftz.f32 %f400, %f399, %f396;
mov.b32 %r1363, %f398;
shfl.sync.bfly.b32 %r1367|%p98, %r1363, %r1337, %r1336, %r1338;
mov.b32 %f401, %r1367;
add.ftz.f32 %f402, %f398, %f401;
mov.b32 %r1368, %f402;
shfl.sync.bfly.b32 %r1370|%p99, %r1368, %r1341, %r1336, %r1338;
mov.b32 %f403, %r1370;
add.ftz.f32 %f404, %f402, %f403;
mov.b32 %r1371, %f404;
shfl.sync.bfly.b32 %r1373|%p100, %r1371, %r1344, %r1336, %r1338;
mov.b32 %f405, %r1373;
add.ftz.f32 %f406, %f404, %f405;
mov.b32 %r1374, %f406;
shfl.sync.bfly.b32 %r1376|%p101, %r1374, %r1347, %r1336, %r1338;
mov.b32 %f407, %r1376;
add.ftz.f32 %f408, %f406, %f407;
mov.b32 %r1377, %f408;
shfl.sync.bfly.b32 %r1379|%p102, %r1377, %r1350, %r1336, %r1338;
mov.b32 %f409, %r1379;
add.ftz.f32 %f46, %f408, %f409;
mov.b32 %r1380, %f400;
shfl.sync.bfly.b32 %r1381|%p103, %r1380, %r1337, %r1336, %r1338;
mov.b32 %f410, %r1381;
add.ftz.f32 %f411, %f400, %f410;
mov.b32 %r1382, %f411;
shfl.sync.bfly.b32 %r1383|%p104, %r1382, %r1341, %r1336, %r1338;
mov.b32 %f412, %r1383;
add.ftz.f32 %f413, %f411, %f412;
mov.b32 %r1384, %f413;
shfl.sync.bfly.b32 %r1385|%p105, %r1384, %r1344, %r1336, %r1338;
mov.b32 %f414, %r1385;
add.ftz.f32 %f415, %f413, %f414;
mov.b32 %r1386, %f415;
shfl.sync.bfly.b32 %r1387|%p106, %r1386, %r1347, %r1336, %r1338;
mov.b32 %f416, %r1387;
add.ftz.f32 %f417, %f415, %f416;
mov.b32 %r1388, %f417;
shfl.sync.bfly.b32 %r1389|%p107, %r1388, %r1350, %r1336, %r1338;
mov.b32 %f418, %r1389;
add.ftz.f32 %f47, %f417, %f418;
add.s32 %r1390, %r4, 3;
setp.ge.u32 %p108, %r1390, %r57;
setp.ne.s32 %p109, %r2, 3;
or.pred %p110, %p109, %p108;
@%p110 bra $L__BB142_78;

ld.param.f32 %f1050, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f419, %f5, %f46;
add.ftz.f32 %f420, %f4, %f47;
min.ftz.f32 %f421, %f420, %f1050;
min.ftz.f32 %f422, %f419, %f1050;
max.ftz.f32 %f423, %f422, %f35;
mul.ftz.f32 %f424, %f421, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f425, %f424;
add.ftz.f32 %f426, %f425, 0f3F800000;
div.approx.ftz.f32 %f427, %f421, %f426;
mul.ftz.f32 %f428, %f423, %f427;
st.global.f32 [%rd92+12], %f428;
bra.uni $L__BB142_78;

$L__BB142_70:
ld.shared.f32 %f917, [%r39];
add.ftz.f32 %f918, %f917, %f30;
ld.shared.f32 %f919, [%r41];
add.ftz.f32 %f920, %f919, %f34;
ld.shared.f32 %f921, [%r39+512];
add.ftz.f32 %f922, %f921, %f918;
ld.shared.f32 %f923, [%r41+512];
add.ftz.f32 %f924, %f923, %f920;
ld.shared.f32 %f925, [%r39+1024];
add.ftz.f32 %f926, %f925, %f922;
ld.shared.f32 %f927, [%r41+1024];
add.ftz.f32 %f928, %f927, %f924;
mov.b32 %r1724, %f926;
mov.u32 %r1725, 31;
mov.u32 %r1726, 16;
mov.u32 %r1727, -1;
shfl.sync.bfly.b32 %r1728|%p267, %r1724, %r1726, %r1725, %r1727;
mov.b32 %f929, %r1728;
add.ftz.f32 %f930, %f926, %f929;
mov.b32 %r1729, %f930;
mov.u32 %r1730, 8;
shfl.sync.bfly.b32 %r1731|%p268, %r1729, %r1730, %r1725, %r1727;
mov.b32 %f931, %r1731;
add.ftz.f32 %f932, %f930, %f931;
mov.b32 %r1732, %f932;
mov.u32 %r1733, 4;
shfl.sync.bfly.b32 %r1734|%p269, %r1732, %r1733, %r1725, %r1727;
mov.b32 %f933, %r1734;
add.ftz.f32 %f934, %f932, %f933;
mov.b32 %r1735, %f934;
mov.u32 %r1736, 2;
shfl.sync.bfly.b32 %r1737|%p270, %r1735, %r1736, %r1725, %r1727;
mov.b32 %f935, %r1737;
add.ftz.f32 %f936, %f934, %f935;
mov.b32 %r1738, %f936;
mov.u32 %r1739, 1;
shfl.sync.bfly.b32 %r1740|%p271, %r1738, %r1739, %r1725, %r1727;
mov.b32 %f937, %r1740;
add.ftz.f32 %f72, %f936, %f937;
mov.b32 %r1741, %f928;
shfl.sync.bfly.b32 %r1742|%p272, %r1741, %r1726, %r1725, %r1727;
mov.b32 %f938, %r1742;
add.ftz.f32 %f939, %f928, %f938;
mov.b32 %r1743, %f939;
shfl.sync.bfly.b32 %r1744|%p273, %r1743, %r1730, %r1725, %r1727;
mov.b32 %f940, %r1744;
add.ftz.f32 %f941, %f939, %f940;
mov.b32 %r1745, %f941;
shfl.sync.bfly.b32 %r1746|%p274, %r1745, %r1733, %r1725, %r1727;
mov.b32 %f942, %r1746;
add.ftz.f32 %f943, %f941, %f942;
mov.b32 %r1747, %f943;
shfl.sync.bfly.b32 %r1748|%p275, %r1747, %r1736, %r1725, %r1727;
mov.b32 %f944, %r1748;
add.ftz.f32 %f945, %f943, %f944;
mov.b32 %r1749, %f945;
shfl.sync.bfly.b32 %r1750|%p276, %r1749, %r1739, %r1725, %r1727;
mov.b32 %f946, %r1750;
add.ftz.f32 %f73, %f945, %f946;
setp.ge.u32 %p277, %r4, %r57;
setp.ne.s32 %p278, %r2, 0;
or.pred %p279, %p278, %p277;
@%p279 bra $L__BB142_72;

add.ftz.f32 %f947, %f5, %f72;
add.ftz.f32 %f948, %f4, %f73;
mul.ftz.f32 %f949, %f947, %f948;
st.global.f32 [%rd92], %f949;

$L__BB142_72:
ld.shared.f32 %f950, [%r39+128];
add.ftz.f32 %f951, %f950, %f29;
ld.shared.f32 %f952, [%r41+128];
add.ftz.f32 %f953, %f952, %f33;
ld.shared.f32 %f954, [%r39+640];
add.ftz.f32 %f955, %f954, %f951;
ld.shared.f32 %f956, [%r41+640];
add.ftz.f32 %f957, %f956, %f953;
ld.shared.f32 %f958, [%r39+1152];
add.ftz.f32 %f959, %f958, %f955;
ld.shared.f32 %f960, [%r41+1152];
add.ftz.f32 %f961, %f960, %f957;
mov.b32 %r1751, %f959;
shfl.sync.bfly.b32 %r1755|%p280, %r1751, %r1726, %r1725, %r1727;
mov.b32 %f962, %r1755;
add.ftz.f32 %f963, %f959, %f962;
mov.b32 %r1756, %f963;
shfl.sync.bfly.b32 %r1758|%p281, %r1756, %r1730, %r1725, %r1727;
mov.b32 %f964, %r1758;
add.ftz.f32 %f965, %f963, %f964;
mov.b32 %r1759, %f965;
shfl.sync.bfly.b32 %r1761|%p282, %r1759, %r1733, %r1725, %r1727;
mov.b32 %f966, %r1761;
add.ftz.f32 %f967, %f965, %f966;
mov.b32 %r1762, %f967;
shfl.sync.bfly.b32 %r1764|%p283, %r1762, %r1736, %r1725, %r1727;
mov.b32 %f968, %r1764;
add.ftz.f32 %f969, %f967, %f968;
mov.b32 %r1765, %f969;
shfl.sync.bfly.b32 %r1767|%p284, %r1765, %r1739, %r1725, %r1727;
mov.b32 %f970, %r1767;
add.ftz.f32 %f74, %f969, %f970;
mov.b32 %r1768, %f961;
shfl.sync.bfly.b32 %r1769|%p285, %r1768, %r1726, %r1725, %r1727;
mov.b32 %f971, %r1769;
add.ftz.f32 %f972, %f961, %f971;
mov.b32 %r1770, %f972;
shfl.sync.bfly.b32 %r1771|%p286, %r1770, %r1730, %r1725, %r1727;
mov.b32 %f973, %r1771;
add.ftz.f32 %f974, %f972, %f973;
mov.b32 %r1772, %f974;
shfl.sync.bfly.b32 %r1773|%p287, %r1772, %r1733, %r1725, %r1727;
mov.b32 %f975, %r1773;
add.ftz.f32 %f976, %f974, %f975;
mov.b32 %r1774, %f976;
shfl.sync.bfly.b32 %r1775|%p288, %r1774, %r1736, %r1725, %r1727;
mov.b32 %f977, %r1775;
add.ftz.f32 %f978, %f976, %f977;
mov.b32 %r1776, %f978;
shfl.sync.bfly.b32 %r1777|%p289, %r1776, %r1739, %r1725, %r1727;
mov.b32 %f979, %r1777;
add.ftz.f32 %f75, %f978, %f979;
add.s32 %r1778, %r4, 1;
setp.ge.u32 %p290, %r1778, %r57;
setp.ne.s32 %p291, %r2, 1;
or.pred %p292, %p291, %p290;
@%p292 bra $L__BB142_74;

add.ftz.f32 %f980, %f5, %f74;
add.ftz.f32 %f981, %f4, %f75;
mul.ftz.f32 %f982, %f980, %f981;
st.global.f32 [%rd92+4], %f982;

$L__BB142_74:
ld.shared.f32 %f983, [%r39+256];
add.ftz.f32 %f984, %f983, %f28;
ld.shared.f32 %f985, [%r41+256];
add.ftz.f32 %f986, %f985, %f32;
ld.shared.f32 %f987, [%r39+768];
add.ftz.f32 %f988, %f987, %f984;
ld.shared.f32 %f989, [%r41+768];
add.ftz.f32 %f990, %f989, %f986;
ld.shared.f32 %f991, [%r39+1280];
add.ftz.f32 %f992, %f991, %f988;
ld.shared.f32 %f993, [%r41+1280];
add.ftz.f32 %f994, %f993, %f990;
mov.b32 %r1779, %f992;
mov.u32 %r1780, 31;
mov.u32 %r1781, 16;
mov.u32 %r1782, -1;
shfl.sync.bfly.b32 %r1783|%p293, %r1779, %r1781, %r1780, %r1782;
mov.b32 %f995, %r1783;
add.ftz.f32 %f996, %f992, %f995;
mov.b32 %r1784, %f996;
mov.u32 %r1785, 8;
shfl.sync.bfly.b32 %r1786|%p294, %r1784, %r1785, %r1780, %r1782;
mov.b32 %f997, %r1786;
add.ftz.f32 %f998, %f996, %f997;
mov.b32 %r1787, %f998;
mov.u32 %r1788, 4;
shfl.sync.bfly.b32 %r1789|%p295, %r1787, %r1788, %r1780, %r1782;
mov.b32 %f999, %r1789;
add.ftz.f32 %f1000, %f998, %f999;
mov.b32 %r1790, %f1000;
mov.u32 %r1791, 2;
shfl.sync.bfly.b32 %r1792|%p296, %r1790, %r1791, %r1780, %r1782;
mov.b32 %f1001, %r1792;
add.ftz.f32 %f1002, %f1000, %f1001;
mov.b32 %r1793, %f1002;
mov.u32 %r1794, 1;
shfl.sync.bfly.b32 %r1795|%p297, %r1793, %r1794, %r1780, %r1782;
mov.b32 %f1003, %r1795;
add.ftz.f32 %f76, %f1002, %f1003;
mov.b32 %r1796, %f994;
shfl.sync.bfly.b32 %r1797|%p298, %r1796, %r1781, %r1780, %r1782;
mov.b32 %f1004, %r1797;
add.ftz.f32 %f1005, %f994, %f1004;
mov.b32 %r1798, %f1005;
shfl.sync.bfly.b32 %r1799|%p299, %r1798, %r1785, %r1780, %r1782;
mov.b32 %f1006, %r1799;
add.ftz.f32 %f1007, %f1005, %f1006;
mov.b32 %r1800, %f1007;
shfl.sync.bfly.b32 %r1801|%p300, %r1800, %r1788, %r1780, %r1782;
mov.b32 %f1008, %r1801;
add.ftz.f32 %f1009, %f1007, %f1008;
mov.b32 %r1802, %f1009;
shfl.sync.bfly.b32 %r1803|%p301, %r1802, %r1791, %r1780, %r1782;
mov.b32 %f1010, %r1803;
add.ftz.f32 %f1011, %f1009, %f1010;
mov.b32 %r1804, %f1011;
shfl.sync.bfly.b32 %r1805|%p302, %r1804, %r1794, %r1780, %r1782;
mov.b32 %f1012, %r1805;
add.ftz.f32 %f77, %f1011, %f1012;
add.s32 %r1806, %r4, 2;
setp.ge.u32 %p303, %r1806, %r57;
setp.ne.s32 %p304, %r2, 2;
or.pred %p305, %p304, %p303;
@%p305 bra $L__BB142_76;

add.ftz.f32 %f1013, %f5, %f76;
add.ftz.f32 %f1014, %f4, %f77;
mul.ftz.f32 %f1015, %f1013, %f1014;
st.global.f32 [%rd92+8], %f1015;

$L__BB142_76:
ld.shared.f32 %f1016, [%r39+384];
add.ftz.f32 %f1017, %f1016, %f27;
ld.shared.f32 %f1018, [%r41+384];
add.ftz.f32 %f1019, %f1018, %f31;
ld.shared.f32 %f1020, [%r39+896];
add.ftz.f32 %f1021, %f1020, %f1017;
ld.shared.f32 %f1022, [%r41+896];
add.ftz.f32 %f1023, %f1022, %f1019;
ld.shared.f32 %f1024, [%r39+1408];
add.ftz.f32 %f1025, %f1024, %f1021;
ld.shared.f32 %f1026, [%r41+1408];
add.ftz.f32 %f1027, %f1026, %f1023;
mov.b32 %r1807, %f1025;
shfl.sync.bfly.b32 %r1811|%p306, %r1807, %r1781, %r1780, %r1782;
mov.b32 %f1028, %r1811;
add.ftz.f32 %f1029, %f1025, %f1028;
mov.b32 %r1812, %f1029;
shfl.sync.bfly.b32 %r1814|%p307, %r1812, %r1785, %r1780, %r1782;
mov.b32 %f1030, %r1814;
add.ftz.f32 %f1031, %f1029, %f1030;
mov.b32 %r1815, %f1031;
shfl.sync.bfly.b32 %r1817|%p308, %r1815, %r1788, %r1780, %r1782;
mov.b32 %f1032, %r1817;
add.ftz.f32 %f1033, %f1031, %f1032;
mov.b32 %r1818, %f1033;
shfl.sync.bfly.b32 %r1820|%p309, %r1818, %r1791, %r1780, %r1782;
mov.b32 %f1034, %r1820;
add.ftz.f32 %f1035, %f1033, %f1034;
mov.b32 %r1821, %f1035;
shfl.sync.bfly.b32 %r1823|%p310, %r1821, %r1794, %r1780, %r1782;
mov.b32 %f1036, %r1823;
add.ftz.f32 %f78, %f1035, %f1036;
mov.b32 %r1824, %f1027;
shfl.sync.bfly.b32 %r1825|%p311, %r1824, %r1781, %r1780, %r1782;
mov.b32 %f1037, %r1825;
add.ftz.f32 %f1038, %f1027, %f1037;
mov.b32 %r1826, %f1038;
shfl.sync.bfly.b32 %r1827|%p312, %r1826, %r1785, %r1780, %r1782;
mov.b32 %f1039, %r1827;
add.ftz.f32 %f1040, %f1038, %f1039;
mov.b32 %r1828, %f1040;
shfl.sync.bfly.b32 %r1829|%p313, %r1828, %r1788, %r1780, %r1782;
mov.b32 %f1041, %r1829;
add.ftz.f32 %f1042, %f1040, %f1041;
mov.b32 %r1830, %f1042;
shfl.sync.bfly.b32 %r1831|%p314, %r1830, %r1791, %r1780, %r1782;
mov.b32 %f1043, %r1831;
add.ftz.f32 %f1044, %f1042, %f1043;
mov.b32 %r1832, %f1044;
shfl.sync.bfly.b32 %r1833|%p315, %r1832, %r1794, %r1780, %r1782;
mov.b32 %f1045, %r1833;
add.ftz.f32 %f79, %f1044, %f1045;
add.s32 %r1834, %r4, 3;
setp.ge.u32 %p316, %r1834, %r57;
setp.ne.s32 %p317, %r2, 3;
or.pred %p318, %p317, %p316;
@%p318 bra $L__BB142_78;

add.ftz.f32 %f1046, %f5, %f78;
add.ftz.f32 %f1047, %f4, %f79;
mul.ftz.f32 %f1048, %f1046, %f1047;
st.global.f32 [%rd92+12], %f1048;
bra.uni $L__BB142_78;

$L__BB142_62:
ld.shared.f32 %f753, [%r39];
add.ftz.f32 %f754, %f753, %f30;
ld.shared.f32 %f755, [%r41];
add.ftz.f32 %f756, %f755, %f34;
ld.shared.f32 %f757, [%r39+512];
add.ftz.f32 %f758, %f757, %f754;
ld.shared.f32 %f759, [%r41+512];
add.ftz.f32 %f760, %f759, %f756;
ld.shared.f32 %f761, [%r39+1024];
add.ftz.f32 %f762, %f761, %f758;
ld.shared.f32 %f763, [%r41+1024];
add.ftz.f32 %f764, %f763, %f760;
mov.b32 %r1613, %f762;
mov.u32 %r1614, 31;
mov.u32 %r1615, 16;
mov.u32 %r1616, -1;
shfl.sync.bfly.b32 %r1617|%p215, %r1613, %r1615, %r1614, %r1616;
mov.b32 %f765, %r1617;
add.ftz.f32 %f766, %f762, %f765;
mov.b32 %r1618, %f766;
mov.u32 %r1619, 8;
shfl.sync.bfly.b32 %r1620|%p216, %r1618, %r1619, %r1614, %r1616;
mov.b32 %f767, %r1620;
add.ftz.f32 %f768, %f766, %f767;
mov.b32 %r1621, %f768;
mov.u32 %r1622, 4;
shfl.sync.bfly.b32 %r1623|%p217, %r1621, %r1622, %r1614, %r1616;
mov.b32 %f769, %r1623;
add.ftz.f32 %f770, %f768, %f769;
mov.b32 %r1624, %f770;
mov.u32 %r1625, 2;
shfl.sync.bfly.b32 %r1626|%p218, %r1624, %r1625, %r1614, %r1616;
mov.b32 %f771, %r1626;
add.ftz.f32 %f772, %f770, %f771;
mov.b32 %r1627, %f772;
mov.u32 %r1628, 1;
shfl.sync.bfly.b32 %r1629|%p219, %r1627, %r1628, %r1614, %r1616;
mov.b32 %f773, %r1629;
add.ftz.f32 %f64, %f772, %f773;
mov.b32 %r1630, %f764;
shfl.sync.bfly.b32 %r1631|%p220, %r1630, %r1615, %r1614, %r1616;
mov.b32 %f774, %r1631;
add.ftz.f32 %f775, %f764, %f774;
mov.b32 %r1632, %f775;
shfl.sync.bfly.b32 %r1633|%p221, %r1632, %r1619, %r1614, %r1616;
mov.b32 %f776, %r1633;
add.ftz.f32 %f777, %f775, %f776;
mov.b32 %r1634, %f777;
shfl.sync.bfly.b32 %r1635|%p222, %r1634, %r1622, %r1614, %r1616;
mov.b32 %f778, %r1635;
add.ftz.f32 %f779, %f777, %f778;
mov.b32 %r1636, %f779;
shfl.sync.bfly.b32 %r1637|%p223, %r1636, %r1625, %r1614, %r1616;
mov.b32 %f780, %r1637;
add.ftz.f32 %f781, %f779, %f780;
mov.b32 %r1638, %f781;
shfl.sync.bfly.b32 %r1639|%p224, %r1638, %r1628, %r1614, %r1616;
mov.b32 %f782, %r1639;
add.ftz.f32 %f65, %f781, %f782;
setp.ge.u32 %p225, %r4, %r57;
setp.ne.s32 %p226, %r2, 0;
or.pred %p227, %p226, %p225;
@%p227 bra $L__BB142_64;

add.ftz.f32 %f783, %f5, %f64;
add.ftz.f32 %f784, %f4, %f65;
mul.ftz.f32 %f785, %f784, 0f3F000000;
mul.ftz.f32 %f786, %f784, 0f3F4C422A;
mul.ftz.f32 %f787, %f784, 0f3D372713;
fma.rn.ftz.f32 %f788, %f784, %f787, 0f3F800000;
mul.ftz.f32 %f789, %f786, %f788;
tanh.approx.f32 %f790, %f789;
add.ftz.f32 %f791, %f790, 0f3F800000;
mul.ftz.f32 %f792, %f785, %f791;
mul.ftz.f32 %f793, %f783, %f792;
st.global.f32 [%rd92], %f793;

$L__BB142_64:
ld.shared.f32 %f794, [%r39+128];
add.ftz.f32 %f795, %f794, %f29;
ld.shared.f32 %f796, [%r41+128];
add.ftz.f32 %f797, %f796, %f33;
ld.shared.f32 %f798, [%r39+640];
add.ftz.f32 %f799, %f798, %f795;
ld.shared.f32 %f800, [%r41+640];
add.ftz.f32 %f801, %f800, %f797;
ld.shared.f32 %f802, [%r39+1152];
add.ftz.f32 %f803, %f802, %f799;
ld.shared.f32 %f804, [%r41+1152];
add.ftz.f32 %f805, %f804, %f801;
mov.b32 %r1640, %f803;
shfl.sync.bfly.b32 %r1644|%p228, %r1640, %r1615, %r1614, %r1616;
mov.b32 %f806, %r1644;
add.ftz.f32 %f807, %f803, %f806;
mov.b32 %r1645, %f807;
shfl.sync.bfly.b32 %r1647|%p229, %r1645, %r1619, %r1614, %r1616;
mov.b32 %f808, %r1647;
add.ftz.f32 %f809, %f807, %f808;
mov.b32 %r1648, %f809;
shfl.sync.bfly.b32 %r1650|%p230, %r1648, %r1622, %r1614, %r1616;
mov.b32 %f810, %r1650;
add.ftz.f32 %f811, %f809, %f810;
mov.b32 %r1651, %f811;
shfl.sync.bfly.b32 %r1653|%p231, %r1651, %r1625, %r1614, %r1616;
mov.b32 %f812, %r1653;
add.ftz.f32 %f813, %f811, %f812;
mov.b32 %r1654, %f813;
shfl.sync.bfly.b32 %r1656|%p232, %r1654, %r1628, %r1614, %r1616;
mov.b32 %f814, %r1656;
add.ftz.f32 %f66, %f813, %f814;
mov.b32 %r1657, %f805;
shfl.sync.bfly.b32 %r1658|%p233, %r1657, %r1615, %r1614, %r1616;
mov.b32 %f815, %r1658;
add.ftz.f32 %f816, %f805, %f815;
mov.b32 %r1659, %f816;
shfl.sync.bfly.b32 %r1660|%p234, %r1659, %r1619, %r1614, %r1616;
mov.b32 %f817, %r1660;
add.ftz.f32 %f818, %f816, %f817;
mov.b32 %r1661, %f818;
shfl.sync.bfly.b32 %r1662|%p235, %r1661, %r1622, %r1614, %r1616;
mov.b32 %f819, %r1662;
add.ftz.f32 %f820, %f818, %f819;
mov.b32 %r1663, %f820;
shfl.sync.bfly.b32 %r1664|%p236, %r1663, %r1625, %r1614, %r1616;
mov.b32 %f821, %r1664;
add.ftz.f32 %f822, %f820, %f821;
mov.b32 %r1665, %f822;
shfl.sync.bfly.b32 %r1666|%p237, %r1665, %r1628, %r1614, %r1616;
mov.b32 %f823, %r1666;
add.ftz.f32 %f67, %f822, %f823;
add.s32 %r1667, %r4, 1;
setp.ge.u32 %p238, %r1667, %r57;
setp.ne.s32 %p239, %r2, 1;
or.pred %p240, %p239, %p238;
@%p240 bra $L__BB142_66;

add.ftz.f32 %f824, %f5, %f66;
add.ftz.f32 %f825, %f4, %f67;
mul.ftz.f32 %f826, %f825, 0f3F000000;
mul.ftz.f32 %f827, %f825, 0f3F4C422A;
mul.ftz.f32 %f828, %f825, 0f3D372713;
fma.rn.ftz.f32 %f829, %f825, %f828, 0f3F800000;
mul.ftz.f32 %f830, %f827, %f829;
tanh.approx.f32 %f831, %f830;
add.ftz.f32 %f832, %f831, 0f3F800000;
mul.ftz.f32 %f833, %f826, %f832;
mul.ftz.f32 %f834, %f824, %f833;
st.global.f32 [%rd92+4], %f834;

$L__BB142_66:
ld.shared.f32 %f835, [%r39+256];
add.ftz.f32 %f836, %f835, %f28;
ld.shared.f32 %f837, [%r41+256];
add.ftz.f32 %f838, %f837, %f32;
ld.shared.f32 %f839, [%r39+768];
add.ftz.f32 %f840, %f839, %f836;
ld.shared.f32 %f841, [%r41+768];
add.ftz.f32 %f842, %f841, %f838;
ld.shared.f32 %f843, [%r39+1280];
add.ftz.f32 %f844, %f843, %f840;
ld.shared.f32 %f845, [%r41+1280];
add.ftz.f32 %f846, %f845, %f842;
mov.b32 %r1668, %f844;
mov.u32 %r1669, 31;
mov.u32 %r1670, 16;
mov.u32 %r1671, -1;
shfl.sync.bfly.b32 %r1672|%p241, %r1668, %r1670, %r1669, %r1671;
mov.b32 %f847, %r1672;
add.ftz.f32 %f848, %f844, %f847;
mov.b32 %r1673, %f848;
mov.u32 %r1674, 8;
shfl.sync.bfly.b32 %r1675|%p242, %r1673, %r1674, %r1669, %r1671;
mov.b32 %f849, %r1675;
add.ftz.f32 %f850, %f848, %f849;
mov.b32 %r1676, %f850;
mov.u32 %r1677, 4;
shfl.sync.bfly.b32 %r1678|%p243, %r1676, %r1677, %r1669, %r1671;
mov.b32 %f851, %r1678;
add.ftz.f32 %f852, %f850, %f851;
mov.b32 %r1679, %f852;
mov.u32 %r1680, 2;
shfl.sync.bfly.b32 %r1681|%p244, %r1679, %r1680, %r1669, %r1671;
mov.b32 %f853, %r1681;
add.ftz.f32 %f854, %f852, %f853;
mov.b32 %r1682, %f854;
mov.u32 %r1683, 1;
shfl.sync.bfly.b32 %r1684|%p245, %r1682, %r1683, %r1669, %r1671;
mov.b32 %f855, %r1684;
add.ftz.f32 %f68, %f854, %f855;
mov.b32 %r1685, %f846;
shfl.sync.bfly.b32 %r1686|%p246, %r1685, %r1670, %r1669, %r1671;
mov.b32 %f856, %r1686;
add.ftz.f32 %f857, %f846, %f856;
mov.b32 %r1687, %f857;
shfl.sync.bfly.b32 %r1688|%p247, %r1687, %r1674, %r1669, %r1671;
mov.b32 %f858, %r1688;
add.ftz.f32 %f859, %f857, %f858;
mov.b32 %r1689, %f859;
shfl.sync.bfly.b32 %r1690|%p248, %r1689, %r1677, %r1669, %r1671;
mov.b32 %f860, %r1690;
add.ftz.f32 %f861, %f859, %f860;
mov.b32 %r1691, %f861;
shfl.sync.bfly.b32 %r1692|%p249, %r1691, %r1680, %r1669, %r1671;
mov.b32 %f862, %r1692;
add.ftz.f32 %f863, %f861, %f862;
mov.b32 %r1693, %f863;
shfl.sync.bfly.b32 %r1694|%p250, %r1693, %r1683, %r1669, %r1671;
mov.b32 %f864, %r1694;
add.ftz.f32 %f69, %f863, %f864;
add.s32 %r1695, %r4, 2;
setp.ge.u32 %p251, %r1695, %r57;
setp.ne.s32 %p252, %r2, 2;
or.pred %p253, %p252, %p251;
@%p253 bra $L__BB142_68;

add.ftz.f32 %f865, %f5, %f68;
add.ftz.f32 %f866, %f4, %f69;
mul.ftz.f32 %f867, %f866, 0f3F000000;
mul.ftz.f32 %f868, %f866, 0f3F4C422A;
mul.ftz.f32 %f869, %f866, 0f3D372713;
fma.rn.ftz.f32 %f870, %f866, %f869, 0f3F800000;
mul.ftz.f32 %f871, %f868, %f870;
tanh.approx.f32 %f872, %f871;
add.ftz.f32 %f873, %f872, 0f3F800000;
mul.ftz.f32 %f874, %f867, %f873;
mul.ftz.f32 %f875, %f865, %f874;
st.global.f32 [%rd92+8], %f875;

$L__BB142_68:
ld.shared.f32 %f876, [%r39+384];
add.ftz.f32 %f877, %f876, %f27;
ld.shared.f32 %f878, [%r41+384];
add.ftz.f32 %f879, %f878, %f31;
ld.shared.f32 %f880, [%r39+896];
add.ftz.f32 %f881, %f880, %f877;
ld.shared.f32 %f882, [%r41+896];
add.ftz.f32 %f883, %f882, %f879;
ld.shared.f32 %f884, [%r39+1408];
add.ftz.f32 %f885, %f884, %f881;
ld.shared.f32 %f886, [%r41+1408];
add.ftz.f32 %f887, %f886, %f883;
mov.b32 %r1696, %f885;
shfl.sync.bfly.b32 %r1700|%p254, %r1696, %r1670, %r1669, %r1671;
mov.b32 %f888, %r1700;
add.ftz.f32 %f889, %f885, %f888;
mov.b32 %r1701, %f889;
shfl.sync.bfly.b32 %r1703|%p255, %r1701, %r1674, %r1669, %r1671;
mov.b32 %f890, %r1703;
add.ftz.f32 %f891, %f889, %f890;
mov.b32 %r1704, %f891;
shfl.sync.bfly.b32 %r1706|%p256, %r1704, %r1677, %r1669, %r1671;
mov.b32 %f892, %r1706;
add.ftz.f32 %f893, %f891, %f892;
mov.b32 %r1707, %f893;
shfl.sync.bfly.b32 %r1709|%p257, %r1707, %r1680, %r1669, %r1671;
mov.b32 %f894, %r1709;
add.ftz.f32 %f895, %f893, %f894;
mov.b32 %r1710, %f895;
shfl.sync.bfly.b32 %r1712|%p258, %r1710, %r1683, %r1669, %r1671;
mov.b32 %f896, %r1712;
add.ftz.f32 %f70, %f895, %f896;
mov.b32 %r1713, %f887;
shfl.sync.bfly.b32 %r1714|%p259, %r1713, %r1670, %r1669, %r1671;
mov.b32 %f897, %r1714;
add.ftz.f32 %f898, %f887, %f897;
mov.b32 %r1715, %f898;
shfl.sync.bfly.b32 %r1716|%p260, %r1715, %r1674, %r1669, %r1671;
mov.b32 %f899, %r1716;
add.ftz.f32 %f900, %f898, %f899;
mov.b32 %r1717, %f900;
shfl.sync.bfly.b32 %r1718|%p261, %r1717, %r1677, %r1669, %r1671;
mov.b32 %f901, %r1718;
add.ftz.f32 %f902, %f900, %f901;
mov.b32 %r1719, %f902;
shfl.sync.bfly.b32 %r1720|%p262, %r1719, %r1680, %r1669, %r1671;
mov.b32 %f903, %r1720;
add.ftz.f32 %f904, %f902, %f903;
mov.b32 %r1721, %f904;
shfl.sync.bfly.b32 %r1722|%p263, %r1721, %r1683, %r1669, %r1671;
mov.b32 %f905, %r1722;
add.ftz.f32 %f71, %f904, %f905;
add.s32 %r1723, %r4, 3;
setp.ge.u32 %p264, %r1723, %r57;
setp.ne.s32 %p265, %r2, 3;
or.pred %p266, %p265, %p264;
@%p266 bra $L__BB142_78;

add.ftz.f32 %f906, %f5, %f70;
add.ftz.f32 %f907, %f4, %f71;
mul.ftz.f32 %f908, %f907, 0f3F000000;
mul.ftz.f32 %f909, %f907, 0f3F4C422A;
mul.ftz.f32 %f910, %f907, 0f3D372713;
fma.rn.ftz.f32 %f911, %f907, %f910, 0f3F800000;
mul.ftz.f32 %f912, %f909, %f911;
tanh.approx.f32 %f913, %f912;
add.ftz.f32 %f914, %f913, 0f3F800000;
mul.ftz.f32 %f915, %f908, %f914;
mul.ftz.f32 %f916, %f906, %f915;
st.global.f32 [%rd92+12], %f916;
bra.uni $L__BB142_78;

$L__BB142_54:
ld.shared.f32 %f429, [%r39];
add.ftz.f32 %f430, %f429, %f30;
ld.shared.f32 %f431, [%r41];
add.ftz.f32 %f432, %f431, %f34;
ld.shared.f32 %f433, [%r39+512];
add.ftz.f32 %f434, %f433, %f430;
ld.shared.f32 %f435, [%r41+512];
add.ftz.f32 %f436, %f435, %f432;
ld.shared.f32 %f437, [%r39+1024];
add.ftz.f32 %f438, %f437, %f434;
ld.shared.f32 %f439, [%r41+1024];
add.ftz.f32 %f440, %f439, %f436;
mov.b32 %r1391, %f438;
mov.u32 %r1392, 31;
mov.u32 %r1393, 16;
mov.u32 %r1394, -1;
shfl.sync.bfly.b32 %r1395|%p111, %r1391, %r1393, %r1392, %r1394;
mov.b32 %f441, %r1395;
add.ftz.f32 %f442, %f438, %f441;
mov.b32 %r1396, %f442;
mov.u32 %r1397, 8;
shfl.sync.bfly.b32 %r1398|%p112, %r1396, %r1397, %r1392, %r1394;
mov.b32 %f443, %r1398;
add.ftz.f32 %f444, %f442, %f443;
mov.b32 %r1399, %f444;
mov.u32 %r1400, 4;
shfl.sync.bfly.b32 %r1401|%p113, %r1399, %r1400, %r1392, %r1394;
mov.b32 %f445, %r1401;
add.ftz.f32 %f446, %f444, %f445;
mov.b32 %r1402, %f446;
mov.u32 %r1403, 2;
shfl.sync.bfly.b32 %r1404|%p114, %r1402, %r1403, %r1392, %r1394;
mov.b32 %f447, %r1404;
add.ftz.f32 %f448, %f446, %f447;
mov.b32 %r1405, %f448;
mov.u32 %r1406, 1;
shfl.sync.bfly.b32 %r1407|%p115, %r1405, %r1406, %r1392, %r1394;
mov.b32 %f449, %r1407;
add.ftz.f32 %f48, %f448, %f449;
mov.b32 %r1408, %f440;
shfl.sync.bfly.b32 %r1409|%p116, %r1408, %r1393, %r1392, %r1394;
mov.b32 %f450, %r1409;
add.ftz.f32 %f451, %f440, %f450;
mov.b32 %r1410, %f451;
shfl.sync.bfly.b32 %r1411|%p117, %r1410, %r1397, %r1392, %r1394;
mov.b32 %f452, %r1411;
add.ftz.f32 %f453, %f451, %f452;
mov.b32 %r1412, %f453;
shfl.sync.bfly.b32 %r1413|%p118, %r1412, %r1400, %r1392, %r1394;
mov.b32 %f454, %r1413;
add.ftz.f32 %f455, %f453, %f454;
mov.b32 %r1414, %f455;
shfl.sync.bfly.b32 %r1415|%p119, %r1414, %r1403, %r1392, %r1394;
mov.b32 %f456, %r1415;
add.ftz.f32 %f457, %f455, %f456;
mov.b32 %r1416, %f457;
shfl.sync.bfly.b32 %r1417|%p120, %r1416, %r1406, %r1392, %r1394;
mov.b32 %f458, %r1417;
add.ftz.f32 %f49, %f457, %f458;
setp.ge.u32 %p121, %r4, %r57;
setp.ne.s32 %p122, %r2, 0;
or.pred %p123, %p122, %p121;
@%p123 bra $L__BB142_56;

add.ftz.f32 %f459, %f5, %f48;
add.ftz.f32 %f460, %f4, %f49;
mov.f32 %f461, 0f40E00000;
min.ftz.f32 %f462, %f460, %f461;
min.ftz.f32 %f463, %f459, %f461;
mov.f32 %f464, 0fC0E00000;
max.ftz.f32 %f465, %f463, %f464;
mul.ftz.f32 %f466, %f462, 0f3FD9DB23;
mul.ftz.f32 %f467, %f466, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f468, %f467;
add.ftz.f32 %f469, %f468, 0f3F800000;
div.approx.ftz.f32 %f470, %f462, %f469;
add.ftz.f32 %f471, %f465, 0f3F800000;
mul.ftz.f32 %f472, %f471, %f470;
st.global.f32 [%rd92], %f472;

$L__BB142_56:
ld.shared.f32 %f473, [%r39+128];
add.ftz.f32 %f474, %f473, %f29;
ld.shared.f32 %f475, [%r41+128];
add.ftz.f32 %f476, %f475, %f33;
ld.shared.f32 %f477, [%r39+640];
add.ftz.f32 %f478, %f477, %f474;
ld.shared.f32 %f479, [%r41+640];
add.ftz.f32 %f480, %f479, %f476;
ld.shared.f32 %f481, [%r39+1152];
add.ftz.f32 %f482, %f481, %f478;
ld.shared.f32 %f483, [%r41+1152];
add.ftz.f32 %f484, %f483, %f480;
mov.b32 %r1418, %f482;
shfl.sync.bfly.b32 %r1422|%p124, %r1418, %r1393, %r1392, %r1394;
mov.b32 %f485, %r1422;
add.ftz.f32 %f486, %f482, %f485;
mov.b32 %r1423, %f486;
shfl.sync.bfly.b32 %r1425|%p125, %r1423, %r1397, %r1392, %r1394;
mov.b32 %f487, %r1425;
add.ftz.f32 %f488, %f486, %f487;
mov.b32 %r1426, %f488;
shfl.sync.bfly.b32 %r1428|%p126, %r1426, %r1400, %r1392, %r1394;
mov.b32 %f489, %r1428;
add.ftz.f32 %f490, %f488, %f489;
mov.b32 %r1429, %f490;
shfl.sync.bfly.b32 %r1431|%p127, %r1429, %r1403, %r1392, %r1394;
mov.b32 %f491, %r1431;
add.ftz.f32 %f492, %f490, %f491;
mov.b32 %r1432, %f492;
shfl.sync.bfly.b32 %r1434|%p128, %r1432, %r1406, %r1392, %r1394;
mov.b32 %f493, %r1434;
add.ftz.f32 %f50, %f492, %f493;
mov.b32 %r1435, %f484;
shfl.sync.bfly.b32 %r1436|%p129, %r1435, %r1393, %r1392, %r1394;
mov.b32 %f494, %r1436;
add.ftz.f32 %f495, %f484, %f494;
mov.b32 %r1437, %f495;
shfl.sync.bfly.b32 %r1438|%p130, %r1437, %r1397, %r1392, %r1394;
mov.b32 %f496, %r1438;
add.ftz.f32 %f497, %f495, %f496;
mov.b32 %r1439, %f497;
shfl.sync.bfly.b32 %r1440|%p131, %r1439, %r1400, %r1392, %r1394;
mov.b32 %f498, %r1440;
add.ftz.f32 %f499, %f497, %f498;
mov.b32 %r1441, %f499;
shfl.sync.bfly.b32 %r1442|%p132, %r1441, %r1403, %r1392, %r1394;
mov.b32 %f500, %r1442;
add.ftz.f32 %f501, %f499, %f500;
mov.b32 %r1443, %f501;
shfl.sync.bfly.b32 %r1444|%p133, %r1443, %r1406, %r1392, %r1394;
mov.b32 %f502, %r1444;
add.ftz.f32 %f51, %f501, %f502;
add.s32 %r1445, %r4, 1;
setp.ge.u32 %p134, %r1445, %r57;
setp.ne.s32 %p135, %r2, 1;
or.pred %p136, %p135, %p134;
@%p136 bra $L__BB142_58;

add.ftz.f32 %f503, %f5, %f50;
add.ftz.f32 %f504, %f4, %f51;
mov.f32 %f505, 0f40E00000;
min.ftz.f32 %f506, %f504, %f505;
min.ftz.f32 %f507, %f503, %f505;
mov.f32 %f508, 0fC0E00000;
max.ftz.f32 %f509, %f507, %f508;
mul.ftz.f32 %f510, %f506, 0f3FD9DB23;
mul.ftz.f32 %f511, %f510, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f512, %f511;
add.ftz.f32 %f513, %f512, 0f3F800000;
div.approx.ftz.f32 %f514, %f506, %f513;
add.ftz.f32 %f515, %f509, 0f3F800000;
mul.ftz.f32 %f516, %f515, %f514;
st.global.f32 [%rd92+4], %f516;

$L__BB142_58:
ld.shared.f32 %f517, [%r39+256];
add.ftz.f32 %f518, %f517, %f28;
ld.shared.f32 %f519, [%r41+256];
add.ftz.f32 %f520, %f519, %f32;
ld.shared.f32 %f521, [%r39+768];
add.ftz.f32 %f522, %f521, %f518;
ld.shared.f32 %f523, [%r41+768];
add.ftz.f32 %f524, %f523, %f520;
ld.shared.f32 %f525, [%r39+1280];
add.ftz.f32 %f526, %f525, %f522;
ld.shared.f32 %f527, [%r41+1280];
add.ftz.f32 %f528, %f527, %f524;
mov.b32 %r1446, %f526;
mov.u32 %r1447, 31;
mov.u32 %r1448, 16;
mov.u32 %r1449, -1;
shfl.sync.bfly.b32 %r1450|%p137, %r1446, %r1448, %r1447, %r1449;
mov.b32 %f529, %r1450;
add.ftz.f32 %f530, %f526, %f529;
mov.b32 %r1451, %f530;
mov.u32 %r1452, 8;
shfl.sync.bfly.b32 %r1453|%p138, %r1451, %r1452, %r1447, %r1449;
mov.b32 %f531, %r1453;
add.ftz.f32 %f532, %f530, %f531;
mov.b32 %r1454, %f532;
mov.u32 %r1455, 4;
shfl.sync.bfly.b32 %r1456|%p139, %r1454, %r1455, %r1447, %r1449;
mov.b32 %f533, %r1456;
add.ftz.f32 %f534, %f532, %f533;
mov.b32 %r1457, %f534;
mov.u32 %r1458, 2;
shfl.sync.bfly.b32 %r1459|%p140, %r1457, %r1458, %r1447, %r1449;
mov.b32 %f535, %r1459;
add.ftz.f32 %f536, %f534, %f535;
mov.b32 %r1460, %f536;
mov.u32 %r1461, 1;
shfl.sync.bfly.b32 %r1462|%p141, %r1460, %r1461, %r1447, %r1449;
mov.b32 %f537, %r1462;
add.ftz.f32 %f52, %f536, %f537;
mov.b32 %r1463, %f528;
shfl.sync.bfly.b32 %r1464|%p142, %r1463, %r1448, %r1447, %r1449;
mov.b32 %f538, %r1464;
add.ftz.f32 %f539, %f528, %f538;
mov.b32 %r1465, %f539;
shfl.sync.bfly.b32 %r1466|%p143, %r1465, %r1452, %r1447, %r1449;
mov.b32 %f540, %r1466;
add.ftz.f32 %f541, %f539, %f540;
mov.b32 %r1467, %f541;
shfl.sync.bfly.b32 %r1468|%p144, %r1467, %r1455, %r1447, %r1449;
mov.b32 %f542, %r1468;
add.ftz.f32 %f543, %f541, %f542;
mov.b32 %r1469, %f543;
shfl.sync.bfly.b32 %r1470|%p145, %r1469, %r1458, %r1447, %r1449;
mov.b32 %f544, %r1470;
add.ftz.f32 %f545, %f543, %f544;
mov.b32 %r1471, %f545;
shfl.sync.bfly.b32 %r1472|%p146, %r1471, %r1461, %r1447, %r1449;
mov.b32 %f546, %r1472;
add.ftz.f32 %f53, %f545, %f546;
add.s32 %r1473, %r4, 2;
setp.ge.u32 %p147, %r1473, %r57;
setp.ne.s32 %p148, %r2, 2;
or.pred %p149, %p148, %p147;
@%p149 bra $L__BB142_60;

add.ftz.f32 %f547, %f5, %f52;
add.ftz.f32 %f548, %f4, %f53;
mov.f32 %f549, 0f40E00000;
min.ftz.f32 %f550, %f548, %f549;
min.ftz.f32 %f551, %f547, %f549;
mov.f32 %f552, 0fC0E00000;
max.ftz.f32 %f553, %f551, %f552;
mul.ftz.f32 %f554, %f550, 0f3FD9DB23;
mul.ftz.f32 %f555, %f554, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f556, %f555;
add.ftz.f32 %f557, %f556, 0f3F800000;
div.approx.ftz.f32 %f558, %f550, %f557;
add.ftz.f32 %f559, %f553, 0f3F800000;
mul.ftz.f32 %f560, %f559, %f558;
st.global.f32 [%rd92+8], %f560;

$L__BB142_60:
ld.shared.f32 %f561, [%r39+384];
add.ftz.f32 %f562, %f561, %f27;
ld.shared.f32 %f563, [%r41+384];
add.ftz.f32 %f564, %f563, %f31;
ld.shared.f32 %f565, [%r39+896];
add.ftz.f32 %f566, %f565, %f562;
ld.shared.f32 %f567, [%r41+896];
add.ftz.f32 %f568, %f567, %f564;
ld.shared.f32 %f569, [%r39+1408];
add.ftz.f32 %f570, %f569, %f566;
ld.shared.f32 %f571, [%r41+1408];
add.ftz.f32 %f572, %f571, %f568;
mov.b32 %r1474, %f570;
shfl.sync.bfly.b32 %r1478|%p150, %r1474, %r1448, %r1447, %r1449;
mov.b32 %f573, %r1478;
add.ftz.f32 %f574, %f570, %f573;
mov.b32 %r1479, %f574;
shfl.sync.bfly.b32 %r1481|%p151, %r1479, %r1452, %r1447, %r1449;
mov.b32 %f575, %r1481;
add.ftz.f32 %f576, %f574, %f575;
mov.b32 %r1482, %f576;
shfl.sync.bfly.b32 %r1484|%p152, %r1482, %r1455, %r1447, %r1449;
mov.b32 %f577, %r1484;
add.ftz.f32 %f578, %f576, %f577;
mov.b32 %r1485, %f578;
shfl.sync.bfly.b32 %r1487|%p153, %r1485, %r1458, %r1447, %r1449;
mov.b32 %f579, %r1487;
add.ftz.f32 %f580, %f578, %f579;
mov.b32 %r1488, %f580;
shfl.sync.bfly.b32 %r1490|%p154, %r1488, %r1461, %r1447, %r1449;
mov.b32 %f581, %r1490;
add.ftz.f32 %f54, %f580, %f581;
mov.b32 %r1491, %f572;
shfl.sync.bfly.b32 %r1492|%p155, %r1491, %r1448, %r1447, %r1449;
mov.b32 %f582, %r1492;
add.ftz.f32 %f583, %f572, %f582;
mov.b32 %r1493, %f583;
shfl.sync.bfly.b32 %r1494|%p156, %r1493, %r1452, %r1447, %r1449;
mov.b32 %f584, %r1494;
add.ftz.f32 %f585, %f583, %f584;
mov.b32 %r1495, %f585;
shfl.sync.bfly.b32 %r1496|%p157, %r1495, %r1455, %r1447, %r1449;
mov.b32 %f586, %r1496;
add.ftz.f32 %f587, %f585, %f586;
mov.b32 %r1497, %f587;
shfl.sync.bfly.b32 %r1498|%p158, %r1497, %r1458, %r1447, %r1449;
mov.b32 %f588, %r1498;
add.ftz.f32 %f589, %f587, %f588;
mov.b32 %r1499, %f589;
shfl.sync.bfly.b32 %r1500|%p159, %r1499, %r1461, %r1447, %r1449;
mov.b32 %f590, %r1500;
add.ftz.f32 %f55, %f589, %f590;
add.s32 %r1501, %r4, 3;
setp.ge.u32 %p160, %r1501, %r57;
setp.ne.s32 %p161, %r2, 3;
or.pred %p162, %p161, %p160;
@%p162 bra $L__BB142_78;

add.ftz.f32 %f591, %f5, %f54;
add.ftz.f32 %f592, %f4, %f55;
mov.f32 %f593, 0f40E00000;
min.ftz.f32 %f594, %f592, %f593;
min.ftz.f32 %f595, %f591, %f593;
mov.f32 %f596, 0fC0E00000;
max.ftz.f32 %f597, %f595, %f596;
mul.ftz.f32 %f598, %f594, 0f3FD9DB23;
mul.ftz.f32 %f599, %f598, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f600, %f599;
add.ftz.f32 %f601, %f600, 0f3F800000;
div.approx.ftz.f32 %f602, %f594, %f601;
add.ftz.f32 %f603, %f597, 0f3F800000;
mul.ftz.f32 %f604, %f603, %f602;
st.global.f32 [%rd92+12], %f604;
bra.uni $L__BB142_78;

}
.entry _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<38>;
.reg .b16 %rs<53>;
.reg .f32 %f<145>;
.reg .b32 %r<773>;
.reg .b64 %rd<112>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

ld.param.u64 %rd23, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd24, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd25, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
mov.u32 %r43, %tid.y;
shl.b32 %r44, %r43, 5;
mov.u32 %r45, %tid.x;
add.s32 %r1, %r44, %r45;
shr.u32 %r2, %r42, 8;
mov.u32 %r770, %ctaid.y;
setp.eq.s64 %p1, %rd25, 0;
@%p1 bra $L__BB143_2;

cvta.to.global.u64 %rd27, %rd25;
mul.wide.u32 %rd28, %r770, 4;
add.s64 %rd29, %rd27, %rd28;
ld.global.nc.u32 %r769, [%rd29];
mul.hi.u32 %r46, %r770, %r25;
add.s32 %r47, %r46, %r770;
shr.u32 %r48, %r47, %r26;
mul.lo.s32 %r49, %r48, %r27;
sub.s32 %r770, %r770, %r49;
bra.uni $L__BB143_3;

$L__BB143_2:
mul.hi.u32 %r50, %r770, %r30;
add.s32 %r51, %r50, %r770;
shr.u32 %r769, %r51, %r31;

$L__BB143_3:
shr.s32 %r52, %r1, 31;
shr.u32 %r53, %r52, 28;
add.s32 %r54, %r1, %r53;
shr.s32 %r772, %r54, 4;
setp.ge.s32 %p2, %r772, %r2;
mov.f32 %f141, 0f00000000;
mov.f32 %f142, %f141;
mov.f32 %f143, %f141;
mov.f32 %f144, %f141;
@%p2 bra $L__BB143_6;

mov.u32 %r55, %ctaid.z;
and.b32 %r59, %r54, -16;
sub.s32 %r60, %r1, %r59;
shr.s32 %r61, %r60, 31;
shr.u32 %r62, %r61, 29;
add.s32 %r63, %r60, %r62;
and.b32 %r64, %r63, -8;
sub.s32 %r65, %r60, %r64;
shr.s32 %r66, %r65, 31;
shr.u32 %r67, %r66, 30;
add.s32 %r68, %r65, %r67;
shr.s32 %r69, %r68, 2;
add.s32 %r70, %r64, %r69;
shr.s32 %r71, %r70, 31;
shr.u32 %r72, %r71, 29;
add.s32 %r73, %r70, %r72;
and.b32 %r74, %r73, -8;
sub.s32 %r75, %r70, %r74;
shr.s32 %r76, %r73, 3;
shl.b32 %r10, %r76, 2;
cvt.s64.s32 %rd30, %r75;
shr.s32 %r77, %r63, 3;
shl.b32 %r78, %r77, 2;
shl.b32 %r79, %r60, 1;
shl.b32 %r80, %r65, 1;
cvt.s64.s32 %rd1, %r65;
cvt.s64.s32 %rd2, %r78;
shr.u32 %r81, %r71, 30;
add.s32 %r82, %r70, %r81;
and.b32 %r83, %r82, -4;
sub.s32 %r84, %r70, %r83;
shr.s32 %r85, %r82, 2;
shl.b32 %r11, %r85, 1;
add.s32 %r86, %r84, 8;
cvt.s64.s32 %rd31, %r86;
add.s32 %r87, %r70, 2;
shr.s32 %r88, %r87, 31;
shr.u32 %r89, %r88, 29;
add.s32 %r90, %r87, %r89;
and.b32 %r91, %r90, -8;
sub.s32 %r92, %r87, %r91;
shr.s32 %r93, %r90, 3;
shl.b32 %r12, %r93, 2;
cvt.s64.s32 %rd32, %r92;
shr.u32 %r94, %r88, 30;
add.s32 %r95, %r87, %r94;
and.b32 %r96, %r95, -4;
sub.s32 %r97, %r87, %r96;
shr.s32 %r98, %r95, 2;
shl.b32 %r13, %r98, 1;
add.s32 %r99, %r97, 8;
cvt.s64.s32 %rd33, %r99;
add.s32 %r100, %r70, 4;
shr.s32 %r101, %r100, 31;
shr.u32 %r102, %r101, 29;
add.s32 %r103, %r100, %r102;
and.b32 %r104, %r103, -8;
sub.s32 %r105, %r100, %r104;
shr.s32 %r106, %r103, 3;
shl.b32 %r14, %r106, 2;
cvt.s64.s32 %rd34, %r105;
shr.u32 %r107, %r101, 30;
add.s32 %r108, %r100, %r107;
and.b32 %r109, %r108, -4;
sub.s32 %r110, %r100, %r109;
shr.s32 %r111, %r108, 2;
shl.b32 %r15, %r111, 1;
add.s32 %r112, %r110, 8;
cvt.s64.s32 %rd35, %r112;
add.s32 %r113, %r70, 6;
shr.s32 %r114, %r113, 31;
shr.u32 %r115, %r114, 29;
add.s32 %r116, %r113, %r115;
and.b32 %r117, %r116, -8;
sub.s32 %r118, %r113, %r117;
shr.s32 %r119, %r116, 3;
shl.b32 %r16, %r119, 2;
cvt.s64.s32 %rd36, %r118;
shr.u32 %r120, %r114, 30;
add.s32 %r121, %r113, %r120;
and.b32 %r122, %r121, -4;
sub.s32 %r123, %r113, %r122;
shr.s32 %r124, %r121, 2;
shl.b32 %r17, %r124, 1;
add.s32 %r125, %r123, 8;
cvt.s64.s32 %rd37, %r125;
shl.b32 %r771, %r772, 3;
mov.u32 %r126, %ctaid.x;
shl.b32 %r127, %r126, 2;
mad.lo.s32 %r128, %r127, %r28, %r772;
mul.hi.u32 %r129, %r55, %r36;
add.s32 %r130, %r129, %r55;
shr.u32 %r131, %r130, %r37;
mad.lo.s32 %r132, %r131, %r39, %r128;
mad.lo.s32 %r133, %r769, %r33, %r132;
mul.wide.s32 %rd3, %r133, 110;
mul.wide.s32 %rd38, %r79, 2;
add.s64 %rd4, %rd3, %rd38;
add.s64 %rd5, %rd4, 34;
mul.wide.s32 %rd39, %r80, 2;
add.s64 %rd40, %rd3, %rd39;
add.s64 %rd6, %rd40, 2;
add.s64 %rd41, %rd3, %rd30;
add.s64 %rd7, %rd41, 96;
add.s64 %rd42, %rd3, %rd31;
add.s64 %rd8, %rd42, 96;
add.s64 %rd43, %rd3, %rd32;
add.s64 %rd9, %rd43, 96;
add.s64 %rd44, %rd3, %rd37;
add.s64 %rd10, %rd44, 96;
add.s64 %rd45, %rd3, %rd36;
add.s64 %rd11, %rd45, 96;
add.s64 %rd46, %rd3, %rd35;
add.s64 %rd12, %rd46, 96;
add.s64 %rd47, %rd3, %rd34;
add.s64 %rd13, %rd47, 96;
add.s64 %rd48, %rd3, %rd33;
add.s64 %rd14, %rd48, 96;
mul.wide.s32 %rd15, %r28, 110;
add.s64 %rd16, %rd15, 2;
cvta.to.global.u64 %rd17, %rd24;
cvta.to.global.u64 %rd111, %rd23;
mul.lo.s32 %r134, %r770, %r34;
cvt.u64.u32 %rd49, %r134;
mul.lo.s32 %r135, %r55, %r40;
cvt.u64.u32 %rd50, %r135;
add.s64 %rd19, %rd49, %rd50;
mov.f32 %f141, 0f00000000;
shl.b64 %rd62, %rd1, 2;

$L__BB143_5:
cvt.u64.u32 %rd51, %r771;
add.s64 %rd52, %rd19, %rd51;
add.s64 %rd53, %rd111, %rd3;
add.s64 %rd54, %rd53, 108;
ld.global.nc.u16 %rs1, [%rd53+108];

{ cvt.f32.f16 %f25, %rs1;}


add.s64 %rd55, %rd111, %rd4;
add.s64 %rd56, %rd111, %rd5;
add.s64 %rd57, %rd111, %rd6;
add.s64 %rd58, %rd57, -2;
ld.global.nc.u16 %rs5, [%rd57+-2];
ld.global.nc.u16 %rs6, [%rd57];
mov.b32 %r280, {%rs5, %rs6};
not.b32 %r281, %r280;
ld.global.nc.u16 %rs7, [%rd55+34];
ld.global.nc.u16 %rs8, [%rd55+32];
add.s64 %rd59, %rd52, %rd2;
mul.lo.s64 %rd60, %rd59, 36;
add.s64 %rd61, %rd17, %rd60;
add.s64 %rd63, %rd61, %rd62;
ld.global.nc.u32 %r146, [%rd63+4];
ld.global.nc.u32 %r136, [%rd61];

{.reg .f16 low,high;
mov.b32 {low,high},%r136;
cvt.f32.f16 %f26, low;}


ld.global.nc.u32 %r154, [%rd63+40];
ld.global.nc.u32 %r137, [%rd61+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r137;
cvt.f32.f16 %f27, low;}


ld.global.nc.u32 %r162, [%rd63+76];
ld.global.nc.u32 %r138, [%rd61+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r138;
cvt.f32.f16 %f28, low;}


ld.global.nc.u32 %r170, [%rd63+112];
ld.global.nc.u32 %r139, [%rd61+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r139;
cvt.f32.f16 %f29, low;}


mov.b32 %r282, {%rs8, %rs7};
cvt.u32.u64 %r283, %rd2;
shr.s32 %r284, %r281, %r283;
add.s64 %rd64, %rd111, %rd7;
ld.global.nc.u8 %rs9, [%rd64];
cvt.u32.u16 %r285, %rs9;
and.b32 %r286, %r285, 255;
shr.u32 %r287, %r286, %r10;
and.b32 %r288, %r287, 15;
add.s64 %rd65, %rd111, %rd8;
ld.global.nc.u8 %rs10, [%rd65];
cvt.u32.u16 %r289, %rs10;
and.b32 %r290, %r289, 255;
shr.u32 %r291, %r290, %r11;
and.b32 %r292, %r291, 3;
bfi.b32 %r293, %r292, %r288, 4, 2;
add.s32 %r294, %r293, -32;
and.b32 %r142, %r282, 50529027;
shl.b32 %r295, %r284, 2;
and.b32 %r296, %r295, 67372036;
or.b32 %r297, %r142, -2139062144;
sub.s32 %r298, %r297, %r296;
xor.b32 %r299, %r298, -2139062144;
or.b32 %r300, %r142, %r296;
xor.b32 %r301, %r298, %r142;
and.b32 %r140, %r301, %r300;

prmt.b32 %r140, %r140, 0, 0xba98;


prmt.b32 %r142, %r142, 0, 0xba98;

xor.b32 %r302, %r142, 2139062143;
not.b32 %r303, %r140;
and.b32 %r304, %r299, %r303;
and.b32 %r305, %r302, %r140;
or.b32 %r145, %r305, %r304;
mov.u32 %r279, 0;

dp4a.s32.s32 %r144, %r145, %r146, %r279;

mul.lo.s32 %r306, %r294, %r144;
cvt.rn.f32.s32 %f45, %r306;
fma.rn.ftz.f32 %f46, %f26, %f45, 0f00000000;
add.s64 %rd66, %rd111, %rd9;
ld.global.nc.u8 %rs11, [%rd66];
cvt.u32.u16 %r307, %rs11;
and.b32 %r308, %r307, 255;
shr.u32 %r309, %r308, %r12;
and.b32 %r310, %r309, 15;
add.s64 %rd67, %rd111, %rd14;
ld.global.nc.u8 %rs12, [%rd67];
cvt.u32.u16 %r311, %rs12;
and.b32 %r312, %r311, 255;
shr.u32 %r313, %r312, %r13;
and.b32 %r314, %r313, 3;
bfi.b32 %r315, %r314, %r310, 4, 2;
add.s32 %r316, %r315, -32;
shr.u32 %r317, %r282, 2;
and.b32 %r150, %r317, 50529027;
shl.b32 %r318, %r284, 1;
and.b32 %r319, %r318, 67372036;
or.b32 %r320, %r150, -2139062144;
sub.s32 %r321, %r320, %r319;
xor.b32 %r322, %r321, -2139062144;
or.b32 %r323, %r150, %r319;
xor.b32 %r324, %r321, %r150;
and.b32 %r148, %r324, %r323;

prmt.b32 %r148, %r148, 0, 0xba98;


prmt.b32 %r150, %r150, 0, 0xba98;

xor.b32 %r325, %r150, 2139062143;
not.b32 %r326, %r148;
and.b32 %r327, %r322, %r326;
and.b32 %r328, %r325, %r148;
or.b32 %r153, %r328, %r327;

dp4a.s32.s32 %r152, %r153, %r154, %r279;

mul.lo.s32 %r329, %r316, %r152;
cvt.rn.f32.s32 %f47, %r329;
fma.rn.ftz.f32 %f48, %f27, %f47, %f46;
add.s64 %rd68, %rd111, %rd13;
ld.global.nc.u8 %rs13, [%rd68];
cvt.u32.u16 %r330, %rs13;
and.b32 %r331, %r330, 255;
shr.u32 %r332, %r331, %r14;
and.b32 %r333, %r332, 15;
add.s64 %rd69, %rd111, %rd12;
ld.global.nc.u8 %rs14, [%rd69];
cvt.u32.u16 %r334, %rs14;
and.b32 %r335, %r334, 255;
shr.u32 %r336, %r335, %r15;
and.b32 %r337, %r336, 3;
bfi.b32 %r338, %r337, %r333, 4, 2;
add.s32 %r339, %r338, -32;
shr.u32 %r340, %r282, 4;
and.b32 %r158, %r340, 50529027;
and.b32 %r341, %r284, 67372036;
or.b32 %r342, %r158, -2139062144;
sub.s32 %r343, %r342, %r341;
xor.b32 %r344, %r343, -2139062144;
or.b32 %r345, %r158, %r341;
xor.b32 %r346, %r343, %r158;
and.b32 %r156, %r346, %r345;

prmt.b32 %r156, %r156, 0, 0xba98;


prmt.b32 %r158, %r158, 0, 0xba98;

xor.b32 %r347, %r158, 2139062143;
not.b32 %r348, %r156;
and.b32 %r349, %r344, %r348;
and.b32 %r350, %r347, %r156;
or.b32 %r161, %r350, %r349;

dp4a.s32.s32 %r160, %r161, %r162, %r279;

mul.lo.s32 %r351, %r339, %r160;
cvt.rn.f32.s32 %f49, %r351;
fma.rn.ftz.f32 %f50, %f28, %f49, %f48;
add.s64 %rd70, %rd111, %rd11;
ld.global.nc.u8 %rs15, [%rd70];
cvt.u32.u16 %r352, %rs15;
and.b32 %r353, %r352, 255;
shr.u32 %r354, %r353, %r16;
and.b32 %r355, %r354, 15;
add.s64 %rd71, %rd111, %rd10;
ld.global.nc.u8 %rs16, [%rd71];
cvt.u32.u16 %r356, %rs16;
and.b32 %r357, %r356, 255;
shr.u32 %r358, %r357, %r17;
and.b32 %r359, %r358, 3;
bfi.b32 %r360, %r359, %r355, 4, 2;
add.s32 %r361, %r360, -32;
shr.u32 %r362, %r282, 6;
and.b32 %r166, %r362, 50529027;
shr.u32 %r363, %r284, 1;
and.b32 %r364, %r363, 67372036;
or.b32 %r365, %r166, -2139062144;
sub.s32 %r366, %r365, %r364;
xor.b32 %r367, %r366, -2139062144;
or.b32 %r368, %r166, %r364;
xor.b32 %r369, %r366, %r166;
and.b32 %r164, %r369, %r368;

prmt.b32 %r164, %r164, 0, 0xba98;


prmt.b32 %r166, %r166, 0, 0xba98;

xor.b32 %r370, %r166, 2139062143;
not.b32 %r371, %r164;
and.b32 %r372, %r367, %r371;
and.b32 %r373, %r370, %r164;
or.b32 %r169, %r373, %r372;

dp4a.s32.s32 %r168, %r169, %r170, %r279;

mul.lo.s32 %r374, %r361, %r168;
cvt.rn.f32.s32 %f51, %r374;
fma.rn.ftz.f32 %f52, %f29, %f51, %f50;
fma.rn.ftz.f32 %f144, %f25, %f52, %f144;
add.s64 %rd72, %rd54, %rd15;
ld.global.nc.u16 %rs2, [%rd72];

{ cvt.f32.f16 %f30, %rs2;}


add.s64 %rd73, %rd58, %rd15;
ld.global.nc.u16 %rs17, [%rd73];
ld.global.nc.u16 %rs18, [%rd73+2];
mov.b32 %r375, {%rs17, %rs18};
not.b32 %r376, %r375;
add.s64 %rd74, %rd56, %rd15;
add.s64 %rd75, %rd74, -2;
add.s64 %rd76, %rd55, %rd15;
ld.global.nc.u16 %rs19, [%rd76+34];
ld.global.nc.u16 %rs20, [%rd76+32];

{.reg .f16 low,high;
mov.b32 {low,high},%r136;
cvt.f32.f16 %f31, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r137;
cvt.f32.f16 %f32, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r138;
cvt.f32.f16 %f33, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r139;
cvt.f32.f16 %f34, low;}


mov.b32 %r377, {%rs20, %rs19};
shr.s32 %r378, %r376, %r283;
add.s64 %rd77, %rd64, %rd15;
ld.global.nc.u8 %rs21, [%rd77];
cvt.u32.u16 %r379, %rs21;
and.b32 %r380, %r379, 255;
shr.u32 %r381, %r380, %r10;
and.b32 %r382, %r381, 15;
add.s64 %rd78, %rd65, %rd15;
ld.global.nc.u8 %rs22, [%rd78];
cvt.u32.u16 %r383, %rs22;
and.b32 %r384, %r383, 255;
shr.u32 %r385, %r384, %r11;
and.b32 %r386, %r385, 3;
bfi.b32 %r387, %r386, %r382, 4, 2;
add.s32 %r388, %r387, -32;
and.b32 %r178, %r377, 50529027;
shl.b32 %r389, %r378, 2;
and.b32 %r390, %r389, 67372036;
or.b32 %r391, %r178, -2139062144;
sub.s32 %r392, %r391, %r390;
xor.b32 %r393, %r392, -2139062144;
or.b32 %r394, %r178, %r390;
xor.b32 %r395, %r392, %r178;
and.b32 %r176, %r395, %r394;

prmt.b32 %r176, %r176, 0, 0xba98;


prmt.b32 %r178, %r178, 0, 0xba98;

xor.b32 %r396, %r178, 2139062143;
not.b32 %r397, %r176;
and.b32 %r398, %r393, %r397;
and.b32 %r399, %r396, %r176;
or.b32 %r181, %r399, %r398;

dp4a.s32.s32 %r180, %r181, %r146, %r279;

mul.lo.s32 %r400, %r388, %r180;
cvt.rn.f32.s32 %f53, %r400;
fma.rn.ftz.f32 %f54, %f31, %f53, 0f00000000;
add.s64 %rd79, %rd66, %rd15;
ld.global.nc.u8 %rs23, [%rd79];
cvt.u32.u16 %r401, %rs23;
and.b32 %r402, %r401, 255;
shr.u32 %r403, %r402, %r12;
and.b32 %r404, %r403, 15;
add.s64 %rd80, %rd67, %rd15;
ld.global.nc.u8 %rs24, [%rd80];
cvt.u32.u16 %r405, %rs24;
and.b32 %r406, %r405, 255;
shr.u32 %r407, %r406, %r13;
and.b32 %r408, %r407, 3;
bfi.b32 %r409, %r408, %r404, 4, 2;
add.s32 %r410, %r409, -32;
shr.u32 %r411, %r377, 2;
and.b32 %r186, %r411, 50529027;
shl.b32 %r412, %r378, 1;
and.b32 %r413, %r412, 67372036;
or.b32 %r414, %r186, -2139062144;
sub.s32 %r415, %r414, %r413;
xor.b32 %r416, %r415, -2139062144;
or.b32 %r417, %r186, %r413;
xor.b32 %r418, %r415, %r186;
and.b32 %r184, %r418, %r417;

prmt.b32 %r184, %r184, 0, 0xba98;


prmt.b32 %r186, %r186, 0, 0xba98;

xor.b32 %r419, %r186, 2139062143;
not.b32 %r420, %r184;
and.b32 %r421, %r416, %r420;
and.b32 %r422, %r419, %r184;
or.b32 %r189, %r422, %r421;

dp4a.s32.s32 %r188, %r189, %r154, %r279;

mul.lo.s32 %r423, %r410, %r188;
cvt.rn.f32.s32 %f55, %r423;
fma.rn.ftz.f32 %f56, %f32, %f55, %f54;
add.s64 %rd81, %rd68, %rd15;
ld.global.nc.u8 %rs25, [%rd81];
cvt.u32.u16 %r424, %rs25;
and.b32 %r425, %r424, 255;
shr.u32 %r426, %r425, %r14;
and.b32 %r427, %r426, 15;
add.s64 %rd82, %rd69, %rd15;
ld.global.nc.u8 %rs26, [%rd82];
cvt.u32.u16 %r428, %rs26;
and.b32 %r429, %r428, 255;
shr.u32 %r430, %r429, %r15;
and.b32 %r431, %r430, 3;
bfi.b32 %r432, %r431, %r427, 4, 2;
add.s32 %r433, %r432, -32;
shr.u32 %r434, %r377, 4;
and.b32 %r194, %r434, 50529027;
and.b32 %r435, %r378, 67372036;
or.b32 %r436, %r194, -2139062144;
sub.s32 %r437, %r436, %r435;
xor.b32 %r438, %r437, -2139062144;
or.b32 %r439, %r194, %r435;
xor.b32 %r440, %r437, %r194;
and.b32 %r192, %r440, %r439;

prmt.b32 %r192, %r192, 0, 0xba98;


prmt.b32 %r194, %r194, 0, 0xba98;

xor.b32 %r441, %r194, 2139062143;
not.b32 %r442, %r192;
and.b32 %r443, %r438, %r442;
and.b32 %r444, %r441, %r192;
or.b32 %r197, %r444, %r443;

dp4a.s32.s32 %r196, %r197, %r162, %r279;

mul.lo.s32 %r445, %r433, %r196;
cvt.rn.f32.s32 %f57, %r445;
fma.rn.ftz.f32 %f58, %f33, %f57, %f56;
add.s64 %rd83, %rd70, %rd15;
ld.global.nc.u8 %rs27, [%rd83];
cvt.u32.u16 %r446, %rs27;
and.b32 %r447, %r446, 255;
shr.u32 %r448, %r447, %r16;
and.b32 %r449, %r448, 15;
add.s64 %rd84, %rd71, %rd15;
ld.global.nc.u8 %rs28, [%rd84];
cvt.u32.u16 %r450, %rs28;
and.b32 %r451, %r450, 255;
shr.u32 %r452, %r451, %r17;
and.b32 %r453, %r452, 3;
bfi.b32 %r454, %r453, %r449, 4, 2;
add.s32 %r455, %r454, -32;
shr.u32 %r456, %r377, 6;
and.b32 %r202, %r456, 50529027;
shr.u32 %r457, %r378, 1;
and.b32 %r458, %r457, 67372036;
or.b32 %r459, %r202, -2139062144;
sub.s32 %r460, %r459, %r458;
xor.b32 %r461, %r460, -2139062144;
or.b32 %r462, %r202, %r458;
xor.b32 %r463, %r460, %r202;
and.b32 %r200, %r463, %r462;

prmt.b32 %r200, %r200, 0, 0xba98;


prmt.b32 %r202, %r202, 0, 0xba98;

xor.b32 %r464, %r202, 2139062143;
not.b32 %r465, %r200;
and.b32 %r466, %r461, %r465;
and.b32 %r467, %r464, %r200;
or.b32 %r205, %r467, %r466;

dp4a.s32.s32 %r204, %r205, %r170, %r279;

mul.lo.s32 %r468, %r455, %r204;
cvt.rn.f32.s32 %f59, %r468;
fma.rn.ftz.f32 %f60, %f34, %f59, %f58;
fma.rn.ftz.f32 %f143, %f30, %f60, %f143;
add.s64 %rd85, %rd72, %rd15;
ld.global.nc.u16 %rs3, [%rd85];

{ cvt.f32.f16 %f35, %rs3;}


add.s64 %rd86, %rd73, %rd15;
ld.global.nc.u16 %rs29, [%rd86];
ld.global.nc.u16 %rs30, [%rd86+2];
mov.b32 %r469, {%rs29, %rs30};
not.b32 %r470, %r469;
add.s64 %rd87, %rd75, %rd15;
add.s64 %rd88, %rd75, %rd16;
ld.global.nc.u16 %rs31, [%rd88];
ld.global.nc.u16 %rs32, [%rd88+-2];

{.reg .f16 low,high;
mov.b32 {low,high},%r136;
cvt.f32.f16 %f36, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r137;
cvt.f32.f16 %f37, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r138;
cvt.f32.f16 %f38, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r139;
cvt.f32.f16 %f39, low;}


mov.b32 %r471, {%rs32, %rs31};
shr.s32 %r472, %r470, %r283;
add.s64 %rd89, %rd77, %rd15;
ld.global.nc.u8 %rs33, [%rd89];
cvt.u32.u16 %r473, %rs33;
and.b32 %r474, %r473, 255;
shr.u32 %r475, %r474, %r10;
and.b32 %r476, %r475, 15;
add.s64 %rd90, %rd78, %rd15;
ld.global.nc.u8 %rs34, [%rd90];
cvt.u32.u16 %r477, %rs34;
and.b32 %r478, %r477, 255;
shr.u32 %r479, %r478, %r11;
and.b32 %r480, %r479, 3;
bfi.b32 %r481, %r480, %r476, 4, 2;
add.s32 %r482, %r481, -32;
and.b32 %r214, %r471, 50529027;
shl.b32 %r483, %r472, 2;
and.b32 %r484, %r483, 67372036;
or.b32 %r485, %r214, -2139062144;
sub.s32 %r486, %r485, %r484;
xor.b32 %r487, %r486, -2139062144;
or.b32 %r488, %r214, %r484;
xor.b32 %r489, %r486, %r214;
and.b32 %r212, %r489, %r488;

prmt.b32 %r212, %r212, 0, 0xba98;


prmt.b32 %r214, %r214, 0, 0xba98;

xor.b32 %r490, %r214, 2139062143;
not.b32 %r491, %r212;
and.b32 %r492, %r487, %r491;
and.b32 %r493, %r490, %r212;
or.b32 %r217, %r493, %r492;

dp4a.s32.s32 %r216, %r217, %r146, %r279;

mul.lo.s32 %r494, %r482, %r216;
cvt.rn.f32.s32 %f61, %r494;
fma.rn.ftz.f32 %f62, %f36, %f61, 0f00000000;
add.s64 %rd91, %rd79, %rd15;
ld.global.nc.u8 %rs35, [%rd91];
cvt.u32.u16 %r495, %rs35;
and.b32 %r496, %r495, 255;
shr.u32 %r497, %r496, %r12;
and.b32 %r498, %r497, 15;
add.s64 %rd92, %rd80, %rd15;
ld.global.nc.u8 %rs36, [%rd92];
cvt.u32.u16 %r499, %rs36;
and.b32 %r500, %r499, 255;
shr.u32 %r501, %r500, %r13;
and.b32 %r502, %r501, 3;
bfi.b32 %r503, %r502, %r498, 4, 2;
add.s32 %r504, %r503, -32;
shr.u32 %r505, %r471, 2;
and.b32 %r222, %r505, 50529027;
shl.b32 %r506, %r472, 1;
and.b32 %r507, %r506, 67372036;
or.b32 %r508, %r222, -2139062144;
sub.s32 %r509, %r508, %r507;
xor.b32 %r510, %r509, -2139062144;
or.b32 %r511, %r222, %r507;
xor.b32 %r512, %r509, %r222;
and.b32 %r220, %r512, %r511;

prmt.b32 %r220, %r220, 0, 0xba98;


prmt.b32 %r222, %r222, 0, 0xba98;

xor.b32 %r513, %r222, 2139062143;
not.b32 %r514, %r220;
and.b32 %r515, %r510, %r514;
and.b32 %r516, %r513, %r220;
or.b32 %r225, %r516, %r515;

dp4a.s32.s32 %r224, %r225, %r154, %r279;

mul.lo.s32 %r517, %r504, %r224;
cvt.rn.f32.s32 %f63, %r517;
fma.rn.ftz.f32 %f64, %f37, %f63, %f62;
add.s64 %rd93, %rd81, %rd15;
ld.global.nc.u8 %rs37, [%rd93];
cvt.u32.u16 %r518, %rs37;
and.b32 %r519, %r518, 255;
shr.u32 %r520, %r519, %r14;
and.b32 %r521, %r520, 15;
add.s64 %rd94, %rd82, %rd15;
ld.global.nc.u8 %rs38, [%rd94];
cvt.u32.u16 %r522, %rs38;
and.b32 %r523, %r522, 255;
shr.u32 %r524, %r523, %r15;
and.b32 %r525, %r524, 3;
bfi.b32 %r526, %r525, %r521, 4, 2;
add.s32 %r527, %r526, -32;
shr.u32 %r528, %r471, 4;
and.b32 %r230, %r528, 50529027;
and.b32 %r529, %r472, 67372036;
or.b32 %r530, %r230, -2139062144;
sub.s32 %r531, %r530, %r529;
xor.b32 %r532, %r531, -2139062144;
or.b32 %r533, %r230, %r529;
xor.b32 %r534, %r531, %r230;
and.b32 %r228, %r534, %r533;

prmt.b32 %r228, %r228, 0, 0xba98;


prmt.b32 %r230, %r230, 0, 0xba98;

xor.b32 %r535, %r230, 2139062143;
not.b32 %r536, %r228;
and.b32 %r537, %r532, %r536;
and.b32 %r538, %r535, %r228;
or.b32 %r233, %r538, %r537;

dp4a.s32.s32 %r232, %r233, %r162, %r279;

mul.lo.s32 %r539, %r527, %r232;
cvt.rn.f32.s32 %f65, %r539;
fma.rn.ftz.f32 %f66, %f38, %f65, %f64;
add.s64 %rd95, %rd83, %rd15;
ld.global.nc.u8 %rs39, [%rd95];
cvt.u32.u16 %r540, %rs39;
and.b32 %r541, %r540, 255;
shr.u32 %r542, %r541, %r16;
and.b32 %r543, %r542, 15;
add.s64 %rd96, %rd84, %rd15;
ld.global.nc.u8 %rs40, [%rd96];
cvt.u32.u16 %r544, %rs40;
and.b32 %r545, %r544, 255;
shr.u32 %r546, %r545, %r17;
and.b32 %r547, %r546, 3;
bfi.b32 %r548, %r547, %r543, 4, 2;
add.s32 %r549, %r548, -32;
shr.u32 %r550, %r471, 6;
and.b32 %r238, %r550, 50529027;
shr.u32 %r551, %r472, 1;
and.b32 %r552, %r551, 67372036;
or.b32 %r553, %r238, -2139062144;
sub.s32 %r554, %r553, %r552;
xor.b32 %r555, %r554, -2139062144;
or.b32 %r556, %r238, %r552;
xor.b32 %r557, %r554, %r238;
and.b32 %r236, %r557, %r556;

prmt.b32 %r236, %r236, 0, 0xba98;


prmt.b32 %r238, %r238, 0, 0xba98;

xor.b32 %r558, %r238, 2139062143;
not.b32 %r559, %r236;
and.b32 %r560, %r555, %r559;
and.b32 %r561, %r558, %r236;
or.b32 %r241, %r561, %r560;

dp4a.s32.s32 %r240, %r241, %r170, %r279;

mul.lo.s32 %r562, %r549, %r240;
cvt.rn.f32.s32 %f67, %r562;
fma.rn.ftz.f32 %f68, %f39, %f67, %f66;
fma.rn.ftz.f32 %f142, %f35, %f68, %f142;
add.s64 %rd97, %rd85, %rd15;
ld.global.nc.u16 %rs4, [%rd97];

{ cvt.f32.f16 %f40, %rs4;}


add.s64 %rd98, %rd86, %rd15;
ld.global.nc.u16 %rs41, [%rd98];
ld.global.nc.u16 %rs42, [%rd98+2];
mov.b32 %r563, {%rs41, %rs42};
not.b32 %r564, %r563;
add.s64 %rd99, %rd87, %rd16;
ld.global.nc.u16 %rs43, [%rd99];
ld.global.nc.u16 %rs44, [%rd99+-2];

{.reg .f16 low,high;
mov.b32 {low,high},%r136;
cvt.f32.f16 %f41, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r137;
cvt.f32.f16 %f42, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r138;
cvt.f32.f16 %f43, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r139;
cvt.f32.f16 %f44, low;}


mov.b32 %r565, {%rs44, %rs43};
shr.s32 %r566, %r564, %r283;
add.s64 %rd100, %rd89, %rd15;
ld.global.nc.u8 %rs45, [%rd100];
cvt.u32.u16 %r567, %rs45;
and.b32 %r568, %r567, 255;
shr.u32 %r569, %r568, %r10;
and.b32 %r570, %r569, 15;
add.s64 %rd101, %rd90, %rd15;
ld.global.nc.u8 %rs46, [%rd101];
cvt.u32.u16 %r571, %rs46;
and.b32 %r572, %r571, 255;
shr.u32 %r573, %r572, %r11;
and.b32 %r574, %r573, 3;
bfi.b32 %r575, %r574, %r570, 4, 2;
add.s32 %r576, %r575, -32;
and.b32 %r250, %r565, 50529027;
shl.b32 %r577, %r566, 2;
and.b32 %r578, %r577, 67372036;
or.b32 %r579, %r250, -2139062144;
sub.s32 %r580, %r579, %r578;
xor.b32 %r581, %r580, -2139062144;
or.b32 %r582, %r250, %r578;
xor.b32 %r583, %r580, %r250;
and.b32 %r248, %r583, %r582;

prmt.b32 %r248, %r248, 0, 0xba98;


prmt.b32 %r250, %r250, 0, 0xba98;

xor.b32 %r584, %r250, 2139062143;
not.b32 %r585, %r248;
and.b32 %r586, %r581, %r585;
and.b32 %r587, %r584, %r248;
or.b32 %r253, %r587, %r586;

dp4a.s32.s32 %r252, %r253, %r146, %r279;

mul.lo.s32 %r588, %r576, %r252;
cvt.rn.f32.s32 %f69, %r588;
fma.rn.ftz.f32 %f70, %f41, %f69, 0f00000000;
add.s64 %rd102, %rd91, %rd15;
ld.global.nc.u8 %rs47, [%rd102];
cvt.u32.u16 %r589, %rs47;
and.b32 %r590, %r589, 255;
shr.u32 %r591, %r590, %r12;
and.b32 %r592, %r591, 15;
add.s64 %rd103, %rd92, %rd15;
ld.global.nc.u8 %rs48, [%rd103];
cvt.u32.u16 %r593, %rs48;
and.b32 %r594, %r593, 255;
shr.u32 %r595, %r594, %r13;
and.b32 %r596, %r595, 3;
bfi.b32 %r597, %r596, %r592, 4, 2;
add.s32 %r598, %r597, -32;
shr.u32 %r599, %r565, 2;
and.b32 %r258, %r599, 50529027;
shl.b32 %r600, %r566, 1;
and.b32 %r601, %r600, 67372036;
or.b32 %r602, %r258, -2139062144;
sub.s32 %r603, %r602, %r601;
xor.b32 %r604, %r603, -2139062144;
or.b32 %r605, %r258, %r601;
xor.b32 %r606, %r603, %r258;
and.b32 %r256, %r606, %r605;

prmt.b32 %r256, %r256, 0, 0xba98;


prmt.b32 %r258, %r258, 0, 0xba98;

xor.b32 %r607, %r258, 2139062143;
not.b32 %r608, %r256;
and.b32 %r609, %r604, %r608;
and.b32 %r610, %r607, %r256;
or.b32 %r261, %r610, %r609;

dp4a.s32.s32 %r260, %r261, %r154, %r279;

mul.lo.s32 %r611, %r598, %r260;
cvt.rn.f32.s32 %f71, %r611;
fma.rn.ftz.f32 %f72, %f42, %f71, %f70;
add.s64 %rd104, %rd93, %rd15;
ld.global.nc.u8 %rs49, [%rd104];
cvt.u32.u16 %r612, %rs49;
and.b32 %r613, %r612, 255;
shr.u32 %r614, %r613, %r14;
and.b32 %r615, %r614, 15;
add.s64 %rd105, %rd94, %rd15;
ld.global.nc.u8 %rs50, [%rd105];
cvt.u32.u16 %r616, %rs50;
and.b32 %r617, %r616, 255;
shr.u32 %r618, %r617, %r15;
and.b32 %r619, %r618, 3;
bfi.b32 %r620, %r619, %r615, 4, 2;
add.s32 %r621, %r620, -32;
shr.u32 %r622, %r565, 4;
and.b32 %r266, %r622, 50529027;
and.b32 %r623, %r566, 67372036;
or.b32 %r624, %r266, -2139062144;
sub.s32 %r625, %r624, %r623;
xor.b32 %r626, %r625, -2139062144;
or.b32 %r627, %r266, %r623;
xor.b32 %r628, %r625, %r266;
and.b32 %r264, %r628, %r627;

prmt.b32 %r264, %r264, 0, 0xba98;


prmt.b32 %r266, %r266, 0, 0xba98;

xor.b32 %r629, %r266, 2139062143;
not.b32 %r630, %r264;
and.b32 %r631, %r626, %r630;
and.b32 %r632, %r629, %r264;
or.b32 %r269, %r632, %r631;

dp4a.s32.s32 %r268, %r269, %r162, %r279;

mul.lo.s32 %r633, %r621, %r268;
cvt.rn.f32.s32 %f73, %r633;
fma.rn.ftz.f32 %f74, %f43, %f73, %f72;
add.s64 %rd106, %rd95, %rd15;
ld.global.nc.u8 %rs51, [%rd106];
cvt.u32.u16 %r634, %rs51;
and.b32 %r635, %r634, 255;
shr.u32 %r636, %r635, %r16;
and.b32 %r637, %r636, 15;
add.s64 %rd107, %rd96, %rd15;
ld.global.nc.u8 %rs52, [%rd107];
cvt.u32.u16 %r638, %rs52;
and.b32 %r639, %r638, 255;
shr.u32 %r640, %r639, %r17;
and.b32 %r641, %r640, 3;
bfi.b32 %r642, %r641, %r637, 4, 2;
add.s32 %r643, %r642, -32;
shr.u32 %r644, %r565, 6;
and.b32 %r274, %r644, 50529027;
shr.u32 %r645, %r566, 1;
and.b32 %r646, %r645, 67372036;
or.b32 %r647, %r274, -2139062144;
sub.s32 %r648, %r647, %r646;
xor.b32 %r649, %r648, -2139062144;
or.b32 %r650, %r274, %r646;
xor.b32 %r651, %r648, %r274;
and.b32 %r272, %r651, %r650;

prmt.b32 %r272, %r272, 0, 0xba98;


prmt.b32 %r274, %r274, 0, 0xba98;

xor.b32 %r652, %r274, 2139062143;
not.b32 %r653, %r272;
and.b32 %r654, %r649, %r653;
and.b32 %r655, %r652, %r272;
or.b32 %r277, %r655, %r654;

dp4a.s32.s32 %r276, %r277, %r170, %r279;

mul.lo.s32 %r656, %r643, %r276;
cvt.rn.f32.s32 %f75, %r656;
fma.rn.ftz.f32 %f76, %f44, %f75, %f74;
fma.rn.ftz.f32 %f141, %f40, %f76, %f141;
add.s32 %r771, %r771, 64;
add.s64 %rd111, %rd111, 880;
add.s32 %r772, %r772, 8;
setp.lt.s32 %p3, %r772, %r2;
@%p3 bra $L__BB143_5;

$L__BB143_6:
mov.u32 %r755, %tid.y;
setp.eq.s32 %p4, %r755, 0;
@%p4 bra $L__BB143_8;

mov.u32 %r764, %tid.x;
mov.u32 %r756, %tid.y;
shl.b32 %r660, %r756, 9;
mov.u32 %r661, _ZZ13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r662, %r661, %r660;
shl.b32 %r663, %r764, 2;
add.s32 %r664, %r662, %r663;
st.shared.f32 [%r664+-512], %f144;
st.shared.f32 [%r664+-384], %f143;
st.shared.f32 [%r664+-256], %f142;
st.shared.f32 [%r664+-128], %f141;

$L__BB143_8:
mov.u32 %r757, %tid.y;
setp.ne.s32 %p5, %r757, 0;
bar.sync 0;
@%p5 bra $L__BB143_17;

mov.u32 %r765, %tid.x;
ld.param.u64 %rd110, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r760, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r759, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r758, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
mov.u32 %r666, %ctaid.z;
mov.u32 %r667, %ctaid.y;
mov.u32 %r668, %ctaid.x;
shl.b32 %r669, %r668, 2;
mov.u32 %r670, 2;
mad.lo.s32 %r671, %r667, %r758, %r669;
mad.lo.s32 %r672, %r666, %r759, %r671;
shl.b32 %r674, %r765, 2;
mov.u32 %r675, _ZZ13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r23, %r675, %r674;
ld.shared.f32 %f77, [%r23];
add.ftz.f32 %f78, %f77, %f144;
ld.shared.f32 %f79, [%r23+512];
add.ftz.f32 %f80, %f79, %f78;
ld.shared.f32 %f81, [%r23+1024];
add.ftz.f32 %f82, %f81, %f80;
mov.b32 %r676, %f82;
mov.u32 %r677, 31;
mov.u32 %r678, 16;
mov.u32 %r679, -1;
shfl.sync.bfly.b32 %r680|%p6, %r676, %r678, %r677, %r679;
mov.b32 %f83, %r680;
add.ftz.f32 %f84, %f82, %f83;
mov.b32 %r681, %f84;
mov.u32 %r682, 8;
shfl.sync.bfly.b32 %r683|%p7, %r681, %r682, %r677, %r679;
mov.b32 %f85, %r683;
add.ftz.f32 %f86, %f84, %f85;
mov.b32 %r684, %f86;
mov.u32 %r685, 4;
shfl.sync.bfly.b32 %r686|%p8, %r684, %r685, %r677, %r679;
mov.b32 %f87, %r686;
add.ftz.f32 %f88, %f86, %f87;
mov.b32 %r687, %f88;
shfl.sync.bfly.b32 %r688|%p9, %r687, %r670, %r677, %r679;
mov.b32 %f89, %r688;
add.ftz.f32 %f90, %f88, %f89;
mov.b32 %r689, %f90;
mov.u32 %r690, 1;
shfl.sync.bfly.b32 %r691|%p10, %r689, %r690, %r677, %r679;
mov.b32 %f91, %r691;
add.ftz.f32 %f13, %f90, %f91;
setp.ne.s32 %p11, %r765, 0;
setp.ge.u32 %p12, %r669, %r760;
cvta.to.global.u64 %rd108, %rd110;
mul.wide.u32 %rd109, %r672, 4;
add.s64 %rd22, %rd108, %rd109;
or.pred %p13, %p11, %p12;
@%p13 bra $L__BB143_11;

st.global.f32 [%rd22], %f13;

$L__BB143_11:
mov.u32 %r766, %tid.x;
ld.param.u32 %r761, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.shared.f32 %f92, [%r23+128];
add.ftz.f32 %f93, %f92, %f143;
ld.shared.f32 %f94, [%r23+640];
add.ftz.f32 %f95, %f94, %f93;
ld.shared.f32 %f96, [%r23+1152];
add.ftz.f32 %f97, %f96, %f95;
mov.b32 %r692, %f97;
shfl.sync.bfly.b32 %r696|%p14, %r692, %r678, %r677, %r679;
mov.b32 %f98, %r696;
add.ftz.f32 %f99, %f97, %f98;
mov.b32 %r697, %f99;
shfl.sync.bfly.b32 %r699|%p15, %r697, %r682, %r677, %r679;
mov.b32 %f100, %r699;
add.ftz.f32 %f101, %f99, %f100;
mov.b32 %r700, %f101;
shfl.sync.bfly.b32 %r702|%p16, %r700, %r685, %r677, %r679;
mov.b32 %f102, %r702;
add.ftz.f32 %f103, %f101, %f102;
mov.b32 %r703, %f103;
shfl.sync.bfly.b32 %r705|%p17, %r703, %r670, %r677, %r679;
mov.b32 %f104, %r705;
add.ftz.f32 %f105, %f103, %f104;
mov.b32 %r706, %f105;
shfl.sync.bfly.b32 %r708|%p18, %r706, %r690, %r677, %r679;
mov.b32 %f106, %r708;
add.ftz.f32 %f14, %f105, %f106;
setp.ne.s32 %p19, %r766, 1;
or.b32 %r712, %r669, 1;
setp.ge.u32 %p20, %r712, %r761;
or.pred %p21, %p19, %p20;
@%p21 bra $L__BB143_13;

st.global.f32 [%rd22+4], %f14;

$L__BB143_13:
mov.u32 %r767, %tid.x;
ld.param.u32 %r762, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.shared.f32 %f107, [%r23+256];
add.ftz.f32 %f108, %f107, %f142;
ld.shared.f32 %f109, [%r23+768];
add.ftz.f32 %f110, %f109, %f108;
ld.shared.f32 %f111, [%r23+1280];
add.ftz.f32 %f112, %f111, %f110;
mov.b32 %r713, %f112;
mov.u32 %r714, 31;
mov.u32 %r715, 16;
mov.u32 %r716, -1;
shfl.sync.bfly.b32 %r717|%p22, %r713, %r715, %r714, %r716;
mov.b32 %f113, %r717;
add.ftz.f32 %f114, %f112, %f113;
mov.b32 %r718, %f114;
mov.u32 %r719, 8;
shfl.sync.bfly.b32 %r720|%p23, %r718, %r719, %r714, %r716;
mov.b32 %f115, %r720;
add.ftz.f32 %f116, %f114, %f115;
mov.b32 %r721, %f116;
mov.u32 %r722, 4;
shfl.sync.bfly.b32 %r723|%p24, %r721, %r722, %r714, %r716;
mov.b32 %f117, %r723;
add.ftz.f32 %f118, %f116, %f117;
mov.b32 %r724, %f118;
mov.u32 %r725, 2;
shfl.sync.bfly.b32 %r726|%p25, %r724, %r725, %r714, %r716;
mov.b32 %f119, %r726;
add.ftz.f32 %f120, %f118, %f119;
mov.b32 %r727, %f120;
mov.u32 %r728, 1;
shfl.sync.bfly.b32 %r729|%p26, %r727, %r728, %r714, %r716;
mov.b32 %f121, %r729;
add.ftz.f32 %f15, %f120, %f121;
setp.ne.s32 %p27, %r767, 2;
or.b32 %r733, %r669, 2;
setp.ge.u32 %p28, %r733, %r762;
or.pred %p29, %p27, %p28;
@%p29 bra $L__BB143_15;

st.global.f32 [%rd22+8], %f15;

$L__BB143_15:
mov.u32 %r768, %tid.x;
ld.param.u32 %r763, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.shared.f32 %f122, [%r23+384];
add.ftz.f32 %f123, %f122, %f141;
ld.shared.f32 %f124, [%r23+896];
add.ftz.f32 %f125, %f124, %f123;
ld.shared.f32 %f126, [%r23+1408];
add.ftz.f32 %f127, %f126, %f125;
mov.b32 %r734, %f127;
shfl.sync.bfly.b32 %r738|%p30, %r734, %r715, %r714, %r716;
mov.b32 %f128, %r738;
add.ftz.f32 %f129, %f127, %f128;
mov.b32 %r739, %f129;
shfl.sync.bfly.b32 %r741|%p31, %r739, %r719, %r714, %r716;
mov.b32 %f130, %r741;
add.ftz.f32 %f131, %f129, %f130;
mov.b32 %r742, %f131;
shfl.sync.bfly.b32 %r744|%p32, %r742, %r722, %r714, %r716;
mov.b32 %f132, %r744;
add.ftz.f32 %f133, %f131, %f132;
mov.b32 %r745, %f133;
shfl.sync.bfly.b32 %r747|%p33, %r745, %r725, %r714, %r716;
mov.b32 %f134, %r747;
add.ftz.f32 %f135, %f133, %f134;
mov.b32 %r748, %f135;
shfl.sync.bfly.b32 %r750|%p34, %r748, %r728, %r714, %r716;
mov.b32 %f136, %r750;
add.ftz.f32 %f16, %f135, %f136;
setp.ne.s32 %p35, %r768, 3;
or.b32 %r754, %r669, 3;
setp.ge.u32 %p36, %r754, %r763;
or.pred %p37, %p35, %p36;
@%p37 bra $L__BB143_17;

st.global.f32 [%rd22+12], %f16;

$L__BB143_17:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<33>;
.reg .b16 %rs<19>;
.reg .f32 %f<139>;
.reg .b32 %r<487>;
.reg .b64 %rd<124>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[384];

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate[384];

ld.param.u64 %rd13, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd14, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u32 %r49, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r53, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r57, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r58, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r59, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r63, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r64, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r65, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r61, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r60, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r55, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r54, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r52, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r51, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r50, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+16];
ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+8];
ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3];
mov.u32 %r1, %tid.y;
shl.b32 %r66, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r66, %r2;
mov.u32 %r4, %ctaid.x;
mov.u32 %r5, %ctaid.y;
setp.eq.s64 %p2, %rd15, 0;
@%p2 bra $L__BB144_2;

cvta.to.global.u64 %rd22, %rd15;
mul.wide.u32 %rd23, %r5, 4;
add.s64 %rd24, %rd22, %rd23;
ld.global.nc.u32 %r483, [%rd24];
mul.hi.u32 %r67, %r5, %r50;
add.s32 %r68, %r67, %r5;
shr.u32 %r69, %r68, %r51;
mul.lo.s32 %r70, %r69, %r52;
sub.s32 %r484, %r5, %r70;
bra.uni $L__BB144_3;

$L__BB144_2:
mul.hi.u32 %r71, %r5, %r54;
add.s32 %r72, %r71, %r5;
shr.u32 %r483, %r72, %r55;
mov.u32 %r484, %r5;

$L__BB144_3:
mov.u32 %r11, %ctaid.z;
selp.b32 %r12, %r5, %r483, %p2;
or.b32 %r73, %r2, %r1;
setp.ne.s32 %p4, %r73, 0;
mov.f32 %f132, 0f00000000;
mov.f32 %f128, %f132;
mov.f32 %f129, %f132;
@%p4 bra $L__BB144_8;

setp.eq.s64 %p5, %rd16, 0;
mov.f32 %f128, 0f00000000;
mov.f32 %f129, %f128;
@%p5 bra $L__BB144_6;

cvta.to.global.u64 %rd25, %rd16;
mul.lo.s32 %r74, %r11, %r65;
cvt.u64.u32 %rd26, %r74;
mul.lo.s32 %r75, %r12, %r59;
cvt.u64.u32 %rd27, %r75;
cvt.u64.u32 %rd28, %r2;
cvt.s64.s32 %rd29, %r4;
add.s64 %rd30, %rd28, %rd29;
add.s64 %rd31, %rd30, %rd26;
add.s64 %rd32, %rd31, %rd27;
shl.b64 %rd33, %rd32, 2;
add.s64 %rd34, %rd25, %rd33;
ld.global.f32 %f129, [%rd34];

$L__BB144_6:
setp.eq.s64 %p6, %rd18, 0;
setp.eq.s64 %p7, %rd17, 0;
or.pred %p8, %p6, %p7;
@%p8 bra $L__BB144_8;

cvta.to.global.u64 %rd35, %rd18;
mul.lo.s32 %r76, %r11, %r65;
cvt.u64.u32 %rd36, %r76;
mul.lo.s32 %r77, %r12, %r59;
cvt.u64.u32 %rd37, %r77;
cvt.u64.u32 %rd38, %r2;
cvt.s64.s32 %rd39, %r4;
add.s64 %rd40, %rd38, %rd39;
add.s64 %rd41, %rd40, %rd36;
add.s64 %rd42, %rd41, %rd37;
shl.b64 %rd43, %rd42, 2;
add.s64 %rd44, %rd35, %rd43;
ld.global.f32 %f128, [%rd44];

$L__BB144_8:
shr.s32 %r78, %r3, 31;
shr.u32 %r79, %r78, 28;
add.s32 %r80, %r3, %r79;
shr.s32 %r486, %r80, 4;
shr.u32 %r81, %r49, 8;
setp.ge.s32 %p9, %r486, %r81;
mov.f32 %f134, %f132;
@%p9 bra $L__BB144_13;

cvta.to.global.u64 %rd45, %rd17;
and.b32 %r85, %r80, -16;
sub.s32 %r86, %r3, %r85;
shr.s32 %r87, %r86, 31;
shr.u32 %r88, %r87, 29;
add.s32 %r89, %r86, %r88;
shr.s32 %r90, %r89, 3;
shl.b32 %r91, %r90, 2;
and.b32 %r92, %r89, -8;
sub.s32 %r14, %r86, %r92;
shr.s32 %r93, %r14, 31;
shr.u32 %r94, %r93, 30;
add.s32 %r95, %r14, %r94;
shr.s32 %r96, %r95, 2;
add.s32 %r97, %r92, %r96;
shl.b32 %r15, %r86, 1;
shl.b32 %r98, %r14, 1;
cvt.s64.s32 %rd4, %r91;
shr.s32 %r99, %r97, 31;
shr.u32 %r100, %r99, 29;
add.s32 %r101, %r97, %r100;
and.b32 %r102, %r101, -8;
sub.s32 %r16, %r97, %r102;
shr.s32 %r103, %r101, 3;
shl.b32 %r17, %r103, 2;
shr.u32 %r104, %r99, 30;
add.s32 %r105, %r97, %r104;
and.b32 %r106, %r105, -4;
sub.s32 %r107, %r97, %r106;
shr.s32 %r108, %r105, 2;
shl.b32 %r18, %r108, 1;
add.s32 %r19, %r107, 8;
add.s32 %r109, %r97, 2;
shr.s32 %r110, %r109, 31;
shr.u32 %r111, %r110, 29;
add.s32 %r112, %r109, %r111;
and.b32 %r113, %r112, -8;
sub.s32 %r20, %r109, %r113;
shr.s32 %r114, %r112, 3;
shl.b32 %r21, %r114, 2;
shr.u32 %r115, %r110, 30;
add.s32 %r116, %r109, %r115;
and.b32 %r117, %r116, -4;
sub.s32 %r118, %r109, %r117;
shr.s32 %r119, %r116, 2;
shl.b32 %r22, %r119, 1;
add.s32 %r23, %r118, 8;
add.s32 %r120, %r97, 4;
shr.s32 %r121, %r120, 31;
shr.u32 %r122, %r121, 29;
add.s32 %r123, %r120, %r122;
and.b32 %r124, %r123, -8;
sub.s32 %r24, %r120, %r124;
shr.s32 %r125, %r123, 3;
shl.b32 %r25, %r125, 2;
shr.u32 %r126, %r121, 30;
add.s32 %r127, %r120, %r126;
and.b32 %r128, %r127, -4;
sub.s32 %r129, %r120, %r128;
shr.s32 %r130, %r127, 2;
shl.b32 %r26, %r130, 1;
add.s32 %r27, %r129, 8;
cvt.s64.s32 %rd46, %r27;
add.s32 %r131, %r97, 6;
shr.s32 %r132, %r131, 31;
shr.u32 %r133, %r132, 29;
add.s32 %r134, %r131, %r133;
and.b32 %r135, %r134, -8;
sub.s32 %r28, %r131, %r135;
shr.s32 %r136, %r134, 3;
shl.b32 %r29, %r136, 2;
shr.u32 %r137, %r132, 30;
add.s32 %r138, %r131, %r137;
and.b32 %r139, %r138, -4;
sub.s32 %r140, %r131, %r139;
shr.s32 %r141, %r138, 2;
shl.b32 %r30, %r141, 1;
add.s32 %r31, %r140, 8;
cvt.s64.s32 %rd47, %r31;
add.s64 %rd48, %rd45, %rd47;
add.s64 %rd5, %rd48, 96;
mul.hi.u32 %r142, %r11, %r60;
add.s32 %r143, %r142, %r11;
shr.u32 %r144, %r143, %r61;
mad.lo.s32 %r145, %r144, %r63, %r486;
mad.lo.s32 %r146, %r483, %r57, %r145;
mad.lo.s32 %r147, %r4, %r53, %r146;
mul.wide.s32 %rd123, %r147, 110;
add.s64 %rd49, %rd45, %rd46;
add.s64 %rd7, %rd49, 96;
shl.b32 %r485, %r486, 3;
mul.wide.s32 %rd50, %r15, 2;
add.s64 %rd51, %rd45, %rd50;
add.s64 %rd8, %rd51, 32;
mul.wide.s32 %rd52, %r98, 2;
add.s64 %rd9, %rd45, %rd52;
mul.lo.s32 %r148, %r484, %r58;
cvt.u64.u32 %rd53, %r148;
mul.lo.s32 %r149, %r11, %r64;
cvt.u64.u32 %rd54, %r149;
add.s64 %rd10, %rd53, %rd54;
mov.f32 %f134, 0f00000000;
cvta.to.global.u64 %rd55, %rd14;
cvta.to.global.u64 %rd58, %rd13;
mov.f32 %f132, %f134;

$L__BB144_10:
cvt.u64.u32 %rd56, %r485;
add.s64 %rd57, %rd10, %rd56;
add.s64 %rd59, %rd58, %rd123;
ld.global.nc.u16 %rs1, [%rd59+108];

{ cvt.f32.f16 %f40, %rs1;}


add.s64 %rd61, %rd58, %rd50;
add.s64 %rd62, %rd61, %rd123;
add.s64 %rd64, %rd58, %rd52;
add.s64 %rd65, %rd64, %rd123;
ld.global.nc.u16 %rs2, [%rd65];
ld.global.nc.u16 %rs3, [%rd65+2];
mov.b32 %r187, {%rs2, %rs3};
not.b32 %r188, %r187;
ld.global.nc.u16 %rs4, [%rd62+34];
ld.global.nc.u16 %rs5, [%rd62+32];
add.s64 %rd66, %rd57, %rd4;
mul.lo.s64 %rd67, %rd66, 36;
add.s64 %rd68, %rd55, %rd67;
mul.wide.s32 %rd69, %r14, 4;
add.s64 %rd70, %rd68, %rd69;
ld.global.nc.u32 %r35, [%rd70+4];
ld.global.nc.u32 %r36, [%rd68];

{.reg .f16 low,high;
mov.b32 {low,high},%r36;
cvt.f32.f16 %f41, low;}


ld.global.nc.u32 %r37, [%rd70+40];
ld.global.nc.u32 %r38, [%rd68+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r38;
cvt.f32.f16 %f42, low;}


ld.global.nc.u32 %r39, [%rd70+76];
ld.global.nc.u32 %r40, [%rd68+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r40;
cvt.f32.f16 %f43, low;}


ld.global.nc.u32 %r41, [%rd70+112];
ld.global.nc.u32 %r42, [%rd68+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r42;
cvt.f32.f16 %f44, low;}


mov.b32 %r189, {%rs5, %rs4};
cvt.u32.u64 %r190, %rd4;
shr.s32 %r191, %r188, %r190;
cvt.s64.s32 %rd71, %r16;
add.s64 %rd72, %rd58, %rd71;
add.s64 %rd73, %rd72, %rd123;
ld.global.nc.u8 %rs6, [%rd73+96];
cvt.u32.u16 %r192, %rs6;
and.b32 %r193, %r192, 255;
shr.u32 %r194, %r193, %r17;
and.b32 %r195, %r194, 15;
cvt.s64.s32 %rd74, %r19;
add.s64 %rd75, %rd58, %rd74;
add.s64 %rd76, %rd75, %rd123;
ld.global.nc.u8 %rs7, [%rd76+96];
cvt.u32.u16 %r196, %rs7;
and.b32 %r197, %r196, 255;
shr.u32 %r198, %r197, %r18;
and.b32 %r199, %r198, 3;
bfi.b32 %r200, %r199, %r195, 4, 2;
add.s32 %r201, %r200, -32;
and.b32 %r156, %r189, 50529027;
shl.b32 %r202, %r191, 2;
and.b32 %r203, %r202, 67372036;
or.b32 %r204, %r156, -2139062144;
sub.s32 %r205, %r204, %r203;
xor.b32 %r206, %r205, -2139062144;
or.b32 %r207, %r156, %r203;
xor.b32 %r208, %r205, %r156;
and.b32 %r154, %r208, %r207;

prmt.b32 %r154, %r154, 0, 0xba98;


prmt.b32 %r156, %r156, 0, 0xba98;

xor.b32 %r209, %r156, 2139062143;
not.b32 %r210, %r154;
and.b32 %r211, %r206, %r210;
and.b32 %r212, %r209, %r154;
or.b32 %r159, %r212, %r211;
mov.u32 %r185, 0;

dp4a.s32.s32 %r158, %r159, %r35, %r185;

mul.lo.s32 %r213, %r201, %r158;
cvt.rn.f32.s32 %f45, %r213;
fma.rn.ftz.f32 %f46, %f41, %f45, 0f00000000;
cvt.s64.s32 %rd77, %r20;
add.s64 %rd78, %rd58, %rd77;
add.s64 %rd79, %rd78, %rd123;
ld.global.nc.u8 %rs8, [%rd79+96];
cvt.u32.u16 %r214, %rs8;
and.b32 %r215, %r214, 255;
shr.u32 %r216, %r215, %r21;
and.b32 %r217, %r216, 15;
cvt.s64.s32 %rd80, %r23;
add.s64 %rd81, %rd58, %rd80;
add.s64 %rd82, %rd81, %rd123;
ld.global.nc.u8 %rs9, [%rd82+96];
cvt.u32.u16 %r218, %rs9;
and.b32 %r219, %r218, 255;
shr.u32 %r220, %r219, %r22;
and.b32 %r221, %r220, 3;
bfi.b32 %r222, %r221, %r217, 4, 2;
add.s32 %r223, %r222, -32;
shr.u32 %r224, %r189, 2;
and.b32 %r164, %r224, 50529027;
shl.b32 %r225, %r191, 1;
and.b32 %r226, %r225, 67372036;
or.b32 %r227, %r164, -2139062144;
sub.s32 %r228, %r227, %r226;
xor.b32 %r229, %r228, -2139062144;
or.b32 %r230, %r164, %r226;
xor.b32 %r231, %r228, %r164;
and.b32 %r162, %r231, %r230;

prmt.b32 %r162, %r162, 0, 0xba98;


prmt.b32 %r164, %r164, 0, 0xba98;

xor.b32 %r232, %r164, 2139062143;
not.b32 %r233, %r162;
and.b32 %r234, %r229, %r233;
and.b32 %r235, %r232, %r162;
or.b32 %r167, %r235, %r234;

dp4a.s32.s32 %r166, %r167, %r37, %r185;

mul.lo.s32 %r236, %r223, %r166;
cvt.rn.f32.s32 %f47, %r236;
fma.rn.ftz.f32 %f48, %f42, %f47, %f46;
cvt.s64.s32 %rd83, %r24;
add.s64 %rd84, %rd58, %rd83;
add.s64 %rd85, %rd84, %rd123;
ld.global.nc.u8 %rs10, [%rd85+96];
cvt.u32.u16 %r237, %rs10;
and.b32 %r238, %r237, 255;
shr.u32 %r239, %r238, %r25;
and.b32 %r240, %r239, 15;
add.s64 %rd87, %rd58, %rd46;
add.s64 %rd88, %rd87, %rd123;
ld.global.nc.u8 %rs11, [%rd88+96];
cvt.u32.u16 %r241, %rs11;
and.b32 %r242, %r241, 255;
shr.u32 %r243, %r242, %r26;
and.b32 %r244, %r243, 3;
bfi.b32 %r245, %r244, %r240, 4, 2;
add.s32 %r246, %r245, -32;
shr.u32 %r247, %r189, 4;
and.b32 %r172, %r247, 50529027;
and.b32 %r248, %r191, 67372036;
or.b32 %r249, %r172, -2139062144;
sub.s32 %r250, %r249, %r248;
xor.b32 %r251, %r250, -2139062144;
or.b32 %r252, %r172, %r248;
xor.b32 %r253, %r250, %r172;
and.b32 %r170, %r253, %r252;

prmt.b32 %r170, %r170, 0, 0xba98;


prmt.b32 %r172, %r172, 0, 0xba98;

xor.b32 %r254, %r172, 2139062143;
not.b32 %r255, %r170;
and.b32 %r256, %r251, %r255;
and.b32 %r257, %r254, %r170;
or.b32 %r175, %r257, %r256;

dp4a.s32.s32 %r174, %r175, %r39, %r185;

mul.lo.s32 %r258, %r246, %r174;
cvt.rn.f32.s32 %f49, %r258;
fma.rn.ftz.f32 %f50, %f43, %f49, %f48;
cvt.s64.s32 %rd89, %r28;
add.s64 %rd90, %rd58, %rd89;
add.s64 %rd91, %rd90, %rd123;
ld.global.nc.u8 %rs12, [%rd91+96];
cvt.u32.u16 %r259, %rs12;
and.b32 %r260, %r259, 255;
shr.u32 %r261, %r260, %r29;
and.b32 %r262, %r261, 15;
add.s64 %rd93, %rd58, %rd47;
add.s64 %rd94, %rd93, %rd123;
ld.global.nc.u8 %rs13, [%rd94+96];
cvt.u32.u16 %r263, %rs13;
and.b32 %r264, %r263, 255;
shr.u32 %r265, %r264, %r30;
and.b32 %r266, %r265, 3;
bfi.b32 %r267, %r266, %r262, 4, 2;
add.s32 %r268, %r267, -32;
shr.u32 %r269, %r189, 6;
and.b32 %r180, %r269, 50529027;
shr.u32 %r270, %r191, 1;
and.b32 %r271, %r270, 67372036;
or.b32 %r272, %r180, -2139062144;
sub.s32 %r273, %r272, %r271;
xor.b32 %r274, %r273, -2139062144;
or.b32 %r275, %r180, %r271;
xor.b32 %r276, %r273, %r180;
and.b32 %r178, %r276, %r275;

prmt.b32 %r178, %r178, 0, 0xba98;


prmt.b32 %r180, %r180, 0, 0xba98;

xor.b32 %r277, %r180, 2139062143;
not.b32 %r278, %r178;
and.b32 %r279, %r274, %r278;
and.b32 %r280, %r277, %r178;
or.b32 %r183, %r280, %r279;

dp4a.s32.s32 %r182, %r183, %r41, %r185;

mul.lo.s32 %r281, %r268, %r182;
cvt.rn.f32.s32 %f51, %r281;
fma.rn.ftz.f32 %f52, %f44, %f51, %f50;
fma.rn.ftz.f32 %f134, %f40, %f52, %f134;
setp.eq.s64 %p10, %rd17, 0;
@%p10 bra $L__BB144_12;

add.s64 %rd96, %rd45, %rd123;
ld.global.u16 %rs14, [%rd96+108];

{ cvt.f32.f16 %f53, %rs14;}


add.s64 %rd97, %rd8, %rd123;
add.s64 %rd98, %rd9, %rd123;
ld.global.u16 %rs15, [%rd98];
ld.global.u16 %rs16, [%rd98+2];
mov.b32 %r319, {%rs15, %rs16};
not.b32 %r320, %r319;
ld.global.u16 %rs17, [%rd97+2];
ld.global.u16 %rs18, [%rd97];

{.reg .f16 low,high;
mov.b32 {low,high},%r36;
cvt.f32.f16 %f54, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r38;
cvt.f32.f16 %f55, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r40;
cvt.f32.f16 %f56, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r42;
cvt.f32.f16 %f57, low;}


mov.b32 %r321, {%rs18, %rs17};
shr.s32 %r322, %r320, %r190;
add.s64 %rd100, %rd45, %rd71;
add.s64 %rd101, %rd100, %rd123;
ld.global.u8 %r323, [%rd101+96];
shr.u32 %r324, %r323, %r17;
and.b32 %r325, %r324, 15;
add.s64 %rd103, %rd45, %rd74;
add.s64 %rd104, %rd103, %rd123;
ld.global.u8 %r326, [%rd104+96];
shr.u32 %r327, %r326, %r18;
and.b32 %r328, %r327, 3;
bfi.b32 %r329, %r328, %r325, 4, 2;
add.s32 %r330, %r329, -32;
and.b32 %r288, %r321, 50529027;
shl.b32 %r331, %r322, 2;
and.b32 %r332, %r331, 67372036;
or.b32 %r333, %r288, -2139062144;
sub.s32 %r334, %r333, %r332;
xor.b32 %r335, %r334, -2139062144;
or.b32 %r336, %r288, %r332;
xor.b32 %r337, %r334, %r288;
and.b32 %r286, %r337, %r336;

prmt.b32 %r286, %r286, 0, 0xba98;


prmt.b32 %r288, %r288, 0, 0xba98;

xor.b32 %r338, %r288, 2139062143;
not.b32 %r339, %r286;
and.b32 %r340, %r335, %r339;
and.b32 %r341, %r338, %r286;
or.b32 %r291, %r341, %r340;

dp4a.s32.s32 %r290, %r291, %r35, %r185;

mul.lo.s32 %r342, %r330, %r290;
cvt.rn.f32.s32 %f58, %r342;
fma.rn.ftz.f32 %f59, %f54, %f58, 0f00000000;
add.s64 %rd106, %rd45, %rd77;
add.s64 %rd107, %rd106, %rd123;
ld.global.u8 %r343, [%rd107+96];
shr.u32 %r344, %r343, %r21;
and.b32 %r345, %r344, 15;
add.s64 %rd109, %rd45, %rd80;
add.s64 %rd110, %rd109, %rd123;
ld.global.u8 %r346, [%rd110+96];
shr.u32 %r347, %r346, %r22;
and.b32 %r348, %r347, 3;
bfi.b32 %r349, %r348, %r345, 4, 2;
add.s32 %r350, %r349, -32;
shr.u32 %r351, %r321, 2;
and.b32 %r296, %r351, 50529027;
shl.b32 %r352, %r322, 1;
and.b32 %r353, %r352, 67372036;
or.b32 %r354, %r296, -2139062144;
sub.s32 %r355, %r354, %r353;
xor.b32 %r356, %r355, -2139062144;
or.b32 %r357, %r296, %r353;
xor.b32 %r358, %r355, %r296;
and.b32 %r294, %r358, %r357;

prmt.b32 %r294, %r294, 0, 0xba98;


prmt.b32 %r296, %r296, 0, 0xba98;

xor.b32 %r359, %r296, 2139062143;
not.b32 %r360, %r294;
and.b32 %r361, %r356, %r360;
and.b32 %r362, %r359, %r294;
or.b32 %r299, %r362, %r361;

dp4a.s32.s32 %r298, %r299, %r37, %r185;

mul.lo.s32 %r363, %r350, %r298;
cvt.rn.f32.s32 %f60, %r363;
fma.rn.ftz.f32 %f61, %f55, %f60, %f59;
add.s64 %rd112, %rd45, %rd83;
add.s64 %rd113, %rd112, %rd123;
ld.global.u8 %r364, [%rd113+96];
shr.u32 %r365, %r364, %r25;
and.b32 %r366, %r365, 15;
add.s64 %rd114, %rd7, %rd123;
ld.global.u8 %r367, [%rd114];
shr.u32 %r368, %r367, %r26;
and.b32 %r369, %r368, 3;
bfi.b32 %r370, %r369, %r366, 4, 2;
add.s32 %r371, %r370, -32;
shr.u32 %r372, %r321, 4;
and.b32 %r304, %r372, 50529027;
and.b32 %r373, %r322, 67372036;
or.b32 %r374, %r304, -2139062144;
sub.s32 %r375, %r374, %r373;
xor.b32 %r376, %r375, -2139062144;
or.b32 %r377, %r304, %r373;
xor.b32 %r378, %r375, %r304;
and.b32 %r302, %r378, %r377;

prmt.b32 %r302, %r302, 0, 0xba98;


prmt.b32 %r304, %r304, 0, 0xba98;

xor.b32 %r379, %r304, 2139062143;
not.b32 %r380, %r302;
and.b32 %r381, %r376, %r380;
and.b32 %r382, %r379, %r302;
or.b32 %r307, %r382, %r381;

dp4a.s32.s32 %r306, %r307, %r39, %r185;

mul.lo.s32 %r383, %r371, %r306;
cvt.rn.f32.s32 %f62, %r383;
fma.rn.ftz.f32 %f63, %f56, %f62, %f61;
add.s64 %rd116, %rd45, %rd89;
add.s64 %rd117, %rd116, %rd123;
ld.global.u8 %r384, [%rd117+96];
shr.u32 %r385, %r384, %r29;
and.b32 %r386, %r385, 15;
add.s64 %rd118, %rd5, %rd123;
ld.global.u8 %r387, [%rd118];
shr.u32 %r388, %r387, %r30;
and.b32 %r389, %r388, 3;
bfi.b32 %r390, %r389, %r386, 4, 2;
add.s32 %r391, %r390, -32;
shr.u32 %r392, %r321, 6;
and.b32 %r312, %r392, 50529027;
shr.u32 %r393, %r322, 1;
and.b32 %r394, %r393, 67372036;
or.b32 %r395, %r312, -2139062144;
sub.s32 %r396, %r395, %r394;
xor.b32 %r397, %r396, -2139062144;
or.b32 %r398, %r312, %r394;
xor.b32 %r399, %r396, %r312;
and.b32 %r310, %r399, %r398;

prmt.b32 %r310, %r310, 0, 0xba98;


prmt.b32 %r312, %r312, 0, 0xba98;

xor.b32 %r400, %r312, 2139062143;
not.b32 %r401, %r310;
and.b32 %r402, %r397, %r401;
and.b32 %r403, %r400, %r310;
or.b32 %r315, %r403, %r402;

dp4a.s32.s32 %r314, %r315, %r41, %r185;

mul.lo.s32 %r404, %r391, %r314;
cvt.rn.f32.s32 %f64, %r404;
fma.rn.ftz.f32 %f65, %f57, %f64, %f63;
fma.rn.ftz.f32 %f132, %f53, %f65, %f132;

$L__BB144_12:
add.s64 %rd123, %rd123, 880;
add.s32 %r485, %r485, 64;
add.s32 %r486, %r486, 8;
setp.lt.s32 %p11, %r486, %r81;
@%p11 bra $L__BB144_10;

$L__BB144_13:
mov.u32 %r468, %tid.y;
setp.eq.s32 %p12, %r468, 0;
@%p12 bra $L__BB144_16;

mov.u32 %r470, %tid.y;
mov.u32 %r469, %tid.x;
setp.eq.s64 %p13, %rd17, 0;
add.s32 %r45, %r470, -1;
shl.b32 %r409, %r470, 7;
mov.u32 %r410, _ZZ13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r411, %r410, %r409;
shl.b32 %r412, %r469, 2;
add.s32 %r413, %r411, %r412;
st.shared.f32 [%r413+-128], %f134;
@%p13 bra $L__BB144_16;

shl.b32 %r415, %r45, 7;
mov.u32 %r416, _ZZ13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r417, %r416, %r415;
add.s32 %r419, %r417, %r412;
st.shared.f32 [%r419], %f132;

$L__BB144_16:
mov.u32 %r471, %tid.y;
setp.ne.s32 %p14, %r471, 0;
bar.sync 0;
@%p14 bra $L__BB144_35;

mov.u32 %r472, %tid.x;
setp.ne.s64 %p15, %rd17, 0;
shl.b32 %r422, %r472, 2;
mov.u32 %r423, _ZZ13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r46, %r423, %r422;
ld.shared.f32 %f66, [%r46];
add.ftz.f32 %f13, %f134, %f66;
@%p15 bra $L__BB144_19;
bra.uni $L__BB144_18;

$L__BB144_19:
mov.u32 %r426, _ZZ13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r427, %r426, %r422;
ld.shared.f32 %f70, [%r427];
add.ftz.f32 %f71, %f132, %f70;
ld.shared.f32 %f72, [%r46+128];
add.ftz.f32 %f73, %f13, %f72;
ld.shared.f32 %f74, [%r427+128];
add.ftz.f32 %f75, %f71, %f74;
ld.shared.f32 %f76, [%r46+256];
add.ftz.f32 %f136, %f73, %f76;
ld.shared.f32 %f77, [%r427+256];
add.ftz.f32 %f132, %f75, %f77;
bra.uni $L__BB144_20;

$L__BB144_18:
ld.shared.f32 %f67, [%r46+128];
add.ftz.f32 %f68, %f13, %f67;
ld.shared.f32 %f69, [%r46+256];
add.ftz.f32 %f136, %f68, %f69;

$L__BB144_20:
mov.b32 %r428, %f136;
mov.u32 %r429, 31;
mov.u32 %r430, 16;
mov.u32 %r431, -1;
shfl.sync.bfly.b32 %r432|%p16, %r428, %r430, %r429, %r431;
mov.b32 %f78, %r432;
add.ftz.f32 %f79, %f136, %f78;
mov.b32 %r433, %f79;
mov.u32 %r434, 8;
shfl.sync.bfly.b32 %r435|%p17, %r433, %r434, %r429, %r431;
mov.b32 %f80, %r435;
add.ftz.f32 %f81, %f79, %f80;
mov.b32 %r436, %f81;
mov.u32 %r437, 4;
shfl.sync.bfly.b32 %r438|%p18, %r436, %r437, %r429, %r431;
mov.b32 %f82, %r438;
add.ftz.f32 %f83, %f81, %f82;
mov.b32 %r439, %f83;
mov.u32 %r440, 2;
shfl.sync.bfly.b32 %r441|%p19, %r439, %r440, %r429, %r431;
mov.b32 %f84, %r441;
add.ftz.f32 %f19, %f83, %f84;
mov.b32 %r442, %f19;
mov.u32 %r443, 1;
shfl.sync.bfly.b32 %r47|%p1, %r442, %r443, %r429, %r431;
setp.eq.s64 %p20, %rd17, 0;
@%p20 bra $L__BB144_22;

mov.b32 %r444, %f132;
shfl.sync.bfly.b32 %r448|%p21, %r444, %r430, %r429, %r431;
mov.b32 %f85, %r448;
add.ftz.f32 %f86, %f132, %f85;
mov.b32 %r449, %f86;
shfl.sync.bfly.b32 %r451|%p22, %r449, %r434, %r429, %r431;
mov.b32 %f87, %r451;
add.ftz.f32 %f88, %f86, %f87;
mov.b32 %r452, %f88;
shfl.sync.bfly.b32 %r454|%p23, %r452, %r437, %r429, %r431;
mov.b32 %f89, %r454;
add.ftz.f32 %f90, %f88, %f89;
mov.b32 %r455, %f90;
shfl.sync.bfly.b32 %r457|%p24, %r455, %r440, %r429, %r431;
mov.b32 %f91, %r457;
add.ftz.f32 %f92, %f90, %f91;
mov.b32 %r458, %f92;
shfl.sync.bfly.b32 %r460|%p25, %r458, %r443, %r429, %r431;
mov.b32 %f93, %r460;
add.ftz.f32 %f132, %f92, %f93;

$L__BB144_22:
mov.u32 %r473, %tid.x;
setp.ne.s32 %p26, %r473, 0;
mov.b32 %f94, %r47;
add.ftz.f32 %f22, %f19, %f94;
@%p26 bra $L__BB144_35;

add.ftz.f32 %f138, %f129, %f22;
@%p20 bra $L__BB144_34;

ld.param.u32 %r478, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
add.ftz.f32 %f24, %f128, %f132;
setp.gt.s32 %p28, %r478, 2;
@%p28 bra $L__BB144_28;

ld.param.u32 %r481, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p31, %r481, 1;
@%p31 bra $L__BB144_32;

ld.param.u32 %r482, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p32, %r482, 2;
@%p32 bra $L__BB144_27;
bra.uni $L__BB144_33;

$L__BB144_27:
mul.ftz.f32 %f114, %f24, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f115, %f114;
add.ftz.f32 %f116, %f115, 0f3F800000;
div.approx.ftz.f32 %f117, %f24, %f116;
mul.ftz.f32 %f138, %f138, %f117;
bra.uni $L__BB144_34;

$L__BB144_28:
ld.param.u32 %r479, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p29, %r479, 3;
@%p29 bra $L__BB144_31;

ld.param.u32 %r480, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.ne.s32 %p30, %r480, 6;
@%p30 bra $L__BB144_33;

ld.param.f32 %f126, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
min.ftz.f32 %f95, %f24, %f126;
neg.ftz.f32 %f96, %f126;
min.ftz.f32 %f97, %f138, %f126;
max.ftz.f32 %f98, %f97, %f96;
mul.ftz.f32 %f99, %f95, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f100, %f99;
add.ftz.f32 %f101, %f100, 0f3F800000;
div.approx.ftz.f32 %f102, %f95, %f101;
mul.ftz.f32 %f138, %f98, %f102;
bra.uni $L__BB144_34;

$L__BB144_32:
mul.ftz.f32 %f118, %f24, 0f3F000000;
mul.ftz.f32 %f119, %f24, 0f3D372713;
fma.rn.ftz.f32 %f120, %f24, %f119, 0f3F800000;
mul.ftz.f32 %f121, %f24, 0f3F4C422A;
mul.ftz.f32 %f122, %f121, %f120;
tanh.approx.f32 %f123, %f122;
add.ftz.f32 %f124, %f123, 0f3F800000;
mul.ftz.f32 %f125, %f118, %f124;
mul.ftz.f32 %f138, %f138, %f125;
bra.uni $L__BB144_34;

$L__BB144_31:
mov.f32 %f103, 0f40E00000;
min.ftz.f32 %f104, %f24, %f103;
min.ftz.f32 %f105, %f138, %f103;
mov.f32 %f106, 0fC0E00000;
max.ftz.f32 %f107, %f105, %f106;
mul.ftz.f32 %f108, %f104, 0f3FD9DB23;
mul.ftz.f32 %f109, %f108, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f110, %f109;
add.ftz.f32 %f111, %f110, 0f3F800000;
div.approx.ftz.f32 %f112, %f104, %f111;
add.ftz.f32 %f113, %f107, 0f3F800000;
mul.ftz.f32 %f138, %f113, %f112;
bra.uni $L__BB144_34;

$L__BB144_33:
mul.ftz.f32 %f138, %f138, %f24;

$L__BB144_34:
ld.param.u64 %rd122, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
mov.u32 %r477, %ctaid.x;
ld.param.u32 %r476, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r475, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
mov.u32 %r474, %ctaid.z;
cvta.to.global.u64 %rd119, %rd122;
mov.u32 %r464, %ctaid.y;
mad.lo.s32 %r466, %r464, %r476, %r477;
mad.lo.s32 %r467, %r474, %r475, %r466;
mul.wide.u32 %rd120, %r467, 4;
add.s64 %rd121, %rd119, %rd120;
st.global.f32 [%rd121], %f138;

$L__BB144_35:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<12>;
.reg .b16 %rs<14>;
.reg .f32 %f<37>;
.reg .b32 %r<297>;
.reg .b64 %rd<50>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[384];

ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
mov.u32 %r1, %tid.y;
shl.b32 %r44, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r44, %r2;
mov.u32 %r4, %ctaid.x;
mov.u32 %r294, %ctaid.y;
setp.eq.s64 %p1, %rd18, 0;
@%p1 bra $L__BB145_2;

cvta.to.global.u64 %rd20, %rd18;
mul.wide.u32 %rd21, %r294, 4;
add.s64 %rd22, %rd20, %rd21;
ld.global.nc.u32 %r293, [%rd22];
mul.hi.u32 %r45, %r294, %r28;
add.s32 %r46, %r45, %r294;
shr.u32 %r47, %r46, %r29;
mul.lo.s32 %r48, %r47, %r30;
sub.s32 %r294, %r294, %r48;
bra.uni $L__BB145_3;

$L__BB145_2:
mul.hi.u32 %r49, %r294, %r32;
add.s32 %r50, %r49, %r294;
shr.u32 %r293, %r50, %r33;

$L__BB145_3:
shr.s32 %r51, %r3, 31;
shr.u32 %r52, %r51, 28;
add.s32 %r53, %r3, %r52;
shr.s32 %r296, %r53, 4;
shr.u32 %r54, %r27, 8;
setp.ge.s32 %p2, %r296, %r54;
mov.f32 %f36, 0f00000000;
@%p2 bra $L__BB145_6;

mov.u32 %r55, %ctaid.z;
and.b32 %r59, %r53, -16;
sub.s32 %r60, %r3, %r59;
shr.s32 %r61, %r60, 31;
shr.u32 %r62, %r61, 29;
add.s32 %r63, %r60, %r62;
shr.s32 %r64, %r63, 3;
shl.b32 %r65, %r64, 2;
and.b32 %r66, %r63, -8;
sub.s32 %r12, %r60, %r66;
shr.s32 %r67, %r12, 31;
shr.u32 %r68, %r67, 30;
add.s32 %r69, %r12, %r68;
shr.s32 %r70, %r69, 2;
add.s32 %r71, %r66, %r70;
shl.b32 %r72, %r60, 1;
shr.s32 %r73, %r71, 31;
shr.u32 %r74, %r73, 29;
add.s32 %r75, %r71, %r74;
and.b32 %r76, %r75, -8;
sub.s32 %r77, %r71, %r76;
shr.s32 %r78, %r75, 3;
shl.b32 %r13, %r78, 2;
cvt.s64.s32 %rd1, %r77;
cvt.s64.s32 %rd2, %r65;
shr.u32 %r79, %r73, 30;
add.s32 %r80, %r71, %r79;
and.b32 %r81, %r80, -4;
sub.s32 %r82, %r71, %r81;
shr.s32 %r83, %r80, 2;
shl.b32 %r14, %r83, 1;
add.s32 %r84, %r82, 8;
cvt.s64.s32 %rd3, %r84;
add.s32 %r85, %r71, 2;
shr.s32 %r86, %r85, 31;
shr.u32 %r87, %r86, 29;
add.s32 %r88, %r85, %r87;
and.b32 %r89, %r88, -8;
sub.s32 %r90, %r85, %r89;
shr.s32 %r91, %r88, 3;
shl.b32 %r15, %r91, 2;
cvt.s64.s32 %rd4, %r90;
shr.u32 %r92, %r86, 30;
add.s32 %r93, %r85, %r92;
and.b32 %r94, %r93, -4;
sub.s32 %r95, %r85, %r94;
shr.s32 %r96, %r93, 2;
shl.b32 %r16, %r96, 1;
add.s32 %r97, %r95, 8;
cvt.s64.s32 %rd5, %r97;
add.s32 %r98, %r71, 4;
shr.s32 %r99, %r98, 31;
shr.u32 %r100, %r99, 29;
add.s32 %r101, %r98, %r100;
and.b32 %r102, %r101, -8;
sub.s32 %r103, %r98, %r102;
shr.s32 %r104, %r101, 3;
shl.b32 %r17, %r104, 2;
cvt.s64.s32 %rd6, %r103;
shr.u32 %r105, %r99, 30;
add.s32 %r106, %r98, %r105;
and.b32 %r107, %r106, -4;
sub.s32 %r108, %r98, %r107;
shr.s32 %r109, %r106, 2;
shl.b32 %r18, %r109, 1;
add.s32 %r110, %r108, 8;
cvt.s64.s32 %rd7, %r110;
add.s32 %r111, %r71, 6;
shr.s32 %r112, %r111, 31;
shr.u32 %r113, %r112, 29;
add.s32 %r114, %r111, %r113;
and.b32 %r115, %r114, -8;
sub.s32 %r116, %r111, %r115;
shr.s32 %r117, %r114, 3;
shl.b32 %r19, %r117, 2;
cvt.s64.s32 %rd8, %r116;
shr.u32 %r118, %r112, 30;
add.s32 %r119, %r111, %r118;
and.b32 %r120, %r119, -4;
sub.s32 %r121, %r111, %r120;
shr.s32 %r122, %r119, 2;
shl.b32 %r20, %r122, 1;
add.s32 %r123, %r121, 8;
cvt.s64.s32 %rd9, %r123;
mul.hi.u32 %r124, %r55, %r38;
add.s32 %r125, %r124, %r55;
shr.u32 %r126, %r125, %r39;
mad.lo.s32 %r127, %r126, %r41, %r296;
mad.lo.s32 %r128, %r293, %r35, %r127;
mad.lo.s32 %r129, %r4, %r31, %r128;
mul.wide.s32 %rd23, %r129, 110;
cvta.to.global.u64 %rd24, %rd16;
add.s64 %rd25, %rd24, %rd23;
add.s64 %rd49, %rd25, 108;
shl.b32 %r295, %r296, 3;
mul.wide.s32 %rd11, %r72, 2;
cvta.to.global.u64 %rd12, %rd17;
mul.lo.s32 %r130, %r294, %r36;
cvt.u64.u32 %rd26, %r130;
mul.lo.s32 %r131, %r55, %r42;
cvt.u64.u32 %rd27, %r131;
add.s64 %rd13, %rd26, %rd27;
mov.f32 %f36, 0f00000000;

$L__BB145_5:
cvt.u64.u32 %rd28, %r295;
add.s64 %rd29, %rd13, %rd28;
ld.global.nc.u16 %rs1, [%rd49];

{ cvt.f32.f16 %f7, %rs1;}


add.s64 %rd30, %rd49, %rd11;
shl.b32 %r168, %r12, 1;
mul.wide.s32 %rd31, %r168, 2;
add.s64 %rd32, %rd49, %rd31;
ld.global.nc.u16 %rs2, [%rd32+-108];
ld.global.nc.u16 %rs3, [%rd32+-106];
mov.b32 %r169, {%rs2, %rs3};
not.b32 %r170, %r169;
ld.global.nc.u16 %rs4, [%rd30+-74];
ld.global.nc.u16 %rs5, [%rd30+-76];
add.s64 %rd33, %rd29, %rd2;
mul.lo.s64 %rd34, %rd33, 36;
add.s64 %rd35, %rd12, %rd34;
mul.wide.s32 %rd36, %r12, 4;
add.s64 %rd37, %rd35, %rd36;
ld.global.nc.u32 %r142, [%rd37+4];
ld.global.nc.u32 %r132, [%rd35];

{.reg .f16 low,high;
mov.b32 {low,high},%r132;
cvt.f32.f16 %f8, low;}


ld.global.nc.u32 %r150, [%rd37+40];
ld.global.nc.u32 %r133, [%rd35+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r133;
cvt.f32.f16 %f9, low;}


ld.global.nc.u32 %r158, [%rd37+76];
ld.global.nc.u32 %r134, [%rd35+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r134;
cvt.f32.f16 %f10, low;}


ld.global.nc.u32 %r166, [%rd37+112];
ld.global.nc.u32 %r135, [%rd35+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r135;
cvt.f32.f16 %f11, low;}


mov.b32 %r171, {%rs5, %rs4};
cvt.u32.u64 %r172, %rd2;
shr.s32 %r173, %r170, %r172;
add.s64 %rd38, %rd49, %rd1;
ld.global.nc.u8 %rs6, [%rd38+-12];
cvt.u32.u16 %r174, %rs6;
and.b32 %r175, %r174, 255;
shr.u32 %r176, %r175, %r13;
and.b32 %r177, %r176, 15;
add.s64 %rd39, %rd49, %rd3;
ld.global.nc.u8 %rs7, [%rd39+-12];
cvt.u32.u16 %r178, %rs7;
and.b32 %r179, %r178, 255;
shr.u32 %r180, %r179, %r14;
and.b32 %r181, %r180, 3;
bfi.b32 %r182, %r181, %r177, 4, 2;
add.s32 %r183, %r182, -32;
and.b32 %r138, %r171, 50529027;
shl.b32 %r184, %r173, 2;
and.b32 %r185, %r184, 67372036;
or.b32 %r186, %r138, -2139062144;
sub.s32 %r187, %r186, %r185;
xor.b32 %r188, %r187, -2139062144;
or.b32 %r189, %r138, %r185;
xor.b32 %r190, %r187, %r138;
and.b32 %r136, %r190, %r189;

prmt.b32 %r136, %r136, 0, 0xba98;


prmt.b32 %r138, %r138, 0, 0xba98;

xor.b32 %r191, %r138, 2139062143;
not.b32 %r192, %r136;
and.b32 %r193, %r188, %r192;
and.b32 %r194, %r191, %r136;
or.b32 %r141, %r194, %r193;
mov.u32 %r167, 0;

dp4a.s32.s32 %r140, %r141, %r142, %r167;

mul.lo.s32 %r195, %r183, %r140;
cvt.rn.f32.s32 %f12, %r195;
fma.rn.ftz.f32 %f13, %f8, %f12, 0f00000000;
add.s64 %rd40, %rd49, %rd4;
ld.global.nc.u8 %rs8, [%rd40+-12];
cvt.u32.u16 %r196, %rs8;
and.b32 %r197, %r196, 255;
shr.u32 %r198, %r197, %r15;
and.b32 %r199, %r198, 15;
add.s64 %rd41, %rd49, %rd5;
ld.global.nc.u8 %rs9, [%rd41+-12];
cvt.u32.u16 %r200, %rs9;
and.b32 %r201, %r200, 255;
shr.u32 %r202, %r201, %r16;
and.b32 %r203, %r202, 3;
bfi.b32 %r204, %r203, %r199, 4, 2;
add.s32 %r205, %r204, -32;
shr.u32 %r206, %r171, 2;
and.b32 %r146, %r206, 50529027;
shl.b32 %r207, %r173, 1;
and.b32 %r208, %r207, 67372036;
or.b32 %r209, %r146, -2139062144;
sub.s32 %r210, %r209, %r208;
xor.b32 %r211, %r210, -2139062144;
or.b32 %r212, %r146, %r208;
xor.b32 %r213, %r210, %r146;
and.b32 %r144, %r213, %r212;

prmt.b32 %r144, %r144, 0, 0xba98;


prmt.b32 %r146, %r146, 0, 0xba98;

xor.b32 %r214, %r146, 2139062143;
not.b32 %r215, %r144;
and.b32 %r216, %r211, %r215;
and.b32 %r217, %r214, %r144;
or.b32 %r149, %r217, %r216;

dp4a.s32.s32 %r148, %r149, %r150, %r167;

mul.lo.s32 %r218, %r205, %r148;
cvt.rn.f32.s32 %f14, %r218;
fma.rn.ftz.f32 %f15, %f9, %f14, %f13;
add.s64 %rd42, %rd49, %rd6;
ld.global.nc.u8 %rs10, [%rd42+-12];
cvt.u32.u16 %r219, %rs10;
and.b32 %r220, %r219, 255;
shr.u32 %r221, %r220, %r17;
and.b32 %r222, %r221, 15;
add.s64 %rd43, %rd49, %rd7;
ld.global.nc.u8 %rs11, [%rd43+-12];
cvt.u32.u16 %r223, %rs11;
and.b32 %r224, %r223, 255;
shr.u32 %r225, %r224, %r18;
and.b32 %r226, %r225, 3;
bfi.b32 %r227, %r226, %r222, 4, 2;
add.s32 %r228, %r227, -32;
shr.u32 %r229, %r171, 4;
and.b32 %r154, %r229, 50529027;
and.b32 %r230, %r173, 67372036;
or.b32 %r231, %r154, -2139062144;
sub.s32 %r232, %r231, %r230;
xor.b32 %r233, %r232, -2139062144;
or.b32 %r234, %r154, %r230;
xor.b32 %r235, %r232, %r154;
and.b32 %r152, %r235, %r234;

prmt.b32 %r152, %r152, 0, 0xba98;


prmt.b32 %r154, %r154, 0, 0xba98;

xor.b32 %r236, %r154, 2139062143;
not.b32 %r237, %r152;
and.b32 %r238, %r233, %r237;
and.b32 %r239, %r236, %r152;
or.b32 %r157, %r239, %r238;

dp4a.s32.s32 %r156, %r157, %r158, %r167;

mul.lo.s32 %r240, %r228, %r156;
cvt.rn.f32.s32 %f16, %r240;
fma.rn.ftz.f32 %f17, %f10, %f16, %f15;
add.s64 %rd44, %rd49, %rd8;
ld.global.nc.u8 %rs12, [%rd44+-12];
cvt.u32.u16 %r241, %rs12;
and.b32 %r242, %r241, 255;
shr.u32 %r243, %r242, %r19;
and.b32 %r244, %r243, 15;
add.s64 %rd45, %rd49, %rd9;
ld.global.nc.u8 %rs13, [%rd45+-12];
cvt.u32.u16 %r245, %rs13;
and.b32 %r246, %r245, 255;
shr.u32 %r247, %r246, %r20;
and.b32 %r248, %r247, 3;
bfi.b32 %r249, %r248, %r244, 4, 2;
add.s32 %r250, %r249, -32;
shr.u32 %r251, %r171, 6;
and.b32 %r162, %r251, 50529027;
shr.u32 %r252, %r173, 1;
and.b32 %r253, %r252, 67372036;
or.b32 %r254, %r162, -2139062144;
sub.s32 %r255, %r254, %r253;
xor.b32 %r256, %r255, -2139062144;
or.b32 %r257, %r162, %r253;
xor.b32 %r258, %r255, %r162;
and.b32 %r160, %r258, %r257;

prmt.b32 %r160, %r160, 0, 0xba98;


prmt.b32 %r162, %r162, 0, 0xba98;

xor.b32 %r259, %r162, 2139062143;
not.b32 %r260, %r160;
and.b32 %r261, %r256, %r260;
and.b32 %r262, %r259, %r160;
or.b32 %r165, %r262, %r261;

dp4a.s32.s32 %r164, %r165, %r166, %r167;

mul.lo.s32 %r263, %r250, %r164;
cvt.rn.f32.s32 %f18, %r263;
fma.rn.ftz.f32 %f19, %f11, %f18, %f17;
fma.rn.ftz.f32 %f36, %f7, %f19, %f36;
add.s64 %rd49, %rd49, 880;
add.s32 %r295, %r295, 64;
add.s32 %r296, %r296, 8;
setp.lt.s32 %p3, %r296, %r54;
@%p3 bra $L__BB145_5;

$L__BB145_6:
setp.eq.s32 %p4, %r1, 0;
@%p4 bra $L__BB145_8;

shl.b32 %r265, %r1, 7;
mov.u32 %r266, _ZZ13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r267, %r266, %r265;
shl.b32 %r268, %r2, 2;
add.s32 %r269, %r267, %r268;
st.shared.f32 [%r269+-128], %f36;

$L__BB145_8:
setp.ne.s32 %p5, %r1, 0;
bar.sync 0;
@%p5 bra $L__BB145_11;

mov.u32 %r270, %ctaid.z;
mov.u32 %r271, %ctaid.y;
mad.lo.s32 %r272, %r271, %r37, %r4;
mad.lo.s32 %r26, %r270, %r43, %r272;
shl.b32 %r273, %r2, 2;
mov.u32 %r274, _ZZ13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r275, %r274, %r273;
mov.u32 %r276, 2;
ld.shared.f32 %f20, [%r275];
add.ftz.f32 %f21, %f36, %f20;
ld.shared.f32 %f22, [%r275+128];
add.ftz.f32 %f23, %f21, %f22;
ld.shared.f32 %f24, [%r275+256];
add.ftz.f32 %f25, %f23, %f24;
mov.b32 %r277, %f25;
mov.u32 %r278, 31;
mov.u32 %r279, 16;
mov.u32 %r280, -1;
shfl.sync.bfly.b32 %r281|%p6, %r277, %r279, %r278, %r280;
mov.b32 %f26, %r281;
add.ftz.f32 %f27, %f25, %f26;
mov.b32 %r282, %f27;
mov.u32 %r283, 8;
shfl.sync.bfly.b32 %r284|%p7, %r282, %r283, %r278, %r280;
mov.b32 %f28, %r284;
add.ftz.f32 %f29, %f27, %f28;
mov.b32 %r285, %f29;
mov.u32 %r286, 4;
shfl.sync.bfly.b32 %r287|%p8, %r285, %r286, %r278, %r280;
mov.b32 %f30, %r287;
add.ftz.f32 %f31, %f29, %f30;
mov.b32 %r288, %f31;
shfl.sync.bfly.b32 %r289|%p9, %r288, %r276, %r278, %r280;
mov.b32 %f32, %r289;
add.ftz.f32 %f33, %f31, %f32;
mov.b32 %r290, %f33;
mov.u32 %r291, 1;
shfl.sync.bfly.b32 %r292|%p10, %r290, %r291, %r278, %r280;
mov.b32 %f34, %r292;
add.ftz.f32 %f4, %f33, %f34;
setp.ne.s32 %p11, %r2, 0;
@%p11 bra $L__BB145_11;

cvta.to.global.u64 %rd46, %rd19;
mul.wide.u32 %rd47, %r26, 4;
add.s64 %rd48, %rd46, %rd47;
st.global.f32 [%rd48], %f4;

$L__BB145_11:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type11ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type11ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type11ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type11ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type11ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type11ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type11ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type11ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type11ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<35>;
.reg .b16 %rs<29>;
.reg .f32 %f<145>;
.reg .b32 %r<615>;
.reg .b64 %rd<114>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type11ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

ld.param.u64 %rd31, [_Z13mul_mat_vec_qIL9ggml_type11ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd32, [_Z13mul_mat_vec_qIL9ggml_type11ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type11ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type11ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type11ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type11ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type11ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type11ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type11ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type11ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type11ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type11ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type11ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r37, %tid.y;
shl.b32 %r38, %r37, 5;
mov.u32 %r39, %tid.x;
add.s32 %r1, %r38, %r39;
shr.s32 %r40, %r1, 31;
shr.u32 %r41, %r40, 28;
add.s32 %r42, %r1, %r41;
shr.s32 %r614, %r42, 4;
shr.u32 %r43, %r21, 8;
setp.ge.s32 %p3, %r614, %r43;
mov.f32 %f141, 0f00000000;
mov.f32 %f142, %f141;
mov.f32 %f143, %f141;
mov.f32 %f144, %f141;
@%p3 bra $L__BB146_3;

mov.u32 %r44, %ctaid.x;
shl.b32 %r45, %r44, 1;
mov.u32 %r46, %ctaid.y;
mov.u32 %r47, %ctaid.z;
mul.lo.s32 %r48, %r46, %r29;
cvt.u64.u32 %rd34, %r48;
mul.lo.s32 %r49, %r47, %r35;
cvt.u64.u32 %rd35, %r49;
add.s64 %rd1, %rd35, %rd34;
mul.hi.u32 %r50, %r47, %r31;
add.s32 %r51, %r50, %r47;
shr.u32 %r52, %r51, %r32;
mul.lo.s32 %r53, %r52, %r34;
mul.hi.u32 %r54, %r46, %r25;
add.s32 %r55, %r54, %r46;
shr.u32 %r56, %r55, %r26;
mul.lo.s32 %r57, %r56, %r28;
cvta.to.global.u64 %rd113, %rd31;
and.b32 %r61, %r42, -16;
sub.s32 %r62, %r1, %r61;
shr.s32 %r63, %r62, 31;
shr.u32 %r64, %r63, 29;
add.s32 %r65, %r62, %r64;
shr.s32 %r66, %r65, 3;
shl.b32 %r67, %r66, 2;
and.b32 %r68, %r65, -8;
sub.s32 %r69, %r62, %r68;
shr.s32 %r70, %r69, 31;
shr.u32 %r71, %r70, 30;
add.s32 %r72, %r69, %r71;
shr.s32 %r73, %r72, 2;
add.s32 %r74, %r68, %r73;
shl.b32 %r75, %r62, 1;
shl.b32 %r76, %r69, 1;
cvt.s64.s32 %rd3, %r69;
or.b32 %r77, %r45, 1;
shr.s32 %r78, %r74, 31;
shr.u32 %r79, %r78, 29;
add.s32 %r80, %r74, %r79;
and.b32 %r81, %r80, -8;
sub.s32 %r82, %r74, %r81;
cvt.s64.s32 %rd36, %r82;
shr.s32 %r83, %r80, 3;
shl.b32 %r3, %r83, 2;
cvt.s64.s32 %rd4, %r67;
shr.u32 %r84, %r78, 30;
add.s32 %r85, %r74, %r84;
and.b32 %r86, %r85, -4;
sub.s32 %r87, %r74, %r86;
shr.s32 %r88, %r85, 2;
shl.b32 %r4, %r88, 1;
add.s32 %r89, %r87, 8;
cvt.s64.s32 %rd37, %r89;
add.s32 %r90, %r74, 2;
shr.s32 %r91, %r90, 31;
shr.u32 %r92, %r91, 29;
add.s32 %r93, %r90, %r92;
and.b32 %r94, %r93, -8;
sub.s32 %r95, %r90, %r94;
shr.s32 %r96, %r93, 3;
shl.b32 %r5, %r96, 2;
cvt.s64.s32 %rd38, %r95;
shr.u32 %r97, %r91, 30;
add.s32 %r98, %r90, %r97;
and.b32 %r99, %r98, -4;
sub.s32 %r100, %r90, %r99;
shr.s32 %r101, %r98, 2;
shl.b32 %r6, %r101, 1;
add.s32 %r102, %r100, 8;
cvt.s64.s32 %rd39, %r102;
add.s32 %r103, %r74, 4;
shr.s32 %r104, %r103, 31;
shr.u32 %r105, %r104, 29;
add.s32 %r106, %r103, %r105;
and.b32 %r107, %r106, -8;
sub.s32 %r108, %r103, %r107;
shr.s32 %r109, %r106, 3;
shl.b32 %r7, %r109, 2;
cvt.s64.s32 %rd40, %r108;
shr.u32 %r110, %r104, 30;
add.s32 %r111, %r103, %r110;
and.b32 %r112, %r111, -4;
sub.s32 %r113, %r103, %r112;
shr.s32 %r114, %r111, 2;
shl.b32 %r8, %r114, 1;
add.s32 %r115, %r113, 8;
cvt.s64.s32 %rd41, %r115;
add.s32 %r116, %r74, 6;
shr.s32 %r117, %r116, 31;
shr.u32 %r118, %r117, 29;
add.s32 %r119, %r116, %r118;
and.b32 %r120, %r119, -8;
sub.s32 %r121, %r116, %r120;
shr.s32 %r122, %r119, 3;
shl.b32 %r9, %r122, 2;
cvt.s64.s32 %rd42, %r121;
shr.u32 %r123, %r117, 30;
add.s32 %r124, %r116, %r123;
and.b32 %r125, %r124, -4;
sub.s32 %r126, %r116, %r125;
shr.s32 %r127, %r124, 2;
shl.b32 %r10, %r127, 1;
add.s32 %r128, %r126, 8;
cvt.s64.s32 %rd43, %r128;
shl.b32 %r612, %r614, 3;
add.s32 %r613, %r23, %r612;
add.s32 %r129, %r614, %r53;
add.s32 %r130, %r129, %r57;
mad.lo.s32 %r131, %r77, %r22, %r130;
mul.wide.s32 %rd5, %r131, 110;
add.s64 %rd44, %rd5, %rd43;
add.s64 %rd6, %rd44, 96;
add.s64 %rd45, %rd5, %rd42;
add.s64 %rd7, %rd45, 96;
add.s64 %rd46, %rd5, %rd41;
add.s64 %rd8, %rd46, 96;
add.s64 %rd47, %rd5, %rd40;
add.s64 %rd9, %rd47, 96;
add.s64 %rd48, %rd5, %rd39;
add.s64 %rd10, %rd48, 96;
add.s64 %rd49, %rd5, %rd38;
add.s64 %rd11, %rd49, 96;
add.s64 %rd50, %rd5, %rd37;
add.s64 %rd12, %rd50, 96;
add.s64 %rd51, %rd5, %rd36;
add.s64 %rd13, %rd51, 96;
mad.lo.s32 %r132, %r45, %r22, %r614;
add.s32 %r133, %r132, %r53;
add.s32 %r134, %r133, %r57;
mul.wide.s32 %rd14, %r134, 110;
mul.wide.s32 %rd52, %r75, 2;
add.s64 %rd15, %rd14, %rd52;
mul.wide.s32 %rd53, %r76, 2;
add.s64 %rd16, %rd5, %rd53;
add.s64 %rd54, %rd14, %rd53;
add.s64 %rd17, %rd54, 2;
add.s64 %rd55, %rd5, %rd52;
add.s64 %rd18, %rd55, 32;
add.s64 %rd56, %rd14, %rd43;
add.s64 %rd19, %rd56, 96;
add.s64 %rd57, %rd14, %rd42;
add.s64 %rd20, %rd57, 96;
add.s64 %rd58, %rd14, %rd41;
add.s64 %rd21, %rd58, 96;
add.s64 %rd59, %rd14, %rd40;
add.s64 %rd22, %rd59, 96;
add.s64 %rd60, %rd14, %rd39;
add.s64 %rd23, %rd60, 96;
add.s64 %rd61, %rd14, %rd38;
add.s64 %rd24, %rd61, 96;
add.s64 %rd62, %rd14, %rd37;
add.s64 %rd25, %rd62, 96;
add.s64 %rd63, %rd14, %rd36;
add.s64 %rd26, %rd63, 96;
mov.f32 %f141, 0f00000000;
cvta.to.global.u64 %rd64, %rd32;
shl.b64 %rd73, %rd3, 2;

$L__BB146_2:
cvt.u64.u32 %rd65, %r612;
add.s64 %rd66, %rd1, %rd65;
add.s64 %rd67, %rd113, %rd14;
ld.global.nc.u16 %rs1, [%rd67+108];

{ cvt.f32.f16 %f25, %rs1;}


add.s64 %rd68, %rd113, %rd15;
add.s64 %rd69, %rd113, %rd17;
ld.global.nc.u16 %rs5, [%rd69+-2];
ld.global.nc.u16 %rs6, [%rd69];
mov.b32 %r279, {%rs5, %rs6};
not.b32 %r280, %r279;
ld.global.nc.u16 %rs7, [%rd68+34];
ld.global.nc.u16 %rs8, [%rd68+32];
add.s64 %rd70, %rd66, %rd4;
mul.lo.s64 %rd71, %rd70, 36;
add.s64 %rd72, %rd64, %rd71;
add.s64 %rd74, %rd72, %rd73;
ld.global.nc.u32 %r145, [%rd74+4];
ld.global.nc.u32 %r135, [%rd72];

{.reg .f16 low,high;
mov.b32 {low,high},%r135;
cvt.f32.f16 %f26, low;}


ld.global.nc.u32 %r153, [%rd74+40];
ld.global.nc.u32 %r136, [%rd72+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r136;
cvt.f32.f16 %f27, low;}


ld.global.nc.u32 %r161, [%rd74+76];
ld.global.nc.u32 %r137, [%rd72+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r137;
cvt.f32.f16 %f28, low;}


ld.global.nc.u32 %r169, [%rd74+112];
ld.global.nc.u32 %r138, [%rd72+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r138;
cvt.f32.f16 %f29, low;}


mov.b32 %r281, {%rs8, %rs7};
cvt.u32.u64 %r282, %rd4;
shr.s32 %r283, %r280, %r282;
add.s64 %rd75, %rd113, %rd26;
ld.global.nc.u8 %rs9, [%rd75];
cvt.u32.u16 %r284, %rs9;
and.b32 %r285, %r284, 255;
shr.u32 %r286, %r285, %r3;
and.b32 %r287, %r286, 15;
add.s64 %rd76, %rd113, %rd25;
ld.global.nc.u8 %rs10, [%rd76];
cvt.u32.u16 %r288, %rs10;
and.b32 %r289, %r288, 255;
shr.u32 %r290, %r289, %r4;
and.b32 %r291, %r290, 3;
bfi.b32 %r292, %r291, %r287, 4, 2;
add.s32 %r293, %r292, -32;
and.b32 %r213, %r281, 50529027;
shl.b32 %r294, %r283, 2;
and.b32 %r295, %r294, 67372036;
or.b32 %r296, %r213, -2139062144;
sub.s32 %r297, %r296, %r295;
xor.b32 %r298, %r297, -2139062144;
or.b32 %r299, %r213, %r295;
xor.b32 %r300, %r297, %r213;
and.b32 %r211, %r300, %r299;
mov.u32 %r139, %r211;

prmt.b32 %r139, %r139, 0, 0xba98;

mov.u32 %r141, %r213;

prmt.b32 %r141, %r141, 0, 0xba98;

xor.b32 %r301, %r141, 2139062143;
not.b32 %r302, %r139;
and.b32 %r303, %r298, %r302;
and.b32 %r304, %r301, %r139;
or.b32 %r144, %r304, %r303;
mov.u32 %r278, 0;

dp4a.s32.s32 %r143, %r144, %r145, %r278;

mul.lo.s32 %r305, %r293, %r143;
cvt.rn.f32.s32 %f45, %r305;
fma.rn.ftz.f32 %f46, %f26, %f45, 0f00000000;
add.s64 %rd77, %rd113, %rd24;
ld.global.nc.u8 %rs11, [%rd77];
cvt.u32.u16 %r306, %rs11;
and.b32 %r307, %r306, 255;
shr.u32 %r308, %r307, %r5;
and.b32 %r309, %r308, 15;
add.s64 %rd78, %rd113, %rd23;
ld.global.nc.u8 %rs12, [%rd78];
cvt.u32.u16 %r310, %rs12;
and.b32 %r311, %r310, 255;
shr.u32 %r312, %r311, %r6;
and.b32 %r313, %r312, 3;
bfi.b32 %r314, %r313, %r309, 4, 2;
add.s32 %r315, %r314, -32;
shr.u32 %r316, %r281, 2;
and.b32 %r221, %r316, 50529027;
shl.b32 %r317, %r283, 1;
and.b32 %r318, %r317, 67372036;
or.b32 %r319, %r221, -2139062144;
sub.s32 %r320, %r319, %r318;
xor.b32 %r321, %r320, -2139062144;
or.b32 %r322, %r221, %r318;
xor.b32 %r323, %r320, %r221;
and.b32 %r219, %r323, %r322;
mov.u32 %r147, %r219;

prmt.b32 %r147, %r147, 0, 0xba98;

mov.u32 %r149, %r221;

prmt.b32 %r149, %r149, 0, 0xba98;

xor.b32 %r324, %r149, 2139062143;
not.b32 %r325, %r147;
and.b32 %r326, %r321, %r325;
and.b32 %r327, %r324, %r147;
or.b32 %r152, %r327, %r326;

dp4a.s32.s32 %r151, %r152, %r153, %r278;

mul.lo.s32 %r328, %r315, %r151;
cvt.rn.f32.s32 %f47, %r328;
fma.rn.ftz.f32 %f48, %f27, %f47, %f46;
add.s64 %rd79, %rd113, %rd22;
ld.global.nc.u8 %rs13, [%rd79];
cvt.u32.u16 %r329, %rs13;
and.b32 %r330, %r329, 255;
shr.u32 %r331, %r330, %r7;
and.b32 %r332, %r331, 15;
add.s64 %rd80, %rd113, %rd21;
ld.global.nc.u8 %rs14, [%rd80];
cvt.u32.u16 %r333, %rs14;
and.b32 %r334, %r333, 255;
shr.u32 %r335, %r334, %r8;
and.b32 %r336, %r335, 3;
bfi.b32 %r337, %r336, %r332, 4, 2;
add.s32 %r338, %r337, -32;
shr.u32 %r339, %r281, 4;
and.b32 %r229, %r339, 50529027;
and.b32 %r340, %r283, 67372036;
or.b32 %r341, %r229, -2139062144;
sub.s32 %r342, %r341, %r340;
xor.b32 %r343, %r342, -2139062144;
or.b32 %r344, %r229, %r340;
xor.b32 %r345, %r342, %r229;
and.b32 %r227, %r345, %r344;
mov.u32 %r155, %r227;

prmt.b32 %r155, %r155, 0, 0xba98;

mov.u32 %r157, %r229;

prmt.b32 %r157, %r157, 0, 0xba98;

xor.b32 %r346, %r157, 2139062143;
not.b32 %r347, %r155;
and.b32 %r348, %r343, %r347;
and.b32 %r349, %r346, %r155;
or.b32 %r160, %r349, %r348;

dp4a.s32.s32 %r159, %r160, %r161, %r278;

mul.lo.s32 %r350, %r338, %r159;
cvt.rn.f32.s32 %f49, %r350;
fma.rn.ftz.f32 %f50, %f28, %f49, %f48;
add.s64 %rd81, %rd113, %rd20;
ld.global.nc.u8 %rs15, [%rd81];
cvt.u32.u16 %r351, %rs15;
and.b32 %r352, %r351, 255;
shr.u32 %r353, %r352, %r9;
and.b32 %r354, %r353, 15;
add.s64 %rd82, %rd113, %rd19;
ld.global.nc.u8 %rs16, [%rd82];
cvt.u32.u16 %r355, %rs16;
and.b32 %r356, %r355, 255;
shr.u32 %r357, %r356, %r10;
and.b32 %r358, %r357, 3;
bfi.b32 %r359, %r358, %r354, 4, 2;
add.s32 %r360, %r359, -32;
shr.u32 %r361, %r281, 6;
and.b32 %r237, %r361, 50529027;
shr.u32 %r362, %r283, 1;
and.b32 %r363, %r362, 67372036;
or.b32 %r364, %r237, -2139062144;
sub.s32 %r365, %r364, %r363;
xor.b32 %r366, %r365, -2139062144;
or.b32 %r367, %r237, %r363;
xor.b32 %r368, %r365, %r237;
and.b32 %r235, %r368, %r367;
mov.u32 %r163, %r235;

prmt.b32 %r163, %r163, 0, 0xba98;

mov.u32 %r165, %r237;

prmt.b32 %r165, %r165, 0, 0xba98;

xor.b32 %r369, %r165, 2139062143;
not.b32 %r370, %r163;
and.b32 %r371, %r366, %r370;
and.b32 %r372, %r369, %r163;
or.b32 %r168, %r372, %r371;

dp4a.s32.s32 %r167, %r168, %r169, %r278;

mul.lo.s32 %r373, %r360, %r167;
cvt.rn.f32.s32 %f51, %r373;
fma.rn.ftz.f32 %f52, %f29, %f51, %f50;
fma.rn.ftz.f32 %f144, %f25, %f52, %f144;
add.s64 %rd83, %rd113, %rd5;
ld.global.nc.u16 %rs2, [%rd83+108];

{ cvt.f32.f16 %f30, %rs2;}


add.s64 %rd84, %rd113, %rd18;
add.s64 %rd85, %rd113, %rd16;
ld.global.nc.u16 %rs17, [%rd85];
ld.global.nc.u16 %rs18, [%rd85+2];
mov.b32 %r374, {%rs17, %rs18};
not.b32 %r375, %r374;
ld.global.nc.u16 %rs19, [%rd84+2];
ld.global.nc.u16 %rs20, [%rd84];

{.reg .f16 low,high;
mov.b32 {low,high},%r135;
cvt.f32.f16 %f31, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r136;
cvt.f32.f16 %f32, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r137;
cvt.f32.f16 %f33, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r138;
cvt.f32.f16 %f34, low;}


mov.b32 %r376, {%rs20, %rs19};
shr.s32 %r377, %r375, %r282;
add.s64 %rd86, %rd113, %rd13;
ld.global.nc.u8 %rs21, [%rd86];
cvt.u32.u16 %r378, %rs21;
and.b32 %r379, %r378, 255;
shr.u32 %r380, %r379, %r3;
and.b32 %r381, %r380, 15;
add.s64 %rd87, %rd113, %rd12;
ld.global.nc.u8 %rs22, [%rd87];
cvt.u32.u16 %r382, %rs22;
and.b32 %r383, %r382, 255;
shr.u32 %r384, %r383, %r4;
and.b32 %r385, %r384, 3;
bfi.b32 %r386, %r385, %r381, 4, 2;
add.s32 %r387, %r386, -32;
and.b32 %r249, %r376, 50529027;
shl.b32 %r388, %r377, 2;
and.b32 %r389, %r388, 67372036;
or.b32 %r390, %r249, -2139062144;
sub.s32 %r391, %r390, %r389;
xor.b32 %r392, %r391, -2139062144;
or.b32 %r393, %r249, %r389;
xor.b32 %r394, %r391, %r249;
and.b32 %r247, %r394, %r393;
mov.u32 %r175, %r247;

prmt.b32 %r175, %r175, 0, 0xba98;

mov.u32 %r177, %r249;

prmt.b32 %r177, %r177, 0, 0xba98;

xor.b32 %r395, %r177, 2139062143;
not.b32 %r396, %r175;
and.b32 %r397, %r392, %r396;
and.b32 %r398, %r395, %r175;
or.b32 %r180, %r398, %r397;

dp4a.s32.s32 %r179, %r180, %r145, %r278;

mul.lo.s32 %r399, %r387, %r179;
cvt.rn.f32.s32 %f53, %r399;
fma.rn.ftz.f32 %f54, %f31, %f53, 0f00000000;
add.s64 %rd88, %rd113, %rd11;
ld.global.nc.u8 %rs23, [%rd88];
cvt.u32.u16 %r400, %rs23;
and.b32 %r401, %r400, 255;
shr.u32 %r402, %r401, %r5;
and.b32 %r403, %r402, 15;
add.s64 %rd89, %rd113, %rd10;
ld.global.nc.u8 %rs24, [%rd89];
cvt.u32.u16 %r404, %rs24;
and.b32 %r405, %r404, 255;
shr.u32 %r406, %r405, %r6;
and.b32 %r407, %r406, 3;
bfi.b32 %r408, %r407, %r403, 4, 2;
add.s32 %r409, %r408, -32;
shr.u32 %r410, %r376, 2;
and.b32 %r257, %r410, 50529027;
shl.b32 %r411, %r377, 1;
and.b32 %r412, %r411, 67372036;
or.b32 %r413, %r257, -2139062144;
sub.s32 %r414, %r413, %r412;
xor.b32 %r415, %r414, -2139062144;
or.b32 %r416, %r257, %r412;
xor.b32 %r417, %r414, %r257;
and.b32 %r255, %r417, %r416;
mov.u32 %r183, %r255;

prmt.b32 %r183, %r183, 0, 0xba98;

mov.u32 %r185, %r257;

prmt.b32 %r185, %r185, 0, 0xba98;

xor.b32 %r418, %r185, 2139062143;
not.b32 %r419, %r183;
and.b32 %r420, %r415, %r419;
and.b32 %r421, %r418, %r183;
or.b32 %r188, %r421, %r420;

dp4a.s32.s32 %r187, %r188, %r153, %r278;

mul.lo.s32 %r422, %r409, %r187;
cvt.rn.f32.s32 %f55, %r422;
fma.rn.ftz.f32 %f56, %f32, %f55, %f54;
add.s64 %rd90, %rd113, %rd9;
ld.global.nc.u8 %rs25, [%rd90];
cvt.u32.u16 %r423, %rs25;
and.b32 %r424, %r423, 255;
shr.u32 %r425, %r424, %r7;
and.b32 %r426, %r425, 15;
add.s64 %rd91, %rd113, %rd8;
ld.global.nc.u8 %rs26, [%rd91];
cvt.u32.u16 %r427, %rs26;
and.b32 %r428, %r427, 255;
shr.u32 %r429, %r428, %r8;
and.b32 %r430, %r429, 3;
bfi.b32 %r431, %r430, %r426, 4, 2;
add.s32 %r432, %r431, -32;
shr.u32 %r433, %r376, 4;
and.b32 %r265, %r433, 50529027;
and.b32 %r434, %r377, 67372036;
or.b32 %r435, %r265, -2139062144;
sub.s32 %r436, %r435, %r434;
xor.b32 %r437, %r436, -2139062144;
or.b32 %r438, %r265, %r434;
xor.b32 %r439, %r436, %r265;
and.b32 %r263, %r439, %r438;
mov.u32 %r191, %r263;

prmt.b32 %r191, %r191, 0, 0xba98;

mov.u32 %r193, %r265;

prmt.b32 %r193, %r193, 0, 0xba98;

xor.b32 %r440, %r193, 2139062143;
not.b32 %r441, %r191;
and.b32 %r442, %r437, %r441;
and.b32 %r443, %r440, %r191;
or.b32 %r196, %r443, %r442;

dp4a.s32.s32 %r195, %r196, %r161, %r278;

mul.lo.s32 %r444, %r432, %r195;
cvt.rn.f32.s32 %f57, %r444;
fma.rn.ftz.f32 %f58, %f33, %f57, %f56;
add.s64 %rd92, %rd113, %rd7;
ld.global.nc.u8 %rs27, [%rd92];
cvt.u32.u16 %r445, %rs27;
and.b32 %r446, %r445, 255;
shr.u32 %r447, %r446, %r9;
and.b32 %r448, %r447, 15;
add.s64 %rd93, %rd113, %rd6;
ld.global.nc.u8 %rs28, [%rd93];
cvt.u32.u16 %r449, %rs28;
and.b32 %r450, %r449, 255;
shr.u32 %r451, %r450, %r10;
and.b32 %r452, %r451, 3;
bfi.b32 %r453, %r452, %r448, 4, 2;
add.s32 %r454, %r453, -32;
shr.u32 %r455, %r376, 6;
and.b32 %r273, %r455, 50529027;
shr.u32 %r456, %r377, 1;
and.b32 %r457, %r456, 67372036;
or.b32 %r458, %r273, -2139062144;
sub.s32 %r459, %r458, %r457;
xor.b32 %r460, %r459, -2139062144;
or.b32 %r461, %r273, %r457;
xor.b32 %r462, %r459, %r273;
and.b32 %r271, %r462, %r461;
mov.u32 %r199, %r271;

prmt.b32 %r199, %r199, 0, 0xba98;

mov.u32 %r201, %r273;

prmt.b32 %r201, %r201, 0, 0xba98;

xor.b32 %r463, %r201, 2139062143;
not.b32 %r464, %r199;
and.b32 %r465, %r460, %r464;
and.b32 %r466, %r463, %r199;
or.b32 %r204, %r466, %r465;

dp4a.s32.s32 %r203, %r204, %r169, %r278;

mul.lo.s32 %r467, %r454, %r203;
cvt.rn.f32.s32 %f59, %r467;
fma.rn.ftz.f32 %f60, %f34, %f59, %f58;
fma.rn.ftz.f32 %f143, %f30, %f60, %f143;
cvt.u64.u32 %rd94, %r613;
add.s64 %rd95, %rd1, %rd94;

{ cvt.f32.f16 %f35, %rs1;}


add.s64 %rd96, %rd95, %rd4;
mul.lo.s64 %rd97, %rd96, 36;
add.s64 %rd98, %rd64, %rd97;
add.s64 %rd99, %rd98, %rd73;
ld.global.nc.u32 %r217, [%rd99+4];
ld.global.nc.u32 %r207, [%rd98];

{.reg .f16 low,high;
mov.b32 {low,high},%r207;
cvt.f32.f16 %f36, low;}


ld.global.nc.u32 %r225, [%rd99+40];
ld.global.nc.u32 %r208, [%rd98+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r208;
cvt.f32.f16 %f37, low;}


ld.global.nc.u32 %r233, [%rd99+76];
ld.global.nc.u32 %r209, [%rd98+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r209;
cvt.f32.f16 %f38, low;}


ld.global.nc.u32 %r241, [%rd99+112];
ld.global.nc.u32 %r210, [%rd98+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r210;
cvt.f32.f16 %f39, low;}



prmt.b32 %r211, %r211, 0, 0xba98;


prmt.b32 %r213, %r213, 0, 0xba98;

xor.b32 %r468, %r213, 2139062143;
not.b32 %r469, %r211;
and.b32 %r470, %r298, %r469;
and.b32 %r471, %r468, %r211;
or.b32 %r216, %r471, %r470;

dp4a.s32.s32 %r215, %r216, %r217, %r278;

mul.lo.s32 %r472, %r293, %r215;
cvt.rn.f32.s32 %f61, %r472;
fma.rn.ftz.f32 %f62, %f36, %f61, 0f00000000;

prmt.b32 %r219, %r219, 0, 0xba98;


prmt.b32 %r221, %r221, 0, 0xba98;

xor.b32 %r473, %r221, 2139062143;
not.b32 %r474, %r219;
and.b32 %r475, %r321, %r474;
and.b32 %r476, %r473, %r219;
or.b32 %r224, %r476, %r475;

dp4a.s32.s32 %r223, %r224, %r225, %r278;

mul.lo.s32 %r477, %r315, %r223;
cvt.rn.f32.s32 %f63, %r477;
fma.rn.ftz.f32 %f64, %f37, %f63, %f62;

prmt.b32 %r227, %r227, 0, 0xba98;


prmt.b32 %r229, %r229, 0, 0xba98;

xor.b32 %r478, %r229, 2139062143;
not.b32 %r479, %r227;
and.b32 %r480, %r343, %r479;
and.b32 %r481, %r478, %r227;
or.b32 %r232, %r481, %r480;

dp4a.s32.s32 %r231, %r232, %r233, %r278;

mul.lo.s32 %r482, %r338, %r231;
cvt.rn.f32.s32 %f65, %r482;
fma.rn.ftz.f32 %f66, %f38, %f65, %f64;

prmt.b32 %r235, %r235, 0, 0xba98;


prmt.b32 %r237, %r237, 0, 0xba98;

xor.b32 %r483, %r237, 2139062143;
not.b32 %r484, %r235;
and.b32 %r485, %r366, %r484;
and.b32 %r486, %r483, %r235;
or.b32 %r240, %r486, %r485;

dp4a.s32.s32 %r239, %r240, %r241, %r278;

mul.lo.s32 %r487, %r360, %r239;
cvt.rn.f32.s32 %f67, %r487;
fma.rn.ftz.f32 %f68, %f39, %f67, %f66;
fma.rn.ftz.f32 %f142, %f35, %f68, %f142;

{ cvt.f32.f16 %f40, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r207;
cvt.f32.f16 %f41, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r208;
cvt.f32.f16 %f42, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r209;
cvt.f32.f16 %f43, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r210;
cvt.f32.f16 %f44, low;}



prmt.b32 %r247, %r247, 0, 0xba98;


prmt.b32 %r249, %r249, 0, 0xba98;

xor.b32 %r488, %r249, 2139062143;
not.b32 %r489, %r247;
and.b32 %r490, %r392, %r489;
and.b32 %r491, %r488, %r247;
or.b32 %r252, %r491, %r490;

dp4a.s32.s32 %r251, %r252, %r217, %r278;

mul.lo.s32 %r492, %r387, %r251;
cvt.rn.f32.s32 %f69, %r492;
fma.rn.ftz.f32 %f70, %f41, %f69, 0f00000000;

prmt.b32 %r255, %r255, 0, 0xba98;


prmt.b32 %r257, %r257, 0, 0xba98;

xor.b32 %r493, %r257, 2139062143;
not.b32 %r494, %r255;
and.b32 %r495, %r415, %r494;
and.b32 %r496, %r493, %r255;
or.b32 %r260, %r496, %r495;

dp4a.s32.s32 %r259, %r260, %r225, %r278;

mul.lo.s32 %r497, %r409, %r259;
cvt.rn.f32.s32 %f71, %r497;
fma.rn.ftz.f32 %f72, %f42, %f71, %f70;

prmt.b32 %r263, %r263, 0, 0xba98;


prmt.b32 %r265, %r265, 0, 0xba98;

xor.b32 %r498, %r265, 2139062143;
not.b32 %r499, %r263;
and.b32 %r500, %r437, %r499;
and.b32 %r501, %r498, %r263;
or.b32 %r268, %r501, %r500;

dp4a.s32.s32 %r267, %r268, %r233, %r278;

mul.lo.s32 %r502, %r432, %r267;
cvt.rn.f32.s32 %f73, %r502;
fma.rn.ftz.f32 %f74, %f43, %f73, %f72;

prmt.b32 %r271, %r271, 0, 0xba98;


prmt.b32 %r273, %r273, 0, 0xba98;

xor.b32 %r503, %r273, 2139062143;
not.b32 %r504, %r271;
and.b32 %r505, %r460, %r504;
and.b32 %r506, %r503, %r271;
or.b32 %r276, %r506, %r505;

dp4a.s32.s32 %r275, %r276, %r241, %r278;

mul.lo.s32 %r507, %r454, %r275;
cvt.rn.f32.s32 %f75, %r507;
fma.rn.ftz.f32 %f76, %f44, %f75, %f74;
fma.rn.ftz.f32 %f141, %f40, %f76, %f141;
add.s32 %r613, %r613, 64;
add.s32 %r612, %r612, 64;
add.s64 %rd113, %rd113, 880;
add.s32 %r614, %r614, 8;
setp.lt.s32 %p4, %r614, %r43;
@%p4 bra $L__BB146_2;

$L__BB146_3:
mov.u32 %r600, %tid.y;
setp.eq.s32 %p5, %r600, 0;
@%p5 bra $L__BB146_5;

mov.u32 %r602, %tid.x;
mov.u32 %r601, %tid.y;
shl.b32 %r512, %r601, 9;
mov.u32 %r513, _ZZ13mul_mat_vec_qIL9ggml_type11ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r514, %r513, %r512;
shl.b32 %r515, %r602, 2;
add.s32 %r516, %r514, %r515;
st.shared.f32 [%r516+-512], %f144;
st.shared.f32 [%r516+-384], %f143;
st.shared.f32 [%r516+-256], %f142;
st.shared.f32 [%r516+-128], %f141;

$L__BB146_5:
mov.u32 %r603, %tid.y;
setp.ne.s32 %p6, %r603, 0;
bar.sync 0;
@%p6 bra $L__BB146_14;

ld.param.u32 %r608, [_Z13mul_mat_vec_qIL9ggml_type11ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u64 %rd112, [_Z13mul_mat_vec_qIL9ggml_type11ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r606, [_Z13mul_mat_vec_qIL9ggml_type11ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r605, [_Z13mul_mat_vec_qIL9ggml_type11ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
mov.u32 %r604, %tid.x;
mov.u32 %r518, %ctaid.z;
mov.u32 %r519, %ctaid.y;
mov.u32 %r520, %ctaid.x;
shl.b32 %r521, %r520, 1;
mov.u32 %r522, 1;
mad.lo.s32 %r523, %r519, %r605, %r521;
mad.lo.s32 %r524, %r518, %r606, %r523;
cvt.u64.u32 %rd29, %r524;
shl.b32 %r526, %r604, 2;
mov.u32 %r527, _ZZ13mul_mat_vec_qIL9ggml_type11ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r19, %r527, %r526;
mov.u32 %r528, 2;
ld.shared.f32 %f77, [%r19];
add.ftz.f32 %f78, %f77, %f144;
ld.shared.f32 %f79, [%r19+512];
add.ftz.f32 %f80, %f79, %f78;
ld.shared.f32 %f81, [%r19+1024];
add.ftz.f32 %f82, %f81, %f80;
mov.b32 %r529, %f82;
mov.u32 %r530, 31;
mov.u32 %r531, 16;
mov.u32 %r532, -1;
shfl.sync.bfly.b32 %r533|%p7, %r529, %r531, %r530, %r532;
mov.b32 %f83, %r533;
add.ftz.f32 %f84, %f82, %f83;
mov.b32 %r534, %f84;
mov.u32 %r535, 8;
shfl.sync.bfly.b32 %r536|%p8, %r534, %r535, %r530, %r532;
mov.b32 %f85, %r536;
add.ftz.f32 %f86, %f84, %f85;
mov.b32 %r537, %f86;
mov.u32 %r538, 4;
shfl.sync.bfly.b32 %r539|%p9, %r537, %r538, %r530, %r532;
mov.b32 %f87, %r539;
add.ftz.f32 %f88, %f86, %f87;
mov.b32 %r540, %f88;
shfl.sync.bfly.b32 %r541|%p10, %r540, %r528, %r530, %r532;
mov.b32 %f89, %r541;
add.ftz.f32 %f90, %f88, %f89;
mov.b32 %r542, %f90;
shfl.sync.bfly.b32 %r543|%p11, %r542, %r522, %r530, %r532;
mov.b32 %f91, %r543;
add.ftz.f32 %f13, %f90, %f91;
setp.eq.s32 %p12, %r604, 0;
setp.lt.u32 %p13, %r521, %r608;
and.pred %p1, %p12, %p13;
cvta.to.global.u64 %rd100, %rd112;
mul.wide.u32 %rd101, %r524, 4;
add.s64 %rd30, %rd100, %rd101;
not.pred %p14, %p1;
@%p14 bra $L__BB146_8;

st.global.f32 [%rd30], %f13;

$L__BB146_8:
ld.param.u32 %r609, [_Z13mul_mat_vec_qIL9ggml_type11ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
mov.u32 %r607, %tid.x;
ld.shared.f32 %f92, [%r19+128];
add.ftz.f32 %f93, %f92, %f143;
ld.shared.f32 %f94, [%r19+640];
add.ftz.f32 %f95, %f94, %f93;
ld.shared.f32 %f96, [%r19+1152];
add.ftz.f32 %f97, %f96, %f95;
mov.b32 %r544, %f97;
shfl.sync.bfly.b32 %r548|%p15, %r544, %r531, %r530, %r532;
mov.b32 %f98, %r548;
add.ftz.f32 %f99, %f97, %f98;
mov.b32 %r549, %f99;
shfl.sync.bfly.b32 %r551|%p16, %r549, %r535, %r530, %r532;
mov.b32 %f100, %r551;
add.ftz.f32 %f101, %f99, %f100;
mov.b32 %r552, %f101;
shfl.sync.bfly.b32 %r554|%p17, %r552, %r538, %r530, %r532;
mov.b32 %f102, %r554;
add.ftz.f32 %f103, %f101, %f102;
mov.b32 %r555, %f103;
shfl.sync.bfly.b32 %r557|%p18, %r555, %r528, %r530, %r532;
mov.b32 %f104, %r557;
add.ftz.f32 %f105, %f103, %f104;
mov.b32 %r558, %f105;
shfl.sync.bfly.b32 %r560|%p19, %r558, %r522, %r530, %r532;
mov.b32 %f106, %r560;
add.ftz.f32 %f14, %f105, %f106;
setp.eq.s32 %p20, %r607, 1;
or.b32 %r564, %r521, 1;
setp.lt.u32 %p21, %r564, %r609;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB146_10;

st.global.f32 [%rd30+4], %f14;

$L__BB146_10:
ld.shared.f32 %f107, [%r19+256];
add.ftz.f32 %f108, %f107, %f142;
ld.shared.f32 %f109, [%r19+768];
add.ftz.f32 %f110, %f109, %f108;
ld.shared.f32 %f111, [%r19+1280];
add.ftz.f32 %f112, %f111, %f110;
mov.b32 %r565, %f112;
mov.u32 %r566, 31;
mov.u32 %r567, 16;
mov.u32 %r568, -1;
shfl.sync.bfly.b32 %r569|%p23, %r565, %r567, %r566, %r568;
mov.b32 %f113, %r569;
add.ftz.f32 %f114, %f112, %f113;
mov.b32 %r570, %f114;
mov.u32 %r571, 8;
shfl.sync.bfly.b32 %r572|%p24, %r570, %r571, %r566, %r568;
mov.b32 %f115, %r572;
add.ftz.f32 %f116, %f114, %f115;
mov.b32 %r573, %f116;
mov.u32 %r574, 4;
shfl.sync.bfly.b32 %r575|%p25, %r573, %r574, %r566, %r568;
mov.b32 %f117, %r575;
add.ftz.f32 %f118, %f116, %f117;
mov.b32 %r576, %f118;
mov.u32 %r577, 2;
shfl.sync.bfly.b32 %r578|%p26, %r576, %r577, %r566, %r568;
mov.b32 %f119, %r578;
add.ftz.f32 %f120, %f118, %f119;
mov.b32 %r579, %f120;
mov.u32 %r580, 1;
shfl.sync.bfly.b32 %r581|%p27, %r579, %r580, %r566, %r568;
mov.b32 %f121, %r581;
add.ftz.f32 %f15, %f120, %f121;
@%p14 bra $L__BB146_12;

ld.param.u32 %r611, [_Z13mul_mat_vec_qIL9ggml_type11ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
cvt.u64.u32 %rd102, %r611;
add.s64 %rd103, %rd102, %rd29;
shl.b64 %rd105, %rd103, 2;
add.s64 %rd106, %rd100, %rd105;
st.global.f32 [%rd106], %f15;

$L__BB146_12:
ld.shared.f32 %f122, [%r19+384];
add.ftz.f32 %f123, %f122, %f141;
ld.shared.f32 %f124, [%r19+896];
add.ftz.f32 %f125, %f124, %f123;
ld.shared.f32 %f126, [%r19+1408];
add.ftz.f32 %f127, %f126, %f125;
mov.b32 %r582, %f127;
shfl.sync.bfly.b32 %r586|%p29, %r582, %r567, %r566, %r568;
mov.b32 %f128, %r586;
add.ftz.f32 %f129, %f127, %f128;
mov.b32 %r587, %f129;
shfl.sync.bfly.b32 %r589|%p30, %r587, %r571, %r566, %r568;
mov.b32 %f130, %r589;
add.ftz.f32 %f131, %f129, %f130;
mov.b32 %r590, %f131;
shfl.sync.bfly.b32 %r592|%p31, %r590, %r574, %r566, %r568;
mov.b32 %f132, %r592;
add.ftz.f32 %f133, %f131, %f132;
mov.b32 %r593, %f133;
shfl.sync.bfly.b32 %r595|%p32, %r593, %r577, %r566, %r568;
mov.b32 %f134, %r595;
add.ftz.f32 %f135, %f133, %f134;
mov.b32 %r596, %f135;
shfl.sync.bfly.b32 %r598|%p33, %r596, %r580, %r566, %r568;
mov.b32 %f136, %r598;
add.ftz.f32 %f16, %f135, %f136;
@%p22 bra $L__BB146_14;

ld.param.u32 %r610, [_Z13mul_mat_vec_qIL9ggml_type11ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
add.s32 %r599, %r610, 1;
cvt.u64.u32 %rd107, %r599;
add.s64 %rd108, %rd107, %rd29;
shl.b64 %rd110, %rd108, 2;
add.s64 %rd111, %rd100, %rd110;
st.global.f32 [%rd111], %f16;

$L__BB146_14:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type11ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type11ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type11ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type11ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type11ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type11ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type11ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type11ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type11ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<47>;
.reg .b16 %rs<31>;
.reg .f32 %f<217>;
.reg .b32 %r<767>;
.reg .b64 %rd<125>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type11ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[2304];

ld.param.u64 %rd32, [_Z13mul_mat_vec_qIL9ggml_type11ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd33, [_Z13mul_mat_vec_qIL9ggml_type11ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type11ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type11ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type11ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type11ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type11ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type11ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type11ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type11ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type11ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type11ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type11ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r40, %tid.y;
shl.b32 %r41, %r40, 5;
mov.u32 %r42, %tid.x;
add.s32 %r1, %r41, %r42;
shr.s32 %r43, %r1, 31;
shr.u32 %r44, %r43, 28;
add.s32 %r45, %r1, %r44;
shr.s32 %r766, %r45, 4;
shr.u32 %r46, %r24, 8;
setp.ge.s32 %p3, %r766, %r46;
mov.f32 %f211, 0f00000000;
mov.f32 %f212, %f211;
mov.f32 %f213, %f211;
mov.f32 %f214, %f211;
mov.f32 %f215, %f211;
mov.f32 %f216, %f211;
@%p3 bra $L__BB147_3;

mov.u32 %r47, %ctaid.x;
shl.b32 %r48, %r47, 1;
mov.u32 %r49, %ctaid.y;
mov.u32 %r50, %ctaid.z;
mul.lo.s32 %r51, %r49, %r32;
cvt.u64.u32 %rd35, %r51;
mul.lo.s32 %r52, %r50, %r38;
cvt.u64.u32 %rd36, %r52;
add.s64 %rd1, %rd36, %rd35;
mul.hi.u32 %r53, %r50, %r34;
add.s32 %r54, %r53, %r50;
shr.u32 %r55, %r54, %r35;
mul.lo.s32 %r56, %r55, %r37;
mul.hi.u32 %r57, %r49, %r28;
add.s32 %r58, %r57, %r49;
shr.u32 %r59, %r58, %r29;
mul.lo.s32 %r60, %r59, %r31;
cvta.to.global.u64 %rd124, %rd32;
and.b32 %r64, %r45, -16;
sub.s32 %r65, %r1, %r64;
shr.s32 %r66, %r65, 31;
shr.u32 %r67, %r66, 29;
add.s32 %r68, %r65, %r67;
shr.s32 %r69, %r68, 3;
shl.b32 %r70, %r69, 2;
and.b32 %r71, %r68, -8;
sub.s32 %r72, %r65, %r71;
shr.s32 %r73, %r72, 31;
shr.u32 %r74, %r73, 30;
add.s32 %r75, %r72, %r74;
shr.s32 %r76, %r75, 2;
add.s32 %r77, %r71, %r76;
shl.b32 %r78, %r65, 1;
shl.b32 %r79, %r72, 1;
cvt.s64.s32 %rd3, %r72;
or.b32 %r80, %r48, 1;
shr.s32 %r81, %r77, 31;
shr.u32 %r82, %r81, 29;
add.s32 %r83, %r77, %r82;
and.b32 %r84, %r83, -8;
sub.s32 %r85, %r77, %r84;
cvt.s64.s32 %rd37, %r85;
shr.s32 %r86, %r83, 3;
shl.b32 %r3, %r86, 2;
cvt.s64.s32 %rd4, %r70;
shr.u32 %r87, %r81, 30;
add.s32 %r88, %r77, %r87;
and.b32 %r89, %r88, -4;
sub.s32 %r90, %r77, %r89;
shr.s32 %r91, %r88, 2;
shl.b32 %r4, %r91, 1;
add.s32 %r92, %r90, 8;
cvt.s64.s32 %rd38, %r92;
add.s32 %r93, %r77, 2;
shr.s32 %r94, %r93, 31;
shr.u32 %r95, %r94, 29;
add.s32 %r96, %r93, %r95;
and.b32 %r97, %r96, -8;
sub.s32 %r98, %r93, %r97;
shr.s32 %r99, %r96, 3;
shl.b32 %r5, %r99, 2;
cvt.s64.s32 %rd39, %r98;
shr.u32 %r100, %r94, 30;
add.s32 %r101, %r93, %r100;
and.b32 %r102, %r101, -4;
sub.s32 %r103, %r93, %r102;
shr.s32 %r104, %r101, 2;
shl.b32 %r6, %r104, 1;
add.s32 %r105, %r103, 8;
cvt.s64.s32 %rd40, %r105;
add.s32 %r106, %r77, 4;
shr.s32 %r107, %r106, 31;
shr.u32 %r108, %r107, 29;
add.s32 %r109, %r106, %r108;
and.b32 %r110, %r109, -8;
sub.s32 %r111, %r106, %r110;
shr.s32 %r112, %r109, 3;
shl.b32 %r7, %r112, 2;
cvt.s64.s32 %rd41, %r111;
shr.u32 %r113, %r107, 30;
add.s32 %r114, %r106, %r113;
and.b32 %r115, %r114, -4;
sub.s32 %r116, %r106, %r115;
shr.s32 %r117, %r114, 2;
shl.b32 %r8, %r117, 1;
add.s32 %r118, %r116, 8;
cvt.s64.s32 %rd42, %r118;
add.s32 %r119, %r77, 6;
shr.s32 %r120, %r119, 31;
shr.u32 %r121, %r120, 29;
add.s32 %r122, %r119, %r121;
and.b32 %r123, %r122, -8;
sub.s32 %r124, %r119, %r123;
shr.s32 %r125, %r122, 3;
shl.b32 %r9, %r125, 2;
cvt.s64.s32 %rd43, %r124;
shr.u32 %r126, %r120, 30;
add.s32 %r127, %r119, %r126;
and.b32 %r128, %r127, -4;
sub.s32 %r129, %r119, %r128;
shr.s32 %r130, %r127, 2;
shl.b32 %r10, %r130, 1;
add.s32 %r131, %r129, 8;
cvt.s64.s32 %rd44, %r131;
shl.b32 %r132, %r26, 1;
shl.b32 %r763, %r766, 3;
add.s32 %r765, %r763, %r132;
add.s32 %r764, %r26, %r763;
add.s32 %r133, %r766, %r56;
add.s32 %r134, %r133, %r60;
mad.lo.s32 %r135, %r80, %r25, %r134;
mul.wide.s32 %rd5, %r135, 110;
add.s64 %rd45, %rd5, %rd44;
add.s64 %rd6, %rd45, 96;
add.s64 %rd46, %rd5, %rd43;
add.s64 %rd7, %rd46, 96;
add.s64 %rd47, %rd5, %rd42;
add.s64 %rd8, %rd47, 96;
add.s64 %rd48, %rd5, %rd41;
add.s64 %rd9, %rd48, 96;
add.s64 %rd49, %rd5, %rd40;
add.s64 %rd10, %rd49, 96;
add.s64 %rd50, %rd5, %rd39;
add.s64 %rd11, %rd50, 96;
add.s64 %rd51, %rd5, %rd38;
add.s64 %rd12, %rd51, 96;
add.s64 %rd52, %rd5, %rd37;
add.s64 %rd13, %rd52, 96;
mad.lo.s32 %r136, %r48, %r25, %r766;
add.s32 %r137, %r136, %r56;
add.s32 %r138, %r137, %r60;
mul.wide.s32 %rd14, %r138, 110;
mul.wide.s32 %rd53, %r78, 2;
add.s64 %rd15, %rd14, %rd53;
mul.wide.s32 %rd54, %r79, 2;
add.s64 %rd16, %rd5, %rd54;
add.s64 %rd55, %rd14, %rd54;
add.s64 %rd17, %rd55, 2;
add.s64 %rd56, %rd5, %rd53;
add.s64 %rd18, %rd56, 32;
add.s64 %rd57, %rd14, %rd44;
add.s64 %rd19, %rd57, 96;
add.s64 %rd58, %rd14, %rd43;
add.s64 %rd20, %rd58, 96;
add.s64 %rd59, %rd14, %rd42;
add.s64 %rd21, %rd59, 96;
add.s64 %rd60, %rd14, %rd41;
add.s64 %rd22, %rd60, 96;
add.s64 %rd61, %rd14, %rd40;
add.s64 %rd23, %rd61, 96;
add.s64 %rd62, %rd14, %rd39;
add.s64 %rd24, %rd62, 96;
add.s64 %rd63, %rd14, %rd38;
add.s64 %rd25, %rd63, 96;
add.s64 %rd64, %rd14, %rd37;
add.s64 %rd26, %rd64, 96;
mov.f32 %f211, 0f00000000;
cvta.to.global.u64 %rd65, %rd33;
shl.b64 %rd74, %rd3, 2;

$L__BB147_2:
cvt.u64.u32 %rd66, %r763;
add.s64 %rd67, %rd1, %rd66;
add.s64 %rd68, %rd124, %rd14;
ld.global.nc.u16 %rs1, [%rd68+108];

{ cvt.f32.f16 %f37, %rs1;}


add.s64 %rd69, %rd124, %rd15;
add.s64 %rd70, %rd124, %rd17;
ld.global.nc.u16 %rs7, [%rd70+-2];
ld.global.nc.u16 %rs8, [%rd70];
mov.b32 %r355, {%rs7, %rs8};
not.b32 %r356, %r355;
ld.global.nc.u16 %rs9, [%rd69+34];
ld.global.nc.u16 %rs10, [%rd69+32];
add.s64 %rd71, %rd67, %rd4;
mul.lo.s64 %rd72, %rd71, 36;
add.s64 %rd73, %rd65, %rd72;
add.s64 %rd75, %rd73, %rd74;
ld.global.nc.u32 %r149, [%rd75+4];
ld.global.nc.u32 %r139, [%rd73];

{.reg .f16 low,high;
mov.b32 {low,high},%r139;
cvt.f32.f16 %f38, low;}


ld.global.nc.u32 %r157, [%rd75+40];
ld.global.nc.u32 %r140, [%rd73+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r140;
cvt.f32.f16 %f39, low;}


ld.global.nc.u32 %r165, [%rd75+76];
ld.global.nc.u32 %r141, [%rd73+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r141;
cvt.f32.f16 %f40, low;}


ld.global.nc.u32 %r173, [%rd75+112];
ld.global.nc.u32 %r142, [%rd73+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r142;
cvt.f32.f16 %f41, low;}


mov.b32 %r357, {%rs10, %rs9};
cvt.u32.u64 %r358, %rd4;
shr.s32 %r359, %r356, %r358;
add.s64 %rd76, %rd124, %rd26;
ld.global.nc.u8 %rs11, [%rd76];
cvt.u32.u16 %r360, %rs11;
and.b32 %r361, %r360, 255;
shr.u32 %r362, %r361, %r3;
and.b32 %r363, %r362, 15;
add.s64 %rd77, %rd124, %rd25;
ld.global.nc.u8 %rs12, [%rd77];
cvt.u32.u16 %r364, %rs12;
and.b32 %r365, %r364, 255;
shr.u32 %r366, %r365, %r4;
and.b32 %r367, %r366, 3;
bfi.b32 %r368, %r367, %r363, 4, 2;
add.s32 %r369, %r368, -32;
and.b32 %r289, %r357, 50529027;
shl.b32 %r370, %r359, 2;
and.b32 %r371, %r370, 67372036;
or.b32 %r372, %r289, -2139062144;
sub.s32 %r373, %r372, %r371;
xor.b32 %r374, %r373, -2139062144;
or.b32 %r375, %r289, %r371;
xor.b32 %r376, %r373, %r289;
and.b32 %r287, %r376, %r375;
mov.u32 %r143, %r287;

prmt.b32 %r143, %r143, 0, 0xba98;

mov.u32 %r145, %r289;

prmt.b32 %r145, %r145, 0, 0xba98;

xor.b32 %r377, %r145, 2139062143;
not.b32 %r378, %r143;
and.b32 %r379, %r374, %r378;
and.b32 %r380, %r377, %r143;
or.b32 %r148, %r380, %r379;
mov.u32 %r354, 0;

dp4a.s32.s32 %r147, %r148, %r149, %r354;

mul.lo.s32 %r381, %r369, %r147;
cvt.rn.f32.s32 %f67, %r381;
fma.rn.ftz.f32 %f68, %f38, %f67, 0f00000000;
add.s64 %rd78, %rd124, %rd24;
ld.global.nc.u8 %rs13, [%rd78];
cvt.u32.u16 %r382, %rs13;
and.b32 %r383, %r382, 255;
shr.u32 %r384, %r383, %r5;
and.b32 %r385, %r384, 15;
add.s64 %rd79, %rd124, %rd23;
ld.global.nc.u8 %rs14, [%rd79];
cvt.u32.u16 %r386, %rs14;
and.b32 %r387, %r386, 255;
shr.u32 %r388, %r387, %r6;
and.b32 %r389, %r388, 3;
bfi.b32 %r390, %r389, %r385, 4, 2;
add.s32 %r391, %r390, -32;
shr.u32 %r392, %r357, 2;
and.b32 %r297, %r392, 50529027;
shl.b32 %r393, %r359, 1;
and.b32 %r394, %r393, 67372036;
or.b32 %r395, %r297, -2139062144;
sub.s32 %r396, %r395, %r394;
xor.b32 %r397, %r396, -2139062144;
or.b32 %r398, %r297, %r394;
xor.b32 %r399, %r396, %r297;
and.b32 %r295, %r399, %r398;
mov.u32 %r151, %r295;

prmt.b32 %r151, %r151, 0, 0xba98;

mov.u32 %r153, %r297;

prmt.b32 %r153, %r153, 0, 0xba98;

xor.b32 %r400, %r153, 2139062143;
not.b32 %r401, %r151;
and.b32 %r402, %r397, %r401;
and.b32 %r403, %r400, %r151;
or.b32 %r156, %r403, %r402;

dp4a.s32.s32 %r155, %r156, %r157, %r354;

mul.lo.s32 %r404, %r391, %r155;
cvt.rn.f32.s32 %f69, %r404;
fma.rn.ftz.f32 %f70, %f39, %f69, %f68;
add.s64 %rd80, %rd124, %rd22;
ld.global.nc.u8 %rs15, [%rd80];
cvt.u32.u16 %r405, %rs15;
and.b32 %r406, %r405, 255;
shr.u32 %r407, %r406, %r7;
and.b32 %r408, %r407, 15;
add.s64 %rd81, %rd124, %rd21;
ld.global.nc.u8 %rs16, [%rd81];
cvt.u32.u16 %r409, %rs16;
and.b32 %r410, %r409, 255;
shr.u32 %r411, %r410, %r8;
and.b32 %r412, %r411, 3;
bfi.b32 %r413, %r412, %r408, 4, 2;
add.s32 %r414, %r413, -32;
shr.u32 %r415, %r357, 4;
and.b32 %r305, %r415, 50529027;
and.b32 %r416, %r359, 67372036;
or.b32 %r417, %r305, -2139062144;
sub.s32 %r418, %r417, %r416;
xor.b32 %r419, %r418, -2139062144;
or.b32 %r420, %r305, %r416;
xor.b32 %r421, %r418, %r305;
and.b32 %r303, %r421, %r420;
mov.u32 %r159, %r303;

prmt.b32 %r159, %r159, 0, 0xba98;

mov.u32 %r161, %r305;

prmt.b32 %r161, %r161, 0, 0xba98;

xor.b32 %r422, %r161, 2139062143;
not.b32 %r423, %r159;
and.b32 %r424, %r419, %r423;
and.b32 %r425, %r422, %r159;
or.b32 %r164, %r425, %r424;

dp4a.s32.s32 %r163, %r164, %r165, %r354;

mul.lo.s32 %r426, %r414, %r163;
cvt.rn.f32.s32 %f71, %r426;
fma.rn.ftz.f32 %f72, %f40, %f71, %f70;
add.s64 %rd82, %rd124, %rd20;
ld.global.nc.u8 %rs17, [%rd82];
cvt.u32.u16 %r427, %rs17;
and.b32 %r428, %r427, 255;
shr.u32 %r429, %r428, %r9;
and.b32 %r430, %r429, 15;
add.s64 %rd83, %rd124, %rd19;
ld.global.nc.u8 %rs18, [%rd83];
cvt.u32.u16 %r431, %rs18;
and.b32 %r432, %r431, 255;
shr.u32 %r433, %r432, %r10;
and.b32 %r434, %r433, 3;
bfi.b32 %r435, %r434, %r430, 4, 2;
add.s32 %r436, %r435, -32;
shr.u32 %r437, %r357, 6;
and.b32 %r313, %r437, 50529027;
shr.u32 %r438, %r359, 1;
and.b32 %r439, %r438, 67372036;
or.b32 %r440, %r313, -2139062144;
sub.s32 %r441, %r440, %r439;
xor.b32 %r442, %r441, -2139062144;
or.b32 %r443, %r313, %r439;
xor.b32 %r444, %r441, %r313;
and.b32 %r311, %r444, %r443;
mov.u32 %r167, %r311;

prmt.b32 %r167, %r167, 0, 0xba98;

mov.u32 %r169, %r313;

prmt.b32 %r169, %r169, 0, 0xba98;

xor.b32 %r445, %r169, 2139062143;
not.b32 %r446, %r167;
and.b32 %r447, %r442, %r446;
and.b32 %r448, %r445, %r167;
or.b32 %r172, %r448, %r447;

dp4a.s32.s32 %r171, %r172, %r173, %r354;

mul.lo.s32 %r449, %r436, %r171;
cvt.rn.f32.s32 %f73, %r449;
fma.rn.ftz.f32 %f74, %f41, %f73, %f72;
fma.rn.ftz.f32 %f216, %f37, %f74, %f216;
add.s64 %rd84, %rd124, %rd5;
ld.global.nc.u16 %rs2, [%rd84+108];

{ cvt.f32.f16 %f42, %rs2;}


add.s64 %rd85, %rd124, %rd18;
add.s64 %rd86, %rd124, %rd16;
ld.global.nc.u16 %rs19, [%rd86];
ld.global.nc.u16 %rs20, [%rd86+2];
mov.b32 %r450, {%rs19, %rs20};
not.b32 %r451, %r450;
ld.global.nc.u16 %rs21, [%rd85+2];
ld.global.nc.u16 %rs22, [%rd85];

{.reg .f16 low,high;
mov.b32 {low,high},%r139;
cvt.f32.f16 %f43, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r140;
cvt.f32.f16 %f44, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r141;
cvt.f32.f16 %f45, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r142;
cvt.f32.f16 %f46, low;}


mov.b32 %r452, {%rs22, %rs21};
shr.s32 %r453, %r451, %r358;
add.s64 %rd87, %rd124, %rd13;
ld.global.nc.u8 %rs23, [%rd87];
cvt.u32.u16 %r454, %rs23;
and.b32 %r455, %r454, 255;
shr.u32 %r456, %r455, %r3;
and.b32 %r457, %r456, 15;
add.s64 %rd88, %rd124, %rd12;
ld.global.nc.u8 %rs24, [%rd88];
cvt.u32.u16 %r458, %rs24;
and.b32 %r459, %r458, 255;
shr.u32 %r460, %r459, %r4;
and.b32 %r461, %r460, 3;
bfi.b32 %r462, %r461, %r457, 4, 2;
add.s32 %r463, %r462, -32;
and.b32 %r325, %r452, 50529027;
shl.b32 %r464, %r453, 2;
and.b32 %r465, %r464, 67372036;
or.b32 %r466, %r325, -2139062144;
sub.s32 %r467, %r466, %r465;
xor.b32 %r468, %r467, -2139062144;
or.b32 %r469, %r325, %r465;
xor.b32 %r470, %r467, %r325;
and.b32 %r323, %r470, %r469;
mov.u32 %r179, %r323;

prmt.b32 %r179, %r179, 0, 0xba98;

mov.u32 %r181, %r325;

prmt.b32 %r181, %r181, 0, 0xba98;

xor.b32 %r471, %r181, 2139062143;
not.b32 %r472, %r179;
and.b32 %r473, %r468, %r472;
and.b32 %r474, %r471, %r179;
or.b32 %r184, %r474, %r473;

dp4a.s32.s32 %r183, %r184, %r149, %r354;

mul.lo.s32 %r475, %r463, %r183;
cvt.rn.f32.s32 %f75, %r475;
fma.rn.ftz.f32 %f76, %f43, %f75, 0f00000000;
add.s64 %rd89, %rd124, %rd11;
ld.global.nc.u8 %rs25, [%rd89];
cvt.u32.u16 %r476, %rs25;
and.b32 %r477, %r476, 255;
shr.u32 %r478, %r477, %r5;
and.b32 %r479, %r478, 15;
add.s64 %rd90, %rd124, %rd10;
ld.global.nc.u8 %rs26, [%rd90];
cvt.u32.u16 %r480, %rs26;
and.b32 %r481, %r480, 255;
shr.u32 %r482, %r481, %r6;
and.b32 %r483, %r482, 3;
bfi.b32 %r484, %r483, %r479, 4, 2;
add.s32 %r485, %r484, -32;
shr.u32 %r486, %r452, 2;
and.b32 %r333, %r486, 50529027;
shl.b32 %r487, %r453, 1;
and.b32 %r488, %r487, 67372036;
or.b32 %r489, %r333, -2139062144;
sub.s32 %r490, %r489, %r488;
xor.b32 %r491, %r490, -2139062144;
or.b32 %r492, %r333, %r488;
xor.b32 %r493, %r490, %r333;
and.b32 %r331, %r493, %r492;
mov.u32 %r187, %r331;

prmt.b32 %r187, %r187, 0, 0xba98;

mov.u32 %r189, %r333;

prmt.b32 %r189, %r189, 0, 0xba98;

xor.b32 %r494, %r189, 2139062143;
not.b32 %r495, %r187;
and.b32 %r496, %r491, %r495;
and.b32 %r497, %r494, %r187;
or.b32 %r192, %r497, %r496;

dp4a.s32.s32 %r191, %r192, %r157, %r354;

mul.lo.s32 %r498, %r485, %r191;
cvt.rn.f32.s32 %f77, %r498;
fma.rn.ftz.f32 %f78, %f44, %f77, %f76;
add.s64 %rd91, %rd124, %rd9;
ld.global.nc.u8 %rs27, [%rd91];
cvt.u32.u16 %r499, %rs27;
and.b32 %r500, %r499, 255;
shr.u32 %r501, %r500, %r7;
and.b32 %r502, %r501, 15;
add.s64 %rd92, %rd124, %rd8;
ld.global.nc.u8 %rs28, [%rd92];
cvt.u32.u16 %r503, %rs28;
and.b32 %r504, %r503, 255;
shr.u32 %r505, %r504, %r8;
and.b32 %r506, %r505, 3;
bfi.b32 %r507, %r506, %r502, 4, 2;
add.s32 %r508, %r507, -32;
shr.u32 %r509, %r452, 4;
and.b32 %r341, %r509, 50529027;
and.b32 %r510, %r453, 67372036;
or.b32 %r511, %r341, -2139062144;
sub.s32 %r512, %r511, %r510;
xor.b32 %r513, %r512, -2139062144;
or.b32 %r514, %r341, %r510;
xor.b32 %r515, %r512, %r341;
and.b32 %r339, %r515, %r514;
mov.u32 %r195, %r339;

prmt.b32 %r195, %r195, 0, 0xba98;

mov.u32 %r197, %r341;

prmt.b32 %r197, %r197, 0, 0xba98;

xor.b32 %r516, %r197, 2139062143;
not.b32 %r517, %r195;
and.b32 %r518, %r513, %r517;
and.b32 %r519, %r516, %r195;
or.b32 %r200, %r519, %r518;

dp4a.s32.s32 %r199, %r200, %r165, %r354;

mul.lo.s32 %r520, %r508, %r199;
cvt.rn.f32.s32 %f79, %r520;
fma.rn.ftz.f32 %f80, %f45, %f79, %f78;
add.s64 %rd93, %rd124, %rd7;
ld.global.nc.u8 %rs29, [%rd93];
cvt.u32.u16 %r521, %rs29;
and.b32 %r522, %r521, 255;
shr.u32 %r523, %r522, %r9;
and.b32 %r524, %r523, 15;
add.s64 %rd94, %rd124, %rd6;
ld.global.nc.u8 %rs30, [%rd94];
cvt.u32.u16 %r525, %rs30;
and.b32 %r526, %r525, 255;
shr.u32 %r527, %r526, %r10;
and.b32 %r528, %r527, 3;
bfi.b32 %r529, %r528, %r524, 4, 2;
add.s32 %r530, %r529, -32;
shr.u32 %r531, %r452, 6;
and.b32 %r349, %r531, 50529027;
shr.u32 %r532, %r453, 1;
and.b32 %r533, %r532, 67372036;
or.b32 %r534, %r349, -2139062144;
sub.s32 %r535, %r534, %r533;
xor.b32 %r536, %r535, -2139062144;
or.b32 %r537, %r349, %r533;
xor.b32 %r538, %r535, %r349;
and.b32 %r347, %r538, %r537;
mov.u32 %r203, %r347;

prmt.b32 %r203, %r203, 0, 0xba98;

mov.u32 %r205, %r349;

prmt.b32 %r205, %r205, 0, 0xba98;

xor.b32 %r539, %r205, 2139062143;
not.b32 %r540, %r203;
and.b32 %r541, %r536, %r540;
and.b32 %r542, %r539, %r203;
or.b32 %r208, %r542, %r541;

dp4a.s32.s32 %r207, %r208, %r173, %r354;

mul.lo.s32 %r543, %r530, %r207;
cvt.rn.f32.s32 %f81, %r543;
fma.rn.ftz.f32 %f82, %f46, %f81, %f80;
fma.rn.ftz.f32 %f215, %f42, %f82, %f215;
cvt.u64.u32 %rd95, %r764;
add.s64 %rd96, %rd1, %rd95;

{ cvt.f32.f16 %f47, %rs1;}


add.s64 %rd97, %rd96, %rd4;
mul.lo.s64 %rd98, %rd97, 36;
add.s64 %rd99, %rd65, %rd98;
add.s64 %rd100, %rd99, %rd74;
ld.global.nc.u32 %r221, [%rd100+4];
ld.global.nc.u32 %r211, [%rd99];

{.reg .f16 low,high;
mov.b32 {low,high},%r211;
cvt.f32.f16 %f48, low;}


ld.global.nc.u32 %r229, [%rd100+40];
ld.global.nc.u32 %r212, [%rd99+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r212;
cvt.f32.f16 %f49, low;}


ld.global.nc.u32 %r237, [%rd100+76];
ld.global.nc.u32 %r213, [%rd99+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r213;
cvt.f32.f16 %f50, low;}


ld.global.nc.u32 %r245, [%rd100+112];
ld.global.nc.u32 %r214, [%rd99+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r214;
cvt.f32.f16 %f51, low;}


mov.u32 %r215, %r287;

prmt.b32 %r215, %r215, 0, 0xba98;

mov.u32 %r217, %r289;

prmt.b32 %r217, %r217, 0, 0xba98;

xor.b32 %r544, %r217, 2139062143;
not.b32 %r545, %r215;
and.b32 %r546, %r374, %r545;
and.b32 %r547, %r544, %r215;
or.b32 %r220, %r547, %r546;

dp4a.s32.s32 %r219, %r220, %r221, %r354;

mul.lo.s32 %r548, %r369, %r219;
cvt.rn.f32.s32 %f83, %r548;
fma.rn.ftz.f32 %f84, %f48, %f83, 0f00000000;
mov.u32 %r223, %r295;

prmt.b32 %r223, %r223, 0, 0xba98;

mov.u32 %r225, %r297;

prmt.b32 %r225, %r225, 0, 0xba98;

xor.b32 %r549, %r225, 2139062143;
not.b32 %r550, %r223;
and.b32 %r551, %r397, %r550;
and.b32 %r552, %r549, %r223;
or.b32 %r228, %r552, %r551;

dp4a.s32.s32 %r227, %r228, %r229, %r354;

mul.lo.s32 %r553, %r391, %r227;
cvt.rn.f32.s32 %f85, %r553;
fma.rn.ftz.f32 %f86, %f49, %f85, %f84;
mov.u32 %r231, %r303;

prmt.b32 %r231, %r231, 0, 0xba98;

mov.u32 %r233, %r305;

prmt.b32 %r233, %r233, 0, 0xba98;

xor.b32 %r554, %r233, 2139062143;
not.b32 %r555, %r231;
and.b32 %r556, %r419, %r555;
and.b32 %r557, %r554, %r231;
or.b32 %r236, %r557, %r556;

dp4a.s32.s32 %r235, %r236, %r237, %r354;

mul.lo.s32 %r558, %r414, %r235;
cvt.rn.f32.s32 %f87, %r558;
fma.rn.ftz.f32 %f88, %f50, %f87, %f86;
mov.u32 %r239, %r311;

prmt.b32 %r239, %r239, 0, 0xba98;

mov.u32 %r241, %r313;

prmt.b32 %r241, %r241, 0, 0xba98;

xor.b32 %r559, %r241, 2139062143;
not.b32 %r560, %r239;
and.b32 %r561, %r442, %r560;
and.b32 %r562, %r559, %r239;
or.b32 %r244, %r562, %r561;

dp4a.s32.s32 %r243, %r244, %r245, %r354;

mul.lo.s32 %r563, %r436, %r243;
cvt.rn.f32.s32 %f89, %r563;
fma.rn.ftz.f32 %f90, %f51, %f89, %f88;
fma.rn.ftz.f32 %f214, %f47, %f90, %f214;

{ cvt.f32.f16 %f52, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r211;
cvt.f32.f16 %f53, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r212;
cvt.f32.f16 %f54, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r213;
cvt.f32.f16 %f55, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r214;
cvt.f32.f16 %f56, low;}


mov.u32 %r251, %r323;

prmt.b32 %r251, %r251, 0, 0xba98;

mov.u32 %r253, %r325;

prmt.b32 %r253, %r253, 0, 0xba98;

xor.b32 %r564, %r253, 2139062143;
not.b32 %r565, %r251;
and.b32 %r566, %r468, %r565;
and.b32 %r567, %r564, %r251;
or.b32 %r256, %r567, %r566;

dp4a.s32.s32 %r255, %r256, %r221, %r354;

mul.lo.s32 %r568, %r463, %r255;
cvt.rn.f32.s32 %f91, %r568;
fma.rn.ftz.f32 %f92, %f53, %f91, 0f00000000;
mov.u32 %r259, %r331;

prmt.b32 %r259, %r259, 0, 0xba98;

mov.u32 %r261, %r333;

prmt.b32 %r261, %r261, 0, 0xba98;

xor.b32 %r569, %r261, 2139062143;
not.b32 %r570, %r259;
and.b32 %r571, %r491, %r570;
and.b32 %r572, %r569, %r259;
or.b32 %r264, %r572, %r571;

dp4a.s32.s32 %r263, %r264, %r229, %r354;

mul.lo.s32 %r573, %r485, %r263;
cvt.rn.f32.s32 %f93, %r573;
fma.rn.ftz.f32 %f94, %f54, %f93, %f92;
mov.u32 %r267, %r339;

prmt.b32 %r267, %r267, 0, 0xba98;

mov.u32 %r269, %r341;

prmt.b32 %r269, %r269, 0, 0xba98;

xor.b32 %r574, %r269, 2139062143;
not.b32 %r575, %r267;
and.b32 %r576, %r513, %r575;
and.b32 %r577, %r574, %r267;
or.b32 %r272, %r577, %r576;

dp4a.s32.s32 %r271, %r272, %r237, %r354;

mul.lo.s32 %r578, %r508, %r271;
cvt.rn.f32.s32 %f95, %r578;
fma.rn.ftz.f32 %f96, %f55, %f95, %f94;
mov.u32 %r275, %r347;

prmt.b32 %r275, %r275, 0, 0xba98;

mov.u32 %r277, %r349;

prmt.b32 %r277, %r277, 0, 0xba98;

xor.b32 %r579, %r277, 2139062143;
not.b32 %r580, %r275;
and.b32 %r581, %r536, %r580;
and.b32 %r582, %r579, %r275;
or.b32 %r280, %r582, %r581;

dp4a.s32.s32 %r279, %r280, %r245, %r354;

mul.lo.s32 %r583, %r530, %r279;
cvt.rn.f32.s32 %f97, %r583;
fma.rn.ftz.f32 %f98, %f56, %f97, %f96;
fma.rn.ftz.f32 %f213, %f52, %f98, %f213;
cvt.u64.u32 %rd101, %r765;
add.s64 %rd102, %rd1, %rd101;

{ cvt.f32.f16 %f57, %rs1;}


add.s64 %rd103, %rd102, %rd4;
mul.lo.s64 %rd104, %rd103, 36;
add.s64 %rd105, %rd65, %rd104;
add.s64 %rd106, %rd105, %rd74;
ld.global.nc.u32 %r293, [%rd106+4];
ld.global.nc.u32 %r283, [%rd105];

{.reg .f16 low,high;
mov.b32 {low,high},%r283;
cvt.f32.f16 %f58, low;}


ld.global.nc.u32 %r301, [%rd106+40];
ld.global.nc.u32 %r284, [%rd105+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r284;
cvt.f32.f16 %f59, low;}


ld.global.nc.u32 %r309, [%rd106+76];
ld.global.nc.u32 %r285, [%rd105+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r285;
cvt.f32.f16 %f60, low;}


ld.global.nc.u32 %r317, [%rd106+112];
ld.global.nc.u32 %r286, [%rd105+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r286;
cvt.f32.f16 %f61, low;}



prmt.b32 %r287, %r287, 0, 0xba98;


prmt.b32 %r289, %r289, 0, 0xba98;

xor.b32 %r584, %r289, 2139062143;
not.b32 %r585, %r287;
and.b32 %r586, %r374, %r585;
and.b32 %r587, %r584, %r287;
or.b32 %r292, %r587, %r586;

dp4a.s32.s32 %r291, %r292, %r293, %r354;

mul.lo.s32 %r588, %r369, %r291;
cvt.rn.f32.s32 %f99, %r588;
fma.rn.ftz.f32 %f100, %f58, %f99, 0f00000000;

prmt.b32 %r295, %r295, 0, 0xba98;


prmt.b32 %r297, %r297, 0, 0xba98;

xor.b32 %r589, %r297, 2139062143;
not.b32 %r590, %r295;
and.b32 %r591, %r397, %r590;
and.b32 %r592, %r589, %r295;
or.b32 %r300, %r592, %r591;

dp4a.s32.s32 %r299, %r300, %r301, %r354;

mul.lo.s32 %r593, %r391, %r299;
cvt.rn.f32.s32 %f101, %r593;
fma.rn.ftz.f32 %f102, %f59, %f101, %f100;

prmt.b32 %r303, %r303, 0, 0xba98;


prmt.b32 %r305, %r305, 0, 0xba98;

xor.b32 %r594, %r305, 2139062143;
not.b32 %r595, %r303;
and.b32 %r596, %r419, %r595;
and.b32 %r597, %r594, %r303;
or.b32 %r308, %r597, %r596;

dp4a.s32.s32 %r307, %r308, %r309, %r354;

mul.lo.s32 %r598, %r414, %r307;
cvt.rn.f32.s32 %f103, %r598;
fma.rn.ftz.f32 %f104, %f60, %f103, %f102;

prmt.b32 %r311, %r311, 0, 0xba98;


prmt.b32 %r313, %r313, 0, 0xba98;

xor.b32 %r599, %r313, 2139062143;
not.b32 %r600, %r311;
and.b32 %r601, %r442, %r600;
and.b32 %r602, %r599, %r311;
or.b32 %r316, %r602, %r601;

dp4a.s32.s32 %r315, %r316, %r317, %r354;

mul.lo.s32 %r603, %r436, %r315;
cvt.rn.f32.s32 %f105, %r603;
fma.rn.ftz.f32 %f106, %f61, %f105, %f104;
fma.rn.ftz.f32 %f212, %f57, %f106, %f212;

{ cvt.f32.f16 %f62, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r283;
cvt.f32.f16 %f63, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r284;
cvt.f32.f16 %f64, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r285;
cvt.f32.f16 %f65, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r286;
cvt.f32.f16 %f66, low;}



prmt.b32 %r323, %r323, 0, 0xba98;


prmt.b32 %r325, %r325, 0, 0xba98;

xor.b32 %r604, %r325, 2139062143;
not.b32 %r605, %r323;
and.b32 %r606, %r468, %r605;
and.b32 %r607, %r604, %r323;
or.b32 %r328, %r607, %r606;

dp4a.s32.s32 %r327, %r328, %r293, %r354;

mul.lo.s32 %r608, %r463, %r327;
cvt.rn.f32.s32 %f107, %r608;
fma.rn.ftz.f32 %f108, %f63, %f107, 0f00000000;

prmt.b32 %r331, %r331, 0, 0xba98;


prmt.b32 %r333, %r333, 0, 0xba98;

xor.b32 %r609, %r333, 2139062143;
not.b32 %r610, %r331;
and.b32 %r611, %r491, %r610;
and.b32 %r612, %r609, %r331;
or.b32 %r336, %r612, %r611;

dp4a.s32.s32 %r335, %r336, %r301, %r354;

mul.lo.s32 %r613, %r485, %r335;
cvt.rn.f32.s32 %f109, %r613;
fma.rn.ftz.f32 %f110, %f64, %f109, %f108;

prmt.b32 %r339, %r339, 0, 0xba98;


prmt.b32 %r341, %r341, 0, 0xba98;

xor.b32 %r614, %r341, 2139062143;
not.b32 %r615, %r339;
and.b32 %r616, %r513, %r615;
and.b32 %r617, %r614, %r339;
or.b32 %r344, %r617, %r616;

dp4a.s32.s32 %r343, %r344, %r309, %r354;

mul.lo.s32 %r618, %r508, %r343;
cvt.rn.f32.s32 %f111, %r618;
fma.rn.ftz.f32 %f112, %f65, %f111, %f110;

prmt.b32 %r347, %r347, 0, 0xba98;


prmt.b32 %r349, %r349, 0, 0xba98;

xor.b32 %r619, %r349, 2139062143;
not.b32 %r620, %r347;
and.b32 %r621, %r536, %r620;
and.b32 %r622, %r619, %r347;
or.b32 %r352, %r622, %r621;

dp4a.s32.s32 %r351, %r352, %r317, %r354;

mul.lo.s32 %r623, %r530, %r351;
cvt.rn.f32.s32 %f113, %r623;
fma.rn.ftz.f32 %f114, %f66, %f113, %f112;
fma.rn.ftz.f32 %f211, %f62, %f114, %f211;
add.s32 %r765, %r765, 64;
add.s32 %r764, %r764, 64;
add.s32 %r763, %r763, 64;
add.s64 %rd124, %rd124, 880;
add.s32 %r766, %r766, 8;
setp.lt.s32 %p4, %r766, %r46;
@%p4 bra $L__BB147_2;

$L__BB147_3:
mov.u32 %r750, %tid.y;
setp.eq.s32 %p5, %r750, 0;
@%p5 bra $L__BB147_5;

mov.u32 %r752, %tid.x;
mov.u32 %r751, %tid.y;
mov.u32 %r628, _ZZ13mul_mat_vec_qIL9ggml_type11ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r629, %r751, 768, %r628;
shl.b32 %r630, %r752, 2;
add.s32 %r631, %r629, %r630;
st.shared.f32 [%r631+-768], %f216;
st.shared.f32 [%r631+-640], %f215;
st.shared.f32 [%r631+-512], %f214;
st.shared.f32 [%r631+-384], %f213;
st.shared.f32 [%r631+-256], %f212;
st.shared.f32 [%r631+-128], %f211;

$L__BB147_5:
mov.u32 %r753, %tid.y;
setp.ne.s32 %p6, %r753, 0;
bar.sync 0;
@%p6 bra $L__BB147_18;

ld.param.u32 %r758, [_Z13mul_mat_vec_qIL9ggml_type11ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u64 %rd123, [_Z13mul_mat_vec_qIL9ggml_type11ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r756, [_Z13mul_mat_vec_qIL9ggml_type11ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r755, [_Z13mul_mat_vec_qIL9ggml_type11ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
mov.u32 %r754, %tid.x;
mov.u32 %r633, %ctaid.z;
mov.u32 %r634, %ctaid.y;
mov.u32 %r635, %ctaid.x;
shl.b32 %r636, %r635, 1;
mov.u32 %r637, 1;
mad.lo.s32 %r638, %r634, %r755, %r636;
mad.lo.s32 %r639, %r633, %r756, %r638;
cvt.u64.u32 %rd29, %r639;
shl.b32 %r641, %r754, 2;
mov.u32 %r642, _ZZ13mul_mat_vec_qIL9ggml_type11ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r22, %r642, %r641;
mov.u32 %r643, 2;
ld.shared.f32 %f115, [%r22];
add.ftz.f32 %f116, %f115, %f216;
ld.shared.f32 %f117, [%r22+768];
add.ftz.f32 %f118, %f117, %f116;
ld.shared.f32 %f119, [%r22+1536];
add.ftz.f32 %f120, %f119, %f118;
mov.b32 %r644, %f120;
mov.u32 %r645, 31;
mov.u32 %r646, 16;
mov.u32 %r647, -1;
shfl.sync.bfly.b32 %r648|%p7, %r644, %r646, %r645, %r647;
mov.b32 %f121, %r648;
add.ftz.f32 %f122, %f120, %f121;
mov.b32 %r649, %f122;
mov.u32 %r650, 8;
shfl.sync.bfly.b32 %r651|%p8, %r649, %r650, %r645, %r647;
mov.b32 %f123, %r651;
add.ftz.f32 %f124, %f122, %f123;
mov.b32 %r652, %f124;
mov.u32 %r653, 4;
shfl.sync.bfly.b32 %r654|%p9, %r652, %r653, %r645, %r647;
mov.b32 %f125, %r654;
add.ftz.f32 %f126, %f124, %f125;
mov.b32 %r655, %f126;
shfl.sync.bfly.b32 %r656|%p10, %r655, %r643, %r645, %r647;
mov.b32 %f127, %r656;
add.ftz.f32 %f128, %f126, %f127;
mov.b32 %r657, %f128;
shfl.sync.bfly.b32 %r658|%p11, %r657, %r637, %r645, %r647;
mov.b32 %f129, %r658;
add.ftz.f32 %f19, %f128, %f129;
setp.eq.s32 %p12, %r754, 0;
setp.lt.u32 %p13, %r636, %r758;
and.pred %p1, %p12, %p13;
cvta.to.global.u64 %rd107, %rd123;
mul.wide.u32 %rd108, %r639, 4;
add.s64 %rd30, %rd107, %rd108;
not.pred %p14, %p1;
@%p14 bra $L__BB147_8;

st.global.f32 [%rd30], %f19;

$L__BB147_8:
ld.param.u32 %r759, [_Z13mul_mat_vec_qIL9ggml_type11ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
mov.u32 %r757, %tid.x;
ld.shared.f32 %f130, [%r22+128];
add.ftz.f32 %f131, %f130, %f215;
ld.shared.f32 %f132, [%r22+896];
add.ftz.f32 %f133, %f132, %f131;
ld.shared.f32 %f134, [%r22+1664];
add.ftz.f32 %f135, %f134, %f133;
mov.b32 %r659, %f135;
shfl.sync.bfly.b32 %r663|%p15, %r659, %r646, %r645, %r647;
mov.b32 %f136, %r663;
add.ftz.f32 %f137, %f135, %f136;
mov.b32 %r664, %f137;
shfl.sync.bfly.b32 %r666|%p16, %r664, %r650, %r645, %r647;
mov.b32 %f138, %r666;
add.ftz.f32 %f139, %f137, %f138;
mov.b32 %r667, %f139;
shfl.sync.bfly.b32 %r669|%p17, %r667, %r653, %r645, %r647;
mov.b32 %f140, %r669;
add.ftz.f32 %f141, %f139, %f140;
mov.b32 %r670, %f141;
shfl.sync.bfly.b32 %r672|%p18, %r670, %r643, %r645, %r647;
mov.b32 %f142, %r672;
add.ftz.f32 %f143, %f141, %f142;
mov.b32 %r673, %f143;
shfl.sync.bfly.b32 %r675|%p19, %r673, %r637, %r645, %r647;
mov.b32 %f144, %r675;
add.ftz.f32 %f20, %f143, %f144;
setp.eq.s32 %p20, %r757, 1;
or.b32 %r679, %r636, 1;
setp.lt.u32 %p21, %r679, %r759;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB147_10;

st.global.f32 [%rd30+4], %f20;

$L__BB147_10:
ld.shared.f32 %f145, [%r22+256];
add.ftz.f32 %f146, %f145, %f214;
ld.shared.f32 %f147, [%r22+1024];
add.ftz.f32 %f148, %f147, %f146;
ld.shared.f32 %f149, [%r22+1792];
add.ftz.f32 %f150, %f149, %f148;
mov.b32 %r680, %f150;
mov.u32 %r681, 31;
mov.u32 %r682, 16;
mov.u32 %r683, -1;
shfl.sync.bfly.b32 %r684|%p23, %r680, %r682, %r681, %r683;
mov.b32 %f151, %r684;
add.ftz.f32 %f152, %f150, %f151;
mov.b32 %r685, %f152;
mov.u32 %r686, 8;
shfl.sync.bfly.b32 %r687|%p24, %r685, %r686, %r681, %r683;
mov.b32 %f153, %r687;
add.ftz.f32 %f154, %f152, %f153;
mov.b32 %r688, %f154;
mov.u32 %r689, 4;
shfl.sync.bfly.b32 %r690|%p25, %r688, %r689, %r681, %r683;
mov.b32 %f155, %r690;
add.ftz.f32 %f156, %f154, %f155;
mov.b32 %r691, %f156;
mov.u32 %r692, 2;
shfl.sync.bfly.b32 %r693|%p26, %r691, %r692, %r681, %r683;
mov.b32 %f157, %r693;
add.ftz.f32 %f158, %f156, %f157;
mov.b32 %r694, %f158;
mov.u32 %r695, 1;
shfl.sync.bfly.b32 %r696|%p27, %r694, %r695, %r681, %r683;
mov.b32 %f159, %r696;
add.ftz.f32 %f21, %f158, %f159;
@%p14 bra $L__BB147_12;

ld.param.u32 %r762, [_Z13mul_mat_vec_qIL9ggml_type11ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
cvt.u64.u32 %rd109, %r762;
add.s64 %rd110, %rd109, %rd29;
shl.b64 %rd112, %rd110, 2;
add.s64 %rd113, %rd107, %rd112;
st.global.f32 [%rd113], %f21;

$L__BB147_12:
ld.shared.f32 %f160, [%r22+384];
add.ftz.f32 %f161, %f160, %f213;
ld.shared.f32 %f162, [%r22+1152];
add.ftz.f32 %f163, %f162, %f161;
ld.shared.f32 %f164, [%r22+1920];
add.ftz.f32 %f165, %f164, %f163;
mov.b32 %r697, %f165;
shfl.sync.bfly.b32 %r701|%p29, %r697, %r682, %r681, %r683;
mov.b32 %f166, %r701;
add.ftz.f32 %f167, %f165, %f166;
mov.b32 %r702, %f167;
shfl.sync.bfly.b32 %r704|%p30, %r702, %r686, %r681, %r683;
mov.b32 %f168, %r704;
add.ftz.f32 %f169, %f167, %f168;
mov.b32 %r705, %f169;
shfl.sync.bfly.b32 %r707|%p31, %r705, %r689, %r681, %r683;
mov.b32 %f170, %r707;
add.ftz.f32 %f171, %f169, %f170;
mov.b32 %r708, %f171;
shfl.sync.bfly.b32 %r710|%p32, %r708, %r692, %r681, %r683;
mov.b32 %f172, %r710;
add.ftz.f32 %f173, %f171, %f172;
mov.b32 %r711, %f173;
shfl.sync.bfly.b32 %r713|%p33, %r711, %r695, %r681, %r683;
mov.b32 %f174, %r713;
add.ftz.f32 %f22, %f173, %f174;
@%p22 bra $L__BB147_14;

ld.param.u32 %r761, [_Z13mul_mat_vec_qIL9ggml_type11ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
add.s32 %r714, %r761, 1;
cvt.u64.u32 %rd114, %r714;
add.s64 %rd115, %rd114, %rd29;
shl.b64 %rd117, %rd115, 2;
add.s64 %rd118, %rd107, %rd117;
st.global.f32 [%rd118], %f22;

$L__BB147_14:
ld.param.u32 %r760, [_Z13mul_mat_vec_qIL9ggml_type11ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
shl.b32 %r715, %r760, 1;
mov.u32 %r716, 1;
ld.shared.f32 %f175, [%r22+512];
add.ftz.f32 %f176, %f175, %f212;
ld.shared.f32 %f177, [%r22+1280];
add.ftz.f32 %f178, %f177, %f176;
ld.shared.f32 %f179, [%r22+2048];
add.ftz.f32 %f180, %f179, %f178;
mov.b32 %r717, %f180;
mov.u32 %r718, 31;
mov.u32 %r719, 16;
mov.u32 %r720, -1;
shfl.sync.bfly.b32 %r721|%p35, %r717, %r719, %r718, %r720;
mov.b32 %f181, %r721;
add.ftz.f32 %f182, %f180, %f181;
mov.b32 %r722, %f182;
mov.u32 %r723, 8;
shfl.sync.bfly.b32 %r724|%p36, %r722, %r723, %r718, %r720;
mov.b32 %f183, %r724;
add.ftz.f32 %f184, %f182, %f183;
mov.b32 %r725, %f184;
mov.u32 %r726, 4;
shfl.sync.bfly.b32 %r727|%p37, %r725, %r726, %r718, %r720;
mov.b32 %f185, %r727;
add.ftz.f32 %f186, %f184, %f185;
mov.b32 %r728, %f186;
mov.u32 %r729, 2;
shfl.sync.bfly.b32 %r730|%p38, %r728, %r729, %r718, %r720;
mov.b32 %f187, %r730;
add.ftz.f32 %f188, %f186, %f187;
mov.b32 %r731, %f188;
shfl.sync.bfly.b32 %r732|%p39, %r731, %r716, %r718, %r720;
mov.b32 %f189, %r732;
add.ftz.f32 %f23, %f188, %f189;
cvt.u64.u32 %rd119, %r715;
add.s64 %rd120, %rd119, %rd29;
shl.b64 %rd122, %rd120, 2;
add.s64 %rd31, %rd107, %rd122;
@%p14 bra $L__BB147_16;

st.global.f32 [%rd31], %f23;

$L__BB147_16:
ld.shared.f32 %f190, [%r22+640];
add.ftz.f32 %f191, %f190, %f211;
ld.shared.f32 %f192, [%r22+1408];
add.ftz.f32 %f193, %f192, %f191;
ld.shared.f32 %f194, [%r22+2176];
add.ftz.f32 %f195, %f194, %f193;
mov.b32 %r733, %f195;
shfl.sync.bfly.b32 %r737|%p41, %r733, %r719, %r718, %r720;
mov.b32 %f196, %r737;
add.ftz.f32 %f197, %f195, %f196;
mov.b32 %r738, %f197;
shfl.sync.bfly.b32 %r740|%p42, %r738, %r723, %r718, %r720;
mov.b32 %f198, %r740;
add.ftz.f32 %f199, %f197, %f198;
mov.b32 %r741, %f199;
shfl.sync.bfly.b32 %r743|%p43, %r741, %r726, %r718, %r720;
mov.b32 %f200, %r743;
add.ftz.f32 %f201, %f199, %f200;
mov.b32 %r744, %f201;
shfl.sync.bfly.b32 %r746|%p44, %r744, %r729, %r718, %r720;
mov.b32 %f202, %r746;
add.ftz.f32 %f203, %f201, %f202;
mov.b32 %r747, %f203;
shfl.sync.bfly.b32 %r749|%p45, %r747, %r716, %r718, %r720;
mov.b32 %f204, %r749;
add.ftz.f32 %f24, %f203, %f204;
@%p22 bra $L__BB147_18;

st.global.f32 [%rd31+4], %f24;

$L__BB147_18:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type11ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type11ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type11ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type11ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type11ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type11ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type11ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type11ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type11ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<59>;
.reg .b16 %rs<33>;
.reg .f32 %f<289>;
.reg .b32 %r<921>;
.reg .b64 %rd<141>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type11ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[3072];

ld.param.u64 %rd32, [_Z13mul_mat_vec_qIL9ggml_type11ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd33, [_Z13mul_mat_vec_qIL9ggml_type11ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type11ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type11ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type11ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type11ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type11ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type11ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type11ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type11ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type11ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type11ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type11ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r44, %tid.y;
shl.b32 %r45, %r44, 5;
mov.u32 %r46, %tid.x;
add.s32 %r1, %r45, %r46;
shr.s32 %r47, %r1, 31;
shr.u32 %r48, %r47, 28;
add.s32 %r49, %r1, %r48;
shr.s32 %r920, %r49, 4;
shr.u32 %r50, %r28, 8;
setp.ge.s32 %p3, %r920, %r50;
mov.f32 %f281, 0f00000000;
mov.f32 %f282, %f281;
mov.f32 %f283, %f281;
mov.f32 %f284, %f281;
mov.f32 %f285, %f281;
mov.f32 %f286, %f281;
mov.f32 %f287, %f281;
mov.f32 %f288, %f281;
@%p3 bra $L__BB148_3;

mov.u32 %r51, %ctaid.x;
shl.b32 %r52, %r51, 1;
mov.u32 %r53, %ctaid.y;
mov.u32 %r54, %ctaid.z;
mul.lo.s32 %r55, %r53, %r36;
cvt.u64.u32 %rd35, %r55;
mul.lo.s32 %r56, %r54, %r42;
cvt.u64.u32 %rd36, %r56;
add.s64 %rd1, %rd36, %rd35;
mul.hi.u32 %r57, %r54, %r38;
add.s32 %r58, %r57, %r54;
shr.u32 %r59, %r58, %r39;
mul.lo.s32 %r60, %r59, %r41;
mul.hi.u32 %r61, %r53, %r32;
add.s32 %r62, %r61, %r53;
shr.u32 %r63, %r62, %r33;
mul.lo.s32 %r64, %r63, %r35;
cvta.to.global.u64 %rd140, %rd32;
or.b32 %r65, %r52, 1;
and.b32 %r69, %r49, -16;
sub.s32 %r70, %r1, %r69;
shr.s32 %r71, %r70, 31;
shr.u32 %r72, %r71, 29;
add.s32 %r73, %r70, %r72;
and.b32 %r74, %r73, -8;
sub.s32 %r75, %r70, %r74;
shr.s32 %r76, %r75, 31;
shr.u32 %r77, %r76, 30;
add.s32 %r78, %r75, %r77;
shr.s32 %r79, %r78, 2;
add.s32 %r80, %r74, %r79;
shr.s32 %r81, %r80, 31;
shr.u32 %r82, %r81, 29;
add.s32 %r83, %r80, %r82;
and.b32 %r84, %r83, -8;
sub.s32 %r85, %r80, %r84;
cvt.s64.s32 %rd37, %r85;
shr.s32 %r86, %r73, 3;
shl.b32 %r87, %r86, 2;
shl.b32 %r88, %r70, 1;
shl.b32 %r89, %r75, 1;
cvt.s64.s32 %rd3, %r75;
shr.s32 %r90, %r83, 3;
shl.b32 %r3, %r90, 2;
cvt.s64.s32 %rd4, %r87;
shr.u32 %r91, %r81, 30;
add.s32 %r92, %r80, %r91;
and.b32 %r93, %r92, -4;
sub.s32 %r94, %r80, %r93;
shr.s32 %r95, %r92, 2;
shl.b32 %r4, %r95, 1;
add.s32 %r96, %r94, 8;
cvt.s64.s32 %rd38, %r96;
add.s32 %r97, %r80, 2;
shr.s32 %r98, %r97, 31;
shr.u32 %r99, %r98, 29;
add.s32 %r100, %r97, %r99;
and.b32 %r101, %r100, -8;
sub.s32 %r102, %r97, %r101;
shr.s32 %r103, %r100, 3;
shl.b32 %r5, %r103, 2;
cvt.s64.s32 %rd39, %r102;
shr.u32 %r104, %r98, 30;
add.s32 %r105, %r97, %r104;
and.b32 %r106, %r105, -4;
sub.s32 %r107, %r97, %r106;
shr.s32 %r108, %r105, 2;
shl.b32 %r6, %r108, 1;
add.s32 %r109, %r107, 8;
cvt.s64.s32 %rd40, %r109;
add.s32 %r110, %r80, 4;
shr.s32 %r111, %r110, 31;
shr.u32 %r112, %r111, 29;
add.s32 %r113, %r110, %r112;
and.b32 %r114, %r113, -8;
sub.s32 %r115, %r110, %r114;
shr.s32 %r116, %r113, 3;
shl.b32 %r7, %r116, 2;
cvt.s64.s32 %rd41, %r115;
shr.u32 %r117, %r111, 30;
add.s32 %r118, %r110, %r117;
and.b32 %r119, %r118, -4;
sub.s32 %r120, %r110, %r119;
shr.s32 %r121, %r118, 2;
shl.b32 %r8, %r121, 1;
add.s32 %r122, %r120, 8;
cvt.s64.s32 %rd42, %r122;
add.s32 %r123, %r80, 6;
shr.s32 %r124, %r123, 31;
shr.u32 %r125, %r124, 29;
add.s32 %r126, %r123, %r125;
and.b32 %r127, %r126, -8;
sub.s32 %r128, %r123, %r127;
shr.s32 %r129, %r126, 3;
shl.b32 %r9, %r129, 2;
cvt.s64.s32 %rd43, %r128;
shr.u32 %r130, %r124, 30;
add.s32 %r131, %r123, %r130;
and.b32 %r132, %r131, -4;
sub.s32 %r133, %r123, %r132;
shr.s32 %r134, %r131, 2;
shl.b32 %r10, %r134, 1;
add.s32 %r135, %r133, 8;
cvt.s64.s32 %rd44, %r135;
shl.b32 %r136, %r30, 1;
shl.b32 %r916, %r920, 3;
add.s32 %r919, %r916, %r136;
mad.lo.s32 %r918, %r30, 3, %r916;
add.s32 %r917, %r30, %r916;
add.s32 %r137, %r920, %r60;
add.s32 %r138, %r137, %r64;
mad.lo.s32 %r139, %r65, %r29, %r138;
mul.wide.s32 %rd5, %r139, 110;
add.s64 %rd45, %rd5, %rd44;
add.s64 %rd6, %rd45, 96;
add.s64 %rd46, %rd5, %rd43;
add.s64 %rd7, %rd46, 96;
add.s64 %rd47, %rd5, %rd42;
add.s64 %rd8, %rd47, 96;
add.s64 %rd48, %rd5, %rd41;
add.s64 %rd9, %rd48, 96;
add.s64 %rd49, %rd5, %rd40;
add.s64 %rd10, %rd49, 96;
add.s64 %rd50, %rd5, %rd39;
add.s64 %rd11, %rd50, 96;
add.s64 %rd51, %rd5, %rd38;
add.s64 %rd12, %rd51, 96;
add.s64 %rd52, %rd5, %rd37;
add.s64 %rd13, %rd52, 96;
mad.lo.s32 %r140, %r52, %r29, %r920;
add.s32 %r141, %r140, %r60;
add.s32 %r142, %r141, %r64;
mul.wide.s32 %rd14, %r142, 110;
mul.wide.s32 %rd53, %r88, 2;
add.s64 %rd15, %rd14, %rd53;
mul.wide.s32 %rd54, %r89, 2;
add.s64 %rd16, %rd5, %rd54;
add.s64 %rd55, %rd14, %rd54;
add.s64 %rd17, %rd55, 2;
add.s64 %rd56, %rd5, %rd53;
add.s64 %rd18, %rd56, 32;
add.s64 %rd57, %rd14, %rd44;
add.s64 %rd19, %rd57, 96;
add.s64 %rd58, %rd14, %rd43;
add.s64 %rd20, %rd58, 96;
add.s64 %rd59, %rd14, %rd42;
add.s64 %rd21, %rd59, 96;
add.s64 %rd60, %rd14, %rd41;
add.s64 %rd22, %rd60, 96;
add.s64 %rd61, %rd14, %rd40;
add.s64 %rd23, %rd61, 96;
add.s64 %rd62, %rd14, %rd39;
add.s64 %rd24, %rd62, 96;
add.s64 %rd63, %rd14, %rd38;
add.s64 %rd25, %rd63, 96;
add.s64 %rd64, %rd14, %rd37;
add.s64 %rd26, %rd64, 96;
mov.f32 %f281, 0f00000000;
cvta.to.global.u64 %rd65, %rd33;
shl.b64 %rd74, %rd3, 2;

$L__BB148_2:
cvt.u64.u32 %rd66, %r916;
add.s64 %rd67, %rd1, %rd66;
add.s64 %rd68, %rd140, %rd14;
ld.global.nc.u16 %rs1, [%rd68+108];

{ cvt.f32.f16 %f49, %rs1;}


add.s64 %rd69, %rd140, %rd15;
add.s64 %rd70, %rd140, %rd17;
ld.global.nc.u16 %rs9, [%rd70+-2];
ld.global.nc.u16 %rs10, [%rd70];
mov.b32 %r431, {%rs9, %rs10};
not.b32 %r432, %r431;
ld.global.nc.u16 %rs11, [%rd69+34];
ld.global.nc.u16 %rs12, [%rd69+32];
add.s64 %rd71, %rd67, %rd4;
mul.lo.s64 %rd72, %rd71, 36;
add.s64 %rd73, %rd65, %rd72;
add.s64 %rd75, %rd73, %rd74;
ld.global.nc.u32 %r153, [%rd75+4];
ld.global.nc.u32 %r143, [%rd73];

{.reg .f16 low,high;
mov.b32 {low,high},%r143;
cvt.f32.f16 %f50, low;}


ld.global.nc.u32 %r161, [%rd75+40];
ld.global.nc.u32 %r144, [%rd73+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r144;
cvt.f32.f16 %f51, low;}


ld.global.nc.u32 %r169, [%rd75+76];
ld.global.nc.u32 %r145, [%rd73+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r145;
cvt.f32.f16 %f52, low;}


ld.global.nc.u32 %r177, [%rd75+112];
ld.global.nc.u32 %r146, [%rd73+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r146;
cvt.f32.f16 %f53, low;}


mov.b32 %r433, {%rs12, %rs11};
cvt.u32.u64 %r434, %rd4;
shr.s32 %r435, %r432, %r434;
add.s64 %rd76, %rd140, %rd26;
ld.global.nc.u8 %rs13, [%rd76];
cvt.u32.u16 %r436, %rs13;
and.b32 %r437, %r436, 255;
shr.u32 %r438, %r437, %r3;
and.b32 %r439, %r438, 15;
add.s64 %rd77, %rd140, %rd25;
ld.global.nc.u8 %rs14, [%rd77];
cvt.u32.u16 %r440, %rs14;
and.b32 %r441, %r440, 255;
shr.u32 %r442, %r441, %r4;
and.b32 %r443, %r442, 3;
bfi.b32 %r444, %r443, %r439, 4, 2;
add.s32 %r445, %r444, -32;
and.b32 %r365, %r433, 50529027;
shl.b32 %r446, %r435, 2;
and.b32 %r447, %r446, 67372036;
or.b32 %r448, %r365, -2139062144;
sub.s32 %r449, %r448, %r447;
xor.b32 %r450, %r449, -2139062144;
or.b32 %r451, %r365, %r447;
xor.b32 %r452, %r449, %r365;
and.b32 %r363, %r452, %r451;
mov.u32 %r147, %r363;

prmt.b32 %r147, %r147, 0, 0xba98;

mov.u32 %r149, %r365;

prmt.b32 %r149, %r149, 0, 0xba98;

xor.b32 %r453, %r149, 2139062143;
not.b32 %r454, %r147;
and.b32 %r455, %r450, %r454;
and.b32 %r456, %r453, %r147;
or.b32 %r152, %r456, %r455;
mov.u32 %r430, 0;

dp4a.s32.s32 %r151, %r152, %r153, %r430;

mul.lo.s32 %r457, %r445, %r151;
cvt.rn.f32.s32 %f89, %r457;
fma.rn.ftz.f32 %f90, %f50, %f89, 0f00000000;
add.s64 %rd78, %rd140, %rd24;
ld.global.nc.u8 %rs15, [%rd78];
cvt.u32.u16 %r458, %rs15;
and.b32 %r459, %r458, 255;
shr.u32 %r460, %r459, %r5;
and.b32 %r461, %r460, 15;
add.s64 %rd79, %rd140, %rd23;
ld.global.nc.u8 %rs16, [%rd79];
cvt.u32.u16 %r462, %rs16;
and.b32 %r463, %r462, 255;
shr.u32 %r464, %r463, %r6;
and.b32 %r465, %r464, 3;
bfi.b32 %r466, %r465, %r461, 4, 2;
add.s32 %r467, %r466, -32;
shr.u32 %r468, %r433, 2;
and.b32 %r373, %r468, 50529027;
shl.b32 %r469, %r435, 1;
and.b32 %r470, %r469, 67372036;
or.b32 %r471, %r373, -2139062144;
sub.s32 %r472, %r471, %r470;
xor.b32 %r473, %r472, -2139062144;
or.b32 %r474, %r373, %r470;
xor.b32 %r475, %r472, %r373;
and.b32 %r371, %r475, %r474;
mov.u32 %r155, %r371;

prmt.b32 %r155, %r155, 0, 0xba98;

mov.u32 %r157, %r373;

prmt.b32 %r157, %r157, 0, 0xba98;

xor.b32 %r476, %r157, 2139062143;
not.b32 %r477, %r155;
and.b32 %r478, %r473, %r477;
and.b32 %r479, %r476, %r155;
or.b32 %r160, %r479, %r478;

dp4a.s32.s32 %r159, %r160, %r161, %r430;

mul.lo.s32 %r480, %r467, %r159;
cvt.rn.f32.s32 %f91, %r480;
fma.rn.ftz.f32 %f92, %f51, %f91, %f90;
add.s64 %rd80, %rd140, %rd22;
ld.global.nc.u8 %rs17, [%rd80];
cvt.u32.u16 %r481, %rs17;
and.b32 %r482, %r481, 255;
shr.u32 %r483, %r482, %r7;
and.b32 %r484, %r483, 15;
add.s64 %rd81, %rd140, %rd21;
ld.global.nc.u8 %rs18, [%rd81];
cvt.u32.u16 %r485, %rs18;
and.b32 %r486, %r485, 255;
shr.u32 %r487, %r486, %r8;
and.b32 %r488, %r487, 3;
bfi.b32 %r489, %r488, %r484, 4, 2;
add.s32 %r490, %r489, -32;
shr.u32 %r491, %r433, 4;
and.b32 %r381, %r491, 50529027;
and.b32 %r492, %r435, 67372036;
or.b32 %r493, %r381, -2139062144;
sub.s32 %r494, %r493, %r492;
xor.b32 %r495, %r494, -2139062144;
or.b32 %r496, %r381, %r492;
xor.b32 %r497, %r494, %r381;
and.b32 %r379, %r497, %r496;
mov.u32 %r163, %r379;

prmt.b32 %r163, %r163, 0, 0xba98;

mov.u32 %r165, %r381;

prmt.b32 %r165, %r165, 0, 0xba98;

xor.b32 %r498, %r165, 2139062143;
not.b32 %r499, %r163;
and.b32 %r500, %r495, %r499;
and.b32 %r501, %r498, %r163;
or.b32 %r168, %r501, %r500;

dp4a.s32.s32 %r167, %r168, %r169, %r430;

mul.lo.s32 %r502, %r490, %r167;
cvt.rn.f32.s32 %f93, %r502;
fma.rn.ftz.f32 %f94, %f52, %f93, %f92;
add.s64 %rd82, %rd140, %rd20;
ld.global.nc.u8 %rs19, [%rd82];
cvt.u32.u16 %r503, %rs19;
and.b32 %r504, %r503, 255;
shr.u32 %r505, %r504, %r9;
and.b32 %r506, %r505, 15;
add.s64 %rd83, %rd140, %rd19;
ld.global.nc.u8 %rs20, [%rd83];
cvt.u32.u16 %r507, %rs20;
and.b32 %r508, %r507, 255;
shr.u32 %r509, %r508, %r10;
and.b32 %r510, %r509, 3;
bfi.b32 %r511, %r510, %r506, 4, 2;
add.s32 %r512, %r511, -32;
shr.u32 %r513, %r433, 6;
and.b32 %r389, %r513, 50529027;
shr.u32 %r514, %r435, 1;
and.b32 %r515, %r514, 67372036;
or.b32 %r516, %r389, -2139062144;
sub.s32 %r517, %r516, %r515;
xor.b32 %r518, %r517, -2139062144;
or.b32 %r519, %r389, %r515;
xor.b32 %r520, %r517, %r389;
and.b32 %r387, %r520, %r519;
mov.u32 %r171, %r387;

prmt.b32 %r171, %r171, 0, 0xba98;

mov.u32 %r173, %r389;

prmt.b32 %r173, %r173, 0, 0xba98;

xor.b32 %r521, %r173, 2139062143;
not.b32 %r522, %r171;
and.b32 %r523, %r518, %r522;
and.b32 %r524, %r521, %r171;
or.b32 %r176, %r524, %r523;

dp4a.s32.s32 %r175, %r176, %r177, %r430;

mul.lo.s32 %r525, %r512, %r175;
cvt.rn.f32.s32 %f95, %r525;
fma.rn.ftz.f32 %f96, %f53, %f95, %f94;
fma.rn.ftz.f32 %f288, %f49, %f96, %f288;
add.s64 %rd84, %rd140, %rd5;
ld.global.nc.u16 %rs2, [%rd84+108];

{ cvt.f32.f16 %f54, %rs2;}


add.s64 %rd85, %rd140, %rd18;
add.s64 %rd86, %rd140, %rd16;
ld.global.nc.u16 %rs21, [%rd86];
ld.global.nc.u16 %rs22, [%rd86+2];
mov.b32 %r526, {%rs21, %rs22};
not.b32 %r527, %r526;
ld.global.nc.u16 %rs23, [%rd85+2];
ld.global.nc.u16 %rs24, [%rd85];

{.reg .f16 low,high;
mov.b32 {low,high},%r143;
cvt.f32.f16 %f55, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r144;
cvt.f32.f16 %f56, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r145;
cvt.f32.f16 %f57, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r146;
cvt.f32.f16 %f58, low;}


mov.b32 %r528, {%rs24, %rs23};
shr.s32 %r529, %r527, %r434;
add.s64 %rd87, %rd140, %rd13;
ld.global.nc.u8 %rs25, [%rd87];
cvt.u32.u16 %r530, %rs25;
and.b32 %r531, %r530, 255;
shr.u32 %r532, %r531, %r3;
and.b32 %r533, %r532, 15;
add.s64 %rd88, %rd140, %rd12;
ld.global.nc.u8 %rs26, [%rd88];
cvt.u32.u16 %r534, %rs26;
and.b32 %r535, %r534, 255;
shr.u32 %r536, %r535, %r4;
and.b32 %r537, %r536, 3;
bfi.b32 %r538, %r537, %r533, 4, 2;
add.s32 %r539, %r538, -32;
and.b32 %r401, %r528, 50529027;
shl.b32 %r540, %r529, 2;
and.b32 %r541, %r540, 67372036;
or.b32 %r542, %r401, -2139062144;
sub.s32 %r543, %r542, %r541;
xor.b32 %r544, %r543, -2139062144;
or.b32 %r545, %r401, %r541;
xor.b32 %r546, %r543, %r401;
and.b32 %r399, %r546, %r545;
mov.u32 %r183, %r399;

prmt.b32 %r183, %r183, 0, 0xba98;

mov.u32 %r185, %r401;

prmt.b32 %r185, %r185, 0, 0xba98;

xor.b32 %r547, %r185, 2139062143;
not.b32 %r548, %r183;
and.b32 %r549, %r544, %r548;
and.b32 %r550, %r547, %r183;
or.b32 %r188, %r550, %r549;

dp4a.s32.s32 %r187, %r188, %r153, %r430;

mul.lo.s32 %r551, %r539, %r187;
cvt.rn.f32.s32 %f97, %r551;
fma.rn.ftz.f32 %f98, %f55, %f97, 0f00000000;
add.s64 %rd89, %rd140, %rd11;
ld.global.nc.u8 %rs27, [%rd89];
cvt.u32.u16 %r552, %rs27;
and.b32 %r553, %r552, 255;
shr.u32 %r554, %r553, %r5;
and.b32 %r555, %r554, 15;
add.s64 %rd90, %rd140, %rd10;
ld.global.nc.u8 %rs28, [%rd90];
cvt.u32.u16 %r556, %rs28;
and.b32 %r557, %r556, 255;
shr.u32 %r558, %r557, %r6;
and.b32 %r559, %r558, 3;
bfi.b32 %r560, %r559, %r555, 4, 2;
add.s32 %r561, %r560, -32;
shr.u32 %r562, %r528, 2;
and.b32 %r409, %r562, 50529027;
shl.b32 %r563, %r529, 1;
and.b32 %r564, %r563, 67372036;
or.b32 %r565, %r409, -2139062144;
sub.s32 %r566, %r565, %r564;
xor.b32 %r567, %r566, -2139062144;
or.b32 %r568, %r409, %r564;
xor.b32 %r569, %r566, %r409;
and.b32 %r407, %r569, %r568;
mov.u32 %r191, %r407;

prmt.b32 %r191, %r191, 0, 0xba98;

mov.u32 %r193, %r409;

prmt.b32 %r193, %r193, 0, 0xba98;

xor.b32 %r570, %r193, 2139062143;
not.b32 %r571, %r191;
and.b32 %r572, %r567, %r571;
and.b32 %r573, %r570, %r191;
or.b32 %r196, %r573, %r572;

dp4a.s32.s32 %r195, %r196, %r161, %r430;

mul.lo.s32 %r574, %r561, %r195;
cvt.rn.f32.s32 %f99, %r574;
fma.rn.ftz.f32 %f100, %f56, %f99, %f98;
add.s64 %rd91, %rd140, %rd9;
ld.global.nc.u8 %rs29, [%rd91];
cvt.u32.u16 %r575, %rs29;
and.b32 %r576, %r575, 255;
shr.u32 %r577, %r576, %r7;
and.b32 %r578, %r577, 15;
add.s64 %rd92, %rd140, %rd8;
ld.global.nc.u8 %rs30, [%rd92];
cvt.u32.u16 %r579, %rs30;
and.b32 %r580, %r579, 255;
shr.u32 %r581, %r580, %r8;
and.b32 %r582, %r581, 3;
bfi.b32 %r583, %r582, %r578, 4, 2;
add.s32 %r584, %r583, -32;
shr.u32 %r585, %r528, 4;
and.b32 %r417, %r585, 50529027;
and.b32 %r586, %r529, 67372036;
or.b32 %r587, %r417, -2139062144;
sub.s32 %r588, %r587, %r586;
xor.b32 %r589, %r588, -2139062144;
or.b32 %r590, %r417, %r586;
xor.b32 %r591, %r588, %r417;
and.b32 %r415, %r591, %r590;
mov.u32 %r199, %r415;

prmt.b32 %r199, %r199, 0, 0xba98;

mov.u32 %r201, %r417;

prmt.b32 %r201, %r201, 0, 0xba98;

xor.b32 %r592, %r201, 2139062143;
not.b32 %r593, %r199;
and.b32 %r594, %r589, %r593;
and.b32 %r595, %r592, %r199;
or.b32 %r204, %r595, %r594;

dp4a.s32.s32 %r203, %r204, %r169, %r430;

mul.lo.s32 %r596, %r584, %r203;
cvt.rn.f32.s32 %f101, %r596;
fma.rn.ftz.f32 %f102, %f57, %f101, %f100;
add.s64 %rd93, %rd140, %rd7;
ld.global.nc.u8 %rs31, [%rd93];
cvt.u32.u16 %r597, %rs31;
and.b32 %r598, %r597, 255;
shr.u32 %r599, %r598, %r9;
and.b32 %r600, %r599, 15;
add.s64 %rd94, %rd140, %rd6;
ld.global.nc.u8 %rs32, [%rd94];
cvt.u32.u16 %r601, %rs32;
and.b32 %r602, %r601, 255;
shr.u32 %r603, %r602, %r10;
and.b32 %r604, %r603, 3;
bfi.b32 %r605, %r604, %r600, 4, 2;
add.s32 %r606, %r605, -32;
shr.u32 %r607, %r528, 6;
and.b32 %r425, %r607, 50529027;
shr.u32 %r608, %r529, 1;
and.b32 %r609, %r608, 67372036;
or.b32 %r610, %r425, -2139062144;
sub.s32 %r611, %r610, %r609;
xor.b32 %r612, %r611, -2139062144;
or.b32 %r613, %r425, %r609;
xor.b32 %r614, %r611, %r425;
and.b32 %r423, %r614, %r613;
mov.u32 %r207, %r423;

prmt.b32 %r207, %r207, 0, 0xba98;

mov.u32 %r209, %r425;

prmt.b32 %r209, %r209, 0, 0xba98;

xor.b32 %r615, %r209, 2139062143;
not.b32 %r616, %r207;
and.b32 %r617, %r612, %r616;
and.b32 %r618, %r615, %r207;
or.b32 %r212, %r618, %r617;

dp4a.s32.s32 %r211, %r212, %r177, %r430;

mul.lo.s32 %r619, %r606, %r211;
cvt.rn.f32.s32 %f103, %r619;
fma.rn.ftz.f32 %f104, %f58, %f103, %f102;
fma.rn.ftz.f32 %f287, %f54, %f104, %f287;
cvt.u64.u32 %rd95, %r917;
add.s64 %rd96, %rd1, %rd95;

{ cvt.f32.f16 %f59, %rs1;}


add.s64 %rd97, %rd96, %rd4;
mul.lo.s64 %rd98, %rd97, 36;
add.s64 %rd99, %rd65, %rd98;
add.s64 %rd100, %rd99, %rd74;
ld.global.nc.u32 %r225, [%rd100+4];
ld.global.nc.u32 %r215, [%rd99];

{.reg .f16 low,high;
mov.b32 {low,high},%r215;
cvt.f32.f16 %f60, low;}


ld.global.nc.u32 %r233, [%rd100+40];
ld.global.nc.u32 %r216, [%rd99+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r216;
cvt.f32.f16 %f61, low;}


ld.global.nc.u32 %r241, [%rd100+76];
ld.global.nc.u32 %r217, [%rd99+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r217;
cvt.f32.f16 %f62, low;}


ld.global.nc.u32 %r249, [%rd100+112];
ld.global.nc.u32 %r218, [%rd99+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r218;
cvt.f32.f16 %f63, low;}


mov.u32 %r219, %r363;

prmt.b32 %r219, %r219, 0, 0xba98;

mov.u32 %r221, %r365;

prmt.b32 %r221, %r221, 0, 0xba98;

xor.b32 %r620, %r221, 2139062143;
not.b32 %r621, %r219;
and.b32 %r622, %r450, %r621;
and.b32 %r623, %r620, %r219;
or.b32 %r224, %r623, %r622;

dp4a.s32.s32 %r223, %r224, %r225, %r430;

mul.lo.s32 %r624, %r445, %r223;
cvt.rn.f32.s32 %f105, %r624;
fma.rn.ftz.f32 %f106, %f60, %f105, 0f00000000;
mov.u32 %r227, %r371;

prmt.b32 %r227, %r227, 0, 0xba98;

mov.u32 %r229, %r373;

prmt.b32 %r229, %r229, 0, 0xba98;

xor.b32 %r625, %r229, 2139062143;
not.b32 %r626, %r227;
and.b32 %r627, %r473, %r626;
and.b32 %r628, %r625, %r227;
or.b32 %r232, %r628, %r627;

dp4a.s32.s32 %r231, %r232, %r233, %r430;

mul.lo.s32 %r629, %r467, %r231;
cvt.rn.f32.s32 %f107, %r629;
fma.rn.ftz.f32 %f108, %f61, %f107, %f106;
mov.u32 %r235, %r379;

prmt.b32 %r235, %r235, 0, 0xba98;

mov.u32 %r237, %r381;

prmt.b32 %r237, %r237, 0, 0xba98;

xor.b32 %r630, %r237, 2139062143;
not.b32 %r631, %r235;
and.b32 %r632, %r495, %r631;
and.b32 %r633, %r630, %r235;
or.b32 %r240, %r633, %r632;

dp4a.s32.s32 %r239, %r240, %r241, %r430;

mul.lo.s32 %r634, %r490, %r239;
cvt.rn.f32.s32 %f109, %r634;
fma.rn.ftz.f32 %f110, %f62, %f109, %f108;
mov.u32 %r243, %r387;

prmt.b32 %r243, %r243, 0, 0xba98;

mov.u32 %r245, %r389;

prmt.b32 %r245, %r245, 0, 0xba98;

xor.b32 %r635, %r245, 2139062143;
not.b32 %r636, %r243;
and.b32 %r637, %r518, %r636;
and.b32 %r638, %r635, %r243;
or.b32 %r248, %r638, %r637;

dp4a.s32.s32 %r247, %r248, %r249, %r430;

mul.lo.s32 %r639, %r512, %r247;
cvt.rn.f32.s32 %f111, %r639;
fma.rn.ftz.f32 %f112, %f63, %f111, %f110;
fma.rn.ftz.f32 %f286, %f59, %f112, %f286;

{ cvt.f32.f16 %f64, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r215;
cvt.f32.f16 %f65, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r216;
cvt.f32.f16 %f66, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r217;
cvt.f32.f16 %f67, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r218;
cvt.f32.f16 %f68, low;}


mov.u32 %r255, %r399;

prmt.b32 %r255, %r255, 0, 0xba98;

mov.u32 %r257, %r401;

prmt.b32 %r257, %r257, 0, 0xba98;

xor.b32 %r640, %r257, 2139062143;
not.b32 %r641, %r255;
and.b32 %r642, %r544, %r641;
and.b32 %r643, %r640, %r255;
or.b32 %r260, %r643, %r642;

dp4a.s32.s32 %r259, %r260, %r225, %r430;

mul.lo.s32 %r644, %r539, %r259;
cvt.rn.f32.s32 %f113, %r644;
fma.rn.ftz.f32 %f114, %f65, %f113, 0f00000000;
mov.u32 %r263, %r407;

prmt.b32 %r263, %r263, 0, 0xba98;

mov.u32 %r265, %r409;

prmt.b32 %r265, %r265, 0, 0xba98;

xor.b32 %r645, %r265, 2139062143;
not.b32 %r646, %r263;
and.b32 %r647, %r567, %r646;
and.b32 %r648, %r645, %r263;
or.b32 %r268, %r648, %r647;

dp4a.s32.s32 %r267, %r268, %r233, %r430;

mul.lo.s32 %r649, %r561, %r267;
cvt.rn.f32.s32 %f115, %r649;
fma.rn.ftz.f32 %f116, %f66, %f115, %f114;
mov.u32 %r271, %r415;

prmt.b32 %r271, %r271, 0, 0xba98;

mov.u32 %r273, %r417;

prmt.b32 %r273, %r273, 0, 0xba98;

xor.b32 %r650, %r273, 2139062143;
not.b32 %r651, %r271;
and.b32 %r652, %r589, %r651;
and.b32 %r653, %r650, %r271;
or.b32 %r276, %r653, %r652;

dp4a.s32.s32 %r275, %r276, %r241, %r430;

mul.lo.s32 %r654, %r584, %r275;
cvt.rn.f32.s32 %f117, %r654;
fma.rn.ftz.f32 %f118, %f67, %f117, %f116;
mov.u32 %r279, %r423;

prmt.b32 %r279, %r279, 0, 0xba98;

mov.u32 %r281, %r425;

prmt.b32 %r281, %r281, 0, 0xba98;

xor.b32 %r655, %r281, 2139062143;
not.b32 %r656, %r279;
and.b32 %r657, %r612, %r656;
and.b32 %r658, %r655, %r279;
or.b32 %r284, %r658, %r657;

dp4a.s32.s32 %r283, %r284, %r249, %r430;

mul.lo.s32 %r659, %r606, %r283;
cvt.rn.f32.s32 %f119, %r659;
fma.rn.ftz.f32 %f120, %f68, %f119, %f118;
fma.rn.ftz.f32 %f285, %f64, %f120, %f285;
cvt.u64.u32 %rd101, %r919;
add.s64 %rd102, %rd1, %rd101;

{ cvt.f32.f16 %f69, %rs1;}


add.s64 %rd103, %rd102, %rd4;
mul.lo.s64 %rd104, %rd103, 36;
add.s64 %rd105, %rd65, %rd104;
add.s64 %rd106, %rd105, %rd74;
ld.global.nc.u32 %r297, [%rd106+4];
ld.global.nc.u32 %r287, [%rd105];

{.reg .f16 low,high;
mov.b32 {low,high},%r287;
cvt.f32.f16 %f70, low;}


ld.global.nc.u32 %r305, [%rd106+40];
ld.global.nc.u32 %r288, [%rd105+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r288;
cvt.f32.f16 %f71, low;}


ld.global.nc.u32 %r313, [%rd106+76];
ld.global.nc.u32 %r289, [%rd105+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r289;
cvt.f32.f16 %f72, low;}


ld.global.nc.u32 %r321, [%rd106+112];
ld.global.nc.u32 %r290, [%rd105+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r290;
cvt.f32.f16 %f73, low;}


mov.u32 %r291, %r363;

prmt.b32 %r291, %r291, 0, 0xba98;

mov.u32 %r293, %r365;

prmt.b32 %r293, %r293, 0, 0xba98;

xor.b32 %r660, %r293, 2139062143;
not.b32 %r661, %r291;
and.b32 %r662, %r450, %r661;
and.b32 %r663, %r660, %r291;
or.b32 %r296, %r663, %r662;

dp4a.s32.s32 %r295, %r296, %r297, %r430;

mul.lo.s32 %r664, %r445, %r295;
cvt.rn.f32.s32 %f121, %r664;
fma.rn.ftz.f32 %f122, %f70, %f121, 0f00000000;
mov.u32 %r299, %r371;

prmt.b32 %r299, %r299, 0, 0xba98;

mov.u32 %r301, %r373;

prmt.b32 %r301, %r301, 0, 0xba98;

xor.b32 %r665, %r301, 2139062143;
not.b32 %r666, %r299;
and.b32 %r667, %r473, %r666;
and.b32 %r668, %r665, %r299;
or.b32 %r304, %r668, %r667;

dp4a.s32.s32 %r303, %r304, %r305, %r430;

mul.lo.s32 %r669, %r467, %r303;
cvt.rn.f32.s32 %f123, %r669;
fma.rn.ftz.f32 %f124, %f71, %f123, %f122;
mov.u32 %r307, %r379;

prmt.b32 %r307, %r307, 0, 0xba98;

mov.u32 %r309, %r381;

prmt.b32 %r309, %r309, 0, 0xba98;

xor.b32 %r670, %r309, 2139062143;
not.b32 %r671, %r307;
and.b32 %r672, %r495, %r671;
and.b32 %r673, %r670, %r307;
or.b32 %r312, %r673, %r672;

dp4a.s32.s32 %r311, %r312, %r313, %r430;

mul.lo.s32 %r674, %r490, %r311;
cvt.rn.f32.s32 %f125, %r674;
fma.rn.ftz.f32 %f126, %f72, %f125, %f124;
mov.u32 %r315, %r387;

prmt.b32 %r315, %r315, 0, 0xba98;

mov.u32 %r317, %r389;

prmt.b32 %r317, %r317, 0, 0xba98;

xor.b32 %r675, %r317, 2139062143;
not.b32 %r676, %r315;
and.b32 %r677, %r518, %r676;
and.b32 %r678, %r675, %r315;
or.b32 %r320, %r678, %r677;

dp4a.s32.s32 %r319, %r320, %r321, %r430;

mul.lo.s32 %r679, %r512, %r319;
cvt.rn.f32.s32 %f127, %r679;
fma.rn.ftz.f32 %f128, %f73, %f127, %f126;
fma.rn.ftz.f32 %f284, %f69, %f128, %f284;

{ cvt.f32.f16 %f74, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r287;
cvt.f32.f16 %f75, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r288;
cvt.f32.f16 %f76, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r289;
cvt.f32.f16 %f77, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r290;
cvt.f32.f16 %f78, low;}


mov.u32 %r327, %r399;

prmt.b32 %r327, %r327, 0, 0xba98;

mov.u32 %r329, %r401;

prmt.b32 %r329, %r329, 0, 0xba98;

xor.b32 %r680, %r329, 2139062143;
not.b32 %r681, %r327;
and.b32 %r682, %r544, %r681;
and.b32 %r683, %r680, %r327;
or.b32 %r332, %r683, %r682;

dp4a.s32.s32 %r331, %r332, %r297, %r430;

mul.lo.s32 %r684, %r539, %r331;
cvt.rn.f32.s32 %f129, %r684;
fma.rn.ftz.f32 %f130, %f75, %f129, 0f00000000;
mov.u32 %r335, %r407;

prmt.b32 %r335, %r335, 0, 0xba98;

mov.u32 %r337, %r409;

prmt.b32 %r337, %r337, 0, 0xba98;

xor.b32 %r685, %r337, 2139062143;
not.b32 %r686, %r335;
and.b32 %r687, %r567, %r686;
and.b32 %r688, %r685, %r335;
or.b32 %r340, %r688, %r687;

dp4a.s32.s32 %r339, %r340, %r305, %r430;

mul.lo.s32 %r689, %r561, %r339;
cvt.rn.f32.s32 %f131, %r689;
fma.rn.ftz.f32 %f132, %f76, %f131, %f130;
mov.u32 %r343, %r415;

prmt.b32 %r343, %r343, 0, 0xba98;

mov.u32 %r345, %r417;

prmt.b32 %r345, %r345, 0, 0xba98;

xor.b32 %r690, %r345, 2139062143;
not.b32 %r691, %r343;
and.b32 %r692, %r589, %r691;
and.b32 %r693, %r690, %r343;
or.b32 %r348, %r693, %r692;

dp4a.s32.s32 %r347, %r348, %r313, %r430;

mul.lo.s32 %r694, %r584, %r347;
cvt.rn.f32.s32 %f133, %r694;
fma.rn.ftz.f32 %f134, %f77, %f133, %f132;
mov.u32 %r351, %r423;

prmt.b32 %r351, %r351, 0, 0xba98;

mov.u32 %r353, %r425;

prmt.b32 %r353, %r353, 0, 0xba98;

xor.b32 %r695, %r353, 2139062143;
not.b32 %r696, %r351;
and.b32 %r697, %r612, %r696;
and.b32 %r698, %r695, %r351;
or.b32 %r356, %r698, %r697;

dp4a.s32.s32 %r355, %r356, %r321, %r430;

mul.lo.s32 %r699, %r606, %r355;
cvt.rn.f32.s32 %f135, %r699;
fma.rn.ftz.f32 %f136, %f78, %f135, %f134;
fma.rn.ftz.f32 %f283, %f74, %f136, %f283;
cvt.u64.u32 %rd107, %r918;
add.s64 %rd108, %rd1, %rd107;

{ cvt.f32.f16 %f79, %rs1;}


add.s64 %rd109, %rd108, %rd4;
mul.lo.s64 %rd110, %rd109, 36;
add.s64 %rd111, %rd65, %rd110;
add.s64 %rd112, %rd111, %rd74;
ld.global.nc.u32 %r369, [%rd112+4];
ld.global.nc.u32 %r359, [%rd111];

{.reg .f16 low,high;
mov.b32 {low,high},%r359;
cvt.f32.f16 %f80, low;}


ld.global.nc.u32 %r377, [%rd112+40];
ld.global.nc.u32 %r360, [%rd111+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r360;
cvt.f32.f16 %f81, low;}


ld.global.nc.u32 %r385, [%rd112+76];
ld.global.nc.u32 %r361, [%rd111+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r361;
cvt.f32.f16 %f82, low;}


ld.global.nc.u32 %r393, [%rd112+112];
ld.global.nc.u32 %r362, [%rd111+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r362;
cvt.f32.f16 %f83, low;}



prmt.b32 %r363, %r363, 0, 0xba98;


prmt.b32 %r365, %r365, 0, 0xba98;

xor.b32 %r700, %r365, 2139062143;
not.b32 %r701, %r363;
and.b32 %r702, %r450, %r701;
and.b32 %r703, %r700, %r363;
or.b32 %r368, %r703, %r702;

dp4a.s32.s32 %r367, %r368, %r369, %r430;

mul.lo.s32 %r704, %r445, %r367;
cvt.rn.f32.s32 %f137, %r704;
fma.rn.ftz.f32 %f138, %f80, %f137, 0f00000000;

prmt.b32 %r371, %r371, 0, 0xba98;


prmt.b32 %r373, %r373, 0, 0xba98;

xor.b32 %r705, %r373, 2139062143;
not.b32 %r706, %r371;
and.b32 %r707, %r473, %r706;
and.b32 %r708, %r705, %r371;
or.b32 %r376, %r708, %r707;

dp4a.s32.s32 %r375, %r376, %r377, %r430;

mul.lo.s32 %r709, %r467, %r375;
cvt.rn.f32.s32 %f139, %r709;
fma.rn.ftz.f32 %f140, %f81, %f139, %f138;

prmt.b32 %r379, %r379, 0, 0xba98;


prmt.b32 %r381, %r381, 0, 0xba98;

xor.b32 %r710, %r381, 2139062143;
not.b32 %r711, %r379;
and.b32 %r712, %r495, %r711;
and.b32 %r713, %r710, %r379;
or.b32 %r384, %r713, %r712;

dp4a.s32.s32 %r383, %r384, %r385, %r430;

mul.lo.s32 %r714, %r490, %r383;
cvt.rn.f32.s32 %f141, %r714;
fma.rn.ftz.f32 %f142, %f82, %f141, %f140;

prmt.b32 %r387, %r387, 0, 0xba98;


prmt.b32 %r389, %r389, 0, 0xba98;

xor.b32 %r715, %r389, 2139062143;
not.b32 %r716, %r387;
and.b32 %r717, %r518, %r716;
and.b32 %r718, %r715, %r387;
or.b32 %r392, %r718, %r717;

dp4a.s32.s32 %r391, %r392, %r393, %r430;

mul.lo.s32 %r719, %r512, %r391;
cvt.rn.f32.s32 %f143, %r719;
fma.rn.ftz.f32 %f144, %f83, %f143, %f142;
fma.rn.ftz.f32 %f282, %f79, %f144, %f282;

{ cvt.f32.f16 %f84, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r359;
cvt.f32.f16 %f85, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r360;
cvt.f32.f16 %f86, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r361;
cvt.f32.f16 %f87, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r362;
cvt.f32.f16 %f88, low;}



prmt.b32 %r399, %r399, 0, 0xba98;


prmt.b32 %r401, %r401, 0, 0xba98;

xor.b32 %r720, %r401, 2139062143;
not.b32 %r721, %r399;
and.b32 %r722, %r544, %r721;
and.b32 %r723, %r720, %r399;
or.b32 %r404, %r723, %r722;

dp4a.s32.s32 %r403, %r404, %r369, %r430;

mul.lo.s32 %r724, %r539, %r403;
cvt.rn.f32.s32 %f145, %r724;
fma.rn.ftz.f32 %f146, %f85, %f145, 0f00000000;

prmt.b32 %r407, %r407, 0, 0xba98;


prmt.b32 %r409, %r409, 0, 0xba98;

xor.b32 %r725, %r409, 2139062143;
not.b32 %r726, %r407;
and.b32 %r727, %r567, %r726;
and.b32 %r728, %r725, %r407;
or.b32 %r412, %r728, %r727;

dp4a.s32.s32 %r411, %r412, %r377, %r430;

mul.lo.s32 %r729, %r561, %r411;
cvt.rn.f32.s32 %f147, %r729;
fma.rn.ftz.f32 %f148, %f86, %f147, %f146;

prmt.b32 %r415, %r415, 0, 0xba98;


prmt.b32 %r417, %r417, 0, 0xba98;

xor.b32 %r730, %r417, 2139062143;
not.b32 %r731, %r415;
and.b32 %r732, %r589, %r731;
and.b32 %r733, %r730, %r415;
or.b32 %r420, %r733, %r732;

dp4a.s32.s32 %r419, %r420, %r385, %r430;

mul.lo.s32 %r734, %r584, %r419;
cvt.rn.f32.s32 %f149, %r734;
fma.rn.ftz.f32 %f150, %f87, %f149, %f148;

prmt.b32 %r423, %r423, 0, 0xba98;


prmt.b32 %r425, %r425, 0, 0xba98;

xor.b32 %r735, %r425, 2139062143;
not.b32 %r736, %r423;
and.b32 %r737, %r612, %r736;
and.b32 %r738, %r735, %r423;
or.b32 %r428, %r738, %r737;

dp4a.s32.s32 %r427, %r428, %r393, %r430;

mul.lo.s32 %r739, %r606, %r427;
cvt.rn.f32.s32 %f151, %r739;
fma.rn.ftz.f32 %f152, %f88, %f151, %f150;
fma.rn.ftz.f32 %f281, %f84, %f152, %f281;
add.s32 %r919, %r919, 64;
add.s32 %r918, %r918, 64;
add.s32 %r917, %r917, 64;
add.s32 %r916, %r916, 64;
add.s64 %rd140, %rd140, 880;
add.s32 %r920, %r920, 8;
setp.lt.s32 %p4, %r920, %r50;
@%p4 bra $L__BB148_2;

$L__BB148_3:
mov.u32 %r902, %tid.y;
setp.eq.s32 %p5, %r902, 0;
@%p5 bra $L__BB148_5;

mov.u32 %r904, %tid.x;
mov.u32 %r903, %tid.y;
shl.b32 %r744, %r903, 10;
mov.u32 %r745, _ZZ13mul_mat_vec_qIL9ggml_type11ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r746, %r745, %r744;
shl.b32 %r747, %r904, 2;
add.s32 %r748, %r746, %r747;
st.shared.f32 [%r748+-1024], %f288;
st.shared.f32 [%r748+-896], %f287;
st.shared.f32 [%r748+-768], %f286;
st.shared.f32 [%r748+-640], %f285;
st.shared.f32 [%r748+-512], %f284;
st.shared.f32 [%r748+-384], %f283;
st.shared.f32 [%r748+-256], %f282;
st.shared.f32 [%r748+-128], %f281;

$L__BB148_5:
mov.u32 %r905, %tid.y;
setp.ne.s32 %p6, %r905, 0;
bar.sync 0;
@%p6 bra $L__BB148_22;

ld.param.u32 %r910, [_Z13mul_mat_vec_qIL9ggml_type11ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u64 %rd139, [_Z13mul_mat_vec_qIL9ggml_type11ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r908, [_Z13mul_mat_vec_qIL9ggml_type11ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r907, [_Z13mul_mat_vec_qIL9ggml_type11ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
mov.u32 %r906, %tid.x;
mov.u32 %r750, %ctaid.z;
mov.u32 %r751, %ctaid.y;
mov.u32 %r752, %ctaid.x;
shl.b32 %r753, %r752, 1;
mov.u32 %r754, 1;
mad.lo.s32 %r755, %r751, %r907, %r753;
mad.lo.s32 %r756, %r750, %r908, %r755;
cvt.u64.u32 %rd29, %r756;
shl.b32 %r758, %r906, 2;
mov.u32 %r759, _ZZ13mul_mat_vec_qIL9ggml_type11ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r25, %r759, %r758;
mov.u32 %r760, 2;
ld.shared.f32 %f153, [%r25];
add.ftz.f32 %f154, %f153, %f288;
ld.shared.f32 %f155, [%r25+1024];
add.ftz.f32 %f156, %f155, %f154;
ld.shared.f32 %f157, [%r25+2048];
add.ftz.f32 %f158, %f157, %f156;
mov.b32 %r761, %f158;
mov.u32 %r762, 31;
mov.u32 %r763, 16;
mov.u32 %r764, -1;
shfl.sync.bfly.b32 %r765|%p7, %r761, %r763, %r762, %r764;
mov.b32 %f159, %r765;
add.ftz.f32 %f160, %f158, %f159;
mov.b32 %r766, %f160;
mov.u32 %r767, 8;
shfl.sync.bfly.b32 %r768|%p8, %r766, %r767, %r762, %r764;
mov.b32 %f161, %r768;
add.ftz.f32 %f162, %f160, %f161;
mov.b32 %r769, %f162;
mov.u32 %r770, 4;
shfl.sync.bfly.b32 %r771|%p9, %r769, %r770, %r762, %r764;
mov.b32 %f163, %r771;
add.ftz.f32 %f164, %f162, %f163;
mov.b32 %r772, %f164;
shfl.sync.bfly.b32 %r773|%p10, %r772, %r760, %r762, %r764;
mov.b32 %f165, %r773;
add.ftz.f32 %f166, %f164, %f165;
mov.b32 %r774, %f166;
shfl.sync.bfly.b32 %r775|%p11, %r774, %r754, %r762, %r764;
mov.b32 %f167, %r775;
add.ftz.f32 %f25, %f166, %f167;
setp.eq.s32 %p12, %r906, 0;
setp.lt.u32 %p13, %r753, %r910;
and.pred %p1, %p12, %p13;
cvta.to.global.u64 %rd113, %rd139;
mul.wide.u32 %rd114, %r756, 4;
add.s64 %rd30, %rd113, %rd114;
not.pred %p14, %p1;
@%p14 bra $L__BB148_8;

st.global.f32 [%rd30], %f25;

$L__BB148_8:
ld.param.u32 %r911, [_Z13mul_mat_vec_qIL9ggml_type11ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
mov.u32 %r909, %tid.x;
ld.shared.f32 %f168, [%r25+128];
add.ftz.f32 %f169, %f168, %f287;
ld.shared.f32 %f170, [%r25+1152];
add.ftz.f32 %f171, %f170, %f169;
ld.shared.f32 %f172, [%r25+2176];
add.ftz.f32 %f173, %f172, %f171;
mov.b32 %r776, %f173;
shfl.sync.bfly.b32 %r780|%p15, %r776, %r763, %r762, %r764;
mov.b32 %f174, %r780;
add.ftz.f32 %f175, %f173, %f174;
mov.b32 %r781, %f175;
shfl.sync.bfly.b32 %r783|%p16, %r781, %r767, %r762, %r764;
mov.b32 %f176, %r783;
add.ftz.f32 %f177, %f175, %f176;
mov.b32 %r784, %f177;
shfl.sync.bfly.b32 %r786|%p17, %r784, %r770, %r762, %r764;
mov.b32 %f178, %r786;
add.ftz.f32 %f179, %f177, %f178;
mov.b32 %r787, %f179;
shfl.sync.bfly.b32 %r789|%p18, %r787, %r760, %r762, %r764;
mov.b32 %f180, %r789;
add.ftz.f32 %f181, %f179, %f180;
mov.b32 %r790, %f181;
shfl.sync.bfly.b32 %r792|%p19, %r790, %r754, %r762, %r764;
mov.b32 %f182, %r792;
add.ftz.f32 %f26, %f181, %f182;
setp.eq.s32 %p20, %r909, 1;
or.b32 %r796, %r753, 1;
setp.lt.u32 %p21, %r796, %r911;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB148_10;

st.global.f32 [%rd30+4], %f26;

$L__BB148_10:
ld.shared.f32 %f183, [%r25+256];
add.ftz.f32 %f184, %f183, %f286;
ld.shared.f32 %f185, [%r25+1280];
add.ftz.f32 %f186, %f185, %f184;
ld.shared.f32 %f187, [%r25+2304];
add.ftz.f32 %f188, %f187, %f186;
mov.b32 %r797, %f188;
mov.u32 %r798, 31;
mov.u32 %r799, 16;
mov.u32 %r800, -1;
shfl.sync.bfly.b32 %r801|%p23, %r797, %r799, %r798, %r800;
mov.b32 %f189, %r801;
add.ftz.f32 %f190, %f188, %f189;
mov.b32 %r802, %f190;
mov.u32 %r803, 8;
shfl.sync.bfly.b32 %r804|%p24, %r802, %r803, %r798, %r800;
mov.b32 %f191, %r804;
add.ftz.f32 %f192, %f190, %f191;
mov.b32 %r805, %f192;
mov.u32 %r806, 4;
shfl.sync.bfly.b32 %r807|%p25, %r805, %r806, %r798, %r800;
mov.b32 %f193, %r807;
add.ftz.f32 %f194, %f192, %f193;
mov.b32 %r808, %f194;
mov.u32 %r809, 2;
shfl.sync.bfly.b32 %r810|%p26, %r808, %r809, %r798, %r800;
mov.b32 %f195, %r810;
add.ftz.f32 %f196, %f194, %f195;
mov.b32 %r811, %f196;
mov.u32 %r812, 1;
shfl.sync.bfly.b32 %r813|%p27, %r811, %r812, %r798, %r800;
mov.b32 %f197, %r813;
add.ftz.f32 %f27, %f196, %f197;
@%p14 bra $L__BB148_12;

ld.param.u32 %r915, [_Z13mul_mat_vec_qIL9ggml_type11ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
cvt.u64.u32 %rd115, %r915;
add.s64 %rd116, %rd115, %rd29;
shl.b64 %rd118, %rd116, 2;
add.s64 %rd119, %rd113, %rd118;
st.global.f32 [%rd119], %f27;

$L__BB148_12:
ld.shared.f32 %f198, [%r25+384];
add.ftz.f32 %f199, %f198, %f285;
ld.shared.f32 %f200, [%r25+1408];
add.ftz.f32 %f201, %f200, %f199;
ld.shared.f32 %f202, [%r25+2432];
add.ftz.f32 %f203, %f202, %f201;
mov.b32 %r814, %f203;
shfl.sync.bfly.b32 %r818|%p29, %r814, %r799, %r798, %r800;
mov.b32 %f204, %r818;
add.ftz.f32 %f205, %f203, %f204;
mov.b32 %r819, %f205;
shfl.sync.bfly.b32 %r821|%p30, %r819, %r803, %r798, %r800;
mov.b32 %f206, %r821;
add.ftz.f32 %f207, %f205, %f206;
mov.b32 %r822, %f207;
shfl.sync.bfly.b32 %r824|%p31, %r822, %r806, %r798, %r800;
mov.b32 %f208, %r824;
add.ftz.f32 %f209, %f207, %f208;
mov.b32 %r825, %f209;
shfl.sync.bfly.b32 %r827|%p32, %r825, %r809, %r798, %r800;
mov.b32 %f210, %r827;
add.ftz.f32 %f211, %f209, %f210;
mov.b32 %r828, %f211;
shfl.sync.bfly.b32 %r830|%p33, %r828, %r812, %r798, %r800;
mov.b32 %f212, %r830;
add.ftz.f32 %f28, %f211, %f212;
@%p22 bra $L__BB148_14;

ld.param.u32 %r914, [_Z13mul_mat_vec_qIL9ggml_type11ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
add.s32 %r831, %r914, 1;
cvt.u64.u32 %rd120, %r831;
add.s64 %rd121, %rd120, %rd29;
shl.b64 %rd123, %rd121, 2;
add.s64 %rd124, %rd113, %rd123;
st.global.f32 [%rd124], %f28;

$L__BB148_14:
ld.param.u32 %r912, [_Z13mul_mat_vec_qIL9ggml_type11ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
shl.b32 %r832, %r912, 1;
mov.u32 %r833, 1;
ld.shared.f32 %f213, [%r25+512];
add.ftz.f32 %f214, %f213, %f284;
ld.shared.f32 %f215, [%r25+1536];
add.ftz.f32 %f216, %f215, %f214;
ld.shared.f32 %f217, [%r25+2560];
add.ftz.f32 %f218, %f217, %f216;
mov.b32 %r834, %f218;
mov.u32 %r835, 31;
mov.u32 %r836, 16;
mov.u32 %r837, -1;
shfl.sync.bfly.b32 %r838|%p35, %r834, %r836, %r835, %r837;
mov.b32 %f219, %r838;
add.ftz.f32 %f220, %f218, %f219;
mov.b32 %r839, %f220;
mov.u32 %r840, 8;
shfl.sync.bfly.b32 %r841|%p36, %r839, %r840, %r835, %r837;
mov.b32 %f221, %r841;
add.ftz.f32 %f222, %f220, %f221;
mov.b32 %r842, %f222;
mov.u32 %r843, 4;
shfl.sync.bfly.b32 %r844|%p37, %r842, %r843, %r835, %r837;
mov.b32 %f223, %r844;
add.ftz.f32 %f224, %f222, %f223;
mov.b32 %r845, %f224;
mov.u32 %r846, 2;
shfl.sync.bfly.b32 %r847|%p38, %r845, %r846, %r835, %r837;
mov.b32 %f225, %r847;
add.ftz.f32 %f226, %f224, %f225;
mov.b32 %r848, %f226;
shfl.sync.bfly.b32 %r849|%p39, %r848, %r833, %r835, %r837;
mov.b32 %f227, %r849;
add.ftz.f32 %f29, %f226, %f227;
cvt.u64.u32 %rd125, %r832;
add.s64 %rd126, %rd125, %rd29;
shl.b64 %rd128, %rd126, 2;
add.s64 %rd31, %rd113, %rd128;
@%p14 bra $L__BB148_16;

st.global.f32 [%rd31], %f29;

$L__BB148_16:
ld.shared.f32 %f228, [%r25+640];
add.ftz.f32 %f229, %f228, %f283;
ld.shared.f32 %f230, [%r25+1664];
add.ftz.f32 %f231, %f230, %f229;
ld.shared.f32 %f232, [%r25+2688];
add.ftz.f32 %f233, %f232, %f231;
mov.b32 %r850, %f233;
shfl.sync.bfly.b32 %r854|%p41, %r850, %r836, %r835, %r837;
mov.b32 %f234, %r854;
add.ftz.f32 %f235, %f233, %f234;
mov.b32 %r855, %f235;
shfl.sync.bfly.b32 %r857|%p42, %r855, %r840, %r835, %r837;
mov.b32 %f236, %r857;
add.ftz.f32 %f237, %f235, %f236;
mov.b32 %r858, %f237;
shfl.sync.bfly.b32 %r860|%p43, %r858, %r843, %r835, %r837;
mov.b32 %f238, %r860;
add.ftz.f32 %f239, %f237, %f238;
mov.b32 %r861, %f239;
shfl.sync.bfly.b32 %r863|%p44, %r861, %r846, %r835, %r837;
mov.b32 %f240, %r863;
add.ftz.f32 %f241, %f239, %f240;
mov.b32 %r864, %f241;
shfl.sync.bfly.b32 %r866|%p45, %r864, %r833, %r835, %r837;
mov.b32 %f242, %r866;
add.ftz.f32 %f30, %f241, %f242;
@%p22 bra $L__BB148_18;

st.global.f32 [%rd31+4], %f30;

$L__BB148_18:
ld.param.u32 %r913, [_Z13mul_mat_vec_qIL9ggml_type11ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
mul.lo.s32 %r27, %r913, 3;
ld.shared.f32 %f243, [%r25+768];
add.ftz.f32 %f244, %f243, %f282;
ld.shared.f32 %f245, [%r25+1792];
add.ftz.f32 %f246, %f245, %f244;
ld.shared.f32 %f247, [%r25+2816];
add.ftz.f32 %f248, %f247, %f246;
mov.b32 %r867, %f248;
mov.u32 %r868, 31;
mov.u32 %r869, 16;
mov.u32 %r870, -1;
shfl.sync.bfly.b32 %r871|%p47, %r867, %r869, %r868, %r870;
mov.b32 %f249, %r871;
add.ftz.f32 %f250, %f248, %f249;
mov.b32 %r872, %f250;
mov.u32 %r873, 8;
shfl.sync.bfly.b32 %r874|%p48, %r872, %r873, %r868, %r870;
mov.b32 %f251, %r874;
add.ftz.f32 %f252, %f250, %f251;
mov.b32 %r875, %f252;
mov.u32 %r876, 4;
shfl.sync.bfly.b32 %r877|%p49, %r875, %r876, %r868, %r870;
mov.b32 %f253, %r877;
add.ftz.f32 %f254, %f252, %f253;
mov.b32 %r878, %f254;
mov.u32 %r879, 2;
shfl.sync.bfly.b32 %r880|%p50, %r878, %r879, %r868, %r870;
mov.b32 %f255, %r880;
add.ftz.f32 %f256, %f254, %f255;
mov.b32 %r881, %f256;
mov.u32 %r882, 1;
shfl.sync.bfly.b32 %r883|%p51, %r881, %r882, %r868, %r870;
mov.b32 %f257, %r883;
add.ftz.f32 %f31, %f256, %f257;
@%p14 bra $L__BB148_20;

cvt.u64.u32 %rd129, %r27;
add.s64 %rd130, %rd129, %rd29;
shl.b64 %rd132, %rd130, 2;
add.s64 %rd133, %rd113, %rd132;
st.global.f32 [%rd133], %f31;

$L__BB148_20:
ld.shared.f32 %f258, [%r25+896];
add.ftz.f32 %f259, %f258, %f281;
ld.shared.f32 %f260, [%r25+1920];
add.ftz.f32 %f261, %f260, %f259;
ld.shared.f32 %f262, [%r25+2944];
add.ftz.f32 %f263, %f262, %f261;
mov.b32 %r884, %f263;
shfl.sync.bfly.b32 %r888|%p53, %r884, %r869, %r868, %r870;
mov.b32 %f264, %r888;
add.ftz.f32 %f265, %f263, %f264;
mov.b32 %r889, %f265;
shfl.sync.bfly.b32 %r891|%p54, %r889, %r873, %r868, %r870;
mov.b32 %f266, %r891;
add.ftz.f32 %f267, %f265, %f266;
mov.b32 %r892, %f267;
shfl.sync.bfly.b32 %r894|%p55, %r892, %r876, %r868, %r870;
mov.b32 %f268, %r894;
add.ftz.f32 %f269, %f267, %f268;
mov.b32 %r895, %f269;
shfl.sync.bfly.b32 %r897|%p56, %r895, %r879, %r868, %r870;
mov.b32 %f270, %r897;
add.ftz.f32 %f271, %f269, %f270;
mov.b32 %r898, %f271;
shfl.sync.bfly.b32 %r900|%p57, %r898, %r882, %r868, %r870;
mov.b32 %f272, %r900;
add.ftz.f32 %f32, %f271, %f272;
@%p22 bra $L__BB148_22;

add.s32 %r901, %r27, 1;
cvt.u64.u32 %rd134, %r901;
add.s64 %rd135, %rd134, %rd29;
shl.b64 %rd137, %rd135, 2;
add.s64 %rd138, %rd113, %rd137;
st.global.f32 [%rd138], %f32;

$L__BB148_22:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type11ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type11ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type11ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type11ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type11ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type11ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type11ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type11ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type11ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<71>;
.reg .b16 %rs<35>;
.reg .f32 %f<321>;
.reg .b32 %r<1068>;
.reg .b64 %rd<145>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type11ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1280];

ld.param.u64 %rd35, [_Z13mul_mat_vec_qIL9ggml_type11ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd36, [_Z13mul_mat_vec_qIL9ggml_type11ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd37, [_Z13mul_mat_vec_qIL9ggml_type11ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r47, [_Z13mul_mat_vec_qIL9ggml_type11ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type11ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type11ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type11ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type11ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type11ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r45, [_Z13mul_mat_vec_qIL9ggml_type11ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type11ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type11ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type11ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type11ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
cvta.to.global.u64 %rd1, %rd37;
mov.u32 %r1, %tid.y;
shl.b32 %r48, %r1, 5;
mov.u32 %r49, %tid.x;
add.s32 %r2, %r48, %r49;
shr.s32 %r50, %r2, 31;
shr.u32 %r51, %r50, 28;
add.s32 %r52, %r2, %r51;
shr.s32 %r1067, %r52, 4;
shr.u32 %r4, %r47, 8;
setp.ge.s32 %p3, %r1067, %r4;
mov.f32 %f311, 0f00000000;
mov.f32 %f312, %f311;
mov.f32 %f313, %f311;
mov.f32 %f314, %f311;
mov.f32 %f315, %f311;
mov.f32 %f316, %f311;
mov.f32 %f317, %f311;
mov.f32 %f318, %f311;
mov.f32 %f319, %f311;
mov.f32 %f320, %f311;
@%p3 bra $L__BB149_3;

mov.u32 %r53, %ctaid.x;
shl.b32 %r54, %r53, 1;
mov.u32 %r55, %ctaid.y;
mov.u32 %r56, %ctaid.z;
mul.lo.s32 %r57, %r55, %r39;
cvt.u64.u32 %rd38, %r57;
mul.lo.s32 %r58, %r56, %r45;
cvt.u64.u32 %rd39, %r58;
add.s64 %rd2, %rd39, %rd38;
mul.hi.u32 %r59, %r56, %r41;
add.s32 %r60, %r59, %r56;
shr.u32 %r61, %r60, %r42;
mul.lo.s32 %r62, %r61, %r44;
mul.hi.u32 %r63, %r55, %r35;
add.s32 %r64, %r63, %r55;
shr.u32 %r65, %r64, %r36;
mul.lo.s32 %r66, %r65, %r38;
cvta.to.global.u64 %rd144, %rd35;
or.b32 %r67, %r54, 1;
and.b32 %r71, %r52, -16;
sub.s32 %r72, %r2, %r71;
shr.s32 %r73, %r72, 31;
shr.u32 %r74, %r73, 29;
add.s32 %r75, %r72, %r74;
and.b32 %r76, %r75, -8;
sub.s32 %r77, %r72, %r76;
shr.s32 %r78, %r77, 31;
shr.u32 %r79, %r78, 30;
add.s32 %r80, %r77, %r79;
shr.s32 %r81, %r80, 2;
add.s32 %r82, %r76, %r81;
shr.s32 %r83, %r82, 31;
shr.u32 %r84, %r83, 29;
add.s32 %r85, %r82, %r84;
and.b32 %r86, %r85, -8;
sub.s32 %r87, %r82, %r86;
cvt.s64.s32 %rd40, %r87;
shr.s32 %r88, %r75, 3;
shl.b32 %r89, %r88, 2;
shl.b32 %r90, %r72, 1;
shl.b32 %r91, %r77, 1;
cvt.s64.s32 %rd4, %r77;
shr.s32 %r92, %r85, 3;
shl.b32 %r5, %r92, 2;
cvt.s64.s32 %rd5, %r89;
shr.u32 %r93, %r83, 30;
add.s32 %r94, %r82, %r93;
and.b32 %r95, %r94, -4;
sub.s32 %r96, %r82, %r95;
shr.s32 %r97, %r94, 2;
shl.b32 %r6, %r97, 1;
add.s32 %r98, %r96, 8;
cvt.s64.s32 %rd41, %r98;
add.s32 %r99, %r82, 2;
shr.s32 %r100, %r99, 31;
shr.u32 %r101, %r100, 29;
add.s32 %r102, %r99, %r101;
and.b32 %r103, %r102, -8;
sub.s32 %r104, %r99, %r103;
shr.s32 %r105, %r102, 3;
shl.b32 %r7, %r105, 2;
cvt.s64.s32 %rd42, %r104;
shr.u32 %r106, %r100, 30;
add.s32 %r107, %r99, %r106;
and.b32 %r108, %r107, -4;
sub.s32 %r109, %r99, %r108;
shr.s32 %r110, %r107, 2;
shl.b32 %r8, %r110, 1;
add.s32 %r111, %r109, 8;
cvt.s64.s32 %rd43, %r111;
add.s32 %r112, %r82, 4;
shr.s32 %r113, %r112, 31;
shr.u32 %r114, %r113, 29;
add.s32 %r115, %r112, %r114;
and.b32 %r116, %r115, -8;
sub.s32 %r117, %r112, %r116;
shr.s32 %r118, %r115, 3;
shl.b32 %r9, %r118, 2;
cvt.s64.s32 %rd44, %r117;
shr.u32 %r119, %r113, 30;
add.s32 %r120, %r112, %r119;
and.b32 %r121, %r120, -4;
sub.s32 %r122, %r112, %r121;
shr.s32 %r123, %r120, 2;
shl.b32 %r10, %r123, 1;
add.s32 %r124, %r122, 8;
cvt.s64.s32 %rd45, %r124;
add.s32 %r125, %r82, 6;
shr.s32 %r126, %r125, 31;
shr.u32 %r127, %r126, 29;
add.s32 %r128, %r125, %r127;
and.b32 %r129, %r128, -8;
sub.s32 %r130, %r125, %r129;
shr.s32 %r131, %r128, 3;
shl.b32 %r11, %r131, 2;
cvt.s64.s32 %rd46, %r130;
shr.u32 %r132, %r126, 30;
add.s32 %r133, %r125, %r132;
and.b32 %r134, %r133, -4;
sub.s32 %r135, %r125, %r134;
shr.s32 %r136, %r133, 2;
shl.b32 %r12, %r136, 1;
add.s32 %r137, %r135, 8;
cvt.s64.s32 %rd47, %r137;
shl.b32 %r138, %r33, 1;
shl.b32 %r1062, %r1067, 3;
add.s32 %r1066, %r1062, %r138;
mad.lo.s32 %r1065, %r33, 3, %r1062;
shl.b32 %r139, %r33, 2;
add.s32 %r1064, %r1062, %r139;
add.s32 %r1063, %r33, %r1062;
add.s32 %r140, %r1067, %r62;
add.s32 %r141, %r140, %r66;
mad.lo.s32 %r142, %r67, %r32, %r141;
mul.wide.s32 %rd6, %r142, 110;
add.s64 %rd48, %rd6, %rd47;
add.s64 %rd7, %rd48, 96;
add.s64 %rd49, %rd6, %rd46;
add.s64 %rd8, %rd49, 96;
add.s64 %rd50, %rd6, %rd45;
add.s64 %rd9, %rd50, 96;
add.s64 %rd51, %rd6, %rd44;
add.s64 %rd10, %rd51, 96;
add.s64 %rd52, %rd6, %rd43;
add.s64 %rd11, %rd52, 96;
add.s64 %rd53, %rd6, %rd42;
add.s64 %rd12, %rd53, 96;
add.s64 %rd54, %rd6, %rd41;
add.s64 %rd13, %rd54, 96;
add.s64 %rd55, %rd6, %rd40;
add.s64 %rd14, %rd55, 96;
mad.lo.s32 %r143, %r54, %r32, %r1067;
add.s32 %r144, %r143, %r62;
add.s32 %r145, %r144, %r66;
mul.wide.s32 %rd15, %r145, 110;
mul.wide.s32 %rd56, %r90, 2;
add.s64 %rd16, %rd15, %rd56;
mul.wide.s32 %rd57, %r91, 2;
add.s64 %rd17, %rd6, %rd57;
add.s64 %rd58, %rd15, %rd57;
add.s64 %rd18, %rd58, 2;
add.s64 %rd59, %rd6, %rd56;
add.s64 %rd19, %rd59, 32;
add.s64 %rd60, %rd15, %rd47;
add.s64 %rd20, %rd60, 96;
add.s64 %rd61, %rd15, %rd46;
add.s64 %rd21, %rd61, 96;
add.s64 %rd62, %rd15, %rd45;
add.s64 %rd22, %rd62, 96;
add.s64 %rd63, %rd15, %rd44;
add.s64 %rd23, %rd63, 96;
add.s64 %rd64, %rd15, %rd43;
add.s64 %rd24, %rd64, 96;
add.s64 %rd65, %rd15, %rd42;
add.s64 %rd25, %rd65, 96;
add.s64 %rd66, %rd15, %rd41;
add.s64 %rd26, %rd66, 96;
add.s64 %rd67, %rd15, %rd40;
add.s64 %rd27, %rd67, 96;
cvta.to.global.u64 %rd28, %rd36;
mov.f32 %f311, 0f00000000;
shl.b64 %rd76, %rd4, 2;

$L__BB149_2:
cvt.u64.u32 %rd68, %r1062;
add.s64 %rd69, %rd2, %rd68;
add.s64 %rd70, %rd144, %rd15;
ld.global.nc.u16 %rs1, [%rd70+108];

{ cvt.f32.f16 %f61, %rs1;}


add.s64 %rd71, %rd144, %rd16;
add.s64 %rd72, %rd144, %rd18;
ld.global.nc.u16 %rs11, [%rd72+-2];
ld.global.nc.u16 %rs12, [%rd72];
mov.b32 %r506, {%rs11, %rs12};
not.b32 %r507, %r506;
ld.global.nc.u16 %rs13, [%rd71+34];
ld.global.nc.u16 %rs14, [%rd71+32];
add.s64 %rd73, %rd69, %rd5;
mul.lo.s64 %rd74, %rd73, 36;
add.s64 %rd75, %rd28, %rd74;
add.s64 %rd77, %rd75, %rd76;
ld.global.nc.u32 %r156, [%rd77+4];
ld.global.nc.u32 %r146, [%rd75];

{.reg .f16 low,high;
mov.b32 {low,high},%r146;
cvt.f32.f16 %f62, low;}


ld.global.nc.u32 %r164, [%rd77+40];
ld.global.nc.u32 %r147, [%rd75+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r147;
cvt.f32.f16 %f63, low;}


ld.global.nc.u32 %r172, [%rd77+76];
ld.global.nc.u32 %r148, [%rd75+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r148;
cvt.f32.f16 %f64, low;}


ld.global.nc.u32 %r180, [%rd77+112];
ld.global.nc.u32 %r149, [%rd75+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r149;
cvt.f32.f16 %f65, low;}


mov.b32 %r508, {%rs14, %rs13};
cvt.u32.u64 %r509, %rd5;
shr.s32 %r510, %r507, %r509;
add.s64 %rd78, %rd144, %rd27;
ld.global.nc.u8 %rs15, [%rd78];
cvt.u32.u16 %r511, %rs15;
and.b32 %r512, %r511, 255;
shr.u32 %r513, %r512, %r5;
and.b32 %r514, %r513, 15;
add.s64 %rd79, %rd144, %rd26;
ld.global.nc.u8 %rs16, [%rd79];
cvt.u32.u16 %r515, %rs16;
and.b32 %r516, %r515, 255;
shr.u32 %r517, %r516, %r6;
and.b32 %r518, %r517, 3;
bfi.b32 %r519, %r518, %r514, 4, 2;
add.s32 %r520, %r519, -32;
and.b32 %r440, %r508, 50529027;
shl.b32 %r521, %r510, 2;
and.b32 %r522, %r521, 67372036;
or.b32 %r523, %r440, -2139062144;
sub.s32 %r524, %r523, %r522;
xor.b32 %r525, %r524, -2139062144;
or.b32 %r526, %r440, %r522;
xor.b32 %r527, %r524, %r440;
and.b32 %r438, %r527, %r526;
mov.u32 %r150, %r438;

prmt.b32 %r150, %r150, 0, 0xba98;

mov.u32 %r152, %r440;

prmt.b32 %r152, %r152, 0, 0xba98;

xor.b32 %r528, %r152, 2139062143;
not.b32 %r529, %r150;
and.b32 %r530, %r525, %r529;
and.b32 %r531, %r528, %r150;
or.b32 %r155, %r531, %r530;
mov.u32 %r505, 0;

dp4a.s32.s32 %r154, %r155, %r156, %r505;

mul.lo.s32 %r532, %r520, %r154;
cvt.rn.f32.s32 %f111, %r532;
fma.rn.ftz.f32 %f112, %f62, %f111, 0f00000000;
add.s64 %rd80, %rd144, %rd25;
ld.global.nc.u8 %rs17, [%rd80];
cvt.u32.u16 %r533, %rs17;
and.b32 %r534, %r533, 255;
shr.u32 %r535, %r534, %r7;
and.b32 %r536, %r535, 15;
add.s64 %rd81, %rd144, %rd24;
ld.global.nc.u8 %rs18, [%rd81];
cvt.u32.u16 %r537, %rs18;
and.b32 %r538, %r537, 255;
shr.u32 %r539, %r538, %r8;
and.b32 %r540, %r539, 3;
bfi.b32 %r541, %r540, %r536, 4, 2;
add.s32 %r542, %r541, -32;
shr.u32 %r543, %r508, 2;
and.b32 %r448, %r543, 50529027;
shl.b32 %r544, %r510, 1;
and.b32 %r545, %r544, 67372036;
or.b32 %r546, %r448, -2139062144;
sub.s32 %r547, %r546, %r545;
xor.b32 %r548, %r547, -2139062144;
or.b32 %r549, %r448, %r545;
xor.b32 %r550, %r547, %r448;
and.b32 %r446, %r550, %r549;
mov.u32 %r158, %r446;

prmt.b32 %r158, %r158, 0, 0xba98;

mov.u32 %r160, %r448;

prmt.b32 %r160, %r160, 0, 0xba98;

xor.b32 %r551, %r160, 2139062143;
not.b32 %r552, %r158;
and.b32 %r553, %r548, %r552;
and.b32 %r554, %r551, %r158;
or.b32 %r163, %r554, %r553;

dp4a.s32.s32 %r162, %r163, %r164, %r505;

mul.lo.s32 %r555, %r542, %r162;
cvt.rn.f32.s32 %f113, %r555;
fma.rn.ftz.f32 %f114, %f63, %f113, %f112;
add.s64 %rd82, %rd144, %rd23;
ld.global.nc.u8 %rs19, [%rd82];
cvt.u32.u16 %r556, %rs19;
and.b32 %r557, %r556, 255;
shr.u32 %r558, %r557, %r9;
and.b32 %r559, %r558, 15;
add.s64 %rd83, %rd144, %rd22;
ld.global.nc.u8 %rs20, [%rd83];
cvt.u32.u16 %r560, %rs20;
and.b32 %r561, %r560, 255;
shr.u32 %r562, %r561, %r10;
and.b32 %r563, %r562, 3;
bfi.b32 %r564, %r563, %r559, 4, 2;
add.s32 %r565, %r564, -32;
shr.u32 %r566, %r508, 4;
and.b32 %r456, %r566, 50529027;
and.b32 %r567, %r510, 67372036;
or.b32 %r568, %r456, -2139062144;
sub.s32 %r569, %r568, %r567;
xor.b32 %r570, %r569, -2139062144;
or.b32 %r571, %r456, %r567;
xor.b32 %r572, %r569, %r456;
and.b32 %r454, %r572, %r571;
mov.u32 %r166, %r454;

prmt.b32 %r166, %r166, 0, 0xba98;

mov.u32 %r168, %r456;

prmt.b32 %r168, %r168, 0, 0xba98;

xor.b32 %r573, %r168, 2139062143;
not.b32 %r574, %r166;
and.b32 %r575, %r570, %r574;
and.b32 %r576, %r573, %r166;
or.b32 %r171, %r576, %r575;

dp4a.s32.s32 %r170, %r171, %r172, %r505;

mul.lo.s32 %r577, %r565, %r170;
cvt.rn.f32.s32 %f115, %r577;
fma.rn.ftz.f32 %f116, %f64, %f115, %f114;
add.s64 %rd84, %rd144, %rd21;
ld.global.nc.u8 %rs21, [%rd84];
cvt.u32.u16 %r578, %rs21;
and.b32 %r579, %r578, 255;
shr.u32 %r580, %r579, %r11;
and.b32 %r581, %r580, 15;
add.s64 %rd85, %rd144, %rd20;
ld.global.nc.u8 %rs22, [%rd85];
cvt.u32.u16 %r582, %rs22;
and.b32 %r583, %r582, 255;
shr.u32 %r584, %r583, %r12;
and.b32 %r585, %r584, 3;
bfi.b32 %r586, %r585, %r581, 4, 2;
add.s32 %r587, %r586, -32;
shr.u32 %r588, %r508, 6;
and.b32 %r464, %r588, 50529027;
shr.u32 %r589, %r510, 1;
and.b32 %r590, %r589, 67372036;
or.b32 %r591, %r464, -2139062144;
sub.s32 %r592, %r591, %r590;
xor.b32 %r593, %r592, -2139062144;
or.b32 %r594, %r464, %r590;
xor.b32 %r595, %r592, %r464;
and.b32 %r462, %r595, %r594;
mov.u32 %r174, %r462;

prmt.b32 %r174, %r174, 0, 0xba98;

mov.u32 %r176, %r464;

prmt.b32 %r176, %r176, 0, 0xba98;

xor.b32 %r596, %r176, 2139062143;
not.b32 %r597, %r174;
and.b32 %r598, %r593, %r597;
and.b32 %r599, %r596, %r174;
or.b32 %r179, %r599, %r598;

dp4a.s32.s32 %r178, %r179, %r180, %r505;

mul.lo.s32 %r600, %r587, %r178;
cvt.rn.f32.s32 %f117, %r600;
fma.rn.ftz.f32 %f118, %f65, %f117, %f116;
fma.rn.ftz.f32 %f320, %f61, %f118, %f320;
add.s64 %rd86, %rd144, %rd6;
ld.global.nc.u16 %rs2, [%rd86+108];

{ cvt.f32.f16 %f66, %rs2;}


add.s64 %rd87, %rd144, %rd19;
add.s64 %rd88, %rd144, %rd17;
ld.global.nc.u16 %rs23, [%rd88];
ld.global.nc.u16 %rs24, [%rd88+2];
mov.b32 %r601, {%rs23, %rs24};
not.b32 %r602, %r601;
ld.global.nc.u16 %rs25, [%rd87+2];
ld.global.nc.u16 %rs26, [%rd87];

{.reg .f16 low,high;
mov.b32 {low,high},%r146;
cvt.f32.f16 %f67, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r147;
cvt.f32.f16 %f68, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r148;
cvt.f32.f16 %f69, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r149;
cvt.f32.f16 %f70, low;}


mov.b32 %r603, {%rs26, %rs25};
shr.s32 %r604, %r602, %r509;
add.s64 %rd89, %rd144, %rd14;
ld.global.nc.u8 %rs27, [%rd89];
cvt.u32.u16 %r605, %rs27;
and.b32 %r606, %r605, 255;
shr.u32 %r607, %r606, %r5;
and.b32 %r608, %r607, 15;
add.s64 %rd90, %rd144, %rd13;
ld.global.nc.u8 %rs28, [%rd90];
cvt.u32.u16 %r609, %rs28;
and.b32 %r610, %r609, 255;
shr.u32 %r611, %r610, %r6;
and.b32 %r612, %r611, 3;
bfi.b32 %r613, %r612, %r608, 4, 2;
add.s32 %r614, %r613, -32;
and.b32 %r476, %r603, 50529027;
shl.b32 %r615, %r604, 2;
and.b32 %r616, %r615, 67372036;
or.b32 %r617, %r476, -2139062144;
sub.s32 %r618, %r617, %r616;
xor.b32 %r619, %r618, -2139062144;
or.b32 %r620, %r476, %r616;
xor.b32 %r621, %r618, %r476;
and.b32 %r474, %r621, %r620;
mov.u32 %r186, %r474;

prmt.b32 %r186, %r186, 0, 0xba98;

mov.u32 %r188, %r476;

prmt.b32 %r188, %r188, 0, 0xba98;

xor.b32 %r622, %r188, 2139062143;
not.b32 %r623, %r186;
and.b32 %r624, %r619, %r623;
and.b32 %r625, %r622, %r186;
or.b32 %r191, %r625, %r624;

dp4a.s32.s32 %r190, %r191, %r156, %r505;

mul.lo.s32 %r626, %r614, %r190;
cvt.rn.f32.s32 %f119, %r626;
fma.rn.ftz.f32 %f120, %f67, %f119, 0f00000000;
add.s64 %rd91, %rd144, %rd12;
ld.global.nc.u8 %rs29, [%rd91];
cvt.u32.u16 %r627, %rs29;
and.b32 %r628, %r627, 255;
shr.u32 %r629, %r628, %r7;
and.b32 %r630, %r629, 15;
add.s64 %rd92, %rd144, %rd11;
ld.global.nc.u8 %rs30, [%rd92];
cvt.u32.u16 %r631, %rs30;
and.b32 %r632, %r631, 255;
shr.u32 %r633, %r632, %r8;
and.b32 %r634, %r633, 3;
bfi.b32 %r635, %r634, %r630, 4, 2;
add.s32 %r636, %r635, -32;
shr.u32 %r637, %r603, 2;
and.b32 %r484, %r637, 50529027;
shl.b32 %r638, %r604, 1;
and.b32 %r639, %r638, 67372036;
or.b32 %r640, %r484, -2139062144;
sub.s32 %r641, %r640, %r639;
xor.b32 %r642, %r641, -2139062144;
or.b32 %r643, %r484, %r639;
xor.b32 %r644, %r641, %r484;
and.b32 %r482, %r644, %r643;
mov.u32 %r194, %r482;

prmt.b32 %r194, %r194, 0, 0xba98;

mov.u32 %r196, %r484;

prmt.b32 %r196, %r196, 0, 0xba98;

xor.b32 %r645, %r196, 2139062143;
not.b32 %r646, %r194;
and.b32 %r647, %r642, %r646;
and.b32 %r648, %r645, %r194;
or.b32 %r199, %r648, %r647;

dp4a.s32.s32 %r198, %r199, %r164, %r505;

mul.lo.s32 %r649, %r636, %r198;
cvt.rn.f32.s32 %f121, %r649;
fma.rn.ftz.f32 %f122, %f68, %f121, %f120;
add.s64 %rd93, %rd144, %rd10;
ld.global.nc.u8 %rs31, [%rd93];
cvt.u32.u16 %r650, %rs31;
and.b32 %r651, %r650, 255;
shr.u32 %r652, %r651, %r9;
and.b32 %r653, %r652, 15;
add.s64 %rd94, %rd144, %rd9;
ld.global.nc.u8 %rs32, [%rd94];
cvt.u32.u16 %r654, %rs32;
and.b32 %r655, %r654, 255;
shr.u32 %r656, %r655, %r10;
and.b32 %r657, %r656, 3;
bfi.b32 %r658, %r657, %r653, 4, 2;
add.s32 %r659, %r658, -32;
shr.u32 %r660, %r603, 4;
and.b32 %r492, %r660, 50529027;
and.b32 %r661, %r604, 67372036;
or.b32 %r662, %r492, -2139062144;
sub.s32 %r663, %r662, %r661;
xor.b32 %r664, %r663, -2139062144;
or.b32 %r665, %r492, %r661;
xor.b32 %r666, %r663, %r492;
and.b32 %r490, %r666, %r665;
mov.u32 %r202, %r490;

prmt.b32 %r202, %r202, 0, 0xba98;

mov.u32 %r204, %r492;

prmt.b32 %r204, %r204, 0, 0xba98;

xor.b32 %r667, %r204, 2139062143;
not.b32 %r668, %r202;
and.b32 %r669, %r664, %r668;
and.b32 %r670, %r667, %r202;
or.b32 %r207, %r670, %r669;

dp4a.s32.s32 %r206, %r207, %r172, %r505;

mul.lo.s32 %r671, %r659, %r206;
cvt.rn.f32.s32 %f123, %r671;
fma.rn.ftz.f32 %f124, %f69, %f123, %f122;
add.s64 %rd95, %rd144, %rd8;
ld.global.nc.u8 %rs33, [%rd95];
cvt.u32.u16 %r672, %rs33;
and.b32 %r673, %r672, 255;
shr.u32 %r674, %r673, %r11;
and.b32 %r675, %r674, 15;
add.s64 %rd96, %rd144, %rd7;
ld.global.nc.u8 %rs34, [%rd96];
cvt.u32.u16 %r676, %rs34;
and.b32 %r677, %r676, 255;
shr.u32 %r678, %r677, %r12;
and.b32 %r679, %r678, 3;
bfi.b32 %r680, %r679, %r675, 4, 2;
add.s32 %r681, %r680, -32;
shr.u32 %r682, %r603, 6;
and.b32 %r500, %r682, 50529027;
shr.u32 %r683, %r604, 1;
and.b32 %r684, %r683, 67372036;
or.b32 %r685, %r500, -2139062144;
sub.s32 %r686, %r685, %r684;
xor.b32 %r687, %r686, -2139062144;
or.b32 %r688, %r500, %r684;
xor.b32 %r689, %r686, %r500;
and.b32 %r498, %r689, %r688;
mov.u32 %r210, %r498;

prmt.b32 %r210, %r210, 0, 0xba98;

mov.u32 %r212, %r500;

prmt.b32 %r212, %r212, 0, 0xba98;

xor.b32 %r690, %r212, 2139062143;
not.b32 %r691, %r210;
and.b32 %r692, %r687, %r691;
and.b32 %r693, %r690, %r210;
or.b32 %r215, %r693, %r692;

dp4a.s32.s32 %r214, %r215, %r180, %r505;

mul.lo.s32 %r694, %r681, %r214;
cvt.rn.f32.s32 %f125, %r694;
fma.rn.ftz.f32 %f126, %f70, %f125, %f124;
fma.rn.ftz.f32 %f319, %f66, %f126, %f319;
cvt.u64.u32 %rd97, %r1063;
add.s64 %rd98, %rd2, %rd97;

{ cvt.f32.f16 %f71, %rs1;}


add.s64 %rd99, %rd98, %rd5;
mul.lo.s64 %rd100, %rd99, 36;
add.s64 %rd101, %rd28, %rd100;
add.s64 %rd102, %rd101, %rd76;
ld.global.nc.u32 %r228, [%rd102+4];
ld.global.nc.u32 %r218, [%rd101];

{.reg .f16 low,high;
mov.b32 {low,high},%r218;
cvt.f32.f16 %f72, low;}


ld.global.nc.u32 %r236, [%rd102+40];
ld.global.nc.u32 %r219, [%rd101+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r219;
cvt.f32.f16 %f73, low;}


ld.global.nc.u32 %r244, [%rd102+76];
ld.global.nc.u32 %r220, [%rd101+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r220;
cvt.f32.f16 %f74, low;}


ld.global.nc.u32 %r252, [%rd102+112];
ld.global.nc.u32 %r221, [%rd101+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r221;
cvt.f32.f16 %f75, low;}


mov.u32 %r222, %r438;

prmt.b32 %r222, %r222, 0, 0xba98;

mov.u32 %r224, %r440;

prmt.b32 %r224, %r224, 0, 0xba98;

xor.b32 %r695, %r224, 2139062143;
not.b32 %r696, %r222;
and.b32 %r697, %r525, %r696;
and.b32 %r698, %r695, %r222;
or.b32 %r227, %r698, %r697;

dp4a.s32.s32 %r226, %r227, %r228, %r505;

mul.lo.s32 %r699, %r520, %r226;
cvt.rn.f32.s32 %f127, %r699;
fma.rn.ftz.f32 %f128, %f72, %f127, 0f00000000;
mov.u32 %r230, %r446;

prmt.b32 %r230, %r230, 0, 0xba98;

mov.u32 %r232, %r448;

prmt.b32 %r232, %r232, 0, 0xba98;

xor.b32 %r700, %r232, 2139062143;
not.b32 %r701, %r230;
and.b32 %r702, %r548, %r701;
and.b32 %r703, %r700, %r230;
or.b32 %r235, %r703, %r702;

dp4a.s32.s32 %r234, %r235, %r236, %r505;

mul.lo.s32 %r704, %r542, %r234;
cvt.rn.f32.s32 %f129, %r704;
fma.rn.ftz.f32 %f130, %f73, %f129, %f128;
mov.u32 %r238, %r454;

prmt.b32 %r238, %r238, 0, 0xba98;

mov.u32 %r240, %r456;

prmt.b32 %r240, %r240, 0, 0xba98;

xor.b32 %r705, %r240, 2139062143;
not.b32 %r706, %r238;
and.b32 %r707, %r570, %r706;
and.b32 %r708, %r705, %r238;
or.b32 %r243, %r708, %r707;

dp4a.s32.s32 %r242, %r243, %r244, %r505;

mul.lo.s32 %r709, %r565, %r242;
cvt.rn.f32.s32 %f131, %r709;
fma.rn.ftz.f32 %f132, %f74, %f131, %f130;
mov.u32 %r246, %r462;

prmt.b32 %r246, %r246, 0, 0xba98;

mov.u32 %r248, %r464;

prmt.b32 %r248, %r248, 0, 0xba98;

xor.b32 %r710, %r248, 2139062143;
not.b32 %r711, %r246;
and.b32 %r712, %r593, %r711;
and.b32 %r713, %r710, %r246;
or.b32 %r251, %r713, %r712;

dp4a.s32.s32 %r250, %r251, %r252, %r505;

mul.lo.s32 %r714, %r587, %r250;
cvt.rn.f32.s32 %f133, %r714;
fma.rn.ftz.f32 %f134, %f75, %f133, %f132;
fma.rn.ftz.f32 %f318, %f71, %f134, %f318;

{ cvt.f32.f16 %f76, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r218;
cvt.f32.f16 %f77, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r219;
cvt.f32.f16 %f78, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r220;
cvt.f32.f16 %f79, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r221;
cvt.f32.f16 %f80, low;}


mov.u32 %r258, %r474;

prmt.b32 %r258, %r258, 0, 0xba98;

mov.u32 %r260, %r476;

prmt.b32 %r260, %r260, 0, 0xba98;

xor.b32 %r715, %r260, 2139062143;
not.b32 %r716, %r258;
and.b32 %r717, %r619, %r716;
and.b32 %r718, %r715, %r258;
or.b32 %r263, %r718, %r717;

dp4a.s32.s32 %r262, %r263, %r228, %r505;

mul.lo.s32 %r719, %r614, %r262;
cvt.rn.f32.s32 %f135, %r719;
fma.rn.ftz.f32 %f136, %f77, %f135, 0f00000000;
mov.u32 %r266, %r482;

prmt.b32 %r266, %r266, 0, 0xba98;

mov.u32 %r268, %r484;

prmt.b32 %r268, %r268, 0, 0xba98;

xor.b32 %r720, %r268, 2139062143;
not.b32 %r721, %r266;
and.b32 %r722, %r642, %r721;
and.b32 %r723, %r720, %r266;
or.b32 %r271, %r723, %r722;

dp4a.s32.s32 %r270, %r271, %r236, %r505;

mul.lo.s32 %r724, %r636, %r270;
cvt.rn.f32.s32 %f137, %r724;
fma.rn.ftz.f32 %f138, %f78, %f137, %f136;
mov.u32 %r274, %r490;

prmt.b32 %r274, %r274, 0, 0xba98;

mov.u32 %r276, %r492;

prmt.b32 %r276, %r276, 0, 0xba98;

xor.b32 %r725, %r276, 2139062143;
not.b32 %r726, %r274;
and.b32 %r727, %r664, %r726;
and.b32 %r728, %r725, %r274;
or.b32 %r279, %r728, %r727;

dp4a.s32.s32 %r278, %r279, %r244, %r505;

mul.lo.s32 %r729, %r659, %r278;
cvt.rn.f32.s32 %f139, %r729;
fma.rn.ftz.f32 %f140, %f79, %f139, %f138;
mov.u32 %r282, %r498;

prmt.b32 %r282, %r282, 0, 0xba98;

mov.u32 %r284, %r500;

prmt.b32 %r284, %r284, 0, 0xba98;

xor.b32 %r730, %r284, 2139062143;
not.b32 %r731, %r282;
and.b32 %r732, %r687, %r731;
and.b32 %r733, %r730, %r282;
or.b32 %r287, %r733, %r732;

dp4a.s32.s32 %r286, %r287, %r252, %r505;

mul.lo.s32 %r734, %r681, %r286;
cvt.rn.f32.s32 %f141, %r734;
fma.rn.ftz.f32 %f142, %f80, %f141, %f140;
fma.rn.ftz.f32 %f317, %f76, %f142, %f317;
cvt.u64.u32 %rd103, %r1066;
add.s64 %rd104, %rd2, %rd103;

{ cvt.f32.f16 %f81, %rs1;}


add.s64 %rd105, %rd104, %rd5;
mul.lo.s64 %rd106, %rd105, 36;
add.s64 %rd107, %rd28, %rd106;
add.s64 %rd108, %rd107, %rd76;
ld.global.nc.u32 %r300, [%rd108+4];
ld.global.nc.u32 %r290, [%rd107];

{.reg .f16 low,high;
mov.b32 {low,high},%r290;
cvt.f32.f16 %f82, low;}


ld.global.nc.u32 %r308, [%rd108+40];
ld.global.nc.u32 %r291, [%rd107+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r291;
cvt.f32.f16 %f83, low;}


ld.global.nc.u32 %r316, [%rd108+76];
ld.global.nc.u32 %r292, [%rd107+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r292;
cvt.f32.f16 %f84, low;}


ld.global.nc.u32 %r324, [%rd108+112];
ld.global.nc.u32 %r293, [%rd107+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r293;
cvt.f32.f16 %f85, low;}


mov.u32 %r294, %r438;

prmt.b32 %r294, %r294, 0, 0xba98;

mov.u32 %r296, %r440;

prmt.b32 %r296, %r296, 0, 0xba98;

xor.b32 %r735, %r296, 2139062143;
not.b32 %r736, %r294;
and.b32 %r737, %r525, %r736;
and.b32 %r738, %r735, %r294;
or.b32 %r299, %r738, %r737;

dp4a.s32.s32 %r298, %r299, %r300, %r505;

mul.lo.s32 %r739, %r520, %r298;
cvt.rn.f32.s32 %f143, %r739;
fma.rn.ftz.f32 %f144, %f82, %f143, 0f00000000;
mov.u32 %r302, %r446;

prmt.b32 %r302, %r302, 0, 0xba98;

mov.u32 %r304, %r448;

prmt.b32 %r304, %r304, 0, 0xba98;

xor.b32 %r740, %r304, 2139062143;
not.b32 %r741, %r302;
and.b32 %r742, %r548, %r741;
and.b32 %r743, %r740, %r302;
or.b32 %r307, %r743, %r742;

dp4a.s32.s32 %r306, %r307, %r308, %r505;

mul.lo.s32 %r744, %r542, %r306;
cvt.rn.f32.s32 %f145, %r744;
fma.rn.ftz.f32 %f146, %f83, %f145, %f144;
mov.u32 %r310, %r454;

prmt.b32 %r310, %r310, 0, 0xba98;

mov.u32 %r312, %r456;

prmt.b32 %r312, %r312, 0, 0xba98;

xor.b32 %r745, %r312, 2139062143;
not.b32 %r746, %r310;
and.b32 %r747, %r570, %r746;
and.b32 %r748, %r745, %r310;
or.b32 %r315, %r748, %r747;

dp4a.s32.s32 %r314, %r315, %r316, %r505;

mul.lo.s32 %r749, %r565, %r314;
cvt.rn.f32.s32 %f147, %r749;
fma.rn.ftz.f32 %f148, %f84, %f147, %f146;
mov.u32 %r318, %r462;

prmt.b32 %r318, %r318, 0, 0xba98;

mov.u32 %r320, %r464;

prmt.b32 %r320, %r320, 0, 0xba98;

xor.b32 %r750, %r320, 2139062143;
not.b32 %r751, %r318;
and.b32 %r752, %r593, %r751;
and.b32 %r753, %r750, %r318;
or.b32 %r323, %r753, %r752;

dp4a.s32.s32 %r322, %r323, %r324, %r505;

mul.lo.s32 %r754, %r587, %r322;
cvt.rn.f32.s32 %f149, %r754;
fma.rn.ftz.f32 %f150, %f85, %f149, %f148;
fma.rn.ftz.f32 %f316, %f81, %f150, %f316;

{ cvt.f32.f16 %f86, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r290;
cvt.f32.f16 %f87, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r291;
cvt.f32.f16 %f88, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r292;
cvt.f32.f16 %f89, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r293;
cvt.f32.f16 %f90, low;}


mov.u32 %r330, %r474;

prmt.b32 %r330, %r330, 0, 0xba98;

mov.u32 %r332, %r476;

prmt.b32 %r332, %r332, 0, 0xba98;

xor.b32 %r755, %r332, 2139062143;
not.b32 %r756, %r330;
and.b32 %r757, %r619, %r756;
and.b32 %r758, %r755, %r330;
or.b32 %r335, %r758, %r757;

dp4a.s32.s32 %r334, %r335, %r300, %r505;

mul.lo.s32 %r759, %r614, %r334;
cvt.rn.f32.s32 %f151, %r759;
fma.rn.ftz.f32 %f152, %f87, %f151, 0f00000000;
mov.u32 %r338, %r482;

prmt.b32 %r338, %r338, 0, 0xba98;

mov.u32 %r340, %r484;

prmt.b32 %r340, %r340, 0, 0xba98;

xor.b32 %r760, %r340, 2139062143;
not.b32 %r761, %r338;
and.b32 %r762, %r642, %r761;
and.b32 %r763, %r760, %r338;
or.b32 %r343, %r763, %r762;

dp4a.s32.s32 %r342, %r343, %r308, %r505;

mul.lo.s32 %r764, %r636, %r342;
cvt.rn.f32.s32 %f153, %r764;
fma.rn.ftz.f32 %f154, %f88, %f153, %f152;
mov.u32 %r346, %r490;

prmt.b32 %r346, %r346, 0, 0xba98;

mov.u32 %r348, %r492;

prmt.b32 %r348, %r348, 0, 0xba98;

xor.b32 %r765, %r348, 2139062143;
not.b32 %r766, %r346;
and.b32 %r767, %r664, %r766;
and.b32 %r768, %r765, %r346;
or.b32 %r351, %r768, %r767;

dp4a.s32.s32 %r350, %r351, %r316, %r505;

mul.lo.s32 %r769, %r659, %r350;
cvt.rn.f32.s32 %f155, %r769;
fma.rn.ftz.f32 %f156, %f89, %f155, %f154;
mov.u32 %r354, %r498;

prmt.b32 %r354, %r354, 0, 0xba98;

mov.u32 %r356, %r500;

prmt.b32 %r356, %r356, 0, 0xba98;

xor.b32 %r770, %r356, 2139062143;
not.b32 %r771, %r354;
and.b32 %r772, %r687, %r771;
and.b32 %r773, %r770, %r354;
or.b32 %r359, %r773, %r772;

dp4a.s32.s32 %r358, %r359, %r324, %r505;

mul.lo.s32 %r774, %r681, %r358;
cvt.rn.f32.s32 %f157, %r774;
fma.rn.ftz.f32 %f158, %f90, %f157, %f156;
fma.rn.ftz.f32 %f315, %f86, %f158, %f315;
cvt.u64.u32 %rd109, %r1065;
add.s64 %rd110, %rd2, %rd109;

{ cvt.f32.f16 %f91, %rs1;}


add.s64 %rd111, %rd110, %rd5;
mul.lo.s64 %rd112, %rd111, 36;
add.s64 %rd113, %rd28, %rd112;
add.s64 %rd114, %rd113, %rd76;
ld.global.nc.u32 %r372, [%rd114+4];
ld.global.nc.u32 %r362, [%rd113];

{.reg .f16 low,high;
mov.b32 {low,high},%r362;
cvt.f32.f16 %f92, low;}


ld.global.nc.u32 %r380, [%rd114+40];
ld.global.nc.u32 %r363, [%rd113+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r363;
cvt.f32.f16 %f93, low;}


ld.global.nc.u32 %r388, [%rd114+76];
ld.global.nc.u32 %r364, [%rd113+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r364;
cvt.f32.f16 %f94, low;}


ld.global.nc.u32 %r396, [%rd114+112];
ld.global.nc.u32 %r365, [%rd113+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r365;
cvt.f32.f16 %f95, low;}


mov.u32 %r366, %r438;

prmt.b32 %r366, %r366, 0, 0xba98;

mov.u32 %r368, %r440;

prmt.b32 %r368, %r368, 0, 0xba98;

xor.b32 %r775, %r368, 2139062143;
not.b32 %r776, %r366;
and.b32 %r777, %r525, %r776;
and.b32 %r778, %r775, %r366;
or.b32 %r371, %r778, %r777;

dp4a.s32.s32 %r370, %r371, %r372, %r505;

mul.lo.s32 %r779, %r520, %r370;
cvt.rn.f32.s32 %f159, %r779;
fma.rn.ftz.f32 %f160, %f92, %f159, 0f00000000;
mov.u32 %r374, %r446;

prmt.b32 %r374, %r374, 0, 0xba98;

mov.u32 %r376, %r448;

prmt.b32 %r376, %r376, 0, 0xba98;

xor.b32 %r780, %r376, 2139062143;
not.b32 %r781, %r374;
and.b32 %r782, %r548, %r781;
and.b32 %r783, %r780, %r374;
or.b32 %r379, %r783, %r782;

dp4a.s32.s32 %r378, %r379, %r380, %r505;

mul.lo.s32 %r784, %r542, %r378;
cvt.rn.f32.s32 %f161, %r784;
fma.rn.ftz.f32 %f162, %f93, %f161, %f160;
mov.u32 %r382, %r454;

prmt.b32 %r382, %r382, 0, 0xba98;

mov.u32 %r384, %r456;

prmt.b32 %r384, %r384, 0, 0xba98;

xor.b32 %r785, %r384, 2139062143;
not.b32 %r786, %r382;
and.b32 %r787, %r570, %r786;
and.b32 %r788, %r785, %r382;
or.b32 %r387, %r788, %r787;

dp4a.s32.s32 %r386, %r387, %r388, %r505;

mul.lo.s32 %r789, %r565, %r386;
cvt.rn.f32.s32 %f163, %r789;
fma.rn.ftz.f32 %f164, %f94, %f163, %f162;
mov.u32 %r390, %r462;

prmt.b32 %r390, %r390, 0, 0xba98;

mov.u32 %r392, %r464;

prmt.b32 %r392, %r392, 0, 0xba98;

xor.b32 %r790, %r392, 2139062143;
not.b32 %r791, %r390;
and.b32 %r792, %r593, %r791;
and.b32 %r793, %r790, %r390;
or.b32 %r395, %r793, %r792;

dp4a.s32.s32 %r394, %r395, %r396, %r505;

mul.lo.s32 %r794, %r587, %r394;
cvt.rn.f32.s32 %f165, %r794;
fma.rn.ftz.f32 %f166, %f95, %f165, %f164;
fma.rn.ftz.f32 %f314, %f91, %f166, %f314;

{ cvt.f32.f16 %f96, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r362;
cvt.f32.f16 %f97, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r363;
cvt.f32.f16 %f98, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r364;
cvt.f32.f16 %f99, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r365;
cvt.f32.f16 %f100, low;}


mov.u32 %r402, %r474;

prmt.b32 %r402, %r402, 0, 0xba98;

mov.u32 %r404, %r476;

prmt.b32 %r404, %r404, 0, 0xba98;

xor.b32 %r795, %r404, 2139062143;
not.b32 %r796, %r402;
and.b32 %r797, %r619, %r796;
and.b32 %r798, %r795, %r402;
or.b32 %r407, %r798, %r797;

dp4a.s32.s32 %r406, %r407, %r372, %r505;

mul.lo.s32 %r799, %r614, %r406;
cvt.rn.f32.s32 %f167, %r799;
fma.rn.ftz.f32 %f168, %f97, %f167, 0f00000000;
mov.u32 %r410, %r482;

prmt.b32 %r410, %r410, 0, 0xba98;

mov.u32 %r412, %r484;

prmt.b32 %r412, %r412, 0, 0xba98;

xor.b32 %r800, %r412, 2139062143;
not.b32 %r801, %r410;
and.b32 %r802, %r642, %r801;
and.b32 %r803, %r800, %r410;
or.b32 %r415, %r803, %r802;

dp4a.s32.s32 %r414, %r415, %r380, %r505;

mul.lo.s32 %r804, %r636, %r414;
cvt.rn.f32.s32 %f169, %r804;
fma.rn.ftz.f32 %f170, %f98, %f169, %f168;
mov.u32 %r418, %r490;

prmt.b32 %r418, %r418, 0, 0xba98;

mov.u32 %r420, %r492;

prmt.b32 %r420, %r420, 0, 0xba98;

xor.b32 %r805, %r420, 2139062143;
not.b32 %r806, %r418;
and.b32 %r807, %r664, %r806;
and.b32 %r808, %r805, %r418;
or.b32 %r423, %r808, %r807;

dp4a.s32.s32 %r422, %r423, %r388, %r505;

mul.lo.s32 %r809, %r659, %r422;
cvt.rn.f32.s32 %f171, %r809;
fma.rn.ftz.f32 %f172, %f99, %f171, %f170;
mov.u32 %r426, %r498;

prmt.b32 %r426, %r426, 0, 0xba98;

mov.u32 %r428, %r500;

prmt.b32 %r428, %r428, 0, 0xba98;

xor.b32 %r810, %r428, 2139062143;
not.b32 %r811, %r426;
and.b32 %r812, %r687, %r811;
and.b32 %r813, %r810, %r426;
or.b32 %r431, %r813, %r812;

dp4a.s32.s32 %r430, %r431, %r396, %r505;

mul.lo.s32 %r814, %r681, %r430;
cvt.rn.f32.s32 %f173, %r814;
fma.rn.ftz.f32 %f174, %f100, %f173, %f172;
fma.rn.ftz.f32 %f313, %f96, %f174, %f313;
cvt.u64.u32 %rd115, %r1064;
add.s64 %rd116, %rd2, %rd115;

{ cvt.f32.f16 %f101, %rs1;}


add.s64 %rd117, %rd116, %rd5;
mul.lo.s64 %rd118, %rd117, 36;
add.s64 %rd119, %rd28, %rd118;
add.s64 %rd120, %rd119, %rd76;
ld.global.nc.u32 %r444, [%rd120+4];
ld.global.nc.u32 %r434, [%rd119];

{.reg .f16 low,high;
mov.b32 {low,high},%r434;
cvt.f32.f16 %f102, low;}


ld.global.nc.u32 %r452, [%rd120+40];
ld.global.nc.u32 %r435, [%rd119+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r435;
cvt.f32.f16 %f103, low;}


ld.global.nc.u32 %r460, [%rd120+76];
ld.global.nc.u32 %r436, [%rd119+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r436;
cvt.f32.f16 %f104, low;}


ld.global.nc.u32 %r468, [%rd120+112];
ld.global.nc.u32 %r437, [%rd119+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r437;
cvt.f32.f16 %f105, low;}



prmt.b32 %r438, %r438, 0, 0xba98;


prmt.b32 %r440, %r440, 0, 0xba98;

xor.b32 %r815, %r440, 2139062143;
not.b32 %r816, %r438;
and.b32 %r817, %r525, %r816;
and.b32 %r818, %r815, %r438;
or.b32 %r443, %r818, %r817;

dp4a.s32.s32 %r442, %r443, %r444, %r505;

mul.lo.s32 %r819, %r520, %r442;
cvt.rn.f32.s32 %f175, %r819;
fma.rn.ftz.f32 %f176, %f102, %f175, 0f00000000;

prmt.b32 %r446, %r446, 0, 0xba98;


prmt.b32 %r448, %r448, 0, 0xba98;

xor.b32 %r820, %r448, 2139062143;
not.b32 %r821, %r446;
and.b32 %r822, %r548, %r821;
and.b32 %r823, %r820, %r446;
or.b32 %r451, %r823, %r822;

dp4a.s32.s32 %r450, %r451, %r452, %r505;

mul.lo.s32 %r824, %r542, %r450;
cvt.rn.f32.s32 %f177, %r824;
fma.rn.ftz.f32 %f178, %f103, %f177, %f176;

prmt.b32 %r454, %r454, 0, 0xba98;


prmt.b32 %r456, %r456, 0, 0xba98;

xor.b32 %r825, %r456, 2139062143;
not.b32 %r826, %r454;
and.b32 %r827, %r570, %r826;
and.b32 %r828, %r825, %r454;
or.b32 %r459, %r828, %r827;

dp4a.s32.s32 %r458, %r459, %r460, %r505;

mul.lo.s32 %r829, %r565, %r458;
cvt.rn.f32.s32 %f179, %r829;
fma.rn.ftz.f32 %f180, %f104, %f179, %f178;

prmt.b32 %r462, %r462, 0, 0xba98;


prmt.b32 %r464, %r464, 0, 0xba98;

xor.b32 %r830, %r464, 2139062143;
not.b32 %r831, %r462;
and.b32 %r832, %r593, %r831;
and.b32 %r833, %r830, %r462;
or.b32 %r467, %r833, %r832;

dp4a.s32.s32 %r466, %r467, %r468, %r505;

mul.lo.s32 %r834, %r587, %r466;
cvt.rn.f32.s32 %f181, %r834;
fma.rn.ftz.f32 %f182, %f105, %f181, %f180;
fma.rn.ftz.f32 %f312, %f101, %f182, %f312;

{ cvt.f32.f16 %f106, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r434;
cvt.f32.f16 %f107, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r435;
cvt.f32.f16 %f108, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r436;
cvt.f32.f16 %f109, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r437;
cvt.f32.f16 %f110, low;}



prmt.b32 %r474, %r474, 0, 0xba98;


prmt.b32 %r476, %r476, 0, 0xba98;

xor.b32 %r835, %r476, 2139062143;
not.b32 %r836, %r474;
and.b32 %r837, %r619, %r836;
and.b32 %r838, %r835, %r474;
or.b32 %r479, %r838, %r837;

dp4a.s32.s32 %r478, %r479, %r444, %r505;

mul.lo.s32 %r839, %r614, %r478;
cvt.rn.f32.s32 %f183, %r839;
fma.rn.ftz.f32 %f184, %f107, %f183, 0f00000000;

prmt.b32 %r482, %r482, 0, 0xba98;


prmt.b32 %r484, %r484, 0, 0xba98;

xor.b32 %r840, %r484, 2139062143;
not.b32 %r841, %r482;
and.b32 %r842, %r642, %r841;
and.b32 %r843, %r840, %r482;
or.b32 %r487, %r843, %r842;

dp4a.s32.s32 %r486, %r487, %r452, %r505;

mul.lo.s32 %r844, %r636, %r486;
cvt.rn.f32.s32 %f185, %r844;
fma.rn.ftz.f32 %f186, %f108, %f185, %f184;

prmt.b32 %r490, %r490, 0, 0xba98;


prmt.b32 %r492, %r492, 0, 0xba98;

xor.b32 %r845, %r492, 2139062143;
not.b32 %r846, %r490;
and.b32 %r847, %r664, %r846;
and.b32 %r848, %r845, %r490;
or.b32 %r495, %r848, %r847;

dp4a.s32.s32 %r494, %r495, %r460, %r505;

mul.lo.s32 %r849, %r659, %r494;
cvt.rn.f32.s32 %f187, %r849;
fma.rn.ftz.f32 %f188, %f109, %f187, %f186;

prmt.b32 %r498, %r498, 0, 0xba98;


prmt.b32 %r500, %r500, 0, 0xba98;

xor.b32 %r850, %r500, 2139062143;
not.b32 %r851, %r498;
and.b32 %r852, %r687, %r851;
and.b32 %r853, %r850, %r498;
or.b32 %r503, %r853, %r852;

dp4a.s32.s32 %r502, %r503, %r468, %r505;

mul.lo.s32 %r854, %r681, %r502;
cvt.rn.f32.s32 %f189, %r854;
fma.rn.ftz.f32 %f190, %f110, %f189, %f188;
fma.rn.ftz.f32 %f311, %f106, %f190, %f311;
add.s32 %r1066, %r1066, 32;
add.s32 %r1065, %r1065, 32;
add.s32 %r1064, %r1064, 32;
add.s32 %r1063, %r1063, 32;
add.s32 %r1062, %r1062, 32;
add.s64 %rd144, %rd144, 440;
add.s32 %r1067, %r1067, 4;
setp.lt.s32 %p4, %r1067, %r4;
@%p4 bra $L__BB149_2;

$L__BB149_3:
mov.u32 %r1047, %tid.y;
setp.eq.s32 %p5, %r1047, 0;
@%p5 bra $L__BB149_5;

mov.u32 %r1049, %tid.x;
mov.u32 %r1048, %tid.y;
mov.u32 %r856, _ZZ13mul_mat_vec_qIL9ggml_type11ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r857, %r1048, 1280, %r856;
shl.b32 %r858, %r1049, 2;
add.s32 %r859, %r857, %r858;
st.shared.f32 [%r859+-1280], %f320;
st.shared.f32 [%r859+-1152], %f319;
st.shared.f32 [%r859+-1024], %f318;
st.shared.f32 [%r859+-896], %f317;
st.shared.f32 [%r859+-768], %f316;
st.shared.f32 [%r859+-640], %f315;
st.shared.f32 [%r859+-512], %f314;
st.shared.f32 [%r859+-384], %f313;
st.shared.f32 [%r859+-256], %f312;
st.shared.f32 [%r859+-128], %f311;

$L__BB149_5:
mov.u32 %r1050, %tid.y;
setp.ne.s32 %p6, %r1050, 0;
bar.sync 0;
@%p6 bra $L__BB149_26;

ld.param.u32 %r1054, [_Z13mul_mat_vec_qIL9ggml_type11ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r1053, [_Z13mul_mat_vec_qIL9ggml_type11ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r1052, [_Z13mul_mat_vec_qIL9ggml_type11ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
mov.u32 %r1051, %tid.x;
mov.u32 %r860, %ctaid.z;
mov.u32 %r861, %ctaid.y;
mov.u32 %r862, %ctaid.x;
shl.b32 %r863, %r862, 1;
mov.u32 %r864, 1;
mad.lo.s32 %r865, %r861, %r1052, %r863;
mad.lo.s32 %r866, %r860, %r1053, %r865;
cvt.u64.u32 %rd31, %r866;
shl.b32 %r868, %r1051, 2;
mov.u32 %r869, _ZZ13mul_mat_vec_qIL9ggml_type11ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r30, %r869, %r868;
mov.u32 %r870, 2;
ld.shared.f32 %f191, [%r30];
add.ftz.f32 %f192, %f191, %f320;
mov.b32 %r871, %f192;
mov.u32 %r872, 31;
mov.u32 %r873, 16;
mov.u32 %r874, -1;
shfl.sync.bfly.b32 %r875|%p7, %r871, %r873, %r872, %r874;
mov.b32 %f193, %r875;
add.ftz.f32 %f194, %f192, %f193;
mov.b32 %r876, %f194;
mov.u32 %r877, 8;
shfl.sync.bfly.b32 %r878|%p8, %r876, %r877, %r872, %r874;
mov.b32 %f195, %r878;
add.ftz.f32 %f196, %f194, %f195;
mov.b32 %r879, %f196;
mov.u32 %r880, 4;
shfl.sync.bfly.b32 %r881|%p9, %r879, %r880, %r872, %r874;
mov.b32 %f197, %r881;
add.ftz.f32 %f198, %f196, %f197;
mov.b32 %r882, %f198;
shfl.sync.bfly.b32 %r883|%p10, %r882, %r870, %r872, %r874;
mov.b32 %f199, %r883;
add.ftz.f32 %f200, %f198, %f199;
mov.b32 %r884, %f200;
shfl.sync.bfly.b32 %r885|%p11, %r884, %r864, %r872, %r874;
mov.b32 %f201, %r885;
add.ftz.f32 %f31, %f200, %f201;
setp.eq.s32 %p12, %r1051, 0;
setp.lt.u32 %p13, %r863, %r1054;
and.pred %p1, %p12, %p13;
mul.wide.u32 %rd121, %r866, 4;
add.s64 %rd32, %rd1, %rd121;
not.pred %p14, %p1;
@%p14 bra $L__BB149_8;

st.global.f32 [%rd32], %f31;

$L__BB149_8:
ld.param.u32 %r1056, [_Z13mul_mat_vec_qIL9ggml_type11ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
mov.u32 %r1055, %tid.x;
ld.shared.f32 %f202, [%r30+128];
add.ftz.f32 %f203, %f202, %f319;
mov.b32 %r886, %f203;
shfl.sync.bfly.b32 %r890|%p15, %r886, %r873, %r872, %r874;
mov.b32 %f204, %r890;
add.ftz.f32 %f205, %f203, %f204;
mov.b32 %r891, %f205;
shfl.sync.bfly.b32 %r893|%p16, %r891, %r877, %r872, %r874;
mov.b32 %f206, %r893;
add.ftz.f32 %f207, %f205, %f206;
mov.b32 %r894, %f207;
shfl.sync.bfly.b32 %r896|%p17, %r894, %r880, %r872, %r874;
mov.b32 %f208, %r896;
add.ftz.f32 %f209, %f207, %f208;
mov.b32 %r897, %f209;
shfl.sync.bfly.b32 %r899|%p18, %r897, %r870, %r872, %r874;
mov.b32 %f210, %r899;
add.ftz.f32 %f211, %f209, %f210;
mov.b32 %r900, %f211;
shfl.sync.bfly.b32 %r902|%p19, %r900, %r864, %r872, %r874;
mov.b32 %f212, %r902;
add.ftz.f32 %f32, %f211, %f212;
setp.eq.s32 %p20, %r1055, 1;
or.b32 %r906, %r863, 1;
setp.lt.u32 %p21, %r906, %r1056;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB149_10;

st.global.f32 [%rd32+4], %f32;

$L__BB149_10:
ld.shared.f32 %f213, [%r30+256];
add.ftz.f32 %f214, %f213, %f318;
mov.b32 %r907, %f214;
mov.u32 %r908, 31;
mov.u32 %r909, 16;
mov.u32 %r910, -1;
shfl.sync.bfly.b32 %r911|%p23, %r907, %r909, %r908, %r910;
mov.b32 %f215, %r911;
add.ftz.f32 %f216, %f214, %f215;
mov.b32 %r912, %f216;
mov.u32 %r913, 8;
shfl.sync.bfly.b32 %r914|%p24, %r912, %r913, %r908, %r910;
mov.b32 %f217, %r914;
add.ftz.f32 %f218, %f216, %f217;
mov.b32 %r915, %f218;
mov.u32 %r916, 4;
shfl.sync.bfly.b32 %r917|%p25, %r915, %r916, %r908, %r910;
mov.b32 %f219, %r917;
add.ftz.f32 %f220, %f218, %f219;
mov.b32 %r918, %f220;
mov.u32 %r919, 2;
shfl.sync.bfly.b32 %r920|%p26, %r918, %r919, %r908, %r910;
mov.b32 %f221, %r920;
add.ftz.f32 %f222, %f220, %f221;
mov.b32 %r921, %f222;
mov.u32 %r922, 1;
shfl.sync.bfly.b32 %r923|%p27, %r921, %r922, %r908, %r910;
mov.b32 %f223, %r923;
add.ftz.f32 %f33, %f222, %f223;
@%p14 bra $L__BB149_12;

ld.param.u32 %r1061, [_Z13mul_mat_vec_qIL9ggml_type11ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
cvt.u64.u32 %rd122, %r1061;
add.s64 %rd123, %rd122, %rd31;
shl.b64 %rd124, %rd123, 2;
add.s64 %rd125, %rd1, %rd124;
st.global.f32 [%rd125], %f33;

$L__BB149_12:
ld.shared.f32 %f224, [%r30+384];
add.ftz.f32 %f225, %f224, %f317;
mov.b32 %r924, %f225;
shfl.sync.bfly.b32 %r928|%p29, %r924, %r909, %r908, %r910;
mov.b32 %f226, %r928;
add.ftz.f32 %f227, %f225, %f226;
mov.b32 %r929, %f227;
shfl.sync.bfly.b32 %r931|%p30, %r929, %r913, %r908, %r910;
mov.b32 %f228, %r931;
add.ftz.f32 %f229, %f227, %f228;
mov.b32 %r932, %f229;
shfl.sync.bfly.b32 %r934|%p31, %r932, %r916, %r908, %r910;
mov.b32 %f230, %r934;
add.ftz.f32 %f231, %f229, %f230;
mov.b32 %r935, %f231;
shfl.sync.bfly.b32 %r937|%p32, %r935, %r919, %r908, %r910;
mov.b32 %f232, %r937;
add.ftz.f32 %f233, %f231, %f232;
mov.b32 %r938, %f233;
shfl.sync.bfly.b32 %r940|%p33, %r938, %r922, %r908, %r910;
mov.b32 %f234, %r940;
add.ftz.f32 %f34, %f233, %f234;
@%p22 bra $L__BB149_14;

ld.param.u32 %r1060, [_Z13mul_mat_vec_qIL9ggml_type11ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
add.s32 %r941, %r1060, 1;
cvt.u64.u32 %rd126, %r941;
add.s64 %rd127, %rd126, %rd31;
shl.b64 %rd128, %rd127, 2;
add.s64 %rd129, %rd1, %rd128;
st.global.f32 [%rd129], %f34;

$L__BB149_14:
ld.param.u32 %r1057, [_Z13mul_mat_vec_qIL9ggml_type11ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
shl.b32 %r942, %r1057, 1;
mov.u32 %r943, 1;
ld.shared.f32 %f235, [%r30+512];
add.ftz.f32 %f236, %f235, %f316;
mov.b32 %r944, %f236;
mov.u32 %r945, 31;
mov.u32 %r946, 16;
mov.u32 %r947, -1;
shfl.sync.bfly.b32 %r948|%p35, %r944, %r946, %r945, %r947;
mov.b32 %f237, %r948;
add.ftz.f32 %f238, %f236, %f237;
mov.b32 %r949, %f238;
mov.u32 %r950, 8;
shfl.sync.bfly.b32 %r951|%p36, %r949, %r950, %r945, %r947;
mov.b32 %f239, %r951;
add.ftz.f32 %f240, %f238, %f239;
mov.b32 %r952, %f240;
mov.u32 %r953, 4;
shfl.sync.bfly.b32 %r954|%p37, %r952, %r953, %r945, %r947;
mov.b32 %f241, %r954;
add.ftz.f32 %f242, %f240, %f241;
mov.b32 %r955, %f242;
mov.u32 %r956, 2;
shfl.sync.bfly.b32 %r957|%p38, %r955, %r956, %r945, %r947;
mov.b32 %f243, %r957;
add.ftz.f32 %f244, %f242, %f243;
mov.b32 %r958, %f244;
shfl.sync.bfly.b32 %r959|%p39, %r958, %r943, %r945, %r947;
mov.b32 %f245, %r959;
add.ftz.f32 %f35, %f244, %f245;
cvt.u64.u32 %rd130, %r942;
add.s64 %rd131, %rd130, %rd31;
shl.b64 %rd132, %rd131, 2;
add.s64 %rd33, %rd1, %rd132;
@%p14 bra $L__BB149_16;

st.global.f32 [%rd33], %f35;

$L__BB149_16:
ld.shared.f32 %f246, [%r30+640];
add.ftz.f32 %f247, %f246, %f315;
mov.b32 %r960, %f247;
shfl.sync.bfly.b32 %r964|%p41, %r960, %r946, %r945, %r947;
mov.b32 %f248, %r964;
add.ftz.f32 %f249, %f247, %f248;
mov.b32 %r965, %f249;
shfl.sync.bfly.b32 %r967|%p42, %r965, %r950, %r945, %r947;
mov.b32 %f250, %r967;
add.ftz.f32 %f251, %f249, %f250;
mov.b32 %r968, %f251;
shfl.sync.bfly.b32 %r970|%p43, %r968, %r953, %r945, %r947;
mov.b32 %f252, %r970;
add.ftz.f32 %f253, %f251, %f252;
mov.b32 %r971, %f253;
shfl.sync.bfly.b32 %r973|%p44, %r971, %r956, %r945, %r947;
mov.b32 %f254, %r973;
add.ftz.f32 %f255, %f253, %f254;
mov.b32 %r974, %f255;
shfl.sync.bfly.b32 %r976|%p45, %r974, %r943, %r945, %r947;
mov.b32 %f256, %r976;
add.ftz.f32 %f36, %f255, %f256;
@%p22 bra $L__BB149_18;

st.global.f32 [%rd33+4], %f36;

$L__BB149_18:
ld.param.u32 %r1058, [_Z13mul_mat_vec_qIL9ggml_type11ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
mul.lo.s32 %r31, %r1058, 3;
ld.shared.f32 %f257, [%r30+768];
add.ftz.f32 %f258, %f257, %f314;
mov.b32 %r977, %f258;
mov.u32 %r978, 31;
mov.u32 %r979, 16;
mov.u32 %r980, -1;
shfl.sync.bfly.b32 %r981|%p47, %r977, %r979, %r978, %r980;
mov.b32 %f259, %r981;
add.ftz.f32 %f260, %f258, %f259;
mov.b32 %r982, %f260;
mov.u32 %r983, 8;
shfl.sync.bfly.b32 %r984|%p48, %r982, %r983, %r978, %r980;
mov.b32 %f261, %r984;
add.ftz.f32 %f262, %f260, %f261;
mov.b32 %r985, %f262;
mov.u32 %r986, 4;
shfl.sync.bfly.b32 %r987|%p49, %r985, %r986, %r978, %r980;
mov.b32 %f263, %r987;
add.ftz.f32 %f264, %f262, %f263;
mov.b32 %r988, %f264;
mov.u32 %r989, 2;
shfl.sync.bfly.b32 %r990|%p50, %r988, %r989, %r978, %r980;
mov.b32 %f265, %r990;
add.ftz.f32 %f266, %f264, %f265;
mov.b32 %r991, %f266;
mov.u32 %r992, 1;
shfl.sync.bfly.b32 %r993|%p51, %r991, %r992, %r978, %r980;
mov.b32 %f267, %r993;
add.ftz.f32 %f37, %f266, %f267;
@%p14 bra $L__BB149_20;

cvt.u64.u32 %rd133, %r31;
add.s64 %rd134, %rd133, %rd31;
shl.b64 %rd135, %rd134, 2;
add.s64 %rd136, %rd1, %rd135;
st.global.f32 [%rd136], %f37;

$L__BB149_20:
ld.shared.f32 %f268, [%r30+896];
add.ftz.f32 %f269, %f268, %f313;
mov.b32 %r994, %f269;
shfl.sync.bfly.b32 %r998|%p53, %r994, %r979, %r978, %r980;
mov.b32 %f270, %r998;
add.ftz.f32 %f271, %f269, %f270;
mov.b32 %r999, %f271;
shfl.sync.bfly.b32 %r1001|%p54, %r999, %r983, %r978, %r980;
mov.b32 %f272, %r1001;
add.ftz.f32 %f273, %f271, %f272;
mov.b32 %r1002, %f273;
shfl.sync.bfly.b32 %r1004|%p55, %r1002, %r986, %r978, %r980;
mov.b32 %f274, %r1004;
add.ftz.f32 %f275, %f273, %f274;
mov.b32 %r1005, %f275;
shfl.sync.bfly.b32 %r1007|%p56, %r1005, %r989, %r978, %r980;
mov.b32 %f276, %r1007;
add.ftz.f32 %f277, %f275, %f276;
mov.b32 %r1008, %f277;
shfl.sync.bfly.b32 %r1010|%p57, %r1008, %r992, %r978, %r980;
mov.b32 %f278, %r1010;
add.ftz.f32 %f38, %f277, %f278;
@%p22 bra $L__BB149_22;

add.s32 %r1011, %r31, 1;
cvt.u64.u32 %rd137, %r1011;
add.s64 %rd138, %rd137, %rd31;
shl.b64 %rd139, %rd138, 2;
add.s64 %rd140, %rd1, %rd139;
st.global.f32 [%rd140], %f38;

$L__BB149_22:
ld.param.u32 %r1059, [_Z13mul_mat_vec_qIL9ggml_type11ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
shl.b32 %r1012, %r1059, 2;
mov.u32 %r1013, 2;
ld.shared.f32 %f279, [%r30+1024];
add.ftz.f32 %f280, %f279, %f312;
mov.b32 %r1014, %f280;
mov.u32 %r1015, 31;
mov.u32 %r1016, 16;
mov.u32 %r1017, -1;
shfl.sync.bfly.b32 %r1018|%p59, %r1014, %r1016, %r1015, %r1017;
mov.b32 %f281, %r1018;
add.ftz.f32 %f282, %f280, %f281;
mov.b32 %r1019, %f282;
mov.u32 %r1020, 8;
shfl.sync.bfly.b32 %r1021|%p60, %r1019, %r1020, %r1015, %r1017;
mov.b32 %f283, %r1021;
add.ftz.f32 %f284, %f282, %f283;
mov.b32 %r1022, %f284;
mov.u32 %r1023, 4;
shfl.sync.bfly.b32 %r1024|%p61, %r1022, %r1023, %r1015, %r1017;
mov.b32 %f285, %r1024;
add.ftz.f32 %f286, %f284, %f285;
mov.b32 %r1025, %f286;
shfl.sync.bfly.b32 %r1026|%p62, %r1025, %r1013, %r1015, %r1017;
mov.b32 %f287, %r1026;
add.ftz.f32 %f288, %f286, %f287;
mov.b32 %r1027, %f288;
mov.u32 %r1028, 1;
shfl.sync.bfly.b32 %r1029|%p63, %r1027, %r1028, %r1015, %r1017;
mov.b32 %f289, %r1029;
add.ftz.f32 %f39, %f288, %f289;
cvt.u64.u32 %rd141, %r1012;
add.s64 %rd142, %rd141, %rd31;
shl.b64 %rd143, %rd142, 2;
add.s64 %rd34, %rd1, %rd143;
@%p14 bra $L__BB149_24;

st.global.f32 [%rd34], %f39;

$L__BB149_24:
ld.shared.f32 %f290, [%r30+1152];
add.ftz.f32 %f291, %f290, %f311;
mov.b32 %r1030, %f291;
shfl.sync.bfly.b32 %r1034|%p65, %r1030, %r1016, %r1015, %r1017;
mov.b32 %f292, %r1034;
add.ftz.f32 %f293, %f291, %f292;
mov.b32 %r1035, %f293;
shfl.sync.bfly.b32 %r1037|%p66, %r1035, %r1020, %r1015, %r1017;
mov.b32 %f294, %r1037;
add.ftz.f32 %f295, %f293, %f294;
mov.b32 %r1038, %f295;
shfl.sync.bfly.b32 %r1040|%p67, %r1038, %r1023, %r1015, %r1017;
mov.b32 %f296, %r1040;
add.ftz.f32 %f297, %f295, %f296;
mov.b32 %r1041, %f297;
shfl.sync.bfly.b32 %r1043|%p68, %r1041, %r1013, %r1015, %r1017;
mov.b32 %f298, %r1043;
add.ftz.f32 %f299, %f297, %f298;
mov.b32 %r1044, %f299;
shfl.sync.bfly.b32 %r1046|%p69, %r1044, %r1028, %r1015, %r1017;
mov.b32 %f300, %r1046;
add.ftz.f32 %f40, %f299, %f300;
@%p22 bra $L__BB149_26;

st.global.f32 [%rd34+4], %f40;

$L__BB149_26:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type11ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type11ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type11ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type11ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type11ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type11ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type11ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type11ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type11ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<83>;
.reg .b16 %rs<37>;
.reg .f32 %f<385>;
.reg .b32 %r<1221>;
.reg .b64 %rd<159>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type11ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

ld.param.u64 %rd35, [_Z13mul_mat_vec_qIL9ggml_type11ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd36, [_Z13mul_mat_vec_qIL9ggml_type11ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd37, [_Z13mul_mat_vec_qIL9ggml_type11ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r51, [_Z13mul_mat_vec_qIL9ggml_type11ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type11ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type11ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type11ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type11ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r48, [_Z13mul_mat_vec_qIL9ggml_type11ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r49, [_Z13mul_mat_vec_qIL9ggml_type11ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r46, [_Z13mul_mat_vec_qIL9ggml_type11ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r45, [_Z13mul_mat_vec_qIL9ggml_type11ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type11ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type11ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
cvta.to.global.u64 %rd1, %rd37;
mov.u32 %r1, %tid.y;
shl.b32 %r52, %r1, 5;
mov.u32 %r53, %tid.x;
add.s32 %r2, %r52, %r53;
shr.s32 %r54, %r2, 31;
shr.u32 %r55, %r54, 28;
add.s32 %r56, %r2, %r55;
shr.s32 %r1220, %r56, 4;
shr.u32 %r4, %r51, 8;
setp.ge.s32 %p3, %r1220, %r4;
mov.f32 %f373, 0f00000000;
mov.f32 %f374, %f373;
mov.f32 %f375, %f373;
mov.f32 %f376, %f373;
mov.f32 %f377, %f373;
mov.f32 %f378, %f373;
mov.f32 %f379, %f373;
mov.f32 %f380, %f373;
mov.f32 %f381, %f373;
mov.f32 %f382, %f373;
mov.f32 %f383, %f373;
mov.f32 %f384, %f373;
@%p3 bra $L__BB150_3;

mov.u32 %r57, %ctaid.x;
shl.b32 %r58, %r57, 1;
mov.u32 %r59, %ctaid.y;
mov.u32 %r60, %ctaid.z;
mul.lo.s32 %r61, %r59, %r43;
cvt.u64.u32 %rd38, %r61;
mul.lo.s32 %r62, %r60, %r49;
cvt.u64.u32 %rd39, %r62;
add.s64 %rd2, %rd39, %rd38;
mul.hi.u32 %r63, %r60, %r45;
add.s32 %r64, %r63, %r60;
shr.u32 %r65, %r64, %r46;
mul.lo.s32 %r66, %r65, %r48;
mul.hi.u32 %r67, %r59, %r39;
add.s32 %r68, %r67, %r59;
shr.u32 %r69, %r68, %r40;
mul.lo.s32 %r70, %r69, %r42;
cvta.to.global.u64 %rd158, %rd35;
or.b32 %r71, %r58, 1;
and.b32 %r75, %r56, -16;
sub.s32 %r76, %r2, %r75;
shr.s32 %r77, %r76, 31;
shr.u32 %r78, %r77, 29;
add.s32 %r79, %r76, %r78;
and.b32 %r80, %r79, -8;
sub.s32 %r81, %r76, %r80;
shr.s32 %r82, %r81, 31;
shr.u32 %r83, %r82, 30;
add.s32 %r84, %r81, %r83;
shr.s32 %r85, %r84, 2;
add.s32 %r86, %r80, %r85;
shr.s32 %r87, %r86, 31;
shr.u32 %r88, %r87, 29;
add.s32 %r89, %r86, %r88;
and.b32 %r90, %r89, -8;
sub.s32 %r91, %r86, %r90;
cvt.s64.s32 %rd40, %r91;
shr.s32 %r92, %r79, 3;
shl.b32 %r93, %r92, 2;
shl.b32 %r94, %r76, 1;
shl.b32 %r95, %r81, 1;
cvt.s64.s32 %rd4, %r81;
shr.s32 %r96, %r89, 3;
shl.b32 %r5, %r96, 2;
cvt.s64.s32 %rd5, %r93;
shr.u32 %r97, %r87, 30;
add.s32 %r98, %r86, %r97;
and.b32 %r99, %r98, -4;
sub.s32 %r100, %r86, %r99;
shr.s32 %r101, %r98, 2;
shl.b32 %r6, %r101, 1;
add.s32 %r102, %r100, 8;
cvt.s64.s32 %rd41, %r102;
add.s32 %r103, %r86, 2;
shr.s32 %r104, %r103, 31;
shr.u32 %r105, %r104, 29;
add.s32 %r106, %r103, %r105;
and.b32 %r107, %r106, -8;
sub.s32 %r108, %r103, %r107;
shr.s32 %r109, %r106, 3;
shl.b32 %r7, %r109, 2;
cvt.s64.s32 %rd42, %r108;
shr.u32 %r110, %r104, 30;
add.s32 %r111, %r103, %r110;
and.b32 %r112, %r111, -4;
sub.s32 %r113, %r103, %r112;
shr.s32 %r114, %r111, 2;
shl.b32 %r8, %r114, 1;
add.s32 %r115, %r113, 8;
cvt.s64.s32 %rd43, %r115;
add.s32 %r116, %r86, 4;
shr.s32 %r117, %r116, 31;
shr.u32 %r118, %r117, 29;
add.s32 %r119, %r116, %r118;
and.b32 %r120, %r119, -8;
sub.s32 %r121, %r116, %r120;
shr.s32 %r122, %r119, 3;
shl.b32 %r9, %r122, 2;
cvt.s64.s32 %rd44, %r121;
shr.u32 %r123, %r117, 30;
add.s32 %r124, %r116, %r123;
and.b32 %r125, %r124, -4;
sub.s32 %r126, %r116, %r125;
shr.s32 %r127, %r124, 2;
shl.b32 %r10, %r127, 1;
add.s32 %r128, %r126, 8;
cvt.s64.s32 %rd45, %r128;
add.s32 %r129, %r86, 6;
shr.s32 %r130, %r129, 31;
shr.u32 %r131, %r130, 29;
add.s32 %r132, %r129, %r131;
and.b32 %r133, %r132, -8;
sub.s32 %r134, %r129, %r133;
shr.s32 %r135, %r132, 3;
shl.b32 %r11, %r135, 2;
cvt.s64.s32 %rd46, %r134;
shr.u32 %r136, %r130, 30;
add.s32 %r137, %r129, %r136;
and.b32 %r138, %r137, -4;
sub.s32 %r139, %r129, %r138;
shr.s32 %r140, %r137, 2;
shl.b32 %r12, %r140, 1;
add.s32 %r141, %r139, 8;
cvt.s64.s32 %rd47, %r141;
shl.b32 %r142, %r37, 1;
shl.b32 %r1214, %r1220, 3;
add.s32 %r1219, %r1214, %r142;
mad.lo.s32 %r1218, %r37, 3, %r1214;
shl.b32 %r143, %r37, 2;
add.s32 %r1217, %r1214, %r143;
mad.lo.s32 %r1216, %r37, 5, %r1214;
add.s32 %r1215, %r37, %r1214;
add.s32 %r144, %r1220, %r66;
add.s32 %r145, %r144, %r70;
mad.lo.s32 %r146, %r71, %r36, %r145;
mul.wide.s32 %rd6, %r146, 110;
add.s64 %rd48, %rd6, %rd47;
add.s64 %rd7, %rd48, 96;
add.s64 %rd49, %rd6, %rd46;
add.s64 %rd8, %rd49, 96;
add.s64 %rd50, %rd6, %rd45;
add.s64 %rd9, %rd50, 96;
add.s64 %rd51, %rd6, %rd44;
add.s64 %rd10, %rd51, 96;
add.s64 %rd52, %rd6, %rd43;
add.s64 %rd11, %rd52, 96;
add.s64 %rd53, %rd6, %rd42;
add.s64 %rd12, %rd53, 96;
add.s64 %rd54, %rd6, %rd41;
add.s64 %rd13, %rd54, 96;
add.s64 %rd55, %rd6, %rd40;
add.s64 %rd14, %rd55, 96;
mad.lo.s32 %r147, %r58, %r36, %r1220;
add.s32 %r148, %r147, %r66;
add.s32 %r149, %r148, %r70;
mul.wide.s32 %rd15, %r149, 110;
mul.wide.s32 %rd56, %r94, 2;
add.s64 %rd16, %rd15, %rd56;
mul.wide.s32 %rd57, %r95, 2;
add.s64 %rd17, %rd6, %rd57;
add.s64 %rd58, %rd15, %rd57;
add.s64 %rd18, %rd58, 2;
add.s64 %rd59, %rd6, %rd56;
add.s64 %rd19, %rd59, 32;
add.s64 %rd60, %rd15, %rd47;
add.s64 %rd20, %rd60, 96;
add.s64 %rd61, %rd15, %rd46;
add.s64 %rd21, %rd61, 96;
add.s64 %rd62, %rd15, %rd45;
add.s64 %rd22, %rd62, 96;
add.s64 %rd63, %rd15, %rd44;
add.s64 %rd23, %rd63, 96;
add.s64 %rd64, %rd15, %rd43;
add.s64 %rd24, %rd64, 96;
add.s64 %rd65, %rd15, %rd42;
add.s64 %rd25, %rd65, 96;
add.s64 %rd66, %rd15, %rd41;
add.s64 %rd26, %rd66, 96;
add.s64 %rd67, %rd15, %rd40;
add.s64 %rd27, %rd67, 96;
cvta.to.global.u64 %rd28, %rd36;
mov.f32 %f373, 0f00000000;
shl.b64 %rd76, %rd4, 2;

$L__BB150_2:
cvt.u64.u32 %rd68, %r1214;
add.s64 %rd69, %rd2, %rd68;
add.s64 %rd70, %rd158, %rd15;
ld.global.nc.u16 %rs1, [%rd70+108];

{ cvt.f32.f16 %f73, %rs1;}


add.s64 %rd71, %rd158, %rd16;
add.s64 %rd72, %rd158, %rd18;
ld.global.nc.u16 %rs13, [%rd72+-2];
ld.global.nc.u16 %rs14, [%rd72];
mov.b32 %r582, {%rs13, %rs14};
not.b32 %r583, %r582;
ld.global.nc.u16 %rs15, [%rd71+34];
ld.global.nc.u16 %rs16, [%rd71+32];
add.s64 %rd73, %rd69, %rd5;
mul.lo.s64 %rd74, %rd73, 36;
add.s64 %rd75, %rd28, %rd74;
add.s64 %rd77, %rd75, %rd76;
ld.global.nc.u32 %r160, [%rd77+4];
ld.global.nc.u32 %r150, [%rd75];

{.reg .f16 low,high;
mov.b32 {low,high},%r150;
cvt.f32.f16 %f74, low;}


ld.global.nc.u32 %r168, [%rd77+40];
ld.global.nc.u32 %r151, [%rd75+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r151;
cvt.f32.f16 %f75, low;}


ld.global.nc.u32 %r176, [%rd77+76];
ld.global.nc.u32 %r152, [%rd75+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r152;
cvt.f32.f16 %f76, low;}


ld.global.nc.u32 %r184, [%rd77+112];
ld.global.nc.u32 %r153, [%rd75+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r153;
cvt.f32.f16 %f77, low;}


mov.b32 %r584, {%rs16, %rs15};
cvt.u32.u64 %r585, %rd5;
shr.s32 %r586, %r583, %r585;
add.s64 %rd78, %rd158, %rd27;
ld.global.nc.u8 %rs17, [%rd78];
cvt.u32.u16 %r587, %rs17;
and.b32 %r588, %r587, 255;
shr.u32 %r589, %r588, %r5;
and.b32 %r590, %r589, 15;
add.s64 %rd79, %rd158, %rd26;
ld.global.nc.u8 %rs18, [%rd79];
cvt.u32.u16 %r591, %rs18;
and.b32 %r592, %r591, 255;
shr.u32 %r593, %r592, %r6;
and.b32 %r594, %r593, 3;
bfi.b32 %r595, %r594, %r590, 4, 2;
add.s32 %r596, %r595, -32;
and.b32 %r516, %r584, 50529027;
shl.b32 %r597, %r586, 2;
and.b32 %r598, %r597, 67372036;
or.b32 %r599, %r516, -2139062144;
sub.s32 %r600, %r599, %r598;
xor.b32 %r601, %r600, -2139062144;
or.b32 %r602, %r516, %r598;
xor.b32 %r603, %r600, %r516;
and.b32 %r514, %r603, %r602;
mov.u32 %r154, %r514;

prmt.b32 %r154, %r154, 0, 0xba98;

mov.u32 %r156, %r516;

prmt.b32 %r156, %r156, 0, 0xba98;

xor.b32 %r604, %r156, 2139062143;
not.b32 %r605, %r154;
and.b32 %r606, %r601, %r605;
and.b32 %r607, %r604, %r154;
or.b32 %r159, %r607, %r606;
mov.u32 %r581, 0;

dp4a.s32.s32 %r158, %r159, %r160, %r581;

mul.lo.s32 %r608, %r596, %r158;
cvt.rn.f32.s32 %f133, %r608;
fma.rn.ftz.f32 %f134, %f74, %f133, 0f00000000;
add.s64 %rd80, %rd158, %rd25;
ld.global.nc.u8 %rs19, [%rd80];
cvt.u32.u16 %r609, %rs19;
and.b32 %r610, %r609, 255;
shr.u32 %r611, %r610, %r7;
and.b32 %r612, %r611, 15;
add.s64 %rd81, %rd158, %rd24;
ld.global.nc.u8 %rs20, [%rd81];
cvt.u32.u16 %r613, %rs20;
and.b32 %r614, %r613, 255;
shr.u32 %r615, %r614, %r8;
and.b32 %r616, %r615, 3;
bfi.b32 %r617, %r616, %r612, 4, 2;
add.s32 %r618, %r617, -32;
shr.u32 %r619, %r584, 2;
and.b32 %r524, %r619, 50529027;
shl.b32 %r620, %r586, 1;
and.b32 %r621, %r620, 67372036;
or.b32 %r622, %r524, -2139062144;
sub.s32 %r623, %r622, %r621;
xor.b32 %r624, %r623, -2139062144;
or.b32 %r625, %r524, %r621;
xor.b32 %r626, %r623, %r524;
and.b32 %r522, %r626, %r625;
mov.u32 %r162, %r522;

prmt.b32 %r162, %r162, 0, 0xba98;

mov.u32 %r164, %r524;

prmt.b32 %r164, %r164, 0, 0xba98;

xor.b32 %r627, %r164, 2139062143;
not.b32 %r628, %r162;
and.b32 %r629, %r624, %r628;
and.b32 %r630, %r627, %r162;
or.b32 %r167, %r630, %r629;

dp4a.s32.s32 %r166, %r167, %r168, %r581;

mul.lo.s32 %r631, %r618, %r166;
cvt.rn.f32.s32 %f135, %r631;
fma.rn.ftz.f32 %f136, %f75, %f135, %f134;
add.s64 %rd82, %rd158, %rd23;
ld.global.nc.u8 %rs21, [%rd82];
cvt.u32.u16 %r632, %rs21;
and.b32 %r633, %r632, 255;
shr.u32 %r634, %r633, %r9;
and.b32 %r635, %r634, 15;
add.s64 %rd83, %rd158, %rd22;
ld.global.nc.u8 %rs22, [%rd83];
cvt.u32.u16 %r636, %rs22;
and.b32 %r637, %r636, 255;
shr.u32 %r638, %r637, %r10;
and.b32 %r639, %r638, 3;
bfi.b32 %r640, %r639, %r635, 4, 2;
add.s32 %r641, %r640, -32;
shr.u32 %r642, %r584, 4;
and.b32 %r532, %r642, 50529027;
and.b32 %r643, %r586, 67372036;
or.b32 %r644, %r532, -2139062144;
sub.s32 %r645, %r644, %r643;
xor.b32 %r646, %r645, -2139062144;
or.b32 %r647, %r532, %r643;
xor.b32 %r648, %r645, %r532;
and.b32 %r530, %r648, %r647;
mov.u32 %r170, %r530;

prmt.b32 %r170, %r170, 0, 0xba98;

mov.u32 %r172, %r532;

prmt.b32 %r172, %r172, 0, 0xba98;

xor.b32 %r649, %r172, 2139062143;
not.b32 %r650, %r170;
and.b32 %r651, %r646, %r650;
and.b32 %r652, %r649, %r170;
or.b32 %r175, %r652, %r651;

dp4a.s32.s32 %r174, %r175, %r176, %r581;

mul.lo.s32 %r653, %r641, %r174;
cvt.rn.f32.s32 %f137, %r653;
fma.rn.ftz.f32 %f138, %f76, %f137, %f136;
add.s64 %rd84, %rd158, %rd21;
ld.global.nc.u8 %rs23, [%rd84];
cvt.u32.u16 %r654, %rs23;
and.b32 %r655, %r654, 255;
shr.u32 %r656, %r655, %r11;
and.b32 %r657, %r656, 15;
add.s64 %rd85, %rd158, %rd20;
ld.global.nc.u8 %rs24, [%rd85];
cvt.u32.u16 %r658, %rs24;
and.b32 %r659, %r658, 255;
shr.u32 %r660, %r659, %r12;
and.b32 %r661, %r660, 3;
bfi.b32 %r662, %r661, %r657, 4, 2;
add.s32 %r663, %r662, -32;
shr.u32 %r664, %r584, 6;
and.b32 %r540, %r664, 50529027;
shr.u32 %r665, %r586, 1;
and.b32 %r666, %r665, 67372036;
or.b32 %r667, %r540, -2139062144;
sub.s32 %r668, %r667, %r666;
xor.b32 %r669, %r668, -2139062144;
or.b32 %r670, %r540, %r666;
xor.b32 %r671, %r668, %r540;
and.b32 %r538, %r671, %r670;
mov.u32 %r178, %r538;

prmt.b32 %r178, %r178, 0, 0xba98;

mov.u32 %r180, %r540;

prmt.b32 %r180, %r180, 0, 0xba98;

xor.b32 %r672, %r180, 2139062143;
not.b32 %r673, %r178;
and.b32 %r674, %r669, %r673;
and.b32 %r675, %r672, %r178;
or.b32 %r183, %r675, %r674;

dp4a.s32.s32 %r182, %r183, %r184, %r581;

mul.lo.s32 %r676, %r663, %r182;
cvt.rn.f32.s32 %f139, %r676;
fma.rn.ftz.f32 %f140, %f77, %f139, %f138;
fma.rn.ftz.f32 %f384, %f73, %f140, %f384;
add.s64 %rd86, %rd158, %rd6;
ld.global.nc.u16 %rs2, [%rd86+108];

{ cvt.f32.f16 %f78, %rs2;}


add.s64 %rd87, %rd158, %rd19;
add.s64 %rd88, %rd158, %rd17;
ld.global.nc.u16 %rs25, [%rd88];
ld.global.nc.u16 %rs26, [%rd88+2];
mov.b32 %r677, {%rs25, %rs26};
not.b32 %r678, %r677;
ld.global.nc.u16 %rs27, [%rd87+2];
ld.global.nc.u16 %rs28, [%rd87];

{.reg .f16 low,high;
mov.b32 {low,high},%r150;
cvt.f32.f16 %f79, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r151;
cvt.f32.f16 %f80, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r152;
cvt.f32.f16 %f81, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r153;
cvt.f32.f16 %f82, low;}


mov.b32 %r679, {%rs28, %rs27};
shr.s32 %r680, %r678, %r585;
add.s64 %rd89, %rd158, %rd14;
ld.global.nc.u8 %rs29, [%rd89];
cvt.u32.u16 %r681, %rs29;
and.b32 %r682, %r681, 255;
shr.u32 %r683, %r682, %r5;
and.b32 %r684, %r683, 15;
add.s64 %rd90, %rd158, %rd13;
ld.global.nc.u8 %rs30, [%rd90];
cvt.u32.u16 %r685, %rs30;
and.b32 %r686, %r685, 255;
shr.u32 %r687, %r686, %r6;
and.b32 %r688, %r687, 3;
bfi.b32 %r689, %r688, %r684, 4, 2;
add.s32 %r690, %r689, -32;
and.b32 %r552, %r679, 50529027;
shl.b32 %r691, %r680, 2;
and.b32 %r692, %r691, 67372036;
or.b32 %r693, %r552, -2139062144;
sub.s32 %r694, %r693, %r692;
xor.b32 %r695, %r694, -2139062144;
or.b32 %r696, %r552, %r692;
xor.b32 %r697, %r694, %r552;
and.b32 %r550, %r697, %r696;
mov.u32 %r190, %r550;

prmt.b32 %r190, %r190, 0, 0xba98;

mov.u32 %r192, %r552;

prmt.b32 %r192, %r192, 0, 0xba98;

xor.b32 %r698, %r192, 2139062143;
not.b32 %r699, %r190;
and.b32 %r700, %r695, %r699;
and.b32 %r701, %r698, %r190;
or.b32 %r195, %r701, %r700;

dp4a.s32.s32 %r194, %r195, %r160, %r581;

mul.lo.s32 %r702, %r690, %r194;
cvt.rn.f32.s32 %f141, %r702;
fma.rn.ftz.f32 %f142, %f79, %f141, 0f00000000;
add.s64 %rd91, %rd158, %rd12;
ld.global.nc.u8 %rs31, [%rd91];
cvt.u32.u16 %r703, %rs31;
and.b32 %r704, %r703, 255;
shr.u32 %r705, %r704, %r7;
and.b32 %r706, %r705, 15;
add.s64 %rd92, %rd158, %rd11;
ld.global.nc.u8 %rs32, [%rd92];
cvt.u32.u16 %r707, %rs32;
and.b32 %r708, %r707, 255;
shr.u32 %r709, %r708, %r8;
and.b32 %r710, %r709, 3;
bfi.b32 %r711, %r710, %r706, 4, 2;
add.s32 %r712, %r711, -32;
shr.u32 %r713, %r679, 2;
and.b32 %r560, %r713, 50529027;
shl.b32 %r714, %r680, 1;
and.b32 %r715, %r714, 67372036;
or.b32 %r716, %r560, -2139062144;
sub.s32 %r717, %r716, %r715;
xor.b32 %r718, %r717, -2139062144;
or.b32 %r719, %r560, %r715;
xor.b32 %r720, %r717, %r560;
and.b32 %r558, %r720, %r719;
mov.u32 %r198, %r558;

prmt.b32 %r198, %r198, 0, 0xba98;

mov.u32 %r200, %r560;

prmt.b32 %r200, %r200, 0, 0xba98;

xor.b32 %r721, %r200, 2139062143;
not.b32 %r722, %r198;
and.b32 %r723, %r718, %r722;
and.b32 %r724, %r721, %r198;
or.b32 %r203, %r724, %r723;

dp4a.s32.s32 %r202, %r203, %r168, %r581;

mul.lo.s32 %r725, %r712, %r202;
cvt.rn.f32.s32 %f143, %r725;
fma.rn.ftz.f32 %f144, %f80, %f143, %f142;
add.s64 %rd93, %rd158, %rd10;
ld.global.nc.u8 %rs33, [%rd93];
cvt.u32.u16 %r726, %rs33;
and.b32 %r727, %r726, 255;
shr.u32 %r728, %r727, %r9;
and.b32 %r729, %r728, 15;
add.s64 %rd94, %rd158, %rd9;
ld.global.nc.u8 %rs34, [%rd94];
cvt.u32.u16 %r730, %rs34;
and.b32 %r731, %r730, 255;
shr.u32 %r732, %r731, %r10;
and.b32 %r733, %r732, 3;
bfi.b32 %r734, %r733, %r729, 4, 2;
add.s32 %r735, %r734, -32;
shr.u32 %r736, %r679, 4;
and.b32 %r568, %r736, 50529027;
and.b32 %r737, %r680, 67372036;
or.b32 %r738, %r568, -2139062144;
sub.s32 %r739, %r738, %r737;
xor.b32 %r740, %r739, -2139062144;
or.b32 %r741, %r568, %r737;
xor.b32 %r742, %r739, %r568;
and.b32 %r566, %r742, %r741;
mov.u32 %r206, %r566;

prmt.b32 %r206, %r206, 0, 0xba98;

mov.u32 %r208, %r568;

prmt.b32 %r208, %r208, 0, 0xba98;

xor.b32 %r743, %r208, 2139062143;
not.b32 %r744, %r206;
and.b32 %r745, %r740, %r744;
and.b32 %r746, %r743, %r206;
or.b32 %r211, %r746, %r745;

dp4a.s32.s32 %r210, %r211, %r176, %r581;

mul.lo.s32 %r747, %r735, %r210;
cvt.rn.f32.s32 %f145, %r747;
fma.rn.ftz.f32 %f146, %f81, %f145, %f144;
add.s64 %rd95, %rd158, %rd8;
ld.global.nc.u8 %rs35, [%rd95];
cvt.u32.u16 %r748, %rs35;
and.b32 %r749, %r748, 255;
shr.u32 %r750, %r749, %r11;
and.b32 %r751, %r750, 15;
add.s64 %rd96, %rd158, %rd7;
ld.global.nc.u8 %rs36, [%rd96];
cvt.u32.u16 %r752, %rs36;
and.b32 %r753, %r752, 255;
shr.u32 %r754, %r753, %r12;
and.b32 %r755, %r754, 3;
bfi.b32 %r756, %r755, %r751, 4, 2;
add.s32 %r757, %r756, -32;
shr.u32 %r758, %r679, 6;
and.b32 %r576, %r758, 50529027;
shr.u32 %r759, %r680, 1;
and.b32 %r760, %r759, 67372036;
or.b32 %r761, %r576, -2139062144;
sub.s32 %r762, %r761, %r760;
xor.b32 %r763, %r762, -2139062144;
or.b32 %r764, %r576, %r760;
xor.b32 %r765, %r762, %r576;
and.b32 %r574, %r765, %r764;
mov.u32 %r214, %r574;

prmt.b32 %r214, %r214, 0, 0xba98;

mov.u32 %r216, %r576;

prmt.b32 %r216, %r216, 0, 0xba98;

xor.b32 %r766, %r216, 2139062143;
not.b32 %r767, %r214;
and.b32 %r768, %r763, %r767;
and.b32 %r769, %r766, %r214;
or.b32 %r219, %r769, %r768;

dp4a.s32.s32 %r218, %r219, %r184, %r581;

mul.lo.s32 %r770, %r757, %r218;
cvt.rn.f32.s32 %f147, %r770;
fma.rn.ftz.f32 %f148, %f82, %f147, %f146;
fma.rn.ftz.f32 %f383, %f78, %f148, %f383;
cvt.u64.u32 %rd97, %r1215;
add.s64 %rd98, %rd2, %rd97;

{ cvt.f32.f16 %f83, %rs1;}


add.s64 %rd99, %rd98, %rd5;
mul.lo.s64 %rd100, %rd99, 36;
add.s64 %rd101, %rd28, %rd100;
add.s64 %rd102, %rd101, %rd76;
ld.global.nc.u32 %r232, [%rd102+4];
ld.global.nc.u32 %r222, [%rd101];

{.reg .f16 low,high;
mov.b32 {low,high},%r222;
cvt.f32.f16 %f84, low;}


ld.global.nc.u32 %r240, [%rd102+40];
ld.global.nc.u32 %r223, [%rd101+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r223;
cvt.f32.f16 %f85, low;}


ld.global.nc.u32 %r248, [%rd102+76];
ld.global.nc.u32 %r224, [%rd101+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r224;
cvt.f32.f16 %f86, low;}


ld.global.nc.u32 %r256, [%rd102+112];
ld.global.nc.u32 %r225, [%rd101+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r225;
cvt.f32.f16 %f87, low;}


mov.u32 %r226, %r514;

prmt.b32 %r226, %r226, 0, 0xba98;

mov.u32 %r228, %r516;

prmt.b32 %r228, %r228, 0, 0xba98;

xor.b32 %r771, %r228, 2139062143;
not.b32 %r772, %r226;
and.b32 %r773, %r601, %r772;
and.b32 %r774, %r771, %r226;
or.b32 %r231, %r774, %r773;

dp4a.s32.s32 %r230, %r231, %r232, %r581;

mul.lo.s32 %r775, %r596, %r230;
cvt.rn.f32.s32 %f149, %r775;
fma.rn.ftz.f32 %f150, %f84, %f149, 0f00000000;
mov.u32 %r234, %r522;

prmt.b32 %r234, %r234, 0, 0xba98;

mov.u32 %r236, %r524;

prmt.b32 %r236, %r236, 0, 0xba98;

xor.b32 %r776, %r236, 2139062143;
not.b32 %r777, %r234;
and.b32 %r778, %r624, %r777;
and.b32 %r779, %r776, %r234;
or.b32 %r239, %r779, %r778;

dp4a.s32.s32 %r238, %r239, %r240, %r581;

mul.lo.s32 %r780, %r618, %r238;
cvt.rn.f32.s32 %f151, %r780;
fma.rn.ftz.f32 %f152, %f85, %f151, %f150;
mov.u32 %r242, %r530;

prmt.b32 %r242, %r242, 0, 0xba98;

mov.u32 %r244, %r532;

prmt.b32 %r244, %r244, 0, 0xba98;

xor.b32 %r781, %r244, 2139062143;
not.b32 %r782, %r242;
and.b32 %r783, %r646, %r782;
and.b32 %r784, %r781, %r242;
or.b32 %r247, %r784, %r783;

dp4a.s32.s32 %r246, %r247, %r248, %r581;

mul.lo.s32 %r785, %r641, %r246;
cvt.rn.f32.s32 %f153, %r785;
fma.rn.ftz.f32 %f154, %f86, %f153, %f152;
mov.u32 %r250, %r538;

prmt.b32 %r250, %r250, 0, 0xba98;

mov.u32 %r252, %r540;

prmt.b32 %r252, %r252, 0, 0xba98;

xor.b32 %r786, %r252, 2139062143;
not.b32 %r787, %r250;
and.b32 %r788, %r669, %r787;
and.b32 %r789, %r786, %r250;
or.b32 %r255, %r789, %r788;

dp4a.s32.s32 %r254, %r255, %r256, %r581;

mul.lo.s32 %r790, %r663, %r254;
cvt.rn.f32.s32 %f155, %r790;
fma.rn.ftz.f32 %f156, %f87, %f155, %f154;
fma.rn.ftz.f32 %f382, %f83, %f156, %f382;

{ cvt.f32.f16 %f88, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r222;
cvt.f32.f16 %f89, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r223;
cvt.f32.f16 %f90, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r224;
cvt.f32.f16 %f91, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r225;
cvt.f32.f16 %f92, low;}


mov.u32 %r262, %r550;

prmt.b32 %r262, %r262, 0, 0xba98;

mov.u32 %r264, %r552;

prmt.b32 %r264, %r264, 0, 0xba98;

xor.b32 %r791, %r264, 2139062143;
not.b32 %r792, %r262;
and.b32 %r793, %r695, %r792;
and.b32 %r794, %r791, %r262;
or.b32 %r267, %r794, %r793;

dp4a.s32.s32 %r266, %r267, %r232, %r581;

mul.lo.s32 %r795, %r690, %r266;
cvt.rn.f32.s32 %f157, %r795;
fma.rn.ftz.f32 %f158, %f89, %f157, 0f00000000;
mov.u32 %r270, %r558;

prmt.b32 %r270, %r270, 0, 0xba98;

mov.u32 %r272, %r560;

prmt.b32 %r272, %r272, 0, 0xba98;

xor.b32 %r796, %r272, 2139062143;
not.b32 %r797, %r270;
and.b32 %r798, %r718, %r797;
and.b32 %r799, %r796, %r270;
or.b32 %r275, %r799, %r798;

dp4a.s32.s32 %r274, %r275, %r240, %r581;

mul.lo.s32 %r800, %r712, %r274;
cvt.rn.f32.s32 %f159, %r800;
fma.rn.ftz.f32 %f160, %f90, %f159, %f158;
mov.u32 %r278, %r566;

prmt.b32 %r278, %r278, 0, 0xba98;

mov.u32 %r280, %r568;

prmt.b32 %r280, %r280, 0, 0xba98;

xor.b32 %r801, %r280, 2139062143;
not.b32 %r802, %r278;
and.b32 %r803, %r740, %r802;
and.b32 %r804, %r801, %r278;
or.b32 %r283, %r804, %r803;

dp4a.s32.s32 %r282, %r283, %r248, %r581;

mul.lo.s32 %r805, %r735, %r282;
cvt.rn.f32.s32 %f161, %r805;
fma.rn.ftz.f32 %f162, %f91, %f161, %f160;
mov.u32 %r286, %r574;

prmt.b32 %r286, %r286, 0, 0xba98;

mov.u32 %r288, %r576;

prmt.b32 %r288, %r288, 0, 0xba98;

xor.b32 %r806, %r288, 2139062143;
not.b32 %r807, %r286;
and.b32 %r808, %r763, %r807;
and.b32 %r809, %r806, %r286;
or.b32 %r291, %r809, %r808;

dp4a.s32.s32 %r290, %r291, %r256, %r581;

mul.lo.s32 %r810, %r757, %r290;
cvt.rn.f32.s32 %f163, %r810;
fma.rn.ftz.f32 %f164, %f92, %f163, %f162;
fma.rn.ftz.f32 %f381, %f88, %f164, %f381;
cvt.u64.u32 %rd103, %r1219;
add.s64 %rd104, %rd2, %rd103;

{ cvt.f32.f16 %f93, %rs1;}


add.s64 %rd105, %rd104, %rd5;
mul.lo.s64 %rd106, %rd105, 36;
add.s64 %rd107, %rd28, %rd106;
add.s64 %rd108, %rd107, %rd76;
ld.global.nc.u32 %r304, [%rd108+4];
ld.global.nc.u32 %r294, [%rd107];

{.reg .f16 low,high;
mov.b32 {low,high},%r294;
cvt.f32.f16 %f94, low;}


ld.global.nc.u32 %r312, [%rd108+40];
ld.global.nc.u32 %r295, [%rd107+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r295;
cvt.f32.f16 %f95, low;}


ld.global.nc.u32 %r320, [%rd108+76];
ld.global.nc.u32 %r296, [%rd107+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r296;
cvt.f32.f16 %f96, low;}


ld.global.nc.u32 %r328, [%rd108+112];
ld.global.nc.u32 %r297, [%rd107+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r297;
cvt.f32.f16 %f97, low;}


mov.u32 %r298, %r514;

prmt.b32 %r298, %r298, 0, 0xba98;

mov.u32 %r300, %r516;

prmt.b32 %r300, %r300, 0, 0xba98;

xor.b32 %r811, %r300, 2139062143;
not.b32 %r812, %r298;
and.b32 %r813, %r601, %r812;
and.b32 %r814, %r811, %r298;
or.b32 %r303, %r814, %r813;

dp4a.s32.s32 %r302, %r303, %r304, %r581;

mul.lo.s32 %r815, %r596, %r302;
cvt.rn.f32.s32 %f165, %r815;
fma.rn.ftz.f32 %f166, %f94, %f165, 0f00000000;
mov.u32 %r306, %r522;

prmt.b32 %r306, %r306, 0, 0xba98;

mov.u32 %r308, %r524;

prmt.b32 %r308, %r308, 0, 0xba98;

xor.b32 %r816, %r308, 2139062143;
not.b32 %r817, %r306;
and.b32 %r818, %r624, %r817;
and.b32 %r819, %r816, %r306;
or.b32 %r311, %r819, %r818;

dp4a.s32.s32 %r310, %r311, %r312, %r581;

mul.lo.s32 %r820, %r618, %r310;
cvt.rn.f32.s32 %f167, %r820;
fma.rn.ftz.f32 %f168, %f95, %f167, %f166;
mov.u32 %r314, %r530;

prmt.b32 %r314, %r314, 0, 0xba98;

mov.u32 %r316, %r532;

prmt.b32 %r316, %r316, 0, 0xba98;

xor.b32 %r821, %r316, 2139062143;
not.b32 %r822, %r314;
and.b32 %r823, %r646, %r822;
and.b32 %r824, %r821, %r314;
or.b32 %r319, %r824, %r823;

dp4a.s32.s32 %r318, %r319, %r320, %r581;

mul.lo.s32 %r825, %r641, %r318;
cvt.rn.f32.s32 %f169, %r825;
fma.rn.ftz.f32 %f170, %f96, %f169, %f168;
mov.u32 %r322, %r538;

prmt.b32 %r322, %r322, 0, 0xba98;

mov.u32 %r324, %r540;

prmt.b32 %r324, %r324, 0, 0xba98;

xor.b32 %r826, %r324, 2139062143;
not.b32 %r827, %r322;
and.b32 %r828, %r669, %r827;
and.b32 %r829, %r826, %r322;
or.b32 %r327, %r829, %r828;

dp4a.s32.s32 %r326, %r327, %r328, %r581;

mul.lo.s32 %r830, %r663, %r326;
cvt.rn.f32.s32 %f171, %r830;
fma.rn.ftz.f32 %f172, %f97, %f171, %f170;
fma.rn.ftz.f32 %f380, %f93, %f172, %f380;

{ cvt.f32.f16 %f98, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r294;
cvt.f32.f16 %f99, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r295;
cvt.f32.f16 %f100, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r296;
cvt.f32.f16 %f101, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r297;
cvt.f32.f16 %f102, low;}


mov.u32 %r334, %r550;

prmt.b32 %r334, %r334, 0, 0xba98;

mov.u32 %r336, %r552;

prmt.b32 %r336, %r336, 0, 0xba98;

xor.b32 %r831, %r336, 2139062143;
not.b32 %r832, %r334;
and.b32 %r833, %r695, %r832;
and.b32 %r834, %r831, %r334;
or.b32 %r339, %r834, %r833;

dp4a.s32.s32 %r338, %r339, %r304, %r581;

mul.lo.s32 %r835, %r690, %r338;
cvt.rn.f32.s32 %f173, %r835;
fma.rn.ftz.f32 %f174, %f99, %f173, 0f00000000;
mov.u32 %r342, %r558;

prmt.b32 %r342, %r342, 0, 0xba98;

mov.u32 %r344, %r560;

prmt.b32 %r344, %r344, 0, 0xba98;

xor.b32 %r836, %r344, 2139062143;
not.b32 %r837, %r342;
and.b32 %r838, %r718, %r837;
and.b32 %r839, %r836, %r342;
or.b32 %r347, %r839, %r838;

dp4a.s32.s32 %r346, %r347, %r312, %r581;

mul.lo.s32 %r840, %r712, %r346;
cvt.rn.f32.s32 %f175, %r840;
fma.rn.ftz.f32 %f176, %f100, %f175, %f174;
mov.u32 %r350, %r566;

prmt.b32 %r350, %r350, 0, 0xba98;

mov.u32 %r352, %r568;

prmt.b32 %r352, %r352, 0, 0xba98;

xor.b32 %r841, %r352, 2139062143;
not.b32 %r842, %r350;
and.b32 %r843, %r740, %r842;
and.b32 %r844, %r841, %r350;
or.b32 %r355, %r844, %r843;

dp4a.s32.s32 %r354, %r355, %r320, %r581;

mul.lo.s32 %r845, %r735, %r354;
cvt.rn.f32.s32 %f177, %r845;
fma.rn.ftz.f32 %f178, %f101, %f177, %f176;
mov.u32 %r358, %r574;

prmt.b32 %r358, %r358, 0, 0xba98;

mov.u32 %r360, %r576;

prmt.b32 %r360, %r360, 0, 0xba98;

xor.b32 %r846, %r360, 2139062143;
not.b32 %r847, %r358;
and.b32 %r848, %r763, %r847;
and.b32 %r849, %r846, %r358;
or.b32 %r363, %r849, %r848;

dp4a.s32.s32 %r362, %r363, %r328, %r581;

mul.lo.s32 %r850, %r757, %r362;
cvt.rn.f32.s32 %f179, %r850;
fma.rn.ftz.f32 %f180, %f102, %f179, %f178;
fma.rn.ftz.f32 %f379, %f98, %f180, %f379;
cvt.u64.u32 %rd109, %r1218;
add.s64 %rd110, %rd2, %rd109;

{ cvt.f32.f16 %f103, %rs1;}


add.s64 %rd111, %rd110, %rd5;
mul.lo.s64 %rd112, %rd111, 36;
add.s64 %rd113, %rd28, %rd112;
add.s64 %rd114, %rd113, %rd76;
ld.global.nc.u32 %r376, [%rd114+4];
ld.global.nc.u32 %r366, [%rd113];

{.reg .f16 low,high;
mov.b32 {low,high},%r366;
cvt.f32.f16 %f104, low;}


ld.global.nc.u32 %r384, [%rd114+40];
ld.global.nc.u32 %r367, [%rd113+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r367;
cvt.f32.f16 %f105, low;}


ld.global.nc.u32 %r392, [%rd114+76];
ld.global.nc.u32 %r368, [%rd113+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r368;
cvt.f32.f16 %f106, low;}


ld.global.nc.u32 %r400, [%rd114+112];
ld.global.nc.u32 %r369, [%rd113+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r369;
cvt.f32.f16 %f107, low;}


mov.u32 %r370, %r514;

prmt.b32 %r370, %r370, 0, 0xba98;

mov.u32 %r372, %r516;

prmt.b32 %r372, %r372, 0, 0xba98;

xor.b32 %r851, %r372, 2139062143;
not.b32 %r852, %r370;
and.b32 %r853, %r601, %r852;
and.b32 %r854, %r851, %r370;
or.b32 %r375, %r854, %r853;

dp4a.s32.s32 %r374, %r375, %r376, %r581;

mul.lo.s32 %r855, %r596, %r374;
cvt.rn.f32.s32 %f181, %r855;
fma.rn.ftz.f32 %f182, %f104, %f181, 0f00000000;
mov.u32 %r378, %r522;

prmt.b32 %r378, %r378, 0, 0xba98;

mov.u32 %r380, %r524;

prmt.b32 %r380, %r380, 0, 0xba98;

xor.b32 %r856, %r380, 2139062143;
not.b32 %r857, %r378;
and.b32 %r858, %r624, %r857;
and.b32 %r859, %r856, %r378;
or.b32 %r383, %r859, %r858;

dp4a.s32.s32 %r382, %r383, %r384, %r581;

mul.lo.s32 %r860, %r618, %r382;
cvt.rn.f32.s32 %f183, %r860;
fma.rn.ftz.f32 %f184, %f105, %f183, %f182;
mov.u32 %r386, %r530;

prmt.b32 %r386, %r386, 0, 0xba98;

mov.u32 %r388, %r532;

prmt.b32 %r388, %r388, 0, 0xba98;

xor.b32 %r861, %r388, 2139062143;
not.b32 %r862, %r386;
and.b32 %r863, %r646, %r862;
and.b32 %r864, %r861, %r386;
or.b32 %r391, %r864, %r863;

dp4a.s32.s32 %r390, %r391, %r392, %r581;

mul.lo.s32 %r865, %r641, %r390;
cvt.rn.f32.s32 %f185, %r865;
fma.rn.ftz.f32 %f186, %f106, %f185, %f184;
mov.u32 %r394, %r538;

prmt.b32 %r394, %r394, 0, 0xba98;

mov.u32 %r396, %r540;

prmt.b32 %r396, %r396, 0, 0xba98;

xor.b32 %r866, %r396, 2139062143;
not.b32 %r867, %r394;
and.b32 %r868, %r669, %r867;
and.b32 %r869, %r866, %r394;
or.b32 %r399, %r869, %r868;

dp4a.s32.s32 %r398, %r399, %r400, %r581;

mul.lo.s32 %r870, %r663, %r398;
cvt.rn.f32.s32 %f187, %r870;
fma.rn.ftz.f32 %f188, %f107, %f187, %f186;
fma.rn.ftz.f32 %f378, %f103, %f188, %f378;

{ cvt.f32.f16 %f108, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r366;
cvt.f32.f16 %f109, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r367;
cvt.f32.f16 %f110, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r368;
cvt.f32.f16 %f111, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r369;
cvt.f32.f16 %f112, low;}


mov.u32 %r406, %r550;

prmt.b32 %r406, %r406, 0, 0xba98;

mov.u32 %r408, %r552;

prmt.b32 %r408, %r408, 0, 0xba98;

xor.b32 %r871, %r408, 2139062143;
not.b32 %r872, %r406;
and.b32 %r873, %r695, %r872;
and.b32 %r874, %r871, %r406;
or.b32 %r411, %r874, %r873;

dp4a.s32.s32 %r410, %r411, %r376, %r581;

mul.lo.s32 %r875, %r690, %r410;
cvt.rn.f32.s32 %f189, %r875;
fma.rn.ftz.f32 %f190, %f109, %f189, 0f00000000;
mov.u32 %r414, %r558;

prmt.b32 %r414, %r414, 0, 0xba98;

mov.u32 %r416, %r560;

prmt.b32 %r416, %r416, 0, 0xba98;

xor.b32 %r876, %r416, 2139062143;
not.b32 %r877, %r414;
and.b32 %r878, %r718, %r877;
and.b32 %r879, %r876, %r414;
or.b32 %r419, %r879, %r878;

dp4a.s32.s32 %r418, %r419, %r384, %r581;

mul.lo.s32 %r880, %r712, %r418;
cvt.rn.f32.s32 %f191, %r880;
fma.rn.ftz.f32 %f192, %f110, %f191, %f190;
mov.u32 %r422, %r566;

prmt.b32 %r422, %r422, 0, 0xba98;

mov.u32 %r424, %r568;

prmt.b32 %r424, %r424, 0, 0xba98;

xor.b32 %r881, %r424, 2139062143;
not.b32 %r882, %r422;
and.b32 %r883, %r740, %r882;
and.b32 %r884, %r881, %r422;
or.b32 %r427, %r884, %r883;

dp4a.s32.s32 %r426, %r427, %r392, %r581;

mul.lo.s32 %r885, %r735, %r426;
cvt.rn.f32.s32 %f193, %r885;
fma.rn.ftz.f32 %f194, %f111, %f193, %f192;
mov.u32 %r430, %r574;

prmt.b32 %r430, %r430, 0, 0xba98;

mov.u32 %r432, %r576;

prmt.b32 %r432, %r432, 0, 0xba98;

xor.b32 %r886, %r432, 2139062143;
not.b32 %r887, %r430;
and.b32 %r888, %r763, %r887;
and.b32 %r889, %r886, %r430;
or.b32 %r435, %r889, %r888;

dp4a.s32.s32 %r434, %r435, %r400, %r581;

mul.lo.s32 %r890, %r757, %r434;
cvt.rn.f32.s32 %f195, %r890;
fma.rn.ftz.f32 %f196, %f112, %f195, %f194;
fma.rn.ftz.f32 %f377, %f108, %f196, %f377;
cvt.u64.u32 %rd115, %r1217;
add.s64 %rd116, %rd2, %rd115;

{ cvt.f32.f16 %f113, %rs1;}


add.s64 %rd117, %rd116, %rd5;
mul.lo.s64 %rd118, %rd117, 36;
add.s64 %rd119, %rd28, %rd118;
add.s64 %rd120, %rd119, %rd76;
ld.global.nc.u32 %r448, [%rd120+4];
ld.global.nc.u32 %r438, [%rd119];

{.reg .f16 low,high;
mov.b32 {low,high},%r438;
cvt.f32.f16 %f114, low;}


ld.global.nc.u32 %r456, [%rd120+40];
ld.global.nc.u32 %r439, [%rd119+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r439;
cvt.f32.f16 %f115, low;}


ld.global.nc.u32 %r464, [%rd120+76];
ld.global.nc.u32 %r440, [%rd119+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r440;
cvt.f32.f16 %f116, low;}


ld.global.nc.u32 %r472, [%rd120+112];
ld.global.nc.u32 %r441, [%rd119+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r441;
cvt.f32.f16 %f117, low;}


mov.u32 %r442, %r514;

prmt.b32 %r442, %r442, 0, 0xba98;

mov.u32 %r444, %r516;

prmt.b32 %r444, %r444, 0, 0xba98;

xor.b32 %r891, %r444, 2139062143;
not.b32 %r892, %r442;
and.b32 %r893, %r601, %r892;
and.b32 %r894, %r891, %r442;
or.b32 %r447, %r894, %r893;

dp4a.s32.s32 %r446, %r447, %r448, %r581;

mul.lo.s32 %r895, %r596, %r446;
cvt.rn.f32.s32 %f197, %r895;
fma.rn.ftz.f32 %f198, %f114, %f197, 0f00000000;
mov.u32 %r450, %r522;

prmt.b32 %r450, %r450, 0, 0xba98;

mov.u32 %r452, %r524;

prmt.b32 %r452, %r452, 0, 0xba98;

xor.b32 %r896, %r452, 2139062143;
not.b32 %r897, %r450;
and.b32 %r898, %r624, %r897;
and.b32 %r899, %r896, %r450;
or.b32 %r455, %r899, %r898;

dp4a.s32.s32 %r454, %r455, %r456, %r581;

mul.lo.s32 %r900, %r618, %r454;
cvt.rn.f32.s32 %f199, %r900;
fma.rn.ftz.f32 %f200, %f115, %f199, %f198;
mov.u32 %r458, %r530;

prmt.b32 %r458, %r458, 0, 0xba98;

mov.u32 %r460, %r532;

prmt.b32 %r460, %r460, 0, 0xba98;

xor.b32 %r901, %r460, 2139062143;
not.b32 %r902, %r458;
and.b32 %r903, %r646, %r902;
and.b32 %r904, %r901, %r458;
or.b32 %r463, %r904, %r903;

dp4a.s32.s32 %r462, %r463, %r464, %r581;

mul.lo.s32 %r905, %r641, %r462;
cvt.rn.f32.s32 %f201, %r905;
fma.rn.ftz.f32 %f202, %f116, %f201, %f200;
mov.u32 %r466, %r538;

prmt.b32 %r466, %r466, 0, 0xba98;

mov.u32 %r468, %r540;

prmt.b32 %r468, %r468, 0, 0xba98;

xor.b32 %r906, %r468, 2139062143;
not.b32 %r907, %r466;
and.b32 %r908, %r669, %r907;
and.b32 %r909, %r906, %r466;
or.b32 %r471, %r909, %r908;

dp4a.s32.s32 %r470, %r471, %r472, %r581;

mul.lo.s32 %r910, %r663, %r470;
cvt.rn.f32.s32 %f203, %r910;
fma.rn.ftz.f32 %f204, %f117, %f203, %f202;
fma.rn.ftz.f32 %f376, %f113, %f204, %f376;

{ cvt.f32.f16 %f118, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r438;
cvt.f32.f16 %f119, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r439;
cvt.f32.f16 %f120, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r440;
cvt.f32.f16 %f121, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r441;
cvt.f32.f16 %f122, low;}


mov.u32 %r478, %r550;

prmt.b32 %r478, %r478, 0, 0xba98;

mov.u32 %r480, %r552;

prmt.b32 %r480, %r480, 0, 0xba98;

xor.b32 %r911, %r480, 2139062143;
not.b32 %r912, %r478;
and.b32 %r913, %r695, %r912;
and.b32 %r914, %r911, %r478;
or.b32 %r483, %r914, %r913;

dp4a.s32.s32 %r482, %r483, %r448, %r581;

mul.lo.s32 %r915, %r690, %r482;
cvt.rn.f32.s32 %f205, %r915;
fma.rn.ftz.f32 %f206, %f119, %f205, 0f00000000;
mov.u32 %r486, %r558;

prmt.b32 %r486, %r486, 0, 0xba98;

mov.u32 %r488, %r560;

prmt.b32 %r488, %r488, 0, 0xba98;

xor.b32 %r916, %r488, 2139062143;
not.b32 %r917, %r486;
and.b32 %r918, %r718, %r917;
and.b32 %r919, %r916, %r486;
or.b32 %r491, %r919, %r918;

dp4a.s32.s32 %r490, %r491, %r456, %r581;

mul.lo.s32 %r920, %r712, %r490;
cvt.rn.f32.s32 %f207, %r920;
fma.rn.ftz.f32 %f208, %f120, %f207, %f206;
mov.u32 %r494, %r566;

prmt.b32 %r494, %r494, 0, 0xba98;

mov.u32 %r496, %r568;

prmt.b32 %r496, %r496, 0, 0xba98;

xor.b32 %r921, %r496, 2139062143;
not.b32 %r922, %r494;
and.b32 %r923, %r740, %r922;
and.b32 %r924, %r921, %r494;
or.b32 %r499, %r924, %r923;

dp4a.s32.s32 %r498, %r499, %r464, %r581;

mul.lo.s32 %r925, %r735, %r498;
cvt.rn.f32.s32 %f209, %r925;
fma.rn.ftz.f32 %f210, %f121, %f209, %f208;
mov.u32 %r502, %r574;

prmt.b32 %r502, %r502, 0, 0xba98;

mov.u32 %r504, %r576;

prmt.b32 %r504, %r504, 0, 0xba98;

xor.b32 %r926, %r504, 2139062143;
not.b32 %r927, %r502;
and.b32 %r928, %r763, %r927;
and.b32 %r929, %r926, %r502;
or.b32 %r507, %r929, %r928;

dp4a.s32.s32 %r506, %r507, %r472, %r581;

mul.lo.s32 %r930, %r757, %r506;
cvt.rn.f32.s32 %f211, %r930;
fma.rn.ftz.f32 %f212, %f122, %f211, %f210;
fma.rn.ftz.f32 %f375, %f118, %f212, %f375;
cvt.u64.u32 %rd121, %r1216;
add.s64 %rd122, %rd2, %rd121;

{ cvt.f32.f16 %f123, %rs1;}


add.s64 %rd123, %rd122, %rd5;
mul.lo.s64 %rd124, %rd123, 36;
add.s64 %rd125, %rd28, %rd124;
add.s64 %rd126, %rd125, %rd76;
ld.global.nc.u32 %r520, [%rd126+4];
ld.global.nc.u32 %r510, [%rd125];

{.reg .f16 low,high;
mov.b32 {low,high},%r510;
cvt.f32.f16 %f124, low;}


ld.global.nc.u32 %r528, [%rd126+40];
ld.global.nc.u32 %r511, [%rd125+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r511;
cvt.f32.f16 %f125, low;}


ld.global.nc.u32 %r536, [%rd126+76];
ld.global.nc.u32 %r512, [%rd125+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r512;
cvt.f32.f16 %f126, low;}


ld.global.nc.u32 %r544, [%rd126+112];
ld.global.nc.u32 %r513, [%rd125+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r513;
cvt.f32.f16 %f127, low;}



prmt.b32 %r514, %r514, 0, 0xba98;


prmt.b32 %r516, %r516, 0, 0xba98;

xor.b32 %r931, %r516, 2139062143;
not.b32 %r932, %r514;
and.b32 %r933, %r601, %r932;
and.b32 %r934, %r931, %r514;
or.b32 %r519, %r934, %r933;

dp4a.s32.s32 %r518, %r519, %r520, %r581;

mul.lo.s32 %r935, %r596, %r518;
cvt.rn.f32.s32 %f213, %r935;
fma.rn.ftz.f32 %f214, %f124, %f213, 0f00000000;

prmt.b32 %r522, %r522, 0, 0xba98;


prmt.b32 %r524, %r524, 0, 0xba98;

xor.b32 %r936, %r524, 2139062143;
not.b32 %r937, %r522;
and.b32 %r938, %r624, %r937;
and.b32 %r939, %r936, %r522;
or.b32 %r527, %r939, %r938;

dp4a.s32.s32 %r526, %r527, %r528, %r581;

mul.lo.s32 %r940, %r618, %r526;
cvt.rn.f32.s32 %f215, %r940;
fma.rn.ftz.f32 %f216, %f125, %f215, %f214;

prmt.b32 %r530, %r530, 0, 0xba98;


prmt.b32 %r532, %r532, 0, 0xba98;

xor.b32 %r941, %r532, 2139062143;
not.b32 %r942, %r530;
and.b32 %r943, %r646, %r942;
and.b32 %r944, %r941, %r530;
or.b32 %r535, %r944, %r943;

dp4a.s32.s32 %r534, %r535, %r536, %r581;

mul.lo.s32 %r945, %r641, %r534;
cvt.rn.f32.s32 %f217, %r945;
fma.rn.ftz.f32 %f218, %f126, %f217, %f216;

prmt.b32 %r538, %r538, 0, 0xba98;


prmt.b32 %r540, %r540, 0, 0xba98;

xor.b32 %r946, %r540, 2139062143;
not.b32 %r947, %r538;
and.b32 %r948, %r669, %r947;
and.b32 %r949, %r946, %r538;
or.b32 %r543, %r949, %r948;

dp4a.s32.s32 %r542, %r543, %r544, %r581;

mul.lo.s32 %r950, %r663, %r542;
cvt.rn.f32.s32 %f219, %r950;
fma.rn.ftz.f32 %f220, %f127, %f219, %f218;
fma.rn.ftz.f32 %f374, %f123, %f220, %f374;

{ cvt.f32.f16 %f128, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r510;
cvt.f32.f16 %f129, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r511;
cvt.f32.f16 %f130, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r512;
cvt.f32.f16 %f131, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r513;
cvt.f32.f16 %f132, low;}



prmt.b32 %r550, %r550, 0, 0xba98;


prmt.b32 %r552, %r552, 0, 0xba98;

xor.b32 %r951, %r552, 2139062143;
not.b32 %r952, %r550;
and.b32 %r953, %r695, %r952;
and.b32 %r954, %r951, %r550;
or.b32 %r555, %r954, %r953;

dp4a.s32.s32 %r554, %r555, %r520, %r581;

mul.lo.s32 %r955, %r690, %r554;
cvt.rn.f32.s32 %f221, %r955;
fma.rn.ftz.f32 %f222, %f129, %f221, 0f00000000;

prmt.b32 %r558, %r558, 0, 0xba98;


prmt.b32 %r560, %r560, 0, 0xba98;

xor.b32 %r956, %r560, 2139062143;
not.b32 %r957, %r558;
and.b32 %r958, %r718, %r957;
and.b32 %r959, %r956, %r558;
or.b32 %r563, %r959, %r958;

dp4a.s32.s32 %r562, %r563, %r528, %r581;

mul.lo.s32 %r960, %r712, %r562;
cvt.rn.f32.s32 %f223, %r960;
fma.rn.ftz.f32 %f224, %f130, %f223, %f222;

prmt.b32 %r566, %r566, 0, 0xba98;


prmt.b32 %r568, %r568, 0, 0xba98;

xor.b32 %r961, %r568, 2139062143;
not.b32 %r962, %r566;
and.b32 %r963, %r740, %r962;
and.b32 %r964, %r961, %r566;
or.b32 %r571, %r964, %r963;

dp4a.s32.s32 %r570, %r571, %r536, %r581;

mul.lo.s32 %r965, %r735, %r570;
cvt.rn.f32.s32 %f225, %r965;
fma.rn.ftz.f32 %f226, %f131, %f225, %f224;

prmt.b32 %r574, %r574, 0, 0xba98;


prmt.b32 %r576, %r576, 0, 0xba98;

xor.b32 %r966, %r576, 2139062143;
not.b32 %r967, %r574;
and.b32 %r968, %r763, %r967;
and.b32 %r969, %r966, %r574;
or.b32 %r579, %r969, %r968;

dp4a.s32.s32 %r578, %r579, %r544, %r581;

mul.lo.s32 %r970, %r757, %r578;
cvt.rn.f32.s32 %f227, %r970;
fma.rn.ftz.f32 %f228, %f132, %f227, %f226;
fma.rn.ftz.f32 %f373, %f128, %f228, %f373;
add.s32 %r1219, %r1219, 32;
add.s32 %r1218, %r1218, 32;
add.s32 %r1217, %r1217, 32;
add.s32 %r1216, %r1216, 32;
add.s32 %r1215, %r1215, 32;
add.s32 %r1214, %r1214, 32;
add.s64 %rd158, %rd158, 440;
add.s32 %r1220, %r1220, 4;
setp.lt.s32 %p4, %r1220, %r4;
@%p4 bra $L__BB150_2;

$L__BB150_3:
mov.u32 %r1198, %tid.y;
setp.eq.s32 %p5, %r1198, 0;
@%p5 bra $L__BB150_5;

mov.u32 %r1200, %tid.x;
mov.u32 %r1199, %tid.y;
mov.u32 %r972, _ZZ13mul_mat_vec_qIL9ggml_type11ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r973, %r1199, 1536, %r972;
shl.b32 %r974, %r1200, 2;
add.s32 %r975, %r973, %r974;
st.shared.f32 [%r975+-1536], %f384;
st.shared.f32 [%r975+-1408], %f383;
st.shared.f32 [%r975+-1280], %f382;
st.shared.f32 [%r975+-1152], %f381;
st.shared.f32 [%r975+-1024], %f380;
st.shared.f32 [%r975+-896], %f379;
st.shared.f32 [%r975+-768], %f378;
st.shared.f32 [%r975+-640], %f377;
st.shared.f32 [%r975+-512], %f376;
st.shared.f32 [%r975+-384], %f375;
st.shared.f32 [%r975+-256], %f374;
st.shared.f32 [%r975+-128], %f373;

$L__BB150_5:
mov.u32 %r1201, %tid.y;
setp.ne.s32 %p6, %r1201, 0;
bar.sync 0;
@%p6 bra $L__BB150_30;

ld.param.u32 %r1205, [_Z13mul_mat_vec_qIL9ggml_type11ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r1204, [_Z13mul_mat_vec_qIL9ggml_type11ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r1203, [_Z13mul_mat_vec_qIL9ggml_type11ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
mov.u32 %r1202, %tid.x;
mov.u32 %r976, %ctaid.z;
mov.u32 %r977, %ctaid.y;
mov.u32 %r978, %ctaid.x;
shl.b32 %r979, %r978, 1;
mov.u32 %r980, 1;
mad.lo.s32 %r981, %r977, %r1203, %r979;
mad.lo.s32 %r982, %r976, %r1204, %r981;
cvt.u64.u32 %rd31, %r982;
shl.b32 %r984, %r1202, 2;
mov.u32 %r985, _ZZ13mul_mat_vec_qIL9ggml_type11ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r33, %r985, %r984;
mov.u32 %r986, 2;
ld.shared.f32 %f229, [%r33];
add.ftz.f32 %f230, %f229, %f384;
mov.b32 %r987, %f230;
mov.u32 %r988, 31;
mov.u32 %r989, 16;
mov.u32 %r990, -1;
shfl.sync.bfly.b32 %r991|%p7, %r987, %r989, %r988, %r990;
mov.b32 %f231, %r991;
add.ftz.f32 %f232, %f230, %f231;
mov.b32 %r992, %f232;
mov.u32 %r993, 8;
shfl.sync.bfly.b32 %r994|%p8, %r992, %r993, %r988, %r990;
mov.b32 %f233, %r994;
add.ftz.f32 %f234, %f232, %f233;
mov.b32 %r995, %f234;
mov.u32 %r996, 4;
shfl.sync.bfly.b32 %r997|%p9, %r995, %r996, %r988, %r990;
mov.b32 %f235, %r997;
add.ftz.f32 %f236, %f234, %f235;
mov.b32 %r998, %f236;
shfl.sync.bfly.b32 %r999|%p10, %r998, %r986, %r988, %r990;
mov.b32 %f237, %r999;
add.ftz.f32 %f238, %f236, %f237;
mov.b32 %r1000, %f238;
shfl.sync.bfly.b32 %r1001|%p11, %r1000, %r980, %r988, %r990;
mov.b32 %f239, %r1001;
add.ftz.f32 %f37, %f238, %f239;
setp.eq.s32 %p12, %r1202, 0;
setp.lt.u32 %p13, %r979, %r1205;
and.pred %p1, %p12, %p13;
mul.wide.u32 %rd127, %r982, 4;
add.s64 %rd32, %rd1, %rd127;
not.pred %p14, %p1;
@%p14 bra $L__BB150_8;

st.global.f32 [%rd32], %f37;

$L__BB150_8:
ld.param.u32 %r1207, [_Z13mul_mat_vec_qIL9ggml_type11ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
mov.u32 %r1206, %tid.x;
ld.shared.f32 %f240, [%r33+128];
add.ftz.f32 %f241, %f240, %f383;
mov.b32 %r1002, %f241;
shfl.sync.bfly.b32 %r1006|%p15, %r1002, %r989, %r988, %r990;
mov.b32 %f242, %r1006;
add.ftz.f32 %f243, %f241, %f242;
mov.b32 %r1007, %f243;
shfl.sync.bfly.b32 %r1009|%p16, %r1007, %r993, %r988, %r990;
mov.b32 %f244, %r1009;
add.ftz.f32 %f245, %f243, %f244;
mov.b32 %r1010, %f245;
shfl.sync.bfly.b32 %r1012|%p17, %r1010, %r996, %r988, %r990;
mov.b32 %f246, %r1012;
add.ftz.f32 %f247, %f245, %f246;
mov.b32 %r1013, %f247;
shfl.sync.bfly.b32 %r1015|%p18, %r1013, %r986, %r988, %r990;
mov.b32 %f248, %r1015;
add.ftz.f32 %f249, %f247, %f248;
mov.b32 %r1016, %f249;
shfl.sync.bfly.b32 %r1018|%p19, %r1016, %r980, %r988, %r990;
mov.b32 %f250, %r1018;
add.ftz.f32 %f38, %f249, %f250;
setp.eq.s32 %p20, %r1206, 1;
or.b32 %r1022, %r979, 1;
setp.lt.u32 %p21, %r1022, %r1207;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB150_10;

st.global.f32 [%rd32+4], %f38;

$L__BB150_10:
ld.shared.f32 %f251, [%r33+256];
add.ftz.f32 %f252, %f251, %f382;
mov.b32 %r1023, %f252;
mov.u32 %r1024, 31;
mov.u32 %r1025, 16;
mov.u32 %r1026, -1;
shfl.sync.bfly.b32 %r1027|%p23, %r1023, %r1025, %r1024, %r1026;
mov.b32 %f253, %r1027;
add.ftz.f32 %f254, %f252, %f253;
mov.b32 %r1028, %f254;
mov.u32 %r1029, 8;
shfl.sync.bfly.b32 %r1030|%p24, %r1028, %r1029, %r1024, %r1026;
mov.b32 %f255, %r1030;
add.ftz.f32 %f256, %f254, %f255;
mov.b32 %r1031, %f256;
mov.u32 %r1032, 4;
shfl.sync.bfly.b32 %r1033|%p25, %r1031, %r1032, %r1024, %r1026;
mov.b32 %f257, %r1033;
add.ftz.f32 %f258, %f256, %f257;
mov.b32 %r1034, %f258;
mov.u32 %r1035, 2;
shfl.sync.bfly.b32 %r1036|%p26, %r1034, %r1035, %r1024, %r1026;
mov.b32 %f259, %r1036;
add.ftz.f32 %f260, %f258, %f259;
mov.b32 %r1037, %f260;
mov.u32 %r1038, 1;
shfl.sync.bfly.b32 %r1039|%p27, %r1037, %r1038, %r1024, %r1026;
mov.b32 %f261, %r1039;
add.ftz.f32 %f39, %f260, %f261;
@%p14 bra $L__BB150_12;

ld.param.u32 %r1213, [_Z13mul_mat_vec_qIL9ggml_type11ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
cvt.u64.u32 %rd128, %r1213;
add.s64 %rd129, %rd128, %rd31;
shl.b64 %rd130, %rd129, 2;
add.s64 %rd131, %rd1, %rd130;
st.global.f32 [%rd131], %f39;

$L__BB150_12:
ld.shared.f32 %f262, [%r33+384];
add.ftz.f32 %f263, %f262, %f381;
mov.b32 %r1040, %f263;
shfl.sync.bfly.b32 %r1044|%p29, %r1040, %r1025, %r1024, %r1026;
mov.b32 %f264, %r1044;
add.ftz.f32 %f265, %f263, %f264;
mov.b32 %r1045, %f265;
shfl.sync.bfly.b32 %r1047|%p30, %r1045, %r1029, %r1024, %r1026;
mov.b32 %f266, %r1047;
add.ftz.f32 %f267, %f265, %f266;
mov.b32 %r1048, %f267;
shfl.sync.bfly.b32 %r1050|%p31, %r1048, %r1032, %r1024, %r1026;
mov.b32 %f268, %r1050;
add.ftz.f32 %f269, %f267, %f268;
mov.b32 %r1051, %f269;
shfl.sync.bfly.b32 %r1053|%p32, %r1051, %r1035, %r1024, %r1026;
mov.b32 %f270, %r1053;
add.ftz.f32 %f271, %f269, %f270;
mov.b32 %r1054, %f271;
shfl.sync.bfly.b32 %r1056|%p33, %r1054, %r1038, %r1024, %r1026;
mov.b32 %f272, %r1056;
add.ftz.f32 %f40, %f271, %f272;
@%p22 bra $L__BB150_14;

ld.param.u32 %r1212, [_Z13mul_mat_vec_qIL9ggml_type11ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
add.s32 %r1057, %r1212, 1;
cvt.u64.u32 %rd132, %r1057;
add.s64 %rd133, %rd132, %rd31;
shl.b64 %rd134, %rd133, 2;
add.s64 %rd135, %rd1, %rd134;
st.global.f32 [%rd135], %f40;

$L__BB150_14:
ld.param.u32 %r1208, [_Z13mul_mat_vec_qIL9ggml_type11ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
shl.b32 %r1058, %r1208, 1;
mov.u32 %r1059, 1;
ld.shared.f32 %f273, [%r33+512];
add.ftz.f32 %f274, %f273, %f380;
mov.b32 %r1060, %f274;
mov.u32 %r1061, 31;
mov.u32 %r1062, 16;
mov.u32 %r1063, -1;
shfl.sync.bfly.b32 %r1064|%p35, %r1060, %r1062, %r1061, %r1063;
mov.b32 %f275, %r1064;
add.ftz.f32 %f276, %f274, %f275;
mov.b32 %r1065, %f276;
mov.u32 %r1066, 8;
shfl.sync.bfly.b32 %r1067|%p36, %r1065, %r1066, %r1061, %r1063;
mov.b32 %f277, %r1067;
add.ftz.f32 %f278, %f276, %f277;
mov.b32 %r1068, %f278;
mov.u32 %r1069, 4;
shfl.sync.bfly.b32 %r1070|%p37, %r1068, %r1069, %r1061, %r1063;
mov.b32 %f279, %r1070;
add.ftz.f32 %f280, %f278, %f279;
mov.b32 %r1071, %f280;
mov.u32 %r1072, 2;
shfl.sync.bfly.b32 %r1073|%p38, %r1071, %r1072, %r1061, %r1063;
mov.b32 %f281, %r1073;
add.ftz.f32 %f282, %f280, %f281;
mov.b32 %r1074, %f282;
shfl.sync.bfly.b32 %r1075|%p39, %r1074, %r1059, %r1061, %r1063;
mov.b32 %f283, %r1075;
add.ftz.f32 %f41, %f282, %f283;
cvt.u64.u32 %rd136, %r1058;
add.s64 %rd137, %rd136, %rd31;
shl.b64 %rd138, %rd137, 2;
add.s64 %rd33, %rd1, %rd138;
@%p14 bra $L__BB150_16;

st.global.f32 [%rd33], %f41;

$L__BB150_16:
ld.shared.f32 %f284, [%r33+640];
add.ftz.f32 %f285, %f284, %f379;
mov.b32 %r1076, %f285;
shfl.sync.bfly.b32 %r1080|%p41, %r1076, %r1062, %r1061, %r1063;
mov.b32 %f286, %r1080;
add.ftz.f32 %f287, %f285, %f286;
mov.b32 %r1081, %f287;
shfl.sync.bfly.b32 %r1083|%p42, %r1081, %r1066, %r1061, %r1063;
mov.b32 %f288, %r1083;
add.ftz.f32 %f289, %f287, %f288;
mov.b32 %r1084, %f289;
shfl.sync.bfly.b32 %r1086|%p43, %r1084, %r1069, %r1061, %r1063;
mov.b32 %f290, %r1086;
add.ftz.f32 %f291, %f289, %f290;
mov.b32 %r1087, %f291;
shfl.sync.bfly.b32 %r1089|%p44, %r1087, %r1072, %r1061, %r1063;
mov.b32 %f292, %r1089;
add.ftz.f32 %f293, %f291, %f292;
mov.b32 %r1090, %f293;
shfl.sync.bfly.b32 %r1092|%p45, %r1090, %r1059, %r1061, %r1063;
mov.b32 %f294, %r1092;
add.ftz.f32 %f42, %f293, %f294;
@%p22 bra $L__BB150_18;

st.global.f32 [%rd33+4], %f42;

$L__BB150_18:
ld.param.u32 %r1209, [_Z13mul_mat_vec_qIL9ggml_type11ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
mul.lo.s32 %r34, %r1209, 3;
ld.shared.f32 %f295, [%r33+768];
add.ftz.f32 %f296, %f295, %f378;
mov.b32 %r1093, %f296;
mov.u32 %r1094, 31;
mov.u32 %r1095, 16;
mov.u32 %r1096, -1;
shfl.sync.bfly.b32 %r1097|%p47, %r1093, %r1095, %r1094, %r1096;
mov.b32 %f297, %r1097;
add.ftz.f32 %f298, %f296, %f297;
mov.b32 %r1098, %f298;
mov.u32 %r1099, 8;
shfl.sync.bfly.b32 %r1100|%p48, %r1098, %r1099, %r1094, %r1096;
mov.b32 %f299, %r1100;
add.ftz.f32 %f300, %f298, %f299;
mov.b32 %r1101, %f300;
mov.u32 %r1102, 4;
shfl.sync.bfly.b32 %r1103|%p49, %r1101, %r1102, %r1094, %r1096;
mov.b32 %f301, %r1103;
add.ftz.f32 %f302, %f300, %f301;
mov.b32 %r1104, %f302;
mov.u32 %r1105, 2;
shfl.sync.bfly.b32 %r1106|%p50, %r1104, %r1105, %r1094, %r1096;
mov.b32 %f303, %r1106;
add.ftz.f32 %f304, %f302, %f303;
mov.b32 %r1107, %f304;
mov.u32 %r1108, 1;
shfl.sync.bfly.b32 %r1109|%p51, %r1107, %r1108, %r1094, %r1096;
mov.b32 %f305, %r1109;
add.ftz.f32 %f43, %f304, %f305;
@%p14 bra $L__BB150_20;

cvt.u64.u32 %rd139, %r34;
add.s64 %rd140, %rd139, %rd31;
shl.b64 %rd141, %rd140, 2;
add.s64 %rd142, %rd1, %rd141;
st.global.f32 [%rd142], %f43;

$L__BB150_20:
ld.shared.f32 %f306, [%r33+896];
add.ftz.f32 %f307, %f306, %f377;
mov.b32 %r1110, %f307;
shfl.sync.bfly.b32 %r1114|%p53, %r1110, %r1095, %r1094, %r1096;
mov.b32 %f308, %r1114;
add.ftz.f32 %f309, %f307, %f308;
mov.b32 %r1115, %f309;
shfl.sync.bfly.b32 %r1117|%p54, %r1115, %r1099, %r1094, %r1096;
mov.b32 %f310, %r1117;
add.ftz.f32 %f311, %f309, %f310;
mov.b32 %r1118, %f311;
shfl.sync.bfly.b32 %r1120|%p55, %r1118, %r1102, %r1094, %r1096;
mov.b32 %f312, %r1120;
add.ftz.f32 %f313, %f311, %f312;
mov.b32 %r1121, %f313;
shfl.sync.bfly.b32 %r1123|%p56, %r1121, %r1105, %r1094, %r1096;
mov.b32 %f314, %r1123;
add.ftz.f32 %f315, %f313, %f314;
mov.b32 %r1124, %f315;
shfl.sync.bfly.b32 %r1126|%p57, %r1124, %r1108, %r1094, %r1096;
mov.b32 %f316, %r1126;
add.ftz.f32 %f44, %f315, %f316;
@%p22 bra $L__BB150_22;

add.s32 %r1127, %r34, 1;
cvt.u64.u32 %rd143, %r1127;
add.s64 %rd144, %rd143, %rd31;
shl.b64 %rd145, %rd144, 2;
add.s64 %rd146, %rd1, %rd145;
st.global.f32 [%rd146], %f44;

$L__BB150_22:
ld.param.u32 %r1210, [_Z13mul_mat_vec_qIL9ggml_type11ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
shl.b32 %r1128, %r1210, 2;
mov.u32 %r1129, 2;
ld.shared.f32 %f317, [%r33+1024];
add.ftz.f32 %f318, %f317, %f376;
mov.b32 %r1130, %f318;
mov.u32 %r1131, 31;
mov.u32 %r1132, 16;
mov.u32 %r1133, -1;
shfl.sync.bfly.b32 %r1134|%p59, %r1130, %r1132, %r1131, %r1133;
mov.b32 %f319, %r1134;
add.ftz.f32 %f320, %f318, %f319;
mov.b32 %r1135, %f320;
mov.u32 %r1136, 8;
shfl.sync.bfly.b32 %r1137|%p60, %r1135, %r1136, %r1131, %r1133;
mov.b32 %f321, %r1137;
add.ftz.f32 %f322, %f320, %f321;
mov.b32 %r1138, %f322;
mov.u32 %r1139, 4;
shfl.sync.bfly.b32 %r1140|%p61, %r1138, %r1139, %r1131, %r1133;
mov.b32 %f323, %r1140;
add.ftz.f32 %f324, %f322, %f323;
mov.b32 %r1141, %f324;
shfl.sync.bfly.b32 %r1142|%p62, %r1141, %r1129, %r1131, %r1133;
mov.b32 %f325, %r1142;
add.ftz.f32 %f326, %f324, %f325;
mov.b32 %r1143, %f326;
mov.u32 %r1144, 1;
shfl.sync.bfly.b32 %r1145|%p63, %r1143, %r1144, %r1131, %r1133;
mov.b32 %f327, %r1145;
add.ftz.f32 %f45, %f326, %f327;
cvt.u64.u32 %rd147, %r1128;
add.s64 %rd148, %rd147, %rd31;
shl.b64 %rd149, %rd148, 2;
add.s64 %rd34, %rd1, %rd149;
@%p14 bra $L__BB150_24;

st.global.f32 [%rd34], %f45;

$L__BB150_24:
ld.shared.f32 %f328, [%r33+1152];
add.ftz.f32 %f329, %f328, %f375;
mov.b32 %r1146, %f329;
shfl.sync.bfly.b32 %r1150|%p65, %r1146, %r1132, %r1131, %r1133;
mov.b32 %f330, %r1150;
add.ftz.f32 %f331, %f329, %f330;
mov.b32 %r1151, %f331;
shfl.sync.bfly.b32 %r1153|%p66, %r1151, %r1136, %r1131, %r1133;
mov.b32 %f332, %r1153;
add.ftz.f32 %f333, %f331, %f332;
mov.b32 %r1154, %f333;
shfl.sync.bfly.b32 %r1156|%p67, %r1154, %r1139, %r1131, %r1133;
mov.b32 %f334, %r1156;
add.ftz.f32 %f335, %f333, %f334;
mov.b32 %r1157, %f335;
shfl.sync.bfly.b32 %r1159|%p68, %r1157, %r1129, %r1131, %r1133;
mov.b32 %f336, %r1159;
add.ftz.f32 %f337, %f335, %f336;
mov.b32 %r1160, %f337;
shfl.sync.bfly.b32 %r1162|%p69, %r1160, %r1144, %r1131, %r1133;
mov.b32 %f338, %r1162;
add.ftz.f32 %f46, %f337, %f338;
@%p22 bra $L__BB150_26;

st.global.f32 [%rd34+4], %f46;

$L__BB150_26:
ld.param.u32 %r1211, [_Z13mul_mat_vec_qIL9ggml_type11ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
mul.lo.s32 %r35, %r1211, 5;
ld.shared.f32 %f339, [%r33+1280];
add.ftz.f32 %f340, %f339, %f374;
mov.b32 %r1163, %f340;
mov.u32 %r1164, 31;
mov.u32 %r1165, 16;
mov.u32 %r1166, -1;
shfl.sync.bfly.b32 %r1167|%p71, %r1163, %r1165, %r1164, %r1166;
mov.b32 %f341, %r1167;
add.ftz.f32 %f342, %f340, %f341;
mov.b32 %r1168, %f342;
mov.u32 %r1169, 8;
shfl.sync.bfly.b32 %r1170|%p72, %r1168, %r1169, %r1164, %r1166;
mov.b32 %f343, %r1170;
add.ftz.f32 %f344, %f342, %f343;
mov.b32 %r1171, %f344;
mov.u32 %r1172, 4;
shfl.sync.bfly.b32 %r1173|%p73, %r1171, %r1172, %r1164, %r1166;
mov.b32 %f345, %r1173;
add.ftz.f32 %f346, %f344, %f345;
mov.b32 %r1174, %f346;
mov.u32 %r1175, 2;
shfl.sync.bfly.b32 %r1176|%p74, %r1174, %r1175, %r1164, %r1166;
mov.b32 %f347, %r1176;
add.ftz.f32 %f348, %f346, %f347;
mov.b32 %r1177, %f348;
mov.u32 %r1178, 1;
shfl.sync.bfly.b32 %r1179|%p75, %r1177, %r1178, %r1164, %r1166;
mov.b32 %f349, %r1179;
add.ftz.f32 %f47, %f348, %f349;
@%p14 bra $L__BB150_28;

cvt.u64.u32 %rd150, %r35;
add.s64 %rd151, %rd150, %rd31;
shl.b64 %rd152, %rd151, 2;
add.s64 %rd153, %rd1, %rd152;
st.global.f32 [%rd153], %f47;

$L__BB150_28:
ld.shared.f32 %f350, [%r33+1408];
add.ftz.f32 %f351, %f350, %f373;
mov.b32 %r1180, %f351;
shfl.sync.bfly.b32 %r1184|%p77, %r1180, %r1165, %r1164, %r1166;
mov.b32 %f352, %r1184;
add.ftz.f32 %f353, %f351, %f352;
mov.b32 %r1185, %f353;
shfl.sync.bfly.b32 %r1187|%p78, %r1185, %r1169, %r1164, %r1166;
mov.b32 %f354, %r1187;
add.ftz.f32 %f355, %f353, %f354;
mov.b32 %r1188, %f355;
shfl.sync.bfly.b32 %r1190|%p79, %r1188, %r1172, %r1164, %r1166;
mov.b32 %f356, %r1190;
add.ftz.f32 %f357, %f355, %f356;
mov.b32 %r1191, %f357;
shfl.sync.bfly.b32 %r1193|%p80, %r1191, %r1175, %r1164, %r1166;
mov.b32 %f358, %r1193;
add.ftz.f32 %f359, %f357, %f358;
mov.b32 %r1194, %f359;
shfl.sync.bfly.b32 %r1196|%p81, %r1194, %r1178, %r1164, %r1166;
mov.b32 %f360, %r1196;
add.ftz.f32 %f48, %f359, %f360;
@%p22 bra $L__BB150_30;

add.s32 %r1197, %r35, 1;
cvt.u64.u32 %rd154, %r1197;
add.s64 %rd155, %rd154, %rd31;
shl.b64 %rd156, %rd155, 2;
add.s64 %rd157, %rd1, %rd156;
st.global.f32 [%rd157], %f48;

$L__BB150_30:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type11ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type11ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type11ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type11ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type11ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type11ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type11ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type11ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type11ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<95>;
.reg .b16 %rs<39>;
.reg .f32 %f<449>;
.reg .b32 %r<1373>;
.reg .b64 %rd<169>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type11ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1792];

ld.param.u64 %rd36, [_Z13mul_mat_vec_qIL9ggml_type11ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd37, [_Z13mul_mat_vec_qIL9ggml_type11ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd38, [_Z13mul_mat_vec_qIL9ggml_type11ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r54, [_Z13mul_mat_vec_qIL9ggml_type11ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type11ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type11ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r45, [_Z13mul_mat_vec_qIL9ggml_type11ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r46, [_Z13mul_mat_vec_qIL9ggml_type11ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r51, [_Z13mul_mat_vec_qIL9ggml_type11ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r52, [_Z13mul_mat_vec_qIL9ggml_type11ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r49, [_Z13mul_mat_vec_qIL9ggml_type11ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r48, [_Z13mul_mat_vec_qIL9ggml_type11ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type11ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type11ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
cvta.to.global.u64 %rd1, %rd38;
mov.u32 %r1, %tid.y;
shl.b32 %r55, %r1, 5;
mov.u32 %r56, %tid.x;
add.s32 %r2, %r55, %r56;
shr.s32 %r57, %r2, 31;
shr.u32 %r58, %r57, 28;
add.s32 %r59, %r2, %r58;
shr.s32 %r1372, %r59, 4;
shr.u32 %r4, %r54, 8;
setp.ge.s32 %p3, %r1372, %r4;
mov.f32 %f435, 0f00000000;
mov.f32 %f436, %f435;
mov.f32 %f437, %f435;
mov.f32 %f438, %f435;
mov.f32 %f439, %f435;
mov.f32 %f440, %f435;
mov.f32 %f441, %f435;
mov.f32 %f442, %f435;
mov.f32 %f443, %f435;
mov.f32 %f444, %f435;
mov.f32 %f445, %f435;
mov.f32 %f446, %f435;
mov.f32 %f447, %f435;
mov.f32 %f448, %f435;
@%p3 bra $L__BB151_3;

mov.u32 %r60, %ctaid.x;
shl.b32 %r61, %r60, 1;
mov.u32 %r62, %ctaid.y;
mov.u32 %r63, %ctaid.z;
mul.lo.s32 %r64, %r62, %r46;
cvt.u64.u32 %rd39, %r64;
mul.lo.s32 %r65, %r63, %r52;
cvt.u64.u32 %rd40, %r65;
add.s64 %rd2, %rd40, %rd39;
mul.hi.u32 %r66, %r63, %r48;
add.s32 %r67, %r66, %r63;
shr.u32 %r68, %r67, %r49;
mul.lo.s32 %r69, %r68, %r51;
mul.hi.u32 %r70, %r62, %r42;
add.s32 %r71, %r70, %r62;
shr.u32 %r72, %r71, %r43;
mul.lo.s32 %r73, %r72, %r45;
cvta.to.global.u64 %rd168, %rd36;
or.b32 %r74, %r61, 1;
and.b32 %r78, %r59, -16;
sub.s32 %r79, %r2, %r78;
shr.s32 %r80, %r79, 31;
shr.u32 %r81, %r80, 29;
add.s32 %r82, %r79, %r81;
and.b32 %r83, %r82, -8;
sub.s32 %r84, %r79, %r83;
shr.s32 %r85, %r84, 31;
shr.u32 %r86, %r85, 30;
add.s32 %r87, %r84, %r86;
shr.s32 %r88, %r87, 2;
add.s32 %r89, %r83, %r88;
shr.s32 %r90, %r89, 31;
shr.u32 %r91, %r90, 29;
add.s32 %r92, %r89, %r91;
and.b32 %r93, %r92, -8;
sub.s32 %r94, %r89, %r93;
cvt.s64.s32 %rd41, %r94;
shr.s32 %r95, %r82, 3;
shl.b32 %r96, %r95, 2;
shl.b32 %r97, %r79, 1;
shl.b32 %r98, %r84, 1;
cvt.s64.s32 %rd4, %r84;
shr.s32 %r99, %r92, 3;
shl.b32 %r5, %r99, 2;
cvt.s64.s32 %rd5, %r96;
shr.u32 %r100, %r90, 30;
add.s32 %r101, %r89, %r100;
and.b32 %r102, %r101, -4;
sub.s32 %r103, %r89, %r102;
shr.s32 %r104, %r101, 2;
shl.b32 %r6, %r104, 1;
add.s32 %r105, %r103, 8;
cvt.s64.s32 %rd42, %r105;
add.s32 %r106, %r89, 2;
shr.s32 %r107, %r106, 31;
shr.u32 %r108, %r107, 29;
add.s32 %r109, %r106, %r108;
and.b32 %r110, %r109, -8;
sub.s32 %r111, %r106, %r110;
shr.s32 %r112, %r109, 3;
shl.b32 %r7, %r112, 2;
cvt.s64.s32 %rd43, %r111;
shr.u32 %r113, %r107, 30;
add.s32 %r114, %r106, %r113;
and.b32 %r115, %r114, -4;
sub.s32 %r116, %r106, %r115;
shr.s32 %r117, %r114, 2;
shl.b32 %r8, %r117, 1;
add.s32 %r118, %r116, 8;
cvt.s64.s32 %rd44, %r118;
add.s32 %r119, %r89, 4;
shr.s32 %r120, %r119, 31;
shr.u32 %r121, %r120, 29;
add.s32 %r122, %r119, %r121;
and.b32 %r123, %r122, -8;
sub.s32 %r124, %r119, %r123;
shr.s32 %r125, %r122, 3;
shl.b32 %r9, %r125, 2;
cvt.s64.s32 %rd45, %r124;
shr.u32 %r126, %r120, 30;
add.s32 %r127, %r119, %r126;
and.b32 %r128, %r127, -4;
sub.s32 %r129, %r119, %r128;
shr.s32 %r130, %r127, 2;
shl.b32 %r10, %r130, 1;
add.s32 %r131, %r129, 8;
cvt.s64.s32 %rd46, %r131;
add.s32 %r132, %r89, 6;
shr.s32 %r133, %r132, 31;
shr.u32 %r134, %r133, 29;
add.s32 %r135, %r132, %r134;
and.b32 %r136, %r135, -8;
sub.s32 %r137, %r132, %r136;
shr.s32 %r138, %r135, 3;
shl.b32 %r11, %r138, 2;
cvt.s64.s32 %rd47, %r137;
shr.u32 %r139, %r133, 30;
add.s32 %r140, %r132, %r139;
and.b32 %r141, %r140, -4;
sub.s32 %r142, %r132, %r141;
shr.s32 %r143, %r140, 2;
shl.b32 %r12, %r143, 1;
add.s32 %r144, %r142, 8;
cvt.s64.s32 %rd48, %r144;
shl.b32 %r145, %r40, 1;
shl.b32 %r1365, %r1372, 3;
add.s32 %r1371, %r1365, %r145;
mad.lo.s32 %r1370, %r40, 3, %r1365;
shl.b32 %r146, %r40, 2;
add.s32 %r1369, %r1365, %r146;
mad.lo.s32 %r1368, %r40, 5, %r1365;
mad.lo.s32 %r1367, %r40, 6, %r1365;
add.s32 %r1366, %r40, %r1365;
add.s32 %r147, %r1372, %r69;
add.s32 %r148, %r147, %r73;
mad.lo.s32 %r149, %r74, %r39, %r148;
mul.wide.s32 %rd6, %r149, 110;
add.s64 %rd49, %rd6, %rd48;
add.s64 %rd7, %rd49, 96;
add.s64 %rd50, %rd6, %rd47;
add.s64 %rd8, %rd50, 96;
add.s64 %rd51, %rd6, %rd46;
add.s64 %rd9, %rd51, 96;
add.s64 %rd52, %rd6, %rd45;
add.s64 %rd10, %rd52, 96;
add.s64 %rd53, %rd6, %rd44;
add.s64 %rd11, %rd53, 96;
add.s64 %rd54, %rd6, %rd43;
add.s64 %rd12, %rd54, 96;
add.s64 %rd55, %rd6, %rd42;
add.s64 %rd13, %rd55, 96;
add.s64 %rd56, %rd6, %rd41;
add.s64 %rd14, %rd56, 96;
mad.lo.s32 %r150, %r61, %r39, %r1372;
add.s32 %r151, %r150, %r69;
add.s32 %r152, %r151, %r73;
mul.wide.s32 %rd15, %r152, 110;
mul.wide.s32 %rd57, %r97, 2;
add.s64 %rd16, %rd15, %rd57;
mul.wide.s32 %rd58, %r98, 2;
add.s64 %rd17, %rd6, %rd58;
add.s64 %rd59, %rd15, %rd58;
add.s64 %rd18, %rd59, 2;
add.s64 %rd60, %rd6, %rd57;
add.s64 %rd19, %rd60, 32;
add.s64 %rd61, %rd15, %rd48;
add.s64 %rd20, %rd61, 96;
add.s64 %rd62, %rd15, %rd47;
add.s64 %rd21, %rd62, 96;
add.s64 %rd63, %rd15, %rd46;
add.s64 %rd22, %rd63, 96;
add.s64 %rd64, %rd15, %rd45;
add.s64 %rd23, %rd64, 96;
add.s64 %rd65, %rd15, %rd44;
add.s64 %rd24, %rd65, 96;
add.s64 %rd66, %rd15, %rd43;
add.s64 %rd25, %rd66, 96;
add.s64 %rd67, %rd15, %rd42;
add.s64 %rd26, %rd67, 96;
add.s64 %rd68, %rd15, %rd41;
add.s64 %rd27, %rd68, 96;
cvta.to.global.u64 %rd28, %rd37;
mov.f32 %f435, 0f00000000;
shl.b64 %rd77, %rd4, 2;

$L__BB151_2:
cvt.u64.u32 %rd69, %r1365;
add.s64 %rd70, %rd2, %rd69;
add.s64 %rd71, %rd168, %rd15;
ld.global.nc.u16 %rs1, [%rd71+108];

{ cvt.f32.f16 %f85, %rs1;}


add.s64 %rd72, %rd168, %rd16;
add.s64 %rd73, %rd168, %rd18;
ld.global.nc.u16 %rs15, [%rd73+-2];
ld.global.nc.u16 %rs16, [%rd73];
mov.b32 %r657, {%rs15, %rs16};
not.b32 %r658, %r657;
ld.global.nc.u16 %rs17, [%rd72+34];
ld.global.nc.u16 %rs18, [%rd72+32];
add.s64 %rd74, %rd70, %rd5;
mul.lo.s64 %rd75, %rd74, 36;
add.s64 %rd76, %rd28, %rd75;
add.s64 %rd78, %rd76, %rd77;
ld.global.nc.u32 %r163, [%rd78+4];
ld.global.nc.u32 %r153, [%rd76];

{.reg .f16 low,high;
mov.b32 {low,high},%r153;
cvt.f32.f16 %f86, low;}


ld.global.nc.u32 %r171, [%rd78+40];
ld.global.nc.u32 %r154, [%rd76+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r154;
cvt.f32.f16 %f87, low;}


ld.global.nc.u32 %r179, [%rd78+76];
ld.global.nc.u32 %r155, [%rd76+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r155;
cvt.f32.f16 %f88, low;}


ld.global.nc.u32 %r187, [%rd78+112];
ld.global.nc.u32 %r156, [%rd76+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r156;
cvt.f32.f16 %f89, low;}


mov.b32 %r659, {%rs18, %rs17};
cvt.u32.u64 %r660, %rd5;
shr.s32 %r661, %r658, %r660;
add.s64 %rd79, %rd168, %rd27;
ld.global.nc.u8 %rs19, [%rd79];
cvt.u32.u16 %r662, %rs19;
and.b32 %r663, %r662, 255;
shr.u32 %r664, %r663, %r5;
and.b32 %r665, %r664, 15;
add.s64 %rd80, %rd168, %rd26;
ld.global.nc.u8 %rs20, [%rd80];
cvt.u32.u16 %r666, %rs20;
and.b32 %r667, %r666, 255;
shr.u32 %r668, %r667, %r6;
and.b32 %r669, %r668, 3;
bfi.b32 %r670, %r669, %r665, 4, 2;
add.s32 %r671, %r670, -32;
and.b32 %r591, %r659, 50529027;
shl.b32 %r672, %r661, 2;
and.b32 %r673, %r672, 67372036;
or.b32 %r674, %r591, -2139062144;
sub.s32 %r675, %r674, %r673;
xor.b32 %r676, %r675, -2139062144;
or.b32 %r677, %r591, %r673;
xor.b32 %r678, %r675, %r591;
and.b32 %r589, %r678, %r677;
mov.u32 %r157, %r589;

prmt.b32 %r157, %r157, 0, 0xba98;

mov.u32 %r159, %r591;

prmt.b32 %r159, %r159, 0, 0xba98;

xor.b32 %r679, %r159, 2139062143;
not.b32 %r680, %r157;
and.b32 %r681, %r676, %r680;
and.b32 %r682, %r679, %r157;
or.b32 %r162, %r682, %r681;
mov.u32 %r656, 0;

dp4a.s32.s32 %r161, %r162, %r163, %r656;

mul.lo.s32 %r683, %r671, %r161;
cvt.rn.f32.s32 %f155, %r683;
fma.rn.ftz.f32 %f156, %f86, %f155, 0f00000000;
add.s64 %rd81, %rd168, %rd25;
ld.global.nc.u8 %rs21, [%rd81];
cvt.u32.u16 %r684, %rs21;
and.b32 %r685, %r684, 255;
shr.u32 %r686, %r685, %r7;
and.b32 %r687, %r686, 15;
add.s64 %rd82, %rd168, %rd24;
ld.global.nc.u8 %rs22, [%rd82];
cvt.u32.u16 %r688, %rs22;
and.b32 %r689, %r688, 255;
shr.u32 %r690, %r689, %r8;
and.b32 %r691, %r690, 3;
bfi.b32 %r692, %r691, %r687, 4, 2;
add.s32 %r693, %r692, -32;
shr.u32 %r694, %r659, 2;
and.b32 %r599, %r694, 50529027;
shl.b32 %r695, %r661, 1;
and.b32 %r696, %r695, 67372036;
or.b32 %r697, %r599, -2139062144;
sub.s32 %r698, %r697, %r696;
xor.b32 %r699, %r698, -2139062144;
or.b32 %r700, %r599, %r696;
xor.b32 %r701, %r698, %r599;
and.b32 %r597, %r701, %r700;
mov.u32 %r165, %r597;

prmt.b32 %r165, %r165, 0, 0xba98;

mov.u32 %r167, %r599;

prmt.b32 %r167, %r167, 0, 0xba98;

xor.b32 %r702, %r167, 2139062143;
not.b32 %r703, %r165;
and.b32 %r704, %r699, %r703;
and.b32 %r705, %r702, %r165;
or.b32 %r170, %r705, %r704;

dp4a.s32.s32 %r169, %r170, %r171, %r656;

mul.lo.s32 %r706, %r693, %r169;
cvt.rn.f32.s32 %f157, %r706;
fma.rn.ftz.f32 %f158, %f87, %f157, %f156;
add.s64 %rd83, %rd168, %rd23;
ld.global.nc.u8 %rs23, [%rd83];
cvt.u32.u16 %r707, %rs23;
and.b32 %r708, %r707, 255;
shr.u32 %r709, %r708, %r9;
and.b32 %r710, %r709, 15;
add.s64 %rd84, %rd168, %rd22;
ld.global.nc.u8 %rs24, [%rd84];
cvt.u32.u16 %r711, %rs24;
and.b32 %r712, %r711, 255;
shr.u32 %r713, %r712, %r10;
and.b32 %r714, %r713, 3;
bfi.b32 %r715, %r714, %r710, 4, 2;
add.s32 %r716, %r715, -32;
shr.u32 %r717, %r659, 4;
and.b32 %r607, %r717, 50529027;
and.b32 %r718, %r661, 67372036;
or.b32 %r719, %r607, -2139062144;
sub.s32 %r720, %r719, %r718;
xor.b32 %r721, %r720, -2139062144;
or.b32 %r722, %r607, %r718;
xor.b32 %r723, %r720, %r607;
and.b32 %r605, %r723, %r722;
mov.u32 %r173, %r605;

prmt.b32 %r173, %r173, 0, 0xba98;

mov.u32 %r175, %r607;

prmt.b32 %r175, %r175, 0, 0xba98;

xor.b32 %r724, %r175, 2139062143;
not.b32 %r725, %r173;
and.b32 %r726, %r721, %r725;
and.b32 %r727, %r724, %r173;
or.b32 %r178, %r727, %r726;

dp4a.s32.s32 %r177, %r178, %r179, %r656;

mul.lo.s32 %r728, %r716, %r177;
cvt.rn.f32.s32 %f159, %r728;
fma.rn.ftz.f32 %f160, %f88, %f159, %f158;
add.s64 %rd85, %rd168, %rd21;
ld.global.nc.u8 %rs25, [%rd85];
cvt.u32.u16 %r729, %rs25;
and.b32 %r730, %r729, 255;
shr.u32 %r731, %r730, %r11;
and.b32 %r732, %r731, 15;
add.s64 %rd86, %rd168, %rd20;
ld.global.nc.u8 %rs26, [%rd86];
cvt.u32.u16 %r733, %rs26;
and.b32 %r734, %r733, 255;
shr.u32 %r735, %r734, %r12;
and.b32 %r736, %r735, 3;
bfi.b32 %r737, %r736, %r732, 4, 2;
add.s32 %r738, %r737, -32;
shr.u32 %r739, %r659, 6;
and.b32 %r615, %r739, 50529027;
shr.u32 %r740, %r661, 1;
and.b32 %r741, %r740, 67372036;
or.b32 %r742, %r615, -2139062144;
sub.s32 %r743, %r742, %r741;
xor.b32 %r744, %r743, -2139062144;
or.b32 %r745, %r615, %r741;
xor.b32 %r746, %r743, %r615;
and.b32 %r613, %r746, %r745;
mov.u32 %r181, %r613;

prmt.b32 %r181, %r181, 0, 0xba98;

mov.u32 %r183, %r615;

prmt.b32 %r183, %r183, 0, 0xba98;

xor.b32 %r747, %r183, 2139062143;
not.b32 %r748, %r181;
and.b32 %r749, %r744, %r748;
and.b32 %r750, %r747, %r181;
or.b32 %r186, %r750, %r749;

dp4a.s32.s32 %r185, %r186, %r187, %r656;

mul.lo.s32 %r751, %r738, %r185;
cvt.rn.f32.s32 %f161, %r751;
fma.rn.ftz.f32 %f162, %f89, %f161, %f160;
fma.rn.ftz.f32 %f448, %f85, %f162, %f448;
add.s64 %rd87, %rd168, %rd6;
ld.global.nc.u16 %rs2, [%rd87+108];

{ cvt.f32.f16 %f90, %rs2;}


add.s64 %rd88, %rd168, %rd19;
add.s64 %rd89, %rd168, %rd17;
ld.global.nc.u16 %rs27, [%rd89];
ld.global.nc.u16 %rs28, [%rd89+2];
mov.b32 %r752, {%rs27, %rs28};
not.b32 %r753, %r752;
ld.global.nc.u16 %rs29, [%rd88+2];
ld.global.nc.u16 %rs30, [%rd88];

{.reg .f16 low,high;
mov.b32 {low,high},%r153;
cvt.f32.f16 %f91, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r154;
cvt.f32.f16 %f92, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r155;
cvt.f32.f16 %f93, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r156;
cvt.f32.f16 %f94, low;}


mov.b32 %r754, {%rs30, %rs29};
shr.s32 %r755, %r753, %r660;
add.s64 %rd90, %rd168, %rd14;
ld.global.nc.u8 %rs31, [%rd90];
cvt.u32.u16 %r756, %rs31;
and.b32 %r757, %r756, 255;
shr.u32 %r758, %r757, %r5;
and.b32 %r759, %r758, 15;
add.s64 %rd91, %rd168, %rd13;
ld.global.nc.u8 %rs32, [%rd91];
cvt.u32.u16 %r760, %rs32;
and.b32 %r761, %r760, 255;
shr.u32 %r762, %r761, %r6;
and.b32 %r763, %r762, 3;
bfi.b32 %r764, %r763, %r759, 4, 2;
add.s32 %r765, %r764, -32;
and.b32 %r627, %r754, 50529027;
shl.b32 %r766, %r755, 2;
and.b32 %r767, %r766, 67372036;
or.b32 %r768, %r627, -2139062144;
sub.s32 %r769, %r768, %r767;
xor.b32 %r770, %r769, -2139062144;
or.b32 %r771, %r627, %r767;
xor.b32 %r772, %r769, %r627;
and.b32 %r625, %r772, %r771;
mov.u32 %r193, %r625;

prmt.b32 %r193, %r193, 0, 0xba98;

mov.u32 %r195, %r627;

prmt.b32 %r195, %r195, 0, 0xba98;

xor.b32 %r773, %r195, 2139062143;
not.b32 %r774, %r193;
and.b32 %r775, %r770, %r774;
and.b32 %r776, %r773, %r193;
or.b32 %r198, %r776, %r775;

dp4a.s32.s32 %r197, %r198, %r163, %r656;

mul.lo.s32 %r777, %r765, %r197;
cvt.rn.f32.s32 %f163, %r777;
fma.rn.ftz.f32 %f164, %f91, %f163, 0f00000000;
add.s64 %rd92, %rd168, %rd12;
ld.global.nc.u8 %rs33, [%rd92];
cvt.u32.u16 %r778, %rs33;
and.b32 %r779, %r778, 255;
shr.u32 %r780, %r779, %r7;
and.b32 %r781, %r780, 15;
add.s64 %rd93, %rd168, %rd11;
ld.global.nc.u8 %rs34, [%rd93];
cvt.u32.u16 %r782, %rs34;
and.b32 %r783, %r782, 255;
shr.u32 %r784, %r783, %r8;
and.b32 %r785, %r784, 3;
bfi.b32 %r786, %r785, %r781, 4, 2;
add.s32 %r787, %r786, -32;
shr.u32 %r788, %r754, 2;
and.b32 %r635, %r788, 50529027;
shl.b32 %r789, %r755, 1;
and.b32 %r790, %r789, 67372036;
or.b32 %r791, %r635, -2139062144;
sub.s32 %r792, %r791, %r790;
xor.b32 %r793, %r792, -2139062144;
or.b32 %r794, %r635, %r790;
xor.b32 %r795, %r792, %r635;
and.b32 %r633, %r795, %r794;
mov.u32 %r201, %r633;

prmt.b32 %r201, %r201, 0, 0xba98;

mov.u32 %r203, %r635;

prmt.b32 %r203, %r203, 0, 0xba98;

xor.b32 %r796, %r203, 2139062143;
not.b32 %r797, %r201;
and.b32 %r798, %r793, %r797;
and.b32 %r799, %r796, %r201;
or.b32 %r206, %r799, %r798;

dp4a.s32.s32 %r205, %r206, %r171, %r656;

mul.lo.s32 %r800, %r787, %r205;
cvt.rn.f32.s32 %f165, %r800;
fma.rn.ftz.f32 %f166, %f92, %f165, %f164;
add.s64 %rd94, %rd168, %rd10;
ld.global.nc.u8 %rs35, [%rd94];
cvt.u32.u16 %r801, %rs35;
and.b32 %r802, %r801, 255;
shr.u32 %r803, %r802, %r9;
and.b32 %r804, %r803, 15;
add.s64 %rd95, %rd168, %rd9;
ld.global.nc.u8 %rs36, [%rd95];
cvt.u32.u16 %r805, %rs36;
and.b32 %r806, %r805, 255;
shr.u32 %r807, %r806, %r10;
and.b32 %r808, %r807, 3;
bfi.b32 %r809, %r808, %r804, 4, 2;
add.s32 %r810, %r809, -32;
shr.u32 %r811, %r754, 4;
and.b32 %r643, %r811, 50529027;
and.b32 %r812, %r755, 67372036;
or.b32 %r813, %r643, -2139062144;
sub.s32 %r814, %r813, %r812;
xor.b32 %r815, %r814, -2139062144;
or.b32 %r816, %r643, %r812;
xor.b32 %r817, %r814, %r643;
and.b32 %r641, %r817, %r816;
mov.u32 %r209, %r641;

prmt.b32 %r209, %r209, 0, 0xba98;

mov.u32 %r211, %r643;

prmt.b32 %r211, %r211, 0, 0xba98;

xor.b32 %r818, %r211, 2139062143;
not.b32 %r819, %r209;
and.b32 %r820, %r815, %r819;
and.b32 %r821, %r818, %r209;
or.b32 %r214, %r821, %r820;

dp4a.s32.s32 %r213, %r214, %r179, %r656;

mul.lo.s32 %r822, %r810, %r213;
cvt.rn.f32.s32 %f167, %r822;
fma.rn.ftz.f32 %f168, %f93, %f167, %f166;
add.s64 %rd96, %rd168, %rd8;
ld.global.nc.u8 %rs37, [%rd96];
cvt.u32.u16 %r823, %rs37;
and.b32 %r824, %r823, 255;
shr.u32 %r825, %r824, %r11;
and.b32 %r826, %r825, 15;
add.s64 %rd97, %rd168, %rd7;
ld.global.nc.u8 %rs38, [%rd97];
cvt.u32.u16 %r827, %rs38;
and.b32 %r828, %r827, 255;
shr.u32 %r829, %r828, %r12;
and.b32 %r830, %r829, 3;
bfi.b32 %r831, %r830, %r826, 4, 2;
add.s32 %r832, %r831, -32;
shr.u32 %r833, %r754, 6;
and.b32 %r651, %r833, 50529027;
shr.u32 %r834, %r755, 1;
and.b32 %r835, %r834, 67372036;
or.b32 %r836, %r651, -2139062144;
sub.s32 %r837, %r836, %r835;
xor.b32 %r838, %r837, -2139062144;
or.b32 %r839, %r651, %r835;
xor.b32 %r840, %r837, %r651;
and.b32 %r649, %r840, %r839;
mov.u32 %r217, %r649;

prmt.b32 %r217, %r217, 0, 0xba98;

mov.u32 %r219, %r651;

prmt.b32 %r219, %r219, 0, 0xba98;

xor.b32 %r841, %r219, 2139062143;
not.b32 %r842, %r217;
and.b32 %r843, %r838, %r842;
and.b32 %r844, %r841, %r217;
or.b32 %r222, %r844, %r843;

dp4a.s32.s32 %r221, %r222, %r187, %r656;

mul.lo.s32 %r845, %r832, %r221;
cvt.rn.f32.s32 %f169, %r845;
fma.rn.ftz.f32 %f170, %f94, %f169, %f168;
fma.rn.ftz.f32 %f447, %f90, %f170, %f447;
cvt.u64.u32 %rd98, %r1366;
add.s64 %rd99, %rd2, %rd98;

{ cvt.f32.f16 %f95, %rs1;}


add.s64 %rd100, %rd99, %rd5;
mul.lo.s64 %rd101, %rd100, 36;
add.s64 %rd102, %rd28, %rd101;
add.s64 %rd103, %rd102, %rd77;
ld.global.nc.u32 %r235, [%rd103+4];
ld.global.nc.u32 %r225, [%rd102];

{.reg .f16 low,high;
mov.b32 {low,high},%r225;
cvt.f32.f16 %f96, low;}


ld.global.nc.u32 %r243, [%rd103+40];
ld.global.nc.u32 %r226, [%rd102+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r226;
cvt.f32.f16 %f97, low;}


ld.global.nc.u32 %r251, [%rd103+76];
ld.global.nc.u32 %r227, [%rd102+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r227;
cvt.f32.f16 %f98, low;}


ld.global.nc.u32 %r259, [%rd103+112];
ld.global.nc.u32 %r228, [%rd102+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r228;
cvt.f32.f16 %f99, low;}


mov.u32 %r229, %r589;

prmt.b32 %r229, %r229, 0, 0xba98;

mov.u32 %r231, %r591;

prmt.b32 %r231, %r231, 0, 0xba98;

xor.b32 %r846, %r231, 2139062143;
not.b32 %r847, %r229;
and.b32 %r848, %r676, %r847;
and.b32 %r849, %r846, %r229;
or.b32 %r234, %r849, %r848;

dp4a.s32.s32 %r233, %r234, %r235, %r656;

mul.lo.s32 %r850, %r671, %r233;
cvt.rn.f32.s32 %f171, %r850;
fma.rn.ftz.f32 %f172, %f96, %f171, 0f00000000;
mov.u32 %r237, %r597;

prmt.b32 %r237, %r237, 0, 0xba98;

mov.u32 %r239, %r599;

prmt.b32 %r239, %r239, 0, 0xba98;

xor.b32 %r851, %r239, 2139062143;
not.b32 %r852, %r237;
and.b32 %r853, %r699, %r852;
and.b32 %r854, %r851, %r237;
or.b32 %r242, %r854, %r853;

dp4a.s32.s32 %r241, %r242, %r243, %r656;

mul.lo.s32 %r855, %r693, %r241;
cvt.rn.f32.s32 %f173, %r855;
fma.rn.ftz.f32 %f174, %f97, %f173, %f172;
mov.u32 %r245, %r605;

prmt.b32 %r245, %r245, 0, 0xba98;

mov.u32 %r247, %r607;

prmt.b32 %r247, %r247, 0, 0xba98;

xor.b32 %r856, %r247, 2139062143;
not.b32 %r857, %r245;
and.b32 %r858, %r721, %r857;
and.b32 %r859, %r856, %r245;
or.b32 %r250, %r859, %r858;

dp4a.s32.s32 %r249, %r250, %r251, %r656;

mul.lo.s32 %r860, %r716, %r249;
cvt.rn.f32.s32 %f175, %r860;
fma.rn.ftz.f32 %f176, %f98, %f175, %f174;
mov.u32 %r253, %r613;

prmt.b32 %r253, %r253, 0, 0xba98;

mov.u32 %r255, %r615;

prmt.b32 %r255, %r255, 0, 0xba98;

xor.b32 %r861, %r255, 2139062143;
not.b32 %r862, %r253;
and.b32 %r863, %r744, %r862;
and.b32 %r864, %r861, %r253;
or.b32 %r258, %r864, %r863;

dp4a.s32.s32 %r257, %r258, %r259, %r656;

mul.lo.s32 %r865, %r738, %r257;
cvt.rn.f32.s32 %f177, %r865;
fma.rn.ftz.f32 %f178, %f99, %f177, %f176;
fma.rn.ftz.f32 %f446, %f95, %f178, %f446;

{ cvt.f32.f16 %f100, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r225;
cvt.f32.f16 %f101, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r226;
cvt.f32.f16 %f102, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r227;
cvt.f32.f16 %f103, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r228;
cvt.f32.f16 %f104, low;}


mov.u32 %r265, %r625;

prmt.b32 %r265, %r265, 0, 0xba98;

mov.u32 %r267, %r627;

prmt.b32 %r267, %r267, 0, 0xba98;

xor.b32 %r866, %r267, 2139062143;
not.b32 %r867, %r265;
and.b32 %r868, %r770, %r867;
and.b32 %r869, %r866, %r265;
or.b32 %r270, %r869, %r868;

dp4a.s32.s32 %r269, %r270, %r235, %r656;

mul.lo.s32 %r870, %r765, %r269;
cvt.rn.f32.s32 %f179, %r870;
fma.rn.ftz.f32 %f180, %f101, %f179, 0f00000000;
mov.u32 %r273, %r633;

prmt.b32 %r273, %r273, 0, 0xba98;

mov.u32 %r275, %r635;

prmt.b32 %r275, %r275, 0, 0xba98;

xor.b32 %r871, %r275, 2139062143;
not.b32 %r872, %r273;
and.b32 %r873, %r793, %r872;
and.b32 %r874, %r871, %r273;
or.b32 %r278, %r874, %r873;

dp4a.s32.s32 %r277, %r278, %r243, %r656;

mul.lo.s32 %r875, %r787, %r277;
cvt.rn.f32.s32 %f181, %r875;
fma.rn.ftz.f32 %f182, %f102, %f181, %f180;
mov.u32 %r281, %r641;

prmt.b32 %r281, %r281, 0, 0xba98;

mov.u32 %r283, %r643;

prmt.b32 %r283, %r283, 0, 0xba98;

xor.b32 %r876, %r283, 2139062143;
not.b32 %r877, %r281;
and.b32 %r878, %r815, %r877;
and.b32 %r879, %r876, %r281;
or.b32 %r286, %r879, %r878;

dp4a.s32.s32 %r285, %r286, %r251, %r656;

mul.lo.s32 %r880, %r810, %r285;
cvt.rn.f32.s32 %f183, %r880;
fma.rn.ftz.f32 %f184, %f103, %f183, %f182;
mov.u32 %r289, %r649;

prmt.b32 %r289, %r289, 0, 0xba98;

mov.u32 %r291, %r651;

prmt.b32 %r291, %r291, 0, 0xba98;

xor.b32 %r881, %r291, 2139062143;
not.b32 %r882, %r289;
and.b32 %r883, %r838, %r882;
and.b32 %r884, %r881, %r289;
or.b32 %r294, %r884, %r883;

dp4a.s32.s32 %r293, %r294, %r259, %r656;

mul.lo.s32 %r885, %r832, %r293;
cvt.rn.f32.s32 %f185, %r885;
fma.rn.ftz.f32 %f186, %f104, %f185, %f184;
fma.rn.ftz.f32 %f445, %f100, %f186, %f445;
cvt.u64.u32 %rd104, %r1371;
add.s64 %rd105, %rd2, %rd104;

{ cvt.f32.f16 %f105, %rs1;}


add.s64 %rd106, %rd105, %rd5;
mul.lo.s64 %rd107, %rd106, 36;
add.s64 %rd108, %rd28, %rd107;
add.s64 %rd109, %rd108, %rd77;
ld.global.nc.u32 %r307, [%rd109+4];
ld.global.nc.u32 %r297, [%rd108];

{.reg .f16 low,high;
mov.b32 {low,high},%r297;
cvt.f32.f16 %f106, low;}


ld.global.nc.u32 %r315, [%rd109+40];
ld.global.nc.u32 %r298, [%rd108+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r298;
cvt.f32.f16 %f107, low;}


ld.global.nc.u32 %r323, [%rd109+76];
ld.global.nc.u32 %r299, [%rd108+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r299;
cvt.f32.f16 %f108, low;}


ld.global.nc.u32 %r331, [%rd109+112];
ld.global.nc.u32 %r300, [%rd108+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r300;
cvt.f32.f16 %f109, low;}


mov.u32 %r301, %r589;

prmt.b32 %r301, %r301, 0, 0xba98;

mov.u32 %r303, %r591;

prmt.b32 %r303, %r303, 0, 0xba98;

xor.b32 %r886, %r303, 2139062143;
not.b32 %r887, %r301;
and.b32 %r888, %r676, %r887;
and.b32 %r889, %r886, %r301;
or.b32 %r306, %r889, %r888;

dp4a.s32.s32 %r305, %r306, %r307, %r656;

mul.lo.s32 %r890, %r671, %r305;
cvt.rn.f32.s32 %f187, %r890;
fma.rn.ftz.f32 %f188, %f106, %f187, 0f00000000;
mov.u32 %r309, %r597;

prmt.b32 %r309, %r309, 0, 0xba98;

mov.u32 %r311, %r599;

prmt.b32 %r311, %r311, 0, 0xba98;

xor.b32 %r891, %r311, 2139062143;
not.b32 %r892, %r309;
and.b32 %r893, %r699, %r892;
and.b32 %r894, %r891, %r309;
or.b32 %r314, %r894, %r893;

dp4a.s32.s32 %r313, %r314, %r315, %r656;

mul.lo.s32 %r895, %r693, %r313;
cvt.rn.f32.s32 %f189, %r895;
fma.rn.ftz.f32 %f190, %f107, %f189, %f188;
mov.u32 %r317, %r605;

prmt.b32 %r317, %r317, 0, 0xba98;

mov.u32 %r319, %r607;

prmt.b32 %r319, %r319, 0, 0xba98;

xor.b32 %r896, %r319, 2139062143;
not.b32 %r897, %r317;
and.b32 %r898, %r721, %r897;
and.b32 %r899, %r896, %r317;
or.b32 %r322, %r899, %r898;

dp4a.s32.s32 %r321, %r322, %r323, %r656;

mul.lo.s32 %r900, %r716, %r321;
cvt.rn.f32.s32 %f191, %r900;
fma.rn.ftz.f32 %f192, %f108, %f191, %f190;
mov.u32 %r325, %r613;

prmt.b32 %r325, %r325, 0, 0xba98;

mov.u32 %r327, %r615;

prmt.b32 %r327, %r327, 0, 0xba98;

xor.b32 %r901, %r327, 2139062143;
not.b32 %r902, %r325;
and.b32 %r903, %r744, %r902;
and.b32 %r904, %r901, %r325;
or.b32 %r330, %r904, %r903;

dp4a.s32.s32 %r329, %r330, %r331, %r656;

mul.lo.s32 %r905, %r738, %r329;
cvt.rn.f32.s32 %f193, %r905;
fma.rn.ftz.f32 %f194, %f109, %f193, %f192;
fma.rn.ftz.f32 %f444, %f105, %f194, %f444;

{ cvt.f32.f16 %f110, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r297;
cvt.f32.f16 %f111, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r298;
cvt.f32.f16 %f112, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r299;
cvt.f32.f16 %f113, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r300;
cvt.f32.f16 %f114, low;}


mov.u32 %r337, %r625;

prmt.b32 %r337, %r337, 0, 0xba98;

mov.u32 %r339, %r627;

prmt.b32 %r339, %r339, 0, 0xba98;

xor.b32 %r906, %r339, 2139062143;
not.b32 %r907, %r337;
and.b32 %r908, %r770, %r907;
and.b32 %r909, %r906, %r337;
or.b32 %r342, %r909, %r908;

dp4a.s32.s32 %r341, %r342, %r307, %r656;

mul.lo.s32 %r910, %r765, %r341;
cvt.rn.f32.s32 %f195, %r910;
fma.rn.ftz.f32 %f196, %f111, %f195, 0f00000000;
mov.u32 %r345, %r633;

prmt.b32 %r345, %r345, 0, 0xba98;

mov.u32 %r347, %r635;

prmt.b32 %r347, %r347, 0, 0xba98;

xor.b32 %r911, %r347, 2139062143;
not.b32 %r912, %r345;
and.b32 %r913, %r793, %r912;
and.b32 %r914, %r911, %r345;
or.b32 %r350, %r914, %r913;

dp4a.s32.s32 %r349, %r350, %r315, %r656;

mul.lo.s32 %r915, %r787, %r349;
cvt.rn.f32.s32 %f197, %r915;
fma.rn.ftz.f32 %f198, %f112, %f197, %f196;
mov.u32 %r353, %r641;

prmt.b32 %r353, %r353, 0, 0xba98;

mov.u32 %r355, %r643;

prmt.b32 %r355, %r355, 0, 0xba98;

xor.b32 %r916, %r355, 2139062143;
not.b32 %r917, %r353;
and.b32 %r918, %r815, %r917;
and.b32 %r919, %r916, %r353;
or.b32 %r358, %r919, %r918;

dp4a.s32.s32 %r357, %r358, %r323, %r656;

mul.lo.s32 %r920, %r810, %r357;
cvt.rn.f32.s32 %f199, %r920;
fma.rn.ftz.f32 %f200, %f113, %f199, %f198;
mov.u32 %r361, %r649;

prmt.b32 %r361, %r361, 0, 0xba98;

mov.u32 %r363, %r651;

prmt.b32 %r363, %r363, 0, 0xba98;

xor.b32 %r921, %r363, 2139062143;
not.b32 %r922, %r361;
and.b32 %r923, %r838, %r922;
and.b32 %r924, %r921, %r361;
or.b32 %r366, %r924, %r923;

dp4a.s32.s32 %r365, %r366, %r331, %r656;

mul.lo.s32 %r925, %r832, %r365;
cvt.rn.f32.s32 %f201, %r925;
fma.rn.ftz.f32 %f202, %f114, %f201, %f200;
fma.rn.ftz.f32 %f443, %f110, %f202, %f443;
cvt.u64.u32 %rd110, %r1370;
add.s64 %rd111, %rd2, %rd110;

{ cvt.f32.f16 %f115, %rs1;}


add.s64 %rd112, %rd111, %rd5;
mul.lo.s64 %rd113, %rd112, 36;
add.s64 %rd114, %rd28, %rd113;
add.s64 %rd115, %rd114, %rd77;
ld.global.nc.u32 %r379, [%rd115+4];
ld.global.nc.u32 %r369, [%rd114];

{.reg .f16 low,high;
mov.b32 {low,high},%r369;
cvt.f32.f16 %f116, low;}


ld.global.nc.u32 %r387, [%rd115+40];
ld.global.nc.u32 %r370, [%rd114+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r370;
cvt.f32.f16 %f117, low;}


ld.global.nc.u32 %r395, [%rd115+76];
ld.global.nc.u32 %r371, [%rd114+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r371;
cvt.f32.f16 %f118, low;}


ld.global.nc.u32 %r403, [%rd115+112];
ld.global.nc.u32 %r372, [%rd114+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r372;
cvt.f32.f16 %f119, low;}


mov.u32 %r373, %r589;

prmt.b32 %r373, %r373, 0, 0xba98;

mov.u32 %r375, %r591;

prmt.b32 %r375, %r375, 0, 0xba98;

xor.b32 %r926, %r375, 2139062143;
not.b32 %r927, %r373;
and.b32 %r928, %r676, %r927;
and.b32 %r929, %r926, %r373;
or.b32 %r378, %r929, %r928;

dp4a.s32.s32 %r377, %r378, %r379, %r656;

mul.lo.s32 %r930, %r671, %r377;
cvt.rn.f32.s32 %f203, %r930;
fma.rn.ftz.f32 %f204, %f116, %f203, 0f00000000;
mov.u32 %r381, %r597;

prmt.b32 %r381, %r381, 0, 0xba98;

mov.u32 %r383, %r599;

prmt.b32 %r383, %r383, 0, 0xba98;

xor.b32 %r931, %r383, 2139062143;
not.b32 %r932, %r381;
and.b32 %r933, %r699, %r932;
and.b32 %r934, %r931, %r381;
or.b32 %r386, %r934, %r933;

dp4a.s32.s32 %r385, %r386, %r387, %r656;

mul.lo.s32 %r935, %r693, %r385;
cvt.rn.f32.s32 %f205, %r935;
fma.rn.ftz.f32 %f206, %f117, %f205, %f204;
mov.u32 %r389, %r605;

prmt.b32 %r389, %r389, 0, 0xba98;

mov.u32 %r391, %r607;

prmt.b32 %r391, %r391, 0, 0xba98;

xor.b32 %r936, %r391, 2139062143;
not.b32 %r937, %r389;
and.b32 %r938, %r721, %r937;
and.b32 %r939, %r936, %r389;
or.b32 %r394, %r939, %r938;

dp4a.s32.s32 %r393, %r394, %r395, %r656;

mul.lo.s32 %r940, %r716, %r393;
cvt.rn.f32.s32 %f207, %r940;
fma.rn.ftz.f32 %f208, %f118, %f207, %f206;
mov.u32 %r397, %r613;

prmt.b32 %r397, %r397, 0, 0xba98;

mov.u32 %r399, %r615;

prmt.b32 %r399, %r399, 0, 0xba98;

xor.b32 %r941, %r399, 2139062143;
not.b32 %r942, %r397;
and.b32 %r943, %r744, %r942;
and.b32 %r944, %r941, %r397;
or.b32 %r402, %r944, %r943;

dp4a.s32.s32 %r401, %r402, %r403, %r656;

mul.lo.s32 %r945, %r738, %r401;
cvt.rn.f32.s32 %f209, %r945;
fma.rn.ftz.f32 %f210, %f119, %f209, %f208;
fma.rn.ftz.f32 %f442, %f115, %f210, %f442;

{ cvt.f32.f16 %f120, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r369;
cvt.f32.f16 %f121, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r370;
cvt.f32.f16 %f122, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r371;
cvt.f32.f16 %f123, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r372;
cvt.f32.f16 %f124, low;}


mov.u32 %r409, %r625;

prmt.b32 %r409, %r409, 0, 0xba98;

mov.u32 %r411, %r627;

prmt.b32 %r411, %r411, 0, 0xba98;

xor.b32 %r946, %r411, 2139062143;
not.b32 %r947, %r409;
and.b32 %r948, %r770, %r947;
and.b32 %r949, %r946, %r409;
or.b32 %r414, %r949, %r948;

dp4a.s32.s32 %r413, %r414, %r379, %r656;

mul.lo.s32 %r950, %r765, %r413;
cvt.rn.f32.s32 %f211, %r950;
fma.rn.ftz.f32 %f212, %f121, %f211, 0f00000000;
mov.u32 %r417, %r633;

prmt.b32 %r417, %r417, 0, 0xba98;

mov.u32 %r419, %r635;

prmt.b32 %r419, %r419, 0, 0xba98;

xor.b32 %r951, %r419, 2139062143;
not.b32 %r952, %r417;
and.b32 %r953, %r793, %r952;
and.b32 %r954, %r951, %r417;
or.b32 %r422, %r954, %r953;

dp4a.s32.s32 %r421, %r422, %r387, %r656;

mul.lo.s32 %r955, %r787, %r421;
cvt.rn.f32.s32 %f213, %r955;
fma.rn.ftz.f32 %f214, %f122, %f213, %f212;
mov.u32 %r425, %r641;

prmt.b32 %r425, %r425, 0, 0xba98;

mov.u32 %r427, %r643;

prmt.b32 %r427, %r427, 0, 0xba98;

xor.b32 %r956, %r427, 2139062143;
not.b32 %r957, %r425;
and.b32 %r958, %r815, %r957;
and.b32 %r959, %r956, %r425;
or.b32 %r430, %r959, %r958;

dp4a.s32.s32 %r429, %r430, %r395, %r656;

mul.lo.s32 %r960, %r810, %r429;
cvt.rn.f32.s32 %f215, %r960;
fma.rn.ftz.f32 %f216, %f123, %f215, %f214;
mov.u32 %r433, %r649;

prmt.b32 %r433, %r433, 0, 0xba98;

mov.u32 %r435, %r651;

prmt.b32 %r435, %r435, 0, 0xba98;

xor.b32 %r961, %r435, 2139062143;
not.b32 %r962, %r433;
and.b32 %r963, %r838, %r962;
and.b32 %r964, %r961, %r433;
or.b32 %r438, %r964, %r963;

dp4a.s32.s32 %r437, %r438, %r403, %r656;

mul.lo.s32 %r965, %r832, %r437;
cvt.rn.f32.s32 %f217, %r965;
fma.rn.ftz.f32 %f218, %f124, %f217, %f216;
fma.rn.ftz.f32 %f441, %f120, %f218, %f441;
cvt.u64.u32 %rd116, %r1369;
add.s64 %rd117, %rd2, %rd116;

{ cvt.f32.f16 %f125, %rs1;}


add.s64 %rd118, %rd117, %rd5;
mul.lo.s64 %rd119, %rd118, 36;
add.s64 %rd120, %rd28, %rd119;
add.s64 %rd121, %rd120, %rd77;
ld.global.nc.u32 %r451, [%rd121+4];
ld.global.nc.u32 %r441, [%rd120];

{.reg .f16 low,high;
mov.b32 {low,high},%r441;
cvt.f32.f16 %f126, low;}


ld.global.nc.u32 %r459, [%rd121+40];
ld.global.nc.u32 %r442, [%rd120+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r442;
cvt.f32.f16 %f127, low;}


ld.global.nc.u32 %r467, [%rd121+76];
ld.global.nc.u32 %r443, [%rd120+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r443;
cvt.f32.f16 %f128, low;}


ld.global.nc.u32 %r475, [%rd121+112];
ld.global.nc.u32 %r444, [%rd120+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r444;
cvt.f32.f16 %f129, low;}


mov.u32 %r445, %r589;

prmt.b32 %r445, %r445, 0, 0xba98;

mov.u32 %r447, %r591;

prmt.b32 %r447, %r447, 0, 0xba98;

xor.b32 %r966, %r447, 2139062143;
not.b32 %r967, %r445;
and.b32 %r968, %r676, %r967;
and.b32 %r969, %r966, %r445;
or.b32 %r450, %r969, %r968;

dp4a.s32.s32 %r449, %r450, %r451, %r656;

mul.lo.s32 %r970, %r671, %r449;
cvt.rn.f32.s32 %f219, %r970;
fma.rn.ftz.f32 %f220, %f126, %f219, 0f00000000;
mov.u32 %r453, %r597;

prmt.b32 %r453, %r453, 0, 0xba98;

mov.u32 %r455, %r599;

prmt.b32 %r455, %r455, 0, 0xba98;

xor.b32 %r971, %r455, 2139062143;
not.b32 %r972, %r453;
and.b32 %r973, %r699, %r972;
and.b32 %r974, %r971, %r453;
or.b32 %r458, %r974, %r973;

dp4a.s32.s32 %r457, %r458, %r459, %r656;

mul.lo.s32 %r975, %r693, %r457;
cvt.rn.f32.s32 %f221, %r975;
fma.rn.ftz.f32 %f222, %f127, %f221, %f220;
mov.u32 %r461, %r605;

prmt.b32 %r461, %r461, 0, 0xba98;

mov.u32 %r463, %r607;

prmt.b32 %r463, %r463, 0, 0xba98;

xor.b32 %r976, %r463, 2139062143;
not.b32 %r977, %r461;
and.b32 %r978, %r721, %r977;
and.b32 %r979, %r976, %r461;
or.b32 %r466, %r979, %r978;

dp4a.s32.s32 %r465, %r466, %r467, %r656;

mul.lo.s32 %r980, %r716, %r465;
cvt.rn.f32.s32 %f223, %r980;
fma.rn.ftz.f32 %f224, %f128, %f223, %f222;
mov.u32 %r469, %r613;

prmt.b32 %r469, %r469, 0, 0xba98;

mov.u32 %r471, %r615;

prmt.b32 %r471, %r471, 0, 0xba98;

xor.b32 %r981, %r471, 2139062143;
not.b32 %r982, %r469;
and.b32 %r983, %r744, %r982;
and.b32 %r984, %r981, %r469;
or.b32 %r474, %r984, %r983;

dp4a.s32.s32 %r473, %r474, %r475, %r656;

mul.lo.s32 %r985, %r738, %r473;
cvt.rn.f32.s32 %f225, %r985;
fma.rn.ftz.f32 %f226, %f129, %f225, %f224;
fma.rn.ftz.f32 %f440, %f125, %f226, %f440;

{ cvt.f32.f16 %f130, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r441;
cvt.f32.f16 %f131, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r442;
cvt.f32.f16 %f132, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r443;
cvt.f32.f16 %f133, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r444;
cvt.f32.f16 %f134, low;}


mov.u32 %r481, %r625;

prmt.b32 %r481, %r481, 0, 0xba98;

mov.u32 %r483, %r627;

prmt.b32 %r483, %r483, 0, 0xba98;

xor.b32 %r986, %r483, 2139062143;
not.b32 %r987, %r481;
and.b32 %r988, %r770, %r987;
and.b32 %r989, %r986, %r481;
or.b32 %r486, %r989, %r988;

dp4a.s32.s32 %r485, %r486, %r451, %r656;

mul.lo.s32 %r990, %r765, %r485;
cvt.rn.f32.s32 %f227, %r990;
fma.rn.ftz.f32 %f228, %f131, %f227, 0f00000000;
mov.u32 %r489, %r633;

prmt.b32 %r489, %r489, 0, 0xba98;

mov.u32 %r491, %r635;

prmt.b32 %r491, %r491, 0, 0xba98;

xor.b32 %r991, %r491, 2139062143;
not.b32 %r992, %r489;
and.b32 %r993, %r793, %r992;
and.b32 %r994, %r991, %r489;
or.b32 %r494, %r994, %r993;

dp4a.s32.s32 %r493, %r494, %r459, %r656;

mul.lo.s32 %r995, %r787, %r493;
cvt.rn.f32.s32 %f229, %r995;
fma.rn.ftz.f32 %f230, %f132, %f229, %f228;
mov.u32 %r497, %r641;

prmt.b32 %r497, %r497, 0, 0xba98;

mov.u32 %r499, %r643;

prmt.b32 %r499, %r499, 0, 0xba98;

xor.b32 %r996, %r499, 2139062143;
not.b32 %r997, %r497;
and.b32 %r998, %r815, %r997;
and.b32 %r999, %r996, %r497;
or.b32 %r502, %r999, %r998;

dp4a.s32.s32 %r501, %r502, %r467, %r656;

mul.lo.s32 %r1000, %r810, %r501;
cvt.rn.f32.s32 %f231, %r1000;
fma.rn.ftz.f32 %f232, %f133, %f231, %f230;
mov.u32 %r505, %r649;

prmt.b32 %r505, %r505, 0, 0xba98;

mov.u32 %r507, %r651;

prmt.b32 %r507, %r507, 0, 0xba98;

xor.b32 %r1001, %r507, 2139062143;
not.b32 %r1002, %r505;
and.b32 %r1003, %r838, %r1002;
and.b32 %r1004, %r1001, %r505;
or.b32 %r510, %r1004, %r1003;

dp4a.s32.s32 %r509, %r510, %r475, %r656;

mul.lo.s32 %r1005, %r832, %r509;
cvt.rn.f32.s32 %f233, %r1005;
fma.rn.ftz.f32 %f234, %f134, %f233, %f232;
fma.rn.ftz.f32 %f439, %f130, %f234, %f439;
cvt.u64.u32 %rd122, %r1368;
add.s64 %rd123, %rd2, %rd122;

{ cvt.f32.f16 %f135, %rs1;}


add.s64 %rd124, %rd123, %rd5;
mul.lo.s64 %rd125, %rd124, 36;
add.s64 %rd126, %rd28, %rd125;
add.s64 %rd127, %rd126, %rd77;
ld.global.nc.u32 %r523, [%rd127+4];
ld.global.nc.u32 %r513, [%rd126];

{.reg .f16 low,high;
mov.b32 {low,high},%r513;
cvt.f32.f16 %f136, low;}


ld.global.nc.u32 %r531, [%rd127+40];
ld.global.nc.u32 %r514, [%rd126+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r514;
cvt.f32.f16 %f137, low;}


ld.global.nc.u32 %r539, [%rd127+76];
ld.global.nc.u32 %r515, [%rd126+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r515;
cvt.f32.f16 %f138, low;}


ld.global.nc.u32 %r547, [%rd127+112];
ld.global.nc.u32 %r516, [%rd126+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r516;
cvt.f32.f16 %f139, low;}


mov.u32 %r517, %r589;

prmt.b32 %r517, %r517, 0, 0xba98;

mov.u32 %r519, %r591;

prmt.b32 %r519, %r519, 0, 0xba98;

xor.b32 %r1006, %r519, 2139062143;
not.b32 %r1007, %r517;
and.b32 %r1008, %r676, %r1007;
and.b32 %r1009, %r1006, %r517;
or.b32 %r522, %r1009, %r1008;

dp4a.s32.s32 %r521, %r522, %r523, %r656;

mul.lo.s32 %r1010, %r671, %r521;
cvt.rn.f32.s32 %f235, %r1010;
fma.rn.ftz.f32 %f236, %f136, %f235, 0f00000000;
mov.u32 %r525, %r597;

prmt.b32 %r525, %r525, 0, 0xba98;

mov.u32 %r527, %r599;

prmt.b32 %r527, %r527, 0, 0xba98;

xor.b32 %r1011, %r527, 2139062143;
not.b32 %r1012, %r525;
and.b32 %r1013, %r699, %r1012;
and.b32 %r1014, %r1011, %r525;
or.b32 %r530, %r1014, %r1013;

dp4a.s32.s32 %r529, %r530, %r531, %r656;

mul.lo.s32 %r1015, %r693, %r529;
cvt.rn.f32.s32 %f237, %r1015;
fma.rn.ftz.f32 %f238, %f137, %f237, %f236;
mov.u32 %r533, %r605;

prmt.b32 %r533, %r533, 0, 0xba98;

mov.u32 %r535, %r607;

prmt.b32 %r535, %r535, 0, 0xba98;

xor.b32 %r1016, %r535, 2139062143;
not.b32 %r1017, %r533;
and.b32 %r1018, %r721, %r1017;
and.b32 %r1019, %r1016, %r533;
or.b32 %r538, %r1019, %r1018;

dp4a.s32.s32 %r537, %r538, %r539, %r656;

mul.lo.s32 %r1020, %r716, %r537;
cvt.rn.f32.s32 %f239, %r1020;
fma.rn.ftz.f32 %f240, %f138, %f239, %f238;
mov.u32 %r541, %r613;

prmt.b32 %r541, %r541, 0, 0xba98;

mov.u32 %r543, %r615;

prmt.b32 %r543, %r543, 0, 0xba98;

xor.b32 %r1021, %r543, 2139062143;
not.b32 %r1022, %r541;
and.b32 %r1023, %r744, %r1022;
and.b32 %r1024, %r1021, %r541;
or.b32 %r546, %r1024, %r1023;

dp4a.s32.s32 %r545, %r546, %r547, %r656;

mul.lo.s32 %r1025, %r738, %r545;
cvt.rn.f32.s32 %f241, %r1025;
fma.rn.ftz.f32 %f242, %f139, %f241, %f240;
fma.rn.ftz.f32 %f438, %f135, %f242, %f438;

{ cvt.f32.f16 %f140, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r513;
cvt.f32.f16 %f141, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r514;
cvt.f32.f16 %f142, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r515;
cvt.f32.f16 %f143, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r516;
cvt.f32.f16 %f144, low;}


mov.u32 %r553, %r625;

prmt.b32 %r553, %r553, 0, 0xba98;

mov.u32 %r555, %r627;

prmt.b32 %r555, %r555, 0, 0xba98;

xor.b32 %r1026, %r555, 2139062143;
not.b32 %r1027, %r553;
and.b32 %r1028, %r770, %r1027;
and.b32 %r1029, %r1026, %r553;
or.b32 %r558, %r1029, %r1028;

dp4a.s32.s32 %r557, %r558, %r523, %r656;

mul.lo.s32 %r1030, %r765, %r557;
cvt.rn.f32.s32 %f243, %r1030;
fma.rn.ftz.f32 %f244, %f141, %f243, 0f00000000;
mov.u32 %r561, %r633;

prmt.b32 %r561, %r561, 0, 0xba98;

mov.u32 %r563, %r635;

prmt.b32 %r563, %r563, 0, 0xba98;

xor.b32 %r1031, %r563, 2139062143;
not.b32 %r1032, %r561;
and.b32 %r1033, %r793, %r1032;
and.b32 %r1034, %r1031, %r561;
or.b32 %r566, %r1034, %r1033;

dp4a.s32.s32 %r565, %r566, %r531, %r656;

mul.lo.s32 %r1035, %r787, %r565;
cvt.rn.f32.s32 %f245, %r1035;
fma.rn.ftz.f32 %f246, %f142, %f245, %f244;
mov.u32 %r569, %r641;

prmt.b32 %r569, %r569, 0, 0xba98;

mov.u32 %r571, %r643;

prmt.b32 %r571, %r571, 0, 0xba98;

xor.b32 %r1036, %r571, 2139062143;
not.b32 %r1037, %r569;
and.b32 %r1038, %r815, %r1037;
and.b32 %r1039, %r1036, %r569;
or.b32 %r574, %r1039, %r1038;

dp4a.s32.s32 %r573, %r574, %r539, %r656;

mul.lo.s32 %r1040, %r810, %r573;
cvt.rn.f32.s32 %f247, %r1040;
fma.rn.ftz.f32 %f248, %f143, %f247, %f246;
mov.u32 %r577, %r649;

prmt.b32 %r577, %r577, 0, 0xba98;

mov.u32 %r579, %r651;

prmt.b32 %r579, %r579, 0, 0xba98;

xor.b32 %r1041, %r579, 2139062143;
not.b32 %r1042, %r577;
and.b32 %r1043, %r838, %r1042;
and.b32 %r1044, %r1041, %r577;
or.b32 %r582, %r1044, %r1043;

dp4a.s32.s32 %r581, %r582, %r547, %r656;

mul.lo.s32 %r1045, %r832, %r581;
cvt.rn.f32.s32 %f249, %r1045;
fma.rn.ftz.f32 %f250, %f144, %f249, %f248;
fma.rn.ftz.f32 %f437, %f140, %f250, %f437;
cvt.u64.u32 %rd128, %r1367;
add.s64 %rd129, %rd2, %rd128;

{ cvt.f32.f16 %f145, %rs1;}


add.s64 %rd130, %rd129, %rd5;
mul.lo.s64 %rd131, %rd130, 36;
add.s64 %rd132, %rd28, %rd131;
add.s64 %rd133, %rd132, %rd77;
ld.global.nc.u32 %r595, [%rd133+4];
ld.global.nc.u32 %r585, [%rd132];

{.reg .f16 low,high;
mov.b32 {low,high},%r585;
cvt.f32.f16 %f146, low;}


ld.global.nc.u32 %r603, [%rd133+40];
ld.global.nc.u32 %r586, [%rd132+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r586;
cvt.f32.f16 %f147, low;}


ld.global.nc.u32 %r611, [%rd133+76];
ld.global.nc.u32 %r587, [%rd132+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r587;
cvt.f32.f16 %f148, low;}


ld.global.nc.u32 %r619, [%rd133+112];
ld.global.nc.u32 %r588, [%rd132+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r588;
cvt.f32.f16 %f149, low;}



prmt.b32 %r589, %r589, 0, 0xba98;


prmt.b32 %r591, %r591, 0, 0xba98;

xor.b32 %r1046, %r591, 2139062143;
not.b32 %r1047, %r589;
and.b32 %r1048, %r676, %r1047;
and.b32 %r1049, %r1046, %r589;
or.b32 %r594, %r1049, %r1048;

dp4a.s32.s32 %r593, %r594, %r595, %r656;

mul.lo.s32 %r1050, %r671, %r593;
cvt.rn.f32.s32 %f251, %r1050;
fma.rn.ftz.f32 %f252, %f146, %f251, 0f00000000;

prmt.b32 %r597, %r597, 0, 0xba98;


prmt.b32 %r599, %r599, 0, 0xba98;

xor.b32 %r1051, %r599, 2139062143;
not.b32 %r1052, %r597;
and.b32 %r1053, %r699, %r1052;
and.b32 %r1054, %r1051, %r597;
or.b32 %r602, %r1054, %r1053;

dp4a.s32.s32 %r601, %r602, %r603, %r656;

mul.lo.s32 %r1055, %r693, %r601;
cvt.rn.f32.s32 %f253, %r1055;
fma.rn.ftz.f32 %f254, %f147, %f253, %f252;

prmt.b32 %r605, %r605, 0, 0xba98;


prmt.b32 %r607, %r607, 0, 0xba98;

xor.b32 %r1056, %r607, 2139062143;
not.b32 %r1057, %r605;
and.b32 %r1058, %r721, %r1057;
and.b32 %r1059, %r1056, %r605;
or.b32 %r610, %r1059, %r1058;

dp4a.s32.s32 %r609, %r610, %r611, %r656;

mul.lo.s32 %r1060, %r716, %r609;
cvt.rn.f32.s32 %f255, %r1060;
fma.rn.ftz.f32 %f256, %f148, %f255, %f254;

prmt.b32 %r613, %r613, 0, 0xba98;


prmt.b32 %r615, %r615, 0, 0xba98;

xor.b32 %r1061, %r615, 2139062143;
not.b32 %r1062, %r613;
and.b32 %r1063, %r744, %r1062;
and.b32 %r1064, %r1061, %r613;
or.b32 %r618, %r1064, %r1063;

dp4a.s32.s32 %r617, %r618, %r619, %r656;

mul.lo.s32 %r1065, %r738, %r617;
cvt.rn.f32.s32 %f257, %r1065;
fma.rn.ftz.f32 %f258, %f149, %f257, %f256;
fma.rn.ftz.f32 %f436, %f145, %f258, %f436;

{ cvt.f32.f16 %f150, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r585;
cvt.f32.f16 %f151, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r586;
cvt.f32.f16 %f152, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r587;
cvt.f32.f16 %f153, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r588;
cvt.f32.f16 %f154, low;}



prmt.b32 %r625, %r625, 0, 0xba98;


prmt.b32 %r627, %r627, 0, 0xba98;

xor.b32 %r1066, %r627, 2139062143;
not.b32 %r1067, %r625;
and.b32 %r1068, %r770, %r1067;
and.b32 %r1069, %r1066, %r625;
or.b32 %r630, %r1069, %r1068;

dp4a.s32.s32 %r629, %r630, %r595, %r656;

mul.lo.s32 %r1070, %r765, %r629;
cvt.rn.f32.s32 %f259, %r1070;
fma.rn.ftz.f32 %f260, %f151, %f259, 0f00000000;

prmt.b32 %r633, %r633, 0, 0xba98;


prmt.b32 %r635, %r635, 0, 0xba98;

xor.b32 %r1071, %r635, 2139062143;
not.b32 %r1072, %r633;
and.b32 %r1073, %r793, %r1072;
and.b32 %r1074, %r1071, %r633;
or.b32 %r638, %r1074, %r1073;

dp4a.s32.s32 %r637, %r638, %r603, %r656;

mul.lo.s32 %r1075, %r787, %r637;
cvt.rn.f32.s32 %f261, %r1075;
fma.rn.ftz.f32 %f262, %f152, %f261, %f260;

prmt.b32 %r641, %r641, 0, 0xba98;


prmt.b32 %r643, %r643, 0, 0xba98;

xor.b32 %r1076, %r643, 2139062143;
not.b32 %r1077, %r641;
and.b32 %r1078, %r815, %r1077;
and.b32 %r1079, %r1076, %r641;
or.b32 %r646, %r1079, %r1078;

dp4a.s32.s32 %r645, %r646, %r611, %r656;

mul.lo.s32 %r1080, %r810, %r645;
cvt.rn.f32.s32 %f263, %r1080;
fma.rn.ftz.f32 %f264, %f153, %f263, %f262;

prmt.b32 %r649, %r649, 0, 0xba98;


prmt.b32 %r651, %r651, 0, 0xba98;

xor.b32 %r1081, %r651, 2139062143;
not.b32 %r1082, %r649;
and.b32 %r1083, %r838, %r1082;
and.b32 %r1084, %r1081, %r649;
or.b32 %r654, %r1084, %r1083;

dp4a.s32.s32 %r653, %r654, %r619, %r656;

mul.lo.s32 %r1085, %r832, %r653;
cvt.rn.f32.s32 %f265, %r1085;
fma.rn.ftz.f32 %f266, %f154, %f265, %f264;
fma.rn.ftz.f32 %f435, %f150, %f266, %f435;
add.s32 %r1371, %r1371, 32;
add.s32 %r1370, %r1370, 32;
add.s32 %r1369, %r1369, 32;
add.s32 %r1368, %r1368, 32;
add.s32 %r1367, %r1367, 32;
add.s32 %r1366, %r1366, 32;
add.s32 %r1365, %r1365, 32;
add.s64 %rd168, %rd168, 440;
add.s32 %r1372, %r1372, 4;
setp.lt.s32 %p4, %r1372, %r4;
@%p4 bra $L__BB151_2;

$L__BB151_3:
mov.u32 %r1348, %tid.y;
setp.eq.s32 %p5, %r1348, 0;
@%p5 bra $L__BB151_5;

mov.u32 %r1350, %tid.x;
mov.u32 %r1349, %tid.y;
mov.u32 %r1087, _ZZ13mul_mat_vec_qIL9ggml_type11ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r1088, %r1349, 1792, %r1087;
shl.b32 %r1089, %r1350, 2;
add.s32 %r1090, %r1088, %r1089;
st.shared.f32 [%r1090+-1792], %f448;
st.shared.f32 [%r1090+-1664], %f447;
st.shared.f32 [%r1090+-1536], %f446;
st.shared.f32 [%r1090+-1408], %f445;
st.shared.f32 [%r1090+-1280], %f444;
st.shared.f32 [%r1090+-1152], %f443;
st.shared.f32 [%r1090+-1024], %f442;
st.shared.f32 [%r1090+-896], %f441;
st.shared.f32 [%r1090+-768], %f440;
st.shared.f32 [%r1090+-640], %f439;
st.shared.f32 [%r1090+-512], %f438;
st.shared.f32 [%r1090+-384], %f437;
st.shared.f32 [%r1090+-256], %f436;
st.shared.f32 [%r1090+-128], %f435;

$L__BB151_5:
mov.u32 %r1351, %tid.y;
setp.ne.s32 %p6, %r1351, 0;
bar.sync 0;
@%p6 bra $L__BB151_34;

ld.param.u32 %r1355, [_Z13mul_mat_vec_qIL9ggml_type11ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r1354, [_Z13mul_mat_vec_qIL9ggml_type11ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r1353, [_Z13mul_mat_vec_qIL9ggml_type11ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
mov.u32 %r1352, %tid.x;
mov.u32 %r1091, %ctaid.z;
mov.u32 %r1092, %ctaid.y;
mov.u32 %r1093, %ctaid.x;
shl.b32 %r1094, %r1093, 1;
mov.u32 %r1095, 1;
mad.lo.s32 %r1096, %r1092, %r1353, %r1094;
mad.lo.s32 %r1097, %r1091, %r1354, %r1096;
cvt.u64.u32 %rd31, %r1097;
shl.b32 %r1099, %r1352, 2;
mov.u32 %r1100, _ZZ13mul_mat_vec_qIL9ggml_type11ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r36, %r1100, %r1099;
mov.u32 %r1101, 2;
ld.shared.f32 %f267, [%r36];
add.ftz.f32 %f268, %f267, %f448;
mov.b32 %r1102, %f268;
mov.u32 %r1103, 31;
mov.u32 %r1104, 16;
mov.u32 %r1105, -1;
shfl.sync.bfly.b32 %r1106|%p7, %r1102, %r1104, %r1103, %r1105;
mov.b32 %f269, %r1106;
add.ftz.f32 %f270, %f268, %f269;
mov.b32 %r1107, %f270;
mov.u32 %r1108, 8;
shfl.sync.bfly.b32 %r1109|%p8, %r1107, %r1108, %r1103, %r1105;
mov.b32 %f271, %r1109;
add.ftz.f32 %f272, %f270, %f271;
mov.b32 %r1110, %f272;
mov.u32 %r1111, 4;
shfl.sync.bfly.b32 %r1112|%p9, %r1110, %r1111, %r1103, %r1105;
mov.b32 %f273, %r1112;
add.ftz.f32 %f274, %f272, %f273;
mov.b32 %r1113, %f274;
shfl.sync.bfly.b32 %r1114|%p10, %r1113, %r1101, %r1103, %r1105;
mov.b32 %f275, %r1114;
add.ftz.f32 %f276, %f274, %f275;
mov.b32 %r1115, %f276;
shfl.sync.bfly.b32 %r1116|%p11, %r1115, %r1095, %r1103, %r1105;
mov.b32 %f277, %r1116;
add.ftz.f32 %f43, %f276, %f277;
setp.eq.s32 %p12, %r1352, 0;
setp.lt.u32 %p13, %r1094, %r1355;
and.pred %p1, %p12, %p13;
mul.wide.u32 %rd134, %r1097, 4;
add.s64 %rd32, %rd1, %rd134;
not.pred %p14, %p1;
@%p14 bra $L__BB151_8;

st.global.f32 [%rd32], %f43;

$L__BB151_8:
ld.param.u32 %r1357, [_Z13mul_mat_vec_qIL9ggml_type11ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
mov.u32 %r1356, %tid.x;
ld.shared.f32 %f278, [%r36+128];
add.ftz.f32 %f279, %f278, %f447;
mov.b32 %r1117, %f279;
shfl.sync.bfly.b32 %r1121|%p15, %r1117, %r1104, %r1103, %r1105;
mov.b32 %f280, %r1121;
add.ftz.f32 %f281, %f279, %f280;
mov.b32 %r1122, %f281;
shfl.sync.bfly.b32 %r1124|%p16, %r1122, %r1108, %r1103, %r1105;
mov.b32 %f282, %r1124;
add.ftz.f32 %f283, %f281, %f282;
mov.b32 %r1125, %f283;
shfl.sync.bfly.b32 %r1127|%p17, %r1125, %r1111, %r1103, %r1105;
mov.b32 %f284, %r1127;
add.ftz.f32 %f285, %f283, %f284;
mov.b32 %r1128, %f285;
shfl.sync.bfly.b32 %r1130|%p18, %r1128, %r1101, %r1103, %r1105;
mov.b32 %f286, %r1130;
add.ftz.f32 %f287, %f285, %f286;
mov.b32 %r1131, %f287;
shfl.sync.bfly.b32 %r1133|%p19, %r1131, %r1095, %r1103, %r1105;
mov.b32 %f288, %r1133;
add.ftz.f32 %f44, %f287, %f288;
setp.eq.s32 %p20, %r1356, 1;
or.b32 %r1137, %r1094, 1;
setp.lt.u32 %p21, %r1137, %r1357;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB151_10;

st.global.f32 [%rd32+4], %f44;

$L__BB151_10:
ld.shared.f32 %f289, [%r36+256];
add.ftz.f32 %f290, %f289, %f446;
mov.b32 %r1138, %f290;
mov.u32 %r1139, 31;
mov.u32 %r1140, 16;
mov.u32 %r1141, -1;
shfl.sync.bfly.b32 %r1142|%p23, %r1138, %r1140, %r1139, %r1141;
mov.b32 %f291, %r1142;
add.ftz.f32 %f292, %f290, %f291;
mov.b32 %r1143, %f292;
mov.u32 %r1144, 8;
shfl.sync.bfly.b32 %r1145|%p24, %r1143, %r1144, %r1139, %r1141;
mov.b32 %f293, %r1145;
add.ftz.f32 %f294, %f292, %f293;
mov.b32 %r1146, %f294;
mov.u32 %r1147, 4;
shfl.sync.bfly.b32 %r1148|%p25, %r1146, %r1147, %r1139, %r1141;
mov.b32 %f295, %r1148;
add.ftz.f32 %f296, %f294, %f295;
mov.b32 %r1149, %f296;
mov.u32 %r1150, 2;
shfl.sync.bfly.b32 %r1151|%p26, %r1149, %r1150, %r1139, %r1141;
mov.b32 %f297, %r1151;
add.ftz.f32 %f298, %f296, %f297;
mov.b32 %r1152, %f298;
mov.u32 %r1153, 1;
shfl.sync.bfly.b32 %r1154|%p27, %r1152, %r1153, %r1139, %r1141;
mov.b32 %f299, %r1154;
add.ftz.f32 %f45, %f298, %f299;
@%p14 bra $L__BB151_12;

ld.param.u32 %r1364, [_Z13mul_mat_vec_qIL9ggml_type11ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
cvt.u64.u32 %rd135, %r1364;
add.s64 %rd136, %rd135, %rd31;
shl.b64 %rd137, %rd136, 2;
add.s64 %rd138, %rd1, %rd137;
st.global.f32 [%rd138], %f45;

$L__BB151_12:
ld.shared.f32 %f300, [%r36+384];
add.ftz.f32 %f301, %f300, %f445;
mov.b32 %r1155, %f301;
shfl.sync.bfly.b32 %r1159|%p29, %r1155, %r1140, %r1139, %r1141;
mov.b32 %f302, %r1159;
add.ftz.f32 %f303, %f301, %f302;
mov.b32 %r1160, %f303;
shfl.sync.bfly.b32 %r1162|%p30, %r1160, %r1144, %r1139, %r1141;
mov.b32 %f304, %r1162;
add.ftz.f32 %f305, %f303, %f304;
mov.b32 %r1163, %f305;
shfl.sync.bfly.b32 %r1165|%p31, %r1163, %r1147, %r1139, %r1141;
mov.b32 %f306, %r1165;
add.ftz.f32 %f307, %f305, %f306;
mov.b32 %r1166, %f307;
shfl.sync.bfly.b32 %r1168|%p32, %r1166, %r1150, %r1139, %r1141;
mov.b32 %f308, %r1168;
add.ftz.f32 %f309, %f307, %f308;
mov.b32 %r1169, %f309;
shfl.sync.bfly.b32 %r1171|%p33, %r1169, %r1153, %r1139, %r1141;
mov.b32 %f310, %r1171;
add.ftz.f32 %f46, %f309, %f310;
@%p22 bra $L__BB151_14;

ld.param.u32 %r1363, [_Z13mul_mat_vec_qIL9ggml_type11ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
add.s32 %r1172, %r1363, 1;
cvt.u64.u32 %rd139, %r1172;
add.s64 %rd140, %rd139, %rd31;
shl.b64 %rd141, %rd140, 2;
add.s64 %rd142, %rd1, %rd141;
st.global.f32 [%rd142], %f46;

$L__BB151_14:
ld.param.u32 %r1358, [_Z13mul_mat_vec_qIL9ggml_type11ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
shl.b32 %r1173, %r1358, 1;
mov.u32 %r1174, 1;
ld.shared.f32 %f311, [%r36+512];
add.ftz.f32 %f312, %f311, %f444;
mov.b32 %r1175, %f312;
mov.u32 %r1176, 31;
mov.u32 %r1177, 16;
mov.u32 %r1178, -1;
shfl.sync.bfly.b32 %r1179|%p35, %r1175, %r1177, %r1176, %r1178;
mov.b32 %f313, %r1179;
add.ftz.f32 %f314, %f312, %f313;
mov.b32 %r1180, %f314;
mov.u32 %r1181, 8;
shfl.sync.bfly.b32 %r1182|%p36, %r1180, %r1181, %r1176, %r1178;
mov.b32 %f315, %r1182;
add.ftz.f32 %f316, %f314, %f315;
mov.b32 %r1183, %f316;
mov.u32 %r1184, 4;
shfl.sync.bfly.b32 %r1185|%p37, %r1183, %r1184, %r1176, %r1178;
mov.b32 %f317, %r1185;
add.ftz.f32 %f318, %f316, %f317;
mov.b32 %r1186, %f318;
mov.u32 %r1187, 2;
shfl.sync.bfly.b32 %r1188|%p38, %r1186, %r1187, %r1176, %r1178;
mov.b32 %f319, %r1188;
add.ftz.f32 %f320, %f318, %f319;
mov.b32 %r1189, %f320;
shfl.sync.bfly.b32 %r1190|%p39, %r1189, %r1174, %r1176, %r1178;
mov.b32 %f321, %r1190;
add.ftz.f32 %f47, %f320, %f321;
cvt.u64.u32 %rd143, %r1173;
add.s64 %rd144, %rd143, %rd31;
shl.b64 %rd145, %rd144, 2;
add.s64 %rd33, %rd1, %rd145;
@%p14 bra $L__BB151_16;

st.global.f32 [%rd33], %f47;

$L__BB151_16:
ld.shared.f32 %f322, [%r36+640];
add.ftz.f32 %f323, %f322, %f443;
mov.b32 %r1191, %f323;
shfl.sync.bfly.b32 %r1195|%p41, %r1191, %r1177, %r1176, %r1178;
mov.b32 %f324, %r1195;
add.ftz.f32 %f325, %f323, %f324;
mov.b32 %r1196, %f325;
shfl.sync.bfly.b32 %r1198|%p42, %r1196, %r1181, %r1176, %r1178;
mov.b32 %f326, %r1198;
add.ftz.f32 %f327, %f325, %f326;
mov.b32 %r1199, %f327;
shfl.sync.bfly.b32 %r1201|%p43, %r1199, %r1184, %r1176, %r1178;
mov.b32 %f328, %r1201;
add.ftz.f32 %f329, %f327, %f328;
mov.b32 %r1202, %f329;
shfl.sync.bfly.b32 %r1204|%p44, %r1202, %r1187, %r1176, %r1178;
mov.b32 %f330, %r1204;
add.ftz.f32 %f331, %f329, %f330;
mov.b32 %r1205, %f331;
shfl.sync.bfly.b32 %r1207|%p45, %r1205, %r1174, %r1176, %r1178;
mov.b32 %f332, %r1207;
add.ftz.f32 %f48, %f331, %f332;
@%p22 bra $L__BB151_18;

st.global.f32 [%rd33+4], %f48;

$L__BB151_18:
ld.param.u32 %r1359, [_Z13mul_mat_vec_qIL9ggml_type11ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
mul.lo.s32 %r37, %r1359, 3;
ld.shared.f32 %f333, [%r36+768];
add.ftz.f32 %f334, %f333, %f442;
mov.b32 %r1208, %f334;
mov.u32 %r1209, 31;
mov.u32 %r1210, 16;
mov.u32 %r1211, -1;
shfl.sync.bfly.b32 %r1212|%p47, %r1208, %r1210, %r1209, %r1211;
mov.b32 %f335, %r1212;
add.ftz.f32 %f336, %f334, %f335;
mov.b32 %r1213, %f336;
mov.u32 %r1214, 8;
shfl.sync.bfly.b32 %r1215|%p48, %r1213, %r1214, %r1209, %r1211;
mov.b32 %f337, %r1215;
add.ftz.f32 %f338, %f336, %f337;
mov.b32 %r1216, %f338;
mov.u32 %r1217, 4;
shfl.sync.bfly.b32 %r1218|%p49, %r1216, %r1217, %r1209, %r1211;
mov.b32 %f339, %r1218;
add.ftz.f32 %f340, %f338, %f339;
mov.b32 %r1219, %f340;
mov.u32 %r1220, 2;
shfl.sync.bfly.b32 %r1221|%p50, %r1219, %r1220, %r1209, %r1211;
mov.b32 %f341, %r1221;
add.ftz.f32 %f342, %f340, %f341;
mov.b32 %r1222, %f342;
mov.u32 %r1223, 1;
shfl.sync.bfly.b32 %r1224|%p51, %r1222, %r1223, %r1209, %r1211;
mov.b32 %f343, %r1224;
add.ftz.f32 %f49, %f342, %f343;
@%p14 bra $L__BB151_20;

cvt.u64.u32 %rd146, %r37;
add.s64 %rd147, %rd146, %rd31;
shl.b64 %rd148, %rd147, 2;
add.s64 %rd149, %rd1, %rd148;
st.global.f32 [%rd149], %f49;

$L__BB151_20:
ld.shared.f32 %f344, [%r36+896];
add.ftz.f32 %f345, %f344, %f441;
mov.b32 %r1225, %f345;
shfl.sync.bfly.b32 %r1229|%p53, %r1225, %r1210, %r1209, %r1211;
mov.b32 %f346, %r1229;
add.ftz.f32 %f347, %f345, %f346;
mov.b32 %r1230, %f347;
shfl.sync.bfly.b32 %r1232|%p54, %r1230, %r1214, %r1209, %r1211;
mov.b32 %f348, %r1232;
add.ftz.f32 %f349, %f347, %f348;
mov.b32 %r1233, %f349;
shfl.sync.bfly.b32 %r1235|%p55, %r1233, %r1217, %r1209, %r1211;
mov.b32 %f350, %r1235;
add.ftz.f32 %f351, %f349, %f350;
mov.b32 %r1236, %f351;
shfl.sync.bfly.b32 %r1238|%p56, %r1236, %r1220, %r1209, %r1211;
mov.b32 %f352, %r1238;
add.ftz.f32 %f353, %f351, %f352;
mov.b32 %r1239, %f353;
shfl.sync.bfly.b32 %r1241|%p57, %r1239, %r1223, %r1209, %r1211;
mov.b32 %f354, %r1241;
add.ftz.f32 %f50, %f353, %f354;
@%p22 bra $L__BB151_22;

add.s32 %r1242, %r37, 1;
cvt.u64.u32 %rd150, %r1242;
add.s64 %rd151, %rd150, %rd31;
shl.b64 %rd152, %rd151, 2;
add.s64 %rd153, %rd1, %rd152;
st.global.f32 [%rd153], %f50;

$L__BB151_22:
ld.param.u32 %r1360, [_Z13mul_mat_vec_qIL9ggml_type11ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
shl.b32 %r1243, %r1360, 2;
mov.u32 %r1244, 2;
ld.shared.f32 %f355, [%r36+1024];
add.ftz.f32 %f356, %f355, %f440;
mov.b32 %r1245, %f356;
mov.u32 %r1246, 31;
mov.u32 %r1247, 16;
mov.u32 %r1248, -1;
shfl.sync.bfly.b32 %r1249|%p59, %r1245, %r1247, %r1246, %r1248;
mov.b32 %f357, %r1249;
add.ftz.f32 %f358, %f356, %f357;
mov.b32 %r1250, %f358;
mov.u32 %r1251, 8;
shfl.sync.bfly.b32 %r1252|%p60, %r1250, %r1251, %r1246, %r1248;
mov.b32 %f359, %r1252;
add.ftz.f32 %f360, %f358, %f359;
mov.b32 %r1253, %f360;
mov.u32 %r1254, 4;
shfl.sync.bfly.b32 %r1255|%p61, %r1253, %r1254, %r1246, %r1248;
mov.b32 %f361, %r1255;
add.ftz.f32 %f362, %f360, %f361;
mov.b32 %r1256, %f362;
shfl.sync.bfly.b32 %r1257|%p62, %r1256, %r1244, %r1246, %r1248;
mov.b32 %f363, %r1257;
add.ftz.f32 %f364, %f362, %f363;
mov.b32 %r1258, %f364;
mov.u32 %r1259, 1;
shfl.sync.bfly.b32 %r1260|%p63, %r1258, %r1259, %r1246, %r1248;
mov.b32 %f365, %r1260;
add.ftz.f32 %f51, %f364, %f365;
cvt.u64.u32 %rd154, %r1243;
add.s64 %rd155, %rd154, %rd31;
shl.b64 %rd156, %rd155, 2;
add.s64 %rd34, %rd1, %rd156;
@%p14 bra $L__BB151_24;

st.global.f32 [%rd34], %f51;

$L__BB151_24:
ld.shared.f32 %f366, [%r36+1152];
add.ftz.f32 %f367, %f366, %f439;
mov.b32 %r1261, %f367;
shfl.sync.bfly.b32 %r1265|%p65, %r1261, %r1247, %r1246, %r1248;
mov.b32 %f368, %r1265;
add.ftz.f32 %f369, %f367, %f368;
mov.b32 %r1266, %f369;
shfl.sync.bfly.b32 %r1268|%p66, %r1266, %r1251, %r1246, %r1248;
mov.b32 %f370, %r1268;
add.ftz.f32 %f371, %f369, %f370;
mov.b32 %r1269, %f371;
shfl.sync.bfly.b32 %r1271|%p67, %r1269, %r1254, %r1246, %r1248;
mov.b32 %f372, %r1271;
add.ftz.f32 %f373, %f371, %f372;
mov.b32 %r1272, %f373;
shfl.sync.bfly.b32 %r1274|%p68, %r1272, %r1244, %r1246, %r1248;
mov.b32 %f374, %r1274;
add.ftz.f32 %f375, %f373, %f374;
mov.b32 %r1275, %f375;
shfl.sync.bfly.b32 %r1277|%p69, %r1275, %r1259, %r1246, %r1248;
mov.b32 %f376, %r1277;
add.ftz.f32 %f52, %f375, %f376;
@%p22 bra $L__BB151_26;

st.global.f32 [%rd34+4], %f52;

$L__BB151_26:
ld.param.u32 %r1361, [_Z13mul_mat_vec_qIL9ggml_type11ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
mul.lo.s32 %r38, %r1361, 5;
ld.shared.f32 %f377, [%r36+1280];
add.ftz.f32 %f378, %f377, %f438;
mov.b32 %r1278, %f378;
mov.u32 %r1279, 31;
mov.u32 %r1280, 16;
mov.u32 %r1281, -1;
shfl.sync.bfly.b32 %r1282|%p71, %r1278, %r1280, %r1279, %r1281;
mov.b32 %f379, %r1282;
add.ftz.f32 %f380, %f378, %f379;
mov.b32 %r1283, %f380;
mov.u32 %r1284, 8;
shfl.sync.bfly.b32 %r1285|%p72, %r1283, %r1284, %r1279, %r1281;
mov.b32 %f381, %r1285;
add.ftz.f32 %f382, %f380, %f381;
mov.b32 %r1286, %f382;
mov.u32 %r1287, 4;
shfl.sync.bfly.b32 %r1288|%p73, %r1286, %r1287, %r1279, %r1281;
mov.b32 %f383, %r1288;
add.ftz.f32 %f384, %f382, %f383;
mov.b32 %r1289, %f384;
mov.u32 %r1290, 2;
shfl.sync.bfly.b32 %r1291|%p74, %r1289, %r1290, %r1279, %r1281;
mov.b32 %f385, %r1291;
add.ftz.f32 %f386, %f384, %f385;
mov.b32 %r1292, %f386;
mov.u32 %r1293, 1;
shfl.sync.bfly.b32 %r1294|%p75, %r1292, %r1293, %r1279, %r1281;
mov.b32 %f387, %r1294;
add.ftz.f32 %f53, %f386, %f387;
@%p14 bra $L__BB151_28;

cvt.u64.u32 %rd157, %r38;
add.s64 %rd158, %rd157, %rd31;
shl.b64 %rd159, %rd158, 2;
add.s64 %rd160, %rd1, %rd159;
st.global.f32 [%rd160], %f53;

$L__BB151_28:
ld.shared.f32 %f388, [%r36+1408];
add.ftz.f32 %f389, %f388, %f437;
mov.b32 %r1295, %f389;
shfl.sync.bfly.b32 %r1299|%p77, %r1295, %r1280, %r1279, %r1281;
mov.b32 %f390, %r1299;
add.ftz.f32 %f391, %f389, %f390;
mov.b32 %r1300, %f391;
shfl.sync.bfly.b32 %r1302|%p78, %r1300, %r1284, %r1279, %r1281;
mov.b32 %f392, %r1302;
add.ftz.f32 %f393, %f391, %f392;
mov.b32 %r1303, %f393;
shfl.sync.bfly.b32 %r1305|%p79, %r1303, %r1287, %r1279, %r1281;
mov.b32 %f394, %r1305;
add.ftz.f32 %f395, %f393, %f394;
mov.b32 %r1306, %f395;
shfl.sync.bfly.b32 %r1308|%p80, %r1306, %r1290, %r1279, %r1281;
mov.b32 %f396, %r1308;
add.ftz.f32 %f397, %f395, %f396;
mov.b32 %r1309, %f397;
shfl.sync.bfly.b32 %r1311|%p81, %r1309, %r1293, %r1279, %r1281;
mov.b32 %f398, %r1311;
add.ftz.f32 %f54, %f397, %f398;
@%p22 bra $L__BB151_30;

add.s32 %r1312, %r38, 1;
cvt.u64.u32 %rd161, %r1312;
add.s64 %rd162, %rd161, %rd31;
shl.b64 %rd163, %rd162, 2;
add.s64 %rd164, %rd1, %rd163;
st.global.f32 [%rd164], %f54;

$L__BB151_30:
ld.param.u32 %r1362, [_Z13mul_mat_vec_qIL9ggml_type11ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
mul.lo.s32 %r1313, %r1362, 6;
ld.shared.f32 %f399, [%r36+1536];
add.ftz.f32 %f400, %f399, %f436;
mov.b32 %r1314, %f400;
mov.u32 %r1315, 31;
mov.u32 %r1316, 16;
mov.u32 %r1317, -1;
shfl.sync.bfly.b32 %r1318|%p83, %r1314, %r1316, %r1315, %r1317;
mov.b32 %f401, %r1318;
add.ftz.f32 %f402, %f400, %f401;
mov.b32 %r1319, %f402;
mov.u32 %r1320, 8;
shfl.sync.bfly.b32 %r1321|%p84, %r1319, %r1320, %r1315, %r1317;
mov.b32 %f403, %r1321;
add.ftz.f32 %f404, %f402, %f403;
mov.b32 %r1322, %f404;
mov.u32 %r1323, 4;
shfl.sync.bfly.b32 %r1324|%p85, %r1322, %r1323, %r1315, %r1317;
mov.b32 %f405, %r1324;
add.ftz.f32 %f406, %f404, %f405;
mov.b32 %r1325, %f406;
mov.u32 %r1326, 2;
shfl.sync.bfly.b32 %r1327|%p86, %r1325, %r1326, %r1315, %r1317;
mov.b32 %f407, %r1327;
add.ftz.f32 %f408, %f406, %f407;
mov.b32 %r1328, %f408;
mov.u32 %r1329, 1;
shfl.sync.bfly.b32 %r1330|%p87, %r1328, %r1329, %r1315, %r1317;
mov.b32 %f409, %r1330;
add.ftz.f32 %f55, %f408, %f409;
cvt.u64.u32 %rd165, %r1313;
add.s64 %rd166, %rd165, %rd31;
shl.b64 %rd167, %rd166, 2;
add.s64 %rd35, %rd1, %rd167;
@%p14 bra $L__BB151_32;

st.global.f32 [%rd35], %f55;

$L__BB151_32:
ld.shared.f32 %f410, [%r36+1664];
add.ftz.f32 %f411, %f410, %f435;
mov.b32 %r1331, %f411;
shfl.sync.bfly.b32 %r1335|%p89, %r1331, %r1316, %r1315, %r1317;
mov.b32 %f412, %r1335;
add.ftz.f32 %f413, %f411, %f412;
mov.b32 %r1336, %f413;
shfl.sync.bfly.b32 %r1338|%p90, %r1336, %r1320, %r1315, %r1317;
mov.b32 %f414, %r1338;
add.ftz.f32 %f415, %f413, %f414;
mov.b32 %r1339, %f415;
shfl.sync.bfly.b32 %r1341|%p91, %r1339, %r1323, %r1315, %r1317;
mov.b32 %f416, %r1341;
add.ftz.f32 %f417, %f415, %f416;
mov.b32 %r1342, %f417;
shfl.sync.bfly.b32 %r1344|%p92, %r1342, %r1326, %r1315, %r1317;
mov.b32 %f418, %r1344;
add.ftz.f32 %f419, %f417, %f418;
mov.b32 %r1345, %f419;
shfl.sync.bfly.b32 %r1347|%p93, %r1345, %r1329, %r1315, %r1317;
mov.b32 %f420, %r1347;
add.ftz.f32 %f56, %f419, %f420;
@%p22 bra $L__BB151_34;

st.global.f32 [%rd35+4], %f56;

$L__BB151_34:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type11ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type11ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type11ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type11ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type11ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type11ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type11ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type11ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type11ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type11ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<107>;
.reg .b16 %rs<41>;
.reg .f32 %f<513>;
.reg .b32 %r<1527>;
.reg .b64 %rd<183>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type11ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[2048];

ld.param.u64 %rd36, [_Z13mul_mat_vec_qIL9ggml_type11ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd37, [_Z13mul_mat_vec_qIL9ggml_type11ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd38, [_Z13mul_mat_vec_qIL9ggml_type11ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r58, [_Z13mul_mat_vec_qIL9ggml_type11ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type11ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type11ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r49, [_Z13mul_mat_vec_qIL9ggml_type11ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r50, [_Z13mul_mat_vec_qIL9ggml_type11ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r55, [_Z13mul_mat_vec_qIL9ggml_type11ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r56, [_Z13mul_mat_vec_qIL9ggml_type11ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r53, [_Z13mul_mat_vec_qIL9ggml_type11ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r52, [_Z13mul_mat_vec_qIL9ggml_type11ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r47, [_Z13mul_mat_vec_qIL9ggml_type11ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r46, [_Z13mul_mat_vec_qIL9ggml_type11ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
cvta.to.global.u64 %rd1, %rd38;
mov.u32 %r1, %tid.y;
shl.b32 %r59, %r1, 5;
mov.u32 %r60, %tid.x;
add.s32 %r2, %r59, %r60;
shr.s32 %r61, %r2, 31;
shr.u32 %r62, %r61, 28;
add.s32 %r63, %r2, %r62;
shr.s32 %r1526, %r63, 4;
shr.u32 %r4, %r58, 8;
setp.ge.s32 %p3, %r1526, %r4;
mov.f32 %f497, 0f00000000;
mov.f32 %f498, %f497;
mov.f32 %f499, %f497;
mov.f32 %f500, %f497;
mov.f32 %f501, %f497;
mov.f32 %f502, %f497;
mov.f32 %f503, %f497;
mov.f32 %f504, %f497;
mov.f32 %f505, %f497;
mov.f32 %f506, %f497;
mov.f32 %f507, %f497;
mov.f32 %f508, %f497;
mov.f32 %f509, %f497;
mov.f32 %f510, %f497;
mov.f32 %f511, %f497;
mov.f32 %f512, %f497;
@%p3 bra $L__BB152_3;

mov.u32 %r64, %ctaid.x;
shl.b32 %r65, %r64, 1;
mov.u32 %r66, %ctaid.y;
mov.u32 %r67, %ctaid.z;
mul.lo.s32 %r68, %r66, %r50;
cvt.u64.u32 %rd39, %r68;
mul.lo.s32 %r69, %r67, %r56;
cvt.u64.u32 %rd40, %r69;
add.s64 %rd2, %rd40, %rd39;
mul.hi.u32 %r70, %r67, %r52;
add.s32 %r71, %r70, %r67;
shr.u32 %r72, %r71, %r53;
mul.lo.s32 %r73, %r72, %r55;
mul.hi.u32 %r74, %r66, %r46;
add.s32 %r75, %r74, %r66;
shr.u32 %r76, %r75, %r47;
mul.lo.s32 %r77, %r76, %r49;
cvta.to.global.u64 %rd182, %rd36;
or.b32 %r78, %r65, 1;
and.b32 %r82, %r63, -16;
sub.s32 %r83, %r2, %r82;
shr.s32 %r84, %r83, 31;
shr.u32 %r85, %r84, 29;
add.s32 %r86, %r83, %r85;
and.b32 %r87, %r86, -8;
sub.s32 %r88, %r83, %r87;
shr.s32 %r89, %r88, 31;
shr.u32 %r90, %r89, 30;
add.s32 %r91, %r88, %r90;
shr.s32 %r92, %r91, 2;
add.s32 %r93, %r87, %r92;
shr.s32 %r94, %r93, 31;
shr.u32 %r95, %r94, 29;
add.s32 %r96, %r93, %r95;
and.b32 %r97, %r96, -8;
sub.s32 %r98, %r93, %r97;
cvt.s64.s32 %rd41, %r98;
shr.s32 %r99, %r86, 3;
shl.b32 %r100, %r99, 2;
shl.b32 %r101, %r83, 1;
shl.b32 %r102, %r88, 1;
cvt.s64.s32 %rd4, %r88;
shr.s32 %r103, %r96, 3;
shl.b32 %r5, %r103, 2;
cvt.s64.s32 %rd5, %r100;
shr.u32 %r104, %r94, 30;
add.s32 %r105, %r93, %r104;
and.b32 %r106, %r105, -4;
sub.s32 %r107, %r93, %r106;
shr.s32 %r108, %r105, 2;
shl.b32 %r6, %r108, 1;
add.s32 %r109, %r107, 8;
cvt.s64.s32 %rd42, %r109;
add.s32 %r110, %r93, 2;
shr.s32 %r111, %r110, 31;
shr.u32 %r112, %r111, 29;
add.s32 %r113, %r110, %r112;
and.b32 %r114, %r113, -8;
sub.s32 %r115, %r110, %r114;
shr.s32 %r116, %r113, 3;
shl.b32 %r7, %r116, 2;
cvt.s64.s32 %rd43, %r115;
shr.u32 %r117, %r111, 30;
add.s32 %r118, %r110, %r117;
and.b32 %r119, %r118, -4;
sub.s32 %r120, %r110, %r119;
shr.s32 %r121, %r118, 2;
shl.b32 %r8, %r121, 1;
add.s32 %r122, %r120, 8;
cvt.s64.s32 %rd44, %r122;
add.s32 %r123, %r93, 4;
shr.s32 %r124, %r123, 31;
shr.u32 %r125, %r124, 29;
add.s32 %r126, %r123, %r125;
and.b32 %r127, %r126, -8;
sub.s32 %r128, %r123, %r127;
shr.s32 %r129, %r126, 3;
shl.b32 %r9, %r129, 2;
cvt.s64.s32 %rd45, %r128;
shr.u32 %r130, %r124, 30;
add.s32 %r131, %r123, %r130;
and.b32 %r132, %r131, -4;
sub.s32 %r133, %r123, %r132;
shr.s32 %r134, %r131, 2;
shl.b32 %r10, %r134, 1;
add.s32 %r135, %r133, 8;
cvt.s64.s32 %rd46, %r135;
add.s32 %r136, %r93, 6;
shr.s32 %r137, %r136, 31;
shr.u32 %r138, %r137, 29;
add.s32 %r139, %r136, %r138;
and.b32 %r140, %r139, -8;
sub.s32 %r141, %r136, %r140;
shr.s32 %r142, %r139, 3;
shl.b32 %r11, %r142, 2;
cvt.s64.s32 %rd47, %r141;
shr.u32 %r143, %r137, 30;
add.s32 %r144, %r136, %r143;
and.b32 %r145, %r144, -4;
sub.s32 %r146, %r136, %r145;
shr.s32 %r147, %r144, 2;
shl.b32 %r12, %r147, 1;
add.s32 %r148, %r146, 8;
cvt.s64.s32 %rd48, %r148;
shl.b32 %r149, %r44, 1;
shl.b32 %r1518, %r1526, 3;
add.s32 %r1525, %r1518, %r149;
mad.lo.s32 %r1524, %r44, 3, %r1518;
shl.b32 %r150, %r44, 2;
add.s32 %r1523, %r1518, %r150;
mad.lo.s32 %r1522, %r44, 5, %r1518;
mad.lo.s32 %r1521, %r44, 6, %r1518;
mad.lo.s32 %r1520, %r44, 7, %r1518;
add.s32 %r1519, %r44, %r1518;
add.s32 %r151, %r1526, %r73;
add.s32 %r152, %r151, %r77;
mad.lo.s32 %r153, %r78, %r43, %r152;
mul.wide.s32 %rd6, %r153, 110;
add.s64 %rd49, %rd6, %rd48;
add.s64 %rd7, %rd49, 96;
add.s64 %rd50, %rd6, %rd47;
add.s64 %rd8, %rd50, 96;
add.s64 %rd51, %rd6, %rd46;
add.s64 %rd9, %rd51, 96;
add.s64 %rd52, %rd6, %rd45;
add.s64 %rd10, %rd52, 96;
add.s64 %rd53, %rd6, %rd44;
add.s64 %rd11, %rd53, 96;
add.s64 %rd54, %rd6, %rd43;
add.s64 %rd12, %rd54, 96;
add.s64 %rd55, %rd6, %rd42;
add.s64 %rd13, %rd55, 96;
add.s64 %rd56, %rd6, %rd41;
add.s64 %rd14, %rd56, 96;
mad.lo.s32 %r154, %r65, %r43, %r1526;
add.s32 %r155, %r154, %r73;
add.s32 %r156, %r155, %r77;
mul.wide.s32 %rd15, %r156, 110;
mul.wide.s32 %rd57, %r101, 2;
add.s64 %rd16, %rd15, %rd57;
mul.wide.s32 %rd58, %r102, 2;
add.s64 %rd17, %rd6, %rd58;
add.s64 %rd59, %rd15, %rd58;
add.s64 %rd18, %rd59, 2;
add.s64 %rd60, %rd6, %rd57;
add.s64 %rd19, %rd60, 32;
add.s64 %rd61, %rd15, %rd48;
add.s64 %rd20, %rd61, 96;
add.s64 %rd62, %rd15, %rd47;
add.s64 %rd21, %rd62, 96;
add.s64 %rd63, %rd15, %rd46;
add.s64 %rd22, %rd63, 96;
add.s64 %rd64, %rd15, %rd45;
add.s64 %rd23, %rd64, 96;
add.s64 %rd65, %rd15, %rd44;
add.s64 %rd24, %rd65, 96;
add.s64 %rd66, %rd15, %rd43;
add.s64 %rd25, %rd66, 96;
add.s64 %rd67, %rd15, %rd42;
add.s64 %rd26, %rd67, 96;
add.s64 %rd68, %rd15, %rd41;
add.s64 %rd27, %rd68, 96;
cvta.to.global.u64 %rd28, %rd37;
mov.f32 %f497, 0f00000000;
shl.b64 %rd77, %rd4, 2;

$L__BB152_2:
cvt.u64.u32 %rd69, %r1518;
add.s64 %rd70, %rd2, %rd69;
add.s64 %rd71, %rd182, %rd15;
ld.global.nc.u16 %rs1, [%rd71+108];

{ cvt.f32.f16 %f97, %rs1;}


add.s64 %rd72, %rd182, %rd16;
add.s64 %rd73, %rd182, %rd18;
ld.global.nc.u16 %rs17, [%rd73+-2];
ld.global.nc.u16 %rs18, [%rd73];
mov.b32 %r733, {%rs17, %rs18};
not.b32 %r734, %r733;
ld.global.nc.u16 %rs19, [%rd72+34];
ld.global.nc.u16 %rs20, [%rd72+32];
add.s64 %rd74, %rd70, %rd5;
mul.lo.s64 %rd75, %rd74, 36;
add.s64 %rd76, %rd28, %rd75;
add.s64 %rd78, %rd76, %rd77;
ld.global.nc.u32 %r167, [%rd78+4];
ld.global.nc.u32 %r157, [%rd76];

{.reg .f16 low,high;
mov.b32 {low,high},%r157;
cvt.f32.f16 %f98, low;}


ld.global.nc.u32 %r175, [%rd78+40];
ld.global.nc.u32 %r158, [%rd76+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r158;
cvt.f32.f16 %f99, low;}


ld.global.nc.u32 %r183, [%rd78+76];
ld.global.nc.u32 %r159, [%rd76+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r159;
cvt.f32.f16 %f100, low;}


ld.global.nc.u32 %r191, [%rd78+112];
ld.global.nc.u32 %r160, [%rd76+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r160;
cvt.f32.f16 %f101, low;}


mov.b32 %r735, {%rs20, %rs19};
cvt.u32.u64 %r736, %rd5;
shr.s32 %r737, %r734, %r736;
add.s64 %rd79, %rd182, %rd27;
ld.global.nc.u8 %rs21, [%rd79];
cvt.u32.u16 %r738, %rs21;
and.b32 %r739, %r738, 255;
shr.u32 %r740, %r739, %r5;
and.b32 %r741, %r740, 15;
add.s64 %rd80, %rd182, %rd26;
ld.global.nc.u8 %rs22, [%rd80];
cvt.u32.u16 %r742, %rs22;
and.b32 %r743, %r742, 255;
shr.u32 %r744, %r743, %r6;
and.b32 %r745, %r744, 3;
bfi.b32 %r746, %r745, %r741, 4, 2;
add.s32 %r747, %r746, -32;
and.b32 %r667, %r735, 50529027;
shl.b32 %r748, %r737, 2;
and.b32 %r749, %r748, 67372036;
or.b32 %r750, %r667, -2139062144;
sub.s32 %r751, %r750, %r749;
xor.b32 %r752, %r751, -2139062144;
or.b32 %r753, %r667, %r749;
xor.b32 %r754, %r751, %r667;
and.b32 %r665, %r754, %r753;
mov.u32 %r161, %r665;

prmt.b32 %r161, %r161, 0, 0xba98;

mov.u32 %r163, %r667;

prmt.b32 %r163, %r163, 0, 0xba98;

xor.b32 %r755, %r163, 2139062143;
not.b32 %r756, %r161;
and.b32 %r757, %r752, %r756;
and.b32 %r758, %r755, %r161;
or.b32 %r166, %r758, %r757;
mov.u32 %r732, 0;

dp4a.s32.s32 %r165, %r166, %r167, %r732;

mul.lo.s32 %r759, %r747, %r165;
cvt.rn.f32.s32 %f177, %r759;
fma.rn.ftz.f32 %f178, %f98, %f177, 0f00000000;
add.s64 %rd81, %rd182, %rd25;
ld.global.nc.u8 %rs23, [%rd81];
cvt.u32.u16 %r760, %rs23;
and.b32 %r761, %r760, 255;
shr.u32 %r762, %r761, %r7;
and.b32 %r763, %r762, 15;
add.s64 %rd82, %rd182, %rd24;
ld.global.nc.u8 %rs24, [%rd82];
cvt.u32.u16 %r764, %rs24;
and.b32 %r765, %r764, 255;
shr.u32 %r766, %r765, %r8;
and.b32 %r767, %r766, 3;
bfi.b32 %r768, %r767, %r763, 4, 2;
add.s32 %r769, %r768, -32;
shr.u32 %r770, %r735, 2;
and.b32 %r675, %r770, 50529027;
shl.b32 %r771, %r737, 1;
and.b32 %r772, %r771, 67372036;
or.b32 %r773, %r675, -2139062144;
sub.s32 %r774, %r773, %r772;
xor.b32 %r775, %r774, -2139062144;
or.b32 %r776, %r675, %r772;
xor.b32 %r777, %r774, %r675;
and.b32 %r673, %r777, %r776;
mov.u32 %r169, %r673;

prmt.b32 %r169, %r169, 0, 0xba98;

mov.u32 %r171, %r675;

prmt.b32 %r171, %r171, 0, 0xba98;

xor.b32 %r778, %r171, 2139062143;
not.b32 %r779, %r169;
and.b32 %r780, %r775, %r779;
and.b32 %r781, %r778, %r169;
or.b32 %r174, %r781, %r780;

dp4a.s32.s32 %r173, %r174, %r175, %r732;

mul.lo.s32 %r782, %r769, %r173;
cvt.rn.f32.s32 %f179, %r782;
fma.rn.ftz.f32 %f180, %f99, %f179, %f178;
add.s64 %rd83, %rd182, %rd23;
ld.global.nc.u8 %rs25, [%rd83];
cvt.u32.u16 %r783, %rs25;
and.b32 %r784, %r783, 255;
shr.u32 %r785, %r784, %r9;
and.b32 %r786, %r785, 15;
add.s64 %rd84, %rd182, %rd22;
ld.global.nc.u8 %rs26, [%rd84];
cvt.u32.u16 %r787, %rs26;
and.b32 %r788, %r787, 255;
shr.u32 %r789, %r788, %r10;
and.b32 %r790, %r789, 3;
bfi.b32 %r791, %r790, %r786, 4, 2;
add.s32 %r792, %r791, -32;
shr.u32 %r793, %r735, 4;
and.b32 %r683, %r793, 50529027;
and.b32 %r794, %r737, 67372036;
or.b32 %r795, %r683, -2139062144;
sub.s32 %r796, %r795, %r794;
xor.b32 %r797, %r796, -2139062144;
or.b32 %r798, %r683, %r794;
xor.b32 %r799, %r796, %r683;
and.b32 %r681, %r799, %r798;
mov.u32 %r177, %r681;

prmt.b32 %r177, %r177, 0, 0xba98;

mov.u32 %r179, %r683;

prmt.b32 %r179, %r179, 0, 0xba98;

xor.b32 %r800, %r179, 2139062143;
not.b32 %r801, %r177;
and.b32 %r802, %r797, %r801;
and.b32 %r803, %r800, %r177;
or.b32 %r182, %r803, %r802;

dp4a.s32.s32 %r181, %r182, %r183, %r732;

mul.lo.s32 %r804, %r792, %r181;
cvt.rn.f32.s32 %f181, %r804;
fma.rn.ftz.f32 %f182, %f100, %f181, %f180;
add.s64 %rd85, %rd182, %rd21;
ld.global.nc.u8 %rs27, [%rd85];
cvt.u32.u16 %r805, %rs27;
and.b32 %r806, %r805, 255;
shr.u32 %r807, %r806, %r11;
and.b32 %r808, %r807, 15;
add.s64 %rd86, %rd182, %rd20;
ld.global.nc.u8 %rs28, [%rd86];
cvt.u32.u16 %r809, %rs28;
and.b32 %r810, %r809, 255;
shr.u32 %r811, %r810, %r12;
and.b32 %r812, %r811, 3;
bfi.b32 %r813, %r812, %r808, 4, 2;
add.s32 %r814, %r813, -32;
shr.u32 %r815, %r735, 6;
and.b32 %r691, %r815, 50529027;
shr.u32 %r816, %r737, 1;
and.b32 %r817, %r816, 67372036;
or.b32 %r818, %r691, -2139062144;
sub.s32 %r819, %r818, %r817;
xor.b32 %r820, %r819, -2139062144;
or.b32 %r821, %r691, %r817;
xor.b32 %r822, %r819, %r691;
and.b32 %r689, %r822, %r821;
mov.u32 %r185, %r689;

prmt.b32 %r185, %r185, 0, 0xba98;

mov.u32 %r187, %r691;

prmt.b32 %r187, %r187, 0, 0xba98;

xor.b32 %r823, %r187, 2139062143;
not.b32 %r824, %r185;
and.b32 %r825, %r820, %r824;
and.b32 %r826, %r823, %r185;
or.b32 %r190, %r826, %r825;

dp4a.s32.s32 %r189, %r190, %r191, %r732;

mul.lo.s32 %r827, %r814, %r189;
cvt.rn.f32.s32 %f183, %r827;
fma.rn.ftz.f32 %f184, %f101, %f183, %f182;
fma.rn.ftz.f32 %f512, %f97, %f184, %f512;
add.s64 %rd87, %rd182, %rd6;
ld.global.nc.u16 %rs2, [%rd87+108];

{ cvt.f32.f16 %f102, %rs2;}


add.s64 %rd88, %rd182, %rd19;
add.s64 %rd89, %rd182, %rd17;
ld.global.nc.u16 %rs29, [%rd89];
ld.global.nc.u16 %rs30, [%rd89+2];
mov.b32 %r828, {%rs29, %rs30};
not.b32 %r829, %r828;
ld.global.nc.u16 %rs31, [%rd88+2];
ld.global.nc.u16 %rs32, [%rd88];

{.reg .f16 low,high;
mov.b32 {low,high},%r157;
cvt.f32.f16 %f103, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r158;
cvt.f32.f16 %f104, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r159;
cvt.f32.f16 %f105, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r160;
cvt.f32.f16 %f106, low;}


mov.b32 %r830, {%rs32, %rs31};
shr.s32 %r831, %r829, %r736;
add.s64 %rd90, %rd182, %rd14;
ld.global.nc.u8 %rs33, [%rd90];
cvt.u32.u16 %r832, %rs33;
and.b32 %r833, %r832, 255;
shr.u32 %r834, %r833, %r5;
and.b32 %r835, %r834, 15;
add.s64 %rd91, %rd182, %rd13;
ld.global.nc.u8 %rs34, [%rd91];
cvt.u32.u16 %r836, %rs34;
and.b32 %r837, %r836, 255;
shr.u32 %r838, %r837, %r6;
and.b32 %r839, %r838, 3;
bfi.b32 %r840, %r839, %r835, 4, 2;
add.s32 %r841, %r840, -32;
and.b32 %r703, %r830, 50529027;
shl.b32 %r842, %r831, 2;
and.b32 %r843, %r842, 67372036;
or.b32 %r844, %r703, -2139062144;
sub.s32 %r845, %r844, %r843;
xor.b32 %r846, %r845, -2139062144;
or.b32 %r847, %r703, %r843;
xor.b32 %r848, %r845, %r703;
and.b32 %r701, %r848, %r847;
mov.u32 %r197, %r701;

prmt.b32 %r197, %r197, 0, 0xba98;

mov.u32 %r199, %r703;

prmt.b32 %r199, %r199, 0, 0xba98;

xor.b32 %r849, %r199, 2139062143;
not.b32 %r850, %r197;
and.b32 %r851, %r846, %r850;
and.b32 %r852, %r849, %r197;
or.b32 %r202, %r852, %r851;

dp4a.s32.s32 %r201, %r202, %r167, %r732;

mul.lo.s32 %r853, %r841, %r201;
cvt.rn.f32.s32 %f185, %r853;
fma.rn.ftz.f32 %f186, %f103, %f185, 0f00000000;
add.s64 %rd92, %rd182, %rd12;
ld.global.nc.u8 %rs35, [%rd92];
cvt.u32.u16 %r854, %rs35;
and.b32 %r855, %r854, 255;
shr.u32 %r856, %r855, %r7;
and.b32 %r857, %r856, 15;
add.s64 %rd93, %rd182, %rd11;
ld.global.nc.u8 %rs36, [%rd93];
cvt.u32.u16 %r858, %rs36;
and.b32 %r859, %r858, 255;
shr.u32 %r860, %r859, %r8;
and.b32 %r861, %r860, 3;
bfi.b32 %r862, %r861, %r857, 4, 2;
add.s32 %r863, %r862, -32;
shr.u32 %r864, %r830, 2;
and.b32 %r711, %r864, 50529027;
shl.b32 %r865, %r831, 1;
and.b32 %r866, %r865, 67372036;
or.b32 %r867, %r711, -2139062144;
sub.s32 %r868, %r867, %r866;
xor.b32 %r869, %r868, -2139062144;
or.b32 %r870, %r711, %r866;
xor.b32 %r871, %r868, %r711;
and.b32 %r709, %r871, %r870;
mov.u32 %r205, %r709;

prmt.b32 %r205, %r205, 0, 0xba98;

mov.u32 %r207, %r711;

prmt.b32 %r207, %r207, 0, 0xba98;

xor.b32 %r872, %r207, 2139062143;
not.b32 %r873, %r205;
and.b32 %r874, %r869, %r873;
and.b32 %r875, %r872, %r205;
or.b32 %r210, %r875, %r874;

dp4a.s32.s32 %r209, %r210, %r175, %r732;

mul.lo.s32 %r876, %r863, %r209;
cvt.rn.f32.s32 %f187, %r876;
fma.rn.ftz.f32 %f188, %f104, %f187, %f186;
add.s64 %rd94, %rd182, %rd10;
ld.global.nc.u8 %rs37, [%rd94];
cvt.u32.u16 %r877, %rs37;
and.b32 %r878, %r877, 255;
shr.u32 %r879, %r878, %r9;
and.b32 %r880, %r879, 15;
add.s64 %rd95, %rd182, %rd9;
ld.global.nc.u8 %rs38, [%rd95];
cvt.u32.u16 %r881, %rs38;
and.b32 %r882, %r881, 255;
shr.u32 %r883, %r882, %r10;
and.b32 %r884, %r883, 3;
bfi.b32 %r885, %r884, %r880, 4, 2;
add.s32 %r886, %r885, -32;
shr.u32 %r887, %r830, 4;
and.b32 %r719, %r887, 50529027;
and.b32 %r888, %r831, 67372036;
or.b32 %r889, %r719, -2139062144;
sub.s32 %r890, %r889, %r888;
xor.b32 %r891, %r890, -2139062144;
or.b32 %r892, %r719, %r888;
xor.b32 %r893, %r890, %r719;
and.b32 %r717, %r893, %r892;
mov.u32 %r213, %r717;

prmt.b32 %r213, %r213, 0, 0xba98;

mov.u32 %r215, %r719;

prmt.b32 %r215, %r215, 0, 0xba98;

xor.b32 %r894, %r215, 2139062143;
not.b32 %r895, %r213;
and.b32 %r896, %r891, %r895;
and.b32 %r897, %r894, %r213;
or.b32 %r218, %r897, %r896;

dp4a.s32.s32 %r217, %r218, %r183, %r732;

mul.lo.s32 %r898, %r886, %r217;
cvt.rn.f32.s32 %f189, %r898;
fma.rn.ftz.f32 %f190, %f105, %f189, %f188;
add.s64 %rd96, %rd182, %rd8;
ld.global.nc.u8 %rs39, [%rd96];
cvt.u32.u16 %r899, %rs39;
and.b32 %r900, %r899, 255;
shr.u32 %r901, %r900, %r11;
and.b32 %r902, %r901, 15;
add.s64 %rd97, %rd182, %rd7;
ld.global.nc.u8 %rs40, [%rd97];
cvt.u32.u16 %r903, %rs40;
and.b32 %r904, %r903, 255;
shr.u32 %r905, %r904, %r12;
and.b32 %r906, %r905, 3;
bfi.b32 %r907, %r906, %r902, 4, 2;
add.s32 %r908, %r907, -32;
shr.u32 %r909, %r830, 6;
and.b32 %r727, %r909, 50529027;
shr.u32 %r910, %r831, 1;
and.b32 %r911, %r910, 67372036;
or.b32 %r912, %r727, -2139062144;
sub.s32 %r913, %r912, %r911;
xor.b32 %r914, %r913, -2139062144;
or.b32 %r915, %r727, %r911;
xor.b32 %r916, %r913, %r727;
and.b32 %r725, %r916, %r915;
mov.u32 %r221, %r725;

prmt.b32 %r221, %r221, 0, 0xba98;

mov.u32 %r223, %r727;

prmt.b32 %r223, %r223, 0, 0xba98;

xor.b32 %r917, %r223, 2139062143;
not.b32 %r918, %r221;
and.b32 %r919, %r914, %r918;
and.b32 %r920, %r917, %r221;
or.b32 %r226, %r920, %r919;

dp4a.s32.s32 %r225, %r226, %r191, %r732;

mul.lo.s32 %r921, %r908, %r225;
cvt.rn.f32.s32 %f191, %r921;
fma.rn.ftz.f32 %f192, %f106, %f191, %f190;
fma.rn.ftz.f32 %f511, %f102, %f192, %f511;
cvt.u64.u32 %rd98, %r1519;
add.s64 %rd99, %rd2, %rd98;

{ cvt.f32.f16 %f107, %rs1;}


add.s64 %rd100, %rd99, %rd5;
mul.lo.s64 %rd101, %rd100, 36;
add.s64 %rd102, %rd28, %rd101;
add.s64 %rd103, %rd102, %rd77;
ld.global.nc.u32 %r239, [%rd103+4];
ld.global.nc.u32 %r229, [%rd102];

{.reg .f16 low,high;
mov.b32 {low,high},%r229;
cvt.f32.f16 %f108, low;}


ld.global.nc.u32 %r247, [%rd103+40];
ld.global.nc.u32 %r230, [%rd102+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r230;
cvt.f32.f16 %f109, low;}


ld.global.nc.u32 %r255, [%rd103+76];
ld.global.nc.u32 %r231, [%rd102+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r231;
cvt.f32.f16 %f110, low;}


ld.global.nc.u32 %r263, [%rd103+112];
ld.global.nc.u32 %r232, [%rd102+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r232;
cvt.f32.f16 %f111, low;}


mov.u32 %r233, %r665;

prmt.b32 %r233, %r233, 0, 0xba98;

mov.u32 %r235, %r667;

prmt.b32 %r235, %r235, 0, 0xba98;

xor.b32 %r922, %r235, 2139062143;
not.b32 %r923, %r233;
and.b32 %r924, %r752, %r923;
and.b32 %r925, %r922, %r233;
or.b32 %r238, %r925, %r924;

dp4a.s32.s32 %r237, %r238, %r239, %r732;

mul.lo.s32 %r926, %r747, %r237;
cvt.rn.f32.s32 %f193, %r926;
fma.rn.ftz.f32 %f194, %f108, %f193, 0f00000000;
mov.u32 %r241, %r673;

prmt.b32 %r241, %r241, 0, 0xba98;

mov.u32 %r243, %r675;

prmt.b32 %r243, %r243, 0, 0xba98;

xor.b32 %r927, %r243, 2139062143;
not.b32 %r928, %r241;
and.b32 %r929, %r775, %r928;
and.b32 %r930, %r927, %r241;
or.b32 %r246, %r930, %r929;

dp4a.s32.s32 %r245, %r246, %r247, %r732;

mul.lo.s32 %r931, %r769, %r245;
cvt.rn.f32.s32 %f195, %r931;
fma.rn.ftz.f32 %f196, %f109, %f195, %f194;
mov.u32 %r249, %r681;

prmt.b32 %r249, %r249, 0, 0xba98;

mov.u32 %r251, %r683;

prmt.b32 %r251, %r251, 0, 0xba98;

xor.b32 %r932, %r251, 2139062143;
not.b32 %r933, %r249;
and.b32 %r934, %r797, %r933;
and.b32 %r935, %r932, %r249;
or.b32 %r254, %r935, %r934;

dp4a.s32.s32 %r253, %r254, %r255, %r732;

mul.lo.s32 %r936, %r792, %r253;
cvt.rn.f32.s32 %f197, %r936;
fma.rn.ftz.f32 %f198, %f110, %f197, %f196;
mov.u32 %r257, %r689;

prmt.b32 %r257, %r257, 0, 0xba98;

mov.u32 %r259, %r691;

prmt.b32 %r259, %r259, 0, 0xba98;

xor.b32 %r937, %r259, 2139062143;
not.b32 %r938, %r257;
and.b32 %r939, %r820, %r938;
and.b32 %r940, %r937, %r257;
or.b32 %r262, %r940, %r939;

dp4a.s32.s32 %r261, %r262, %r263, %r732;

mul.lo.s32 %r941, %r814, %r261;
cvt.rn.f32.s32 %f199, %r941;
fma.rn.ftz.f32 %f200, %f111, %f199, %f198;
fma.rn.ftz.f32 %f510, %f107, %f200, %f510;

{ cvt.f32.f16 %f112, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r229;
cvt.f32.f16 %f113, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r230;
cvt.f32.f16 %f114, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r231;
cvt.f32.f16 %f115, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r232;
cvt.f32.f16 %f116, low;}


mov.u32 %r269, %r701;

prmt.b32 %r269, %r269, 0, 0xba98;

mov.u32 %r271, %r703;

prmt.b32 %r271, %r271, 0, 0xba98;

xor.b32 %r942, %r271, 2139062143;
not.b32 %r943, %r269;
and.b32 %r944, %r846, %r943;
and.b32 %r945, %r942, %r269;
or.b32 %r274, %r945, %r944;

dp4a.s32.s32 %r273, %r274, %r239, %r732;

mul.lo.s32 %r946, %r841, %r273;
cvt.rn.f32.s32 %f201, %r946;
fma.rn.ftz.f32 %f202, %f113, %f201, 0f00000000;
mov.u32 %r277, %r709;

prmt.b32 %r277, %r277, 0, 0xba98;

mov.u32 %r279, %r711;

prmt.b32 %r279, %r279, 0, 0xba98;

xor.b32 %r947, %r279, 2139062143;
not.b32 %r948, %r277;
and.b32 %r949, %r869, %r948;
and.b32 %r950, %r947, %r277;
or.b32 %r282, %r950, %r949;

dp4a.s32.s32 %r281, %r282, %r247, %r732;

mul.lo.s32 %r951, %r863, %r281;
cvt.rn.f32.s32 %f203, %r951;
fma.rn.ftz.f32 %f204, %f114, %f203, %f202;
mov.u32 %r285, %r717;

prmt.b32 %r285, %r285, 0, 0xba98;

mov.u32 %r287, %r719;

prmt.b32 %r287, %r287, 0, 0xba98;

xor.b32 %r952, %r287, 2139062143;
not.b32 %r953, %r285;
and.b32 %r954, %r891, %r953;
and.b32 %r955, %r952, %r285;
or.b32 %r290, %r955, %r954;

dp4a.s32.s32 %r289, %r290, %r255, %r732;

mul.lo.s32 %r956, %r886, %r289;
cvt.rn.f32.s32 %f205, %r956;
fma.rn.ftz.f32 %f206, %f115, %f205, %f204;
mov.u32 %r293, %r725;

prmt.b32 %r293, %r293, 0, 0xba98;

mov.u32 %r295, %r727;

prmt.b32 %r295, %r295, 0, 0xba98;

xor.b32 %r957, %r295, 2139062143;
not.b32 %r958, %r293;
and.b32 %r959, %r914, %r958;
and.b32 %r960, %r957, %r293;
or.b32 %r298, %r960, %r959;

dp4a.s32.s32 %r297, %r298, %r263, %r732;

mul.lo.s32 %r961, %r908, %r297;
cvt.rn.f32.s32 %f207, %r961;
fma.rn.ftz.f32 %f208, %f116, %f207, %f206;
fma.rn.ftz.f32 %f509, %f112, %f208, %f509;
cvt.u64.u32 %rd104, %r1525;
add.s64 %rd105, %rd2, %rd104;

{ cvt.f32.f16 %f117, %rs1;}


add.s64 %rd106, %rd105, %rd5;
mul.lo.s64 %rd107, %rd106, 36;
add.s64 %rd108, %rd28, %rd107;
add.s64 %rd109, %rd108, %rd77;
ld.global.nc.u32 %r311, [%rd109+4];
ld.global.nc.u32 %r301, [%rd108];

{.reg .f16 low,high;
mov.b32 {low,high},%r301;
cvt.f32.f16 %f118, low;}


ld.global.nc.u32 %r319, [%rd109+40];
ld.global.nc.u32 %r302, [%rd108+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r302;
cvt.f32.f16 %f119, low;}


ld.global.nc.u32 %r327, [%rd109+76];
ld.global.nc.u32 %r303, [%rd108+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r303;
cvt.f32.f16 %f120, low;}


ld.global.nc.u32 %r335, [%rd109+112];
ld.global.nc.u32 %r304, [%rd108+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r304;
cvt.f32.f16 %f121, low;}


mov.u32 %r305, %r665;

prmt.b32 %r305, %r305, 0, 0xba98;

mov.u32 %r307, %r667;

prmt.b32 %r307, %r307, 0, 0xba98;

xor.b32 %r962, %r307, 2139062143;
not.b32 %r963, %r305;
and.b32 %r964, %r752, %r963;
and.b32 %r965, %r962, %r305;
or.b32 %r310, %r965, %r964;

dp4a.s32.s32 %r309, %r310, %r311, %r732;

mul.lo.s32 %r966, %r747, %r309;
cvt.rn.f32.s32 %f209, %r966;
fma.rn.ftz.f32 %f210, %f118, %f209, 0f00000000;
mov.u32 %r313, %r673;

prmt.b32 %r313, %r313, 0, 0xba98;

mov.u32 %r315, %r675;

prmt.b32 %r315, %r315, 0, 0xba98;

xor.b32 %r967, %r315, 2139062143;
not.b32 %r968, %r313;
and.b32 %r969, %r775, %r968;
and.b32 %r970, %r967, %r313;
or.b32 %r318, %r970, %r969;

dp4a.s32.s32 %r317, %r318, %r319, %r732;

mul.lo.s32 %r971, %r769, %r317;
cvt.rn.f32.s32 %f211, %r971;
fma.rn.ftz.f32 %f212, %f119, %f211, %f210;
mov.u32 %r321, %r681;

prmt.b32 %r321, %r321, 0, 0xba98;

mov.u32 %r323, %r683;

prmt.b32 %r323, %r323, 0, 0xba98;

xor.b32 %r972, %r323, 2139062143;
not.b32 %r973, %r321;
and.b32 %r974, %r797, %r973;
and.b32 %r975, %r972, %r321;
or.b32 %r326, %r975, %r974;

dp4a.s32.s32 %r325, %r326, %r327, %r732;

mul.lo.s32 %r976, %r792, %r325;
cvt.rn.f32.s32 %f213, %r976;
fma.rn.ftz.f32 %f214, %f120, %f213, %f212;
mov.u32 %r329, %r689;

prmt.b32 %r329, %r329, 0, 0xba98;

mov.u32 %r331, %r691;

prmt.b32 %r331, %r331, 0, 0xba98;

xor.b32 %r977, %r331, 2139062143;
not.b32 %r978, %r329;
and.b32 %r979, %r820, %r978;
and.b32 %r980, %r977, %r329;
or.b32 %r334, %r980, %r979;

dp4a.s32.s32 %r333, %r334, %r335, %r732;

mul.lo.s32 %r981, %r814, %r333;
cvt.rn.f32.s32 %f215, %r981;
fma.rn.ftz.f32 %f216, %f121, %f215, %f214;
fma.rn.ftz.f32 %f508, %f117, %f216, %f508;

{ cvt.f32.f16 %f122, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r301;
cvt.f32.f16 %f123, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r302;
cvt.f32.f16 %f124, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r303;
cvt.f32.f16 %f125, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r304;
cvt.f32.f16 %f126, low;}


mov.u32 %r341, %r701;

prmt.b32 %r341, %r341, 0, 0xba98;

mov.u32 %r343, %r703;

prmt.b32 %r343, %r343, 0, 0xba98;

xor.b32 %r982, %r343, 2139062143;
not.b32 %r983, %r341;
and.b32 %r984, %r846, %r983;
and.b32 %r985, %r982, %r341;
or.b32 %r346, %r985, %r984;

dp4a.s32.s32 %r345, %r346, %r311, %r732;

mul.lo.s32 %r986, %r841, %r345;
cvt.rn.f32.s32 %f217, %r986;
fma.rn.ftz.f32 %f218, %f123, %f217, 0f00000000;
mov.u32 %r349, %r709;

prmt.b32 %r349, %r349, 0, 0xba98;

mov.u32 %r351, %r711;

prmt.b32 %r351, %r351, 0, 0xba98;

xor.b32 %r987, %r351, 2139062143;
not.b32 %r988, %r349;
and.b32 %r989, %r869, %r988;
and.b32 %r990, %r987, %r349;
or.b32 %r354, %r990, %r989;

dp4a.s32.s32 %r353, %r354, %r319, %r732;

mul.lo.s32 %r991, %r863, %r353;
cvt.rn.f32.s32 %f219, %r991;
fma.rn.ftz.f32 %f220, %f124, %f219, %f218;
mov.u32 %r357, %r717;

prmt.b32 %r357, %r357, 0, 0xba98;

mov.u32 %r359, %r719;

prmt.b32 %r359, %r359, 0, 0xba98;

xor.b32 %r992, %r359, 2139062143;
not.b32 %r993, %r357;
and.b32 %r994, %r891, %r993;
and.b32 %r995, %r992, %r357;
or.b32 %r362, %r995, %r994;

dp4a.s32.s32 %r361, %r362, %r327, %r732;

mul.lo.s32 %r996, %r886, %r361;
cvt.rn.f32.s32 %f221, %r996;
fma.rn.ftz.f32 %f222, %f125, %f221, %f220;
mov.u32 %r365, %r725;

prmt.b32 %r365, %r365, 0, 0xba98;

mov.u32 %r367, %r727;

prmt.b32 %r367, %r367, 0, 0xba98;

xor.b32 %r997, %r367, 2139062143;
not.b32 %r998, %r365;
and.b32 %r999, %r914, %r998;
and.b32 %r1000, %r997, %r365;
or.b32 %r370, %r1000, %r999;

dp4a.s32.s32 %r369, %r370, %r335, %r732;

mul.lo.s32 %r1001, %r908, %r369;
cvt.rn.f32.s32 %f223, %r1001;
fma.rn.ftz.f32 %f224, %f126, %f223, %f222;
fma.rn.ftz.f32 %f507, %f122, %f224, %f507;
cvt.u64.u32 %rd110, %r1524;
add.s64 %rd111, %rd2, %rd110;

{ cvt.f32.f16 %f127, %rs1;}


add.s64 %rd112, %rd111, %rd5;
mul.lo.s64 %rd113, %rd112, 36;
add.s64 %rd114, %rd28, %rd113;
add.s64 %rd115, %rd114, %rd77;
ld.global.nc.u32 %r383, [%rd115+4];
ld.global.nc.u32 %r373, [%rd114];

{.reg .f16 low,high;
mov.b32 {low,high},%r373;
cvt.f32.f16 %f128, low;}


ld.global.nc.u32 %r391, [%rd115+40];
ld.global.nc.u32 %r374, [%rd114+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r374;
cvt.f32.f16 %f129, low;}


ld.global.nc.u32 %r399, [%rd115+76];
ld.global.nc.u32 %r375, [%rd114+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r375;
cvt.f32.f16 %f130, low;}


ld.global.nc.u32 %r407, [%rd115+112];
ld.global.nc.u32 %r376, [%rd114+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r376;
cvt.f32.f16 %f131, low;}


mov.u32 %r377, %r665;

prmt.b32 %r377, %r377, 0, 0xba98;

mov.u32 %r379, %r667;

prmt.b32 %r379, %r379, 0, 0xba98;

xor.b32 %r1002, %r379, 2139062143;
not.b32 %r1003, %r377;
and.b32 %r1004, %r752, %r1003;
and.b32 %r1005, %r1002, %r377;
or.b32 %r382, %r1005, %r1004;

dp4a.s32.s32 %r381, %r382, %r383, %r732;

mul.lo.s32 %r1006, %r747, %r381;
cvt.rn.f32.s32 %f225, %r1006;
fma.rn.ftz.f32 %f226, %f128, %f225, 0f00000000;
mov.u32 %r385, %r673;

prmt.b32 %r385, %r385, 0, 0xba98;

mov.u32 %r387, %r675;

prmt.b32 %r387, %r387, 0, 0xba98;

xor.b32 %r1007, %r387, 2139062143;
not.b32 %r1008, %r385;
and.b32 %r1009, %r775, %r1008;
and.b32 %r1010, %r1007, %r385;
or.b32 %r390, %r1010, %r1009;

dp4a.s32.s32 %r389, %r390, %r391, %r732;

mul.lo.s32 %r1011, %r769, %r389;
cvt.rn.f32.s32 %f227, %r1011;
fma.rn.ftz.f32 %f228, %f129, %f227, %f226;
mov.u32 %r393, %r681;

prmt.b32 %r393, %r393, 0, 0xba98;

mov.u32 %r395, %r683;

prmt.b32 %r395, %r395, 0, 0xba98;

xor.b32 %r1012, %r395, 2139062143;
not.b32 %r1013, %r393;
and.b32 %r1014, %r797, %r1013;
and.b32 %r1015, %r1012, %r393;
or.b32 %r398, %r1015, %r1014;

dp4a.s32.s32 %r397, %r398, %r399, %r732;

mul.lo.s32 %r1016, %r792, %r397;
cvt.rn.f32.s32 %f229, %r1016;
fma.rn.ftz.f32 %f230, %f130, %f229, %f228;
mov.u32 %r401, %r689;

prmt.b32 %r401, %r401, 0, 0xba98;

mov.u32 %r403, %r691;

prmt.b32 %r403, %r403, 0, 0xba98;

xor.b32 %r1017, %r403, 2139062143;
not.b32 %r1018, %r401;
and.b32 %r1019, %r820, %r1018;
and.b32 %r1020, %r1017, %r401;
or.b32 %r406, %r1020, %r1019;

dp4a.s32.s32 %r405, %r406, %r407, %r732;

mul.lo.s32 %r1021, %r814, %r405;
cvt.rn.f32.s32 %f231, %r1021;
fma.rn.ftz.f32 %f232, %f131, %f231, %f230;
fma.rn.ftz.f32 %f506, %f127, %f232, %f506;

{ cvt.f32.f16 %f132, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r373;
cvt.f32.f16 %f133, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r374;
cvt.f32.f16 %f134, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r375;
cvt.f32.f16 %f135, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r376;
cvt.f32.f16 %f136, low;}


mov.u32 %r413, %r701;

prmt.b32 %r413, %r413, 0, 0xba98;

mov.u32 %r415, %r703;

prmt.b32 %r415, %r415, 0, 0xba98;

xor.b32 %r1022, %r415, 2139062143;
not.b32 %r1023, %r413;
and.b32 %r1024, %r846, %r1023;
and.b32 %r1025, %r1022, %r413;
or.b32 %r418, %r1025, %r1024;

dp4a.s32.s32 %r417, %r418, %r383, %r732;

mul.lo.s32 %r1026, %r841, %r417;
cvt.rn.f32.s32 %f233, %r1026;
fma.rn.ftz.f32 %f234, %f133, %f233, 0f00000000;
mov.u32 %r421, %r709;

prmt.b32 %r421, %r421, 0, 0xba98;

mov.u32 %r423, %r711;

prmt.b32 %r423, %r423, 0, 0xba98;

xor.b32 %r1027, %r423, 2139062143;
not.b32 %r1028, %r421;
and.b32 %r1029, %r869, %r1028;
and.b32 %r1030, %r1027, %r421;
or.b32 %r426, %r1030, %r1029;

dp4a.s32.s32 %r425, %r426, %r391, %r732;

mul.lo.s32 %r1031, %r863, %r425;
cvt.rn.f32.s32 %f235, %r1031;
fma.rn.ftz.f32 %f236, %f134, %f235, %f234;
mov.u32 %r429, %r717;

prmt.b32 %r429, %r429, 0, 0xba98;

mov.u32 %r431, %r719;

prmt.b32 %r431, %r431, 0, 0xba98;

xor.b32 %r1032, %r431, 2139062143;
not.b32 %r1033, %r429;
and.b32 %r1034, %r891, %r1033;
and.b32 %r1035, %r1032, %r429;
or.b32 %r434, %r1035, %r1034;

dp4a.s32.s32 %r433, %r434, %r399, %r732;

mul.lo.s32 %r1036, %r886, %r433;
cvt.rn.f32.s32 %f237, %r1036;
fma.rn.ftz.f32 %f238, %f135, %f237, %f236;
mov.u32 %r437, %r725;

prmt.b32 %r437, %r437, 0, 0xba98;

mov.u32 %r439, %r727;

prmt.b32 %r439, %r439, 0, 0xba98;

xor.b32 %r1037, %r439, 2139062143;
not.b32 %r1038, %r437;
and.b32 %r1039, %r914, %r1038;
and.b32 %r1040, %r1037, %r437;
or.b32 %r442, %r1040, %r1039;

dp4a.s32.s32 %r441, %r442, %r407, %r732;

mul.lo.s32 %r1041, %r908, %r441;
cvt.rn.f32.s32 %f239, %r1041;
fma.rn.ftz.f32 %f240, %f136, %f239, %f238;
fma.rn.ftz.f32 %f505, %f132, %f240, %f505;
cvt.u64.u32 %rd116, %r1523;
add.s64 %rd117, %rd2, %rd116;

{ cvt.f32.f16 %f137, %rs1;}


add.s64 %rd118, %rd117, %rd5;
mul.lo.s64 %rd119, %rd118, 36;
add.s64 %rd120, %rd28, %rd119;
add.s64 %rd121, %rd120, %rd77;
ld.global.nc.u32 %r455, [%rd121+4];
ld.global.nc.u32 %r445, [%rd120];

{.reg .f16 low,high;
mov.b32 {low,high},%r445;
cvt.f32.f16 %f138, low;}


ld.global.nc.u32 %r463, [%rd121+40];
ld.global.nc.u32 %r446, [%rd120+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r446;
cvt.f32.f16 %f139, low;}


ld.global.nc.u32 %r471, [%rd121+76];
ld.global.nc.u32 %r447, [%rd120+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r447;
cvt.f32.f16 %f140, low;}


ld.global.nc.u32 %r479, [%rd121+112];
ld.global.nc.u32 %r448, [%rd120+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r448;
cvt.f32.f16 %f141, low;}


mov.u32 %r449, %r665;

prmt.b32 %r449, %r449, 0, 0xba98;

mov.u32 %r451, %r667;

prmt.b32 %r451, %r451, 0, 0xba98;

xor.b32 %r1042, %r451, 2139062143;
not.b32 %r1043, %r449;
and.b32 %r1044, %r752, %r1043;
and.b32 %r1045, %r1042, %r449;
or.b32 %r454, %r1045, %r1044;

dp4a.s32.s32 %r453, %r454, %r455, %r732;

mul.lo.s32 %r1046, %r747, %r453;
cvt.rn.f32.s32 %f241, %r1046;
fma.rn.ftz.f32 %f242, %f138, %f241, 0f00000000;
mov.u32 %r457, %r673;

prmt.b32 %r457, %r457, 0, 0xba98;

mov.u32 %r459, %r675;

prmt.b32 %r459, %r459, 0, 0xba98;

xor.b32 %r1047, %r459, 2139062143;
not.b32 %r1048, %r457;
and.b32 %r1049, %r775, %r1048;
and.b32 %r1050, %r1047, %r457;
or.b32 %r462, %r1050, %r1049;

dp4a.s32.s32 %r461, %r462, %r463, %r732;

mul.lo.s32 %r1051, %r769, %r461;
cvt.rn.f32.s32 %f243, %r1051;
fma.rn.ftz.f32 %f244, %f139, %f243, %f242;
mov.u32 %r465, %r681;

prmt.b32 %r465, %r465, 0, 0xba98;

mov.u32 %r467, %r683;

prmt.b32 %r467, %r467, 0, 0xba98;

xor.b32 %r1052, %r467, 2139062143;
not.b32 %r1053, %r465;
and.b32 %r1054, %r797, %r1053;
and.b32 %r1055, %r1052, %r465;
or.b32 %r470, %r1055, %r1054;

dp4a.s32.s32 %r469, %r470, %r471, %r732;

mul.lo.s32 %r1056, %r792, %r469;
cvt.rn.f32.s32 %f245, %r1056;
fma.rn.ftz.f32 %f246, %f140, %f245, %f244;
mov.u32 %r473, %r689;

prmt.b32 %r473, %r473, 0, 0xba98;

mov.u32 %r475, %r691;

prmt.b32 %r475, %r475, 0, 0xba98;

xor.b32 %r1057, %r475, 2139062143;
not.b32 %r1058, %r473;
and.b32 %r1059, %r820, %r1058;
and.b32 %r1060, %r1057, %r473;
or.b32 %r478, %r1060, %r1059;

dp4a.s32.s32 %r477, %r478, %r479, %r732;

mul.lo.s32 %r1061, %r814, %r477;
cvt.rn.f32.s32 %f247, %r1061;
fma.rn.ftz.f32 %f248, %f141, %f247, %f246;
fma.rn.ftz.f32 %f504, %f137, %f248, %f504;

{ cvt.f32.f16 %f142, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r445;
cvt.f32.f16 %f143, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r446;
cvt.f32.f16 %f144, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r447;
cvt.f32.f16 %f145, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r448;
cvt.f32.f16 %f146, low;}


mov.u32 %r485, %r701;

prmt.b32 %r485, %r485, 0, 0xba98;

mov.u32 %r487, %r703;

prmt.b32 %r487, %r487, 0, 0xba98;

xor.b32 %r1062, %r487, 2139062143;
not.b32 %r1063, %r485;
and.b32 %r1064, %r846, %r1063;
and.b32 %r1065, %r1062, %r485;
or.b32 %r490, %r1065, %r1064;

dp4a.s32.s32 %r489, %r490, %r455, %r732;

mul.lo.s32 %r1066, %r841, %r489;
cvt.rn.f32.s32 %f249, %r1066;
fma.rn.ftz.f32 %f250, %f143, %f249, 0f00000000;
mov.u32 %r493, %r709;

prmt.b32 %r493, %r493, 0, 0xba98;

mov.u32 %r495, %r711;

prmt.b32 %r495, %r495, 0, 0xba98;

xor.b32 %r1067, %r495, 2139062143;
not.b32 %r1068, %r493;
and.b32 %r1069, %r869, %r1068;
and.b32 %r1070, %r1067, %r493;
or.b32 %r498, %r1070, %r1069;

dp4a.s32.s32 %r497, %r498, %r463, %r732;

mul.lo.s32 %r1071, %r863, %r497;
cvt.rn.f32.s32 %f251, %r1071;
fma.rn.ftz.f32 %f252, %f144, %f251, %f250;
mov.u32 %r501, %r717;

prmt.b32 %r501, %r501, 0, 0xba98;

mov.u32 %r503, %r719;

prmt.b32 %r503, %r503, 0, 0xba98;

xor.b32 %r1072, %r503, 2139062143;
not.b32 %r1073, %r501;
and.b32 %r1074, %r891, %r1073;
and.b32 %r1075, %r1072, %r501;
or.b32 %r506, %r1075, %r1074;

dp4a.s32.s32 %r505, %r506, %r471, %r732;

mul.lo.s32 %r1076, %r886, %r505;
cvt.rn.f32.s32 %f253, %r1076;
fma.rn.ftz.f32 %f254, %f145, %f253, %f252;
mov.u32 %r509, %r725;

prmt.b32 %r509, %r509, 0, 0xba98;

mov.u32 %r511, %r727;

prmt.b32 %r511, %r511, 0, 0xba98;

xor.b32 %r1077, %r511, 2139062143;
not.b32 %r1078, %r509;
and.b32 %r1079, %r914, %r1078;
and.b32 %r1080, %r1077, %r509;
or.b32 %r514, %r1080, %r1079;

dp4a.s32.s32 %r513, %r514, %r479, %r732;

mul.lo.s32 %r1081, %r908, %r513;
cvt.rn.f32.s32 %f255, %r1081;
fma.rn.ftz.f32 %f256, %f146, %f255, %f254;
fma.rn.ftz.f32 %f503, %f142, %f256, %f503;
cvt.u64.u32 %rd122, %r1522;
add.s64 %rd123, %rd2, %rd122;

{ cvt.f32.f16 %f147, %rs1;}


add.s64 %rd124, %rd123, %rd5;
mul.lo.s64 %rd125, %rd124, 36;
add.s64 %rd126, %rd28, %rd125;
add.s64 %rd127, %rd126, %rd77;
ld.global.nc.u32 %r527, [%rd127+4];
ld.global.nc.u32 %r517, [%rd126];

{.reg .f16 low,high;
mov.b32 {low,high},%r517;
cvt.f32.f16 %f148, low;}


ld.global.nc.u32 %r535, [%rd127+40];
ld.global.nc.u32 %r518, [%rd126+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r518;
cvt.f32.f16 %f149, low;}


ld.global.nc.u32 %r543, [%rd127+76];
ld.global.nc.u32 %r519, [%rd126+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r519;
cvt.f32.f16 %f150, low;}


ld.global.nc.u32 %r551, [%rd127+112];
ld.global.nc.u32 %r520, [%rd126+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r520;
cvt.f32.f16 %f151, low;}


mov.u32 %r521, %r665;

prmt.b32 %r521, %r521, 0, 0xba98;

mov.u32 %r523, %r667;

prmt.b32 %r523, %r523, 0, 0xba98;

xor.b32 %r1082, %r523, 2139062143;
not.b32 %r1083, %r521;
and.b32 %r1084, %r752, %r1083;
and.b32 %r1085, %r1082, %r521;
or.b32 %r526, %r1085, %r1084;

dp4a.s32.s32 %r525, %r526, %r527, %r732;

mul.lo.s32 %r1086, %r747, %r525;
cvt.rn.f32.s32 %f257, %r1086;
fma.rn.ftz.f32 %f258, %f148, %f257, 0f00000000;
mov.u32 %r529, %r673;

prmt.b32 %r529, %r529, 0, 0xba98;

mov.u32 %r531, %r675;

prmt.b32 %r531, %r531, 0, 0xba98;

xor.b32 %r1087, %r531, 2139062143;
not.b32 %r1088, %r529;
and.b32 %r1089, %r775, %r1088;
and.b32 %r1090, %r1087, %r529;
or.b32 %r534, %r1090, %r1089;

dp4a.s32.s32 %r533, %r534, %r535, %r732;

mul.lo.s32 %r1091, %r769, %r533;
cvt.rn.f32.s32 %f259, %r1091;
fma.rn.ftz.f32 %f260, %f149, %f259, %f258;
mov.u32 %r537, %r681;

prmt.b32 %r537, %r537, 0, 0xba98;

mov.u32 %r539, %r683;

prmt.b32 %r539, %r539, 0, 0xba98;

xor.b32 %r1092, %r539, 2139062143;
not.b32 %r1093, %r537;
and.b32 %r1094, %r797, %r1093;
and.b32 %r1095, %r1092, %r537;
or.b32 %r542, %r1095, %r1094;

dp4a.s32.s32 %r541, %r542, %r543, %r732;

mul.lo.s32 %r1096, %r792, %r541;
cvt.rn.f32.s32 %f261, %r1096;
fma.rn.ftz.f32 %f262, %f150, %f261, %f260;
mov.u32 %r545, %r689;

prmt.b32 %r545, %r545, 0, 0xba98;

mov.u32 %r547, %r691;

prmt.b32 %r547, %r547, 0, 0xba98;

xor.b32 %r1097, %r547, 2139062143;
not.b32 %r1098, %r545;
and.b32 %r1099, %r820, %r1098;
and.b32 %r1100, %r1097, %r545;
or.b32 %r550, %r1100, %r1099;

dp4a.s32.s32 %r549, %r550, %r551, %r732;

mul.lo.s32 %r1101, %r814, %r549;
cvt.rn.f32.s32 %f263, %r1101;
fma.rn.ftz.f32 %f264, %f151, %f263, %f262;
fma.rn.ftz.f32 %f502, %f147, %f264, %f502;

{ cvt.f32.f16 %f152, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r517;
cvt.f32.f16 %f153, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r518;
cvt.f32.f16 %f154, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r519;
cvt.f32.f16 %f155, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r520;
cvt.f32.f16 %f156, low;}


mov.u32 %r557, %r701;

prmt.b32 %r557, %r557, 0, 0xba98;

mov.u32 %r559, %r703;

prmt.b32 %r559, %r559, 0, 0xba98;

xor.b32 %r1102, %r559, 2139062143;
not.b32 %r1103, %r557;
and.b32 %r1104, %r846, %r1103;
and.b32 %r1105, %r1102, %r557;
or.b32 %r562, %r1105, %r1104;

dp4a.s32.s32 %r561, %r562, %r527, %r732;

mul.lo.s32 %r1106, %r841, %r561;
cvt.rn.f32.s32 %f265, %r1106;
fma.rn.ftz.f32 %f266, %f153, %f265, 0f00000000;
mov.u32 %r565, %r709;

prmt.b32 %r565, %r565, 0, 0xba98;

mov.u32 %r567, %r711;

prmt.b32 %r567, %r567, 0, 0xba98;

xor.b32 %r1107, %r567, 2139062143;
not.b32 %r1108, %r565;
and.b32 %r1109, %r869, %r1108;
and.b32 %r1110, %r1107, %r565;
or.b32 %r570, %r1110, %r1109;

dp4a.s32.s32 %r569, %r570, %r535, %r732;

mul.lo.s32 %r1111, %r863, %r569;
cvt.rn.f32.s32 %f267, %r1111;
fma.rn.ftz.f32 %f268, %f154, %f267, %f266;
mov.u32 %r573, %r717;

prmt.b32 %r573, %r573, 0, 0xba98;

mov.u32 %r575, %r719;

prmt.b32 %r575, %r575, 0, 0xba98;

xor.b32 %r1112, %r575, 2139062143;
not.b32 %r1113, %r573;
and.b32 %r1114, %r891, %r1113;
and.b32 %r1115, %r1112, %r573;
or.b32 %r578, %r1115, %r1114;

dp4a.s32.s32 %r577, %r578, %r543, %r732;

mul.lo.s32 %r1116, %r886, %r577;
cvt.rn.f32.s32 %f269, %r1116;
fma.rn.ftz.f32 %f270, %f155, %f269, %f268;
mov.u32 %r581, %r725;

prmt.b32 %r581, %r581, 0, 0xba98;

mov.u32 %r583, %r727;

prmt.b32 %r583, %r583, 0, 0xba98;

xor.b32 %r1117, %r583, 2139062143;
not.b32 %r1118, %r581;
and.b32 %r1119, %r914, %r1118;
and.b32 %r1120, %r1117, %r581;
or.b32 %r586, %r1120, %r1119;

dp4a.s32.s32 %r585, %r586, %r551, %r732;

mul.lo.s32 %r1121, %r908, %r585;
cvt.rn.f32.s32 %f271, %r1121;
fma.rn.ftz.f32 %f272, %f156, %f271, %f270;
fma.rn.ftz.f32 %f501, %f152, %f272, %f501;
cvt.u64.u32 %rd128, %r1521;
add.s64 %rd129, %rd2, %rd128;

{ cvt.f32.f16 %f157, %rs1;}


add.s64 %rd130, %rd129, %rd5;
mul.lo.s64 %rd131, %rd130, 36;
add.s64 %rd132, %rd28, %rd131;
add.s64 %rd133, %rd132, %rd77;
ld.global.nc.u32 %r599, [%rd133+4];
ld.global.nc.u32 %r589, [%rd132];

{.reg .f16 low,high;
mov.b32 {low,high},%r589;
cvt.f32.f16 %f158, low;}


ld.global.nc.u32 %r607, [%rd133+40];
ld.global.nc.u32 %r590, [%rd132+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r590;
cvt.f32.f16 %f159, low;}


ld.global.nc.u32 %r615, [%rd133+76];
ld.global.nc.u32 %r591, [%rd132+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r591;
cvt.f32.f16 %f160, low;}


ld.global.nc.u32 %r623, [%rd133+112];
ld.global.nc.u32 %r592, [%rd132+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r592;
cvt.f32.f16 %f161, low;}


mov.u32 %r593, %r665;

prmt.b32 %r593, %r593, 0, 0xba98;

mov.u32 %r595, %r667;

prmt.b32 %r595, %r595, 0, 0xba98;

xor.b32 %r1122, %r595, 2139062143;
not.b32 %r1123, %r593;
and.b32 %r1124, %r752, %r1123;
and.b32 %r1125, %r1122, %r593;
or.b32 %r598, %r1125, %r1124;

dp4a.s32.s32 %r597, %r598, %r599, %r732;

mul.lo.s32 %r1126, %r747, %r597;
cvt.rn.f32.s32 %f273, %r1126;
fma.rn.ftz.f32 %f274, %f158, %f273, 0f00000000;
mov.u32 %r601, %r673;

prmt.b32 %r601, %r601, 0, 0xba98;

mov.u32 %r603, %r675;

prmt.b32 %r603, %r603, 0, 0xba98;

xor.b32 %r1127, %r603, 2139062143;
not.b32 %r1128, %r601;
and.b32 %r1129, %r775, %r1128;
and.b32 %r1130, %r1127, %r601;
or.b32 %r606, %r1130, %r1129;

dp4a.s32.s32 %r605, %r606, %r607, %r732;

mul.lo.s32 %r1131, %r769, %r605;
cvt.rn.f32.s32 %f275, %r1131;
fma.rn.ftz.f32 %f276, %f159, %f275, %f274;
mov.u32 %r609, %r681;

prmt.b32 %r609, %r609, 0, 0xba98;

mov.u32 %r611, %r683;

prmt.b32 %r611, %r611, 0, 0xba98;

xor.b32 %r1132, %r611, 2139062143;
not.b32 %r1133, %r609;
and.b32 %r1134, %r797, %r1133;
and.b32 %r1135, %r1132, %r609;
or.b32 %r614, %r1135, %r1134;

dp4a.s32.s32 %r613, %r614, %r615, %r732;

mul.lo.s32 %r1136, %r792, %r613;
cvt.rn.f32.s32 %f277, %r1136;
fma.rn.ftz.f32 %f278, %f160, %f277, %f276;
mov.u32 %r617, %r689;

prmt.b32 %r617, %r617, 0, 0xba98;

mov.u32 %r619, %r691;

prmt.b32 %r619, %r619, 0, 0xba98;

xor.b32 %r1137, %r619, 2139062143;
not.b32 %r1138, %r617;
and.b32 %r1139, %r820, %r1138;
and.b32 %r1140, %r1137, %r617;
or.b32 %r622, %r1140, %r1139;

dp4a.s32.s32 %r621, %r622, %r623, %r732;

mul.lo.s32 %r1141, %r814, %r621;
cvt.rn.f32.s32 %f279, %r1141;
fma.rn.ftz.f32 %f280, %f161, %f279, %f278;
fma.rn.ftz.f32 %f500, %f157, %f280, %f500;

{ cvt.f32.f16 %f162, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r589;
cvt.f32.f16 %f163, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r590;
cvt.f32.f16 %f164, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r591;
cvt.f32.f16 %f165, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r592;
cvt.f32.f16 %f166, low;}


mov.u32 %r629, %r701;

prmt.b32 %r629, %r629, 0, 0xba98;

mov.u32 %r631, %r703;

prmt.b32 %r631, %r631, 0, 0xba98;

xor.b32 %r1142, %r631, 2139062143;
not.b32 %r1143, %r629;
and.b32 %r1144, %r846, %r1143;
and.b32 %r1145, %r1142, %r629;
or.b32 %r634, %r1145, %r1144;

dp4a.s32.s32 %r633, %r634, %r599, %r732;

mul.lo.s32 %r1146, %r841, %r633;
cvt.rn.f32.s32 %f281, %r1146;
fma.rn.ftz.f32 %f282, %f163, %f281, 0f00000000;
mov.u32 %r637, %r709;

prmt.b32 %r637, %r637, 0, 0xba98;

mov.u32 %r639, %r711;

prmt.b32 %r639, %r639, 0, 0xba98;

xor.b32 %r1147, %r639, 2139062143;
not.b32 %r1148, %r637;
and.b32 %r1149, %r869, %r1148;
and.b32 %r1150, %r1147, %r637;
or.b32 %r642, %r1150, %r1149;

dp4a.s32.s32 %r641, %r642, %r607, %r732;

mul.lo.s32 %r1151, %r863, %r641;
cvt.rn.f32.s32 %f283, %r1151;
fma.rn.ftz.f32 %f284, %f164, %f283, %f282;
mov.u32 %r645, %r717;

prmt.b32 %r645, %r645, 0, 0xba98;

mov.u32 %r647, %r719;

prmt.b32 %r647, %r647, 0, 0xba98;

xor.b32 %r1152, %r647, 2139062143;
not.b32 %r1153, %r645;
and.b32 %r1154, %r891, %r1153;
and.b32 %r1155, %r1152, %r645;
or.b32 %r650, %r1155, %r1154;

dp4a.s32.s32 %r649, %r650, %r615, %r732;

mul.lo.s32 %r1156, %r886, %r649;
cvt.rn.f32.s32 %f285, %r1156;
fma.rn.ftz.f32 %f286, %f165, %f285, %f284;
mov.u32 %r653, %r725;

prmt.b32 %r653, %r653, 0, 0xba98;

mov.u32 %r655, %r727;

prmt.b32 %r655, %r655, 0, 0xba98;

xor.b32 %r1157, %r655, 2139062143;
not.b32 %r1158, %r653;
and.b32 %r1159, %r914, %r1158;
and.b32 %r1160, %r1157, %r653;
or.b32 %r658, %r1160, %r1159;

dp4a.s32.s32 %r657, %r658, %r623, %r732;

mul.lo.s32 %r1161, %r908, %r657;
cvt.rn.f32.s32 %f287, %r1161;
fma.rn.ftz.f32 %f288, %f166, %f287, %f286;
fma.rn.ftz.f32 %f499, %f162, %f288, %f499;
cvt.u64.u32 %rd134, %r1520;
add.s64 %rd135, %rd2, %rd134;

{ cvt.f32.f16 %f167, %rs1;}


add.s64 %rd136, %rd135, %rd5;
mul.lo.s64 %rd137, %rd136, 36;
add.s64 %rd138, %rd28, %rd137;
add.s64 %rd139, %rd138, %rd77;
ld.global.nc.u32 %r671, [%rd139+4];
ld.global.nc.u32 %r661, [%rd138];

{.reg .f16 low,high;
mov.b32 {low,high},%r661;
cvt.f32.f16 %f168, low;}


ld.global.nc.u32 %r679, [%rd139+40];
ld.global.nc.u32 %r662, [%rd138+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r662;
cvt.f32.f16 %f169, low;}


ld.global.nc.u32 %r687, [%rd139+76];
ld.global.nc.u32 %r663, [%rd138+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r663;
cvt.f32.f16 %f170, low;}


ld.global.nc.u32 %r695, [%rd139+112];
ld.global.nc.u32 %r664, [%rd138+108];

{.reg .f16 low,high;
mov.b32 {low,high},%r664;
cvt.f32.f16 %f171, low;}



prmt.b32 %r665, %r665, 0, 0xba98;


prmt.b32 %r667, %r667, 0, 0xba98;

xor.b32 %r1162, %r667, 2139062143;
not.b32 %r1163, %r665;
and.b32 %r1164, %r752, %r1163;
and.b32 %r1165, %r1162, %r665;
or.b32 %r670, %r1165, %r1164;

dp4a.s32.s32 %r669, %r670, %r671, %r732;

mul.lo.s32 %r1166, %r747, %r669;
cvt.rn.f32.s32 %f289, %r1166;
fma.rn.ftz.f32 %f290, %f168, %f289, 0f00000000;

prmt.b32 %r673, %r673, 0, 0xba98;


prmt.b32 %r675, %r675, 0, 0xba98;

xor.b32 %r1167, %r675, 2139062143;
not.b32 %r1168, %r673;
and.b32 %r1169, %r775, %r1168;
and.b32 %r1170, %r1167, %r673;
or.b32 %r678, %r1170, %r1169;

dp4a.s32.s32 %r677, %r678, %r679, %r732;

mul.lo.s32 %r1171, %r769, %r677;
cvt.rn.f32.s32 %f291, %r1171;
fma.rn.ftz.f32 %f292, %f169, %f291, %f290;

prmt.b32 %r681, %r681, 0, 0xba98;


prmt.b32 %r683, %r683, 0, 0xba98;

xor.b32 %r1172, %r683, 2139062143;
not.b32 %r1173, %r681;
and.b32 %r1174, %r797, %r1173;
and.b32 %r1175, %r1172, %r681;
or.b32 %r686, %r1175, %r1174;

dp4a.s32.s32 %r685, %r686, %r687, %r732;

mul.lo.s32 %r1176, %r792, %r685;
cvt.rn.f32.s32 %f293, %r1176;
fma.rn.ftz.f32 %f294, %f170, %f293, %f292;

prmt.b32 %r689, %r689, 0, 0xba98;


prmt.b32 %r691, %r691, 0, 0xba98;

xor.b32 %r1177, %r691, 2139062143;
not.b32 %r1178, %r689;
and.b32 %r1179, %r820, %r1178;
and.b32 %r1180, %r1177, %r689;
or.b32 %r694, %r1180, %r1179;

dp4a.s32.s32 %r693, %r694, %r695, %r732;

mul.lo.s32 %r1181, %r814, %r693;
cvt.rn.f32.s32 %f295, %r1181;
fma.rn.ftz.f32 %f296, %f171, %f295, %f294;
fma.rn.ftz.f32 %f498, %f167, %f296, %f498;

{ cvt.f32.f16 %f172, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r661;
cvt.f32.f16 %f173, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r662;
cvt.f32.f16 %f174, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r663;
cvt.f32.f16 %f175, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r664;
cvt.f32.f16 %f176, low;}



prmt.b32 %r701, %r701, 0, 0xba98;


prmt.b32 %r703, %r703, 0, 0xba98;

xor.b32 %r1182, %r703, 2139062143;
not.b32 %r1183, %r701;
and.b32 %r1184, %r846, %r1183;
and.b32 %r1185, %r1182, %r701;
or.b32 %r706, %r1185, %r1184;

dp4a.s32.s32 %r705, %r706, %r671, %r732;

mul.lo.s32 %r1186, %r841, %r705;
cvt.rn.f32.s32 %f297, %r1186;
fma.rn.ftz.f32 %f298, %f173, %f297, 0f00000000;

prmt.b32 %r709, %r709, 0, 0xba98;


prmt.b32 %r711, %r711, 0, 0xba98;

xor.b32 %r1187, %r711, 2139062143;
not.b32 %r1188, %r709;
and.b32 %r1189, %r869, %r1188;
and.b32 %r1190, %r1187, %r709;
or.b32 %r714, %r1190, %r1189;

dp4a.s32.s32 %r713, %r714, %r679, %r732;

mul.lo.s32 %r1191, %r863, %r713;
cvt.rn.f32.s32 %f299, %r1191;
fma.rn.ftz.f32 %f300, %f174, %f299, %f298;

prmt.b32 %r717, %r717, 0, 0xba98;


prmt.b32 %r719, %r719, 0, 0xba98;

xor.b32 %r1192, %r719, 2139062143;
not.b32 %r1193, %r717;
and.b32 %r1194, %r891, %r1193;
and.b32 %r1195, %r1192, %r717;
or.b32 %r722, %r1195, %r1194;

dp4a.s32.s32 %r721, %r722, %r687, %r732;

mul.lo.s32 %r1196, %r886, %r721;
cvt.rn.f32.s32 %f301, %r1196;
fma.rn.ftz.f32 %f302, %f175, %f301, %f300;

prmt.b32 %r725, %r725, 0, 0xba98;


prmt.b32 %r727, %r727, 0, 0xba98;

xor.b32 %r1197, %r727, 2139062143;
not.b32 %r1198, %r725;
and.b32 %r1199, %r914, %r1198;
and.b32 %r1200, %r1197, %r725;
or.b32 %r730, %r1200, %r1199;

dp4a.s32.s32 %r729, %r730, %r695, %r732;

mul.lo.s32 %r1201, %r908, %r729;
cvt.rn.f32.s32 %f303, %r1201;
fma.rn.ftz.f32 %f304, %f176, %f303, %f302;
fma.rn.ftz.f32 %f497, %f172, %f304, %f497;
add.s32 %r1525, %r1525, 32;
add.s32 %r1524, %r1524, 32;
add.s32 %r1523, %r1523, 32;
add.s32 %r1522, %r1522, 32;
add.s32 %r1521, %r1521, 32;
add.s32 %r1520, %r1520, 32;
add.s32 %r1519, %r1519, 32;
add.s32 %r1518, %r1518, 32;
add.s64 %rd182, %rd182, 440;
add.s32 %r1526, %r1526, 4;
setp.lt.s32 %p4, %r1526, %r4;
@%p4 bra $L__BB152_2;

$L__BB152_3:
mov.u32 %r1500, %tid.y;
setp.eq.s32 %p5, %r1500, 0;
@%p5 bra $L__BB152_5;

mov.u32 %r1502, %tid.x;
mov.u32 %r1501, %tid.y;
shl.b32 %r1203, %r1501, 11;
mov.u32 %r1204, _ZZ13mul_mat_vec_qIL9ggml_type11ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r1205, %r1204, %r1203;
shl.b32 %r1206, %r1502, 2;
add.s32 %r1207, %r1205, %r1206;
st.shared.f32 [%r1207+-2048], %f512;
st.shared.f32 [%r1207+-1920], %f511;
st.shared.f32 [%r1207+-1792], %f510;
st.shared.f32 [%r1207+-1664], %f509;
st.shared.f32 [%r1207+-1536], %f508;
st.shared.f32 [%r1207+-1408], %f507;
st.shared.f32 [%r1207+-1280], %f506;
st.shared.f32 [%r1207+-1152], %f505;
st.shared.f32 [%r1207+-1024], %f504;
st.shared.f32 [%r1207+-896], %f503;
st.shared.f32 [%r1207+-768], %f502;
st.shared.f32 [%r1207+-640], %f501;
st.shared.f32 [%r1207+-512], %f500;
st.shared.f32 [%r1207+-384], %f499;
st.shared.f32 [%r1207+-256], %f498;
st.shared.f32 [%r1207+-128], %f497;

$L__BB152_5:
mov.u32 %r1503, %tid.y;
setp.ne.s32 %p6, %r1503, 0;
bar.sync 0;
@%p6 bra $L__BB152_38;

ld.param.u32 %r1507, [_Z13mul_mat_vec_qIL9ggml_type11ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r1506, [_Z13mul_mat_vec_qIL9ggml_type11ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r1505, [_Z13mul_mat_vec_qIL9ggml_type11ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
mov.u32 %r1504, %tid.x;
mov.u32 %r1208, %ctaid.z;
mov.u32 %r1209, %ctaid.y;
mov.u32 %r1210, %ctaid.x;
shl.b32 %r1211, %r1210, 1;
mov.u32 %r1212, 1;
mad.lo.s32 %r1213, %r1209, %r1505, %r1211;
mad.lo.s32 %r1214, %r1208, %r1506, %r1213;
cvt.u64.u32 %rd31, %r1214;
shl.b32 %r1216, %r1504, 2;
mov.u32 %r1217, _ZZ13mul_mat_vec_qIL9ggml_type11ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r39, %r1217, %r1216;
mov.u32 %r1218, 2;
ld.shared.f32 %f305, [%r39];
add.ftz.f32 %f306, %f305, %f512;
mov.b32 %r1219, %f306;
mov.u32 %r1220, 31;
mov.u32 %r1221, 16;
mov.u32 %r1222, -1;
shfl.sync.bfly.b32 %r1223|%p7, %r1219, %r1221, %r1220, %r1222;
mov.b32 %f307, %r1223;
add.ftz.f32 %f308, %f306, %f307;
mov.b32 %r1224, %f308;
mov.u32 %r1225, 8;
shfl.sync.bfly.b32 %r1226|%p8, %r1224, %r1225, %r1220, %r1222;
mov.b32 %f309, %r1226;
add.ftz.f32 %f310, %f308, %f309;
mov.b32 %r1227, %f310;
mov.u32 %r1228, 4;
shfl.sync.bfly.b32 %r1229|%p9, %r1227, %r1228, %r1220, %r1222;
mov.b32 %f311, %r1229;
add.ftz.f32 %f312, %f310, %f311;
mov.b32 %r1230, %f312;
shfl.sync.bfly.b32 %r1231|%p10, %r1230, %r1218, %r1220, %r1222;
mov.b32 %f313, %r1231;
add.ftz.f32 %f314, %f312, %f313;
mov.b32 %r1232, %f314;
shfl.sync.bfly.b32 %r1233|%p11, %r1232, %r1212, %r1220, %r1222;
mov.b32 %f315, %r1233;
add.ftz.f32 %f49, %f314, %f315;
setp.eq.s32 %p12, %r1504, 0;
setp.lt.u32 %p13, %r1211, %r1507;
and.pred %p1, %p12, %p13;
mul.wide.u32 %rd140, %r1214, 4;
add.s64 %rd32, %rd1, %rd140;
not.pred %p14, %p1;
@%p14 bra $L__BB152_8;

st.global.f32 [%rd32], %f49;

$L__BB152_8:
ld.param.u32 %r1509, [_Z13mul_mat_vec_qIL9ggml_type11ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
mov.u32 %r1508, %tid.x;
ld.shared.f32 %f316, [%r39+128];
add.ftz.f32 %f317, %f316, %f511;
mov.b32 %r1234, %f317;
shfl.sync.bfly.b32 %r1238|%p15, %r1234, %r1221, %r1220, %r1222;
mov.b32 %f318, %r1238;
add.ftz.f32 %f319, %f317, %f318;
mov.b32 %r1239, %f319;
shfl.sync.bfly.b32 %r1241|%p16, %r1239, %r1225, %r1220, %r1222;
mov.b32 %f320, %r1241;
add.ftz.f32 %f321, %f319, %f320;
mov.b32 %r1242, %f321;
shfl.sync.bfly.b32 %r1244|%p17, %r1242, %r1228, %r1220, %r1222;
mov.b32 %f322, %r1244;
add.ftz.f32 %f323, %f321, %f322;
mov.b32 %r1245, %f323;
shfl.sync.bfly.b32 %r1247|%p18, %r1245, %r1218, %r1220, %r1222;
mov.b32 %f324, %r1247;
add.ftz.f32 %f325, %f323, %f324;
mov.b32 %r1248, %f325;
shfl.sync.bfly.b32 %r1250|%p19, %r1248, %r1212, %r1220, %r1222;
mov.b32 %f326, %r1250;
add.ftz.f32 %f50, %f325, %f326;
setp.eq.s32 %p20, %r1508, 1;
or.b32 %r1254, %r1211, 1;
setp.lt.u32 %p21, %r1254, %r1509;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB152_10;

st.global.f32 [%rd32+4], %f50;

$L__BB152_10:
ld.shared.f32 %f327, [%r39+256];
add.ftz.f32 %f328, %f327, %f510;
mov.b32 %r1255, %f328;
mov.u32 %r1256, 31;
mov.u32 %r1257, 16;
mov.u32 %r1258, -1;
shfl.sync.bfly.b32 %r1259|%p23, %r1255, %r1257, %r1256, %r1258;
mov.b32 %f329, %r1259;
add.ftz.f32 %f330, %f328, %f329;
mov.b32 %r1260, %f330;
mov.u32 %r1261, 8;
shfl.sync.bfly.b32 %r1262|%p24, %r1260, %r1261, %r1256, %r1258;
mov.b32 %f331, %r1262;
add.ftz.f32 %f332, %f330, %f331;
mov.b32 %r1263, %f332;
mov.u32 %r1264, 4;
shfl.sync.bfly.b32 %r1265|%p25, %r1263, %r1264, %r1256, %r1258;
mov.b32 %f333, %r1265;
add.ftz.f32 %f334, %f332, %f333;
mov.b32 %r1266, %f334;
mov.u32 %r1267, 2;
shfl.sync.bfly.b32 %r1268|%p26, %r1266, %r1267, %r1256, %r1258;
mov.b32 %f335, %r1268;
add.ftz.f32 %f336, %f334, %f335;
mov.b32 %r1269, %f336;
mov.u32 %r1270, 1;
shfl.sync.bfly.b32 %r1271|%p27, %r1269, %r1270, %r1256, %r1258;
mov.b32 %f337, %r1271;
add.ftz.f32 %f51, %f336, %f337;
@%p14 bra $L__BB152_12;

ld.param.u32 %r1517, [_Z13mul_mat_vec_qIL9ggml_type11ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
cvt.u64.u32 %rd141, %r1517;
add.s64 %rd142, %rd141, %rd31;
shl.b64 %rd143, %rd142, 2;
add.s64 %rd144, %rd1, %rd143;
st.global.f32 [%rd144], %f51;

$L__BB152_12:
ld.shared.f32 %f338, [%r39+384];
add.ftz.f32 %f339, %f338, %f509;
mov.b32 %r1272, %f339;
shfl.sync.bfly.b32 %r1276|%p29, %r1272, %r1257, %r1256, %r1258;
mov.b32 %f340, %r1276;
add.ftz.f32 %f341, %f339, %f340;
mov.b32 %r1277, %f341;
shfl.sync.bfly.b32 %r1279|%p30, %r1277, %r1261, %r1256, %r1258;
mov.b32 %f342, %r1279;
add.ftz.f32 %f343, %f341, %f342;
mov.b32 %r1280, %f343;
shfl.sync.bfly.b32 %r1282|%p31, %r1280, %r1264, %r1256, %r1258;
mov.b32 %f344, %r1282;
add.ftz.f32 %f345, %f343, %f344;
mov.b32 %r1283, %f345;
shfl.sync.bfly.b32 %r1285|%p32, %r1283, %r1267, %r1256, %r1258;
mov.b32 %f346, %r1285;
add.ftz.f32 %f347, %f345, %f346;
mov.b32 %r1286, %f347;
shfl.sync.bfly.b32 %r1288|%p33, %r1286, %r1270, %r1256, %r1258;
mov.b32 %f348, %r1288;
add.ftz.f32 %f52, %f347, %f348;
@%p22 bra $L__BB152_14;

ld.param.u32 %r1516, [_Z13mul_mat_vec_qIL9ggml_type11ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
add.s32 %r1289, %r1516, 1;
cvt.u64.u32 %rd145, %r1289;
add.s64 %rd146, %rd145, %rd31;
shl.b64 %rd147, %rd146, 2;
add.s64 %rd148, %rd1, %rd147;
st.global.f32 [%rd148], %f52;

$L__BB152_14:
ld.param.u32 %r1510, [_Z13mul_mat_vec_qIL9ggml_type11ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
shl.b32 %r1290, %r1510, 1;
mov.u32 %r1291, 1;
ld.shared.f32 %f349, [%r39+512];
add.ftz.f32 %f350, %f349, %f508;
mov.b32 %r1292, %f350;
mov.u32 %r1293, 31;
mov.u32 %r1294, 16;
mov.u32 %r1295, -1;
shfl.sync.bfly.b32 %r1296|%p35, %r1292, %r1294, %r1293, %r1295;
mov.b32 %f351, %r1296;
add.ftz.f32 %f352, %f350, %f351;
mov.b32 %r1297, %f352;
mov.u32 %r1298, 8;
shfl.sync.bfly.b32 %r1299|%p36, %r1297, %r1298, %r1293, %r1295;
mov.b32 %f353, %r1299;
add.ftz.f32 %f354, %f352, %f353;
mov.b32 %r1300, %f354;
mov.u32 %r1301, 4;
shfl.sync.bfly.b32 %r1302|%p37, %r1300, %r1301, %r1293, %r1295;
mov.b32 %f355, %r1302;
add.ftz.f32 %f356, %f354, %f355;
mov.b32 %r1303, %f356;
mov.u32 %r1304, 2;
shfl.sync.bfly.b32 %r1305|%p38, %r1303, %r1304, %r1293, %r1295;
mov.b32 %f357, %r1305;
add.ftz.f32 %f358, %f356, %f357;
mov.b32 %r1306, %f358;
shfl.sync.bfly.b32 %r1307|%p39, %r1306, %r1291, %r1293, %r1295;
mov.b32 %f359, %r1307;
add.ftz.f32 %f53, %f358, %f359;
cvt.u64.u32 %rd149, %r1290;
add.s64 %rd150, %rd149, %rd31;
shl.b64 %rd151, %rd150, 2;
add.s64 %rd33, %rd1, %rd151;
@%p14 bra $L__BB152_16;

st.global.f32 [%rd33], %f53;

$L__BB152_16:
ld.shared.f32 %f360, [%r39+640];
add.ftz.f32 %f361, %f360, %f507;
mov.b32 %r1308, %f361;
shfl.sync.bfly.b32 %r1312|%p41, %r1308, %r1294, %r1293, %r1295;
mov.b32 %f362, %r1312;
add.ftz.f32 %f363, %f361, %f362;
mov.b32 %r1313, %f363;
shfl.sync.bfly.b32 %r1315|%p42, %r1313, %r1298, %r1293, %r1295;
mov.b32 %f364, %r1315;
add.ftz.f32 %f365, %f363, %f364;
mov.b32 %r1316, %f365;
shfl.sync.bfly.b32 %r1318|%p43, %r1316, %r1301, %r1293, %r1295;
mov.b32 %f366, %r1318;
add.ftz.f32 %f367, %f365, %f366;
mov.b32 %r1319, %f367;
shfl.sync.bfly.b32 %r1321|%p44, %r1319, %r1304, %r1293, %r1295;
mov.b32 %f368, %r1321;
add.ftz.f32 %f369, %f367, %f368;
mov.b32 %r1322, %f369;
shfl.sync.bfly.b32 %r1324|%p45, %r1322, %r1291, %r1293, %r1295;
mov.b32 %f370, %r1324;
add.ftz.f32 %f54, %f369, %f370;
@%p22 bra $L__BB152_18;

st.global.f32 [%rd33+4], %f54;

$L__BB152_18:
ld.param.u32 %r1511, [_Z13mul_mat_vec_qIL9ggml_type11ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
mul.lo.s32 %r40, %r1511, 3;
ld.shared.f32 %f371, [%r39+768];
add.ftz.f32 %f372, %f371, %f506;
mov.b32 %r1325, %f372;
mov.u32 %r1326, 31;
mov.u32 %r1327, 16;
mov.u32 %r1328, -1;
shfl.sync.bfly.b32 %r1329|%p47, %r1325, %r1327, %r1326, %r1328;
mov.b32 %f373, %r1329;
add.ftz.f32 %f374, %f372, %f373;
mov.b32 %r1330, %f374;
mov.u32 %r1331, 8;
shfl.sync.bfly.b32 %r1332|%p48, %r1330, %r1331, %r1326, %r1328;
mov.b32 %f375, %r1332;
add.ftz.f32 %f376, %f374, %f375;
mov.b32 %r1333, %f376;
mov.u32 %r1334, 4;
shfl.sync.bfly.b32 %r1335|%p49, %r1333, %r1334, %r1326, %r1328;
mov.b32 %f377, %r1335;
add.ftz.f32 %f378, %f376, %f377;
mov.b32 %r1336, %f378;
mov.u32 %r1337, 2;
shfl.sync.bfly.b32 %r1338|%p50, %r1336, %r1337, %r1326, %r1328;
mov.b32 %f379, %r1338;
add.ftz.f32 %f380, %f378, %f379;
mov.b32 %r1339, %f380;
mov.u32 %r1340, 1;
shfl.sync.bfly.b32 %r1341|%p51, %r1339, %r1340, %r1326, %r1328;
mov.b32 %f381, %r1341;
add.ftz.f32 %f55, %f380, %f381;
@%p14 bra $L__BB152_20;

cvt.u64.u32 %rd152, %r40;
add.s64 %rd153, %rd152, %rd31;
shl.b64 %rd154, %rd153, 2;
add.s64 %rd155, %rd1, %rd154;
st.global.f32 [%rd155], %f55;

$L__BB152_20:
ld.shared.f32 %f382, [%r39+896];
add.ftz.f32 %f383, %f382, %f505;
mov.b32 %r1342, %f383;
shfl.sync.bfly.b32 %r1346|%p53, %r1342, %r1327, %r1326, %r1328;
mov.b32 %f384, %r1346;
add.ftz.f32 %f385, %f383, %f384;
mov.b32 %r1347, %f385;
shfl.sync.bfly.b32 %r1349|%p54, %r1347, %r1331, %r1326, %r1328;
mov.b32 %f386, %r1349;
add.ftz.f32 %f387, %f385, %f386;
mov.b32 %r1350, %f387;
shfl.sync.bfly.b32 %r1352|%p55, %r1350, %r1334, %r1326, %r1328;
mov.b32 %f388, %r1352;
add.ftz.f32 %f389, %f387, %f388;
mov.b32 %r1353, %f389;
shfl.sync.bfly.b32 %r1355|%p56, %r1353, %r1337, %r1326, %r1328;
mov.b32 %f390, %r1355;
add.ftz.f32 %f391, %f389, %f390;
mov.b32 %r1356, %f391;
shfl.sync.bfly.b32 %r1358|%p57, %r1356, %r1340, %r1326, %r1328;
mov.b32 %f392, %r1358;
add.ftz.f32 %f56, %f391, %f392;
@%p22 bra $L__BB152_22;

add.s32 %r1359, %r40, 1;
cvt.u64.u32 %rd156, %r1359;
add.s64 %rd157, %rd156, %rd31;
shl.b64 %rd158, %rd157, 2;
add.s64 %rd159, %rd1, %rd158;
st.global.f32 [%rd159], %f56;

$L__BB152_22:
ld.param.u32 %r1512, [_Z13mul_mat_vec_qIL9ggml_type11ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
shl.b32 %r1360, %r1512, 2;
mov.u32 %r1361, 2;
ld.shared.f32 %f393, [%r39+1024];
add.ftz.f32 %f394, %f393, %f504;
mov.b32 %r1362, %f394;
mov.u32 %r1363, 31;
mov.u32 %r1364, 16;
mov.u32 %r1365, -1;
shfl.sync.bfly.b32 %r1366|%p59, %r1362, %r1364, %r1363, %r1365;
mov.b32 %f395, %r1366;
add.ftz.f32 %f396, %f394, %f395;
mov.b32 %r1367, %f396;
mov.u32 %r1368, 8;
shfl.sync.bfly.b32 %r1369|%p60, %r1367, %r1368, %r1363, %r1365;
mov.b32 %f397, %r1369;
add.ftz.f32 %f398, %f396, %f397;
mov.b32 %r1370, %f398;
mov.u32 %r1371, 4;
shfl.sync.bfly.b32 %r1372|%p61, %r1370, %r1371, %r1363, %r1365;
mov.b32 %f399, %r1372;
add.ftz.f32 %f400, %f398, %f399;
mov.b32 %r1373, %f400;
shfl.sync.bfly.b32 %r1374|%p62, %r1373, %r1361, %r1363, %r1365;
mov.b32 %f401, %r1374;
add.ftz.f32 %f402, %f400, %f401;
mov.b32 %r1375, %f402;
mov.u32 %r1376, 1;
shfl.sync.bfly.b32 %r1377|%p63, %r1375, %r1376, %r1363, %r1365;
mov.b32 %f403, %r1377;
add.ftz.f32 %f57, %f402, %f403;
cvt.u64.u32 %rd160, %r1360;
add.s64 %rd161, %rd160, %rd31;
shl.b64 %rd162, %rd161, 2;
add.s64 %rd34, %rd1, %rd162;
@%p14 bra $L__BB152_24;

st.global.f32 [%rd34], %f57;

$L__BB152_24:
ld.shared.f32 %f404, [%r39+1152];
add.ftz.f32 %f405, %f404, %f503;
mov.b32 %r1378, %f405;
shfl.sync.bfly.b32 %r1382|%p65, %r1378, %r1364, %r1363, %r1365;
mov.b32 %f406, %r1382;
add.ftz.f32 %f407, %f405, %f406;
mov.b32 %r1383, %f407;
shfl.sync.bfly.b32 %r1385|%p66, %r1383, %r1368, %r1363, %r1365;
mov.b32 %f408, %r1385;
add.ftz.f32 %f409, %f407, %f408;
mov.b32 %r1386, %f409;
shfl.sync.bfly.b32 %r1388|%p67, %r1386, %r1371, %r1363, %r1365;
mov.b32 %f410, %r1388;
add.ftz.f32 %f411, %f409, %f410;
mov.b32 %r1389, %f411;
shfl.sync.bfly.b32 %r1391|%p68, %r1389, %r1361, %r1363, %r1365;
mov.b32 %f412, %r1391;
add.ftz.f32 %f413, %f411, %f412;
mov.b32 %r1392, %f413;
shfl.sync.bfly.b32 %r1394|%p69, %r1392, %r1376, %r1363, %r1365;
mov.b32 %f414, %r1394;
add.ftz.f32 %f58, %f413, %f414;
@%p22 bra $L__BB152_26;

st.global.f32 [%rd34+4], %f58;

$L__BB152_26:
ld.param.u32 %r1513, [_Z13mul_mat_vec_qIL9ggml_type11ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
mul.lo.s32 %r41, %r1513, 5;
ld.shared.f32 %f415, [%r39+1280];
add.ftz.f32 %f416, %f415, %f502;
mov.b32 %r1395, %f416;
mov.u32 %r1396, 31;
mov.u32 %r1397, 16;
mov.u32 %r1398, -1;
shfl.sync.bfly.b32 %r1399|%p71, %r1395, %r1397, %r1396, %r1398;
mov.b32 %f417, %r1399;
add.ftz.f32 %f418, %f416, %f417;
mov.b32 %r1400, %f418;
mov.u32 %r1401, 8;
shfl.sync.bfly.b32 %r1402|%p72, %r1400, %r1401, %r1396, %r1398;
mov.b32 %f419, %r1402;
add.ftz.f32 %f420, %f418, %f419;
mov.b32 %r1403, %f420;
mov.u32 %r1404, 4;
shfl.sync.bfly.b32 %r1405|%p73, %r1403, %r1404, %r1396, %r1398;
mov.b32 %f421, %r1405;
add.ftz.f32 %f422, %f420, %f421;
mov.b32 %r1406, %f422;
mov.u32 %r1407, 2;
shfl.sync.bfly.b32 %r1408|%p74, %r1406, %r1407, %r1396, %r1398;
mov.b32 %f423, %r1408;
add.ftz.f32 %f424, %f422, %f423;
mov.b32 %r1409, %f424;
mov.u32 %r1410, 1;
shfl.sync.bfly.b32 %r1411|%p75, %r1409, %r1410, %r1396, %r1398;
mov.b32 %f425, %r1411;
add.ftz.f32 %f59, %f424, %f425;
@%p14 bra $L__BB152_28;

cvt.u64.u32 %rd163, %r41;
add.s64 %rd164, %rd163, %rd31;
shl.b64 %rd165, %rd164, 2;
add.s64 %rd166, %rd1, %rd165;
st.global.f32 [%rd166], %f59;

$L__BB152_28:
ld.shared.f32 %f426, [%r39+1408];
add.ftz.f32 %f427, %f426, %f501;
mov.b32 %r1412, %f427;
shfl.sync.bfly.b32 %r1416|%p77, %r1412, %r1397, %r1396, %r1398;
mov.b32 %f428, %r1416;
add.ftz.f32 %f429, %f427, %f428;
mov.b32 %r1417, %f429;
shfl.sync.bfly.b32 %r1419|%p78, %r1417, %r1401, %r1396, %r1398;
mov.b32 %f430, %r1419;
add.ftz.f32 %f431, %f429, %f430;
mov.b32 %r1420, %f431;
shfl.sync.bfly.b32 %r1422|%p79, %r1420, %r1404, %r1396, %r1398;
mov.b32 %f432, %r1422;
add.ftz.f32 %f433, %f431, %f432;
mov.b32 %r1423, %f433;
shfl.sync.bfly.b32 %r1425|%p80, %r1423, %r1407, %r1396, %r1398;
mov.b32 %f434, %r1425;
add.ftz.f32 %f435, %f433, %f434;
mov.b32 %r1426, %f435;
shfl.sync.bfly.b32 %r1428|%p81, %r1426, %r1410, %r1396, %r1398;
mov.b32 %f436, %r1428;
add.ftz.f32 %f60, %f435, %f436;
@%p22 bra $L__BB152_30;

add.s32 %r1429, %r41, 1;
cvt.u64.u32 %rd167, %r1429;
add.s64 %rd168, %rd167, %rd31;
shl.b64 %rd169, %rd168, 2;
add.s64 %rd170, %rd1, %rd169;
st.global.f32 [%rd170], %f60;

$L__BB152_30:
ld.param.u32 %r1514, [_Z13mul_mat_vec_qIL9ggml_type11ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
mul.lo.s32 %r1430, %r1514, 6;
ld.shared.f32 %f437, [%r39+1536];
add.ftz.f32 %f438, %f437, %f500;
mov.b32 %r1431, %f438;
mov.u32 %r1432, 31;
mov.u32 %r1433, 16;
mov.u32 %r1434, -1;
shfl.sync.bfly.b32 %r1435|%p83, %r1431, %r1433, %r1432, %r1434;
mov.b32 %f439, %r1435;
add.ftz.f32 %f440, %f438, %f439;
mov.b32 %r1436, %f440;
mov.u32 %r1437, 8;
shfl.sync.bfly.b32 %r1438|%p84, %r1436, %r1437, %r1432, %r1434;
mov.b32 %f441, %r1438;
add.ftz.f32 %f442, %f440, %f441;
mov.b32 %r1439, %f442;
mov.u32 %r1440, 4;
shfl.sync.bfly.b32 %r1441|%p85, %r1439, %r1440, %r1432, %r1434;
mov.b32 %f443, %r1441;
add.ftz.f32 %f444, %f442, %f443;
mov.b32 %r1442, %f444;
mov.u32 %r1443, 2;
shfl.sync.bfly.b32 %r1444|%p86, %r1442, %r1443, %r1432, %r1434;
mov.b32 %f445, %r1444;
add.ftz.f32 %f446, %f444, %f445;
mov.b32 %r1445, %f446;
mov.u32 %r1446, 1;
shfl.sync.bfly.b32 %r1447|%p87, %r1445, %r1446, %r1432, %r1434;
mov.b32 %f447, %r1447;
add.ftz.f32 %f61, %f446, %f447;
cvt.u64.u32 %rd171, %r1430;
add.s64 %rd172, %rd171, %rd31;
shl.b64 %rd173, %rd172, 2;
add.s64 %rd35, %rd1, %rd173;
@%p14 bra $L__BB152_32;

st.global.f32 [%rd35], %f61;

$L__BB152_32:
ld.shared.f32 %f448, [%r39+1664];
add.ftz.f32 %f449, %f448, %f499;
mov.b32 %r1448, %f449;
shfl.sync.bfly.b32 %r1452|%p89, %r1448, %r1433, %r1432, %r1434;
mov.b32 %f450, %r1452;
add.ftz.f32 %f451, %f449, %f450;
mov.b32 %r1453, %f451;
shfl.sync.bfly.b32 %r1455|%p90, %r1453, %r1437, %r1432, %r1434;
mov.b32 %f452, %r1455;
add.ftz.f32 %f453, %f451, %f452;
mov.b32 %r1456, %f453;
shfl.sync.bfly.b32 %r1458|%p91, %r1456, %r1440, %r1432, %r1434;
mov.b32 %f454, %r1458;
add.ftz.f32 %f455, %f453, %f454;
mov.b32 %r1459, %f455;
shfl.sync.bfly.b32 %r1461|%p92, %r1459, %r1443, %r1432, %r1434;
mov.b32 %f456, %r1461;
add.ftz.f32 %f457, %f455, %f456;
mov.b32 %r1462, %f457;
shfl.sync.bfly.b32 %r1464|%p93, %r1462, %r1446, %r1432, %r1434;
mov.b32 %f458, %r1464;
add.ftz.f32 %f62, %f457, %f458;
@%p22 bra $L__BB152_34;

st.global.f32 [%rd35+4], %f62;

$L__BB152_34:
ld.param.u32 %r1515, [_Z13mul_mat_vec_qIL9ggml_type11ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
mul.lo.s32 %r42, %r1515, 7;
ld.shared.f32 %f459, [%r39+1792];
add.ftz.f32 %f460, %f459, %f498;
mov.b32 %r1465, %f460;
mov.u32 %r1466, 31;
mov.u32 %r1467, 16;
mov.u32 %r1468, -1;
shfl.sync.bfly.b32 %r1469|%p95, %r1465, %r1467, %r1466, %r1468;
mov.b32 %f461, %r1469;
add.ftz.f32 %f462, %f460, %f461;
mov.b32 %r1470, %f462;
mov.u32 %r1471, 8;
shfl.sync.bfly.b32 %r1472|%p96, %r1470, %r1471, %r1466, %r1468;
mov.b32 %f463, %r1472;
add.ftz.f32 %f464, %f462, %f463;
mov.b32 %r1473, %f464;
mov.u32 %r1474, 4;
shfl.sync.bfly.b32 %r1475|%p97, %r1473, %r1474, %r1466, %r1468;
mov.b32 %f465, %r1475;
add.ftz.f32 %f466, %f464, %f465;
mov.b32 %r1476, %f466;
mov.u32 %r1477, 2;
shfl.sync.bfly.b32 %r1478|%p98, %r1476, %r1477, %r1466, %r1468;
mov.b32 %f467, %r1478;
add.ftz.f32 %f468, %f466, %f467;
mov.b32 %r1479, %f468;
mov.u32 %r1480, 1;
shfl.sync.bfly.b32 %r1481|%p99, %r1479, %r1480, %r1466, %r1468;
mov.b32 %f469, %r1481;
add.ftz.f32 %f63, %f468, %f469;
@%p14 bra $L__BB152_36;

cvt.u64.u32 %rd174, %r42;
add.s64 %rd175, %rd174, %rd31;
shl.b64 %rd176, %rd175, 2;
add.s64 %rd177, %rd1, %rd176;
st.global.f32 [%rd177], %f63;

$L__BB152_36:
ld.shared.f32 %f470, [%r39+1920];
add.ftz.f32 %f471, %f470, %f497;
mov.b32 %r1482, %f471;
shfl.sync.bfly.b32 %r1486|%p101, %r1482, %r1467, %r1466, %r1468;
mov.b32 %f472, %r1486;
add.ftz.f32 %f473, %f471, %f472;
mov.b32 %r1487, %f473;
shfl.sync.bfly.b32 %r1489|%p102, %r1487, %r1471, %r1466, %r1468;
mov.b32 %f474, %r1489;
add.ftz.f32 %f475, %f473, %f474;
mov.b32 %r1490, %f475;
shfl.sync.bfly.b32 %r1492|%p103, %r1490, %r1474, %r1466, %r1468;
mov.b32 %f476, %r1492;
add.ftz.f32 %f477, %f475, %f476;
mov.b32 %r1493, %f477;
shfl.sync.bfly.b32 %r1495|%p104, %r1493, %r1477, %r1466, %r1468;
mov.b32 %f478, %r1495;
add.ftz.f32 %f479, %f477, %f478;
mov.b32 %r1496, %f479;
shfl.sync.bfly.b32 %r1498|%p105, %r1496, %r1480, %r1466, %r1468;
mov.b32 %f480, %r1498;
add.ftz.f32 %f64, %f479, %f480;
@%p22 bra $L__BB152_38;

add.s32 %r1499, %r42, 1;
cvt.u64.u32 %rd178, %r1499;
add.s64 %rd179, %rd178, %rd31;
shl.b64 %rd180, %rd179, 2;
add.s64 %rd181, %rd1, %rd180;
st.global.f32 [%rd181], %f64;

$L__BB152_38:
ret;

}
.entry _Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj(
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2,
.param .align 8 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3[48],
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5,
.param .align 4 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6[12],
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15
)
.maxntid 256, 1, 1
.minnctapersm 1
{
.local .align 8 .b8 __local_depot153[16];
.reg .b64 %SP;
.reg .b64 %SPL;
.reg .pred %p<52>;
.reg .b16 %rs<23>;
.reg .f32 %f<213>;
.reg .b32 %r<446>;
.reg .b64 %rd<114>;


mov.u64 %SPL, __local_depot153;
ld.param.u64 %rd17, [_Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0];
ld.param.u64 %rd18, [_Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1];
ld.param.u64 %rd19, [_Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2];
ld.param.u32 %r18, [_Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5];
ld.param.u32 %r23, [_Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8];
ld.param.u32 %r24, [_Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9];
ld.param.u32 %r26, [_Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11];
ld.param.u32 %r27, [_Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12];
ld.param.u32 %r30, [_Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14];
ld.param.u32 %r29, [_Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15];
ld.param.u32 %r21, [_Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+8];
ld.param.u32 %r20, [_Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+4];
ld.param.u32 %r19, [_Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6];
ld.param.u64 %rd21, [_Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+8];
add.u64 %rd1, %SPL, 0;
add.u64 %rd2, %SPL, 8;
mov.u32 %r1, %tid.y;
setp.ge.u32 %p1, %r1, %r30;
@%p1 bra $L__BB153_29;

mov.u32 %r2, %ctaid.y;
cvta.to.global.u64 %rd28, %rd19;
mad.lo.s32 %r31, %r1, %r29, %r2;
mul.wide.u32 %rd29, %r31, 4;
add.s64 %rd30, %rd28, %rd29;
ld.global.nc.u32 %r3, [%rd30];
mov.f32 %f206, 0f00000000;
st.local.v2.f32 [%rd1], {%f206, %f206};
st.local.v2.f32 [%rd2], {%f206, %f206};
mov.u32 %r4, %tid.x;
shr.u32 %r445, %r4, 4;
shr.u32 %r32, %r18, 8;
setp.ge.u32 %p2, %r445, %r32;
mov.f32 %f207, %f206;
mov.f32 %f208, %f206;
mov.f32 %f209, %f206;
@%p2 bra $L__BB153_9;

shr.u32 %r33, %r4, 1;
and.b32 %r34, %r33, 6;
shl.b32 %r35, %r4, 3;
and.b32 %r36, %r35, 96;
shl.b32 %r37, %r4, 2;
and.b32 %r38, %r37, 12;
or.b32 %r39, %r36, %r38;
cvt.u64.u32 %rd31, %r39;
cvt.u64.u32 %rd32, %r34;
mov.u32 %r40, %ctaid.x;
shl.b32 %r41, %r40, 1;
mad.lo.s32 %r42, %r41, %r23, %r445;
mad.lo.s32 %r43, %r3, %r26, %r42;
mul.wide.s32 %rd33, %r43, 144;
add.s64 %rd34, %rd33, %rd31;
add.s64 %rd4, %rd34, 32;
cvt.u64.u32 %rd35, %r33;
and.b64 %rd36, %rd35, 2147483640;
or.b64 %rd37, %rd36, %rd32;
mul.lo.s32 %r44, %r1, %r24;
cvt.u64.u32 %rd38, %r44;
add.s64 %rd39, %rd37, %rd38;
mul.hi.u32 %r45, %r2, %r19;
add.s32 %r46, %r45, %r2;
shr.u32 %r47, %r46, %r20;
mul.lo.s32 %r48, %r47, %r21;
sub.s32 %r49, %r2, %r48;
mul.lo.s32 %r50, %r49, %r27;
cvt.u64.u32 %rd40, %r50;
add.s64 %rd41, %rd39, %rd40;
mul.lo.s64 %rd42, %rd41, 36;
cvta.to.global.u64 %rd43, %rd18;
add.s64 %rd112, %rd43, %rd42;
cvta.to.global.u64 %rd113, %rd17;
cvta.to.global.u64 %rd111, %rd21;
mov.f32 %f206, 0f00000000;

$L__BB153_3:
and.b32 %r88, %r4, 15;
setp.gt.u32 %p3, %r88, 7;
add.s64 %rd44, %rd113, %rd4;
ld.global.nc.u32 %r89, [%rd44+-16];
ld.global.nc.u32 %r90, [%rd44];
shr.u32 %r91, %r88, 1;
cvt.u64.u32 %rd45, %r91;
and.b64 %rd46, %rd45, 2;
or.b64 %rd48, %rd46, %rd33;
add.s64 %rd49, %rd113, %rd48;
ld.global.nc.u16 %rs1, [%rd49+4];
cvt.u32.u16 %r97, %rs1;
ld.global.nc.u16 %rs2, [%rd49+8];
cvt.u32.u16 %r98, %rs2;
ld.global.nc.u16 %rs3, [%rd49+12];
cvt.u32.u16 %r99, %rs3;
and.b32 %r100, %r97, 16191;
and.b32 %r101, %r99, 3855;
and.b32 %r102, %r97, 49344;
shr.u32 %r103, %r102, 2;
or.b32 %r104, %r101, %r103;
and.b32 %r105, %r98, 16191;
shr.u32 %r106, %r99, 4;
and.b32 %r107, %r106, 3855;
and.b32 %r108, %r98, 49344;
shr.u32 %r109, %r108, 2;
or.b32 %r110, %r107, %r109;
ld.global.nc.u32 %r7, [%rd112];

{.reg .f16 low,high;
mov.b32 {low,high},%r7;
cvt.f32.f16 %f42, low;}


cvt.u64.u32 %rd50, %r37;
and.b64 %rd51, %rd50, 12;
add.s64 %rd52, %rd112, %rd51;
ld.global.nc.u32 %r8, [%rd52+4];
ld.global.nc.u32 %r9, [%rd52+20];
ld.global.nc.u32 %r10, [%rd112+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r10;
cvt.f32.f16 %f43, low;}


ld.global.nc.u32 %r11, [%rd52+40];
ld.global.nc.u32 %r12, [%rd52+56];
add.s64 %rd53, %rd113, %rd33;
and.b32 %r54, %r89, 252645135;
and.b32 %r58, %r90, 252645135;
mov.u32 %r80, 0;

dp4a.s32.s32 %r53, %r54, %r8, %r80;


dp4a.s32.s32 %r57, %r58, %r9, %r53;

mov.u32 %r82, 16843009;

dp4a.s32.s32 %r61, %r82, %r8, %r80;


dp4a.s32.s32 %r65, %r82, %r9, %r61;

selp.b32 %r112, %r104, %r100, %p3;
cvt.u16.u32 %rs4, %r112;
and.b32 %r113, %r112, 63;
mul.lo.s32 %r114, %r57, %r113;
cvt.rn.f32.s32 %f46, %r114;
fma.rn.ftz.f32 %f47, %f42, %f46, 0f00000000;
selp.b32 %r115, %r110, %r105, %p3;
cvt.u16.u32 %rs5, %r115;
and.b32 %r116, %r115, 63;
mul.lo.s32 %r117, %r65, %r116;
cvt.rn.f32.s32 %f48, %r117;
fma.rn.ftz.f32 %f49, %f42, %f48, 0f00000000;
shr.u32 %r118, %r89, 4;
and.b32 %r70, %r118, 252645135;
shr.u32 %r119, %r90, 4;
and.b32 %r74, %r119, 252645135;

dp4a.s32.s32 %r69, %r70, %r11, %r80;


dp4a.s32.s32 %r73, %r74, %r12, %r69;


dp4a.s32.s32 %r77, %r82, %r11, %r80;


dp4a.s32.s32 %r81, %r82, %r12, %r77;

shr.u16 %rs6, %rs4, 8;
cvt.u32.u16 %r120, %rs6;
mul.lo.s32 %r121, %r73, %r120;
cvt.rn.f32.s32 %f50, %r121;
fma.rn.ftz.f32 %f51, %f43, %f50, %f47;
shr.u16 %rs7, %rs5, 8;
cvt.u32.u16 %r122, %rs7;
mul.lo.s32 %r123, %r81, %r122;
cvt.rn.f32.s32 %f52, %r123;
fma.rn.ftz.f32 %f53, %f43, %f52, %f49;
ld.global.nc.u32 %r85, [%rd53];

{.reg .f16 low,high;
mov.b32 {low,high},%r85;
cvt.f32.f16 %f44, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r85;
cvt.f32.f16 %f45, high;}


mul.ftz.f32 %f54, %f51, %f44;
mul.ftz.f32 %f55, %f53, %f45;
sub.ftz.f32 %f56, %f54, %f55;
add.ftz.f32 %f209, %f56, %f209;
setp.eq.s64 %p4, %rd21, 0;
@%p4 bra $L__BB153_5;

add.s64 %rd54, %rd111, %rd4;
ld.global.u32 %r162, [%rd54+-16];
ld.global.u32 %r163, [%rd54];
shr.u32 %r165, %r4, 4;
mad.lo.s32 %r168, %r41, %r23, %r165;
mad.lo.s32 %r169, %r3, %r26, %r168;
mul.wide.s32 %rd57, %r169, 144;
or.b64 %rd58, %rd46, %rd57;
add.s64 %rd59, %rd111, %rd58;
ld.global.u16 %r170, [%rd59+4];
ld.global.u16 %r171, [%rd59+8];
ld.global.u16 %r172, [%rd59+12];
and.b32 %r173, %r170, 16191;
and.b32 %r174, %r172, 3855;
and.b32 %r175, %r170, 49344;
shr.u32 %r176, %r175, 2;
or.b32 %r177, %r174, %r176;
and.b32 %r178, %r171, 16191;
shr.u32 %r179, %r172, 4;
and.b32 %r180, %r179, 3855;
and.b32 %r181, %r171, 49344;
shr.u32 %r182, %r181, 2;
or.b32 %r183, %r180, %r182;

{.reg .f16 low,high;
mov.b32 {low,high},%r7;
cvt.f32.f16 %f57, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r10;
cvt.f32.f16 %f58, low;}


add.s64 %rd60, %rd111, %rd57;
and.b32 %r127, %r162, 252645135;
and.b32 %r131, %r163, 252645135;

dp4a.s32.s32 %r126, %r127, %r8, %r80;


dp4a.s32.s32 %r130, %r131, %r9, %r126;


dp4a.s32.s32 %r134, %r82, %r8, %r80;


dp4a.s32.s32 %r138, %r82, %r9, %r134;

selp.b32 %r184, %r177, %r173, %p3;
cvt.u16.u32 %rs8, %r184;
and.b32 %r185, %r184, 63;
mul.lo.s32 %r186, %r130, %r185;
cvt.rn.f32.s32 %f61, %r186;
fma.rn.ftz.f32 %f62, %f57, %f61, 0f00000000;
selp.b32 %r187, %r183, %r178, %p3;
cvt.u16.u32 %rs9, %r187;
and.b32 %r188, %r187, 63;
mul.lo.s32 %r189, %r138, %r188;
cvt.rn.f32.s32 %f63, %r189;
fma.rn.ftz.f32 %f64, %f57, %f63, 0f00000000;
shr.u32 %r190, %r162, 4;
and.b32 %r143, %r190, 252645135;
shr.u32 %r191, %r163, 4;
and.b32 %r147, %r191, 252645135;

dp4a.s32.s32 %r142, %r143, %r11, %r80;


dp4a.s32.s32 %r146, %r147, %r12, %r142;


dp4a.s32.s32 %r150, %r82, %r11, %r80;


dp4a.s32.s32 %r154, %r82, %r12, %r150;

shr.u16 %rs10, %rs8, 8;
cvt.u32.u16 %r192, %rs10;
mul.lo.s32 %r193, %r146, %r192;
cvt.rn.f32.s32 %f65, %r193;
fma.rn.ftz.f32 %f66, %f58, %f65, %f62;
shr.u16 %rs11, %rs9, 8;
cvt.u32.u16 %r194, %rs11;
mul.lo.s32 %r195, %r154, %r194;
cvt.rn.f32.s32 %f67, %r195;
fma.rn.ftz.f32 %f68, %f58, %f67, %f64;
ld.global.u32 %r159, [%rd60];

{.reg .f16 low,high;
mov.b32 {low,high},%r159;
cvt.f32.f16 %f59, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r159;
cvt.f32.f16 %f60, high;}


mul.ftz.f32 %f69, %f66, %f59;
mul.ftz.f32 %f70, %f68, %f60;
sub.ftz.f32 %f71, %f69, %f70;
add.ftz.f32 %f207, %f71, %f207;

$L__BB153_5:
shr.u32 %r239, %r4, 4;
mad.lo.s32 %r240, %r3, %r26, %r239;
or.b32 %r243, %r41, 1;
mad.lo.s32 %r244, %r23, %r243, %r240;
mul.wide.s32 %rd62, %r244, 144;
add.s64 %rd63, %rd62, %rd31;
add.s64 %rd64, %rd113, %rd63;
ld.global.nc.u32 %r245, [%rd64+16];
ld.global.nc.u32 %r246, [%rd64+32];
or.b64 %rd67, %rd46, %rd62;
add.s64 %rd68, %rd113, %rd67;
ld.global.nc.u16 %rs12, [%rd68+4];
cvt.u32.u16 %r248, %rs12;
ld.global.nc.u16 %rs13, [%rd68+8];
cvt.u32.u16 %r249, %rs13;
ld.global.nc.u16 %rs14, [%rd68+12];
cvt.u32.u16 %r250, %rs14;
and.b32 %r251, %r248, 16191;
and.b32 %r252, %r250, 3855;
and.b32 %r253, %r248, 49344;
shr.u32 %r254, %r253, 2;
or.b32 %r255, %r252, %r254;
and.b32 %r256, %r249, 16191;
shr.u32 %r257, %r250, 4;
and.b32 %r258, %r257, 3855;
and.b32 %r259, %r249, 49344;
shr.u32 %r260, %r259, 2;
or.b32 %r261, %r258, %r260;

{.reg .f16 low,high;
mov.b32 {low,high},%r7;
cvt.f32.f16 %f72, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r10;
cvt.f32.f16 %f73, low;}


add.s64 %rd69, %rd113, %rd62;
and.b32 %r199, %r245, 252645135;
and.b32 %r203, %r246, 252645135;

dp4a.s32.s32 %r198, %r199, %r8, %r80;


dp4a.s32.s32 %r202, %r203, %r9, %r198;


dp4a.s32.s32 %r206, %r82, %r8, %r80;


dp4a.s32.s32 %r210, %r82, %r9, %r206;

selp.b32 %r262, %r255, %r251, %p3;
cvt.u16.u32 %rs15, %r262;
and.b32 %r263, %r262, 63;
mul.lo.s32 %r264, %r202, %r263;
cvt.rn.f32.s32 %f76, %r264;
fma.rn.ftz.f32 %f77, %f72, %f76, 0f00000000;
selp.b32 %r265, %r261, %r256, %p3;
cvt.u16.u32 %rs16, %r265;
and.b32 %r266, %r265, 63;
mul.lo.s32 %r267, %r210, %r266;
cvt.rn.f32.s32 %f78, %r267;
fma.rn.ftz.f32 %f79, %f72, %f78, 0f00000000;
shr.u32 %r268, %r245, 4;
and.b32 %r215, %r268, 252645135;
shr.u32 %r269, %r246, 4;
and.b32 %r219, %r269, 252645135;

dp4a.s32.s32 %r214, %r215, %r11, %r80;


dp4a.s32.s32 %r218, %r219, %r12, %r214;


dp4a.s32.s32 %r222, %r82, %r11, %r80;


dp4a.s32.s32 %r226, %r82, %r12, %r222;

shr.u16 %rs17, %rs15, 8;
cvt.u32.u16 %r270, %rs17;
mul.lo.s32 %r271, %r218, %r270;
cvt.rn.f32.s32 %f80, %r271;
fma.rn.ftz.f32 %f81, %f73, %f80, %f77;
shr.u16 %rs18, %rs16, 8;
cvt.u32.u16 %r272, %rs18;
mul.lo.s32 %r273, %r226, %r272;
cvt.rn.f32.s32 %f82, %r273;
fma.rn.ftz.f32 %f83, %f73, %f82, %f79;
ld.global.nc.u32 %r230, [%rd69];

{.reg .f16 low,high;
mov.b32 {low,high},%r230;
cvt.f32.f16 %f74, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r230;
cvt.f32.f16 %f75, high;}


mul.ftz.f32 %f84, %f81, %f74;
mul.ftz.f32 %f85, %f83, %f75;
sub.ftz.f32 %f86, %f84, %f85;
add.ftz.f32 %f208, %f86, %f208;
@%p4 bra $L__BB153_7;

add.s64 %rd73, %rd111, %rd63;
ld.global.u32 %r323, [%rd73+16];
ld.global.u32 %r324, [%rd73+32];
add.s64 %rd77, %rd111, %rd67;
ld.global.u16 %r326, [%rd77+4];
ld.global.u16 %r327, [%rd77+8];
ld.global.u16 %r328, [%rd77+12];
and.b32 %r329, %r326, 16191;
and.b32 %r330, %r328, 3855;
and.b32 %r331, %r326, 49344;
shr.u32 %r332, %r331, 2;
or.b32 %r333, %r330, %r332;
and.b32 %r334, %r327, 16191;
shr.u32 %r335, %r328, 4;
and.b32 %r336, %r335, 3855;
and.b32 %r337, %r327, 49344;
shr.u32 %r338, %r337, 2;
or.b32 %r339, %r336, %r338;

{.reg .f16 low,high;
mov.b32 {low,high},%r7;
cvt.f32.f16 %f87, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r10;
cvt.f32.f16 %f88, low;}


add.s64 %rd78, %rd111, %rd62;
and.b32 %r277, %r323, 252645135;
and.b32 %r281, %r324, 252645135;
mov.u32 %r303, 0;

dp4a.s32.s32 %r276, %r277, %r8, %r303;


dp4a.s32.s32 %r280, %r281, %r9, %r276;

mov.u32 %r305, 16843009;

dp4a.s32.s32 %r284, %r305, %r8, %r303;


dp4a.s32.s32 %r288, %r305, %r9, %r284;

selp.b32 %r340, %r333, %r329, %p3;
cvt.u16.u32 %rs19, %r340;
and.b32 %r341, %r340, 63;
mul.lo.s32 %r342, %r280, %r341;
cvt.rn.f32.s32 %f91, %r342;
fma.rn.ftz.f32 %f92, %f87, %f91, 0f00000000;
selp.b32 %r343, %r339, %r334, %p3;
cvt.u16.u32 %rs20, %r343;
and.b32 %r344, %r343, 63;
mul.lo.s32 %r345, %r288, %r344;
cvt.rn.f32.s32 %f93, %r345;
fma.rn.ftz.f32 %f94, %f87, %f93, 0f00000000;
shr.u32 %r346, %r323, 4;
and.b32 %r293, %r346, 252645135;
shr.u32 %r347, %r324, 4;
and.b32 %r297, %r347, 252645135;

dp4a.s32.s32 %r292, %r293, %r11, %r303;


dp4a.s32.s32 %r296, %r297, %r12, %r292;


dp4a.s32.s32 %r300, %r305, %r11, %r303;


dp4a.s32.s32 %r304, %r305, %r12, %r300;

shr.u16 %rs21, %rs19, 8;
cvt.u32.u16 %r348, %rs21;
mul.lo.s32 %r349, %r296, %r348;
cvt.rn.f32.s32 %f95, %r349;
fma.rn.ftz.f32 %f96, %f88, %f95, %f92;
shr.u16 %rs22, %rs20, 8;
cvt.u32.u16 %r350, %rs22;
mul.lo.s32 %r351, %r304, %r350;
cvt.rn.f32.s32 %f97, %r351;
fma.rn.ftz.f32 %f98, %f88, %f97, %f94;
ld.global.u32 %r309, [%rd78];

{.reg .f16 low,high;
mov.b32 {low,high},%r309;
cvt.f32.f16 %f89, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r309;
cvt.f32.f16 %f90, high;}


mul.ftz.f32 %f99, %f96, %f89;
mul.ftz.f32 %f100, %f98, %f90;
sub.ftz.f32 %f101, %f99, %f100;
add.ftz.f32 %f206, %f101, %f206;

$L__BB153_7:
add.s64 %rd113, %rd113, 288;
add.s64 %rd112, %rd112, 576;
add.s64 %rd111, %rd111, 288;
add.s32 %r445, %r445, 2;
setp.lt.u32 %p9, %r445, %r32;
@%p9 bra $L__BB153_3;

add.u64 %rd99, %SPL, 8;
add.u64 %rd97, %SPL, 0;
st.local.v2.f32 [%rd97], {%f209, %f208};
st.local.v2.f32 [%rd99], {%f207, %f206};

$L__BB153_9:
mov.b32 %r14, %f209;
setp.ne.s64 %p10, %rd21, 0;
@%p10 bra $L__BB153_11;
bra.uni $L__BB153_10;

$L__BB153_11:
add.u64 %rd109, %SPL, 8;
add.u64 %rd107, %SPL, 0;
mov.u32 %r379, 31;
mov.u32 %r380, 16;
mov.u32 %r381, -1;
shfl.sync.bfly.b32 %r382|%p21, %r14, %r380, %r379, %r381;
mov.b32 %f122, %r382;
add.ftz.f32 %f123, %f209, %f122;
mov.b32 %r383, %f123;
mov.u32 %r384, 8;
shfl.sync.bfly.b32 %r385|%p22, %r383, %r384, %r379, %r381;
mov.b32 %f124, %r385;
add.ftz.f32 %f125, %f123, %f124;
mov.b32 %r386, %f125;
mov.u32 %r387, 4;
shfl.sync.bfly.b32 %r388|%p23, %r386, %r387, %r379, %r381;
mov.b32 %f126, %r388;
add.ftz.f32 %f127, %f125, %f126;
mov.b32 %r389, %f127;
mov.u32 %r390, 2;
shfl.sync.bfly.b32 %r391|%p24, %r389, %r390, %r379, %r381;
mov.b32 %f128, %r391;
add.ftz.f32 %f129, %f127, %f128;
mov.b32 %r392, %f129;
mov.u32 %r393, 1;
shfl.sync.bfly.b32 %r394|%p25, %r392, %r393, %r379, %r381;
mov.b32 %f130, %r394;
add.ftz.f32 %f131, %f129, %f130;
st.local.f32 [%rd107], %f131;
mov.b32 %r395, %f207;
shfl.sync.bfly.b32 %r396|%p26, %r395, %r380, %r379, %r381;
mov.b32 %f132, %r396;
add.ftz.f32 %f133, %f207, %f132;
mov.b32 %r397, %f133;
shfl.sync.bfly.b32 %r398|%p27, %r397, %r384, %r379, %r381;
mov.b32 %f134, %r398;
add.ftz.f32 %f135, %f133, %f134;
mov.b32 %r399, %f135;
shfl.sync.bfly.b32 %r400|%p28, %r399, %r387, %r379, %r381;
mov.b32 %f136, %r400;
add.ftz.f32 %f137, %f135, %f136;
mov.b32 %r401, %f137;
shfl.sync.bfly.b32 %r402|%p29, %r401, %r390, %r379, %r381;
mov.b32 %f138, %r402;
add.ftz.f32 %f139, %f137, %f138;
mov.b32 %r403, %f139;
shfl.sync.bfly.b32 %r404|%p30, %r403, %r393, %r379, %r381;
mov.b32 %f140, %r404;
add.ftz.f32 %f141, %f139, %f140;
st.local.f32 [%rd109], %f141;
mov.b32 %r405, %f208;
shfl.sync.bfly.b32 %r406|%p31, %r405, %r380, %r379, %r381;
mov.b32 %f142, %r406;
add.ftz.f32 %f143, %f208, %f142;
mov.b32 %r407, %f143;
shfl.sync.bfly.b32 %r408|%p32, %r407, %r384, %r379, %r381;
mov.b32 %f144, %r408;
add.ftz.f32 %f145, %f143, %f144;
mov.b32 %r409, %f145;
shfl.sync.bfly.b32 %r410|%p33, %r409, %r387, %r379, %r381;
mov.b32 %f146, %r410;
add.ftz.f32 %f147, %f145, %f146;
mov.b32 %r411, %f147;
shfl.sync.bfly.b32 %r412|%p34, %r411, %r390, %r379, %r381;
mov.b32 %f148, %r412;
add.ftz.f32 %f149, %f147, %f148;
mov.b32 %r413, %f149;
shfl.sync.bfly.b32 %r414|%p35, %r413, %r393, %r379, %r381;
mov.b32 %f150, %r414;
add.ftz.f32 %f151, %f149, %f150;
st.local.f32 [%rd107+4], %f151;
mov.b32 %r415, %f206;
shfl.sync.bfly.b32 %r416|%p36, %r415, %r380, %r379, %r381;
mov.b32 %f152, %r416;
add.ftz.f32 %f153, %f206, %f152;
mov.b32 %r417, %f153;
shfl.sync.bfly.b32 %r418|%p37, %r417, %r384, %r379, %r381;
mov.b32 %f154, %r418;
add.ftz.f32 %f155, %f153, %f154;
mov.b32 %r419, %f155;
shfl.sync.bfly.b32 %r420|%p38, %r419, %r387, %r379, %r381;
mov.b32 %f156, %r420;
add.ftz.f32 %f157, %f155, %f156;
mov.b32 %r421, %f157;
shfl.sync.bfly.b32 %r422|%p39, %r421, %r390, %r379, %r381;
mov.b32 %f158, %r422;
add.ftz.f32 %f159, %f157, %f158;
mov.b32 %r423, %f159;
shfl.sync.bfly.b32 %r424|%p40, %r423, %r393, %r379, %r381;
mov.b32 %f160, %r424;
add.ftz.f32 %f161, %f159, %f160;
st.local.f32 [%rd109+4], %f161;
bra.uni $L__BB153_12;

$L__BB153_10:
add.u64 %rd101, %SPL, 0;
mov.u32 %r353, 31;
mov.u32 %r354, 16;
mov.u32 %r355, -1;
shfl.sync.bfly.b32 %r356|%p11, %r14, %r354, %r353, %r355;
mov.b32 %f102, %r356;
add.ftz.f32 %f103, %f209, %f102;
mov.b32 %r357, %f103;
mov.u32 %r358, 8;
shfl.sync.bfly.b32 %r359|%p12, %r357, %r358, %r353, %r355;
mov.b32 %f104, %r359;
add.ftz.f32 %f105, %f103, %f104;
mov.b32 %r360, %f105;
mov.u32 %r361, 4;
shfl.sync.bfly.b32 %r362|%p13, %r360, %r361, %r353, %r355;
mov.b32 %f106, %r362;
add.ftz.f32 %f107, %f105, %f106;
mov.b32 %r363, %f107;
mov.u32 %r364, 2;
shfl.sync.bfly.b32 %r365|%p14, %r363, %r364, %r353, %r355;
mov.b32 %f108, %r365;
add.ftz.f32 %f109, %f107, %f108;
mov.b32 %r366, %f109;
mov.u32 %r367, 1;
shfl.sync.bfly.b32 %r368|%p15, %r366, %r367, %r353, %r355;
mov.b32 %f110, %r368;
add.ftz.f32 %f111, %f109, %f110;
st.local.f32 [%rd101], %f111;
mov.b32 %r369, %f208;
shfl.sync.bfly.b32 %r370|%p16, %r369, %r354, %r353, %r355;
mov.b32 %f112, %r370;
add.ftz.f32 %f113, %f208, %f112;
mov.b32 %r371, %f113;
shfl.sync.bfly.b32 %r372|%p17, %r371, %r358, %r353, %r355;
mov.b32 %f114, %r372;
add.ftz.f32 %f115, %f113, %f114;
mov.b32 %r373, %f115;
shfl.sync.bfly.b32 %r374|%p18, %r373, %r361, %r353, %r355;
mov.b32 %f116, %r374;
add.ftz.f32 %f117, %f115, %f116;
mov.b32 %r375, %f117;
shfl.sync.bfly.b32 %r376|%p19, %r375, %r364, %r353, %r355;
mov.b32 %f118, %r376;
add.ftz.f32 %f119, %f117, %f118;
mov.b32 %r377, %f119;
shfl.sync.bfly.b32 %r378|%p20, %r377, %r367, %r353, %r355;
mov.b32 %f120, %r378;
add.ftz.f32 %f121, %f119, %f120;
st.local.f32 [%rd101+4], %f121;

$L__BB153_12:
ld.param.u32 %r434, [_Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7];
setp.gt.u32 %p41, %r4, 1;
mov.u32 %r426, %ctaid.x;
shl.b32 %r427, %r426, 1;
add.s32 %r15, %r427, %r4;
setp.ge.u32 %p42, %r15, %r434;
or.pred %p43, %p41, %p42;
@%p43 bra $L__BB153_29;

add.u64 %rd103, %SPL, 0;
ld.param.u64 %rd92, [_Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3];
ld.param.u32 %r435, [_Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13];
cvt.u64.u32 %rd15, %r4;
mul.wide.u32 %rd79, %r4, 4;
add.s64 %rd80, %rd103, %rd79;
ld.local.f32 %f212, [%rd80];
mad.lo.s32 %r16, %r3, %r435, %r15;
setp.eq.s64 %p44, %rd92, 0;
@%p44 bra $L__BB153_15;

ld.param.u64 %rd93, [_Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3];
cvta.to.global.u64 %rd81, %rd93;
mul.wide.u32 %rd82, %r16, 4;
add.s64 %rd83, %rd81, %rd82;
ld.global.f32 %f162, [%rd83];
add.ftz.f32 %f212, %f212, %f162;

$L__BB153_15:
setp.eq.s64 %p45, %rd21, 0;
@%p45 bra $L__BB153_28;

add.u64 %rd105, %SPL, 8;
ld.param.u64 %rd95, [_Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+16];
shl.b64 %rd84, %rd15, 2;
add.s64 %rd85, %rd105, %rd84;
ld.local.f32 %f24, [%rd85];
setp.eq.s64 %p46, %rd95, 0;
@%p46 bra $L__BB153_18;

ld.param.u64 %rd96, [_Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+16];
cvta.to.global.u64 %rd86, %rd96;
mul.wide.u32 %rd87, %r16, 4;
add.s64 %rd88, %rd86, %rd87;
ld.global.f32 %f163, [%rd88];
add.ftz.f32 %f24, %f24, %f163;

$L__BB153_18:
ld.param.u32 %r440, [_Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.gt.s32 %p47, %r440, 2;
@%p47 bra $L__BB153_22;

ld.param.u32 %r443, [_Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.eq.s32 %p50, %r443, 1;
@%p50 bra $L__BB153_26;

ld.param.u32 %r444, [_Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.eq.s32 %p51, %r444, 2;
@%p51 bra $L__BB153_21;
bra.uni $L__BB153_27;

$L__BB153_21:
mul.ftz.f32 %f183, %f24, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f184, %f183;
add.ftz.f32 %f185, %f184, 0f3F800000;
div.approx.ftz.f32 %f186, %f24, %f185;
mul.ftz.f32 %f212, %f212, %f186;
bra.uni $L__BB153_28;

$L__BB153_22:
ld.param.u32 %r441, [_Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.eq.s32 %p48, %r441, 3;
@%p48 bra $L__BB153_25;

ld.param.u32 %r442, [_Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.ne.s32 %p49, %r442, 6;
@%p49 bra $L__BB153_27;

ld.param.f32 %f195, [_Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+44];
min.ftz.f32 %f164, %f24, %f195;
neg.ftz.f32 %f165, %f195;
min.ftz.f32 %f166, %f212, %f195;
max.ftz.f32 %f167, %f166, %f165;
mul.ftz.f32 %f168, %f164, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f169, %f168;
add.ftz.f32 %f170, %f169, 0f3F800000;
div.approx.ftz.f32 %f171, %f164, %f170;
mul.ftz.f32 %f212, %f167, %f171;
bra.uni $L__BB153_28;

$L__BB153_27:
mul.ftz.f32 %f212, %f212, %f24;
bra.uni $L__BB153_28;

$L__BB153_26:
mul.ftz.f32 %f187, %f24, 0f3F000000;
mul.ftz.f32 %f188, %f24, 0f3D372713;
fma.rn.ftz.f32 %f189, %f24, %f188, 0f3F800000;
mul.ftz.f32 %f190, %f24, 0f3F4C422A;
mul.ftz.f32 %f191, %f190, %f189;
tanh.approx.f32 %f192, %f191;
add.ftz.f32 %f193, %f192, 0f3F800000;
mul.ftz.f32 %f194, %f187, %f193;
mul.ftz.f32 %f212, %f212, %f194;
bra.uni $L__BB153_28;

$L__BB153_25:
mov.f32 %f172, 0f40E00000;
min.ftz.f32 %f173, %f24, %f172;
min.ftz.f32 %f174, %f212, %f172;
mov.f32 %f175, 0fC0E00000;
max.ftz.f32 %f176, %f174, %f175;
mul.ftz.f32 %f177, %f173, 0f3FD9DB23;
mul.ftz.f32 %f178, %f177, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f179, %f178;
add.ftz.f32 %f180, %f179, 0f3F800000;
div.approx.ftz.f32 %f181, %f173, %f180;
add.ftz.f32 %f182, %f176, 0f3F800000;
mul.ftz.f32 %f212, %f182, %f181;

$L__BB153_28:
ld.param.u64 %rd94, [_Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4];
ld.param.u32 %r439, [_Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10];
ld.param.u32 %r438, [_Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13];
mov.u32 %r437, %ctaid.y;
mov.u32 %r436, %tid.y;
mad.lo.s32 %r432, %r436, %r439, %r15;
mad.lo.s32 %r433, %r437, %r438, %r432;
cvta.to.global.u64 %rd89, %rd94;
mul.wide.u32 %rd90, %r433, 4;
add.s64 %rd91, %rd89, %rd90;
st.global.f32 [%rd91], %f212;

$L__BB153_29:
ret;

}
.entry _Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj(
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2,
.param .align 8 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3[48],
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5,
.param .align 4 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6[12],
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15
)
.maxntid 256, 1, 1
.minnctapersm 1
{
.local .align 8 .b8 __local_depot154[8];
.reg .b64 %SP;
.reg .b64 %SPL;
.reg .pred %p<18>;
.reg .b16 %rs<15>;
.reg .f32 %f<66>;
.reg .b32 %r<209>;
.reg .b64 %rd<56>;


mov.u64 %SPL, __local_depot154;
ld.param.u64 %rd16, [_Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0];
ld.param.u64 %rd17, [_Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1];
ld.param.u64 %rd18, [_Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2];
ld.param.u64 %rd19, [_Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4];
ld.param.u32 %r20, [_Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5];
ld.param.u32 %r12, [_Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7];
ld.param.u32 %r13, [_Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8];
ld.param.u32 %r14, [_Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9];
ld.param.u32 %r15, [_Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10];
ld.param.u32 %r16, [_Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11];
ld.param.u32 %r17, [_Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12];
ld.param.u32 %r18, [_Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13];
ld.param.u32 %r21, [_Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14];
ld.param.u32 %r19, [_Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15];
ld.param.u32 %r11, [_Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+8];
ld.param.u32 %r10, [_Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+4];
ld.param.u32 %r9, [_Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6];
add.u64 %rd1, %SPL, 0;
shr.u32 %r1, %r20, 8;
mov.u32 %r2, %tid.y;
setp.ge.u32 %p1, %r2, %r21;
@%p1 bra $L__BB154_7;

mov.f32 %f64, 0f00000000;
st.local.v2.f32 [%rd1], {%f64, %f64};
mov.u32 %r3, %tid.x;
shr.u32 %r208, %r3, 4;
setp.ge.u32 %p2, %r208, %r1;
mov.f32 %f65, %f64;
@%p2 bra $L__BB154_5;

mov.u32 %r22, %ctaid.x;
shl.b32 %r23, %r22, 1;
mov.u32 %r24, %ctaid.y;
mad.lo.s32 %r25, %r2, %r19, %r24;
cvta.to.global.u64 %rd21, %rd18;
mul.wide.u32 %rd22, %r25, 4;
add.s64 %rd23, %rd21, %rd22;
ld.global.nc.u32 %r26, [%rd23];
mul.lo.s32 %r27, %r26, %r16;
mul.lo.s32 %r28, %r2, %r14;
cvt.u64.u32 %rd24, %r28;
and.b32 %r5, %r3, 15;
shr.u32 %r29, %r5, 2;
shl.b32 %r30, %r29, 1;
shl.b32 %r31, %r29, 5;
shl.b32 %r32, %r3, 2;
and.b32 %r33, %r32, 12;
or.b32 %r34, %r31, %r33;
cvt.u64.u32 %rd25, %r34;
cvt.u64.u32 %rd26, %r30;
mad.lo.s32 %r35, %r23, %r13, %r208;
add.s32 %r36, %r35, %r27;
mul.wide.s32 %rd3, %r36, 144;
add.s64 %rd27, %rd3, %rd25;
add.s64 %rd4, %rd27, 32;
shr.u32 %r37, %r3, 1;
cvt.u64.u32 %rd28, %r37;
and.b64 %rd29, %rd28, 2147483640;
or.b64 %rd30, %rd29, %rd26;
add.s64 %rd31, %rd30, %rd24;
mul.hi.u32 %r38, %r24, %r9;
add.s32 %r39, %r38, %r24;
shr.u32 %r40, %r39, %r10;
mul.lo.s32 %r41, %r40, %r11;
sub.s32 %r42, %r24, %r41;
mul.lo.s32 %r43, %r42, %r17;
cvt.u64.u32 %rd32, %r43;
add.s64 %rd33, %rd31, %rd32;
mul.lo.s64 %rd34, %rd33, 36;
cvta.to.global.u64 %rd35, %rd17;
add.s64 %rd54, %rd35, %rd34;
and.b64 %rd36, %rd26, 2;
or.b64 %rd37, %rd36, %rd3;
or.b64 %rd6, %rd37, 12;
mul.wide.u32 %rd38, %r3, 4;
and.b64 %rd39, %rd38, 12;
add.s64 %rd7, %rd39, 40;
or.b32 %r44, %r23, 1;
add.s32 %r45, %r208, %r27;
mad.lo.s32 %r46, %r13, %r44, %r45;
mul.wide.s32 %rd8, %r46, 144;
add.s64 %rd40, %rd8, %rd25;
add.s64 %rd9, %rd40, 32;
or.b64 %rd41, %rd36, %rd8;
or.b64 %rd10, %rd41, 8;
cvta.to.global.u64 %rd55, %rd16;
mov.f32 %f64, 0f00000000;

$L__BB154_3:
add.s64 %rd42, %rd55, %rd4;
ld.global.nc.u32 %r119, [%rd42+-16];
ld.global.nc.u32 %r120, [%rd42];
add.s64 %rd43, %rd55, %rd6;
ld.global.nc.u16 %rs1, [%rd43+-8];
cvt.u32.u16 %r121, %rs1;
ld.global.nc.u16 %rs2, [%rd43+-4];
cvt.u32.u16 %r122, %rs2;
ld.global.nc.u16 %rs3, [%rd43];
cvt.u32.u16 %r123, %rs3;
and.b32 %r124, %r121, 16191;
and.b32 %r125, %r123, 3855;
and.b32 %r126, %r121, 49344;
shr.u32 %r127, %r126, 2;
or.b32 %r128, %r125, %r127;
setp.gt.u32 %p3, %r5, 7;
and.b32 %r129, %r122, 16191;
shr.u32 %r130, %r123, 4;
and.b32 %r131, %r130, 3855;
and.b32 %r132, %r122, 49344;
shr.u32 %r133, %r132, 2;
or.b32 %r134, %r131, %r133;
ld.global.nc.u32 %r47, [%rd54];

{.reg .f16 low,high;
mov.b32 {low,high},%r47;
cvt.f32.f16 %f11, low;}


add.s64 %rd44, %rd54, %rd7;
ld.global.nc.u32 %r51, [%rd44+-36];
ld.global.nc.u32 %r55, [%rd44+-20];
ld.global.nc.u32 %r48, [%rd54+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r48;
cvt.f32.f16 %f12, low;}


ld.global.nc.u32 %r67, [%rd44];
ld.global.nc.u32 %r71, [%rd44+16];
add.s64 %rd45, %rd55, %rd3;
and.b32 %r50, %r119, 252645135;
and.b32 %r54, %r120, 252645135;
mov.u32 %r112, 0;

dp4a.s32.s32 %r49, %r50, %r51, %r112;


dp4a.s32.s32 %r53, %r54, %r55, %r49;

mov.u32 %r114, 16843009;

dp4a.s32.s32 %r57, %r114, %r51, %r112;


dp4a.s32.s32 %r61, %r114, %r55, %r57;

selp.b32 %r135, %r128, %r124, %p3;
cvt.u16.u32 %rs4, %r135;
and.b32 %r136, %r135, 63;
mul.lo.s32 %r137, %r53, %r136;
cvt.rn.f32.s32 %f19, %r137;
fma.rn.ftz.f32 %f20, %f11, %f19, 0f00000000;
selp.b32 %r138, %r134, %r129, %p3;
cvt.u16.u32 %rs5, %r138;
and.b32 %r139, %r138, 63;
mul.lo.s32 %r140, %r61, %r139;
cvt.rn.f32.s32 %f21, %r140;
fma.rn.ftz.f32 %f22, %f11, %f21, 0f00000000;
shr.u32 %r141, %r119, 4;
and.b32 %r66, %r141, 252645135;
shr.u32 %r142, %r120, 4;
and.b32 %r70, %r142, 252645135;

dp4a.s32.s32 %r65, %r66, %r67, %r112;


dp4a.s32.s32 %r69, %r70, %r71, %r65;


dp4a.s32.s32 %r73, %r114, %r67, %r112;


dp4a.s32.s32 %r77, %r114, %r71, %r73;

shr.u16 %rs6, %rs4, 8;
cvt.u32.u16 %r143, %rs6;
mul.lo.s32 %r144, %r69, %r143;
cvt.rn.f32.s32 %f23, %r144;
fma.rn.ftz.f32 %f24, %f12, %f23, %f20;
shr.u16 %rs7, %rs5, 8;
cvt.u32.u16 %r145, %rs7;
mul.lo.s32 %r146, %r77, %r145;
cvt.rn.f32.s32 %f25, %r146;
fma.rn.ftz.f32 %f26, %f12, %f25, %f22;
ld.global.nc.u32 %r81, [%rd45];

{.reg .f16 low,high;
mov.b32 {low,high},%r81;
cvt.f32.f16 %f13, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r81;
cvt.f32.f16 %f14, high;}


mul.ftz.f32 %f27, %f24, %f13;
mul.ftz.f32 %f28, %f26, %f14;
sub.ftz.f32 %f29, %f27, %f28;
add.ftz.f32 %f65, %f29, %f65;
add.s64 %rd46, %rd55, %rd9;
ld.global.nc.u32 %r147, [%rd46+-16];
ld.global.nc.u32 %r148, [%rd46];
add.s64 %rd47, %rd55, %rd10;
ld.global.nc.u16 %rs8, [%rd47+-4];
cvt.u32.u16 %r149, %rs8;
ld.global.nc.u16 %rs9, [%rd47];
cvt.u32.u16 %r150, %rs9;
ld.global.nc.u16 %rs10, [%rd47+4];
cvt.u32.u16 %r151, %rs10;
and.b32 %r152, %r149, 16191;
and.b32 %r153, %r151, 3855;
and.b32 %r154, %r149, 49344;
shr.u32 %r155, %r154, 2;
or.b32 %r156, %r153, %r155;
and.b32 %r157, %r150, 16191;
shr.u32 %r158, %r151, 4;
and.b32 %r159, %r158, 3855;
and.b32 %r160, %r150, 49344;
shr.u32 %r161, %r160, 2;
or.b32 %r162, %r159, %r161;

{.reg .f16 low,high;
mov.b32 {low,high},%r47;
cvt.f32.f16 %f15, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r48;
cvt.f32.f16 %f16, low;}


add.s64 %rd48, %rd55, %rd8;
and.b32 %r86, %r147, 252645135;
and.b32 %r90, %r148, 252645135;

dp4a.s32.s32 %r85, %r86, %r51, %r112;


dp4a.s32.s32 %r89, %r90, %r55, %r85;


dp4a.s32.s32 %r93, %r114, %r51, %r112;


dp4a.s32.s32 %r97, %r114, %r55, %r93;

selp.b32 %r163, %r156, %r152, %p3;
cvt.u16.u32 %rs11, %r163;
and.b32 %r164, %r163, 63;
mul.lo.s32 %r165, %r89, %r164;
cvt.rn.f32.s32 %f30, %r165;
fma.rn.ftz.f32 %f31, %f15, %f30, 0f00000000;
selp.b32 %r166, %r162, %r157, %p3;
cvt.u16.u32 %rs12, %r166;
and.b32 %r167, %r166, 63;
mul.lo.s32 %r168, %r97, %r167;
cvt.rn.f32.s32 %f32, %r168;
fma.rn.ftz.f32 %f33, %f15, %f32, 0f00000000;
shr.u32 %r169, %r147, 4;
and.b32 %r102, %r169, 252645135;
shr.u32 %r170, %r148, 4;
and.b32 %r106, %r170, 252645135;

dp4a.s32.s32 %r101, %r102, %r67, %r112;


dp4a.s32.s32 %r105, %r106, %r71, %r101;


dp4a.s32.s32 %r109, %r114, %r67, %r112;


dp4a.s32.s32 %r113, %r114, %r71, %r109;

shr.u16 %rs13, %rs11, 8;
cvt.u32.u16 %r171, %rs13;
mul.lo.s32 %r172, %r105, %r171;
cvt.rn.f32.s32 %f34, %r172;
fma.rn.ftz.f32 %f35, %f16, %f34, %f31;
shr.u16 %rs14, %rs12, 8;
cvt.u32.u16 %r173, %rs14;
mul.lo.s32 %r174, %r113, %r173;
cvt.rn.f32.s32 %f36, %r174;
fma.rn.ftz.f32 %f37, %f16, %f36, %f33;
ld.global.nc.u32 %r117, [%rd48];

{.reg .f16 low,high;
mov.b32 {low,high},%r117;
cvt.f32.f16 %f17, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r117;
cvt.f32.f16 %f18, high;}


mul.ftz.f32 %f38, %f35, %f17;
mul.ftz.f32 %f39, %f37, %f18;
sub.ftz.f32 %f40, %f38, %f39;
add.ftz.f32 %f64, %f40, %f64;
add.s64 %rd55, %rd55, 288;
add.s64 %rd54, %rd54, 576;
add.s32 %r208, %r208, 2;
setp.lt.u32 %p4, %r208, %r1;
@%p4 bra $L__BB154_3;

st.local.v2.f32 [%rd1], {%f65, %f64};

$L__BB154_5:
mov.b32 %r175, %f65;
mov.u32 %r176, 31;
mov.u32 %r177, 16;
mov.u32 %r178, -1;
shfl.sync.bfly.b32 %r179|%p5, %r175, %r177, %r176, %r178;
mov.b32 %f41, %r179;
add.ftz.f32 %f42, %f65, %f41;
mov.b32 %r180, %f42;
mov.u32 %r181, 8;
shfl.sync.bfly.b32 %r182|%p6, %r180, %r181, %r176, %r178;
mov.b32 %f43, %r182;
add.ftz.f32 %f44, %f42, %f43;
mov.b32 %r183, %f44;
mov.u32 %r184, 4;
shfl.sync.bfly.b32 %r185|%p7, %r183, %r184, %r176, %r178;
mov.b32 %f45, %r185;
add.ftz.f32 %f46, %f44, %f45;
mov.b32 %r186, %f46;
mov.u32 %r187, 2;
shfl.sync.bfly.b32 %r188|%p8, %r186, %r187, %r176, %r178;
mov.b32 %f47, %r188;
add.ftz.f32 %f48, %f46, %f47;
mov.b32 %r189, %f48;
mov.u32 %r190, 1;
shfl.sync.bfly.b32 %r191|%p9, %r189, %r190, %r176, %r178;
mov.b32 %f49, %r191;
add.ftz.f32 %f50, %f48, %f49;
st.local.f32 [%rd1], %f50;
mov.b32 %r192, %f64;
shfl.sync.bfly.b32 %r193|%p10, %r192, %r177, %r176, %r178;
mov.b32 %f51, %r193;
add.ftz.f32 %f52, %f64, %f51;
mov.b32 %r194, %f52;
shfl.sync.bfly.b32 %r195|%p11, %r194, %r181, %r176, %r178;
mov.b32 %f53, %r195;
add.ftz.f32 %f54, %f52, %f53;
mov.b32 %r196, %f54;
shfl.sync.bfly.b32 %r197|%p12, %r196, %r184, %r176, %r178;
mov.b32 %f55, %r197;
add.ftz.f32 %f56, %f54, %f55;
mov.b32 %r198, %f56;
shfl.sync.bfly.b32 %r199|%p13, %r198, %r187, %r176, %r178;
mov.b32 %f57, %r199;
add.ftz.f32 %f58, %f56, %f57;
mov.b32 %r200, %f58;
shfl.sync.bfly.b32 %r201|%p14, %r200, %r190, %r176, %r178;
mov.b32 %f59, %r201;
add.ftz.f32 %f60, %f58, %f59;
st.local.f32 [%rd1+4], %f60;
mov.u32 %r202, %ctaid.x;
shl.b32 %r203, %r202, 1;
add.s32 %r8, %r203, %r3;
setp.ge.u32 %p15, %r8, %r12;
setp.gt.u32 %p16, %r3, 1;
or.pred %p17, %p16, %p15;
@%p17 bra $L__BB154_7;

mul.wide.u32 %rd49, %r3, 4;
add.s64 %rd50, %rd1, %rd49;
ld.local.f32 %f61, [%rd50];
mov.u32 %r204, %ctaid.y;
mad.lo.s32 %r206, %r2, %r15, %r8;
mad.lo.s32 %r207, %r204, %r18, %r206;
cvta.to.global.u64 %rd51, %rd19;
mul.wide.u32 %rd52, %r207, 4;
add.s64 %rd53, %rd51, %rd52;
st.global.f32 [%rd53], %f61;

$L__BB154_7:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<328>;
.reg .b16 %rs<45>;
.reg .f32 %f<1093>;
.reg .b32 %r<1369>;
.reg .b64 %rd<115>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate[1536];

ld.param.u64 %rd36, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd37, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd38, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r47, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r48, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r52, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r53, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r54, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r58, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r59, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r60, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r56, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r55, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r50, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r49, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r46, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r45, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
ld.param.u64 %rd41, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+16];
ld.param.u64 %rd40, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+8];
ld.param.u64 %rd39, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3];
mov.u32 %r1, %tid.y;
shl.b32 %r61, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r61, %r2;
mov.u32 %r62, %ctaid.x;
shl.b32 %r4, %r62, 2;
mov.u32 %r5, %ctaid.y;
setp.eq.s64 %p1, %rd38, 0;
@%p1 bra $L__BB155_2;

cvta.to.global.u64 %rd45, %rd38;
mul.wide.u32 %rd46, %r5, 4;
add.s64 %rd47, %rd45, %rd46;
ld.global.nc.u32 %r1365, [%rd47];
mul.hi.u32 %r63, %r5, %r44;
add.s32 %r64, %r63, %r5;
shr.u32 %r65, %r64, %r45;
mul.lo.s32 %r66, %r65, %r46;
sub.s32 %r1366, %r5, %r66;
bra.uni $L__BB155_3;

$L__BB155_2:
mul.hi.u32 %r67, %r5, %r49;
add.s32 %r68, %r67, %r5;
shr.u32 %r1365, %r68, %r50;
mov.u32 %r1366, %r5;

$L__BB155_3:
mov.u32 %r11, %ctaid.z;
selp.b32 %r12, %r5, %r1365, %p1;
setp.ne.s32 %p3, %r1, 0;
setp.gt.u32 %p4, %r2, 3;
or.pred %p5, %p4, %p3;
add.s32 %r69, %r4, %r2;
setp.ge.u32 %p6, %r69, %r48;
mov.f32 %f27, 0f00000000;
or.pred %p7, %p5, %p6;
mov.f32 %f4, %f27;
mov.f32 %f5, %f27;
@%p7 bra $L__BB155_8;

setp.eq.s64 %p8, %rd39, 0;
mov.f32 %f4, 0f00000000;
mov.f32 %f5, %f4;
@%p8 bra $L__BB155_6;

cvta.to.global.u64 %rd48, %rd39;
mul.lo.s32 %r70, %r11, %r60;
cvt.u64.u32 %rd49, %r70;
mul.lo.s32 %r71, %r12, %r54;
cvt.u64.u32 %rd50, %r71;
cvt.u64.u32 %rd51, %r2;
cvt.s64.s32 %rd52, %r4;
add.s64 %rd53, %rd52, %rd51;
add.s64 %rd54, %rd53, %rd49;
add.s64 %rd55, %rd54, %rd50;
shl.b64 %rd56, %rd55, 2;
add.s64 %rd57, %rd48, %rd56;
ld.global.f32 %f5, [%rd57];

$L__BB155_6:
setp.eq.s64 %p9, %rd41, 0;
setp.eq.s64 %p10, %rd40, 0;
or.pred %p11, %p9, %p10;
@%p11 bra $L__BB155_8;

cvta.to.global.u64 %rd58, %rd41;
mul.lo.s32 %r72, %r11, %r60;
cvt.u64.u32 %rd59, %r72;
mul.lo.s32 %r73, %r12, %r54;
cvt.u64.u32 %rd60, %r73;
cvt.u64.u32 %rd61, %r2;
cvt.s64.s32 %rd62, %r4;
add.s64 %rd63, %rd62, %rd61;
add.s64 %rd64, %rd63, %rd59;
add.s64 %rd65, %rd64, %rd60;
shl.b64 %rd66, %rd65, 2;
add.s64 %rd67, %rd58, %rd66;
ld.global.f32 %f4, [%rd67];

$L__BB155_8:
mul.lo.s32 %r74, %r11, %r59;
cvt.u64.u32 %rd68, %r74;
mul.lo.s32 %r75, %r1366, %r53;
cvt.u64.u32 %rd69, %r75;
add.s64 %rd4, %rd69, %rd68;
mul.hi.u32 %r76, %r11, %r55;
add.s32 %r77, %r76, %r11;
shr.u32 %r13, %r77, %r56;
shr.s32 %r78, %r3, 31;
shr.u32 %r79, %r78, 28;
add.s32 %r80, %r3, %r79;
shr.s32 %r1368, %r80, 4;
shr.u32 %r81, %r43, 8;
setp.ge.s32 %p12, %r1368, %r81;
mov.f32 %f28, %f27;
mov.f32 %f29, %f27;
mov.f32 %f30, %f27;
mov.f32 %f31, %f27;
mov.f32 %f32, %f27;
mov.f32 %f33, %f27;
mov.f32 %f34, %f27;
@%p12 bra $L__BB155_19;

mul.lo.s32 %r82, %r13, %r58;
mul.lo.s32 %r83, %r1365, %r52;
and.b32 %r87, %r80, -16;
sub.s32 %r15, %r3, %r87;
shr.s32 %r88, %r15, 31;
shr.u32 %r89, %r88, 30;
add.s32 %r90, %r15, %r89;
shr.s32 %r91, %r90, 2;
shl.b32 %r92, %r91, 1;
shl.b32 %r93, %r91, 5;
and.b32 %r94, %r90, -4;
sub.s32 %r95, %r15, %r94;
shl.b32 %r96, %r95, 2;
add.s32 %r97, %r93, %r96;
cvt.s64.s32 %rd70, %r97;
cvt.s64.s32 %rd5, %r92;
cvt.s64.s32 %rd6, %r95;
shl.b32 %r1367, %r1368, 3;
add.s32 %r98, %r1368, %r82;
add.s32 %r99, %r98, %r83;
add.s32 %r100, %r4, 1;
mad.lo.s32 %r101, %r47, %r100, %r99;
mul.wide.s32 %rd7, %r101, 144;
mad.lo.s32 %r102, %r4, %r47, %r1368;
add.s32 %r103, %r102, %r82;
add.s32 %r104, %r103, %r83;
mul.wide.s32 %rd8, %r104, 144;
add.s64 %rd71, %rd8, %rd70;
add.s64 %rd9, %rd71, 32;
cvt.u64.u32 %rd72, %r92;
and.b64 %rd73, %rd72, 2;
or.b64 %rd74, %rd73, %rd7;
or.b64 %rd10, %rd74, 8;
or.b64 %rd75, %rd73, %rd8;
or.b64 %rd11, %rd75, 12;
add.s64 %rd76, %rd7, %rd70;
add.s64 %rd12, %rd76, 16;
add.s32 %r105, %r4, 2;
mad.lo.s32 %r106, %r47, %r105, %r99;
mul.wide.s32 %rd13, %r106, 144;
add.s32 %r107, %r4, 3;
mad.lo.s32 %r108, %r47, %r107, %r99;
mul.wide.s32 %rd14, %r108, 144;
add.s64 %rd77, %rd14, %rd70;
add.s64 %rd15, %rd77, 32;
or.b64 %rd78, %rd73, %rd13;
or.b64 %rd16, %rd78, 8;
or.b64 %rd79, %rd73, %rd14;
or.b64 %rd17, %rd79, 12;
add.s64 %rd80, %rd13, %rd70;
add.s64 %rd18, %rd80, 16;
mul.wide.s32 %rd19, %r47, 144;
cvta.to.global.u64 %rd114, %rd40;
cvta.to.global.u64 %rd113, %rd36;
mov.f32 %f27, 0f00000000;
cvta.to.global.u64 %rd81, %rd37;
shl.b64 %rd89, %rd6, 2;
mov.f32 %f28, %f27;
mov.f32 %f29, %f27;
mov.f32 %f30, %f27;
mov.f32 %f31, %f27;
mov.f32 %f32, %f27;
mov.f32 %f33, %f27;
mov.f32 %f34, %f27;

$L__BB155_10:
cvt.u64.u32 %rd82, %r1367;
add.s64 %rd83, %rd4, %rd82;
add.s64 %rd84, %rd113, %rd9;
ld.global.nc.u32 %r145, [%rd84+-16];
ld.global.nc.u32 %r146, [%rd84];
add.s64 %rd85, %rd113, %rd11;
ld.global.nc.u16 %rs1, [%rd85+-8];
cvt.u32.u16 %r147, %rs1;
ld.global.nc.u16 %rs2, [%rd85+-4];
cvt.u32.u16 %r148, %rs2;
ld.global.nc.u16 %rs3, [%rd85];
cvt.u32.u16 %r149, %rs3;
and.b32 %r150, %r147, 16191;
and.b32 %r151, %r149, 3855;
and.b32 %r152, %r147, 49344;
shr.u32 %r153, %r152, 2;
or.b32 %r154, %r151, %r153;
setp.gt.s32 %p13, %r15, 7;
and.b32 %r155, %r148, 16191;
shr.u32 %r156, %r149, 4;
and.b32 %r157, %r156, 3855;
and.b32 %r158, %r148, 49344;
shr.u32 %r159, %r158, 2;
or.b32 %r160, %r157, %r159;
add.s64 %rd86, %rd83, %rd5;
mul.lo.s64 %rd87, %rd86, 36;
add.s64 %rd88, %rd81, %rd87;
ld.global.nc.u32 %r19, [%rd88];

{.reg .f16 low,high;
mov.b32 {low,high},%r19;
cvt.f32.f16 %f101, low;}


add.s64 %rd90, %rd88, %rd89;
ld.global.nc.u32 %r20, [%rd90+4];
ld.global.nc.u32 %r21, [%rd90+20];
ld.global.nc.u32 %r22, [%rd88+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r22;
cvt.f32.f16 %f102, low;}


ld.global.nc.u32 %r23, [%rd90+40];
ld.global.nc.u32 %r24, [%rd90+56];
add.s64 %rd26, %rd113, %rd8;
and.b32 %r112, %r145, 252645135;
and.b32 %r116, %r146, 252645135;
mov.u32 %r138, 0;

dp4a.s32.s32 %r111, %r112, %r20, %r138;


dp4a.s32.s32 %r115, %r116, %r21, %r111;

mov.u32 %r140, 16843009;

dp4a.s32.s32 %r119, %r140, %r20, %r138;


dp4a.s32.s32 %r123, %r140, %r21, %r119;

selp.b32 %r161, %r154, %r150, %p13;
cvt.u16.u32 %rs4, %r161;
and.b32 %r162, %r161, 63;
mul.lo.s32 %r163, %r115, %r162;
cvt.rn.f32.s32 %f105, %r163;
fma.rn.ftz.f32 %f106, %f101, %f105, 0f00000000;
selp.b32 %r164, %r160, %r155, %p13;
cvt.u16.u32 %rs5, %r164;
and.b32 %r165, %r164, 63;
mul.lo.s32 %r166, %r123, %r165;
cvt.rn.f32.s32 %f107, %r166;
fma.rn.ftz.f32 %f108, %f101, %f107, 0f00000000;
shr.u32 %r167, %r145, 4;
and.b32 %r128, %r167, 252645135;
shr.u32 %r168, %r146, 4;
and.b32 %r132, %r168, 252645135;

dp4a.s32.s32 %r127, %r128, %r23, %r138;


dp4a.s32.s32 %r131, %r132, %r24, %r127;


dp4a.s32.s32 %r135, %r140, %r23, %r138;


dp4a.s32.s32 %r139, %r140, %r24, %r135;

shr.u16 %rs6, %rs4, 8;
cvt.u32.u16 %r169, %rs6;
mul.lo.s32 %r170, %r131, %r169;
cvt.rn.f32.s32 %f109, %r170;
fma.rn.ftz.f32 %f110, %f102, %f109, %f106;
shr.u16 %rs7, %rs5, 8;
cvt.u32.u16 %r171, %rs7;
mul.lo.s32 %r172, %r139, %r171;
cvt.rn.f32.s32 %f111, %r172;
fma.rn.ftz.f32 %f112, %f102, %f111, %f108;
ld.global.nc.u32 %r143, [%rd26];

{.reg .f16 low,high;
mov.b32 {low,high},%r143;
cvt.f32.f16 %f103, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r143;
cvt.f32.f16 %f104, high;}


mul.ftz.f32 %f113, %f110, %f103;
mul.ftz.f32 %f114, %f112, %f104;
sub.ftz.f32 %f115, %f113, %f114;
add.ftz.f32 %f30, %f115, %f30;
setp.eq.s64 %p14, %rd40, 0;
@%p14 bra $L__BB155_12;

add.s64 %rd91, %rd114, %rd9;
ld.global.u32 %r209, [%rd91+-16];
ld.global.u32 %r210, [%rd91];
add.s64 %rd92, %rd114, %rd11;
ld.global.u16 %r211, [%rd92+-8];
ld.global.u16 %r212, [%rd92+-4];
ld.global.u16 %r213, [%rd92];
and.b32 %r214, %r211, 16191;
and.b32 %r215, %r213, 3855;
and.b32 %r216, %r211, 49344;
shr.u32 %r217, %r216, 2;
or.b32 %r218, %r215, %r217;
and.b32 %r219, %r212, 16191;
shr.u32 %r220, %r213, 4;
and.b32 %r221, %r220, 3855;
and.b32 %r222, %r212, 49344;
shr.u32 %r223, %r222, 2;
or.b32 %r224, %r221, %r223;

{.reg .f16 low,high;
mov.b32 {low,high},%r19;
cvt.f32.f16 %f116, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r22;
cvt.f32.f16 %f117, low;}


add.s64 %rd93, %rd114, %rd8;
and.b32 %r176, %r209, 252645135;
and.b32 %r180, %r210, 252645135;

dp4a.s32.s32 %r175, %r176, %r20, %r138;


dp4a.s32.s32 %r179, %r180, %r21, %r175;


dp4a.s32.s32 %r183, %r140, %r20, %r138;


dp4a.s32.s32 %r187, %r140, %r21, %r183;

selp.b32 %r225, %r218, %r214, %p13;
cvt.u16.u32 %rs8, %r225;
and.b32 %r226, %r225, 63;
mul.lo.s32 %r227, %r179, %r226;
cvt.rn.f32.s32 %f120, %r227;
fma.rn.ftz.f32 %f121, %f116, %f120, 0f00000000;
selp.b32 %r228, %r224, %r219, %p13;
cvt.u16.u32 %rs9, %r228;
and.b32 %r229, %r228, 63;
mul.lo.s32 %r230, %r187, %r229;
cvt.rn.f32.s32 %f122, %r230;
fma.rn.ftz.f32 %f123, %f116, %f122, 0f00000000;
shr.u32 %r231, %r209, 4;
and.b32 %r192, %r231, 252645135;
shr.u32 %r232, %r210, 4;
and.b32 %r196, %r232, 252645135;

dp4a.s32.s32 %r191, %r192, %r23, %r138;


dp4a.s32.s32 %r195, %r196, %r24, %r191;


dp4a.s32.s32 %r199, %r140, %r23, %r138;


dp4a.s32.s32 %r203, %r140, %r24, %r199;

shr.u16 %rs10, %rs8, 8;
cvt.u32.u16 %r233, %rs10;
mul.lo.s32 %r234, %r195, %r233;
cvt.rn.f32.s32 %f124, %r234;
fma.rn.ftz.f32 %f125, %f117, %f124, %f121;
shr.u16 %rs11, %rs9, 8;
cvt.u32.u16 %r235, %rs11;
mul.lo.s32 %r236, %r203, %r235;
cvt.rn.f32.s32 %f126, %r236;
fma.rn.ftz.f32 %f127, %f117, %f126, %f123;
ld.global.u32 %r208, [%rd93];

{.reg .f16 low,high;
mov.b32 {low,high},%r208;
cvt.f32.f16 %f118, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r208;
cvt.f32.f16 %f119, high;}


mul.ftz.f32 %f128, %f125, %f118;
mul.ftz.f32 %f129, %f127, %f119;
sub.ftz.f32 %f130, %f128, %f129;
add.ftz.f32 %f34, %f130, %f34;

$L__BB155_12:
add.s64 %rd111, %rd113, %rd11;
add.s64 %rd110, %rd111, -8;
add.s64 %rd109, %rd113, %rd9;
add.s64 %rd108, %rd109, -16;
add.s64 %rd27, %rd108, %rd19;
ld.global.nc.u32 %r273, [%rd27];
ld.global.nc.u32 %r274, [%rd27+16];
add.s64 %rd28, %rd110, %rd19;
ld.global.nc.u16 %rs12, [%rd28];
cvt.u32.u16 %r275, %rs12;
ld.global.nc.u16 %rs13, [%rd28+4];
cvt.u32.u16 %r276, %rs13;
ld.global.nc.u16 %rs14, [%rd28+8];
cvt.u32.u16 %r277, %rs14;
and.b32 %r278, %r275, 16191;
and.b32 %r279, %r277, 3855;
and.b32 %r280, %r275, 49344;
shr.u32 %r281, %r280, 2;
or.b32 %r282, %r279, %r281;
and.b32 %r283, %r276, 16191;
shr.u32 %r284, %r277, 4;
and.b32 %r285, %r284, 3855;
and.b32 %r286, %r276, 49344;
shr.u32 %r287, %r286, 2;
or.b32 %r288, %r285, %r287;

{.reg .f16 low,high;
mov.b32 {low,high},%r19;
cvt.f32.f16 %f131, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r22;
cvt.f32.f16 %f132, low;}


and.b32 %r240, %r273, 252645135;
and.b32 %r244, %r274, 252645135;

dp4a.s32.s32 %r239, %r240, %r20, %r138;


dp4a.s32.s32 %r243, %r244, %r21, %r239;


dp4a.s32.s32 %r247, %r140, %r20, %r138;


dp4a.s32.s32 %r251, %r140, %r21, %r247;

selp.b32 %r289, %r282, %r278, %p13;
cvt.u16.u32 %rs15, %r289;
and.b32 %r290, %r289, 63;
mul.lo.s32 %r291, %r243, %r290;
cvt.rn.f32.s32 %f135, %r291;
fma.rn.ftz.f32 %f136, %f131, %f135, 0f00000000;
selp.b32 %r292, %r288, %r283, %p13;
cvt.u16.u32 %rs16, %r292;
and.b32 %r293, %r292, 63;
mul.lo.s32 %r294, %r251, %r293;
cvt.rn.f32.s32 %f137, %r294;
fma.rn.ftz.f32 %f138, %f131, %f137, 0f00000000;
shr.u32 %r295, %r273, 4;
and.b32 %r256, %r295, 252645135;
shr.u32 %r296, %r274, 4;
and.b32 %r260, %r296, 252645135;

dp4a.s32.s32 %r255, %r256, %r23, %r138;


dp4a.s32.s32 %r259, %r260, %r24, %r255;


dp4a.s32.s32 %r263, %r140, %r23, %r138;


dp4a.s32.s32 %r267, %r140, %r24, %r263;

shr.u16 %rs17, %rs15, 8;
cvt.u32.u16 %r297, %rs17;
mul.lo.s32 %r298, %r259, %r297;
cvt.rn.f32.s32 %f139, %r298;
fma.rn.ftz.f32 %f140, %f132, %f139, %f136;
shr.u16 %rs18, %rs16, 8;
cvt.u32.u16 %r299, %rs18;
mul.lo.s32 %r300, %r267, %r299;
cvt.rn.f32.s32 %f141, %r300;
fma.rn.ftz.f32 %f142, %f132, %f141, %f138;
add.s64 %rd29, %rd26, %rd19;
ld.global.nc.u32 %r271, [%rd29];

{.reg .f16 low,high;
mov.b32 {low,high},%r271;
cvt.f32.f16 %f133, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r271;
cvt.f32.f16 %f134, high;}


mul.ftz.f32 %f143, %f140, %f133;
mul.ftz.f32 %f144, %f142, %f134;
sub.ftz.f32 %f145, %f143, %f144;
add.ftz.f32 %f29, %f145, %f29;
@%p14 bra $L__BB155_14;

add.s64 %rd94, %rd114, %rd12;
ld.global.u32 %r337, [%rd94];
ld.global.u32 %r338, [%rd94+16];
add.s64 %rd95, %rd114, %rd10;
ld.global.u16 %r339, [%rd95+-4];
ld.global.u16 %r340, [%rd95];
ld.global.u16 %r341, [%rd95+4];
and.b32 %r342, %r339, 16191;
and.b32 %r343, %r341, 3855;
and.b32 %r344, %r339, 49344;
shr.u32 %r345, %r344, 2;
or.b32 %r346, %r343, %r345;
and.b32 %r347, %r340, 16191;
shr.u32 %r348, %r341, 4;
and.b32 %r349, %r348, 3855;
and.b32 %r350, %r340, 49344;
shr.u32 %r351, %r350, 2;
or.b32 %r352, %r349, %r351;

{.reg .f16 low,high;
mov.b32 {low,high},%r19;
cvt.f32.f16 %f146, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r22;
cvt.f32.f16 %f147, low;}


add.s64 %rd96, %rd114, %rd7;
and.b32 %r304, %r337, 252645135;
and.b32 %r308, %r338, 252645135;
mov.u32 %r330, 0;

dp4a.s32.s32 %r303, %r304, %r20, %r330;


dp4a.s32.s32 %r307, %r308, %r21, %r303;

mov.u32 %r332, 16843009;

dp4a.s32.s32 %r311, %r332, %r20, %r330;


dp4a.s32.s32 %r315, %r332, %r21, %r311;

selp.b32 %r353, %r346, %r342, %p13;
cvt.u16.u32 %rs19, %r353;
and.b32 %r354, %r353, 63;
mul.lo.s32 %r355, %r307, %r354;
cvt.rn.f32.s32 %f150, %r355;
fma.rn.ftz.f32 %f151, %f146, %f150, 0f00000000;
selp.b32 %r356, %r352, %r347, %p13;
cvt.u16.u32 %rs20, %r356;
and.b32 %r357, %r356, 63;
mul.lo.s32 %r358, %r315, %r357;
cvt.rn.f32.s32 %f152, %r358;
fma.rn.ftz.f32 %f153, %f146, %f152, 0f00000000;
shr.u32 %r359, %r337, 4;
and.b32 %r320, %r359, 252645135;
shr.u32 %r360, %r338, 4;
and.b32 %r324, %r360, 252645135;

dp4a.s32.s32 %r319, %r320, %r23, %r330;


dp4a.s32.s32 %r323, %r324, %r24, %r319;


dp4a.s32.s32 %r327, %r332, %r23, %r330;


dp4a.s32.s32 %r331, %r332, %r24, %r327;

shr.u16 %rs21, %rs19, 8;
cvt.u32.u16 %r361, %rs21;
mul.lo.s32 %r362, %r323, %r361;
cvt.rn.f32.s32 %f154, %r362;
fma.rn.ftz.f32 %f155, %f147, %f154, %f151;
shr.u16 %rs22, %rs20, 8;
cvt.u32.u16 %r363, %rs22;
mul.lo.s32 %r364, %r331, %r363;
cvt.rn.f32.s32 %f156, %r364;
fma.rn.ftz.f32 %f157, %f147, %f156, %f153;
ld.global.u32 %r336, [%rd96];

{.reg .f16 low,high;
mov.b32 {low,high},%r336;
cvt.f32.f16 %f148, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r336;
cvt.f32.f16 %f149, high;}


mul.ftz.f32 %f158, %f155, %f148;
mul.ftz.f32 %f159, %f157, %f149;
sub.ftz.f32 %f160, %f158, %f159;
add.ftz.f32 %f33, %f160, %f33;

$L__BB155_14:
add.s64 %rd30, %rd27, %rd19;
ld.global.nc.u32 %r401, [%rd30];
ld.global.nc.u32 %r402, [%rd30+16];
add.s64 %rd31, %rd28, %rd19;
ld.global.nc.u16 %rs23, [%rd31];
cvt.u32.u16 %r403, %rs23;
ld.global.nc.u16 %rs24, [%rd31+4];
cvt.u32.u16 %r404, %rs24;
ld.global.nc.u16 %rs25, [%rd31+8];
cvt.u32.u16 %r405, %rs25;
and.b32 %r406, %r403, 16191;
and.b32 %r407, %r405, 3855;
and.b32 %r408, %r403, 49344;
shr.u32 %r409, %r408, 2;
or.b32 %r410, %r407, %r409;
and.b32 %r411, %r404, 16191;
shr.u32 %r412, %r405, 4;
and.b32 %r413, %r412, 3855;
and.b32 %r414, %r404, 49344;
shr.u32 %r415, %r414, 2;
or.b32 %r416, %r413, %r415;

{.reg .f16 low,high;
mov.b32 {low,high},%r19;
cvt.f32.f16 %f161, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r22;
cvt.f32.f16 %f162, low;}


and.b32 %r368, %r401, 252645135;
and.b32 %r372, %r402, 252645135;
mov.u32 %r394, 0;

dp4a.s32.s32 %r367, %r368, %r20, %r394;


dp4a.s32.s32 %r371, %r372, %r21, %r367;

mov.u32 %r396, 16843009;

dp4a.s32.s32 %r375, %r396, %r20, %r394;


dp4a.s32.s32 %r379, %r396, %r21, %r375;

selp.b32 %r417, %r410, %r406, %p13;
cvt.u16.u32 %rs26, %r417;
and.b32 %r418, %r417, 63;
mul.lo.s32 %r419, %r371, %r418;
cvt.rn.f32.s32 %f165, %r419;
fma.rn.ftz.f32 %f166, %f161, %f165, 0f00000000;
selp.b32 %r420, %r416, %r411, %p13;
cvt.u16.u32 %rs27, %r420;
and.b32 %r421, %r420, 63;
mul.lo.s32 %r422, %r379, %r421;
cvt.rn.f32.s32 %f167, %r422;
fma.rn.ftz.f32 %f168, %f161, %f167, 0f00000000;
shr.u32 %r423, %r401, 4;
and.b32 %r384, %r423, 252645135;
shr.u32 %r424, %r402, 4;
and.b32 %r388, %r424, 252645135;

dp4a.s32.s32 %r383, %r384, %r23, %r394;


dp4a.s32.s32 %r387, %r388, %r24, %r383;


dp4a.s32.s32 %r391, %r396, %r23, %r394;


dp4a.s32.s32 %r395, %r396, %r24, %r391;

shr.u16 %rs28, %rs26, 8;
cvt.u32.u16 %r425, %rs28;
mul.lo.s32 %r426, %r387, %r425;
cvt.rn.f32.s32 %f169, %r426;
fma.rn.ftz.f32 %f170, %f162, %f169, %f166;
shr.u16 %rs29, %rs27, 8;
cvt.u32.u16 %r427, %rs29;
mul.lo.s32 %r428, %r395, %r427;
cvt.rn.f32.s32 %f171, %r428;
fma.rn.ftz.f32 %f172, %f162, %f171, %f168;
add.s64 %rd32, %rd29, %rd19;
ld.global.nc.u32 %r399, [%rd32];

{.reg .f16 low,high;
mov.b32 {low,high},%r399;
cvt.f32.f16 %f163, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r399;
cvt.f32.f16 %f164, high;}


mul.ftz.f32 %f173, %f170, %f163;
mul.ftz.f32 %f174, %f172, %f164;
sub.ftz.f32 %f175, %f173, %f174;
add.ftz.f32 %f28, %f175, %f28;
@%p14 bra $L__BB155_16;

add.s64 %rd97, %rd114, %rd18;
ld.global.u32 %r465, [%rd97];
ld.global.u32 %r466, [%rd97+16];
add.s64 %rd98, %rd114, %rd16;
ld.global.u16 %r467, [%rd98+-4];
ld.global.u16 %r468, [%rd98];
ld.global.u16 %r469, [%rd98+4];
and.b32 %r470, %r467, 16191;
and.b32 %r471, %r469, 3855;
and.b32 %r472, %r467, 49344;
shr.u32 %r473, %r472, 2;
or.b32 %r474, %r471, %r473;
and.b32 %r475, %r468, 16191;
shr.u32 %r476, %r469, 4;
and.b32 %r477, %r476, 3855;
and.b32 %r478, %r468, 49344;
shr.u32 %r479, %r478, 2;
or.b32 %r480, %r477, %r479;

{.reg .f16 low,high;
mov.b32 {low,high},%r19;
cvt.f32.f16 %f176, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r22;
cvt.f32.f16 %f177, low;}


add.s64 %rd99, %rd114, %rd13;
and.b32 %r432, %r465, 252645135;
and.b32 %r436, %r466, 252645135;

dp4a.s32.s32 %r431, %r432, %r20, %r394;


dp4a.s32.s32 %r435, %r436, %r21, %r431;


dp4a.s32.s32 %r439, %r396, %r20, %r394;


dp4a.s32.s32 %r443, %r396, %r21, %r439;

selp.b32 %r481, %r474, %r470, %p13;
cvt.u16.u32 %rs30, %r481;
and.b32 %r482, %r481, 63;
mul.lo.s32 %r483, %r435, %r482;
cvt.rn.f32.s32 %f180, %r483;
fma.rn.ftz.f32 %f181, %f176, %f180, 0f00000000;
selp.b32 %r484, %r480, %r475, %p13;
cvt.u16.u32 %rs31, %r484;
and.b32 %r485, %r484, 63;
mul.lo.s32 %r486, %r443, %r485;
cvt.rn.f32.s32 %f182, %r486;
fma.rn.ftz.f32 %f183, %f176, %f182, 0f00000000;
shr.u32 %r487, %r465, 4;
and.b32 %r448, %r487, 252645135;
shr.u32 %r488, %r466, 4;
and.b32 %r452, %r488, 252645135;

dp4a.s32.s32 %r447, %r448, %r23, %r394;


dp4a.s32.s32 %r451, %r452, %r24, %r447;


dp4a.s32.s32 %r455, %r396, %r23, %r394;


dp4a.s32.s32 %r459, %r396, %r24, %r455;

shr.u16 %rs32, %rs30, 8;
cvt.u32.u16 %r489, %rs32;
mul.lo.s32 %r490, %r451, %r489;
cvt.rn.f32.s32 %f184, %r490;
fma.rn.ftz.f32 %f185, %f177, %f184, %f181;
shr.u16 %rs33, %rs31, 8;
cvt.u32.u16 %r491, %rs33;
mul.lo.s32 %r492, %r459, %r491;
cvt.rn.f32.s32 %f186, %r492;
fma.rn.ftz.f32 %f187, %f177, %f186, %f183;
ld.global.u32 %r464, [%rd99];

{.reg .f16 low,high;
mov.b32 {low,high},%r464;
cvt.f32.f16 %f178, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r464;
cvt.f32.f16 %f179, high;}


mul.ftz.f32 %f188, %f185, %f178;
mul.ftz.f32 %f189, %f187, %f179;
sub.ftz.f32 %f190, %f188, %f189;
add.ftz.f32 %f32, %f190, %f32;

$L__BB155_16:
add.s64 %rd100, %rd30, %rd19;
ld.global.nc.u32 %r529, [%rd100];
ld.global.nc.u32 %r530, [%rd100+16];
add.s64 %rd101, %rd31, %rd19;
ld.global.nc.u16 %rs34, [%rd101];
cvt.u32.u16 %r531, %rs34;
ld.global.nc.u16 %rs35, [%rd101+4];
cvt.u32.u16 %r532, %rs35;
ld.global.nc.u16 %rs36, [%rd101+8];
cvt.u32.u16 %r533, %rs36;
and.b32 %r534, %r531, 16191;
and.b32 %r535, %r533, 3855;
and.b32 %r536, %r531, 49344;
shr.u32 %r537, %r536, 2;
or.b32 %r538, %r535, %r537;
and.b32 %r539, %r532, 16191;
shr.u32 %r540, %r533, 4;
and.b32 %r541, %r540, 3855;
and.b32 %r542, %r532, 49344;
shr.u32 %r543, %r542, 2;
or.b32 %r544, %r541, %r543;

{.reg .f16 low,high;
mov.b32 {low,high},%r19;
cvt.f32.f16 %f191, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r22;
cvt.f32.f16 %f192, low;}


and.b32 %r496, %r529, 252645135;
and.b32 %r500, %r530, 252645135;

dp4a.s32.s32 %r495, %r496, %r20, %r394;


dp4a.s32.s32 %r499, %r500, %r21, %r495;


dp4a.s32.s32 %r503, %r396, %r20, %r394;


dp4a.s32.s32 %r507, %r396, %r21, %r503;

selp.b32 %r545, %r538, %r534, %p13;
cvt.u16.u32 %rs37, %r545;
and.b32 %r546, %r545, 63;
mul.lo.s32 %r547, %r499, %r546;
cvt.rn.f32.s32 %f195, %r547;
fma.rn.ftz.f32 %f196, %f191, %f195, 0f00000000;
selp.b32 %r548, %r544, %r539, %p13;
cvt.u16.u32 %rs38, %r548;
and.b32 %r549, %r548, 63;
mul.lo.s32 %r550, %r507, %r549;
cvt.rn.f32.s32 %f197, %r550;
fma.rn.ftz.f32 %f198, %f191, %f197, 0f00000000;
shr.u32 %r551, %r529, 4;
and.b32 %r512, %r551, 252645135;
shr.u32 %r552, %r530, 4;
and.b32 %r516, %r552, 252645135;

dp4a.s32.s32 %r511, %r512, %r23, %r394;


dp4a.s32.s32 %r515, %r516, %r24, %r511;


dp4a.s32.s32 %r519, %r396, %r23, %r394;


dp4a.s32.s32 %r523, %r396, %r24, %r519;

shr.u16 %rs39, %rs37, 8;
cvt.u32.u16 %r553, %rs39;
mul.lo.s32 %r554, %r515, %r553;
cvt.rn.f32.s32 %f199, %r554;
fma.rn.ftz.f32 %f200, %f192, %f199, %f196;
shr.u16 %rs40, %rs38, 8;
cvt.u32.u16 %r555, %rs40;
mul.lo.s32 %r556, %r523, %r555;
cvt.rn.f32.s32 %f201, %r556;
fma.rn.ftz.f32 %f202, %f192, %f201, %f198;
add.s64 %rd102, %rd32, %rd19;
ld.global.nc.u32 %r527, [%rd102];

{.reg .f16 low,high;
mov.b32 {low,high},%r527;
cvt.f32.f16 %f193, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r527;
cvt.f32.f16 %f194, high;}


mul.ftz.f32 %f203, %f200, %f193;
mul.ftz.f32 %f204, %f202, %f194;
sub.ftz.f32 %f205, %f203, %f204;
add.ftz.f32 %f27, %f205, %f27;
@%p14 bra $L__BB155_18;

add.s64 %rd103, %rd114, %rd15;
ld.global.u32 %r593, [%rd103+-16];
ld.global.u32 %r594, [%rd103];
add.s64 %rd104, %rd114, %rd17;
ld.global.u16 %r595, [%rd104+-8];
ld.global.u16 %r596, [%rd104+-4];
ld.global.u16 %r597, [%rd104];
and.b32 %r598, %r595, 16191;
and.b32 %r599, %r597, 3855;
and.b32 %r600, %r595, 49344;
shr.u32 %r601, %r600, 2;
or.b32 %r602, %r599, %r601;
and.b32 %r603, %r596, 16191;
shr.u32 %r604, %r597, 4;
and.b32 %r605, %r604, 3855;
and.b32 %r606, %r596, 49344;
shr.u32 %r607, %r606, 2;
or.b32 %r608, %r605, %r607;

{.reg .f16 low,high;
mov.b32 {low,high},%r19;
cvt.f32.f16 %f206, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r22;
cvt.f32.f16 %f207, low;}


add.s64 %rd105, %rd114, %rd14;
and.b32 %r560, %r593, 252645135;
and.b32 %r564, %r594, 252645135;
mov.u32 %r586, 0;

dp4a.s32.s32 %r559, %r560, %r20, %r586;


dp4a.s32.s32 %r563, %r564, %r21, %r559;

mov.u32 %r588, 16843009;

dp4a.s32.s32 %r567, %r588, %r20, %r586;


dp4a.s32.s32 %r571, %r588, %r21, %r567;

selp.b32 %r609, %r602, %r598, %p13;
cvt.u16.u32 %rs41, %r609;
and.b32 %r610, %r609, 63;
mul.lo.s32 %r611, %r563, %r610;
cvt.rn.f32.s32 %f210, %r611;
fma.rn.ftz.f32 %f211, %f206, %f210, 0f00000000;
selp.b32 %r612, %r608, %r603, %p13;
cvt.u16.u32 %rs42, %r612;
and.b32 %r613, %r612, 63;
mul.lo.s32 %r614, %r571, %r613;
cvt.rn.f32.s32 %f212, %r614;
fma.rn.ftz.f32 %f213, %f206, %f212, 0f00000000;
shr.u32 %r615, %r593, 4;
and.b32 %r576, %r615, 252645135;
shr.u32 %r616, %r594, 4;
and.b32 %r580, %r616, 252645135;

dp4a.s32.s32 %r575, %r576, %r23, %r586;


dp4a.s32.s32 %r579, %r580, %r24, %r575;


dp4a.s32.s32 %r583, %r588, %r23, %r586;


dp4a.s32.s32 %r587, %r588, %r24, %r583;

shr.u16 %rs43, %rs41, 8;
cvt.u32.u16 %r617, %rs43;
mul.lo.s32 %r618, %r579, %r617;
cvt.rn.f32.s32 %f214, %r618;
fma.rn.ftz.f32 %f215, %f207, %f214, %f211;
shr.u16 %rs44, %rs42, 8;
cvt.u32.u16 %r619, %rs44;
mul.lo.s32 %r620, %r587, %r619;
cvt.rn.f32.s32 %f216, %r620;
fma.rn.ftz.f32 %f217, %f207, %f216, %f213;
ld.global.u32 %r592, [%rd105];

{.reg .f16 low,high;
mov.b32 {low,high},%r592;
cvt.f32.f16 %f208, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r592;
cvt.f32.f16 %f209, high;}


mul.ftz.f32 %f218, %f215, %f208;
mul.ftz.f32 %f219, %f217, %f209;
sub.ftz.f32 %f220, %f218, %f219;
add.ftz.f32 %f31, %f220, %f31;

$L__BB155_18:
add.s32 %r1367, %r1367, 64;
add.s64 %rd114, %rd114, 1152;
add.s64 %rd113, %rd113, 1152;
add.s32 %r1368, %r1368, 8;
setp.lt.s32 %p25, %r1368, %r81;
@%p25 bra $L__BB155_10;

$L__BB155_19:
mov.u32 %r1354, %tid.y;
setp.eq.s32 %p26, %r1354, 0;
@%p26 bra $L__BB155_23;

mov.u32 %r1355, %tid.y;
setp.ne.s64 %p27, %rd40, 0;
add.s32 %r27, %r1355, -1;
shl.b32 %r625, %r1355, 9;
mov.u32 %r626, _ZZ13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r627, %r626, %r625;
shl.b32 %r628, %r2, 2;
add.s32 %r629, %r627, %r628;
st.shared.f32 [%r629+-512], %f30;
@%p27 bra $L__BB155_22;
bra.uni $L__BB155_21;

$L__BB155_22:
shl.b32 %r631, %r27, 9;
mov.u32 %r632, _ZZ13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r633, %r632, %r631;
add.s32 %r635, %r633, %r628;
st.shared.f32 [%r635], %f34;
add.s32 %r1351, %r629, -512;
st.shared.f32 [%r1351+128], %f29;
st.shared.f32 [%r635+128], %f33;
add.s32 %r1352, %r629, -512;
st.shared.f32 [%r1352+256], %f28;
st.shared.f32 [%r635+256], %f32;
add.s32 %r1353, %r629, -512;
st.shared.f32 [%r1353+384], %f27;
st.shared.f32 [%r635+384], %f31;
bra.uni $L__BB155_23;

$L__BB155_21:
add.s32 %r1348, %r629, -512;
st.shared.f32 [%r1348+128], %f29;
add.s32 %r1349, %r629, -512;
st.shared.f32 [%r1349+256], %f28;
add.s32 %r1350, %r629, -512;
st.shared.f32 [%r1350+384], %f27;

$L__BB155_23:
mov.u32 %r1364, %tid.y;
setp.ne.s32 %p327, %r1364, 0;
bar.sync 0;
@%p327 bra $L__BB155_78;
bra.uni $L__BB155_24;

$L__BB155_78:
ret;

$L__BB155_24:
ld.param.u64 %rd112, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.f32 %f1065, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
ld.param.u32 %r1358, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r1357, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
mov.u32 %r1356, %ctaid.z;
setp.ne.s64 %p29, %rd40, 0;
mov.u32 %r638, %ctaid.y;
mad.lo.s32 %r641, %r638, %r1358, %r4;
mad.lo.s32 %r642, %r1356, %r1357, %r641;
neg.ftz.f32 %f35, %f1065;
shl.b32 %r644, %r2, 2;
mov.u32 %r645, _ZZ13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r29, %r645, %r644;
cvta.to.global.u64 %rd106, %rd112;
mul.wide.u32 %rd107, %r642, 4;
add.s64 %rd35, %rd106, %rd107;
@%p29 bra $L__BB155_33;
bra.uni $L__BB155_25;

$L__BB155_33:
ld.param.u32 %r1359, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
mov.u32 %r732, _ZZ13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r31, %r732, %r644;
setp.gt.s32 %p62, %r1359, 2;
@%p62 bra $L__BB155_44;

ld.param.u32 %r1362, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p65, %r1362, 1;
@%p65 bra $L__BB155_62;

ld.param.u32 %r1363, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p66, %r1363, 2;
@%p66 bra $L__BB155_36;
bra.uni $L__BB155_70;

$L__BB155_36:
ld.shared.f32 %f621, [%r29];
add.ftz.f32 %f622, %f621, %f30;
ld.shared.f32 %f623, [%r31];
add.ftz.f32 %f624, %f623, %f34;
ld.shared.f32 %f625, [%r29+512];
add.ftz.f32 %f626, %f625, %f622;
ld.shared.f32 %f627, [%r31+512];
add.ftz.f32 %f628, %f627, %f624;
ld.shared.f32 %f629, [%r29+1024];
add.ftz.f32 %f630, %f629, %f626;
ld.shared.f32 %f631, [%r31+1024];
add.ftz.f32 %f632, %f631, %f628;
mov.b32 %r979, %f630;
mov.u32 %r980, 31;
mov.u32 %r981, 16;
mov.u32 %r982, -1;
shfl.sync.bfly.b32 %r983|%p171, %r979, %r981, %r980, %r982;
mov.b32 %f633, %r983;
add.ftz.f32 %f634, %f630, %f633;
mov.b32 %r984, %f634;
mov.u32 %r985, 8;
shfl.sync.bfly.b32 %r986|%p172, %r984, %r985, %r980, %r982;
mov.b32 %f635, %r986;
add.ftz.f32 %f636, %f634, %f635;
mov.b32 %r987, %f636;
mov.u32 %r988, 4;
shfl.sync.bfly.b32 %r989|%p173, %r987, %r988, %r980, %r982;
mov.b32 %f637, %r989;
add.ftz.f32 %f638, %f636, %f637;
mov.b32 %r990, %f638;
mov.u32 %r991, 2;
shfl.sync.bfly.b32 %r992|%p174, %r990, %r991, %r980, %r982;
mov.b32 %f639, %r992;
add.ftz.f32 %f640, %f638, %f639;
mov.b32 %r993, %f640;
mov.u32 %r994, 1;
shfl.sync.bfly.b32 %r995|%p175, %r993, %r994, %r980, %r982;
mov.b32 %f641, %r995;
add.ftz.f32 %f56, %f640, %f641;
mov.b32 %r996, %f632;
shfl.sync.bfly.b32 %r997|%p176, %r996, %r981, %r980, %r982;
mov.b32 %f642, %r997;
add.ftz.f32 %f643, %f632, %f642;
mov.b32 %r998, %f643;
shfl.sync.bfly.b32 %r999|%p177, %r998, %r985, %r980, %r982;
mov.b32 %f644, %r999;
add.ftz.f32 %f645, %f643, %f644;
mov.b32 %r1000, %f645;
shfl.sync.bfly.b32 %r1001|%p178, %r1000, %r988, %r980, %r982;
mov.b32 %f646, %r1001;
add.ftz.f32 %f647, %f645, %f646;
mov.b32 %r1002, %f647;
shfl.sync.bfly.b32 %r1003|%p179, %r1002, %r991, %r980, %r982;
mov.b32 %f648, %r1003;
add.ftz.f32 %f649, %f647, %f648;
mov.b32 %r1004, %f649;
shfl.sync.bfly.b32 %r1005|%p180, %r1004, %r994, %r980, %r982;
mov.b32 %f650, %r1005;
add.ftz.f32 %f57, %f649, %f650;
setp.ne.s32 %p181, %r2, 0;
setp.ge.u32 %p182, %r4, %r48;
or.pred %p183, %p181, %p182;
@%p183 bra $L__BB155_38;

add.ftz.f32 %f651, %f5, %f56;
add.ftz.f32 %f652, %f4, %f57;
mul.ftz.f32 %f653, %f652, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f654, %f653;
add.ftz.f32 %f655, %f654, 0f3F800000;
div.approx.ftz.f32 %f656, %f652, %f655;
mul.ftz.f32 %f657, %f651, %f656;
st.global.f32 [%rd35], %f657;

$L__BB155_38:
ld.shared.f32 %f658, [%r29+128];
add.ftz.f32 %f659, %f658, %f29;
ld.shared.f32 %f660, [%r31+128];
add.ftz.f32 %f661, %f660, %f33;
ld.shared.f32 %f662, [%r29+640];
add.ftz.f32 %f663, %f662, %f659;
ld.shared.f32 %f664, [%r31+640];
add.ftz.f32 %f665, %f664, %f661;
ld.shared.f32 %f666, [%r29+1152];
add.ftz.f32 %f667, %f666, %f663;
ld.shared.f32 %f668, [%r31+1152];
add.ftz.f32 %f669, %f668, %f665;
mov.b32 %r1009, %f667;
shfl.sync.bfly.b32 %r1013|%p184, %r1009, %r981, %r980, %r982;
mov.b32 %f670, %r1013;
add.ftz.f32 %f671, %f667, %f670;
mov.b32 %r1014, %f671;
shfl.sync.bfly.b32 %r1016|%p185, %r1014, %r985, %r980, %r982;
mov.b32 %f672, %r1016;
add.ftz.f32 %f673, %f671, %f672;
mov.b32 %r1017, %f673;
shfl.sync.bfly.b32 %r1019|%p186, %r1017, %r988, %r980, %r982;
mov.b32 %f674, %r1019;
add.ftz.f32 %f675, %f673, %f674;
mov.b32 %r1020, %f675;
shfl.sync.bfly.b32 %r1022|%p187, %r1020, %r991, %r980, %r982;
mov.b32 %f676, %r1022;
add.ftz.f32 %f677, %f675, %f676;
mov.b32 %r1023, %f677;
shfl.sync.bfly.b32 %r1025|%p188, %r1023, %r994, %r980, %r982;
mov.b32 %f678, %r1025;
add.ftz.f32 %f58, %f677, %f678;
mov.b32 %r1026, %f669;
shfl.sync.bfly.b32 %r1027|%p189, %r1026, %r981, %r980, %r982;
mov.b32 %f679, %r1027;
add.ftz.f32 %f680, %f669, %f679;
mov.b32 %r1028, %f680;
shfl.sync.bfly.b32 %r1029|%p190, %r1028, %r985, %r980, %r982;
mov.b32 %f681, %r1029;
add.ftz.f32 %f682, %f680, %f681;
mov.b32 %r1030, %f682;
shfl.sync.bfly.b32 %r1031|%p191, %r1030, %r988, %r980, %r982;
mov.b32 %f683, %r1031;
add.ftz.f32 %f684, %f682, %f683;
mov.b32 %r1032, %f684;
shfl.sync.bfly.b32 %r1033|%p192, %r1032, %r991, %r980, %r982;
mov.b32 %f685, %r1033;
add.ftz.f32 %f686, %f684, %f685;
mov.b32 %r1034, %f686;
shfl.sync.bfly.b32 %r1035|%p193, %r1034, %r994, %r980, %r982;
mov.b32 %f687, %r1035;
add.ftz.f32 %f59, %f686, %f687;
setp.ne.s32 %p194, %r2, 1;
or.b32 %r1039, %r4, 1;
setp.ge.u32 %p195, %r1039, %r48;
or.pred %p196, %p194, %p195;
@%p196 bra $L__BB155_40;

add.ftz.f32 %f688, %f5, %f58;
add.ftz.f32 %f689, %f4, %f59;
mul.ftz.f32 %f690, %f689, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f691, %f690;
add.ftz.f32 %f692, %f691, 0f3F800000;
div.approx.ftz.f32 %f693, %f689, %f692;
mul.ftz.f32 %f694, %f688, %f693;
st.global.f32 [%rd35+4], %f694;

$L__BB155_40:
ld.shared.f32 %f695, [%r29+256];
add.ftz.f32 %f696, %f695, %f28;
ld.shared.f32 %f697, [%r31+256];
add.ftz.f32 %f698, %f697, %f32;
ld.shared.f32 %f699, [%r29+768];
add.ftz.f32 %f700, %f699, %f696;
ld.shared.f32 %f701, [%r31+768];
add.ftz.f32 %f702, %f701, %f698;
ld.shared.f32 %f703, [%r29+1280];
add.ftz.f32 %f704, %f703, %f700;
ld.shared.f32 %f705, [%r31+1280];
add.ftz.f32 %f706, %f705, %f702;
mov.b32 %r1040, %f704;
mov.u32 %r1041, 31;
mov.u32 %r1042, 16;
mov.u32 %r1043, -1;
shfl.sync.bfly.b32 %r1044|%p197, %r1040, %r1042, %r1041, %r1043;
mov.b32 %f707, %r1044;
add.ftz.f32 %f708, %f704, %f707;
mov.b32 %r1045, %f708;
mov.u32 %r1046, 8;
shfl.sync.bfly.b32 %r1047|%p198, %r1045, %r1046, %r1041, %r1043;
mov.b32 %f709, %r1047;
add.ftz.f32 %f710, %f708, %f709;
mov.b32 %r1048, %f710;
mov.u32 %r1049, 4;
shfl.sync.bfly.b32 %r1050|%p199, %r1048, %r1049, %r1041, %r1043;
mov.b32 %f711, %r1050;
add.ftz.f32 %f712, %f710, %f711;
mov.b32 %r1051, %f712;
mov.u32 %r1052, 2;
shfl.sync.bfly.b32 %r1053|%p200, %r1051, %r1052, %r1041, %r1043;
mov.b32 %f713, %r1053;
add.ftz.f32 %f714, %f712, %f713;
mov.b32 %r1054, %f714;
mov.u32 %r1055, 1;
shfl.sync.bfly.b32 %r1056|%p201, %r1054, %r1055, %r1041, %r1043;
mov.b32 %f715, %r1056;
add.ftz.f32 %f60, %f714, %f715;
mov.b32 %r1057, %f706;
shfl.sync.bfly.b32 %r1058|%p202, %r1057, %r1042, %r1041, %r1043;
mov.b32 %f716, %r1058;
add.ftz.f32 %f717, %f706, %f716;
mov.b32 %r1059, %f717;
shfl.sync.bfly.b32 %r1060|%p203, %r1059, %r1046, %r1041, %r1043;
mov.b32 %f718, %r1060;
add.ftz.f32 %f719, %f717, %f718;
mov.b32 %r1061, %f719;
shfl.sync.bfly.b32 %r1062|%p204, %r1061, %r1049, %r1041, %r1043;
mov.b32 %f720, %r1062;
add.ftz.f32 %f721, %f719, %f720;
mov.b32 %r1063, %f721;
shfl.sync.bfly.b32 %r1064|%p205, %r1063, %r1052, %r1041, %r1043;
mov.b32 %f722, %r1064;
add.ftz.f32 %f723, %f721, %f722;
mov.b32 %r1065, %f723;
shfl.sync.bfly.b32 %r1066|%p206, %r1065, %r1055, %r1041, %r1043;
mov.b32 %f724, %r1066;
add.ftz.f32 %f61, %f723, %f724;
setp.ne.s32 %p207, %r2, 2;
or.b32 %r1070, %r4, 2;
setp.ge.u32 %p208, %r1070, %r48;
or.pred %p209, %p207, %p208;
@%p209 bra $L__BB155_42;

add.ftz.f32 %f725, %f5, %f60;
add.ftz.f32 %f726, %f4, %f61;
mul.ftz.f32 %f727, %f726, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f728, %f727;
add.ftz.f32 %f729, %f728, 0f3F800000;
div.approx.ftz.f32 %f730, %f726, %f729;
mul.ftz.f32 %f731, %f725, %f730;
st.global.f32 [%rd35+8], %f731;

$L__BB155_42:
ld.shared.f32 %f732, [%r29+384];
add.ftz.f32 %f733, %f732, %f27;
ld.shared.f32 %f734, [%r31+384];
add.ftz.f32 %f735, %f734, %f31;
ld.shared.f32 %f736, [%r29+896];
add.ftz.f32 %f737, %f736, %f733;
ld.shared.f32 %f738, [%r31+896];
add.ftz.f32 %f739, %f738, %f735;
ld.shared.f32 %f740, [%r29+1408];
add.ftz.f32 %f741, %f740, %f737;
ld.shared.f32 %f742, [%r31+1408];
add.ftz.f32 %f743, %f742, %f739;
mov.b32 %r1071, %f741;
shfl.sync.bfly.b32 %r1075|%p210, %r1071, %r1042, %r1041, %r1043;
mov.b32 %f744, %r1075;
add.ftz.f32 %f745, %f741, %f744;
mov.b32 %r1076, %f745;
shfl.sync.bfly.b32 %r1078|%p211, %r1076, %r1046, %r1041, %r1043;
mov.b32 %f746, %r1078;
add.ftz.f32 %f747, %f745, %f746;
mov.b32 %r1079, %f747;
shfl.sync.bfly.b32 %r1081|%p212, %r1079, %r1049, %r1041, %r1043;
mov.b32 %f748, %r1081;
add.ftz.f32 %f749, %f747, %f748;
mov.b32 %r1082, %f749;
shfl.sync.bfly.b32 %r1084|%p213, %r1082, %r1052, %r1041, %r1043;
mov.b32 %f750, %r1084;
add.ftz.f32 %f751, %f749, %f750;
mov.b32 %r1085, %f751;
shfl.sync.bfly.b32 %r1087|%p214, %r1085, %r1055, %r1041, %r1043;
mov.b32 %f752, %r1087;
add.ftz.f32 %f62, %f751, %f752;
mov.b32 %r1088, %f743;
shfl.sync.bfly.b32 %r1089|%p215, %r1088, %r1042, %r1041, %r1043;
mov.b32 %f753, %r1089;
add.ftz.f32 %f754, %f743, %f753;
mov.b32 %r1090, %f754;
shfl.sync.bfly.b32 %r1091|%p216, %r1090, %r1046, %r1041, %r1043;
mov.b32 %f755, %r1091;
add.ftz.f32 %f756, %f754, %f755;
mov.b32 %r1092, %f756;
shfl.sync.bfly.b32 %r1093|%p217, %r1092, %r1049, %r1041, %r1043;
mov.b32 %f757, %r1093;
add.ftz.f32 %f758, %f756, %f757;
mov.b32 %r1094, %f758;
shfl.sync.bfly.b32 %r1095|%p218, %r1094, %r1052, %r1041, %r1043;
mov.b32 %f759, %r1095;
add.ftz.f32 %f760, %f758, %f759;
mov.b32 %r1096, %f760;
shfl.sync.bfly.b32 %r1097|%p219, %r1096, %r1055, %r1041, %r1043;
mov.b32 %f761, %r1097;
add.ftz.f32 %f63, %f760, %f761;
setp.ne.s32 %p220, %r2, 3;
or.b32 %r1101, %r4, 3;
setp.ge.u32 %p221, %r1101, %r48;
or.pred %p222, %p220, %p221;
@%p222 bra $L__BB155_78;

add.ftz.f32 %f762, %f5, %f62;
add.ftz.f32 %f763, %f4, %f63;
mul.ftz.f32 %f764, %f763, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f765, %f764;
add.ftz.f32 %f766, %f765, 0f3F800000;
div.approx.ftz.f32 %f767, %f763, %f766;
mul.ftz.f32 %f768, %f762, %f767;
st.global.f32 [%rd35+12], %f768;
bra.uni $L__BB155_78;

$L__BB155_25:
ld.shared.f32 %f221, [%r29];
add.ftz.f32 %f222, %f221, %f30;
ld.shared.f32 %f223, [%r29+512];
add.ftz.f32 %f224, %f223, %f222;
ld.shared.f32 %f225, [%r29+1024];
add.ftz.f32 %f226, %f225, %f224;
mov.b32 %r646, %f226;
mov.u32 %r647, 31;
mov.u32 %r648, 16;
mov.u32 %r649, -1;
shfl.sync.bfly.b32 %r650|%p30, %r646, %r648, %r647, %r649;
mov.b32 %f227, %r650;
add.ftz.f32 %f228, %f226, %f227;
mov.b32 %r651, %f228;
mov.u32 %r652, 8;
shfl.sync.bfly.b32 %r653|%p31, %r651, %r652, %r647, %r649;
mov.b32 %f229, %r653;
add.ftz.f32 %f230, %f228, %f229;
mov.b32 %r654, %f230;
mov.u32 %r655, 4;
shfl.sync.bfly.b32 %r656|%p32, %r654, %r655, %r647, %r649;
mov.b32 %f231, %r656;
add.ftz.f32 %f232, %f230, %f231;
mov.b32 %r657, %f232;
mov.u32 %r658, 2;
shfl.sync.bfly.b32 %r659|%p33, %r657, %r658, %r647, %r649;
mov.b32 %f233, %r659;
add.ftz.f32 %f234, %f232, %f233;
mov.b32 %r660, %f234;
mov.u32 %r661, 1;
shfl.sync.bfly.b32 %r662|%p34, %r660, %r661, %r647, %r649;
mov.b32 %f235, %r662;
add.ftz.f32 %f36, %f234, %f235;
setp.ne.s32 %p35, %r2, 0;
setp.ge.u32 %p36, %r4, %r48;
or.pred %p37, %p35, %p36;
@%p37 bra $L__BB155_27;

add.ftz.f32 %f236, %f5, %f36;
st.global.f32 [%rd35], %f236;

$L__BB155_27:
ld.shared.f32 %f237, [%r29+128];
add.ftz.f32 %f238, %f237, %f29;
ld.shared.f32 %f239, [%r29+640];
add.ftz.f32 %f240, %f239, %f238;
ld.shared.f32 %f241, [%r29+1152];
add.ftz.f32 %f242, %f241, %f240;
mov.b32 %r666, %f242;
shfl.sync.bfly.b32 %r670|%p38, %r666, %r648, %r647, %r649;
mov.b32 %f243, %r670;
add.ftz.f32 %f244, %f242, %f243;
mov.b32 %r671, %f244;
shfl.sync.bfly.b32 %r673|%p39, %r671, %r652, %r647, %r649;
mov.b32 %f245, %r673;
add.ftz.f32 %f246, %f244, %f245;
mov.b32 %r674, %f246;
shfl.sync.bfly.b32 %r676|%p40, %r674, %r655, %r647, %r649;
mov.b32 %f247, %r676;
add.ftz.f32 %f248, %f246, %f247;
mov.b32 %r677, %f248;
shfl.sync.bfly.b32 %r679|%p41, %r677, %r658, %r647, %r649;
mov.b32 %f249, %r679;
add.ftz.f32 %f250, %f248, %f249;
mov.b32 %r680, %f250;
shfl.sync.bfly.b32 %r682|%p42, %r680, %r661, %r647, %r649;
mov.b32 %f251, %r682;
add.ftz.f32 %f37, %f250, %f251;
setp.ne.s32 %p43, %r2, 1;
or.b32 %r686, %r4, 1;
setp.ge.u32 %p44, %r686, %r48;
or.pred %p45, %p43, %p44;
@%p45 bra $L__BB155_29;

add.ftz.f32 %f252, %f5, %f37;
st.global.f32 [%rd35+4], %f252;

$L__BB155_29:
ld.shared.f32 %f253, [%r29+256];
add.ftz.f32 %f254, %f253, %f28;
ld.shared.f32 %f255, [%r29+768];
add.ftz.f32 %f256, %f255, %f254;
ld.shared.f32 %f257, [%r29+1280];
add.ftz.f32 %f258, %f257, %f256;
mov.b32 %r687, %f258;
mov.u32 %r688, 31;
mov.u32 %r689, 16;
mov.u32 %r690, -1;
shfl.sync.bfly.b32 %r691|%p46, %r687, %r689, %r688, %r690;
mov.b32 %f259, %r691;
add.ftz.f32 %f260, %f258, %f259;
mov.b32 %r692, %f260;
mov.u32 %r693, 8;
shfl.sync.bfly.b32 %r694|%p47, %r692, %r693, %r688, %r690;
mov.b32 %f261, %r694;
add.ftz.f32 %f262, %f260, %f261;
mov.b32 %r695, %f262;
mov.u32 %r696, 4;
shfl.sync.bfly.b32 %r697|%p48, %r695, %r696, %r688, %r690;
mov.b32 %f263, %r697;
add.ftz.f32 %f264, %f262, %f263;
mov.b32 %r698, %f264;
mov.u32 %r699, 2;
shfl.sync.bfly.b32 %r700|%p49, %r698, %r699, %r688, %r690;
mov.b32 %f265, %r700;
add.ftz.f32 %f266, %f264, %f265;
mov.b32 %r701, %f266;
mov.u32 %r702, 1;
shfl.sync.bfly.b32 %r703|%p50, %r701, %r702, %r688, %r690;
mov.b32 %f267, %r703;
add.ftz.f32 %f38, %f266, %f267;
setp.ne.s32 %p51, %r2, 2;
or.b32 %r707, %r4, 2;
setp.ge.u32 %p52, %r707, %r48;
or.pred %p53, %p51, %p52;
@%p53 bra $L__BB155_31;

add.ftz.f32 %f268, %f5, %f38;
st.global.f32 [%rd35+8], %f268;

$L__BB155_31:
ld.shared.f32 %f269, [%r29+384];
add.ftz.f32 %f270, %f269, %f27;
ld.shared.f32 %f271, [%r29+896];
add.ftz.f32 %f272, %f271, %f270;
ld.shared.f32 %f273, [%r29+1408];
add.ftz.f32 %f274, %f273, %f272;
mov.b32 %r708, %f274;
shfl.sync.bfly.b32 %r712|%p54, %r708, %r689, %r688, %r690;
mov.b32 %f275, %r712;
add.ftz.f32 %f276, %f274, %f275;
mov.b32 %r713, %f276;
shfl.sync.bfly.b32 %r715|%p55, %r713, %r693, %r688, %r690;
mov.b32 %f277, %r715;
add.ftz.f32 %f278, %f276, %f277;
mov.b32 %r716, %f278;
shfl.sync.bfly.b32 %r718|%p56, %r716, %r696, %r688, %r690;
mov.b32 %f279, %r718;
add.ftz.f32 %f280, %f278, %f279;
mov.b32 %r719, %f280;
shfl.sync.bfly.b32 %r721|%p57, %r719, %r699, %r688, %r690;
mov.b32 %f281, %r721;
add.ftz.f32 %f282, %f280, %f281;
mov.b32 %r722, %f282;
shfl.sync.bfly.b32 %r724|%p58, %r722, %r702, %r688, %r690;
mov.b32 %f283, %r724;
add.ftz.f32 %f39, %f282, %f283;
setp.ne.s32 %p59, %r2, 3;
or.b32 %r728, %r4, 3;
setp.ge.u32 %p60, %r728, %r48;
or.pred %p61, %p59, %p60;
@%p61 bra $L__BB155_78;

add.ftz.f32 %f284, %f5, %f39;
st.global.f32 [%rd35+12], %f284;
bra.uni $L__BB155_78;

$L__BB155_44:
ld.param.u32 %r1360, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p63, %r1360, 3;
@%p63 bra $L__BB155_54;

ld.param.u32 %r1361, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.ne.s32 %p64, %r1361, 6;
@%p64 bra $L__BB155_70;

ld.shared.f32 %f285, [%r29];
add.ftz.f32 %f286, %f285, %f30;
ld.shared.f32 %f287, [%r31];
add.ftz.f32 %f288, %f287, %f34;
ld.shared.f32 %f289, [%r29+512];
add.ftz.f32 %f290, %f289, %f286;
ld.shared.f32 %f291, [%r31+512];
add.ftz.f32 %f292, %f291, %f288;
ld.shared.f32 %f293, [%r29+1024];
add.ftz.f32 %f294, %f293, %f290;
ld.shared.f32 %f295, [%r31+1024];
add.ftz.f32 %f296, %f295, %f292;
mov.b32 %r733, %f294;
mov.u32 %r734, 31;
mov.u32 %r735, 16;
mov.u32 %r736, -1;
shfl.sync.bfly.b32 %r737|%p67, %r733, %r735, %r734, %r736;
mov.b32 %f297, %r737;
add.ftz.f32 %f298, %f294, %f297;
mov.b32 %r738, %f298;
mov.u32 %r739, 8;
shfl.sync.bfly.b32 %r740|%p68, %r738, %r739, %r734, %r736;
mov.b32 %f299, %r740;
add.ftz.f32 %f300, %f298, %f299;
mov.b32 %r741, %f300;
mov.u32 %r742, 4;
shfl.sync.bfly.b32 %r743|%p69, %r741, %r742, %r734, %r736;
mov.b32 %f301, %r743;
add.ftz.f32 %f302, %f300, %f301;
mov.b32 %r744, %f302;
mov.u32 %r745, 2;
shfl.sync.bfly.b32 %r746|%p70, %r744, %r745, %r734, %r736;
mov.b32 %f303, %r746;
add.ftz.f32 %f304, %f302, %f303;
mov.b32 %r747, %f304;
mov.u32 %r748, 1;
shfl.sync.bfly.b32 %r749|%p71, %r747, %r748, %r734, %r736;
mov.b32 %f305, %r749;
add.ftz.f32 %f40, %f304, %f305;
mov.b32 %r750, %f296;
shfl.sync.bfly.b32 %r751|%p72, %r750, %r735, %r734, %r736;
mov.b32 %f306, %r751;
add.ftz.f32 %f307, %f296, %f306;
mov.b32 %r752, %f307;
shfl.sync.bfly.b32 %r753|%p73, %r752, %r739, %r734, %r736;
mov.b32 %f308, %r753;
add.ftz.f32 %f309, %f307, %f308;
mov.b32 %r754, %f309;
shfl.sync.bfly.b32 %r755|%p74, %r754, %r742, %r734, %r736;
mov.b32 %f310, %r755;
add.ftz.f32 %f311, %f309, %f310;
mov.b32 %r756, %f311;
shfl.sync.bfly.b32 %r757|%p75, %r756, %r745, %r734, %r736;
mov.b32 %f312, %r757;
add.ftz.f32 %f313, %f311, %f312;
mov.b32 %r758, %f313;
shfl.sync.bfly.b32 %r759|%p76, %r758, %r748, %r734, %r736;
mov.b32 %f314, %r759;
add.ftz.f32 %f41, %f313, %f314;
setp.ne.s32 %p77, %r2, 0;
setp.ge.u32 %p78, %r4, %r48;
or.pred %p79, %p77, %p78;
@%p79 bra $L__BB155_48;

ld.param.f32 %f1069, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f315, %f5, %f40;
add.ftz.f32 %f316, %f4, %f41;
min.ftz.f32 %f317, %f316, %f1069;
min.ftz.f32 %f318, %f315, %f1069;
max.ftz.f32 %f319, %f318, %f35;
mul.ftz.f32 %f320, %f317, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f321, %f320;
add.ftz.f32 %f322, %f321, 0f3F800000;
div.approx.ftz.f32 %f323, %f317, %f322;
mul.ftz.f32 %f324, %f319, %f323;
st.global.f32 [%rd35], %f324;

$L__BB155_48:
ld.shared.f32 %f325, [%r29+128];
add.ftz.f32 %f326, %f325, %f29;
ld.shared.f32 %f327, [%r31+128];
add.ftz.f32 %f328, %f327, %f33;
ld.shared.f32 %f329, [%r29+640];
add.ftz.f32 %f330, %f329, %f326;
ld.shared.f32 %f331, [%r31+640];
add.ftz.f32 %f332, %f331, %f328;
ld.shared.f32 %f333, [%r29+1152];
add.ftz.f32 %f334, %f333, %f330;
ld.shared.f32 %f335, [%r31+1152];
add.ftz.f32 %f336, %f335, %f332;
mov.b32 %r763, %f334;
shfl.sync.bfly.b32 %r767|%p80, %r763, %r735, %r734, %r736;
mov.b32 %f337, %r767;
add.ftz.f32 %f338, %f334, %f337;
mov.b32 %r768, %f338;
shfl.sync.bfly.b32 %r770|%p81, %r768, %r739, %r734, %r736;
mov.b32 %f339, %r770;
add.ftz.f32 %f340, %f338, %f339;
mov.b32 %r771, %f340;
shfl.sync.bfly.b32 %r773|%p82, %r771, %r742, %r734, %r736;
mov.b32 %f341, %r773;
add.ftz.f32 %f342, %f340, %f341;
mov.b32 %r774, %f342;
shfl.sync.bfly.b32 %r776|%p83, %r774, %r745, %r734, %r736;
mov.b32 %f343, %r776;
add.ftz.f32 %f344, %f342, %f343;
mov.b32 %r777, %f344;
shfl.sync.bfly.b32 %r779|%p84, %r777, %r748, %r734, %r736;
mov.b32 %f345, %r779;
add.ftz.f32 %f42, %f344, %f345;
mov.b32 %r780, %f336;
shfl.sync.bfly.b32 %r781|%p85, %r780, %r735, %r734, %r736;
mov.b32 %f346, %r781;
add.ftz.f32 %f347, %f336, %f346;
mov.b32 %r782, %f347;
shfl.sync.bfly.b32 %r783|%p86, %r782, %r739, %r734, %r736;
mov.b32 %f348, %r783;
add.ftz.f32 %f349, %f347, %f348;
mov.b32 %r784, %f349;
shfl.sync.bfly.b32 %r785|%p87, %r784, %r742, %r734, %r736;
mov.b32 %f350, %r785;
add.ftz.f32 %f351, %f349, %f350;
mov.b32 %r786, %f351;
shfl.sync.bfly.b32 %r787|%p88, %r786, %r745, %r734, %r736;
mov.b32 %f352, %r787;
add.ftz.f32 %f353, %f351, %f352;
mov.b32 %r788, %f353;
shfl.sync.bfly.b32 %r789|%p89, %r788, %r748, %r734, %r736;
mov.b32 %f354, %r789;
add.ftz.f32 %f43, %f353, %f354;
setp.ne.s32 %p90, %r2, 1;
or.b32 %r793, %r4, 1;
setp.ge.u32 %p91, %r793, %r48;
or.pred %p92, %p90, %p91;
@%p92 bra $L__BB155_50;

ld.param.f32 %f1068, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f355, %f5, %f42;
add.ftz.f32 %f356, %f4, %f43;
min.ftz.f32 %f357, %f356, %f1068;
min.ftz.f32 %f358, %f355, %f1068;
max.ftz.f32 %f359, %f358, %f35;
mul.ftz.f32 %f360, %f357, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f361, %f360;
add.ftz.f32 %f362, %f361, 0f3F800000;
div.approx.ftz.f32 %f363, %f357, %f362;
mul.ftz.f32 %f364, %f359, %f363;
st.global.f32 [%rd35+4], %f364;

$L__BB155_50:
ld.shared.f32 %f365, [%r29+256];
add.ftz.f32 %f366, %f365, %f28;
ld.shared.f32 %f367, [%r31+256];
add.ftz.f32 %f368, %f367, %f32;
ld.shared.f32 %f369, [%r29+768];
add.ftz.f32 %f370, %f369, %f366;
ld.shared.f32 %f371, [%r31+768];
add.ftz.f32 %f372, %f371, %f368;
ld.shared.f32 %f373, [%r29+1280];
add.ftz.f32 %f374, %f373, %f370;
ld.shared.f32 %f375, [%r31+1280];
add.ftz.f32 %f376, %f375, %f372;
mov.b32 %r794, %f374;
mov.u32 %r795, 31;
mov.u32 %r796, 16;
mov.u32 %r797, -1;
shfl.sync.bfly.b32 %r798|%p93, %r794, %r796, %r795, %r797;
mov.b32 %f377, %r798;
add.ftz.f32 %f378, %f374, %f377;
mov.b32 %r799, %f378;
mov.u32 %r800, 8;
shfl.sync.bfly.b32 %r801|%p94, %r799, %r800, %r795, %r797;
mov.b32 %f379, %r801;
add.ftz.f32 %f380, %f378, %f379;
mov.b32 %r802, %f380;
mov.u32 %r803, 4;
shfl.sync.bfly.b32 %r804|%p95, %r802, %r803, %r795, %r797;
mov.b32 %f381, %r804;
add.ftz.f32 %f382, %f380, %f381;
mov.b32 %r805, %f382;
mov.u32 %r806, 2;
shfl.sync.bfly.b32 %r807|%p96, %r805, %r806, %r795, %r797;
mov.b32 %f383, %r807;
add.ftz.f32 %f384, %f382, %f383;
mov.b32 %r808, %f384;
mov.u32 %r809, 1;
shfl.sync.bfly.b32 %r810|%p97, %r808, %r809, %r795, %r797;
mov.b32 %f385, %r810;
add.ftz.f32 %f44, %f384, %f385;
mov.b32 %r811, %f376;
shfl.sync.bfly.b32 %r812|%p98, %r811, %r796, %r795, %r797;
mov.b32 %f386, %r812;
add.ftz.f32 %f387, %f376, %f386;
mov.b32 %r813, %f387;
shfl.sync.bfly.b32 %r814|%p99, %r813, %r800, %r795, %r797;
mov.b32 %f388, %r814;
add.ftz.f32 %f389, %f387, %f388;
mov.b32 %r815, %f389;
shfl.sync.bfly.b32 %r816|%p100, %r815, %r803, %r795, %r797;
mov.b32 %f390, %r816;
add.ftz.f32 %f391, %f389, %f390;
mov.b32 %r817, %f391;
shfl.sync.bfly.b32 %r818|%p101, %r817, %r806, %r795, %r797;
mov.b32 %f392, %r818;
add.ftz.f32 %f393, %f391, %f392;
mov.b32 %r819, %f393;
shfl.sync.bfly.b32 %r820|%p102, %r819, %r809, %r795, %r797;
mov.b32 %f394, %r820;
add.ftz.f32 %f45, %f393, %f394;
setp.ne.s32 %p103, %r2, 2;
or.b32 %r824, %r4, 2;
setp.ge.u32 %p104, %r824, %r48;
or.pred %p105, %p103, %p104;
@%p105 bra $L__BB155_52;

ld.param.f32 %f1067, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f395, %f5, %f44;
add.ftz.f32 %f396, %f4, %f45;
min.ftz.f32 %f397, %f396, %f1067;
min.ftz.f32 %f398, %f395, %f1067;
max.ftz.f32 %f399, %f398, %f35;
mul.ftz.f32 %f400, %f397, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f401, %f400;
add.ftz.f32 %f402, %f401, 0f3F800000;
div.approx.ftz.f32 %f403, %f397, %f402;
mul.ftz.f32 %f404, %f399, %f403;
st.global.f32 [%rd35+8], %f404;

$L__BB155_52:
ld.shared.f32 %f405, [%r29+384];
add.ftz.f32 %f406, %f405, %f27;
ld.shared.f32 %f407, [%r31+384];
add.ftz.f32 %f408, %f407, %f31;
ld.shared.f32 %f409, [%r29+896];
add.ftz.f32 %f410, %f409, %f406;
ld.shared.f32 %f411, [%r31+896];
add.ftz.f32 %f412, %f411, %f408;
ld.shared.f32 %f413, [%r29+1408];
add.ftz.f32 %f414, %f413, %f410;
ld.shared.f32 %f415, [%r31+1408];
add.ftz.f32 %f416, %f415, %f412;
mov.b32 %r825, %f414;
shfl.sync.bfly.b32 %r829|%p106, %r825, %r796, %r795, %r797;
mov.b32 %f417, %r829;
add.ftz.f32 %f418, %f414, %f417;
mov.b32 %r830, %f418;
shfl.sync.bfly.b32 %r832|%p107, %r830, %r800, %r795, %r797;
mov.b32 %f419, %r832;
add.ftz.f32 %f420, %f418, %f419;
mov.b32 %r833, %f420;
shfl.sync.bfly.b32 %r835|%p108, %r833, %r803, %r795, %r797;
mov.b32 %f421, %r835;
add.ftz.f32 %f422, %f420, %f421;
mov.b32 %r836, %f422;
shfl.sync.bfly.b32 %r838|%p109, %r836, %r806, %r795, %r797;
mov.b32 %f423, %r838;
add.ftz.f32 %f424, %f422, %f423;
mov.b32 %r839, %f424;
shfl.sync.bfly.b32 %r841|%p110, %r839, %r809, %r795, %r797;
mov.b32 %f425, %r841;
add.ftz.f32 %f46, %f424, %f425;
mov.b32 %r842, %f416;
shfl.sync.bfly.b32 %r843|%p111, %r842, %r796, %r795, %r797;
mov.b32 %f426, %r843;
add.ftz.f32 %f427, %f416, %f426;
mov.b32 %r844, %f427;
shfl.sync.bfly.b32 %r845|%p112, %r844, %r800, %r795, %r797;
mov.b32 %f428, %r845;
add.ftz.f32 %f429, %f427, %f428;
mov.b32 %r846, %f429;
shfl.sync.bfly.b32 %r847|%p113, %r846, %r803, %r795, %r797;
mov.b32 %f430, %r847;
add.ftz.f32 %f431, %f429, %f430;
mov.b32 %r848, %f431;
shfl.sync.bfly.b32 %r849|%p114, %r848, %r806, %r795, %r797;
mov.b32 %f432, %r849;
add.ftz.f32 %f433, %f431, %f432;
mov.b32 %r850, %f433;
shfl.sync.bfly.b32 %r851|%p115, %r850, %r809, %r795, %r797;
mov.b32 %f434, %r851;
add.ftz.f32 %f47, %f433, %f434;
setp.ne.s32 %p116, %r2, 3;
or.b32 %r855, %r4, 3;
setp.ge.u32 %p117, %r855, %r48;
or.pred %p118, %p116, %p117;
@%p118 bra $L__BB155_78;

ld.param.f32 %f1066, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f435, %f5, %f46;
add.ftz.f32 %f436, %f4, %f47;
min.ftz.f32 %f437, %f436, %f1066;
min.ftz.f32 %f438, %f435, %f1066;
max.ftz.f32 %f439, %f438, %f35;
mul.ftz.f32 %f440, %f437, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f441, %f440;
add.ftz.f32 %f442, %f441, 0f3F800000;
div.approx.ftz.f32 %f443, %f437, %f442;
mul.ftz.f32 %f444, %f439, %f443;
st.global.f32 [%rd35+12], %f444;
bra.uni $L__BB155_78;

$L__BB155_70:
ld.shared.f32 %f933, [%r29];
add.ftz.f32 %f934, %f933, %f30;
ld.shared.f32 %f935, [%r31];
add.ftz.f32 %f936, %f935, %f34;
ld.shared.f32 %f937, [%r29+512];
add.ftz.f32 %f938, %f937, %f934;
ld.shared.f32 %f939, [%r31+512];
add.ftz.f32 %f940, %f939, %f936;
ld.shared.f32 %f941, [%r29+1024];
add.ftz.f32 %f942, %f941, %f938;
ld.shared.f32 %f943, [%r31+1024];
add.ftz.f32 %f944, %f943, %f940;
mov.b32 %r1225, %f942;
mov.u32 %r1226, 31;
mov.u32 %r1227, 16;
mov.u32 %r1228, -1;
shfl.sync.bfly.b32 %r1229|%p275, %r1225, %r1227, %r1226, %r1228;
mov.b32 %f945, %r1229;
add.ftz.f32 %f946, %f942, %f945;
mov.b32 %r1230, %f946;
mov.u32 %r1231, 8;
shfl.sync.bfly.b32 %r1232|%p276, %r1230, %r1231, %r1226, %r1228;
mov.b32 %f947, %r1232;
add.ftz.f32 %f948, %f946, %f947;
mov.b32 %r1233, %f948;
mov.u32 %r1234, 4;
shfl.sync.bfly.b32 %r1235|%p277, %r1233, %r1234, %r1226, %r1228;
mov.b32 %f949, %r1235;
add.ftz.f32 %f950, %f948, %f949;
mov.b32 %r1236, %f950;
mov.u32 %r1237, 2;
shfl.sync.bfly.b32 %r1238|%p278, %r1236, %r1237, %r1226, %r1228;
mov.b32 %f951, %r1238;
add.ftz.f32 %f952, %f950, %f951;
mov.b32 %r1239, %f952;
mov.u32 %r1240, 1;
shfl.sync.bfly.b32 %r1241|%p279, %r1239, %r1240, %r1226, %r1228;
mov.b32 %f953, %r1241;
add.ftz.f32 %f72, %f952, %f953;
mov.b32 %r1242, %f944;
shfl.sync.bfly.b32 %r1243|%p280, %r1242, %r1227, %r1226, %r1228;
mov.b32 %f954, %r1243;
add.ftz.f32 %f955, %f944, %f954;
mov.b32 %r1244, %f955;
shfl.sync.bfly.b32 %r1245|%p281, %r1244, %r1231, %r1226, %r1228;
mov.b32 %f956, %r1245;
add.ftz.f32 %f957, %f955, %f956;
mov.b32 %r1246, %f957;
shfl.sync.bfly.b32 %r1247|%p282, %r1246, %r1234, %r1226, %r1228;
mov.b32 %f958, %r1247;
add.ftz.f32 %f959, %f957, %f958;
mov.b32 %r1248, %f959;
shfl.sync.bfly.b32 %r1249|%p283, %r1248, %r1237, %r1226, %r1228;
mov.b32 %f960, %r1249;
add.ftz.f32 %f961, %f959, %f960;
mov.b32 %r1250, %f961;
shfl.sync.bfly.b32 %r1251|%p284, %r1250, %r1240, %r1226, %r1228;
mov.b32 %f962, %r1251;
add.ftz.f32 %f73, %f961, %f962;
setp.ne.s32 %p285, %r2, 0;
setp.ge.u32 %p286, %r4, %r48;
or.pred %p287, %p285, %p286;
@%p287 bra $L__BB155_72;

add.ftz.f32 %f963, %f5, %f72;
add.ftz.f32 %f964, %f4, %f73;
mul.ftz.f32 %f965, %f963, %f964;
st.global.f32 [%rd35], %f965;

$L__BB155_72:
ld.shared.f32 %f966, [%r29+128];
add.ftz.f32 %f967, %f966, %f29;
ld.shared.f32 %f968, [%r31+128];
add.ftz.f32 %f969, %f968, %f33;
ld.shared.f32 %f970, [%r29+640];
add.ftz.f32 %f971, %f970, %f967;
ld.shared.f32 %f972, [%r31+640];
add.ftz.f32 %f973, %f972, %f969;
ld.shared.f32 %f974, [%r29+1152];
add.ftz.f32 %f975, %f974, %f971;
ld.shared.f32 %f976, [%r31+1152];
add.ftz.f32 %f977, %f976, %f973;
mov.b32 %r1255, %f975;
shfl.sync.bfly.b32 %r1259|%p288, %r1255, %r1227, %r1226, %r1228;
mov.b32 %f978, %r1259;
add.ftz.f32 %f979, %f975, %f978;
mov.b32 %r1260, %f979;
shfl.sync.bfly.b32 %r1262|%p289, %r1260, %r1231, %r1226, %r1228;
mov.b32 %f980, %r1262;
add.ftz.f32 %f981, %f979, %f980;
mov.b32 %r1263, %f981;
shfl.sync.bfly.b32 %r1265|%p290, %r1263, %r1234, %r1226, %r1228;
mov.b32 %f982, %r1265;
add.ftz.f32 %f983, %f981, %f982;
mov.b32 %r1266, %f983;
shfl.sync.bfly.b32 %r1268|%p291, %r1266, %r1237, %r1226, %r1228;
mov.b32 %f984, %r1268;
add.ftz.f32 %f985, %f983, %f984;
mov.b32 %r1269, %f985;
shfl.sync.bfly.b32 %r1271|%p292, %r1269, %r1240, %r1226, %r1228;
mov.b32 %f986, %r1271;
add.ftz.f32 %f74, %f985, %f986;
mov.b32 %r1272, %f977;
shfl.sync.bfly.b32 %r1273|%p293, %r1272, %r1227, %r1226, %r1228;
mov.b32 %f987, %r1273;
add.ftz.f32 %f988, %f977, %f987;
mov.b32 %r1274, %f988;
shfl.sync.bfly.b32 %r1275|%p294, %r1274, %r1231, %r1226, %r1228;
mov.b32 %f989, %r1275;
add.ftz.f32 %f990, %f988, %f989;
mov.b32 %r1276, %f990;
shfl.sync.bfly.b32 %r1277|%p295, %r1276, %r1234, %r1226, %r1228;
mov.b32 %f991, %r1277;
add.ftz.f32 %f992, %f990, %f991;
mov.b32 %r1278, %f992;
shfl.sync.bfly.b32 %r1279|%p296, %r1278, %r1237, %r1226, %r1228;
mov.b32 %f993, %r1279;
add.ftz.f32 %f994, %f992, %f993;
mov.b32 %r1280, %f994;
shfl.sync.bfly.b32 %r1281|%p297, %r1280, %r1240, %r1226, %r1228;
mov.b32 %f995, %r1281;
add.ftz.f32 %f75, %f994, %f995;
setp.ne.s32 %p298, %r2, 1;
or.b32 %r1285, %r4, 1;
setp.ge.u32 %p299, %r1285, %r48;
or.pred %p300, %p298, %p299;
@%p300 bra $L__BB155_74;

add.ftz.f32 %f996, %f5, %f74;
add.ftz.f32 %f997, %f4, %f75;
mul.ftz.f32 %f998, %f996, %f997;
st.global.f32 [%rd35+4], %f998;

$L__BB155_74:
ld.shared.f32 %f999, [%r29+256];
add.ftz.f32 %f1000, %f999, %f28;
ld.shared.f32 %f1001, [%r31+256];
add.ftz.f32 %f1002, %f1001, %f32;
ld.shared.f32 %f1003, [%r29+768];
add.ftz.f32 %f1004, %f1003, %f1000;
ld.shared.f32 %f1005, [%r31+768];
add.ftz.f32 %f1006, %f1005, %f1002;
ld.shared.f32 %f1007, [%r29+1280];
add.ftz.f32 %f1008, %f1007, %f1004;
ld.shared.f32 %f1009, [%r31+1280];
add.ftz.f32 %f1010, %f1009, %f1006;
mov.b32 %r1286, %f1008;
mov.u32 %r1287, 31;
mov.u32 %r1288, 16;
mov.u32 %r1289, -1;
shfl.sync.bfly.b32 %r1290|%p301, %r1286, %r1288, %r1287, %r1289;
mov.b32 %f1011, %r1290;
add.ftz.f32 %f1012, %f1008, %f1011;
mov.b32 %r1291, %f1012;
mov.u32 %r1292, 8;
shfl.sync.bfly.b32 %r1293|%p302, %r1291, %r1292, %r1287, %r1289;
mov.b32 %f1013, %r1293;
add.ftz.f32 %f1014, %f1012, %f1013;
mov.b32 %r1294, %f1014;
mov.u32 %r1295, 4;
shfl.sync.bfly.b32 %r1296|%p303, %r1294, %r1295, %r1287, %r1289;
mov.b32 %f1015, %r1296;
add.ftz.f32 %f1016, %f1014, %f1015;
mov.b32 %r1297, %f1016;
mov.u32 %r1298, 2;
shfl.sync.bfly.b32 %r1299|%p304, %r1297, %r1298, %r1287, %r1289;
mov.b32 %f1017, %r1299;
add.ftz.f32 %f1018, %f1016, %f1017;
mov.b32 %r1300, %f1018;
mov.u32 %r1301, 1;
shfl.sync.bfly.b32 %r1302|%p305, %r1300, %r1301, %r1287, %r1289;
mov.b32 %f1019, %r1302;
add.ftz.f32 %f76, %f1018, %f1019;
mov.b32 %r1303, %f1010;
shfl.sync.bfly.b32 %r1304|%p306, %r1303, %r1288, %r1287, %r1289;
mov.b32 %f1020, %r1304;
add.ftz.f32 %f1021, %f1010, %f1020;
mov.b32 %r1305, %f1021;
shfl.sync.bfly.b32 %r1306|%p307, %r1305, %r1292, %r1287, %r1289;
mov.b32 %f1022, %r1306;
add.ftz.f32 %f1023, %f1021, %f1022;
mov.b32 %r1307, %f1023;
shfl.sync.bfly.b32 %r1308|%p308, %r1307, %r1295, %r1287, %r1289;
mov.b32 %f1024, %r1308;
add.ftz.f32 %f1025, %f1023, %f1024;
mov.b32 %r1309, %f1025;
shfl.sync.bfly.b32 %r1310|%p309, %r1309, %r1298, %r1287, %r1289;
mov.b32 %f1026, %r1310;
add.ftz.f32 %f1027, %f1025, %f1026;
mov.b32 %r1311, %f1027;
shfl.sync.bfly.b32 %r1312|%p310, %r1311, %r1301, %r1287, %r1289;
mov.b32 %f1028, %r1312;
add.ftz.f32 %f77, %f1027, %f1028;
setp.ne.s32 %p311, %r2, 2;
or.b32 %r1316, %r4, 2;
setp.ge.u32 %p312, %r1316, %r48;
or.pred %p313, %p311, %p312;
@%p313 bra $L__BB155_76;

add.ftz.f32 %f1029, %f5, %f76;
add.ftz.f32 %f1030, %f4, %f77;
mul.ftz.f32 %f1031, %f1029, %f1030;
st.global.f32 [%rd35+8], %f1031;

$L__BB155_76:
ld.shared.f32 %f1032, [%r29+384];
add.ftz.f32 %f1033, %f1032, %f27;
ld.shared.f32 %f1034, [%r31+384];
add.ftz.f32 %f1035, %f1034, %f31;
ld.shared.f32 %f1036, [%r29+896];
add.ftz.f32 %f1037, %f1036, %f1033;
ld.shared.f32 %f1038, [%r31+896];
add.ftz.f32 %f1039, %f1038, %f1035;
ld.shared.f32 %f1040, [%r29+1408];
add.ftz.f32 %f1041, %f1040, %f1037;
ld.shared.f32 %f1042, [%r31+1408];
add.ftz.f32 %f1043, %f1042, %f1039;
mov.b32 %r1317, %f1041;
shfl.sync.bfly.b32 %r1321|%p314, %r1317, %r1288, %r1287, %r1289;
mov.b32 %f1044, %r1321;
add.ftz.f32 %f1045, %f1041, %f1044;
mov.b32 %r1322, %f1045;
shfl.sync.bfly.b32 %r1324|%p315, %r1322, %r1292, %r1287, %r1289;
mov.b32 %f1046, %r1324;
add.ftz.f32 %f1047, %f1045, %f1046;
mov.b32 %r1325, %f1047;
shfl.sync.bfly.b32 %r1327|%p316, %r1325, %r1295, %r1287, %r1289;
mov.b32 %f1048, %r1327;
add.ftz.f32 %f1049, %f1047, %f1048;
mov.b32 %r1328, %f1049;
shfl.sync.bfly.b32 %r1330|%p317, %r1328, %r1298, %r1287, %r1289;
mov.b32 %f1050, %r1330;
add.ftz.f32 %f1051, %f1049, %f1050;
mov.b32 %r1331, %f1051;
shfl.sync.bfly.b32 %r1333|%p318, %r1331, %r1301, %r1287, %r1289;
mov.b32 %f1052, %r1333;
add.ftz.f32 %f78, %f1051, %f1052;
mov.b32 %r1334, %f1043;
shfl.sync.bfly.b32 %r1335|%p319, %r1334, %r1288, %r1287, %r1289;
mov.b32 %f1053, %r1335;
add.ftz.f32 %f1054, %f1043, %f1053;
mov.b32 %r1336, %f1054;
shfl.sync.bfly.b32 %r1337|%p320, %r1336, %r1292, %r1287, %r1289;
mov.b32 %f1055, %r1337;
add.ftz.f32 %f1056, %f1054, %f1055;
mov.b32 %r1338, %f1056;
shfl.sync.bfly.b32 %r1339|%p321, %r1338, %r1295, %r1287, %r1289;
mov.b32 %f1057, %r1339;
add.ftz.f32 %f1058, %f1056, %f1057;
mov.b32 %r1340, %f1058;
shfl.sync.bfly.b32 %r1341|%p322, %r1340, %r1298, %r1287, %r1289;
mov.b32 %f1059, %r1341;
add.ftz.f32 %f1060, %f1058, %f1059;
mov.b32 %r1342, %f1060;
shfl.sync.bfly.b32 %r1343|%p323, %r1342, %r1301, %r1287, %r1289;
mov.b32 %f1061, %r1343;
add.ftz.f32 %f79, %f1060, %f1061;
setp.ne.s32 %p324, %r2, 3;
or.b32 %r1347, %r4, 3;
setp.ge.u32 %p325, %r1347, %r48;
or.pred %p326, %p324, %p325;
@%p326 bra $L__BB155_78;

add.ftz.f32 %f1062, %f5, %f78;
add.ftz.f32 %f1063, %f4, %f79;
mul.ftz.f32 %f1064, %f1062, %f1063;
st.global.f32 [%rd35+12], %f1064;
bra.uni $L__BB155_78;

$L__BB155_62:
ld.shared.f32 %f769, [%r29];
add.ftz.f32 %f770, %f769, %f30;
ld.shared.f32 %f771, [%r31];
add.ftz.f32 %f772, %f771, %f34;
ld.shared.f32 %f773, [%r29+512];
add.ftz.f32 %f774, %f773, %f770;
ld.shared.f32 %f775, [%r31+512];
add.ftz.f32 %f776, %f775, %f772;
ld.shared.f32 %f777, [%r29+1024];
add.ftz.f32 %f778, %f777, %f774;
ld.shared.f32 %f779, [%r31+1024];
add.ftz.f32 %f780, %f779, %f776;
mov.b32 %r1102, %f778;
mov.u32 %r1103, 31;
mov.u32 %r1104, 16;
mov.u32 %r1105, -1;
shfl.sync.bfly.b32 %r1106|%p223, %r1102, %r1104, %r1103, %r1105;
mov.b32 %f781, %r1106;
add.ftz.f32 %f782, %f778, %f781;
mov.b32 %r1107, %f782;
mov.u32 %r1108, 8;
shfl.sync.bfly.b32 %r1109|%p224, %r1107, %r1108, %r1103, %r1105;
mov.b32 %f783, %r1109;
add.ftz.f32 %f784, %f782, %f783;
mov.b32 %r1110, %f784;
mov.u32 %r1111, 4;
shfl.sync.bfly.b32 %r1112|%p225, %r1110, %r1111, %r1103, %r1105;
mov.b32 %f785, %r1112;
add.ftz.f32 %f786, %f784, %f785;
mov.b32 %r1113, %f786;
mov.u32 %r1114, 2;
shfl.sync.bfly.b32 %r1115|%p226, %r1113, %r1114, %r1103, %r1105;
mov.b32 %f787, %r1115;
add.ftz.f32 %f788, %f786, %f787;
mov.b32 %r1116, %f788;
mov.u32 %r1117, 1;
shfl.sync.bfly.b32 %r1118|%p227, %r1116, %r1117, %r1103, %r1105;
mov.b32 %f789, %r1118;
add.ftz.f32 %f64, %f788, %f789;
mov.b32 %r1119, %f780;
shfl.sync.bfly.b32 %r1120|%p228, %r1119, %r1104, %r1103, %r1105;
mov.b32 %f790, %r1120;
add.ftz.f32 %f791, %f780, %f790;
mov.b32 %r1121, %f791;
shfl.sync.bfly.b32 %r1122|%p229, %r1121, %r1108, %r1103, %r1105;
mov.b32 %f792, %r1122;
add.ftz.f32 %f793, %f791, %f792;
mov.b32 %r1123, %f793;
shfl.sync.bfly.b32 %r1124|%p230, %r1123, %r1111, %r1103, %r1105;
mov.b32 %f794, %r1124;
add.ftz.f32 %f795, %f793, %f794;
mov.b32 %r1125, %f795;
shfl.sync.bfly.b32 %r1126|%p231, %r1125, %r1114, %r1103, %r1105;
mov.b32 %f796, %r1126;
add.ftz.f32 %f797, %f795, %f796;
mov.b32 %r1127, %f797;
shfl.sync.bfly.b32 %r1128|%p232, %r1127, %r1117, %r1103, %r1105;
mov.b32 %f798, %r1128;
add.ftz.f32 %f65, %f797, %f798;
setp.ne.s32 %p233, %r2, 0;
setp.ge.u32 %p234, %r4, %r48;
or.pred %p235, %p233, %p234;
@%p235 bra $L__BB155_64;

add.ftz.f32 %f799, %f5, %f64;
add.ftz.f32 %f800, %f4, %f65;
mul.ftz.f32 %f801, %f800, 0f3F000000;
mul.ftz.f32 %f802, %f800, 0f3F4C422A;
mul.ftz.f32 %f803, %f800, 0f3D372713;
fma.rn.ftz.f32 %f804, %f800, %f803, 0f3F800000;
mul.ftz.f32 %f805, %f802, %f804;
tanh.approx.f32 %f806, %f805;
add.ftz.f32 %f807, %f806, 0f3F800000;
mul.ftz.f32 %f808, %f801, %f807;
mul.ftz.f32 %f809, %f799, %f808;
st.global.f32 [%rd35], %f809;

$L__BB155_64:
ld.shared.f32 %f810, [%r29+128];
add.ftz.f32 %f811, %f810, %f29;
ld.shared.f32 %f812, [%r31+128];
add.ftz.f32 %f813, %f812, %f33;
ld.shared.f32 %f814, [%r29+640];
add.ftz.f32 %f815, %f814, %f811;
ld.shared.f32 %f816, [%r31+640];
add.ftz.f32 %f817, %f816, %f813;
ld.shared.f32 %f818, [%r29+1152];
add.ftz.f32 %f819, %f818, %f815;
ld.shared.f32 %f820, [%r31+1152];
add.ftz.f32 %f821, %f820, %f817;
mov.b32 %r1132, %f819;
shfl.sync.bfly.b32 %r1136|%p236, %r1132, %r1104, %r1103, %r1105;
mov.b32 %f822, %r1136;
add.ftz.f32 %f823, %f819, %f822;
mov.b32 %r1137, %f823;
shfl.sync.bfly.b32 %r1139|%p237, %r1137, %r1108, %r1103, %r1105;
mov.b32 %f824, %r1139;
add.ftz.f32 %f825, %f823, %f824;
mov.b32 %r1140, %f825;
shfl.sync.bfly.b32 %r1142|%p238, %r1140, %r1111, %r1103, %r1105;
mov.b32 %f826, %r1142;
add.ftz.f32 %f827, %f825, %f826;
mov.b32 %r1143, %f827;
shfl.sync.bfly.b32 %r1145|%p239, %r1143, %r1114, %r1103, %r1105;
mov.b32 %f828, %r1145;
add.ftz.f32 %f829, %f827, %f828;
mov.b32 %r1146, %f829;
shfl.sync.bfly.b32 %r1148|%p240, %r1146, %r1117, %r1103, %r1105;
mov.b32 %f830, %r1148;
add.ftz.f32 %f66, %f829, %f830;
mov.b32 %r1149, %f821;
shfl.sync.bfly.b32 %r1150|%p241, %r1149, %r1104, %r1103, %r1105;
mov.b32 %f831, %r1150;
add.ftz.f32 %f832, %f821, %f831;
mov.b32 %r1151, %f832;
shfl.sync.bfly.b32 %r1152|%p242, %r1151, %r1108, %r1103, %r1105;
mov.b32 %f833, %r1152;
add.ftz.f32 %f834, %f832, %f833;
mov.b32 %r1153, %f834;
shfl.sync.bfly.b32 %r1154|%p243, %r1153, %r1111, %r1103, %r1105;
mov.b32 %f835, %r1154;
add.ftz.f32 %f836, %f834, %f835;
mov.b32 %r1155, %f836;
shfl.sync.bfly.b32 %r1156|%p244, %r1155, %r1114, %r1103, %r1105;
mov.b32 %f837, %r1156;
add.ftz.f32 %f838, %f836, %f837;
mov.b32 %r1157, %f838;
shfl.sync.bfly.b32 %r1158|%p245, %r1157, %r1117, %r1103, %r1105;
mov.b32 %f839, %r1158;
add.ftz.f32 %f67, %f838, %f839;
setp.ne.s32 %p246, %r2, 1;
or.b32 %r1162, %r4, 1;
setp.ge.u32 %p247, %r1162, %r48;
or.pred %p248, %p246, %p247;
@%p248 bra $L__BB155_66;

add.ftz.f32 %f840, %f5, %f66;
add.ftz.f32 %f841, %f4, %f67;
mul.ftz.f32 %f842, %f841, 0f3F000000;
mul.ftz.f32 %f843, %f841, 0f3F4C422A;
mul.ftz.f32 %f844, %f841, 0f3D372713;
fma.rn.ftz.f32 %f845, %f841, %f844, 0f3F800000;
mul.ftz.f32 %f846, %f843, %f845;
tanh.approx.f32 %f847, %f846;
add.ftz.f32 %f848, %f847, 0f3F800000;
mul.ftz.f32 %f849, %f842, %f848;
mul.ftz.f32 %f850, %f840, %f849;
st.global.f32 [%rd35+4], %f850;

$L__BB155_66:
ld.shared.f32 %f851, [%r29+256];
add.ftz.f32 %f852, %f851, %f28;
ld.shared.f32 %f853, [%r31+256];
add.ftz.f32 %f854, %f853, %f32;
ld.shared.f32 %f855, [%r29+768];
add.ftz.f32 %f856, %f855, %f852;
ld.shared.f32 %f857, [%r31+768];
add.ftz.f32 %f858, %f857, %f854;
ld.shared.f32 %f859, [%r29+1280];
add.ftz.f32 %f860, %f859, %f856;
ld.shared.f32 %f861, [%r31+1280];
add.ftz.f32 %f862, %f861, %f858;
mov.b32 %r1163, %f860;
mov.u32 %r1164, 31;
mov.u32 %r1165, 16;
mov.u32 %r1166, -1;
shfl.sync.bfly.b32 %r1167|%p249, %r1163, %r1165, %r1164, %r1166;
mov.b32 %f863, %r1167;
add.ftz.f32 %f864, %f860, %f863;
mov.b32 %r1168, %f864;
mov.u32 %r1169, 8;
shfl.sync.bfly.b32 %r1170|%p250, %r1168, %r1169, %r1164, %r1166;
mov.b32 %f865, %r1170;
add.ftz.f32 %f866, %f864, %f865;
mov.b32 %r1171, %f866;
mov.u32 %r1172, 4;
shfl.sync.bfly.b32 %r1173|%p251, %r1171, %r1172, %r1164, %r1166;
mov.b32 %f867, %r1173;
add.ftz.f32 %f868, %f866, %f867;
mov.b32 %r1174, %f868;
mov.u32 %r1175, 2;
shfl.sync.bfly.b32 %r1176|%p252, %r1174, %r1175, %r1164, %r1166;
mov.b32 %f869, %r1176;
add.ftz.f32 %f870, %f868, %f869;
mov.b32 %r1177, %f870;
mov.u32 %r1178, 1;
shfl.sync.bfly.b32 %r1179|%p253, %r1177, %r1178, %r1164, %r1166;
mov.b32 %f871, %r1179;
add.ftz.f32 %f68, %f870, %f871;
mov.b32 %r1180, %f862;
shfl.sync.bfly.b32 %r1181|%p254, %r1180, %r1165, %r1164, %r1166;
mov.b32 %f872, %r1181;
add.ftz.f32 %f873, %f862, %f872;
mov.b32 %r1182, %f873;
shfl.sync.bfly.b32 %r1183|%p255, %r1182, %r1169, %r1164, %r1166;
mov.b32 %f874, %r1183;
add.ftz.f32 %f875, %f873, %f874;
mov.b32 %r1184, %f875;
shfl.sync.bfly.b32 %r1185|%p256, %r1184, %r1172, %r1164, %r1166;
mov.b32 %f876, %r1185;
add.ftz.f32 %f877, %f875, %f876;
mov.b32 %r1186, %f877;
shfl.sync.bfly.b32 %r1187|%p257, %r1186, %r1175, %r1164, %r1166;
mov.b32 %f878, %r1187;
add.ftz.f32 %f879, %f877, %f878;
mov.b32 %r1188, %f879;
shfl.sync.bfly.b32 %r1189|%p258, %r1188, %r1178, %r1164, %r1166;
mov.b32 %f880, %r1189;
add.ftz.f32 %f69, %f879, %f880;
setp.ne.s32 %p259, %r2, 2;
or.b32 %r1193, %r4, 2;
setp.ge.u32 %p260, %r1193, %r48;
or.pred %p261, %p259, %p260;
@%p261 bra $L__BB155_68;

add.ftz.f32 %f881, %f5, %f68;
add.ftz.f32 %f882, %f4, %f69;
mul.ftz.f32 %f883, %f882, 0f3F000000;
mul.ftz.f32 %f884, %f882, 0f3F4C422A;
mul.ftz.f32 %f885, %f882, 0f3D372713;
fma.rn.ftz.f32 %f886, %f882, %f885, 0f3F800000;
mul.ftz.f32 %f887, %f884, %f886;
tanh.approx.f32 %f888, %f887;
add.ftz.f32 %f889, %f888, 0f3F800000;
mul.ftz.f32 %f890, %f883, %f889;
mul.ftz.f32 %f891, %f881, %f890;
st.global.f32 [%rd35+8], %f891;

$L__BB155_68:
ld.shared.f32 %f892, [%r29+384];
add.ftz.f32 %f893, %f892, %f27;
ld.shared.f32 %f894, [%r31+384];
add.ftz.f32 %f895, %f894, %f31;
ld.shared.f32 %f896, [%r29+896];
add.ftz.f32 %f897, %f896, %f893;
ld.shared.f32 %f898, [%r31+896];
add.ftz.f32 %f899, %f898, %f895;
ld.shared.f32 %f900, [%r29+1408];
add.ftz.f32 %f901, %f900, %f897;
ld.shared.f32 %f902, [%r31+1408];
add.ftz.f32 %f903, %f902, %f899;
mov.b32 %r1194, %f901;
shfl.sync.bfly.b32 %r1198|%p262, %r1194, %r1165, %r1164, %r1166;
mov.b32 %f904, %r1198;
add.ftz.f32 %f905, %f901, %f904;
mov.b32 %r1199, %f905;
shfl.sync.bfly.b32 %r1201|%p263, %r1199, %r1169, %r1164, %r1166;
mov.b32 %f906, %r1201;
add.ftz.f32 %f907, %f905, %f906;
mov.b32 %r1202, %f907;
shfl.sync.bfly.b32 %r1204|%p264, %r1202, %r1172, %r1164, %r1166;
mov.b32 %f908, %r1204;
add.ftz.f32 %f909, %f907, %f908;
mov.b32 %r1205, %f909;
shfl.sync.bfly.b32 %r1207|%p265, %r1205, %r1175, %r1164, %r1166;
mov.b32 %f910, %r1207;
add.ftz.f32 %f911, %f909, %f910;
mov.b32 %r1208, %f911;
shfl.sync.bfly.b32 %r1210|%p266, %r1208, %r1178, %r1164, %r1166;
mov.b32 %f912, %r1210;
add.ftz.f32 %f70, %f911, %f912;
mov.b32 %r1211, %f903;
shfl.sync.bfly.b32 %r1212|%p267, %r1211, %r1165, %r1164, %r1166;
mov.b32 %f913, %r1212;
add.ftz.f32 %f914, %f903, %f913;
mov.b32 %r1213, %f914;
shfl.sync.bfly.b32 %r1214|%p268, %r1213, %r1169, %r1164, %r1166;
mov.b32 %f915, %r1214;
add.ftz.f32 %f916, %f914, %f915;
mov.b32 %r1215, %f916;
shfl.sync.bfly.b32 %r1216|%p269, %r1215, %r1172, %r1164, %r1166;
mov.b32 %f917, %r1216;
add.ftz.f32 %f918, %f916, %f917;
mov.b32 %r1217, %f918;
shfl.sync.bfly.b32 %r1218|%p270, %r1217, %r1175, %r1164, %r1166;
mov.b32 %f919, %r1218;
add.ftz.f32 %f920, %f918, %f919;
mov.b32 %r1219, %f920;
shfl.sync.bfly.b32 %r1220|%p271, %r1219, %r1178, %r1164, %r1166;
mov.b32 %f921, %r1220;
add.ftz.f32 %f71, %f920, %f921;
setp.ne.s32 %p272, %r2, 3;
or.b32 %r1224, %r4, 3;
setp.ge.u32 %p273, %r1224, %r48;
or.pred %p274, %p272, %p273;
@%p274 bra $L__BB155_78;

add.ftz.f32 %f922, %f5, %f70;
add.ftz.f32 %f923, %f4, %f71;
mul.ftz.f32 %f924, %f923, 0f3F000000;
mul.ftz.f32 %f925, %f923, 0f3F4C422A;
mul.ftz.f32 %f926, %f923, 0f3D372713;
fma.rn.ftz.f32 %f927, %f923, %f926, 0f3F800000;
mul.ftz.f32 %f928, %f925, %f927;
tanh.approx.f32 %f929, %f928;
add.ftz.f32 %f930, %f929, 0f3F800000;
mul.ftz.f32 %f931, %f924, %f930;
mul.ftz.f32 %f932, %f922, %f931;
st.global.f32 [%rd35+12], %f932;
bra.uni $L__BB155_78;

$L__BB155_54:
ld.shared.f32 %f445, [%r29];
add.ftz.f32 %f446, %f445, %f30;
ld.shared.f32 %f447, [%r31];
add.ftz.f32 %f448, %f447, %f34;
ld.shared.f32 %f449, [%r29+512];
add.ftz.f32 %f450, %f449, %f446;
ld.shared.f32 %f451, [%r31+512];
add.ftz.f32 %f452, %f451, %f448;
ld.shared.f32 %f453, [%r29+1024];
add.ftz.f32 %f454, %f453, %f450;
ld.shared.f32 %f455, [%r31+1024];
add.ftz.f32 %f456, %f455, %f452;
mov.b32 %r856, %f454;
mov.u32 %r857, 31;
mov.u32 %r858, 16;
mov.u32 %r859, -1;
shfl.sync.bfly.b32 %r860|%p119, %r856, %r858, %r857, %r859;
mov.b32 %f457, %r860;
add.ftz.f32 %f458, %f454, %f457;
mov.b32 %r861, %f458;
mov.u32 %r862, 8;
shfl.sync.bfly.b32 %r863|%p120, %r861, %r862, %r857, %r859;
mov.b32 %f459, %r863;
add.ftz.f32 %f460, %f458, %f459;
mov.b32 %r864, %f460;
mov.u32 %r865, 4;
shfl.sync.bfly.b32 %r866|%p121, %r864, %r865, %r857, %r859;
mov.b32 %f461, %r866;
add.ftz.f32 %f462, %f460, %f461;
mov.b32 %r867, %f462;
mov.u32 %r868, 2;
shfl.sync.bfly.b32 %r869|%p122, %r867, %r868, %r857, %r859;
mov.b32 %f463, %r869;
add.ftz.f32 %f464, %f462, %f463;
mov.b32 %r870, %f464;
mov.u32 %r871, 1;
shfl.sync.bfly.b32 %r872|%p123, %r870, %r871, %r857, %r859;
mov.b32 %f465, %r872;
add.ftz.f32 %f48, %f464, %f465;
mov.b32 %r873, %f456;
shfl.sync.bfly.b32 %r874|%p124, %r873, %r858, %r857, %r859;
mov.b32 %f466, %r874;
add.ftz.f32 %f467, %f456, %f466;
mov.b32 %r875, %f467;
shfl.sync.bfly.b32 %r876|%p125, %r875, %r862, %r857, %r859;
mov.b32 %f468, %r876;
add.ftz.f32 %f469, %f467, %f468;
mov.b32 %r877, %f469;
shfl.sync.bfly.b32 %r878|%p126, %r877, %r865, %r857, %r859;
mov.b32 %f470, %r878;
add.ftz.f32 %f471, %f469, %f470;
mov.b32 %r879, %f471;
shfl.sync.bfly.b32 %r880|%p127, %r879, %r868, %r857, %r859;
mov.b32 %f472, %r880;
add.ftz.f32 %f473, %f471, %f472;
mov.b32 %r881, %f473;
shfl.sync.bfly.b32 %r882|%p128, %r881, %r871, %r857, %r859;
mov.b32 %f474, %r882;
add.ftz.f32 %f49, %f473, %f474;
setp.ne.s32 %p129, %r2, 0;
setp.ge.u32 %p130, %r4, %r48;
or.pred %p131, %p129, %p130;
@%p131 bra $L__BB155_56;

add.ftz.f32 %f475, %f5, %f48;
add.ftz.f32 %f476, %f4, %f49;
mov.f32 %f477, 0f40E00000;
min.ftz.f32 %f478, %f476, %f477;
min.ftz.f32 %f479, %f475, %f477;
mov.f32 %f480, 0fC0E00000;
max.ftz.f32 %f481, %f479, %f480;
mul.ftz.f32 %f482, %f478, 0f3FD9DB23;
mul.ftz.f32 %f483, %f482, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f484, %f483;
add.ftz.f32 %f485, %f484, 0f3F800000;
div.approx.ftz.f32 %f486, %f478, %f485;
add.ftz.f32 %f487, %f481, 0f3F800000;
mul.ftz.f32 %f488, %f487, %f486;
st.global.f32 [%rd35], %f488;

$L__BB155_56:
ld.shared.f32 %f489, [%r29+128];
add.ftz.f32 %f490, %f489, %f29;
ld.shared.f32 %f491, [%r31+128];
add.ftz.f32 %f492, %f491, %f33;
ld.shared.f32 %f493, [%r29+640];
add.ftz.f32 %f494, %f493, %f490;
ld.shared.f32 %f495, [%r31+640];
add.ftz.f32 %f496, %f495, %f492;
ld.shared.f32 %f497, [%r29+1152];
add.ftz.f32 %f498, %f497, %f494;
ld.shared.f32 %f499, [%r31+1152];
add.ftz.f32 %f500, %f499, %f496;
mov.b32 %r886, %f498;
shfl.sync.bfly.b32 %r890|%p132, %r886, %r858, %r857, %r859;
mov.b32 %f501, %r890;
add.ftz.f32 %f502, %f498, %f501;
mov.b32 %r891, %f502;
shfl.sync.bfly.b32 %r893|%p133, %r891, %r862, %r857, %r859;
mov.b32 %f503, %r893;
add.ftz.f32 %f504, %f502, %f503;
mov.b32 %r894, %f504;
shfl.sync.bfly.b32 %r896|%p134, %r894, %r865, %r857, %r859;
mov.b32 %f505, %r896;
add.ftz.f32 %f506, %f504, %f505;
mov.b32 %r897, %f506;
shfl.sync.bfly.b32 %r899|%p135, %r897, %r868, %r857, %r859;
mov.b32 %f507, %r899;
add.ftz.f32 %f508, %f506, %f507;
mov.b32 %r900, %f508;
shfl.sync.bfly.b32 %r902|%p136, %r900, %r871, %r857, %r859;
mov.b32 %f509, %r902;
add.ftz.f32 %f50, %f508, %f509;
mov.b32 %r903, %f500;
shfl.sync.bfly.b32 %r904|%p137, %r903, %r858, %r857, %r859;
mov.b32 %f510, %r904;
add.ftz.f32 %f511, %f500, %f510;
mov.b32 %r905, %f511;
shfl.sync.bfly.b32 %r906|%p138, %r905, %r862, %r857, %r859;
mov.b32 %f512, %r906;
add.ftz.f32 %f513, %f511, %f512;
mov.b32 %r907, %f513;
shfl.sync.bfly.b32 %r908|%p139, %r907, %r865, %r857, %r859;
mov.b32 %f514, %r908;
add.ftz.f32 %f515, %f513, %f514;
mov.b32 %r909, %f515;
shfl.sync.bfly.b32 %r910|%p140, %r909, %r868, %r857, %r859;
mov.b32 %f516, %r910;
add.ftz.f32 %f517, %f515, %f516;
mov.b32 %r911, %f517;
shfl.sync.bfly.b32 %r912|%p141, %r911, %r871, %r857, %r859;
mov.b32 %f518, %r912;
add.ftz.f32 %f51, %f517, %f518;
setp.ne.s32 %p142, %r2, 1;
or.b32 %r916, %r4, 1;
setp.ge.u32 %p143, %r916, %r48;
or.pred %p144, %p142, %p143;
@%p144 bra $L__BB155_58;

add.ftz.f32 %f519, %f5, %f50;
add.ftz.f32 %f520, %f4, %f51;
mov.f32 %f521, 0f40E00000;
min.ftz.f32 %f522, %f520, %f521;
min.ftz.f32 %f523, %f519, %f521;
mov.f32 %f524, 0fC0E00000;
max.ftz.f32 %f525, %f523, %f524;
mul.ftz.f32 %f526, %f522, 0f3FD9DB23;
mul.ftz.f32 %f527, %f526, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f528, %f527;
add.ftz.f32 %f529, %f528, 0f3F800000;
div.approx.ftz.f32 %f530, %f522, %f529;
add.ftz.f32 %f531, %f525, 0f3F800000;
mul.ftz.f32 %f532, %f531, %f530;
st.global.f32 [%rd35+4], %f532;

$L__BB155_58:
ld.shared.f32 %f533, [%r29+256];
add.ftz.f32 %f534, %f533, %f28;
ld.shared.f32 %f535, [%r31+256];
add.ftz.f32 %f536, %f535, %f32;
ld.shared.f32 %f537, [%r29+768];
add.ftz.f32 %f538, %f537, %f534;
ld.shared.f32 %f539, [%r31+768];
add.ftz.f32 %f540, %f539, %f536;
ld.shared.f32 %f541, [%r29+1280];
add.ftz.f32 %f542, %f541, %f538;
ld.shared.f32 %f543, [%r31+1280];
add.ftz.f32 %f544, %f543, %f540;
mov.b32 %r917, %f542;
mov.u32 %r918, 31;
mov.u32 %r919, 16;
mov.u32 %r920, -1;
shfl.sync.bfly.b32 %r921|%p145, %r917, %r919, %r918, %r920;
mov.b32 %f545, %r921;
add.ftz.f32 %f546, %f542, %f545;
mov.b32 %r922, %f546;
mov.u32 %r923, 8;
shfl.sync.bfly.b32 %r924|%p146, %r922, %r923, %r918, %r920;
mov.b32 %f547, %r924;
add.ftz.f32 %f548, %f546, %f547;
mov.b32 %r925, %f548;
mov.u32 %r926, 4;
shfl.sync.bfly.b32 %r927|%p147, %r925, %r926, %r918, %r920;
mov.b32 %f549, %r927;
add.ftz.f32 %f550, %f548, %f549;
mov.b32 %r928, %f550;
mov.u32 %r929, 2;
shfl.sync.bfly.b32 %r930|%p148, %r928, %r929, %r918, %r920;
mov.b32 %f551, %r930;
add.ftz.f32 %f552, %f550, %f551;
mov.b32 %r931, %f552;
mov.u32 %r932, 1;
shfl.sync.bfly.b32 %r933|%p149, %r931, %r932, %r918, %r920;
mov.b32 %f553, %r933;
add.ftz.f32 %f52, %f552, %f553;
mov.b32 %r934, %f544;
shfl.sync.bfly.b32 %r935|%p150, %r934, %r919, %r918, %r920;
mov.b32 %f554, %r935;
add.ftz.f32 %f555, %f544, %f554;
mov.b32 %r936, %f555;
shfl.sync.bfly.b32 %r937|%p151, %r936, %r923, %r918, %r920;
mov.b32 %f556, %r937;
add.ftz.f32 %f557, %f555, %f556;
mov.b32 %r938, %f557;
shfl.sync.bfly.b32 %r939|%p152, %r938, %r926, %r918, %r920;
mov.b32 %f558, %r939;
add.ftz.f32 %f559, %f557, %f558;
mov.b32 %r940, %f559;
shfl.sync.bfly.b32 %r941|%p153, %r940, %r929, %r918, %r920;
mov.b32 %f560, %r941;
add.ftz.f32 %f561, %f559, %f560;
mov.b32 %r942, %f561;
shfl.sync.bfly.b32 %r943|%p154, %r942, %r932, %r918, %r920;
mov.b32 %f562, %r943;
add.ftz.f32 %f53, %f561, %f562;
setp.ne.s32 %p155, %r2, 2;
or.b32 %r947, %r4, 2;
setp.ge.u32 %p156, %r947, %r48;
or.pred %p157, %p155, %p156;
@%p157 bra $L__BB155_60;

add.ftz.f32 %f563, %f5, %f52;
add.ftz.f32 %f564, %f4, %f53;
mov.f32 %f565, 0f40E00000;
min.ftz.f32 %f566, %f564, %f565;
min.ftz.f32 %f567, %f563, %f565;
mov.f32 %f568, 0fC0E00000;
max.ftz.f32 %f569, %f567, %f568;
mul.ftz.f32 %f570, %f566, 0f3FD9DB23;
mul.ftz.f32 %f571, %f570, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f572, %f571;
add.ftz.f32 %f573, %f572, 0f3F800000;
div.approx.ftz.f32 %f574, %f566, %f573;
add.ftz.f32 %f575, %f569, 0f3F800000;
mul.ftz.f32 %f576, %f575, %f574;
st.global.f32 [%rd35+8], %f576;

$L__BB155_60:
ld.shared.f32 %f577, [%r29+384];
add.ftz.f32 %f578, %f577, %f27;
ld.shared.f32 %f579, [%r31+384];
add.ftz.f32 %f580, %f579, %f31;
ld.shared.f32 %f581, [%r29+896];
add.ftz.f32 %f582, %f581, %f578;
ld.shared.f32 %f583, [%r31+896];
add.ftz.f32 %f584, %f583, %f580;
ld.shared.f32 %f585, [%r29+1408];
add.ftz.f32 %f586, %f585, %f582;
ld.shared.f32 %f587, [%r31+1408];
add.ftz.f32 %f588, %f587, %f584;
mov.b32 %r948, %f586;
shfl.sync.bfly.b32 %r952|%p158, %r948, %r919, %r918, %r920;
mov.b32 %f589, %r952;
add.ftz.f32 %f590, %f586, %f589;
mov.b32 %r953, %f590;
shfl.sync.bfly.b32 %r955|%p159, %r953, %r923, %r918, %r920;
mov.b32 %f591, %r955;
add.ftz.f32 %f592, %f590, %f591;
mov.b32 %r956, %f592;
shfl.sync.bfly.b32 %r958|%p160, %r956, %r926, %r918, %r920;
mov.b32 %f593, %r958;
add.ftz.f32 %f594, %f592, %f593;
mov.b32 %r959, %f594;
shfl.sync.bfly.b32 %r961|%p161, %r959, %r929, %r918, %r920;
mov.b32 %f595, %r961;
add.ftz.f32 %f596, %f594, %f595;
mov.b32 %r962, %f596;
shfl.sync.bfly.b32 %r964|%p162, %r962, %r932, %r918, %r920;
mov.b32 %f597, %r964;
add.ftz.f32 %f54, %f596, %f597;
mov.b32 %r965, %f588;
shfl.sync.bfly.b32 %r966|%p163, %r965, %r919, %r918, %r920;
mov.b32 %f598, %r966;
add.ftz.f32 %f599, %f588, %f598;
mov.b32 %r967, %f599;
shfl.sync.bfly.b32 %r968|%p164, %r967, %r923, %r918, %r920;
mov.b32 %f600, %r968;
add.ftz.f32 %f601, %f599, %f600;
mov.b32 %r969, %f601;
shfl.sync.bfly.b32 %r970|%p165, %r969, %r926, %r918, %r920;
mov.b32 %f602, %r970;
add.ftz.f32 %f603, %f601, %f602;
mov.b32 %r971, %f603;
shfl.sync.bfly.b32 %r972|%p166, %r971, %r929, %r918, %r920;
mov.b32 %f604, %r972;
add.ftz.f32 %f605, %f603, %f604;
mov.b32 %r973, %f605;
shfl.sync.bfly.b32 %r974|%p167, %r973, %r932, %r918, %r920;
mov.b32 %f606, %r974;
add.ftz.f32 %f55, %f605, %f606;
setp.ne.s32 %p168, %r2, 3;
or.b32 %r978, %r4, 3;
setp.ge.u32 %p169, %r978, %r48;
or.pred %p170, %p168, %p169;
@%p170 bra $L__BB155_78;

add.ftz.f32 %f607, %f5, %f54;
add.ftz.f32 %f608, %f4, %f55;
mov.f32 %f609, 0f40E00000;
min.ftz.f32 %f610, %f608, %f609;
min.ftz.f32 %f611, %f607, %f609;
mov.f32 %f612, 0fC0E00000;
max.ftz.f32 %f613, %f611, %f612;
mul.ftz.f32 %f614, %f610, 0f3FD9DB23;
mul.ftz.f32 %f615, %f614, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f616, %f615;
add.ftz.f32 %f617, %f616, 0f3F800000;
div.approx.ftz.f32 %f618, %f610, %f617;
add.ftz.f32 %f619, %f613, 0f3F800000;
mul.ftz.f32 %f620, %f619, %f618;
st.global.f32 [%rd35+12], %f620;
bra.uni $L__BB155_78;

}
.entry _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<39>;
.reg .b16 %rs<29>;
.reg .f32 %f<153>;
.reg .b32 %r<415>;
.reg .b64 %rd<51>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

ld.param.u64 %rd12, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd13, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd14, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
mov.u32 %r1, %tid.y;
shl.b32 %r39, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r39, %r2;
mov.u32 %r40, %ctaid.x;
shl.b32 %r4, %r40, 2;
shr.u32 %r5, %r38, 8;
mov.u32 %r412, %ctaid.y;
setp.eq.s64 %p1, %rd14, 0;
@%p1 bra $L__BB156_2;

cvta.to.global.u64 %rd16, %rd14;
mul.wide.u32 %rd17, %r412, 4;
add.s64 %rd18, %rd16, %rd17;
ld.global.nc.u32 %r411, [%rd18];
mul.hi.u32 %r41, %r412, %r21;
add.s32 %r42, %r41, %r412;
shr.u32 %r43, %r42, %r22;
mul.lo.s32 %r44, %r43, %r23;
sub.s32 %r412, %r412, %r44;
bra.uni $L__BB156_3;

$L__BB156_2:
mul.hi.u32 %r45, %r412, %r26;
add.s32 %r46, %r45, %r412;
shr.u32 %r411, %r46, %r27;

$L__BB156_3:
shr.s32 %r47, %r3, 31;
shr.u32 %r48, %r47, 28;
add.s32 %r49, %r3, %r48;
shr.s32 %r414, %r49, 4;
setp.ge.s32 %p2, %r414, %r5;
mov.f32 %f149, 0f00000000;
mov.f32 %f150, %f149;
mov.f32 %f151, %f149;
mov.f32 %f152, %f149;
@%p2 bra $L__BB156_6;

mov.u32 %r50, %ctaid.z;
mul.lo.s32 %r51, %r412, %r30;
cvt.u64.u32 %rd19, %r51;
and.b32 %r55, %r49, -16;
sub.s32 %r13, %r3, %r55;
shr.s32 %r56, %r13, 31;
shr.u32 %r57, %r56, 30;
add.s32 %r58, %r13, %r57;
shr.s32 %r59, %r58, 2;
shl.b32 %r60, %r59, 1;
shl.b32 %r61, %r59, 5;
and.b32 %r62, %r58, -4;
sub.s32 %r63, %r13, %r62;
shl.b32 %r64, %r63, 2;
add.s32 %r65, %r61, %r64;
cvt.s64.s32 %rd20, %r65;
cvt.s64.s32 %rd21, %r60;
mul.lo.s32 %r66, %r50, %r36;
cvt.u64.u32 %rd22, %r66;
add.s64 %rd23, %rd22, %rd21;
add.s64 %rd1, %rd23, %rd19;
cvt.s64.s32 %rd2, %r63;
shl.b32 %r413, %r414, 3;
mad.lo.s32 %r67, %r4, %r24, %r414;
mul.hi.u32 %r68, %r50, %r32;
add.s32 %r69, %r68, %r50;
shr.u32 %r70, %r69, %r33;
mad.lo.s32 %r71, %r70, %r35, %r67;
mad.lo.s32 %r72, %r411, %r29, %r71;
mul.wide.s32 %rd3, %r72, 144;
add.s64 %rd24, %rd3, %rd20;
add.s64 %rd4, %rd24, 32;
cvt.u64.u32 %rd25, %r60;
and.b64 %rd26, %rd25, 2;
or.b64 %rd27, %rd26, %rd3;
or.b64 %rd5, %rd27, 12;
mul.wide.s32 %rd6, %r24, 144;
cvta.to.global.u64 %rd7, %rd13;
cvta.to.global.u64 %rd50, %rd12;
mov.f32 %f149, 0f00000000;
shl.b64 %rd36, %rd2, 2;

$L__BB156_5:
add.s64 %rd28, %rd50, %rd4;
add.s64 %rd29, %rd28, -16;
ld.global.nc.u32 %r217, [%rd28+-16];
ld.global.nc.u32 %r218, [%rd28];
add.s64 %rd30, %rd50, %rd5;
add.s64 %rd31, %rd30, -8;
ld.global.nc.u16 %rs1, [%rd30+-8];
cvt.u32.u16 %r219, %rs1;
ld.global.nc.u16 %rs2, [%rd30+-4];
cvt.u32.u16 %r220, %rs2;
ld.global.nc.u16 %rs3, [%rd30];
cvt.u32.u16 %r221, %rs3;
and.b32 %r222, %r219, 16191;
and.b32 %r223, %r221, 3855;
and.b32 %r224, %r219, 49344;
shr.u32 %r225, %r224, 2;
or.b32 %r226, %r223, %r225;
setp.gt.s32 %p3, %r13, 7;
and.b32 %r227, %r220, 16191;
shr.u32 %r228, %r221, 4;
and.b32 %r229, %r228, 3855;
and.b32 %r230, %r220, 49344;
shr.u32 %r231, %r230, 2;
or.b32 %r232, %r229, %r231;
cvt.u64.u32 %rd32, %r413;
add.s64 %rd33, %rd1, %rd32;
mul.lo.s64 %rd34, %rd33, 36;
add.s64 %rd35, %rd7, %rd34;
ld.global.nc.u32 %r73, [%rd35];

{.reg .f16 low,high;
mov.b32 {low,high},%r73;
cvt.f32.f16 %f25, low;}


add.s64 %rd37, %rd35, %rd36;
ld.global.nc.u32 %r77, [%rd37+4];
ld.global.nc.u32 %r81, [%rd37+20];
ld.global.nc.u32 %r74, [%rd35+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r74;
cvt.f32.f16 %f26, low;}


ld.global.nc.u32 %r93, [%rd37+40];
ld.global.nc.u32 %r97, [%rd37+56];
add.s64 %rd38, %rd50, %rd3;
and.b32 %r76, %r217, 252645135;
and.b32 %r80, %r218, 252645135;
mov.u32 %r210, 0;

dp4a.s32.s32 %r75, %r76, %r77, %r210;


dp4a.s32.s32 %r79, %r80, %r81, %r75;

mov.u32 %r212, 16843009;

dp4a.s32.s32 %r83, %r212, %r77, %r210;


dp4a.s32.s32 %r87, %r212, %r81, %r83;

selp.b32 %r233, %r226, %r222, %p3;
cvt.u16.u32 %rs4, %r233;
and.b32 %r234, %r233, 63;
mul.lo.s32 %r235, %r79, %r234;
cvt.rn.f32.s32 %f41, %r235;
fma.rn.ftz.f32 %f42, %f25, %f41, 0f00000000;
selp.b32 %r236, %r232, %r227, %p3;
cvt.u16.u32 %rs5, %r236;
and.b32 %r237, %r236, 63;
mul.lo.s32 %r238, %r87, %r237;
cvt.rn.f32.s32 %f43, %r238;
fma.rn.ftz.f32 %f44, %f25, %f43, 0f00000000;
shr.u32 %r239, %r217, 4;
and.b32 %r92, %r239, 252645135;
shr.u32 %r240, %r218, 4;
and.b32 %r96, %r240, 252645135;

dp4a.s32.s32 %r91, %r92, %r93, %r210;


dp4a.s32.s32 %r95, %r96, %r97, %r91;


dp4a.s32.s32 %r99, %r212, %r93, %r210;


dp4a.s32.s32 %r103, %r212, %r97, %r99;

shr.u16 %rs6, %rs4, 8;
cvt.u32.u16 %r241, %rs6;
mul.lo.s32 %r242, %r95, %r241;
cvt.rn.f32.s32 %f45, %r242;
fma.rn.ftz.f32 %f46, %f26, %f45, %f42;
shr.u16 %rs7, %rs5, 8;
cvt.u32.u16 %r243, %rs7;
mul.lo.s32 %r244, %r103, %r243;
cvt.rn.f32.s32 %f47, %r244;
fma.rn.ftz.f32 %f48, %f26, %f47, %f44;
ld.global.nc.u32 %r107, [%rd38];

{.reg .f16 low,high;
mov.b32 {low,high},%r107;
cvt.f32.f16 %f27, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r107;
cvt.f32.f16 %f28, high;}


mul.ftz.f32 %f49, %f46, %f27;
mul.ftz.f32 %f50, %f48, %f28;
sub.ftz.f32 %f51, %f49, %f50;
add.ftz.f32 %f152, %f51, %f152;
add.s64 %rd39, %rd29, %rd6;
ld.global.nc.u32 %r245, [%rd39];
ld.global.nc.u32 %r246, [%rd39+16];
add.s64 %rd40, %rd31, %rd6;
ld.global.nc.u16 %rs8, [%rd40];
cvt.u32.u16 %r247, %rs8;
ld.global.nc.u16 %rs9, [%rd40+4];
cvt.u32.u16 %r248, %rs9;
ld.global.nc.u16 %rs10, [%rd40+8];
cvt.u32.u16 %r249, %rs10;
and.b32 %r250, %r247, 16191;
and.b32 %r251, %r249, 3855;
and.b32 %r252, %r247, 49344;
shr.u32 %r253, %r252, 2;
or.b32 %r254, %r251, %r253;
and.b32 %r255, %r248, 16191;
shr.u32 %r256, %r249, 4;
and.b32 %r257, %r256, 3855;
and.b32 %r258, %r248, 49344;
shr.u32 %r259, %r258, 2;
or.b32 %r260, %r257, %r259;

{.reg .f16 low,high;
mov.b32 {low,high},%r73;
cvt.f32.f16 %f29, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r74;
cvt.f32.f16 %f30, low;}


and.b32 %r112, %r245, 252645135;
and.b32 %r116, %r246, 252645135;

dp4a.s32.s32 %r111, %r112, %r77, %r210;


dp4a.s32.s32 %r115, %r116, %r81, %r111;


dp4a.s32.s32 %r119, %r212, %r77, %r210;


dp4a.s32.s32 %r123, %r212, %r81, %r119;

selp.b32 %r261, %r254, %r250, %p3;
cvt.u16.u32 %rs11, %r261;
and.b32 %r262, %r261, 63;
mul.lo.s32 %r263, %r115, %r262;
cvt.rn.f32.s32 %f52, %r263;
fma.rn.ftz.f32 %f53, %f29, %f52, 0f00000000;
selp.b32 %r264, %r260, %r255, %p3;
cvt.u16.u32 %rs12, %r264;
and.b32 %r265, %r264, 63;
mul.lo.s32 %r266, %r123, %r265;
cvt.rn.f32.s32 %f54, %r266;
fma.rn.ftz.f32 %f55, %f29, %f54, 0f00000000;
shr.u32 %r267, %r245, 4;
and.b32 %r128, %r267, 252645135;
shr.u32 %r268, %r246, 4;
and.b32 %r132, %r268, 252645135;

dp4a.s32.s32 %r127, %r128, %r93, %r210;


dp4a.s32.s32 %r131, %r132, %r97, %r127;


dp4a.s32.s32 %r135, %r212, %r93, %r210;


dp4a.s32.s32 %r139, %r212, %r97, %r135;

shr.u16 %rs13, %rs11, 8;
cvt.u32.u16 %r269, %rs13;
mul.lo.s32 %r270, %r131, %r269;
cvt.rn.f32.s32 %f56, %r270;
fma.rn.ftz.f32 %f57, %f30, %f56, %f53;
shr.u16 %rs14, %rs12, 8;
cvt.u32.u16 %r271, %rs14;
mul.lo.s32 %r272, %r139, %r271;
cvt.rn.f32.s32 %f58, %r272;
fma.rn.ftz.f32 %f59, %f30, %f58, %f55;
add.s64 %rd41, %rd38, %rd6;
ld.global.nc.u32 %r143, [%rd41];

{.reg .f16 low,high;
mov.b32 {low,high},%r143;
cvt.f32.f16 %f31, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r143;
cvt.f32.f16 %f32, high;}


mul.ftz.f32 %f60, %f57, %f31;
mul.ftz.f32 %f61, %f59, %f32;
sub.ftz.f32 %f62, %f60, %f61;
add.ftz.f32 %f151, %f62, %f151;
add.s64 %rd42, %rd39, %rd6;
ld.global.nc.u32 %r273, [%rd42];
ld.global.nc.u32 %r274, [%rd42+16];
add.s64 %rd43, %rd40, %rd6;
ld.global.nc.u16 %rs15, [%rd43];
cvt.u32.u16 %r275, %rs15;
ld.global.nc.u16 %rs16, [%rd43+4];
cvt.u32.u16 %r276, %rs16;
ld.global.nc.u16 %rs17, [%rd43+8];
cvt.u32.u16 %r277, %rs17;
and.b32 %r278, %r275, 16191;
and.b32 %r279, %r277, 3855;
and.b32 %r280, %r275, 49344;
shr.u32 %r281, %r280, 2;
or.b32 %r282, %r279, %r281;
and.b32 %r283, %r276, 16191;
shr.u32 %r284, %r277, 4;
and.b32 %r285, %r284, 3855;
and.b32 %r286, %r276, 49344;
shr.u32 %r287, %r286, 2;
or.b32 %r288, %r285, %r287;

{.reg .f16 low,high;
mov.b32 {low,high},%r73;
cvt.f32.f16 %f33, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r74;
cvt.f32.f16 %f34, low;}


and.b32 %r148, %r273, 252645135;
and.b32 %r152, %r274, 252645135;

dp4a.s32.s32 %r147, %r148, %r77, %r210;


dp4a.s32.s32 %r151, %r152, %r81, %r147;


dp4a.s32.s32 %r155, %r212, %r77, %r210;


dp4a.s32.s32 %r159, %r212, %r81, %r155;

selp.b32 %r289, %r282, %r278, %p3;
cvt.u16.u32 %rs18, %r289;
and.b32 %r290, %r289, 63;
mul.lo.s32 %r291, %r151, %r290;
cvt.rn.f32.s32 %f63, %r291;
fma.rn.ftz.f32 %f64, %f33, %f63, 0f00000000;
selp.b32 %r292, %r288, %r283, %p3;
cvt.u16.u32 %rs19, %r292;
and.b32 %r293, %r292, 63;
mul.lo.s32 %r294, %r159, %r293;
cvt.rn.f32.s32 %f65, %r294;
fma.rn.ftz.f32 %f66, %f33, %f65, 0f00000000;
shr.u32 %r295, %r273, 4;
and.b32 %r164, %r295, 252645135;
shr.u32 %r296, %r274, 4;
and.b32 %r168, %r296, 252645135;

dp4a.s32.s32 %r163, %r164, %r93, %r210;


dp4a.s32.s32 %r167, %r168, %r97, %r163;


dp4a.s32.s32 %r171, %r212, %r93, %r210;


dp4a.s32.s32 %r175, %r212, %r97, %r171;

shr.u16 %rs20, %rs18, 8;
cvt.u32.u16 %r297, %rs20;
mul.lo.s32 %r298, %r167, %r297;
cvt.rn.f32.s32 %f67, %r298;
fma.rn.ftz.f32 %f68, %f34, %f67, %f64;
shr.u16 %rs21, %rs19, 8;
cvt.u32.u16 %r299, %rs21;
mul.lo.s32 %r300, %r175, %r299;
cvt.rn.f32.s32 %f69, %r300;
fma.rn.ftz.f32 %f70, %f34, %f69, %f66;
add.s64 %rd44, %rd41, %rd6;
ld.global.nc.u32 %r179, [%rd44];

{.reg .f16 low,high;
mov.b32 {low,high},%r179;
cvt.f32.f16 %f35, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r179;
cvt.f32.f16 %f36, high;}


mul.ftz.f32 %f71, %f68, %f35;
mul.ftz.f32 %f72, %f70, %f36;
sub.ftz.f32 %f73, %f71, %f72;
add.ftz.f32 %f150, %f73, %f150;
add.s64 %rd45, %rd42, %rd6;
ld.global.nc.u32 %r301, [%rd45];
ld.global.nc.u32 %r302, [%rd45+16];
add.s64 %rd46, %rd43, %rd6;
ld.global.nc.u16 %rs22, [%rd46];
cvt.u32.u16 %r303, %rs22;
ld.global.nc.u16 %rs23, [%rd46+4];
cvt.u32.u16 %r304, %rs23;
ld.global.nc.u16 %rs24, [%rd46+8];
cvt.u32.u16 %r305, %rs24;
and.b32 %r306, %r303, 16191;
and.b32 %r307, %r305, 3855;
and.b32 %r308, %r303, 49344;
shr.u32 %r309, %r308, 2;
or.b32 %r310, %r307, %r309;
and.b32 %r311, %r304, 16191;
shr.u32 %r312, %r305, 4;
and.b32 %r313, %r312, 3855;
and.b32 %r314, %r304, 49344;
shr.u32 %r315, %r314, 2;
or.b32 %r316, %r313, %r315;

{.reg .f16 low,high;
mov.b32 {low,high},%r73;
cvt.f32.f16 %f37, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r74;
cvt.f32.f16 %f38, low;}


and.b32 %r184, %r301, 252645135;
and.b32 %r188, %r302, 252645135;

dp4a.s32.s32 %r183, %r184, %r77, %r210;


dp4a.s32.s32 %r187, %r188, %r81, %r183;


dp4a.s32.s32 %r191, %r212, %r77, %r210;


dp4a.s32.s32 %r195, %r212, %r81, %r191;

selp.b32 %r317, %r310, %r306, %p3;
cvt.u16.u32 %rs25, %r317;
and.b32 %r318, %r317, 63;
mul.lo.s32 %r319, %r187, %r318;
cvt.rn.f32.s32 %f74, %r319;
fma.rn.ftz.f32 %f75, %f37, %f74, 0f00000000;
selp.b32 %r320, %r316, %r311, %p3;
cvt.u16.u32 %rs26, %r320;
and.b32 %r321, %r320, 63;
mul.lo.s32 %r322, %r195, %r321;
cvt.rn.f32.s32 %f76, %r322;
fma.rn.ftz.f32 %f77, %f37, %f76, 0f00000000;
shr.u32 %r323, %r301, 4;
and.b32 %r200, %r323, 252645135;
shr.u32 %r324, %r302, 4;
and.b32 %r204, %r324, 252645135;

dp4a.s32.s32 %r199, %r200, %r93, %r210;


dp4a.s32.s32 %r203, %r204, %r97, %r199;


dp4a.s32.s32 %r207, %r212, %r93, %r210;


dp4a.s32.s32 %r211, %r212, %r97, %r207;

shr.u16 %rs27, %rs25, 8;
cvt.u32.u16 %r325, %rs27;
mul.lo.s32 %r326, %r203, %r325;
cvt.rn.f32.s32 %f78, %r326;
fma.rn.ftz.f32 %f79, %f38, %f78, %f75;
shr.u16 %rs28, %rs26, 8;
cvt.u32.u16 %r327, %rs28;
mul.lo.s32 %r328, %r211, %r327;
cvt.rn.f32.s32 %f80, %r328;
fma.rn.ftz.f32 %f81, %f38, %f80, %f77;
add.s64 %rd47, %rd44, %rd6;
ld.global.nc.u32 %r215, [%rd47];

{.reg .f16 low,high;
mov.b32 {low,high},%r215;
cvt.f32.f16 %f39, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r215;
cvt.f32.f16 %f40, high;}


mul.ftz.f32 %f82, %f79, %f39;
mul.ftz.f32 %f83, %f81, %f40;
sub.ftz.f32 %f84, %f82, %f83;
add.ftz.f32 %f149, %f84, %f149;
add.s32 %r413, %r413, 64;
add.s64 %rd50, %rd50, 1152;
add.s32 %r414, %r414, 8;
setp.lt.s32 %p4, %r414, %r5;
@%p4 bra $L__BB156_5;

$L__BB156_6:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB156_8;

shl.b32 %r329, %r1, 9;
mov.u32 %r330, _ZZ13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r331, %r330, %r329;
shl.b32 %r332, %r2, 2;
add.s32 %r333, %r331, %r332;
st.shared.f32 [%r333+-512], %f152;
st.shared.f32 [%r333+-384], %f151;
st.shared.f32 [%r333+-256], %f150;
st.shared.f32 [%r333+-128], %f149;

$L__BB156_8:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB156_17;

mov.u32 %r334, %ctaid.z;
mov.u32 %r335, %ctaid.y;
mad.lo.s32 %r336, %r335, %r31, %r4;
mad.lo.s32 %r337, %r334, %r37, %r336;
shl.b32 %r338, %r2, 2;
mov.u32 %r339, _ZZ13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r19, %r339, %r338;
mov.u32 %r340, 2;
ld.shared.f32 %f85, [%r19];
add.ftz.f32 %f86, %f85, %f152;
ld.shared.f32 %f87, [%r19+512];
add.ftz.f32 %f88, %f87, %f86;
ld.shared.f32 %f89, [%r19+1024];
add.ftz.f32 %f90, %f89, %f88;
mov.b32 %r341, %f90;
mov.u32 %r342, 31;
mov.u32 %r343, 16;
mov.u32 %r344, -1;
shfl.sync.bfly.b32 %r345|%p7, %r341, %r343, %r342, %r344;
mov.b32 %f91, %r345;
add.ftz.f32 %f92, %f90, %f91;
mov.b32 %r346, %f92;
mov.u32 %r347, 8;
shfl.sync.bfly.b32 %r348|%p8, %r346, %r347, %r342, %r344;
mov.b32 %f93, %r348;
add.ftz.f32 %f94, %f92, %f93;
mov.b32 %r349, %f94;
mov.u32 %r350, 4;
shfl.sync.bfly.b32 %r351|%p9, %r349, %r350, %r342, %r344;
mov.b32 %f95, %r351;
add.ftz.f32 %f96, %f94, %f95;
mov.b32 %r352, %f96;
shfl.sync.bfly.b32 %r353|%p10, %r352, %r340, %r342, %r344;
mov.b32 %f97, %r353;
add.ftz.f32 %f98, %f96, %f97;
mov.b32 %r354, %f98;
mov.u32 %r355, 1;
shfl.sync.bfly.b32 %r356|%p11, %r354, %r355, %r342, %r344;
mov.b32 %f99, %r356;
add.ftz.f32 %f13, %f98, %f99;
setp.ge.u32 %p12, %r4, %r25;
setp.ne.s32 %p13, %r2, 0;
cvta.to.global.u64 %rd48, %rd15;
mul.wide.u32 %rd49, %r337, 4;
add.s64 %rd11, %rd48, %rd49;
or.pred %p14, %p13, %p12;
@%p14 bra $L__BB156_11;

st.global.f32 [%rd11], %f13;

$L__BB156_11:
ld.shared.f32 %f100, [%r19+128];
add.ftz.f32 %f101, %f100, %f151;
ld.shared.f32 %f102, [%r19+640];
add.ftz.f32 %f103, %f102, %f101;
ld.shared.f32 %f104, [%r19+1152];
add.ftz.f32 %f105, %f104, %f103;
mov.b32 %r357, %f105;
shfl.sync.bfly.b32 %r361|%p15, %r357, %r343, %r342, %r344;
mov.b32 %f106, %r361;
add.ftz.f32 %f107, %f105, %f106;
mov.b32 %r362, %f107;
shfl.sync.bfly.b32 %r364|%p16, %r362, %r347, %r342, %r344;
mov.b32 %f108, %r364;
add.ftz.f32 %f109, %f107, %f108;
mov.b32 %r365, %f109;
shfl.sync.bfly.b32 %r367|%p17, %r365, %r350, %r342, %r344;
mov.b32 %f110, %r367;
add.ftz.f32 %f111, %f109, %f110;
mov.b32 %r368, %f111;
shfl.sync.bfly.b32 %r370|%p18, %r368, %r340, %r342, %r344;
mov.b32 %f112, %r370;
add.ftz.f32 %f113, %f111, %f112;
mov.b32 %r371, %f113;
shfl.sync.bfly.b32 %r373|%p19, %r371, %r355, %r342, %r344;
mov.b32 %f114, %r373;
add.ftz.f32 %f14, %f113, %f114;
add.s32 %r374, %r4, 1;
setp.ge.u32 %p20, %r374, %r25;
setp.ne.s32 %p21, %r2, 1;
or.pred %p22, %p21, %p20;
@%p22 bra $L__BB156_13;

st.global.f32 [%rd11+4], %f14;

$L__BB156_13:
ld.shared.f32 %f115, [%r19+256];
add.ftz.f32 %f116, %f115, %f150;
ld.shared.f32 %f117, [%r19+768];
add.ftz.f32 %f118, %f117, %f116;
ld.shared.f32 %f119, [%r19+1280];
add.ftz.f32 %f120, %f119, %f118;
mov.b32 %r375, %f120;
mov.u32 %r376, 31;
mov.u32 %r377, 16;
mov.u32 %r378, -1;
shfl.sync.bfly.b32 %r379|%p23, %r375, %r377, %r376, %r378;
mov.b32 %f121, %r379;
add.ftz.f32 %f122, %f120, %f121;
mov.b32 %r380, %f122;
mov.u32 %r381, 8;
shfl.sync.bfly.b32 %r382|%p24, %r380, %r381, %r376, %r378;
mov.b32 %f123, %r382;
add.ftz.f32 %f124, %f122, %f123;
mov.b32 %r383, %f124;
mov.u32 %r384, 4;
shfl.sync.bfly.b32 %r385|%p25, %r383, %r384, %r376, %r378;
mov.b32 %f125, %r385;
add.ftz.f32 %f126, %f124, %f125;
mov.b32 %r386, %f126;
mov.u32 %r387, 2;
shfl.sync.bfly.b32 %r388|%p26, %r386, %r387, %r376, %r378;
mov.b32 %f127, %r388;
add.ftz.f32 %f128, %f126, %f127;
mov.b32 %r389, %f128;
mov.u32 %r390, 1;
shfl.sync.bfly.b32 %r391|%p27, %r389, %r390, %r376, %r378;
mov.b32 %f129, %r391;
add.ftz.f32 %f15, %f128, %f129;
add.s32 %r392, %r4, 2;
setp.ge.u32 %p28, %r392, %r25;
setp.ne.s32 %p29, %r2, 2;
or.pred %p30, %p29, %p28;
@%p30 bra $L__BB156_15;

st.global.f32 [%rd11+8], %f15;

$L__BB156_15:
ld.shared.f32 %f130, [%r19+384];
add.ftz.f32 %f131, %f130, %f149;
ld.shared.f32 %f132, [%r19+896];
add.ftz.f32 %f133, %f132, %f131;
ld.shared.f32 %f134, [%r19+1408];
add.ftz.f32 %f135, %f134, %f133;
mov.b32 %r393, %f135;
shfl.sync.bfly.b32 %r397|%p31, %r393, %r377, %r376, %r378;
mov.b32 %f136, %r397;
add.ftz.f32 %f137, %f135, %f136;
mov.b32 %r398, %f137;
shfl.sync.bfly.b32 %r400|%p32, %r398, %r381, %r376, %r378;
mov.b32 %f138, %r400;
add.ftz.f32 %f139, %f137, %f138;
mov.b32 %r401, %f139;
shfl.sync.bfly.b32 %r403|%p33, %r401, %r384, %r376, %r378;
mov.b32 %f140, %r403;
add.ftz.f32 %f141, %f139, %f140;
mov.b32 %r404, %f141;
shfl.sync.bfly.b32 %r406|%p34, %r404, %r387, %r376, %r378;
mov.b32 %f142, %r406;
add.ftz.f32 %f143, %f141, %f142;
mov.b32 %r407, %f143;
shfl.sync.bfly.b32 %r409|%p35, %r407, %r390, %r376, %r378;
mov.b32 %f144, %r409;
add.ftz.f32 %f16, %f143, %f144;
add.s32 %r410, %r4, 3;
setp.ge.u32 %p36, %r410, %r25;
setp.ne.s32 %p37, %r2, 3;
or.pred %p38, %p37, %p36;
@%p38 bra $L__BB156_17;

st.global.f32 [%rd11+12], %f16;

$L__BB156_17:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<35>;
.reg .b16 %rs<12>;
.reg .f32 %f<142>;
.reg .b32 %r<286>;
.reg .b64 %rd<73>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[384];

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate[384];

ld.param.u64 %rd8, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd9, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd10, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r46, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r47, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r48, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
ld.param.f32 %f31, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
ld.param.u64 %rd13, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+16];
ld.param.u64 %rd12, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+8];
ld.param.u64 %rd11, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3];
mov.u32 %r1, %tid.y;
shl.b32 %r49, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r49, %r2;
mov.u32 %r4, %ctaid.x;
mov.u32 %r5, %ctaid.y;
setp.eq.s64 %p2, %rd10, 0;
@%p2 bra $L__BB157_2;

cvta.to.global.u64 %rd17, %rd10;
mul.wide.u32 %rd18, %r5, 4;
add.s64 %rd19, %rd17, %rd18;
ld.global.nc.u32 %r282, [%rd19];
mul.hi.u32 %r50, %r5, %r33;
add.s32 %r51, %r50, %r5;
shr.u32 %r52, %r51, %r34;
mul.lo.s32 %r53, %r52, %r35;
sub.s32 %r283, %r5, %r53;
bra.uni $L__BB157_3;

$L__BB157_2:
mul.hi.u32 %r54, %r5, %r37;
add.s32 %r55, %r54, %r5;
shr.u32 %r282, %r55, %r38;
mov.u32 %r283, %r5;

$L__BB157_3:
mov.u32 %r11, %ctaid.z;
selp.b32 %r12, %r5, %r282, %p2;
or.b32 %r56, %r2, %r1;
setp.ne.s32 %p4, %r56, 0;
mov.f32 %f136, 0f00000000;
mov.f32 %f131, %f136;
mov.f32 %f132, %f136;
@%p4 bra $L__BB157_8;

setp.eq.s64 %p5, %rd11, 0;
mov.f32 %f131, 0f00000000;
mov.f32 %f132, %f131;
@%p5 bra $L__BB157_6;

cvta.to.global.u64 %rd20, %rd11;
mul.lo.s32 %r57, %r11, %r48;
cvt.u64.u32 %rd21, %r57;
mul.lo.s32 %r58, %r12, %r42;
cvt.u64.u32 %rd22, %r58;
cvt.u64.u32 %rd23, %r2;
cvt.s64.s32 %rd24, %r4;
add.s64 %rd25, %rd23, %rd24;
add.s64 %rd26, %rd25, %rd21;
add.s64 %rd27, %rd26, %rd22;
shl.b64 %rd28, %rd27, 2;
add.s64 %rd29, %rd20, %rd28;
ld.global.f32 %f132, [%rd29];

$L__BB157_6:
setp.eq.s64 %p6, %rd13, 0;
setp.eq.s64 %p7, %rd12, 0;
or.pred %p8, %p6, %p7;
@%p8 bra $L__BB157_8;

cvta.to.global.u64 %rd30, %rd13;
mul.lo.s32 %r60, %r11, %r48;
cvt.u64.u32 %rd31, %r60;
mul.lo.s32 %r61, %r12, %r42;
cvt.u64.u32 %rd32, %r61;
cvt.u64.u32 %rd33, %r2;
cvt.s64.s32 %rd34, %r4;
add.s64 %rd35, %rd33, %rd34;
add.s64 %rd36, %rd35, %rd31;
add.s64 %rd37, %rd36, %rd32;
shl.b64 %rd38, %rd37, 2;
add.s64 %rd39, %rd30, %rd38;
ld.global.f32 %f131, [%rd39];

$L__BB157_8:
shr.u32 %r63, %r32, 8;
shr.s32 %r64, %r3, 31;
shr.u32 %r65, %r64, 28;
add.s32 %r66, %r3, %r65;
shr.s32 %r285, %r66, 4;
setp.ge.s32 %p9, %r285, %r63;
mov.f32 %f135, %f136;
@%p9 bra $L__BB157_13;

and.b32 %r70, %r66, -16;
sub.s32 %r14, %r3, %r70;
shr.s32 %r71, %r14, 31;
shr.u32 %r72, %r71, 30;
add.s32 %r73, %r14, %r72;
shr.s32 %r15, %r73, 2;
and.b32 %r74, %r73, -4;
sub.s32 %r16, %r14, %r74;
shl.b32 %r284, %r285, 3;
mul.hi.u32 %r75, %r11, %r43;
add.s32 %r76, %r75, %r11;
shr.u32 %r77, %r76, %r44;
mad.lo.s32 %r78, %r77, %r46, %r285;
mad.lo.s32 %r79, %r282, %r40, %r78;
mad.lo.s32 %r80, %r4, %r36, %r79;
mul.wide.s32 %rd72, %r80, 144;
mul.lo.s32 %r81, %r283, %r41;
cvt.u64.u32 %rd40, %r81;
mul.lo.s32 %r82, %r11, %r47;
cvt.u64.u32 %rd41, %r82;
add.s64 %rd5, %rd40, %rd41;
mov.f32 %f135, 0f00000000;
cvta.to.global.u64 %rd42, %rd9;
cvta.to.global.u64 %rd46, %rd8;
mul.wide.s32 %rd57, %r16, 4;
mov.f32 %f136, %f135;

$L__BB157_10:
cvt.u64.u32 %rd43, %r284;
add.s64 %rd44, %rd5, %rd43;
shl.b32 %r119, %r16, 2;
shl.b32 %r120, %r15, 5;
add.s32 %r121, %r120, %r119;
cvt.s64.s32 %rd45, %r121;
add.s64 %rd47, %rd46, %rd45;
add.s64 %rd48, %rd47, %rd72;
ld.global.nc.u32 %r122, [%rd48+16];
ld.global.nc.u32 %r123, [%rd48+32];
shl.b32 %r124, %r15, 1;
cvt.u64.u32 %rd49, %r124;
and.b64 %rd50, %rd49, 2;
add.s64 %rd51, %rd46, %rd50;
add.s64 %rd52, %rd51, %rd72;
ld.global.nc.u16 %rs1, [%rd52+4];
cvt.u32.u16 %r125, %rs1;
ld.global.nc.u16 %rs2, [%rd52+8];
cvt.u32.u16 %r126, %rs2;
ld.global.nc.u16 %rs3, [%rd52+12];
cvt.u32.u16 %r127, %rs3;
and.b32 %r128, %r125, 16191;
and.b32 %r129, %r127, 3855;
and.b32 %r130, %r125, 49344;
shr.u32 %r131, %r130, 2;
or.b32 %r132, %r129, %r131;
setp.gt.s32 %p10, %r14, 7;
and.b32 %r133, %r126, 16191;
shr.u32 %r134, %r127, 4;
and.b32 %r135, %r134, 3855;
and.b32 %r136, %r126, 49344;
shr.u32 %r137, %r136, 2;
or.b32 %r138, %r135, %r137;
cvt.s64.s32 %rd53, %r124;
add.s64 %rd54, %rd44, %rd53;
mul.lo.s64 %rd55, %rd54, 36;
add.s64 %rd56, %rd42, %rd55;
ld.global.nc.u32 %r20, [%rd56];

{.reg .f16 low,high;
mov.b32 {low,high},%r20;
cvt.f32.f16 %f40, low;}


add.s64 %rd58, %rd56, %rd57;
ld.global.nc.u32 %r21, [%rd58+4];
ld.global.nc.u32 %r22, [%rd58+20];
ld.global.nc.u32 %r23, [%rd56+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r23;
cvt.f32.f16 %f41, low;}


ld.global.nc.u32 %r24, [%rd58+40];
ld.global.nc.u32 %r25, [%rd58+56];
add.s64 %rd59, %rd46, %rd72;
and.b32 %r86, %r122, 252645135;
and.b32 %r90, %r123, 252645135;
mov.u32 %r112, 0;

dp4a.s32.s32 %r85, %r86, %r21, %r112;


dp4a.s32.s32 %r89, %r90, %r22, %r85;

mov.u32 %r114, 16843009;

dp4a.s32.s32 %r93, %r114, %r21, %r112;


dp4a.s32.s32 %r97, %r114, %r22, %r93;

selp.b32 %r139, %r132, %r128, %p10;
cvt.u16.u32 %rs4, %r139;
and.b32 %r140, %r139, 63;
mul.lo.s32 %r141, %r89, %r140;
cvt.rn.f32.s32 %f44, %r141;
fma.rn.ftz.f32 %f45, %f40, %f44, 0f00000000;
selp.b32 %r142, %r138, %r133, %p10;
cvt.u16.u32 %rs5, %r142;
and.b32 %r143, %r142, 63;
mul.lo.s32 %r144, %r97, %r143;
cvt.rn.f32.s32 %f46, %r144;
fma.rn.ftz.f32 %f47, %f40, %f46, 0f00000000;
shr.u32 %r145, %r122, 4;
and.b32 %r102, %r145, 252645135;
shr.u32 %r146, %r123, 4;
and.b32 %r106, %r146, 252645135;

dp4a.s32.s32 %r101, %r102, %r24, %r112;


dp4a.s32.s32 %r105, %r106, %r25, %r101;


dp4a.s32.s32 %r109, %r114, %r24, %r112;


dp4a.s32.s32 %r113, %r114, %r25, %r109;

shr.u16 %rs6, %rs4, 8;
cvt.u32.u16 %r147, %rs6;
mul.lo.s32 %r148, %r105, %r147;
cvt.rn.f32.s32 %f48, %r148;
fma.rn.ftz.f32 %f49, %f41, %f48, %f45;
shr.u16 %rs7, %rs5, 8;
cvt.u32.u16 %r149, %rs7;
mul.lo.s32 %r150, %r113, %r149;
cvt.rn.f32.s32 %f50, %r150;
fma.rn.ftz.f32 %f51, %f41, %f50, %f47;
ld.global.nc.u32 %r117, [%rd59];

{.reg .f16 low,high;
mov.b32 {low,high},%r117;
cvt.f32.f16 %f42, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r117;
cvt.f32.f16 %f43, high;}


mul.ftz.f32 %f52, %f49, %f42;
mul.ftz.f32 %f53, %f51, %f43;
sub.ftz.f32 %f54, %f52, %f53;
add.ftz.f32 %f136, %f136, %f54;
setp.eq.s64 %p11, %rd12, 0;
@%p11 bra $L__BB157_12;

cvta.to.global.u64 %rd61, %rd12;
add.s64 %rd62, %rd61, %rd45;
add.s64 %rd63, %rd62, %rd72;
ld.global.u32 %r190, [%rd63+16];
ld.global.u32 %r191, [%rd63+32];
add.s64 %rd66, %rd61, %rd50;
add.s64 %rd67, %rd66, %rd72;
ld.global.u16 %r193, [%rd67+4];
ld.global.u16 %r194, [%rd67+8];
ld.global.u16 %r195, [%rd67+12];
and.b32 %r196, %r193, 16191;
and.b32 %r197, %r195, 3855;
and.b32 %r198, %r193, 49344;
shr.u32 %r199, %r198, 2;
or.b32 %r200, %r197, %r199;
and.b32 %r201, %r194, 16191;
shr.u32 %r202, %r195, 4;
and.b32 %r203, %r202, 3855;
and.b32 %r204, %r194, 49344;
shr.u32 %r205, %r204, 2;
or.b32 %r206, %r203, %r205;

{.reg .f16 low,high;
mov.b32 {low,high},%r20;
cvt.f32.f16 %f55, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r23;
cvt.f32.f16 %f56, low;}


add.s64 %rd68, %rd61, %rd72;
and.b32 %r154, %r190, 252645135;
and.b32 %r158, %r191, 252645135;

dp4a.s32.s32 %r153, %r154, %r21, %r112;


dp4a.s32.s32 %r157, %r158, %r22, %r153;


dp4a.s32.s32 %r161, %r114, %r21, %r112;


dp4a.s32.s32 %r165, %r114, %r22, %r161;

selp.b32 %r207, %r200, %r196, %p10;
cvt.u16.u32 %rs8, %r207;
and.b32 %r208, %r207, 63;
mul.lo.s32 %r209, %r157, %r208;
cvt.rn.f32.s32 %f59, %r209;
fma.rn.ftz.f32 %f60, %f55, %f59, 0f00000000;
selp.b32 %r210, %r206, %r201, %p10;
cvt.u16.u32 %rs9, %r210;
and.b32 %r211, %r210, 63;
mul.lo.s32 %r212, %r165, %r211;
cvt.rn.f32.s32 %f61, %r212;
fma.rn.ftz.f32 %f62, %f55, %f61, 0f00000000;
shr.u32 %r213, %r190, 4;
and.b32 %r170, %r213, 252645135;
shr.u32 %r214, %r191, 4;
and.b32 %r174, %r214, 252645135;

dp4a.s32.s32 %r169, %r170, %r24, %r112;


dp4a.s32.s32 %r173, %r174, %r25, %r169;


dp4a.s32.s32 %r177, %r114, %r24, %r112;


dp4a.s32.s32 %r181, %r114, %r25, %r177;

shr.u16 %rs10, %rs8, 8;
cvt.u32.u16 %r215, %rs10;
mul.lo.s32 %r216, %r173, %r215;
cvt.rn.f32.s32 %f63, %r216;
fma.rn.ftz.f32 %f64, %f56, %f63, %f60;
shr.u16 %rs11, %rs9, 8;
cvt.u32.u16 %r217, %rs11;
mul.lo.s32 %r218, %r181, %r217;
cvt.rn.f32.s32 %f65, %r218;
fma.rn.ftz.f32 %f66, %f56, %f65, %f62;
ld.global.u32 %r186, [%rd68];

{.reg .f16 low,high;
mov.b32 {low,high},%r186;
cvt.f32.f16 %f57, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r186;
cvt.f32.f16 %f58, high;}


mul.ftz.f32 %f67, %f64, %f57;
mul.ftz.f32 %f68, %f66, %f58;
sub.ftz.f32 %f69, %f67, %f68;
add.ftz.f32 %f135, %f135, %f69;

$L__BB157_12:
add.s32 %r284, %r284, 64;
add.s64 %rd72, %rd72, 1152;
add.s32 %r285, %r285, 8;
setp.lt.s32 %p13, %r285, %r63;
@%p13 bra $L__BB157_10;

$L__BB157_13:
setp.eq.s32 %p14, %r1, 0;
@%p14 bra $L__BB157_16;

setp.eq.s64 %p15, %rd12, 0;
add.s32 %r28, %r1, -1;
shl.b32 %r223, %r1, 7;
mov.u32 %r224, _ZZ13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r225, %r224, %r223;
shl.b32 %r226, %r2, 2;
add.s32 %r227, %r225, %r226;
st.shared.f32 [%r227+-128], %f136;
@%p15 bra $L__BB157_16;

shl.b32 %r229, %r28, 7;
mov.u32 %r230, _ZZ13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r231, %r230, %r229;
add.s32 %r233, %r231, %r226;
st.shared.f32 [%r233], %f135;

$L__BB157_16:
setp.ne.s32 %p16, %r1, 0;
bar.sync 0;
@%p16 bra $L__BB157_35;

setp.ne.s64 %p17, %rd12, 0;
shl.b32 %r236, %r2, 2;
mov.u32 %r237, _ZZ13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r29, %r237, %r236;
ld.shared.f32 %f70, [%r29];
add.ftz.f32 %f13, %f136, %f70;
@%p17 bra $L__BB157_19;
bra.uni $L__BB157_18;

$L__BB157_19:
mov.u32 %r240, _ZZ13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r241, %r240, %r236;
ld.shared.f32 %f74, [%r241];
add.ftz.f32 %f75, %f135, %f74;
ld.shared.f32 %f76, [%r29+128];
add.ftz.f32 %f77, %f13, %f76;
ld.shared.f32 %f78, [%r241+128];
add.ftz.f32 %f79, %f75, %f78;
ld.shared.f32 %f80, [%r29+256];
add.ftz.f32 %f139, %f77, %f80;
ld.shared.f32 %f81, [%r241+256];
add.ftz.f32 %f135, %f79, %f81;
bra.uni $L__BB157_20;

$L__BB157_18:
ld.shared.f32 %f71, [%r29+128];
add.ftz.f32 %f72, %f13, %f71;
ld.shared.f32 %f73, [%r29+256];
add.ftz.f32 %f139, %f72, %f73;

$L__BB157_20:
mov.b32 %r242, %f139;
mov.u32 %r243, 31;
mov.u32 %r244, 16;
mov.u32 %r245, -1;
shfl.sync.bfly.b32 %r246|%p18, %r242, %r244, %r243, %r245;
mov.b32 %f82, %r246;
add.ftz.f32 %f83, %f139, %f82;
mov.b32 %r247, %f83;
mov.u32 %r248, 8;
shfl.sync.bfly.b32 %r249|%p19, %r247, %r248, %r243, %r245;
mov.b32 %f84, %r249;
add.ftz.f32 %f85, %f83, %f84;
mov.b32 %r250, %f85;
mov.u32 %r251, 4;
shfl.sync.bfly.b32 %r252|%p20, %r250, %r251, %r243, %r245;
mov.b32 %f86, %r252;
add.ftz.f32 %f87, %f85, %f86;
mov.b32 %r253, %f87;
mov.u32 %r254, 2;
shfl.sync.bfly.b32 %r255|%p21, %r253, %r254, %r243, %r245;
mov.b32 %f88, %r255;
add.ftz.f32 %f19, %f87, %f88;
mov.b32 %r256, %f19;
mov.u32 %r257, 1;
shfl.sync.bfly.b32 %r30|%p1, %r256, %r257, %r243, %r245;
setp.eq.s64 %p22, %rd12, 0;
@%p22 bra $L__BB157_22;

mov.b32 %r258, %f135;
shfl.sync.bfly.b32 %r262|%p23, %r258, %r244, %r243, %r245;
mov.b32 %f89, %r262;
add.ftz.f32 %f90, %f135, %f89;
mov.b32 %r263, %f90;
shfl.sync.bfly.b32 %r265|%p24, %r263, %r248, %r243, %r245;
mov.b32 %f91, %r265;
add.ftz.f32 %f92, %f90, %f91;
mov.b32 %r266, %f92;
shfl.sync.bfly.b32 %r268|%p25, %r266, %r251, %r243, %r245;
mov.b32 %f93, %r268;
add.ftz.f32 %f94, %f92, %f93;
mov.b32 %r269, %f94;
shfl.sync.bfly.b32 %r271|%p26, %r269, %r254, %r243, %r245;
mov.b32 %f95, %r271;
add.ftz.f32 %f96, %f94, %f95;
mov.b32 %r272, %f96;
shfl.sync.bfly.b32 %r274|%p27, %r272, %r257, %r243, %r245;
mov.b32 %f97, %r274;
add.ftz.f32 %f135, %f96, %f97;

$L__BB157_22:
setp.ne.s32 %p28, %r2, 0;
mov.b32 %f98, %r30;
add.ftz.f32 %f22, %f19, %f98;
@%p28 bra $L__BB157_35;

add.ftz.f32 %f141, %f132, %f22;
@%p22 bra $L__BB157_34;

add.ftz.f32 %f24, %f131, %f135;
setp.gt.s32 %p30, %r31, 2;
@%p30 bra $L__BB157_28;

setp.eq.s32 %p33, %r31, 1;
@%p33 bra $L__BB157_32;

setp.eq.s32 %p34, %r31, 2;
@%p34 bra $L__BB157_27;
bra.uni $L__BB157_33;

$L__BB157_27:
mul.ftz.f32 %f118, %f24, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f119, %f118;
add.ftz.f32 %f120, %f119, 0f3F800000;
div.approx.ftz.f32 %f121, %f24, %f120;
mul.ftz.f32 %f141, %f141, %f121;
bra.uni $L__BB157_34;

$L__BB157_28:
setp.eq.s32 %p31, %r31, 3;
@%p31 bra $L__BB157_31;

setp.ne.s32 %p32, %r31, 6;
@%p32 bra $L__BB157_33;

min.ftz.f32 %f99, %f24, %f31;
neg.ftz.f32 %f100, %f31;
min.ftz.f32 %f101, %f141, %f31;
max.ftz.f32 %f102, %f101, %f100;
mul.ftz.f32 %f103, %f99, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f104, %f103;
add.ftz.f32 %f105, %f104, 0f3F800000;
div.approx.ftz.f32 %f106, %f99, %f105;
mul.ftz.f32 %f141, %f102, %f106;
bra.uni $L__BB157_34;

$L__BB157_32:
mul.ftz.f32 %f122, %f24, 0f3F000000;
mul.ftz.f32 %f123, %f24, 0f3D372713;
fma.rn.ftz.f32 %f124, %f24, %f123, 0f3F800000;
mul.ftz.f32 %f125, %f24, 0f3F4C422A;
mul.ftz.f32 %f126, %f125, %f124;
tanh.approx.f32 %f127, %f126;
add.ftz.f32 %f128, %f127, 0f3F800000;
mul.ftz.f32 %f129, %f122, %f128;
mul.ftz.f32 %f141, %f141, %f129;
bra.uni $L__BB157_34;

$L__BB157_31:
mov.f32 %f107, 0f40E00000;
min.ftz.f32 %f108, %f24, %f107;
min.ftz.f32 %f109, %f141, %f107;
mov.f32 %f110, 0fC0E00000;
max.ftz.f32 %f111, %f109, %f110;
mul.ftz.f32 %f112, %f108, 0f3FD9DB23;
mul.ftz.f32 %f113, %f112, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f114, %f113;
add.ftz.f32 %f115, %f114, 0f3F800000;
div.approx.ftz.f32 %f116, %f108, %f115;
add.ftz.f32 %f117, %f111, 0f3F800000;
mul.ftz.f32 %f141, %f117, %f116;
bra.uni $L__BB157_34;

$L__BB157_33:
mul.ftz.f32 %f141, %f141, %f24;

$L__BB157_34:
cvta.to.global.u64 %rd69, %rd16;
mov.u32 %r278, %ctaid.y;
mad.lo.s32 %r280, %r278, %r42, %r4;
mad.lo.s32 %r281, %r11, %r48, %r280;
mul.wide.u32 %rd70, %r281, 4;
add.s64 %rd71, %rd69, %rd70;
st.global.f32 [%rd71], %f141;

$L__BB157_35:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<16>;
.reg .b16 %rs<22>;
.reg .f32 %f<76>;
.reg .b32 %r<312>;
.reg .b64 %rd<55>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[384];

ld.param.u64 %rd14, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd12, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd13, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
cvta.to.global.u64 %rd1, %rd15;
cvta.to.global.u64 %rd54, %rd14;
mov.u32 %r1, %tid.y;
shl.b32 %r44, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r44, %r2;
shr.u32 %r4, %r27, 8;
mov.u32 %r308, %ctaid.y;
setp.eq.s64 %p1, %rd12, 0;
@%p1 bra $L__BB158_2;

cvta.to.global.u64 %rd16, %rd12;
mul.wide.u32 %rd17, %r308, 4;
add.s64 %rd18, %rd16, %rd17;
ld.global.nc.u32 %r307, [%rd18];
mul.hi.u32 %r45, %r308, %r28;
add.s32 %r46, %r45, %r308;
shr.u32 %r47, %r46, %r29;
mul.lo.s32 %r48, %r47, %r30;
sub.s32 %r308, %r308, %r48;
bra.uni $L__BB158_3;

$L__BB158_2:
mul.hi.u32 %r49, %r308, %r32;
add.s32 %r50, %r49, %r308;
shr.u32 %r307, %r50, %r33;

$L__BB158_3:
mov.u32 %r11, %ctaid.z;
mul.hi.u32 %r51, %r11, %r38;
add.s32 %r52, %r51, %r11;
shr.u32 %r53, %r52, %r39;
mul.lo.s32 %r12, %r53, %r41;
mul.lo.s32 %r13, %r307, %r35;
mov.u32 %r54, %ctaid.x;
mul.lo.s32 %r14, %r54, %r31;
shr.s32 %r55, %r3, 31;
shr.u32 %r56, %r55, 28;
add.s32 %r57, %r3, %r56;
shr.s32 %r311, %r57, 4;
setp.ge.s32 %p2, %r311, %r4;
mov.f32 %f75, 0f00000000;
@%p2 bra $L__BB158_9;

mul.lo.s32 %r58, %r308, %r36;
cvt.u64.u32 %rd19, %r58;
and.b32 %r62, %r57, -16;
sub.s32 %r16, %r3, %r62;
shr.s32 %r63, %r16, 31;
shr.u32 %r64, %r63, 30;
add.s32 %r65, %r16, %r64;
shr.s32 %r66, %r65, 2;
shl.b32 %r67, %r66, 1;
shl.b32 %r68, %r66, 5;
and.b32 %r69, %r65, -4;
sub.s32 %r70, %r16, %r69;
shl.b32 %r71, %r70, 2;
add.s32 %r72, %r68, %r71;
cvt.s64.s32 %rd3, %r72;
and.b32 %r73, %r66, 1;
cvt.u64.u32 %rd4, %r73;
cvt.s64.s32 %rd20, %r67;
mul.lo.s32 %r74, %r11, %r42;
cvt.u64.u32 %rd21, %r74;
add.s64 %rd22, %rd21, %rd20;
add.s64 %rd5, %rd22, %rd19;
cvt.s64.s32 %rd6, %r70;
not.b32 %r75, %r311;
add.s32 %r17, %r4, %r75;
and.b32 %r76, %r17, 8;
setp.ne.s32 %p3, %r76, 0;
mov.f32 %f75, 0f00000000;
@%p3 bra $L__BB158_6;

setp.gt.s32 %p4, %r16, 7;
shl.b32 %r113, %r311, 3;
cvt.u64.u32 %rd23, %r113;
add.s32 %r114, %r13, %r14;
add.s32 %r115, %r114, %r12;
add.s32 %r116, %r115, %r311;
mul.wide.s32 %rd24, %r116, 144;
add.s64 %rd25, %rd54, %rd24;
add.s64 %rd26, %rd25, %rd3;
ld.global.nc.u32 %r117, [%rd26+16];
ld.global.nc.u32 %r118, [%rd26+32];
shl.b64 %rd27, %rd4, 1;
add.s64 %rd28, %rd25, %rd27;
ld.global.nc.u16 %rs1, [%rd28+4];
cvt.u32.u16 %r119, %rs1;
ld.global.nc.u16 %rs2, [%rd28+8];
cvt.u32.u16 %r120, %rs2;
ld.global.nc.u16 %rs3, [%rd28+12];
cvt.u32.u16 %r121, %rs3;
and.b32 %r122, %r119, 16191;
and.b32 %r123, %r121, 3855;
and.b32 %r124, %r119, 49344;
shr.u32 %r125, %r124, 2;
or.b32 %r126, %r123, %r125;
and.b32 %r127, %r120, 16191;
shr.u32 %r128, %r121, 4;
and.b32 %r129, %r128, 3855;
and.b32 %r130, %r120, 49344;
shr.u32 %r131, %r130, 2;
or.b32 %r132, %r129, %r131;
add.s64 %rd29, %rd5, %rd23;
mul.lo.s64 %rd30, %rd29, 36;
add.s64 %rd31, %rd1, %rd30;
ld.global.nc.u32 %r77, [%rd31];

{.reg .f16 low,high;
mov.b32 {low,high},%r77;
cvt.f32.f16 %f11, low;}


shl.b64 %rd32, %rd6, 2;
add.s64 %rd33, %rd31, %rd32;
ld.global.nc.u32 %r81, [%rd33+4];
ld.global.nc.u32 %r85, [%rd33+20];
ld.global.nc.u32 %r78, [%rd31+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r78;
cvt.f32.f16 %f12, low;}


ld.global.nc.u32 %r97, [%rd33+40];
ld.global.nc.u32 %r101, [%rd33+56];
and.b32 %r80, %r117, 252645135;
and.b32 %r84, %r118, 252645135;
mov.u32 %r106, 0;

dp4a.s32.s32 %r79, %r80, %r81, %r106;


dp4a.s32.s32 %r83, %r84, %r85, %r79;

mov.u32 %r108, 16843009;

dp4a.s32.s32 %r87, %r108, %r81, %r106;


dp4a.s32.s32 %r91, %r108, %r85, %r87;

selp.b32 %r133, %r126, %r122, %p4;
cvt.u16.u32 %rs4, %r133;
and.b32 %r134, %r133, 63;
mul.lo.s32 %r135, %r83, %r134;
cvt.rn.f32.s32 %f15, %r135;
fma.rn.ftz.f32 %f16, %f11, %f15, 0f00000000;
selp.b32 %r136, %r132, %r127, %p4;
cvt.u16.u32 %rs5, %r136;
and.b32 %r137, %r136, 63;
mul.lo.s32 %r138, %r91, %r137;
cvt.rn.f32.s32 %f17, %r138;
fma.rn.ftz.f32 %f18, %f11, %f17, 0f00000000;
shr.u32 %r139, %r117, 4;
and.b32 %r96, %r139, 252645135;
shr.u32 %r140, %r118, 4;
and.b32 %r100, %r140, 252645135;

dp4a.s32.s32 %r95, %r96, %r97, %r106;


dp4a.s32.s32 %r99, %r100, %r101, %r95;


dp4a.s32.s32 %r103, %r108, %r97, %r106;


dp4a.s32.s32 %r107, %r108, %r101, %r103;

shr.u16 %rs6, %rs4, 8;
cvt.u32.u16 %r141, %rs6;
mul.lo.s32 %r142, %r99, %r141;
cvt.rn.f32.s32 %f19, %r142;
fma.rn.ftz.f32 %f20, %f12, %f19, %f16;
shr.u16 %rs7, %rs5, 8;
cvt.u32.u16 %r143, %rs7;
mul.lo.s32 %r144, %r107, %r143;
cvt.rn.f32.s32 %f21, %r144;
fma.rn.ftz.f32 %f22, %f12, %f21, %f18;
ld.global.nc.u32 %r111, [%rd25];

{.reg .f16 low,high;
mov.b32 {low,high},%r111;
cvt.f32.f16 %f13, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r111;
cvt.f32.f16 %f14, high;}


mul.ftz.f32 %f23, %f20, %f13;
mul.ftz.f32 %f24, %f22, %f14;
sub.ftz.f32 %f25, %f23, %f24;
add.ftz.f32 %f75, %f25, 0f00000000;
add.s32 %r311, %r311, 8;

$L__BB158_6:
and.b32 %r145, %r17, -8;
setp.eq.s32 %p5, %r145, 0;
@%p5 bra $L__BB158_9;

add.s32 %r146, %r311, %r12;
add.s32 %r147, %r146, %r13;
add.s32 %r148, %r147, %r14;
mul.wide.s32 %rd7, %r148, 144;
shl.b32 %r149, %r311, 3;
add.s32 %r310, %r149, 64;
shl.b64 %rd34, %rd4, 1;
or.b64 %rd35, %rd34, %rd7;
or.b64 %rd8, %rd35, 8;
add.s64 %rd36, %rd7, %rd3;
add.s64 %rd9, %rd36, 16;

$L__BB158_8:
add.s32 %r222, %r310, -64;
cvt.u64.u32 %rd37, %r222;
add.s64 %rd38, %rd54, %rd9;
ld.global.nc.u32 %r223, [%rd38];
ld.global.nc.u32 %r224, [%rd38+16];
add.s64 %rd39, %rd54, %rd8;
ld.global.nc.u16 %rs8, [%rd39+-4];
cvt.u32.u16 %r225, %rs8;
ld.global.nc.u16 %rs9, [%rd39];
cvt.u32.u16 %r226, %rs9;
ld.global.nc.u16 %rs10, [%rd39+4];
cvt.u32.u16 %r227, %rs10;
and.b32 %r228, %r225, 16191;
and.b32 %r229, %r227, 3855;
and.b32 %r230, %r225, 49344;
shr.u32 %r231, %r230, 2;
or.b32 %r232, %r229, %r231;
setp.gt.s32 %p6, %r16, 7;
and.b32 %r233, %r226, 16191;
shr.u32 %r234, %r227, 4;
and.b32 %r235, %r234, 3855;
and.b32 %r236, %r226, 49344;
shr.u32 %r237, %r236, 2;
or.b32 %r238, %r235, %r237;
add.s64 %rd40, %rd5, %rd37;
mul.lo.s64 %rd41, %rd40, 36;
add.s64 %rd42, %rd1, %rd41;
ld.global.nc.u32 %r150, [%rd42];

{.reg .f16 low,high;
mov.b32 {low,high},%r150;
cvt.f32.f16 %f26, low;}


shl.b64 %rd43, %rd6, 2;
add.s64 %rd44, %rd42, %rd43;
ld.global.nc.u32 %r154, [%rd44+4];
ld.global.nc.u32 %r158, [%rd44+20];
ld.global.nc.u32 %r151, [%rd42+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r151;
cvt.f32.f16 %f27, low;}


ld.global.nc.u32 %r170, [%rd44+40];
ld.global.nc.u32 %r174, [%rd44+56];
add.s64 %rd45, %rd54, %rd7;
and.b32 %r153, %r223, 252645135;
and.b32 %r157, %r224, 252645135;
mov.u32 %r215, 0;

dp4a.s32.s32 %r152, %r153, %r154, %r215;


dp4a.s32.s32 %r156, %r157, %r158, %r152;

mov.u32 %r217, 16843009;

dp4a.s32.s32 %r160, %r217, %r154, %r215;


dp4a.s32.s32 %r164, %r217, %r158, %r160;

selp.b32 %r239, %r232, %r228, %p6;
cvt.u16.u32 %rs11, %r239;
and.b32 %r240, %r239, 63;
mul.lo.s32 %r241, %r156, %r240;
cvt.rn.f32.s32 %f34, %r241;
fma.rn.ftz.f32 %f35, %f26, %f34, 0f00000000;
selp.b32 %r242, %r238, %r233, %p6;
cvt.u16.u32 %rs12, %r242;
and.b32 %r243, %r242, 63;
mul.lo.s32 %r244, %r164, %r243;
cvt.rn.f32.s32 %f36, %r244;
fma.rn.ftz.f32 %f37, %f26, %f36, 0f00000000;
shr.u32 %r245, %r223, 4;
and.b32 %r169, %r245, 252645135;
shr.u32 %r246, %r224, 4;
and.b32 %r173, %r246, 252645135;

dp4a.s32.s32 %r168, %r169, %r170, %r215;


dp4a.s32.s32 %r172, %r173, %r174, %r168;


dp4a.s32.s32 %r176, %r217, %r170, %r215;


dp4a.s32.s32 %r180, %r217, %r174, %r176;

shr.u16 %rs13, %rs11, 8;
cvt.u32.u16 %r247, %rs13;
mul.lo.s32 %r248, %r172, %r247;
cvt.rn.f32.s32 %f38, %r248;
fma.rn.ftz.f32 %f39, %f27, %f38, %f35;
shr.u16 %rs14, %rs12, 8;
cvt.u32.u16 %r249, %rs14;
mul.lo.s32 %r250, %r180, %r249;
cvt.rn.f32.s32 %f40, %r250;
fma.rn.ftz.f32 %f41, %f27, %f40, %f37;
ld.global.nc.u32 %r184, [%rd45];

{.reg .f16 low,high;
mov.b32 {low,high},%r184;
cvt.f32.f16 %f28, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r184;
cvt.f32.f16 %f29, high;}


mul.ftz.f32 %f42, %f39, %f28;
mul.ftz.f32 %f43, %f41, %f29;
sub.ftz.f32 %f44, %f42, %f43;
add.ftz.f32 %f45, %f75, %f44;
ld.global.nc.u32 %r251, [%rd38+1152];
ld.global.nc.u32 %r252, [%rd38+1168];
ld.global.nc.u16 %rs15, [%rd39+1148];
cvt.u32.u16 %r253, %rs15;
ld.global.nc.u16 %rs16, [%rd39+1152];
cvt.u32.u16 %r254, %rs16;
ld.global.nc.u16 %rs17, [%rd39+1156];
cvt.u32.u16 %r255, %rs17;
and.b32 %r256, %r253, 16191;
and.b32 %r257, %r255, 3855;
and.b32 %r258, %r253, 49344;
shr.u32 %r259, %r258, 2;
or.b32 %r260, %r257, %r259;
and.b32 %r261, %r254, 16191;
shr.u32 %r262, %r255, 4;
and.b32 %r263, %r262, 3855;
and.b32 %r264, %r254, 49344;
shr.u32 %r265, %r264, 2;
or.b32 %r266, %r263, %r265;
cvt.u64.u32 %rd46, %r310;
add.s64 %rd47, %rd5, %rd46;
mul.lo.s64 %rd48, %rd47, 36;
add.s64 %rd49, %rd1, %rd48;
ld.global.nc.u32 %r186, [%rd49];

{.reg .f16 low,high;
mov.b32 {low,high},%r186;
cvt.f32.f16 %f30, low;}


add.s64 %rd50, %rd49, %rd43;
ld.global.nc.u32 %r190, [%rd50+4];
ld.global.nc.u32 %r194, [%rd50+20];
ld.global.nc.u32 %r187, [%rd49+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r187;
cvt.f32.f16 %f31, low;}


ld.global.nc.u32 %r206, [%rd50+40];
ld.global.nc.u32 %r210, [%rd50+56];
and.b32 %r189, %r251, 252645135;
and.b32 %r193, %r252, 252645135;

dp4a.s32.s32 %r188, %r189, %r190, %r215;


dp4a.s32.s32 %r192, %r193, %r194, %r188;


dp4a.s32.s32 %r196, %r217, %r190, %r215;


dp4a.s32.s32 %r200, %r217, %r194, %r196;

selp.b32 %r267, %r260, %r256, %p6;
cvt.u16.u32 %rs18, %r267;
and.b32 %r268, %r267, 63;
mul.lo.s32 %r269, %r192, %r268;
cvt.rn.f32.s32 %f46, %r269;
fma.rn.ftz.f32 %f47, %f30, %f46, 0f00000000;
selp.b32 %r270, %r266, %r261, %p6;
cvt.u16.u32 %rs19, %r270;
and.b32 %r271, %r270, 63;
mul.lo.s32 %r272, %r200, %r271;
cvt.rn.f32.s32 %f48, %r272;
fma.rn.ftz.f32 %f49, %f30, %f48, 0f00000000;
shr.u32 %r273, %r251, 4;
and.b32 %r205, %r273, 252645135;
shr.u32 %r274, %r252, 4;
and.b32 %r209, %r274, 252645135;

dp4a.s32.s32 %r204, %r205, %r206, %r215;


dp4a.s32.s32 %r208, %r209, %r210, %r204;


dp4a.s32.s32 %r212, %r217, %r206, %r215;


dp4a.s32.s32 %r216, %r217, %r210, %r212;

shr.u16 %rs20, %rs18, 8;
cvt.u32.u16 %r275, %rs20;
mul.lo.s32 %r276, %r208, %r275;
cvt.rn.f32.s32 %f50, %r276;
fma.rn.ftz.f32 %f51, %f31, %f50, %f47;
shr.u16 %rs21, %rs19, 8;
cvt.u32.u16 %r277, %rs21;
mul.lo.s32 %r278, %r216, %r277;
cvt.rn.f32.s32 %f52, %r278;
fma.rn.ftz.f32 %f53, %f31, %f52, %f49;
ld.global.nc.u32 %r220, [%rd45+1152];

{.reg .f16 low,high;
mov.b32 {low,high},%r220;
cvt.f32.f16 %f32, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r220;
cvt.f32.f16 %f33, high;}


mul.ftz.f32 %f54, %f51, %f32;
mul.ftz.f32 %f55, %f53, %f33;
sub.ftz.f32 %f56, %f54, %f55;
add.ftz.f32 %f75, %f45, %f56;
add.s64 %rd54, %rd54, 2304;
add.s32 %r310, %r310, 128;
add.s32 %r311, %r311, 16;
setp.lt.s32 %p7, %r311, %r4;
@%p7 bra $L__BB158_8;

$L__BB158_9:
setp.eq.s32 %p8, %r1, 0;
@%p8 bra $L__BB158_11;

shl.b32 %r279, %r1, 7;
mov.u32 %r280, _ZZ13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r281, %r280, %r279;
shl.b32 %r282, %r2, 2;
add.s32 %r283, %r281, %r282;
st.shared.f32 [%r283+-128], %f75;

$L__BB158_11:
setp.ne.s32 %p9, %r1, 0;
bar.sync 0;
@%p9 bra $L__BB158_14;

mov.u32 %r284, %ctaid.y;
mad.lo.s32 %r286, %r284, %r37, %r54;
mad.lo.s32 %r26, %r11, %r43, %r286;
shl.b32 %r287, %r2, 2;
mov.u32 %r288, _ZZ13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r289, %r288, %r287;
mov.u32 %r290, 2;
ld.shared.f32 %f57, [%r289];
add.ftz.f32 %f58, %f75, %f57;
ld.shared.f32 %f59, [%r289+128];
add.ftz.f32 %f60, %f58, %f59;
ld.shared.f32 %f61, [%r289+256];
add.ftz.f32 %f62, %f60, %f61;
mov.b32 %r291, %f62;
mov.u32 %r292, 31;
mov.u32 %r293, 16;
mov.u32 %r294, -1;
shfl.sync.bfly.b32 %r295|%p10, %r291, %r293, %r292, %r294;
mov.b32 %f63, %r295;
add.ftz.f32 %f64, %f62, %f63;
mov.b32 %r296, %f64;
mov.u32 %r297, 8;
shfl.sync.bfly.b32 %r298|%p11, %r296, %r297, %r292, %r294;
mov.b32 %f65, %r298;
add.ftz.f32 %f66, %f64, %f65;
mov.b32 %r299, %f66;
mov.u32 %r300, 4;
shfl.sync.bfly.b32 %r301|%p12, %r299, %r300, %r292, %r294;
mov.b32 %f67, %r301;
add.ftz.f32 %f68, %f66, %f67;
mov.b32 %r302, %f68;
shfl.sync.bfly.b32 %r303|%p13, %r302, %r290, %r292, %r294;
mov.b32 %f69, %r303;
add.ftz.f32 %f70, %f68, %f69;
mov.b32 %r304, %f70;
mov.u32 %r305, 1;
shfl.sync.bfly.b32 %r306|%p14, %r304, %r305, %r292, %r294;
mov.b32 %f71, %r306;
add.ftz.f32 %f7, %f70, %f71;
setp.ne.s32 %p15, %r2, 0;
@%p15 bra $L__BB158_14;

cvta.to.global.u64 %rd51, %rd13;
mul.wide.u32 %rd52, %r26, 4;
add.s64 %rd53, %rd51, %rd52;
st.global.f32 [%rd53], %f7;

$L__BB158_14:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<35>;
.reg .b16 %rs<12>;
.reg .f32 %f<142>;
.reg .b32 %r<286>;
.reg .b64 %rd<73>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[384];

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate[384];

ld.param.u64 %rd8, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd9, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd10, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r46, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r47, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r48, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
ld.param.f32 %f31, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
ld.param.u64 %rd13, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+16];
ld.param.u64 %rd12, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+8];
ld.param.u64 %rd11, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3];
mov.u32 %r1, %tid.y;
shl.b32 %r49, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r49, %r2;
mov.u32 %r4, %ctaid.x;
mov.u32 %r5, %ctaid.y;
setp.eq.s64 %p2, %rd10, 0;
@%p2 bra $L__BB159_2;

cvta.to.global.u64 %rd17, %rd10;
mul.wide.u32 %rd18, %r5, 4;
add.s64 %rd19, %rd17, %rd18;
ld.global.nc.u32 %r282, [%rd19];
mul.hi.u32 %r50, %r5, %r33;
add.s32 %r51, %r50, %r5;
shr.u32 %r52, %r51, %r34;
mul.lo.s32 %r53, %r52, %r35;
sub.s32 %r283, %r5, %r53;
bra.uni $L__BB159_3;

$L__BB159_2:
mul.hi.u32 %r54, %r5, %r37;
add.s32 %r55, %r54, %r5;
shr.u32 %r282, %r55, %r38;
mov.u32 %r283, %r5;

$L__BB159_3:
mov.u32 %r11, %ctaid.z;
selp.b32 %r12, %r5, %r282, %p2;
or.b32 %r56, %r2, %r1;
setp.ne.s32 %p4, %r56, 0;
mov.f32 %f136, 0f00000000;
mov.f32 %f131, %f136;
mov.f32 %f132, %f136;
@%p4 bra $L__BB159_8;

setp.eq.s64 %p5, %rd11, 0;
mov.f32 %f131, 0f00000000;
mov.f32 %f132, %f131;
@%p5 bra $L__BB159_6;

cvta.to.global.u64 %rd20, %rd11;
mul.lo.s32 %r57, %r11, %r48;
cvt.u64.u32 %rd21, %r57;
mul.lo.s32 %r58, %r12, %r42;
cvt.u64.u32 %rd22, %r58;
cvt.u64.u32 %rd23, %r2;
cvt.s64.s32 %rd24, %r4;
add.s64 %rd25, %rd23, %rd24;
add.s64 %rd26, %rd25, %rd21;
add.s64 %rd27, %rd26, %rd22;
shl.b64 %rd28, %rd27, 2;
add.s64 %rd29, %rd20, %rd28;
ld.global.f32 %f132, [%rd29];

$L__BB159_6:
setp.eq.s64 %p6, %rd13, 0;
setp.eq.s64 %p7, %rd12, 0;
or.pred %p8, %p6, %p7;
@%p8 bra $L__BB159_8;

cvta.to.global.u64 %rd30, %rd13;
mul.lo.s32 %r60, %r11, %r48;
cvt.u64.u32 %rd31, %r60;
mul.lo.s32 %r61, %r12, %r42;
cvt.u64.u32 %rd32, %r61;
cvt.u64.u32 %rd33, %r2;
cvt.s64.s32 %rd34, %r4;
add.s64 %rd35, %rd33, %rd34;
add.s64 %rd36, %rd35, %rd31;
add.s64 %rd37, %rd36, %rd32;
shl.b64 %rd38, %rd37, 2;
add.s64 %rd39, %rd30, %rd38;
ld.global.f32 %f131, [%rd39];

$L__BB159_8:
shr.u32 %r63, %r32, 8;
shr.s32 %r64, %r3, 31;
shr.u32 %r65, %r64, 28;
add.s32 %r66, %r3, %r65;
shr.s32 %r285, %r66, 4;
setp.ge.s32 %p9, %r285, %r63;
mov.f32 %f135, %f136;
@%p9 bra $L__BB159_13;

and.b32 %r70, %r66, -16;
sub.s32 %r14, %r3, %r70;
shr.s32 %r71, %r14, 31;
shr.u32 %r72, %r71, 30;
add.s32 %r73, %r14, %r72;
shr.s32 %r15, %r73, 2;
and.b32 %r74, %r73, -4;
sub.s32 %r16, %r14, %r74;
shl.b32 %r284, %r285, 3;
mul.hi.u32 %r75, %r11, %r43;
add.s32 %r76, %r75, %r11;
shr.u32 %r77, %r76, %r44;
mad.lo.s32 %r78, %r77, %r46, %r285;
mad.lo.s32 %r79, %r282, %r40, %r78;
mad.lo.s32 %r80, %r4, %r36, %r79;
mul.wide.s32 %rd72, %r80, 144;
mul.lo.s32 %r81, %r283, %r41;
cvt.u64.u32 %rd40, %r81;
mul.lo.s32 %r82, %r11, %r47;
cvt.u64.u32 %rd41, %r82;
add.s64 %rd5, %rd40, %rd41;
mov.f32 %f135, 0f00000000;
cvta.to.global.u64 %rd42, %rd9;
cvta.to.global.u64 %rd46, %rd8;
mul.wide.s32 %rd57, %r16, 4;
mov.f32 %f136, %f135;

$L__BB159_10:
cvt.u64.u32 %rd43, %r284;
add.s64 %rd44, %rd5, %rd43;
shl.b32 %r119, %r16, 2;
shl.b32 %r120, %r15, 5;
add.s32 %r121, %r120, %r119;
cvt.s64.s32 %rd45, %r121;
add.s64 %rd47, %rd46, %rd45;
add.s64 %rd48, %rd47, %rd72;
ld.global.nc.u32 %r122, [%rd48+16];
ld.global.nc.u32 %r123, [%rd48+32];
shl.b32 %r124, %r15, 1;
cvt.u64.u32 %rd49, %r124;
and.b64 %rd50, %rd49, 2;
add.s64 %rd51, %rd46, %rd50;
add.s64 %rd52, %rd51, %rd72;
ld.global.nc.u16 %rs1, [%rd52+4];
cvt.u32.u16 %r125, %rs1;
ld.global.nc.u16 %rs2, [%rd52+8];
cvt.u32.u16 %r126, %rs2;
ld.global.nc.u16 %rs3, [%rd52+12];
cvt.u32.u16 %r127, %rs3;
and.b32 %r128, %r125, 16191;
and.b32 %r129, %r127, 3855;
and.b32 %r130, %r125, 49344;
shr.u32 %r131, %r130, 2;
or.b32 %r132, %r129, %r131;
setp.gt.s32 %p10, %r14, 7;
and.b32 %r133, %r126, 16191;
shr.u32 %r134, %r127, 4;
and.b32 %r135, %r134, 3855;
and.b32 %r136, %r126, 49344;
shr.u32 %r137, %r136, 2;
or.b32 %r138, %r135, %r137;
cvt.s64.s32 %rd53, %r124;
add.s64 %rd54, %rd44, %rd53;
mul.lo.s64 %rd55, %rd54, 36;
add.s64 %rd56, %rd42, %rd55;
ld.global.nc.u32 %r20, [%rd56];

{.reg .f16 low,high;
mov.b32 {low,high},%r20;
cvt.f32.f16 %f40, low;}


add.s64 %rd58, %rd56, %rd57;
ld.global.nc.u32 %r21, [%rd58+4];
ld.global.nc.u32 %r22, [%rd58+20];
ld.global.nc.u32 %r23, [%rd56+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r23;
cvt.f32.f16 %f41, low;}


ld.global.nc.u32 %r24, [%rd58+40];
ld.global.nc.u32 %r25, [%rd58+56];
add.s64 %rd59, %rd46, %rd72;
and.b32 %r86, %r122, 252645135;
and.b32 %r90, %r123, 252645135;
mov.u32 %r112, 0;

dp4a.s32.s32 %r85, %r86, %r21, %r112;


dp4a.s32.s32 %r89, %r90, %r22, %r85;

mov.u32 %r114, 16843009;

dp4a.s32.s32 %r93, %r114, %r21, %r112;


dp4a.s32.s32 %r97, %r114, %r22, %r93;

selp.b32 %r139, %r132, %r128, %p10;
cvt.u16.u32 %rs4, %r139;
and.b32 %r140, %r139, 63;
mul.lo.s32 %r141, %r89, %r140;
cvt.rn.f32.s32 %f44, %r141;
fma.rn.ftz.f32 %f45, %f40, %f44, 0f00000000;
selp.b32 %r142, %r138, %r133, %p10;
cvt.u16.u32 %rs5, %r142;
and.b32 %r143, %r142, 63;
mul.lo.s32 %r144, %r97, %r143;
cvt.rn.f32.s32 %f46, %r144;
fma.rn.ftz.f32 %f47, %f40, %f46, 0f00000000;
shr.u32 %r145, %r122, 4;
and.b32 %r102, %r145, 252645135;
shr.u32 %r146, %r123, 4;
and.b32 %r106, %r146, 252645135;

dp4a.s32.s32 %r101, %r102, %r24, %r112;


dp4a.s32.s32 %r105, %r106, %r25, %r101;


dp4a.s32.s32 %r109, %r114, %r24, %r112;


dp4a.s32.s32 %r113, %r114, %r25, %r109;

shr.u16 %rs6, %rs4, 8;
cvt.u32.u16 %r147, %rs6;
mul.lo.s32 %r148, %r105, %r147;
cvt.rn.f32.s32 %f48, %r148;
fma.rn.ftz.f32 %f49, %f41, %f48, %f45;
shr.u16 %rs7, %rs5, 8;
cvt.u32.u16 %r149, %rs7;
mul.lo.s32 %r150, %r113, %r149;
cvt.rn.f32.s32 %f50, %r150;
fma.rn.ftz.f32 %f51, %f41, %f50, %f47;
ld.global.nc.u32 %r117, [%rd59];

{.reg .f16 low,high;
mov.b32 {low,high},%r117;
cvt.f32.f16 %f42, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r117;
cvt.f32.f16 %f43, high;}


mul.ftz.f32 %f52, %f49, %f42;
mul.ftz.f32 %f53, %f51, %f43;
sub.ftz.f32 %f54, %f52, %f53;
add.ftz.f32 %f136, %f136, %f54;
setp.eq.s64 %p11, %rd12, 0;
@%p11 bra $L__BB159_12;

cvta.to.global.u64 %rd61, %rd12;
add.s64 %rd62, %rd61, %rd45;
add.s64 %rd63, %rd62, %rd72;
ld.global.u32 %r190, [%rd63+16];
ld.global.u32 %r191, [%rd63+32];
add.s64 %rd66, %rd61, %rd50;
add.s64 %rd67, %rd66, %rd72;
ld.global.u16 %r193, [%rd67+4];
ld.global.u16 %r194, [%rd67+8];
ld.global.u16 %r195, [%rd67+12];
and.b32 %r196, %r193, 16191;
and.b32 %r197, %r195, 3855;
and.b32 %r198, %r193, 49344;
shr.u32 %r199, %r198, 2;
or.b32 %r200, %r197, %r199;
and.b32 %r201, %r194, 16191;
shr.u32 %r202, %r195, 4;
and.b32 %r203, %r202, 3855;
and.b32 %r204, %r194, 49344;
shr.u32 %r205, %r204, 2;
or.b32 %r206, %r203, %r205;

{.reg .f16 low,high;
mov.b32 {low,high},%r20;
cvt.f32.f16 %f55, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r23;
cvt.f32.f16 %f56, low;}


add.s64 %rd68, %rd61, %rd72;
and.b32 %r154, %r190, 252645135;
and.b32 %r158, %r191, 252645135;

dp4a.s32.s32 %r153, %r154, %r21, %r112;


dp4a.s32.s32 %r157, %r158, %r22, %r153;


dp4a.s32.s32 %r161, %r114, %r21, %r112;


dp4a.s32.s32 %r165, %r114, %r22, %r161;

selp.b32 %r207, %r200, %r196, %p10;
cvt.u16.u32 %rs8, %r207;
and.b32 %r208, %r207, 63;
mul.lo.s32 %r209, %r157, %r208;
cvt.rn.f32.s32 %f59, %r209;
fma.rn.ftz.f32 %f60, %f55, %f59, 0f00000000;
selp.b32 %r210, %r206, %r201, %p10;
cvt.u16.u32 %rs9, %r210;
and.b32 %r211, %r210, 63;
mul.lo.s32 %r212, %r165, %r211;
cvt.rn.f32.s32 %f61, %r212;
fma.rn.ftz.f32 %f62, %f55, %f61, 0f00000000;
shr.u32 %r213, %r190, 4;
and.b32 %r170, %r213, 252645135;
shr.u32 %r214, %r191, 4;
and.b32 %r174, %r214, 252645135;

dp4a.s32.s32 %r169, %r170, %r24, %r112;


dp4a.s32.s32 %r173, %r174, %r25, %r169;


dp4a.s32.s32 %r177, %r114, %r24, %r112;


dp4a.s32.s32 %r181, %r114, %r25, %r177;

shr.u16 %rs10, %rs8, 8;
cvt.u32.u16 %r215, %rs10;
mul.lo.s32 %r216, %r173, %r215;
cvt.rn.f32.s32 %f63, %r216;
fma.rn.ftz.f32 %f64, %f56, %f63, %f60;
shr.u16 %rs11, %rs9, 8;
cvt.u32.u16 %r217, %rs11;
mul.lo.s32 %r218, %r181, %r217;
cvt.rn.f32.s32 %f65, %r218;
fma.rn.ftz.f32 %f66, %f56, %f65, %f62;
ld.global.u32 %r186, [%rd68];

{.reg .f16 low,high;
mov.b32 {low,high},%r186;
cvt.f32.f16 %f57, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r186;
cvt.f32.f16 %f58, high;}


mul.ftz.f32 %f67, %f64, %f57;
mul.ftz.f32 %f68, %f66, %f58;
sub.ftz.f32 %f69, %f67, %f68;
add.ftz.f32 %f135, %f135, %f69;

$L__BB159_12:
add.s32 %r284, %r284, 64;
add.s64 %rd72, %rd72, 1152;
add.s32 %r285, %r285, 8;
setp.lt.s32 %p13, %r285, %r63;
@%p13 bra $L__BB159_10;

$L__BB159_13:
setp.eq.s32 %p14, %r1, 0;
@%p14 bra $L__BB159_16;

setp.eq.s64 %p15, %rd12, 0;
add.s32 %r28, %r1, -1;
shl.b32 %r223, %r1, 7;
mov.u32 %r224, _ZZ13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r225, %r224, %r223;
shl.b32 %r226, %r2, 2;
add.s32 %r227, %r225, %r226;
st.shared.f32 [%r227+-128], %f136;
@%p15 bra $L__BB159_16;

shl.b32 %r229, %r28, 7;
mov.u32 %r230, _ZZ13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r231, %r230, %r229;
add.s32 %r233, %r231, %r226;
st.shared.f32 [%r233], %f135;

$L__BB159_16:
setp.ne.s32 %p16, %r1, 0;
bar.sync 0;
@%p16 bra $L__BB159_35;

setp.ne.s64 %p17, %rd12, 0;
shl.b32 %r236, %r2, 2;
mov.u32 %r237, _ZZ13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r29, %r237, %r236;
ld.shared.f32 %f70, [%r29];
add.ftz.f32 %f13, %f136, %f70;
@%p17 bra $L__BB159_19;
bra.uni $L__BB159_18;

$L__BB159_19:
mov.u32 %r240, _ZZ13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r241, %r240, %r236;
ld.shared.f32 %f74, [%r241];
add.ftz.f32 %f75, %f135, %f74;
ld.shared.f32 %f76, [%r29+128];
add.ftz.f32 %f77, %f13, %f76;
ld.shared.f32 %f78, [%r241+128];
add.ftz.f32 %f79, %f75, %f78;
ld.shared.f32 %f80, [%r29+256];
add.ftz.f32 %f139, %f77, %f80;
ld.shared.f32 %f81, [%r241+256];
add.ftz.f32 %f135, %f79, %f81;
bra.uni $L__BB159_20;

$L__BB159_18:
ld.shared.f32 %f71, [%r29+128];
add.ftz.f32 %f72, %f13, %f71;
ld.shared.f32 %f73, [%r29+256];
add.ftz.f32 %f139, %f72, %f73;

$L__BB159_20:
mov.b32 %r242, %f139;
mov.u32 %r243, 31;
mov.u32 %r244, 16;
mov.u32 %r245, -1;
shfl.sync.bfly.b32 %r246|%p18, %r242, %r244, %r243, %r245;
mov.b32 %f82, %r246;
add.ftz.f32 %f83, %f139, %f82;
mov.b32 %r247, %f83;
mov.u32 %r248, 8;
shfl.sync.bfly.b32 %r249|%p19, %r247, %r248, %r243, %r245;
mov.b32 %f84, %r249;
add.ftz.f32 %f85, %f83, %f84;
mov.b32 %r250, %f85;
mov.u32 %r251, 4;
shfl.sync.bfly.b32 %r252|%p20, %r250, %r251, %r243, %r245;
mov.b32 %f86, %r252;
add.ftz.f32 %f87, %f85, %f86;
mov.b32 %r253, %f87;
mov.u32 %r254, 2;
shfl.sync.bfly.b32 %r255|%p21, %r253, %r254, %r243, %r245;
mov.b32 %f88, %r255;
add.ftz.f32 %f19, %f87, %f88;
mov.b32 %r256, %f19;
mov.u32 %r257, 1;
shfl.sync.bfly.b32 %r30|%p1, %r256, %r257, %r243, %r245;
setp.eq.s64 %p22, %rd12, 0;
@%p22 bra $L__BB159_22;

mov.b32 %r258, %f135;
shfl.sync.bfly.b32 %r262|%p23, %r258, %r244, %r243, %r245;
mov.b32 %f89, %r262;
add.ftz.f32 %f90, %f135, %f89;
mov.b32 %r263, %f90;
shfl.sync.bfly.b32 %r265|%p24, %r263, %r248, %r243, %r245;
mov.b32 %f91, %r265;
add.ftz.f32 %f92, %f90, %f91;
mov.b32 %r266, %f92;
shfl.sync.bfly.b32 %r268|%p25, %r266, %r251, %r243, %r245;
mov.b32 %f93, %r268;
add.ftz.f32 %f94, %f92, %f93;
mov.b32 %r269, %f94;
shfl.sync.bfly.b32 %r271|%p26, %r269, %r254, %r243, %r245;
mov.b32 %f95, %r271;
add.ftz.f32 %f96, %f94, %f95;
mov.b32 %r272, %f96;
shfl.sync.bfly.b32 %r274|%p27, %r272, %r257, %r243, %r245;
mov.b32 %f97, %r274;
add.ftz.f32 %f135, %f96, %f97;

$L__BB159_22:
setp.ne.s32 %p28, %r2, 0;
mov.b32 %f98, %r30;
add.ftz.f32 %f22, %f19, %f98;
@%p28 bra $L__BB159_35;

add.ftz.f32 %f141, %f132, %f22;
@%p22 bra $L__BB159_34;

add.ftz.f32 %f24, %f131, %f135;
setp.gt.s32 %p30, %r31, 2;
@%p30 bra $L__BB159_28;

setp.eq.s32 %p33, %r31, 1;
@%p33 bra $L__BB159_32;

setp.eq.s32 %p34, %r31, 2;
@%p34 bra $L__BB159_27;
bra.uni $L__BB159_33;

$L__BB159_27:
mul.ftz.f32 %f118, %f24, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f119, %f118;
add.ftz.f32 %f120, %f119, 0f3F800000;
div.approx.ftz.f32 %f121, %f24, %f120;
mul.ftz.f32 %f141, %f141, %f121;
bra.uni $L__BB159_34;

$L__BB159_28:
setp.eq.s32 %p31, %r31, 3;
@%p31 bra $L__BB159_31;

setp.ne.s32 %p32, %r31, 6;
@%p32 bra $L__BB159_33;

min.ftz.f32 %f99, %f24, %f31;
neg.ftz.f32 %f100, %f31;
min.ftz.f32 %f101, %f141, %f31;
max.ftz.f32 %f102, %f101, %f100;
mul.ftz.f32 %f103, %f99, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f104, %f103;
add.ftz.f32 %f105, %f104, 0f3F800000;
div.approx.ftz.f32 %f106, %f99, %f105;
mul.ftz.f32 %f141, %f102, %f106;
bra.uni $L__BB159_34;

$L__BB159_32:
mul.ftz.f32 %f122, %f24, 0f3F000000;
mul.ftz.f32 %f123, %f24, 0f3D372713;
fma.rn.ftz.f32 %f124, %f24, %f123, 0f3F800000;
mul.ftz.f32 %f125, %f24, 0f3F4C422A;
mul.ftz.f32 %f126, %f125, %f124;
tanh.approx.f32 %f127, %f126;
add.ftz.f32 %f128, %f127, 0f3F800000;
mul.ftz.f32 %f129, %f122, %f128;
mul.ftz.f32 %f141, %f141, %f129;
bra.uni $L__BB159_34;

$L__BB159_31:
mov.f32 %f107, 0f40E00000;
min.ftz.f32 %f108, %f24, %f107;
min.ftz.f32 %f109, %f141, %f107;
mov.f32 %f110, 0fC0E00000;
max.ftz.f32 %f111, %f109, %f110;
mul.ftz.f32 %f112, %f108, 0f3FD9DB23;
mul.ftz.f32 %f113, %f112, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f114, %f113;
add.ftz.f32 %f115, %f114, 0f3F800000;
div.approx.ftz.f32 %f116, %f108, %f115;
add.ftz.f32 %f117, %f111, 0f3F800000;
mul.ftz.f32 %f141, %f117, %f116;
bra.uni $L__BB159_34;

$L__BB159_33:
mul.ftz.f32 %f141, %f141, %f24;

$L__BB159_34:
cvta.to.global.u64 %rd69, %rd16;
mov.u32 %r278, %ctaid.y;
mad.lo.s32 %r280, %r278, %r42, %r4;
mad.lo.s32 %r281, %r11, %r48, %r280;
mul.wide.u32 %rd70, %r281, 4;
add.s64 %rd71, %rd69, %rd70;
st.global.f32 [%rd71], %f141;

$L__BB159_35:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<16>;
.reg .b16 %rs<22>;
.reg .f32 %f<76>;
.reg .b32 %r<312>;
.reg .b64 %rd<55>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[384];

ld.param.u64 %rd14, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd12, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd13, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
cvta.to.global.u64 %rd1, %rd15;
cvta.to.global.u64 %rd54, %rd14;
mov.u32 %r1, %tid.y;
shl.b32 %r44, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r44, %r2;
shr.u32 %r4, %r27, 8;
mov.u32 %r308, %ctaid.y;
setp.eq.s64 %p1, %rd12, 0;
@%p1 bra $L__BB160_2;

cvta.to.global.u64 %rd16, %rd12;
mul.wide.u32 %rd17, %r308, 4;
add.s64 %rd18, %rd16, %rd17;
ld.global.nc.u32 %r307, [%rd18];
mul.hi.u32 %r45, %r308, %r28;
add.s32 %r46, %r45, %r308;
shr.u32 %r47, %r46, %r29;
mul.lo.s32 %r48, %r47, %r30;
sub.s32 %r308, %r308, %r48;
bra.uni $L__BB160_3;

$L__BB160_2:
mul.hi.u32 %r49, %r308, %r32;
add.s32 %r50, %r49, %r308;
shr.u32 %r307, %r50, %r33;

$L__BB160_3:
mov.u32 %r11, %ctaid.z;
mul.hi.u32 %r51, %r11, %r38;
add.s32 %r52, %r51, %r11;
shr.u32 %r53, %r52, %r39;
mul.lo.s32 %r12, %r53, %r41;
mul.lo.s32 %r13, %r307, %r35;
mov.u32 %r54, %ctaid.x;
mul.lo.s32 %r14, %r54, %r31;
shr.s32 %r55, %r3, 31;
shr.u32 %r56, %r55, 28;
add.s32 %r57, %r3, %r56;
shr.s32 %r311, %r57, 4;
setp.ge.s32 %p2, %r311, %r4;
mov.f32 %f75, 0f00000000;
@%p2 bra $L__BB160_9;

mul.lo.s32 %r58, %r308, %r36;
cvt.u64.u32 %rd19, %r58;
and.b32 %r62, %r57, -16;
sub.s32 %r16, %r3, %r62;
shr.s32 %r63, %r16, 31;
shr.u32 %r64, %r63, 30;
add.s32 %r65, %r16, %r64;
shr.s32 %r66, %r65, 2;
shl.b32 %r67, %r66, 1;
shl.b32 %r68, %r66, 5;
and.b32 %r69, %r65, -4;
sub.s32 %r70, %r16, %r69;
shl.b32 %r71, %r70, 2;
add.s32 %r72, %r68, %r71;
cvt.s64.s32 %rd3, %r72;
and.b32 %r73, %r66, 1;
cvt.u64.u32 %rd4, %r73;
cvt.s64.s32 %rd20, %r67;
mul.lo.s32 %r74, %r11, %r42;
cvt.u64.u32 %rd21, %r74;
add.s64 %rd22, %rd21, %rd20;
add.s64 %rd5, %rd22, %rd19;
cvt.s64.s32 %rd6, %r70;
not.b32 %r75, %r311;
add.s32 %r17, %r4, %r75;
and.b32 %r76, %r17, 8;
setp.ne.s32 %p3, %r76, 0;
mov.f32 %f75, 0f00000000;
@%p3 bra $L__BB160_6;

setp.gt.s32 %p4, %r16, 7;
shl.b32 %r113, %r311, 3;
cvt.u64.u32 %rd23, %r113;
add.s32 %r114, %r13, %r14;
add.s32 %r115, %r114, %r12;
add.s32 %r116, %r115, %r311;
mul.wide.s32 %rd24, %r116, 144;
add.s64 %rd25, %rd54, %rd24;
add.s64 %rd26, %rd25, %rd3;
ld.global.nc.u32 %r117, [%rd26+16];
ld.global.nc.u32 %r118, [%rd26+32];
shl.b64 %rd27, %rd4, 1;
add.s64 %rd28, %rd25, %rd27;
ld.global.nc.u16 %rs1, [%rd28+4];
cvt.u32.u16 %r119, %rs1;
ld.global.nc.u16 %rs2, [%rd28+8];
cvt.u32.u16 %r120, %rs2;
ld.global.nc.u16 %rs3, [%rd28+12];
cvt.u32.u16 %r121, %rs3;
and.b32 %r122, %r119, 16191;
and.b32 %r123, %r121, 3855;
and.b32 %r124, %r119, 49344;
shr.u32 %r125, %r124, 2;
or.b32 %r126, %r123, %r125;
and.b32 %r127, %r120, 16191;
shr.u32 %r128, %r121, 4;
and.b32 %r129, %r128, 3855;
and.b32 %r130, %r120, 49344;
shr.u32 %r131, %r130, 2;
or.b32 %r132, %r129, %r131;
add.s64 %rd29, %rd5, %rd23;
mul.lo.s64 %rd30, %rd29, 36;
add.s64 %rd31, %rd1, %rd30;
ld.global.nc.u32 %r77, [%rd31];

{.reg .f16 low,high;
mov.b32 {low,high},%r77;
cvt.f32.f16 %f11, low;}


shl.b64 %rd32, %rd6, 2;
add.s64 %rd33, %rd31, %rd32;
ld.global.nc.u32 %r81, [%rd33+4];
ld.global.nc.u32 %r85, [%rd33+20];
ld.global.nc.u32 %r78, [%rd31+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r78;
cvt.f32.f16 %f12, low;}


ld.global.nc.u32 %r97, [%rd33+40];
ld.global.nc.u32 %r101, [%rd33+56];
and.b32 %r80, %r117, 252645135;
and.b32 %r84, %r118, 252645135;
mov.u32 %r106, 0;

dp4a.s32.s32 %r79, %r80, %r81, %r106;


dp4a.s32.s32 %r83, %r84, %r85, %r79;

mov.u32 %r108, 16843009;

dp4a.s32.s32 %r87, %r108, %r81, %r106;


dp4a.s32.s32 %r91, %r108, %r85, %r87;

selp.b32 %r133, %r126, %r122, %p4;
cvt.u16.u32 %rs4, %r133;
and.b32 %r134, %r133, 63;
mul.lo.s32 %r135, %r83, %r134;
cvt.rn.f32.s32 %f15, %r135;
fma.rn.ftz.f32 %f16, %f11, %f15, 0f00000000;
selp.b32 %r136, %r132, %r127, %p4;
cvt.u16.u32 %rs5, %r136;
and.b32 %r137, %r136, 63;
mul.lo.s32 %r138, %r91, %r137;
cvt.rn.f32.s32 %f17, %r138;
fma.rn.ftz.f32 %f18, %f11, %f17, 0f00000000;
shr.u32 %r139, %r117, 4;
and.b32 %r96, %r139, 252645135;
shr.u32 %r140, %r118, 4;
and.b32 %r100, %r140, 252645135;

dp4a.s32.s32 %r95, %r96, %r97, %r106;


dp4a.s32.s32 %r99, %r100, %r101, %r95;


dp4a.s32.s32 %r103, %r108, %r97, %r106;


dp4a.s32.s32 %r107, %r108, %r101, %r103;

shr.u16 %rs6, %rs4, 8;
cvt.u32.u16 %r141, %rs6;
mul.lo.s32 %r142, %r99, %r141;
cvt.rn.f32.s32 %f19, %r142;
fma.rn.ftz.f32 %f20, %f12, %f19, %f16;
shr.u16 %rs7, %rs5, 8;
cvt.u32.u16 %r143, %rs7;
mul.lo.s32 %r144, %r107, %r143;
cvt.rn.f32.s32 %f21, %r144;
fma.rn.ftz.f32 %f22, %f12, %f21, %f18;
ld.global.nc.u32 %r111, [%rd25];

{.reg .f16 low,high;
mov.b32 {low,high},%r111;
cvt.f32.f16 %f13, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r111;
cvt.f32.f16 %f14, high;}


mul.ftz.f32 %f23, %f20, %f13;
mul.ftz.f32 %f24, %f22, %f14;
sub.ftz.f32 %f25, %f23, %f24;
add.ftz.f32 %f75, %f25, 0f00000000;
add.s32 %r311, %r311, 8;

$L__BB160_6:
and.b32 %r145, %r17, -8;
setp.eq.s32 %p5, %r145, 0;
@%p5 bra $L__BB160_9;

add.s32 %r146, %r311, %r12;
add.s32 %r147, %r146, %r13;
add.s32 %r148, %r147, %r14;
mul.wide.s32 %rd7, %r148, 144;
shl.b32 %r149, %r311, 3;
add.s32 %r310, %r149, 64;
shl.b64 %rd34, %rd4, 1;
or.b64 %rd35, %rd34, %rd7;
or.b64 %rd8, %rd35, 8;
add.s64 %rd36, %rd7, %rd3;
add.s64 %rd9, %rd36, 16;

$L__BB160_8:
add.s32 %r222, %r310, -64;
cvt.u64.u32 %rd37, %r222;
add.s64 %rd38, %rd54, %rd9;
ld.global.nc.u32 %r223, [%rd38];
ld.global.nc.u32 %r224, [%rd38+16];
add.s64 %rd39, %rd54, %rd8;
ld.global.nc.u16 %rs8, [%rd39+-4];
cvt.u32.u16 %r225, %rs8;
ld.global.nc.u16 %rs9, [%rd39];
cvt.u32.u16 %r226, %rs9;
ld.global.nc.u16 %rs10, [%rd39+4];
cvt.u32.u16 %r227, %rs10;
and.b32 %r228, %r225, 16191;
and.b32 %r229, %r227, 3855;
and.b32 %r230, %r225, 49344;
shr.u32 %r231, %r230, 2;
or.b32 %r232, %r229, %r231;
setp.gt.s32 %p6, %r16, 7;
and.b32 %r233, %r226, 16191;
shr.u32 %r234, %r227, 4;
and.b32 %r235, %r234, 3855;
and.b32 %r236, %r226, 49344;
shr.u32 %r237, %r236, 2;
or.b32 %r238, %r235, %r237;
add.s64 %rd40, %rd5, %rd37;
mul.lo.s64 %rd41, %rd40, 36;
add.s64 %rd42, %rd1, %rd41;
ld.global.nc.u32 %r150, [%rd42];

{.reg .f16 low,high;
mov.b32 {low,high},%r150;
cvt.f32.f16 %f26, low;}


shl.b64 %rd43, %rd6, 2;
add.s64 %rd44, %rd42, %rd43;
ld.global.nc.u32 %r154, [%rd44+4];
ld.global.nc.u32 %r158, [%rd44+20];
ld.global.nc.u32 %r151, [%rd42+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r151;
cvt.f32.f16 %f27, low;}


ld.global.nc.u32 %r170, [%rd44+40];
ld.global.nc.u32 %r174, [%rd44+56];
add.s64 %rd45, %rd54, %rd7;
and.b32 %r153, %r223, 252645135;
and.b32 %r157, %r224, 252645135;
mov.u32 %r215, 0;

dp4a.s32.s32 %r152, %r153, %r154, %r215;


dp4a.s32.s32 %r156, %r157, %r158, %r152;

mov.u32 %r217, 16843009;

dp4a.s32.s32 %r160, %r217, %r154, %r215;


dp4a.s32.s32 %r164, %r217, %r158, %r160;

selp.b32 %r239, %r232, %r228, %p6;
cvt.u16.u32 %rs11, %r239;
and.b32 %r240, %r239, 63;
mul.lo.s32 %r241, %r156, %r240;
cvt.rn.f32.s32 %f34, %r241;
fma.rn.ftz.f32 %f35, %f26, %f34, 0f00000000;
selp.b32 %r242, %r238, %r233, %p6;
cvt.u16.u32 %rs12, %r242;
and.b32 %r243, %r242, 63;
mul.lo.s32 %r244, %r164, %r243;
cvt.rn.f32.s32 %f36, %r244;
fma.rn.ftz.f32 %f37, %f26, %f36, 0f00000000;
shr.u32 %r245, %r223, 4;
and.b32 %r169, %r245, 252645135;
shr.u32 %r246, %r224, 4;
and.b32 %r173, %r246, 252645135;

dp4a.s32.s32 %r168, %r169, %r170, %r215;


dp4a.s32.s32 %r172, %r173, %r174, %r168;


dp4a.s32.s32 %r176, %r217, %r170, %r215;


dp4a.s32.s32 %r180, %r217, %r174, %r176;

shr.u16 %rs13, %rs11, 8;
cvt.u32.u16 %r247, %rs13;
mul.lo.s32 %r248, %r172, %r247;
cvt.rn.f32.s32 %f38, %r248;
fma.rn.ftz.f32 %f39, %f27, %f38, %f35;
shr.u16 %rs14, %rs12, 8;
cvt.u32.u16 %r249, %rs14;
mul.lo.s32 %r250, %r180, %r249;
cvt.rn.f32.s32 %f40, %r250;
fma.rn.ftz.f32 %f41, %f27, %f40, %f37;
ld.global.nc.u32 %r184, [%rd45];

{.reg .f16 low,high;
mov.b32 {low,high},%r184;
cvt.f32.f16 %f28, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r184;
cvt.f32.f16 %f29, high;}


mul.ftz.f32 %f42, %f39, %f28;
mul.ftz.f32 %f43, %f41, %f29;
sub.ftz.f32 %f44, %f42, %f43;
add.ftz.f32 %f45, %f75, %f44;
ld.global.nc.u32 %r251, [%rd38+1152];
ld.global.nc.u32 %r252, [%rd38+1168];
ld.global.nc.u16 %rs15, [%rd39+1148];
cvt.u32.u16 %r253, %rs15;
ld.global.nc.u16 %rs16, [%rd39+1152];
cvt.u32.u16 %r254, %rs16;
ld.global.nc.u16 %rs17, [%rd39+1156];
cvt.u32.u16 %r255, %rs17;
and.b32 %r256, %r253, 16191;
and.b32 %r257, %r255, 3855;
and.b32 %r258, %r253, 49344;
shr.u32 %r259, %r258, 2;
or.b32 %r260, %r257, %r259;
and.b32 %r261, %r254, 16191;
shr.u32 %r262, %r255, 4;
and.b32 %r263, %r262, 3855;
and.b32 %r264, %r254, 49344;
shr.u32 %r265, %r264, 2;
or.b32 %r266, %r263, %r265;
cvt.u64.u32 %rd46, %r310;
add.s64 %rd47, %rd5, %rd46;
mul.lo.s64 %rd48, %rd47, 36;
add.s64 %rd49, %rd1, %rd48;
ld.global.nc.u32 %r186, [%rd49];

{.reg .f16 low,high;
mov.b32 {low,high},%r186;
cvt.f32.f16 %f30, low;}


add.s64 %rd50, %rd49, %rd43;
ld.global.nc.u32 %r190, [%rd50+4];
ld.global.nc.u32 %r194, [%rd50+20];
ld.global.nc.u32 %r187, [%rd49+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r187;
cvt.f32.f16 %f31, low;}


ld.global.nc.u32 %r206, [%rd50+40];
ld.global.nc.u32 %r210, [%rd50+56];
and.b32 %r189, %r251, 252645135;
and.b32 %r193, %r252, 252645135;

dp4a.s32.s32 %r188, %r189, %r190, %r215;


dp4a.s32.s32 %r192, %r193, %r194, %r188;


dp4a.s32.s32 %r196, %r217, %r190, %r215;


dp4a.s32.s32 %r200, %r217, %r194, %r196;

selp.b32 %r267, %r260, %r256, %p6;
cvt.u16.u32 %rs18, %r267;
and.b32 %r268, %r267, 63;
mul.lo.s32 %r269, %r192, %r268;
cvt.rn.f32.s32 %f46, %r269;
fma.rn.ftz.f32 %f47, %f30, %f46, 0f00000000;
selp.b32 %r270, %r266, %r261, %p6;
cvt.u16.u32 %rs19, %r270;
and.b32 %r271, %r270, 63;
mul.lo.s32 %r272, %r200, %r271;
cvt.rn.f32.s32 %f48, %r272;
fma.rn.ftz.f32 %f49, %f30, %f48, 0f00000000;
shr.u32 %r273, %r251, 4;
and.b32 %r205, %r273, 252645135;
shr.u32 %r274, %r252, 4;
and.b32 %r209, %r274, 252645135;

dp4a.s32.s32 %r204, %r205, %r206, %r215;


dp4a.s32.s32 %r208, %r209, %r210, %r204;


dp4a.s32.s32 %r212, %r217, %r206, %r215;


dp4a.s32.s32 %r216, %r217, %r210, %r212;

shr.u16 %rs20, %rs18, 8;
cvt.u32.u16 %r275, %rs20;
mul.lo.s32 %r276, %r208, %r275;
cvt.rn.f32.s32 %f50, %r276;
fma.rn.ftz.f32 %f51, %f31, %f50, %f47;
shr.u16 %rs21, %rs19, 8;
cvt.u32.u16 %r277, %rs21;
mul.lo.s32 %r278, %r216, %r277;
cvt.rn.f32.s32 %f52, %r278;
fma.rn.ftz.f32 %f53, %f31, %f52, %f49;
ld.global.nc.u32 %r220, [%rd45+1152];

{.reg .f16 low,high;
mov.b32 {low,high},%r220;
cvt.f32.f16 %f32, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r220;
cvt.f32.f16 %f33, high;}


mul.ftz.f32 %f54, %f51, %f32;
mul.ftz.f32 %f55, %f53, %f33;
sub.ftz.f32 %f56, %f54, %f55;
add.ftz.f32 %f75, %f45, %f56;
add.s64 %rd54, %rd54, 2304;
add.s32 %r310, %r310, 128;
add.s32 %r311, %r311, 16;
setp.lt.s32 %p7, %r311, %r4;
@%p7 bra $L__BB160_8;

$L__BB160_9:
setp.eq.s32 %p8, %r1, 0;
@%p8 bra $L__BB160_11;

shl.b32 %r279, %r1, 7;
mov.u32 %r280, _ZZ13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r281, %r280, %r279;
shl.b32 %r282, %r2, 2;
add.s32 %r283, %r281, %r282;
st.shared.f32 [%r283+-128], %f75;

$L__BB160_11:
setp.ne.s32 %p9, %r1, 0;
bar.sync 0;
@%p9 bra $L__BB160_14;

mov.u32 %r284, %ctaid.y;
mad.lo.s32 %r286, %r284, %r37, %r54;
mad.lo.s32 %r26, %r11, %r43, %r286;
shl.b32 %r287, %r2, 2;
mov.u32 %r288, _ZZ13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r289, %r288, %r287;
mov.u32 %r290, 2;
ld.shared.f32 %f57, [%r289];
add.ftz.f32 %f58, %f75, %f57;
ld.shared.f32 %f59, [%r289+128];
add.ftz.f32 %f60, %f58, %f59;
ld.shared.f32 %f61, [%r289+256];
add.ftz.f32 %f62, %f60, %f61;
mov.b32 %r291, %f62;
mov.u32 %r292, 31;
mov.u32 %r293, 16;
mov.u32 %r294, -1;
shfl.sync.bfly.b32 %r295|%p10, %r291, %r293, %r292, %r294;
mov.b32 %f63, %r295;
add.ftz.f32 %f64, %f62, %f63;
mov.b32 %r296, %f64;
mov.u32 %r297, 8;
shfl.sync.bfly.b32 %r298|%p11, %r296, %r297, %r292, %r294;
mov.b32 %f65, %r298;
add.ftz.f32 %f66, %f64, %f65;
mov.b32 %r299, %f66;
mov.u32 %r300, 4;
shfl.sync.bfly.b32 %r301|%p12, %r299, %r300, %r292, %r294;
mov.b32 %f67, %r301;
add.ftz.f32 %f68, %f66, %f67;
mov.b32 %r302, %f68;
shfl.sync.bfly.b32 %r303|%p13, %r302, %r290, %r292, %r294;
mov.b32 %f69, %r303;
add.ftz.f32 %f70, %f68, %f69;
mov.b32 %r304, %f70;
mov.u32 %r305, 1;
shfl.sync.bfly.b32 %r306|%p14, %r304, %r305, %r292, %r294;
mov.b32 %f71, %r306;
add.ftz.f32 %f7, %f70, %f71;
setp.ne.s32 %p15, %r2, 0;
@%p15 bra $L__BB160_14;

cvta.to.global.u64 %rd51, %rd13;
mul.wide.u32 %rd52, %r26, 4;
add.s64 %rd53, %rd51, %rd52;
st.global.f32 [%rd53], %f7;

$L__BB160_14:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type12ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type12ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type12ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type12ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type12ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type12ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type12ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type12ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type12ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<36>;
.reg .b16 %rs<15>;
.reg .f32 %f<153>;
.reg .b32 %r<363>;
.reg .b64 %rd<57>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type12ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type12ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type12ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type12ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type12ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r19, [_Z13mul_mat_vec_qIL9ggml_type12ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type12ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type12ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type12ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type12ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type12ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type12ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type12ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type12ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type12ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type12ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type12ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type12ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
cvta.to.global.u64 %rd1, %rd18;
mov.u32 %r1, %tid.y;
shl.b32 %r35, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r35, %r2;
mov.u32 %r36, %ctaid.x;
shl.b32 %r4, %r36, 1;
mov.u32 %r5, %ctaid.y;
shr.s32 %r37, %r3, 31;
shr.u32 %r38, %r37, 28;
add.s32 %r39, %r3, %r38;
shr.s32 %r362, %r39, 4;
shr.u32 %r7, %r34, 8;
setp.ge.s32 %p3, %r362, %r7;
mov.f32 %f149, 0f00000000;
mov.f32 %f150, %f149;
mov.f32 %f151, %f149;
mov.f32 %f152, %f149;
@%p3 bra $L__BB161_3;

mov.u32 %r40, %ctaid.z;
mul.hi.u32 %r41, %r40, %r28;
add.s32 %r42, %r41, %r40;
shr.u32 %r43, %r42, %r29;
mul.lo.s32 %r44, %r43, %r31;
mul.hi.u32 %r45, %r5, %r22;
add.s32 %r46, %r45, %r5;
shr.u32 %r47, %r46, %r23;
mul.lo.s32 %r48, %r47, %r25;
mul.lo.s32 %r49, %r5, %r26;
cvt.u64.u32 %rd19, %r49;
cvta.to.global.u64 %rd56, %rd16;
and.b32 %r53, %r39, -16;
sub.s32 %r8, %r3, %r53;
shr.s32 %r54, %r8, 31;
shr.u32 %r55, %r54, 30;
add.s32 %r56, %r8, %r55;
shr.s32 %r57, %r56, 2;
shl.b32 %r58, %r57, 1;
shl.b32 %r59, %r57, 5;
and.b32 %r60, %r56, -4;
sub.s32 %r61, %r8, %r60;
shl.b32 %r62, %r61, 2;
add.s32 %r63, %r59, %r62;
cvt.s64.s32 %rd20, %r63;
cvt.s64.s32 %rd21, %r58;
add.s64 %rd22, %rd19, %rd21;
mul.lo.s32 %r64, %r40, %r32;
cvt.u64.u32 %rd23, %r64;
add.s64 %rd3, %rd22, %rd23;
cvt.s64.s32 %rd4, %r61;
shl.b32 %r360, %r362, 3;
add.s32 %r361, %r20, %r360;
add.s32 %r65, %r362, %r44;
add.s32 %r66, %r65, %r48;
add.s32 %r67, %r4, 1;
mad.lo.s32 %r68, %r19, %r67, %r66;
mul.wide.s32 %rd5, %r68, 144;
mad.lo.s32 %r69, %r4, %r19, %r362;
add.s32 %r70, %r69, %r44;
add.s32 %r71, %r70, %r48;
mul.wide.s32 %rd6, %r71, 144;
add.s64 %rd24, %rd6, %rd20;
add.s64 %rd7, %rd24, 32;
cvt.u64.u32 %rd25, %r58;
and.b64 %rd26, %rd25, 2;
or.b64 %rd27, %rd26, %rd5;
or.b64 %rd8, %rd27, 8;
or.b64 %rd28, %rd26, %rd6;
or.b64 %rd9, %rd28, 12;
add.s64 %rd29, %rd5, %rd20;
add.s64 %rd10, %rd29, 16;
cvta.to.global.u64 %rd11, %rd17;
mov.f32 %f149, 0f00000000;
shl.b64 %rd36, %rd4, 2;

$L__BB161_2:
cvt.u64.u32 %rd30, %r360;
add.s64 %rd31, %rd3, %rd30;
add.s64 %rd32, %rd56, %rd7;
ld.global.nc.u32 %r216, [%rd32+-16];
ld.global.nc.u32 %r217, [%rd32];
add.s64 %rd33, %rd56, %rd9;
ld.global.nc.u16 %rs1, [%rd33+-8];
cvt.u32.u16 %r218, %rs1;
ld.global.nc.u16 %rs2, [%rd33+-4];
cvt.u32.u16 %r219, %rs2;
ld.global.nc.u16 %rs3, [%rd33];
cvt.u32.u16 %r220, %rs3;
and.b32 %r221, %r218, 16191;
and.b32 %r222, %r220, 3855;
and.b32 %r223, %r218, 49344;
shr.u32 %r224, %r223, 2;
or.b32 %r225, %r222, %r224;
setp.gt.s32 %p4, %r8, 7;
and.b32 %r226, %r219, 16191;
shr.u32 %r227, %r220, 4;
and.b32 %r228, %r227, 3855;
and.b32 %r229, %r219, 49344;
shr.u32 %r230, %r229, 2;
or.b32 %r231, %r228, %r230;
mul.lo.s64 %rd34, %rd31, 36;
add.s64 %rd35, %rd11, %rd34;
ld.global.nc.u32 %r72, [%rd35];

{.reg .f16 low,high;
mov.b32 {low,high},%r72;
cvt.f32.f16 %f25, low;}


add.s64 %rd37, %rd35, %rd36;
ld.global.nc.u32 %r76, [%rd37+4];
ld.global.nc.u32 %r80, [%rd37+20];
ld.global.nc.u32 %r73, [%rd35+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r73;
cvt.f32.f16 %f26, low;}


ld.global.nc.u32 %r92, [%rd37+40];
ld.global.nc.u32 %r96, [%rd37+56];
add.s64 %rd38, %rd56, %rd6;
and.b32 %r147, %r216, 252645135;
and.b32 %r151, %r217, 252645135;
mov.u32 %r209, 0;

dp4a.s32.s32 %r74, %r147, %r76, %r209;


dp4a.s32.s32 %r78, %r151, %r80, %r74;

mov.u32 %r211, 16843009;

dp4a.s32.s32 %r82, %r211, %r76, %r209;


dp4a.s32.s32 %r86, %r211, %r80, %r82;

selp.b32 %r232, %r225, %r221, %p4;
cvt.u16.u32 %rs4, %r232;
and.b32 %r233, %r232, 63;
mul.lo.s32 %r234, %r78, %r233;
cvt.rn.f32.s32 %f41, %r234;
fma.rn.ftz.f32 %f42, %f25, %f41, 0f00000000;
selp.b32 %r235, %r231, %r226, %p4;
cvt.u16.u32 %rs5, %r235;
and.b32 %r236, %r235, 63;
mul.lo.s32 %r237, %r86, %r236;
cvt.rn.f32.s32 %f43, %r237;
fma.rn.ftz.f32 %f44, %f25, %f43, 0f00000000;
shr.u32 %r238, %r216, 4;
and.b32 %r163, %r238, 252645135;
shr.u32 %r239, %r217, 4;
and.b32 %r167, %r239, 252645135;

dp4a.s32.s32 %r90, %r163, %r92, %r209;


dp4a.s32.s32 %r94, %r167, %r96, %r90;


dp4a.s32.s32 %r98, %r211, %r92, %r209;


dp4a.s32.s32 %r102, %r211, %r96, %r98;

shr.u16 %rs6, %rs4, 8;
cvt.u32.u16 %r240, %rs6;
mul.lo.s32 %r241, %r94, %r240;
cvt.rn.f32.s32 %f45, %r241;
fma.rn.ftz.f32 %f46, %f26, %f45, %f42;
shr.u16 %rs7, %rs5, 8;
cvt.u32.u16 %r242, %rs7;
mul.lo.s32 %r243, %r102, %r242;
cvt.rn.f32.s32 %f47, %r243;
fma.rn.ftz.f32 %f48, %f26, %f47, %f44;
ld.global.nc.u32 %r106, [%rd38];

{.reg .f16 low,high;
mov.b32 {low,high},%r106;
cvt.f32.f16 %f27, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r106;
cvt.f32.f16 %f28, high;}


mul.ftz.f32 %f49, %f46, %f27;
mul.ftz.f32 %f50, %f48, %f28;
sub.ftz.f32 %f51, %f49, %f50;
add.ftz.f32 %f152, %f51, %f152;
add.s64 %rd39, %rd56, %rd10;
ld.global.nc.u32 %r244, [%rd39];
ld.global.nc.u32 %r245, [%rd39+16];
add.s64 %rd40, %rd56, %rd8;
ld.global.nc.u16 %rs8, [%rd40+-4];
cvt.u32.u16 %r246, %rs8;
ld.global.nc.u16 %rs9, [%rd40];
cvt.u32.u16 %r247, %rs9;
ld.global.nc.u16 %rs10, [%rd40+4];
cvt.u32.u16 %r248, %rs10;
and.b32 %r249, %r246, 16191;
and.b32 %r250, %r248, 3855;
and.b32 %r251, %r246, 49344;
shr.u32 %r252, %r251, 2;
or.b32 %r253, %r250, %r252;
and.b32 %r254, %r247, 16191;
shr.u32 %r255, %r248, 4;
and.b32 %r256, %r255, 3855;
and.b32 %r257, %r247, 49344;
shr.u32 %r258, %r257, 2;
or.b32 %r259, %r256, %r258;

{.reg .f16 low,high;
mov.b32 {low,high},%r72;
cvt.f32.f16 %f29, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r73;
cvt.f32.f16 %f30, low;}


add.s64 %rd41, %rd56, %rd5;
and.b32 %r183, %r244, 252645135;
and.b32 %r187, %r245, 252645135;

dp4a.s32.s32 %r110, %r183, %r76, %r209;


dp4a.s32.s32 %r114, %r187, %r80, %r110;


dp4a.s32.s32 %r118, %r211, %r76, %r209;


dp4a.s32.s32 %r122, %r211, %r80, %r118;

selp.b32 %r260, %r253, %r249, %p4;
cvt.u16.u32 %rs11, %r260;
and.b32 %r261, %r260, 63;
mul.lo.s32 %r262, %r114, %r261;
cvt.rn.f32.s32 %f52, %r262;
fma.rn.ftz.f32 %f53, %f29, %f52, 0f00000000;
selp.b32 %r263, %r259, %r254, %p4;
cvt.u16.u32 %rs12, %r263;
and.b32 %r264, %r263, 63;
mul.lo.s32 %r265, %r122, %r264;
cvt.rn.f32.s32 %f54, %r265;
fma.rn.ftz.f32 %f55, %f29, %f54, 0f00000000;
shr.u32 %r266, %r244, 4;
and.b32 %r199, %r266, 252645135;
shr.u32 %r267, %r245, 4;
and.b32 %r203, %r267, 252645135;

dp4a.s32.s32 %r126, %r199, %r92, %r209;


dp4a.s32.s32 %r130, %r203, %r96, %r126;


dp4a.s32.s32 %r134, %r211, %r92, %r209;


dp4a.s32.s32 %r138, %r211, %r96, %r134;

shr.u16 %rs13, %rs11, 8;
cvt.u32.u16 %r268, %rs13;
mul.lo.s32 %r269, %r130, %r268;
cvt.rn.f32.s32 %f56, %r269;
fma.rn.ftz.f32 %f57, %f30, %f56, %f53;
shr.u16 %rs14, %rs12, 8;
cvt.u32.u16 %r270, %rs14;
mul.lo.s32 %r271, %r138, %r270;
cvt.rn.f32.s32 %f58, %r271;
fma.rn.ftz.f32 %f59, %f30, %f58, %f55;
ld.global.nc.u32 %r142, [%rd41];

{.reg .f16 low,high;
mov.b32 {low,high},%r142;
cvt.f32.f16 %f31, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r142;
cvt.f32.f16 %f32, high;}


mul.ftz.f32 %f60, %f57, %f31;
mul.ftz.f32 %f61, %f59, %f32;
sub.ftz.f32 %f62, %f60, %f61;
add.ftz.f32 %f151, %f62, %f151;
cvt.u64.u32 %rd42, %r361;
add.s64 %rd43, %rd3, %rd42;
mul.lo.s64 %rd44, %rd43, 36;
add.s64 %rd45, %rd11, %rd44;
ld.global.nc.u32 %r144, [%rd45];

{.reg .f16 low,high;
mov.b32 {low,high},%r144;
cvt.f32.f16 %f33, low;}


add.s64 %rd46, %rd45, %rd36;
ld.global.nc.u32 %r148, [%rd46+4];
ld.global.nc.u32 %r152, [%rd46+20];
ld.global.nc.u32 %r145, [%rd45+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r145;
cvt.f32.f16 %f34, low;}


ld.global.nc.u32 %r164, [%rd46+40];
ld.global.nc.u32 %r168, [%rd46+56];

dp4a.s32.s32 %r146, %r147, %r148, %r209;


dp4a.s32.s32 %r150, %r151, %r152, %r146;


dp4a.s32.s32 %r154, %r211, %r148, %r209;


dp4a.s32.s32 %r158, %r211, %r152, %r154;

mul.lo.s32 %r272, %r150, %r233;
cvt.rn.f32.s32 %f63, %r272;
fma.rn.ftz.f32 %f64, %f33, %f63, 0f00000000;
mul.lo.s32 %r273, %r158, %r236;
cvt.rn.f32.s32 %f65, %r273;
fma.rn.ftz.f32 %f66, %f33, %f65, 0f00000000;

dp4a.s32.s32 %r162, %r163, %r164, %r209;


dp4a.s32.s32 %r166, %r167, %r168, %r162;


dp4a.s32.s32 %r170, %r211, %r164, %r209;


dp4a.s32.s32 %r174, %r211, %r168, %r170;

mul.lo.s32 %r274, %r166, %r240;
cvt.rn.f32.s32 %f67, %r274;
fma.rn.ftz.f32 %f68, %f34, %f67, %f64;
mul.lo.s32 %r275, %r174, %r242;
cvt.rn.f32.s32 %f69, %r275;
fma.rn.ftz.f32 %f70, %f34, %f69, %f66;

{.reg .f16 low,high;
mov.b32 {low,high},%r106;
cvt.f32.f16 %f35, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r106;
cvt.f32.f16 %f36, high;}


mul.ftz.f32 %f71, %f68, %f35;
mul.ftz.f32 %f72, %f70, %f36;
sub.ftz.f32 %f73, %f71, %f72;
add.ftz.f32 %f150, %f73, %f150;

{.reg .f16 low,high;
mov.b32 {low,high},%r144;
cvt.f32.f16 %f37, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r145;
cvt.f32.f16 %f38, low;}



dp4a.s32.s32 %r182, %r183, %r148, %r209;


dp4a.s32.s32 %r186, %r187, %r152, %r182;


dp4a.s32.s32 %r190, %r211, %r148, %r209;


dp4a.s32.s32 %r194, %r211, %r152, %r190;

mul.lo.s32 %r276, %r186, %r261;
cvt.rn.f32.s32 %f74, %r276;
fma.rn.ftz.f32 %f75, %f37, %f74, 0f00000000;
mul.lo.s32 %r277, %r194, %r264;
cvt.rn.f32.s32 %f76, %r277;
fma.rn.ftz.f32 %f77, %f37, %f76, 0f00000000;

dp4a.s32.s32 %r198, %r199, %r164, %r209;


dp4a.s32.s32 %r202, %r203, %r168, %r198;


dp4a.s32.s32 %r206, %r211, %r164, %r209;


dp4a.s32.s32 %r210, %r211, %r168, %r206;

mul.lo.s32 %r278, %r202, %r268;
cvt.rn.f32.s32 %f78, %r278;
fma.rn.ftz.f32 %f79, %f38, %f78, %f75;
mul.lo.s32 %r279, %r210, %r270;
cvt.rn.f32.s32 %f80, %r279;
fma.rn.ftz.f32 %f81, %f38, %f80, %f77;

{.reg .f16 low,high;
mov.b32 {low,high},%r142;
cvt.f32.f16 %f39, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r142;
cvt.f32.f16 %f40, high;}


mul.ftz.f32 %f82, %f79, %f39;
mul.ftz.f32 %f83, %f81, %f40;
sub.ftz.f32 %f84, %f82, %f83;
add.ftz.f32 %f149, %f84, %f149;
add.s32 %r361, %r361, 64;
add.s32 %r360, %r360, 64;
add.s64 %rd56, %rd56, 1152;
add.s32 %r362, %r362, 8;
setp.lt.s32 %p5, %r362, %r7;
@%p5 bra $L__BB161_2;

$L__BB161_3:
setp.eq.s32 %p6, %r1, 0;
@%p6 bra $L__BB161_5;

shl.b32 %r280, %r1, 9;
mov.u32 %r281, _ZZ13mul_mat_vec_qIL9ggml_type12ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r282, %r281, %r280;
shl.b32 %r283, %r2, 2;
add.s32 %r284, %r282, %r283;
st.shared.f32 [%r284+-512], %f152;
st.shared.f32 [%r284+-384], %f151;
st.shared.f32 [%r284+-256], %f150;
st.shared.f32 [%r284+-128], %f149;

$L__BB161_5:
setp.ne.s32 %p7, %r1, 0;
bar.sync 0;
@%p7 bra $L__BB161_14;

mov.u32 %r285, %ctaid.z;
mad.lo.s32 %r286, %r5, %r27, %r4;
mad.lo.s32 %r287, %r285, %r33, %r286;
cvt.u64.u32 %rd14, %r287;
shl.b32 %r288, %r2, 2;
mov.u32 %r289, _ZZ13mul_mat_vec_qIL9ggml_type12ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r17, %r289, %r288;
mov.u32 %r290, 2;
ld.shared.f32 %f85, [%r17];
add.ftz.f32 %f86, %f85, %f152;
ld.shared.f32 %f87, [%r17+512];
add.ftz.f32 %f88, %f87, %f86;
ld.shared.f32 %f89, [%r17+1024];
add.ftz.f32 %f90, %f89, %f88;
mov.b32 %r291, %f90;
mov.u32 %r292, 31;
mov.u32 %r293, 16;
mov.u32 %r294, -1;
shfl.sync.bfly.b32 %r295|%p8, %r291, %r293, %r292, %r294;
mov.b32 %f91, %r295;
add.ftz.f32 %f92, %f90, %f91;
mov.b32 %r296, %f92;
mov.u32 %r297, 8;
shfl.sync.bfly.b32 %r298|%p9, %r296, %r297, %r292, %r294;
mov.b32 %f93, %r298;
add.ftz.f32 %f94, %f92, %f93;
mov.b32 %r299, %f94;
mov.u32 %r300, 4;
shfl.sync.bfly.b32 %r301|%p10, %r299, %r300, %r292, %r294;
mov.b32 %f95, %r301;
add.ftz.f32 %f96, %f94, %f95;
mov.b32 %r302, %f96;
shfl.sync.bfly.b32 %r303|%p11, %r302, %r290, %r292, %r294;
mov.b32 %f97, %r303;
add.ftz.f32 %f98, %f96, %f97;
mov.b32 %r304, %f98;
mov.u32 %r305, 1;
shfl.sync.bfly.b32 %r306|%p12, %r304, %r305, %r292, %r294;
mov.b32 %f99, %r306;
add.ftz.f32 %f13, %f98, %f99;
setp.lt.u32 %p13, %r4, %r21;
setp.eq.s32 %p14, %r2, 0;
and.pred %p1, %p14, %p13;
mul.wide.u32 %rd47, %r287, 4;
add.s64 %rd15, %rd1, %rd47;
not.pred %p15, %p1;
@%p15 bra $L__BB161_8;

st.global.f32 [%rd15], %f13;

$L__BB161_8:
ld.shared.f32 %f100, [%r17+128];
add.ftz.f32 %f101, %f100, %f151;
ld.shared.f32 %f102, [%r17+640];
add.ftz.f32 %f103, %f102, %f101;
ld.shared.f32 %f104, [%r17+1152];
add.ftz.f32 %f105, %f104, %f103;
mov.b32 %r307, %f105;
shfl.sync.bfly.b32 %r311|%p16, %r307, %r293, %r292, %r294;
mov.b32 %f106, %r311;
add.ftz.f32 %f107, %f105, %f106;
mov.b32 %r312, %f107;
shfl.sync.bfly.b32 %r314|%p17, %r312, %r297, %r292, %r294;
mov.b32 %f108, %r314;
add.ftz.f32 %f109, %f107, %f108;
mov.b32 %r315, %f109;
shfl.sync.bfly.b32 %r317|%p18, %r315, %r300, %r292, %r294;
mov.b32 %f110, %r317;
add.ftz.f32 %f111, %f109, %f110;
mov.b32 %r318, %f111;
shfl.sync.bfly.b32 %r320|%p19, %r318, %r290, %r292, %r294;
mov.b32 %f112, %r320;
add.ftz.f32 %f113, %f111, %f112;
mov.b32 %r321, %f113;
shfl.sync.bfly.b32 %r323|%p20, %r321, %r305, %r292, %r294;
mov.b32 %f114, %r323;
add.ftz.f32 %f14, %f113, %f114;
add.s32 %r324, %r4, 1;
setp.lt.u32 %p21, %r324, %r21;
setp.eq.s32 %p22, %r2, 1;
and.pred %p2, %p22, %p21;
not.pred %p23, %p2;
@%p23 bra $L__BB161_10;

st.global.f32 [%rd15+4], %f14;

$L__BB161_10:
ld.shared.f32 %f115, [%r17+256];
add.ftz.f32 %f116, %f115, %f150;
ld.shared.f32 %f117, [%r17+768];
add.ftz.f32 %f118, %f117, %f116;
ld.shared.f32 %f119, [%r17+1280];
add.ftz.f32 %f120, %f119, %f118;
mov.b32 %r325, %f120;
mov.u32 %r326, 31;
mov.u32 %r327, 16;
mov.u32 %r328, -1;
shfl.sync.bfly.b32 %r329|%p24, %r325, %r327, %r326, %r328;
mov.b32 %f121, %r329;
add.ftz.f32 %f122, %f120, %f121;
mov.b32 %r330, %f122;
mov.u32 %r331, 8;
shfl.sync.bfly.b32 %r332|%p25, %r330, %r331, %r326, %r328;
mov.b32 %f123, %r332;
add.ftz.f32 %f124, %f122, %f123;
mov.b32 %r333, %f124;
mov.u32 %r334, 4;
shfl.sync.bfly.b32 %r335|%p26, %r333, %r334, %r326, %r328;
mov.b32 %f125, %r335;
add.ftz.f32 %f126, %f124, %f125;
mov.b32 %r336, %f126;
mov.u32 %r337, 2;
shfl.sync.bfly.b32 %r338|%p27, %r336, %r337, %r326, %r328;
mov.b32 %f127, %r338;
add.ftz.f32 %f128, %f126, %f127;
mov.b32 %r339, %f128;
mov.u32 %r340, 1;
shfl.sync.bfly.b32 %r341|%p28, %r339, %r340, %r326, %r328;
mov.b32 %f129, %r341;
add.ftz.f32 %f15, %f128, %f129;
@%p15 bra $L__BB161_12;

cvt.u64.u32 %rd48, %r21;
add.s64 %rd49, %rd48, %rd14;
shl.b64 %rd50, %rd49, 2;
add.s64 %rd51, %rd1, %rd50;
st.global.f32 [%rd51], %f15;

$L__BB161_12:
ld.shared.f32 %f130, [%r17+384];
add.ftz.f32 %f131, %f130, %f149;
ld.shared.f32 %f132, [%r17+896];
add.ftz.f32 %f133, %f132, %f131;
ld.shared.f32 %f134, [%r17+1408];
add.ftz.f32 %f135, %f134, %f133;
mov.b32 %r342, %f135;
shfl.sync.bfly.b32 %r346|%p30, %r342, %r327, %r326, %r328;
mov.b32 %f136, %r346;
add.ftz.f32 %f137, %f135, %f136;
mov.b32 %r347, %f137;
shfl.sync.bfly.b32 %r349|%p31, %r347, %r331, %r326, %r328;
mov.b32 %f138, %r349;
add.ftz.f32 %f139, %f137, %f138;
mov.b32 %r350, %f139;
shfl.sync.bfly.b32 %r352|%p32, %r350, %r334, %r326, %r328;
mov.b32 %f140, %r352;
add.ftz.f32 %f141, %f139, %f140;
mov.b32 %r353, %f141;
shfl.sync.bfly.b32 %r355|%p33, %r353, %r337, %r326, %r328;
mov.b32 %f142, %r355;
add.ftz.f32 %f143, %f141, %f142;
mov.b32 %r356, %f143;
shfl.sync.bfly.b32 %r358|%p34, %r356, %r340, %r326, %r328;
mov.b32 %f144, %r358;
add.ftz.f32 %f16, %f143, %f144;
@%p23 bra $L__BB161_14;

add.s32 %r359, %r21, 1;
cvt.u64.u32 %rd52, %r359;
add.s64 %rd53, %rd52, %rd14;
shl.b64 %rd54, %rd53, 2;
add.s64 %rd55, %rd1, %rd54;
st.global.f32 [%rd55], %f16;

$L__BB161_14:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type12ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type12ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type12ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type12ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type12ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type12ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type12ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type12ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type12ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<48>;
.reg .b16 %rs<15>;
.reg .f32 %f<229>;
.reg .b32 %r<482>;
.reg .b64 %rd<66>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type12ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[2304];

ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type12ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type12ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type12ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type12ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type12ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type12ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type12ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type12ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type12ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type12ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type12ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type12ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type12ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type12ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type12ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type12ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type12ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
cvta.to.global.u64 %rd1, %rd19;
mov.u32 %r1, %tid.y;
shl.b32 %r38, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r38, %r2;
mov.u32 %r39, %ctaid.x;
shl.b32 %r4, %r39, 1;
mov.u32 %r5, %ctaid.y;
shr.s32 %r40, %r3, 31;
shr.u32 %r41, %r40, 28;
add.s32 %r42, %r3, %r41;
shr.s32 %r481, %r42, 4;
shr.u32 %r7, %r37, 8;
setp.ge.s32 %p3, %r481, %r7;
mov.f32 %f223, 0f00000000;
mov.f32 %f224, %f223;
mov.f32 %f225, %f223;
mov.f32 %f226, %f223;
mov.f32 %f227, %f223;
mov.f32 %f228, %f223;
@%p3 bra $L__BB162_3;

mov.u32 %r43, %ctaid.z;
mul.hi.u32 %r44, %r43, %r31;
add.s32 %r45, %r44, %r43;
shr.u32 %r46, %r45, %r32;
mul.lo.s32 %r47, %r46, %r34;
mul.hi.u32 %r48, %r5, %r25;
add.s32 %r49, %r48, %r5;
shr.u32 %r50, %r49, %r26;
mul.lo.s32 %r51, %r50, %r28;
mul.lo.s32 %r52, %r5, %r29;
cvt.u64.u32 %rd20, %r52;
cvta.to.global.u64 %rd65, %rd17;
and.b32 %r56, %r42, -16;
sub.s32 %r8, %r3, %r56;
shr.s32 %r57, %r8, 31;
shr.u32 %r58, %r57, 30;
add.s32 %r59, %r8, %r58;
shr.s32 %r60, %r59, 2;
shl.b32 %r61, %r60, 1;
shl.b32 %r62, %r60, 5;
and.b32 %r63, %r59, -4;
sub.s32 %r64, %r8, %r63;
shl.b32 %r65, %r64, 2;
add.s32 %r66, %r62, %r65;
cvt.s64.s32 %rd21, %r66;
cvt.s64.s32 %rd22, %r61;
add.s64 %rd23, %rd20, %rd22;
mul.lo.s32 %r67, %r43, %r35;
cvt.u64.u32 %rd24, %r67;
add.s64 %rd3, %rd23, %rd24;
cvt.s64.s32 %rd4, %r64;
shl.b32 %r68, %r23, 1;
shl.b32 %r478, %r481, 3;
add.s32 %r480, %r478, %r68;
add.s32 %r479, %r23, %r478;
add.s32 %r69, %r481, %r47;
add.s32 %r70, %r69, %r51;
add.s32 %r71, %r4, 1;
mad.lo.s32 %r72, %r22, %r71, %r70;
mul.wide.s32 %rd5, %r72, 144;
mad.lo.s32 %r73, %r4, %r22, %r481;
add.s32 %r74, %r73, %r47;
add.s32 %r75, %r74, %r51;
mul.wide.s32 %rd6, %r75, 144;
add.s64 %rd25, %rd6, %rd21;
add.s64 %rd7, %rd25, 32;
cvt.u64.u32 %rd26, %r61;
and.b64 %rd27, %rd26, 2;
or.b64 %rd28, %rd27, %rd5;
or.b64 %rd8, %rd28, 8;
or.b64 %rd29, %rd27, %rd6;
or.b64 %rd9, %rd29, 12;
add.s64 %rd30, %rd5, %rd21;
add.s64 %rd10, %rd30, 16;
cvta.to.global.u64 %rd11, %rd18;
mov.f32 %f223, 0f00000000;
shl.b64 %rd37, %rd4, 2;

$L__BB162_2:
cvt.u64.u32 %rd31, %r478;
add.s64 %rd32, %rd3, %rd31;
add.s64 %rd33, %rd65, %rd7;
ld.global.nc.u32 %r292, [%rd33+-16];
ld.global.nc.u32 %r293, [%rd33];
add.s64 %rd34, %rd65, %rd9;
ld.global.nc.u16 %rs1, [%rd34+-8];
cvt.u32.u16 %r294, %rs1;
ld.global.nc.u16 %rs2, [%rd34+-4];
cvt.u32.u16 %r295, %rs2;
ld.global.nc.u16 %rs3, [%rd34];
cvt.u32.u16 %r296, %rs3;
and.b32 %r297, %r294, 16191;
and.b32 %r298, %r296, 3855;
and.b32 %r299, %r294, 49344;
shr.u32 %r300, %r299, 2;
or.b32 %r301, %r298, %r300;
setp.gt.s32 %p4, %r8, 7;
and.b32 %r302, %r295, 16191;
shr.u32 %r303, %r296, 4;
and.b32 %r304, %r303, 3855;
and.b32 %r305, %r295, 49344;
shr.u32 %r306, %r305, 2;
or.b32 %r307, %r304, %r306;
mul.lo.s64 %rd35, %rd32, 36;
add.s64 %rd36, %rd11, %rd35;
ld.global.nc.u32 %r76, [%rd36];

{.reg .f16 low,high;
mov.b32 {low,high},%r76;
cvt.f32.f16 %f37, low;}


add.s64 %rd38, %rd36, %rd37;
ld.global.nc.u32 %r80, [%rd38+4];
ld.global.nc.u32 %r84, [%rd38+20];
ld.global.nc.u32 %r77, [%rd36+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r77;
cvt.f32.f16 %f38, low;}


ld.global.nc.u32 %r96, [%rd38+40];
ld.global.nc.u32 %r100, [%rd38+56];
add.s64 %rd39, %rd65, %rd6;
and.b32 %r223, %r292, 252645135;
and.b32 %r227, %r293, 252645135;
mov.u32 %r285, 0;

dp4a.s32.s32 %r78, %r223, %r80, %r285;


dp4a.s32.s32 %r82, %r227, %r84, %r78;

mov.u32 %r287, 16843009;

dp4a.s32.s32 %r86, %r287, %r80, %r285;


dp4a.s32.s32 %r90, %r287, %r84, %r86;

selp.b32 %r308, %r301, %r297, %p4;
cvt.u16.u32 %rs4, %r308;
and.b32 %r309, %r308, 63;
mul.lo.s32 %r310, %r82, %r309;
cvt.rn.f32.s32 %f61, %r310;
fma.rn.ftz.f32 %f62, %f37, %f61, 0f00000000;
selp.b32 %r311, %r307, %r302, %p4;
cvt.u16.u32 %rs5, %r311;
and.b32 %r312, %r311, 63;
mul.lo.s32 %r313, %r90, %r312;
cvt.rn.f32.s32 %f63, %r313;
fma.rn.ftz.f32 %f64, %f37, %f63, 0f00000000;
shr.u32 %r314, %r292, 4;
and.b32 %r239, %r314, 252645135;
shr.u32 %r315, %r293, 4;
and.b32 %r243, %r315, 252645135;

dp4a.s32.s32 %r94, %r239, %r96, %r285;


dp4a.s32.s32 %r98, %r243, %r100, %r94;


dp4a.s32.s32 %r102, %r287, %r96, %r285;


dp4a.s32.s32 %r106, %r287, %r100, %r102;

shr.u16 %rs6, %rs4, 8;
cvt.u32.u16 %r316, %rs6;
mul.lo.s32 %r317, %r98, %r316;
cvt.rn.f32.s32 %f65, %r317;
fma.rn.ftz.f32 %f66, %f38, %f65, %f62;
shr.u16 %rs7, %rs5, 8;
cvt.u32.u16 %r318, %rs7;
mul.lo.s32 %r319, %r106, %r318;
cvt.rn.f32.s32 %f67, %r319;
fma.rn.ftz.f32 %f68, %f38, %f67, %f64;
ld.global.nc.u32 %r110, [%rd39];

{.reg .f16 low,high;
mov.b32 {low,high},%r110;
cvt.f32.f16 %f39, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r110;
cvt.f32.f16 %f40, high;}


mul.ftz.f32 %f69, %f66, %f39;
mul.ftz.f32 %f70, %f68, %f40;
sub.ftz.f32 %f71, %f69, %f70;
add.ftz.f32 %f228, %f71, %f228;
add.s64 %rd40, %rd65, %rd10;
ld.global.nc.u32 %r320, [%rd40];
ld.global.nc.u32 %r321, [%rd40+16];
add.s64 %rd41, %rd65, %rd8;
ld.global.nc.u16 %rs8, [%rd41+-4];
cvt.u32.u16 %r322, %rs8;
ld.global.nc.u16 %rs9, [%rd41];
cvt.u32.u16 %r323, %rs9;
ld.global.nc.u16 %rs10, [%rd41+4];
cvt.u32.u16 %r324, %rs10;
and.b32 %r325, %r322, 16191;
and.b32 %r326, %r324, 3855;
and.b32 %r327, %r322, 49344;
shr.u32 %r328, %r327, 2;
or.b32 %r329, %r326, %r328;
and.b32 %r330, %r323, 16191;
shr.u32 %r331, %r324, 4;
and.b32 %r332, %r331, 3855;
and.b32 %r333, %r323, 49344;
shr.u32 %r334, %r333, 2;
or.b32 %r335, %r332, %r334;

{.reg .f16 low,high;
mov.b32 {low,high},%r76;
cvt.f32.f16 %f41, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r77;
cvt.f32.f16 %f42, low;}


add.s64 %rd42, %rd65, %rd5;
and.b32 %r259, %r320, 252645135;
and.b32 %r263, %r321, 252645135;

dp4a.s32.s32 %r114, %r259, %r80, %r285;


dp4a.s32.s32 %r118, %r263, %r84, %r114;


dp4a.s32.s32 %r122, %r287, %r80, %r285;


dp4a.s32.s32 %r126, %r287, %r84, %r122;

selp.b32 %r336, %r329, %r325, %p4;
cvt.u16.u32 %rs11, %r336;
and.b32 %r337, %r336, 63;
mul.lo.s32 %r338, %r118, %r337;
cvt.rn.f32.s32 %f72, %r338;
fma.rn.ftz.f32 %f73, %f41, %f72, 0f00000000;
selp.b32 %r339, %r335, %r330, %p4;
cvt.u16.u32 %rs12, %r339;
and.b32 %r340, %r339, 63;
mul.lo.s32 %r341, %r126, %r340;
cvt.rn.f32.s32 %f74, %r341;
fma.rn.ftz.f32 %f75, %f41, %f74, 0f00000000;
shr.u32 %r342, %r320, 4;
and.b32 %r275, %r342, 252645135;
shr.u32 %r343, %r321, 4;
and.b32 %r279, %r343, 252645135;

dp4a.s32.s32 %r130, %r275, %r96, %r285;


dp4a.s32.s32 %r134, %r279, %r100, %r130;


dp4a.s32.s32 %r138, %r287, %r96, %r285;


dp4a.s32.s32 %r142, %r287, %r100, %r138;

shr.u16 %rs13, %rs11, 8;
cvt.u32.u16 %r344, %rs13;
mul.lo.s32 %r345, %r134, %r344;
cvt.rn.f32.s32 %f76, %r345;
fma.rn.ftz.f32 %f77, %f42, %f76, %f73;
shr.u16 %rs14, %rs12, 8;
cvt.u32.u16 %r346, %rs14;
mul.lo.s32 %r347, %r142, %r346;
cvt.rn.f32.s32 %f78, %r347;
fma.rn.ftz.f32 %f79, %f42, %f78, %f75;
ld.global.nc.u32 %r146, [%rd42];

{.reg .f16 low,high;
mov.b32 {low,high},%r146;
cvt.f32.f16 %f43, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r146;
cvt.f32.f16 %f44, high;}


mul.ftz.f32 %f80, %f77, %f43;
mul.ftz.f32 %f81, %f79, %f44;
sub.ftz.f32 %f82, %f80, %f81;
add.ftz.f32 %f227, %f82, %f227;
cvt.u64.u32 %rd43, %r479;
add.s64 %rd44, %rd3, %rd43;
mul.lo.s64 %rd45, %rd44, 36;
add.s64 %rd46, %rd11, %rd45;
ld.global.nc.u32 %r148, [%rd46];

{.reg .f16 low,high;
mov.b32 {low,high},%r148;
cvt.f32.f16 %f45, low;}


add.s64 %rd47, %rd46, %rd37;
ld.global.nc.u32 %r152, [%rd47+4];
ld.global.nc.u32 %r156, [%rd47+20];
ld.global.nc.u32 %r149, [%rd46+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r149;
cvt.f32.f16 %f46, low;}


ld.global.nc.u32 %r168, [%rd47+40];
ld.global.nc.u32 %r172, [%rd47+56];

dp4a.s32.s32 %r150, %r223, %r152, %r285;


dp4a.s32.s32 %r154, %r227, %r156, %r150;


dp4a.s32.s32 %r158, %r287, %r152, %r285;


dp4a.s32.s32 %r162, %r287, %r156, %r158;

mul.lo.s32 %r348, %r154, %r309;
cvt.rn.f32.s32 %f83, %r348;
fma.rn.ftz.f32 %f84, %f45, %f83, 0f00000000;
mul.lo.s32 %r349, %r162, %r312;
cvt.rn.f32.s32 %f85, %r349;
fma.rn.ftz.f32 %f86, %f45, %f85, 0f00000000;

dp4a.s32.s32 %r166, %r239, %r168, %r285;


dp4a.s32.s32 %r170, %r243, %r172, %r166;


dp4a.s32.s32 %r174, %r287, %r168, %r285;


dp4a.s32.s32 %r178, %r287, %r172, %r174;

mul.lo.s32 %r350, %r170, %r316;
cvt.rn.f32.s32 %f87, %r350;
fma.rn.ftz.f32 %f88, %f46, %f87, %f84;
mul.lo.s32 %r351, %r178, %r318;
cvt.rn.f32.s32 %f89, %r351;
fma.rn.ftz.f32 %f90, %f46, %f89, %f86;

{.reg .f16 low,high;
mov.b32 {low,high},%r110;
cvt.f32.f16 %f47, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r110;
cvt.f32.f16 %f48, high;}


mul.ftz.f32 %f91, %f88, %f47;
mul.ftz.f32 %f92, %f90, %f48;
sub.ftz.f32 %f93, %f91, %f92;
add.ftz.f32 %f226, %f93, %f226;

{.reg .f16 low,high;
mov.b32 {low,high},%r148;
cvt.f32.f16 %f49, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r149;
cvt.f32.f16 %f50, low;}



dp4a.s32.s32 %r186, %r259, %r152, %r285;


dp4a.s32.s32 %r190, %r263, %r156, %r186;


dp4a.s32.s32 %r194, %r287, %r152, %r285;


dp4a.s32.s32 %r198, %r287, %r156, %r194;

mul.lo.s32 %r352, %r190, %r337;
cvt.rn.f32.s32 %f94, %r352;
fma.rn.ftz.f32 %f95, %f49, %f94, 0f00000000;
mul.lo.s32 %r353, %r198, %r340;
cvt.rn.f32.s32 %f96, %r353;
fma.rn.ftz.f32 %f97, %f49, %f96, 0f00000000;

dp4a.s32.s32 %r202, %r275, %r168, %r285;


dp4a.s32.s32 %r206, %r279, %r172, %r202;


dp4a.s32.s32 %r210, %r287, %r168, %r285;


dp4a.s32.s32 %r214, %r287, %r172, %r210;

mul.lo.s32 %r354, %r206, %r344;
cvt.rn.f32.s32 %f98, %r354;
fma.rn.ftz.f32 %f99, %f50, %f98, %f95;
mul.lo.s32 %r355, %r214, %r346;
cvt.rn.f32.s32 %f100, %r355;
fma.rn.ftz.f32 %f101, %f50, %f100, %f97;

{.reg .f16 low,high;
mov.b32 {low,high},%r146;
cvt.f32.f16 %f51, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r146;
cvt.f32.f16 %f52, high;}


mul.ftz.f32 %f102, %f99, %f51;
mul.ftz.f32 %f103, %f101, %f52;
sub.ftz.f32 %f104, %f102, %f103;
add.ftz.f32 %f225, %f104, %f225;
cvt.u64.u32 %rd48, %r480;
add.s64 %rd49, %rd3, %rd48;
mul.lo.s64 %rd50, %rd49, 36;
add.s64 %rd51, %rd11, %rd50;
ld.global.nc.u32 %r220, [%rd51];

{.reg .f16 low,high;
mov.b32 {low,high},%r220;
cvt.f32.f16 %f53, low;}


add.s64 %rd52, %rd51, %rd37;
ld.global.nc.u32 %r224, [%rd52+4];
ld.global.nc.u32 %r228, [%rd52+20];
ld.global.nc.u32 %r221, [%rd51+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r221;
cvt.f32.f16 %f54, low;}


ld.global.nc.u32 %r240, [%rd52+40];
ld.global.nc.u32 %r244, [%rd52+56];

dp4a.s32.s32 %r222, %r223, %r224, %r285;


dp4a.s32.s32 %r226, %r227, %r228, %r222;


dp4a.s32.s32 %r230, %r287, %r224, %r285;


dp4a.s32.s32 %r234, %r287, %r228, %r230;

mul.lo.s32 %r356, %r226, %r309;
cvt.rn.f32.s32 %f105, %r356;
fma.rn.ftz.f32 %f106, %f53, %f105, 0f00000000;
mul.lo.s32 %r357, %r234, %r312;
cvt.rn.f32.s32 %f107, %r357;
fma.rn.ftz.f32 %f108, %f53, %f107, 0f00000000;

dp4a.s32.s32 %r238, %r239, %r240, %r285;


dp4a.s32.s32 %r242, %r243, %r244, %r238;


dp4a.s32.s32 %r246, %r287, %r240, %r285;


dp4a.s32.s32 %r250, %r287, %r244, %r246;

mul.lo.s32 %r358, %r242, %r316;
cvt.rn.f32.s32 %f109, %r358;
fma.rn.ftz.f32 %f110, %f54, %f109, %f106;
mul.lo.s32 %r359, %r250, %r318;
cvt.rn.f32.s32 %f111, %r359;
fma.rn.ftz.f32 %f112, %f54, %f111, %f108;

{.reg .f16 low,high;
mov.b32 {low,high},%r110;
cvt.f32.f16 %f55, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r110;
cvt.f32.f16 %f56, high;}


mul.ftz.f32 %f113, %f110, %f55;
mul.ftz.f32 %f114, %f112, %f56;
sub.ftz.f32 %f115, %f113, %f114;
add.ftz.f32 %f224, %f115, %f224;

{.reg .f16 low,high;
mov.b32 {low,high},%r220;
cvt.f32.f16 %f57, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r221;
cvt.f32.f16 %f58, low;}



dp4a.s32.s32 %r258, %r259, %r224, %r285;


dp4a.s32.s32 %r262, %r263, %r228, %r258;


dp4a.s32.s32 %r266, %r287, %r224, %r285;


dp4a.s32.s32 %r270, %r287, %r228, %r266;

mul.lo.s32 %r360, %r262, %r337;
cvt.rn.f32.s32 %f116, %r360;
fma.rn.ftz.f32 %f117, %f57, %f116, 0f00000000;
mul.lo.s32 %r361, %r270, %r340;
cvt.rn.f32.s32 %f118, %r361;
fma.rn.ftz.f32 %f119, %f57, %f118, 0f00000000;

dp4a.s32.s32 %r274, %r275, %r240, %r285;


dp4a.s32.s32 %r278, %r279, %r244, %r274;


dp4a.s32.s32 %r282, %r287, %r240, %r285;


dp4a.s32.s32 %r286, %r287, %r244, %r282;

mul.lo.s32 %r362, %r278, %r344;
cvt.rn.f32.s32 %f120, %r362;
fma.rn.ftz.f32 %f121, %f58, %f120, %f117;
mul.lo.s32 %r363, %r286, %r346;
cvt.rn.f32.s32 %f122, %r363;
fma.rn.ftz.f32 %f123, %f58, %f122, %f119;

{.reg .f16 low,high;
mov.b32 {low,high},%r146;
cvt.f32.f16 %f59, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r146;
cvt.f32.f16 %f60, high;}


mul.ftz.f32 %f124, %f121, %f59;
mul.ftz.f32 %f125, %f123, %f60;
sub.ftz.f32 %f126, %f124, %f125;
add.ftz.f32 %f223, %f126, %f223;
add.s32 %r480, %r480, 64;
add.s32 %r479, %r479, 64;
add.s32 %r478, %r478, 64;
add.s64 %rd65, %rd65, 1152;
add.s32 %r481, %r481, 8;
setp.lt.s32 %p5, %r481, %r7;
@%p5 bra $L__BB162_2;

$L__BB162_3:
setp.eq.s32 %p6, %r1, 0;
@%p6 bra $L__BB162_5;

mov.u32 %r364, _ZZ13mul_mat_vec_qIL9ggml_type12ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r365, %r1, 768, %r364;
shl.b32 %r366, %r2, 2;
add.s32 %r367, %r365, %r366;
st.shared.f32 [%r367+-768], %f228;
st.shared.f32 [%r367+-640], %f227;
st.shared.f32 [%r367+-512], %f226;
st.shared.f32 [%r367+-384], %f225;
st.shared.f32 [%r367+-256], %f224;
st.shared.f32 [%r367+-128], %f223;

$L__BB162_5:
setp.ne.s32 %p7, %r1, 0;
bar.sync 0;
@%p7 bra $L__BB162_18;

mov.u32 %r368, %ctaid.z;
mad.lo.s32 %r369, %r5, %r30, %r4;
mad.lo.s32 %r370, %r368, %r36, %r369;
cvt.u64.u32 %rd14, %r370;
shl.b32 %r371, %r2, 2;
mov.u32 %r372, _ZZ13mul_mat_vec_qIL9ggml_type12ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r20, %r372, %r371;
mov.u32 %r373, 2;
ld.shared.f32 %f127, [%r20];
add.ftz.f32 %f128, %f127, %f228;
ld.shared.f32 %f129, [%r20+768];
add.ftz.f32 %f130, %f129, %f128;
ld.shared.f32 %f131, [%r20+1536];
add.ftz.f32 %f132, %f131, %f130;
mov.b32 %r374, %f132;
mov.u32 %r375, 31;
mov.u32 %r376, 16;
mov.u32 %r377, -1;
shfl.sync.bfly.b32 %r378|%p8, %r374, %r376, %r375, %r377;
mov.b32 %f133, %r378;
add.ftz.f32 %f134, %f132, %f133;
mov.b32 %r379, %f134;
mov.u32 %r380, 8;
shfl.sync.bfly.b32 %r381|%p9, %r379, %r380, %r375, %r377;
mov.b32 %f135, %r381;
add.ftz.f32 %f136, %f134, %f135;
mov.b32 %r382, %f136;
mov.u32 %r383, 4;
shfl.sync.bfly.b32 %r384|%p10, %r382, %r383, %r375, %r377;
mov.b32 %f137, %r384;
add.ftz.f32 %f138, %f136, %f137;
mov.b32 %r385, %f138;
shfl.sync.bfly.b32 %r386|%p11, %r385, %r373, %r375, %r377;
mov.b32 %f139, %r386;
add.ftz.f32 %f140, %f138, %f139;
mov.b32 %r387, %f140;
mov.u32 %r388, 1;
shfl.sync.bfly.b32 %r389|%p12, %r387, %r388, %r375, %r377;
mov.b32 %f141, %r389;
add.ftz.f32 %f19, %f140, %f141;
setp.lt.u32 %p13, %r4, %r24;
setp.eq.s32 %p14, %r2, 0;
and.pred %p1, %p14, %p13;
mul.wide.u32 %rd53, %r370, 4;
add.s64 %rd15, %rd1, %rd53;
not.pred %p15, %p1;
@%p15 bra $L__BB162_8;

st.global.f32 [%rd15], %f19;

$L__BB162_8:
ld.shared.f32 %f142, [%r20+128];
add.ftz.f32 %f143, %f142, %f227;
ld.shared.f32 %f144, [%r20+896];
add.ftz.f32 %f145, %f144, %f143;
ld.shared.f32 %f146, [%r20+1664];
add.ftz.f32 %f147, %f146, %f145;
mov.b32 %r390, %f147;
shfl.sync.bfly.b32 %r394|%p16, %r390, %r376, %r375, %r377;
mov.b32 %f148, %r394;
add.ftz.f32 %f149, %f147, %f148;
mov.b32 %r395, %f149;
shfl.sync.bfly.b32 %r397|%p17, %r395, %r380, %r375, %r377;
mov.b32 %f150, %r397;
add.ftz.f32 %f151, %f149, %f150;
mov.b32 %r398, %f151;
shfl.sync.bfly.b32 %r400|%p18, %r398, %r383, %r375, %r377;
mov.b32 %f152, %r400;
add.ftz.f32 %f153, %f151, %f152;
mov.b32 %r401, %f153;
shfl.sync.bfly.b32 %r403|%p19, %r401, %r373, %r375, %r377;
mov.b32 %f154, %r403;
add.ftz.f32 %f155, %f153, %f154;
mov.b32 %r404, %f155;
shfl.sync.bfly.b32 %r406|%p20, %r404, %r388, %r375, %r377;
mov.b32 %f156, %r406;
add.ftz.f32 %f20, %f155, %f156;
add.s32 %r407, %r4, 1;
setp.lt.u32 %p21, %r407, %r24;
setp.eq.s32 %p22, %r2, 1;
and.pred %p2, %p22, %p21;
not.pred %p23, %p2;
@%p23 bra $L__BB162_10;

st.global.f32 [%rd15+4], %f20;

$L__BB162_10:
ld.shared.f32 %f157, [%r20+256];
add.ftz.f32 %f158, %f157, %f226;
ld.shared.f32 %f159, [%r20+1024];
add.ftz.f32 %f160, %f159, %f158;
ld.shared.f32 %f161, [%r20+1792];
add.ftz.f32 %f162, %f161, %f160;
mov.b32 %r408, %f162;
mov.u32 %r409, 31;
mov.u32 %r410, 16;
mov.u32 %r411, -1;
shfl.sync.bfly.b32 %r412|%p24, %r408, %r410, %r409, %r411;
mov.b32 %f163, %r412;
add.ftz.f32 %f164, %f162, %f163;
mov.b32 %r413, %f164;
mov.u32 %r414, 8;
shfl.sync.bfly.b32 %r415|%p25, %r413, %r414, %r409, %r411;
mov.b32 %f165, %r415;
add.ftz.f32 %f166, %f164, %f165;
mov.b32 %r416, %f166;
mov.u32 %r417, 4;
shfl.sync.bfly.b32 %r418|%p26, %r416, %r417, %r409, %r411;
mov.b32 %f167, %r418;
add.ftz.f32 %f168, %f166, %f167;
mov.b32 %r419, %f168;
mov.u32 %r420, 2;
shfl.sync.bfly.b32 %r421|%p27, %r419, %r420, %r409, %r411;
mov.b32 %f169, %r421;
add.ftz.f32 %f170, %f168, %f169;
mov.b32 %r422, %f170;
mov.u32 %r423, 1;
shfl.sync.bfly.b32 %r424|%p28, %r422, %r423, %r409, %r411;
mov.b32 %f171, %r424;
add.ftz.f32 %f21, %f170, %f171;
@%p15 bra $L__BB162_12;

cvt.u64.u32 %rd54, %r24;
add.s64 %rd55, %rd54, %rd14;
shl.b64 %rd56, %rd55, 2;
add.s64 %rd57, %rd1, %rd56;
st.global.f32 [%rd57], %f21;

$L__BB162_12:
ld.shared.f32 %f172, [%r20+384];
add.ftz.f32 %f173, %f172, %f225;
ld.shared.f32 %f174, [%r20+1152];
add.ftz.f32 %f175, %f174, %f173;
ld.shared.f32 %f176, [%r20+1920];
add.ftz.f32 %f177, %f176, %f175;
mov.b32 %r425, %f177;
shfl.sync.bfly.b32 %r429|%p30, %r425, %r410, %r409, %r411;
mov.b32 %f178, %r429;
add.ftz.f32 %f179, %f177, %f178;
mov.b32 %r430, %f179;
shfl.sync.bfly.b32 %r432|%p31, %r430, %r414, %r409, %r411;
mov.b32 %f180, %r432;
add.ftz.f32 %f181, %f179, %f180;
mov.b32 %r433, %f181;
shfl.sync.bfly.b32 %r435|%p32, %r433, %r417, %r409, %r411;
mov.b32 %f182, %r435;
add.ftz.f32 %f183, %f181, %f182;
mov.b32 %r436, %f183;
shfl.sync.bfly.b32 %r438|%p33, %r436, %r420, %r409, %r411;
mov.b32 %f184, %r438;
add.ftz.f32 %f185, %f183, %f184;
mov.b32 %r439, %f185;
shfl.sync.bfly.b32 %r441|%p34, %r439, %r423, %r409, %r411;
mov.b32 %f186, %r441;
add.ftz.f32 %f22, %f185, %f186;
@%p23 bra $L__BB162_14;

add.s32 %r442, %r24, 1;
cvt.u64.u32 %rd58, %r442;
add.s64 %rd59, %rd58, %rd14;
shl.b64 %rd60, %rd59, 2;
add.s64 %rd61, %rd1, %rd60;
st.global.f32 [%rd61], %f22;

$L__BB162_14:
shl.b32 %r443, %r24, 1;
mov.u32 %r444, 1;
ld.shared.f32 %f187, [%r20+512];
add.ftz.f32 %f188, %f187, %f224;
ld.shared.f32 %f189, [%r20+1280];
add.ftz.f32 %f190, %f189, %f188;
ld.shared.f32 %f191, [%r20+2048];
add.ftz.f32 %f192, %f191, %f190;
mov.b32 %r445, %f192;
mov.u32 %r446, 31;
mov.u32 %r447, 16;
mov.u32 %r448, -1;
shfl.sync.bfly.b32 %r449|%p36, %r445, %r447, %r446, %r448;
mov.b32 %f193, %r449;
add.ftz.f32 %f194, %f192, %f193;
mov.b32 %r450, %f194;
mov.u32 %r451, 8;
shfl.sync.bfly.b32 %r452|%p37, %r450, %r451, %r446, %r448;
mov.b32 %f195, %r452;
add.ftz.f32 %f196, %f194, %f195;
mov.b32 %r453, %f196;
mov.u32 %r454, 4;
shfl.sync.bfly.b32 %r455|%p38, %r453, %r454, %r446, %r448;
mov.b32 %f197, %r455;
add.ftz.f32 %f198, %f196, %f197;
mov.b32 %r456, %f198;
mov.u32 %r457, 2;
shfl.sync.bfly.b32 %r458|%p39, %r456, %r457, %r446, %r448;
mov.b32 %f199, %r458;
add.ftz.f32 %f200, %f198, %f199;
mov.b32 %r459, %f200;
shfl.sync.bfly.b32 %r460|%p40, %r459, %r444, %r446, %r448;
mov.b32 %f201, %r460;
add.ftz.f32 %f23, %f200, %f201;
cvt.u64.u32 %rd62, %r443;
add.s64 %rd63, %rd62, %rd14;
shl.b64 %rd64, %rd63, 2;
add.s64 %rd16, %rd1, %rd64;
@%p15 bra $L__BB162_16;

st.global.f32 [%rd16], %f23;

$L__BB162_16:
ld.shared.f32 %f202, [%r20+640];
add.ftz.f32 %f203, %f202, %f223;
ld.shared.f32 %f204, [%r20+1408];
add.ftz.f32 %f205, %f204, %f203;
ld.shared.f32 %f206, [%r20+2176];
add.ftz.f32 %f207, %f206, %f205;
mov.b32 %r461, %f207;
shfl.sync.bfly.b32 %r465|%p42, %r461, %r447, %r446, %r448;
mov.b32 %f208, %r465;
add.ftz.f32 %f209, %f207, %f208;
mov.b32 %r466, %f209;
shfl.sync.bfly.b32 %r468|%p43, %r466, %r451, %r446, %r448;
mov.b32 %f210, %r468;
add.ftz.f32 %f211, %f209, %f210;
mov.b32 %r469, %f211;
shfl.sync.bfly.b32 %r471|%p44, %r469, %r454, %r446, %r448;
mov.b32 %f212, %r471;
add.ftz.f32 %f213, %f211, %f212;
mov.b32 %r472, %f213;
shfl.sync.bfly.b32 %r474|%p45, %r472, %r457, %r446, %r448;
mov.b32 %f214, %r474;
add.ftz.f32 %f215, %f213, %f214;
mov.b32 %r475, %f215;
shfl.sync.bfly.b32 %r477|%p46, %r475, %r444, %r446, %r448;
mov.b32 %f216, %r477;
add.ftz.f32 %f24, %f215, %f216;
@%p23 bra $L__BB162_18;

st.global.f32 [%rd16+4], %f24;

$L__BB162_18:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type12ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type12ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type12ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type12ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type12ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type12ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type12ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type12ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type12ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<60>;
.reg .b16 %rs<15>;
.reg .f32 %f<305>;
.reg .b32 %r<603>;
.reg .b64 %rd<79>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type12ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[3072];

ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type12ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type12ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type12ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type12ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type12ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type12ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type12ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type12ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type12ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type12ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type12ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type12ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type12ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type12ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type12ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type12ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type12ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
cvta.to.global.u64 %rd1, %rd19;
mov.u32 %r1, %tid.y;
shl.b32 %r42, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r42, %r2;
mov.u32 %r43, %ctaid.x;
shl.b32 %r4, %r43, 1;
mov.u32 %r5, %ctaid.y;
shr.s32 %r44, %r3, 31;
shr.u32 %r45, %r44, 28;
add.s32 %r46, %r3, %r45;
shr.s32 %r602, %r46, 4;
shr.u32 %r7, %r41, 8;
setp.ge.s32 %p3, %r602, %r7;
mov.f32 %f297, 0f00000000;
mov.f32 %f298, %f297;
mov.f32 %f299, %f297;
mov.f32 %f300, %f297;
mov.f32 %f301, %f297;
mov.f32 %f302, %f297;
mov.f32 %f303, %f297;
mov.f32 %f304, %f297;
@%p3 bra $L__BB163_3;

mov.u32 %r47, %ctaid.z;
mul.hi.u32 %r48, %r47, %r35;
add.s32 %r49, %r48, %r47;
shr.u32 %r50, %r49, %r36;
mul.lo.s32 %r51, %r50, %r38;
mul.hi.u32 %r52, %r5, %r29;
add.s32 %r53, %r52, %r5;
shr.u32 %r54, %r53, %r30;
mul.lo.s32 %r55, %r54, %r32;
mul.lo.s32 %r56, %r5, %r33;
cvt.u64.u32 %rd20, %r56;
cvta.to.global.u64 %rd78, %rd17;
and.b32 %r60, %r46, -16;
sub.s32 %r8, %r3, %r60;
shr.s32 %r61, %r8, 31;
shr.u32 %r62, %r61, 30;
add.s32 %r63, %r8, %r62;
shr.s32 %r64, %r63, 2;
shl.b32 %r65, %r64, 1;
shl.b32 %r66, %r64, 5;
and.b32 %r67, %r63, -4;
sub.s32 %r68, %r8, %r67;
shl.b32 %r69, %r68, 2;
add.s32 %r70, %r66, %r69;
cvt.s64.s32 %rd21, %r70;
cvt.s64.s32 %rd22, %r65;
add.s64 %rd23, %rd20, %rd22;
mul.lo.s32 %r71, %r47, %r39;
cvt.u64.u32 %rd24, %r71;
add.s64 %rd3, %rd23, %rd24;
cvt.s64.s32 %rd4, %r68;
shl.b32 %r72, %r27, 1;
shl.b32 %r598, %r602, 3;
add.s32 %r601, %r598, %r72;
mad.lo.s32 %r600, %r27, 3, %r598;
add.s32 %r599, %r27, %r598;
add.s32 %r73, %r602, %r51;
add.s32 %r74, %r73, %r55;
add.s32 %r75, %r4, 1;
mad.lo.s32 %r76, %r26, %r75, %r74;
mul.wide.s32 %rd5, %r76, 144;
mad.lo.s32 %r77, %r4, %r26, %r602;
add.s32 %r78, %r77, %r51;
add.s32 %r79, %r78, %r55;
mul.wide.s32 %rd6, %r79, 144;
add.s64 %rd25, %rd6, %rd21;
add.s64 %rd7, %rd25, 32;
cvt.u64.u32 %rd26, %r65;
and.b64 %rd27, %rd26, 2;
or.b64 %rd28, %rd27, %rd5;
or.b64 %rd8, %rd28, 8;
or.b64 %rd29, %rd27, %rd6;
or.b64 %rd9, %rd29, 12;
add.s64 %rd30, %rd5, %rd21;
add.s64 %rd10, %rd30, 16;
cvta.to.global.u64 %rd11, %rd18;
mov.f32 %f297, 0f00000000;
shl.b64 %rd37, %rd4, 2;

$L__BB163_2:
cvt.u64.u32 %rd31, %r598;
add.s64 %rd32, %rd3, %rd31;
add.s64 %rd33, %rd78, %rd7;
ld.global.nc.u32 %r368, [%rd33+-16];
ld.global.nc.u32 %r369, [%rd33];
add.s64 %rd34, %rd78, %rd9;
ld.global.nc.u16 %rs1, [%rd34+-8];
cvt.u32.u16 %r370, %rs1;
ld.global.nc.u16 %rs2, [%rd34+-4];
cvt.u32.u16 %r371, %rs2;
ld.global.nc.u16 %rs3, [%rd34];
cvt.u32.u16 %r372, %rs3;
and.b32 %r373, %r370, 16191;
and.b32 %r374, %r372, 3855;
and.b32 %r375, %r370, 49344;
shr.u32 %r376, %r375, 2;
or.b32 %r377, %r374, %r376;
setp.gt.s32 %p4, %r8, 7;
and.b32 %r378, %r371, 16191;
shr.u32 %r379, %r372, 4;
and.b32 %r380, %r379, 3855;
and.b32 %r381, %r371, 49344;
shr.u32 %r382, %r381, 2;
or.b32 %r383, %r380, %r382;
mul.lo.s64 %rd35, %rd32, 36;
add.s64 %rd36, %rd11, %rd35;
ld.global.nc.u32 %r80, [%rd36];

{.reg .f16 low,high;
mov.b32 {low,high},%r80;
cvt.f32.f16 %f49, low;}


add.s64 %rd38, %rd36, %rd37;
ld.global.nc.u32 %r84, [%rd38+4];
ld.global.nc.u32 %r88, [%rd38+20];
ld.global.nc.u32 %r81, [%rd36+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r81;
cvt.f32.f16 %f50, low;}


ld.global.nc.u32 %r100, [%rd38+40];
ld.global.nc.u32 %r104, [%rd38+56];
add.s64 %rd39, %rd78, %rd6;
and.b32 %r299, %r368, 252645135;
and.b32 %r303, %r369, 252645135;
mov.u32 %r361, 0;

dp4a.s32.s32 %r82, %r299, %r84, %r361;


dp4a.s32.s32 %r86, %r303, %r88, %r82;

mov.u32 %r363, 16843009;

dp4a.s32.s32 %r90, %r363, %r84, %r361;


dp4a.s32.s32 %r94, %r363, %r88, %r90;

selp.b32 %r384, %r377, %r373, %p4;
cvt.u16.u32 %rs4, %r384;
and.b32 %r385, %r384, 63;
mul.lo.s32 %r386, %r86, %r385;
cvt.rn.f32.s32 %f81, %r386;
fma.rn.ftz.f32 %f82, %f49, %f81, 0f00000000;
selp.b32 %r387, %r383, %r378, %p4;
cvt.u16.u32 %rs5, %r387;
and.b32 %r388, %r387, 63;
mul.lo.s32 %r389, %r94, %r388;
cvt.rn.f32.s32 %f83, %r389;
fma.rn.ftz.f32 %f84, %f49, %f83, 0f00000000;
shr.u32 %r390, %r368, 4;
and.b32 %r315, %r390, 252645135;
shr.u32 %r391, %r369, 4;
and.b32 %r319, %r391, 252645135;

dp4a.s32.s32 %r98, %r315, %r100, %r361;


dp4a.s32.s32 %r102, %r319, %r104, %r98;


dp4a.s32.s32 %r106, %r363, %r100, %r361;


dp4a.s32.s32 %r110, %r363, %r104, %r106;

shr.u16 %rs6, %rs4, 8;
cvt.u32.u16 %r392, %rs6;
mul.lo.s32 %r393, %r102, %r392;
cvt.rn.f32.s32 %f85, %r393;
fma.rn.ftz.f32 %f86, %f50, %f85, %f82;
shr.u16 %rs7, %rs5, 8;
cvt.u32.u16 %r394, %rs7;
mul.lo.s32 %r395, %r110, %r394;
cvt.rn.f32.s32 %f87, %r395;
fma.rn.ftz.f32 %f88, %f50, %f87, %f84;
ld.global.nc.u32 %r114, [%rd39];

{.reg .f16 low,high;
mov.b32 {low,high},%r114;
cvt.f32.f16 %f51, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r114;
cvt.f32.f16 %f52, high;}


mul.ftz.f32 %f89, %f86, %f51;
mul.ftz.f32 %f90, %f88, %f52;
sub.ftz.f32 %f91, %f89, %f90;
add.ftz.f32 %f304, %f91, %f304;
add.s64 %rd40, %rd78, %rd10;
ld.global.nc.u32 %r396, [%rd40];
ld.global.nc.u32 %r397, [%rd40+16];
add.s64 %rd41, %rd78, %rd8;
ld.global.nc.u16 %rs8, [%rd41+-4];
cvt.u32.u16 %r398, %rs8;
ld.global.nc.u16 %rs9, [%rd41];
cvt.u32.u16 %r399, %rs9;
ld.global.nc.u16 %rs10, [%rd41+4];
cvt.u32.u16 %r400, %rs10;
and.b32 %r401, %r398, 16191;
and.b32 %r402, %r400, 3855;
and.b32 %r403, %r398, 49344;
shr.u32 %r404, %r403, 2;
or.b32 %r405, %r402, %r404;
and.b32 %r406, %r399, 16191;
shr.u32 %r407, %r400, 4;
and.b32 %r408, %r407, 3855;
and.b32 %r409, %r399, 49344;
shr.u32 %r410, %r409, 2;
or.b32 %r411, %r408, %r410;

{.reg .f16 low,high;
mov.b32 {low,high},%r80;
cvt.f32.f16 %f53, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r81;
cvt.f32.f16 %f54, low;}


add.s64 %rd42, %rd78, %rd5;
and.b32 %r335, %r396, 252645135;
and.b32 %r339, %r397, 252645135;

dp4a.s32.s32 %r118, %r335, %r84, %r361;


dp4a.s32.s32 %r122, %r339, %r88, %r118;


dp4a.s32.s32 %r126, %r363, %r84, %r361;


dp4a.s32.s32 %r130, %r363, %r88, %r126;

selp.b32 %r412, %r405, %r401, %p4;
cvt.u16.u32 %rs11, %r412;
and.b32 %r413, %r412, 63;
mul.lo.s32 %r414, %r122, %r413;
cvt.rn.f32.s32 %f92, %r414;
fma.rn.ftz.f32 %f93, %f53, %f92, 0f00000000;
selp.b32 %r415, %r411, %r406, %p4;
cvt.u16.u32 %rs12, %r415;
and.b32 %r416, %r415, 63;
mul.lo.s32 %r417, %r130, %r416;
cvt.rn.f32.s32 %f94, %r417;
fma.rn.ftz.f32 %f95, %f53, %f94, 0f00000000;
shr.u32 %r418, %r396, 4;
and.b32 %r351, %r418, 252645135;
shr.u32 %r419, %r397, 4;
and.b32 %r355, %r419, 252645135;

dp4a.s32.s32 %r134, %r351, %r100, %r361;


dp4a.s32.s32 %r138, %r355, %r104, %r134;


dp4a.s32.s32 %r142, %r363, %r100, %r361;


dp4a.s32.s32 %r146, %r363, %r104, %r142;

shr.u16 %rs13, %rs11, 8;
cvt.u32.u16 %r420, %rs13;
mul.lo.s32 %r421, %r138, %r420;
cvt.rn.f32.s32 %f96, %r421;
fma.rn.ftz.f32 %f97, %f54, %f96, %f93;
shr.u16 %rs14, %rs12, 8;
cvt.u32.u16 %r422, %rs14;
mul.lo.s32 %r423, %r146, %r422;
cvt.rn.f32.s32 %f98, %r423;
fma.rn.ftz.f32 %f99, %f54, %f98, %f95;
ld.global.nc.u32 %r150, [%rd42];

{.reg .f16 low,high;
mov.b32 {low,high},%r150;
cvt.f32.f16 %f55, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r150;
cvt.f32.f16 %f56, high;}


mul.ftz.f32 %f100, %f97, %f55;
mul.ftz.f32 %f101, %f99, %f56;
sub.ftz.f32 %f102, %f100, %f101;
add.ftz.f32 %f303, %f102, %f303;
cvt.u64.u32 %rd43, %r599;
add.s64 %rd44, %rd3, %rd43;
mul.lo.s64 %rd45, %rd44, 36;
add.s64 %rd46, %rd11, %rd45;
ld.global.nc.u32 %r152, [%rd46];

{.reg .f16 low,high;
mov.b32 {low,high},%r152;
cvt.f32.f16 %f57, low;}


add.s64 %rd47, %rd46, %rd37;
ld.global.nc.u32 %r156, [%rd47+4];
ld.global.nc.u32 %r160, [%rd47+20];
ld.global.nc.u32 %r153, [%rd46+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r153;
cvt.f32.f16 %f58, low;}


ld.global.nc.u32 %r172, [%rd47+40];
ld.global.nc.u32 %r176, [%rd47+56];

dp4a.s32.s32 %r154, %r299, %r156, %r361;


dp4a.s32.s32 %r158, %r303, %r160, %r154;


dp4a.s32.s32 %r162, %r363, %r156, %r361;


dp4a.s32.s32 %r166, %r363, %r160, %r162;

mul.lo.s32 %r424, %r158, %r385;
cvt.rn.f32.s32 %f103, %r424;
fma.rn.ftz.f32 %f104, %f57, %f103, 0f00000000;
mul.lo.s32 %r425, %r166, %r388;
cvt.rn.f32.s32 %f105, %r425;
fma.rn.ftz.f32 %f106, %f57, %f105, 0f00000000;

dp4a.s32.s32 %r170, %r315, %r172, %r361;


dp4a.s32.s32 %r174, %r319, %r176, %r170;


dp4a.s32.s32 %r178, %r363, %r172, %r361;


dp4a.s32.s32 %r182, %r363, %r176, %r178;

mul.lo.s32 %r426, %r174, %r392;
cvt.rn.f32.s32 %f107, %r426;
fma.rn.ftz.f32 %f108, %f58, %f107, %f104;
mul.lo.s32 %r427, %r182, %r394;
cvt.rn.f32.s32 %f109, %r427;
fma.rn.ftz.f32 %f110, %f58, %f109, %f106;

{.reg .f16 low,high;
mov.b32 {low,high},%r114;
cvt.f32.f16 %f59, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r114;
cvt.f32.f16 %f60, high;}


mul.ftz.f32 %f111, %f108, %f59;
mul.ftz.f32 %f112, %f110, %f60;
sub.ftz.f32 %f113, %f111, %f112;
add.ftz.f32 %f302, %f113, %f302;

{.reg .f16 low,high;
mov.b32 {low,high},%r152;
cvt.f32.f16 %f61, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r153;
cvt.f32.f16 %f62, low;}



dp4a.s32.s32 %r190, %r335, %r156, %r361;


dp4a.s32.s32 %r194, %r339, %r160, %r190;


dp4a.s32.s32 %r198, %r363, %r156, %r361;


dp4a.s32.s32 %r202, %r363, %r160, %r198;

mul.lo.s32 %r428, %r194, %r413;
cvt.rn.f32.s32 %f114, %r428;
fma.rn.ftz.f32 %f115, %f61, %f114, 0f00000000;
mul.lo.s32 %r429, %r202, %r416;
cvt.rn.f32.s32 %f116, %r429;
fma.rn.ftz.f32 %f117, %f61, %f116, 0f00000000;

dp4a.s32.s32 %r206, %r351, %r172, %r361;


dp4a.s32.s32 %r210, %r355, %r176, %r206;


dp4a.s32.s32 %r214, %r363, %r172, %r361;


dp4a.s32.s32 %r218, %r363, %r176, %r214;

mul.lo.s32 %r430, %r210, %r420;
cvt.rn.f32.s32 %f118, %r430;
fma.rn.ftz.f32 %f119, %f62, %f118, %f115;
mul.lo.s32 %r431, %r218, %r422;
cvt.rn.f32.s32 %f120, %r431;
fma.rn.ftz.f32 %f121, %f62, %f120, %f117;

{.reg .f16 low,high;
mov.b32 {low,high},%r150;
cvt.f32.f16 %f63, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r150;
cvt.f32.f16 %f64, high;}


mul.ftz.f32 %f122, %f119, %f63;
mul.ftz.f32 %f123, %f121, %f64;
sub.ftz.f32 %f124, %f122, %f123;
add.ftz.f32 %f301, %f124, %f301;
cvt.u64.u32 %rd48, %r601;
add.s64 %rd49, %rd3, %rd48;
mul.lo.s64 %rd50, %rd49, 36;
add.s64 %rd51, %rd11, %rd50;
ld.global.nc.u32 %r224, [%rd51];

{.reg .f16 low,high;
mov.b32 {low,high},%r224;
cvt.f32.f16 %f65, low;}


add.s64 %rd52, %rd51, %rd37;
ld.global.nc.u32 %r228, [%rd52+4];
ld.global.nc.u32 %r232, [%rd52+20];
ld.global.nc.u32 %r225, [%rd51+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r225;
cvt.f32.f16 %f66, low;}


ld.global.nc.u32 %r244, [%rd52+40];
ld.global.nc.u32 %r248, [%rd52+56];

dp4a.s32.s32 %r226, %r299, %r228, %r361;


dp4a.s32.s32 %r230, %r303, %r232, %r226;


dp4a.s32.s32 %r234, %r363, %r228, %r361;


dp4a.s32.s32 %r238, %r363, %r232, %r234;

mul.lo.s32 %r432, %r230, %r385;
cvt.rn.f32.s32 %f125, %r432;
fma.rn.ftz.f32 %f126, %f65, %f125, 0f00000000;
mul.lo.s32 %r433, %r238, %r388;
cvt.rn.f32.s32 %f127, %r433;
fma.rn.ftz.f32 %f128, %f65, %f127, 0f00000000;

dp4a.s32.s32 %r242, %r315, %r244, %r361;


dp4a.s32.s32 %r246, %r319, %r248, %r242;


dp4a.s32.s32 %r250, %r363, %r244, %r361;


dp4a.s32.s32 %r254, %r363, %r248, %r250;

mul.lo.s32 %r434, %r246, %r392;
cvt.rn.f32.s32 %f129, %r434;
fma.rn.ftz.f32 %f130, %f66, %f129, %f126;
mul.lo.s32 %r435, %r254, %r394;
cvt.rn.f32.s32 %f131, %r435;
fma.rn.ftz.f32 %f132, %f66, %f131, %f128;

{.reg .f16 low,high;
mov.b32 {low,high},%r114;
cvt.f32.f16 %f67, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r114;
cvt.f32.f16 %f68, high;}


mul.ftz.f32 %f133, %f130, %f67;
mul.ftz.f32 %f134, %f132, %f68;
sub.ftz.f32 %f135, %f133, %f134;
add.ftz.f32 %f300, %f135, %f300;

{.reg .f16 low,high;
mov.b32 {low,high},%r224;
cvt.f32.f16 %f69, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r225;
cvt.f32.f16 %f70, low;}



dp4a.s32.s32 %r262, %r335, %r228, %r361;


dp4a.s32.s32 %r266, %r339, %r232, %r262;


dp4a.s32.s32 %r270, %r363, %r228, %r361;


dp4a.s32.s32 %r274, %r363, %r232, %r270;

mul.lo.s32 %r436, %r266, %r413;
cvt.rn.f32.s32 %f136, %r436;
fma.rn.ftz.f32 %f137, %f69, %f136, 0f00000000;
mul.lo.s32 %r437, %r274, %r416;
cvt.rn.f32.s32 %f138, %r437;
fma.rn.ftz.f32 %f139, %f69, %f138, 0f00000000;

dp4a.s32.s32 %r278, %r351, %r244, %r361;


dp4a.s32.s32 %r282, %r355, %r248, %r278;


dp4a.s32.s32 %r286, %r363, %r244, %r361;


dp4a.s32.s32 %r290, %r363, %r248, %r286;

mul.lo.s32 %r438, %r282, %r420;
cvt.rn.f32.s32 %f140, %r438;
fma.rn.ftz.f32 %f141, %f70, %f140, %f137;
mul.lo.s32 %r439, %r290, %r422;
cvt.rn.f32.s32 %f142, %r439;
fma.rn.ftz.f32 %f143, %f70, %f142, %f139;

{.reg .f16 low,high;
mov.b32 {low,high},%r150;
cvt.f32.f16 %f71, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r150;
cvt.f32.f16 %f72, high;}


mul.ftz.f32 %f144, %f141, %f71;
mul.ftz.f32 %f145, %f143, %f72;
sub.ftz.f32 %f146, %f144, %f145;
add.ftz.f32 %f299, %f146, %f299;
cvt.u64.u32 %rd53, %r600;
add.s64 %rd54, %rd3, %rd53;
mul.lo.s64 %rd55, %rd54, 36;
add.s64 %rd56, %rd11, %rd55;
ld.global.nc.u32 %r296, [%rd56];

{.reg .f16 low,high;
mov.b32 {low,high},%r296;
cvt.f32.f16 %f73, low;}


add.s64 %rd57, %rd56, %rd37;
ld.global.nc.u32 %r300, [%rd57+4];
ld.global.nc.u32 %r304, [%rd57+20];
ld.global.nc.u32 %r297, [%rd56+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r297;
cvt.f32.f16 %f74, low;}


ld.global.nc.u32 %r316, [%rd57+40];
ld.global.nc.u32 %r320, [%rd57+56];

dp4a.s32.s32 %r298, %r299, %r300, %r361;


dp4a.s32.s32 %r302, %r303, %r304, %r298;


dp4a.s32.s32 %r306, %r363, %r300, %r361;


dp4a.s32.s32 %r310, %r363, %r304, %r306;

mul.lo.s32 %r440, %r302, %r385;
cvt.rn.f32.s32 %f147, %r440;
fma.rn.ftz.f32 %f148, %f73, %f147, 0f00000000;
mul.lo.s32 %r441, %r310, %r388;
cvt.rn.f32.s32 %f149, %r441;
fma.rn.ftz.f32 %f150, %f73, %f149, 0f00000000;

dp4a.s32.s32 %r314, %r315, %r316, %r361;


dp4a.s32.s32 %r318, %r319, %r320, %r314;


dp4a.s32.s32 %r322, %r363, %r316, %r361;


dp4a.s32.s32 %r326, %r363, %r320, %r322;

mul.lo.s32 %r442, %r318, %r392;
cvt.rn.f32.s32 %f151, %r442;
fma.rn.ftz.f32 %f152, %f74, %f151, %f148;
mul.lo.s32 %r443, %r326, %r394;
cvt.rn.f32.s32 %f153, %r443;
fma.rn.ftz.f32 %f154, %f74, %f153, %f150;

{.reg .f16 low,high;
mov.b32 {low,high},%r114;
cvt.f32.f16 %f75, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r114;
cvt.f32.f16 %f76, high;}


mul.ftz.f32 %f155, %f152, %f75;
mul.ftz.f32 %f156, %f154, %f76;
sub.ftz.f32 %f157, %f155, %f156;
add.ftz.f32 %f298, %f157, %f298;

{.reg .f16 low,high;
mov.b32 {low,high},%r296;
cvt.f32.f16 %f77, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r297;
cvt.f32.f16 %f78, low;}



dp4a.s32.s32 %r334, %r335, %r300, %r361;


dp4a.s32.s32 %r338, %r339, %r304, %r334;


dp4a.s32.s32 %r342, %r363, %r300, %r361;


dp4a.s32.s32 %r346, %r363, %r304, %r342;

mul.lo.s32 %r444, %r338, %r413;
cvt.rn.f32.s32 %f158, %r444;
fma.rn.ftz.f32 %f159, %f77, %f158, 0f00000000;
mul.lo.s32 %r445, %r346, %r416;
cvt.rn.f32.s32 %f160, %r445;
fma.rn.ftz.f32 %f161, %f77, %f160, 0f00000000;

dp4a.s32.s32 %r350, %r351, %r316, %r361;


dp4a.s32.s32 %r354, %r355, %r320, %r350;


dp4a.s32.s32 %r358, %r363, %r316, %r361;


dp4a.s32.s32 %r362, %r363, %r320, %r358;

mul.lo.s32 %r446, %r354, %r420;
cvt.rn.f32.s32 %f162, %r446;
fma.rn.ftz.f32 %f163, %f78, %f162, %f159;
mul.lo.s32 %r447, %r362, %r422;
cvt.rn.f32.s32 %f164, %r447;
fma.rn.ftz.f32 %f165, %f78, %f164, %f161;

{.reg .f16 low,high;
mov.b32 {low,high},%r150;
cvt.f32.f16 %f79, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r150;
cvt.f32.f16 %f80, high;}


mul.ftz.f32 %f166, %f163, %f79;
mul.ftz.f32 %f167, %f165, %f80;
sub.ftz.f32 %f168, %f166, %f167;
add.ftz.f32 %f297, %f168, %f297;
add.s32 %r601, %r601, 64;
add.s32 %r600, %r600, 64;
add.s32 %r599, %r599, 64;
add.s32 %r598, %r598, 64;
add.s64 %rd78, %rd78, 1152;
add.s32 %r602, %r602, 8;
setp.lt.s32 %p5, %r602, %r7;
@%p5 bra $L__BB163_2;

$L__BB163_3:
setp.eq.s32 %p6, %r1, 0;
@%p6 bra $L__BB163_5;

shl.b32 %r448, %r1, 10;
mov.u32 %r449, _ZZ13mul_mat_vec_qIL9ggml_type12ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r450, %r449, %r448;
shl.b32 %r451, %r2, 2;
add.s32 %r452, %r450, %r451;
st.shared.f32 [%r452+-1024], %f304;
st.shared.f32 [%r452+-896], %f303;
st.shared.f32 [%r452+-768], %f302;
st.shared.f32 [%r452+-640], %f301;
st.shared.f32 [%r452+-512], %f300;
st.shared.f32 [%r452+-384], %f299;
st.shared.f32 [%r452+-256], %f298;
st.shared.f32 [%r452+-128], %f297;

$L__BB163_5:
setp.ne.s32 %p7, %r1, 0;
bar.sync 0;
@%p7 bra $L__BB163_22;

mov.u32 %r453, %ctaid.z;
mad.lo.s32 %r454, %r5, %r34, %r4;
mad.lo.s32 %r455, %r453, %r40, %r454;
cvt.u64.u32 %rd14, %r455;
shl.b32 %r456, %r2, 2;
mov.u32 %r457, _ZZ13mul_mat_vec_qIL9ggml_type12ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r23, %r457, %r456;
mov.u32 %r458, 2;
ld.shared.f32 %f169, [%r23];
add.ftz.f32 %f170, %f169, %f304;
ld.shared.f32 %f171, [%r23+1024];
add.ftz.f32 %f172, %f171, %f170;
ld.shared.f32 %f173, [%r23+2048];
add.ftz.f32 %f174, %f173, %f172;
mov.b32 %r459, %f174;
mov.u32 %r460, 31;
mov.u32 %r461, 16;
mov.u32 %r462, -1;
shfl.sync.bfly.b32 %r463|%p8, %r459, %r461, %r460, %r462;
mov.b32 %f175, %r463;
add.ftz.f32 %f176, %f174, %f175;
mov.b32 %r464, %f176;
mov.u32 %r465, 8;
shfl.sync.bfly.b32 %r466|%p9, %r464, %r465, %r460, %r462;
mov.b32 %f177, %r466;
add.ftz.f32 %f178, %f176, %f177;
mov.b32 %r467, %f178;
mov.u32 %r468, 4;
shfl.sync.bfly.b32 %r469|%p10, %r467, %r468, %r460, %r462;
mov.b32 %f179, %r469;
add.ftz.f32 %f180, %f178, %f179;
mov.b32 %r470, %f180;
shfl.sync.bfly.b32 %r471|%p11, %r470, %r458, %r460, %r462;
mov.b32 %f181, %r471;
add.ftz.f32 %f182, %f180, %f181;
mov.b32 %r472, %f182;
mov.u32 %r473, 1;
shfl.sync.bfly.b32 %r474|%p12, %r472, %r473, %r460, %r462;
mov.b32 %f183, %r474;
add.ftz.f32 %f25, %f182, %f183;
setp.lt.u32 %p13, %r4, %r28;
setp.eq.s32 %p14, %r2, 0;
and.pred %p1, %p14, %p13;
mul.wide.u32 %rd58, %r455, 4;
add.s64 %rd15, %rd1, %rd58;
not.pred %p15, %p1;
@%p15 bra $L__BB163_8;

st.global.f32 [%rd15], %f25;

$L__BB163_8:
ld.shared.f32 %f184, [%r23+128];
add.ftz.f32 %f185, %f184, %f303;
ld.shared.f32 %f186, [%r23+1152];
add.ftz.f32 %f187, %f186, %f185;
ld.shared.f32 %f188, [%r23+2176];
add.ftz.f32 %f189, %f188, %f187;
mov.b32 %r475, %f189;
shfl.sync.bfly.b32 %r479|%p16, %r475, %r461, %r460, %r462;
mov.b32 %f190, %r479;
add.ftz.f32 %f191, %f189, %f190;
mov.b32 %r480, %f191;
shfl.sync.bfly.b32 %r482|%p17, %r480, %r465, %r460, %r462;
mov.b32 %f192, %r482;
add.ftz.f32 %f193, %f191, %f192;
mov.b32 %r483, %f193;
shfl.sync.bfly.b32 %r485|%p18, %r483, %r468, %r460, %r462;
mov.b32 %f194, %r485;
add.ftz.f32 %f195, %f193, %f194;
mov.b32 %r486, %f195;
shfl.sync.bfly.b32 %r488|%p19, %r486, %r458, %r460, %r462;
mov.b32 %f196, %r488;
add.ftz.f32 %f197, %f195, %f196;
mov.b32 %r489, %f197;
shfl.sync.bfly.b32 %r491|%p20, %r489, %r473, %r460, %r462;
mov.b32 %f198, %r491;
add.ftz.f32 %f26, %f197, %f198;
add.s32 %r492, %r4, 1;
setp.lt.u32 %p21, %r492, %r28;
setp.eq.s32 %p22, %r2, 1;
and.pred %p2, %p22, %p21;
not.pred %p23, %p2;
@%p23 bra $L__BB163_10;

st.global.f32 [%rd15+4], %f26;

$L__BB163_10:
ld.shared.f32 %f199, [%r23+256];
add.ftz.f32 %f200, %f199, %f302;
ld.shared.f32 %f201, [%r23+1280];
add.ftz.f32 %f202, %f201, %f200;
ld.shared.f32 %f203, [%r23+2304];
add.ftz.f32 %f204, %f203, %f202;
mov.b32 %r493, %f204;
mov.u32 %r494, 31;
mov.u32 %r495, 16;
mov.u32 %r496, -1;
shfl.sync.bfly.b32 %r497|%p24, %r493, %r495, %r494, %r496;
mov.b32 %f205, %r497;
add.ftz.f32 %f206, %f204, %f205;
mov.b32 %r498, %f206;
mov.u32 %r499, 8;
shfl.sync.bfly.b32 %r500|%p25, %r498, %r499, %r494, %r496;
mov.b32 %f207, %r500;
add.ftz.f32 %f208, %f206, %f207;
mov.b32 %r501, %f208;
mov.u32 %r502, 4;
shfl.sync.bfly.b32 %r503|%p26, %r501, %r502, %r494, %r496;
mov.b32 %f209, %r503;
add.ftz.f32 %f210, %f208, %f209;
mov.b32 %r504, %f210;
mov.u32 %r505, 2;
shfl.sync.bfly.b32 %r506|%p27, %r504, %r505, %r494, %r496;
mov.b32 %f211, %r506;
add.ftz.f32 %f212, %f210, %f211;
mov.b32 %r507, %f212;
mov.u32 %r508, 1;
shfl.sync.bfly.b32 %r509|%p28, %r507, %r508, %r494, %r496;
mov.b32 %f213, %r509;
add.ftz.f32 %f27, %f212, %f213;
@%p15 bra $L__BB163_12;

cvt.u64.u32 %rd59, %r28;
add.s64 %rd60, %rd59, %rd14;
shl.b64 %rd61, %rd60, 2;
add.s64 %rd62, %rd1, %rd61;
st.global.f32 [%rd62], %f27;

$L__BB163_12:
ld.shared.f32 %f214, [%r23+384];
add.ftz.f32 %f215, %f214, %f301;
ld.shared.f32 %f216, [%r23+1408];
add.ftz.f32 %f217, %f216, %f215;
ld.shared.f32 %f218, [%r23+2432];
add.ftz.f32 %f219, %f218, %f217;
mov.b32 %r510, %f219;
shfl.sync.bfly.b32 %r514|%p30, %r510, %r495, %r494, %r496;
mov.b32 %f220, %r514;
add.ftz.f32 %f221, %f219, %f220;
mov.b32 %r515, %f221;
shfl.sync.bfly.b32 %r517|%p31, %r515, %r499, %r494, %r496;
mov.b32 %f222, %r517;
add.ftz.f32 %f223, %f221, %f222;
mov.b32 %r518, %f223;
shfl.sync.bfly.b32 %r520|%p32, %r518, %r502, %r494, %r496;
mov.b32 %f224, %r520;
add.ftz.f32 %f225, %f223, %f224;
mov.b32 %r521, %f225;
shfl.sync.bfly.b32 %r523|%p33, %r521, %r505, %r494, %r496;
mov.b32 %f226, %r523;
add.ftz.f32 %f227, %f225, %f226;
mov.b32 %r524, %f227;
shfl.sync.bfly.b32 %r526|%p34, %r524, %r508, %r494, %r496;
mov.b32 %f228, %r526;
add.ftz.f32 %f28, %f227, %f228;
@%p23 bra $L__BB163_14;

add.s32 %r527, %r28, 1;
cvt.u64.u32 %rd63, %r527;
add.s64 %rd64, %rd63, %rd14;
shl.b64 %rd65, %rd64, 2;
add.s64 %rd66, %rd1, %rd65;
st.global.f32 [%rd66], %f28;

$L__BB163_14:
shl.b32 %r528, %r28, 1;
mov.u32 %r529, 1;
ld.shared.f32 %f229, [%r23+512];
add.ftz.f32 %f230, %f229, %f300;
ld.shared.f32 %f231, [%r23+1536];
add.ftz.f32 %f232, %f231, %f230;
ld.shared.f32 %f233, [%r23+2560];
add.ftz.f32 %f234, %f233, %f232;
mov.b32 %r530, %f234;
mov.u32 %r531, 31;
mov.u32 %r532, 16;
mov.u32 %r533, -1;
shfl.sync.bfly.b32 %r534|%p36, %r530, %r532, %r531, %r533;
mov.b32 %f235, %r534;
add.ftz.f32 %f236, %f234, %f235;
mov.b32 %r535, %f236;
mov.u32 %r536, 8;
shfl.sync.bfly.b32 %r537|%p37, %r535, %r536, %r531, %r533;
mov.b32 %f237, %r537;
add.ftz.f32 %f238, %f236, %f237;
mov.b32 %r538, %f238;
mov.u32 %r539, 4;
shfl.sync.bfly.b32 %r540|%p38, %r538, %r539, %r531, %r533;
mov.b32 %f239, %r540;
add.ftz.f32 %f240, %f238, %f239;
mov.b32 %r541, %f240;
mov.u32 %r542, 2;
shfl.sync.bfly.b32 %r543|%p39, %r541, %r542, %r531, %r533;
mov.b32 %f241, %r543;
add.ftz.f32 %f242, %f240, %f241;
mov.b32 %r544, %f242;
shfl.sync.bfly.b32 %r545|%p40, %r544, %r529, %r531, %r533;
mov.b32 %f243, %r545;
add.ftz.f32 %f29, %f242, %f243;
cvt.u64.u32 %rd67, %r528;
add.s64 %rd68, %rd67, %rd14;
shl.b64 %rd69, %rd68, 2;
add.s64 %rd16, %rd1, %rd69;
@%p15 bra $L__BB163_16;

st.global.f32 [%rd16], %f29;

$L__BB163_16:
ld.shared.f32 %f244, [%r23+640];
add.ftz.f32 %f245, %f244, %f299;
ld.shared.f32 %f246, [%r23+1664];
add.ftz.f32 %f247, %f246, %f245;
ld.shared.f32 %f248, [%r23+2688];
add.ftz.f32 %f249, %f248, %f247;
mov.b32 %r546, %f249;
shfl.sync.bfly.b32 %r550|%p42, %r546, %r532, %r531, %r533;
mov.b32 %f250, %r550;
add.ftz.f32 %f251, %f249, %f250;
mov.b32 %r551, %f251;
shfl.sync.bfly.b32 %r553|%p43, %r551, %r536, %r531, %r533;
mov.b32 %f252, %r553;
add.ftz.f32 %f253, %f251, %f252;
mov.b32 %r554, %f253;
shfl.sync.bfly.b32 %r556|%p44, %r554, %r539, %r531, %r533;
mov.b32 %f254, %r556;
add.ftz.f32 %f255, %f253, %f254;
mov.b32 %r557, %f255;
shfl.sync.bfly.b32 %r559|%p45, %r557, %r542, %r531, %r533;
mov.b32 %f256, %r559;
add.ftz.f32 %f257, %f255, %f256;
mov.b32 %r560, %f257;
shfl.sync.bfly.b32 %r562|%p46, %r560, %r529, %r531, %r533;
mov.b32 %f258, %r562;
add.ftz.f32 %f30, %f257, %f258;
@%p23 bra $L__BB163_18;

st.global.f32 [%rd16+4], %f30;

$L__BB163_18:
mul.lo.s32 %r25, %r28, 3;
ld.shared.f32 %f259, [%r23+768];
add.ftz.f32 %f260, %f259, %f298;
ld.shared.f32 %f261, [%r23+1792];
add.ftz.f32 %f262, %f261, %f260;
ld.shared.f32 %f263, [%r23+2816];
add.ftz.f32 %f264, %f263, %f262;
mov.b32 %r563, %f264;
mov.u32 %r564, 31;
mov.u32 %r565, 16;
mov.u32 %r566, -1;
shfl.sync.bfly.b32 %r567|%p48, %r563, %r565, %r564, %r566;
mov.b32 %f265, %r567;
add.ftz.f32 %f266, %f264, %f265;
mov.b32 %r568, %f266;
mov.u32 %r569, 8;
shfl.sync.bfly.b32 %r570|%p49, %r568, %r569, %r564, %r566;
mov.b32 %f267, %r570;
add.ftz.f32 %f268, %f266, %f267;
mov.b32 %r571, %f268;
mov.u32 %r572, 4;
shfl.sync.bfly.b32 %r573|%p50, %r571, %r572, %r564, %r566;
mov.b32 %f269, %r573;
add.ftz.f32 %f270, %f268, %f269;
mov.b32 %r574, %f270;
mov.u32 %r575, 2;
shfl.sync.bfly.b32 %r576|%p51, %r574, %r575, %r564, %r566;
mov.b32 %f271, %r576;
add.ftz.f32 %f272, %f270, %f271;
mov.b32 %r577, %f272;
mov.u32 %r578, 1;
shfl.sync.bfly.b32 %r579|%p52, %r577, %r578, %r564, %r566;
mov.b32 %f273, %r579;
add.ftz.f32 %f31, %f272, %f273;
@%p15 bra $L__BB163_20;

cvt.u64.u32 %rd70, %r25;
add.s64 %rd71, %rd70, %rd14;
shl.b64 %rd72, %rd71, 2;
add.s64 %rd73, %rd1, %rd72;
st.global.f32 [%rd73], %f31;

$L__BB163_20:
ld.shared.f32 %f274, [%r23+896];
add.ftz.f32 %f275, %f274, %f297;
ld.shared.f32 %f276, [%r23+1920];
add.ftz.f32 %f277, %f276, %f275;
ld.shared.f32 %f278, [%r23+2944];
add.ftz.f32 %f279, %f278, %f277;
mov.b32 %r580, %f279;
shfl.sync.bfly.b32 %r584|%p54, %r580, %r565, %r564, %r566;
mov.b32 %f280, %r584;
add.ftz.f32 %f281, %f279, %f280;
mov.b32 %r585, %f281;
shfl.sync.bfly.b32 %r587|%p55, %r585, %r569, %r564, %r566;
mov.b32 %f282, %r587;
add.ftz.f32 %f283, %f281, %f282;
mov.b32 %r588, %f283;
shfl.sync.bfly.b32 %r590|%p56, %r588, %r572, %r564, %r566;
mov.b32 %f284, %r590;
add.ftz.f32 %f285, %f283, %f284;
mov.b32 %r591, %f285;
shfl.sync.bfly.b32 %r593|%p57, %r591, %r575, %r564, %r566;
mov.b32 %f286, %r593;
add.ftz.f32 %f287, %f285, %f286;
mov.b32 %r594, %f287;
shfl.sync.bfly.b32 %r596|%p58, %r594, %r578, %r564, %r566;
mov.b32 %f288, %r596;
add.ftz.f32 %f32, %f287, %f288;
@%p23 bra $L__BB163_22;

add.s32 %r597, %r25, 1;
cvt.u64.u32 %rd74, %r597;
add.s64 %rd75, %rd74, %rd14;
shl.b64 %rd76, %rd75, 2;
add.s64 %rd77, %rd1, %rd76;
st.global.f32 [%rd77], %f32;

$L__BB163_22:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type12ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type12ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type12ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type12ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type12ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type12ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type12ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type12ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type12ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<72>;
.reg .b16 %rs<15>;
.reg .f32 %f<341>;
.reg .b32 %r<727>;
.reg .b64 %rd<94>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type12ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1280];

ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type12ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type12ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type12ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type12ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type12ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type12ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type12ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type12ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type12ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type12ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type12ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type12ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type12ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type12ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type12ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type12ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type12ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r43, %r1, 5;
mov.u32 %r44, %tid.x;
add.s32 %r2, %r43, %r44;
mov.u32 %r3, %ctaid.y;
mov.u32 %r4, %ctaid.z;
shr.s32 %r45, %r2, 31;
shr.u32 %r46, %r45, 28;
add.s32 %r47, %r2, %r46;
shr.s32 %r726, %r47, 4;
shr.u32 %r6, %r42, 8;
setp.ge.s32 %p3, %r726, %r6;
mov.f32 %f331, 0f00000000;
mov.f32 %f332, %f331;
mov.f32 %f333, %f331;
mov.f32 %f334, %f331;
mov.f32 %f335, %f331;
mov.f32 %f336, %f331;
mov.f32 %f337, %f331;
mov.f32 %f338, %f331;
mov.f32 %f339, %f331;
mov.f32 %f340, %f331;
@%p3 bra $L__BB164_3;

mov.u32 %r48, %ctaid.x;
shl.b32 %r49, %r48, 1;
mul.hi.u32 %r50, %r4, %r36;
add.s32 %r51, %r50, %r4;
shr.u32 %r52, %r51, %r37;
mul.lo.s32 %r53, %r52, %r39;
mul.hi.u32 %r54, %r3, %r30;
add.s32 %r55, %r54, %r3;
shr.u32 %r56, %r55, %r31;
mul.lo.s32 %r57, %r56, %r33;
mul.lo.s32 %r58, %r3, %r34;
cvt.u64.u32 %rd20, %r58;
cvta.to.global.u64 %rd93, %rd17;
and.b32 %r62, %r47, -16;
sub.s32 %r7, %r2, %r62;
shr.s32 %r63, %r7, 31;
shr.u32 %r64, %r63, 30;
add.s32 %r65, %r7, %r64;
shr.s32 %r66, %r65, 2;
shl.b32 %r67, %r66, 1;
shl.b32 %r68, %r66, 5;
and.b32 %r69, %r65, -4;
sub.s32 %r70, %r7, %r69;
shl.b32 %r71, %r70, 2;
add.s32 %r72, %r68, %r71;
cvt.s64.s32 %rd21, %r72;
cvt.s64.s32 %rd22, %r67;
add.s64 %rd23, %rd20, %rd22;
mul.lo.s32 %r73, %r4, %r40;
cvt.u64.u32 %rd24, %r73;
add.s64 %rd2, %rd23, %rd24;
cvt.s64.s32 %rd3, %r70;
shl.b32 %r74, %r28, 1;
shl.b32 %r721, %r726, 3;
add.s32 %r725, %r721, %r74;
mad.lo.s32 %r724, %r28, 3, %r721;
shl.b32 %r75, %r28, 2;
add.s32 %r723, %r721, %r75;
add.s32 %r722, %r28, %r721;
add.s32 %r76, %r726, %r53;
add.s32 %r77, %r76, %r57;
or.b32 %r78, %r49, 1;
mad.lo.s32 %r79, %r27, %r78, %r77;
mul.wide.s32 %rd4, %r79, 144;
mad.lo.s32 %r80, %r49, %r27, %r726;
add.s32 %r81, %r80, %r53;
add.s32 %r82, %r81, %r57;
mul.wide.s32 %rd5, %r82, 144;
add.s64 %rd25, %rd5, %rd21;
add.s64 %rd6, %rd25, 32;
cvt.u64.u32 %rd26, %r67;
and.b64 %rd27, %rd26, 2;
or.b64 %rd28, %rd27, %rd4;
or.b64 %rd7, %rd28, 8;
or.b64 %rd29, %rd27, %rd5;
or.b64 %rd8, %rd29, 12;
add.s64 %rd30, %rd4, %rd21;
add.s64 %rd9, %rd30, 16;
cvta.to.global.u64 %rd10, %rd18;
mov.f32 %f331, 0f00000000;
shl.b64 %rd37, %rd3, 2;

$L__BB164_2:
cvt.u64.u32 %rd31, %r721;
add.s64 %rd32, %rd2, %rd31;
add.s64 %rd33, %rd93, %rd6;
ld.global.nc.u32 %r443, [%rd33+-16];
ld.global.nc.u32 %r444, [%rd33];
add.s64 %rd34, %rd93, %rd8;
ld.global.nc.u16 %rs1, [%rd34+-8];
cvt.u32.u16 %r445, %rs1;
ld.global.nc.u16 %rs2, [%rd34+-4];
cvt.u32.u16 %r446, %rs2;
ld.global.nc.u16 %rs3, [%rd34];
cvt.u32.u16 %r447, %rs3;
and.b32 %r448, %r445, 16191;
and.b32 %r449, %r447, 3855;
and.b32 %r450, %r445, 49344;
shr.u32 %r451, %r450, 2;
or.b32 %r452, %r449, %r451;
setp.gt.s32 %p4, %r7, 7;
and.b32 %r453, %r446, 16191;
shr.u32 %r454, %r447, 4;
and.b32 %r455, %r454, 3855;
and.b32 %r456, %r446, 49344;
shr.u32 %r457, %r456, 2;
or.b32 %r458, %r455, %r457;
mul.lo.s64 %rd35, %rd32, 36;
add.s64 %rd36, %rd10, %rd35;
ld.global.nc.u32 %r83, [%rd36];

{.reg .f16 low,high;
mov.b32 {low,high},%r83;
cvt.f32.f16 %f61, low;}


add.s64 %rd38, %rd36, %rd37;
ld.global.nc.u32 %r87, [%rd38+4];
ld.global.nc.u32 %r91, [%rd38+20];
ld.global.nc.u32 %r84, [%rd36+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r84;
cvt.f32.f16 %f62, low;}


ld.global.nc.u32 %r103, [%rd38+40];
ld.global.nc.u32 %r107, [%rd38+56];
add.s64 %rd39, %rd93, %rd5;
and.b32 %r374, %r443, 252645135;
and.b32 %r378, %r444, 252645135;
mov.u32 %r436, 0;

dp4a.s32.s32 %r85, %r374, %r87, %r436;


dp4a.s32.s32 %r89, %r378, %r91, %r85;

mov.u32 %r438, 16843009;

dp4a.s32.s32 %r93, %r438, %r87, %r436;


dp4a.s32.s32 %r97, %r438, %r91, %r93;

selp.b32 %r459, %r452, %r448, %p4;
cvt.u16.u32 %rs4, %r459;
and.b32 %r460, %r459, 63;
mul.lo.s32 %r461, %r89, %r460;
cvt.rn.f32.s32 %f101, %r461;
fma.rn.ftz.f32 %f102, %f61, %f101, 0f00000000;
selp.b32 %r462, %r458, %r453, %p4;
cvt.u16.u32 %rs5, %r462;
and.b32 %r463, %r462, 63;
mul.lo.s32 %r464, %r97, %r463;
cvt.rn.f32.s32 %f103, %r464;
fma.rn.ftz.f32 %f104, %f61, %f103, 0f00000000;
shr.u32 %r465, %r443, 4;
and.b32 %r390, %r465, 252645135;
shr.u32 %r466, %r444, 4;
and.b32 %r394, %r466, 252645135;

dp4a.s32.s32 %r101, %r390, %r103, %r436;


dp4a.s32.s32 %r105, %r394, %r107, %r101;


dp4a.s32.s32 %r109, %r438, %r103, %r436;


dp4a.s32.s32 %r113, %r438, %r107, %r109;

shr.u16 %rs6, %rs4, 8;
cvt.u32.u16 %r467, %rs6;
mul.lo.s32 %r468, %r105, %r467;
cvt.rn.f32.s32 %f105, %r468;
fma.rn.ftz.f32 %f106, %f62, %f105, %f102;
shr.u16 %rs7, %rs5, 8;
cvt.u32.u16 %r469, %rs7;
mul.lo.s32 %r470, %r113, %r469;
cvt.rn.f32.s32 %f107, %r470;
fma.rn.ftz.f32 %f108, %f62, %f107, %f104;
ld.global.nc.u32 %r117, [%rd39];

{.reg .f16 low,high;
mov.b32 {low,high},%r117;
cvt.f32.f16 %f63, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r117;
cvt.f32.f16 %f64, high;}


mul.ftz.f32 %f109, %f106, %f63;
mul.ftz.f32 %f110, %f108, %f64;
sub.ftz.f32 %f111, %f109, %f110;
add.ftz.f32 %f340, %f111, %f340;
add.s64 %rd40, %rd93, %rd9;
ld.global.nc.u32 %r471, [%rd40];
ld.global.nc.u32 %r472, [%rd40+16];
add.s64 %rd41, %rd93, %rd7;
ld.global.nc.u16 %rs8, [%rd41+-4];
cvt.u32.u16 %r473, %rs8;
ld.global.nc.u16 %rs9, [%rd41];
cvt.u32.u16 %r474, %rs9;
ld.global.nc.u16 %rs10, [%rd41+4];
cvt.u32.u16 %r475, %rs10;
and.b32 %r476, %r473, 16191;
and.b32 %r477, %r475, 3855;
and.b32 %r478, %r473, 49344;
shr.u32 %r479, %r478, 2;
or.b32 %r480, %r477, %r479;
and.b32 %r481, %r474, 16191;
shr.u32 %r482, %r475, 4;
and.b32 %r483, %r482, 3855;
and.b32 %r484, %r474, 49344;
shr.u32 %r485, %r484, 2;
or.b32 %r486, %r483, %r485;

{.reg .f16 low,high;
mov.b32 {low,high},%r83;
cvt.f32.f16 %f65, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r84;
cvt.f32.f16 %f66, low;}


add.s64 %rd42, %rd93, %rd4;
and.b32 %r410, %r471, 252645135;
and.b32 %r414, %r472, 252645135;

dp4a.s32.s32 %r121, %r410, %r87, %r436;


dp4a.s32.s32 %r125, %r414, %r91, %r121;


dp4a.s32.s32 %r129, %r438, %r87, %r436;


dp4a.s32.s32 %r133, %r438, %r91, %r129;

selp.b32 %r487, %r480, %r476, %p4;
cvt.u16.u32 %rs11, %r487;
and.b32 %r488, %r487, 63;
mul.lo.s32 %r489, %r125, %r488;
cvt.rn.f32.s32 %f112, %r489;
fma.rn.ftz.f32 %f113, %f65, %f112, 0f00000000;
selp.b32 %r490, %r486, %r481, %p4;
cvt.u16.u32 %rs12, %r490;
and.b32 %r491, %r490, 63;
mul.lo.s32 %r492, %r133, %r491;
cvt.rn.f32.s32 %f114, %r492;
fma.rn.ftz.f32 %f115, %f65, %f114, 0f00000000;
shr.u32 %r493, %r471, 4;
and.b32 %r426, %r493, 252645135;
shr.u32 %r494, %r472, 4;
and.b32 %r430, %r494, 252645135;

dp4a.s32.s32 %r137, %r426, %r103, %r436;


dp4a.s32.s32 %r141, %r430, %r107, %r137;


dp4a.s32.s32 %r145, %r438, %r103, %r436;


dp4a.s32.s32 %r149, %r438, %r107, %r145;

shr.u16 %rs13, %rs11, 8;
cvt.u32.u16 %r495, %rs13;
mul.lo.s32 %r496, %r141, %r495;
cvt.rn.f32.s32 %f116, %r496;
fma.rn.ftz.f32 %f117, %f66, %f116, %f113;
shr.u16 %rs14, %rs12, 8;
cvt.u32.u16 %r497, %rs14;
mul.lo.s32 %r498, %r149, %r497;
cvt.rn.f32.s32 %f118, %r498;
fma.rn.ftz.f32 %f119, %f66, %f118, %f115;
ld.global.nc.u32 %r153, [%rd42];

{.reg .f16 low,high;
mov.b32 {low,high},%r153;
cvt.f32.f16 %f67, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r153;
cvt.f32.f16 %f68, high;}


mul.ftz.f32 %f120, %f117, %f67;
mul.ftz.f32 %f121, %f119, %f68;
sub.ftz.f32 %f122, %f120, %f121;
add.ftz.f32 %f339, %f122, %f339;
cvt.u64.u32 %rd43, %r722;
add.s64 %rd44, %rd2, %rd43;
mul.lo.s64 %rd45, %rd44, 36;
add.s64 %rd46, %rd10, %rd45;
ld.global.nc.u32 %r155, [%rd46];

{.reg .f16 low,high;
mov.b32 {low,high},%r155;
cvt.f32.f16 %f69, low;}


add.s64 %rd47, %rd46, %rd37;
ld.global.nc.u32 %r159, [%rd47+4];
ld.global.nc.u32 %r163, [%rd47+20];
ld.global.nc.u32 %r156, [%rd46+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r156;
cvt.f32.f16 %f70, low;}


ld.global.nc.u32 %r175, [%rd47+40];
ld.global.nc.u32 %r179, [%rd47+56];

dp4a.s32.s32 %r157, %r374, %r159, %r436;


dp4a.s32.s32 %r161, %r378, %r163, %r157;


dp4a.s32.s32 %r165, %r438, %r159, %r436;


dp4a.s32.s32 %r169, %r438, %r163, %r165;

mul.lo.s32 %r499, %r161, %r460;
cvt.rn.f32.s32 %f123, %r499;
fma.rn.ftz.f32 %f124, %f69, %f123, 0f00000000;
mul.lo.s32 %r500, %r169, %r463;
cvt.rn.f32.s32 %f125, %r500;
fma.rn.ftz.f32 %f126, %f69, %f125, 0f00000000;

dp4a.s32.s32 %r173, %r390, %r175, %r436;


dp4a.s32.s32 %r177, %r394, %r179, %r173;


dp4a.s32.s32 %r181, %r438, %r175, %r436;


dp4a.s32.s32 %r185, %r438, %r179, %r181;

mul.lo.s32 %r501, %r177, %r467;
cvt.rn.f32.s32 %f127, %r501;
fma.rn.ftz.f32 %f128, %f70, %f127, %f124;
mul.lo.s32 %r502, %r185, %r469;
cvt.rn.f32.s32 %f129, %r502;
fma.rn.ftz.f32 %f130, %f70, %f129, %f126;

{.reg .f16 low,high;
mov.b32 {low,high},%r117;
cvt.f32.f16 %f71, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r117;
cvt.f32.f16 %f72, high;}


mul.ftz.f32 %f131, %f128, %f71;
mul.ftz.f32 %f132, %f130, %f72;
sub.ftz.f32 %f133, %f131, %f132;
add.ftz.f32 %f338, %f133, %f338;

{.reg .f16 low,high;
mov.b32 {low,high},%r155;
cvt.f32.f16 %f73, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r156;
cvt.f32.f16 %f74, low;}



dp4a.s32.s32 %r193, %r410, %r159, %r436;


dp4a.s32.s32 %r197, %r414, %r163, %r193;


dp4a.s32.s32 %r201, %r438, %r159, %r436;


dp4a.s32.s32 %r205, %r438, %r163, %r201;

mul.lo.s32 %r503, %r197, %r488;
cvt.rn.f32.s32 %f134, %r503;
fma.rn.ftz.f32 %f135, %f73, %f134, 0f00000000;
mul.lo.s32 %r504, %r205, %r491;
cvt.rn.f32.s32 %f136, %r504;
fma.rn.ftz.f32 %f137, %f73, %f136, 0f00000000;

dp4a.s32.s32 %r209, %r426, %r175, %r436;


dp4a.s32.s32 %r213, %r430, %r179, %r209;


dp4a.s32.s32 %r217, %r438, %r175, %r436;


dp4a.s32.s32 %r221, %r438, %r179, %r217;

mul.lo.s32 %r505, %r213, %r495;
cvt.rn.f32.s32 %f138, %r505;
fma.rn.ftz.f32 %f139, %f74, %f138, %f135;
mul.lo.s32 %r506, %r221, %r497;
cvt.rn.f32.s32 %f140, %r506;
fma.rn.ftz.f32 %f141, %f74, %f140, %f137;

{.reg .f16 low,high;
mov.b32 {low,high},%r153;
cvt.f32.f16 %f75, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r153;
cvt.f32.f16 %f76, high;}


mul.ftz.f32 %f142, %f139, %f75;
mul.ftz.f32 %f143, %f141, %f76;
sub.ftz.f32 %f144, %f142, %f143;
add.ftz.f32 %f337, %f144, %f337;
cvt.u64.u32 %rd48, %r725;
add.s64 %rd49, %rd2, %rd48;
mul.lo.s64 %rd50, %rd49, 36;
add.s64 %rd51, %rd10, %rd50;
ld.global.nc.u32 %r227, [%rd51];

{.reg .f16 low,high;
mov.b32 {low,high},%r227;
cvt.f32.f16 %f77, low;}


add.s64 %rd52, %rd51, %rd37;
ld.global.nc.u32 %r231, [%rd52+4];
ld.global.nc.u32 %r235, [%rd52+20];
ld.global.nc.u32 %r228, [%rd51+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r228;
cvt.f32.f16 %f78, low;}


ld.global.nc.u32 %r247, [%rd52+40];
ld.global.nc.u32 %r251, [%rd52+56];

dp4a.s32.s32 %r229, %r374, %r231, %r436;


dp4a.s32.s32 %r233, %r378, %r235, %r229;


dp4a.s32.s32 %r237, %r438, %r231, %r436;


dp4a.s32.s32 %r241, %r438, %r235, %r237;

mul.lo.s32 %r507, %r233, %r460;
cvt.rn.f32.s32 %f145, %r507;
fma.rn.ftz.f32 %f146, %f77, %f145, 0f00000000;
mul.lo.s32 %r508, %r241, %r463;
cvt.rn.f32.s32 %f147, %r508;
fma.rn.ftz.f32 %f148, %f77, %f147, 0f00000000;

dp4a.s32.s32 %r245, %r390, %r247, %r436;


dp4a.s32.s32 %r249, %r394, %r251, %r245;


dp4a.s32.s32 %r253, %r438, %r247, %r436;


dp4a.s32.s32 %r257, %r438, %r251, %r253;

mul.lo.s32 %r509, %r249, %r467;
cvt.rn.f32.s32 %f149, %r509;
fma.rn.ftz.f32 %f150, %f78, %f149, %f146;
mul.lo.s32 %r510, %r257, %r469;
cvt.rn.f32.s32 %f151, %r510;
fma.rn.ftz.f32 %f152, %f78, %f151, %f148;

{.reg .f16 low,high;
mov.b32 {low,high},%r117;
cvt.f32.f16 %f79, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r117;
cvt.f32.f16 %f80, high;}


mul.ftz.f32 %f153, %f150, %f79;
mul.ftz.f32 %f154, %f152, %f80;
sub.ftz.f32 %f155, %f153, %f154;
add.ftz.f32 %f336, %f155, %f336;

{.reg .f16 low,high;
mov.b32 {low,high},%r227;
cvt.f32.f16 %f81, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r228;
cvt.f32.f16 %f82, low;}



dp4a.s32.s32 %r265, %r410, %r231, %r436;


dp4a.s32.s32 %r269, %r414, %r235, %r265;


dp4a.s32.s32 %r273, %r438, %r231, %r436;


dp4a.s32.s32 %r277, %r438, %r235, %r273;

mul.lo.s32 %r511, %r269, %r488;
cvt.rn.f32.s32 %f156, %r511;
fma.rn.ftz.f32 %f157, %f81, %f156, 0f00000000;
mul.lo.s32 %r512, %r277, %r491;
cvt.rn.f32.s32 %f158, %r512;
fma.rn.ftz.f32 %f159, %f81, %f158, 0f00000000;

dp4a.s32.s32 %r281, %r426, %r247, %r436;


dp4a.s32.s32 %r285, %r430, %r251, %r281;


dp4a.s32.s32 %r289, %r438, %r247, %r436;


dp4a.s32.s32 %r293, %r438, %r251, %r289;

mul.lo.s32 %r513, %r285, %r495;
cvt.rn.f32.s32 %f160, %r513;
fma.rn.ftz.f32 %f161, %f82, %f160, %f157;
mul.lo.s32 %r514, %r293, %r497;
cvt.rn.f32.s32 %f162, %r514;
fma.rn.ftz.f32 %f163, %f82, %f162, %f159;

{.reg .f16 low,high;
mov.b32 {low,high},%r153;
cvt.f32.f16 %f83, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r153;
cvt.f32.f16 %f84, high;}


mul.ftz.f32 %f164, %f161, %f83;
mul.ftz.f32 %f165, %f163, %f84;
sub.ftz.f32 %f166, %f164, %f165;
add.ftz.f32 %f335, %f166, %f335;
cvt.u64.u32 %rd53, %r724;
add.s64 %rd54, %rd2, %rd53;
mul.lo.s64 %rd55, %rd54, 36;
add.s64 %rd56, %rd10, %rd55;
ld.global.nc.u32 %r299, [%rd56];

{.reg .f16 low,high;
mov.b32 {low,high},%r299;
cvt.f32.f16 %f85, low;}


add.s64 %rd57, %rd56, %rd37;
ld.global.nc.u32 %r303, [%rd57+4];
ld.global.nc.u32 %r307, [%rd57+20];
ld.global.nc.u32 %r300, [%rd56+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r300;
cvt.f32.f16 %f86, low;}


ld.global.nc.u32 %r319, [%rd57+40];
ld.global.nc.u32 %r323, [%rd57+56];

dp4a.s32.s32 %r301, %r374, %r303, %r436;


dp4a.s32.s32 %r305, %r378, %r307, %r301;


dp4a.s32.s32 %r309, %r438, %r303, %r436;


dp4a.s32.s32 %r313, %r438, %r307, %r309;

mul.lo.s32 %r515, %r305, %r460;
cvt.rn.f32.s32 %f167, %r515;
fma.rn.ftz.f32 %f168, %f85, %f167, 0f00000000;
mul.lo.s32 %r516, %r313, %r463;
cvt.rn.f32.s32 %f169, %r516;
fma.rn.ftz.f32 %f170, %f85, %f169, 0f00000000;

dp4a.s32.s32 %r317, %r390, %r319, %r436;


dp4a.s32.s32 %r321, %r394, %r323, %r317;


dp4a.s32.s32 %r325, %r438, %r319, %r436;


dp4a.s32.s32 %r329, %r438, %r323, %r325;

mul.lo.s32 %r517, %r321, %r467;
cvt.rn.f32.s32 %f171, %r517;
fma.rn.ftz.f32 %f172, %f86, %f171, %f168;
mul.lo.s32 %r518, %r329, %r469;
cvt.rn.f32.s32 %f173, %r518;
fma.rn.ftz.f32 %f174, %f86, %f173, %f170;

{.reg .f16 low,high;
mov.b32 {low,high},%r117;
cvt.f32.f16 %f87, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r117;
cvt.f32.f16 %f88, high;}


mul.ftz.f32 %f175, %f172, %f87;
mul.ftz.f32 %f176, %f174, %f88;
sub.ftz.f32 %f177, %f175, %f176;
add.ftz.f32 %f334, %f177, %f334;

{.reg .f16 low,high;
mov.b32 {low,high},%r299;
cvt.f32.f16 %f89, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r300;
cvt.f32.f16 %f90, low;}



dp4a.s32.s32 %r337, %r410, %r303, %r436;


dp4a.s32.s32 %r341, %r414, %r307, %r337;


dp4a.s32.s32 %r345, %r438, %r303, %r436;


dp4a.s32.s32 %r349, %r438, %r307, %r345;

mul.lo.s32 %r519, %r341, %r488;
cvt.rn.f32.s32 %f178, %r519;
fma.rn.ftz.f32 %f179, %f89, %f178, 0f00000000;
mul.lo.s32 %r520, %r349, %r491;
cvt.rn.f32.s32 %f180, %r520;
fma.rn.ftz.f32 %f181, %f89, %f180, 0f00000000;

dp4a.s32.s32 %r353, %r426, %r319, %r436;


dp4a.s32.s32 %r357, %r430, %r323, %r353;


dp4a.s32.s32 %r361, %r438, %r319, %r436;


dp4a.s32.s32 %r365, %r438, %r323, %r361;

mul.lo.s32 %r521, %r357, %r495;
cvt.rn.f32.s32 %f182, %r521;
fma.rn.ftz.f32 %f183, %f90, %f182, %f179;
mul.lo.s32 %r522, %r365, %r497;
cvt.rn.f32.s32 %f184, %r522;
fma.rn.ftz.f32 %f185, %f90, %f184, %f181;

{.reg .f16 low,high;
mov.b32 {low,high},%r153;
cvt.f32.f16 %f91, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r153;
cvt.f32.f16 %f92, high;}


mul.ftz.f32 %f186, %f183, %f91;
mul.ftz.f32 %f187, %f185, %f92;
sub.ftz.f32 %f188, %f186, %f187;
add.ftz.f32 %f333, %f188, %f333;
cvt.u64.u32 %rd58, %r723;
add.s64 %rd59, %rd2, %rd58;
mul.lo.s64 %rd60, %rd59, 36;
add.s64 %rd61, %rd10, %rd60;
ld.global.nc.u32 %r371, [%rd61];

{.reg .f16 low,high;
mov.b32 {low,high},%r371;
cvt.f32.f16 %f93, low;}


add.s64 %rd62, %rd61, %rd37;
ld.global.nc.u32 %r375, [%rd62+4];
ld.global.nc.u32 %r379, [%rd62+20];
ld.global.nc.u32 %r372, [%rd61+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r372;
cvt.f32.f16 %f94, low;}


ld.global.nc.u32 %r391, [%rd62+40];
ld.global.nc.u32 %r395, [%rd62+56];

dp4a.s32.s32 %r373, %r374, %r375, %r436;


dp4a.s32.s32 %r377, %r378, %r379, %r373;


dp4a.s32.s32 %r381, %r438, %r375, %r436;


dp4a.s32.s32 %r385, %r438, %r379, %r381;

mul.lo.s32 %r523, %r377, %r460;
cvt.rn.f32.s32 %f189, %r523;
fma.rn.ftz.f32 %f190, %f93, %f189, 0f00000000;
mul.lo.s32 %r524, %r385, %r463;
cvt.rn.f32.s32 %f191, %r524;
fma.rn.ftz.f32 %f192, %f93, %f191, 0f00000000;

dp4a.s32.s32 %r389, %r390, %r391, %r436;


dp4a.s32.s32 %r393, %r394, %r395, %r389;


dp4a.s32.s32 %r397, %r438, %r391, %r436;


dp4a.s32.s32 %r401, %r438, %r395, %r397;

mul.lo.s32 %r525, %r393, %r467;
cvt.rn.f32.s32 %f193, %r525;
fma.rn.ftz.f32 %f194, %f94, %f193, %f190;
mul.lo.s32 %r526, %r401, %r469;
cvt.rn.f32.s32 %f195, %r526;
fma.rn.ftz.f32 %f196, %f94, %f195, %f192;

{.reg .f16 low,high;
mov.b32 {low,high},%r117;
cvt.f32.f16 %f95, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r117;
cvt.f32.f16 %f96, high;}


mul.ftz.f32 %f197, %f194, %f95;
mul.ftz.f32 %f198, %f196, %f96;
sub.ftz.f32 %f199, %f197, %f198;
add.ftz.f32 %f332, %f199, %f332;

{.reg .f16 low,high;
mov.b32 {low,high},%r371;
cvt.f32.f16 %f97, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r372;
cvt.f32.f16 %f98, low;}



dp4a.s32.s32 %r409, %r410, %r375, %r436;


dp4a.s32.s32 %r413, %r414, %r379, %r409;


dp4a.s32.s32 %r417, %r438, %r375, %r436;


dp4a.s32.s32 %r421, %r438, %r379, %r417;

mul.lo.s32 %r527, %r413, %r488;
cvt.rn.f32.s32 %f200, %r527;
fma.rn.ftz.f32 %f201, %f97, %f200, 0f00000000;
mul.lo.s32 %r528, %r421, %r491;
cvt.rn.f32.s32 %f202, %r528;
fma.rn.ftz.f32 %f203, %f97, %f202, 0f00000000;

dp4a.s32.s32 %r425, %r426, %r391, %r436;


dp4a.s32.s32 %r429, %r430, %r395, %r425;


dp4a.s32.s32 %r433, %r438, %r391, %r436;


dp4a.s32.s32 %r437, %r438, %r395, %r433;

mul.lo.s32 %r529, %r429, %r495;
cvt.rn.f32.s32 %f204, %r529;
fma.rn.ftz.f32 %f205, %f98, %f204, %f201;
mul.lo.s32 %r530, %r437, %r497;
cvt.rn.f32.s32 %f206, %r530;
fma.rn.ftz.f32 %f207, %f98, %f206, %f203;

{.reg .f16 low,high;
mov.b32 {low,high},%r153;
cvt.f32.f16 %f99, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r153;
cvt.f32.f16 %f100, high;}


mul.ftz.f32 %f208, %f205, %f99;
mul.ftz.f32 %f209, %f207, %f100;
sub.ftz.f32 %f210, %f208, %f209;
add.ftz.f32 %f331, %f210, %f331;
add.s32 %r725, %r725, 32;
add.s32 %r724, %r724, 32;
add.s32 %r723, %r723, 32;
add.s32 %r722, %r722, 32;
add.s32 %r721, %r721, 32;
add.s64 %rd93, %rd93, 576;
add.s32 %r726, %r726, 4;
setp.lt.s32 %p5, %r726, %r6;
@%p5 bra $L__BB164_2;

$L__BB164_3:
setp.eq.s32 %p6, %r1, 0;
@%p6 bra $L__BB164_5;

mov.u32 %r532, _ZZ13mul_mat_vec_qIL9ggml_type12ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r533, %r1, 1280, %r532;
shl.b32 %r534, %r44, 2;
add.s32 %r535, %r533, %r534;
st.shared.f32 [%r535+-1280], %f340;
st.shared.f32 [%r535+-1152], %f339;
st.shared.f32 [%r535+-1024], %f338;
st.shared.f32 [%r535+-896], %f337;
st.shared.f32 [%r535+-768], %f336;
st.shared.f32 [%r535+-640], %f335;
st.shared.f32 [%r535+-512], %f334;
st.shared.f32 [%r535+-384], %f333;
st.shared.f32 [%r535+-256], %f332;
st.shared.f32 [%r535+-128], %f331;

$L__BB164_5:
setp.ne.s32 %p7, %r1, 0;
bar.sync 0;
@%p7 bra $L__BB164_26;

mov.u32 %r536, %ctaid.x;
shl.b32 %r537, %r536, 1;
mov.u32 %r538, 1;
mad.lo.s32 %r539, %r3, %r35, %r537;
mad.lo.s32 %r540, %r4, %r41, %r539;
cvt.u64.u32 %rd13, %r540;
shl.b32 %r542, %r44, 2;
mov.u32 %r543, _ZZ13mul_mat_vec_qIL9ggml_type12ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r25, %r543, %r542;
mov.u32 %r544, 2;
ld.shared.f32 %f211, [%r25];
add.ftz.f32 %f212, %f211, %f340;
mov.b32 %r545, %f212;
mov.u32 %r546, 31;
mov.u32 %r547, 16;
mov.u32 %r548, -1;
shfl.sync.bfly.b32 %r549|%p8, %r545, %r547, %r546, %r548;
mov.b32 %f213, %r549;
add.ftz.f32 %f214, %f212, %f213;
mov.b32 %r550, %f214;
mov.u32 %r551, 8;
shfl.sync.bfly.b32 %r552|%p9, %r550, %r551, %r546, %r548;
mov.b32 %f215, %r552;
add.ftz.f32 %f216, %f214, %f215;
mov.b32 %r553, %f216;
mov.u32 %r554, 4;
shfl.sync.bfly.b32 %r555|%p10, %r553, %r554, %r546, %r548;
mov.b32 %f217, %r555;
add.ftz.f32 %f218, %f216, %f217;
mov.b32 %r556, %f218;
shfl.sync.bfly.b32 %r557|%p11, %r556, %r544, %r546, %r548;
mov.b32 %f219, %r557;
add.ftz.f32 %f220, %f218, %f219;
mov.b32 %r558, %f220;
shfl.sync.bfly.b32 %r559|%p12, %r558, %r538, %r546, %r548;
mov.b32 %f221, %r559;
add.ftz.f32 %f31, %f220, %f221;
setp.eq.s32 %p13, %r44, 0;
setp.lt.u32 %p14, %r537, %r29;
and.pred %p1, %p13, %p14;
cvta.to.global.u64 %rd63, %rd19;
mul.wide.u32 %rd64, %r540, 4;
add.s64 %rd14, %rd63, %rd64;
not.pred %p15, %p1;
@%p15 bra $L__BB164_8;

st.global.f32 [%rd14], %f31;

$L__BB164_8:
ld.shared.f32 %f222, [%r25+128];
add.ftz.f32 %f223, %f222, %f339;
mov.b32 %r560, %f223;
shfl.sync.bfly.b32 %r564|%p16, %r560, %r547, %r546, %r548;
mov.b32 %f224, %r564;
add.ftz.f32 %f225, %f223, %f224;
mov.b32 %r565, %f225;
shfl.sync.bfly.b32 %r567|%p17, %r565, %r551, %r546, %r548;
mov.b32 %f226, %r567;
add.ftz.f32 %f227, %f225, %f226;
mov.b32 %r568, %f227;
shfl.sync.bfly.b32 %r570|%p18, %r568, %r554, %r546, %r548;
mov.b32 %f228, %r570;
add.ftz.f32 %f229, %f227, %f228;
mov.b32 %r571, %f229;
shfl.sync.bfly.b32 %r573|%p19, %r571, %r544, %r546, %r548;
mov.b32 %f230, %r573;
add.ftz.f32 %f231, %f229, %f230;
mov.b32 %r574, %f231;
shfl.sync.bfly.b32 %r576|%p20, %r574, %r538, %r546, %r548;
mov.b32 %f232, %r576;
add.ftz.f32 %f32, %f231, %f232;
setp.eq.s32 %p21, %r44, 1;
or.b32 %r580, %r537, 1;
setp.lt.u32 %p22, %r580, %r29;
and.pred %p2, %p21, %p22;
not.pred %p23, %p2;
@%p23 bra $L__BB164_10;

st.global.f32 [%rd14+4], %f32;

$L__BB164_10:
ld.shared.f32 %f233, [%r25+256];
add.ftz.f32 %f234, %f233, %f338;
mov.b32 %r581, %f234;
mov.u32 %r582, 31;
mov.u32 %r583, 16;
mov.u32 %r584, -1;
shfl.sync.bfly.b32 %r585|%p24, %r581, %r583, %r582, %r584;
mov.b32 %f235, %r585;
add.ftz.f32 %f236, %f234, %f235;
mov.b32 %r586, %f236;
mov.u32 %r587, 8;
shfl.sync.bfly.b32 %r588|%p25, %r586, %r587, %r582, %r584;
mov.b32 %f237, %r588;
add.ftz.f32 %f238, %f236, %f237;
mov.b32 %r589, %f238;
mov.u32 %r590, 4;
shfl.sync.bfly.b32 %r591|%p26, %r589, %r590, %r582, %r584;
mov.b32 %f239, %r591;
add.ftz.f32 %f240, %f238, %f239;
mov.b32 %r592, %f240;
mov.u32 %r593, 2;
shfl.sync.bfly.b32 %r594|%p27, %r592, %r593, %r582, %r584;
mov.b32 %f241, %r594;
add.ftz.f32 %f242, %f240, %f241;
mov.b32 %r595, %f242;
mov.u32 %r596, 1;
shfl.sync.bfly.b32 %r597|%p28, %r595, %r596, %r582, %r584;
mov.b32 %f243, %r597;
add.ftz.f32 %f33, %f242, %f243;
@%p15 bra $L__BB164_12;

cvt.u64.u32 %rd65, %r29;
add.s64 %rd66, %rd65, %rd13;
shl.b64 %rd68, %rd66, 2;
add.s64 %rd69, %rd63, %rd68;
st.global.f32 [%rd69], %f33;

$L__BB164_12:
ld.shared.f32 %f244, [%r25+384];
add.ftz.f32 %f245, %f244, %f337;
mov.b32 %r598, %f245;
shfl.sync.bfly.b32 %r602|%p30, %r598, %r583, %r582, %r584;
mov.b32 %f246, %r602;
add.ftz.f32 %f247, %f245, %f246;
mov.b32 %r603, %f247;
shfl.sync.bfly.b32 %r605|%p31, %r603, %r587, %r582, %r584;
mov.b32 %f248, %r605;
add.ftz.f32 %f249, %f247, %f248;
mov.b32 %r606, %f249;
shfl.sync.bfly.b32 %r608|%p32, %r606, %r590, %r582, %r584;
mov.b32 %f250, %r608;
add.ftz.f32 %f251, %f249, %f250;
mov.b32 %r609, %f251;
shfl.sync.bfly.b32 %r611|%p33, %r609, %r593, %r582, %r584;
mov.b32 %f252, %r611;
add.ftz.f32 %f253, %f251, %f252;
mov.b32 %r612, %f253;
shfl.sync.bfly.b32 %r614|%p34, %r612, %r596, %r582, %r584;
mov.b32 %f254, %r614;
add.ftz.f32 %f34, %f253, %f254;
@%p23 bra $L__BB164_14;

add.s32 %r615, %r29, 1;
cvt.u64.u32 %rd70, %r615;
add.s64 %rd71, %rd70, %rd13;
shl.b64 %rd73, %rd71, 2;
add.s64 %rd74, %rd63, %rd73;
st.global.f32 [%rd74], %f34;

$L__BB164_14:
shl.b32 %r616, %r29, 1;
mov.u32 %r617, 1;
ld.shared.f32 %f255, [%r25+512];
add.ftz.f32 %f256, %f255, %f336;
mov.b32 %r618, %f256;
mov.u32 %r619, 31;
mov.u32 %r620, 16;
mov.u32 %r621, -1;
shfl.sync.bfly.b32 %r622|%p36, %r618, %r620, %r619, %r621;
mov.b32 %f257, %r622;
add.ftz.f32 %f258, %f256, %f257;
mov.b32 %r623, %f258;
mov.u32 %r624, 8;
shfl.sync.bfly.b32 %r625|%p37, %r623, %r624, %r619, %r621;
mov.b32 %f259, %r625;
add.ftz.f32 %f260, %f258, %f259;
mov.b32 %r626, %f260;
mov.u32 %r627, 4;
shfl.sync.bfly.b32 %r628|%p38, %r626, %r627, %r619, %r621;
mov.b32 %f261, %r628;
add.ftz.f32 %f262, %f260, %f261;
mov.b32 %r629, %f262;
mov.u32 %r630, 2;
shfl.sync.bfly.b32 %r631|%p39, %r629, %r630, %r619, %r621;
mov.b32 %f263, %r631;
add.ftz.f32 %f264, %f262, %f263;
mov.b32 %r632, %f264;
shfl.sync.bfly.b32 %r633|%p40, %r632, %r617, %r619, %r621;
mov.b32 %f265, %r633;
add.ftz.f32 %f35, %f264, %f265;
cvt.u64.u32 %rd75, %r616;
add.s64 %rd76, %rd75, %rd13;
shl.b64 %rd78, %rd76, 2;
add.s64 %rd15, %rd63, %rd78;
@%p15 bra $L__BB164_16;

st.global.f32 [%rd15], %f35;

$L__BB164_16:
ld.shared.f32 %f266, [%r25+640];
add.ftz.f32 %f267, %f266, %f335;
mov.b32 %r634, %f267;
shfl.sync.bfly.b32 %r638|%p42, %r634, %r620, %r619, %r621;
mov.b32 %f268, %r638;
add.ftz.f32 %f269, %f267, %f268;
mov.b32 %r639, %f269;
shfl.sync.bfly.b32 %r641|%p43, %r639, %r624, %r619, %r621;
mov.b32 %f270, %r641;
add.ftz.f32 %f271, %f269, %f270;
mov.b32 %r642, %f271;
shfl.sync.bfly.b32 %r644|%p44, %r642, %r627, %r619, %r621;
mov.b32 %f272, %r644;
add.ftz.f32 %f273, %f271, %f272;
mov.b32 %r645, %f273;
shfl.sync.bfly.b32 %r647|%p45, %r645, %r630, %r619, %r621;
mov.b32 %f274, %r647;
add.ftz.f32 %f275, %f273, %f274;
mov.b32 %r648, %f275;
shfl.sync.bfly.b32 %r650|%p46, %r648, %r617, %r619, %r621;
mov.b32 %f276, %r650;
add.ftz.f32 %f36, %f275, %f276;
@%p23 bra $L__BB164_18;

st.global.f32 [%rd15+4], %f36;

$L__BB164_18:
mul.lo.s32 %r26, %r29, 3;
ld.shared.f32 %f277, [%r25+768];
add.ftz.f32 %f278, %f277, %f334;
mov.b32 %r651, %f278;
mov.u32 %r652, 31;
mov.u32 %r653, 16;
mov.u32 %r654, -1;
shfl.sync.bfly.b32 %r655|%p48, %r651, %r653, %r652, %r654;
mov.b32 %f279, %r655;
add.ftz.f32 %f280, %f278, %f279;
mov.b32 %r656, %f280;
mov.u32 %r657, 8;
shfl.sync.bfly.b32 %r658|%p49, %r656, %r657, %r652, %r654;
mov.b32 %f281, %r658;
add.ftz.f32 %f282, %f280, %f281;
mov.b32 %r659, %f282;
mov.u32 %r660, 4;
shfl.sync.bfly.b32 %r661|%p50, %r659, %r660, %r652, %r654;
mov.b32 %f283, %r661;
add.ftz.f32 %f284, %f282, %f283;
mov.b32 %r662, %f284;
mov.u32 %r663, 2;
shfl.sync.bfly.b32 %r664|%p51, %r662, %r663, %r652, %r654;
mov.b32 %f285, %r664;
add.ftz.f32 %f286, %f284, %f285;
mov.b32 %r665, %f286;
mov.u32 %r666, 1;
shfl.sync.bfly.b32 %r667|%p52, %r665, %r666, %r652, %r654;
mov.b32 %f287, %r667;
add.ftz.f32 %f37, %f286, %f287;
@%p15 bra $L__BB164_20;

cvt.u64.u32 %rd79, %r26;
add.s64 %rd80, %rd79, %rd13;
shl.b64 %rd82, %rd80, 2;
add.s64 %rd83, %rd63, %rd82;
st.global.f32 [%rd83], %f37;

$L__BB164_20:
ld.shared.f32 %f288, [%r25+896];
add.ftz.f32 %f289, %f288, %f333;
mov.b32 %r668, %f289;
shfl.sync.bfly.b32 %r672|%p54, %r668, %r653, %r652, %r654;
mov.b32 %f290, %r672;
add.ftz.f32 %f291, %f289, %f290;
mov.b32 %r673, %f291;
shfl.sync.bfly.b32 %r675|%p55, %r673, %r657, %r652, %r654;
mov.b32 %f292, %r675;
add.ftz.f32 %f293, %f291, %f292;
mov.b32 %r676, %f293;
shfl.sync.bfly.b32 %r678|%p56, %r676, %r660, %r652, %r654;
mov.b32 %f294, %r678;
add.ftz.f32 %f295, %f293, %f294;
mov.b32 %r679, %f295;
shfl.sync.bfly.b32 %r681|%p57, %r679, %r663, %r652, %r654;
mov.b32 %f296, %r681;
add.ftz.f32 %f297, %f295, %f296;
mov.b32 %r682, %f297;
shfl.sync.bfly.b32 %r684|%p58, %r682, %r666, %r652, %r654;
mov.b32 %f298, %r684;
add.ftz.f32 %f38, %f297, %f298;
@%p23 bra $L__BB164_22;

add.s32 %r685, %r26, 1;
cvt.u64.u32 %rd84, %r685;
add.s64 %rd85, %rd84, %rd13;
shl.b64 %rd87, %rd85, 2;
add.s64 %rd88, %rd63, %rd87;
st.global.f32 [%rd88], %f38;

$L__BB164_22:
shl.b32 %r686, %r29, 2;
mov.u32 %r687, 2;
ld.shared.f32 %f299, [%r25+1024];
add.ftz.f32 %f300, %f299, %f332;
mov.b32 %r688, %f300;
mov.u32 %r689, 31;
mov.u32 %r690, 16;
mov.u32 %r691, -1;
shfl.sync.bfly.b32 %r692|%p60, %r688, %r690, %r689, %r691;
mov.b32 %f301, %r692;
add.ftz.f32 %f302, %f300, %f301;
mov.b32 %r693, %f302;
mov.u32 %r694, 8;
shfl.sync.bfly.b32 %r695|%p61, %r693, %r694, %r689, %r691;
mov.b32 %f303, %r695;
add.ftz.f32 %f304, %f302, %f303;
mov.b32 %r696, %f304;
mov.u32 %r697, 4;
shfl.sync.bfly.b32 %r698|%p62, %r696, %r697, %r689, %r691;
mov.b32 %f305, %r698;
add.ftz.f32 %f306, %f304, %f305;
mov.b32 %r699, %f306;
shfl.sync.bfly.b32 %r700|%p63, %r699, %r687, %r689, %r691;
mov.b32 %f307, %r700;
add.ftz.f32 %f308, %f306, %f307;
mov.b32 %r701, %f308;
mov.u32 %r702, 1;
shfl.sync.bfly.b32 %r703|%p64, %r701, %r702, %r689, %r691;
mov.b32 %f309, %r703;
add.ftz.f32 %f39, %f308, %f309;
cvt.u64.u32 %rd89, %r686;
add.s64 %rd90, %rd89, %rd13;
shl.b64 %rd92, %rd90, 2;
add.s64 %rd16, %rd63, %rd92;
@%p15 bra $L__BB164_24;

st.global.f32 [%rd16], %f39;

$L__BB164_24:
ld.shared.f32 %f310, [%r25+1152];
add.ftz.f32 %f311, %f310, %f331;
mov.b32 %r704, %f311;
shfl.sync.bfly.b32 %r708|%p66, %r704, %r690, %r689, %r691;
mov.b32 %f312, %r708;
add.ftz.f32 %f313, %f311, %f312;
mov.b32 %r709, %f313;
shfl.sync.bfly.b32 %r711|%p67, %r709, %r694, %r689, %r691;
mov.b32 %f314, %r711;
add.ftz.f32 %f315, %f313, %f314;
mov.b32 %r712, %f315;
shfl.sync.bfly.b32 %r714|%p68, %r712, %r697, %r689, %r691;
mov.b32 %f316, %r714;
add.ftz.f32 %f317, %f315, %f316;
mov.b32 %r715, %f317;
shfl.sync.bfly.b32 %r717|%p69, %r715, %r687, %r689, %r691;
mov.b32 %f318, %r717;
add.ftz.f32 %f319, %f317, %f318;
mov.b32 %r718, %f319;
shfl.sync.bfly.b32 %r720|%p70, %r718, %r702, %r689, %r691;
mov.b32 %f320, %r720;
add.ftz.f32 %f40, %f319, %f320;
@%p23 bra $L__BB164_26;

st.global.f32 [%rd16+4], %f40;

$L__BB164_26:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type12ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type12ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type12ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type12ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type12ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type12ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type12ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type12ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type12ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<84>;
.reg .b16 %rs<15>;
.reg .f32 %f<409>;
.reg .b32 %r<861>;
.reg .b64 %rd<112>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type12ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type12ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type12ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type12ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type12ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type12ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type12ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type12ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type12ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type12ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type12ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type12ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type12ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type12ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type12ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type12ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type12ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type12ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r45, %r1, 5;
mov.u32 %r46, %tid.x;
add.s32 %r2, %r45, %r46;
shr.s32 %r47, %r2, 31;
shr.u32 %r48, %r47, 28;
add.s32 %r49, %r2, %r48;
shr.s32 %r860, %r49, 4;
shr.u32 %r4, %r44, 8;
setp.ge.s32 %p3, %r860, %r4;
mov.f32 %f397, 0f00000000;
mov.f32 %f398, %f397;
mov.f32 %f399, %f397;
mov.f32 %f400, %f397;
mov.f32 %f401, %f397;
mov.f32 %f402, %f397;
mov.f32 %f403, %f397;
mov.f32 %f404, %f397;
mov.f32 %f405, %f397;
mov.f32 %f406, %f397;
mov.f32 %f407, %f397;
mov.f32 %f408, %f397;
@%p3 bra $L__BB165_3;

mov.u32 %r50, %ctaid.x;
shl.b32 %r51, %r50, 1;
mov.u32 %r52, %ctaid.y;
mov.u32 %r53, %ctaid.z;
mul.hi.u32 %r54, %r53, %r38;
add.s32 %r55, %r54, %r53;
shr.u32 %r56, %r55, %r39;
mul.lo.s32 %r57, %r56, %r41;
mul.hi.u32 %r58, %r52, %r32;
add.s32 %r59, %r58, %r52;
shr.u32 %r60, %r59, %r33;
mul.lo.s32 %r61, %r60, %r35;
mul.lo.s32 %r62, %r52, %r36;
cvt.u64.u32 %rd19, %r62;
cvta.to.global.u64 %rd111, %rd16;
and.b32 %r66, %r49, -16;
sub.s32 %r5, %r2, %r66;
shr.s32 %r67, %r5, 31;
shr.u32 %r68, %r67, 30;
add.s32 %r69, %r5, %r68;
shr.s32 %r70, %r69, 2;
shl.b32 %r71, %r70, 1;
shl.b32 %r72, %r70, 5;
and.b32 %r73, %r69, -4;
sub.s32 %r74, %r5, %r73;
shl.b32 %r75, %r74, 2;
add.s32 %r76, %r72, %r75;
cvt.s64.s32 %rd20, %r76;
cvt.s64.s32 %rd21, %r71;
add.s64 %rd22, %rd19, %rd21;
mul.lo.s32 %r77, %r53, %r42;
cvt.u64.u32 %rd23, %r77;
add.s64 %rd2, %rd22, %rd23;
cvt.s64.s32 %rd3, %r74;
shl.b32 %r78, %r30, 1;
shl.b32 %r854, %r860, 3;
add.s32 %r859, %r854, %r78;
mad.lo.s32 %r858, %r30, 3, %r854;
shl.b32 %r79, %r30, 2;
add.s32 %r857, %r854, %r79;
mad.lo.s32 %r856, %r30, 5, %r854;
add.s32 %r855, %r30, %r854;
add.s32 %r80, %r860, %r57;
add.s32 %r81, %r80, %r61;
or.b32 %r82, %r51, 1;
mad.lo.s32 %r83, %r29, %r82, %r81;
mul.wide.s32 %rd4, %r83, 144;
mad.lo.s32 %r84, %r51, %r29, %r860;
add.s32 %r85, %r84, %r57;
add.s32 %r86, %r85, %r61;
mul.wide.s32 %rd5, %r86, 144;
add.s64 %rd24, %rd5, %rd20;
add.s64 %rd6, %rd24, 32;
cvt.u64.u32 %rd25, %r71;
and.b64 %rd26, %rd25, 2;
or.b64 %rd27, %rd26, %rd4;
or.b64 %rd7, %rd27, 8;
or.b64 %rd28, %rd26, %rd5;
or.b64 %rd8, %rd28, 12;
add.s64 %rd29, %rd4, %rd20;
add.s64 %rd9, %rd29, 16;
cvta.to.global.u64 %rd10, %rd17;
mov.f32 %f397, 0f00000000;
shl.b64 %rd36, %rd3, 2;

$L__BB165_2:
cvt.u64.u32 %rd30, %r854;
add.s64 %rd31, %rd2, %rd30;
add.s64 %rd32, %rd111, %rd6;
ld.global.nc.u32 %r519, [%rd32+-16];
ld.global.nc.u32 %r520, [%rd32];
add.s64 %rd33, %rd111, %rd8;
ld.global.nc.u16 %rs1, [%rd33+-8];
cvt.u32.u16 %r521, %rs1;
ld.global.nc.u16 %rs2, [%rd33+-4];
cvt.u32.u16 %r522, %rs2;
ld.global.nc.u16 %rs3, [%rd33];
cvt.u32.u16 %r523, %rs3;
and.b32 %r524, %r521, 16191;
and.b32 %r525, %r523, 3855;
and.b32 %r526, %r521, 49344;
shr.u32 %r527, %r526, 2;
or.b32 %r528, %r525, %r527;
setp.gt.s32 %p4, %r5, 7;
and.b32 %r529, %r522, 16191;
shr.u32 %r530, %r523, 4;
and.b32 %r531, %r530, 3855;
and.b32 %r532, %r522, 49344;
shr.u32 %r533, %r532, 2;
or.b32 %r534, %r531, %r533;
mul.lo.s64 %rd34, %rd31, 36;
add.s64 %rd35, %rd10, %rd34;
ld.global.nc.u32 %r87, [%rd35];

{.reg .f16 low,high;
mov.b32 {low,high},%r87;
cvt.f32.f16 %f73, low;}


add.s64 %rd37, %rd35, %rd36;
ld.global.nc.u32 %r91, [%rd37+4];
ld.global.nc.u32 %r95, [%rd37+20];
ld.global.nc.u32 %r88, [%rd35+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r88;
cvt.f32.f16 %f74, low;}


ld.global.nc.u32 %r107, [%rd37+40];
ld.global.nc.u32 %r111, [%rd37+56];
add.s64 %rd38, %rd111, %rd5;
and.b32 %r450, %r519, 252645135;
and.b32 %r454, %r520, 252645135;
mov.u32 %r512, 0;

dp4a.s32.s32 %r89, %r450, %r91, %r512;


dp4a.s32.s32 %r93, %r454, %r95, %r89;

mov.u32 %r514, 16843009;

dp4a.s32.s32 %r97, %r514, %r91, %r512;


dp4a.s32.s32 %r101, %r514, %r95, %r97;

selp.b32 %r535, %r528, %r524, %p4;
cvt.u16.u32 %rs4, %r535;
and.b32 %r536, %r535, 63;
mul.lo.s32 %r537, %r93, %r536;
cvt.rn.f32.s32 %f121, %r537;
fma.rn.ftz.f32 %f122, %f73, %f121, 0f00000000;
selp.b32 %r538, %r534, %r529, %p4;
cvt.u16.u32 %rs5, %r538;
and.b32 %r539, %r538, 63;
mul.lo.s32 %r540, %r101, %r539;
cvt.rn.f32.s32 %f123, %r540;
fma.rn.ftz.f32 %f124, %f73, %f123, 0f00000000;
shr.u32 %r541, %r519, 4;
and.b32 %r466, %r541, 252645135;
shr.u32 %r542, %r520, 4;
and.b32 %r470, %r542, 252645135;

dp4a.s32.s32 %r105, %r466, %r107, %r512;


dp4a.s32.s32 %r109, %r470, %r111, %r105;


dp4a.s32.s32 %r113, %r514, %r107, %r512;


dp4a.s32.s32 %r117, %r514, %r111, %r113;

shr.u16 %rs6, %rs4, 8;
cvt.u32.u16 %r543, %rs6;
mul.lo.s32 %r544, %r109, %r543;
cvt.rn.f32.s32 %f125, %r544;
fma.rn.ftz.f32 %f126, %f74, %f125, %f122;
shr.u16 %rs7, %rs5, 8;
cvt.u32.u16 %r545, %rs7;
mul.lo.s32 %r546, %r117, %r545;
cvt.rn.f32.s32 %f127, %r546;
fma.rn.ftz.f32 %f128, %f74, %f127, %f124;
ld.global.nc.u32 %r121, [%rd38];

{.reg .f16 low,high;
mov.b32 {low,high},%r121;
cvt.f32.f16 %f75, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r121;
cvt.f32.f16 %f76, high;}


mul.ftz.f32 %f129, %f126, %f75;
mul.ftz.f32 %f130, %f128, %f76;
sub.ftz.f32 %f131, %f129, %f130;
add.ftz.f32 %f408, %f131, %f408;
add.s64 %rd39, %rd111, %rd9;
ld.global.nc.u32 %r547, [%rd39];
ld.global.nc.u32 %r548, [%rd39+16];
add.s64 %rd40, %rd111, %rd7;
ld.global.nc.u16 %rs8, [%rd40+-4];
cvt.u32.u16 %r549, %rs8;
ld.global.nc.u16 %rs9, [%rd40];
cvt.u32.u16 %r550, %rs9;
ld.global.nc.u16 %rs10, [%rd40+4];
cvt.u32.u16 %r551, %rs10;
and.b32 %r552, %r549, 16191;
and.b32 %r553, %r551, 3855;
and.b32 %r554, %r549, 49344;
shr.u32 %r555, %r554, 2;
or.b32 %r556, %r553, %r555;
and.b32 %r557, %r550, 16191;
shr.u32 %r558, %r551, 4;
and.b32 %r559, %r558, 3855;
and.b32 %r560, %r550, 49344;
shr.u32 %r561, %r560, 2;
or.b32 %r562, %r559, %r561;

{.reg .f16 low,high;
mov.b32 {low,high},%r87;
cvt.f32.f16 %f77, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r88;
cvt.f32.f16 %f78, low;}


add.s64 %rd41, %rd111, %rd4;
and.b32 %r486, %r547, 252645135;
and.b32 %r490, %r548, 252645135;

dp4a.s32.s32 %r125, %r486, %r91, %r512;


dp4a.s32.s32 %r129, %r490, %r95, %r125;


dp4a.s32.s32 %r133, %r514, %r91, %r512;


dp4a.s32.s32 %r137, %r514, %r95, %r133;

selp.b32 %r563, %r556, %r552, %p4;
cvt.u16.u32 %rs11, %r563;
and.b32 %r564, %r563, 63;
mul.lo.s32 %r565, %r129, %r564;
cvt.rn.f32.s32 %f132, %r565;
fma.rn.ftz.f32 %f133, %f77, %f132, 0f00000000;
selp.b32 %r566, %r562, %r557, %p4;
cvt.u16.u32 %rs12, %r566;
and.b32 %r567, %r566, 63;
mul.lo.s32 %r568, %r137, %r567;
cvt.rn.f32.s32 %f134, %r568;
fma.rn.ftz.f32 %f135, %f77, %f134, 0f00000000;
shr.u32 %r569, %r547, 4;
and.b32 %r502, %r569, 252645135;
shr.u32 %r570, %r548, 4;
and.b32 %r506, %r570, 252645135;

dp4a.s32.s32 %r141, %r502, %r107, %r512;


dp4a.s32.s32 %r145, %r506, %r111, %r141;


dp4a.s32.s32 %r149, %r514, %r107, %r512;


dp4a.s32.s32 %r153, %r514, %r111, %r149;

shr.u16 %rs13, %rs11, 8;
cvt.u32.u16 %r571, %rs13;
mul.lo.s32 %r572, %r145, %r571;
cvt.rn.f32.s32 %f136, %r572;
fma.rn.ftz.f32 %f137, %f78, %f136, %f133;
shr.u16 %rs14, %rs12, 8;
cvt.u32.u16 %r573, %rs14;
mul.lo.s32 %r574, %r153, %r573;
cvt.rn.f32.s32 %f138, %r574;
fma.rn.ftz.f32 %f139, %f78, %f138, %f135;
ld.global.nc.u32 %r157, [%rd41];

{.reg .f16 low,high;
mov.b32 {low,high},%r157;
cvt.f32.f16 %f79, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r157;
cvt.f32.f16 %f80, high;}


mul.ftz.f32 %f140, %f137, %f79;
mul.ftz.f32 %f141, %f139, %f80;
sub.ftz.f32 %f142, %f140, %f141;
add.ftz.f32 %f407, %f142, %f407;
cvt.u64.u32 %rd42, %r855;
add.s64 %rd43, %rd2, %rd42;
mul.lo.s64 %rd44, %rd43, 36;
add.s64 %rd45, %rd10, %rd44;
ld.global.nc.u32 %r159, [%rd45];

{.reg .f16 low,high;
mov.b32 {low,high},%r159;
cvt.f32.f16 %f81, low;}


add.s64 %rd46, %rd45, %rd36;
ld.global.nc.u32 %r163, [%rd46+4];
ld.global.nc.u32 %r167, [%rd46+20];
ld.global.nc.u32 %r160, [%rd45+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r160;
cvt.f32.f16 %f82, low;}


ld.global.nc.u32 %r179, [%rd46+40];
ld.global.nc.u32 %r183, [%rd46+56];

dp4a.s32.s32 %r161, %r450, %r163, %r512;


dp4a.s32.s32 %r165, %r454, %r167, %r161;


dp4a.s32.s32 %r169, %r514, %r163, %r512;


dp4a.s32.s32 %r173, %r514, %r167, %r169;

mul.lo.s32 %r575, %r165, %r536;
cvt.rn.f32.s32 %f143, %r575;
fma.rn.ftz.f32 %f144, %f81, %f143, 0f00000000;
mul.lo.s32 %r576, %r173, %r539;
cvt.rn.f32.s32 %f145, %r576;
fma.rn.ftz.f32 %f146, %f81, %f145, 0f00000000;

dp4a.s32.s32 %r177, %r466, %r179, %r512;


dp4a.s32.s32 %r181, %r470, %r183, %r177;


dp4a.s32.s32 %r185, %r514, %r179, %r512;


dp4a.s32.s32 %r189, %r514, %r183, %r185;

mul.lo.s32 %r577, %r181, %r543;
cvt.rn.f32.s32 %f147, %r577;
fma.rn.ftz.f32 %f148, %f82, %f147, %f144;
mul.lo.s32 %r578, %r189, %r545;
cvt.rn.f32.s32 %f149, %r578;
fma.rn.ftz.f32 %f150, %f82, %f149, %f146;

{.reg .f16 low,high;
mov.b32 {low,high},%r121;
cvt.f32.f16 %f83, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r121;
cvt.f32.f16 %f84, high;}


mul.ftz.f32 %f151, %f148, %f83;
mul.ftz.f32 %f152, %f150, %f84;
sub.ftz.f32 %f153, %f151, %f152;
add.ftz.f32 %f406, %f153, %f406;

{.reg .f16 low,high;
mov.b32 {low,high},%r159;
cvt.f32.f16 %f85, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r160;
cvt.f32.f16 %f86, low;}



dp4a.s32.s32 %r197, %r486, %r163, %r512;


dp4a.s32.s32 %r201, %r490, %r167, %r197;


dp4a.s32.s32 %r205, %r514, %r163, %r512;


dp4a.s32.s32 %r209, %r514, %r167, %r205;

mul.lo.s32 %r579, %r201, %r564;
cvt.rn.f32.s32 %f154, %r579;
fma.rn.ftz.f32 %f155, %f85, %f154, 0f00000000;
mul.lo.s32 %r580, %r209, %r567;
cvt.rn.f32.s32 %f156, %r580;
fma.rn.ftz.f32 %f157, %f85, %f156, 0f00000000;

dp4a.s32.s32 %r213, %r502, %r179, %r512;


dp4a.s32.s32 %r217, %r506, %r183, %r213;


dp4a.s32.s32 %r221, %r514, %r179, %r512;


dp4a.s32.s32 %r225, %r514, %r183, %r221;

mul.lo.s32 %r581, %r217, %r571;
cvt.rn.f32.s32 %f158, %r581;
fma.rn.ftz.f32 %f159, %f86, %f158, %f155;
mul.lo.s32 %r582, %r225, %r573;
cvt.rn.f32.s32 %f160, %r582;
fma.rn.ftz.f32 %f161, %f86, %f160, %f157;

{.reg .f16 low,high;
mov.b32 {low,high},%r157;
cvt.f32.f16 %f87, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r157;
cvt.f32.f16 %f88, high;}


mul.ftz.f32 %f162, %f159, %f87;
mul.ftz.f32 %f163, %f161, %f88;
sub.ftz.f32 %f164, %f162, %f163;
add.ftz.f32 %f405, %f164, %f405;
cvt.u64.u32 %rd47, %r859;
add.s64 %rd48, %rd2, %rd47;
mul.lo.s64 %rd49, %rd48, 36;
add.s64 %rd50, %rd10, %rd49;
ld.global.nc.u32 %r231, [%rd50];

{.reg .f16 low,high;
mov.b32 {low,high},%r231;
cvt.f32.f16 %f89, low;}


add.s64 %rd51, %rd50, %rd36;
ld.global.nc.u32 %r235, [%rd51+4];
ld.global.nc.u32 %r239, [%rd51+20];
ld.global.nc.u32 %r232, [%rd50+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r232;
cvt.f32.f16 %f90, low;}


ld.global.nc.u32 %r251, [%rd51+40];
ld.global.nc.u32 %r255, [%rd51+56];

dp4a.s32.s32 %r233, %r450, %r235, %r512;


dp4a.s32.s32 %r237, %r454, %r239, %r233;


dp4a.s32.s32 %r241, %r514, %r235, %r512;


dp4a.s32.s32 %r245, %r514, %r239, %r241;

mul.lo.s32 %r583, %r237, %r536;
cvt.rn.f32.s32 %f165, %r583;
fma.rn.ftz.f32 %f166, %f89, %f165, 0f00000000;
mul.lo.s32 %r584, %r245, %r539;
cvt.rn.f32.s32 %f167, %r584;
fma.rn.ftz.f32 %f168, %f89, %f167, 0f00000000;

dp4a.s32.s32 %r249, %r466, %r251, %r512;


dp4a.s32.s32 %r253, %r470, %r255, %r249;


dp4a.s32.s32 %r257, %r514, %r251, %r512;


dp4a.s32.s32 %r261, %r514, %r255, %r257;

mul.lo.s32 %r585, %r253, %r543;
cvt.rn.f32.s32 %f169, %r585;
fma.rn.ftz.f32 %f170, %f90, %f169, %f166;
mul.lo.s32 %r586, %r261, %r545;
cvt.rn.f32.s32 %f171, %r586;
fma.rn.ftz.f32 %f172, %f90, %f171, %f168;

{.reg .f16 low,high;
mov.b32 {low,high},%r121;
cvt.f32.f16 %f91, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r121;
cvt.f32.f16 %f92, high;}


mul.ftz.f32 %f173, %f170, %f91;
mul.ftz.f32 %f174, %f172, %f92;
sub.ftz.f32 %f175, %f173, %f174;
add.ftz.f32 %f404, %f175, %f404;

{.reg .f16 low,high;
mov.b32 {low,high},%r231;
cvt.f32.f16 %f93, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r232;
cvt.f32.f16 %f94, low;}



dp4a.s32.s32 %r269, %r486, %r235, %r512;


dp4a.s32.s32 %r273, %r490, %r239, %r269;


dp4a.s32.s32 %r277, %r514, %r235, %r512;


dp4a.s32.s32 %r281, %r514, %r239, %r277;

mul.lo.s32 %r587, %r273, %r564;
cvt.rn.f32.s32 %f176, %r587;
fma.rn.ftz.f32 %f177, %f93, %f176, 0f00000000;
mul.lo.s32 %r588, %r281, %r567;
cvt.rn.f32.s32 %f178, %r588;
fma.rn.ftz.f32 %f179, %f93, %f178, 0f00000000;

dp4a.s32.s32 %r285, %r502, %r251, %r512;


dp4a.s32.s32 %r289, %r506, %r255, %r285;


dp4a.s32.s32 %r293, %r514, %r251, %r512;


dp4a.s32.s32 %r297, %r514, %r255, %r293;

mul.lo.s32 %r589, %r289, %r571;
cvt.rn.f32.s32 %f180, %r589;
fma.rn.ftz.f32 %f181, %f94, %f180, %f177;
mul.lo.s32 %r590, %r297, %r573;
cvt.rn.f32.s32 %f182, %r590;
fma.rn.ftz.f32 %f183, %f94, %f182, %f179;

{.reg .f16 low,high;
mov.b32 {low,high},%r157;
cvt.f32.f16 %f95, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r157;
cvt.f32.f16 %f96, high;}


mul.ftz.f32 %f184, %f181, %f95;
mul.ftz.f32 %f185, %f183, %f96;
sub.ftz.f32 %f186, %f184, %f185;
add.ftz.f32 %f403, %f186, %f403;
cvt.u64.u32 %rd52, %r858;
add.s64 %rd53, %rd2, %rd52;
mul.lo.s64 %rd54, %rd53, 36;
add.s64 %rd55, %rd10, %rd54;
ld.global.nc.u32 %r303, [%rd55];

{.reg .f16 low,high;
mov.b32 {low,high},%r303;
cvt.f32.f16 %f97, low;}


add.s64 %rd56, %rd55, %rd36;
ld.global.nc.u32 %r307, [%rd56+4];
ld.global.nc.u32 %r311, [%rd56+20];
ld.global.nc.u32 %r304, [%rd55+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r304;
cvt.f32.f16 %f98, low;}


ld.global.nc.u32 %r323, [%rd56+40];
ld.global.nc.u32 %r327, [%rd56+56];

dp4a.s32.s32 %r305, %r450, %r307, %r512;


dp4a.s32.s32 %r309, %r454, %r311, %r305;


dp4a.s32.s32 %r313, %r514, %r307, %r512;


dp4a.s32.s32 %r317, %r514, %r311, %r313;

mul.lo.s32 %r591, %r309, %r536;
cvt.rn.f32.s32 %f187, %r591;
fma.rn.ftz.f32 %f188, %f97, %f187, 0f00000000;
mul.lo.s32 %r592, %r317, %r539;
cvt.rn.f32.s32 %f189, %r592;
fma.rn.ftz.f32 %f190, %f97, %f189, 0f00000000;

dp4a.s32.s32 %r321, %r466, %r323, %r512;


dp4a.s32.s32 %r325, %r470, %r327, %r321;


dp4a.s32.s32 %r329, %r514, %r323, %r512;


dp4a.s32.s32 %r333, %r514, %r327, %r329;

mul.lo.s32 %r593, %r325, %r543;
cvt.rn.f32.s32 %f191, %r593;
fma.rn.ftz.f32 %f192, %f98, %f191, %f188;
mul.lo.s32 %r594, %r333, %r545;
cvt.rn.f32.s32 %f193, %r594;
fma.rn.ftz.f32 %f194, %f98, %f193, %f190;

{.reg .f16 low,high;
mov.b32 {low,high},%r121;
cvt.f32.f16 %f99, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r121;
cvt.f32.f16 %f100, high;}


mul.ftz.f32 %f195, %f192, %f99;
mul.ftz.f32 %f196, %f194, %f100;
sub.ftz.f32 %f197, %f195, %f196;
add.ftz.f32 %f402, %f197, %f402;

{.reg .f16 low,high;
mov.b32 {low,high},%r303;
cvt.f32.f16 %f101, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r304;
cvt.f32.f16 %f102, low;}



dp4a.s32.s32 %r341, %r486, %r307, %r512;


dp4a.s32.s32 %r345, %r490, %r311, %r341;


dp4a.s32.s32 %r349, %r514, %r307, %r512;


dp4a.s32.s32 %r353, %r514, %r311, %r349;

mul.lo.s32 %r595, %r345, %r564;
cvt.rn.f32.s32 %f198, %r595;
fma.rn.ftz.f32 %f199, %f101, %f198, 0f00000000;
mul.lo.s32 %r596, %r353, %r567;
cvt.rn.f32.s32 %f200, %r596;
fma.rn.ftz.f32 %f201, %f101, %f200, 0f00000000;

dp4a.s32.s32 %r357, %r502, %r323, %r512;


dp4a.s32.s32 %r361, %r506, %r327, %r357;


dp4a.s32.s32 %r365, %r514, %r323, %r512;


dp4a.s32.s32 %r369, %r514, %r327, %r365;

mul.lo.s32 %r597, %r361, %r571;
cvt.rn.f32.s32 %f202, %r597;
fma.rn.ftz.f32 %f203, %f102, %f202, %f199;
mul.lo.s32 %r598, %r369, %r573;
cvt.rn.f32.s32 %f204, %r598;
fma.rn.ftz.f32 %f205, %f102, %f204, %f201;

{.reg .f16 low,high;
mov.b32 {low,high},%r157;
cvt.f32.f16 %f103, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r157;
cvt.f32.f16 %f104, high;}


mul.ftz.f32 %f206, %f203, %f103;
mul.ftz.f32 %f207, %f205, %f104;
sub.ftz.f32 %f208, %f206, %f207;
add.ftz.f32 %f401, %f208, %f401;
cvt.u64.u32 %rd57, %r857;
add.s64 %rd58, %rd2, %rd57;
mul.lo.s64 %rd59, %rd58, 36;
add.s64 %rd60, %rd10, %rd59;
ld.global.nc.u32 %r375, [%rd60];

{.reg .f16 low,high;
mov.b32 {low,high},%r375;
cvt.f32.f16 %f105, low;}


add.s64 %rd61, %rd60, %rd36;
ld.global.nc.u32 %r379, [%rd61+4];
ld.global.nc.u32 %r383, [%rd61+20];
ld.global.nc.u32 %r376, [%rd60+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r376;
cvt.f32.f16 %f106, low;}


ld.global.nc.u32 %r395, [%rd61+40];
ld.global.nc.u32 %r399, [%rd61+56];

dp4a.s32.s32 %r377, %r450, %r379, %r512;


dp4a.s32.s32 %r381, %r454, %r383, %r377;


dp4a.s32.s32 %r385, %r514, %r379, %r512;


dp4a.s32.s32 %r389, %r514, %r383, %r385;

mul.lo.s32 %r599, %r381, %r536;
cvt.rn.f32.s32 %f209, %r599;
fma.rn.ftz.f32 %f210, %f105, %f209, 0f00000000;
mul.lo.s32 %r600, %r389, %r539;
cvt.rn.f32.s32 %f211, %r600;
fma.rn.ftz.f32 %f212, %f105, %f211, 0f00000000;

dp4a.s32.s32 %r393, %r466, %r395, %r512;


dp4a.s32.s32 %r397, %r470, %r399, %r393;


dp4a.s32.s32 %r401, %r514, %r395, %r512;


dp4a.s32.s32 %r405, %r514, %r399, %r401;

mul.lo.s32 %r601, %r397, %r543;
cvt.rn.f32.s32 %f213, %r601;
fma.rn.ftz.f32 %f214, %f106, %f213, %f210;
mul.lo.s32 %r602, %r405, %r545;
cvt.rn.f32.s32 %f215, %r602;
fma.rn.ftz.f32 %f216, %f106, %f215, %f212;

{.reg .f16 low,high;
mov.b32 {low,high},%r121;
cvt.f32.f16 %f107, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r121;
cvt.f32.f16 %f108, high;}


mul.ftz.f32 %f217, %f214, %f107;
mul.ftz.f32 %f218, %f216, %f108;
sub.ftz.f32 %f219, %f217, %f218;
add.ftz.f32 %f400, %f219, %f400;

{.reg .f16 low,high;
mov.b32 {low,high},%r375;
cvt.f32.f16 %f109, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r376;
cvt.f32.f16 %f110, low;}



dp4a.s32.s32 %r413, %r486, %r379, %r512;


dp4a.s32.s32 %r417, %r490, %r383, %r413;


dp4a.s32.s32 %r421, %r514, %r379, %r512;


dp4a.s32.s32 %r425, %r514, %r383, %r421;

mul.lo.s32 %r603, %r417, %r564;
cvt.rn.f32.s32 %f220, %r603;
fma.rn.ftz.f32 %f221, %f109, %f220, 0f00000000;
mul.lo.s32 %r604, %r425, %r567;
cvt.rn.f32.s32 %f222, %r604;
fma.rn.ftz.f32 %f223, %f109, %f222, 0f00000000;

dp4a.s32.s32 %r429, %r502, %r395, %r512;


dp4a.s32.s32 %r433, %r506, %r399, %r429;


dp4a.s32.s32 %r437, %r514, %r395, %r512;


dp4a.s32.s32 %r441, %r514, %r399, %r437;

mul.lo.s32 %r605, %r433, %r571;
cvt.rn.f32.s32 %f224, %r605;
fma.rn.ftz.f32 %f225, %f110, %f224, %f221;
mul.lo.s32 %r606, %r441, %r573;
cvt.rn.f32.s32 %f226, %r606;
fma.rn.ftz.f32 %f227, %f110, %f226, %f223;

{.reg .f16 low,high;
mov.b32 {low,high},%r157;
cvt.f32.f16 %f111, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r157;
cvt.f32.f16 %f112, high;}


mul.ftz.f32 %f228, %f225, %f111;
mul.ftz.f32 %f229, %f227, %f112;
sub.ftz.f32 %f230, %f228, %f229;
add.ftz.f32 %f399, %f230, %f399;
cvt.u64.u32 %rd62, %r856;
add.s64 %rd63, %rd2, %rd62;
mul.lo.s64 %rd64, %rd63, 36;
add.s64 %rd65, %rd10, %rd64;
ld.global.nc.u32 %r447, [%rd65];

{.reg .f16 low,high;
mov.b32 {low,high},%r447;
cvt.f32.f16 %f113, low;}


add.s64 %rd66, %rd65, %rd36;
ld.global.nc.u32 %r451, [%rd66+4];
ld.global.nc.u32 %r455, [%rd66+20];
ld.global.nc.u32 %r448, [%rd65+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r448;
cvt.f32.f16 %f114, low;}


ld.global.nc.u32 %r467, [%rd66+40];
ld.global.nc.u32 %r471, [%rd66+56];

dp4a.s32.s32 %r449, %r450, %r451, %r512;


dp4a.s32.s32 %r453, %r454, %r455, %r449;


dp4a.s32.s32 %r457, %r514, %r451, %r512;


dp4a.s32.s32 %r461, %r514, %r455, %r457;

mul.lo.s32 %r607, %r453, %r536;
cvt.rn.f32.s32 %f231, %r607;
fma.rn.ftz.f32 %f232, %f113, %f231, 0f00000000;
mul.lo.s32 %r608, %r461, %r539;
cvt.rn.f32.s32 %f233, %r608;
fma.rn.ftz.f32 %f234, %f113, %f233, 0f00000000;

dp4a.s32.s32 %r465, %r466, %r467, %r512;


dp4a.s32.s32 %r469, %r470, %r471, %r465;


dp4a.s32.s32 %r473, %r514, %r467, %r512;


dp4a.s32.s32 %r477, %r514, %r471, %r473;

mul.lo.s32 %r609, %r469, %r543;
cvt.rn.f32.s32 %f235, %r609;
fma.rn.ftz.f32 %f236, %f114, %f235, %f232;
mul.lo.s32 %r610, %r477, %r545;
cvt.rn.f32.s32 %f237, %r610;
fma.rn.ftz.f32 %f238, %f114, %f237, %f234;

{.reg .f16 low,high;
mov.b32 {low,high},%r121;
cvt.f32.f16 %f115, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r121;
cvt.f32.f16 %f116, high;}


mul.ftz.f32 %f239, %f236, %f115;
mul.ftz.f32 %f240, %f238, %f116;
sub.ftz.f32 %f241, %f239, %f240;
add.ftz.f32 %f398, %f241, %f398;

{.reg .f16 low,high;
mov.b32 {low,high},%r447;
cvt.f32.f16 %f117, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r448;
cvt.f32.f16 %f118, low;}



dp4a.s32.s32 %r485, %r486, %r451, %r512;


dp4a.s32.s32 %r489, %r490, %r455, %r485;


dp4a.s32.s32 %r493, %r514, %r451, %r512;


dp4a.s32.s32 %r497, %r514, %r455, %r493;

mul.lo.s32 %r611, %r489, %r564;
cvt.rn.f32.s32 %f242, %r611;
fma.rn.ftz.f32 %f243, %f117, %f242, 0f00000000;
mul.lo.s32 %r612, %r497, %r567;
cvt.rn.f32.s32 %f244, %r612;
fma.rn.ftz.f32 %f245, %f117, %f244, 0f00000000;

dp4a.s32.s32 %r501, %r502, %r467, %r512;


dp4a.s32.s32 %r505, %r506, %r471, %r501;


dp4a.s32.s32 %r509, %r514, %r467, %r512;


dp4a.s32.s32 %r513, %r514, %r471, %r509;

mul.lo.s32 %r613, %r505, %r571;
cvt.rn.f32.s32 %f246, %r613;
fma.rn.ftz.f32 %f247, %f118, %f246, %f243;
mul.lo.s32 %r614, %r513, %r573;
cvt.rn.f32.s32 %f248, %r614;
fma.rn.ftz.f32 %f249, %f118, %f248, %f245;

{.reg .f16 low,high;
mov.b32 {low,high},%r157;
cvt.f32.f16 %f119, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r157;
cvt.f32.f16 %f120, high;}


mul.ftz.f32 %f250, %f247, %f119;
mul.ftz.f32 %f251, %f249, %f120;
sub.ftz.f32 %f252, %f250, %f251;
add.ftz.f32 %f397, %f252, %f397;
add.s32 %r859, %r859, 32;
add.s32 %r858, %r858, 32;
add.s32 %r857, %r857, 32;
add.s32 %r856, %r856, 32;
add.s32 %r855, %r855, 32;
add.s32 %r854, %r854, 32;
add.s64 %rd111, %rd111, 576;
add.s32 %r860, %r860, 4;
setp.lt.s32 %p5, %r860, %r4;
@%p5 bra $L__BB165_2;

$L__BB165_3:
setp.eq.s32 %p6, %r1, 0;
@%p6 bra $L__BB165_5;

mov.u32 %r616, _ZZ13mul_mat_vec_qIL9ggml_type12ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r617, %r1, 1536, %r616;
shl.b32 %r618, %r46, 2;
add.s32 %r619, %r617, %r618;
st.shared.f32 [%r619+-1536], %f408;
st.shared.f32 [%r619+-1408], %f407;
st.shared.f32 [%r619+-1280], %f406;
st.shared.f32 [%r619+-1152], %f405;
st.shared.f32 [%r619+-1024], %f404;
st.shared.f32 [%r619+-896], %f403;
st.shared.f32 [%r619+-768], %f402;
st.shared.f32 [%r619+-640], %f401;
st.shared.f32 [%r619+-512], %f400;
st.shared.f32 [%r619+-384], %f399;
st.shared.f32 [%r619+-256], %f398;
st.shared.f32 [%r619+-128], %f397;

$L__BB165_5:
setp.ne.s32 %p7, %r1, 0;
bar.sync 0;
@%p7 bra $L__BB165_30;

mov.u32 %r620, %ctaid.z;
mov.u32 %r621, %ctaid.y;
mov.u32 %r622, %ctaid.x;
shl.b32 %r623, %r622, 1;
mov.u32 %r624, 1;
mad.lo.s32 %r625, %r621, %r37, %r623;
mad.lo.s32 %r626, %r620, %r43, %r625;
cvt.u64.u32 %rd13, %r626;
shl.b32 %r628, %r46, 2;
mov.u32 %r629, _ZZ13mul_mat_vec_qIL9ggml_type12ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r26, %r629, %r628;
mov.u32 %r630, 2;
ld.shared.f32 %f253, [%r26];
add.ftz.f32 %f254, %f253, %f408;
mov.b32 %r631, %f254;
mov.u32 %r632, 31;
mov.u32 %r633, 16;
mov.u32 %r634, -1;
shfl.sync.bfly.b32 %r635|%p8, %r631, %r633, %r632, %r634;
mov.b32 %f255, %r635;
add.ftz.f32 %f256, %f254, %f255;
mov.b32 %r636, %f256;
mov.u32 %r637, 8;
shfl.sync.bfly.b32 %r638|%p9, %r636, %r637, %r632, %r634;
mov.b32 %f257, %r638;
add.ftz.f32 %f258, %f256, %f257;
mov.b32 %r639, %f258;
mov.u32 %r640, 4;
shfl.sync.bfly.b32 %r641|%p10, %r639, %r640, %r632, %r634;
mov.b32 %f259, %r641;
add.ftz.f32 %f260, %f258, %f259;
mov.b32 %r642, %f260;
shfl.sync.bfly.b32 %r643|%p11, %r642, %r630, %r632, %r634;
mov.b32 %f261, %r643;
add.ftz.f32 %f262, %f260, %f261;
mov.b32 %r644, %f262;
shfl.sync.bfly.b32 %r645|%p12, %r644, %r624, %r632, %r634;
mov.b32 %f263, %r645;
add.ftz.f32 %f37, %f262, %f263;
setp.eq.s32 %p13, %r46, 0;
setp.lt.u32 %p14, %r623, %r31;
and.pred %p1, %p13, %p14;
not.pred %p15, %p1;
@%p15 bra $L__BB165_8;

cvta.to.global.u64 %rd67, %rd18;
mul.wide.u32 %rd68, %r626, 4;
add.s64 %rd69, %rd67, %rd68;
st.global.f32 [%rd69], %f37;

$L__BB165_8:
ld.shared.f32 %f264, [%r26+128];
add.ftz.f32 %f265, %f264, %f407;
mov.b32 %r652, %f265;
shfl.sync.bfly.b32 %r656|%p16, %r652, %r633, %r632, %r634;
mov.b32 %f266, %r656;
add.ftz.f32 %f267, %f265, %f266;
mov.b32 %r657, %f267;
shfl.sync.bfly.b32 %r659|%p17, %r657, %r637, %r632, %r634;
mov.b32 %f268, %r659;
add.ftz.f32 %f269, %f267, %f268;
mov.b32 %r660, %f269;
shfl.sync.bfly.b32 %r662|%p18, %r660, %r640, %r632, %r634;
mov.b32 %f270, %r662;
add.ftz.f32 %f271, %f269, %f270;
mov.b32 %r663, %f271;
shfl.sync.bfly.b32 %r665|%p19, %r663, %r630, %r632, %r634;
mov.b32 %f272, %r665;
add.ftz.f32 %f273, %f271, %f272;
mov.b32 %r666, %f273;
shfl.sync.bfly.b32 %r668|%p20, %r666, %r624, %r632, %r634;
mov.b32 %f274, %r668;
add.ftz.f32 %f38, %f273, %f274;
setp.eq.s32 %p21, %r46, 1;
or.b32 %r672, %r623, 1;
setp.lt.u32 %p22, %r672, %r31;
and.pred %p2, %p21, %p22;
not.pred %p23, %p2;
@%p23 bra $L__BB165_10;

cvta.to.global.u64 %rd70, %rd18;
mul.wide.u32 %rd71, %r626, 4;
add.s64 %rd72, %rd70, %rd71;
st.global.f32 [%rd72+4], %f38;

$L__BB165_10:
ld.shared.f32 %f275, [%r26+256];
add.ftz.f32 %f276, %f275, %f406;
mov.b32 %r679, %f276;
mov.u32 %r680, 31;
mov.u32 %r681, 16;
mov.u32 %r682, -1;
shfl.sync.bfly.b32 %r683|%p24, %r679, %r681, %r680, %r682;
mov.b32 %f277, %r683;
add.ftz.f32 %f278, %f276, %f277;
mov.b32 %r684, %f278;
mov.u32 %r685, 8;
shfl.sync.bfly.b32 %r686|%p25, %r684, %r685, %r680, %r682;
mov.b32 %f279, %r686;
add.ftz.f32 %f280, %f278, %f279;
mov.b32 %r687, %f280;
mov.u32 %r688, 4;
shfl.sync.bfly.b32 %r689|%p26, %r687, %r688, %r680, %r682;
mov.b32 %f281, %r689;
add.ftz.f32 %f282, %f280, %f281;
mov.b32 %r690, %f282;
mov.u32 %r691, 2;
shfl.sync.bfly.b32 %r692|%p27, %r690, %r691, %r680, %r682;
mov.b32 %f283, %r692;
add.ftz.f32 %f284, %f282, %f283;
mov.b32 %r693, %f284;
mov.u32 %r694, 1;
shfl.sync.bfly.b32 %r695|%p28, %r693, %r694, %r680, %r682;
mov.b32 %f285, %r695;
add.ftz.f32 %f39, %f284, %f285;
@%p15 bra $L__BB165_12;

cvt.u64.u32 %rd73, %r31;
add.s64 %rd74, %rd73, %rd13;
cvta.to.global.u64 %rd75, %rd18;
shl.b64 %rd76, %rd74, 2;
add.s64 %rd77, %rd75, %rd76;
st.global.f32 [%rd77], %f39;

$L__BB165_12:
ld.shared.f32 %f286, [%r26+384];
add.ftz.f32 %f287, %f286, %f405;
mov.b32 %r696, %f287;
shfl.sync.bfly.b32 %r700|%p30, %r696, %r681, %r680, %r682;
mov.b32 %f288, %r700;
add.ftz.f32 %f289, %f287, %f288;
mov.b32 %r701, %f289;
shfl.sync.bfly.b32 %r703|%p31, %r701, %r685, %r680, %r682;
mov.b32 %f290, %r703;
add.ftz.f32 %f291, %f289, %f290;
mov.b32 %r704, %f291;
shfl.sync.bfly.b32 %r706|%p32, %r704, %r688, %r680, %r682;
mov.b32 %f292, %r706;
add.ftz.f32 %f293, %f291, %f292;
mov.b32 %r707, %f293;
shfl.sync.bfly.b32 %r709|%p33, %r707, %r691, %r680, %r682;
mov.b32 %f294, %r709;
add.ftz.f32 %f295, %f293, %f294;
mov.b32 %r710, %f295;
shfl.sync.bfly.b32 %r712|%p34, %r710, %r694, %r680, %r682;
mov.b32 %f296, %r712;
add.ftz.f32 %f40, %f295, %f296;
@%p23 bra $L__BB165_14;

add.s32 %r713, %r31, 1;
cvt.u64.u32 %rd78, %r713;
add.s64 %rd79, %rd78, %rd13;
cvta.to.global.u64 %rd80, %rd18;
shl.b64 %rd81, %rd79, 2;
add.s64 %rd82, %rd80, %rd81;
st.global.f32 [%rd82], %f40;

$L__BB165_14:
shl.b32 %r714, %r31, 1;
mov.u32 %r715, 1;
ld.shared.f32 %f297, [%r26+512];
add.ftz.f32 %f298, %f297, %f404;
mov.b32 %r716, %f298;
mov.u32 %r717, 31;
mov.u32 %r718, 16;
mov.u32 %r719, -1;
shfl.sync.bfly.b32 %r720|%p36, %r716, %r718, %r717, %r719;
mov.b32 %f299, %r720;
add.ftz.f32 %f300, %f298, %f299;
mov.b32 %r721, %f300;
mov.u32 %r722, 8;
shfl.sync.bfly.b32 %r723|%p37, %r721, %r722, %r717, %r719;
mov.b32 %f301, %r723;
add.ftz.f32 %f302, %f300, %f301;
mov.b32 %r724, %f302;
mov.u32 %r725, 4;
shfl.sync.bfly.b32 %r726|%p38, %r724, %r725, %r717, %r719;
mov.b32 %f303, %r726;
add.ftz.f32 %f304, %f302, %f303;
mov.b32 %r727, %f304;
mov.u32 %r728, 2;
shfl.sync.bfly.b32 %r729|%p39, %r727, %r728, %r717, %r719;
mov.b32 %f305, %r729;
add.ftz.f32 %f306, %f304, %f305;
mov.b32 %r730, %f306;
shfl.sync.bfly.b32 %r731|%p40, %r730, %r715, %r717, %r719;
mov.b32 %f307, %r731;
add.ftz.f32 %f41, %f306, %f307;
cvt.u64.u32 %rd83, %r714;
add.s64 %rd84, %rd83, %rd13;
cvta.to.global.u64 %rd85, %rd18;
shl.b64 %rd86, %rd84, 2;
add.s64 %rd14, %rd85, %rd86;
@%p15 bra $L__BB165_16;

st.global.f32 [%rd14], %f41;

$L__BB165_16:
ld.shared.f32 %f308, [%r26+640];
add.ftz.f32 %f309, %f308, %f403;
mov.b32 %r732, %f309;
shfl.sync.bfly.b32 %r736|%p42, %r732, %r718, %r717, %r719;
mov.b32 %f310, %r736;
add.ftz.f32 %f311, %f309, %f310;
mov.b32 %r737, %f311;
shfl.sync.bfly.b32 %r739|%p43, %r737, %r722, %r717, %r719;
mov.b32 %f312, %r739;
add.ftz.f32 %f313, %f311, %f312;
mov.b32 %r740, %f313;
shfl.sync.bfly.b32 %r742|%p44, %r740, %r725, %r717, %r719;
mov.b32 %f314, %r742;
add.ftz.f32 %f315, %f313, %f314;
mov.b32 %r743, %f315;
shfl.sync.bfly.b32 %r745|%p45, %r743, %r728, %r717, %r719;
mov.b32 %f316, %r745;
add.ftz.f32 %f317, %f315, %f316;
mov.b32 %r746, %f317;
shfl.sync.bfly.b32 %r748|%p46, %r746, %r715, %r717, %r719;
mov.b32 %f318, %r748;
add.ftz.f32 %f42, %f317, %f318;
@%p23 bra $L__BB165_18;

st.global.f32 [%rd14+4], %f42;

$L__BB165_18:
mul.lo.s32 %r27, %r31, 3;
ld.shared.f32 %f319, [%r26+768];
add.ftz.f32 %f320, %f319, %f402;
mov.b32 %r749, %f320;
mov.u32 %r750, 31;
mov.u32 %r751, 16;
mov.u32 %r752, -1;
shfl.sync.bfly.b32 %r753|%p48, %r749, %r751, %r750, %r752;
mov.b32 %f321, %r753;
add.ftz.f32 %f322, %f320, %f321;
mov.b32 %r754, %f322;
mov.u32 %r755, 8;
shfl.sync.bfly.b32 %r756|%p49, %r754, %r755, %r750, %r752;
mov.b32 %f323, %r756;
add.ftz.f32 %f324, %f322, %f323;
mov.b32 %r757, %f324;
mov.u32 %r758, 4;
shfl.sync.bfly.b32 %r759|%p50, %r757, %r758, %r750, %r752;
mov.b32 %f325, %r759;
add.ftz.f32 %f326, %f324, %f325;
mov.b32 %r760, %f326;
mov.u32 %r761, 2;
shfl.sync.bfly.b32 %r762|%p51, %r760, %r761, %r750, %r752;
mov.b32 %f327, %r762;
add.ftz.f32 %f328, %f326, %f327;
mov.b32 %r763, %f328;
mov.u32 %r764, 1;
shfl.sync.bfly.b32 %r765|%p52, %r763, %r764, %r750, %r752;
mov.b32 %f329, %r765;
add.ftz.f32 %f43, %f328, %f329;
@%p15 bra $L__BB165_20;

cvt.u64.u32 %rd87, %r27;
add.s64 %rd88, %rd87, %rd13;
shl.b64 %rd90, %rd88, 2;
add.s64 %rd91, %rd85, %rd90;
st.global.f32 [%rd91], %f43;

$L__BB165_20:
ld.shared.f32 %f330, [%r26+896];
add.ftz.f32 %f331, %f330, %f401;
mov.b32 %r766, %f331;
shfl.sync.bfly.b32 %r770|%p54, %r766, %r751, %r750, %r752;
mov.b32 %f332, %r770;
add.ftz.f32 %f333, %f331, %f332;
mov.b32 %r771, %f333;
shfl.sync.bfly.b32 %r773|%p55, %r771, %r755, %r750, %r752;
mov.b32 %f334, %r773;
add.ftz.f32 %f335, %f333, %f334;
mov.b32 %r774, %f335;
shfl.sync.bfly.b32 %r776|%p56, %r774, %r758, %r750, %r752;
mov.b32 %f336, %r776;
add.ftz.f32 %f337, %f335, %f336;
mov.b32 %r777, %f337;
shfl.sync.bfly.b32 %r779|%p57, %r777, %r761, %r750, %r752;
mov.b32 %f338, %r779;
add.ftz.f32 %f339, %f337, %f338;
mov.b32 %r780, %f339;
shfl.sync.bfly.b32 %r782|%p58, %r780, %r764, %r750, %r752;
mov.b32 %f340, %r782;
add.ftz.f32 %f44, %f339, %f340;
@%p23 bra $L__BB165_22;

add.s32 %r783, %r27, 1;
cvt.u64.u32 %rd92, %r783;
add.s64 %rd93, %rd92, %rd13;
shl.b64 %rd95, %rd93, 2;
add.s64 %rd96, %rd85, %rd95;
st.global.f32 [%rd96], %f44;

$L__BB165_22:
shl.b32 %r784, %r31, 2;
mov.u32 %r785, 2;
ld.shared.f32 %f341, [%r26+1024];
add.ftz.f32 %f342, %f341, %f400;
mov.b32 %r786, %f342;
mov.u32 %r787, 31;
mov.u32 %r788, 16;
mov.u32 %r789, -1;
shfl.sync.bfly.b32 %r790|%p60, %r786, %r788, %r787, %r789;
mov.b32 %f343, %r790;
add.ftz.f32 %f344, %f342, %f343;
mov.b32 %r791, %f344;
mov.u32 %r792, 8;
shfl.sync.bfly.b32 %r793|%p61, %r791, %r792, %r787, %r789;
mov.b32 %f345, %r793;
add.ftz.f32 %f346, %f344, %f345;
mov.b32 %r794, %f346;
mov.u32 %r795, 4;
shfl.sync.bfly.b32 %r796|%p62, %r794, %r795, %r787, %r789;
mov.b32 %f347, %r796;
add.ftz.f32 %f348, %f346, %f347;
mov.b32 %r797, %f348;
shfl.sync.bfly.b32 %r798|%p63, %r797, %r785, %r787, %r789;
mov.b32 %f349, %r798;
add.ftz.f32 %f350, %f348, %f349;
mov.b32 %r799, %f350;
mov.u32 %r800, 1;
shfl.sync.bfly.b32 %r801|%p64, %r799, %r800, %r787, %r789;
mov.b32 %f351, %r801;
add.ftz.f32 %f45, %f350, %f351;
cvt.u64.u32 %rd97, %r784;
add.s64 %rd98, %rd97, %rd13;
shl.b64 %rd100, %rd98, 2;
add.s64 %rd15, %rd85, %rd100;
@%p15 bra $L__BB165_24;

st.global.f32 [%rd15], %f45;

$L__BB165_24:
ld.shared.f32 %f352, [%r26+1152];
add.ftz.f32 %f353, %f352, %f399;
mov.b32 %r802, %f353;
shfl.sync.bfly.b32 %r806|%p66, %r802, %r788, %r787, %r789;
mov.b32 %f354, %r806;
add.ftz.f32 %f355, %f353, %f354;
mov.b32 %r807, %f355;
shfl.sync.bfly.b32 %r809|%p67, %r807, %r792, %r787, %r789;
mov.b32 %f356, %r809;
add.ftz.f32 %f357, %f355, %f356;
mov.b32 %r810, %f357;
shfl.sync.bfly.b32 %r812|%p68, %r810, %r795, %r787, %r789;
mov.b32 %f358, %r812;
add.ftz.f32 %f359, %f357, %f358;
mov.b32 %r813, %f359;
shfl.sync.bfly.b32 %r815|%p69, %r813, %r785, %r787, %r789;
mov.b32 %f360, %r815;
add.ftz.f32 %f361, %f359, %f360;
mov.b32 %r816, %f361;
shfl.sync.bfly.b32 %r818|%p70, %r816, %r800, %r787, %r789;
mov.b32 %f362, %r818;
add.ftz.f32 %f46, %f361, %f362;
@%p23 bra $L__BB165_26;

st.global.f32 [%rd15+4], %f46;

$L__BB165_26:
mul.lo.s32 %r28, %r31, 5;
ld.shared.f32 %f363, [%r26+1280];
add.ftz.f32 %f364, %f363, %f398;
mov.b32 %r819, %f364;
mov.u32 %r820, 31;
mov.u32 %r821, 16;
mov.u32 %r822, -1;
shfl.sync.bfly.b32 %r823|%p72, %r819, %r821, %r820, %r822;
mov.b32 %f365, %r823;
add.ftz.f32 %f366, %f364, %f365;
mov.b32 %r824, %f366;
mov.u32 %r825, 8;
shfl.sync.bfly.b32 %r826|%p73, %r824, %r825, %r820, %r822;
mov.b32 %f367, %r826;
add.ftz.f32 %f368, %f366, %f367;
mov.b32 %r827, %f368;
mov.u32 %r828, 4;
shfl.sync.bfly.b32 %r829|%p74, %r827, %r828, %r820, %r822;
mov.b32 %f369, %r829;
add.ftz.f32 %f370, %f368, %f369;
mov.b32 %r830, %f370;
mov.u32 %r831, 2;
shfl.sync.bfly.b32 %r832|%p75, %r830, %r831, %r820, %r822;
mov.b32 %f371, %r832;
add.ftz.f32 %f372, %f370, %f371;
mov.b32 %r833, %f372;
mov.u32 %r834, 1;
shfl.sync.bfly.b32 %r835|%p76, %r833, %r834, %r820, %r822;
mov.b32 %f373, %r835;
add.ftz.f32 %f47, %f372, %f373;
@%p15 bra $L__BB165_28;

cvt.u64.u32 %rd101, %r28;
add.s64 %rd102, %rd101, %rd13;
shl.b64 %rd104, %rd102, 2;
add.s64 %rd105, %rd85, %rd104;
st.global.f32 [%rd105], %f47;

$L__BB165_28:
ld.shared.f32 %f374, [%r26+1408];
add.ftz.f32 %f375, %f374, %f397;
mov.b32 %r836, %f375;
shfl.sync.bfly.b32 %r840|%p78, %r836, %r821, %r820, %r822;
mov.b32 %f376, %r840;
add.ftz.f32 %f377, %f375, %f376;
mov.b32 %r841, %f377;
shfl.sync.bfly.b32 %r843|%p79, %r841, %r825, %r820, %r822;
mov.b32 %f378, %r843;
add.ftz.f32 %f379, %f377, %f378;
mov.b32 %r844, %f379;
shfl.sync.bfly.b32 %r846|%p80, %r844, %r828, %r820, %r822;
mov.b32 %f380, %r846;
add.ftz.f32 %f381, %f379, %f380;
mov.b32 %r847, %f381;
shfl.sync.bfly.b32 %r849|%p81, %r847, %r831, %r820, %r822;
mov.b32 %f382, %r849;
add.ftz.f32 %f383, %f381, %f382;
mov.b32 %r850, %f383;
shfl.sync.bfly.b32 %r852|%p82, %r850, %r834, %r820, %r822;
mov.b32 %f384, %r852;
add.ftz.f32 %f48, %f383, %f384;
@%p23 bra $L__BB165_30;

add.s32 %r853, %r28, 1;
cvt.u64.u32 %rd106, %r853;
add.s64 %rd107, %rd106, %rd13;
shl.b64 %rd109, %rd107, 2;
add.s64 %rd110, %rd85, %rd109;
st.global.f32 [%rd110], %f48;

$L__BB165_30:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type12ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type12ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type12ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type12ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type12ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type12ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type12ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type12ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type12ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<96>;
.reg .b16 %rs<15>;
.reg .f32 %f<477>;
.reg .b32 %r<964>;
.reg .b64 %rd<119>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type12ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1792];

ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type12ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type12ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd20, [_Z13mul_mat_vec_qIL9ggml_type12ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r49, [_Z13mul_mat_vec_qIL9ggml_type12ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type12ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type12ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type12ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type12ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type12ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type12ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r46, [_Z13mul_mat_vec_qIL9ggml_type12ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r47, [_Z13mul_mat_vec_qIL9ggml_type12ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r48, [_Z13mul_mat_vec_qIL9ggml_type12ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type12ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type12ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type12ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type12ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r50, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r50, %r2;
mov.u32 %r51, %ctaid.x;
shl.b32 %r4, %r51, 1;
shr.s32 %r52, %r3, 31;
shr.u32 %r53, %r52, 28;
add.s32 %r54, %r3, %r53;
shr.s32 %r963, %r54, 4;
shr.u32 %r6, %r49, 8;
setp.ge.s32 %p3, %r963, %r6;
mov.f32 %f463, 0f00000000;
mov.f32 %f464, %f463;
mov.f32 %f465, %f463;
mov.f32 %f466, %f463;
mov.f32 %f467, %f463;
mov.f32 %f468, %f463;
mov.f32 %f469, %f463;
mov.f32 %f470, %f463;
mov.f32 %f471, %f463;
mov.f32 %f472, %f463;
mov.f32 %f473, %f463;
mov.f32 %f474, %f463;
mov.f32 %f475, %f463;
mov.f32 %f476, %f463;
@%p3 bra $L__BB166_3;

mov.u32 %r55, %ctaid.y;
mov.u32 %r56, %ctaid.z;
mul.hi.u32 %r57, %r56, %r43;
add.s32 %r58, %r57, %r56;
shr.u32 %r59, %r58, %r44;
mul.lo.s32 %r60, %r59, %r46;
mul.hi.u32 %r61, %r55, %r37;
add.s32 %r62, %r61, %r55;
shr.u32 %r63, %r62, %r38;
mul.lo.s32 %r64, %r63, %r40;
mul.lo.s32 %r65, %r55, %r41;
cvt.u64.u32 %rd21, %r65;
cvta.to.global.u64 %rd118, %rd18;
and.b32 %r69, %r54, -16;
sub.s32 %r7, %r3, %r69;
shr.s32 %r70, %r7, 31;
shr.u32 %r71, %r70, 30;
add.s32 %r72, %r7, %r71;
shr.s32 %r73, %r72, 2;
shl.b32 %r74, %r73, 1;
shl.b32 %r75, %r73, 5;
and.b32 %r76, %r72, -4;
sub.s32 %r77, %r7, %r76;
shl.b32 %r78, %r77, 2;
add.s32 %r79, %r75, %r78;
cvt.s64.s32 %rd22, %r79;
cvt.s64.s32 %rd23, %r74;
add.s64 %rd24, %rd21, %rd23;
mul.lo.s32 %r80, %r56, %r47;
cvt.u64.u32 %rd25, %r80;
add.s64 %rd2, %rd24, %rd25;
cvt.s64.s32 %rd3, %r77;
shl.b32 %r81, %r35, 1;
shl.b32 %r956, %r963, 3;
add.s32 %r962, %r956, %r81;
mad.lo.s32 %r961, %r35, 3, %r956;
shl.b32 %r82, %r35, 2;
add.s32 %r960, %r956, %r82;
mad.lo.s32 %r959, %r35, 5, %r956;
mad.lo.s32 %r958, %r35, 6, %r956;
add.s32 %r957, %r35, %r956;
add.s32 %r83, %r963, %r60;
add.s32 %r84, %r83, %r64;
add.s32 %r85, %r4, 1;
mad.lo.s32 %r86, %r34, %r85, %r84;
mul.wide.s32 %rd4, %r86, 144;
mad.lo.s32 %r87, %r4, %r34, %r963;
add.s32 %r88, %r87, %r60;
add.s32 %r89, %r88, %r64;
mul.wide.s32 %rd5, %r89, 144;
add.s64 %rd26, %rd5, %rd22;
add.s64 %rd6, %rd26, 32;
cvt.u64.u32 %rd27, %r74;
and.b64 %rd28, %rd27, 2;
or.b64 %rd29, %rd28, %rd4;
or.b64 %rd7, %rd29, 8;
or.b64 %rd30, %rd28, %rd5;
or.b64 %rd8, %rd30, 12;
add.s64 %rd31, %rd4, %rd22;
add.s64 %rd9, %rd31, 16;
cvta.to.global.u64 %rd10, %rd19;
mov.f32 %f463, 0f00000000;
shl.b64 %rd38, %rd3, 2;

$L__BB166_2:
cvt.u64.u32 %rd32, %r956;
add.s64 %rd33, %rd2, %rd32;
add.s64 %rd34, %rd118, %rd6;
ld.global.nc.u32 %r594, [%rd34+-16];
ld.global.nc.u32 %r595, [%rd34];
add.s64 %rd35, %rd118, %rd8;
ld.global.nc.u16 %rs1, [%rd35+-8];
cvt.u32.u16 %r596, %rs1;
ld.global.nc.u16 %rs2, [%rd35+-4];
cvt.u32.u16 %r597, %rs2;
ld.global.nc.u16 %rs3, [%rd35];
cvt.u32.u16 %r598, %rs3;
and.b32 %r599, %r596, 16191;
and.b32 %r600, %r598, 3855;
and.b32 %r601, %r596, 49344;
shr.u32 %r602, %r601, 2;
or.b32 %r603, %r600, %r602;
setp.gt.s32 %p4, %r7, 7;
and.b32 %r604, %r597, 16191;
shr.u32 %r605, %r598, 4;
and.b32 %r606, %r605, 3855;
and.b32 %r607, %r597, 49344;
shr.u32 %r608, %r607, 2;
or.b32 %r609, %r606, %r608;
mul.lo.s64 %rd36, %rd33, 36;
add.s64 %rd37, %rd10, %rd36;
ld.global.nc.u32 %r90, [%rd37];

{.reg .f16 low,high;
mov.b32 {low,high},%r90;
cvt.f32.f16 %f85, low;}


add.s64 %rd39, %rd37, %rd38;
ld.global.nc.u32 %r94, [%rd39+4];
ld.global.nc.u32 %r98, [%rd39+20];
ld.global.nc.u32 %r91, [%rd37+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r91;
cvt.f32.f16 %f86, low;}


ld.global.nc.u32 %r110, [%rd39+40];
ld.global.nc.u32 %r114, [%rd39+56];
add.s64 %rd40, %rd118, %rd5;
and.b32 %r525, %r594, 252645135;
and.b32 %r529, %r595, 252645135;
mov.u32 %r587, 0;

dp4a.s32.s32 %r92, %r525, %r94, %r587;


dp4a.s32.s32 %r96, %r529, %r98, %r92;

mov.u32 %r589, 16843009;

dp4a.s32.s32 %r100, %r589, %r94, %r587;


dp4a.s32.s32 %r104, %r589, %r98, %r100;

selp.b32 %r610, %r603, %r599, %p4;
cvt.u16.u32 %rs4, %r610;
and.b32 %r611, %r610, 63;
mul.lo.s32 %r612, %r96, %r611;
cvt.rn.f32.s32 %f141, %r612;
fma.rn.ftz.f32 %f142, %f85, %f141, 0f00000000;
selp.b32 %r613, %r609, %r604, %p4;
cvt.u16.u32 %rs5, %r613;
and.b32 %r614, %r613, 63;
mul.lo.s32 %r615, %r104, %r614;
cvt.rn.f32.s32 %f143, %r615;
fma.rn.ftz.f32 %f144, %f85, %f143, 0f00000000;
shr.u32 %r616, %r594, 4;
and.b32 %r541, %r616, 252645135;
shr.u32 %r617, %r595, 4;
and.b32 %r545, %r617, 252645135;

dp4a.s32.s32 %r108, %r541, %r110, %r587;


dp4a.s32.s32 %r112, %r545, %r114, %r108;


dp4a.s32.s32 %r116, %r589, %r110, %r587;


dp4a.s32.s32 %r120, %r589, %r114, %r116;

shr.u16 %rs6, %rs4, 8;
cvt.u32.u16 %r618, %rs6;
mul.lo.s32 %r619, %r112, %r618;
cvt.rn.f32.s32 %f145, %r619;
fma.rn.ftz.f32 %f146, %f86, %f145, %f142;
shr.u16 %rs7, %rs5, 8;
cvt.u32.u16 %r620, %rs7;
mul.lo.s32 %r621, %r120, %r620;
cvt.rn.f32.s32 %f147, %r621;
fma.rn.ftz.f32 %f148, %f86, %f147, %f144;
ld.global.nc.u32 %r124, [%rd40];

{.reg .f16 low,high;
mov.b32 {low,high},%r124;
cvt.f32.f16 %f87, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r124;
cvt.f32.f16 %f88, high;}


mul.ftz.f32 %f149, %f146, %f87;
mul.ftz.f32 %f150, %f148, %f88;
sub.ftz.f32 %f151, %f149, %f150;
add.ftz.f32 %f476, %f151, %f476;
add.s64 %rd41, %rd118, %rd9;
ld.global.nc.u32 %r622, [%rd41];
ld.global.nc.u32 %r623, [%rd41+16];
add.s64 %rd42, %rd118, %rd7;
ld.global.nc.u16 %rs8, [%rd42+-4];
cvt.u32.u16 %r624, %rs8;
ld.global.nc.u16 %rs9, [%rd42];
cvt.u32.u16 %r625, %rs9;
ld.global.nc.u16 %rs10, [%rd42+4];
cvt.u32.u16 %r626, %rs10;
and.b32 %r627, %r624, 16191;
and.b32 %r628, %r626, 3855;
and.b32 %r629, %r624, 49344;
shr.u32 %r630, %r629, 2;
or.b32 %r631, %r628, %r630;
and.b32 %r632, %r625, 16191;
shr.u32 %r633, %r626, 4;
and.b32 %r634, %r633, 3855;
and.b32 %r635, %r625, 49344;
shr.u32 %r636, %r635, 2;
or.b32 %r637, %r634, %r636;

{.reg .f16 low,high;
mov.b32 {low,high},%r90;
cvt.f32.f16 %f89, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r91;
cvt.f32.f16 %f90, low;}


add.s64 %rd43, %rd118, %rd4;
and.b32 %r561, %r622, 252645135;
and.b32 %r565, %r623, 252645135;

dp4a.s32.s32 %r128, %r561, %r94, %r587;


dp4a.s32.s32 %r132, %r565, %r98, %r128;


dp4a.s32.s32 %r136, %r589, %r94, %r587;


dp4a.s32.s32 %r140, %r589, %r98, %r136;

selp.b32 %r638, %r631, %r627, %p4;
cvt.u16.u32 %rs11, %r638;
and.b32 %r639, %r638, 63;
mul.lo.s32 %r640, %r132, %r639;
cvt.rn.f32.s32 %f152, %r640;
fma.rn.ftz.f32 %f153, %f89, %f152, 0f00000000;
selp.b32 %r641, %r637, %r632, %p4;
cvt.u16.u32 %rs12, %r641;
and.b32 %r642, %r641, 63;
mul.lo.s32 %r643, %r140, %r642;
cvt.rn.f32.s32 %f154, %r643;
fma.rn.ftz.f32 %f155, %f89, %f154, 0f00000000;
shr.u32 %r644, %r622, 4;
and.b32 %r577, %r644, 252645135;
shr.u32 %r645, %r623, 4;
and.b32 %r581, %r645, 252645135;

dp4a.s32.s32 %r144, %r577, %r110, %r587;


dp4a.s32.s32 %r148, %r581, %r114, %r144;


dp4a.s32.s32 %r152, %r589, %r110, %r587;


dp4a.s32.s32 %r156, %r589, %r114, %r152;

shr.u16 %rs13, %rs11, 8;
cvt.u32.u16 %r646, %rs13;
mul.lo.s32 %r647, %r148, %r646;
cvt.rn.f32.s32 %f156, %r647;
fma.rn.ftz.f32 %f157, %f90, %f156, %f153;
shr.u16 %rs14, %rs12, 8;
cvt.u32.u16 %r648, %rs14;
mul.lo.s32 %r649, %r156, %r648;
cvt.rn.f32.s32 %f158, %r649;
fma.rn.ftz.f32 %f159, %f90, %f158, %f155;
ld.global.nc.u32 %r160, [%rd43];

{.reg .f16 low,high;
mov.b32 {low,high},%r160;
cvt.f32.f16 %f91, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r160;
cvt.f32.f16 %f92, high;}


mul.ftz.f32 %f160, %f157, %f91;
mul.ftz.f32 %f161, %f159, %f92;
sub.ftz.f32 %f162, %f160, %f161;
add.ftz.f32 %f475, %f162, %f475;
cvt.u64.u32 %rd44, %r957;
add.s64 %rd45, %rd2, %rd44;
mul.lo.s64 %rd46, %rd45, 36;
add.s64 %rd47, %rd10, %rd46;
ld.global.nc.u32 %r162, [%rd47];

{.reg .f16 low,high;
mov.b32 {low,high},%r162;
cvt.f32.f16 %f93, low;}


add.s64 %rd48, %rd47, %rd38;
ld.global.nc.u32 %r166, [%rd48+4];
ld.global.nc.u32 %r170, [%rd48+20];
ld.global.nc.u32 %r163, [%rd47+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r163;
cvt.f32.f16 %f94, low;}


ld.global.nc.u32 %r182, [%rd48+40];
ld.global.nc.u32 %r186, [%rd48+56];

dp4a.s32.s32 %r164, %r525, %r166, %r587;


dp4a.s32.s32 %r168, %r529, %r170, %r164;


dp4a.s32.s32 %r172, %r589, %r166, %r587;


dp4a.s32.s32 %r176, %r589, %r170, %r172;

mul.lo.s32 %r650, %r168, %r611;
cvt.rn.f32.s32 %f163, %r650;
fma.rn.ftz.f32 %f164, %f93, %f163, 0f00000000;
mul.lo.s32 %r651, %r176, %r614;
cvt.rn.f32.s32 %f165, %r651;
fma.rn.ftz.f32 %f166, %f93, %f165, 0f00000000;

dp4a.s32.s32 %r180, %r541, %r182, %r587;


dp4a.s32.s32 %r184, %r545, %r186, %r180;


dp4a.s32.s32 %r188, %r589, %r182, %r587;


dp4a.s32.s32 %r192, %r589, %r186, %r188;

mul.lo.s32 %r652, %r184, %r618;
cvt.rn.f32.s32 %f167, %r652;
fma.rn.ftz.f32 %f168, %f94, %f167, %f164;
mul.lo.s32 %r653, %r192, %r620;
cvt.rn.f32.s32 %f169, %r653;
fma.rn.ftz.f32 %f170, %f94, %f169, %f166;

{.reg .f16 low,high;
mov.b32 {low,high},%r124;
cvt.f32.f16 %f95, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r124;
cvt.f32.f16 %f96, high;}


mul.ftz.f32 %f171, %f168, %f95;
mul.ftz.f32 %f172, %f170, %f96;
sub.ftz.f32 %f173, %f171, %f172;
add.ftz.f32 %f474, %f173, %f474;

{.reg .f16 low,high;
mov.b32 {low,high},%r162;
cvt.f32.f16 %f97, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r163;
cvt.f32.f16 %f98, low;}



dp4a.s32.s32 %r200, %r561, %r166, %r587;


dp4a.s32.s32 %r204, %r565, %r170, %r200;


dp4a.s32.s32 %r208, %r589, %r166, %r587;


dp4a.s32.s32 %r212, %r589, %r170, %r208;

mul.lo.s32 %r654, %r204, %r639;
cvt.rn.f32.s32 %f174, %r654;
fma.rn.ftz.f32 %f175, %f97, %f174, 0f00000000;
mul.lo.s32 %r655, %r212, %r642;
cvt.rn.f32.s32 %f176, %r655;
fma.rn.ftz.f32 %f177, %f97, %f176, 0f00000000;

dp4a.s32.s32 %r216, %r577, %r182, %r587;


dp4a.s32.s32 %r220, %r581, %r186, %r216;


dp4a.s32.s32 %r224, %r589, %r182, %r587;


dp4a.s32.s32 %r228, %r589, %r186, %r224;

mul.lo.s32 %r656, %r220, %r646;
cvt.rn.f32.s32 %f178, %r656;
fma.rn.ftz.f32 %f179, %f98, %f178, %f175;
mul.lo.s32 %r657, %r228, %r648;
cvt.rn.f32.s32 %f180, %r657;
fma.rn.ftz.f32 %f181, %f98, %f180, %f177;

{.reg .f16 low,high;
mov.b32 {low,high},%r160;
cvt.f32.f16 %f99, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r160;
cvt.f32.f16 %f100, high;}


mul.ftz.f32 %f182, %f179, %f99;
mul.ftz.f32 %f183, %f181, %f100;
sub.ftz.f32 %f184, %f182, %f183;
add.ftz.f32 %f473, %f184, %f473;
cvt.u64.u32 %rd49, %r962;
add.s64 %rd50, %rd2, %rd49;
mul.lo.s64 %rd51, %rd50, 36;
add.s64 %rd52, %rd10, %rd51;
ld.global.nc.u32 %r234, [%rd52];

{.reg .f16 low,high;
mov.b32 {low,high},%r234;
cvt.f32.f16 %f101, low;}


add.s64 %rd53, %rd52, %rd38;
ld.global.nc.u32 %r238, [%rd53+4];
ld.global.nc.u32 %r242, [%rd53+20];
ld.global.nc.u32 %r235, [%rd52+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r235;
cvt.f32.f16 %f102, low;}


ld.global.nc.u32 %r254, [%rd53+40];
ld.global.nc.u32 %r258, [%rd53+56];

dp4a.s32.s32 %r236, %r525, %r238, %r587;


dp4a.s32.s32 %r240, %r529, %r242, %r236;


dp4a.s32.s32 %r244, %r589, %r238, %r587;


dp4a.s32.s32 %r248, %r589, %r242, %r244;

mul.lo.s32 %r658, %r240, %r611;
cvt.rn.f32.s32 %f185, %r658;
fma.rn.ftz.f32 %f186, %f101, %f185, 0f00000000;
mul.lo.s32 %r659, %r248, %r614;
cvt.rn.f32.s32 %f187, %r659;
fma.rn.ftz.f32 %f188, %f101, %f187, 0f00000000;

dp4a.s32.s32 %r252, %r541, %r254, %r587;


dp4a.s32.s32 %r256, %r545, %r258, %r252;


dp4a.s32.s32 %r260, %r589, %r254, %r587;


dp4a.s32.s32 %r264, %r589, %r258, %r260;

mul.lo.s32 %r660, %r256, %r618;
cvt.rn.f32.s32 %f189, %r660;
fma.rn.ftz.f32 %f190, %f102, %f189, %f186;
mul.lo.s32 %r661, %r264, %r620;
cvt.rn.f32.s32 %f191, %r661;
fma.rn.ftz.f32 %f192, %f102, %f191, %f188;

{.reg .f16 low,high;
mov.b32 {low,high},%r124;
cvt.f32.f16 %f103, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r124;
cvt.f32.f16 %f104, high;}


mul.ftz.f32 %f193, %f190, %f103;
mul.ftz.f32 %f194, %f192, %f104;
sub.ftz.f32 %f195, %f193, %f194;
add.ftz.f32 %f472, %f195, %f472;

{.reg .f16 low,high;
mov.b32 {low,high},%r234;
cvt.f32.f16 %f105, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r235;
cvt.f32.f16 %f106, low;}



dp4a.s32.s32 %r272, %r561, %r238, %r587;


dp4a.s32.s32 %r276, %r565, %r242, %r272;


dp4a.s32.s32 %r280, %r589, %r238, %r587;


dp4a.s32.s32 %r284, %r589, %r242, %r280;

mul.lo.s32 %r662, %r276, %r639;
cvt.rn.f32.s32 %f196, %r662;
fma.rn.ftz.f32 %f197, %f105, %f196, 0f00000000;
mul.lo.s32 %r663, %r284, %r642;
cvt.rn.f32.s32 %f198, %r663;
fma.rn.ftz.f32 %f199, %f105, %f198, 0f00000000;

dp4a.s32.s32 %r288, %r577, %r254, %r587;


dp4a.s32.s32 %r292, %r581, %r258, %r288;


dp4a.s32.s32 %r296, %r589, %r254, %r587;


dp4a.s32.s32 %r300, %r589, %r258, %r296;

mul.lo.s32 %r664, %r292, %r646;
cvt.rn.f32.s32 %f200, %r664;
fma.rn.ftz.f32 %f201, %f106, %f200, %f197;
mul.lo.s32 %r665, %r300, %r648;
cvt.rn.f32.s32 %f202, %r665;
fma.rn.ftz.f32 %f203, %f106, %f202, %f199;

{.reg .f16 low,high;
mov.b32 {low,high},%r160;
cvt.f32.f16 %f107, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r160;
cvt.f32.f16 %f108, high;}


mul.ftz.f32 %f204, %f201, %f107;
mul.ftz.f32 %f205, %f203, %f108;
sub.ftz.f32 %f206, %f204, %f205;
add.ftz.f32 %f471, %f206, %f471;
cvt.u64.u32 %rd54, %r961;
add.s64 %rd55, %rd2, %rd54;
mul.lo.s64 %rd56, %rd55, 36;
add.s64 %rd57, %rd10, %rd56;
ld.global.nc.u32 %r306, [%rd57];

{.reg .f16 low,high;
mov.b32 {low,high},%r306;
cvt.f32.f16 %f109, low;}


add.s64 %rd58, %rd57, %rd38;
ld.global.nc.u32 %r310, [%rd58+4];
ld.global.nc.u32 %r314, [%rd58+20];
ld.global.nc.u32 %r307, [%rd57+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r307;
cvt.f32.f16 %f110, low;}


ld.global.nc.u32 %r326, [%rd58+40];
ld.global.nc.u32 %r330, [%rd58+56];

dp4a.s32.s32 %r308, %r525, %r310, %r587;


dp4a.s32.s32 %r312, %r529, %r314, %r308;


dp4a.s32.s32 %r316, %r589, %r310, %r587;


dp4a.s32.s32 %r320, %r589, %r314, %r316;

mul.lo.s32 %r666, %r312, %r611;
cvt.rn.f32.s32 %f207, %r666;
fma.rn.ftz.f32 %f208, %f109, %f207, 0f00000000;
mul.lo.s32 %r667, %r320, %r614;
cvt.rn.f32.s32 %f209, %r667;
fma.rn.ftz.f32 %f210, %f109, %f209, 0f00000000;

dp4a.s32.s32 %r324, %r541, %r326, %r587;


dp4a.s32.s32 %r328, %r545, %r330, %r324;


dp4a.s32.s32 %r332, %r589, %r326, %r587;


dp4a.s32.s32 %r336, %r589, %r330, %r332;

mul.lo.s32 %r668, %r328, %r618;
cvt.rn.f32.s32 %f211, %r668;
fma.rn.ftz.f32 %f212, %f110, %f211, %f208;
mul.lo.s32 %r669, %r336, %r620;
cvt.rn.f32.s32 %f213, %r669;
fma.rn.ftz.f32 %f214, %f110, %f213, %f210;

{.reg .f16 low,high;
mov.b32 {low,high},%r124;
cvt.f32.f16 %f111, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r124;
cvt.f32.f16 %f112, high;}


mul.ftz.f32 %f215, %f212, %f111;
mul.ftz.f32 %f216, %f214, %f112;
sub.ftz.f32 %f217, %f215, %f216;
add.ftz.f32 %f470, %f217, %f470;

{.reg .f16 low,high;
mov.b32 {low,high},%r306;
cvt.f32.f16 %f113, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r307;
cvt.f32.f16 %f114, low;}



dp4a.s32.s32 %r344, %r561, %r310, %r587;


dp4a.s32.s32 %r348, %r565, %r314, %r344;


dp4a.s32.s32 %r352, %r589, %r310, %r587;


dp4a.s32.s32 %r356, %r589, %r314, %r352;

mul.lo.s32 %r670, %r348, %r639;
cvt.rn.f32.s32 %f218, %r670;
fma.rn.ftz.f32 %f219, %f113, %f218, 0f00000000;
mul.lo.s32 %r671, %r356, %r642;
cvt.rn.f32.s32 %f220, %r671;
fma.rn.ftz.f32 %f221, %f113, %f220, 0f00000000;

dp4a.s32.s32 %r360, %r577, %r326, %r587;


dp4a.s32.s32 %r364, %r581, %r330, %r360;


dp4a.s32.s32 %r368, %r589, %r326, %r587;


dp4a.s32.s32 %r372, %r589, %r330, %r368;

mul.lo.s32 %r672, %r364, %r646;
cvt.rn.f32.s32 %f222, %r672;
fma.rn.ftz.f32 %f223, %f114, %f222, %f219;
mul.lo.s32 %r673, %r372, %r648;
cvt.rn.f32.s32 %f224, %r673;
fma.rn.ftz.f32 %f225, %f114, %f224, %f221;

{.reg .f16 low,high;
mov.b32 {low,high},%r160;
cvt.f32.f16 %f115, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r160;
cvt.f32.f16 %f116, high;}


mul.ftz.f32 %f226, %f223, %f115;
mul.ftz.f32 %f227, %f225, %f116;
sub.ftz.f32 %f228, %f226, %f227;
add.ftz.f32 %f469, %f228, %f469;
cvt.u64.u32 %rd59, %r960;
add.s64 %rd60, %rd2, %rd59;
mul.lo.s64 %rd61, %rd60, 36;
add.s64 %rd62, %rd10, %rd61;
ld.global.nc.u32 %r378, [%rd62];

{.reg .f16 low,high;
mov.b32 {low,high},%r378;
cvt.f32.f16 %f117, low;}


add.s64 %rd63, %rd62, %rd38;
ld.global.nc.u32 %r382, [%rd63+4];
ld.global.nc.u32 %r386, [%rd63+20];
ld.global.nc.u32 %r379, [%rd62+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r379;
cvt.f32.f16 %f118, low;}


ld.global.nc.u32 %r398, [%rd63+40];
ld.global.nc.u32 %r402, [%rd63+56];

dp4a.s32.s32 %r380, %r525, %r382, %r587;


dp4a.s32.s32 %r384, %r529, %r386, %r380;


dp4a.s32.s32 %r388, %r589, %r382, %r587;


dp4a.s32.s32 %r392, %r589, %r386, %r388;

mul.lo.s32 %r674, %r384, %r611;
cvt.rn.f32.s32 %f229, %r674;
fma.rn.ftz.f32 %f230, %f117, %f229, 0f00000000;
mul.lo.s32 %r675, %r392, %r614;
cvt.rn.f32.s32 %f231, %r675;
fma.rn.ftz.f32 %f232, %f117, %f231, 0f00000000;

dp4a.s32.s32 %r396, %r541, %r398, %r587;


dp4a.s32.s32 %r400, %r545, %r402, %r396;


dp4a.s32.s32 %r404, %r589, %r398, %r587;


dp4a.s32.s32 %r408, %r589, %r402, %r404;

mul.lo.s32 %r676, %r400, %r618;
cvt.rn.f32.s32 %f233, %r676;
fma.rn.ftz.f32 %f234, %f118, %f233, %f230;
mul.lo.s32 %r677, %r408, %r620;
cvt.rn.f32.s32 %f235, %r677;
fma.rn.ftz.f32 %f236, %f118, %f235, %f232;

{.reg .f16 low,high;
mov.b32 {low,high},%r124;
cvt.f32.f16 %f119, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r124;
cvt.f32.f16 %f120, high;}


mul.ftz.f32 %f237, %f234, %f119;
mul.ftz.f32 %f238, %f236, %f120;
sub.ftz.f32 %f239, %f237, %f238;
add.ftz.f32 %f468, %f239, %f468;

{.reg .f16 low,high;
mov.b32 {low,high},%r378;
cvt.f32.f16 %f121, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r379;
cvt.f32.f16 %f122, low;}



dp4a.s32.s32 %r416, %r561, %r382, %r587;


dp4a.s32.s32 %r420, %r565, %r386, %r416;


dp4a.s32.s32 %r424, %r589, %r382, %r587;


dp4a.s32.s32 %r428, %r589, %r386, %r424;

mul.lo.s32 %r678, %r420, %r639;
cvt.rn.f32.s32 %f240, %r678;
fma.rn.ftz.f32 %f241, %f121, %f240, 0f00000000;
mul.lo.s32 %r679, %r428, %r642;
cvt.rn.f32.s32 %f242, %r679;
fma.rn.ftz.f32 %f243, %f121, %f242, 0f00000000;

dp4a.s32.s32 %r432, %r577, %r398, %r587;


dp4a.s32.s32 %r436, %r581, %r402, %r432;


dp4a.s32.s32 %r440, %r589, %r398, %r587;


dp4a.s32.s32 %r444, %r589, %r402, %r440;

mul.lo.s32 %r680, %r436, %r646;
cvt.rn.f32.s32 %f244, %r680;
fma.rn.ftz.f32 %f245, %f122, %f244, %f241;
mul.lo.s32 %r681, %r444, %r648;
cvt.rn.f32.s32 %f246, %r681;
fma.rn.ftz.f32 %f247, %f122, %f246, %f243;

{.reg .f16 low,high;
mov.b32 {low,high},%r160;
cvt.f32.f16 %f123, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r160;
cvt.f32.f16 %f124, high;}


mul.ftz.f32 %f248, %f245, %f123;
mul.ftz.f32 %f249, %f247, %f124;
sub.ftz.f32 %f250, %f248, %f249;
add.ftz.f32 %f467, %f250, %f467;
cvt.u64.u32 %rd64, %r959;
add.s64 %rd65, %rd2, %rd64;
mul.lo.s64 %rd66, %rd65, 36;
add.s64 %rd67, %rd10, %rd66;
ld.global.nc.u32 %r450, [%rd67];

{.reg .f16 low,high;
mov.b32 {low,high},%r450;
cvt.f32.f16 %f125, low;}


add.s64 %rd68, %rd67, %rd38;
ld.global.nc.u32 %r454, [%rd68+4];
ld.global.nc.u32 %r458, [%rd68+20];
ld.global.nc.u32 %r451, [%rd67+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r451;
cvt.f32.f16 %f126, low;}


ld.global.nc.u32 %r470, [%rd68+40];
ld.global.nc.u32 %r474, [%rd68+56];

dp4a.s32.s32 %r452, %r525, %r454, %r587;


dp4a.s32.s32 %r456, %r529, %r458, %r452;


dp4a.s32.s32 %r460, %r589, %r454, %r587;


dp4a.s32.s32 %r464, %r589, %r458, %r460;

mul.lo.s32 %r682, %r456, %r611;
cvt.rn.f32.s32 %f251, %r682;
fma.rn.ftz.f32 %f252, %f125, %f251, 0f00000000;
mul.lo.s32 %r683, %r464, %r614;
cvt.rn.f32.s32 %f253, %r683;
fma.rn.ftz.f32 %f254, %f125, %f253, 0f00000000;

dp4a.s32.s32 %r468, %r541, %r470, %r587;


dp4a.s32.s32 %r472, %r545, %r474, %r468;


dp4a.s32.s32 %r476, %r589, %r470, %r587;


dp4a.s32.s32 %r480, %r589, %r474, %r476;

mul.lo.s32 %r684, %r472, %r618;
cvt.rn.f32.s32 %f255, %r684;
fma.rn.ftz.f32 %f256, %f126, %f255, %f252;
mul.lo.s32 %r685, %r480, %r620;
cvt.rn.f32.s32 %f257, %r685;
fma.rn.ftz.f32 %f258, %f126, %f257, %f254;

{.reg .f16 low,high;
mov.b32 {low,high},%r124;
cvt.f32.f16 %f127, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r124;
cvt.f32.f16 %f128, high;}


mul.ftz.f32 %f259, %f256, %f127;
mul.ftz.f32 %f260, %f258, %f128;
sub.ftz.f32 %f261, %f259, %f260;
add.ftz.f32 %f466, %f261, %f466;

{.reg .f16 low,high;
mov.b32 {low,high},%r450;
cvt.f32.f16 %f129, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r451;
cvt.f32.f16 %f130, low;}



dp4a.s32.s32 %r488, %r561, %r454, %r587;


dp4a.s32.s32 %r492, %r565, %r458, %r488;


dp4a.s32.s32 %r496, %r589, %r454, %r587;


dp4a.s32.s32 %r500, %r589, %r458, %r496;

mul.lo.s32 %r686, %r492, %r639;
cvt.rn.f32.s32 %f262, %r686;
fma.rn.ftz.f32 %f263, %f129, %f262, 0f00000000;
mul.lo.s32 %r687, %r500, %r642;
cvt.rn.f32.s32 %f264, %r687;
fma.rn.ftz.f32 %f265, %f129, %f264, 0f00000000;

dp4a.s32.s32 %r504, %r577, %r470, %r587;


dp4a.s32.s32 %r508, %r581, %r474, %r504;


dp4a.s32.s32 %r512, %r589, %r470, %r587;


dp4a.s32.s32 %r516, %r589, %r474, %r512;

mul.lo.s32 %r688, %r508, %r646;
cvt.rn.f32.s32 %f266, %r688;
fma.rn.ftz.f32 %f267, %f130, %f266, %f263;
mul.lo.s32 %r689, %r516, %r648;
cvt.rn.f32.s32 %f268, %r689;
fma.rn.ftz.f32 %f269, %f130, %f268, %f265;

{.reg .f16 low,high;
mov.b32 {low,high},%r160;
cvt.f32.f16 %f131, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r160;
cvt.f32.f16 %f132, high;}


mul.ftz.f32 %f270, %f267, %f131;
mul.ftz.f32 %f271, %f269, %f132;
sub.ftz.f32 %f272, %f270, %f271;
add.ftz.f32 %f465, %f272, %f465;
cvt.u64.u32 %rd69, %r958;
add.s64 %rd70, %rd2, %rd69;
mul.lo.s64 %rd71, %rd70, 36;
add.s64 %rd72, %rd10, %rd71;
ld.global.nc.u32 %r522, [%rd72];

{.reg .f16 low,high;
mov.b32 {low,high},%r522;
cvt.f32.f16 %f133, low;}


add.s64 %rd73, %rd72, %rd38;
ld.global.nc.u32 %r526, [%rd73+4];
ld.global.nc.u32 %r530, [%rd73+20];
ld.global.nc.u32 %r523, [%rd72+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r523;
cvt.f32.f16 %f134, low;}


ld.global.nc.u32 %r542, [%rd73+40];
ld.global.nc.u32 %r546, [%rd73+56];

dp4a.s32.s32 %r524, %r525, %r526, %r587;


dp4a.s32.s32 %r528, %r529, %r530, %r524;


dp4a.s32.s32 %r532, %r589, %r526, %r587;


dp4a.s32.s32 %r536, %r589, %r530, %r532;

mul.lo.s32 %r690, %r528, %r611;
cvt.rn.f32.s32 %f273, %r690;
fma.rn.ftz.f32 %f274, %f133, %f273, 0f00000000;
mul.lo.s32 %r691, %r536, %r614;
cvt.rn.f32.s32 %f275, %r691;
fma.rn.ftz.f32 %f276, %f133, %f275, 0f00000000;

dp4a.s32.s32 %r540, %r541, %r542, %r587;


dp4a.s32.s32 %r544, %r545, %r546, %r540;


dp4a.s32.s32 %r548, %r589, %r542, %r587;


dp4a.s32.s32 %r552, %r589, %r546, %r548;

mul.lo.s32 %r692, %r544, %r618;
cvt.rn.f32.s32 %f277, %r692;
fma.rn.ftz.f32 %f278, %f134, %f277, %f274;
mul.lo.s32 %r693, %r552, %r620;
cvt.rn.f32.s32 %f279, %r693;
fma.rn.ftz.f32 %f280, %f134, %f279, %f276;

{.reg .f16 low,high;
mov.b32 {low,high},%r124;
cvt.f32.f16 %f135, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r124;
cvt.f32.f16 %f136, high;}


mul.ftz.f32 %f281, %f278, %f135;
mul.ftz.f32 %f282, %f280, %f136;
sub.ftz.f32 %f283, %f281, %f282;
add.ftz.f32 %f464, %f283, %f464;

{.reg .f16 low,high;
mov.b32 {low,high},%r522;
cvt.f32.f16 %f137, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r523;
cvt.f32.f16 %f138, low;}



dp4a.s32.s32 %r560, %r561, %r526, %r587;


dp4a.s32.s32 %r564, %r565, %r530, %r560;


dp4a.s32.s32 %r568, %r589, %r526, %r587;


dp4a.s32.s32 %r572, %r589, %r530, %r568;

mul.lo.s32 %r694, %r564, %r639;
cvt.rn.f32.s32 %f284, %r694;
fma.rn.ftz.f32 %f285, %f137, %f284, 0f00000000;
mul.lo.s32 %r695, %r572, %r642;
cvt.rn.f32.s32 %f286, %r695;
fma.rn.ftz.f32 %f287, %f137, %f286, 0f00000000;

dp4a.s32.s32 %r576, %r577, %r542, %r587;


dp4a.s32.s32 %r580, %r581, %r546, %r576;


dp4a.s32.s32 %r584, %r589, %r542, %r587;


dp4a.s32.s32 %r588, %r589, %r546, %r584;

mul.lo.s32 %r696, %r580, %r646;
cvt.rn.f32.s32 %f288, %r696;
fma.rn.ftz.f32 %f289, %f138, %f288, %f285;
mul.lo.s32 %r697, %r588, %r648;
cvt.rn.f32.s32 %f290, %r697;
fma.rn.ftz.f32 %f291, %f138, %f290, %f287;

{.reg .f16 low,high;
mov.b32 {low,high},%r160;
cvt.f32.f16 %f139, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r160;
cvt.f32.f16 %f140, high;}


mul.ftz.f32 %f292, %f289, %f139;
mul.ftz.f32 %f293, %f291, %f140;
sub.ftz.f32 %f294, %f292, %f293;
add.ftz.f32 %f463, %f294, %f463;
add.s32 %r962, %r962, 32;
add.s32 %r961, %r961, 32;
add.s32 %r960, %r960, 32;
add.s32 %r959, %r959, 32;
add.s32 %r958, %r958, 32;
add.s32 %r957, %r957, 32;
add.s32 %r956, %r956, 32;
add.s64 %rd118, %rd118, 576;
add.s32 %r963, %r963, 4;
setp.lt.s32 %p5, %r963, %r6;
@%p5 bra $L__BB166_2;

$L__BB166_3:
setp.eq.s32 %p6, %r1, 0;
@%p6 bra $L__BB166_5;

mov.u32 %r698, _ZZ13mul_mat_vec_qIL9ggml_type12ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r699, %r1, 1792, %r698;
shl.b32 %r700, %r2, 2;
add.s32 %r701, %r699, %r700;
st.shared.f32 [%r701+-1792], %f476;
st.shared.f32 [%r701+-1664], %f475;
st.shared.f32 [%r701+-1536], %f474;
st.shared.f32 [%r701+-1408], %f473;
st.shared.f32 [%r701+-1280], %f472;
st.shared.f32 [%r701+-1152], %f471;
st.shared.f32 [%r701+-1024], %f470;
st.shared.f32 [%r701+-896], %f469;
st.shared.f32 [%r701+-768], %f468;
st.shared.f32 [%r701+-640], %f467;
st.shared.f32 [%r701+-512], %f466;
st.shared.f32 [%r701+-384], %f465;
st.shared.f32 [%r701+-256], %f464;
st.shared.f32 [%r701+-128], %f463;

$L__BB166_5:
setp.ne.s32 %p7, %r1, 0;
bar.sync 0;
@%p7 bra $L__BB166_34;

mov.u32 %r702, %ctaid.z;
mov.u32 %r703, %ctaid.y;
mad.lo.s32 %r704, %r703, %r42, %r4;
mad.lo.s32 %r705, %r702, %r48, %r704;
cvt.u64.u32 %rd13, %r705;
shl.b32 %r706, %r2, 2;
mov.u32 %r707, _ZZ13mul_mat_vec_qIL9ggml_type12ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r31, %r707, %r706;
mov.u32 %r708, 2;
ld.shared.f32 %f295, [%r31];
add.ftz.f32 %f296, %f295, %f476;
mov.b32 %r709, %f296;
mov.u32 %r710, 31;
mov.u32 %r711, 16;
mov.u32 %r712, -1;
shfl.sync.bfly.b32 %r713|%p8, %r709, %r711, %r710, %r712;
mov.b32 %f297, %r713;
add.ftz.f32 %f298, %f296, %f297;
mov.b32 %r714, %f298;
mov.u32 %r715, 8;
shfl.sync.bfly.b32 %r716|%p9, %r714, %r715, %r710, %r712;
mov.b32 %f299, %r716;
add.ftz.f32 %f300, %f298, %f299;
mov.b32 %r717, %f300;
mov.u32 %r718, 4;
shfl.sync.bfly.b32 %r719|%p10, %r717, %r718, %r710, %r712;
mov.b32 %f301, %r719;
add.ftz.f32 %f302, %f300, %f301;
mov.b32 %r720, %f302;
shfl.sync.bfly.b32 %r721|%p11, %r720, %r708, %r710, %r712;
mov.b32 %f303, %r721;
add.ftz.f32 %f304, %f302, %f303;
mov.b32 %r722, %f304;
mov.u32 %r723, 1;
shfl.sync.bfly.b32 %r724|%p12, %r722, %r723, %r710, %r712;
mov.b32 %f305, %r724;
add.ftz.f32 %f43, %f304, %f305;
setp.lt.u32 %p13, %r4, %r36;
setp.eq.s32 %p14, %r2, 0;
and.pred %p1, %p14, %p13;
cvta.to.global.u64 %rd74, %rd20;
mul.wide.u32 %rd75, %r705, 4;
add.s64 %rd14, %rd74, %rd75;
not.pred %p15, %p1;
@%p15 bra $L__BB166_8;

st.global.f32 [%rd14], %f43;

$L__BB166_8:
ld.shared.f32 %f306, [%r31+128];
add.ftz.f32 %f307, %f306, %f475;
mov.b32 %r725, %f307;
shfl.sync.bfly.b32 %r729|%p16, %r725, %r711, %r710, %r712;
mov.b32 %f308, %r729;
add.ftz.f32 %f309, %f307, %f308;
mov.b32 %r730, %f309;
shfl.sync.bfly.b32 %r732|%p17, %r730, %r715, %r710, %r712;
mov.b32 %f310, %r732;
add.ftz.f32 %f311, %f309, %f310;
mov.b32 %r733, %f311;
shfl.sync.bfly.b32 %r735|%p18, %r733, %r718, %r710, %r712;
mov.b32 %f312, %r735;
add.ftz.f32 %f313, %f311, %f312;
mov.b32 %r736, %f313;
shfl.sync.bfly.b32 %r738|%p19, %r736, %r708, %r710, %r712;
mov.b32 %f314, %r738;
add.ftz.f32 %f315, %f313, %f314;
mov.b32 %r739, %f315;
shfl.sync.bfly.b32 %r741|%p20, %r739, %r723, %r710, %r712;
mov.b32 %f316, %r741;
add.ftz.f32 %f44, %f315, %f316;
setp.eq.s32 %p21, %r2, 1;
or.b32 %r745, %r4, 1;
setp.lt.u32 %p22, %r745, %r36;
and.pred %p2, %p21, %p22;
not.pred %p23, %p2;
@%p23 bra $L__BB166_10;

st.global.f32 [%rd14+4], %f44;

$L__BB166_10:
ld.shared.f32 %f317, [%r31+256];
add.ftz.f32 %f318, %f317, %f474;
mov.b32 %r746, %f318;
mov.u32 %r747, 31;
mov.u32 %r748, 16;
mov.u32 %r749, -1;
shfl.sync.bfly.b32 %r750|%p24, %r746, %r748, %r747, %r749;
mov.b32 %f319, %r750;
add.ftz.f32 %f320, %f318, %f319;
mov.b32 %r751, %f320;
mov.u32 %r752, 8;
shfl.sync.bfly.b32 %r753|%p25, %r751, %r752, %r747, %r749;
mov.b32 %f321, %r753;
add.ftz.f32 %f322, %f320, %f321;
mov.b32 %r754, %f322;
mov.u32 %r755, 4;
shfl.sync.bfly.b32 %r756|%p26, %r754, %r755, %r747, %r749;
mov.b32 %f323, %r756;
add.ftz.f32 %f324, %f322, %f323;
mov.b32 %r757, %f324;
mov.u32 %r758, 2;
shfl.sync.bfly.b32 %r759|%p27, %r757, %r758, %r747, %r749;
mov.b32 %f325, %r759;
add.ftz.f32 %f326, %f324, %f325;
mov.b32 %r760, %f326;
mov.u32 %r761, 1;
shfl.sync.bfly.b32 %r762|%p28, %r760, %r761, %r747, %r749;
mov.b32 %f327, %r762;
add.ftz.f32 %f45, %f326, %f327;
@%p15 bra $L__BB166_12;

cvt.u64.u32 %rd76, %r36;
add.s64 %rd77, %rd76, %rd13;
shl.b64 %rd79, %rd77, 2;
add.s64 %rd80, %rd74, %rd79;
st.global.f32 [%rd80], %f45;

$L__BB166_12:
ld.shared.f32 %f328, [%r31+384];
add.ftz.f32 %f329, %f328, %f473;
mov.b32 %r763, %f329;
shfl.sync.bfly.b32 %r767|%p30, %r763, %r748, %r747, %r749;
mov.b32 %f330, %r767;
add.ftz.f32 %f331, %f329, %f330;
mov.b32 %r768, %f331;
shfl.sync.bfly.b32 %r770|%p31, %r768, %r752, %r747, %r749;
mov.b32 %f332, %r770;
add.ftz.f32 %f333, %f331, %f332;
mov.b32 %r771, %f333;
shfl.sync.bfly.b32 %r773|%p32, %r771, %r755, %r747, %r749;
mov.b32 %f334, %r773;
add.ftz.f32 %f335, %f333, %f334;
mov.b32 %r774, %f335;
shfl.sync.bfly.b32 %r776|%p33, %r774, %r758, %r747, %r749;
mov.b32 %f336, %r776;
add.ftz.f32 %f337, %f335, %f336;
mov.b32 %r777, %f337;
shfl.sync.bfly.b32 %r779|%p34, %r777, %r761, %r747, %r749;
mov.b32 %f338, %r779;
add.ftz.f32 %f46, %f337, %f338;
@%p23 bra $L__BB166_14;

add.s32 %r780, %r36, 1;
cvt.u64.u32 %rd81, %r780;
add.s64 %rd82, %rd81, %rd13;
shl.b64 %rd84, %rd82, 2;
add.s64 %rd85, %rd74, %rd84;
st.global.f32 [%rd85], %f46;

$L__BB166_14:
shl.b32 %r781, %r36, 1;
mov.u32 %r782, 1;
ld.shared.f32 %f339, [%r31+512];
add.ftz.f32 %f340, %f339, %f472;
mov.b32 %r783, %f340;
mov.u32 %r784, 31;
mov.u32 %r785, 16;
mov.u32 %r786, -1;
shfl.sync.bfly.b32 %r787|%p36, %r783, %r785, %r784, %r786;
mov.b32 %f341, %r787;
add.ftz.f32 %f342, %f340, %f341;
mov.b32 %r788, %f342;
mov.u32 %r789, 8;
shfl.sync.bfly.b32 %r790|%p37, %r788, %r789, %r784, %r786;
mov.b32 %f343, %r790;
add.ftz.f32 %f344, %f342, %f343;
mov.b32 %r791, %f344;
mov.u32 %r792, 4;
shfl.sync.bfly.b32 %r793|%p38, %r791, %r792, %r784, %r786;
mov.b32 %f345, %r793;
add.ftz.f32 %f346, %f344, %f345;
mov.b32 %r794, %f346;
mov.u32 %r795, 2;
shfl.sync.bfly.b32 %r796|%p39, %r794, %r795, %r784, %r786;
mov.b32 %f347, %r796;
add.ftz.f32 %f348, %f346, %f347;
mov.b32 %r797, %f348;
shfl.sync.bfly.b32 %r798|%p40, %r797, %r782, %r784, %r786;
mov.b32 %f349, %r798;
add.ftz.f32 %f47, %f348, %f349;
cvt.u64.u32 %rd86, %r781;
add.s64 %rd87, %rd86, %rd13;
shl.b64 %rd89, %rd87, 2;
add.s64 %rd15, %rd74, %rd89;
@%p15 bra $L__BB166_16;

st.global.f32 [%rd15], %f47;

$L__BB166_16:
ld.shared.f32 %f350, [%r31+640];
add.ftz.f32 %f351, %f350, %f471;
mov.b32 %r799, %f351;
shfl.sync.bfly.b32 %r803|%p42, %r799, %r785, %r784, %r786;
mov.b32 %f352, %r803;
add.ftz.f32 %f353, %f351, %f352;
mov.b32 %r804, %f353;
shfl.sync.bfly.b32 %r806|%p43, %r804, %r789, %r784, %r786;
mov.b32 %f354, %r806;
add.ftz.f32 %f355, %f353, %f354;
mov.b32 %r807, %f355;
shfl.sync.bfly.b32 %r809|%p44, %r807, %r792, %r784, %r786;
mov.b32 %f356, %r809;
add.ftz.f32 %f357, %f355, %f356;
mov.b32 %r810, %f357;
shfl.sync.bfly.b32 %r812|%p45, %r810, %r795, %r784, %r786;
mov.b32 %f358, %r812;
add.ftz.f32 %f359, %f357, %f358;
mov.b32 %r813, %f359;
shfl.sync.bfly.b32 %r815|%p46, %r813, %r782, %r784, %r786;
mov.b32 %f360, %r815;
add.ftz.f32 %f48, %f359, %f360;
@%p23 bra $L__BB166_18;

st.global.f32 [%rd15+4], %f48;

$L__BB166_18:
mul.lo.s32 %r32, %r36, 3;
ld.shared.f32 %f361, [%r31+768];
add.ftz.f32 %f362, %f361, %f470;
mov.b32 %r816, %f362;
mov.u32 %r817, 31;
mov.u32 %r818, 16;
mov.u32 %r819, -1;
shfl.sync.bfly.b32 %r820|%p48, %r816, %r818, %r817, %r819;
mov.b32 %f363, %r820;
add.ftz.f32 %f364, %f362, %f363;
mov.b32 %r821, %f364;
mov.u32 %r822, 8;
shfl.sync.bfly.b32 %r823|%p49, %r821, %r822, %r817, %r819;
mov.b32 %f365, %r823;
add.ftz.f32 %f366, %f364, %f365;
mov.b32 %r824, %f366;
mov.u32 %r825, 4;
shfl.sync.bfly.b32 %r826|%p50, %r824, %r825, %r817, %r819;
mov.b32 %f367, %r826;
add.ftz.f32 %f368, %f366, %f367;
mov.b32 %r827, %f368;
mov.u32 %r828, 2;
shfl.sync.bfly.b32 %r829|%p51, %r827, %r828, %r817, %r819;
mov.b32 %f369, %r829;
add.ftz.f32 %f370, %f368, %f369;
mov.b32 %r830, %f370;
mov.u32 %r831, 1;
shfl.sync.bfly.b32 %r832|%p52, %r830, %r831, %r817, %r819;
mov.b32 %f371, %r832;
add.ftz.f32 %f49, %f370, %f371;
@%p15 bra $L__BB166_20;

cvt.u64.u32 %rd90, %r32;
add.s64 %rd91, %rd90, %rd13;
shl.b64 %rd93, %rd91, 2;
add.s64 %rd94, %rd74, %rd93;
st.global.f32 [%rd94], %f49;

$L__BB166_20:
ld.shared.f32 %f372, [%r31+896];
add.ftz.f32 %f373, %f372, %f469;
mov.b32 %r833, %f373;
shfl.sync.bfly.b32 %r837|%p54, %r833, %r818, %r817, %r819;
mov.b32 %f374, %r837;
add.ftz.f32 %f375, %f373, %f374;
mov.b32 %r838, %f375;
shfl.sync.bfly.b32 %r840|%p55, %r838, %r822, %r817, %r819;
mov.b32 %f376, %r840;
add.ftz.f32 %f377, %f375, %f376;
mov.b32 %r841, %f377;
shfl.sync.bfly.b32 %r843|%p56, %r841, %r825, %r817, %r819;
mov.b32 %f378, %r843;
add.ftz.f32 %f379, %f377, %f378;
mov.b32 %r844, %f379;
shfl.sync.bfly.b32 %r846|%p57, %r844, %r828, %r817, %r819;
mov.b32 %f380, %r846;
add.ftz.f32 %f381, %f379, %f380;
mov.b32 %r847, %f381;
shfl.sync.bfly.b32 %r849|%p58, %r847, %r831, %r817, %r819;
mov.b32 %f382, %r849;
add.ftz.f32 %f50, %f381, %f382;
@%p23 bra $L__BB166_22;

add.s32 %r850, %r32, 1;
cvt.u64.u32 %rd95, %r850;
add.s64 %rd96, %rd95, %rd13;
shl.b64 %rd98, %rd96, 2;
add.s64 %rd99, %rd74, %rd98;
st.global.f32 [%rd99], %f50;

$L__BB166_22:
shl.b32 %r851, %r36, 2;
mov.u32 %r852, 2;
ld.shared.f32 %f383, [%r31+1024];
add.ftz.f32 %f384, %f383, %f468;
mov.b32 %r853, %f384;
mov.u32 %r854, 31;
mov.u32 %r855, 16;
mov.u32 %r856, -1;
shfl.sync.bfly.b32 %r857|%p60, %r853, %r855, %r854, %r856;
mov.b32 %f385, %r857;
add.ftz.f32 %f386, %f384, %f385;
mov.b32 %r858, %f386;
mov.u32 %r859, 8;
shfl.sync.bfly.b32 %r860|%p61, %r858, %r859, %r854, %r856;
mov.b32 %f387, %r860;
add.ftz.f32 %f388, %f386, %f387;
mov.b32 %r861, %f388;
mov.u32 %r862, 4;
shfl.sync.bfly.b32 %r863|%p62, %r861, %r862, %r854, %r856;
mov.b32 %f389, %r863;
add.ftz.f32 %f390, %f388, %f389;
mov.b32 %r864, %f390;
shfl.sync.bfly.b32 %r865|%p63, %r864, %r852, %r854, %r856;
mov.b32 %f391, %r865;
add.ftz.f32 %f392, %f390, %f391;
mov.b32 %r866, %f392;
mov.u32 %r867, 1;
shfl.sync.bfly.b32 %r868|%p64, %r866, %r867, %r854, %r856;
mov.b32 %f393, %r868;
add.ftz.f32 %f51, %f392, %f393;
cvt.u64.u32 %rd100, %r851;
add.s64 %rd101, %rd100, %rd13;
shl.b64 %rd103, %rd101, 2;
add.s64 %rd16, %rd74, %rd103;
@%p15 bra $L__BB166_24;

st.global.f32 [%rd16], %f51;

$L__BB166_24:
ld.shared.f32 %f394, [%r31+1152];
add.ftz.f32 %f395, %f394, %f467;
mov.b32 %r869, %f395;
shfl.sync.bfly.b32 %r873|%p66, %r869, %r855, %r854, %r856;
mov.b32 %f396, %r873;
add.ftz.f32 %f397, %f395, %f396;
mov.b32 %r874, %f397;
shfl.sync.bfly.b32 %r876|%p67, %r874, %r859, %r854, %r856;
mov.b32 %f398, %r876;
add.ftz.f32 %f399, %f397, %f398;
mov.b32 %r877, %f399;
shfl.sync.bfly.b32 %r879|%p68, %r877, %r862, %r854, %r856;
mov.b32 %f400, %r879;
add.ftz.f32 %f401, %f399, %f400;
mov.b32 %r880, %f401;
shfl.sync.bfly.b32 %r882|%p69, %r880, %r852, %r854, %r856;
mov.b32 %f402, %r882;
add.ftz.f32 %f403, %f401, %f402;
mov.b32 %r883, %f403;
shfl.sync.bfly.b32 %r885|%p70, %r883, %r867, %r854, %r856;
mov.b32 %f404, %r885;
add.ftz.f32 %f52, %f403, %f404;
@%p23 bra $L__BB166_26;

st.global.f32 [%rd16+4], %f52;

$L__BB166_26:
mul.lo.s32 %r33, %r36, 5;
ld.shared.f32 %f405, [%r31+1280];
add.ftz.f32 %f406, %f405, %f466;
mov.b32 %r886, %f406;
mov.u32 %r887, 31;
mov.u32 %r888, 16;
mov.u32 %r889, -1;
shfl.sync.bfly.b32 %r890|%p72, %r886, %r888, %r887, %r889;
mov.b32 %f407, %r890;
add.ftz.f32 %f408, %f406, %f407;
mov.b32 %r891, %f408;
mov.u32 %r892, 8;
shfl.sync.bfly.b32 %r893|%p73, %r891, %r892, %r887, %r889;
mov.b32 %f409, %r893;
add.ftz.f32 %f410, %f408, %f409;
mov.b32 %r894, %f410;
mov.u32 %r895, 4;
shfl.sync.bfly.b32 %r896|%p74, %r894, %r895, %r887, %r889;
mov.b32 %f411, %r896;
add.ftz.f32 %f412, %f410, %f411;
mov.b32 %r897, %f412;
mov.u32 %r898, 2;
shfl.sync.bfly.b32 %r899|%p75, %r897, %r898, %r887, %r889;
mov.b32 %f413, %r899;
add.ftz.f32 %f414, %f412, %f413;
mov.b32 %r900, %f414;
mov.u32 %r901, 1;
shfl.sync.bfly.b32 %r902|%p76, %r900, %r901, %r887, %r889;
mov.b32 %f415, %r902;
add.ftz.f32 %f53, %f414, %f415;
@%p15 bra $L__BB166_28;

cvt.u64.u32 %rd104, %r33;
add.s64 %rd105, %rd104, %rd13;
shl.b64 %rd107, %rd105, 2;
add.s64 %rd108, %rd74, %rd107;
st.global.f32 [%rd108], %f53;

$L__BB166_28:
ld.shared.f32 %f416, [%r31+1408];
add.ftz.f32 %f417, %f416, %f465;
mov.b32 %r903, %f417;
shfl.sync.bfly.b32 %r907|%p78, %r903, %r888, %r887, %r889;
mov.b32 %f418, %r907;
add.ftz.f32 %f419, %f417, %f418;
mov.b32 %r908, %f419;
shfl.sync.bfly.b32 %r910|%p79, %r908, %r892, %r887, %r889;
mov.b32 %f420, %r910;
add.ftz.f32 %f421, %f419, %f420;
mov.b32 %r911, %f421;
shfl.sync.bfly.b32 %r913|%p80, %r911, %r895, %r887, %r889;
mov.b32 %f422, %r913;
add.ftz.f32 %f423, %f421, %f422;
mov.b32 %r914, %f423;
shfl.sync.bfly.b32 %r916|%p81, %r914, %r898, %r887, %r889;
mov.b32 %f424, %r916;
add.ftz.f32 %f425, %f423, %f424;
mov.b32 %r917, %f425;
shfl.sync.bfly.b32 %r919|%p82, %r917, %r901, %r887, %r889;
mov.b32 %f426, %r919;
add.ftz.f32 %f54, %f425, %f426;
@%p23 bra $L__BB166_30;

add.s32 %r920, %r33, 1;
cvt.u64.u32 %rd109, %r920;
add.s64 %rd110, %rd109, %rd13;
shl.b64 %rd112, %rd110, 2;
add.s64 %rd113, %rd74, %rd112;
st.global.f32 [%rd113], %f54;

$L__BB166_30:
mul.lo.s32 %r921, %r36, 6;
ld.shared.f32 %f427, [%r31+1536];
add.ftz.f32 %f428, %f427, %f464;
mov.b32 %r922, %f428;
mov.u32 %r923, 31;
mov.u32 %r924, 16;
mov.u32 %r925, -1;
shfl.sync.bfly.b32 %r926|%p84, %r922, %r924, %r923, %r925;
mov.b32 %f429, %r926;
add.ftz.f32 %f430, %f428, %f429;
mov.b32 %r927, %f430;
mov.u32 %r928, 8;
shfl.sync.bfly.b32 %r929|%p85, %r927, %r928, %r923, %r925;
mov.b32 %f431, %r929;
add.ftz.f32 %f432, %f430, %f431;
mov.b32 %r930, %f432;
mov.u32 %r931, 4;
shfl.sync.bfly.b32 %r932|%p86, %r930, %r931, %r923, %r925;
mov.b32 %f433, %r932;
add.ftz.f32 %f434, %f432, %f433;
mov.b32 %r933, %f434;
mov.u32 %r934, 2;
shfl.sync.bfly.b32 %r935|%p87, %r933, %r934, %r923, %r925;
mov.b32 %f435, %r935;
add.ftz.f32 %f436, %f434, %f435;
mov.b32 %r936, %f436;
mov.u32 %r937, 1;
shfl.sync.bfly.b32 %r938|%p88, %r936, %r937, %r923, %r925;
mov.b32 %f437, %r938;
add.ftz.f32 %f55, %f436, %f437;
cvt.u64.u32 %rd114, %r921;
add.s64 %rd115, %rd114, %rd13;
shl.b64 %rd117, %rd115, 2;
add.s64 %rd17, %rd74, %rd117;
@%p15 bra $L__BB166_32;

st.global.f32 [%rd17], %f55;

$L__BB166_32:
ld.shared.f32 %f438, [%r31+1664];
add.ftz.f32 %f439, %f438, %f463;
mov.b32 %r939, %f439;
shfl.sync.bfly.b32 %r943|%p90, %r939, %r924, %r923, %r925;
mov.b32 %f440, %r943;
add.ftz.f32 %f441, %f439, %f440;
mov.b32 %r944, %f441;
shfl.sync.bfly.b32 %r946|%p91, %r944, %r928, %r923, %r925;
mov.b32 %f442, %r946;
add.ftz.f32 %f443, %f441, %f442;
mov.b32 %r947, %f443;
shfl.sync.bfly.b32 %r949|%p92, %r947, %r931, %r923, %r925;
mov.b32 %f444, %r949;
add.ftz.f32 %f445, %f443, %f444;
mov.b32 %r950, %f445;
shfl.sync.bfly.b32 %r952|%p93, %r950, %r934, %r923, %r925;
mov.b32 %f446, %r952;
add.ftz.f32 %f447, %f445, %f446;
mov.b32 %r953, %f447;
shfl.sync.bfly.b32 %r955|%p94, %r953, %r937, %r923, %r925;
mov.b32 %f448, %r955;
add.ftz.f32 %f56, %f447, %f448;
@%p23 bra $L__BB166_34;

st.global.f32 [%rd17+4], %f56;

$L__BB166_34:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type12ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type12ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type12ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type12ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type12ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type12ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type12ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type12ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type12ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type12ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<108>;
.reg .b16 %rs<15>;
.reg .f32 %f<545>;
.reg .b32 %r<1085>;
.reg .b64 %rd<134>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type12ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[2048];

ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type12ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type12ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd20, [_Z13mul_mat_vec_qIL9ggml_type12ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r53, [_Z13mul_mat_vec_qIL9ggml_type12ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type12ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type12ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type12ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type12ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r45, [_Z13mul_mat_vec_qIL9ggml_type12ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r46, [_Z13mul_mat_vec_qIL9ggml_type12ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r50, [_Z13mul_mat_vec_qIL9ggml_type12ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r51, [_Z13mul_mat_vec_qIL9ggml_type12ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r52, [_Z13mul_mat_vec_qIL9ggml_type12ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r48, [_Z13mul_mat_vec_qIL9ggml_type12ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r47, [_Z13mul_mat_vec_qIL9ggml_type12ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type12ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type12ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r54, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r54, %r2;
mov.u32 %r55, %ctaid.x;
shl.b32 %r4, %r55, 1;
shr.s32 %r56, %r3, 31;
shr.u32 %r57, %r56, 28;
add.s32 %r58, %r3, %r57;
shr.s32 %r1084, %r58, 4;
shr.u32 %r6, %r53, 8;
setp.ge.s32 %p3, %r1084, %r6;
mov.f32 %f529, 0f00000000;
mov.f32 %f530, %f529;
mov.f32 %f531, %f529;
mov.f32 %f532, %f529;
mov.f32 %f533, %f529;
mov.f32 %f534, %f529;
mov.f32 %f535, %f529;
mov.f32 %f536, %f529;
mov.f32 %f537, %f529;
mov.f32 %f538, %f529;
mov.f32 %f539, %f529;
mov.f32 %f540, %f529;
mov.f32 %f541, %f529;
mov.f32 %f542, %f529;
mov.f32 %f543, %f529;
mov.f32 %f544, %f529;
@%p3 bra $L__BB167_3;

mov.u32 %r59, %ctaid.y;
mov.u32 %r60, %ctaid.z;
mul.hi.u32 %r61, %r60, %r47;
add.s32 %r62, %r61, %r60;
shr.u32 %r63, %r62, %r48;
mul.lo.s32 %r64, %r63, %r50;
mul.hi.u32 %r65, %r59, %r41;
add.s32 %r66, %r65, %r59;
shr.u32 %r67, %r66, %r42;
mul.lo.s32 %r68, %r67, %r44;
mul.lo.s32 %r69, %r59, %r45;
cvt.u64.u32 %rd21, %r69;
cvta.to.global.u64 %rd133, %rd18;
and.b32 %r73, %r58, -16;
sub.s32 %r7, %r3, %r73;
shr.s32 %r74, %r7, 31;
shr.u32 %r75, %r74, 30;
add.s32 %r76, %r7, %r75;
shr.s32 %r77, %r76, 2;
shl.b32 %r78, %r77, 1;
shl.b32 %r79, %r77, 5;
and.b32 %r80, %r76, -4;
sub.s32 %r81, %r7, %r80;
shl.b32 %r82, %r81, 2;
add.s32 %r83, %r79, %r82;
cvt.s64.s32 %rd22, %r83;
cvt.s64.s32 %rd23, %r78;
add.s64 %rd24, %rd21, %rd23;
mul.lo.s32 %r84, %r60, %r51;
cvt.u64.u32 %rd25, %r84;
add.s64 %rd2, %rd24, %rd25;
cvt.s64.s32 %rd3, %r81;
shl.b32 %r85, %r39, 1;
shl.b32 %r1076, %r1084, 3;
add.s32 %r1083, %r1076, %r85;
mad.lo.s32 %r1082, %r39, 3, %r1076;
shl.b32 %r86, %r39, 2;
add.s32 %r1081, %r1076, %r86;
mad.lo.s32 %r1080, %r39, 5, %r1076;
mad.lo.s32 %r1079, %r39, 6, %r1076;
mad.lo.s32 %r1078, %r39, 7, %r1076;
add.s32 %r1077, %r39, %r1076;
add.s32 %r87, %r1084, %r64;
add.s32 %r88, %r87, %r68;
add.s32 %r89, %r4, 1;
mad.lo.s32 %r90, %r38, %r89, %r88;
mul.wide.s32 %rd4, %r90, 144;
mad.lo.s32 %r91, %r4, %r38, %r1084;
add.s32 %r92, %r91, %r64;
add.s32 %r93, %r92, %r68;
mul.wide.s32 %rd5, %r93, 144;
add.s64 %rd26, %rd5, %rd22;
add.s64 %rd6, %rd26, 32;
cvt.u64.u32 %rd27, %r78;
and.b64 %rd28, %rd27, 2;
or.b64 %rd29, %rd28, %rd4;
or.b64 %rd7, %rd29, 8;
or.b64 %rd30, %rd28, %rd5;
or.b64 %rd8, %rd30, 12;
add.s64 %rd31, %rd4, %rd22;
add.s64 %rd9, %rd31, 16;
cvta.to.global.u64 %rd10, %rd19;
mov.f32 %f529, 0f00000000;
shl.b64 %rd38, %rd3, 2;

$L__BB167_2:
cvt.u64.u32 %rd32, %r1076;
add.s64 %rd33, %rd2, %rd32;
add.s64 %rd34, %rd133, %rd6;
ld.global.nc.u32 %r670, [%rd34+-16];
ld.global.nc.u32 %r671, [%rd34];
add.s64 %rd35, %rd133, %rd8;
ld.global.nc.u16 %rs1, [%rd35+-8];
cvt.u32.u16 %r672, %rs1;
ld.global.nc.u16 %rs2, [%rd35+-4];
cvt.u32.u16 %r673, %rs2;
ld.global.nc.u16 %rs3, [%rd35];
cvt.u32.u16 %r674, %rs3;
and.b32 %r675, %r672, 16191;
and.b32 %r676, %r674, 3855;
and.b32 %r677, %r672, 49344;
shr.u32 %r678, %r677, 2;
or.b32 %r679, %r676, %r678;
setp.gt.s32 %p4, %r7, 7;
and.b32 %r680, %r673, 16191;
shr.u32 %r681, %r674, 4;
and.b32 %r682, %r681, 3855;
and.b32 %r683, %r673, 49344;
shr.u32 %r684, %r683, 2;
or.b32 %r685, %r682, %r684;
mul.lo.s64 %rd36, %rd33, 36;
add.s64 %rd37, %rd10, %rd36;
ld.global.nc.u32 %r94, [%rd37];

{.reg .f16 low,high;
mov.b32 {low,high},%r94;
cvt.f32.f16 %f97, low;}


add.s64 %rd39, %rd37, %rd38;
ld.global.nc.u32 %r98, [%rd39+4];
ld.global.nc.u32 %r102, [%rd39+20];
ld.global.nc.u32 %r95, [%rd37+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r95;
cvt.f32.f16 %f98, low;}


ld.global.nc.u32 %r114, [%rd39+40];
ld.global.nc.u32 %r118, [%rd39+56];
add.s64 %rd40, %rd133, %rd5;
and.b32 %r601, %r670, 252645135;
and.b32 %r605, %r671, 252645135;
mov.u32 %r663, 0;

dp4a.s32.s32 %r96, %r601, %r98, %r663;


dp4a.s32.s32 %r100, %r605, %r102, %r96;

mov.u32 %r665, 16843009;

dp4a.s32.s32 %r104, %r665, %r98, %r663;


dp4a.s32.s32 %r108, %r665, %r102, %r104;

selp.b32 %r686, %r679, %r675, %p4;
cvt.u16.u32 %rs4, %r686;
and.b32 %r687, %r686, 63;
mul.lo.s32 %r688, %r100, %r687;
cvt.rn.f32.s32 %f161, %r688;
fma.rn.ftz.f32 %f162, %f97, %f161, 0f00000000;
selp.b32 %r689, %r685, %r680, %p4;
cvt.u16.u32 %rs5, %r689;
and.b32 %r690, %r689, 63;
mul.lo.s32 %r691, %r108, %r690;
cvt.rn.f32.s32 %f163, %r691;
fma.rn.ftz.f32 %f164, %f97, %f163, 0f00000000;
shr.u32 %r692, %r670, 4;
and.b32 %r617, %r692, 252645135;
shr.u32 %r693, %r671, 4;
and.b32 %r621, %r693, 252645135;

dp4a.s32.s32 %r112, %r617, %r114, %r663;


dp4a.s32.s32 %r116, %r621, %r118, %r112;


dp4a.s32.s32 %r120, %r665, %r114, %r663;


dp4a.s32.s32 %r124, %r665, %r118, %r120;

shr.u16 %rs6, %rs4, 8;
cvt.u32.u16 %r694, %rs6;
mul.lo.s32 %r695, %r116, %r694;
cvt.rn.f32.s32 %f165, %r695;
fma.rn.ftz.f32 %f166, %f98, %f165, %f162;
shr.u16 %rs7, %rs5, 8;
cvt.u32.u16 %r696, %rs7;
mul.lo.s32 %r697, %r124, %r696;
cvt.rn.f32.s32 %f167, %r697;
fma.rn.ftz.f32 %f168, %f98, %f167, %f164;
ld.global.nc.u32 %r128, [%rd40];

{.reg .f16 low,high;
mov.b32 {low,high},%r128;
cvt.f32.f16 %f99, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r128;
cvt.f32.f16 %f100, high;}


mul.ftz.f32 %f169, %f166, %f99;
mul.ftz.f32 %f170, %f168, %f100;
sub.ftz.f32 %f171, %f169, %f170;
add.ftz.f32 %f544, %f171, %f544;
add.s64 %rd41, %rd133, %rd9;
ld.global.nc.u32 %r698, [%rd41];
ld.global.nc.u32 %r699, [%rd41+16];
add.s64 %rd42, %rd133, %rd7;
ld.global.nc.u16 %rs8, [%rd42+-4];
cvt.u32.u16 %r700, %rs8;
ld.global.nc.u16 %rs9, [%rd42];
cvt.u32.u16 %r701, %rs9;
ld.global.nc.u16 %rs10, [%rd42+4];
cvt.u32.u16 %r702, %rs10;
and.b32 %r703, %r700, 16191;
and.b32 %r704, %r702, 3855;
and.b32 %r705, %r700, 49344;
shr.u32 %r706, %r705, 2;
or.b32 %r707, %r704, %r706;
and.b32 %r708, %r701, 16191;
shr.u32 %r709, %r702, 4;
and.b32 %r710, %r709, 3855;
and.b32 %r711, %r701, 49344;
shr.u32 %r712, %r711, 2;
or.b32 %r713, %r710, %r712;

{.reg .f16 low,high;
mov.b32 {low,high},%r94;
cvt.f32.f16 %f101, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r95;
cvt.f32.f16 %f102, low;}


add.s64 %rd43, %rd133, %rd4;
and.b32 %r637, %r698, 252645135;
and.b32 %r641, %r699, 252645135;

dp4a.s32.s32 %r132, %r637, %r98, %r663;


dp4a.s32.s32 %r136, %r641, %r102, %r132;


dp4a.s32.s32 %r140, %r665, %r98, %r663;


dp4a.s32.s32 %r144, %r665, %r102, %r140;

selp.b32 %r714, %r707, %r703, %p4;
cvt.u16.u32 %rs11, %r714;
and.b32 %r715, %r714, 63;
mul.lo.s32 %r716, %r136, %r715;
cvt.rn.f32.s32 %f172, %r716;
fma.rn.ftz.f32 %f173, %f101, %f172, 0f00000000;
selp.b32 %r717, %r713, %r708, %p4;
cvt.u16.u32 %rs12, %r717;
and.b32 %r718, %r717, 63;
mul.lo.s32 %r719, %r144, %r718;
cvt.rn.f32.s32 %f174, %r719;
fma.rn.ftz.f32 %f175, %f101, %f174, 0f00000000;
shr.u32 %r720, %r698, 4;
and.b32 %r653, %r720, 252645135;
shr.u32 %r721, %r699, 4;
and.b32 %r657, %r721, 252645135;

dp4a.s32.s32 %r148, %r653, %r114, %r663;


dp4a.s32.s32 %r152, %r657, %r118, %r148;


dp4a.s32.s32 %r156, %r665, %r114, %r663;


dp4a.s32.s32 %r160, %r665, %r118, %r156;

shr.u16 %rs13, %rs11, 8;
cvt.u32.u16 %r722, %rs13;
mul.lo.s32 %r723, %r152, %r722;
cvt.rn.f32.s32 %f176, %r723;
fma.rn.ftz.f32 %f177, %f102, %f176, %f173;
shr.u16 %rs14, %rs12, 8;
cvt.u32.u16 %r724, %rs14;
mul.lo.s32 %r725, %r160, %r724;
cvt.rn.f32.s32 %f178, %r725;
fma.rn.ftz.f32 %f179, %f102, %f178, %f175;
ld.global.nc.u32 %r164, [%rd43];

{.reg .f16 low,high;
mov.b32 {low,high},%r164;
cvt.f32.f16 %f103, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r164;
cvt.f32.f16 %f104, high;}


mul.ftz.f32 %f180, %f177, %f103;
mul.ftz.f32 %f181, %f179, %f104;
sub.ftz.f32 %f182, %f180, %f181;
add.ftz.f32 %f543, %f182, %f543;
cvt.u64.u32 %rd44, %r1077;
add.s64 %rd45, %rd2, %rd44;
mul.lo.s64 %rd46, %rd45, 36;
add.s64 %rd47, %rd10, %rd46;
ld.global.nc.u32 %r166, [%rd47];

{.reg .f16 low,high;
mov.b32 {low,high},%r166;
cvt.f32.f16 %f105, low;}


add.s64 %rd48, %rd47, %rd38;
ld.global.nc.u32 %r170, [%rd48+4];
ld.global.nc.u32 %r174, [%rd48+20];
ld.global.nc.u32 %r167, [%rd47+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r167;
cvt.f32.f16 %f106, low;}


ld.global.nc.u32 %r186, [%rd48+40];
ld.global.nc.u32 %r190, [%rd48+56];

dp4a.s32.s32 %r168, %r601, %r170, %r663;


dp4a.s32.s32 %r172, %r605, %r174, %r168;


dp4a.s32.s32 %r176, %r665, %r170, %r663;


dp4a.s32.s32 %r180, %r665, %r174, %r176;

mul.lo.s32 %r726, %r172, %r687;
cvt.rn.f32.s32 %f183, %r726;
fma.rn.ftz.f32 %f184, %f105, %f183, 0f00000000;
mul.lo.s32 %r727, %r180, %r690;
cvt.rn.f32.s32 %f185, %r727;
fma.rn.ftz.f32 %f186, %f105, %f185, 0f00000000;

dp4a.s32.s32 %r184, %r617, %r186, %r663;


dp4a.s32.s32 %r188, %r621, %r190, %r184;


dp4a.s32.s32 %r192, %r665, %r186, %r663;


dp4a.s32.s32 %r196, %r665, %r190, %r192;

mul.lo.s32 %r728, %r188, %r694;
cvt.rn.f32.s32 %f187, %r728;
fma.rn.ftz.f32 %f188, %f106, %f187, %f184;
mul.lo.s32 %r729, %r196, %r696;
cvt.rn.f32.s32 %f189, %r729;
fma.rn.ftz.f32 %f190, %f106, %f189, %f186;

{.reg .f16 low,high;
mov.b32 {low,high},%r128;
cvt.f32.f16 %f107, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r128;
cvt.f32.f16 %f108, high;}


mul.ftz.f32 %f191, %f188, %f107;
mul.ftz.f32 %f192, %f190, %f108;
sub.ftz.f32 %f193, %f191, %f192;
add.ftz.f32 %f542, %f193, %f542;

{.reg .f16 low,high;
mov.b32 {low,high},%r166;
cvt.f32.f16 %f109, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r167;
cvt.f32.f16 %f110, low;}



dp4a.s32.s32 %r204, %r637, %r170, %r663;


dp4a.s32.s32 %r208, %r641, %r174, %r204;


dp4a.s32.s32 %r212, %r665, %r170, %r663;


dp4a.s32.s32 %r216, %r665, %r174, %r212;

mul.lo.s32 %r730, %r208, %r715;
cvt.rn.f32.s32 %f194, %r730;
fma.rn.ftz.f32 %f195, %f109, %f194, 0f00000000;
mul.lo.s32 %r731, %r216, %r718;
cvt.rn.f32.s32 %f196, %r731;
fma.rn.ftz.f32 %f197, %f109, %f196, 0f00000000;

dp4a.s32.s32 %r220, %r653, %r186, %r663;


dp4a.s32.s32 %r224, %r657, %r190, %r220;


dp4a.s32.s32 %r228, %r665, %r186, %r663;


dp4a.s32.s32 %r232, %r665, %r190, %r228;

mul.lo.s32 %r732, %r224, %r722;
cvt.rn.f32.s32 %f198, %r732;
fma.rn.ftz.f32 %f199, %f110, %f198, %f195;
mul.lo.s32 %r733, %r232, %r724;
cvt.rn.f32.s32 %f200, %r733;
fma.rn.ftz.f32 %f201, %f110, %f200, %f197;

{.reg .f16 low,high;
mov.b32 {low,high},%r164;
cvt.f32.f16 %f111, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r164;
cvt.f32.f16 %f112, high;}


mul.ftz.f32 %f202, %f199, %f111;
mul.ftz.f32 %f203, %f201, %f112;
sub.ftz.f32 %f204, %f202, %f203;
add.ftz.f32 %f541, %f204, %f541;
cvt.u64.u32 %rd49, %r1083;
add.s64 %rd50, %rd2, %rd49;
mul.lo.s64 %rd51, %rd50, 36;
add.s64 %rd52, %rd10, %rd51;
ld.global.nc.u32 %r238, [%rd52];

{.reg .f16 low,high;
mov.b32 {low,high},%r238;
cvt.f32.f16 %f113, low;}


add.s64 %rd53, %rd52, %rd38;
ld.global.nc.u32 %r242, [%rd53+4];
ld.global.nc.u32 %r246, [%rd53+20];
ld.global.nc.u32 %r239, [%rd52+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r239;
cvt.f32.f16 %f114, low;}


ld.global.nc.u32 %r258, [%rd53+40];
ld.global.nc.u32 %r262, [%rd53+56];

dp4a.s32.s32 %r240, %r601, %r242, %r663;


dp4a.s32.s32 %r244, %r605, %r246, %r240;


dp4a.s32.s32 %r248, %r665, %r242, %r663;


dp4a.s32.s32 %r252, %r665, %r246, %r248;

mul.lo.s32 %r734, %r244, %r687;
cvt.rn.f32.s32 %f205, %r734;
fma.rn.ftz.f32 %f206, %f113, %f205, 0f00000000;
mul.lo.s32 %r735, %r252, %r690;
cvt.rn.f32.s32 %f207, %r735;
fma.rn.ftz.f32 %f208, %f113, %f207, 0f00000000;

dp4a.s32.s32 %r256, %r617, %r258, %r663;


dp4a.s32.s32 %r260, %r621, %r262, %r256;


dp4a.s32.s32 %r264, %r665, %r258, %r663;


dp4a.s32.s32 %r268, %r665, %r262, %r264;

mul.lo.s32 %r736, %r260, %r694;
cvt.rn.f32.s32 %f209, %r736;
fma.rn.ftz.f32 %f210, %f114, %f209, %f206;
mul.lo.s32 %r737, %r268, %r696;
cvt.rn.f32.s32 %f211, %r737;
fma.rn.ftz.f32 %f212, %f114, %f211, %f208;

{.reg .f16 low,high;
mov.b32 {low,high},%r128;
cvt.f32.f16 %f115, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r128;
cvt.f32.f16 %f116, high;}


mul.ftz.f32 %f213, %f210, %f115;
mul.ftz.f32 %f214, %f212, %f116;
sub.ftz.f32 %f215, %f213, %f214;
add.ftz.f32 %f540, %f215, %f540;

{.reg .f16 low,high;
mov.b32 {low,high},%r238;
cvt.f32.f16 %f117, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r239;
cvt.f32.f16 %f118, low;}



dp4a.s32.s32 %r276, %r637, %r242, %r663;


dp4a.s32.s32 %r280, %r641, %r246, %r276;


dp4a.s32.s32 %r284, %r665, %r242, %r663;


dp4a.s32.s32 %r288, %r665, %r246, %r284;

mul.lo.s32 %r738, %r280, %r715;
cvt.rn.f32.s32 %f216, %r738;
fma.rn.ftz.f32 %f217, %f117, %f216, 0f00000000;
mul.lo.s32 %r739, %r288, %r718;
cvt.rn.f32.s32 %f218, %r739;
fma.rn.ftz.f32 %f219, %f117, %f218, 0f00000000;

dp4a.s32.s32 %r292, %r653, %r258, %r663;


dp4a.s32.s32 %r296, %r657, %r262, %r292;


dp4a.s32.s32 %r300, %r665, %r258, %r663;


dp4a.s32.s32 %r304, %r665, %r262, %r300;

mul.lo.s32 %r740, %r296, %r722;
cvt.rn.f32.s32 %f220, %r740;
fma.rn.ftz.f32 %f221, %f118, %f220, %f217;
mul.lo.s32 %r741, %r304, %r724;
cvt.rn.f32.s32 %f222, %r741;
fma.rn.ftz.f32 %f223, %f118, %f222, %f219;

{.reg .f16 low,high;
mov.b32 {low,high},%r164;
cvt.f32.f16 %f119, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r164;
cvt.f32.f16 %f120, high;}


mul.ftz.f32 %f224, %f221, %f119;
mul.ftz.f32 %f225, %f223, %f120;
sub.ftz.f32 %f226, %f224, %f225;
add.ftz.f32 %f539, %f226, %f539;
cvt.u64.u32 %rd54, %r1082;
add.s64 %rd55, %rd2, %rd54;
mul.lo.s64 %rd56, %rd55, 36;
add.s64 %rd57, %rd10, %rd56;
ld.global.nc.u32 %r310, [%rd57];

{.reg .f16 low,high;
mov.b32 {low,high},%r310;
cvt.f32.f16 %f121, low;}


add.s64 %rd58, %rd57, %rd38;
ld.global.nc.u32 %r314, [%rd58+4];
ld.global.nc.u32 %r318, [%rd58+20];
ld.global.nc.u32 %r311, [%rd57+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r311;
cvt.f32.f16 %f122, low;}


ld.global.nc.u32 %r330, [%rd58+40];
ld.global.nc.u32 %r334, [%rd58+56];

dp4a.s32.s32 %r312, %r601, %r314, %r663;


dp4a.s32.s32 %r316, %r605, %r318, %r312;


dp4a.s32.s32 %r320, %r665, %r314, %r663;


dp4a.s32.s32 %r324, %r665, %r318, %r320;

mul.lo.s32 %r742, %r316, %r687;
cvt.rn.f32.s32 %f227, %r742;
fma.rn.ftz.f32 %f228, %f121, %f227, 0f00000000;
mul.lo.s32 %r743, %r324, %r690;
cvt.rn.f32.s32 %f229, %r743;
fma.rn.ftz.f32 %f230, %f121, %f229, 0f00000000;

dp4a.s32.s32 %r328, %r617, %r330, %r663;


dp4a.s32.s32 %r332, %r621, %r334, %r328;


dp4a.s32.s32 %r336, %r665, %r330, %r663;


dp4a.s32.s32 %r340, %r665, %r334, %r336;

mul.lo.s32 %r744, %r332, %r694;
cvt.rn.f32.s32 %f231, %r744;
fma.rn.ftz.f32 %f232, %f122, %f231, %f228;
mul.lo.s32 %r745, %r340, %r696;
cvt.rn.f32.s32 %f233, %r745;
fma.rn.ftz.f32 %f234, %f122, %f233, %f230;

{.reg .f16 low,high;
mov.b32 {low,high},%r128;
cvt.f32.f16 %f123, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r128;
cvt.f32.f16 %f124, high;}


mul.ftz.f32 %f235, %f232, %f123;
mul.ftz.f32 %f236, %f234, %f124;
sub.ftz.f32 %f237, %f235, %f236;
add.ftz.f32 %f538, %f237, %f538;

{.reg .f16 low,high;
mov.b32 {low,high},%r310;
cvt.f32.f16 %f125, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r311;
cvt.f32.f16 %f126, low;}



dp4a.s32.s32 %r348, %r637, %r314, %r663;


dp4a.s32.s32 %r352, %r641, %r318, %r348;


dp4a.s32.s32 %r356, %r665, %r314, %r663;


dp4a.s32.s32 %r360, %r665, %r318, %r356;

mul.lo.s32 %r746, %r352, %r715;
cvt.rn.f32.s32 %f238, %r746;
fma.rn.ftz.f32 %f239, %f125, %f238, 0f00000000;
mul.lo.s32 %r747, %r360, %r718;
cvt.rn.f32.s32 %f240, %r747;
fma.rn.ftz.f32 %f241, %f125, %f240, 0f00000000;

dp4a.s32.s32 %r364, %r653, %r330, %r663;


dp4a.s32.s32 %r368, %r657, %r334, %r364;


dp4a.s32.s32 %r372, %r665, %r330, %r663;


dp4a.s32.s32 %r376, %r665, %r334, %r372;

mul.lo.s32 %r748, %r368, %r722;
cvt.rn.f32.s32 %f242, %r748;
fma.rn.ftz.f32 %f243, %f126, %f242, %f239;
mul.lo.s32 %r749, %r376, %r724;
cvt.rn.f32.s32 %f244, %r749;
fma.rn.ftz.f32 %f245, %f126, %f244, %f241;

{.reg .f16 low,high;
mov.b32 {low,high},%r164;
cvt.f32.f16 %f127, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r164;
cvt.f32.f16 %f128, high;}


mul.ftz.f32 %f246, %f243, %f127;
mul.ftz.f32 %f247, %f245, %f128;
sub.ftz.f32 %f248, %f246, %f247;
add.ftz.f32 %f537, %f248, %f537;
cvt.u64.u32 %rd59, %r1081;
add.s64 %rd60, %rd2, %rd59;
mul.lo.s64 %rd61, %rd60, 36;
add.s64 %rd62, %rd10, %rd61;
ld.global.nc.u32 %r382, [%rd62];

{.reg .f16 low,high;
mov.b32 {low,high},%r382;
cvt.f32.f16 %f129, low;}


add.s64 %rd63, %rd62, %rd38;
ld.global.nc.u32 %r386, [%rd63+4];
ld.global.nc.u32 %r390, [%rd63+20];
ld.global.nc.u32 %r383, [%rd62+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r383;
cvt.f32.f16 %f130, low;}


ld.global.nc.u32 %r402, [%rd63+40];
ld.global.nc.u32 %r406, [%rd63+56];

dp4a.s32.s32 %r384, %r601, %r386, %r663;


dp4a.s32.s32 %r388, %r605, %r390, %r384;


dp4a.s32.s32 %r392, %r665, %r386, %r663;


dp4a.s32.s32 %r396, %r665, %r390, %r392;

mul.lo.s32 %r750, %r388, %r687;
cvt.rn.f32.s32 %f249, %r750;
fma.rn.ftz.f32 %f250, %f129, %f249, 0f00000000;
mul.lo.s32 %r751, %r396, %r690;
cvt.rn.f32.s32 %f251, %r751;
fma.rn.ftz.f32 %f252, %f129, %f251, 0f00000000;

dp4a.s32.s32 %r400, %r617, %r402, %r663;


dp4a.s32.s32 %r404, %r621, %r406, %r400;


dp4a.s32.s32 %r408, %r665, %r402, %r663;


dp4a.s32.s32 %r412, %r665, %r406, %r408;

mul.lo.s32 %r752, %r404, %r694;
cvt.rn.f32.s32 %f253, %r752;
fma.rn.ftz.f32 %f254, %f130, %f253, %f250;
mul.lo.s32 %r753, %r412, %r696;
cvt.rn.f32.s32 %f255, %r753;
fma.rn.ftz.f32 %f256, %f130, %f255, %f252;

{.reg .f16 low,high;
mov.b32 {low,high},%r128;
cvt.f32.f16 %f131, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r128;
cvt.f32.f16 %f132, high;}


mul.ftz.f32 %f257, %f254, %f131;
mul.ftz.f32 %f258, %f256, %f132;
sub.ftz.f32 %f259, %f257, %f258;
add.ftz.f32 %f536, %f259, %f536;

{.reg .f16 low,high;
mov.b32 {low,high},%r382;
cvt.f32.f16 %f133, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r383;
cvt.f32.f16 %f134, low;}



dp4a.s32.s32 %r420, %r637, %r386, %r663;


dp4a.s32.s32 %r424, %r641, %r390, %r420;


dp4a.s32.s32 %r428, %r665, %r386, %r663;


dp4a.s32.s32 %r432, %r665, %r390, %r428;

mul.lo.s32 %r754, %r424, %r715;
cvt.rn.f32.s32 %f260, %r754;
fma.rn.ftz.f32 %f261, %f133, %f260, 0f00000000;
mul.lo.s32 %r755, %r432, %r718;
cvt.rn.f32.s32 %f262, %r755;
fma.rn.ftz.f32 %f263, %f133, %f262, 0f00000000;

dp4a.s32.s32 %r436, %r653, %r402, %r663;


dp4a.s32.s32 %r440, %r657, %r406, %r436;


dp4a.s32.s32 %r444, %r665, %r402, %r663;


dp4a.s32.s32 %r448, %r665, %r406, %r444;

mul.lo.s32 %r756, %r440, %r722;
cvt.rn.f32.s32 %f264, %r756;
fma.rn.ftz.f32 %f265, %f134, %f264, %f261;
mul.lo.s32 %r757, %r448, %r724;
cvt.rn.f32.s32 %f266, %r757;
fma.rn.ftz.f32 %f267, %f134, %f266, %f263;

{.reg .f16 low,high;
mov.b32 {low,high},%r164;
cvt.f32.f16 %f135, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r164;
cvt.f32.f16 %f136, high;}


mul.ftz.f32 %f268, %f265, %f135;
mul.ftz.f32 %f269, %f267, %f136;
sub.ftz.f32 %f270, %f268, %f269;
add.ftz.f32 %f535, %f270, %f535;
cvt.u64.u32 %rd64, %r1080;
add.s64 %rd65, %rd2, %rd64;
mul.lo.s64 %rd66, %rd65, 36;
add.s64 %rd67, %rd10, %rd66;
ld.global.nc.u32 %r454, [%rd67];

{.reg .f16 low,high;
mov.b32 {low,high},%r454;
cvt.f32.f16 %f137, low;}


add.s64 %rd68, %rd67, %rd38;
ld.global.nc.u32 %r458, [%rd68+4];
ld.global.nc.u32 %r462, [%rd68+20];
ld.global.nc.u32 %r455, [%rd67+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r455;
cvt.f32.f16 %f138, low;}


ld.global.nc.u32 %r474, [%rd68+40];
ld.global.nc.u32 %r478, [%rd68+56];

dp4a.s32.s32 %r456, %r601, %r458, %r663;


dp4a.s32.s32 %r460, %r605, %r462, %r456;


dp4a.s32.s32 %r464, %r665, %r458, %r663;


dp4a.s32.s32 %r468, %r665, %r462, %r464;

mul.lo.s32 %r758, %r460, %r687;
cvt.rn.f32.s32 %f271, %r758;
fma.rn.ftz.f32 %f272, %f137, %f271, 0f00000000;
mul.lo.s32 %r759, %r468, %r690;
cvt.rn.f32.s32 %f273, %r759;
fma.rn.ftz.f32 %f274, %f137, %f273, 0f00000000;

dp4a.s32.s32 %r472, %r617, %r474, %r663;


dp4a.s32.s32 %r476, %r621, %r478, %r472;


dp4a.s32.s32 %r480, %r665, %r474, %r663;


dp4a.s32.s32 %r484, %r665, %r478, %r480;

mul.lo.s32 %r760, %r476, %r694;
cvt.rn.f32.s32 %f275, %r760;
fma.rn.ftz.f32 %f276, %f138, %f275, %f272;
mul.lo.s32 %r761, %r484, %r696;
cvt.rn.f32.s32 %f277, %r761;
fma.rn.ftz.f32 %f278, %f138, %f277, %f274;

{.reg .f16 low,high;
mov.b32 {low,high},%r128;
cvt.f32.f16 %f139, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r128;
cvt.f32.f16 %f140, high;}


mul.ftz.f32 %f279, %f276, %f139;
mul.ftz.f32 %f280, %f278, %f140;
sub.ftz.f32 %f281, %f279, %f280;
add.ftz.f32 %f534, %f281, %f534;

{.reg .f16 low,high;
mov.b32 {low,high},%r454;
cvt.f32.f16 %f141, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r455;
cvt.f32.f16 %f142, low;}



dp4a.s32.s32 %r492, %r637, %r458, %r663;


dp4a.s32.s32 %r496, %r641, %r462, %r492;


dp4a.s32.s32 %r500, %r665, %r458, %r663;


dp4a.s32.s32 %r504, %r665, %r462, %r500;

mul.lo.s32 %r762, %r496, %r715;
cvt.rn.f32.s32 %f282, %r762;
fma.rn.ftz.f32 %f283, %f141, %f282, 0f00000000;
mul.lo.s32 %r763, %r504, %r718;
cvt.rn.f32.s32 %f284, %r763;
fma.rn.ftz.f32 %f285, %f141, %f284, 0f00000000;

dp4a.s32.s32 %r508, %r653, %r474, %r663;


dp4a.s32.s32 %r512, %r657, %r478, %r508;


dp4a.s32.s32 %r516, %r665, %r474, %r663;


dp4a.s32.s32 %r520, %r665, %r478, %r516;

mul.lo.s32 %r764, %r512, %r722;
cvt.rn.f32.s32 %f286, %r764;
fma.rn.ftz.f32 %f287, %f142, %f286, %f283;
mul.lo.s32 %r765, %r520, %r724;
cvt.rn.f32.s32 %f288, %r765;
fma.rn.ftz.f32 %f289, %f142, %f288, %f285;

{.reg .f16 low,high;
mov.b32 {low,high},%r164;
cvt.f32.f16 %f143, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r164;
cvt.f32.f16 %f144, high;}


mul.ftz.f32 %f290, %f287, %f143;
mul.ftz.f32 %f291, %f289, %f144;
sub.ftz.f32 %f292, %f290, %f291;
add.ftz.f32 %f533, %f292, %f533;
cvt.u64.u32 %rd69, %r1079;
add.s64 %rd70, %rd2, %rd69;
mul.lo.s64 %rd71, %rd70, 36;
add.s64 %rd72, %rd10, %rd71;
ld.global.nc.u32 %r526, [%rd72];

{.reg .f16 low,high;
mov.b32 {low,high},%r526;
cvt.f32.f16 %f145, low;}


add.s64 %rd73, %rd72, %rd38;
ld.global.nc.u32 %r530, [%rd73+4];
ld.global.nc.u32 %r534, [%rd73+20];
ld.global.nc.u32 %r527, [%rd72+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r527;
cvt.f32.f16 %f146, low;}


ld.global.nc.u32 %r546, [%rd73+40];
ld.global.nc.u32 %r550, [%rd73+56];

dp4a.s32.s32 %r528, %r601, %r530, %r663;


dp4a.s32.s32 %r532, %r605, %r534, %r528;


dp4a.s32.s32 %r536, %r665, %r530, %r663;


dp4a.s32.s32 %r540, %r665, %r534, %r536;

mul.lo.s32 %r766, %r532, %r687;
cvt.rn.f32.s32 %f293, %r766;
fma.rn.ftz.f32 %f294, %f145, %f293, 0f00000000;
mul.lo.s32 %r767, %r540, %r690;
cvt.rn.f32.s32 %f295, %r767;
fma.rn.ftz.f32 %f296, %f145, %f295, 0f00000000;

dp4a.s32.s32 %r544, %r617, %r546, %r663;


dp4a.s32.s32 %r548, %r621, %r550, %r544;


dp4a.s32.s32 %r552, %r665, %r546, %r663;


dp4a.s32.s32 %r556, %r665, %r550, %r552;

mul.lo.s32 %r768, %r548, %r694;
cvt.rn.f32.s32 %f297, %r768;
fma.rn.ftz.f32 %f298, %f146, %f297, %f294;
mul.lo.s32 %r769, %r556, %r696;
cvt.rn.f32.s32 %f299, %r769;
fma.rn.ftz.f32 %f300, %f146, %f299, %f296;

{.reg .f16 low,high;
mov.b32 {low,high},%r128;
cvt.f32.f16 %f147, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r128;
cvt.f32.f16 %f148, high;}


mul.ftz.f32 %f301, %f298, %f147;
mul.ftz.f32 %f302, %f300, %f148;
sub.ftz.f32 %f303, %f301, %f302;
add.ftz.f32 %f532, %f303, %f532;

{.reg .f16 low,high;
mov.b32 {low,high},%r526;
cvt.f32.f16 %f149, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r527;
cvt.f32.f16 %f150, low;}



dp4a.s32.s32 %r564, %r637, %r530, %r663;


dp4a.s32.s32 %r568, %r641, %r534, %r564;


dp4a.s32.s32 %r572, %r665, %r530, %r663;


dp4a.s32.s32 %r576, %r665, %r534, %r572;

mul.lo.s32 %r770, %r568, %r715;
cvt.rn.f32.s32 %f304, %r770;
fma.rn.ftz.f32 %f305, %f149, %f304, 0f00000000;
mul.lo.s32 %r771, %r576, %r718;
cvt.rn.f32.s32 %f306, %r771;
fma.rn.ftz.f32 %f307, %f149, %f306, 0f00000000;

dp4a.s32.s32 %r580, %r653, %r546, %r663;


dp4a.s32.s32 %r584, %r657, %r550, %r580;


dp4a.s32.s32 %r588, %r665, %r546, %r663;


dp4a.s32.s32 %r592, %r665, %r550, %r588;

mul.lo.s32 %r772, %r584, %r722;
cvt.rn.f32.s32 %f308, %r772;
fma.rn.ftz.f32 %f309, %f150, %f308, %f305;
mul.lo.s32 %r773, %r592, %r724;
cvt.rn.f32.s32 %f310, %r773;
fma.rn.ftz.f32 %f311, %f150, %f310, %f307;

{.reg .f16 low,high;
mov.b32 {low,high},%r164;
cvt.f32.f16 %f151, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r164;
cvt.f32.f16 %f152, high;}


mul.ftz.f32 %f312, %f309, %f151;
mul.ftz.f32 %f313, %f311, %f152;
sub.ftz.f32 %f314, %f312, %f313;
add.ftz.f32 %f531, %f314, %f531;
cvt.u64.u32 %rd74, %r1078;
add.s64 %rd75, %rd2, %rd74;
mul.lo.s64 %rd76, %rd75, 36;
add.s64 %rd77, %rd10, %rd76;
ld.global.nc.u32 %r598, [%rd77];

{.reg .f16 low,high;
mov.b32 {low,high},%r598;
cvt.f32.f16 %f153, low;}


add.s64 %rd78, %rd77, %rd38;
ld.global.nc.u32 %r602, [%rd78+4];
ld.global.nc.u32 %r606, [%rd78+20];
ld.global.nc.u32 %r599, [%rd77+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r599;
cvt.f32.f16 %f154, low;}


ld.global.nc.u32 %r618, [%rd78+40];
ld.global.nc.u32 %r622, [%rd78+56];

dp4a.s32.s32 %r600, %r601, %r602, %r663;


dp4a.s32.s32 %r604, %r605, %r606, %r600;


dp4a.s32.s32 %r608, %r665, %r602, %r663;


dp4a.s32.s32 %r612, %r665, %r606, %r608;

mul.lo.s32 %r774, %r604, %r687;
cvt.rn.f32.s32 %f315, %r774;
fma.rn.ftz.f32 %f316, %f153, %f315, 0f00000000;
mul.lo.s32 %r775, %r612, %r690;
cvt.rn.f32.s32 %f317, %r775;
fma.rn.ftz.f32 %f318, %f153, %f317, 0f00000000;

dp4a.s32.s32 %r616, %r617, %r618, %r663;


dp4a.s32.s32 %r620, %r621, %r622, %r616;


dp4a.s32.s32 %r624, %r665, %r618, %r663;


dp4a.s32.s32 %r628, %r665, %r622, %r624;

mul.lo.s32 %r776, %r620, %r694;
cvt.rn.f32.s32 %f319, %r776;
fma.rn.ftz.f32 %f320, %f154, %f319, %f316;
mul.lo.s32 %r777, %r628, %r696;
cvt.rn.f32.s32 %f321, %r777;
fma.rn.ftz.f32 %f322, %f154, %f321, %f318;

{.reg .f16 low,high;
mov.b32 {low,high},%r128;
cvt.f32.f16 %f155, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r128;
cvt.f32.f16 %f156, high;}


mul.ftz.f32 %f323, %f320, %f155;
mul.ftz.f32 %f324, %f322, %f156;
sub.ftz.f32 %f325, %f323, %f324;
add.ftz.f32 %f530, %f325, %f530;

{.reg .f16 low,high;
mov.b32 {low,high},%r598;
cvt.f32.f16 %f157, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r599;
cvt.f32.f16 %f158, low;}



dp4a.s32.s32 %r636, %r637, %r602, %r663;


dp4a.s32.s32 %r640, %r641, %r606, %r636;


dp4a.s32.s32 %r644, %r665, %r602, %r663;


dp4a.s32.s32 %r648, %r665, %r606, %r644;

mul.lo.s32 %r778, %r640, %r715;
cvt.rn.f32.s32 %f326, %r778;
fma.rn.ftz.f32 %f327, %f157, %f326, 0f00000000;
mul.lo.s32 %r779, %r648, %r718;
cvt.rn.f32.s32 %f328, %r779;
fma.rn.ftz.f32 %f329, %f157, %f328, 0f00000000;

dp4a.s32.s32 %r652, %r653, %r618, %r663;


dp4a.s32.s32 %r656, %r657, %r622, %r652;


dp4a.s32.s32 %r660, %r665, %r618, %r663;


dp4a.s32.s32 %r664, %r665, %r622, %r660;

mul.lo.s32 %r780, %r656, %r722;
cvt.rn.f32.s32 %f330, %r780;
fma.rn.ftz.f32 %f331, %f158, %f330, %f327;
mul.lo.s32 %r781, %r664, %r724;
cvt.rn.f32.s32 %f332, %r781;
fma.rn.ftz.f32 %f333, %f158, %f332, %f329;

{.reg .f16 low,high;
mov.b32 {low,high},%r164;
cvt.f32.f16 %f159, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r164;
cvt.f32.f16 %f160, high;}


mul.ftz.f32 %f334, %f331, %f159;
mul.ftz.f32 %f335, %f333, %f160;
sub.ftz.f32 %f336, %f334, %f335;
add.ftz.f32 %f529, %f336, %f529;
add.s32 %r1083, %r1083, 32;
add.s32 %r1082, %r1082, 32;
add.s32 %r1081, %r1081, 32;
add.s32 %r1080, %r1080, 32;
add.s32 %r1079, %r1079, 32;
add.s32 %r1078, %r1078, 32;
add.s32 %r1077, %r1077, 32;
add.s32 %r1076, %r1076, 32;
add.s64 %rd133, %rd133, 576;
add.s32 %r1084, %r1084, 4;
setp.lt.s32 %p5, %r1084, %r6;
@%p5 bra $L__BB167_2;

$L__BB167_3:
setp.eq.s32 %p6, %r1, 0;
@%p6 bra $L__BB167_5;

shl.b32 %r782, %r1, 11;
mov.u32 %r783, _ZZ13mul_mat_vec_qIL9ggml_type12ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r784, %r783, %r782;
shl.b32 %r785, %r2, 2;
add.s32 %r786, %r784, %r785;
st.shared.f32 [%r786+-2048], %f544;
st.shared.f32 [%r786+-1920], %f543;
st.shared.f32 [%r786+-1792], %f542;
st.shared.f32 [%r786+-1664], %f541;
st.shared.f32 [%r786+-1536], %f540;
st.shared.f32 [%r786+-1408], %f539;
st.shared.f32 [%r786+-1280], %f538;
st.shared.f32 [%r786+-1152], %f537;
st.shared.f32 [%r786+-1024], %f536;
st.shared.f32 [%r786+-896], %f535;
st.shared.f32 [%r786+-768], %f534;
st.shared.f32 [%r786+-640], %f533;
st.shared.f32 [%r786+-512], %f532;
st.shared.f32 [%r786+-384], %f531;
st.shared.f32 [%r786+-256], %f530;
st.shared.f32 [%r786+-128], %f529;

$L__BB167_5:
setp.ne.s32 %p7, %r1, 0;
bar.sync 0;
@%p7 bra $L__BB167_38;

mov.u32 %r787, %ctaid.z;
mov.u32 %r788, %ctaid.y;
mad.lo.s32 %r789, %r788, %r46, %r4;
mad.lo.s32 %r790, %r787, %r52, %r789;
cvt.u64.u32 %rd13, %r790;
shl.b32 %r791, %r2, 2;
mov.u32 %r792, _ZZ13mul_mat_vec_qIL9ggml_type12ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r34, %r792, %r791;
mov.u32 %r793, 2;
ld.shared.f32 %f337, [%r34];
add.ftz.f32 %f338, %f337, %f544;
mov.b32 %r794, %f338;
mov.u32 %r795, 31;
mov.u32 %r796, 16;
mov.u32 %r797, -1;
shfl.sync.bfly.b32 %r798|%p8, %r794, %r796, %r795, %r797;
mov.b32 %f339, %r798;
add.ftz.f32 %f340, %f338, %f339;
mov.b32 %r799, %f340;
mov.u32 %r800, 8;
shfl.sync.bfly.b32 %r801|%p9, %r799, %r800, %r795, %r797;
mov.b32 %f341, %r801;
add.ftz.f32 %f342, %f340, %f341;
mov.b32 %r802, %f342;
mov.u32 %r803, 4;
shfl.sync.bfly.b32 %r804|%p10, %r802, %r803, %r795, %r797;
mov.b32 %f343, %r804;
add.ftz.f32 %f344, %f342, %f343;
mov.b32 %r805, %f344;
shfl.sync.bfly.b32 %r806|%p11, %r805, %r793, %r795, %r797;
mov.b32 %f345, %r806;
add.ftz.f32 %f346, %f344, %f345;
mov.b32 %r807, %f346;
mov.u32 %r808, 1;
shfl.sync.bfly.b32 %r809|%p12, %r807, %r808, %r795, %r797;
mov.b32 %f347, %r809;
add.ftz.f32 %f49, %f346, %f347;
setp.lt.u32 %p13, %r4, %r40;
setp.eq.s32 %p14, %r2, 0;
and.pred %p1, %p14, %p13;
cvta.to.global.u64 %rd79, %rd20;
mul.wide.u32 %rd80, %r790, 4;
add.s64 %rd14, %rd79, %rd80;
not.pred %p15, %p1;
@%p15 bra $L__BB167_8;

st.global.f32 [%rd14], %f49;

$L__BB167_8:
ld.shared.f32 %f348, [%r34+128];
add.ftz.f32 %f349, %f348, %f543;
mov.b32 %r810, %f349;
shfl.sync.bfly.b32 %r814|%p16, %r810, %r796, %r795, %r797;
mov.b32 %f350, %r814;
add.ftz.f32 %f351, %f349, %f350;
mov.b32 %r815, %f351;
shfl.sync.bfly.b32 %r817|%p17, %r815, %r800, %r795, %r797;
mov.b32 %f352, %r817;
add.ftz.f32 %f353, %f351, %f352;
mov.b32 %r818, %f353;
shfl.sync.bfly.b32 %r820|%p18, %r818, %r803, %r795, %r797;
mov.b32 %f354, %r820;
add.ftz.f32 %f355, %f353, %f354;
mov.b32 %r821, %f355;
shfl.sync.bfly.b32 %r823|%p19, %r821, %r793, %r795, %r797;
mov.b32 %f356, %r823;
add.ftz.f32 %f357, %f355, %f356;
mov.b32 %r824, %f357;
shfl.sync.bfly.b32 %r826|%p20, %r824, %r808, %r795, %r797;
mov.b32 %f358, %r826;
add.ftz.f32 %f50, %f357, %f358;
setp.eq.s32 %p21, %r2, 1;
or.b32 %r830, %r4, 1;
setp.lt.u32 %p22, %r830, %r40;
and.pred %p2, %p21, %p22;
not.pred %p23, %p2;
@%p23 bra $L__BB167_10;

st.global.f32 [%rd14+4], %f50;

$L__BB167_10:
ld.shared.f32 %f359, [%r34+256];
add.ftz.f32 %f360, %f359, %f542;
mov.b32 %r831, %f360;
mov.u32 %r832, 31;
mov.u32 %r833, 16;
mov.u32 %r834, -1;
shfl.sync.bfly.b32 %r835|%p24, %r831, %r833, %r832, %r834;
mov.b32 %f361, %r835;
add.ftz.f32 %f362, %f360, %f361;
mov.b32 %r836, %f362;
mov.u32 %r837, 8;
shfl.sync.bfly.b32 %r838|%p25, %r836, %r837, %r832, %r834;
mov.b32 %f363, %r838;
add.ftz.f32 %f364, %f362, %f363;
mov.b32 %r839, %f364;
mov.u32 %r840, 4;
shfl.sync.bfly.b32 %r841|%p26, %r839, %r840, %r832, %r834;
mov.b32 %f365, %r841;
add.ftz.f32 %f366, %f364, %f365;
mov.b32 %r842, %f366;
mov.u32 %r843, 2;
shfl.sync.bfly.b32 %r844|%p27, %r842, %r843, %r832, %r834;
mov.b32 %f367, %r844;
add.ftz.f32 %f368, %f366, %f367;
mov.b32 %r845, %f368;
mov.u32 %r846, 1;
shfl.sync.bfly.b32 %r847|%p28, %r845, %r846, %r832, %r834;
mov.b32 %f369, %r847;
add.ftz.f32 %f51, %f368, %f369;
@%p15 bra $L__BB167_12;

cvt.u64.u32 %rd81, %r40;
add.s64 %rd82, %rd81, %rd13;
shl.b64 %rd84, %rd82, 2;
add.s64 %rd85, %rd79, %rd84;
st.global.f32 [%rd85], %f51;

$L__BB167_12:
ld.shared.f32 %f370, [%r34+384];
add.ftz.f32 %f371, %f370, %f541;
mov.b32 %r848, %f371;
shfl.sync.bfly.b32 %r852|%p30, %r848, %r833, %r832, %r834;
mov.b32 %f372, %r852;
add.ftz.f32 %f373, %f371, %f372;
mov.b32 %r853, %f373;
shfl.sync.bfly.b32 %r855|%p31, %r853, %r837, %r832, %r834;
mov.b32 %f374, %r855;
add.ftz.f32 %f375, %f373, %f374;
mov.b32 %r856, %f375;
shfl.sync.bfly.b32 %r858|%p32, %r856, %r840, %r832, %r834;
mov.b32 %f376, %r858;
add.ftz.f32 %f377, %f375, %f376;
mov.b32 %r859, %f377;
shfl.sync.bfly.b32 %r861|%p33, %r859, %r843, %r832, %r834;
mov.b32 %f378, %r861;
add.ftz.f32 %f379, %f377, %f378;
mov.b32 %r862, %f379;
shfl.sync.bfly.b32 %r864|%p34, %r862, %r846, %r832, %r834;
mov.b32 %f380, %r864;
add.ftz.f32 %f52, %f379, %f380;
@%p23 bra $L__BB167_14;

add.s32 %r865, %r40, 1;
cvt.u64.u32 %rd86, %r865;
add.s64 %rd87, %rd86, %rd13;
shl.b64 %rd89, %rd87, 2;
add.s64 %rd90, %rd79, %rd89;
st.global.f32 [%rd90], %f52;

$L__BB167_14:
shl.b32 %r866, %r40, 1;
mov.u32 %r867, 1;
ld.shared.f32 %f381, [%r34+512];
add.ftz.f32 %f382, %f381, %f540;
mov.b32 %r868, %f382;
mov.u32 %r869, 31;
mov.u32 %r870, 16;
mov.u32 %r871, -1;
shfl.sync.bfly.b32 %r872|%p36, %r868, %r870, %r869, %r871;
mov.b32 %f383, %r872;
add.ftz.f32 %f384, %f382, %f383;
mov.b32 %r873, %f384;
mov.u32 %r874, 8;
shfl.sync.bfly.b32 %r875|%p37, %r873, %r874, %r869, %r871;
mov.b32 %f385, %r875;
add.ftz.f32 %f386, %f384, %f385;
mov.b32 %r876, %f386;
mov.u32 %r877, 4;
shfl.sync.bfly.b32 %r878|%p38, %r876, %r877, %r869, %r871;
mov.b32 %f387, %r878;
add.ftz.f32 %f388, %f386, %f387;
mov.b32 %r879, %f388;
mov.u32 %r880, 2;
shfl.sync.bfly.b32 %r881|%p39, %r879, %r880, %r869, %r871;
mov.b32 %f389, %r881;
add.ftz.f32 %f390, %f388, %f389;
mov.b32 %r882, %f390;
shfl.sync.bfly.b32 %r883|%p40, %r882, %r867, %r869, %r871;
mov.b32 %f391, %r883;
add.ftz.f32 %f53, %f390, %f391;
cvt.u64.u32 %rd91, %r866;
add.s64 %rd92, %rd91, %rd13;
shl.b64 %rd94, %rd92, 2;
add.s64 %rd15, %rd79, %rd94;
@%p15 bra $L__BB167_16;

st.global.f32 [%rd15], %f53;

$L__BB167_16:
ld.shared.f32 %f392, [%r34+640];
add.ftz.f32 %f393, %f392, %f539;
mov.b32 %r884, %f393;
shfl.sync.bfly.b32 %r888|%p42, %r884, %r870, %r869, %r871;
mov.b32 %f394, %r888;
add.ftz.f32 %f395, %f393, %f394;
mov.b32 %r889, %f395;
shfl.sync.bfly.b32 %r891|%p43, %r889, %r874, %r869, %r871;
mov.b32 %f396, %r891;
add.ftz.f32 %f397, %f395, %f396;
mov.b32 %r892, %f397;
shfl.sync.bfly.b32 %r894|%p44, %r892, %r877, %r869, %r871;
mov.b32 %f398, %r894;
add.ftz.f32 %f399, %f397, %f398;
mov.b32 %r895, %f399;
shfl.sync.bfly.b32 %r897|%p45, %r895, %r880, %r869, %r871;
mov.b32 %f400, %r897;
add.ftz.f32 %f401, %f399, %f400;
mov.b32 %r898, %f401;
shfl.sync.bfly.b32 %r900|%p46, %r898, %r867, %r869, %r871;
mov.b32 %f402, %r900;
add.ftz.f32 %f54, %f401, %f402;
@%p23 bra $L__BB167_18;

st.global.f32 [%rd15+4], %f54;

$L__BB167_18:
mul.lo.s32 %r35, %r40, 3;
ld.shared.f32 %f403, [%r34+768];
add.ftz.f32 %f404, %f403, %f538;
mov.b32 %r901, %f404;
mov.u32 %r902, 31;
mov.u32 %r903, 16;
mov.u32 %r904, -1;
shfl.sync.bfly.b32 %r905|%p48, %r901, %r903, %r902, %r904;
mov.b32 %f405, %r905;
add.ftz.f32 %f406, %f404, %f405;
mov.b32 %r906, %f406;
mov.u32 %r907, 8;
shfl.sync.bfly.b32 %r908|%p49, %r906, %r907, %r902, %r904;
mov.b32 %f407, %r908;
add.ftz.f32 %f408, %f406, %f407;
mov.b32 %r909, %f408;
mov.u32 %r910, 4;
shfl.sync.bfly.b32 %r911|%p50, %r909, %r910, %r902, %r904;
mov.b32 %f409, %r911;
add.ftz.f32 %f410, %f408, %f409;
mov.b32 %r912, %f410;
mov.u32 %r913, 2;
shfl.sync.bfly.b32 %r914|%p51, %r912, %r913, %r902, %r904;
mov.b32 %f411, %r914;
add.ftz.f32 %f412, %f410, %f411;
mov.b32 %r915, %f412;
mov.u32 %r916, 1;
shfl.sync.bfly.b32 %r917|%p52, %r915, %r916, %r902, %r904;
mov.b32 %f413, %r917;
add.ftz.f32 %f55, %f412, %f413;
@%p15 bra $L__BB167_20;

cvt.u64.u32 %rd95, %r35;
add.s64 %rd96, %rd95, %rd13;
shl.b64 %rd98, %rd96, 2;
add.s64 %rd99, %rd79, %rd98;
st.global.f32 [%rd99], %f55;

$L__BB167_20:
ld.shared.f32 %f414, [%r34+896];
add.ftz.f32 %f415, %f414, %f537;
mov.b32 %r918, %f415;
shfl.sync.bfly.b32 %r922|%p54, %r918, %r903, %r902, %r904;
mov.b32 %f416, %r922;
add.ftz.f32 %f417, %f415, %f416;
mov.b32 %r923, %f417;
shfl.sync.bfly.b32 %r925|%p55, %r923, %r907, %r902, %r904;
mov.b32 %f418, %r925;
add.ftz.f32 %f419, %f417, %f418;
mov.b32 %r926, %f419;
shfl.sync.bfly.b32 %r928|%p56, %r926, %r910, %r902, %r904;
mov.b32 %f420, %r928;
add.ftz.f32 %f421, %f419, %f420;
mov.b32 %r929, %f421;
shfl.sync.bfly.b32 %r931|%p57, %r929, %r913, %r902, %r904;
mov.b32 %f422, %r931;
add.ftz.f32 %f423, %f421, %f422;
mov.b32 %r932, %f423;
shfl.sync.bfly.b32 %r934|%p58, %r932, %r916, %r902, %r904;
mov.b32 %f424, %r934;
add.ftz.f32 %f56, %f423, %f424;
@%p23 bra $L__BB167_22;

add.s32 %r935, %r35, 1;
cvt.u64.u32 %rd100, %r935;
add.s64 %rd101, %rd100, %rd13;
shl.b64 %rd103, %rd101, 2;
add.s64 %rd104, %rd79, %rd103;
st.global.f32 [%rd104], %f56;

$L__BB167_22:
shl.b32 %r936, %r40, 2;
mov.u32 %r937, 2;
ld.shared.f32 %f425, [%r34+1024];
add.ftz.f32 %f426, %f425, %f536;
mov.b32 %r938, %f426;
mov.u32 %r939, 31;
mov.u32 %r940, 16;
mov.u32 %r941, -1;
shfl.sync.bfly.b32 %r942|%p60, %r938, %r940, %r939, %r941;
mov.b32 %f427, %r942;
add.ftz.f32 %f428, %f426, %f427;
mov.b32 %r943, %f428;
mov.u32 %r944, 8;
shfl.sync.bfly.b32 %r945|%p61, %r943, %r944, %r939, %r941;
mov.b32 %f429, %r945;
add.ftz.f32 %f430, %f428, %f429;
mov.b32 %r946, %f430;
mov.u32 %r947, 4;
shfl.sync.bfly.b32 %r948|%p62, %r946, %r947, %r939, %r941;
mov.b32 %f431, %r948;
add.ftz.f32 %f432, %f430, %f431;
mov.b32 %r949, %f432;
shfl.sync.bfly.b32 %r950|%p63, %r949, %r937, %r939, %r941;
mov.b32 %f433, %r950;
add.ftz.f32 %f434, %f432, %f433;
mov.b32 %r951, %f434;
mov.u32 %r952, 1;
shfl.sync.bfly.b32 %r953|%p64, %r951, %r952, %r939, %r941;
mov.b32 %f435, %r953;
add.ftz.f32 %f57, %f434, %f435;
cvt.u64.u32 %rd105, %r936;
add.s64 %rd106, %rd105, %rd13;
shl.b64 %rd108, %rd106, 2;
add.s64 %rd16, %rd79, %rd108;
@%p15 bra $L__BB167_24;

st.global.f32 [%rd16], %f57;

$L__BB167_24:
ld.shared.f32 %f436, [%r34+1152];
add.ftz.f32 %f437, %f436, %f535;
mov.b32 %r954, %f437;
shfl.sync.bfly.b32 %r958|%p66, %r954, %r940, %r939, %r941;
mov.b32 %f438, %r958;
add.ftz.f32 %f439, %f437, %f438;
mov.b32 %r959, %f439;
shfl.sync.bfly.b32 %r961|%p67, %r959, %r944, %r939, %r941;
mov.b32 %f440, %r961;
add.ftz.f32 %f441, %f439, %f440;
mov.b32 %r962, %f441;
shfl.sync.bfly.b32 %r964|%p68, %r962, %r947, %r939, %r941;
mov.b32 %f442, %r964;
add.ftz.f32 %f443, %f441, %f442;
mov.b32 %r965, %f443;
shfl.sync.bfly.b32 %r967|%p69, %r965, %r937, %r939, %r941;
mov.b32 %f444, %r967;
add.ftz.f32 %f445, %f443, %f444;
mov.b32 %r968, %f445;
shfl.sync.bfly.b32 %r970|%p70, %r968, %r952, %r939, %r941;
mov.b32 %f446, %r970;
add.ftz.f32 %f58, %f445, %f446;
@%p23 bra $L__BB167_26;

st.global.f32 [%rd16+4], %f58;

$L__BB167_26:
mul.lo.s32 %r36, %r40, 5;
ld.shared.f32 %f447, [%r34+1280];
add.ftz.f32 %f448, %f447, %f534;
mov.b32 %r971, %f448;
mov.u32 %r972, 31;
mov.u32 %r973, 16;
mov.u32 %r974, -1;
shfl.sync.bfly.b32 %r975|%p72, %r971, %r973, %r972, %r974;
mov.b32 %f449, %r975;
add.ftz.f32 %f450, %f448, %f449;
mov.b32 %r976, %f450;
mov.u32 %r977, 8;
shfl.sync.bfly.b32 %r978|%p73, %r976, %r977, %r972, %r974;
mov.b32 %f451, %r978;
add.ftz.f32 %f452, %f450, %f451;
mov.b32 %r979, %f452;
mov.u32 %r980, 4;
shfl.sync.bfly.b32 %r981|%p74, %r979, %r980, %r972, %r974;
mov.b32 %f453, %r981;
add.ftz.f32 %f454, %f452, %f453;
mov.b32 %r982, %f454;
mov.u32 %r983, 2;
shfl.sync.bfly.b32 %r984|%p75, %r982, %r983, %r972, %r974;
mov.b32 %f455, %r984;
add.ftz.f32 %f456, %f454, %f455;
mov.b32 %r985, %f456;
mov.u32 %r986, 1;
shfl.sync.bfly.b32 %r987|%p76, %r985, %r986, %r972, %r974;
mov.b32 %f457, %r987;
add.ftz.f32 %f59, %f456, %f457;
@%p15 bra $L__BB167_28;

cvt.u64.u32 %rd109, %r36;
add.s64 %rd110, %rd109, %rd13;
shl.b64 %rd112, %rd110, 2;
add.s64 %rd113, %rd79, %rd112;
st.global.f32 [%rd113], %f59;

$L__BB167_28:
ld.shared.f32 %f458, [%r34+1408];
add.ftz.f32 %f459, %f458, %f533;
mov.b32 %r988, %f459;
shfl.sync.bfly.b32 %r992|%p78, %r988, %r973, %r972, %r974;
mov.b32 %f460, %r992;
add.ftz.f32 %f461, %f459, %f460;
mov.b32 %r993, %f461;
shfl.sync.bfly.b32 %r995|%p79, %r993, %r977, %r972, %r974;
mov.b32 %f462, %r995;
add.ftz.f32 %f463, %f461, %f462;
mov.b32 %r996, %f463;
shfl.sync.bfly.b32 %r998|%p80, %r996, %r980, %r972, %r974;
mov.b32 %f464, %r998;
add.ftz.f32 %f465, %f463, %f464;
mov.b32 %r999, %f465;
shfl.sync.bfly.b32 %r1001|%p81, %r999, %r983, %r972, %r974;
mov.b32 %f466, %r1001;
add.ftz.f32 %f467, %f465, %f466;
mov.b32 %r1002, %f467;
shfl.sync.bfly.b32 %r1004|%p82, %r1002, %r986, %r972, %r974;
mov.b32 %f468, %r1004;
add.ftz.f32 %f60, %f467, %f468;
@%p23 bra $L__BB167_30;

add.s32 %r1005, %r36, 1;
cvt.u64.u32 %rd114, %r1005;
add.s64 %rd115, %rd114, %rd13;
shl.b64 %rd117, %rd115, 2;
add.s64 %rd118, %rd79, %rd117;
st.global.f32 [%rd118], %f60;

$L__BB167_30:
mul.lo.s32 %r1006, %r40, 6;
ld.shared.f32 %f469, [%r34+1536];
add.ftz.f32 %f470, %f469, %f532;
mov.b32 %r1007, %f470;
mov.u32 %r1008, 31;
mov.u32 %r1009, 16;
mov.u32 %r1010, -1;
shfl.sync.bfly.b32 %r1011|%p84, %r1007, %r1009, %r1008, %r1010;
mov.b32 %f471, %r1011;
add.ftz.f32 %f472, %f470, %f471;
mov.b32 %r1012, %f472;
mov.u32 %r1013, 8;
shfl.sync.bfly.b32 %r1014|%p85, %r1012, %r1013, %r1008, %r1010;
mov.b32 %f473, %r1014;
add.ftz.f32 %f474, %f472, %f473;
mov.b32 %r1015, %f474;
mov.u32 %r1016, 4;
shfl.sync.bfly.b32 %r1017|%p86, %r1015, %r1016, %r1008, %r1010;
mov.b32 %f475, %r1017;
add.ftz.f32 %f476, %f474, %f475;
mov.b32 %r1018, %f476;
mov.u32 %r1019, 2;
shfl.sync.bfly.b32 %r1020|%p87, %r1018, %r1019, %r1008, %r1010;
mov.b32 %f477, %r1020;
add.ftz.f32 %f478, %f476, %f477;
mov.b32 %r1021, %f478;
mov.u32 %r1022, 1;
shfl.sync.bfly.b32 %r1023|%p88, %r1021, %r1022, %r1008, %r1010;
mov.b32 %f479, %r1023;
add.ftz.f32 %f61, %f478, %f479;
cvt.u64.u32 %rd119, %r1006;
add.s64 %rd120, %rd119, %rd13;
shl.b64 %rd122, %rd120, 2;
add.s64 %rd17, %rd79, %rd122;
@%p15 bra $L__BB167_32;

st.global.f32 [%rd17], %f61;

$L__BB167_32:
ld.shared.f32 %f480, [%r34+1664];
add.ftz.f32 %f481, %f480, %f531;
mov.b32 %r1024, %f481;
shfl.sync.bfly.b32 %r1028|%p90, %r1024, %r1009, %r1008, %r1010;
mov.b32 %f482, %r1028;
add.ftz.f32 %f483, %f481, %f482;
mov.b32 %r1029, %f483;
shfl.sync.bfly.b32 %r1031|%p91, %r1029, %r1013, %r1008, %r1010;
mov.b32 %f484, %r1031;
add.ftz.f32 %f485, %f483, %f484;
mov.b32 %r1032, %f485;
shfl.sync.bfly.b32 %r1034|%p92, %r1032, %r1016, %r1008, %r1010;
mov.b32 %f486, %r1034;
add.ftz.f32 %f487, %f485, %f486;
mov.b32 %r1035, %f487;
shfl.sync.bfly.b32 %r1037|%p93, %r1035, %r1019, %r1008, %r1010;
mov.b32 %f488, %r1037;
add.ftz.f32 %f489, %f487, %f488;
mov.b32 %r1038, %f489;
shfl.sync.bfly.b32 %r1040|%p94, %r1038, %r1022, %r1008, %r1010;
mov.b32 %f490, %r1040;
add.ftz.f32 %f62, %f489, %f490;
@%p23 bra $L__BB167_34;

st.global.f32 [%rd17+4], %f62;

$L__BB167_34:
mul.lo.s32 %r37, %r40, 7;
ld.shared.f32 %f491, [%r34+1792];
add.ftz.f32 %f492, %f491, %f530;
mov.b32 %r1041, %f492;
mov.u32 %r1042, 31;
mov.u32 %r1043, 16;
mov.u32 %r1044, -1;
shfl.sync.bfly.b32 %r1045|%p96, %r1041, %r1043, %r1042, %r1044;
mov.b32 %f493, %r1045;
add.ftz.f32 %f494, %f492, %f493;
mov.b32 %r1046, %f494;
mov.u32 %r1047, 8;
shfl.sync.bfly.b32 %r1048|%p97, %r1046, %r1047, %r1042, %r1044;
mov.b32 %f495, %r1048;
add.ftz.f32 %f496, %f494, %f495;
mov.b32 %r1049, %f496;
mov.u32 %r1050, 4;
shfl.sync.bfly.b32 %r1051|%p98, %r1049, %r1050, %r1042, %r1044;
mov.b32 %f497, %r1051;
add.ftz.f32 %f498, %f496, %f497;
mov.b32 %r1052, %f498;
mov.u32 %r1053, 2;
shfl.sync.bfly.b32 %r1054|%p99, %r1052, %r1053, %r1042, %r1044;
mov.b32 %f499, %r1054;
add.ftz.f32 %f500, %f498, %f499;
mov.b32 %r1055, %f500;
mov.u32 %r1056, 1;
shfl.sync.bfly.b32 %r1057|%p100, %r1055, %r1056, %r1042, %r1044;
mov.b32 %f501, %r1057;
add.ftz.f32 %f63, %f500, %f501;
@%p15 bra $L__BB167_36;

cvt.u64.u32 %rd123, %r37;
add.s64 %rd124, %rd123, %rd13;
shl.b64 %rd126, %rd124, 2;
add.s64 %rd127, %rd79, %rd126;
st.global.f32 [%rd127], %f63;

$L__BB167_36:
ld.shared.f32 %f502, [%r34+1920];
add.ftz.f32 %f503, %f502, %f529;
mov.b32 %r1058, %f503;
shfl.sync.bfly.b32 %r1062|%p102, %r1058, %r1043, %r1042, %r1044;
mov.b32 %f504, %r1062;
add.ftz.f32 %f505, %f503, %f504;
mov.b32 %r1063, %f505;
shfl.sync.bfly.b32 %r1065|%p103, %r1063, %r1047, %r1042, %r1044;
mov.b32 %f506, %r1065;
add.ftz.f32 %f507, %f505, %f506;
mov.b32 %r1066, %f507;
shfl.sync.bfly.b32 %r1068|%p104, %r1066, %r1050, %r1042, %r1044;
mov.b32 %f508, %r1068;
add.ftz.f32 %f509, %f507, %f508;
mov.b32 %r1069, %f509;
shfl.sync.bfly.b32 %r1071|%p105, %r1069, %r1053, %r1042, %r1044;
mov.b32 %f510, %r1071;
add.ftz.f32 %f511, %f509, %f510;
mov.b32 %r1072, %f511;
shfl.sync.bfly.b32 %r1074|%p106, %r1072, %r1056, %r1042, %r1044;
mov.b32 %f512, %r1074;
add.ftz.f32 %f64, %f511, %f512;
@%p23 bra $L__BB167_38;

add.s32 %r1075, %r37, 1;
cvt.u64.u32 %rd128, %r1075;
add.s64 %rd129, %rd128, %rd13;
shl.b64 %rd131, %rd129, 2;
add.s64 %rd132, %rd79, %rd131;
st.global.f32 [%rd132], %f64;

$L__BB167_38:
ret;

}
.entry _Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj(
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2,
.param .align 8 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3[48],
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5,
.param .align 4 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6[12],
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15
)
.maxntid 256, 1, 1
.minnctapersm 1
{
.local .align 8 .b8 __local_depot168[16];
.reg .b64 %SP;
.reg .b64 %SPL;
.reg .pred %p<52>;
.reg .b16 %rs<23>;
.reg .f32 %f<213>;
.reg .b32 %r<472>;
.reg .b64 %rd<111>;


mov.u64 %SPL, __local_depot168;
ld.param.u64 %rd26, [_Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0];
ld.param.u64 %rd27, [_Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1];
ld.param.u64 %rd28, [_Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2];
ld.param.u32 %r33, [_Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5];
ld.param.u32 %r26, [_Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8];
ld.param.u32 %r27, [_Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9];
ld.param.u32 %r29, [_Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11];
ld.param.u32 %r30, [_Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12];
ld.param.u32 %r34, [_Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14];
ld.param.u32 %r32, [_Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15];
ld.param.u32 %r24, [_Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+8];
ld.param.u32 %r23, [_Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+4];
ld.param.u32 %r22, [_Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6];
ld.param.u64 %rd30, [_Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+8];
add.u64 %rd1, %SPL, 0;
add.u64 %rd2, %SPL, 8;
mov.u32 %r35, %ctaid.x;
shl.b32 %r1, %r35, 1;
shr.u32 %r2, %r33, 8;
mov.u32 %r3, %tid.y;
setp.ge.u32 %p1, %r3, %r34;
@%p1 bra $L__BB168_29;

mov.u32 %r4, %ctaid.y;
cvta.to.global.u64 %rd37, %rd28;
mad.lo.s32 %r36, %r3, %r32, %r4;
mul.wide.u32 %rd38, %r36, 4;
add.s64 %rd39, %rd37, %rd38;
ld.global.nc.u32 %r5, [%rd39];
mov.f32 %f206, 0f00000000;
st.local.v2.f32 [%rd1], {%f206, %f206};
st.local.v2.f32 [%rd2], {%f206, %f206};
mov.u32 %r6, %tid.x;
shr.u32 %r471, %r6, 4;
setp.ge.u32 %p2, %r471, %r2;
mov.f32 %f207, %f206;
mov.f32 %f208, %f206;
mov.f32 %f209, %f206;
@%p2 bra $L__BB168_9;

mul.lo.s32 %r37, %r5, %r29;
and.b32 %r8, %r6, 15;
shr.u32 %r38, %r8, 2;
shl.b32 %r9, %r38, 1;
shl.b32 %r39, %r38, 5;
shl.b32 %r40, %r6, 2;
and.b32 %r41, %r40, 12;
or.b32 %r42, %r39, %r41;
cvt.u64.u32 %rd40, %r42;
cvt.u64.u32 %rd41, %r41;
cvt.u64.u32 %rd42, %r9;
mad.lo.s32 %r43, %r1, %r26, %r471;
add.s32 %r44, %r43, %r37;
mul.wide.s32 %rd6, %r44, 176;
add.s64 %rd43, %rd6, %rd40;
add.s64 %rd7, %rd43, 64;
shr.u32 %r45, %r6, 1;
cvt.u64.u32 %rd44, %r45;
and.b64 %rd45, %rd44, 2147483640;
or.b64 %rd46, %rd45, %rd42;
mul.lo.s32 %r46, %r3, %r27;
cvt.u64.u32 %rd47, %r46;
add.s64 %rd48, %rd46, %rd47;
mul.hi.u32 %r47, %r4, %r22;
add.s32 %r48, %r47, %r4;
shr.u32 %r49, %r48, %r23;
mul.lo.s32 %r50, %r49, %r24;
sub.s32 %r51, %r4, %r50;
mul.lo.s32 %r52, %r51, %r30;
cvt.u64.u32 %rd49, %r52;
add.s64 %rd50, %rd48, %rd49;
mul.lo.s64 %rd51, %rd50, 36;
cvta.to.global.u64 %rd52, %rd27;
add.s64 %rd53, %rd52, %rd51;
add.s64 %rd109, %rd53, 56;
or.b64 %rd54, %rd41, %rd6;
add.s64 %rd9, %rd54, 32;
and.b64 %rd55, %rd42, 2;
or.b64 %rd56, %rd55, %rd6;
or.b64 %rd10, %rd56, 12;
add.s32 %r53, %r1, 1;
add.s32 %r54, %r471, %r37;
mad.lo.s32 %r55, %r26, %r53, %r54;
mul.wide.s32 %rd11, %r55, 176;
or.b64 %rd57, %rd55, %rd11;
or.b64 %rd12, %rd57, 8;
or.b64 %rd13, %rd41, %rd11;
add.s64 %rd14, %rd11, %rd40;
cvta.to.global.u64 %rd110, %rd26;
cvta.to.global.u64 %rd108, %rd30;
mov.f32 %f206, 0f00000000;

$L__BB168_3:
add.s64 %rd58, %rd110, %rd7;
ld.global.nc.u32 %r92, [%rd58+-16];
ld.global.nc.u32 %r93, [%rd58];
add.s64 %rd59, %rd110, %rd9;
ld.global.nc.u32 %r94, [%rd59+-16];
shr.s32 %r95, %r94, %r9;
ld.global.nc.u32 %r96, [%rd59];
shr.s32 %r97, %r96, %r9;
add.s64 %rd60, %rd110, %rd10;
ld.global.nc.u16 %rs1, [%rd60+-8];
cvt.u32.u16 %r98, %rs1;
ld.global.nc.u16 %rs2, [%rd60+-4];
cvt.u32.u16 %r99, %rs2;
ld.global.nc.u16 %rs3, [%rd60];
cvt.u32.u16 %r100, %rs3;
and.b32 %r101, %r98, 16191;
and.b32 %r102, %r100, 3855;
and.b32 %r103, %r98, 49344;
shr.u32 %r104, %r103, 2;
or.b32 %r105, %r102, %r104;
setp.gt.u32 %p3, %r8, 7;
and.b32 %r106, %r99, 16191;
shr.u32 %r107, %r100, 4;
and.b32 %r108, %r107, 3855;
and.b32 %r109, %r99, 49344;
shr.u32 %r110, %r109, 2;
or.b32 %r111, %r108, %r110;
ld.global.nc.u32 %r11, [%rd109+-56];

{.reg .f16 low,high;
mov.b32 {low,high},%r11;
cvt.f32.f16 %f42, low;}


add.s64 %rd62, %rd109, %rd41;
ld.global.nc.u32 %r12, [%rd62+-52];
ld.global.nc.u32 %r13, [%rd62+-36];
ld.global.nc.u32 %r14, [%rd109+-20];

{.reg .f16 low,high;
mov.b32 {low,high},%r14;
cvt.f32.f16 %f43, low;}


ld.global.nc.u32 %r15, [%rd62+-16];
ld.global.nc.u32 %r16, [%rd62];
add.s64 %rd63, %rd110, %rd6;
and.b32 %r115, %r92, 252645135;
and.b32 %r116, %r93, 252645135;
shl.b32 %r117, %r95, 4;
and.b32 %r118, %r117, 269488144;
shl.b32 %r119, %r97, 4;
and.b32 %r120, %r119, 269488144;
or.b32 %r63, %r115, %r118;
or.b32 %r59, %r116, %r120;
mov.u32 %r85, 0;

dp4a.s32.s32 %r58, %r59, %r13, %r85;


dp4a.s32.s32 %r62, %r63, %r12, %r58;

mov.u32 %r87, 16843009;

dp4a.s32.s32 %r66, %r87, %r13, %r85;


dp4a.s32.s32 %r70, %r87, %r12, %r66;

selp.b32 %r121, %r105, %r101, %p3;
cvt.u16.u32 %rs4, %r121;
and.b32 %r122, %r121, 63;
mul.lo.s32 %r123, %r62, %r122;
cvt.rn.f32.s32 %f46, %r123;
fma.rn.ftz.f32 %f47, %f42, %f46, 0f00000000;
selp.b32 %r124, %r111, %r106, %p3;
cvt.u16.u32 %rs5, %r124;
and.b32 %r125, %r124, 63;
mul.lo.s32 %r126, %r70, %r125;
cvt.rn.f32.s32 %f48, %r126;
fma.rn.ftz.f32 %f49, %f42, %f48, 0f00000000;
shr.u32 %r127, %r92, 4;
and.b32 %r128, %r127, 252645135;
shr.u32 %r129, %r93, 4;
and.b32 %r130, %r129, 252645135;
shl.b32 %r131, %r95, 3;
and.b32 %r132, %r131, 269488144;
shl.b32 %r133, %r97, 3;
and.b32 %r134, %r133, 269488144;
or.b32 %r79, %r128, %r132;
or.b32 %r75, %r130, %r134;

dp4a.s32.s32 %r74, %r75, %r16, %r85;


dp4a.s32.s32 %r78, %r79, %r15, %r74;


dp4a.s32.s32 %r82, %r87, %r16, %r85;


dp4a.s32.s32 %r86, %r87, %r15, %r82;

shr.u16 %rs6, %rs4, 8;
cvt.u32.u16 %r135, %rs6;
mul.lo.s32 %r136, %r78, %r135;
cvt.rn.f32.s32 %f50, %r136;
fma.rn.ftz.f32 %f51, %f43, %f50, %f47;
shr.u16 %rs7, %rs5, 8;
cvt.u32.u16 %r137, %rs7;
mul.lo.s32 %r138, %r86, %r137;
cvt.rn.f32.s32 %f52, %r138;
fma.rn.ftz.f32 %f53, %f43, %f52, %f49;
ld.global.nc.u32 %r90, [%rd63];

{.reg .f16 low,high;
mov.b32 {low,high},%r90;
cvt.f32.f16 %f44, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r90;
cvt.f32.f16 %f45, high;}


mul.ftz.f32 %f54, %f51, %f44;
mul.ftz.f32 %f55, %f53, %f45;
sub.ftz.f32 %f56, %f54, %f55;
add.ftz.f32 %f209, %f56, %f209;
setp.eq.s64 %p4, %rd30, 0;
@%p4 bra $L__BB168_5;

add.s64 %rd64, %rd108, %rd7;
ld.global.u32 %r175, [%rd64+-16];
ld.global.u32 %r176, [%rd64];
add.s64 %rd65, %rd108, %rd9;
ld.global.u32 %r177, [%rd65+-16];
shr.s32 %r178, %r177, %r9;
ld.global.u32 %r179, [%rd65];
shr.s32 %r180, %r179, %r9;
add.s64 %rd66, %rd108, %rd10;
ld.global.u16 %r181, [%rd66+-8];
ld.global.u16 %r182, [%rd66+-4];
ld.global.u16 %r183, [%rd66];
and.b32 %r184, %r181, 16191;
and.b32 %r185, %r183, 3855;
and.b32 %r186, %r181, 49344;
shr.u32 %r187, %r186, 2;
or.b32 %r188, %r185, %r187;
and.b32 %r189, %r182, 16191;
shr.u32 %r190, %r183, 4;
and.b32 %r191, %r190, 3855;
and.b32 %r192, %r182, 49344;
shr.u32 %r193, %r192, 2;
or.b32 %r194, %r191, %r193;

{.reg .f16 low,high;
mov.b32 {low,high},%r11;
cvt.f32.f16 %f57, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r14;
cvt.f32.f16 %f58, low;}


add.s64 %rd67, %rd108, %rd6;
and.b32 %r195, %r175, 252645135;
and.b32 %r196, %r176, 252645135;
shl.b32 %r197, %r178, 4;
and.b32 %r198, %r197, 269488144;
shl.b32 %r199, %r180, 4;
and.b32 %r200, %r199, 269488144;
or.b32 %r146, %r195, %r198;
or.b32 %r142, %r196, %r200;

dp4a.s32.s32 %r141, %r142, %r13, %r85;


dp4a.s32.s32 %r145, %r146, %r12, %r141;


dp4a.s32.s32 %r149, %r87, %r13, %r85;


dp4a.s32.s32 %r153, %r87, %r12, %r149;

selp.b32 %r201, %r188, %r184, %p3;
cvt.u16.u32 %rs8, %r201;
and.b32 %r202, %r201, 63;
mul.lo.s32 %r203, %r145, %r202;
cvt.rn.f32.s32 %f61, %r203;
fma.rn.ftz.f32 %f62, %f57, %f61, 0f00000000;
selp.b32 %r204, %r194, %r189, %p3;
cvt.u16.u32 %rs9, %r204;
and.b32 %r205, %r204, 63;
mul.lo.s32 %r206, %r153, %r205;
cvt.rn.f32.s32 %f63, %r206;
fma.rn.ftz.f32 %f64, %f57, %f63, 0f00000000;
shr.u32 %r207, %r175, 4;
and.b32 %r208, %r207, 252645135;
shr.u32 %r209, %r176, 4;
and.b32 %r210, %r209, 252645135;
shl.b32 %r211, %r178, 3;
and.b32 %r212, %r211, 269488144;
shl.b32 %r213, %r180, 3;
and.b32 %r214, %r213, 269488144;
or.b32 %r162, %r208, %r212;
or.b32 %r158, %r210, %r214;

dp4a.s32.s32 %r157, %r158, %r16, %r85;


dp4a.s32.s32 %r161, %r162, %r15, %r157;


dp4a.s32.s32 %r165, %r87, %r16, %r85;


dp4a.s32.s32 %r169, %r87, %r15, %r165;

shr.u16 %rs10, %rs8, 8;
cvt.u32.u16 %r215, %rs10;
mul.lo.s32 %r216, %r161, %r215;
cvt.rn.f32.s32 %f65, %r216;
fma.rn.ftz.f32 %f66, %f58, %f65, %f62;
shr.u16 %rs11, %rs9, 8;
cvt.u32.u16 %r217, %rs11;
mul.lo.s32 %r218, %r169, %r217;
cvt.rn.f32.s32 %f67, %r218;
fma.rn.ftz.f32 %f68, %f58, %f67, %f64;
ld.global.u32 %r174, [%rd67];

{.reg .f16 low,high;
mov.b32 {low,high},%r174;
cvt.f32.f16 %f59, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r174;
cvt.f32.f16 %f60, high;}


mul.ftz.f32 %f69, %f66, %f59;
mul.ftz.f32 %f70, %f68, %f60;
sub.ftz.f32 %f71, %f69, %f70;
add.ftz.f32 %f207, %f71, %f207;

$L__BB168_5:
add.s64 %rd68, %rd110, %rd14;
add.s64 %rd69, %rd110, %rd13;
ld.global.nc.u32 %r255, [%rd68+48];
ld.global.nc.u32 %r256, [%rd68+64];
ld.global.nc.u32 %r257, [%rd69+16];
shr.s32 %r258, %r257, %r9;
ld.global.nc.u32 %r259, [%rd69+32];
shr.s32 %r260, %r259, %r9;
add.s64 %rd70, %rd110, %rd12;
ld.global.nc.u16 %rs12, [%rd70+-4];
cvt.u32.u16 %r261, %rs12;
ld.global.nc.u16 %rs13, [%rd70];
cvt.u32.u16 %r262, %rs13;
ld.global.nc.u16 %rs14, [%rd70+4];
cvt.u32.u16 %r263, %rs14;
and.b32 %r264, %r261, 16191;
and.b32 %r265, %r263, 3855;
and.b32 %r266, %r261, 49344;
shr.u32 %r267, %r266, 2;
or.b32 %r268, %r265, %r267;
and.b32 %r269, %r262, 16191;
shr.u32 %r270, %r263, 4;
and.b32 %r271, %r270, 3855;
and.b32 %r272, %r262, 49344;
shr.u32 %r273, %r272, 2;
or.b32 %r274, %r271, %r273;

{.reg .f16 low,high;
mov.b32 {low,high},%r11;
cvt.f32.f16 %f72, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r14;
cvt.f32.f16 %f73, low;}


add.s64 %rd71, %rd110, %rd11;
and.b32 %r275, %r255, 252645135;
and.b32 %r276, %r256, 252645135;
shl.b32 %r277, %r258, 4;
and.b32 %r278, %r277, 269488144;
shl.b32 %r279, %r260, 4;
and.b32 %r280, %r279, 269488144;
or.b32 %r226, %r275, %r278;
or.b32 %r222, %r276, %r280;

dp4a.s32.s32 %r221, %r222, %r13, %r85;


dp4a.s32.s32 %r225, %r226, %r12, %r221;


dp4a.s32.s32 %r229, %r87, %r13, %r85;


dp4a.s32.s32 %r233, %r87, %r12, %r229;

selp.b32 %r281, %r268, %r264, %p3;
cvt.u16.u32 %rs15, %r281;
and.b32 %r282, %r281, 63;
mul.lo.s32 %r283, %r225, %r282;
cvt.rn.f32.s32 %f76, %r283;
fma.rn.ftz.f32 %f77, %f72, %f76, 0f00000000;
selp.b32 %r284, %r274, %r269, %p3;
cvt.u16.u32 %rs16, %r284;
and.b32 %r285, %r284, 63;
mul.lo.s32 %r286, %r233, %r285;
cvt.rn.f32.s32 %f78, %r286;
fma.rn.ftz.f32 %f79, %f72, %f78, 0f00000000;
shr.u32 %r287, %r255, 4;
and.b32 %r288, %r287, 252645135;
shr.u32 %r289, %r256, 4;
and.b32 %r290, %r289, 252645135;
shl.b32 %r291, %r258, 3;
and.b32 %r292, %r291, 269488144;
shl.b32 %r293, %r260, 3;
and.b32 %r294, %r293, 269488144;
or.b32 %r242, %r288, %r292;
or.b32 %r238, %r290, %r294;

dp4a.s32.s32 %r237, %r238, %r16, %r85;


dp4a.s32.s32 %r241, %r242, %r15, %r237;


dp4a.s32.s32 %r245, %r87, %r16, %r85;


dp4a.s32.s32 %r249, %r87, %r15, %r245;

shr.u16 %rs17, %rs15, 8;
cvt.u32.u16 %r295, %rs17;
mul.lo.s32 %r296, %r241, %r295;
cvt.rn.f32.s32 %f80, %r296;
fma.rn.ftz.f32 %f81, %f73, %f80, %f77;
shr.u16 %rs18, %rs16, 8;
cvt.u32.u16 %r297, %rs18;
mul.lo.s32 %r298, %r249, %r297;
cvt.rn.f32.s32 %f82, %r298;
fma.rn.ftz.f32 %f83, %f73, %f82, %f79;
ld.global.nc.u32 %r253, [%rd71];

{.reg .f16 low,high;
mov.b32 {low,high},%r253;
cvt.f32.f16 %f74, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r253;
cvt.f32.f16 %f75, high;}


mul.ftz.f32 %f84, %f81, %f74;
mul.ftz.f32 %f85, %f83, %f75;
sub.ftz.f32 %f86, %f84, %f85;
add.ftz.f32 %f208, %f86, %f208;
@%p4 bra $L__BB168_7;

add.s64 %rd72, %rd108, %rd14;
add.s64 %rd73, %rd108, %rd13;
ld.global.u32 %r335, [%rd72+48];
ld.global.u32 %r336, [%rd72+64];
ld.global.u32 %r337, [%rd73+16];
shr.s32 %r338, %r337, %r9;
ld.global.u32 %r339, [%rd73+32];
shr.s32 %r340, %r339, %r9;
add.s64 %rd74, %rd108, %rd12;
ld.global.u16 %r341, [%rd74+-4];
ld.global.u16 %r342, [%rd74];
ld.global.u16 %r343, [%rd74+4];
and.b32 %r344, %r341, 16191;
and.b32 %r345, %r343, 3855;
and.b32 %r346, %r341, 49344;
shr.u32 %r347, %r346, 2;
or.b32 %r348, %r345, %r347;
and.b32 %r349, %r342, 16191;
shr.u32 %r350, %r343, 4;
and.b32 %r351, %r350, 3855;
and.b32 %r352, %r342, 49344;
shr.u32 %r353, %r352, 2;
or.b32 %r354, %r351, %r353;

{.reg .f16 low,high;
mov.b32 {low,high},%r11;
cvt.f32.f16 %f87, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r14;
cvt.f32.f16 %f88, low;}


add.s64 %rd75, %rd108, %rd11;
and.b32 %r355, %r335, 252645135;
and.b32 %r356, %r336, 252645135;
shl.b32 %r357, %r338, 4;
and.b32 %r358, %r357, 269488144;
shl.b32 %r359, %r340, 4;
and.b32 %r360, %r359, 269488144;
or.b32 %r306, %r355, %r358;
or.b32 %r302, %r356, %r360;
mov.u32 %r328, 0;

dp4a.s32.s32 %r301, %r302, %r13, %r328;


dp4a.s32.s32 %r305, %r306, %r12, %r301;

mov.u32 %r330, 16843009;

dp4a.s32.s32 %r309, %r330, %r13, %r328;


dp4a.s32.s32 %r313, %r330, %r12, %r309;

selp.b32 %r361, %r348, %r344, %p3;
cvt.u16.u32 %rs19, %r361;
and.b32 %r362, %r361, 63;
mul.lo.s32 %r363, %r305, %r362;
cvt.rn.f32.s32 %f91, %r363;
fma.rn.ftz.f32 %f92, %f87, %f91, 0f00000000;
selp.b32 %r364, %r354, %r349, %p3;
cvt.u16.u32 %rs20, %r364;
and.b32 %r365, %r364, 63;
mul.lo.s32 %r366, %r313, %r365;
cvt.rn.f32.s32 %f93, %r366;
fma.rn.ftz.f32 %f94, %f87, %f93, 0f00000000;
shr.u32 %r367, %r335, 4;
and.b32 %r368, %r367, 252645135;
shr.u32 %r369, %r336, 4;
and.b32 %r370, %r369, 252645135;
shl.b32 %r371, %r338, 3;
and.b32 %r372, %r371, 269488144;
shl.b32 %r373, %r340, 3;
and.b32 %r374, %r373, 269488144;
or.b32 %r322, %r368, %r372;
or.b32 %r318, %r370, %r374;

dp4a.s32.s32 %r317, %r318, %r16, %r328;


dp4a.s32.s32 %r321, %r322, %r15, %r317;


dp4a.s32.s32 %r325, %r330, %r16, %r328;


dp4a.s32.s32 %r329, %r330, %r15, %r325;

shr.u16 %rs21, %rs19, 8;
cvt.u32.u16 %r375, %rs21;
mul.lo.s32 %r376, %r321, %r375;
cvt.rn.f32.s32 %f95, %r376;
fma.rn.ftz.f32 %f96, %f88, %f95, %f92;
shr.u16 %rs22, %rs20, 8;
cvt.u32.u16 %r377, %rs22;
mul.lo.s32 %r378, %r329, %r377;
cvt.rn.f32.s32 %f97, %r378;
fma.rn.ftz.f32 %f98, %f88, %f97, %f94;
ld.global.u32 %r334, [%rd75];

{.reg .f16 low,high;
mov.b32 {low,high},%r334;
cvt.f32.f16 %f89, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r334;
cvt.f32.f16 %f90, high;}


mul.ftz.f32 %f99, %f96, %f89;
mul.ftz.f32 %f100, %f98, %f90;
sub.ftz.f32 %f101, %f99, %f100;
add.ftz.f32 %f206, %f101, %f206;

$L__BB168_7:
add.s64 %rd110, %rd110, 352;
add.s64 %rd109, %rd109, 576;
add.s64 %rd108, %rd108, 352;
add.s32 %r471, %r471, 2;
setp.lt.u32 %p9, %r471, %r2;
@%p9 bra $L__BB168_3;

add.u64 %rd96, %SPL, 8;
add.u64 %rd94, %SPL, 0;
st.local.v2.f32 [%rd94], {%f209, %f208};
st.local.v2.f32 [%rd96], {%f207, %f206};

$L__BB168_9:
mov.b32 %r18, %f209;
setp.ne.s64 %p10, %rd30, 0;
@%p10 bra $L__BB168_11;
bra.uni $L__BB168_10;

$L__BB168_11:
add.u64 %rd106, %SPL, 8;
add.u64 %rd104, %SPL, 0;
mov.u32 %r405, 31;
mov.u32 %r406, 16;
mov.u32 %r407, -1;
shfl.sync.bfly.b32 %r408|%p21, %r18, %r406, %r405, %r407;
mov.b32 %f122, %r408;
add.ftz.f32 %f123, %f209, %f122;
mov.b32 %r409, %f123;
mov.u32 %r410, 8;
shfl.sync.bfly.b32 %r411|%p22, %r409, %r410, %r405, %r407;
mov.b32 %f124, %r411;
add.ftz.f32 %f125, %f123, %f124;
mov.b32 %r412, %f125;
mov.u32 %r413, 4;
shfl.sync.bfly.b32 %r414|%p23, %r412, %r413, %r405, %r407;
mov.b32 %f126, %r414;
add.ftz.f32 %f127, %f125, %f126;
mov.b32 %r415, %f127;
mov.u32 %r416, 2;
shfl.sync.bfly.b32 %r417|%p24, %r415, %r416, %r405, %r407;
mov.b32 %f128, %r417;
add.ftz.f32 %f129, %f127, %f128;
mov.b32 %r418, %f129;
mov.u32 %r419, 1;
shfl.sync.bfly.b32 %r420|%p25, %r418, %r419, %r405, %r407;
mov.b32 %f130, %r420;
add.ftz.f32 %f131, %f129, %f130;
st.local.f32 [%rd104], %f131;
mov.b32 %r421, %f207;
shfl.sync.bfly.b32 %r422|%p26, %r421, %r406, %r405, %r407;
mov.b32 %f132, %r422;
add.ftz.f32 %f133, %f207, %f132;
mov.b32 %r423, %f133;
shfl.sync.bfly.b32 %r424|%p27, %r423, %r410, %r405, %r407;
mov.b32 %f134, %r424;
add.ftz.f32 %f135, %f133, %f134;
mov.b32 %r425, %f135;
shfl.sync.bfly.b32 %r426|%p28, %r425, %r413, %r405, %r407;
mov.b32 %f136, %r426;
add.ftz.f32 %f137, %f135, %f136;
mov.b32 %r427, %f137;
shfl.sync.bfly.b32 %r428|%p29, %r427, %r416, %r405, %r407;
mov.b32 %f138, %r428;
add.ftz.f32 %f139, %f137, %f138;
mov.b32 %r429, %f139;
shfl.sync.bfly.b32 %r430|%p30, %r429, %r419, %r405, %r407;
mov.b32 %f140, %r430;
add.ftz.f32 %f141, %f139, %f140;
st.local.f32 [%rd106], %f141;
mov.b32 %r431, %f208;
shfl.sync.bfly.b32 %r432|%p31, %r431, %r406, %r405, %r407;
mov.b32 %f142, %r432;
add.ftz.f32 %f143, %f208, %f142;
mov.b32 %r433, %f143;
shfl.sync.bfly.b32 %r434|%p32, %r433, %r410, %r405, %r407;
mov.b32 %f144, %r434;
add.ftz.f32 %f145, %f143, %f144;
mov.b32 %r435, %f145;
shfl.sync.bfly.b32 %r436|%p33, %r435, %r413, %r405, %r407;
mov.b32 %f146, %r436;
add.ftz.f32 %f147, %f145, %f146;
mov.b32 %r437, %f147;
shfl.sync.bfly.b32 %r438|%p34, %r437, %r416, %r405, %r407;
mov.b32 %f148, %r438;
add.ftz.f32 %f149, %f147, %f148;
mov.b32 %r439, %f149;
shfl.sync.bfly.b32 %r440|%p35, %r439, %r419, %r405, %r407;
mov.b32 %f150, %r440;
add.ftz.f32 %f151, %f149, %f150;
st.local.f32 [%rd104+4], %f151;
mov.b32 %r441, %f206;
shfl.sync.bfly.b32 %r442|%p36, %r441, %r406, %r405, %r407;
mov.b32 %f152, %r442;
add.ftz.f32 %f153, %f206, %f152;
mov.b32 %r443, %f153;
shfl.sync.bfly.b32 %r444|%p37, %r443, %r410, %r405, %r407;
mov.b32 %f154, %r444;
add.ftz.f32 %f155, %f153, %f154;
mov.b32 %r445, %f155;
shfl.sync.bfly.b32 %r446|%p38, %r445, %r413, %r405, %r407;
mov.b32 %f156, %r446;
add.ftz.f32 %f157, %f155, %f156;
mov.b32 %r447, %f157;
shfl.sync.bfly.b32 %r448|%p39, %r447, %r416, %r405, %r407;
mov.b32 %f158, %r448;
add.ftz.f32 %f159, %f157, %f158;
mov.b32 %r449, %f159;
shfl.sync.bfly.b32 %r450|%p40, %r449, %r419, %r405, %r407;
mov.b32 %f160, %r450;
add.ftz.f32 %f161, %f159, %f160;
st.local.f32 [%rd106+4], %f161;
bra.uni $L__BB168_12;

$L__BB168_10:
add.u64 %rd98, %SPL, 0;
mov.u32 %r379, 31;
mov.u32 %r380, 16;
mov.u32 %r381, -1;
shfl.sync.bfly.b32 %r382|%p11, %r18, %r380, %r379, %r381;
mov.b32 %f102, %r382;
add.ftz.f32 %f103, %f209, %f102;
mov.b32 %r383, %f103;
mov.u32 %r384, 8;
shfl.sync.bfly.b32 %r385|%p12, %r383, %r384, %r379, %r381;
mov.b32 %f104, %r385;
add.ftz.f32 %f105, %f103, %f104;
mov.b32 %r386, %f105;
mov.u32 %r387, 4;
shfl.sync.bfly.b32 %r388|%p13, %r386, %r387, %r379, %r381;
mov.b32 %f106, %r388;
add.ftz.f32 %f107, %f105, %f106;
mov.b32 %r389, %f107;
mov.u32 %r390, 2;
shfl.sync.bfly.b32 %r391|%p14, %r389, %r390, %r379, %r381;
mov.b32 %f108, %r391;
add.ftz.f32 %f109, %f107, %f108;
mov.b32 %r392, %f109;
mov.u32 %r393, 1;
shfl.sync.bfly.b32 %r394|%p15, %r392, %r393, %r379, %r381;
mov.b32 %f110, %r394;
add.ftz.f32 %f111, %f109, %f110;
st.local.f32 [%rd98], %f111;
mov.b32 %r395, %f208;
shfl.sync.bfly.b32 %r396|%p16, %r395, %r380, %r379, %r381;
mov.b32 %f112, %r396;
add.ftz.f32 %f113, %f208, %f112;
mov.b32 %r397, %f113;
shfl.sync.bfly.b32 %r398|%p17, %r397, %r384, %r379, %r381;
mov.b32 %f114, %r398;
add.ftz.f32 %f115, %f113, %f114;
mov.b32 %r399, %f115;
shfl.sync.bfly.b32 %r400|%p18, %r399, %r387, %r379, %r381;
mov.b32 %f116, %r400;
add.ftz.f32 %f117, %f115, %f116;
mov.b32 %r401, %f117;
shfl.sync.bfly.b32 %r402|%p19, %r401, %r390, %r379, %r381;
mov.b32 %f118, %r402;
add.ftz.f32 %f119, %f117, %f118;
mov.b32 %r403, %f119;
shfl.sync.bfly.b32 %r404|%p20, %r403, %r393, %r379, %r381;
mov.b32 %f120, %r404;
add.ftz.f32 %f121, %f119, %f120;
st.local.f32 [%rd98+4], %f121;

$L__BB168_12:
ld.param.u32 %r459, [_Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7];
mov.u32 %r458, %ctaid.x;
shl.b32 %r457, %r458, 1;
mov.u32 %r456, %tid.x;
setp.gt.u32 %p41, %r456, 1;
add.s32 %r19, %r457, %r456;
setp.ge.u32 %p42, %r19, %r459;
or.pred %p43, %p41, %p42;
@%p43 bra $L__BB168_29;

add.u64 %rd100, %SPL, 0;
ld.param.u64 %rd89, [_Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3];
ld.param.u32 %r461, [_Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13];
mov.u32 %r460, %tid.x;
cvt.u64.u32 %rd24, %r460;
mul.wide.u32 %rd76, %r460, 4;
add.s64 %rd77, %rd100, %rd76;
ld.local.f32 %f212, [%rd77];
mad.lo.s32 %r20, %r5, %r461, %r19;
setp.eq.s64 %p44, %rd89, 0;
@%p44 bra $L__BB168_15;

ld.param.u64 %rd90, [_Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3];
cvta.to.global.u64 %rd78, %rd90;
mul.wide.u32 %rd79, %r20, 4;
add.s64 %rd80, %rd78, %rd79;
ld.global.f32 %f162, [%rd80];
add.ftz.f32 %f212, %f212, %f162;

$L__BB168_15:
setp.eq.s64 %p45, %rd30, 0;
@%p45 bra $L__BB168_28;

add.u64 %rd102, %SPL, 8;
ld.param.u64 %rd92, [_Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+16];
shl.b64 %rd81, %rd24, 2;
add.s64 %rd82, %rd102, %rd81;
ld.local.f32 %f24, [%rd82];
setp.eq.s64 %p46, %rd92, 0;
@%p46 bra $L__BB168_18;

ld.param.u64 %rd93, [_Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+16];
cvta.to.global.u64 %rd83, %rd93;
mul.wide.u32 %rd84, %r20, 4;
add.s64 %rd85, %rd83, %rd84;
ld.global.f32 %f163, [%rd85];
add.ftz.f32 %f24, %f24, %f163;

$L__BB168_18:
ld.param.u32 %r466, [_Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.gt.s32 %p47, %r466, 2;
@%p47 bra $L__BB168_22;

ld.param.u32 %r469, [_Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.eq.s32 %p50, %r469, 1;
@%p50 bra $L__BB168_26;

ld.param.u32 %r470, [_Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.eq.s32 %p51, %r470, 2;
@%p51 bra $L__BB168_21;
bra.uni $L__BB168_27;

$L__BB168_21:
mul.ftz.f32 %f183, %f24, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f184, %f183;
add.ftz.f32 %f185, %f184, 0f3F800000;
div.approx.ftz.f32 %f186, %f24, %f185;
mul.ftz.f32 %f212, %f212, %f186;
bra.uni $L__BB168_28;

$L__BB168_22:
ld.param.u32 %r467, [_Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.eq.s32 %p48, %r467, 3;
@%p48 bra $L__BB168_25;

ld.param.u32 %r468, [_Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.ne.s32 %p49, %r468, 6;
@%p49 bra $L__BB168_27;

ld.param.f32 %f195, [_Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+44];
min.ftz.f32 %f164, %f24, %f195;
neg.ftz.f32 %f165, %f195;
min.ftz.f32 %f166, %f212, %f195;
max.ftz.f32 %f167, %f166, %f165;
mul.ftz.f32 %f168, %f164, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f169, %f168;
add.ftz.f32 %f170, %f169, 0f3F800000;
div.approx.ftz.f32 %f171, %f164, %f170;
mul.ftz.f32 %f212, %f167, %f171;
bra.uni $L__BB168_28;

$L__BB168_27:
mul.ftz.f32 %f212, %f212, %f24;
bra.uni $L__BB168_28;

$L__BB168_26:
mul.ftz.f32 %f187, %f24, 0f3F000000;
mul.ftz.f32 %f188, %f24, 0f3D372713;
fma.rn.ftz.f32 %f189, %f24, %f188, 0f3F800000;
mul.ftz.f32 %f190, %f24, 0f3F4C422A;
mul.ftz.f32 %f191, %f190, %f189;
tanh.approx.f32 %f192, %f191;
add.ftz.f32 %f193, %f192, 0f3F800000;
mul.ftz.f32 %f194, %f187, %f193;
mul.ftz.f32 %f212, %f212, %f194;
bra.uni $L__BB168_28;

$L__BB168_25:
mov.f32 %f172, 0f40E00000;
min.ftz.f32 %f173, %f24, %f172;
min.ftz.f32 %f174, %f212, %f172;
mov.f32 %f175, 0fC0E00000;
max.ftz.f32 %f176, %f174, %f175;
mul.ftz.f32 %f177, %f173, 0f3FD9DB23;
mul.ftz.f32 %f178, %f177, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f179, %f178;
add.ftz.f32 %f180, %f179, 0f3F800000;
div.approx.ftz.f32 %f181, %f173, %f180;
add.ftz.f32 %f182, %f176, 0f3F800000;
mul.ftz.f32 %f212, %f182, %f181;

$L__BB168_28:
ld.param.u64 %rd91, [_Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4];
ld.param.u32 %r465, [_Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10];
ld.param.u32 %r464, [_Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13];
mov.u32 %r463, %ctaid.y;
mov.u32 %r462, %tid.y;
mad.lo.s32 %r454, %r462, %r465, %r19;
mad.lo.s32 %r455, %r463, %r464, %r454;
cvta.to.global.u64 %rd86, %rd91;
mul.wide.u32 %rd87, %r455, 4;
add.s64 %rd88, %rd86, %rd87;
st.global.f32 [%rd88], %f212;

$L__BB168_29:
ret;

}
.entry _Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj(
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2,
.param .align 8 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3[48],
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5,
.param .align 4 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6[12],
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15
)
.maxntid 256, 1, 1
.minnctapersm 1
{
.local .align 8 .b8 __local_depot169[8];
.reg .b64 %SP;
.reg .b64 %SPL;
.reg .pred %p<18>;
.reg .b16 %rs<15>;
.reg .f32 %f<66>;
.reg .b32 %r<239>;
.reg .b64 %rd<61>;


mov.u64 %SPL, __local_depot169;
ld.param.u64 %rd18, [_Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0];
ld.param.u64 %rd19, [_Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1];
ld.param.u64 %rd20, [_Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2];
ld.param.u64 %rd21, [_Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4];
ld.param.u32 %r22, [_Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5];
ld.param.u32 %r14, [_Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7];
ld.param.u32 %r15, [_Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8];
ld.param.u32 %r16, [_Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9];
ld.param.u32 %r17, [_Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10];
ld.param.u32 %r18, [_Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11];
ld.param.u32 %r19, [_Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12];
ld.param.u32 %r20, [_Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13];
ld.param.u32 %r23, [_Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14];
ld.param.u32 %r21, [_Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15];
ld.param.u32 %r13, [_Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+8];
ld.param.u32 %r12, [_Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+4];
ld.param.u32 %r11, [_Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6];
add.u64 %rd1, %SPL, 0;
mov.u32 %r24, %ctaid.x;
shl.b32 %r1, %r24, 1;
shr.u32 %r2, %r22, 8;
mov.u32 %r3, %tid.y;
setp.ge.u32 %p1, %r3, %r23;
@%p1 bra $L__BB169_7;

mov.f32 %f64, 0f00000000;
st.local.v2.f32 [%rd1], {%f64, %f64};
mov.u32 %r4, %tid.x;
shr.u32 %r238, %r4, 4;
setp.ge.u32 %p2, %r238, %r2;
mov.f32 %f65, %f64;
@%p2 bra $L__BB169_5;

mov.u32 %r25, %ctaid.y;
mad.lo.s32 %r26, %r3, %r21, %r25;
cvta.to.global.u64 %rd23, %rd20;
mul.wide.u32 %rd24, %r26, 4;
add.s64 %rd25, %rd23, %rd24;
ld.global.nc.u32 %r27, [%rd25];
mul.lo.s32 %r28, %r27, %r18;
mul.lo.s32 %r29, %r3, %r16;
cvt.u64.u32 %rd26, %r29;
and.b32 %r6, %r4, 15;
shr.u32 %r30, %r6, 2;
shl.b32 %r7, %r30, 1;
shl.b32 %r31, %r30, 5;
shl.b32 %r32, %r4, 2;
and.b32 %r33, %r32, 12;
or.b32 %r34, %r31, %r33;
cvt.u64.u32 %rd27, %r34;
cvt.u64.u32 %rd3, %r33;
cvt.u64.u32 %rd28, %r7;
mad.lo.s32 %r35, %r1, %r15, %r238;
add.s32 %r36, %r35, %r28;
mul.wide.s32 %rd4, %r36, 176;
add.s64 %rd29, %rd4, %rd27;
add.s64 %rd5, %rd29, 64;
shr.u32 %r37, %r4, 1;
cvt.u64.u32 %rd30, %r37;
and.b64 %rd31, %rd30, 2147483640;
or.b64 %rd32, %rd31, %rd28;
add.s64 %rd33, %rd32, %rd26;
mul.hi.u32 %r38, %r25, %r11;
add.s32 %r39, %r38, %r25;
shr.u32 %r40, %r39, %r12;
mul.lo.s32 %r41, %r40, %r13;
sub.s32 %r42, %r25, %r41;
mul.lo.s32 %r43, %r42, %r19;
cvt.u64.u32 %rd34, %r43;
add.s64 %rd35, %rd33, %rd34;
mul.lo.s64 %rd36, %rd35, 36;
cvta.to.global.u64 %rd37, %rd19;
add.s64 %rd38, %rd37, %rd36;
add.s64 %rd59, %rd38, 56;
or.b64 %rd39, %rd3, %rd4;
add.s64 %rd7, %rd39, 32;
and.b64 %rd40, %rd28, 2;
or.b64 %rd41, %rd40, %rd4;
or.b64 %rd8, %rd41, 12;
add.s32 %r44, %r1, 1;
add.s32 %r45, %r238, %r28;
mad.lo.s32 %r46, %r15, %r44, %r45;
mul.wide.s32 %rd9, %r46, 176;
add.s64 %rd42, %rd9, %rd27;
add.s64 %rd10, %rd42, 64;
or.b64 %rd43, %rd40, %rd9;
or.b64 %rd11, %rd43, 8;
or.b64 %rd44, %rd3, %rd9;
add.s64 %rd12, %rd44, 32;
cvta.to.global.u64 %rd60, %rd18;
mov.f32 %f64, 0f00000000;

$L__BB169_3:
add.s64 %rd45, %rd60, %rd5;
ld.global.nc.u32 %r119, [%rd45+-16];
ld.global.nc.u32 %r120, [%rd45];
add.s64 %rd46, %rd60, %rd7;
ld.global.nc.u32 %r121, [%rd46+-16];
shr.s32 %r122, %r121, %r7;
ld.global.nc.u32 %r123, [%rd46];
shr.s32 %r124, %r123, %r7;
add.s64 %rd47, %rd60, %rd8;
ld.global.nc.u16 %rs1, [%rd47+-8];
cvt.u32.u16 %r125, %rs1;
ld.global.nc.u16 %rs2, [%rd47+-4];
cvt.u32.u16 %r126, %rs2;
ld.global.nc.u16 %rs3, [%rd47];
cvt.u32.u16 %r127, %rs3;
and.b32 %r128, %r125, 16191;
and.b32 %r129, %r127, 3855;
and.b32 %r130, %r125, 49344;
shr.u32 %r131, %r130, 2;
or.b32 %r132, %r129, %r131;
setp.gt.u32 %p3, %r6, 7;
and.b32 %r133, %r126, 16191;
shr.u32 %r134, %r127, 4;
and.b32 %r135, %r134, 3855;
and.b32 %r136, %r126, 49344;
shr.u32 %r137, %r136, 2;
or.b32 %r138, %r135, %r137;
ld.global.nc.u32 %r47, [%rd59+-56];

{.reg .f16 low,high;
mov.b32 {low,high},%r47;
cvt.f32.f16 %f11, low;}


add.s64 %rd48, %rd59, %rd3;
ld.global.nc.u32 %r55, [%rd48+-52];
ld.global.nc.u32 %r51, [%rd48+-36];
ld.global.nc.u32 %r48, [%rd59+-20];

{.reg .f16 low,high;
mov.b32 {low,high},%r48;
cvt.f32.f16 %f12, low;}


ld.global.nc.u32 %r71, [%rd48+-16];
ld.global.nc.u32 %r67, [%rd48];
add.s64 %rd49, %rd60, %rd4;
and.b32 %r139, %r119, 252645135;
and.b32 %r140, %r120, 252645135;
shl.b32 %r141, %r122, 4;
and.b32 %r142, %r141, 269488144;
shl.b32 %r143, %r124, 4;
and.b32 %r144, %r143, 269488144;
or.b32 %r54, %r139, %r142;
or.b32 %r50, %r140, %r144;
mov.u32 %r112, 0;

dp4a.s32.s32 %r49, %r50, %r51, %r112;


dp4a.s32.s32 %r53, %r54, %r55, %r49;

mov.u32 %r114, 16843009;

dp4a.s32.s32 %r57, %r114, %r51, %r112;


dp4a.s32.s32 %r61, %r114, %r55, %r57;

selp.b32 %r145, %r132, %r128, %p3;
cvt.u16.u32 %rs4, %r145;
and.b32 %r146, %r145, 63;
mul.lo.s32 %r147, %r53, %r146;
cvt.rn.f32.s32 %f19, %r147;
fma.rn.ftz.f32 %f20, %f11, %f19, 0f00000000;
selp.b32 %r148, %r138, %r133, %p3;
cvt.u16.u32 %rs5, %r148;
and.b32 %r149, %r148, 63;
mul.lo.s32 %r150, %r61, %r149;
cvt.rn.f32.s32 %f21, %r150;
fma.rn.ftz.f32 %f22, %f11, %f21, 0f00000000;
shr.u32 %r151, %r119, 4;
and.b32 %r152, %r151, 252645135;
shr.u32 %r153, %r120, 4;
and.b32 %r154, %r153, 252645135;
shl.b32 %r155, %r122, 3;
and.b32 %r156, %r155, 269488144;
shl.b32 %r157, %r124, 3;
and.b32 %r158, %r157, 269488144;
or.b32 %r70, %r152, %r156;
or.b32 %r66, %r154, %r158;

dp4a.s32.s32 %r65, %r66, %r67, %r112;


dp4a.s32.s32 %r69, %r70, %r71, %r65;


dp4a.s32.s32 %r73, %r114, %r67, %r112;


dp4a.s32.s32 %r77, %r114, %r71, %r73;

shr.u16 %rs6, %rs4, 8;
cvt.u32.u16 %r159, %rs6;
mul.lo.s32 %r160, %r69, %r159;
cvt.rn.f32.s32 %f23, %r160;
fma.rn.ftz.f32 %f24, %f12, %f23, %f20;
shr.u16 %rs7, %rs5, 8;
cvt.u32.u16 %r161, %rs7;
mul.lo.s32 %r162, %r77, %r161;
cvt.rn.f32.s32 %f25, %r162;
fma.rn.ftz.f32 %f26, %f12, %f25, %f22;
ld.global.nc.u32 %r81, [%rd49];

{.reg .f16 low,high;
mov.b32 {low,high},%r81;
cvt.f32.f16 %f13, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r81;
cvt.f32.f16 %f14, high;}


mul.ftz.f32 %f27, %f24, %f13;
mul.ftz.f32 %f28, %f26, %f14;
sub.ftz.f32 %f29, %f27, %f28;
add.ftz.f32 %f65, %f29, %f65;
add.s64 %rd50, %rd60, %rd10;
ld.global.nc.u32 %r163, [%rd50+-16];
ld.global.nc.u32 %r164, [%rd50];
add.s64 %rd51, %rd60, %rd12;
ld.global.nc.u32 %r165, [%rd51+-16];
shr.s32 %r166, %r165, %r7;
ld.global.nc.u32 %r167, [%rd51];
shr.s32 %r168, %r167, %r7;
add.s64 %rd52, %rd60, %rd11;
ld.global.nc.u16 %rs8, [%rd52+-4];
cvt.u32.u16 %r169, %rs8;
ld.global.nc.u16 %rs9, [%rd52];
cvt.u32.u16 %r170, %rs9;
ld.global.nc.u16 %rs10, [%rd52+4];
cvt.u32.u16 %r171, %rs10;
and.b32 %r172, %r169, 16191;
and.b32 %r173, %r171, 3855;
and.b32 %r174, %r169, 49344;
shr.u32 %r175, %r174, 2;
or.b32 %r176, %r173, %r175;
and.b32 %r177, %r170, 16191;
shr.u32 %r178, %r171, 4;
and.b32 %r179, %r178, 3855;
and.b32 %r180, %r170, 49344;
shr.u32 %r181, %r180, 2;
or.b32 %r182, %r179, %r181;

{.reg .f16 low,high;
mov.b32 {low,high},%r47;
cvt.f32.f16 %f15, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r48;
cvt.f32.f16 %f16, low;}


add.s64 %rd53, %rd60, %rd9;
and.b32 %r183, %r163, 252645135;
and.b32 %r184, %r164, 252645135;
shl.b32 %r185, %r166, 4;
and.b32 %r186, %r185, 269488144;
shl.b32 %r187, %r168, 4;
and.b32 %r188, %r187, 269488144;
or.b32 %r90, %r183, %r186;
or.b32 %r86, %r184, %r188;

dp4a.s32.s32 %r85, %r86, %r51, %r112;


dp4a.s32.s32 %r89, %r90, %r55, %r85;


dp4a.s32.s32 %r93, %r114, %r51, %r112;


dp4a.s32.s32 %r97, %r114, %r55, %r93;

selp.b32 %r189, %r176, %r172, %p3;
cvt.u16.u32 %rs11, %r189;
and.b32 %r190, %r189, 63;
mul.lo.s32 %r191, %r89, %r190;
cvt.rn.f32.s32 %f30, %r191;
fma.rn.ftz.f32 %f31, %f15, %f30, 0f00000000;
selp.b32 %r192, %r182, %r177, %p3;
cvt.u16.u32 %rs12, %r192;
and.b32 %r193, %r192, 63;
mul.lo.s32 %r194, %r97, %r193;
cvt.rn.f32.s32 %f32, %r194;
fma.rn.ftz.f32 %f33, %f15, %f32, 0f00000000;
shr.u32 %r195, %r163, 4;
and.b32 %r196, %r195, 252645135;
shr.u32 %r197, %r164, 4;
and.b32 %r198, %r197, 252645135;
shl.b32 %r199, %r166, 3;
and.b32 %r200, %r199, 269488144;
shl.b32 %r201, %r168, 3;
and.b32 %r202, %r201, 269488144;
or.b32 %r106, %r196, %r200;
or.b32 %r102, %r198, %r202;

dp4a.s32.s32 %r101, %r102, %r67, %r112;


dp4a.s32.s32 %r105, %r106, %r71, %r101;


dp4a.s32.s32 %r109, %r114, %r67, %r112;


dp4a.s32.s32 %r113, %r114, %r71, %r109;

shr.u16 %rs13, %rs11, 8;
cvt.u32.u16 %r203, %rs13;
mul.lo.s32 %r204, %r105, %r203;
cvt.rn.f32.s32 %f34, %r204;
fma.rn.ftz.f32 %f35, %f16, %f34, %f31;
shr.u16 %rs14, %rs12, 8;
cvt.u32.u16 %r205, %rs14;
mul.lo.s32 %r206, %r113, %r205;
cvt.rn.f32.s32 %f36, %r206;
fma.rn.ftz.f32 %f37, %f16, %f36, %f33;
ld.global.nc.u32 %r117, [%rd53];

{.reg .f16 low,high;
mov.b32 {low,high},%r117;
cvt.f32.f16 %f17, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r117;
cvt.f32.f16 %f18, high;}


mul.ftz.f32 %f38, %f35, %f17;
mul.ftz.f32 %f39, %f37, %f18;
sub.ftz.f32 %f40, %f38, %f39;
add.ftz.f32 %f64, %f40, %f64;
add.s64 %rd60, %rd60, 352;
add.s64 %rd59, %rd59, 576;
add.s32 %r238, %r238, 2;
setp.lt.u32 %p4, %r238, %r2;
@%p4 bra $L__BB169_3;

st.local.v2.f32 [%rd1], {%f65, %f64};

$L__BB169_5:
mov.b32 %r207, %f65;
mov.u32 %r208, 31;
mov.u32 %r209, 16;
mov.u32 %r210, -1;
shfl.sync.bfly.b32 %r211|%p5, %r207, %r209, %r208, %r210;
mov.b32 %f41, %r211;
add.ftz.f32 %f42, %f65, %f41;
mov.b32 %r212, %f42;
mov.u32 %r213, 8;
shfl.sync.bfly.b32 %r214|%p6, %r212, %r213, %r208, %r210;
mov.b32 %f43, %r214;
add.ftz.f32 %f44, %f42, %f43;
mov.b32 %r215, %f44;
mov.u32 %r216, 4;
shfl.sync.bfly.b32 %r217|%p7, %r215, %r216, %r208, %r210;
mov.b32 %f45, %r217;
add.ftz.f32 %f46, %f44, %f45;
mov.b32 %r218, %f46;
mov.u32 %r219, 2;
shfl.sync.bfly.b32 %r220|%p8, %r218, %r219, %r208, %r210;
mov.b32 %f47, %r220;
add.ftz.f32 %f48, %f46, %f47;
mov.b32 %r221, %f48;
mov.u32 %r222, 1;
shfl.sync.bfly.b32 %r223|%p9, %r221, %r222, %r208, %r210;
mov.b32 %f49, %r223;
add.ftz.f32 %f50, %f48, %f49;
st.local.f32 [%rd1], %f50;
mov.b32 %r224, %f64;
shfl.sync.bfly.b32 %r225|%p10, %r224, %r209, %r208, %r210;
mov.b32 %f51, %r225;
add.ftz.f32 %f52, %f64, %f51;
mov.b32 %r226, %f52;
shfl.sync.bfly.b32 %r227|%p11, %r226, %r213, %r208, %r210;
mov.b32 %f53, %r227;
add.ftz.f32 %f54, %f52, %f53;
mov.b32 %r228, %f54;
shfl.sync.bfly.b32 %r229|%p12, %r228, %r216, %r208, %r210;
mov.b32 %f55, %r229;
add.ftz.f32 %f56, %f54, %f55;
mov.b32 %r230, %f56;
shfl.sync.bfly.b32 %r231|%p13, %r230, %r219, %r208, %r210;
mov.b32 %f57, %r231;
add.ftz.f32 %f58, %f56, %f57;
mov.b32 %r232, %f58;
shfl.sync.bfly.b32 %r233|%p14, %r232, %r222, %r208, %r210;
mov.b32 %f59, %r233;
add.ftz.f32 %f60, %f58, %f59;
st.local.f32 [%rd1+4], %f60;
add.s32 %r10, %r1, %r4;
setp.ge.u32 %p15, %r10, %r14;
setp.gt.u32 %p16, %r4, 1;
or.pred %p17, %p16, %p15;
@%p17 bra $L__BB169_7;

mul.wide.u32 %rd54, %r4, 4;
add.s64 %rd55, %rd1, %rd54;
ld.local.f32 %f61, [%rd55];
mov.u32 %r234, %ctaid.y;
mad.lo.s32 %r236, %r3, %r17, %r10;
mad.lo.s32 %r237, %r234, %r20, %r236;
cvta.to.global.u64 %rd56, %rd21;
mul.wide.u32 %rd57, %r237, 4;
add.s64 %rd58, %rd56, %rd57;
st.global.f32 [%rd58], %f61;

$L__BB169_7:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<328>;
.reg .b16 %rs<45>;
.reg .f32 %f<1093>;
.reg .b32 %r<1524>;
.reg .b64 %rd<142>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate[1536];

ld.param.u64 %rd39, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd40, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd41, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r48, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r49, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r53, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r54, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r55, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r59, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r60, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r61, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r57, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r56, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r51, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r50, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r47, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r46, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r45, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
ld.param.u64 %rd44, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+16];
ld.param.u64 %rd43, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+8];
ld.param.u64 %rd42, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3];
mov.u32 %r1, %tid.y;
shl.b32 %r62, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r62, %r2;
mov.u32 %r63, %ctaid.x;
shl.b32 %r4, %r63, 2;
mov.u32 %r5, %ctaid.y;
setp.eq.s64 %p1, %rd41, 0;
@%p1 bra $L__BB170_2;

cvta.to.global.u64 %rd48, %rd41;
mul.wide.u32 %rd49, %r5, 4;
add.s64 %rd50, %rd48, %rd49;
ld.global.nc.u32 %r1520, [%rd50];
mul.hi.u32 %r64, %r5, %r45;
add.s32 %r65, %r64, %r5;
shr.u32 %r66, %r65, %r46;
mul.lo.s32 %r67, %r66, %r47;
sub.s32 %r1521, %r5, %r67;
bra.uni $L__BB170_3;

$L__BB170_2:
mul.hi.u32 %r68, %r5, %r50;
add.s32 %r69, %r68, %r5;
shr.u32 %r1520, %r69, %r51;
mov.u32 %r1521, %r5;

$L__BB170_3:
mov.u32 %r11, %ctaid.z;
selp.b32 %r12, %r5, %r1520, %p1;
setp.ne.s32 %p3, %r1, 0;
setp.gt.u32 %p4, %r2, 3;
or.pred %p5, %p4, %p3;
add.s32 %r70, %r4, %r2;
setp.ge.u32 %p6, %r70, %r49;
mov.f32 %f27, 0f00000000;
or.pred %p7, %p5, %p6;
mov.f32 %f4, %f27;
mov.f32 %f5, %f27;
@%p7 bra $L__BB170_8;

setp.eq.s64 %p8, %rd42, 0;
mov.f32 %f5, 0f00000000;
mov.f32 %f4, %f5;
@%p8 bra $L__BB170_6;

cvta.to.global.u64 %rd51, %rd42;
mul.lo.s32 %r71, %r11, %r61;
cvt.u64.u32 %rd52, %r71;
mul.lo.s32 %r72, %r12, %r55;
cvt.u64.u32 %rd53, %r72;
cvt.u64.u32 %rd54, %r2;
cvt.s64.s32 %rd55, %r4;
add.s64 %rd56, %rd55, %rd54;
add.s64 %rd57, %rd56, %rd52;
add.s64 %rd58, %rd57, %rd53;
shl.b64 %rd59, %rd58, 2;
add.s64 %rd60, %rd51, %rd59;
ld.global.f32 %f4, [%rd60];

$L__BB170_6:
setp.eq.s64 %p9, %rd44, 0;
setp.eq.s64 %p10, %rd43, 0;
or.pred %p11, %p9, %p10;
@%p11 bra $L__BB170_8;

cvta.to.global.u64 %rd61, %rd44;
mul.lo.s32 %r73, %r11, %r61;
cvt.u64.u32 %rd62, %r73;
mul.lo.s32 %r74, %r12, %r55;
cvt.u64.u32 %rd63, %r74;
cvt.u64.u32 %rd64, %r2;
cvt.s64.s32 %rd65, %r4;
add.s64 %rd66, %rd65, %rd64;
add.s64 %rd67, %rd66, %rd62;
add.s64 %rd68, %rd67, %rd63;
shl.b64 %rd69, %rd68, 2;
add.s64 %rd70, %rd61, %rd69;
ld.global.f32 %f5, [%rd70];

$L__BB170_8:
mul.lo.s32 %r75, %r11, %r60;
cvt.u64.u32 %rd71, %r75;
mul.lo.s32 %r76, %r1521, %r54;
cvt.u64.u32 %rd72, %r76;
add.s64 %rd4, %rd72, %rd71;
mul.hi.u32 %r77, %r11, %r56;
add.s32 %r78, %r77, %r11;
shr.u32 %r13, %r78, %r57;
shr.s32 %r79, %r3, 31;
shr.u32 %r80, %r79, 28;
add.s32 %r81, %r3, %r80;
shr.s32 %r1523, %r81, 4;
shr.u32 %r82, %r44, 8;
setp.ge.s32 %p12, %r1523, %r82;
mov.f32 %f28, %f27;
mov.f32 %f29, %f27;
mov.f32 %f30, %f27;
mov.f32 %f31, %f27;
mov.f32 %f32, %f27;
mov.f32 %f33, %f27;
mov.f32 %f34, %f27;
@%p12 bra $L__BB170_19;

mul.lo.s32 %r83, %r13, %r59;
mul.lo.s32 %r84, %r1520, %r53;
and.b32 %r88, %r81, -16;
sub.s32 %r15, %r3, %r88;
shr.s32 %r89, %r15, 31;
shr.u32 %r90, %r89, 30;
add.s32 %r91, %r15, %r90;
shr.s32 %r92, %r91, 2;
shl.b32 %r93, %r92, 1;
shl.b32 %r94, %r92, 5;
and.b32 %r95, %r91, -4;
sub.s32 %r96, %r15, %r95;
shl.b32 %r97, %r96, 2;
add.s32 %r98, %r94, %r97;
cvt.s64.s32 %rd73, %r98;
cvt.s64.s32 %rd74, %r97;
cvt.s64.s32 %rd5, %r93;
cvt.s64.s32 %rd6, %r96;
shl.b32 %r1522, %r1523, 3;
add.s32 %r99, %r1523, %r83;
add.s32 %r17, %r99, %r84;
add.s32 %r100, %r4, 1;
mad.lo.s32 %r101, %r48, %r100, %r17;
mul.wide.s32 %rd75, %r101, 176;
mad.lo.s32 %r102, %r4, %r48, %r1523;
add.s32 %r103, %r102, %r83;
add.s32 %r104, %r103, %r84;
mul.wide.s32 %rd7, %r104, 176;
add.s64 %rd76, %rd7, %rd73;
add.s64 %rd8, %rd76, 64;
cvt.u64.u32 %rd77, %r93;
and.b64 %rd78, %rd77, 2;
or.b64 %rd79, %rd78, %rd75;
or.b64 %rd9, %rd79, 8;
add.s64 %rd80, %rd7, %rd74;
add.s64 %rd10, %rd80, 32;
or.b64 %rd81, %rd78, %rd7;
or.b64 %rd11, %rd81, 12;
add.s64 %rd82, %rd75, %rd74;
add.s64 %rd12, %rd82, 16;
add.s64 %rd83, %rd75, %rd73;
add.s64 %rd13, %rd83, 48;
add.s32 %r105, %r4, 2;
mad.lo.s32 %r106, %r48, %r105, %r17;
mul.wide.s32 %rd84, %r106, 176;
add.s32 %r107, %r4, 3;
mad.lo.s32 %r108, %r48, %r107, %r17;
mul.wide.s32 %rd85, %r108, 176;
add.s64 %rd86, %rd85, %rd73;
add.s64 %rd14, %rd86, 64;
or.b64 %rd87, %rd78, %rd84;
or.b64 %rd15, %rd87, 8;
add.s64 %rd88, %rd85, %rd74;
add.s64 %rd16, %rd88, 32;
or.b64 %rd89, %rd78, %rd85;
or.b64 %rd17, %rd89, 12;
add.s64 %rd90, %rd84, %rd74;
add.s64 %rd18, %rd90, 16;
add.s64 %rd91, %rd84, %rd73;
add.s64 %rd19, %rd91, 48;
cvta.to.global.u64 %rd141, %rd43;
cvta.to.global.u64 %rd140, %rd39;
mov.f32 %f27, 0f00000000;
cvta.to.global.u64 %rd92, %rd40;
shl.b64 %rd101, %rd6, 2;
mov.f32 %f28, %f27;
mov.f32 %f29, %f27;
mov.f32 %f30, %f27;
mov.f32 %f31, %f27;
mov.f32 %f32, %f27;
mov.f32 %f33, %f27;
mov.f32 %f34, %f27;

$L__BB170_10:
cvt.u64.u32 %rd93, %r1522;
add.s64 %rd94, %rd4, %rd93;
add.s64 %rd95, %rd140, %rd8;
add.s64 %rd96, %rd140, %rd10;
ld.global.nc.u32 %r145, [%rd95+-16];
ld.global.nc.u32 %r146, [%rd95];
ld.global.nc.u32 %r147, [%rd96+-16];
cvt.u32.u64 %r148, %rd5;
shr.s32 %r149, %r147, %r148;
ld.global.nc.u32 %r150, [%rd96];
shr.s32 %r151, %r150, %r148;
add.s64 %rd97, %rd140, %rd11;
ld.global.nc.u16 %rs1, [%rd97+-8];
cvt.u32.u16 %r152, %rs1;
ld.global.nc.u16 %rs2, [%rd97+-4];
cvt.u32.u16 %r153, %rs2;
ld.global.nc.u16 %rs3, [%rd97];
cvt.u32.u16 %r154, %rs3;
and.b32 %r155, %r152, 16191;
and.b32 %r156, %r154, 3855;
and.b32 %r157, %r152, 49344;
shr.u32 %r158, %r157, 2;
or.b32 %r159, %r156, %r158;
setp.gt.s32 %p13, %r15, 7;
and.b32 %r160, %r153, 16191;
shr.u32 %r161, %r154, 4;
and.b32 %r162, %r161, 3855;
and.b32 %r163, %r153, 49344;
shr.u32 %r164, %r163, 2;
or.b32 %r165, %r162, %r164;
add.s64 %rd98, %rd94, %rd5;
mul.lo.s64 %rd99, %rd98, 36;
add.s64 %rd100, %rd92, %rd99;
ld.global.nc.u32 %r20, [%rd100];

{.reg .f16 low,high;
mov.b32 {low,high},%r20;
cvt.f32.f16 %f101, low;}


add.s64 %rd102, %rd100, %rd101;
ld.global.nc.u32 %r21, [%rd102+4];
ld.global.nc.u32 %r22, [%rd102+20];
ld.global.nc.u32 %r23, [%rd100+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r23;
cvt.f32.f16 %f102, low;}


ld.global.nc.u32 %r24, [%rd102+40];
ld.global.nc.u32 %r25, [%rd102+56];
add.s64 %rd27, %rd140, %rd7;
and.b32 %r166, %r145, 252645135;
and.b32 %r167, %r146, 252645135;
shl.b32 %r168, %r149, 4;
and.b32 %r169, %r168, 269488144;
shl.b32 %r170, %r151, 4;
and.b32 %r171, %r170, 269488144;
or.b32 %r116, %r166, %r169;
or.b32 %r112, %r167, %r171;
mov.u32 %r138, 0;

dp4a.s32.s32 %r111, %r112, %r22, %r138;


dp4a.s32.s32 %r115, %r116, %r21, %r111;

mov.u32 %r140, 16843009;

dp4a.s32.s32 %r119, %r140, %r22, %r138;


dp4a.s32.s32 %r123, %r140, %r21, %r119;

selp.b32 %r172, %r159, %r155, %p13;
cvt.u16.u32 %rs4, %r172;
and.b32 %r173, %r172, 63;
mul.lo.s32 %r174, %r115, %r173;
cvt.rn.f32.s32 %f105, %r174;
fma.rn.ftz.f32 %f106, %f101, %f105, 0f00000000;
selp.b32 %r175, %r165, %r160, %p13;
cvt.u16.u32 %rs5, %r175;
and.b32 %r176, %r175, 63;
mul.lo.s32 %r177, %r123, %r176;
cvt.rn.f32.s32 %f107, %r177;
fma.rn.ftz.f32 %f108, %f101, %f107, 0f00000000;
shr.u32 %r178, %r145, 4;
and.b32 %r179, %r178, 252645135;
shr.u32 %r180, %r146, 4;
and.b32 %r181, %r180, 252645135;
shl.b32 %r182, %r149, 3;
and.b32 %r183, %r182, 269488144;
shl.b32 %r184, %r151, 3;
and.b32 %r185, %r184, 269488144;
or.b32 %r132, %r179, %r183;
or.b32 %r128, %r181, %r185;

dp4a.s32.s32 %r127, %r128, %r25, %r138;


dp4a.s32.s32 %r131, %r132, %r24, %r127;


dp4a.s32.s32 %r135, %r140, %r25, %r138;


dp4a.s32.s32 %r139, %r140, %r24, %r135;

shr.u16 %rs6, %rs4, 8;
cvt.u32.u16 %r186, %rs6;
mul.lo.s32 %r187, %r131, %r186;
cvt.rn.f32.s32 %f109, %r187;
fma.rn.ftz.f32 %f110, %f102, %f109, %f106;
shr.u16 %rs7, %rs5, 8;
cvt.u32.u16 %r188, %rs7;
mul.lo.s32 %r189, %r139, %r188;
cvt.rn.f32.s32 %f111, %r189;
fma.rn.ftz.f32 %f112, %f102, %f111, %f108;
ld.global.nc.u32 %r143, [%rd27];

{.reg .f16 low,high;
mov.b32 {low,high},%r143;
cvt.f32.f16 %f103, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r143;
cvt.f32.f16 %f104, high;}


mul.ftz.f32 %f113, %f110, %f103;
mul.ftz.f32 %f114, %f112, %f104;
sub.ftz.f32 %f115, %f113, %f114;
add.ftz.f32 %f30, %f115, %f30;
setp.eq.s64 %p14, %rd43, 0;
@%p14 bra $L__BB170_12;

cvt.u32.u64 %r1517, %rd5;
add.s64 %rd103, %rd141, %rd8;
ld.global.u32 %r227, [%rd103+-16];
ld.global.u32 %r228, [%rd103];
add.s64 %rd104, %rd141, %rd10;
ld.global.u32 %r229, [%rd104+-16];
shr.s32 %r230, %r229, %r1517;
ld.global.u32 %r231, [%rd104];
shr.s32 %r232, %r231, %r1517;
add.s64 %rd105, %rd141, %rd11;
ld.global.u16 %r233, [%rd105+-8];
ld.global.u16 %r234, [%rd105+-4];
ld.global.u16 %r235, [%rd105];
and.b32 %r236, %r233, 16191;
and.b32 %r237, %r235, 3855;
and.b32 %r238, %r233, 49344;
shr.u32 %r239, %r238, 2;
or.b32 %r240, %r237, %r239;
and.b32 %r241, %r234, 16191;
shr.u32 %r242, %r235, 4;
and.b32 %r243, %r242, 3855;
and.b32 %r244, %r234, 49344;
shr.u32 %r245, %r244, 2;
or.b32 %r246, %r243, %r245;

{.reg .f16 low,high;
mov.b32 {low,high},%r20;
cvt.f32.f16 %f116, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r23;
cvt.f32.f16 %f117, low;}


add.s64 %rd106, %rd141, %rd7;
and.b32 %r247, %r227, 252645135;
and.b32 %r248, %r228, 252645135;
shl.b32 %r249, %r230, 4;
and.b32 %r250, %r249, 269488144;
shl.b32 %r251, %r232, 4;
and.b32 %r252, %r251, 269488144;
or.b32 %r197, %r247, %r250;
or.b32 %r193, %r248, %r252;

dp4a.s32.s32 %r192, %r193, %r22, %r138;


dp4a.s32.s32 %r196, %r197, %r21, %r192;


dp4a.s32.s32 %r200, %r140, %r22, %r138;


dp4a.s32.s32 %r204, %r140, %r21, %r200;

selp.b32 %r253, %r240, %r236, %p13;
cvt.u16.u32 %rs8, %r253;
and.b32 %r254, %r253, 63;
mul.lo.s32 %r255, %r196, %r254;
cvt.rn.f32.s32 %f120, %r255;
fma.rn.ftz.f32 %f121, %f116, %f120, 0f00000000;
selp.b32 %r256, %r246, %r241, %p13;
cvt.u16.u32 %rs9, %r256;
and.b32 %r257, %r256, 63;
mul.lo.s32 %r258, %r204, %r257;
cvt.rn.f32.s32 %f122, %r258;
fma.rn.ftz.f32 %f123, %f116, %f122, 0f00000000;
shr.u32 %r259, %r227, 4;
and.b32 %r260, %r259, 252645135;
shr.u32 %r261, %r228, 4;
and.b32 %r262, %r261, 252645135;
shl.b32 %r263, %r230, 3;
and.b32 %r264, %r263, 269488144;
shl.b32 %r265, %r232, 3;
and.b32 %r266, %r265, 269488144;
or.b32 %r213, %r260, %r264;
or.b32 %r209, %r262, %r266;

dp4a.s32.s32 %r208, %r209, %r25, %r138;


dp4a.s32.s32 %r212, %r213, %r24, %r208;


dp4a.s32.s32 %r216, %r140, %r25, %r138;


dp4a.s32.s32 %r220, %r140, %r24, %r216;

shr.u16 %rs10, %rs8, 8;
cvt.u32.u16 %r267, %rs10;
mul.lo.s32 %r268, %r212, %r267;
cvt.rn.f32.s32 %f124, %r268;
fma.rn.ftz.f32 %f125, %f117, %f124, %f121;
shr.u16 %rs11, %rs9, 8;
cvt.u32.u16 %r269, %rs11;
mul.lo.s32 %r270, %r220, %r269;
cvt.rn.f32.s32 %f126, %r270;
fma.rn.ftz.f32 %f127, %f117, %f126, %f123;
ld.global.u32 %r225, [%rd106];

{.reg .f16 low,high;
mov.b32 {low,high},%r225;
cvt.f32.f16 %f118, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r225;
cvt.f32.f16 %f119, high;}


mul.ftz.f32 %f128, %f125, %f118;
mul.ftz.f32 %f129, %f127, %f119;
sub.ftz.f32 %f130, %f128, %f129;
add.ftz.f32 %f34, %f130, %f34;

$L__BB170_12:
add.s64 %rd139, %rd140, %rd11;
add.s64 %rd138, %rd139, -8;
add.s64 %rd137, %rd140, %rd10;
add.s64 %rd136, %rd137, -16;
add.s64 %rd135, %rd140, %rd8;
add.s64 %rd134, %rd135, -16;
cvt.u32.u64 %r1514, %rd5;
mul.wide.s32 %rd107, %r48, 176;
add.s64 %rd28, %rd134, %rd107;
ld.global.nc.u32 %r307, [%rd28];
ld.global.nc.u32 %r308, [%rd28+16];
add.s64 %rd29, %rd136, %rd107;
ld.global.nc.u32 %r309, [%rd29];
shr.s32 %r311, %r309, %r1514;
ld.global.nc.u32 %r312, [%rd29+16];
shr.s32 %r313, %r312, %r1514;
add.s64 %rd30, %rd138, %rd107;
ld.global.nc.u16 %rs12, [%rd30];
cvt.u32.u16 %r314, %rs12;
ld.global.nc.u16 %rs13, [%rd30+4];
cvt.u32.u16 %r315, %rs13;
ld.global.nc.u16 %rs14, [%rd30+8];
cvt.u32.u16 %r316, %rs14;
and.b32 %r317, %r314, 16191;
and.b32 %r318, %r316, 3855;
and.b32 %r319, %r314, 49344;
shr.u32 %r320, %r319, 2;
or.b32 %r321, %r318, %r320;
and.b32 %r322, %r315, 16191;
shr.u32 %r323, %r316, 4;
and.b32 %r324, %r323, 3855;
and.b32 %r325, %r315, 49344;
shr.u32 %r326, %r325, 2;
or.b32 %r327, %r324, %r326;

{.reg .f16 low,high;
mov.b32 {low,high},%r20;
cvt.f32.f16 %f131, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r23;
cvt.f32.f16 %f132, low;}


and.b32 %r328, %r307, 252645135;
and.b32 %r329, %r308, 252645135;
shl.b32 %r330, %r311, 4;
and.b32 %r331, %r330, 269488144;
shl.b32 %r332, %r313, 4;
and.b32 %r333, %r332, 269488144;
or.b32 %r278, %r328, %r331;
or.b32 %r274, %r329, %r333;

dp4a.s32.s32 %r273, %r274, %r22, %r138;


dp4a.s32.s32 %r277, %r278, %r21, %r273;


dp4a.s32.s32 %r281, %r140, %r22, %r138;


dp4a.s32.s32 %r285, %r140, %r21, %r281;

selp.b32 %r334, %r321, %r317, %p13;
cvt.u16.u32 %rs15, %r334;
and.b32 %r335, %r334, 63;
mul.lo.s32 %r336, %r277, %r335;
cvt.rn.f32.s32 %f135, %r336;
fma.rn.ftz.f32 %f136, %f131, %f135, 0f00000000;
selp.b32 %r337, %r327, %r322, %p13;
cvt.u16.u32 %rs16, %r337;
and.b32 %r338, %r337, 63;
mul.lo.s32 %r339, %r285, %r338;
cvt.rn.f32.s32 %f137, %r339;
fma.rn.ftz.f32 %f138, %f131, %f137, 0f00000000;
shr.u32 %r340, %r307, 4;
and.b32 %r341, %r340, 252645135;
shr.u32 %r342, %r308, 4;
and.b32 %r343, %r342, 252645135;
shl.b32 %r344, %r311, 3;
and.b32 %r345, %r344, 269488144;
shl.b32 %r346, %r313, 3;
and.b32 %r347, %r346, 269488144;
or.b32 %r294, %r341, %r345;
or.b32 %r290, %r343, %r347;

dp4a.s32.s32 %r289, %r290, %r25, %r138;


dp4a.s32.s32 %r293, %r294, %r24, %r289;


dp4a.s32.s32 %r297, %r140, %r25, %r138;


dp4a.s32.s32 %r301, %r140, %r24, %r297;

shr.u16 %rs17, %rs15, 8;
cvt.u32.u16 %r348, %rs17;
mul.lo.s32 %r349, %r293, %r348;
cvt.rn.f32.s32 %f139, %r349;
fma.rn.ftz.f32 %f140, %f132, %f139, %f136;
shr.u16 %rs18, %rs16, 8;
cvt.u32.u16 %r350, %rs18;
mul.lo.s32 %r351, %r301, %r350;
cvt.rn.f32.s32 %f141, %r351;
fma.rn.ftz.f32 %f142, %f132, %f141, %f138;
add.s64 %rd31, %rd27, %rd107;
ld.global.nc.u32 %r305, [%rd31];

{.reg .f16 low,high;
mov.b32 {low,high},%r305;
cvt.f32.f16 %f133, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r305;
cvt.f32.f16 %f134, high;}


mul.ftz.f32 %f143, %f140, %f133;
mul.ftz.f32 %f144, %f142, %f134;
sub.ftz.f32 %f145, %f143, %f144;
add.ftz.f32 %f29, %f145, %f29;
@%p14 bra $L__BB170_14;

cvt.u32.u64 %r1516, %rd5;
add.s64 %rd108, %rd141, %rd13;
add.s64 %rd109, %rd141, %rd12;
ld.global.u32 %r389, [%rd108];
ld.global.u32 %r390, [%rd108+16];
ld.global.u32 %r391, [%rd109];
shr.s32 %r392, %r391, %r1516;
ld.global.u32 %r393, [%rd109+16];
shr.s32 %r394, %r393, %r1516;
add.s64 %rd110, %rd141, %rd9;
ld.global.u16 %r395, [%rd110+-4];
ld.global.u16 %r396, [%rd110];
ld.global.u16 %r397, [%rd110+4];
and.b32 %r398, %r395, 16191;
and.b32 %r399, %r397, 3855;
and.b32 %r400, %r395, 49344;
shr.u32 %r401, %r400, 2;
or.b32 %r402, %r399, %r401;
and.b32 %r403, %r396, 16191;
shr.u32 %r404, %r397, 4;
and.b32 %r405, %r404, 3855;
and.b32 %r406, %r396, 49344;
shr.u32 %r407, %r406, 2;
or.b32 %r408, %r405, %r407;

{.reg .f16 low,high;
mov.b32 {low,high},%r20;
cvt.f32.f16 %f146, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r23;
cvt.f32.f16 %f147, low;}


or.b32 %r411, %r4, 1;
mad.lo.s32 %r412, %r48, %r411, %r17;
mul.wide.s32 %rd111, %r412, 176;
add.s64 %rd112, %rd141, %rd111;
and.b32 %r413, %r389, 252645135;
and.b32 %r414, %r390, 252645135;
shl.b32 %r415, %r392, 4;
and.b32 %r416, %r415, 269488144;
shl.b32 %r417, %r394, 4;
and.b32 %r418, %r417, 269488144;
or.b32 %r359, %r413, %r416;
or.b32 %r355, %r414, %r418;
mov.u32 %r381, 0;

dp4a.s32.s32 %r354, %r355, %r22, %r381;


dp4a.s32.s32 %r358, %r359, %r21, %r354;

mov.u32 %r383, 16843009;

dp4a.s32.s32 %r362, %r383, %r22, %r381;


dp4a.s32.s32 %r366, %r383, %r21, %r362;

selp.b32 %r419, %r402, %r398, %p13;
cvt.u16.u32 %rs19, %r419;
and.b32 %r420, %r419, 63;
mul.lo.s32 %r421, %r358, %r420;
cvt.rn.f32.s32 %f150, %r421;
fma.rn.ftz.f32 %f151, %f146, %f150, 0f00000000;
selp.b32 %r422, %r408, %r403, %p13;
cvt.u16.u32 %rs20, %r422;
and.b32 %r423, %r422, 63;
mul.lo.s32 %r424, %r366, %r423;
cvt.rn.f32.s32 %f152, %r424;
fma.rn.ftz.f32 %f153, %f146, %f152, 0f00000000;
shr.u32 %r425, %r389, 4;
and.b32 %r426, %r425, 252645135;
shr.u32 %r427, %r390, 4;
and.b32 %r428, %r427, 252645135;
shl.b32 %r429, %r392, 3;
and.b32 %r430, %r429, 269488144;
shl.b32 %r431, %r394, 3;
and.b32 %r432, %r431, 269488144;
or.b32 %r375, %r426, %r430;
or.b32 %r371, %r428, %r432;

dp4a.s32.s32 %r370, %r371, %r25, %r381;


dp4a.s32.s32 %r374, %r375, %r24, %r370;


dp4a.s32.s32 %r378, %r383, %r25, %r381;


dp4a.s32.s32 %r382, %r383, %r24, %r378;

shr.u16 %rs21, %rs19, 8;
cvt.u32.u16 %r433, %rs21;
mul.lo.s32 %r434, %r374, %r433;
cvt.rn.f32.s32 %f154, %r434;
fma.rn.ftz.f32 %f155, %f147, %f154, %f151;
shr.u16 %rs22, %rs20, 8;
cvt.u32.u16 %r435, %rs22;
mul.lo.s32 %r436, %r382, %r435;
cvt.rn.f32.s32 %f156, %r436;
fma.rn.ftz.f32 %f157, %f147, %f156, %f153;
ld.global.u32 %r387, [%rd112];

{.reg .f16 low,high;
mov.b32 {low,high},%r387;
cvt.f32.f16 %f148, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r387;
cvt.f32.f16 %f149, high;}


mul.ftz.f32 %f158, %f155, %f148;
mul.ftz.f32 %f159, %f157, %f149;
sub.ftz.f32 %f160, %f158, %f159;
add.ftz.f32 %f33, %f160, %f33;

$L__BB170_14:
mul.wide.s32 %rd133, %r48, 176;
cvt.u32.u64 %r1515, %rd5;
add.s64 %rd32, %rd28, %rd133;
ld.global.nc.u32 %r473, [%rd32];
ld.global.nc.u32 %r474, [%rd32+16];
add.s64 %rd33, %rd29, %rd133;
ld.global.nc.u32 %r475, [%rd33];
shr.s32 %r477, %r475, %r1515;
ld.global.nc.u32 %r478, [%rd33+16];
shr.s32 %r479, %r478, %r1515;
add.s64 %rd34, %rd30, %rd133;
ld.global.nc.u16 %rs23, [%rd34];
cvt.u32.u16 %r480, %rs23;
ld.global.nc.u16 %rs24, [%rd34+4];
cvt.u32.u16 %r481, %rs24;
ld.global.nc.u16 %rs25, [%rd34+8];
cvt.u32.u16 %r482, %rs25;
and.b32 %r483, %r480, 16191;
and.b32 %r484, %r482, 3855;
and.b32 %r485, %r480, 49344;
shr.u32 %r486, %r485, 2;
or.b32 %r487, %r484, %r486;
and.b32 %r488, %r481, 16191;
shr.u32 %r489, %r482, 4;
and.b32 %r490, %r489, 3855;
and.b32 %r491, %r481, 49344;
shr.u32 %r492, %r491, 2;
or.b32 %r493, %r490, %r492;

{.reg .f16 low,high;
mov.b32 {low,high},%r20;
cvt.f32.f16 %f161, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r23;
cvt.f32.f16 %f162, low;}


and.b32 %r494, %r473, 252645135;
and.b32 %r495, %r474, 252645135;
shl.b32 %r496, %r477, 4;
and.b32 %r497, %r496, 269488144;
shl.b32 %r498, %r479, 4;
and.b32 %r499, %r498, 269488144;
or.b32 %r444, %r494, %r497;
or.b32 %r440, %r495, %r499;
mov.u32 %r466, 0;

dp4a.s32.s32 %r439, %r440, %r22, %r466;


dp4a.s32.s32 %r443, %r444, %r21, %r439;

mov.u32 %r468, 16843009;

dp4a.s32.s32 %r447, %r468, %r22, %r466;


dp4a.s32.s32 %r451, %r468, %r21, %r447;

selp.b32 %r500, %r487, %r483, %p13;
cvt.u16.u32 %rs26, %r500;
and.b32 %r501, %r500, 63;
mul.lo.s32 %r502, %r443, %r501;
cvt.rn.f32.s32 %f165, %r502;
fma.rn.ftz.f32 %f166, %f161, %f165, 0f00000000;
selp.b32 %r503, %r493, %r488, %p13;
cvt.u16.u32 %rs27, %r503;
and.b32 %r504, %r503, 63;
mul.lo.s32 %r505, %r451, %r504;
cvt.rn.f32.s32 %f167, %r505;
fma.rn.ftz.f32 %f168, %f161, %f167, 0f00000000;
shr.u32 %r506, %r473, 4;
and.b32 %r507, %r506, 252645135;
shr.u32 %r508, %r474, 4;
and.b32 %r509, %r508, 252645135;
shl.b32 %r510, %r477, 3;
and.b32 %r511, %r510, 269488144;
shl.b32 %r512, %r479, 3;
and.b32 %r513, %r512, 269488144;
or.b32 %r460, %r507, %r511;
or.b32 %r456, %r509, %r513;

dp4a.s32.s32 %r455, %r456, %r25, %r466;


dp4a.s32.s32 %r459, %r460, %r24, %r455;


dp4a.s32.s32 %r463, %r468, %r25, %r466;


dp4a.s32.s32 %r467, %r468, %r24, %r463;

shr.u16 %rs28, %rs26, 8;
cvt.u32.u16 %r514, %rs28;
mul.lo.s32 %r515, %r459, %r514;
cvt.rn.f32.s32 %f169, %r515;
fma.rn.ftz.f32 %f170, %f162, %f169, %f166;
shr.u16 %rs29, %rs27, 8;
cvt.u32.u16 %r516, %rs29;
mul.lo.s32 %r517, %r467, %r516;
cvt.rn.f32.s32 %f171, %r517;
fma.rn.ftz.f32 %f172, %f162, %f171, %f168;
add.s64 %rd35, %rd31, %rd133;
ld.global.nc.u32 %r471, [%rd35];

{.reg .f16 low,high;
mov.b32 {low,high},%r471;
cvt.f32.f16 %f163, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r471;
cvt.f32.f16 %f164, high;}


mul.ftz.f32 %f173, %f170, %f163;
mul.ftz.f32 %f174, %f172, %f164;
sub.ftz.f32 %f175, %f173, %f174;
add.ftz.f32 %f28, %f175, %f28;
@%p14 bra $L__BB170_16;

cvt.u32.u64 %r1519, %rd5;
add.s64 %rd114, %rd141, %rd19;
add.s64 %rd115, %rd141, %rd18;
ld.global.u32 %r555, [%rd114];
ld.global.u32 %r556, [%rd114+16];
ld.global.u32 %r557, [%rd115];
shr.s32 %r558, %r557, %r1519;
ld.global.u32 %r559, [%rd115+16];
shr.s32 %r560, %r559, %r1519;
add.s64 %rd116, %rd141, %rd15;
ld.global.u16 %r561, [%rd116+-4];
ld.global.u16 %r562, [%rd116];
ld.global.u16 %r563, [%rd116+4];
and.b32 %r564, %r561, 16191;
and.b32 %r565, %r563, 3855;
and.b32 %r566, %r561, 49344;
shr.u32 %r567, %r566, 2;
or.b32 %r568, %r565, %r567;
and.b32 %r569, %r562, 16191;
shr.u32 %r570, %r563, 4;
and.b32 %r571, %r570, 3855;
and.b32 %r572, %r562, 49344;
shr.u32 %r573, %r572, 2;
or.b32 %r574, %r571, %r573;

{.reg .f16 low,high;
mov.b32 {low,high},%r20;
cvt.f32.f16 %f176, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r23;
cvt.f32.f16 %f177, low;}


or.b32 %r577, %r4, 2;
mad.lo.s32 %r578, %r48, %r577, %r17;
mul.wide.s32 %rd117, %r578, 176;
add.s64 %rd118, %rd141, %rd117;
and.b32 %r579, %r555, 252645135;
and.b32 %r580, %r556, 252645135;
shl.b32 %r581, %r558, 4;
and.b32 %r582, %r581, 269488144;
shl.b32 %r583, %r560, 4;
and.b32 %r584, %r583, 269488144;
or.b32 %r525, %r579, %r582;
or.b32 %r521, %r580, %r584;

dp4a.s32.s32 %r520, %r521, %r22, %r466;


dp4a.s32.s32 %r524, %r525, %r21, %r520;


dp4a.s32.s32 %r528, %r468, %r22, %r466;


dp4a.s32.s32 %r532, %r468, %r21, %r528;

selp.b32 %r585, %r568, %r564, %p13;
cvt.u16.u32 %rs30, %r585;
and.b32 %r586, %r585, 63;
mul.lo.s32 %r587, %r524, %r586;
cvt.rn.f32.s32 %f180, %r587;
fma.rn.ftz.f32 %f181, %f176, %f180, 0f00000000;
selp.b32 %r588, %r574, %r569, %p13;
cvt.u16.u32 %rs31, %r588;
and.b32 %r589, %r588, 63;
mul.lo.s32 %r590, %r532, %r589;
cvt.rn.f32.s32 %f182, %r590;
fma.rn.ftz.f32 %f183, %f176, %f182, 0f00000000;
shr.u32 %r591, %r555, 4;
and.b32 %r592, %r591, 252645135;
shr.u32 %r593, %r556, 4;
and.b32 %r594, %r593, 252645135;
shl.b32 %r595, %r558, 3;
and.b32 %r596, %r595, 269488144;
shl.b32 %r597, %r560, 3;
and.b32 %r598, %r597, 269488144;
or.b32 %r541, %r592, %r596;
or.b32 %r537, %r594, %r598;

dp4a.s32.s32 %r536, %r537, %r25, %r466;


dp4a.s32.s32 %r540, %r541, %r24, %r536;


dp4a.s32.s32 %r544, %r468, %r25, %r466;


dp4a.s32.s32 %r548, %r468, %r24, %r544;

shr.u16 %rs32, %rs30, 8;
cvt.u32.u16 %r599, %rs32;
mul.lo.s32 %r600, %r540, %r599;
cvt.rn.f32.s32 %f184, %r600;
fma.rn.ftz.f32 %f185, %f177, %f184, %f181;
shr.u16 %rs33, %rs31, 8;
cvt.u32.u16 %r601, %rs33;
mul.lo.s32 %r602, %r548, %r601;
cvt.rn.f32.s32 %f186, %r602;
fma.rn.ftz.f32 %f187, %f177, %f186, %f183;
ld.global.u32 %r553, [%rd118];

{.reg .f16 low,high;
mov.b32 {low,high},%r553;
cvt.f32.f16 %f178, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r553;
cvt.f32.f16 %f179, high;}


mul.ftz.f32 %f188, %f185, %f178;
mul.ftz.f32 %f189, %f187, %f179;
sub.ftz.f32 %f190, %f188, %f189;
add.ftz.f32 %f32, %f190, %f32;

$L__BB170_16:
mul.wide.s32 %rd131, %r48, 176;
cvt.u32.u64 %r1502, %rd5;
add.s64 %rd120, %rd32, %rd131;
ld.global.nc.u32 %r639, [%rd120];
ld.global.nc.u32 %r640, [%rd120+16];
add.s64 %rd121, %rd33, %rd131;
ld.global.nc.u32 %r641, [%rd121];
shr.s32 %r643, %r641, %r1502;
ld.global.nc.u32 %r644, [%rd121+16];
shr.s32 %r645, %r644, %r1502;
add.s64 %rd122, %rd34, %rd131;
ld.global.nc.u16 %rs34, [%rd122];
cvt.u32.u16 %r646, %rs34;
ld.global.nc.u16 %rs35, [%rd122+4];
cvt.u32.u16 %r647, %rs35;
ld.global.nc.u16 %rs36, [%rd122+8];
cvt.u32.u16 %r648, %rs36;
and.b32 %r649, %r646, 16191;
and.b32 %r650, %r648, 3855;
and.b32 %r651, %r646, 49344;
shr.u32 %r652, %r651, 2;
or.b32 %r653, %r650, %r652;
and.b32 %r654, %r647, 16191;
shr.u32 %r655, %r648, 4;
and.b32 %r656, %r655, 3855;
and.b32 %r657, %r647, 49344;
shr.u32 %r658, %r657, 2;
or.b32 %r659, %r656, %r658;

{.reg .f16 low,high;
mov.b32 {low,high},%r20;
cvt.f32.f16 %f191, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r23;
cvt.f32.f16 %f192, low;}


and.b32 %r660, %r639, 252645135;
and.b32 %r661, %r640, 252645135;
shl.b32 %r662, %r643, 4;
and.b32 %r663, %r662, 269488144;
shl.b32 %r664, %r645, 4;
and.b32 %r665, %r664, 269488144;
or.b32 %r610, %r660, %r663;
or.b32 %r606, %r661, %r665;

dp4a.s32.s32 %r605, %r606, %r22, %r466;


dp4a.s32.s32 %r609, %r610, %r21, %r605;


dp4a.s32.s32 %r613, %r468, %r22, %r466;


dp4a.s32.s32 %r617, %r468, %r21, %r613;

selp.b32 %r666, %r653, %r649, %p13;
cvt.u16.u32 %rs37, %r666;
and.b32 %r667, %r666, 63;
mul.lo.s32 %r668, %r609, %r667;
cvt.rn.f32.s32 %f195, %r668;
fma.rn.ftz.f32 %f196, %f191, %f195, 0f00000000;
selp.b32 %r669, %r659, %r654, %p13;
cvt.u16.u32 %rs38, %r669;
and.b32 %r670, %r669, 63;
mul.lo.s32 %r671, %r617, %r670;
cvt.rn.f32.s32 %f197, %r671;
fma.rn.ftz.f32 %f198, %f191, %f197, 0f00000000;
shr.u32 %r672, %r639, 4;
and.b32 %r673, %r672, 252645135;
shr.u32 %r674, %r640, 4;
and.b32 %r675, %r674, 252645135;
shl.b32 %r676, %r643, 3;
and.b32 %r677, %r676, 269488144;
shl.b32 %r678, %r645, 3;
and.b32 %r679, %r678, 269488144;
or.b32 %r626, %r673, %r677;
or.b32 %r622, %r675, %r679;

dp4a.s32.s32 %r621, %r622, %r25, %r466;


dp4a.s32.s32 %r625, %r626, %r24, %r621;


dp4a.s32.s32 %r629, %r468, %r25, %r466;


dp4a.s32.s32 %r633, %r468, %r24, %r629;

shr.u16 %rs39, %rs37, 8;
cvt.u32.u16 %r680, %rs39;
mul.lo.s32 %r681, %r625, %r680;
cvt.rn.f32.s32 %f199, %r681;
fma.rn.ftz.f32 %f200, %f192, %f199, %f196;
shr.u16 %rs40, %rs38, 8;
cvt.u32.u16 %r682, %rs40;
mul.lo.s32 %r683, %r633, %r682;
cvt.rn.f32.s32 %f201, %r683;
fma.rn.ftz.f32 %f202, %f192, %f201, %f198;
add.s64 %rd123, %rd35, %rd131;
ld.global.nc.u32 %r637, [%rd123];

{.reg .f16 low,high;
mov.b32 {low,high},%r637;
cvt.f32.f16 %f193, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r637;
cvt.f32.f16 %f194, high;}


mul.ftz.f32 %f203, %f200, %f193;
mul.ftz.f32 %f204, %f202, %f194;
sub.ftz.f32 %f205, %f203, %f204;
add.ftz.f32 %f27, %f205, %f27;
@%p14 bra $L__BB170_18;

cvt.u32.u64 %r1518, %rd5;
add.s64 %rd124, %rd141, %rd14;
ld.global.u32 %r721, [%rd124+-16];
ld.global.u32 %r722, [%rd124];
add.s64 %rd125, %rd141, %rd16;
ld.global.u32 %r723, [%rd125+-16];
shr.s32 %r724, %r723, %r1518;
ld.global.u32 %r725, [%rd125];
shr.s32 %r726, %r725, %r1518;
add.s64 %rd126, %rd141, %rd17;
ld.global.u16 %r727, [%rd126+-8];
ld.global.u16 %r728, [%rd126+-4];
ld.global.u16 %r729, [%rd126];
and.b32 %r730, %r727, 16191;
and.b32 %r731, %r729, 3855;
and.b32 %r732, %r727, 49344;
shr.u32 %r733, %r732, 2;
or.b32 %r734, %r731, %r733;
and.b32 %r735, %r728, 16191;
shr.u32 %r736, %r729, 4;
and.b32 %r737, %r736, 3855;
and.b32 %r738, %r728, 49344;
shr.u32 %r739, %r738, 2;
or.b32 %r740, %r737, %r739;

{.reg .f16 low,high;
mov.b32 {low,high},%r20;
cvt.f32.f16 %f206, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r23;
cvt.f32.f16 %f207, low;}


or.b32 %r743, %r4, 3;
mad.lo.s32 %r744, %r48, %r743, %r17;
mul.wide.s32 %rd127, %r744, 176;
add.s64 %rd128, %rd141, %rd127;
and.b32 %r745, %r721, 252645135;
and.b32 %r746, %r722, 252645135;
shl.b32 %r747, %r724, 4;
and.b32 %r748, %r747, 269488144;
shl.b32 %r749, %r726, 4;
and.b32 %r750, %r749, 269488144;
or.b32 %r691, %r745, %r748;
or.b32 %r687, %r746, %r750;
mov.u32 %r713, 0;

dp4a.s32.s32 %r686, %r687, %r22, %r713;


dp4a.s32.s32 %r690, %r691, %r21, %r686;

mov.u32 %r715, 16843009;

dp4a.s32.s32 %r694, %r715, %r22, %r713;


dp4a.s32.s32 %r698, %r715, %r21, %r694;

selp.b32 %r751, %r734, %r730, %p13;
cvt.u16.u32 %rs41, %r751;
and.b32 %r752, %r751, 63;
mul.lo.s32 %r753, %r690, %r752;
cvt.rn.f32.s32 %f210, %r753;
fma.rn.ftz.f32 %f211, %f206, %f210, 0f00000000;
selp.b32 %r754, %r740, %r735, %p13;
cvt.u16.u32 %rs42, %r754;
and.b32 %r755, %r754, 63;
mul.lo.s32 %r756, %r698, %r755;
cvt.rn.f32.s32 %f212, %r756;
fma.rn.ftz.f32 %f213, %f206, %f212, 0f00000000;
shr.u32 %r757, %r721, 4;
and.b32 %r758, %r757, 252645135;
shr.u32 %r759, %r722, 4;
and.b32 %r760, %r759, 252645135;
shl.b32 %r761, %r724, 3;
and.b32 %r762, %r761, 269488144;
shl.b32 %r763, %r726, 3;
and.b32 %r764, %r763, 269488144;
or.b32 %r707, %r758, %r762;
or.b32 %r703, %r760, %r764;

dp4a.s32.s32 %r702, %r703, %r25, %r713;


dp4a.s32.s32 %r706, %r707, %r24, %r702;


dp4a.s32.s32 %r710, %r715, %r25, %r713;


dp4a.s32.s32 %r714, %r715, %r24, %r710;

shr.u16 %rs43, %rs41, 8;
cvt.u32.u16 %r765, %rs43;
mul.lo.s32 %r766, %r706, %r765;
cvt.rn.f32.s32 %f214, %r766;
fma.rn.ftz.f32 %f215, %f207, %f214, %f211;
shr.u16 %rs44, %rs42, 8;
cvt.u32.u16 %r767, %rs44;
mul.lo.s32 %r768, %r714, %r767;
cvt.rn.f32.s32 %f216, %r768;
fma.rn.ftz.f32 %f217, %f207, %f216, %f213;
ld.global.u32 %r719, [%rd128];

{.reg .f16 low,high;
mov.b32 {low,high},%r719;
cvt.f32.f16 %f208, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r719;
cvt.f32.f16 %f209, high;}


mul.ftz.f32 %f218, %f215, %f208;
mul.ftz.f32 %f219, %f217, %f209;
sub.ftz.f32 %f220, %f218, %f219;
add.ftz.f32 %f31, %f220, %f31;

$L__BB170_18:
add.s32 %r1522, %r1522, 64;
add.s64 %rd141, %rd141, 1408;
add.s64 %rd140, %rd140, 1408;
add.s32 %r1523, %r1523, 8;
setp.lt.s32 %p25, %r1523, %r82;
@%p25 bra $L__BB170_10;

$L__BB170_19:
mov.u32 %r1503, %tid.y;
setp.eq.s32 %p26, %r1503, 0;
@%p26 bra $L__BB170_23;

mov.u32 %r1504, %tid.y;
setp.ne.s64 %p27, %rd43, 0;
add.s32 %r28, %r1504, -1;
shl.b32 %r773, %r1504, 9;
mov.u32 %r774, _ZZ13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r775, %r774, %r773;
shl.b32 %r776, %r2, 2;
add.s32 %r777, %r775, %r776;
st.shared.f32 [%r777+-512], %f30;
@%p27 bra $L__BB170_22;
bra.uni $L__BB170_21;

$L__BB170_22:
shl.b32 %r779, %r28, 9;
mov.u32 %r780, _ZZ13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r781, %r780, %r779;
add.s32 %r783, %r781, %r776;
st.shared.f32 [%r783], %f34;
add.s32 %r1499, %r777, -512;
st.shared.f32 [%r1499+128], %f29;
st.shared.f32 [%r783+128], %f33;
add.s32 %r1500, %r777, -512;
st.shared.f32 [%r1500+256], %f28;
st.shared.f32 [%r783+256], %f32;
add.s32 %r1501, %r777, -512;
st.shared.f32 [%r1501+384], %f27;
st.shared.f32 [%r783+384], %f31;
bra.uni $L__BB170_23;

$L__BB170_21:
add.s32 %r1496, %r777, -512;
st.shared.f32 [%r1496+128], %f29;
add.s32 %r1497, %r777, -512;
st.shared.f32 [%r1497+256], %f28;
add.s32 %r1498, %r777, -512;
st.shared.f32 [%r1498+384], %f27;

$L__BB170_23:
mov.u32 %r1505, %tid.y;
setp.ne.s32 %p327, %r1505, 0;
bar.sync 0;
@%p327 bra $L__BB170_78;
bra.uni $L__BB170_24;

$L__BB170_78:
ret;

$L__BB170_24:
ld.param.u64 %rd132, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.f32 %f1065, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
ld.param.u32 %r1508, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r1507, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
mov.u32 %r1506, %ctaid.z;
setp.ne.s64 %p29, %rd43, 0;
mov.u32 %r786, %ctaid.y;
mad.lo.s32 %r789, %r786, %r1508, %r4;
mad.lo.s32 %r790, %r1506, %r1507, %r789;
neg.ftz.f32 %f35, %f1065;
shl.b32 %r792, %r2, 2;
mov.u32 %r793, _ZZ13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r30, %r793, %r792;
cvta.to.global.u64 %rd129, %rd132;
mul.wide.u32 %rd130, %r790, 4;
add.s64 %rd38, %rd129, %rd130;
@%p29 bra $L__BB170_33;
bra.uni $L__BB170_25;

$L__BB170_33:
ld.param.u32 %r1509, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
mov.u32 %r880, _ZZ13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r32, %r880, %r792;
setp.gt.s32 %p62, %r1509, 2;
@%p62 bra $L__BB170_44;

ld.param.u32 %r1512, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p65, %r1512, 1;
@%p65 bra $L__BB170_62;

ld.param.u32 %r1513, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p66, %r1513, 2;
@%p66 bra $L__BB170_36;
bra.uni $L__BB170_70;

$L__BB170_36:
ld.shared.f32 %f621, [%r30];
add.ftz.f32 %f622, %f621, %f30;
ld.shared.f32 %f623, [%r32];
add.ftz.f32 %f624, %f623, %f34;
ld.shared.f32 %f625, [%r30+512];
add.ftz.f32 %f626, %f625, %f622;
ld.shared.f32 %f627, [%r32+512];
add.ftz.f32 %f628, %f627, %f624;
ld.shared.f32 %f629, [%r30+1024];
add.ftz.f32 %f630, %f629, %f626;
ld.shared.f32 %f631, [%r32+1024];
add.ftz.f32 %f632, %f631, %f628;
mov.b32 %r1127, %f630;
mov.u32 %r1128, 31;
mov.u32 %r1129, 16;
mov.u32 %r1130, -1;
shfl.sync.bfly.b32 %r1131|%p171, %r1127, %r1129, %r1128, %r1130;
mov.b32 %f633, %r1131;
add.ftz.f32 %f634, %f630, %f633;
mov.b32 %r1132, %f634;
mov.u32 %r1133, 8;
shfl.sync.bfly.b32 %r1134|%p172, %r1132, %r1133, %r1128, %r1130;
mov.b32 %f635, %r1134;
add.ftz.f32 %f636, %f634, %f635;
mov.b32 %r1135, %f636;
mov.u32 %r1136, 4;
shfl.sync.bfly.b32 %r1137|%p173, %r1135, %r1136, %r1128, %r1130;
mov.b32 %f637, %r1137;
add.ftz.f32 %f638, %f636, %f637;
mov.b32 %r1138, %f638;
mov.u32 %r1139, 2;
shfl.sync.bfly.b32 %r1140|%p174, %r1138, %r1139, %r1128, %r1130;
mov.b32 %f639, %r1140;
add.ftz.f32 %f640, %f638, %f639;
mov.b32 %r1141, %f640;
mov.u32 %r1142, 1;
shfl.sync.bfly.b32 %r1143|%p175, %r1141, %r1142, %r1128, %r1130;
mov.b32 %f641, %r1143;
add.ftz.f32 %f56, %f640, %f641;
mov.b32 %r1144, %f632;
shfl.sync.bfly.b32 %r1145|%p176, %r1144, %r1129, %r1128, %r1130;
mov.b32 %f642, %r1145;
add.ftz.f32 %f643, %f632, %f642;
mov.b32 %r1146, %f643;
shfl.sync.bfly.b32 %r1147|%p177, %r1146, %r1133, %r1128, %r1130;
mov.b32 %f644, %r1147;
add.ftz.f32 %f645, %f643, %f644;
mov.b32 %r1148, %f645;
shfl.sync.bfly.b32 %r1149|%p178, %r1148, %r1136, %r1128, %r1130;
mov.b32 %f646, %r1149;
add.ftz.f32 %f647, %f645, %f646;
mov.b32 %r1150, %f647;
shfl.sync.bfly.b32 %r1151|%p179, %r1150, %r1139, %r1128, %r1130;
mov.b32 %f648, %r1151;
add.ftz.f32 %f649, %f647, %f648;
mov.b32 %r1152, %f649;
shfl.sync.bfly.b32 %r1153|%p180, %r1152, %r1142, %r1128, %r1130;
mov.b32 %f650, %r1153;
add.ftz.f32 %f57, %f649, %f650;
setp.ne.s32 %p181, %r2, 0;
setp.ge.u32 %p182, %r4, %r49;
or.pred %p183, %p181, %p182;
@%p183 bra $L__BB170_38;

add.ftz.f32 %f651, %f4, %f56;
add.ftz.f32 %f652, %f5, %f57;
mul.ftz.f32 %f653, %f652, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f654, %f653;
add.ftz.f32 %f655, %f654, 0f3F800000;
div.approx.ftz.f32 %f656, %f652, %f655;
mul.ftz.f32 %f657, %f651, %f656;
st.global.f32 [%rd38], %f657;

$L__BB170_38:
ld.shared.f32 %f658, [%r30+128];
add.ftz.f32 %f659, %f658, %f29;
ld.shared.f32 %f660, [%r32+128];
add.ftz.f32 %f661, %f660, %f33;
ld.shared.f32 %f662, [%r30+640];
add.ftz.f32 %f663, %f662, %f659;
ld.shared.f32 %f664, [%r32+640];
add.ftz.f32 %f665, %f664, %f661;
ld.shared.f32 %f666, [%r30+1152];
add.ftz.f32 %f667, %f666, %f663;
ld.shared.f32 %f668, [%r32+1152];
add.ftz.f32 %f669, %f668, %f665;
mov.b32 %r1157, %f667;
shfl.sync.bfly.b32 %r1161|%p184, %r1157, %r1129, %r1128, %r1130;
mov.b32 %f670, %r1161;
add.ftz.f32 %f671, %f667, %f670;
mov.b32 %r1162, %f671;
shfl.sync.bfly.b32 %r1164|%p185, %r1162, %r1133, %r1128, %r1130;
mov.b32 %f672, %r1164;
add.ftz.f32 %f673, %f671, %f672;
mov.b32 %r1165, %f673;
shfl.sync.bfly.b32 %r1167|%p186, %r1165, %r1136, %r1128, %r1130;
mov.b32 %f674, %r1167;
add.ftz.f32 %f675, %f673, %f674;
mov.b32 %r1168, %f675;
shfl.sync.bfly.b32 %r1170|%p187, %r1168, %r1139, %r1128, %r1130;
mov.b32 %f676, %r1170;
add.ftz.f32 %f677, %f675, %f676;
mov.b32 %r1171, %f677;
shfl.sync.bfly.b32 %r1173|%p188, %r1171, %r1142, %r1128, %r1130;
mov.b32 %f678, %r1173;
add.ftz.f32 %f58, %f677, %f678;
mov.b32 %r1174, %f669;
shfl.sync.bfly.b32 %r1175|%p189, %r1174, %r1129, %r1128, %r1130;
mov.b32 %f679, %r1175;
add.ftz.f32 %f680, %f669, %f679;
mov.b32 %r1176, %f680;
shfl.sync.bfly.b32 %r1177|%p190, %r1176, %r1133, %r1128, %r1130;
mov.b32 %f681, %r1177;
add.ftz.f32 %f682, %f680, %f681;
mov.b32 %r1178, %f682;
shfl.sync.bfly.b32 %r1179|%p191, %r1178, %r1136, %r1128, %r1130;
mov.b32 %f683, %r1179;
add.ftz.f32 %f684, %f682, %f683;
mov.b32 %r1180, %f684;
shfl.sync.bfly.b32 %r1181|%p192, %r1180, %r1139, %r1128, %r1130;
mov.b32 %f685, %r1181;
add.ftz.f32 %f686, %f684, %f685;
mov.b32 %r1182, %f686;
shfl.sync.bfly.b32 %r1183|%p193, %r1182, %r1142, %r1128, %r1130;
mov.b32 %f687, %r1183;
add.ftz.f32 %f59, %f686, %f687;
setp.ne.s32 %p194, %r2, 1;
or.b32 %r1187, %r4, 1;
setp.ge.u32 %p195, %r1187, %r49;
or.pred %p196, %p194, %p195;
@%p196 bra $L__BB170_40;

add.ftz.f32 %f688, %f4, %f58;
add.ftz.f32 %f689, %f5, %f59;
mul.ftz.f32 %f690, %f689, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f691, %f690;
add.ftz.f32 %f692, %f691, 0f3F800000;
div.approx.ftz.f32 %f693, %f689, %f692;
mul.ftz.f32 %f694, %f688, %f693;
st.global.f32 [%rd38+4], %f694;

$L__BB170_40:
ld.shared.f32 %f695, [%r30+256];
add.ftz.f32 %f696, %f695, %f28;
ld.shared.f32 %f697, [%r32+256];
add.ftz.f32 %f698, %f697, %f32;
ld.shared.f32 %f699, [%r30+768];
add.ftz.f32 %f700, %f699, %f696;
ld.shared.f32 %f701, [%r32+768];
add.ftz.f32 %f702, %f701, %f698;
ld.shared.f32 %f703, [%r30+1280];
add.ftz.f32 %f704, %f703, %f700;
ld.shared.f32 %f705, [%r32+1280];
add.ftz.f32 %f706, %f705, %f702;
mov.b32 %r1188, %f704;
mov.u32 %r1189, 31;
mov.u32 %r1190, 16;
mov.u32 %r1191, -1;
shfl.sync.bfly.b32 %r1192|%p197, %r1188, %r1190, %r1189, %r1191;
mov.b32 %f707, %r1192;
add.ftz.f32 %f708, %f704, %f707;
mov.b32 %r1193, %f708;
mov.u32 %r1194, 8;
shfl.sync.bfly.b32 %r1195|%p198, %r1193, %r1194, %r1189, %r1191;
mov.b32 %f709, %r1195;
add.ftz.f32 %f710, %f708, %f709;
mov.b32 %r1196, %f710;
mov.u32 %r1197, 4;
shfl.sync.bfly.b32 %r1198|%p199, %r1196, %r1197, %r1189, %r1191;
mov.b32 %f711, %r1198;
add.ftz.f32 %f712, %f710, %f711;
mov.b32 %r1199, %f712;
mov.u32 %r1200, 2;
shfl.sync.bfly.b32 %r1201|%p200, %r1199, %r1200, %r1189, %r1191;
mov.b32 %f713, %r1201;
add.ftz.f32 %f714, %f712, %f713;
mov.b32 %r1202, %f714;
mov.u32 %r1203, 1;
shfl.sync.bfly.b32 %r1204|%p201, %r1202, %r1203, %r1189, %r1191;
mov.b32 %f715, %r1204;
add.ftz.f32 %f60, %f714, %f715;
mov.b32 %r1205, %f706;
shfl.sync.bfly.b32 %r1206|%p202, %r1205, %r1190, %r1189, %r1191;
mov.b32 %f716, %r1206;
add.ftz.f32 %f717, %f706, %f716;
mov.b32 %r1207, %f717;
shfl.sync.bfly.b32 %r1208|%p203, %r1207, %r1194, %r1189, %r1191;
mov.b32 %f718, %r1208;
add.ftz.f32 %f719, %f717, %f718;
mov.b32 %r1209, %f719;
shfl.sync.bfly.b32 %r1210|%p204, %r1209, %r1197, %r1189, %r1191;
mov.b32 %f720, %r1210;
add.ftz.f32 %f721, %f719, %f720;
mov.b32 %r1211, %f721;
shfl.sync.bfly.b32 %r1212|%p205, %r1211, %r1200, %r1189, %r1191;
mov.b32 %f722, %r1212;
add.ftz.f32 %f723, %f721, %f722;
mov.b32 %r1213, %f723;
shfl.sync.bfly.b32 %r1214|%p206, %r1213, %r1203, %r1189, %r1191;
mov.b32 %f724, %r1214;
add.ftz.f32 %f61, %f723, %f724;
setp.ne.s32 %p207, %r2, 2;
or.b32 %r1218, %r4, 2;
setp.ge.u32 %p208, %r1218, %r49;
or.pred %p209, %p207, %p208;
@%p209 bra $L__BB170_42;

add.ftz.f32 %f725, %f4, %f60;
add.ftz.f32 %f726, %f5, %f61;
mul.ftz.f32 %f727, %f726, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f728, %f727;
add.ftz.f32 %f729, %f728, 0f3F800000;
div.approx.ftz.f32 %f730, %f726, %f729;
mul.ftz.f32 %f731, %f725, %f730;
st.global.f32 [%rd38+8], %f731;

$L__BB170_42:
ld.shared.f32 %f732, [%r30+384];
add.ftz.f32 %f733, %f732, %f27;
ld.shared.f32 %f734, [%r32+384];
add.ftz.f32 %f735, %f734, %f31;
ld.shared.f32 %f736, [%r30+896];
add.ftz.f32 %f737, %f736, %f733;
ld.shared.f32 %f738, [%r32+896];
add.ftz.f32 %f739, %f738, %f735;
ld.shared.f32 %f740, [%r30+1408];
add.ftz.f32 %f741, %f740, %f737;
ld.shared.f32 %f742, [%r32+1408];
add.ftz.f32 %f743, %f742, %f739;
mov.b32 %r1219, %f741;
shfl.sync.bfly.b32 %r1223|%p210, %r1219, %r1190, %r1189, %r1191;
mov.b32 %f744, %r1223;
add.ftz.f32 %f745, %f741, %f744;
mov.b32 %r1224, %f745;
shfl.sync.bfly.b32 %r1226|%p211, %r1224, %r1194, %r1189, %r1191;
mov.b32 %f746, %r1226;
add.ftz.f32 %f747, %f745, %f746;
mov.b32 %r1227, %f747;
shfl.sync.bfly.b32 %r1229|%p212, %r1227, %r1197, %r1189, %r1191;
mov.b32 %f748, %r1229;
add.ftz.f32 %f749, %f747, %f748;
mov.b32 %r1230, %f749;
shfl.sync.bfly.b32 %r1232|%p213, %r1230, %r1200, %r1189, %r1191;
mov.b32 %f750, %r1232;
add.ftz.f32 %f751, %f749, %f750;
mov.b32 %r1233, %f751;
shfl.sync.bfly.b32 %r1235|%p214, %r1233, %r1203, %r1189, %r1191;
mov.b32 %f752, %r1235;
add.ftz.f32 %f62, %f751, %f752;
mov.b32 %r1236, %f743;
shfl.sync.bfly.b32 %r1237|%p215, %r1236, %r1190, %r1189, %r1191;
mov.b32 %f753, %r1237;
add.ftz.f32 %f754, %f743, %f753;
mov.b32 %r1238, %f754;
shfl.sync.bfly.b32 %r1239|%p216, %r1238, %r1194, %r1189, %r1191;
mov.b32 %f755, %r1239;
add.ftz.f32 %f756, %f754, %f755;
mov.b32 %r1240, %f756;
shfl.sync.bfly.b32 %r1241|%p217, %r1240, %r1197, %r1189, %r1191;
mov.b32 %f757, %r1241;
add.ftz.f32 %f758, %f756, %f757;
mov.b32 %r1242, %f758;
shfl.sync.bfly.b32 %r1243|%p218, %r1242, %r1200, %r1189, %r1191;
mov.b32 %f759, %r1243;
add.ftz.f32 %f760, %f758, %f759;
mov.b32 %r1244, %f760;
shfl.sync.bfly.b32 %r1245|%p219, %r1244, %r1203, %r1189, %r1191;
mov.b32 %f761, %r1245;
add.ftz.f32 %f63, %f760, %f761;
setp.ne.s32 %p220, %r2, 3;
or.b32 %r1249, %r4, 3;
setp.ge.u32 %p221, %r1249, %r49;
or.pred %p222, %p220, %p221;
@%p222 bra $L__BB170_78;

add.ftz.f32 %f762, %f4, %f62;
add.ftz.f32 %f763, %f5, %f63;
mul.ftz.f32 %f764, %f763, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f765, %f764;
add.ftz.f32 %f766, %f765, 0f3F800000;
div.approx.ftz.f32 %f767, %f763, %f766;
mul.ftz.f32 %f768, %f762, %f767;
st.global.f32 [%rd38+12], %f768;
bra.uni $L__BB170_78;

$L__BB170_25:
ld.shared.f32 %f221, [%r30];
add.ftz.f32 %f222, %f221, %f30;
ld.shared.f32 %f223, [%r30+512];
add.ftz.f32 %f224, %f223, %f222;
ld.shared.f32 %f225, [%r30+1024];
add.ftz.f32 %f226, %f225, %f224;
mov.b32 %r794, %f226;
mov.u32 %r795, 31;
mov.u32 %r796, 16;
mov.u32 %r797, -1;
shfl.sync.bfly.b32 %r798|%p30, %r794, %r796, %r795, %r797;
mov.b32 %f227, %r798;
add.ftz.f32 %f228, %f226, %f227;
mov.b32 %r799, %f228;
mov.u32 %r800, 8;
shfl.sync.bfly.b32 %r801|%p31, %r799, %r800, %r795, %r797;
mov.b32 %f229, %r801;
add.ftz.f32 %f230, %f228, %f229;
mov.b32 %r802, %f230;
mov.u32 %r803, 4;
shfl.sync.bfly.b32 %r804|%p32, %r802, %r803, %r795, %r797;
mov.b32 %f231, %r804;
add.ftz.f32 %f232, %f230, %f231;
mov.b32 %r805, %f232;
mov.u32 %r806, 2;
shfl.sync.bfly.b32 %r807|%p33, %r805, %r806, %r795, %r797;
mov.b32 %f233, %r807;
add.ftz.f32 %f234, %f232, %f233;
mov.b32 %r808, %f234;
mov.u32 %r809, 1;
shfl.sync.bfly.b32 %r810|%p34, %r808, %r809, %r795, %r797;
mov.b32 %f235, %r810;
add.ftz.f32 %f36, %f234, %f235;
setp.ne.s32 %p35, %r2, 0;
setp.ge.u32 %p36, %r4, %r49;
or.pred %p37, %p35, %p36;
@%p37 bra $L__BB170_27;

add.ftz.f32 %f236, %f4, %f36;
st.global.f32 [%rd38], %f236;

$L__BB170_27:
ld.shared.f32 %f237, [%r30+128];
add.ftz.f32 %f238, %f237, %f29;
ld.shared.f32 %f239, [%r30+640];
add.ftz.f32 %f240, %f239, %f238;
ld.shared.f32 %f241, [%r30+1152];
add.ftz.f32 %f242, %f241, %f240;
mov.b32 %r814, %f242;
shfl.sync.bfly.b32 %r818|%p38, %r814, %r796, %r795, %r797;
mov.b32 %f243, %r818;
add.ftz.f32 %f244, %f242, %f243;
mov.b32 %r819, %f244;
shfl.sync.bfly.b32 %r821|%p39, %r819, %r800, %r795, %r797;
mov.b32 %f245, %r821;
add.ftz.f32 %f246, %f244, %f245;
mov.b32 %r822, %f246;
shfl.sync.bfly.b32 %r824|%p40, %r822, %r803, %r795, %r797;
mov.b32 %f247, %r824;
add.ftz.f32 %f248, %f246, %f247;
mov.b32 %r825, %f248;
shfl.sync.bfly.b32 %r827|%p41, %r825, %r806, %r795, %r797;
mov.b32 %f249, %r827;
add.ftz.f32 %f250, %f248, %f249;
mov.b32 %r828, %f250;
shfl.sync.bfly.b32 %r830|%p42, %r828, %r809, %r795, %r797;
mov.b32 %f251, %r830;
add.ftz.f32 %f37, %f250, %f251;
setp.ne.s32 %p43, %r2, 1;
or.b32 %r834, %r4, 1;
setp.ge.u32 %p44, %r834, %r49;
or.pred %p45, %p43, %p44;
@%p45 bra $L__BB170_29;

add.ftz.f32 %f252, %f4, %f37;
st.global.f32 [%rd38+4], %f252;

$L__BB170_29:
ld.shared.f32 %f253, [%r30+256];
add.ftz.f32 %f254, %f253, %f28;
ld.shared.f32 %f255, [%r30+768];
add.ftz.f32 %f256, %f255, %f254;
ld.shared.f32 %f257, [%r30+1280];
add.ftz.f32 %f258, %f257, %f256;
mov.b32 %r835, %f258;
mov.u32 %r836, 31;
mov.u32 %r837, 16;
mov.u32 %r838, -1;
shfl.sync.bfly.b32 %r839|%p46, %r835, %r837, %r836, %r838;
mov.b32 %f259, %r839;
add.ftz.f32 %f260, %f258, %f259;
mov.b32 %r840, %f260;
mov.u32 %r841, 8;
shfl.sync.bfly.b32 %r842|%p47, %r840, %r841, %r836, %r838;
mov.b32 %f261, %r842;
add.ftz.f32 %f262, %f260, %f261;
mov.b32 %r843, %f262;
mov.u32 %r844, 4;
shfl.sync.bfly.b32 %r845|%p48, %r843, %r844, %r836, %r838;
mov.b32 %f263, %r845;
add.ftz.f32 %f264, %f262, %f263;
mov.b32 %r846, %f264;
mov.u32 %r847, 2;
shfl.sync.bfly.b32 %r848|%p49, %r846, %r847, %r836, %r838;
mov.b32 %f265, %r848;
add.ftz.f32 %f266, %f264, %f265;
mov.b32 %r849, %f266;
mov.u32 %r850, 1;
shfl.sync.bfly.b32 %r851|%p50, %r849, %r850, %r836, %r838;
mov.b32 %f267, %r851;
add.ftz.f32 %f38, %f266, %f267;
setp.ne.s32 %p51, %r2, 2;
or.b32 %r855, %r4, 2;
setp.ge.u32 %p52, %r855, %r49;
or.pred %p53, %p51, %p52;
@%p53 bra $L__BB170_31;

add.ftz.f32 %f268, %f4, %f38;
st.global.f32 [%rd38+8], %f268;

$L__BB170_31:
ld.shared.f32 %f269, [%r30+384];
add.ftz.f32 %f270, %f269, %f27;
ld.shared.f32 %f271, [%r30+896];
add.ftz.f32 %f272, %f271, %f270;
ld.shared.f32 %f273, [%r30+1408];
add.ftz.f32 %f274, %f273, %f272;
mov.b32 %r856, %f274;
shfl.sync.bfly.b32 %r860|%p54, %r856, %r837, %r836, %r838;
mov.b32 %f275, %r860;
add.ftz.f32 %f276, %f274, %f275;
mov.b32 %r861, %f276;
shfl.sync.bfly.b32 %r863|%p55, %r861, %r841, %r836, %r838;
mov.b32 %f277, %r863;
add.ftz.f32 %f278, %f276, %f277;
mov.b32 %r864, %f278;
shfl.sync.bfly.b32 %r866|%p56, %r864, %r844, %r836, %r838;
mov.b32 %f279, %r866;
add.ftz.f32 %f280, %f278, %f279;
mov.b32 %r867, %f280;
shfl.sync.bfly.b32 %r869|%p57, %r867, %r847, %r836, %r838;
mov.b32 %f281, %r869;
add.ftz.f32 %f282, %f280, %f281;
mov.b32 %r870, %f282;
shfl.sync.bfly.b32 %r872|%p58, %r870, %r850, %r836, %r838;
mov.b32 %f283, %r872;
add.ftz.f32 %f39, %f282, %f283;
setp.ne.s32 %p59, %r2, 3;
or.b32 %r876, %r4, 3;
setp.ge.u32 %p60, %r876, %r49;
or.pred %p61, %p59, %p60;
@%p61 bra $L__BB170_78;

add.ftz.f32 %f284, %f4, %f39;
st.global.f32 [%rd38+12], %f284;
bra.uni $L__BB170_78;

$L__BB170_44:
ld.param.u32 %r1510, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p63, %r1510, 3;
@%p63 bra $L__BB170_54;

ld.param.u32 %r1511, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.ne.s32 %p64, %r1511, 6;
@%p64 bra $L__BB170_70;

ld.shared.f32 %f285, [%r30];
add.ftz.f32 %f286, %f285, %f30;
ld.shared.f32 %f287, [%r32];
add.ftz.f32 %f288, %f287, %f34;
ld.shared.f32 %f289, [%r30+512];
add.ftz.f32 %f290, %f289, %f286;
ld.shared.f32 %f291, [%r32+512];
add.ftz.f32 %f292, %f291, %f288;
ld.shared.f32 %f293, [%r30+1024];
add.ftz.f32 %f294, %f293, %f290;
ld.shared.f32 %f295, [%r32+1024];
add.ftz.f32 %f296, %f295, %f292;
mov.b32 %r881, %f294;
mov.u32 %r882, 31;
mov.u32 %r883, 16;
mov.u32 %r884, -1;
shfl.sync.bfly.b32 %r885|%p67, %r881, %r883, %r882, %r884;
mov.b32 %f297, %r885;
add.ftz.f32 %f298, %f294, %f297;
mov.b32 %r886, %f298;
mov.u32 %r887, 8;
shfl.sync.bfly.b32 %r888|%p68, %r886, %r887, %r882, %r884;
mov.b32 %f299, %r888;
add.ftz.f32 %f300, %f298, %f299;
mov.b32 %r889, %f300;
mov.u32 %r890, 4;
shfl.sync.bfly.b32 %r891|%p69, %r889, %r890, %r882, %r884;
mov.b32 %f301, %r891;
add.ftz.f32 %f302, %f300, %f301;
mov.b32 %r892, %f302;
mov.u32 %r893, 2;
shfl.sync.bfly.b32 %r894|%p70, %r892, %r893, %r882, %r884;
mov.b32 %f303, %r894;
add.ftz.f32 %f304, %f302, %f303;
mov.b32 %r895, %f304;
mov.u32 %r896, 1;
shfl.sync.bfly.b32 %r897|%p71, %r895, %r896, %r882, %r884;
mov.b32 %f305, %r897;
add.ftz.f32 %f40, %f304, %f305;
mov.b32 %r898, %f296;
shfl.sync.bfly.b32 %r899|%p72, %r898, %r883, %r882, %r884;
mov.b32 %f306, %r899;
add.ftz.f32 %f307, %f296, %f306;
mov.b32 %r900, %f307;
shfl.sync.bfly.b32 %r901|%p73, %r900, %r887, %r882, %r884;
mov.b32 %f308, %r901;
add.ftz.f32 %f309, %f307, %f308;
mov.b32 %r902, %f309;
shfl.sync.bfly.b32 %r903|%p74, %r902, %r890, %r882, %r884;
mov.b32 %f310, %r903;
add.ftz.f32 %f311, %f309, %f310;
mov.b32 %r904, %f311;
shfl.sync.bfly.b32 %r905|%p75, %r904, %r893, %r882, %r884;
mov.b32 %f312, %r905;
add.ftz.f32 %f313, %f311, %f312;
mov.b32 %r906, %f313;
shfl.sync.bfly.b32 %r907|%p76, %r906, %r896, %r882, %r884;
mov.b32 %f314, %r907;
add.ftz.f32 %f41, %f313, %f314;
setp.ne.s32 %p77, %r2, 0;
setp.ge.u32 %p78, %r4, %r49;
or.pred %p79, %p77, %p78;
@%p79 bra $L__BB170_48;

ld.param.f32 %f1069, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f315, %f4, %f40;
add.ftz.f32 %f316, %f5, %f41;
min.ftz.f32 %f317, %f316, %f1069;
min.ftz.f32 %f318, %f315, %f1069;
max.ftz.f32 %f319, %f318, %f35;
mul.ftz.f32 %f320, %f317, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f321, %f320;
add.ftz.f32 %f322, %f321, 0f3F800000;
div.approx.ftz.f32 %f323, %f317, %f322;
mul.ftz.f32 %f324, %f319, %f323;
st.global.f32 [%rd38], %f324;

$L__BB170_48:
ld.shared.f32 %f325, [%r30+128];
add.ftz.f32 %f326, %f325, %f29;
ld.shared.f32 %f327, [%r32+128];
add.ftz.f32 %f328, %f327, %f33;
ld.shared.f32 %f329, [%r30+640];
add.ftz.f32 %f330, %f329, %f326;
ld.shared.f32 %f331, [%r32+640];
add.ftz.f32 %f332, %f331, %f328;
ld.shared.f32 %f333, [%r30+1152];
add.ftz.f32 %f334, %f333, %f330;
ld.shared.f32 %f335, [%r32+1152];
add.ftz.f32 %f336, %f335, %f332;
mov.b32 %r911, %f334;
shfl.sync.bfly.b32 %r915|%p80, %r911, %r883, %r882, %r884;
mov.b32 %f337, %r915;
add.ftz.f32 %f338, %f334, %f337;
mov.b32 %r916, %f338;
shfl.sync.bfly.b32 %r918|%p81, %r916, %r887, %r882, %r884;
mov.b32 %f339, %r918;
add.ftz.f32 %f340, %f338, %f339;
mov.b32 %r919, %f340;
shfl.sync.bfly.b32 %r921|%p82, %r919, %r890, %r882, %r884;
mov.b32 %f341, %r921;
add.ftz.f32 %f342, %f340, %f341;
mov.b32 %r922, %f342;
shfl.sync.bfly.b32 %r924|%p83, %r922, %r893, %r882, %r884;
mov.b32 %f343, %r924;
add.ftz.f32 %f344, %f342, %f343;
mov.b32 %r925, %f344;
shfl.sync.bfly.b32 %r927|%p84, %r925, %r896, %r882, %r884;
mov.b32 %f345, %r927;
add.ftz.f32 %f42, %f344, %f345;
mov.b32 %r928, %f336;
shfl.sync.bfly.b32 %r929|%p85, %r928, %r883, %r882, %r884;
mov.b32 %f346, %r929;
add.ftz.f32 %f347, %f336, %f346;
mov.b32 %r930, %f347;
shfl.sync.bfly.b32 %r931|%p86, %r930, %r887, %r882, %r884;
mov.b32 %f348, %r931;
add.ftz.f32 %f349, %f347, %f348;
mov.b32 %r932, %f349;
shfl.sync.bfly.b32 %r933|%p87, %r932, %r890, %r882, %r884;
mov.b32 %f350, %r933;
add.ftz.f32 %f351, %f349, %f350;
mov.b32 %r934, %f351;
shfl.sync.bfly.b32 %r935|%p88, %r934, %r893, %r882, %r884;
mov.b32 %f352, %r935;
add.ftz.f32 %f353, %f351, %f352;
mov.b32 %r936, %f353;
shfl.sync.bfly.b32 %r937|%p89, %r936, %r896, %r882, %r884;
mov.b32 %f354, %r937;
add.ftz.f32 %f43, %f353, %f354;
setp.ne.s32 %p90, %r2, 1;
or.b32 %r941, %r4, 1;
setp.ge.u32 %p91, %r941, %r49;
or.pred %p92, %p90, %p91;
@%p92 bra $L__BB170_50;

ld.param.f32 %f1068, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f355, %f4, %f42;
add.ftz.f32 %f356, %f5, %f43;
min.ftz.f32 %f357, %f356, %f1068;
min.ftz.f32 %f358, %f355, %f1068;
max.ftz.f32 %f359, %f358, %f35;
mul.ftz.f32 %f360, %f357, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f361, %f360;
add.ftz.f32 %f362, %f361, 0f3F800000;
div.approx.ftz.f32 %f363, %f357, %f362;
mul.ftz.f32 %f364, %f359, %f363;
st.global.f32 [%rd38+4], %f364;

$L__BB170_50:
ld.shared.f32 %f365, [%r30+256];
add.ftz.f32 %f366, %f365, %f28;
ld.shared.f32 %f367, [%r32+256];
add.ftz.f32 %f368, %f367, %f32;
ld.shared.f32 %f369, [%r30+768];
add.ftz.f32 %f370, %f369, %f366;
ld.shared.f32 %f371, [%r32+768];
add.ftz.f32 %f372, %f371, %f368;
ld.shared.f32 %f373, [%r30+1280];
add.ftz.f32 %f374, %f373, %f370;
ld.shared.f32 %f375, [%r32+1280];
add.ftz.f32 %f376, %f375, %f372;
mov.b32 %r942, %f374;
mov.u32 %r943, 31;
mov.u32 %r944, 16;
mov.u32 %r945, -1;
shfl.sync.bfly.b32 %r946|%p93, %r942, %r944, %r943, %r945;
mov.b32 %f377, %r946;
add.ftz.f32 %f378, %f374, %f377;
mov.b32 %r947, %f378;
mov.u32 %r948, 8;
shfl.sync.bfly.b32 %r949|%p94, %r947, %r948, %r943, %r945;
mov.b32 %f379, %r949;
add.ftz.f32 %f380, %f378, %f379;
mov.b32 %r950, %f380;
mov.u32 %r951, 4;
shfl.sync.bfly.b32 %r952|%p95, %r950, %r951, %r943, %r945;
mov.b32 %f381, %r952;
add.ftz.f32 %f382, %f380, %f381;
mov.b32 %r953, %f382;
mov.u32 %r954, 2;
shfl.sync.bfly.b32 %r955|%p96, %r953, %r954, %r943, %r945;
mov.b32 %f383, %r955;
add.ftz.f32 %f384, %f382, %f383;
mov.b32 %r956, %f384;
mov.u32 %r957, 1;
shfl.sync.bfly.b32 %r958|%p97, %r956, %r957, %r943, %r945;
mov.b32 %f385, %r958;
add.ftz.f32 %f44, %f384, %f385;
mov.b32 %r959, %f376;
shfl.sync.bfly.b32 %r960|%p98, %r959, %r944, %r943, %r945;
mov.b32 %f386, %r960;
add.ftz.f32 %f387, %f376, %f386;
mov.b32 %r961, %f387;
shfl.sync.bfly.b32 %r962|%p99, %r961, %r948, %r943, %r945;
mov.b32 %f388, %r962;
add.ftz.f32 %f389, %f387, %f388;
mov.b32 %r963, %f389;
shfl.sync.bfly.b32 %r964|%p100, %r963, %r951, %r943, %r945;
mov.b32 %f390, %r964;
add.ftz.f32 %f391, %f389, %f390;
mov.b32 %r965, %f391;
shfl.sync.bfly.b32 %r966|%p101, %r965, %r954, %r943, %r945;
mov.b32 %f392, %r966;
add.ftz.f32 %f393, %f391, %f392;
mov.b32 %r967, %f393;
shfl.sync.bfly.b32 %r968|%p102, %r967, %r957, %r943, %r945;
mov.b32 %f394, %r968;
add.ftz.f32 %f45, %f393, %f394;
setp.ne.s32 %p103, %r2, 2;
or.b32 %r972, %r4, 2;
setp.ge.u32 %p104, %r972, %r49;
or.pred %p105, %p103, %p104;
@%p105 bra $L__BB170_52;

ld.param.f32 %f1067, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f395, %f4, %f44;
add.ftz.f32 %f396, %f5, %f45;
min.ftz.f32 %f397, %f396, %f1067;
min.ftz.f32 %f398, %f395, %f1067;
max.ftz.f32 %f399, %f398, %f35;
mul.ftz.f32 %f400, %f397, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f401, %f400;
add.ftz.f32 %f402, %f401, 0f3F800000;
div.approx.ftz.f32 %f403, %f397, %f402;
mul.ftz.f32 %f404, %f399, %f403;
st.global.f32 [%rd38+8], %f404;

$L__BB170_52:
ld.shared.f32 %f405, [%r30+384];
add.ftz.f32 %f406, %f405, %f27;
ld.shared.f32 %f407, [%r32+384];
add.ftz.f32 %f408, %f407, %f31;
ld.shared.f32 %f409, [%r30+896];
add.ftz.f32 %f410, %f409, %f406;
ld.shared.f32 %f411, [%r32+896];
add.ftz.f32 %f412, %f411, %f408;
ld.shared.f32 %f413, [%r30+1408];
add.ftz.f32 %f414, %f413, %f410;
ld.shared.f32 %f415, [%r32+1408];
add.ftz.f32 %f416, %f415, %f412;
mov.b32 %r973, %f414;
shfl.sync.bfly.b32 %r977|%p106, %r973, %r944, %r943, %r945;
mov.b32 %f417, %r977;
add.ftz.f32 %f418, %f414, %f417;
mov.b32 %r978, %f418;
shfl.sync.bfly.b32 %r980|%p107, %r978, %r948, %r943, %r945;
mov.b32 %f419, %r980;
add.ftz.f32 %f420, %f418, %f419;
mov.b32 %r981, %f420;
shfl.sync.bfly.b32 %r983|%p108, %r981, %r951, %r943, %r945;
mov.b32 %f421, %r983;
add.ftz.f32 %f422, %f420, %f421;
mov.b32 %r984, %f422;
shfl.sync.bfly.b32 %r986|%p109, %r984, %r954, %r943, %r945;
mov.b32 %f423, %r986;
add.ftz.f32 %f424, %f422, %f423;
mov.b32 %r987, %f424;
shfl.sync.bfly.b32 %r989|%p110, %r987, %r957, %r943, %r945;
mov.b32 %f425, %r989;
add.ftz.f32 %f46, %f424, %f425;
mov.b32 %r990, %f416;
shfl.sync.bfly.b32 %r991|%p111, %r990, %r944, %r943, %r945;
mov.b32 %f426, %r991;
add.ftz.f32 %f427, %f416, %f426;
mov.b32 %r992, %f427;
shfl.sync.bfly.b32 %r993|%p112, %r992, %r948, %r943, %r945;
mov.b32 %f428, %r993;
add.ftz.f32 %f429, %f427, %f428;
mov.b32 %r994, %f429;
shfl.sync.bfly.b32 %r995|%p113, %r994, %r951, %r943, %r945;
mov.b32 %f430, %r995;
add.ftz.f32 %f431, %f429, %f430;
mov.b32 %r996, %f431;
shfl.sync.bfly.b32 %r997|%p114, %r996, %r954, %r943, %r945;
mov.b32 %f432, %r997;
add.ftz.f32 %f433, %f431, %f432;
mov.b32 %r998, %f433;
shfl.sync.bfly.b32 %r999|%p115, %r998, %r957, %r943, %r945;
mov.b32 %f434, %r999;
add.ftz.f32 %f47, %f433, %f434;
setp.ne.s32 %p116, %r2, 3;
or.b32 %r1003, %r4, 3;
setp.ge.u32 %p117, %r1003, %r49;
or.pred %p118, %p116, %p117;
@%p118 bra $L__BB170_78;

ld.param.f32 %f1066, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f435, %f4, %f46;
add.ftz.f32 %f436, %f5, %f47;
min.ftz.f32 %f437, %f436, %f1066;
min.ftz.f32 %f438, %f435, %f1066;
max.ftz.f32 %f439, %f438, %f35;
mul.ftz.f32 %f440, %f437, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f441, %f440;
add.ftz.f32 %f442, %f441, 0f3F800000;
div.approx.ftz.f32 %f443, %f437, %f442;
mul.ftz.f32 %f444, %f439, %f443;
st.global.f32 [%rd38+12], %f444;
bra.uni $L__BB170_78;

$L__BB170_70:
ld.shared.f32 %f933, [%r30];
add.ftz.f32 %f934, %f933, %f30;
ld.shared.f32 %f935, [%r32];
add.ftz.f32 %f936, %f935, %f34;
ld.shared.f32 %f937, [%r30+512];
add.ftz.f32 %f938, %f937, %f934;
ld.shared.f32 %f939, [%r32+512];
add.ftz.f32 %f940, %f939, %f936;
ld.shared.f32 %f941, [%r30+1024];
add.ftz.f32 %f942, %f941, %f938;
ld.shared.f32 %f943, [%r32+1024];
add.ftz.f32 %f944, %f943, %f940;
mov.b32 %r1373, %f942;
mov.u32 %r1374, 31;
mov.u32 %r1375, 16;
mov.u32 %r1376, -1;
shfl.sync.bfly.b32 %r1377|%p275, %r1373, %r1375, %r1374, %r1376;
mov.b32 %f945, %r1377;
add.ftz.f32 %f946, %f942, %f945;
mov.b32 %r1378, %f946;
mov.u32 %r1379, 8;
shfl.sync.bfly.b32 %r1380|%p276, %r1378, %r1379, %r1374, %r1376;
mov.b32 %f947, %r1380;
add.ftz.f32 %f948, %f946, %f947;
mov.b32 %r1381, %f948;
mov.u32 %r1382, 4;
shfl.sync.bfly.b32 %r1383|%p277, %r1381, %r1382, %r1374, %r1376;
mov.b32 %f949, %r1383;
add.ftz.f32 %f950, %f948, %f949;
mov.b32 %r1384, %f950;
mov.u32 %r1385, 2;
shfl.sync.bfly.b32 %r1386|%p278, %r1384, %r1385, %r1374, %r1376;
mov.b32 %f951, %r1386;
add.ftz.f32 %f952, %f950, %f951;
mov.b32 %r1387, %f952;
mov.u32 %r1388, 1;
shfl.sync.bfly.b32 %r1389|%p279, %r1387, %r1388, %r1374, %r1376;
mov.b32 %f953, %r1389;
add.ftz.f32 %f72, %f952, %f953;
mov.b32 %r1390, %f944;
shfl.sync.bfly.b32 %r1391|%p280, %r1390, %r1375, %r1374, %r1376;
mov.b32 %f954, %r1391;
add.ftz.f32 %f955, %f944, %f954;
mov.b32 %r1392, %f955;
shfl.sync.bfly.b32 %r1393|%p281, %r1392, %r1379, %r1374, %r1376;
mov.b32 %f956, %r1393;
add.ftz.f32 %f957, %f955, %f956;
mov.b32 %r1394, %f957;
shfl.sync.bfly.b32 %r1395|%p282, %r1394, %r1382, %r1374, %r1376;
mov.b32 %f958, %r1395;
add.ftz.f32 %f959, %f957, %f958;
mov.b32 %r1396, %f959;
shfl.sync.bfly.b32 %r1397|%p283, %r1396, %r1385, %r1374, %r1376;
mov.b32 %f960, %r1397;
add.ftz.f32 %f961, %f959, %f960;
mov.b32 %r1398, %f961;
shfl.sync.bfly.b32 %r1399|%p284, %r1398, %r1388, %r1374, %r1376;
mov.b32 %f962, %r1399;
add.ftz.f32 %f73, %f961, %f962;
setp.ne.s32 %p285, %r2, 0;
setp.ge.u32 %p286, %r4, %r49;
or.pred %p287, %p285, %p286;
@%p287 bra $L__BB170_72;

add.ftz.f32 %f963, %f4, %f72;
add.ftz.f32 %f964, %f5, %f73;
mul.ftz.f32 %f965, %f963, %f964;
st.global.f32 [%rd38], %f965;

$L__BB170_72:
ld.shared.f32 %f966, [%r30+128];
add.ftz.f32 %f967, %f966, %f29;
ld.shared.f32 %f968, [%r32+128];
add.ftz.f32 %f969, %f968, %f33;
ld.shared.f32 %f970, [%r30+640];
add.ftz.f32 %f971, %f970, %f967;
ld.shared.f32 %f972, [%r32+640];
add.ftz.f32 %f973, %f972, %f969;
ld.shared.f32 %f974, [%r30+1152];
add.ftz.f32 %f975, %f974, %f971;
ld.shared.f32 %f976, [%r32+1152];
add.ftz.f32 %f977, %f976, %f973;
mov.b32 %r1403, %f975;
shfl.sync.bfly.b32 %r1407|%p288, %r1403, %r1375, %r1374, %r1376;
mov.b32 %f978, %r1407;
add.ftz.f32 %f979, %f975, %f978;
mov.b32 %r1408, %f979;
shfl.sync.bfly.b32 %r1410|%p289, %r1408, %r1379, %r1374, %r1376;
mov.b32 %f980, %r1410;
add.ftz.f32 %f981, %f979, %f980;
mov.b32 %r1411, %f981;
shfl.sync.bfly.b32 %r1413|%p290, %r1411, %r1382, %r1374, %r1376;
mov.b32 %f982, %r1413;
add.ftz.f32 %f983, %f981, %f982;
mov.b32 %r1414, %f983;
shfl.sync.bfly.b32 %r1416|%p291, %r1414, %r1385, %r1374, %r1376;
mov.b32 %f984, %r1416;
add.ftz.f32 %f985, %f983, %f984;
mov.b32 %r1417, %f985;
shfl.sync.bfly.b32 %r1419|%p292, %r1417, %r1388, %r1374, %r1376;
mov.b32 %f986, %r1419;
add.ftz.f32 %f74, %f985, %f986;
mov.b32 %r1420, %f977;
shfl.sync.bfly.b32 %r1421|%p293, %r1420, %r1375, %r1374, %r1376;
mov.b32 %f987, %r1421;
add.ftz.f32 %f988, %f977, %f987;
mov.b32 %r1422, %f988;
shfl.sync.bfly.b32 %r1423|%p294, %r1422, %r1379, %r1374, %r1376;
mov.b32 %f989, %r1423;
add.ftz.f32 %f990, %f988, %f989;
mov.b32 %r1424, %f990;
shfl.sync.bfly.b32 %r1425|%p295, %r1424, %r1382, %r1374, %r1376;
mov.b32 %f991, %r1425;
add.ftz.f32 %f992, %f990, %f991;
mov.b32 %r1426, %f992;
shfl.sync.bfly.b32 %r1427|%p296, %r1426, %r1385, %r1374, %r1376;
mov.b32 %f993, %r1427;
add.ftz.f32 %f994, %f992, %f993;
mov.b32 %r1428, %f994;
shfl.sync.bfly.b32 %r1429|%p297, %r1428, %r1388, %r1374, %r1376;
mov.b32 %f995, %r1429;
add.ftz.f32 %f75, %f994, %f995;
setp.ne.s32 %p298, %r2, 1;
or.b32 %r1433, %r4, 1;
setp.ge.u32 %p299, %r1433, %r49;
or.pred %p300, %p298, %p299;
@%p300 bra $L__BB170_74;

add.ftz.f32 %f996, %f4, %f74;
add.ftz.f32 %f997, %f5, %f75;
mul.ftz.f32 %f998, %f996, %f997;
st.global.f32 [%rd38+4], %f998;

$L__BB170_74:
ld.shared.f32 %f999, [%r30+256];
add.ftz.f32 %f1000, %f999, %f28;
ld.shared.f32 %f1001, [%r32+256];
add.ftz.f32 %f1002, %f1001, %f32;
ld.shared.f32 %f1003, [%r30+768];
add.ftz.f32 %f1004, %f1003, %f1000;
ld.shared.f32 %f1005, [%r32+768];
add.ftz.f32 %f1006, %f1005, %f1002;
ld.shared.f32 %f1007, [%r30+1280];
add.ftz.f32 %f1008, %f1007, %f1004;
ld.shared.f32 %f1009, [%r32+1280];
add.ftz.f32 %f1010, %f1009, %f1006;
mov.b32 %r1434, %f1008;
mov.u32 %r1435, 31;
mov.u32 %r1436, 16;
mov.u32 %r1437, -1;
shfl.sync.bfly.b32 %r1438|%p301, %r1434, %r1436, %r1435, %r1437;
mov.b32 %f1011, %r1438;
add.ftz.f32 %f1012, %f1008, %f1011;
mov.b32 %r1439, %f1012;
mov.u32 %r1440, 8;
shfl.sync.bfly.b32 %r1441|%p302, %r1439, %r1440, %r1435, %r1437;
mov.b32 %f1013, %r1441;
add.ftz.f32 %f1014, %f1012, %f1013;
mov.b32 %r1442, %f1014;
mov.u32 %r1443, 4;
shfl.sync.bfly.b32 %r1444|%p303, %r1442, %r1443, %r1435, %r1437;
mov.b32 %f1015, %r1444;
add.ftz.f32 %f1016, %f1014, %f1015;
mov.b32 %r1445, %f1016;
mov.u32 %r1446, 2;
shfl.sync.bfly.b32 %r1447|%p304, %r1445, %r1446, %r1435, %r1437;
mov.b32 %f1017, %r1447;
add.ftz.f32 %f1018, %f1016, %f1017;
mov.b32 %r1448, %f1018;
mov.u32 %r1449, 1;
shfl.sync.bfly.b32 %r1450|%p305, %r1448, %r1449, %r1435, %r1437;
mov.b32 %f1019, %r1450;
add.ftz.f32 %f76, %f1018, %f1019;
mov.b32 %r1451, %f1010;
shfl.sync.bfly.b32 %r1452|%p306, %r1451, %r1436, %r1435, %r1437;
mov.b32 %f1020, %r1452;
add.ftz.f32 %f1021, %f1010, %f1020;
mov.b32 %r1453, %f1021;
shfl.sync.bfly.b32 %r1454|%p307, %r1453, %r1440, %r1435, %r1437;
mov.b32 %f1022, %r1454;
add.ftz.f32 %f1023, %f1021, %f1022;
mov.b32 %r1455, %f1023;
shfl.sync.bfly.b32 %r1456|%p308, %r1455, %r1443, %r1435, %r1437;
mov.b32 %f1024, %r1456;
add.ftz.f32 %f1025, %f1023, %f1024;
mov.b32 %r1457, %f1025;
shfl.sync.bfly.b32 %r1458|%p309, %r1457, %r1446, %r1435, %r1437;
mov.b32 %f1026, %r1458;
add.ftz.f32 %f1027, %f1025, %f1026;
mov.b32 %r1459, %f1027;
shfl.sync.bfly.b32 %r1460|%p310, %r1459, %r1449, %r1435, %r1437;
mov.b32 %f1028, %r1460;
add.ftz.f32 %f77, %f1027, %f1028;
setp.ne.s32 %p311, %r2, 2;
or.b32 %r1464, %r4, 2;
setp.ge.u32 %p312, %r1464, %r49;
or.pred %p313, %p311, %p312;
@%p313 bra $L__BB170_76;

add.ftz.f32 %f1029, %f4, %f76;
add.ftz.f32 %f1030, %f5, %f77;
mul.ftz.f32 %f1031, %f1029, %f1030;
st.global.f32 [%rd38+8], %f1031;

$L__BB170_76:
ld.shared.f32 %f1032, [%r30+384];
add.ftz.f32 %f1033, %f1032, %f27;
ld.shared.f32 %f1034, [%r32+384];
add.ftz.f32 %f1035, %f1034, %f31;
ld.shared.f32 %f1036, [%r30+896];
add.ftz.f32 %f1037, %f1036, %f1033;
ld.shared.f32 %f1038, [%r32+896];
add.ftz.f32 %f1039, %f1038, %f1035;
ld.shared.f32 %f1040, [%r30+1408];
add.ftz.f32 %f1041, %f1040, %f1037;
ld.shared.f32 %f1042, [%r32+1408];
add.ftz.f32 %f1043, %f1042, %f1039;
mov.b32 %r1465, %f1041;
shfl.sync.bfly.b32 %r1469|%p314, %r1465, %r1436, %r1435, %r1437;
mov.b32 %f1044, %r1469;
add.ftz.f32 %f1045, %f1041, %f1044;
mov.b32 %r1470, %f1045;
shfl.sync.bfly.b32 %r1472|%p315, %r1470, %r1440, %r1435, %r1437;
mov.b32 %f1046, %r1472;
add.ftz.f32 %f1047, %f1045, %f1046;
mov.b32 %r1473, %f1047;
shfl.sync.bfly.b32 %r1475|%p316, %r1473, %r1443, %r1435, %r1437;
mov.b32 %f1048, %r1475;
add.ftz.f32 %f1049, %f1047, %f1048;
mov.b32 %r1476, %f1049;
shfl.sync.bfly.b32 %r1478|%p317, %r1476, %r1446, %r1435, %r1437;
mov.b32 %f1050, %r1478;
add.ftz.f32 %f1051, %f1049, %f1050;
mov.b32 %r1479, %f1051;
shfl.sync.bfly.b32 %r1481|%p318, %r1479, %r1449, %r1435, %r1437;
mov.b32 %f1052, %r1481;
add.ftz.f32 %f78, %f1051, %f1052;
mov.b32 %r1482, %f1043;
shfl.sync.bfly.b32 %r1483|%p319, %r1482, %r1436, %r1435, %r1437;
mov.b32 %f1053, %r1483;
add.ftz.f32 %f1054, %f1043, %f1053;
mov.b32 %r1484, %f1054;
shfl.sync.bfly.b32 %r1485|%p320, %r1484, %r1440, %r1435, %r1437;
mov.b32 %f1055, %r1485;
add.ftz.f32 %f1056, %f1054, %f1055;
mov.b32 %r1486, %f1056;
shfl.sync.bfly.b32 %r1487|%p321, %r1486, %r1443, %r1435, %r1437;
mov.b32 %f1057, %r1487;
add.ftz.f32 %f1058, %f1056, %f1057;
mov.b32 %r1488, %f1058;
shfl.sync.bfly.b32 %r1489|%p322, %r1488, %r1446, %r1435, %r1437;
mov.b32 %f1059, %r1489;
add.ftz.f32 %f1060, %f1058, %f1059;
mov.b32 %r1490, %f1060;
shfl.sync.bfly.b32 %r1491|%p323, %r1490, %r1449, %r1435, %r1437;
mov.b32 %f1061, %r1491;
add.ftz.f32 %f79, %f1060, %f1061;
setp.ne.s32 %p324, %r2, 3;
or.b32 %r1495, %r4, 3;
setp.ge.u32 %p325, %r1495, %r49;
or.pred %p326, %p324, %p325;
@%p326 bra $L__BB170_78;

add.ftz.f32 %f1062, %f4, %f78;
add.ftz.f32 %f1063, %f5, %f79;
mul.ftz.f32 %f1064, %f1062, %f1063;
st.global.f32 [%rd38+12], %f1064;
bra.uni $L__BB170_78;

$L__BB170_62:
ld.shared.f32 %f769, [%r30];
add.ftz.f32 %f770, %f769, %f30;
ld.shared.f32 %f771, [%r32];
add.ftz.f32 %f772, %f771, %f34;
ld.shared.f32 %f773, [%r30+512];
add.ftz.f32 %f774, %f773, %f770;
ld.shared.f32 %f775, [%r32+512];
add.ftz.f32 %f776, %f775, %f772;
ld.shared.f32 %f777, [%r30+1024];
add.ftz.f32 %f778, %f777, %f774;
ld.shared.f32 %f779, [%r32+1024];
add.ftz.f32 %f780, %f779, %f776;
mov.b32 %r1250, %f778;
mov.u32 %r1251, 31;
mov.u32 %r1252, 16;
mov.u32 %r1253, -1;
shfl.sync.bfly.b32 %r1254|%p223, %r1250, %r1252, %r1251, %r1253;
mov.b32 %f781, %r1254;
add.ftz.f32 %f782, %f778, %f781;
mov.b32 %r1255, %f782;
mov.u32 %r1256, 8;
shfl.sync.bfly.b32 %r1257|%p224, %r1255, %r1256, %r1251, %r1253;
mov.b32 %f783, %r1257;
add.ftz.f32 %f784, %f782, %f783;
mov.b32 %r1258, %f784;
mov.u32 %r1259, 4;
shfl.sync.bfly.b32 %r1260|%p225, %r1258, %r1259, %r1251, %r1253;
mov.b32 %f785, %r1260;
add.ftz.f32 %f786, %f784, %f785;
mov.b32 %r1261, %f786;
mov.u32 %r1262, 2;
shfl.sync.bfly.b32 %r1263|%p226, %r1261, %r1262, %r1251, %r1253;
mov.b32 %f787, %r1263;
add.ftz.f32 %f788, %f786, %f787;
mov.b32 %r1264, %f788;
mov.u32 %r1265, 1;
shfl.sync.bfly.b32 %r1266|%p227, %r1264, %r1265, %r1251, %r1253;
mov.b32 %f789, %r1266;
add.ftz.f32 %f64, %f788, %f789;
mov.b32 %r1267, %f780;
shfl.sync.bfly.b32 %r1268|%p228, %r1267, %r1252, %r1251, %r1253;
mov.b32 %f790, %r1268;
add.ftz.f32 %f791, %f780, %f790;
mov.b32 %r1269, %f791;
shfl.sync.bfly.b32 %r1270|%p229, %r1269, %r1256, %r1251, %r1253;
mov.b32 %f792, %r1270;
add.ftz.f32 %f793, %f791, %f792;
mov.b32 %r1271, %f793;
shfl.sync.bfly.b32 %r1272|%p230, %r1271, %r1259, %r1251, %r1253;
mov.b32 %f794, %r1272;
add.ftz.f32 %f795, %f793, %f794;
mov.b32 %r1273, %f795;
shfl.sync.bfly.b32 %r1274|%p231, %r1273, %r1262, %r1251, %r1253;
mov.b32 %f796, %r1274;
add.ftz.f32 %f797, %f795, %f796;
mov.b32 %r1275, %f797;
shfl.sync.bfly.b32 %r1276|%p232, %r1275, %r1265, %r1251, %r1253;
mov.b32 %f798, %r1276;
add.ftz.f32 %f65, %f797, %f798;
setp.ne.s32 %p233, %r2, 0;
setp.ge.u32 %p234, %r4, %r49;
or.pred %p235, %p233, %p234;
@%p235 bra $L__BB170_64;

add.ftz.f32 %f799, %f4, %f64;
add.ftz.f32 %f800, %f5, %f65;
mul.ftz.f32 %f801, %f800, 0f3F000000;
mul.ftz.f32 %f802, %f800, 0f3F4C422A;
mul.ftz.f32 %f803, %f800, 0f3D372713;
fma.rn.ftz.f32 %f804, %f800, %f803, 0f3F800000;
mul.ftz.f32 %f805, %f802, %f804;
tanh.approx.f32 %f806, %f805;
add.ftz.f32 %f807, %f806, 0f3F800000;
mul.ftz.f32 %f808, %f801, %f807;
mul.ftz.f32 %f809, %f799, %f808;
st.global.f32 [%rd38], %f809;

$L__BB170_64:
ld.shared.f32 %f810, [%r30+128];
add.ftz.f32 %f811, %f810, %f29;
ld.shared.f32 %f812, [%r32+128];
add.ftz.f32 %f813, %f812, %f33;
ld.shared.f32 %f814, [%r30+640];
add.ftz.f32 %f815, %f814, %f811;
ld.shared.f32 %f816, [%r32+640];
add.ftz.f32 %f817, %f816, %f813;
ld.shared.f32 %f818, [%r30+1152];
add.ftz.f32 %f819, %f818, %f815;
ld.shared.f32 %f820, [%r32+1152];
add.ftz.f32 %f821, %f820, %f817;
mov.b32 %r1280, %f819;
shfl.sync.bfly.b32 %r1284|%p236, %r1280, %r1252, %r1251, %r1253;
mov.b32 %f822, %r1284;
add.ftz.f32 %f823, %f819, %f822;
mov.b32 %r1285, %f823;
shfl.sync.bfly.b32 %r1287|%p237, %r1285, %r1256, %r1251, %r1253;
mov.b32 %f824, %r1287;
add.ftz.f32 %f825, %f823, %f824;
mov.b32 %r1288, %f825;
shfl.sync.bfly.b32 %r1290|%p238, %r1288, %r1259, %r1251, %r1253;
mov.b32 %f826, %r1290;
add.ftz.f32 %f827, %f825, %f826;
mov.b32 %r1291, %f827;
shfl.sync.bfly.b32 %r1293|%p239, %r1291, %r1262, %r1251, %r1253;
mov.b32 %f828, %r1293;
add.ftz.f32 %f829, %f827, %f828;
mov.b32 %r1294, %f829;
shfl.sync.bfly.b32 %r1296|%p240, %r1294, %r1265, %r1251, %r1253;
mov.b32 %f830, %r1296;
add.ftz.f32 %f66, %f829, %f830;
mov.b32 %r1297, %f821;
shfl.sync.bfly.b32 %r1298|%p241, %r1297, %r1252, %r1251, %r1253;
mov.b32 %f831, %r1298;
add.ftz.f32 %f832, %f821, %f831;
mov.b32 %r1299, %f832;
shfl.sync.bfly.b32 %r1300|%p242, %r1299, %r1256, %r1251, %r1253;
mov.b32 %f833, %r1300;
add.ftz.f32 %f834, %f832, %f833;
mov.b32 %r1301, %f834;
shfl.sync.bfly.b32 %r1302|%p243, %r1301, %r1259, %r1251, %r1253;
mov.b32 %f835, %r1302;
add.ftz.f32 %f836, %f834, %f835;
mov.b32 %r1303, %f836;
shfl.sync.bfly.b32 %r1304|%p244, %r1303, %r1262, %r1251, %r1253;
mov.b32 %f837, %r1304;
add.ftz.f32 %f838, %f836, %f837;
mov.b32 %r1305, %f838;
shfl.sync.bfly.b32 %r1306|%p245, %r1305, %r1265, %r1251, %r1253;
mov.b32 %f839, %r1306;
add.ftz.f32 %f67, %f838, %f839;
setp.ne.s32 %p246, %r2, 1;
or.b32 %r1310, %r4, 1;
setp.ge.u32 %p247, %r1310, %r49;
or.pred %p248, %p246, %p247;
@%p248 bra $L__BB170_66;

add.ftz.f32 %f840, %f4, %f66;
add.ftz.f32 %f841, %f5, %f67;
mul.ftz.f32 %f842, %f841, 0f3F000000;
mul.ftz.f32 %f843, %f841, 0f3F4C422A;
mul.ftz.f32 %f844, %f841, 0f3D372713;
fma.rn.ftz.f32 %f845, %f841, %f844, 0f3F800000;
mul.ftz.f32 %f846, %f843, %f845;
tanh.approx.f32 %f847, %f846;
add.ftz.f32 %f848, %f847, 0f3F800000;
mul.ftz.f32 %f849, %f842, %f848;
mul.ftz.f32 %f850, %f840, %f849;
st.global.f32 [%rd38+4], %f850;

$L__BB170_66:
ld.shared.f32 %f851, [%r30+256];
add.ftz.f32 %f852, %f851, %f28;
ld.shared.f32 %f853, [%r32+256];
add.ftz.f32 %f854, %f853, %f32;
ld.shared.f32 %f855, [%r30+768];
add.ftz.f32 %f856, %f855, %f852;
ld.shared.f32 %f857, [%r32+768];
add.ftz.f32 %f858, %f857, %f854;
ld.shared.f32 %f859, [%r30+1280];
add.ftz.f32 %f860, %f859, %f856;
ld.shared.f32 %f861, [%r32+1280];
add.ftz.f32 %f862, %f861, %f858;
mov.b32 %r1311, %f860;
mov.u32 %r1312, 31;
mov.u32 %r1313, 16;
mov.u32 %r1314, -1;
shfl.sync.bfly.b32 %r1315|%p249, %r1311, %r1313, %r1312, %r1314;
mov.b32 %f863, %r1315;
add.ftz.f32 %f864, %f860, %f863;
mov.b32 %r1316, %f864;
mov.u32 %r1317, 8;
shfl.sync.bfly.b32 %r1318|%p250, %r1316, %r1317, %r1312, %r1314;
mov.b32 %f865, %r1318;
add.ftz.f32 %f866, %f864, %f865;
mov.b32 %r1319, %f866;
mov.u32 %r1320, 4;
shfl.sync.bfly.b32 %r1321|%p251, %r1319, %r1320, %r1312, %r1314;
mov.b32 %f867, %r1321;
add.ftz.f32 %f868, %f866, %f867;
mov.b32 %r1322, %f868;
mov.u32 %r1323, 2;
shfl.sync.bfly.b32 %r1324|%p252, %r1322, %r1323, %r1312, %r1314;
mov.b32 %f869, %r1324;
add.ftz.f32 %f870, %f868, %f869;
mov.b32 %r1325, %f870;
mov.u32 %r1326, 1;
shfl.sync.bfly.b32 %r1327|%p253, %r1325, %r1326, %r1312, %r1314;
mov.b32 %f871, %r1327;
add.ftz.f32 %f68, %f870, %f871;
mov.b32 %r1328, %f862;
shfl.sync.bfly.b32 %r1329|%p254, %r1328, %r1313, %r1312, %r1314;
mov.b32 %f872, %r1329;
add.ftz.f32 %f873, %f862, %f872;
mov.b32 %r1330, %f873;
shfl.sync.bfly.b32 %r1331|%p255, %r1330, %r1317, %r1312, %r1314;
mov.b32 %f874, %r1331;
add.ftz.f32 %f875, %f873, %f874;
mov.b32 %r1332, %f875;
shfl.sync.bfly.b32 %r1333|%p256, %r1332, %r1320, %r1312, %r1314;
mov.b32 %f876, %r1333;
add.ftz.f32 %f877, %f875, %f876;
mov.b32 %r1334, %f877;
shfl.sync.bfly.b32 %r1335|%p257, %r1334, %r1323, %r1312, %r1314;
mov.b32 %f878, %r1335;
add.ftz.f32 %f879, %f877, %f878;
mov.b32 %r1336, %f879;
shfl.sync.bfly.b32 %r1337|%p258, %r1336, %r1326, %r1312, %r1314;
mov.b32 %f880, %r1337;
add.ftz.f32 %f69, %f879, %f880;
setp.ne.s32 %p259, %r2, 2;
or.b32 %r1341, %r4, 2;
setp.ge.u32 %p260, %r1341, %r49;
or.pred %p261, %p259, %p260;
@%p261 bra $L__BB170_68;

add.ftz.f32 %f881, %f4, %f68;
add.ftz.f32 %f882, %f5, %f69;
mul.ftz.f32 %f883, %f882, 0f3F000000;
mul.ftz.f32 %f884, %f882, 0f3F4C422A;
mul.ftz.f32 %f885, %f882, 0f3D372713;
fma.rn.ftz.f32 %f886, %f882, %f885, 0f3F800000;
mul.ftz.f32 %f887, %f884, %f886;
tanh.approx.f32 %f888, %f887;
add.ftz.f32 %f889, %f888, 0f3F800000;
mul.ftz.f32 %f890, %f883, %f889;
mul.ftz.f32 %f891, %f881, %f890;
st.global.f32 [%rd38+8], %f891;

$L__BB170_68:
ld.shared.f32 %f892, [%r30+384];
add.ftz.f32 %f893, %f892, %f27;
ld.shared.f32 %f894, [%r32+384];
add.ftz.f32 %f895, %f894, %f31;
ld.shared.f32 %f896, [%r30+896];
add.ftz.f32 %f897, %f896, %f893;
ld.shared.f32 %f898, [%r32+896];
add.ftz.f32 %f899, %f898, %f895;
ld.shared.f32 %f900, [%r30+1408];
add.ftz.f32 %f901, %f900, %f897;
ld.shared.f32 %f902, [%r32+1408];
add.ftz.f32 %f903, %f902, %f899;
mov.b32 %r1342, %f901;
shfl.sync.bfly.b32 %r1346|%p262, %r1342, %r1313, %r1312, %r1314;
mov.b32 %f904, %r1346;
add.ftz.f32 %f905, %f901, %f904;
mov.b32 %r1347, %f905;
shfl.sync.bfly.b32 %r1349|%p263, %r1347, %r1317, %r1312, %r1314;
mov.b32 %f906, %r1349;
add.ftz.f32 %f907, %f905, %f906;
mov.b32 %r1350, %f907;
shfl.sync.bfly.b32 %r1352|%p264, %r1350, %r1320, %r1312, %r1314;
mov.b32 %f908, %r1352;
add.ftz.f32 %f909, %f907, %f908;
mov.b32 %r1353, %f909;
shfl.sync.bfly.b32 %r1355|%p265, %r1353, %r1323, %r1312, %r1314;
mov.b32 %f910, %r1355;
add.ftz.f32 %f911, %f909, %f910;
mov.b32 %r1356, %f911;
shfl.sync.bfly.b32 %r1358|%p266, %r1356, %r1326, %r1312, %r1314;
mov.b32 %f912, %r1358;
add.ftz.f32 %f70, %f911, %f912;
mov.b32 %r1359, %f903;
shfl.sync.bfly.b32 %r1360|%p267, %r1359, %r1313, %r1312, %r1314;
mov.b32 %f913, %r1360;
add.ftz.f32 %f914, %f903, %f913;
mov.b32 %r1361, %f914;
shfl.sync.bfly.b32 %r1362|%p268, %r1361, %r1317, %r1312, %r1314;
mov.b32 %f915, %r1362;
add.ftz.f32 %f916, %f914, %f915;
mov.b32 %r1363, %f916;
shfl.sync.bfly.b32 %r1364|%p269, %r1363, %r1320, %r1312, %r1314;
mov.b32 %f917, %r1364;
add.ftz.f32 %f918, %f916, %f917;
mov.b32 %r1365, %f918;
shfl.sync.bfly.b32 %r1366|%p270, %r1365, %r1323, %r1312, %r1314;
mov.b32 %f919, %r1366;
add.ftz.f32 %f920, %f918, %f919;
mov.b32 %r1367, %f920;
shfl.sync.bfly.b32 %r1368|%p271, %r1367, %r1326, %r1312, %r1314;
mov.b32 %f921, %r1368;
add.ftz.f32 %f71, %f920, %f921;
setp.ne.s32 %p272, %r2, 3;
or.b32 %r1372, %r4, 3;
setp.ge.u32 %p273, %r1372, %r49;
or.pred %p274, %p272, %p273;
@%p274 bra $L__BB170_78;

add.ftz.f32 %f922, %f4, %f70;
add.ftz.f32 %f923, %f5, %f71;
mul.ftz.f32 %f924, %f923, 0f3F000000;
mul.ftz.f32 %f925, %f923, 0f3F4C422A;
mul.ftz.f32 %f926, %f923, 0f3D372713;
fma.rn.ftz.f32 %f927, %f923, %f926, 0f3F800000;
mul.ftz.f32 %f928, %f925, %f927;
tanh.approx.f32 %f929, %f928;
add.ftz.f32 %f930, %f929, 0f3F800000;
mul.ftz.f32 %f931, %f924, %f930;
mul.ftz.f32 %f932, %f922, %f931;
st.global.f32 [%rd38+12], %f932;
bra.uni $L__BB170_78;

$L__BB170_54:
ld.shared.f32 %f445, [%r30];
add.ftz.f32 %f446, %f445, %f30;
ld.shared.f32 %f447, [%r32];
add.ftz.f32 %f448, %f447, %f34;
ld.shared.f32 %f449, [%r30+512];
add.ftz.f32 %f450, %f449, %f446;
ld.shared.f32 %f451, [%r32+512];
add.ftz.f32 %f452, %f451, %f448;
ld.shared.f32 %f453, [%r30+1024];
add.ftz.f32 %f454, %f453, %f450;
ld.shared.f32 %f455, [%r32+1024];
add.ftz.f32 %f456, %f455, %f452;
mov.b32 %r1004, %f454;
mov.u32 %r1005, 31;
mov.u32 %r1006, 16;
mov.u32 %r1007, -1;
shfl.sync.bfly.b32 %r1008|%p119, %r1004, %r1006, %r1005, %r1007;
mov.b32 %f457, %r1008;
add.ftz.f32 %f458, %f454, %f457;
mov.b32 %r1009, %f458;
mov.u32 %r1010, 8;
shfl.sync.bfly.b32 %r1011|%p120, %r1009, %r1010, %r1005, %r1007;
mov.b32 %f459, %r1011;
add.ftz.f32 %f460, %f458, %f459;
mov.b32 %r1012, %f460;
mov.u32 %r1013, 4;
shfl.sync.bfly.b32 %r1014|%p121, %r1012, %r1013, %r1005, %r1007;
mov.b32 %f461, %r1014;
add.ftz.f32 %f462, %f460, %f461;
mov.b32 %r1015, %f462;
mov.u32 %r1016, 2;
shfl.sync.bfly.b32 %r1017|%p122, %r1015, %r1016, %r1005, %r1007;
mov.b32 %f463, %r1017;
add.ftz.f32 %f464, %f462, %f463;
mov.b32 %r1018, %f464;
mov.u32 %r1019, 1;
shfl.sync.bfly.b32 %r1020|%p123, %r1018, %r1019, %r1005, %r1007;
mov.b32 %f465, %r1020;
add.ftz.f32 %f48, %f464, %f465;
mov.b32 %r1021, %f456;
shfl.sync.bfly.b32 %r1022|%p124, %r1021, %r1006, %r1005, %r1007;
mov.b32 %f466, %r1022;
add.ftz.f32 %f467, %f456, %f466;
mov.b32 %r1023, %f467;
shfl.sync.bfly.b32 %r1024|%p125, %r1023, %r1010, %r1005, %r1007;
mov.b32 %f468, %r1024;
add.ftz.f32 %f469, %f467, %f468;
mov.b32 %r1025, %f469;
shfl.sync.bfly.b32 %r1026|%p126, %r1025, %r1013, %r1005, %r1007;
mov.b32 %f470, %r1026;
add.ftz.f32 %f471, %f469, %f470;
mov.b32 %r1027, %f471;
shfl.sync.bfly.b32 %r1028|%p127, %r1027, %r1016, %r1005, %r1007;
mov.b32 %f472, %r1028;
add.ftz.f32 %f473, %f471, %f472;
mov.b32 %r1029, %f473;
shfl.sync.bfly.b32 %r1030|%p128, %r1029, %r1019, %r1005, %r1007;
mov.b32 %f474, %r1030;
add.ftz.f32 %f49, %f473, %f474;
setp.ne.s32 %p129, %r2, 0;
setp.ge.u32 %p130, %r4, %r49;
or.pred %p131, %p129, %p130;
@%p131 bra $L__BB170_56;

add.ftz.f32 %f475, %f4, %f48;
add.ftz.f32 %f476, %f5, %f49;
mov.f32 %f477, 0f40E00000;
min.ftz.f32 %f478, %f476, %f477;
min.ftz.f32 %f479, %f475, %f477;
mov.f32 %f480, 0fC0E00000;
max.ftz.f32 %f481, %f479, %f480;
mul.ftz.f32 %f482, %f478, 0f3FD9DB23;
mul.ftz.f32 %f483, %f482, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f484, %f483;
add.ftz.f32 %f485, %f484, 0f3F800000;
div.approx.ftz.f32 %f486, %f478, %f485;
add.ftz.f32 %f487, %f481, 0f3F800000;
mul.ftz.f32 %f488, %f487, %f486;
st.global.f32 [%rd38], %f488;

$L__BB170_56:
ld.shared.f32 %f489, [%r30+128];
add.ftz.f32 %f490, %f489, %f29;
ld.shared.f32 %f491, [%r32+128];
add.ftz.f32 %f492, %f491, %f33;
ld.shared.f32 %f493, [%r30+640];
add.ftz.f32 %f494, %f493, %f490;
ld.shared.f32 %f495, [%r32+640];
add.ftz.f32 %f496, %f495, %f492;
ld.shared.f32 %f497, [%r30+1152];
add.ftz.f32 %f498, %f497, %f494;
ld.shared.f32 %f499, [%r32+1152];
add.ftz.f32 %f500, %f499, %f496;
mov.b32 %r1034, %f498;
shfl.sync.bfly.b32 %r1038|%p132, %r1034, %r1006, %r1005, %r1007;
mov.b32 %f501, %r1038;
add.ftz.f32 %f502, %f498, %f501;
mov.b32 %r1039, %f502;
shfl.sync.bfly.b32 %r1041|%p133, %r1039, %r1010, %r1005, %r1007;
mov.b32 %f503, %r1041;
add.ftz.f32 %f504, %f502, %f503;
mov.b32 %r1042, %f504;
shfl.sync.bfly.b32 %r1044|%p134, %r1042, %r1013, %r1005, %r1007;
mov.b32 %f505, %r1044;
add.ftz.f32 %f506, %f504, %f505;
mov.b32 %r1045, %f506;
shfl.sync.bfly.b32 %r1047|%p135, %r1045, %r1016, %r1005, %r1007;
mov.b32 %f507, %r1047;
add.ftz.f32 %f508, %f506, %f507;
mov.b32 %r1048, %f508;
shfl.sync.bfly.b32 %r1050|%p136, %r1048, %r1019, %r1005, %r1007;
mov.b32 %f509, %r1050;
add.ftz.f32 %f50, %f508, %f509;
mov.b32 %r1051, %f500;
shfl.sync.bfly.b32 %r1052|%p137, %r1051, %r1006, %r1005, %r1007;
mov.b32 %f510, %r1052;
add.ftz.f32 %f511, %f500, %f510;
mov.b32 %r1053, %f511;
shfl.sync.bfly.b32 %r1054|%p138, %r1053, %r1010, %r1005, %r1007;
mov.b32 %f512, %r1054;
add.ftz.f32 %f513, %f511, %f512;
mov.b32 %r1055, %f513;
shfl.sync.bfly.b32 %r1056|%p139, %r1055, %r1013, %r1005, %r1007;
mov.b32 %f514, %r1056;
add.ftz.f32 %f515, %f513, %f514;
mov.b32 %r1057, %f515;
shfl.sync.bfly.b32 %r1058|%p140, %r1057, %r1016, %r1005, %r1007;
mov.b32 %f516, %r1058;
add.ftz.f32 %f517, %f515, %f516;
mov.b32 %r1059, %f517;
shfl.sync.bfly.b32 %r1060|%p141, %r1059, %r1019, %r1005, %r1007;
mov.b32 %f518, %r1060;
add.ftz.f32 %f51, %f517, %f518;
setp.ne.s32 %p142, %r2, 1;
or.b32 %r1064, %r4, 1;
setp.ge.u32 %p143, %r1064, %r49;
or.pred %p144, %p142, %p143;
@%p144 bra $L__BB170_58;

add.ftz.f32 %f519, %f4, %f50;
add.ftz.f32 %f520, %f5, %f51;
mov.f32 %f521, 0f40E00000;
min.ftz.f32 %f522, %f520, %f521;
min.ftz.f32 %f523, %f519, %f521;
mov.f32 %f524, 0fC0E00000;
max.ftz.f32 %f525, %f523, %f524;
mul.ftz.f32 %f526, %f522, 0f3FD9DB23;
mul.ftz.f32 %f527, %f526, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f528, %f527;
add.ftz.f32 %f529, %f528, 0f3F800000;
div.approx.ftz.f32 %f530, %f522, %f529;
add.ftz.f32 %f531, %f525, 0f3F800000;
mul.ftz.f32 %f532, %f531, %f530;
st.global.f32 [%rd38+4], %f532;

$L__BB170_58:
ld.shared.f32 %f533, [%r30+256];
add.ftz.f32 %f534, %f533, %f28;
ld.shared.f32 %f535, [%r32+256];
add.ftz.f32 %f536, %f535, %f32;
ld.shared.f32 %f537, [%r30+768];
add.ftz.f32 %f538, %f537, %f534;
ld.shared.f32 %f539, [%r32+768];
add.ftz.f32 %f540, %f539, %f536;
ld.shared.f32 %f541, [%r30+1280];
add.ftz.f32 %f542, %f541, %f538;
ld.shared.f32 %f543, [%r32+1280];
add.ftz.f32 %f544, %f543, %f540;
mov.b32 %r1065, %f542;
mov.u32 %r1066, 31;
mov.u32 %r1067, 16;
mov.u32 %r1068, -1;
shfl.sync.bfly.b32 %r1069|%p145, %r1065, %r1067, %r1066, %r1068;
mov.b32 %f545, %r1069;
add.ftz.f32 %f546, %f542, %f545;
mov.b32 %r1070, %f546;
mov.u32 %r1071, 8;
shfl.sync.bfly.b32 %r1072|%p146, %r1070, %r1071, %r1066, %r1068;
mov.b32 %f547, %r1072;
add.ftz.f32 %f548, %f546, %f547;
mov.b32 %r1073, %f548;
mov.u32 %r1074, 4;
shfl.sync.bfly.b32 %r1075|%p147, %r1073, %r1074, %r1066, %r1068;
mov.b32 %f549, %r1075;
add.ftz.f32 %f550, %f548, %f549;
mov.b32 %r1076, %f550;
mov.u32 %r1077, 2;
shfl.sync.bfly.b32 %r1078|%p148, %r1076, %r1077, %r1066, %r1068;
mov.b32 %f551, %r1078;
add.ftz.f32 %f552, %f550, %f551;
mov.b32 %r1079, %f552;
mov.u32 %r1080, 1;
shfl.sync.bfly.b32 %r1081|%p149, %r1079, %r1080, %r1066, %r1068;
mov.b32 %f553, %r1081;
add.ftz.f32 %f52, %f552, %f553;
mov.b32 %r1082, %f544;
shfl.sync.bfly.b32 %r1083|%p150, %r1082, %r1067, %r1066, %r1068;
mov.b32 %f554, %r1083;
add.ftz.f32 %f555, %f544, %f554;
mov.b32 %r1084, %f555;
shfl.sync.bfly.b32 %r1085|%p151, %r1084, %r1071, %r1066, %r1068;
mov.b32 %f556, %r1085;
add.ftz.f32 %f557, %f555, %f556;
mov.b32 %r1086, %f557;
shfl.sync.bfly.b32 %r1087|%p152, %r1086, %r1074, %r1066, %r1068;
mov.b32 %f558, %r1087;
add.ftz.f32 %f559, %f557, %f558;
mov.b32 %r1088, %f559;
shfl.sync.bfly.b32 %r1089|%p153, %r1088, %r1077, %r1066, %r1068;
mov.b32 %f560, %r1089;
add.ftz.f32 %f561, %f559, %f560;
mov.b32 %r1090, %f561;
shfl.sync.bfly.b32 %r1091|%p154, %r1090, %r1080, %r1066, %r1068;
mov.b32 %f562, %r1091;
add.ftz.f32 %f53, %f561, %f562;
setp.ne.s32 %p155, %r2, 2;
or.b32 %r1095, %r4, 2;
setp.ge.u32 %p156, %r1095, %r49;
or.pred %p157, %p155, %p156;
@%p157 bra $L__BB170_60;

add.ftz.f32 %f563, %f4, %f52;
add.ftz.f32 %f564, %f5, %f53;
mov.f32 %f565, 0f40E00000;
min.ftz.f32 %f566, %f564, %f565;
min.ftz.f32 %f567, %f563, %f565;
mov.f32 %f568, 0fC0E00000;
max.ftz.f32 %f569, %f567, %f568;
mul.ftz.f32 %f570, %f566, 0f3FD9DB23;
mul.ftz.f32 %f571, %f570, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f572, %f571;
add.ftz.f32 %f573, %f572, 0f3F800000;
div.approx.ftz.f32 %f574, %f566, %f573;
add.ftz.f32 %f575, %f569, 0f3F800000;
mul.ftz.f32 %f576, %f575, %f574;
st.global.f32 [%rd38+8], %f576;

$L__BB170_60:
ld.shared.f32 %f577, [%r30+384];
add.ftz.f32 %f578, %f577, %f27;
ld.shared.f32 %f579, [%r32+384];
add.ftz.f32 %f580, %f579, %f31;
ld.shared.f32 %f581, [%r30+896];
add.ftz.f32 %f582, %f581, %f578;
ld.shared.f32 %f583, [%r32+896];
add.ftz.f32 %f584, %f583, %f580;
ld.shared.f32 %f585, [%r30+1408];
add.ftz.f32 %f586, %f585, %f582;
ld.shared.f32 %f587, [%r32+1408];
add.ftz.f32 %f588, %f587, %f584;
mov.b32 %r1096, %f586;
shfl.sync.bfly.b32 %r1100|%p158, %r1096, %r1067, %r1066, %r1068;
mov.b32 %f589, %r1100;
add.ftz.f32 %f590, %f586, %f589;
mov.b32 %r1101, %f590;
shfl.sync.bfly.b32 %r1103|%p159, %r1101, %r1071, %r1066, %r1068;
mov.b32 %f591, %r1103;
add.ftz.f32 %f592, %f590, %f591;
mov.b32 %r1104, %f592;
shfl.sync.bfly.b32 %r1106|%p160, %r1104, %r1074, %r1066, %r1068;
mov.b32 %f593, %r1106;
add.ftz.f32 %f594, %f592, %f593;
mov.b32 %r1107, %f594;
shfl.sync.bfly.b32 %r1109|%p161, %r1107, %r1077, %r1066, %r1068;
mov.b32 %f595, %r1109;
add.ftz.f32 %f596, %f594, %f595;
mov.b32 %r1110, %f596;
shfl.sync.bfly.b32 %r1112|%p162, %r1110, %r1080, %r1066, %r1068;
mov.b32 %f597, %r1112;
add.ftz.f32 %f54, %f596, %f597;
mov.b32 %r1113, %f588;
shfl.sync.bfly.b32 %r1114|%p163, %r1113, %r1067, %r1066, %r1068;
mov.b32 %f598, %r1114;
add.ftz.f32 %f599, %f588, %f598;
mov.b32 %r1115, %f599;
shfl.sync.bfly.b32 %r1116|%p164, %r1115, %r1071, %r1066, %r1068;
mov.b32 %f600, %r1116;
add.ftz.f32 %f601, %f599, %f600;
mov.b32 %r1117, %f601;
shfl.sync.bfly.b32 %r1118|%p165, %r1117, %r1074, %r1066, %r1068;
mov.b32 %f602, %r1118;
add.ftz.f32 %f603, %f601, %f602;
mov.b32 %r1119, %f603;
shfl.sync.bfly.b32 %r1120|%p166, %r1119, %r1077, %r1066, %r1068;
mov.b32 %f604, %r1120;
add.ftz.f32 %f605, %f603, %f604;
mov.b32 %r1121, %f605;
shfl.sync.bfly.b32 %r1122|%p167, %r1121, %r1080, %r1066, %r1068;
mov.b32 %f606, %r1122;
add.ftz.f32 %f55, %f605, %f606;
setp.ne.s32 %p168, %r2, 3;
or.b32 %r1126, %r4, 3;
setp.ge.u32 %p169, %r1126, %r49;
or.pred %p170, %p168, %p169;
@%p170 bra $L__BB170_78;

add.ftz.f32 %f607, %f4, %f54;
add.ftz.f32 %f608, %f5, %f55;
mov.f32 %f609, 0f40E00000;
min.ftz.f32 %f610, %f608, %f609;
min.ftz.f32 %f611, %f607, %f609;
mov.f32 %f612, 0fC0E00000;
max.ftz.f32 %f613, %f611, %f612;
mul.ftz.f32 %f614, %f610, 0f3FD9DB23;
mul.ftz.f32 %f615, %f614, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f616, %f615;
add.ftz.f32 %f617, %f616, 0f3F800000;
div.approx.ftz.f32 %f618, %f610, %f617;
add.ftz.f32 %f619, %f613, 0f3F800000;
mul.ftz.f32 %f620, %f619, %f618;
st.global.f32 [%rd38+12], %f620;
bra.uni $L__BB170_78;

}
.entry _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<39>;
.reg .b16 %rs<29>;
.reg .f32 %f<153>;
.reg .b32 %r<479>;
.reg .b64 %rd<59>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

ld.param.u64 %rd13, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd14, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
mov.u32 %r1, %tid.y;
shl.b32 %r40, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r40, %r2;
mov.u32 %r41, %ctaid.x;
shl.b32 %r4, %r41, 2;
shr.u32 %r5, %r39, 8;
mov.u32 %r476, %ctaid.y;
setp.eq.s64 %p1, %rd15, 0;
@%p1 bra $L__BB171_2;

cvta.to.global.u64 %rd17, %rd15;
mul.wide.u32 %rd18, %r476, 4;
add.s64 %rd19, %rd17, %rd18;
ld.global.nc.u32 %r475, [%rd19];
mul.hi.u32 %r42, %r476, %r22;
add.s32 %r43, %r42, %r476;
shr.u32 %r44, %r43, %r23;
mul.lo.s32 %r45, %r44, %r24;
sub.s32 %r476, %r476, %r45;
bra.uni $L__BB171_3;

$L__BB171_2:
mul.hi.u32 %r46, %r476, %r27;
add.s32 %r47, %r46, %r476;
shr.u32 %r475, %r47, %r28;

$L__BB171_3:
shr.s32 %r48, %r3, 31;
shr.u32 %r49, %r48, 28;
add.s32 %r50, %r3, %r49;
shr.s32 %r478, %r50, 4;
setp.ge.s32 %p2, %r478, %r5;
mov.f32 %f149, 0f00000000;
mov.f32 %f150, %f149;
mov.f32 %f151, %f149;
mov.f32 %f152, %f149;
@%p2 bra $L__BB171_6;

mov.u32 %r51, %ctaid.z;
mul.lo.s32 %r52, %r476, %r31;
cvt.u64.u32 %rd20, %r52;
and.b32 %r56, %r50, -16;
sub.s32 %r13, %r3, %r56;
shr.s32 %r57, %r13, 31;
shr.u32 %r58, %r57, 30;
add.s32 %r59, %r13, %r58;
shr.s32 %r60, %r59, 2;
shl.b32 %r14, %r60, 1;
shl.b32 %r61, %r60, 5;
and.b32 %r62, %r59, -4;
sub.s32 %r63, %r13, %r62;
shl.b32 %r64, %r63, 2;
add.s32 %r65, %r61, %r64;
cvt.s64.s32 %rd21, %r65;
cvt.s64.s32 %rd22, %r64;
cvt.s64.s32 %rd23, %r14;
mul.lo.s32 %r66, %r51, %r37;
cvt.u64.u32 %rd24, %r66;
add.s64 %rd25, %rd24, %rd23;
add.s64 %rd1, %rd25, %rd20;
cvt.s64.s32 %rd2, %r63;
shl.b32 %r477, %r478, 3;
mad.lo.s32 %r67, %r4, %r25, %r478;
mul.hi.u32 %r68, %r51, %r33;
add.s32 %r69, %r68, %r51;
shr.u32 %r70, %r69, %r34;
mad.lo.s32 %r71, %r70, %r36, %r67;
mad.lo.s32 %r72, %r475, %r30, %r71;
mul.wide.s32 %rd3, %r72, 176;
add.s64 %rd26, %rd3, %rd21;
add.s64 %rd4, %rd26, 64;
cvt.u64.u32 %rd27, %r14;
and.b64 %rd28, %rd27, 2;
add.s64 %rd29, %rd3, %rd22;
add.s64 %rd5, %rd29, 32;
or.b64 %rd30, %rd28, %rd3;
or.b64 %rd6, %rd30, 12;
mul.wide.s32 %rd7, %r25, 176;
cvta.to.global.u64 %rd8, %rd14;
cvta.to.global.u64 %rd58, %rd13;
mov.f32 %f149, 0f00000000;
shl.b64 %rd41, %rd2, 2;

$L__BB171_5:
add.s64 %rd31, %rd58, %rd4;
add.s64 %rd32, %rd31, -16;
add.s64 %rd33, %rd58, %rd5;
add.s64 %rd34, %rd33, -16;
ld.global.nc.u32 %r217, [%rd31+-16];
ld.global.nc.u32 %r218, [%rd31];
ld.global.nc.u32 %r219, [%rd33+-16];
shr.s32 %r220, %r219, %r14;
ld.global.nc.u32 %r221, [%rd33];
shr.s32 %r222, %r221, %r14;
add.s64 %rd35, %rd58, %rd6;
add.s64 %rd36, %rd35, -8;
ld.global.nc.u16 %rs1, [%rd35+-8];
cvt.u32.u16 %r223, %rs1;
ld.global.nc.u16 %rs2, [%rd35+-4];
cvt.u32.u16 %r224, %rs2;
ld.global.nc.u16 %rs3, [%rd35];
cvt.u32.u16 %r225, %rs3;
and.b32 %r226, %r223, 16191;
and.b32 %r227, %r225, 3855;
and.b32 %r228, %r223, 49344;
shr.u32 %r229, %r228, 2;
or.b32 %r230, %r227, %r229;
setp.gt.s32 %p3, %r13, 7;
and.b32 %r231, %r224, 16191;
shr.u32 %r232, %r225, 4;
and.b32 %r233, %r232, 3855;
and.b32 %r234, %r224, 49344;
shr.u32 %r235, %r234, 2;
or.b32 %r236, %r233, %r235;
cvt.u64.u32 %rd37, %r477;
add.s64 %rd38, %rd1, %rd37;
mul.lo.s64 %rd39, %rd38, 36;
add.s64 %rd40, %rd8, %rd39;
ld.global.nc.u32 %r73, [%rd40];

{.reg .f16 low,high;
mov.b32 {low,high},%r73;
cvt.f32.f16 %f25, low;}


add.s64 %rd42, %rd40, %rd41;
ld.global.nc.u32 %r81, [%rd42+4];
ld.global.nc.u32 %r77, [%rd42+20];
ld.global.nc.u32 %r74, [%rd40+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r74;
cvt.f32.f16 %f26, low;}


ld.global.nc.u32 %r97, [%rd42+40];
ld.global.nc.u32 %r93, [%rd42+56];
add.s64 %rd43, %rd58, %rd3;
and.b32 %r237, %r217, 252645135;
and.b32 %r238, %r218, 252645135;
shl.b32 %r239, %r220, 4;
and.b32 %r240, %r239, 269488144;
shl.b32 %r241, %r222, 4;
and.b32 %r242, %r241, 269488144;
or.b32 %r80, %r237, %r240;
or.b32 %r76, %r238, %r242;
mov.u32 %r210, 0;

dp4a.s32.s32 %r75, %r76, %r77, %r210;


dp4a.s32.s32 %r79, %r80, %r81, %r75;

mov.u32 %r212, 16843009;

dp4a.s32.s32 %r83, %r212, %r77, %r210;


dp4a.s32.s32 %r87, %r212, %r81, %r83;

selp.b32 %r243, %r230, %r226, %p3;
cvt.u16.u32 %rs4, %r243;
and.b32 %r244, %r243, 63;
mul.lo.s32 %r245, %r79, %r244;
cvt.rn.f32.s32 %f41, %r245;
fma.rn.ftz.f32 %f42, %f25, %f41, 0f00000000;
selp.b32 %r246, %r236, %r231, %p3;
cvt.u16.u32 %rs5, %r246;
and.b32 %r247, %r246, 63;
mul.lo.s32 %r248, %r87, %r247;
cvt.rn.f32.s32 %f43, %r248;
fma.rn.ftz.f32 %f44, %f25, %f43, 0f00000000;
shr.u32 %r249, %r217, 4;
and.b32 %r250, %r249, 252645135;
shr.u32 %r251, %r218, 4;
and.b32 %r252, %r251, 252645135;
shl.b32 %r253, %r220, 3;
and.b32 %r254, %r253, 269488144;
shl.b32 %r255, %r222, 3;
and.b32 %r256, %r255, 269488144;
or.b32 %r96, %r250, %r254;
or.b32 %r92, %r252, %r256;

dp4a.s32.s32 %r91, %r92, %r93, %r210;


dp4a.s32.s32 %r95, %r96, %r97, %r91;


dp4a.s32.s32 %r99, %r212, %r93, %r210;


dp4a.s32.s32 %r103, %r212, %r97, %r99;

shr.u16 %rs6, %rs4, 8;
cvt.u32.u16 %r257, %rs6;
mul.lo.s32 %r258, %r95, %r257;
cvt.rn.f32.s32 %f45, %r258;
fma.rn.ftz.f32 %f46, %f26, %f45, %f42;
shr.u16 %rs7, %rs5, 8;
cvt.u32.u16 %r259, %rs7;
mul.lo.s32 %r260, %r103, %r259;
cvt.rn.f32.s32 %f47, %r260;
fma.rn.ftz.f32 %f48, %f26, %f47, %f44;
ld.global.nc.u32 %r107, [%rd43];

{.reg .f16 low,high;
mov.b32 {low,high},%r107;
cvt.f32.f16 %f27, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r107;
cvt.f32.f16 %f28, high;}


mul.ftz.f32 %f49, %f46, %f27;
mul.ftz.f32 %f50, %f48, %f28;
sub.ftz.f32 %f51, %f49, %f50;
add.ftz.f32 %f152, %f51, %f152;
add.s64 %rd44, %rd32, %rd7;
ld.global.nc.u32 %r261, [%rd44];
ld.global.nc.u32 %r262, [%rd44+16];
add.s64 %rd45, %rd34, %rd7;
ld.global.nc.u32 %r263, [%rd45];
shr.s32 %r264, %r263, %r14;
ld.global.nc.u32 %r265, [%rd45+16];
shr.s32 %r266, %r265, %r14;
add.s64 %rd46, %rd36, %rd7;
ld.global.nc.u16 %rs8, [%rd46];
cvt.u32.u16 %r267, %rs8;
ld.global.nc.u16 %rs9, [%rd46+4];
cvt.u32.u16 %r268, %rs9;
ld.global.nc.u16 %rs10, [%rd46+8];
cvt.u32.u16 %r269, %rs10;
and.b32 %r270, %r267, 16191;
and.b32 %r271, %r269, 3855;
and.b32 %r272, %r267, 49344;
shr.u32 %r273, %r272, 2;
or.b32 %r274, %r271, %r273;
and.b32 %r275, %r268, 16191;
shr.u32 %r276, %r269, 4;
and.b32 %r277, %r276, 3855;
and.b32 %r278, %r268, 49344;
shr.u32 %r279, %r278, 2;
or.b32 %r280, %r277, %r279;

{.reg .f16 low,high;
mov.b32 {low,high},%r73;
cvt.f32.f16 %f29, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r74;
cvt.f32.f16 %f30, low;}


and.b32 %r281, %r261, 252645135;
and.b32 %r282, %r262, 252645135;
shl.b32 %r283, %r264, 4;
and.b32 %r284, %r283, 269488144;
shl.b32 %r285, %r266, 4;
and.b32 %r286, %r285, 269488144;
or.b32 %r116, %r281, %r284;
or.b32 %r112, %r282, %r286;

dp4a.s32.s32 %r111, %r112, %r77, %r210;


dp4a.s32.s32 %r115, %r116, %r81, %r111;


dp4a.s32.s32 %r119, %r212, %r77, %r210;


dp4a.s32.s32 %r123, %r212, %r81, %r119;

selp.b32 %r287, %r274, %r270, %p3;
cvt.u16.u32 %rs11, %r287;
and.b32 %r288, %r287, 63;
mul.lo.s32 %r289, %r115, %r288;
cvt.rn.f32.s32 %f52, %r289;
fma.rn.ftz.f32 %f53, %f29, %f52, 0f00000000;
selp.b32 %r290, %r280, %r275, %p3;
cvt.u16.u32 %rs12, %r290;
and.b32 %r291, %r290, 63;
mul.lo.s32 %r292, %r123, %r291;
cvt.rn.f32.s32 %f54, %r292;
fma.rn.ftz.f32 %f55, %f29, %f54, 0f00000000;
shr.u32 %r293, %r261, 4;
and.b32 %r294, %r293, 252645135;
shr.u32 %r295, %r262, 4;
and.b32 %r296, %r295, 252645135;
shl.b32 %r297, %r264, 3;
and.b32 %r298, %r297, 269488144;
shl.b32 %r299, %r266, 3;
and.b32 %r300, %r299, 269488144;
or.b32 %r132, %r294, %r298;
or.b32 %r128, %r296, %r300;

dp4a.s32.s32 %r127, %r128, %r93, %r210;


dp4a.s32.s32 %r131, %r132, %r97, %r127;


dp4a.s32.s32 %r135, %r212, %r93, %r210;


dp4a.s32.s32 %r139, %r212, %r97, %r135;

shr.u16 %rs13, %rs11, 8;
cvt.u32.u16 %r301, %rs13;
mul.lo.s32 %r302, %r131, %r301;
cvt.rn.f32.s32 %f56, %r302;
fma.rn.ftz.f32 %f57, %f30, %f56, %f53;
shr.u16 %rs14, %rs12, 8;
cvt.u32.u16 %r303, %rs14;
mul.lo.s32 %r304, %r139, %r303;
cvt.rn.f32.s32 %f58, %r304;
fma.rn.ftz.f32 %f59, %f30, %f58, %f55;
add.s64 %rd47, %rd43, %rd7;
ld.global.nc.u32 %r143, [%rd47];

{.reg .f16 low,high;
mov.b32 {low,high},%r143;
cvt.f32.f16 %f31, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r143;
cvt.f32.f16 %f32, high;}


mul.ftz.f32 %f60, %f57, %f31;
mul.ftz.f32 %f61, %f59, %f32;
sub.ftz.f32 %f62, %f60, %f61;
add.ftz.f32 %f151, %f62, %f151;
add.s64 %rd48, %rd44, %rd7;
ld.global.nc.u32 %r305, [%rd48];
ld.global.nc.u32 %r306, [%rd48+16];
add.s64 %rd49, %rd45, %rd7;
ld.global.nc.u32 %r307, [%rd49];
shr.s32 %r308, %r307, %r14;
ld.global.nc.u32 %r309, [%rd49+16];
shr.s32 %r310, %r309, %r14;
add.s64 %rd50, %rd46, %rd7;
ld.global.nc.u16 %rs15, [%rd50];
cvt.u32.u16 %r311, %rs15;
ld.global.nc.u16 %rs16, [%rd50+4];
cvt.u32.u16 %r312, %rs16;
ld.global.nc.u16 %rs17, [%rd50+8];
cvt.u32.u16 %r313, %rs17;
and.b32 %r314, %r311, 16191;
and.b32 %r315, %r313, 3855;
and.b32 %r316, %r311, 49344;
shr.u32 %r317, %r316, 2;
or.b32 %r318, %r315, %r317;
and.b32 %r319, %r312, 16191;
shr.u32 %r320, %r313, 4;
and.b32 %r321, %r320, 3855;
and.b32 %r322, %r312, 49344;
shr.u32 %r323, %r322, 2;
or.b32 %r324, %r321, %r323;

{.reg .f16 low,high;
mov.b32 {low,high},%r73;
cvt.f32.f16 %f33, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r74;
cvt.f32.f16 %f34, low;}


and.b32 %r325, %r305, 252645135;
and.b32 %r326, %r306, 252645135;
shl.b32 %r327, %r308, 4;
and.b32 %r328, %r327, 269488144;
shl.b32 %r329, %r310, 4;
and.b32 %r330, %r329, 269488144;
or.b32 %r152, %r325, %r328;
or.b32 %r148, %r326, %r330;

dp4a.s32.s32 %r147, %r148, %r77, %r210;


dp4a.s32.s32 %r151, %r152, %r81, %r147;


dp4a.s32.s32 %r155, %r212, %r77, %r210;


dp4a.s32.s32 %r159, %r212, %r81, %r155;

selp.b32 %r331, %r318, %r314, %p3;
cvt.u16.u32 %rs18, %r331;
and.b32 %r332, %r331, 63;
mul.lo.s32 %r333, %r151, %r332;
cvt.rn.f32.s32 %f63, %r333;
fma.rn.ftz.f32 %f64, %f33, %f63, 0f00000000;
selp.b32 %r334, %r324, %r319, %p3;
cvt.u16.u32 %rs19, %r334;
and.b32 %r335, %r334, 63;
mul.lo.s32 %r336, %r159, %r335;
cvt.rn.f32.s32 %f65, %r336;
fma.rn.ftz.f32 %f66, %f33, %f65, 0f00000000;
shr.u32 %r337, %r305, 4;
and.b32 %r338, %r337, 252645135;
shr.u32 %r339, %r306, 4;
and.b32 %r340, %r339, 252645135;
shl.b32 %r341, %r308, 3;
and.b32 %r342, %r341, 269488144;
shl.b32 %r343, %r310, 3;
and.b32 %r344, %r343, 269488144;
or.b32 %r168, %r338, %r342;
or.b32 %r164, %r340, %r344;

dp4a.s32.s32 %r163, %r164, %r93, %r210;


dp4a.s32.s32 %r167, %r168, %r97, %r163;


dp4a.s32.s32 %r171, %r212, %r93, %r210;


dp4a.s32.s32 %r175, %r212, %r97, %r171;

shr.u16 %rs20, %rs18, 8;
cvt.u32.u16 %r345, %rs20;
mul.lo.s32 %r346, %r167, %r345;
cvt.rn.f32.s32 %f67, %r346;
fma.rn.ftz.f32 %f68, %f34, %f67, %f64;
shr.u16 %rs21, %rs19, 8;
cvt.u32.u16 %r347, %rs21;
mul.lo.s32 %r348, %r175, %r347;
cvt.rn.f32.s32 %f69, %r348;
fma.rn.ftz.f32 %f70, %f34, %f69, %f66;
add.s64 %rd51, %rd47, %rd7;
ld.global.nc.u32 %r179, [%rd51];

{.reg .f16 low,high;
mov.b32 {low,high},%r179;
cvt.f32.f16 %f35, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r179;
cvt.f32.f16 %f36, high;}


mul.ftz.f32 %f71, %f68, %f35;
mul.ftz.f32 %f72, %f70, %f36;
sub.ftz.f32 %f73, %f71, %f72;
add.ftz.f32 %f150, %f73, %f150;
add.s64 %rd52, %rd48, %rd7;
ld.global.nc.u32 %r349, [%rd52];
ld.global.nc.u32 %r350, [%rd52+16];
add.s64 %rd53, %rd49, %rd7;
ld.global.nc.u32 %r351, [%rd53];
shr.s32 %r352, %r351, %r14;
ld.global.nc.u32 %r353, [%rd53+16];
shr.s32 %r354, %r353, %r14;
add.s64 %rd54, %rd50, %rd7;
ld.global.nc.u16 %rs22, [%rd54];
cvt.u32.u16 %r355, %rs22;
ld.global.nc.u16 %rs23, [%rd54+4];
cvt.u32.u16 %r356, %rs23;
ld.global.nc.u16 %rs24, [%rd54+8];
cvt.u32.u16 %r357, %rs24;
and.b32 %r358, %r355, 16191;
and.b32 %r359, %r357, 3855;
and.b32 %r360, %r355, 49344;
shr.u32 %r361, %r360, 2;
or.b32 %r362, %r359, %r361;
and.b32 %r363, %r356, 16191;
shr.u32 %r364, %r357, 4;
and.b32 %r365, %r364, 3855;
and.b32 %r366, %r356, 49344;
shr.u32 %r367, %r366, 2;
or.b32 %r368, %r365, %r367;

{.reg .f16 low,high;
mov.b32 {low,high},%r73;
cvt.f32.f16 %f37, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r74;
cvt.f32.f16 %f38, low;}


and.b32 %r369, %r349, 252645135;
and.b32 %r370, %r350, 252645135;
shl.b32 %r371, %r352, 4;
and.b32 %r372, %r371, 269488144;
shl.b32 %r373, %r354, 4;
and.b32 %r374, %r373, 269488144;
or.b32 %r188, %r369, %r372;
or.b32 %r184, %r370, %r374;

dp4a.s32.s32 %r183, %r184, %r77, %r210;


dp4a.s32.s32 %r187, %r188, %r81, %r183;


dp4a.s32.s32 %r191, %r212, %r77, %r210;


dp4a.s32.s32 %r195, %r212, %r81, %r191;

selp.b32 %r375, %r362, %r358, %p3;
cvt.u16.u32 %rs25, %r375;
and.b32 %r376, %r375, 63;
mul.lo.s32 %r377, %r187, %r376;
cvt.rn.f32.s32 %f74, %r377;
fma.rn.ftz.f32 %f75, %f37, %f74, 0f00000000;
selp.b32 %r378, %r368, %r363, %p3;
cvt.u16.u32 %rs26, %r378;
and.b32 %r379, %r378, 63;
mul.lo.s32 %r380, %r195, %r379;
cvt.rn.f32.s32 %f76, %r380;
fma.rn.ftz.f32 %f77, %f37, %f76, 0f00000000;
shr.u32 %r381, %r349, 4;
and.b32 %r382, %r381, 252645135;
shr.u32 %r383, %r350, 4;
and.b32 %r384, %r383, 252645135;
shl.b32 %r385, %r352, 3;
and.b32 %r386, %r385, 269488144;
shl.b32 %r387, %r354, 3;
and.b32 %r388, %r387, 269488144;
or.b32 %r204, %r382, %r386;
or.b32 %r200, %r384, %r388;

dp4a.s32.s32 %r199, %r200, %r93, %r210;


dp4a.s32.s32 %r203, %r204, %r97, %r199;


dp4a.s32.s32 %r207, %r212, %r93, %r210;


dp4a.s32.s32 %r211, %r212, %r97, %r207;

shr.u16 %rs27, %rs25, 8;
cvt.u32.u16 %r389, %rs27;
mul.lo.s32 %r390, %r203, %r389;
cvt.rn.f32.s32 %f78, %r390;
fma.rn.ftz.f32 %f79, %f38, %f78, %f75;
shr.u16 %rs28, %rs26, 8;
cvt.u32.u16 %r391, %rs28;
mul.lo.s32 %r392, %r211, %r391;
cvt.rn.f32.s32 %f80, %r392;
fma.rn.ftz.f32 %f81, %f38, %f80, %f77;
add.s64 %rd55, %rd51, %rd7;
ld.global.nc.u32 %r215, [%rd55];

{.reg .f16 low,high;
mov.b32 {low,high},%r215;
cvt.f32.f16 %f39, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r215;
cvt.f32.f16 %f40, high;}


mul.ftz.f32 %f82, %f79, %f39;
mul.ftz.f32 %f83, %f81, %f40;
sub.ftz.f32 %f84, %f82, %f83;
add.ftz.f32 %f149, %f84, %f149;
add.s32 %r477, %r477, 64;
add.s64 %rd58, %rd58, 1408;
add.s32 %r478, %r478, 8;
setp.lt.s32 %p4, %r478, %r5;
@%p4 bra $L__BB171_5;

$L__BB171_6:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB171_8;

shl.b32 %r393, %r1, 9;
mov.u32 %r394, _ZZ13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r395, %r394, %r393;
shl.b32 %r396, %r2, 2;
add.s32 %r397, %r395, %r396;
st.shared.f32 [%r397+-512], %f152;
st.shared.f32 [%r397+-384], %f151;
st.shared.f32 [%r397+-256], %f150;
st.shared.f32 [%r397+-128], %f149;

$L__BB171_8:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB171_17;

mov.u32 %r398, %ctaid.z;
mov.u32 %r399, %ctaid.y;
mad.lo.s32 %r400, %r399, %r32, %r4;
mad.lo.s32 %r401, %r398, %r38, %r400;
shl.b32 %r402, %r2, 2;
mov.u32 %r403, _ZZ13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r20, %r403, %r402;
mov.u32 %r404, 2;
ld.shared.f32 %f85, [%r20];
add.ftz.f32 %f86, %f85, %f152;
ld.shared.f32 %f87, [%r20+512];
add.ftz.f32 %f88, %f87, %f86;
ld.shared.f32 %f89, [%r20+1024];
add.ftz.f32 %f90, %f89, %f88;
mov.b32 %r405, %f90;
mov.u32 %r406, 31;
mov.u32 %r407, 16;
mov.u32 %r408, -1;
shfl.sync.bfly.b32 %r409|%p7, %r405, %r407, %r406, %r408;
mov.b32 %f91, %r409;
add.ftz.f32 %f92, %f90, %f91;
mov.b32 %r410, %f92;
mov.u32 %r411, 8;
shfl.sync.bfly.b32 %r412|%p8, %r410, %r411, %r406, %r408;
mov.b32 %f93, %r412;
add.ftz.f32 %f94, %f92, %f93;
mov.b32 %r413, %f94;
mov.u32 %r414, 4;
shfl.sync.bfly.b32 %r415|%p9, %r413, %r414, %r406, %r408;
mov.b32 %f95, %r415;
add.ftz.f32 %f96, %f94, %f95;
mov.b32 %r416, %f96;
shfl.sync.bfly.b32 %r417|%p10, %r416, %r404, %r406, %r408;
mov.b32 %f97, %r417;
add.ftz.f32 %f98, %f96, %f97;
mov.b32 %r418, %f98;
mov.u32 %r419, 1;
shfl.sync.bfly.b32 %r420|%p11, %r418, %r419, %r406, %r408;
mov.b32 %f99, %r420;
add.ftz.f32 %f13, %f98, %f99;
setp.ge.u32 %p12, %r4, %r26;
setp.ne.s32 %p13, %r2, 0;
cvta.to.global.u64 %rd56, %rd16;
mul.wide.u32 %rd57, %r401, 4;
add.s64 %rd12, %rd56, %rd57;
or.pred %p14, %p13, %p12;
@%p14 bra $L__BB171_11;

st.global.f32 [%rd12], %f13;

$L__BB171_11:
ld.shared.f32 %f100, [%r20+128];
add.ftz.f32 %f101, %f100, %f151;
ld.shared.f32 %f102, [%r20+640];
add.ftz.f32 %f103, %f102, %f101;
ld.shared.f32 %f104, [%r20+1152];
add.ftz.f32 %f105, %f104, %f103;
mov.b32 %r421, %f105;
shfl.sync.bfly.b32 %r425|%p15, %r421, %r407, %r406, %r408;
mov.b32 %f106, %r425;
add.ftz.f32 %f107, %f105, %f106;
mov.b32 %r426, %f107;
shfl.sync.bfly.b32 %r428|%p16, %r426, %r411, %r406, %r408;
mov.b32 %f108, %r428;
add.ftz.f32 %f109, %f107, %f108;
mov.b32 %r429, %f109;
shfl.sync.bfly.b32 %r431|%p17, %r429, %r414, %r406, %r408;
mov.b32 %f110, %r431;
add.ftz.f32 %f111, %f109, %f110;
mov.b32 %r432, %f111;
shfl.sync.bfly.b32 %r434|%p18, %r432, %r404, %r406, %r408;
mov.b32 %f112, %r434;
add.ftz.f32 %f113, %f111, %f112;
mov.b32 %r435, %f113;
shfl.sync.bfly.b32 %r437|%p19, %r435, %r419, %r406, %r408;
mov.b32 %f114, %r437;
add.ftz.f32 %f14, %f113, %f114;
add.s32 %r438, %r4, 1;
setp.ge.u32 %p20, %r438, %r26;
setp.ne.s32 %p21, %r2, 1;
or.pred %p22, %p21, %p20;
@%p22 bra $L__BB171_13;

st.global.f32 [%rd12+4], %f14;

$L__BB171_13:
ld.shared.f32 %f115, [%r20+256];
add.ftz.f32 %f116, %f115, %f150;
ld.shared.f32 %f117, [%r20+768];
add.ftz.f32 %f118, %f117, %f116;
ld.shared.f32 %f119, [%r20+1280];
add.ftz.f32 %f120, %f119, %f118;
mov.b32 %r439, %f120;
mov.u32 %r440, 31;
mov.u32 %r441, 16;
mov.u32 %r442, -1;
shfl.sync.bfly.b32 %r443|%p23, %r439, %r441, %r440, %r442;
mov.b32 %f121, %r443;
add.ftz.f32 %f122, %f120, %f121;
mov.b32 %r444, %f122;
mov.u32 %r445, 8;
shfl.sync.bfly.b32 %r446|%p24, %r444, %r445, %r440, %r442;
mov.b32 %f123, %r446;
add.ftz.f32 %f124, %f122, %f123;
mov.b32 %r447, %f124;
mov.u32 %r448, 4;
shfl.sync.bfly.b32 %r449|%p25, %r447, %r448, %r440, %r442;
mov.b32 %f125, %r449;
add.ftz.f32 %f126, %f124, %f125;
mov.b32 %r450, %f126;
mov.u32 %r451, 2;
shfl.sync.bfly.b32 %r452|%p26, %r450, %r451, %r440, %r442;
mov.b32 %f127, %r452;
add.ftz.f32 %f128, %f126, %f127;
mov.b32 %r453, %f128;
mov.u32 %r454, 1;
shfl.sync.bfly.b32 %r455|%p27, %r453, %r454, %r440, %r442;
mov.b32 %f129, %r455;
add.ftz.f32 %f15, %f128, %f129;
add.s32 %r456, %r4, 2;
setp.ge.u32 %p28, %r456, %r26;
setp.ne.s32 %p29, %r2, 2;
or.pred %p30, %p29, %p28;
@%p30 bra $L__BB171_15;

st.global.f32 [%rd12+8], %f15;

$L__BB171_15:
ld.shared.f32 %f130, [%r20+384];
add.ftz.f32 %f131, %f130, %f149;
ld.shared.f32 %f132, [%r20+896];
add.ftz.f32 %f133, %f132, %f131;
ld.shared.f32 %f134, [%r20+1408];
add.ftz.f32 %f135, %f134, %f133;
mov.b32 %r457, %f135;
shfl.sync.bfly.b32 %r461|%p31, %r457, %r441, %r440, %r442;
mov.b32 %f136, %r461;
add.ftz.f32 %f137, %f135, %f136;
mov.b32 %r462, %f137;
shfl.sync.bfly.b32 %r464|%p32, %r462, %r445, %r440, %r442;
mov.b32 %f138, %r464;
add.ftz.f32 %f139, %f137, %f138;
mov.b32 %r465, %f139;
shfl.sync.bfly.b32 %r467|%p33, %r465, %r448, %r440, %r442;
mov.b32 %f140, %r467;
add.ftz.f32 %f141, %f139, %f140;
mov.b32 %r468, %f141;
shfl.sync.bfly.b32 %r470|%p34, %r468, %r451, %r440, %r442;
mov.b32 %f142, %r470;
add.ftz.f32 %f143, %f141, %f142;
mov.b32 %r471, %f143;
shfl.sync.bfly.b32 %r473|%p35, %r471, %r454, %r440, %r442;
mov.b32 %f144, %r473;
add.ftz.f32 %f16, %f143, %f144;
add.s32 %r474, %r4, 3;
setp.ge.u32 %p36, %r474, %r26;
setp.ne.s32 %p37, %r2, 3;
or.pred %p38, %p37, %p36;
@%p38 bra $L__BB171_17;

st.global.f32 [%rd12+12], %f16;

$L__BB171_17:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<35>;
.reg .b16 %rs<12>;
.reg .f32 %f<142>;
.reg .b32 %r<318>;
.reg .b64 %rd<79>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[384];

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate[384];

ld.param.u64 %rd8, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd9, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd10, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r46, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r47, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r48, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
ld.param.f32 %f31, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
ld.param.u64 %rd13, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+16];
ld.param.u64 %rd12, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+8];
ld.param.u64 %rd11, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3];
mov.u32 %r1, %tid.y;
shl.b32 %r49, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r49, %r2;
mov.u32 %r4, %ctaid.x;
mov.u32 %r5, %ctaid.y;
setp.eq.s64 %p2, %rd10, 0;
@%p2 bra $L__BB172_2;

cvta.to.global.u64 %rd17, %rd10;
mul.wide.u32 %rd18, %r5, 4;
add.s64 %rd19, %rd17, %rd18;
ld.global.nc.u32 %r314, [%rd19];
mul.hi.u32 %r50, %r5, %r33;
add.s32 %r51, %r50, %r5;
shr.u32 %r52, %r51, %r34;
mul.lo.s32 %r53, %r52, %r35;
sub.s32 %r315, %r5, %r53;
bra.uni $L__BB172_3;

$L__BB172_2:
mul.hi.u32 %r54, %r5, %r37;
add.s32 %r55, %r54, %r5;
shr.u32 %r314, %r55, %r38;
mov.u32 %r315, %r5;

$L__BB172_3:
mov.u32 %r11, %ctaid.z;
selp.b32 %r12, %r5, %r314, %p2;
or.b32 %r56, %r2, %r1;
setp.ne.s32 %p4, %r56, 0;
mov.f32 %f136, 0f00000000;
mov.f32 %f131, %f136;
mov.f32 %f132, %f136;
@%p4 bra $L__BB172_8;

setp.eq.s64 %p5, %rd11, 0;
mov.f32 %f132, 0f00000000;
mov.f32 %f131, %f132;
@%p5 bra $L__BB172_6;

cvta.to.global.u64 %rd20, %rd11;
mul.lo.s32 %r57, %r11, %r48;
cvt.u64.u32 %rd21, %r57;
mul.lo.s32 %r58, %r12, %r42;
cvt.u64.u32 %rd22, %r58;
cvt.u64.u32 %rd23, %r2;
cvt.s64.s32 %rd24, %r4;
add.s64 %rd25, %rd23, %rd24;
add.s64 %rd26, %rd25, %rd21;
add.s64 %rd27, %rd26, %rd22;
shl.b64 %rd28, %rd27, 2;
add.s64 %rd29, %rd20, %rd28;
ld.global.f32 %f131, [%rd29];

$L__BB172_6:
setp.eq.s64 %p6, %rd13, 0;
setp.eq.s64 %p7, %rd12, 0;
or.pred %p8, %p6, %p7;
@%p8 bra $L__BB172_8;

cvta.to.global.u64 %rd30, %rd13;
mul.lo.s32 %r60, %r11, %r48;
cvt.u64.u32 %rd31, %r60;
mul.lo.s32 %r61, %r12, %r42;
cvt.u64.u32 %rd32, %r61;
cvt.u64.u32 %rd33, %r2;
cvt.s64.s32 %rd34, %r4;
add.s64 %rd35, %rd33, %rd34;
add.s64 %rd36, %rd35, %rd31;
add.s64 %rd37, %rd36, %rd32;
shl.b64 %rd38, %rd37, 2;
add.s64 %rd39, %rd30, %rd38;
ld.global.f32 %f132, [%rd39];

$L__BB172_8:
shr.u32 %r63, %r32, 8;
shr.s32 %r64, %r3, 31;
shr.u32 %r65, %r64, 28;
add.s32 %r66, %r3, %r65;
shr.s32 %r317, %r66, 4;
setp.ge.s32 %p9, %r317, %r63;
mov.f32 %f135, %f136;
@%p9 bra $L__BB172_13;

and.b32 %r70, %r66, -16;
sub.s32 %r14, %r3, %r70;
shr.s32 %r71, %r14, 31;
shr.u32 %r72, %r71, 30;
add.s32 %r73, %r14, %r72;
shr.s32 %r15, %r73, 2;
and.b32 %r74, %r73, -4;
sub.s32 %r16, %r14, %r74;
shl.b32 %r316, %r317, 3;
mul.hi.u32 %r75, %r11, %r43;
add.s32 %r76, %r75, %r11;
shr.u32 %r77, %r76, %r44;
mad.lo.s32 %r78, %r77, %r46, %r317;
mad.lo.s32 %r79, %r314, %r40, %r78;
mad.lo.s32 %r80, %r4, %r36, %r79;
mul.wide.s32 %rd78, %r80, 176;
mul.lo.s32 %r81, %r315, %r41;
cvt.u64.u32 %rd40, %r81;
mul.lo.s32 %r82, %r11, %r47;
cvt.u64.u32 %rd41, %r82;
add.s64 %rd5, %rd40, %rd41;
mov.f32 %f135, 0f00000000;
cvta.to.global.u64 %rd43, %rd9;
cvta.to.global.u64 %rd47, %rd8;
mul.wide.s32 %rd60, %r16, 4;
mov.f32 %f136, %f135;

$L__BB172_10:
shl.b32 %r119, %r15, 1;
cvt.s64.s32 %rd42, %r119;
cvt.u64.u32 %rd44, %r316;
add.s64 %rd45, %rd5, %rd44;
shl.b32 %r120, %r16, 2;
shl.b32 %r121, %r15, 5;
add.s32 %r122, %r121, %r120;
cvt.s64.s32 %rd46, %r122;
add.s64 %rd48, %rd47, %rd46;
add.s64 %rd49, %rd48, %rd78;
cvt.s64.s32 %rd50, %r120;
add.s64 %rd51, %rd47, %rd50;
add.s64 %rd52, %rd51, %rd78;
ld.global.nc.u32 %r123, [%rd49+48];
ld.global.nc.u32 %r124, [%rd49+64];
ld.global.nc.u32 %r125, [%rd52+16];
shr.s32 %r126, %r125, %r119;
ld.global.nc.u32 %r127, [%rd52+32];
shr.s32 %r128, %r127, %r119;
cvt.u64.u32 %rd53, %r119;
and.b64 %rd54, %rd53, 2;
add.s64 %rd55, %rd47, %rd54;
add.s64 %rd56, %rd55, %rd78;
ld.global.nc.u16 %rs1, [%rd56+4];
cvt.u32.u16 %r129, %rs1;
ld.global.nc.u16 %rs2, [%rd56+8];
cvt.u32.u16 %r130, %rs2;
ld.global.nc.u16 %rs3, [%rd56+12];
cvt.u32.u16 %r131, %rs3;
and.b32 %r132, %r129, 16191;
and.b32 %r133, %r131, 3855;
and.b32 %r134, %r129, 49344;
shr.u32 %r135, %r134, 2;
or.b32 %r136, %r133, %r135;
setp.gt.s32 %p10, %r14, 7;
and.b32 %r137, %r130, 16191;
shr.u32 %r138, %r131, 4;
and.b32 %r139, %r138, 3855;
and.b32 %r140, %r130, 49344;
shr.u32 %r141, %r140, 2;
or.b32 %r142, %r139, %r141;
add.s64 %rd57, %rd45, %rd42;
mul.lo.s64 %rd58, %rd57, 36;
add.s64 %rd59, %rd43, %rd58;
ld.global.nc.u32 %r20, [%rd59];

{.reg .f16 low,high;
mov.b32 {low,high},%r20;
cvt.f32.f16 %f40, low;}


add.s64 %rd61, %rd59, %rd60;
ld.global.nc.u32 %r21, [%rd61+4];
ld.global.nc.u32 %r22, [%rd61+20];
ld.global.nc.u32 %r23, [%rd59+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r23;
cvt.f32.f16 %f41, low;}


ld.global.nc.u32 %r24, [%rd61+40];
ld.global.nc.u32 %r25, [%rd61+56];
add.s64 %rd62, %rd47, %rd78;
and.b32 %r143, %r123, 252645135;
and.b32 %r144, %r124, 252645135;
shl.b32 %r145, %r126, 4;
and.b32 %r146, %r145, 269488144;
shl.b32 %r147, %r128, 4;
and.b32 %r148, %r147, 269488144;
or.b32 %r90, %r143, %r146;
or.b32 %r86, %r144, %r148;
mov.u32 %r112, 0;

dp4a.s32.s32 %r85, %r86, %r22, %r112;


dp4a.s32.s32 %r89, %r90, %r21, %r85;

mov.u32 %r114, 16843009;

dp4a.s32.s32 %r93, %r114, %r22, %r112;


dp4a.s32.s32 %r97, %r114, %r21, %r93;

selp.b32 %r149, %r136, %r132, %p10;
cvt.u16.u32 %rs4, %r149;
and.b32 %r150, %r149, 63;
mul.lo.s32 %r151, %r89, %r150;
cvt.rn.f32.s32 %f44, %r151;
fma.rn.ftz.f32 %f45, %f40, %f44, 0f00000000;
selp.b32 %r152, %r142, %r137, %p10;
cvt.u16.u32 %rs5, %r152;
and.b32 %r153, %r152, 63;
mul.lo.s32 %r154, %r97, %r153;
cvt.rn.f32.s32 %f46, %r154;
fma.rn.ftz.f32 %f47, %f40, %f46, 0f00000000;
shr.u32 %r155, %r123, 4;
and.b32 %r156, %r155, 252645135;
shr.u32 %r157, %r124, 4;
and.b32 %r158, %r157, 252645135;
shl.b32 %r159, %r126, 3;
and.b32 %r160, %r159, 269488144;
shl.b32 %r161, %r128, 3;
and.b32 %r162, %r161, 269488144;
or.b32 %r106, %r156, %r160;
or.b32 %r102, %r158, %r162;

dp4a.s32.s32 %r101, %r102, %r25, %r112;


dp4a.s32.s32 %r105, %r106, %r24, %r101;


dp4a.s32.s32 %r109, %r114, %r25, %r112;


dp4a.s32.s32 %r113, %r114, %r24, %r109;

shr.u16 %rs6, %rs4, 8;
cvt.u32.u16 %r163, %rs6;
mul.lo.s32 %r164, %r105, %r163;
cvt.rn.f32.s32 %f48, %r164;
fma.rn.ftz.f32 %f49, %f41, %f48, %f45;
shr.u16 %rs7, %rs5, 8;
cvt.u32.u16 %r165, %rs7;
mul.lo.s32 %r166, %r113, %r165;
cvt.rn.f32.s32 %f50, %r166;
fma.rn.ftz.f32 %f51, %f41, %f50, %f47;
ld.global.nc.u32 %r117, [%rd62];

{.reg .f16 low,high;
mov.b32 {low,high},%r117;
cvt.f32.f16 %f42, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r117;
cvt.f32.f16 %f43, high;}


mul.ftz.f32 %f52, %f49, %f42;
mul.ftz.f32 %f53, %f51, %f43;
sub.ftz.f32 %f54, %f52, %f53;
add.ftz.f32 %f136, %f136, %f54;
setp.eq.s64 %p11, %rd12, 0;
@%p11 bra $L__BB172_12;

cvta.to.global.u64 %rd64, %rd12;
add.s64 %rd65, %rd64, %rd46;
add.s64 %rd66, %rd65, %rd78;
add.s64 %rd68, %rd64, %rd50;
add.s64 %rd69, %rd68, %rd78;
ld.global.u32 %r207, [%rd66+48];
ld.global.u32 %r208, [%rd66+64];
ld.global.u32 %r209, [%rd69+16];
shr.s32 %r210, %r209, %r119;
ld.global.u32 %r211, [%rd69+32];
shr.s32 %r212, %r211, %r119;
add.s64 %rd72, %rd64, %rd54;
add.s64 %rd73, %rd72, %rd78;
ld.global.u16 %r213, [%rd73+4];
ld.global.u16 %r214, [%rd73+8];
ld.global.u16 %r215, [%rd73+12];
and.b32 %r216, %r213, 16191;
and.b32 %r217, %r215, 3855;
and.b32 %r218, %r213, 49344;
shr.u32 %r219, %r218, 2;
or.b32 %r220, %r217, %r219;
and.b32 %r221, %r214, 16191;
shr.u32 %r222, %r215, 4;
and.b32 %r223, %r222, 3855;
and.b32 %r224, %r214, 49344;
shr.u32 %r225, %r224, 2;
or.b32 %r226, %r223, %r225;

{.reg .f16 low,high;
mov.b32 {low,high},%r20;
cvt.f32.f16 %f55, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r23;
cvt.f32.f16 %f56, low;}


add.s64 %rd74, %rd64, %rd78;
and.b32 %r227, %r207, 252645135;
and.b32 %r228, %r208, 252645135;
shl.b32 %r229, %r210, 4;
and.b32 %r230, %r229, 269488144;
shl.b32 %r231, %r212, 4;
and.b32 %r232, %r231, 269488144;
or.b32 %r174, %r227, %r230;
or.b32 %r170, %r228, %r232;

dp4a.s32.s32 %r169, %r170, %r22, %r112;


dp4a.s32.s32 %r173, %r174, %r21, %r169;


dp4a.s32.s32 %r177, %r114, %r22, %r112;


dp4a.s32.s32 %r181, %r114, %r21, %r177;

selp.b32 %r233, %r220, %r216, %p10;
cvt.u16.u32 %rs8, %r233;
and.b32 %r234, %r233, 63;
mul.lo.s32 %r235, %r173, %r234;
cvt.rn.f32.s32 %f59, %r235;
fma.rn.ftz.f32 %f60, %f55, %f59, 0f00000000;
selp.b32 %r236, %r226, %r221, %p10;
cvt.u16.u32 %rs9, %r236;
and.b32 %r237, %r236, 63;
mul.lo.s32 %r238, %r181, %r237;
cvt.rn.f32.s32 %f61, %r238;
fma.rn.ftz.f32 %f62, %f55, %f61, 0f00000000;
shr.u32 %r239, %r207, 4;
and.b32 %r240, %r239, 252645135;
shr.u32 %r241, %r208, 4;
and.b32 %r242, %r241, 252645135;
shl.b32 %r243, %r210, 3;
and.b32 %r244, %r243, 269488144;
shl.b32 %r245, %r212, 3;
and.b32 %r246, %r245, 269488144;
or.b32 %r190, %r240, %r244;
or.b32 %r186, %r242, %r246;

dp4a.s32.s32 %r185, %r186, %r25, %r112;


dp4a.s32.s32 %r189, %r190, %r24, %r185;


dp4a.s32.s32 %r193, %r114, %r25, %r112;


dp4a.s32.s32 %r197, %r114, %r24, %r193;

shr.u16 %rs10, %rs8, 8;
cvt.u32.u16 %r247, %rs10;
mul.lo.s32 %r248, %r189, %r247;
cvt.rn.f32.s32 %f63, %r248;
fma.rn.ftz.f32 %f64, %f56, %f63, %f60;
shr.u16 %rs11, %rs9, 8;
cvt.u32.u16 %r249, %rs11;
mul.lo.s32 %r250, %r197, %r249;
cvt.rn.f32.s32 %f65, %r250;
fma.rn.ftz.f32 %f66, %f56, %f65, %f62;
ld.global.u32 %r202, [%rd74];

{.reg .f16 low,high;
mov.b32 {low,high},%r202;
cvt.f32.f16 %f57, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r202;
cvt.f32.f16 %f58, high;}


mul.ftz.f32 %f67, %f64, %f57;
mul.ftz.f32 %f68, %f66, %f58;
sub.ftz.f32 %f69, %f67, %f68;
add.ftz.f32 %f135, %f135, %f69;

$L__BB172_12:
add.s32 %r316, %r316, 64;
add.s64 %rd78, %rd78, 1408;
add.s32 %r317, %r317, 8;
setp.lt.s32 %p13, %r317, %r63;
@%p13 bra $L__BB172_10;

$L__BB172_13:
setp.eq.s32 %p14, %r1, 0;
@%p14 bra $L__BB172_16;

setp.eq.s64 %p15, %rd12, 0;
add.s32 %r28, %r1, -1;
shl.b32 %r255, %r1, 7;
mov.u32 %r256, _ZZ13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r257, %r256, %r255;
shl.b32 %r258, %r2, 2;
add.s32 %r259, %r257, %r258;
st.shared.f32 [%r259+-128], %f136;
@%p15 bra $L__BB172_16;

shl.b32 %r261, %r28, 7;
mov.u32 %r262, _ZZ13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r263, %r262, %r261;
add.s32 %r265, %r263, %r258;
st.shared.f32 [%r265], %f135;

$L__BB172_16:
setp.ne.s32 %p16, %r1, 0;
bar.sync 0;
@%p16 bra $L__BB172_35;

setp.ne.s64 %p17, %rd12, 0;
shl.b32 %r268, %r2, 2;
mov.u32 %r269, _ZZ13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r29, %r269, %r268;
ld.shared.f32 %f70, [%r29];
add.ftz.f32 %f13, %f136, %f70;
@%p17 bra $L__BB172_19;
bra.uni $L__BB172_18;

$L__BB172_19:
mov.u32 %r272, _ZZ13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r273, %r272, %r268;
ld.shared.f32 %f74, [%r273];
add.ftz.f32 %f75, %f135, %f74;
ld.shared.f32 %f76, [%r29+128];
add.ftz.f32 %f77, %f13, %f76;
ld.shared.f32 %f78, [%r273+128];
add.ftz.f32 %f79, %f75, %f78;
ld.shared.f32 %f80, [%r29+256];
add.ftz.f32 %f139, %f77, %f80;
ld.shared.f32 %f81, [%r273+256];
add.ftz.f32 %f135, %f79, %f81;
bra.uni $L__BB172_20;

$L__BB172_18:
ld.shared.f32 %f71, [%r29+128];
add.ftz.f32 %f72, %f13, %f71;
ld.shared.f32 %f73, [%r29+256];
add.ftz.f32 %f139, %f72, %f73;

$L__BB172_20:
mov.b32 %r274, %f139;
mov.u32 %r275, 31;
mov.u32 %r276, 16;
mov.u32 %r277, -1;
shfl.sync.bfly.b32 %r278|%p18, %r274, %r276, %r275, %r277;
mov.b32 %f82, %r278;
add.ftz.f32 %f83, %f139, %f82;
mov.b32 %r279, %f83;
mov.u32 %r280, 8;
shfl.sync.bfly.b32 %r281|%p19, %r279, %r280, %r275, %r277;
mov.b32 %f84, %r281;
add.ftz.f32 %f85, %f83, %f84;
mov.b32 %r282, %f85;
mov.u32 %r283, 4;
shfl.sync.bfly.b32 %r284|%p20, %r282, %r283, %r275, %r277;
mov.b32 %f86, %r284;
add.ftz.f32 %f87, %f85, %f86;
mov.b32 %r285, %f87;
mov.u32 %r286, 2;
shfl.sync.bfly.b32 %r287|%p21, %r285, %r286, %r275, %r277;
mov.b32 %f88, %r287;
add.ftz.f32 %f19, %f87, %f88;
mov.b32 %r288, %f19;
mov.u32 %r289, 1;
shfl.sync.bfly.b32 %r30|%p1, %r288, %r289, %r275, %r277;
setp.eq.s64 %p22, %rd12, 0;
@%p22 bra $L__BB172_22;

mov.b32 %r290, %f135;
shfl.sync.bfly.b32 %r294|%p23, %r290, %r276, %r275, %r277;
mov.b32 %f89, %r294;
add.ftz.f32 %f90, %f135, %f89;
mov.b32 %r295, %f90;
shfl.sync.bfly.b32 %r297|%p24, %r295, %r280, %r275, %r277;
mov.b32 %f91, %r297;
add.ftz.f32 %f92, %f90, %f91;
mov.b32 %r298, %f92;
shfl.sync.bfly.b32 %r300|%p25, %r298, %r283, %r275, %r277;
mov.b32 %f93, %r300;
add.ftz.f32 %f94, %f92, %f93;
mov.b32 %r301, %f94;
shfl.sync.bfly.b32 %r303|%p26, %r301, %r286, %r275, %r277;
mov.b32 %f95, %r303;
add.ftz.f32 %f96, %f94, %f95;
mov.b32 %r304, %f96;
shfl.sync.bfly.b32 %r306|%p27, %r304, %r289, %r275, %r277;
mov.b32 %f97, %r306;
add.ftz.f32 %f135, %f96, %f97;

$L__BB172_22:
setp.ne.s32 %p28, %r2, 0;
mov.b32 %f98, %r30;
add.ftz.f32 %f22, %f19, %f98;
@%p28 bra $L__BB172_35;

add.ftz.f32 %f141, %f131, %f22;
@%p22 bra $L__BB172_34;

add.ftz.f32 %f24, %f132, %f135;
setp.gt.s32 %p30, %r31, 2;
@%p30 bra $L__BB172_28;

setp.eq.s32 %p33, %r31, 1;
@%p33 bra $L__BB172_32;

setp.eq.s32 %p34, %r31, 2;
@%p34 bra $L__BB172_27;
bra.uni $L__BB172_33;

$L__BB172_27:
mul.ftz.f32 %f118, %f24, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f119, %f118;
add.ftz.f32 %f120, %f119, 0f3F800000;
div.approx.ftz.f32 %f121, %f24, %f120;
mul.ftz.f32 %f141, %f141, %f121;
bra.uni $L__BB172_34;

$L__BB172_28:
setp.eq.s32 %p31, %r31, 3;
@%p31 bra $L__BB172_31;

setp.ne.s32 %p32, %r31, 6;
@%p32 bra $L__BB172_33;

min.ftz.f32 %f99, %f24, %f31;
neg.ftz.f32 %f100, %f31;
min.ftz.f32 %f101, %f141, %f31;
max.ftz.f32 %f102, %f101, %f100;
mul.ftz.f32 %f103, %f99, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f104, %f103;
add.ftz.f32 %f105, %f104, 0f3F800000;
div.approx.ftz.f32 %f106, %f99, %f105;
mul.ftz.f32 %f141, %f102, %f106;
bra.uni $L__BB172_34;

$L__BB172_32:
mul.ftz.f32 %f122, %f24, 0f3F000000;
mul.ftz.f32 %f123, %f24, 0f3D372713;
fma.rn.ftz.f32 %f124, %f24, %f123, 0f3F800000;
mul.ftz.f32 %f125, %f24, 0f3F4C422A;
mul.ftz.f32 %f126, %f125, %f124;
tanh.approx.f32 %f127, %f126;
add.ftz.f32 %f128, %f127, 0f3F800000;
mul.ftz.f32 %f129, %f122, %f128;
mul.ftz.f32 %f141, %f141, %f129;
bra.uni $L__BB172_34;

$L__BB172_31:
mov.f32 %f107, 0f40E00000;
min.ftz.f32 %f108, %f24, %f107;
min.ftz.f32 %f109, %f141, %f107;
mov.f32 %f110, 0fC0E00000;
max.ftz.f32 %f111, %f109, %f110;
mul.ftz.f32 %f112, %f108, 0f3FD9DB23;
mul.ftz.f32 %f113, %f112, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f114, %f113;
add.ftz.f32 %f115, %f114, 0f3F800000;
div.approx.ftz.f32 %f116, %f108, %f115;
add.ftz.f32 %f117, %f111, 0f3F800000;
mul.ftz.f32 %f141, %f117, %f116;
bra.uni $L__BB172_34;

$L__BB172_33:
mul.ftz.f32 %f141, %f141, %f24;

$L__BB172_34:
cvta.to.global.u64 %rd75, %rd16;
mov.u32 %r310, %ctaid.y;
mad.lo.s32 %r312, %r310, %r42, %r4;
mad.lo.s32 %r313, %r11, %r48, %r312;
mul.wide.u32 %rd76, %r313, 4;
add.s64 %rd77, %rd75, %rd76;
st.global.f32 [%rd77], %f141;

$L__BB172_35:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<13>;
.reg .b16 %rs<8>;
.reg .f32 %f<39>;
.reg .b32 %r<183>;
.reg .b64 %rd<40>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[384];

ld.param.u64 %rd10, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd11, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd12, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd13, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
mov.u32 %r1, %tid.y;
shl.b32 %r38, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r38, %r2;
mov.u32 %r180, %ctaid.y;
setp.eq.s64 %p1, %rd12, 0;
@%p1 bra $L__BB173_2;

cvta.to.global.u64 %rd14, %rd12;
mul.wide.u32 %rd15, %r180, 4;
add.s64 %rd16, %rd14, %rd15;
ld.global.nc.u32 %r179, [%rd16];
mul.hi.u32 %r39, %r180, %r22;
add.s32 %r40, %r39, %r180;
shr.u32 %r41, %r40, %r23;
mul.lo.s32 %r42, %r41, %r24;
sub.s32 %r180, %r180, %r42;
bra.uni $L__BB173_3;

$L__BB173_2:
mul.hi.u32 %r43, %r180, %r26;
add.s32 %r44, %r43, %r180;
shr.u32 %r179, %r44, %r27;

$L__BB173_3:
mov.u32 %r10, %ctaid.z;
shr.s32 %r45, %r3, 31;
shr.u32 %r46, %r45, 28;
add.s32 %r47, %r3, %r46;
shr.s32 %r182, %r47, 4;
shr.u32 %r48, %r21, 8;
setp.ge.s32 %p2, %r182, %r48;
mov.f32 %f38, 0f00000000;
@%p2 bra $L__BB173_6;

mul.lo.s32 %r49, %r180, %r30;
cvt.u64.u32 %rd17, %r49;
and.b32 %r53, %r47, -16;
sub.s32 %r12, %r3, %r53;
shr.s32 %r54, %r12, 31;
shr.u32 %r55, %r54, 30;
add.s32 %r56, %r12, %r55;
shr.s32 %r57, %r56, 2;
shl.b32 %r13, %r57, 1;
shl.b32 %r58, %r57, 5;
and.b32 %r59, %r56, -4;
sub.s32 %r60, %r12, %r59;
shl.b32 %r61, %r60, 2;
add.s32 %r62, %r58, %r61;
cvt.s64.s32 %rd18, %r62;
cvt.s64.s32 %rd19, %r61;
cvt.s64.s32 %rd20, %r13;
mul.lo.s32 %r63, %r10, %r36;
cvt.u64.u32 %rd21, %r63;
add.s64 %rd22, %rd21, %rd20;
add.s64 %rd1, %rd22, %rd17;
cvt.s64.s32 %rd2, %r60;
mul.hi.u32 %r64, %r10, %r32;
add.s32 %r65, %r64, %r10;
shr.u32 %r66, %r65, %r33;
mad.lo.s32 %r67, %r66, %r35, %r182;
mad.lo.s32 %r68, %r179, %r29, %r67;
mov.u32 %r69, %ctaid.x;
mad.lo.s32 %r70, %r69, %r25, %r68;
mul.wide.s32 %rd23, %r70, 176;
cvta.to.global.u64 %rd24, %rd10;
add.s64 %rd39, %rd24, %rd23;
shl.b32 %r181, %r182, 3;
cvt.u64.u32 %rd25, %r13;
and.b64 %rd26, %rd25, 2;
or.b64 %rd4, %rd26, 8;
add.s64 %rd5, %rd18, 64;
add.s64 %rd6, %rd19, 16;
cvta.to.global.u64 %rd7, %rd11;
mov.f32 %f38, 0f00000000;
shl.b64 %rd34, %rd2, 2;

$L__BB173_5:
add.s64 %rd27, %rd39, %rd6;
add.s64 %rd28, %rd39, %rd5;
ld.global.nc.u32 %r107, [%rd28+-16];
ld.global.nc.u32 %r108, [%rd28];
ld.global.nc.u32 %r109, [%rd27];
shr.s32 %r110, %r109, %r13;
ld.global.nc.u32 %r111, [%rd27+16];
shr.s32 %r112, %r111, %r13;
add.s64 %rd29, %rd39, %rd4;
ld.global.nc.u16 %rs1, [%rd29+-4];
cvt.u32.u16 %r113, %rs1;
ld.global.nc.u16 %rs2, [%rd29];
cvt.u32.u16 %r114, %rs2;
ld.global.nc.u16 %rs3, [%rd29+4];
cvt.u32.u16 %r115, %rs3;
and.b32 %r116, %r113, 16191;
and.b32 %r117, %r115, 3855;
and.b32 %r118, %r113, 49344;
shr.u32 %r119, %r118, 2;
or.b32 %r120, %r117, %r119;
setp.gt.s32 %p3, %r12, 7;
and.b32 %r121, %r114, 16191;
shr.u32 %r122, %r115, 4;
and.b32 %r123, %r122, 3855;
and.b32 %r124, %r114, 49344;
shr.u32 %r125, %r124, 2;
or.b32 %r126, %r123, %r125;
cvt.u64.u32 %rd30, %r181;
add.s64 %rd31, %rd1, %rd30;
mul.lo.s64 %rd32, %rd31, 36;
add.s64 %rd33, %rd7, %rd32;
ld.global.nc.u32 %r71, [%rd33];

{.reg .f16 low,high;
mov.b32 {low,high},%r71;
cvt.f32.f16 %f7, low;}


add.s64 %rd35, %rd33, %rd34;
ld.global.nc.u32 %r79, [%rd35+4];
ld.global.nc.u32 %r75, [%rd35+20];
ld.global.nc.u32 %r72, [%rd33+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r72;
cvt.f32.f16 %f8, low;}


ld.global.nc.u32 %r95, [%rd35+40];
ld.global.nc.u32 %r91, [%rd35+56];
and.b32 %r127, %r107, 252645135;
and.b32 %r128, %r108, 252645135;
shl.b32 %r129, %r110, 4;
and.b32 %r130, %r129, 269488144;
shl.b32 %r131, %r112, 4;
and.b32 %r132, %r131, 269488144;
or.b32 %r78, %r127, %r130;
or.b32 %r74, %r128, %r132;
mov.u32 %r100, 0;

dp4a.s32.s32 %r73, %r74, %r75, %r100;


dp4a.s32.s32 %r77, %r78, %r79, %r73;

mov.u32 %r102, 16843009;

dp4a.s32.s32 %r81, %r102, %r75, %r100;


dp4a.s32.s32 %r85, %r102, %r79, %r81;

selp.b32 %r133, %r120, %r116, %p3;
cvt.u16.u32 %rs4, %r133;
and.b32 %r134, %r133, 63;
mul.lo.s32 %r135, %r77, %r134;
cvt.rn.f32.s32 %f11, %r135;
fma.rn.ftz.f32 %f12, %f7, %f11, 0f00000000;
selp.b32 %r136, %r126, %r121, %p3;
cvt.u16.u32 %rs5, %r136;
and.b32 %r137, %r136, 63;
mul.lo.s32 %r138, %r85, %r137;
cvt.rn.f32.s32 %f13, %r138;
fma.rn.ftz.f32 %f14, %f7, %f13, 0f00000000;
shr.u32 %r139, %r107, 4;
and.b32 %r140, %r139, 252645135;
shr.u32 %r141, %r108, 4;
and.b32 %r142, %r141, 252645135;
shl.b32 %r143, %r110, 3;
and.b32 %r144, %r143, 269488144;
shl.b32 %r145, %r112, 3;
and.b32 %r146, %r145, 269488144;
or.b32 %r94, %r140, %r144;
or.b32 %r90, %r142, %r146;

dp4a.s32.s32 %r89, %r90, %r91, %r100;


dp4a.s32.s32 %r93, %r94, %r95, %r89;


dp4a.s32.s32 %r97, %r102, %r91, %r100;


dp4a.s32.s32 %r101, %r102, %r95, %r97;

shr.u16 %rs6, %rs4, 8;
cvt.u32.u16 %r147, %rs6;
mul.lo.s32 %r148, %r93, %r147;
cvt.rn.f32.s32 %f15, %r148;
fma.rn.ftz.f32 %f16, %f8, %f15, %f12;
shr.u16 %rs7, %rs5, 8;
cvt.u32.u16 %r149, %rs7;
mul.lo.s32 %r150, %r101, %r149;
cvt.rn.f32.s32 %f17, %r150;
fma.rn.ftz.f32 %f18, %f8, %f17, %f14;
ld.global.nc.u32 %r105, [%rd39];

{.reg .f16 low,high;
mov.b32 {low,high},%r105;
cvt.f32.f16 %f9, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r105;
cvt.f32.f16 %f10, high;}


mul.ftz.f32 %f19, %f16, %f9;
mul.ftz.f32 %f20, %f18, %f10;
sub.ftz.f32 %f21, %f19, %f20;
add.ftz.f32 %f38, %f38, %f21;
add.s64 %rd39, %rd39, 1408;
add.s32 %r181, %r181, 64;
add.s32 %r182, %r182, 8;
setp.lt.s32 %p4, %r182, %r48;
@%p4 bra $L__BB173_5;

$L__BB173_6:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB173_8;

shl.b32 %r151, %r1, 7;
mov.u32 %r152, _ZZ13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r153, %r152, %r151;
shl.b32 %r154, %r2, 2;
add.s32 %r155, %r153, %r154;
st.shared.f32 [%r155+-128], %f38;

$L__BB173_8:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB173_11;

mov.u32 %r156, %ctaid.y;
mov.u32 %r157, %ctaid.x;
mad.lo.s32 %r158, %r156, %r31, %r157;
mad.lo.s32 %r20, %r10, %r37, %r158;
shl.b32 %r159, %r2, 2;
mov.u32 %r160, _ZZ13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r161, %r160, %r159;
mov.u32 %r162, 2;
ld.shared.f32 %f22, [%r161];
add.ftz.f32 %f23, %f38, %f22;
ld.shared.f32 %f24, [%r161+128];
add.ftz.f32 %f25, %f23, %f24;
ld.shared.f32 %f26, [%r161+256];
add.ftz.f32 %f27, %f25, %f26;
mov.b32 %r163, %f27;
mov.u32 %r164, 31;
mov.u32 %r165, 16;
mov.u32 %r166, -1;
shfl.sync.bfly.b32 %r167|%p7, %r163, %r165, %r164, %r166;
mov.b32 %f28, %r167;
add.ftz.f32 %f29, %f27, %f28;
mov.b32 %r168, %f29;
mov.u32 %r169, 8;
shfl.sync.bfly.b32 %r170|%p8, %r168, %r169, %r164, %r166;
mov.b32 %f30, %r170;
add.ftz.f32 %f31, %f29, %f30;
mov.b32 %r171, %f31;
mov.u32 %r172, 4;
shfl.sync.bfly.b32 %r173|%p9, %r171, %r172, %r164, %r166;
mov.b32 %f32, %r173;
add.ftz.f32 %f33, %f31, %f32;
mov.b32 %r174, %f33;
shfl.sync.bfly.b32 %r175|%p10, %r174, %r162, %r164, %r166;
mov.b32 %f34, %r175;
add.ftz.f32 %f35, %f33, %f34;
mov.b32 %r176, %f35;
mov.u32 %r177, 1;
shfl.sync.bfly.b32 %r178|%p11, %r176, %r177, %r164, %r166;
mov.b32 %f36, %r178;
add.ftz.f32 %f4, %f35, %f36;
setp.ne.s32 %p12, %r2, 0;
@%p12 bra $L__BB173_11;

cvta.to.global.u64 %rd36, %rd13;
mul.wide.u32 %rd37, %r20, 4;
add.s64 %rd38, %rd36, %rd37;
st.global.f32 [%rd38], %f4;

$L__BB173_11:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<35>;
.reg .b16 %rs<12>;
.reg .f32 %f<142>;
.reg .b32 %r<318>;
.reg .b64 %rd<79>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[384];

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate[384];

ld.param.u64 %rd8, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd9, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd10, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r46, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r47, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r48, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
ld.param.f32 %f31, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
ld.param.u64 %rd13, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+16];
ld.param.u64 %rd12, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+8];
ld.param.u64 %rd11, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3];
mov.u32 %r1, %tid.y;
shl.b32 %r49, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r49, %r2;
mov.u32 %r4, %ctaid.x;
mov.u32 %r5, %ctaid.y;
setp.eq.s64 %p2, %rd10, 0;
@%p2 bra $L__BB174_2;

cvta.to.global.u64 %rd17, %rd10;
mul.wide.u32 %rd18, %r5, 4;
add.s64 %rd19, %rd17, %rd18;
ld.global.nc.u32 %r314, [%rd19];
mul.hi.u32 %r50, %r5, %r33;
add.s32 %r51, %r50, %r5;
shr.u32 %r52, %r51, %r34;
mul.lo.s32 %r53, %r52, %r35;
sub.s32 %r315, %r5, %r53;
bra.uni $L__BB174_3;

$L__BB174_2:
mul.hi.u32 %r54, %r5, %r37;
add.s32 %r55, %r54, %r5;
shr.u32 %r314, %r55, %r38;
mov.u32 %r315, %r5;

$L__BB174_3:
mov.u32 %r11, %ctaid.z;
selp.b32 %r12, %r5, %r314, %p2;
or.b32 %r56, %r2, %r1;
setp.ne.s32 %p4, %r56, 0;
mov.f32 %f136, 0f00000000;
mov.f32 %f131, %f136;
mov.f32 %f132, %f136;
@%p4 bra $L__BB174_8;

setp.eq.s64 %p5, %rd11, 0;
mov.f32 %f132, 0f00000000;
mov.f32 %f131, %f132;
@%p5 bra $L__BB174_6;

cvta.to.global.u64 %rd20, %rd11;
mul.lo.s32 %r57, %r11, %r48;
cvt.u64.u32 %rd21, %r57;
mul.lo.s32 %r58, %r12, %r42;
cvt.u64.u32 %rd22, %r58;
cvt.u64.u32 %rd23, %r2;
cvt.s64.s32 %rd24, %r4;
add.s64 %rd25, %rd23, %rd24;
add.s64 %rd26, %rd25, %rd21;
add.s64 %rd27, %rd26, %rd22;
shl.b64 %rd28, %rd27, 2;
add.s64 %rd29, %rd20, %rd28;
ld.global.f32 %f131, [%rd29];

$L__BB174_6:
setp.eq.s64 %p6, %rd13, 0;
setp.eq.s64 %p7, %rd12, 0;
or.pred %p8, %p6, %p7;
@%p8 bra $L__BB174_8;

cvta.to.global.u64 %rd30, %rd13;
mul.lo.s32 %r60, %r11, %r48;
cvt.u64.u32 %rd31, %r60;
mul.lo.s32 %r61, %r12, %r42;
cvt.u64.u32 %rd32, %r61;
cvt.u64.u32 %rd33, %r2;
cvt.s64.s32 %rd34, %r4;
add.s64 %rd35, %rd33, %rd34;
add.s64 %rd36, %rd35, %rd31;
add.s64 %rd37, %rd36, %rd32;
shl.b64 %rd38, %rd37, 2;
add.s64 %rd39, %rd30, %rd38;
ld.global.f32 %f132, [%rd39];

$L__BB174_8:
shr.u32 %r63, %r32, 8;
shr.s32 %r64, %r3, 31;
shr.u32 %r65, %r64, 28;
add.s32 %r66, %r3, %r65;
shr.s32 %r317, %r66, 4;
setp.ge.s32 %p9, %r317, %r63;
mov.f32 %f135, %f136;
@%p9 bra $L__BB174_13;

and.b32 %r70, %r66, -16;
sub.s32 %r14, %r3, %r70;
shr.s32 %r71, %r14, 31;
shr.u32 %r72, %r71, 30;
add.s32 %r73, %r14, %r72;
shr.s32 %r15, %r73, 2;
and.b32 %r74, %r73, -4;
sub.s32 %r16, %r14, %r74;
shl.b32 %r316, %r317, 3;
mul.hi.u32 %r75, %r11, %r43;
add.s32 %r76, %r75, %r11;
shr.u32 %r77, %r76, %r44;
mad.lo.s32 %r78, %r77, %r46, %r317;
mad.lo.s32 %r79, %r314, %r40, %r78;
mad.lo.s32 %r80, %r4, %r36, %r79;
mul.wide.s32 %rd78, %r80, 176;
mul.lo.s32 %r81, %r315, %r41;
cvt.u64.u32 %rd40, %r81;
mul.lo.s32 %r82, %r11, %r47;
cvt.u64.u32 %rd41, %r82;
add.s64 %rd5, %rd40, %rd41;
mov.f32 %f135, 0f00000000;
cvta.to.global.u64 %rd43, %rd9;
cvta.to.global.u64 %rd47, %rd8;
mul.wide.s32 %rd60, %r16, 4;
mov.f32 %f136, %f135;

$L__BB174_10:
shl.b32 %r119, %r15, 1;
cvt.s64.s32 %rd42, %r119;
cvt.u64.u32 %rd44, %r316;
add.s64 %rd45, %rd5, %rd44;
shl.b32 %r120, %r16, 2;
shl.b32 %r121, %r15, 5;
add.s32 %r122, %r121, %r120;
cvt.s64.s32 %rd46, %r122;
add.s64 %rd48, %rd47, %rd46;
add.s64 %rd49, %rd48, %rd78;
cvt.s64.s32 %rd50, %r120;
add.s64 %rd51, %rd47, %rd50;
add.s64 %rd52, %rd51, %rd78;
ld.global.nc.u32 %r123, [%rd49+48];
ld.global.nc.u32 %r124, [%rd49+64];
ld.global.nc.u32 %r125, [%rd52+16];
shr.s32 %r126, %r125, %r119;
ld.global.nc.u32 %r127, [%rd52+32];
shr.s32 %r128, %r127, %r119;
cvt.u64.u32 %rd53, %r119;
and.b64 %rd54, %rd53, 2;
add.s64 %rd55, %rd47, %rd54;
add.s64 %rd56, %rd55, %rd78;
ld.global.nc.u16 %rs1, [%rd56+4];
cvt.u32.u16 %r129, %rs1;
ld.global.nc.u16 %rs2, [%rd56+8];
cvt.u32.u16 %r130, %rs2;
ld.global.nc.u16 %rs3, [%rd56+12];
cvt.u32.u16 %r131, %rs3;
and.b32 %r132, %r129, 16191;
and.b32 %r133, %r131, 3855;
and.b32 %r134, %r129, 49344;
shr.u32 %r135, %r134, 2;
or.b32 %r136, %r133, %r135;
setp.gt.s32 %p10, %r14, 7;
and.b32 %r137, %r130, 16191;
shr.u32 %r138, %r131, 4;
and.b32 %r139, %r138, 3855;
and.b32 %r140, %r130, 49344;
shr.u32 %r141, %r140, 2;
or.b32 %r142, %r139, %r141;
add.s64 %rd57, %rd45, %rd42;
mul.lo.s64 %rd58, %rd57, 36;
add.s64 %rd59, %rd43, %rd58;
ld.global.nc.u32 %r20, [%rd59];

{.reg .f16 low,high;
mov.b32 {low,high},%r20;
cvt.f32.f16 %f40, low;}


add.s64 %rd61, %rd59, %rd60;
ld.global.nc.u32 %r21, [%rd61+4];
ld.global.nc.u32 %r22, [%rd61+20];
ld.global.nc.u32 %r23, [%rd59+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r23;
cvt.f32.f16 %f41, low;}


ld.global.nc.u32 %r24, [%rd61+40];
ld.global.nc.u32 %r25, [%rd61+56];
add.s64 %rd62, %rd47, %rd78;
and.b32 %r143, %r123, 252645135;
and.b32 %r144, %r124, 252645135;
shl.b32 %r145, %r126, 4;
and.b32 %r146, %r145, 269488144;
shl.b32 %r147, %r128, 4;
and.b32 %r148, %r147, 269488144;
or.b32 %r90, %r143, %r146;
or.b32 %r86, %r144, %r148;
mov.u32 %r112, 0;

dp4a.s32.s32 %r85, %r86, %r22, %r112;


dp4a.s32.s32 %r89, %r90, %r21, %r85;

mov.u32 %r114, 16843009;

dp4a.s32.s32 %r93, %r114, %r22, %r112;


dp4a.s32.s32 %r97, %r114, %r21, %r93;

selp.b32 %r149, %r136, %r132, %p10;
cvt.u16.u32 %rs4, %r149;
and.b32 %r150, %r149, 63;
mul.lo.s32 %r151, %r89, %r150;
cvt.rn.f32.s32 %f44, %r151;
fma.rn.ftz.f32 %f45, %f40, %f44, 0f00000000;
selp.b32 %r152, %r142, %r137, %p10;
cvt.u16.u32 %rs5, %r152;
and.b32 %r153, %r152, 63;
mul.lo.s32 %r154, %r97, %r153;
cvt.rn.f32.s32 %f46, %r154;
fma.rn.ftz.f32 %f47, %f40, %f46, 0f00000000;
shr.u32 %r155, %r123, 4;
and.b32 %r156, %r155, 252645135;
shr.u32 %r157, %r124, 4;
and.b32 %r158, %r157, 252645135;
shl.b32 %r159, %r126, 3;
and.b32 %r160, %r159, 269488144;
shl.b32 %r161, %r128, 3;
and.b32 %r162, %r161, 269488144;
or.b32 %r106, %r156, %r160;
or.b32 %r102, %r158, %r162;

dp4a.s32.s32 %r101, %r102, %r25, %r112;


dp4a.s32.s32 %r105, %r106, %r24, %r101;


dp4a.s32.s32 %r109, %r114, %r25, %r112;


dp4a.s32.s32 %r113, %r114, %r24, %r109;

shr.u16 %rs6, %rs4, 8;
cvt.u32.u16 %r163, %rs6;
mul.lo.s32 %r164, %r105, %r163;
cvt.rn.f32.s32 %f48, %r164;
fma.rn.ftz.f32 %f49, %f41, %f48, %f45;
shr.u16 %rs7, %rs5, 8;
cvt.u32.u16 %r165, %rs7;
mul.lo.s32 %r166, %r113, %r165;
cvt.rn.f32.s32 %f50, %r166;
fma.rn.ftz.f32 %f51, %f41, %f50, %f47;
ld.global.nc.u32 %r117, [%rd62];

{.reg .f16 low,high;
mov.b32 {low,high},%r117;
cvt.f32.f16 %f42, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r117;
cvt.f32.f16 %f43, high;}


mul.ftz.f32 %f52, %f49, %f42;
mul.ftz.f32 %f53, %f51, %f43;
sub.ftz.f32 %f54, %f52, %f53;
add.ftz.f32 %f136, %f136, %f54;
setp.eq.s64 %p11, %rd12, 0;
@%p11 bra $L__BB174_12;

cvta.to.global.u64 %rd64, %rd12;
add.s64 %rd65, %rd64, %rd46;
add.s64 %rd66, %rd65, %rd78;
add.s64 %rd68, %rd64, %rd50;
add.s64 %rd69, %rd68, %rd78;
ld.global.u32 %r207, [%rd66+48];
ld.global.u32 %r208, [%rd66+64];
ld.global.u32 %r209, [%rd69+16];
shr.s32 %r210, %r209, %r119;
ld.global.u32 %r211, [%rd69+32];
shr.s32 %r212, %r211, %r119;
add.s64 %rd72, %rd64, %rd54;
add.s64 %rd73, %rd72, %rd78;
ld.global.u16 %r213, [%rd73+4];
ld.global.u16 %r214, [%rd73+8];
ld.global.u16 %r215, [%rd73+12];
and.b32 %r216, %r213, 16191;
and.b32 %r217, %r215, 3855;
and.b32 %r218, %r213, 49344;
shr.u32 %r219, %r218, 2;
or.b32 %r220, %r217, %r219;
and.b32 %r221, %r214, 16191;
shr.u32 %r222, %r215, 4;
and.b32 %r223, %r222, 3855;
and.b32 %r224, %r214, 49344;
shr.u32 %r225, %r224, 2;
or.b32 %r226, %r223, %r225;

{.reg .f16 low,high;
mov.b32 {low,high},%r20;
cvt.f32.f16 %f55, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r23;
cvt.f32.f16 %f56, low;}


add.s64 %rd74, %rd64, %rd78;
and.b32 %r227, %r207, 252645135;
and.b32 %r228, %r208, 252645135;
shl.b32 %r229, %r210, 4;
and.b32 %r230, %r229, 269488144;
shl.b32 %r231, %r212, 4;
and.b32 %r232, %r231, 269488144;
or.b32 %r174, %r227, %r230;
or.b32 %r170, %r228, %r232;

dp4a.s32.s32 %r169, %r170, %r22, %r112;


dp4a.s32.s32 %r173, %r174, %r21, %r169;


dp4a.s32.s32 %r177, %r114, %r22, %r112;


dp4a.s32.s32 %r181, %r114, %r21, %r177;

selp.b32 %r233, %r220, %r216, %p10;
cvt.u16.u32 %rs8, %r233;
and.b32 %r234, %r233, 63;
mul.lo.s32 %r235, %r173, %r234;
cvt.rn.f32.s32 %f59, %r235;
fma.rn.ftz.f32 %f60, %f55, %f59, 0f00000000;
selp.b32 %r236, %r226, %r221, %p10;
cvt.u16.u32 %rs9, %r236;
and.b32 %r237, %r236, 63;
mul.lo.s32 %r238, %r181, %r237;
cvt.rn.f32.s32 %f61, %r238;
fma.rn.ftz.f32 %f62, %f55, %f61, 0f00000000;
shr.u32 %r239, %r207, 4;
and.b32 %r240, %r239, 252645135;
shr.u32 %r241, %r208, 4;
and.b32 %r242, %r241, 252645135;
shl.b32 %r243, %r210, 3;
and.b32 %r244, %r243, 269488144;
shl.b32 %r245, %r212, 3;
and.b32 %r246, %r245, 269488144;
or.b32 %r190, %r240, %r244;
or.b32 %r186, %r242, %r246;

dp4a.s32.s32 %r185, %r186, %r25, %r112;


dp4a.s32.s32 %r189, %r190, %r24, %r185;


dp4a.s32.s32 %r193, %r114, %r25, %r112;


dp4a.s32.s32 %r197, %r114, %r24, %r193;

shr.u16 %rs10, %rs8, 8;
cvt.u32.u16 %r247, %rs10;
mul.lo.s32 %r248, %r189, %r247;
cvt.rn.f32.s32 %f63, %r248;
fma.rn.ftz.f32 %f64, %f56, %f63, %f60;
shr.u16 %rs11, %rs9, 8;
cvt.u32.u16 %r249, %rs11;
mul.lo.s32 %r250, %r197, %r249;
cvt.rn.f32.s32 %f65, %r250;
fma.rn.ftz.f32 %f66, %f56, %f65, %f62;
ld.global.u32 %r202, [%rd74];

{.reg .f16 low,high;
mov.b32 {low,high},%r202;
cvt.f32.f16 %f57, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r202;
cvt.f32.f16 %f58, high;}


mul.ftz.f32 %f67, %f64, %f57;
mul.ftz.f32 %f68, %f66, %f58;
sub.ftz.f32 %f69, %f67, %f68;
add.ftz.f32 %f135, %f135, %f69;

$L__BB174_12:
add.s32 %r316, %r316, 64;
add.s64 %rd78, %rd78, 1408;
add.s32 %r317, %r317, 8;
setp.lt.s32 %p13, %r317, %r63;
@%p13 bra $L__BB174_10;

$L__BB174_13:
setp.eq.s32 %p14, %r1, 0;
@%p14 bra $L__BB174_16;

setp.eq.s64 %p15, %rd12, 0;
add.s32 %r28, %r1, -1;
shl.b32 %r255, %r1, 7;
mov.u32 %r256, _ZZ13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r257, %r256, %r255;
shl.b32 %r258, %r2, 2;
add.s32 %r259, %r257, %r258;
st.shared.f32 [%r259+-128], %f136;
@%p15 bra $L__BB174_16;

shl.b32 %r261, %r28, 7;
mov.u32 %r262, _ZZ13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r263, %r262, %r261;
add.s32 %r265, %r263, %r258;
st.shared.f32 [%r265], %f135;

$L__BB174_16:
setp.ne.s32 %p16, %r1, 0;
bar.sync 0;
@%p16 bra $L__BB174_35;

setp.ne.s64 %p17, %rd12, 0;
shl.b32 %r268, %r2, 2;
mov.u32 %r269, _ZZ13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r29, %r269, %r268;
ld.shared.f32 %f70, [%r29];
add.ftz.f32 %f13, %f136, %f70;
@%p17 bra $L__BB174_19;
bra.uni $L__BB174_18;

$L__BB174_19:
mov.u32 %r272, _ZZ13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r273, %r272, %r268;
ld.shared.f32 %f74, [%r273];
add.ftz.f32 %f75, %f135, %f74;
ld.shared.f32 %f76, [%r29+128];
add.ftz.f32 %f77, %f13, %f76;
ld.shared.f32 %f78, [%r273+128];
add.ftz.f32 %f79, %f75, %f78;
ld.shared.f32 %f80, [%r29+256];
add.ftz.f32 %f139, %f77, %f80;
ld.shared.f32 %f81, [%r273+256];
add.ftz.f32 %f135, %f79, %f81;
bra.uni $L__BB174_20;

$L__BB174_18:
ld.shared.f32 %f71, [%r29+128];
add.ftz.f32 %f72, %f13, %f71;
ld.shared.f32 %f73, [%r29+256];
add.ftz.f32 %f139, %f72, %f73;

$L__BB174_20:
mov.b32 %r274, %f139;
mov.u32 %r275, 31;
mov.u32 %r276, 16;
mov.u32 %r277, -1;
shfl.sync.bfly.b32 %r278|%p18, %r274, %r276, %r275, %r277;
mov.b32 %f82, %r278;
add.ftz.f32 %f83, %f139, %f82;
mov.b32 %r279, %f83;
mov.u32 %r280, 8;
shfl.sync.bfly.b32 %r281|%p19, %r279, %r280, %r275, %r277;
mov.b32 %f84, %r281;
add.ftz.f32 %f85, %f83, %f84;
mov.b32 %r282, %f85;
mov.u32 %r283, 4;
shfl.sync.bfly.b32 %r284|%p20, %r282, %r283, %r275, %r277;
mov.b32 %f86, %r284;
add.ftz.f32 %f87, %f85, %f86;
mov.b32 %r285, %f87;
mov.u32 %r286, 2;
shfl.sync.bfly.b32 %r287|%p21, %r285, %r286, %r275, %r277;
mov.b32 %f88, %r287;
add.ftz.f32 %f19, %f87, %f88;
mov.b32 %r288, %f19;
mov.u32 %r289, 1;
shfl.sync.bfly.b32 %r30|%p1, %r288, %r289, %r275, %r277;
setp.eq.s64 %p22, %rd12, 0;
@%p22 bra $L__BB174_22;

mov.b32 %r290, %f135;
shfl.sync.bfly.b32 %r294|%p23, %r290, %r276, %r275, %r277;
mov.b32 %f89, %r294;
add.ftz.f32 %f90, %f135, %f89;
mov.b32 %r295, %f90;
shfl.sync.bfly.b32 %r297|%p24, %r295, %r280, %r275, %r277;
mov.b32 %f91, %r297;
add.ftz.f32 %f92, %f90, %f91;
mov.b32 %r298, %f92;
shfl.sync.bfly.b32 %r300|%p25, %r298, %r283, %r275, %r277;
mov.b32 %f93, %r300;
add.ftz.f32 %f94, %f92, %f93;
mov.b32 %r301, %f94;
shfl.sync.bfly.b32 %r303|%p26, %r301, %r286, %r275, %r277;
mov.b32 %f95, %r303;
add.ftz.f32 %f96, %f94, %f95;
mov.b32 %r304, %f96;
shfl.sync.bfly.b32 %r306|%p27, %r304, %r289, %r275, %r277;
mov.b32 %f97, %r306;
add.ftz.f32 %f135, %f96, %f97;

$L__BB174_22:
setp.ne.s32 %p28, %r2, 0;
mov.b32 %f98, %r30;
add.ftz.f32 %f22, %f19, %f98;
@%p28 bra $L__BB174_35;

add.ftz.f32 %f141, %f131, %f22;
@%p22 bra $L__BB174_34;

add.ftz.f32 %f24, %f132, %f135;
setp.gt.s32 %p30, %r31, 2;
@%p30 bra $L__BB174_28;

setp.eq.s32 %p33, %r31, 1;
@%p33 bra $L__BB174_32;

setp.eq.s32 %p34, %r31, 2;
@%p34 bra $L__BB174_27;
bra.uni $L__BB174_33;

$L__BB174_27:
mul.ftz.f32 %f118, %f24, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f119, %f118;
add.ftz.f32 %f120, %f119, 0f3F800000;
div.approx.ftz.f32 %f121, %f24, %f120;
mul.ftz.f32 %f141, %f141, %f121;
bra.uni $L__BB174_34;

$L__BB174_28:
setp.eq.s32 %p31, %r31, 3;
@%p31 bra $L__BB174_31;

setp.ne.s32 %p32, %r31, 6;
@%p32 bra $L__BB174_33;

min.ftz.f32 %f99, %f24, %f31;
neg.ftz.f32 %f100, %f31;
min.ftz.f32 %f101, %f141, %f31;
max.ftz.f32 %f102, %f101, %f100;
mul.ftz.f32 %f103, %f99, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f104, %f103;
add.ftz.f32 %f105, %f104, 0f3F800000;
div.approx.ftz.f32 %f106, %f99, %f105;
mul.ftz.f32 %f141, %f102, %f106;
bra.uni $L__BB174_34;

$L__BB174_32:
mul.ftz.f32 %f122, %f24, 0f3F000000;
mul.ftz.f32 %f123, %f24, 0f3D372713;
fma.rn.ftz.f32 %f124, %f24, %f123, 0f3F800000;
mul.ftz.f32 %f125, %f24, 0f3F4C422A;
mul.ftz.f32 %f126, %f125, %f124;
tanh.approx.f32 %f127, %f126;
add.ftz.f32 %f128, %f127, 0f3F800000;
mul.ftz.f32 %f129, %f122, %f128;
mul.ftz.f32 %f141, %f141, %f129;
bra.uni $L__BB174_34;

$L__BB174_31:
mov.f32 %f107, 0f40E00000;
min.ftz.f32 %f108, %f24, %f107;
min.ftz.f32 %f109, %f141, %f107;
mov.f32 %f110, 0fC0E00000;
max.ftz.f32 %f111, %f109, %f110;
mul.ftz.f32 %f112, %f108, 0f3FD9DB23;
mul.ftz.f32 %f113, %f112, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f114, %f113;
add.ftz.f32 %f115, %f114, 0f3F800000;
div.approx.ftz.f32 %f116, %f108, %f115;
add.ftz.f32 %f117, %f111, 0f3F800000;
mul.ftz.f32 %f141, %f117, %f116;
bra.uni $L__BB174_34;

$L__BB174_33:
mul.ftz.f32 %f141, %f141, %f24;

$L__BB174_34:
cvta.to.global.u64 %rd75, %rd16;
mov.u32 %r310, %ctaid.y;
mad.lo.s32 %r312, %r310, %r42, %r4;
mad.lo.s32 %r313, %r11, %r48, %r312;
mul.wide.u32 %rd76, %r313, 4;
add.s64 %rd77, %rd75, %rd76;
st.global.f32 [%rd77], %f141;

$L__BB174_35:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<13>;
.reg .b16 %rs<8>;
.reg .f32 %f<39>;
.reg .b32 %r<183>;
.reg .b64 %rd<40>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[384];

ld.param.u64 %rd10, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd11, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd12, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd13, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
mov.u32 %r1, %tid.y;
shl.b32 %r38, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r38, %r2;
mov.u32 %r180, %ctaid.y;
setp.eq.s64 %p1, %rd12, 0;
@%p1 bra $L__BB175_2;

cvta.to.global.u64 %rd14, %rd12;
mul.wide.u32 %rd15, %r180, 4;
add.s64 %rd16, %rd14, %rd15;
ld.global.nc.u32 %r179, [%rd16];
mul.hi.u32 %r39, %r180, %r22;
add.s32 %r40, %r39, %r180;
shr.u32 %r41, %r40, %r23;
mul.lo.s32 %r42, %r41, %r24;
sub.s32 %r180, %r180, %r42;
bra.uni $L__BB175_3;

$L__BB175_2:
mul.hi.u32 %r43, %r180, %r26;
add.s32 %r44, %r43, %r180;
shr.u32 %r179, %r44, %r27;

$L__BB175_3:
mov.u32 %r10, %ctaid.z;
shr.s32 %r45, %r3, 31;
shr.u32 %r46, %r45, 28;
add.s32 %r47, %r3, %r46;
shr.s32 %r182, %r47, 4;
shr.u32 %r48, %r21, 8;
setp.ge.s32 %p2, %r182, %r48;
mov.f32 %f38, 0f00000000;
@%p2 bra $L__BB175_6;

mul.lo.s32 %r49, %r180, %r30;
cvt.u64.u32 %rd17, %r49;
and.b32 %r53, %r47, -16;
sub.s32 %r12, %r3, %r53;
shr.s32 %r54, %r12, 31;
shr.u32 %r55, %r54, 30;
add.s32 %r56, %r12, %r55;
shr.s32 %r57, %r56, 2;
shl.b32 %r13, %r57, 1;
shl.b32 %r58, %r57, 5;
and.b32 %r59, %r56, -4;
sub.s32 %r60, %r12, %r59;
shl.b32 %r61, %r60, 2;
add.s32 %r62, %r58, %r61;
cvt.s64.s32 %rd18, %r62;
cvt.s64.s32 %rd19, %r61;
cvt.s64.s32 %rd20, %r13;
mul.lo.s32 %r63, %r10, %r36;
cvt.u64.u32 %rd21, %r63;
add.s64 %rd22, %rd21, %rd20;
add.s64 %rd1, %rd22, %rd17;
cvt.s64.s32 %rd2, %r60;
mul.hi.u32 %r64, %r10, %r32;
add.s32 %r65, %r64, %r10;
shr.u32 %r66, %r65, %r33;
mad.lo.s32 %r67, %r66, %r35, %r182;
mad.lo.s32 %r68, %r179, %r29, %r67;
mov.u32 %r69, %ctaid.x;
mad.lo.s32 %r70, %r69, %r25, %r68;
mul.wide.s32 %rd23, %r70, 176;
cvta.to.global.u64 %rd24, %rd10;
add.s64 %rd39, %rd24, %rd23;
shl.b32 %r181, %r182, 3;
cvt.u64.u32 %rd25, %r13;
and.b64 %rd26, %rd25, 2;
or.b64 %rd4, %rd26, 8;
add.s64 %rd5, %rd18, 64;
add.s64 %rd6, %rd19, 16;
cvta.to.global.u64 %rd7, %rd11;
mov.f32 %f38, 0f00000000;
shl.b64 %rd34, %rd2, 2;

$L__BB175_5:
add.s64 %rd27, %rd39, %rd6;
add.s64 %rd28, %rd39, %rd5;
ld.global.nc.u32 %r107, [%rd28+-16];
ld.global.nc.u32 %r108, [%rd28];
ld.global.nc.u32 %r109, [%rd27];
shr.s32 %r110, %r109, %r13;
ld.global.nc.u32 %r111, [%rd27+16];
shr.s32 %r112, %r111, %r13;
add.s64 %rd29, %rd39, %rd4;
ld.global.nc.u16 %rs1, [%rd29+-4];
cvt.u32.u16 %r113, %rs1;
ld.global.nc.u16 %rs2, [%rd29];
cvt.u32.u16 %r114, %rs2;
ld.global.nc.u16 %rs3, [%rd29+4];
cvt.u32.u16 %r115, %rs3;
and.b32 %r116, %r113, 16191;
and.b32 %r117, %r115, 3855;
and.b32 %r118, %r113, 49344;
shr.u32 %r119, %r118, 2;
or.b32 %r120, %r117, %r119;
setp.gt.s32 %p3, %r12, 7;
and.b32 %r121, %r114, 16191;
shr.u32 %r122, %r115, 4;
and.b32 %r123, %r122, 3855;
and.b32 %r124, %r114, 49344;
shr.u32 %r125, %r124, 2;
or.b32 %r126, %r123, %r125;
cvt.u64.u32 %rd30, %r181;
add.s64 %rd31, %rd1, %rd30;
mul.lo.s64 %rd32, %rd31, 36;
add.s64 %rd33, %rd7, %rd32;
ld.global.nc.u32 %r71, [%rd33];

{.reg .f16 low,high;
mov.b32 {low,high},%r71;
cvt.f32.f16 %f7, low;}


add.s64 %rd35, %rd33, %rd34;
ld.global.nc.u32 %r79, [%rd35+4];
ld.global.nc.u32 %r75, [%rd35+20];
ld.global.nc.u32 %r72, [%rd33+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r72;
cvt.f32.f16 %f8, low;}


ld.global.nc.u32 %r95, [%rd35+40];
ld.global.nc.u32 %r91, [%rd35+56];
and.b32 %r127, %r107, 252645135;
and.b32 %r128, %r108, 252645135;
shl.b32 %r129, %r110, 4;
and.b32 %r130, %r129, 269488144;
shl.b32 %r131, %r112, 4;
and.b32 %r132, %r131, 269488144;
or.b32 %r78, %r127, %r130;
or.b32 %r74, %r128, %r132;
mov.u32 %r100, 0;

dp4a.s32.s32 %r73, %r74, %r75, %r100;


dp4a.s32.s32 %r77, %r78, %r79, %r73;

mov.u32 %r102, 16843009;

dp4a.s32.s32 %r81, %r102, %r75, %r100;


dp4a.s32.s32 %r85, %r102, %r79, %r81;

selp.b32 %r133, %r120, %r116, %p3;
cvt.u16.u32 %rs4, %r133;
and.b32 %r134, %r133, 63;
mul.lo.s32 %r135, %r77, %r134;
cvt.rn.f32.s32 %f11, %r135;
fma.rn.ftz.f32 %f12, %f7, %f11, 0f00000000;
selp.b32 %r136, %r126, %r121, %p3;
cvt.u16.u32 %rs5, %r136;
and.b32 %r137, %r136, 63;
mul.lo.s32 %r138, %r85, %r137;
cvt.rn.f32.s32 %f13, %r138;
fma.rn.ftz.f32 %f14, %f7, %f13, 0f00000000;
shr.u32 %r139, %r107, 4;
and.b32 %r140, %r139, 252645135;
shr.u32 %r141, %r108, 4;
and.b32 %r142, %r141, 252645135;
shl.b32 %r143, %r110, 3;
and.b32 %r144, %r143, 269488144;
shl.b32 %r145, %r112, 3;
and.b32 %r146, %r145, 269488144;
or.b32 %r94, %r140, %r144;
or.b32 %r90, %r142, %r146;

dp4a.s32.s32 %r89, %r90, %r91, %r100;


dp4a.s32.s32 %r93, %r94, %r95, %r89;


dp4a.s32.s32 %r97, %r102, %r91, %r100;


dp4a.s32.s32 %r101, %r102, %r95, %r97;

shr.u16 %rs6, %rs4, 8;
cvt.u32.u16 %r147, %rs6;
mul.lo.s32 %r148, %r93, %r147;
cvt.rn.f32.s32 %f15, %r148;
fma.rn.ftz.f32 %f16, %f8, %f15, %f12;
shr.u16 %rs7, %rs5, 8;
cvt.u32.u16 %r149, %rs7;
mul.lo.s32 %r150, %r101, %r149;
cvt.rn.f32.s32 %f17, %r150;
fma.rn.ftz.f32 %f18, %f8, %f17, %f14;
ld.global.nc.u32 %r105, [%rd39];

{.reg .f16 low,high;
mov.b32 {low,high},%r105;
cvt.f32.f16 %f9, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r105;
cvt.f32.f16 %f10, high;}


mul.ftz.f32 %f19, %f16, %f9;
mul.ftz.f32 %f20, %f18, %f10;
sub.ftz.f32 %f21, %f19, %f20;
add.ftz.f32 %f38, %f38, %f21;
add.s64 %rd39, %rd39, 1408;
add.s32 %r181, %r181, 64;
add.s32 %r182, %r182, 8;
setp.lt.s32 %p4, %r182, %r48;
@%p4 bra $L__BB175_5;

$L__BB175_6:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB175_8;

shl.b32 %r151, %r1, 7;
mov.u32 %r152, _ZZ13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r153, %r152, %r151;
shl.b32 %r154, %r2, 2;
add.s32 %r155, %r153, %r154;
st.shared.f32 [%r155+-128], %f38;

$L__BB175_8:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB175_11;

mov.u32 %r156, %ctaid.y;
mov.u32 %r157, %ctaid.x;
mad.lo.s32 %r158, %r156, %r31, %r157;
mad.lo.s32 %r20, %r10, %r37, %r158;
shl.b32 %r159, %r2, 2;
mov.u32 %r160, _ZZ13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r161, %r160, %r159;
mov.u32 %r162, 2;
ld.shared.f32 %f22, [%r161];
add.ftz.f32 %f23, %f38, %f22;
ld.shared.f32 %f24, [%r161+128];
add.ftz.f32 %f25, %f23, %f24;
ld.shared.f32 %f26, [%r161+256];
add.ftz.f32 %f27, %f25, %f26;
mov.b32 %r163, %f27;
mov.u32 %r164, 31;
mov.u32 %r165, 16;
mov.u32 %r166, -1;
shfl.sync.bfly.b32 %r167|%p7, %r163, %r165, %r164, %r166;
mov.b32 %f28, %r167;
add.ftz.f32 %f29, %f27, %f28;
mov.b32 %r168, %f29;
mov.u32 %r169, 8;
shfl.sync.bfly.b32 %r170|%p8, %r168, %r169, %r164, %r166;
mov.b32 %f30, %r170;
add.ftz.f32 %f31, %f29, %f30;
mov.b32 %r171, %f31;
mov.u32 %r172, 4;
shfl.sync.bfly.b32 %r173|%p9, %r171, %r172, %r164, %r166;
mov.b32 %f32, %r173;
add.ftz.f32 %f33, %f31, %f32;
mov.b32 %r174, %f33;
shfl.sync.bfly.b32 %r175|%p10, %r174, %r162, %r164, %r166;
mov.b32 %f34, %r175;
add.ftz.f32 %f35, %f33, %f34;
mov.b32 %r176, %f35;
mov.u32 %r177, 1;
shfl.sync.bfly.b32 %r178|%p11, %r176, %r177, %r164, %r166;
mov.b32 %f36, %r178;
add.ftz.f32 %f4, %f35, %f36;
setp.ne.s32 %p12, %r2, 0;
@%p12 bra $L__BB175_11;

cvta.to.global.u64 %rd36, %rd13;
mul.wide.u32 %rd37, %r20, 4;
add.s64 %rd38, %rd36, %rd37;
st.global.f32 [%rd38], %f4;

$L__BB175_11:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type13ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type13ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type13ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type13ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type13ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type13ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type13ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type13ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type13ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<36>;
.reg .b16 %rs<15>;
.reg .f32 %f<153>;
.reg .b32 %r<399>;
.reg .b64 %rd<64>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type13ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type13ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type13ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd20, [_Z13mul_mat_vec_qIL9ggml_type13ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type13ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r18, [_Z13mul_mat_vec_qIL9ggml_type13ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r19, [_Z13mul_mat_vec_qIL9ggml_type13ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type13ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type13ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type13ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type13ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type13ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type13ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type13ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type13ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type13ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type13ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type13ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
cvta.to.global.u64 %rd1, %rd20;
mov.u32 %r1, %tid.y;
shl.b32 %r34, %r1, 5;
mov.u32 %r35, %tid.x;
add.s32 %r2, %r34, %r35;
mov.u32 %r36, %ctaid.x;
shl.b32 %r3, %r36, 1;
shr.s32 %r37, %r2, 31;
shr.u32 %r38, %r37, 28;
add.s32 %r39, %r2, %r38;
shr.s32 %r398, %r39, 4;
shr.u32 %r5, %r33, 8;
setp.ge.s32 %p3, %r398, %r5;
mov.f32 %f149, 0f00000000;
mov.f32 %f150, %f149;
mov.f32 %f151, %f149;
mov.f32 %f152, %f149;
@%p3 bra $L__BB176_3;

mov.u32 %r40, %ctaid.y;
mov.u32 %r41, %ctaid.z;
mul.hi.u32 %r42, %r41, %r27;
add.s32 %r43, %r42, %r41;
shr.u32 %r44, %r43, %r28;
mul.lo.s32 %r45, %r44, %r30;
mul.hi.u32 %r46, %r40, %r21;
add.s32 %r47, %r46, %r40;
shr.u32 %r48, %r47, %r22;
mul.lo.s32 %r49, %r48, %r24;
mul.lo.s32 %r50, %r40, %r25;
cvt.u64.u32 %rd21, %r50;
cvta.to.global.u64 %rd63, %rd18;
and.b32 %r54, %r39, -16;
sub.s32 %r6, %r2, %r54;
shr.s32 %r55, %r6, 31;
shr.u32 %r56, %r55, 30;
add.s32 %r57, %r6, %r56;
shr.s32 %r58, %r57, 2;
shl.b32 %r7, %r58, 1;
shl.b32 %r59, %r58, 5;
and.b32 %r60, %r57, -4;
sub.s32 %r61, %r6, %r60;
shl.b32 %r62, %r61, 2;
add.s32 %r63, %r59, %r62;
cvt.s64.s32 %rd22, %r63;
cvt.s64.s32 %rd23, %r62;
cvt.s64.s32 %rd24, %r7;
add.s64 %rd25, %rd21, %rd24;
mul.lo.s32 %r64, %r41, %r31;
cvt.u64.u32 %rd26, %r64;
add.s64 %rd3, %rd25, %rd26;
cvt.s64.s32 %rd4, %r61;
shl.b32 %r396, %r398, 3;
add.s32 %r397, %r19, %r396;
add.s32 %r65, %r398, %r45;
add.s32 %r66, %r65, %r49;
add.s32 %r67, %r3, 1;
mad.lo.s32 %r68, %r18, %r67, %r66;
mul.wide.s32 %rd5, %r68, 176;
mad.lo.s32 %r69, %r3, %r18, %r398;
add.s32 %r70, %r69, %r45;
add.s32 %r71, %r70, %r49;
mul.wide.s32 %rd6, %r71, 176;
add.s64 %rd27, %rd6, %rd22;
add.s64 %rd7, %rd27, 64;
cvt.u64.u32 %rd28, %r7;
and.b64 %rd29, %rd28, 2;
or.b64 %rd30, %rd29, %rd5;
or.b64 %rd8, %rd30, 8;
add.s64 %rd31, %rd6, %rd23;
add.s64 %rd9, %rd31, 32;
or.b64 %rd32, %rd29, %rd6;
or.b64 %rd10, %rd32, 12;
add.s64 %rd33, %rd5, %rd23;
add.s64 %rd11, %rd33, 16;
add.s64 %rd34, %rd5, %rd22;
add.s64 %rd12, %rd34, 48;
cvta.to.global.u64 %rd13, %rd19;
mov.f32 %f149, 0f00000000;
shl.b64 %rd42, %rd4, 2;

$L__BB176_2:
cvt.u64.u32 %rd35, %r396;
add.s64 %rd36, %rd3, %rd35;
add.s64 %rd37, %rd63, %rd7;
ld.global.nc.u32 %r216, [%rd37+-16];
ld.global.nc.u32 %r217, [%rd37];
add.s64 %rd38, %rd63, %rd9;
ld.global.nc.u32 %r218, [%rd38+-16];
shr.s32 %r219, %r218, %r7;
ld.global.nc.u32 %r220, [%rd38];
shr.s32 %r221, %r220, %r7;
add.s64 %rd39, %rd63, %rd10;
ld.global.nc.u16 %rs1, [%rd39+-8];
cvt.u32.u16 %r222, %rs1;
ld.global.nc.u16 %rs2, [%rd39+-4];
cvt.u32.u16 %r223, %rs2;
ld.global.nc.u16 %rs3, [%rd39];
cvt.u32.u16 %r224, %rs3;
and.b32 %r225, %r222, 16191;
and.b32 %r226, %r224, 3855;
and.b32 %r227, %r222, 49344;
shr.u32 %r228, %r227, 2;
or.b32 %r229, %r226, %r228;
setp.gt.s32 %p4, %r6, 7;
and.b32 %r230, %r223, 16191;
shr.u32 %r231, %r224, 4;
and.b32 %r232, %r231, 3855;
and.b32 %r233, %r223, 49344;
shr.u32 %r234, %r233, 2;
or.b32 %r235, %r232, %r234;
mul.lo.s64 %rd40, %rd36, 36;
add.s64 %rd41, %rd13, %rd40;
ld.global.nc.u32 %r72, [%rd41];

{.reg .f16 low,high;
mov.b32 {low,high},%r72;
cvt.f32.f16 %f25, low;}


add.s64 %rd43, %rd41, %rd42;
ld.global.nc.u32 %r80, [%rd43+4];
ld.global.nc.u32 %r76, [%rd43+20];
ld.global.nc.u32 %r73, [%rd41+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r73;
cvt.f32.f16 %f26, low;}


ld.global.nc.u32 %r96, [%rd43+40];
ld.global.nc.u32 %r92, [%rd43+56];
add.s64 %rd44, %rd63, %rd6;
and.b32 %r236, %r216, 252645135;
and.b32 %r237, %r217, 252645135;
shl.b32 %r238, %r219, 4;
and.b32 %r239, %r238, 269488144;
shl.b32 %r240, %r221, 4;
and.b32 %r241, %r240, 269488144;
or.b32 %r151, %r236, %r239;
or.b32 %r147, %r237, %r241;
mov.u32 %r209, 0;

dp4a.s32.s32 %r74, %r147, %r76, %r209;


dp4a.s32.s32 %r78, %r151, %r80, %r74;

mov.u32 %r211, 16843009;

dp4a.s32.s32 %r82, %r211, %r76, %r209;


dp4a.s32.s32 %r86, %r211, %r80, %r82;

selp.b32 %r242, %r229, %r225, %p4;
cvt.u16.u32 %rs4, %r242;
and.b32 %r243, %r242, 63;
mul.lo.s32 %r244, %r78, %r243;
cvt.rn.f32.s32 %f41, %r244;
fma.rn.ftz.f32 %f42, %f25, %f41, 0f00000000;
selp.b32 %r245, %r235, %r230, %p4;
cvt.u16.u32 %rs5, %r245;
and.b32 %r246, %r245, 63;
mul.lo.s32 %r247, %r86, %r246;
cvt.rn.f32.s32 %f43, %r247;
fma.rn.ftz.f32 %f44, %f25, %f43, 0f00000000;
shr.u32 %r248, %r216, 4;
and.b32 %r249, %r248, 252645135;
shr.u32 %r250, %r217, 4;
and.b32 %r251, %r250, 252645135;
shl.b32 %r252, %r219, 3;
and.b32 %r253, %r252, 269488144;
shl.b32 %r254, %r221, 3;
and.b32 %r255, %r254, 269488144;
or.b32 %r167, %r249, %r253;
or.b32 %r163, %r251, %r255;

dp4a.s32.s32 %r90, %r163, %r92, %r209;


dp4a.s32.s32 %r94, %r167, %r96, %r90;


dp4a.s32.s32 %r98, %r211, %r92, %r209;


dp4a.s32.s32 %r102, %r211, %r96, %r98;

shr.u16 %rs6, %rs4, 8;
cvt.u32.u16 %r256, %rs6;
mul.lo.s32 %r257, %r94, %r256;
cvt.rn.f32.s32 %f45, %r257;
fma.rn.ftz.f32 %f46, %f26, %f45, %f42;
shr.u16 %rs7, %rs5, 8;
cvt.u32.u16 %r258, %rs7;
mul.lo.s32 %r259, %r102, %r258;
cvt.rn.f32.s32 %f47, %r259;
fma.rn.ftz.f32 %f48, %f26, %f47, %f44;
ld.global.nc.u32 %r106, [%rd44];

{.reg .f16 low,high;
mov.b32 {low,high},%r106;
cvt.f32.f16 %f27, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r106;
cvt.f32.f16 %f28, high;}


mul.ftz.f32 %f49, %f46, %f27;
mul.ftz.f32 %f50, %f48, %f28;
sub.ftz.f32 %f51, %f49, %f50;
add.ftz.f32 %f152, %f51, %f152;
add.s64 %rd45, %rd63, %rd12;
add.s64 %rd46, %rd63, %rd11;
ld.global.nc.u32 %r260, [%rd45];
ld.global.nc.u32 %r261, [%rd45+16];
ld.global.nc.u32 %r262, [%rd46];
shr.s32 %r263, %r262, %r7;
ld.global.nc.u32 %r264, [%rd46+16];
shr.s32 %r265, %r264, %r7;
add.s64 %rd47, %rd63, %rd8;
ld.global.nc.u16 %rs8, [%rd47+-4];
cvt.u32.u16 %r266, %rs8;
ld.global.nc.u16 %rs9, [%rd47];
cvt.u32.u16 %r267, %rs9;
ld.global.nc.u16 %rs10, [%rd47+4];
cvt.u32.u16 %r268, %rs10;
and.b32 %r269, %r266, 16191;
and.b32 %r270, %r268, 3855;
and.b32 %r271, %r266, 49344;
shr.u32 %r272, %r271, 2;
or.b32 %r273, %r270, %r272;
and.b32 %r274, %r267, 16191;
shr.u32 %r275, %r268, 4;
and.b32 %r276, %r275, 3855;
and.b32 %r277, %r267, 49344;
shr.u32 %r278, %r277, 2;
or.b32 %r279, %r276, %r278;

{.reg .f16 low,high;
mov.b32 {low,high},%r72;
cvt.f32.f16 %f29, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r73;
cvt.f32.f16 %f30, low;}


add.s64 %rd48, %rd63, %rd5;
and.b32 %r280, %r260, 252645135;
and.b32 %r281, %r261, 252645135;
shl.b32 %r282, %r263, 4;
and.b32 %r283, %r282, 269488144;
shl.b32 %r284, %r265, 4;
and.b32 %r285, %r284, 269488144;
or.b32 %r187, %r280, %r283;
or.b32 %r183, %r281, %r285;

dp4a.s32.s32 %r110, %r183, %r76, %r209;


dp4a.s32.s32 %r114, %r187, %r80, %r110;


dp4a.s32.s32 %r118, %r211, %r76, %r209;


dp4a.s32.s32 %r122, %r211, %r80, %r118;

selp.b32 %r286, %r273, %r269, %p4;
cvt.u16.u32 %rs11, %r286;
and.b32 %r287, %r286, 63;
mul.lo.s32 %r288, %r114, %r287;
cvt.rn.f32.s32 %f52, %r288;
fma.rn.ftz.f32 %f53, %f29, %f52, 0f00000000;
selp.b32 %r289, %r279, %r274, %p4;
cvt.u16.u32 %rs12, %r289;
and.b32 %r290, %r289, 63;
mul.lo.s32 %r291, %r122, %r290;
cvt.rn.f32.s32 %f54, %r291;
fma.rn.ftz.f32 %f55, %f29, %f54, 0f00000000;
shr.u32 %r292, %r260, 4;
and.b32 %r293, %r292, 252645135;
shr.u32 %r294, %r261, 4;
and.b32 %r295, %r294, 252645135;
shl.b32 %r296, %r263, 3;
and.b32 %r297, %r296, 269488144;
shl.b32 %r298, %r265, 3;
and.b32 %r299, %r298, 269488144;
or.b32 %r203, %r293, %r297;
or.b32 %r199, %r295, %r299;

dp4a.s32.s32 %r126, %r199, %r92, %r209;


dp4a.s32.s32 %r130, %r203, %r96, %r126;


dp4a.s32.s32 %r134, %r211, %r92, %r209;


dp4a.s32.s32 %r138, %r211, %r96, %r134;

shr.u16 %rs13, %rs11, 8;
cvt.u32.u16 %r300, %rs13;
mul.lo.s32 %r301, %r130, %r300;
cvt.rn.f32.s32 %f56, %r301;
fma.rn.ftz.f32 %f57, %f30, %f56, %f53;
shr.u16 %rs14, %rs12, 8;
cvt.u32.u16 %r302, %rs14;
mul.lo.s32 %r303, %r138, %r302;
cvt.rn.f32.s32 %f58, %r303;
fma.rn.ftz.f32 %f59, %f30, %f58, %f55;
ld.global.nc.u32 %r142, [%rd48];

{.reg .f16 low,high;
mov.b32 {low,high},%r142;
cvt.f32.f16 %f31, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r142;
cvt.f32.f16 %f32, high;}


mul.ftz.f32 %f60, %f57, %f31;
mul.ftz.f32 %f61, %f59, %f32;
sub.ftz.f32 %f62, %f60, %f61;
add.ftz.f32 %f151, %f62, %f151;
cvt.u64.u32 %rd49, %r397;
add.s64 %rd50, %rd3, %rd49;
mul.lo.s64 %rd51, %rd50, 36;
add.s64 %rd52, %rd13, %rd51;
ld.global.nc.u32 %r144, [%rd52];

{.reg .f16 low,high;
mov.b32 {low,high},%r144;
cvt.f32.f16 %f33, low;}


add.s64 %rd53, %rd52, %rd42;
ld.global.nc.u32 %r152, [%rd53+4];
ld.global.nc.u32 %r148, [%rd53+20];
ld.global.nc.u32 %r145, [%rd52+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r145;
cvt.f32.f16 %f34, low;}


ld.global.nc.u32 %r168, [%rd53+40];
ld.global.nc.u32 %r164, [%rd53+56];

dp4a.s32.s32 %r146, %r147, %r148, %r209;


dp4a.s32.s32 %r150, %r151, %r152, %r146;


dp4a.s32.s32 %r154, %r211, %r148, %r209;


dp4a.s32.s32 %r158, %r211, %r152, %r154;

mul.lo.s32 %r304, %r150, %r243;
cvt.rn.f32.s32 %f63, %r304;
fma.rn.ftz.f32 %f64, %f33, %f63, 0f00000000;
mul.lo.s32 %r305, %r158, %r246;
cvt.rn.f32.s32 %f65, %r305;
fma.rn.ftz.f32 %f66, %f33, %f65, 0f00000000;

dp4a.s32.s32 %r162, %r163, %r164, %r209;


dp4a.s32.s32 %r166, %r167, %r168, %r162;


dp4a.s32.s32 %r170, %r211, %r164, %r209;


dp4a.s32.s32 %r174, %r211, %r168, %r170;

mul.lo.s32 %r306, %r166, %r256;
cvt.rn.f32.s32 %f67, %r306;
fma.rn.ftz.f32 %f68, %f34, %f67, %f64;
mul.lo.s32 %r307, %r174, %r258;
cvt.rn.f32.s32 %f69, %r307;
fma.rn.ftz.f32 %f70, %f34, %f69, %f66;

{.reg .f16 low,high;
mov.b32 {low,high},%r106;
cvt.f32.f16 %f35, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r106;
cvt.f32.f16 %f36, high;}


mul.ftz.f32 %f71, %f68, %f35;
mul.ftz.f32 %f72, %f70, %f36;
sub.ftz.f32 %f73, %f71, %f72;
add.ftz.f32 %f150, %f73, %f150;

{.reg .f16 low,high;
mov.b32 {low,high},%r144;
cvt.f32.f16 %f37, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r145;
cvt.f32.f16 %f38, low;}



dp4a.s32.s32 %r182, %r183, %r148, %r209;


dp4a.s32.s32 %r186, %r187, %r152, %r182;


dp4a.s32.s32 %r190, %r211, %r148, %r209;


dp4a.s32.s32 %r194, %r211, %r152, %r190;

mul.lo.s32 %r308, %r186, %r287;
cvt.rn.f32.s32 %f74, %r308;
fma.rn.ftz.f32 %f75, %f37, %f74, 0f00000000;
mul.lo.s32 %r309, %r194, %r290;
cvt.rn.f32.s32 %f76, %r309;
fma.rn.ftz.f32 %f77, %f37, %f76, 0f00000000;

dp4a.s32.s32 %r198, %r199, %r164, %r209;


dp4a.s32.s32 %r202, %r203, %r168, %r198;


dp4a.s32.s32 %r206, %r211, %r164, %r209;


dp4a.s32.s32 %r210, %r211, %r168, %r206;

mul.lo.s32 %r310, %r202, %r300;
cvt.rn.f32.s32 %f78, %r310;
fma.rn.ftz.f32 %f79, %f38, %f78, %f75;
mul.lo.s32 %r311, %r210, %r302;
cvt.rn.f32.s32 %f80, %r311;
fma.rn.ftz.f32 %f81, %f38, %f80, %f77;

{.reg .f16 low,high;
mov.b32 {low,high},%r142;
cvt.f32.f16 %f39, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r142;
cvt.f32.f16 %f40, high;}


mul.ftz.f32 %f82, %f79, %f39;
mul.ftz.f32 %f83, %f81, %f40;
sub.ftz.f32 %f84, %f82, %f83;
add.ftz.f32 %f149, %f84, %f149;
add.s32 %r397, %r397, 64;
add.s32 %r396, %r396, 64;
add.s64 %rd63, %rd63, 1408;
add.s32 %r398, %r398, 8;
setp.lt.s32 %p5, %r398, %r5;
@%p5 bra $L__BB176_2;

$L__BB176_3:
setp.eq.s32 %p6, %r1, 0;
@%p6 bra $L__BB176_5;

shl.b32 %r313, %r1, 9;
mov.u32 %r314, _ZZ13mul_mat_vec_qIL9ggml_type13ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r315, %r314, %r313;
shl.b32 %r316, %r35, 2;
add.s32 %r317, %r315, %r316;
st.shared.f32 [%r317+-512], %f152;
st.shared.f32 [%r317+-384], %f151;
st.shared.f32 [%r317+-256], %f150;
st.shared.f32 [%r317+-128], %f149;

$L__BB176_5:
setp.ne.s32 %p7, %r1, 0;
bar.sync 0;
@%p7 bra $L__BB176_14;

mov.u32 %r318, %ctaid.z;
mov.u32 %r319, %ctaid.y;
mad.lo.s32 %r320, %r319, %r26, %r3;
mad.lo.s32 %r321, %r318, %r32, %r320;
cvt.u64.u32 %rd16, %r321;
shl.b32 %r323, %r35, 2;
mov.u32 %r324, _ZZ13mul_mat_vec_qIL9ggml_type13ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r16, %r324, %r323;
mov.u32 %r325, 2;
ld.shared.f32 %f85, [%r16];
add.ftz.f32 %f86, %f85, %f152;
ld.shared.f32 %f87, [%r16+512];
add.ftz.f32 %f88, %f87, %f86;
ld.shared.f32 %f89, [%r16+1024];
add.ftz.f32 %f90, %f89, %f88;
mov.b32 %r326, %f90;
mov.u32 %r327, 31;
mov.u32 %r328, 16;
mov.u32 %r329, -1;
shfl.sync.bfly.b32 %r330|%p8, %r326, %r328, %r327, %r329;
mov.b32 %f91, %r330;
add.ftz.f32 %f92, %f90, %f91;
mov.b32 %r331, %f92;
mov.u32 %r332, 8;
shfl.sync.bfly.b32 %r333|%p9, %r331, %r332, %r327, %r329;
mov.b32 %f93, %r333;
add.ftz.f32 %f94, %f92, %f93;
mov.b32 %r334, %f94;
mov.u32 %r335, 4;
shfl.sync.bfly.b32 %r336|%p10, %r334, %r335, %r327, %r329;
mov.b32 %f95, %r336;
add.ftz.f32 %f96, %f94, %f95;
mov.b32 %r337, %f96;
shfl.sync.bfly.b32 %r338|%p11, %r337, %r325, %r327, %r329;
mov.b32 %f97, %r338;
add.ftz.f32 %f98, %f96, %f97;
mov.b32 %r339, %f98;
mov.u32 %r340, 1;
shfl.sync.bfly.b32 %r341|%p12, %r339, %r340, %r327, %r329;
mov.b32 %f99, %r341;
add.ftz.f32 %f13, %f98, %f99;
setp.eq.s32 %p13, %r35, 0;
setp.lt.u32 %p14, %r3, %r20;
and.pred %p1, %p13, %p14;
mul.wide.u32 %rd54, %r321, 4;
add.s64 %rd17, %rd1, %rd54;
not.pred %p15, %p1;
@%p15 bra $L__BB176_8;

st.global.f32 [%rd17], %f13;

$L__BB176_8:
ld.shared.f32 %f100, [%r16+128];
add.ftz.f32 %f101, %f100, %f151;
ld.shared.f32 %f102, [%r16+640];
add.ftz.f32 %f103, %f102, %f101;
ld.shared.f32 %f104, [%r16+1152];
add.ftz.f32 %f105, %f104, %f103;
mov.b32 %r342, %f105;
shfl.sync.bfly.b32 %r346|%p16, %r342, %r328, %r327, %r329;
mov.b32 %f106, %r346;
add.ftz.f32 %f107, %f105, %f106;
mov.b32 %r347, %f107;
shfl.sync.bfly.b32 %r349|%p17, %r347, %r332, %r327, %r329;
mov.b32 %f108, %r349;
add.ftz.f32 %f109, %f107, %f108;
mov.b32 %r350, %f109;
shfl.sync.bfly.b32 %r352|%p18, %r350, %r335, %r327, %r329;
mov.b32 %f110, %r352;
add.ftz.f32 %f111, %f109, %f110;
mov.b32 %r353, %f111;
shfl.sync.bfly.b32 %r355|%p19, %r353, %r325, %r327, %r329;
mov.b32 %f112, %r355;
add.ftz.f32 %f113, %f111, %f112;
mov.b32 %r356, %f113;
shfl.sync.bfly.b32 %r358|%p20, %r356, %r340, %r327, %r329;
mov.b32 %f114, %r358;
add.ftz.f32 %f14, %f113, %f114;
setp.eq.s32 %p21, %r35, 1;
add.s32 %r360, %r3, 1;
setp.lt.u32 %p22, %r360, %r20;
and.pred %p2, %p21, %p22;
not.pred %p23, %p2;
@%p23 bra $L__BB176_10;

st.global.f32 [%rd17+4], %f14;

$L__BB176_10:
ld.shared.f32 %f115, [%r16+256];
add.ftz.f32 %f116, %f115, %f150;
ld.shared.f32 %f117, [%r16+768];
add.ftz.f32 %f118, %f117, %f116;
ld.shared.f32 %f119, [%r16+1280];
add.ftz.f32 %f120, %f119, %f118;
mov.b32 %r361, %f120;
mov.u32 %r362, 31;
mov.u32 %r363, 16;
mov.u32 %r364, -1;
shfl.sync.bfly.b32 %r365|%p24, %r361, %r363, %r362, %r364;
mov.b32 %f121, %r365;
add.ftz.f32 %f122, %f120, %f121;
mov.b32 %r366, %f122;
mov.u32 %r367, 8;
shfl.sync.bfly.b32 %r368|%p25, %r366, %r367, %r362, %r364;
mov.b32 %f123, %r368;
add.ftz.f32 %f124, %f122, %f123;
mov.b32 %r369, %f124;
mov.u32 %r370, 4;
shfl.sync.bfly.b32 %r371|%p26, %r369, %r370, %r362, %r364;
mov.b32 %f125, %r371;
add.ftz.f32 %f126, %f124, %f125;
mov.b32 %r372, %f126;
mov.u32 %r373, 2;
shfl.sync.bfly.b32 %r374|%p27, %r372, %r373, %r362, %r364;
mov.b32 %f127, %r374;
add.ftz.f32 %f128, %f126, %f127;
mov.b32 %r375, %f128;
mov.u32 %r376, 1;
shfl.sync.bfly.b32 %r377|%p28, %r375, %r376, %r362, %r364;
mov.b32 %f129, %r377;
add.ftz.f32 %f15, %f128, %f129;
@%p15 bra $L__BB176_12;

cvt.u64.u32 %rd55, %r20;
add.s64 %rd56, %rd55, %rd16;
shl.b64 %rd57, %rd56, 2;
add.s64 %rd58, %rd1, %rd57;
st.global.f32 [%rd58], %f15;

$L__BB176_12:
ld.shared.f32 %f130, [%r16+384];
add.ftz.f32 %f131, %f130, %f149;
ld.shared.f32 %f132, [%r16+896];
add.ftz.f32 %f133, %f132, %f131;
ld.shared.f32 %f134, [%r16+1408];
add.ftz.f32 %f135, %f134, %f133;
mov.b32 %r378, %f135;
shfl.sync.bfly.b32 %r382|%p30, %r378, %r363, %r362, %r364;
mov.b32 %f136, %r382;
add.ftz.f32 %f137, %f135, %f136;
mov.b32 %r383, %f137;
shfl.sync.bfly.b32 %r385|%p31, %r383, %r367, %r362, %r364;
mov.b32 %f138, %r385;
add.ftz.f32 %f139, %f137, %f138;
mov.b32 %r386, %f139;
shfl.sync.bfly.b32 %r388|%p32, %r386, %r370, %r362, %r364;
mov.b32 %f140, %r388;
add.ftz.f32 %f141, %f139, %f140;
mov.b32 %r389, %f141;
shfl.sync.bfly.b32 %r391|%p33, %r389, %r373, %r362, %r364;
mov.b32 %f142, %r391;
add.ftz.f32 %f143, %f141, %f142;
mov.b32 %r392, %f143;
shfl.sync.bfly.b32 %r394|%p34, %r392, %r376, %r362, %r364;
mov.b32 %f144, %r394;
add.ftz.f32 %f16, %f143, %f144;
@%p23 bra $L__BB176_14;

add.s32 %r395, %r20, 1;
cvt.u64.u32 %rd59, %r395;
add.s64 %rd60, %rd59, %rd16;
shl.b64 %rd61, %rd60, 2;
add.s64 %rd62, %rd1, %rd61;
st.global.f32 [%rd62], %f16;

$L__BB176_14:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type13ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type13ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type13ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type13ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type13ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type13ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type13ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type13ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type13ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<48>;
.reg .b16 %rs<15>;
.reg .f32 %f<229>;
.reg .b32 %r<518>;
.reg .b64 %rd<73>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type13ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[2304];

ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type13ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd20, [_Z13mul_mat_vec_qIL9ggml_type13ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd21, [_Z13mul_mat_vec_qIL9ggml_type13ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type13ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type13ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type13ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type13ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type13ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type13ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type13ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type13ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type13ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type13ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type13ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type13ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type13ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type13ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
cvta.to.global.u64 %rd1, %rd21;
mov.u32 %r1, %tid.y;
shl.b32 %r37, %r1, 5;
mov.u32 %r38, %tid.x;
add.s32 %r2, %r37, %r38;
mov.u32 %r39, %ctaid.x;
shl.b32 %r3, %r39, 1;
shr.s32 %r40, %r2, 31;
shr.u32 %r41, %r40, 28;
add.s32 %r42, %r2, %r41;
shr.s32 %r517, %r42, 4;
shr.u32 %r5, %r36, 8;
setp.ge.s32 %p3, %r517, %r5;
mov.f32 %f223, 0f00000000;
mov.f32 %f224, %f223;
mov.f32 %f225, %f223;
mov.f32 %f226, %f223;
mov.f32 %f227, %f223;
mov.f32 %f228, %f223;
@%p3 bra $L__BB177_3;

mov.u32 %r43, %ctaid.y;
mov.u32 %r44, %ctaid.z;
mul.hi.u32 %r45, %r44, %r30;
add.s32 %r46, %r45, %r44;
shr.u32 %r47, %r46, %r31;
mul.lo.s32 %r48, %r47, %r33;
mul.hi.u32 %r49, %r43, %r24;
add.s32 %r50, %r49, %r43;
shr.u32 %r51, %r50, %r25;
mul.lo.s32 %r52, %r51, %r27;
mul.lo.s32 %r53, %r43, %r28;
cvt.u64.u32 %rd22, %r53;
cvta.to.global.u64 %rd72, %rd19;
and.b32 %r57, %r42, -16;
sub.s32 %r6, %r2, %r57;
shr.s32 %r58, %r6, 31;
shr.u32 %r59, %r58, 30;
add.s32 %r60, %r6, %r59;
shr.s32 %r61, %r60, 2;
shl.b32 %r7, %r61, 1;
shl.b32 %r62, %r61, 5;
and.b32 %r63, %r60, -4;
sub.s32 %r64, %r6, %r63;
shl.b32 %r65, %r64, 2;
add.s32 %r66, %r62, %r65;
cvt.s64.s32 %rd23, %r66;
cvt.s64.s32 %rd24, %r65;
cvt.s64.s32 %rd25, %r7;
add.s64 %rd26, %rd22, %rd25;
mul.lo.s32 %r67, %r44, %r34;
cvt.u64.u32 %rd27, %r67;
add.s64 %rd3, %rd26, %rd27;
cvt.s64.s32 %rd4, %r64;
shl.b32 %r68, %r22, 1;
shl.b32 %r514, %r517, 3;
add.s32 %r516, %r514, %r68;
add.s32 %r515, %r22, %r514;
add.s32 %r69, %r517, %r48;
add.s32 %r70, %r69, %r52;
add.s32 %r71, %r3, 1;
mad.lo.s32 %r72, %r21, %r71, %r70;
mul.wide.s32 %rd5, %r72, 176;
mad.lo.s32 %r73, %r3, %r21, %r517;
add.s32 %r74, %r73, %r48;
add.s32 %r75, %r74, %r52;
mul.wide.s32 %rd6, %r75, 176;
add.s64 %rd28, %rd6, %rd23;
add.s64 %rd7, %rd28, 64;
cvt.u64.u32 %rd29, %r7;
and.b64 %rd30, %rd29, 2;
or.b64 %rd31, %rd30, %rd5;
or.b64 %rd8, %rd31, 8;
add.s64 %rd32, %rd6, %rd24;
add.s64 %rd9, %rd32, 32;
or.b64 %rd33, %rd30, %rd6;
or.b64 %rd10, %rd33, 12;
add.s64 %rd34, %rd5, %rd24;
add.s64 %rd11, %rd34, 16;
add.s64 %rd35, %rd5, %rd23;
add.s64 %rd12, %rd35, 48;
cvta.to.global.u64 %rd13, %rd20;
mov.f32 %f223, 0f00000000;
shl.b64 %rd43, %rd4, 2;

$L__BB177_2:
cvt.u64.u32 %rd36, %r514;
add.s64 %rd37, %rd3, %rd36;
add.s64 %rd38, %rd72, %rd7;
ld.global.nc.u32 %r292, [%rd38+-16];
ld.global.nc.u32 %r293, [%rd38];
add.s64 %rd39, %rd72, %rd9;
ld.global.nc.u32 %r294, [%rd39+-16];
shr.s32 %r295, %r294, %r7;
ld.global.nc.u32 %r296, [%rd39];
shr.s32 %r297, %r296, %r7;
add.s64 %rd40, %rd72, %rd10;
ld.global.nc.u16 %rs1, [%rd40+-8];
cvt.u32.u16 %r298, %rs1;
ld.global.nc.u16 %rs2, [%rd40+-4];
cvt.u32.u16 %r299, %rs2;
ld.global.nc.u16 %rs3, [%rd40];
cvt.u32.u16 %r300, %rs3;
and.b32 %r301, %r298, 16191;
and.b32 %r302, %r300, 3855;
and.b32 %r303, %r298, 49344;
shr.u32 %r304, %r303, 2;
or.b32 %r305, %r302, %r304;
setp.gt.s32 %p4, %r6, 7;
and.b32 %r306, %r299, 16191;
shr.u32 %r307, %r300, 4;
and.b32 %r308, %r307, 3855;
and.b32 %r309, %r299, 49344;
shr.u32 %r310, %r309, 2;
or.b32 %r311, %r308, %r310;
mul.lo.s64 %rd41, %rd37, 36;
add.s64 %rd42, %rd13, %rd41;
ld.global.nc.u32 %r76, [%rd42];

{.reg .f16 low,high;
mov.b32 {low,high},%r76;
cvt.f32.f16 %f37, low;}


add.s64 %rd44, %rd42, %rd43;
ld.global.nc.u32 %r84, [%rd44+4];
ld.global.nc.u32 %r80, [%rd44+20];
ld.global.nc.u32 %r77, [%rd42+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r77;
cvt.f32.f16 %f38, low;}


ld.global.nc.u32 %r100, [%rd44+40];
ld.global.nc.u32 %r96, [%rd44+56];
add.s64 %rd45, %rd72, %rd6;
and.b32 %r312, %r292, 252645135;
and.b32 %r313, %r293, 252645135;
shl.b32 %r314, %r295, 4;
and.b32 %r315, %r314, 269488144;
shl.b32 %r316, %r297, 4;
and.b32 %r317, %r316, 269488144;
or.b32 %r227, %r312, %r315;
or.b32 %r223, %r313, %r317;
mov.u32 %r285, 0;

dp4a.s32.s32 %r78, %r223, %r80, %r285;


dp4a.s32.s32 %r82, %r227, %r84, %r78;

mov.u32 %r287, 16843009;

dp4a.s32.s32 %r86, %r287, %r80, %r285;


dp4a.s32.s32 %r90, %r287, %r84, %r86;

selp.b32 %r318, %r305, %r301, %p4;
cvt.u16.u32 %rs4, %r318;
and.b32 %r319, %r318, 63;
mul.lo.s32 %r320, %r82, %r319;
cvt.rn.f32.s32 %f61, %r320;
fma.rn.ftz.f32 %f62, %f37, %f61, 0f00000000;
selp.b32 %r321, %r311, %r306, %p4;
cvt.u16.u32 %rs5, %r321;
and.b32 %r322, %r321, 63;
mul.lo.s32 %r323, %r90, %r322;
cvt.rn.f32.s32 %f63, %r323;
fma.rn.ftz.f32 %f64, %f37, %f63, 0f00000000;
shr.u32 %r324, %r292, 4;
and.b32 %r325, %r324, 252645135;
shr.u32 %r326, %r293, 4;
and.b32 %r327, %r326, 252645135;
shl.b32 %r328, %r295, 3;
and.b32 %r329, %r328, 269488144;
shl.b32 %r330, %r297, 3;
and.b32 %r331, %r330, 269488144;
or.b32 %r243, %r325, %r329;
or.b32 %r239, %r327, %r331;

dp4a.s32.s32 %r94, %r239, %r96, %r285;


dp4a.s32.s32 %r98, %r243, %r100, %r94;


dp4a.s32.s32 %r102, %r287, %r96, %r285;


dp4a.s32.s32 %r106, %r287, %r100, %r102;

shr.u16 %rs6, %rs4, 8;
cvt.u32.u16 %r332, %rs6;
mul.lo.s32 %r333, %r98, %r332;
cvt.rn.f32.s32 %f65, %r333;
fma.rn.ftz.f32 %f66, %f38, %f65, %f62;
shr.u16 %rs7, %rs5, 8;
cvt.u32.u16 %r334, %rs7;
mul.lo.s32 %r335, %r106, %r334;
cvt.rn.f32.s32 %f67, %r335;
fma.rn.ftz.f32 %f68, %f38, %f67, %f64;
ld.global.nc.u32 %r110, [%rd45];

{.reg .f16 low,high;
mov.b32 {low,high},%r110;
cvt.f32.f16 %f39, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r110;
cvt.f32.f16 %f40, high;}


mul.ftz.f32 %f69, %f66, %f39;
mul.ftz.f32 %f70, %f68, %f40;
sub.ftz.f32 %f71, %f69, %f70;
add.ftz.f32 %f228, %f71, %f228;
add.s64 %rd46, %rd72, %rd12;
add.s64 %rd47, %rd72, %rd11;
ld.global.nc.u32 %r336, [%rd46];
ld.global.nc.u32 %r337, [%rd46+16];
ld.global.nc.u32 %r338, [%rd47];
shr.s32 %r339, %r338, %r7;
ld.global.nc.u32 %r340, [%rd47+16];
shr.s32 %r341, %r340, %r7;
add.s64 %rd48, %rd72, %rd8;
ld.global.nc.u16 %rs8, [%rd48+-4];
cvt.u32.u16 %r342, %rs8;
ld.global.nc.u16 %rs9, [%rd48];
cvt.u32.u16 %r343, %rs9;
ld.global.nc.u16 %rs10, [%rd48+4];
cvt.u32.u16 %r344, %rs10;
and.b32 %r345, %r342, 16191;
and.b32 %r346, %r344, 3855;
and.b32 %r347, %r342, 49344;
shr.u32 %r348, %r347, 2;
or.b32 %r349, %r346, %r348;
and.b32 %r350, %r343, 16191;
shr.u32 %r351, %r344, 4;
and.b32 %r352, %r351, 3855;
and.b32 %r353, %r343, 49344;
shr.u32 %r354, %r353, 2;
or.b32 %r355, %r352, %r354;

{.reg .f16 low,high;
mov.b32 {low,high},%r76;
cvt.f32.f16 %f41, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r77;
cvt.f32.f16 %f42, low;}


add.s64 %rd49, %rd72, %rd5;
and.b32 %r356, %r336, 252645135;
and.b32 %r357, %r337, 252645135;
shl.b32 %r358, %r339, 4;
and.b32 %r359, %r358, 269488144;
shl.b32 %r360, %r341, 4;
and.b32 %r361, %r360, 269488144;
or.b32 %r263, %r356, %r359;
or.b32 %r259, %r357, %r361;

dp4a.s32.s32 %r114, %r259, %r80, %r285;


dp4a.s32.s32 %r118, %r263, %r84, %r114;


dp4a.s32.s32 %r122, %r287, %r80, %r285;


dp4a.s32.s32 %r126, %r287, %r84, %r122;

selp.b32 %r362, %r349, %r345, %p4;
cvt.u16.u32 %rs11, %r362;
and.b32 %r363, %r362, 63;
mul.lo.s32 %r364, %r118, %r363;
cvt.rn.f32.s32 %f72, %r364;
fma.rn.ftz.f32 %f73, %f41, %f72, 0f00000000;
selp.b32 %r365, %r355, %r350, %p4;
cvt.u16.u32 %rs12, %r365;
and.b32 %r366, %r365, 63;
mul.lo.s32 %r367, %r126, %r366;
cvt.rn.f32.s32 %f74, %r367;
fma.rn.ftz.f32 %f75, %f41, %f74, 0f00000000;
shr.u32 %r368, %r336, 4;
and.b32 %r369, %r368, 252645135;
shr.u32 %r370, %r337, 4;
and.b32 %r371, %r370, 252645135;
shl.b32 %r372, %r339, 3;
and.b32 %r373, %r372, 269488144;
shl.b32 %r374, %r341, 3;
and.b32 %r375, %r374, 269488144;
or.b32 %r279, %r369, %r373;
or.b32 %r275, %r371, %r375;

dp4a.s32.s32 %r130, %r275, %r96, %r285;


dp4a.s32.s32 %r134, %r279, %r100, %r130;


dp4a.s32.s32 %r138, %r287, %r96, %r285;


dp4a.s32.s32 %r142, %r287, %r100, %r138;

shr.u16 %rs13, %rs11, 8;
cvt.u32.u16 %r376, %rs13;
mul.lo.s32 %r377, %r134, %r376;
cvt.rn.f32.s32 %f76, %r377;
fma.rn.ftz.f32 %f77, %f42, %f76, %f73;
shr.u16 %rs14, %rs12, 8;
cvt.u32.u16 %r378, %rs14;
mul.lo.s32 %r379, %r142, %r378;
cvt.rn.f32.s32 %f78, %r379;
fma.rn.ftz.f32 %f79, %f42, %f78, %f75;
ld.global.nc.u32 %r146, [%rd49];

{.reg .f16 low,high;
mov.b32 {low,high},%r146;
cvt.f32.f16 %f43, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r146;
cvt.f32.f16 %f44, high;}


mul.ftz.f32 %f80, %f77, %f43;
mul.ftz.f32 %f81, %f79, %f44;
sub.ftz.f32 %f82, %f80, %f81;
add.ftz.f32 %f227, %f82, %f227;
cvt.u64.u32 %rd50, %r515;
add.s64 %rd51, %rd3, %rd50;
mul.lo.s64 %rd52, %rd51, 36;
add.s64 %rd53, %rd13, %rd52;
ld.global.nc.u32 %r148, [%rd53];

{.reg .f16 low,high;
mov.b32 {low,high},%r148;
cvt.f32.f16 %f45, low;}


add.s64 %rd54, %rd53, %rd43;
ld.global.nc.u32 %r156, [%rd54+4];
ld.global.nc.u32 %r152, [%rd54+20];
ld.global.nc.u32 %r149, [%rd53+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r149;
cvt.f32.f16 %f46, low;}


ld.global.nc.u32 %r172, [%rd54+40];
ld.global.nc.u32 %r168, [%rd54+56];

dp4a.s32.s32 %r150, %r223, %r152, %r285;


dp4a.s32.s32 %r154, %r227, %r156, %r150;


dp4a.s32.s32 %r158, %r287, %r152, %r285;


dp4a.s32.s32 %r162, %r287, %r156, %r158;

mul.lo.s32 %r380, %r154, %r319;
cvt.rn.f32.s32 %f83, %r380;
fma.rn.ftz.f32 %f84, %f45, %f83, 0f00000000;
mul.lo.s32 %r381, %r162, %r322;
cvt.rn.f32.s32 %f85, %r381;
fma.rn.ftz.f32 %f86, %f45, %f85, 0f00000000;

dp4a.s32.s32 %r166, %r239, %r168, %r285;


dp4a.s32.s32 %r170, %r243, %r172, %r166;


dp4a.s32.s32 %r174, %r287, %r168, %r285;


dp4a.s32.s32 %r178, %r287, %r172, %r174;

mul.lo.s32 %r382, %r170, %r332;
cvt.rn.f32.s32 %f87, %r382;
fma.rn.ftz.f32 %f88, %f46, %f87, %f84;
mul.lo.s32 %r383, %r178, %r334;
cvt.rn.f32.s32 %f89, %r383;
fma.rn.ftz.f32 %f90, %f46, %f89, %f86;

{.reg .f16 low,high;
mov.b32 {low,high},%r110;
cvt.f32.f16 %f47, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r110;
cvt.f32.f16 %f48, high;}


mul.ftz.f32 %f91, %f88, %f47;
mul.ftz.f32 %f92, %f90, %f48;
sub.ftz.f32 %f93, %f91, %f92;
add.ftz.f32 %f226, %f93, %f226;

{.reg .f16 low,high;
mov.b32 {low,high},%r148;
cvt.f32.f16 %f49, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r149;
cvt.f32.f16 %f50, low;}



dp4a.s32.s32 %r186, %r259, %r152, %r285;


dp4a.s32.s32 %r190, %r263, %r156, %r186;


dp4a.s32.s32 %r194, %r287, %r152, %r285;


dp4a.s32.s32 %r198, %r287, %r156, %r194;

mul.lo.s32 %r384, %r190, %r363;
cvt.rn.f32.s32 %f94, %r384;
fma.rn.ftz.f32 %f95, %f49, %f94, 0f00000000;
mul.lo.s32 %r385, %r198, %r366;
cvt.rn.f32.s32 %f96, %r385;
fma.rn.ftz.f32 %f97, %f49, %f96, 0f00000000;

dp4a.s32.s32 %r202, %r275, %r168, %r285;


dp4a.s32.s32 %r206, %r279, %r172, %r202;


dp4a.s32.s32 %r210, %r287, %r168, %r285;


dp4a.s32.s32 %r214, %r287, %r172, %r210;

mul.lo.s32 %r386, %r206, %r376;
cvt.rn.f32.s32 %f98, %r386;
fma.rn.ftz.f32 %f99, %f50, %f98, %f95;
mul.lo.s32 %r387, %r214, %r378;
cvt.rn.f32.s32 %f100, %r387;
fma.rn.ftz.f32 %f101, %f50, %f100, %f97;

{.reg .f16 low,high;
mov.b32 {low,high},%r146;
cvt.f32.f16 %f51, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r146;
cvt.f32.f16 %f52, high;}


mul.ftz.f32 %f102, %f99, %f51;
mul.ftz.f32 %f103, %f101, %f52;
sub.ftz.f32 %f104, %f102, %f103;
add.ftz.f32 %f225, %f104, %f225;
cvt.u64.u32 %rd55, %r516;
add.s64 %rd56, %rd3, %rd55;
mul.lo.s64 %rd57, %rd56, 36;
add.s64 %rd58, %rd13, %rd57;
ld.global.nc.u32 %r220, [%rd58];

{.reg .f16 low,high;
mov.b32 {low,high},%r220;
cvt.f32.f16 %f53, low;}


add.s64 %rd59, %rd58, %rd43;
ld.global.nc.u32 %r228, [%rd59+4];
ld.global.nc.u32 %r224, [%rd59+20];
ld.global.nc.u32 %r221, [%rd58+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r221;
cvt.f32.f16 %f54, low;}


ld.global.nc.u32 %r244, [%rd59+40];
ld.global.nc.u32 %r240, [%rd59+56];

dp4a.s32.s32 %r222, %r223, %r224, %r285;


dp4a.s32.s32 %r226, %r227, %r228, %r222;


dp4a.s32.s32 %r230, %r287, %r224, %r285;


dp4a.s32.s32 %r234, %r287, %r228, %r230;

mul.lo.s32 %r388, %r226, %r319;
cvt.rn.f32.s32 %f105, %r388;
fma.rn.ftz.f32 %f106, %f53, %f105, 0f00000000;
mul.lo.s32 %r389, %r234, %r322;
cvt.rn.f32.s32 %f107, %r389;
fma.rn.ftz.f32 %f108, %f53, %f107, 0f00000000;

dp4a.s32.s32 %r238, %r239, %r240, %r285;


dp4a.s32.s32 %r242, %r243, %r244, %r238;


dp4a.s32.s32 %r246, %r287, %r240, %r285;


dp4a.s32.s32 %r250, %r287, %r244, %r246;

mul.lo.s32 %r390, %r242, %r332;
cvt.rn.f32.s32 %f109, %r390;
fma.rn.ftz.f32 %f110, %f54, %f109, %f106;
mul.lo.s32 %r391, %r250, %r334;
cvt.rn.f32.s32 %f111, %r391;
fma.rn.ftz.f32 %f112, %f54, %f111, %f108;

{.reg .f16 low,high;
mov.b32 {low,high},%r110;
cvt.f32.f16 %f55, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r110;
cvt.f32.f16 %f56, high;}


mul.ftz.f32 %f113, %f110, %f55;
mul.ftz.f32 %f114, %f112, %f56;
sub.ftz.f32 %f115, %f113, %f114;
add.ftz.f32 %f224, %f115, %f224;

{.reg .f16 low,high;
mov.b32 {low,high},%r220;
cvt.f32.f16 %f57, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r221;
cvt.f32.f16 %f58, low;}



dp4a.s32.s32 %r258, %r259, %r224, %r285;


dp4a.s32.s32 %r262, %r263, %r228, %r258;


dp4a.s32.s32 %r266, %r287, %r224, %r285;


dp4a.s32.s32 %r270, %r287, %r228, %r266;

mul.lo.s32 %r392, %r262, %r363;
cvt.rn.f32.s32 %f116, %r392;
fma.rn.ftz.f32 %f117, %f57, %f116, 0f00000000;
mul.lo.s32 %r393, %r270, %r366;
cvt.rn.f32.s32 %f118, %r393;
fma.rn.ftz.f32 %f119, %f57, %f118, 0f00000000;

dp4a.s32.s32 %r274, %r275, %r240, %r285;


dp4a.s32.s32 %r278, %r279, %r244, %r274;


dp4a.s32.s32 %r282, %r287, %r240, %r285;


dp4a.s32.s32 %r286, %r287, %r244, %r282;

mul.lo.s32 %r394, %r278, %r376;
cvt.rn.f32.s32 %f120, %r394;
fma.rn.ftz.f32 %f121, %f58, %f120, %f117;
mul.lo.s32 %r395, %r286, %r378;
cvt.rn.f32.s32 %f122, %r395;
fma.rn.ftz.f32 %f123, %f58, %f122, %f119;

{.reg .f16 low,high;
mov.b32 {low,high},%r146;
cvt.f32.f16 %f59, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r146;
cvt.f32.f16 %f60, high;}


mul.ftz.f32 %f124, %f121, %f59;
mul.ftz.f32 %f125, %f123, %f60;
sub.ftz.f32 %f126, %f124, %f125;
add.ftz.f32 %f223, %f126, %f223;
add.s32 %r516, %r516, 64;
add.s32 %r515, %r515, 64;
add.s32 %r514, %r514, 64;
add.s64 %rd72, %rd72, 1408;
add.s32 %r517, %r517, 8;
setp.lt.s32 %p5, %r517, %r5;
@%p5 bra $L__BB177_2;

$L__BB177_3:
setp.eq.s32 %p6, %r1, 0;
@%p6 bra $L__BB177_5;

mov.u32 %r397, _ZZ13mul_mat_vec_qIL9ggml_type13ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r398, %r1, 768, %r397;
shl.b32 %r399, %r38, 2;
add.s32 %r400, %r398, %r399;
st.shared.f32 [%r400+-768], %f228;
st.shared.f32 [%r400+-640], %f227;
st.shared.f32 [%r400+-512], %f226;
st.shared.f32 [%r400+-384], %f225;
st.shared.f32 [%r400+-256], %f224;
st.shared.f32 [%r400+-128], %f223;

$L__BB177_5:
setp.ne.s32 %p7, %r1, 0;
bar.sync 0;
@%p7 bra $L__BB177_18;

mov.u32 %r401, %ctaid.z;
mov.u32 %r402, %ctaid.y;
mad.lo.s32 %r403, %r402, %r29, %r3;
mad.lo.s32 %r404, %r401, %r35, %r403;
cvt.u64.u32 %rd16, %r404;
shl.b32 %r406, %r38, 2;
mov.u32 %r407, _ZZ13mul_mat_vec_qIL9ggml_type13ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r19, %r407, %r406;
mov.u32 %r408, 2;
ld.shared.f32 %f127, [%r19];
add.ftz.f32 %f128, %f127, %f228;
ld.shared.f32 %f129, [%r19+768];
add.ftz.f32 %f130, %f129, %f128;
ld.shared.f32 %f131, [%r19+1536];
add.ftz.f32 %f132, %f131, %f130;
mov.b32 %r409, %f132;
mov.u32 %r410, 31;
mov.u32 %r411, 16;
mov.u32 %r412, -1;
shfl.sync.bfly.b32 %r413|%p8, %r409, %r411, %r410, %r412;
mov.b32 %f133, %r413;
add.ftz.f32 %f134, %f132, %f133;
mov.b32 %r414, %f134;
mov.u32 %r415, 8;
shfl.sync.bfly.b32 %r416|%p9, %r414, %r415, %r410, %r412;
mov.b32 %f135, %r416;
add.ftz.f32 %f136, %f134, %f135;
mov.b32 %r417, %f136;
mov.u32 %r418, 4;
shfl.sync.bfly.b32 %r419|%p10, %r417, %r418, %r410, %r412;
mov.b32 %f137, %r419;
add.ftz.f32 %f138, %f136, %f137;
mov.b32 %r420, %f138;
shfl.sync.bfly.b32 %r421|%p11, %r420, %r408, %r410, %r412;
mov.b32 %f139, %r421;
add.ftz.f32 %f140, %f138, %f139;
mov.b32 %r422, %f140;
mov.u32 %r423, 1;
shfl.sync.bfly.b32 %r424|%p12, %r422, %r423, %r410, %r412;
mov.b32 %f141, %r424;
add.ftz.f32 %f19, %f140, %f141;
setp.eq.s32 %p13, %r38, 0;
setp.lt.u32 %p14, %r3, %r23;
and.pred %p1, %p13, %p14;
mul.wide.u32 %rd60, %r404, 4;
add.s64 %rd17, %rd1, %rd60;
not.pred %p15, %p1;
@%p15 bra $L__BB177_8;

st.global.f32 [%rd17], %f19;

$L__BB177_8:
ld.shared.f32 %f142, [%r19+128];
add.ftz.f32 %f143, %f142, %f227;
ld.shared.f32 %f144, [%r19+896];
add.ftz.f32 %f145, %f144, %f143;
ld.shared.f32 %f146, [%r19+1664];
add.ftz.f32 %f147, %f146, %f145;
mov.b32 %r425, %f147;
shfl.sync.bfly.b32 %r429|%p16, %r425, %r411, %r410, %r412;
mov.b32 %f148, %r429;
add.ftz.f32 %f149, %f147, %f148;
mov.b32 %r430, %f149;
shfl.sync.bfly.b32 %r432|%p17, %r430, %r415, %r410, %r412;
mov.b32 %f150, %r432;
add.ftz.f32 %f151, %f149, %f150;
mov.b32 %r433, %f151;
shfl.sync.bfly.b32 %r435|%p18, %r433, %r418, %r410, %r412;
mov.b32 %f152, %r435;
add.ftz.f32 %f153, %f151, %f152;
mov.b32 %r436, %f153;
shfl.sync.bfly.b32 %r438|%p19, %r436, %r408, %r410, %r412;
mov.b32 %f154, %r438;
add.ftz.f32 %f155, %f153, %f154;
mov.b32 %r439, %f155;
shfl.sync.bfly.b32 %r441|%p20, %r439, %r423, %r410, %r412;
mov.b32 %f156, %r441;
add.ftz.f32 %f20, %f155, %f156;
setp.eq.s32 %p21, %r38, 1;
add.s32 %r443, %r3, 1;
setp.lt.u32 %p22, %r443, %r23;
and.pred %p2, %p21, %p22;
not.pred %p23, %p2;
@%p23 bra $L__BB177_10;

st.global.f32 [%rd17+4], %f20;

$L__BB177_10:
ld.shared.f32 %f157, [%r19+256];
add.ftz.f32 %f158, %f157, %f226;
ld.shared.f32 %f159, [%r19+1024];
add.ftz.f32 %f160, %f159, %f158;
ld.shared.f32 %f161, [%r19+1792];
add.ftz.f32 %f162, %f161, %f160;
mov.b32 %r444, %f162;
mov.u32 %r445, 31;
mov.u32 %r446, 16;
mov.u32 %r447, -1;
shfl.sync.bfly.b32 %r448|%p24, %r444, %r446, %r445, %r447;
mov.b32 %f163, %r448;
add.ftz.f32 %f164, %f162, %f163;
mov.b32 %r449, %f164;
mov.u32 %r450, 8;
shfl.sync.bfly.b32 %r451|%p25, %r449, %r450, %r445, %r447;
mov.b32 %f165, %r451;
add.ftz.f32 %f166, %f164, %f165;
mov.b32 %r452, %f166;
mov.u32 %r453, 4;
shfl.sync.bfly.b32 %r454|%p26, %r452, %r453, %r445, %r447;
mov.b32 %f167, %r454;
add.ftz.f32 %f168, %f166, %f167;
mov.b32 %r455, %f168;
mov.u32 %r456, 2;
shfl.sync.bfly.b32 %r457|%p27, %r455, %r456, %r445, %r447;
mov.b32 %f169, %r457;
add.ftz.f32 %f170, %f168, %f169;
mov.b32 %r458, %f170;
mov.u32 %r459, 1;
shfl.sync.bfly.b32 %r460|%p28, %r458, %r459, %r445, %r447;
mov.b32 %f171, %r460;
add.ftz.f32 %f21, %f170, %f171;
@%p15 bra $L__BB177_12;

cvt.u64.u32 %rd61, %r23;
add.s64 %rd62, %rd61, %rd16;
shl.b64 %rd63, %rd62, 2;
add.s64 %rd64, %rd1, %rd63;
st.global.f32 [%rd64], %f21;

$L__BB177_12:
ld.shared.f32 %f172, [%r19+384];
add.ftz.f32 %f173, %f172, %f225;
ld.shared.f32 %f174, [%r19+1152];
add.ftz.f32 %f175, %f174, %f173;
ld.shared.f32 %f176, [%r19+1920];
add.ftz.f32 %f177, %f176, %f175;
mov.b32 %r461, %f177;
shfl.sync.bfly.b32 %r465|%p30, %r461, %r446, %r445, %r447;
mov.b32 %f178, %r465;
add.ftz.f32 %f179, %f177, %f178;
mov.b32 %r466, %f179;
shfl.sync.bfly.b32 %r468|%p31, %r466, %r450, %r445, %r447;
mov.b32 %f180, %r468;
add.ftz.f32 %f181, %f179, %f180;
mov.b32 %r469, %f181;
shfl.sync.bfly.b32 %r471|%p32, %r469, %r453, %r445, %r447;
mov.b32 %f182, %r471;
add.ftz.f32 %f183, %f181, %f182;
mov.b32 %r472, %f183;
shfl.sync.bfly.b32 %r474|%p33, %r472, %r456, %r445, %r447;
mov.b32 %f184, %r474;
add.ftz.f32 %f185, %f183, %f184;
mov.b32 %r475, %f185;
shfl.sync.bfly.b32 %r477|%p34, %r475, %r459, %r445, %r447;
mov.b32 %f186, %r477;
add.ftz.f32 %f22, %f185, %f186;
@%p23 bra $L__BB177_14;

add.s32 %r478, %r23, 1;
cvt.u64.u32 %rd65, %r478;
add.s64 %rd66, %rd65, %rd16;
shl.b64 %rd67, %rd66, 2;
add.s64 %rd68, %rd1, %rd67;
st.global.f32 [%rd68], %f22;

$L__BB177_14:
shl.b32 %r479, %r23, 1;
mov.u32 %r480, 1;
ld.shared.f32 %f187, [%r19+512];
add.ftz.f32 %f188, %f187, %f224;
ld.shared.f32 %f189, [%r19+1280];
add.ftz.f32 %f190, %f189, %f188;
ld.shared.f32 %f191, [%r19+2048];
add.ftz.f32 %f192, %f191, %f190;
mov.b32 %r481, %f192;
mov.u32 %r482, 31;
mov.u32 %r483, 16;
mov.u32 %r484, -1;
shfl.sync.bfly.b32 %r485|%p36, %r481, %r483, %r482, %r484;
mov.b32 %f193, %r485;
add.ftz.f32 %f194, %f192, %f193;
mov.b32 %r486, %f194;
mov.u32 %r487, 8;
shfl.sync.bfly.b32 %r488|%p37, %r486, %r487, %r482, %r484;
mov.b32 %f195, %r488;
add.ftz.f32 %f196, %f194, %f195;
mov.b32 %r489, %f196;
mov.u32 %r490, 4;
shfl.sync.bfly.b32 %r491|%p38, %r489, %r490, %r482, %r484;
mov.b32 %f197, %r491;
add.ftz.f32 %f198, %f196, %f197;
mov.b32 %r492, %f198;
mov.u32 %r493, 2;
shfl.sync.bfly.b32 %r494|%p39, %r492, %r493, %r482, %r484;
mov.b32 %f199, %r494;
add.ftz.f32 %f200, %f198, %f199;
mov.b32 %r495, %f200;
shfl.sync.bfly.b32 %r496|%p40, %r495, %r480, %r482, %r484;
mov.b32 %f201, %r496;
add.ftz.f32 %f23, %f200, %f201;
cvt.u64.u32 %rd69, %r479;
add.s64 %rd70, %rd69, %rd16;
shl.b64 %rd71, %rd70, 2;
add.s64 %rd18, %rd1, %rd71;
@%p15 bra $L__BB177_16;

st.global.f32 [%rd18], %f23;

$L__BB177_16:
ld.shared.f32 %f202, [%r19+640];
add.ftz.f32 %f203, %f202, %f223;
ld.shared.f32 %f204, [%r19+1408];
add.ftz.f32 %f205, %f204, %f203;
ld.shared.f32 %f206, [%r19+2176];
add.ftz.f32 %f207, %f206, %f205;
mov.b32 %r497, %f207;
shfl.sync.bfly.b32 %r501|%p42, %r497, %r483, %r482, %r484;
mov.b32 %f208, %r501;
add.ftz.f32 %f209, %f207, %f208;
mov.b32 %r502, %f209;
shfl.sync.bfly.b32 %r504|%p43, %r502, %r487, %r482, %r484;
mov.b32 %f210, %r504;
add.ftz.f32 %f211, %f209, %f210;
mov.b32 %r505, %f211;
shfl.sync.bfly.b32 %r507|%p44, %r505, %r490, %r482, %r484;
mov.b32 %f212, %r507;
add.ftz.f32 %f213, %f211, %f212;
mov.b32 %r508, %f213;
shfl.sync.bfly.b32 %r510|%p45, %r508, %r493, %r482, %r484;
mov.b32 %f214, %r510;
add.ftz.f32 %f215, %f213, %f214;
mov.b32 %r511, %f215;
shfl.sync.bfly.b32 %r513|%p46, %r511, %r480, %r482, %r484;
mov.b32 %f216, %r513;
add.ftz.f32 %f24, %f215, %f216;
@%p23 bra $L__BB177_18;

st.global.f32 [%rd18+4], %f24;

$L__BB177_18:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type13ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type13ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type13ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type13ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type13ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type13ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type13ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type13ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type13ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<60>;
.reg .b16 %rs<15>;
.reg .f32 %f<305>;
.reg .b32 %r<639>;
.reg .b64 %rd<86>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type13ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[3072];

ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type13ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd20, [_Z13mul_mat_vec_qIL9ggml_type13ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd21, [_Z13mul_mat_vec_qIL9ggml_type13ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type13ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type13ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type13ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type13ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type13ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type13ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type13ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type13ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type13ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type13ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type13ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type13ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type13ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type13ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
cvta.to.global.u64 %rd1, %rd21;
mov.u32 %r1, %tid.y;
shl.b32 %r41, %r1, 5;
mov.u32 %r42, %tid.x;
add.s32 %r2, %r41, %r42;
mov.u32 %r43, %ctaid.x;
shl.b32 %r3, %r43, 1;
shr.s32 %r44, %r2, 31;
shr.u32 %r45, %r44, 28;
add.s32 %r46, %r2, %r45;
shr.s32 %r638, %r46, 4;
shr.u32 %r5, %r40, 8;
setp.ge.s32 %p3, %r638, %r5;
mov.f32 %f297, 0f00000000;
mov.f32 %f298, %f297;
mov.f32 %f299, %f297;
mov.f32 %f300, %f297;
mov.f32 %f301, %f297;
mov.f32 %f302, %f297;
mov.f32 %f303, %f297;
mov.f32 %f304, %f297;
@%p3 bra $L__BB178_3;

mov.u32 %r47, %ctaid.y;
mov.u32 %r48, %ctaid.z;
mul.hi.u32 %r49, %r48, %r34;
add.s32 %r50, %r49, %r48;
shr.u32 %r51, %r50, %r35;
mul.lo.s32 %r52, %r51, %r37;
mul.hi.u32 %r53, %r47, %r28;
add.s32 %r54, %r53, %r47;
shr.u32 %r55, %r54, %r29;
mul.lo.s32 %r56, %r55, %r31;
mul.lo.s32 %r57, %r47, %r32;
cvt.u64.u32 %rd22, %r57;
cvta.to.global.u64 %rd85, %rd19;
and.b32 %r61, %r46, -16;
sub.s32 %r6, %r2, %r61;
shr.s32 %r62, %r6, 31;
shr.u32 %r63, %r62, 30;
add.s32 %r64, %r6, %r63;
shr.s32 %r65, %r64, 2;
shl.b32 %r7, %r65, 1;
shl.b32 %r66, %r65, 5;
and.b32 %r67, %r64, -4;
sub.s32 %r68, %r6, %r67;
shl.b32 %r69, %r68, 2;
add.s32 %r70, %r66, %r69;
cvt.s64.s32 %rd23, %r70;
cvt.s64.s32 %rd24, %r69;
cvt.s64.s32 %rd25, %r7;
add.s64 %rd26, %rd22, %rd25;
mul.lo.s32 %r71, %r48, %r38;
cvt.u64.u32 %rd27, %r71;
add.s64 %rd3, %rd26, %rd27;
cvt.s64.s32 %rd4, %r68;
shl.b32 %r72, %r26, 1;
shl.b32 %r634, %r638, 3;
add.s32 %r637, %r634, %r72;
mad.lo.s32 %r636, %r26, 3, %r634;
add.s32 %r635, %r26, %r634;
add.s32 %r73, %r638, %r52;
add.s32 %r74, %r73, %r56;
add.s32 %r75, %r3, 1;
mad.lo.s32 %r76, %r25, %r75, %r74;
mul.wide.s32 %rd5, %r76, 176;
mad.lo.s32 %r77, %r3, %r25, %r638;
add.s32 %r78, %r77, %r52;
add.s32 %r79, %r78, %r56;
mul.wide.s32 %rd6, %r79, 176;
add.s64 %rd28, %rd6, %rd23;
add.s64 %rd7, %rd28, 64;
cvt.u64.u32 %rd29, %r7;
and.b64 %rd30, %rd29, 2;
or.b64 %rd31, %rd30, %rd5;
or.b64 %rd8, %rd31, 8;
add.s64 %rd32, %rd6, %rd24;
add.s64 %rd9, %rd32, 32;
or.b64 %rd33, %rd30, %rd6;
or.b64 %rd10, %rd33, 12;
add.s64 %rd34, %rd5, %rd24;
add.s64 %rd11, %rd34, 16;
add.s64 %rd35, %rd5, %rd23;
add.s64 %rd12, %rd35, 48;
cvta.to.global.u64 %rd13, %rd20;
mov.f32 %f297, 0f00000000;
shl.b64 %rd43, %rd4, 2;

$L__BB178_2:
cvt.u64.u32 %rd36, %r634;
add.s64 %rd37, %rd3, %rd36;
add.s64 %rd38, %rd85, %rd7;
ld.global.nc.u32 %r368, [%rd38+-16];
ld.global.nc.u32 %r369, [%rd38];
add.s64 %rd39, %rd85, %rd9;
ld.global.nc.u32 %r370, [%rd39+-16];
shr.s32 %r371, %r370, %r7;
ld.global.nc.u32 %r372, [%rd39];
shr.s32 %r373, %r372, %r7;
add.s64 %rd40, %rd85, %rd10;
ld.global.nc.u16 %rs1, [%rd40+-8];
cvt.u32.u16 %r374, %rs1;
ld.global.nc.u16 %rs2, [%rd40+-4];
cvt.u32.u16 %r375, %rs2;
ld.global.nc.u16 %rs3, [%rd40];
cvt.u32.u16 %r376, %rs3;
and.b32 %r377, %r374, 16191;
and.b32 %r378, %r376, 3855;
and.b32 %r379, %r374, 49344;
shr.u32 %r380, %r379, 2;
or.b32 %r381, %r378, %r380;
setp.gt.s32 %p4, %r6, 7;
and.b32 %r382, %r375, 16191;
shr.u32 %r383, %r376, 4;
and.b32 %r384, %r383, 3855;
and.b32 %r385, %r375, 49344;
shr.u32 %r386, %r385, 2;
or.b32 %r387, %r384, %r386;
mul.lo.s64 %rd41, %rd37, 36;
add.s64 %rd42, %rd13, %rd41;
ld.global.nc.u32 %r80, [%rd42];

{.reg .f16 low,high;
mov.b32 {low,high},%r80;
cvt.f32.f16 %f49, low;}


add.s64 %rd44, %rd42, %rd43;
ld.global.nc.u32 %r88, [%rd44+4];
ld.global.nc.u32 %r84, [%rd44+20];
ld.global.nc.u32 %r81, [%rd42+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r81;
cvt.f32.f16 %f50, low;}


ld.global.nc.u32 %r104, [%rd44+40];
ld.global.nc.u32 %r100, [%rd44+56];
add.s64 %rd45, %rd85, %rd6;
and.b32 %r388, %r368, 252645135;
and.b32 %r389, %r369, 252645135;
shl.b32 %r390, %r371, 4;
and.b32 %r391, %r390, 269488144;
shl.b32 %r392, %r373, 4;
and.b32 %r393, %r392, 269488144;
or.b32 %r303, %r388, %r391;
or.b32 %r299, %r389, %r393;
mov.u32 %r361, 0;

dp4a.s32.s32 %r82, %r299, %r84, %r361;


dp4a.s32.s32 %r86, %r303, %r88, %r82;

mov.u32 %r363, 16843009;

dp4a.s32.s32 %r90, %r363, %r84, %r361;


dp4a.s32.s32 %r94, %r363, %r88, %r90;

selp.b32 %r394, %r381, %r377, %p4;
cvt.u16.u32 %rs4, %r394;
and.b32 %r395, %r394, 63;
mul.lo.s32 %r396, %r86, %r395;
cvt.rn.f32.s32 %f81, %r396;
fma.rn.ftz.f32 %f82, %f49, %f81, 0f00000000;
selp.b32 %r397, %r387, %r382, %p4;
cvt.u16.u32 %rs5, %r397;
and.b32 %r398, %r397, 63;
mul.lo.s32 %r399, %r94, %r398;
cvt.rn.f32.s32 %f83, %r399;
fma.rn.ftz.f32 %f84, %f49, %f83, 0f00000000;
shr.u32 %r400, %r368, 4;
and.b32 %r401, %r400, 252645135;
shr.u32 %r402, %r369, 4;
and.b32 %r403, %r402, 252645135;
shl.b32 %r404, %r371, 3;
and.b32 %r405, %r404, 269488144;
shl.b32 %r406, %r373, 3;
and.b32 %r407, %r406, 269488144;
or.b32 %r319, %r401, %r405;
or.b32 %r315, %r403, %r407;

dp4a.s32.s32 %r98, %r315, %r100, %r361;


dp4a.s32.s32 %r102, %r319, %r104, %r98;


dp4a.s32.s32 %r106, %r363, %r100, %r361;


dp4a.s32.s32 %r110, %r363, %r104, %r106;

shr.u16 %rs6, %rs4, 8;
cvt.u32.u16 %r408, %rs6;
mul.lo.s32 %r409, %r102, %r408;
cvt.rn.f32.s32 %f85, %r409;
fma.rn.ftz.f32 %f86, %f50, %f85, %f82;
shr.u16 %rs7, %rs5, 8;
cvt.u32.u16 %r410, %rs7;
mul.lo.s32 %r411, %r110, %r410;
cvt.rn.f32.s32 %f87, %r411;
fma.rn.ftz.f32 %f88, %f50, %f87, %f84;
ld.global.nc.u32 %r114, [%rd45];

{.reg .f16 low,high;
mov.b32 {low,high},%r114;
cvt.f32.f16 %f51, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r114;
cvt.f32.f16 %f52, high;}


mul.ftz.f32 %f89, %f86, %f51;
mul.ftz.f32 %f90, %f88, %f52;
sub.ftz.f32 %f91, %f89, %f90;
add.ftz.f32 %f304, %f91, %f304;
add.s64 %rd46, %rd85, %rd12;
add.s64 %rd47, %rd85, %rd11;
ld.global.nc.u32 %r412, [%rd46];
ld.global.nc.u32 %r413, [%rd46+16];
ld.global.nc.u32 %r414, [%rd47];
shr.s32 %r415, %r414, %r7;
ld.global.nc.u32 %r416, [%rd47+16];
shr.s32 %r417, %r416, %r7;
add.s64 %rd48, %rd85, %rd8;
ld.global.nc.u16 %rs8, [%rd48+-4];
cvt.u32.u16 %r418, %rs8;
ld.global.nc.u16 %rs9, [%rd48];
cvt.u32.u16 %r419, %rs9;
ld.global.nc.u16 %rs10, [%rd48+4];
cvt.u32.u16 %r420, %rs10;
and.b32 %r421, %r418, 16191;
and.b32 %r422, %r420, 3855;
and.b32 %r423, %r418, 49344;
shr.u32 %r424, %r423, 2;
or.b32 %r425, %r422, %r424;
and.b32 %r426, %r419, 16191;
shr.u32 %r427, %r420, 4;
and.b32 %r428, %r427, 3855;
and.b32 %r429, %r419, 49344;
shr.u32 %r430, %r429, 2;
or.b32 %r431, %r428, %r430;

{.reg .f16 low,high;
mov.b32 {low,high},%r80;
cvt.f32.f16 %f53, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r81;
cvt.f32.f16 %f54, low;}


add.s64 %rd49, %rd85, %rd5;
and.b32 %r432, %r412, 252645135;
and.b32 %r433, %r413, 252645135;
shl.b32 %r434, %r415, 4;
and.b32 %r435, %r434, 269488144;
shl.b32 %r436, %r417, 4;
and.b32 %r437, %r436, 269488144;
or.b32 %r339, %r432, %r435;
or.b32 %r335, %r433, %r437;

dp4a.s32.s32 %r118, %r335, %r84, %r361;


dp4a.s32.s32 %r122, %r339, %r88, %r118;


dp4a.s32.s32 %r126, %r363, %r84, %r361;


dp4a.s32.s32 %r130, %r363, %r88, %r126;

selp.b32 %r438, %r425, %r421, %p4;
cvt.u16.u32 %rs11, %r438;
and.b32 %r439, %r438, 63;
mul.lo.s32 %r440, %r122, %r439;
cvt.rn.f32.s32 %f92, %r440;
fma.rn.ftz.f32 %f93, %f53, %f92, 0f00000000;
selp.b32 %r441, %r431, %r426, %p4;
cvt.u16.u32 %rs12, %r441;
and.b32 %r442, %r441, 63;
mul.lo.s32 %r443, %r130, %r442;
cvt.rn.f32.s32 %f94, %r443;
fma.rn.ftz.f32 %f95, %f53, %f94, 0f00000000;
shr.u32 %r444, %r412, 4;
and.b32 %r445, %r444, 252645135;
shr.u32 %r446, %r413, 4;
and.b32 %r447, %r446, 252645135;
shl.b32 %r448, %r415, 3;
and.b32 %r449, %r448, 269488144;
shl.b32 %r450, %r417, 3;
and.b32 %r451, %r450, 269488144;
or.b32 %r355, %r445, %r449;
or.b32 %r351, %r447, %r451;

dp4a.s32.s32 %r134, %r351, %r100, %r361;


dp4a.s32.s32 %r138, %r355, %r104, %r134;


dp4a.s32.s32 %r142, %r363, %r100, %r361;


dp4a.s32.s32 %r146, %r363, %r104, %r142;

shr.u16 %rs13, %rs11, 8;
cvt.u32.u16 %r452, %rs13;
mul.lo.s32 %r453, %r138, %r452;
cvt.rn.f32.s32 %f96, %r453;
fma.rn.ftz.f32 %f97, %f54, %f96, %f93;
shr.u16 %rs14, %rs12, 8;
cvt.u32.u16 %r454, %rs14;
mul.lo.s32 %r455, %r146, %r454;
cvt.rn.f32.s32 %f98, %r455;
fma.rn.ftz.f32 %f99, %f54, %f98, %f95;
ld.global.nc.u32 %r150, [%rd49];

{.reg .f16 low,high;
mov.b32 {low,high},%r150;
cvt.f32.f16 %f55, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r150;
cvt.f32.f16 %f56, high;}


mul.ftz.f32 %f100, %f97, %f55;
mul.ftz.f32 %f101, %f99, %f56;
sub.ftz.f32 %f102, %f100, %f101;
add.ftz.f32 %f303, %f102, %f303;
cvt.u64.u32 %rd50, %r635;
add.s64 %rd51, %rd3, %rd50;
mul.lo.s64 %rd52, %rd51, 36;
add.s64 %rd53, %rd13, %rd52;
ld.global.nc.u32 %r152, [%rd53];

{.reg .f16 low,high;
mov.b32 {low,high},%r152;
cvt.f32.f16 %f57, low;}


add.s64 %rd54, %rd53, %rd43;
ld.global.nc.u32 %r160, [%rd54+4];
ld.global.nc.u32 %r156, [%rd54+20];
ld.global.nc.u32 %r153, [%rd53+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r153;
cvt.f32.f16 %f58, low;}


ld.global.nc.u32 %r176, [%rd54+40];
ld.global.nc.u32 %r172, [%rd54+56];

dp4a.s32.s32 %r154, %r299, %r156, %r361;


dp4a.s32.s32 %r158, %r303, %r160, %r154;


dp4a.s32.s32 %r162, %r363, %r156, %r361;


dp4a.s32.s32 %r166, %r363, %r160, %r162;

mul.lo.s32 %r456, %r158, %r395;
cvt.rn.f32.s32 %f103, %r456;
fma.rn.ftz.f32 %f104, %f57, %f103, 0f00000000;
mul.lo.s32 %r457, %r166, %r398;
cvt.rn.f32.s32 %f105, %r457;
fma.rn.ftz.f32 %f106, %f57, %f105, 0f00000000;

dp4a.s32.s32 %r170, %r315, %r172, %r361;


dp4a.s32.s32 %r174, %r319, %r176, %r170;


dp4a.s32.s32 %r178, %r363, %r172, %r361;


dp4a.s32.s32 %r182, %r363, %r176, %r178;

mul.lo.s32 %r458, %r174, %r408;
cvt.rn.f32.s32 %f107, %r458;
fma.rn.ftz.f32 %f108, %f58, %f107, %f104;
mul.lo.s32 %r459, %r182, %r410;
cvt.rn.f32.s32 %f109, %r459;
fma.rn.ftz.f32 %f110, %f58, %f109, %f106;

{.reg .f16 low,high;
mov.b32 {low,high},%r114;
cvt.f32.f16 %f59, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r114;
cvt.f32.f16 %f60, high;}


mul.ftz.f32 %f111, %f108, %f59;
mul.ftz.f32 %f112, %f110, %f60;
sub.ftz.f32 %f113, %f111, %f112;
add.ftz.f32 %f302, %f113, %f302;

{.reg .f16 low,high;
mov.b32 {low,high},%r152;
cvt.f32.f16 %f61, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r153;
cvt.f32.f16 %f62, low;}



dp4a.s32.s32 %r190, %r335, %r156, %r361;


dp4a.s32.s32 %r194, %r339, %r160, %r190;


dp4a.s32.s32 %r198, %r363, %r156, %r361;


dp4a.s32.s32 %r202, %r363, %r160, %r198;

mul.lo.s32 %r460, %r194, %r439;
cvt.rn.f32.s32 %f114, %r460;
fma.rn.ftz.f32 %f115, %f61, %f114, 0f00000000;
mul.lo.s32 %r461, %r202, %r442;
cvt.rn.f32.s32 %f116, %r461;
fma.rn.ftz.f32 %f117, %f61, %f116, 0f00000000;

dp4a.s32.s32 %r206, %r351, %r172, %r361;


dp4a.s32.s32 %r210, %r355, %r176, %r206;


dp4a.s32.s32 %r214, %r363, %r172, %r361;


dp4a.s32.s32 %r218, %r363, %r176, %r214;

mul.lo.s32 %r462, %r210, %r452;
cvt.rn.f32.s32 %f118, %r462;
fma.rn.ftz.f32 %f119, %f62, %f118, %f115;
mul.lo.s32 %r463, %r218, %r454;
cvt.rn.f32.s32 %f120, %r463;
fma.rn.ftz.f32 %f121, %f62, %f120, %f117;

{.reg .f16 low,high;
mov.b32 {low,high},%r150;
cvt.f32.f16 %f63, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r150;
cvt.f32.f16 %f64, high;}


mul.ftz.f32 %f122, %f119, %f63;
mul.ftz.f32 %f123, %f121, %f64;
sub.ftz.f32 %f124, %f122, %f123;
add.ftz.f32 %f301, %f124, %f301;
cvt.u64.u32 %rd55, %r637;
add.s64 %rd56, %rd3, %rd55;
mul.lo.s64 %rd57, %rd56, 36;
add.s64 %rd58, %rd13, %rd57;
ld.global.nc.u32 %r224, [%rd58];

{.reg .f16 low,high;
mov.b32 {low,high},%r224;
cvt.f32.f16 %f65, low;}


add.s64 %rd59, %rd58, %rd43;
ld.global.nc.u32 %r232, [%rd59+4];
ld.global.nc.u32 %r228, [%rd59+20];
ld.global.nc.u32 %r225, [%rd58+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r225;
cvt.f32.f16 %f66, low;}


ld.global.nc.u32 %r248, [%rd59+40];
ld.global.nc.u32 %r244, [%rd59+56];

dp4a.s32.s32 %r226, %r299, %r228, %r361;


dp4a.s32.s32 %r230, %r303, %r232, %r226;


dp4a.s32.s32 %r234, %r363, %r228, %r361;


dp4a.s32.s32 %r238, %r363, %r232, %r234;

mul.lo.s32 %r464, %r230, %r395;
cvt.rn.f32.s32 %f125, %r464;
fma.rn.ftz.f32 %f126, %f65, %f125, 0f00000000;
mul.lo.s32 %r465, %r238, %r398;
cvt.rn.f32.s32 %f127, %r465;
fma.rn.ftz.f32 %f128, %f65, %f127, 0f00000000;

dp4a.s32.s32 %r242, %r315, %r244, %r361;


dp4a.s32.s32 %r246, %r319, %r248, %r242;


dp4a.s32.s32 %r250, %r363, %r244, %r361;


dp4a.s32.s32 %r254, %r363, %r248, %r250;

mul.lo.s32 %r466, %r246, %r408;
cvt.rn.f32.s32 %f129, %r466;
fma.rn.ftz.f32 %f130, %f66, %f129, %f126;
mul.lo.s32 %r467, %r254, %r410;
cvt.rn.f32.s32 %f131, %r467;
fma.rn.ftz.f32 %f132, %f66, %f131, %f128;

{.reg .f16 low,high;
mov.b32 {low,high},%r114;
cvt.f32.f16 %f67, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r114;
cvt.f32.f16 %f68, high;}


mul.ftz.f32 %f133, %f130, %f67;
mul.ftz.f32 %f134, %f132, %f68;
sub.ftz.f32 %f135, %f133, %f134;
add.ftz.f32 %f300, %f135, %f300;

{.reg .f16 low,high;
mov.b32 {low,high},%r224;
cvt.f32.f16 %f69, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r225;
cvt.f32.f16 %f70, low;}



dp4a.s32.s32 %r262, %r335, %r228, %r361;


dp4a.s32.s32 %r266, %r339, %r232, %r262;


dp4a.s32.s32 %r270, %r363, %r228, %r361;


dp4a.s32.s32 %r274, %r363, %r232, %r270;

mul.lo.s32 %r468, %r266, %r439;
cvt.rn.f32.s32 %f136, %r468;
fma.rn.ftz.f32 %f137, %f69, %f136, 0f00000000;
mul.lo.s32 %r469, %r274, %r442;
cvt.rn.f32.s32 %f138, %r469;
fma.rn.ftz.f32 %f139, %f69, %f138, 0f00000000;

dp4a.s32.s32 %r278, %r351, %r244, %r361;


dp4a.s32.s32 %r282, %r355, %r248, %r278;


dp4a.s32.s32 %r286, %r363, %r244, %r361;


dp4a.s32.s32 %r290, %r363, %r248, %r286;

mul.lo.s32 %r470, %r282, %r452;
cvt.rn.f32.s32 %f140, %r470;
fma.rn.ftz.f32 %f141, %f70, %f140, %f137;
mul.lo.s32 %r471, %r290, %r454;
cvt.rn.f32.s32 %f142, %r471;
fma.rn.ftz.f32 %f143, %f70, %f142, %f139;

{.reg .f16 low,high;
mov.b32 {low,high},%r150;
cvt.f32.f16 %f71, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r150;
cvt.f32.f16 %f72, high;}


mul.ftz.f32 %f144, %f141, %f71;
mul.ftz.f32 %f145, %f143, %f72;
sub.ftz.f32 %f146, %f144, %f145;
add.ftz.f32 %f299, %f146, %f299;
cvt.u64.u32 %rd60, %r636;
add.s64 %rd61, %rd3, %rd60;
mul.lo.s64 %rd62, %rd61, 36;
add.s64 %rd63, %rd13, %rd62;
ld.global.nc.u32 %r296, [%rd63];

{.reg .f16 low,high;
mov.b32 {low,high},%r296;
cvt.f32.f16 %f73, low;}


add.s64 %rd64, %rd63, %rd43;
ld.global.nc.u32 %r304, [%rd64+4];
ld.global.nc.u32 %r300, [%rd64+20];
ld.global.nc.u32 %r297, [%rd63+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r297;
cvt.f32.f16 %f74, low;}


ld.global.nc.u32 %r320, [%rd64+40];
ld.global.nc.u32 %r316, [%rd64+56];

dp4a.s32.s32 %r298, %r299, %r300, %r361;


dp4a.s32.s32 %r302, %r303, %r304, %r298;


dp4a.s32.s32 %r306, %r363, %r300, %r361;


dp4a.s32.s32 %r310, %r363, %r304, %r306;

mul.lo.s32 %r472, %r302, %r395;
cvt.rn.f32.s32 %f147, %r472;
fma.rn.ftz.f32 %f148, %f73, %f147, 0f00000000;
mul.lo.s32 %r473, %r310, %r398;
cvt.rn.f32.s32 %f149, %r473;
fma.rn.ftz.f32 %f150, %f73, %f149, 0f00000000;

dp4a.s32.s32 %r314, %r315, %r316, %r361;


dp4a.s32.s32 %r318, %r319, %r320, %r314;


dp4a.s32.s32 %r322, %r363, %r316, %r361;


dp4a.s32.s32 %r326, %r363, %r320, %r322;

mul.lo.s32 %r474, %r318, %r408;
cvt.rn.f32.s32 %f151, %r474;
fma.rn.ftz.f32 %f152, %f74, %f151, %f148;
mul.lo.s32 %r475, %r326, %r410;
cvt.rn.f32.s32 %f153, %r475;
fma.rn.ftz.f32 %f154, %f74, %f153, %f150;

{.reg .f16 low,high;
mov.b32 {low,high},%r114;
cvt.f32.f16 %f75, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r114;
cvt.f32.f16 %f76, high;}


mul.ftz.f32 %f155, %f152, %f75;
mul.ftz.f32 %f156, %f154, %f76;
sub.ftz.f32 %f157, %f155, %f156;
add.ftz.f32 %f298, %f157, %f298;

{.reg .f16 low,high;
mov.b32 {low,high},%r296;
cvt.f32.f16 %f77, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r297;
cvt.f32.f16 %f78, low;}



dp4a.s32.s32 %r334, %r335, %r300, %r361;


dp4a.s32.s32 %r338, %r339, %r304, %r334;


dp4a.s32.s32 %r342, %r363, %r300, %r361;


dp4a.s32.s32 %r346, %r363, %r304, %r342;

mul.lo.s32 %r476, %r338, %r439;
cvt.rn.f32.s32 %f158, %r476;
fma.rn.ftz.f32 %f159, %f77, %f158, 0f00000000;
mul.lo.s32 %r477, %r346, %r442;
cvt.rn.f32.s32 %f160, %r477;
fma.rn.ftz.f32 %f161, %f77, %f160, 0f00000000;

dp4a.s32.s32 %r350, %r351, %r316, %r361;


dp4a.s32.s32 %r354, %r355, %r320, %r350;


dp4a.s32.s32 %r358, %r363, %r316, %r361;


dp4a.s32.s32 %r362, %r363, %r320, %r358;

mul.lo.s32 %r478, %r354, %r452;
cvt.rn.f32.s32 %f162, %r478;
fma.rn.ftz.f32 %f163, %f78, %f162, %f159;
mul.lo.s32 %r479, %r362, %r454;
cvt.rn.f32.s32 %f164, %r479;
fma.rn.ftz.f32 %f165, %f78, %f164, %f161;

{.reg .f16 low,high;
mov.b32 {low,high},%r150;
cvt.f32.f16 %f79, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r150;
cvt.f32.f16 %f80, high;}


mul.ftz.f32 %f166, %f163, %f79;
mul.ftz.f32 %f167, %f165, %f80;
sub.ftz.f32 %f168, %f166, %f167;
add.ftz.f32 %f297, %f168, %f297;
add.s32 %r637, %r637, 64;
add.s32 %r636, %r636, 64;
add.s32 %r635, %r635, 64;
add.s32 %r634, %r634, 64;
add.s64 %rd85, %rd85, 1408;
add.s32 %r638, %r638, 8;
setp.lt.s32 %p5, %r638, %r5;
@%p5 bra $L__BB178_2;

$L__BB178_3:
setp.eq.s32 %p6, %r1, 0;
@%p6 bra $L__BB178_5;

shl.b32 %r481, %r1, 10;
mov.u32 %r482, _ZZ13mul_mat_vec_qIL9ggml_type13ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r483, %r482, %r481;
shl.b32 %r484, %r42, 2;
add.s32 %r485, %r483, %r484;
st.shared.f32 [%r485+-1024], %f304;
st.shared.f32 [%r485+-896], %f303;
st.shared.f32 [%r485+-768], %f302;
st.shared.f32 [%r485+-640], %f301;
st.shared.f32 [%r485+-512], %f300;
st.shared.f32 [%r485+-384], %f299;
st.shared.f32 [%r485+-256], %f298;
st.shared.f32 [%r485+-128], %f297;

$L__BB178_5:
setp.ne.s32 %p7, %r1, 0;
bar.sync 0;
@%p7 bra $L__BB178_22;

mov.u32 %r486, %ctaid.z;
mov.u32 %r487, %ctaid.y;
mad.lo.s32 %r488, %r487, %r33, %r3;
mad.lo.s32 %r489, %r486, %r39, %r488;
cvt.u64.u32 %rd16, %r489;
shl.b32 %r491, %r42, 2;
mov.u32 %r492, _ZZ13mul_mat_vec_qIL9ggml_type13ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r22, %r492, %r491;
mov.u32 %r493, 2;
ld.shared.f32 %f169, [%r22];
add.ftz.f32 %f170, %f169, %f304;
ld.shared.f32 %f171, [%r22+1024];
add.ftz.f32 %f172, %f171, %f170;
ld.shared.f32 %f173, [%r22+2048];
add.ftz.f32 %f174, %f173, %f172;
mov.b32 %r494, %f174;
mov.u32 %r495, 31;
mov.u32 %r496, 16;
mov.u32 %r497, -1;
shfl.sync.bfly.b32 %r498|%p8, %r494, %r496, %r495, %r497;
mov.b32 %f175, %r498;
add.ftz.f32 %f176, %f174, %f175;
mov.b32 %r499, %f176;
mov.u32 %r500, 8;
shfl.sync.bfly.b32 %r501|%p9, %r499, %r500, %r495, %r497;
mov.b32 %f177, %r501;
add.ftz.f32 %f178, %f176, %f177;
mov.b32 %r502, %f178;
mov.u32 %r503, 4;
shfl.sync.bfly.b32 %r504|%p10, %r502, %r503, %r495, %r497;
mov.b32 %f179, %r504;
add.ftz.f32 %f180, %f178, %f179;
mov.b32 %r505, %f180;
shfl.sync.bfly.b32 %r506|%p11, %r505, %r493, %r495, %r497;
mov.b32 %f181, %r506;
add.ftz.f32 %f182, %f180, %f181;
mov.b32 %r507, %f182;
mov.u32 %r508, 1;
shfl.sync.bfly.b32 %r509|%p12, %r507, %r508, %r495, %r497;
mov.b32 %f183, %r509;
add.ftz.f32 %f25, %f182, %f183;
setp.eq.s32 %p13, %r42, 0;
setp.lt.u32 %p14, %r3, %r27;
and.pred %p1, %p13, %p14;
mul.wide.u32 %rd65, %r489, 4;
add.s64 %rd17, %rd1, %rd65;
not.pred %p15, %p1;
@%p15 bra $L__BB178_8;

st.global.f32 [%rd17], %f25;

$L__BB178_8:
ld.shared.f32 %f184, [%r22+128];
add.ftz.f32 %f185, %f184, %f303;
ld.shared.f32 %f186, [%r22+1152];
add.ftz.f32 %f187, %f186, %f185;
ld.shared.f32 %f188, [%r22+2176];
add.ftz.f32 %f189, %f188, %f187;
mov.b32 %r510, %f189;
shfl.sync.bfly.b32 %r514|%p16, %r510, %r496, %r495, %r497;
mov.b32 %f190, %r514;
add.ftz.f32 %f191, %f189, %f190;
mov.b32 %r515, %f191;
shfl.sync.bfly.b32 %r517|%p17, %r515, %r500, %r495, %r497;
mov.b32 %f192, %r517;
add.ftz.f32 %f193, %f191, %f192;
mov.b32 %r518, %f193;
shfl.sync.bfly.b32 %r520|%p18, %r518, %r503, %r495, %r497;
mov.b32 %f194, %r520;
add.ftz.f32 %f195, %f193, %f194;
mov.b32 %r521, %f195;
shfl.sync.bfly.b32 %r523|%p19, %r521, %r493, %r495, %r497;
mov.b32 %f196, %r523;
add.ftz.f32 %f197, %f195, %f196;
mov.b32 %r524, %f197;
shfl.sync.bfly.b32 %r526|%p20, %r524, %r508, %r495, %r497;
mov.b32 %f198, %r526;
add.ftz.f32 %f26, %f197, %f198;
setp.eq.s32 %p21, %r42, 1;
add.s32 %r528, %r3, 1;
setp.lt.u32 %p22, %r528, %r27;
and.pred %p2, %p21, %p22;
not.pred %p23, %p2;
@%p23 bra $L__BB178_10;

st.global.f32 [%rd17+4], %f26;

$L__BB178_10:
ld.shared.f32 %f199, [%r22+256];
add.ftz.f32 %f200, %f199, %f302;
ld.shared.f32 %f201, [%r22+1280];
add.ftz.f32 %f202, %f201, %f200;
ld.shared.f32 %f203, [%r22+2304];
add.ftz.f32 %f204, %f203, %f202;
mov.b32 %r529, %f204;
mov.u32 %r530, 31;
mov.u32 %r531, 16;
mov.u32 %r532, -1;
shfl.sync.bfly.b32 %r533|%p24, %r529, %r531, %r530, %r532;
mov.b32 %f205, %r533;
add.ftz.f32 %f206, %f204, %f205;
mov.b32 %r534, %f206;
mov.u32 %r535, 8;
shfl.sync.bfly.b32 %r536|%p25, %r534, %r535, %r530, %r532;
mov.b32 %f207, %r536;
add.ftz.f32 %f208, %f206, %f207;
mov.b32 %r537, %f208;
mov.u32 %r538, 4;
shfl.sync.bfly.b32 %r539|%p26, %r537, %r538, %r530, %r532;
mov.b32 %f209, %r539;
add.ftz.f32 %f210, %f208, %f209;
mov.b32 %r540, %f210;
mov.u32 %r541, 2;
shfl.sync.bfly.b32 %r542|%p27, %r540, %r541, %r530, %r532;
mov.b32 %f211, %r542;
add.ftz.f32 %f212, %f210, %f211;
mov.b32 %r543, %f212;
mov.u32 %r544, 1;
shfl.sync.bfly.b32 %r545|%p28, %r543, %r544, %r530, %r532;
mov.b32 %f213, %r545;
add.ftz.f32 %f27, %f212, %f213;
@%p15 bra $L__BB178_12;

cvt.u64.u32 %rd66, %r27;
add.s64 %rd67, %rd66, %rd16;
shl.b64 %rd68, %rd67, 2;
add.s64 %rd69, %rd1, %rd68;
st.global.f32 [%rd69], %f27;

$L__BB178_12:
ld.shared.f32 %f214, [%r22+384];
add.ftz.f32 %f215, %f214, %f301;
ld.shared.f32 %f216, [%r22+1408];
add.ftz.f32 %f217, %f216, %f215;
ld.shared.f32 %f218, [%r22+2432];
add.ftz.f32 %f219, %f218, %f217;
mov.b32 %r546, %f219;
shfl.sync.bfly.b32 %r550|%p30, %r546, %r531, %r530, %r532;
mov.b32 %f220, %r550;
add.ftz.f32 %f221, %f219, %f220;
mov.b32 %r551, %f221;
shfl.sync.bfly.b32 %r553|%p31, %r551, %r535, %r530, %r532;
mov.b32 %f222, %r553;
add.ftz.f32 %f223, %f221, %f222;
mov.b32 %r554, %f223;
shfl.sync.bfly.b32 %r556|%p32, %r554, %r538, %r530, %r532;
mov.b32 %f224, %r556;
add.ftz.f32 %f225, %f223, %f224;
mov.b32 %r557, %f225;
shfl.sync.bfly.b32 %r559|%p33, %r557, %r541, %r530, %r532;
mov.b32 %f226, %r559;
add.ftz.f32 %f227, %f225, %f226;
mov.b32 %r560, %f227;
shfl.sync.bfly.b32 %r562|%p34, %r560, %r544, %r530, %r532;
mov.b32 %f228, %r562;
add.ftz.f32 %f28, %f227, %f228;
@%p23 bra $L__BB178_14;

add.s32 %r563, %r27, 1;
cvt.u64.u32 %rd70, %r563;
add.s64 %rd71, %rd70, %rd16;
shl.b64 %rd72, %rd71, 2;
add.s64 %rd73, %rd1, %rd72;
st.global.f32 [%rd73], %f28;

$L__BB178_14:
shl.b32 %r564, %r27, 1;
mov.u32 %r565, 1;
ld.shared.f32 %f229, [%r22+512];
add.ftz.f32 %f230, %f229, %f300;
ld.shared.f32 %f231, [%r22+1536];
add.ftz.f32 %f232, %f231, %f230;
ld.shared.f32 %f233, [%r22+2560];
add.ftz.f32 %f234, %f233, %f232;
mov.b32 %r566, %f234;
mov.u32 %r567, 31;
mov.u32 %r568, 16;
mov.u32 %r569, -1;
shfl.sync.bfly.b32 %r570|%p36, %r566, %r568, %r567, %r569;
mov.b32 %f235, %r570;
add.ftz.f32 %f236, %f234, %f235;
mov.b32 %r571, %f236;
mov.u32 %r572, 8;
shfl.sync.bfly.b32 %r573|%p37, %r571, %r572, %r567, %r569;
mov.b32 %f237, %r573;
add.ftz.f32 %f238, %f236, %f237;
mov.b32 %r574, %f238;
mov.u32 %r575, 4;
shfl.sync.bfly.b32 %r576|%p38, %r574, %r575, %r567, %r569;
mov.b32 %f239, %r576;
add.ftz.f32 %f240, %f238, %f239;
mov.b32 %r577, %f240;
mov.u32 %r578, 2;
shfl.sync.bfly.b32 %r579|%p39, %r577, %r578, %r567, %r569;
mov.b32 %f241, %r579;
add.ftz.f32 %f242, %f240, %f241;
mov.b32 %r580, %f242;
shfl.sync.bfly.b32 %r581|%p40, %r580, %r565, %r567, %r569;
mov.b32 %f243, %r581;
add.ftz.f32 %f29, %f242, %f243;
cvt.u64.u32 %rd74, %r564;
add.s64 %rd75, %rd74, %rd16;
shl.b64 %rd76, %rd75, 2;
add.s64 %rd18, %rd1, %rd76;
@%p15 bra $L__BB178_16;

st.global.f32 [%rd18], %f29;

$L__BB178_16:
ld.shared.f32 %f244, [%r22+640];
add.ftz.f32 %f245, %f244, %f299;
ld.shared.f32 %f246, [%r22+1664];
add.ftz.f32 %f247, %f246, %f245;
ld.shared.f32 %f248, [%r22+2688];
add.ftz.f32 %f249, %f248, %f247;
mov.b32 %r582, %f249;
shfl.sync.bfly.b32 %r586|%p42, %r582, %r568, %r567, %r569;
mov.b32 %f250, %r586;
add.ftz.f32 %f251, %f249, %f250;
mov.b32 %r587, %f251;
shfl.sync.bfly.b32 %r589|%p43, %r587, %r572, %r567, %r569;
mov.b32 %f252, %r589;
add.ftz.f32 %f253, %f251, %f252;
mov.b32 %r590, %f253;
shfl.sync.bfly.b32 %r592|%p44, %r590, %r575, %r567, %r569;
mov.b32 %f254, %r592;
add.ftz.f32 %f255, %f253, %f254;
mov.b32 %r593, %f255;
shfl.sync.bfly.b32 %r595|%p45, %r593, %r578, %r567, %r569;
mov.b32 %f256, %r595;
add.ftz.f32 %f257, %f255, %f256;
mov.b32 %r596, %f257;
shfl.sync.bfly.b32 %r598|%p46, %r596, %r565, %r567, %r569;
mov.b32 %f258, %r598;
add.ftz.f32 %f30, %f257, %f258;
@%p23 bra $L__BB178_18;

st.global.f32 [%rd18+4], %f30;

$L__BB178_18:
mul.lo.s32 %r24, %r27, 3;
ld.shared.f32 %f259, [%r22+768];
add.ftz.f32 %f260, %f259, %f298;
ld.shared.f32 %f261, [%r22+1792];
add.ftz.f32 %f262, %f261, %f260;
ld.shared.f32 %f263, [%r22+2816];
add.ftz.f32 %f264, %f263, %f262;
mov.b32 %r599, %f264;
mov.u32 %r600, 31;
mov.u32 %r601, 16;
mov.u32 %r602, -1;
shfl.sync.bfly.b32 %r603|%p48, %r599, %r601, %r600, %r602;
mov.b32 %f265, %r603;
add.ftz.f32 %f266, %f264, %f265;
mov.b32 %r604, %f266;
mov.u32 %r605, 8;
shfl.sync.bfly.b32 %r606|%p49, %r604, %r605, %r600, %r602;
mov.b32 %f267, %r606;
add.ftz.f32 %f268, %f266, %f267;
mov.b32 %r607, %f268;
mov.u32 %r608, 4;
shfl.sync.bfly.b32 %r609|%p50, %r607, %r608, %r600, %r602;
mov.b32 %f269, %r609;
add.ftz.f32 %f270, %f268, %f269;
mov.b32 %r610, %f270;
mov.u32 %r611, 2;
shfl.sync.bfly.b32 %r612|%p51, %r610, %r611, %r600, %r602;
mov.b32 %f271, %r612;
add.ftz.f32 %f272, %f270, %f271;
mov.b32 %r613, %f272;
mov.u32 %r614, 1;
shfl.sync.bfly.b32 %r615|%p52, %r613, %r614, %r600, %r602;
mov.b32 %f273, %r615;
add.ftz.f32 %f31, %f272, %f273;
@%p15 bra $L__BB178_20;

cvt.u64.u32 %rd77, %r24;
add.s64 %rd78, %rd77, %rd16;
shl.b64 %rd79, %rd78, 2;
add.s64 %rd80, %rd1, %rd79;
st.global.f32 [%rd80], %f31;

$L__BB178_20:
ld.shared.f32 %f274, [%r22+896];
add.ftz.f32 %f275, %f274, %f297;
ld.shared.f32 %f276, [%r22+1920];
add.ftz.f32 %f277, %f276, %f275;
ld.shared.f32 %f278, [%r22+2944];
add.ftz.f32 %f279, %f278, %f277;
mov.b32 %r616, %f279;
shfl.sync.bfly.b32 %r620|%p54, %r616, %r601, %r600, %r602;
mov.b32 %f280, %r620;
add.ftz.f32 %f281, %f279, %f280;
mov.b32 %r621, %f281;
shfl.sync.bfly.b32 %r623|%p55, %r621, %r605, %r600, %r602;
mov.b32 %f282, %r623;
add.ftz.f32 %f283, %f281, %f282;
mov.b32 %r624, %f283;
shfl.sync.bfly.b32 %r626|%p56, %r624, %r608, %r600, %r602;
mov.b32 %f284, %r626;
add.ftz.f32 %f285, %f283, %f284;
mov.b32 %r627, %f285;
shfl.sync.bfly.b32 %r629|%p57, %r627, %r611, %r600, %r602;
mov.b32 %f286, %r629;
add.ftz.f32 %f287, %f285, %f286;
mov.b32 %r630, %f287;
shfl.sync.bfly.b32 %r632|%p58, %r630, %r614, %r600, %r602;
mov.b32 %f288, %r632;
add.ftz.f32 %f32, %f287, %f288;
@%p23 bra $L__BB178_22;

add.s32 %r633, %r24, 1;
cvt.u64.u32 %rd81, %r633;
add.s64 %rd82, %rd81, %rd16;
shl.b64 %rd83, %rd82, 2;
add.s64 %rd84, %rd1, %rd83;
st.global.f32 [%rd84], %f32;

$L__BB178_22:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type13ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type13ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type13ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type13ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type13ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type13ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type13ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type13ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type13ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<72>;
.reg .b16 %rs<15>;
.reg .f32 %f<341>;
.reg .b32 %r<753>;
.reg .b64 %rd<95>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type13ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1280];

ld.param.u64 %rd20, [_Z13mul_mat_vec_qIL9ggml_type13ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd21, [_Z13mul_mat_vec_qIL9ggml_type13ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd22, [_Z13mul_mat_vec_qIL9ggml_type13ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type13ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type13ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type13ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type13ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type13ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type13ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type13ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type13ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type13ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type13ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type13ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type13ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type13ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type13ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
cvta.to.global.u64 %rd1, %rd22;
mov.u32 %r1, %tid.y;
shl.b32 %r45, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r45, %r2;
mov.u32 %r46, %ctaid.x;
shl.b32 %r4, %r46, 1;
mov.u32 %r5, %ctaid.y;
shr.s32 %r47, %r3, 31;
shr.u32 %r48, %r47, 28;
add.s32 %r49, %r3, %r48;
shr.s32 %r752, %r49, 4;
shr.u32 %r7, %r44, 8;
setp.ge.s32 %p3, %r752, %r7;
mov.f32 %f331, 0f00000000;
mov.f32 %f332, %f331;
mov.f32 %f333, %f331;
mov.f32 %f334, %f331;
mov.f32 %f335, %f331;
mov.f32 %f336, %f331;
mov.f32 %f337, %f331;
mov.f32 %f338, %f331;
mov.f32 %f339, %f331;
mov.f32 %f340, %f331;
@%p3 bra $L__BB179_3;

mov.u32 %r50, %ctaid.z;
mul.hi.u32 %r51, %r50, %r38;
add.s32 %r52, %r51, %r50;
shr.u32 %r53, %r52, %r39;
mul.lo.s32 %r54, %r53, %r41;
mul.hi.u32 %r55, %r5, %r32;
add.s32 %r56, %r55, %r5;
shr.u32 %r57, %r56, %r33;
mul.lo.s32 %r58, %r57, %r35;
mul.lo.s32 %r59, %r5, %r36;
cvt.u64.u32 %rd23, %r59;
cvta.to.global.u64 %rd94, %rd20;
and.b32 %r63, %r49, -16;
sub.s32 %r8, %r3, %r63;
shr.s32 %r64, %r8, 31;
shr.u32 %r65, %r64, 30;
add.s32 %r66, %r8, %r65;
shr.s32 %r67, %r66, 2;
shl.b32 %r9, %r67, 1;
shl.b32 %r68, %r67, 5;
and.b32 %r69, %r66, -4;
sub.s32 %r70, %r8, %r69;
shl.b32 %r71, %r70, 2;
add.s32 %r72, %r68, %r71;
cvt.s64.s32 %rd24, %r72;
cvt.s64.s32 %rd25, %r71;
cvt.s64.s32 %rd26, %r9;
add.s64 %rd27, %rd23, %rd26;
mul.lo.s32 %r73, %r50, %r42;
cvt.u64.u32 %rd28, %r73;
add.s64 %rd3, %rd27, %rd28;
cvt.s64.s32 %rd4, %r70;
shl.b32 %r74, %r30, 1;
shl.b32 %r747, %r752, 3;
add.s32 %r751, %r747, %r74;
mad.lo.s32 %r750, %r30, 3, %r747;
shl.b32 %r75, %r30, 2;
add.s32 %r749, %r747, %r75;
add.s32 %r748, %r30, %r747;
add.s32 %r76, %r752, %r54;
add.s32 %r77, %r76, %r58;
add.s32 %r78, %r4, 1;
mad.lo.s32 %r79, %r29, %r78, %r77;
mul.wide.s32 %rd5, %r79, 176;
mad.lo.s32 %r80, %r4, %r29, %r752;
add.s32 %r81, %r80, %r54;
add.s32 %r82, %r81, %r58;
mul.wide.s32 %rd6, %r82, 176;
add.s64 %rd29, %rd6, %rd24;
add.s64 %rd7, %rd29, 64;
cvt.u64.u32 %rd30, %r9;
and.b64 %rd31, %rd30, 2;
or.b64 %rd32, %rd31, %rd5;
or.b64 %rd8, %rd32, 8;
add.s64 %rd33, %rd6, %rd25;
add.s64 %rd9, %rd33, 32;
or.b64 %rd34, %rd31, %rd6;
or.b64 %rd10, %rd34, 12;
add.s64 %rd35, %rd5, %rd25;
add.s64 %rd11, %rd35, 16;
add.s64 %rd36, %rd5, %rd24;
add.s64 %rd12, %rd36, 48;
cvta.to.global.u64 %rd13, %rd21;
mov.f32 %f331, 0f00000000;
shl.b64 %rd44, %rd4, 2;

$L__BB179_2:
cvt.u64.u32 %rd37, %r747;
add.s64 %rd38, %rd3, %rd37;
add.s64 %rd39, %rd94, %rd7;
ld.global.nc.u32 %r443, [%rd39+-16];
ld.global.nc.u32 %r444, [%rd39];
add.s64 %rd40, %rd94, %rd9;
ld.global.nc.u32 %r445, [%rd40+-16];
shr.s32 %r446, %r445, %r9;
ld.global.nc.u32 %r447, [%rd40];
shr.s32 %r448, %r447, %r9;
add.s64 %rd41, %rd94, %rd10;
ld.global.nc.u16 %rs1, [%rd41+-8];
cvt.u32.u16 %r449, %rs1;
ld.global.nc.u16 %rs2, [%rd41+-4];
cvt.u32.u16 %r450, %rs2;
ld.global.nc.u16 %rs3, [%rd41];
cvt.u32.u16 %r451, %rs3;
and.b32 %r452, %r449, 16191;
and.b32 %r453, %r451, 3855;
and.b32 %r454, %r449, 49344;
shr.u32 %r455, %r454, 2;
or.b32 %r456, %r453, %r455;
setp.gt.s32 %p4, %r8, 7;
and.b32 %r457, %r450, 16191;
shr.u32 %r458, %r451, 4;
and.b32 %r459, %r458, 3855;
and.b32 %r460, %r450, 49344;
shr.u32 %r461, %r460, 2;
or.b32 %r462, %r459, %r461;
mul.lo.s64 %rd42, %rd38, 36;
add.s64 %rd43, %rd13, %rd42;
ld.global.nc.u32 %r83, [%rd43];

{.reg .f16 low,high;
mov.b32 {low,high},%r83;
cvt.f32.f16 %f61, low;}


add.s64 %rd45, %rd43, %rd44;
ld.global.nc.u32 %r91, [%rd45+4];
ld.global.nc.u32 %r87, [%rd45+20];
ld.global.nc.u32 %r84, [%rd43+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r84;
cvt.f32.f16 %f62, low;}


ld.global.nc.u32 %r107, [%rd45+40];
ld.global.nc.u32 %r103, [%rd45+56];
add.s64 %rd46, %rd94, %rd6;
and.b32 %r463, %r443, 252645135;
and.b32 %r464, %r444, 252645135;
shl.b32 %r465, %r446, 4;
and.b32 %r466, %r465, 269488144;
shl.b32 %r467, %r448, 4;
and.b32 %r468, %r467, 269488144;
or.b32 %r378, %r463, %r466;
or.b32 %r374, %r464, %r468;
mov.u32 %r436, 0;

dp4a.s32.s32 %r85, %r374, %r87, %r436;


dp4a.s32.s32 %r89, %r378, %r91, %r85;

mov.u32 %r438, 16843009;

dp4a.s32.s32 %r93, %r438, %r87, %r436;


dp4a.s32.s32 %r97, %r438, %r91, %r93;

selp.b32 %r469, %r456, %r452, %p4;
cvt.u16.u32 %rs4, %r469;
and.b32 %r470, %r469, 63;
mul.lo.s32 %r471, %r89, %r470;
cvt.rn.f32.s32 %f101, %r471;
fma.rn.ftz.f32 %f102, %f61, %f101, 0f00000000;
selp.b32 %r472, %r462, %r457, %p4;
cvt.u16.u32 %rs5, %r472;
and.b32 %r473, %r472, 63;
mul.lo.s32 %r474, %r97, %r473;
cvt.rn.f32.s32 %f103, %r474;
fma.rn.ftz.f32 %f104, %f61, %f103, 0f00000000;
shr.u32 %r475, %r443, 4;
and.b32 %r476, %r475, 252645135;
shr.u32 %r477, %r444, 4;
and.b32 %r478, %r477, 252645135;
shl.b32 %r479, %r446, 3;
and.b32 %r480, %r479, 269488144;
shl.b32 %r481, %r448, 3;
and.b32 %r482, %r481, 269488144;
or.b32 %r394, %r476, %r480;
or.b32 %r390, %r478, %r482;

dp4a.s32.s32 %r101, %r390, %r103, %r436;


dp4a.s32.s32 %r105, %r394, %r107, %r101;


dp4a.s32.s32 %r109, %r438, %r103, %r436;


dp4a.s32.s32 %r113, %r438, %r107, %r109;

shr.u16 %rs6, %rs4, 8;
cvt.u32.u16 %r483, %rs6;
mul.lo.s32 %r484, %r105, %r483;
cvt.rn.f32.s32 %f105, %r484;
fma.rn.ftz.f32 %f106, %f62, %f105, %f102;
shr.u16 %rs7, %rs5, 8;
cvt.u32.u16 %r485, %rs7;
mul.lo.s32 %r486, %r113, %r485;
cvt.rn.f32.s32 %f107, %r486;
fma.rn.ftz.f32 %f108, %f62, %f107, %f104;
ld.global.nc.u32 %r117, [%rd46];

{.reg .f16 low,high;
mov.b32 {low,high},%r117;
cvt.f32.f16 %f63, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r117;
cvt.f32.f16 %f64, high;}


mul.ftz.f32 %f109, %f106, %f63;
mul.ftz.f32 %f110, %f108, %f64;
sub.ftz.f32 %f111, %f109, %f110;
add.ftz.f32 %f340, %f111, %f340;
add.s64 %rd47, %rd94, %rd12;
add.s64 %rd48, %rd94, %rd11;
ld.global.nc.u32 %r487, [%rd47];
ld.global.nc.u32 %r488, [%rd47+16];
ld.global.nc.u32 %r489, [%rd48];
shr.s32 %r490, %r489, %r9;
ld.global.nc.u32 %r491, [%rd48+16];
shr.s32 %r492, %r491, %r9;
add.s64 %rd49, %rd94, %rd8;
ld.global.nc.u16 %rs8, [%rd49+-4];
cvt.u32.u16 %r493, %rs8;
ld.global.nc.u16 %rs9, [%rd49];
cvt.u32.u16 %r494, %rs9;
ld.global.nc.u16 %rs10, [%rd49+4];
cvt.u32.u16 %r495, %rs10;
and.b32 %r496, %r493, 16191;
and.b32 %r497, %r495, 3855;
and.b32 %r498, %r493, 49344;
shr.u32 %r499, %r498, 2;
or.b32 %r500, %r497, %r499;
and.b32 %r501, %r494, 16191;
shr.u32 %r502, %r495, 4;
and.b32 %r503, %r502, 3855;
and.b32 %r504, %r494, 49344;
shr.u32 %r505, %r504, 2;
or.b32 %r506, %r503, %r505;

{.reg .f16 low,high;
mov.b32 {low,high},%r83;
cvt.f32.f16 %f65, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r84;
cvt.f32.f16 %f66, low;}


add.s64 %rd50, %rd94, %rd5;
and.b32 %r507, %r487, 252645135;
and.b32 %r508, %r488, 252645135;
shl.b32 %r509, %r490, 4;
and.b32 %r510, %r509, 269488144;
shl.b32 %r511, %r492, 4;
and.b32 %r512, %r511, 269488144;
or.b32 %r414, %r507, %r510;
or.b32 %r410, %r508, %r512;

dp4a.s32.s32 %r121, %r410, %r87, %r436;


dp4a.s32.s32 %r125, %r414, %r91, %r121;


dp4a.s32.s32 %r129, %r438, %r87, %r436;


dp4a.s32.s32 %r133, %r438, %r91, %r129;

selp.b32 %r513, %r500, %r496, %p4;
cvt.u16.u32 %rs11, %r513;
and.b32 %r514, %r513, 63;
mul.lo.s32 %r515, %r125, %r514;
cvt.rn.f32.s32 %f112, %r515;
fma.rn.ftz.f32 %f113, %f65, %f112, 0f00000000;
selp.b32 %r516, %r506, %r501, %p4;
cvt.u16.u32 %rs12, %r516;
and.b32 %r517, %r516, 63;
mul.lo.s32 %r518, %r133, %r517;
cvt.rn.f32.s32 %f114, %r518;
fma.rn.ftz.f32 %f115, %f65, %f114, 0f00000000;
shr.u32 %r519, %r487, 4;
and.b32 %r520, %r519, 252645135;
shr.u32 %r521, %r488, 4;
and.b32 %r522, %r521, 252645135;
shl.b32 %r523, %r490, 3;
and.b32 %r524, %r523, 269488144;
shl.b32 %r525, %r492, 3;
and.b32 %r526, %r525, 269488144;
or.b32 %r430, %r520, %r524;
or.b32 %r426, %r522, %r526;

dp4a.s32.s32 %r137, %r426, %r103, %r436;


dp4a.s32.s32 %r141, %r430, %r107, %r137;


dp4a.s32.s32 %r145, %r438, %r103, %r436;


dp4a.s32.s32 %r149, %r438, %r107, %r145;

shr.u16 %rs13, %rs11, 8;
cvt.u32.u16 %r527, %rs13;
mul.lo.s32 %r528, %r141, %r527;
cvt.rn.f32.s32 %f116, %r528;
fma.rn.ftz.f32 %f117, %f66, %f116, %f113;
shr.u16 %rs14, %rs12, 8;
cvt.u32.u16 %r529, %rs14;
mul.lo.s32 %r530, %r149, %r529;
cvt.rn.f32.s32 %f118, %r530;
fma.rn.ftz.f32 %f119, %f66, %f118, %f115;
ld.global.nc.u32 %r153, [%rd50];

{.reg .f16 low,high;
mov.b32 {low,high},%r153;
cvt.f32.f16 %f67, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r153;
cvt.f32.f16 %f68, high;}


mul.ftz.f32 %f120, %f117, %f67;
mul.ftz.f32 %f121, %f119, %f68;
sub.ftz.f32 %f122, %f120, %f121;
add.ftz.f32 %f339, %f122, %f339;
cvt.u64.u32 %rd51, %r748;
add.s64 %rd52, %rd3, %rd51;
mul.lo.s64 %rd53, %rd52, 36;
add.s64 %rd54, %rd13, %rd53;
ld.global.nc.u32 %r155, [%rd54];

{.reg .f16 low,high;
mov.b32 {low,high},%r155;
cvt.f32.f16 %f69, low;}


add.s64 %rd55, %rd54, %rd44;
ld.global.nc.u32 %r163, [%rd55+4];
ld.global.nc.u32 %r159, [%rd55+20];
ld.global.nc.u32 %r156, [%rd54+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r156;
cvt.f32.f16 %f70, low;}


ld.global.nc.u32 %r179, [%rd55+40];
ld.global.nc.u32 %r175, [%rd55+56];

dp4a.s32.s32 %r157, %r374, %r159, %r436;


dp4a.s32.s32 %r161, %r378, %r163, %r157;


dp4a.s32.s32 %r165, %r438, %r159, %r436;


dp4a.s32.s32 %r169, %r438, %r163, %r165;

mul.lo.s32 %r531, %r161, %r470;
cvt.rn.f32.s32 %f123, %r531;
fma.rn.ftz.f32 %f124, %f69, %f123, 0f00000000;
mul.lo.s32 %r532, %r169, %r473;
cvt.rn.f32.s32 %f125, %r532;
fma.rn.ftz.f32 %f126, %f69, %f125, 0f00000000;

dp4a.s32.s32 %r173, %r390, %r175, %r436;


dp4a.s32.s32 %r177, %r394, %r179, %r173;


dp4a.s32.s32 %r181, %r438, %r175, %r436;


dp4a.s32.s32 %r185, %r438, %r179, %r181;

mul.lo.s32 %r533, %r177, %r483;
cvt.rn.f32.s32 %f127, %r533;
fma.rn.ftz.f32 %f128, %f70, %f127, %f124;
mul.lo.s32 %r534, %r185, %r485;
cvt.rn.f32.s32 %f129, %r534;
fma.rn.ftz.f32 %f130, %f70, %f129, %f126;

{.reg .f16 low,high;
mov.b32 {low,high},%r117;
cvt.f32.f16 %f71, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r117;
cvt.f32.f16 %f72, high;}


mul.ftz.f32 %f131, %f128, %f71;
mul.ftz.f32 %f132, %f130, %f72;
sub.ftz.f32 %f133, %f131, %f132;
add.ftz.f32 %f338, %f133, %f338;

{.reg .f16 low,high;
mov.b32 {low,high},%r155;
cvt.f32.f16 %f73, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r156;
cvt.f32.f16 %f74, low;}



dp4a.s32.s32 %r193, %r410, %r159, %r436;


dp4a.s32.s32 %r197, %r414, %r163, %r193;


dp4a.s32.s32 %r201, %r438, %r159, %r436;


dp4a.s32.s32 %r205, %r438, %r163, %r201;

mul.lo.s32 %r535, %r197, %r514;
cvt.rn.f32.s32 %f134, %r535;
fma.rn.ftz.f32 %f135, %f73, %f134, 0f00000000;
mul.lo.s32 %r536, %r205, %r517;
cvt.rn.f32.s32 %f136, %r536;
fma.rn.ftz.f32 %f137, %f73, %f136, 0f00000000;

dp4a.s32.s32 %r209, %r426, %r175, %r436;


dp4a.s32.s32 %r213, %r430, %r179, %r209;


dp4a.s32.s32 %r217, %r438, %r175, %r436;


dp4a.s32.s32 %r221, %r438, %r179, %r217;

mul.lo.s32 %r537, %r213, %r527;
cvt.rn.f32.s32 %f138, %r537;
fma.rn.ftz.f32 %f139, %f74, %f138, %f135;
mul.lo.s32 %r538, %r221, %r529;
cvt.rn.f32.s32 %f140, %r538;
fma.rn.ftz.f32 %f141, %f74, %f140, %f137;

{.reg .f16 low,high;
mov.b32 {low,high},%r153;
cvt.f32.f16 %f75, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r153;
cvt.f32.f16 %f76, high;}


mul.ftz.f32 %f142, %f139, %f75;
mul.ftz.f32 %f143, %f141, %f76;
sub.ftz.f32 %f144, %f142, %f143;
add.ftz.f32 %f337, %f144, %f337;
cvt.u64.u32 %rd56, %r751;
add.s64 %rd57, %rd3, %rd56;
mul.lo.s64 %rd58, %rd57, 36;
add.s64 %rd59, %rd13, %rd58;
ld.global.nc.u32 %r227, [%rd59];

{.reg .f16 low,high;
mov.b32 {low,high},%r227;
cvt.f32.f16 %f77, low;}


add.s64 %rd60, %rd59, %rd44;
ld.global.nc.u32 %r235, [%rd60+4];
ld.global.nc.u32 %r231, [%rd60+20];
ld.global.nc.u32 %r228, [%rd59+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r228;
cvt.f32.f16 %f78, low;}


ld.global.nc.u32 %r251, [%rd60+40];
ld.global.nc.u32 %r247, [%rd60+56];

dp4a.s32.s32 %r229, %r374, %r231, %r436;


dp4a.s32.s32 %r233, %r378, %r235, %r229;


dp4a.s32.s32 %r237, %r438, %r231, %r436;


dp4a.s32.s32 %r241, %r438, %r235, %r237;

mul.lo.s32 %r539, %r233, %r470;
cvt.rn.f32.s32 %f145, %r539;
fma.rn.ftz.f32 %f146, %f77, %f145, 0f00000000;
mul.lo.s32 %r540, %r241, %r473;
cvt.rn.f32.s32 %f147, %r540;
fma.rn.ftz.f32 %f148, %f77, %f147, 0f00000000;

dp4a.s32.s32 %r245, %r390, %r247, %r436;


dp4a.s32.s32 %r249, %r394, %r251, %r245;


dp4a.s32.s32 %r253, %r438, %r247, %r436;


dp4a.s32.s32 %r257, %r438, %r251, %r253;

mul.lo.s32 %r541, %r249, %r483;
cvt.rn.f32.s32 %f149, %r541;
fma.rn.ftz.f32 %f150, %f78, %f149, %f146;
mul.lo.s32 %r542, %r257, %r485;
cvt.rn.f32.s32 %f151, %r542;
fma.rn.ftz.f32 %f152, %f78, %f151, %f148;

{.reg .f16 low,high;
mov.b32 {low,high},%r117;
cvt.f32.f16 %f79, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r117;
cvt.f32.f16 %f80, high;}


mul.ftz.f32 %f153, %f150, %f79;
mul.ftz.f32 %f154, %f152, %f80;
sub.ftz.f32 %f155, %f153, %f154;
add.ftz.f32 %f336, %f155, %f336;

{.reg .f16 low,high;
mov.b32 {low,high},%r227;
cvt.f32.f16 %f81, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r228;
cvt.f32.f16 %f82, low;}



dp4a.s32.s32 %r265, %r410, %r231, %r436;


dp4a.s32.s32 %r269, %r414, %r235, %r265;


dp4a.s32.s32 %r273, %r438, %r231, %r436;


dp4a.s32.s32 %r277, %r438, %r235, %r273;

mul.lo.s32 %r543, %r269, %r514;
cvt.rn.f32.s32 %f156, %r543;
fma.rn.ftz.f32 %f157, %f81, %f156, 0f00000000;
mul.lo.s32 %r544, %r277, %r517;
cvt.rn.f32.s32 %f158, %r544;
fma.rn.ftz.f32 %f159, %f81, %f158, 0f00000000;

dp4a.s32.s32 %r281, %r426, %r247, %r436;


dp4a.s32.s32 %r285, %r430, %r251, %r281;


dp4a.s32.s32 %r289, %r438, %r247, %r436;


dp4a.s32.s32 %r293, %r438, %r251, %r289;

mul.lo.s32 %r545, %r285, %r527;
cvt.rn.f32.s32 %f160, %r545;
fma.rn.ftz.f32 %f161, %f82, %f160, %f157;
mul.lo.s32 %r546, %r293, %r529;
cvt.rn.f32.s32 %f162, %r546;
fma.rn.ftz.f32 %f163, %f82, %f162, %f159;

{.reg .f16 low,high;
mov.b32 {low,high},%r153;
cvt.f32.f16 %f83, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r153;
cvt.f32.f16 %f84, high;}


mul.ftz.f32 %f164, %f161, %f83;
mul.ftz.f32 %f165, %f163, %f84;
sub.ftz.f32 %f166, %f164, %f165;
add.ftz.f32 %f335, %f166, %f335;
cvt.u64.u32 %rd61, %r750;
add.s64 %rd62, %rd3, %rd61;
mul.lo.s64 %rd63, %rd62, 36;
add.s64 %rd64, %rd13, %rd63;
ld.global.nc.u32 %r299, [%rd64];

{.reg .f16 low,high;
mov.b32 {low,high},%r299;
cvt.f32.f16 %f85, low;}


add.s64 %rd65, %rd64, %rd44;
ld.global.nc.u32 %r307, [%rd65+4];
ld.global.nc.u32 %r303, [%rd65+20];
ld.global.nc.u32 %r300, [%rd64+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r300;
cvt.f32.f16 %f86, low;}


ld.global.nc.u32 %r323, [%rd65+40];
ld.global.nc.u32 %r319, [%rd65+56];

dp4a.s32.s32 %r301, %r374, %r303, %r436;


dp4a.s32.s32 %r305, %r378, %r307, %r301;


dp4a.s32.s32 %r309, %r438, %r303, %r436;


dp4a.s32.s32 %r313, %r438, %r307, %r309;

mul.lo.s32 %r547, %r305, %r470;
cvt.rn.f32.s32 %f167, %r547;
fma.rn.ftz.f32 %f168, %f85, %f167, 0f00000000;
mul.lo.s32 %r548, %r313, %r473;
cvt.rn.f32.s32 %f169, %r548;
fma.rn.ftz.f32 %f170, %f85, %f169, 0f00000000;

dp4a.s32.s32 %r317, %r390, %r319, %r436;


dp4a.s32.s32 %r321, %r394, %r323, %r317;


dp4a.s32.s32 %r325, %r438, %r319, %r436;


dp4a.s32.s32 %r329, %r438, %r323, %r325;

mul.lo.s32 %r549, %r321, %r483;
cvt.rn.f32.s32 %f171, %r549;
fma.rn.ftz.f32 %f172, %f86, %f171, %f168;
mul.lo.s32 %r550, %r329, %r485;
cvt.rn.f32.s32 %f173, %r550;
fma.rn.ftz.f32 %f174, %f86, %f173, %f170;

{.reg .f16 low,high;
mov.b32 {low,high},%r117;
cvt.f32.f16 %f87, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r117;
cvt.f32.f16 %f88, high;}


mul.ftz.f32 %f175, %f172, %f87;
mul.ftz.f32 %f176, %f174, %f88;
sub.ftz.f32 %f177, %f175, %f176;
add.ftz.f32 %f334, %f177, %f334;

{.reg .f16 low,high;
mov.b32 {low,high},%r299;
cvt.f32.f16 %f89, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r300;
cvt.f32.f16 %f90, low;}



dp4a.s32.s32 %r337, %r410, %r303, %r436;


dp4a.s32.s32 %r341, %r414, %r307, %r337;


dp4a.s32.s32 %r345, %r438, %r303, %r436;


dp4a.s32.s32 %r349, %r438, %r307, %r345;

mul.lo.s32 %r551, %r341, %r514;
cvt.rn.f32.s32 %f178, %r551;
fma.rn.ftz.f32 %f179, %f89, %f178, 0f00000000;
mul.lo.s32 %r552, %r349, %r517;
cvt.rn.f32.s32 %f180, %r552;
fma.rn.ftz.f32 %f181, %f89, %f180, 0f00000000;

dp4a.s32.s32 %r353, %r426, %r319, %r436;


dp4a.s32.s32 %r357, %r430, %r323, %r353;


dp4a.s32.s32 %r361, %r438, %r319, %r436;


dp4a.s32.s32 %r365, %r438, %r323, %r361;

mul.lo.s32 %r553, %r357, %r527;
cvt.rn.f32.s32 %f182, %r553;
fma.rn.ftz.f32 %f183, %f90, %f182, %f179;
mul.lo.s32 %r554, %r365, %r529;
cvt.rn.f32.s32 %f184, %r554;
fma.rn.ftz.f32 %f185, %f90, %f184, %f181;

{.reg .f16 low,high;
mov.b32 {low,high},%r153;
cvt.f32.f16 %f91, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r153;
cvt.f32.f16 %f92, high;}


mul.ftz.f32 %f186, %f183, %f91;
mul.ftz.f32 %f187, %f185, %f92;
sub.ftz.f32 %f188, %f186, %f187;
add.ftz.f32 %f333, %f188, %f333;
cvt.u64.u32 %rd66, %r749;
add.s64 %rd67, %rd3, %rd66;
mul.lo.s64 %rd68, %rd67, 36;
add.s64 %rd69, %rd13, %rd68;
ld.global.nc.u32 %r371, [%rd69];

{.reg .f16 low,high;
mov.b32 {low,high},%r371;
cvt.f32.f16 %f93, low;}


add.s64 %rd70, %rd69, %rd44;
ld.global.nc.u32 %r379, [%rd70+4];
ld.global.nc.u32 %r375, [%rd70+20];
ld.global.nc.u32 %r372, [%rd69+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r372;
cvt.f32.f16 %f94, low;}


ld.global.nc.u32 %r395, [%rd70+40];
ld.global.nc.u32 %r391, [%rd70+56];

dp4a.s32.s32 %r373, %r374, %r375, %r436;


dp4a.s32.s32 %r377, %r378, %r379, %r373;


dp4a.s32.s32 %r381, %r438, %r375, %r436;


dp4a.s32.s32 %r385, %r438, %r379, %r381;

mul.lo.s32 %r555, %r377, %r470;
cvt.rn.f32.s32 %f189, %r555;
fma.rn.ftz.f32 %f190, %f93, %f189, 0f00000000;
mul.lo.s32 %r556, %r385, %r473;
cvt.rn.f32.s32 %f191, %r556;
fma.rn.ftz.f32 %f192, %f93, %f191, 0f00000000;

dp4a.s32.s32 %r389, %r390, %r391, %r436;


dp4a.s32.s32 %r393, %r394, %r395, %r389;


dp4a.s32.s32 %r397, %r438, %r391, %r436;


dp4a.s32.s32 %r401, %r438, %r395, %r397;

mul.lo.s32 %r557, %r393, %r483;
cvt.rn.f32.s32 %f193, %r557;
fma.rn.ftz.f32 %f194, %f94, %f193, %f190;
mul.lo.s32 %r558, %r401, %r485;
cvt.rn.f32.s32 %f195, %r558;
fma.rn.ftz.f32 %f196, %f94, %f195, %f192;

{.reg .f16 low,high;
mov.b32 {low,high},%r117;
cvt.f32.f16 %f95, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r117;
cvt.f32.f16 %f96, high;}


mul.ftz.f32 %f197, %f194, %f95;
mul.ftz.f32 %f198, %f196, %f96;
sub.ftz.f32 %f199, %f197, %f198;
add.ftz.f32 %f332, %f199, %f332;

{.reg .f16 low,high;
mov.b32 {low,high},%r371;
cvt.f32.f16 %f97, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r372;
cvt.f32.f16 %f98, low;}



dp4a.s32.s32 %r409, %r410, %r375, %r436;


dp4a.s32.s32 %r413, %r414, %r379, %r409;


dp4a.s32.s32 %r417, %r438, %r375, %r436;


dp4a.s32.s32 %r421, %r438, %r379, %r417;

mul.lo.s32 %r559, %r413, %r514;
cvt.rn.f32.s32 %f200, %r559;
fma.rn.ftz.f32 %f201, %f97, %f200, 0f00000000;
mul.lo.s32 %r560, %r421, %r517;
cvt.rn.f32.s32 %f202, %r560;
fma.rn.ftz.f32 %f203, %f97, %f202, 0f00000000;

dp4a.s32.s32 %r425, %r426, %r391, %r436;


dp4a.s32.s32 %r429, %r430, %r395, %r425;


dp4a.s32.s32 %r433, %r438, %r391, %r436;


dp4a.s32.s32 %r437, %r438, %r395, %r433;

mul.lo.s32 %r561, %r429, %r527;
cvt.rn.f32.s32 %f204, %r561;
fma.rn.ftz.f32 %f205, %f98, %f204, %f201;
mul.lo.s32 %r562, %r437, %r529;
cvt.rn.f32.s32 %f206, %r562;
fma.rn.ftz.f32 %f207, %f98, %f206, %f203;

{.reg .f16 low,high;
mov.b32 {low,high},%r153;
cvt.f32.f16 %f99, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r153;
cvt.f32.f16 %f100, high;}


mul.ftz.f32 %f208, %f205, %f99;
mul.ftz.f32 %f209, %f207, %f100;
sub.ftz.f32 %f210, %f208, %f209;
add.ftz.f32 %f331, %f210, %f331;
add.s32 %r751, %r751, 32;
add.s32 %r750, %r750, 32;
add.s32 %r749, %r749, 32;
add.s32 %r748, %r748, 32;
add.s32 %r747, %r747, 32;
add.s64 %rd94, %rd94, 704;
add.s32 %r752, %r752, 4;
setp.lt.s32 %p5, %r752, %r7;
@%p5 bra $L__BB179_2;

$L__BB179_3:
setp.eq.s32 %p6, %r1, 0;
@%p6 bra $L__BB179_5;

mov.u32 %r563, _ZZ13mul_mat_vec_qIL9ggml_type13ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r564, %r1, 1280, %r563;
shl.b32 %r565, %r2, 2;
add.s32 %r566, %r564, %r565;
st.shared.f32 [%r566+-1280], %f340;
st.shared.f32 [%r566+-1152], %f339;
st.shared.f32 [%r566+-1024], %f338;
st.shared.f32 [%r566+-896], %f337;
st.shared.f32 [%r566+-768], %f336;
st.shared.f32 [%r566+-640], %f335;
st.shared.f32 [%r566+-512], %f334;
st.shared.f32 [%r566+-384], %f333;
st.shared.f32 [%r566+-256], %f332;
st.shared.f32 [%r566+-128], %f331;

$L__BB179_5:
setp.ne.s32 %p7, %r1, 0;
bar.sync 0;
@%p7 bra $L__BB179_26;

mov.u32 %r567, %ctaid.z;
mad.lo.s32 %r568, %r5, %r37, %r4;
mad.lo.s32 %r569, %r567, %r43, %r568;
cvt.u64.u32 %rd16, %r569;
shl.b32 %r570, %r2, 2;
mov.u32 %r571, _ZZ13mul_mat_vec_qIL9ggml_type13ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r27, %r571, %r570;
mov.u32 %r572, 2;
ld.shared.f32 %f211, [%r27];
add.ftz.f32 %f212, %f211, %f340;
mov.b32 %r573, %f212;
mov.u32 %r574, 31;
mov.u32 %r575, 16;
mov.u32 %r576, -1;
shfl.sync.bfly.b32 %r577|%p8, %r573, %r575, %r574, %r576;
mov.b32 %f213, %r577;
add.ftz.f32 %f214, %f212, %f213;
mov.b32 %r578, %f214;
mov.u32 %r579, 8;
shfl.sync.bfly.b32 %r580|%p9, %r578, %r579, %r574, %r576;
mov.b32 %f215, %r580;
add.ftz.f32 %f216, %f214, %f215;
mov.b32 %r581, %f216;
mov.u32 %r582, 4;
shfl.sync.bfly.b32 %r583|%p10, %r581, %r582, %r574, %r576;
mov.b32 %f217, %r583;
add.ftz.f32 %f218, %f216, %f217;
mov.b32 %r584, %f218;
shfl.sync.bfly.b32 %r585|%p11, %r584, %r572, %r574, %r576;
mov.b32 %f219, %r585;
add.ftz.f32 %f220, %f218, %f219;
mov.b32 %r586, %f220;
mov.u32 %r587, 1;
shfl.sync.bfly.b32 %r588|%p12, %r586, %r587, %r574, %r576;
mov.b32 %f221, %r588;
add.ftz.f32 %f31, %f220, %f221;
setp.lt.u32 %p13, %r4, %r31;
setp.eq.s32 %p14, %r2, 0;
and.pred %p1, %p14, %p13;
mul.wide.u32 %rd71, %r569, 4;
add.s64 %rd17, %rd1, %rd71;
not.pred %p15, %p1;
@%p15 bra $L__BB179_8;

st.global.f32 [%rd17], %f31;

$L__BB179_8:
ld.shared.f32 %f222, [%r27+128];
add.ftz.f32 %f223, %f222, %f339;
mov.b32 %r589, %f223;
shfl.sync.bfly.b32 %r593|%p16, %r589, %r575, %r574, %r576;
mov.b32 %f224, %r593;
add.ftz.f32 %f225, %f223, %f224;
mov.b32 %r594, %f225;
shfl.sync.bfly.b32 %r596|%p17, %r594, %r579, %r574, %r576;
mov.b32 %f226, %r596;
add.ftz.f32 %f227, %f225, %f226;
mov.b32 %r597, %f227;
shfl.sync.bfly.b32 %r599|%p18, %r597, %r582, %r574, %r576;
mov.b32 %f228, %r599;
add.ftz.f32 %f229, %f227, %f228;
mov.b32 %r600, %f229;
shfl.sync.bfly.b32 %r602|%p19, %r600, %r572, %r574, %r576;
mov.b32 %f230, %r602;
add.ftz.f32 %f231, %f229, %f230;
mov.b32 %r603, %f231;
shfl.sync.bfly.b32 %r605|%p20, %r603, %r587, %r574, %r576;
mov.b32 %f232, %r605;
add.ftz.f32 %f32, %f231, %f232;
add.s32 %r606, %r4, 1;
setp.lt.u32 %p21, %r606, %r31;
setp.eq.s32 %p22, %r2, 1;
and.pred %p2, %p22, %p21;
not.pred %p23, %p2;
@%p23 bra $L__BB179_10;

st.global.f32 [%rd17+4], %f32;

$L__BB179_10:
ld.shared.f32 %f233, [%r27+256];
add.ftz.f32 %f234, %f233, %f338;
mov.b32 %r607, %f234;
mov.u32 %r608, 31;
mov.u32 %r609, 16;
mov.u32 %r610, -1;
shfl.sync.bfly.b32 %r611|%p24, %r607, %r609, %r608, %r610;
mov.b32 %f235, %r611;
add.ftz.f32 %f236, %f234, %f235;
mov.b32 %r612, %f236;
mov.u32 %r613, 8;
shfl.sync.bfly.b32 %r614|%p25, %r612, %r613, %r608, %r610;
mov.b32 %f237, %r614;
add.ftz.f32 %f238, %f236, %f237;
mov.b32 %r615, %f238;
mov.u32 %r616, 4;
shfl.sync.bfly.b32 %r617|%p26, %r615, %r616, %r608, %r610;
mov.b32 %f239, %r617;
add.ftz.f32 %f240, %f238, %f239;
mov.b32 %r618, %f240;
mov.u32 %r619, 2;
shfl.sync.bfly.b32 %r620|%p27, %r618, %r619, %r608, %r610;
mov.b32 %f241, %r620;
add.ftz.f32 %f242, %f240, %f241;
mov.b32 %r621, %f242;
mov.u32 %r622, 1;
shfl.sync.bfly.b32 %r623|%p28, %r621, %r622, %r608, %r610;
mov.b32 %f243, %r623;
add.ftz.f32 %f33, %f242, %f243;
@%p15 bra $L__BB179_12;

cvt.u64.u32 %rd72, %r31;
add.s64 %rd73, %rd72, %rd16;
shl.b64 %rd74, %rd73, 2;
add.s64 %rd75, %rd1, %rd74;
st.global.f32 [%rd75], %f33;

$L__BB179_12:
ld.shared.f32 %f244, [%r27+384];
add.ftz.f32 %f245, %f244, %f337;
mov.b32 %r624, %f245;
shfl.sync.bfly.b32 %r628|%p30, %r624, %r609, %r608, %r610;
mov.b32 %f246, %r628;
add.ftz.f32 %f247, %f245, %f246;
mov.b32 %r629, %f247;
shfl.sync.bfly.b32 %r631|%p31, %r629, %r613, %r608, %r610;
mov.b32 %f248, %r631;
add.ftz.f32 %f249, %f247, %f248;
mov.b32 %r632, %f249;
shfl.sync.bfly.b32 %r634|%p32, %r632, %r616, %r608, %r610;
mov.b32 %f250, %r634;
add.ftz.f32 %f251, %f249, %f250;
mov.b32 %r635, %f251;
shfl.sync.bfly.b32 %r637|%p33, %r635, %r619, %r608, %r610;
mov.b32 %f252, %r637;
add.ftz.f32 %f253, %f251, %f252;
mov.b32 %r638, %f253;
shfl.sync.bfly.b32 %r640|%p34, %r638, %r622, %r608, %r610;
mov.b32 %f254, %r640;
add.ftz.f32 %f34, %f253, %f254;
@%p23 bra $L__BB179_14;

add.s32 %r641, %r31, 1;
cvt.u64.u32 %rd76, %r641;
add.s64 %rd77, %rd76, %rd16;
shl.b64 %rd78, %rd77, 2;
add.s64 %rd79, %rd1, %rd78;
st.global.f32 [%rd79], %f34;

$L__BB179_14:
shl.b32 %r642, %r31, 1;
mov.u32 %r643, 1;
ld.shared.f32 %f255, [%r27+512];
add.ftz.f32 %f256, %f255, %f336;
mov.b32 %r644, %f256;
mov.u32 %r645, 31;
mov.u32 %r646, 16;
mov.u32 %r647, -1;
shfl.sync.bfly.b32 %r648|%p36, %r644, %r646, %r645, %r647;
mov.b32 %f257, %r648;
add.ftz.f32 %f258, %f256, %f257;
mov.b32 %r649, %f258;
mov.u32 %r650, 8;
shfl.sync.bfly.b32 %r651|%p37, %r649, %r650, %r645, %r647;
mov.b32 %f259, %r651;
add.ftz.f32 %f260, %f258, %f259;
mov.b32 %r652, %f260;
mov.u32 %r653, 4;
shfl.sync.bfly.b32 %r654|%p38, %r652, %r653, %r645, %r647;
mov.b32 %f261, %r654;
add.ftz.f32 %f262, %f260, %f261;
mov.b32 %r655, %f262;
mov.u32 %r656, 2;
shfl.sync.bfly.b32 %r657|%p39, %r655, %r656, %r645, %r647;
mov.b32 %f263, %r657;
add.ftz.f32 %f264, %f262, %f263;
mov.b32 %r658, %f264;
shfl.sync.bfly.b32 %r659|%p40, %r658, %r643, %r645, %r647;
mov.b32 %f265, %r659;
add.ftz.f32 %f35, %f264, %f265;
cvt.u64.u32 %rd80, %r642;
add.s64 %rd81, %rd80, %rd16;
shl.b64 %rd82, %rd81, 2;
add.s64 %rd18, %rd1, %rd82;
@%p15 bra $L__BB179_16;

st.global.f32 [%rd18], %f35;

$L__BB179_16:
ld.shared.f32 %f266, [%r27+640];
add.ftz.f32 %f267, %f266, %f335;
mov.b32 %r660, %f267;
shfl.sync.bfly.b32 %r664|%p42, %r660, %r646, %r645, %r647;
mov.b32 %f268, %r664;
add.ftz.f32 %f269, %f267, %f268;
mov.b32 %r665, %f269;
shfl.sync.bfly.b32 %r667|%p43, %r665, %r650, %r645, %r647;
mov.b32 %f270, %r667;
add.ftz.f32 %f271, %f269, %f270;
mov.b32 %r668, %f271;
shfl.sync.bfly.b32 %r670|%p44, %r668, %r653, %r645, %r647;
mov.b32 %f272, %r670;
add.ftz.f32 %f273, %f271, %f272;
mov.b32 %r671, %f273;
shfl.sync.bfly.b32 %r673|%p45, %r671, %r656, %r645, %r647;
mov.b32 %f274, %r673;
add.ftz.f32 %f275, %f273, %f274;
mov.b32 %r674, %f275;
shfl.sync.bfly.b32 %r676|%p46, %r674, %r643, %r645, %r647;
mov.b32 %f276, %r676;
add.ftz.f32 %f36, %f275, %f276;
@%p23 bra $L__BB179_18;

st.global.f32 [%rd18+4], %f36;

$L__BB179_18:
mul.lo.s32 %r28, %r31, 3;
ld.shared.f32 %f277, [%r27+768];
add.ftz.f32 %f278, %f277, %f334;
mov.b32 %r677, %f278;
mov.u32 %r678, 31;
mov.u32 %r679, 16;
mov.u32 %r680, -1;
shfl.sync.bfly.b32 %r681|%p48, %r677, %r679, %r678, %r680;
mov.b32 %f279, %r681;
add.ftz.f32 %f280, %f278, %f279;
mov.b32 %r682, %f280;
mov.u32 %r683, 8;
shfl.sync.bfly.b32 %r684|%p49, %r682, %r683, %r678, %r680;
mov.b32 %f281, %r684;
add.ftz.f32 %f282, %f280, %f281;
mov.b32 %r685, %f282;
mov.u32 %r686, 4;
shfl.sync.bfly.b32 %r687|%p50, %r685, %r686, %r678, %r680;
mov.b32 %f283, %r687;
add.ftz.f32 %f284, %f282, %f283;
mov.b32 %r688, %f284;
mov.u32 %r689, 2;
shfl.sync.bfly.b32 %r690|%p51, %r688, %r689, %r678, %r680;
mov.b32 %f285, %r690;
add.ftz.f32 %f286, %f284, %f285;
mov.b32 %r691, %f286;
mov.u32 %r692, 1;
shfl.sync.bfly.b32 %r693|%p52, %r691, %r692, %r678, %r680;
mov.b32 %f287, %r693;
add.ftz.f32 %f37, %f286, %f287;
@%p15 bra $L__BB179_20;

cvt.u64.u32 %rd83, %r28;
add.s64 %rd84, %rd83, %rd16;
shl.b64 %rd85, %rd84, 2;
add.s64 %rd86, %rd1, %rd85;
st.global.f32 [%rd86], %f37;

$L__BB179_20:
ld.shared.f32 %f288, [%r27+896];
add.ftz.f32 %f289, %f288, %f333;
mov.b32 %r694, %f289;
shfl.sync.bfly.b32 %r698|%p54, %r694, %r679, %r678, %r680;
mov.b32 %f290, %r698;
add.ftz.f32 %f291, %f289, %f290;
mov.b32 %r699, %f291;
shfl.sync.bfly.b32 %r701|%p55, %r699, %r683, %r678, %r680;
mov.b32 %f292, %r701;
add.ftz.f32 %f293, %f291, %f292;
mov.b32 %r702, %f293;
shfl.sync.bfly.b32 %r704|%p56, %r702, %r686, %r678, %r680;
mov.b32 %f294, %r704;
add.ftz.f32 %f295, %f293, %f294;
mov.b32 %r705, %f295;
shfl.sync.bfly.b32 %r707|%p57, %r705, %r689, %r678, %r680;
mov.b32 %f296, %r707;
add.ftz.f32 %f297, %f295, %f296;
mov.b32 %r708, %f297;
shfl.sync.bfly.b32 %r710|%p58, %r708, %r692, %r678, %r680;
mov.b32 %f298, %r710;
add.ftz.f32 %f38, %f297, %f298;
@%p23 bra $L__BB179_22;

add.s32 %r711, %r28, 1;
cvt.u64.u32 %rd87, %r711;
add.s64 %rd88, %rd87, %rd16;
shl.b64 %rd89, %rd88, 2;
add.s64 %rd90, %rd1, %rd89;
st.global.f32 [%rd90], %f38;

$L__BB179_22:
shl.b32 %r712, %r31, 2;
mov.u32 %r713, 2;
ld.shared.f32 %f299, [%r27+1024];
add.ftz.f32 %f300, %f299, %f332;
mov.b32 %r714, %f300;
mov.u32 %r715, 31;
mov.u32 %r716, 16;
mov.u32 %r717, -1;
shfl.sync.bfly.b32 %r718|%p60, %r714, %r716, %r715, %r717;
mov.b32 %f301, %r718;
add.ftz.f32 %f302, %f300, %f301;
mov.b32 %r719, %f302;
mov.u32 %r720, 8;
shfl.sync.bfly.b32 %r721|%p61, %r719, %r720, %r715, %r717;
mov.b32 %f303, %r721;
add.ftz.f32 %f304, %f302, %f303;
mov.b32 %r722, %f304;
mov.u32 %r723, 4;
shfl.sync.bfly.b32 %r724|%p62, %r722, %r723, %r715, %r717;
mov.b32 %f305, %r724;
add.ftz.f32 %f306, %f304, %f305;
mov.b32 %r725, %f306;
shfl.sync.bfly.b32 %r726|%p63, %r725, %r713, %r715, %r717;
mov.b32 %f307, %r726;
add.ftz.f32 %f308, %f306, %f307;
mov.b32 %r727, %f308;
mov.u32 %r728, 1;
shfl.sync.bfly.b32 %r729|%p64, %r727, %r728, %r715, %r717;
mov.b32 %f309, %r729;
add.ftz.f32 %f39, %f308, %f309;
cvt.u64.u32 %rd91, %r712;
add.s64 %rd92, %rd91, %rd16;
shl.b64 %rd93, %rd92, 2;
add.s64 %rd19, %rd1, %rd93;
@%p15 bra $L__BB179_24;

st.global.f32 [%rd19], %f39;

$L__BB179_24:
ld.shared.f32 %f310, [%r27+1152];
add.ftz.f32 %f311, %f310, %f331;
mov.b32 %r730, %f311;
shfl.sync.bfly.b32 %r734|%p66, %r730, %r716, %r715, %r717;
mov.b32 %f312, %r734;
add.ftz.f32 %f313, %f311, %f312;
mov.b32 %r735, %f313;
shfl.sync.bfly.b32 %r737|%p67, %r735, %r720, %r715, %r717;
mov.b32 %f314, %r737;
add.ftz.f32 %f315, %f313, %f314;
mov.b32 %r738, %f315;
shfl.sync.bfly.b32 %r740|%p68, %r738, %r723, %r715, %r717;
mov.b32 %f316, %r740;
add.ftz.f32 %f317, %f315, %f316;
mov.b32 %r741, %f317;
shfl.sync.bfly.b32 %r743|%p69, %r741, %r713, %r715, %r717;
mov.b32 %f318, %r743;
add.ftz.f32 %f319, %f317, %f318;
mov.b32 %r744, %f319;
shfl.sync.bfly.b32 %r746|%p70, %r744, %r728, %r715, %r717;
mov.b32 %f320, %r746;
add.ftz.f32 %f40, %f319, %f320;
@%p23 bra $L__BB179_26;

st.global.f32 [%rd19+4], %f40;

$L__BB179_26:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type13ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type13ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type13ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type13ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type13ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type13ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type13ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type13ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type13ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<84>;
.reg .b16 %rs<15>;
.reg .f32 %f<409>;
.reg .b32 %r<873>;
.reg .b64 %rd<108>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type13ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

ld.param.u64 %rd20, [_Z13mul_mat_vec_qIL9ggml_type13ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd21, [_Z13mul_mat_vec_qIL9ggml_type13ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd22, [_Z13mul_mat_vec_qIL9ggml_type13ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r48, [_Z13mul_mat_vec_qIL9ggml_type13ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type13ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type13ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type13ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type13ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type13ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type13ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r45, [_Z13mul_mat_vec_qIL9ggml_type13ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r46, [_Z13mul_mat_vec_qIL9ggml_type13ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r47, [_Z13mul_mat_vec_qIL9ggml_type13ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type13ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type13ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type13ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type13ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
cvta.to.global.u64 %rd1, %rd22;
mov.u32 %r1, %tid.y;
shl.b32 %r49, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r49, %r2;
mov.u32 %r50, %ctaid.x;
shl.b32 %r4, %r50, 1;
mov.u32 %r5, %ctaid.y;
shr.s32 %r51, %r3, 31;
shr.u32 %r52, %r51, 28;
add.s32 %r53, %r3, %r52;
shr.s32 %r872, %r53, 4;
shr.u32 %r7, %r48, 8;
setp.ge.s32 %p3, %r872, %r7;
mov.f32 %f397, 0f00000000;
mov.f32 %f398, %f397;
mov.f32 %f399, %f397;
mov.f32 %f400, %f397;
mov.f32 %f401, %f397;
mov.f32 %f402, %f397;
mov.f32 %f403, %f397;
mov.f32 %f404, %f397;
mov.f32 %f405, %f397;
mov.f32 %f406, %f397;
mov.f32 %f407, %f397;
mov.f32 %f408, %f397;
@%p3 bra $L__BB180_3;

mov.u32 %r54, %ctaid.z;
mul.hi.u32 %r55, %r54, %r42;
add.s32 %r56, %r55, %r54;
shr.u32 %r57, %r56, %r43;
mul.lo.s32 %r58, %r57, %r45;
mul.hi.u32 %r59, %r5, %r36;
add.s32 %r60, %r59, %r5;
shr.u32 %r61, %r60, %r37;
mul.lo.s32 %r62, %r61, %r39;
mul.lo.s32 %r63, %r5, %r40;
cvt.u64.u32 %rd23, %r63;
cvta.to.global.u64 %rd107, %rd20;
and.b32 %r67, %r53, -16;
sub.s32 %r8, %r3, %r67;
shr.s32 %r68, %r8, 31;
shr.u32 %r69, %r68, 30;
add.s32 %r70, %r8, %r69;
shr.s32 %r71, %r70, 2;
shl.b32 %r9, %r71, 1;
shl.b32 %r72, %r71, 5;
and.b32 %r73, %r70, -4;
sub.s32 %r74, %r8, %r73;
shl.b32 %r75, %r74, 2;
add.s32 %r76, %r72, %r75;
cvt.s64.s32 %rd24, %r76;
cvt.s64.s32 %rd25, %r75;
cvt.s64.s32 %rd26, %r9;
add.s64 %rd27, %rd23, %rd26;
mul.lo.s32 %r77, %r54, %r46;
cvt.u64.u32 %rd28, %r77;
add.s64 %rd3, %rd27, %rd28;
cvt.s64.s32 %rd4, %r74;
shl.b32 %r78, %r34, 1;
shl.b32 %r866, %r872, 3;
add.s32 %r871, %r866, %r78;
mad.lo.s32 %r870, %r34, 3, %r866;
shl.b32 %r79, %r34, 2;
add.s32 %r869, %r866, %r79;
mad.lo.s32 %r868, %r34, 5, %r866;
add.s32 %r867, %r34, %r866;
add.s32 %r80, %r872, %r58;
add.s32 %r81, %r80, %r62;
add.s32 %r82, %r4, 1;
mad.lo.s32 %r83, %r33, %r82, %r81;
mul.wide.s32 %rd5, %r83, 176;
mad.lo.s32 %r84, %r4, %r33, %r872;
add.s32 %r85, %r84, %r58;
add.s32 %r86, %r85, %r62;
mul.wide.s32 %rd6, %r86, 176;
add.s64 %rd29, %rd6, %rd24;
add.s64 %rd7, %rd29, 64;
cvt.u64.u32 %rd30, %r9;
and.b64 %rd31, %rd30, 2;
or.b64 %rd32, %rd31, %rd5;
or.b64 %rd8, %rd32, 8;
add.s64 %rd33, %rd6, %rd25;
add.s64 %rd9, %rd33, 32;
or.b64 %rd34, %rd31, %rd6;
or.b64 %rd10, %rd34, 12;
add.s64 %rd35, %rd5, %rd25;
add.s64 %rd11, %rd35, 16;
add.s64 %rd36, %rd5, %rd24;
add.s64 %rd12, %rd36, 48;
cvta.to.global.u64 %rd13, %rd21;
mov.f32 %f397, 0f00000000;
shl.b64 %rd44, %rd4, 2;

$L__BB180_2:
cvt.u64.u32 %rd37, %r866;
add.s64 %rd38, %rd3, %rd37;
add.s64 %rd39, %rd107, %rd7;
ld.global.nc.u32 %r519, [%rd39+-16];
ld.global.nc.u32 %r520, [%rd39];
add.s64 %rd40, %rd107, %rd9;
ld.global.nc.u32 %r521, [%rd40+-16];
shr.s32 %r522, %r521, %r9;
ld.global.nc.u32 %r523, [%rd40];
shr.s32 %r524, %r523, %r9;
add.s64 %rd41, %rd107, %rd10;
ld.global.nc.u16 %rs1, [%rd41+-8];
cvt.u32.u16 %r525, %rs1;
ld.global.nc.u16 %rs2, [%rd41+-4];
cvt.u32.u16 %r526, %rs2;
ld.global.nc.u16 %rs3, [%rd41];
cvt.u32.u16 %r527, %rs3;
and.b32 %r528, %r525, 16191;
and.b32 %r529, %r527, 3855;
and.b32 %r530, %r525, 49344;
shr.u32 %r531, %r530, 2;
or.b32 %r532, %r529, %r531;
setp.gt.s32 %p4, %r8, 7;
and.b32 %r533, %r526, 16191;
shr.u32 %r534, %r527, 4;
and.b32 %r535, %r534, 3855;
and.b32 %r536, %r526, 49344;
shr.u32 %r537, %r536, 2;
or.b32 %r538, %r535, %r537;
mul.lo.s64 %rd42, %rd38, 36;
add.s64 %rd43, %rd13, %rd42;
ld.global.nc.u32 %r87, [%rd43];

{.reg .f16 low,high;
mov.b32 {low,high},%r87;
cvt.f32.f16 %f73, low;}


add.s64 %rd45, %rd43, %rd44;
ld.global.nc.u32 %r95, [%rd45+4];
ld.global.nc.u32 %r91, [%rd45+20];
ld.global.nc.u32 %r88, [%rd43+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r88;
cvt.f32.f16 %f74, low;}


ld.global.nc.u32 %r111, [%rd45+40];
ld.global.nc.u32 %r107, [%rd45+56];
add.s64 %rd46, %rd107, %rd6;
and.b32 %r539, %r519, 252645135;
and.b32 %r540, %r520, 252645135;
shl.b32 %r541, %r522, 4;
and.b32 %r542, %r541, 269488144;
shl.b32 %r543, %r524, 4;
and.b32 %r544, %r543, 269488144;
or.b32 %r454, %r539, %r542;
or.b32 %r450, %r540, %r544;
mov.u32 %r512, 0;

dp4a.s32.s32 %r89, %r450, %r91, %r512;


dp4a.s32.s32 %r93, %r454, %r95, %r89;

mov.u32 %r514, 16843009;

dp4a.s32.s32 %r97, %r514, %r91, %r512;


dp4a.s32.s32 %r101, %r514, %r95, %r97;

selp.b32 %r545, %r532, %r528, %p4;
cvt.u16.u32 %rs4, %r545;
and.b32 %r546, %r545, 63;
mul.lo.s32 %r547, %r93, %r546;
cvt.rn.f32.s32 %f121, %r547;
fma.rn.ftz.f32 %f122, %f73, %f121, 0f00000000;
selp.b32 %r548, %r538, %r533, %p4;
cvt.u16.u32 %rs5, %r548;
and.b32 %r549, %r548, 63;
mul.lo.s32 %r550, %r101, %r549;
cvt.rn.f32.s32 %f123, %r550;
fma.rn.ftz.f32 %f124, %f73, %f123, 0f00000000;
shr.u32 %r551, %r519, 4;
and.b32 %r552, %r551, 252645135;
shr.u32 %r553, %r520, 4;
and.b32 %r554, %r553, 252645135;
shl.b32 %r555, %r522, 3;
and.b32 %r556, %r555, 269488144;
shl.b32 %r557, %r524, 3;
and.b32 %r558, %r557, 269488144;
or.b32 %r470, %r552, %r556;
or.b32 %r466, %r554, %r558;

dp4a.s32.s32 %r105, %r466, %r107, %r512;


dp4a.s32.s32 %r109, %r470, %r111, %r105;


dp4a.s32.s32 %r113, %r514, %r107, %r512;


dp4a.s32.s32 %r117, %r514, %r111, %r113;

shr.u16 %rs6, %rs4, 8;
cvt.u32.u16 %r559, %rs6;
mul.lo.s32 %r560, %r109, %r559;
cvt.rn.f32.s32 %f125, %r560;
fma.rn.ftz.f32 %f126, %f74, %f125, %f122;
shr.u16 %rs7, %rs5, 8;
cvt.u32.u16 %r561, %rs7;
mul.lo.s32 %r562, %r117, %r561;
cvt.rn.f32.s32 %f127, %r562;
fma.rn.ftz.f32 %f128, %f74, %f127, %f124;
ld.global.nc.u32 %r121, [%rd46];

{.reg .f16 low,high;
mov.b32 {low,high},%r121;
cvt.f32.f16 %f75, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r121;
cvt.f32.f16 %f76, high;}


mul.ftz.f32 %f129, %f126, %f75;
mul.ftz.f32 %f130, %f128, %f76;
sub.ftz.f32 %f131, %f129, %f130;
add.ftz.f32 %f408, %f131, %f408;
add.s64 %rd47, %rd107, %rd12;
add.s64 %rd48, %rd107, %rd11;
ld.global.nc.u32 %r563, [%rd47];
ld.global.nc.u32 %r564, [%rd47+16];
ld.global.nc.u32 %r565, [%rd48];
shr.s32 %r566, %r565, %r9;
ld.global.nc.u32 %r567, [%rd48+16];
shr.s32 %r568, %r567, %r9;
add.s64 %rd49, %rd107, %rd8;
ld.global.nc.u16 %rs8, [%rd49+-4];
cvt.u32.u16 %r569, %rs8;
ld.global.nc.u16 %rs9, [%rd49];
cvt.u32.u16 %r570, %rs9;
ld.global.nc.u16 %rs10, [%rd49+4];
cvt.u32.u16 %r571, %rs10;
and.b32 %r572, %r569, 16191;
and.b32 %r573, %r571, 3855;
and.b32 %r574, %r569, 49344;
shr.u32 %r575, %r574, 2;
or.b32 %r576, %r573, %r575;
and.b32 %r577, %r570, 16191;
shr.u32 %r578, %r571, 4;
and.b32 %r579, %r578, 3855;
and.b32 %r580, %r570, 49344;
shr.u32 %r581, %r580, 2;
or.b32 %r582, %r579, %r581;

{.reg .f16 low,high;
mov.b32 {low,high},%r87;
cvt.f32.f16 %f77, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r88;
cvt.f32.f16 %f78, low;}


add.s64 %rd50, %rd107, %rd5;
and.b32 %r583, %r563, 252645135;
and.b32 %r584, %r564, 252645135;
shl.b32 %r585, %r566, 4;
and.b32 %r586, %r585, 269488144;
shl.b32 %r587, %r568, 4;
and.b32 %r588, %r587, 269488144;
or.b32 %r490, %r583, %r586;
or.b32 %r486, %r584, %r588;

dp4a.s32.s32 %r125, %r486, %r91, %r512;


dp4a.s32.s32 %r129, %r490, %r95, %r125;


dp4a.s32.s32 %r133, %r514, %r91, %r512;


dp4a.s32.s32 %r137, %r514, %r95, %r133;

selp.b32 %r589, %r576, %r572, %p4;
cvt.u16.u32 %rs11, %r589;
and.b32 %r590, %r589, 63;
mul.lo.s32 %r591, %r129, %r590;
cvt.rn.f32.s32 %f132, %r591;
fma.rn.ftz.f32 %f133, %f77, %f132, 0f00000000;
selp.b32 %r592, %r582, %r577, %p4;
cvt.u16.u32 %rs12, %r592;
and.b32 %r593, %r592, 63;
mul.lo.s32 %r594, %r137, %r593;
cvt.rn.f32.s32 %f134, %r594;
fma.rn.ftz.f32 %f135, %f77, %f134, 0f00000000;
shr.u32 %r595, %r563, 4;
and.b32 %r596, %r595, 252645135;
shr.u32 %r597, %r564, 4;
and.b32 %r598, %r597, 252645135;
shl.b32 %r599, %r566, 3;
and.b32 %r600, %r599, 269488144;
shl.b32 %r601, %r568, 3;
and.b32 %r602, %r601, 269488144;
or.b32 %r506, %r596, %r600;
or.b32 %r502, %r598, %r602;

dp4a.s32.s32 %r141, %r502, %r107, %r512;


dp4a.s32.s32 %r145, %r506, %r111, %r141;


dp4a.s32.s32 %r149, %r514, %r107, %r512;


dp4a.s32.s32 %r153, %r514, %r111, %r149;

shr.u16 %rs13, %rs11, 8;
cvt.u32.u16 %r603, %rs13;
mul.lo.s32 %r604, %r145, %r603;
cvt.rn.f32.s32 %f136, %r604;
fma.rn.ftz.f32 %f137, %f78, %f136, %f133;
shr.u16 %rs14, %rs12, 8;
cvt.u32.u16 %r605, %rs14;
mul.lo.s32 %r606, %r153, %r605;
cvt.rn.f32.s32 %f138, %r606;
fma.rn.ftz.f32 %f139, %f78, %f138, %f135;
ld.global.nc.u32 %r157, [%rd50];

{.reg .f16 low,high;
mov.b32 {low,high},%r157;
cvt.f32.f16 %f79, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r157;
cvt.f32.f16 %f80, high;}


mul.ftz.f32 %f140, %f137, %f79;
mul.ftz.f32 %f141, %f139, %f80;
sub.ftz.f32 %f142, %f140, %f141;
add.ftz.f32 %f407, %f142, %f407;
cvt.u64.u32 %rd51, %r867;
add.s64 %rd52, %rd3, %rd51;
mul.lo.s64 %rd53, %rd52, 36;
add.s64 %rd54, %rd13, %rd53;
ld.global.nc.u32 %r159, [%rd54];

{.reg .f16 low,high;
mov.b32 {low,high},%r159;
cvt.f32.f16 %f81, low;}


add.s64 %rd55, %rd54, %rd44;
ld.global.nc.u32 %r167, [%rd55+4];
ld.global.nc.u32 %r163, [%rd55+20];
ld.global.nc.u32 %r160, [%rd54+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r160;
cvt.f32.f16 %f82, low;}


ld.global.nc.u32 %r183, [%rd55+40];
ld.global.nc.u32 %r179, [%rd55+56];

dp4a.s32.s32 %r161, %r450, %r163, %r512;


dp4a.s32.s32 %r165, %r454, %r167, %r161;


dp4a.s32.s32 %r169, %r514, %r163, %r512;


dp4a.s32.s32 %r173, %r514, %r167, %r169;

mul.lo.s32 %r607, %r165, %r546;
cvt.rn.f32.s32 %f143, %r607;
fma.rn.ftz.f32 %f144, %f81, %f143, 0f00000000;
mul.lo.s32 %r608, %r173, %r549;
cvt.rn.f32.s32 %f145, %r608;
fma.rn.ftz.f32 %f146, %f81, %f145, 0f00000000;

dp4a.s32.s32 %r177, %r466, %r179, %r512;


dp4a.s32.s32 %r181, %r470, %r183, %r177;


dp4a.s32.s32 %r185, %r514, %r179, %r512;


dp4a.s32.s32 %r189, %r514, %r183, %r185;

mul.lo.s32 %r609, %r181, %r559;
cvt.rn.f32.s32 %f147, %r609;
fma.rn.ftz.f32 %f148, %f82, %f147, %f144;
mul.lo.s32 %r610, %r189, %r561;
cvt.rn.f32.s32 %f149, %r610;
fma.rn.ftz.f32 %f150, %f82, %f149, %f146;

{.reg .f16 low,high;
mov.b32 {low,high},%r121;
cvt.f32.f16 %f83, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r121;
cvt.f32.f16 %f84, high;}


mul.ftz.f32 %f151, %f148, %f83;
mul.ftz.f32 %f152, %f150, %f84;
sub.ftz.f32 %f153, %f151, %f152;
add.ftz.f32 %f406, %f153, %f406;

{.reg .f16 low,high;
mov.b32 {low,high},%r159;
cvt.f32.f16 %f85, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r160;
cvt.f32.f16 %f86, low;}



dp4a.s32.s32 %r197, %r486, %r163, %r512;


dp4a.s32.s32 %r201, %r490, %r167, %r197;


dp4a.s32.s32 %r205, %r514, %r163, %r512;


dp4a.s32.s32 %r209, %r514, %r167, %r205;

mul.lo.s32 %r611, %r201, %r590;
cvt.rn.f32.s32 %f154, %r611;
fma.rn.ftz.f32 %f155, %f85, %f154, 0f00000000;
mul.lo.s32 %r612, %r209, %r593;
cvt.rn.f32.s32 %f156, %r612;
fma.rn.ftz.f32 %f157, %f85, %f156, 0f00000000;

dp4a.s32.s32 %r213, %r502, %r179, %r512;


dp4a.s32.s32 %r217, %r506, %r183, %r213;


dp4a.s32.s32 %r221, %r514, %r179, %r512;


dp4a.s32.s32 %r225, %r514, %r183, %r221;

mul.lo.s32 %r613, %r217, %r603;
cvt.rn.f32.s32 %f158, %r613;
fma.rn.ftz.f32 %f159, %f86, %f158, %f155;
mul.lo.s32 %r614, %r225, %r605;
cvt.rn.f32.s32 %f160, %r614;
fma.rn.ftz.f32 %f161, %f86, %f160, %f157;

{.reg .f16 low,high;
mov.b32 {low,high},%r157;
cvt.f32.f16 %f87, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r157;
cvt.f32.f16 %f88, high;}


mul.ftz.f32 %f162, %f159, %f87;
mul.ftz.f32 %f163, %f161, %f88;
sub.ftz.f32 %f164, %f162, %f163;
add.ftz.f32 %f405, %f164, %f405;
cvt.u64.u32 %rd56, %r871;
add.s64 %rd57, %rd3, %rd56;
mul.lo.s64 %rd58, %rd57, 36;
add.s64 %rd59, %rd13, %rd58;
ld.global.nc.u32 %r231, [%rd59];

{.reg .f16 low,high;
mov.b32 {low,high},%r231;
cvt.f32.f16 %f89, low;}


add.s64 %rd60, %rd59, %rd44;
ld.global.nc.u32 %r239, [%rd60+4];
ld.global.nc.u32 %r235, [%rd60+20];
ld.global.nc.u32 %r232, [%rd59+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r232;
cvt.f32.f16 %f90, low;}


ld.global.nc.u32 %r255, [%rd60+40];
ld.global.nc.u32 %r251, [%rd60+56];

dp4a.s32.s32 %r233, %r450, %r235, %r512;


dp4a.s32.s32 %r237, %r454, %r239, %r233;


dp4a.s32.s32 %r241, %r514, %r235, %r512;


dp4a.s32.s32 %r245, %r514, %r239, %r241;

mul.lo.s32 %r615, %r237, %r546;
cvt.rn.f32.s32 %f165, %r615;
fma.rn.ftz.f32 %f166, %f89, %f165, 0f00000000;
mul.lo.s32 %r616, %r245, %r549;
cvt.rn.f32.s32 %f167, %r616;
fma.rn.ftz.f32 %f168, %f89, %f167, 0f00000000;

dp4a.s32.s32 %r249, %r466, %r251, %r512;


dp4a.s32.s32 %r253, %r470, %r255, %r249;


dp4a.s32.s32 %r257, %r514, %r251, %r512;


dp4a.s32.s32 %r261, %r514, %r255, %r257;

mul.lo.s32 %r617, %r253, %r559;
cvt.rn.f32.s32 %f169, %r617;
fma.rn.ftz.f32 %f170, %f90, %f169, %f166;
mul.lo.s32 %r618, %r261, %r561;
cvt.rn.f32.s32 %f171, %r618;
fma.rn.ftz.f32 %f172, %f90, %f171, %f168;

{.reg .f16 low,high;
mov.b32 {low,high},%r121;
cvt.f32.f16 %f91, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r121;
cvt.f32.f16 %f92, high;}


mul.ftz.f32 %f173, %f170, %f91;
mul.ftz.f32 %f174, %f172, %f92;
sub.ftz.f32 %f175, %f173, %f174;
add.ftz.f32 %f404, %f175, %f404;

{.reg .f16 low,high;
mov.b32 {low,high},%r231;
cvt.f32.f16 %f93, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r232;
cvt.f32.f16 %f94, low;}



dp4a.s32.s32 %r269, %r486, %r235, %r512;


dp4a.s32.s32 %r273, %r490, %r239, %r269;


dp4a.s32.s32 %r277, %r514, %r235, %r512;


dp4a.s32.s32 %r281, %r514, %r239, %r277;

mul.lo.s32 %r619, %r273, %r590;
cvt.rn.f32.s32 %f176, %r619;
fma.rn.ftz.f32 %f177, %f93, %f176, 0f00000000;
mul.lo.s32 %r620, %r281, %r593;
cvt.rn.f32.s32 %f178, %r620;
fma.rn.ftz.f32 %f179, %f93, %f178, 0f00000000;

dp4a.s32.s32 %r285, %r502, %r251, %r512;


dp4a.s32.s32 %r289, %r506, %r255, %r285;


dp4a.s32.s32 %r293, %r514, %r251, %r512;


dp4a.s32.s32 %r297, %r514, %r255, %r293;

mul.lo.s32 %r621, %r289, %r603;
cvt.rn.f32.s32 %f180, %r621;
fma.rn.ftz.f32 %f181, %f94, %f180, %f177;
mul.lo.s32 %r622, %r297, %r605;
cvt.rn.f32.s32 %f182, %r622;
fma.rn.ftz.f32 %f183, %f94, %f182, %f179;

{.reg .f16 low,high;
mov.b32 {low,high},%r157;
cvt.f32.f16 %f95, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r157;
cvt.f32.f16 %f96, high;}


mul.ftz.f32 %f184, %f181, %f95;
mul.ftz.f32 %f185, %f183, %f96;
sub.ftz.f32 %f186, %f184, %f185;
add.ftz.f32 %f403, %f186, %f403;
cvt.u64.u32 %rd61, %r870;
add.s64 %rd62, %rd3, %rd61;
mul.lo.s64 %rd63, %rd62, 36;
add.s64 %rd64, %rd13, %rd63;
ld.global.nc.u32 %r303, [%rd64];

{.reg .f16 low,high;
mov.b32 {low,high},%r303;
cvt.f32.f16 %f97, low;}


add.s64 %rd65, %rd64, %rd44;
ld.global.nc.u32 %r311, [%rd65+4];
ld.global.nc.u32 %r307, [%rd65+20];
ld.global.nc.u32 %r304, [%rd64+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r304;
cvt.f32.f16 %f98, low;}


ld.global.nc.u32 %r327, [%rd65+40];
ld.global.nc.u32 %r323, [%rd65+56];

dp4a.s32.s32 %r305, %r450, %r307, %r512;


dp4a.s32.s32 %r309, %r454, %r311, %r305;


dp4a.s32.s32 %r313, %r514, %r307, %r512;


dp4a.s32.s32 %r317, %r514, %r311, %r313;

mul.lo.s32 %r623, %r309, %r546;
cvt.rn.f32.s32 %f187, %r623;
fma.rn.ftz.f32 %f188, %f97, %f187, 0f00000000;
mul.lo.s32 %r624, %r317, %r549;
cvt.rn.f32.s32 %f189, %r624;
fma.rn.ftz.f32 %f190, %f97, %f189, 0f00000000;

dp4a.s32.s32 %r321, %r466, %r323, %r512;


dp4a.s32.s32 %r325, %r470, %r327, %r321;


dp4a.s32.s32 %r329, %r514, %r323, %r512;


dp4a.s32.s32 %r333, %r514, %r327, %r329;

mul.lo.s32 %r625, %r325, %r559;
cvt.rn.f32.s32 %f191, %r625;
fma.rn.ftz.f32 %f192, %f98, %f191, %f188;
mul.lo.s32 %r626, %r333, %r561;
cvt.rn.f32.s32 %f193, %r626;
fma.rn.ftz.f32 %f194, %f98, %f193, %f190;

{.reg .f16 low,high;
mov.b32 {low,high},%r121;
cvt.f32.f16 %f99, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r121;
cvt.f32.f16 %f100, high;}


mul.ftz.f32 %f195, %f192, %f99;
mul.ftz.f32 %f196, %f194, %f100;
sub.ftz.f32 %f197, %f195, %f196;
add.ftz.f32 %f402, %f197, %f402;

{.reg .f16 low,high;
mov.b32 {low,high},%r303;
cvt.f32.f16 %f101, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r304;
cvt.f32.f16 %f102, low;}



dp4a.s32.s32 %r341, %r486, %r307, %r512;


dp4a.s32.s32 %r345, %r490, %r311, %r341;


dp4a.s32.s32 %r349, %r514, %r307, %r512;


dp4a.s32.s32 %r353, %r514, %r311, %r349;

mul.lo.s32 %r627, %r345, %r590;
cvt.rn.f32.s32 %f198, %r627;
fma.rn.ftz.f32 %f199, %f101, %f198, 0f00000000;
mul.lo.s32 %r628, %r353, %r593;
cvt.rn.f32.s32 %f200, %r628;
fma.rn.ftz.f32 %f201, %f101, %f200, 0f00000000;

dp4a.s32.s32 %r357, %r502, %r323, %r512;


dp4a.s32.s32 %r361, %r506, %r327, %r357;


dp4a.s32.s32 %r365, %r514, %r323, %r512;


dp4a.s32.s32 %r369, %r514, %r327, %r365;

mul.lo.s32 %r629, %r361, %r603;
cvt.rn.f32.s32 %f202, %r629;
fma.rn.ftz.f32 %f203, %f102, %f202, %f199;
mul.lo.s32 %r630, %r369, %r605;
cvt.rn.f32.s32 %f204, %r630;
fma.rn.ftz.f32 %f205, %f102, %f204, %f201;

{.reg .f16 low,high;
mov.b32 {low,high},%r157;
cvt.f32.f16 %f103, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r157;
cvt.f32.f16 %f104, high;}


mul.ftz.f32 %f206, %f203, %f103;
mul.ftz.f32 %f207, %f205, %f104;
sub.ftz.f32 %f208, %f206, %f207;
add.ftz.f32 %f401, %f208, %f401;
cvt.u64.u32 %rd66, %r869;
add.s64 %rd67, %rd3, %rd66;
mul.lo.s64 %rd68, %rd67, 36;
add.s64 %rd69, %rd13, %rd68;
ld.global.nc.u32 %r375, [%rd69];

{.reg .f16 low,high;
mov.b32 {low,high},%r375;
cvt.f32.f16 %f105, low;}


add.s64 %rd70, %rd69, %rd44;
ld.global.nc.u32 %r383, [%rd70+4];
ld.global.nc.u32 %r379, [%rd70+20];
ld.global.nc.u32 %r376, [%rd69+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r376;
cvt.f32.f16 %f106, low;}


ld.global.nc.u32 %r399, [%rd70+40];
ld.global.nc.u32 %r395, [%rd70+56];

dp4a.s32.s32 %r377, %r450, %r379, %r512;


dp4a.s32.s32 %r381, %r454, %r383, %r377;


dp4a.s32.s32 %r385, %r514, %r379, %r512;


dp4a.s32.s32 %r389, %r514, %r383, %r385;

mul.lo.s32 %r631, %r381, %r546;
cvt.rn.f32.s32 %f209, %r631;
fma.rn.ftz.f32 %f210, %f105, %f209, 0f00000000;
mul.lo.s32 %r632, %r389, %r549;
cvt.rn.f32.s32 %f211, %r632;
fma.rn.ftz.f32 %f212, %f105, %f211, 0f00000000;

dp4a.s32.s32 %r393, %r466, %r395, %r512;


dp4a.s32.s32 %r397, %r470, %r399, %r393;


dp4a.s32.s32 %r401, %r514, %r395, %r512;


dp4a.s32.s32 %r405, %r514, %r399, %r401;

mul.lo.s32 %r633, %r397, %r559;
cvt.rn.f32.s32 %f213, %r633;
fma.rn.ftz.f32 %f214, %f106, %f213, %f210;
mul.lo.s32 %r634, %r405, %r561;
cvt.rn.f32.s32 %f215, %r634;
fma.rn.ftz.f32 %f216, %f106, %f215, %f212;

{.reg .f16 low,high;
mov.b32 {low,high},%r121;
cvt.f32.f16 %f107, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r121;
cvt.f32.f16 %f108, high;}


mul.ftz.f32 %f217, %f214, %f107;
mul.ftz.f32 %f218, %f216, %f108;
sub.ftz.f32 %f219, %f217, %f218;
add.ftz.f32 %f400, %f219, %f400;

{.reg .f16 low,high;
mov.b32 {low,high},%r375;
cvt.f32.f16 %f109, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r376;
cvt.f32.f16 %f110, low;}



dp4a.s32.s32 %r413, %r486, %r379, %r512;


dp4a.s32.s32 %r417, %r490, %r383, %r413;


dp4a.s32.s32 %r421, %r514, %r379, %r512;


dp4a.s32.s32 %r425, %r514, %r383, %r421;

mul.lo.s32 %r635, %r417, %r590;
cvt.rn.f32.s32 %f220, %r635;
fma.rn.ftz.f32 %f221, %f109, %f220, 0f00000000;
mul.lo.s32 %r636, %r425, %r593;
cvt.rn.f32.s32 %f222, %r636;
fma.rn.ftz.f32 %f223, %f109, %f222, 0f00000000;

dp4a.s32.s32 %r429, %r502, %r395, %r512;


dp4a.s32.s32 %r433, %r506, %r399, %r429;


dp4a.s32.s32 %r437, %r514, %r395, %r512;


dp4a.s32.s32 %r441, %r514, %r399, %r437;

mul.lo.s32 %r637, %r433, %r603;
cvt.rn.f32.s32 %f224, %r637;
fma.rn.ftz.f32 %f225, %f110, %f224, %f221;
mul.lo.s32 %r638, %r441, %r605;
cvt.rn.f32.s32 %f226, %r638;
fma.rn.ftz.f32 %f227, %f110, %f226, %f223;

{.reg .f16 low,high;
mov.b32 {low,high},%r157;
cvt.f32.f16 %f111, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r157;
cvt.f32.f16 %f112, high;}


mul.ftz.f32 %f228, %f225, %f111;
mul.ftz.f32 %f229, %f227, %f112;
sub.ftz.f32 %f230, %f228, %f229;
add.ftz.f32 %f399, %f230, %f399;
cvt.u64.u32 %rd71, %r868;
add.s64 %rd72, %rd3, %rd71;
mul.lo.s64 %rd73, %rd72, 36;
add.s64 %rd74, %rd13, %rd73;
ld.global.nc.u32 %r447, [%rd74];

{.reg .f16 low,high;
mov.b32 {low,high},%r447;
cvt.f32.f16 %f113, low;}


add.s64 %rd75, %rd74, %rd44;
ld.global.nc.u32 %r455, [%rd75+4];
ld.global.nc.u32 %r451, [%rd75+20];
ld.global.nc.u32 %r448, [%rd74+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r448;
cvt.f32.f16 %f114, low;}


ld.global.nc.u32 %r471, [%rd75+40];
ld.global.nc.u32 %r467, [%rd75+56];

dp4a.s32.s32 %r449, %r450, %r451, %r512;


dp4a.s32.s32 %r453, %r454, %r455, %r449;


dp4a.s32.s32 %r457, %r514, %r451, %r512;


dp4a.s32.s32 %r461, %r514, %r455, %r457;

mul.lo.s32 %r639, %r453, %r546;
cvt.rn.f32.s32 %f231, %r639;
fma.rn.ftz.f32 %f232, %f113, %f231, 0f00000000;
mul.lo.s32 %r640, %r461, %r549;
cvt.rn.f32.s32 %f233, %r640;
fma.rn.ftz.f32 %f234, %f113, %f233, 0f00000000;

dp4a.s32.s32 %r465, %r466, %r467, %r512;


dp4a.s32.s32 %r469, %r470, %r471, %r465;


dp4a.s32.s32 %r473, %r514, %r467, %r512;


dp4a.s32.s32 %r477, %r514, %r471, %r473;

mul.lo.s32 %r641, %r469, %r559;
cvt.rn.f32.s32 %f235, %r641;
fma.rn.ftz.f32 %f236, %f114, %f235, %f232;
mul.lo.s32 %r642, %r477, %r561;
cvt.rn.f32.s32 %f237, %r642;
fma.rn.ftz.f32 %f238, %f114, %f237, %f234;

{.reg .f16 low,high;
mov.b32 {low,high},%r121;
cvt.f32.f16 %f115, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r121;
cvt.f32.f16 %f116, high;}


mul.ftz.f32 %f239, %f236, %f115;
mul.ftz.f32 %f240, %f238, %f116;
sub.ftz.f32 %f241, %f239, %f240;
add.ftz.f32 %f398, %f241, %f398;

{.reg .f16 low,high;
mov.b32 {low,high},%r447;
cvt.f32.f16 %f117, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r448;
cvt.f32.f16 %f118, low;}



dp4a.s32.s32 %r485, %r486, %r451, %r512;


dp4a.s32.s32 %r489, %r490, %r455, %r485;


dp4a.s32.s32 %r493, %r514, %r451, %r512;


dp4a.s32.s32 %r497, %r514, %r455, %r493;

mul.lo.s32 %r643, %r489, %r590;
cvt.rn.f32.s32 %f242, %r643;
fma.rn.ftz.f32 %f243, %f117, %f242, 0f00000000;
mul.lo.s32 %r644, %r497, %r593;
cvt.rn.f32.s32 %f244, %r644;
fma.rn.ftz.f32 %f245, %f117, %f244, 0f00000000;

dp4a.s32.s32 %r501, %r502, %r467, %r512;


dp4a.s32.s32 %r505, %r506, %r471, %r501;


dp4a.s32.s32 %r509, %r514, %r467, %r512;


dp4a.s32.s32 %r513, %r514, %r471, %r509;

mul.lo.s32 %r645, %r505, %r603;
cvt.rn.f32.s32 %f246, %r645;
fma.rn.ftz.f32 %f247, %f118, %f246, %f243;
mul.lo.s32 %r646, %r513, %r605;
cvt.rn.f32.s32 %f248, %r646;
fma.rn.ftz.f32 %f249, %f118, %f248, %f245;

{.reg .f16 low,high;
mov.b32 {low,high},%r157;
cvt.f32.f16 %f119, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r157;
cvt.f32.f16 %f120, high;}


mul.ftz.f32 %f250, %f247, %f119;
mul.ftz.f32 %f251, %f249, %f120;
sub.ftz.f32 %f252, %f250, %f251;
add.ftz.f32 %f397, %f252, %f397;
add.s32 %r871, %r871, 32;
add.s32 %r870, %r870, 32;
add.s32 %r869, %r869, 32;
add.s32 %r868, %r868, 32;
add.s32 %r867, %r867, 32;
add.s32 %r866, %r866, 32;
add.s64 %rd107, %rd107, 704;
add.s32 %r872, %r872, 4;
setp.lt.s32 %p5, %r872, %r7;
@%p5 bra $L__BB180_2;

$L__BB180_3:
setp.eq.s32 %p6, %r1, 0;
@%p6 bra $L__BB180_5;

mov.u32 %r647, _ZZ13mul_mat_vec_qIL9ggml_type13ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r648, %r1, 1536, %r647;
shl.b32 %r649, %r2, 2;
add.s32 %r650, %r648, %r649;
st.shared.f32 [%r650+-1536], %f408;
st.shared.f32 [%r650+-1408], %f407;
st.shared.f32 [%r650+-1280], %f406;
st.shared.f32 [%r650+-1152], %f405;
st.shared.f32 [%r650+-1024], %f404;
st.shared.f32 [%r650+-896], %f403;
st.shared.f32 [%r650+-768], %f402;
st.shared.f32 [%r650+-640], %f401;
st.shared.f32 [%r650+-512], %f400;
st.shared.f32 [%r650+-384], %f399;
st.shared.f32 [%r650+-256], %f398;
st.shared.f32 [%r650+-128], %f397;

$L__BB180_5:
setp.ne.s32 %p7, %r1, 0;
bar.sync 0;
@%p7 bra $L__BB180_30;

mov.u32 %r651, %ctaid.z;
mad.lo.s32 %r652, %r5, %r41, %r4;
mad.lo.s32 %r653, %r651, %r47, %r652;
cvt.u64.u32 %rd16, %r653;
shl.b32 %r654, %r2, 2;
mov.u32 %r655, _ZZ13mul_mat_vec_qIL9ggml_type13ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r30, %r655, %r654;
mov.u32 %r656, 2;
ld.shared.f32 %f253, [%r30];
add.ftz.f32 %f254, %f253, %f408;
mov.b32 %r657, %f254;
mov.u32 %r658, 31;
mov.u32 %r659, 16;
mov.u32 %r660, -1;
shfl.sync.bfly.b32 %r661|%p8, %r657, %r659, %r658, %r660;
mov.b32 %f255, %r661;
add.ftz.f32 %f256, %f254, %f255;
mov.b32 %r662, %f256;
mov.u32 %r663, 8;
shfl.sync.bfly.b32 %r664|%p9, %r662, %r663, %r658, %r660;
mov.b32 %f257, %r664;
add.ftz.f32 %f258, %f256, %f257;
mov.b32 %r665, %f258;
mov.u32 %r666, 4;
shfl.sync.bfly.b32 %r667|%p10, %r665, %r666, %r658, %r660;
mov.b32 %f259, %r667;
add.ftz.f32 %f260, %f258, %f259;
mov.b32 %r668, %f260;
shfl.sync.bfly.b32 %r669|%p11, %r668, %r656, %r658, %r660;
mov.b32 %f261, %r669;
add.ftz.f32 %f262, %f260, %f261;
mov.b32 %r670, %f262;
mov.u32 %r671, 1;
shfl.sync.bfly.b32 %r672|%p12, %r670, %r671, %r658, %r660;
mov.b32 %f263, %r672;
add.ftz.f32 %f37, %f262, %f263;
setp.lt.u32 %p13, %r4, %r35;
setp.eq.s32 %p14, %r2, 0;
and.pred %p1, %p14, %p13;
mul.wide.u32 %rd76, %r653, 4;
add.s64 %rd17, %rd1, %rd76;
not.pred %p15, %p1;
@%p15 bra $L__BB180_8;

st.global.f32 [%rd17], %f37;

$L__BB180_8:
ld.shared.f32 %f264, [%r30+128];
add.ftz.f32 %f265, %f264, %f407;
mov.b32 %r673, %f265;
shfl.sync.bfly.b32 %r677|%p16, %r673, %r659, %r658, %r660;
mov.b32 %f266, %r677;
add.ftz.f32 %f267, %f265, %f266;
mov.b32 %r678, %f267;
shfl.sync.bfly.b32 %r680|%p17, %r678, %r663, %r658, %r660;
mov.b32 %f268, %r680;
add.ftz.f32 %f269, %f267, %f268;
mov.b32 %r681, %f269;
shfl.sync.bfly.b32 %r683|%p18, %r681, %r666, %r658, %r660;
mov.b32 %f270, %r683;
add.ftz.f32 %f271, %f269, %f270;
mov.b32 %r684, %f271;
shfl.sync.bfly.b32 %r686|%p19, %r684, %r656, %r658, %r660;
mov.b32 %f272, %r686;
add.ftz.f32 %f273, %f271, %f272;
mov.b32 %r687, %f273;
shfl.sync.bfly.b32 %r689|%p20, %r687, %r671, %r658, %r660;
mov.b32 %f274, %r689;
add.ftz.f32 %f38, %f273, %f274;
add.s32 %r690, %r4, 1;
setp.lt.u32 %p21, %r690, %r35;
setp.eq.s32 %p22, %r2, 1;
and.pred %p2, %p22, %p21;
not.pred %p23, %p2;
@%p23 bra $L__BB180_10;

st.global.f32 [%rd17+4], %f38;

$L__BB180_10:
ld.shared.f32 %f275, [%r30+256];
add.ftz.f32 %f276, %f275, %f406;
mov.b32 %r691, %f276;
mov.u32 %r692, 31;
mov.u32 %r693, 16;
mov.u32 %r694, -1;
shfl.sync.bfly.b32 %r695|%p24, %r691, %r693, %r692, %r694;
mov.b32 %f277, %r695;
add.ftz.f32 %f278, %f276, %f277;
mov.b32 %r696, %f278;
mov.u32 %r697, 8;
shfl.sync.bfly.b32 %r698|%p25, %r696, %r697, %r692, %r694;
mov.b32 %f279, %r698;
add.ftz.f32 %f280, %f278, %f279;
mov.b32 %r699, %f280;
mov.u32 %r700, 4;
shfl.sync.bfly.b32 %r701|%p26, %r699, %r700, %r692, %r694;
mov.b32 %f281, %r701;
add.ftz.f32 %f282, %f280, %f281;
mov.b32 %r702, %f282;
mov.u32 %r703, 2;
shfl.sync.bfly.b32 %r704|%p27, %r702, %r703, %r692, %r694;
mov.b32 %f283, %r704;
add.ftz.f32 %f284, %f282, %f283;
mov.b32 %r705, %f284;
mov.u32 %r706, 1;
shfl.sync.bfly.b32 %r707|%p28, %r705, %r706, %r692, %r694;
mov.b32 %f285, %r707;
add.ftz.f32 %f39, %f284, %f285;
@%p15 bra $L__BB180_12;

cvt.u64.u32 %rd77, %r35;
add.s64 %rd78, %rd77, %rd16;
shl.b64 %rd79, %rd78, 2;
add.s64 %rd80, %rd1, %rd79;
st.global.f32 [%rd80], %f39;

$L__BB180_12:
ld.shared.f32 %f286, [%r30+384];
add.ftz.f32 %f287, %f286, %f405;
mov.b32 %r708, %f287;
shfl.sync.bfly.b32 %r712|%p30, %r708, %r693, %r692, %r694;
mov.b32 %f288, %r712;
add.ftz.f32 %f289, %f287, %f288;
mov.b32 %r713, %f289;
shfl.sync.bfly.b32 %r715|%p31, %r713, %r697, %r692, %r694;
mov.b32 %f290, %r715;
add.ftz.f32 %f291, %f289, %f290;
mov.b32 %r716, %f291;
shfl.sync.bfly.b32 %r718|%p32, %r716, %r700, %r692, %r694;
mov.b32 %f292, %r718;
add.ftz.f32 %f293, %f291, %f292;
mov.b32 %r719, %f293;
shfl.sync.bfly.b32 %r721|%p33, %r719, %r703, %r692, %r694;
mov.b32 %f294, %r721;
add.ftz.f32 %f295, %f293, %f294;
mov.b32 %r722, %f295;
shfl.sync.bfly.b32 %r724|%p34, %r722, %r706, %r692, %r694;
mov.b32 %f296, %r724;
add.ftz.f32 %f40, %f295, %f296;
@%p23 bra $L__BB180_14;

add.s32 %r725, %r35, 1;
cvt.u64.u32 %rd81, %r725;
add.s64 %rd82, %rd81, %rd16;
shl.b64 %rd83, %rd82, 2;
add.s64 %rd84, %rd1, %rd83;
st.global.f32 [%rd84], %f40;

$L__BB180_14:
shl.b32 %r726, %r35, 1;
mov.u32 %r727, 1;
ld.shared.f32 %f297, [%r30+512];
add.ftz.f32 %f298, %f297, %f404;
mov.b32 %r728, %f298;
mov.u32 %r729, 31;
mov.u32 %r730, 16;
mov.u32 %r731, -1;
shfl.sync.bfly.b32 %r732|%p36, %r728, %r730, %r729, %r731;
mov.b32 %f299, %r732;
add.ftz.f32 %f300, %f298, %f299;
mov.b32 %r733, %f300;
mov.u32 %r734, 8;
shfl.sync.bfly.b32 %r735|%p37, %r733, %r734, %r729, %r731;
mov.b32 %f301, %r735;
add.ftz.f32 %f302, %f300, %f301;
mov.b32 %r736, %f302;
mov.u32 %r737, 4;
shfl.sync.bfly.b32 %r738|%p38, %r736, %r737, %r729, %r731;
mov.b32 %f303, %r738;
add.ftz.f32 %f304, %f302, %f303;
mov.b32 %r739, %f304;
mov.u32 %r740, 2;
shfl.sync.bfly.b32 %r741|%p39, %r739, %r740, %r729, %r731;
mov.b32 %f305, %r741;
add.ftz.f32 %f306, %f304, %f305;
mov.b32 %r742, %f306;
shfl.sync.bfly.b32 %r743|%p40, %r742, %r727, %r729, %r731;
mov.b32 %f307, %r743;
add.ftz.f32 %f41, %f306, %f307;
cvt.u64.u32 %rd85, %r726;
add.s64 %rd86, %rd85, %rd16;
shl.b64 %rd87, %rd86, 2;
add.s64 %rd18, %rd1, %rd87;
@%p15 bra $L__BB180_16;

st.global.f32 [%rd18], %f41;

$L__BB180_16:
ld.shared.f32 %f308, [%r30+640];
add.ftz.f32 %f309, %f308, %f403;
mov.b32 %r744, %f309;
shfl.sync.bfly.b32 %r748|%p42, %r744, %r730, %r729, %r731;
mov.b32 %f310, %r748;
add.ftz.f32 %f311, %f309, %f310;
mov.b32 %r749, %f311;
shfl.sync.bfly.b32 %r751|%p43, %r749, %r734, %r729, %r731;
mov.b32 %f312, %r751;
add.ftz.f32 %f313, %f311, %f312;
mov.b32 %r752, %f313;
shfl.sync.bfly.b32 %r754|%p44, %r752, %r737, %r729, %r731;
mov.b32 %f314, %r754;
add.ftz.f32 %f315, %f313, %f314;
mov.b32 %r755, %f315;
shfl.sync.bfly.b32 %r757|%p45, %r755, %r740, %r729, %r731;
mov.b32 %f316, %r757;
add.ftz.f32 %f317, %f315, %f316;
mov.b32 %r758, %f317;
shfl.sync.bfly.b32 %r760|%p46, %r758, %r727, %r729, %r731;
mov.b32 %f318, %r760;
add.ftz.f32 %f42, %f317, %f318;
@%p23 bra $L__BB180_18;

st.global.f32 [%rd18+4], %f42;

$L__BB180_18:
mul.lo.s32 %r31, %r35, 3;
ld.shared.f32 %f319, [%r30+768];
add.ftz.f32 %f320, %f319, %f402;
mov.b32 %r761, %f320;
mov.u32 %r762, 31;
mov.u32 %r763, 16;
mov.u32 %r764, -1;
shfl.sync.bfly.b32 %r765|%p48, %r761, %r763, %r762, %r764;
mov.b32 %f321, %r765;
add.ftz.f32 %f322, %f320, %f321;
mov.b32 %r766, %f322;
mov.u32 %r767, 8;
shfl.sync.bfly.b32 %r768|%p49, %r766, %r767, %r762, %r764;
mov.b32 %f323, %r768;
add.ftz.f32 %f324, %f322, %f323;
mov.b32 %r769, %f324;
mov.u32 %r770, 4;
shfl.sync.bfly.b32 %r771|%p50, %r769, %r770, %r762, %r764;
mov.b32 %f325, %r771;
add.ftz.f32 %f326, %f324, %f325;
mov.b32 %r772, %f326;
mov.u32 %r773, 2;
shfl.sync.bfly.b32 %r774|%p51, %r772, %r773, %r762, %r764;
mov.b32 %f327, %r774;
add.ftz.f32 %f328, %f326, %f327;
mov.b32 %r775, %f328;
mov.u32 %r776, 1;
shfl.sync.bfly.b32 %r777|%p52, %r775, %r776, %r762, %r764;
mov.b32 %f329, %r777;
add.ftz.f32 %f43, %f328, %f329;
@%p15 bra $L__BB180_20;

cvt.u64.u32 %rd88, %r31;
add.s64 %rd89, %rd88, %rd16;
shl.b64 %rd90, %rd89, 2;
add.s64 %rd91, %rd1, %rd90;
st.global.f32 [%rd91], %f43;

$L__BB180_20:
ld.shared.f32 %f330, [%r30+896];
add.ftz.f32 %f331, %f330, %f401;
mov.b32 %r778, %f331;
shfl.sync.bfly.b32 %r782|%p54, %r778, %r763, %r762, %r764;
mov.b32 %f332, %r782;
add.ftz.f32 %f333, %f331, %f332;
mov.b32 %r783, %f333;
shfl.sync.bfly.b32 %r785|%p55, %r783, %r767, %r762, %r764;
mov.b32 %f334, %r785;
add.ftz.f32 %f335, %f333, %f334;
mov.b32 %r786, %f335;
shfl.sync.bfly.b32 %r788|%p56, %r786, %r770, %r762, %r764;
mov.b32 %f336, %r788;
add.ftz.f32 %f337, %f335, %f336;
mov.b32 %r789, %f337;
shfl.sync.bfly.b32 %r791|%p57, %r789, %r773, %r762, %r764;
mov.b32 %f338, %r791;
add.ftz.f32 %f339, %f337, %f338;
mov.b32 %r792, %f339;
shfl.sync.bfly.b32 %r794|%p58, %r792, %r776, %r762, %r764;
mov.b32 %f340, %r794;
add.ftz.f32 %f44, %f339, %f340;
@%p23 bra $L__BB180_22;

add.s32 %r795, %r31, 1;
cvt.u64.u32 %rd92, %r795;
add.s64 %rd93, %rd92, %rd16;
shl.b64 %rd94, %rd93, 2;
add.s64 %rd95, %rd1, %rd94;
st.global.f32 [%rd95], %f44;

$L__BB180_22:
shl.b32 %r796, %r35, 2;
mov.u32 %r797, 2;
ld.shared.f32 %f341, [%r30+1024];
add.ftz.f32 %f342, %f341, %f400;
mov.b32 %r798, %f342;
mov.u32 %r799, 31;
mov.u32 %r800, 16;
mov.u32 %r801, -1;
shfl.sync.bfly.b32 %r802|%p60, %r798, %r800, %r799, %r801;
mov.b32 %f343, %r802;
add.ftz.f32 %f344, %f342, %f343;
mov.b32 %r803, %f344;
mov.u32 %r804, 8;
shfl.sync.bfly.b32 %r805|%p61, %r803, %r804, %r799, %r801;
mov.b32 %f345, %r805;
add.ftz.f32 %f346, %f344, %f345;
mov.b32 %r806, %f346;
mov.u32 %r807, 4;
shfl.sync.bfly.b32 %r808|%p62, %r806, %r807, %r799, %r801;
mov.b32 %f347, %r808;
add.ftz.f32 %f348, %f346, %f347;
mov.b32 %r809, %f348;
shfl.sync.bfly.b32 %r810|%p63, %r809, %r797, %r799, %r801;
mov.b32 %f349, %r810;
add.ftz.f32 %f350, %f348, %f349;
mov.b32 %r811, %f350;
mov.u32 %r812, 1;
shfl.sync.bfly.b32 %r813|%p64, %r811, %r812, %r799, %r801;
mov.b32 %f351, %r813;
add.ftz.f32 %f45, %f350, %f351;
cvt.u64.u32 %rd96, %r796;
add.s64 %rd97, %rd96, %rd16;
shl.b64 %rd98, %rd97, 2;
add.s64 %rd19, %rd1, %rd98;
@%p15 bra $L__BB180_24;

st.global.f32 [%rd19], %f45;

$L__BB180_24:
ld.shared.f32 %f352, [%r30+1152];
add.ftz.f32 %f353, %f352, %f399;
mov.b32 %r814, %f353;
shfl.sync.bfly.b32 %r818|%p66, %r814, %r800, %r799, %r801;
mov.b32 %f354, %r818;
add.ftz.f32 %f355, %f353, %f354;
mov.b32 %r819, %f355;
shfl.sync.bfly.b32 %r821|%p67, %r819, %r804, %r799, %r801;
mov.b32 %f356, %r821;
add.ftz.f32 %f357, %f355, %f356;
mov.b32 %r822, %f357;
shfl.sync.bfly.b32 %r824|%p68, %r822, %r807, %r799, %r801;
mov.b32 %f358, %r824;
add.ftz.f32 %f359, %f357, %f358;
mov.b32 %r825, %f359;
shfl.sync.bfly.b32 %r827|%p69, %r825, %r797, %r799, %r801;
mov.b32 %f360, %r827;
add.ftz.f32 %f361, %f359, %f360;
mov.b32 %r828, %f361;
shfl.sync.bfly.b32 %r830|%p70, %r828, %r812, %r799, %r801;
mov.b32 %f362, %r830;
add.ftz.f32 %f46, %f361, %f362;
@%p23 bra $L__BB180_26;

st.global.f32 [%rd19+4], %f46;

$L__BB180_26:
mul.lo.s32 %r32, %r35, 5;
ld.shared.f32 %f363, [%r30+1280];
add.ftz.f32 %f364, %f363, %f398;
mov.b32 %r831, %f364;
mov.u32 %r832, 31;
mov.u32 %r833, 16;
mov.u32 %r834, -1;
shfl.sync.bfly.b32 %r835|%p72, %r831, %r833, %r832, %r834;
mov.b32 %f365, %r835;
add.ftz.f32 %f366, %f364, %f365;
mov.b32 %r836, %f366;
mov.u32 %r837, 8;
shfl.sync.bfly.b32 %r838|%p73, %r836, %r837, %r832, %r834;
mov.b32 %f367, %r838;
add.ftz.f32 %f368, %f366, %f367;
mov.b32 %r839, %f368;
mov.u32 %r840, 4;
shfl.sync.bfly.b32 %r841|%p74, %r839, %r840, %r832, %r834;
mov.b32 %f369, %r841;
add.ftz.f32 %f370, %f368, %f369;
mov.b32 %r842, %f370;
mov.u32 %r843, 2;
shfl.sync.bfly.b32 %r844|%p75, %r842, %r843, %r832, %r834;
mov.b32 %f371, %r844;
add.ftz.f32 %f372, %f370, %f371;
mov.b32 %r845, %f372;
mov.u32 %r846, 1;
shfl.sync.bfly.b32 %r847|%p76, %r845, %r846, %r832, %r834;
mov.b32 %f373, %r847;
add.ftz.f32 %f47, %f372, %f373;
@%p15 bra $L__BB180_28;

cvt.u64.u32 %rd99, %r32;
add.s64 %rd100, %rd99, %rd16;
shl.b64 %rd101, %rd100, 2;
add.s64 %rd102, %rd1, %rd101;
st.global.f32 [%rd102], %f47;

$L__BB180_28:
ld.shared.f32 %f374, [%r30+1408];
add.ftz.f32 %f375, %f374, %f397;
mov.b32 %r848, %f375;
shfl.sync.bfly.b32 %r852|%p78, %r848, %r833, %r832, %r834;
mov.b32 %f376, %r852;
add.ftz.f32 %f377, %f375, %f376;
mov.b32 %r853, %f377;
shfl.sync.bfly.b32 %r855|%p79, %r853, %r837, %r832, %r834;
mov.b32 %f378, %r855;
add.ftz.f32 %f379, %f377, %f378;
mov.b32 %r856, %f379;
shfl.sync.bfly.b32 %r858|%p80, %r856, %r840, %r832, %r834;
mov.b32 %f380, %r858;
add.ftz.f32 %f381, %f379, %f380;
mov.b32 %r859, %f381;
shfl.sync.bfly.b32 %r861|%p81, %r859, %r843, %r832, %r834;
mov.b32 %f382, %r861;
add.ftz.f32 %f383, %f381, %f382;
mov.b32 %r862, %f383;
shfl.sync.bfly.b32 %r864|%p82, %r862, %r846, %r832, %r834;
mov.b32 %f384, %r864;
add.ftz.f32 %f48, %f383, %f384;
@%p23 bra $L__BB180_30;

add.s32 %r865, %r32, 1;
cvt.u64.u32 %rd103, %r865;
add.s64 %rd104, %rd103, %rd16;
shl.b64 %rd105, %rd104, 2;
add.s64 %rd106, %rd1, %rd105;
st.global.f32 [%rd106], %f48;

$L__BB180_30:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type13ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type13ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type13ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type13ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type13ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type13ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type13ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type13ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type13ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<96>;
.reg .b16 %rs<15>;
.reg .f32 %f<477>;
.reg .b32 %r<996>;
.reg .b64 %rd<126>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type13ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1792];

ld.param.u64 %rd20, [_Z13mul_mat_vec_qIL9ggml_type13ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd21, [_Z13mul_mat_vec_qIL9ggml_type13ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd22, [_Z13mul_mat_vec_qIL9ggml_type13ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r50, [_Z13mul_mat_vec_qIL9ggml_type13ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type13ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type13ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type13ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type13ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type13ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type13ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r47, [_Z13mul_mat_vec_qIL9ggml_type13ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r48, [_Z13mul_mat_vec_qIL9ggml_type13ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r49, [_Z13mul_mat_vec_qIL9ggml_type13ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r45, [_Z13mul_mat_vec_qIL9ggml_type13ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type13ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type13ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type13ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r51, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r51, %r2;
mov.u32 %r52, %ctaid.x;
shl.b32 %r4, %r52, 1;
shr.s32 %r53, %r3, 31;
shr.u32 %r54, %r53, 28;
add.s32 %r55, %r3, %r54;
shr.s32 %r995, %r55, 4;
shr.u32 %r6, %r50, 8;
setp.ge.s32 %p3, %r995, %r6;
mov.f32 %f463, 0f00000000;
mov.f32 %f464, %f463;
mov.f32 %f465, %f463;
mov.f32 %f466, %f463;
mov.f32 %f467, %f463;
mov.f32 %f468, %f463;
mov.f32 %f469, %f463;
mov.f32 %f470, %f463;
mov.f32 %f471, %f463;
mov.f32 %f472, %f463;
mov.f32 %f473, %f463;
mov.f32 %f474, %f463;
mov.f32 %f475, %f463;
mov.f32 %f476, %f463;
@%p3 bra $L__BB181_3;

mov.u32 %r56, %ctaid.y;
mov.u32 %r57, %ctaid.z;
mul.hi.u32 %r58, %r57, %r44;
add.s32 %r59, %r58, %r57;
shr.u32 %r60, %r59, %r45;
mul.lo.s32 %r61, %r60, %r47;
mul.hi.u32 %r62, %r56, %r38;
add.s32 %r63, %r62, %r56;
shr.u32 %r64, %r63, %r39;
mul.lo.s32 %r65, %r64, %r41;
mul.lo.s32 %r66, %r56, %r42;
cvt.u64.u32 %rd23, %r66;
cvta.to.global.u64 %rd125, %rd20;
and.b32 %r70, %r55, -16;
sub.s32 %r7, %r3, %r70;
shr.s32 %r71, %r7, 31;
shr.u32 %r72, %r71, 30;
add.s32 %r73, %r7, %r72;
shr.s32 %r74, %r73, 2;
shl.b32 %r8, %r74, 1;
shl.b32 %r75, %r74, 5;
and.b32 %r76, %r73, -4;
sub.s32 %r77, %r7, %r76;
shl.b32 %r78, %r77, 2;
add.s32 %r79, %r75, %r78;
cvt.s64.s32 %rd24, %r79;
cvt.s64.s32 %rd25, %r78;
cvt.s64.s32 %rd26, %r8;
add.s64 %rd27, %rd23, %rd26;
mul.lo.s32 %r80, %r57, %r48;
cvt.u64.u32 %rd28, %r80;
add.s64 %rd2, %rd27, %rd28;
cvt.s64.s32 %rd3, %r77;
shl.b32 %r81, %r36, 1;
shl.b32 %r988, %r995, 3;
add.s32 %r994, %r988, %r81;
mad.lo.s32 %r993, %r36, 3, %r988;
shl.b32 %r82, %r36, 2;
add.s32 %r992, %r988, %r82;
mad.lo.s32 %r991, %r36, 5, %r988;
mad.lo.s32 %r990, %r36, 6, %r988;
add.s32 %r989, %r36, %r988;
add.s32 %r83, %r995, %r61;
add.s32 %r84, %r83, %r65;
add.s32 %r85, %r4, 1;
mad.lo.s32 %r86, %r35, %r85, %r84;
mul.wide.s32 %rd4, %r86, 176;
mad.lo.s32 %r87, %r4, %r35, %r995;
add.s32 %r88, %r87, %r61;
add.s32 %r89, %r88, %r65;
mul.wide.s32 %rd5, %r89, 176;
add.s64 %rd29, %rd5, %rd24;
add.s64 %rd6, %rd29, 64;
cvt.u64.u32 %rd30, %r8;
and.b64 %rd31, %rd30, 2;
or.b64 %rd32, %rd31, %rd4;
or.b64 %rd7, %rd32, 8;
add.s64 %rd33, %rd5, %rd25;
add.s64 %rd8, %rd33, 32;
or.b64 %rd34, %rd31, %rd5;
or.b64 %rd9, %rd34, 12;
add.s64 %rd35, %rd4, %rd25;
add.s64 %rd10, %rd35, 16;
add.s64 %rd36, %rd4, %rd24;
add.s64 %rd11, %rd36, 48;
cvta.to.global.u64 %rd12, %rd21;
mov.f32 %f463, 0f00000000;
shl.b64 %rd44, %rd3, 2;

$L__BB181_2:
cvt.u64.u32 %rd37, %r988;
add.s64 %rd38, %rd2, %rd37;
add.s64 %rd39, %rd125, %rd6;
ld.global.nc.u32 %r594, [%rd39+-16];
ld.global.nc.u32 %r595, [%rd39];
add.s64 %rd40, %rd125, %rd8;
ld.global.nc.u32 %r596, [%rd40+-16];
shr.s32 %r597, %r596, %r8;
ld.global.nc.u32 %r598, [%rd40];
shr.s32 %r599, %r598, %r8;
add.s64 %rd41, %rd125, %rd9;
ld.global.nc.u16 %rs1, [%rd41+-8];
cvt.u32.u16 %r600, %rs1;
ld.global.nc.u16 %rs2, [%rd41+-4];
cvt.u32.u16 %r601, %rs2;
ld.global.nc.u16 %rs3, [%rd41];
cvt.u32.u16 %r602, %rs3;
and.b32 %r603, %r600, 16191;
and.b32 %r604, %r602, 3855;
and.b32 %r605, %r600, 49344;
shr.u32 %r606, %r605, 2;
or.b32 %r607, %r604, %r606;
setp.gt.s32 %p4, %r7, 7;
and.b32 %r608, %r601, 16191;
shr.u32 %r609, %r602, 4;
and.b32 %r610, %r609, 3855;
and.b32 %r611, %r601, 49344;
shr.u32 %r612, %r611, 2;
or.b32 %r613, %r610, %r612;
mul.lo.s64 %rd42, %rd38, 36;
add.s64 %rd43, %rd12, %rd42;
ld.global.nc.u32 %r90, [%rd43];

{.reg .f16 low,high;
mov.b32 {low,high},%r90;
cvt.f32.f16 %f85, low;}


add.s64 %rd45, %rd43, %rd44;
ld.global.nc.u32 %r98, [%rd45+4];
ld.global.nc.u32 %r94, [%rd45+20];
ld.global.nc.u32 %r91, [%rd43+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r91;
cvt.f32.f16 %f86, low;}


ld.global.nc.u32 %r114, [%rd45+40];
ld.global.nc.u32 %r110, [%rd45+56];
add.s64 %rd46, %rd125, %rd5;
and.b32 %r614, %r594, 252645135;
and.b32 %r615, %r595, 252645135;
shl.b32 %r616, %r597, 4;
and.b32 %r617, %r616, 269488144;
shl.b32 %r618, %r599, 4;
and.b32 %r619, %r618, 269488144;
or.b32 %r529, %r614, %r617;
or.b32 %r525, %r615, %r619;
mov.u32 %r587, 0;

dp4a.s32.s32 %r92, %r525, %r94, %r587;


dp4a.s32.s32 %r96, %r529, %r98, %r92;

mov.u32 %r589, 16843009;

dp4a.s32.s32 %r100, %r589, %r94, %r587;


dp4a.s32.s32 %r104, %r589, %r98, %r100;

selp.b32 %r620, %r607, %r603, %p4;
cvt.u16.u32 %rs4, %r620;
and.b32 %r621, %r620, 63;
mul.lo.s32 %r622, %r96, %r621;
cvt.rn.f32.s32 %f141, %r622;
fma.rn.ftz.f32 %f142, %f85, %f141, 0f00000000;
selp.b32 %r623, %r613, %r608, %p4;
cvt.u16.u32 %rs5, %r623;
and.b32 %r624, %r623, 63;
mul.lo.s32 %r625, %r104, %r624;
cvt.rn.f32.s32 %f143, %r625;
fma.rn.ftz.f32 %f144, %f85, %f143, 0f00000000;
shr.u32 %r626, %r594, 4;
and.b32 %r627, %r626, 252645135;
shr.u32 %r628, %r595, 4;
and.b32 %r629, %r628, 252645135;
shl.b32 %r630, %r597, 3;
and.b32 %r631, %r630, 269488144;
shl.b32 %r632, %r599, 3;
and.b32 %r633, %r632, 269488144;
or.b32 %r545, %r627, %r631;
or.b32 %r541, %r629, %r633;

dp4a.s32.s32 %r108, %r541, %r110, %r587;


dp4a.s32.s32 %r112, %r545, %r114, %r108;


dp4a.s32.s32 %r116, %r589, %r110, %r587;


dp4a.s32.s32 %r120, %r589, %r114, %r116;

shr.u16 %rs6, %rs4, 8;
cvt.u32.u16 %r634, %rs6;
mul.lo.s32 %r635, %r112, %r634;
cvt.rn.f32.s32 %f145, %r635;
fma.rn.ftz.f32 %f146, %f86, %f145, %f142;
shr.u16 %rs7, %rs5, 8;
cvt.u32.u16 %r636, %rs7;
mul.lo.s32 %r637, %r120, %r636;
cvt.rn.f32.s32 %f147, %r637;
fma.rn.ftz.f32 %f148, %f86, %f147, %f144;
ld.global.nc.u32 %r124, [%rd46];

{.reg .f16 low,high;
mov.b32 {low,high},%r124;
cvt.f32.f16 %f87, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r124;
cvt.f32.f16 %f88, high;}


mul.ftz.f32 %f149, %f146, %f87;
mul.ftz.f32 %f150, %f148, %f88;
sub.ftz.f32 %f151, %f149, %f150;
add.ftz.f32 %f476, %f151, %f476;
add.s64 %rd47, %rd125, %rd11;
add.s64 %rd48, %rd125, %rd10;
ld.global.nc.u32 %r638, [%rd47];
ld.global.nc.u32 %r639, [%rd47+16];
ld.global.nc.u32 %r640, [%rd48];
shr.s32 %r641, %r640, %r8;
ld.global.nc.u32 %r642, [%rd48+16];
shr.s32 %r643, %r642, %r8;
add.s64 %rd49, %rd125, %rd7;
ld.global.nc.u16 %rs8, [%rd49+-4];
cvt.u32.u16 %r644, %rs8;
ld.global.nc.u16 %rs9, [%rd49];
cvt.u32.u16 %r645, %rs9;
ld.global.nc.u16 %rs10, [%rd49+4];
cvt.u32.u16 %r646, %rs10;
and.b32 %r647, %r644, 16191;
and.b32 %r648, %r646, 3855;
and.b32 %r649, %r644, 49344;
shr.u32 %r650, %r649, 2;
or.b32 %r651, %r648, %r650;
and.b32 %r652, %r645, 16191;
shr.u32 %r653, %r646, 4;
and.b32 %r654, %r653, 3855;
and.b32 %r655, %r645, 49344;
shr.u32 %r656, %r655, 2;
or.b32 %r657, %r654, %r656;

{.reg .f16 low,high;
mov.b32 {low,high},%r90;
cvt.f32.f16 %f89, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r91;
cvt.f32.f16 %f90, low;}


add.s64 %rd50, %rd125, %rd4;
and.b32 %r658, %r638, 252645135;
and.b32 %r659, %r639, 252645135;
shl.b32 %r660, %r641, 4;
and.b32 %r661, %r660, 269488144;
shl.b32 %r662, %r643, 4;
and.b32 %r663, %r662, 269488144;
or.b32 %r565, %r658, %r661;
or.b32 %r561, %r659, %r663;

dp4a.s32.s32 %r128, %r561, %r94, %r587;


dp4a.s32.s32 %r132, %r565, %r98, %r128;


dp4a.s32.s32 %r136, %r589, %r94, %r587;


dp4a.s32.s32 %r140, %r589, %r98, %r136;

selp.b32 %r664, %r651, %r647, %p4;
cvt.u16.u32 %rs11, %r664;
and.b32 %r665, %r664, 63;
mul.lo.s32 %r666, %r132, %r665;
cvt.rn.f32.s32 %f152, %r666;
fma.rn.ftz.f32 %f153, %f89, %f152, 0f00000000;
selp.b32 %r667, %r657, %r652, %p4;
cvt.u16.u32 %rs12, %r667;
and.b32 %r668, %r667, 63;
mul.lo.s32 %r669, %r140, %r668;
cvt.rn.f32.s32 %f154, %r669;
fma.rn.ftz.f32 %f155, %f89, %f154, 0f00000000;
shr.u32 %r670, %r638, 4;
and.b32 %r671, %r670, 252645135;
shr.u32 %r672, %r639, 4;
and.b32 %r673, %r672, 252645135;
shl.b32 %r674, %r641, 3;
and.b32 %r675, %r674, 269488144;
shl.b32 %r676, %r643, 3;
and.b32 %r677, %r676, 269488144;
or.b32 %r581, %r671, %r675;
or.b32 %r577, %r673, %r677;

dp4a.s32.s32 %r144, %r577, %r110, %r587;


dp4a.s32.s32 %r148, %r581, %r114, %r144;


dp4a.s32.s32 %r152, %r589, %r110, %r587;


dp4a.s32.s32 %r156, %r589, %r114, %r152;

shr.u16 %rs13, %rs11, 8;
cvt.u32.u16 %r678, %rs13;
mul.lo.s32 %r679, %r148, %r678;
cvt.rn.f32.s32 %f156, %r679;
fma.rn.ftz.f32 %f157, %f90, %f156, %f153;
shr.u16 %rs14, %rs12, 8;
cvt.u32.u16 %r680, %rs14;
mul.lo.s32 %r681, %r156, %r680;
cvt.rn.f32.s32 %f158, %r681;
fma.rn.ftz.f32 %f159, %f90, %f158, %f155;
ld.global.nc.u32 %r160, [%rd50];

{.reg .f16 low,high;
mov.b32 {low,high},%r160;
cvt.f32.f16 %f91, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r160;
cvt.f32.f16 %f92, high;}


mul.ftz.f32 %f160, %f157, %f91;
mul.ftz.f32 %f161, %f159, %f92;
sub.ftz.f32 %f162, %f160, %f161;
add.ftz.f32 %f475, %f162, %f475;
cvt.u64.u32 %rd51, %r989;
add.s64 %rd52, %rd2, %rd51;
mul.lo.s64 %rd53, %rd52, 36;
add.s64 %rd54, %rd12, %rd53;
ld.global.nc.u32 %r162, [%rd54];

{.reg .f16 low,high;
mov.b32 {low,high},%r162;
cvt.f32.f16 %f93, low;}


add.s64 %rd55, %rd54, %rd44;
ld.global.nc.u32 %r170, [%rd55+4];
ld.global.nc.u32 %r166, [%rd55+20];
ld.global.nc.u32 %r163, [%rd54+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r163;
cvt.f32.f16 %f94, low;}


ld.global.nc.u32 %r186, [%rd55+40];
ld.global.nc.u32 %r182, [%rd55+56];

dp4a.s32.s32 %r164, %r525, %r166, %r587;


dp4a.s32.s32 %r168, %r529, %r170, %r164;


dp4a.s32.s32 %r172, %r589, %r166, %r587;


dp4a.s32.s32 %r176, %r589, %r170, %r172;

mul.lo.s32 %r682, %r168, %r621;
cvt.rn.f32.s32 %f163, %r682;
fma.rn.ftz.f32 %f164, %f93, %f163, 0f00000000;
mul.lo.s32 %r683, %r176, %r624;
cvt.rn.f32.s32 %f165, %r683;
fma.rn.ftz.f32 %f166, %f93, %f165, 0f00000000;

dp4a.s32.s32 %r180, %r541, %r182, %r587;


dp4a.s32.s32 %r184, %r545, %r186, %r180;


dp4a.s32.s32 %r188, %r589, %r182, %r587;


dp4a.s32.s32 %r192, %r589, %r186, %r188;

mul.lo.s32 %r684, %r184, %r634;
cvt.rn.f32.s32 %f167, %r684;
fma.rn.ftz.f32 %f168, %f94, %f167, %f164;
mul.lo.s32 %r685, %r192, %r636;
cvt.rn.f32.s32 %f169, %r685;
fma.rn.ftz.f32 %f170, %f94, %f169, %f166;

{.reg .f16 low,high;
mov.b32 {low,high},%r124;
cvt.f32.f16 %f95, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r124;
cvt.f32.f16 %f96, high;}


mul.ftz.f32 %f171, %f168, %f95;
mul.ftz.f32 %f172, %f170, %f96;
sub.ftz.f32 %f173, %f171, %f172;
add.ftz.f32 %f474, %f173, %f474;

{.reg .f16 low,high;
mov.b32 {low,high},%r162;
cvt.f32.f16 %f97, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r163;
cvt.f32.f16 %f98, low;}



dp4a.s32.s32 %r200, %r561, %r166, %r587;


dp4a.s32.s32 %r204, %r565, %r170, %r200;


dp4a.s32.s32 %r208, %r589, %r166, %r587;


dp4a.s32.s32 %r212, %r589, %r170, %r208;

mul.lo.s32 %r686, %r204, %r665;
cvt.rn.f32.s32 %f174, %r686;
fma.rn.ftz.f32 %f175, %f97, %f174, 0f00000000;
mul.lo.s32 %r687, %r212, %r668;
cvt.rn.f32.s32 %f176, %r687;
fma.rn.ftz.f32 %f177, %f97, %f176, 0f00000000;

dp4a.s32.s32 %r216, %r577, %r182, %r587;


dp4a.s32.s32 %r220, %r581, %r186, %r216;


dp4a.s32.s32 %r224, %r589, %r182, %r587;


dp4a.s32.s32 %r228, %r589, %r186, %r224;

mul.lo.s32 %r688, %r220, %r678;
cvt.rn.f32.s32 %f178, %r688;
fma.rn.ftz.f32 %f179, %f98, %f178, %f175;
mul.lo.s32 %r689, %r228, %r680;
cvt.rn.f32.s32 %f180, %r689;
fma.rn.ftz.f32 %f181, %f98, %f180, %f177;

{.reg .f16 low,high;
mov.b32 {low,high},%r160;
cvt.f32.f16 %f99, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r160;
cvt.f32.f16 %f100, high;}


mul.ftz.f32 %f182, %f179, %f99;
mul.ftz.f32 %f183, %f181, %f100;
sub.ftz.f32 %f184, %f182, %f183;
add.ftz.f32 %f473, %f184, %f473;
cvt.u64.u32 %rd56, %r994;
add.s64 %rd57, %rd2, %rd56;
mul.lo.s64 %rd58, %rd57, 36;
add.s64 %rd59, %rd12, %rd58;
ld.global.nc.u32 %r234, [%rd59];

{.reg .f16 low,high;
mov.b32 {low,high},%r234;
cvt.f32.f16 %f101, low;}


add.s64 %rd60, %rd59, %rd44;
ld.global.nc.u32 %r242, [%rd60+4];
ld.global.nc.u32 %r238, [%rd60+20];
ld.global.nc.u32 %r235, [%rd59+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r235;
cvt.f32.f16 %f102, low;}


ld.global.nc.u32 %r258, [%rd60+40];
ld.global.nc.u32 %r254, [%rd60+56];

dp4a.s32.s32 %r236, %r525, %r238, %r587;


dp4a.s32.s32 %r240, %r529, %r242, %r236;


dp4a.s32.s32 %r244, %r589, %r238, %r587;


dp4a.s32.s32 %r248, %r589, %r242, %r244;

mul.lo.s32 %r690, %r240, %r621;
cvt.rn.f32.s32 %f185, %r690;
fma.rn.ftz.f32 %f186, %f101, %f185, 0f00000000;
mul.lo.s32 %r691, %r248, %r624;
cvt.rn.f32.s32 %f187, %r691;
fma.rn.ftz.f32 %f188, %f101, %f187, 0f00000000;

dp4a.s32.s32 %r252, %r541, %r254, %r587;


dp4a.s32.s32 %r256, %r545, %r258, %r252;


dp4a.s32.s32 %r260, %r589, %r254, %r587;


dp4a.s32.s32 %r264, %r589, %r258, %r260;

mul.lo.s32 %r692, %r256, %r634;
cvt.rn.f32.s32 %f189, %r692;
fma.rn.ftz.f32 %f190, %f102, %f189, %f186;
mul.lo.s32 %r693, %r264, %r636;
cvt.rn.f32.s32 %f191, %r693;
fma.rn.ftz.f32 %f192, %f102, %f191, %f188;

{.reg .f16 low,high;
mov.b32 {low,high},%r124;
cvt.f32.f16 %f103, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r124;
cvt.f32.f16 %f104, high;}


mul.ftz.f32 %f193, %f190, %f103;
mul.ftz.f32 %f194, %f192, %f104;
sub.ftz.f32 %f195, %f193, %f194;
add.ftz.f32 %f472, %f195, %f472;

{.reg .f16 low,high;
mov.b32 {low,high},%r234;
cvt.f32.f16 %f105, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r235;
cvt.f32.f16 %f106, low;}



dp4a.s32.s32 %r272, %r561, %r238, %r587;


dp4a.s32.s32 %r276, %r565, %r242, %r272;


dp4a.s32.s32 %r280, %r589, %r238, %r587;


dp4a.s32.s32 %r284, %r589, %r242, %r280;

mul.lo.s32 %r694, %r276, %r665;
cvt.rn.f32.s32 %f196, %r694;
fma.rn.ftz.f32 %f197, %f105, %f196, 0f00000000;
mul.lo.s32 %r695, %r284, %r668;
cvt.rn.f32.s32 %f198, %r695;
fma.rn.ftz.f32 %f199, %f105, %f198, 0f00000000;

dp4a.s32.s32 %r288, %r577, %r254, %r587;


dp4a.s32.s32 %r292, %r581, %r258, %r288;


dp4a.s32.s32 %r296, %r589, %r254, %r587;


dp4a.s32.s32 %r300, %r589, %r258, %r296;

mul.lo.s32 %r696, %r292, %r678;
cvt.rn.f32.s32 %f200, %r696;
fma.rn.ftz.f32 %f201, %f106, %f200, %f197;
mul.lo.s32 %r697, %r300, %r680;
cvt.rn.f32.s32 %f202, %r697;
fma.rn.ftz.f32 %f203, %f106, %f202, %f199;

{.reg .f16 low,high;
mov.b32 {low,high},%r160;
cvt.f32.f16 %f107, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r160;
cvt.f32.f16 %f108, high;}


mul.ftz.f32 %f204, %f201, %f107;
mul.ftz.f32 %f205, %f203, %f108;
sub.ftz.f32 %f206, %f204, %f205;
add.ftz.f32 %f471, %f206, %f471;
cvt.u64.u32 %rd61, %r993;
add.s64 %rd62, %rd2, %rd61;
mul.lo.s64 %rd63, %rd62, 36;
add.s64 %rd64, %rd12, %rd63;
ld.global.nc.u32 %r306, [%rd64];

{.reg .f16 low,high;
mov.b32 {low,high},%r306;
cvt.f32.f16 %f109, low;}


add.s64 %rd65, %rd64, %rd44;
ld.global.nc.u32 %r314, [%rd65+4];
ld.global.nc.u32 %r310, [%rd65+20];
ld.global.nc.u32 %r307, [%rd64+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r307;
cvt.f32.f16 %f110, low;}


ld.global.nc.u32 %r330, [%rd65+40];
ld.global.nc.u32 %r326, [%rd65+56];

dp4a.s32.s32 %r308, %r525, %r310, %r587;


dp4a.s32.s32 %r312, %r529, %r314, %r308;


dp4a.s32.s32 %r316, %r589, %r310, %r587;


dp4a.s32.s32 %r320, %r589, %r314, %r316;

mul.lo.s32 %r698, %r312, %r621;
cvt.rn.f32.s32 %f207, %r698;
fma.rn.ftz.f32 %f208, %f109, %f207, 0f00000000;
mul.lo.s32 %r699, %r320, %r624;
cvt.rn.f32.s32 %f209, %r699;
fma.rn.ftz.f32 %f210, %f109, %f209, 0f00000000;

dp4a.s32.s32 %r324, %r541, %r326, %r587;


dp4a.s32.s32 %r328, %r545, %r330, %r324;


dp4a.s32.s32 %r332, %r589, %r326, %r587;


dp4a.s32.s32 %r336, %r589, %r330, %r332;

mul.lo.s32 %r700, %r328, %r634;
cvt.rn.f32.s32 %f211, %r700;
fma.rn.ftz.f32 %f212, %f110, %f211, %f208;
mul.lo.s32 %r701, %r336, %r636;
cvt.rn.f32.s32 %f213, %r701;
fma.rn.ftz.f32 %f214, %f110, %f213, %f210;

{.reg .f16 low,high;
mov.b32 {low,high},%r124;
cvt.f32.f16 %f111, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r124;
cvt.f32.f16 %f112, high;}


mul.ftz.f32 %f215, %f212, %f111;
mul.ftz.f32 %f216, %f214, %f112;
sub.ftz.f32 %f217, %f215, %f216;
add.ftz.f32 %f470, %f217, %f470;

{.reg .f16 low,high;
mov.b32 {low,high},%r306;
cvt.f32.f16 %f113, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r307;
cvt.f32.f16 %f114, low;}



dp4a.s32.s32 %r344, %r561, %r310, %r587;


dp4a.s32.s32 %r348, %r565, %r314, %r344;


dp4a.s32.s32 %r352, %r589, %r310, %r587;


dp4a.s32.s32 %r356, %r589, %r314, %r352;

mul.lo.s32 %r702, %r348, %r665;
cvt.rn.f32.s32 %f218, %r702;
fma.rn.ftz.f32 %f219, %f113, %f218, 0f00000000;
mul.lo.s32 %r703, %r356, %r668;
cvt.rn.f32.s32 %f220, %r703;
fma.rn.ftz.f32 %f221, %f113, %f220, 0f00000000;

dp4a.s32.s32 %r360, %r577, %r326, %r587;


dp4a.s32.s32 %r364, %r581, %r330, %r360;


dp4a.s32.s32 %r368, %r589, %r326, %r587;


dp4a.s32.s32 %r372, %r589, %r330, %r368;

mul.lo.s32 %r704, %r364, %r678;
cvt.rn.f32.s32 %f222, %r704;
fma.rn.ftz.f32 %f223, %f114, %f222, %f219;
mul.lo.s32 %r705, %r372, %r680;
cvt.rn.f32.s32 %f224, %r705;
fma.rn.ftz.f32 %f225, %f114, %f224, %f221;

{.reg .f16 low,high;
mov.b32 {low,high},%r160;
cvt.f32.f16 %f115, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r160;
cvt.f32.f16 %f116, high;}


mul.ftz.f32 %f226, %f223, %f115;
mul.ftz.f32 %f227, %f225, %f116;
sub.ftz.f32 %f228, %f226, %f227;
add.ftz.f32 %f469, %f228, %f469;
cvt.u64.u32 %rd66, %r992;
add.s64 %rd67, %rd2, %rd66;
mul.lo.s64 %rd68, %rd67, 36;
add.s64 %rd69, %rd12, %rd68;
ld.global.nc.u32 %r378, [%rd69];

{.reg .f16 low,high;
mov.b32 {low,high},%r378;
cvt.f32.f16 %f117, low;}


add.s64 %rd70, %rd69, %rd44;
ld.global.nc.u32 %r386, [%rd70+4];
ld.global.nc.u32 %r382, [%rd70+20];
ld.global.nc.u32 %r379, [%rd69+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r379;
cvt.f32.f16 %f118, low;}


ld.global.nc.u32 %r402, [%rd70+40];
ld.global.nc.u32 %r398, [%rd70+56];

dp4a.s32.s32 %r380, %r525, %r382, %r587;


dp4a.s32.s32 %r384, %r529, %r386, %r380;


dp4a.s32.s32 %r388, %r589, %r382, %r587;


dp4a.s32.s32 %r392, %r589, %r386, %r388;

mul.lo.s32 %r706, %r384, %r621;
cvt.rn.f32.s32 %f229, %r706;
fma.rn.ftz.f32 %f230, %f117, %f229, 0f00000000;
mul.lo.s32 %r707, %r392, %r624;
cvt.rn.f32.s32 %f231, %r707;
fma.rn.ftz.f32 %f232, %f117, %f231, 0f00000000;

dp4a.s32.s32 %r396, %r541, %r398, %r587;


dp4a.s32.s32 %r400, %r545, %r402, %r396;


dp4a.s32.s32 %r404, %r589, %r398, %r587;


dp4a.s32.s32 %r408, %r589, %r402, %r404;

mul.lo.s32 %r708, %r400, %r634;
cvt.rn.f32.s32 %f233, %r708;
fma.rn.ftz.f32 %f234, %f118, %f233, %f230;
mul.lo.s32 %r709, %r408, %r636;
cvt.rn.f32.s32 %f235, %r709;
fma.rn.ftz.f32 %f236, %f118, %f235, %f232;

{.reg .f16 low,high;
mov.b32 {low,high},%r124;
cvt.f32.f16 %f119, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r124;
cvt.f32.f16 %f120, high;}


mul.ftz.f32 %f237, %f234, %f119;
mul.ftz.f32 %f238, %f236, %f120;
sub.ftz.f32 %f239, %f237, %f238;
add.ftz.f32 %f468, %f239, %f468;

{.reg .f16 low,high;
mov.b32 {low,high},%r378;
cvt.f32.f16 %f121, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r379;
cvt.f32.f16 %f122, low;}



dp4a.s32.s32 %r416, %r561, %r382, %r587;


dp4a.s32.s32 %r420, %r565, %r386, %r416;


dp4a.s32.s32 %r424, %r589, %r382, %r587;


dp4a.s32.s32 %r428, %r589, %r386, %r424;

mul.lo.s32 %r710, %r420, %r665;
cvt.rn.f32.s32 %f240, %r710;
fma.rn.ftz.f32 %f241, %f121, %f240, 0f00000000;
mul.lo.s32 %r711, %r428, %r668;
cvt.rn.f32.s32 %f242, %r711;
fma.rn.ftz.f32 %f243, %f121, %f242, 0f00000000;

dp4a.s32.s32 %r432, %r577, %r398, %r587;


dp4a.s32.s32 %r436, %r581, %r402, %r432;


dp4a.s32.s32 %r440, %r589, %r398, %r587;


dp4a.s32.s32 %r444, %r589, %r402, %r440;

mul.lo.s32 %r712, %r436, %r678;
cvt.rn.f32.s32 %f244, %r712;
fma.rn.ftz.f32 %f245, %f122, %f244, %f241;
mul.lo.s32 %r713, %r444, %r680;
cvt.rn.f32.s32 %f246, %r713;
fma.rn.ftz.f32 %f247, %f122, %f246, %f243;

{.reg .f16 low,high;
mov.b32 {low,high},%r160;
cvt.f32.f16 %f123, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r160;
cvt.f32.f16 %f124, high;}


mul.ftz.f32 %f248, %f245, %f123;
mul.ftz.f32 %f249, %f247, %f124;
sub.ftz.f32 %f250, %f248, %f249;
add.ftz.f32 %f467, %f250, %f467;
cvt.u64.u32 %rd71, %r991;
add.s64 %rd72, %rd2, %rd71;
mul.lo.s64 %rd73, %rd72, 36;
add.s64 %rd74, %rd12, %rd73;
ld.global.nc.u32 %r450, [%rd74];

{.reg .f16 low,high;
mov.b32 {low,high},%r450;
cvt.f32.f16 %f125, low;}


add.s64 %rd75, %rd74, %rd44;
ld.global.nc.u32 %r458, [%rd75+4];
ld.global.nc.u32 %r454, [%rd75+20];
ld.global.nc.u32 %r451, [%rd74+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r451;
cvt.f32.f16 %f126, low;}


ld.global.nc.u32 %r474, [%rd75+40];
ld.global.nc.u32 %r470, [%rd75+56];

dp4a.s32.s32 %r452, %r525, %r454, %r587;


dp4a.s32.s32 %r456, %r529, %r458, %r452;


dp4a.s32.s32 %r460, %r589, %r454, %r587;


dp4a.s32.s32 %r464, %r589, %r458, %r460;

mul.lo.s32 %r714, %r456, %r621;
cvt.rn.f32.s32 %f251, %r714;
fma.rn.ftz.f32 %f252, %f125, %f251, 0f00000000;
mul.lo.s32 %r715, %r464, %r624;
cvt.rn.f32.s32 %f253, %r715;
fma.rn.ftz.f32 %f254, %f125, %f253, 0f00000000;

dp4a.s32.s32 %r468, %r541, %r470, %r587;


dp4a.s32.s32 %r472, %r545, %r474, %r468;


dp4a.s32.s32 %r476, %r589, %r470, %r587;


dp4a.s32.s32 %r480, %r589, %r474, %r476;

mul.lo.s32 %r716, %r472, %r634;
cvt.rn.f32.s32 %f255, %r716;
fma.rn.ftz.f32 %f256, %f126, %f255, %f252;
mul.lo.s32 %r717, %r480, %r636;
cvt.rn.f32.s32 %f257, %r717;
fma.rn.ftz.f32 %f258, %f126, %f257, %f254;

{.reg .f16 low,high;
mov.b32 {low,high},%r124;
cvt.f32.f16 %f127, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r124;
cvt.f32.f16 %f128, high;}


mul.ftz.f32 %f259, %f256, %f127;
mul.ftz.f32 %f260, %f258, %f128;
sub.ftz.f32 %f261, %f259, %f260;
add.ftz.f32 %f466, %f261, %f466;

{.reg .f16 low,high;
mov.b32 {low,high},%r450;
cvt.f32.f16 %f129, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r451;
cvt.f32.f16 %f130, low;}



dp4a.s32.s32 %r488, %r561, %r454, %r587;


dp4a.s32.s32 %r492, %r565, %r458, %r488;


dp4a.s32.s32 %r496, %r589, %r454, %r587;


dp4a.s32.s32 %r500, %r589, %r458, %r496;

mul.lo.s32 %r718, %r492, %r665;
cvt.rn.f32.s32 %f262, %r718;
fma.rn.ftz.f32 %f263, %f129, %f262, 0f00000000;
mul.lo.s32 %r719, %r500, %r668;
cvt.rn.f32.s32 %f264, %r719;
fma.rn.ftz.f32 %f265, %f129, %f264, 0f00000000;

dp4a.s32.s32 %r504, %r577, %r470, %r587;


dp4a.s32.s32 %r508, %r581, %r474, %r504;


dp4a.s32.s32 %r512, %r589, %r470, %r587;


dp4a.s32.s32 %r516, %r589, %r474, %r512;

mul.lo.s32 %r720, %r508, %r678;
cvt.rn.f32.s32 %f266, %r720;
fma.rn.ftz.f32 %f267, %f130, %f266, %f263;
mul.lo.s32 %r721, %r516, %r680;
cvt.rn.f32.s32 %f268, %r721;
fma.rn.ftz.f32 %f269, %f130, %f268, %f265;

{.reg .f16 low,high;
mov.b32 {low,high},%r160;
cvt.f32.f16 %f131, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r160;
cvt.f32.f16 %f132, high;}


mul.ftz.f32 %f270, %f267, %f131;
mul.ftz.f32 %f271, %f269, %f132;
sub.ftz.f32 %f272, %f270, %f271;
add.ftz.f32 %f465, %f272, %f465;
cvt.u64.u32 %rd76, %r990;
add.s64 %rd77, %rd2, %rd76;
mul.lo.s64 %rd78, %rd77, 36;
add.s64 %rd79, %rd12, %rd78;
ld.global.nc.u32 %r522, [%rd79];

{.reg .f16 low,high;
mov.b32 {low,high},%r522;
cvt.f32.f16 %f133, low;}


add.s64 %rd80, %rd79, %rd44;
ld.global.nc.u32 %r530, [%rd80+4];
ld.global.nc.u32 %r526, [%rd80+20];
ld.global.nc.u32 %r523, [%rd79+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r523;
cvt.f32.f16 %f134, low;}


ld.global.nc.u32 %r546, [%rd80+40];
ld.global.nc.u32 %r542, [%rd80+56];

dp4a.s32.s32 %r524, %r525, %r526, %r587;


dp4a.s32.s32 %r528, %r529, %r530, %r524;


dp4a.s32.s32 %r532, %r589, %r526, %r587;


dp4a.s32.s32 %r536, %r589, %r530, %r532;

mul.lo.s32 %r722, %r528, %r621;
cvt.rn.f32.s32 %f273, %r722;
fma.rn.ftz.f32 %f274, %f133, %f273, 0f00000000;
mul.lo.s32 %r723, %r536, %r624;
cvt.rn.f32.s32 %f275, %r723;
fma.rn.ftz.f32 %f276, %f133, %f275, 0f00000000;

dp4a.s32.s32 %r540, %r541, %r542, %r587;


dp4a.s32.s32 %r544, %r545, %r546, %r540;


dp4a.s32.s32 %r548, %r589, %r542, %r587;


dp4a.s32.s32 %r552, %r589, %r546, %r548;

mul.lo.s32 %r724, %r544, %r634;
cvt.rn.f32.s32 %f277, %r724;
fma.rn.ftz.f32 %f278, %f134, %f277, %f274;
mul.lo.s32 %r725, %r552, %r636;
cvt.rn.f32.s32 %f279, %r725;
fma.rn.ftz.f32 %f280, %f134, %f279, %f276;

{.reg .f16 low,high;
mov.b32 {low,high},%r124;
cvt.f32.f16 %f135, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r124;
cvt.f32.f16 %f136, high;}


mul.ftz.f32 %f281, %f278, %f135;
mul.ftz.f32 %f282, %f280, %f136;
sub.ftz.f32 %f283, %f281, %f282;
add.ftz.f32 %f464, %f283, %f464;

{.reg .f16 low,high;
mov.b32 {low,high},%r522;
cvt.f32.f16 %f137, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r523;
cvt.f32.f16 %f138, low;}



dp4a.s32.s32 %r560, %r561, %r526, %r587;


dp4a.s32.s32 %r564, %r565, %r530, %r560;


dp4a.s32.s32 %r568, %r589, %r526, %r587;


dp4a.s32.s32 %r572, %r589, %r530, %r568;

mul.lo.s32 %r726, %r564, %r665;
cvt.rn.f32.s32 %f284, %r726;
fma.rn.ftz.f32 %f285, %f137, %f284, 0f00000000;
mul.lo.s32 %r727, %r572, %r668;
cvt.rn.f32.s32 %f286, %r727;
fma.rn.ftz.f32 %f287, %f137, %f286, 0f00000000;

dp4a.s32.s32 %r576, %r577, %r542, %r587;


dp4a.s32.s32 %r580, %r581, %r546, %r576;


dp4a.s32.s32 %r584, %r589, %r542, %r587;


dp4a.s32.s32 %r588, %r589, %r546, %r584;

mul.lo.s32 %r728, %r580, %r678;
cvt.rn.f32.s32 %f288, %r728;
fma.rn.ftz.f32 %f289, %f138, %f288, %f285;
mul.lo.s32 %r729, %r588, %r680;
cvt.rn.f32.s32 %f290, %r729;
fma.rn.ftz.f32 %f291, %f138, %f290, %f287;

{.reg .f16 low,high;
mov.b32 {low,high},%r160;
cvt.f32.f16 %f139, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r160;
cvt.f32.f16 %f140, high;}


mul.ftz.f32 %f292, %f289, %f139;
mul.ftz.f32 %f293, %f291, %f140;
sub.ftz.f32 %f294, %f292, %f293;
add.ftz.f32 %f463, %f294, %f463;
add.s32 %r994, %r994, 32;
add.s32 %r993, %r993, 32;
add.s32 %r992, %r992, 32;
add.s32 %r991, %r991, 32;
add.s32 %r990, %r990, 32;
add.s32 %r989, %r989, 32;
add.s32 %r988, %r988, 32;
add.s64 %rd125, %rd125, 704;
add.s32 %r995, %r995, 4;
setp.lt.s32 %p5, %r995, %r6;
@%p5 bra $L__BB181_2;

$L__BB181_3:
setp.eq.s32 %p6, %r1, 0;
@%p6 bra $L__BB181_5;

mov.u32 %r730, _ZZ13mul_mat_vec_qIL9ggml_type13ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r731, %r1, 1792, %r730;
shl.b32 %r732, %r2, 2;
add.s32 %r733, %r731, %r732;
st.shared.f32 [%r733+-1792], %f476;
st.shared.f32 [%r733+-1664], %f475;
st.shared.f32 [%r733+-1536], %f474;
st.shared.f32 [%r733+-1408], %f473;
st.shared.f32 [%r733+-1280], %f472;
st.shared.f32 [%r733+-1152], %f471;
st.shared.f32 [%r733+-1024], %f470;
st.shared.f32 [%r733+-896], %f469;
st.shared.f32 [%r733+-768], %f468;
st.shared.f32 [%r733+-640], %f467;
st.shared.f32 [%r733+-512], %f466;
st.shared.f32 [%r733+-384], %f465;
st.shared.f32 [%r733+-256], %f464;
st.shared.f32 [%r733+-128], %f463;

$L__BB181_5:
setp.ne.s32 %p7, %r1, 0;
bar.sync 0;
@%p7 bra $L__BB181_34;

mov.u32 %r734, %ctaid.z;
mov.u32 %r735, %ctaid.y;
mad.lo.s32 %r736, %r735, %r43, %r4;
mad.lo.s32 %r737, %r734, %r49, %r736;
cvt.u64.u32 %rd15, %r737;
shl.b32 %r738, %r2, 2;
mov.u32 %r739, _ZZ13mul_mat_vec_qIL9ggml_type13ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r32, %r739, %r738;
mov.u32 %r740, 2;
ld.shared.f32 %f295, [%r32];
add.ftz.f32 %f296, %f295, %f476;
mov.b32 %r741, %f296;
mov.u32 %r742, 31;
mov.u32 %r743, 16;
mov.u32 %r744, -1;
shfl.sync.bfly.b32 %r745|%p8, %r741, %r743, %r742, %r744;
mov.b32 %f297, %r745;
add.ftz.f32 %f298, %f296, %f297;
mov.b32 %r746, %f298;
mov.u32 %r747, 8;
shfl.sync.bfly.b32 %r748|%p9, %r746, %r747, %r742, %r744;
mov.b32 %f299, %r748;
add.ftz.f32 %f300, %f298, %f299;
mov.b32 %r749, %f300;
mov.u32 %r750, 4;
shfl.sync.bfly.b32 %r751|%p10, %r749, %r750, %r742, %r744;
mov.b32 %f301, %r751;
add.ftz.f32 %f302, %f300, %f301;
mov.b32 %r752, %f302;
shfl.sync.bfly.b32 %r753|%p11, %r752, %r740, %r742, %r744;
mov.b32 %f303, %r753;
add.ftz.f32 %f304, %f302, %f303;
mov.b32 %r754, %f304;
mov.u32 %r755, 1;
shfl.sync.bfly.b32 %r756|%p12, %r754, %r755, %r742, %r744;
mov.b32 %f305, %r756;
add.ftz.f32 %f43, %f304, %f305;
setp.lt.u32 %p13, %r4, %r37;
setp.eq.s32 %p14, %r2, 0;
and.pred %p1, %p14, %p13;
cvta.to.global.u64 %rd81, %rd22;
mul.wide.u32 %rd82, %r737, 4;
add.s64 %rd16, %rd81, %rd82;
not.pred %p15, %p1;
@%p15 bra $L__BB181_8;

st.global.f32 [%rd16], %f43;

$L__BB181_8:
ld.shared.f32 %f306, [%r32+128];
add.ftz.f32 %f307, %f306, %f475;
mov.b32 %r757, %f307;
shfl.sync.bfly.b32 %r761|%p16, %r757, %r743, %r742, %r744;
mov.b32 %f308, %r761;
add.ftz.f32 %f309, %f307, %f308;
mov.b32 %r762, %f309;
shfl.sync.bfly.b32 %r764|%p17, %r762, %r747, %r742, %r744;
mov.b32 %f310, %r764;
add.ftz.f32 %f311, %f309, %f310;
mov.b32 %r765, %f311;
shfl.sync.bfly.b32 %r767|%p18, %r765, %r750, %r742, %r744;
mov.b32 %f312, %r767;
add.ftz.f32 %f313, %f311, %f312;
mov.b32 %r768, %f313;
shfl.sync.bfly.b32 %r770|%p19, %r768, %r740, %r742, %r744;
mov.b32 %f314, %r770;
add.ftz.f32 %f315, %f313, %f314;
mov.b32 %r771, %f315;
shfl.sync.bfly.b32 %r773|%p20, %r771, %r755, %r742, %r744;
mov.b32 %f316, %r773;
add.ftz.f32 %f44, %f315, %f316;
setp.eq.s32 %p21, %r2, 1;
or.b32 %r777, %r4, 1;
setp.lt.u32 %p22, %r777, %r37;
and.pred %p2, %p21, %p22;
not.pred %p23, %p2;
@%p23 bra $L__BB181_10;

st.global.f32 [%rd16+4], %f44;

$L__BB181_10:
ld.shared.f32 %f317, [%r32+256];
add.ftz.f32 %f318, %f317, %f474;
mov.b32 %r778, %f318;
mov.u32 %r779, 31;
mov.u32 %r780, 16;
mov.u32 %r781, -1;
shfl.sync.bfly.b32 %r782|%p24, %r778, %r780, %r779, %r781;
mov.b32 %f319, %r782;
add.ftz.f32 %f320, %f318, %f319;
mov.b32 %r783, %f320;
mov.u32 %r784, 8;
shfl.sync.bfly.b32 %r785|%p25, %r783, %r784, %r779, %r781;
mov.b32 %f321, %r785;
add.ftz.f32 %f322, %f320, %f321;
mov.b32 %r786, %f322;
mov.u32 %r787, 4;
shfl.sync.bfly.b32 %r788|%p26, %r786, %r787, %r779, %r781;
mov.b32 %f323, %r788;
add.ftz.f32 %f324, %f322, %f323;
mov.b32 %r789, %f324;
mov.u32 %r790, 2;
shfl.sync.bfly.b32 %r791|%p27, %r789, %r790, %r779, %r781;
mov.b32 %f325, %r791;
add.ftz.f32 %f326, %f324, %f325;
mov.b32 %r792, %f326;
mov.u32 %r793, 1;
shfl.sync.bfly.b32 %r794|%p28, %r792, %r793, %r779, %r781;
mov.b32 %f327, %r794;
add.ftz.f32 %f45, %f326, %f327;
@%p15 bra $L__BB181_12;

cvt.u64.u32 %rd83, %r37;
add.s64 %rd84, %rd83, %rd15;
shl.b64 %rd86, %rd84, 2;
add.s64 %rd87, %rd81, %rd86;
st.global.f32 [%rd87], %f45;

$L__BB181_12:
ld.shared.f32 %f328, [%r32+384];
add.ftz.f32 %f329, %f328, %f473;
mov.b32 %r795, %f329;
shfl.sync.bfly.b32 %r799|%p30, %r795, %r780, %r779, %r781;
mov.b32 %f330, %r799;
add.ftz.f32 %f331, %f329, %f330;
mov.b32 %r800, %f331;
shfl.sync.bfly.b32 %r802|%p31, %r800, %r784, %r779, %r781;
mov.b32 %f332, %r802;
add.ftz.f32 %f333, %f331, %f332;
mov.b32 %r803, %f333;
shfl.sync.bfly.b32 %r805|%p32, %r803, %r787, %r779, %r781;
mov.b32 %f334, %r805;
add.ftz.f32 %f335, %f333, %f334;
mov.b32 %r806, %f335;
shfl.sync.bfly.b32 %r808|%p33, %r806, %r790, %r779, %r781;
mov.b32 %f336, %r808;
add.ftz.f32 %f337, %f335, %f336;
mov.b32 %r809, %f337;
shfl.sync.bfly.b32 %r811|%p34, %r809, %r793, %r779, %r781;
mov.b32 %f338, %r811;
add.ftz.f32 %f46, %f337, %f338;
@%p23 bra $L__BB181_14;

add.s32 %r812, %r37, 1;
cvt.u64.u32 %rd88, %r812;
add.s64 %rd89, %rd88, %rd15;
shl.b64 %rd91, %rd89, 2;
add.s64 %rd92, %rd81, %rd91;
st.global.f32 [%rd92], %f46;

$L__BB181_14:
shl.b32 %r813, %r37, 1;
mov.u32 %r814, 1;
ld.shared.f32 %f339, [%r32+512];
add.ftz.f32 %f340, %f339, %f472;
mov.b32 %r815, %f340;
mov.u32 %r816, 31;
mov.u32 %r817, 16;
mov.u32 %r818, -1;
shfl.sync.bfly.b32 %r819|%p36, %r815, %r817, %r816, %r818;
mov.b32 %f341, %r819;
add.ftz.f32 %f342, %f340, %f341;
mov.b32 %r820, %f342;
mov.u32 %r821, 8;
shfl.sync.bfly.b32 %r822|%p37, %r820, %r821, %r816, %r818;
mov.b32 %f343, %r822;
add.ftz.f32 %f344, %f342, %f343;
mov.b32 %r823, %f344;
mov.u32 %r824, 4;
shfl.sync.bfly.b32 %r825|%p38, %r823, %r824, %r816, %r818;
mov.b32 %f345, %r825;
add.ftz.f32 %f346, %f344, %f345;
mov.b32 %r826, %f346;
mov.u32 %r827, 2;
shfl.sync.bfly.b32 %r828|%p39, %r826, %r827, %r816, %r818;
mov.b32 %f347, %r828;
add.ftz.f32 %f348, %f346, %f347;
mov.b32 %r829, %f348;
shfl.sync.bfly.b32 %r830|%p40, %r829, %r814, %r816, %r818;
mov.b32 %f349, %r830;
add.ftz.f32 %f47, %f348, %f349;
cvt.u64.u32 %rd93, %r813;
add.s64 %rd94, %rd93, %rd15;
shl.b64 %rd96, %rd94, 2;
add.s64 %rd17, %rd81, %rd96;
@%p15 bra $L__BB181_16;

st.global.f32 [%rd17], %f47;

$L__BB181_16:
ld.shared.f32 %f350, [%r32+640];
add.ftz.f32 %f351, %f350, %f471;
mov.b32 %r831, %f351;
shfl.sync.bfly.b32 %r835|%p42, %r831, %r817, %r816, %r818;
mov.b32 %f352, %r835;
add.ftz.f32 %f353, %f351, %f352;
mov.b32 %r836, %f353;
shfl.sync.bfly.b32 %r838|%p43, %r836, %r821, %r816, %r818;
mov.b32 %f354, %r838;
add.ftz.f32 %f355, %f353, %f354;
mov.b32 %r839, %f355;
shfl.sync.bfly.b32 %r841|%p44, %r839, %r824, %r816, %r818;
mov.b32 %f356, %r841;
add.ftz.f32 %f357, %f355, %f356;
mov.b32 %r842, %f357;
shfl.sync.bfly.b32 %r844|%p45, %r842, %r827, %r816, %r818;
mov.b32 %f358, %r844;
add.ftz.f32 %f359, %f357, %f358;
mov.b32 %r845, %f359;
shfl.sync.bfly.b32 %r847|%p46, %r845, %r814, %r816, %r818;
mov.b32 %f360, %r847;
add.ftz.f32 %f48, %f359, %f360;
@%p23 bra $L__BB181_18;

st.global.f32 [%rd17+4], %f48;

$L__BB181_18:
mul.lo.s32 %r33, %r37, 3;
ld.shared.f32 %f361, [%r32+768];
add.ftz.f32 %f362, %f361, %f470;
mov.b32 %r848, %f362;
mov.u32 %r849, 31;
mov.u32 %r850, 16;
mov.u32 %r851, -1;
shfl.sync.bfly.b32 %r852|%p48, %r848, %r850, %r849, %r851;
mov.b32 %f363, %r852;
add.ftz.f32 %f364, %f362, %f363;
mov.b32 %r853, %f364;
mov.u32 %r854, 8;
shfl.sync.bfly.b32 %r855|%p49, %r853, %r854, %r849, %r851;
mov.b32 %f365, %r855;
add.ftz.f32 %f366, %f364, %f365;
mov.b32 %r856, %f366;
mov.u32 %r857, 4;
shfl.sync.bfly.b32 %r858|%p50, %r856, %r857, %r849, %r851;
mov.b32 %f367, %r858;
add.ftz.f32 %f368, %f366, %f367;
mov.b32 %r859, %f368;
mov.u32 %r860, 2;
shfl.sync.bfly.b32 %r861|%p51, %r859, %r860, %r849, %r851;
mov.b32 %f369, %r861;
add.ftz.f32 %f370, %f368, %f369;
mov.b32 %r862, %f370;
mov.u32 %r863, 1;
shfl.sync.bfly.b32 %r864|%p52, %r862, %r863, %r849, %r851;
mov.b32 %f371, %r864;
add.ftz.f32 %f49, %f370, %f371;
@%p15 bra $L__BB181_20;

cvt.u64.u32 %rd97, %r33;
add.s64 %rd98, %rd97, %rd15;
shl.b64 %rd100, %rd98, 2;
add.s64 %rd101, %rd81, %rd100;
st.global.f32 [%rd101], %f49;

$L__BB181_20:
ld.shared.f32 %f372, [%r32+896];
add.ftz.f32 %f373, %f372, %f469;
mov.b32 %r865, %f373;
shfl.sync.bfly.b32 %r869|%p54, %r865, %r850, %r849, %r851;
mov.b32 %f374, %r869;
add.ftz.f32 %f375, %f373, %f374;
mov.b32 %r870, %f375;
shfl.sync.bfly.b32 %r872|%p55, %r870, %r854, %r849, %r851;
mov.b32 %f376, %r872;
add.ftz.f32 %f377, %f375, %f376;
mov.b32 %r873, %f377;
shfl.sync.bfly.b32 %r875|%p56, %r873, %r857, %r849, %r851;
mov.b32 %f378, %r875;
add.ftz.f32 %f379, %f377, %f378;
mov.b32 %r876, %f379;
shfl.sync.bfly.b32 %r878|%p57, %r876, %r860, %r849, %r851;
mov.b32 %f380, %r878;
add.ftz.f32 %f381, %f379, %f380;
mov.b32 %r879, %f381;
shfl.sync.bfly.b32 %r881|%p58, %r879, %r863, %r849, %r851;
mov.b32 %f382, %r881;
add.ftz.f32 %f50, %f381, %f382;
@%p23 bra $L__BB181_22;

add.s32 %r882, %r33, 1;
cvt.u64.u32 %rd102, %r882;
add.s64 %rd103, %rd102, %rd15;
shl.b64 %rd105, %rd103, 2;
add.s64 %rd106, %rd81, %rd105;
st.global.f32 [%rd106], %f50;

$L__BB181_22:
shl.b32 %r883, %r37, 2;
mov.u32 %r884, 2;
ld.shared.f32 %f383, [%r32+1024];
add.ftz.f32 %f384, %f383, %f468;
mov.b32 %r885, %f384;
mov.u32 %r886, 31;
mov.u32 %r887, 16;
mov.u32 %r888, -1;
shfl.sync.bfly.b32 %r889|%p60, %r885, %r887, %r886, %r888;
mov.b32 %f385, %r889;
add.ftz.f32 %f386, %f384, %f385;
mov.b32 %r890, %f386;
mov.u32 %r891, 8;
shfl.sync.bfly.b32 %r892|%p61, %r890, %r891, %r886, %r888;
mov.b32 %f387, %r892;
add.ftz.f32 %f388, %f386, %f387;
mov.b32 %r893, %f388;
mov.u32 %r894, 4;
shfl.sync.bfly.b32 %r895|%p62, %r893, %r894, %r886, %r888;
mov.b32 %f389, %r895;
add.ftz.f32 %f390, %f388, %f389;
mov.b32 %r896, %f390;
shfl.sync.bfly.b32 %r897|%p63, %r896, %r884, %r886, %r888;
mov.b32 %f391, %r897;
add.ftz.f32 %f392, %f390, %f391;
mov.b32 %r898, %f392;
mov.u32 %r899, 1;
shfl.sync.bfly.b32 %r900|%p64, %r898, %r899, %r886, %r888;
mov.b32 %f393, %r900;
add.ftz.f32 %f51, %f392, %f393;
cvt.u64.u32 %rd107, %r883;
add.s64 %rd108, %rd107, %rd15;
shl.b64 %rd110, %rd108, 2;
add.s64 %rd18, %rd81, %rd110;
@%p15 bra $L__BB181_24;

st.global.f32 [%rd18], %f51;

$L__BB181_24:
ld.shared.f32 %f394, [%r32+1152];
add.ftz.f32 %f395, %f394, %f467;
mov.b32 %r901, %f395;
shfl.sync.bfly.b32 %r905|%p66, %r901, %r887, %r886, %r888;
mov.b32 %f396, %r905;
add.ftz.f32 %f397, %f395, %f396;
mov.b32 %r906, %f397;
shfl.sync.bfly.b32 %r908|%p67, %r906, %r891, %r886, %r888;
mov.b32 %f398, %r908;
add.ftz.f32 %f399, %f397, %f398;
mov.b32 %r909, %f399;
shfl.sync.bfly.b32 %r911|%p68, %r909, %r894, %r886, %r888;
mov.b32 %f400, %r911;
add.ftz.f32 %f401, %f399, %f400;
mov.b32 %r912, %f401;
shfl.sync.bfly.b32 %r914|%p69, %r912, %r884, %r886, %r888;
mov.b32 %f402, %r914;
add.ftz.f32 %f403, %f401, %f402;
mov.b32 %r915, %f403;
shfl.sync.bfly.b32 %r917|%p70, %r915, %r899, %r886, %r888;
mov.b32 %f404, %r917;
add.ftz.f32 %f52, %f403, %f404;
@%p23 bra $L__BB181_26;

st.global.f32 [%rd18+4], %f52;

$L__BB181_26:
mul.lo.s32 %r34, %r37, 5;
ld.shared.f32 %f405, [%r32+1280];
add.ftz.f32 %f406, %f405, %f466;
mov.b32 %r918, %f406;
mov.u32 %r919, 31;
mov.u32 %r920, 16;
mov.u32 %r921, -1;
shfl.sync.bfly.b32 %r922|%p72, %r918, %r920, %r919, %r921;
mov.b32 %f407, %r922;
add.ftz.f32 %f408, %f406, %f407;
mov.b32 %r923, %f408;
mov.u32 %r924, 8;
shfl.sync.bfly.b32 %r925|%p73, %r923, %r924, %r919, %r921;
mov.b32 %f409, %r925;
add.ftz.f32 %f410, %f408, %f409;
mov.b32 %r926, %f410;
mov.u32 %r927, 4;
shfl.sync.bfly.b32 %r928|%p74, %r926, %r927, %r919, %r921;
mov.b32 %f411, %r928;
add.ftz.f32 %f412, %f410, %f411;
mov.b32 %r929, %f412;
mov.u32 %r930, 2;
shfl.sync.bfly.b32 %r931|%p75, %r929, %r930, %r919, %r921;
mov.b32 %f413, %r931;
add.ftz.f32 %f414, %f412, %f413;
mov.b32 %r932, %f414;
mov.u32 %r933, 1;
shfl.sync.bfly.b32 %r934|%p76, %r932, %r933, %r919, %r921;
mov.b32 %f415, %r934;
add.ftz.f32 %f53, %f414, %f415;
@%p15 bra $L__BB181_28;

cvt.u64.u32 %rd111, %r34;
add.s64 %rd112, %rd111, %rd15;
shl.b64 %rd114, %rd112, 2;
add.s64 %rd115, %rd81, %rd114;
st.global.f32 [%rd115], %f53;

$L__BB181_28:
ld.shared.f32 %f416, [%r32+1408];
add.ftz.f32 %f417, %f416, %f465;
mov.b32 %r935, %f417;
shfl.sync.bfly.b32 %r939|%p78, %r935, %r920, %r919, %r921;
mov.b32 %f418, %r939;
add.ftz.f32 %f419, %f417, %f418;
mov.b32 %r940, %f419;
shfl.sync.bfly.b32 %r942|%p79, %r940, %r924, %r919, %r921;
mov.b32 %f420, %r942;
add.ftz.f32 %f421, %f419, %f420;
mov.b32 %r943, %f421;
shfl.sync.bfly.b32 %r945|%p80, %r943, %r927, %r919, %r921;
mov.b32 %f422, %r945;
add.ftz.f32 %f423, %f421, %f422;
mov.b32 %r946, %f423;
shfl.sync.bfly.b32 %r948|%p81, %r946, %r930, %r919, %r921;
mov.b32 %f424, %r948;
add.ftz.f32 %f425, %f423, %f424;
mov.b32 %r949, %f425;
shfl.sync.bfly.b32 %r951|%p82, %r949, %r933, %r919, %r921;
mov.b32 %f426, %r951;
add.ftz.f32 %f54, %f425, %f426;
@%p23 bra $L__BB181_30;

add.s32 %r952, %r34, 1;
cvt.u64.u32 %rd116, %r952;
add.s64 %rd117, %rd116, %rd15;
shl.b64 %rd119, %rd117, 2;
add.s64 %rd120, %rd81, %rd119;
st.global.f32 [%rd120], %f54;

$L__BB181_30:
mul.lo.s32 %r953, %r37, 6;
ld.shared.f32 %f427, [%r32+1536];
add.ftz.f32 %f428, %f427, %f464;
mov.b32 %r954, %f428;
mov.u32 %r955, 31;
mov.u32 %r956, 16;
mov.u32 %r957, -1;
shfl.sync.bfly.b32 %r958|%p84, %r954, %r956, %r955, %r957;
mov.b32 %f429, %r958;
add.ftz.f32 %f430, %f428, %f429;
mov.b32 %r959, %f430;
mov.u32 %r960, 8;
shfl.sync.bfly.b32 %r961|%p85, %r959, %r960, %r955, %r957;
mov.b32 %f431, %r961;
add.ftz.f32 %f432, %f430, %f431;
mov.b32 %r962, %f432;
mov.u32 %r963, 4;
shfl.sync.bfly.b32 %r964|%p86, %r962, %r963, %r955, %r957;
mov.b32 %f433, %r964;
add.ftz.f32 %f434, %f432, %f433;
mov.b32 %r965, %f434;
mov.u32 %r966, 2;
shfl.sync.bfly.b32 %r967|%p87, %r965, %r966, %r955, %r957;
mov.b32 %f435, %r967;
add.ftz.f32 %f436, %f434, %f435;
mov.b32 %r968, %f436;
mov.u32 %r969, 1;
shfl.sync.bfly.b32 %r970|%p88, %r968, %r969, %r955, %r957;
mov.b32 %f437, %r970;
add.ftz.f32 %f55, %f436, %f437;
cvt.u64.u32 %rd121, %r953;
add.s64 %rd122, %rd121, %rd15;
shl.b64 %rd124, %rd122, 2;
add.s64 %rd19, %rd81, %rd124;
@%p15 bra $L__BB181_32;

st.global.f32 [%rd19], %f55;

$L__BB181_32:
ld.shared.f32 %f438, [%r32+1664];
add.ftz.f32 %f439, %f438, %f463;
mov.b32 %r971, %f439;
shfl.sync.bfly.b32 %r975|%p90, %r971, %r956, %r955, %r957;
mov.b32 %f440, %r975;
add.ftz.f32 %f441, %f439, %f440;
mov.b32 %r976, %f441;
shfl.sync.bfly.b32 %r978|%p91, %r976, %r960, %r955, %r957;
mov.b32 %f442, %r978;
add.ftz.f32 %f443, %f441, %f442;
mov.b32 %r979, %f443;
shfl.sync.bfly.b32 %r981|%p92, %r979, %r963, %r955, %r957;
mov.b32 %f444, %r981;
add.ftz.f32 %f445, %f443, %f444;
mov.b32 %r982, %f445;
shfl.sync.bfly.b32 %r984|%p93, %r982, %r966, %r955, %r957;
mov.b32 %f446, %r984;
add.ftz.f32 %f447, %f445, %f446;
mov.b32 %r985, %f447;
shfl.sync.bfly.b32 %r987|%p94, %r985, %r969, %r955, %r957;
mov.b32 %f448, %r987;
add.ftz.f32 %f56, %f447, %f448;
@%p23 bra $L__BB181_34;

st.global.f32 [%rd19+4], %f56;

$L__BB181_34:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type13ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type13ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type13ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type13ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type13ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type13ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type13ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type13ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type13ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type13ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<108>;
.reg .b16 %rs<15>;
.reg .f32 %f<545>;
.reg .b32 %r<1117>;
.reg .b64 %rd<141>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type13ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[2048];

ld.param.u64 %rd20, [_Z13mul_mat_vec_qIL9ggml_type13ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd21, [_Z13mul_mat_vec_qIL9ggml_type13ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd22, [_Z13mul_mat_vec_qIL9ggml_type13ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r54, [_Z13mul_mat_vec_qIL9ggml_type13ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type13ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type13ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type13ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r45, [_Z13mul_mat_vec_qIL9ggml_type13ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r46, [_Z13mul_mat_vec_qIL9ggml_type13ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r47, [_Z13mul_mat_vec_qIL9ggml_type13ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r51, [_Z13mul_mat_vec_qIL9ggml_type13ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r52, [_Z13mul_mat_vec_qIL9ggml_type13ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r53, [_Z13mul_mat_vec_qIL9ggml_type13ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r49, [_Z13mul_mat_vec_qIL9ggml_type13ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r48, [_Z13mul_mat_vec_qIL9ggml_type13ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type13ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type13ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r55, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r55, %r2;
mov.u32 %r56, %ctaid.x;
shl.b32 %r4, %r56, 1;
shr.s32 %r57, %r3, 31;
shr.u32 %r58, %r57, 28;
add.s32 %r59, %r3, %r58;
shr.s32 %r1116, %r59, 4;
shr.u32 %r6, %r54, 8;
setp.ge.s32 %p3, %r1116, %r6;
mov.f32 %f529, 0f00000000;
mov.f32 %f530, %f529;
mov.f32 %f531, %f529;
mov.f32 %f532, %f529;
mov.f32 %f533, %f529;
mov.f32 %f534, %f529;
mov.f32 %f535, %f529;
mov.f32 %f536, %f529;
mov.f32 %f537, %f529;
mov.f32 %f538, %f529;
mov.f32 %f539, %f529;
mov.f32 %f540, %f529;
mov.f32 %f541, %f529;
mov.f32 %f542, %f529;
mov.f32 %f543, %f529;
mov.f32 %f544, %f529;
@%p3 bra $L__BB182_3;

mov.u32 %r60, %ctaid.y;
mov.u32 %r61, %ctaid.z;
mul.hi.u32 %r62, %r61, %r48;
add.s32 %r63, %r62, %r61;
shr.u32 %r64, %r63, %r49;
mul.lo.s32 %r65, %r64, %r51;
mul.hi.u32 %r66, %r60, %r42;
add.s32 %r67, %r66, %r60;
shr.u32 %r68, %r67, %r43;
mul.lo.s32 %r69, %r68, %r45;
mul.lo.s32 %r70, %r60, %r46;
cvt.u64.u32 %rd23, %r70;
cvta.to.global.u64 %rd140, %rd20;
and.b32 %r74, %r59, -16;
sub.s32 %r7, %r3, %r74;
shr.s32 %r75, %r7, 31;
shr.u32 %r76, %r75, 30;
add.s32 %r77, %r7, %r76;
shr.s32 %r78, %r77, 2;
shl.b32 %r8, %r78, 1;
shl.b32 %r79, %r78, 5;
and.b32 %r80, %r77, -4;
sub.s32 %r81, %r7, %r80;
shl.b32 %r82, %r81, 2;
add.s32 %r83, %r79, %r82;
cvt.s64.s32 %rd24, %r83;
cvt.s64.s32 %rd25, %r82;
cvt.s64.s32 %rd26, %r8;
add.s64 %rd27, %rd23, %rd26;
mul.lo.s32 %r84, %r61, %r52;
cvt.u64.u32 %rd28, %r84;
add.s64 %rd2, %rd27, %rd28;
cvt.s64.s32 %rd3, %r81;
shl.b32 %r85, %r40, 1;
shl.b32 %r1108, %r1116, 3;
add.s32 %r1115, %r1108, %r85;
mad.lo.s32 %r1114, %r40, 3, %r1108;
shl.b32 %r86, %r40, 2;
add.s32 %r1113, %r1108, %r86;
mad.lo.s32 %r1112, %r40, 5, %r1108;
mad.lo.s32 %r1111, %r40, 6, %r1108;
mad.lo.s32 %r1110, %r40, 7, %r1108;
add.s32 %r1109, %r40, %r1108;
add.s32 %r87, %r1116, %r65;
add.s32 %r88, %r87, %r69;
add.s32 %r89, %r4, 1;
mad.lo.s32 %r90, %r39, %r89, %r88;
mul.wide.s32 %rd4, %r90, 176;
mad.lo.s32 %r91, %r4, %r39, %r1116;
add.s32 %r92, %r91, %r65;
add.s32 %r93, %r92, %r69;
mul.wide.s32 %rd5, %r93, 176;
add.s64 %rd29, %rd5, %rd24;
add.s64 %rd6, %rd29, 64;
cvt.u64.u32 %rd30, %r8;
and.b64 %rd31, %rd30, 2;
or.b64 %rd32, %rd31, %rd4;
or.b64 %rd7, %rd32, 8;
add.s64 %rd33, %rd5, %rd25;
add.s64 %rd8, %rd33, 32;
or.b64 %rd34, %rd31, %rd5;
or.b64 %rd9, %rd34, 12;
add.s64 %rd35, %rd4, %rd25;
add.s64 %rd10, %rd35, 16;
add.s64 %rd36, %rd4, %rd24;
add.s64 %rd11, %rd36, 48;
cvta.to.global.u64 %rd12, %rd21;
mov.f32 %f529, 0f00000000;
shl.b64 %rd44, %rd3, 2;

$L__BB182_2:
cvt.u64.u32 %rd37, %r1108;
add.s64 %rd38, %rd2, %rd37;
add.s64 %rd39, %rd140, %rd6;
ld.global.nc.u32 %r670, [%rd39+-16];
ld.global.nc.u32 %r671, [%rd39];
add.s64 %rd40, %rd140, %rd8;
ld.global.nc.u32 %r672, [%rd40+-16];
shr.s32 %r673, %r672, %r8;
ld.global.nc.u32 %r674, [%rd40];
shr.s32 %r675, %r674, %r8;
add.s64 %rd41, %rd140, %rd9;
ld.global.nc.u16 %rs1, [%rd41+-8];
cvt.u32.u16 %r676, %rs1;
ld.global.nc.u16 %rs2, [%rd41+-4];
cvt.u32.u16 %r677, %rs2;
ld.global.nc.u16 %rs3, [%rd41];
cvt.u32.u16 %r678, %rs3;
and.b32 %r679, %r676, 16191;
and.b32 %r680, %r678, 3855;
and.b32 %r681, %r676, 49344;
shr.u32 %r682, %r681, 2;
or.b32 %r683, %r680, %r682;
setp.gt.s32 %p4, %r7, 7;
and.b32 %r684, %r677, 16191;
shr.u32 %r685, %r678, 4;
and.b32 %r686, %r685, 3855;
and.b32 %r687, %r677, 49344;
shr.u32 %r688, %r687, 2;
or.b32 %r689, %r686, %r688;
mul.lo.s64 %rd42, %rd38, 36;
add.s64 %rd43, %rd12, %rd42;
ld.global.nc.u32 %r94, [%rd43];

{.reg .f16 low,high;
mov.b32 {low,high},%r94;
cvt.f32.f16 %f97, low;}


add.s64 %rd45, %rd43, %rd44;
ld.global.nc.u32 %r102, [%rd45+4];
ld.global.nc.u32 %r98, [%rd45+20];
ld.global.nc.u32 %r95, [%rd43+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r95;
cvt.f32.f16 %f98, low;}


ld.global.nc.u32 %r118, [%rd45+40];
ld.global.nc.u32 %r114, [%rd45+56];
add.s64 %rd46, %rd140, %rd5;
and.b32 %r690, %r670, 252645135;
and.b32 %r691, %r671, 252645135;
shl.b32 %r692, %r673, 4;
and.b32 %r693, %r692, 269488144;
shl.b32 %r694, %r675, 4;
and.b32 %r695, %r694, 269488144;
or.b32 %r605, %r690, %r693;
or.b32 %r601, %r691, %r695;
mov.u32 %r663, 0;

dp4a.s32.s32 %r96, %r601, %r98, %r663;


dp4a.s32.s32 %r100, %r605, %r102, %r96;

mov.u32 %r665, 16843009;

dp4a.s32.s32 %r104, %r665, %r98, %r663;


dp4a.s32.s32 %r108, %r665, %r102, %r104;

selp.b32 %r696, %r683, %r679, %p4;
cvt.u16.u32 %rs4, %r696;
and.b32 %r697, %r696, 63;
mul.lo.s32 %r698, %r100, %r697;
cvt.rn.f32.s32 %f161, %r698;
fma.rn.ftz.f32 %f162, %f97, %f161, 0f00000000;
selp.b32 %r699, %r689, %r684, %p4;
cvt.u16.u32 %rs5, %r699;
and.b32 %r700, %r699, 63;
mul.lo.s32 %r701, %r108, %r700;
cvt.rn.f32.s32 %f163, %r701;
fma.rn.ftz.f32 %f164, %f97, %f163, 0f00000000;
shr.u32 %r702, %r670, 4;
and.b32 %r703, %r702, 252645135;
shr.u32 %r704, %r671, 4;
and.b32 %r705, %r704, 252645135;
shl.b32 %r706, %r673, 3;
and.b32 %r707, %r706, 269488144;
shl.b32 %r708, %r675, 3;
and.b32 %r709, %r708, 269488144;
or.b32 %r621, %r703, %r707;
or.b32 %r617, %r705, %r709;

dp4a.s32.s32 %r112, %r617, %r114, %r663;


dp4a.s32.s32 %r116, %r621, %r118, %r112;


dp4a.s32.s32 %r120, %r665, %r114, %r663;


dp4a.s32.s32 %r124, %r665, %r118, %r120;

shr.u16 %rs6, %rs4, 8;
cvt.u32.u16 %r710, %rs6;
mul.lo.s32 %r711, %r116, %r710;
cvt.rn.f32.s32 %f165, %r711;
fma.rn.ftz.f32 %f166, %f98, %f165, %f162;
shr.u16 %rs7, %rs5, 8;
cvt.u32.u16 %r712, %rs7;
mul.lo.s32 %r713, %r124, %r712;
cvt.rn.f32.s32 %f167, %r713;
fma.rn.ftz.f32 %f168, %f98, %f167, %f164;
ld.global.nc.u32 %r128, [%rd46];

{.reg .f16 low,high;
mov.b32 {low,high},%r128;
cvt.f32.f16 %f99, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r128;
cvt.f32.f16 %f100, high;}


mul.ftz.f32 %f169, %f166, %f99;
mul.ftz.f32 %f170, %f168, %f100;
sub.ftz.f32 %f171, %f169, %f170;
add.ftz.f32 %f544, %f171, %f544;
add.s64 %rd47, %rd140, %rd11;
add.s64 %rd48, %rd140, %rd10;
ld.global.nc.u32 %r714, [%rd47];
ld.global.nc.u32 %r715, [%rd47+16];
ld.global.nc.u32 %r716, [%rd48];
shr.s32 %r717, %r716, %r8;
ld.global.nc.u32 %r718, [%rd48+16];
shr.s32 %r719, %r718, %r8;
add.s64 %rd49, %rd140, %rd7;
ld.global.nc.u16 %rs8, [%rd49+-4];
cvt.u32.u16 %r720, %rs8;
ld.global.nc.u16 %rs9, [%rd49];
cvt.u32.u16 %r721, %rs9;
ld.global.nc.u16 %rs10, [%rd49+4];
cvt.u32.u16 %r722, %rs10;
and.b32 %r723, %r720, 16191;
and.b32 %r724, %r722, 3855;
and.b32 %r725, %r720, 49344;
shr.u32 %r726, %r725, 2;
or.b32 %r727, %r724, %r726;
and.b32 %r728, %r721, 16191;
shr.u32 %r729, %r722, 4;
and.b32 %r730, %r729, 3855;
and.b32 %r731, %r721, 49344;
shr.u32 %r732, %r731, 2;
or.b32 %r733, %r730, %r732;

{.reg .f16 low,high;
mov.b32 {low,high},%r94;
cvt.f32.f16 %f101, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r95;
cvt.f32.f16 %f102, low;}


add.s64 %rd50, %rd140, %rd4;
and.b32 %r734, %r714, 252645135;
and.b32 %r735, %r715, 252645135;
shl.b32 %r736, %r717, 4;
and.b32 %r737, %r736, 269488144;
shl.b32 %r738, %r719, 4;
and.b32 %r739, %r738, 269488144;
or.b32 %r641, %r734, %r737;
or.b32 %r637, %r735, %r739;

dp4a.s32.s32 %r132, %r637, %r98, %r663;


dp4a.s32.s32 %r136, %r641, %r102, %r132;


dp4a.s32.s32 %r140, %r665, %r98, %r663;


dp4a.s32.s32 %r144, %r665, %r102, %r140;

selp.b32 %r740, %r727, %r723, %p4;
cvt.u16.u32 %rs11, %r740;
and.b32 %r741, %r740, 63;
mul.lo.s32 %r742, %r136, %r741;
cvt.rn.f32.s32 %f172, %r742;
fma.rn.ftz.f32 %f173, %f101, %f172, 0f00000000;
selp.b32 %r743, %r733, %r728, %p4;
cvt.u16.u32 %rs12, %r743;
and.b32 %r744, %r743, 63;
mul.lo.s32 %r745, %r144, %r744;
cvt.rn.f32.s32 %f174, %r745;
fma.rn.ftz.f32 %f175, %f101, %f174, 0f00000000;
shr.u32 %r746, %r714, 4;
and.b32 %r747, %r746, 252645135;
shr.u32 %r748, %r715, 4;
and.b32 %r749, %r748, 252645135;
shl.b32 %r750, %r717, 3;
and.b32 %r751, %r750, 269488144;
shl.b32 %r752, %r719, 3;
and.b32 %r753, %r752, 269488144;
or.b32 %r657, %r747, %r751;
or.b32 %r653, %r749, %r753;

dp4a.s32.s32 %r148, %r653, %r114, %r663;


dp4a.s32.s32 %r152, %r657, %r118, %r148;


dp4a.s32.s32 %r156, %r665, %r114, %r663;


dp4a.s32.s32 %r160, %r665, %r118, %r156;

shr.u16 %rs13, %rs11, 8;
cvt.u32.u16 %r754, %rs13;
mul.lo.s32 %r755, %r152, %r754;
cvt.rn.f32.s32 %f176, %r755;
fma.rn.ftz.f32 %f177, %f102, %f176, %f173;
shr.u16 %rs14, %rs12, 8;
cvt.u32.u16 %r756, %rs14;
mul.lo.s32 %r757, %r160, %r756;
cvt.rn.f32.s32 %f178, %r757;
fma.rn.ftz.f32 %f179, %f102, %f178, %f175;
ld.global.nc.u32 %r164, [%rd50];

{.reg .f16 low,high;
mov.b32 {low,high},%r164;
cvt.f32.f16 %f103, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r164;
cvt.f32.f16 %f104, high;}


mul.ftz.f32 %f180, %f177, %f103;
mul.ftz.f32 %f181, %f179, %f104;
sub.ftz.f32 %f182, %f180, %f181;
add.ftz.f32 %f543, %f182, %f543;
cvt.u64.u32 %rd51, %r1109;
add.s64 %rd52, %rd2, %rd51;
mul.lo.s64 %rd53, %rd52, 36;
add.s64 %rd54, %rd12, %rd53;
ld.global.nc.u32 %r166, [%rd54];

{.reg .f16 low,high;
mov.b32 {low,high},%r166;
cvt.f32.f16 %f105, low;}


add.s64 %rd55, %rd54, %rd44;
ld.global.nc.u32 %r174, [%rd55+4];
ld.global.nc.u32 %r170, [%rd55+20];
ld.global.nc.u32 %r167, [%rd54+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r167;
cvt.f32.f16 %f106, low;}


ld.global.nc.u32 %r190, [%rd55+40];
ld.global.nc.u32 %r186, [%rd55+56];

dp4a.s32.s32 %r168, %r601, %r170, %r663;


dp4a.s32.s32 %r172, %r605, %r174, %r168;


dp4a.s32.s32 %r176, %r665, %r170, %r663;


dp4a.s32.s32 %r180, %r665, %r174, %r176;

mul.lo.s32 %r758, %r172, %r697;
cvt.rn.f32.s32 %f183, %r758;
fma.rn.ftz.f32 %f184, %f105, %f183, 0f00000000;
mul.lo.s32 %r759, %r180, %r700;
cvt.rn.f32.s32 %f185, %r759;
fma.rn.ftz.f32 %f186, %f105, %f185, 0f00000000;

dp4a.s32.s32 %r184, %r617, %r186, %r663;


dp4a.s32.s32 %r188, %r621, %r190, %r184;


dp4a.s32.s32 %r192, %r665, %r186, %r663;


dp4a.s32.s32 %r196, %r665, %r190, %r192;

mul.lo.s32 %r760, %r188, %r710;
cvt.rn.f32.s32 %f187, %r760;
fma.rn.ftz.f32 %f188, %f106, %f187, %f184;
mul.lo.s32 %r761, %r196, %r712;
cvt.rn.f32.s32 %f189, %r761;
fma.rn.ftz.f32 %f190, %f106, %f189, %f186;

{.reg .f16 low,high;
mov.b32 {low,high},%r128;
cvt.f32.f16 %f107, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r128;
cvt.f32.f16 %f108, high;}


mul.ftz.f32 %f191, %f188, %f107;
mul.ftz.f32 %f192, %f190, %f108;
sub.ftz.f32 %f193, %f191, %f192;
add.ftz.f32 %f542, %f193, %f542;

{.reg .f16 low,high;
mov.b32 {low,high},%r166;
cvt.f32.f16 %f109, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r167;
cvt.f32.f16 %f110, low;}



dp4a.s32.s32 %r204, %r637, %r170, %r663;


dp4a.s32.s32 %r208, %r641, %r174, %r204;


dp4a.s32.s32 %r212, %r665, %r170, %r663;


dp4a.s32.s32 %r216, %r665, %r174, %r212;

mul.lo.s32 %r762, %r208, %r741;
cvt.rn.f32.s32 %f194, %r762;
fma.rn.ftz.f32 %f195, %f109, %f194, 0f00000000;
mul.lo.s32 %r763, %r216, %r744;
cvt.rn.f32.s32 %f196, %r763;
fma.rn.ftz.f32 %f197, %f109, %f196, 0f00000000;

dp4a.s32.s32 %r220, %r653, %r186, %r663;


dp4a.s32.s32 %r224, %r657, %r190, %r220;


dp4a.s32.s32 %r228, %r665, %r186, %r663;


dp4a.s32.s32 %r232, %r665, %r190, %r228;

mul.lo.s32 %r764, %r224, %r754;
cvt.rn.f32.s32 %f198, %r764;
fma.rn.ftz.f32 %f199, %f110, %f198, %f195;
mul.lo.s32 %r765, %r232, %r756;
cvt.rn.f32.s32 %f200, %r765;
fma.rn.ftz.f32 %f201, %f110, %f200, %f197;

{.reg .f16 low,high;
mov.b32 {low,high},%r164;
cvt.f32.f16 %f111, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r164;
cvt.f32.f16 %f112, high;}


mul.ftz.f32 %f202, %f199, %f111;
mul.ftz.f32 %f203, %f201, %f112;
sub.ftz.f32 %f204, %f202, %f203;
add.ftz.f32 %f541, %f204, %f541;
cvt.u64.u32 %rd56, %r1115;
add.s64 %rd57, %rd2, %rd56;
mul.lo.s64 %rd58, %rd57, 36;
add.s64 %rd59, %rd12, %rd58;
ld.global.nc.u32 %r238, [%rd59];

{.reg .f16 low,high;
mov.b32 {low,high},%r238;
cvt.f32.f16 %f113, low;}


add.s64 %rd60, %rd59, %rd44;
ld.global.nc.u32 %r246, [%rd60+4];
ld.global.nc.u32 %r242, [%rd60+20];
ld.global.nc.u32 %r239, [%rd59+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r239;
cvt.f32.f16 %f114, low;}


ld.global.nc.u32 %r262, [%rd60+40];
ld.global.nc.u32 %r258, [%rd60+56];

dp4a.s32.s32 %r240, %r601, %r242, %r663;


dp4a.s32.s32 %r244, %r605, %r246, %r240;


dp4a.s32.s32 %r248, %r665, %r242, %r663;


dp4a.s32.s32 %r252, %r665, %r246, %r248;

mul.lo.s32 %r766, %r244, %r697;
cvt.rn.f32.s32 %f205, %r766;
fma.rn.ftz.f32 %f206, %f113, %f205, 0f00000000;
mul.lo.s32 %r767, %r252, %r700;
cvt.rn.f32.s32 %f207, %r767;
fma.rn.ftz.f32 %f208, %f113, %f207, 0f00000000;

dp4a.s32.s32 %r256, %r617, %r258, %r663;


dp4a.s32.s32 %r260, %r621, %r262, %r256;


dp4a.s32.s32 %r264, %r665, %r258, %r663;


dp4a.s32.s32 %r268, %r665, %r262, %r264;

mul.lo.s32 %r768, %r260, %r710;
cvt.rn.f32.s32 %f209, %r768;
fma.rn.ftz.f32 %f210, %f114, %f209, %f206;
mul.lo.s32 %r769, %r268, %r712;
cvt.rn.f32.s32 %f211, %r769;
fma.rn.ftz.f32 %f212, %f114, %f211, %f208;

{.reg .f16 low,high;
mov.b32 {low,high},%r128;
cvt.f32.f16 %f115, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r128;
cvt.f32.f16 %f116, high;}


mul.ftz.f32 %f213, %f210, %f115;
mul.ftz.f32 %f214, %f212, %f116;
sub.ftz.f32 %f215, %f213, %f214;
add.ftz.f32 %f540, %f215, %f540;

{.reg .f16 low,high;
mov.b32 {low,high},%r238;
cvt.f32.f16 %f117, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r239;
cvt.f32.f16 %f118, low;}



dp4a.s32.s32 %r276, %r637, %r242, %r663;


dp4a.s32.s32 %r280, %r641, %r246, %r276;


dp4a.s32.s32 %r284, %r665, %r242, %r663;


dp4a.s32.s32 %r288, %r665, %r246, %r284;

mul.lo.s32 %r770, %r280, %r741;
cvt.rn.f32.s32 %f216, %r770;
fma.rn.ftz.f32 %f217, %f117, %f216, 0f00000000;
mul.lo.s32 %r771, %r288, %r744;
cvt.rn.f32.s32 %f218, %r771;
fma.rn.ftz.f32 %f219, %f117, %f218, 0f00000000;

dp4a.s32.s32 %r292, %r653, %r258, %r663;


dp4a.s32.s32 %r296, %r657, %r262, %r292;


dp4a.s32.s32 %r300, %r665, %r258, %r663;


dp4a.s32.s32 %r304, %r665, %r262, %r300;

mul.lo.s32 %r772, %r296, %r754;
cvt.rn.f32.s32 %f220, %r772;
fma.rn.ftz.f32 %f221, %f118, %f220, %f217;
mul.lo.s32 %r773, %r304, %r756;
cvt.rn.f32.s32 %f222, %r773;
fma.rn.ftz.f32 %f223, %f118, %f222, %f219;

{.reg .f16 low,high;
mov.b32 {low,high},%r164;
cvt.f32.f16 %f119, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r164;
cvt.f32.f16 %f120, high;}


mul.ftz.f32 %f224, %f221, %f119;
mul.ftz.f32 %f225, %f223, %f120;
sub.ftz.f32 %f226, %f224, %f225;
add.ftz.f32 %f539, %f226, %f539;
cvt.u64.u32 %rd61, %r1114;
add.s64 %rd62, %rd2, %rd61;
mul.lo.s64 %rd63, %rd62, 36;
add.s64 %rd64, %rd12, %rd63;
ld.global.nc.u32 %r310, [%rd64];

{.reg .f16 low,high;
mov.b32 {low,high},%r310;
cvt.f32.f16 %f121, low;}


add.s64 %rd65, %rd64, %rd44;
ld.global.nc.u32 %r318, [%rd65+4];
ld.global.nc.u32 %r314, [%rd65+20];
ld.global.nc.u32 %r311, [%rd64+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r311;
cvt.f32.f16 %f122, low;}


ld.global.nc.u32 %r334, [%rd65+40];
ld.global.nc.u32 %r330, [%rd65+56];

dp4a.s32.s32 %r312, %r601, %r314, %r663;


dp4a.s32.s32 %r316, %r605, %r318, %r312;


dp4a.s32.s32 %r320, %r665, %r314, %r663;


dp4a.s32.s32 %r324, %r665, %r318, %r320;

mul.lo.s32 %r774, %r316, %r697;
cvt.rn.f32.s32 %f227, %r774;
fma.rn.ftz.f32 %f228, %f121, %f227, 0f00000000;
mul.lo.s32 %r775, %r324, %r700;
cvt.rn.f32.s32 %f229, %r775;
fma.rn.ftz.f32 %f230, %f121, %f229, 0f00000000;

dp4a.s32.s32 %r328, %r617, %r330, %r663;


dp4a.s32.s32 %r332, %r621, %r334, %r328;


dp4a.s32.s32 %r336, %r665, %r330, %r663;


dp4a.s32.s32 %r340, %r665, %r334, %r336;

mul.lo.s32 %r776, %r332, %r710;
cvt.rn.f32.s32 %f231, %r776;
fma.rn.ftz.f32 %f232, %f122, %f231, %f228;
mul.lo.s32 %r777, %r340, %r712;
cvt.rn.f32.s32 %f233, %r777;
fma.rn.ftz.f32 %f234, %f122, %f233, %f230;

{.reg .f16 low,high;
mov.b32 {low,high},%r128;
cvt.f32.f16 %f123, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r128;
cvt.f32.f16 %f124, high;}


mul.ftz.f32 %f235, %f232, %f123;
mul.ftz.f32 %f236, %f234, %f124;
sub.ftz.f32 %f237, %f235, %f236;
add.ftz.f32 %f538, %f237, %f538;

{.reg .f16 low,high;
mov.b32 {low,high},%r310;
cvt.f32.f16 %f125, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r311;
cvt.f32.f16 %f126, low;}



dp4a.s32.s32 %r348, %r637, %r314, %r663;


dp4a.s32.s32 %r352, %r641, %r318, %r348;


dp4a.s32.s32 %r356, %r665, %r314, %r663;


dp4a.s32.s32 %r360, %r665, %r318, %r356;

mul.lo.s32 %r778, %r352, %r741;
cvt.rn.f32.s32 %f238, %r778;
fma.rn.ftz.f32 %f239, %f125, %f238, 0f00000000;
mul.lo.s32 %r779, %r360, %r744;
cvt.rn.f32.s32 %f240, %r779;
fma.rn.ftz.f32 %f241, %f125, %f240, 0f00000000;

dp4a.s32.s32 %r364, %r653, %r330, %r663;


dp4a.s32.s32 %r368, %r657, %r334, %r364;


dp4a.s32.s32 %r372, %r665, %r330, %r663;


dp4a.s32.s32 %r376, %r665, %r334, %r372;

mul.lo.s32 %r780, %r368, %r754;
cvt.rn.f32.s32 %f242, %r780;
fma.rn.ftz.f32 %f243, %f126, %f242, %f239;
mul.lo.s32 %r781, %r376, %r756;
cvt.rn.f32.s32 %f244, %r781;
fma.rn.ftz.f32 %f245, %f126, %f244, %f241;

{.reg .f16 low,high;
mov.b32 {low,high},%r164;
cvt.f32.f16 %f127, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r164;
cvt.f32.f16 %f128, high;}


mul.ftz.f32 %f246, %f243, %f127;
mul.ftz.f32 %f247, %f245, %f128;
sub.ftz.f32 %f248, %f246, %f247;
add.ftz.f32 %f537, %f248, %f537;
cvt.u64.u32 %rd66, %r1113;
add.s64 %rd67, %rd2, %rd66;
mul.lo.s64 %rd68, %rd67, 36;
add.s64 %rd69, %rd12, %rd68;
ld.global.nc.u32 %r382, [%rd69];

{.reg .f16 low,high;
mov.b32 {low,high},%r382;
cvt.f32.f16 %f129, low;}


add.s64 %rd70, %rd69, %rd44;
ld.global.nc.u32 %r390, [%rd70+4];
ld.global.nc.u32 %r386, [%rd70+20];
ld.global.nc.u32 %r383, [%rd69+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r383;
cvt.f32.f16 %f130, low;}


ld.global.nc.u32 %r406, [%rd70+40];
ld.global.nc.u32 %r402, [%rd70+56];

dp4a.s32.s32 %r384, %r601, %r386, %r663;


dp4a.s32.s32 %r388, %r605, %r390, %r384;


dp4a.s32.s32 %r392, %r665, %r386, %r663;


dp4a.s32.s32 %r396, %r665, %r390, %r392;

mul.lo.s32 %r782, %r388, %r697;
cvt.rn.f32.s32 %f249, %r782;
fma.rn.ftz.f32 %f250, %f129, %f249, 0f00000000;
mul.lo.s32 %r783, %r396, %r700;
cvt.rn.f32.s32 %f251, %r783;
fma.rn.ftz.f32 %f252, %f129, %f251, 0f00000000;

dp4a.s32.s32 %r400, %r617, %r402, %r663;


dp4a.s32.s32 %r404, %r621, %r406, %r400;


dp4a.s32.s32 %r408, %r665, %r402, %r663;


dp4a.s32.s32 %r412, %r665, %r406, %r408;

mul.lo.s32 %r784, %r404, %r710;
cvt.rn.f32.s32 %f253, %r784;
fma.rn.ftz.f32 %f254, %f130, %f253, %f250;
mul.lo.s32 %r785, %r412, %r712;
cvt.rn.f32.s32 %f255, %r785;
fma.rn.ftz.f32 %f256, %f130, %f255, %f252;

{.reg .f16 low,high;
mov.b32 {low,high},%r128;
cvt.f32.f16 %f131, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r128;
cvt.f32.f16 %f132, high;}


mul.ftz.f32 %f257, %f254, %f131;
mul.ftz.f32 %f258, %f256, %f132;
sub.ftz.f32 %f259, %f257, %f258;
add.ftz.f32 %f536, %f259, %f536;

{.reg .f16 low,high;
mov.b32 {low,high},%r382;
cvt.f32.f16 %f133, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r383;
cvt.f32.f16 %f134, low;}



dp4a.s32.s32 %r420, %r637, %r386, %r663;


dp4a.s32.s32 %r424, %r641, %r390, %r420;


dp4a.s32.s32 %r428, %r665, %r386, %r663;


dp4a.s32.s32 %r432, %r665, %r390, %r428;

mul.lo.s32 %r786, %r424, %r741;
cvt.rn.f32.s32 %f260, %r786;
fma.rn.ftz.f32 %f261, %f133, %f260, 0f00000000;
mul.lo.s32 %r787, %r432, %r744;
cvt.rn.f32.s32 %f262, %r787;
fma.rn.ftz.f32 %f263, %f133, %f262, 0f00000000;

dp4a.s32.s32 %r436, %r653, %r402, %r663;


dp4a.s32.s32 %r440, %r657, %r406, %r436;


dp4a.s32.s32 %r444, %r665, %r402, %r663;


dp4a.s32.s32 %r448, %r665, %r406, %r444;

mul.lo.s32 %r788, %r440, %r754;
cvt.rn.f32.s32 %f264, %r788;
fma.rn.ftz.f32 %f265, %f134, %f264, %f261;
mul.lo.s32 %r789, %r448, %r756;
cvt.rn.f32.s32 %f266, %r789;
fma.rn.ftz.f32 %f267, %f134, %f266, %f263;

{.reg .f16 low,high;
mov.b32 {low,high},%r164;
cvt.f32.f16 %f135, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r164;
cvt.f32.f16 %f136, high;}


mul.ftz.f32 %f268, %f265, %f135;
mul.ftz.f32 %f269, %f267, %f136;
sub.ftz.f32 %f270, %f268, %f269;
add.ftz.f32 %f535, %f270, %f535;
cvt.u64.u32 %rd71, %r1112;
add.s64 %rd72, %rd2, %rd71;
mul.lo.s64 %rd73, %rd72, 36;
add.s64 %rd74, %rd12, %rd73;
ld.global.nc.u32 %r454, [%rd74];

{.reg .f16 low,high;
mov.b32 {low,high},%r454;
cvt.f32.f16 %f137, low;}


add.s64 %rd75, %rd74, %rd44;
ld.global.nc.u32 %r462, [%rd75+4];
ld.global.nc.u32 %r458, [%rd75+20];
ld.global.nc.u32 %r455, [%rd74+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r455;
cvt.f32.f16 %f138, low;}


ld.global.nc.u32 %r478, [%rd75+40];
ld.global.nc.u32 %r474, [%rd75+56];

dp4a.s32.s32 %r456, %r601, %r458, %r663;


dp4a.s32.s32 %r460, %r605, %r462, %r456;


dp4a.s32.s32 %r464, %r665, %r458, %r663;


dp4a.s32.s32 %r468, %r665, %r462, %r464;

mul.lo.s32 %r790, %r460, %r697;
cvt.rn.f32.s32 %f271, %r790;
fma.rn.ftz.f32 %f272, %f137, %f271, 0f00000000;
mul.lo.s32 %r791, %r468, %r700;
cvt.rn.f32.s32 %f273, %r791;
fma.rn.ftz.f32 %f274, %f137, %f273, 0f00000000;

dp4a.s32.s32 %r472, %r617, %r474, %r663;


dp4a.s32.s32 %r476, %r621, %r478, %r472;


dp4a.s32.s32 %r480, %r665, %r474, %r663;


dp4a.s32.s32 %r484, %r665, %r478, %r480;

mul.lo.s32 %r792, %r476, %r710;
cvt.rn.f32.s32 %f275, %r792;
fma.rn.ftz.f32 %f276, %f138, %f275, %f272;
mul.lo.s32 %r793, %r484, %r712;
cvt.rn.f32.s32 %f277, %r793;
fma.rn.ftz.f32 %f278, %f138, %f277, %f274;

{.reg .f16 low,high;
mov.b32 {low,high},%r128;
cvt.f32.f16 %f139, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r128;
cvt.f32.f16 %f140, high;}


mul.ftz.f32 %f279, %f276, %f139;
mul.ftz.f32 %f280, %f278, %f140;
sub.ftz.f32 %f281, %f279, %f280;
add.ftz.f32 %f534, %f281, %f534;

{.reg .f16 low,high;
mov.b32 {low,high},%r454;
cvt.f32.f16 %f141, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r455;
cvt.f32.f16 %f142, low;}



dp4a.s32.s32 %r492, %r637, %r458, %r663;


dp4a.s32.s32 %r496, %r641, %r462, %r492;


dp4a.s32.s32 %r500, %r665, %r458, %r663;


dp4a.s32.s32 %r504, %r665, %r462, %r500;

mul.lo.s32 %r794, %r496, %r741;
cvt.rn.f32.s32 %f282, %r794;
fma.rn.ftz.f32 %f283, %f141, %f282, 0f00000000;
mul.lo.s32 %r795, %r504, %r744;
cvt.rn.f32.s32 %f284, %r795;
fma.rn.ftz.f32 %f285, %f141, %f284, 0f00000000;

dp4a.s32.s32 %r508, %r653, %r474, %r663;


dp4a.s32.s32 %r512, %r657, %r478, %r508;


dp4a.s32.s32 %r516, %r665, %r474, %r663;


dp4a.s32.s32 %r520, %r665, %r478, %r516;

mul.lo.s32 %r796, %r512, %r754;
cvt.rn.f32.s32 %f286, %r796;
fma.rn.ftz.f32 %f287, %f142, %f286, %f283;
mul.lo.s32 %r797, %r520, %r756;
cvt.rn.f32.s32 %f288, %r797;
fma.rn.ftz.f32 %f289, %f142, %f288, %f285;

{.reg .f16 low,high;
mov.b32 {low,high},%r164;
cvt.f32.f16 %f143, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r164;
cvt.f32.f16 %f144, high;}


mul.ftz.f32 %f290, %f287, %f143;
mul.ftz.f32 %f291, %f289, %f144;
sub.ftz.f32 %f292, %f290, %f291;
add.ftz.f32 %f533, %f292, %f533;
cvt.u64.u32 %rd76, %r1111;
add.s64 %rd77, %rd2, %rd76;
mul.lo.s64 %rd78, %rd77, 36;
add.s64 %rd79, %rd12, %rd78;
ld.global.nc.u32 %r526, [%rd79];

{.reg .f16 low,high;
mov.b32 {low,high},%r526;
cvt.f32.f16 %f145, low;}


add.s64 %rd80, %rd79, %rd44;
ld.global.nc.u32 %r534, [%rd80+4];
ld.global.nc.u32 %r530, [%rd80+20];
ld.global.nc.u32 %r527, [%rd79+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r527;
cvt.f32.f16 %f146, low;}


ld.global.nc.u32 %r550, [%rd80+40];
ld.global.nc.u32 %r546, [%rd80+56];

dp4a.s32.s32 %r528, %r601, %r530, %r663;


dp4a.s32.s32 %r532, %r605, %r534, %r528;


dp4a.s32.s32 %r536, %r665, %r530, %r663;


dp4a.s32.s32 %r540, %r665, %r534, %r536;

mul.lo.s32 %r798, %r532, %r697;
cvt.rn.f32.s32 %f293, %r798;
fma.rn.ftz.f32 %f294, %f145, %f293, 0f00000000;
mul.lo.s32 %r799, %r540, %r700;
cvt.rn.f32.s32 %f295, %r799;
fma.rn.ftz.f32 %f296, %f145, %f295, 0f00000000;

dp4a.s32.s32 %r544, %r617, %r546, %r663;


dp4a.s32.s32 %r548, %r621, %r550, %r544;


dp4a.s32.s32 %r552, %r665, %r546, %r663;


dp4a.s32.s32 %r556, %r665, %r550, %r552;

mul.lo.s32 %r800, %r548, %r710;
cvt.rn.f32.s32 %f297, %r800;
fma.rn.ftz.f32 %f298, %f146, %f297, %f294;
mul.lo.s32 %r801, %r556, %r712;
cvt.rn.f32.s32 %f299, %r801;
fma.rn.ftz.f32 %f300, %f146, %f299, %f296;

{.reg .f16 low,high;
mov.b32 {low,high},%r128;
cvt.f32.f16 %f147, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r128;
cvt.f32.f16 %f148, high;}


mul.ftz.f32 %f301, %f298, %f147;
mul.ftz.f32 %f302, %f300, %f148;
sub.ftz.f32 %f303, %f301, %f302;
add.ftz.f32 %f532, %f303, %f532;

{.reg .f16 low,high;
mov.b32 {low,high},%r526;
cvt.f32.f16 %f149, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r527;
cvt.f32.f16 %f150, low;}



dp4a.s32.s32 %r564, %r637, %r530, %r663;


dp4a.s32.s32 %r568, %r641, %r534, %r564;


dp4a.s32.s32 %r572, %r665, %r530, %r663;


dp4a.s32.s32 %r576, %r665, %r534, %r572;

mul.lo.s32 %r802, %r568, %r741;
cvt.rn.f32.s32 %f304, %r802;
fma.rn.ftz.f32 %f305, %f149, %f304, 0f00000000;
mul.lo.s32 %r803, %r576, %r744;
cvt.rn.f32.s32 %f306, %r803;
fma.rn.ftz.f32 %f307, %f149, %f306, 0f00000000;

dp4a.s32.s32 %r580, %r653, %r546, %r663;


dp4a.s32.s32 %r584, %r657, %r550, %r580;


dp4a.s32.s32 %r588, %r665, %r546, %r663;


dp4a.s32.s32 %r592, %r665, %r550, %r588;

mul.lo.s32 %r804, %r584, %r754;
cvt.rn.f32.s32 %f308, %r804;
fma.rn.ftz.f32 %f309, %f150, %f308, %f305;
mul.lo.s32 %r805, %r592, %r756;
cvt.rn.f32.s32 %f310, %r805;
fma.rn.ftz.f32 %f311, %f150, %f310, %f307;

{.reg .f16 low,high;
mov.b32 {low,high},%r164;
cvt.f32.f16 %f151, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r164;
cvt.f32.f16 %f152, high;}


mul.ftz.f32 %f312, %f309, %f151;
mul.ftz.f32 %f313, %f311, %f152;
sub.ftz.f32 %f314, %f312, %f313;
add.ftz.f32 %f531, %f314, %f531;
cvt.u64.u32 %rd81, %r1110;
add.s64 %rd82, %rd2, %rd81;
mul.lo.s64 %rd83, %rd82, 36;
add.s64 %rd84, %rd12, %rd83;
ld.global.nc.u32 %r598, [%rd84];

{.reg .f16 low,high;
mov.b32 {low,high},%r598;
cvt.f32.f16 %f153, low;}


add.s64 %rd85, %rd84, %rd44;
ld.global.nc.u32 %r606, [%rd85+4];
ld.global.nc.u32 %r602, [%rd85+20];
ld.global.nc.u32 %r599, [%rd84+36];

{.reg .f16 low,high;
mov.b32 {low,high},%r599;
cvt.f32.f16 %f154, low;}


ld.global.nc.u32 %r622, [%rd85+40];
ld.global.nc.u32 %r618, [%rd85+56];

dp4a.s32.s32 %r600, %r601, %r602, %r663;


dp4a.s32.s32 %r604, %r605, %r606, %r600;


dp4a.s32.s32 %r608, %r665, %r602, %r663;


dp4a.s32.s32 %r612, %r665, %r606, %r608;

mul.lo.s32 %r806, %r604, %r697;
cvt.rn.f32.s32 %f315, %r806;
fma.rn.ftz.f32 %f316, %f153, %f315, 0f00000000;
mul.lo.s32 %r807, %r612, %r700;
cvt.rn.f32.s32 %f317, %r807;
fma.rn.ftz.f32 %f318, %f153, %f317, 0f00000000;

dp4a.s32.s32 %r616, %r617, %r618, %r663;


dp4a.s32.s32 %r620, %r621, %r622, %r616;


dp4a.s32.s32 %r624, %r665, %r618, %r663;


dp4a.s32.s32 %r628, %r665, %r622, %r624;

mul.lo.s32 %r808, %r620, %r710;
cvt.rn.f32.s32 %f319, %r808;
fma.rn.ftz.f32 %f320, %f154, %f319, %f316;
mul.lo.s32 %r809, %r628, %r712;
cvt.rn.f32.s32 %f321, %r809;
fma.rn.ftz.f32 %f322, %f154, %f321, %f318;

{.reg .f16 low,high;
mov.b32 {low,high},%r128;
cvt.f32.f16 %f155, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r128;
cvt.f32.f16 %f156, high;}


mul.ftz.f32 %f323, %f320, %f155;
mul.ftz.f32 %f324, %f322, %f156;
sub.ftz.f32 %f325, %f323, %f324;
add.ftz.f32 %f530, %f325, %f530;

{.reg .f16 low,high;
mov.b32 {low,high},%r598;
cvt.f32.f16 %f157, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r599;
cvt.f32.f16 %f158, low;}



dp4a.s32.s32 %r636, %r637, %r602, %r663;


dp4a.s32.s32 %r640, %r641, %r606, %r636;


dp4a.s32.s32 %r644, %r665, %r602, %r663;


dp4a.s32.s32 %r648, %r665, %r606, %r644;

mul.lo.s32 %r810, %r640, %r741;
cvt.rn.f32.s32 %f326, %r810;
fma.rn.ftz.f32 %f327, %f157, %f326, 0f00000000;
mul.lo.s32 %r811, %r648, %r744;
cvt.rn.f32.s32 %f328, %r811;
fma.rn.ftz.f32 %f329, %f157, %f328, 0f00000000;

dp4a.s32.s32 %r652, %r653, %r618, %r663;


dp4a.s32.s32 %r656, %r657, %r622, %r652;


dp4a.s32.s32 %r660, %r665, %r618, %r663;


dp4a.s32.s32 %r664, %r665, %r622, %r660;

mul.lo.s32 %r812, %r656, %r754;
cvt.rn.f32.s32 %f330, %r812;
fma.rn.ftz.f32 %f331, %f158, %f330, %f327;
mul.lo.s32 %r813, %r664, %r756;
cvt.rn.f32.s32 %f332, %r813;
fma.rn.ftz.f32 %f333, %f158, %f332, %f329;

{.reg .f16 low,high;
mov.b32 {low,high},%r164;
cvt.f32.f16 %f159, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r164;
cvt.f32.f16 %f160, high;}


mul.ftz.f32 %f334, %f331, %f159;
mul.ftz.f32 %f335, %f333, %f160;
sub.ftz.f32 %f336, %f334, %f335;
add.ftz.f32 %f529, %f336, %f529;
add.s32 %r1115, %r1115, 32;
add.s32 %r1114, %r1114, 32;
add.s32 %r1113, %r1113, 32;
add.s32 %r1112, %r1112, 32;
add.s32 %r1111, %r1111, 32;
add.s32 %r1110, %r1110, 32;
add.s32 %r1109, %r1109, 32;
add.s32 %r1108, %r1108, 32;
add.s64 %rd140, %rd140, 704;
add.s32 %r1116, %r1116, 4;
setp.lt.s32 %p5, %r1116, %r6;
@%p5 bra $L__BB182_2;

$L__BB182_3:
setp.eq.s32 %p6, %r1, 0;
@%p6 bra $L__BB182_5;

shl.b32 %r814, %r1, 11;
mov.u32 %r815, _ZZ13mul_mat_vec_qIL9ggml_type13ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r816, %r815, %r814;
shl.b32 %r817, %r2, 2;
add.s32 %r818, %r816, %r817;
st.shared.f32 [%r818+-2048], %f544;
st.shared.f32 [%r818+-1920], %f543;
st.shared.f32 [%r818+-1792], %f542;
st.shared.f32 [%r818+-1664], %f541;
st.shared.f32 [%r818+-1536], %f540;
st.shared.f32 [%r818+-1408], %f539;
st.shared.f32 [%r818+-1280], %f538;
st.shared.f32 [%r818+-1152], %f537;
st.shared.f32 [%r818+-1024], %f536;
st.shared.f32 [%r818+-896], %f535;
st.shared.f32 [%r818+-768], %f534;
st.shared.f32 [%r818+-640], %f533;
st.shared.f32 [%r818+-512], %f532;
st.shared.f32 [%r818+-384], %f531;
st.shared.f32 [%r818+-256], %f530;
st.shared.f32 [%r818+-128], %f529;

$L__BB182_5:
setp.ne.s32 %p7, %r1, 0;
bar.sync 0;
@%p7 bra $L__BB182_38;

mov.u32 %r819, %ctaid.z;
mov.u32 %r820, %ctaid.y;
mad.lo.s32 %r821, %r820, %r47, %r4;
mad.lo.s32 %r822, %r819, %r53, %r821;
cvt.u64.u32 %rd15, %r822;
shl.b32 %r823, %r2, 2;
mov.u32 %r824, _ZZ13mul_mat_vec_qIL9ggml_type13ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r35, %r824, %r823;
mov.u32 %r825, 2;
ld.shared.f32 %f337, [%r35];
add.ftz.f32 %f338, %f337, %f544;
mov.b32 %r826, %f338;
mov.u32 %r827, 31;
mov.u32 %r828, 16;
mov.u32 %r829, -1;
shfl.sync.bfly.b32 %r830|%p8, %r826, %r828, %r827, %r829;
mov.b32 %f339, %r830;
add.ftz.f32 %f340, %f338, %f339;
mov.b32 %r831, %f340;
mov.u32 %r832, 8;
shfl.sync.bfly.b32 %r833|%p9, %r831, %r832, %r827, %r829;
mov.b32 %f341, %r833;
add.ftz.f32 %f342, %f340, %f341;
mov.b32 %r834, %f342;
mov.u32 %r835, 4;
shfl.sync.bfly.b32 %r836|%p10, %r834, %r835, %r827, %r829;
mov.b32 %f343, %r836;
add.ftz.f32 %f344, %f342, %f343;
mov.b32 %r837, %f344;
shfl.sync.bfly.b32 %r838|%p11, %r837, %r825, %r827, %r829;
mov.b32 %f345, %r838;
add.ftz.f32 %f346, %f344, %f345;
mov.b32 %r839, %f346;
mov.u32 %r840, 1;
shfl.sync.bfly.b32 %r841|%p12, %r839, %r840, %r827, %r829;
mov.b32 %f347, %r841;
add.ftz.f32 %f49, %f346, %f347;
setp.lt.u32 %p13, %r4, %r41;
setp.eq.s32 %p14, %r2, 0;
and.pred %p1, %p14, %p13;
cvta.to.global.u64 %rd86, %rd22;
mul.wide.u32 %rd87, %r822, 4;
add.s64 %rd16, %rd86, %rd87;
not.pred %p15, %p1;
@%p15 bra $L__BB182_8;

st.global.f32 [%rd16], %f49;

$L__BB182_8:
ld.shared.f32 %f348, [%r35+128];
add.ftz.f32 %f349, %f348, %f543;
mov.b32 %r842, %f349;
shfl.sync.bfly.b32 %r846|%p16, %r842, %r828, %r827, %r829;
mov.b32 %f350, %r846;
add.ftz.f32 %f351, %f349, %f350;
mov.b32 %r847, %f351;
shfl.sync.bfly.b32 %r849|%p17, %r847, %r832, %r827, %r829;
mov.b32 %f352, %r849;
add.ftz.f32 %f353, %f351, %f352;
mov.b32 %r850, %f353;
shfl.sync.bfly.b32 %r852|%p18, %r850, %r835, %r827, %r829;
mov.b32 %f354, %r852;
add.ftz.f32 %f355, %f353, %f354;
mov.b32 %r853, %f355;
shfl.sync.bfly.b32 %r855|%p19, %r853, %r825, %r827, %r829;
mov.b32 %f356, %r855;
add.ftz.f32 %f357, %f355, %f356;
mov.b32 %r856, %f357;
shfl.sync.bfly.b32 %r858|%p20, %r856, %r840, %r827, %r829;
mov.b32 %f358, %r858;
add.ftz.f32 %f50, %f357, %f358;
setp.eq.s32 %p21, %r2, 1;
or.b32 %r862, %r4, 1;
setp.lt.u32 %p22, %r862, %r41;
and.pred %p2, %p21, %p22;
not.pred %p23, %p2;
@%p23 bra $L__BB182_10;

st.global.f32 [%rd16+4], %f50;

$L__BB182_10:
ld.shared.f32 %f359, [%r35+256];
add.ftz.f32 %f360, %f359, %f542;
mov.b32 %r863, %f360;
mov.u32 %r864, 31;
mov.u32 %r865, 16;
mov.u32 %r866, -1;
shfl.sync.bfly.b32 %r867|%p24, %r863, %r865, %r864, %r866;
mov.b32 %f361, %r867;
add.ftz.f32 %f362, %f360, %f361;
mov.b32 %r868, %f362;
mov.u32 %r869, 8;
shfl.sync.bfly.b32 %r870|%p25, %r868, %r869, %r864, %r866;
mov.b32 %f363, %r870;
add.ftz.f32 %f364, %f362, %f363;
mov.b32 %r871, %f364;
mov.u32 %r872, 4;
shfl.sync.bfly.b32 %r873|%p26, %r871, %r872, %r864, %r866;
mov.b32 %f365, %r873;
add.ftz.f32 %f366, %f364, %f365;
mov.b32 %r874, %f366;
mov.u32 %r875, 2;
shfl.sync.bfly.b32 %r876|%p27, %r874, %r875, %r864, %r866;
mov.b32 %f367, %r876;
add.ftz.f32 %f368, %f366, %f367;
mov.b32 %r877, %f368;
mov.u32 %r878, 1;
shfl.sync.bfly.b32 %r879|%p28, %r877, %r878, %r864, %r866;
mov.b32 %f369, %r879;
add.ftz.f32 %f51, %f368, %f369;
@%p15 bra $L__BB182_12;

cvt.u64.u32 %rd88, %r41;
add.s64 %rd89, %rd88, %rd15;
shl.b64 %rd91, %rd89, 2;
add.s64 %rd92, %rd86, %rd91;
st.global.f32 [%rd92], %f51;

$L__BB182_12:
ld.shared.f32 %f370, [%r35+384];
add.ftz.f32 %f371, %f370, %f541;
mov.b32 %r880, %f371;
shfl.sync.bfly.b32 %r884|%p30, %r880, %r865, %r864, %r866;
mov.b32 %f372, %r884;
add.ftz.f32 %f373, %f371, %f372;
mov.b32 %r885, %f373;
shfl.sync.bfly.b32 %r887|%p31, %r885, %r869, %r864, %r866;
mov.b32 %f374, %r887;
add.ftz.f32 %f375, %f373, %f374;
mov.b32 %r888, %f375;
shfl.sync.bfly.b32 %r890|%p32, %r888, %r872, %r864, %r866;
mov.b32 %f376, %r890;
add.ftz.f32 %f377, %f375, %f376;
mov.b32 %r891, %f377;
shfl.sync.bfly.b32 %r893|%p33, %r891, %r875, %r864, %r866;
mov.b32 %f378, %r893;
add.ftz.f32 %f379, %f377, %f378;
mov.b32 %r894, %f379;
shfl.sync.bfly.b32 %r896|%p34, %r894, %r878, %r864, %r866;
mov.b32 %f380, %r896;
add.ftz.f32 %f52, %f379, %f380;
@%p23 bra $L__BB182_14;

add.s32 %r897, %r41, 1;
cvt.u64.u32 %rd93, %r897;
add.s64 %rd94, %rd93, %rd15;
shl.b64 %rd96, %rd94, 2;
add.s64 %rd97, %rd86, %rd96;
st.global.f32 [%rd97], %f52;

$L__BB182_14:
shl.b32 %r898, %r41, 1;
mov.u32 %r899, 1;
ld.shared.f32 %f381, [%r35+512];
add.ftz.f32 %f382, %f381, %f540;
mov.b32 %r900, %f382;
mov.u32 %r901, 31;
mov.u32 %r902, 16;
mov.u32 %r903, -1;
shfl.sync.bfly.b32 %r904|%p36, %r900, %r902, %r901, %r903;
mov.b32 %f383, %r904;
add.ftz.f32 %f384, %f382, %f383;
mov.b32 %r905, %f384;
mov.u32 %r906, 8;
shfl.sync.bfly.b32 %r907|%p37, %r905, %r906, %r901, %r903;
mov.b32 %f385, %r907;
add.ftz.f32 %f386, %f384, %f385;
mov.b32 %r908, %f386;
mov.u32 %r909, 4;
shfl.sync.bfly.b32 %r910|%p38, %r908, %r909, %r901, %r903;
mov.b32 %f387, %r910;
add.ftz.f32 %f388, %f386, %f387;
mov.b32 %r911, %f388;
mov.u32 %r912, 2;
shfl.sync.bfly.b32 %r913|%p39, %r911, %r912, %r901, %r903;
mov.b32 %f389, %r913;
add.ftz.f32 %f390, %f388, %f389;
mov.b32 %r914, %f390;
shfl.sync.bfly.b32 %r915|%p40, %r914, %r899, %r901, %r903;
mov.b32 %f391, %r915;
add.ftz.f32 %f53, %f390, %f391;
cvt.u64.u32 %rd98, %r898;
add.s64 %rd99, %rd98, %rd15;
shl.b64 %rd101, %rd99, 2;
add.s64 %rd17, %rd86, %rd101;
@%p15 bra $L__BB182_16;

st.global.f32 [%rd17], %f53;

$L__BB182_16:
ld.shared.f32 %f392, [%r35+640];
add.ftz.f32 %f393, %f392, %f539;
mov.b32 %r916, %f393;
shfl.sync.bfly.b32 %r920|%p42, %r916, %r902, %r901, %r903;
mov.b32 %f394, %r920;
add.ftz.f32 %f395, %f393, %f394;
mov.b32 %r921, %f395;
shfl.sync.bfly.b32 %r923|%p43, %r921, %r906, %r901, %r903;
mov.b32 %f396, %r923;
add.ftz.f32 %f397, %f395, %f396;
mov.b32 %r924, %f397;
shfl.sync.bfly.b32 %r926|%p44, %r924, %r909, %r901, %r903;
mov.b32 %f398, %r926;
add.ftz.f32 %f399, %f397, %f398;
mov.b32 %r927, %f399;
shfl.sync.bfly.b32 %r929|%p45, %r927, %r912, %r901, %r903;
mov.b32 %f400, %r929;
add.ftz.f32 %f401, %f399, %f400;
mov.b32 %r930, %f401;
shfl.sync.bfly.b32 %r932|%p46, %r930, %r899, %r901, %r903;
mov.b32 %f402, %r932;
add.ftz.f32 %f54, %f401, %f402;
@%p23 bra $L__BB182_18;

st.global.f32 [%rd17+4], %f54;

$L__BB182_18:
mul.lo.s32 %r36, %r41, 3;
ld.shared.f32 %f403, [%r35+768];
add.ftz.f32 %f404, %f403, %f538;
mov.b32 %r933, %f404;
mov.u32 %r934, 31;
mov.u32 %r935, 16;
mov.u32 %r936, -1;
shfl.sync.bfly.b32 %r937|%p48, %r933, %r935, %r934, %r936;
mov.b32 %f405, %r937;
add.ftz.f32 %f406, %f404, %f405;
mov.b32 %r938, %f406;
mov.u32 %r939, 8;
shfl.sync.bfly.b32 %r940|%p49, %r938, %r939, %r934, %r936;
mov.b32 %f407, %r940;
add.ftz.f32 %f408, %f406, %f407;
mov.b32 %r941, %f408;
mov.u32 %r942, 4;
shfl.sync.bfly.b32 %r943|%p50, %r941, %r942, %r934, %r936;
mov.b32 %f409, %r943;
add.ftz.f32 %f410, %f408, %f409;
mov.b32 %r944, %f410;
mov.u32 %r945, 2;
shfl.sync.bfly.b32 %r946|%p51, %r944, %r945, %r934, %r936;
mov.b32 %f411, %r946;
add.ftz.f32 %f412, %f410, %f411;
mov.b32 %r947, %f412;
mov.u32 %r948, 1;
shfl.sync.bfly.b32 %r949|%p52, %r947, %r948, %r934, %r936;
mov.b32 %f413, %r949;
add.ftz.f32 %f55, %f412, %f413;
@%p15 bra $L__BB182_20;

cvt.u64.u32 %rd102, %r36;
add.s64 %rd103, %rd102, %rd15;
shl.b64 %rd105, %rd103, 2;
add.s64 %rd106, %rd86, %rd105;
st.global.f32 [%rd106], %f55;

$L__BB182_20:
ld.shared.f32 %f414, [%r35+896];
add.ftz.f32 %f415, %f414, %f537;
mov.b32 %r950, %f415;
shfl.sync.bfly.b32 %r954|%p54, %r950, %r935, %r934, %r936;
mov.b32 %f416, %r954;
add.ftz.f32 %f417, %f415, %f416;
mov.b32 %r955, %f417;
shfl.sync.bfly.b32 %r957|%p55, %r955, %r939, %r934, %r936;
mov.b32 %f418, %r957;
add.ftz.f32 %f419, %f417, %f418;
mov.b32 %r958, %f419;
shfl.sync.bfly.b32 %r960|%p56, %r958, %r942, %r934, %r936;
mov.b32 %f420, %r960;
add.ftz.f32 %f421, %f419, %f420;
mov.b32 %r961, %f421;
shfl.sync.bfly.b32 %r963|%p57, %r961, %r945, %r934, %r936;
mov.b32 %f422, %r963;
add.ftz.f32 %f423, %f421, %f422;
mov.b32 %r964, %f423;
shfl.sync.bfly.b32 %r966|%p58, %r964, %r948, %r934, %r936;
mov.b32 %f424, %r966;
add.ftz.f32 %f56, %f423, %f424;
@%p23 bra $L__BB182_22;

add.s32 %r967, %r36, 1;
cvt.u64.u32 %rd107, %r967;
add.s64 %rd108, %rd107, %rd15;
shl.b64 %rd110, %rd108, 2;
add.s64 %rd111, %rd86, %rd110;
st.global.f32 [%rd111], %f56;

$L__BB182_22:
shl.b32 %r968, %r41, 2;
mov.u32 %r969, 2;
ld.shared.f32 %f425, [%r35+1024];
add.ftz.f32 %f426, %f425, %f536;
mov.b32 %r970, %f426;
mov.u32 %r971, 31;
mov.u32 %r972, 16;
mov.u32 %r973, -1;
shfl.sync.bfly.b32 %r974|%p60, %r970, %r972, %r971, %r973;
mov.b32 %f427, %r974;
add.ftz.f32 %f428, %f426, %f427;
mov.b32 %r975, %f428;
mov.u32 %r976, 8;
shfl.sync.bfly.b32 %r977|%p61, %r975, %r976, %r971, %r973;
mov.b32 %f429, %r977;
add.ftz.f32 %f430, %f428, %f429;
mov.b32 %r978, %f430;
mov.u32 %r979, 4;
shfl.sync.bfly.b32 %r980|%p62, %r978, %r979, %r971, %r973;
mov.b32 %f431, %r980;
add.ftz.f32 %f432, %f430, %f431;
mov.b32 %r981, %f432;
shfl.sync.bfly.b32 %r982|%p63, %r981, %r969, %r971, %r973;
mov.b32 %f433, %r982;
add.ftz.f32 %f434, %f432, %f433;
mov.b32 %r983, %f434;
mov.u32 %r984, 1;
shfl.sync.bfly.b32 %r985|%p64, %r983, %r984, %r971, %r973;
mov.b32 %f435, %r985;
add.ftz.f32 %f57, %f434, %f435;
cvt.u64.u32 %rd112, %r968;
add.s64 %rd113, %rd112, %rd15;
shl.b64 %rd115, %rd113, 2;
add.s64 %rd18, %rd86, %rd115;
@%p15 bra $L__BB182_24;

st.global.f32 [%rd18], %f57;

$L__BB182_24:
ld.shared.f32 %f436, [%r35+1152];
add.ftz.f32 %f437, %f436, %f535;
mov.b32 %r986, %f437;
shfl.sync.bfly.b32 %r990|%p66, %r986, %r972, %r971, %r973;
mov.b32 %f438, %r990;
add.ftz.f32 %f439, %f437, %f438;
mov.b32 %r991, %f439;
shfl.sync.bfly.b32 %r993|%p67, %r991, %r976, %r971, %r973;
mov.b32 %f440, %r993;
add.ftz.f32 %f441, %f439, %f440;
mov.b32 %r994, %f441;
shfl.sync.bfly.b32 %r996|%p68, %r994, %r979, %r971, %r973;
mov.b32 %f442, %r996;
add.ftz.f32 %f443, %f441, %f442;
mov.b32 %r997, %f443;
shfl.sync.bfly.b32 %r999|%p69, %r997, %r969, %r971, %r973;
mov.b32 %f444, %r999;
add.ftz.f32 %f445, %f443, %f444;
mov.b32 %r1000, %f445;
shfl.sync.bfly.b32 %r1002|%p70, %r1000, %r984, %r971, %r973;
mov.b32 %f446, %r1002;
add.ftz.f32 %f58, %f445, %f446;
@%p23 bra $L__BB182_26;

st.global.f32 [%rd18+4], %f58;

$L__BB182_26:
mul.lo.s32 %r37, %r41, 5;
ld.shared.f32 %f447, [%r35+1280];
add.ftz.f32 %f448, %f447, %f534;
mov.b32 %r1003, %f448;
mov.u32 %r1004, 31;
mov.u32 %r1005, 16;
mov.u32 %r1006, -1;
shfl.sync.bfly.b32 %r1007|%p72, %r1003, %r1005, %r1004, %r1006;
mov.b32 %f449, %r1007;
add.ftz.f32 %f450, %f448, %f449;
mov.b32 %r1008, %f450;
mov.u32 %r1009, 8;
shfl.sync.bfly.b32 %r1010|%p73, %r1008, %r1009, %r1004, %r1006;
mov.b32 %f451, %r1010;
add.ftz.f32 %f452, %f450, %f451;
mov.b32 %r1011, %f452;
mov.u32 %r1012, 4;
shfl.sync.bfly.b32 %r1013|%p74, %r1011, %r1012, %r1004, %r1006;
mov.b32 %f453, %r1013;
add.ftz.f32 %f454, %f452, %f453;
mov.b32 %r1014, %f454;
mov.u32 %r1015, 2;
shfl.sync.bfly.b32 %r1016|%p75, %r1014, %r1015, %r1004, %r1006;
mov.b32 %f455, %r1016;
add.ftz.f32 %f456, %f454, %f455;
mov.b32 %r1017, %f456;
mov.u32 %r1018, 1;
shfl.sync.bfly.b32 %r1019|%p76, %r1017, %r1018, %r1004, %r1006;
mov.b32 %f457, %r1019;
add.ftz.f32 %f59, %f456, %f457;
@%p15 bra $L__BB182_28;

cvt.u64.u32 %rd116, %r37;
add.s64 %rd117, %rd116, %rd15;
shl.b64 %rd119, %rd117, 2;
add.s64 %rd120, %rd86, %rd119;
st.global.f32 [%rd120], %f59;

$L__BB182_28:
ld.shared.f32 %f458, [%r35+1408];
add.ftz.f32 %f459, %f458, %f533;
mov.b32 %r1020, %f459;
shfl.sync.bfly.b32 %r1024|%p78, %r1020, %r1005, %r1004, %r1006;
mov.b32 %f460, %r1024;
add.ftz.f32 %f461, %f459, %f460;
mov.b32 %r1025, %f461;
shfl.sync.bfly.b32 %r1027|%p79, %r1025, %r1009, %r1004, %r1006;
mov.b32 %f462, %r1027;
add.ftz.f32 %f463, %f461, %f462;
mov.b32 %r1028, %f463;
shfl.sync.bfly.b32 %r1030|%p80, %r1028, %r1012, %r1004, %r1006;
mov.b32 %f464, %r1030;
add.ftz.f32 %f465, %f463, %f464;
mov.b32 %r1031, %f465;
shfl.sync.bfly.b32 %r1033|%p81, %r1031, %r1015, %r1004, %r1006;
mov.b32 %f466, %r1033;
add.ftz.f32 %f467, %f465, %f466;
mov.b32 %r1034, %f467;
shfl.sync.bfly.b32 %r1036|%p82, %r1034, %r1018, %r1004, %r1006;
mov.b32 %f468, %r1036;
add.ftz.f32 %f60, %f467, %f468;
@%p23 bra $L__BB182_30;

add.s32 %r1037, %r37, 1;
cvt.u64.u32 %rd121, %r1037;
add.s64 %rd122, %rd121, %rd15;
shl.b64 %rd124, %rd122, 2;
add.s64 %rd125, %rd86, %rd124;
st.global.f32 [%rd125], %f60;

$L__BB182_30:
mul.lo.s32 %r1038, %r41, 6;
ld.shared.f32 %f469, [%r35+1536];
add.ftz.f32 %f470, %f469, %f532;
mov.b32 %r1039, %f470;
mov.u32 %r1040, 31;
mov.u32 %r1041, 16;
mov.u32 %r1042, -1;
shfl.sync.bfly.b32 %r1043|%p84, %r1039, %r1041, %r1040, %r1042;
mov.b32 %f471, %r1043;
add.ftz.f32 %f472, %f470, %f471;
mov.b32 %r1044, %f472;
mov.u32 %r1045, 8;
shfl.sync.bfly.b32 %r1046|%p85, %r1044, %r1045, %r1040, %r1042;
mov.b32 %f473, %r1046;
add.ftz.f32 %f474, %f472, %f473;
mov.b32 %r1047, %f474;
mov.u32 %r1048, 4;
shfl.sync.bfly.b32 %r1049|%p86, %r1047, %r1048, %r1040, %r1042;
mov.b32 %f475, %r1049;
add.ftz.f32 %f476, %f474, %f475;
mov.b32 %r1050, %f476;
mov.u32 %r1051, 2;
shfl.sync.bfly.b32 %r1052|%p87, %r1050, %r1051, %r1040, %r1042;
mov.b32 %f477, %r1052;
add.ftz.f32 %f478, %f476, %f477;
mov.b32 %r1053, %f478;
mov.u32 %r1054, 1;
shfl.sync.bfly.b32 %r1055|%p88, %r1053, %r1054, %r1040, %r1042;
mov.b32 %f479, %r1055;
add.ftz.f32 %f61, %f478, %f479;
cvt.u64.u32 %rd126, %r1038;
add.s64 %rd127, %rd126, %rd15;
shl.b64 %rd129, %rd127, 2;
add.s64 %rd19, %rd86, %rd129;
@%p15 bra $L__BB182_32;

st.global.f32 [%rd19], %f61;

$L__BB182_32:
ld.shared.f32 %f480, [%r35+1664];
add.ftz.f32 %f481, %f480, %f531;
mov.b32 %r1056, %f481;
shfl.sync.bfly.b32 %r1060|%p90, %r1056, %r1041, %r1040, %r1042;
mov.b32 %f482, %r1060;
add.ftz.f32 %f483, %f481, %f482;
mov.b32 %r1061, %f483;
shfl.sync.bfly.b32 %r1063|%p91, %r1061, %r1045, %r1040, %r1042;
mov.b32 %f484, %r1063;
add.ftz.f32 %f485, %f483, %f484;
mov.b32 %r1064, %f485;
shfl.sync.bfly.b32 %r1066|%p92, %r1064, %r1048, %r1040, %r1042;
mov.b32 %f486, %r1066;
add.ftz.f32 %f487, %f485, %f486;
mov.b32 %r1067, %f487;
shfl.sync.bfly.b32 %r1069|%p93, %r1067, %r1051, %r1040, %r1042;
mov.b32 %f488, %r1069;
add.ftz.f32 %f489, %f487, %f488;
mov.b32 %r1070, %f489;
shfl.sync.bfly.b32 %r1072|%p94, %r1070, %r1054, %r1040, %r1042;
mov.b32 %f490, %r1072;
add.ftz.f32 %f62, %f489, %f490;
@%p23 bra $L__BB182_34;

st.global.f32 [%rd19+4], %f62;

$L__BB182_34:
mul.lo.s32 %r38, %r41, 7;
ld.shared.f32 %f491, [%r35+1792];
add.ftz.f32 %f492, %f491, %f530;
mov.b32 %r1073, %f492;
mov.u32 %r1074, 31;
mov.u32 %r1075, 16;
mov.u32 %r1076, -1;
shfl.sync.bfly.b32 %r1077|%p96, %r1073, %r1075, %r1074, %r1076;
mov.b32 %f493, %r1077;
add.ftz.f32 %f494, %f492, %f493;
mov.b32 %r1078, %f494;
mov.u32 %r1079, 8;
shfl.sync.bfly.b32 %r1080|%p97, %r1078, %r1079, %r1074, %r1076;
mov.b32 %f495, %r1080;
add.ftz.f32 %f496, %f494, %f495;
mov.b32 %r1081, %f496;
mov.u32 %r1082, 4;
shfl.sync.bfly.b32 %r1083|%p98, %r1081, %r1082, %r1074, %r1076;
mov.b32 %f497, %r1083;
add.ftz.f32 %f498, %f496, %f497;
mov.b32 %r1084, %f498;
mov.u32 %r1085, 2;
shfl.sync.bfly.b32 %r1086|%p99, %r1084, %r1085, %r1074, %r1076;
mov.b32 %f499, %r1086;
add.ftz.f32 %f500, %f498, %f499;
mov.b32 %r1087, %f500;
mov.u32 %r1088, 1;
shfl.sync.bfly.b32 %r1089|%p100, %r1087, %r1088, %r1074, %r1076;
mov.b32 %f501, %r1089;
add.ftz.f32 %f63, %f500, %f501;
@%p15 bra $L__BB182_36;

cvt.u64.u32 %rd130, %r38;
add.s64 %rd131, %rd130, %rd15;
shl.b64 %rd133, %rd131, 2;
add.s64 %rd134, %rd86, %rd133;
st.global.f32 [%rd134], %f63;

$L__BB182_36:
ld.shared.f32 %f502, [%r35+1920];
add.ftz.f32 %f503, %f502, %f529;
mov.b32 %r1090, %f503;
shfl.sync.bfly.b32 %r1094|%p102, %r1090, %r1075, %r1074, %r1076;
mov.b32 %f504, %r1094;
add.ftz.f32 %f505, %f503, %f504;
mov.b32 %r1095, %f505;
shfl.sync.bfly.b32 %r1097|%p103, %r1095, %r1079, %r1074, %r1076;
mov.b32 %f506, %r1097;
add.ftz.f32 %f507, %f505, %f506;
mov.b32 %r1098, %f507;
shfl.sync.bfly.b32 %r1100|%p104, %r1098, %r1082, %r1074, %r1076;
mov.b32 %f508, %r1100;
add.ftz.f32 %f509, %f507, %f508;
mov.b32 %r1101, %f509;
shfl.sync.bfly.b32 %r1103|%p105, %r1101, %r1085, %r1074, %r1076;
mov.b32 %f510, %r1103;
add.ftz.f32 %f511, %f509, %f510;
mov.b32 %r1104, %f511;
shfl.sync.bfly.b32 %r1106|%p106, %r1104, %r1088, %r1074, %r1076;
mov.b32 %f512, %r1106;
add.ftz.f32 %f64, %f511, %f512;
@%p23 bra $L__BB182_38;

add.s32 %r1107, %r38, 1;
cvt.u64.u32 %rd135, %r1107;
add.s64 %rd136, %rd135, %rd15;
shl.b64 %rd138, %rd136, 2;
add.s64 %rd139, %rd86, %rd138;
st.global.f32 [%rd139], %f64;

$L__BB182_38:
ret;

}
.entry _Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj(
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2,
.param .align 8 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3[48],
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5,
.param .align 4 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6[12],
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15
)
.maxntid 256, 1, 1
.minnctapersm 1
{
.local .align 8 .b8 __local_depot183[16];
.reg .b64 %SP;
.reg .b64 %SPL;
.reg .pred %p<48>;
.reg .b16 %rs<25>;
.reg .f32 %f<181>;
.reg .b32 %r<355>;
.reg .b64 %rd<120>;


mov.u64 %SPL, __local_depot183;
ld.param.u64 %rd29, [_Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0];
ld.param.u64 %rd30, [_Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1];
ld.param.u64 %rd31, [_Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2];
ld.param.u32 %r30, [_Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5];
ld.param.u32 %r23, [_Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8];
ld.param.u32 %r24, [_Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9];
ld.param.u32 %r26, [_Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11];
ld.param.u32 %r27, [_Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12];
ld.param.u32 %r31, [_Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14];
ld.param.u32 %r29, [_Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15];
ld.param.u32 %r21, [_Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+8];
ld.param.u32 %r20, [_Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+4];
ld.param.u32 %r19, [_Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6];
ld.param.u64 %rd33, [_Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+8];
add.u64 %rd1, %SPL, 0;
add.u64 %rd2, %SPL, 8;
mov.u32 %r32, %ctaid.x;
shl.b32 %r1, %r32, 1;
shr.u32 %r2, %r30, 8;
mov.u32 %r3, %tid.y;
setp.ge.u32 %p1, %r3, %r31;
@%p1 bra $L__BB183_29;

mov.u32 %r4, %ctaid.y;
cvta.to.global.u64 %rd40, %rd31;
mad.lo.s32 %r33, %r3, %r29, %r4;
mul.wide.u32 %rd41, %r33, 4;
add.s64 %rd42, %rd40, %rd41;
ld.global.nc.u32 %r5, [%rd42];
mov.f32 %f174, 0f00000000;
st.local.v2.f32 [%rd1], {%f174, %f174};
st.local.v2.f32 [%rd2], {%f174, %f174};
mov.u32 %r6, %tid.x;
shr.u32 %r354, %r6, 5;
setp.ge.u32 %p2, %r354, %r2;
mov.f32 %f175, %f174;
mov.f32 %f176, %f174;
mov.f32 %f177, %f174;
@%p2 bra $L__BB183_9;

cvta.to.global.u64 %rd43, %rd30;
mul.lo.s32 %r34, %r5, %r26;
and.b32 %r35, %r6, 31;
shr.u32 %r36, %r35, 4;
shl.b32 %r37, %r36, 2;
and.b32 %r38, %r6, 8;
shr.u32 %r39, %r38, 3;
or.b32 %r40, %r37, %r39;
shl.b32 %r41, %r36, 3;
and.b32 %r42, %r6, 12;
shr.u32 %r43, %r42, 2;
or.b32 %r44, %r41, %r43;
shr.u32 %r8, %r38, 2;
and.b32 %r45, %r6, 7;
bfi.b32 %r46, %r36, %r45, 3, 29;
shl.b32 %r47, %r46, 1;
cvt.u64.u32 %rd44, %r44;
cvt.u64.u32 %rd45, %r40;
or.b32 %r48, %r40, 2;
cvt.u64.u32 %rd46, %r48;
shr.u32 %r49, %r6, 2;
cvt.u64.u32 %rd47, %r49;
and.b64 %rd48, %rd47, 1073741816;
mul.lo.s32 %r50, %r3, %r24;
cvt.u64.u32 %rd49, %r50;
add.s64 %rd50, %rd48, %rd49;
mul.hi.u32 %r51, %r4, %r19;
add.s32 %r52, %r51, %r4;
shr.u32 %r53, %r52, %r20;
mul.lo.s32 %r54, %r53, %r21;
sub.s32 %r55, %r4, %r54;
mul.lo.s32 %r56, %r55, %r27;
cvt.u64.u32 %rd51, %r56;
add.s64 %rd52, %rd50, %rd51;
add.s64 %rd53, %rd52, %rd46;
mul.lo.s64 %rd54, %rd53, 36;
add.s64 %rd119, %rd43, %rd54;
add.s64 %rd55, %rd52, %rd45;
mul.lo.s64 %rd56, %rd55, 36;
add.s64 %rd118, %rd43, %rd56;
mul.wide.u32 %rd57, %r35, 4;
mad.lo.s32 %r57, %r1, %r23, %r354;
add.s32 %r58, %r57, %r34;
mul.wide.s32 %rd8, %r58, 210;
add.s64 %rd9, %rd57, %rd8;
add.s32 %r59, %r1, 1;
add.s32 %r60, %r354, %r34;
mad.lo.s32 %r61, %r23, %r59, %r60;
mul.wide.s32 %rd10, %r61, 210;
add.s64 %rd58, %rd10, %rd44;
add.s64 %rd11, %rd58, 192;
mul.wide.s32 %rd59, %r47, 2;
add.s64 %rd60, %rd59, %rd8;
add.s64 %rd12, %rd60, 130;
add.s64 %rd61, %rd59, %rd10;
add.s64 %rd13, %rd61, 128;
add.s64 %rd14, %rd57, %rd10;
add.s64 %rd62, %rd8, %rd44;
add.s64 %rd15, %rd62, 196;
cvta.to.global.u64 %rd117, %rd29;
cvta.to.global.u64 %rd116, %rd33;
mov.f32 %f174, 0f00000000;

$L__BB183_3:
add.s64 %rd63, %rd117, %rd9;
ld.global.nc.u16 %rs2, [%rd63];
ld.global.nc.u16 %rs3, [%rd63+2];
mov.b32 %r80, {%rs2, %rs3};
add.s64 %rd64, %rd117, %rd12;
ld.global.nc.u16 %rs4, [%rd64+-2];
ld.global.nc.u16 %rs5, [%rd64];
mov.b32 %r81, {%rs4, %rs5};
shr.s32 %r82, %r81, %r8;
shl.b32 %r84, %r6, 2;
cvt.u64.u32 %rd65, %r84;
and.b64 %rd66, %rd65, 28;
add.s64 %rd67, %rd66, 4;
add.s64 %rd68, %rd118, %rd67;
ld.global.nc.u32 %r10, [%rd68];
ld.global.nc.u32 %r11, [%rd118];

{.reg .f16 low,high;
mov.b32 {low,high},%r11;
cvt.f32.f16 %f42, low;}


add.s64 %rd69, %rd119, %rd67;
ld.global.nc.u32 %r12, [%rd69];
ld.global.nc.u32 %r13, [%rd119];

{.reg .f16 low,high;
mov.b32 {low,high},%r13;
cvt.f32.f16 %f43, low;}


add.s64 %rd70, %rd117, %rd8;
ld.global.nc.u16 %rs1, [%rd70+208];

{ cvt.f32.f16 %f44, %rs1;}


add.s64 %rd71, %rd117, %rd15;
ld.global.nc.u8 %rs6, [%rd71+-4];
cvt.u32.u16 %r85, %rs6;
cvt.s32.s8 %r86, %r85;
and.b32 %r87, %r80, 252645135;
shl.b32 %r88, %r82, 4;
and.b32 %r89, %r88, 808464432;
or.b32 %r66, %r89, %r87;
or.b32 %r90, %r66, -2139062144;
add.s32 %r91, %r90, -538976288;
xor.b32 %r92, %r91, -2139062144;
xor.b32 %r93, %r66, 538976288;
xor.b32 %r94, %r91, %r66;
and.b32 %r64, %r94, %r93;

prmt.b32 %r64, %r64, 0, 0xba98;


prmt.b32 %r66, %r66, 0, 0xba98;

xor.b32 %r95, %r66, 2139062143;
not.b32 %r96, %r64;
and.b32 %r97, %r92, %r96;
and.b32 %r98, %r95, %r64;
or.b32 %r69, %r98, %r97;
mov.u32 %r79, 0;

dp4a.s32.s32 %r68, %r69, %r10, %r79;

mul.lo.s32 %r99, %r68, %r86;
cvt.rn.f32.s32 %f45, %r99;
fma.rn.ftz.f32 %f46, %f42, %f45, 0f00000000;
ld.global.nc.u8 %rs7, [%rd71];
cvt.u32.u16 %r100, %rs7;
cvt.s32.s8 %r101, %r100;
shr.u32 %r102, %r80, 4;
and.b32 %r103, %r102, 252645135;
and.b32 %r104, %r82, 808464432;
or.b32 %r74, %r104, %r103;
or.b32 %r105, %r74, -2139062144;
add.s32 %r106, %r105, -538976288;
xor.b32 %r107, %r106, -2139062144;
xor.b32 %r108, %r74, 538976288;
xor.b32 %r109, %r106, %r74;
and.b32 %r72, %r109, %r108;

prmt.b32 %r72, %r72, 0, 0xba98;


prmt.b32 %r74, %r74, 0, 0xba98;

xor.b32 %r110, %r74, 2139062143;
not.b32 %r111, %r72;
and.b32 %r112, %r107, %r111;
and.b32 %r113, %r110, %r72;
or.b32 %r77, %r113, %r112;

dp4a.s32.s32 %r76, %r77, %r12, %r79;

mul.lo.s32 %r114, %r76, %r101;
cvt.rn.f32.s32 %f47, %r114;
fma.rn.ftz.f32 %f48, %f43, %f47, %f46;
fma.rn.ftz.f32 %f177, %f44, %f48, %f177;
setp.eq.s64 %p3, %rd33, 0;
@%p3 bra $L__BB183_5;

add.s64 %rd72, %rd116, %rd9;
ld.global.u16 %rs9, [%rd72];
ld.global.u16 %rs10, [%rd72+2];
mov.b32 %r133, {%rs9, %rs10};
add.s64 %rd73, %rd116, %rd12;
ld.global.u16 %rs11, [%rd73+-2];
ld.global.u16 %rs12, [%rd73];
mov.b32 %r134, {%rs11, %rs12};
shr.s32 %r135, %r134, %r8;

{.reg .f16 low,high;
mov.b32 {low,high},%r11;
cvt.f32.f16 %f49, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r13;
cvt.f32.f16 %f50, low;}


add.s64 %rd74, %rd116, %rd8;
ld.global.u16 %rs8, [%rd74+208];

{ cvt.f32.f16 %f51, %rs8;}


add.s64 %rd75, %rd116, %rd15;
ld.global.s8 %r136, [%rd75+-4];
and.b32 %r137, %r133, 252645135;
shl.b32 %r138, %r135, 4;
and.b32 %r139, %r138, 808464432;
or.b32 %r119, %r139, %r137;
or.b32 %r140, %r119, -2139062144;
add.s32 %r141, %r140, -538976288;
xor.b32 %r142, %r141, -2139062144;
xor.b32 %r143, %r119, 538976288;
xor.b32 %r144, %r141, %r119;
and.b32 %r117, %r144, %r143;

prmt.b32 %r117, %r117, 0, 0xba98;


prmt.b32 %r119, %r119, 0, 0xba98;

xor.b32 %r145, %r119, 2139062143;
not.b32 %r146, %r117;
and.b32 %r147, %r142, %r146;
and.b32 %r148, %r145, %r117;
or.b32 %r122, %r148, %r147;

dp4a.s32.s32 %r121, %r122, %r10, %r79;

mul.lo.s32 %r149, %r121, %r136;
cvt.rn.f32.s32 %f52, %r149;
fma.rn.ftz.f32 %f53, %f49, %f52, 0f00000000;
ld.global.s8 %r150, [%rd75];
shr.u32 %r151, %r133, 4;
and.b32 %r152, %r151, 252645135;
and.b32 %r153, %r135, 808464432;
or.b32 %r127, %r153, %r152;
or.b32 %r154, %r127, -2139062144;
add.s32 %r155, %r154, -538976288;
xor.b32 %r156, %r155, -2139062144;
xor.b32 %r157, %r127, 538976288;
xor.b32 %r158, %r155, %r127;
and.b32 %r125, %r158, %r157;

prmt.b32 %r125, %r125, 0, 0xba98;


prmt.b32 %r127, %r127, 0, 0xba98;

xor.b32 %r159, %r127, 2139062143;
not.b32 %r160, %r125;
and.b32 %r161, %r156, %r160;
and.b32 %r162, %r159, %r125;
or.b32 %r130, %r162, %r161;

dp4a.s32.s32 %r129, %r130, %r12, %r79;

mul.lo.s32 %r163, %r129, %r150;
cvt.rn.f32.s32 %f54, %r163;
fma.rn.ftz.f32 %f55, %f50, %f54, %f53;
fma.rn.ftz.f32 %f175, %f51, %f55, %f175;

$L__BB183_5:
add.s64 %rd76, %rd117, %rd14;
ld.global.nc.u16 %rs14, [%rd76];
ld.global.nc.u16 %rs15, [%rd76+2];
mov.b32 %r182, {%rs14, %rs15};
add.s64 %rd77, %rd117, %rd13;
ld.global.nc.u16 %rs16, [%rd77];
ld.global.nc.u16 %rs17, [%rd77+2];
mov.b32 %r183, {%rs16, %rs17};
shr.s32 %r184, %r183, %r8;

{.reg .f16 low,high;
mov.b32 {low,high},%r11;
cvt.f32.f16 %f56, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r13;
cvt.f32.f16 %f57, low;}


add.s64 %rd78, %rd117, %rd10;
ld.global.nc.u16 %rs13, [%rd78+208];

{ cvt.f32.f16 %f58, %rs13;}


add.s64 %rd79, %rd117, %rd11;
ld.global.nc.u8 %rs18, [%rd79];
cvt.u32.u16 %r185, %rs18;
cvt.s32.s8 %r186, %r185;
and.b32 %r187, %r182, 252645135;
shl.b32 %r188, %r184, 4;
and.b32 %r189, %r188, 808464432;
or.b32 %r168, %r189, %r187;
or.b32 %r190, %r168, -2139062144;
add.s32 %r191, %r190, -538976288;
xor.b32 %r192, %r191, -2139062144;
xor.b32 %r193, %r168, 538976288;
xor.b32 %r194, %r191, %r168;
and.b32 %r166, %r194, %r193;

prmt.b32 %r166, %r166, 0, 0xba98;


prmt.b32 %r168, %r168, 0, 0xba98;

xor.b32 %r195, %r168, 2139062143;
not.b32 %r196, %r166;
and.b32 %r197, %r192, %r196;
and.b32 %r198, %r195, %r166;
or.b32 %r171, %r198, %r197;

dp4a.s32.s32 %r170, %r171, %r10, %r79;

mul.lo.s32 %r199, %r170, %r186;
cvt.rn.f32.s32 %f59, %r199;
fma.rn.ftz.f32 %f60, %f56, %f59, 0f00000000;
ld.global.nc.u8 %rs19, [%rd79+4];
cvt.u32.u16 %r200, %rs19;
cvt.s32.s8 %r201, %r200;
shr.u32 %r202, %r182, 4;
and.b32 %r203, %r202, 252645135;
and.b32 %r204, %r184, 808464432;
or.b32 %r176, %r204, %r203;
or.b32 %r205, %r176, -2139062144;
add.s32 %r206, %r205, -538976288;
xor.b32 %r207, %r206, -2139062144;
xor.b32 %r208, %r176, 538976288;
xor.b32 %r209, %r206, %r176;
and.b32 %r174, %r209, %r208;

prmt.b32 %r174, %r174, 0, 0xba98;


prmt.b32 %r176, %r176, 0, 0xba98;

xor.b32 %r210, %r176, 2139062143;
not.b32 %r211, %r174;
and.b32 %r212, %r207, %r211;
and.b32 %r213, %r210, %r174;
or.b32 %r179, %r213, %r212;

dp4a.s32.s32 %r178, %r179, %r12, %r79;

mul.lo.s32 %r214, %r178, %r201;
cvt.rn.f32.s32 %f61, %r214;
fma.rn.ftz.f32 %f62, %f57, %f61, %f60;
fma.rn.ftz.f32 %f176, %f58, %f62, %f176;
@%p3 bra $L__BB183_7;

add.s64 %rd80, %rd116, %rd14;
ld.global.u16 %rs21, [%rd80];
ld.global.u16 %rs22, [%rd80+2];
mov.b32 %r233, {%rs21, %rs22};
add.s64 %rd81, %rd116, %rd13;
ld.global.u16 %rs23, [%rd81];
ld.global.u16 %rs24, [%rd81+2];
mov.b32 %r234, {%rs23, %rs24};
shr.s32 %r235, %r234, %r8;

{.reg .f16 low,high;
mov.b32 {low,high},%r11;
cvt.f32.f16 %f63, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r13;
cvt.f32.f16 %f64, low;}


add.s64 %rd82, %rd116, %rd10;
ld.global.u16 %rs20, [%rd82+208];

{ cvt.f32.f16 %f65, %rs20;}


add.s64 %rd83, %rd116, %rd11;
ld.global.s8 %r236, [%rd83];
and.b32 %r237, %r233, 252645135;
shl.b32 %r238, %r235, 4;
and.b32 %r239, %r238, 808464432;
or.b32 %r219, %r239, %r237;
or.b32 %r240, %r219, -2139062144;
add.s32 %r241, %r240, -538976288;
xor.b32 %r242, %r241, -2139062144;
xor.b32 %r243, %r219, 538976288;
xor.b32 %r244, %r241, %r219;
and.b32 %r217, %r244, %r243;

prmt.b32 %r217, %r217, 0, 0xba98;


prmt.b32 %r219, %r219, 0, 0xba98;

xor.b32 %r245, %r219, 2139062143;
not.b32 %r246, %r217;
and.b32 %r247, %r242, %r246;
and.b32 %r248, %r245, %r217;
or.b32 %r222, %r248, %r247;
mov.u32 %r232, 0;

dp4a.s32.s32 %r221, %r222, %r10, %r232;

mul.lo.s32 %r249, %r221, %r236;
cvt.rn.f32.s32 %f66, %r249;
fma.rn.ftz.f32 %f67, %f63, %f66, 0f00000000;
ld.global.s8 %r250, [%rd83+4];
shr.u32 %r251, %r233, 4;
and.b32 %r252, %r251, 252645135;
and.b32 %r253, %r235, 808464432;
or.b32 %r227, %r253, %r252;
or.b32 %r254, %r227, -2139062144;
add.s32 %r255, %r254, -538976288;
xor.b32 %r256, %r255, -2139062144;
xor.b32 %r257, %r227, 538976288;
xor.b32 %r258, %r255, %r227;
and.b32 %r225, %r258, %r257;

prmt.b32 %r225, %r225, 0, 0xba98;


prmt.b32 %r227, %r227, 0, 0xba98;

xor.b32 %r259, %r227, 2139062143;
not.b32 %r260, %r225;
and.b32 %r261, %r256, %r260;
and.b32 %r262, %r259, %r225;
or.b32 %r230, %r262, %r261;

dp4a.s32.s32 %r229, %r230, %r12, %r232;

mul.lo.s32 %r263, %r229, %r250;
cvt.rn.f32.s32 %f68, %r263;
fma.rn.ftz.f32 %f69, %f64, %f68, %f67;
fma.rn.ftz.f32 %f174, %f65, %f69, %f174;

$L__BB183_7:
add.s64 %rd119, %rd119, 288;
add.s64 %rd118, %rd118, 288;
add.s64 %rd117, %rd117, 210;
add.s64 %rd116, %rd116, 210;
add.s32 %r354, %r354, 1;
setp.lt.u32 %p5, %r354, %r2;
@%p5 bra $L__BB183_3;

add.u64 %rd108, %SPL, 0;
add.u64 %rd102, %SPL, 8;
st.local.v2.f32 [%rd108], {%f177, %f176};
st.local.v2.f32 [%rd102], {%f175, %f174};

$L__BB183_9:
mov.b32 %r15, %f177;
setp.ne.s64 %p6, %rd33, 0;
@%p6 bra $L__BB183_11;
bra.uni $L__BB183_10;

$L__BB183_11:
add.u64 %rd114, %SPL, 0;
add.u64 %rd106, %SPL, 8;
mov.u32 %r290, 31;
mov.u32 %r291, 16;
mov.u32 %r292, -1;
shfl.sync.bfly.b32 %r293|%p17, %r15, %r291, %r290, %r292;
mov.b32 %f90, %r293;
add.ftz.f32 %f91, %f177, %f90;
mov.b32 %r294, %f91;
mov.u32 %r295, 8;
shfl.sync.bfly.b32 %r296|%p18, %r294, %r295, %r290, %r292;
mov.b32 %f92, %r296;
add.ftz.f32 %f93, %f91, %f92;
mov.b32 %r297, %f93;
mov.u32 %r298, 4;
shfl.sync.bfly.b32 %r299|%p19, %r297, %r298, %r290, %r292;
mov.b32 %f94, %r299;
add.ftz.f32 %f95, %f93, %f94;
mov.b32 %r300, %f95;
mov.u32 %r301, 2;
shfl.sync.bfly.b32 %r302|%p20, %r300, %r301, %r290, %r292;
mov.b32 %f96, %r302;
add.ftz.f32 %f97, %f95, %f96;
mov.b32 %r303, %f97;
mov.u32 %r304, 1;
shfl.sync.bfly.b32 %r305|%p21, %r303, %r304, %r290, %r292;
mov.b32 %f98, %r305;
add.ftz.f32 %f99, %f97, %f98;
st.local.f32 [%rd114], %f99;
mov.b32 %r306, %f175;
shfl.sync.bfly.b32 %r307|%p22, %r306, %r291, %r290, %r292;
mov.b32 %f100, %r307;
add.ftz.f32 %f101, %f175, %f100;
mov.b32 %r308, %f101;
shfl.sync.bfly.b32 %r309|%p23, %r308, %r295, %r290, %r292;
mov.b32 %f102, %r309;
add.ftz.f32 %f103, %f101, %f102;
mov.b32 %r310, %f103;
shfl.sync.bfly.b32 %r311|%p24, %r310, %r298, %r290, %r292;
mov.b32 %f104, %r311;
add.ftz.f32 %f105, %f103, %f104;
mov.b32 %r312, %f105;
shfl.sync.bfly.b32 %r313|%p25, %r312, %r301, %r290, %r292;
mov.b32 %f106, %r313;
add.ftz.f32 %f107, %f105, %f106;
mov.b32 %r314, %f107;
shfl.sync.bfly.b32 %r315|%p26, %r314, %r304, %r290, %r292;
mov.b32 %f108, %r315;
add.ftz.f32 %f109, %f107, %f108;
st.local.f32 [%rd106], %f109;
mov.b32 %r316, %f176;
shfl.sync.bfly.b32 %r317|%p27, %r316, %r291, %r290, %r292;
mov.b32 %f110, %r317;
add.ftz.f32 %f111, %f176, %f110;
mov.b32 %r318, %f111;
shfl.sync.bfly.b32 %r319|%p28, %r318, %r295, %r290, %r292;
mov.b32 %f112, %r319;
add.ftz.f32 %f113, %f111, %f112;
mov.b32 %r320, %f113;
shfl.sync.bfly.b32 %r321|%p29, %r320, %r298, %r290, %r292;
mov.b32 %f114, %r321;
add.ftz.f32 %f115, %f113, %f114;
mov.b32 %r322, %f115;
shfl.sync.bfly.b32 %r323|%p30, %r322, %r301, %r290, %r292;
mov.b32 %f116, %r323;
add.ftz.f32 %f117, %f115, %f116;
mov.b32 %r324, %f117;
shfl.sync.bfly.b32 %r325|%p31, %r324, %r304, %r290, %r292;
mov.b32 %f118, %r325;
add.ftz.f32 %f119, %f117, %f118;
st.local.f32 [%rd114+4], %f119;
mov.b32 %r326, %f174;
shfl.sync.bfly.b32 %r327|%p32, %r326, %r291, %r290, %r292;
mov.b32 %f120, %r327;
add.ftz.f32 %f121, %f174, %f120;
mov.b32 %r328, %f121;
shfl.sync.bfly.b32 %r329|%p33, %r328, %r295, %r290, %r292;
mov.b32 %f122, %r329;
add.ftz.f32 %f123, %f121, %f122;
mov.b32 %r330, %f123;
shfl.sync.bfly.b32 %r331|%p34, %r330, %r298, %r290, %r292;
mov.b32 %f124, %r331;
add.ftz.f32 %f125, %f123, %f124;
mov.b32 %r332, %f125;
shfl.sync.bfly.b32 %r333|%p35, %r332, %r301, %r290, %r292;
mov.b32 %f126, %r333;
add.ftz.f32 %f127, %f125, %f126;
mov.b32 %r334, %f127;
shfl.sync.bfly.b32 %r335|%p36, %r334, %r304, %r290, %r292;
mov.b32 %f128, %r335;
add.ftz.f32 %f129, %f127, %f128;
st.local.f32 [%rd106+4], %f129;
bra.uni $L__BB183_12;

$L__BB183_10:
add.u64 %rd110, %SPL, 0;
mov.u32 %r264, 31;
mov.u32 %r265, 16;
mov.u32 %r266, -1;
shfl.sync.bfly.b32 %r267|%p7, %r15, %r265, %r264, %r266;
mov.b32 %f70, %r267;
add.ftz.f32 %f71, %f177, %f70;
mov.b32 %r268, %f71;
mov.u32 %r269, 8;
shfl.sync.bfly.b32 %r270|%p8, %r268, %r269, %r264, %r266;
mov.b32 %f72, %r270;
add.ftz.f32 %f73, %f71, %f72;
mov.b32 %r271, %f73;
mov.u32 %r272, 4;
shfl.sync.bfly.b32 %r273|%p9, %r271, %r272, %r264, %r266;
mov.b32 %f74, %r273;
add.ftz.f32 %f75, %f73, %f74;
mov.b32 %r274, %f75;
mov.u32 %r275, 2;
shfl.sync.bfly.b32 %r276|%p10, %r274, %r275, %r264, %r266;
mov.b32 %f76, %r276;
add.ftz.f32 %f77, %f75, %f76;
mov.b32 %r277, %f77;
mov.u32 %r278, 1;
shfl.sync.bfly.b32 %r279|%p11, %r277, %r278, %r264, %r266;
mov.b32 %f78, %r279;
add.ftz.f32 %f79, %f77, %f78;
st.local.f32 [%rd110], %f79;
mov.b32 %r280, %f176;
shfl.sync.bfly.b32 %r281|%p12, %r280, %r265, %r264, %r266;
mov.b32 %f80, %r281;
add.ftz.f32 %f81, %f176, %f80;
mov.b32 %r282, %f81;
shfl.sync.bfly.b32 %r283|%p13, %r282, %r269, %r264, %r266;
mov.b32 %f82, %r283;
add.ftz.f32 %f83, %f81, %f82;
mov.b32 %r284, %f83;
shfl.sync.bfly.b32 %r285|%p14, %r284, %r272, %r264, %r266;
mov.b32 %f84, %r285;
add.ftz.f32 %f85, %f83, %f84;
mov.b32 %r286, %f85;
shfl.sync.bfly.b32 %r287|%p15, %r286, %r275, %r264, %r266;
mov.b32 %f86, %r287;
add.ftz.f32 %f87, %f85, %f86;
mov.b32 %r288, %f87;
shfl.sync.bfly.b32 %r289|%p16, %r288, %r278, %r264, %r266;
mov.b32 %f88, %r289;
add.ftz.f32 %f89, %f87, %f88;
st.local.f32 [%rd110+4], %f89;

$L__BB183_12:
ld.param.u32 %r343, [_Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7];
mov.u32 %r342, %ctaid.x;
shl.b32 %r341, %r342, 1;
setp.gt.u32 %p37, %r6, 1;
add.s32 %r16, %r341, %r6;
setp.ge.u32 %p38, %r16, %r343;
or.pred %p39, %p37, %p38;
@%p39 bra $L__BB183_29;

add.u64 %rd112, %SPL, 0;
ld.param.u64 %rd97, [_Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3];
ld.param.u32 %r344, [_Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13];
cvt.u64.u32 %rd27, %r6;
mul.wide.u32 %rd84, %r6, 4;
add.s64 %rd85, %rd112, %rd84;
ld.local.f32 %f180, [%rd85];
mad.lo.s32 %r17, %r5, %r344, %r16;
setp.eq.s64 %p40, %rd97, 0;
@%p40 bra $L__BB183_15;

ld.param.u64 %rd98, [_Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3];
cvta.to.global.u64 %rd86, %rd98;
mul.wide.u32 %rd87, %r17, 4;
add.s64 %rd88, %rd86, %rd87;
ld.global.f32 %f130, [%rd88];
add.ftz.f32 %f180, %f180, %f130;

$L__BB183_15:
setp.eq.s64 %p41, %rd33, 0;
@%p41 bra $L__BB183_28;

add.u64 %rd104, %SPL, 8;
ld.param.u64 %rd100, [_Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+16];
shl.b64 %rd89, %rd27, 2;
add.s64 %rd90, %rd104, %rd89;
ld.local.f32 %f24, [%rd90];
setp.eq.s64 %p42, %rd100, 0;
@%p42 bra $L__BB183_18;

ld.param.u64 %rd101, [_Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+16];
cvta.to.global.u64 %rd91, %rd101;
mul.wide.u32 %rd92, %r17, 4;
add.s64 %rd93, %rd91, %rd92;
ld.global.f32 %f131, [%rd93];
add.ftz.f32 %f24, %f24, %f131;

$L__BB183_18:
ld.param.u32 %r349, [_Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.gt.s32 %p43, %r349, 2;
@%p43 bra $L__BB183_22;

ld.param.u32 %r352, [_Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.eq.s32 %p46, %r352, 1;
@%p46 bra $L__BB183_26;

ld.param.u32 %r353, [_Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.eq.s32 %p47, %r353, 2;
@%p47 bra $L__BB183_21;
bra.uni $L__BB183_27;

$L__BB183_21:
mul.ftz.f32 %f151, %f24, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f152, %f151;
add.ftz.f32 %f153, %f152, 0f3F800000;
div.approx.ftz.f32 %f154, %f24, %f153;
mul.ftz.f32 %f180, %f180, %f154;
bra.uni $L__BB183_28;

$L__BB183_22:
ld.param.u32 %r350, [_Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.eq.s32 %p44, %r350, 3;
@%p44 bra $L__BB183_25;

ld.param.u32 %r351, [_Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.ne.s32 %p45, %r351, 6;
@%p45 bra $L__BB183_27;

ld.param.f32 %f163, [_Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+44];
min.ftz.f32 %f132, %f24, %f163;
neg.ftz.f32 %f133, %f163;
min.ftz.f32 %f134, %f180, %f163;
max.ftz.f32 %f135, %f134, %f133;
mul.ftz.f32 %f136, %f132, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f137, %f136;
add.ftz.f32 %f138, %f137, 0f3F800000;
div.approx.ftz.f32 %f139, %f132, %f138;
mul.ftz.f32 %f180, %f135, %f139;
bra.uni $L__BB183_28;

$L__BB183_27:
mul.ftz.f32 %f180, %f180, %f24;
bra.uni $L__BB183_28;

$L__BB183_26:
mul.ftz.f32 %f155, %f24, 0f3F000000;
mul.ftz.f32 %f156, %f24, 0f3D372713;
fma.rn.ftz.f32 %f157, %f24, %f156, 0f3F800000;
mul.ftz.f32 %f158, %f24, 0f3F4C422A;
mul.ftz.f32 %f159, %f158, %f157;
tanh.approx.f32 %f160, %f159;
add.ftz.f32 %f161, %f160, 0f3F800000;
mul.ftz.f32 %f162, %f155, %f161;
mul.ftz.f32 %f180, %f180, %f162;
bra.uni $L__BB183_28;

$L__BB183_25:
mov.f32 %f140, 0f40E00000;
min.ftz.f32 %f141, %f24, %f140;
min.ftz.f32 %f142, %f180, %f140;
mov.f32 %f143, 0fC0E00000;
max.ftz.f32 %f144, %f142, %f143;
mul.ftz.f32 %f145, %f141, 0f3FD9DB23;
mul.ftz.f32 %f146, %f145, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f147, %f146;
add.ftz.f32 %f148, %f147, 0f3F800000;
div.approx.ftz.f32 %f149, %f141, %f148;
add.ftz.f32 %f150, %f144, 0f3F800000;
mul.ftz.f32 %f180, %f150, %f149;

$L__BB183_28:
ld.param.u64 %rd99, [_Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4];
ld.param.u32 %r348, [_Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10];
ld.param.u32 %r347, [_Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13];
mov.u32 %r346, %ctaid.y;
mov.u32 %r345, %tid.y;
mad.lo.s32 %r339, %r345, %r348, %r16;
mad.lo.s32 %r340, %r346, %r347, %r339;
cvta.to.global.u64 %rd94, %rd99;
mul.wide.u32 %rd95, %r340, 4;
add.s64 %rd96, %rd94, %rd95;
st.global.f32 [%rd96], %f180;

$L__BB183_29:
ret;

}
.entry _Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj(
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2,
.param .align 8 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3[48],
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5,
.param .align 4 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6[12],
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15
)
.maxntid 256, 1, 1
.minnctapersm 1
{
.local .align 8 .b8 __local_depot184[8];
.reg .b64 %SP;
.reg .b64 %SPL;
.reg .pred %p<17>;
.reg .b16 %rs<15>;
.reg .f32 %f<50>;
.reg .b32 %r<192>;
.reg .b64 %rd<72>;


mov.u64 %SPL, __local_depot184;
ld.param.u64 %rd19, [_Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0];
ld.param.u64 %rd20, [_Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1];
ld.param.u64 %rd21, [_Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2];
ld.param.u64 %rd22, [_Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4];
ld.param.u32 %r10, [_Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5];
ld.param.u32 %r14, [_Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7];
ld.param.u32 %r15, [_Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8];
ld.param.u32 %r16, [_Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9];
ld.param.u32 %r17, [_Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10];
ld.param.u32 %r18, [_Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11];
ld.param.u32 %r19, [_Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12];
ld.param.u32 %r20, [_Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13];
ld.param.u32 %r22, [_Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14];
ld.param.u32 %r21, [_Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15];
ld.param.u32 %r13, [_Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+8];
ld.param.u32 %r12, [_Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+4];
ld.param.u32 %r11, [_Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6];
add.u64 %rd1, %SPL, 0;
mov.u32 %r23, %ctaid.x;
shl.b32 %r1, %r23, 1;
mov.u32 %r2, %tid.y;
setp.ge.u32 %p1, %r2, %r22;
@%p1 bra $L__BB184_7;

mov.f32 %f48, 0f00000000;
st.local.v2.f32 [%rd1], {%f48, %f48};
mov.u32 %r3, %tid.x;
shr.u32 %r191, %r3, 5;
shr.u32 %r24, %r10, 8;
setp.ge.u32 %p2, %r191, %r24;
mov.f32 %f49, %f48;
@%p2 bra $L__BB184_5;

mov.u32 %r25, %ctaid.y;
mad.lo.s32 %r26, %r2, %r21, %r25;
cvta.to.global.u64 %rd24, %rd21;
mul.wide.u32 %rd25, %r26, 4;
add.s64 %rd26, %rd24, %rd25;
ld.global.nc.u32 %r27, [%rd26];
mul.lo.s32 %r28, %r27, %r18;
and.b32 %r29, %r3, 31;
shr.u32 %r30, %r29, 4;
shl.b32 %r31, %r30, 2;
and.b32 %r32, %r3, 8;
shr.u32 %r33, %r32, 3;
or.b32 %r34, %r31, %r33;
shl.b32 %r35, %r30, 3;
and.b32 %r36, %r3, 12;
shr.u32 %r37, %r36, 2;
or.b32 %r38, %r35, %r37;
and.b32 %r39, %r3, 7;
bfi.b32 %r40, %r30, %r39, 3, 29;
shl.b32 %r41, %r40, 1;
cvt.u64.u32 %rd27, %r38;
cvt.u64.u32 %rd28, %r34;
or.b32 %r42, %r34, 2;
cvt.u64.u32 %rd29, %r42;
shr.u32 %r43, %r3, 2;
cvt.u64.u32 %rd30, %r43;
and.b64 %rd31, %rd30, 1073741816;
mul.lo.s32 %r44, %r2, %r16;
cvt.u64.u32 %rd32, %r44;
add.s64 %rd33, %rd31, %rd32;
mul.hi.u32 %r45, %r25, %r11;
add.s32 %r46, %r45, %r25;
shr.u32 %r47, %r46, %r12;
mul.lo.s32 %r48, %r47, %r13;
sub.s32 %r49, %r25, %r48;
mul.lo.s32 %r50, %r49, %r19;
cvt.u64.u32 %rd34, %r50;
add.s64 %rd35, %rd33, %rd34;
add.s64 %rd36, %rd35, %rd29;
mul.lo.s64 %rd37, %rd36, 36;
cvta.to.global.u64 %rd38, %rd20;
add.s64 %rd71, %rd38, %rd37;
mul.wide.u32 %rd39, %r3, 4;
and.b64 %rd40, %rd39, 28;
add.s64 %rd4, %rd40, 4;
add.s64 %rd41, %rd35, %rd28;
mul.lo.s64 %rd42, %rd41, 36;
add.s64 %rd70, %rd38, %rd42;
mul.wide.u32 %rd43, %r29, 4;
mad.lo.s32 %r51, %r1, %r15, %r191;
add.s32 %r5, %r51, %r28;
mul.wide.s32 %rd44, %r5, 210;
add.s64 %rd45, %rd43, %rd44;
add.s64 %rd6, %rd45, 2;
add.s32 %r52, %r1, 1;
add.s32 %r53, %r191, %r28;
mad.lo.s32 %r6, %r15, %r52, %r53;
mul.wide.s32 %rd46, %r6, 210;
add.s64 %rd47, %rd46, %rd27;
add.s64 %rd7, %rd47, 192;
mul.wide.s32 %rd48, %r41, 2;
add.s64 %rd49, %rd48, %rd44;
add.s64 %rd8, %rd49, 130;
add.s64 %rd50, %rd48, %rd46;
add.s64 %rd9, %rd50, 128;
add.s64 %rd10, %rd43, %rd46;
add.s64 %rd51, %rd44, %rd27;
add.s64 %rd11, %rd51, 192;
cvta.to.global.u64 %rd69, %rd19;
mov.f32 %f48, 0f00000000;

$L__BB184_3:
add.s64 %rd52, %rd69, %rd6;
ld.global.nc.u16 %rs3, [%rd52+-2];
ld.global.nc.u16 %rs4, [%rd52];
mov.b32 %r90, {%rs3, %rs4};
add.s64 %rd53, %rd69, %rd8;
ld.global.nc.u16 %rs5, [%rd53+-2];
ld.global.nc.u16 %rs6, [%rd53];
mov.b32 %r91, {%rs5, %rs6};
shr.u32 %r94, %r32, 2;
shr.s32 %r95, %r91, %r94;
add.s64 %rd54, %rd70, %rd4;
ld.global.nc.u32 %r62, [%rd54];
ld.global.nc.u32 %r54, [%rd70];

{.reg .f16 low,high;
mov.b32 {low,high},%r54;
cvt.f32.f16 %f11, low;}


add.s64 %rd55, %rd71, %rd4;
ld.global.nc.u32 %r70, [%rd55];
ld.global.nc.u32 %r55, [%rd71];

{.reg .f16 low,high;
mov.b32 {low,high},%r55;
cvt.f32.f16 %f12, low;}


add.s64 %rd57, %rd69, %rd44;
ld.global.nc.u16 %rs1, [%rd57+208];

{ cvt.f32.f16 %f13, %rs1;}


add.s64 %rd58, %rd69, %rd11;
ld.global.nc.u8 %rs7, [%rd58];
cvt.u32.u16 %r96, %rs7;
cvt.s32.s8 %r97, %r96;
and.b32 %r98, %r90, 252645135;
shl.b32 %r99, %r95, 4;
and.b32 %r100, %r99, 808464432;
or.b32 %r58, %r100, %r98;
or.b32 %r101, %r58, -2139062144;
add.s32 %r102, %r101, -538976288;
xor.b32 %r103, %r102, -2139062144;
xor.b32 %r104, %r58, 538976288;
xor.b32 %r105, %r102, %r58;
and.b32 %r56, %r105, %r104;

prmt.b32 %r56, %r56, 0, 0xba98;


prmt.b32 %r58, %r58, 0, 0xba98;

xor.b32 %r106, %r58, 2139062143;
not.b32 %r107, %r56;
and.b32 %r108, %r103, %r107;
and.b32 %r109, %r106, %r56;
or.b32 %r61, %r109, %r108;
mov.u32 %r89, 0;

dp4a.s32.s32 %r60, %r61, %r62, %r89;

mul.lo.s32 %r110, %r60, %r97;
cvt.rn.f32.s32 %f17, %r110;
fma.rn.ftz.f32 %f18, %f11, %f17, 0f00000000;
ld.global.nc.u8 %rs8, [%rd58+4];
cvt.u32.u16 %r111, %rs8;
cvt.s32.s8 %r112, %r111;
shr.u32 %r113, %r90, 4;
and.b32 %r114, %r113, 252645135;
and.b32 %r115, %r95, 808464432;
or.b32 %r66, %r115, %r114;
or.b32 %r116, %r66, -2139062144;
add.s32 %r117, %r116, -538976288;
xor.b32 %r118, %r117, -2139062144;
xor.b32 %r119, %r66, 538976288;
xor.b32 %r120, %r117, %r66;
and.b32 %r64, %r120, %r119;

prmt.b32 %r64, %r64, 0, 0xba98;


prmt.b32 %r66, %r66, 0, 0xba98;

xor.b32 %r121, %r66, 2139062143;
not.b32 %r122, %r64;
and.b32 %r123, %r118, %r122;
and.b32 %r124, %r121, %r64;
or.b32 %r69, %r124, %r123;

dp4a.s32.s32 %r68, %r69, %r70, %r89;

mul.lo.s32 %r125, %r68, %r112;
cvt.rn.f32.s32 %f19, %r125;
fma.rn.ftz.f32 %f20, %f12, %f19, %f18;
fma.rn.ftz.f32 %f49, %f13, %f20, %f49;
add.s64 %rd59, %rd69, %rd10;
ld.global.nc.u16 %rs9, [%rd59];
ld.global.nc.u16 %rs10, [%rd59+2];
mov.b32 %r126, {%rs9, %rs10};
add.s64 %rd60, %rd69, %rd9;
ld.global.nc.u16 %rs11, [%rd60];
ld.global.nc.u16 %rs12, [%rd60+2];
mov.b32 %r127, {%rs11, %rs12};
shr.s32 %r128, %r127, %r94;

{.reg .f16 low,high;
mov.b32 {low,high},%r54;
cvt.f32.f16 %f14, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r55;
cvt.f32.f16 %f15, low;}


add.s64 %rd62, %rd69, %rd46;
ld.global.nc.u16 %rs2, [%rd62+208];

{ cvt.f32.f16 %f16, %rs2;}


add.s64 %rd63, %rd69, %rd7;
ld.global.nc.u8 %rs13, [%rd63];
cvt.u32.u16 %r129, %rs13;
cvt.s32.s8 %r130, %r129;
and.b32 %r131, %r126, 252645135;
shl.b32 %r132, %r128, 4;
and.b32 %r133, %r132, 808464432;
or.b32 %r76, %r133, %r131;
or.b32 %r134, %r76, -2139062144;
add.s32 %r135, %r134, -538976288;
xor.b32 %r136, %r135, -2139062144;
xor.b32 %r137, %r76, 538976288;
xor.b32 %r138, %r135, %r76;
and.b32 %r74, %r138, %r137;

prmt.b32 %r74, %r74, 0, 0xba98;


prmt.b32 %r76, %r76, 0, 0xba98;

xor.b32 %r139, %r76, 2139062143;
not.b32 %r140, %r74;
and.b32 %r141, %r136, %r140;
and.b32 %r142, %r139, %r74;
or.b32 %r79, %r142, %r141;

dp4a.s32.s32 %r78, %r79, %r62, %r89;

mul.lo.s32 %r143, %r78, %r130;
cvt.rn.f32.s32 %f21, %r143;
fma.rn.ftz.f32 %f22, %f14, %f21, 0f00000000;
ld.global.nc.u8 %rs14, [%rd63+4];
cvt.u32.u16 %r144, %rs14;
cvt.s32.s8 %r145, %r144;
shr.u32 %r146, %r126, 4;
and.b32 %r147, %r146, 252645135;
and.b32 %r148, %r128, 808464432;
or.b32 %r84, %r148, %r147;
or.b32 %r149, %r84, -2139062144;
add.s32 %r150, %r149, -538976288;
xor.b32 %r151, %r150, -2139062144;
xor.b32 %r152, %r84, 538976288;
xor.b32 %r153, %r150, %r84;
and.b32 %r82, %r153, %r152;

prmt.b32 %r82, %r82, 0, 0xba98;


prmt.b32 %r84, %r84, 0, 0xba98;

xor.b32 %r154, %r84, 2139062143;
not.b32 %r155, %r82;
and.b32 %r156, %r151, %r155;
and.b32 %r157, %r154, %r82;
or.b32 %r87, %r157, %r156;

dp4a.s32.s32 %r86, %r87, %r70, %r89;

mul.lo.s32 %r158, %r86, %r145;
cvt.rn.f32.s32 %f23, %r158;
fma.rn.ftz.f32 %f24, %f15, %f23, %f22;
fma.rn.ftz.f32 %f48, %f16, %f24, %f48;
add.s64 %rd71, %rd71, 288;
add.s64 %rd70, %rd70, 288;
add.s64 %rd69, %rd69, 210;
add.s32 %r191, %r191, 1;
setp.lt.u32 %p3, %r191, %r24;
@%p3 bra $L__BB184_3;

st.local.v2.f32 [%rd1], {%f49, %f48};

$L__BB184_5:
mov.b32 %r160, %f49;
mov.u32 %r161, 31;
mov.u32 %r162, 16;
mov.u32 %r163, -1;
shfl.sync.bfly.b32 %r164|%p4, %r160, %r162, %r161, %r163;
mov.b32 %f25, %r164;
add.ftz.f32 %f26, %f49, %f25;
mov.b32 %r165, %f26;
mov.u32 %r166, 8;
shfl.sync.bfly.b32 %r167|%p5, %r165, %r166, %r161, %r163;
mov.b32 %f27, %r167;
add.ftz.f32 %f28, %f26, %f27;
mov.b32 %r168, %f28;
mov.u32 %r169, 4;
shfl.sync.bfly.b32 %r170|%p6, %r168, %r169, %r161, %r163;
mov.b32 %f29, %r170;
add.ftz.f32 %f30, %f28, %f29;
mov.b32 %r171, %f30;
mov.u32 %r172, 2;
shfl.sync.bfly.b32 %r173|%p7, %r171, %r172, %r161, %r163;
mov.b32 %f31, %r173;
add.ftz.f32 %f32, %f30, %f31;
mov.b32 %r174, %f32;
mov.u32 %r175, 1;
shfl.sync.bfly.b32 %r176|%p8, %r174, %r175, %r161, %r163;
mov.b32 %f33, %r176;
add.ftz.f32 %f34, %f32, %f33;
st.local.f32 [%rd1], %f34;
mov.b32 %r177, %f48;
shfl.sync.bfly.b32 %r178|%p9, %r177, %r162, %r161, %r163;
mov.b32 %f35, %r178;
add.ftz.f32 %f36, %f48, %f35;
mov.b32 %r179, %f36;
shfl.sync.bfly.b32 %r180|%p10, %r179, %r166, %r161, %r163;
mov.b32 %f37, %r180;
add.ftz.f32 %f38, %f36, %f37;
mov.b32 %r181, %f38;
shfl.sync.bfly.b32 %r182|%p11, %r181, %r169, %r161, %r163;
mov.b32 %f39, %r182;
add.ftz.f32 %f40, %f38, %f39;
mov.b32 %r183, %f40;
shfl.sync.bfly.b32 %r184|%p12, %r183, %r172, %r161, %r163;
mov.b32 %f41, %r184;
add.ftz.f32 %f42, %f40, %f41;
mov.b32 %r185, %f42;
shfl.sync.bfly.b32 %r186|%p13, %r185, %r175, %r161, %r163;
mov.b32 %f43, %r186;
add.ftz.f32 %f44, %f42, %f43;
st.local.f32 [%rd1+4], %f44;
add.s32 %r9, %r1, %r3;
setp.ge.u32 %p14, %r9, %r14;
setp.gt.u32 %p15, %r3, 1;
or.pred %p16, %p15, %p14;
@%p16 bra $L__BB184_7;

mul.wide.u32 %rd64, %r3, 4;
add.s64 %rd65, %rd1, %rd64;
ld.local.f32 %f45, [%rd65];
mov.u32 %r187, %ctaid.y;
mad.lo.s32 %r189, %r2, %r17, %r9;
mad.lo.s32 %r190, %r187, %r20, %r189;
cvta.to.global.u64 %rd66, %rd22;
mul.wide.u32 %rd67, %r190, 4;
add.s64 %rd68, %rd66, %rd67;
st.global.f32 [%rd68], %f45;

$L__BB184_7:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<320>;
.reg .b16 %rs<49>;
.reg .f32 %f<1029>;
.reg .b32 %r<1282>;
.reg .b64 %rd<140>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate[1536];

ld.param.u64 %rd39, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd40, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd41, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r46, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r47, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r51, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r52, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r53, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r57, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r58, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r59, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r55, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r54, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r49, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r48, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r45, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
ld.param.u64 %rd44, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+16];
ld.param.u64 %rd43, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+8];
ld.param.u64 %rd42, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3];
mov.u32 %r1, %tid.y;
shl.b32 %r60, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r60, %r2;
mov.u32 %r61, %ctaid.x;
shl.b32 %r4, %r61, 2;
mov.u32 %r5, %ctaid.y;
setp.eq.s64 %p1, %rd41, 0;
@%p1 bra $L__BB185_2;

cvta.to.global.u64 %rd48, %rd41;
mul.wide.u32 %rd49, %r5, 4;
add.s64 %rd50, %rd48, %rd49;
ld.global.nc.u32 %r1278, [%rd50];
mul.hi.u32 %r62, %r5, %r43;
add.s32 %r63, %r62, %r5;
shr.u32 %r64, %r63, %r44;
mul.lo.s32 %r65, %r64, %r45;
sub.s32 %r1279, %r5, %r65;
bra.uni $L__BB185_3;

$L__BB185_2:
mul.hi.u32 %r66, %r5, %r48;
add.s32 %r67, %r66, %r5;
shr.u32 %r1278, %r67, %r49;
mov.u32 %r1279, %r5;

$L__BB185_3:
mov.u32 %r11, %ctaid.z;
selp.b32 %r12, %r5, %r1278, %p1;
setp.ne.s32 %p3, %r1, 0;
setp.gt.u32 %p4, %r2, 3;
or.pred %p5, %p4, %p3;
add.s32 %r68, %r4, %r2;
setp.ge.u32 %p6, %r68, %r47;
mov.f32 %f27, 0f00000000;
or.pred %p7, %p5, %p6;
mov.f32 %f4, %f27;
mov.f32 %f5, %f27;
@%p7 bra $L__BB185_8;

setp.eq.s64 %p8, %rd42, 0;
mov.f32 %f4, 0f00000000;
mov.f32 %f5, %f4;
@%p8 bra $L__BB185_6;

cvta.to.global.u64 %rd51, %rd42;
mul.lo.s32 %r69, %r11, %r59;
cvt.u64.u32 %rd52, %r69;
mul.lo.s32 %r70, %r12, %r53;
cvt.u64.u32 %rd53, %r70;
cvt.u64.u32 %rd54, %r2;
cvt.s64.s32 %rd55, %r4;
add.s64 %rd56, %rd55, %rd54;
add.s64 %rd57, %rd56, %rd52;
add.s64 %rd58, %rd57, %rd53;
shl.b64 %rd59, %rd58, 2;
add.s64 %rd60, %rd51, %rd59;
ld.global.f32 %f5, [%rd60];

$L__BB185_6:
setp.eq.s64 %p9, %rd44, 0;
setp.eq.s64 %p10, %rd43, 0;
or.pred %p11, %p9, %p10;
@%p11 bra $L__BB185_8;

cvta.to.global.u64 %rd61, %rd44;
mul.lo.s32 %r71, %r11, %r59;
cvt.u64.u32 %rd62, %r71;
mul.lo.s32 %r72, %r12, %r53;
cvt.u64.u32 %rd63, %r72;
cvt.u64.u32 %rd64, %r2;
cvt.s64.s32 %rd65, %r4;
add.s64 %rd66, %rd65, %rd64;
add.s64 %rd67, %rd66, %rd62;
add.s64 %rd68, %rd67, %rd63;
shl.b64 %rd69, %rd68, 2;
add.s64 %rd70, %rd61, %rd69;
ld.global.f32 %f4, [%rd70];

$L__BB185_8:
mul.lo.s32 %r73, %r11, %r58;
cvt.u64.u32 %rd71, %r73;
mul.lo.s32 %r74, %r1279, %r52;
cvt.u64.u32 %rd72, %r74;
add.s64 %rd4, %rd72, %rd71;
mul.hi.u32 %r75, %r11, %r54;
add.s32 %r76, %r75, %r11;
shr.u32 %r13, %r76, %r55;
shr.s32 %r77, %r3, 31;
shr.u32 %r78, %r77, 27;
add.s32 %r79, %r3, %r78;
shr.s32 %r1281, %r79, 5;
shr.u32 %r80, %r42, 8;
setp.ge.s32 %p12, %r1281, %r80;
mov.f32 %f28, %f27;
mov.f32 %f29, %f27;
mov.f32 %f30, %f27;
mov.f32 %f31, %f27;
mov.f32 %f32, %f27;
mov.f32 %f33, %f27;
mov.f32 %f34, %f27;
@%p12 bra $L__BB185_19;

mul.lo.s32 %r81, %r13, %r57;
mul.lo.s32 %r82, %r1278, %r51;
and.b32 %r86, %r79, -32;
sub.s32 %r87, %r3, %r86;
shr.s32 %r88, %r87, 31;
shr.u32 %r89, %r88, 28;
add.s32 %r90, %r87, %r89;
shr.s32 %r91, %r90, 4;
shl.b32 %r92, %r91, 2;
and.b32 %r93, %r90, -16;
sub.s32 %r94, %r87, %r93;
shr.s32 %r95, %r94, 31;
shr.u32 %r96, %r95, 29;
add.s32 %r97, %r94, %r96;
shr.s32 %r98, %r97, 3;
add.s32 %r99, %r92, %r98;
shl.b32 %r100, %r91, 3;
shr.u32 %r101, %r95, 30;
add.s32 %r102, %r94, %r101;
shr.s32 %r103, %r102, 2;
add.s32 %r104, %r100, %r103;
shl.b32 %r15, %r98, 1;
shl.b32 %r105, %r87, 1;
shr.u32 %r106, %r88, 29;
add.s32 %r107, %r87, %r106;
and.b32 %r108, %r107, -8;
sub.s32 %r109, %r87, %r108;
add.s32 %r110, %r100, %r109;
shl.b32 %r111, %r110, 1;
cvt.s64.s32 %rd73, %r104;
cvt.s64.s32 %rd5, %r109;
cvt.s64.s32 %rd6, %r99;
shl.b32 %r1280, %r1281, 3;
mad.lo.s32 %r112, %r4, %r46, %r1281;
add.s32 %r113, %r112, %r81;
add.s32 %r114, %r113, %r82;
mul.wide.s32 %rd7, %r114, 210;
mul.wide.s32 %rd74, %r105, 2;
add.s64 %rd75, %rd7, %rd74;
add.s64 %rd8, %rd75, 2;
add.s32 %r115, %r1281, %r81;
add.s32 %r17, %r115, %r82;
add.s32 %r116, %r4, 1;
mad.lo.s32 %r117, %r46, %r116, %r17;
mul.wide.s32 %rd76, %r117, 210;
add.s64 %rd77, %rd76, %rd73;
add.s64 %rd9, %rd77, 192;
mul.wide.s32 %rd78, %r111, 2;
add.s64 %rd79, %rd7, %rd78;
add.s64 %rd10, %rd79, 130;
add.s64 %rd80, %rd76, %rd78;
add.s64 %rd11, %rd80, 128;
add.s64 %rd12, %rd76, %rd74;
add.s64 %rd81, %rd7, %rd73;
add.s64 %rd13, %rd81, 196;
add.s32 %r118, %r4, 2;
mad.lo.s32 %r119, %r46, %r118, %r17;
mul.wide.s32 %rd82, %r119, 210;
add.s64 %rd83, %rd82, %rd73;
add.s64 %rd14, %rd83, 192;
add.s64 %rd84, %rd82, %rd78;
add.s64 %rd15, %rd84, 128;
add.s32 %r120, %r4, 3;
mad.lo.s32 %r121, %r46, %r120, %r17;
mul.wide.s32 %rd85, %r121, 210;
add.s64 %rd86, %rd85, %rd74;
add.s64 %rd16, %rd86, 2;
add.s64 %rd17, %rd82, %rd74;
add.s64 %rd18, %rd85, %rd78;
add.s64 %rd19, %rd85, %rd73;
cvta.to.global.u64 %rd138, %rd43;
cvta.to.global.u64 %rd139, %rd39;
mov.f32 %f27, 0f00000000;
cvta.to.global.u64 %rd87, %rd40;
shl.b64 %rd95, %rd5, 2;
mov.f32 %f28, %f27;
mov.f32 %f29, %f27;
mov.f32 %f30, %f27;
mov.f32 %f31, %f27;
mov.f32 %f32, %f27;
mov.f32 %f33, %f27;
mov.f32 %f34, %f27;

$L__BB185_10:
cvt.u64.u32 %rd88, %r1280;
add.s64 %rd89, %rd4, %rd88;
add.s64 %rd90, %rd139, %rd8;
ld.global.nc.u16 %rs2, [%rd90+-2];
ld.global.nc.u16 %rs3, [%rd90];
mov.b32 %r140, {%rs2, %rs3};
add.s64 %rd91, %rd139, %rd10;
ld.global.nc.u16 %rs4, [%rd91+-2];
ld.global.nc.u16 %rs5, [%rd91];
mov.b32 %r141, {%rs4, %rs5};
shr.s32 %r142, %r141, %r15;
add.s64 %rd92, %rd89, %rd6;
mul.lo.s64 %rd93, %rd92, 36;
add.s64 %rd94, %rd87, %rd93;
add.s64 %rd96, %rd94, %rd95;
ld.global.nc.u32 %r20, [%rd96+4];
ld.global.nc.u32 %r21, [%rd94];

{.reg .f16 low,high;
mov.b32 {low,high},%r21;
cvt.f32.f16 %f101, low;}


ld.global.nc.u32 %r22, [%rd96+76];
ld.global.nc.u32 %r23, [%rd94+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r23;
cvt.f32.f16 %f102, low;}


add.s64 %rd97, %rd139, %rd7;
ld.global.nc.u16 %rs1, [%rd97+208];

{ cvt.f32.f16 %f103, %rs1;}


add.s64 %rd98, %rd139, %rd13;
ld.global.nc.u8 %rs6, [%rd98+-4];
cvt.u32.u16 %r143, %rs6;
cvt.s32.s8 %r144, %r143;
and.b32 %r145, %r140, 252645135;
shl.b32 %r146, %r142, 4;
and.b32 %r147, %r146, 808464432;
or.b32 %r126, %r147, %r145;
or.b32 %r148, %r126, -2139062144;
add.s32 %r149, %r148, -538976288;
xor.b32 %r150, %r149, -2139062144;
xor.b32 %r151, %r126, 538976288;
xor.b32 %r152, %r149, %r126;
and.b32 %r124, %r152, %r151;

prmt.b32 %r124, %r124, 0, 0xba98;


prmt.b32 %r126, %r126, 0, 0xba98;

xor.b32 %r153, %r126, 2139062143;
not.b32 %r154, %r124;
and.b32 %r155, %r150, %r154;
and.b32 %r156, %r153, %r124;
or.b32 %r129, %r156, %r155;
mov.u32 %r139, 0;

dp4a.s32.s32 %r128, %r129, %r20, %r139;

mul.lo.s32 %r157, %r128, %r144;
cvt.rn.f32.s32 %f104, %r157;
fma.rn.ftz.f32 %f105, %f101, %f104, 0f00000000;
ld.global.nc.u8 %rs7, [%rd98];
cvt.u32.u16 %r158, %rs7;
cvt.s32.s8 %r159, %r158;
shr.u32 %r160, %r140, 4;
and.b32 %r161, %r160, 252645135;
and.b32 %r162, %r142, 808464432;
or.b32 %r134, %r162, %r161;
or.b32 %r163, %r134, -2139062144;
add.s32 %r164, %r163, -538976288;
xor.b32 %r165, %r164, -2139062144;
xor.b32 %r166, %r134, 538976288;
xor.b32 %r167, %r164, %r134;
and.b32 %r132, %r167, %r166;

prmt.b32 %r132, %r132, 0, 0xba98;


prmt.b32 %r134, %r134, 0, 0xba98;

xor.b32 %r168, %r134, 2139062143;
not.b32 %r169, %r132;
and.b32 %r170, %r165, %r169;
and.b32 %r171, %r168, %r132;
or.b32 %r137, %r171, %r170;

dp4a.s32.s32 %r136, %r137, %r22, %r139;

mul.lo.s32 %r172, %r136, %r159;
cvt.rn.f32.s32 %f106, %r172;
fma.rn.ftz.f32 %f107, %f102, %f106, %f105;
fma.rn.ftz.f32 %f30, %f103, %f107, %f30;
setp.eq.s64 %p13, %rd43, 0;
@%p13 bra $L__BB185_12;

add.s64 %rd99, %rd138, %rd8;
ld.global.u16 %rs9, [%rd99+-2];
ld.global.u16 %rs10, [%rd99];
mov.b32 %r191, {%rs9, %rs10};
add.s64 %rd100, %rd138, %rd10;
ld.global.u16 %rs11, [%rd100+-2];
ld.global.u16 %rs12, [%rd100];
mov.b32 %r192, {%rs11, %rs12};
shr.s32 %r193, %r192, %r15;

{.reg .f16 low,high;
mov.b32 {low,high},%r21;
cvt.f32.f16 %f108, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r23;
cvt.f32.f16 %f109, low;}


add.s64 %rd101, %rd138, %rd7;
ld.global.u16 %rs8, [%rd101+208];

{ cvt.f32.f16 %f110, %rs8;}


add.s64 %rd102, %rd138, %rd13;
ld.global.s8 %r194, [%rd102+-4];
and.b32 %r195, %r191, 252645135;
shl.b32 %r196, %r193, 4;
and.b32 %r197, %r196, 808464432;
or.b32 %r177, %r197, %r195;
or.b32 %r198, %r177, -2139062144;
add.s32 %r199, %r198, -538976288;
xor.b32 %r200, %r199, -2139062144;
xor.b32 %r201, %r177, 538976288;
xor.b32 %r202, %r199, %r177;
and.b32 %r175, %r202, %r201;

prmt.b32 %r175, %r175, 0, 0xba98;


prmt.b32 %r177, %r177, 0, 0xba98;

xor.b32 %r203, %r177, 2139062143;
not.b32 %r204, %r175;
and.b32 %r205, %r200, %r204;
and.b32 %r206, %r203, %r175;
or.b32 %r180, %r206, %r205;

dp4a.s32.s32 %r179, %r180, %r20, %r139;

mul.lo.s32 %r207, %r179, %r194;
cvt.rn.f32.s32 %f111, %r207;
fma.rn.ftz.f32 %f112, %f108, %f111, 0f00000000;
ld.global.s8 %r208, [%rd102];
shr.u32 %r209, %r191, 4;
and.b32 %r210, %r209, 252645135;
and.b32 %r211, %r193, 808464432;
or.b32 %r185, %r211, %r210;
or.b32 %r212, %r185, -2139062144;
add.s32 %r213, %r212, -538976288;
xor.b32 %r214, %r213, -2139062144;
xor.b32 %r215, %r185, 538976288;
xor.b32 %r216, %r213, %r185;
and.b32 %r183, %r216, %r215;

prmt.b32 %r183, %r183, 0, 0xba98;


prmt.b32 %r185, %r185, 0, 0xba98;

xor.b32 %r217, %r185, 2139062143;
not.b32 %r218, %r183;
and.b32 %r219, %r214, %r218;
and.b32 %r220, %r217, %r183;
or.b32 %r188, %r220, %r219;

dp4a.s32.s32 %r187, %r188, %r22, %r139;

mul.lo.s32 %r221, %r187, %r208;
cvt.rn.f32.s32 %f113, %r221;
fma.rn.ftz.f32 %f114, %f109, %f113, %f112;
fma.rn.ftz.f32 %f34, %f110, %f114, %f34;

$L__BB185_12:
add.s64 %rd137, %rd139, %rd13;
add.s64 %rd136, %rd137, -4;
add.s64 %rd135, %rd139, %rd7;
add.s64 %rd134, %rd135, 208;
add.s64 %rd133, %rd139, %rd10;
add.s64 %rd132, %rd133, -2;
add.s64 %rd131, %rd139, %rd8;
add.s64 %rd130, %rd131, -2;
mul.wide.s32 %rd103, %r46, 210;
add.s64 %rd28, %rd130, %rd103;
ld.global.nc.u16 %rs14, [%rd28];
ld.global.nc.u16 %rs15, [%rd28+2];
mov.b32 %r240, {%rs14, %rs15};
add.s64 %rd29, %rd132, %rd103;
ld.global.nc.u16 %rs16, [%rd29];
ld.global.nc.u16 %rs17, [%rd29+2];
mov.b32 %r241, {%rs16, %rs17};
shr.s32 %r242, %r241, %r15;

{.reg .f16 low,high;
mov.b32 {low,high},%r21;
cvt.f32.f16 %f115, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r23;
cvt.f32.f16 %f116, low;}


add.s64 %rd30, %rd134, %rd103;
ld.global.nc.u16 %rs13, [%rd30];

{ cvt.f32.f16 %f117, %rs13;}


add.s64 %rd31, %rd136, %rd103;
ld.global.nc.u8 %rs18, [%rd31];
cvt.u32.u16 %r243, %rs18;
cvt.s32.s8 %r244, %r243;
and.b32 %r245, %r240, 252645135;
shl.b32 %r246, %r242, 4;
and.b32 %r247, %r246, 808464432;
or.b32 %r226, %r247, %r245;
or.b32 %r248, %r226, -2139062144;
add.s32 %r249, %r248, -538976288;
xor.b32 %r250, %r249, -2139062144;
xor.b32 %r251, %r226, 538976288;
xor.b32 %r252, %r249, %r226;
and.b32 %r224, %r252, %r251;

prmt.b32 %r224, %r224, 0, 0xba98;


prmt.b32 %r226, %r226, 0, 0xba98;

xor.b32 %r253, %r226, 2139062143;
not.b32 %r254, %r224;
and.b32 %r255, %r250, %r254;
and.b32 %r256, %r253, %r224;
or.b32 %r229, %r256, %r255;

dp4a.s32.s32 %r228, %r229, %r20, %r139;

mul.lo.s32 %r257, %r228, %r244;
cvt.rn.f32.s32 %f118, %r257;
fma.rn.ftz.f32 %f119, %f115, %f118, 0f00000000;
ld.global.nc.u8 %rs19, [%rd31+4];
cvt.u32.u16 %r258, %rs19;
cvt.s32.s8 %r259, %r258;
shr.u32 %r260, %r240, 4;
and.b32 %r261, %r260, 252645135;
and.b32 %r262, %r242, 808464432;
or.b32 %r234, %r262, %r261;
or.b32 %r263, %r234, -2139062144;
add.s32 %r264, %r263, -538976288;
xor.b32 %r265, %r264, -2139062144;
xor.b32 %r266, %r234, 538976288;
xor.b32 %r267, %r264, %r234;
and.b32 %r232, %r267, %r266;

prmt.b32 %r232, %r232, 0, 0xba98;


prmt.b32 %r234, %r234, 0, 0xba98;

xor.b32 %r268, %r234, 2139062143;
not.b32 %r269, %r232;
and.b32 %r270, %r265, %r269;
and.b32 %r271, %r268, %r232;
or.b32 %r237, %r271, %r270;

dp4a.s32.s32 %r236, %r237, %r22, %r139;

mul.lo.s32 %r272, %r236, %r259;
cvt.rn.f32.s32 %f120, %r272;
fma.rn.ftz.f32 %f121, %f116, %f120, %f119;
fma.rn.ftz.f32 %f29, %f117, %f121, %f29;
@%p13 bra $L__BB185_14;

add.s64 %rd104, %rd138, %rd12;
ld.global.u16 %rs21, [%rd104];
ld.global.u16 %rs22, [%rd104+2];
mov.b32 %r291, {%rs21, %rs22};
add.s64 %rd105, %rd138, %rd11;
ld.global.u16 %rs23, [%rd105];
ld.global.u16 %rs24, [%rd105+2];
mov.b32 %r292, {%rs23, %rs24};
shr.s32 %r293, %r292, %r15;

{.reg .f16 low,high;
mov.b32 {low,high},%r21;
cvt.f32.f16 %f122, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r23;
cvt.f32.f16 %f123, low;}


or.b32 %r296, %r4, 1;
mad.lo.s32 %r297, %r46, %r296, %r17;
mul.wide.s32 %rd106, %r297, 210;
add.s64 %rd107, %rd138, %rd106;
ld.global.u16 %rs20, [%rd107+208];

{ cvt.f32.f16 %f124, %rs20;}


add.s64 %rd108, %rd138, %rd9;
ld.global.s8 %r298, [%rd108];
and.b32 %r299, %r291, 252645135;
shl.b32 %r300, %r293, 4;
and.b32 %r301, %r300, 808464432;
or.b32 %r277, %r301, %r299;
or.b32 %r302, %r277, -2139062144;
add.s32 %r303, %r302, -538976288;
xor.b32 %r304, %r303, -2139062144;
xor.b32 %r305, %r277, 538976288;
xor.b32 %r306, %r303, %r277;
and.b32 %r275, %r306, %r305;

prmt.b32 %r275, %r275, 0, 0xba98;


prmt.b32 %r277, %r277, 0, 0xba98;

xor.b32 %r307, %r277, 2139062143;
not.b32 %r308, %r275;
and.b32 %r309, %r304, %r308;
and.b32 %r310, %r307, %r275;
or.b32 %r280, %r310, %r309;
mov.u32 %r290, 0;

dp4a.s32.s32 %r279, %r280, %r20, %r290;

mul.lo.s32 %r311, %r279, %r298;
cvt.rn.f32.s32 %f125, %r311;
fma.rn.ftz.f32 %f126, %f122, %f125, 0f00000000;
ld.global.s8 %r312, [%rd108+4];
shr.u32 %r313, %r291, 4;
and.b32 %r314, %r313, 252645135;
and.b32 %r315, %r293, 808464432;
or.b32 %r285, %r315, %r314;
or.b32 %r316, %r285, -2139062144;
add.s32 %r317, %r316, -538976288;
xor.b32 %r318, %r317, -2139062144;
xor.b32 %r319, %r285, 538976288;
xor.b32 %r320, %r317, %r285;
and.b32 %r283, %r320, %r319;

prmt.b32 %r283, %r283, 0, 0xba98;


prmt.b32 %r285, %r285, 0, 0xba98;

xor.b32 %r321, %r285, 2139062143;
not.b32 %r322, %r283;
and.b32 %r323, %r318, %r322;
and.b32 %r324, %r321, %r283;
or.b32 %r288, %r324, %r323;

dp4a.s32.s32 %r287, %r288, %r22, %r290;

mul.lo.s32 %r325, %r287, %r312;
cvt.rn.f32.s32 %f127, %r325;
fma.rn.ftz.f32 %f128, %f123, %f127, %f126;
fma.rn.ftz.f32 %f33, %f124, %f128, %f33;

$L__BB185_14:
mul.wide.s32 %rd129, %r46, 210;
add.s64 %rd32, %rd28, %rd129;
ld.global.nc.u16 %rs26, [%rd32];
ld.global.nc.u16 %rs27, [%rd32+2];
mov.b32 %r344, {%rs26, %rs27};
add.s64 %rd33, %rd29, %rd129;
ld.global.nc.u16 %rs28, [%rd33];
ld.global.nc.u16 %rs29, [%rd33+2];
mov.b32 %r345, {%rs28, %rs29};
shr.s32 %r346, %r345, %r15;

{.reg .f16 low,high;
mov.b32 {low,high},%r21;
cvt.f32.f16 %f129, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r23;
cvt.f32.f16 %f130, low;}


add.s64 %rd34, %rd30, %rd129;
ld.global.nc.u16 %rs25, [%rd34];

{ cvt.f32.f16 %f131, %rs25;}


add.s64 %rd35, %rd31, %rd129;
ld.global.nc.u8 %rs30, [%rd35];
cvt.u32.u16 %r347, %rs30;
cvt.s32.s8 %r348, %r347;
and.b32 %r349, %r344, 252645135;
shl.b32 %r350, %r346, 4;
and.b32 %r351, %r350, 808464432;
or.b32 %r330, %r351, %r349;
or.b32 %r352, %r330, -2139062144;
add.s32 %r353, %r352, -538976288;
xor.b32 %r354, %r353, -2139062144;
xor.b32 %r355, %r330, 538976288;
xor.b32 %r356, %r353, %r330;
and.b32 %r328, %r356, %r355;

prmt.b32 %r328, %r328, 0, 0xba98;


prmt.b32 %r330, %r330, 0, 0xba98;

xor.b32 %r357, %r330, 2139062143;
not.b32 %r358, %r328;
and.b32 %r359, %r354, %r358;
and.b32 %r360, %r357, %r328;
or.b32 %r333, %r360, %r359;
mov.u32 %r343, 0;

dp4a.s32.s32 %r332, %r333, %r20, %r343;

mul.lo.s32 %r361, %r332, %r348;
cvt.rn.f32.s32 %f132, %r361;
fma.rn.ftz.f32 %f133, %f129, %f132, 0f00000000;
ld.global.nc.u8 %rs31, [%rd35+4];
cvt.u32.u16 %r362, %rs31;
cvt.s32.s8 %r363, %r362;
shr.u32 %r364, %r344, 4;
and.b32 %r365, %r364, 252645135;
and.b32 %r366, %r346, 808464432;
or.b32 %r338, %r366, %r365;
or.b32 %r367, %r338, -2139062144;
add.s32 %r368, %r367, -538976288;
xor.b32 %r369, %r368, -2139062144;
xor.b32 %r370, %r338, 538976288;
xor.b32 %r371, %r368, %r338;
and.b32 %r336, %r371, %r370;

prmt.b32 %r336, %r336, 0, 0xba98;


prmt.b32 %r338, %r338, 0, 0xba98;

xor.b32 %r372, %r338, 2139062143;
not.b32 %r373, %r336;
and.b32 %r374, %r369, %r373;
and.b32 %r375, %r372, %r336;
or.b32 %r341, %r375, %r374;

dp4a.s32.s32 %r340, %r341, %r22, %r343;

mul.lo.s32 %r376, %r340, %r363;
cvt.rn.f32.s32 %f134, %r376;
fma.rn.ftz.f32 %f135, %f130, %f134, %f133;
fma.rn.ftz.f32 %f28, %f131, %f135, %f28;
@%p13 bra $L__BB185_16;

add.s64 %rd110, %rd138, %rd17;
ld.global.u16 %rs33, [%rd110];
ld.global.u16 %rs34, [%rd110+2];
mov.b32 %r395, {%rs33, %rs34};
add.s64 %rd111, %rd138, %rd15;
ld.global.u16 %rs35, [%rd111];
ld.global.u16 %rs36, [%rd111+2];
mov.b32 %r396, {%rs35, %rs36};
shr.s32 %r397, %r396, %r15;

{.reg .f16 low,high;
mov.b32 {low,high},%r21;
cvt.f32.f16 %f136, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r23;
cvt.f32.f16 %f137, low;}


or.b32 %r400, %r4, 2;
mad.lo.s32 %r401, %r46, %r400, %r17;
mul.wide.s32 %rd112, %r401, 210;
add.s64 %rd113, %rd138, %rd112;
ld.global.u16 %rs32, [%rd113+208];

{ cvt.f32.f16 %f138, %rs32;}


add.s64 %rd114, %rd138, %rd14;
ld.global.s8 %r402, [%rd114];
and.b32 %r403, %r395, 252645135;
shl.b32 %r404, %r397, 4;
and.b32 %r405, %r404, 808464432;
or.b32 %r381, %r405, %r403;
or.b32 %r406, %r381, -2139062144;
add.s32 %r407, %r406, -538976288;
xor.b32 %r408, %r407, -2139062144;
xor.b32 %r409, %r381, 538976288;
xor.b32 %r410, %r407, %r381;
and.b32 %r379, %r410, %r409;

prmt.b32 %r379, %r379, 0, 0xba98;


prmt.b32 %r381, %r381, 0, 0xba98;

xor.b32 %r411, %r381, 2139062143;
not.b32 %r412, %r379;
and.b32 %r413, %r408, %r412;
and.b32 %r414, %r411, %r379;
or.b32 %r384, %r414, %r413;

dp4a.s32.s32 %r383, %r384, %r20, %r343;

mul.lo.s32 %r415, %r383, %r402;
cvt.rn.f32.s32 %f139, %r415;
fma.rn.ftz.f32 %f140, %f136, %f139, 0f00000000;
ld.global.s8 %r416, [%rd114+4];
shr.u32 %r417, %r395, 4;
and.b32 %r418, %r417, 252645135;
and.b32 %r419, %r397, 808464432;
or.b32 %r389, %r419, %r418;
or.b32 %r420, %r389, -2139062144;
add.s32 %r421, %r420, -538976288;
xor.b32 %r422, %r421, -2139062144;
xor.b32 %r423, %r389, 538976288;
xor.b32 %r424, %r421, %r389;
and.b32 %r387, %r424, %r423;

prmt.b32 %r387, %r387, 0, 0xba98;


prmt.b32 %r389, %r389, 0, 0xba98;

xor.b32 %r425, %r389, 2139062143;
not.b32 %r426, %r387;
and.b32 %r427, %r422, %r426;
and.b32 %r428, %r425, %r387;
or.b32 %r392, %r428, %r427;

dp4a.s32.s32 %r391, %r392, %r22, %r343;

mul.lo.s32 %r429, %r391, %r416;
cvt.rn.f32.s32 %f141, %r429;
fma.rn.ftz.f32 %f142, %f137, %f141, %f140;
fma.rn.ftz.f32 %f32, %f138, %f142, %f32;

$L__BB185_16:
mul.wide.s32 %rd127, %r46, 210;
add.s64 %rd116, %rd32, %rd127;
ld.global.nc.u16 %rs38, [%rd116];
ld.global.nc.u16 %rs39, [%rd116+2];
mov.b32 %r448, {%rs38, %rs39};
add.s64 %rd117, %rd33, %rd127;
ld.global.nc.u16 %rs40, [%rd117];
ld.global.nc.u16 %rs41, [%rd117+2];
mov.b32 %r449, {%rs40, %rs41};
shr.s32 %r450, %r449, %r15;

{.reg .f16 low,high;
mov.b32 {low,high},%r21;
cvt.f32.f16 %f143, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r23;
cvt.f32.f16 %f144, low;}


add.s64 %rd118, %rd34, %rd127;
ld.global.nc.u16 %rs37, [%rd118];

{ cvt.f32.f16 %f145, %rs37;}


add.s64 %rd119, %rd35, %rd127;
ld.global.nc.u8 %rs42, [%rd119];
cvt.u32.u16 %r451, %rs42;
cvt.s32.s8 %r452, %r451;
and.b32 %r453, %r448, 252645135;
shl.b32 %r454, %r450, 4;
and.b32 %r455, %r454, 808464432;
or.b32 %r434, %r455, %r453;
or.b32 %r456, %r434, -2139062144;
add.s32 %r457, %r456, -538976288;
xor.b32 %r458, %r457, -2139062144;
xor.b32 %r459, %r434, 538976288;
xor.b32 %r460, %r457, %r434;
and.b32 %r432, %r460, %r459;

prmt.b32 %r432, %r432, 0, 0xba98;


prmt.b32 %r434, %r434, 0, 0xba98;

xor.b32 %r461, %r434, 2139062143;
not.b32 %r462, %r432;
and.b32 %r463, %r458, %r462;
and.b32 %r464, %r461, %r432;
or.b32 %r437, %r464, %r463;

dp4a.s32.s32 %r436, %r437, %r20, %r343;

mul.lo.s32 %r465, %r436, %r452;
cvt.rn.f32.s32 %f146, %r465;
fma.rn.ftz.f32 %f147, %f143, %f146, 0f00000000;
ld.global.nc.u8 %rs43, [%rd119+4];
cvt.u32.u16 %r466, %rs43;
cvt.s32.s8 %r467, %r466;
shr.u32 %r468, %r448, 4;
and.b32 %r469, %r468, 252645135;
and.b32 %r470, %r450, 808464432;
or.b32 %r442, %r470, %r469;
or.b32 %r471, %r442, -2139062144;
add.s32 %r472, %r471, -538976288;
xor.b32 %r473, %r472, -2139062144;
xor.b32 %r474, %r442, 538976288;
xor.b32 %r475, %r472, %r442;
and.b32 %r440, %r475, %r474;

prmt.b32 %r440, %r440, 0, 0xba98;


prmt.b32 %r442, %r442, 0, 0xba98;

xor.b32 %r476, %r442, 2139062143;
not.b32 %r477, %r440;
and.b32 %r478, %r473, %r477;
and.b32 %r479, %r476, %r440;
or.b32 %r445, %r479, %r478;

dp4a.s32.s32 %r444, %r445, %r22, %r343;

mul.lo.s32 %r480, %r444, %r467;
cvt.rn.f32.s32 %f148, %r480;
fma.rn.ftz.f32 %f149, %f144, %f148, %f147;
fma.rn.ftz.f32 %f27, %f145, %f149, %f27;
@%p13 bra $L__BB185_18;

add.s64 %rd120, %rd138, %rd16;
ld.global.u16 %rs45, [%rd120+-2];
ld.global.u16 %rs46, [%rd120];
mov.b32 %r499, {%rs45, %rs46};
add.s64 %rd121, %rd138, %rd18;
ld.global.u16 %rs47, [%rd121+128];
ld.global.u16 %rs48, [%rd121+130];
mov.b32 %r500, {%rs47, %rs48};
shr.s32 %r501, %r500, %r15;

{.reg .f16 low,high;
mov.b32 {low,high},%r21;
cvt.f32.f16 %f150, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r23;
cvt.f32.f16 %f151, low;}


or.b32 %r504, %r4, 3;
mad.lo.s32 %r505, %r46, %r504, %r17;
mul.wide.s32 %rd122, %r505, 210;
add.s64 %rd123, %rd138, %rd122;
ld.global.u16 %rs44, [%rd123+208];

{ cvt.f32.f16 %f152, %rs44;}


add.s64 %rd124, %rd138, %rd19;
ld.global.s8 %r506, [%rd124+192];
and.b32 %r507, %r499, 252645135;
shl.b32 %r508, %r501, 4;
and.b32 %r509, %r508, 808464432;
or.b32 %r485, %r509, %r507;
or.b32 %r510, %r485, -2139062144;
add.s32 %r511, %r510, -538976288;
xor.b32 %r512, %r511, -2139062144;
xor.b32 %r513, %r485, 538976288;
xor.b32 %r514, %r511, %r485;
and.b32 %r483, %r514, %r513;

prmt.b32 %r483, %r483, 0, 0xba98;


prmt.b32 %r485, %r485, 0, 0xba98;

xor.b32 %r515, %r485, 2139062143;
not.b32 %r516, %r483;
and.b32 %r517, %r512, %r516;
and.b32 %r518, %r515, %r483;
or.b32 %r488, %r518, %r517;
mov.u32 %r498, 0;

dp4a.s32.s32 %r487, %r488, %r20, %r498;

mul.lo.s32 %r519, %r487, %r506;
cvt.rn.f32.s32 %f153, %r519;
fma.rn.ftz.f32 %f154, %f150, %f153, 0f00000000;
ld.global.s8 %r520, [%rd124+196];
shr.u32 %r521, %r499, 4;
and.b32 %r522, %r521, 252645135;
and.b32 %r523, %r501, 808464432;
or.b32 %r493, %r523, %r522;
or.b32 %r524, %r493, -2139062144;
add.s32 %r525, %r524, -538976288;
xor.b32 %r526, %r525, -2139062144;
xor.b32 %r527, %r493, 538976288;
xor.b32 %r528, %r525, %r493;
and.b32 %r491, %r528, %r527;

prmt.b32 %r491, %r491, 0, 0xba98;


prmt.b32 %r493, %r493, 0, 0xba98;

xor.b32 %r529, %r493, 2139062143;
not.b32 %r530, %r491;
and.b32 %r531, %r526, %r530;
and.b32 %r532, %r529, %r491;
or.b32 %r496, %r532, %r531;

dp4a.s32.s32 %r495, %r496, %r22, %r498;

mul.lo.s32 %r533, %r495, %r520;
cvt.rn.f32.s32 %f155, %r533;
fma.rn.ftz.f32 %f156, %f151, %f155, %f154;
fma.rn.ftz.f32 %f31, %f152, %f156, %f31;

$L__BB185_18:
add.s32 %r1280, %r1280, 32;
add.s64 %rd139, %rd139, 840;
add.s64 %rd138, %rd138, 840;
add.s32 %r1281, %r1281, 4;
setp.lt.s32 %p17, %r1281, %r80;
@%p17 bra $L__BB185_10;

$L__BB185_19:
mov.u32 %r1267, %tid.y;
setp.eq.s32 %p18, %r1267, 0;
@%p18 bra $L__BB185_23;

mov.u32 %r1268, %tid.y;
setp.ne.s64 %p19, %rd43, 0;
add.s32 %r26, %r1268, -1;
shl.b32 %r538, %r1268, 9;
mov.u32 %r539, _ZZ13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r540, %r539, %r538;
shl.b32 %r541, %r2, 2;
add.s32 %r542, %r540, %r541;
st.shared.f32 [%r542+-512], %f30;
@%p19 bra $L__BB185_22;
bra.uni $L__BB185_21;

$L__BB185_22:
shl.b32 %r544, %r26, 9;
mov.u32 %r545, _ZZ13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r546, %r545, %r544;
add.s32 %r548, %r546, %r541;
st.shared.f32 [%r548], %f34;
add.s32 %r1264, %r542, -512;
st.shared.f32 [%r1264+128], %f29;
st.shared.f32 [%r548+128], %f33;
add.s32 %r1265, %r542, -512;
st.shared.f32 [%r1265+256], %f28;
st.shared.f32 [%r548+256], %f32;
add.s32 %r1266, %r542, -512;
st.shared.f32 [%r1266+384], %f27;
st.shared.f32 [%r548+384], %f31;
bra.uni $L__BB185_23;

$L__BB185_21:
add.s32 %r1261, %r542, -512;
st.shared.f32 [%r1261+128], %f29;
add.s32 %r1262, %r542, -512;
st.shared.f32 [%r1262+256], %f28;
add.s32 %r1263, %r542, -512;
st.shared.f32 [%r1263+384], %f27;

$L__BB185_23:
mov.u32 %r1269, %tid.y;
setp.ne.s32 %p319, %r1269, 0;
bar.sync 0;
@%p319 bra $L__BB185_78;
bra.uni $L__BB185_24;

$L__BB185_78:
ret;

$L__BB185_24:
ld.param.u64 %rd128, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.f32 %f1001, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
ld.param.u32 %r1272, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r1271, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
mov.u32 %r1270, %ctaid.z;
setp.ne.s64 %p21, %rd43, 0;
mov.u32 %r551, %ctaid.y;
mad.lo.s32 %r554, %r551, %r1272, %r4;
mad.lo.s32 %r555, %r1270, %r1271, %r554;
neg.ftz.f32 %f35, %f1001;
shl.b32 %r557, %r2, 2;
mov.u32 %r558, _ZZ13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r28, %r558, %r557;
cvta.to.global.u64 %rd125, %rd128;
mul.wide.u32 %rd126, %r555, 4;
add.s64 %rd38, %rd125, %rd126;
@%p21 bra $L__BB185_33;
bra.uni $L__BB185_25;

$L__BB185_33:
ld.param.u32 %r1273, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
mov.u32 %r645, _ZZ13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r30, %r645, %r557;
setp.gt.s32 %p54, %r1273, 2;
@%p54 bra $L__BB185_44;

ld.param.u32 %r1276, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p57, %r1276, 1;
@%p57 bra $L__BB185_62;

ld.param.u32 %r1277, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p58, %r1277, 2;
@%p58 bra $L__BB185_36;
bra.uni $L__BB185_70;

$L__BB185_36:
ld.shared.f32 %f557, [%r28];
add.ftz.f32 %f558, %f557, %f30;
ld.shared.f32 %f559, [%r30];
add.ftz.f32 %f560, %f559, %f34;
ld.shared.f32 %f561, [%r28+512];
add.ftz.f32 %f562, %f561, %f558;
ld.shared.f32 %f563, [%r30+512];
add.ftz.f32 %f564, %f563, %f560;
ld.shared.f32 %f565, [%r28+1024];
add.ftz.f32 %f566, %f565, %f562;
ld.shared.f32 %f567, [%r30+1024];
add.ftz.f32 %f568, %f567, %f564;
mov.b32 %r892, %f566;
mov.u32 %r893, 31;
mov.u32 %r894, 16;
mov.u32 %r895, -1;
shfl.sync.bfly.b32 %r896|%p163, %r892, %r894, %r893, %r895;
mov.b32 %f569, %r896;
add.ftz.f32 %f570, %f566, %f569;
mov.b32 %r897, %f570;
mov.u32 %r898, 8;
shfl.sync.bfly.b32 %r899|%p164, %r897, %r898, %r893, %r895;
mov.b32 %f571, %r899;
add.ftz.f32 %f572, %f570, %f571;
mov.b32 %r900, %f572;
mov.u32 %r901, 4;
shfl.sync.bfly.b32 %r902|%p165, %r900, %r901, %r893, %r895;
mov.b32 %f573, %r902;
add.ftz.f32 %f574, %f572, %f573;
mov.b32 %r903, %f574;
mov.u32 %r904, 2;
shfl.sync.bfly.b32 %r905|%p166, %r903, %r904, %r893, %r895;
mov.b32 %f575, %r905;
add.ftz.f32 %f576, %f574, %f575;
mov.b32 %r906, %f576;
mov.u32 %r907, 1;
shfl.sync.bfly.b32 %r908|%p167, %r906, %r907, %r893, %r895;
mov.b32 %f577, %r908;
add.ftz.f32 %f56, %f576, %f577;
mov.b32 %r909, %f568;
shfl.sync.bfly.b32 %r910|%p168, %r909, %r894, %r893, %r895;
mov.b32 %f578, %r910;
add.ftz.f32 %f579, %f568, %f578;
mov.b32 %r911, %f579;
shfl.sync.bfly.b32 %r912|%p169, %r911, %r898, %r893, %r895;
mov.b32 %f580, %r912;
add.ftz.f32 %f581, %f579, %f580;
mov.b32 %r913, %f581;
shfl.sync.bfly.b32 %r914|%p170, %r913, %r901, %r893, %r895;
mov.b32 %f582, %r914;
add.ftz.f32 %f583, %f581, %f582;
mov.b32 %r915, %f583;
shfl.sync.bfly.b32 %r916|%p171, %r915, %r904, %r893, %r895;
mov.b32 %f584, %r916;
add.ftz.f32 %f585, %f583, %f584;
mov.b32 %r917, %f585;
shfl.sync.bfly.b32 %r918|%p172, %r917, %r907, %r893, %r895;
mov.b32 %f586, %r918;
add.ftz.f32 %f57, %f585, %f586;
setp.ne.s32 %p173, %r2, 0;
setp.ge.u32 %p174, %r4, %r47;
or.pred %p175, %p173, %p174;
@%p175 bra $L__BB185_38;

add.ftz.f32 %f587, %f5, %f56;
add.ftz.f32 %f588, %f4, %f57;
mul.ftz.f32 %f589, %f588, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f590, %f589;
add.ftz.f32 %f591, %f590, 0f3F800000;
div.approx.ftz.f32 %f592, %f588, %f591;
mul.ftz.f32 %f593, %f587, %f592;
st.global.f32 [%rd38], %f593;

$L__BB185_38:
ld.shared.f32 %f594, [%r28+128];
add.ftz.f32 %f595, %f594, %f29;
ld.shared.f32 %f596, [%r30+128];
add.ftz.f32 %f597, %f596, %f33;
ld.shared.f32 %f598, [%r28+640];
add.ftz.f32 %f599, %f598, %f595;
ld.shared.f32 %f600, [%r30+640];
add.ftz.f32 %f601, %f600, %f597;
ld.shared.f32 %f602, [%r28+1152];
add.ftz.f32 %f603, %f602, %f599;
ld.shared.f32 %f604, [%r30+1152];
add.ftz.f32 %f605, %f604, %f601;
mov.b32 %r922, %f603;
shfl.sync.bfly.b32 %r926|%p176, %r922, %r894, %r893, %r895;
mov.b32 %f606, %r926;
add.ftz.f32 %f607, %f603, %f606;
mov.b32 %r927, %f607;
shfl.sync.bfly.b32 %r929|%p177, %r927, %r898, %r893, %r895;
mov.b32 %f608, %r929;
add.ftz.f32 %f609, %f607, %f608;
mov.b32 %r930, %f609;
shfl.sync.bfly.b32 %r932|%p178, %r930, %r901, %r893, %r895;
mov.b32 %f610, %r932;
add.ftz.f32 %f611, %f609, %f610;
mov.b32 %r933, %f611;
shfl.sync.bfly.b32 %r935|%p179, %r933, %r904, %r893, %r895;
mov.b32 %f612, %r935;
add.ftz.f32 %f613, %f611, %f612;
mov.b32 %r936, %f613;
shfl.sync.bfly.b32 %r938|%p180, %r936, %r907, %r893, %r895;
mov.b32 %f614, %r938;
add.ftz.f32 %f58, %f613, %f614;
mov.b32 %r939, %f605;
shfl.sync.bfly.b32 %r940|%p181, %r939, %r894, %r893, %r895;
mov.b32 %f615, %r940;
add.ftz.f32 %f616, %f605, %f615;
mov.b32 %r941, %f616;
shfl.sync.bfly.b32 %r942|%p182, %r941, %r898, %r893, %r895;
mov.b32 %f617, %r942;
add.ftz.f32 %f618, %f616, %f617;
mov.b32 %r943, %f618;
shfl.sync.bfly.b32 %r944|%p183, %r943, %r901, %r893, %r895;
mov.b32 %f619, %r944;
add.ftz.f32 %f620, %f618, %f619;
mov.b32 %r945, %f620;
shfl.sync.bfly.b32 %r946|%p184, %r945, %r904, %r893, %r895;
mov.b32 %f621, %r946;
add.ftz.f32 %f622, %f620, %f621;
mov.b32 %r947, %f622;
shfl.sync.bfly.b32 %r948|%p185, %r947, %r907, %r893, %r895;
mov.b32 %f623, %r948;
add.ftz.f32 %f59, %f622, %f623;
setp.ne.s32 %p186, %r2, 1;
or.b32 %r952, %r4, 1;
setp.ge.u32 %p187, %r952, %r47;
or.pred %p188, %p186, %p187;
@%p188 bra $L__BB185_40;

add.ftz.f32 %f624, %f5, %f58;
add.ftz.f32 %f625, %f4, %f59;
mul.ftz.f32 %f626, %f625, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f627, %f626;
add.ftz.f32 %f628, %f627, 0f3F800000;
div.approx.ftz.f32 %f629, %f625, %f628;
mul.ftz.f32 %f630, %f624, %f629;
st.global.f32 [%rd38+4], %f630;

$L__BB185_40:
ld.shared.f32 %f631, [%r28+256];
add.ftz.f32 %f632, %f631, %f28;
ld.shared.f32 %f633, [%r30+256];
add.ftz.f32 %f634, %f633, %f32;
ld.shared.f32 %f635, [%r28+768];
add.ftz.f32 %f636, %f635, %f632;
ld.shared.f32 %f637, [%r30+768];
add.ftz.f32 %f638, %f637, %f634;
ld.shared.f32 %f639, [%r28+1280];
add.ftz.f32 %f640, %f639, %f636;
ld.shared.f32 %f641, [%r30+1280];
add.ftz.f32 %f642, %f641, %f638;
mov.b32 %r953, %f640;
mov.u32 %r954, 31;
mov.u32 %r955, 16;
mov.u32 %r956, -1;
shfl.sync.bfly.b32 %r957|%p189, %r953, %r955, %r954, %r956;
mov.b32 %f643, %r957;
add.ftz.f32 %f644, %f640, %f643;
mov.b32 %r958, %f644;
mov.u32 %r959, 8;
shfl.sync.bfly.b32 %r960|%p190, %r958, %r959, %r954, %r956;
mov.b32 %f645, %r960;
add.ftz.f32 %f646, %f644, %f645;
mov.b32 %r961, %f646;
mov.u32 %r962, 4;
shfl.sync.bfly.b32 %r963|%p191, %r961, %r962, %r954, %r956;
mov.b32 %f647, %r963;
add.ftz.f32 %f648, %f646, %f647;
mov.b32 %r964, %f648;
mov.u32 %r965, 2;
shfl.sync.bfly.b32 %r966|%p192, %r964, %r965, %r954, %r956;
mov.b32 %f649, %r966;
add.ftz.f32 %f650, %f648, %f649;
mov.b32 %r967, %f650;
mov.u32 %r968, 1;
shfl.sync.bfly.b32 %r969|%p193, %r967, %r968, %r954, %r956;
mov.b32 %f651, %r969;
add.ftz.f32 %f60, %f650, %f651;
mov.b32 %r970, %f642;
shfl.sync.bfly.b32 %r971|%p194, %r970, %r955, %r954, %r956;
mov.b32 %f652, %r971;
add.ftz.f32 %f653, %f642, %f652;
mov.b32 %r972, %f653;
shfl.sync.bfly.b32 %r973|%p195, %r972, %r959, %r954, %r956;
mov.b32 %f654, %r973;
add.ftz.f32 %f655, %f653, %f654;
mov.b32 %r974, %f655;
shfl.sync.bfly.b32 %r975|%p196, %r974, %r962, %r954, %r956;
mov.b32 %f656, %r975;
add.ftz.f32 %f657, %f655, %f656;
mov.b32 %r976, %f657;
shfl.sync.bfly.b32 %r977|%p197, %r976, %r965, %r954, %r956;
mov.b32 %f658, %r977;
add.ftz.f32 %f659, %f657, %f658;
mov.b32 %r978, %f659;
shfl.sync.bfly.b32 %r979|%p198, %r978, %r968, %r954, %r956;
mov.b32 %f660, %r979;
add.ftz.f32 %f61, %f659, %f660;
setp.ne.s32 %p199, %r2, 2;
or.b32 %r983, %r4, 2;
setp.ge.u32 %p200, %r983, %r47;
or.pred %p201, %p199, %p200;
@%p201 bra $L__BB185_42;

add.ftz.f32 %f661, %f5, %f60;
add.ftz.f32 %f662, %f4, %f61;
mul.ftz.f32 %f663, %f662, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f664, %f663;
add.ftz.f32 %f665, %f664, 0f3F800000;
div.approx.ftz.f32 %f666, %f662, %f665;
mul.ftz.f32 %f667, %f661, %f666;
st.global.f32 [%rd38+8], %f667;

$L__BB185_42:
ld.shared.f32 %f668, [%r28+384];
add.ftz.f32 %f669, %f668, %f27;
ld.shared.f32 %f670, [%r30+384];
add.ftz.f32 %f671, %f670, %f31;
ld.shared.f32 %f672, [%r28+896];
add.ftz.f32 %f673, %f672, %f669;
ld.shared.f32 %f674, [%r30+896];
add.ftz.f32 %f675, %f674, %f671;
ld.shared.f32 %f676, [%r28+1408];
add.ftz.f32 %f677, %f676, %f673;
ld.shared.f32 %f678, [%r30+1408];
add.ftz.f32 %f679, %f678, %f675;
mov.b32 %r984, %f677;
shfl.sync.bfly.b32 %r988|%p202, %r984, %r955, %r954, %r956;
mov.b32 %f680, %r988;
add.ftz.f32 %f681, %f677, %f680;
mov.b32 %r989, %f681;
shfl.sync.bfly.b32 %r991|%p203, %r989, %r959, %r954, %r956;
mov.b32 %f682, %r991;
add.ftz.f32 %f683, %f681, %f682;
mov.b32 %r992, %f683;
shfl.sync.bfly.b32 %r994|%p204, %r992, %r962, %r954, %r956;
mov.b32 %f684, %r994;
add.ftz.f32 %f685, %f683, %f684;
mov.b32 %r995, %f685;
shfl.sync.bfly.b32 %r997|%p205, %r995, %r965, %r954, %r956;
mov.b32 %f686, %r997;
add.ftz.f32 %f687, %f685, %f686;
mov.b32 %r998, %f687;
shfl.sync.bfly.b32 %r1000|%p206, %r998, %r968, %r954, %r956;
mov.b32 %f688, %r1000;
add.ftz.f32 %f62, %f687, %f688;
mov.b32 %r1001, %f679;
shfl.sync.bfly.b32 %r1002|%p207, %r1001, %r955, %r954, %r956;
mov.b32 %f689, %r1002;
add.ftz.f32 %f690, %f679, %f689;
mov.b32 %r1003, %f690;
shfl.sync.bfly.b32 %r1004|%p208, %r1003, %r959, %r954, %r956;
mov.b32 %f691, %r1004;
add.ftz.f32 %f692, %f690, %f691;
mov.b32 %r1005, %f692;
shfl.sync.bfly.b32 %r1006|%p209, %r1005, %r962, %r954, %r956;
mov.b32 %f693, %r1006;
add.ftz.f32 %f694, %f692, %f693;
mov.b32 %r1007, %f694;
shfl.sync.bfly.b32 %r1008|%p210, %r1007, %r965, %r954, %r956;
mov.b32 %f695, %r1008;
add.ftz.f32 %f696, %f694, %f695;
mov.b32 %r1009, %f696;
shfl.sync.bfly.b32 %r1010|%p211, %r1009, %r968, %r954, %r956;
mov.b32 %f697, %r1010;
add.ftz.f32 %f63, %f696, %f697;
setp.ne.s32 %p212, %r2, 3;
or.b32 %r1014, %r4, 3;
setp.ge.u32 %p213, %r1014, %r47;
or.pred %p214, %p212, %p213;
@%p214 bra $L__BB185_78;

add.ftz.f32 %f698, %f5, %f62;
add.ftz.f32 %f699, %f4, %f63;
mul.ftz.f32 %f700, %f699, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f701, %f700;
add.ftz.f32 %f702, %f701, 0f3F800000;
div.approx.ftz.f32 %f703, %f699, %f702;
mul.ftz.f32 %f704, %f698, %f703;
st.global.f32 [%rd38+12], %f704;
bra.uni $L__BB185_78;

$L__BB185_25:
ld.shared.f32 %f157, [%r28];
add.ftz.f32 %f158, %f157, %f30;
ld.shared.f32 %f159, [%r28+512];
add.ftz.f32 %f160, %f159, %f158;
ld.shared.f32 %f161, [%r28+1024];
add.ftz.f32 %f162, %f161, %f160;
mov.b32 %r559, %f162;
mov.u32 %r560, 31;
mov.u32 %r561, 16;
mov.u32 %r562, -1;
shfl.sync.bfly.b32 %r563|%p22, %r559, %r561, %r560, %r562;
mov.b32 %f163, %r563;
add.ftz.f32 %f164, %f162, %f163;
mov.b32 %r564, %f164;
mov.u32 %r565, 8;
shfl.sync.bfly.b32 %r566|%p23, %r564, %r565, %r560, %r562;
mov.b32 %f165, %r566;
add.ftz.f32 %f166, %f164, %f165;
mov.b32 %r567, %f166;
mov.u32 %r568, 4;
shfl.sync.bfly.b32 %r569|%p24, %r567, %r568, %r560, %r562;
mov.b32 %f167, %r569;
add.ftz.f32 %f168, %f166, %f167;
mov.b32 %r570, %f168;
mov.u32 %r571, 2;
shfl.sync.bfly.b32 %r572|%p25, %r570, %r571, %r560, %r562;
mov.b32 %f169, %r572;
add.ftz.f32 %f170, %f168, %f169;
mov.b32 %r573, %f170;
mov.u32 %r574, 1;
shfl.sync.bfly.b32 %r575|%p26, %r573, %r574, %r560, %r562;
mov.b32 %f171, %r575;
add.ftz.f32 %f36, %f170, %f171;
setp.ne.s32 %p27, %r2, 0;
setp.ge.u32 %p28, %r4, %r47;
or.pred %p29, %p27, %p28;
@%p29 bra $L__BB185_27;

add.ftz.f32 %f172, %f5, %f36;
st.global.f32 [%rd38], %f172;

$L__BB185_27:
ld.shared.f32 %f173, [%r28+128];
add.ftz.f32 %f174, %f173, %f29;
ld.shared.f32 %f175, [%r28+640];
add.ftz.f32 %f176, %f175, %f174;
ld.shared.f32 %f177, [%r28+1152];
add.ftz.f32 %f178, %f177, %f176;
mov.b32 %r579, %f178;
shfl.sync.bfly.b32 %r583|%p30, %r579, %r561, %r560, %r562;
mov.b32 %f179, %r583;
add.ftz.f32 %f180, %f178, %f179;
mov.b32 %r584, %f180;
shfl.sync.bfly.b32 %r586|%p31, %r584, %r565, %r560, %r562;
mov.b32 %f181, %r586;
add.ftz.f32 %f182, %f180, %f181;
mov.b32 %r587, %f182;
shfl.sync.bfly.b32 %r589|%p32, %r587, %r568, %r560, %r562;
mov.b32 %f183, %r589;
add.ftz.f32 %f184, %f182, %f183;
mov.b32 %r590, %f184;
shfl.sync.bfly.b32 %r592|%p33, %r590, %r571, %r560, %r562;
mov.b32 %f185, %r592;
add.ftz.f32 %f186, %f184, %f185;
mov.b32 %r593, %f186;
shfl.sync.bfly.b32 %r595|%p34, %r593, %r574, %r560, %r562;
mov.b32 %f187, %r595;
add.ftz.f32 %f37, %f186, %f187;
setp.ne.s32 %p35, %r2, 1;
or.b32 %r599, %r4, 1;
setp.ge.u32 %p36, %r599, %r47;
or.pred %p37, %p35, %p36;
@%p37 bra $L__BB185_29;

add.ftz.f32 %f188, %f5, %f37;
st.global.f32 [%rd38+4], %f188;

$L__BB185_29:
ld.shared.f32 %f189, [%r28+256];
add.ftz.f32 %f190, %f189, %f28;
ld.shared.f32 %f191, [%r28+768];
add.ftz.f32 %f192, %f191, %f190;
ld.shared.f32 %f193, [%r28+1280];
add.ftz.f32 %f194, %f193, %f192;
mov.b32 %r600, %f194;
mov.u32 %r601, 31;
mov.u32 %r602, 16;
mov.u32 %r603, -1;
shfl.sync.bfly.b32 %r604|%p38, %r600, %r602, %r601, %r603;
mov.b32 %f195, %r604;
add.ftz.f32 %f196, %f194, %f195;
mov.b32 %r605, %f196;
mov.u32 %r606, 8;
shfl.sync.bfly.b32 %r607|%p39, %r605, %r606, %r601, %r603;
mov.b32 %f197, %r607;
add.ftz.f32 %f198, %f196, %f197;
mov.b32 %r608, %f198;
mov.u32 %r609, 4;
shfl.sync.bfly.b32 %r610|%p40, %r608, %r609, %r601, %r603;
mov.b32 %f199, %r610;
add.ftz.f32 %f200, %f198, %f199;
mov.b32 %r611, %f200;
mov.u32 %r612, 2;
shfl.sync.bfly.b32 %r613|%p41, %r611, %r612, %r601, %r603;
mov.b32 %f201, %r613;
add.ftz.f32 %f202, %f200, %f201;
mov.b32 %r614, %f202;
mov.u32 %r615, 1;
shfl.sync.bfly.b32 %r616|%p42, %r614, %r615, %r601, %r603;
mov.b32 %f203, %r616;
add.ftz.f32 %f38, %f202, %f203;
setp.ne.s32 %p43, %r2, 2;
or.b32 %r620, %r4, 2;
setp.ge.u32 %p44, %r620, %r47;
or.pred %p45, %p43, %p44;
@%p45 bra $L__BB185_31;

add.ftz.f32 %f204, %f5, %f38;
st.global.f32 [%rd38+8], %f204;

$L__BB185_31:
ld.shared.f32 %f205, [%r28+384];
add.ftz.f32 %f206, %f205, %f27;
ld.shared.f32 %f207, [%r28+896];
add.ftz.f32 %f208, %f207, %f206;
ld.shared.f32 %f209, [%r28+1408];
add.ftz.f32 %f210, %f209, %f208;
mov.b32 %r621, %f210;
shfl.sync.bfly.b32 %r625|%p46, %r621, %r602, %r601, %r603;
mov.b32 %f211, %r625;
add.ftz.f32 %f212, %f210, %f211;
mov.b32 %r626, %f212;
shfl.sync.bfly.b32 %r628|%p47, %r626, %r606, %r601, %r603;
mov.b32 %f213, %r628;
add.ftz.f32 %f214, %f212, %f213;
mov.b32 %r629, %f214;
shfl.sync.bfly.b32 %r631|%p48, %r629, %r609, %r601, %r603;
mov.b32 %f215, %r631;
add.ftz.f32 %f216, %f214, %f215;
mov.b32 %r632, %f216;
shfl.sync.bfly.b32 %r634|%p49, %r632, %r612, %r601, %r603;
mov.b32 %f217, %r634;
add.ftz.f32 %f218, %f216, %f217;
mov.b32 %r635, %f218;
shfl.sync.bfly.b32 %r637|%p50, %r635, %r615, %r601, %r603;
mov.b32 %f219, %r637;
add.ftz.f32 %f39, %f218, %f219;
setp.ne.s32 %p51, %r2, 3;
or.b32 %r641, %r4, 3;
setp.ge.u32 %p52, %r641, %r47;
or.pred %p53, %p51, %p52;
@%p53 bra $L__BB185_78;

add.ftz.f32 %f220, %f5, %f39;
st.global.f32 [%rd38+12], %f220;
bra.uni $L__BB185_78;

$L__BB185_44:
ld.param.u32 %r1274, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p55, %r1274, 3;
@%p55 bra $L__BB185_54;

ld.param.u32 %r1275, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.ne.s32 %p56, %r1275, 6;
@%p56 bra $L__BB185_70;

ld.shared.f32 %f221, [%r28];
add.ftz.f32 %f222, %f221, %f30;
ld.shared.f32 %f223, [%r30];
add.ftz.f32 %f224, %f223, %f34;
ld.shared.f32 %f225, [%r28+512];
add.ftz.f32 %f226, %f225, %f222;
ld.shared.f32 %f227, [%r30+512];
add.ftz.f32 %f228, %f227, %f224;
ld.shared.f32 %f229, [%r28+1024];
add.ftz.f32 %f230, %f229, %f226;
ld.shared.f32 %f231, [%r30+1024];
add.ftz.f32 %f232, %f231, %f228;
mov.b32 %r646, %f230;
mov.u32 %r647, 31;
mov.u32 %r648, 16;
mov.u32 %r649, -1;
shfl.sync.bfly.b32 %r650|%p59, %r646, %r648, %r647, %r649;
mov.b32 %f233, %r650;
add.ftz.f32 %f234, %f230, %f233;
mov.b32 %r651, %f234;
mov.u32 %r652, 8;
shfl.sync.bfly.b32 %r653|%p60, %r651, %r652, %r647, %r649;
mov.b32 %f235, %r653;
add.ftz.f32 %f236, %f234, %f235;
mov.b32 %r654, %f236;
mov.u32 %r655, 4;
shfl.sync.bfly.b32 %r656|%p61, %r654, %r655, %r647, %r649;
mov.b32 %f237, %r656;
add.ftz.f32 %f238, %f236, %f237;
mov.b32 %r657, %f238;
mov.u32 %r658, 2;
shfl.sync.bfly.b32 %r659|%p62, %r657, %r658, %r647, %r649;
mov.b32 %f239, %r659;
add.ftz.f32 %f240, %f238, %f239;
mov.b32 %r660, %f240;
mov.u32 %r661, 1;
shfl.sync.bfly.b32 %r662|%p63, %r660, %r661, %r647, %r649;
mov.b32 %f241, %r662;
add.ftz.f32 %f40, %f240, %f241;
mov.b32 %r663, %f232;
shfl.sync.bfly.b32 %r664|%p64, %r663, %r648, %r647, %r649;
mov.b32 %f242, %r664;
add.ftz.f32 %f243, %f232, %f242;
mov.b32 %r665, %f243;
shfl.sync.bfly.b32 %r666|%p65, %r665, %r652, %r647, %r649;
mov.b32 %f244, %r666;
add.ftz.f32 %f245, %f243, %f244;
mov.b32 %r667, %f245;
shfl.sync.bfly.b32 %r668|%p66, %r667, %r655, %r647, %r649;
mov.b32 %f246, %r668;
add.ftz.f32 %f247, %f245, %f246;
mov.b32 %r669, %f247;
shfl.sync.bfly.b32 %r670|%p67, %r669, %r658, %r647, %r649;
mov.b32 %f248, %r670;
add.ftz.f32 %f249, %f247, %f248;
mov.b32 %r671, %f249;
shfl.sync.bfly.b32 %r672|%p68, %r671, %r661, %r647, %r649;
mov.b32 %f250, %r672;
add.ftz.f32 %f41, %f249, %f250;
setp.ne.s32 %p69, %r2, 0;
setp.ge.u32 %p70, %r4, %r47;
or.pred %p71, %p69, %p70;
@%p71 bra $L__BB185_48;

ld.param.f32 %f1005, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f251, %f5, %f40;
add.ftz.f32 %f252, %f4, %f41;
min.ftz.f32 %f253, %f252, %f1005;
min.ftz.f32 %f254, %f251, %f1005;
max.ftz.f32 %f255, %f254, %f35;
mul.ftz.f32 %f256, %f253, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f257, %f256;
add.ftz.f32 %f258, %f257, 0f3F800000;
div.approx.ftz.f32 %f259, %f253, %f258;
mul.ftz.f32 %f260, %f255, %f259;
st.global.f32 [%rd38], %f260;

$L__BB185_48:
ld.shared.f32 %f261, [%r28+128];
add.ftz.f32 %f262, %f261, %f29;
ld.shared.f32 %f263, [%r30+128];
add.ftz.f32 %f264, %f263, %f33;
ld.shared.f32 %f265, [%r28+640];
add.ftz.f32 %f266, %f265, %f262;
ld.shared.f32 %f267, [%r30+640];
add.ftz.f32 %f268, %f267, %f264;
ld.shared.f32 %f269, [%r28+1152];
add.ftz.f32 %f270, %f269, %f266;
ld.shared.f32 %f271, [%r30+1152];
add.ftz.f32 %f272, %f271, %f268;
mov.b32 %r676, %f270;
shfl.sync.bfly.b32 %r680|%p72, %r676, %r648, %r647, %r649;
mov.b32 %f273, %r680;
add.ftz.f32 %f274, %f270, %f273;
mov.b32 %r681, %f274;
shfl.sync.bfly.b32 %r683|%p73, %r681, %r652, %r647, %r649;
mov.b32 %f275, %r683;
add.ftz.f32 %f276, %f274, %f275;
mov.b32 %r684, %f276;
shfl.sync.bfly.b32 %r686|%p74, %r684, %r655, %r647, %r649;
mov.b32 %f277, %r686;
add.ftz.f32 %f278, %f276, %f277;
mov.b32 %r687, %f278;
shfl.sync.bfly.b32 %r689|%p75, %r687, %r658, %r647, %r649;
mov.b32 %f279, %r689;
add.ftz.f32 %f280, %f278, %f279;
mov.b32 %r690, %f280;
shfl.sync.bfly.b32 %r692|%p76, %r690, %r661, %r647, %r649;
mov.b32 %f281, %r692;
add.ftz.f32 %f42, %f280, %f281;
mov.b32 %r693, %f272;
shfl.sync.bfly.b32 %r694|%p77, %r693, %r648, %r647, %r649;
mov.b32 %f282, %r694;
add.ftz.f32 %f283, %f272, %f282;
mov.b32 %r695, %f283;
shfl.sync.bfly.b32 %r696|%p78, %r695, %r652, %r647, %r649;
mov.b32 %f284, %r696;
add.ftz.f32 %f285, %f283, %f284;
mov.b32 %r697, %f285;
shfl.sync.bfly.b32 %r698|%p79, %r697, %r655, %r647, %r649;
mov.b32 %f286, %r698;
add.ftz.f32 %f287, %f285, %f286;
mov.b32 %r699, %f287;
shfl.sync.bfly.b32 %r700|%p80, %r699, %r658, %r647, %r649;
mov.b32 %f288, %r700;
add.ftz.f32 %f289, %f287, %f288;
mov.b32 %r701, %f289;
shfl.sync.bfly.b32 %r702|%p81, %r701, %r661, %r647, %r649;
mov.b32 %f290, %r702;
add.ftz.f32 %f43, %f289, %f290;
setp.ne.s32 %p82, %r2, 1;
or.b32 %r706, %r4, 1;
setp.ge.u32 %p83, %r706, %r47;
or.pred %p84, %p82, %p83;
@%p84 bra $L__BB185_50;

ld.param.f32 %f1004, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f291, %f5, %f42;
add.ftz.f32 %f292, %f4, %f43;
min.ftz.f32 %f293, %f292, %f1004;
min.ftz.f32 %f294, %f291, %f1004;
max.ftz.f32 %f295, %f294, %f35;
mul.ftz.f32 %f296, %f293, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f297, %f296;
add.ftz.f32 %f298, %f297, 0f3F800000;
div.approx.ftz.f32 %f299, %f293, %f298;
mul.ftz.f32 %f300, %f295, %f299;
st.global.f32 [%rd38+4], %f300;

$L__BB185_50:
ld.shared.f32 %f301, [%r28+256];
add.ftz.f32 %f302, %f301, %f28;
ld.shared.f32 %f303, [%r30+256];
add.ftz.f32 %f304, %f303, %f32;
ld.shared.f32 %f305, [%r28+768];
add.ftz.f32 %f306, %f305, %f302;
ld.shared.f32 %f307, [%r30+768];
add.ftz.f32 %f308, %f307, %f304;
ld.shared.f32 %f309, [%r28+1280];
add.ftz.f32 %f310, %f309, %f306;
ld.shared.f32 %f311, [%r30+1280];
add.ftz.f32 %f312, %f311, %f308;
mov.b32 %r707, %f310;
mov.u32 %r708, 31;
mov.u32 %r709, 16;
mov.u32 %r710, -1;
shfl.sync.bfly.b32 %r711|%p85, %r707, %r709, %r708, %r710;
mov.b32 %f313, %r711;
add.ftz.f32 %f314, %f310, %f313;
mov.b32 %r712, %f314;
mov.u32 %r713, 8;
shfl.sync.bfly.b32 %r714|%p86, %r712, %r713, %r708, %r710;
mov.b32 %f315, %r714;
add.ftz.f32 %f316, %f314, %f315;
mov.b32 %r715, %f316;
mov.u32 %r716, 4;
shfl.sync.bfly.b32 %r717|%p87, %r715, %r716, %r708, %r710;
mov.b32 %f317, %r717;
add.ftz.f32 %f318, %f316, %f317;
mov.b32 %r718, %f318;
mov.u32 %r719, 2;
shfl.sync.bfly.b32 %r720|%p88, %r718, %r719, %r708, %r710;
mov.b32 %f319, %r720;
add.ftz.f32 %f320, %f318, %f319;
mov.b32 %r721, %f320;
mov.u32 %r722, 1;
shfl.sync.bfly.b32 %r723|%p89, %r721, %r722, %r708, %r710;
mov.b32 %f321, %r723;
add.ftz.f32 %f44, %f320, %f321;
mov.b32 %r724, %f312;
shfl.sync.bfly.b32 %r725|%p90, %r724, %r709, %r708, %r710;
mov.b32 %f322, %r725;
add.ftz.f32 %f323, %f312, %f322;
mov.b32 %r726, %f323;
shfl.sync.bfly.b32 %r727|%p91, %r726, %r713, %r708, %r710;
mov.b32 %f324, %r727;
add.ftz.f32 %f325, %f323, %f324;
mov.b32 %r728, %f325;
shfl.sync.bfly.b32 %r729|%p92, %r728, %r716, %r708, %r710;
mov.b32 %f326, %r729;
add.ftz.f32 %f327, %f325, %f326;
mov.b32 %r730, %f327;
shfl.sync.bfly.b32 %r731|%p93, %r730, %r719, %r708, %r710;
mov.b32 %f328, %r731;
add.ftz.f32 %f329, %f327, %f328;
mov.b32 %r732, %f329;
shfl.sync.bfly.b32 %r733|%p94, %r732, %r722, %r708, %r710;
mov.b32 %f330, %r733;
add.ftz.f32 %f45, %f329, %f330;
setp.ne.s32 %p95, %r2, 2;
or.b32 %r737, %r4, 2;
setp.ge.u32 %p96, %r737, %r47;
or.pred %p97, %p95, %p96;
@%p97 bra $L__BB185_52;

ld.param.f32 %f1003, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f331, %f5, %f44;
add.ftz.f32 %f332, %f4, %f45;
min.ftz.f32 %f333, %f332, %f1003;
min.ftz.f32 %f334, %f331, %f1003;
max.ftz.f32 %f335, %f334, %f35;
mul.ftz.f32 %f336, %f333, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f337, %f336;
add.ftz.f32 %f338, %f337, 0f3F800000;
div.approx.ftz.f32 %f339, %f333, %f338;
mul.ftz.f32 %f340, %f335, %f339;
st.global.f32 [%rd38+8], %f340;

$L__BB185_52:
ld.shared.f32 %f341, [%r28+384];
add.ftz.f32 %f342, %f341, %f27;
ld.shared.f32 %f343, [%r30+384];
add.ftz.f32 %f344, %f343, %f31;
ld.shared.f32 %f345, [%r28+896];
add.ftz.f32 %f346, %f345, %f342;
ld.shared.f32 %f347, [%r30+896];
add.ftz.f32 %f348, %f347, %f344;
ld.shared.f32 %f349, [%r28+1408];
add.ftz.f32 %f350, %f349, %f346;
ld.shared.f32 %f351, [%r30+1408];
add.ftz.f32 %f352, %f351, %f348;
mov.b32 %r738, %f350;
shfl.sync.bfly.b32 %r742|%p98, %r738, %r709, %r708, %r710;
mov.b32 %f353, %r742;
add.ftz.f32 %f354, %f350, %f353;
mov.b32 %r743, %f354;
shfl.sync.bfly.b32 %r745|%p99, %r743, %r713, %r708, %r710;
mov.b32 %f355, %r745;
add.ftz.f32 %f356, %f354, %f355;
mov.b32 %r746, %f356;
shfl.sync.bfly.b32 %r748|%p100, %r746, %r716, %r708, %r710;
mov.b32 %f357, %r748;
add.ftz.f32 %f358, %f356, %f357;
mov.b32 %r749, %f358;
shfl.sync.bfly.b32 %r751|%p101, %r749, %r719, %r708, %r710;
mov.b32 %f359, %r751;
add.ftz.f32 %f360, %f358, %f359;
mov.b32 %r752, %f360;
shfl.sync.bfly.b32 %r754|%p102, %r752, %r722, %r708, %r710;
mov.b32 %f361, %r754;
add.ftz.f32 %f46, %f360, %f361;
mov.b32 %r755, %f352;
shfl.sync.bfly.b32 %r756|%p103, %r755, %r709, %r708, %r710;
mov.b32 %f362, %r756;
add.ftz.f32 %f363, %f352, %f362;
mov.b32 %r757, %f363;
shfl.sync.bfly.b32 %r758|%p104, %r757, %r713, %r708, %r710;
mov.b32 %f364, %r758;
add.ftz.f32 %f365, %f363, %f364;
mov.b32 %r759, %f365;
shfl.sync.bfly.b32 %r760|%p105, %r759, %r716, %r708, %r710;
mov.b32 %f366, %r760;
add.ftz.f32 %f367, %f365, %f366;
mov.b32 %r761, %f367;
shfl.sync.bfly.b32 %r762|%p106, %r761, %r719, %r708, %r710;
mov.b32 %f368, %r762;
add.ftz.f32 %f369, %f367, %f368;
mov.b32 %r763, %f369;
shfl.sync.bfly.b32 %r764|%p107, %r763, %r722, %r708, %r710;
mov.b32 %f370, %r764;
add.ftz.f32 %f47, %f369, %f370;
setp.ne.s32 %p108, %r2, 3;
or.b32 %r768, %r4, 3;
setp.ge.u32 %p109, %r768, %r47;
or.pred %p110, %p108, %p109;
@%p110 bra $L__BB185_78;

ld.param.f32 %f1002, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f371, %f5, %f46;
add.ftz.f32 %f372, %f4, %f47;
min.ftz.f32 %f373, %f372, %f1002;
min.ftz.f32 %f374, %f371, %f1002;
max.ftz.f32 %f375, %f374, %f35;
mul.ftz.f32 %f376, %f373, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f377, %f376;
add.ftz.f32 %f378, %f377, 0f3F800000;
div.approx.ftz.f32 %f379, %f373, %f378;
mul.ftz.f32 %f380, %f375, %f379;
st.global.f32 [%rd38+12], %f380;
bra.uni $L__BB185_78;

$L__BB185_70:
ld.shared.f32 %f869, [%r28];
add.ftz.f32 %f870, %f869, %f30;
ld.shared.f32 %f871, [%r30];
add.ftz.f32 %f872, %f871, %f34;
ld.shared.f32 %f873, [%r28+512];
add.ftz.f32 %f874, %f873, %f870;
ld.shared.f32 %f875, [%r30+512];
add.ftz.f32 %f876, %f875, %f872;
ld.shared.f32 %f877, [%r28+1024];
add.ftz.f32 %f878, %f877, %f874;
ld.shared.f32 %f879, [%r30+1024];
add.ftz.f32 %f880, %f879, %f876;
mov.b32 %r1138, %f878;
mov.u32 %r1139, 31;
mov.u32 %r1140, 16;
mov.u32 %r1141, -1;
shfl.sync.bfly.b32 %r1142|%p267, %r1138, %r1140, %r1139, %r1141;
mov.b32 %f881, %r1142;
add.ftz.f32 %f882, %f878, %f881;
mov.b32 %r1143, %f882;
mov.u32 %r1144, 8;
shfl.sync.bfly.b32 %r1145|%p268, %r1143, %r1144, %r1139, %r1141;
mov.b32 %f883, %r1145;
add.ftz.f32 %f884, %f882, %f883;
mov.b32 %r1146, %f884;
mov.u32 %r1147, 4;
shfl.sync.bfly.b32 %r1148|%p269, %r1146, %r1147, %r1139, %r1141;
mov.b32 %f885, %r1148;
add.ftz.f32 %f886, %f884, %f885;
mov.b32 %r1149, %f886;
mov.u32 %r1150, 2;
shfl.sync.bfly.b32 %r1151|%p270, %r1149, %r1150, %r1139, %r1141;
mov.b32 %f887, %r1151;
add.ftz.f32 %f888, %f886, %f887;
mov.b32 %r1152, %f888;
mov.u32 %r1153, 1;
shfl.sync.bfly.b32 %r1154|%p271, %r1152, %r1153, %r1139, %r1141;
mov.b32 %f889, %r1154;
add.ftz.f32 %f72, %f888, %f889;
mov.b32 %r1155, %f880;
shfl.sync.bfly.b32 %r1156|%p272, %r1155, %r1140, %r1139, %r1141;
mov.b32 %f890, %r1156;
add.ftz.f32 %f891, %f880, %f890;
mov.b32 %r1157, %f891;
shfl.sync.bfly.b32 %r1158|%p273, %r1157, %r1144, %r1139, %r1141;
mov.b32 %f892, %r1158;
add.ftz.f32 %f893, %f891, %f892;
mov.b32 %r1159, %f893;
shfl.sync.bfly.b32 %r1160|%p274, %r1159, %r1147, %r1139, %r1141;
mov.b32 %f894, %r1160;
add.ftz.f32 %f895, %f893, %f894;
mov.b32 %r1161, %f895;
shfl.sync.bfly.b32 %r1162|%p275, %r1161, %r1150, %r1139, %r1141;
mov.b32 %f896, %r1162;
add.ftz.f32 %f897, %f895, %f896;
mov.b32 %r1163, %f897;
shfl.sync.bfly.b32 %r1164|%p276, %r1163, %r1153, %r1139, %r1141;
mov.b32 %f898, %r1164;
add.ftz.f32 %f73, %f897, %f898;
setp.ne.s32 %p277, %r2, 0;
setp.ge.u32 %p278, %r4, %r47;
or.pred %p279, %p277, %p278;
@%p279 bra $L__BB185_72;

add.ftz.f32 %f899, %f5, %f72;
add.ftz.f32 %f900, %f4, %f73;
mul.ftz.f32 %f901, %f899, %f900;
st.global.f32 [%rd38], %f901;

$L__BB185_72:
ld.shared.f32 %f902, [%r28+128];
add.ftz.f32 %f903, %f902, %f29;
ld.shared.f32 %f904, [%r30+128];
add.ftz.f32 %f905, %f904, %f33;
ld.shared.f32 %f906, [%r28+640];
add.ftz.f32 %f907, %f906, %f903;
ld.shared.f32 %f908, [%r30+640];
add.ftz.f32 %f909, %f908, %f905;
ld.shared.f32 %f910, [%r28+1152];
add.ftz.f32 %f911, %f910, %f907;
ld.shared.f32 %f912, [%r30+1152];
add.ftz.f32 %f913, %f912, %f909;
mov.b32 %r1168, %f911;
shfl.sync.bfly.b32 %r1172|%p280, %r1168, %r1140, %r1139, %r1141;
mov.b32 %f914, %r1172;
add.ftz.f32 %f915, %f911, %f914;
mov.b32 %r1173, %f915;
shfl.sync.bfly.b32 %r1175|%p281, %r1173, %r1144, %r1139, %r1141;
mov.b32 %f916, %r1175;
add.ftz.f32 %f917, %f915, %f916;
mov.b32 %r1176, %f917;
shfl.sync.bfly.b32 %r1178|%p282, %r1176, %r1147, %r1139, %r1141;
mov.b32 %f918, %r1178;
add.ftz.f32 %f919, %f917, %f918;
mov.b32 %r1179, %f919;
shfl.sync.bfly.b32 %r1181|%p283, %r1179, %r1150, %r1139, %r1141;
mov.b32 %f920, %r1181;
add.ftz.f32 %f921, %f919, %f920;
mov.b32 %r1182, %f921;
shfl.sync.bfly.b32 %r1184|%p284, %r1182, %r1153, %r1139, %r1141;
mov.b32 %f922, %r1184;
add.ftz.f32 %f74, %f921, %f922;
mov.b32 %r1185, %f913;
shfl.sync.bfly.b32 %r1186|%p285, %r1185, %r1140, %r1139, %r1141;
mov.b32 %f923, %r1186;
add.ftz.f32 %f924, %f913, %f923;
mov.b32 %r1187, %f924;
shfl.sync.bfly.b32 %r1188|%p286, %r1187, %r1144, %r1139, %r1141;
mov.b32 %f925, %r1188;
add.ftz.f32 %f926, %f924, %f925;
mov.b32 %r1189, %f926;
shfl.sync.bfly.b32 %r1190|%p287, %r1189, %r1147, %r1139, %r1141;
mov.b32 %f927, %r1190;
add.ftz.f32 %f928, %f926, %f927;
mov.b32 %r1191, %f928;
shfl.sync.bfly.b32 %r1192|%p288, %r1191, %r1150, %r1139, %r1141;
mov.b32 %f929, %r1192;
add.ftz.f32 %f930, %f928, %f929;
mov.b32 %r1193, %f930;
shfl.sync.bfly.b32 %r1194|%p289, %r1193, %r1153, %r1139, %r1141;
mov.b32 %f931, %r1194;
add.ftz.f32 %f75, %f930, %f931;
setp.ne.s32 %p290, %r2, 1;
or.b32 %r1198, %r4, 1;
setp.ge.u32 %p291, %r1198, %r47;
or.pred %p292, %p290, %p291;
@%p292 bra $L__BB185_74;

add.ftz.f32 %f932, %f5, %f74;
add.ftz.f32 %f933, %f4, %f75;
mul.ftz.f32 %f934, %f932, %f933;
st.global.f32 [%rd38+4], %f934;

$L__BB185_74:
ld.shared.f32 %f935, [%r28+256];
add.ftz.f32 %f936, %f935, %f28;
ld.shared.f32 %f937, [%r30+256];
add.ftz.f32 %f938, %f937, %f32;
ld.shared.f32 %f939, [%r28+768];
add.ftz.f32 %f940, %f939, %f936;
ld.shared.f32 %f941, [%r30+768];
add.ftz.f32 %f942, %f941, %f938;
ld.shared.f32 %f943, [%r28+1280];
add.ftz.f32 %f944, %f943, %f940;
ld.shared.f32 %f945, [%r30+1280];
add.ftz.f32 %f946, %f945, %f942;
mov.b32 %r1199, %f944;
mov.u32 %r1200, 31;
mov.u32 %r1201, 16;
mov.u32 %r1202, -1;
shfl.sync.bfly.b32 %r1203|%p293, %r1199, %r1201, %r1200, %r1202;
mov.b32 %f947, %r1203;
add.ftz.f32 %f948, %f944, %f947;
mov.b32 %r1204, %f948;
mov.u32 %r1205, 8;
shfl.sync.bfly.b32 %r1206|%p294, %r1204, %r1205, %r1200, %r1202;
mov.b32 %f949, %r1206;
add.ftz.f32 %f950, %f948, %f949;
mov.b32 %r1207, %f950;
mov.u32 %r1208, 4;
shfl.sync.bfly.b32 %r1209|%p295, %r1207, %r1208, %r1200, %r1202;
mov.b32 %f951, %r1209;
add.ftz.f32 %f952, %f950, %f951;
mov.b32 %r1210, %f952;
mov.u32 %r1211, 2;
shfl.sync.bfly.b32 %r1212|%p296, %r1210, %r1211, %r1200, %r1202;
mov.b32 %f953, %r1212;
add.ftz.f32 %f954, %f952, %f953;
mov.b32 %r1213, %f954;
mov.u32 %r1214, 1;
shfl.sync.bfly.b32 %r1215|%p297, %r1213, %r1214, %r1200, %r1202;
mov.b32 %f955, %r1215;
add.ftz.f32 %f76, %f954, %f955;
mov.b32 %r1216, %f946;
shfl.sync.bfly.b32 %r1217|%p298, %r1216, %r1201, %r1200, %r1202;
mov.b32 %f956, %r1217;
add.ftz.f32 %f957, %f946, %f956;
mov.b32 %r1218, %f957;
shfl.sync.bfly.b32 %r1219|%p299, %r1218, %r1205, %r1200, %r1202;
mov.b32 %f958, %r1219;
add.ftz.f32 %f959, %f957, %f958;
mov.b32 %r1220, %f959;
shfl.sync.bfly.b32 %r1221|%p300, %r1220, %r1208, %r1200, %r1202;
mov.b32 %f960, %r1221;
add.ftz.f32 %f961, %f959, %f960;
mov.b32 %r1222, %f961;
shfl.sync.bfly.b32 %r1223|%p301, %r1222, %r1211, %r1200, %r1202;
mov.b32 %f962, %r1223;
add.ftz.f32 %f963, %f961, %f962;
mov.b32 %r1224, %f963;
shfl.sync.bfly.b32 %r1225|%p302, %r1224, %r1214, %r1200, %r1202;
mov.b32 %f964, %r1225;
add.ftz.f32 %f77, %f963, %f964;
setp.ne.s32 %p303, %r2, 2;
or.b32 %r1229, %r4, 2;
setp.ge.u32 %p304, %r1229, %r47;
or.pred %p305, %p303, %p304;
@%p305 bra $L__BB185_76;

add.ftz.f32 %f965, %f5, %f76;
add.ftz.f32 %f966, %f4, %f77;
mul.ftz.f32 %f967, %f965, %f966;
st.global.f32 [%rd38+8], %f967;

$L__BB185_76:
ld.shared.f32 %f968, [%r28+384];
add.ftz.f32 %f969, %f968, %f27;
ld.shared.f32 %f970, [%r30+384];
add.ftz.f32 %f971, %f970, %f31;
ld.shared.f32 %f972, [%r28+896];
add.ftz.f32 %f973, %f972, %f969;
ld.shared.f32 %f974, [%r30+896];
add.ftz.f32 %f975, %f974, %f971;
ld.shared.f32 %f976, [%r28+1408];
add.ftz.f32 %f977, %f976, %f973;
ld.shared.f32 %f978, [%r30+1408];
add.ftz.f32 %f979, %f978, %f975;
mov.b32 %r1230, %f977;
shfl.sync.bfly.b32 %r1234|%p306, %r1230, %r1201, %r1200, %r1202;
mov.b32 %f980, %r1234;
add.ftz.f32 %f981, %f977, %f980;
mov.b32 %r1235, %f981;
shfl.sync.bfly.b32 %r1237|%p307, %r1235, %r1205, %r1200, %r1202;
mov.b32 %f982, %r1237;
add.ftz.f32 %f983, %f981, %f982;
mov.b32 %r1238, %f983;
shfl.sync.bfly.b32 %r1240|%p308, %r1238, %r1208, %r1200, %r1202;
mov.b32 %f984, %r1240;
add.ftz.f32 %f985, %f983, %f984;
mov.b32 %r1241, %f985;
shfl.sync.bfly.b32 %r1243|%p309, %r1241, %r1211, %r1200, %r1202;
mov.b32 %f986, %r1243;
add.ftz.f32 %f987, %f985, %f986;
mov.b32 %r1244, %f987;
shfl.sync.bfly.b32 %r1246|%p310, %r1244, %r1214, %r1200, %r1202;
mov.b32 %f988, %r1246;
add.ftz.f32 %f78, %f987, %f988;
mov.b32 %r1247, %f979;
shfl.sync.bfly.b32 %r1248|%p311, %r1247, %r1201, %r1200, %r1202;
mov.b32 %f989, %r1248;
add.ftz.f32 %f990, %f979, %f989;
mov.b32 %r1249, %f990;
shfl.sync.bfly.b32 %r1250|%p312, %r1249, %r1205, %r1200, %r1202;
mov.b32 %f991, %r1250;
add.ftz.f32 %f992, %f990, %f991;
mov.b32 %r1251, %f992;
shfl.sync.bfly.b32 %r1252|%p313, %r1251, %r1208, %r1200, %r1202;
mov.b32 %f993, %r1252;
add.ftz.f32 %f994, %f992, %f993;
mov.b32 %r1253, %f994;
shfl.sync.bfly.b32 %r1254|%p314, %r1253, %r1211, %r1200, %r1202;
mov.b32 %f995, %r1254;
add.ftz.f32 %f996, %f994, %f995;
mov.b32 %r1255, %f996;
shfl.sync.bfly.b32 %r1256|%p315, %r1255, %r1214, %r1200, %r1202;
mov.b32 %f997, %r1256;
add.ftz.f32 %f79, %f996, %f997;
setp.ne.s32 %p316, %r2, 3;
or.b32 %r1260, %r4, 3;
setp.ge.u32 %p317, %r1260, %r47;
or.pred %p318, %p316, %p317;
@%p318 bra $L__BB185_78;

add.ftz.f32 %f998, %f5, %f78;
add.ftz.f32 %f999, %f4, %f79;
mul.ftz.f32 %f1000, %f998, %f999;
st.global.f32 [%rd38+12], %f1000;
bra.uni $L__BB185_78;

$L__BB185_62:
ld.shared.f32 %f705, [%r28];
add.ftz.f32 %f706, %f705, %f30;
ld.shared.f32 %f707, [%r30];
add.ftz.f32 %f708, %f707, %f34;
ld.shared.f32 %f709, [%r28+512];
add.ftz.f32 %f710, %f709, %f706;
ld.shared.f32 %f711, [%r30+512];
add.ftz.f32 %f712, %f711, %f708;
ld.shared.f32 %f713, [%r28+1024];
add.ftz.f32 %f714, %f713, %f710;
ld.shared.f32 %f715, [%r30+1024];
add.ftz.f32 %f716, %f715, %f712;
mov.b32 %r1015, %f714;
mov.u32 %r1016, 31;
mov.u32 %r1017, 16;
mov.u32 %r1018, -1;
shfl.sync.bfly.b32 %r1019|%p215, %r1015, %r1017, %r1016, %r1018;
mov.b32 %f717, %r1019;
add.ftz.f32 %f718, %f714, %f717;
mov.b32 %r1020, %f718;
mov.u32 %r1021, 8;
shfl.sync.bfly.b32 %r1022|%p216, %r1020, %r1021, %r1016, %r1018;
mov.b32 %f719, %r1022;
add.ftz.f32 %f720, %f718, %f719;
mov.b32 %r1023, %f720;
mov.u32 %r1024, 4;
shfl.sync.bfly.b32 %r1025|%p217, %r1023, %r1024, %r1016, %r1018;
mov.b32 %f721, %r1025;
add.ftz.f32 %f722, %f720, %f721;
mov.b32 %r1026, %f722;
mov.u32 %r1027, 2;
shfl.sync.bfly.b32 %r1028|%p218, %r1026, %r1027, %r1016, %r1018;
mov.b32 %f723, %r1028;
add.ftz.f32 %f724, %f722, %f723;
mov.b32 %r1029, %f724;
mov.u32 %r1030, 1;
shfl.sync.bfly.b32 %r1031|%p219, %r1029, %r1030, %r1016, %r1018;
mov.b32 %f725, %r1031;
add.ftz.f32 %f64, %f724, %f725;
mov.b32 %r1032, %f716;
shfl.sync.bfly.b32 %r1033|%p220, %r1032, %r1017, %r1016, %r1018;
mov.b32 %f726, %r1033;
add.ftz.f32 %f727, %f716, %f726;
mov.b32 %r1034, %f727;
shfl.sync.bfly.b32 %r1035|%p221, %r1034, %r1021, %r1016, %r1018;
mov.b32 %f728, %r1035;
add.ftz.f32 %f729, %f727, %f728;
mov.b32 %r1036, %f729;
shfl.sync.bfly.b32 %r1037|%p222, %r1036, %r1024, %r1016, %r1018;
mov.b32 %f730, %r1037;
add.ftz.f32 %f731, %f729, %f730;
mov.b32 %r1038, %f731;
shfl.sync.bfly.b32 %r1039|%p223, %r1038, %r1027, %r1016, %r1018;
mov.b32 %f732, %r1039;
add.ftz.f32 %f733, %f731, %f732;
mov.b32 %r1040, %f733;
shfl.sync.bfly.b32 %r1041|%p224, %r1040, %r1030, %r1016, %r1018;
mov.b32 %f734, %r1041;
add.ftz.f32 %f65, %f733, %f734;
setp.ne.s32 %p225, %r2, 0;
setp.ge.u32 %p226, %r4, %r47;
or.pred %p227, %p225, %p226;
@%p227 bra $L__BB185_64;

add.ftz.f32 %f735, %f5, %f64;
add.ftz.f32 %f736, %f4, %f65;
mul.ftz.f32 %f737, %f736, 0f3F000000;
mul.ftz.f32 %f738, %f736, 0f3F4C422A;
mul.ftz.f32 %f739, %f736, 0f3D372713;
fma.rn.ftz.f32 %f740, %f736, %f739, 0f3F800000;
mul.ftz.f32 %f741, %f738, %f740;
tanh.approx.f32 %f742, %f741;
add.ftz.f32 %f743, %f742, 0f3F800000;
mul.ftz.f32 %f744, %f737, %f743;
mul.ftz.f32 %f745, %f735, %f744;
st.global.f32 [%rd38], %f745;

$L__BB185_64:
ld.shared.f32 %f746, [%r28+128];
add.ftz.f32 %f747, %f746, %f29;
ld.shared.f32 %f748, [%r30+128];
add.ftz.f32 %f749, %f748, %f33;
ld.shared.f32 %f750, [%r28+640];
add.ftz.f32 %f751, %f750, %f747;
ld.shared.f32 %f752, [%r30+640];
add.ftz.f32 %f753, %f752, %f749;
ld.shared.f32 %f754, [%r28+1152];
add.ftz.f32 %f755, %f754, %f751;
ld.shared.f32 %f756, [%r30+1152];
add.ftz.f32 %f757, %f756, %f753;
mov.b32 %r1045, %f755;
shfl.sync.bfly.b32 %r1049|%p228, %r1045, %r1017, %r1016, %r1018;
mov.b32 %f758, %r1049;
add.ftz.f32 %f759, %f755, %f758;
mov.b32 %r1050, %f759;
shfl.sync.bfly.b32 %r1052|%p229, %r1050, %r1021, %r1016, %r1018;
mov.b32 %f760, %r1052;
add.ftz.f32 %f761, %f759, %f760;
mov.b32 %r1053, %f761;
shfl.sync.bfly.b32 %r1055|%p230, %r1053, %r1024, %r1016, %r1018;
mov.b32 %f762, %r1055;
add.ftz.f32 %f763, %f761, %f762;
mov.b32 %r1056, %f763;
shfl.sync.bfly.b32 %r1058|%p231, %r1056, %r1027, %r1016, %r1018;
mov.b32 %f764, %r1058;
add.ftz.f32 %f765, %f763, %f764;
mov.b32 %r1059, %f765;
shfl.sync.bfly.b32 %r1061|%p232, %r1059, %r1030, %r1016, %r1018;
mov.b32 %f766, %r1061;
add.ftz.f32 %f66, %f765, %f766;
mov.b32 %r1062, %f757;
shfl.sync.bfly.b32 %r1063|%p233, %r1062, %r1017, %r1016, %r1018;
mov.b32 %f767, %r1063;
add.ftz.f32 %f768, %f757, %f767;
mov.b32 %r1064, %f768;
shfl.sync.bfly.b32 %r1065|%p234, %r1064, %r1021, %r1016, %r1018;
mov.b32 %f769, %r1065;
add.ftz.f32 %f770, %f768, %f769;
mov.b32 %r1066, %f770;
shfl.sync.bfly.b32 %r1067|%p235, %r1066, %r1024, %r1016, %r1018;
mov.b32 %f771, %r1067;
add.ftz.f32 %f772, %f770, %f771;
mov.b32 %r1068, %f772;
shfl.sync.bfly.b32 %r1069|%p236, %r1068, %r1027, %r1016, %r1018;
mov.b32 %f773, %r1069;
add.ftz.f32 %f774, %f772, %f773;
mov.b32 %r1070, %f774;
shfl.sync.bfly.b32 %r1071|%p237, %r1070, %r1030, %r1016, %r1018;
mov.b32 %f775, %r1071;
add.ftz.f32 %f67, %f774, %f775;
setp.ne.s32 %p238, %r2, 1;
or.b32 %r1075, %r4, 1;
setp.ge.u32 %p239, %r1075, %r47;
or.pred %p240, %p238, %p239;
@%p240 bra $L__BB185_66;

add.ftz.f32 %f776, %f5, %f66;
add.ftz.f32 %f777, %f4, %f67;
mul.ftz.f32 %f778, %f777, 0f3F000000;
mul.ftz.f32 %f779, %f777, 0f3F4C422A;
mul.ftz.f32 %f780, %f777, 0f3D372713;
fma.rn.ftz.f32 %f781, %f777, %f780, 0f3F800000;
mul.ftz.f32 %f782, %f779, %f781;
tanh.approx.f32 %f783, %f782;
add.ftz.f32 %f784, %f783, 0f3F800000;
mul.ftz.f32 %f785, %f778, %f784;
mul.ftz.f32 %f786, %f776, %f785;
st.global.f32 [%rd38+4], %f786;

$L__BB185_66:
ld.shared.f32 %f787, [%r28+256];
add.ftz.f32 %f788, %f787, %f28;
ld.shared.f32 %f789, [%r30+256];
add.ftz.f32 %f790, %f789, %f32;
ld.shared.f32 %f791, [%r28+768];
add.ftz.f32 %f792, %f791, %f788;
ld.shared.f32 %f793, [%r30+768];
add.ftz.f32 %f794, %f793, %f790;
ld.shared.f32 %f795, [%r28+1280];
add.ftz.f32 %f796, %f795, %f792;
ld.shared.f32 %f797, [%r30+1280];
add.ftz.f32 %f798, %f797, %f794;
mov.b32 %r1076, %f796;
mov.u32 %r1077, 31;
mov.u32 %r1078, 16;
mov.u32 %r1079, -1;
shfl.sync.bfly.b32 %r1080|%p241, %r1076, %r1078, %r1077, %r1079;
mov.b32 %f799, %r1080;
add.ftz.f32 %f800, %f796, %f799;
mov.b32 %r1081, %f800;
mov.u32 %r1082, 8;
shfl.sync.bfly.b32 %r1083|%p242, %r1081, %r1082, %r1077, %r1079;
mov.b32 %f801, %r1083;
add.ftz.f32 %f802, %f800, %f801;
mov.b32 %r1084, %f802;
mov.u32 %r1085, 4;
shfl.sync.bfly.b32 %r1086|%p243, %r1084, %r1085, %r1077, %r1079;
mov.b32 %f803, %r1086;
add.ftz.f32 %f804, %f802, %f803;
mov.b32 %r1087, %f804;
mov.u32 %r1088, 2;
shfl.sync.bfly.b32 %r1089|%p244, %r1087, %r1088, %r1077, %r1079;
mov.b32 %f805, %r1089;
add.ftz.f32 %f806, %f804, %f805;
mov.b32 %r1090, %f806;
mov.u32 %r1091, 1;
shfl.sync.bfly.b32 %r1092|%p245, %r1090, %r1091, %r1077, %r1079;
mov.b32 %f807, %r1092;
add.ftz.f32 %f68, %f806, %f807;
mov.b32 %r1093, %f798;
shfl.sync.bfly.b32 %r1094|%p246, %r1093, %r1078, %r1077, %r1079;
mov.b32 %f808, %r1094;
add.ftz.f32 %f809, %f798, %f808;
mov.b32 %r1095, %f809;
shfl.sync.bfly.b32 %r1096|%p247, %r1095, %r1082, %r1077, %r1079;
mov.b32 %f810, %r1096;
add.ftz.f32 %f811, %f809, %f810;
mov.b32 %r1097, %f811;
shfl.sync.bfly.b32 %r1098|%p248, %r1097, %r1085, %r1077, %r1079;
mov.b32 %f812, %r1098;
add.ftz.f32 %f813, %f811, %f812;
mov.b32 %r1099, %f813;
shfl.sync.bfly.b32 %r1100|%p249, %r1099, %r1088, %r1077, %r1079;
mov.b32 %f814, %r1100;
add.ftz.f32 %f815, %f813, %f814;
mov.b32 %r1101, %f815;
shfl.sync.bfly.b32 %r1102|%p250, %r1101, %r1091, %r1077, %r1079;
mov.b32 %f816, %r1102;
add.ftz.f32 %f69, %f815, %f816;
setp.ne.s32 %p251, %r2, 2;
or.b32 %r1106, %r4, 2;
setp.ge.u32 %p252, %r1106, %r47;
or.pred %p253, %p251, %p252;
@%p253 bra $L__BB185_68;

add.ftz.f32 %f817, %f5, %f68;
add.ftz.f32 %f818, %f4, %f69;
mul.ftz.f32 %f819, %f818, 0f3F000000;
mul.ftz.f32 %f820, %f818, 0f3F4C422A;
mul.ftz.f32 %f821, %f818, 0f3D372713;
fma.rn.ftz.f32 %f822, %f818, %f821, 0f3F800000;
mul.ftz.f32 %f823, %f820, %f822;
tanh.approx.f32 %f824, %f823;
add.ftz.f32 %f825, %f824, 0f3F800000;
mul.ftz.f32 %f826, %f819, %f825;
mul.ftz.f32 %f827, %f817, %f826;
st.global.f32 [%rd38+8], %f827;

$L__BB185_68:
ld.shared.f32 %f828, [%r28+384];
add.ftz.f32 %f829, %f828, %f27;
ld.shared.f32 %f830, [%r30+384];
add.ftz.f32 %f831, %f830, %f31;
ld.shared.f32 %f832, [%r28+896];
add.ftz.f32 %f833, %f832, %f829;
ld.shared.f32 %f834, [%r30+896];
add.ftz.f32 %f835, %f834, %f831;
ld.shared.f32 %f836, [%r28+1408];
add.ftz.f32 %f837, %f836, %f833;
ld.shared.f32 %f838, [%r30+1408];
add.ftz.f32 %f839, %f838, %f835;
mov.b32 %r1107, %f837;
shfl.sync.bfly.b32 %r1111|%p254, %r1107, %r1078, %r1077, %r1079;
mov.b32 %f840, %r1111;
add.ftz.f32 %f841, %f837, %f840;
mov.b32 %r1112, %f841;
shfl.sync.bfly.b32 %r1114|%p255, %r1112, %r1082, %r1077, %r1079;
mov.b32 %f842, %r1114;
add.ftz.f32 %f843, %f841, %f842;
mov.b32 %r1115, %f843;
shfl.sync.bfly.b32 %r1117|%p256, %r1115, %r1085, %r1077, %r1079;
mov.b32 %f844, %r1117;
add.ftz.f32 %f845, %f843, %f844;
mov.b32 %r1118, %f845;
shfl.sync.bfly.b32 %r1120|%p257, %r1118, %r1088, %r1077, %r1079;
mov.b32 %f846, %r1120;
add.ftz.f32 %f847, %f845, %f846;
mov.b32 %r1121, %f847;
shfl.sync.bfly.b32 %r1123|%p258, %r1121, %r1091, %r1077, %r1079;
mov.b32 %f848, %r1123;
add.ftz.f32 %f70, %f847, %f848;
mov.b32 %r1124, %f839;
shfl.sync.bfly.b32 %r1125|%p259, %r1124, %r1078, %r1077, %r1079;
mov.b32 %f849, %r1125;
add.ftz.f32 %f850, %f839, %f849;
mov.b32 %r1126, %f850;
shfl.sync.bfly.b32 %r1127|%p260, %r1126, %r1082, %r1077, %r1079;
mov.b32 %f851, %r1127;
add.ftz.f32 %f852, %f850, %f851;
mov.b32 %r1128, %f852;
shfl.sync.bfly.b32 %r1129|%p261, %r1128, %r1085, %r1077, %r1079;
mov.b32 %f853, %r1129;
add.ftz.f32 %f854, %f852, %f853;
mov.b32 %r1130, %f854;
shfl.sync.bfly.b32 %r1131|%p262, %r1130, %r1088, %r1077, %r1079;
mov.b32 %f855, %r1131;
add.ftz.f32 %f856, %f854, %f855;
mov.b32 %r1132, %f856;
shfl.sync.bfly.b32 %r1133|%p263, %r1132, %r1091, %r1077, %r1079;
mov.b32 %f857, %r1133;
add.ftz.f32 %f71, %f856, %f857;
setp.ne.s32 %p264, %r2, 3;
or.b32 %r1137, %r4, 3;
setp.ge.u32 %p265, %r1137, %r47;
or.pred %p266, %p264, %p265;
@%p266 bra $L__BB185_78;

add.ftz.f32 %f858, %f5, %f70;
add.ftz.f32 %f859, %f4, %f71;
mul.ftz.f32 %f860, %f859, 0f3F000000;
mul.ftz.f32 %f861, %f859, 0f3F4C422A;
mul.ftz.f32 %f862, %f859, 0f3D372713;
fma.rn.ftz.f32 %f863, %f859, %f862, 0f3F800000;
mul.ftz.f32 %f864, %f861, %f863;
tanh.approx.f32 %f865, %f864;
add.ftz.f32 %f866, %f865, 0f3F800000;
mul.ftz.f32 %f867, %f860, %f866;
mul.ftz.f32 %f868, %f858, %f867;
st.global.f32 [%rd38+12], %f868;
bra.uni $L__BB185_78;

$L__BB185_54:
ld.shared.f32 %f381, [%r28];
add.ftz.f32 %f382, %f381, %f30;
ld.shared.f32 %f383, [%r30];
add.ftz.f32 %f384, %f383, %f34;
ld.shared.f32 %f385, [%r28+512];
add.ftz.f32 %f386, %f385, %f382;
ld.shared.f32 %f387, [%r30+512];
add.ftz.f32 %f388, %f387, %f384;
ld.shared.f32 %f389, [%r28+1024];
add.ftz.f32 %f390, %f389, %f386;
ld.shared.f32 %f391, [%r30+1024];
add.ftz.f32 %f392, %f391, %f388;
mov.b32 %r769, %f390;
mov.u32 %r770, 31;
mov.u32 %r771, 16;
mov.u32 %r772, -1;
shfl.sync.bfly.b32 %r773|%p111, %r769, %r771, %r770, %r772;
mov.b32 %f393, %r773;
add.ftz.f32 %f394, %f390, %f393;
mov.b32 %r774, %f394;
mov.u32 %r775, 8;
shfl.sync.bfly.b32 %r776|%p112, %r774, %r775, %r770, %r772;
mov.b32 %f395, %r776;
add.ftz.f32 %f396, %f394, %f395;
mov.b32 %r777, %f396;
mov.u32 %r778, 4;
shfl.sync.bfly.b32 %r779|%p113, %r777, %r778, %r770, %r772;
mov.b32 %f397, %r779;
add.ftz.f32 %f398, %f396, %f397;
mov.b32 %r780, %f398;
mov.u32 %r781, 2;
shfl.sync.bfly.b32 %r782|%p114, %r780, %r781, %r770, %r772;
mov.b32 %f399, %r782;
add.ftz.f32 %f400, %f398, %f399;
mov.b32 %r783, %f400;
mov.u32 %r784, 1;
shfl.sync.bfly.b32 %r785|%p115, %r783, %r784, %r770, %r772;
mov.b32 %f401, %r785;
add.ftz.f32 %f48, %f400, %f401;
mov.b32 %r786, %f392;
shfl.sync.bfly.b32 %r787|%p116, %r786, %r771, %r770, %r772;
mov.b32 %f402, %r787;
add.ftz.f32 %f403, %f392, %f402;
mov.b32 %r788, %f403;
shfl.sync.bfly.b32 %r789|%p117, %r788, %r775, %r770, %r772;
mov.b32 %f404, %r789;
add.ftz.f32 %f405, %f403, %f404;
mov.b32 %r790, %f405;
shfl.sync.bfly.b32 %r791|%p118, %r790, %r778, %r770, %r772;
mov.b32 %f406, %r791;
add.ftz.f32 %f407, %f405, %f406;
mov.b32 %r792, %f407;
shfl.sync.bfly.b32 %r793|%p119, %r792, %r781, %r770, %r772;
mov.b32 %f408, %r793;
add.ftz.f32 %f409, %f407, %f408;
mov.b32 %r794, %f409;
shfl.sync.bfly.b32 %r795|%p120, %r794, %r784, %r770, %r772;
mov.b32 %f410, %r795;
add.ftz.f32 %f49, %f409, %f410;
setp.ne.s32 %p121, %r2, 0;
setp.ge.u32 %p122, %r4, %r47;
or.pred %p123, %p121, %p122;
@%p123 bra $L__BB185_56;

add.ftz.f32 %f411, %f5, %f48;
add.ftz.f32 %f412, %f4, %f49;
mov.f32 %f413, 0f40E00000;
min.ftz.f32 %f414, %f412, %f413;
min.ftz.f32 %f415, %f411, %f413;
mov.f32 %f416, 0fC0E00000;
max.ftz.f32 %f417, %f415, %f416;
mul.ftz.f32 %f418, %f414, 0f3FD9DB23;
mul.ftz.f32 %f419, %f418, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f420, %f419;
add.ftz.f32 %f421, %f420, 0f3F800000;
div.approx.ftz.f32 %f422, %f414, %f421;
add.ftz.f32 %f423, %f417, 0f3F800000;
mul.ftz.f32 %f424, %f423, %f422;
st.global.f32 [%rd38], %f424;

$L__BB185_56:
ld.shared.f32 %f425, [%r28+128];
add.ftz.f32 %f426, %f425, %f29;
ld.shared.f32 %f427, [%r30+128];
add.ftz.f32 %f428, %f427, %f33;
ld.shared.f32 %f429, [%r28+640];
add.ftz.f32 %f430, %f429, %f426;
ld.shared.f32 %f431, [%r30+640];
add.ftz.f32 %f432, %f431, %f428;
ld.shared.f32 %f433, [%r28+1152];
add.ftz.f32 %f434, %f433, %f430;
ld.shared.f32 %f435, [%r30+1152];
add.ftz.f32 %f436, %f435, %f432;
mov.b32 %r799, %f434;
shfl.sync.bfly.b32 %r803|%p124, %r799, %r771, %r770, %r772;
mov.b32 %f437, %r803;
add.ftz.f32 %f438, %f434, %f437;
mov.b32 %r804, %f438;
shfl.sync.bfly.b32 %r806|%p125, %r804, %r775, %r770, %r772;
mov.b32 %f439, %r806;
add.ftz.f32 %f440, %f438, %f439;
mov.b32 %r807, %f440;
shfl.sync.bfly.b32 %r809|%p126, %r807, %r778, %r770, %r772;
mov.b32 %f441, %r809;
add.ftz.f32 %f442, %f440, %f441;
mov.b32 %r810, %f442;
shfl.sync.bfly.b32 %r812|%p127, %r810, %r781, %r770, %r772;
mov.b32 %f443, %r812;
add.ftz.f32 %f444, %f442, %f443;
mov.b32 %r813, %f444;
shfl.sync.bfly.b32 %r815|%p128, %r813, %r784, %r770, %r772;
mov.b32 %f445, %r815;
add.ftz.f32 %f50, %f444, %f445;
mov.b32 %r816, %f436;
shfl.sync.bfly.b32 %r817|%p129, %r816, %r771, %r770, %r772;
mov.b32 %f446, %r817;
add.ftz.f32 %f447, %f436, %f446;
mov.b32 %r818, %f447;
shfl.sync.bfly.b32 %r819|%p130, %r818, %r775, %r770, %r772;
mov.b32 %f448, %r819;
add.ftz.f32 %f449, %f447, %f448;
mov.b32 %r820, %f449;
shfl.sync.bfly.b32 %r821|%p131, %r820, %r778, %r770, %r772;
mov.b32 %f450, %r821;
add.ftz.f32 %f451, %f449, %f450;
mov.b32 %r822, %f451;
shfl.sync.bfly.b32 %r823|%p132, %r822, %r781, %r770, %r772;
mov.b32 %f452, %r823;
add.ftz.f32 %f453, %f451, %f452;
mov.b32 %r824, %f453;
shfl.sync.bfly.b32 %r825|%p133, %r824, %r784, %r770, %r772;
mov.b32 %f454, %r825;
add.ftz.f32 %f51, %f453, %f454;
setp.ne.s32 %p134, %r2, 1;
or.b32 %r829, %r4, 1;
setp.ge.u32 %p135, %r829, %r47;
or.pred %p136, %p134, %p135;
@%p136 bra $L__BB185_58;

add.ftz.f32 %f455, %f5, %f50;
add.ftz.f32 %f456, %f4, %f51;
mov.f32 %f457, 0f40E00000;
min.ftz.f32 %f458, %f456, %f457;
min.ftz.f32 %f459, %f455, %f457;
mov.f32 %f460, 0fC0E00000;
max.ftz.f32 %f461, %f459, %f460;
mul.ftz.f32 %f462, %f458, 0f3FD9DB23;
mul.ftz.f32 %f463, %f462, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f464, %f463;
add.ftz.f32 %f465, %f464, 0f3F800000;
div.approx.ftz.f32 %f466, %f458, %f465;
add.ftz.f32 %f467, %f461, 0f3F800000;
mul.ftz.f32 %f468, %f467, %f466;
st.global.f32 [%rd38+4], %f468;

$L__BB185_58:
ld.shared.f32 %f469, [%r28+256];
add.ftz.f32 %f470, %f469, %f28;
ld.shared.f32 %f471, [%r30+256];
add.ftz.f32 %f472, %f471, %f32;
ld.shared.f32 %f473, [%r28+768];
add.ftz.f32 %f474, %f473, %f470;
ld.shared.f32 %f475, [%r30+768];
add.ftz.f32 %f476, %f475, %f472;
ld.shared.f32 %f477, [%r28+1280];
add.ftz.f32 %f478, %f477, %f474;
ld.shared.f32 %f479, [%r30+1280];
add.ftz.f32 %f480, %f479, %f476;
mov.b32 %r830, %f478;
mov.u32 %r831, 31;
mov.u32 %r832, 16;
mov.u32 %r833, -1;
shfl.sync.bfly.b32 %r834|%p137, %r830, %r832, %r831, %r833;
mov.b32 %f481, %r834;
add.ftz.f32 %f482, %f478, %f481;
mov.b32 %r835, %f482;
mov.u32 %r836, 8;
shfl.sync.bfly.b32 %r837|%p138, %r835, %r836, %r831, %r833;
mov.b32 %f483, %r837;
add.ftz.f32 %f484, %f482, %f483;
mov.b32 %r838, %f484;
mov.u32 %r839, 4;
shfl.sync.bfly.b32 %r840|%p139, %r838, %r839, %r831, %r833;
mov.b32 %f485, %r840;
add.ftz.f32 %f486, %f484, %f485;
mov.b32 %r841, %f486;
mov.u32 %r842, 2;
shfl.sync.bfly.b32 %r843|%p140, %r841, %r842, %r831, %r833;
mov.b32 %f487, %r843;
add.ftz.f32 %f488, %f486, %f487;
mov.b32 %r844, %f488;
mov.u32 %r845, 1;
shfl.sync.bfly.b32 %r846|%p141, %r844, %r845, %r831, %r833;
mov.b32 %f489, %r846;
add.ftz.f32 %f52, %f488, %f489;
mov.b32 %r847, %f480;
shfl.sync.bfly.b32 %r848|%p142, %r847, %r832, %r831, %r833;
mov.b32 %f490, %r848;
add.ftz.f32 %f491, %f480, %f490;
mov.b32 %r849, %f491;
shfl.sync.bfly.b32 %r850|%p143, %r849, %r836, %r831, %r833;
mov.b32 %f492, %r850;
add.ftz.f32 %f493, %f491, %f492;
mov.b32 %r851, %f493;
shfl.sync.bfly.b32 %r852|%p144, %r851, %r839, %r831, %r833;
mov.b32 %f494, %r852;
add.ftz.f32 %f495, %f493, %f494;
mov.b32 %r853, %f495;
shfl.sync.bfly.b32 %r854|%p145, %r853, %r842, %r831, %r833;
mov.b32 %f496, %r854;
add.ftz.f32 %f497, %f495, %f496;
mov.b32 %r855, %f497;
shfl.sync.bfly.b32 %r856|%p146, %r855, %r845, %r831, %r833;
mov.b32 %f498, %r856;
add.ftz.f32 %f53, %f497, %f498;
setp.ne.s32 %p147, %r2, 2;
or.b32 %r860, %r4, 2;
setp.ge.u32 %p148, %r860, %r47;
or.pred %p149, %p147, %p148;
@%p149 bra $L__BB185_60;

add.ftz.f32 %f499, %f5, %f52;
add.ftz.f32 %f500, %f4, %f53;
mov.f32 %f501, 0f40E00000;
min.ftz.f32 %f502, %f500, %f501;
min.ftz.f32 %f503, %f499, %f501;
mov.f32 %f504, 0fC0E00000;
max.ftz.f32 %f505, %f503, %f504;
mul.ftz.f32 %f506, %f502, 0f3FD9DB23;
mul.ftz.f32 %f507, %f506, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f508, %f507;
add.ftz.f32 %f509, %f508, 0f3F800000;
div.approx.ftz.f32 %f510, %f502, %f509;
add.ftz.f32 %f511, %f505, 0f3F800000;
mul.ftz.f32 %f512, %f511, %f510;
st.global.f32 [%rd38+8], %f512;

$L__BB185_60:
ld.shared.f32 %f513, [%r28+384];
add.ftz.f32 %f514, %f513, %f27;
ld.shared.f32 %f515, [%r30+384];
add.ftz.f32 %f516, %f515, %f31;
ld.shared.f32 %f517, [%r28+896];
add.ftz.f32 %f518, %f517, %f514;
ld.shared.f32 %f519, [%r30+896];
add.ftz.f32 %f520, %f519, %f516;
ld.shared.f32 %f521, [%r28+1408];
add.ftz.f32 %f522, %f521, %f518;
ld.shared.f32 %f523, [%r30+1408];
add.ftz.f32 %f524, %f523, %f520;
mov.b32 %r861, %f522;
shfl.sync.bfly.b32 %r865|%p150, %r861, %r832, %r831, %r833;
mov.b32 %f525, %r865;
add.ftz.f32 %f526, %f522, %f525;
mov.b32 %r866, %f526;
shfl.sync.bfly.b32 %r868|%p151, %r866, %r836, %r831, %r833;
mov.b32 %f527, %r868;
add.ftz.f32 %f528, %f526, %f527;
mov.b32 %r869, %f528;
shfl.sync.bfly.b32 %r871|%p152, %r869, %r839, %r831, %r833;
mov.b32 %f529, %r871;
add.ftz.f32 %f530, %f528, %f529;
mov.b32 %r872, %f530;
shfl.sync.bfly.b32 %r874|%p153, %r872, %r842, %r831, %r833;
mov.b32 %f531, %r874;
add.ftz.f32 %f532, %f530, %f531;
mov.b32 %r875, %f532;
shfl.sync.bfly.b32 %r877|%p154, %r875, %r845, %r831, %r833;
mov.b32 %f533, %r877;
add.ftz.f32 %f54, %f532, %f533;
mov.b32 %r878, %f524;
shfl.sync.bfly.b32 %r879|%p155, %r878, %r832, %r831, %r833;
mov.b32 %f534, %r879;
add.ftz.f32 %f535, %f524, %f534;
mov.b32 %r880, %f535;
shfl.sync.bfly.b32 %r881|%p156, %r880, %r836, %r831, %r833;
mov.b32 %f536, %r881;
add.ftz.f32 %f537, %f535, %f536;
mov.b32 %r882, %f537;
shfl.sync.bfly.b32 %r883|%p157, %r882, %r839, %r831, %r833;
mov.b32 %f538, %r883;
add.ftz.f32 %f539, %f537, %f538;
mov.b32 %r884, %f539;
shfl.sync.bfly.b32 %r885|%p158, %r884, %r842, %r831, %r833;
mov.b32 %f540, %r885;
add.ftz.f32 %f541, %f539, %f540;
mov.b32 %r886, %f541;
shfl.sync.bfly.b32 %r887|%p159, %r886, %r845, %r831, %r833;
mov.b32 %f542, %r887;
add.ftz.f32 %f55, %f541, %f542;
setp.ne.s32 %p160, %r2, 3;
or.b32 %r891, %r4, 3;
setp.ge.u32 %p161, %r891, %r47;
or.pred %p162, %p160, %p161;
@%p162 bra $L__BB185_78;

add.ftz.f32 %f543, %f5, %f54;
add.ftz.f32 %f544, %f4, %f55;
mov.f32 %f545, 0f40E00000;
min.ftz.f32 %f546, %f544, %f545;
min.ftz.f32 %f547, %f543, %f545;
mov.f32 %f548, 0fC0E00000;
max.ftz.f32 %f549, %f547, %f548;
mul.ftz.f32 %f550, %f546, 0f3FD9DB23;
mul.ftz.f32 %f551, %f550, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f552, %f551;
add.ftz.f32 %f553, %f552, 0f3F800000;
div.approx.ftz.f32 %f554, %f546, %f553;
add.ftz.f32 %f555, %f549, 0f3F800000;
mul.ftz.f32 %f556, %f555, %f554;
st.global.f32 [%rd38+12], %f556;
bra.uni $L__BB185_78;

}
.entry _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<38>;
.reg .b16 %rs<29>;
.reg .f32 %f<121>;
.reg .b32 %r<378>;
.reg .b64 %rd<59>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

ld.param.u64 %rd14, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
mov.u32 %r1, %tid.y;
shl.b32 %r39, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r39, %r2;
mov.u32 %r40, %ctaid.x;
shl.b32 %r4, %r40, 2;
shr.u32 %r5, %r38, 8;
mov.u32 %r375, %ctaid.y;
setp.eq.s64 %p1, %rd16, 0;
@%p1 bra $L__BB186_2;

cvta.to.global.u64 %rd18, %rd16;
mul.wide.u32 %rd19, %r375, 4;
add.s64 %rd20, %rd18, %rd19;
ld.global.nc.u32 %r374, [%rd20];
mul.hi.u32 %r41, %r375, %r21;
add.s32 %r42, %r41, %r375;
shr.u32 %r43, %r42, %r22;
mul.lo.s32 %r44, %r43, %r23;
sub.s32 %r375, %r375, %r44;
bra.uni $L__BB186_3;

$L__BB186_2:
mul.hi.u32 %r45, %r375, %r26;
add.s32 %r46, %r45, %r375;
shr.u32 %r374, %r46, %r27;

$L__BB186_3:
shr.s32 %r47, %r3, 31;
shr.u32 %r48, %r47, 27;
add.s32 %r49, %r3, %r48;
shr.s32 %r377, %r49, 5;
setp.ge.s32 %p2, %r377, %r5;
mov.f32 %f117, 0f00000000;
mov.f32 %f118, %f117;
mov.f32 %f119, %f117;
mov.f32 %f120, %f117;
@%p2 bra $L__BB186_6;

mov.u32 %r50, %ctaid.z;
and.b32 %r54, %r49, -32;
sub.s32 %r55, %r3, %r54;
shr.s32 %r56, %r55, 31;
shr.u32 %r57, %r56, 28;
add.s32 %r58, %r55, %r57;
shr.s32 %r59, %r58, 4;
shl.b32 %r60, %r59, 2;
and.b32 %r61, %r58, -16;
sub.s32 %r62, %r55, %r61;
shr.s32 %r63, %r62, 31;
shr.u32 %r64, %r63, 29;
add.s32 %r65, %r62, %r64;
shr.s32 %r66, %r65, 3;
add.s32 %r67, %r60, %r66;
shl.b32 %r68, %r59, 3;
shr.u32 %r69, %r63, 30;
add.s32 %r70, %r62, %r69;
shr.s32 %r71, %r70, 2;
add.s32 %r72, %r68, %r71;
shl.b32 %r13, %r66, 1;
shl.b32 %r73, %r55, 1;
shr.u32 %r74, %r56, 29;
add.s32 %r75, %r55, %r74;
and.b32 %r76, %r75, -8;
sub.s32 %r77, %r55, %r76;
add.s32 %r78, %r68, %r77;
shl.b32 %r79, %r78, 1;
cvt.s64.s32 %rd21, %r72;
cvt.s64.s32 %rd1, %r77;
cvt.s64.s32 %rd2, %r67;
shl.b32 %r376, %r377, 3;
mad.lo.s32 %r80, %r4, %r24, %r377;
mul.hi.u32 %r81, %r50, %r32;
add.s32 %r82, %r81, %r50;
shr.u32 %r83, %r82, %r33;
mad.lo.s32 %r84, %r83, %r35, %r80;
mad.lo.s32 %r85, %r374, %r29, %r84;
mul.wide.s32 %rd3, %r85, 210;
mul.wide.s32 %rd22, %r73, 2;
add.s64 %rd23, %rd3, %rd22;
add.s64 %rd4, %rd23, 2;
mul.wide.s32 %rd24, %r79, 2;
add.s64 %rd25, %rd3, %rd24;
add.s64 %rd5, %rd25, 130;
add.s64 %rd26, %rd3, %rd21;
add.s64 %rd6, %rd26, 196;
mul.wide.s32 %rd7, %r24, 210;
cvta.to.global.u64 %rd8, %rd15;
cvta.to.global.u64 %rd58, %rd14;
mul.lo.s32 %r86, %r375, %r30;
cvt.u64.u32 %rd27, %r86;
mul.lo.s32 %r87, %r50, %r36;
cvt.u64.u32 %rd28, %r87;
add.s64 %rd10, %rd27, %rd28;
mov.f32 %f117, 0f00000000;
shl.b64 %rd38, %rd1, 2;

$L__BB186_5:
cvt.u64.u32 %rd29, %r376;
add.s64 %rd30, %rd10, %rd29;
add.s64 %rd31, %rd58, %rd4;
add.s64 %rd32, %rd31, -2;
ld.global.nc.u16 %rs5, [%rd31+-2];
ld.global.nc.u16 %rs6, [%rd31];
mov.b32 %r160, {%rs5, %rs6};
add.s64 %rd33, %rd58, %rd5;
add.s64 %rd34, %rd33, -2;
ld.global.nc.u16 %rs7, [%rd33+-2];
ld.global.nc.u16 %rs8, [%rd33];
mov.b32 %r161, {%rs7, %rs8};
shr.s32 %r162, %r161, %r13;
add.s64 %rd35, %rd30, %rd2;
mul.lo.s64 %rd36, %rd35, 36;
add.s64 %rd37, %rd8, %rd36;
add.s64 %rd39, %rd37, %rd38;
ld.global.nc.u32 %r96, [%rd39+4];
ld.global.nc.u32 %r88, [%rd37];

{.reg .f16 low,high;
mov.b32 {low,high},%r88;
cvt.f32.f16 %f25, low;}


ld.global.nc.u32 %r104, [%rd39+76];
ld.global.nc.u32 %r89, [%rd37+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r89;
cvt.f32.f16 %f26, low;}


add.s64 %rd40, %rd58, %rd3;
add.s64 %rd41, %rd40, 208;
ld.global.nc.u16 %rs1, [%rd40+208];

{ cvt.f32.f16 %f27, %rs1;}


add.s64 %rd42, %rd58, %rd6;
add.s64 %rd43, %rd42, -4;
ld.global.nc.u8 %rs9, [%rd42+-4];
cvt.u32.u16 %r163, %rs9;
cvt.s32.s8 %r164, %r163;
and.b32 %r165, %r160, 252645135;
shl.b32 %r166, %r162, 4;
and.b32 %r167, %r166, 808464432;
or.b32 %r92, %r167, %r165;
or.b32 %r168, %r92, -2139062144;
add.s32 %r169, %r168, -538976288;
xor.b32 %r170, %r169, -2139062144;
xor.b32 %r171, %r92, 538976288;
xor.b32 %r172, %r169, %r92;
and.b32 %r90, %r172, %r171;

prmt.b32 %r90, %r90, 0, 0xba98;


prmt.b32 %r92, %r92, 0, 0xba98;

xor.b32 %r173, %r92, 2139062143;
not.b32 %r174, %r90;
and.b32 %r175, %r170, %r174;
and.b32 %r176, %r173, %r90;
or.b32 %r95, %r176, %r175;
mov.u32 %r159, 0;

dp4a.s32.s32 %r94, %r95, %r96, %r159;

mul.lo.s32 %r177, %r94, %r164;
cvt.rn.f32.s32 %f37, %r177;
fma.rn.ftz.f32 %f38, %f25, %f37, 0f00000000;
ld.global.nc.u8 %rs10, [%rd42];
cvt.u32.u16 %r178, %rs10;
cvt.s32.s8 %r179, %r178;
shr.u32 %r180, %r160, 4;
and.b32 %r181, %r180, 252645135;
and.b32 %r182, %r162, 808464432;
or.b32 %r100, %r182, %r181;
or.b32 %r183, %r100, -2139062144;
add.s32 %r184, %r183, -538976288;
xor.b32 %r185, %r184, -2139062144;
xor.b32 %r186, %r100, 538976288;
xor.b32 %r187, %r184, %r100;
and.b32 %r98, %r187, %r186;

prmt.b32 %r98, %r98, 0, 0xba98;


prmt.b32 %r100, %r100, 0, 0xba98;

xor.b32 %r188, %r100, 2139062143;
not.b32 %r189, %r98;
and.b32 %r190, %r185, %r189;
and.b32 %r191, %r188, %r98;
or.b32 %r103, %r191, %r190;

dp4a.s32.s32 %r102, %r103, %r104, %r159;

mul.lo.s32 %r192, %r102, %r179;
cvt.rn.f32.s32 %f39, %r192;
fma.rn.ftz.f32 %f40, %f26, %f39, %f38;
fma.rn.ftz.f32 %f120, %f27, %f40, %f120;
add.s64 %rd44, %rd32, %rd7;
ld.global.nc.u16 %rs11, [%rd44];
ld.global.nc.u16 %rs12, [%rd44+2];
mov.b32 %r193, {%rs11, %rs12};
add.s64 %rd45, %rd34, %rd7;
ld.global.nc.u16 %rs13, [%rd45];
ld.global.nc.u16 %rs14, [%rd45+2];
mov.b32 %r194, {%rs13, %rs14};
shr.s32 %r195, %r194, %r13;

{.reg .f16 low,high;
mov.b32 {low,high},%r88;
cvt.f32.f16 %f28, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r89;
cvt.f32.f16 %f29, low;}


add.s64 %rd46, %rd41, %rd7;
ld.global.nc.u16 %rs2, [%rd46];

{ cvt.f32.f16 %f30, %rs2;}


add.s64 %rd47, %rd43, %rd7;
ld.global.nc.u8 %rs15, [%rd47];
cvt.u32.u16 %r196, %rs15;
cvt.s32.s8 %r197, %r196;
and.b32 %r198, %r193, 252645135;
shl.b32 %r199, %r195, 4;
and.b32 %r200, %r199, 808464432;
or.b32 %r110, %r200, %r198;
or.b32 %r201, %r110, -2139062144;
add.s32 %r202, %r201, -538976288;
xor.b32 %r203, %r202, -2139062144;
xor.b32 %r204, %r110, 538976288;
xor.b32 %r205, %r202, %r110;
and.b32 %r108, %r205, %r204;

prmt.b32 %r108, %r108, 0, 0xba98;


prmt.b32 %r110, %r110, 0, 0xba98;

xor.b32 %r206, %r110, 2139062143;
not.b32 %r207, %r108;
and.b32 %r208, %r203, %r207;
and.b32 %r209, %r206, %r108;
or.b32 %r113, %r209, %r208;

dp4a.s32.s32 %r112, %r113, %r96, %r159;

mul.lo.s32 %r210, %r112, %r197;
cvt.rn.f32.s32 %f41, %r210;
fma.rn.ftz.f32 %f42, %f28, %f41, 0f00000000;
ld.global.nc.u8 %rs16, [%rd47+4];
cvt.u32.u16 %r211, %rs16;
cvt.s32.s8 %r212, %r211;
shr.u32 %r213, %r193, 4;
and.b32 %r214, %r213, 252645135;
and.b32 %r215, %r195, 808464432;
or.b32 %r118, %r215, %r214;
or.b32 %r216, %r118, -2139062144;
add.s32 %r217, %r216, -538976288;
xor.b32 %r218, %r217, -2139062144;
xor.b32 %r219, %r118, 538976288;
xor.b32 %r220, %r217, %r118;
and.b32 %r116, %r220, %r219;

prmt.b32 %r116, %r116, 0, 0xba98;


prmt.b32 %r118, %r118, 0, 0xba98;

xor.b32 %r221, %r118, 2139062143;
not.b32 %r222, %r116;
and.b32 %r223, %r218, %r222;
and.b32 %r224, %r221, %r116;
or.b32 %r121, %r224, %r223;

dp4a.s32.s32 %r120, %r121, %r104, %r159;

mul.lo.s32 %r225, %r120, %r212;
cvt.rn.f32.s32 %f43, %r225;
fma.rn.ftz.f32 %f44, %f29, %f43, %f42;
fma.rn.ftz.f32 %f119, %f30, %f44, %f119;
add.s64 %rd48, %rd44, %rd7;
ld.global.nc.u16 %rs17, [%rd48];
ld.global.nc.u16 %rs18, [%rd48+2];
mov.b32 %r226, {%rs17, %rs18};
add.s64 %rd49, %rd45, %rd7;
ld.global.nc.u16 %rs19, [%rd49];
ld.global.nc.u16 %rs20, [%rd49+2];
mov.b32 %r227, {%rs19, %rs20};
shr.s32 %r228, %r227, %r13;

{.reg .f16 low,high;
mov.b32 {low,high},%r88;
cvt.f32.f16 %f31, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r89;
cvt.f32.f16 %f32, low;}


add.s64 %rd50, %rd46, %rd7;
ld.global.nc.u16 %rs3, [%rd50];

{ cvt.f32.f16 %f33, %rs3;}


add.s64 %rd51, %rd47, %rd7;
ld.global.nc.u8 %rs21, [%rd51];
cvt.u32.u16 %r229, %rs21;
cvt.s32.s8 %r230, %r229;
and.b32 %r231, %r226, 252645135;
shl.b32 %r232, %r228, 4;
and.b32 %r233, %r232, 808464432;
or.b32 %r128, %r233, %r231;
or.b32 %r234, %r128, -2139062144;
add.s32 %r235, %r234, -538976288;
xor.b32 %r236, %r235, -2139062144;
xor.b32 %r237, %r128, 538976288;
xor.b32 %r238, %r235, %r128;
and.b32 %r126, %r238, %r237;

prmt.b32 %r126, %r126, 0, 0xba98;


prmt.b32 %r128, %r128, 0, 0xba98;

xor.b32 %r239, %r128, 2139062143;
not.b32 %r240, %r126;
and.b32 %r241, %r236, %r240;
and.b32 %r242, %r239, %r126;
or.b32 %r131, %r242, %r241;

dp4a.s32.s32 %r130, %r131, %r96, %r159;

mul.lo.s32 %r243, %r130, %r230;
cvt.rn.f32.s32 %f45, %r243;
fma.rn.ftz.f32 %f46, %f31, %f45, 0f00000000;
ld.global.nc.u8 %rs22, [%rd51+4];
cvt.u32.u16 %r244, %rs22;
cvt.s32.s8 %r245, %r244;
shr.u32 %r246, %r226, 4;
and.b32 %r247, %r246, 252645135;
and.b32 %r248, %r228, 808464432;
or.b32 %r136, %r248, %r247;
or.b32 %r249, %r136, -2139062144;
add.s32 %r250, %r249, -538976288;
xor.b32 %r251, %r250, -2139062144;
xor.b32 %r252, %r136, 538976288;
xor.b32 %r253, %r250, %r136;
and.b32 %r134, %r253, %r252;

prmt.b32 %r134, %r134, 0, 0xba98;


prmt.b32 %r136, %r136, 0, 0xba98;

xor.b32 %r254, %r136, 2139062143;
not.b32 %r255, %r134;
and.b32 %r256, %r251, %r255;
and.b32 %r257, %r254, %r134;
or.b32 %r139, %r257, %r256;

dp4a.s32.s32 %r138, %r139, %r104, %r159;

mul.lo.s32 %r258, %r138, %r245;
cvt.rn.f32.s32 %f47, %r258;
fma.rn.ftz.f32 %f48, %f32, %f47, %f46;
fma.rn.ftz.f32 %f118, %f33, %f48, %f118;
add.s64 %rd52, %rd48, %rd7;
ld.global.nc.u16 %rs23, [%rd52];
ld.global.nc.u16 %rs24, [%rd52+2];
mov.b32 %r259, {%rs23, %rs24};
add.s64 %rd53, %rd49, %rd7;
ld.global.nc.u16 %rs25, [%rd53];
ld.global.nc.u16 %rs26, [%rd53+2];
mov.b32 %r260, {%rs25, %rs26};
shr.s32 %r261, %r260, %r13;

{.reg .f16 low,high;
mov.b32 {low,high},%r88;
cvt.f32.f16 %f34, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r89;
cvt.f32.f16 %f35, low;}


add.s64 %rd54, %rd50, %rd7;
ld.global.nc.u16 %rs4, [%rd54];

{ cvt.f32.f16 %f36, %rs4;}


add.s64 %rd55, %rd51, %rd7;
ld.global.nc.u8 %rs27, [%rd55];
cvt.u32.u16 %r262, %rs27;
cvt.s32.s8 %r263, %r262;
and.b32 %r264, %r259, 252645135;
shl.b32 %r265, %r261, 4;
and.b32 %r266, %r265, 808464432;
or.b32 %r146, %r266, %r264;
or.b32 %r267, %r146, -2139062144;
add.s32 %r268, %r267, -538976288;
xor.b32 %r269, %r268, -2139062144;
xor.b32 %r270, %r146, 538976288;
xor.b32 %r271, %r268, %r146;
and.b32 %r144, %r271, %r270;

prmt.b32 %r144, %r144, 0, 0xba98;


prmt.b32 %r146, %r146, 0, 0xba98;

xor.b32 %r272, %r146, 2139062143;
not.b32 %r273, %r144;
and.b32 %r274, %r269, %r273;
and.b32 %r275, %r272, %r144;
or.b32 %r149, %r275, %r274;

dp4a.s32.s32 %r148, %r149, %r96, %r159;

mul.lo.s32 %r276, %r148, %r263;
cvt.rn.f32.s32 %f49, %r276;
fma.rn.ftz.f32 %f50, %f34, %f49, 0f00000000;
ld.global.nc.u8 %rs28, [%rd55+4];
cvt.u32.u16 %r277, %rs28;
cvt.s32.s8 %r278, %r277;
shr.u32 %r279, %r259, 4;
and.b32 %r280, %r279, 252645135;
and.b32 %r281, %r261, 808464432;
or.b32 %r154, %r281, %r280;
or.b32 %r282, %r154, -2139062144;
add.s32 %r283, %r282, -538976288;
xor.b32 %r284, %r283, -2139062144;
xor.b32 %r285, %r154, 538976288;
xor.b32 %r286, %r283, %r154;
and.b32 %r152, %r286, %r285;

prmt.b32 %r152, %r152, 0, 0xba98;


prmt.b32 %r154, %r154, 0, 0xba98;

xor.b32 %r287, %r154, 2139062143;
not.b32 %r288, %r152;
and.b32 %r289, %r284, %r288;
and.b32 %r290, %r287, %r152;
or.b32 %r157, %r290, %r289;

dp4a.s32.s32 %r156, %r157, %r104, %r159;

mul.lo.s32 %r291, %r156, %r278;
cvt.rn.f32.s32 %f51, %r291;
fma.rn.ftz.f32 %f52, %f35, %f51, %f50;
fma.rn.ftz.f32 %f117, %f36, %f52, %f117;
add.s32 %r376, %r376, 32;
add.s64 %rd58, %rd58, 840;
add.s32 %r377, %r377, 4;
setp.lt.s32 %p3, %r377, %r5;
@%p3 bra $L__BB186_5;

$L__BB186_6:
setp.eq.s32 %p4, %r1, 0;
@%p4 bra $L__BB186_8;

shl.b32 %r292, %r1, 9;
mov.u32 %r293, _ZZ13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r294, %r293, %r292;
shl.b32 %r295, %r2, 2;
add.s32 %r296, %r294, %r295;
st.shared.f32 [%r296+-512], %f120;
st.shared.f32 [%r296+-384], %f119;
st.shared.f32 [%r296+-256], %f118;
st.shared.f32 [%r296+-128], %f117;

$L__BB186_8:
setp.ne.s32 %p5, %r1, 0;
bar.sync 0;
@%p5 bra $L__BB186_17;

mov.u32 %r297, %ctaid.z;
mov.u32 %r298, %ctaid.y;
mad.lo.s32 %r299, %r298, %r31, %r4;
mad.lo.s32 %r300, %r297, %r37, %r299;
shl.b32 %r301, %r2, 2;
mov.u32 %r302, _ZZ13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r19, %r302, %r301;
mov.u32 %r303, 2;
ld.shared.f32 %f53, [%r19];
add.ftz.f32 %f54, %f53, %f120;
ld.shared.f32 %f55, [%r19+512];
add.ftz.f32 %f56, %f55, %f54;
ld.shared.f32 %f57, [%r19+1024];
add.ftz.f32 %f58, %f57, %f56;
mov.b32 %r304, %f58;
mov.u32 %r305, 31;
mov.u32 %r306, 16;
mov.u32 %r307, -1;
shfl.sync.bfly.b32 %r308|%p6, %r304, %r306, %r305, %r307;
mov.b32 %f59, %r308;
add.ftz.f32 %f60, %f58, %f59;
mov.b32 %r309, %f60;
mov.u32 %r310, 8;
shfl.sync.bfly.b32 %r311|%p7, %r309, %r310, %r305, %r307;
mov.b32 %f61, %r311;
add.ftz.f32 %f62, %f60, %f61;
mov.b32 %r312, %f62;
mov.u32 %r313, 4;
shfl.sync.bfly.b32 %r314|%p8, %r312, %r313, %r305, %r307;
mov.b32 %f63, %r314;
add.ftz.f32 %f64, %f62, %f63;
mov.b32 %r315, %f64;
shfl.sync.bfly.b32 %r316|%p9, %r315, %r303, %r305, %r307;
mov.b32 %f65, %r316;
add.ftz.f32 %f66, %f64, %f65;
mov.b32 %r317, %f66;
mov.u32 %r318, 1;
shfl.sync.bfly.b32 %r319|%p10, %r317, %r318, %r305, %r307;
mov.b32 %f67, %r319;
add.ftz.f32 %f13, %f66, %f67;
setp.ge.u32 %p11, %r4, %r25;
setp.ne.s32 %p12, %r2, 0;
cvta.to.global.u64 %rd56, %rd17;
mul.wide.u32 %rd57, %r300, 4;
add.s64 %rd13, %rd56, %rd57;
or.pred %p13, %p12, %p11;
@%p13 bra $L__BB186_11;

st.global.f32 [%rd13], %f13;

$L__BB186_11:
ld.shared.f32 %f68, [%r19+128];
add.ftz.f32 %f69, %f68, %f119;
ld.shared.f32 %f70, [%r19+640];
add.ftz.f32 %f71, %f70, %f69;
ld.shared.f32 %f72, [%r19+1152];
add.ftz.f32 %f73, %f72, %f71;
mov.b32 %r320, %f73;
shfl.sync.bfly.b32 %r324|%p14, %r320, %r306, %r305, %r307;
mov.b32 %f74, %r324;
add.ftz.f32 %f75, %f73, %f74;
mov.b32 %r325, %f75;
shfl.sync.bfly.b32 %r327|%p15, %r325, %r310, %r305, %r307;
mov.b32 %f76, %r327;
add.ftz.f32 %f77, %f75, %f76;
mov.b32 %r328, %f77;
shfl.sync.bfly.b32 %r330|%p16, %r328, %r313, %r305, %r307;
mov.b32 %f78, %r330;
add.ftz.f32 %f79, %f77, %f78;
mov.b32 %r331, %f79;
shfl.sync.bfly.b32 %r333|%p17, %r331, %r303, %r305, %r307;
mov.b32 %f80, %r333;
add.ftz.f32 %f81, %f79, %f80;
mov.b32 %r334, %f81;
shfl.sync.bfly.b32 %r336|%p18, %r334, %r318, %r305, %r307;
mov.b32 %f82, %r336;
add.ftz.f32 %f14, %f81, %f82;
add.s32 %r337, %r4, 1;
setp.ge.u32 %p19, %r337, %r25;
setp.ne.s32 %p20, %r2, 1;
or.pred %p21, %p20, %p19;
@%p21 bra $L__BB186_13;

st.global.f32 [%rd13+4], %f14;

$L__BB186_13:
ld.shared.f32 %f83, [%r19+256];
add.ftz.f32 %f84, %f83, %f118;
ld.shared.f32 %f85, [%r19+768];
add.ftz.f32 %f86, %f85, %f84;
ld.shared.f32 %f87, [%r19+1280];
add.ftz.f32 %f88, %f87, %f86;
mov.b32 %r338, %f88;
mov.u32 %r339, 31;
mov.u32 %r340, 16;
mov.u32 %r341, -1;
shfl.sync.bfly.b32 %r342|%p22, %r338, %r340, %r339, %r341;
mov.b32 %f89, %r342;
add.ftz.f32 %f90, %f88, %f89;
mov.b32 %r343, %f90;
mov.u32 %r344, 8;
shfl.sync.bfly.b32 %r345|%p23, %r343, %r344, %r339, %r341;
mov.b32 %f91, %r345;
add.ftz.f32 %f92, %f90, %f91;
mov.b32 %r346, %f92;
mov.u32 %r347, 4;
shfl.sync.bfly.b32 %r348|%p24, %r346, %r347, %r339, %r341;
mov.b32 %f93, %r348;
add.ftz.f32 %f94, %f92, %f93;
mov.b32 %r349, %f94;
mov.u32 %r350, 2;
shfl.sync.bfly.b32 %r351|%p25, %r349, %r350, %r339, %r341;
mov.b32 %f95, %r351;
add.ftz.f32 %f96, %f94, %f95;
mov.b32 %r352, %f96;
mov.u32 %r353, 1;
shfl.sync.bfly.b32 %r354|%p26, %r352, %r353, %r339, %r341;
mov.b32 %f97, %r354;
add.ftz.f32 %f15, %f96, %f97;
add.s32 %r355, %r4, 2;
setp.ge.u32 %p27, %r355, %r25;
setp.ne.s32 %p28, %r2, 2;
or.pred %p29, %p28, %p27;
@%p29 bra $L__BB186_15;

st.global.f32 [%rd13+8], %f15;

$L__BB186_15:
ld.shared.f32 %f98, [%r19+384];
add.ftz.f32 %f99, %f98, %f117;
ld.shared.f32 %f100, [%r19+896];
add.ftz.f32 %f101, %f100, %f99;
ld.shared.f32 %f102, [%r19+1408];
add.ftz.f32 %f103, %f102, %f101;
mov.b32 %r356, %f103;
shfl.sync.bfly.b32 %r360|%p30, %r356, %r340, %r339, %r341;
mov.b32 %f104, %r360;
add.ftz.f32 %f105, %f103, %f104;
mov.b32 %r361, %f105;
shfl.sync.bfly.b32 %r363|%p31, %r361, %r344, %r339, %r341;
mov.b32 %f106, %r363;
add.ftz.f32 %f107, %f105, %f106;
mov.b32 %r364, %f107;
shfl.sync.bfly.b32 %r366|%p32, %r364, %r347, %r339, %r341;
mov.b32 %f108, %r366;
add.ftz.f32 %f109, %f107, %f108;
mov.b32 %r367, %f109;
shfl.sync.bfly.b32 %r369|%p33, %r367, %r350, %r339, %r341;
mov.b32 %f110, %r369;
add.ftz.f32 %f111, %f109, %f110;
mov.b32 %r370, %f111;
shfl.sync.bfly.b32 %r372|%p34, %r370, %r353, %r339, %r341;
mov.b32 %f112, %r372;
add.ftz.f32 %f16, %f111, %f112;
add.s32 %r373, %r4, 3;
setp.ge.u32 %p35, %r373, %r25;
setp.ne.s32 %p36, %r2, 3;
or.pred %p37, %p36, %p35;
@%p37 bra $L__BB186_17;

st.global.f32 [%rd13+12], %f16;

$L__BB186_17:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<33>;
.reg .b16 %rs<13>;
.reg .f32 %f<126>;
.reg .b32 %r<267>;
.reg .b64 %rd<78>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[384];

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate[384];

ld.param.u64 %rd11, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd12, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd13, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r45, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r46, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
ld.param.f32 %f31, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+16];
ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+8];
ld.param.u64 %rd14, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3];
mov.u32 %r1, %tid.y;
shl.b32 %r47, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r47, %r2;
mov.u32 %r4, %ctaid.x;
mov.u32 %r5, %ctaid.y;
setp.eq.s64 %p2, %rd13, 0;
@%p2 bra $L__BB187_2;

cvta.to.global.u64 %rd20, %rd13;
mul.wide.u32 %rd21, %r5, 4;
add.s64 %rd22, %rd20, %rd21;
ld.global.nc.u32 %r263, [%rd22];
mul.hi.u32 %r48, %r5, %r31;
add.s32 %r49, %r48, %r5;
shr.u32 %r50, %r49, %r32;
mul.lo.s32 %r51, %r50, %r33;
sub.s32 %r264, %r5, %r51;
bra.uni $L__BB187_3;

$L__BB187_2:
mul.hi.u32 %r52, %r5, %r35;
add.s32 %r53, %r52, %r5;
shr.u32 %r263, %r53, %r36;
mov.u32 %r264, %r5;

$L__BB187_3:
mov.u32 %r11, %ctaid.z;
selp.b32 %r12, %r5, %r263, %p2;
or.b32 %r54, %r2, %r1;
setp.ne.s32 %p4, %r54, 0;
mov.f32 %f119, 0f00000000;
mov.f32 %f115, %f119;
mov.f32 %f116, %f119;
@%p4 bra $L__BB187_8;

setp.eq.s64 %p5, %rd14, 0;
mov.f32 %f115, 0f00000000;
mov.f32 %f116, %f115;
@%p5 bra $L__BB187_6;

cvta.to.global.u64 %rd23, %rd14;
mul.lo.s32 %r55, %r11, %r46;
cvt.u64.u32 %rd24, %r55;
mul.lo.s32 %r56, %r12, %r40;
cvt.u64.u32 %rd25, %r56;
cvt.u64.u32 %rd26, %r2;
cvt.s64.s32 %rd27, %r4;
add.s64 %rd28, %rd26, %rd27;
add.s64 %rd29, %rd28, %rd24;
add.s64 %rd30, %rd29, %rd25;
shl.b64 %rd31, %rd30, 2;
add.s64 %rd32, %rd23, %rd31;
ld.global.f32 %f116, [%rd32];

$L__BB187_6:
setp.eq.s64 %p6, %rd16, 0;
setp.eq.s64 %p7, %rd15, 0;
or.pred %p8, %p6, %p7;
@%p8 bra $L__BB187_8;

cvta.to.global.u64 %rd33, %rd16;
mul.lo.s32 %r58, %r11, %r46;
cvt.u64.u32 %rd34, %r58;
mul.lo.s32 %r59, %r12, %r40;
cvt.u64.u32 %rd35, %r59;
cvt.u64.u32 %rd36, %r2;
cvt.s64.s32 %rd37, %r4;
add.s64 %rd38, %rd36, %rd37;
add.s64 %rd39, %rd38, %rd34;
add.s64 %rd40, %rd39, %rd35;
shl.b64 %rd41, %rd40, 2;
add.s64 %rd42, %rd33, %rd41;
ld.global.f32 %f115, [%rd42];

$L__BB187_8:
shr.u32 %r61, %r30, 8;
shr.s32 %r62, %r3, 31;
shr.u32 %r63, %r62, 27;
add.s32 %r64, %r3, %r63;
shr.s32 %r266, %r64, 5;
setp.ge.s32 %p9, %r266, %r61;
mov.f32 %f121, %f119;
@%p9 bra $L__BB187_13;

and.b32 %r68, %r64, -32;
sub.s32 %r69, %r3, %r68;
shr.s32 %r70, %r69, 31;
shr.u32 %r71, %r70, 28;
add.s32 %r72, %r69, %r71;
shr.s32 %r73, %r72, 4;
shl.b32 %r74, %r73, 2;
and.b32 %r75, %r72, -16;
sub.s32 %r76, %r69, %r75;
shr.s32 %r77, %r76, 31;
shr.u32 %r78, %r77, 29;
add.s32 %r79, %r76, %r78;
shr.s32 %r80, %r79, 3;
add.s32 %r81, %r74, %r80;
shl.b32 %r82, %r73, 3;
shr.u32 %r83, %r77, 30;
add.s32 %r84, %r76, %r83;
shr.s32 %r85, %r84, 2;
add.s32 %r86, %r82, %r85;
shl.b32 %r14, %r80, 1;
shl.b32 %r15, %r69, 1;
shr.u32 %r87, %r70, 29;
add.s32 %r88, %r69, %r87;
and.b32 %r89, %r88, -8;
sub.s32 %r90, %r69, %r89;
add.s32 %r91, %r82, %r90;
shl.b32 %r16, %r91, 1;
cvt.s64.s32 %rd43, %r86;
cvt.s64.s32 %rd4, %r90;
cvt.s64.s32 %rd5, %r81;
shl.b32 %r265, %r266, 3;
mul.hi.u32 %r92, %r11, %r41;
add.s32 %r93, %r92, %r11;
shr.u32 %r94, %r93, %r42;
mad.lo.s32 %r95, %r94, %r44, %r266;
mad.lo.s32 %r96, %r263, %r38, %r95;
mad.lo.s32 %r97, %r4, %r34, %r96;
mul.wide.s32 %rd77, %r97, 210;
add.s64 %rd7, %rd43, 192;
mul.lo.s32 %r98, %r264, %r39;
cvt.u64.u32 %rd44, %r98;
mul.lo.s32 %r99, %r11, %r45;
cvt.u64.u32 %rd45, %r99;
add.s64 %rd8, %rd44, %rd45;
mov.f32 %f121, 0f00000000;
cvta.to.global.u64 %rd46, %rd12;
cvta.to.global.u64 %rd50, %rd11;
shl.b64 %rd59, %rd4, 2;
mov.f32 %f119, %f121;

$L__BB187_10:
cvt.u64.u32 %rd47, %r265;
add.s64 %rd48, %rd8, %rd47;
mul.wide.s32 %rd49, %r15, 2;
add.s64 %rd51, %rd50, %rd49;
add.s64 %rd52, %rd51, %rd77;
ld.global.nc.u16 %rs2, [%rd52];
ld.global.nc.u16 %rs3, [%rd52+2];
mov.b32 %r118, {%rs2, %rs3};
mul.wide.s32 %rd53, %r16, 2;
add.s64 %rd54, %rd50, %rd53;
add.s64 %rd55, %rd54, %rd77;
ld.global.nc.u16 %rs4, [%rd55+128];
ld.global.nc.u16 %rs5, [%rd55+130];
mov.b32 %r119, {%rs4, %rs5};
shr.s32 %r120, %r119, %r14;
add.s64 %rd56, %rd48, %rd5;
mul.lo.s64 %rd57, %rd56, 36;
add.s64 %rd58, %rd46, %rd57;
add.s64 %rd60, %rd58, %rd59;
ld.global.nc.u32 %r20, [%rd60+4];
ld.global.nc.u32 %r21, [%rd58];

{.reg .f16 low,high;
mov.b32 {low,high},%r21;
cvt.f32.f16 %f40, low;}


ld.global.nc.u32 %r22, [%rd60+76];
ld.global.nc.u32 %r23, [%rd58+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r23;
cvt.f32.f16 %f41, low;}


add.s64 %rd61, %rd50, %rd77;
ld.global.nc.u16 %rs1, [%rd61+208];

{ cvt.f32.f16 %f42, %rs1;}


add.s64 %rd62, %rd50, %rd7;
add.s64 %rd63, %rd62, %rd77;
ld.global.nc.u8 %rs6, [%rd63];
cvt.u32.u16 %r121, %rs6;
cvt.s32.s8 %r122, %r121;
and.b32 %r123, %r118, 252645135;
shl.b32 %r124, %r120, 4;
and.b32 %r125, %r124, 808464432;
or.b32 %r104, %r125, %r123;
or.b32 %r126, %r104, -2139062144;
add.s32 %r127, %r126, -538976288;
xor.b32 %r128, %r127, -2139062144;
xor.b32 %r129, %r104, 538976288;
xor.b32 %r130, %r127, %r104;
and.b32 %r102, %r130, %r129;

prmt.b32 %r102, %r102, 0, 0xba98;


prmt.b32 %r104, %r104, 0, 0xba98;

xor.b32 %r131, %r104, 2139062143;
not.b32 %r132, %r102;
and.b32 %r133, %r128, %r132;
and.b32 %r134, %r131, %r102;
or.b32 %r107, %r134, %r133;
mov.u32 %r117, 0;

dp4a.s32.s32 %r106, %r107, %r20, %r117;

mul.lo.s32 %r135, %r106, %r122;
cvt.rn.f32.s32 %f43, %r135;
fma.rn.ftz.f32 %f44, %f40, %f43, 0f00000000;
ld.global.nc.u8 %rs7, [%rd63+4];
cvt.u32.u16 %r136, %rs7;
cvt.s32.s8 %r137, %r136;
shr.u32 %r138, %r118, 4;
and.b32 %r139, %r138, 252645135;
and.b32 %r140, %r120, 808464432;
or.b32 %r112, %r140, %r139;
or.b32 %r141, %r112, -2139062144;
add.s32 %r142, %r141, -538976288;
xor.b32 %r143, %r142, -2139062144;
xor.b32 %r144, %r112, 538976288;
xor.b32 %r145, %r142, %r112;
and.b32 %r110, %r145, %r144;

prmt.b32 %r110, %r110, 0, 0xba98;


prmt.b32 %r112, %r112, 0, 0xba98;

xor.b32 %r146, %r112, 2139062143;
not.b32 %r147, %r110;
and.b32 %r148, %r143, %r147;
and.b32 %r149, %r146, %r110;
or.b32 %r115, %r149, %r148;

dp4a.s32.s32 %r114, %r115, %r22, %r117;

mul.lo.s32 %r150, %r114, %r137;
cvt.rn.f32.s32 %f45, %r150;
fma.rn.ftz.f32 %f46, %f41, %f45, %f44;
fma.rn.ftz.f32 %f121, %f42, %f46, %f121;
setp.eq.s64 %p10, %rd15, 0;
@%p10 bra $L__BB187_12;

cvta.to.global.u64 %rd65, %rd15;
add.s64 %rd66, %rd65, %rd49;
add.s64 %rd67, %rd66, %rd77;
ld.global.u16 %rs9, [%rd67];
ld.global.u16 %rs10, [%rd67+2];
mov.b32 %r169, {%rs9, %rs10};
add.s64 %rd69, %rd65, %rd53;
add.s64 %rd70, %rd69, %rd77;
ld.global.u16 %rs11, [%rd70+128];
ld.global.u16 %rs12, [%rd70+130];
mov.b32 %r170, {%rs11, %rs12};
shr.s32 %r171, %r170, %r14;

{.reg .f16 low,high;
mov.b32 {low,high},%r21;
cvt.f32.f16 %f47, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r23;
cvt.f32.f16 %f48, low;}


add.s64 %rd71, %rd65, %rd77;
ld.global.u16 %rs8, [%rd71+208];

{ cvt.f32.f16 %f49, %rs8;}


add.s64 %rd72, %rd65, %rd7;
add.s64 %rd73, %rd72, %rd77;
ld.global.s8 %r172, [%rd73];
and.b32 %r173, %r169, 252645135;
shl.b32 %r174, %r171, 4;
and.b32 %r175, %r174, 808464432;
or.b32 %r155, %r175, %r173;
or.b32 %r176, %r155, -2139062144;
add.s32 %r177, %r176, -538976288;
xor.b32 %r178, %r177, -2139062144;
xor.b32 %r179, %r155, 538976288;
xor.b32 %r180, %r177, %r155;
and.b32 %r153, %r180, %r179;

prmt.b32 %r153, %r153, 0, 0xba98;


prmt.b32 %r155, %r155, 0, 0xba98;

xor.b32 %r181, %r155, 2139062143;
not.b32 %r182, %r153;
and.b32 %r183, %r178, %r182;
and.b32 %r184, %r181, %r153;
or.b32 %r158, %r184, %r183;

dp4a.s32.s32 %r157, %r158, %r20, %r117;

mul.lo.s32 %r185, %r157, %r172;
cvt.rn.f32.s32 %f50, %r185;
fma.rn.ftz.f32 %f51, %f47, %f50, 0f00000000;
ld.global.s8 %r186, [%rd73+4];
shr.u32 %r187, %r169, 4;
and.b32 %r188, %r187, 252645135;
and.b32 %r189, %r171, 808464432;
or.b32 %r163, %r189, %r188;
or.b32 %r190, %r163, -2139062144;
add.s32 %r191, %r190, -538976288;
xor.b32 %r192, %r191, -2139062144;
xor.b32 %r193, %r163, 538976288;
xor.b32 %r194, %r191, %r163;
and.b32 %r161, %r194, %r193;

prmt.b32 %r161, %r161, 0, 0xba98;


prmt.b32 %r163, %r163, 0, 0xba98;

xor.b32 %r195, %r163, 2139062143;
not.b32 %r196, %r161;
and.b32 %r197, %r192, %r196;
and.b32 %r198, %r195, %r161;
or.b32 %r166, %r198, %r197;

dp4a.s32.s32 %r165, %r166, %r22, %r117;

mul.lo.s32 %r199, %r165, %r186;
cvt.rn.f32.s32 %f52, %r199;
fma.rn.ftz.f32 %f53, %f48, %f52, %f51;
fma.rn.ftz.f32 %f119, %f49, %f53, %f119;

$L__BB187_12:
add.s32 %r265, %r265, 32;
add.s64 %rd77, %rd77, 840;
add.s32 %r266, %r266, 4;
setp.lt.s32 %p11, %r266, %r61;
@%p11 bra $L__BB187_10;

$L__BB187_13:
setp.eq.s32 %p12, %r1, 0;
@%p12 bra $L__BB187_16;

setp.eq.s64 %p13, %rd15, 0;
add.s32 %r26, %r1, -1;
shl.b32 %r204, %r1, 7;
mov.u32 %r205, _ZZ13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r206, %r205, %r204;
shl.b32 %r207, %r2, 2;
add.s32 %r208, %r206, %r207;
st.shared.f32 [%r208+-128], %f121;
@%p13 bra $L__BB187_16;

shl.b32 %r210, %r26, 7;
mov.u32 %r211, _ZZ13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r212, %r211, %r210;
add.s32 %r214, %r212, %r207;
st.shared.f32 [%r214], %f119;

$L__BB187_16:
setp.ne.s32 %p14, %r1, 0;
bar.sync 0;
@%p14 bra $L__BB187_35;

setp.ne.s64 %p15, %rd15, 0;
shl.b32 %r217, %r2, 2;
mov.u32 %r218, _ZZ13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r27, %r218, %r217;
ld.shared.f32 %f54, [%r27];
add.ftz.f32 %f13, %f121, %f54;
@%p15 bra $L__BB187_19;
bra.uni $L__BB187_18;

$L__BB187_19:
mov.u32 %r221, _ZZ13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r222, %r221, %r217;
ld.shared.f32 %f58, [%r222];
add.ftz.f32 %f59, %f119, %f58;
ld.shared.f32 %f60, [%r27+128];
add.ftz.f32 %f61, %f13, %f60;
ld.shared.f32 %f62, [%r222+128];
add.ftz.f32 %f63, %f59, %f62;
ld.shared.f32 %f64, [%r27+256];
add.ftz.f32 %f123, %f61, %f64;
ld.shared.f32 %f65, [%r222+256];
add.ftz.f32 %f119, %f63, %f65;
bra.uni $L__BB187_20;

$L__BB187_18:
ld.shared.f32 %f55, [%r27+128];
add.ftz.f32 %f56, %f13, %f55;
ld.shared.f32 %f57, [%r27+256];
add.ftz.f32 %f123, %f56, %f57;

$L__BB187_20:
mov.b32 %r223, %f123;
mov.u32 %r224, 31;
mov.u32 %r225, 16;
mov.u32 %r226, -1;
shfl.sync.bfly.b32 %r227|%p16, %r223, %r225, %r224, %r226;
mov.b32 %f66, %r227;
add.ftz.f32 %f67, %f123, %f66;
mov.b32 %r228, %f67;
mov.u32 %r229, 8;
shfl.sync.bfly.b32 %r230|%p17, %r228, %r229, %r224, %r226;
mov.b32 %f68, %r230;
add.ftz.f32 %f69, %f67, %f68;
mov.b32 %r231, %f69;
mov.u32 %r232, 4;
shfl.sync.bfly.b32 %r233|%p18, %r231, %r232, %r224, %r226;
mov.b32 %f70, %r233;
add.ftz.f32 %f71, %f69, %f70;
mov.b32 %r234, %f71;
mov.u32 %r235, 2;
shfl.sync.bfly.b32 %r236|%p19, %r234, %r235, %r224, %r226;
mov.b32 %f72, %r236;
add.ftz.f32 %f19, %f71, %f72;
mov.b32 %r237, %f19;
mov.u32 %r238, 1;
shfl.sync.bfly.b32 %r28|%p1, %r237, %r238, %r224, %r226;
setp.eq.s64 %p20, %rd15, 0;
@%p20 bra $L__BB187_22;

mov.b32 %r239, %f119;
shfl.sync.bfly.b32 %r243|%p21, %r239, %r225, %r224, %r226;
mov.b32 %f73, %r243;
add.ftz.f32 %f74, %f119, %f73;
mov.b32 %r244, %f74;
shfl.sync.bfly.b32 %r246|%p22, %r244, %r229, %r224, %r226;
mov.b32 %f75, %r246;
add.ftz.f32 %f76, %f74, %f75;
mov.b32 %r247, %f76;
shfl.sync.bfly.b32 %r249|%p23, %r247, %r232, %r224, %r226;
mov.b32 %f77, %r249;
add.ftz.f32 %f78, %f76, %f77;
mov.b32 %r250, %f78;
shfl.sync.bfly.b32 %r252|%p24, %r250, %r235, %r224, %r226;
mov.b32 %f79, %r252;
add.ftz.f32 %f80, %f78, %f79;
mov.b32 %r253, %f80;
shfl.sync.bfly.b32 %r255|%p25, %r253, %r238, %r224, %r226;
mov.b32 %f81, %r255;
add.ftz.f32 %f119, %f80, %f81;

$L__BB187_22:
setp.ne.s32 %p26, %r2, 0;
mov.b32 %f82, %r28;
add.ftz.f32 %f22, %f19, %f82;
@%p26 bra $L__BB187_35;

add.ftz.f32 %f125, %f116, %f22;
@%p20 bra $L__BB187_34;

add.ftz.f32 %f24, %f115, %f119;
setp.gt.s32 %p28, %r29, 2;
@%p28 bra $L__BB187_28;

setp.eq.s32 %p31, %r29, 1;
@%p31 bra $L__BB187_32;

setp.eq.s32 %p32, %r29, 2;
@%p32 bra $L__BB187_27;
bra.uni $L__BB187_33;

$L__BB187_27:
mul.ftz.f32 %f102, %f24, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f103, %f102;
add.ftz.f32 %f104, %f103, 0f3F800000;
div.approx.ftz.f32 %f105, %f24, %f104;
mul.ftz.f32 %f125, %f125, %f105;
bra.uni $L__BB187_34;

$L__BB187_28:
setp.eq.s32 %p29, %r29, 3;
@%p29 bra $L__BB187_31;

setp.ne.s32 %p30, %r29, 6;
@%p30 bra $L__BB187_33;

min.ftz.f32 %f83, %f24, %f31;
neg.ftz.f32 %f84, %f31;
min.ftz.f32 %f85, %f125, %f31;
max.ftz.f32 %f86, %f85, %f84;
mul.ftz.f32 %f87, %f83, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f88, %f87;
add.ftz.f32 %f89, %f88, 0f3F800000;
div.approx.ftz.f32 %f90, %f83, %f89;
mul.ftz.f32 %f125, %f86, %f90;
bra.uni $L__BB187_34;

$L__BB187_32:
mul.ftz.f32 %f106, %f24, 0f3F000000;
mul.ftz.f32 %f107, %f24, 0f3D372713;
fma.rn.ftz.f32 %f108, %f24, %f107, 0f3F800000;
mul.ftz.f32 %f109, %f24, 0f3F4C422A;
mul.ftz.f32 %f110, %f109, %f108;
tanh.approx.f32 %f111, %f110;
add.ftz.f32 %f112, %f111, 0f3F800000;
mul.ftz.f32 %f113, %f106, %f112;
mul.ftz.f32 %f125, %f125, %f113;
bra.uni $L__BB187_34;

$L__BB187_31:
mov.f32 %f91, 0f40E00000;
min.ftz.f32 %f92, %f24, %f91;
min.ftz.f32 %f93, %f125, %f91;
mov.f32 %f94, 0fC0E00000;
max.ftz.f32 %f95, %f93, %f94;
mul.ftz.f32 %f96, %f92, 0f3FD9DB23;
mul.ftz.f32 %f97, %f96, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f98, %f97;
add.ftz.f32 %f99, %f98, 0f3F800000;
div.approx.ftz.f32 %f100, %f92, %f99;
add.ftz.f32 %f101, %f95, 0f3F800000;
mul.ftz.f32 %f125, %f101, %f100;
bra.uni $L__BB187_34;

$L__BB187_33:
mul.ftz.f32 %f125, %f125, %f24;

$L__BB187_34:
cvta.to.global.u64 %rd74, %rd19;
mov.u32 %r259, %ctaid.y;
mad.lo.s32 %r261, %r259, %r40, %r4;
mad.lo.s32 %r262, %r11, %r46, %r261;
mul.wide.u32 %rd75, %r262, 4;
add.s64 %rd76, %rd74, %rd75;
st.global.f32 [%rd76], %f125;

$L__BB187_35:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<12>;
.reg .b16 %rs<8>;
.reg .f32 %f<31>;
.reg .b32 %r<169>;
.reg .b64 %rd<42>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[384];

ld.param.u64 %rd13, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd14, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
mov.u32 %r1, %tid.y;
shl.b32 %r37, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r37, %r2;
mov.u32 %r166, %ctaid.y;
setp.eq.s64 %p1, %rd15, 0;
@%p1 bra $L__BB188_2;

cvta.to.global.u64 %rd17, %rd15;
mul.wide.u32 %rd18, %r166, 4;
add.s64 %rd19, %rd17, %rd18;
ld.global.nc.u32 %r165, [%rd19];
mul.hi.u32 %r38, %r166, %r21;
add.s32 %r39, %r38, %r166;
shr.u32 %r40, %r39, %r22;
mul.lo.s32 %r41, %r40, %r23;
sub.s32 %r166, %r166, %r41;
bra.uni $L__BB188_3;

$L__BB188_2:
mul.hi.u32 %r42, %r166, %r25;
add.s32 %r43, %r42, %r166;
shr.u32 %r165, %r43, %r26;

$L__BB188_3:
mov.u32 %r10, %ctaid.z;
shr.s32 %r44, %r3, 31;
shr.u32 %r45, %r44, 27;
add.s32 %r46, %r3, %r45;
shr.s32 %r168, %r46, 5;
shr.u32 %r47, %r20, 8;
setp.ge.s32 %p2, %r168, %r47;
mov.f32 %f30, 0f00000000;
@%p2 bra $L__BB188_6;

cvta.to.global.u64 %rd20, %rd13;
and.b32 %r51, %r46, -32;
sub.s32 %r52, %r3, %r51;
shr.s32 %r53, %r52, 31;
shr.u32 %r54, %r53, 28;
add.s32 %r55, %r52, %r54;
shr.s32 %r56, %r55, 4;
shl.b32 %r57, %r56, 2;
and.b32 %r58, %r55, -16;
sub.s32 %r59, %r52, %r58;
shr.s32 %r60, %r59, 31;
shr.u32 %r61, %r60, 29;
add.s32 %r62, %r59, %r61;
shr.s32 %r63, %r62, 3;
add.s32 %r64, %r57, %r63;
shl.b32 %r65, %r56, 3;
shr.u32 %r66, %r60, 30;
add.s32 %r67, %r59, %r66;
shr.s32 %r68, %r67, 2;
add.s32 %r69, %r65, %r68;
shl.b32 %r12, %r63, 1;
shl.b32 %r70, %r52, 1;
shr.u32 %r71, %r53, 29;
add.s32 %r72, %r52, %r71;
and.b32 %r73, %r72, -8;
sub.s32 %r74, %r52, %r73;
add.s32 %r75, %r65, %r74;
shl.b32 %r76, %r75, 1;
cvt.s64.s32 %rd21, %r69;
cvt.s64.s32 %rd1, %r74;
cvt.s64.s32 %rd2, %r64;
shl.b32 %r167, %r168, 3;
mul.hi.u32 %r77, %r10, %r31;
add.s32 %r78, %r77, %r10;
shr.u32 %r79, %r78, %r32;
mad.lo.s32 %r80, %r79, %r34, %r168;
mad.lo.s32 %r81, %r165, %r28, %r80;
mov.u32 %r82, %ctaid.x;
mad.lo.s32 %r83, %r82, %r24, %r81;
mul.wide.s32 %rd22, %r83, 210;
add.s64 %rd23, %rd22, %rd21;
add.s64 %rd24, %rd20, %rd23;
add.s64 %rd41, %rd24, 196;
add.s64 %rd40, %rd20, %rd22;
mul.wide.s32 %rd25, %r76, 2;
add.s64 %rd5, %rd25, 128;
mul.wide.s32 %rd6, %r70, 2;
cvta.to.global.u64 %rd7, %rd14;
mul.lo.s32 %r84, %r166, %r29;
cvt.u64.u32 %rd26, %r84;
mul.lo.s32 %r85, %r10, %r35;
cvt.u64.u32 %rd27, %r85;
add.s64 %rd8, %rd26, %rd27;
mov.f32 %f30, 0f00000000;
shl.b64 %rd35, %rd1, 2;

$L__BB188_5:
cvt.u64.u32 %rd28, %r167;
add.s64 %rd29, %rd8, %rd28;
add.s64 %rd30, %rd40, %rd6;
ld.global.nc.u16 %rs2, [%rd30];
ld.global.nc.u16 %rs3, [%rd30+2];
mov.b32 %r104, {%rs2, %rs3};
add.s64 %rd31, %rd40, %rd5;
ld.global.nc.u16 %rs4, [%rd31];
ld.global.nc.u16 %rs5, [%rd31+2];
mov.b32 %r105, {%rs4, %rs5};
shr.s32 %r106, %r105, %r12;
add.s64 %rd32, %rd29, %rd2;
mul.lo.s64 %rd33, %rd32, 36;
add.s64 %rd34, %rd7, %rd33;
add.s64 %rd36, %rd34, %rd35;
ld.global.nc.u32 %r94, [%rd36+4];
ld.global.nc.u32 %r86, [%rd34];

{.reg .f16 low,high;
mov.b32 {low,high},%r86;
cvt.f32.f16 %f7, low;}


ld.global.nc.u32 %r102, [%rd36+76];
ld.global.nc.u32 %r87, [%rd34+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r87;
cvt.f32.f16 %f8, low;}


ld.global.nc.u16 %rs1, [%rd40+208];

{ cvt.f32.f16 %f9, %rs1;}


ld.global.nc.u8 %rs6, [%rd41+-4];
cvt.u32.u16 %r107, %rs6;
cvt.s32.s8 %r108, %r107;
and.b32 %r109, %r104, 252645135;
shl.b32 %r110, %r106, 4;
and.b32 %r111, %r110, 808464432;
or.b32 %r90, %r111, %r109;
or.b32 %r112, %r90, -2139062144;
add.s32 %r113, %r112, -538976288;
xor.b32 %r114, %r113, -2139062144;
xor.b32 %r115, %r90, 538976288;
xor.b32 %r116, %r113, %r90;
and.b32 %r88, %r116, %r115;

prmt.b32 %r88, %r88, 0, 0xba98;


prmt.b32 %r90, %r90, 0, 0xba98;

xor.b32 %r117, %r90, 2139062143;
not.b32 %r118, %r88;
and.b32 %r119, %r114, %r118;
and.b32 %r120, %r117, %r88;
or.b32 %r93, %r120, %r119;
mov.u32 %r103, 0;

dp4a.s32.s32 %r92, %r93, %r94, %r103;

mul.lo.s32 %r121, %r92, %r108;
cvt.rn.f32.s32 %f10, %r121;
fma.rn.ftz.f32 %f11, %f7, %f10, 0f00000000;
ld.global.nc.u8 %rs7, [%rd41];
cvt.u32.u16 %r122, %rs7;
cvt.s32.s8 %r123, %r122;
shr.u32 %r124, %r104, 4;
and.b32 %r125, %r124, 252645135;
and.b32 %r126, %r106, 808464432;
or.b32 %r98, %r126, %r125;
or.b32 %r127, %r98, -2139062144;
add.s32 %r128, %r127, -538976288;
xor.b32 %r129, %r128, -2139062144;
xor.b32 %r130, %r98, 538976288;
xor.b32 %r131, %r128, %r98;
and.b32 %r96, %r131, %r130;

prmt.b32 %r96, %r96, 0, 0xba98;


prmt.b32 %r98, %r98, 0, 0xba98;

xor.b32 %r132, %r98, 2139062143;
not.b32 %r133, %r96;
and.b32 %r134, %r129, %r133;
and.b32 %r135, %r132, %r96;
or.b32 %r101, %r135, %r134;

dp4a.s32.s32 %r100, %r101, %r102, %r103;

mul.lo.s32 %r136, %r100, %r123;
cvt.rn.f32.s32 %f12, %r136;
fma.rn.ftz.f32 %f13, %f8, %f12, %f11;
fma.rn.ftz.f32 %f30, %f9, %f13, %f30;
add.s32 %r167, %r167, 32;
add.s64 %rd41, %rd41, 840;
add.s64 %rd40, %rd40, 840;
add.s32 %r168, %r168, 4;
setp.lt.s32 %p3, %r168, %r47;
@%p3 bra $L__BB188_5;

$L__BB188_6:
setp.eq.s32 %p4, %r1, 0;
@%p4 bra $L__BB188_8;

shl.b32 %r137, %r1, 7;
mov.u32 %r138, _ZZ13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r139, %r138, %r137;
shl.b32 %r140, %r2, 2;
add.s32 %r141, %r139, %r140;
st.shared.f32 [%r141+-128], %f30;

$L__BB188_8:
setp.ne.s32 %p5, %r1, 0;
bar.sync 0;
@%p5 bra $L__BB188_11;

mov.u32 %r142, %ctaid.y;
mov.u32 %r143, %ctaid.x;
mad.lo.s32 %r144, %r142, %r30, %r143;
mad.lo.s32 %r19, %r10, %r36, %r144;
shl.b32 %r145, %r2, 2;
mov.u32 %r146, _ZZ13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r147, %r146, %r145;
mov.u32 %r148, 2;
ld.shared.f32 %f14, [%r147];
add.ftz.f32 %f15, %f30, %f14;
ld.shared.f32 %f16, [%r147+128];
add.ftz.f32 %f17, %f15, %f16;
ld.shared.f32 %f18, [%r147+256];
add.ftz.f32 %f19, %f17, %f18;
mov.b32 %r149, %f19;
mov.u32 %r150, 31;
mov.u32 %r151, 16;
mov.u32 %r152, -1;
shfl.sync.bfly.b32 %r153|%p6, %r149, %r151, %r150, %r152;
mov.b32 %f20, %r153;
add.ftz.f32 %f21, %f19, %f20;
mov.b32 %r154, %f21;
mov.u32 %r155, 8;
shfl.sync.bfly.b32 %r156|%p7, %r154, %r155, %r150, %r152;
mov.b32 %f22, %r156;
add.ftz.f32 %f23, %f21, %f22;
mov.b32 %r157, %f23;
mov.u32 %r158, 4;
shfl.sync.bfly.b32 %r159|%p8, %r157, %r158, %r150, %r152;
mov.b32 %f24, %r159;
add.ftz.f32 %f25, %f23, %f24;
mov.b32 %r160, %f25;
shfl.sync.bfly.b32 %r161|%p9, %r160, %r148, %r150, %r152;
mov.b32 %f26, %r161;
add.ftz.f32 %f27, %f25, %f26;
mov.b32 %r162, %f27;
mov.u32 %r163, 1;
shfl.sync.bfly.b32 %r164|%p10, %r162, %r163, %r150, %r152;
mov.b32 %f28, %r164;
add.ftz.f32 %f4, %f27, %f28;
setp.ne.s32 %p11, %r2, 0;
@%p11 bra $L__BB188_11;

cvta.to.global.u64 %rd37, %rd16;
mul.wide.u32 %rd38, %r19, 4;
add.s64 %rd39, %rd37, %rd38;
st.global.f32 [%rd39], %f4;

$L__BB188_11:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<33>;
.reg .b16 %rs<13>;
.reg .f32 %f<126>;
.reg .b32 %r<267>;
.reg .b64 %rd<78>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[384];

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate[384];

ld.param.u64 %rd11, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd12, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd13, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r45, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r46, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
ld.param.f32 %f31, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+16];
ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+8];
ld.param.u64 %rd14, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3];
mov.u32 %r1, %tid.y;
shl.b32 %r47, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r47, %r2;
mov.u32 %r4, %ctaid.x;
mov.u32 %r5, %ctaid.y;
setp.eq.s64 %p2, %rd13, 0;
@%p2 bra $L__BB189_2;

cvta.to.global.u64 %rd20, %rd13;
mul.wide.u32 %rd21, %r5, 4;
add.s64 %rd22, %rd20, %rd21;
ld.global.nc.u32 %r263, [%rd22];
mul.hi.u32 %r48, %r5, %r31;
add.s32 %r49, %r48, %r5;
shr.u32 %r50, %r49, %r32;
mul.lo.s32 %r51, %r50, %r33;
sub.s32 %r264, %r5, %r51;
bra.uni $L__BB189_3;

$L__BB189_2:
mul.hi.u32 %r52, %r5, %r35;
add.s32 %r53, %r52, %r5;
shr.u32 %r263, %r53, %r36;
mov.u32 %r264, %r5;

$L__BB189_3:
mov.u32 %r11, %ctaid.z;
selp.b32 %r12, %r5, %r263, %p2;
or.b32 %r54, %r2, %r1;
setp.ne.s32 %p4, %r54, 0;
mov.f32 %f119, 0f00000000;
mov.f32 %f115, %f119;
mov.f32 %f116, %f119;
@%p4 bra $L__BB189_8;

setp.eq.s64 %p5, %rd14, 0;
mov.f32 %f115, 0f00000000;
mov.f32 %f116, %f115;
@%p5 bra $L__BB189_6;

cvta.to.global.u64 %rd23, %rd14;
mul.lo.s32 %r55, %r11, %r46;
cvt.u64.u32 %rd24, %r55;
mul.lo.s32 %r56, %r12, %r40;
cvt.u64.u32 %rd25, %r56;
cvt.u64.u32 %rd26, %r2;
cvt.s64.s32 %rd27, %r4;
add.s64 %rd28, %rd26, %rd27;
add.s64 %rd29, %rd28, %rd24;
add.s64 %rd30, %rd29, %rd25;
shl.b64 %rd31, %rd30, 2;
add.s64 %rd32, %rd23, %rd31;
ld.global.f32 %f116, [%rd32];

$L__BB189_6:
setp.eq.s64 %p6, %rd16, 0;
setp.eq.s64 %p7, %rd15, 0;
or.pred %p8, %p6, %p7;
@%p8 bra $L__BB189_8;

cvta.to.global.u64 %rd33, %rd16;
mul.lo.s32 %r58, %r11, %r46;
cvt.u64.u32 %rd34, %r58;
mul.lo.s32 %r59, %r12, %r40;
cvt.u64.u32 %rd35, %r59;
cvt.u64.u32 %rd36, %r2;
cvt.s64.s32 %rd37, %r4;
add.s64 %rd38, %rd36, %rd37;
add.s64 %rd39, %rd38, %rd34;
add.s64 %rd40, %rd39, %rd35;
shl.b64 %rd41, %rd40, 2;
add.s64 %rd42, %rd33, %rd41;
ld.global.f32 %f115, [%rd42];

$L__BB189_8:
shr.u32 %r61, %r30, 8;
shr.s32 %r62, %r3, 31;
shr.u32 %r63, %r62, 27;
add.s32 %r64, %r3, %r63;
shr.s32 %r266, %r64, 5;
setp.ge.s32 %p9, %r266, %r61;
mov.f32 %f121, %f119;
@%p9 bra $L__BB189_13;

and.b32 %r68, %r64, -32;
sub.s32 %r69, %r3, %r68;
shr.s32 %r70, %r69, 31;
shr.u32 %r71, %r70, 28;
add.s32 %r72, %r69, %r71;
shr.s32 %r73, %r72, 4;
shl.b32 %r74, %r73, 2;
and.b32 %r75, %r72, -16;
sub.s32 %r76, %r69, %r75;
shr.s32 %r77, %r76, 31;
shr.u32 %r78, %r77, 29;
add.s32 %r79, %r76, %r78;
shr.s32 %r80, %r79, 3;
add.s32 %r81, %r74, %r80;
shl.b32 %r82, %r73, 3;
shr.u32 %r83, %r77, 30;
add.s32 %r84, %r76, %r83;
shr.s32 %r85, %r84, 2;
add.s32 %r86, %r82, %r85;
shl.b32 %r14, %r80, 1;
shl.b32 %r15, %r69, 1;
shr.u32 %r87, %r70, 29;
add.s32 %r88, %r69, %r87;
and.b32 %r89, %r88, -8;
sub.s32 %r90, %r69, %r89;
add.s32 %r91, %r82, %r90;
shl.b32 %r16, %r91, 1;
cvt.s64.s32 %rd43, %r86;
cvt.s64.s32 %rd4, %r90;
cvt.s64.s32 %rd5, %r81;
shl.b32 %r265, %r266, 3;
mul.hi.u32 %r92, %r11, %r41;
add.s32 %r93, %r92, %r11;
shr.u32 %r94, %r93, %r42;
mad.lo.s32 %r95, %r94, %r44, %r266;
mad.lo.s32 %r96, %r263, %r38, %r95;
mad.lo.s32 %r97, %r4, %r34, %r96;
mul.wide.s32 %rd77, %r97, 210;
add.s64 %rd7, %rd43, 192;
mul.lo.s32 %r98, %r264, %r39;
cvt.u64.u32 %rd44, %r98;
mul.lo.s32 %r99, %r11, %r45;
cvt.u64.u32 %rd45, %r99;
add.s64 %rd8, %rd44, %rd45;
mov.f32 %f121, 0f00000000;
cvta.to.global.u64 %rd46, %rd12;
cvta.to.global.u64 %rd50, %rd11;
shl.b64 %rd59, %rd4, 2;
mov.f32 %f119, %f121;

$L__BB189_10:
cvt.u64.u32 %rd47, %r265;
add.s64 %rd48, %rd8, %rd47;
mul.wide.s32 %rd49, %r15, 2;
add.s64 %rd51, %rd50, %rd49;
add.s64 %rd52, %rd51, %rd77;
ld.global.nc.u16 %rs2, [%rd52];
ld.global.nc.u16 %rs3, [%rd52+2];
mov.b32 %r118, {%rs2, %rs3};
mul.wide.s32 %rd53, %r16, 2;
add.s64 %rd54, %rd50, %rd53;
add.s64 %rd55, %rd54, %rd77;
ld.global.nc.u16 %rs4, [%rd55+128];
ld.global.nc.u16 %rs5, [%rd55+130];
mov.b32 %r119, {%rs4, %rs5};
shr.s32 %r120, %r119, %r14;
add.s64 %rd56, %rd48, %rd5;
mul.lo.s64 %rd57, %rd56, 36;
add.s64 %rd58, %rd46, %rd57;
add.s64 %rd60, %rd58, %rd59;
ld.global.nc.u32 %r20, [%rd60+4];
ld.global.nc.u32 %r21, [%rd58];

{.reg .f16 low,high;
mov.b32 {low,high},%r21;
cvt.f32.f16 %f40, low;}


ld.global.nc.u32 %r22, [%rd60+76];
ld.global.nc.u32 %r23, [%rd58+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r23;
cvt.f32.f16 %f41, low;}


add.s64 %rd61, %rd50, %rd77;
ld.global.nc.u16 %rs1, [%rd61+208];

{ cvt.f32.f16 %f42, %rs1;}


add.s64 %rd62, %rd50, %rd7;
add.s64 %rd63, %rd62, %rd77;
ld.global.nc.u8 %rs6, [%rd63];
cvt.u32.u16 %r121, %rs6;
cvt.s32.s8 %r122, %r121;
and.b32 %r123, %r118, 252645135;
shl.b32 %r124, %r120, 4;
and.b32 %r125, %r124, 808464432;
or.b32 %r104, %r125, %r123;
or.b32 %r126, %r104, -2139062144;
add.s32 %r127, %r126, -538976288;
xor.b32 %r128, %r127, -2139062144;
xor.b32 %r129, %r104, 538976288;
xor.b32 %r130, %r127, %r104;
and.b32 %r102, %r130, %r129;

prmt.b32 %r102, %r102, 0, 0xba98;


prmt.b32 %r104, %r104, 0, 0xba98;

xor.b32 %r131, %r104, 2139062143;
not.b32 %r132, %r102;
and.b32 %r133, %r128, %r132;
and.b32 %r134, %r131, %r102;
or.b32 %r107, %r134, %r133;
mov.u32 %r117, 0;

dp4a.s32.s32 %r106, %r107, %r20, %r117;

mul.lo.s32 %r135, %r106, %r122;
cvt.rn.f32.s32 %f43, %r135;
fma.rn.ftz.f32 %f44, %f40, %f43, 0f00000000;
ld.global.nc.u8 %rs7, [%rd63+4];
cvt.u32.u16 %r136, %rs7;
cvt.s32.s8 %r137, %r136;
shr.u32 %r138, %r118, 4;
and.b32 %r139, %r138, 252645135;
and.b32 %r140, %r120, 808464432;
or.b32 %r112, %r140, %r139;
or.b32 %r141, %r112, -2139062144;
add.s32 %r142, %r141, -538976288;
xor.b32 %r143, %r142, -2139062144;
xor.b32 %r144, %r112, 538976288;
xor.b32 %r145, %r142, %r112;
and.b32 %r110, %r145, %r144;

prmt.b32 %r110, %r110, 0, 0xba98;


prmt.b32 %r112, %r112, 0, 0xba98;

xor.b32 %r146, %r112, 2139062143;
not.b32 %r147, %r110;
and.b32 %r148, %r143, %r147;
and.b32 %r149, %r146, %r110;
or.b32 %r115, %r149, %r148;

dp4a.s32.s32 %r114, %r115, %r22, %r117;

mul.lo.s32 %r150, %r114, %r137;
cvt.rn.f32.s32 %f45, %r150;
fma.rn.ftz.f32 %f46, %f41, %f45, %f44;
fma.rn.ftz.f32 %f121, %f42, %f46, %f121;
setp.eq.s64 %p10, %rd15, 0;
@%p10 bra $L__BB189_12;

cvta.to.global.u64 %rd65, %rd15;
add.s64 %rd66, %rd65, %rd49;
add.s64 %rd67, %rd66, %rd77;
ld.global.u16 %rs9, [%rd67];
ld.global.u16 %rs10, [%rd67+2];
mov.b32 %r169, {%rs9, %rs10};
add.s64 %rd69, %rd65, %rd53;
add.s64 %rd70, %rd69, %rd77;
ld.global.u16 %rs11, [%rd70+128];
ld.global.u16 %rs12, [%rd70+130];
mov.b32 %r170, {%rs11, %rs12};
shr.s32 %r171, %r170, %r14;

{.reg .f16 low,high;
mov.b32 {low,high},%r21;
cvt.f32.f16 %f47, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r23;
cvt.f32.f16 %f48, low;}


add.s64 %rd71, %rd65, %rd77;
ld.global.u16 %rs8, [%rd71+208];

{ cvt.f32.f16 %f49, %rs8;}


add.s64 %rd72, %rd65, %rd7;
add.s64 %rd73, %rd72, %rd77;
ld.global.s8 %r172, [%rd73];
and.b32 %r173, %r169, 252645135;
shl.b32 %r174, %r171, 4;
and.b32 %r175, %r174, 808464432;
or.b32 %r155, %r175, %r173;
or.b32 %r176, %r155, -2139062144;
add.s32 %r177, %r176, -538976288;
xor.b32 %r178, %r177, -2139062144;
xor.b32 %r179, %r155, 538976288;
xor.b32 %r180, %r177, %r155;
and.b32 %r153, %r180, %r179;

prmt.b32 %r153, %r153, 0, 0xba98;


prmt.b32 %r155, %r155, 0, 0xba98;

xor.b32 %r181, %r155, 2139062143;
not.b32 %r182, %r153;
and.b32 %r183, %r178, %r182;
and.b32 %r184, %r181, %r153;
or.b32 %r158, %r184, %r183;

dp4a.s32.s32 %r157, %r158, %r20, %r117;

mul.lo.s32 %r185, %r157, %r172;
cvt.rn.f32.s32 %f50, %r185;
fma.rn.ftz.f32 %f51, %f47, %f50, 0f00000000;
ld.global.s8 %r186, [%rd73+4];
shr.u32 %r187, %r169, 4;
and.b32 %r188, %r187, 252645135;
and.b32 %r189, %r171, 808464432;
or.b32 %r163, %r189, %r188;
or.b32 %r190, %r163, -2139062144;
add.s32 %r191, %r190, -538976288;
xor.b32 %r192, %r191, -2139062144;
xor.b32 %r193, %r163, 538976288;
xor.b32 %r194, %r191, %r163;
and.b32 %r161, %r194, %r193;

prmt.b32 %r161, %r161, 0, 0xba98;


prmt.b32 %r163, %r163, 0, 0xba98;

xor.b32 %r195, %r163, 2139062143;
not.b32 %r196, %r161;
and.b32 %r197, %r192, %r196;
and.b32 %r198, %r195, %r161;
or.b32 %r166, %r198, %r197;

dp4a.s32.s32 %r165, %r166, %r22, %r117;

mul.lo.s32 %r199, %r165, %r186;
cvt.rn.f32.s32 %f52, %r199;
fma.rn.ftz.f32 %f53, %f48, %f52, %f51;
fma.rn.ftz.f32 %f119, %f49, %f53, %f119;

$L__BB189_12:
add.s32 %r265, %r265, 32;
add.s64 %rd77, %rd77, 840;
add.s32 %r266, %r266, 4;
setp.lt.s32 %p11, %r266, %r61;
@%p11 bra $L__BB189_10;

$L__BB189_13:
setp.eq.s32 %p12, %r1, 0;
@%p12 bra $L__BB189_16;

setp.eq.s64 %p13, %rd15, 0;
add.s32 %r26, %r1, -1;
shl.b32 %r204, %r1, 7;
mov.u32 %r205, _ZZ13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r206, %r205, %r204;
shl.b32 %r207, %r2, 2;
add.s32 %r208, %r206, %r207;
st.shared.f32 [%r208+-128], %f121;
@%p13 bra $L__BB189_16;

shl.b32 %r210, %r26, 7;
mov.u32 %r211, _ZZ13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r212, %r211, %r210;
add.s32 %r214, %r212, %r207;
st.shared.f32 [%r214], %f119;

$L__BB189_16:
setp.ne.s32 %p14, %r1, 0;
bar.sync 0;
@%p14 bra $L__BB189_35;

setp.ne.s64 %p15, %rd15, 0;
shl.b32 %r217, %r2, 2;
mov.u32 %r218, _ZZ13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r27, %r218, %r217;
ld.shared.f32 %f54, [%r27];
add.ftz.f32 %f13, %f121, %f54;
@%p15 bra $L__BB189_19;
bra.uni $L__BB189_18;

$L__BB189_19:
mov.u32 %r221, _ZZ13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r222, %r221, %r217;
ld.shared.f32 %f58, [%r222];
add.ftz.f32 %f59, %f119, %f58;
ld.shared.f32 %f60, [%r27+128];
add.ftz.f32 %f61, %f13, %f60;
ld.shared.f32 %f62, [%r222+128];
add.ftz.f32 %f63, %f59, %f62;
ld.shared.f32 %f64, [%r27+256];
add.ftz.f32 %f123, %f61, %f64;
ld.shared.f32 %f65, [%r222+256];
add.ftz.f32 %f119, %f63, %f65;
bra.uni $L__BB189_20;

$L__BB189_18:
ld.shared.f32 %f55, [%r27+128];
add.ftz.f32 %f56, %f13, %f55;
ld.shared.f32 %f57, [%r27+256];
add.ftz.f32 %f123, %f56, %f57;

$L__BB189_20:
mov.b32 %r223, %f123;
mov.u32 %r224, 31;
mov.u32 %r225, 16;
mov.u32 %r226, -1;
shfl.sync.bfly.b32 %r227|%p16, %r223, %r225, %r224, %r226;
mov.b32 %f66, %r227;
add.ftz.f32 %f67, %f123, %f66;
mov.b32 %r228, %f67;
mov.u32 %r229, 8;
shfl.sync.bfly.b32 %r230|%p17, %r228, %r229, %r224, %r226;
mov.b32 %f68, %r230;
add.ftz.f32 %f69, %f67, %f68;
mov.b32 %r231, %f69;
mov.u32 %r232, 4;
shfl.sync.bfly.b32 %r233|%p18, %r231, %r232, %r224, %r226;
mov.b32 %f70, %r233;
add.ftz.f32 %f71, %f69, %f70;
mov.b32 %r234, %f71;
mov.u32 %r235, 2;
shfl.sync.bfly.b32 %r236|%p19, %r234, %r235, %r224, %r226;
mov.b32 %f72, %r236;
add.ftz.f32 %f19, %f71, %f72;
mov.b32 %r237, %f19;
mov.u32 %r238, 1;
shfl.sync.bfly.b32 %r28|%p1, %r237, %r238, %r224, %r226;
setp.eq.s64 %p20, %rd15, 0;
@%p20 bra $L__BB189_22;

mov.b32 %r239, %f119;
shfl.sync.bfly.b32 %r243|%p21, %r239, %r225, %r224, %r226;
mov.b32 %f73, %r243;
add.ftz.f32 %f74, %f119, %f73;
mov.b32 %r244, %f74;
shfl.sync.bfly.b32 %r246|%p22, %r244, %r229, %r224, %r226;
mov.b32 %f75, %r246;
add.ftz.f32 %f76, %f74, %f75;
mov.b32 %r247, %f76;
shfl.sync.bfly.b32 %r249|%p23, %r247, %r232, %r224, %r226;
mov.b32 %f77, %r249;
add.ftz.f32 %f78, %f76, %f77;
mov.b32 %r250, %f78;
shfl.sync.bfly.b32 %r252|%p24, %r250, %r235, %r224, %r226;
mov.b32 %f79, %r252;
add.ftz.f32 %f80, %f78, %f79;
mov.b32 %r253, %f80;
shfl.sync.bfly.b32 %r255|%p25, %r253, %r238, %r224, %r226;
mov.b32 %f81, %r255;
add.ftz.f32 %f119, %f80, %f81;

$L__BB189_22:
setp.ne.s32 %p26, %r2, 0;
mov.b32 %f82, %r28;
add.ftz.f32 %f22, %f19, %f82;
@%p26 bra $L__BB189_35;

add.ftz.f32 %f125, %f116, %f22;
@%p20 bra $L__BB189_34;

add.ftz.f32 %f24, %f115, %f119;
setp.gt.s32 %p28, %r29, 2;
@%p28 bra $L__BB189_28;

setp.eq.s32 %p31, %r29, 1;
@%p31 bra $L__BB189_32;

setp.eq.s32 %p32, %r29, 2;
@%p32 bra $L__BB189_27;
bra.uni $L__BB189_33;

$L__BB189_27:
mul.ftz.f32 %f102, %f24, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f103, %f102;
add.ftz.f32 %f104, %f103, 0f3F800000;
div.approx.ftz.f32 %f105, %f24, %f104;
mul.ftz.f32 %f125, %f125, %f105;
bra.uni $L__BB189_34;

$L__BB189_28:
setp.eq.s32 %p29, %r29, 3;
@%p29 bra $L__BB189_31;

setp.ne.s32 %p30, %r29, 6;
@%p30 bra $L__BB189_33;

min.ftz.f32 %f83, %f24, %f31;
neg.ftz.f32 %f84, %f31;
min.ftz.f32 %f85, %f125, %f31;
max.ftz.f32 %f86, %f85, %f84;
mul.ftz.f32 %f87, %f83, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f88, %f87;
add.ftz.f32 %f89, %f88, 0f3F800000;
div.approx.ftz.f32 %f90, %f83, %f89;
mul.ftz.f32 %f125, %f86, %f90;
bra.uni $L__BB189_34;

$L__BB189_32:
mul.ftz.f32 %f106, %f24, 0f3F000000;
mul.ftz.f32 %f107, %f24, 0f3D372713;
fma.rn.ftz.f32 %f108, %f24, %f107, 0f3F800000;
mul.ftz.f32 %f109, %f24, 0f3F4C422A;
mul.ftz.f32 %f110, %f109, %f108;
tanh.approx.f32 %f111, %f110;
add.ftz.f32 %f112, %f111, 0f3F800000;
mul.ftz.f32 %f113, %f106, %f112;
mul.ftz.f32 %f125, %f125, %f113;
bra.uni $L__BB189_34;

$L__BB189_31:
mov.f32 %f91, 0f40E00000;
min.ftz.f32 %f92, %f24, %f91;
min.ftz.f32 %f93, %f125, %f91;
mov.f32 %f94, 0fC0E00000;
max.ftz.f32 %f95, %f93, %f94;
mul.ftz.f32 %f96, %f92, 0f3FD9DB23;
mul.ftz.f32 %f97, %f96, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f98, %f97;
add.ftz.f32 %f99, %f98, 0f3F800000;
div.approx.ftz.f32 %f100, %f92, %f99;
add.ftz.f32 %f101, %f95, 0f3F800000;
mul.ftz.f32 %f125, %f101, %f100;
bra.uni $L__BB189_34;

$L__BB189_33:
mul.ftz.f32 %f125, %f125, %f24;

$L__BB189_34:
cvta.to.global.u64 %rd74, %rd19;
mov.u32 %r259, %ctaid.y;
mad.lo.s32 %r261, %r259, %r40, %r4;
mad.lo.s32 %r262, %r11, %r46, %r261;
mul.wide.u32 %rd75, %r262, 4;
add.s64 %rd76, %rd74, %rd75;
st.global.f32 [%rd76], %f125;

$L__BB189_35:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<12>;
.reg .b16 %rs<8>;
.reg .f32 %f<31>;
.reg .b32 %r<169>;
.reg .b64 %rd<42>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[384];

ld.param.u64 %rd13, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd14, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
mov.u32 %r1, %tid.y;
shl.b32 %r37, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r37, %r2;
mov.u32 %r166, %ctaid.y;
setp.eq.s64 %p1, %rd15, 0;
@%p1 bra $L__BB190_2;

cvta.to.global.u64 %rd17, %rd15;
mul.wide.u32 %rd18, %r166, 4;
add.s64 %rd19, %rd17, %rd18;
ld.global.nc.u32 %r165, [%rd19];
mul.hi.u32 %r38, %r166, %r21;
add.s32 %r39, %r38, %r166;
shr.u32 %r40, %r39, %r22;
mul.lo.s32 %r41, %r40, %r23;
sub.s32 %r166, %r166, %r41;
bra.uni $L__BB190_3;

$L__BB190_2:
mul.hi.u32 %r42, %r166, %r25;
add.s32 %r43, %r42, %r166;
shr.u32 %r165, %r43, %r26;

$L__BB190_3:
mov.u32 %r10, %ctaid.z;
shr.s32 %r44, %r3, 31;
shr.u32 %r45, %r44, 27;
add.s32 %r46, %r3, %r45;
shr.s32 %r168, %r46, 5;
shr.u32 %r47, %r20, 8;
setp.ge.s32 %p2, %r168, %r47;
mov.f32 %f30, 0f00000000;
@%p2 bra $L__BB190_6;

cvta.to.global.u64 %rd20, %rd13;
and.b32 %r51, %r46, -32;
sub.s32 %r52, %r3, %r51;
shr.s32 %r53, %r52, 31;
shr.u32 %r54, %r53, 28;
add.s32 %r55, %r52, %r54;
shr.s32 %r56, %r55, 4;
shl.b32 %r57, %r56, 2;
and.b32 %r58, %r55, -16;
sub.s32 %r59, %r52, %r58;
shr.s32 %r60, %r59, 31;
shr.u32 %r61, %r60, 29;
add.s32 %r62, %r59, %r61;
shr.s32 %r63, %r62, 3;
add.s32 %r64, %r57, %r63;
shl.b32 %r65, %r56, 3;
shr.u32 %r66, %r60, 30;
add.s32 %r67, %r59, %r66;
shr.s32 %r68, %r67, 2;
add.s32 %r69, %r65, %r68;
shl.b32 %r12, %r63, 1;
shl.b32 %r70, %r52, 1;
shr.u32 %r71, %r53, 29;
add.s32 %r72, %r52, %r71;
and.b32 %r73, %r72, -8;
sub.s32 %r74, %r52, %r73;
add.s32 %r75, %r65, %r74;
shl.b32 %r76, %r75, 1;
cvt.s64.s32 %rd21, %r69;
cvt.s64.s32 %rd1, %r74;
cvt.s64.s32 %rd2, %r64;
shl.b32 %r167, %r168, 3;
mul.hi.u32 %r77, %r10, %r31;
add.s32 %r78, %r77, %r10;
shr.u32 %r79, %r78, %r32;
mad.lo.s32 %r80, %r79, %r34, %r168;
mad.lo.s32 %r81, %r165, %r28, %r80;
mov.u32 %r82, %ctaid.x;
mad.lo.s32 %r83, %r82, %r24, %r81;
mul.wide.s32 %rd22, %r83, 210;
add.s64 %rd23, %rd22, %rd21;
add.s64 %rd24, %rd20, %rd23;
add.s64 %rd41, %rd24, 196;
add.s64 %rd40, %rd20, %rd22;
mul.wide.s32 %rd25, %r76, 2;
add.s64 %rd5, %rd25, 128;
mul.wide.s32 %rd6, %r70, 2;
cvta.to.global.u64 %rd7, %rd14;
mul.lo.s32 %r84, %r166, %r29;
cvt.u64.u32 %rd26, %r84;
mul.lo.s32 %r85, %r10, %r35;
cvt.u64.u32 %rd27, %r85;
add.s64 %rd8, %rd26, %rd27;
mov.f32 %f30, 0f00000000;
shl.b64 %rd35, %rd1, 2;

$L__BB190_5:
cvt.u64.u32 %rd28, %r167;
add.s64 %rd29, %rd8, %rd28;
add.s64 %rd30, %rd40, %rd6;
ld.global.nc.u16 %rs2, [%rd30];
ld.global.nc.u16 %rs3, [%rd30+2];
mov.b32 %r104, {%rs2, %rs3};
add.s64 %rd31, %rd40, %rd5;
ld.global.nc.u16 %rs4, [%rd31];
ld.global.nc.u16 %rs5, [%rd31+2];
mov.b32 %r105, {%rs4, %rs5};
shr.s32 %r106, %r105, %r12;
add.s64 %rd32, %rd29, %rd2;
mul.lo.s64 %rd33, %rd32, 36;
add.s64 %rd34, %rd7, %rd33;
add.s64 %rd36, %rd34, %rd35;
ld.global.nc.u32 %r94, [%rd36+4];
ld.global.nc.u32 %r86, [%rd34];

{.reg .f16 low,high;
mov.b32 {low,high},%r86;
cvt.f32.f16 %f7, low;}


ld.global.nc.u32 %r102, [%rd36+76];
ld.global.nc.u32 %r87, [%rd34+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r87;
cvt.f32.f16 %f8, low;}


ld.global.nc.u16 %rs1, [%rd40+208];

{ cvt.f32.f16 %f9, %rs1;}


ld.global.nc.u8 %rs6, [%rd41+-4];
cvt.u32.u16 %r107, %rs6;
cvt.s32.s8 %r108, %r107;
and.b32 %r109, %r104, 252645135;
shl.b32 %r110, %r106, 4;
and.b32 %r111, %r110, 808464432;
or.b32 %r90, %r111, %r109;
or.b32 %r112, %r90, -2139062144;
add.s32 %r113, %r112, -538976288;
xor.b32 %r114, %r113, -2139062144;
xor.b32 %r115, %r90, 538976288;
xor.b32 %r116, %r113, %r90;
and.b32 %r88, %r116, %r115;

prmt.b32 %r88, %r88, 0, 0xba98;


prmt.b32 %r90, %r90, 0, 0xba98;

xor.b32 %r117, %r90, 2139062143;
not.b32 %r118, %r88;
and.b32 %r119, %r114, %r118;
and.b32 %r120, %r117, %r88;
or.b32 %r93, %r120, %r119;
mov.u32 %r103, 0;

dp4a.s32.s32 %r92, %r93, %r94, %r103;

mul.lo.s32 %r121, %r92, %r108;
cvt.rn.f32.s32 %f10, %r121;
fma.rn.ftz.f32 %f11, %f7, %f10, 0f00000000;
ld.global.nc.u8 %rs7, [%rd41];
cvt.u32.u16 %r122, %rs7;
cvt.s32.s8 %r123, %r122;
shr.u32 %r124, %r104, 4;
and.b32 %r125, %r124, 252645135;
and.b32 %r126, %r106, 808464432;
or.b32 %r98, %r126, %r125;
or.b32 %r127, %r98, -2139062144;
add.s32 %r128, %r127, -538976288;
xor.b32 %r129, %r128, -2139062144;
xor.b32 %r130, %r98, 538976288;
xor.b32 %r131, %r128, %r98;
and.b32 %r96, %r131, %r130;

prmt.b32 %r96, %r96, 0, 0xba98;


prmt.b32 %r98, %r98, 0, 0xba98;

xor.b32 %r132, %r98, 2139062143;
not.b32 %r133, %r96;
and.b32 %r134, %r129, %r133;
and.b32 %r135, %r132, %r96;
or.b32 %r101, %r135, %r134;

dp4a.s32.s32 %r100, %r101, %r102, %r103;

mul.lo.s32 %r136, %r100, %r123;
cvt.rn.f32.s32 %f12, %r136;
fma.rn.ftz.f32 %f13, %f8, %f12, %f11;
fma.rn.ftz.f32 %f30, %f9, %f13, %f30;
add.s32 %r167, %r167, 32;
add.s64 %rd41, %rd41, 840;
add.s64 %rd40, %rd40, 840;
add.s32 %r168, %r168, 4;
setp.lt.s32 %p3, %r168, %r47;
@%p3 bra $L__BB190_5;

$L__BB190_6:
setp.eq.s32 %p4, %r1, 0;
@%p4 bra $L__BB190_8;

shl.b32 %r137, %r1, 7;
mov.u32 %r138, _ZZ13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r139, %r138, %r137;
shl.b32 %r140, %r2, 2;
add.s32 %r141, %r139, %r140;
st.shared.f32 [%r141+-128], %f30;

$L__BB190_8:
setp.ne.s32 %p5, %r1, 0;
bar.sync 0;
@%p5 bra $L__BB190_11;

mov.u32 %r142, %ctaid.y;
mov.u32 %r143, %ctaid.x;
mad.lo.s32 %r144, %r142, %r30, %r143;
mad.lo.s32 %r19, %r10, %r36, %r144;
shl.b32 %r145, %r2, 2;
mov.u32 %r146, _ZZ13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r147, %r146, %r145;
mov.u32 %r148, 2;
ld.shared.f32 %f14, [%r147];
add.ftz.f32 %f15, %f30, %f14;
ld.shared.f32 %f16, [%r147+128];
add.ftz.f32 %f17, %f15, %f16;
ld.shared.f32 %f18, [%r147+256];
add.ftz.f32 %f19, %f17, %f18;
mov.b32 %r149, %f19;
mov.u32 %r150, 31;
mov.u32 %r151, 16;
mov.u32 %r152, -1;
shfl.sync.bfly.b32 %r153|%p6, %r149, %r151, %r150, %r152;
mov.b32 %f20, %r153;
add.ftz.f32 %f21, %f19, %f20;
mov.b32 %r154, %f21;
mov.u32 %r155, 8;
shfl.sync.bfly.b32 %r156|%p7, %r154, %r155, %r150, %r152;
mov.b32 %f22, %r156;
add.ftz.f32 %f23, %f21, %f22;
mov.b32 %r157, %f23;
mov.u32 %r158, 4;
shfl.sync.bfly.b32 %r159|%p8, %r157, %r158, %r150, %r152;
mov.b32 %f24, %r159;
add.ftz.f32 %f25, %f23, %f24;
mov.b32 %r160, %f25;
shfl.sync.bfly.b32 %r161|%p9, %r160, %r148, %r150, %r152;
mov.b32 %f26, %r161;
add.ftz.f32 %f27, %f25, %f26;
mov.b32 %r162, %f27;
mov.u32 %r163, 1;
shfl.sync.bfly.b32 %r164|%p10, %r162, %r163, %r150, %r152;
mov.b32 %f28, %r164;
add.ftz.f32 %f4, %f27, %f28;
setp.ne.s32 %p11, %r2, 0;
@%p11 bra $L__BB190_11;

cvta.to.global.u64 %rd37, %rd16;
mul.wide.u32 %rd38, %r19, 4;
add.s64 %rd39, %rd37, %rd38;
st.global.f32 [%rd39], %f4;

$L__BB190_11:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type14ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type14ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type14ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type14ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type14ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type14ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type14ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type14ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type14ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<35>;
.reg .b16 %rs<17>;
.reg .f32 %f<121>;
.reg .b32 %r<328>;
.reg .b64 %rd<63>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type14ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type14ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd20, [_Z13mul_mat_vec_qIL9ggml_type14ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd21, [_Z13mul_mat_vec_qIL9ggml_type14ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type14ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r19, [_Z13mul_mat_vec_qIL9ggml_type14ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type14ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type14ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type14ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type14ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type14ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type14ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type14ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type14ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type14ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type14ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type14ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type14ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
cvta.to.global.u64 %rd1, %rd21;
mov.u32 %r1, %tid.y;
shl.b32 %r35, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r35, %r2;
mov.u32 %r36, %ctaid.x;
shl.b32 %r4, %r36, 1;
mov.u32 %r5, %ctaid.y;
shr.s32 %r37, %r3, 31;
shr.u32 %r38, %r37, 27;
add.s32 %r39, %r3, %r38;
shr.s32 %r327, %r39, 5;
shr.u32 %r7, %r34, 8;
setp.ge.s32 %p3, %r327, %r7;
mov.f32 %f117, 0f00000000;
mov.f32 %f118, %f117;
mov.f32 %f119, %f117;
mov.f32 %f120, %f117;
@%p3 bra $L__BB191_3;

mov.u32 %r40, %ctaid.z;
mul.lo.s32 %r41, %r5, %r26;
cvt.u64.u32 %rd22, %r41;
mul.lo.s32 %r42, %r40, %r32;
cvt.u64.u32 %rd23, %r42;
add.s64 %rd2, %rd23, %rd22;
mul.hi.u32 %r43, %r40, %r28;
add.s32 %r44, %r43, %r40;
shr.u32 %r45, %r44, %r29;
mul.lo.s32 %r46, %r45, %r31;
mul.hi.u32 %r47, %r5, %r22;
add.s32 %r48, %r47, %r5;
shr.u32 %r49, %r48, %r23;
mul.lo.s32 %r50, %r49, %r25;
cvta.to.global.u64 %rd62, %rd19;
and.b32 %r54, %r39, -32;
sub.s32 %r55, %r3, %r54;
shr.s32 %r56, %r55, 31;
shr.u32 %r57, %r56, 28;
add.s32 %r58, %r55, %r57;
shr.s32 %r59, %r58, 4;
shl.b32 %r60, %r59, 2;
and.b32 %r61, %r58, -16;
sub.s32 %r62, %r55, %r61;
shr.s32 %r63, %r62, 31;
shr.u32 %r64, %r63, 29;
add.s32 %r65, %r62, %r64;
shr.s32 %r66, %r65, 3;
add.s32 %r67, %r60, %r66;
shl.b32 %r68, %r59, 3;
shr.u32 %r69, %r63, 30;
add.s32 %r70, %r62, %r69;
shr.s32 %r71, %r70, 2;
add.s32 %r72, %r68, %r71;
shl.b32 %r8, %r66, 1;
shl.b32 %r73, %r55, 1;
shr.u32 %r74, %r56, 29;
add.s32 %r75, %r55, %r74;
and.b32 %r76, %r75, -8;
sub.s32 %r77, %r55, %r76;
add.s32 %r78, %r68, %r77;
shl.b32 %r79, %r78, 1;
cvt.s64.s32 %rd24, %r72;
cvt.s64.s32 %rd4, %r77;
add.s32 %r80, %r4, 1;
cvt.s64.s32 %rd5, %r67;
shl.b32 %r325, %r327, 3;
add.s32 %r326, %r20, %r325;
mad.lo.s32 %r81, %r4, %r19, %r327;
add.s32 %r82, %r81, %r46;
add.s32 %r83, %r82, %r50;
mul.wide.s32 %rd6, %r83, 210;
mul.wide.s32 %rd25, %r73, 2;
add.s64 %rd26, %rd6, %rd25;
add.s64 %rd7, %rd26, 2;
add.s32 %r84, %r327, %r46;
add.s32 %r85, %r84, %r50;
mad.lo.s32 %r86, %r80, %r19, %r85;
mul.wide.s32 %rd8, %r86, 210;
add.s64 %rd27, %rd8, %rd24;
add.s64 %rd9, %rd27, 192;
mul.wide.s32 %rd28, %r79, 2;
add.s64 %rd29, %rd6, %rd28;
add.s64 %rd10, %rd29, 130;
add.s64 %rd30, %rd8, %rd28;
add.s64 %rd11, %rd30, 128;
add.s64 %rd12, %rd8, %rd25;
add.s64 %rd31, %rd6, %rd24;
add.s64 %rd13, %rd31, 192;
cvta.to.global.u64 %rd14, %rd20;
mov.f32 %f117, 0f00000000;
shl.b64 %rd39, %rd4, 2;

$L__BB191_2:
cvt.u64.u32 %rd32, %r325;
add.s64 %rd33, %rd2, %rd32;
add.s64 %rd34, %rd62, %rd7;
ld.global.nc.u16 %rs5, [%rd34+-2];
ld.global.nc.u16 %rs6, [%rd34];
mov.b32 %r159, {%rs5, %rs6};
add.s64 %rd35, %rd62, %rd10;
ld.global.nc.u16 %rs7, [%rd35+-2];
ld.global.nc.u16 %rs8, [%rd35];
mov.b32 %r160, {%rs7, %rs8};
shr.s32 %r161, %r160, %r8;
add.s64 %rd36, %rd33, %rd5;
mul.lo.s64 %rd37, %rd36, 36;
add.s64 %rd38, %rd14, %rd37;
add.s64 %rd40, %rd38, %rd39;
ld.global.nc.u32 %r95, [%rd40+4];
ld.global.nc.u32 %r87, [%rd38];

{.reg .f16 low,high;
mov.b32 {low,high},%r87;
cvt.f32.f16 %f25, low;}


ld.global.nc.u32 %r103, [%rd40+76];
ld.global.nc.u32 %r88, [%rd38+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r88;
cvt.f32.f16 %f26, low;}


add.s64 %rd41, %rd62, %rd6;
ld.global.nc.u16 %rs1, [%rd41+208];

{ cvt.f32.f16 %f27, %rs1;}


add.s64 %rd42, %rd62, %rd13;
ld.global.nc.u8 %rs9, [%rd42];
cvt.u32.u16 %r162, %rs9;
cvt.s32.s8 %r163, %r162;
and.b32 %r164, %r159, 252645135;
shl.b32 %r165, %r161, 4;
and.b32 %r166, %r165, 808464432;
or.b32 %r127, %r166, %r164;
or.b32 %r167, %r127, -2139062144;
add.s32 %r168, %r167, -538976288;
xor.b32 %r169, %r168, -2139062144;
xor.b32 %r170, %r127, 538976288;
xor.b32 %r171, %r168, %r127;
and.b32 %r125, %r171, %r170;
mov.u32 %r89, %r125;

prmt.b32 %r89, %r89, 0, 0xba98;

mov.u32 %r91, %r127;

prmt.b32 %r91, %r91, 0, 0xba98;

xor.b32 %r172, %r91, 2139062143;
not.b32 %r173, %r89;
and.b32 %r174, %r169, %r173;
and.b32 %r175, %r172, %r89;
or.b32 %r94, %r175, %r174;
mov.u32 %r158, 0;

dp4a.s32.s32 %r93, %r94, %r95, %r158;

mul.lo.s32 %r176, %r93, %r163;
cvt.rn.f32.s32 %f37, %r176;
fma.rn.ftz.f32 %f38, %f25, %f37, 0f00000000;
ld.global.nc.u8 %rs10, [%rd42+4];
cvt.u32.u16 %r177, %rs10;
cvt.s32.s8 %r178, %r177;
shr.u32 %r179, %r159, 4;
and.b32 %r180, %r179, 252645135;
and.b32 %r181, %r161, 808464432;
or.b32 %r135, %r181, %r180;
or.b32 %r182, %r135, -2139062144;
add.s32 %r183, %r182, -538976288;
xor.b32 %r184, %r183, -2139062144;
xor.b32 %r185, %r135, 538976288;
xor.b32 %r186, %r183, %r135;
and.b32 %r133, %r186, %r185;
mov.u32 %r97, %r133;

prmt.b32 %r97, %r97, 0, 0xba98;

mov.u32 %r99, %r135;

prmt.b32 %r99, %r99, 0, 0xba98;

xor.b32 %r187, %r99, 2139062143;
not.b32 %r188, %r97;
and.b32 %r189, %r184, %r188;
and.b32 %r190, %r187, %r97;
or.b32 %r102, %r190, %r189;

dp4a.s32.s32 %r101, %r102, %r103, %r158;

mul.lo.s32 %r191, %r101, %r178;
cvt.rn.f32.s32 %f39, %r191;
fma.rn.ftz.f32 %f40, %f26, %f39, %f38;
fma.rn.ftz.f32 %f120, %f27, %f40, %f120;
add.s64 %rd43, %rd62, %rd12;
ld.global.nc.u16 %rs11, [%rd43];
ld.global.nc.u16 %rs12, [%rd43+2];
mov.b32 %r192, {%rs11, %rs12};
add.s64 %rd44, %rd62, %rd11;
ld.global.nc.u16 %rs13, [%rd44];
ld.global.nc.u16 %rs14, [%rd44+2];
mov.b32 %r193, {%rs13, %rs14};
shr.s32 %r194, %r193, %r8;

{.reg .f16 low,high;
mov.b32 {low,high},%r87;
cvt.f32.f16 %f28, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r88;
cvt.f32.f16 %f29, low;}


add.s64 %rd45, %rd62, %rd8;
ld.global.nc.u16 %rs2, [%rd45+208];

{ cvt.f32.f16 %f30, %rs2;}


add.s64 %rd46, %rd62, %rd9;
ld.global.nc.u8 %rs15, [%rd46];
cvt.u32.u16 %r195, %rs15;
cvt.s32.s8 %r196, %r195;
and.b32 %r197, %r192, 252645135;
shl.b32 %r198, %r194, 4;
and.b32 %r199, %r198, 808464432;
or.b32 %r145, %r199, %r197;
or.b32 %r200, %r145, -2139062144;
add.s32 %r201, %r200, -538976288;
xor.b32 %r202, %r201, -2139062144;
xor.b32 %r203, %r145, 538976288;
xor.b32 %r204, %r201, %r145;
and.b32 %r143, %r204, %r203;
mov.u32 %r107, %r143;

prmt.b32 %r107, %r107, 0, 0xba98;

mov.u32 %r109, %r145;

prmt.b32 %r109, %r109, 0, 0xba98;

xor.b32 %r205, %r109, 2139062143;
not.b32 %r206, %r107;
and.b32 %r207, %r202, %r206;
and.b32 %r208, %r205, %r107;
or.b32 %r112, %r208, %r207;

dp4a.s32.s32 %r111, %r112, %r95, %r158;

mul.lo.s32 %r209, %r111, %r196;
cvt.rn.f32.s32 %f41, %r209;
fma.rn.ftz.f32 %f42, %f28, %f41, 0f00000000;
ld.global.nc.u8 %rs16, [%rd46+4];
cvt.u32.u16 %r210, %rs16;
cvt.s32.s8 %r211, %r210;
shr.u32 %r212, %r192, 4;
and.b32 %r213, %r212, 252645135;
and.b32 %r214, %r194, 808464432;
or.b32 %r153, %r214, %r213;
or.b32 %r215, %r153, -2139062144;
add.s32 %r216, %r215, -538976288;
xor.b32 %r217, %r216, -2139062144;
xor.b32 %r218, %r153, 538976288;
xor.b32 %r219, %r216, %r153;
and.b32 %r151, %r219, %r218;
mov.u32 %r115, %r151;

prmt.b32 %r115, %r115, 0, 0xba98;

mov.u32 %r117, %r153;

prmt.b32 %r117, %r117, 0, 0xba98;

xor.b32 %r220, %r117, 2139062143;
not.b32 %r221, %r115;
and.b32 %r222, %r217, %r221;
and.b32 %r223, %r220, %r115;
or.b32 %r120, %r223, %r222;

dp4a.s32.s32 %r119, %r120, %r103, %r158;

mul.lo.s32 %r224, %r119, %r211;
cvt.rn.f32.s32 %f43, %r224;
fma.rn.ftz.f32 %f44, %f29, %f43, %f42;
fma.rn.ftz.f32 %f119, %f30, %f44, %f119;
cvt.u64.u32 %rd47, %r326;
add.s64 %rd48, %rd2, %rd47;
add.s64 %rd49, %rd48, %rd5;
mul.lo.s64 %rd50, %rd49, 36;
add.s64 %rd51, %rd14, %rd50;
add.s64 %rd52, %rd51, %rd39;
ld.global.nc.u32 %r131, [%rd52+4];
ld.global.nc.u32 %r123, [%rd51];

{.reg .f16 low,high;
mov.b32 {low,high},%r123;
cvt.f32.f16 %f31, low;}


ld.global.nc.u32 %r139, [%rd52+76];
ld.global.nc.u32 %r124, [%rd51+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r124;
cvt.f32.f16 %f32, low;}



{ cvt.f32.f16 %f33, %rs1;}



prmt.b32 %r125, %r125, 0, 0xba98;


prmt.b32 %r127, %r127, 0, 0xba98;

xor.b32 %r225, %r127, 2139062143;
not.b32 %r226, %r125;
and.b32 %r227, %r169, %r226;
and.b32 %r228, %r225, %r125;
or.b32 %r130, %r228, %r227;

dp4a.s32.s32 %r129, %r130, %r131, %r158;

mul.lo.s32 %r229, %r129, %r163;
cvt.rn.f32.s32 %f45, %r229;
fma.rn.ftz.f32 %f46, %f31, %f45, 0f00000000;

prmt.b32 %r133, %r133, 0, 0xba98;


prmt.b32 %r135, %r135, 0, 0xba98;

xor.b32 %r230, %r135, 2139062143;
not.b32 %r231, %r133;
and.b32 %r232, %r184, %r231;
and.b32 %r233, %r230, %r133;
or.b32 %r138, %r233, %r232;

dp4a.s32.s32 %r137, %r138, %r139, %r158;

mul.lo.s32 %r234, %r137, %r178;
cvt.rn.f32.s32 %f47, %r234;
fma.rn.ftz.f32 %f48, %f32, %f47, %f46;
fma.rn.ftz.f32 %f118, %f33, %f48, %f118;

{.reg .f16 low,high;
mov.b32 {low,high},%r123;
cvt.f32.f16 %f34, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r124;
cvt.f32.f16 %f35, low;}



{ cvt.f32.f16 %f36, %rs2;}



prmt.b32 %r143, %r143, 0, 0xba98;


prmt.b32 %r145, %r145, 0, 0xba98;

xor.b32 %r235, %r145, 2139062143;
not.b32 %r236, %r143;
and.b32 %r237, %r202, %r236;
and.b32 %r238, %r235, %r143;
or.b32 %r148, %r238, %r237;

dp4a.s32.s32 %r147, %r148, %r131, %r158;

mul.lo.s32 %r239, %r147, %r196;
cvt.rn.f32.s32 %f49, %r239;
fma.rn.ftz.f32 %f50, %f34, %f49, 0f00000000;

prmt.b32 %r151, %r151, 0, 0xba98;


prmt.b32 %r153, %r153, 0, 0xba98;

xor.b32 %r240, %r153, 2139062143;
not.b32 %r241, %r151;
and.b32 %r242, %r217, %r241;
and.b32 %r243, %r240, %r151;
or.b32 %r156, %r243, %r242;

dp4a.s32.s32 %r155, %r156, %r139, %r158;

mul.lo.s32 %r244, %r155, %r211;
cvt.rn.f32.s32 %f51, %r244;
fma.rn.ftz.f32 %f52, %f35, %f51, %f50;
fma.rn.ftz.f32 %f117, %f36, %f52, %f117;
add.s32 %r326, %r326, 32;
add.s32 %r325, %r325, 32;
add.s64 %rd62, %rd62, 840;
add.s32 %r327, %r327, 4;
setp.lt.s32 %p4, %r327, %r7;
@%p4 bra $L__BB191_2;

$L__BB191_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB191_5;

shl.b32 %r245, %r1, 9;
mov.u32 %r246, _ZZ13mul_mat_vec_qIL9ggml_type14ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r247, %r246, %r245;
shl.b32 %r248, %r2, 2;
add.s32 %r249, %r247, %r248;
st.shared.f32 [%r249+-512], %f120;
st.shared.f32 [%r249+-384], %f119;
st.shared.f32 [%r249+-256], %f118;
st.shared.f32 [%r249+-128], %f117;

$L__BB191_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB191_14;

mov.u32 %r250, %ctaid.z;
mad.lo.s32 %r251, %r5, %r27, %r4;
mad.lo.s32 %r252, %r250, %r33, %r251;
cvt.u64.u32 %rd17, %r252;
shl.b32 %r253, %r2, 2;
mov.u32 %r254, _ZZ13mul_mat_vec_qIL9ggml_type14ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r17, %r254, %r253;
mov.u32 %r255, 2;
ld.shared.f32 %f53, [%r17];
add.ftz.f32 %f54, %f53, %f120;
ld.shared.f32 %f55, [%r17+512];
add.ftz.f32 %f56, %f55, %f54;
ld.shared.f32 %f57, [%r17+1024];
add.ftz.f32 %f58, %f57, %f56;
mov.b32 %r256, %f58;
mov.u32 %r257, 31;
mov.u32 %r258, 16;
mov.u32 %r259, -1;
shfl.sync.bfly.b32 %r260|%p7, %r256, %r258, %r257, %r259;
mov.b32 %f59, %r260;
add.ftz.f32 %f60, %f58, %f59;
mov.b32 %r261, %f60;
mov.u32 %r262, 8;
shfl.sync.bfly.b32 %r263|%p8, %r261, %r262, %r257, %r259;
mov.b32 %f61, %r263;
add.ftz.f32 %f62, %f60, %f61;
mov.b32 %r264, %f62;
mov.u32 %r265, 4;
shfl.sync.bfly.b32 %r266|%p9, %r264, %r265, %r257, %r259;
mov.b32 %f63, %r266;
add.ftz.f32 %f64, %f62, %f63;
mov.b32 %r267, %f64;
shfl.sync.bfly.b32 %r268|%p10, %r267, %r255, %r257, %r259;
mov.b32 %f65, %r268;
add.ftz.f32 %f66, %f64, %f65;
mov.b32 %r269, %f66;
mov.u32 %r270, 1;
shfl.sync.bfly.b32 %r271|%p11, %r269, %r270, %r257, %r259;
mov.b32 %f67, %r271;
add.ftz.f32 %f13, %f66, %f67;
setp.lt.u32 %p12, %r4, %r21;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
mul.wide.u32 %rd53, %r252, 4;
add.s64 %rd18, %rd1, %rd53;
not.pred %p14, %p1;
@%p14 bra $L__BB191_8;

st.global.f32 [%rd18], %f13;

$L__BB191_8:
ld.shared.f32 %f68, [%r17+128];
add.ftz.f32 %f69, %f68, %f119;
ld.shared.f32 %f70, [%r17+640];
add.ftz.f32 %f71, %f70, %f69;
ld.shared.f32 %f72, [%r17+1152];
add.ftz.f32 %f73, %f72, %f71;
mov.b32 %r272, %f73;
shfl.sync.bfly.b32 %r276|%p15, %r272, %r258, %r257, %r259;
mov.b32 %f74, %r276;
add.ftz.f32 %f75, %f73, %f74;
mov.b32 %r277, %f75;
shfl.sync.bfly.b32 %r279|%p16, %r277, %r262, %r257, %r259;
mov.b32 %f76, %r279;
add.ftz.f32 %f77, %f75, %f76;
mov.b32 %r280, %f77;
shfl.sync.bfly.b32 %r282|%p17, %r280, %r265, %r257, %r259;
mov.b32 %f78, %r282;
add.ftz.f32 %f79, %f77, %f78;
mov.b32 %r283, %f79;
shfl.sync.bfly.b32 %r285|%p18, %r283, %r255, %r257, %r259;
mov.b32 %f80, %r285;
add.ftz.f32 %f81, %f79, %f80;
mov.b32 %r286, %f81;
shfl.sync.bfly.b32 %r288|%p19, %r286, %r270, %r257, %r259;
mov.b32 %f82, %r288;
add.ftz.f32 %f14, %f81, %f82;
add.s32 %r289, %r4, 1;
setp.lt.u32 %p20, %r289, %r21;
setp.eq.s32 %p21, %r2, 1;
and.pred %p2, %p21, %p20;
not.pred %p22, %p2;
@%p22 bra $L__BB191_10;

st.global.f32 [%rd18+4], %f14;

$L__BB191_10:
ld.shared.f32 %f83, [%r17+256];
add.ftz.f32 %f84, %f83, %f118;
ld.shared.f32 %f85, [%r17+768];
add.ftz.f32 %f86, %f85, %f84;
ld.shared.f32 %f87, [%r17+1280];
add.ftz.f32 %f88, %f87, %f86;
mov.b32 %r290, %f88;
mov.u32 %r291, 31;
mov.u32 %r292, 16;
mov.u32 %r293, -1;
shfl.sync.bfly.b32 %r294|%p23, %r290, %r292, %r291, %r293;
mov.b32 %f89, %r294;
add.ftz.f32 %f90, %f88, %f89;
mov.b32 %r295, %f90;
mov.u32 %r296, 8;
shfl.sync.bfly.b32 %r297|%p24, %r295, %r296, %r291, %r293;
mov.b32 %f91, %r297;
add.ftz.f32 %f92, %f90, %f91;
mov.b32 %r298, %f92;
mov.u32 %r299, 4;
shfl.sync.bfly.b32 %r300|%p25, %r298, %r299, %r291, %r293;
mov.b32 %f93, %r300;
add.ftz.f32 %f94, %f92, %f93;
mov.b32 %r301, %f94;
mov.u32 %r302, 2;
shfl.sync.bfly.b32 %r303|%p26, %r301, %r302, %r291, %r293;
mov.b32 %f95, %r303;
add.ftz.f32 %f96, %f94, %f95;
mov.b32 %r304, %f96;
mov.u32 %r305, 1;
shfl.sync.bfly.b32 %r306|%p27, %r304, %r305, %r291, %r293;
mov.b32 %f97, %r306;
add.ftz.f32 %f15, %f96, %f97;
@%p14 bra $L__BB191_12;

cvt.u64.u32 %rd54, %r21;
add.s64 %rd55, %rd54, %rd17;
shl.b64 %rd56, %rd55, 2;
add.s64 %rd57, %rd1, %rd56;
st.global.f32 [%rd57], %f15;

$L__BB191_12:
ld.shared.f32 %f98, [%r17+384];
add.ftz.f32 %f99, %f98, %f117;
ld.shared.f32 %f100, [%r17+896];
add.ftz.f32 %f101, %f100, %f99;
ld.shared.f32 %f102, [%r17+1408];
add.ftz.f32 %f103, %f102, %f101;
mov.b32 %r307, %f103;
shfl.sync.bfly.b32 %r311|%p29, %r307, %r292, %r291, %r293;
mov.b32 %f104, %r311;
add.ftz.f32 %f105, %f103, %f104;
mov.b32 %r312, %f105;
shfl.sync.bfly.b32 %r314|%p30, %r312, %r296, %r291, %r293;
mov.b32 %f106, %r314;
add.ftz.f32 %f107, %f105, %f106;
mov.b32 %r315, %f107;
shfl.sync.bfly.b32 %r317|%p31, %r315, %r299, %r291, %r293;
mov.b32 %f108, %r317;
add.ftz.f32 %f109, %f107, %f108;
mov.b32 %r318, %f109;
shfl.sync.bfly.b32 %r320|%p32, %r318, %r302, %r291, %r293;
mov.b32 %f110, %r320;
add.ftz.f32 %f111, %f109, %f110;
mov.b32 %r321, %f111;
shfl.sync.bfly.b32 %r323|%p33, %r321, %r305, %r291, %r293;
mov.b32 %f112, %r323;
add.ftz.f32 %f16, %f111, %f112;
@%p22 bra $L__BB191_14;

add.s32 %r324, %r21, 1;
cvt.u64.u32 %rd58, %r324;
add.s64 %rd59, %rd58, %rd17;
shl.b64 %rd60, %rd59, 2;
add.s64 %rd61, %rd1, %rd60;
st.global.f32 [%rd61], %f16;

$L__BB191_14:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type14ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type14ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type14ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type14ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type14ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type14ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type14ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type14ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type14ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<47>;
.reg .b16 %rs<19>;
.reg .f32 %f<181>;
.reg .b32 %r<423>;
.reg .b64 %rd<73>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type14ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[2304];

ld.param.u64 %rd20, [_Z13mul_mat_vec_qIL9ggml_type14ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd21, [_Z13mul_mat_vec_qIL9ggml_type14ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd22, [_Z13mul_mat_vec_qIL9ggml_type14ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type14ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type14ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type14ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type14ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type14ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type14ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type14ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type14ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type14ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type14ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type14ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type14ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type14ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type14ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
cvta.to.global.u64 %rd1, %rd22;
mov.u32 %r1, %tid.y;
shl.b32 %r38, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r38, %r2;
mov.u32 %r39, %ctaid.x;
shl.b32 %r4, %r39, 1;
mov.u32 %r5, %ctaid.y;
shr.s32 %r40, %r3, 31;
shr.u32 %r41, %r40, 27;
add.s32 %r42, %r3, %r41;
shr.s32 %r422, %r42, 5;
shr.u32 %r7, %r37, 8;
setp.ge.s32 %p3, %r422, %r7;
mov.f32 %f175, 0f00000000;
mov.f32 %f176, %f175;
mov.f32 %f177, %f175;
mov.f32 %f178, %f175;
mov.f32 %f179, %f175;
mov.f32 %f180, %f175;
@%p3 bra $L__BB192_3;

mov.u32 %r43, %ctaid.z;
mul.lo.s32 %r44, %r5, %r29;
cvt.u64.u32 %rd23, %r44;
mul.lo.s32 %r45, %r43, %r35;
cvt.u64.u32 %rd24, %r45;
add.s64 %rd2, %rd24, %rd23;
mul.hi.u32 %r46, %r43, %r31;
add.s32 %r47, %r46, %r43;
shr.u32 %r48, %r47, %r32;
mul.lo.s32 %r49, %r48, %r34;
mul.hi.u32 %r50, %r5, %r25;
add.s32 %r51, %r50, %r5;
shr.u32 %r52, %r51, %r26;
mul.lo.s32 %r53, %r52, %r28;
cvta.to.global.u64 %rd72, %rd20;
and.b32 %r57, %r42, -32;
sub.s32 %r58, %r3, %r57;
shr.s32 %r59, %r58, 31;
shr.u32 %r60, %r59, 28;
add.s32 %r61, %r58, %r60;
shr.s32 %r62, %r61, 4;
shl.b32 %r63, %r62, 2;
and.b32 %r64, %r61, -16;
sub.s32 %r65, %r58, %r64;
shr.s32 %r66, %r65, 31;
shr.u32 %r67, %r66, 29;
add.s32 %r68, %r65, %r67;
shr.s32 %r69, %r68, 3;
add.s32 %r70, %r63, %r69;
shl.b32 %r71, %r62, 3;
shr.u32 %r72, %r66, 30;
add.s32 %r73, %r65, %r72;
shr.s32 %r74, %r73, 2;
add.s32 %r75, %r71, %r74;
shl.b32 %r8, %r69, 1;
shl.b32 %r76, %r58, 1;
shr.u32 %r77, %r59, 29;
add.s32 %r78, %r58, %r77;
and.b32 %r79, %r78, -8;
sub.s32 %r80, %r58, %r79;
add.s32 %r81, %r71, %r80;
shl.b32 %r82, %r81, 1;
cvt.s64.s32 %rd25, %r75;
cvt.s64.s32 %rd4, %r80;
add.s32 %r83, %r4, 1;
cvt.s64.s32 %rd5, %r70;
shl.b32 %r84, %r23, 1;
shl.b32 %r419, %r422, 3;
add.s32 %r421, %r419, %r84;
add.s32 %r420, %r23, %r419;
mad.lo.s32 %r85, %r4, %r22, %r422;
add.s32 %r86, %r85, %r49;
add.s32 %r87, %r86, %r53;
mul.wide.s32 %rd6, %r87, 210;
mul.wide.s32 %rd26, %r76, 2;
add.s64 %rd27, %rd6, %rd26;
add.s64 %rd7, %rd27, 2;
add.s32 %r88, %r422, %r49;
add.s32 %r89, %r88, %r53;
mad.lo.s32 %r90, %r83, %r22, %r89;
mul.wide.s32 %rd8, %r90, 210;
add.s64 %rd28, %rd8, %rd25;
add.s64 %rd9, %rd28, 192;
mul.wide.s32 %rd29, %r82, 2;
add.s64 %rd30, %rd6, %rd29;
add.s64 %rd10, %rd30, 130;
add.s64 %rd31, %rd8, %rd29;
add.s64 %rd11, %rd31, 128;
add.s64 %rd12, %rd8, %rd26;
add.s64 %rd32, %rd6, %rd25;
add.s64 %rd13, %rd32, 192;
cvta.to.global.u64 %rd14, %rd21;
mov.f32 %f175, 0f00000000;
shl.b64 %rd40, %rd4, 2;

$L__BB192_2:
cvt.u64.u32 %rd33, %r419;
add.s64 %rd34, %rd2, %rd33;
add.s64 %rd35, %rd72, %rd7;
ld.global.nc.u16 %rs7, [%rd35+-2];
ld.global.nc.u16 %rs8, [%rd35];
mov.b32 %r199, {%rs7, %rs8};
add.s64 %rd36, %rd72, %rd10;
ld.global.nc.u16 %rs9, [%rd36+-2];
ld.global.nc.u16 %rs10, [%rd36];
mov.b32 %r200, {%rs9, %rs10};
shr.s32 %r201, %r200, %r8;
add.s64 %rd37, %rd34, %rd5;
mul.lo.s64 %rd38, %rd37, 36;
add.s64 %rd39, %rd14, %rd38;
add.s64 %rd41, %rd39, %rd40;
ld.global.nc.u32 %r99, [%rd41+4];
ld.global.nc.u32 %r91, [%rd39];

{.reg .f16 low,high;
mov.b32 {low,high},%r91;
cvt.f32.f16 %f37, low;}


ld.global.nc.u32 %r107, [%rd41+76];
ld.global.nc.u32 %r92, [%rd39+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r92;
cvt.f32.f16 %f38, low;}


add.s64 %rd42, %rd72, %rd6;
ld.global.nc.u16 %rs1, [%rd42+208];

{ cvt.f32.f16 %f39, %rs1;}


add.s64 %rd43, %rd72, %rd13;
ld.global.nc.u8 %rs11, [%rd43];
cvt.u32.u16 %r202, %rs11;
cvt.s32.s8 %r203, %r202;
and.b32 %r204, %r199, 252645135;
shl.b32 %r205, %r201, 4;
and.b32 %r206, %r205, 808464432;
or.b32 %r167, %r206, %r204;
or.b32 %r207, %r167, -2139062144;
add.s32 %r208, %r207, -538976288;
xor.b32 %r209, %r208, -2139062144;
xor.b32 %r210, %r167, 538976288;
xor.b32 %r211, %r208, %r167;
and.b32 %r165, %r211, %r210;
mov.u32 %r93, %r165;

prmt.b32 %r93, %r93, 0, 0xba98;

mov.u32 %r95, %r167;

prmt.b32 %r95, %r95, 0, 0xba98;

xor.b32 %r212, %r95, 2139062143;
not.b32 %r213, %r93;
and.b32 %r214, %r209, %r213;
and.b32 %r215, %r212, %r93;
or.b32 %r98, %r215, %r214;
mov.u32 %r198, 0;

dp4a.s32.s32 %r97, %r98, %r99, %r198;

mul.lo.s32 %r216, %r97, %r203;
cvt.rn.f32.s32 %f55, %r216;
fma.rn.ftz.f32 %f56, %f37, %f55, 0f00000000;
ld.global.nc.u8 %rs12, [%rd43+4];
cvt.u32.u16 %r217, %rs12;
cvt.s32.s8 %r218, %r217;
shr.u32 %r219, %r199, 4;
and.b32 %r220, %r219, 252645135;
and.b32 %r221, %r201, 808464432;
or.b32 %r175, %r221, %r220;
or.b32 %r222, %r175, -2139062144;
add.s32 %r223, %r222, -538976288;
xor.b32 %r224, %r223, -2139062144;
xor.b32 %r225, %r175, 538976288;
xor.b32 %r226, %r223, %r175;
and.b32 %r173, %r226, %r225;
mov.u32 %r101, %r173;

prmt.b32 %r101, %r101, 0, 0xba98;

mov.u32 %r103, %r175;

prmt.b32 %r103, %r103, 0, 0xba98;

xor.b32 %r227, %r103, 2139062143;
not.b32 %r228, %r101;
and.b32 %r229, %r224, %r228;
and.b32 %r230, %r227, %r101;
or.b32 %r106, %r230, %r229;

dp4a.s32.s32 %r105, %r106, %r107, %r198;

mul.lo.s32 %r231, %r105, %r218;
cvt.rn.f32.s32 %f57, %r231;
fma.rn.ftz.f32 %f58, %f38, %f57, %f56;
fma.rn.ftz.f32 %f180, %f39, %f58, %f180;
add.s64 %rd44, %rd72, %rd12;
ld.global.nc.u16 %rs13, [%rd44];
ld.global.nc.u16 %rs14, [%rd44+2];
mov.b32 %r232, {%rs13, %rs14};
add.s64 %rd45, %rd72, %rd11;
ld.global.nc.u16 %rs15, [%rd45];
ld.global.nc.u16 %rs16, [%rd45+2];
mov.b32 %r233, {%rs15, %rs16};
shr.s32 %r234, %r233, %r8;

{.reg .f16 low,high;
mov.b32 {low,high},%r91;
cvt.f32.f16 %f40, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r92;
cvt.f32.f16 %f41, low;}


add.s64 %rd46, %rd72, %rd8;
ld.global.nc.u16 %rs2, [%rd46+208];

{ cvt.f32.f16 %f42, %rs2;}


add.s64 %rd47, %rd72, %rd9;
ld.global.nc.u8 %rs17, [%rd47];
cvt.u32.u16 %r235, %rs17;
cvt.s32.s8 %r236, %r235;
and.b32 %r237, %r232, 252645135;
shl.b32 %r238, %r234, 4;
and.b32 %r239, %r238, 808464432;
or.b32 %r185, %r239, %r237;
or.b32 %r240, %r185, -2139062144;
add.s32 %r241, %r240, -538976288;
xor.b32 %r242, %r241, -2139062144;
xor.b32 %r243, %r185, 538976288;
xor.b32 %r244, %r241, %r185;
and.b32 %r183, %r244, %r243;
mov.u32 %r111, %r183;

prmt.b32 %r111, %r111, 0, 0xba98;

mov.u32 %r113, %r185;

prmt.b32 %r113, %r113, 0, 0xba98;

xor.b32 %r245, %r113, 2139062143;
not.b32 %r246, %r111;
and.b32 %r247, %r242, %r246;
and.b32 %r248, %r245, %r111;
or.b32 %r116, %r248, %r247;

dp4a.s32.s32 %r115, %r116, %r99, %r198;

mul.lo.s32 %r249, %r115, %r236;
cvt.rn.f32.s32 %f59, %r249;
fma.rn.ftz.f32 %f60, %f40, %f59, 0f00000000;
ld.global.nc.u8 %rs18, [%rd47+4];
cvt.u32.u16 %r250, %rs18;
cvt.s32.s8 %r251, %r250;
shr.u32 %r252, %r232, 4;
and.b32 %r253, %r252, 252645135;
and.b32 %r254, %r234, 808464432;
or.b32 %r193, %r254, %r253;
or.b32 %r255, %r193, -2139062144;
add.s32 %r256, %r255, -538976288;
xor.b32 %r257, %r256, -2139062144;
xor.b32 %r258, %r193, 538976288;
xor.b32 %r259, %r256, %r193;
and.b32 %r191, %r259, %r258;
mov.u32 %r119, %r191;

prmt.b32 %r119, %r119, 0, 0xba98;

mov.u32 %r121, %r193;

prmt.b32 %r121, %r121, 0, 0xba98;

xor.b32 %r260, %r121, 2139062143;
not.b32 %r261, %r119;
and.b32 %r262, %r257, %r261;
and.b32 %r263, %r260, %r119;
or.b32 %r124, %r263, %r262;

dp4a.s32.s32 %r123, %r124, %r107, %r198;

mul.lo.s32 %r264, %r123, %r251;
cvt.rn.f32.s32 %f61, %r264;
fma.rn.ftz.f32 %f62, %f41, %f61, %f60;
fma.rn.ftz.f32 %f179, %f42, %f62, %f179;
cvt.u64.u32 %rd48, %r420;
add.s64 %rd49, %rd2, %rd48;
add.s64 %rd50, %rd49, %rd5;
mul.lo.s64 %rd51, %rd50, 36;
add.s64 %rd52, %rd14, %rd51;
add.s64 %rd53, %rd52, %rd40;
ld.global.nc.u32 %r135, [%rd53+4];
ld.global.nc.u32 %r127, [%rd52];

{.reg .f16 low,high;
mov.b32 {low,high},%r127;
cvt.f32.f16 %f43, low;}


ld.global.nc.u32 %r143, [%rd53+76];
ld.global.nc.u32 %r128, [%rd52+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r128;
cvt.f32.f16 %f44, low;}



{ cvt.f32.f16 %f45, %rs1;}


mov.u32 %r129, %r165;

prmt.b32 %r129, %r129, 0, 0xba98;

mov.u32 %r131, %r167;

prmt.b32 %r131, %r131, 0, 0xba98;

xor.b32 %r265, %r131, 2139062143;
not.b32 %r266, %r129;
and.b32 %r267, %r209, %r266;
and.b32 %r268, %r265, %r129;
or.b32 %r134, %r268, %r267;

dp4a.s32.s32 %r133, %r134, %r135, %r198;

mul.lo.s32 %r269, %r133, %r203;
cvt.rn.f32.s32 %f63, %r269;
fma.rn.ftz.f32 %f64, %f43, %f63, 0f00000000;
mov.u32 %r137, %r173;

prmt.b32 %r137, %r137, 0, 0xba98;

mov.u32 %r139, %r175;

prmt.b32 %r139, %r139, 0, 0xba98;

xor.b32 %r270, %r139, 2139062143;
not.b32 %r271, %r137;
and.b32 %r272, %r224, %r271;
and.b32 %r273, %r270, %r137;
or.b32 %r142, %r273, %r272;

dp4a.s32.s32 %r141, %r142, %r143, %r198;

mul.lo.s32 %r274, %r141, %r218;
cvt.rn.f32.s32 %f65, %r274;
fma.rn.ftz.f32 %f66, %f44, %f65, %f64;
fma.rn.ftz.f32 %f178, %f45, %f66, %f178;

{.reg .f16 low,high;
mov.b32 {low,high},%r127;
cvt.f32.f16 %f46, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r128;
cvt.f32.f16 %f47, low;}



{ cvt.f32.f16 %f48, %rs2;}


mov.u32 %r147, %r183;

prmt.b32 %r147, %r147, 0, 0xba98;

mov.u32 %r149, %r185;

prmt.b32 %r149, %r149, 0, 0xba98;

xor.b32 %r275, %r149, 2139062143;
not.b32 %r276, %r147;
and.b32 %r277, %r242, %r276;
and.b32 %r278, %r275, %r147;
or.b32 %r152, %r278, %r277;

dp4a.s32.s32 %r151, %r152, %r135, %r198;

mul.lo.s32 %r279, %r151, %r236;
cvt.rn.f32.s32 %f67, %r279;
fma.rn.ftz.f32 %f68, %f46, %f67, 0f00000000;
mov.u32 %r155, %r191;

prmt.b32 %r155, %r155, 0, 0xba98;

mov.u32 %r157, %r193;

prmt.b32 %r157, %r157, 0, 0xba98;

xor.b32 %r280, %r157, 2139062143;
not.b32 %r281, %r155;
and.b32 %r282, %r257, %r281;
and.b32 %r283, %r280, %r155;
or.b32 %r160, %r283, %r282;

dp4a.s32.s32 %r159, %r160, %r143, %r198;

mul.lo.s32 %r284, %r159, %r251;
cvt.rn.f32.s32 %f69, %r284;
fma.rn.ftz.f32 %f70, %f47, %f69, %f68;
fma.rn.ftz.f32 %f177, %f48, %f70, %f177;
cvt.u64.u32 %rd54, %r421;
add.s64 %rd55, %rd2, %rd54;
add.s64 %rd56, %rd55, %rd5;
mul.lo.s64 %rd57, %rd56, 36;
add.s64 %rd58, %rd14, %rd57;
add.s64 %rd59, %rd58, %rd40;
ld.global.nc.u32 %r171, [%rd59+4];
ld.global.nc.u32 %r163, [%rd58];

{.reg .f16 low,high;
mov.b32 {low,high},%r163;
cvt.f32.f16 %f49, low;}


ld.global.nc.u32 %r179, [%rd59+76];
ld.global.nc.u32 %r164, [%rd58+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r164;
cvt.f32.f16 %f50, low;}



{ cvt.f32.f16 %f51, %rs1;}



prmt.b32 %r165, %r165, 0, 0xba98;


prmt.b32 %r167, %r167, 0, 0xba98;

xor.b32 %r285, %r167, 2139062143;
not.b32 %r286, %r165;
and.b32 %r287, %r209, %r286;
and.b32 %r288, %r285, %r165;
or.b32 %r170, %r288, %r287;

dp4a.s32.s32 %r169, %r170, %r171, %r198;

mul.lo.s32 %r289, %r169, %r203;
cvt.rn.f32.s32 %f71, %r289;
fma.rn.ftz.f32 %f72, %f49, %f71, 0f00000000;

prmt.b32 %r173, %r173, 0, 0xba98;


prmt.b32 %r175, %r175, 0, 0xba98;

xor.b32 %r290, %r175, 2139062143;
not.b32 %r291, %r173;
and.b32 %r292, %r224, %r291;
and.b32 %r293, %r290, %r173;
or.b32 %r178, %r293, %r292;

dp4a.s32.s32 %r177, %r178, %r179, %r198;

mul.lo.s32 %r294, %r177, %r218;
cvt.rn.f32.s32 %f73, %r294;
fma.rn.ftz.f32 %f74, %f50, %f73, %f72;
fma.rn.ftz.f32 %f176, %f51, %f74, %f176;

{.reg .f16 low,high;
mov.b32 {low,high},%r163;
cvt.f32.f16 %f52, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r164;
cvt.f32.f16 %f53, low;}



{ cvt.f32.f16 %f54, %rs2;}



prmt.b32 %r183, %r183, 0, 0xba98;


prmt.b32 %r185, %r185, 0, 0xba98;

xor.b32 %r295, %r185, 2139062143;
not.b32 %r296, %r183;
and.b32 %r297, %r242, %r296;
and.b32 %r298, %r295, %r183;
or.b32 %r188, %r298, %r297;

dp4a.s32.s32 %r187, %r188, %r171, %r198;

mul.lo.s32 %r299, %r187, %r236;
cvt.rn.f32.s32 %f75, %r299;
fma.rn.ftz.f32 %f76, %f52, %f75, 0f00000000;

prmt.b32 %r191, %r191, 0, 0xba98;


prmt.b32 %r193, %r193, 0, 0xba98;

xor.b32 %r300, %r193, 2139062143;
not.b32 %r301, %r191;
and.b32 %r302, %r257, %r301;
and.b32 %r303, %r300, %r191;
or.b32 %r196, %r303, %r302;

dp4a.s32.s32 %r195, %r196, %r179, %r198;

mul.lo.s32 %r304, %r195, %r251;
cvt.rn.f32.s32 %f77, %r304;
fma.rn.ftz.f32 %f78, %f53, %f77, %f76;
fma.rn.ftz.f32 %f175, %f54, %f78, %f175;
add.s32 %r421, %r421, 32;
add.s32 %r420, %r420, 32;
add.s32 %r419, %r419, 32;
add.s64 %rd72, %rd72, 840;
add.s32 %r422, %r422, 4;
setp.lt.s32 %p4, %r422, %r7;
@%p4 bra $L__BB192_2;

$L__BB192_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB192_5;

mov.u32 %r305, _ZZ13mul_mat_vec_qIL9ggml_type14ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r306, %r1, 768, %r305;
shl.b32 %r307, %r2, 2;
add.s32 %r308, %r306, %r307;
st.shared.f32 [%r308+-768], %f180;
st.shared.f32 [%r308+-640], %f179;
st.shared.f32 [%r308+-512], %f178;
st.shared.f32 [%r308+-384], %f177;
st.shared.f32 [%r308+-256], %f176;
st.shared.f32 [%r308+-128], %f175;

$L__BB192_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB192_18;

mov.u32 %r309, %ctaid.z;
mad.lo.s32 %r310, %r5, %r30, %r4;
mad.lo.s32 %r311, %r309, %r36, %r310;
cvt.u64.u32 %rd17, %r311;
shl.b32 %r312, %r2, 2;
mov.u32 %r313, _ZZ13mul_mat_vec_qIL9ggml_type14ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r20, %r313, %r312;
mov.u32 %r314, 2;
ld.shared.f32 %f79, [%r20];
add.ftz.f32 %f80, %f79, %f180;
ld.shared.f32 %f81, [%r20+768];
add.ftz.f32 %f82, %f81, %f80;
ld.shared.f32 %f83, [%r20+1536];
add.ftz.f32 %f84, %f83, %f82;
mov.b32 %r315, %f84;
mov.u32 %r316, 31;
mov.u32 %r317, 16;
mov.u32 %r318, -1;
shfl.sync.bfly.b32 %r319|%p7, %r315, %r317, %r316, %r318;
mov.b32 %f85, %r319;
add.ftz.f32 %f86, %f84, %f85;
mov.b32 %r320, %f86;
mov.u32 %r321, 8;
shfl.sync.bfly.b32 %r322|%p8, %r320, %r321, %r316, %r318;
mov.b32 %f87, %r322;
add.ftz.f32 %f88, %f86, %f87;
mov.b32 %r323, %f88;
mov.u32 %r324, 4;
shfl.sync.bfly.b32 %r325|%p9, %r323, %r324, %r316, %r318;
mov.b32 %f89, %r325;
add.ftz.f32 %f90, %f88, %f89;
mov.b32 %r326, %f90;
shfl.sync.bfly.b32 %r327|%p10, %r326, %r314, %r316, %r318;
mov.b32 %f91, %r327;
add.ftz.f32 %f92, %f90, %f91;
mov.b32 %r328, %f92;
mov.u32 %r329, 1;
shfl.sync.bfly.b32 %r330|%p11, %r328, %r329, %r316, %r318;
mov.b32 %f93, %r330;
add.ftz.f32 %f19, %f92, %f93;
setp.lt.u32 %p12, %r4, %r24;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
mul.wide.u32 %rd60, %r311, 4;
add.s64 %rd18, %rd1, %rd60;
not.pred %p14, %p1;
@%p14 bra $L__BB192_8;

st.global.f32 [%rd18], %f19;

$L__BB192_8:
ld.shared.f32 %f94, [%r20+128];
add.ftz.f32 %f95, %f94, %f179;
ld.shared.f32 %f96, [%r20+896];
add.ftz.f32 %f97, %f96, %f95;
ld.shared.f32 %f98, [%r20+1664];
add.ftz.f32 %f99, %f98, %f97;
mov.b32 %r331, %f99;
shfl.sync.bfly.b32 %r335|%p15, %r331, %r317, %r316, %r318;
mov.b32 %f100, %r335;
add.ftz.f32 %f101, %f99, %f100;
mov.b32 %r336, %f101;
shfl.sync.bfly.b32 %r338|%p16, %r336, %r321, %r316, %r318;
mov.b32 %f102, %r338;
add.ftz.f32 %f103, %f101, %f102;
mov.b32 %r339, %f103;
shfl.sync.bfly.b32 %r341|%p17, %r339, %r324, %r316, %r318;
mov.b32 %f104, %r341;
add.ftz.f32 %f105, %f103, %f104;
mov.b32 %r342, %f105;
shfl.sync.bfly.b32 %r344|%p18, %r342, %r314, %r316, %r318;
mov.b32 %f106, %r344;
add.ftz.f32 %f107, %f105, %f106;
mov.b32 %r345, %f107;
shfl.sync.bfly.b32 %r347|%p19, %r345, %r329, %r316, %r318;
mov.b32 %f108, %r347;
add.ftz.f32 %f20, %f107, %f108;
add.s32 %r348, %r4, 1;
setp.lt.u32 %p20, %r348, %r24;
setp.eq.s32 %p21, %r2, 1;
and.pred %p2, %p21, %p20;
not.pred %p22, %p2;
@%p22 bra $L__BB192_10;

st.global.f32 [%rd18+4], %f20;

$L__BB192_10:
ld.shared.f32 %f109, [%r20+256];
add.ftz.f32 %f110, %f109, %f178;
ld.shared.f32 %f111, [%r20+1024];
add.ftz.f32 %f112, %f111, %f110;
ld.shared.f32 %f113, [%r20+1792];
add.ftz.f32 %f114, %f113, %f112;
mov.b32 %r349, %f114;
mov.u32 %r350, 31;
mov.u32 %r351, 16;
mov.u32 %r352, -1;
shfl.sync.bfly.b32 %r353|%p23, %r349, %r351, %r350, %r352;
mov.b32 %f115, %r353;
add.ftz.f32 %f116, %f114, %f115;
mov.b32 %r354, %f116;
mov.u32 %r355, 8;
shfl.sync.bfly.b32 %r356|%p24, %r354, %r355, %r350, %r352;
mov.b32 %f117, %r356;
add.ftz.f32 %f118, %f116, %f117;
mov.b32 %r357, %f118;
mov.u32 %r358, 4;
shfl.sync.bfly.b32 %r359|%p25, %r357, %r358, %r350, %r352;
mov.b32 %f119, %r359;
add.ftz.f32 %f120, %f118, %f119;
mov.b32 %r360, %f120;
mov.u32 %r361, 2;
shfl.sync.bfly.b32 %r362|%p26, %r360, %r361, %r350, %r352;
mov.b32 %f121, %r362;
add.ftz.f32 %f122, %f120, %f121;
mov.b32 %r363, %f122;
mov.u32 %r364, 1;
shfl.sync.bfly.b32 %r365|%p27, %r363, %r364, %r350, %r352;
mov.b32 %f123, %r365;
add.ftz.f32 %f21, %f122, %f123;
@%p14 bra $L__BB192_12;

cvt.u64.u32 %rd61, %r24;
add.s64 %rd62, %rd61, %rd17;
shl.b64 %rd63, %rd62, 2;
add.s64 %rd64, %rd1, %rd63;
st.global.f32 [%rd64], %f21;

$L__BB192_12:
ld.shared.f32 %f124, [%r20+384];
add.ftz.f32 %f125, %f124, %f177;
ld.shared.f32 %f126, [%r20+1152];
add.ftz.f32 %f127, %f126, %f125;
ld.shared.f32 %f128, [%r20+1920];
add.ftz.f32 %f129, %f128, %f127;
mov.b32 %r366, %f129;
shfl.sync.bfly.b32 %r370|%p29, %r366, %r351, %r350, %r352;
mov.b32 %f130, %r370;
add.ftz.f32 %f131, %f129, %f130;
mov.b32 %r371, %f131;
shfl.sync.bfly.b32 %r373|%p30, %r371, %r355, %r350, %r352;
mov.b32 %f132, %r373;
add.ftz.f32 %f133, %f131, %f132;
mov.b32 %r374, %f133;
shfl.sync.bfly.b32 %r376|%p31, %r374, %r358, %r350, %r352;
mov.b32 %f134, %r376;
add.ftz.f32 %f135, %f133, %f134;
mov.b32 %r377, %f135;
shfl.sync.bfly.b32 %r379|%p32, %r377, %r361, %r350, %r352;
mov.b32 %f136, %r379;
add.ftz.f32 %f137, %f135, %f136;
mov.b32 %r380, %f137;
shfl.sync.bfly.b32 %r382|%p33, %r380, %r364, %r350, %r352;
mov.b32 %f138, %r382;
add.ftz.f32 %f22, %f137, %f138;
@%p22 bra $L__BB192_14;

add.s32 %r383, %r24, 1;
cvt.u64.u32 %rd65, %r383;
add.s64 %rd66, %rd65, %rd17;
shl.b64 %rd67, %rd66, 2;
add.s64 %rd68, %rd1, %rd67;
st.global.f32 [%rd68], %f22;

$L__BB192_14:
shl.b32 %r384, %r24, 1;
mov.u32 %r385, 1;
ld.shared.f32 %f139, [%r20+512];
add.ftz.f32 %f140, %f139, %f176;
ld.shared.f32 %f141, [%r20+1280];
add.ftz.f32 %f142, %f141, %f140;
ld.shared.f32 %f143, [%r20+2048];
add.ftz.f32 %f144, %f143, %f142;
mov.b32 %r386, %f144;
mov.u32 %r387, 31;
mov.u32 %r388, 16;
mov.u32 %r389, -1;
shfl.sync.bfly.b32 %r390|%p35, %r386, %r388, %r387, %r389;
mov.b32 %f145, %r390;
add.ftz.f32 %f146, %f144, %f145;
mov.b32 %r391, %f146;
mov.u32 %r392, 8;
shfl.sync.bfly.b32 %r393|%p36, %r391, %r392, %r387, %r389;
mov.b32 %f147, %r393;
add.ftz.f32 %f148, %f146, %f147;
mov.b32 %r394, %f148;
mov.u32 %r395, 4;
shfl.sync.bfly.b32 %r396|%p37, %r394, %r395, %r387, %r389;
mov.b32 %f149, %r396;
add.ftz.f32 %f150, %f148, %f149;
mov.b32 %r397, %f150;
mov.u32 %r398, 2;
shfl.sync.bfly.b32 %r399|%p38, %r397, %r398, %r387, %r389;
mov.b32 %f151, %r399;
add.ftz.f32 %f152, %f150, %f151;
mov.b32 %r400, %f152;
shfl.sync.bfly.b32 %r401|%p39, %r400, %r385, %r387, %r389;
mov.b32 %f153, %r401;
add.ftz.f32 %f23, %f152, %f153;
cvt.u64.u32 %rd69, %r384;
add.s64 %rd70, %rd69, %rd17;
shl.b64 %rd71, %rd70, 2;
add.s64 %rd19, %rd1, %rd71;
@%p14 bra $L__BB192_16;

st.global.f32 [%rd19], %f23;

$L__BB192_16:
ld.shared.f32 %f154, [%r20+640];
add.ftz.f32 %f155, %f154, %f175;
ld.shared.f32 %f156, [%r20+1408];
add.ftz.f32 %f157, %f156, %f155;
ld.shared.f32 %f158, [%r20+2176];
add.ftz.f32 %f159, %f158, %f157;
mov.b32 %r402, %f159;
shfl.sync.bfly.b32 %r406|%p41, %r402, %r388, %r387, %r389;
mov.b32 %f160, %r406;
add.ftz.f32 %f161, %f159, %f160;
mov.b32 %r407, %f161;
shfl.sync.bfly.b32 %r409|%p42, %r407, %r392, %r387, %r389;
mov.b32 %f162, %r409;
add.ftz.f32 %f163, %f161, %f162;
mov.b32 %r410, %f163;
shfl.sync.bfly.b32 %r412|%p43, %r410, %r395, %r387, %r389;
mov.b32 %f164, %r412;
add.ftz.f32 %f165, %f163, %f164;
mov.b32 %r413, %f165;
shfl.sync.bfly.b32 %r415|%p44, %r413, %r398, %r387, %r389;
mov.b32 %f166, %r415;
add.ftz.f32 %f167, %f165, %f166;
mov.b32 %r416, %f167;
shfl.sync.bfly.b32 %r418|%p45, %r416, %r385, %r387, %r389;
mov.b32 %f168, %r418;
add.ftz.f32 %f24, %f167, %f168;
@%p22 bra $L__BB192_18;

st.global.f32 [%rd19+4], %f24;

$L__BB192_18:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type14ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type14ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type14ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type14ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type14ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type14ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type14ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type14ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type14ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<59>;
.reg .b16 %rs<21>;
.reg .f32 %f<241>;
.reg .b32 %r<520>;
.reg .b64 %rd<87>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type14ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[3072];

ld.param.u64 %rd20, [_Z13mul_mat_vec_qIL9ggml_type14ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd21, [_Z13mul_mat_vec_qIL9ggml_type14ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd22, [_Z13mul_mat_vec_qIL9ggml_type14ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type14ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type14ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type14ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type14ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type14ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type14ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type14ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type14ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type14ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type14ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type14ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type14ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type14ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type14ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
cvta.to.global.u64 %rd1, %rd22;
mov.u32 %r1, %tid.y;
shl.b32 %r42, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r42, %r2;
mov.u32 %r43, %ctaid.x;
shl.b32 %r4, %r43, 1;
mov.u32 %r5, %ctaid.y;
shr.s32 %r44, %r3, 31;
shr.u32 %r45, %r44, 27;
add.s32 %r46, %r3, %r45;
shr.s32 %r519, %r46, 5;
shr.u32 %r7, %r41, 8;
setp.ge.s32 %p3, %r519, %r7;
mov.f32 %f233, 0f00000000;
mov.f32 %f234, %f233;
mov.f32 %f235, %f233;
mov.f32 %f236, %f233;
mov.f32 %f237, %f233;
mov.f32 %f238, %f233;
mov.f32 %f239, %f233;
mov.f32 %f240, %f233;
@%p3 bra $L__BB193_3;

mov.u32 %r47, %ctaid.z;
mul.lo.s32 %r48, %r5, %r33;
cvt.u64.u32 %rd23, %r48;
mul.lo.s32 %r49, %r47, %r39;
cvt.u64.u32 %rd24, %r49;
add.s64 %rd2, %rd24, %rd23;
mul.hi.u32 %r50, %r47, %r35;
add.s32 %r51, %r50, %r47;
shr.u32 %r52, %r51, %r36;
mul.lo.s32 %r53, %r52, %r38;
mul.hi.u32 %r54, %r5, %r29;
add.s32 %r55, %r54, %r5;
shr.u32 %r56, %r55, %r30;
mul.lo.s32 %r57, %r56, %r32;
cvta.to.global.u64 %rd86, %rd20;
add.s32 %r58, %r4, 1;
and.b32 %r62, %r46, -32;
sub.s32 %r63, %r3, %r62;
shr.s32 %r64, %r63, 31;
shr.u32 %r65, %r64, 28;
add.s32 %r66, %r63, %r65;
shr.s32 %r67, %r66, 4;
shl.b32 %r68, %r67, 3;
and.b32 %r69, %r66, -16;
sub.s32 %r70, %r63, %r69;
shr.s32 %r71, %r70, 31;
shr.u32 %r72, %r71, 30;
add.s32 %r73, %r70, %r72;
shr.s32 %r74, %r73, 2;
add.s32 %r75, %r68, %r74;
cvt.s64.s32 %rd25, %r75;
shl.b32 %r76, %r67, 2;
shr.u32 %r77, %r71, 29;
add.s32 %r78, %r70, %r77;
shr.s32 %r79, %r78, 3;
add.s32 %r80, %r76, %r79;
shl.b32 %r8, %r79, 1;
shl.b32 %r81, %r63, 1;
shr.u32 %r82, %r64, 29;
add.s32 %r83, %r63, %r82;
and.b32 %r84, %r83, -8;
sub.s32 %r85, %r63, %r84;
add.s32 %r86, %r68, %r85;
shl.b32 %r87, %r86, 1;
cvt.s64.s32 %rd4, %r85;
cvt.s64.s32 %rd5, %r80;
shl.b32 %r88, %r27, 1;
shl.b32 %r515, %r519, 3;
add.s32 %r518, %r515, %r88;
mad.lo.s32 %r517, %r27, 3, %r515;
add.s32 %r516, %r27, %r515;
mad.lo.s32 %r89, %r4, %r26, %r519;
add.s32 %r90, %r89, %r53;
add.s32 %r91, %r90, %r57;
mul.wide.s32 %rd6, %r91, 210;
mul.wide.s32 %rd26, %r81, 2;
add.s64 %rd27, %rd6, %rd26;
add.s64 %rd7, %rd27, 2;
add.s32 %r92, %r519, %r53;
add.s32 %r93, %r92, %r57;
mad.lo.s32 %r94, %r58, %r26, %r93;
mul.wide.s32 %rd8, %r94, 210;
add.s64 %rd28, %rd8, %rd25;
add.s64 %rd9, %rd28, 192;
mul.wide.s32 %rd29, %r87, 2;
add.s64 %rd30, %rd6, %rd29;
add.s64 %rd10, %rd30, 130;
add.s64 %rd31, %rd8, %rd29;
add.s64 %rd11, %rd31, 128;
add.s64 %rd12, %rd8, %rd26;
add.s64 %rd32, %rd6, %rd25;
add.s64 %rd13, %rd32, 192;
cvta.to.global.u64 %rd14, %rd21;
mov.f32 %f233, 0f00000000;
shl.b64 %rd40, %rd4, 2;

$L__BB193_2:
cvt.u64.u32 %rd33, %r515;
add.s64 %rd34, %rd2, %rd33;
add.s64 %rd35, %rd86, %rd7;
ld.global.nc.u16 %rs9, [%rd35+-2];
ld.global.nc.u16 %rs10, [%rd35];
mov.b32 %r239, {%rs9, %rs10};
add.s64 %rd36, %rd86, %rd10;
ld.global.nc.u16 %rs11, [%rd36+-2];
ld.global.nc.u16 %rs12, [%rd36];
mov.b32 %r240, {%rs11, %rs12};
shr.s32 %r241, %r240, %r8;
add.s64 %rd37, %rd34, %rd5;
mul.lo.s64 %rd38, %rd37, 36;
add.s64 %rd39, %rd14, %rd38;
add.s64 %rd41, %rd39, %rd40;
ld.global.nc.u32 %r103, [%rd41+4];
ld.global.nc.u32 %r95, [%rd39];

{.reg .f16 low,high;
mov.b32 {low,high},%r95;
cvt.f32.f16 %f49, low;}


ld.global.nc.u32 %r111, [%rd41+76];
ld.global.nc.u32 %r96, [%rd39+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r96;
cvt.f32.f16 %f50, low;}


add.s64 %rd42, %rd86, %rd6;
ld.global.nc.u16 %rs1, [%rd42+208];

{ cvt.f32.f16 %f51, %rs1;}


add.s64 %rd43, %rd86, %rd13;
ld.global.nc.u8 %rs13, [%rd43];
cvt.u32.u16 %r242, %rs13;
cvt.s32.s8 %r243, %r242;
and.b32 %r244, %r239, 252645135;
shl.b32 %r245, %r241, 4;
and.b32 %r246, %r245, 808464432;
or.b32 %r207, %r246, %r244;
or.b32 %r247, %r207, -2139062144;
add.s32 %r248, %r247, -538976288;
xor.b32 %r249, %r248, -2139062144;
xor.b32 %r250, %r207, 538976288;
xor.b32 %r251, %r248, %r207;
and.b32 %r205, %r251, %r250;
mov.u32 %r97, %r205;

prmt.b32 %r97, %r97, 0, 0xba98;

mov.u32 %r99, %r207;

prmt.b32 %r99, %r99, 0, 0xba98;

xor.b32 %r252, %r99, 2139062143;
not.b32 %r253, %r97;
and.b32 %r254, %r249, %r253;
and.b32 %r255, %r252, %r97;
or.b32 %r102, %r255, %r254;
mov.u32 %r238, 0;

dp4a.s32.s32 %r101, %r102, %r103, %r238;

mul.lo.s32 %r256, %r101, %r243;
cvt.rn.f32.s32 %f73, %r256;
fma.rn.ftz.f32 %f74, %f49, %f73, 0f00000000;
ld.global.nc.u8 %rs14, [%rd43+4];
cvt.u32.u16 %r257, %rs14;
cvt.s32.s8 %r258, %r257;
shr.u32 %r259, %r239, 4;
and.b32 %r260, %r259, 252645135;
and.b32 %r261, %r241, 808464432;
or.b32 %r215, %r261, %r260;
or.b32 %r262, %r215, -2139062144;
add.s32 %r263, %r262, -538976288;
xor.b32 %r264, %r263, -2139062144;
xor.b32 %r265, %r215, 538976288;
xor.b32 %r266, %r263, %r215;
and.b32 %r213, %r266, %r265;
mov.u32 %r105, %r213;

prmt.b32 %r105, %r105, 0, 0xba98;

mov.u32 %r107, %r215;

prmt.b32 %r107, %r107, 0, 0xba98;

xor.b32 %r267, %r107, 2139062143;
not.b32 %r268, %r105;
and.b32 %r269, %r264, %r268;
and.b32 %r270, %r267, %r105;
or.b32 %r110, %r270, %r269;

dp4a.s32.s32 %r109, %r110, %r111, %r238;

mul.lo.s32 %r271, %r109, %r258;
cvt.rn.f32.s32 %f75, %r271;
fma.rn.ftz.f32 %f76, %f50, %f75, %f74;
fma.rn.ftz.f32 %f240, %f51, %f76, %f240;
add.s64 %rd44, %rd86, %rd12;
ld.global.nc.u16 %rs15, [%rd44];
ld.global.nc.u16 %rs16, [%rd44+2];
mov.b32 %r272, {%rs15, %rs16};
add.s64 %rd45, %rd86, %rd11;
ld.global.nc.u16 %rs17, [%rd45];
ld.global.nc.u16 %rs18, [%rd45+2];
mov.b32 %r273, {%rs17, %rs18};
shr.s32 %r274, %r273, %r8;

{.reg .f16 low,high;
mov.b32 {low,high},%r95;
cvt.f32.f16 %f52, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r96;
cvt.f32.f16 %f53, low;}


add.s64 %rd46, %rd86, %rd8;
ld.global.nc.u16 %rs2, [%rd46+208];

{ cvt.f32.f16 %f54, %rs2;}


add.s64 %rd47, %rd86, %rd9;
ld.global.nc.u8 %rs19, [%rd47];
cvt.u32.u16 %r275, %rs19;
cvt.s32.s8 %r276, %r275;
and.b32 %r277, %r272, 252645135;
shl.b32 %r278, %r274, 4;
and.b32 %r279, %r278, 808464432;
or.b32 %r225, %r279, %r277;
or.b32 %r280, %r225, -2139062144;
add.s32 %r281, %r280, -538976288;
xor.b32 %r282, %r281, -2139062144;
xor.b32 %r283, %r225, 538976288;
xor.b32 %r284, %r281, %r225;
and.b32 %r223, %r284, %r283;
mov.u32 %r115, %r223;

prmt.b32 %r115, %r115, 0, 0xba98;

mov.u32 %r117, %r225;

prmt.b32 %r117, %r117, 0, 0xba98;

xor.b32 %r285, %r117, 2139062143;
not.b32 %r286, %r115;
and.b32 %r287, %r282, %r286;
and.b32 %r288, %r285, %r115;
or.b32 %r120, %r288, %r287;

dp4a.s32.s32 %r119, %r120, %r103, %r238;

mul.lo.s32 %r289, %r119, %r276;
cvt.rn.f32.s32 %f77, %r289;
fma.rn.ftz.f32 %f78, %f52, %f77, 0f00000000;
ld.global.nc.u8 %rs20, [%rd47+4];
cvt.u32.u16 %r290, %rs20;
cvt.s32.s8 %r291, %r290;
shr.u32 %r292, %r272, 4;
and.b32 %r293, %r292, 252645135;
and.b32 %r294, %r274, 808464432;
or.b32 %r233, %r294, %r293;
or.b32 %r295, %r233, -2139062144;
add.s32 %r296, %r295, -538976288;
xor.b32 %r297, %r296, -2139062144;
xor.b32 %r298, %r233, 538976288;
xor.b32 %r299, %r296, %r233;
and.b32 %r231, %r299, %r298;
mov.u32 %r123, %r231;

prmt.b32 %r123, %r123, 0, 0xba98;

mov.u32 %r125, %r233;

prmt.b32 %r125, %r125, 0, 0xba98;

xor.b32 %r300, %r125, 2139062143;
not.b32 %r301, %r123;
and.b32 %r302, %r297, %r301;
and.b32 %r303, %r300, %r123;
or.b32 %r128, %r303, %r302;

dp4a.s32.s32 %r127, %r128, %r111, %r238;

mul.lo.s32 %r304, %r127, %r291;
cvt.rn.f32.s32 %f79, %r304;
fma.rn.ftz.f32 %f80, %f53, %f79, %f78;
fma.rn.ftz.f32 %f239, %f54, %f80, %f239;
cvt.u64.u32 %rd48, %r516;
add.s64 %rd49, %rd2, %rd48;
add.s64 %rd50, %rd49, %rd5;
mul.lo.s64 %rd51, %rd50, 36;
add.s64 %rd52, %rd14, %rd51;
add.s64 %rd53, %rd52, %rd40;
ld.global.nc.u32 %r139, [%rd53+4];
ld.global.nc.u32 %r131, [%rd52];

{.reg .f16 low,high;
mov.b32 {low,high},%r131;
cvt.f32.f16 %f55, low;}


ld.global.nc.u32 %r147, [%rd53+76];
ld.global.nc.u32 %r132, [%rd52+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r132;
cvt.f32.f16 %f56, low;}



{ cvt.f32.f16 %f57, %rs1;}


mov.u32 %r133, %r205;

prmt.b32 %r133, %r133, 0, 0xba98;

mov.u32 %r135, %r207;

prmt.b32 %r135, %r135, 0, 0xba98;

xor.b32 %r305, %r135, 2139062143;
not.b32 %r306, %r133;
and.b32 %r307, %r249, %r306;
and.b32 %r308, %r305, %r133;
or.b32 %r138, %r308, %r307;

dp4a.s32.s32 %r137, %r138, %r139, %r238;

mul.lo.s32 %r309, %r137, %r243;
cvt.rn.f32.s32 %f81, %r309;
fma.rn.ftz.f32 %f82, %f55, %f81, 0f00000000;
mov.u32 %r141, %r213;

prmt.b32 %r141, %r141, 0, 0xba98;

mov.u32 %r143, %r215;

prmt.b32 %r143, %r143, 0, 0xba98;

xor.b32 %r310, %r143, 2139062143;
not.b32 %r311, %r141;
and.b32 %r312, %r264, %r311;
and.b32 %r313, %r310, %r141;
or.b32 %r146, %r313, %r312;

dp4a.s32.s32 %r145, %r146, %r147, %r238;

mul.lo.s32 %r314, %r145, %r258;
cvt.rn.f32.s32 %f83, %r314;
fma.rn.ftz.f32 %f84, %f56, %f83, %f82;
fma.rn.ftz.f32 %f238, %f57, %f84, %f238;

{.reg .f16 low,high;
mov.b32 {low,high},%r131;
cvt.f32.f16 %f58, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r132;
cvt.f32.f16 %f59, low;}



{ cvt.f32.f16 %f60, %rs2;}


mov.u32 %r151, %r223;

prmt.b32 %r151, %r151, 0, 0xba98;

mov.u32 %r153, %r225;

prmt.b32 %r153, %r153, 0, 0xba98;

xor.b32 %r315, %r153, 2139062143;
not.b32 %r316, %r151;
and.b32 %r317, %r282, %r316;
and.b32 %r318, %r315, %r151;
or.b32 %r156, %r318, %r317;

dp4a.s32.s32 %r155, %r156, %r139, %r238;

mul.lo.s32 %r319, %r155, %r276;
cvt.rn.f32.s32 %f85, %r319;
fma.rn.ftz.f32 %f86, %f58, %f85, 0f00000000;
mov.u32 %r159, %r231;

prmt.b32 %r159, %r159, 0, 0xba98;

mov.u32 %r161, %r233;

prmt.b32 %r161, %r161, 0, 0xba98;

xor.b32 %r320, %r161, 2139062143;
not.b32 %r321, %r159;
and.b32 %r322, %r297, %r321;
and.b32 %r323, %r320, %r159;
or.b32 %r164, %r323, %r322;

dp4a.s32.s32 %r163, %r164, %r147, %r238;

mul.lo.s32 %r324, %r163, %r291;
cvt.rn.f32.s32 %f87, %r324;
fma.rn.ftz.f32 %f88, %f59, %f87, %f86;
fma.rn.ftz.f32 %f237, %f60, %f88, %f237;
cvt.u64.u32 %rd54, %r518;
add.s64 %rd55, %rd2, %rd54;
add.s64 %rd56, %rd55, %rd5;
mul.lo.s64 %rd57, %rd56, 36;
add.s64 %rd58, %rd14, %rd57;
add.s64 %rd59, %rd58, %rd40;
ld.global.nc.u32 %r175, [%rd59+4];
ld.global.nc.u32 %r167, [%rd58];

{.reg .f16 low,high;
mov.b32 {low,high},%r167;
cvt.f32.f16 %f61, low;}


ld.global.nc.u32 %r183, [%rd59+76];
ld.global.nc.u32 %r168, [%rd58+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r168;
cvt.f32.f16 %f62, low;}



{ cvt.f32.f16 %f63, %rs1;}


mov.u32 %r169, %r205;

prmt.b32 %r169, %r169, 0, 0xba98;

mov.u32 %r171, %r207;

prmt.b32 %r171, %r171, 0, 0xba98;

xor.b32 %r325, %r171, 2139062143;
not.b32 %r326, %r169;
and.b32 %r327, %r249, %r326;
and.b32 %r328, %r325, %r169;
or.b32 %r174, %r328, %r327;

dp4a.s32.s32 %r173, %r174, %r175, %r238;

mul.lo.s32 %r329, %r173, %r243;
cvt.rn.f32.s32 %f89, %r329;
fma.rn.ftz.f32 %f90, %f61, %f89, 0f00000000;
mov.u32 %r177, %r213;

prmt.b32 %r177, %r177, 0, 0xba98;

mov.u32 %r179, %r215;

prmt.b32 %r179, %r179, 0, 0xba98;

xor.b32 %r330, %r179, 2139062143;
not.b32 %r331, %r177;
and.b32 %r332, %r264, %r331;
and.b32 %r333, %r330, %r177;
or.b32 %r182, %r333, %r332;

dp4a.s32.s32 %r181, %r182, %r183, %r238;

mul.lo.s32 %r334, %r181, %r258;
cvt.rn.f32.s32 %f91, %r334;
fma.rn.ftz.f32 %f92, %f62, %f91, %f90;
fma.rn.ftz.f32 %f236, %f63, %f92, %f236;

{.reg .f16 low,high;
mov.b32 {low,high},%r167;
cvt.f32.f16 %f64, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r168;
cvt.f32.f16 %f65, low;}



{ cvt.f32.f16 %f66, %rs2;}


mov.u32 %r187, %r223;

prmt.b32 %r187, %r187, 0, 0xba98;

mov.u32 %r189, %r225;

prmt.b32 %r189, %r189, 0, 0xba98;

xor.b32 %r335, %r189, 2139062143;
not.b32 %r336, %r187;
and.b32 %r337, %r282, %r336;
and.b32 %r338, %r335, %r187;
or.b32 %r192, %r338, %r337;

dp4a.s32.s32 %r191, %r192, %r175, %r238;

mul.lo.s32 %r339, %r191, %r276;
cvt.rn.f32.s32 %f93, %r339;
fma.rn.ftz.f32 %f94, %f64, %f93, 0f00000000;
mov.u32 %r195, %r231;

prmt.b32 %r195, %r195, 0, 0xba98;

mov.u32 %r197, %r233;

prmt.b32 %r197, %r197, 0, 0xba98;

xor.b32 %r340, %r197, 2139062143;
not.b32 %r341, %r195;
and.b32 %r342, %r297, %r341;
and.b32 %r343, %r340, %r195;
or.b32 %r200, %r343, %r342;

dp4a.s32.s32 %r199, %r200, %r183, %r238;

mul.lo.s32 %r344, %r199, %r291;
cvt.rn.f32.s32 %f95, %r344;
fma.rn.ftz.f32 %f96, %f65, %f95, %f94;
fma.rn.ftz.f32 %f235, %f66, %f96, %f235;
cvt.u64.u32 %rd60, %r517;
add.s64 %rd61, %rd2, %rd60;
add.s64 %rd62, %rd61, %rd5;
mul.lo.s64 %rd63, %rd62, 36;
add.s64 %rd64, %rd14, %rd63;
add.s64 %rd65, %rd64, %rd40;
ld.global.nc.u32 %r211, [%rd65+4];
ld.global.nc.u32 %r203, [%rd64];

{.reg .f16 low,high;
mov.b32 {low,high},%r203;
cvt.f32.f16 %f67, low;}


ld.global.nc.u32 %r219, [%rd65+76];
ld.global.nc.u32 %r204, [%rd64+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r204;
cvt.f32.f16 %f68, low;}



{ cvt.f32.f16 %f69, %rs1;}



prmt.b32 %r205, %r205, 0, 0xba98;


prmt.b32 %r207, %r207, 0, 0xba98;

xor.b32 %r345, %r207, 2139062143;
not.b32 %r346, %r205;
and.b32 %r347, %r249, %r346;
and.b32 %r348, %r345, %r205;
or.b32 %r210, %r348, %r347;

dp4a.s32.s32 %r209, %r210, %r211, %r238;

mul.lo.s32 %r349, %r209, %r243;
cvt.rn.f32.s32 %f97, %r349;
fma.rn.ftz.f32 %f98, %f67, %f97, 0f00000000;

prmt.b32 %r213, %r213, 0, 0xba98;


prmt.b32 %r215, %r215, 0, 0xba98;

xor.b32 %r350, %r215, 2139062143;
not.b32 %r351, %r213;
and.b32 %r352, %r264, %r351;
and.b32 %r353, %r350, %r213;
or.b32 %r218, %r353, %r352;

dp4a.s32.s32 %r217, %r218, %r219, %r238;

mul.lo.s32 %r354, %r217, %r258;
cvt.rn.f32.s32 %f99, %r354;
fma.rn.ftz.f32 %f100, %f68, %f99, %f98;
fma.rn.ftz.f32 %f234, %f69, %f100, %f234;

{.reg .f16 low,high;
mov.b32 {low,high},%r203;
cvt.f32.f16 %f70, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r204;
cvt.f32.f16 %f71, low;}



{ cvt.f32.f16 %f72, %rs2;}



prmt.b32 %r223, %r223, 0, 0xba98;


prmt.b32 %r225, %r225, 0, 0xba98;

xor.b32 %r355, %r225, 2139062143;
not.b32 %r356, %r223;
and.b32 %r357, %r282, %r356;
and.b32 %r358, %r355, %r223;
or.b32 %r228, %r358, %r357;

dp4a.s32.s32 %r227, %r228, %r211, %r238;

mul.lo.s32 %r359, %r227, %r276;
cvt.rn.f32.s32 %f101, %r359;
fma.rn.ftz.f32 %f102, %f70, %f101, 0f00000000;

prmt.b32 %r231, %r231, 0, 0xba98;


prmt.b32 %r233, %r233, 0, 0xba98;

xor.b32 %r360, %r233, 2139062143;
not.b32 %r361, %r231;
and.b32 %r362, %r297, %r361;
and.b32 %r363, %r360, %r231;
or.b32 %r236, %r363, %r362;

dp4a.s32.s32 %r235, %r236, %r219, %r238;

mul.lo.s32 %r364, %r235, %r291;
cvt.rn.f32.s32 %f103, %r364;
fma.rn.ftz.f32 %f104, %f71, %f103, %f102;
fma.rn.ftz.f32 %f233, %f72, %f104, %f233;
add.s32 %r518, %r518, 32;
add.s32 %r517, %r517, 32;
add.s32 %r516, %r516, 32;
add.s32 %r515, %r515, 32;
add.s64 %rd86, %rd86, 840;
add.s32 %r519, %r519, 4;
setp.lt.s32 %p4, %r519, %r7;
@%p4 bra $L__BB193_2;

$L__BB193_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB193_5;

shl.b32 %r365, %r1, 10;
mov.u32 %r366, _ZZ13mul_mat_vec_qIL9ggml_type14ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r367, %r366, %r365;
shl.b32 %r368, %r2, 2;
add.s32 %r369, %r367, %r368;
st.shared.f32 [%r369+-1024], %f240;
st.shared.f32 [%r369+-896], %f239;
st.shared.f32 [%r369+-768], %f238;
st.shared.f32 [%r369+-640], %f237;
st.shared.f32 [%r369+-512], %f236;
st.shared.f32 [%r369+-384], %f235;
st.shared.f32 [%r369+-256], %f234;
st.shared.f32 [%r369+-128], %f233;

$L__BB193_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB193_22;

mov.u32 %r370, %ctaid.z;
mad.lo.s32 %r371, %r5, %r34, %r4;
mad.lo.s32 %r372, %r370, %r40, %r371;
cvt.u64.u32 %rd17, %r372;
shl.b32 %r373, %r2, 2;
mov.u32 %r374, _ZZ13mul_mat_vec_qIL9ggml_type14ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r23, %r374, %r373;
mov.u32 %r375, 2;
ld.shared.f32 %f105, [%r23];
add.ftz.f32 %f106, %f105, %f240;
ld.shared.f32 %f107, [%r23+1024];
add.ftz.f32 %f108, %f107, %f106;
ld.shared.f32 %f109, [%r23+2048];
add.ftz.f32 %f110, %f109, %f108;
mov.b32 %r376, %f110;
mov.u32 %r377, 31;
mov.u32 %r378, 16;
mov.u32 %r379, -1;
shfl.sync.bfly.b32 %r380|%p7, %r376, %r378, %r377, %r379;
mov.b32 %f111, %r380;
add.ftz.f32 %f112, %f110, %f111;
mov.b32 %r381, %f112;
mov.u32 %r382, 8;
shfl.sync.bfly.b32 %r383|%p8, %r381, %r382, %r377, %r379;
mov.b32 %f113, %r383;
add.ftz.f32 %f114, %f112, %f113;
mov.b32 %r384, %f114;
mov.u32 %r385, 4;
shfl.sync.bfly.b32 %r386|%p9, %r384, %r385, %r377, %r379;
mov.b32 %f115, %r386;
add.ftz.f32 %f116, %f114, %f115;
mov.b32 %r387, %f116;
shfl.sync.bfly.b32 %r388|%p10, %r387, %r375, %r377, %r379;
mov.b32 %f117, %r388;
add.ftz.f32 %f118, %f116, %f117;
mov.b32 %r389, %f118;
mov.u32 %r390, 1;
shfl.sync.bfly.b32 %r391|%p11, %r389, %r390, %r377, %r379;
mov.b32 %f119, %r391;
add.ftz.f32 %f25, %f118, %f119;
setp.lt.u32 %p12, %r4, %r28;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
mul.wide.u32 %rd66, %r372, 4;
add.s64 %rd18, %rd1, %rd66;
not.pred %p14, %p1;
@%p14 bra $L__BB193_8;

st.global.f32 [%rd18], %f25;

$L__BB193_8:
ld.shared.f32 %f120, [%r23+128];
add.ftz.f32 %f121, %f120, %f239;
ld.shared.f32 %f122, [%r23+1152];
add.ftz.f32 %f123, %f122, %f121;
ld.shared.f32 %f124, [%r23+2176];
add.ftz.f32 %f125, %f124, %f123;
mov.b32 %r392, %f125;
shfl.sync.bfly.b32 %r396|%p15, %r392, %r378, %r377, %r379;
mov.b32 %f126, %r396;
add.ftz.f32 %f127, %f125, %f126;
mov.b32 %r397, %f127;
shfl.sync.bfly.b32 %r399|%p16, %r397, %r382, %r377, %r379;
mov.b32 %f128, %r399;
add.ftz.f32 %f129, %f127, %f128;
mov.b32 %r400, %f129;
shfl.sync.bfly.b32 %r402|%p17, %r400, %r385, %r377, %r379;
mov.b32 %f130, %r402;
add.ftz.f32 %f131, %f129, %f130;
mov.b32 %r403, %f131;
shfl.sync.bfly.b32 %r405|%p18, %r403, %r375, %r377, %r379;
mov.b32 %f132, %r405;
add.ftz.f32 %f133, %f131, %f132;
mov.b32 %r406, %f133;
shfl.sync.bfly.b32 %r408|%p19, %r406, %r390, %r377, %r379;
mov.b32 %f134, %r408;
add.ftz.f32 %f26, %f133, %f134;
add.s32 %r409, %r4, 1;
setp.lt.u32 %p20, %r409, %r28;
setp.eq.s32 %p21, %r2, 1;
and.pred %p2, %p21, %p20;
not.pred %p22, %p2;
@%p22 bra $L__BB193_10;

st.global.f32 [%rd18+4], %f26;

$L__BB193_10:
ld.shared.f32 %f135, [%r23+256];
add.ftz.f32 %f136, %f135, %f238;
ld.shared.f32 %f137, [%r23+1280];
add.ftz.f32 %f138, %f137, %f136;
ld.shared.f32 %f139, [%r23+2304];
add.ftz.f32 %f140, %f139, %f138;
mov.b32 %r410, %f140;
mov.u32 %r411, 31;
mov.u32 %r412, 16;
mov.u32 %r413, -1;
shfl.sync.bfly.b32 %r414|%p23, %r410, %r412, %r411, %r413;
mov.b32 %f141, %r414;
add.ftz.f32 %f142, %f140, %f141;
mov.b32 %r415, %f142;
mov.u32 %r416, 8;
shfl.sync.bfly.b32 %r417|%p24, %r415, %r416, %r411, %r413;
mov.b32 %f143, %r417;
add.ftz.f32 %f144, %f142, %f143;
mov.b32 %r418, %f144;
mov.u32 %r419, 4;
shfl.sync.bfly.b32 %r420|%p25, %r418, %r419, %r411, %r413;
mov.b32 %f145, %r420;
add.ftz.f32 %f146, %f144, %f145;
mov.b32 %r421, %f146;
mov.u32 %r422, 2;
shfl.sync.bfly.b32 %r423|%p26, %r421, %r422, %r411, %r413;
mov.b32 %f147, %r423;
add.ftz.f32 %f148, %f146, %f147;
mov.b32 %r424, %f148;
mov.u32 %r425, 1;
shfl.sync.bfly.b32 %r426|%p27, %r424, %r425, %r411, %r413;
mov.b32 %f149, %r426;
add.ftz.f32 %f27, %f148, %f149;
@%p14 bra $L__BB193_12;

cvt.u64.u32 %rd67, %r28;
add.s64 %rd68, %rd67, %rd17;
shl.b64 %rd69, %rd68, 2;
add.s64 %rd70, %rd1, %rd69;
st.global.f32 [%rd70], %f27;

$L__BB193_12:
ld.shared.f32 %f150, [%r23+384];
add.ftz.f32 %f151, %f150, %f237;
ld.shared.f32 %f152, [%r23+1408];
add.ftz.f32 %f153, %f152, %f151;
ld.shared.f32 %f154, [%r23+2432];
add.ftz.f32 %f155, %f154, %f153;
mov.b32 %r427, %f155;
shfl.sync.bfly.b32 %r431|%p29, %r427, %r412, %r411, %r413;
mov.b32 %f156, %r431;
add.ftz.f32 %f157, %f155, %f156;
mov.b32 %r432, %f157;
shfl.sync.bfly.b32 %r434|%p30, %r432, %r416, %r411, %r413;
mov.b32 %f158, %r434;
add.ftz.f32 %f159, %f157, %f158;
mov.b32 %r435, %f159;
shfl.sync.bfly.b32 %r437|%p31, %r435, %r419, %r411, %r413;
mov.b32 %f160, %r437;
add.ftz.f32 %f161, %f159, %f160;
mov.b32 %r438, %f161;
shfl.sync.bfly.b32 %r440|%p32, %r438, %r422, %r411, %r413;
mov.b32 %f162, %r440;
add.ftz.f32 %f163, %f161, %f162;
mov.b32 %r441, %f163;
shfl.sync.bfly.b32 %r443|%p33, %r441, %r425, %r411, %r413;
mov.b32 %f164, %r443;
add.ftz.f32 %f28, %f163, %f164;
@%p22 bra $L__BB193_14;

add.s32 %r444, %r28, 1;
cvt.u64.u32 %rd71, %r444;
add.s64 %rd72, %rd71, %rd17;
shl.b64 %rd73, %rd72, 2;
add.s64 %rd74, %rd1, %rd73;
st.global.f32 [%rd74], %f28;

$L__BB193_14:
shl.b32 %r445, %r28, 1;
mov.u32 %r446, 1;
ld.shared.f32 %f165, [%r23+512];
add.ftz.f32 %f166, %f165, %f236;
ld.shared.f32 %f167, [%r23+1536];
add.ftz.f32 %f168, %f167, %f166;
ld.shared.f32 %f169, [%r23+2560];
add.ftz.f32 %f170, %f169, %f168;
mov.b32 %r447, %f170;
mov.u32 %r448, 31;
mov.u32 %r449, 16;
mov.u32 %r450, -1;
shfl.sync.bfly.b32 %r451|%p35, %r447, %r449, %r448, %r450;
mov.b32 %f171, %r451;
add.ftz.f32 %f172, %f170, %f171;
mov.b32 %r452, %f172;
mov.u32 %r453, 8;
shfl.sync.bfly.b32 %r454|%p36, %r452, %r453, %r448, %r450;
mov.b32 %f173, %r454;
add.ftz.f32 %f174, %f172, %f173;
mov.b32 %r455, %f174;
mov.u32 %r456, 4;
shfl.sync.bfly.b32 %r457|%p37, %r455, %r456, %r448, %r450;
mov.b32 %f175, %r457;
add.ftz.f32 %f176, %f174, %f175;
mov.b32 %r458, %f176;
mov.u32 %r459, 2;
shfl.sync.bfly.b32 %r460|%p38, %r458, %r459, %r448, %r450;
mov.b32 %f177, %r460;
add.ftz.f32 %f178, %f176, %f177;
mov.b32 %r461, %f178;
shfl.sync.bfly.b32 %r462|%p39, %r461, %r446, %r448, %r450;
mov.b32 %f179, %r462;
add.ftz.f32 %f29, %f178, %f179;
cvt.u64.u32 %rd75, %r445;
add.s64 %rd76, %rd75, %rd17;
shl.b64 %rd77, %rd76, 2;
add.s64 %rd19, %rd1, %rd77;
@%p14 bra $L__BB193_16;

st.global.f32 [%rd19], %f29;

$L__BB193_16:
ld.shared.f32 %f180, [%r23+640];
add.ftz.f32 %f181, %f180, %f235;
ld.shared.f32 %f182, [%r23+1664];
add.ftz.f32 %f183, %f182, %f181;
ld.shared.f32 %f184, [%r23+2688];
add.ftz.f32 %f185, %f184, %f183;
mov.b32 %r463, %f185;
shfl.sync.bfly.b32 %r467|%p41, %r463, %r449, %r448, %r450;
mov.b32 %f186, %r467;
add.ftz.f32 %f187, %f185, %f186;
mov.b32 %r468, %f187;
shfl.sync.bfly.b32 %r470|%p42, %r468, %r453, %r448, %r450;
mov.b32 %f188, %r470;
add.ftz.f32 %f189, %f187, %f188;
mov.b32 %r471, %f189;
shfl.sync.bfly.b32 %r473|%p43, %r471, %r456, %r448, %r450;
mov.b32 %f190, %r473;
add.ftz.f32 %f191, %f189, %f190;
mov.b32 %r474, %f191;
shfl.sync.bfly.b32 %r476|%p44, %r474, %r459, %r448, %r450;
mov.b32 %f192, %r476;
add.ftz.f32 %f193, %f191, %f192;
mov.b32 %r477, %f193;
shfl.sync.bfly.b32 %r479|%p45, %r477, %r446, %r448, %r450;
mov.b32 %f194, %r479;
add.ftz.f32 %f30, %f193, %f194;
@%p22 bra $L__BB193_18;

st.global.f32 [%rd19+4], %f30;

$L__BB193_18:
mul.lo.s32 %r25, %r28, 3;
ld.shared.f32 %f195, [%r23+768];
add.ftz.f32 %f196, %f195, %f234;
ld.shared.f32 %f197, [%r23+1792];
add.ftz.f32 %f198, %f197, %f196;
ld.shared.f32 %f199, [%r23+2816];
add.ftz.f32 %f200, %f199, %f198;
mov.b32 %r480, %f200;
mov.u32 %r481, 31;
mov.u32 %r482, 16;
mov.u32 %r483, -1;
shfl.sync.bfly.b32 %r484|%p47, %r480, %r482, %r481, %r483;
mov.b32 %f201, %r484;
add.ftz.f32 %f202, %f200, %f201;
mov.b32 %r485, %f202;
mov.u32 %r486, 8;
shfl.sync.bfly.b32 %r487|%p48, %r485, %r486, %r481, %r483;
mov.b32 %f203, %r487;
add.ftz.f32 %f204, %f202, %f203;
mov.b32 %r488, %f204;
mov.u32 %r489, 4;
shfl.sync.bfly.b32 %r490|%p49, %r488, %r489, %r481, %r483;
mov.b32 %f205, %r490;
add.ftz.f32 %f206, %f204, %f205;
mov.b32 %r491, %f206;
mov.u32 %r492, 2;
shfl.sync.bfly.b32 %r493|%p50, %r491, %r492, %r481, %r483;
mov.b32 %f207, %r493;
add.ftz.f32 %f208, %f206, %f207;
mov.b32 %r494, %f208;
mov.u32 %r495, 1;
shfl.sync.bfly.b32 %r496|%p51, %r494, %r495, %r481, %r483;
mov.b32 %f209, %r496;
add.ftz.f32 %f31, %f208, %f209;
@%p14 bra $L__BB193_20;

cvt.u64.u32 %rd78, %r25;
add.s64 %rd79, %rd78, %rd17;
shl.b64 %rd80, %rd79, 2;
add.s64 %rd81, %rd1, %rd80;
st.global.f32 [%rd81], %f31;

$L__BB193_20:
ld.shared.f32 %f210, [%r23+896];
add.ftz.f32 %f211, %f210, %f233;
ld.shared.f32 %f212, [%r23+1920];
add.ftz.f32 %f213, %f212, %f211;
ld.shared.f32 %f214, [%r23+2944];
add.ftz.f32 %f215, %f214, %f213;
mov.b32 %r497, %f215;
shfl.sync.bfly.b32 %r501|%p53, %r497, %r482, %r481, %r483;
mov.b32 %f216, %r501;
add.ftz.f32 %f217, %f215, %f216;
mov.b32 %r502, %f217;
shfl.sync.bfly.b32 %r504|%p54, %r502, %r486, %r481, %r483;
mov.b32 %f218, %r504;
add.ftz.f32 %f219, %f217, %f218;
mov.b32 %r505, %f219;
shfl.sync.bfly.b32 %r507|%p55, %r505, %r489, %r481, %r483;
mov.b32 %f220, %r507;
add.ftz.f32 %f221, %f219, %f220;
mov.b32 %r508, %f221;
shfl.sync.bfly.b32 %r510|%p56, %r508, %r492, %r481, %r483;
mov.b32 %f222, %r510;
add.ftz.f32 %f223, %f221, %f222;
mov.b32 %r511, %f223;
shfl.sync.bfly.b32 %r513|%p57, %r511, %r495, %r481, %r483;
mov.b32 %f224, %r513;
add.ftz.f32 %f32, %f223, %f224;
@%p22 bra $L__BB193_22;

add.s32 %r514, %r25, 1;
cvt.u64.u32 %rd82, %r514;
add.s64 %rd83, %rd82, %rd17;
shl.b64 %rd84, %rd83, 2;
add.s64 %rd85, %rd1, %rd84;
st.global.f32 [%rd85], %f32;

$L__BB193_22:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type14ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type14ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type14ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type14ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type14ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type14ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type14ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type14ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type14ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<71>;
.reg .b16 %rs<23>;
.reg .f32 %f<261>;
.reg .b32 %r<614>;
.reg .b64 %rd<97>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type14ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1280];

ld.param.u64 %rd21, [_Z13mul_mat_vec_qIL9ggml_type14ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd22, [_Z13mul_mat_vec_qIL9ggml_type14ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd23, [_Z13mul_mat_vec_qIL9ggml_type14ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type14ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type14ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type14ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type14ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type14ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type14ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type14ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type14ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type14ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type14ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type14ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type14ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type14ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type14ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
cvta.to.global.u64 %rd1, %rd23;
mov.u32 %r1, %tid.y;
shl.b32 %r44, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r44, %r2;
mov.u32 %r45, %ctaid.x;
shl.b32 %r4, %r45, 1;
mov.u32 %r5, %ctaid.y;
shr.s32 %r46, %r3, 31;
shr.u32 %r47, %r46, 27;
add.s32 %r48, %r3, %r47;
shr.s32 %r613, %r48, 5;
shr.u32 %r7, %r43, 8;
setp.ge.s32 %p3, %r613, %r7;
mov.f32 %f251, 0f00000000;
mov.f32 %f252, %f251;
mov.f32 %f253, %f251;
mov.f32 %f254, %f251;
mov.f32 %f255, %f251;
mov.f32 %f256, %f251;
mov.f32 %f257, %f251;
mov.f32 %f258, %f251;
mov.f32 %f259, %f251;
mov.f32 %f260, %f251;
@%p3 bra $L__BB194_3;

mov.u32 %r49, %ctaid.z;
mul.lo.s32 %r50, %r5, %r35;
cvt.u64.u32 %rd24, %r50;
mul.lo.s32 %r51, %r49, %r41;
cvt.u64.u32 %rd25, %r51;
add.s64 %rd2, %rd25, %rd24;
mul.hi.u32 %r52, %r49, %r37;
add.s32 %r53, %r52, %r49;
shr.u32 %r54, %r53, %r38;
mul.lo.s32 %r55, %r54, %r40;
mul.hi.u32 %r56, %r5, %r31;
add.s32 %r57, %r56, %r5;
shr.u32 %r58, %r57, %r32;
mul.lo.s32 %r59, %r58, %r34;
cvta.to.global.u64 %rd96, %rd21;
add.s32 %r60, %r4, 1;
and.b32 %r64, %r48, -32;
sub.s32 %r65, %r3, %r64;
shr.s32 %r66, %r65, 31;
shr.u32 %r67, %r66, 28;
add.s32 %r68, %r65, %r67;
shr.s32 %r69, %r68, 4;
shl.b32 %r70, %r69, 3;
and.b32 %r71, %r68, -16;
sub.s32 %r72, %r65, %r71;
shr.s32 %r73, %r72, 31;
shr.u32 %r74, %r73, 30;
add.s32 %r75, %r72, %r74;
shr.s32 %r76, %r75, 2;
add.s32 %r77, %r70, %r76;
cvt.s64.s32 %rd26, %r77;
shl.b32 %r78, %r69, 2;
shr.u32 %r79, %r73, 29;
add.s32 %r80, %r72, %r79;
shr.s32 %r81, %r80, 3;
add.s32 %r82, %r78, %r81;
shl.b32 %r8, %r81, 1;
shl.b32 %r83, %r65, 1;
shr.u32 %r84, %r66, 29;
add.s32 %r85, %r65, %r84;
and.b32 %r86, %r85, -8;
sub.s32 %r87, %r65, %r86;
add.s32 %r88, %r70, %r87;
shl.b32 %r89, %r88, 1;
cvt.s64.s32 %rd4, %r87;
cvt.s64.s32 %rd5, %r82;
shl.b32 %r90, %r29, 1;
shl.b32 %r608, %r613, 3;
add.s32 %r612, %r608, %r90;
mad.lo.s32 %r611, %r29, 3, %r608;
shl.b32 %r91, %r29, 2;
add.s32 %r610, %r608, %r91;
add.s32 %r609, %r29, %r608;
mad.lo.s32 %r92, %r4, %r28, %r613;
add.s32 %r93, %r92, %r55;
add.s32 %r94, %r93, %r59;
mul.wide.s32 %rd6, %r94, 210;
mul.wide.s32 %rd27, %r83, 2;
add.s64 %rd28, %rd6, %rd27;
add.s64 %rd7, %rd28, 2;
add.s32 %r95, %r613, %r55;
add.s32 %r96, %r95, %r59;
mad.lo.s32 %r97, %r60, %r28, %r96;
mul.wide.s32 %rd8, %r97, 210;
add.s64 %rd29, %rd8, %rd26;
add.s64 %rd9, %rd29, 192;
mul.wide.s32 %rd30, %r89, 2;
add.s64 %rd31, %rd6, %rd30;
add.s64 %rd10, %rd31, 130;
add.s64 %rd32, %rd8, %rd30;
add.s64 %rd11, %rd32, 128;
add.s64 %rd12, %rd8, %rd27;
add.s64 %rd33, %rd6, %rd26;
add.s64 %rd13, %rd33, 192;
cvta.to.global.u64 %rd14, %rd22;
mov.f32 %f251, 0f00000000;
shl.b64 %rd41, %rd4, 2;

$L__BB194_2:
cvt.u64.u32 %rd34, %r608;
add.s64 %rd35, %rd2, %rd34;
add.s64 %rd36, %rd96, %rd7;
ld.global.nc.u16 %rs11, [%rd36+-2];
ld.global.nc.u16 %rs12, [%rd36];
mov.b32 %r278, {%rs11, %rs12};
add.s64 %rd37, %rd96, %rd10;
ld.global.nc.u16 %rs13, [%rd37+-2];
ld.global.nc.u16 %rs14, [%rd37];
mov.b32 %r279, {%rs13, %rs14};
shr.s32 %r280, %r279, %r8;
add.s64 %rd38, %rd35, %rd5;
mul.lo.s64 %rd39, %rd38, 36;
add.s64 %rd40, %rd14, %rd39;
add.s64 %rd42, %rd40, %rd41;
ld.global.nc.u32 %r106, [%rd42+4];
ld.global.nc.u32 %r98, [%rd40];

{.reg .f16 low,high;
mov.b32 {low,high},%r98;
cvt.f32.f16 %f61, low;}


ld.global.nc.u32 %r114, [%rd42+76];
ld.global.nc.u32 %r99, [%rd40+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r99;
cvt.f32.f16 %f62, low;}


add.s64 %rd43, %rd96, %rd6;
ld.global.nc.u16 %rs1, [%rd43+208];

{ cvt.f32.f16 %f63, %rs1;}


add.s64 %rd44, %rd96, %rd13;
ld.global.nc.u8 %rs15, [%rd44];
cvt.u32.u16 %r281, %rs15;
cvt.s32.s8 %r282, %r281;
and.b32 %r283, %r278, 252645135;
shl.b32 %r284, %r280, 4;
and.b32 %r285, %r284, 808464432;
or.b32 %r246, %r285, %r283;
or.b32 %r286, %r246, -2139062144;
add.s32 %r287, %r286, -538976288;
xor.b32 %r288, %r287, -2139062144;
xor.b32 %r289, %r246, 538976288;
xor.b32 %r290, %r287, %r246;
and.b32 %r244, %r290, %r289;
mov.u32 %r100, %r244;

prmt.b32 %r100, %r100, 0, 0xba98;

mov.u32 %r102, %r246;

prmt.b32 %r102, %r102, 0, 0xba98;

xor.b32 %r291, %r102, 2139062143;
not.b32 %r292, %r100;
and.b32 %r293, %r288, %r292;
and.b32 %r294, %r291, %r100;
or.b32 %r105, %r294, %r293;
mov.u32 %r277, 0;

dp4a.s32.s32 %r104, %r105, %r106, %r277;

mul.lo.s32 %r295, %r104, %r282;
cvt.rn.f32.s32 %f91, %r295;
fma.rn.ftz.f32 %f92, %f61, %f91, 0f00000000;
ld.global.nc.u8 %rs16, [%rd44+4];
cvt.u32.u16 %r296, %rs16;
cvt.s32.s8 %r297, %r296;
shr.u32 %r298, %r278, 4;
and.b32 %r299, %r298, 252645135;
and.b32 %r300, %r280, 808464432;
or.b32 %r254, %r300, %r299;
or.b32 %r301, %r254, -2139062144;
add.s32 %r302, %r301, -538976288;
xor.b32 %r303, %r302, -2139062144;
xor.b32 %r304, %r254, 538976288;
xor.b32 %r305, %r302, %r254;
and.b32 %r252, %r305, %r304;
mov.u32 %r108, %r252;

prmt.b32 %r108, %r108, 0, 0xba98;

mov.u32 %r110, %r254;

prmt.b32 %r110, %r110, 0, 0xba98;

xor.b32 %r306, %r110, 2139062143;
not.b32 %r307, %r108;
and.b32 %r308, %r303, %r307;
and.b32 %r309, %r306, %r108;
or.b32 %r113, %r309, %r308;

dp4a.s32.s32 %r112, %r113, %r114, %r277;

mul.lo.s32 %r310, %r112, %r297;
cvt.rn.f32.s32 %f93, %r310;
fma.rn.ftz.f32 %f94, %f62, %f93, %f92;
fma.rn.ftz.f32 %f260, %f63, %f94, %f260;
add.s64 %rd45, %rd96, %rd12;
ld.global.nc.u16 %rs17, [%rd45];
ld.global.nc.u16 %rs18, [%rd45+2];
mov.b32 %r311, {%rs17, %rs18};
add.s64 %rd46, %rd96, %rd11;
ld.global.nc.u16 %rs19, [%rd46];
ld.global.nc.u16 %rs20, [%rd46+2];
mov.b32 %r312, {%rs19, %rs20};
shr.s32 %r313, %r312, %r8;

{.reg .f16 low,high;
mov.b32 {low,high},%r98;
cvt.f32.f16 %f64, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r99;
cvt.f32.f16 %f65, low;}


add.s64 %rd47, %rd96, %rd8;
ld.global.nc.u16 %rs2, [%rd47+208];

{ cvt.f32.f16 %f66, %rs2;}


add.s64 %rd48, %rd96, %rd9;
ld.global.nc.u8 %rs21, [%rd48];
cvt.u32.u16 %r314, %rs21;
cvt.s32.s8 %r315, %r314;
and.b32 %r316, %r311, 252645135;
shl.b32 %r317, %r313, 4;
and.b32 %r318, %r317, 808464432;
or.b32 %r264, %r318, %r316;
or.b32 %r319, %r264, -2139062144;
add.s32 %r320, %r319, -538976288;
xor.b32 %r321, %r320, -2139062144;
xor.b32 %r322, %r264, 538976288;
xor.b32 %r323, %r320, %r264;
and.b32 %r262, %r323, %r322;
mov.u32 %r118, %r262;

prmt.b32 %r118, %r118, 0, 0xba98;

mov.u32 %r120, %r264;

prmt.b32 %r120, %r120, 0, 0xba98;

xor.b32 %r324, %r120, 2139062143;
not.b32 %r325, %r118;
and.b32 %r326, %r321, %r325;
and.b32 %r327, %r324, %r118;
or.b32 %r123, %r327, %r326;

dp4a.s32.s32 %r122, %r123, %r106, %r277;

mul.lo.s32 %r328, %r122, %r315;
cvt.rn.f32.s32 %f95, %r328;
fma.rn.ftz.f32 %f96, %f64, %f95, 0f00000000;
ld.global.nc.u8 %rs22, [%rd48+4];
cvt.u32.u16 %r329, %rs22;
cvt.s32.s8 %r330, %r329;
shr.u32 %r331, %r311, 4;
and.b32 %r332, %r331, 252645135;
and.b32 %r333, %r313, 808464432;
or.b32 %r272, %r333, %r332;
or.b32 %r334, %r272, -2139062144;
add.s32 %r335, %r334, -538976288;
xor.b32 %r336, %r335, -2139062144;
xor.b32 %r337, %r272, 538976288;
xor.b32 %r338, %r335, %r272;
and.b32 %r270, %r338, %r337;
mov.u32 %r126, %r270;

prmt.b32 %r126, %r126, 0, 0xba98;

mov.u32 %r128, %r272;

prmt.b32 %r128, %r128, 0, 0xba98;

xor.b32 %r339, %r128, 2139062143;
not.b32 %r340, %r126;
and.b32 %r341, %r336, %r340;
and.b32 %r342, %r339, %r126;
or.b32 %r131, %r342, %r341;

dp4a.s32.s32 %r130, %r131, %r114, %r277;

mul.lo.s32 %r343, %r130, %r330;
cvt.rn.f32.s32 %f97, %r343;
fma.rn.ftz.f32 %f98, %f65, %f97, %f96;
fma.rn.ftz.f32 %f259, %f66, %f98, %f259;
cvt.u64.u32 %rd49, %r609;
add.s64 %rd50, %rd2, %rd49;
add.s64 %rd51, %rd50, %rd5;
mul.lo.s64 %rd52, %rd51, 36;
add.s64 %rd53, %rd14, %rd52;
add.s64 %rd54, %rd53, %rd41;
ld.global.nc.u32 %r142, [%rd54+4];
ld.global.nc.u32 %r134, [%rd53];

{.reg .f16 low,high;
mov.b32 {low,high},%r134;
cvt.f32.f16 %f67, low;}


ld.global.nc.u32 %r150, [%rd54+76];
ld.global.nc.u32 %r135, [%rd53+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r135;
cvt.f32.f16 %f68, low;}



{ cvt.f32.f16 %f69, %rs1;}


mov.u32 %r136, %r244;

prmt.b32 %r136, %r136, 0, 0xba98;

mov.u32 %r138, %r246;

prmt.b32 %r138, %r138, 0, 0xba98;

xor.b32 %r344, %r138, 2139062143;
not.b32 %r345, %r136;
and.b32 %r346, %r288, %r345;
and.b32 %r347, %r344, %r136;
or.b32 %r141, %r347, %r346;

dp4a.s32.s32 %r140, %r141, %r142, %r277;

mul.lo.s32 %r348, %r140, %r282;
cvt.rn.f32.s32 %f99, %r348;
fma.rn.ftz.f32 %f100, %f67, %f99, 0f00000000;
mov.u32 %r144, %r252;

prmt.b32 %r144, %r144, 0, 0xba98;

mov.u32 %r146, %r254;

prmt.b32 %r146, %r146, 0, 0xba98;

xor.b32 %r349, %r146, 2139062143;
not.b32 %r350, %r144;
and.b32 %r351, %r303, %r350;
and.b32 %r352, %r349, %r144;
or.b32 %r149, %r352, %r351;

dp4a.s32.s32 %r148, %r149, %r150, %r277;

mul.lo.s32 %r353, %r148, %r297;
cvt.rn.f32.s32 %f101, %r353;
fma.rn.ftz.f32 %f102, %f68, %f101, %f100;
fma.rn.ftz.f32 %f258, %f69, %f102, %f258;

{.reg .f16 low,high;
mov.b32 {low,high},%r134;
cvt.f32.f16 %f70, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r135;
cvt.f32.f16 %f71, low;}



{ cvt.f32.f16 %f72, %rs2;}


mov.u32 %r154, %r262;

prmt.b32 %r154, %r154, 0, 0xba98;

mov.u32 %r156, %r264;

prmt.b32 %r156, %r156, 0, 0xba98;

xor.b32 %r354, %r156, 2139062143;
not.b32 %r355, %r154;
and.b32 %r356, %r321, %r355;
and.b32 %r357, %r354, %r154;
or.b32 %r159, %r357, %r356;

dp4a.s32.s32 %r158, %r159, %r142, %r277;

mul.lo.s32 %r358, %r158, %r315;
cvt.rn.f32.s32 %f103, %r358;
fma.rn.ftz.f32 %f104, %f70, %f103, 0f00000000;
mov.u32 %r162, %r270;

prmt.b32 %r162, %r162, 0, 0xba98;

mov.u32 %r164, %r272;

prmt.b32 %r164, %r164, 0, 0xba98;

xor.b32 %r359, %r164, 2139062143;
not.b32 %r360, %r162;
and.b32 %r361, %r336, %r360;
and.b32 %r362, %r359, %r162;
or.b32 %r167, %r362, %r361;

dp4a.s32.s32 %r166, %r167, %r150, %r277;

mul.lo.s32 %r363, %r166, %r330;
cvt.rn.f32.s32 %f105, %r363;
fma.rn.ftz.f32 %f106, %f71, %f105, %f104;
fma.rn.ftz.f32 %f257, %f72, %f106, %f257;
cvt.u64.u32 %rd55, %r612;
add.s64 %rd56, %rd2, %rd55;
add.s64 %rd57, %rd56, %rd5;
mul.lo.s64 %rd58, %rd57, 36;
add.s64 %rd59, %rd14, %rd58;
add.s64 %rd60, %rd59, %rd41;
ld.global.nc.u32 %r178, [%rd60+4];
ld.global.nc.u32 %r170, [%rd59];

{.reg .f16 low,high;
mov.b32 {low,high},%r170;
cvt.f32.f16 %f73, low;}


ld.global.nc.u32 %r186, [%rd60+76];
ld.global.nc.u32 %r171, [%rd59+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r171;
cvt.f32.f16 %f74, low;}



{ cvt.f32.f16 %f75, %rs1;}


mov.u32 %r172, %r244;

prmt.b32 %r172, %r172, 0, 0xba98;

mov.u32 %r174, %r246;

prmt.b32 %r174, %r174, 0, 0xba98;

xor.b32 %r364, %r174, 2139062143;
not.b32 %r365, %r172;
and.b32 %r366, %r288, %r365;
and.b32 %r367, %r364, %r172;
or.b32 %r177, %r367, %r366;

dp4a.s32.s32 %r176, %r177, %r178, %r277;

mul.lo.s32 %r368, %r176, %r282;
cvt.rn.f32.s32 %f107, %r368;
fma.rn.ftz.f32 %f108, %f73, %f107, 0f00000000;
mov.u32 %r180, %r252;

prmt.b32 %r180, %r180, 0, 0xba98;

mov.u32 %r182, %r254;

prmt.b32 %r182, %r182, 0, 0xba98;

xor.b32 %r369, %r182, 2139062143;
not.b32 %r370, %r180;
and.b32 %r371, %r303, %r370;
and.b32 %r372, %r369, %r180;
or.b32 %r185, %r372, %r371;

dp4a.s32.s32 %r184, %r185, %r186, %r277;

mul.lo.s32 %r373, %r184, %r297;
cvt.rn.f32.s32 %f109, %r373;
fma.rn.ftz.f32 %f110, %f74, %f109, %f108;
fma.rn.ftz.f32 %f256, %f75, %f110, %f256;

{.reg .f16 low,high;
mov.b32 {low,high},%r170;
cvt.f32.f16 %f76, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r171;
cvt.f32.f16 %f77, low;}



{ cvt.f32.f16 %f78, %rs2;}


mov.u32 %r190, %r262;

prmt.b32 %r190, %r190, 0, 0xba98;

mov.u32 %r192, %r264;

prmt.b32 %r192, %r192, 0, 0xba98;

xor.b32 %r374, %r192, 2139062143;
not.b32 %r375, %r190;
and.b32 %r376, %r321, %r375;
and.b32 %r377, %r374, %r190;
or.b32 %r195, %r377, %r376;

dp4a.s32.s32 %r194, %r195, %r178, %r277;

mul.lo.s32 %r378, %r194, %r315;
cvt.rn.f32.s32 %f111, %r378;
fma.rn.ftz.f32 %f112, %f76, %f111, 0f00000000;
mov.u32 %r198, %r270;

prmt.b32 %r198, %r198, 0, 0xba98;

mov.u32 %r200, %r272;

prmt.b32 %r200, %r200, 0, 0xba98;

xor.b32 %r379, %r200, 2139062143;
not.b32 %r380, %r198;
and.b32 %r381, %r336, %r380;
and.b32 %r382, %r379, %r198;
or.b32 %r203, %r382, %r381;

dp4a.s32.s32 %r202, %r203, %r186, %r277;

mul.lo.s32 %r383, %r202, %r330;
cvt.rn.f32.s32 %f113, %r383;
fma.rn.ftz.f32 %f114, %f77, %f113, %f112;
fma.rn.ftz.f32 %f255, %f78, %f114, %f255;
cvt.u64.u32 %rd61, %r611;
add.s64 %rd62, %rd2, %rd61;
add.s64 %rd63, %rd62, %rd5;
mul.lo.s64 %rd64, %rd63, 36;
add.s64 %rd65, %rd14, %rd64;
add.s64 %rd66, %rd65, %rd41;
ld.global.nc.u32 %r214, [%rd66+4];
ld.global.nc.u32 %r206, [%rd65];

{.reg .f16 low,high;
mov.b32 {low,high},%r206;
cvt.f32.f16 %f79, low;}


ld.global.nc.u32 %r222, [%rd66+76];
ld.global.nc.u32 %r207, [%rd65+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r207;
cvt.f32.f16 %f80, low;}



{ cvt.f32.f16 %f81, %rs1;}


mov.u32 %r208, %r244;

prmt.b32 %r208, %r208, 0, 0xba98;

mov.u32 %r210, %r246;

prmt.b32 %r210, %r210, 0, 0xba98;

xor.b32 %r384, %r210, 2139062143;
not.b32 %r385, %r208;
and.b32 %r386, %r288, %r385;
and.b32 %r387, %r384, %r208;
or.b32 %r213, %r387, %r386;

dp4a.s32.s32 %r212, %r213, %r214, %r277;

mul.lo.s32 %r388, %r212, %r282;
cvt.rn.f32.s32 %f115, %r388;
fma.rn.ftz.f32 %f116, %f79, %f115, 0f00000000;
mov.u32 %r216, %r252;

prmt.b32 %r216, %r216, 0, 0xba98;

mov.u32 %r218, %r254;

prmt.b32 %r218, %r218, 0, 0xba98;

xor.b32 %r389, %r218, 2139062143;
not.b32 %r390, %r216;
and.b32 %r391, %r303, %r390;
and.b32 %r392, %r389, %r216;
or.b32 %r221, %r392, %r391;

dp4a.s32.s32 %r220, %r221, %r222, %r277;

mul.lo.s32 %r393, %r220, %r297;
cvt.rn.f32.s32 %f117, %r393;
fma.rn.ftz.f32 %f118, %f80, %f117, %f116;
fma.rn.ftz.f32 %f254, %f81, %f118, %f254;

{.reg .f16 low,high;
mov.b32 {low,high},%r206;
cvt.f32.f16 %f82, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r207;
cvt.f32.f16 %f83, low;}



{ cvt.f32.f16 %f84, %rs2;}


mov.u32 %r226, %r262;

prmt.b32 %r226, %r226, 0, 0xba98;

mov.u32 %r228, %r264;

prmt.b32 %r228, %r228, 0, 0xba98;

xor.b32 %r394, %r228, 2139062143;
not.b32 %r395, %r226;
and.b32 %r396, %r321, %r395;
and.b32 %r397, %r394, %r226;
or.b32 %r231, %r397, %r396;

dp4a.s32.s32 %r230, %r231, %r214, %r277;

mul.lo.s32 %r398, %r230, %r315;
cvt.rn.f32.s32 %f119, %r398;
fma.rn.ftz.f32 %f120, %f82, %f119, 0f00000000;
mov.u32 %r234, %r270;

prmt.b32 %r234, %r234, 0, 0xba98;

mov.u32 %r236, %r272;

prmt.b32 %r236, %r236, 0, 0xba98;

xor.b32 %r399, %r236, 2139062143;
not.b32 %r400, %r234;
and.b32 %r401, %r336, %r400;
and.b32 %r402, %r399, %r234;
or.b32 %r239, %r402, %r401;

dp4a.s32.s32 %r238, %r239, %r222, %r277;

mul.lo.s32 %r403, %r238, %r330;
cvt.rn.f32.s32 %f121, %r403;
fma.rn.ftz.f32 %f122, %f83, %f121, %f120;
fma.rn.ftz.f32 %f253, %f84, %f122, %f253;
cvt.u64.u32 %rd67, %r610;
add.s64 %rd68, %rd2, %rd67;
add.s64 %rd69, %rd68, %rd5;
mul.lo.s64 %rd70, %rd69, 36;
add.s64 %rd71, %rd14, %rd70;
add.s64 %rd72, %rd71, %rd41;
ld.global.nc.u32 %r250, [%rd72+4];
ld.global.nc.u32 %r242, [%rd71];

{.reg .f16 low,high;
mov.b32 {low,high},%r242;
cvt.f32.f16 %f85, low;}


ld.global.nc.u32 %r258, [%rd72+76];
ld.global.nc.u32 %r243, [%rd71+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r243;
cvt.f32.f16 %f86, low;}



{ cvt.f32.f16 %f87, %rs1;}



prmt.b32 %r244, %r244, 0, 0xba98;


prmt.b32 %r246, %r246, 0, 0xba98;

xor.b32 %r404, %r246, 2139062143;
not.b32 %r405, %r244;
and.b32 %r406, %r288, %r405;
and.b32 %r407, %r404, %r244;
or.b32 %r249, %r407, %r406;

dp4a.s32.s32 %r248, %r249, %r250, %r277;

mul.lo.s32 %r408, %r248, %r282;
cvt.rn.f32.s32 %f123, %r408;
fma.rn.ftz.f32 %f124, %f85, %f123, 0f00000000;

prmt.b32 %r252, %r252, 0, 0xba98;


prmt.b32 %r254, %r254, 0, 0xba98;

xor.b32 %r409, %r254, 2139062143;
not.b32 %r410, %r252;
and.b32 %r411, %r303, %r410;
and.b32 %r412, %r409, %r252;
or.b32 %r257, %r412, %r411;

dp4a.s32.s32 %r256, %r257, %r258, %r277;

mul.lo.s32 %r413, %r256, %r297;
cvt.rn.f32.s32 %f125, %r413;
fma.rn.ftz.f32 %f126, %f86, %f125, %f124;
fma.rn.ftz.f32 %f252, %f87, %f126, %f252;

{.reg .f16 low,high;
mov.b32 {low,high},%r242;
cvt.f32.f16 %f88, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r243;
cvt.f32.f16 %f89, low;}



{ cvt.f32.f16 %f90, %rs2;}



prmt.b32 %r262, %r262, 0, 0xba98;


prmt.b32 %r264, %r264, 0, 0xba98;

xor.b32 %r414, %r264, 2139062143;
not.b32 %r415, %r262;
and.b32 %r416, %r321, %r415;
and.b32 %r417, %r414, %r262;
or.b32 %r267, %r417, %r416;

dp4a.s32.s32 %r266, %r267, %r250, %r277;

mul.lo.s32 %r418, %r266, %r315;
cvt.rn.f32.s32 %f127, %r418;
fma.rn.ftz.f32 %f128, %f88, %f127, 0f00000000;

prmt.b32 %r270, %r270, 0, 0xba98;


prmt.b32 %r272, %r272, 0, 0xba98;

xor.b32 %r419, %r272, 2139062143;
not.b32 %r420, %r270;
and.b32 %r421, %r336, %r420;
and.b32 %r422, %r419, %r270;
or.b32 %r275, %r422, %r421;

dp4a.s32.s32 %r274, %r275, %r258, %r277;

mul.lo.s32 %r423, %r274, %r330;
cvt.rn.f32.s32 %f129, %r423;
fma.rn.ftz.f32 %f130, %f89, %f129, %f128;
fma.rn.ftz.f32 %f251, %f90, %f130, %f251;
add.s32 %r612, %r612, 16;
add.s32 %r611, %r611, 16;
add.s32 %r610, %r610, 16;
add.s32 %r609, %r609, 16;
add.s32 %r608, %r608, 16;
add.s64 %rd96, %rd96, 420;
add.s32 %r613, %r613, 2;
setp.lt.s32 %p4, %r613, %r7;
@%p4 bra $L__BB194_2;

$L__BB194_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB194_5;

mov.u32 %r424, _ZZ13mul_mat_vec_qIL9ggml_type14ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r425, %r1, 1280, %r424;
shl.b32 %r426, %r2, 2;
add.s32 %r427, %r425, %r426;
st.shared.f32 [%r427+-1280], %f260;
st.shared.f32 [%r427+-1152], %f259;
st.shared.f32 [%r427+-1024], %f258;
st.shared.f32 [%r427+-896], %f257;
st.shared.f32 [%r427+-768], %f256;
st.shared.f32 [%r427+-640], %f255;
st.shared.f32 [%r427+-512], %f254;
st.shared.f32 [%r427+-384], %f253;
st.shared.f32 [%r427+-256], %f252;
st.shared.f32 [%r427+-128], %f251;

$L__BB194_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB194_26;

mov.u32 %r428, %ctaid.z;
mad.lo.s32 %r429, %r5, %r36, %r4;
mad.lo.s32 %r430, %r428, %r42, %r429;
cvt.u64.u32 %rd17, %r430;
shl.b32 %r431, %r2, 2;
mov.u32 %r432, _ZZ13mul_mat_vec_qIL9ggml_type14ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r26, %r432, %r431;
mov.u32 %r433, 2;
ld.shared.f32 %f131, [%r26];
add.ftz.f32 %f132, %f131, %f260;
mov.b32 %r434, %f132;
mov.u32 %r435, 31;
mov.u32 %r436, 16;
mov.u32 %r437, -1;
shfl.sync.bfly.b32 %r438|%p7, %r434, %r436, %r435, %r437;
mov.b32 %f133, %r438;
add.ftz.f32 %f134, %f132, %f133;
mov.b32 %r439, %f134;
mov.u32 %r440, 8;
shfl.sync.bfly.b32 %r441|%p8, %r439, %r440, %r435, %r437;
mov.b32 %f135, %r441;
add.ftz.f32 %f136, %f134, %f135;
mov.b32 %r442, %f136;
mov.u32 %r443, 4;
shfl.sync.bfly.b32 %r444|%p9, %r442, %r443, %r435, %r437;
mov.b32 %f137, %r444;
add.ftz.f32 %f138, %f136, %f137;
mov.b32 %r445, %f138;
shfl.sync.bfly.b32 %r446|%p10, %r445, %r433, %r435, %r437;
mov.b32 %f139, %r446;
add.ftz.f32 %f140, %f138, %f139;
mov.b32 %r447, %f140;
mov.u32 %r448, 1;
shfl.sync.bfly.b32 %r449|%p11, %r447, %r448, %r435, %r437;
mov.b32 %f141, %r449;
add.ftz.f32 %f31, %f140, %f141;
setp.lt.u32 %p12, %r4, %r30;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
mul.wide.u32 %rd73, %r430, 4;
add.s64 %rd18, %rd1, %rd73;
not.pred %p14, %p1;
@%p14 bra $L__BB194_8;

st.global.f32 [%rd18], %f31;

$L__BB194_8:
ld.shared.f32 %f142, [%r26+128];
add.ftz.f32 %f143, %f142, %f259;
mov.b32 %r450, %f143;
shfl.sync.bfly.b32 %r454|%p15, %r450, %r436, %r435, %r437;
mov.b32 %f144, %r454;
add.ftz.f32 %f145, %f143, %f144;
mov.b32 %r455, %f145;
shfl.sync.bfly.b32 %r457|%p16, %r455, %r440, %r435, %r437;
mov.b32 %f146, %r457;
add.ftz.f32 %f147, %f145, %f146;
mov.b32 %r458, %f147;
shfl.sync.bfly.b32 %r460|%p17, %r458, %r443, %r435, %r437;
mov.b32 %f148, %r460;
add.ftz.f32 %f149, %f147, %f148;
mov.b32 %r461, %f149;
shfl.sync.bfly.b32 %r463|%p18, %r461, %r433, %r435, %r437;
mov.b32 %f150, %r463;
add.ftz.f32 %f151, %f149, %f150;
mov.b32 %r464, %f151;
shfl.sync.bfly.b32 %r466|%p19, %r464, %r448, %r435, %r437;
mov.b32 %f152, %r466;
add.ftz.f32 %f32, %f151, %f152;
add.s32 %r467, %r4, 1;
setp.lt.u32 %p20, %r467, %r30;
setp.eq.s32 %p21, %r2, 1;
and.pred %p2, %p21, %p20;
not.pred %p22, %p2;
@%p22 bra $L__BB194_10;

st.global.f32 [%rd18+4], %f32;

$L__BB194_10:
ld.shared.f32 %f153, [%r26+256];
add.ftz.f32 %f154, %f153, %f258;
mov.b32 %r468, %f154;
mov.u32 %r469, 31;
mov.u32 %r470, 16;
mov.u32 %r471, -1;
shfl.sync.bfly.b32 %r472|%p23, %r468, %r470, %r469, %r471;
mov.b32 %f155, %r472;
add.ftz.f32 %f156, %f154, %f155;
mov.b32 %r473, %f156;
mov.u32 %r474, 8;
shfl.sync.bfly.b32 %r475|%p24, %r473, %r474, %r469, %r471;
mov.b32 %f157, %r475;
add.ftz.f32 %f158, %f156, %f157;
mov.b32 %r476, %f158;
mov.u32 %r477, 4;
shfl.sync.bfly.b32 %r478|%p25, %r476, %r477, %r469, %r471;
mov.b32 %f159, %r478;
add.ftz.f32 %f160, %f158, %f159;
mov.b32 %r479, %f160;
mov.u32 %r480, 2;
shfl.sync.bfly.b32 %r481|%p26, %r479, %r480, %r469, %r471;
mov.b32 %f161, %r481;
add.ftz.f32 %f162, %f160, %f161;
mov.b32 %r482, %f162;
mov.u32 %r483, 1;
shfl.sync.bfly.b32 %r484|%p27, %r482, %r483, %r469, %r471;
mov.b32 %f163, %r484;
add.ftz.f32 %f33, %f162, %f163;
@%p14 bra $L__BB194_12;

cvt.u64.u32 %rd74, %r30;
add.s64 %rd75, %rd74, %rd17;
shl.b64 %rd76, %rd75, 2;
add.s64 %rd77, %rd1, %rd76;
st.global.f32 [%rd77], %f33;

$L__BB194_12:
ld.shared.f32 %f164, [%r26+384];
add.ftz.f32 %f165, %f164, %f257;
mov.b32 %r485, %f165;
shfl.sync.bfly.b32 %r489|%p29, %r485, %r470, %r469, %r471;
mov.b32 %f166, %r489;
add.ftz.f32 %f167, %f165, %f166;
mov.b32 %r490, %f167;
shfl.sync.bfly.b32 %r492|%p30, %r490, %r474, %r469, %r471;
mov.b32 %f168, %r492;
add.ftz.f32 %f169, %f167, %f168;
mov.b32 %r493, %f169;
shfl.sync.bfly.b32 %r495|%p31, %r493, %r477, %r469, %r471;
mov.b32 %f170, %r495;
add.ftz.f32 %f171, %f169, %f170;
mov.b32 %r496, %f171;
shfl.sync.bfly.b32 %r498|%p32, %r496, %r480, %r469, %r471;
mov.b32 %f172, %r498;
add.ftz.f32 %f173, %f171, %f172;
mov.b32 %r499, %f173;
shfl.sync.bfly.b32 %r501|%p33, %r499, %r483, %r469, %r471;
mov.b32 %f174, %r501;
add.ftz.f32 %f34, %f173, %f174;
@%p22 bra $L__BB194_14;

add.s32 %r502, %r30, 1;
cvt.u64.u32 %rd78, %r502;
add.s64 %rd79, %rd78, %rd17;
shl.b64 %rd80, %rd79, 2;
add.s64 %rd81, %rd1, %rd80;
st.global.f32 [%rd81], %f34;

$L__BB194_14:
shl.b32 %r503, %r30, 1;
mov.u32 %r504, 1;
ld.shared.f32 %f175, [%r26+512];
add.ftz.f32 %f176, %f175, %f256;
mov.b32 %r505, %f176;
mov.u32 %r506, 31;
mov.u32 %r507, 16;
mov.u32 %r508, -1;
shfl.sync.bfly.b32 %r509|%p35, %r505, %r507, %r506, %r508;
mov.b32 %f177, %r509;
add.ftz.f32 %f178, %f176, %f177;
mov.b32 %r510, %f178;
mov.u32 %r511, 8;
shfl.sync.bfly.b32 %r512|%p36, %r510, %r511, %r506, %r508;
mov.b32 %f179, %r512;
add.ftz.f32 %f180, %f178, %f179;
mov.b32 %r513, %f180;
mov.u32 %r514, 4;
shfl.sync.bfly.b32 %r515|%p37, %r513, %r514, %r506, %r508;
mov.b32 %f181, %r515;
add.ftz.f32 %f182, %f180, %f181;
mov.b32 %r516, %f182;
mov.u32 %r517, 2;
shfl.sync.bfly.b32 %r518|%p38, %r516, %r517, %r506, %r508;
mov.b32 %f183, %r518;
add.ftz.f32 %f184, %f182, %f183;
mov.b32 %r519, %f184;
shfl.sync.bfly.b32 %r520|%p39, %r519, %r504, %r506, %r508;
mov.b32 %f185, %r520;
add.ftz.f32 %f35, %f184, %f185;
cvt.u64.u32 %rd82, %r503;
add.s64 %rd83, %rd82, %rd17;
shl.b64 %rd84, %rd83, 2;
add.s64 %rd19, %rd1, %rd84;
@%p14 bra $L__BB194_16;

st.global.f32 [%rd19], %f35;

$L__BB194_16:
ld.shared.f32 %f186, [%r26+640];
add.ftz.f32 %f187, %f186, %f255;
mov.b32 %r521, %f187;
shfl.sync.bfly.b32 %r525|%p41, %r521, %r507, %r506, %r508;
mov.b32 %f188, %r525;
add.ftz.f32 %f189, %f187, %f188;
mov.b32 %r526, %f189;
shfl.sync.bfly.b32 %r528|%p42, %r526, %r511, %r506, %r508;
mov.b32 %f190, %r528;
add.ftz.f32 %f191, %f189, %f190;
mov.b32 %r529, %f191;
shfl.sync.bfly.b32 %r531|%p43, %r529, %r514, %r506, %r508;
mov.b32 %f192, %r531;
add.ftz.f32 %f193, %f191, %f192;
mov.b32 %r532, %f193;
shfl.sync.bfly.b32 %r534|%p44, %r532, %r517, %r506, %r508;
mov.b32 %f194, %r534;
add.ftz.f32 %f195, %f193, %f194;
mov.b32 %r535, %f195;
shfl.sync.bfly.b32 %r537|%p45, %r535, %r504, %r506, %r508;
mov.b32 %f196, %r537;
add.ftz.f32 %f36, %f195, %f196;
@%p22 bra $L__BB194_18;

st.global.f32 [%rd19+4], %f36;

$L__BB194_18:
mul.lo.s32 %r27, %r30, 3;
ld.shared.f32 %f197, [%r26+768];
add.ftz.f32 %f198, %f197, %f254;
mov.b32 %r538, %f198;
mov.u32 %r539, 31;
mov.u32 %r540, 16;
mov.u32 %r541, -1;
shfl.sync.bfly.b32 %r542|%p47, %r538, %r540, %r539, %r541;
mov.b32 %f199, %r542;
add.ftz.f32 %f200, %f198, %f199;
mov.b32 %r543, %f200;
mov.u32 %r544, 8;
shfl.sync.bfly.b32 %r545|%p48, %r543, %r544, %r539, %r541;
mov.b32 %f201, %r545;
add.ftz.f32 %f202, %f200, %f201;
mov.b32 %r546, %f202;
mov.u32 %r547, 4;
shfl.sync.bfly.b32 %r548|%p49, %r546, %r547, %r539, %r541;
mov.b32 %f203, %r548;
add.ftz.f32 %f204, %f202, %f203;
mov.b32 %r549, %f204;
mov.u32 %r550, 2;
shfl.sync.bfly.b32 %r551|%p50, %r549, %r550, %r539, %r541;
mov.b32 %f205, %r551;
add.ftz.f32 %f206, %f204, %f205;
mov.b32 %r552, %f206;
mov.u32 %r553, 1;
shfl.sync.bfly.b32 %r554|%p51, %r552, %r553, %r539, %r541;
mov.b32 %f207, %r554;
add.ftz.f32 %f37, %f206, %f207;
@%p14 bra $L__BB194_20;

cvt.u64.u32 %rd85, %r27;
add.s64 %rd86, %rd85, %rd17;
shl.b64 %rd87, %rd86, 2;
add.s64 %rd88, %rd1, %rd87;
st.global.f32 [%rd88], %f37;

$L__BB194_20:
ld.shared.f32 %f208, [%r26+896];
add.ftz.f32 %f209, %f208, %f253;
mov.b32 %r555, %f209;
shfl.sync.bfly.b32 %r559|%p53, %r555, %r540, %r539, %r541;
mov.b32 %f210, %r559;
add.ftz.f32 %f211, %f209, %f210;
mov.b32 %r560, %f211;
shfl.sync.bfly.b32 %r562|%p54, %r560, %r544, %r539, %r541;
mov.b32 %f212, %r562;
add.ftz.f32 %f213, %f211, %f212;
mov.b32 %r563, %f213;
shfl.sync.bfly.b32 %r565|%p55, %r563, %r547, %r539, %r541;
mov.b32 %f214, %r565;
add.ftz.f32 %f215, %f213, %f214;
mov.b32 %r566, %f215;
shfl.sync.bfly.b32 %r568|%p56, %r566, %r550, %r539, %r541;
mov.b32 %f216, %r568;
add.ftz.f32 %f217, %f215, %f216;
mov.b32 %r569, %f217;
shfl.sync.bfly.b32 %r571|%p57, %r569, %r553, %r539, %r541;
mov.b32 %f218, %r571;
add.ftz.f32 %f38, %f217, %f218;
@%p22 bra $L__BB194_22;

add.s32 %r572, %r27, 1;
cvt.u64.u32 %rd89, %r572;
add.s64 %rd90, %rd89, %rd17;
shl.b64 %rd91, %rd90, 2;
add.s64 %rd92, %rd1, %rd91;
st.global.f32 [%rd92], %f38;

$L__BB194_22:
shl.b32 %r573, %r30, 2;
mov.u32 %r574, 2;
ld.shared.f32 %f219, [%r26+1024];
add.ftz.f32 %f220, %f219, %f252;
mov.b32 %r575, %f220;
mov.u32 %r576, 31;
mov.u32 %r577, 16;
mov.u32 %r578, -1;
shfl.sync.bfly.b32 %r579|%p59, %r575, %r577, %r576, %r578;
mov.b32 %f221, %r579;
add.ftz.f32 %f222, %f220, %f221;
mov.b32 %r580, %f222;
mov.u32 %r581, 8;
shfl.sync.bfly.b32 %r582|%p60, %r580, %r581, %r576, %r578;
mov.b32 %f223, %r582;
add.ftz.f32 %f224, %f222, %f223;
mov.b32 %r583, %f224;
mov.u32 %r584, 4;
shfl.sync.bfly.b32 %r585|%p61, %r583, %r584, %r576, %r578;
mov.b32 %f225, %r585;
add.ftz.f32 %f226, %f224, %f225;
mov.b32 %r586, %f226;
shfl.sync.bfly.b32 %r587|%p62, %r586, %r574, %r576, %r578;
mov.b32 %f227, %r587;
add.ftz.f32 %f228, %f226, %f227;
mov.b32 %r588, %f228;
mov.u32 %r589, 1;
shfl.sync.bfly.b32 %r590|%p63, %r588, %r589, %r576, %r578;
mov.b32 %f229, %r590;
add.ftz.f32 %f39, %f228, %f229;
cvt.u64.u32 %rd93, %r573;
add.s64 %rd94, %rd93, %rd17;
shl.b64 %rd95, %rd94, 2;
add.s64 %rd20, %rd1, %rd95;
@%p14 bra $L__BB194_24;

st.global.f32 [%rd20], %f39;

$L__BB194_24:
ld.shared.f32 %f230, [%r26+1152];
add.ftz.f32 %f231, %f230, %f251;
mov.b32 %r591, %f231;
shfl.sync.bfly.b32 %r595|%p65, %r591, %r577, %r576, %r578;
mov.b32 %f232, %r595;
add.ftz.f32 %f233, %f231, %f232;
mov.b32 %r596, %f233;
shfl.sync.bfly.b32 %r598|%p66, %r596, %r581, %r576, %r578;
mov.b32 %f234, %r598;
add.ftz.f32 %f235, %f233, %f234;
mov.b32 %r599, %f235;
shfl.sync.bfly.b32 %r601|%p67, %r599, %r584, %r576, %r578;
mov.b32 %f236, %r601;
add.ftz.f32 %f237, %f235, %f236;
mov.b32 %r602, %f237;
shfl.sync.bfly.b32 %r604|%p68, %r602, %r574, %r576, %r578;
mov.b32 %f238, %r604;
add.ftz.f32 %f239, %f237, %f238;
mov.b32 %r605, %f239;
shfl.sync.bfly.b32 %r607|%p69, %r605, %r589, %r576, %r578;
mov.b32 %f240, %r607;
add.ftz.f32 %f40, %f239, %f240;
@%p22 bra $L__BB194_26;

st.global.f32 [%rd20+4], %f40;

$L__BB194_26:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type14ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type14ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type14ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type14ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type14ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type14ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type14ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type14ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type14ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<83>;
.reg .b16 %rs<25>;
.reg .f32 %f<313>;
.reg .b32 %r<710>;
.reg .b64 %rd<111>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type14ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

ld.param.u64 %rd21, [_Z13mul_mat_vec_qIL9ggml_type14ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd22, [_Z13mul_mat_vec_qIL9ggml_type14ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd23, [_Z13mul_mat_vec_qIL9ggml_type14ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r47, [_Z13mul_mat_vec_qIL9ggml_type14ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type14ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type14ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type14ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type14ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type14ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type14ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type14ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r45, [_Z13mul_mat_vec_qIL9ggml_type14ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r46, [_Z13mul_mat_vec_qIL9ggml_type14ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type14ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type14ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type14ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type14ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
cvta.to.global.u64 %rd1, %rd23;
mov.u32 %r1, %tid.y;
shl.b32 %r48, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r48, %r2;
mov.u32 %r49, %ctaid.x;
shl.b32 %r4, %r49, 1;
mov.u32 %r5, %ctaid.y;
shr.s32 %r50, %r3, 31;
shr.u32 %r51, %r50, 27;
add.s32 %r52, %r3, %r51;
shr.s32 %r709, %r52, 5;
shr.u32 %r7, %r47, 8;
setp.ge.s32 %p3, %r709, %r7;
mov.f32 %f301, 0f00000000;
mov.f32 %f302, %f301;
mov.f32 %f303, %f301;
mov.f32 %f304, %f301;
mov.f32 %f305, %f301;
mov.f32 %f306, %f301;
mov.f32 %f307, %f301;
mov.f32 %f308, %f301;
mov.f32 %f309, %f301;
mov.f32 %f310, %f301;
mov.f32 %f311, %f301;
mov.f32 %f312, %f301;
@%p3 bra $L__BB195_3;

mov.u32 %r53, %ctaid.z;
mul.lo.s32 %r54, %r5, %r39;
cvt.u64.u32 %rd24, %r54;
mul.lo.s32 %r55, %r53, %r45;
cvt.u64.u32 %rd25, %r55;
add.s64 %rd2, %rd25, %rd24;
mul.hi.u32 %r56, %r53, %r41;
add.s32 %r57, %r56, %r53;
shr.u32 %r58, %r57, %r42;
mul.lo.s32 %r59, %r58, %r44;
mul.hi.u32 %r60, %r5, %r35;
add.s32 %r61, %r60, %r5;
shr.u32 %r62, %r61, %r36;
mul.lo.s32 %r63, %r62, %r38;
cvta.to.global.u64 %rd110, %rd21;
add.s32 %r64, %r4, 1;
and.b32 %r68, %r52, -32;
sub.s32 %r69, %r3, %r68;
shr.s32 %r70, %r69, 31;
shr.u32 %r71, %r70, 28;
add.s32 %r72, %r69, %r71;
shr.s32 %r73, %r72, 4;
shl.b32 %r74, %r73, 3;
and.b32 %r75, %r72, -16;
sub.s32 %r76, %r69, %r75;
shr.s32 %r77, %r76, 31;
shr.u32 %r78, %r77, 30;
add.s32 %r79, %r76, %r78;
shr.s32 %r80, %r79, 2;
add.s32 %r81, %r74, %r80;
cvt.s64.s32 %rd26, %r81;
shl.b32 %r82, %r73, 2;
shr.u32 %r83, %r77, 29;
add.s32 %r84, %r76, %r83;
shr.s32 %r85, %r84, 3;
add.s32 %r86, %r82, %r85;
shl.b32 %r8, %r85, 1;
shl.b32 %r87, %r69, 1;
shr.u32 %r88, %r70, 29;
add.s32 %r89, %r69, %r88;
and.b32 %r90, %r89, -8;
sub.s32 %r91, %r69, %r90;
add.s32 %r92, %r74, %r91;
shl.b32 %r93, %r92, 1;
cvt.s64.s32 %rd4, %r91;
cvt.s64.s32 %rd5, %r86;
shl.b32 %r94, %r33, 1;
shl.b32 %r703, %r709, 3;
add.s32 %r708, %r703, %r94;
mad.lo.s32 %r707, %r33, 3, %r703;
shl.b32 %r95, %r33, 2;
add.s32 %r706, %r703, %r95;
mad.lo.s32 %r705, %r33, 5, %r703;
add.s32 %r704, %r33, %r703;
mad.lo.s32 %r96, %r4, %r32, %r709;
add.s32 %r97, %r96, %r59;
add.s32 %r98, %r97, %r63;
mul.wide.s32 %rd6, %r98, 210;
mul.wide.s32 %rd27, %r87, 2;
add.s64 %rd28, %rd6, %rd27;
add.s64 %rd7, %rd28, 2;
add.s32 %r99, %r709, %r59;
add.s32 %r100, %r99, %r63;
mad.lo.s32 %r101, %r64, %r32, %r100;
mul.wide.s32 %rd8, %r101, 210;
add.s64 %rd29, %rd8, %rd26;
add.s64 %rd9, %rd29, 192;
mul.wide.s32 %rd30, %r93, 2;
add.s64 %rd31, %rd6, %rd30;
add.s64 %rd10, %rd31, 130;
add.s64 %rd32, %rd8, %rd30;
add.s64 %rd11, %rd32, 128;
add.s64 %rd12, %rd8, %rd27;
add.s64 %rd33, %rd6, %rd26;
add.s64 %rd13, %rd33, 192;
cvta.to.global.u64 %rd14, %rd22;
mov.f32 %f301, 0f00000000;
shl.b64 %rd41, %rd4, 2;

$L__BB195_2:
cvt.u64.u32 %rd34, %r703;
add.s64 %rd35, %rd2, %rd34;
add.s64 %rd36, %rd110, %rd7;
ld.global.nc.u16 %rs13, [%rd36+-2];
ld.global.nc.u16 %rs14, [%rd36];
mov.b32 %r318, {%rs13, %rs14};
add.s64 %rd37, %rd110, %rd10;
ld.global.nc.u16 %rs15, [%rd37+-2];
ld.global.nc.u16 %rs16, [%rd37];
mov.b32 %r319, {%rs15, %rs16};
shr.s32 %r320, %r319, %r8;
add.s64 %rd38, %rd35, %rd5;
mul.lo.s64 %rd39, %rd38, 36;
add.s64 %rd40, %rd14, %rd39;
add.s64 %rd42, %rd40, %rd41;
ld.global.nc.u32 %r110, [%rd42+4];
ld.global.nc.u32 %r102, [%rd40];

{.reg .f16 low,high;
mov.b32 {low,high},%r102;
cvt.f32.f16 %f73, low;}


ld.global.nc.u32 %r118, [%rd42+76];
ld.global.nc.u32 %r103, [%rd40+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r103;
cvt.f32.f16 %f74, low;}


add.s64 %rd43, %rd110, %rd6;
ld.global.nc.u16 %rs1, [%rd43+208];

{ cvt.f32.f16 %f75, %rs1;}


add.s64 %rd44, %rd110, %rd13;
ld.global.nc.u8 %rs17, [%rd44];
cvt.u32.u16 %r321, %rs17;
cvt.s32.s8 %r322, %r321;
and.b32 %r323, %r318, 252645135;
shl.b32 %r324, %r320, 4;
and.b32 %r325, %r324, 808464432;
or.b32 %r286, %r325, %r323;
or.b32 %r326, %r286, -2139062144;
add.s32 %r327, %r326, -538976288;
xor.b32 %r328, %r327, -2139062144;
xor.b32 %r329, %r286, 538976288;
xor.b32 %r330, %r327, %r286;
and.b32 %r284, %r330, %r329;
mov.u32 %r104, %r284;

prmt.b32 %r104, %r104, 0, 0xba98;

mov.u32 %r106, %r286;

prmt.b32 %r106, %r106, 0, 0xba98;

xor.b32 %r331, %r106, 2139062143;
not.b32 %r332, %r104;
and.b32 %r333, %r328, %r332;
and.b32 %r334, %r331, %r104;
or.b32 %r109, %r334, %r333;
mov.u32 %r317, 0;

dp4a.s32.s32 %r108, %r109, %r110, %r317;

mul.lo.s32 %r335, %r108, %r322;
cvt.rn.f32.s32 %f109, %r335;
fma.rn.ftz.f32 %f110, %f73, %f109, 0f00000000;
ld.global.nc.u8 %rs18, [%rd44+4];
cvt.u32.u16 %r336, %rs18;
cvt.s32.s8 %r337, %r336;
shr.u32 %r338, %r318, 4;
and.b32 %r339, %r338, 252645135;
and.b32 %r340, %r320, 808464432;
or.b32 %r294, %r340, %r339;
or.b32 %r341, %r294, -2139062144;
add.s32 %r342, %r341, -538976288;
xor.b32 %r343, %r342, -2139062144;
xor.b32 %r344, %r294, 538976288;
xor.b32 %r345, %r342, %r294;
and.b32 %r292, %r345, %r344;
mov.u32 %r112, %r292;

prmt.b32 %r112, %r112, 0, 0xba98;

mov.u32 %r114, %r294;

prmt.b32 %r114, %r114, 0, 0xba98;

xor.b32 %r346, %r114, 2139062143;
not.b32 %r347, %r112;
and.b32 %r348, %r343, %r347;
and.b32 %r349, %r346, %r112;
or.b32 %r117, %r349, %r348;

dp4a.s32.s32 %r116, %r117, %r118, %r317;

mul.lo.s32 %r350, %r116, %r337;
cvt.rn.f32.s32 %f111, %r350;
fma.rn.ftz.f32 %f112, %f74, %f111, %f110;
fma.rn.ftz.f32 %f312, %f75, %f112, %f312;
add.s64 %rd45, %rd110, %rd12;
ld.global.nc.u16 %rs19, [%rd45];
ld.global.nc.u16 %rs20, [%rd45+2];
mov.b32 %r351, {%rs19, %rs20};
add.s64 %rd46, %rd110, %rd11;
ld.global.nc.u16 %rs21, [%rd46];
ld.global.nc.u16 %rs22, [%rd46+2];
mov.b32 %r352, {%rs21, %rs22};
shr.s32 %r353, %r352, %r8;

{.reg .f16 low,high;
mov.b32 {low,high},%r102;
cvt.f32.f16 %f76, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r103;
cvt.f32.f16 %f77, low;}


add.s64 %rd47, %rd110, %rd8;
ld.global.nc.u16 %rs2, [%rd47+208];

{ cvt.f32.f16 %f78, %rs2;}


add.s64 %rd48, %rd110, %rd9;
ld.global.nc.u8 %rs23, [%rd48];
cvt.u32.u16 %r354, %rs23;
cvt.s32.s8 %r355, %r354;
and.b32 %r356, %r351, 252645135;
shl.b32 %r357, %r353, 4;
and.b32 %r358, %r357, 808464432;
or.b32 %r304, %r358, %r356;
or.b32 %r359, %r304, -2139062144;
add.s32 %r360, %r359, -538976288;
xor.b32 %r361, %r360, -2139062144;
xor.b32 %r362, %r304, 538976288;
xor.b32 %r363, %r360, %r304;
and.b32 %r302, %r363, %r362;
mov.u32 %r122, %r302;

prmt.b32 %r122, %r122, 0, 0xba98;

mov.u32 %r124, %r304;

prmt.b32 %r124, %r124, 0, 0xba98;

xor.b32 %r364, %r124, 2139062143;
not.b32 %r365, %r122;
and.b32 %r366, %r361, %r365;
and.b32 %r367, %r364, %r122;
or.b32 %r127, %r367, %r366;

dp4a.s32.s32 %r126, %r127, %r110, %r317;

mul.lo.s32 %r368, %r126, %r355;
cvt.rn.f32.s32 %f113, %r368;
fma.rn.ftz.f32 %f114, %f76, %f113, 0f00000000;
ld.global.nc.u8 %rs24, [%rd48+4];
cvt.u32.u16 %r369, %rs24;
cvt.s32.s8 %r370, %r369;
shr.u32 %r371, %r351, 4;
and.b32 %r372, %r371, 252645135;
and.b32 %r373, %r353, 808464432;
or.b32 %r312, %r373, %r372;
or.b32 %r374, %r312, -2139062144;
add.s32 %r375, %r374, -538976288;
xor.b32 %r376, %r375, -2139062144;
xor.b32 %r377, %r312, 538976288;
xor.b32 %r378, %r375, %r312;
and.b32 %r310, %r378, %r377;
mov.u32 %r130, %r310;

prmt.b32 %r130, %r130, 0, 0xba98;

mov.u32 %r132, %r312;

prmt.b32 %r132, %r132, 0, 0xba98;

xor.b32 %r379, %r132, 2139062143;
not.b32 %r380, %r130;
and.b32 %r381, %r376, %r380;
and.b32 %r382, %r379, %r130;
or.b32 %r135, %r382, %r381;

dp4a.s32.s32 %r134, %r135, %r118, %r317;

mul.lo.s32 %r383, %r134, %r370;
cvt.rn.f32.s32 %f115, %r383;
fma.rn.ftz.f32 %f116, %f77, %f115, %f114;
fma.rn.ftz.f32 %f311, %f78, %f116, %f311;
cvt.u64.u32 %rd49, %r704;
add.s64 %rd50, %rd2, %rd49;
add.s64 %rd51, %rd50, %rd5;
mul.lo.s64 %rd52, %rd51, 36;
add.s64 %rd53, %rd14, %rd52;
add.s64 %rd54, %rd53, %rd41;
ld.global.nc.u32 %r146, [%rd54+4];
ld.global.nc.u32 %r138, [%rd53];

{.reg .f16 low,high;
mov.b32 {low,high},%r138;
cvt.f32.f16 %f79, low;}


ld.global.nc.u32 %r154, [%rd54+76];
ld.global.nc.u32 %r139, [%rd53+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r139;
cvt.f32.f16 %f80, low;}



{ cvt.f32.f16 %f81, %rs1;}


mov.u32 %r140, %r284;

prmt.b32 %r140, %r140, 0, 0xba98;

mov.u32 %r142, %r286;

prmt.b32 %r142, %r142, 0, 0xba98;

xor.b32 %r384, %r142, 2139062143;
not.b32 %r385, %r140;
and.b32 %r386, %r328, %r385;
and.b32 %r387, %r384, %r140;
or.b32 %r145, %r387, %r386;

dp4a.s32.s32 %r144, %r145, %r146, %r317;

mul.lo.s32 %r388, %r144, %r322;
cvt.rn.f32.s32 %f117, %r388;
fma.rn.ftz.f32 %f118, %f79, %f117, 0f00000000;
mov.u32 %r148, %r292;

prmt.b32 %r148, %r148, 0, 0xba98;

mov.u32 %r150, %r294;

prmt.b32 %r150, %r150, 0, 0xba98;

xor.b32 %r389, %r150, 2139062143;
not.b32 %r390, %r148;
and.b32 %r391, %r343, %r390;
and.b32 %r392, %r389, %r148;
or.b32 %r153, %r392, %r391;

dp4a.s32.s32 %r152, %r153, %r154, %r317;

mul.lo.s32 %r393, %r152, %r337;
cvt.rn.f32.s32 %f119, %r393;
fma.rn.ftz.f32 %f120, %f80, %f119, %f118;
fma.rn.ftz.f32 %f310, %f81, %f120, %f310;

{.reg .f16 low,high;
mov.b32 {low,high},%r138;
cvt.f32.f16 %f82, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r139;
cvt.f32.f16 %f83, low;}



{ cvt.f32.f16 %f84, %rs2;}


mov.u32 %r158, %r302;

prmt.b32 %r158, %r158, 0, 0xba98;

mov.u32 %r160, %r304;

prmt.b32 %r160, %r160, 0, 0xba98;

xor.b32 %r394, %r160, 2139062143;
not.b32 %r395, %r158;
and.b32 %r396, %r361, %r395;
and.b32 %r397, %r394, %r158;
or.b32 %r163, %r397, %r396;

dp4a.s32.s32 %r162, %r163, %r146, %r317;

mul.lo.s32 %r398, %r162, %r355;
cvt.rn.f32.s32 %f121, %r398;
fma.rn.ftz.f32 %f122, %f82, %f121, 0f00000000;
mov.u32 %r166, %r310;

prmt.b32 %r166, %r166, 0, 0xba98;

mov.u32 %r168, %r312;

prmt.b32 %r168, %r168, 0, 0xba98;

xor.b32 %r399, %r168, 2139062143;
not.b32 %r400, %r166;
and.b32 %r401, %r376, %r400;
and.b32 %r402, %r399, %r166;
or.b32 %r171, %r402, %r401;

dp4a.s32.s32 %r170, %r171, %r154, %r317;

mul.lo.s32 %r403, %r170, %r370;
cvt.rn.f32.s32 %f123, %r403;
fma.rn.ftz.f32 %f124, %f83, %f123, %f122;
fma.rn.ftz.f32 %f309, %f84, %f124, %f309;
cvt.u64.u32 %rd55, %r708;
add.s64 %rd56, %rd2, %rd55;
add.s64 %rd57, %rd56, %rd5;
mul.lo.s64 %rd58, %rd57, 36;
add.s64 %rd59, %rd14, %rd58;
add.s64 %rd60, %rd59, %rd41;
ld.global.nc.u32 %r182, [%rd60+4];
ld.global.nc.u32 %r174, [%rd59];

{.reg .f16 low,high;
mov.b32 {low,high},%r174;
cvt.f32.f16 %f85, low;}


ld.global.nc.u32 %r190, [%rd60+76];
ld.global.nc.u32 %r175, [%rd59+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r175;
cvt.f32.f16 %f86, low;}



{ cvt.f32.f16 %f87, %rs1;}


mov.u32 %r176, %r284;

prmt.b32 %r176, %r176, 0, 0xba98;

mov.u32 %r178, %r286;

prmt.b32 %r178, %r178, 0, 0xba98;

xor.b32 %r404, %r178, 2139062143;
not.b32 %r405, %r176;
and.b32 %r406, %r328, %r405;
and.b32 %r407, %r404, %r176;
or.b32 %r181, %r407, %r406;

dp4a.s32.s32 %r180, %r181, %r182, %r317;

mul.lo.s32 %r408, %r180, %r322;
cvt.rn.f32.s32 %f125, %r408;
fma.rn.ftz.f32 %f126, %f85, %f125, 0f00000000;
mov.u32 %r184, %r292;

prmt.b32 %r184, %r184, 0, 0xba98;

mov.u32 %r186, %r294;

prmt.b32 %r186, %r186, 0, 0xba98;

xor.b32 %r409, %r186, 2139062143;
not.b32 %r410, %r184;
and.b32 %r411, %r343, %r410;
and.b32 %r412, %r409, %r184;
or.b32 %r189, %r412, %r411;

dp4a.s32.s32 %r188, %r189, %r190, %r317;

mul.lo.s32 %r413, %r188, %r337;
cvt.rn.f32.s32 %f127, %r413;
fma.rn.ftz.f32 %f128, %f86, %f127, %f126;
fma.rn.ftz.f32 %f308, %f87, %f128, %f308;

{.reg .f16 low,high;
mov.b32 {low,high},%r174;
cvt.f32.f16 %f88, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r175;
cvt.f32.f16 %f89, low;}



{ cvt.f32.f16 %f90, %rs2;}


mov.u32 %r194, %r302;

prmt.b32 %r194, %r194, 0, 0xba98;

mov.u32 %r196, %r304;

prmt.b32 %r196, %r196, 0, 0xba98;

xor.b32 %r414, %r196, 2139062143;
not.b32 %r415, %r194;
and.b32 %r416, %r361, %r415;
and.b32 %r417, %r414, %r194;
or.b32 %r199, %r417, %r416;

dp4a.s32.s32 %r198, %r199, %r182, %r317;

mul.lo.s32 %r418, %r198, %r355;
cvt.rn.f32.s32 %f129, %r418;
fma.rn.ftz.f32 %f130, %f88, %f129, 0f00000000;
mov.u32 %r202, %r310;

prmt.b32 %r202, %r202, 0, 0xba98;

mov.u32 %r204, %r312;

prmt.b32 %r204, %r204, 0, 0xba98;

xor.b32 %r419, %r204, 2139062143;
not.b32 %r420, %r202;
and.b32 %r421, %r376, %r420;
and.b32 %r422, %r419, %r202;
or.b32 %r207, %r422, %r421;

dp4a.s32.s32 %r206, %r207, %r190, %r317;

mul.lo.s32 %r423, %r206, %r370;
cvt.rn.f32.s32 %f131, %r423;
fma.rn.ftz.f32 %f132, %f89, %f131, %f130;
fma.rn.ftz.f32 %f307, %f90, %f132, %f307;
cvt.u64.u32 %rd61, %r707;
add.s64 %rd62, %rd2, %rd61;
add.s64 %rd63, %rd62, %rd5;
mul.lo.s64 %rd64, %rd63, 36;
add.s64 %rd65, %rd14, %rd64;
add.s64 %rd66, %rd65, %rd41;
ld.global.nc.u32 %r218, [%rd66+4];
ld.global.nc.u32 %r210, [%rd65];

{.reg .f16 low,high;
mov.b32 {low,high},%r210;
cvt.f32.f16 %f91, low;}


ld.global.nc.u32 %r226, [%rd66+76];
ld.global.nc.u32 %r211, [%rd65+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r211;
cvt.f32.f16 %f92, low;}



{ cvt.f32.f16 %f93, %rs1;}


mov.u32 %r212, %r284;

prmt.b32 %r212, %r212, 0, 0xba98;

mov.u32 %r214, %r286;

prmt.b32 %r214, %r214, 0, 0xba98;

xor.b32 %r424, %r214, 2139062143;
not.b32 %r425, %r212;
and.b32 %r426, %r328, %r425;
and.b32 %r427, %r424, %r212;
or.b32 %r217, %r427, %r426;

dp4a.s32.s32 %r216, %r217, %r218, %r317;

mul.lo.s32 %r428, %r216, %r322;
cvt.rn.f32.s32 %f133, %r428;
fma.rn.ftz.f32 %f134, %f91, %f133, 0f00000000;
mov.u32 %r220, %r292;

prmt.b32 %r220, %r220, 0, 0xba98;

mov.u32 %r222, %r294;

prmt.b32 %r222, %r222, 0, 0xba98;

xor.b32 %r429, %r222, 2139062143;
not.b32 %r430, %r220;
and.b32 %r431, %r343, %r430;
and.b32 %r432, %r429, %r220;
or.b32 %r225, %r432, %r431;

dp4a.s32.s32 %r224, %r225, %r226, %r317;

mul.lo.s32 %r433, %r224, %r337;
cvt.rn.f32.s32 %f135, %r433;
fma.rn.ftz.f32 %f136, %f92, %f135, %f134;
fma.rn.ftz.f32 %f306, %f93, %f136, %f306;

{.reg .f16 low,high;
mov.b32 {low,high},%r210;
cvt.f32.f16 %f94, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r211;
cvt.f32.f16 %f95, low;}



{ cvt.f32.f16 %f96, %rs2;}


mov.u32 %r230, %r302;

prmt.b32 %r230, %r230, 0, 0xba98;

mov.u32 %r232, %r304;

prmt.b32 %r232, %r232, 0, 0xba98;

xor.b32 %r434, %r232, 2139062143;
not.b32 %r435, %r230;
and.b32 %r436, %r361, %r435;
and.b32 %r437, %r434, %r230;
or.b32 %r235, %r437, %r436;

dp4a.s32.s32 %r234, %r235, %r218, %r317;

mul.lo.s32 %r438, %r234, %r355;
cvt.rn.f32.s32 %f137, %r438;
fma.rn.ftz.f32 %f138, %f94, %f137, 0f00000000;
mov.u32 %r238, %r310;

prmt.b32 %r238, %r238, 0, 0xba98;

mov.u32 %r240, %r312;

prmt.b32 %r240, %r240, 0, 0xba98;

xor.b32 %r439, %r240, 2139062143;
not.b32 %r440, %r238;
and.b32 %r441, %r376, %r440;
and.b32 %r442, %r439, %r238;
or.b32 %r243, %r442, %r441;

dp4a.s32.s32 %r242, %r243, %r226, %r317;

mul.lo.s32 %r443, %r242, %r370;
cvt.rn.f32.s32 %f139, %r443;
fma.rn.ftz.f32 %f140, %f95, %f139, %f138;
fma.rn.ftz.f32 %f305, %f96, %f140, %f305;
cvt.u64.u32 %rd67, %r706;
add.s64 %rd68, %rd2, %rd67;
add.s64 %rd69, %rd68, %rd5;
mul.lo.s64 %rd70, %rd69, 36;
add.s64 %rd71, %rd14, %rd70;
add.s64 %rd72, %rd71, %rd41;
ld.global.nc.u32 %r254, [%rd72+4];
ld.global.nc.u32 %r246, [%rd71];

{.reg .f16 low,high;
mov.b32 {low,high},%r246;
cvt.f32.f16 %f97, low;}


ld.global.nc.u32 %r262, [%rd72+76];
ld.global.nc.u32 %r247, [%rd71+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r247;
cvt.f32.f16 %f98, low;}



{ cvt.f32.f16 %f99, %rs1;}


mov.u32 %r248, %r284;

prmt.b32 %r248, %r248, 0, 0xba98;

mov.u32 %r250, %r286;

prmt.b32 %r250, %r250, 0, 0xba98;

xor.b32 %r444, %r250, 2139062143;
not.b32 %r445, %r248;
and.b32 %r446, %r328, %r445;
and.b32 %r447, %r444, %r248;
or.b32 %r253, %r447, %r446;

dp4a.s32.s32 %r252, %r253, %r254, %r317;

mul.lo.s32 %r448, %r252, %r322;
cvt.rn.f32.s32 %f141, %r448;
fma.rn.ftz.f32 %f142, %f97, %f141, 0f00000000;
mov.u32 %r256, %r292;

prmt.b32 %r256, %r256, 0, 0xba98;

mov.u32 %r258, %r294;

prmt.b32 %r258, %r258, 0, 0xba98;

xor.b32 %r449, %r258, 2139062143;
not.b32 %r450, %r256;
and.b32 %r451, %r343, %r450;
and.b32 %r452, %r449, %r256;
or.b32 %r261, %r452, %r451;

dp4a.s32.s32 %r260, %r261, %r262, %r317;

mul.lo.s32 %r453, %r260, %r337;
cvt.rn.f32.s32 %f143, %r453;
fma.rn.ftz.f32 %f144, %f98, %f143, %f142;
fma.rn.ftz.f32 %f304, %f99, %f144, %f304;

{.reg .f16 low,high;
mov.b32 {low,high},%r246;
cvt.f32.f16 %f100, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r247;
cvt.f32.f16 %f101, low;}



{ cvt.f32.f16 %f102, %rs2;}


mov.u32 %r266, %r302;

prmt.b32 %r266, %r266, 0, 0xba98;

mov.u32 %r268, %r304;

prmt.b32 %r268, %r268, 0, 0xba98;

xor.b32 %r454, %r268, 2139062143;
not.b32 %r455, %r266;
and.b32 %r456, %r361, %r455;
and.b32 %r457, %r454, %r266;
or.b32 %r271, %r457, %r456;

dp4a.s32.s32 %r270, %r271, %r254, %r317;

mul.lo.s32 %r458, %r270, %r355;
cvt.rn.f32.s32 %f145, %r458;
fma.rn.ftz.f32 %f146, %f100, %f145, 0f00000000;
mov.u32 %r274, %r310;

prmt.b32 %r274, %r274, 0, 0xba98;

mov.u32 %r276, %r312;

prmt.b32 %r276, %r276, 0, 0xba98;

xor.b32 %r459, %r276, 2139062143;
not.b32 %r460, %r274;
and.b32 %r461, %r376, %r460;
and.b32 %r462, %r459, %r274;
or.b32 %r279, %r462, %r461;

dp4a.s32.s32 %r278, %r279, %r262, %r317;

mul.lo.s32 %r463, %r278, %r370;
cvt.rn.f32.s32 %f147, %r463;
fma.rn.ftz.f32 %f148, %f101, %f147, %f146;
fma.rn.ftz.f32 %f303, %f102, %f148, %f303;
cvt.u64.u32 %rd73, %r705;
add.s64 %rd74, %rd2, %rd73;
add.s64 %rd75, %rd74, %rd5;
mul.lo.s64 %rd76, %rd75, 36;
add.s64 %rd77, %rd14, %rd76;
add.s64 %rd78, %rd77, %rd41;
ld.global.nc.u32 %r290, [%rd78+4];
ld.global.nc.u32 %r282, [%rd77];

{.reg .f16 low,high;
mov.b32 {low,high},%r282;
cvt.f32.f16 %f103, low;}


ld.global.nc.u32 %r298, [%rd78+76];
ld.global.nc.u32 %r283, [%rd77+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r283;
cvt.f32.f16 %f104, low;}



{ cvt.f32.f16 %f105, %rs1;}



prmt.b32 %r284, %r284, 0, 0xba98;


prmt.b32 %r286, %r286, 0, 0xba98;

xor.b32 %r464, %r286, 2139062143;
not.b32 %r465, %r284;
and.b32 %r466, %r328, %r465;
and.b32 %r467, %r464, %r284;
or.b32 %r289, %r467, %r466;

dp4a.s32.s32 %r288, %r289, %r290, %r317;

mul.lo.s32 %r468, %r288, %r322;
cvt.rn.f32.s32 %f149, %r468;
fma.rn.ftz.f32 %f150, %f103, %f149, 0f00000000;

prmt.b32 %r292, %r292, 0, 0xba98;


prmt.b32 %r294, %r294, 0, 0xba98;

xor.b32 %r469, %r294, 2139062143;
not.b32 %r470, %r292;
and.b32 %r471, %r343, %r470;
and.b32 %r472, %r469, %r292;
or.b32 %r297, %r472, %r471;

dp4a.s32.s32 %r296, %r297, %r298, %r317;

mul.lo.s32 %r473, %r296, %r337;
cvt.rn.f32.s32 %f151, %r473;
fma.rn.ftz.f32 %f152, %f104, %f151, %f150;
fma.rn.ftz.f32 %f302, %f105, %f152, %f302;

{.reg .f16 low,high;
mov.b32 {low,high},%r282;
cvt.f32.f16 %f106, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r283;
cvt.f32.f16 %f107, low;}



{ cvt.f32.f16 %f108, %rs2;}



prmt.b32 %r302, %r302, 0, 0xba98;


prmt.b32 %r304, %r304, 0, 0xba98;

xor.b32 %r474, %r304, 2139062143;
not.b32 %r475, %r302;
and.b32 %r476, %r361, %r475;
and.b32 %r477, %r474, %r302;
or.b32 %r307, %r477, %r476;

dp4a.s32.s32 %r306, %r307, %r290, %r317;

mul.lo.s32 %r478, %r306, %r355;
cvt.rn.f32.s32 %f153, %r478;
fma.rn.ftz.f32 %f154, %f106, %f153, 0f00000000;

prmt.b32 %r310, %r310, 0, 0xba98;


prmt.b32 %r312, %r312, 0, 0xba98;

xor.b32 %r479, %r312, 2139062143;
not.b32 %r480, %r310;
and.b32 %r481, %r376, %r480;
and.b32 %r482, %r479, %r310;
or.b32 %r315, %r482, %r481;

dp4a.s32.s32 %r314, %r315, %r298, %r317;

mul.lo.s32 %r483, %r314, %r370;
cvt.rn.f32.s32 %f155, %r483;
fma.rn.ftz.f32 %f156, %f107, %f155, %f154;
fma.rn.ftz.f32 %f301, %f108, %f156, %f301;
add.s32 %r708, %r708, 16;
add.s32 %r707, %r707, 16;
add.s32 %r706, %r706, 16;
add.s32 %r705, %r705, 16;
add.s32 %r704, %r704, 16;
add.s32 %r703, %r703, 16;
add.s64 %rd110, %rd110, 420;
add.s32 %r709, %r709, 2;
setp.lt.s32 %p4, %r709, %r7;
@%p4 bra $L__BB195_2;

$L__BB195_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB195_5;

mov.u32 %r484, _ZZ13mul_mat_vec_qIL9ggml_type14ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r485, %r1, 1536, %r484;
shl.b32 %r486, %r2, 2;
add.s32 %r487, %r485, %r486;
st.shared.f32 [%r487+-1536], %f312;
st.shared.f32 [%r487+-1408], %f311;
st.shared.f32 [%r487+-1280], %f310;
st.shared.f32 [%r487+-1152], %f309;
st.shared.f32 [%r487+-1024], %f308;
st.shared.f32 [%r487+-896], %f307;
st.shared.f32 [%r487+-768], %f306;
st.shared.f32 [%r487+-640], %f305;
st.shared.f32 [%r487+-512], %f304;
st.shared.f32 [%r487+-384], %f303;
st.shared.f32 [%r487+-256], %f302;
st.shared.f32 [%r487+-128], %f301;

$L__BB195_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB195_30;

mov.u32 %r488, %ctaid.z;
mad.lo.s32 %r489, %r5, %r40, %r4;
mad.lo.s32 %r490, %r488, %r46, %r489;
cvt.u64.u32 %rd17, %r490;
shl.b32 %r491, %r2, 2;
mov.u32 %r492, _ZZ13mul_mat_vec_qIL9ggml_type14ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r29, %r492, %r491;
mov.u32 %r493, 2;
ld.shared.f32 %f157, [%r29];
add.ftz.f32 %f158, %f157, %f312;
mov.b32 %r494, %f158;
mov.u32 %r495, 31;
mov.u32 %r496, 16;
mov.u32 %r497, -1;
shfl.sync.bfly.b32 %r498|%p7, %r494, %r496, %r495, %r497;
mov.b32 %f159, %r498;
add.ftz.f32 %f160, %f158, %f159;
mov.b32 %r499, %f160;
mov.u32 %r500, 8;
shfl.sync.bfly.b32 %r501|%p8, %r499, %r500, %r495, %r497;
mov.b32 %f161, %r501;
add.ftz.f32 %f162, %f160, %f161;
mov.b32 %r502, %f162;
mov.u32 %r503, 4;
shfl.sync.bfly.b32 %r504|%p9, %r502, %r503, %r495, %r497;
mov.b32 %f163, %r504;
add.ftz.f32 %f164, %f162, %f163;
mov.b32 %r505, %f164;
shfl.sync.bfly.b32 %r506|%p10, %r505, %r493, %r495, %r497;
mov.b32 %f165, %r506;
add.ftz.f32 %f166, %f164, %f165;
mov.b32 %r507, %f166;
mov.u32 %r508, 1;
shfl.sync.bfly.b32 %r509|%p11, %r507, %r508, %r495, %r497;
mov.b32 %f167, %r509;
add.ftz.f32 %f37, %f166, %f167;
setp.lt.u32 %p12, %r4, %r34;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
mul.wide.u32 %rd79, %r490, 4;
add.s64 %rd18, %rd1, %rd79;
not.pred %p14, %p1;
@%p14 bra $L__BB195_8;

st.global.f32 [%rd18], %f37;

$L__BB195_8:
ld.shared.f32 %f168, [%r29+128];
add.ftz.f32 %f169, %f168, %f311;
mov.b32 %r510, %f169;
shfl.sync.bfly.b32 %r514|%p15, %r510, %r496, %r495, %r497;
mov.b32 %f170, %r514;
add.ftz.f32 %f171, %f169, %f170;
mov.b32 %r515, %f171;
shfl.sync.bfly.b32 %r517|%p16, %r515, %r500, %r495, %r497;
mov.b32 %f172, %r517;
add.ftz.f32 %f173, %f171, %f172;
mov.b32 %r518, %f173;
shfl.sync.bfly.b32 %r520|%p17, %r518, %r503, %r495, %r497;
mov.b32 %f174, %r520;
add.ftz.f32 %f175, %f173, %f174;
mov.b32 %r521, %f175;
shfl.sync.bfly.b32 %r523|%p18, %r521, %r493, %r495, %r497;
mov.b32 %f176, %r523;
add.ftz.f32 %f177, %f175, %f176;
mov.b32 %r524, %f177;
shfl.sync.bfly.b32 %r526|%p19, %r524, %r508, %r495, %r497;
mov.b32 %f178, %r526;
add.ftz.f32 %f38, %f177, %f178;
add.s32 %r527, %r4, 1;
setp.lt.u32 %p20, %r527, %r34;
setp.eq.s32 %p21, %r2, 1;
and.pred %p2, %p21, %p20;
not.pred %p22, %p2;
@%p22 bra $L__BB195_10;

st.global.f32 [%rd18+4], %f38;

$L__BB195_10:
ld.shared.f32 %f179, [%r29+256];
add.ftz.f32 %f180, %f179, %f310;
mov.b32 %r528, %f180;
mov.u32 %r529, 31;
mov.u32 %r530, 16;
mov.u32 %r531, -1;
shfl.sync.bfly.b32 %r532|%p23, %r528, %r530, %r529, %r531;
mov.b32 %f181, %r532;
add.ftz.f32 %f182, %f180, %f181;
mov.b32 %r533, %f182;
mov.u32 %r534, 8;
shfl.sync.bfly.b32 %r535|%p24, %r533, %r534, %r529, %r531;
mov.b32 %f183, %r535;
add.ftz.f32 %f184, %f182, %f183;
mov.b32 %r536, %f184;
mov.u32 %r537, 4;
shfl.sync.bfly.b32 %r538|%p25, %r536, %r537, %r529, %r531;
mov.b32 %f185, %r538;
add.ftz.f32 %f186, %f184, %f185;
mov.b32 %r539, %f186;
mov.u32 %r540, 2;
shfl.sync.bfly.b32 %r541|%p26, %r539, %r540, %r529, %r531;
mov.b32 %f187, %r541;
add.ftz.f32 %f188, %f186, %f187;
mov.b32 %r542, %f188;
mov.u32 %r543, 1;
shfl.sync.bfly.b32 %r544|%p27, %r542, %r543, %r529, %r531;
mov.b32 %f189, %r544;
add.ftz.f32 %f39, %f188, %f189;
@%p14 bra $L__BB195_12;

cvt.u64.u32 %rd80, %r34;
add.s64 %rd81, %rd80, %rd17;
shl.b64 %rd82, %rd81, 2;
add.s64 %rd83, %rd1, %rd82;
st.global.f32 [%rd83], %f39;

$L__BB195_12:
ld.shared.f32 %f190, [%r29+384];
add.ftz.f32 %f191, %f190, %f309;
mov.b32 %r545, %f191;
shfl.sync.bfly.b32 %r549|%p29, %r545, %r530, %r529, %r531;
mov.b32 %f192, %r549;
add.ftz.f32 %f193, %f191, %f192;
mov.b32 %r550, %f193;
shfl.sync.bfly.b32 %r552|%p30, %r550, %r534, %r529, %r531;
mov.b32 %f194, %r552;
add.ftz.f32 %f195, %f193, %f194;
mov.b32 %r553, %f195;
shfl.sync.bfly.b32 %r555|%p31, %r553, %r537, %r529, %r531;
mov.b32 %f196, %r555;
add.ftz.f32 %f197, %f195, %f196;
mov.b32 %r556, %f197;
shfl.sync.bfly.b32 %r558|%p32, %r556, %r540, %r529, %r531;
mov.b32 %f198, %r558;
add.ftz.f32 %f199, %f197, %f198;
mov.b32 %r559, %f199;
shfl.sync.bfly.b32 %r561|%p33, %r559, %r543, %r529, %r531;
mov.b32 %f200, %r561;
add.ftz.f32 %f40, %f199, %f200;
@%p22 bra $L__BB195_14;

add.s32 %r562, %r34, 1;
cvt.u64.u32 %rd84, %r562;
add.s64 %rd85, %rd84, %rd17;
shl.b64 %rd86, %rd85, 2;
add.s64 %rd87, %rd1, %rd86;
st.global.f32 [%rd87], %f40;

$L__BB195_14:
shl.b32 %r563, %r34, 1;
mov.u32 %r564, 1;
ld.shared.f32 %f201, [%r29+512];
add.ftz.f32 %f202, %f201, %f308;
mov.b32 %r565, %f202;
mov.u32 %r566, 31;
mov.u32 %r567, 16;
mov.u32 %r568, -1;
shfl.sync.bfly.b32 %r569|%p35, %r565, %r567, %r566, %r568;
mov.b32 %f203, %r569;
add.ftz.f32 %f204, %f202, %f203;
mov.b32 %r570, %f204;
mov.u32 %r571, 8;
shfl.sync.bfly.b32 %r572|%p36, %r570, %r571, %r566, %r568;
mov.b32 %f205, %r572;
add.ftz.f32 %f206, %f204, %f205;
mov.b32 %r573, %f206;
mov.u32 %r574, 4;
shfl.sync.bfly.b32 %r575|%p37, %r573, %r574, %r566, %r568;
mov.b32 %f207, %r575;
add.ftz.f32 %f208, %f206, %f207;
mov.b32 %r576, %f208;
mov.u32 %r577, 2;
shfl.sync.bfly.b32 %r578|%p38, %r576, %r577, %r566, %r568;
mov.b32 %f209, %r578;
add.ftz.f32 %f210, %f208, %f209;
mov.b32 %r579, %f210;
shfl.sync.bfly.b32 %r580|%p39, %r579, %r564, %r566, %r568;
mov.b32 %f211, %r580;
add.ftz.f32 %f41, %f210, %f211;
cvt.u64.u32 %rd88, %r563;
add.s64 %rd89, %rd88, %rd17;
shl.b64 %rd90, %rd89, 2;
add.s64 %rd19, %rd1, %rd90;
@%p14 bra $L__BB195_16;

st.global.f32 [%rd19], %f41;

$L__BB195_16:
ld.shared.f32 %f212, [%r29+640];
add.ftz.f32 %f213, %f212, %f307;
mov.b32 %r581, %f213;
shfl.sync.bfly.b32 %r585|%p41, %r581, %r567, %r566, %r568;
mov.b32 %f214, %r585;
add.ftz.f32 %f215, %f213, %f214;
mov.b32 %r586, %f215;
shfl.sync.bfly.b32 %r588|%p42, %r586, %r571, %r566, %r568;
mov.b32 %f216, %r588;
add.ftz.f32 %f217, %f215, %f216;
mov.b32 %r589, %f217;
shfl.sync.bfly.b32 %r591|%p43, %r589, %r574, %r566, %r568;
mov.b32 %f218, %r591;
add.ftz.f32 %f219, %f217, %f218;
mov.b32 %r592, %f219;
shfl.sync.bfly.b32 %r594|%p44, %r592, %r577, %r566, %r568;
mov.b32 %f220, %r594;
add.ftz.f32 %f221, %f219, %f220;
mov.b32 %r595, %f221;
shfl.sync.bfly.b32 %r597|%p45, %r595, %r564, %r566, %r568;
mov.b32 %f222, %r597;
add.ftz.f32 %f42, %f221, %f222;
@%p22 bra $L__BB195_18;

st.global.f32 [%rd19+4], %f42;

$L__BB195_18:
mul.lo.s32 %r30, %r34, 3;
ld.shared.f32 %f223, [%r29+768];
add.ftz.f32 %f224, %f223, %f306;
mov.b32 %r598, %f224;
mov.u32 %r599, 31;
mov.u32 %r600, 16;
mov.u32 %r601, -1;
shfl.sync.bfly.b32 %r602|%p47, %r598, %r600, %r599, %r601;
mov.b32 %f225, %r602;
add.ftz.f32 %f226, %f224, %f225;
mov.b32 %r603, %f226;
mov.u32 %r604, 8;
shfl.sync.bfly.b32 %r605|%p48, %r603, %r604, %r599, %r601;
mov.b32 %f227, %r605;
add.ftz.f32 %f228, %f226, %f227;
mov.b32 %r606, %f228;
mov.u32 %r607, 4;
shfl.sync.bfly.b32 %r608|%p49, %r606, %r607, %r599, %r601;
mov.b32 %f229, %r608;
add.ftz.f32 %f230, %f228, %f229;
mov.b32 %r609, %f230;
mov.u32 %r610, 2;
shfl.sync.bfly.b32 %r611|%p50, %r609, %r610, %r599, %r601;
mov.b32 %f231, %r611;
add.ftz.f32 %f232, %f230, %f231;
mov.b32 %r612, %f232;
mov.u32 %r613, 1;
shfl.sync.bfly.b32 %r614|%p51, %r612, %r613, %r599, %r601;
mov.b32 %f233, %r614;
add.ftz.f32 %f43, %f232, %f233;
@%p14 bra $L__BB195_20;

cvt.u64.u32 %rd91, %r30;
add.s64 %rd92, %rd91, %rd17;
shl.b64 %rd93, %rd92, 2;
add.s64 %rd94, %rd1, %rd93;
st.global.f32 [%rd94], %f43;

$L__BB195_20:
ld.shared.f32 %f234, [%r29+896];
add.ftz.f32 %f235, %f234, %f305;
mov.b32 %r615, %f235;
shfl.sync.bfly.b32 %r619|%p53, %r615, %r600, %r599, %r601;
mov.b32 %f236, %r619;
add.ftz.f32 %f237, %f235, %f236;
mov.b32 %r620, %f237;
shfl.sync.bfly.b32 %r622|%p54, %r620, %r604, %r599, %r601;
mov.b32 %f238, %r622;
add.ftz.f32 %f239, %f237, %f238;
mov.b32 %r623, %f239;
shfl.sync.bfly.b32 %r625|%p55, %r623, %r607, %r599, %r601;
mov.b32 %f240, %r625;
add.ftz.f32 %f241, %f239, %f240;
mov.b32 %r626, %f241;
shfl.sync.bfly.b32 %r628|%p56, %r626, %r610, %r599, %r601;
mov.b32 %f242, %r628;
add.ftz.f32 %f243, %f241, %f242;
mov.b32 %r629, %f243;
shfl.sync.bfly.b32 %r631|%p57, %r629, %r613, %r599, %r601;
mov.b32 %f244, %r631;
add.ftz.f32 %f44, %f243, %f244;
@%p22 bra $L__BB195_22;

add.s32 %r632, %r30, 1;
cvt.u64.u32 %rd95, %r632;
add.s64 %rd96, %rd95, %rd17;
shl.b64 %rd97, %rd96, 2;
add.s64 %rd98, %rd1, %rd97;
st.global.f32 [%rd98], %f44;

$L__BB195_22:
shl.b32 %r633, %r34, 2;
mov.u32 %r634, 2;
ld.shared.f32 %f245, [%r29+1024];
add.ftz.f32 %f246, %f245, %f304;
mov.b32 %r635, %f246;
mov.u32 %r636, 31;
mov.u32 %r637, 16;
mov.u32 %r638, -1;
shfl.sync.bfly.b32 %r639|%p59, %r635, %r637, %r636, %r638;
mov.b32 %f247, %r639;
add.ftz.f32 %f248, %f246, %f247;
mov.b32 %r640, %f248;
mov.u32 %r641, 8;
shfl.sync.bfly.b32 %r642|%p60, %r640, %r641, %r636, %r638;
mov.b32 %f249, %r642;
add.ftz.f32 %f250, %f248, %f249;
mov.b32 %r643, %f250;
mov.u32 %r644, 4;
shfl.sync.bfly.b32 %r645|%p61, %r643, %r644, %r636, %r638;
mov.b32 %f251, %r645;
add.ftz.f32 %f252, %f250, %f251;
mov.b32 %r646, %f252;
shfl.sync.bfly.b32 %r647|%p62, %r646, %r634, %r636, %r638;
mov.b32 %f253, %r647;
add.ftz.f32 %f254, %f252, %f253;
mov.b32 %r648, %f254;
mov.u32 %r649, 1;
shfl.sync.bfly.b32 %r650|%p63, %r648, %r649, %r636, %r638;
mov.b32 %f255, %r650;
add.ftz.f32 %f45, %f254, %f255;
cvt.u64.u32 %rd99, %r633;
add.s64 %rd100, %rd99, %rd17;
shl.b64 %rd101, %rd100, 2;
add.s64 %rd20, %rd1, %rd101;
@%p14 bra $L__BB195_24;

st.global.f32 [%rd20], %f45;

$L__BB195_24:
ld.shared.f32 %f256, [%r29+1152];
add.ftz.f32 %f257, %f256, %f303;
mov.b32 %r651, %f257;
shfl.sync.bfly.b32 %r655|%p65, %r651, %r637, %r636, %r638;
mov.b32 %f258, %r655;
add.ftz.f32 %f259, %f257, %f258;
mov.b32 %r656, %f259;
shfl.sync.bfly.b32 %r658|%p66, %r656, %r641, %r636, %r638;
mov.b32 %f260, %r658;
add.ftz.f32 %f261, %f259, %f260;
mov.b32 %r659, %f261;
shfl.sync.bfly.b32 %r661|%p67, %r659, %r644, %r636, %r638;
mov.b32 %f262, %r661;
add.ftz.f32 %f263, %f261, %f262;
mov.b32 %r662, %f263;
shfl.sync.bfly.b32 %r664|%p68, %r662, %r634, %r636, %r638;
mov.b32 %f264, %r664;
add.ftz.f32 %f265, %f263, %f264;
mov.b32 %r665, %f265;
shfl.sync.bfly.b32 %r667|%p69, %r665, %r649, %r636, %r638;
mov.b32 %f266, %r667;
add.ftz.f32 %f46, %f265, %f266;
@%p22 bra $L__BB195_26;

st.global.f32 [%rd20+4], %f46;

$L__BB195_26:
mul.lo.s32 %r31, %r34, 5;
ld.shared.f32 %f267, [%r29+1280];
add.ftz.f32 %f268, %f267, %f302;
mov.b32 %r668, %f268;
mov.u32 %r669, 31;
mov.u32 %r670, 16;
mov.u32 %r671, -1;
shfl.sync.bfly.b32 %r672|%p71, %r668, %r670, %r669, %r671;
mov.b32 %f269, %r672;
add.ftz.f32 %f270, %f268, %f269;
mov.b32 %r673, %f270;
mov.u32 %r674, 8;
shfl.sync.bfly.b32 %r675|%p72, %r673, %r674, %r669, %r671;
mov.b32 %f271, %r675;
add.ftz.f32 %f272, %f270, %f271;
mov.b32 %r676, %f272;
mov.u32 %r677, 4;
shfl.sync.bfly.b32 %r678|%p73, %r676, %r677, %r669, %r671;
mov.b32 %f273, %r678;
add.ftz.f32 %f274, %f272, %f273;
mov.b32 %r679, %f274;
mov.u32 %r680, 2;
shfl.sync.bfly.b32 %r681|%p74, %r679, %r680, %r669, %r671;
mov.b32 %f275, %r681;
add.ftz.f32 %f276, %f274, %f275;
mov.b32 %r682, %f276;
mov.u32 %r683, 1;
shfl.sync.bfly.b32 %r684|%p75, %r682, %r683, %r669, %r671;
mov.b32 %f277, %r684;
add.ftz.f32 %f47, %f276, %f277;
@%p14 bra $L__BB195_28;

cvt.u64.u32 %rd102, %r31;
add.s64 %rd103, %rd102, %rd17;
shl.b64 %rd104, %rd103, 2;
add.s64 %rd105, %rd1, %rd104;
st.global.f32 [%rd105], %f47;

$L__BB195_28:
ld.shared.f32 %f278, [%r29+1408];
add.ftz.f32 %f279, %f278, %f301;
mov.b32 %r685, %f279;
shfl.sync.bfly.b32 %r689|%p77, %r685, %r670, %r669, %r671;
mov.b32 %f280, %r689;
add.ftz.f32 %f281, %f279, %f280;
mov.b32 %r690, %f281;
shfl.sync.bfly.b32 %r692|%p78, %r690, %r674, %r669, %r671;
mov.b32 %f282, %r692;
add.ftz.f32 %f283, %f281, %f282;
mov.b32 %r693, %f283;
shfl.sync.bfly.b32 %r695|%p79, %r693, %r677, %r669, %r671;
mov.b32 %f284, %r695;
add.ftz.f32 %f285, %f283, %f284;
mov.b32 %r696, %f285;
shfl.sync.bfly.b32 %r698|%p80, %r696, %r680, %r669, %r671;
mov.b32 %f286, %r698;
add.ftz.f32 %f287, %f285, %f286;
mov.b32 %r699, %f287;
shfl.sync.bfly.b32 %r701|%p81, %r699, %r683, %r669, %r671;
mov.b32 %f288, %r701;
add.ftz.f32 %f48, %f287, %f288;
@%p22 bra $L__BB195_30;

add.s32 %r702, %r31, 1;
cvt.u64.u32 %rd106, %r702;
add.s64 %rd107, %rd106, %rd17;
shl.b64 %rd108, %rd107, 2;
add.s64 %rd109, %rd1, %rd108;
st.global.f32 [%rd109], %f48;

$L__BB195_30:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type14ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type14ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type14ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type14ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type14ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type14ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type14ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type14ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type14ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<95>;
.reg .b16 %rs<27>;
.reg .f32 %f<365>;
.reg .b32 %r<808>;
.reg .b64 %rd<121>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type14ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1792];

ld.param.u64 %rd22, [_Z13mul_mat_vec_qIL9ggml_type14ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd23, [_Z13mul_mat_vec_qIL9ggml_type14ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd24, [_Z13mul_mat_vec_qIL9ggml_type14ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r50, [_Z13mul_mat_vec_qIL9ggml_type14ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type14ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type14ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type14ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type14ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type14ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type14ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r47, [_Z13mul_mat_vec_qIL9ggml_type14ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r48, [_Z13mul_mat_vec_qIL9ggml_type14ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r49, [_Z13mul_mat_vec_qIL9ggml_type14ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r45, [_Z13mul_mat_vec_qIL9ggml_type14ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type14ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type14ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type14ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
cvta.to.global.u64 %rd1, %rd24;
mov.u32 %r1, %tid.y;
shl.b32 %r51, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r51, %r2;
mov.u32 %r52, %ctaid.x;
shl.b32 %r4, %r52, 1;
mov.u32 %r5, %ctaid.y;
shr.s32 %r53, %r3, 31;
shr.u32 %r54, %r53, 27;
add.s32 %r55, %r3, %r54;
shr.s32 %r807, %r55, 5;
shr.u32 %r7, %r50, 8;
setp.ge.s32 %p3, %r807, %r7;
mov.f32 %f351, 0f00000000;
mov.f32 %f352, %f351;
mov.f32 %f353, %f351;
mov.f32 %f354, %f351;
mov.f32 %f355, %f351;
mov.f32 %f356, %f351;
mov.f32 %f357, %f351;
mov.f32 %f358, %f351;
mov.f32 %f359, %f351;
mov.f32 %f360, %f351;
mov.f32 %f361, %f351;
mov.f32 %f362, %f351;
mov.f32 %f363, %f351;
mov.f32 %f364, %f351;
@%p3 bra $L__BB196_3;

mov.u32 %r56, %ctaid.z;
mul.lo.s32 %r57, %r5, %r42;
cvt.u64.u32 %rd25, %r57;
mul.lo.s32 %r58, %r56, %r48;
cvt.u64.u32 %rd26, %r58;
add.s64 %rd2, %rd26, %rd25;
mul.hi.u32 %r59, %r56, %r44;
add.s32 %r60, %r59, %r56;
shr.u32 %r61, %r60, %r45;
mul.lo.s32 %r62, %r61, %r47;
mul.hi.u32 %r63, %r5, %r38;
add.s32 %r64, %r63, %r5;
shr.u32 %r65, %r64, %r39;
mul.lo.s32 %r66, %r65, %r41;
cvta.to.global.u64 %rd120, %rd22;
add.s32 %r67, %r4, 1;
and.b32 %r71, %r55, -32;
sub.s32 %r72, %r3, %r71;
shr.s32 %r73, %r72, 31;
shr.u32 %r74, %r73, 28;
add.s32 %r75, %r72, %r74;
shr.s32 %r76, %r75, 4;
shl.b32 %r77, %r76, 3;
and.b32 %r78, %r75, -16;
sub.s32 %r79, %r72, %r78;
shr.s32 %r80, %r79, 31;
shr.u32 %r81, %r80, 30;
add.s32 %r82, %r79, %r81;
shr.s32 %r83, %r82, 2;
add.s32 %r84, %r77, %r83;
cvt.s64.s32 %rd27, %r84;
shl.b32 %r85, %r76, 2;
shr.u32 %r86, %r80, 29;
add.s32 %r87, %r79, %r86;
shr.s32 %r88, %r87, 3;
add.s32 %r89, %r85, %r88;
shl.b32 %r8, %r88, 1;
shl.b32 %r90, %r72, 1;
shr.u32 %r91, %r73, 29;
add.s32 %r92, %r72, %r91;
and.b32 %r93, %r92, -8;
sub.s32 %r94, %r72, %r93;
add.s32 %r95, %r77, %r94;
shl.b32 %r96, %r95, 1;
cvt.s64.s32 %rd4, %r94;
cvt.s64.s32 %rd5, %r89;
shl.b32 %r97, %r36, 1;
shl.b32 %r800, %r807, 3;
add.s32 %r806, %r800, %r97;
mad.lo.s32 %r805, %r36, 3, %r800;
shl.b32 %r98, %r36, 2;
add.s32 %r804, %r800, %r98;
mad.lo.s32 %r803, %r36, 5, %r800;
mad.lo.s32 %r802, %r36, 6, %r800;
add.s32 %r801, %r36, %r800;
mad.lo.s32 %r99, %r4, %r35, %r807;
add.s32 %r100, %r99, %r62;
add.s32 %r101, %r100, %r66;
mul.wide.s32 %rd6, %r101, 210;
mul.wide.s32 %rd28, %r90, 2;
add.s64 %rd29, %rd6, %rd28;
add.s64 %rd7, %rd29, 2;
add.s32 %r102, %r807, %r62;
add.s32 %r103, %r102, %r66;
mad.lo.s32 %r104, %r67, %r35, %r103;
mul.wide.s32 %rd8, %r104, 210;
add.s64 %rd30, %rd8, %rd27;
add.s64 %rd9, %rd30, 192;
mul.wide.s32 %rd31, %r96, 2;
add.s64 %rd32, %rd6, %rd31;
add.s64 %rd10, %rd32, 130;
add.s64 %rd33, %rd8, %rd31;
add.s64 %rd11, %rd33, 128;
add.s64 %rd12, %rd8, %rd28;
add.s64 %rd34, %rd6, %rd27;
add.s64 %rd13, %rd34, 192;
cvta.to.global.u64 %rd14, %rd23;
mov.f32 %f351, 0f00000000;
shl.b64 %rd42, %rd4, 2;

$L__BB196_2:
cvt.u64.u32 %rd35, %r800;
add.s64 %rd36, %rd2, %rd35;
add.s64 %rd37, %rd120, %rd7;
ld.global.nc.u16 %rs15, [%rd37+-2];
ld.global.nc.u16 %rs16, [%rd37];
mov.b32 %r357, {%rs15, %rs16};
add.s64 %rd38, %rd120, %rd10;
ld.global.nc.u16 %rs17, [%rd38+-2];
ld.global.nc.u16 %rs18, [%rd38];
mov.b32 %r358, {%rs17, %rs18};
shr.s32 %r359, %r358, %r8;
add.s64 %rd39, %rd36, %rd5;
mul.lo.s64 %rd40, %rd39, 36;
add.s64 %rd41, %rd14, %rd40;
add.s64 %rd43, %rd41, %rd42;
ld.global.nc.u32 %r113, [%rd43+4];
ld.global.nc.u32 %r105, [%rd41];

{.reg .f16 low,high;
mov.b32 {low,high},%r105;
cvt.f32.f16 %f85, low;}


ld.global.nc.u32 %r121, [%rd43+76];
ld.global.nc.u32 %r106, [%rd41+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r106;
cvt.f32.f16 %f86, low;}


add.s64 %rd44, %rd120, %rd6;
ld.global.nc.u16 %rs1, [%rd44+208];

{ cvt.f32.f16 %f87, %rs1;}


add.s64 %rd45, %rd120, %rd13;
ld.global.nc.u8 %rs19, [%rd45];
cvt.u32.u16 %r360, %rs19;
cvt.s32.s8 %r361, %r360;
and.b32 %r362, %r357, 252645135;
shl.b32 %r363, %r359, 4;
and.b32 %r364, %r363, 808464432;
or.b32 %r325, %r364, %r362;
or.b32 %r365, %r325, -2139062144;
add.s32 %r366, %r365, -538976288;
xor.b32 %r367, %r366, -2139062144;
xor.b32 %r368, %r325, 538976288;
xor.b32 %r369, %r366, %r325;
and.b32 %r323, %r369, %r368;
mov.u32 %r107, %r323;

prmt.b32 %r107, %r107, 0, 0xba98;

mov.u32 %r109, %r325;

prmt.b32 %r109, %r109, 0, 0xba98;

xor.b32 %r370, %r109, 2139062143;
not.b32 %r371, %r107;
and.b32 %r372, %r367, %r371;
and.b32 %r373, %r370, %r107;
or.b32 %r112, %r373, %r372;
mov.u32 %r356, 0;

dp4a.s32.s32 %r111, %r112, %r113, %r356;

mul.lo.s32 %r374, %r111, %r361;
cvt.rn.f32.s32 %f127, %r374;
fma.rn.ftz.f32 %f128, %f85, %f127, 0f00000000;
ld.global.nc.u8 %rs20, [%rd45+4];
cvt.u32.u16 %r375, %rs20;
cvt.s32.s8 %r376, %r375;
shr.u32 %r377, %r357, 4;
and.b32 %r378, %r377, 252645135;
and.b32 %r379, %r359, 808464432;
or.b32 %r333, %r379, %r378;
or.b32 %r380, %r333, -2139062144;
add.s32 %r381, %r380, -538976288;
xor.b32 %r382, %r381, -2139062144;
xor.b32 %r383, %r333, 538976288;
xor.b32 %r384, %r381, %r333;
and.b32 %r331, %r384, %r383;
mov.u32 %r115, %r331;

prmt.b32 %r115, %r115, 0, 0xba98;

mov.u32 %r117, %r333;

prmt.b32 %r117, %r117, 0, 0xba98;

xor.b32 %r385, %r117, 2139062143;
not.b32 %r386, %r115;
and.b32 %r387, %r382, %r386;
and.b32 %r388, %r385, %r115;
or.b32 %r120, %r388, %r387;

dp4a.s32.s32 %r119, %r120, %r121, %r356;

mul.lo.s32 %r389, %r119, %r376;
cvt.rn.f32.s32 %f129, %r389;
fma.rn.ftz.f32 %f130, %f86, %f129, %f128;
fma.rn.ftz.f32 %f364, %f87, %f130, %f364;
add.s64 %rd46, %rd120, %rd12;
ld.global.nc.u16 %rs21, [%rd46];
ld.global.nc.u16 %rs22, [%rd46+2];
mov.b32 %r390, {%rs21, %rs22};
add.s64 %rd47, %rd120, %rd11;
ld.global.nc.u16 %rs23, [%rd47];
ld.global.nc.u16 %rs24, [%rd47+2];
mov.b32 %r391, {%rs23, %rs24};
shr.s32 %r392, %r391, %r8;

{.reg .f16 low,high;
mov.b32 {low,high},%r105;
cvt.f32.f16 %f88, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r106;
cvt.f32.f16 %f89, low;}


add.s64 %rd48, %rd120, %rd8;
ld.global.nc.u16 %rs2, [%rd48+208];

{ cvt.f32.f16 %f90, %rs2;}


add.s64 %rd49, %rd120, %rd9;
ld.global.nc.u8 %rs25, [%rd49];
cvt.u32.u16 %r393, %rs25;
cvt.s32.s8 %r394, %r393;
and.b32 %r395, %r390, 252645135;
shl.b32 %r396, %r392, 4;
and.b32 %r397, %r396, 808464432;
or.b32 %r343, %r397, %r395;
or.b32 %r398, %r343, -2139062144;
add.s32 %r399, %r398, -538976288;
xor.b32 %r400, %r399, -2139062144;
xor.b32 %r401, %r343, 538976288;
xor.b32 %r402, %r399, %r343;
and.b32 %r341, %r402, %r401;
mov.u32 %r125, %r341;

prmt.b32 %r125, %r125, 0, 0xba98;

mov.u32 %r127, %r343;

prmt.b32 %r127, %r127, 0, 0xba98;

xor.b32 %r403, %r127, 2139062143;
not.b32 %r404, %r125;
and.b32 %r405, %r400, %r404;
and.b32 %r406, %r403, %r125;
or.b32 %r130, %r406, %r405;

dp4a.s32.s32 %r129, %r130, %r113, %r356;

mul.lo.s32 %r407, %r129, %r394;
cvt.rn.f32.s32 %f131, %r407;
fma.rn.ftz.f32 %f132, %f88, %f131, 0f00000000;
ld.global.nc.u8 %rs26, [%rd49+4];
cvt.u32.u16 %r408, %rs26;
cvt.s32.s8 %r409, %r408;
shr.u32 %r410, %r390, 4;
and.b32 %r411, %r410, 252645135;
and.b32 %r412, %r392, 808464432;
or.b32 %r351, %r412, %r411;
or.b32 %r413, %r351, -2139062144;
add.s32 %r414, %r413, -538976288;
xor.b32 %r415, %r414, -2139062144;
xor.b32 %r416, %r351, 538976288;
xor.b32 %r417, %r414, %r351;
and.b32 %r349, %r417, %r416;
mov.u32 %r133, %r349;

prmt.b32 %r133, %r133, 0, 0xba98;

mov.u32 %r135, %r351;

prmt.b32 %r135, %r135, 0, 0xba98;

xor.b32 %r418, %r135, 2139062143;
not.b32 %r419, %r133;
and.b32 %r420, %r415, %r419;
and.b32 %r421, %r418, %r133;
or.b32 %r138, %r421, %r420;

dp4a.s32.s32 %r137, %r138, %r121, %r356;

mul.lo.s32 %r422, %r137, %r409;
cvt.rn.f32.s32 %f133, %r422;
fma.rn.ftz.f32 %f134, %f89, %f133, %f132;
fma.rn.ftz.f32 %f363, %f90, %f134, %f363;
cvt.u64.u32 %rd50, %r801;
add.s64 %rd51, %rd2, %rd50;
add.s64 %rd52, %rd51, %rd5;
mul.lo.s64 %rd53, %rd52, 36;
add.s64 %rd54, %rd14, %rd53;
add.s64 %rd55, %rd54, %rd42;
ld.global.nc.u32 %r149, [%rd55+4];
ld.global.nc.u32 %r141, [%rd54];

{.reg .f16 low,high;
mov.b32 {low,high},%r141;
cvt.f32.f16 %f91, low;}


ld.global.nc.u32 %r157, [%rd55+76];
ld.global.nc.u32 %r142, [%rd54+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r142;
cvt.f32.f16 %f92, low;}



{ cvt.f32.f16 %f93, %rs1;}


mov.u32 %r143, %r323;

prmt.b32 %r143, %r143, 0, 0xba98;

mov.u32 %r145, %r325;

prmt.b32 %r145, %r145, 0, 0xba98;

xor.b32 %r423, %r145, 2139062143;
not.b32 %r424, %r143;
and.b32 %r425, %r367, %r424;
and.b32 %r426, %r423, %r143;
or.b32 %r148, %r426, %r425;

dp4a.s32.s32 %r147, %r148, %r149, %r356;

mul.lo.s32 %r427, %r147, %r361;
cvt.rn.f32.s32 %f135, %r427;
fma.rn.ftz.f32 %f136, %f91, %f135, 0f00000000;
mov.u32 %r151, %r331;

prmt.b32 %r151, %r151, 0, 0xba98;

mov.u32 %r153, %r333;

prmt.b32 %r153, %r153, 0, 0xba98;

xor.b32 %r428, %r153, 2139062143;
not.b32 %r429, %r151;
and.b32 %r430, %r382, %r429;
and.b32 %r431, %r428, %r151;
or.b32 %r156, %r431, %r430;

dp4a.s32.s32 %r155, %r156, %r157, %r356;

mul.lo.s32 %r432, %r155, %r376;
cvt.rn.f32.s32 %f137, %r432;
fma.rn.ftz.f32 %f138, %f92, %f137, %f136;
fma.rn.ftz.f32 %f362, %f93, %f138, %f362;

{.reg .f16 low,high;
mov.b32 {low,high},%r141;
cvt.f32.f16 %f94, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r142;
cvt.f32.f16 %f95, low;}



{ cvt.f32.f16 %f96, %rs2;}


mov.u32 %r161, %r341;

prmt.b32 %r161, %r161, 0, 0xba98;

mov.u32 %r163, %r343;

prmt.b32 %r163, %r163, 0, 0xba98;

xor.b32 %r433, %r163, 2139062143;
not.b32 %r434, %r161;
and.b32 %r435, %r400, %r434;
and.b32 %r436, %r433, %r161;
or.b32 %r166, %r436, %r435;

dp4a.s32.s32 %r165, %r166, %r149, %r356;

mul.lo.s32 %r437, %r165, %r394;
cvt.rn.f32.s32 %f139, %r437;
fma.rn.ftz.f32 %f140, %f94, %f139, 0f00000000;
mov.u32 %r169, %r349;

prmt.b32 %r169, %r169, 0, 0xba98;

mov.u32 %r171, %r351;

prmt.b32 %r171, %r171, 0, 0xba98;

xor.b32 %r438, %r171, 2139062143;
not.b32 %r439, %r169;
and.b32 %r440, %r415, %r439;
and.b32 %r441, %r438, %r169;
or.b32 %r174, %r441, %r440;

dp4a.s32.s32 %r173, %r174, %r157, %r356;

mul.lo.s32 %r442, %r173, %r409;
cvt.rn.f32.s32 %f141, %r442;
fma.rn.ftz.f32 %f142, %f95, %f141, %f140;
fma.rn.ftz.f32 %f361, %f96, %f142, %f361;
cvt.u64.u32 %rd56, %r806;
add.s64 %rd57, %rd2, %rd56;
add.s64 %rd58, %rd57, %rd5;
mul.lo.s64 %rd59, %rd58, 36;
add.s64 %rd60, %rd14, %rd59;
add.s64 %rd61, %rd60, %rd42;
ld.global.nc.u32 %r185, [%rd61+4];
ld.global.nc.u32 %r177, [%rd60];

{.reg .f16 low,high;
mov.b32 {low,high},%r177;
cvt.f32.f16 %f97, low;}


ld.global.nc.u32 %r193, [%rd61+76];
ld.global.nc.u32 %r178, [%rd60+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r178;
cvt.f32.f16 %f98, low;}



{ cvt.f32.f16 %f99, %rs1;}


mov.u32 %r179, %r323;

prmt.b32 %r179, %r179, 0, 0xba98;

mov.u32 %r181, %r325;

prmt.b32 %r181, %r181, 0, 0xba98;

xor.b32 %r443, %r181, 2139062143;
not.b32 %r444, %r179;
and.b32 %r445, %r367, %r444;
and.b32 %r446, %r443, %r179;
or.b32 %r184, %r446, %r445;

dp4a.s32.s32 %r183, %r184, %r185, %r356;

mul.lo.s32 %r447, %r183, %r361;
cvt.rn.f32.s32 %f143, %r447;
fma.rn.ftz.f32 %f144, %f97, %f143, 0f00000000;
mov.u32 %r187, %r331;

prmt.b32 %r187, %r187, 0, 0xba98;

mov.u32 %r189, %r333;

prmt.b32 %r189, %r189, 0, 0xba98;

xor.b32 %r448, %r189, 2139062143;
not.b32 %r449, %r187;
and.b32 %r450, %r382, %r449;
and.b32 %r451, %r448, %r187;
or.b32 %r192, %r451, %r450;

dp4a.s32.s32 %r191, %r192, %r193, %r356;

mul.lo.s32 %r452, %r191, %r376;
cvt.rn.f32.s32 %f145, %r452;
fma.rn.ftz.f32 %f146, %f98, %f145, %f144;
fma.rn.ftz.f32 %f360, %f99, %f146, %f360;

{.reg .f16 low,high;
mov.b32 {low,high},%r177;
cvt.f32.f16 %f100, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r178;
cvt.f32.f16 %f101, low;}



{ cvt.f32.f16 %f102, %rs2;}


mov.u32 %r197, %r341;

prmt.b32 %r197, %r197, 0, 0xba98;

mov.u32 %r199, %r343;

prmt.b32 %r199, %r199, 0, 0xba98;

xor.b32 %r453, %r199, 2139062143;
not.b32 %r454, %r197;
and.b32 %r455, %r400, %r454;
and.b32 %r456, %r453, %r197;
or.b32 %r202, %r456, %r455;

dp4a.s32.s32 %r201, %r202, %r185, %r356;

mul.lo.s32 %r457, %r201, %r394;
cvt.rn.f32.s32 %f147, %r457;
fma.rn.ftz.f32 %f148, %f100, %f147, 0f00000000;
mov.u32 %r205, %r349;

prmt.b32 %r205, %r205, 0, 0xba98;

mov.u32 %r207, %r351;

prmt.b32 %r207, %r207, 0, 0xba98;

xor.b32 %r458, %r207, 2139062143;
not.b32 %r459, %r205;
and.b32 %r460, %r415, %r459;
and.b32 %r461, %r458, %r205;
or.b32 %r210, %r461, %r460;

dp4a.s32.s32 %r209, %r210, %r193, %r356;

mul.lo.s32 %r462, %r209, %r409;
cvt.rn.f32.s32 %f149, %r462;
fma.rn.ftz.f32 %f150, %f101, %f149, %f148;
fma.rn.ftz.f32 %f359, %f102, %f150, %f359;
cvt.u64.u32 %rd62, %r805;
add.s64 %rd63, %rd2, %rd62;
add.s64 %rd64, %rd63, %rd5;
mul.lo.s64 %rd65, %rd64, 36;
add.s64 %rd66, %rd14, %rd65;
add.s64 %rd67, %rd66, %rd42;
ld.global.nc.u32 %r221, [%rd67+4];
ld.global.nc.u32 %r213, [%rd66];

{.reg .f16 low,high;
mov.b32 {low,high},%r213;
cvt.f32.f16 %f103, low;}


ld.global.nc.u32 %r229, [%rd67+76];
ld.global.nc.u32 %r214, [%rd66+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r214;
cvt.f32.f16 %f104, low;}



{ cvt.f32.f16 %f105, %rs1;}


mov.u32 %r215, %r323;

prmt.b32 %r215, %r215, 0, 0xba98;

mov.u32 %r217, %r325;

prmt.b32 %r217, %r217, 0, 0xba98;

xor.b32 %r463, %r217, 2139062143;
not.b32 %r464, %r215;
and.b32 %r465, %r367, %r464;
and.b32 %r466, %r463, %r215;
or.b32 %r220, %r466, %r465;

dp4a.s32.s32 %r219, %r220, %r221, %r356;

mul.lo.s32 %r467, %r219, %r361;
cvt.rn.f32.s32 %f151, %r467;
fma.rn.ftz.f32 %f152, %f103, %f151, 0f00000000;
mov.u32 %r223, %r331;

prmt.b32 %r223, %r223, 0, 0xba98;

mov.u32 %r225, %r333;

prmt.b32 %r225, %r225, 0, 0xba98;

xor.b32 %r468, %r225, 2139062143;
not.b32 %r469, %r223;
and.b32 %r470, %r382, %r469;
and.b32 %r471, %r468, %r223;
or.b32 %r228, %r471, %r470;

dp4a.s32.s32 %r227, %r228, %r229, %r356;

mul.lo.s32 %r472, %r227, %r376;
cvt.rn.f32.s32 %f153, %r472;
fma.rn.ftz.f32 %f154, %f104, %f153, %f152;
fma.rn.ftz.f32 %f358, %f105, %f154, %f358;

{.reg .f16 low,high;
mov.b32 {low,high},%r213;
cvt.f32.f16 %f106, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r214;
cvt.f32.f16 %f107, low;}



{ cvt.f32.f16 %f108, %rs2;}


mov.u32 %r233, %r341;

prmt.b32 %r233, %r233, 0, 0xba98;

mov.u32 %r235, %r343;

prmt.b32 %r235, %r235, 0, 0xba98;

xor.b32 %r473, %r235, 2139062143;
not.b32 %r474, %r233;
and.b32 %r475, %r400, %r474;
and.b32 %r476, %r473, %r233;
or.b32 %r238, %r476, %r475;

dp4a.s32.s32 %r237, %r238, %r221, %r356;

mul.lo.s32 %r477, %r237, %r394;
cvt.rn.f32.s32 %f155, %r477;
fma.rn.ftz.f32 %f156, %f106, %f155, 0f00000000;
mov.u32 %r241, %r349;

prmt.b32 %r241, %r241, 0, 0xba98;

mov.u32 %r243, %r351;

prmt.b32 %r243, %r243, 0, 0xba98;

xor.b32 %r478, %r243, 2139062143;
not.b32 %r479, %r241;
and.b32 %r480, %r415, %r479;
and.b32 %r481, %r478, %r241;
or.b32 %r246, %r481, %r480;

dp4a.s32.s32 %r245, %r246, %r229, %r356;

mul.lo.s32 %r482, %r245, %r409;
cvt.rn.f32.s32 %f157, %r482;
fma.rn.ftz.f32 %f158, %f107, %f157, %f156;
fma.rn.ftz.f32 %f357, %f108, %f158, %f357;
cvt.u64.u32 %rd68, %r804;
add.s64 %rd69, %rd2, %rd68;
add.s64 %rd70, %rd69, %rd5;
mul.lo.s64 %rd71, %rd70, 36;
add.s64 %rd72, %rd14, %rd71;
add.s64 %rd73, %rd72, %rd42;
ld.global.nc.u32 %r257, [%rd73+4];
ld.global.nc.u32 %r249, [%rd72];

{.reg .f16 low,high;
mov.b32 {low,high},%r249;
cvt.f32.f16 %f109, low;}


ld.global.nc.u32 %r265, [%rd73+76];
ld.global.nc.u32 %r250, [%rd72+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r250;
cvt.f32.f16 %f110, low;}



{ cvt.f32.f16 %f111, %rs1;}


mov.u32 %r251, %r323;

prmt.b32 %r251, %r251, 0, 0xba98;

mov.u32 %r253, %r325;

prmt.b32 %r253, %r253, 0, 0xba98;

xor.b32 %r483, %r253, 2139062143;
not.b32 %r484, %r251;
and.b32 %r485, %r367, %r484;
and.b32 %r486, %r483, %r251;
or.b32 %r256, %r486, %r485;

dp4a.s32.s32 %r255, %r256, %r257, %r356;

mul.lo.s32 %r487, %r255, %r361;
cvt.rn.f32.s32 %f159, %r487;
fma.rn.ftz.f32 %f160, %f109, %f159, 0f00000000;
mov.u32 %r259, %r331;

prmt.b32 %r259, %r259, 0, 0xba98;

mov.u32 %r261, %r333;

prmt.b32 %r261, %r261, 0, 0xba98;

xor.b32 %r488, %r261, 2139062143;
not.b32 %r489, %r259;
and.b32 %r490, %r382, %r489;
and.b32 %r491, %r488, %r259;
or.b32 %r264, %r491, %r490;

dp4a.s32.s32 %r263, %r264, %r265, %r356;

mul.lo.s32 %r492, %r263, %r376;
cvt.rn.f32.s32 %f161, %r492;
fma.rn.ftz.f32 %f162, %f110, %f161, %f160;
fma.rn.ftz.f32 %f356, %f111, %f162, %f356;

{.reg .f16 low,high;
mov.b32 {low,high},%r249;
cvt.f32.f16 %f112, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r250;
cvt.f32.f16 %f113, low;}



{ cvt.f32.f16 %f114, %rs2;}


mov.u32 %r269, %r341;

prmt.b32 %r269, %r269, 0, 0xba98;

mov.u32 %r271, %r343;

prmt.b32 %r271, %r271, 0, 0xba98;

xor.b32 %r493, %r271, 2139062143;
not.b32 %r494, %r269;
and.b32 %r495, %r400, %r494;
and.b32 %r496, %r493, %r269;
or.b32 %r274, %r496, %r495;

dp4a.s32.s32 %r273, %r274, %r257, %r356;

mul.lo.s32 %r497, %r273, %r394;
cvt.rn.f32.s32 %f163, %r497;
fma.rn.ftz.f32 %f164, %f112, %f163, 0f00000000;
mov.u32 %r277, %r349;

prmt.b32 %r277, %r277, 0, 0xba98;

mov.u32 %r279, %r351;

prmt.b32 %r279, %r279, 0, 0xba98;

xor.b32 %r498, %r279, 2139062143;
not.b32 %r499, %r277;
and.b32 %r500, %r415, %r499;
and.b32 %r501, %r498, %r277;
or.b32 %r282, %r501, %r500;

dp4a.s32.s32 %r281, %r282, %r265, %r356;

mul.lo.s32 %r502, %r281, %r409;
cvt.rn.f32.s32 %f165, %r502;
fma.rn.ftz.f32 %f166, %f113, %f165, %f164;
fma.rn.ftz.f32 %f355, %f114, %f166, %f355;
cvt.u64.u32 %rd74, %r803;
add.s64 %rd75, %rd2, %rd74;
add.s64 %rd76, %rd75, %rd5;
mul.lo.s64 %rd77, %rd76, 36;
add.s64 %rd78, %rd14, %rd77;
add.s64 %rd79, %rd78, %rd42;
ld.global.nc.u32 %r293, [%rd79+4];
ld.global.nc.u32 %r285, [%rd78];

{.reg .f16 low,high;
mov.b32 {low,high},%r285;
cvt.f32.f16 %f115, low;}


ld.global.nc.u32 %r301, [%rd79+76];
ld.global.nc.u32 %r286, [%rd78+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r286;
cvt.f32.f16 %f116, low;}



{ cvt.f32.f16 %f117, %rs1;}


mov.u32 %r287, %r323;

prmt.b32 %r287, %r287, 0, 0xba98;

mov.u32 %r289, %r325;

prmt.b32 %r289, %r289, 0, 0xba98;

xor.b32 %r503, %r289, 2139062143;
not.b32 %r504, %r287;
and.b32 %r505, %r367, %r504;
and.b32 %r506, %r503, %r287;
or.b32 %r292, %r506, %r505;

dp4a.s32.s32 %r291, %r292, %r293, %r356;

mul.lo.s32 %r507, %r291, %r361;
cvt.rn.f32.s32 %f167, %r507;
fma.rn.ftz.f32 %f168, %f115, %f167, 0f00000000;
mov.u32 %r295, %r331;

prmt.b32 %r295, %r295, 0, 0xba98;

mov.u32 %r297, %r333;

prmt.b32 %r297, %r297, 0, 0xba98;

xor.b32 %r508, %r297, 2139062143;
not.b32 %r509, %r295;
and.b32 %r510, %r382, %r509;
and.b32 %r511, %r508, %r295;
or.b32 %r300, %r511, %r510;

dp4a.s32.s32 %r299, %r300, %r301, %r356;

mul.lo.s32 %r512, %r299, %r376;
cvt.rn.f32.s32 %f169, %r512;
fma.rn.ftz.f32 %f170, %f116, %f169, %f168;
fma.rn.ftz.f32 %f354, %f117, %f170, %f354;

{.reg .f16 low,high;
mov.b32 {low,high},%r285;
cvt.f32.f16 %f118, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r286;
cvt.f32.f16 %f119, low;}



{ cvt.f32.f16 %f120, %rs2;}


mov.u32 %r305, %r341;

prmt.b32 %r305, %r305, 0, 0xba98;

mov.u32 %r307, %r343;

prmt.b32 %r307, %r307, 0, 0xba98;

xor.b32 %r513, %r307, 2139062143;
not.b32 %r514, %r305;
and.b32 %r515, %r400, %r514;
and.b32 %r516, %r513, %r305;
or.b32 %r310, %r516, %r515;

dp4a.s32.s32 %r309, %r310, %r293, %r356;

mul.lo.s32 %r517, %r309, %r394;
cvt.rn.f32.s32 %f171, %r517;
fma.rn.ftz.f32 %f172, %f118, %f171, 0f00000000;
mov.u32 %r313, %r349;

prmt.b32 %r313, %r313, 0, 0xba98;

mov.u32 %r315, %r351;

prmt.b32 %r315, %r315, 0, 0xba98;

xor.b32 %r518, %r315, 2139062143;
not.b32 %r519, %r313;
and.b32 %r520, %r415, %r519;
and.b32 %r521, %r518, %r313;
or.b32 %r318, %r521, %r520;

dp4a.s32.s32 %r317, %r318, %r301, %r356;

mul.lo.s32 %r522, %r317, %r409;
cvt.rn.f32.s32 %f173, %r522;
fma.rn.ftz.f32 %f174, %f119, %f173, %f172;
fma.rn.ftz.f32 %f353, %f120, %f174, %f353;
cvt.u64.u32 %rd80, %r802;
add.s64 %rd81, %rd2, %rd80;
add.s64 %rd82, %rd81, %rd5;
mul.lo.s64 %rd83, %rd82, 36;
add.s64 %rd84, %rd14, %rd83;
add.s64 %rd85, %rd84, %rd42;
ld.global.nc.u32 %r329, [%rd85+4];
ld.global.nc.u32 %r321, [%rd84];

{.reg .f16 low,high;
mov.b32 {low,high},%r321;
cvt.f32.f16 %f121, low;}


ld.global.nc.u32 %r337, [%rd85+76];
ld.global.nc.u32 %r322, [%rd84+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r322;
cvt.f32.f16 %f122, low;}



{ cvt.f32.f16 %f123, %rs1;}



prmt.b32 %r323, %r323, 0, 0xba98;


prmt.b32 %r325, %r325, 0, 0xba98;

xor.b32 %r523, %r325, 2139062143;
not.b32 %r524, %r323;
and.b32 %r525, %r367, %r524;
and.b32 %r526, %r523, %r323;
or.b32 %r328, %r526, %r525;

dp4a.s32.s32 %r327, %r328, %r329, %r356;

mul.lo.s32 %r527, %r327, %r361;
cvt.rn.f32.s32 %f175, %r527;
fma.rn.ftz.f32 %f176, %f121, %f175, 0f00000000;

prmt.b32 %r331, %r331, 0, 0xba98;


prmt.b32 %r333, %r333, 0, 0xba98;

xor.b32 %r528, %r333, 2139062143;
not.b32 %r529, %r331;
and.b32 %r530, %r382, %r529;
and.b32 %r531, %r528, %r331;
or.b32 %r336, %r531, %r530;

dp4a.s32.s32 %r335, %r336, %r337, %r356;

mul.lo.s32 %r532, %r335, %r376;
cvt.rn.f32.s32 %f177, %r532;
fma.rn.ftz.f32 %f178, %f122, %f177, %f176;
fma.rn.ftz.f32 %f352, %f123, %f178, %f352;

{.reg .f16 low,high;
mov.b32 {low,high},%r321;
cvt.f32.f16 %f124, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r322;
cvt.f32.f16 %f125, low;}



{ cvt.f32.f16 %f126, %rs2;}



prmt.b32 %r341, %r341, 0, 0xba98;


prmt.b32 %r343, %r343, 0, 0xba98;

xor.b32 %r533, %r343, 2139062143;
not.b32 %r534, %r341;
and.b32 %r535, %r400, %r534;
and.b32 %r536, %r533, %r341;
or.b32 %r346, %r536, %r535;

dp4a.s32.s32 %r345, %r346, %r329, %r356;

mul.lo.s32 %r537, %r345, %r394;
cvt.rn.f32.s32 %f179, %r537;
fma.rn.ftz.f32 %f180, %f124, %f179, 0f00000000;

prmt.b32 %r349, %r349, 0, 0xba98;


prmt.b32 %r351, %r351, 0, 0xba98;

xor.b32 %r538, %r351, 2139062143;
not.b32 %r539, %r349;
and.b32 %r540, %r415, %r539;
and.b32 %r541, %r538, %r349;
or.b32 %r354, %r541, %r540;

dp4a.s32.s32 %r353, %r354, %r337, %r356;

mul.lo.s32 %r542, %r353, %r409;
cvt.rn.f32.s32 %f181, %r542;
fma.rn.ftz.f32 %f182, %f125, %f181, %f180;
fma.rn.ftz.f32 %f351, %f126, %f182, %f351;
add.s32 %r806, %r806, 16;
add.s32 %r805, %r805, 16;
add.s32 %r804, %r804, 16;
add.s32 %r803, %r803, 16;
add.s32 %r802, %r802, 16;
add.s32 %r801, %r801, 16;
add.s32 %r800, %r800, 16;
add.s64 %rd120, %rd120, 420;
add.s32 %r807, %r807, 2;
setp.lt.s32 %p4, %r807, %r7;
@%p4 bra $L__BB196_2;

$L__BB196_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB196_5;

mov.u32 %r543, _ZZ13mul_mat_vec_qIL9ggml_type14ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r544, %r1, 1792, %r543;
shl.b32 %r545, %r2, 2;
add.s32 %r546, %r544, %r545;
st.shared.f32 [%r546+-1792], %f364;
st.shared.f32 [%r546+-1664], %f363;
st.shared.f32 [%r546+-1536], %f362;
st.shared.f32 [%r546+-1408], %f361;
st.shared.f32 [%r546+-1280], %f360;
st.shared.f32 [%r546+-1152], %f359;
st.shared.f32 [%r546+-1024], %f358;
st.shared.f32 [%r546+-896], %f357;
st.shared.f32 [%r546+-768], %f356;
st.shared.f32 [%r546+-640], %f355;
st.shared.f32 [%r546+-512], %f354;
st.shared.f32 [%r546+-384], %f353;
st.shared.f32 [%r546+-256], %f352;
st.shared.f32 [%r546+-128], %f351;

$L__BB196_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB196_34;

mov.u32 %r547, %ctaid.z;
mad.lo.s32 %r548, %r5, %r43, %r4;
mad.lo.s32 %r549, %r547, %r49, %r548;
cvt.u64.u32 %rd17, %r549;
shl.b32 %r550, %r2, 2;
mov.u32 %r551, _ZZ13mul_mat_vec_qIL9ggml_type14ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r32, %r551, %r550;
mov.u32 %r552, 2;
ld.shared.f32 %f183, [%r32];
add.ftz.f32 %f184, %f183, %f364;
mov.b32 %r553, %f184;
mov.u32 %r554, 31;
mov.u32 %r555, 16;
mov.u32 %r556, -1;
shfl.sync.bfly.b32 %r557|%p7, %r553, %r555, %r554, %r556;
mov.b32 %f185, %r557;
add.ftz.f32 %f186, %f184, %f185;
mov.b32 %r558, %f186;
mov.u32 %r559, 8;
shfl.sync.bfly.b32 %r560|%p8, %r558, %r559, %r554, %r556;
mov.b32 %f187, %r560;
add.ftz.f32 %f188, %f186, %f187;
mov.b32 %r561, %f188;
mov.u32 %r562, 4;
shfl.sync.bfly.b32 %r563|%p9, %r561, %r562, %r554, %r556;
mov.b32 %f189, %r563;
add.ftz.f32 %f190, %f188, %f189;
mov.b32 %r564, %f190;
shfl.sync.bfly.b32 %r565|%p10, %r564, %r552, %r554, %r556;
mov.b32 %f191, %r565;
add.ftz.f32 %f192, %f190, %f191;
mov.b32 %r566, %f192;
mov.u32 %r567, 1;
shfl.sync.bfly.b32 %r568|%p11, %r566, %r567, %r554, %r556;
mov.b32 %f193, %r568;
add.ftz.f32 %f43, %f192, %f193;
setp.lt.u32 %p12, %r4, %r37;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
mul.wide.u32 %rd86, %r549, 4;
add.s64 %rd18, %rd1, %rd86;
not.pred %p14, %p1;
@%p14 bra $L__BB196_8;

st.global.f32 [%rd18], %f43;

$L__BB196_8:
ld.shared.f32 %f194, [%r32+128];
add.ftz.f32 %f195, %f194, %f363;
mov.b32 %r569, %f195;
shfl.sync.bfly.b32 %r573|%p15, %r569, %r555, %r554, %r556;
mov.b32 %f196, %r573;
add.ftz.f32 %f197, %f195, %f196;
mov.b32 %r574, %f197;
shfl.sync.bfly.b32 %r576|%p16, %r574, %r559, %r554, %r556;
mov.b32 %f198, %r576;
add.ftz.f32 %f199, %f197, %f198;
mov.b32 %r577, %f199;
shfl.sync.bfly.b32 %r579|%p17, %r577, %r562, %r554, %r556;
mov.b32 %f200, %r579;
add.ftz.f32 %f201, %f199, %f200;
mov.b32 %r580, %f201;
shfl.sync.bfly.b32 %r582|%p18, %r580, %r552, %r554, %r556;
mov.b32 %f202, %r582;
add.ftz.f32 %f203, %f201, %f202;
mov.b32 %r583, %f203;
shfl.sync.bfly.b32 %r585|%p19, %r583, %r567, %r554, %r556;
mov.b32 %f204, %r585;
add.ftz.f32 %f44, %f203, %f204;
setp.eq.s32 %p20, %r2, 1;
or.b32 %r589, %r4, 1;
setp.lt.u32 %p21, %r589, %r37;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB196_10;

st.global.f32 [%rd18+4], %f44;

$L__BB196_10:
ld.shared.f32 %f205, [%r32+256];
add.ftz.f32 %f206, %f205, %f362;
mov.b32 %r590, %f206;
mov.u32 %r591, 31;
mov.u32 %r592, 16;
mov.u32 %r593, -1;
shfl.sync.bfly.b32 %r594|%p23, %r590, %r592, %r591, %r593;
mov.b32 %f207, %r594;
add.ftz.f32 %f208, %f206, %f207;
mov.b32 %r595, %f208;
mov.u32 %r596, 8;
shfl.sync.bfly.b32 %r597|%p24, %r595, %r596, %r591, %r593;
mov.b32 %f209, %r597;
add.ftz.f32 %f210, %f208, %f209;
mov.b32 %r598, %f210;
mov.u32 %r599, 4;
shfl.sync.bfly.b32 %r600|%p25, %r598, %r599, %r591, %r593;
mov.b32 %f211, %r600;
add.ftz.f32 %f212, %f210, %f211;
mov.b32 %r601, %f212;
mov.u32 %r602, 2;
shfl.sync.bfly.b32 %r603|%p26, %r601, %r602, %r591, %r593;
mov.b32 %f213, %r603;
add.ftz.f32 %f214, %f212, %f213;
mov.b32 %r604, %f214;
mov.u32 %r605, 1;
shfl.sync.bfly.b32 %r606|%p27, %r604, %r605, %r591, %r593;
mov.b32 %f215, %r606;
add.ftz.f32 %f45, %f214, %f215;
@%p14 bra $L__BB196_12;

cvt.u64.u32 %rd87, %r37;
add.s64 %rd88, %rd87, %rd17;
shl.b64 %rd89, %rd88, 2;
add.s64 %rd90, %rd1, %rd89;
st.global.f32 [%rd90], %f45;

$L__BB196_12:
ld.shared.f32 %f216, [%r32+384];
add.ftz.f32 %f217, %f216, %f361;
mov.b32 %r607, %f217;
shfl.sync.bfly.b32 %r611|%p29, %r607, %r592, %r591, %r593;
mov.b32 %f218, %r611;
add.ftz.f32 %f219, %f217, %f218;
mov.b32 %r612, %f219;
shfl.sync.bfly.b32 %r614|%p30, %r612, %r596, %r591, %r593;
mov.b32 %f220, %r614;
add.ftz.f32 %f221, %f219, %f220;
mov.b32 %r615, %f221;
shfl.sync.bfly.b32 %r617|%p31, %r615, %r599, %r591, %r593;
mov.b32 %f222, %r617;
add.ftz.f32 %f223, %f221, %f222;
mov.b32 %r618, %f223;
shfl.sync.bfly.b32 %r620|%p32, %r618, %r602, %r591, %r593;
mov.b32 %f224, %r620;
add.ftz.f32 %f225, %f223, %f224;
mov.b32 %r621, %f225;
shfl.sync.bfly.b32 %r623|%p33, %r621, %r605, %r591, %r593;
mov.b32 %f226, %r623;
add.ftz.f32 %f46, %f225, %f226;
@%p22 bra $L__BB196_14;

add.s32 %r624, %r37, 1;
cvt.u64.u32 %rd91, %r624;
add.s64 %rd92, %rd91, %rd17;
shl.b64 %rd93, %rd92, 2;
add.s64 %rd94, %rd1, %rd93;
st.global.f32 [%rd94], %f46;

$L__BB196_14:
shl.b32 %r625, %r37, 1;
mov.u32 %r626, 1;
ld.shared.f32 %f227, [%r32+512];
add.ftz.f32 %f228, %f227, %f360;
mov.b32 %r627, %f228;
mov.u32 %r628, 31;
mov.u32 %r629, 16;
mov.u32 %r630, -1;
shfl.sync.bfly.b32 %r631|%p35, %r627, %r629, %r628, %r630;
mov.b32 %f229, %r631;
add.ftz.f32 %f230, %f228, %f229;
mov.b32 %r632, %f230;
mov.u32 %r633, 8;
shfl.sync.bfly.b32 %r634|%p36, %r632, %r633, %r628, %r630;
mov.b32 %f231, %r634;
add.ftz.f32 %f232, %f230, %f231;
mov.b32 %r635, %f232;
mov.u32 %r636, 4;
shfl.sync.bfly.b32 %r637|%p37, %r635, %r636, %r628, %r630;
mov.b32 %f233, %r637;
add.ftz.f32 %f234, %f232, %f233;
mov.b32 %r638, %f234;
mov.u32 %r639, 2;
shfl.sync.bfly.b32 %r640|%p38, %r638, %r639, %r628, %r630;
mov.b32 %f235, %r640;
add.ftz.f32 %f236, %f234, %f235;
mov.b32 %r641, %f236;
shfl.sync.bfly.b32 %r642|%p39, %r641, %r626, %r628, %r630;
mov.b32 %f237, %r642;
add.ftz.f32 %f47, %f236, %f237;
cvt.u64.u32 %rd95, %r625;
add.s64 %rd96, %rd95, %rd17;
shl.b64 %rd97, %rd96, 2;
add.s64 %rd19, %rd1, %rd97;
@%p14 bra $L__BB196_16;

st.global.f32 [%rd19], %f47;

$L__BB196_16:
ld.shared.f32 %f238, [%r32+640];
add.ftz.f32 %f239, %f238, %f359;
mov.b32 %r643, %f239;
shfl.sync.bfly.b32 %r647|%p41, %r643, %r629, %r628, %r630;
mov.b32 %f240, %r647;
add.ftz.f32 %f241, %f239, %f240;
mov.b32 %r648, %f241;
shfl.sync.bfly.b32 %r650|%p42, %r648, %r633, %r628, %r630;
mov.b32 %f242, %r650;
add.ftz.f32 %f243, %f241, %f242;
mov.b32 %r651, %f243;
shfl.sync.bfly.b32 %r653|%p43, %r651, %r636, %r628, %r630;
mov.b32 %f244, %r653;
add.ftz.f32 %f245, %f243, %f244;
mov.b32 %r654, %f245;
shfl.sync.bfly.b32 %r656|%p44, %r654, %r639, %r628, %r630;
mov.b32 %f246, %r656;
add.ftz.f32 %f247, %f245, %f246;
mov.b32 %r657, %f247;
shfl.sync.bfly.b32 %r659|%p45, %r657, %r626, %r628, %r630;
mov.b32 %f248, %r659;
add.ftz.f32 %f48, %f247, %f248;
@%p22 bra $L__BB196_18;

st.global.f32 [%rd19+4], %f48;

$L__BB196_18:
mul.lo.s32 %r33, %r37, 3;
ld.shared.f32 %f249, [%r32+768];
add.ftz.f32 %f250, %f249, %f358;
mov.b32 %r660, %f250;
mov.u32 %r661, 31;
mov.u32 %r662, 16;
mov.u32 %r663, -1;
shfl.sync.bfly.b32 %r664|%p47, %r660, %r662, %r661, %r663;
mov.b32 %f251, %r664;
add.ftz.f32 %f252, %f250, %f251;
mov.b32 %r665, %f252;
mov.u32 %r666, 8;
shfl.sync.bfly.b32 %r667|%p48, %r665, %r666, %r661, %r663;
mov.b32 %f253, %r667;
add.ftz.f32 %f254, %f252, %f253;
mov.b32 %r668, %f254;
mov.u32 %r669, 4;
shfl.sync.bfly.b32 %r670|%p49, %r668, %r669, %r661, %r663;
mov.b32 %f255, %r670;
add.ftz.f32 %f256, %f254, %f255;
mov.b32 %r671, %f256;
mov.u32 %r672, 2;
shfl.sync.bfly.b32 %r673|%p50, %r671, %r672, %r661, %r663;
mov.b32 %f257, %r673;
add.ftz.f32 %f258, %f256, %f257;
mov.b32 %r674, %f258;
mov.u32 %r675, 1;
shfl.sync.bfly.b32 %r676|%p51, %r674, %r675, %r661, %r663;
mov.b32 %f259, %r676;
add.ftz.f32 %f49, %f258, %f259;
@%p14 bra $L__BB196_20;

cvt.u64.u32 %rd98, %r33;
add.s64 %rd99, %rd98, %rd17;
shl.b64 %rd100, %rd99, 2;
add.s64 %rd101, %rd1, %rd100;
st.global.f32 [%rd101], %f49;

$L__BB196_20:
ld.shared.f32 %f260, [%r32+896];
add.ftz.f32 %f261, %f260, %f357;
mov.b32 %r677, %f261;
shfl.sync.bfly.b32 %r681|%p53, %r677, %r662, %r661, %r663;
mov.b32 %f262, %r681;
add.ftz.f32 %f263, %f261, %f262;
mov.b32 %r682, %f263;
shfl.sync.bfly.b32 %r684|%p54, %r682, %r666, %r661, %r663;
mov.b32 %f264, %r684;
add.ftz.f32 %f265, %f263, %f264;
mov.b32 %r685, %f265;
shfl.sync.bfly.b32 %r687|%p55, %r685, %r669, %r661, %r663;
mov.b32 %f266, %r687;
add.ftz.f32 %f267, %f265, %f266;
mov.b32 %r688, %f267;
shfl.sync.bfly.b32 %r690|%p56, %r688, %r672, %r661, %r663;
mov.b32 %f268, %r690;
add.ftz.f32 %f269, %f267, %f268;
mov.b32 %r691, %f269;
shfl.sync.bfly.b32 %r693|%p57, %r691, %r675, %r661, %r663;
mov.b32 %f270, %r693;
add.ftz.f32 %f50, %f269, %f270;
@%p22 bra $L__BB196_22;

add.s32 %r694, %r33, 1;
cvt.u64.u32 %rd102, %r694;
add.s64 %rd103, %rd102, %rd17;
shl.b64 %rd104, %rd103, 2;
add.s64 %rd105, %rd1, %rd104;
st.global.f32 [%rd105], %f50;

$L__BB196_22:
shl.b32 %r695, %r37, 2;
mov.u32 %r696, 2;
ld.shared.f32 %f271, [%r32+1024];
add.ftz.f32 %f272, %f271, %f356;
mov.b32 %r697, %f272;
mov.u32 %r698, 31;
mov.u32 %r699, 16;
mov.u32 %r700, -1;
shfl.sync.bfly.b32 %r701|%p59, %r697, %r699, %r698, %r700;
mov.b32 %f273, %r701;
add.ftz.f32 %f274, %f272, %f273;
mov.b32 %r702, %f274;
mov.u32 %r703, 8;
shfl.sync.bfly.b32 %r704|%p60, %r702, %r703, %r698, %r700;
mov.b32 %f275, %r704;
add.ftz.f32 %f276, %f274, %f275;
mov.b32 %r705, %f276;
mov.u32 %r706, 4;
shfl.sync.bfly.b32 %r707|%p61, %r705, %r706, %r698, %r700;
mov.b32 %f277, %r707;
add.ftz.f32 %f278, %f276, %f277;
mov.b32 %r708, %f278;
shfl.sync.bfly.b32 %r709|%p62, %r708, %r696, %r698, %r700;
mov.b32 %f279, %r709;
add.ftz.f32 %f280, %f278, %f279;
mov.b32 %r710, %f280;
mov.u32 %r711, 1;
shfl.sync.bfly.b32 %r712|%p63, %r710, %r711, %r698, %r700;
mov.b32 %f281, %r712;
add.ftz.f32 %f51, %f280, %f281;
cvt.u64.u32 %rd106, %r695;
add.s64 %rd107, %rd106, %rd17;
shl.b64 %rd108, %rd107, 2;
add.s64 %rd20, %rd1, %rd108;
@%p14 bra $L__BB196_24;

st.global.f32 [%rd20], %f51;

$L__BB196_24:
ld.shared.f32 %f282, [%r32+1152];
add.ftz.f32 %f283, %f282, %f355;
mov.b32 %r713, %f283;
shfl.sync.bfly.b32 %r717|%p65, %r713, %r699, %r698, %r700;
mov.b32 %f284, %r717;
add.ftz.f32 %f285, %f283, %f284;
mov.b32 %r718, %f285;
shfl.sync.bfly.b32 %r720|%p66, %r718, %r703, %r698, %r700;
mov.b32 %f286, %r720;
add.ftz.f32 %f287, %f285, %f286;
mov.b32 %r721, %f287;
shfl.sync.bfly.b32 %r723|%p67, %r721, %r706, %r698, %r700;
mov.b32 %f288, %r723;
add.ftz.f32 %f289, %f287, %f288;
mov.b32 %r724, %f289;
shfl.sync.bfly.b32 %r726|%p68, %r724, %r696, %r698, %r700;
mov.b32 %f290, %r726;
add.ftz.f32 %f291, %f289, %f290;
mov.b32 %r727, %f291;
shfl.sync.bfly.b32 %r729|%p69, %r727, %r711, %r698, %r700;
mov.b32 %f292, %r729;
add.ftz.f32 %f52, %f291, %f292;
@%p22 bra $L__BB196_26;

st.global.f32 [%rd20+4], %f52;

$L__BB196_26:
mul.lo.s32 %r34, %r37, 5;
ld.shared.f32 %f293, [%r32+1280];
add.ftz.f32 %f294, %f293, %f354;
mov.b32 %r730, %f294;
mov.u32 %r731, 31;
mov.u32 %r732, 16;
mov.u32 %r733, -1;
shfl.sync.bfly.b32 %r734|%p71, %r730, %r732, %r731, %r733;
mov.b32 %f295, %r734;
add.ftz.f32 %f296, %f294, %f295;
mov.b32 %r735, %f296;
mov.u32 %r736, 8;
shfl.sync.bfly.b32 %r737|%p72, %r735, %r736, %r731, %r733;
mov.b32 %f297, %r737;
add.ftz.f32 %f298, %f296, %f297;
mov.b32 %r738, %f298;
mov.u32 %r739, 4;
shfl.sync.bfly.b32 %r740|%p73, %r738, %r739, %r731, %r733;
mov.b32 %f299, %r740;
add.ftz.f32 %f300, %f298, %f299;
mov.b32 %r741, %f300;
mov.u32 %r742, 2;
shfl.sync.bfly.b32 %r743|%p74, %r741, %r742, %r731, %r733;
mov.b32 %f301, %r743;
add.ftz.f32 %f302, %f300, %f301;
mov.b32 %r744, %f302;
mov.u32 %r745, 1;
shfl.sync.bfly.b32 %r746|%p75, %r744, %r745, %r731, %r733;
mov.b32 %f303, %r746;
add.ftz.f32 %f53, %f302, %f303;
@%p14 bra $L__BB196_28;

cvt.u64.u32 %rd109, %r34;
add.s64 %rd110, %rd109, %rd17;
shl.b64 %rd111, %rd110, 2;
add.s64 %rd112, %rd1, %rd111;
st.global.f32 [%rd112], %f53;

$L__BB196_28:
ld.shared.f32 %f304, [%r32+1408];
add.ftz.f32 %f305, %f304, %f353;
mov.b32 %r747, %f305;
shfl.sync.bfly.b32 %r751|%p77, %r747, %r732, %r731, %r733;
mov.b32 %f306, %r751;
add.ftz.f32 %f307, %f305, %f306;
mov.b32 %r752, %f307;
shfl.sync.bfly.b32 %r754|%p78, %r752, %r736, %r731, %r733;
mov.b32 %f308, %r754;
add.ftz.f32 %f309, %f307, %f308;
mov.b32 %r755, %f309;
shfl.sync.bfly.b32 %r757|%p79, %r755, %r739, %r731, %r733;
mov.b32 %f310, %r757;
add.ftz.f32 %f311, %f309, %f310;
mov.b32 %r758, %f311;
shfl.sync.bfly.b32 %r760|%p80, %r758, %r742, %r731, %r733;
mov.b32 %f312, %r760;
add.ftz.f32 %f313, %f311, %f312;
mov.b32 %r761, %f313;
shfl.sync.bfly.b32 %r763|%p81, %r761, %r745, %r731, %r733;
mov.b32 %f314, %r763;
add.ftz.f32 %f54, %f313, %f314;
@%p22 bra $L__BB196_30;

add.s32 %r764, %r34, 1;
cvt.u64.u32 %rd113, %r764;
add.s64 %rd114, %rd113, %rd17;
shl.b64 %rd115, %rd114, 2;
add.s64 %rd116, %rd1, %rd115;
st.global.f32 [%rd116], %f54;

$L__BB196_30:
mul.lo.s32 %r765, %r37, 6;
ld.shared.f32 %f315, [%r32+1536];
add.ftz.f32 %f316, %f315, %f352;
mov.b32 %r766, %f316;
mov.u32 %r767, 31;
mov.u32 %r768, 16;
mov.u32 %r769, -1;
shfl.sync.bfly.b32 %r770|%p83, %r766, %r768, %r767, %r769;
mov.b32 %f317, %r770;
add.ftz.f32 %f318, %f316, %f317;
mov.b32 %r771, %f318;
mov.u32 %r772, 8;
shfl.sync.bfly.b32 %r773|%p84, %r771, %r772, %r767, %r769;
mov.b32 %f319, %r773;
add.ftz.f32 %f320, %f318, %f319;
mov.b32 %r774, %f320;
mov.u32 %r775, 4;
shfl.sync.bfly.b32 %r776|%p85, %r774, %r775, %r767, %r769;
mov.b32 %f321, %r776;
add.ftz.f32 %f322, %f320, %f321;
mov.b32 %r777, %f322;
mov.u32 %r778, 2;
shfl.sync.bfly.b32 %r779|%p86, %r777, %r778, %r767, %r769;
mov.b32 %f323, %r779;
add.ftz.f32 %f324, %f322, %f323;
mov.b32 %r780, %f324;
mov.u32 %r781, 1;
shfl.sync.bfly.b32 %r782|%p87, %r780, %r781, %r767, %r769;
mov.b32 %f325, %r782;
add.ftz.f32 %f55, %f324, %f325;
cvt.u64.u32 %rd117, %r765;
add.s64 %rd118, %rd117, %rd17;
shl.b64 %rd119, %rd118, 2;
add.s64 %rd21, %rd1, %rd119;
@%p14 bra $L__BB196_32;

st.global.f32 [%rd21], %f55;

$L__BB196_32:
ld.shared.f32 %f326, [%r32+1664];
add.ftz.f32 %f327, %f326, %f351;
mov.b32 %r783, %f327;
shfl.sync.bfly.b32 %r787|%p89, %r783, %r768, %r767, %r769;
mov.b32 %f328, %r787;
add.ftz.f32 %f329, %f327, %f328;
mov.b32 %r788, %f329;
shfl.sync.bfly.b32 %r790|%p90, %r788, %r772, %r767, %r769;
mov.b32 %f330, %r790;
add.ftz.f32 %f331, %f329, %f330;
mov.b32 %r791, %f331;
shfl.sync.bfly.b32 %r793|%p91, %r791, %r775, %r767, %r769;
mov.b32 %f332, %r793;
add.ftz.f32 %f333, %f331, %f332;
mov.b32 %r794, %f333;
shfl.sync.bfly.b32 %r796|%p92, %r794, %r778, %r767, %r769;
mov.b32 %f334, %r796;
add.ftz.f32 %f335, %f333, %f334;
mov.b32 %r797, %f335;
shfl.sync.bfly.b32 %r799|%p93, %r797, %r781, %r767, %r769;
mov.b32 %f336, %r799;
add.ftz.f32 %f56, %f335, %f336;
@%p22 bra $L__BB196_34;

st.global.f32 [%rd21+4], %f56;

$L__BB196_34:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type14ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type14ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type14ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type14ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type14ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type14ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type14ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type14ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type14ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type14ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<107>;
.reg .b16 %rs<29>;
.reg .f32 %f<417>;
.reg .b32 %r<920>;
.reg .b64 %rd<149>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type14ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[2048];

ld.param.u64 %rd20, [_Z13mul_mat_vec_qIL9ggml_type14ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd21, [_Z13mul_mat_vec_qIL9ggml_type14ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd22, [_Z13mul_mat_vec_qIL9ggml_type14ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r52, [_Z13mul_mat_vec_qIL9ggml_type14ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type14ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type14ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type14ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type14ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type14ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r45, [_Z13mul_mat_vec_qIL9ggml_type14ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r49, [_Z13mul_mat_vec_qIL9ggml_type14ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r50, [_Z13mul_mat_vec_qIL9ggml_type14ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r51, [_Z13mul_mat_vec_qIL9ggml_type14ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r47, [_Z13mul_mat_vec_qIL9ggml_type14ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r46, [_Z13mul_mat_vec_qIL9ggml_type14ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type14ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type14ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r53, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r53, %r2;
shr.s32 %r54, %r3, 31;
shr.u32 %r55, %r54, 27;
add.s32 %r56, %r3, %r55;
shr.s32 %r919, %r56, 5;
shr.u32 %r5, %r52, 8;
setp.ge.s32 %p3, %r919, %r5;
mov.f32 %f401, 0f00000000;
mov.f32 %f402, %f401;
mov.f32 %f403, %f401;
mov.f32 %f404, %f401;
mov.f32 %f405, %f401;
mov.f32 %f406, %f401;
mov.f32 %f407, %f401;
mov.f32 %f408, %f401;
mov.f32 %f409, %f401;
mov.f32 %f410, %f401;
mov.f32 %f411, %f401;
mov.f32 %f412, %f401;
mov.f32 %f413, %f401;
mov.f32 %f414, %f401;
mov.f32 %f415, %f401;
mov.f32 %f416, %f401;
@%p3 bra $L__BB197_3;

mov.u32 %r57, %ctaid.x;
shl.b32 %r58, %r57, 1;
mov.u32 %r59, %ctaid.y;
mov.u32 %r60, %ctaid.z;
mul.lo.s32 %r61, %r59, %r44;
cvt.u64.u32 %rd23, %r61;
mul.lo.s32 %r62, %r60, %r50;
cvt.u64.u32 %rd24, %r62;
add.s64 %rd1, %rd24, %rd23;
mul.hi.u32 %r63, %r60, %r46;
add.s32 %r64, %r63, %r60;
shr.u32 %r65, %r64, %r47;
mul.lo.s32 %r66, %r65, %r49;
mul.hi.u32 %r67, %r59, %r40;
add.s32 %r68, %r67, %r59;
shr.u32 %r69, %r68, %r41;
mul.lo.s32 %r70, %r69, %r43;
cvta.to.global.u64 %rd148, %rd20;
or.b32 %r71, %r58, 1;
and.b32 %r75, %r56, -32;
sub.s32 %r76, %r3, %r75;
shr.s32 %r77, %r76, 31;
shr.u32 %r78, %r77, 28;
add.s32 %r79, %r76, %r78;
shr.s32 %r80, %r79, 4;
shl.b32 %r81, %r80, 3;
and.b32 %r82, %r79, -16;
sub.s32 %r83, %r76, %r82;
shr.s32 %r84, %r83, 31;
shr.u32 %r85, %r84, 30;
add.s32 %r86, %r83, %r85;
shr.s32 %r87, %r86, 2;
add.s32 %r88, %r81, %r87;
cvt.s64.s32 %rd25, %r88;
shl.b32 %r89, %r80, 2;
shr.u32 %r90, %r84, 29;
add.s32 %r91, %r83, %r90;
shr.s32 %r92, %r91, 3;
add.s32 %r93, %r89, %r92;
shl.b32 %r6, %r92, 1;
shl.b32 %r94, %r76, 1;
shr.u32 %r95, %r77, 29;
add.s32 %r96, %r76, %r95;
and.b32 %r97, %r96, -8;
sub.s32 %r98, %r76, %r97;
add.s32 %r99, %r81, %r98;
shl.b32 %r100, %r99, 1;
cvt.s64.s32 %rd3, %r98;
cvt.s64.s32 %rd4, %r93;
shl.b32 %r101, %r38, 1;
shl.b32 %r911, %r919, 3;
add.s32 %r918, %r911, %r101;
mad.lo.s32 %r917, %r38, 3, %r911;
shl.b32 %r102, %r38, 2;
add.s32 %r916, %r911, %r102;
mad.lo.s32 %r915, %r38, 5, %r911;
mad.lo.s32 %r914, %r38, 6, %r911;
mad.lo.s32 %r913, %r38, 7, %r911;
add.s32 %r912, %r38, %r911;
mad.lo.s32 %r103, %r58, %r37, %r919;
add.s32 %r104, %r103, %r66;
add.s32 %r105, %r104, %r70;
mul.wide.s32 %rd5, %r105, 210;
mul.wide.s32 %rd26, %r94, 2;
add.s64 %rd27, %rd5, %rd26;
add.s64 %rd6, %rd27, 2;
add.s32 %r106, %r919, %r66;
add.s32 %r107, %r106, %r70;
mad.lo.s32 %r108, %r71, %r37, %r107;
mul.wide.s32 %rd7, %r108, 210;
add.s64 %rd28, %rd7, %rd25;
add.s64 %rd8, %rd28, 192;
mul.wide.s32 %rd29, %r100, 2;
add.s64 %rd30, %rd5, %rd29;
add.s64 %rd9, %rd30, 130;
add.s64 %rd31, %rd7, %rd29;
add.s64 %rd10, %rd31, 128;
add.s64 %rd11, %rd7, %rd26;
add.s64 %rd32, %rd5, %rd25;
add.s64 %rd12, %rd32, 192;
cvta.to.global.u64 %rd13, %rd21;
mov.f32 %f401, 0f00000000;
shl.b64 %rd40, %rd3, 2;

$L__BB197_2:
cvt.u64.u32 %rd33, %r911;
add.s64 %rd34, %rd1, %rd33;
add.s64 %rd35, %rd148, %rd6;
ld.global.nc.u16 %rs17, [%rd35+-2];
ld.global.nc.u16 %rs18, [%rd35];
mov.b32 %r397, {%rs17, %rs18};
add.s64 %rd36, %rd148, %rd9;
ld.global.nc.u16 %rs19, [%rd36+-2];
ld.global.nc.u16 %rs20, [%rd36];
mov.b32 %r398, {%rs19, %rs20};
shr.s32 %r399, %r398, %r6;
add.s64 %rd37, %rd34, %rd4;
mul.lo.s64 %rd38, %rd37, 36;
add.s64 %rd39, %rd13, %rd38;
add.s64 %rd41, %rd39, %rd40;
ld.global.nc.u32 %r117, [%rd41+4];
ld.global.nc.u32 %r109, [%rd39];

{.reg .f16 low,high;
mov.b32 {low,high},%r109;
cvt.f32.f16 %f97, low;}


ld.global.nc.u32 %r125, [%rd41+76];
ld.global.nc.u32 %r110, [%rd39+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r110;
cvt.f32.f16 %f98, low;}


add.s64 %rd42, %rd148, %rd5;
ld.global.nc.u16 %rs1, [%rd42+208];

{ cvt.f32.f16 %f99, %rs1;}


add.s64 %rd43, %rd148, %rd12;
ld.global.nc.u8 %rs21, [%rd43];
cvt.u32.u16 %r400, %rs21;
cvt.s32.s8 %r401, %r400;
and.b32 %r402, %r397, 252645135;
shl.b32 %r403, %r399, 4;
and.b32 %r404, %r403, 808464432;
or.b32 %r365, %r404, %r402;
or.b32 %r405, %r365, -2139062144;
add.s32 %r406, %r405, -538976288;
xor.b32 %r407, %r406, -2139062144;
xor.b32 %r408, %r365, 538976288;
xor.b32 %r409, %r406, %r365;
and.b32 %r363, %r409, %r408;
mov.u32 %r111, %r363;

prmt.b32 %r111, %r111, 0, 0xba98;

mov.u32 %r113, %r365;

prmt.b32 %r113, %r113, 0, 0xba98;

xor.b32 %r410, %r113, 2139062143;
not.b32 %r411, %r111;
and.b32 %r412, %r407, %r411;
and.b32 %r413, %r410, %r111;
or.b32 %r116, %r413, %r412;
mov.u32 %r396, 0;

dp4a.s32.s32 %r115, %r116, %r117, %r396;

mul.lo.s32 %r414, %r115, %r401;
cvt.rn.f32.s32 %f145, %r414;
fma.rn.ftz.f32 %f146, %f97, %f145, 0f00000000;
ld.global.nc.u8 %rs22, [%rd43+4];
cvt.u32.u16 %r415, %rs22;
cvt.s32.s8 %r416, %r415;
shr.u32 %r417, %r397, 4;
and.b32 %r418, %r417, 252645135;
and.b32 %r419, %r399, 808464432;
or.b32 %r373, %r419, %r418;
or.b32 %r420, %r373, -2139062144;
add.s32 %r421, %r420, -538976288;
xor.b32 %r422, %r421, -2139062144;
xor.b32 %r423, %r373, 538976288;
xor.b32 %r424, %r421, %r373;
and.b32 %r371, %r424, %r423;
mov.u32 %r119, %r371;

prmt.b32 %r119, %r119, 0, 0xba98;

mov.u32 %r121, %r373;

prmt.b32 %r121, %r121, 0, 0xba98;

xor.b32 %r425, %r121, 2139062143;
not.b32 %r426, %r119;
and.b32 %r427, %r422, %r426;
and.b32 %r428, %r425, %r119;
or.b32 %r124, %r428, %r427;

dp4a.s32.s32 %r123, %r124, %r125, %r396;

mul.lo.s32 %r429, %r123, %r416;
cvt.rn.f32.s32 %f147, %r429;
fma.rn.ftz.f32 %f148, %f98, %f147, %f146;
fma.rn.ftz.f32 %f416, %f99, %f148, %f416;
add.s64 %rd44, %rd148, %rd11;
ld.global.nc.u16 %rs23, [%rd44];
ld.global.nc.u16 %rs24, [%rd44+2];
mov.b32 %r430, {%rs23, %rs24};
add.s64 %rd45, %rd148, %rd10;
ld.global.nc.u16 %rs25, [%rd45];
ld.global.nc.u16 %rs26, [%rd45+2];
mov.b32 %r431, {%rs25, %rs26};
shr.s32 %r432, %r431, %r6;

{.reg .f16 low,high;
mov.b32 {low,high},%r109;
cvt.f32.f16 %f100, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r110;
cvt.f32.f16 %f101, low;}


add.s64 %rd46, %rd148, %rd7;
ld.global.nc.u16 %rs2, [%rd46+208];

{ cvt.f32.f16 %f102, %rs2;}


add.s64 %rd47, %rd148, %rd8;
ld.global.nc.u8 %rs27, [%rd47];
cvt.u32.u16 %r433, %rs27;
cvt.s32.s8 %r434, %r433;
and.b32 %r435, %r430, 252645135;
shl.b32 %r436, %r432, 4;
and.b32 %r437, %r436, 808464432;
or.b32 %r383, %r437, %r435;
or.b32 %r438, %r383, -2139062144;
add.s32 %r439, %r438, -538976288;
xor.b32 %r440, %r439, -2139062144;
xor.b32 %r441, %r383, 538976288;
xor.b32 %r442, %r439, %r383;
and.b32 %r381, %r442, %r441;
mov.u32 %r129, %r381;

prmt.b32 %r129, %r129, 0, 0xba98;

mov.u32 %r131, %r383;

prmt.b32 %r131, %r131, 0, 0xba98;

xor.b32 %r443, %r131, 2139062143;
not.b32 %r444, %r129;
and.b32 %r445, %r440, %r444;
and.b32 %r446, %r443, %r129;
or.b32 %r134, %r446, %r445;

dp4a.s32.s32 %r133, %r134, %r117, %r396;

mul.lo.s32 %r447, %r133, %r434;
cvt.rn.f32.s32 %f149, %r447;
fma.rn.ftz.f32 %f150, %f100, %f149, 0f00000000;
ld.global.nc.u8 %rs28, [%rd47+4];
cvt.u32.u16 %r448, %rs28;
cvt.s32.s8 %r449, %r448;
shr.u32 %r450, %r430, 4;
and.b32 %r451, %r450, 252645135;
and.b32 %r452, %r432, 808464432;
or.b32 %r391, %r452, %r451;
or.b32 %r453, %r391, -2139062144;
add.s32 %r454, %r453, -538976288;
xor.b32 %r455, %r454, -2139062144;
xor.b32 %r456, %r391, 538976288;
xor.b32 %r457, %r454, %r391;
and.b32 %r389, %r457, %r456;
mov.u32 %r137, %r389;

prmt.b32 %r137, %r137, 0, 0xba98;

mov.u32 %r139, %r391;

prmt.b32 %r139, %r139, 0, 0xba98;

xor.b32 %r458, %r139, 2139062143;
not.b32 %r459, %r137;
and.b32 %r460, %r455, %r459;
and.b32 %r461, %r458, %r137;
or.b32 %r142, %r461, %r460;

dp4a.s32.s32 %r141, %r142, %r125, %r396;

mul.lo.s32 %r462, %r141, %r449;
cvt.rn.f32.s32 %f151, %r462;
fma.rn.ftz.f32 %f152, %f101, %f151, %f150;
fma.rn.ftz.f32 %f415, %f102, %f152, %f415;
cvt.u64.u32 %rd48, %r912;
add.s64 %rd49, %rd1, %rd48;
add.s64 %rd50, %rd49, %rd4;
mul.lo.s64 %rd51, %rd50, 36;
add.s64 %rd52, %rd13, %rd51;
add.s64 %rd53, %rd52, %rd40;
ld.global.nc.u32 %r153, [%rd53+4];
ld.global.nc.u32 %r145, [%rd52];

{.reg .f16 low,high;
mov.b32 {low,high},%r145;
cvt.f32.f16 %f103, low;}


ld.global.nc.u32 %r161, [%rd53+76];
ld.global.nc.u32 %r146, [%rd52+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r146;
cvt.f32.f16 %f104, low;}



{ cvt.f32.f16 %f105, %rs1;}


mov.u32 %r147, %r363;

prmt.b32 %r147, %r147, 0, 0xba98;

mov.u32 %r149, %r365;

prmt.b32 %r149, %r149, 0, 0xba98;

xor.b32 %r463, %r149, 2139062143;
not.b32 %r464, %r147;
and.b32 %r465, %r407, %r464;
and.b32 %r466, %r463, %r147;
or.b32 %r152, %r466, %r465;

dp4a.s32.s32 %r151, %r152, %r153, %r396;

mul.lo.s32 %r467, %r151, %r401;
cvt.rn.f32.s32 %f153, %r467;
fma.rn.ftz.f32 %f154, %f103, %f153, 0f00000000;
mov.u32 %r155, %r371;

prmt.b32 %r155, %r155, 0, 0xba98;

mov.u32 %r157, %r373;

prmt.b32 %r157, %r157, 0, 0xba98;

xor.b32 %r468, %r157, 2139062143;
not.b32 %r469, %r155;
and.b32 %r470, %r422, %r469;
and.b32 %r471, %r468, %r155;
or.b32 %r160, %r471, %r470;

dp4a.s32.s32 %r159, %r160, %r161, %r396;

mul.lo.s32 %r472, %r159, %r416;
cvt.rn.f32.s32 %f155, %r472;
fma.rn.ftz.f32 %f156, %f104, %f155, %f154;
fma.rn.ftz.f32 %f414, %f105, %f156, %f414;

{.reg .f16 low,high;
mov.b32 {low,high},%r145;
cvt.f32.f16 %f106, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r146;
cvt.f32.f16 %f107, low;}



{ cvt.f32.f16 %f108, %rs2;}


mov.u32 %r165, %r381;

prmt.b32 %r165, %r165, 0, 0xba98;

mov.u32 %r167, %r383;

prmt.b32 %r167, %r167, 0, 0xba98;

xor.b32 %r473, %r167, 2139062143;
not.b32 %r474, %r165;
and.b32 %r475, %r440, %r474;
and.b32 %r476, %r473, %r165;
or.b32 %r170, %r476, %r475;

dp4a.s32.s32 %r169, %r170, %r153, %r396;

mul.lo.s32 %r477, %r169, %r434;
cvt.rn.f32.s32 %f157, %r477;
fma.rn.ftz.f32 %f158, %f106, %f157, 0f00000000;
mov.u32 %r173, %r389;

prmt.b32 %r173, %r173, 0, 0xba98;

mov.u32 %r175, %r391;

prmt.b32 %r175, %r175, 0, 0xba98;

xor.b32 %r478, %r175, 2139062143;
not.b32 %r479, %r173;
and.b32 %r480, %r455, %r479;
and.b32 %r481, %r478, %r173;
or.b32 %r178, %r481, %r480;

dp4a.s32.s32 %r177, %r178, %r161, %r396;

mul.lo.s32 %r482, %r177, %r449;
cvt.rn.f32.s32 %f159, %r482;
fma.rn.ftz.f32 %f160, %f107, %f159, %f158;
fma.rn.ftz.f32 %f413, %f108, %f160, %f413;
cvt.u64.u32 %rd54, %r918;
add.s64 %rd55, %rd1, %rd54;
add.s64 %rd56, %rd55, %rd4;
mul.lo.s64 %rd57, %rd56, 36;
add.s64 %rd58, %rd13, %rd57;
add.s64 %rd59, %rd58, %rd40;
ld.global.nc.u32 %r189, [%rd59+4];
ld.global.nc.u32 %r181, [%rd58];

{.reg .f16 low,high;
mov.b32 {low,high},%r181;
cvt.f32.f16 %f109, low;}


ld.global.nc.u32 %r197, [%rd59+76];
ld.global.nc.u32 %r182, [%rd58+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r182;
cvt.f32.f16 %f110, low;}



{ cvt.f32.f16 %f111, %rs1;}


mov.u32 %r183, %r363;

prmt.b32 %r183, %r183, 0, 0xba98;

mov.u32 %r185, %r365;

prmt.b32 %r185, %r185, 0, 0xba98;

xor.b32 %r483, %r185, 2139062143;
not.b32 %r484, %r183;
and.b32 %r485, %r407, %r484;
and.b32 %r486, %r483, %r183;
or.b32 %r188, %r486, %r485;

dp4a.s32.s32 %r187, %r188, %r189, %r396;

mul.lo.s32 %r487, %r187, %r401;
cvt.rn.f32.s32 %f161, %r487;
fma.rn.ftz.f32 %f162, %f109, %f161, 0f00000000;
mov.u32 %r191, %r371;

prmt.b32 %r191, %r191, 0, 0xba98;

mov.u32 %r193, %r373;

prmt.b32 %r193, %r193, 0, 0xba98;

xor.b32 %r488, %r193, 2139062143;
not.b32 %r489, %r191;
and.b32 %r490, %r422, %r489;
and.b32 %r491, %r488, %r191;
or.b32 %r196, %r491, %r490;

dp4a.s32.s32 %r195, %r196, %r197, %r396;

mul.lo.s32 %r492, %r195, %r416;
cvt.rn.f32.s32 %f163, %r492;
fma.rn.ftz.f32 %f164, %f110, %f163, %f162;
fma.rn.ftz.f32 %f412, %f111, %f164, %f412;

{.reg .f16 low,high;
mov.b32 {low,high},%r181;
cvt.f32.f16 %f112, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r182;
cvt.f32.f16 %f113, low;}



{ cvt.f32.f16 %f114, %rs2;}


mov.u32 %r201, %r381;

prmt.b32 %r201, %r201, 0, 0xba98;

mov.u32 %r203, %r383;

prmt.b32 %r203, %r203, 0, 0xba98;

xor.b32 %r493, %r203, 2139062143;
not.b32 %r494, %r201;
and.b32 %r495, %r440, %r494;
and.b32 %r496, %r493, %r201;
or.b32 %r206, %r496, %r495;

dp4a.s32.s32 %r205, %r206, %r189, %r396;

mul.lo.s32 %r497, %r205, %r434;
cvt.rn.f32.s32 %f165, %r497;
fma.rn.ftz.f32 %f166, %f112, %f165, 0f00000000;
mov.u32 %r209, %r389;

prmt.b32 %r209, %r209, 0, 0xba98;

mov.u32 %r211, %r391;

prmt.b32 %r211, %r211, 0, 0xba98;

xor.b32 %r498, %r211, 2139062143;
not.b32 %r499, %r209;
and.b32 %r500, %r455, %r499;
and.b32 %r501, %r498, %r209;
or.b32 %r214, %r501, %r500;

dp4a.s32.s32 %r213, %r214, %r197, %r396;

mul.lo.s32 %r502, %r213, %r449;
cvt.rn.f32.s32 %f167, %r502;
fma.rn.ftz.f32 %f168, %f113, %f167, %f166;
fma.rn.ftz.f32 %f411, %f114, %f168, %f411;
cvt.u64.u32 %rd60, %r917;
add.s64 %rd61, %rd1, %rd60;
add.s64 %rd62, %rd61, %rd4;
mul.lo.s64 %rd63, %rd62, 36;
add.s64 %rd64, %rd13, %rd63;
add.s64 %rd65, %rd64, %rd40;
ld.global.nc.u32 %r225, [%rd65+4];
ld.global.nc.u32 %r217, [%rd64];

{.reg .f16 low,high;
mov.b32 {low,high},%r217;
cvt.f32.f16 %f115, low;}


ld.global.nc.u32 %r233, [%rd65+76];
ld.global.nc.u32 %r218, [%rd64+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r218;
cvt.f32.f16 %f116, low;}



{ cvt.f32.f16 %f117, %rs1;}


mov.u32 %r219, %r363;

prmt.b32 %r219, %r219, 0, 0xba98;

mov.u32 %r221, %r365;

prmt.b32 %r221, %r221, 0, 0xba98;

xor.b32 %r503, %r221, 2139062143;
not.b32 %r504, %r219;
and.b32 %r505, %r407, %r504;
and.b32 %r506, %r503, %r219;
or.b32 %r224, %r506, %r505;

dp4a.s32.s32 %r223, %r224, %r225, %r396;

mul.lo.s32 %r507, %r223, %r401;
cvt.rn.f32.s32 %f169, %r507;
fma.rn.ftz.f32 %f170, %f115, %f169, 0f00000000;
mov.u32 %r227, %r371;

prmt.b32 %r227, %r227, 0, 0xba98;

mov.u32 %r229, %r373;

prmt.b32 %r229, %r229, 0, 0xba98;

xor.b32 %r508, %r229, 2139062143;
not.b32 %r509, %r227;
and.b32 %r510, %r422, %r509;
and.b32 %r511, %r508, %r227;
or.b32 %r232, %r511, %r510;

dp4a.s32.s32 %r231, %r232, %r233, %r396;

mul.lo.s32 %r512, %r231, %r416;
cvt.rn.f32.s32 %f171, %r512;
fma.rn.ftz.f32 %f172, %f116, %f171, %f170;
fma.rn.ftz.f32 %f410, %f117, %f172, %f410;

{.reg .f16 low,high;
mov.b32 {low,high},%r217;
cvt.f32.f16 %f118, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r218;
cvt.f32.f16 %f119, low;}



{ cvt.f32.f16 %f120, %rs2;}


mov.u32 %r237, %r381;

prmt.b32 %r237, %r237, 0, 0xba98;

mov.u32 %r239, %r383;

prmt.b32 %r239, %r239, 0, 0xba98;

xor.b32 %r513, %r239, 2139062143;
not.b32 %r514, %r237;
and.b32 %r515, %r440, %r514;
and.b32 %r516, %r513, %r237;
or.b32 %r242, %r516, %r515;

dp4a.s32.s32 %r241, %r242, %r225, %r396;

mul.lo.s32 %r517, %r241, %r434;
cvt.rn.f32.s32 %f173, %r517;
fma.rn.ftz.f32 %f174, %f118, %f173, 0f00000000;
mov.u32 %r245, %r389;

prmt.b32 %r245, %r245, 0, 0xba98;

mov.u32 %r247, %r391;

prmt.b32 %r247, %r247, 0, 0xba98;

xor.b32 %r518, %r247, 2139062143;
not.b32 %r519, %r245;
and.b32 %r520, %r455, %r519;
and.b32 %r521, %r518, %r245;
or.b32 %r250, %r521, %r520;

dp4a.s32.s32 %r249, %r250, %r233, %r396;

mul.lo.s32 %r522, %r249, %r449;
cvt.rn.f32.s32 %f175, %r522;
fma.rn.ftz.f32 %f176, %f119, %f175, %f174;
fma.rn.ftz.f32 %f409, %f120, %f176, %f409;
cvt.u64.u32 %rd66, %r916;
add.s64 %rd67, %rd1, %rd66;
add.s64 %rd68, %rd67, %rd4;
mul.lo.s64 %rd69, %rd68, 36;
add.s64 %rd70, %rd13, %rd69;
add.s64 %rd71, %rd70, %rd40;
ld.global.nc.u32 %r261, [%rd71+4];
ld.global.nc.u32 %r253, [%rd70];

{.reg .f16 low,high;
mov.b32 {low,high},%r253;
cvt.f32.f16 %f121, low;}


ld.global.nc.u32 %r269, [%rd71+76];
ld.global.nc.u32 %r254, [%rd70+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r254;
cvt.f32.f16 %f122, low;}



{ cvt.f32.f16 %f123, %rs1;}


mov.u32 %r255, %r363;

prmt.b32 %r255, %r255, 0, 0xba98;

mov.u32 %r257, %r365;

prmt.b32 %r257, %r257, 0, 0xba98;

xor.b32 %r523, %r257, 2139062143;
not.b32 %r524, %r255;
and.b32 %r525, %r407, %r524;
and.b32 %r526, %r523, %r255;
or.b32 %r260, %r526, %r525;

dp4a.s32.s32 %r259, %r260, %r261, %r396;

mul.lo.s32 %r527, %r259, %r401;
cvt.rn.f32.s32 %f177, %r527;
fma.rn.ftz.f32 %f178, %f121, %f177, 0f00000000;
mov.u32 %r263, %r371;

prmt.b32 %r263, %r263, 0, 0xba98;

mov.u32 %r265, %r373;

prmt.b32 %r265, %r265, 0, 0xba98;

xor.b32 %r528, %r265, 2139062143;
not.b32 %r529, %r263;
and.b32 %r530, %r422, %r529;
and.b32 %r531, %r528, %r263;
or.b32 %r268, %r531, %r530;

dp4a.s32.s32 %r267, %r268, %r269, %r396;

mul.lo.s32 %r532, %r267, %r416;
cvt.rn.f32.s32 %f179, %r532;
fma.rn.ftz.f32 %f180, %f122, %f179, %f178;
fma.rn.ftz.f32 %f408, %f123, %f180, %f408;

{.reg .f16 low,high;
mov.b32 {low,high},%r253;
cvt.f32.f16 %f124, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r254;
cvt.f32.f16 %f125, low;}



{ cvt.f32.f16 %f126, %rs2;}


mov.u32 %r273, %r381;

prmt.b32 %r273, %r273, 0, 0xba98;

mov.u32 %r275, %r383;

prmt.b32 %r275, %r275, 0, 0xba98;

xor.b32 %r533, %r275, 2139062143;
not.b32 %r534, %r273;
and.b32 %r535, %r440, %r534;
and.b32 %r536, %r533, %r273;
or.b32 %r278, %r536, %r535;

dp4a.s32.s32 %r277, %r278, %r261, %r396;

mul.lo.s32 %r537, %r277, %r434;
cvt.rn.f32.s32 %f181, %r537;
fma.rn.ftz.f32 %f182, %f124, %f181, 0f00000000;
mov.u32 %r281, %r389;

prmt.b32 %r281, %r281, 0, 0xba98;

mov.u32 %r283, %r391;

prmt.b32 %r283, %r283, 0, 0xba98;

xor.b32 %r538, %r283, 2139062143;
not.b32 %r539, %r281;
and.b32 %r540, %r455, %r539;
and.b32 %r541, %r538, %r281;
or.b32 %r286, %r541, %r540;

dp4a.s32.s32 %r285, %r286, %r269, %r396;

mul.lo.s32 %r542, %r285, %r449;
cvt.rn.f32.s32 %f183, %r542;
fma.rn.ftz.f32 %f184, %f125, %f183, %f182;
fma.rn.ftz.f32 %f407, %f126, %f184, %f407;
cvt.u64.u32 %rd72, %r915;
add.s64 %rd73, %rd1, %rd72;
add.s64 %rd74, %rd73, %rd4;
mul.lo.s64 %rd75, %rd74, 36;
add.s64 %rd76, %rd13, %rd75;
add.s64 %rd77, %rd76, %rd40;
ld.global.nc.u32 %r297, [%rd77+4];
ld.global.nc.u32 %r289, [%rd76];

{.reg .f16 low,high;
mov.b32 {low,high},%r289;
cvt.f32.f16 %f127, low;}


ld.global.nc.u32 %r305, [%rd77+76];
ld.global.nc.u32 %r290, [%rd76+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r290;
cvt.f32.f16 %f128, low;}



{ cvt.f32.f16 %f129, %rs1;}


mov.u32 %r291, %r363;

prmt.b32 %r291, %r291, 0, 0xba98;

mov.u32 %r293, %r365;

prmt.b32 %r293, %r293, 0, 0xba98;

xor.b32 %r543, %r293, 2139062143;
not.b32 %r544, %r291;
and.b32 %r545, %r407, %r544;
and.b32 %r546, %r543, %r291;
or.b32 %r296, %r546, %r545;

dp4a.s32.s32 %r295, %r296, %r297, %r396;

mul.lo.s32 %r547, %r295, %r401;
cvt.rn.f32.s32 %f185, %r547;
fma.rn.ftz.f32 %f186, %f127, %f185, 0f00000000;
mov.u32 %r299, %r371;

prmt.b32 %r299, %r299, 0, 0xba98;

mov.u32 %r301, %r373;

prmt.b32 %r301, %r301, 0, 0xba98;

xor.b32 %r548, %r301, 2139062143;
not.b32 %r549, %r299;
and.b32 %r550, %r422, %r549;
and.b32 %r551, %r548, %r299;
or.b32 %r304, %r551, %r550;

dp4a.s32.s32 %r303, %r304, %r305, %r396;

mul.lo.s32 %r552, %r303, %r416;
cvt.rn.f32.s32 %f187, %r552;
fma.rn.ftz.f32 %f188, %f128, %f187, %f186;
fma.rn.ftz.f32 %f406, %f129, %f188, %f406;

{.reg .f16 low,high;
mov.b32 {low,high},%r289;
cvt.f32.f16 %f130, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r290;
cvt.f32.f16 %f131, low;}



{ cvt.f32.f16 %f132, %rs2;}


mov.u32 %r309, %r381;

prmt.b32 %r309, %r309, 0, 0xba98;

mov.u32 %r311, %r383;

prmt.b32 %r311, %r311, 0, 0xba98;

xor.b32 %r553, %r311, 2139062143;
not.b32 %r554, %r309;
and.b32 %r555, %r440, %r554;
and.b32 %r556, %r553, %r309;
or.b32 %r314, %r556, %r555;

dp4a.s32.s32 %r313, %r314, %r297, %r396;

mul.lo.s32 %r557, %r313, %r434;
cvt.rn.f32.s32 %f189, %r557;
fma.rn.ftz.f32 %f190, %f130, %f189, 0f00000000;
mov.u32 %r317, %r389;

prmt.b32 %r317, %r317, 0, 0xba98;

mov.u32 %r319, %r391;

prmt.b32 %r319, %r319, 0, 0xba98;

xor.b32 %r558, %r319, 2139062143;
not.b32 %r559, %r317;
and.b32 %r560, %r455, %r559;
and.b32 %r561, %r558, %r317;
or.b32 %r322, %r561, %r560;

dp4a.s32.s32 %r321, %r322, %r305, %r396;

mul.lo.s32 %r562, %r321, %r449;
cvt.rn.f32.s32 %f191, %r562;
fma.rn.ftz.f32 %f192, %f131, %f191, %f190;
fma.rn.ftz.f32 %f405, %f132, %f192, %f405;
cvt.u64.u32 %rd78, %r914;
add.s64 %rd79, %rd1, %rd78;
add.s64 %rd80, %rd79, %rd4;
mul.lo.s64 %rd81, %rd80, 36;
add.s64 %rd82, %rd13, %rd81;
add.s64 %rd83, %rd82, %rd40;
ld.global.nc.u32 %r333, [%rd83+4];
ld.global.nc.u32 %r325, [%rd82];

{.reg .f16 low,high;
mov.b32 {low,high},%r325;
cvt.f32.f16 %f133, low;}


ld.global.nc.u32 %r341, [%rd83+76];
ld.global.nc.u32 %r326, [%rd82+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r326;
cvt.f32.f16 %f134, low;}



{ cvt.f32.f16 %f135, %rs1;}


mov.u32 %r327, %r363;

prmt.b32 %r327, %r327, 0, 0xba98;

mov.u32 %r329, %r365;

prmt.b32 %r329, %r329, 0, 0xba98;

xor.b32 %r563, %r329, 2139062143;
not.b32 %r564, %r327;
and.b32 %r565, %r407, %r564;
and.b32 %r566, %r563, %r327;
or.b32 %r332, %r566, %r565;

dp4a.s32.s32 %r331, %r332, %r333, %r396;

mul.lo.s32 %r567, %r331, %r401;
cvt.rn.f32.s32 %f193, %r567;
fma.rn.ftz.f32 %f194, %f133, %f193, 0f00000000;
mov.u32 %r335, %r371;

prmt.b32 %r335, %r335, 0, 0xba98;

mov.u32 %r337, %r373;

prmt.b32 %r337, %r337, 0, 0xba98;

xor.b32 %r568, %r337, 2139062143;
not.b32 %r569, %r335;
and.b32 %r570, %r422, %r569;
and.b32 %r571, %r568, %r335;
or.b32 %r340, %r571, %r570;

dp4a.s32.s32 %r339, %r340, %r341, %r396;

mul.lo.s32 %r572, %r339, %r416;
cvt.rn.f32.s32 %f195, %r572;
fma.rn.ftz.f32 %f196, %f134, %f195, %f194;
fma.rn.ftz.f32 %f404, %f135, %f196, %f404;

{.reg .f16 low,high;
mov.b32 {low,high},%r325;
cvt.f32.f16 %f136, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r326;
cvt.f32.f16 %f137, low;}



{ cvt.f32.f16 %f138, %rs2;}


mov.u32 %r345, %r381;

prmt.b32 %r345, %r345, 0, 0xba98;

mov.u32 %r347, %r383;

prmt.b32 %r347, %r347, 0, 0xba98;

xor.b32 %r573, %r347, 2139062143;
not.b32 %r574, %r345;
and.b32 %r575, %r440, %r574;
and.b32 %r576, %r573, %r345;
or.b32 %r350, %r576, %r575;

dp4a.s32.s32 %r349, %r350, %r333, %r396;

mul.lo.s32 %r577, %r349, %r434;
cvt.rn.f32.s32 %f197, %r577;
fma.rn.ftz.f32 %f198, %f136, %f197, 0f00000000;
mov.u32 %r353, %r389;

prmt.b32 %r353, %r353, 0, 0xba98;

mov.u32 %r355, %r391;

prmt.b32 %r355, %r355, 0, 0xba98;

xor.b32 %r578, %r355, 2139062143;
not.b32 %r579, %r353;
and.b32 %r580, %r455, %r579;
and.b32 %r581, %r578, %r353;
or.b32 %r358, %r581, %r580;

dp4a.s32.s32 %r357, %r358, %r341, %r396;

mul.lo.s32 %r582, %r357, %r449;
cvt.rn.f32.s32 %f199, %r582;
fma.rn.ftz.f32 %f200, %f137, %f199, %f198;
fma.rn.ftz.f32 %f403, %f138, %f200, %f403;
cvt.u64.u32 %rd84, %r913;
add.s64 %rd85, %rd1, %rd84;
add.s64 %rd86, %rd85, %rd4;
mul.lo.s64 %rd87, %rd86, 36;
add.s64 %rd88, %rd13, %rd87;
add.s64 %rd89, %rd88, %rd40;
ld.global.nc.u32 %r369, [%rd89+4];
ld.global.nc.u32 %r361, [%rd88];

{.reg .f16 low,high;
mov.b32 {low,high},%r361;
cvt.f32.f16 %f139, low;}


ld.global.nc.u32 %r377, [%rd89+76];
ld.global.nc.u32 %r362, [%rd88+72];

{.reg .f16 low,high;
mov.b32 {low,high},%r362;
cvt.f32.f16 %f140, low;}



{ cvt.f32.f16 %f141, %rs1;}



prmt.b32 %r363, %r363, 0, 0xba98;


prmt.b32 %r365, %r365, 0, 0xba98;

xor.b32 %r583, %r365, 2139062143;
not.b32 %r584, %r363;
and.b32 %r585, %r407, %r584;
and.b32 %r586, %r583, %r363;
or.b32 %r368, %r586, %r585;

dp4a.s32.s32 %r367, %r368, %r369, %r396;

mul.lo.s32 %r587, %r367, %r401;
cvt.rn.f32.s32 %f201, %r587;
fma.rn.ftz.f32 %f202, %f139, %f201, 0f00000000;

prmt.b32 %r371, %r371, 0, 0xba98;


prmt.b32 %r373, %r373, 0, 0xba98;

xor.b32 %r588, %r373, 2139062143;
not.b32 %r589, %r371;
and.b32 %r590, %r422, %r589;
and.b32 %r591, %r588, %r371;
or.b32 %r376, %r591, %r590;

dp4a.s32.s32 %r375, %r376, %r377, %r396;

mul.lo.s32 %r592, %r375, %r416;
cvt.rn.f32.s32 %f203, %r592;
fma.rn.ftz.f32 %f204, %f140, %f203, %f202;
fma.rn.ftz.f32 %f402, %f141, %f204, %f402;

{.reg .f16 low,high;
mov.b32 {low,high},%r361;
cvt.f32.f16 %f142, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r362;
cvt.f32.f16 %f143, low;}



{ cvt.f32.f16 %f144, %rs2;}



prmt.b32 %r381, %r381, 0, 0xba98;


prmt.b32 %r383, %r383, 0, 0xba98;

xor.b32 %r593, %r383, 2139062143;
not.b32 %r594, %r381;
and.b32 %r595, %r440, %r594;
and.b32 %r596, %r593, %r381;
or.b32 %r386, %r596, %r595;

dp4a.s32.s32 %r385, %r386, %r369, %r396;

mul.lo.s32 %r597, %r385, %r434;
cvt.rn.f32.s32 %f205, %r597;
fma.rn.ftz.f32 %f206, %f142, %f205, 0f00000000;

prmt.b32 %r389, %r389, 0, 0xba98;


prmt.b32 %r391, %r391, 0, 0xba98;

xor.b32 %r598, %r391, 2139062143;
not.b32 %r599, %r389;
and.b32 %r600, %r455, %r599;
and.b32 %r601, %r598, %r389;
or.b32 %r394, %r601, %r600;

dp4a.s32.s32 %r393, %r394, %r377, %r396;

mul.lo.s32 %r602, %r393, %r449;
cvt.rn.f32.s32 %f207, %r602;
fma.rn.ftz.f32 %f208, %f143, %f207, %f206;
fma.rn.ftz.f32 %f401, %f144, %f208, %f401;
add.s32 %r918, %r918, 16;
add.s32 %r917, %r917, 16;
add.s32 %r916, %r916, 16;
add.s32 %r915, %r915, 16;
add.s32 %r914, %r914, 16;
add.s32 %r913, %r913, 16;
add.s32 %r912, %r912, 16;
add.s32 %r911, %r911, 16;
add.s64 %rd148, %rd148, 420;
add.s32 %r919, %r919, 2;
setp.lt.s32 %p4, %r919, %r5;
@%p4 bra $L__BB197_2;

$L__BB197_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB197_5;

shl.b32 %r603, %r1, 11;
mov.u32 %r604, _ZZ13mul_mat_vec_qIL9ggml_type14ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r605, %r604, %r603;
shl.b32 %r606, %r2, 2;
add.s32 %r607, %r605, %r606;
st.shared.f32 [%r607+-2048], %f416;
st.shared.f32 [%r607+-1920], %f415;
st.shared.f32 [%r607+-1792], %f414;
st.shared.f32 [%r607+-1664], %f413;
st.shared.f32 [%r607+-1536], %f412;
st.shared.f32 [%r607+-1408], %f411;
st.shared.f32 [%r607+-1280], %f410;
st.shared.f32 [%r607+-1152], %f409;
st.shared.f32 [%r607+-1024], %f408;
st.shared.f32 [%r607+-896], %f407;
st.shared.f32 [%r607+-768], %f406;
st.shared.f32 [%r607+-640], %f405;
st.shared.f32 [%r607+-512], %f404;
st.shared.f32 [%r607+-384], %f403;
st.shared.f32 [%r607+-256], %f402;
st.shared.f32 [%r607+-128], %f401;

$L__BB197_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB197_38;

mov.u32 %r608, %ctaid.z;
mov.u32 %r609, %ctaid.y;
mov.u32 %r610, %ctaid.x;
shl.b32 %r611, %r610, 1;
mov.u32 %r612, 1;
mad.lo.s32 %r613, %r609, %r45, %r611;
mad.lo.s32 %r614, %r608, %r51, %r613;
cvt.u64.u32 %rd16, %r614;
shl.b32 %r615, %r2, 2;
mov.u32 %r616, _ZZ13mul_mat_vec_qIL9ggml_type14ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r33, %r616, %r615;
mov.u32 %r617, 2;
ld.shared.f32 %f209, [%r33];
add.ftz.f32 %f210, %f209, %f416;
mov.b32 %r618, %f210;
mov.u32 %r619, 31;
mov.u32 %r620, 16;
mov.u32 %r621, -1;
shfl.sync.bfly.b32 %r622|%p7, %r618, %r620, %r619, %r621;
mov.b32 %f211, %r622;
add.ftz.f32 %f212, %f210, %f211;
mov.b32 %r623, %f212;
mov.u32 %r624, 8;
shfl.sync.bfly.b32 %r625|%p8, %r623, %r624, %r619, %r621;
mov.b32 %f213, %r625;
add.ftz.f32 %f214, %f212, %f213;
mov.b32 %r626, %f214;
mov.u32 %r627, 4;
shfl.sync.bfly.b32 %r628|%p9, %r626, %r627, %r619, %r621;
mov.b32 %f215, %r628;
add.ftz.f32 %f216, %f214, %f215;
mov.b32 %r629, %f216;
shfl.sync.bfly.b32 %r630|%p10, %r629, %r617, %r619, %r621;
mov.b32 %f217, %r630;
add.ftz.f32 %f218, %f216, %f217;
mov.b32 %r631, %f218;
shfl.sync.bfly.b32 %r632|%p11, %r631, %r612, %r619, %r621;
mov.b32 %f219, %r632;
add.ftz.f32 %f49, %f218, %f219;
setp.lt.u32 %p12, %r611, %r39;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
not.pred %p14, %p1;
@%p14 bra $L__BB197_8;

cvta.to.global.u64 %rd90, %rd22;
mul.wide.u32 %rd91, %r614, 4;
add.s64 %rd92, %rd90, %rd91;
st.global.f32 [%rd92], %f49;

$L__BB197_8:
ld.shared.f32 %f220, [%r33+128];
add.ftz.f32 %f221, %f220, %f415;
mov.b32 %r639, %f221;
shfl.sync.bfly.b32 %r643|%p15, %r639, %r620, %r619, %r621;
mov.b32 %f222, %r643;
add.ftz.f32 %f223, %f221, %f222;
mov.b32 %r644, %f223;
shfl.sync.bfly.b32 %r646|%p16, %r644, %r624, %r619, %r621;
mov.b32 %f224, %r646;
add.ftz.f32 %f225, %f223, %f224;
mov.b32 %r647, %f225;
shfl.sync.bfly.b32 %r649|%p17, %r647, %r627, %r619, %r621;
mov.b32 %f226, %r649;
add.ftz.f32 %f227, %f225, %f226;
mov.b32 %r650, %f227;
shfl.sync.bfly.b32 %r652|%p18, %r650, %r617, %r619, %r621;
mov.b32 %f228, %r652;
add.ftz.f32 %f229, %f227, %f228;
mov.b32 %r653, %f229;
shfl.sync.bfly.b32 %r655|%p19, %r653, %r612, %r619, %r621;
mov.b32 %f230, %r655;
add.ftz.f32 %f50, %f229, %f230;
setp.eq.s32 %p20, %r2, 1;
or.b32 %r659, %r611, 1;
setp.lt.u32 %p21, %r659, %r39;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB197_10;

cvta.to.global.u64 %rd93, %rd22;
mul.wide.u32 %rd94, %r614, 4;
add.s64 %rd95, %rd93, %rd94;
st.global.f32 [%rd95+4], %f50;

$L__BB197_10:
ld.shared.f32 %f231, [%r33+256];
add.ftz.f32 %f232, %f231, %f414;
mov.b32 %r666, %f232;
mov.u32 %r667, 31;
mov.u32 %r668, 16;
mov.u32 %r669, -1;
shfl.sync.bfly.b32 %r670|%p23, %r666, %r668, %r667, %r669;
mov.b32 %f233, %r670;
add.ftz.f32 %f234, %f232, %f233;
mov.b32 %r671, %f234;
mov.u32 %r672, 8;
shfl.sync.bfly.b32 %r673|%p24, %r671, %r672, %r667, %r669;
mov.b32 %f235, %r673;
add.ftz.f32 %f236, %f234, %f235;
mov.b32 %r674, %f236;
mov.u32 %r675, 4;
shfl.sync.bfly.b32 %r676|%p25, %r674, %r675, %r667, %r669;
mov.b32 %f237, %r676;
add.ftz.f32 %f238, %f236, %f237;
mov.b32 %r677, %f238;
mov.u32 %r678, 2;
shfl.sync.bfly.b32 %r679|%p26, %r677, %r678, %r667, %r669;
mov.b32 %f239, %r679;
add.ftz.f32 %f240, %f238, %f239;
mov.b32 %r680, %f240;
mov.u32 %r681, 1;
shfl.sync.bfly.b32 %r682|%p27, %r680, %r681, %r667, %r669;
mov.b32 %f241, %r682;
add.ftz.f32 %f51, %f240, %f241;
@%p14 bra $L__BB197_12;

cvt.u64.u32 %rd96, %r39;
add.s64 %rd97, %rd96, %rd16;
cvta.to.global.u64 %rd98, %rd22;
shl.b64 %rd99, %rd97, 2;
add.s64 %rd100, %rd98, %rd99;
st.global.f32 [%rd100], %f51;

$L__BB197_12:
ld.shared.f32 %f242, [%r33+384];
add.ftz.f32 %f243, %f242, %f413;
mov.b32 %r683, %f243;
shfl.sync.bfly.b32 %r687|%p29, %r683, %r668, %r667, %r669;
mov.b32 %f244, %r687;
add.ftz.f32 %f245, %f243, %f244;
mov.b32 %r688, %f245;
shfl.sync.bfly.b32 %r690|%p30, %r688, %r672, %r667, %r669;
mov.b32 %f246, %r690;
add.ftz.f32 %f247, %f245, %f246;
mov.b32 %r691, %f247;
shfl.sync.bfly.b32 %r693|%p31, %r691, %r675, %r667, %r669;
mov.b32 %f248, %r693;
add.ftz.f32 %f249, %f247, %f248;
mov.b32 %r694, %f249;
shfl.sync.bfly.b32 %r696|%p32, %r694, %r678, %r667, %r669;
mov.b32 %f250, %r696;
add.ftz.f32 %f251, %f249, %f250;
mov.b32 %r697, %f251;
shfl.sync.bfly.b32 %r699|%p33, %r697, %r681, %r667, %r669;
mov.b32 %f252, %r699;
add.ftz.f32 %f52, %f251, %f252;
@%p22 bra $L__BB197_14;

add.s32 %r700, %r39, 1;
cvt.u64.u32 %rd101, %r700;
add.s64 %rd102, %rd101, %rd16;
cvta.to.global.u64 %rd103, %rd22;
shl.b64 %rd104, %rd102, 2;
add.s64 %rd105, %rd103, %rd104;
st.global.f32 [%rd105], %f52;

$L__BB197_14:
shl.b32 %r701, %r39, 1;
mov.u32 %r702, 1;
ld.shared.f32 %f253, [%r33+512];
add.ftz.f32 %f254, %f253, %f412;
mov.b32 %r703, %f254;
mov.u32 %r704, 31;
mov.u32 %r705, 16;
mov.u32 %r706, -1;
shfl.sync.bfly.b32 %r707|%p35, %r703, %r705, %r704, %r706;
mov.b32 %f255, %r707;
add.ftz.f32 %f256, %f254, %f255;
mov.b32 %r708, %f256;
mov.u32 %r709, 8;
shfl.sync.bfly.b32 %r710|%p36, %r708, %r709, %r704, %r706;
mov.b32 %f257, %r710;
add.ftz.f32 %f258, %f256, %f257;
mov.b32 %r711, %f258;
mov.u32 %r712, 4;
shfl.sync.bfly.b32 %r713|%p37, %r711, %r712, %r704, %r706;
mov.b32 %f259, %r713;
add.ftz.f32 %f260, %f258, %f259;
mov.b32 %r714, %f260;
mov.u32 %r715, 2;
shfl.sync.bfly.b32 %r716|%p38, %r714, %r715, %r704, %r706;
mov.b32 %f261, %r716;
add.ftz.f32 %f262, %f260, %f261;
mov.b32 %r717, %f262;
shfl.sync.bfly.b32 %r718|%p39, %r717, %r702, %r704, %r706;
mov.b32 %f263, %r718;
add.ftz.f32 %f53, %f262, %f263;
cvt.u64.u32 %rd106, %r701;
add.s64 %rd107, %rd106, %rd16;
cvta.to.global.u64 %rd108, %rd22;
shl.b64 %rd109, %rd107, 2;
add.s64 %rd17, %rd108, %rd109;
@%p14 bra $L__BB197_16;

st.global.f32 [%rd17], %f53;

$L__BB197_16:
ld.shared.f32 %f264, [%r33+640];
add.ftz.f32 %f265, %f264, %f411;
mov.b32 %r719, %f265;
shfl.sync.bfly.b32 %r723|%p41, %r719, %r705, %r704, %r706;
mov.b32 %f266, %r723;
add.ftz.f32 %f267, %f265, %f266;
mov.b32 %r724, %f267;
shfl.sync.bfly.b32 %r726|%p42, %r724, %r709, %r704, %r706;
mov.b32 %f268, %r726;
add.ftz.f32 %f269, %f267, %f268;
mov.b32 %r727, %f269;
shfl.sync.bfly.b32 %r729|%p43, %r727, %r712, %r704, %r706;
mov.b32 %f270, %r729;
add.ftz.f32 %f271, %f269, %f270;
mov.b32 %r730, %f271;
shfl.sync.bfly.b32 %r732|%p44, %r730, %r715, %r704, %r706;
mov.b32 %f272, %r732;
add.ftz.f32 %f273, %f271, %f272;
mov.b32 %r733, %f273;
shfl.sync.bfly.b32 %r735|%p45, %r733, %r702, %r704, %r706;
mov.b32 %f274, %r735;
add.ftz.f32 %f54, %f273, %f274;
@%p22 bra $L__BB197_18;

st.global.f32 [%rd17+4], %f54;

$L__BB197_18:
mul.lo.s32 %r34, %r39, 3;
ld.shared.f32 %f275, [%r33+768];
add.ftz.f32 %f276, %f275, %f410;
mov.b32 %r736, %f276;
mov.u32 %r737, 31;
mov.u32 %r738, 16;
mov.u32 %r739, -1;
shfl.sync.bfly.b32 %r740|%p47, %r736, %r738, %r737, %r739;
mov.b32 %f277, %r740;
add.ftz.f32 %f278, %f276, %f277;
mov.b32 %r741, %f278;
mov.u32 %r742, 8;
shfl.sync.bfly.b32 %r743|%p48, %r741, %r742, %r737, %r739;
mov.b32 %f279, %r743;
add.ftz.f32 %f280, %f278, %f279;
mov.b32 %r744, %f280;
mov.u32 %r745, 4;
shfl.sync.bfly.b32 %r746|%p49, %r744, %r745, %r737, %r739;
mov.b32 %f281, %r746;
add.ftz.f32 %f282, %f280, %f281;
mov.b32 %r747, %f282;
mov.u32 %r748, 2;
shfl.sync.bfly.b32 %r749|%p50, %r747, %r748, %r737, %r739;
mov.b32 %f283, %r749;
add.ftz.f32 %f284, %f282, %f283;
mov.b32 %r750, %f284;
mov.u32 %r751, 1;
shfl.sync.bfly.b32 %r752|%p51, %r750, %r751, %r737, %r739;
mov.b32 %f285, %r752;
add.ftz.f32 %f55, %f284, %f285;
@%p14 bra $L__BB197_20;

cvt.u64.u32 %rd110, %r34;
add.s64 %rd111, %rd110, %rd16;
shl.b64 %rd113, %rd111, 2;
add.s64 %rd114, %rd108, %rd113;
st.global.f32 [%rd114], %f55;

$L__BB197_20:
ld.shared.f32 %f286, [%r33+896];
add.ftz.f32 %f287, %f286, %f409;
mov.b32 %r753, %f287;
shfl.sync.bfly.b32 %r757|%p53, %r753, %r738, %r737, %r739;
mov.b32 %f288, %r757;
add.ftz.f32 %f289, %f287, %f288;
mov.b32 %r758, %f289;
shfl.sync.bfly.b32 %r760|%p54, %r758, %r742, %r737, %r739;
mov.b32 %f290, %r760;
add.ftz.f32 %f291, %f289, %f290;
mov.b32 %r761, %f291;
shfl.sync.bfly.b32 %r763|%p55, %r761, %r745, %r737, %r739;
mov.b32 %f292, %r763;
add.ftz.f32 %f293, %f291, %f292;
mov.b32 %r764, %f293;
shfl.sync.bfly.b32 %r766|%p56, %r764, %r748, %r737, %r739;
mov.b32 %f294, %r766;
add.ftz.f32 %f295, %f293, %f294;
mov.b32 %r767, %f295;
shfl.sync.bfly.b32 %r769|%p57, %r767, %r751, %r737, %r739;
mov.b32 %f296, %r769;
add.ftz.f32 %f56, %f295, %f296;
@%p22 bra $L__BB197_22;

add.s32 %r770, %r34, 1;
cvt.u64.u32 %rd115, %r770;
add.s64 %rd116, %rd115, %rd16;
shl.b64 %rd118, %rd116, 2;
add.s64 %rd119, %rd108, %rd118;
st.global.f32 [%rd119], %f56;

$L__BB197_22:
shl.b32 %r771, %r39, 2;
mov.u32 %r772, 2;
ld.shared.f32 %f297, [%r33+1024];
add.ftz.f32 %f298, %f297, %f408;
mov.b32 %r773, %f298;
mov.u32 %r774, 31;
mov.u32 %r775, 16;
mov.u32 %r776, -1;
shfl.sync.bfly.b32 %r777|%p59, %r773, %r775, %r774, %r776;
mov.b32 %f299, %r777;
add.ftz.f32 %f300, %f298, %f299;
mov.b32 %r778, %f300;
mov.u32 %r779, 8;
shfl.sync.bfly.b32 %r780|%p60, %r778, %r779, %r774, %r776;
mov.b32 %f301, %r780;
add.ftz.f32 %f302, %f300, %f301;
mov.b32 %r781, %f302;
mov.u32 %r782, 4;
shfl.sync.bfly.b32 %r783|%p61, %r781, %r782, %r774, %r776;
mov.b32 %f303, %r783;
add.ftz.f32 %f304, %f302, %f303;
mov.b32 %r784, %f304;
shfl.sync.bfly.b32 %r785|%p62, %r784, %r772, %r774, %r776;
mov.b32 %f305, %r785;
add.ftz.f32 %f306, %f304, %f305;
mov.b32 %r786, %f306;
mov.u32 %r787, 1;
shfl.sync.bfly.b32 %r788|%p63, %r786, %r787, %r774, %r776;
mov.b32 %f307, %r788;
add.ftz.f32 %f57, %f306, %f307;
cvt.u64.u32 %rd120, %r771;
add.s64 %rd121, %rd120, %rd16;
shl.b64 %rd123, %rd121, 2;
add.s64 %rd18, %rd108, %rd123;
@%p14 bra $L__BB197_24;

st.global.f32 [%rd18], %f57;

$L__BB197_24:
ld.shared.f32 %f308, [%r33+1152];
add.ftz.f32 %f309, %f308, %f407;
mov.b32 %r789, %f309;
shfl.sync.bfly.b32 %r793|%p65, %r789, %r775, %r774, %r776;
mov.b32 %f310, %r793;
add.ftz.f32 %f311, %f309, %f310;
mov.b32 %r794, %f311;
shfl.sync.bfly.b32 %r796|%p66, %r794, %r779, %r774, %r776;
mov.b32 %f312, %r796;
add.ftz.f32 %f313, %f311, %f312;
mov.b32 %r797, %f313;
shfl.sync.bfly.b32 %r799|%p67, %r797, %r782, %r774, %r776;
mov.b32 %f314, %r799;
add.ftz.f32 %f315, %f313, %f314;
mov.b32 %r800, %f315;
shfl.sync.bfly.b32 %r802|%p68, %r800, %r772, %r774, %r776;
mov.b32 %f316, %r802;
add.ftz.f32 %f317, %f315, %f316;
mov.b32 %r803, %f317;
shfl.sync.bfly.b32 %r805|%p69, %r803, %r787, %r774, %r776;
mov.b32 %f318, %r805;
add.ftz.f32 %f58, %f317, %f318;
@%p22 bra $L__BB197_26;

st.global.f32 [%rd18+4], %f58;

$L__BB197_26:
mul.lo.s32 %r35, %r39, 5;
ld.shared.f32 %f319, [%r33+1280];
add.ftz.f32 %f320, %f319, %f406;
mov.b32 %r806, %f320;
mov.u32 %r807, 31;
mov.u32 %r808, 16;
mov.u32 %r809, -1;
shfl.sync.bfly.b32 %r810|%p71, %r806, %r808, %r807, %r809;
mov.b32 %f321, %r810;
add.ftz.f32 %f322, %f320, %f321;
mov.b32 %r811, %f322;
mov.u32 %r812, 8;
shfl.sync.bfly.b32 %r813|%p72, %r811, %r812, %r807, %r809;
mov.b32 %f323, %r813;
add.ftz.f32 %f324, %f322, %f323;
mov.b32 %r814, %f324;
mov.u32 %r815, 4;
shfl.sync.bfly.b32 %r816|%p73, %r814, %r815, %r807, %r809;
mov.b32 %f325, %r816;
add.ftz.f32 %f326, %f324, %f325;
mov.b32 %r817, %f326;
mov.u32 %r818, 2;
shfl.sync.bfly.b32 %r819|%p74, %r817, %r818, %r807, %r809;
mov.b32 %f327, %r819;
add.ftz.f32 %f328, %f326, %f327;
mov.b32 %r820, %f328;
mov.u32 %r821, 1;
shfl.sync.bfly.b32 %r822|%p75, %r820, %r821, %r807, %r809;
mov.b32 %f329, %r822;
add.ftz.f32 %f59, %f328, %f329;
@%p14 bra $L__BB197_28;

cvt.u64.u32 %rd124, %r35;
add.s64 %rd125, %rd124, %rd16;
shl.b64 %rd127, %rd125, 2;
add.s64 %rd128, %rd108, %rd127;
st.global.f32 [%rd128], %f59;

$L__BB197_28:
ld.shared.f32 %f330, [%r33+1408];
add.ftz.f32 %f331, %f330, %f405;
mov.b32 %r823, %f331;
shfl.sync.bfly.b32 %r827|%p77, %r823, %r808, %r807, %r809;
mov.b32 %f332, %r827;
add.ftz.f32 %f333, %f331, %f332;
mov.b32 %r828, %f333;
shfl.sync.bfly.b32 %r830|%p78, %r828, %r812, %r807, %r809;
mov.b32 %f334, %r830;
add.ftz.f32 %f335, %f333, %f334;
mov.b32 %r831, %f335;
shfl.sync.bfly.b32 %r833|%p79, %r831, %r815, %r807, %r809;
mov.b32 %f336, %r833;
add.ftz.f32 %f337, %f335, %f336;
mov.b32 %r834, %f337;
shfl.sync.bfly.b32 %r836|%p80, %r834, %r818, %r807, %r809;
mov.b32 %f338, %r836;
add.ftz.f32 %f339, %f337, %f338;
mov.b32 %r837, %f339;
shfl.sync.bfly.b32 %r839|%p81, %r837, %r821, %r807, %r809;
mov.b32 %f340, %r839;
add.ftz.f32 %f60, %f339, %f340;
@%p22 bra $L__BB197_30;

add.s32 %r840, %r35, 1;
cvt.u64.u32 %rd129, %r840;
add.s64 %rd130, %rd129, %rd16;
shl.b64 %rd132, %rd130, 2;
add.s64 %rd133, %rd108, %rd132;
st.global.f32 [%rd133], %f60;

$L__BB197_30:
mul.lo.s32 %r841, %r39, 6;
ld.shared.f32 %f341, [%r33+1536];
add.ftz.f32 %f342, %f341, %f404;
mov.b32 %r842, %f342;
mov.u32 %r843, 31;
mov.u32 %r844, 16;
mov.u32 %r845, -1;
shfl.sync.bfly.b32 %r846|%p83, %r842, %r844, %r843, %r845;
mov.b32 %f343, %r846;
add.ftz.f32 %f344, %f342, %f343;
mov.b32 %r847, %f344;
mov.u32 %r848, 8;
shfl.sync.bfly.b32 %r849|%p84, %r847, %r848, %r843, %r845;
mov.b32 %f345, %r849;
add.ftz.f32 %f346, %f344, %f345;
mov.b32 %r850, %f346;
mov.u32 %r851, 4;
shfl.sync.bfly.b32 %r852|%p85, %r850, %r851, %r843, %r845;
mov.b32 %f347, %r852;
add.ftz.f32 %f348, %f346, %f347;
mov.b32 %r853, %f348;
mov.u32 %r854, 2;
shfl.sync.bfly.b32 %r855|%p86, %r853, %r854, %r843, %r845;
mov.b32 %f349, %r855;
add.ftz.f32 %f350, %f348, %f349;
mov.b32 %r856, %f350;
mov.u32 %r857, 1;
shfl.sync.bfly.b32 %r858|%p87, %r856, %r857, %r843, %r845;
mov.b32 %f351, %r858;
add.ftz.f32 %f61, %f350, %f351;
cvt.u64.u32 %rd134, %r841;
add.s64 %rd135, %rd134, %rd16;
shl.b64 %rd137, %rd135, 2;
add.s64 %rd19, %rd108, %rd137;
@%p14 bra $L__BB197_32;

st.global.f32 [%rd19], %f61;

$L__BB197_32:
ld.shared.f32 %f352, [%r33+1664];
add.ftz.f32 %f353, %f352, %f403;
mov.b32 %r859, %f353;
shfl.sync.bfly.b32 %r863|%p89, %r859, %r844, %r843, %r845;
mov.b32 %f354, %r863;
add.ftz.f32 %f355, %f353, %f354;
mov.b32 %r864, %f355;
shfl.sync.bfly.b32 %r866|%p90, %r864, %r848, %r843, %r845;
mov.b32 %f356, %r866;
add.ftz.f32 %f357, %f355, %f356;
mov.b32 %r867, %f357;
shfl.sync.bfly.b32 %r869|%p91, %r867, %r851, %r843, %r845;
mov.b32 %f358, %r869;
add.ftz.f32 %f359, %f357, %f358;
mov.b32 %r870, %f359;
shfl.sync.bfly.b32 %r872|%p92, %r870, %r854, %r843, %r845;
mov.b32 %f360, %r872;
add.ftz.f32 %f361, %f359, %f360;
mov.b32 %r873, %f361;
shfl.sync.bfly.b32 %r875|%p93, %r873, %r857, %r843, %r845;
mov.b32 %f362, %r875;
add.ftz.f32 %f62, %f361, %f362;
@%p22 bra $L__BB197_34;

st.global.f32 [%rd19+4], %f62;

$L__BB197_34:
mul.lo.s32 %r36, %r39, 7;
ld.shared.f32 %f363, [%r33+1792];
add.ftz.f32 %f364, %f363, %f402;
mov.b32 %r876, %f364;
mov.u32 %r877, 31;
mov.u32 %r878, 16;
mov.u32 %r879, -1;
shfl.sync.bfly.b32 %r880|%p95, %r876, %r878, %r877, %r879;
mov.b32 %f365, %r880;
add.ftz.f32 %f366, %f364, %f365;
mov.b32 %r881, %f366;
mov.u32 %r882, 8;
shfl.sync.bfly.b32 %r883|%p96, %r881, %r882, %r877, %r879;
mov.b32 %f367, %r883;
add.ftz.f32 %f368, %f366, %f367;
mov.b32 %r884, %f368;
mov.u32 %r885, 4;
shfl.sync.bfly.b32 %r886|%p97, %r884, %r885, %r877, %r879;
mov.b32 %f369, %r886;
add.ftz.f32 %f370, %f368, %f369;
mov.b32 %r887, %f370;
mov.u32 %r888, 2;
shfl.sync.bfly.b32 %r889|%p98, %r887, %r888, %r877, %r879;
mov.b32 %f371, %r889;
add.ftz.f32 %f372, %f370, %f371;
mov.b32 %r890, %f372;
mov.u32 %r891, 1;
shfl.sync.bfly.b32 %r892|%p99, %r890, %r891, %r877, %r879;
mov.b32 %f373, %r892;
add.ftz.f32 %f63, %f372, %f373;
@%p14 bra $L__BB197_36;

cvt.u64.u32 %rd138, %r36;
add.s64 %rd139, %rd138, %rd16;
shl.b64 %rd141, %rd139, 2;
add.s64 %rd142, %rd108, %rd141;
st.global.f32 [%rd142], %f63;

$L__BB197_36:
ld.shared.f32 %f374, [%r33+1920];
add.ftz.f32 %f375, %f374, %f401;
mov.b32 %r893, %f375;
shfl.sync.bfly.b32 %r897|%p101, %r893, %r878, %r877, %r879;
mov.b32 %f376, %r897;
add.ftz.f32 %f377, %f375, %f376;
mov.b32 %r898, %f377;
shfl.sync.bfly.b32 %r900|%p102, %r898, %r882, %r877, %r879;
mov.b32 %f378, %r900;
add.ftz.f32 %f379, %f377, %f378;
mov.b32 %r901, %f379;
shfl.sync.bfly.b32 %r903|%p103, %r901, %r885, %r877, %r879;
mov.b32 %f380, %r903;
add.ftz.f32 %f381, %f379, %f380;
mov.b32 %r904, %f381;
shfl.sync.bfly.b32 %r906|%p104, %r904, %r888, %r877, %r879;
mov.b32 %f382, %r906;
add.ftz.f32 %f383, %f381, %f382;
mov.b32 %r907, %f383;
shfl.sync.bfly.b32 %r909|%p105, %r907, %r891, %r877, %r879;
mov.b32 %f384, %r909;
add.ftz.f32 %f64, %f383, %f384;
@%p22 bra $L__BB197_38;

add.s32 %r910, %r36, 1;
cvt.u64.u32 %rd143, %r910;
add.s64 %rd144, %rd143, %rd16;
shl.b64 %rd146, %rd144, 2;
add.s64 %rd147, %rd108, %rd146;
st.global.f32 [%rd147], %f64;

$L__BB197_38:
ret;

}
.entry _Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj(
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2,
.param .align 8 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3[48],
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5,
.param .align 4 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6[12],
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15
)
.maxntid 256, 1, 1
.minnctapersm 1
{
.local .align 8 .b8 __local_depot198[16];
.reg .b64 %SP;
.reg .b64 %SPL;
.reg .pred %p<48>;
.reg .b16 %rs<37>;
.reg .f32 %f<169>;
.reg .b32 %r<855>;
.reg .b64 %rd<134>;


mov.u64 %SPL, __local_depot198;
ld.param.u64 %rd21, [_Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0];
ld.param.u64 %rd22, [_Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1];
ld.param.u64 %rd23, [_Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2];
ld.param.u32 %r36, [_Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5];
ld.param.u32 %r29, [_Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8];
ld.param.u32 %r30, [_Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9];
ld.param.u32 %r32, [_Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11];
ld.param.u32 %r33, [_Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12];
ld.param.u32 %r37, [_Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14];
ld.param.u32 %r35, [_Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15];
ld.param.u32 %r27, [_Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+8];
ld.param.u32 %r26, [_Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+4];
ld.param.u32 %r25, [_Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6];
ld.param.u64 %rd25, [_Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+8];
add.u64 %rd1, %SPL, 0;
add.u64 %rd2, %SPL, 8;
shr.u32 %r1, %r36, 8;
mov.u32 %r2, %tid.y;
setp.ge.u32 %p1, %r2, %r37;
@%p1 bra $L__BB198_29;

mov.u32 %r3, %ctaid.y;
cvta.to.global.u64 %rd32, %rd23;
mad.lo.s32 %r38, %r2, %r35, %r3;
mul.wide.u32 %rd33, %r38, 4;
add.s64 %rd34, %rd32, %rd33;
ld.global.nc.u32 %r4, [%rd34];
mov.f32 %f162, 0f00000000;
st.local.v2.f32 [%rd1], {%f162, %f162};
st.local.v2.f32 [%rd2], {%f162, %f162};
mov.u32 %r5, %tid.x;
shr.u32 %r854, %r5, 3;
setp.ge.u32 %p2, %r854, %r1;
mov.f32 %f163, %f162;
mov.f32 %f164, %f162;
mov.f32 %f165, %f162;
@%p2 bra $L__BB198_9;

mov.u32 %r39, %ctaid.x;
shl.b32 %r40, %r39, 1;
mul.lo.s32 %r41, %r4, %r32;
and.b32 %r42, %r5, 7;
cvt.u64.u32 %rd5, %r42;
shl.b32 %r853, %r854, 3;
or.b32 %r43, %r40, 1;
add.s32 %r44, %r854, %r41;
mad.lo.s32 %r45, %r29, %r43, %r44;
mul.wide.s32 %rd6, %r45, 66;
mad.lo.s32 %r46, %r40, %r29, %r854;
add.s32 %r47, %r46, %r41;
mul.wide.s32 %rd7, %r47, 66;
mul.wide.u32 %rd35, %r42, 8;
add.s64 %rd36, %rd35, %rd7;
add.s64 %rd8, %rd36, 8;
add.s64 %rd37, %rd35, %rd6;
add.s64 %rd9, %rd37, 4;
cvta.to.global.u64 %rd132, %rd21;
cvta.to.global.u64 %rd133, %rd25;
mul.lo.s32 %r48, %r2, %r30;
cvt.u64.u32 %rd38, %r48;
mul.hi.u32 %r49, %r3, %r25;
add.s32 %r50, %r49, %r3;
shr.u32 %r51, %r50, %r26;
mul.lo.s32 %r52, %r51, %r27;
sub.s32 %r53, %r3, %r52;
mul.lo.s32 %r54, %r53, %r33;
cvt.u64.u32 %rd39, %r54;
add.s64 %rd12, %rd39, %rd38;
cvta.to.global.u64 %rd13, %rd22;
mov.f32 %f162, 0f00000000;

$L__BB198_3:
cvt.s64.s32 %rd40, %r853;
add.s64 %rd41, %rd12, %rd40;
add.s64 %rd42, %rd41, %rd5;
mul.lo.s64 %rd43, %rd42, 36;
add.s64 %rd44, %rd132, %rd8;
ld.global.nc.u16 %rs2, [%rd44+-6];
shr.u16 %rs3, %rs2, 8;
ld.global.nc.u16 %rs4, [%rd44+-4];
shr.u16 %rs5, %rs4, 8;
ld.global.nc.u16 %rs6, [%rd44+-2];
cvt.u32.u16 %r104, %rs6;
ld.global.nc.u16 %rs7, [%rd44];
cvt.u32.u16 %r105, %rs7;
mov.b32 %r106, {%rs6, %rs7};
shl.b16 %rs8, %rs2, 3;
cvt.u64.u16 %rd45, %rs8;
and.b64 %rd46, %rd45, 2040;
mov.u64 %rd47, iq2xxs_grid;
add.s64 %rd48, %rd47, %rd46;
ld.global.nc.v2.u32 {%r107, %r108}, [%rd48];
mov.u32 %r60, 0;
and.b32 %r111, %r104, 255;
popc.b32 %r112, %r111;
shl.b32 %r113, %r112, 7;
and.b32 %r114, %r113, 128;
xor.b32 %r115, %r114, %r111;
mul.lo.s32 %r116, %r115, 16843009;
and.b32 %r117, %r116, 134480385;
or.b32 %r118, %r117, -2139062144;
add.s32 %r119, %r118, -16843009;
or.b32 %r55, %r119, %r117;

prmt.b32 %r55, %r55, 0, 0xba98;

xor.b32 %r120, %r107, %r55;
or.b32 %r121, %r120, -2139062144;
and.b32 %r122, %r55, 2139062143;
and.b32 %r123, %r107, -2139062144;
sub.s32 %r124, %r121, %r122;
xor.b32 %r125, %r123, %r124;
xor.b32 %r58, %r125, -2139062144;
add.s64 %rd49, %rd13, %rd43;
ld.global.nc.u32 %r10, [%rd49+4];

dp4a.s32.s32 %r57, %r58, %r10, %r60;

and.b32 %r126, %r116, -2143281136;
or.b32 %r127, %r126, -2139062144;
add.s32 %r128, %r127, -16843009;
or.b32 %r61, %r128, %r126;

prmt.b32 %r61, %r61, 0, 0xba98;

xor.b32 %r129, %r61, %r108;
or.b32 %r130, %r129, -2139062144;
and.b32 %r131, %r61, 2139062143;
and.b32 %r132, %r108, -2139062144;
sub.s32 %r133, %r130, %r131;
xor.b32 %r134, %r132, %r133;
xor.b32 %r64, %r134, -2139062144;
ld.global.nc.u32 %r11, [%rd49+8];

dp4a.s32.s32 %r63, %r64, %r11, %r57;

cvt.u32.u16 %r135, %rs3;
mul.wide.u32 %rd50, %r135, 8;
add.s64 %rd51, %rd47, %rd50;
ld.global.nc.v2.u32 {%r136, %r137}, [%rd51];
shr.u32 %r140, %r104, 7;
and.b32 %r141, %r140, 255;
popc.b32 %r142, %r141;
shl.b32 %r143, %r142, 7;
and.b32 %r144, %r143, 128;
xor.b32 %r145, %r144, %r141;
mul.lo.s32 %r146, %r145, 16843009;
and.b32 %r147, %r146, 134480385;
or.b32 %r148, %r147, -2139062144;
add.s32 %r149, %r148, -16843009;
or.b32 %r67, %r149, %r147;

prmt.b32 %r67, %r67, 0, 0xba98;

xor.b32 %r150, %r136, %r67;
or.b32 %r151, %r150, -2139062144;
and.b32 %r152, %r67, 2139062143;
and.b32 %r153, %r136, -2139062144;
sub.s32 %r154, %r151, %r152;
xor.b32 %r155, %r153, %r154;
xor.b32 %r70, %r155, -2139062144;
ld.global.nc.u32 %r12, [%rd49+12];

dp4a.s32.s32 %r69, %r70, %r12, %r63;

and.b32 %r156, %r146, -2143281136;
or.b32 %r157, %r156, -2139062144;
add.s32 %r158, %r157, -16843009;
or.b32 %r73, %r158, %r156;

prmt.b32 %r73, %r73, 0, 0xba98;

xor.b32 %r159, %r73, %r137;
or.b32 %r160, %r159, -2139062144;
and.b32 %r161, %r73, 2139062143;
and.b32 %r162, %r137, -2139062144;
sub.s32 %r163, %r160, %r161;
xor.b32 %r164, %r162, %r163;
xor.b32 %r76, %r164, -2139062144;
ld.global.nc.u32 %r13, [%rd49+16];

dp4a.s32.s32 %r75, %r76, %r13, %r69;

shl.b16 %rs9, %rs4, 3;
cvt.u64.u16 %rd52, %rs9;
and.b64 %rd53, %rd52, 2040;
add.s64 %rd54, %rd47, %rd53;
ld.global.nc.v2.u32 {%r165, %r166}, [%rd54];
shr.u32 %r169, %r106, 14;
and.b32 %r170, %r169, 255;
popc.b32 %r171, %r170;
shl.b32 %r172, %r171, 7;
and.b32 %r173, %r172, 128;
xor.b32 %r174, %r173, %r170;
mul.lo.s32 %r175, %r174, 16843009;
and.b32 %r176, %r175, 134480385;
or.b32 %r177, %r176, -2139062144;
add.s32 %r178, %r177, -16843009;
or.b32 %r79, %r178, %r176;

prmt.b32 %r79, %r79, 0, 0xba98;

xor.b32 %r179, %r165, %r79;
or.b32 %r180, %r179, -2139062144;
and.b32 %r181, %r79, 2139062143;
and.b32 %r182, %r165, -2139062144;
sub.s32 %r183, %r180, %r181;
xor.b32 %r184, %r182, %r183;
xor.b32 %r82, %r184, -2139062144;
ld.global.nc.u32 %r14, [%rd49+20];

dp4a.s32.s32 %r81, %r82, %r14, %r75;

and.b32 %r185, %r175, -2143281136;
or.b32 %r186, %r185, -2139062144;
add.s32 %r187, %r186, -16843009;
or.b32 %r85, %r187, %r185;

prmt.b32 %r85, %r85, 0, 0xba98;

xor.b32 %r188, %r85, %r166;
or.b32 %r189, %r188, -2139062144;
and.b32 %r190, %r85, 2139062143;
and.b32 %r191, %r166, -2139062144;
sub.s32 %r192, %r189, %r190;
xor.b32 %r193, %r191, %r192;
xor.b32 %r88, %r193, -2139062144;
ld.global.nc.u32 %r15, [%rd49+24];

dp4a.s32.s32 %r87, %r88, %r15, %r81;

cvt.u32.u16 %r194, %rs5;
mul.wide.u32 %rd55, %r194, 8;
add.s64 %rd56, %rd47, %rd55;
ld.global.nc.v2.u32 {%r195, %r196}, [%rd56];
shr.u32 %r199, %r105, 5;
and.b32 %r200, %r199, 255;
popc.b32 %r201, %r200;
shl.b32 %r202, %r201, 7;
and.b32 %r203, %r202, 128;
xor.b32 %r204, %r203, %r200;
mul.lo.s32 %r205, %r204, 16843009;
and.b32 %r206, %r205, 134480385;
or.b32 %r207, %r206, -2139062144;
add.s32 %r208, %r207, -16843009;
or.b32 %r91, %r208, %r206;

prmt.b32 %r91, %r91, 0, 0xba98;

xor.b32 %r209, %r195, %r91;
or.b32 %r210, %r209, -2139062144;
and.b32 %r211, %r91, 2139062143;
and.b32 %r212, %r195, -2139062144;
sub.s32 %r213, %r210, %r211;
xor.b32 %r214, %r212, %r213;
xor.b32 %r94, %r214, -2139062144;
ld.global.nc.u32 %r16, [%rd49+28];

dp4a.s32.s32 %r93, %r94, %r16, %r87;

and.b32 %r215, %r205, -2143281136;
or.b32 %r216, %r215, -2139062144;
add.s32 %r217, %r216, -16843009;
or.b32 %r97, %r217, %r215;

prmt.b32 %r97, %r97, 0, 0xba98;

xor.b32 %r218, %r97, %r196;
or.b32 %r219, %r218, -2139062144;
and.b32 %r220, %r97, 2139062143;
and.b32 %r221, %r196, -2139062144;
sub.s32 %r222, %r219, %r220;
xor.b32 %r223, %r221, %r222;
xor.b32 %r100, %r223, -2139062144;
ld.global.nc.u32 %r17, [%rd49+32];

dp4a.s32.s32 %r99, %r100, %r17, %r93;

shr.u32 %r224, %r105, 11;
or.b32 %r225, %r224, 1;
mul.lo.s32 %r226, %r99, %r225;
shr.s32 %r227, %r226, 31;
shr.u32 %r228, %r227, 29;
add.s32 %r229, %r226, %r228;
shr.s32 %r230, %r229, 3;
add.s64 %rd57, %rd132, %rd7;
ld.global.nc.u16 %rs1, [%rd57];

{ cvt.f32.f16 %f42, %rs1;}


ld.global.nc.u32 %r18, [%rd49];

{.reg .f16 low,high;
mov.b32 {low,high},%r18;
cvt.f32.f16 %f43, low;}


mul.ftz.f32 %f44, %f42, %f43;
cvt.rn.f32.s32 %f45, %r230;
fma.rn.ftz.f32 %f165, %f44, %f45, %f165;
setp.eq.s64 %p3, %rd25, 0;
@%p3 bra $L__BB198_5;

mov.u64 %rd116, iq2xxs_grid;
add.s64 %rd58, %rd133, %rd8;
ld.global.u16 %rs11, [%rd58+-6];
shr.u16 %rs12, %rs11, 8;
ld.global.u16 %rs13, [%rd58+-4];
shr.u16 %rs14, %rs13, 8;
ld.global.u16 %rs15, [%rd58+-2];
cvt.u32.u16 %r280, %rs15;
ld.global.u16 %rs16, [%rd58];
cvt.u32.u16 %r281, %rs16;
mov.b32 %r282, {%rs15, %rs16};
shl.b16 %rs17, %rs11, 3;
cvt.u64.u16 %rd59, %rs17;
and.b64 %rd60, %rd59, 2040;
add.s64 %rd62, %rd116, %rd60;
ld.global.nc.v2.u32 {%r283, %r284}, [%rd62];
and.b32 %r287, %r280, 255;
popc.b32 %r288, %r287;
shl.b32 %r289, %r288, 7;
and.b32 %r290, %r289, 128;
xor.b32 %r291, %r290, %r287;
mul.lo.s32 %r292, %r291, 16843009;
and.b32 %r293, %r292, 134480385;
or.b32 %r294, %r293, -2139062144;
add.s32 %r295, %r294, -16843009;
or.b32 %r231, %r295, %r293;

prmt.b32 %r231, %r231, 0, 0xba98;

xor.b32 %r296, %r283, %r231;
or.b32 %r297, %r296, -2139062144;
and.b32 %r298, %r231, 2139062143;
and.b32 %r299, %r283, -2139062144;
sub.s32 %r300, %r297, %r298;
xor.b32 %r301, %r299, %r300;
xor.b32 %r234, %r301, -2139062144;

dp4a.s32.s32 %r233, %r234, %r10, %r60;

and.b32 %r302, %r292, -2143281136;
or.b32 %r303, %r302, -2139062144;
add.s32 %r304, %r303, -16843009;
or.b32 %r237, %r304, %r302;

prmt.b32 %r237, %r237, 0, 0xba98;

xor.b32 %r305, %r237, %r284;
or.b32 %r306, %r305, -2139062144;
and.b32 %r307, %r237, 2139062143;
and.b32 %r308, %r284, -2139062144;
sub.s32 %r309, %r306, %r307;
xor.b32 %r310, %r308, %r309;
xor.b32 %r240, %r310, -2139062144;

dp4a.s32.s32 %r239, %r240, %r11, %r233;

cvt.u32.u16 %r311, %rs12;
mul.wide.u32 %rd63, %r311, 8;
add.s64 %rd64, %rd116, %rd63;
ld.global.nc.v2.u32 {%r312, %r313}, [%rd64];
shr.u32 %r316, %r280, 7;
and.b32 %r317, %r316, 255;
popc.b32 %r318, %r317;
shl.b32 %r319, %r318, 7;
and.b32 %r320, %r319, 128;
xor.b32 %r321, %r320, %r317;
mul.lo.s32 %r322, %r321, 16843009;
and.b32 %r323, %r322, 134480385;
or.b32 %r324, %r323, -2139062144;
add.s32 %r325, %r324, -16843009;
or.b32 %r243, %r325, %r323;

prmt.b32 %r243, %r243, 0, 0xba98;

xor.b32 %r326, %r312, %r243;
or.b32 %r327, %r326, -2139062144;
and.b32 %r328, %r243, 2139062143;
and.b32 %r329, %r312, -2139062144;
sub.s32 %r330, %r327, %r328;
xor.b32 %r331, %r329, %r330;
xor.b32 %r246, %r331, -2139062144;

dp4a.s32.s32 %r245, %r246, %r12, %r239;

and.b32 %r332, %r322, -2143281136;
or.b32 %r333, %r332, -2139062144;
add.s32 %r334, %r333, -16843009;
or.b32 %r249, %r334, %r332;

prmt.b32 %r249, %r249, 0, 0xba98;

xor.b32 %r335, %r249, %r313;
or.b32 %r336, %r335, -2139062144;
and.b32 %r337, %r249, 2139062143;
and.b32 %r338, %r313, -2139062144;
sub.s32 %r339, %r336, %r337;
xor.b32 %r340, %r338, %r339;
xor.b32 %r252, %r340, -2139062144;

dp4a.s32.s32 %r251, %r252, %r13, %r245;

shl.b16 %rs18, %rs13, 3;
cvt.u64.u16 %rd65, %rs18;
and.b64 %rd66, %rd65, 2040;
add.s64 %rd67, %rd116, %rd66;
ld.global.nc.v2.u32 {%r341, %r342}, [%rd67];
shr.u32 %r345, %r282, 14;
and.b32 %r346, %r345, 255;
popc.b32 %r347, %r346;
shl.b32 %r348, %r347, 7;
and.b32 %r349, %r348, 128;
xor.b32 %r350, %r349, %r346;
mul.lo.s32 %r351, %r350, 16843009;
and.b32 %r352, %r351, 134480385;
or.b32 %r353, %r352, -2139062144;
add.s32 %r354, %r353, -16843009;
or.b32 %r255, %r354, %r352;

prmt.b32 %r255, %r255, 0, 0xba98;

xor.b32 %r355, %r341, %r255;
or.b32 %r356, %r355, -2139062144;
and.b32 %r357, %r255, 2139062143;
and.b32 %r358, %r341, -2139062144;
sub.s32 %r359, %r356, %r357;
xor.b32 %r360, %r358, %r359;
xor.b32 %r258, %r360, -2139062144;

dp4a.s32.s32 %r257, %r258, %r14, %r251;

and.b32 %r361, %r351, -2143281136;
or.b32 %r362, %r361, -2139062144;
add.s32 %r363, %r362, -16843009;
or.b32 %r261, %r363, %r361;

prmt.b32 %r261, %r261, 0, 0xba98;

xor.b32 %r364, %r261, %r342;
or.b32 %r365, %r364, -2139062144;
and.b32 %r366, %r261, 2139062143;
and.b32 %r367, %r342, -2139062144;
sub.s32 %r368, %r365, %r366;
xor.b32 %r369, %r367, %r368;
xor.b32 %r264, %r369, -2139062144;

dp4a.s32.s32 %r263, %r264, %r15, %r257;

cvt.u32.u16 %r370, %rs14;
mul.wide.u32 %rd68, %r370, 8;
add.s64 %rd69, %rd116, %rd68;
ld.global.nc.v2.u32 {%r371, %r372}, [%rd69];
shr.u32 %r375, %r281, 5;
and.b32 %r376, %r375, 255;
popc.b32 %r377, %r376;
shl.b32 %r378, %r377, 7;
and.b32 %r379, %r378, 128;
xor.b32 %r380, %r379, %r376;
mul.lo.s32 %r381, %r380, 16843009;
and.b32 %r382, %r381, 134480385;
or.b32 %r383, %r382, -2139062144;
add.s32 %r384, %r383, -16843009;
or.b32 %r267, %r384, %r382;

prmt.b32 %r267, %r267, 0, 0xba98;

xor.b32 %r385, %r371, %r267;
or.b32 %r386, %r385, -2139062144;
and.b32 %r387, %r267, 2139062143;
and.b32 %r388, %r371, -2139062144;
sub.s32 %r389, %r386, %r387;
xor.b32 %r390, %r388, %r389;
xor.b32 %r270, %r390, -2139062144;

dp4a.s32.s32 %r269, %r270, %r16, %r263;

and.b32 %r391, %r381, -2143281136;
or.b32 %r392, %r391, -2139062144;
add.s32 %r393, %r392, -16843009;
or.b32 %r273, %r393, %r391;

prmt.b32 %r273, %r273, 0, 0xba98;

xor.b32 %r394, %r273, %r372;
or.b32 %r395, %r394, -2139062144;
and.b32 %r396, %r273, 2139062143;
and.b32 %r397, %r372, -2139062144;
sub.s32 %r398, %r395, %r396;
xor.b32 %r399, %r397, %r398;
xor.b32 %r276, %r399, -2139062144;

dp4a.s32.s32 %r275, %r276, %r17, %r269;

shr.u32 %r400, %r281, 11;
or.b32 %r401, %r400, 1;
mul.lo.s32 %r402, %r275, %r401;
shr.s32 %r403, %r402, 31;
shr.u32 %r404, %r403, 29;
add.s32 %r405, %r402, %r404;
shr.s32 %r406, %r405, 3;
add.s64 %rd70, %rd133, %rd7;
ld.global.u16 %rs10, [%rd70];

{ cvt.f32.f16 %f46, %rs10;}



{.reg .f16 low,high;
mov.b32 {low,high},%r18;
cvt.f32.f16 %f47, low;}


mul.ftz.f32 %f48, %f46, %f47;
cvt.rn.f32.s32 %f49, %r406;
fma.rn.ftz.f32 %f163, %f48, %f49, %f163;

$L__BB198_5:
mov.u64 %rd110, iq2xxs_grid;
add.s64 %rd71, %rd132, %rd9;
ld.global.nc.u16 %rs20, [%rd71+-2];
shr.u16 %rs21, %rs20, 8;
ld.global.nc.u16 %rs22, [%rd71];
shr.u16 %rs23, %rs22, 8;
ld.global.nc.u16 %rs24, [%rd71+2];
cvt.u32.u16 %r456, %rs24;
ld.global.nc.u16 %rs25, [%rd71+4];
cvt.u32.u16 %r457, %rs25;
mov.b32 %r458, {%rs24, %rs25};
shl.b16 %rs26, %rs20, 3;
cvt.u64.u16 %rd72, %rs26;
and.b64 %rd73, %rd72, 2040;
add.s64 %rd75, %rd110, %rd73;
ld.global.nc.v2.u32 {%r459, %r460}, [%rd75];
and.b32 %r463, %r456, 255;
popc.b32 %r464, %r463;
shl.b32 %r465, %r464, 7;
and.b32 %r466, %r465, 128;
xor.b32 %r467, %r466, %r463;
mul.lo.s32 %r468, %r467, 16843009;
and.b32 %r469, %r468, 134480385;
or.b32 %r470, %r469, -2139062144;
add.s32 %r471, %r470, -16843009;
or.b32 %r407, %r471, %r469;

prmt.b32 %r407, %r407, 0, 0xba98;

xor.b32 %r472, %r459, %r407;
or.b32 %r473, %r472, -2139062144;
and.b32 %r474, %r407, 2139062143;
and.b32 %r475, %r459, -2139062144;
sub.s32 %r476, %r473, %r474;
xor.b32 %r477, %r475, %r476;
xor.b32 %r410, %r477, -2139062144;

dp4a.s32.s32 %r409, %r410, %r10, %r60;

and.b32 %r478, %r468, -2143281136;
or.b32 %r479, %r478, -2139062144;
add.s32 %r480, %r479, -16843009;
or.b32 %r413, %r480, %r478;

prmt.b32 %r413, %r413, 0, 0xba98;

xor.b32 %r481, %r413, %r460;
or.b32 %r482, %r481, -2139062144;
and.b32 %r483, %r413, 2139062143;
and.b32 %r484, %r460, -2139062144;
sub.s32 %r485, %r482, %r483;
xor.b32 %r486, %r484, %r485;
xor.b32 %r416, %r486, -2139062144;

dp4a.s32.s32 %r415, %r416, %r11, %r409;

cvt.u32.u16 %r487, %rs21;
mul.wide.u32 %rd76, %r487, 8;
add.s64 %rd77, %rd110, %rd76;
ld.global.nc.v2.u32 {%r488, %r489}, [%rd77];
shr.u32 %r492, %r456, 7;
and.b32 %r493, %r492, 255;
popc.b32 %r494, %r493;
shl.b32 %r495, %r494, 7;
and.b32 %r496, %r495, 128;
xor.b32 %r497, %r496, %r493;
mul.lo.s32 %r498, %r497, 16843009;
and.b32 %r499, %r498, 134480385;
or.b32 %r500, %r499, -2139062144;
add.s32 %r501, %r500, -16843009;
or.b32 %r419, %r501, %r499;

prmt.b32 %r419, %r419, 0, 0xba98;

xor.b32 %r502, %r488, %r419;
or.b32 %r503, %r502, -2139062144;
and.b32 %r504, %r419, 2139062143;
and.b32 %r505, %r488, -2139062144;
sub.s32 %r506, %r503, %r504;
xor.b32 %r507, %r505, %r506;
xor.b32 %r422, %r507, -2139062144;

dp4a.s32.s32 %r421, %r422, %r12, %r415;

and.b32 %r508, %r498, -2143281136;
or.b32 %r509, %r508, -2139062144;
add.s32 %r510, %r509, -16843009;
or.b32 %r425, %r510, %r508;

prmt.b32 %r425, %r425, 0, 0xba98;

xor.b32 %r511, %r425, %r489;
or.b32 %r512, %r511, -2139062144;
and.b32 %r513, %r425, 2139062143;
and.b32 %r514, %r489, -2139062144;
sub.s32 %r515, %r512, %r513;
xor.b32 %r516, %r514, %r515;
xor.b32 %r428, %r516, -2139062144;

dp4a.s32.s32 %r427, %r428, %r13, %r421;

shl.b16 %rs27, %rs22, 3;
cvt.u64.u16 %rd78, %rs27;
and.b64 %rd79, %rd78, 2040;
add.s64 %rd80, %rd110, %rd79;
ld.global.nc.v2.u32 {%r517, %r518}, [%rd80];
shr.u32 %r521, %r458, 14;
and.b32 %r522, %r521, 255;
popc.b32 %r523, %r522;
shl.b32 %r524, %r523, 7;
and.b32 %r525, %r524, 128;
xor.b32 %r526, %r525, %r522;
mul.lo.s32 %r527, %r526, 16843009;
and.b32 %r528, %r527, 134480385;
or.b32 %r529, %r528, -2139062144;
add.s32 %r530, %r529, -16843009;
or.b32 %r431, %r530, %r528;

prmt.b32 %r431, %r431, 0, 0xba98;

xor.b32 %r531, %r517, %r431;
or.b32 %r532, %r531, -2139062144;
and.b32 %r533, %r431, 2139062143;
and.b32 %r534, %r517, -2139062144;
sub.s32 %r535, %r532, %r533;
xor.b32 %r536, %r534, %r535;
xor.b32 %r434, %r536, -2139062144;

dp4a.s32.s32 %r433, %r434, %r14, %r427;

and.b32 %r537, %r527, -2143281136;
or.b32 %r538, %r537, -2139062144;
add.s32 %r539, %r538, -16843009;
or.b32 %r437, %r539, %r537;

prmt.b32 %r437, %r437, 0, 0xba98;

xor.b32 %r540, %r437, %r518;
or.b32 %r541, %r540, -2139062144;
and.b32 %r542, %r437, 2139062143;
and.b32 %r543, %r518, -2139062144;
sub.s32 %r544, %r541, %r542;
xor.b32 %r545, %r543, %r544;
xor.b32 %r440, %r545, -2139062144;

dp4a.s32.s32 %r439, %r440, %r15, %r433;

cvt.u32.u16 %r546, %rs23;
mul.wide.u32 %rd81, %r546, 8;
add.s64 %rd82, %rd110, %rd81;
ld.global.nc.v2.u32 {%r547, %r548}, [%rd82];
shr.u32 %r551, %r457, 5;
and.b32 %r552, %r551, 255;
popc.b32 %r553, %r552;
shl.b32 %r554, %r553, 7;
and.b32 %r555, %r554, 128;
xor.b32 %r556, %r555, %r552;
mul.lo.s32 %r557, %r556, 16843009;
and.b32 %r558, %r557, 134480385;
or.b32 %r559, %r558, -2139062144;
add.s32 %r560, %r559, -16843009;
or.b32 %r443, %r560, %r558;

prmt.b32 %r443, %r443, 0, 0xba98;

xor.b32 %r561, %r547, %r443;
or.b32 %r562, %r561, -2139062144;
and.b32 %r563, %r443, 2139062143;
and.b32 %r564, %r547, -2139062144;
sub.s32 %r565, %r562, %r563;
xor.b32 %r566, %r564, %r565;
xor.b32 %r446, %r566, -2139062144;

dp4a.s32.s32 %r445, %r446, %r16, %r439;

and.b32 %r567, %r557, -2143281136;
or.b32 %r568, %r567, -2139062144;
add.s32 %r569, %r568, -16843009;
or.b32 %r449, %r569, %r567;

prmt.b32 %r449, %r449, 0, 0xba98;

xor.b32 %r570, %r449, %r548;
or.b32 %r571, %r570, -2139062144;
and.b32 %r572, %r449, 2139062143;
and.b32 %r573, %r548, -2139062144;
sub.s32 %r574, %r571, %r572;
xor.b32 %r575, %r573, %r574;
xor.b32 %r452, %r575, -2139062144;

dp4a.s32.s32 %r451, %r452, %r17, %r445;

shr.u32 %r576, %r457, 11;
or.b32 %r577, %r576, 1;
mul.lo.s32 %r578, %r451, %r577;
shr.s32 %r579, %r578, 31;
shr.u32 %r580, %r579, 29;
add.s32 %r581, %r578, %r580;
shr.s32 %r582, %r581, 3;
add.s64 %rd83, %rd132, %rd6;
ld.global.nc.u16 %rs19, [%rd83];

{ cvt.f32.f16 %f50, %rs19;}



{.reg .f16 low,high;
mov.b32 {low,high},%r18;
cvt.f32.f16 %f51, low;}


mul.ftz.f32 %f52, %f50, %f51;
cvt.rn.f32.s32 %f53, %r582;
fma.rn.ftz.f32 %f164, %f52, %f53, %f164;
@%p3 bra $L__BB198_7;

mov.u64 %rd117, iq2xxs_grid;
add.s64 %rd84, %rd133, %rd9;
ld.global.u16 %rs29, [%rd84+-2];
shr.u16 %rs30, %rs29, 8;
ld.global.u16 %rs31, [%rd84];
shr.u16 %rs32, %rs31, 8;
ld.global.u16 %rs33, [%rd84+2];
cvt.u32.u16 %r632, %rs33;
ld.global.u16 %rs34, [%rd84+4];
cvt.u32.u16 %r633, %rs34;
mov.b32 %r634, {%rs33, %rs34};
shl.b16 %rs35, %rs29, 3;
cvt.u64.u16 %rd85, %rs35;
and.b64 %rd86, %rd85, 2040;
add.s64 %rd88, %rd117, %rd86;
ld.global.nc.v2.u32 {%r635, %r636}, [%rd88];
mov.u32 %r588, 0;
and.b32 %r639, %r632, 255;
popc.b32 %r640, %r639;
shl.b32 %r641, %r640, 7;
and.b32 %r642, %r641, 128;
xor.b32 %r643, %r642, %r639;
mul.lo.s32 %r644, %r643, 16843009;
and.b32 %r645, %r644, 134480385;
or.b32 %r646, %r645, -2139062144;
add.s32 %r647, %r646, -16843009;
or.b32 %r583, %r647, %r645;

prmt.b32 %r583, %r583, 0, 0xba98;

xor.b32 %r648, %r635, %r583;
or.b32 %r649, %r648, -2139062144;
and.b32 %r650, %r583, 2139062143;
and.b32 %r651, %r635, -2139062144;
sub.s32 %r652, %r649, %r650;
xor.b32 %r653, %r651, %r652;
xor.b32 %r586, %r653, -2139062144;

dp4a.s32.s32 %r585, %r586, %r10, %r588;

and.b32 %r654, %r644, -2143281136;
or.b32 %r655, %r654, -2139062144;
add.s32 %r656, %r655, -16843009;
or.b32 %r589, %r656, %r654;

prmt.b32 %r589, %r589, 0, 0xba98;

xor.b32 %r657, %r589, %r636;
or.b32 %r658, %r657, -2139062144;
and.b32 %r659, %r589, 2139062143;
and.b32 %r660, %r636, -2139062144;
sub.s32 %r661, %r658, %r659;
xor.b32 %r662, %r660, %r661;
xor.b32 %r592, %r662, -2139062144;

dp4a.s32.s32 %r591, %r592, %r11, %r585;

cvt.u32.u16 %r663, %rs30;
mul.wide.u32 %rd89, %r663, 8;
add.s64 %rd90, %rd117, %rd89;
ld.global.nc.v2.u32 {%r664, %r665}, [%rd90];
shr.u32 %r668, %r632, 7;
and.b32 %r669, %r668, 255;
popc.b32 %r670, %r669;
shl.b32 %r671, %r670, 7;
and.b32 %r672, %r671, 128;
xor.b32 %r673, %r672, %r669;
mul.lo.s32 %r674, %r673, 16843009;
and.b32 %r675, %r674, 134480385;
or.b32 %r676, %r675, -2139062144;
add.s32 %r677, %r676, -16843009;
or.b32 %r595, %r677, %r675;

prmt.b32 %r595, %r595, 0, 0xba98;

xor.b32 %r678, %r664, %r595;
or.b32 %r679, %r678, -2139062144;
and.b32 %r680, %r595, 2139062143;
and.b32 %r681, %r664, -2139062144;
sub.s32 %r682, %r679, %r680;
xor.b32 %r683, %r681, %r682;
xor.b32 %r598, %r683, -2139062144;

dp4a.s32.s32 %r597, %r598, %r12, %r591;

and.b32 %r684, %r674, -2143281136;
or.b32 %r685, %r684, -2139062144;
add.s32 %r686, %r685, -16843009;
or.b32 %r601, %r686, %r684;

prmt.b32 %r601, %r601, 0, 0xba98;

xor.b32 %r687, %r601, %r665;
or.b32 %r688, %r687, -2139062144;
and.b32 %r689, %r601, 2139062143;
and.b32 %r690, %r665, -2139062144;
sub.s32 %r691, %r688, %r689;
xor.b32 %r692, %r690, %r691;
xor.b32 %r604, %r692, -2139062144;

dp4a.s32.s32 %r603, %r604, %r13, %r597;

shl.b16 %rs36, %rs31, 3;
cvt.u64.u16 %rd91, %rs36;
and.b64 %rd92, %rd91, 2040;
add.s64 %rd93, %rd117, %rd92;
ld.global.nc.v2.u32 {%r693, %r694}, [%rd93];
shr.u32 %r697, %r634, 14;
and.b32 %r698, %r697, 255;
popc.b32 %r699, %r698;
shl.b32 %r700, %r699, 7;
and.b32 %r701, %r700, 128;
xor.b32 %r702, %r701, %r698;
mul.lo.s32 %r703, %r702, 16843009;
and.b32 %r704, %r703, 134480385;
or.b32 %r705, %r704, -2139062144;
add.s32 %r706, %r705, -16843009;
or.b32 %r607, %r706, %r704;

prmt.b32 %r607, %r607, 0, 0xba98;

xor.b32 %r707, %r693, %r607;
or.b32 %r708, %r707, -2139062144;
and.b32 %r709, %r607, 2139062143;
and.b32 %r710, %r693, -2139062144;
sub.s32 %r711, %r708, %r709;
xor.b32 %r712, %r710, %r711;
xor.b32 %r610, %r712, -2139062144;

dp4a.s32.s32 %r609, %r610, %r14, %r603;

and.b32 %r713, %r703, -2143281136;
or.b32 %r714, %r713, -2139062144;
add.s32 %r715, %r714, -16843009;
or.b32 %r613, %r715, %r713;

prmt.b32 %r613, %r613, 0, 0xba98;

xor.b32 %r716, %r613, %r694;
or.b32 %r717, %r716, -2139062144;
and.b32 %r718, %r613, 2139062143;
and.b32 %r719, %r694, -2139062144;
sub.s32 %r720, %r717, %r718;
xor.b32 %r721, %r719, %r720;
xor.b32 %r616, %r721, -2139062144;

dp4a.s32.s32 %r615, %r616, %r15, %r609;

cvt.u32.u16 %r722, %rs32;
mul.wide.u32 %rd94, %r722, 8;
add.s64 %rd95, %rd117, %rd94;
ld.global.nc.v2.u32 {%r723, %r724}, [%rd95];
shr.u32 %r727, %r633, 5;
and.b32 %r728, %r727, 255;
popc.b32 %r729, %r728;
shl.b32 %r730, %r729, 7;
and.b32 %r731, %r730, 128;
xor.b32 %r732, %r731, %r728;
mul.lo.s32 %r733, %r732, 16843009;
and.b32 %r734, %r733, 134480385;
or.b32 %r735, %r734, -2139062144;
add.s32 %r736, %r735, -16843009;
or.b32 %r619, %r736, %r734;

prmt.b32 %r619, %r619, 0, 0xba98;

xor.b32 %r737, %r723, %r619;
or.b32 %r738, %r737, -2139062144;
and.b32 %r739, %r619, 2139062143;
and.b32 %r740, %r723, -2139062144;
sub.s32 %r741, %r738, %r739;
xor.b32 %r742, %r740, %r741;
xor.b32 %r622, %r742, -2139062144;

dp4a.s32.s32 %r621, %r622, %r16, %r615;

and.b32 %r743, %r733, -2143281136;
or.b32 %r744, %r743, -2139062144;
add.s32 %r745, %r744, -16843009;
or.b32 %r625, %r745, %r743;

prmt.b32 %r625, %r625, 0, 0xba98;

xor.b32 %r746, %r625, %r724;
or.b32 %r747, %r746, -2139062144;
and.b32 %r748, %r625, 2139062143;
and.b32 %r749, %r724, -2139062144;
sub.s32 %r750, %r747, %r748;
xor.b32 %r751, %r749, %r750;
xor.b32 %r628, %r751, -2139062144;

dp4a.s32.s32 %r627, %r628, %r17, %r621;

shr.u32 %r752, %r633, 11;
or.b32 %r753, %r752, 1;
mul.lo.s32 %r754, %r627, %r753;
shr.s32 %r755, %r754, 31;
shr.u32 %r756, %r755, 29;
add.s32 %r757, %r754, %r756;
shr.s32 %r758, %r757, 3;
add.s64 %rd96, %rd133, %rd6;
ld.global.u16 %rs28, [%rd96];

{ cvt.f32.f16 %f54, %rs28;}



{.reg .f16 low,high;
mov.b32 {low,high},%r18;
cvt.f32.f16 %f55, low;}


mul.ftz.f32 %f56, %f54, %f55;
cvt.rn.f32.s32 %f57, %r758;
fma.rn.ftz.f32 %f162, %f56, %f57, %f162;

$L__BB198_7:
add.s32 %r853, %r853, 32;
add.s64 %rd133, %rd133, 264;
add.s64 %rd132, %rd132, 264;
add.s32 %r854, %r854, 4;
setp.lt.u32 %p5, %r854, %r1;
@%p5 bra $L__BB198_3;

add.u64 %rd124, %SPL, 0;
add.u64 %rd118, %SPL, 8;
st.local.v2.f32 [%rd124], {%f165, %f164};
st.local.v2.f32 [%rd118], {%f163, %f162};

$L__BB198_9:
mov.b32 %r21, %f165;
setp.ne.s64 %p6, %rd25, 0;
@%p6 bra $L__BB198_11;
bra.uni $L__BB198_10;

$L__BB198_11:
add.u64 %rd130, %SPL, 0;
add.u64 %rd122, %SPL, 8;
mov.u32 %r785, 31;
mov.u32 %r786, 16;
mov.u32 %r787, -1;
shfl.sync.bfly.b32 %r788|%p17, %r21, %r786, %r785, %r787;
mov.b32 %f78, %r788;
add.ftz.f32 %f79, %f165, %f78;
mov.b32 %r789, %f79;
mov.u32 %r790, 8;
shfl.sync.bfly.b32 %r791|%p18, %r789, %r790, %r785, %r787;
mov.b32 %f80, %r791;
add.ftz.f32 %f81, %f79, %f80;
mov.b32 %r792, %f81;
mov.u32 %r793, 4;
shfl.sync.bfly.b32 %r794|%p19, %r792, %r793, %r785, %r787;
mov.b32 %f82, %r794;
add.ftz.f32 %f83, %f81, %f82;
mov.b32 %r795, %f83;
mov.u32 %r796, 2;
shfl.sync.bfly.b32 %r797|%p20, %r795, %r796, %r785, %r787;
mov.b32 %f84, %r797;
add.ftz.f32 %f85, %f83, %f84;
mov.b32 %r798, %f85;
mov.u32 %r799, 1;
shfl.sync.bfly.b32 %r800|%p21, %r798, %r799, %r785, %r787;
mov.b32 %f86, %r800;
add.ftz.f32 %f87, %f85, %f86;
st.local.f32 [%rd130], %f87;
mov.b32 %r801, %f163;
shfl.sync.bfly.b32 %r802|%p22, %r801, %r786, %r785, %r787;
mov.b32 %f88, %r802;
add.ftz.f32 %f89, %f163, %f88;
mov.b32 %r803, %f89;
shfl.sync.bfly.b32 %r804|%p23, %r803, %r790, %r785, %r787;
mov.b32 %f90, %r804;
add.ftz.f32 %f91, %f89, %f90;
mov.b32 %r805, %f91;
shfl.sync.bfly.b32 %r806|%p24, %r805, %r793, %r785, %r787;
mov.b32 %f92, %r806;
add.ftz.f32 %f93, %f91, %f92;
mov.b32 %r807, %f93;
shfl.sync.bfly.b32 %r808|%p25, %r807, %r796, %r785, %r787;
mov.b32 %f94, %r808;
add.ftz.f32 %f95, %f93, %f94;
mov.b32 %r809, %f95;
shfl.sync.bfly.b32 %r810|%p26, %r809, %r799, %r785, %r787;
mov.b32 %f96, %r810;
add.ftz.f32 %f97, %f95, %f96;
st.local.f32 [%rd122], %f97;
mov.b32 %r811, %f164;
shfl.sync.bfly.b32 %r812|%p27, %r811, %r786, %r785, %r787;
mov.b32 %f98, %r812;
add.ftz.f32 %f99, %f164, %f98;
mov.b32 %r813, %f99;
shfl.sync.bfly.b32 %r814|%p28, %r813, %r790, %r785, %r787;
mov.b32 %f100, %r814;
add.ftz.f32 %f101, %f99, %f100;
mov.b32 %r815, %f101;
shfl.sync.bfly.b32 %r816|%p29, %r815, %r793, %r785, %r787;
mov.b32 %f102, %r816;
add.ftz.f32 %f103, %f101, %f102;
mov.b32 %r817, %f103;
shfl.sync.bfly.b32 %r818|%p30, %r817, %r796, %r785, %r787;
mov.b32 %f104, %r818;
add.ftz.f32 %f105, %f103, %f104;
mov.b32 %r819, %f105;
shfl.sync.bfly.b32 %r820|%p31, %r819, %r799, %r785, %r787;
mov.b32 %f106, %r820;
add.ftz.f32 %f107, %f105, %f106;
st.local.f32 [%rd130+4], %f107;
mov.b32 %r821, %f162;
shfl.sync.bfly.b32 %r822|%p32, %r821, %r786, %r785, %r787;
mov.b32 %f108, %r822;
add.ftz.f32 %f109, %f162, %f108;
mov.b32 %r823, %f109;
shfl.sync.bfly.b32 %r824|%p33, %r823, %r790, %r785, %r787;
mov.b32 %f110, %r824;
add.ftz.f32 %f111, %f109, %f110;
mov.b32 %r825, %f111;
shfl.sync.bfly.b32 %r826|%p34, %r825, %r793, %r785, %r787;
mov.b32 %f112, %r826;
add.ftz.f32 %f113, %f111, %f112;
mov.b32 %r827, %f113;
shfl.sync.bfly.b32 %r828|%p35, %r827, %r796, %r785, %r787;
mov.b32 %f114, %r828;
add.ftz.f32 %f115, %f113, %f114;
mov.b32 %r829, %f115;
shfl.sync.bfly.b32 %r830|%p36, %r829, %r799, %r785, %r787;
mov.b32 %f116, %r830;
add.ftz.f32 %f117, %f115, %f116;
st.local.f32 [%rd122+4], %f117;
bra.uni $L__BB198_12;

$L__BB198_10:
add.u64 %rd126, %SPL, 0;
mov.u32 %r759, 31;
mov.u32 %r760, 16;
mov.u32 %r761, -1;
shfl.sync.bfly.b32 %r762|%p7, %r21, %r760, %r759, %r761;
mov.b32 %f58, %r762;
add.ftz.f32 %f59, %f165, %f58;
mov.b32 %r763, %f59;
mov.u32 %r764, 8;
shfl.sync.bfly.b32 %r765|%p8, %r763, %r764, %r759, %r761;
mov.b32 %f60, %r765;
add.ftz.f32 %f61, %f59, %f60;
mov.b32 %r766, %f61;
mov.u32 %r767, 4;
shfl.sync.bfly.b32 %r768|%p9, %r766, %r767, %r759, %r761;
mov.b32 %f62, %r768;
add.ftz.f32 %f63, %f61, %f62;
mov.b32 %r769, %f63;
mov.u32 %r770, 2;
shfl.sync.bfly.b32 %r771|%p10, %r769, %r770, %r759, %r761;
mov.b32 %f64, %r771;
add.ftz.f32 %f65, %f63, %f64;
mov.b32 %r772, %f65;
mov.u32 %r773, 1;
shfl.sync.bfly.b32 %r774|%p11, %r772, %r773, %r759, %r761;
mov.b32 %f66, %r774;
add.ftz.f32 %f67, %f65, %f66;
st.local.f32 [%rd126], %f67;
mov.b32 %r775, %f164;
shfl.sync.bfly.b32 %r776|%p12, %r775, %r760, %r759, %r761;
mov.b32 %f68, %r776;
add.ftz.f32 %f69, %f164, %f68;
mov.b32 %r777, %f69;
shfl.sync.bfly.b32 %r778|%p13, %r777, %r764, %r759, %r761;
mov.b32 %f70, %r778;
add.ftz.f32 %f71, %f69, %f70;
mov.b32 %r779, %f71;
shfl.sync.bfly.b32 %r780|%p14, %r779, %r767, %r759, %r761;
mov.b32 %f72, %r780;
add.ftz.f32 %f73, %f71, %f72;
mov.b32 %r781, %f73;
shfl.sync.bfly.b32 %r782|%p15, %r781, %r770, %r759, %r761;
mov.b32 %f74, %r782;
add.ftz.f32 %f75, %f73, %f74;
mov.b32 %r783, %f75;
shfl.sync.bfly.b32 %r784|%p16, %r783, %r773, %r759, %r761;
mov.b32 %f76, %r784;
add.ftz.f32 %f77, %f75, %f76;
st.local.f32 [%rd126+4], %f77;

$L__BB198_12:
ld.param.u32 %r841, [_Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7];
mov.u32 %r840, %tid.x;
setp.gt.u32 %p37, %r840, 1;
mov.u32 %r832, %ctaid.x;
shl.b32 %r833, %r832, 1;
add.s32 %r22, %r833, %r840;
setp.ge.u32 %p38, %r22, %r841;
or.pred %p39, %p37, %p38;
@%p39 bra $L__BB198_29;

add.u64 %rd128, %SPL, 0;
ld.param.u64 %rd111, [_Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3];
ld.param.u32 %r843, [_Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13];
mov.u32 %r842, %tid.x;
cvt.u64.u32 %rd19, %r842;
mul.wide.u32 %rd97, %r842, 4;
add.s64 %rd98, %rd128, %rd97;
ld.local.f32 %f168, [%rd98];
mad.lo.s32 %r23, %r4, %r843, %r22;
setp.eq.s64 %p40, %rd111, 0;
@%p40 bra $L__BB198_15;

ld.param.u64 %rd112, [_Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3];
cvta.to.global.u64 %rd99, %rd112;
mul.wide.u32 %rd100, %r23, 4;
add.s64 %rd101, %rd99, %rd100;
ld.global.f32 %f118, [%rd101];
add.ftz.f32 %f168, %f168, %f118;

$L__BB198_15:
setp.eq.s64 %p41, %rd25, 0;
@%p41 bra $L__BB198_28;

add.u64 %rd120, %SPL, 8;
ld.param.u64 %rd114, [_Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+16];
shl.b64 %rd102, %rd19, 2;
add.s64 %rd103, %rd120, %rd102;
ld.local.f32 %f24, [%rd103];
setp.eq.s64 %p42, %rd114, 0;
@%p42 bra $L__BB198_18;

ld.param.u64 %rd115, [_Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+16];
cvta.to.global.u64 %rd104, %rd115;
mul.wide.u32 %rd105, %r23, 4;
add.s64 %rd106, %rd104, %rd105;
ld.global.f32 %f119, [%rd106];
add.ftz.f32 %f24, %f24, %f119;

$L__BB198_18:
ld.param.u32 %r848, [_Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.gt.s32 %p43, %r848, 2;
@%p43 bra $L__BB198_22;

ld.param.u32 %r851, [_Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.eq.s32 %p46, %r851, 1;
@%p46 bra $L__BB198_26;

ld.param.u32 %r852, [_Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.eq.s32 %p47, %r852, 2;
@%p47 bra $L__BB198_21;
bra.uni $L__BB198_27;

$L__BB198_21:
mul.ftz.f32 %f139, %f24, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f140, %f139;
add.ftz.f32 %f141, %f140, 0f3F800000;
div.approx.ftz.f32 %f142, %f24, %f141;
mul.ftz.f32 %f168, %f168, %f142;
bra.uni $L__BB198_28;

$L__BB198_22:
ld.param.u32 %r849, [_Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.eq.s32 %p44, %r849, 3;
@%p44 bra $L__BB198_25;

ld.param.u32 %r850, [_Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.ne.s32 %p45, %r850, 6;
@%p45 bra $L__BB198_27;

ld.param.f32 %f151, [_Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+44];
min.ftz.f32 %f120, %f24, %f151;
neg.ftz.f32 %f121, %f151;
min.ftz.f32 %f122, %f168, %f151;
max.ftz.f32 %f123, %f122, %f121;
mul.ftz.f32 %f124, %f120, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f125, %f124;
add.ftz.f32 %f126, %f125, 0f3F800000;
div.approx.ftz.f32 %f127, %f120, %f126;
mul.ftz.f32 %f168, %f123, %f127;
bra.uni $L__BB198_28;

$L__BB198_27:
mul.ftz.f32 %f168, %f168, %f24;
bra.uni $L__BB198_28;

$L__BB198_26:
mul.ftz.f32 %f143, %f24, 0f3F000000;
mul.ftz.f32 %f144, %f24, 0f3D372713;
fma.rn.ftz.f32 %f145, %f24, %f144, 0f3F800000;
mul.ftz.f32 %f146, %f24, 0f3F4C422A;
mul.ftz.f32 %f147, %f146, %f145;
tanh.approx.f32 %f148, %f147;
add.ftz.f32 %f149, %f148, 0f3F800000;
mul.ftz.f32 %f150, %f143, %f149;
mul.ftz.f32 %f168, %f168, %f150;
bra.uni $L__BB198_28;

$L__BB198_25:
mov.f32 %f128, 0f40E00000;
min.ftz.f32 %f129, %f24, %f128;
min.ftz.f32 %f130, %f168, %f128;
mov.f32 %f131, 0fC0E00000;
max.ftz.f32 %f132, %f130, %f131;
mul.ftz.f32 %f133, %f129, 0f3FD9DB23;
mul.ftz.f32 %f134, %f133, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f135, %f134;
add.ftz.f32 %f136, %f135, 0f3F800000;
div.approx.ftz.f32 %f137, %f129, %f136;
add.ftz.f32 %f138, %f132, 0f3F800000;
mul.ftz.f32 %f168, %f138, %f137;

$L__BB198_28:
ld.param.u64 %rd113, [_Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4];
ld.param.u32 %r847, [_Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10];
ld.param.u32 %r846, [_Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13];
mov.u32 %r845, %ctaid.y;
mov.u32 %r844, %tid.y;
mad.lo.s32 %r838, %r844, %r847, %r22;
mad.lo.s32 %r839, %r845, %r846, %r838;
cvta.to.global.u64 %rd107, %rd113;
mul.wide.u32 %rd108, %r839, 4;
add.s64 %rd109, %rd107, %rd108;
st.global.f32 [%rd109], %f168;

$L__BB198_29:
ret;

}
.entry _Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj(
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2,
.param .align 8 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3[48],
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5,
.param .align 4 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6[12],
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15
)
.maxntid 256, 1, 1
.minnctapersm 1
{
.local .align 8 .b8 __local_depot199[8];
.reg .b64 %SP;
.reg .b64 %SPL;
.reg .pred %p<17>;
.reg .b16 %rs<19>;
.reg .f32 %f<44>;
.reg .b32 %r<432>;
.reg .b64 %rd<67>;


mov.u64 %SPL, __local_depot199;
ld.param.u64 %rd15, [_Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0];
ld.param.u64 %rd16, [_Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1];
ld.param.u64 %rd17, [_Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2];
ld.param.u64 %rd18, [_Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4];
ld.param.u32 %r22, [_Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5];
ld.param.u32 %r14, [_Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7];
ld.param.u32 %r15, [_Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8];
ld.param.u32 %r16, [_Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9];
ld.param.u32 %r17, [_Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10];
ld.param.u32 %r18, [_Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11];
ld.param.u32 %r19, [_Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12];
ld.param.u32 %r20, [_Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13];
ld.param.u32 %r23, [_Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14];
ld.param.u32 %r21, [_Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15];
ld.param.u32 %r13, [_Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+8];
ld.param.u32 %r12, [_Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+4];
ld.param.u32 %r11, [_Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6];
add.u64 %rd1, %SPL, 0;
mov.u32 %r24, %ctaid.x;
shl.b32 %r1, %r24, 1;
shr.u32 %r2, %r22, 8;
mov.u32 %r3, %tid.y;
setp.ge.u32 %p1, %r3, %r23;
@%p1 bra $L__BB199_7;

mov.f32 %f42, 0f00000000;
st.local.v2.f32 [%rd1], {%f42, %f42};
mov.u32 %r25, %tid.x;
shr.u32 %r431, %r25, 3;
setp.ge.u32 %p2, %r431, %r2;
mov.f32 %f43, %f42;
@%p2 bra $L__BB199_5;

mov.u32 %r27, %ctaid.y;
mad.lo.s32 %r28, %r3, %r21, %r27;
cvta.to.global.u64 %rd20, %rd17;
mul.wide.u32 %rd21, %r28, 4;
add.s64 %rd22, %rd20, %rd21;
ld.global.nc.u32 %r29, [%rd22];
mul.lo.s32 %r30, %r29, %r18;
and.b32 %r32, %r25, 7;
cvt.u64.u32 %rd2, %r32;
shl.b32 %r430, %r431, 3;
add.s32 %r33, %r431, %r30;
add.s32 %r34, %r1, 1;
mad.lo.s32 %r35, %r15, %r34, %r33;
mul.wide.s32 %rd23, %r35, 66;
cvta.to.global.u64 %rd24, %rd15;
add.s64 %rd66, %rd24, %rd23;
mad.lo.s32 %r36, %r1, %r15, %r431;
add.s32 %r37, %r36, %r30;
mul.wide.s32 %rd25, %r37, 66;
add.s64 %rd26, %rd24, %rd25;
add.s64 %rd65, %rd26, 4;
mul.wide.u32 %rd27, %r32, 8;
add.s64 %rd28, %rd27, %rd25;
add.s64 %rd29, %rd24, %rd28;
add.s64 %rd64, %rd29, 8;
or.b64 %rd6, %rd27, 4;
mul.lo.s32 %r38, %r3, %r16;
cvt.u64.u32 %rd30, %r38;
mul.hi.u32 %r39, %r27, %r11;
add.s32 %r40, %r39, %r27;
shr.u32 %r41, %r40, %r12;
mul.lo.s32 %r42, %r41, %r13;
sub.s32 %r43, %r27, %r42;
mul.lo.s32 %r44, %r43, %r19;
cvt.u64.u32 %rd31, %r44;
add.s64 %rd7, %rd31, %rd30;
cvta.to.global.u64 %rd8, %rd16;
mov.f32 %f42, 0f00000000;

$L__BB199_3:
cvt.s64.s32 %rd32, %r430;
add.s64 %rd33, %rd7, %rd32;
add.s64 %rd34, %rd33, %rd2;
mul.lo.s64 %rd35, %rd34, 36;
ld.global.nc.u16 %rs3, [%rd64+-6];
shr.u16 %rs4, %rs3, 8;
ld.global.nc.u16 %rs5, [%rd64+-4];
shr.u16 %rs6, %rs5, 8;
ld.global.nc.u16 %rs7, [%rd64+-2];
cvt.u32.u16 %r143, %rs7;
ld.global.nc.u16 %rs8, [%rd64];
cvt.u32.u16 %r144, %rs8;
mov.b32 %r145, {%rs7, %rs8};
shl.b16 %rs9, %rs3, 3;
cvt.u64.u16 %rd36, %rs9;
and.b64 %rd37, %rd36, 2040;
mov.u64 %rd38, iq2xxs_grid;
add.s64 %rd39, %rd38, %rd37;
ld.global.nc.v2.u32 {%r146, %r147}, [%rd39];
mov.u32 %r99, 0;
and.b32 %r150, %r143, 255;
popc.b32 %r151, %r150;
shl.b32 %r152, %r151, 7;
and.b32 %r153, %r152, 128;
xor.b32 %r154, %r153, %r150;
mul.lo.s32 %r155, %r154, 16843009;
and.b32 %r156, %r155, 134480385;
or.b32 %r157, %r156, -2139062144;
add.s32 %r158, %r157, -16843009;
or.b32 %r45, %r158, %r156;

prmt.b32 %r45, %r45, 0, 0xba98;

xor.b32 %r159, %r146, %r45;
or.b32 %r160, %r159, -2139062144;
and.b32 %r161, %r45, 2139062143;
and.b32 %r162, %r146, -2139062144;
sub.s32 %r163, %r160, %r161;
xor.b32 %r164, %r162, %r163;
xor.b32 %r48, %r164, -2139062144;
add.s64 %rd40, %rd8, %rd35;
ld.global.nc.u32 %r49, [%rd40+4];

dp4a.s32.s32 %r47, %r48, %r49, %r99;

and.b32 %r165, %r155, -2143281136;
or.b32 %r166, %r165, -2139062144;
add.s32 %r167, %r166, -16843009;
or.b32 %r51, %r167, %r165;

prmt.b32 %r51, %r51, 0, 0xba98;

xor.b32 %r168, %r51, %r147;
or.b32 %r169, %r168, -2139062144;
and.b32 %r170, %r51, 2139062143;
and.b32 %r171, %r147, -2139062144;
sub.s32 %r172, %r169, %r170;
xor.b32 %r173, %r171, %r172;
xor.b32 %r54, %r173, -2139062144;
ld.global.nc.u32 %r55, [%rd40+8];

dp4a.s32.s32 %r53, %r54, %r55, %r47;

cvt.u32.u16 %r174, %rs4;
mul.wide.u32 %rd41, %r174, 8;
add.s64 %rd42, %rd38, %rd41;
ld.global.nc.v2.u32 {%r175, %r176}, [%rd42];
shr.u32 %r179, %r143, 7;
and.b32 %r180, %r179, 255;
popc.b32 %r181, %r180;
shl.b32 %r182, %r181, 7;
and.b32 %r183, %r182, 128;
xor.b32 %r184, %r183, %r180;
mul.lo.s32 %r185, %r184, 16843009;
and.b32 %r186, %r185, 134480385;
or.b32 %r187, %r186, -2139062144;
add.s32 %r188, %r187, -16843009;
or.b32 %r57, %r188, %r186;

prmt.b32 %r57, %r57, 0, 0xba98;

xor.b32 %r189, %r175, %r57;
or.b32 %r190, %r189, -2139062144;
and.b32 %r191, %r57, 2139062143;
and.b32 %r192, %r175, -2139062144;
sub.s32 %r193, %r190, %r191;
xor.b32 %r194, %r192, %r193;
xor.b32 %r60, %r194, -2139062144;
ld.global.nc.u32 %r61, [%rd40+12];

dp4a.s32.s32 %r59, %r60, %r61, %r53;

and.b32 %r195, %r185, -2143281136;
or.b32 %r196, %r195, -2139062144;
add.s32 %r197, %r196, -16843009;
or.b32 %r63, %r197, %r195;

prmt.b32 %r63, %r63, 0, 0xba98;

xor.b32 %r198, %r63, %r176;
or.b32 %r199, %r198, -2139062144;
and.b32 %r200, %r63, 2139062143;
and.b32 %r201, %r176, -2139062144;
sub.s32 %r202, %r199, %r200;
xor.b32 %r203, %r201, %r202;
xor.b32 %r66, %r203, -2139062144;
ld.global.nc.u32 %r67, [%rd40+16];

dp4a.s32.s32 %r65, %r66, %r67, %r59;

shl.b16 %rs10, %rs5, 3;
cvt.u64.u16 %rd43, %rs10;
and.b64 %rd44, %rd43, 2040;
add.s64 %rd45, %rd38, %rd44;
ld.global.nc.v2.u32 {%r204, %r205}, [%rd45];
shr.u32 %r208, %r145, 14;
and.b32 %r209, %r208, 255;
popc.b32 %r210, %r209;
shl.b32 %r211, %r210, 7;
and.b32 %r212, %r211, 128;
xor.b32 %r213, %r212, %r209;
mul.lo.s32 %r214, %r213, 16843009;
and.b32 %r215, %r214, 134480385;
or.b32 %r216, %r215, -2139062144;
add.s32 %r217, %r216, -16843009;
or.b32 %r69, %r217, %r215;

prmt.b32 %r69, %r69, 0, 0xba98;

xor.b32 %r218, %r204, %r69;
or.b32 %r219, %r218, -2139062144;
and.b32 %r220, %r69, 2139062143;
and.b32 %r221, %r204, -2139062144;
sub.s32 %r222, %r219, %r220;
xor.b32 %r223, %r221, %r222;
xor.b32 %r72, %r223, -2139062144;
ld.global.nc.u32 %r73, [%rd40+20];

dp4a.s32.s32 %r71, %r72, %r73, %r65;

and.b32 %r224, %r214, -2143281136;
or.b32 %r225, %r224, -2139062144;
add.s32 %r226, %r225, -16843009;
or.b32 %r75, %r226, %r224;

prmt.b32 %r75, %r75, 0, 0xba98;

xor.b32 %r227, %r75, %r205;
or.b32 %r228, %r227, -2139062144;
and.b32 %r229, %r75, 2139062143;
and.b32 %r230, %r205, -2139062144;
sub.s32 %r231, %r228, %r229;
xor.b32 %r232, %r230, %r231;
xor.b32 %r78, %r232, -2139062144;
ld.global.nc.u32 %r79, [%rd40+24];

dp4a.s32.s32 %r77, %r78, %r79, %r71;

cvt.u32.u16 %r233, %rs6;
mul.wide.u32 %rd46, %r233, 8;
add.s64 %rd47, %rd38, %rd46;
ld.global.nc.v2.u32 {%r234, %r235}, [%rd47];
shr.u32 %r238, %r144, 5;
and.b32 %r239, %r238, 255;
popc.b32 %r240, %r239;
shl.b32 %r241, %r240, 7;
and.b32 %r242, %r241, 128;
xor.b32 %r243, %r242, %r239;
mul.lo.s32 %r244, %r243, 16843009;
and.b32 %r245, %r244, 134480385;
or.b32 %r246, %r245, -2139062144;
add.s32 %r247, %r246, -16843009;
or.b32 %r81, %r247, %r245;

prmt.b32 %r81, %r81, 0, 0xba98;

xor.b32 %r248, %r234, %r81;
or.b32 %r249, %r248, -2139062144;
and.b32 %r250, %r81, 2139062143;
and.b32 %r251, %r234, -2139062144;
sub.s32 %r252, %r249, %r250;
xor.b32 %r253, %r251, %r252;
xor.b32 %r84, %r253, -2139062144;
ld.global.nc.u32 %r85, [%rd40+28];

dp4a.s32.s32 %r83, %r84, %r85, %r77;

and.b32 %r254, %r244, -2143281136;
or.b32 %r255, %r254, -2139062144;
add.s32 %r256, %r255, -16843009;
or.b32 %r87, %r256, %r254;

prmt.b32 %r87, %r87, 0, 0xba98;

xor.b32 %r257, %r87, %r235;
or.b32 %r258, %r257, -2139062144;
and.b32 %r259, %r87, 2139062143;
and.b32 %r260, %r235, -2139062144;
sub.s32 %r261, %r258, %r259;
xor.b32 %r262, %r260, %r261;
xor.b32 %r90, %r262, -2139062144;
ld.global.nc.u32 %r91, [%rd40+32];

dp4a.s32.s32 %r89, %r90, %r91, %r83;

shr.u32 %r263, %r144, 11;
or.b32 %r264, %r263, 1;
mul.lo.s32 %r265, %r89, %r264;
shr.s32 %r266, %r265, 31;
shr.u32 %r267, %r266, 29;
add.s32 %r268, %r265, %r267;
shr.s32 %r269, %r268, 3;
ld.global.nc.u16 %rs1, [%rd65+-4];

{ cvt.f32.f16 %f11, %rs1;}


ld.global.nc.u32 %r93, [%rd40];

{.reg .f16 low,high;
mov.b32 {low,high},%r93;
cvt.f32.f16 %f12, low;}


mul.ftz.f32 %f15, %f11, %f12;
cvt.rn.f32.s32 %f16, %r269;
fma.rn.ftz.f32 %f43, %f15, %f16, %f43;
add.s64 %rd48, %rd66, %rd6;
ld.global.nc.u16 %rs11, [%rd48+-2];
shr.u16 %rs12, %rs11, 8;
ld.global.nc.u16 %rs13, [%rd48];
shr.u16 %rs14, %rs13, 8;
ld.global.nc.u16 %rs15, [%rd48+2];
cvt.u32.u16 %r270, %rs15;
ld.global.nc.u16 %rs16, [%rd48+4];
cvt.u32.u16 %r271, %rs16;
mov.b32 %r272, {%rs15, %rs16};
shl.b16 %rs17, %rs11, 3;
cvt.u64.u16 %rd49, %rs17;
and.b64 %rd50, %rd49, 2040;
add.s64 %rd51, %rd38, %rd50;
ld.global.nc.v2.u32 {%r273, %r274}, [%rd51];
and.b32 %r277, %r270, 255;
popc.b32 %r278, %r277;
shl.b32 %r279, %r278, 7;
and.b32 %r280, %r279, 128;
xor.b32 %r281, %r280, %r277;
mul.lo.s32 %r282, %r281, 16843009;
and.b32 %r283, %r282, 134480385;
or.b32 %r284, %r283, -2139062144;
add.s32 %r285, %r284, -16843009;
or.b32 %r94, %r285, %r283;

prmt.b32 %r94, %r94, 0, 0xba98;

xor.b32 %r286, %r273, %r94;
or.b32 %r287, %r286, -2139062144;
and.b32 %r288, %r94, 2139062143;
and.b32 %r289, %r273, -2139062144;
sub.s32 %r290, %r287, %r288;
xor.b32 %r291, %r289, %r290;
xor.b32 %r97, %r291, -2139062144;

dp4a.s32.s32 %r96, %r97, %r49, %r99;

and.b32 %r292, %r282, -2143281136;
or.b32 %r293, %r292, -2139062144;
add.s32 %r294, %r293, -16843009;
or.b32 %r100, %r294, %r292;

prmt.b32 %r100, %r100, 0, 0xba98;

xor.b32 %r295, %r100, %r274;
or.b32 %r296, %r295, -2139062144;
and.b32 %r297, %r100, 2139062143;
and.b32 %r298, %r274, -2139062144;
sub.s32 %r299, %r296, %r297;
xor.b32 %r300, %r298, %r299;
xor.b32 %r103, %r300, -2139062144;

dp4a.s32.s32 %r102, %r103, %r55, %r96;

cvt.u32.u16 %r301, %rs12;
mul.wide.u32 %rd52, %r301, 8;
add.s64 %rd53, %rd38, %rd52;
ld.global.nc.v2.u32 {%r302, %r303}, [%rd53];
shr.u32 %r306, %r270, 7;
and.b32 %r307, %r306, 255;
popc.b32 %r308, %r307;
shl.b32 %r309, %r308, 7;
and.b32 %r310, %r309, 128;
xor.b32 %r311, %r310, %r307;
mul.lo.s32 %r312, %r311, 16843009;
and.b32 %r313, %r312, 134480385;
or.b32 %r314, %r313, -2139062144;
add.s32 %r315, %r314, -16843009;
or.b32 %r106, %r315, %r313;

prmt.b32 %r106, %r106, 0, 0xba98;

xor.b32 %r316, %r302, %r106;
or.b32 %r317, %r316, -2139062144;
and.b32 %r318, %r106, 2139062143;
and.b32 %r319, %r302, -2139062144;
sub.s32 %r320, %r317, %r318;
xor.b32 %r321, %r319, %r320;
xor.b32 %r109, %r321, -2139062144;

dp4a.s32.s32 %r108, %r109, %r61, %r102;

and.b32 %r322, %r312, -2143281136;
or.b32 %r323, %r322, -2139062144;
add.s32 %r324, %r323, -16843009;
or.b32 %r112, %r324, %r322;

prmt.b32 %r112, %r112, 0, 0xba98;

xor.b32 %r325, %r112, %r303;
or.b32 %r326, %r325, -2139062144;
and.b32 %r327, %r112, 2139062143;
and.b32 %r328, %r303, -2139062144;
sub.s32 %r329, %r326, %r327;
xor.b32 %r330, %r328, %r329;
xor.b32 %r115, %r330, -2139062144;

dp4a.s32.s32 %r114, %r115, %r67, %r108;

shl.b16 %rs18, %rs13, 3;
cvt.u64.u16 %rd54, %rs18;
and.b64 %rd55, %rd54, 2040;
add.s64 %rd56, %rd38, %rd55;
ld.global.nc.v2.u32 {%r331, %r332}, [%rd56];
shr.u32 %r335, %r272, 14;
and.b32 %r336, %r335, 255;
popc.b32 %r337, %r336;
shl.b32 %r338, %r337, 7;
and.b32 %r339, %r338, 128;
xor.b32 %r340, %r339, %r336;
mul.lo.s32 %r341, %r340, 16843009;
and.b32 %r342, %r341, 134480385;
or.b32 %r343, %r342, -2139062144;
add.s32 %r344, %r343, -16843009;
or.b32 %r118, %r344, %r342;

prmt.b32 %r118, %r118, 0, 0xba98;

xor.b32 %r345, %r331, %r118;
or.b32 %r346, %r345, -2139062144;
and.b32 %r347, %r118, 2139062143;
and.b32 %r348, %r331, -2139062144;
sub.s32 %r349, %r346, %r347;
xor.b32 %r350, %r348, %r349;
xor.b32 %r121, %r350, -2139062144;

dp4a.s32.s32 %r120, %r121, %r73, %r114;

and.b32 %r351, %r341, -2143281136;
or.b32 %r352, %r351, -2139062144;
add.s32 %r353, %r352, -16843009;
or.b32 %r124, %r353, %r351;

prmt.b32 %r124, %r124, 0, 0xba98;

xor.b32 %r354, %r124, %r332;
or.b32 %r355, %r354, -2139062144;
and.b32 %r356, %r124, 2139062143;
and.b32 %r357, %r332, -2139062144;
sub.s32 %r358, %r355, %r356;
xor.b32 %r359, %r357, %r358;
xor.b32 %r127, %r359, -2139062144;

dp4a.s32.s32 %r126, %r127, %r79, %r120;

cvt.u32.u16 %r360, %rs14;
mul.wide.u32 %rd57, %r360, 8;
add.s64 %rd58, %rd38, %rd57;
ld.global.nc.v2.u32 {%r361, %r362}, [%rd58];
shr.u32 %r365, %r271, 5;
and.b32 %r366, %r365, 255;
popc.b32 %r367, %r366;
shl.b32 %r368, %r367, 7;
and.b32 %r369, %r368, 128;
xor.b32 %r370, %r369, %r366;
mul.lo.s32 %r371, %r370, 16843009;
and.b32 %r372, %r371, 134480385;
or.b32 %r373, %r372, -2139062144;
add.s32 %r374, %r373, -16843009;
or.b32 %r130, %r374, %r372;

prmt.b32 %r130, %r130, 0, 0xba98;

xor.b32 %r375, %r361, %r130;
or.b32 %r376, %r375, -2139062144;
and.b32 %r377, %r130, 2139062143;
and.b32 %r378, %r361, -2139062144;
sub.s32 %r379, %r376, %r377;
xor.b32 %r380, %r378, %r379;
xor.b32 %r133, %r380, -2139062144;

dp4a.s32.s32 %r132, %r133, %r85, %r126;

and.b32 %r381, %r371, -2143281136;
or.b32 %r382, %r381, -2139062144;
add.s32 %r383, %r382, -16843009;
or.b32 %r136, %r383, %r381;

prmt.b32 %r136, %r136, 0, 0xba98;

xor.b32 %r384, %r136, %r362;
or.b32 %r385, %r384, -2139062144;
and.b32 %r386, %r136, 2139062143;
and.b32 %r387, %r362, -2139062144;
sub.s32 %r388, %r385, %r386;
xor.b32 %r389, %r387, %r388;
xor.b32 %r139, %r389, -2139062144;

dp4a.s32.s32 %r138, %r139, %r91, %r132;

shr.u32 %r390, %r271, 11;
or.b32 %r391, %r390, 1;
mul.lo.s32 %r392, %r138, %r391;
shr.s32 %r393, %r392, 31;
shr.u32 %r394, %r393, 29;
add.s32 %r395, %r392, %r394;
shr.s32 %r396, %r395, 3;
ld.global.nc.u16 %rs2, [%rd66];

{ cvt.f32.f16 %f13, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r93;
cvt.f32.f16 %f14, low;}


mul.ftz.f32 %f17, %f13, %f14;
cvt.rn.f32.s32 %f18, %r396;
fma.rn.ftz.f32 %f42, %f17, %f18, %f42;
add.s32 %r430, %r430, 32;
add.s64 %rd66, %rd66, 264;
add.s64 %rd65, %rd65, 264;
add.s64 %rd64, %rd64, 264;
add.s32 %r431, %r431, 4;
setp.lt.u32 %p3, %r431, %r2;
@%p3 bra $L__BB199_3;

st.local.v2.f32 [%rd1], {%f43, %f42};

$L__BB199_5:
mov.b32 %r397, %f43;
mov.u32 %r398, 31;
mov.u32 %r399, 16;
mov.u32 %r400, -1;
shfl.sync.bfly.b32 %r401|%p4, %r397, %r399, %r398, %r400;
mov.b32 %f19, %r401;
add.ftz.f32 %f20, %f43, %f19;
mov.b32 %r402, %f20;
mov.u32 %r403, 8;
shfl.sync.bfly.b32 %r404|%p5, %r402, %r403, %r398, %r400;
mov.b32 %f21, %r404;
add.ftz.f32 %f22, %f20, %f21;
mov.b32 %r405, %f22;
mov.u32 %r406, 4;
shfl.sync.bfly.b32 %r407|%p6, %r405, %r406, %r398, %r400;
mov.b32 %f23, %r407;
add.ftz.f32 %f24, %f22, %f23;
mov.b32 %r408, %f24;
mov.u32 %r409, 2;
shfl.sync.bfly.b32 %r410|%p7, %r408, %r409, %r398, %r400;
mov.b32 %f25, %r410;
add.ftz.f32 %f26, %f24, %f25;
mov.b32 %r411, %f26;
mov.u32 %r412, 1;
shfl.sync.bfly.b32 %r413|%p8, %r411, %r412, %r398, %r400;
mov.b32 %f27, %r413;
add.ftz.f32 %f28, %f26, %f27;
st.local.f32 [%rd1], %f28;
mov.b32 %r414, %f42;
shfl.sync.bfly.b32 %r415|%p9, %r414, %r399, %r398, %r400;
mov.b32 %f29, %r415;
add.ftz.f32 %f30, %f42, %f29;
mov.b32 %r416, %f30;
shfl.sync.bfly.b32 %r417|%p10, %r416, %r403, %r398, %r400;
mov.b32 %f31, %r417;
add.ftz.f32 %f32, %f30, %f31;
mov.b32 %r418, %f32;
shfl.sync.bfly.b32 %r419|%p11, %r418, %r406, %r398, %r400;
mov.b32 %f33, %r419;
add.ftz.f32 %f34, %f32, %f33;
mov.b32 %r420, %f34;
shfl.sync.bfly.b32 %r421|%p12, %r420, %r409, %r398, %r400;
mov.b32 %f35, %r421;
add.ftz.f32 %f36, %f34, %f35;
mov.b32 %r422, %f36;
shfl.sync.bfly.b32 %r423|%p13, %r422, %r412, %r398, %r400;
mov.b32 %f37, %r423;
add.ftz.f32 %f38, %f36, %f37;
st.local.f32 [%rd1+4], %f38;
setp.gt.u32 %p14, %r25, 1;
add.s32 %r10, %r1, %r25;
setp.ge.u32 %p15, %r10, %r14;
or.pred %p16, %p14, %p15;
@%p16 bra $L__BB199_7;

mul.wide.u32 %rd59, %r25, 4;
add.s64 %rd60, %rd1, %rd59;
ld.local.f32 %f39, [%rd60];
mov.u32 %r426, %ctaid.y;
mad.lo.s32 %r428, %r3, %r17, %r10;
mad.lo.s32 %r429, %r426, %r20, %r428;
cvta.to.global.u64 %rd61, %rd18;
mul.wide.u32 %rd62, %r429, 4;
add.s64 %rd63, %rd61, %rd62;
st.global.f32 [%rd63], %f39;

$L__BB199_7:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<320>;
.reg .b16 %rs<73>;
.reg .f32 %f<1005>;
.reg .b32 %r<2209>;
.reg .b64 %rd<185>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate[1536];

ld.param.u64 %rd28, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd29, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd30, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u32 %r45, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r49, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r50, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r54, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r55, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r56, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r60, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r61, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r62, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r58, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r57, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r52, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r51, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r48, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r47, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r46, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
ld.param.u64 %rd33, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+16];
ld.param.u64 %rd32, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+8];
ld.param.u64 %rd31, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3];
mov.u32 %r1, %tid.y;
shl.b32 %r63, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r63, %r2;
mov.u32 %r64, %ctaid.x;
shl.b32 %r4, %r64, 2;
mov.u32 %r5, %ctaid.y;
setp.eq.s64 %p1, %rd30, 0;
@%p1 bra $L__BB200_2;

cvta.to.global.u64 %rd37, %rd30;
mul.wide.u32 %rd38, %r5, 4;
add.s64 %rd39, %rd37, %rd38;
ld.global.nc.u32 %r2205, [%rd39];
mul.hi.u32 %r65, %r5, %r46;
add.s32 %r66, %r65, %r5;
shr.u32 %r67, %r66, %r47;
mul.lo.s32 %r68, %r67, %r48;
sub.s32 %r2206, %r5, %r68;
bra.uni $L__BB200_3;

$L__BB200_2:
mul.hi.u32 %r69, %r5, %r51;
add.s32 %r70, %r69, %r5;
shr.u32 %r2205, %r70, %r52;
mov.u32 %r2206, %r5;

$L__BB200_3:
mov.u32 %r11, %ctaid.z;
selp.b32 %r12, %r5, %r2205, %p1;
setp.ne.s32 %p3, %r1, 0;
setp.gt.u32 %p4, %r2, 3;
or.pred %p5, %p4, %p3;
add.s32 %r71, %r4, %r2;
setp.ge.u32 %p6, %r71, %r50;
mov.f32 %f27, 0f00000000;
or.pred %p7, %p5, %p6;
mov.f32 %f4, %f27;
mov.f32 %f5, %f27;
@%p7 bra $L__BB200_8;

setp.eq.s64 %p8, %rd31, 0;
mov.f32 %f4, 0f00000000;
mov.f32 %f5, %f4;
@%p8 bra $L__BB200_6;

cvta.to.global.u64 %rd40, %rd31;
mul.lo.s32 %r72, %r11, %r62;
cvt.u64.u32 %rd41, %r72;
mul.lo.s32 %r73, %r12, %r56;
cvt.u64.u32 %rd42, %r73;
cvt.u64.u32 %rd43, %r2;
cvt.s64.s32 %rd44, %r4;
add.s64 %rd45, %rd44, %rd43;
add.s64 %rd46, %rd45, %rd41;
add.s64 %rd47, %rd46, %rd42;
shl.b64 %rd48, %rd47, 2;
add.s64 %rd49, %rd40, %rd48;
ld.global.f32 %f5, [%rd49];

$L__BB200_6:
setp.eq.s64 %p9, %rd33, 0;
setp.eq.s64 %p10, %rd32, 0;
or.pred %p11, %p9, %p10;
@%p11 bra $L__BB200_8;

cvta.to.global.u64 %rd50, %rd33;
mul.lo.s32 %r74, %r11, %r62;
cvt.u64.u32 %rd51, %r74;
mul.lo.s32 %r75, %r12, %r56;
cvt.u64.u32 %rd52, %r75;
cvt.u64.u32 %rd53, %r2;
cvt.s64.s32 %rd54, %r4;
add.s64 %rd55, %rd54, %rd53;
add.s64 %rd56, %rd55, %rd51;
add.s64 %rd57, %rd56, %rd52;
shl.b64 %rd58, %rd57, 2;
add.s64 %rd59, %rd50, %rd58;
ld.global.f32 %f4, [%rd59];

$L__BB200_8:
mul.lo.s32 %r76, %r11, %r61;
cvt.u64.u32 %rd60, %r76;
mul.lo.s32 %r77, %r2206, %r55;
cvt.u64.u32 %rd61, %r77;
add.s64 %rd4, %rd61, %rd60;
mul.hi.u32 %r78, %r11, %r57;
add.s32 %r79, %r78, %r11;
shr.u32 %r13, %r79, %r58;
shr.s32 %r80, %r3, 31;
shr.u32 %r81, %r80, 29;
add.s32 %r82, %r3, %r81;
shr.s32 %r2208, %r82, 3;
shr.u32 %r83, %r45, 8;
setp.ge.s32 %p12, %r2208, %r83;
mov.f32 %f28, %f27;
mov.f32 %f29, %f27;
mov.f32 %f30, %f27;
mov.f32 %f31, %f27;
mov.f32 %f32, %f27;
mov.f32 %f33, %f27;
mov.f32 %f34, %f27;
@%p12 bra $L__BB200_19;

mul.lo.s32 %r84, %r13, %r60;
mul.lo.s32 %r85, %r2205, %r54;
and.b32 %r89, %r82, -8;
sub.s32 %r90, %r3, %r89;
shl.b32 %r91, %r90, 2;
cvt.s64.s32 %rd5, %r90;
shl.b32 %r2207, %r2208, 3;
add.s32 %r92, %r2208, %r84;
add.s32 %r93, %r92, %r85;
add.s32 %r94, %r4, 1;
mad.lo.s32 %r95, %r49, %r94, %r93;
mul.wide.s32 %rd6, %r95, 66;
mad.lo.s32 %r96, %r4, %r49, %r2208;
add.s32 %r97, %r96, %r84;
add.s32 %r98, %r97, %r85;
mul.wide.s32 %rd7, %r98, 66;
mul.wide.s32 %rd62, %r91, 2;
add.s64 %rd63, %rd7, %rd62;
add.s64 %rd8, %rd63, 8;
add.s64 %rd64, %rd6, %rd62;
add.s64 %rd9, %rd64, 4;
add.s32 %r99, %r4, 3;
mad.lo.s32 %r100, %r49, %r99, %r93;
mul.wide.s32 %rd10, %r100, 66;
add.s64 %rd65, %rd10, %rd62;
add.s64 %rd11, %rd65, 8;
add.s32 %r101, %r4, 2;
mad.lo.s32 %r102, %r49, %r101, %r93;
mul.wide.s32 %rd12, %r102, 66;
add.s64 %rd13, %rd12, %rd62;
mul.wide.s32 %rd14, %r49, 66;
cvta.to.global.u64 %rd184, %rd32;
cvta.to.global.u64 %rd183, %rd28;
mov.f32 %f27, 0f00000000;
cvta.to.global.u64 %rd66, %rd29;
mov.f32 %f28, %f27;
mov.f32 %f29, %f27;
mov.f32 %f30, %f27;
mov.f32 %f31, %f27;
mov.f32 %f32, %f27;
mov.f32 %f33, %f27;
mov.f32 %f34, %f27;

$L__BB200_10:
cvt.u64.u32 %rd67, %r2207;
add.s64 %rd68, %rd4, %rd67;
add.s64 %rd69, %rd183, %rd8;
ld.global.nc.u16 %rs2, [%rd69+-6];
shr.u16 %rs3, %rs2, 8;
ld.global.nc.u16 %rs4, [%rd69+-4];
shr.u16 %rs5, %rs4, 8;
ld.global.nc.u16 %rs6, [%rd69+-2];
cvt.u32.u16 %r152, %rs6;
ld.global.nc.u16 %rs7, [%rd69];
cvt.u32.u16 %r153, %rs7;
mov.b32 %r154, {%rs6, %rs7};
add.s64 %rd70, %rd68, %rd5;
mul.lo.s64 %rd71, %rd70, 36;
shl.b16 %rs8, %rs2, 3;
cvt.u64.u16 %rd72, %rs8;
and.b64 %rd73, %rd72, 2040;
mov.u64 %rd74, iq2xxs_grid;
add.s64 %rd75, %rd74, %rd73;
ld.global.nc.v2.u32 {%r155, %r156}, [%rd75];
mov.u32 %r108, 0;
and.b32 %r159, %r152, 255;
popc.b32 %r160, %r159;
shl.b32 %r161, %r160, 7;
and.b32 %r162, %r161, 128;
xor.b32 %r163, %r162, %r159;
mul.lo.s32 %r164, %r163, 16843009;
and.b32 %r165, %r164, 134480385;
or.b32 %r166, %r165, -2139062144;
add.s32 %r167, %r166, -16843009;
or.b32 %r103, %r167, %r165;

prmt.b32 %r103, %r103, 0, 0xba98;

xor.b32 %r168, %r155, %r103;
or.b32 %r169, %r168, -2139062144;
and.b32 %r170, %r103, 2139062143;
and.b32 %r171, %r155, -2139062144;
sub.s32 %r172, %r169, %r170;
xor.b32 %r173, %r171, %r172;
xor.b32 %r106, %r173, -2139062144;
add.s64 %rd76, %rd66, %rd71;
ld.global.nc.u32 %r18, [%rd76+4];

dp4a.s32.s32 %r105, %r106, %r18, %r108;

and.b32 %r174, %r164, -2143281136;
or.b32 %r175, %r174, -2139062144;
add.s32 %r176, %r175, -16843009;
or.b32 %r109, %r176, %r174;

prmt.b32 %r109, %r109, 0, 0xba98;

xor.b32 %r177, %r109, %r156;
or.b32 %r178, %r177, -2139062144;
and.b32 %r179, %r109, 2139062143;
and.b32 %r180, %r156, -2139062144;
sub.s32 %r181, %r178, %r179;
xor.b32 %r182, %r180, %r181;
xor.b32 %r112, %r182, -2139062144;
ld.global.nc.u32 %r19, [%rd76+8];

dp4a.s32.s32 %r111, %r112, %r19, %r105;

cvt.u32.u16 %r183, %rs3;
mul.wide.u32 %rd77, %r183, 8;
add.s64 %rd78, %rd74, %rd77;
ld.global.nc.v2.u32 {%r184, %r185}, [%rd78];
shr.u32 %r188, %r152, 7;
and.b32 %r189, %r188, 255;
popc.b32 %r190, %r189;
shl.b32 %r191, %r190, 7;
and.b32 %r192, %r191, 128;
xor.b32 %r193, %r192, %r189;
mul.lo.s32 %r194, %r193, 16843009;
and.b32 %r195, %r194, 134480385;
or.b32 %r196, %r195, -2139062144;
add.s32 %r197, %r196, -16843009;
or.b32 %r115, %r197, %r195;

prmt.b32 %r115, %r115, 0, 0xba98;

xor.b32 %r198, %r184, %r115;
or.b32 %r199, %r198, -2139062144;
and.b32 %r200, %r115, 2139062143;
and.b32 %r201, %r184, -2139062144;
sub.s32 %r202, %r199, %r200;
xor.b32 %r203, %r201, %r202;
xor.b32 %r118, %r203, -2139062144;
ld.global.nc.u32 %r20, [%rd76+12];

dp4a.s32.s32 %r117, %r118, %r20, %r111;

and.b32 %r204, %r194, -2143281136;
or.b32 %r205, %r204, -2139062144;
add.s32 %r206, %r205, -16843009;
or.b32 %r121, %r206, %r204;

prmt.b32 %r121, %r121, 0, 0xba98;

xor.b32 %r207, %r121, %r185;
or.b32 %r208, %r207, -2139062144;
and.b32 %r209, %r121, 2139062143;
and.b32 %r210, %r185, -2139062144;
sub.s32 %r211, %r208, %r209;
xor.b32 %r212, %r210, %r211;
xor.b32 %r124, %r212, -2139062144;
ld.global.nc.u32 %r21, [%rd76+16];

dp4a.s32.s32 %r123, %r124, %r21, %r117;

shl.b16 %rs9, %rs4, 3;
cvt.u64.u16 %rd79, %rs9;
and.b64 %rd80, %rd79, 2040;
add.s64 %rd81, %rd74, %rd80;
ld.global.nc.v2.u32 {%r213, %r214}, [%rd81];
shr.u32 %r217, %r154, 14;
and.b32 %r218, %r217, 255;
popc.b32 %r219, %r218;
shl.b32 %r220, %r219, 7;
and.b32 %r221, %r220, 128;
xor.b32 %r222, %r221, %r218;
mul.lo.s32 %r223, %r222, 16843009;
and.b32 %r224, %r223, 134480385;
or.b32 %r225, %r224, -2139062144;
add.s32 %r226, %r225, -16843009;
or.b32 %r127, %r226, %r224;

prmt.b32 %r127, %r127, 0, 0xba98;

xor.b32 %r227, %r213, %r127;
or.b32 %r228, %r227, -2139062144;
and.b32 %r229, %r127, 2139062143;
and.b32 %r230, %r213, -2139062144;
sub.s32 %r231, %r228, %r229;
xor.b32 %r232, %r230, %r231;
xor.b32 %r130, %r232, -2139062144;
ld.global.nc.u32 %r22, [%rd76+20];

dp4a.s32.s32 %r129, %r130, %r22, %r123;

and.b32 %r233, %r223, -2143281136;
or.b32 %r234, %r233, -2139062144;
add.s32 %r235, %r234, -16843009;
or.b32 %r133, %r235, %r233;

prmt.b32 %r133, %r133, 0, 0xba98;

xor.b32 %r236, %r133, %r214;
or.b32 %r237, %r236, -2139062144;
and.b32 %r238, %r133, 2139062143;
and.b32 %r239, %r214, -2139062144;
sub.s32 %r240, %r237, %r238;
xor.b32 %r241, %r239, %r240;
xor.b32 %r136, %r241, -2139062144;
ld.global.nc.u32 %r23, [%rd76+24];

dp4a.s32.s32 %r135, %r136, %r23, %r129;

cvt.u32.u16 %r242, %rs5;
mul.wide.u32 %rd82, %r242, 8;
add.s64 %rd83, %rd74, %rd82;
ld.global.nc.v2.u32 {%r243, %r244}, [%rd83];
shr.u32 %r247, %r153, 5;
and.b32 %r248, %r247, 255;
popc.b32 %r249, %r248;
shl.b32 %r250, %r249, 7;
and.b32 %r251, %r250, 128;
xor.b32 %r252, %r251, %r248;
mul.lo.s32 %r253, %r252, 16843009;
and.b32 %r254, %r253, 134480385;
or.b32 %r255, %r254, -2139062144;
add.s32 %r256, %r255, -16843009;
or.b32 %r139, %r256, %r254;

prmt.b32 %r139, %r139, 0, 0xba98;

xor.b32 %r257, %r243, %r139;
or.b32 %r258, %r257, -2139062144;
and.b32 %r259, %r139, 2139062143;
and.b32 %r260, %r243, -2139062144;
sub.s32 %r261, %r258, %r259;
xor.b32 %r262, %r260, %r261;
xor.b32 %r142, %r262, -2139062144;
ld.global.nc.u32 %r24, [%rd76+28];

dp4a.s32.s32 %r141, %r142, %r24, %r135;

and.b32 %r263, %r253, -2143281136;
or.b32 %r264, %r263, -2139062144;
add.s32 %r265, %r264, -16843009;
or.b32 %r145, %r265, %r263;

prmt.b32 %r145, %r145, 0, 0xba98;

xor.b32 %r266, %r145, %r244;
or.b32 %r267, %r266, -2139062144;
and.b32 %r268, %r145, 2139062143;
and.b32 %r269, %r244, -2139062144;
sub.s32 %r270, %r267, %r268;
xor.b32 %r271, %r269, %r270;
xor.b32 %r148, %r271, -2139062144;
ld.global.nc.u32 %r25, [%rd76+32];

dp4a.s32.s32 %r147, %r148, %r25, %r141;

shr.u32 %r272, %r153, 11;
or.b32 %r273, %r272, 1;
mul.lo.s32 %r274, %r147, %r273;
shr.s32 %r275, %r274, 31;
shr.u32 %r276, %r275, 29;
add.s32 %r277, %r274, %r276;
shr.s32 %r278, %r277, 3;
add.s64 %rd20, %rd183, %rd7;
ld.global.nc.u16 %rs1, [%rd20];

{ cvt.f32.f16 %f101, %rs1;}


ld.global.nc.u32 %r26, [%rd76];

{.reg .f16 low,high;
mov.b32 {low,high},%r26;
cvt.f32.f16 %f102, low;}


mul.ftz.f32 %f103, %f101, %f102;
cvt.rn.f32.s32 %f104, %r278;
fma.rn.ftz.f32 %f30, %f103, %f104, %f30;
setp.eq.s64 %p13, %rd32, 0;
@%p13 bra $L__BB200_12;

mov.u64 %rd182, iq2xxs_grid;
add.s64 %rd84, %rd184, %rd8;
ld.global.u16 %rs11, [%rd84+-6];
shr.u16 %rs12, %rs11, 8;
ld.global.u16 %rs13, [%rd84+-4];
shr.u16 %rs14, %rs13, 8;
ld.global.u16 %rs15, [%rd84+-2];
cvt.u32.u16 %r328, %rs15;
ld.global.u16 %rs16, [%rd84];
cvt.u32.u16 %r329, %rs16;
mov.b32 %r330, {%rs15, %rs16};
shl.b16 %rs17, %rs11, 3;
cvt.u64.u16 %rd85, %rs17;
and.b64 %rd86, %rd85, 2040;
add.s64 %rd88, %rd182, %rd86;
ld.global.nc.v2.u32 {%r331, %r332}, [%rd88];
and.b32 %r335, %r328, 255;
popc.b32 %r336, %r335;
shl.b32 %r337, %r336, 7;
and.b32 %r338, %r337, 128;
xor.b32 %r339, %r338, %r335;
mul.lo.s32 %r340, %r339, 16843009;
and.b32 %r341, %r340, 134480385;
or.b32 %r342, %r341, -2139062144;
add.s32 %r343, %r342, -16843009;
or.b32 %r279, %r343, %r341;

prmt.b32 %r279, %r279, 0, 0xba98;

xor.b32 %r344, %r331, %r279;
or.b32 %r345, %r344, -2139062144;
and.b32 %r346, %r279, 2139062143;
and.b32 %r347, %r331, -2139062144;
sub.s32 %r348, %r345, %r346;
xor.b32 %r349, %r347, %r348;
xor.b32 %r282, %r349, -2139062144;

dp4a.s32.s32 %r281, %r282, %r18, %r108;

and.b32 %r350, %r340, -2143281136;
or.b32 %r351, %r350, -2139062144;
add.s32 %r352, %r351, -16843009;
or.b32 %r285, %r352, %r350;

prmt.b32 %r285, %r285, 0, 0xba98;

xor.b32 %r353, %r285, %r332;
or.b32 %r354, %r353, -2139062144;
and.b32 %r355, %r285, 2139062143;
and.b32 %r356, %r332, -2139062144;
sub.s32 %r357, %r354, %r355;
xor.b32 %r358, %r356, %r357;
xor.b32 %r288, %r358, -2139062144;

dp4a.s32.s32 %r287, %r288, %r19, %r281;

cvt.u32.u16 %r359, %rs12;
mul.wide.u32 %rd89, %r359, 8;
add.s64 %rd90, %rd182, %rd89;
ld.global.nc.v2.u32 {%r360, %r361}, [%rd90];
shr.u32 %r364, %r328, 7;
and.b32 %r365, %r364, 255;
popc.b32 %r366, %r365;
shl.b32 %r367, %r366, 7;
and.b32 %r368, %r367, 128;
xor.b32 %r369, %r368, %r365;
mul.lo.s32 %r370, %r369, 16843009;
and.b32 %r371, %r370, 134480385;
or.b32 %r372, %r371, -2139062144;
add.s32 %r373, %r372, -16843009;
or.b32 %r291, %r373, %r371;

prmt.b32 %r291, %r291, 0, 0xba98;

xor.b32 %r374, %r360, %r291;
or.b32 %r375, %r374, -2139062144;
and.b32 %r376, %r291, 2139062143;
and.b32 %r377, %r360, -2139062144;
sub.s32 %r378, %r375, %r376;
xor.b32 %r379, %r377, %r378;
xor.b32 %r294, %r379, -2139062144;

dp4a.s32.s32 %r293, %r294, %r20, %r287;

and.b32 %r380, %r370, -2143281136;
or.b32 %r381, %r380, -2139062144;
add.s32 %r382, %r381, -16843009;
or.b32 %r297, %r382, %r380;

prmt.b32 %r297, %r297, 0, 0xba98;

xor.b32 %r383, %r297, %r361;
or.b32 %r384, %r383, -2139062144;
and.b32 %r385, %r297, 2139062143;
and.b32 %r386, %r361, -2139062144;
sub.s32 %r387, %r384, %r385;
xor.b32 %r388, %r386, %r387;
xor.b32 %r300, %r388, -2139062144;

dp4a.s32.s32 %r299, %r300, %r21, %r293;

shl.b16 %rs18, %rs13, 3;
cvt.u64.u16 %rd91, %rs18;
and.b64 %rd92, %rd91, 2040;
add.s64 %rd93, %rd182, %rd92;
ld.global.nc.v2.u32 {%r389, %r390}, [%rd93];
shr.u32 %r393, %r330, 14;
and.b32 %r394, %r393, 255;
popc.b32 %r395, %r394;
shl.b32 %r396, %r395, 7;
and.b32 %r397, %r396, 128;
xor.b32 %r398, %r397, %r394;
mul.lo.s32 %r399, %r398, 16843009;
and.b32 %r400, %r399, 134480385;
or.b32 %r401, %r400, -2139062144;
add.s32 %r402, %r401, -16843009;
or.b32 %r303, %r402, %r400;

prmt.b32 %r303, %r303, 0, 0xba98;

xor.b32 %r403, %r389, %r303;
or.b32 %r404, %r403, -2139062144;
and.b32 %r405, %r303, 2139062143;
and.b32 %r406, %r389, -2139062144;
sub.s32 %r407, %r404, %r405;
xor.b32 %r408, %r406, %r407;
xor.b32 %r306, %r408, -2139062144;

dp4a.s32.s32 %r305, %r306, %r22, %r299;

and.b32 %r409, %r399, -2143281136;
or.b32 %r410, %r409, -2139062144;
add.s32 %r411, %r410, -16843009;
or.b32 %r309, %r411, %r409;

prmt.b32 %r309, %r309, 0, 0xba98;

xor.b32 %r412, %r309, %r390;
or.b32 %r413, %r412, -2139062144;
and.b32 %r414, %r309, 2139062143;
and.b32 %r415, %r390, -2139062144;
sub.s32 %r416, %r413, %r414;
xor.b32 %r417, %r415, %r416;
xor.b32 %r312, %r417, -2139062144;

dp4a.s32.s32 %r311, %r312, %r23, %r305;

cvt.u32.u16 %r418, %rs14;
mul.wide.u32 %rd94, %r418, 8;
add.s64 %rd95, %rd182, %rd94;
ld.global.nc.v2.u32 {%r419, %r420}, [%rd95];
shr.u32 %r423, %r329, 5;
and.b32 %r424, %r423, 255;
popc.b32 %r425, %r424;
shl.b32 %r426, %r425, 7;
and.b32 %r427, %r426, 128;
xor.b32 %r428, %r427, %r424;
mul.lo.s32 %r429, %r428, 16843009;
and.b32 %r430, %r429, 134480385;
or.b32 %r431, %r430, -2139062144;
add.s32 %r432, %r431, -16843009;
or.b32 %r315, %r432, %r430;

prmt.b32 %r315, %r315, 0, 0xba98;

xor.b32 %r433, %r419, %r315;
or.b32 %r434, %r433, -2139062144;
and.b32 %r435, %r315, 2139062143;
and.b32 %r436, %r419, -2139062144;
sub.s32 %r437, %r434, %r435;
xor.b32 %r438, %r436, %r437;
xor.b32 %r318, %r438, -2139062144;

dp4a.s32.s32 %r317, %r318, %r24, %r311;

and.b32 %r439, %r429, -2143281136;
or.b32 %r440, %r439, -2139062144;
add.s32 %r441, %r440, -16843009;
or.b32 %r321, %r441, %r439;

prmt.b32 %r321, %r321, 0, 0xba98;

xor.b32 %r442, %r321, %r420;
or.b32 %r443, %r442, -2139062144;
and.b32 %r444, %r321, 2139062143;
and.b32 %r445, %r420, -2139062144;
sub.s32 %r446, %r443, %r444;
xor.b32 %r447, %r445, %r446;
xor.b32 %r324, %r447, -2139062144;

dp4a.s32.s32 %r323, %r324, %r25, %r317;

shr.u32 %r448, %r329, 11;
or.b32 %r449, %r448, 1;
mul.lo.s32 %r450, %r323, %r449;
shr.s32 %r451, %r450, 31;
shr.u32 %r452, %r451, 29;
add.s32 %r453, %r450, %r452;
shr.s32 %r454, %r453, 3;
add.s64 %rd96, %rd184, %rd7;
ld.global.u16 %rs10, [%rd96];

{ cvt.f32.f16 %f105, %rs10;}



{.reg .f16 low,high;
mov.b32 {low,high},%r26;
cvt.f32.f16 %f106, low;}


mul.ftz.f32 %f107, %f105, %f106;
cvt.rn.f32.s32 %f108, %r454;
fma.rn.ftz.f32 %f34, %f107, %f108, %f34;

$L__BB200_12:
mov.u64 %rd179, iq2xxs_grid;
add.s64 %rd174, %rd183, %rd8;
add.s64 %rd173, %rd174, -6;
add.s64 %rd21, %rd173, %rd14;
ld.global.nc.u16 %rs20, [%rd21];
shr.u16 %rs21, %rs20, 8;
ld.global.nc.u16 %rs22, [%rd21+2];
shr.u16 %rs23, %rs22, 8;
ld.global.nc.u16 %rs24, [%rd21+4];
cvt.u32.u16 %r504, %rs24;
ld.global.nc.u16 %rs25, [%rd21+6];
cvt.u32.u16 %r505, %rs25;
mov.b32 %r506, {%rs24, %rs25};
shl.b16 %rs26, %rs20, 3;
cvt.u64.u16 %rd97, %rs26;
and.b64 %rd98, %rd97, 2040;
add.s64 %rd100, %rd179, %rd98;
ld.global.nc.v2.u32 {%r507, %r508}, [%rd100];
and.b32 %r511, %r504, 255;
popc.b32 %r512, %r511;
shl.b32 %r513, %r512, 7;
and.b32 %r514, %r513, 128;
xor.b32 %r515, %r514, %r511;
mul.lo.s32 %r516, %r515, 16843009;
and.b32 %r517, %r516, 134480385;
or.b32 %r518, %r517, -2139062144;
add.s32 %r519, %r518, -16843009;
or.b32 %r455, %r519, %r517;

prmt.b32 %r455, %r455, 0, 0xba98;

xor.b32 %r520, %r507, %r455;
or.b32 %r521, %r520, -2139062144;
and.b32 %r522, %r455, 2139062143;
and.b32 %r523, %r507, -2139062144;
sub.s32 %r524, %r521, %r522;
xor.b32 %r525, %r523, %r524;
xor.b32 %r458, %r525, -2139062144;

dp4a.s32.s32 %r457, %r458, %r18, %r108;

and.b32 %r526, %r516, -2143281136;
or.b32 %r527, %r526, -2139062144;
add.s32 %r528, %r527, -16843009;
or.b32 %r461, %r528, %r526;

prmt.b32 %r461, %r461, 0, 0xba98;

xor.b32 %r529, %r461, %r508;
or.b32 %r530, %r529, -2139062144;
and.b32 %r531, %r461, 2139062143;
and.b32 %r532, %r508, -2139062144;
sub.s32 %r533, %r530, %r531;
xor.b32 %r534, %r532, %r533;
xor.b32 %r464, %r534, -2139062144;

dp4a.s32.s32 %r463, %r464, %r19, %r457;

cvt.u32.u16 %r535, %rs21;
mul.wide.u32 %rd101, %r535, 8;
add.s64 %rd102, %rd179, %rd101;
ld.global.nc.v2.u32 {%r536, %r537}, [%rd102];
shr.u32 %r540, %r504, 7;
and.b32 %r541, %r540, 255;
popc.b32 %r542, %r541;
shl.b32 %r543, %r542, 7;
and.b32 %r544, %r543, 128;
xor.b32 %r545, %r544, %r541;
mul.lo.s32 %r546, %r545, 16843009;
and.b32 %r547, %r546, 134480385;
or.b32 %r548, %r547, -2139062144;
add.s32 %r549, %r548, -16843009;
or.b32 %r467, %r549, %r547;

prmt.b32 %r467, %r467, 0, 0xba98;

xor.b32 %r550, %r536, %r467;
or.b32 %r551, %r550, -2139062144;
and.b32 %r552, %r467, 2139062143;
and.b32 %r553, %r536, -2139062144;
sub.s32 %r554, %r551, %r552;
xor.b32 %r555, %r553, %r554;
xor.b32 %r470, %r555, -2139062144;

dp4a.s32.s32 %r469, %r470, %r20, %r463;

and.b32 %r556, %r546, -2143281136;
or.b32 %r557, %r556, -2139062144;
add.s32 %r558, %r557, -16843009;
or.b32 %r473, %r558, %r556;

prmt.b32 %r473, %r473, 0, 0xba98;

xor.b32 %r559, %r473, %r537;
or.b32 %r560, %r559, -2139062144;
and.b32 %r561, %r473, 2139062143;
and.b32 %r562, %r537, -2139062144;
sub.s32 %r563, %r560, %r561;
xor.b32 %r564, %r562, %r563;
xor.b32 %r476, %r564, -2139062144;

dp4a.s32.s32 %r475, %r476, %r21, %r469;

shl.b16 %rs27, %rs22, 3;
cvt.u64.u16 %rd103, %rs27;
and.b64 %rd104, %rd103, 2040;
add.s64 %rd105, %rd179, %rd104;
ld.global.nc.v2.u32 {%r565, %r566}, [%rd105];
shr.u32 %r569, %r506, 14;
and.b32 %r570, %r569, 255;
popc.b32 %r571, %r570;
shl.b32 %r572, %r571, 7;
and.b32 %r573, %r572, 128;
xor.b32 %r574, %r573, %r570;
mul.lo.s32 %r575, %r574, 16843009;
and.b32 %r576, %r575, 134480385;
or.b32 %r577, %r576, -2139062144;
add.s32 %r578, %r577, -16843009;
or.b32 %r479, %r578, %r576;

prmt.b32 %r479, %r479, 0, 0xba98;

xor.b32 %r579, %r565, %r479;
or.b32 %r580, %r579, -2139062144;
and.b32 %r581, %r479, 2139062143;
and.b32 %r582, %r565, -2139062144;
sub.s32 %r583, %r580, %r581;
xor.b32 %r584, %r582, %r583;
xor.b32 %r482, %r584, -2139062144;

dp4a.s32.s32 %r481, %r482, %r22, %r475;

and.b32 %r585, %r575, -2143281136;
or.b32 %r586, %r585, -2139062144;
add.s32 %r587, %r586, -16843009;
or.b32 %r485, %r587, %r585;

prmt.b32 %r485, %r485, 0, 0xba98;

xor.b32 %r588, %r485, %r566;
or.b32 %r589, %r588, -2139062144;
and.b32 %r590, %r485, 2139062143;
and.b32 %r591, %r566, -2139062144;
sub.s32 %r592, %r589, %r590;
xor.b32 %r593, %r591, %r592;
xor.b32 %r488, %r593, -2139062144;

dp4a.s32.s32 %r487, %r488, %r23, %r481;

cvt.u32.u16 %r594, %rs23;
mul.wide.u32 %rd106, %r594, 8;
add.s64 %rd107, %rd179, %rd106;
ld.global.nc.v2.u32 {%r595, %r596}, [%rd107];
shr.u32 %r599, %r505, 5;
and.b32 %r600, %r599, 255;
popc.b32 %r601, %r600;
shl.b32 %r602, %r601, 7;
and.b32 %r603, %r602, 128;
xor.b32 %r604, %r603, %r600;
mul.lo.s32 %r605, %r604, 16843009;
and.b32 %r606, %r605, 134480385;
or.b32 %r607, %r606, -2139062144;
add.s32 %r608, %r607, -16843009;
or.b32 %r491, %r608, %r606;

prmt.b32 %r491, %r491, 0, 0xba98;

xor.b32 %r609, %r595, %r491;
or.b32 %r610, %r609, -2139062144;
and.b32 %r611, %r491, 2139062143;
and.b32 %r612, %r595, -2139062144;
sub.s32 %r613, %r610, %r611;
xor.b32 %r614, %r612, %r613;
xor.b32 %r494, %r614, -2139062144;

dp4a.s32.s32 %r493, %r494, %r24, %r487;

and.b32 %r615, %r605, -2143281136;
or.b32 %r616, %r615, -2139062144;
add.s32 %r617, %r616, -16843009;
or.b32 %r497, %r617, %r615;

prmt.b32 %r497, %r497, 0, 0xba98;

xor.b32 %r618, %r497, %r596;
or.b32 %r619, %r618, -2139062144;
and.b32 %r620, %r497, 2139062143;
and.b32 %r621, %r596, -2139062144;
sub.s32 %r622, %r619, %r620;
xor.b32 %r623, %r621, %r622;
xor.b32 %r500, %r623, -2139062144;

dp4a.s32.s32 %r499, %r500, %r25, %r493;

shr.u32 %r624, %r505, 11;
or.b32 %r625, %r624, 1;
mul.lo.s32 %r626, %r499, %r625;
shr.s32 %r627, %r626, 31;
shr.u32 %r628, %r627, 29;
add.s32 %r629, %r626, %r628;
shr.s32 %r630, %r629, 3;
add.s64 %rd22, %rd20, %rd14;
ld.global.nc.u16 %rs19, [%rd22];

{ cvt.f32.f16 %f109, %rs19;}



{.reg .f16 low,high;
mov.b32 {low,high},%r26;
cvt.f32.f16 %f110, low;}


mul.ftz.f32 %f111, %f109, %f110;
cvt.rn.f32.s32 %f112, %r630;
fma.rn.ftz.f32 %f29, %f111, %f112, %f29;
@%p13 bra $L__BB200_14;

mov.u64 %rd181, iq2xxs_grid;
add.s64 %rd108, %rd184, %rd9;
ld.global.u16 %rs29, [%rd108+-2];
shr.u16 %rs30, %rs29, 8;
ld.global.u16 %rs31, [%rd108];
shr.u16 %rs32, %rs31, 8;
ld.global.u16 %rs33, [%rd108+2];
cvt.u32.u16 %r680, %rs33;
ld.global.u16 %rs34, [%rd108+4];
cvt.u32.u16 %r681, %rs34;
mov.b32 %r682, {%rs33, %rs34};
shl.b16 %rs35, %rs29, 3;
cvt.u64.u16 %rd109, %rs35;
and.b64 %rd110, %rd109, 2040;
add.s64 %rd112, %rd181, %rd110;
ld.global.nc.v2.u32 {%r683, %r684}, [%rd112];
mov.u32 %r636, 0;
and.b32 %r687, %r680, 255;
popc.b32 %r688, %r687;
shl.b32 %r689, %r688, 7;
and.b32 %r690, %r689, 128;
xor.b32 %r691, %r690, %r687;
mul.lo.s32 %r692, %r691, 16843009;
and.b32 %r693, %r692, 134480385;
or.b32 %r694, %r693, -2139062144;
add.s32 %r695, %r694, -16843009;
or.b32 %r631, %r695, %r693;

prmt.b32 %r631, %r631, 0, 0xba98;

xor.b32 %r696, %r683, %r631;
or.b32 %r697, %r696, -2139062144;
and.b32 %r698, %r631, 2139062143;
and.b32 %r699, %r683, -2139062144;
sub.s32 %r700, %r697, %r698;
xor.b32 %r701, %r699, %r700;
xor.b32 %r634, %r701, -2139062144;

dp4a.s32.s32 %r633, %r634, %r18, %r636;

and.b32 %r702, %r692, -2143281136;
or.b32 %r703, %r702, -2139062144;
add.s32 %r704, %r703, -16843009;
or.b32 %r637, %r704, %r702;

prmt.b32 %r637, %r637, 0, 0xba98;

xor.b32 %r705, %r637, %r684;
or.b32 %r706, %r705, -2139062144;
and.b32 %r707, %r637, 2139062143;
and.b32 %r708, %r684, -2139062144;
sub.s32 %r709, %r706, %r707;
xor.b32 %r710, %r708, %r709;
xor.b32 %r640, %r710, -2139062144;

dp4a.s32.s32 %r639, %r640, %r19, %r633;

cvt.u32.u16 %r711, %rs30;
mul.wide.u32 %rd113, %r711, 8;
add.s64 %rd114, %rd181, %rd113;
ld.global.nc.v2.u32 {%r712, %r713}, [%rd114];
shr.u32 %r716, %r680, 7;
and.b32 %r717, %r716, 255;
popc.b32 %r718, %r717;
shl.b32 %r719, %r718, 7;
and.b32 %r720, %r719, 128;
xor.b32 %r721, %r720, %r717;
mul.lo.s32 %r722, %r721, 16843009;
and.b32 %r723, %r722, 134480385;
or.b32 %r724, %r723, -2139062144;
add.s32 %r725, %r724, -16843009;
or.b32 %r643, %r725, %r723;

prmt.b32 %r643, %r643, 0, 0xba98;

xor.b32 %r726, %r712, %r643;
or.b32 %r727, %r726, -2139062144;
and.b32 %r728, %r643, 2139062143;
and.b32 %r729, %r712, -2139062144;
sub.s32 %r730, %r727, %r728;
xor.b32 %r731, %r729, %r730;
xor.b32 %r646, %r731, -2139062144;

dp4a.s32.s32 %r645, %r646, %r20, %r639;

and.b32 %r732, %r722, -2143281136;
or.b32 %r733, %r732, -2139062144;
add.s32 %r734, %r733, -16843009;
or.b32 %r649, %r734, %r732;

prmt.b32 %r649, %r649, 0, 0xba98;

xor.b32 %r735, %r649, %r713;
or.b32 %r736, %r735, -2139062144;
and.b32 %r737, %r649, 2139062143;
and.b32 %r738, %r713, -2139062144;
sub.s32 %r739, %r736, %r737;
xor.b32 %r740, %r738, %r739;
xor.b32 %r652, %r740, -2139062144;

dp4a.s32.s32 %r651, %r652, %r21, %r645;

shl.b16 %rs36, %rs31, 3;
cvt.u64.u16 %rd115, %rs36;
and.b64 %rd116, %rd115, 2040;
add.s64 %rd117, %rd181, %rd116;
ld.global.nc.v2.u32 {%r741, %r742}, [%rd117];
shr.u32 %r745, %r682, 14;
and.b32 %r746, %r745, 255;
popc.b32 %r747, %r746;
shl.b32 %r748, %r747, 7;
and.b32 %r749, %r748, 128;
xor.b32 %r750, %r749, %r746;
mul.lo.s32 %r751, %r750, 16843009;
and.b32 %r752, %r751, 134480385;
or.b32 %r753, %r752, -2139062144;
add.s32 %r754, %r753, -16843009;
or.b32 %r655, %r754, %r752;

prmt.b32 %r655, %r655, 0, 0xba98;

xor.b32 %r755, %r741, %r655;
or.b32 %r756, %r755, -2139062144;
and.b32 %r757, %r655, 2139062143;
and.b32 %r758, %r741, -2139062144;
sub.s32 %r759, %r756, %r757;
xor.b32 %r760, %r758, %r759;
xor.b32 %r658, %r760, -2139062144;

dp4a.s32.s32 %r657, %r658, %r22, %r651;

and.b32 %r761, %r751, -2143281136;
or.b32 %r762, %r761, -2139062144;
add.s32 %r763, %r762, -16843009;
or.b32 %r661, %r763, %r761;

prmt.b32 %r661, %r661, 0, 0xba98;

xor.b32 %r764, %r661, %r742;
or.b32 %r765, %r764, -2139062144;
and.b32 %r766, %r661, 2139062143;
and.b32 %r767, %r742, -2139062144;
sub.s32 %r768, %r765, %r766;
xor.b32 %r769, %r767, %r768;
xor.b32 %r664, %r769, -2139062144;

dp4a.s32.s32 %r663, %r664, %r23, %r657;

cvt.u32.u16 %r770, %rs32;
mul.wide.u32 %rd118, %r770, 8;
add.s64 %rd119, %rd181, %rd118;
ld.global.nc.v2.u32 {%r771, %r772}, [%rd119];
shr.u32 %r775, %r681, 5;
and.b32 %r776, %r775, 255;
popc.b32 %r777, %r776;
shl.b32 %r778, %r777, 7;
and.b32 %r779, %r778, 128;
xor.b32 %r780, %r779, %r776;
mul.lo.s32 %r781, %r780, 16843009;
and.b32 %r782, %r781, 134480385;
or.b32 %r783, %r782, -2139062144;
add.s32 %r784, %r783, -16843009;
or.b32 %r667, %r784, %r782;

prmt.b32 %r667, %r667, 0, 0xba98;

xor.b32 %r785, %r771, %r667;
or.b32 %r786, %r785, -2139062144;
and.b32 %r787, %r667, 2139062143;
and.b32 %r788, %r771, -2139062144;
sub.s32 %r789, %r786, %r787;
xor.b32 %r790, %r788, %r789;
xor.b32 %r670, %r790, -2139062144;

dp4a.s32.s32 %r669, %r670, %r24, %r663;

and.b32 %r791, %r781, -2143281136;
or.b32 %r792, %r791, -2139062144;
add.s32 %r793, %r792, -16843009;
or.b32 %r673, %r793, %r791;

prmt.b32 %r673, %r673, 0, 0xba98;

xor.b32 %r794, %r673, %r772;
or.b32 %r795, %r794, -2139062144;
and.b32 %r796, %r673, 2139062143;
and.b32 %r797, %r772, -2139062144;
sub.s32 %r798, %r795, %r796;
xor.b32 %r799, %r797, %r798;
xor.b32 %r676, %r799, -2139062144;

dp4a.s32.s32 %r675, %r676, %r25, %r669;

shr.u32 %r800, %r681, 11;
or.b32 %r801, %r800, 1;
mul.lo.s32 %r802, %r675, %r801;
shr.s32 %r803, %r802, 31;
shr.u32 %r804, %r803, 29;
add.s32 %r805, %r802, %r804;
shr.s32 %r806, %r805, 3;
add.s64 %rd120, %rd184, %rd6;
ld.global.u16 %rs28, [%rd120];

{ cvt.f32.f16 %f113, %rs28;}



{.reg .f16 low,high;
mov.b32 {low,high},%r26;
cvt.f32.f16 %f114, low;}


mul.ftz.f32 %f115, %f113, %f114;
cvt.rn.f32.s32 %f116, %r806;
fma.rn.ftz.f32 %f33, %f115, %f116, %f33;

$L__BB200_14:
mov.u64 %rd180, iq2xxs_grid;
add.s64 %rd23, %rd21, %rd14;
ld.global.nc.u16 %rs38, [%rd23];
shr.u16 %rs39, %rs38, 8;
ld.global.nc.u16 %rs40, [%rd23+2];
shr.u16 %rs41, %rs40, 8;
ld.global.nc.u16 %rs42, [%rd23+4];
cvt.u32.u16 %r856, %rs42;
ld.global.nc.u16 %rs43, [%rd23+6];
cvt.u32.u16 %r857, %rs43;
mov.b32 %r858, {%rs42, %rs43};
shl.b16 %rs44, %rs38, 3;
cvt.u64.u16 %rd121, %rs44;
and.b64 %rd122, %rd121, 2040;
add.s64 %rd124, %rd180, %rd122;
ld.global.nc.v2.u32 {%r859, %r860}, [%rd124];
mov.u32 %r812, 0;
and.b32 %r863, %r856, 255;
popc.b32 %r864, %r863;
shl.b32 %r865, %r864, 7;
and.b32 %r866, %r865, 128;
xor.b32 %r867, %r866, %r863;
mul.lo.s32 %r868, %r867, 16843009;
and.b32 %r869, %r868, 134480385;
or.b32 %r870, %r869, -2139062144;
add.s32 %r871, %r870, -16843009;
or.b32 %r807, %r871, %r869;

prmt.b32 %r807, %r807, 0, 0xba98;

xor.b32 %r872, %r859, %r807;
or.b32 %r873, %r872, -2139062144;
and.b32 %r874, %r807, 2139062143;
and.b32 %r875, %r859, -2139062144;
sub.s32 %r876, %r873, %r874;
xor.b32 %r877, %r875, %r876;
xor.b32 %r810, %r877, -2139062144;

dp4a.s32.s32 %r809, %r810, %r18, %r812;

and.b32 %r878, %r868, -2143281136;
or.b32 %r879, %r878, -2139062144;
add.s32 %r880, %r879, -16843009;
or.b32 %r813, %r880, %r878;

prmt.b32 %r813, %r813, 0, 0xba98;

xor.b32 %r881, %r813, %r860;
or.b32 %r882, %r881, -2139062144;
and.b32 %r883, %r813, 2139062143;
and.b32 %r884, %r860, -2139062144;
sub.s32 %r885, %r882, %r883;
xor.b32 %r886, %r884, %r885;
xor.b32 %r816, %r886, -2139062144;

dp4a.s32.s32 %r815, %r816, %r19, %r809;

cvt.u32.u16 %r887, %rs39;
mul.wide.u32 %rd125, %r887, 8;
add.s64 %rd126, %rd180, %rd125;
ld.global.nc.v2.u32 {%r888, %r889}, [%rd126];
shr.u32 %r892, %r856, 7;
and.b32 %r893, %r892, 255;
popc.b32 %r894, %r893;
shl.b32 %r895, %r894, 7;
and.b32 %r896, %r895, 128;
xor.b32 %r897, %r896, %r893;
mul.lo.s32 %r898, %r897, 16843009;
and.b32 %r899, %r898, 134480385;
or.b32 %r900, %r899, -2139062144;
add.s32 %r901, %r900, -16843009;
or.b32 %r819, %r901, %r899;

prmt.b32 %r819, %r819, 0, 0xba98;

xor.b32 %r902, %r888, %r819;
or.b32 %r903, %r902, -2139062144;
and.b32 %r904, %r819, 2139062143;
and.b32 %r905, %r888, -2139062144;
sub.s32 %r906, %r903, %r904;
xor.b32 %r907, %r905, %r906;
xor.b32 %r822, %r907, -2139062144;

dp4a.s32.s32 %r821, %r822, %r20, %r815;

and.b32 %r908, %r898, -2143281136;
or.b32 %r909, %r908, -2139062144;
add.s32 %r910, %r909, -16843009;
or.b32 %r825, %r910, %r908;

prmt.b32 %r825, %r825, 0, 0xba98;

xor.b32 %r911, %r825, %r889;
or.b32 %r912, %r911, -2139062144;
and.b32 %r913, %r825, 2139062143;
and.b32 %r914, %r889, -2139062144;
sub.s32 %r915, %r912, %r913;
xor.b32 %r916, %r914, %r915;
xor.b32 %r828, %r916, -2139062144;

dp4a.s32.s32 %r827, %r828, %r21, %r821;

shl.b16 %rs45, %rs40, 3;
cvt.u64.u16 %rd127, %rs45;
and.b64 %rd128, %rd127, 2040;
add.s64 %rd129, %rd180, %rd128;
ld.global.nc.v2.u32 {%r917, %r918}, [%rd129];
shr.u32 %r921, %r858, 14;
and.b32 %r922, %r921, 255;
popc.b32 %r923, %r922;
shl.b32 %r924, %r923, 7;
and.b32 %r925, %r924, 128;
xor.b32 %r926, %r925, %r922;
mul.lo.s32 %r927, %r926, 16843009;
and.b32 %r928, %r927, 134480385;
or.b32 %r929, %r928, -2139062144;
add.s32 %r930, %r929, -16843009;
or.b32 %r831, %r930, %r928;

prmt.b32 %r831, %r831, 0, 0xba98;

xor.b32 %r931, %r917, %r831;
or.b32 %r932, %r931, -2139062144;
and.b32 %r933, %r831, 2139062143;
and.b32 %r934, %r917, -2139062144;
sub.s32 %r935, %r932, %r933;
xor.b32 %r936, %r934, %r935;
xor.b32 %r834, %r936, -2139062144;

dp4a.s32.s32 %r833, %r834, %r22, %r827;

and.b32 %r937, %r927, -2143281136;
or.b32 %r938, %r937, -2139062144;
add.s32 %r939, %r938, -16843009;
or.b32 %r837, %r939, %r937;

prmt.b32 %r837, %r837, 0, 0xba98;

xor.b32 %r940, %r837, %r918;
or.b32 %r941, %r940, -2139062144;
and.b32 %r942, %r837, 2139062143;
and.b32 %r943, %r918, -2139062144;
sub.s32 %r944, %r941, %r942;
xor.b32 %r945, %r943, %r944;
xor.b32 %r840, %r945, -2139062144;

dp4a.s32.s32 %r839, %r840, %r23, %r833;

cvt.u32.u16 %r946, %rs41;
mul.wide.u32 %rd130, %r946, 8;
add.s64 %rd131, %rd180, %rd130;
ld.global.nc.v2.u32 {%r947, %r948}, [%rd131];
shr.u32 %r951, %r857, 5;
and.b32 %r952, %r951, 255;
popc.b32 %r953, %r952;
shl.b32 %r954, %r953, 7;
and.b32 %r955, %r954, 128;
xor.b32 %r956, %r955, %r952;
mul.lo.s32 %r957, %r956, 16843009;
and.b32 %r958, %r957, 134480385;
or.b32 %r959, %r958, -2139062144;
add.s32 %r960, %r959, -16843009;
or.b32 %r843, %r960, %r958;

prmt.b32 %r843, %r843, 0, 0xba98;

xor.b32 %r961, %r947, %r843;
or.b32 %r962, %r961, -2139062144;
and.b32 %r963, %r843, 2139062143;
and.b32 %r964, %r947, -2139062144;
sub.s32 %r965, %r962, %r963;
xor.b32 %r966, %r964, %r965;
xor.b32 %r846, %r966, -2139062144;

dp4a.s32.s32 %r845, %r846, %r24, %r839;

and.b32 %r967, %r957, -2143281136;
or.b32 %r968, %r967, -2139062144;
add.s32 %r969, %r968, -16843009;
or.b32 %r849, %r969, %r967;

prmt.b32 %r849, %r849, 0, 0xba98;

xor.b32 %r970, %r849, %r948;
or.b32 %r971, %r970, -2139062144;
and.b32 %r972, %r849, 2139062143;
and.b32 %r973, %r948, -2139062144;
sub.s32 %r974, %r971, %r972;
xor.b32 %r975, %r973, %r974;
xor.b32 %r852, %r975, -2139062144;

dp4a.s32.s32 %r851, %r852, %r25, %r845;

shr.u32 %r976, %r857, 11;
or.b32 %r977, %r976, 1;
mul.lo.s32 %r978, %r851, %r977;
shr.s32 %r979, %r978, 31;
shr.u32 %r980, %r979, 29;
add.s32 %r981, %r978, %r980;
shr.s32 %r982, %r981, 3;
add.s64 %rd24, %rd22, %rd14;
ld.global.nc.u16 %rs37, [%rd24];

{ cvt.f32.f16 %f117, %rs37;}



{.reg .f16 low,high;
mov.b32 {low,high},%r26;
cvt.f32.f16 %f118, low;}


mul.ftz.f32 %f119, %f117, %f118;
cvt.rn.f32.s32 %f120, %r982;
fma.rn.ftz.f32 %f28, %f119, %f120, %f28;
@%p13 bra $L__BB200_16;

mov.u64 %rd178, iq2xxs_grid;
add.s64 %rd132, %rd184, %rd13;
ld.global.u16 %rs47, [%rd132+2];
shr.u16 %rs48, %rs47, 8;
ld.global.u16 %rs49, [%rd132+4];
shr.u16 %rs50, %rs49, 8;
ld.global.u16 %rs51, [%rd132+6];
cvt.u32.u16 %r1032, %rs51;
ld.global.u16 %rs52, [%rd132+8];
cvt.u32.u16 %r1033, %rs52;
mov.b32 %r1034, {%rs51, %rs52};
shl.b16 %rs53, %rs47, 3;
cvt.u64.u16 %rd133, %rs53;
and.b64 %rd134, %rd133, 2040;
add.s64 %rd136, %rd178, %rd134;
ld.global.nc.v2.u32 {%r1035, %r1036}, [%rd136];
and.b32 %r1039, %r1032, 255;
popc.b32 %r1040, %r1039;
shl.b32 %r1041, %r1040, 7;
and.b32 %r1042, %r1041, 128;
xor.b32 %r1043, %r1042, %r1039;
mul.lo.s32 %r1044, %r1043, 16843009;
and.b32 %r1045, %r1044, 134480385;
or.b32 %r1046, %r1045, -2139062144;
add.s32 %r1047, %r1046, -16843009;
or.b32 %r983, %r1047, %r1045;

prmt.b32 %r983, %r983, 0, 0xba98;

xor.b32 %r1048, %r1035, %r983;
or.b32 %r1049, %r1048, -2139062144;
and.b32 %r1050, %r983, 2139062143;
and.b32 %r1051, %r1035, -2139062144;
sub.s32 %r1052, %r1049, %r1050;
xor.b32 %r1053, %r1051, %r1052;
xor.b32 %r986, %r1053, -2139062144;

dp4a.s32.s32 %r985, %r986, %r18, %r812;

and.b32 %r1054, %r1044, -2143281136;
or.b32 %r1055, %r1054, -2139062144;
add.s32 %r1056, %r1055, -16843009;
or.b32 %r989, %r1056, %r1054;

prmt.b32 %r989, %r989, 0, 0xba98;

xor.b32 %r1057, %r989, %r1036;
or.b32 %r1058, %r1057, -2139062144;
and.b32 %r1059, %r989, 2139062143;
and.b32 %r1060, %r1036, -2139062144;
sub.s32 %r1061, %r1058, %r1059;
xor.b32 %r1062, %r1060, %r1061;
xor.b32 %r992, %r1062, -2139062144;

dp4a.s32.s32 %r991, %r992, %r19, %r985;

cvt.u32.u16 %r1063, %rs48;
mul.wide.u32 %rd137, %r1063, 8;
add.s64 %rd138, %rd178, %rd137;
ld.global.nc.v2.u32 {%r1064, %r1065}, [%rd138];
shr.u32 %r1068, %r1032, 7;
and.b32 %r1069, %r1068, 255;
popc.b32 %r1070, %r1069;
shl.b32 %r1071, %r1070, 7;
and.b32 %r1072, %r1071, 128;
xor.b32 %r1073, %r1072, %r1069;
mul.lo.s32 %r1074, %r1073, 16843009;
and.b32 %r1075, %r1074, 134480385;
or.b32 %r1076, %r1075, -2139062144;
add.s32 %r1077, %r1076, -16843009;
or.b32 %r995, %r1077, %r1075;

prmt.b32 %r995, %r995, 0, 0xba98;

xor.b32 %r1078, %r1064, %r995;
or.b32 %r1079, %r1078, -2139062144;
and.b32 %r1080, %r995, 2139062143;
and.b32 %r1081, %r1064, -2139062144;
sub.s32 %r1082, %r1079, %r1080;
xor.b32 %r1083, %r1081, %r1082;
xor.b32 %r998, %r1083, -2139062144;

dp4a.s32.s32 %r997, %r998, %r20, %r991;

and.b32 %r1084, %r1074, -2143281136;
or.b32 %r1085, %r1084, -2139062144;
add.s32 %r1086, %r1085, -16843009;
or.b32 %r1001, %r1086, %r1084;

prmt.b32 %r1001, %r1001, 0, 0xba98;

xor.b32 %r1087, %r1001, %r1065;
or.b32 %r1088, %r1087, -2139062144;
and.b32 %r1089, %r1001, 2139062143;
and.b32 %r1090, %r1065, -2139062144;
sub.s32 %r1091, %r1088, %r1089;
xor.b32 %r1092, %r1090, %r1091;
xor.b32 %r1004, %r1092, -2139062144;

dp4a.s32.s32 %r1003, %r1004, %r21, %r997;

shl.b16 %rs54, %rs49, 3;
cvt.u64.u16 %rd139, %rs54;
and.b64 %rd140, %rd139, 2040;
add.s64 %rd141, %rd178, %rd140;
ld.global.nc.v2.u32 {%r1093, %r1094}, [%rd141];
shr.u32 %r1097, %r1034, 14;
and.b32 %r1098, %r1097, 255;
popc.b32 %r1099, %r1098;
shl.b32 %r1100, %r1099, 7;
and.b32 %r1101, %r1100, 128;
xor.b32 %r1102, %r1101, %r1098;
mul.lo.s32 %r1103, %r1102, 16843009;
and.b32 %r1104, %r1103, 134480385;
or.b32 %r1105, %r1104, -2139062144;
add.s32 %r1106, %r1105, -16843009;
or.b32 %r1007, %r1106, %r1104;

prmt.b32 %r1007, %r1007, 0, 0xba98;

xor.b32 %r1107, %r1093, %r1007;
or.b32 %r1108, %r1107, -2139062144;
and.b32 %r1109, %r1007, 2139062143;
and.b32 %r1110, %r1093, -2139062144;
sub.s32 %r1111, %r1108, %r1109;
xor.b32 %r1112, %r1110, %r1111;
xor.b32 %r1010, %r1112, -2139062144;

dp4a.s32.s32 %r1009, %r1010, %r22, %r1003;

and.b32 %r1113, %r1103, -2143281136;
or.b32 %r1114, %r1113, -2139062144;
add.s32 %r1115, %r1114, -16843009;
or.b32 %r1013, %r1115, %r1113;

prmt.b32 %r1013, %r1013, 0, 0xba98;

xor.b32 %r1116, %r1013, %r1094;
or.b32 %r1117, %r1116, -2139062144;
and.b32 %r1118, %r1013, 2139062143;
and.b32 %r1119, %r1094, -2139062144;
sub.s32 %r1120, %r1117, %r1118;
xor.b32 %r1121, %r1119, %r1120;
xor.b32 %r1016, %r1121, -2139062144;

dp4a.s32.s32 %r1015, %r1016, %r23, %r1009;

cvt.u32.u16 %r1122, %rs50;
mul.wide.u32 %rd142, %r1122, 8;
add.s64 %rd143, %rd178, %rd142;
ld.global.nc.v2.u32 {%r1123, %r1124}, [%rd143];
shr.u32 %r1127, %r1033, 5;
and.b32 %r1128, %r1127, 255;
popc.b32 %r1129, %r1128;
shl.b32 %r1130, %r1129, 7;
and.b32 %r1131, %r1130, 128;
xor.b32 %r1132, %r1131, %r1128;
mul.lo.s32 %r1133, %r1132, 16843009;
and.b32 %r1134, %r1133, 134480385;
or.b32 %r1135, %r1134, -2139062144;
add.s32 %r1136, %r1135, -16843009;
or.b32 %r1019, %r1136, %r1134;

prmt.b32 %r1019, %r1019, 0, 0xba98;

xor.b32 %r1137, %r1123, %r1019;
or.b32 %r1138, %r1137, -2139062144;
and.b32 %r1139, %r1019, 2139062143;
and.b32 %r1140, %r1123, -2139062144;
sub.s32 %r1141, %r1138, %r1139;
xor.b32 %r1142, %r1140, %r1141;
xor.b32 %r1022, %r1142, -2139062144;

dp4a.s32.s32 %r1021, %r1022, %r24, %r1015;

and.b32 %r1143, %r1133, -2143281136;
or.b32 %r1144, %r1143, -2139062144;
add.s32 %r1145, %r1144, -16843009;
or.b32 %r1025, %r1145, %r1143;

prmt.b32 %r1025, %r1025, 0, 0xba98;

xor.b32 %r1146, %r1025, %r1124;
or.b32 %r1147, %r1146, -2139062144;
and.b32 %r1148, %r1025, 2139062143;
and.b32 %r1149, %r1124, -2139062144;
sub.s32 %r1150, %r1147, %r1148;
xor.b32 %r1151, %r1149, %r1150;
xor.b32 %r1028, %r1151, -2139062144;

dp4a.s32.s32 %r1027, %r1028, %r25, %r1021;

shr.u32 %r1152, %r1033, 11;
or.b32 %r1153, %r1152, 1;
mul.lo.s32 %r1154, %r1027, %r1153;
shr.s32 %r1155, %r1154, 31;
shr.u32 %r1156, %r1155, 29;
add.s32 %r1157, %r1154, %r1156;
shr.s32 %r1158, %r1157, 3;
add.s64 %rd144, %rd184, %rd12;
ld.global.u16 %rs46, [%rd144];

{ cvt.f32.f16 %f121, %rs46;}



{.reg .f16 low,high;
mov.b32 {low,high},%r26;
cvt.f32.f16 %f122, low;}


mul.ftz.f32 %f123, %f121, %f122;
cvt.rn.f32.s32 %f124, %r1158;
fma.rn.ftz.f32 %f32, %f123, %f124, %f32;

$L__BB200_16:
mov.u64 %rd176, iq2xxs_grid;
add.s64 %rd145, %rd23, %rd14;
ld.global.nc.u16 %rs56, [%rd145];
shr.u16 %rs57, %rs56, 8;
ld.global.nc.u16 %rs58, [%rd145+2];
shr.u16 %rs59, %rs58, 8;
ld.global.nc.u16 %rs60, [%rd145+4];
cvt.u32.u16 %r1208, %rs60;
ld.global.nc.u16 %rs61, [%rd145+6];
cvt.u32.u16 %r1209, %rs61;
mov.b32 %r1210, {%rs60, %rs61};
shl.b16 %rs62, %rs56, 3;
cvt.u64.u16 %rd146, %rs62;
and.b64 %rd147, %rd146, 2040;
add.s64 %rd149, %rd176, %rd147;
ld.global.nc.v2.u32 {%r1211, %r1212}, [%rd149];
and.b32 %r1215, %r1208, 255;
popc.b32 %r1216, %r1215;
shl.b32 %r1217, %r1216, 7;
and.b32 %r1218, %r1217, 128;
xor.b32 %r1219, %r1218, %r1215;
mul.lo.s32 %r1220, %r1219, 16843009;
and.b32 %r1221, %r1220, 134480385;
or.b32 %r1222, %r1221, -2139062144;
add.s32 %r1223, %r1222, -16843009;
or.b32 %r1159, %r1223, %r1221;

prmt.b32 %r1159, %r1159, 0, 0xba98;

xor.b32 %r1224, %r1211, %r1159;
or.b32 %r1225, %r1224, -2139062144;
and.b32 %r1226, %r1159, 2139062143;
and.b32 %r1227, %r1211, -2139062144;
sub.s32 %r1228, %r1225, %r1226;
xor.b32 %r1229, %r1227, %r1228;
xor.b32 %r1162, %r1229, -2139062144;

dp4a.s32.s32 %r1161, %r1162, %r18, %r812;

and.b32 %r1230, %r1220, -2143281136;
or.b32 %r1231, %r1230, -2139062144;
add.s32 %r1232, %r1231, -16843009;
or.b32 %r1165, %r1232, %r1230;

prmt.b32 %r1165, %r1165, 0, 0xba98;

xor.b32 %r1233, %r1165, %r1212;
or.b32 %r1234, %r1233, -2139062144;
and.b32 %r1235, %r1165, 2139062143;
and.b32 %r1236, %r1212, -2139062144;
sub.s32 %r1237, %r1234, %r1235;
xor.b32 %r1238, %r1236, %r1237;
xor.b32 %r1168, %r1238, -2139062144;

dp4a.s32.s32 %r1167, %r1168, %r19, %r1161;

cvt.u32.u16 %r1239, %rs57;
mul.wide.u32 %rd150, %r1239, 8;
add.s64 %rd151, %rd176, %rd150;
ld.global.nc.v2.u32 {%r1240, %r1241}, [%rd151];
shr.u32 %r1244, %r1208, 7;
and.b32 %r1245, %r1244, 255;
popc.b32 %r1246, %r1245;
shl.b32 %r1247, %r1246, 7;
and.b32 %r1248, %r1247, 128;
xor.b32 %r1249, %r1248, %r1245;
mul.lo.s32 %r1250, %r1249, 16843009;
and.b32 %r1251, %r1250, 134480385;
or.b32 %r1252, %r1251, -2139062144;
add.s32 %r1253, %r1252, -16843009;
or.b32 %r1171, %r1253, %r1251;

prmt.b32 %r1171, %r1171, 0, 0xba98;

xor.b32 %r1254, %r1240, %r1171;
or.b32 %r1255, %r1254, -2139062144;
and.b32 %r1256, %r1171, 2139062143;
and.b32 %r1257, %r1240, -2139062144;
sub.s32 %r1258, %r1255, %r1256;
xor.b32 %r1259, %r1257, %r1258;
xor.b32 %r1174, %r1259, -2139062144;

dp4a.s32.s32 %r1173, %r1174, %r20, %r1167;

and.b32 %r1260, %r1250, -2143281136;
or.b32 %r1261, %r1260, -2139062144;
add.s32 %r1262, %r1261, -16843009;
or.b32 %r1177, %r1262, %r1260;

prmt.b32 %r1177, %r1177, 0, 0xba98;

xor.b32 %r1263, %r1177, %r1241;
or.b32 %r1264, %r1263, -2139062144;
and.b32 %r1265, %r1177, 2139062143;
and.b32 %r1266, %r1241, -2139062144;
sub.s32 %r1267, %r1264, %r1265;
xor.b32 %r1268, %r1266, %r1267;
xor.b32 %r1180, %r1268, -2139062144;

dp4a.s32.s32 %r1179, %r1180, %r21, %r1173;

shl.b16 %rs63, %rs58, 3;
cvt.u64.u16 %rd152, %rs63;
and.b64 %rd153, %rd152, 2040;
add.s64 %rd154, %rd176, %rd153;
ld.global.nc.v2.u32 {%r1269, %r1270}, [%rd154];
shr.u32 %r1273, %r1210, 14;
and.b32 %r1274, %r1273, 255;
popc.b32 %r1275, %r1274;
shl.b32 %r1276, %r1275, 7;
and.b32 %r1277, %r1276, 128;
xor.b32 %r1278, %r1277, %r1274;
mul.lo.s32 %r1279, %r1278, 16843009;
and.b32 %r1280, %r1279, 134480385;
or.b32 %r1281, %r1280, -2139062144;
add.s32 %r1282, %r1281, -16843009;
or.b32 %r1183, %r1282, %r1280;

prmt.b32 %r1183, %r1183, 0, 0xba98;

xor.b32 %r1283, %r1269, %r1183;
or.b32 %r1284, %r1283, -2139062144;
and.b32 %r1285, %r1183, 2139062143;
and.b32 %r1286, %r1269, -2139062144;
sub.s32 %r1287, %r1284, %r1285;
xor.b32 %r1288, %r1286, %r1287;
xor.b32 %r1186, %r1288, -2139062144;

dp4a.s32.s32 %r1185, %r1186, %r22, %r1179;

and.b32 %r1289, %r1279, -2143281136;
or.b32 %r1290, %r1289, -2139062144;
add.s32 %r1291, %r1290, -16843009;
or.b32 %r1189, %r1291, %r1289;

prmt.b32 %r1189, %r1189, 0, 0xba98;

xor.b32 %r1292, %r1189, %r1270;
or.b32 %r1293, %r1292, -2139062144;
and.b32 %r1294, %r1189, 2139062143;
and.b32 %r1295, %r1270, -2139062144;
sub.s32 %r1296, %r1293, %r1294;
xor.b32 %r1297, %r1295, %r1296;
xor.b32 %r1192, %r1297, -2139062144;

dp4a.s32.s32 %r1191, %r1192, %r23, %r1185;

cvt.u32.u16 %r1298, %rs59;
mul.wide.u32 %rd155, %r1298, 8;
add.s64 %rd156, %rd176, %rd155;
ld.global.nc.v2.u32 {%r1299, %r1300}, [%rd156];
shr.u32 %r1303, %r1209, 5;
and.b32 %r1304, %r1303, 255;
popc.b32 %r1305, %r1304;
shl.b32 %r1306, %r1305, 7;
and.b32 %r1307, %r1306, 128;
xor.b32 %r1308, %r1307, %r1304;
mul.lo.s32 %r1309, %r1308, 16843009;
and.b32 %r1310, %r1309, 134480385;
or.b32 %r1311, %r1310, -2139062144;
add.s32 %r1312, %r1311, -16843009;
or.b32 %r1195, %r1312, %r1310;

prmt.b32 %r1195, %r1195, 0, 0xba98;

xor.b32 %r1313, %r1299, %r1195;
or.b32 %r1314, %r1313, -2139062144;
and.b32 %r1315, %r1195, 2139062143;
and.b32 %r1316, %r1299, -2139062144;
sub.s32 %r1317, %r1314, %r1315;
xor.b32 %r1318, %r1316, %r1317;
xor.b32 %r1198, %r1318, -2139062144;

dp4a.s32.s32 %r1197, %r1198, %r24, %r1191;

and.b32 %r1319, %r1309, -2143281136;
or.b32 %r1320, %r1319, -2139062144;
add.s32 %r1321, %r1320, -16843009;
or.b32 %r1201, %r1321, %r1319;

prmt.b32 %r1201, %r1201, 0, 0xba98;

xor.b32 %r1322, %r1201, %r1300;
or.b32 %r1323, %r1322, -2139062144;
and.b32 %r1324, %r1201, 2139062143;
and.b32 %r1325, %r1300, -2139062144;
sub.s32 %r1326, %r1323, %r1324;
xor.b32 %r1327, %r1325, %r1326;
xor.b32 %r1204, %r1327, -2139062144;

dp4a.s32.s32 %r1203, %r1204, %r25, %r1197;

shr.u32 %r1328, %r1209, 11;
or.b32 %r1329, %r1328, 1;
mul.lo.s32 %r1330, %r1203, %r1329;
shr.s32 %r1331, %r1330, 31;
shr.u32 %r1332, %r1331, 29;
add.s32 %r1333, %r1330, %r1332;
shr.s32 %r1334, %r1333, 3;
add.s64 %rd157, %rd24, %rd14;
ld.global.nc.u16 %rs55, [%rd157];

{ cvt.f32.f16 %f125, %rs55;}



{.reg .f16 low,high;
mov.b32 {low,high},%r26;
cvt.f32.f16 %f126, low;}


mul.ftz.f32 %f127, %f125, %f126;
cvt.rn.f32.s32 %f128, %r1334;
fma.rn.ftz.f32 %f27, %f127, %f128, %f27;
@%p13 bra $L__BB200_18;

mov.u64 %rd177, iq2xxs_grid;
add.s64 %rd158, %rd184, %rd11;
ld.global.u16 %rs65, [%rd158+-6];
shr.u16 %rs66, %rs65, 8;
ld.global.u16 %rs67, [%rd158+-4];
shr.u16 %rs68, %rs67, 8;
ld.global.u16 %rs69, [%rd158+-2];
cvt.u32.u16 %r1384, %rs69;
ld.global.u16 %rs70, [%rd158];
cvt.u32.u16 %r1385, %rs70;
mov.b32 %r1386, {%rs69, %rs70};
shl.b16 %rs71, %rs65, 3;
cvt.u64.u16 %rd159, %rs71;
and.b64 %rd160, %rd159, 2040;
add.s64 %rd162, %rd177, %rd160;
ld.global.nc.v2.u32 {%r1387, %r1388}, [%rd162];
mov.u32 %r1340, 0;
and.b32 %r1391, %r1384, 255;
popc.b32 %r1392, %r1391;
shl.b32 %r1393, %r1392, 7;
and.b32 %r1394, %r1393, 128;
xor.b32 %r1395, %r1394, %r1391;
mul.lo.s32 %r1396, %r1395, 16843009;
and.b32 %r1397, %r1396, 134480385;
or.b32 %r1398, %r1397, -2139062144;
add.s32 %r1399, %r1398, -16843009;
or.b32 %r1335, %r1399, %r1397;

prmt.b32 %r1335, %r1335, 0, 0xba98;

xor.b32 %r1400, %r1387, %r1335;
or.b32 %r1401, %r1400, -2139062144;
and.b32 %r1402, %r1335, 2139062143;
and.b32 %r1403, %r1387, -2139062144;
sub.s32 %r1404, %r1401, %r1402;
xor.b32 %r1405, %r1403, %r1404;
xor.b32 %r1338, %r1405, -2139062144;

dp4a.s32.s32 %r1337, %r1338, %r18, %r1340;

and.b32 %r1406, %r1396, -2143281136;
or.b32 %r1407, %r1406, -2139062144;
add.s32 %r1408, %r1407, -16843009;
or.b32 %r1341, %r1408, %r1406;

prmt.b32 %r1341, %r1341, 0, 0xba98;

xor.b32 %r1409, %r1341, %r1388;
or.b32 %r1410, %r1409, -2139062144;
and.b32 %r1411, %r1341, 2139062143;
and.b32 %r1412, %r1388, -2139062144;
sub.s32 %r1413, %r1410, %r1411;
xor.b32 %r1414, %r1412, %r1413;
xor.b32 %r1344, %r1414, -2139062144;

dp4a.s32.s32 %r1343, %r1344, %r19, %r1337;

cvt.u32.u16 %r1415, %rs66;
mul.wide.u32 %rd163, %r1415, 8;
add.s64 %rd164, %rd177, %rd163;
ld.global.nc.v2.u32 {%r1416, %r1417}, [%rd164];
shr.u32 %r1420, %r1384, 7;
and.b32 %r1421, %r1420, 255;
popc.b32 %r1422, %r1421;
shl.b32 %r1423, %r1422, 7;
and.b32 %r1424, %r1423, 128;
xor.b32 %r1425, %r1424, %r1421;
mul.lo.s32 %r1426, %r1425, 16843009;
and.b32 %r1427, %r1426, 134480385;
or.b32 %r1428, %r1427, -2139062144;
add.s32 %r1429, %r1428, -16843009;
or.b32 %r1347, %r1429, %r1427;

prmt.b32 %r1347, %r1347, 0, 0xba98;

xor.b32 %r1430, %r1416, %r1347;
or.b32 %r1431, %r1430, -2139062144;
and.b32 %r1432, %r1347, 2139062143;
and.b32 %r1433, %r1416, -2139062144;
sub.s32 %r1434, %r1431, %r1432;
xor.b32 %r1435, %r1433, %r1434;
xor.b32 %r1350, %r1435, -2139062144;

dp4a.s32.s32 %r1349, %r1350, %r20, %r1343;

and.b32 %r1436, %r1426, -2143281136;
or.b32 %r1437, %r1436, -2139062144;
add.s32 %r1438, %r1437, -16843009;
or.b32 %r1353, %r1438, %r1436;

prmt.b32 %r1353, %r1353, 0, 0xba98;

xor.b32 %r1439, %r1353, %r1417;
or.b32 %r1440, %r1439, -2139062144;
and.b32 %r1441, %r1353, 2139062143;
and.b32 %r1442, %r1417, -2139062144;
sub.s32 %r1443, %r1440, %r1441;
xor.b32 %r1444, %r1442, %r1443;
xor.b32 %r1356, %r1444, -2139062144;

dp4a.s32.s32 %r1355, %r1356, %r21, %r1349;

shl.b16 %rs72, %rs67, 3;
cvt.u64.u16 %rd165, %rs72;
and.b64 %rd166, %rd165, 2040;
add.s64 %rd167, %rd177, %rd166;
ld.global.nc.v2.u32 {%r1445, %r1446}, [%rd167];
shr.u32 %r1449, %r1386, 14;
and.b32 %r1450, %r1449, 255;
popc.b32 %r1451, %r1450;
shl.b32 %r1452, %r1451, 7;
and.b32 %r1453, %r1452, 128;
xor.b32 %r1454, %r1453, %r1450;
mul.lo.s32 %r1455, %r1454, 16843009;
and.b32 %r1456, %r1455, 134480385;
or.b32 %r1457, %r1456, -2139062144;
add.s32 %r1458, %r1457, -16843009;
or.b32 %r1359, %r1458, %r1456;

prmt.b32 %r1359, %r1359, 0, 0xba98;

xor.b32 %r1459, %r1445, %r1359;
or.b32 %r1460, %r1459, -2139062144;
and.b32 %r1461, %r1359, 2139062143;
and.b32 %r1462, %r1445, -2139062144;
sub.s32 %r1463, %r1460, %r1461;
xor.b32 %r1464, %r1462, %r1463;
xor.b32 %r1362, %r1464, -2139062144;

dp4a.s32.s32 %r1361, %r1362, %r22, %r1355;

and.b32 %r1465, %r1455, -2143281136;
or.b32 %r1466, %r1465, -2139062144;
add.s32 %r1467, %r1466, -16843009;
or.b32 %r1365, %r1467, %r1465;

prmt.b32 %r1365, %r1365, 0, 0xba98;

xor.b32 %r1468, %r1365, %r1446;
or.b32 %r1469, %r1468, -2139062144;
and.b32 %r1470, %r1365, 2139062143;
and.b32 %r1471, %r1446, -2139062144;
sub.s32 %r1472, %r1469, %r1470;
xor.b32 %r1473, %r1471, %r1472;
xor.b32 %r1368, %r1473, -2139062144;

dp4a.s32.s32 %r1367, %r1368, %r23, %r1361;

cvt.u32.u16 %r1474, %rs68;
mul.wide.u32 %rd168, %r1474, 8;
add.s64 %rd169, %rd177, %rd168;
ld.global.nc.v2.u32 {%r1475, %r1476}, [%rd169];
shr.u32 %r1479, %r1385, 5;
and.b32 %r1480, %r1479, 255;
popc.b32 %r1481, %r1480;
shl.b32 %r1482, %r1481, 7;
and.b32 %r1483, %r1482, 128;
xor.b32 %r1484, %r1483, %r1480;
mul.lo.s32 %r1485, %r1484, 16843009;
and.b32 %r1486, %r1485, 134480385;
or.b32 %r1487, %r1486, -2139062144;
add.s32 %r1488, %r1487, -16843009;
or.b32 %r1371, %r1488, %r1486;

prmt.b32 %r1371, %r1371, 0, 0xba98;

xor.b32 %r1489, %r1475, %r1371;
or.b32 %r1490, %r1489, -2139062144;
and.b32 %r1491, %r1371, 2139062143;
and.b32 %r1492, %r1475, -2139062144;
sub.s32 %r1493, %r1490, %r1491;
xor.b32 %r1494, %r1492, %r1493;
xor.b32 %r1374, %r1494, -2139062144;

dp4a.s32.s32 %r1373, %r1374, %r24, %r1367;

and.b32 %r1495, %r1485, -2143281136;
or.b32 %r1496, %r1495, -2139062144;
add.s32 %r1497, %r1496, -16843009;
or.b32 %r1377, %r1497, %r1495;

prmt.b32 %r1377, %r1377, 0, 0xba98;

xor.b32 %r1498, %r1377, %r1476;
or.b32 %r1499, %r1498, -2139062144;
and.b32 %r1500, %r1377, 2139062143;
and.b32 %r1501, %r1476, -2139062144;
sub.s32 %r1502, %r1499, %r1500;
xor.b32 %r1503, %r1501, %r1502;
xor.b32 %r1380, %r1503, -2139062144;

dp4a.s32.s32 %r1379, %r1380, %r25, %r1373;

shr.u32 %r1504, %r1385, 11;
or.b32 %r1505, %r1504, 1;
mul.lo.s32 %r1506, %r1379, %r1505;
shr.s32 %r1507, %r1506, 31;
shr.u32 %r1508, %r1507, 29;
add.s32 %r1509, %r1506, %r1508;
shr.s32 %r1510, %r1509, 3;
add.s64 %rd170, %rd184, %rd10;
ld.global.u16 %rs64, [%rd170];

{ cvt.f32.f16 %f129, %rs64;}



{.reg .f16 low,high;
mov.b32 {low,high},%r26;
cvt.f32.f16 %f130, low;}


mul.ftz.f32 %f131, %f129, %f130;
cvt.rn.f32.s32 %f132, %r1510;
fma.rn.ftz.f32 %f31, %f131, %f132, %f31;

$L__BB200_18:
add.s32 %r2207, %r2207, 128;
add.s64 %rd184, %rd184, 1056;
add.s64 %rd183, %rd183, 1056;
add.s32 %r2208, %r2208, 16;
setp.lt.s32 %p17, %r2208, %r83;
@%p17 bra $L__BB200_10;

$L__BB200_19:
mov.u32 %r2194, %tid.y;
setp.eq.s32 %p18, %r2194, 0;
@%p18 bra $L__BB200_23;

mov.u32 %r2195, %tid.y;
setp.ne.s64 %p19, %rd32, 0;
add.s32 %r29, %r2195, -1;
shl.b32 %r1515, %r2195, 9;
mov.u32 %r1516, _ZZ13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r1517, %r1516, %r1515;
shl.b32 %r1518, %r2, 2;
add.s32 %r1519, %r1517, %r1518;
st.shared.f32 [%r1519+-512], %f30;
@%p19 bra $L__BB200_22;
bra.uni $L__BB200_21;

$L__BB200_22:
shl.b32 %r1521, %r29, 9;
mov.u32 %r1522, _ZZ13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r1523, %r1522, %r1521;
add.s32 %r1525, %r1523, %r1518;
st.shared.f32 [%r1525], %f34;
add.s32 %r2191, %r1519, -512;
st.shared.f32 [%r2191+128], %f29;
st.shared.f32 [%r1525+128], %f33;
add.s32 %r2192, %r1519, -512;
st.shared.f32 [%r2192+256], %f28;
st.shared.f32 [%r1525+256], %f32;
add.s32 %r2193, %r1519, -512;
st.shared.f32 [%r2193+384], %f27;
st.shared.f32 [%r1525+384], %f31;
bra.uni $L__BB200_23;

$L__BB200_21:
add.s32 %r2188, %r1519, -512;
st.shared.f32 [%r2188+128], %f29;
add.s32 %r2189, %r1519, -512;
st.shared.f32 [%r2189+256], %f28;
add.s32 %r2190, %r1519, -512;
st.shared.f32 [%r2190+384], %f27;

$L__BB200_23:
mov.u32 %r2196, %tid.y;
setp.ne.s32 %p319, %r2196, 0;
bar.sync 0;
@%p319 bra $L__BB200_78;
bra.uni $L__BB200_24;

$L__BB200_78:
ret;

$L__BB200_24:
ld.param.u64 %rd175, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.f32 %f977, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
ld.param.u32 %r2199, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r2198, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
mov.u32 %r2197, %ctaid.z;
setp.ne.s64 %p21, %rd32, 0;
mov.u32 %r1528, %ctaid.y;
mad.lo.s32 %r1529, %r1528, %r2199, %r4;
mad.lo.s32 %r1530, %r2197, %r2198, %r1529;
neg.ftz.f32 %f35, %f977;
shl.b32 %r1532, %r2, 2;
mov.u32 %r1533, _ZZ13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r31, %r1533, %r1532;
cvta.to.global.u64 %rd171, %rd175;
mul.wide.u32 %rd172, %r1530, 4;
add.s64 %rd27, %rd171, %rd172;
@%p21 bra $L__BB200_33;
bra.uni $L__BB200_25;

$L__BB200_33:
ld.param.u32 %r2200, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
mov.u32 %r1612, _ZZ13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r33, %r1612, %r1532;
setp.gt.s32 %p54, %r2200, 2;
@%p54 bra $L__BB200_44;

ld.param.u32 %r2203, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p57, %r2203, 1;
@%p57 bra $L__BB200_62;

ld.param.u32 %r2204, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p58, %r2204, 2;
@%p58 bra $L__BB200_36;
bra.uni $L__BB200_70;

$L__BB200_36:
ld.shared.f32 %f533, [%r31];
add.ftz.f32 %f534, %f533, %f30;
ld.shared.f32 %f535, [%r33];
add.ftz.f32 %f536, %f535, %f34;
ld.shared.f32 %f537, [%r31+512];
add.ftz.f32 %f538, %f537, %f534;
ld.shared.f32 %f539, [%r33+512];
add.ftz.f32 %f540, %f539, %f536;
ld.shared.f32 %f541, [%r31+1024];
add.ftz.f32 %f542, %f541, %f538;
ld.shared.f32 %f543, [%r33+1024];
add.ftz.f32 %f544, %f543, %f540;
mov.b32 %r1843, %f542;
mov.u32 %r1844, 31;
mov.u32 %r1845, 16;
mov.u32 %r1846, -1;
shfl.sync.bfly.b32 %r1847|%p163, %r1843, %r1845, %r1844, %r1846;
mov.b32 %f545, %r1847;
add.ftz.f32 %f546, %f542, %f545;
mov.b32 %r1848, %f546;
mov.u32 %r1849, 8;
shfl.sync.bfly.b32 %r1850|%p164, %r1848, %r1849, %r1844, %r1846;
mov.b32 %f547, %r1850;
add.ftz.f32 %f548, %f546, %f547;
mov.b32 %r1851, %f548;
mov.u32 %r1852, 4;
shfl.sync.bfly.b32 %r1853|%p165, %r1851, %r1852, %r1844, %r1846;
mov.b32 %f549, %r1853;
add.ftz.f32 %f550, %f548, %f549;
mov.b32 %r1854, %f550;
mov.u32 %r1855, 2;
shfl.sync.bfly.b32 %r1856|%p166, %r1854, %r1855, %r1844, %r1846;
mov.b32 %f551, %r1856;
add.ftz.f32 %f552, %f550, %f551;
mov.b32 %r1857, %f552;
mov.u32 %r1858, 1;
shfl.sync.bfly.b32 %r1859|%p167, %r1857, %r1858, %r1844, %r1846;
mov.b32 %f553, %r1859;
add.ftz.f32 %f56, %f552, %f553;
mov.b32 %r1860, %f544;
shfl.sync.bfly.b32 %r1861|%p168, %r1860, %r1845, %r1844, %r1846;
mov.b32 %f554, %r1861;
add.ftz.f32 %f555, %f544, %f554;
mov.b32 %r1862, %f555;
shfl.sync.bfly.b32 %r1863|%p169, %r1862, %r1849, %r1844, %r1846;
mov.b32 %f556, %r1863;
add.ftz.f32 %f557, %f555, %f556;
mov.b32 %r1864, %f557;
shfl.sync.bfly.b32 %r1865|%p170, %r1864, %r1852, %r1844, %r1846;
mov.b32 %f558, %r1865;
add.ftz.f32 %f559, %f557, %f558;
mov.b32 %r1866, %f559;
shfl.sync.bfly.b32 %r1867|%p171, %r1866, %r1855, %r1844, %r1846;
mov.b32 %f560, %r1867;
add.ftz.f32 %f561, %f559, %f560;
mov.b32 %r1868, %f561;
shfl.sync.bfly.b32 %r1869|%p172, %r1868, %r1858, %r1844, %r1846;
mov.b32 %f562, %r1869;
add.ftz.f32 %f57, %f561, %f562;
setp.ne.s32 %p173, %r2, 0;
setp.ge.u32 %p174, %r4, %r50;
or.pred %p175, %p173, %p174;
@%p175 bra $L__BB200_38;

add.ftz.f32 %f563, %f5, %f56;
add.ftz.f32 %f564, %f4, %f57;
mul.ftz.f32 %f565, %f564, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f566, %f565;
add.ftz.f32 %f567, %f566, 0f3F800000;
div.approx.ftz.f32 %f568, %f564, %f567;
mul.ftz.f32 %f569, %f563, %f568;
st.global.f32 [%rd27], %f569;

$L__BB200_38:
ld.shared.f32 %f570, [%r31+128];
add.ftz.f32 %f571, %f570, %f29;
ld.shared.f32 %f572, [%r33+128];
add.ftz.f32 %f573, %f572, %f33;
ld.shared.f32 %f574, [%r31+640];
add.ftz.f32 %f575, %f574, %f571;
ld.shared.f32 %f576, [%r33+640];
add.ftz.f32 %f577, %f576, %f573;
ld.shared.f32 %f578, [%r31+1152];
add.ftz.f32 %f579, %f578, %f575;
ld.shared.f32 %f580, [%r33+1152];
add.ftz.f32 %f581, %f580, %f577;
mov.b32 %r1871, %f579;
shfl.sync.bfly.b32 %r1875|%p176, %r1871, %r1845, %r1844, %r1846;
mov.b32 %f582, %r1875;
add.ftz.f32 %f583, %f579, %f582;
mov.b32 %r1876, %f583;
shfl.sync.bfly.b32 %r1878|%p177, %r1876, %r1849, %r1844, %r1846;
mov.b32 %f584, %r1878;
add.ftz.f32 %f585, %f583, %f584;
mov.b32 %r1879, %f585;
shfl.sync.bfly.b32 %r1881|%p178, %r1879, %r1852, %r1844, %r1846;
mov.b32 %f586, %r1881;
add.ftz.f32 %f587, %f585, %f586;
mov.b32 %r1882, %f587;
shfl.sync.bfly.b32 %r1884|%p179, %r1882, %r1855, %r1844, %r1846;
mov.b32 %f588, %r1884;
add.ftz.f32 %f589, %f587, %f588;
mov.b32 %r1885, %f589;
shfl.sync.bfly.b32 %r1887|%p180, %r1885, %r1858, %r1844, %r1846;
mov.b32 %f590, %r1887;
add.ftz.f32 %f58, %f589, %f590;
mov.b32 %r1888, %f581;
shfl.sync.bfly.b32 %r1889|%p181, %r1888, %r1845, %r1844, %r1846;
mov.b32 %f591, %r1889;
add.ftz.f32 %f592, %f581, %f591;
mov.b32 %r1890, %f592;
shfl.sync.bfly.b32 %r1891|%p182, %r1890, %r1849, %r1844, %r1846;
mov.b32 %f593, %r1891;
add.ftz.f32 %f594, %f592, %f593;
mov.b32 %r1892, %f594;
shfl.sync.bfly.b32 %r1893|%p183, %r1892, %r1852, %r1844, %r1846;
mov.b32 %f595, %r1893;
add.ftz.f32 %f596, %f594, %f595;
mov.b32 %r1894, %f596;
shfl.sync.bfly.b32 %r1895|%p184, %r1894, %r1855, %r1844, %r1846;
mov.b32 %f597, %r1895;
add.ftz.f32 %f598, %f596, %f597;
mov.b32 %r1896, %f598;
shfl.sync.bfly.b32 %r1897|%p185, %r1896, %r1858, %r1844, %r1846;
mov.b32 %f599, %r1897;
add.ftz.f32 %f59, %f598, %f599;
setp.ne.s32 %p186, %r2, 1;
add.s32 %r1899, %r4, 1;
setp.ge.u32 %p187, %r1899, %r50;
or.pred %p188, %p186, %p187;
@%p188 bra $L__BB200_40;

add.ftz.f32 %f600, %f5, %f58;
add.ftz.f32 %f601, %f4, %f59;
mul.ftz.f32 %f602, %f601, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f603, %f602;
add.ftz.f32 %f604, %f603, 0f3F800000;
div.approx.ftz.f32 %f605, %f601, %f604;
mul.ftz.f32 %f606, %f600, %f605;
st.global.f32 [%rd27+4], %f606;

$L__BB200_40:
ld.shared.f32 %f607, [%r31+256];
add.ftz.f32 %f608, %f607, %f28;
ld.shared.f32 %f609, [%r33+256];
add.ftz.f32 %f610, %f609, %f32;
ld.shared.f32 %f611, [%r31+768];
add.ftz.f32 %f612, %f611, %f608;
ld.shared.f32 %f613, [%r33+768];
add.ftz.f32 %f614, %f613, %f610;
ld.shared.f32 %f615, [%r31+1280];
add.ftz.f32 %f616, %f615, %f612;
ld.shared.f32 %f617, [%r33+1280];
add.ftz.f32 %f618, %f617, %f614;
mov.b32 %r1900, %f616;
mov.u32 %r1901, 31;
mov.u32 %r1902, 16;
mov.u32 %r1903, -1;
shfl.sync.bfly.b32 %r1904|%p189, %r1900, %r1902, %r1901, %r1903;
mov.b32 %f619, %r1904;
add.ftz.f32 %f620, %f616, %f619;
mov.b32 %r1905, %f620;
mov.u32 %r1906, 8;
shfl.sync.bfly.b32 %r1907|%p190, %r1905, %r1906, %r1901, %r1903;
mov.b32 %f621, %r1907;
add.ftz.f32 %f622, %f620, %f621;
mov.b32 %r1908, %f622;
mov.u32 %r1909, 4;
shfl.sync.bfly.b32 %r1910|%p191, %r1908, %r1909, %r1901, %r1903;
mov.b32 %f623, %r1910;
add.ftz.f32 %f624, %f622, %f623;
mov.b32 %r1911, %f624;
mov.u32 %r1912, 2;
shfl.sync.bfly.b32 %r1913|%p192, %r1911, %r1912, %r1901, %r1903;
mov.b32 %f625, %r1913;
add.ftz.f32 %f626, %f624, %f625;
mov.b32 %r1914, %f626;
mov.u32 %r1915, 1;
shfl.sync.bfly.b32 %r1916|%p193, %r1914, %r1915, %r1901, %r1903;
mov.b32 %f627, %r1916;
add.ftz.f32 %f60, %f626, %f627;
mov.b32 %r1917, %f618;
shfl.sync.bfly.b32 %r1918|%p194, %r1917, %r1902, %r1901, %r1903;
mov.b32 %f628, %r1918;
add.ftz.f32 %f629, %f618, %f628;
mov.b32 %r1919, %f629;
shfl.sync.bfly.b32 %r1920|%p195, %r1919, %r1906, %r1901, %r1903;
mov.b32 %f630, %r1920;
add.ftz.f32 %f631, %f629, %f630;
mov.b32 %r1921, %f631;
shfl.sync.bfly.b32 %r1922|%p196, %r1921, %r1909, %r1901, %r1903;
mov.b32 %f632, %r1922;
add.ftz.f32 %f633, %f631, %f632;
mov.b32 %r1923, %f633;
shfl.sync.bfly.b32 %r1924|%p197, %r1923, %r1912, %r1901, %r1903;
mov.b32 %f634, %r1924;
add.ftz.f32 %f635, %f633, %f634;
mov.b32 %r1925, %f635;
shfl.sync.bfly.b32 %r1926|%p198, %r1925, %r1915, %r1901, %r1903;
mov.b32 %f636, %r1926;
add.ftz.f32 %f61, %f635, %f636;
setp.ne.s32 %p199, %r2, 2;
add.s32 %r1928, %r4, 2;
setp.ge.u32 %p200, %r1928, %r50;
or.pred %p201, %p199, %p200;
@%p201 bra $L__BB200_42;

add.ftz.f32 %f637, %f5, %f60;
add.ftz.f32 %f638, %f4, %f61;
mul.ftz.f32 %f639, %f638, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f640, %f639;
add.ftz.f32 %f641, %f640, 0f3F800000;
div.approx.ftz.f32 %f642, %f638, %f641;
mul.ftz.f32 %f643, %f637, %f642;
st.global.f32 [%rd27+8], %f643;

$L__BB200_42:
ld.shared.f32 %f644, [%r31+384];
add.ftz.f32 %f645, %f644, %f27;
ld.shared.f32 %f646, [%r33+384];
add.ftz.f32 %f647, %f646, %f31;
ld.shared.f32 %f648, [%r31+896];
add.ftz.f32 %f649, %f648, %f645;
ld.shared.f32 %f650, [%r33+896];
add.ftz.f32 %f651, %f650, %f647;
ld.shared.f32 %f652, [%r31+1408];
add.ftz.f32 %f653, %f652, %f649;
ld.shared.f32 %f654, [%r33+1408];
add.ftz.f32 %f655, %f654, %f651;
mov.b32 %r1929, %f653;
shfl.sync.bfly.b32 %r1933|%p202, %r1929, %r1902, %r1901, %r1903;
mov.b32 %f656, %r1933;
add.ftz.f32 %f657, %f653, %f656;
mov.b32 %r1934, %f657;
shfl.sync.bfly.b32 %r1936|%p203, %r1934, %r1906, %r1901, %r1903;
mov.b32 %f658, %r1936;
add.ftz.f32 %f659, %f657, %f658;
mov.b32 %r1937, %f659;
shfl.sync.bfly.b32 %r1939|%p204, %r1937, %r1909, %r1901, %r1903;
mov.b32 %f660, %r1939;
add.ftz.f32 %f661, %f659, %f660;
mov.b32 %r1940, %f661;
shfl.sync.bfly.b32 %r1942|%p205, %r1940, %r1912, %r1901, %r1903;
mov.b32 %f662, %r1942;
add.ftz.f32 %f663, %f661, %f662;
mov.b32 %r1943, %f663;
shfl.sync.bfly.b32 %r1945|%p206, %r1943, %r1915, %r1901, %r1903;
mov.b32 %f664, %r1945;
add.ftz.f32 %f62, %f663, %f664;
mov.b32 %r1946, %f655;
shfl.sync.bfly.b32 %r1947|%p207, %r1946, %r1902, %r1901, %r1903;
mov.b32 %f665, %r1947;
add.ftz.f32 %f666, %f655, %f665;
mov.b32 %r1948, %f666;
shfl.sync.bfly.b32 %r1949|%p208, %r1948, %r1906, %r1901, %r1903;
mov.b32 %f667, %r1949;
add.ftz.f32 %f668, %f666, %f667;
mov.b32 %r1950, %f668;
shfl.sync.bfly.b32 %r1951|%p209, %r1950, %r1909, %r1901, %r1903;
mov.b32 %f669, %r1951;
add.ftz.f32 %f670, %f668, %f669;
mov.b32 %r1952, %f670;
shfl.sync.bfly.b32 %r1953|%p210, %r1952, %r1912, %r1901, %r1903;
mov.b32 %f671, %r1953;
add.ftz.f32 %f672, %f670, %f671;
mov.b32 %r1954, %f672;
shfl.sync.bfly.b32 %r1955|%p211, %r1954, %r1915, %r1901, %r1903;
mov.b32 %f673, %r1955;
add.ftz.f32 %f63, %f672, %f673;
setp.ne.s32 %p212, %r2, 3;
add.s32 %r1957, %r4, 3;
setp.ge.u32 %p213, %r1957, %r50;
or.pred %p214, %p212, %p213;
@%p214 bra $L__BB200_78;

add.ftz.f32 %f674, %f5, %f62;
add.ftz.f32 %f675, %f4, %f63;
mul.ftz.f32 %f676, %f675, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f677, %f676;
add.ftz.f32 %f678, %f677, 0f3F800000;
div.approx.ftz.f32 %f679, %f675, %f678;
mul.ftz.f32 %f680, %f674, %f679;
st.global.f32 [%rd27+12], %f680;
bra.uni $L__BB200_78;

$L__BB200_25:
ld.shared.f32 %f133, [%r31];
add.ftz.f32 %f134, %f133, %f30;
ld.shared.f32 %f135, [%r31+512];
add.ftz.f32 %f136, %f135, %f134;
ld.shared.f32 %f137, [%r31+1024];
add.ftz.f32 %f138, %f137, %f136;
mov.b32 %r1534, %f138;
mov.u32 %r1535, 31;
mov.u32 %r1536, 16;
mov.u32 %r1537, -1;
shfl.sync.bfly.b32 %r1538|%p22, %r1534, %r1536, %r1535, %r1537;
mov.b32 %f139, %r1538;
add.ftz.f32 %f140, %f138, %f139;
mov.b32 %r1539, %f140;
mov.u32 %r1540, 8;
shfl.sync.bfly.b32 %r1541|%p23, %r1539, %r1540, %r1535, %r1537;
mov.b32 %f141, %r1541;
add.ftz.f32 %f142, %f140, %f141;
mov.b32 %r1542, %f142;
mov.u32 %r1543, 4;
shfl.sync.bfly.b32 %r1544|%p24, %r1542, %r1543, %r1535, %r1537;
mov.b32 %f143, %r1544;
add.ftz.f32 %f144, %f142, %f143;
mov.b32 %r1545, %f144;
mov.u32 %r1546, 2;
shfl.sync.bfly.b32 %r1547|%p25, %r1545, %r1546, %r1535, %r1537;
mov.b32 %f145, %r1547;
add.ftz.f32 %f146, %f144, %f145;
mov.b32 %r1548, %f146;
mov.u32 %r1549, 1;
shfl.sync.bfly.b32 %r1550|%p26, %r1548, %r1549, %r1535, %r1537;
mov.b32 %f147, %r1550;
add.ftz.f32 %f36, %f146, %f147;
setp.ne.s32 %p27, %r2, 0;
setp.ge.u32 %p28, %r4, %r50;
or.pred %p29, %p27, %p28;
@%p29 bra $L__BB200_27;

add.ftz.f32 %f148, %f5, %f36;
st.global.f32 [%rd27], %f148;

$L__BB200_27:
ld.shared.f32 %f149, [%r31+128];
add.ftz.f32 %f150, %f149, %f29;
ld.shared.f32 %f151, [%r31+640];
add.ftz.f32 %f152, %f151, %f150;
ld.shared.f32 %f153, [%r31+1152];
add.ftz.f32 %f154, %f153, %f152;
mov.b32 %r1552, %f154;
shfl.sync.bfly.b32 %r1556|%p30, %r1552, %r1536, %r1535, %r1537;
mov.b32 %f155, %r1556;
add.ftz.f32 %f156, %f154, %f155;
mov.b32 %r1557, %f156;
shfl.sync.bfly.b32 %r1559|%p31, %r1557, %r1540, %r1535, %r1537;
mov.b32 %f157, %r1559;
add.ftz.f32 %f158, %f156, %f157;
mov.b32 %r1560, %f158;
shfl.sync.bfly.b32 %r1562|%p32, %r1560, %r1543, %r1535, %r1537;
mov.b32 %f159, %r1562;
add.ftz.f32 %f160, %f158, %f159;
mov.b32 %r1563, %f160;
shfl.sync.bfly.b32 %r1565|%p33, %r1563, %r1546, %r1535, %r1537;
mov.b32 %f161, %r1565;
add.ftz.f32 %f162, %f160, %f161;
mov.b32 %r1566, %f162;
shfl.sync.bfly.b32 %r1568|%p34, %r1566, %r1549, %r1535, %r1537;
mov.b32 %f163, %r1568;
add.ftz.f32 %f37, %f162, %f163;
setp.ne.s32 %p35, %r2, 1;
add.s32 %r1570, %r4, 1;
setp.ge.u32 %p36, %r1570, %r50;
or.pred %p37, %p35, %p36;
@%p37 bra $L__BB200_29;

add.ftz.f32 %f164, %f5, %f37;
st.global.f32 [%rd27+4], %f164;

$L__BB200_29:
ld.shared.f32 %f165, [%r31+256];
add.ftz.f32 %f166, %f165, %f28;
ld.shared.f32 %f167, [%r31+768];
add.ftz.f32 %f168, %f167, %f166;
ld.shared.f32 %f169, [%r31+1280];
add.ftz.f32 %f170, %f169, %f168;
mov.b32 %r1571, %f170;
mov.u32 %r1572, 31;
mov.u32 %r1573, 16;
mov.u32 %r1574, -1;
shfl.sync.bfly.b32 %r1575|%p38, %r1571, %r1573, %r1572, %r1574;
mov.b32 %f171, %r1575;
add.ftz.f32 %f172, %f170, %f171;
mov.b32 %r1576, %f172;
mov.u32 %r1577, 8;
shfl.sync.bfly.b32 %r1578|%p39, %r1576, %r1577, %r1572, %r1574;
mov.b32 %f173, %r1578;
add.ftz.f32 %f174, %f172, %f173;
mov.b32 %r1579, %f174;
mov.u32 %r1580, 4;
shfl.sync.bfly.b32 %r1581|%p40, %r1579, %r1580, %r1572, %r1574;
mov.b32 %f175, %r1581;
add.ftz.f32 %f176, %f174, %f175;
mov.b32 %r1582, %f176;
mov.u32 %r1583, 2;
shfl.sync.bfly.b32 %r1584|%p41, %r1582, %r1583, %r1572, %r1574;
mov.b32 %f177, %r1584;
add.ftz.f32 %f178, %f176, %f177;
mov.b32 %r1585, %f178;
mov.u32 %r1586, 1;
shfl.sync.bfly.b32 %r1587|%p42, %r1585, %r1586, %r1572, %r1574;
mov.b32 %f179, %r1587;
add.ftz.f32 %f38, %f178, %f179;
setp.ne.s32 %p43, %r2, 2;
add.s32 %r1589, %r4, 2;
setp.ge.u32 %p44, %r1589, %r50;
or.pred %p45, %p43, %p44;
@%p45 bra $L__BB200_31;

add.ftz.f32 %f180, %f5, %f38;
st.global.f32 [%rd27+8], %f180;

$L__BB200_31:
ld.shared.f32 %f181, [%r31+384];
add.ftz.f32 %f182, %f181, %f27;
ld.shared.f32 %f183, [%r31+896];
add.ftz.f32 %f184, %f183, %f182;
ld.shared.f32 %f185, [%r31+1408];
add.ftz.f32 %f186, %f185, %f184;
mov.b32 %r1590, %f186;
shfl.sync.bfly.b32 %r1594|%p46, %r1590, %r1573, %r1572, %r1574;
mov.b32 %f187, %r1594;
add.ftz.f32 %f188, %f186, %f187;
mov.b32 %r1595, %f188;
shfl.sync.bfly.b32 %r1597|%p47, %r1595, %r1577, %r1572, %r1574;
mov.b32 %f189, %r1597;
add.ftz.f32 %f190, %f188, %f189;
mov.b32 %r1598, %f190;
shfl.sync.bfly.b32 %r1600|%p48, %r1598, %r1580, %r1572, %r1574;
mov.b32 %f191, %r1600;
add.ftz.f32 %f192, %f190, %f191;
mov.b32 %r1601, %f192;
shfl.sync.bfly.b32 %r1603|%p49, %r1601, %r1583, %r1572, %r1574;
mov.b32 %f193, %r1603;
add.ftz.f32 %f194, %f192, %f193;
mov.b32 %r1604, %f194;
shfl.sync.bfly.b32 %r1606|%p50, %r1604, %r1586, %r1572, %r1574;
mov.b32 %f195, %r1606;
add.ftz.f32 %f39, %f194, %f195;
setp.ne.s32 %p51, %r2, 3;
add.s32 %r1608, %r4, 3;
setp.ge.u32 %p52, %r1608, %r50;
or.pred %p53, %p51, %p52;
@%p53 bra $L__BB200_78;

add.ftz.f32 %f196, %f5, %f39;
st.global.f32 [%rd27+12], %f196;
bra.uni $L__BB200_78;

$L__BB200_44:
ld.param.u32 %r2201, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p55, %r2201, 3;
@%p55 bra $L__BB200_54;

ld.param.u32 %r2202, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.ne.s32 %p56, %r2202, 6;
@%p56 bra $L__BB200_70;

ld.shared.f32 %f197, [%r31];
add.ftz.f32 %f198, %f197, %f30;
ld.shared.f32 %f199, [%r33];
add.ftz.f32 %f200, %f199, %f34;
ld.shared.f32 %f201, [%r31+512];
add.ftz.f32 %f202, %f201, %f198;
ld.shared.f32 %f203, [%r33+512];
add.ftz.f32 %f204, %f203, %f200;
ld.shared.f32 %f205, [%r31+1024];
add.ftz.f32 %f206, %f205, %f202;
ld.shared.f32 %f207, [%r33+1024];
add.ftz.f32 %f208, %f207, %f204;
mov.b32 %r1613, %f206;
mov.u32 %r1614, 31;
mov.u32 %r1615, 16;
mov.u32 %r1616, -1;
shfl.sync.bfly.b32 %r1617|%p59, %r1613, %r1615, %r1614, %r1616;
mov.b32 %f209, %r1617;
add.ftz.f32 %f210, %f206, %f209;
mov.b32 %r1618, %f210;
mov.u32 %r1619, 8;
shfl.sync.bfly.b32 %r1620|%p60, %r1618, %r1619, %r1614, %r1616;
mov.b32 %f211, %r1620;
add.ftz.f32 %f212, %f210, %f211;
mov.b32 %r1621, %f212;
mov.u32 %r1622, 4;
shfl.sync.bfly.b32 %r1623|%p61, %r1621, %r1622, %r1614, %r1616;
mov.b32 %f213, %r1623;
add.ftz.f32 %f214, %f212, %f213;
mov.b32 %r1624, %f214;
mov.u32 %r1625, 2;
shfl.sync.bfly.b32 %r1626|%p62, %r1624, %r1625, %r1614, %r1616;
mov.b32 %f215, %r1626;
add.ftz.f32 %f216, %f214, %f215;
mov.b32 %r1627, %f216;
mov.u32 %r1628, 1;
shfl.sync.bfly.b32 %r1629|%p63, %r1627, %r1628, %r1614, %r1616;
mov.b32 %f217, %r1629;
add.ftz.f32 %f40, %f216, %f217;
mov.b32 %r1630, %f208;
shfl.sync.bfly.b32 %r1631|%p64, %r1630, %r1615, %r1614, %r1616;
mov.b32 %f218, %r1631;
add.ftz.f32 %f219, %f208, %f218;
mov.b32 %r1632, %f219;
shfl.sync.bfly.b32 %r1633|%p65, %r1632, %r1619, %r1614, %r1616;
mov.b32 %f220, %r1633;
add.ftz.f32 %f221, %f219, %f220;
mov.b32 %r1634, %f221;
shfl.sync.bfly.b32 %r1635|%p66, %r1634, %r1622, %r1614, %r1616;
mov.b32 %f222, %r1635;
add.ftz.f32 %f223, %f221, %f222;
mov.b32 %r1636, %f223;
shfl.sync.bfly.b32 %r1637|%p67, %r1636, %r1625, %r1614, %r1616;
mov.b32 %f224, %r1637;
add.ftz.f32 %f225, %f223, %f224;
mov.b32 %r1638, %f225;
shfl.sync.bfly.b32 %r1639|%p68, %r1638, %r1628, %r1614, %r1616;
mov.b32 %f226, %r1639;
add.ftz.f32 %f41, %f225, %f226;
setp.ne.s32 %p69, %r2, 0;
setp.ge.u32 %p70, %r4, %r50;
or.pred %p71, %p69, %p70;
@%p71 bra $L__BB200_48;

ld.param.f32 %f981, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f227, %f5, %f40;
add.ftz.f32 %f228, %f4, %f41;
min.ftz.f32 %f229, %f228, %f981;
min.ftz.f32 %f230, %f227, %f981;
max.ftz.f32 %f231, %f230, %f35;
mul.ftz.f32 %f232, %f229, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f233, %f232;
add.ftz.f32 %f234, %f233, 0f3F800000;
div.approx.ftz.f32 %f235, %f229, %f234;
mul.ftz.f32 %f236, %f231, %f235;
st.global.f32 [%rd27], %f236;

$L__BB200_48:
ld.shared.f32 %f237, [%r31+128];
add.ftz.f32 %f238, %f237, %f29;
ld.shared.f32 %f239, [%r33+128];
add.ftz.f32 %f240, %f239, %f33;
ld.shared.f32 %f241, [%r31+640];
add.ftz.f32 %f242, %f241, %f238;
ld.shared.f32 %f243, [%r33+640];
add.ftz.f32 %f244, %f243, %f240;
ld.shared.f32 %f245, [%r31+1152];
add.ftz.f32 %f246, %f245, %f242;
ld.shared.f32 %f247, [%r33+1152];
add.ftz.f32 %f248, %f247, %f244;
mov.b32 %r1641, %f246;
shfl.sync.bfly.b32 %r1645|%p72, %r1641, %r1615, %r1614, %r1616;
mov.b32 %f249, %r1645;
add.ftz.f32 %f250, %f246, %f249;
mov.b32 %r1646, %f250;
shfl.sync.bfly.b32 %r1648|%p73, %r1646, %r1619, %r1614, %r1616;
mov.b32 %f251, %r1648;
add.ftz.f32 %f252, %f250, %f251;
mov.b32 %r1649, %f252;
shfl.sync.bfly.b32 %r1651|%p74, %r1649, %r1622, %r1614, %r1616;
mov.b32 %f253, %r1651;
add.ftz.f32 %f254, %f252, %f253;
mov.b32 %r1652, %f254;
shfl.sync.bfly.b32 %r1654|%p75, %r1652, %r1625, %r1614, %r1616;
mov.b32 %f255, %r1654;
add.ftz.f32 %f256, %f254, %f255;
mov.b32 %r1655, %f256;
shfl.sync.bfly.b32 %r1657|%p76, %r1655, %r1628, %r1614, %r1616;
mov.b32 %f257, %r1657;
add.ftz.f32 %f42, %f256, %f257;
mov.b32 %r1658, %f248;
shfl.sync.bfly.b32 %r1659|%p77, %r1658, %r1615, %r1614, %r1616;
mov.b32 %f258, %r1659;
add.ftz.f32 %f259, %f248, %f258;
mov.b32 %r1660, %f259;
shfl.sync.bfly.b32 %r1661|%p78, %r1660, %r1619, %r1614, %r1616;
mov.b32 %f260, %r1661;
add.ftz.f32 %f261, %f259, %f260;
mov.b32 %r1662, %f261;
shfl.sync.bfly.b32 %r1663|%p79, %r1662, %r1622, %r1614, %r1616;
mov.b32 %f262, %r1663;
add.ftz.f32 %f263, %f261, %f262;
mov.b32 %r1664, %f263;
shfl.sync.bfly.b32 %r1665|%p80, %r1664, %r1625, %r1614, %r1616;
mov.b32 %f264, %r1665;
add.ftz.f32 %f265, %f263, %f264;
mov.b32 %r1666, %f265;
shfl.sync.bfly.b32 %r1667|%p81, %r1666, %r1628, %r1614, %r1616;
mov.b32 %f266, %r1667;
add.ftz.f32 %f43, %f265, %f266;
setp.ne.s32 %p82, %r2, 1;
add.s32 %r1669, %r4, 1;
setp.ge.u32 %p83, %r1669, %r50;
or.pred %p84, %p82, %p83;
@%p84 bra $L__BB200_50;

ld.param.f32 %f980, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f267, %f5, %f42;
add.ftz.f32 %f268, %f4, %f43;
min.ftz.f32 %f269, %f268, %f980;
min.ftz.f32 %f270, %f267, %f980;
max.ftz.f32 %f271, %f270, %f35;
mul.ftz.f32 %f272, %f269, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f273, %f272;
add.ftz.f32 %f274, %f273, 0f3F800000;
div.approx.ftz.f32 %f275, %f269, %f274;
mul.ftz.f32 %f276, %f271, %f275;
st.global.f32 [%rd27+4], %f276;

$L__BB200_50:
ld.shared.f32 %f277, [%r31+256];
add.ftz.f32 %f278, %f277, %f28;
ld.shared.f32 %f279, [%r33+256];
add.ftz.f32 %f280, %f279, %f32;
ld.shared.f32 %f281, [%r31+768];
add.ftz.f32 %f282, %f281, %f278;
ld.shared.f32 %f283, [%r33+768];
add.ftz.f32 %f284, %f283, %f280;
ld.shared.f32 %f285, [%r31+1280];
add.ftz.f32 %f286, %f285, %f282;
ld.shared.f32 %f287, [%r33+1280];
add.ftz.f32 %f288, %f287, %f284;
mov.b32 %r1670, %f286;
mov.u32 %r1671, 31;
mov.u32 %r1672, 16;
mov.u32 %r1673, -1;
shfl.sync.bfly.b32 %r1674|%p85, %r1670, %r1672, %r1671, %r1673;
mov.b32 %f289, %r1674;
add.ftz.f32 %f290, %f286, %f289;
mov.b32 %r1675, %f290;
mov.u32 %r1676, 8;
shfl.sync.bfly.b32 %r1677|%p86, %r1675, %r1676, %r1671, %r1673;
mov.b32 %f291, %r1677;
add.ftz.f32 %f292, %f290, %f291;
mov.b32 %r1678, %f292;
mov.u32 %r1679, 4;
shfl.sync.bfly.b32 %r1680|%p87, %r1678, %r1679, %r1671, %r1673;
mov.b32 %f293, %r1680;
add.ftz.f32 %f294, %f292, %f293;
mov.b32 %r1681, %f294;
mov.u32 %r1682, 2;
shfl.sync.bfly.b32 %r1683|%p88, %r1681, %r1682, %r1671, %r1673;
mov.b32 %f295, %r1683;
add.ftz.f32 %f296, %f294, %f295;
mov.b32 %r1684, %f296;
mov.u32 %r1685, 1;
shfl.sync.bfly.b32 %r1686|%p89, %r1684, %r1685, %r1671, %r1673;
mov.b32 %f297, %r1686;
add.ftz.f32 %f44, %f296, %f297;
mov.b32 %r1687, %f288;
shfl.sync.bfly.b32 %r1688|%p90, %r1687, %r1672, %r1671, %r1673;
mov.b32 %f298, %r1688;
add.ftz.f32 %f299, %f288, %f298;
mov.b32 %r1689, %f299;
shfl.sync.bfly.b32 %r1690|%p91, %r1689, %r1676, %r1671, %r1673;
mov.b32 %f300, %r1690;
add.ftz.f32 %f301, %f299, %f300;
mov.b32 %r1691, %f301;
shfl.sync.bfly.b32 %r1692|%p92, %r1691, %r1679, %r1671, %r1673;
mov.b32 %f302, %r1692;
add.ftz.f32 %f303, %f301, %f302;
mov.b32 %r1693, %f303;
shfl.sync.bfly.b32 %r1694|%p93, %r1693, %r1682, %r1671, %r1673;
mov.b32 %f304, %r1694;
add.ftz.f32 %f305, %f303, %f304;
mov.b32 %r1695, %f305;
shfl.sync.bfly.b32 %r1696|%p94, %r1695, %r1685, %r1671, %r1673;
mov.b32 %f306, %r1696;
add.ftz.f32 %f45, %f305, %f306;
setp.ne.s32 %p95, %r2, 2;
add.s32 %r1698, %r4, 2;
setp.ge.u32 %p96, %r1698, %r50;
or.pred %p97, %p95, %p96;
@%p97 bra $L__BB200_52;

ld.param.f32 %f979, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f307, %f5, %f44;
add.ftz.f32 %f308, %f4, %f45;
min.ftz.f32 %f309, %f308, %f979;
min.ftz.f32 %f310, %f307, %f979;
max.ftz.f32 %f311, %f310, %f35;
mul.ftz.f32 %f312, %f309, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f313, %f312;
add.ftz.f32 %f314, %f313, 0f3F800000;
div.approx.ftz.f32 %f315, %f309, %f314;
mul.ftz.f32 %f316, %f311, %f315;
st.global.f32 [%rd27+8], %f316;

$L__BB200_52:
ld.shared.f32 %f317, [%r31+384];
add.ftz.f32 %f318, %f317, %f27;
ld.shared.f32 %f319, [%r33+384];
add.ftz.f32 %f320, %f319, %f31;
ld.shared.f32 %f321, [%r31+896];
add.ftz.f32 %f322, %f321, %f318;
ld.shared.f32 %f323, [%r33+896];
add.ftz.f32 %f324, %f323, %f320;
ld.shared.f32 %f325, [%r31+1408];
add.ftz.f32 %f326, %f325, %f322;
ld.shared.f32 %f327, [%r33+1408];
add.ftz.f32 %f328, %f327, %f324;
mov.b32 %r1699, %f326;
shfl.sync.bfly.b32 %r1703|%p98, %r1699, %r1672, %r1671, %r1673;
mov.b32 %f329, %r1703;
add.ftz.f32 %f330, %f326, %f329;
mov.b32 %r1704, %f330;
shfl.sync.bfly.b32 %r1706|%p99, %r1704, %r1676, %r1671, %r1673;
mov.b32 %f331, %r1706;
add.ftz.f32 %f332, %f330, %f331;
mov.b32 %r1707, %f332;
shfl.sync.bfly.b32 %r1709|%p100, %r1707, %r1679, %r1671, %r1673;
mov.b32 %f333, %r1709;
add.ftz.f32 %f334, %f332, %f333;
mov.b32 %r1710, %f334;
shfl.sync.bfly.b32 %r1712|%p101, %r1710, %r1682, %r1671, %r1673;
mov.b32 %f335, %r1712;
add.ftz.f32 %f336, %f334, %f335;
mov.b32 %r1713, %f336;
shfl.sync.bfly.b32 %r1715|%p102, %r1713, %r1685, %r1671, %r1673;
mov.b32 %f337, %r1715;
add.ftz.f32 %f46, %f336, %f337;
mov.b32 %r1716, %f328;
shfl.sync.bfly.b32 %r1717|%p103, %r1716, %r1672, %r1671, %r1673;
mov.b32 %f338, %r1717;
add.ftz.f32 %f339, %f328, %f338;
mov.b32 %r1718, %f339;
shfl.sync.bfly.b32 %r1719|%p104, %r1718, %r1676, %r1671, %r1673;
mov.b32 %f340, %r1719;
add.ftz.f32 %f341, %f339, %f340;
mov.b32 %r1720, %f341;
shfl.sync.bfly.b32 %r1721|%p105, %r1720, %r1679, %r1671, %r1673;
mov.b32 %f342, %r1721;
add.ftz.f32 %f343, %f341, %f342;
mov.b32 %r1722, %f343;
shfl.sync.bfly.b32 %r1723|%p106, %r1722, %r1682, %r1671, %r1673;
mov.b32 %f344, %r1723;
add.ftz.f32 %f345, %f343, %f344;
mov.b32 %r1724, %f345;
shfl.sync.bfly.b32 %r1725|%p107, %r1724, %r1685, %r1671, %r1673;
mov.b32 %f346, %r1725;
add.ftz.f32 %f47, %f345, %f346;
setp.ne.s32 %p108, %r2, 3;
add.s32 %r1727, %r4, 3;
setp.ge.u32 %p109, %r1727, %r50;
or.pred %p110, %p108, %p109;
@%p110 bra $L__BB200_78;

ld.param.f32 %f978, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f347, %f5, %f46;
add.ftz.f32 %f348, %f4, %f47;
min.ftz.f32 %f349, %f348, %f978;
min.ftz.f32 %f350, %f347, %f978;
max.ftz.f32 %f351, %f350, %f35;
mul.ftz.f32 %f352, %f349, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f353, %f352;
add.ftz.f32 %f354, %f353, 0f3F800000;
div.approx.ftz.f32 %f355, %f349, %f354;
mul.ftz.f32 %f356, %f351, %f355;
st.global.f32 [%rd27+12], %f356;
bra.uni $L__BB200_78;

$L__BB200_70:
ld.shared.f32 %f845, [%r31];
add.ftz.f32 %f846, %f845, %f30;
ld.shared.f32 %f847, [%r33];
add.ftz.f32 %f848, %f847, %f34;
ld.shared.f32 %f849, [%r31+512];
add.ftz.f32 %f850, %f849, %f846;
ld.shared.f32 %f851, [%r33+512];
add.ftz.f32 %f852, %f851, %f848;
ld.shared.f32 %f853, [%r31+1024];
add.ftz.f32 %f854, %f853, %f850;
ld.shared.f32 %f855, [%r33+1024];
add.ftz.f32 %f856, %f855, %f852;
mov.b32 %r2073, %f854;
mov.u32 %r2074, 31;
mov.u32 %r2075, 16;
mov.u32 %r2076, -1;
shfl.sync.bfly.b32 %r2077|%p267, %r2073, %r2075, %r2074, %r2076;
mov.b32 %f857, %r2077;
add.ftz.f32 %f858, %f854, %f857;
mov.b32 %r2078, %f858;
mov.u32 %r2079, 8;
shfl.sync.bfly.b32 %r2080|%p268, %r2078, %r2079, %r2074, %r2076;
mov.b32 %f859, %r2080;
add.ftz.f32 %f860, %f858, %f859;
mov.b32 %r2081, %f860;
mov.u32 %r2082, 4;
shfl.sync.bfly.b32 %r2083|%p269, %r2081, %r2082, %r2074, %r2076;
mov.b32 %f861, %r2083;
add.ftz.f32 %f862, %f860, %f861;
mov.b32 %r2084, %f862;
mov.u32 %r2085, 2;
shfl.sync.bfly.b32 %r2086|%p270, %r2084, %r2085, %r2074, %r2076;
mov.b32 %f863, %r2086;
add.ftz.f32 %f864, %f862, %f863;
mov.b32 %r2087, %f864;
mov.u32 %r2088, 1;
shfl.sync.bfly.b32 %r2089|%p271, %r2087, %r2088, %r2074, %r2076;
mov.b32 %f865, %r2089;
add.ftz.f32 %f72, %f864, %f865;
mov.b32 %r2090, %f856;
shfl.sync.bfly.b32 %r2091|%p272, %r2090, %r2075, %r2074, %r2076;
mov.b32 %f866, %r2091;
add.ftz.f32 %f867, %f856, %f866;
mov.b32 %r2092, %f867;
shfl.sync.bfly.b32 %r2093|%p273, %r2092, %r2079, %r2074, %r2076;
mov.b32 %f868, %r2093;
add.ftz.f32 %f869, %f867, %f868;
mov.b32 %r2094, %f869;
shfl.sync.bfly.b32 %r2095|%p274, %r2094, %r2082, %r2074, %r2076;
mov.b32 %f870, %r2095;
add.ftz.f32 %f871, %f869, %f870;
mov.b32 %r2096, %f871;
shfl.sync.bfly.b32 %r2097|%p275, %r2096, %r2085, %r2074, %r2076;
mov.b32 %f872, %r2097;
add.ftz.f32 %f873, %f871, %f872;
mov.b32 %r2098, %f873;
shfl.sync.bfly.b32 %r2099|%p276, %r2098, %r2088, %r2074, %r2076;
mov.b32 %f874, %r2099;
add.ftz.f32 %f73, %f873, %f874;
setp.ne.s32 %p277, %r2, 0;
setp.ge.u32 %p278, %r4, %r50;
or.pred %p279, %p277, %p278;
@%p279 bra $L__BB200_72;

add.ftz.f32 %f875, %f5, %f72;
add.ftz.f32 %f876, %f4, %f73;
mul.ftz.f32 %f877, %f875, %f876;
st.global.f32 [%rd27], %f877;

$L__BB200_72:
ld.shared.f32 %f878, [%r31+128];
add.ftz.f32 %f879, %f878, %f29;
ld.shared.f32 %f880, [%r33+128];
add.ftz.f32 %f881, %f880, %f33;
ld.shared.f32 %f882, [%r31+640];
add.ftz.f32 %f883, %f882, %f879;
ld.shared.f32 %f884, [%r33+640];
add.ftz.f32 %f885, %f884, %f881;
ld.shared.f32 %f886, [%r31+1152];
add.ftz.f32 %f887, %f886, %f883;
ld.shared.f32 %f888, [%r33+1152];
add.ftz.f32 %f889, %f888, %f885;
mov.b32 %r2101, %f887;
shfl.sync.bfly.b32 %r2105|%p280, %r2101, %r2075, %r2074, %r2076;
mov.b32 %f890, %r2105;
add.ftz.f32 %f891, %f887, %f890;
mov.b32 %r2106, %f891;
shfl.sync.bfly.b32 %r2108|%p281, %r2106, %r2079, %r2074, %r2076;
mov.b32 %f892, %r2108;
add.ftz.f32 %f893, %f891, %f892;
mov.b32 %r2109, %f893;
shfl.sync.bfly.b32 %r2111|%p282, %r2109, %r2082, %r2074, %r2076;
mov.b32 %f894, %r2111;
add.ftz.f32 %f895, %f893, %f894;
mov.b32 %r2112, %f895;
shfl.sync.bfly.b32 %r2114|%p283, %r2112, %r2085, %r2074, %r2076;
mov.b32 %f896, %r2114;
add.ftz.f32 %f897, %f895, %f896;
mov.b32 %r2115, %f897;
shfl.sync.bfly.b32 %r2117|%p284, %r2115, %r2088, %r2074, %r2076;
mov.b32 %f898, %r2117;
add.ftz.f32 %f74, %f897, %f898;
mov.b32 %r2118, %f889;
shfl.sync.bfly.b32 %r2119|%p285, %r2118, %r2075, %r2074, %r2076;
mov.b32 %f899, %r2119;
add.ftz.f32 %f900, %f889, %f899;
mov.b32 %r2120, %f900;
shfl.sync.bfly.b32 %r2121|%p286, %r2120, %r2079, %r2074, %r2076;
mov.b32 %f901, %r2121;
add.ftz.f32 %f902, %f900, %f901;
mov.b32 %r2122, %f902;
shfl.sync.bfly.b32 %r2123|%p287, %r2122, %r2082, %r2074, %r2076;
mov.b32 %f903, %r2123;
add.ftz.f32 %f904, %f902, %f903;
mov.b32 %r2124, %f904;
shfl.sync.bfly.b32 %r2125|%p288, %r2124, %r2085, %r2074, %r2076;
mov.b32 %f905, %r2125;
add.ftz.f32 %f906, %f904, %f905;
mov.b32 %r2126, %f906;
shfl.sync.bfly.b32 %r2127|%p289, %r2126, %r2088, %r2074, %r2076;
mov.b32 %f907, %r2127;
add.ftz.f32 %f75, %f906, %f907;
setp.ne.s32 %p290, %r2, 1;
add.s32 %r2129, %r4, 1;
setp.ge.u32 %p291, %r2129, %r50;
or.pred %p292, %p290, %p291;
@%p292 bra $L__BB200_74;

add.ftz.f32 %f908, %f5, %f74;
add.ftz.f32 %f909, %f4, %f75;
mul.ftz.f32 %f910, %f908, %f909;
st.global.f32 [%rd27+4], %f910;

$L__BB200_74:
ld.shared.f32 %f911, [%r31+256];
add.ftz.f32 %f912, %f911, %f28;
ld.shared.f32 %f913, [%r33+256];
add.ftz.f32 %f914, %f913, %f32;
ld.shared.f32 %f915, [%r31+768];
add.ftz.f32 %f916, %f915, %f912;
ld.shared.f32 %f917, [%r33+768];
add.ftz.f32 %f918, %f917, %f914;
ld.shared.f32 %f919, [%r31+1280];
add.ftz.f32 %f920, %f919, %f916;
ld.shared.f32 %f921, [%r33+1280];
add.ftz.f32 %f922, %f921, %f918;
mov.b32 %r2130, %f920;
mov.u32 %r2131, 31;
mov.u32 %r2132, 16;
mov.u32 %r2133, -1;
shfl.sync.bfly.b32 %r2134|%p293, %r2130, %r2132, %r2131, %r2133;
mov.b32 %f923, %r2134;
add.ftz.f32 %f924, %f920, %f923;
mov.b32 %r2135, %f924;
mov.u32 %r2136, 8;
shfl.sync.bfly.b32 %r2137|%p294, %r2135, %r2136, %r2131, %r2133;
mov.b32 %f925, %r2137;
add.ftz.f32 %f926, %f924, %f925;
mov.b32 %r2138, %f926;
mov.u32 %r2139, 4;
shfl.sync.bfly.b32 %r2140|%p295, %r2138, %r2139, %r2131, %r2133;
mov.b32 %f927, %r2140;
add.ftz.f32 %f928, %f926, %f927;
mov.b32 %r2141, %f928;
mov.u32 %r2142, 2;
shfl.sync.bfly.b32 %r2143|%p296, %r2141, %r2142, %r2131, %r2133;
mov.b32 %f929, %r2143;
add.ftz.f32 %f930, %f928, %f929;
mov.b32 %r2144, %f930;
mov.u32 %r2145, 1;
shfl.sync.bfly.b32 %r2146|%p297, %r2144, %r2145, %r2131, %r2133;
mov.b32 %f931, %r2146;
add.ftz.f32 %f76, %f930, %f931;
mov.b32 %r2147, %f922;
shfl.sync.bfly.b32 %r2148|%p298, %r2147, %r2132, %r2131, %r2133;
mov.b32 %f932, %r2148;
add.ftz.f32 %f933, %f922, %f932;
mov.b32 %r2149, %f933;
shfl.sync.bfly.b32 %r2150|%p299, %r2149, %r2136, %r2131, %r2133;
mov.b32 %f934, %r2150;
add.ftz.f32 %f935, %f933, %f934;
mov.b32 %r2151, %f935;
shfl.sync.bfly.b32 %r2152|%p300, %r2151, %r2139, %r2131, %r2133;
mov.b32 %f936, %r2152;
add.ftz.f32 %f937, %f935, %f936;
mov.b32 %r2153, %f937;
shfl.sync.bfly.b32 %r2154|%p301, %r2153, %r2142, %r2131, %r2133;
mov.b32 %f938, %r2154;
add.ftz.f32 %f939, %f937, %f938;
mov.b32 %r2155, %f939;
shfl.sync.bfly.b32 %r2156|%p302, %r2155, %r2145, %r2131, %r2133;
mov.b32 %f940, %r2156;
add.ftz.f32 %f77, %f939, %f940;
setp.ne.s32 %p303, %r2, 2;
add.s32 %r2158, %r4, 2;
setp.ge.u32 %p304, %r2158, %r50;
or.pred %p305, %p303, %p304;
@%p305 bra $L__BB200_76;

add.ftz.f32 %f941, %f5, %f76;
add.ftz.f32 %f942, %f4, %f77;
mul.ftz.f32 %f943, %f941, %f942;
st.global.f32 [%rd27+8], %f943;

$L__BB200_76:
ld.shared.f32 %f944, [%r31+384];
add.ftz.f32 %f945, %f944, %f27;
ld.shared.f32 %f946, [%r33+384];
add.ftz.f32 %f947, %f946, %f31;
ld.shared.f32 %f948, [%r31+896];
add.ftz.f32 %f949, %f948, %f945;
ld.shared.f32 %f950, [%r33+896];
add.ftz.f32 %f951, %f950, %f947;
ld.shared.f32 %f952, [%r31+1408];
add.ftz.f32 %f953, %f952, %f949;
ld.shared.f32 %f954, [%r33+1408];
add.ftz.f32 %f955, %f954, %f951;
mov.b32 %r2159, %f953;
shfl.sync.bfly.b32 %r2163|%p306, %r2159, %r2132, %r2131, %r2133;
mov.b32 %f956, %r2163;
add.ftz.f32 %f957, %f953, %f956;
mov.b32 %r2164, %f957;
shfl.sync.bfly.b32 %r2166|%p307, %r2164, %r2136, %r2131, %r2133;
mov.b32 %f958, %r2166;
add.ftz.f32 %f959, %f957, %f958;
mov.b32 %r2167, %f959;
shfl.sync.bfly.b32 %r2169|%p308, %r2167, %r2139, %r2131, %r2133;
mov.b32 %f960, %r2169;
add.ftz.f32 %f961, %f959, %f960;
mov.b32 %r2170, %f961;
shfl.sync.bfly.b32 %r2172|%p309, %r2170, %r2142, %r2131, %r2133;
mov.b32 %f962, %r2172;
add.ftz.f32 %f963, %f961, %f962;
mov.b32 %r2173, %f963;
shfl.sync.bfly.b32 %r2175|%p310, %r2173, %r2145, %r2131, %r2133;
mov.b32 %f964, %r2175;
add.ftz.f32 %f78, %f963, %f964;
mov.b32 %r2176, %f955;
shfl.sync.bfly.b32 %r2177|%p311, %r2176, %r2132, %r2131, %r2133;
mov.b32 %f965, %r2177;
add.ftz.f32 %f966, %f955, %f965;
mov.b32 %r2178, %f966;
shfl.sync.bfly.b32 %r2179|%p312, %r2178, %r2136, %r2131, %r2133;
mov.b32 %f967, %r2179;
add.ftz.f32 %f968, %f966, %f967;
mov.b32 %r2180, %f968;
shfl.sync.bfly.b32 %r2181|%p313, %r2180, %r2139, %r2131, %r2133;
mov.b32 %f969, %r2181;
add.ftz.f32 %f970, %f968, %f969;
mov.b32 %r2182, %f970;
shfl.sync.bfly.b32 %r2183|%p314, %r2182, %r2142, %r2131, %r2133;
mov.b32 %f971, %r2183;
add.ftz.f32 %f972, %f970, %f971;
mov.b32 %r2184, %f972;
shfl.sync.bfly.b32 %r2185|%p315, %r2184, %r2145, %r2131, %r2133;
mov.b32 %f973, %r2185;
add.ftz.f32 %f79, %f972, %f973;
setp.ne.s32 %p316, %r2, 3;
add.s32 %r2187, %r4, 3;
setp.ge.u32 %p317, %r2187, %r50;
or.pred %p318, %p316, %p317;
@%p318 bra $L__BB200_78;

add.ftz.f32 %f974, %f5, %f78;
add.ftz.f32 %f975, %f4, %f79;
mul.ftz.f32 %f976, %f974, %f975;
st.global.f32 [%rd27+12], %f976;
bra.uni $L__BB200_78;

$L__BB200_62:
ld.shared.f32 %f681, [%r31];
add.ftz.f32 %f682, %f681, %f30;
ld.shared.f32 %f683, [%r33];
add.ftz.f32 %f684, %f683, %f34;
ld.shared.f32 %f685, [%r31+512];
add.ftz.f32 %f686, %f685, %f682;
ld.shared.f32 %f687, [%r33+512];
add.ftz.f32 %f688, %f687, %f684;
ld.shared.f32 %f689, [%r31+1024];
add.ftz.f32 %f690, %f689, %f686;
ld.shared.f32 %f691, [%r33+1024];
add.ftz.f32 %f692, %f691, %f688;
mov.b32 %r1958, %f690;
mov.u32 %r1959, 31;
mov.u32 %r1960, 16;
mov.u32 %r1961, -1;
shfl.sync.bfly.b32 %r1962|%p215, %r1958, %r1960, %r1959, %r1961;
mov.b32 %f693, %r1962;
add.ftz.f32 %f694, %f690, %f693;
mov.b32 %r1963, %f694;
mov.u32 %r1964, 8;
shfl.sync.bfly.b32 %r1965|%p216, %r1963, %r1964, %r1959, %r1961;
mov.b32 %f695, %r1965;
add.ftz.f32 %f696, %f694, %f695;
mov.b32 %r1966, %f696;
mov.u32 %r1967, 4;
shfl.sync.bfly.b32 %r1968|%p217, %r1966, %r1967, %r1959, %r1961;
mov.b32 %f697, %r1968;
add.ftz.f32 %f698, %f696, %f697;
mov.b32 %r1969, %f698;
mov.u32 %r1970, 2;
shfl.sync.bfly.b32 %r1971|%p218, %r1969, %r1970, %r1959, %r1961;
mov.b32 %f699, %r1971;
add.ftz.f32 %f700, %f698, %f699;
mov.b32 %r1972, %f700;
mov.u32 %r1973, 1;
shfl.sync.bfly.b32 %r1974|%p219, %r1972, %r1973, %r1959, %r1961;
mov.b32 %f701, %r1974;
add.ftz.f32 %f64, %f700, %f701;
mov.b32 %r1975, %f692;
shfl.sync.bfly.b32 %r1976|%p220, %r1975, %r1960, %r1959, %r1961;
mov.b32 %f702, %r1976;
add.ftz.f32 %f703, %f692, %f702;
mov.b32 %r1977, %f703;
shfl.sync.bfly.b32 %r1978|%p221, %r1977, %r1964, %r1959, %r1961;
mov.b32 %f704, %r1978;
add.ftz.f32 %f705, %f703, %f704;
mov.b32 %r1979, %f705;
shfl.sync.bfly.b32 %r1980|%p222, %r1979, %r1967, %r1959, %r1961;
mov.b32 %f706, %r1980;
add.ftz.f32 %f707, %f705, %f706;
mov.b32 %r1981, %f707;
shfl.sync.bfly.b32 %r1982|%p223, %r1981, %r1970, %r1959, %r1961;
mov.b32 %f708, %r1982;
add.ftz.f32 %f709, %f707, %f708;
mov.b32 %r1983, %f709;
shfl.sync.bfly.b32 %r1984|%p224, %r1983, %r1973, %r1959, %r1961;
mov.b32 %f710, %r1984;
add.ftz.f32 %f65, %f709, %f710;
setp.ne.s32 %p225, %r2, 0;
setp.ge.u32 %p226, %r4, %r50;
or.pred %p227, %p225, %p226;
@%p227 bra $L__BB200_64;

add.ftz.f32 %f711, %f5, %f64;
add.ftz.f32 %f712, %f4, %f65;
mul.ftz.f32 %f713, %f712, 0f3F000000;
mul.ftz.f32 %f714, %f712, 0f3F4C422A;
mul.ftz.f32 %f715, %f712, 0f3D372713;
fma.rn.ftz.f32 %f716, %f712, %f715, 0f3F800000;
mul.ftz.f32 %f717, %f714, %f716;
tanh.approx.f32 %f718, %f717;
add.ftz.f32 %f719, %f718, 0f3F800000;
mul.ftz.f32 %f720, %f713, %f719;
mul.ftz.f32 %f721, %f711, %f720;
st.global.f32 [%rd27], %f721;

$L__BB200_64:
ld.shared.f32 %f722, [%r31+128];
add.ftz.f32 %f723, %f722, %f29;
ld.shared.f32 %f724, [%r33+128];
add.ftz.f32 %f725, %f724, %f33;
ld.shared.f32 %f726, [%r31+640];
add.ftz.f32 %f727, %f726, %f723;
ld.shared.f32 %f728, [%r33+640];
add.ftz.f32 %f729, %f728, %f725;
ld.shared.f32 %f730, [%r31+1152];
add.ftz.f32 %f731, %f730, %f727;
ld.shared.f32 %f732, [%r33+1152];
add.ftz.f32 %f733, %f732, %f729;
mov.b32 %r1986, %f731;
shfl.sync.bfly.b32 %r1990|%p228, %r1986, %r1960, %r1959, %r1961;
mov.b32 %f734, %r1990;
add.ftz.f32 %f735, %f731, %f734;
mov.b32 %r1991, %f735;
shfl.sync.bfly.b32 %r1993|%p229, %r1991, %r1964, %r1959, %r1961;
mov.b32 %f736, %r1993;
add.ftz.f32 %f737, %f735, %f736;
mov.b32 %r1994, %f737;
shfl.sync.bfly.b32 %r1996|%p230, %r1994, %r1967, %r1959, %r1961;
mov.b32 %f738, %r1996;
add.ftz.f32 %f739, %f737, %f738;
mov.b32 %r1997, %f739;
shfl.sync.bfly.b32 %r1999|%p231, %r1997, %r1970, %r1959, %r1961;
mov.b32 %f740, %r1999;
add.ftz.f32 %f741, %f739, %f740;
mov.b32 %r2000, %f741;
shfl.sync.bfly.b32 %r2002|%p232, %r2000, %r1973, %r1959, %r1961;
mov.b32 %f742, %r2002;
add.ftz.f32 %f66, %f741, %f742;
mov.b32 %r2003, %f733;
shfl.sync.bfly.b32 %r2004|%p233, %r2003, %r1960, %r1959, %r1961;
mov.b32 %f743, %r2004;
add.ftz.f32 %f744, %f733, %f743;
mov.b32 %r2005, %f744;
shfl.sync.bfly.b32 %r2006|%p234, %r2005, %r1964, %r1959, %r1961;
mov.b32 %f745, %r2006;
add.ftz.f32 %f746, %f744, %f745;
mov.b32 %r2007, %f746;
shfl.sync.bfly.b32 %r2008|%p235, %r2007, %r1967, %r1959, %r1961;
mov.b32 %f747, %r2008;
add.ftz.f32 %f748, %f746, %f747;
mov.b32 %r2009, %f748;
shfl.sync.bfly.b32 %r2010|%p236, %r2009, %r1970, %r1959, %r1961;
mov.b32 %f749, %r2010;
add.ftz.f32 %f750, %f748, %f749;
mov.b32 %r2011, %f750;
shfl.sync.bfly.b32 %r2012|%p237, %r2011, %r1973, %r1959, %r1961;
mov.b32 %f751, %r2012;
add.ftz.f32 %f67, %f750, %f751;
setp.ne.s32 %p238, %r2, 1;
add.s32 %r2014, %r4, 1;
setp.ge.u32 %p239, %r2014, %r50;
or.pred %p240, %p238, %p239;
@%p240 bra $L__BB200_66;

add.ftz.f32 %f752, %f5, %f66;
add.ftz.f32 %f753, %f4, %f67;
mul.ftz.f32 %f754, %f753, 0f3F000000;
mul.ftz.f32 %f755, %f753, 0f3F4C422A;
mul.ftz.f32 %f756, %f753, 0f3D372713;
fma.rn.ftz.f32 %f757, %f753, %f756, 0f3F800000;
mul.ftz.f32 %f758, %f755, %f757;
tanh.approx.f32 %f759, %f758;
add.ftz.f32 %f760, %f759, 0f3F800000;
mul.ftz.f32 %f761, %f754, %f760;
mul.ftz.f32 %f762, %f752, %f761;
st.global.f32 [%rd27+4], %f762;

$L__BB200_66:
ld.shared.f32 %f763, [%r31+256];
add.ftz.f32 %f764, %f763, %f28;
ld.shared.f32 %f765, [%r33+256];
add.ftz.f32 %f766, %f765, %f32;
ld.shared.f32 %f767, [%r31+768];
add.ftz.f32 %f768, %f767, %f764;
ld.shared.f32 %f769, [%r33+768];
add.ftz.f32 %f770, %f769, %f766;
ld.shared.f32 %f771, [%r31+1280];
add.ftz.f32 %f772, %f771, %f768;
ld.shared.f32 %f773, [%r33+1280];
add.ftz.f32 %f774, %f773, %f770;
mov.b32 %r2015, %f772;
mov.u32 %r2016, 31;
mov.u32 %r2017, 16;
mov.u32 %r2018, -1;
shfl.sync.bfly.b32 %r2019|%p241, %r2015, %r2017, %r2016, %r2018;
mov.b32 %f775, %r2019;
add.ftz.f32 %f776, %f772, %f775;
mov.b32 %r2020, %f776;
mov.u32 %r2021, 8;
shfl.sync.bfly.b32 %r2022|%p242, %r2020, %r2021, %r2016, %r2018;
mov.b32 %f777, %r2022;
add.ftz.f32 %f778, %f776, %f777;
mov.b32 %r2023, %f778;
mov.u32 %r2024, 4;
shfl.sync.bfly.b32 %r2025|%p243, %r2023, %r2024, %r2016, %r2018;
mov.b32 %f779, %r2025;
add.ftz.f32 %f780, %f778, %f779;
mov.b32 %r2026, %f780;
mov.u32 %r2027, 2;
shfl.sync.bfly.b32 %r2028|%p244, %r2026, %r2027, %r2016, %r2018;
mov.b32 %f781, %r2028;
add.ftz.f32 %f782, %f780, %f781;
mov.b32 %r2029, %f782;
mov.u32 %r2030, 1;
shfl.sync.bfly.b32 %r2031|%p245, %r2029, %r2030, %r2016, %r2018;
mov.b32 %f783, %r2031;
add.ftz.f32 %f68, %f782, %f783;
mov.b32 %r2032, %f774;
shfl.sync.bfly.b32 %r2033|%p246, %r2032, %r2017, %r2016, %r2018;
mov.b32 %f784, %r2033;
add.ftz.f32 %f785, %f774, %f784;
mov.b32 %r2034, %f785;
shfl.sync.bfly.b32 %r2035|%p247, %r2034, %r2021, %r2016, %r2018;
mov.b32 %f786, %r2035;
add.ftz.f32 %f787, %f785, %f786;
mov.b32 %r2036, %f787;
shfl.sync.bfly.b32 %r2037|%p248, %r2036, %r2024, %r2016, %r2018;
mov.b32 %f788, %r2037;
add.ftz.f32 %f789, %f787, %f788;
mov.b32 %r2038, %f789;
shfl.sync.bfly.b32 %r2039|%p249, %r2038, %r2027, %r2016, %r2018;
mov.b32 %f790, %r2039;
add.ftz.f32 %f791, %f789, %f790;
mov.b32 %r2040, %f791;
shfl.sync.bfly.b32 %r2041|%p250, %r2040, %r2030, %r2016, %r2018;
mov.b32 %f792, %r2041;
add.ftz.f32 %f69, %f791, %f792;
setp.ne.s32 %p251, %r2, 2;
add.s32 %r2043, %r4, 2;
setp.ge.u32 %p252, %r2043, %r50;
or.pred %p253, %p251, %p252;
@%p253 bra $L__BB200_68;

add.ftz.f32 %f793, %f5, %f68;
add.ftz.f32 %f794, %f4, %f69;
mul.ftz.f32 %f795, %f794, 0f3F000000;
mul.ftz.f32 %f796, %f794, 0f3F4C422A;
mul.ftz.f32 %f797, %f794, 0f3D372713;
fma.rn.ftz.f32 %f798, %f794, %f797, 0f3F800000;
mul.ftz.f32 %f799, %f796, %f798;
tanh.approx.f32 %f800, %f799;
add.ftz.f32 %f801, %f800, 0f3F800000;
mul.ftz.f32 %f802, %f795, %f801;
mul.ftz.f32 %f803, %f793, %f802;
st.global.f32 [%rd27+8], %f803;

$L__BB200_68:
ld.shared.f32 %f804, [%r31+384];
add.ftz.f32 %f805, %f804, %f27;
ld.shared.f32 %f806, [%r33+384];
add.ftz.f32 %f807, %f806, %f31;
ld.shared.f32 %f808, [%r31+896];
add.ftz.f32 %f809, %f808, %f805;
ld.shared.f32 %f810, [%r33+896];
add.ftz.f32 %f811, %f810, %f807;
ld.shared.f32 %f812, [%r31+1408];
add.ftz.f32 %f813, %f812, %f809;
ld.shared.f32 %f814, [%r33+1408];
add.ftz.f32 %f815, %f814, %f811;
mov.b32 %r2044, %f813;
shfl.sync.bfly.b32 %r2048|%p254, %r2044, %r2017, %r2016, %r2018;
mov.b32 %f816, %r2048;
add.ftz.f32 %f817, %f813, %f816;
mov.b32 %r2049, %f817;
shfl.sync.bfly.b32 %r2051|%p255, %r2049, %r2021, %r2016, %r2018;
mov.b32 %f818, %r2051;
add.ftz.f32 %f819, %f817, %f818;
mov.b32 %r2052, %f819;
shfl.sync.bfly.b32 %r2054|%p256, %r2052, %r2024, %r2016, %r2018;
mov.b32 %f820, %r2054;
add.ftz.f32 %f821, %f819, %f820;
mov.b32 %r2055, %f821;
shfl.sync.bfly.b32 %r2057|%p257, %r2055, %r2027, %r2016, %r2018;
mov.b32 %f822, %r2057;
add.ftz.f32 %f823, %f821, %f822;
mov.b32 %r2058, %f823;
shfl.sync.bfly.b32 %r2060|%p258, %r2058, %r2030, %r2016, %r2018;
mov.b32 %f824, %r2060;
add.ftz.f32 %f70, %f823, %f824;
mov.b32 %r2061, %f815;
shfl.sync.bfly.b32 %r2062|%p259, %r2061, %r2017, %r2016, %r2018;
mov.b32 %f825, %r2062;
add.ftz.f32 %f826, %f815, %f825;
mov.b32 %r2063, %f826;
shfl.sync.bfly.b32 %r2064|%p260, %r2063, %r2021, %r2016, %r2018;
mov.b32 %f827, %r2064;
add.ftz.f32 %f828, %f826, %f827;
mov.b32 %r2065, %f828;
shfl.sync.bfly.b32 %r2066|%p261, %r2065, %r2024, %r2016, %r2018;
mov.b32 %f829, %r2066;
add.ftz.f32 %f830, %f828, %f829;
mov.b32 %r2067, %f830;
shfl.sync.bfly.b32 %r2068|%p262, %r2067, %r2027, %r2016, %r2018;
mov.b32 %f831, %r2068;
add.ftz.f32 %f832, %f830, %f831;
mov.b32 %r2069, %f832;
shfl.sync.bfly.b32 %r2070|%p263, %r2069, %r2030, %r2016, %r2018;
mov.b32 %f833, %r2070;
add.ftz.f32 %f71, %f832, %f833;
setp.ne.s32 %p264, %r2, 3;
add.s32 %r2072, %r4, 3;
setp.ge.u32 %p265, %r2072, %r50;
or.pred %p266, %p264, %p265;
@%p266 bra $L__BB200_78;

add.ftz.f32 %f834, %f5, %f70;
add.ftz.f32 %f835, %f4, %f71;
mul.ftz.f32 %f836, %f835, 0f3F000000;
mul.ftz.f32 %f837, %f835, 0f3F4C422A;
mul.ftz.f32 %f838, %f835, 0f3D372713;
fma.rn.ftz.f32 %f839, %f835, %f838, 0f3F800000;
mul.ftz.f32 %f840, %f837, %f839;
tanh.approx.f32 %f841, %f840;
add.ftz.f32 %f842, %f841, 0f3F800000;
mul.ftz.f32 %f843, %f836, %f842;
mul.ftz.f32 %f844, %f834, %f843;
st.global.f32 [%rd27+12], %f844;
bra.uni $L__BB200_78;

$L__BB200_54:
ld.shared.f32 %f357, [%r31];
add.ftz.f32 %f358, %f357, %f30;
ld.shared.f32 %f359, [%r33];
add.ftz.f32 %f360, %f359, %f34;
ld.shared.f32 %f361, [%r31+512];
add.ftz.f32 %f362, %f361, %f358;
ld.shared.f32 %f363, [%r33+512];
add.ftz.f32 %f364, %f363, %f360;
ld.shared.f32 %f365, [%r31+1024];
add.ftz.f32 %f366, %f365, %f362;
ld.shared.f32 %f367, [%r33+1024];
add.ftz.f32 %f368, %f367, %f364;
mov.b32 %r1728, %f366;
mov.u32 %r1729, 31;
mov.u32 %r1730, 16;
mov.u32 %r1731, -1;
shfl.sync.bfly.b32 %r1732|%p111, %r1728, %r1730, %r1729, %r1731;
mov.b32 %f369, %r1732;
add.ftz.f32 %f370, %f366, %f369;
mov.b32 %r1733, %f370;
mov.u32 %r1734, 8;
shfl.sync.bfly.b32 %r1735|%p112, %r1733, %r1734, %r1729, %r1731;
mov.b32 %f371, %r1735;
add.ftz.f32 %f372, %f370, %f371;
mov.b32 %r1736, %f372;
mov.u32 %r1737, 4;
shfl.sync.bfly.b32 %r1738|%p113, %r1736, %r1737, %r1729, %r1731;
mov.b32 %f373, %r1738;
add.ftz.f32 %f374, %f372, %f373;
mov.b32 %r1739, %f374;
mov.u32 %r1740, 2;
shfl.sync.bfly.b32 %r1741|%p114, %r1739, %r1740, %r1729, %r1731;
mov.b32 %f375, %r1741;
add.ftz.f32 %f376, %f374, %f375;
mov.b32 %r1742, %f376;
mov.u32 %r1743, 1;
shfl.sync.bfly.b32 %r1744|%p115, %r1742, %r1743, %r1729, %r1731;
mov.b32 %f377, %r1744;
add.ftz.f32 %f48, %f376, %f377;
mov.b32 %r1745, %f368;
shfl.sync.bfly.b32 %r1746|%p116, %r1745, %r1730, %r1729, %r1731;
mov.b32 %f378, %r1746;
add.ftz.f32 %f379, %f368, %f378;
mov.b32 %r1747, %f379;
shfl.sync.bfly.b32 %r1748|%p117, %r1747, %r1734, %r1729, %r1731;
mov.b32 %f380, %r1748;
add.ftz.f32 %f381, %f379, %f380;
mov.b32 %r1749, %f381;
shfl.sync.bfly.b32 %r1750|%p118, %r1749, %r1737, %r1729, %r1731;
mov.b32 %f382, %r1750;
add.ftz.f32 %f383, %f381, %f382;
mov.b32 %r1751, %f383;
shfl.sync.bfly.b32 %r1752|%p119, %r1751, %r1740, %r1729, %r1731;
mov.b32 %f384, %r1752;
add.ftz.f32 %f385, %f383, %f384;
mov.b32 %r1753, %f385;
shfl.sync.bfly.b32 %r1754|%p120, %r1753, %r1743, %r1729, %r1731;
mov.b32 %f386, %r1754;
add.ftz.f32 %f49, %f385, %f386;
setp.ne.s32 %p121, %r2, 0;
setp.ge.u32 %p122, %r4, %r50;
or.pred %p123, %p121, %p122;
@%p123 bra $L__BB200_56;

add.ftz.f32 %f387, %f5, %f48;
add.ftz.f32 %f388, %f4, %f49;
mov.f32 %f389, 0f40E00000;
min.ftz.f32 %f390, %f388, %f389;
min.ftz.f32 %f391, %f387, %f389;
mov.f32 %f392, 0fC0E00000;
max.ftz.f32 %f393, %f391, %f392;
mul.ftz.f32 %f394, %f390, 0f3FD9DB23;
mul.ftz.f32 %f395, %f394, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f396, %f395;
add.ftz.f32 %f397, %f396, 0f3F800000;
div.approx.ftz.f32 %f398, %f390, %f397;
add.ftz.f32 %f399, %f393, 0f3F800000;
mul.ftz.f32 %f400, %f399, %f398;
st.global.f32 [%rd27], %f400;

$L__BB200_56:
ld.shared.f32 %f401, [%r31+128];
add.ftz.f32 %f402, %f401, %f29;
ld.shared.f32 %f403, [%r33+128];
add.ftz.f32 %f404, %f403, %f33;
ld.shared.f32 %f405, [%r31+640];
add.ftz.f32 %f406, %f405, %f402;
ld.shared.f32 %f407, [%r33+640];
add.ftz.f32 %f408, %f407, %f404;
ld.shared.f32 %f409, [%r31+1152];
add.ftz.f32 %f410, %f409, %f406;
ld.shared.f32 %f411, [%r33+1152];
add.ftz.f32 %f412, %f411, %f408;
mov.b32 %r1756, %f410;
shfl.sync.bfly.b32 %r1760|%p124, %r1756, %r1730, %r1729, %r1731;
mov.b32 %f413, %r1760;
add.ftz.f32 %f414, %f410, %f413;
mov.b32 %r1761, %f414;
shfl.sync.bfly.b32 %r1763|%p125, %r1761, %r1734, %r1729, %r1731;
mov.b32 %f415, %r1763;
add.ftz.f32 %f416, %f414, %f415;
mov.b32 %r1764, %f416;
shfl.sync.bfly.b32 %r1766|%p126, %r1764, %r1737, %r1729, %r1731;
mov.b32 %f417, %r1766;
add.ftz.f32 %f418, %f416, %f417;
mov.b32 %r1767, %f418;
shfl.sync.bfly.b32 %r1769|%p127, %r1767, %r1740, %r1729, %r1731;
mov.b32 %f419, %r1769;
add.ftz.f32 %f420, %f418, %f419;
mov.b32 %r1770, %f420;
shfl.sync.bfly.b32 %r1772|%p128, %r1770, %r1743, %r1729, %r1731;
mov.b32 %f421, %r1772;
add.ftz.f32 %f50, %f420, %f421;
mov.b32 %r1773, %f412;
shfl.sync.bfly.b32 %r1774|%p129, %r1773, %r1730, %r1729, %r1731;
mov.b32 %f422, %r1774;
add.ftz.f32 %f423, %f412, %f422;
mov.b32 %r1775, %f423;
shfl.sync.bfly.b32 %r1776|%p130, %r1775, %r1734, %r1729, %r1731;
mov.b32 %f424, %r1776;
add.ftz.f32 %f425, %f423, %f424;
mov.b32 %r1777, %f425;
shfl.sync.bfly.b32 %r1778|%p131, %r1777, %r1737, %r1729, %r1731;
mov.b32 %f426, %r1778;
add.ftz.f32 %f427, %f425, %f426;
mov.b32 %r1779, %f427;
shfl.sync.bfly.b32 %r1780|%p132, %r1779, %r1740, %r1729, %r1731;
mov.b32 %f428, %r1780;
add.ftz.f32 %f429, %f427, %f428;
mov.b32 %r1781, %f429;
shfl.sync.bfly.b32 %r1782|%p133, %r1781, %r1743, %r1729, %r1731;
mov.b32 %f430, %r1782;
add.ftz.f32 %f51, %f429, %f430;
setp.ne.s32 %p134, %r2, 1;
add.s32 %r1784, %r4, 1;
setp.ge.u32 %p135, %r1784, %r50;
or.pred %p136, %p134, %p135;
@%p136 bra $L__BB200_58;

add.ftz.f32 %f431, %f5, %f50;
add.ftz.f32 %f432, %f4, %f51;
mov.f32 %f433, 0f40E00000;
min.ftz.f32 %f434, %f432, %f433;
min.ftz.f32 %f435, %f431, %f433;
mov.f32 %f436, 0fC0E00000;
max.ftz.f32 %f437, %f435, %f436;
mul.ftz.f32 %f438, %f434, 0f3FD9DB23;
mul.ftz.f32 %f439, %f438, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f440, %f439;
add.ftz.f32 %f441, %f440, 0f3F800000;
div.approx.ftz.f32 %f442, %f434, %f441;
add.ftz.f32 %f443, %f437, 0f3F800000;
mul.ftz.f32 %f444, %f443, %f442;
st.global.f32 [%rd27+4], %f444;

$L__BB200_58:
ld.shared.f32 %f445, [%r31+256];
add.ftz.f32 %f446, %f445, %f28;
ld.shared.f32 %f447, [%r33+256];
add.ftz.f32 %f448, %f447, %f32;
ld.shared.f32 %f449, [%r31+768];
add.ftz.f32 %f450, %f449, %f446;
ld.shared.f32 %f451, [%r33+768];
add.ftz.f32 %f452, %f451, %f448;
ld.shared.f32 %f453, [%r31+1280];
add.ftz.f32 %f454, %f453, %f450;
ld.shared.f32 %f455, [%r33+1280];
add.ftz.f32 %f456, %f455, %f452;
mov.b32 %r1785, %f454;
mov.u32 %r1786, 31;
mov.u32 %r1787, 16;
mov.u32 %r1788, -1;
shfl.sync.bfly.b32 %r1789|%p137, %r1785, %r1787, %r1786, %r1788;
mov.b32 %f457, %r1789;
add.ftz.f32 %f458, %f454, %f457;
mov.b32 %r1790, %f458;
mov.u32 %r1791, 8;
shfl.sync.bfly.b32 %r1792|%p138, %r1790, %r1791, %r1786, %r1788;
mov.b32 %f459, %r1792;
add.ftz.f32 %f460, %f458, %f459;
mov.b32 %r1793, %f460;
mov.u32 %r1794, 4;
shfl.sync.bfly.b32 %r1795|%p139, %r1793, %r1794, %r1786, %r1788;
mov.b32 %f461, %r1795;
add.ftz.f32 %f462, %f460, %f461;
mov.b32 %r1796, %f462;
mov.u32 %r1797, 2;
shfl.sync.bfly.b32 %r1798|%p140, %r1796, %r1797, %r1786, %r1788;
mov.b32 %f463, %r1798;
add.ftz.f32 %f464, %f462, %f463;
mov.b32 %r1799, %f464;
mov.u32 %r1800, 1;
shfl.sync.bfly.b32 %r1801|%p141, %r1799, %r1800, %r1786, %r1788;
mov.b32 %f465, %r1801;
add.ftz.f32 %f52, %f464, %f465;
mov.b32 %r1802, %f456;
shfl.sync.bfly.b32 %r1803|%p142, %r1802, %r1787, %r1786, %r1788;
mov.b32 %f466, %r1803;
add.ftz.f32 %f467, %f456, %f466;
mov.b32 %r1804, %f467;
shfl.sync.bfly.b32 %r1805|%p143, %r1804, %r1791, %r1786, %r1788;
mov.b32 %f468, %r1805;
add.ftz.f32 %f469, %f467, %f468;
mov.b32 %r1806, %f469;
shfl.sync.bfly.b32 %r1807|%p144, %r1806, %r1794, %r1786, %r1788;
mov.b32 %f470, %r1807;
add.ftz.f32 %f471, %f469, %f470;
mov.b32 %r1808, %f471;
shfl.sync.bfly.b32 %r1809|%p145, %r1808, %r1797, %r1786, %r1788;
mov.b32 %f472, %r1809;
add.ftz.f32 %f473, %f471, %f472;
mov.b32 %r1810, %f473;
shfl.sync.bfly.b32 %r1811|%p146, %r1810, %r1800, %r1786, %r1788;
mov.b32 %f474, %r1811;
add.ftz.f32 %f53, %f473, %f474;
setp.ne.s32 %p147, %r2, 2;
add.s32 %r1813, %r4, 2;
setp.ge.u32 %p148, %r1813, %r50;
or.pred %p149, %p147, %p148;
@%p149 bra $L__BB200_60;

add.ftz.f32 %f475, %f5, %f52;
add.ftz.f32 %f476, %f4, %f53;
mov.f32 %f477, 0f40E00000;
min.ftz.f32 %f478, %f476, %f477;
min.ftz.f32 %f479, %f475, %f477;
mov.f32 %f480, 0fC0E00000;
max.ftz.f32 %f481, %f479, %f480;
mul.ftz.f32 %f482, %f478, 0f3FD9DB23;
mul.ftz.f32 %f483, %f482, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f484, %f483;
add.ftz.f32 %f485, %f484, 0f3F800000;
div.approx.ftz.f32 %f486, %f478, %f485;
add.ftz.f32 %f487, %f481, 0f3F800000;
mul.ftz.f32 %f488, %f487, %f486;
st.global.f32 [%rd27+8], %f488;

$L__BB200_60:
ld.shared.f32 %f489, [%r31+384];
add.ftz.f32 %f490, %f489, %f27;
ld.shared.f32 %f491, [%r33+384];
add.ftz.f32 %f492, %f491, %f31;
ld.shared.f32 %f493, [%r31+896];
add.ftz.f32 %f494, %f493, %f490;
ld.shared.f32 %f495, [%r33+896];
add.ftz.f32 %f496, %f495, %f492;
ld.shared.f32 %f497, [%r31+1408];
add.ftz.f32 %f498, %f497, %f494;
ld.shared.f32 %f499, [%r33+1408];
add.ftz.f32 %f500, %f499, %f496;
mov.b32 %r1814, %f498;
shfl.sync.bfly.b32 %r1818|%p150, %r1814, %r1787, %r1786, %r1788;
mov.b32 %f501, %r1818;
add.ftz.f32 %f502, %f498, %f501;
mov.b32 %r1819, %f502;
shfl.sync.bfly.b32 %r1821|%p151, %r1819, %r1791, %r1786, %r1788;
mov.b32 %f503, %r1821;
add.ftz.f32 %f504, %f502, %f503;
mov.b32 %r1822, %f504;
shfl.sync.bfly.b32 %r1824|%p152, %r1822, %r1794, %r1786, %r1788;
mov.b32 %f505, %r1824;
add.ftz.f32 %f506, %f504, %f505;
mov.b32 %r1825, %f506;
shfl.sync.bfly.b32 %r1827|%p153, %r1825, %r1797, %r1786, %r1788;
mov.b32 %f507, %r1827;
add.ftz.f32 %f508, %f506, %f507;
mov.b32 %r1828, %f508;
shfl.sync.bfly.b32 %r1830|%p154, %r1828, %r1800, %r1786, %r1788;
mov.b32 %f509, %r1830;
add.ftz.f32 %f54, %f508, %f509;
mov.b32 %r1831, %f500;
shfl.sync.bfly.b32 %r1832|%p155, %r1831, %r1787, %r1786, %r1788;
mov.b32 %f510, %r1832;
add.ftz.f32 %f511, %f500, %f510;
mov.b32 %r1833, %f511;
shfl.sync.bfly.b32 %r1834|%p156, %r1833, %r1791, %r1786, %r1788;
mov.b32 %f512, %r1834;
add.ftz.f32 %f513, %f511, %f512;
mov.b32 %r1835, %f513;
shfl.sync.bfly.b32 %r1836|%p157, %r1835, %r1794, %r1786, %r1788;
mov.b32 %f514, %r1836;
add.ftz.f32 %f515, %f513, %f514;
mov.b32 %r1837, %f515;
shfl.sync.bfly.b32 %r1838|%p158, %r1837, %r1797, %r1786, %r1788;
mov.b32 %f516, %r1838;
add.ftz.f32 %f517, %f515, %f516;
mov.b32 %r1839, %f517;
shfl.sync.bfly.b32 %r1840|%p159, %r1839, %r1800, %r1786, %r1788;
mov.b32 %f518, %r1840;
add.ftz.f32 %f55, %f517, %f518;
setp.ne.s32 %p160, %r2, 3;
add.s32 %r1842, %r4, 3;
setp.ge.u32 %p161, %r1842, %r50;
or.pred %p162, %p160, %p161;
@%p162 bra $L__BB200_78;

add.ftz.f32 %f519, %f5, %f54;
add.ftz.f32 %f520, %f4, %f55;
mov.f32 %f521, 0f40E00000;
min.ftz.f32 %f522, %f520, %f521;
min.ftz.f32 %f523, %f519, %f521;
mov.f32 %f524, 0fC0E00000;
max.ftz.f32 %f525, %f523, %f524;
mul.ftz.f32 %f526, %f522, 0f3FD9DB23;
mul.ftz.f32 %f527, %f526, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f528, %f527;
add.ftz.f32 %f529, %f528, 0f3F800000;
div.approx.ftz.f32 %f530, %f522, %f529;
add.ftz.f32 %f531, %f525, 0f3F800000;
mul.ftz.f32 %f532, %f531, %f530;
st.global.f32 [%rd27+12], %f532;
bra.uni $L__BB200_78;

}
.entry _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<38>;
.reg .b16 %rs<37>;
.reg .f32 %f<109>;
.reg .b32 %r<853>;
.reg .b64 %rd<80>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

ld.param.u64 %rd11, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd12, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd13, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd14, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
mov.u32 %r1, %tid.y;
shl.b32 %r39, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r39, %r2;
mov.u32 %r40, %ctaid.x;
shl.b32 %r4, %r40, 2;
shr.u32 %r5, %r38, 8;
mov.u32 %r850, %ctaid.y;
setp.eq.s64 %p1, %rd13, 0;
@%p1 bra $L__BB201_2;

cvta.to.global.u64 %rd15, %rd13;
mul.wide.u32 %rd16, %r850, 4;
add.s64 %rd17, %rd15, %rd16;
ld.global.nc.u32 %r849, [%rd17];
mul.hi.u32 %r41, %r850, %r21;
add.s32 %r42, %r41, %r850;
shr.u32 %r43, %r42, %r22;
mul.lo.s32 %r44, %r43, %r23;
sub.s32 %r850, %r850, %r44;
bra.uni $L__BB201_3;

$L__BB201_2:
mul.hi.u32 %r45, %r850, %r26;
add.s32 %r46, %r45, %r850;
shr.u32 %r849, %r46, %r27;

$L__BB201_3:
mov.u32 %r12, %ctaid.z;
shr.s32 %r47, %r3, 31;
shr.u32 %r48, %r47, 29;
add.s32 %r49, %r3, %r48;
shr.s32 %r852, %r49, 3;
setp.ge.s32 %p2, %r852, %r5;
mov.f32 %f105, 0f00000000;
mov.f32 %f106, %f105;
mov.f32 %f107, %f105;
mov.f32 %f108, %f105;
@%p2 bra $L__BB201_6;

and.b32 %r53, %r49, -8;
sub.s32 %r54, %r3, %r53;
shl.b32 %r55, %r54, 2;
cvt.s64.s32 %rd1, %r54;
shl.b32 %r851, %r852, 3;
mad.lo.s32 %r56, %r4, %r24, %r852;
mul.hi.u32 %r57, %r12, %r32;
add.s32 %r58, %r57, %r12;
shr.u32 %r59, %r58, %r33;
mad.lo.s32 %r60, %r59, %r35, %r56;
mad.lo.s32 %r61, %r849, %r29, %r60;
mul.wide.s32 %rd2, %r61, 66;
mul.wide.s32 %rd18, %r55, 2;
add.s64 %rd19, %rd2, %rd18;
add.s64 %rd3, %rd19, 8;
mul.wide.s32 %rd4, %r24, 66;
cvta.to.global.u64 %rd79, %rd11;
cvta.to.global.u64 %rd6, %rd12;
mul.lo.s32 %r62, %r850, %r30;
cvt.u64.u32 %rd20, %r62;
mul.lo.s32 %r63, %r12, %r36;
cvt.u64.u32 %rd21, %r63;
add.s64 %rd7, %rd20, %rd21;
mov.f32 %f105, 0f00000000;

$L__BB201_5:
cvt.u64.u32 %rd22, %r851;
add.s64 %rd23, %rd7, %rd22;
add.s64 %rd24, %rd23, %rd1;
mul.lo.s64 %rd25, %rd24, 36;
add.s64 %rd26, %rd79, %rd3;
add.s64 %rd27, %rd26, -6;
ld.global.nc.u16 %rs5, [%rd26+-6];
shr.u16 %rs6, %rs5, 8;
ld.global.nc.u16 %rs7, [%rd26+-4];
shr.u16 %rs8, %rs7, 8;
ld.global.nc.u16 %rs9, [%rd26+-2];
cvt.u32.u16 %r260, %rs9;
ld.global.nc.u16 %rs10, [%rd26];
cvt.u32.u16 %r261, %rs10;
mov.b32 %r262, {%rs9, %rs10};
shl.b16 %rs11, %rs5, 3;
cvt.u64.u16 %rd28, %rs11;
and.b64 %rd29, %rd28, 2040;
mov.u64 %rd30, iq2xxs_grid;
add.s64 %rd31, %rd30, %rd29;
ld.global.nc.v2.u32 {%r263, %r264}, [%rd31];
mov.u32 %r216, 0;
and.b32 %r267, %r260, 255;
popc.b32 %r268, %r267;
shl.b32 %r269, %r268, 7;
and.b32 %r270, %r269, 128;
xor.b32 %r271, %r270, %r267;
mul.lo.s32 %r272, %r271, 16843009;
and.b32 %r273, %r272, 134480385;
or.b32 %r274, %r273, -2139062144;
add.s32 %r275, %r274, -16843009;
or.b32 %r64, %r275, %r273;

prmt.b32 %r64, %r64, 0, 0xba98;

xor.b32 %r276, %r263, %r64;
or.b32 %r277, %r276, -2139062144;
and.b32 %r278, %r64, 2139062143;
and.b32 %r279, %r263, -2139062144;
sub.s32 %r280, %r277, %r278;
xor.b32 %r281, %r279, %r280;
xor.b32 %r67, %r281, -2139062144;
add.s64 %rd32, %rd6, %rd25;
ld.global.nc.u32 %r68, [%rd32+4];

dp4a.s32.s32 %r66, %r67, %r68, %r216;

and.b32 %r282, %r272, -2143281136;
or.b32 %r283, %r282, -2139062144;
add.s32 %r284, %r283, -16843009;
or.b32 %r70, %r284, %r282;

prmt.b32 %r70, %r70, 0, 0xba98;

xor.b32 %r285, %r70, %r264;
or.b32 %r286, %r285, -2139062144;
and.b32 %r287, %r70, 2139062143;
and.b32 %r288, %r264, -2139062144;
sub.s32 %r289, %r286, %r287;
xor.b32 %r290, %r288, %r289;
xor.b32 %r73, %r290, -2139062144;
ld.global.nc.u32 %r74, [%rd32+8];

dp4a.s32.s32 %r72, %r73, %r74, %r66;

cvt.u32.u16 %r291, %rs6;
mul.wide.u32 %rd33, %r291, 8;
add.s64 %rd34, %rd30, %rd33;
ld.global.nc.v2.u32 {%r292, %r293}, [%rd34];
shr.u32 %r296, %r260, 7;
and.b32 %r297, %r296, 255;
popc.b32 %r298, %r297;
shl.b32 %r299, %r298, 7;
and.b32 %r300, %r299, 128;
xor.b32 %r301, %r300, %r297;
mul.lo.s32 %r302, %r301, 16843009;
and.b32 %r303, %r302, 134480385;
or.b32 %r304, %r303, -2139062144;
add.s32 %r305, %r304, -16843009;
or.b32 %r76, %r305, %r303;

prmt.b32 %r76, %r76, 0, 0xba98;

xor.b32 %r306, %r292, %r76;
or.b32 %r307, %r306, -2139062144;
and.b32 %r308, %r76, 2139062143;
and.b32 %r309, %r292, -2139062144;
sub.s32 %r310, %r307, %r308;
xor.b32 %r311, %r309, %r310;
xor.b32 %r79, %r311, -2139062144;
ld.global.nc.u32 %r80, [%rd32+12];

dp4a.s32.s32 %r78, %r79, %r80, %r72;

and.b32 %r312, %r302, -2143281136;
or.b32 %r313, %r312, -2139062144;
add.s32 %r314, %r313, -16843009;
or.b32 %r82, %r314, %r312;

prmt.b32 %r82, %r82, 0, 0xba98;

xor.b32 %r315, %r82, %r293;
or.b32 %r316, %r315, -2139062144;
and.b32 %r317, %r82, 2139062143;
and.b32 %r318, %r293, -2139062144;
sub.s32 %r319, %r316, %r317;
xor.b32 %r320, %r318, %r319;
xor.b32 %r85, %r320, -2139062144;
ld.global.nc.u32 %r86, [%rd32+16];

dp4a.s32.s32 %r84, %r85, %r86, %r78;

shl.b16 %rs12, %rs7, 3;
cvt.u64.u16 %rd35, %rs12;
and.b64 %rd36, %rd35, 2040;
add.s64 %rd37, %rd30, %rd36;
ld.global.nc.v2.u32 {%r321, %r322}, [%rd37];
shr.u32 %r325, %r262, 14;
and.b32 %r326, %r325, 255;
popc.b32 %r327, %r326;
shl.b32 %r328, %r327, 7;
and.b32 %r329, %r328, 128;
xor.b32 %r330, %r329, %r326;
mul.lo.s32 %r331, %r330, 16843009;
and.b32 %r332, %r331, 134480385;
or.b32 %r333, %r332, -2139062144;
add.s32 %r334, %r333, -16843009;
or.b32 %r88, %r334, %r332;

prmt.b32 %r88, %r88, 0, 0xba98;

xor.b32 %r335, %r321, %r88;
or.b32 %r336, %r335, -2139062144;
and.b32 %r337, %r88, 2139062143;
and.b32 %r338, %r321, -2139062144;
sub.s32 %r339, %r336, %r337;
xor.b32 %r340, %r338, %r339;
xor.b32 %r91, %r340, -2139062144;
ld.global.nc.u32 %r92, [%rd32+20];

dp4a.s32.s32 %r90, %r91, %r92, %r84;

and.b32 %r341, %r331, -2143281136;
or.b32 %r342, %r341, -2139062144;
add.s32 %r343, %r342, -16843009;
or.b32 %r94, %r343, %r341;

prmt.b32 %r94, %r94, 0, 0xba98;

xor.b32 %r344, %r94, %r322;
or.b32 %r345, %r344, -2139062144;
and.b32 %r346, %r94, 2139062143;
and.b32 %r347, %r322, -2139062144;
sub.s32 %r348, %r345, %r346;
xor.b32 %r349, %r347, %r348;
xor.b32 %r97, %r349, -2139062144;
ld.global.nc.u32 %r98, [%rd32+24];

dp4a.s32.s32 %r96, %r97, %r98, %r90;

cvt.u32.u16 %r350, %rs8;
mul.wide.u32 %rd38, %r350, 8;
add.s64 %rd39, %rd30, %rd38;
ld.global.nc.v2.u32 {%r351, %r352}, [%rd39];
shr.u32 %r355, %r261, 5;
and.b32 %r356, %r355, 255;
popc.b32 %r357, %r356;
shl.b32 %r358, %r357, 7;
and.b32 %r359, %r358, 128;
xor.b32 %r360, %r359, %r356;
mul.lo.s32 %r361, %r360, 16843009;
and.b32 %r362, %r361, 134480385;
or.b32 %r363, %r362, -2139062144;
add.s32 %r364, %r363, -16843009;
or.b32 %r100, %r364, %r362;

prmt.b32 %r100, %r100, 0, 0xba98;

xor.b32 %r365, %r351, %r100;
or.b32 %r366, %r365, -2139062144;
and.b32 %r367, %r100, 2139062143;
and.b32 %r368, %r351, -2139062144;
sub.s32 %r369, %r366, %r367;
xor.b32 %r370, %r368, %r369;
xor.b32 %r103, %r370, -2139062144;
ld.global.nc.u32 %r104, [%rd32+28];

dp4a.s32.s32 %r102, %r103, %r104, %r96;

and.b32 %r371, %r361, -2143281136;
or.b32 %r372, %r371, -2139062144;
add.s32 %r373, %r372, -16843009;
or.b32 %r106, %r373, %r371;

prmt.b32 %r106, %r106, 0, 0xba98;

xor.b32 %r374, %r106, %r352;
or.b32 %r375, %r374, -2139062144;
and.b32 %r376, %r106, 2139062143;
and.b32 %r377, %r352, -2139062144;
sub.s32 %r378, %r375, %r376;
xor.b32 %r379, %r377, %r378;
xor.b32 %r109, %r379, -2139062144;
ld.global.nc.u32 %r110, [%rd32+32];

dp4a.s32.s32 %r108, %r109, %r110, %r102;

shr.u32 %r380, %r261, 11;
or.b32 %r381, %r380, 1;
mul.lo.s32 %r382, %r108, %r381;
shr.s32 %r383, %r382, 31;
shr.u32 %r384, %r383, 29;
add.s32 %r385, %r382, %r384;
shr.s32 %r386, %r385, 3;
add.s64 %rd40, %rd79, %rd2;
ld.global.nc.u16 %rs1, [%rd40];

{ cvt.f32.f16 %f25, %rs1;}


ld.global.nc.u32 %r112, [%rd32];

{.reg .f16 low,high;
mov.b32 {low,high},%r112;
cvt.f32.f16 %f26, low;}


mul.ftz.f32 %f33, %f25, %f26;
cvt.rn.f32.s32 %f34, %r386;
fma.rn.ftz.f32 %f108, %f33, %f34, %f108;
add.s64 %rd41, %rd27, %rd4;
ld.global.nc.u16 %rs13, [%rd41];
shr.u16 %rs14, %rs13, 8;
ld.global.nc.u16 %rs15, [%rd41+2];
shr.u16 %rs16, %rs15, 8;
ld.global.nc.u16 %rs17, [%rd41+4];
cvt.u32.u16 %r387, %rs17;
ld.global.nc.u16 %rs18, [%rd41+6];
cvt.u32.u16 %r388, %rs18;
mov.b32 %r389, {%rs17, %rs18};
shl.b16 %rs19, %rs13, 3;
cvt.u64.u16 %rd42, %rs19;
and.b64 %rd43, %rd42, 2040;
add.s64 %rd44, %rd30, %rd43;
ld.global.nc.v2.u32 {%r390, %r391}, [%rd44];
and.b32 %r394, %r387, 255;
popc.b32 %r395, %r394;
shl.b32 %r396, %r395, 7;
and.b32 %r397, %r396, 128;
xor.b32 %r398, %r397, %r394;
mul.lo.s32 %r399, %r398, 16843009;
and.b32 %r400, %r399, 134480385;
or.b32 %r401, %r400, -2139062144;
add.s32 %r402, %r401, -16843009;
or.b32 %r113, %r402, %r400;

prmt.b32 %r113, %r113, 0, 0xba98;

xor.b32 %r403, %r390, %r113;
or.b32 %r404, %r403, -2139062144;
and.b32 %r405, %r113, 2139062143;
and.b32 %r406, %r390, -2139062144;
sub.s32 %r407, %r404, %r405;
xor.b32 %r408, %r406, %r407;
xor.b32 %r116, %r408, -2139062144;

dp4a.s32.s32 %r115, %r116, %r68, %r216;

and.b32 %r409, %r399, -2143281136;
or.b32 %r410, %r409, -2139062144;
add.s32 %r411, %r410, -16843009;
or.b32 %r119, %r411, %r409;

prmt.b32 %r119, %r119, 0, 0xba98;

xor.b32 %r412, %r119, %r391;
or.b32 %r413, %r412, -2139062144;
and.b32 %r414, %r119, 2139062143;
and.b32 %r415, %r391, -2139062144;
sub.s32 %r416, %r413, %r414;
xor.b32 %r417, %r415, %r416;
xor.b32 %r122, %r417, -2139062144;

dp4a.s32.s32 %r121, %r122, %r74, %r115;

cvt.u32.u16 %r418, %rs14;
mul.wide.u32 %rd45, %r418, 8;
add.s64 %rd46, %rd30, %rd45;
ld.global.nc.v2.u32 {%r419, %r420}, [%rd46];
shr.u32 %r423, %r387, 7;
and.b32 %r424, %r423, 255;
popc.b32 %r425, %r424;
shl.b32 %r426, %r425, 7;
and.b32 %r427, %r426, 128;
xor.b32 %r428, %r427, %r424;
mul.lo.s32 %r429, %r428, 16843009;
and.b32 %r430, %r429, 134480385;
or.b32 %r431, %r430, -2139062144;
add.s32 %r432, %r431, -16843009;
or.b32 %r125, %r432, %r430;

prmt.b32 %r125, %r125, 0, 0xba98;

xor.b32 %r433, %r419, %r125;
or.b32 %r434, %r433, -2139062144;
and.b32 %r435, %r125, 2139062143;
and.b32 %r436, %r419, -2139062144;
sub.s32 %r437, %r434, %r435;
xor.b32 %r438, %r436, %r437;
xor.b32 %r128, %r438, -2139062144;

dp4a.s32.s32 %r127, %r128, %r80, %r121;

and.b32 %r439, %r429, -2143281136;
or.b32 %r440, %r439, -2139062144;
add.s32 %r441, %r440, -16843009;
or.b32 %r131, %r441, %r439;

prmt.b32 %r131, %r131, 0, 0xba98;

xor.b32 %r442, %r131, %r420;
or.b32 %r443, %r442, -2139062144;
and.b32 %r444, %r131, 2139062143;
and.b32 %r445, %r420, -2139062144;
sub.s32 %r446, %r443, %r444;
xor.b32 %r447, %r445, %r446;
xor.b32 %r134, %r447, -2139062144;

dp4a.s32.s32 %r133, %r134, %r86, %r127;

shl.b16 %rs20, %rs15, 3;
cvt.u64.u16 %rd47, %rs20;
and.b64 %rd48, %rd47, 2040;
add.s64 %rd49, %rd30, %rd48;
ld.global.nc.v2.u32 {%r448, %r449}, [%rd49];
shr.u32 %r452, %r389, 14;
and.b32 %r453, %r452, 255;
popc.b32 %r454, %r453;
shl.b32 %r455, %r454, 7;
and.b32 %r456, %r455, 128;
xor.b32 %r457, %r456, %r453;
mul.lo.s32 %r458, %r457, 16843009;
and.b32 %r459, %r458, 134480385;
or.b32 %r460, %r459, -2139062144;
add.s32 %r461, %r460, -16843009;
or.b32 %r137, %r461, %r459;

prmt.b32 %r137, %r137, 0, 0xba98;

xor.b32 %r462, %r448, %r137;
or.b32 %r463, %r462, -2139062144;
and.b32 %r464, %r137, 2139062143;
and.b32 %r465, %r448, -2139062144;
sub.s32 %r466, %r463, %r464;
xor.b32 %r467, %r465, %r466;
xor.b32 %r140, %r467, -2139062144;

dp4a.s32.s32 %r139, %r140, %r92, %r133;

and.b32 %r468, %r458, -2143281136;
or.b32 %r469, %r468, -2139062144;
add.s32 %r470, %r469, -16843009;
or.b32 %r143, %r470, %r468;

prmt.b32 %r143, %r143, 0, 0xba98;

xor.b32 %r471, %r143, %r449;
or.b32 %r472, %r471, -2139062144;
and.b32 %r473, %r143, 2139062143;
and.b32 %r474, %r449, -2139062144;
sub.s32 %r475, %r472, %r473;
xor.b32 %r476, %r474, %r475;
xor.b32 %r146, %r476, -2139062144;

dp4a.s32.s32 %r145, %r146, %r98, %r139;

cvt.u32.u16 %r477, %rs16;
mul.wide.u32 %rd50, %r477, 8;
add.s64 %rd51, %rd30, %rd50;
ld.global.nc.v2.u32 {%r478, %r479}, [%rd51];
shr.u32 %r482, %r388, 5;
and.b32 %r483, %r482, 255;
popc.b32 %r484, %r483;
shl.b32 %r485, %r484, 7;
and.b32 %r486, %r485, 128;
xor.b32 %r487, %r486, %r483;
mul.lo.s32 %r488, %r487, 16843009;
and.b32 %r489, %r488, 134480385;
or.b32 %r490, %r489, -2139062144;
add.s32 %r491, %r490, -16843009;
or.b32 %r149, %r491, %r489;

prmt.b32 %r149, %r149, 0, 0xba98;

xor.b32 %r492, %r478, %r149;
or.b32 %r493, %r492, -2139062144;
and.b32 %r494, %r149, 2139062143;
and.b32 %r495, %r478, -2139062144;
sub.s32 %r496, %r493, %r494;
xor.b32 %r497, %r495, %r496;
xor.b32 %r152, %r497, -2139062144;

dp4a.s32.s32 %r151, %r152, %r104, %r145;

and.b32 %r498, %r488, -2143281136;
or.b32 %r499, %r498, -2139062144;
add.s32 %r500, %r499, -16843009;
or.b32 %r155, %r500, %r498;

prmt.b32 %r155, %r155, 0, 0xba98;

xor.b32 %r501, %r155, %r479;
or.b32 %r502, %r501, -2139062144;
and.b32 %r503, %r155, 2139062143;
and.b32 %r504, %r479, -2139062144;
sub.s32 %r505, %r502, %r503;
xor.b32 %r506, %r504, %r505;
xor.b32 %r158, %r506, -2139062144;

dp4a.s32.s32 %r157, %r158, %r110, %r151;

shr.u32 %r507, %r388, 11;
or.b32 %r508, %r507, 1;
mul.lo.s32 %r509, %r157, %r508;
shr.s32 %r510, %r509, 31;
shr.u32 %r511, %r510, 29;
add.s32 %r512, %r509, %r511;
shr.s32 %r513, %r512, 3;
add.s64 %rd52, %rd40, %rd4;
ld.global.nc.u16 %rs2, [%rd52];

{ cvt.f32.f16 %f27, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r112;
cvt.f32.f16 %f28, low;}


mul.ftz.f32 %f35, %f27, %f28;
cvt.rn.f32.s32 %f36, %r513;
fma.rn.ftz.f32 %f107, %f35, %f36, %f107;
add.s64 %rd53, %rd41, %rd4;
ld.global.nc.u16 %rs21, [%rd53];
shr.u16 %rs22, %rs21, 8;
ld.global.nc.u16 %rs23, [%rd53+2];
shr.u16 %rs24, %rs23, 8;
ld.global.nc.u16 %rs25, [%rd53+4];
cvt.u32.u16 %r514, %rs25;
ld.global.nc.u16 %rs26, [%rd53+6];
cvt.u32.u16 %r515, %rs26;
mov.b32 %r516, {%rs25, %rs26};
shl.b16 %rs27, %rs21, 3;
cvt.u64.u16 %rd54, %rs27;
and.b64 %rd55, %rd54, 2040;
add.s64 %rd56, %rd30, %rd55;
ld.global.nc.v2.u32 {%r517, %r518}, [%rd56];
and.b32 %r521, %r514, 255;
popc.b32 %r522, %r521;
shl.b32 %r523, %r522, 7;
and.b32 %r524, %r523, 128;
xor.b32 %r525, %r524, %r521;
mul.lo.s32 %r526, %r525, 16843009;
and.b32 %r527, %r526, 134480385;
or.b32 %r528, %r527, -2139062144;
add.s32 %r529, %r528, -16843009;
or.b32 %r162, %r529, %r527;

prmt.b32 %r162, %r162, 0, 0xba98;

xor.b32 %r530, %r517, %r162;
or.b32 %r531, %r530, -2139062144;
and.b32 %r532, %r162, 2139062143;
and.b32 %r533, %r517, -2139062144;
sub.s32 %r534, %r531, %r532;
xor.b32 %r535, %r533, %r534;
xor.b32 %r165, %r535, -2139062144;

dp4a.s32.s32 %r164, %r165, %r68, %r216;

and.b32 %r536, %r526, -2143281136;
or.b32 %r537, %r536, -2139062144;
add.s32 %r538, %r537, -16843009;
or.b32 %r168, %r538, %r536;

prmt.b32 %r168, %r168, 0, 0xba98;

xor.b32 %r539, %r168, %r518;
or.b32 %r540, %r539, -2139062144;
and.b32 %r541, %r168, 2139062143;
and.b32 %r542, %r518, -2139062144;
sub.s32 %r543, %r540, %r541;
xor.b32 %r544, %r542, %r543;
xor.b32 %r171, %r544, -2139062144;

dp4a.s32.s32 %r170, %r171, %r74, %r164;

cvt.u32.u16 %r545, %rs22;
mul.wide.u32 %rd57, %r545, 8;
add.s64 %rd58, %rd30, %rd57;
ld.global.nc.v2.u32 {%r546, %r547}, [%rd58];
shr.u32 %r550, %r514, 7;
and.b32 %r551, %r550, 255;
popc.b32 %r552, %r551;
shl.b32 %r553, %r552, 7;
and.b32 %r554, %r553, 128;
xor.b32 %r555, %r554, %r551;
mul.lo.s32 %r556, %r555, 16843009;
and.b32 %r557, %r556, 134480385;
or.b32 %r558, %r557, -2139062144;
add.s32 %r559, %r558, -16843009;
or.b32 %r174, %r559, %r557;

prmt.b32 %r174, %r174, 0, 0xba98;

xor.b32 %r560, %r546, %r174;
or.b32 %r561, %r560, -2139062144;
and.b32 %r562, %r174, 2139062143;
and.b32 %r563, %r546, -2139062144;
sub.s32 %r564, %r561, %r562;
xor.b32 %r565, %r563, %r564;
xor.b32 %r177, %r565, -2139062144;

dp4a.s32.s32 %r176, %r177, %r80, %r170;

and.b32 %r566, %r556, -2143281136;
or.b32 %r567, %r566, -2139062144;
add.s32 %r568, %r567, -16843009;
or.b32 %r180, %r568, %r566;

prmt.b32 %r180, %r180, 0, 0xba98;

xor.b32 %r569, %r180, %r547;
or.b32 %r570, %r569, -2139062144;
and.b32 %r571, %r180, 2139062143;
and.b32 %r572, %r547, -2139062144;
sub.s32 %r573, %r570, %r571;
xor.b32 %r574, %r572, %r573;
xor.b32 %r183, %r574, -2139062144;

dp4a.s32.s32 %r182, %r183, %r86, %r176;

shl.b16 %rs28, %rs23, 3;
cvt.u64.u16 %rd59, %rs28;
and.b64 %rd60, %rd59, 2040;
add.s64 %rd61, %rd30, %rd60;
ld.global.nc.v2.u32 {%r575, %r576}, [%rd61];
shr.u32 %r579, %r516, 14;
and.b32 %r580, %r579, 255;
popc.b32 %r581, %r580;
shl.b32 %r582, %r581, 7;
and.b32 %r583, %r582, 128;
xor.b32 %r584, %r583, %r580;
mul.lo.s32 %r585, %r584, 16843009;
and.b32 %r586, %r585, 134480385;
or.b32 %r587, %r586, -2139062144;
add.s32 %r588, %r587, -16843009;
or.b32 %r186, %r588, %r586;

prmt.b32 %r186, %r186, 0, 0xba98;

xor.b32 %r589, %r575, %r186;
or.b32 %r590, %r589, -2139062144;
and.b32 %r591, %r186, 2139062143;
and.b32 %r592, %r575, -2139062144;
sub.s32 %r593, %r590, %r591;
xor.b32 %r594, %r592, %r593;
xor.b32 %r189, %r594, -2139062144;

dp4a.s32.s32 %r188, %r189, %r92, %r182;

and.b32 %r595, %r585, -2143281136;
or.b32 %r596, %r595, -2139062144;
add.s32 %r597, %r596, -16843009;
or.b32 %r192, %r597, %r595;

prmt.b32 %r192, %r192, 0, 0xba98;

xor.b32 %r598, %r192, %r576;
or.b32 %r599, %r598, -2139062144;
and.b32 %r600, %r192, 2139062143;
and.b32 %r601, %r576, -2139062144;
sub.s32 %r602, %r599, %r600;
xor.b32 %r603, %r601, %r602;
xor.b32 %r195, %r603, -2139062144;

dp4a.s32.s32 %r194, %r195, %r98, %r188;

cvt.u32.u16 %r604, %rs24;
mul.wide.u32 %rd62, %r604, 8;
add.s64 %rd63, %rd30, %rd62;
ld.global.nc.v2.u32 {%r605, %r606}, [%rd63];
shr.u32 %r609, %r515, 5;
and.b32 %r610, %r609, 255;
popc.b32 %r611, %r610;
shl.b32 %r612, %r611, 7;
and.b32 %r613, %r612, 128;
xor.b32 %r614, %r613, %r610;
mul.lo.s32 %r615, %r614, 16843009;
and.b32 %r616, %r615, 134480385;
or.b32 %r617, %r616, -2139062144;
add.s32 %r618, %r617, -16843009;
or.b32 %r198, %r618, %r616;

prmt.b32 %r198, %r198, 0, 0xba98;

xor.b32 %r619, %r605, %r198;
or.b32 %r620, %r619, -2139062144;
and.b32 %r621, %r198, 2139062143;
and.b32 %r622, %r605, -2139062144;
sub.s32 %r623, %r620, %r621;
xor.b32 %r624, %r622, %r623;
xor.b32 %r201, %r624, -2139062144;

dp4a.s32.s32 %r200, %r201, %r104, %r194;

and.b32 %r625, %r615, -2143281136;
or.b32 %r626, %r625, -2139062144;
add.s32 %r627, %r626, -16843009;
or.b32 %r204, %r627, %r625;

prmt.b32 %r204, %r204, 0, 0xba98;

xor.b32 %r628, %r204, %r606;
or.b32 %r629, %r628, -2139062144;
and.b32 %r630, %r204, 2139062143;
and.b32 %r631, %r606, -2139062144;
sub.s32 %r632, %r629, %r630;
xor.b32 %r633, %r631, %r632;
xor.b32 %r207, %r633, -2139062144;

dp4a.s32.s32 %r206, %r207, %r110, %r200;

shr.u32 %r634, %r515, 11;
or.b32 %r635, %r634, 1;
mul.lo.s32 %r636, %r206, %r635;
shr.s32 %r637, %r636, 31;
shr.u32 %r638, %r637, 29;
add.s32 %r639, %r636, %r638;
shr.s32 %r640, %r639, 3;
add.s64 %rd64, %rd52, %rd4;
ld.global.nc.u16 %rs3, [%rd64];

{ cvt.f32.f16 %f29, %rs3;}



{.reg .f16 low,high;
mov.b32 {low,high},%r112;
cvt.f32.f16 %f30, low;}


mul.ftz.f32 %f37, %f29, %f30;
cvt.rn.f32.s32 %f38, %r640;
fma.rn.ftz.f32 %f106, %f37, %f38, %f106;
add.s64 %rd65, %rd53, %rd4;
ld.global.nc.u16 %rs29, [%rd65];
shr.u16 %rs30, %rs29, 8;
ld.global.nc.u16 %rs31, [%rd65+2];
shr.u16 %rs32, %rs31, 8;
ld.global.nc.u16 %rs33, [%rd65+4];
cvt.u32.u16 %r641, %rs33;
ld.global.nc.u16 %rs34, [%rd65+6];
cvt.u32.u16 %r642, %rs34;
mov.b32 %r643, {%rs33, %rs34};
shl.b16 %rs35, %rs29, 3;
cvt.u64.u16 %rd66, %rs35;
and.b64 %rd67, %rd66, 2040;
add.s64 %rd68, %rd30, %rd67;
ld.global.nc.v2.u32 {%r644, %r645}, [%rd68];
and.b32 %r648, %r641, 255;
popc.b32 %r649, %r648;
shl.b32 %r650, %r649, 7;
and.b32 %r651, %r650, 128;
xor.b32 %r652, %r651, %r648;
mul.lo.s32 %r653, %r652, 16843009;
and.b32 %r654, %r653, 134480385;
or.b32 %r655, %r654, -2139062144;
add.s32 %r656, %r655, -16843009;
or.b32 %r211, %r656, %r654;

prmt.b32 %r211, %r211, 0, 0xba98;

xor.b32 %r657, %r644, %r211;
or.b32 %r658, %r657, -2139062144;
and.b32 %r659, %r211, 2139062143;
and.b32 %r660, %r644, -2139062144;
sub.s32 %r661, %r658, %r659;
xor.b32 %r662, %r660, %r661;
xor.b32 %r214, %r662, -2139062144;

dp4a.s32.s32 %r213, %r214, %r68, %r216;

and.b32 %r663, %r653, -2143281136;
or.b32 %r664, %r663, -2139062144;
add.s32 %r665, %r664, -16843009;
or.b32 %r217, %r665, %r663;

prmt.b32 %r217, %r217, 0, 0xba98;

xor.b32 %r666, %r217, %r645;
or.b32 %r667, %r666, -2139062144;
and.b32 %r668, %r217, 2139062143;
and.b32 %r669, %r645, -2139062144;
sub.s32 %r670, %r667, %r668;
xor.b32 %r671, %r669, %r670;
xor.b32 %r220, %r671, -2139062144;

dp4a.s32.s32 %r219, %r220, %r74, %r213;

cvt.u32.u16 %r672, %rs30;
mul.wide.u32 %rd69, %r672, 8;
add.s64 %rd70, %rd30, %rd69;
ld.global.nc.v2.u32 {%r673, %r674}, [%rd70];
shr.u32 %r677, %r641, 7;
and.b32 %r678, %r677, 255;
popc.b32 %r679, %r678;
shl.b32 %r680, %r679, 7;
and.b32 %r681, %r680, 128;
xor.b32 %r682, %r681, %r678;
mul.lo.s32 %r683, %r682, 16843009;
and.b32 %r684, %r683, 134480385;
or.b32 %r685, %r684, -2139062144;
add.s32 %r686, %r685, -16843009;
or.b32 %r223, %r686, %r684;

prmt.b32 %r223, %r223, 0, 0xba98;

xor.b32 %r687, %r673, %r223;
or.b32 %r688, %r687, -2139062144;
and.b32 %r689, %r223, 2139062143;
and.b32 %r690, %r673, -2139062144;
sub.s32 %r691, %r688, %r689;
xor.b32 %r692, %r690, %r691;
xor.b32 %r226, %r692, -2139062144;

dp4a.s32.s32 %r225, %r226, %r80, %r219;

and.b32 %r693, %r683, -2143281136;
or.b32 %r694, %r693, -2139062144;
add.s32 %r695, %r694, -16843009;
or.b32 %r229, %r695, %r693;

prmt.b32 %r229, %r229, 0, 0xba98;

xor.b32 %r696, %r229, %r674;
or.b32 %r697, %r696, -2139062144;
and.b32 %r698, %r229, 2139062143;
and.b32 %r699, %r674, -2139062144;
sub.s32 %r700, %r697, %r698;
xor.b32 %r701, %r699, %r700;
xor.b32 %r232, %r701, -2139062144;

dp4a.s32.s32 %r231, %r232, %r86, %r225;

shl.b16 %rs36, %rs31, 3;
cvt.u64.u16 %rd71, %rs36;
and.b64 %rd72, %rd71, 2040;
add.s64 %rd73, %rd30, %rd72;
ld.global.nc.v2.u32 {%r702, %r703}, [%rd73];
shr.u32 %r706, %r643, 14;
and.b32 %r707, %r706, 255;
popc.b32 %r708, %r707;
shl.b32 %r709, %r708, 7;
and.b32 %r710, %r709, 128;
xor.b32 %r711, %r710, %r707;
mul.lo.s32 %r712, %r711, 16843009;
and.b32 %r713, %r712, 134480385;
or.b32 %r714, %r713, -2139062144;
add.s32 %r715, %r714, -16843009;
or.b32 %r235, %r715, %r713;

prmt.b32 %r235, %r235, 0, 0xba98;

xor.b32 %r716, %r702, %r235;
or.b32 %r717, %r716, -2139062144;
and.b32 %r718, %r235, 2139062143;
and.b32 %r719, %r702, -2139062144;
sub.s32 %r720, %r717, %r718;
xor.b32 %r721, %r719, %r720;
xor.b32 %r238, %r721, -2139062144;

dp4a.s32.s32 %r237, %r238, %r92, %r231;

and.b32 %r722, %r712, -2143281136;
or.b32 %r723, %r722, -2139062144;
add.s32 %r724, %r723, -16843009;
or.b32 %r241, %r724, %r722;

prmt.b32 %r241, %r241, 0, 0xba98;

xor.b32 %r725, %r241, %r703;
or.b32 %r726, %r725, -2139062144;
and.b32 %r727, %r241, 2139062143;
and.b32 %r728, %r703, -2139062144;
sub.s32 %r729, %r726, %r727;
xor.b32 %r730, %r728, %r729;
xor.b32 %r244, %r730, -2139062144;

dp4a.s32.s32 %r243, %r244, %r98, %r237;

cvt.u32.u16 %r731, %rs32;
mul.wide.u32 %rd74, %r731, 8;
add.s64 %rd75, %rd30, %rd74;
ld.global.nc.v2.u32 {%r732, %r733}, [%rd75];
shr.u32 %r736, %r642, 5;
and.b32 %r737, %r736, 255;
popc.b32 %r738, %r737;
shl.b32 %r739, %r738, 7;
and.b32 %r740, %r739, 128;
xor.b32 %r741, %r740, %r737;
mul.lo.s32 %r742, %r741, 16843009;
and.b32 %r743, %r742, 134480385;
or.b32 %r744, %r743, -2139062144;
add.s32 %r745, %r744, -16843009;
or.b32 %r247, %r745, %r743;

prmt.b32 %r247, %r247, 0, 0xba98;

xor.b32 %r746, %r732, %r247;
or.b32 %r747, %r746, -2139062144;
and.b32 %r748, %r247, 2139062143;
and.b32 %r749, %r732, -2139062144;
sub.s32 %r750, %r747, %r748;
xor.b32 %r751, %r749, %r750;
xor.b32 %r250, %r751, -2139062144;

dp4a.s32.s32 %r249, %r250, %r104, %r243;

and.b32 %r752, %r742, -2143281136;
or.b32 %r753, %r752, -2139062144;
add.s32 %r754, %r753, -16843009;
or.b32 %r253, %r754, %r752;

prmt.b32 %r253, %r253, 0, 0xba98;

xor.b32 %r755, %r253, %r733;
or.b32 %r756, %r755, -2139062144;
and.b32 %r757, %r253, 2139062143;
and.b32 %r758, %r733, -2139062144;
sub.s32 %r759, %r756, %r757;
xor.b32 %r760, %r758, %r759;
xor.b32 %r256, %r760, -2139062144;

dp4a.s32.s32 %r255, %r256, %r110, %r249;

shr.u32 %r761, %r642, 11;
or.b32 %r762, %r761, 1;
mul.lo.s32 %r763, %r255, %r762;
shr.s32 %r764, %r763, 31;
shr.u32 %r765, %r764, 29;
add.s32 %r766, %r763, %r765;
shr.s32 %r767, %r766, 3;
add.s64 %rd76, %rd64, %rd4;
ld.global.nc.u16 %rs4, [%rd76];

{ cvt.f32.f16 %f31, %rs4;}



{.reg .f16 low,high;
mov.b32 {low,high},%r112;
cvt.f32.f16 %f32, low;}


mul.ftz.f32 %f39, %f31, %f32;
cvt.rn.f32.s32 %f40, %r767;
fma.rn.ftz.f32 %f105, %f39, %f40, %f105;
add.s32 %r851, %r851, 128;
add.s64 %rd79, %rd79, 1056;
add.s32 %r852, %r852, 16;
setp.lt.s32 %p3, %r852, %r5;
@%p3 bra $L__BB201_5;

$L__BB201_6:
setp.eq.s32 %p4, %r1, 0;
@%p4 bra $L__BB201_8;

shl.b32 %r768, %r1, 9;
mov.u32 %r769, _ZZ13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r770, %r769, %r768;
shl.b32 %r771, %r2, 2;
add.s32 %r772, %r770, %r771;
st.shared.f32 [%r772+-512], %f108;
st.shared.f32 [%r772+-384], %f107;
st.shared.f32 [%r772+-256], %f106;
st.shared.f32 [%r772+-128], %f105;

$L__BB201_8:
setp.ne.s32 %p5, %r1, 0;
bar.sync 0;
@%p5 bra $L__BB201_17;

mov.u32 %r773, %ctaid.y;
mad.lo.s32 %r774, %r773, %r31, %r4;
mad.lo.s32 %r775, %r12, %r37, %r774;
shl.b32 %r776, %r2, 2;
mov.u32 %r777, _ZZ13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r19, %r777, %r776;
mov.u32 %r778, 2;
ld.shared.f32 %f41, [%r19];
add.ftz.f32 %f42, %f41, %f108;
ld.shared.f32 %f43, [%r19+512];
add.ftz.f32 %f44, %f43, %f42;
ld.shared.f32 %f45, [%r19+1024];
add.ftz.f32 %f46, %f45, %f44;
mov.b32 %r779, %f46;
mov.u32 %r780, 31;
mov.u32 %r781, 16;
mov.u32 %r782, -1;
shfl.sync.bfly.b32 %r783|%p6, %r779, %r781, %r780, %r782;
mov.b32 %f47, %r783;
add.ftz.f32 %f48, %f46, %f47;
mov.b32 %r784, %f48;
mov.u32 %r785, 8;
shfl.sync.bfly.b32 %r786|%p7, %r784, %r785, %r780, %r782;
mov.b32 %f49, %r786;
add.ftz.f32 %f50, %f48, %f49;
mov.b32 %r787, %f50;
mov.u32 %r788, 4;
shfl.sync.bfly.b32 %r789|%p8, %r787, %r788, %r780, %r782;
mov.b32 %f51, %r789;
add.ftz.f32 %f52, %f50, %f51;
mov.b32 %r790, %f52;
shfl.sync.bfly.b32 %r791|%p9, %r790, %r778, %r780, %r782;
mov.b32 %f53, %r791;
add.ftz.f32 %f54, %f52, %f53;
mov.b32 %r792, %f54;
mov.u32 %r793, 1;
shfl.sync.bfly.b32 %r794|%p10, %r792, %r793, %r780, %r782;
mov.b32 %f55, %r794;
add.ftz.f32 %f13, %f54, %f55;
setp.ge.u32 %p11, %r4, %r25;
setp.ne.s32 %p12, %r2, 0;
cvta.to.global.u64 %rd77, %rd14;
mul.wide.u32 %rd78, %r775, 4;
add.s64 %rd10, %rd77, %rd78;
or.pred %p13, %p12, %p11;
@%p13 bra $L__BB201_11;

st.global.f32 [%rd10], %f13;

$L__BB201_11:
ld.shared.f32 %f56, [%r19+128];
add.ftz.f32 %f57, %f56, %f107;
ld.shared.f32 %f58, [%r19+640];
add.ftz.f32 %f59, %f58, %f57;
ld.shared.f32 %f60, [%r19+1152];
add.ftz.f32 %f61, %f60, %f59;
mov.b32 %r795, %f61;
shfl.sync.bfly.b32 %r799|%p14, %r795, %r781, %r780, %r782;
mov.b32 %f62, %r799;
add.ftz.f32 %f63, %f61, %f62;
mov.b32 %r800, %f63;
shfl.sync.bfly.b32 %r802|%p15, %r800, %r785, %r780, %r782;
mov.b32 %f64, %r802;
add.ftz.f32 %f65, %f63, %f64;
mov.b32 %r803, %f65;
shfl.sync.bfly.b32 %r805|%p16, %r803, %r788, %r780, %r782;
mov.b32 %f66, %r805;
add.ftz.f32 %f67, %f65, %f66;
mov.b32 %r806, %f67;
shfl.sync.bfly.b32 %r808|%p17, %r806, %r778, %r780, %r782;
mov.b32 %f68, %r808;
add.ftz.f32 %f69, %f67, %f68;
mov.b32 %r809, %f69;
shfl.sync.bfly.b32 %r811|%p18, %r809, %r793, %r780, %r782;
mov.b32 %f70, %r811;
add.ftz.f32 %f14, %f69, %f70;
add.s32 %r812, %r4, 1;
setp.ge.u32 %p19, %r812, %r25;
setp.ne.s32 %p20, %r2, 1;
or.pred %p21, %p20, %p19;
@%p21 bra $L__BB201_13;

st.global.f32 [%rd10+4], %f14;

$L__BB201_13:
ld.shared.f32 %f71, [%r19+256];
add.ftz.f32 %f72, %f71, %f106;
ld.shared.f32 %f73, [%r19+768];
add.ftz.f32 %f74, %f73, %f72;
ld.shared.f32 %f75, [%r19+1280];
add.ftz.f32 %f76, %f75, %f74;
mov.b32 %r813, %f76;
mov.u32 %r814, 31;
mov.u32 %r815, 16;
mov.u32 %r816, -1;
shfl.sync.bfly.b32 %r817|%p22, %r813, %r815, %r814, %r816;
mov.b32 %f77, %r817;
add.ftz.f32 %f78, %f76, %f77;
mov.b32 %r818, %f78;
mov.u32 %r819, 8;
shfl.sync.bfly.b32 %r820|%p23, %r818, %r819, %r814, %r816;
mov.b32 %f79, %r820;
add.ftz.f32 %f80, %f78, %f79;
mov.b32 %r821, %f80;
mov.u32 %r822, 4;
shfl.sync.bfly.b32 %r823|%p24, %r821, %r822, %r814, %r816;
mov.b32 %f81, %r823;
add.ftz.f32 %f82, %f80, %f81;
mov.b32 %r824, %f82;
mov.u32 %r825, 2;
shfl.sync.bfly.b32 %r826|%p25, %r824, %r825, %r814, %r816;
mov.b32 %f83, %r826;
add.ftz.f32 %f84, %f82, %f83;
mov.b32 %r827, %f84;
mov.u32 %r828, 1;
shfl.sync.bfly.b32 %r829|%p26, %r827, %r828, %r814, %r816;
mov.b32 %f85, %r829;
add.ftz.f32 %f15, %f84, %f85;
add.s32 %r830, %r4, 2;
setp.ge.u32 %p27, %r830, %r25;
setp.ne.s32 %p28, %r2, 2;
or.pred %p29, %p28, %p27;
@%p29 bra $L__BB201_15;

st.global.f32 [%rd10+8], %f15;

$L__BB201_15:
ld.shared.f32 %f86, [%r19+384];
add.ftz.f32 %f87, %f86, %f105;
ld.shared.f32 %f88, [%r19+896];
add.ftz.f32 %f89, %f88, %f87;
ld.shared.f32 %f90, [%r19+1408];
add.ftz.f32 %f91, %f90, %f89;
mov.b32 %r831, %f91;
shfl.sync.bfly.b32 %r835|%p30, %r831, %r815, %r814, %r816;
mov.b32 %f92, %r835;
add.ftz.f32 %f93, %f91, %f92;
mov.b32 %r836, %f93;
shfl.sync.bfly.b32 %r838|%p31, %r836, %r819, %r814, %r816;
mov.b32 %f94, %r838;
add.ftz.f32 %f95, %f93, %f94;
mov.b32 %r839, %f95;
shfl.sync.bfly.b32 %r841|%p32, %r839, %r822, %r814, %r816;
mov.b32 %f96, %r841;
add.ftz.f32 %f97, %f95, %f96;
mov.b32 %r842, %f97;
shfl.sync.bfly.b32 %r844|%p33, %r842, %r825, %r814, %r816;
mov.b32 %f98, %r844;
add.ftz.f32 %f99, %f97, %f98;
mov.b32 %r845, %f99;
shfl.sync.bfly.b32 %r847|%p34, %r845, %r828, %r814, %r816;
mov.b32 %f100, %r847;
add.ftz.f32 %f16, %f99, %f100;
add.s32 %r848, %r4, 3;
setp.ge.u32 %p35, %r848, %r25;
setp.ne.s32 %p36, %r2, 3;
or.pred %p37, %p36, %p35;
@%p37 bra $L__BB201_17;

st.global.f32 [%rd10+12], %f16;

$L__BB201_17:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<34>;
.reg .b16 %rs<19>;
.reg .f32 %f<120>;
.reg .b32 %r<512>;
.reg .b64 %rd<86>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[384];

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate[384];

ld.param.u64 %rd11, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd12, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd13, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
ld.param.f32 %f31, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+16];
ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+8];
ld.param.u64 %rd14, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3];
mov.u32 %r1, %tid.y;
mov.u32 %r2, %ctaid.y;
setp.eq.s64 %p2, %rd13, 0;
@%p2 bra $L__BB202_2;

cvta.to.global.u64 %rd20, %rd13;
mul.wide.u32 %rd21, %r2, 4;
add.s64 %rd22, %rd20, %rd21;
ld.global.nc.u32 %r508, [%rd22];
mul.hi.u32 %r45, %r2, %r29;
add.s32 %r46, %r45, %r2;
shr.u32 %r47, %r46, %r30;
mul.lo.s32 %r48, %r47, %r31;
sub.s32 %r509, %r2, %r48;
bra.uni $L__BB202_3;

$L__BB202_2:
mul.hi.u32 %r49, %r2, %r33;
add.s32 %r50, %r49, %r2;
shr.u32 %r508, %r50, %r34;
mov.u32 %r509, %r2;

$L__BB202_3:
mov.u32 %r51, %tid.x;
or.b32 %r52, %r51, %r1;
setp.ne.s32 %p3, %r52, 0;
mov.f32 %f113, 0f00000000;
mov.f32 %f109, %f113;
mov.f32 %f110, %f113;
@%p3 bra $L__BB202_8;

setp.eq.s64 %p4, %rd14, 0;
mov.f32 %f109, 0f00000000;
mov.f32 %f110, %f109;
@%p4 bra $L__BB202_6;

cvta.to.global.u64 %rd23, %rd14;
mov.u32 %r53, %ctaid.z;
mul.lo.s32 %r54, %r53, %r44;
cvt.u64.u32 %rd24, %r54;
selp.b32 %r55, %r2, %r508, %p2;
mul.lo.s32 %r56, %r55, %r38;
cvt.u64.u32 %rd25, %r56;
mov.u32 %r57, %ctaid.x;
cvt.s64.s32 %rd26, %r57;
cvt.u64.u32 %rd27, %r51;
add.s64 %rd28, %rd27, %rd26;
add.s64 %rd29, %rd28, %rd24;
add.s64 %rd30, %rd29, %rd25;
shl.b64 %rd31, %rd30, 2;
add.s64 %rd32, %rd23, %rd31;
ld.global.f32 %f110, [%rd32];

$L__BB202_6:
setp.eq.s64 %p6, %rd16, 0;
setp.eq.s64 %p7, %rd15, 0;
or.pred %p8, %p6, %p7;
@%p8 bra $L__BB202_8;

cvta.to.global.u64 %rd33, %rd16;
mov.u32 %r59, %ctaid.z;
mul.lo.s32 %r60, %r59, %r44;
cvt.u64.u32 %rd34, %r60;
selp.b32 %r61, %r2, %r508, %p2;
mul.lo.s32 %r62, %r61, %r38;
cvt.u64.u32 %rd35, %r62;
mov.u32 %r63, %ctaid.x;
cvt.s64.s32 %rd36, %r63;
cvt.u64.u32 %rd37, %r51;
add.s64 %rd38, %rd37, %rd36;
add.s64 %rd39, %rd38, %rd34;
add.s64 %rd40, %rd39, %rd35;
shl.b64 %rd41, %rd40, 2;
add.s64 %rd42, %rd33, %rd41;
ld.global.f32 %f109, [%rd42];

$L__BB202_8:
shl.b32 %r66, %r1, 5;
add.s32 %r67, %r66, %r51;
shr.s32 %r68, %r67, 31;
shr.u32 %r69, %r68, 29;
add.s32 %r70, %r67, %r69;
shr.s32 %r511, %r70, 3;
shr.u32 %r71, %r28, 8;
setp.ge.s32 %p10, %r511, %r71;
mov.f32 %f115, %f113;
@%p10 bra $L__BB202_13;

mov.u32 %r72, %ctaid.z;
and.b32 %r79, %r70, -8;
sub.s32 %r9, %r67, %r79;
mul.hi.u32 %r80, %r72, %r39;
add.s32 %r81, %r80, %r72;
shr.u32 %r82, %r81, %r40;
mad.lo.s32 %r83, %r82, %r42, %r511;
mad.lo.s32 %r84, %r508, %r36, %r83;
mov.u32 %r85, %ctaid.x;
mad.lo.s32 %r86, %r85, %r32, %r84;
mul.wide.s32 %rd43, %r86, 66;
cvta.to.global.u64 %rd44, %rd15;
add.s64 %rd85, %rd44, %rd43;
shl.b32 %r510, %r511, 3;
cvta.to.global.u64 %rd45, %rd11;
add.s64 %rd84, %rd45, %rd43;
mul.lo.s32 %r87, %r509, %r37;
cvt.u64.u32 %rd46, %r87;
mul.lo.s32 %r88, %r72, %r43;
cvt.u64.u32 %rd47, %r88;
add.s64 %rd6, %rd46, %rd47;
mov.f32 %f115, 0f00000000;
cvta.to.global.u64 %rd48, %rd12;
mov.f32 %f113, %f115;

$L__BB202_10:
cvt.u64.u32 %rd49, %r510;
add.s64 %rd50, %rd6, %rd49;
shl.b32 %r138, %r9, 2;
mul.wide.s32 %rd51, %r138, 2;
add.s64 %rd52, %rd84, %rd51;
ld.global.nc.u16 %rs2, [%rd52+2];
shr.u16 %rs3, %rs2, 8;
ld.global.nc.u16 %rs4, [%rd52+4];
shr.u16 %rs5, %rs4, 8;
ld.global.nc.u16 %rs6, [%rd52+6];
cvt.u32.u16 %r139, %rs6;
ld.global.nc.u16 %rs7, [%rd52+8];
cvt.u32.u16 %r140, %rs7;
mov.b32 %r141, {%rs6, %rs7};
cvt.s64.s32 %rd53, %r9;
add.s64 %rd54, %rd50, %rd53;
mul.lo.s64 %rd55, %rd54, 36;
shl.b16 %rs8, %rs2, 3;
cvt.u64.u16 %rd56, %rs8;
and.b64 %rd57, %rd56, 2040;
mov.u64 %rd58, iq2xxs_grid;
add.s64 %rd59, %rd58, %rd57;
ld.global.nc.v2.u32 {%r142, %r143}, [%rd59];
mov.u32 %r94, 0;
and.b32 %r146, %r139, 255;
popc.b32 %r147, %r146;
shl.b32 %r148, %r147, 7;
and.b32 %r149, %r148, 128;
xor.b32 %r150, %r149, %r146;
mul.lo.s32 %r151, %r150, 16843009;
and.b32 %r152, %r151, 134480385;
or.b32 %r153, %r152, -2139062144;
add.s32 %r154, %r153, -16843009;
or.b32 %r89, %r154, %r152;

prmt.b32 %r89, %r89, 0, 0xba98;

xor.b32 %r155, %r142, %r89;
or.b32 %r156, %r155, -2139062144;
and.b32 %r157, %r89, 2139062143;
and.b32 %r158, %r142, -2139062144;
sub.s32 %r159, %r156, %r157;
xor.b32 %r160, %r158, %r159;
xor.b32 %r92, %r160, -2139062144;
add.s64 %rd60, %rd48, %rd55;
ld.global.nc.u32 %r13, [%rd60+4];

dp4a.s32.s32 %r91, %r92, %r13, %r94;

and.b32 %r161, %r151, -2143281136;
or.b32 %r162, %r161, -2139062144;
add.s32 %r163, %r162, -16843009;
or.b32 %r95, %r163, %r161;

prmt.b32 %r95, %r95, 0, 0xba98;

xor.b32 %r164, %r95, %r143;
or.b32 %r165, %r164, -2139062144;
and.b32 %r166, %r95, 2139062143;
and.b32 %r167, %r143, -2139062144;
sub.s32 %r168, %r165, %r166;
xor.b32 %r169, %r167, %r168;
xor.b32 %r98, %r169, -2139062144;
ld.global.nc.u32 %r14, [%rd60+8];

dp4a.s32.s32 %r97, %r98, %r14, %r91;

cvt.u32.u16 %r170, %rs3;
mul.wide.u32 %rd61, %r170, 8;
add.s64 %rd62, %rd58, %rd61;
ld.global.nc.v2.u32 {%r171, %r172}, [%rd62];
shr.u32 %r175, %r139, 7;
and.b32 %r176, %r175, 255;
popc.b32 %r177, %r176;
shl.b32 %r178, %r177, 7;
and.b32 %r179, %r178, 128;
xor.b32 %r180, %r179, %r176;
mul.lo.s32 %r181, %r180, 16843009;
and.b32 %r182, %r181, 134480385;
or.b32 %r183, %r182, -2139062144;
add.s32 %r184, %r183, -16843009;
or.b32 %r101, %r184, %r182;

prmt.b32 %r101, %r101, 0, 0xba98;

xor.b32 %r185, %r171, %r101;
or.b32 %r186, %r185, -2139062144;
and.b32 %r187, %r101, 2139062143;
and.b32 %r188, %r171, -2139062144;
sub.s32 %r189, %r186, %r187;
xor.b32 %r190, %r188, %r189;
xor.b32 %r104, %r190, -2139062144;
ld.global.nc.u32 %r15, [%rd60+12];

dp4a.s32.s32 %r103, %r104, %r15, %r97;

and.b32 %r191, %r181, -2143281136;
or.b32 %r192, %r191, -2139062144;
add.s32 %r193, %r192, -16843009;
or.b32 %r107, %r193, %r191;

prmt.b32 %r107, %r107, 0, 0xba98;

xor.b32 %r194, %r107, %r172;
or.b32 %r195, %r194, -2139062144;
and.b32 %r196, %r107, 2139062143;
and.b32 %r197, %r172, -2139062144;
sub.s32 %r198, %r195, %r196;
xor.b32 %r199, %r197, %r198;
xor.b32 %r110, %r199, -2139062144;
ld.global.nc.u32 %r16, [%rd60+16];

dp4a.s32.s32 %r109, %r110, %r16, %r103;

shl.b16 %rs9, %rs4, 3;
cvt.u64.u16 %rd63, %rs9;
and.b64 %rd64, %rd63, 2040;
add.s64 %rd65, %rd58, %rd64;
ld.global.nc.v2.u32 {%r200, %r201}, [%rd65];
shr.u32 %r204, %r141, 14;
and.b32 %r205, %r204, 255;
popc.b32 %r206, %r205;
shl.b32 %r207, %r206, 7;
and.b32 %r208, %r207, 128;
xor.b32 %r209, %r208, %r205;
mul.lo.s32 %r210, %r209, 16843009;
and.b32 %r211, %r210, 134480385;
or.b32 %r212, %r211, -2139062144;
add.s32 %r213, %r212, -16843009;
or.b32 %r113, %r213, %r211;

prmt.b32 %r113, %r113, 0, 0xba98;

xor.b32 %r214, %r200, %r113;
or.b32 %r215, %r214, -2139062144;
and.b32 %r216, %r113, 2139062143;
and.b32 %r217, %r200, -2139062144;
sub.s32 %r218, %r215, %r216;
xor.b32 %r219, %r217, %r218;
xor.b32 %r116, %r219, -2139062144;
ld.global.nc.u32 %r17, [%rd60+20];

dp4a.s32.s32 %r115, %r116, %r17, %r109;

and.b32 %r220, %r210, -2143281136;
or.b32 %r221, %r220, -2139062144;
add.s32 %r222, %r221, -16843009;
or.b32 %r119, %r222, %r220;

prmt.b32 %r119, %r119, 0, 0xba98;

xor.b32 %r223, %r119, %r201;
or.b32 %r224, %r223, -2139062144;
and.b32 %r225, %r119, 2139062143;
and.b32 %r226, %r201, -2139062144;
sub.s32 %r227, %r224, %r225;
xor.b32 %r228, %r226, %r227;
xor.b32 %r122, %r228, -2139062144;
ld.global.nc.u32 %r18, [%rd60+24];

dp4a.s32.s32 %r121, %r122, %r18, %r115;

cvt.u32.u16 %r229, %rs5;
mul.wide.u32 %rd66, %r229, 8;
add.s64 %rd67, %rd58, %rd66;
ld.global.nc.v2.u32 {%r230, %r231}, [%rd67];
shr.u32 %r234, %r140, 5;
and.b32 %r235, %r234, 255;
popc.b32 %r236, %r235;
shl.b32 %r237, %r236, 7;
and.b32 %r238, %r237, 128;
xor.b32 %r239, %r238, %r235;
mul.lo.s32 %r240, %r239, 16843009;
and.b32 %r241, %r240, 134480385;
or.b32 %r242, %r241, -2139062144;
add.s32 %r243, %r242, -16843009;
or.b32 %r125, %r243, %r241;

prmt.b32 %r125, %r125, 0, 0xba98;

xor.b32 %r244, %r230, %r125;
or.b32 %r245, %r244, -2139062144;
and.b32 %r246, %r125, 2139062143;
and.b32 %r247, %r230, -2139062144;
sub.s32 %r248, %r245, %r246;
xor.b32 %r249, %r247, %r248;
xor.b32 %r128, %r249, -2139062144;
ld.global.nc.u32 %r19, [%rd60+28];

dp4a.s32.s32 %r127, %r128, %r19, %r121;

and.b32 %r250, %r240, -2143281136;
or.b32 %r251, %r250, -2139062144;
add.s32 %r252, %r251, -16843009;
or.b32 %r131, %r252, %r250;

prmt.b32 %r131, %r131, 0, 0xba98;

xor.b32 %r253, %r131, %r231;
or.b32 %r254, %r253, -2139062144;
and.b32 %r255, %r131, 2139062143;
and.b32 %r256, %r231, -2139062144;
sub.s32 %r257, %r254, %r255;
xor.b32 %r258, %r256, %r257;
xor.b32 %r134, %r258, -2139062144;
ld.global.nc.u32 %r20, [%rd60+32];

dp4a.s32.s32 %r133, %r134, %r20, %r127;

shr.u32 %r259, %r140, 11;
or.b32 %r260, %r259, 1;
mul.lo.s32 %r261, %r133, %r260;
shr.s32 %r262, %r261, 31;
shr.u32 %r263, %r262, 29;
add.s32 %r264, %r261, %r263;
shr.s32 %r265, %r264, 3;
ld.global.nc.u16 %rs1, [%rd84];

{ cvt.f32.f16 %f40, %rs1;}


ld.global.nc.u32 %r21, [%rd60];

{.reg .f16 low,high;
mov.b32 {low,high},%r21;
cvt.f32.f16 %f41, low;}


mul.ftz.f32 %f42, %f40, %f41;
cvt.rn.f32.s32 %f43, %r265;
fma.rn.ftz.f32 %f115, %f42, %f43, %f115;
setp.eq.s64 %p11, %rd15, 0;
@%p11 bra $L__BB202_12;

add.s64 %rd69, %rd85, %rd51;
ld.global.u16 %rs11, [%rd69+2];
shr.u16 %rs12, %rs11, 8;
ld.global.u16 %rs13, [%rd69+4];
shr.u16 %rs14, %rs13, 8;
ld.global.u16 %rs15, [%rd69+6];
cvt.u32.u16 %r316, %rs15;
ld.global.u16 %rs16, [%rd69+8];
cvt.u32.u16 %r317, %rs16;
mov.b32 %r318, {%rs15, %rs16};
shl.b16 %rs17, %rs11, 3;
cvt.u64.u16 %rd70, %rs17;
and.b64 %rd71, %rd70, 2040;
add.s64 %rd73, %rd58, %rd71;
ld.global.nc.v2.u32 {%r319, %r320}, [%rd73];
and.b32 %r323, %r316, 255;
popc.b32 %r324, %r323;
shl.b32 %r325, %r324, 7;
and.b32 %r326, %r325, 128;
xor.b32 %r327, %r326, %r323;
mul.lo.s32 %r328, %r327, 16843009;
and.b32 %r329, %r328, 134480385;
or.b32 %r330, %r329, -2139062144;
add.s32 %r331, %r330, -16843009;
or.b32 %r266, %r331, %r329;

prmt.b32 %r266, %r266, 0, 0xba98;

xor.b32 %r332, %r319, %r266;
or.b32 %r333, %r332, -2139062144;
and.b32 %r334, %r266, 2139062143;
and.b32 %r335, %r319, -2139062144;
sub.s32 %r336, %r333, %r334;
xor.b32 %r337, %r335, %r336;
xor.b32 %r269, %r337, -2139062144;

dp4a.s32.s32 %r268, %r269, %r13, %r94;

and.b32 %r338, %r328, -2143281136;
or.b32 %r339, %r338, -2139062144;
add.s32 %r340, %r339, -16843009;
or.b32 %r272, %r340, %r338;

prmt.b32 %r272, %r272, 0, 0xba98;

xor.b32 %r341, %r272, %r320;
or.b32 %r342, %r341, -2139062144;
and.b32 %r343, %r272, 2139062143;
and.b32 %r344, %r320, -2139062144;
sub.s32 %r345, %r342, %r343;
xor.b32 %r346, %r344, %r345;
xor.b32 %r275, %r346, -2139062144;

dp4a.s32.s32 %r274, %r275, %r14, %r268;

cvt.u32.u16 %r347, %rs12;
mul.wide.u32 %rd74, %r347, 8;
add.s64 %rd75, %rd58, %rd74;
ld.global.nc.v2.u32 {%r348, %r349}, [%rd75];
shr.u32 %r352, %r316, 7;
and.b32 %r353, %r352, 255;
popc.b32 %r354, %r353;
shl.b32 %r355, %r354, 7;
and.b32 %r356, %r355, 128;
xor.b32 %r357, %r356, %r353;
mul.lo.s32 %r358, %r357, 16843009;
and.b32 %r359, %r358, 134480385;
or.b32 %r360, %r359, -2139062144;
add.s32 %r361, %r360, -16843009;
or.b32 %r278, %r361, %r359;

prmt.b32 %r278, %r278, 0, 0xba98;

xor.b32 %r362, %r348, %r278;
or.b32 %r363, %r362, -2139062144;
and.b32 %r364, %r278, 2139062143;
and.b32 %r365, %r348, -2139062144;
sub.s32 %r366, %r363, %r364;
xor.b32 %r367, %r365, %r366;
xor.b32 %r281, %r367, -2139062144;

dp4a.s32.s32 %r280, %r281, %r15, %r274;

and.b32 %r368, %r358, -2143281136;
or.b32 %r369, %r368, -2139062144;
add.s32 %r370, %r369, -16843009;
or.b32 %r284, %r370, %r368;

prmt.b32 %r284, %r284, 0, 0xba98;

xor.b32 %r371, %r284, %r349;
or.b32 %r372, %r371, -2139062144;
and.b32 %r373, %r284, 2139062143;
and.b32 %r374, %r349, -2139062144;
sub.s32 %r375, %r372, %r373;
xor.b32 %r376, %r374, %r375;
xor.b32 %r287, %r376, -2139062144;

dp4a.s32.s32 %r286, %r287, %r16, %r280;

shl.b16 %rs18, %rs13, 3;
cvt.u64.u16 %rd76, %rs18;
and.b64 %rd77, %rd76, 2040;
add.s64 %rd78, %rd58, %rd77;
ld.global.nc.v2.u32 {%r377, %r378}, [%rd78];
shr.u32 %r381, %r318, 14;
and.b32 %r382, %r381, 255;
popc.b32 %r383, %r382;
shl.b32 %r384, %r383, 7;
and.b32 %r385, %r384, 128;
xor.b32 %r386, %r385, %r382;
mul.lo.s32 %r387, %r386, 16843009;
and.b32 %r388, %r387, 134480385;
or.b32 %r389, %r388, -2139062144;
add.s32 %r390, %r389, -16843009;
or.b32 %r290, %r390, %r388;

prmt.b32 %r290, %r290, 0, 0xba98;

xor.b32 %r391, %r377, %r290;
or.b32 %r392, %r391, -2139062144;
and.b32 %r393, %r290, 2139062143;
and.b32 %r394, %r377, -2139062144;
sub.s32 %r395, %r392, %r393;
xor.b32 %r396, %r394, %r395;
xor.b32 %r293, %r396, -2139062144;

dp4a.s32.s32 %r292, %r293, %r17, %r286;

and.b32 %r397, %r387, -2143281136;
or.b32 %r398, %r397, -2139062144;
add.s32 %r399, %r398, -16843009;
or.b32 %r296, %r399, %r397;

prmt.b32 %r296, %r296, 0, 0xba98;

xor.b32 %r400, %r296, %r378;
or.b32 %r401, %r400, -2139062144;
and.b32 %r402, %r296, 2139062143;
and.b32 %r403, %r378, -2139062144;
sub.s32 %r404, %r401, %r402;
xor.b32 %r405, %r403, %r404;
xor.b32 %r299, %r405, -2139062144;

dp4a.s32.s32 %r298, %r299, %r18, %r292;

cvt.u32.u16 %r406, %rs14;
mul.wide.u32 %rd79, %r406, 8;
add.s64 %rd80, %rd58, %rd79;
ld.global.nc.v2.u32 {%r407, %r408}, [%rd80];
shr.u32 %r411, %r317, 5;
and.b32 %r412, %r411, 255;
popc.b32 %r413, %r412;
shl.b32 %r414, %r413, 7;
and.b32 %r415, %r414, 128;
xor.b32 %r416, %r415, %r412;
mul.lo.s32 %r417, %r416, 16843009;
and.b32 %r418, %r417, 134480385;
or.b32 %r419, %r418, -2139062144;
add.s32 %r420, %r419, -16843009;
or.b32 %r302, %r420, %r418;

prmt.b32 %r302, %r302, 0, 0xba98;

xor.b32 %r421, %r407, %r302;
or.b32 %r422, %r421, -2139062144;
and.b32 %r423, %r302, 2139062143;
and.b32 %r424, %r407, -2139062144;
sub.s32 %r425, %r422, %r423;
xor.b32 %r426, %r424, %r425;
xor.b32 %r305, %r426, -2139062144;

dp4a.s32.s32 %r304, %r305, %r19, %r298;

and.b32 %r427, %r417, -2143281136;
or.b32 %r428, %r427, -2139062144;
add.s32 %r429, %r428, -16843009;
or.b32 %r308, %r429, %r427;

prmt.b32 %r308, %r308, 0, 0xba98;

xor.b32 %r430, %r308, %r408;
or.b32 %r431, %r430, -2139062144;
and.b32 %r432, %r308, 2139062143;
and.b32 %r433, %r408, -2139062144;
sub.s32 %r434, %r431, %r432;
xor.b32 %r435, %r433, %r434;
xor.b32 %r311, %r435, -2139062144;

dp4a.s32.s32 %r310, %r311, %r20, %r304;

shr.u32 %r436, %r317, 11;
or.b32 %r437, %r436, 1;
mul.lo.s32 %r438, %r310, %r437;
shr.s32 %r439, %r438, 31;
shr.u32 %r440, %r439, 29;
add.s32 %r441, %r438, %r440;
shr.s32 %r442, %r441, 3;
ld.global.u16 %rs10, [%rd85];

{ cvt.f32.f16 %f44, %rs10;}



{.reg .f16 low,high;
mov.b32 {low,high},%r21;
cvt.f32.f16 %f45, low;}


mul.ftz.f32 %f46, %f44, %f45;
cvt.rn.f32.s32 %f47, %r442;
fma.rn.ftz.f32 %f113, %f46, %f47, %f113;

$L__BB202_12:
add.s64 %rd85, %rd85, 1056;
add.s32 %r510, %r510, 128;
add.s64 %rd84, %rd84, 1056;
add.s32 %r511, %r511, 16;
setp.lt.s32 %p12, %r511, %r71;
@%p12 bra $L__BB202_10;

$L__BB202_13:
setp.eq.s32 %p13, %r1, 0;
@%p13 bra $L__BB202_16;

setp.eq.s64 %p14, %rd15, 0;
add.s32 %r24, %r1, -1;
shl.b32 %r447, %r1, 7;
mov.u32 %r448, _ZZ13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r449, %r448, %r447;
shl.b32 %r450, %r51, 2;
add.s32 %r451, %r449, %r450;
st.shared.f32 [%r451+-128], %f115;
@%p14 bra $L__BB202_16;

shl.b32 %r453, %r24, 7;
mov.u32 %r454, _ZZ13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r455, %r454, %r453;
add.s32 %r457, %r455, %r450;
st.shared.f32 [%r457], %f113;

$L__BB202_16:
setp.ne.s32 %p15, %r1, 0;
bar.sync 0;
@%p15 bra $L__BB202_35;

setp.ne.s64 %p16, %rd15, 0;
shl.b32 %r460, %r51, 2;
mov.u32 %r461, _ZZ13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r25, %r461, %r460;
ld.shared.f32 %f48, [%r25];
add.ftz.f32 %f13, %f115, %f48;
@%p16 bra $L__BB202_19;
bra.uni $L__BB202_18;

$L__BB202_19:
mov.u32 %r464, _ZZ13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r465, %r464, %r460;
ld.shared.f32 %f52, [%r465];
add.ftz.f32 %f53, %f113, %f52;
ld.shared.f32 %f54, [%r25+128];
add.ftz.f32 %f55, %f13, %f54;
ld.shared.f32 %f56, [%r465+128];
add.ftz.f32 %f57, %f53, %f56;
ld.shared.f32 %f58, [%r25+256];
add.ftz.f32 %f117, %f55, %f58;
ld.shared.f32 %f59, [%r465+256];
add.ftz.f32 %f113, %f57, %f59;
bra.uni $L__BB202_20;

$L__BB202_18:
ld.shared.f32 %f49, [%r25+128];
add.ftz.f32 %f50, %f13, %f49;
ld.shared.f32 %f51, [%r25+256];
add.ftz.f32 %f117, %f50, %f51;

$L__BB202_20:
mov.b32 %r466, %f117;
mov.u32 %r467, 31;
mov.u32 %r468, 16;
mov.u32 %r469, -1;
shfl.sync.bfly.b32 %r470|%p17, %r466, %r468, %r467, %r469;
mov.b32 %f60, %r470;
add.ftz.f32 %f61, %f117, %f60;
mov.b32 %r471, %f61;
mov.u32 %r472, 8;
shfl.sync.bfly.b32 %r473|%p18, %r471, %r472, %r467, %r469;
mov.b32 %f62, %r473;
add.ftz.f32 %f63, %f61, %f62;
mov.b32 %r474, %f63;
mov.u32 %r475, 4;
shfl.sync.bfly.b32 %r476|%p19, %r474, %r475, %r467, %r469;
mov.b32 %f64, %r476;
add.ftz.f32 %f65, %f63, %f64;
mov.b32 %r477, %f65;
mov.u32 %r478, 2;
shfl.sync.bfly.b32 %r479|%p20, %r477, %r478, %r467, %r469;
mov.b32 %f66, %r479;
add.ftz.f32 %f19, %f65, %f66;
mov.b32 %r480, %f19;
mov.u32 %r481, 1;
shfl.sync.bfly.b32 %r26|%p1, %r480, %r481, %r467, %r469;
setp.eq.s64 %p21, %rd15, 0;
@%p21 bra $L__BB202_22;

mov.b32 %r482, %f113;
shfl.sync.bfly.b32 %r486|%p22, %r482, %r468, %r467, %r469;
mov.b32 %f67, %r486;
add.ftz.f32 %f68, %f113, %f67;
mov.b32 %r487, %f68;
shfl.sync.bfly.b32 %r489|%p23, %r487, %r472, %r467, %r469;
mov.b32 %f69, %r489;
add.ftz.f32 %f70, %f68, %f69;
mov.b32 %r490, %f70;
shfl.sync.bfly.b32 %r492|%p24, %r490, %r475, %r467, %r469;
mov.b32 %f71, %r492;
add.ftz.f32 %f72, %f70, %f71;
mov.b32 %r493, %f72;
shfl.sync.bfly.b32 %r495|%p25, %r493, %r478, %r467, %r469;
mov.b32 %f73, %r495;
add.ftz.f32 %f74, %f72, %f73;
mov.b32 %r496, %f74;
shfl.sync.bfly.b32 %r498|%p26, %r496, %r481, %r467, %r469;
mov.b32 %f75, %r498;
add.ftz.f32 %f113, %f74, %f75;

$L__BB202_22:
setp.ne.s32 %p27, %r51, 0;
mov.b32 %f76, %r26;
add.ftz.f32 %f22, %f19, %f76;
@%p27 bra $L__BB202_35;

add.ftz.f32 %f119, %f110, %f22;
@%p21 bra $L__BB202_34;

add.ftz.f32 %f24, %f109, %f113;
setp.gt.s32 %p29, %r27, 2;
@%p29 bra $L__BB202_28;

setp.eq.s32 %p32, %r27, 1;
@%p32 bra $L__BB202_32;

setp.eq.s32 %p33, %r27, 2;
@%p33 bra $L__BB202_27;
bra.uni $L__BB202_33;

$L__BB202_27:
mul.ftz.f32 %f96, %f24, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f97, %f96;
add.ftz.f32 %f98, %f97, 0f3F800000;
div.approx.ftz.f32 %f99, %f24, %f98;
mul.ftz.f32 %f119, %f119, %f99;
bra.uni $L__BB202_34;

$L__BB202_28:
setp.eq.s32 %p30, %r27, 3;
@%p30 bra $L__BB202_31;

setp.ne.s32 %p31, %r27, 6;
@%p31 bra $L__BB202_33;

min.ftz.f32 %f77, %f24, %f31;
neg.ftz.f32 %f78, %f31;
min.ftz.f32 %f79, %f119, %f31;
max.ftz.f32 %f80, %f79, %f78;
mul.ftz.f32 %f81, %f77, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f82, %f81;
add.ftz.f32 %f83, %f82, 0f3F800000;
div.approx.ftz.f32 %f84, %f77, %f83;
mul.ftz.f32 %f119, %f80, %f84;
bra.uni $L__BB202_34;

$L__BB202_32:
mul.ftz.f32 %f100, %f24, 0f3F000000;
mul.ftz.f32 %f101, %f24, 0f3D372713;
fma.rn.ftz.f32 %f102, %f24, %f101, 0f3F800000;
mul.ftz.f32 %f103, %f24, 0f3F4C422A;
mul.ftz.f32 %f104, %f103, %f102;
tanh.approx.f32 %f105, %f104;
add.ftz.f32 %f106, %f105, 0f3F800000;
mul.ftz.f32 %f107, %f100, %f106;
mul.ftz.f32 %f119, %f119, %f107;
bra.uni $L__BB202_34;

$L__BB202_31:
mov.f32 %f85, 0f40E00000;
min.ftz.f32 %f86, %f24, %f85;
min.ftz.f32 %f87, %f119, %f85;
mov.f32 %f88, 0fC0E00000;
max.ftz.f32 %f89, %f87, %f88;
mul.ftz.f32 %f90, %f86, 0f3FD9DB23;
mul.ftz.f32 %f91, %f90, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f92, %f91;
add.ftz.f32 %f93, %f92, 0f3F800000;
div.approx.ftz.f32 %f94, %f86, %f93;
add.ftz.f32 %f95, %f89, 0f3F800000;
mul.ftz.f32 %f119, %f95, %f94;
bra.uni $L__BB202_34;

$L__BB202_33:
mul.ftz.f32 %f119, %f119, %f24;

$L__BB202_34:
ld.param.u32 %r507, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r506, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
cvta.to.global.u64 %rd81, %rd19;
mov.u32 %r501, %ctaid.z;
mov.u32 %r502, %ctaid.y;
mov.u32 %r503, %ctaid.x;
mad.lo.s32 %r504, %r502, %r507, %r503;
mad.lo.s32 %r505, %r501, %r506, %r504;
mul.wide.u32 %rd82, %r505, 4;
add.s64 %rd83, %rd81, %rd82;
st.global.f32 [%rd83], %f119;

$L__BB202_35:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<12>;
.reg .b16 %rs<10>;
.reg .f32 %f<28>;
.reg .b32 %r<272>;
.reg .b64 %rd<45>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[384];

ld.param.u64 %rd10, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd11, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd12, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd13, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r19, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r18, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r17, [_Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
mov.u32 %r1, %tid.y;
shl.b32 %r34, %r1, 5;
mov.u32 %r35, %tid.x;
add.s32 %r2, %r34, %r35;
shr.u32 %r3, %r33, 8;
mov.u32 %r269, %ctaid.y;
setp.eq.s64 %p2, %rd12, 0;
@%p2 bra $L__BB203_2;

cvta.to.global.u64 %rd14, %rd12;
mul.wide.u32 %rd15, %r269, 4;
add.s64 %rd16, %rd14, %rd15;
ld.global.nc.u32 %r268, [%rd16];
mul.hi.u32 %r36, %r269, %r17;
add.s32 %r37, %r36, %r269;
shr.u32 %r38, %r37, %r18;
mul.lo.s32 %r39, %r38, %r19;
sub.s32 %r269, %r269, %r39;
bra.uni $L__BB203_3;

$L__BB203_2:
mul.hi.u32 %r40, %r269, %r21;
add.s32 %r41, %r40, %r269;
shr.u32 %r268, %r41, %r22;

$L__BB203_3:
shr.s32 %r42, %r2, 31;
shr.u32 %r43, %r42, 29;
add.s32 %r44, %r2, %r43;
shr.s32 %r271, %r44, 3;
setp.ge.s32 %p3, %r271, %r3;
mov.f32 %f27, 0f00000000;
@%p3 bra $L__BB203_6;

mov.u32 %r45, %ctaid.z;
and.b32 %r49, %r44, -8;
sub.s32 %r50, %r2, %r49;
shl.b32 %r51, %r50, 2;
cvt.s64.s32 %rd1, %r50;
shl.b32 %r270, %r271, 3;
mul.hi.u32 %r52, %r45, %r27;
add.s32 %r53, %r52, %r45;
shr.u32 %r54, %r53, %r28;
mad.lo.s32 %r55, %r54, %r30, %r271;
mad.lo.s32 %r56, %r268, %r24, %r55;
mov.u32 %r57, %ctaid.x;
mad.lo.s32 %r58, %r57, %r20, %r56;
mul.wide.s32 %rd17, %r58, 66;
mul.wide.s32 %rd18, %r51, 2;
add.s64 %rd19, %rd17, %rd18;
cvta.to.global.u64 %rd20, %rd10;
add.s64 %rd21, %rd20, %rd19;
add.s64 %rd44, %rd21, 4;
add.s64 %rd43, %rd20, %rd17;
cvta.to.global.u64 %rd4, %rd11;
mul.lo.s32 %r59, %r269, %r25;
cvt.u64.u32 %rd22, %r59;
mul.lo.s32 %r60, %r45, %r31;
cvt.u64.u32 %rd23, %r60;
add.s64 %rd5, %rd22, %rd23;
mov.f32 %f27, 0f00000000;

$L__BB203_5:
cvt.u64.u32 %rd24, %r270;
add.s64 %rd25, %rd5, %rd24;
ld.global.nc.u16 %rs2, [%rd44+-2];
shr.u16 %rs3, %rs2, 8;
ld.global.nc.u16 %rs4, [%rd44];
shr.u16 %rs5, %rs4, 8;
ld.global.nc.u16 %rs6, [%rd44+2];
cvt.u32.u16 %r110, %rs6;
ld.global.nc.u16 %rs7, [%rd44+4];
cvt.u32.u16 %r111, %rs7;
mov.b32 %r112, {%rs6, %rs7};
add.s64 %rd26, %rd25, %rd1;
mul.lo.s64 %rd27, %rd26, 36;
shl.b16 %rs8, %rs2, 3;
cvt.u64.u16 %rd28, %rs8;
and.b64 %rd29, %rd28, 2040;
mov.u64 %rd30, iq2xxs_grid;
add.s64 %rd31, %rd30, %rd29;
ld.global.nc.v2.u32 {%r113, %r114}, [%rd31];
mov.u32 %r66, 0;
and.b32 %r117, %r110, 255;
popc.b32 %r118, %r117;
shl.b32 %r119, %r118, 7;
and.b32 %r120, %r119, 128;
xor.b32 %r121, %r120, %r117;
mul.lo.s32 %r122, %r121, 16843009;
and.b32 %r123, %r122, 134480385;
or.b32 %r124, %r123, -2139062144;
add.s32 %r125, %r124, -16843009;
or.b32 %r61, %r125, %r123;

prmt.b32 %r61, %r61, 0, 0xba98;

xor.b32 %r126, %r113, %r61;
or.b32 %r127, %r126, -2139062144;
and.b32 %r128, %r61, 2139062143;
and.b32 %r129, %r113, -2139062144;
sub.s32 %r130, %r127, %r128;
xor.b32 %r131, %r129, %r130;
xor.b32 %r64, %r131, -2139062144;
add.s64 %rd32, %rd4, %rd27;
ld.global.nc.u32 %r65, [%rd32+4];

dp4a.s32.s32 %r63, %r64, %r65, %r66;

and.b32 %r132, %r122, -2143281136;
or.b32 %r133, %r132, -2139062144;
add.s32 %r134, %r133, -16843009;
or.b32 %r67, %r134, %r132;

prmt.b32 %r67, %r67, 0, 0xba98;

xor.b32 %r135, %r67, %r114;
or.b32 %r136, %r135, -2139062144;
and.b32 %r137, %r67, 2139062143;
and.b32 %r138, %r114, -2139062144;
sub.s32 %r139, %r136, %r137;
xor.b32 %r140, %r138, %r139;
xor.b32 %r70, %r140, -2139062144;
ld.global.nc.u32 %r71, [%rd32+8];

dp4a.s32.s32 %r69, %r70, %r71, %r63;

cvt.u32.u16 %r141, %rs3;
mul.wide.u32 %rd33, %r141, 8;
add.s64 %rd34, %rd30, %rd33;
ld.global.nc.v2.u32 {%r142, %r143}, [%rd34];
shr.u32 %r146, %r110, 7;
and.b32 %r147, %r146, 255;
popc.b32 %r148, %r147;
shl.b32 %r149, %r148, 7;
and.b32 %r150, %r149, 128;
xor.b32 %r151, %r150, %r147;
mul.lo.s32 %r152, %r151, 16843009;
and.b32 %r153, %r152, 134480385;
or.b32 %r154, %r153, -2139062144;
add.s32 %r155, %r154, -16843009;
or.b32 %r73, %r155, %r153;

prmt.b32 %r73, %r73, 0, 0xba98;

xor.b32 %r156, %r142, %r73;
or.b32 %r157, %r156, -2139062144;
and.b32 %r158, %r73, 2139062143;
and.b32 %r159, %r142, -2139062144;
sub.s32 %r160, %r157, %r158;
xor.b32 %r161, %r159, %r160;
xor.b32 %r76, %r161, -2139062144;
ld.global.nc.u32 %r77, [%rd32+12];

dp4a.s32.s32 %r75, %r76, %r77, %r69;

and.b32 %r162, %r152, -2143281136;
or.b32 %r163, %r162, -2139062144;
add.s32 %r164, %r163, -16843009;
or.b32 %r79, %r164, %r162;

prmt.b32 %r79, %r79, 0, 0xba98;

xor.b32 %r165, %r79, %r143;
or.b32 %r166, %r165, -2139062144;
and.b32 %r167, %r79, 2139062143;
and.b32 %r168, %r143, -2139062144;
sub.s32 %r169, %r166, %r167;
xor.b32 %r170, %r168, %r169;
xor.b32 %r82, %r170, -2139062144;
ld.global.nc.u32 %r83, [%rd32+16];

dp4a.s32.s32 %r81, %r82, %r83, %r75;

shl.b16 %rs9, %rs4, 3;
cvt.u64.u16 %rd35, %rs9;
and.b64 %rd36, %rd35, 2040;
add.s64 %rd37, %rd30, %rd36;
ld.global.nc.v2.u32 {%r171, %r172}, [%rd37];
shr.u32 %r175, %r112, 14;
and.b32 %r176, %r175, 255;
popc.b32 %r177, %r176;
shl.b32 %r178, %r177, 7;
and.b32 %r179, %r178, 128;
xor.b32 %r180, %r179, %r176;
mul.lo.s32 %r181, %r180, 16843009;
and.b32 %r182, %r181, 134480385;
or.b32 %r183, %r182, -2139062144;
add.s32 %r184, %r183, -16843009;
or.b32 %r85, %r184, %r182;

prmt.b32 %r85, %r85, 0, 0xba98;

xor.b32 %r185, %r171, %r85;
or.b32 %r186, %r185, -2139062144;
and.b32 %r187, %r85, 2139062143;
and.b32 %r188, %r171, -2139062144;
sub.s32 %r189, %r186, %r187;
xor.b32 %r190, %r188, %r189;
xor.b32 %r88, %r190, -2139062144;
ld.global.nc.u32 %r89, [%rd32+20];

dp4a.s32.s32 %r87, %r88, %r89, %r81;

and.b32 %r191, %r181, -2143281136;
or.b32 %r192, %r191, -2139062144;
add.s32 %r193, %r192, -16843009;
or.b32 %r91, %r193, %r191;

prmt.b32 %r91, %r91, 0, 0xba98;

xor.b32 %r194, %r91, %r172;
or.b32 %r195, %r194, -2139062144;
and.b32 %r196, %r91, 2139062143;
and.b32 %r197, %r172, -2139062144;
sub.s32 %r198, %r195, %r196;
xor.b32 %r199, %r197, %r198;
xor.b32 %r94, %r199, -2139062144;
ld.global.nc.u32 %r95, [%rd32+24];

dp4a.s32.s32 %r93, %r94, %r95, %r87;

cvt.u32.u16 %r200, %rs5;
mul.wide.u32 %rd38, %r200, 8;
add.s64 %rd39, %rd30, %rd38;
ld.global.nc.v2.u32 {%r201, %r202}, [%rd39];
shr.u32 %r205, %r111, 5;
and.b32 %r206, %r205, 255;
popc.b32 %r207, %r206;
shl.b32 %r208, %r207, 7;
and.b32 %r209, %r208, 128;
xor.b32 %r210, %r209, %r206;
mul.lo.s32 %r211, %r210, 16843009;
and.b32 %r212, %r211, 134480385;
or.b32 %r213, %r212, -2139062144;
add.s32 %r214, %r213, -16843009;
or.b32 %r97, %r214, %r212;

prmt.b32 %r97, %r97, 0, 0xba98;

xor.b32 %r215, %r201, %r97;
or.b32 %r216, %r215, -2139062144;
and.b32 %r217, %r97, 2139062143;
and.b32 %r218, %r201, -2139062144;
sub.s32 %r219, %r216, %r217;
xor.b32 %r220, %r218, %r219;
xor.b32 %r100, %r220, -2139062144;
ld.global.nc.u32 %r101, [%rd32+28];

dp4a.s32.s32 %r99, %r100, %r101, %r93;

and.b32 %r221, %r211, -2143281136;
or.b32 %r222, %r221, -2139062144;
add.s32 %r223, %r222, -16843009;
or.b32 %r103, %r223, %r221;

prmt.b32 %r103, %r103, 0, 0xba98;

xor.b32 %r224, %r103, %r202;
or.b32 %r225, %r224, -2139062144;
and.b32 %r226, %r103, 2139062143;
and.b32 %r227, %r202, -2139062144;
sub.s32 %r228, %r225, %r226;
xor.b32 %r229, %r227, %r228;
xor.b32 %r106, %r229, -2139062144;
ld.global.nc.u32 %r107, [%rd32+32];

dp4a.s32.s32 %r105, %r106, %r107, %r99;

shr.u32 %r230, %r111, 11;
or.b32 %r231, %r230, 1;
mul.lo.s32 %r232, %r105, %r231;
shr.s32 %r233, %r232, 31;
shr.u32 %r234, %r233, 29;
add.s32 %r235, %r232, %r234;
shr.s32 %r236, %r235, 3;
ld.global.nc.u16 %rs1, [%rd43];

{ cvt.f32.f16 %f7, %rs1;}


ld.global.nc.u32 %r109, [%rd32];

{.reg .f16 low,high;
mov.b32 {low,high},%r109;
cvt.f32.f16 %f8, low;}


mul.ftz.f32 %f9, %f7, %f8;
cvt.rn.f32.s32 %f10, %r236;
fma.rn.ftz.f32 %f27, %f9, %f10, %f27;
add.s32 %r270, %r270, 128;
add.s64 %rd44, %rd44, 1056;
add.s64 %rd43, %rd43, 1056;
add.s32 %r271, %r271, 16;
setp.lt.s32 %p4, %r271, %r3;
@%p4 bra $L__BB203_5;

$L__BB203_6:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB203_8;

shl.b32 %r238, %r1, 7;
mov.u32 %r239, _ZZ13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r240, %r239, %r238;
shl.b32 %r241, %r35, 2;
add.s32 %r242, %r240, %r241;
st.shared.f32 [%r242+-128], %f27;

$L__BB203_8:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB203_11;

shl.b32 %r244, %r35, 2;
mov.u32 %r245, _ZZ13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r246, %r245, %r244;
mov.u32 %r247, 2;
ld.shared.f32 %f11, [%r246];
add.ftz.f32 %f12, %f27, %f11;
ld.shared.f32 %f13, [%r246+128];
add.ftz.f32 %f14, %f12, %f13;
ld.shared.f32 %f15, [%r246+256];
add.ftz.f32 %f16, %f14, %f15;
mov.b32 %r248, %f16;
mov.u32 %r249, 31;
mov.u32 %r250, 16;
mov.u32 %r251, -1;
shfl.sync.bfly.b32 %r252|%p7, %r248, %r250, %r249, %r251;
mov.b32 %f17, %r252;
add.ftz.f32 %f18, %f16, %f17;
mov.b32 %r253, %f18;
mov.u32 %r254, 8;
shfl.sync.bfly.b32 %r255|%p8, %r253, %r254, %r249, %r251;
mov.b32 %f19, %r255;
add.ftz.f32 %f20, %f18, %f19;
mov.b32 %r256, %f20;
mov.u32 %r257, 4;
shfl.sync.bfly.b32 %r258|%p9, %r256, %r257, %r249, %r251;
mov.b32 %f21, %r258;
add.ftz.f32 %f22, %f20, %f21;
mov.b32 %r259, %f22;
shfl.sync.bfly.b32 %r260|%p10, %r259, %r247, %r249, %r251;
mov.b32 %f23, %r260;
add.ftz.f32 %f4, %f22, %f23;
mov.b32 %r261, %f4;
mov.u32 %r262, 1;
shfl.sync.bfly.b32 %r16|%p1, %r261, %r262, %r249, %r251;
setp.ne.s32 %p11, %r35, 0;
@%p11 bra $L__BB203_11;

mov.u32 %r263, %ctaid.x;
mov.u32 %r264, %ctaid.y;
mad.lo.s32 %r265, %r264, %r26, %r263;
mov.u32 %r266, %ctaid.z;
mad.lo.s32 %r267, %r266, %r32, %r265;
cvta.to.global.u64 %rd40, %rd13;
mul.wide.u32 %rd41, %r267, 4;
add.s64 %rd42, %rd40, %rd41;
mov.b32 %f24, %r16;
add.ftz.f32 %f25, %f4, %f24;
st.global.f32 [%rd42], %f25;

$L__BB203_11:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type16ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type16ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type16ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type16ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type16ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type16ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type16ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type16ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type16ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<35>;
.reg .b16 %rs<21>;
.reg .f32 %f<109>;
.reg .b32 %r<680>;
.reg .b64 %rd<75>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type16ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type16ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type16ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type16ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type16ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r15, [_Z13mul_mat_vec_qIL9ggml_type16ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r16, [_Z13mul_mat_vec_qIL9ggml_type16ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r17, [_Z13mul_mat_vec_qIL9ggml_type16ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type16ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type16ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type16ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type16ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type16ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type16ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type16ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type16ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r19, [_Z13mul_mat_vec_qIL9ggml_type16ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r18, [_Z13mul_mat_vec_qIL9ggml_type16ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r31, %r1, 5;
mov.u32 %r32, %tid.x;
add.s32 %r33, %r31, %r32;
mov.u32 %r34, %ctaid.x;
shl.b32 %r2, %r34, 1;
shr.s32 %r35, %r33, 31;
shr.u32 %r36, %r35, 29;
add.s32 %r37, %r33, %r36;
shr.s32 %r679, %r37, 3;
shr.u32 %r4, %r30, 8;
setp.ge.s32 %p3, %r679, %r4;
mov.f32 %f105, 0f00000000;
mov.f32 %f106, %f105;
mov.f32 %f107, %f105;
mov.f32 %f108, %f105;
@%p3 bra $L__BB204_3;

mov.u32 %r38, %ctaid.y;
mov.u32 %r39, %ctaid.z;
mul.lo.s32 %r40, %r38, %r22;
cvt.u64.u32 %rd19, %r40;
mul.lo.s32 %r41, %r39, %r28;
cvt.u64.u32 %rd20, %r41;
add.s64 %rd1, %rd20, %rd19;
mul.hi.u32 %r42, %r39, %r24;
add.s32 %r43, %r42, %r39;
shr.u32 %r44, %r43, %r25;
mul.lo.s32 %r45, %r44, %r27;
mul.hi.u32 %r46, %r38, %r18;
add.s32 %r47, %r46, %r38;
shr.u32 %r48, %r47, %r19;
mul.lo.s32 %r49, %r48, %r21;
cvta.to.global.u64 %rd21, %rd16;
and.b32 %r57, %r37, -8;
sub.s32 %r58, %r33, %r57;
shl.b32 %r59, %r58, 2;
cvt.s64.s32 %rd2, %r58;
shl.b32 %r677, %r679, 3;
add.s32 %r678, %r16, %r677;
add.s32 %r60, %r679, %r45;
add.s32 %r61, %r60, %r49;
add.s32 %r62, %r2, 1;
mad.lo.s32 %r63, %r15, %r62, %r61;
mul.wide.s32 %rd22, %r63, 66;
add.s64 %rd74, %rd21, %rd22;
mad.lo.s32 %r64, %r2, %r15, %r679;
add.s32 %r65, %r64, %r45;
add.s32 %r66, %r65, %r49;
mul.wide.s32 %rd23, %r66, 66;
add.s64 %rd24, %rd21, %rd23;
add.s64 %rd73, %rd24, 4;
mul.wide.s32 %rd25, %r59, 2;
add.s64 %rd26, %rd23, %rd25;
add.s64 %rd27, %rd21, %rd26;
add.s64 %rd72, %rd27, 8;
or.b64 %rd6, %rd25, 4;
cvta.to.global.u64 %rd7, %rd17;
mov.f32 %f105, 0f00000000;

$L__BB204_2:
cvt.u64.u32 %rd28, %r677;
add.s64 %rd29, %rd1, %rd28;
add.s64 %rd30, %rd29, %rd2;
mul.lo.s64 %rd31, %rd30, 36;
ld.global.nc.u16 %rs5, [%rd72+-6];
shr.u16 %rs6, %rs5, 8;
ld.global.nc.u16 %rs7, [%rd72+-4];
shr.u16 %rs8, %rs7, 8;
ld.global.nc.u16 %rs9, [%rd72+-2];
cvt.u32.u16 %r263, %rs9;
ld.global.nc.u16 %rs10, [%rd72];
cvt.u32.u16 %r264, %rs10;
mov.b32 %r265, {%rs9, %rs10};
shl.b16 %rs11, %rs5, 3;
cvt.u64.u16 %rd32, %rs11;
and.b64 %rd33, %rd32, 2040;
mov.u64 %rd34, iq2xxs_grid;
add.s64 %rd35, %rd34, %rd33;
ld.global.nc.v2.u32 {%r266, %r267}, [%rd35];
not.b32 %r269, %r266;
mov.u32 %r219, 0;
not.b32 %r271, %r267;
and.b32 %r272, %r263, 255;
popc.b32 %r273, %r272;
shl.b32 %r274, %r273, 7;
and.b32 %r275, %r274, 128;
xor.b32 %r276, %r275, %r272;
mul.lo.s32 %r277, %r276, 16843009;
and.b32 %r278, %r277, 134480385;
or.b32 %r279, %r278, -2139062144;
add.s32 %r280, %r279, -16843009;
or.b32 %r165, %r280, %r278;
mov.u32 %r67, %r165;

prmt.b32 %r67, %r67, 0, 0xba98;

xor.b32 %r281, %r266, %r67;
or.b32 %r282, %r281, -2139062144;
and.b32 %r283, %r67, 2139062143;
and.b32 %r284, %r269, -2139062144;
sub.s32 %r285, %r282, %r283;
xor.b32 %r70, %r284, %r285;
add.s64 %rd36, %rd7, %rd31;
ld.global.nc.u32 %r71, [%rd36+4];

dp4a.s32.s32 %r69, %r70, %r71, %r219;

and.b32 %r286, %r277, -2143281136;
or.b32 %r287, %r286, -2139062144;
add.s32 %r288, %r287, -16843009;
or.b32 %r171, %r288, %r286;
mov.u32 %r73, %r171;

prmt.b32 %r73, %r73, 0, 0xba98;

xor.b32 %r289, %r73, %r267;
or.b32 %r290, %r289, -2139062144;
and.b32 %r291, %r73, 2139062143;
and.b32 %r292, %r271, -2139062144;
sub.s32 %r293, %r290, %r291;
xor.b32 %r76, %r292, %r293;
ld.global.nc.u32 %r77, [%rd36+8];

dp4a.s32.s32 %r75, %r76, %r77, %r69;

cvt.u32.u16 %r294, %rs6;
mul.wide.u32 %rd37, %r294, 8;
add.s64 %rd38, %rd34, %rd37;
ld.global.nc.v2.u32 {%r295, %r296}, [%rd38];
not.b32 %r298, %r295;
not.b32 %r300, %r296;
shr.u32 %r301, %r263, 7;
and.b32 %r302, %r301, 255;
popc.b32 %r303, %r302;
shl.b32 %r304, %r303, 7;
and.b32 %r305, %r304, 128;
xor.b32 %r306, %r305, %r302;
mul.lo.s32 %r307, %r306, 16843009;
and.b32 %r308, %r307, 134480385;
or.b32 %r309, %r308, -2139062144;
add.s32 %r310, %r309, -16843009;
or.b32 %r177, %r310, %r308;
mov.u32 %r79, %r177;

prmt.b32 %r79, %r79, 0, 0xba98;

xor.b32 %r311, %r295, %r79;
or.b32 %r312, %r311, -2139062144;
and.b32 %r313, %r79, 2139062143;
and.b32 %r314, %r298, -2139062144;
sub.s32 %r315, %r312, %r313;
xor.b32 %r82, %r314, %r315;
ld.global.nc.u32 %r83, [%rd36+12];

dp4a.s32.s32 %r81, %r82, %r83, %r75;

and.b32 %r316, %r307, -2143281136;
or.b32 %r317, %r316, -2139062144;
add.s32 %r318, %r317, -16843009;
or.b32 %r183, %r318, %r316;
mov.u32 %r85, %r183;

prmt.b32 %r85, %r85, 0, 0xba98;

xor.b32 %r319, %r85, %r296;
or.b32 %r320, %r319, -2139062144;
and.b32 %r321, %r85, 2139062143;
and.b32 %r322, %r300, -2139062144;
sub.s32 %r323, %r320, %r321;
xor.b32 %r88, %r322, %r323;
ld.global.nc.u32 %r89, [%rd36+16];

dp4a.s32.s32 %r87, %r88, %r89, %r81;

shl.b16 %rs12, %rs7, 3;
cvt.u64.u16 %rd39, %rs12;
and.b64 %rd40, %rd39, 2040;
add.s64 %rd41, %rd34, %rd40;
ld.global.nc.v2.u32 {%r324, %r325}, [%rd41];
not.b32 %r327, %r324;
not.b32 %r329, %r325;
shr.u32 %r330, %r265, 14;
and.b32 %r331, %r330, 255;
popc.b32 %r332, %r331;
shl.b32 %r333, %r332, 7;
and.b32 %r334, %r333, 128;
xor.b32 %r335, %r334, %r331;
mul.lo.s32 %r336, %r335, 16843009;
and.b32 %r337, %r336, 134480385;
or.b32 %r338, %r337, -2139062144;
add.s32 %r339, %r338, -16843009;
or.b32 %r189, %r339, %r337;
mov.u32 %r91, %r189;

prmt.b32 %r91, %r91, 0, 0xba98;

xor.b32 %r340, %r324, %r91;
or.b32 %r341, %r340, -2139062144;
and.b32 %r342, %r91, 2139062143;
and.b32 %r343, %r327, -2139062144;
sub.s32 %r344, %r341, %r342;
xor.b32 %r94, %r343, %r344;
ld.global.nc.u32 %r95, [%rd36+20];

dp4a.s32.s32 %r93, %r94, %r95, %r87;

and.b32 %r345, %r336, -2143281136;
or.b32 %r346, %r345, -2139062144;
add.s32 %r347, %r346, -16843009;
or.b32 %r195, %r347, %r345;
mov.u32 %r97, %r195;

prmt.b32 %r97, %r97, 0, 0xba98;

xor.b32 %r348, %r97, %r325;
or.b32 %r349, %r348, -2139062144;
and.b32 %r350, %r97, 2139062143;
and.b32 %r351, %r329, -2139062144;
sub.s32 %r352, %r349, %r350;
xor.b32 %r100, %r351, %r352;
ld.global.nc.u32 %r101, [%rd36+24];

dp4a.s32.s32 %r99, %r100, %r101, %r93;

cvt.u32.u16 %r353, %rs8;
mul.wide.u32 %rd42, %r353, 8;
add.s64 %rd43, %rd34, %rd42;
ld.global.nc.v2.u32 {%r354, %r355}, [%rd43];
not.b32 %r357, %r354;
not.b32 %r359, %r355;
shr.u32 %r360, %r264, 5;
and.b32 %r361, %r360, 255;
popc.b32 %r362, %r361;
shl.b32 %r363, %r362, 7;
and.b32 %r364, %r363, 128;
xor.b32 %r365, %r364, %r361;
mul.lo.s32 %r366, %r365, 16843009;
and.b32 %r367, %r366, 134480385;
or.b32 %r368, %r367, -2139062144;
add.s32 %r369, %r368, -16843009;
or.b32 %r201, %r369, %r367;
mov.u32 %r103, %r201;

prmt.b32 %r103, %r103, 0, 0xba98;

xor.b32 %r370, %r354, %r103;
or.b32 %r371, %r370, -2139062144;
and.b32 %r372, %r103, 2139062143;
and.b32 %r373, %r357, -2139062144;
sub.s32 %r374, %r371, %r372;
xor.b32 %r106, %r373, %r374;
ld.global.nc.u32 %r107, [%rd36+28];

dp4a.s32.s32 %r105, %r106, %r107, %r99;

and.b32 %r375, %r366, -2143281136;
or.b32 %r376, %r375, -2139062144;
add.s32 %r377, %r376, -16843009;
or.b32 %r207, %r377, %r375;
mov.u32 %r109, %r207;

prmt.b32 %r109, %r109, 0, 0xba98;

xor.b32 %r378, %r109, %r355;
or.b32 %r379, %r378, -2139062144;
and.b32 %r380, %r109, 2139062143;
and.b32 %r381, %r359, -2139062144;
sub.s32 %r382, %r379, %r380;
xor.b32 %r112, %r381, %r382;
ld.global.nc.u32 %r113, [%rd36+32];

dp4a.s32.s32 %r111, %r112, %r113, %r105;

shr.u32 %r383, %r264, 11;
or.b32 %r384, %r383, 1;
mul.lo.s32 %r385, %r111, %r384;
shr.s32 %r386, %r385, 31;
shr.u32 %r387, %r386, 29;
add.s32 %r388, %r385, %r387;
shr.s32 %r389, %r388, 3;
ld.global.nc.u16 %rs1, [%rd73+-4];

{ cvt.f32.f16 %f25, %rs1;}


ld.global.nc.u32 %r115, [%rd36];

{.reg .f16 low,high;
mov.b32 {low,high},%r115;
cvt.f32.f16 %f26, low;}


mul.ftz.f32 %f33, %f25, %f26;
cvt.rn.f32.s32 %f34, %r389;
fma.rn.ftz.f32 %f108, %f33, %f34, %f108;
add.s64 %rd44, %rd74, %rd6;
ld.global.nc.u16 %rs13, [%rd44+-2];
shr.u16 %rs14, %rs13, 8;
ld.global.nc.u16 %rs15, [%rd44];
shr.u16 %rs16, %rs15, 8;
ld.global.nc.u16 %rs17, [%rd44+2];
cvt.u32.u16 %r390, %rs17;
ld.global.nc.u16 %rs18, [%rd44+4];
cvt.u32.u16 %r391, %rs18;
mov.b32 %r392, {%rs17, %rs18};
shl.b16 %rs19, %rs13, 3;
cvt.u64.u16 %rd45, %rs19;
and.b64 %rd46, %rd45, 2040;
add.s64 %rd47, %rd34, %rd46;
ld.global.nc.v2.u32 {%r393, %r394}, [%rd47];
not.b32 %r396, %r393;
not.b32 %r398, %r394;
and.b32 %r399, %r390, 255;
popc.b32 %r400, %r399;
shl.b32 %r401, %r400, 7;
and.b32 %r402, %r401, 128;
xor.b32 %r403, %r402, %r399;
mul.lo.s32 %r404, %r403, 16843009;
and.b32 %r405, %r404, 134480385;
or.b32 %r406, %r405, -2139062144;
add.s32 %r407, %r406, -16843009;
or.b32 %r214, %r407, %r405;
mov.u32 %r116, %r214;

prmt.b32 %r116, %r116, 0, 0xba98;

xor.b32 %r408, %r393, %r116;
or.b32 %r409, %r408, -2139062144;
and.b32 %r410, %r116, 2139062143;
and.b32 %r411, %r396, -2139062144;
sub.s32 %r412, %r409, %r410;
xor.b32 %r119, %r411, %r412;

dp4a.s32.s32 %r118, %r119, %r71, %r219;

and.b32 %r413, %r404, -2143281136;
or.b32 %r414, %r413, -2139062144;
add.s32 %r415, %r414, -16843009;
or.b32 %r220, %r415, %r413;
mov.u32 %r122, %r220;

prmt.b32 %r122, %r122, 0, 0xba98;

xor.b32 %r416, %r122, %r394;
or.b32 %r417, %r416, -2139062144;
and.b32 %r418, %r122, 2139062143;
and.b32 %r419, %r398, -2139062144;
sub.s32 %r420, %r417, %r418;
xor.b32 %r125, %r419, %r420;

dp4a.s32.s32 %r124, %r125, %r77, %r118;

cvt.u32.u16 %r421, %rs14;
mul.wide.u32 %rd48, %r421, 8;
add.s64 %rd49, %rd34, %rd48;
ld.global.nc.v2.u32 {%r422, %r423}, [%rd49];
not.b32 %r425, %r422;
not.b32 %r427, %r423;
shr.u32 %r428, %r390, 7;
and.b32 %r429, %r428, 255;
popc.b32 %r430, %r429;
shl.b32 %r431, %r430, 7;
and.b32 %r432, %r431, 128;
xor.b32 %r433, %r432, %r429;
mul.lo.s32 %r434, %r433, 16843009;
and.b32 %r435, %r434, 134480385;
or.b32 %r436, %r435, -2139062144;
add.s32 %r437, %r436, -16843009;
or.b32 %r226, %r437, %r435;
mov.u32 %r128, %r226;

prmt.b32 %r128, %r128, 0, 0xba98;

xor.b32 %r438, %r422, %r128;
or.b32 %r439, %r438, -2139062144;
and.b32 %r440, %r128, 2139062143;
and.b32 %r441, %r425, -2139062144;
sub.s32 %r442, %r439, %r440;
xor.b32 %r131, %r441, %r442;

dp4a.s32.s32 %r130, %r131, %r83, %r124;

and.b32 %r443, %r434, -2143281136;
or.b32 %r444, %r443, -2139062144;
add.s32 %r445, %r444, -16843009;
or.b32 %r232, %r445, %r443;
mov.u32 %r134, %r232;

prmt.b32 %r134, %r134, 0, 0xba98;

xor.b32 %r446, %r134, %r423;
or.b32 %r447, %r446, -2139062144;
and.b32 %r448, %r134, 2139062143;
and.b32 %r449, %r427, -2139062144;
sub.s32 %r450, %r447, %r448;
xor.b32 %r137, %r449, %r450;

dp4a.s32.s32 %r136, %r137, %r89, %r130;

shl.b16 %rs20, %rs15, 3;
cvt.u64.u16 %rd50, %rs20;
and.b64 %rd51, %rd50, 2040;
add.s64 %rd52, %rd34, %rd51;
ld.global.nc.v2.u32 {%r451, %r452}, [%rd52];
not.b32 %r454, %r451;
not.b32 %r456, %r452;
shr.u32 %r457, %r392, 14;
and.b32 %r458, %r457, 255;
popc.b32 %r459, %r458;
shl.b32 %r460, %r459, 7;
and.b32 %r461, %r460, 128;
xor.b32 %r462, %r461, %r458;
mul.lo.s32 %r463, %r462, 16843009;
and.b32 %r464, %r463, 134480385;
or.b32 %r465, %r464, -2139062144;
add.s32 %r466, %r465, -16843009;
or.b32 %r238, %r466, %r464;
mov.u32 %r140, %r238;

prmt.b32 %r140, %r140, 0, 0xba98;

xor.b32 %r467, %r451, %r140;
or.b32 %r468, %r467, -2139062144;
and.b32 %r469, %r140, 2139062143;
and.b32 %r470, %r454, -2139062144;
sub.s32 %r471, %r468, %r469;
xor.b32 %r143, %r470, %r471;

dp4a.s32.s32 %r142, %r143, %r95, %r136;

and.b32 %r472, %r463, -2143281136;
or.b32 %r473, %r472, -2139062144;
add.s32 %r474, %r473, -16843009;
or.b32 %r244, %r474, %r472;
mov.u32 %r146, %r244;

prmt.b32 %r146, %r146, 0, 0xba98;

xor.b32 %r475, %r146, %r452;
or.b32 %r476, %r475, -2139062144;
and.b32 %r477, %r146, 2139062143;
and.b32 %r478, %r456, -2139062144;
sub.s32 %r479, %r476, %r477;
xor.b32 %r149, %r478, %r479;

dp4a.s32.s32 %r148, %r149, %r101, %r142;

cvt.u32.u16 %r480, %rs16;
mul.wide.u32 %rd53, %r480, 8;
add.s64 %rd54, %rd34, %rd53;
ld.global.nc.v2.u32 {%r481, %r482}, [%rd54];
not.b32 %r484, %r481;
not.b32 %r486, %r482;
shr.u32 %r487, %r391, 5;
and.b32 %r488, %r487, 255;
popc.b32 %r489, %r488;
shl.b32 %r490, %r489, 7;
and.b32 %r491, %r490, 128;
xor.b32 %r492, %r491, %r488;
mul.lo.s32 %r493, %r492, 16843009;
and.b32 %r494, %r493, 134480385;
or.b32 %r495, %r494, -2139062144;
add.s32 %r496, %r495, -16843009;
or.b32 %r250, %r496, %r494;
mov.u32 %r152, %r250;

prmt.b32 %r152, %r152, 0, 0xba98;

xor.b32 %r497, %r481, %r152;
or.b32 %r498, %r497, -2139062144;
and.b32 %r499, %r152, 2139062143;
and.b32 %r500, %r484, -2139062144;
sub.s32 %r501, %r498, %r499;
xor.b32 %r155, %r500, %r501;

dp4a.s32.s32 %r154, %r155, %r107, %r148;

and.b32 %r502, %r493, -2143281136;
or.b32 %r503, %r502, -2139062144;
add.s32 %r504, %r503, -16843009;
or.b32 %r256, %r504, %r502;
mov.u32 %r158, %r256;

prmt.b32 %r158, %r158, 0, 0xba98;

xor.b32 %r505, %r158, %r482;
or.b32 %r506, %r505, -2139062144;
and.b32 %r507, %r158, 2139062143;
and.b32 %r508, %r486, -2139062144;
sub.s32 %r509, %r506, %r507;
xor.b32 %r161, %r508, %r509;

dp4a.s32.s32 %r160, %r161, %r113, %r154;

shr.u32 %r510, %r391, 11;
or.b32 %r511, %r510, 1;
mul.lo.s32 %r512, %r160, %r511;
shr.s32 %r513, %r512, 31;
shr.u32 %r514, %r513, 29;
add.s32 %r515, %r512, %r514;
shr.s32 %r516, %r515, 3;
ld.global.nc.u16 %rs2, [%rd74];

{ cvt.f32.f16 %f27, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r115;
cvt.f32.f16 %f28, low;}


mul.ftz.f32 %f35, %f27, %f28;
cvt.rn.f32.s32 %f36, %r516;
fma.rn.ftz.f32 %f107, %f35, %f36, %f107;
cvt.u64.u32 %rd55, %r678;
add.s64 %rd56, %rd1, %rd55;
add.s64 %rd57, %rd56, %rd2;
mul.lo.s64 %rd58, %rd57, 36;

prmt.b32 %r165, %r165, 0, 0xba98;

xor.b32 %r517, %r266, %r165;
or.b32 %r518, %r517, -2139062144;
and.b32 %r519, %r165, 2139062143;
sub.s32 %r520, %r518, %r519;
xor.b32 %r168, %r284, %r520;
add.s64 %rd59, %rd7, %rd58;
ld.global.nc.u32 %r169, [%rd59+4];

dp4a.s32.s32 %r167, %r168, %r169, %r219;


prmt.b32 %r171, %r171, 0, 0xba98;

xor.b32 %r521, %r171, %r267;
or.b32 %r522, %r521, -2139062144;
and.b32 %r523, %r171, 2139062143;
sub.s32 %r524, %r522, %r523;
xor.b32 %r174, %r292, %r524;
ld.global.nc.u32 %r175, [%rd59+8];

dp4a.s32.s32 %r173, %r174, %r175, %r167;


prmt.b32 %r177, %r177, 0, 0xba98;

xor.b32 %r525, %r295, %r177;
or.b32 %r526, %r525, -2139062144;
and.b32 %r527, %r177, 2139062143;
sub.s32 %r528, %r526, %r527;
xor.b32 %r180, %r314, %r528;
ld.global.nc.u32 %r181, [%rd59+12];

dp4a.s32.s32 %r179, %r180, %r181, %r173;


prmt.b32 %r183, %r183, 0, 0xba98;

xor.b32 %r529, %r183, %r296;
or.b32 %r530, %r529, -2139062144;
and.b32 %r531, %r183, 2139062143;
sub.s32 %r532, %r530, %r531;
xor.b32 %r186, %r322, %r532;
ld.global.nc.u32 %r187, [%rd59+16];

dp4a.s32.s32 %r185, %r186, %r187, %r179;


prmt.b32 %r189, %r189, 0, 0xba98;

xor.b32 %r533, %r324, %r189;
or.b32 %r534, %r533, -2139062144;
and.b32 %r535, %r189, 2139062143;
sub.s32 %r536, %r534, %r535;
xor.b32 %r192, %r343, %r536;
ld.global.nc.u32 %r193, [%rd59+20];

dp4a.s32.s32 %r191, %r192, %r193, %r185;


prmt.b32 %r195, %r195, 0, 0xba98;

xor.b32 %r537, %r195, %r325;
or.b32 %r538, %r537, -2139062144;
and.b32 %r539, %r195, 2139062143;
sub.s32 %r540, %r538, %r539;
xor.b32 %r198, %r351, %r540;
ld.global.nc.u32 %r199, [%rd59+24];

dp4a.s32.s32 %r197, %r198, %r199, %r191;


prmt.b32 %r201, %r201, 0, 0xba98;

xor.b32 %r541, %r354, %r201;
or.b32 %r542, %r541, -2139062144;
and.b32 %r543, %r201, 2139062143;
sub.s32 %r544, %r542, %r543;
xor.b32 %r204, %r373, %r544;
ld.global.nc.u32 %r205, [%rd59+28];

dp4a.s32.s32 %r203, %r204, %r205, %r197;


prmt.b32 %r207, %r207, 0, 0xba98;

xor.b32 %r545, %r207, %r355;
or.b32 %r546, %r545, -2139062144;
and.b32 %r547, %r207, 2139062143;
sub.s32 %r548, %r546, %r547;
xor.b32 %r210, %r381, %r548;
ld.global.nc.u32 %r211, [%rd59+32];

dp4a.s32.s32 %r209, %r210, %r211, %r203;

mul.lo.s32 %r549, %r209, %r384;
shr.s32 %r550, %r549, 31;
shr.u32 %r551, %r550, 29;
add.s32 %r552, %r549, %r551;
shr.s32 %r553, %r552, 3;

{ cvt.f32.f16 %f29, %rs1;}


ld.global.nc.u32 %r213, [%rd59];

{.reg .f16 low,high;
mov.b32 {low,high},%r213;
cvt.f32.f16 %f30, low;}


mul.ftz.f32 %f37, %f29, %f30;
cvt.rn.f32.s32 %f38, %r553;
fma.rn.ftz.f32 %f106, %f37, %f38, %f106;

prmt.b32 %r214, %r214, 0, 0xba98;

xor.b32 %r554, %r393, %r214;
or.b32 %r555, %r554, -2139062144;
and.b32 %r556, %r214, 2139062143;
sub.s32 %r557, %r555, %r556;
xor.b32 %r217, %r411, %r557;

dp4a.s32.s32 %r216, %r217, %r169, %r219;


prmt.b32 %r220, %r220, 0, 0xba98;

xor.b32 %r558, %r220, %r394;
or.b32 %r559, %r558, -2139062144;
and.b32 %r560, %r220, 2139062143;
sub.s32 %r561, %r559, %r560;
xor.b32 %r223, %r419, %r561;

dp4a.s32.s32 %r222, %r223, %r175, %r216;


prmt.b32 %r226, %r226, 0, 0xba98;

xor.b32 %r562, %r422, %r226;
or.b32 %r563, %r562, -2139062144;
and.b32 %r564, %r226, 2139062143;
sub.s32 %r565, %r563, %r564;
xor.b32 %r229, %r441, %r565;

dp4a.s32.s32 %r228, %r229, %r181, %r222;


prmt.b32 %r232, %r232, 0, 0xba98;

xor.b32 %r566, %r232, %r423;
or.b32 %r567, %r566, -2139062144;
and.b32 %r568, %r232, 2139062143;
sub.s32 %r569, %r567, %r568;
xor.b32 %r235, %r449, %r569;

dp4a.s32.s32 %r234, %r235, %r187, %r228;


prmt.b32 %r238, %r238, 0, 0xba98;

xor.b32 %r570, %r451, %r238;
or.b32 %r571, %r570, -2139062144;
and.b32 %r572, %r238, 2139062143;
sub.s32 %r573, %r571, %r572;
xor.b32 %r241, %r470, %r573;

dp4a.s32.s32 %r240, %r241, %r193, %r234;


prmt.b32 %r244, %r244, 0, 0xba98;

xor.b32 %r574, %r244, %r452;
or.b32 %r575, %r574, -2139062144;
and.b32 %r576, %r244, 2139062143;
sub.s32 %r577, %r575, %r576;
xor.b32 %r247, %r478, %r577;

dp4a.s32.s32 %r246, %r247, %r199, %r240;


prmt.b32 %r250, %r250, 0, 0xba98;

xor.b32 %r578, %r481, %r250;
or.b32 %r579, %r578, -2139062144;
and.b32 %r580, %r250, 2139062143;
sub.s32 %r581, %r579, %r580;
xor.b32 %r253, %r500, %r581;

dp4a.s32.s32 %r252, %r253, %r205, %r246;


prmt.b32 %r256, %r256, 0, 0xba98;

xor.b32 %r582, %r256, %r482;
or.b32 %r583, %r582, -2139062144;
and.b32 %r584, %r256, 2139062143;
sub.s32 %r585, %r583, %r584;
xor.b32 %r259, %r508, %r585;

dp4a.s32.s32 %r258, %r259, %r211, %r252;

mul.lo.s32 %r586, %r258, %r511;
shr.s32 %r587, %r586, 31;
shr.u32 %r588, %r587, 29;
add.s32 %r589, %r586, %r588;
shr.s32 %r590, %r589, 3;

{ cvt.f32.f16 %f31, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r213;
cvt.f32.f16 %f32, low;}


mul.ftz.f32 %f39, %f31, %f32;
cvt.rn.f32.s32 %f40, %r590;
fma.rn.ftz.f32 %f105, %f39, %f40, %f105;
add.s32 %r678, %r678, 128;
add.s32 %r677, %r677, 128;
add.s64 %rd74, %rd74, 1056;
add.s64 %rd73, %rd73, 1056;
add.s64 %rd72, %rd72, 1056;
add.s32 %r679, %r679, 16;
setp.lt.s32 %p4, %r679, %r4;
@%p4 bra $L__BB204_2;

$L__BB204_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB204_5;

shl.b32 %r592, %r1, 9;
mov.u32 %r593, _ZZ13mul_mat_vec_qIL9ggml_type16ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r594, %r593, %r592;
shl.b32 %r595, %r32, 2;
add.s32 %r596, %r594, %r595;
st.shared.f32 [%r596+-512], %f108;
st.shared.f32 [%r596+-384], %f107;
st.shared.f32 [%r596+-256], %f106;
st.shared.f32 [%r596+-128], %f105;

$L__BB204_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB204_14;

mov.u32 %r597, %ctaid.z;
mov.u32 %r598, %ctaid.y;
mad.lo.s32 %r599, %r598, %r23, %r2;
mad.lo.s32 %r600, %r597, %r29, %r599;
cvt.u64.u32 %rd14, %r600;
shl.b32 %r602, %r32, 2;
mov.u32 %r603, _ZZ13mul_mat_vec_qIL9ggml_type16ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r13, %r603, %r602;
mov.u32 %r604, 2;
ld.shared.f32 %f41, [%r13];
add.ftz.f32 %f42, %f41, %f108;
ld.shared.f32 %f43, [%r13+512];
add.ftz.f32 %f44, %f43, %f42;
ld.shared.f32 %f45, [%r13+1024];
add.ftz.f32 %f46, %f45, %f44;
mov.b32 %r605, %f46;
mov.u32 %r606, 31;
mov.u32 %r607, 16;
mov.u32 %r608, -1;
shfl.sync.bfly.b32 %r609|%p7, %r605, %r607, %r606, %r608;
mov.b32 %f47, %r609;
add.ftz.f32 %f48, %f46, %f47;
mov.b32 %r610, %f48;
mov.u32 %r611, 8;
shfl.sync.bfly.b32 %r612|%p8, %r610, %r611, %r606, %r608;
mov.b32 %f49, %r612;
add.ftz.f32 %f50, %f48, %f49;
mov.b32 %r613, %f50;
mov.u32 %r614, 4;
shfl.sync.bfly.b32 %r615|%p9, %r613, %r614, %r606, %r608;
mov.b32 %f51, %r615;
add.ftz.f32 %f52, %f50, %f51;
mov.b32 %r616, %f52;
shfl.sync.bfly.b32 %r617|%p10, %r616, %r604, %r606, %r608;
mov.b32 %f53, %r617;
add.ftz.f32 %f54, %f52, %f53;
mov.b32 %r618, %f54;
mov.u32 %r619, 1;
shfl.sync.bfly.b32 %r620|%p11, %r618, %r619, %r606, %r608;
mov.b32 %f55, %r620;
add.ftz.f32 %f13, %f54, %f55;
setp.eq.s32 %p12, %r32, 0;
setp.lt.u32 %p13, %r2, %r17;
and.pred %p1, %p12, %p13;
cvta.to.global.u64 %rd60, %rd18;
mul.wide.u32 %rd61, %r600, 4;
add.s64 %rd15, %rd60, %rd61;
not.pred %p14, %p1;
@%p14 bra $L__BB204_8;

st.global.f32 [%rd15], %f13;

$L__BB204_8:
ld.shared.f32 %f56, [%r13+128];
add.ftz.f32 %f57, %f56, %f107;
ld.shared.f32 %f58, [%r13+640];
add.ftz.f32 %f59, %f58, %f57;
ld.shared.f32 %f60, [%r13+1152];
add.ftz.f32 %f61, %f60, %f59;
mov.b32 %r621, %f61;
shfl.sync.bfly.b32 %r625|%p15, %r621, %r607, %r606, %r608;
mov.b32 %f62, %r625;
add.ftz.f32 %f63, %f61, %f62;
mov.b32 %r626, %f63;
shfl.sync.bfly.b32 %r628|%p16, %r626, %r611, %r606, %r608;
mov.b32 %f64, %r628;
add.ftz.f32 %f65, %f63, %f64;
mov.b32 %r629, %f65;
shfl.sync.bfly.b32 %r631|%p17, %r629, %r614, %r606, %r608;
mov.b32 %f66, %r631;
add.ftz.f32 %f67, %f65, %f66;
mov.b32 %r632, %f67;
shfl.sync.bfly.b32 %r634|%p18, %r632, %r604, %r606, %r608;
mov.b32 %f68, %r634;
add.ftz.f32 %f69, %f67, %f68;
mov.b32 %r635, %f69;
shfl.sync.bfly.b32 %r637|%p19, %r635, %r619, %r606, %r608;
mov.b32 %f70, %r637;
add.ftz.f32 %f14, %f69, %f70;
setp.eq.s32 %p20, %r32, 1;
or.b32 %r641, %r2, 1;
setp.lt.u32 %p21, %r641, %r17;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB204_10;

st.global.f32 [%rd15+4], %f14;

$L__BB204_10:
ld.shared.f32 %f71, [%r13+256];
add.ftz.f32 %f72, %f71, %f106;
ld.shared.f32 %f73, [%r13+768];
add.ftz.f32 %f74, %f73, %f72;
ld.shared.f32 %f75, [%r13+1280];
add.ftz.f32 %f76, %f75, %f74;
mov.b32 %r642, %f76;
mov.u32 %r643, 31;
mov.u32 %r644, 16;
mov.u32 %r645, -1;
shfl.sync.bfly.b32 %r646|%p23, %r642, %r644, %r643, %r645;
mov.b32 %f77, %r646;
add.ftz.f32 %f78, %f76, %f77;
mov.b32 %r647, %f78;
mov.u32 %r648, 8;
shfl.sync.bfly.b32 %r649|%p24, %r647, %r648, %r643, %r645;
mov.b32 %f79, %r649;
add.ftz.f32 %f80, %f78, %f79;
mov.b32 %r650, %f80;
mov.u32 %r651, 4;
shfl.sync.bfly.b32 %r652|%p25, %r650, %r651, %r643, %r645;
mov.b32 %f81, %r652;
add.ftz.f32 %f82, %f80, %f81;
mov.b32 %r653, %f82;
mov.u32 %r654, 2;
shfl.sync.bfly.b32 %r655|%p26, %r653, %r654, %r643, %r645;
mov.b32 %f83, %r655;
add.ftz.f32 %f84, %f82, %f83;
mov.b32 %r656, %f84;
mov.u32 %r657, 1;
shfl.sync.bfly.b32 %r658|%p27, %r656, %r657, %r643, %r645;
mov.b32 %f85, %r658;
add.ftz.f32 %f15, %f84, %f85;
@%p14 bra $L__BB204_12;

cvt.u64.u32 %rd62, %r17;
add.s64 %rd63, %rd62, %rd14;
shl.b64 %rd65, %rd63, 2;
add.s64 %rd66, %rd60, %rd65;
st.global.f32 [%rd66], %f15;

$L__BB204_12:
ld.shared.f32 %f86, [%r13+384];
add.ftz.f32 %f87, %f86, %f105;
ld.shared.f32 %f88, [%r13+896];
add.ftz.f32 %f89, %f88, %f87;
ld.shared.f32 %f90, [%r13+1408];
add.ftz.f32 %f91, %f90, %f89;
mov.b32 %r659, %f91;
shfl.sync.bfly.b32 %r663|%p29, %r659, %r644, %r643, %r645;
mov.b32 %f92, %r663;
add.ftz.f32 %f93, %f91, %f92;
mov.b32 %r664, %f93;
shfl.sync.bfly.b32 %r666|%p30, %r664, %r648, %r643, %r645;
mov.b32 %f94, %r666;
add.ftz.f32 %f95, %f93, %f94;
mov.b32 %r667, %f95;
shfl.sync.bfly.b32 %r669|%p31, %r667, %r651, %r643, %r645;
mov.b32 %f96, %r669;
add.ftz.f32 %f97, %f95, %f96;
mov.b32 %r670, %f97;
shfl.sync.bfly.b32 %r672|%p32, %r670, %r654, %r643, %r645;
mov.b32 %f98, %r672;
add.ftz.f32 %f99, %f97, %f98;
mov.b32 %r673, %f99;
shfl.sync.bfly.b32 %r675|%p33, %r673, %r657, %r643, %r645;
mov.b32 %f100, %r675;
add.ftz.f32 %f16, %f99, %f100;
@%p22 bra $L__BB204_14;

add.s32 %r676, %r17, 1;
cvt.u64.u32 %rd67, %r676;
add.s64 %rd68, %rd67, %rd14;
shl.b64 %rd70, %rd68, 2;
add.s64 %rd71, %rd60, %rd70;
st.global.f32 [%rd71], %f16;

$L__BB204_14:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type16ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type16ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type16ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type16ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type16ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type16ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type16ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type16ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type16ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<50>;
.reg .b16 %rs<23>;
.reg .f32 %f<163>;
.reg .b32 %r<907>;
.reg .b64 %rd<88>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type16ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[2304];

ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type16ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type16ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type16ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type16ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r18, [_Z13mul_mat_vec_qIL9ggml_type16ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r19, [_Z13mul_mat_vec_qIL9ggml_type16ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type16ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type16ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type16ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type16ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type16ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type16ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type16ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type16ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type16ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type16ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type16ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r34, %r1, 5;
mov.u32 %r35, %tid.x;
add.s32 %r2, %r34, %r35;
shr.s32 %r36, %r2, 31;
shr.u32 %r37, %r36, 29;
add.s32 %r38, %r2, %r37;
shr.s32 %r906, %r38, 3;
shr.u32 %r4, %r33, 8;
setp.ge.s32 %p2, %r906, %r4;
mov.f32 %f157, 0f00000000;
mov.f32 %f158, %f157;
mov.f32 %f159, %f157;
mov.f32 %f160, %f157;
mov.f32 %f161, %f157;
mov.f32 %f162, %f157;
@%p2 bra $L__BB205_3;

mov.u32 %r39, %ctaid.y;
mov.u32 %r40, %ctaid.z;
mul.lo.s32 %r41, %r39, %r25;
cvt.u64.u32 %rd19, %r41;
mul.lo.s32 %r42, %r40, %r31;
cvt.u64.u32 %rd20, %r42;
add.s64 %rd1, %rd20, %rd19;
mul.hi.u32 %r43, %r40, %r27;
add.s32 %r44, %r43, %r40;
shr.u32 %r45, %r44, %r28;
mul.lo.s32 %r46, %r45, %r30;
mul.hi.u32 %r47, %r39, %r21;
add.s32 %r48, %r47, %r39;
shr.u32 %r49, %r48, %r22;
mul.lo.s32 %r50, %r49, %r24;
cvta.to.global.u64 %rd21, %rd16;
and.b32 %r54, %r38, -8;
sub.s32 %r55, %r2, %r54;
shl.b32 %r56, %r55, 2;
cvt.s64.s32 %rd2, %r55;
mov.u32 %r57, %ctaid.x;
shl.b32 %r58, %r57, 1;
shl.b32 %r59, %r19, 1;
shl.b32 %r903, %r906, 3;
add.s32 %r905, %r903, %r59;
add.s32 %r904, %r19, %r903;
add.s32 %r60, %r906, %r46;
add.s32 %r61, %r60, %r50;
or.b32 %r62, %r58, 1;
mad.lo.s32 %r63, %r18, %r62, %r61;
mul.wide.s32 %rd22, %r63, 66;
add.s64 %rd87, %rd21, %rd22;
mad.lo.s32 %r64, %r58, %r18, %r906;
add.s32 %r65, %r64, %r46;
add.s32 %r66, %r65, %r50;
mul.wide.s32 %rd23, %r66, 66;
add.s64 %rd24, %rd21, %rd23;
add.s64 %rd86, %rd24, 4;
mul.wide.s32 %rd25, %r56, 2;
add.s64 %rd26, %rd23, %rd25;
add.s64 %rd27, %rd21, %rd26;
add.s64 %rd85, %rd27, 8;
or.b64 %rd6, %rd25, 4;
cvta.to.global.u64 %rd7, %rd17;
mov.f32 %f157, 0f00000000;

$L__BB205_2:
cvt.u64.u32 %rd28, %r903;
add.s64 %rd29, %rd1, %rd28;
add.s64 %rd30, %rd29, %rd2;
mul.lo.s64 %rd31, %rd30, 36;
ld.global.nc.u16 %rs7, [%rd85+-6];
shr.u16 %rs8, %rs7, 8;
ld.global.nc.u16 %rs9, [%rd85+-4];
shr.u16 %rs10, %rs9, 8;
ld.global.nc.u16 %rs11, [%rd85+-2];
cvt.u32.u16 %r361, %rs11;
ld.global.nc.u16 %rs12, [%rd85];
cvt.u32.u16 %r362, %rs12;
mov.b32 %r363, {%rs11, %rs12};
shl.b16 %rs13, %rs7, 3;
cvt.u64.u16 %rd32, %rs13;
and.b64 %rd33, %rd32, 2040;
mov.u64 %rd34, iq2xxs_grid;
add.s64 %rd35, %rd34, %rd33;
ld.global.nc.v2.u32 {%r364, %r365}, [%rd35];
not.b32 %r367, %r364;
mov.u32 %r317, 0;
not.b32 %r369, %r365;
and.b32 %r370, %r361, 255;
popc.b32 %r371, %r370;
shl.b32 %r372, %r371, 7;
and.b32 %r373, %r372, 128;
xor.b32 %r374, %r373, %r370;
mul.lo.s32 %r375, %r374, 16843009;
and.b32 %r376, %r375, 134480385;
or.b32 %r377, %r376, -2139062144;
add.s32 %r378, %r377, -16843009;
or.b32 %r263, %r378, %r376;
mov.u32 %r67, %r263;

prmt.b32 %r67, %r67, 0, 0xba98;

xor.b32 %r379, %r364, %r67;
or.b32 %r380, %r379, -2139062144;
and.b32 %r381, %r67, 2139062143;
and.b32 %r382, %r367, -2139062144;
sub.s32 %r383, %r380, %r381;
xor.b32 %r70, %r382, %r383;
add.s64 %rd36, %rd7, %rd31;
ld.global.nc.u32 %r71, [%rd36+4];

dp4a.s32.s32 %r69, %r70, %r71, %r317;

and.b32 %r384, %r375, -2143281136;
or.b32 %r385, %r384, -2139062144;
add.s32 %r386, %r385, -16843009;
or.b32 %r269, %r386, %r384;
mov.u32 %r73, %r269;

prmt.b32 %r73, %r73, 0, 0xba98;

xor.b32 %r387, %r73, %r365;
or.b32 %r388, %r387, -2139062144;
and.b32 %r389, %r73, 2139062143;
and.b32 %r390, %r369, -2139062144;
sub.s32 %r391, %r388, %r389;
xor.b32 %r76, %r390, %r391;
ld.global.nc.u32 %r77, [%rd36+8];

dp4a.s32.s32 %r75, %r76, %r77, %r69;

cvt.u32.u16 %r392, %rs8;
mul.wide.u32 %rd37, %r392, 8;
add.s64 %rd38, %rd34, %rd37;
ld.global.nc.v2.u32 {%r393, %r394}, [%rd38];
not.b32 %r396, %r393;
not.b32 %r398, %r394;
shr.u32 %r399, %r361, 7;
and.b32 %r400, %r399, 255;
popc.b32 %r401, %r400;
shl.b32 %r402, %r401, 7;
and.b32 %r403, %r402, 128;
xor.b32 %r404, %r403, %r400;
mul.lo.s32 %r405, %r404, 16843009;
and.b32 %r406, %r405, 134480385;
or.b32 %r407, %r406, -2139062144;
add.s32 %r408, %r407, -16843009;
or.b32 %r275, %r408, %r406;
mov.u32 %r79, %r275;

prmt.b32 %r79, %r79, 0, 0xba98;

xor.b32 %r409, %r393, %r79;
or.b32 %r410, %r409, -2139062144;
and.b32 %r411, %r79, 2139062143;
and.b32 %r412, %r396, -2139062144;
sub.s32 %r413, %r410, %r411;
xor.b32 %r82, %r412, %r413;
ld.global.nc.u32 %r83, [%rd36+12];

dp4a.s32.s32 %r81, %r82, %r83, %r75;

and.b32 %r414, %r405, -2143281136;
or.b32 %r415, %r414, -2139062144;
add.s32 %r416, %r415, -16843009;
or.b32 %r281, %r416, %r414;
mov.u32 %r85, %r281;

prmt.b32 %r85, %r85, 0, 0xba98;

xor.b32 %r417, %r85, %r394;
or.b32 %r418, %r417, -2139062144;
and.b32 %r419, %r85, 2139062143;
and.b32 %r420, %r398, -2139062144;
sub.s32 %r421, %r418, %r419;
xor.b32 %r88, %r420, %r421;
ld.global.nc.u32 %r89, [%rd36+16];

dp4a.s32.s32 %r87, %r88, %r89, %r81;

shl.b16 %rs14, %rs9, 3;
cvt.u64.u16 %rd39, %rs14;
and.b64 %rd40, %rd39, 2040;
add.s64 %rd41, %rd34, %rd40;
ld.global.nc.v2.u32 {%r422, %r423}, [%rd41];
not.b32 %r425, %r422;
not.b32 %r427, %r423;
shr.u32 %r428, %r363, 14;
and.b32 %r429, %r428, 255;
popc.b32 %r430, %r429;
shl.b32 %r431, %r430, 7;
and.b32 %r432, %r431, 128;
xor.b32 %r433, %r432, %r429;
mul.lo.s32 %r434, %r433, 16843009;
and.b32 %r435, %r434, 134480385;
or.b32 %r436, %r435, -2139062144;
add.s32 %r437, %r436, -16843009;
or.b32 %r287, %r437, %r435;
mov.u32 %r91, %r287;

prmt.b32 %r91, %r91, 0, 0xba98;

xor.b32 %r438, %r422, %r91;
or.b32 %r439, %r438, -2139062144;
and.b32 %r440, %r91, 2139062143;
and.b32 %r441, %r425, -2139062144;
sub.s32 %r442, %r439, %r440;
xor.b32 %r94, %r441, %r442;
ld.global.nc.u32 %r95, [%rd36+20];

dp4a.s32.s32 %r93, %r94, %r95, %r87;

and.b32 %r443, %r434, -2143281136;
or.b32 %r444, %r443, -2139062144;
add.s32 %r445, %r444, -16843009;
or.b32 %r293, %r445, %r443;
mov.u32 %r97, %r293;

prmt.b32 %r97, %r97, 0, 0xba98;

xor.b32 %r446, %r97, %r423;
or.b32 %r447, %r446, -2139062144;
and.b32 %r448, %r97, 2139062143;
and.b32 %r449, %r427, -2139062144;
sub.s32 %r450, %r447, %r448;
xor.b32 %r100, %r449, %r450;
ld.global.nc.u32 %r101, [%rd36+24];

dp4a.s32.s32 %r99, %r100, %r101, %r93;

cvt.u32.u16 %r451, %rs10;
mul.wide.u32 %rd42, %r451, 8;
add.s64 %rd43, %rd34, %rd42;
ld.global.nc.v2.u32 {%r452, %r453}, [%rd43];
not.b32 %r455, %r452;
not.b32 %r457, %r453;
shr.u32 %r458, %r362, 5;
and.b32 %r459, %r458, 255;
popc.b32 %r460, %r459;
shl.b32 %r461, %r460, 7;
and.b32 %r462, %r461, 128;
xor.b32 %r463, %r462, %r459;
mul.lo.s32 %r464, %r463, 16843009;
and.b32 %r465, %r464, 134480385;
or.b32 %r466, %r465, -2139062144;
add.s32 %r467, %r466, -16843009;
or.b32 %r299, %r467, %r465;
mov.u32 %r103, %r299;

prmt.b32 %r103, %r103, 0, 0xba98;

xor.b32 %r468, %r452, %r103;
or.b32 %r469, %r468, -2139062144;
and.b32 %r470, %r103, 2139062143;
and.b32 %r471, %r455, -2139062144;
sub.s32 %r472, %r469, %r470;
xor.b32 %r106, %r471, %r472;
ld.global.nc.u32 %r107, [%rd36+28];

dp4a.s32.s32 %r105, %r106, %r107, %r99;

and.b32 %r473, %r464, -2143281136;
or.b32 %r474, %r473, -2139062144;
add.s32 %r475, %r474, -16843009;
or.b32 %r305, %r475, %r473;
mov.u32 %r109, %r305;

prmt.b32 %r109, %r109, 0, 0xba98;

xor.b32 %r476, %r109, %r453;
or.b32 %r477, %r476, -2139062144;
and.b32 %r478, %r109, 2139062143;
and.b32 %r479, %r457, -2139062144;
sub.s32 %r480, %r477, %r478;
xor.b32 %r112, %r479, %r480;
ld.global.nc.u32 %r113, [%rd36+32];

dp4a.s32.s32 %r111, %r112, %r113, %r105;

shr.u32 %r481, %r362, 11;
or.b32 %r482, %r481, 1;
mul.lo.s32 %r483, %r111, %r482;
shr.s32 %r484, %r483, 31;
shr.u32 %r485, %r484, 29;
add.s32 %r486, %r483, %r485;
shr.s32 %r487, %r486, 3;
ld.global.nc.u16 %rs1, [%rd86+-4];

{ cvt.f32.f16 %f37, %rs1;}


ld.global.nc.u32 %r115, [%rd36];

{.reg .f16 low,high;
mov.b32 {low,high},%r115;
cvt.f32.f16 %f38, low;}


mul.ftz.f32 %f49, %f37, %f38;
cvt.rn.f32.s32 %f50, %r487;
fma.rn.ftz.f32 %f162, %f49, %f50, %f162;
add.s64 %rd44, %rd87, %rd6;
ld.global.nc.u16 %rs15, [%rd44+-2];
shr.u16 %rs16, %rs15, 8;
ld.global.nc.u16 %rs17, [%rd44];
shr.u16 %rs18, %rs17, 8;
ld.global.nc.u16 %rs19, [%rd44+2];
cvt.u32.u16 %r488, %rs19;
ld.global.nc.u16 %rs20, [%rd44+4];
cvt.u32.u16 %r489, %rs20;
mov.b32 %r490, {%rs19, %rs20};
shl.b16 %rs21, %rs15, 3;
cvt.u64.u16 %rd45, %rs21;
and.b64 %rd46, %rd45, 2040;
add.s64 %rd47, %rd34, %rd46;
ld.global.nc.v2.u32 {%r491, %r492}, [%rd47];
not.b32 %r494, %r491;
not.b32 %r496, %r492;
and.b32 %r497, %r488, 255;
popc.b32 %r498, %r497;
shl.b32 %r499, %r498, 7;
and.b32 %r500, %r499, 128;
xor.b32 %r501, %r500, %r497;
mul.lo.s32 %r502, %r501, 16843009;
and.b32 %r503, %r502, 134480385;
or.b32 %r504, %r503, -2139062144;
add.s32 %r505, %r504, -16843009;
or.b32 %r312, %r505, %r503;
mov.u32 %r116, %r312;

prmt.b32 %r116, %r116, 0, 0xba98;

xor.b32 %r506, %r491, %r116;
or.b32 %r507, %r506, -2139062144;
and.b32 %r508, %r116, 2139062143;
and.b32 %r509, %r494, -2139062144;
sub.s32 %r510, %r507, %r508;
xor.b32 %r119, %r509, %r510;

dp4a.s32.s32 %r118, %r119, %r71, %r317;

and.b32 %r511, %r502, -2143281136;
or.b32 %r512, %r511, -2139062144;
add.s32 %r513, %r512, -16843009;
or.b32 %r318, %r513, %r511;
mov.u32 %r122, %r318;

prmt.b32 %r122, %r122, 0, 0xba98;

xor.b32 %r514, %r122, %r492;
or.b32 %r515, %r514, -2139062144;
and.b32 %r516, %r122, 2139062143;
and.b32 %r517, %r496, -2139062144;
sub.s32 %r518, %r515, %r516;
xor.b32 %r125, %r517, %r518;

dp4a.s32.s32 %r124, %r125, %r77, %r118;

cvt.u32.u16 %r519, %rs16;
mul.wide.u32 %rd48, %r519, 8;
add.s64 %rd49, %rd34, %rd48;
ld.global.nc.v2.u32 {%r520, %r521}, [%rd49];
not.b32 %r523, %r520;
not.b32 %r525, %r521;
shr.u32 %r526, %r488, 7;
and.b32 %r527, %r526, 255;
popc.b32 %r528, %r527;
shl.b32 %r529, %r528, 7;
and.b32 %r530, %r529, 128;
xor.b32 %r531, %r530, %r527;
mul.lo.s32 %r532, %r531, 16843009;
and.b32 %r533, %r532, 134480385;
or.b32 %r534, %r533, -2139062144;
add.s32 %r535, %r534, -16843009;
or.b32 %r324, %r535, %r533;
mov.u32 %r128, %r324;

prmt.b32 %r128, %r128, 0, 0xba98;

xor.b32 %r536, %r520, %r128;
or.b32 %r537, %r536, -2139062144;
and.b32 %r538, %r128, 2139062143;
and.b32 %r539, %r523, -2139062144;
sub.s32 %r540, %r537, %r538;
xor.b32 %r131, %r539, %r540;

dp4a.s32.s32 %r130, %r131, %r83, %r124;

and.b32 %r541, %r532, -2143281136;
or.b32 %r542, %r541, -2139062144;
add.s32 %r543, %r542, -16843009;
or.b32 %r330, %r543, %r541;
mov.u32 %r134, %r330;

prmt.b32 %r134, %r134, 0, 0xba98;

xor.b32 %r544, %r134, %r521;
or.b32 %r545, %r544, -2139062144;
and.b32 %r546, %r134, 2139062143;
and.b32 %r547, %r525, -2139062144;
sub.s32 %r548, %r545, %r546;
xor.b32 %r137, %r547, %r548;

dp4a.s32.s32 %r136, %r137, %r89, %r130;

shl.b16 %rs22, %rs17, 3;
cvt.u64.u16 %rd50, %rs22;
and.b64 %rd51, %rd50, 2040;
add.s64 %rd52, %rd34, %rd51;
ld.global.nc.v2.u32 {%r549, %r550}, [%rd52];
not.b32 %r552, %r549;
not.b32 %r554, %r550;
shr.u32 %r555, %r490, 14;
and.b32 %r556, %r555, 255;
popc.b32 %r557, %r556;
shl.b32 %r558, %r557, 7;
and.b32 %r559, %r558, 128;
xor.b32 %r560, %r559, %r556;
mul.lo.s32 %r561, %r560, 16843009;
and.b32 %r562, %r561, 134480385;
or.b32 %r563, %r562, -2139062144;
add.s32 %r564, %r563, -16843009;
or.b32 %r336, %r564, %r562;
mov.u32 %r140, %r336;

prmt.b32 %r140, %r140, 0, 0xba98;

xor.b32 %r565, %r549, %r140;
or.b32 %r566, %r565, -2139062144;
and.b32 %r567, %r140, 2139062143;
and.b32 %r568, %r552, -2139062144;
sub.s32 %r569, %r566, %r567;
xor.b32 %r143, %r568, %r569;

dp4a.s32.s32 %r142, %r143, %r95, %r136;

and.b32 %r570, %r561, -2143281136;
or.b32 %r571, %r570, -2139062144;
add.s32 %r572, %r571, -16843009;
or.b32 %r342, %r572, %r570;
mov.u32 %r146, %r342;

prmt.b32 %r146, %r146, 0, 0xba98;

xor.b32 %r573, %r146, %r550;
or.b32 %r574, %r573, -2139062144;
and.b32 %r575, %r146, 2139062143;
and.b32 %r576, %r554, -2139062144;
sub.s32 %r577, %r574, %r575;
xor.b32 %r149, %r576, %r577;

dp4a.s32.s32 %r148, %r149, %r101, %r142;

cvt.u32.u16 %r578, %rs18;
mul.wide.u32 %rd53, %r578, 8;
add.s64 %rd54, %rd34, %rd53;
ld.global.nc.v2.u32 {%r579, %r580}, [%rd54];
not.b32 %r582, %r579;
not.b32 %r584, %r580;
shr.u32 %r585, %r489, 5;
and.b32 %r586, %r585, 255;
popc.b32 %r587, %r586;
shl.b32 %r588, %r587, 7;
and.b32 %r589, %r588, 128;
xor.b32 %r590, %r589, %r586;
mul.lo.s32 %r591, %r590, 16843009;
and.b32 %r592, %r591, 134480385;
or.b32 %r593, %r592, -2139062144;
add.s32 %r594, %r593, -16843009;
or.b32 %r348, %r594, %r592;
mov.u32 %r152, %r348;

prmt.b32 %r152, %r152, 0, 0xba98;

xor.b32 %r595, %r579, %r152;
or.b32 %r596, %r595, -2139062144;
and.b32 %r597, %r152, 2139062143;
and.b32 %r598, %r582, -2139062144;
sub.s32 %r599, %r596, %r597;
xor.b32 %r155, %r598, %r599;

dp4a.s32.s32 %r154, %r155, %r107, %r148;

and.b32 %r600, %r591, -2143281136;
or.b32 %r601, %r600, -2139062144;
add.s32 %r602, %r601, -16843009;
or.b32 %r354, %r602, %r600;
mov.u32 %r158, %r354;

prmt.b32 %r158, %r158, 0, 0xba98;

xor.b32 %r603, %r158, %r580;
or.b32 %r604, %r603, -2139062144;
and.b32 %r605, %r158, 2139062143;
and.b32 %r606, %r584, -2139062144;
sub.s32 %r607, %r604, %r605;
xor.b32 %r161, %r606, %r607;

dp4a.s32.s32 %r160, %r161, %r113, %r154;

shr.u32 %r608, %r489, 11;
or.b32 %r609, %r608, 1;
mul.lo.s32 %r610, %r160, %r609;
shr.s32 %r611, %r610, 31;
shr.u32 %r612, %r611, 29;
add.s32 %r613, %r610, %r612;
shr.s32 %r614, %r613, 3;
ld.global.nc.u16 %rs2, [%rd87];

{ cvt.f32.f16 %f39, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r115;
cvt.f32.f16 %f40, low;}


mul.ftz.f32 %f51, %f39, %f40;
cvt.rn.f32.s32 %f52, %r614;
fma.rn.ftz.f32 %f161, %f51, %f52, %f161;
cvt.u64.u32 %rd55, %r904;
add.s64 %rd56, %rd1, %rd55;
add.s64 %rd57, %rd56, %rd2;
mul.lo.s64 %rd58, %rd57, 36;
mov.u32 %r165, %r263;

prmt.b32 %r165, %r165, 0, 0xba98;

xor.b32 %r615, %r364, %r165;
or.b32 %r616, %r615, -2139062144;
and.b32 %r617, %r165, 2139062143;
sub.s32 %r618, %r616, %r617;
xor.b32 %r168, %r382, %r618;
add.s64 %rd59, %rd7, %rd58;
ld.global.nc.u32 %r169, [%rd59+4];

dp4a.s32.s32 %r167, %r168, %r169, %r317;

mov.u32 %r171, %r269;

prmt.b32 %r171, %r171, 0, 0xba98;

xor.b32 %r619, %r171, %r365;
or.b32 %r620, %r619, -2139062144;
and.b32 %r621, %r171, 2139062143;
sub.s32 %r622, %r620, %r621;
xor.b32 %r174, %r390, %r622;
ld.global.nc.u32 %r175, [%rd59+8];

dp4a.s32.s32 %r173, %r174, %r175, %r167;

mov.u32 %r177, %r275;

prmt.b32 %r177, %r177, 0, 0xba98;

xor.b32 %r623, %r393, %r177;
or.b32 %r624, %r623, -2139062144;
and.b32 %r625, %r177, 2139062143;
sub.s32 %r626, %r624, %r625;
xor.b32 %r180, %r412, %r626;
ld.global.nc.u32 %r181, [%rd59+12];

dp4a.s32.s32 %r179, %r180, %r181, %r173;

mov.u32 %r183, %r281;

prmt.b32 %r183, %r183, 0, 0xba98;

xor.b32 %r627, %r183, %r394;
or.b32 %r628, %r627, -2139062144;
and.b32 %r629, %r183, 2139062143;
sub.s32 %r630, %r628, %r629;
xor.b32 %r186, %r420, %r630;
ld.global.nc.u32 %r187, [%rd59+16];

dp4a.s32.s32 %r185, %r186, %r187, %r179;

mov.u32 %r189, %r287;

prmt.b32 %r189, %r189, 0, 0xba98;

xor.b32 %r631, %r422, %r189;
or.b32 %r632, %r631, -2139062144;
and.b32 %r633, %r189, 2139062143;
sub.s32 %r634, %r632, %r633;
xor.b32 %r192, %r441, %r634;
ld.global.nc.u32 %r193, [%rd59+20];

dp4a.s32.s32 %r191, %r192, %r193, %r185;

mov.u32 %r195, %r293;

prmt.b32 %r195, %r195, 0, 0xba98;

xor.b32 %r635, %r195, %r423;
or.b32 %r636, %r635, -2139062144;
and.b32 %r637, %r195, 2139062143;
sub.s32 %r638, %r636, %r637;
xor.b32 %r198, %r449, %r638;
ld.global.nc.u32 %r199, [%rd59+24];

dp4a.s32.s32 %r197, %r198, %r199, %r191;

mov.u32 %r201, %r299;

prmt.b32 %r201, %r201, 0, 0xba98;

xor.b32 %r639, %r452, %r201;
or.b32 %r640, %r639, -2139062144;
and.b32 %r641, %r201, 2139062143;
sub.s32 %r642, %r640, %r641;
xor.b32 %r204, %r471, %r642;
ld.global.nc.u32 %r205, [%rd59+28];

dp4a.s32.s32 %r203, %r204, %r205, %r197;

mov.u32 %r207, %r305;

prmt.b32 %r207, %r207, 0, 0xba98;

xor.b32 %r643, %r207, %r453;
or.b32 %r644, %r643, -2139062144;
and.b32 %r645, %r207, 2139062143;
sub.s32 %r646, %r644, %r645;
xor.b32 %r210, %r479, %r646;
ld.global.nc.u32 %r211, [%rd59+32];

dp4a.s32.s32 %r209, %r210, %r211, %r203;

mul.lo.s32 %r647, %r209, %r482;
shr.s32 %r648, %r647, 31;
shr.u32 %r649, %r648, 29;
add.s32 %r650, %r647, %r649;
shr.s32 %r651, %r650, 3;

{ cvt.f32.f16 %f41, %rs1;}


ld.global.nc.u32 %r213, [%rd59];

{.reg .f16 low,high;
mov.b32 {low,high},%r213;
cvt.f32.f16 %f42, low;}


mul.ftz.f32 %f53, %f41, %f42;
cvt.rn.f32.s32 %f54, %r651;
fma.rn.ftz.f32 %f160, %f53, %f54, %f160;
mov.u32 %r214, %r312;

prmt.b32 %r214, %r214, 0, 0xba98;

xor.b32 %r652, %r491, %r214;
or.b32 %r653, %r652, -2139062144;
and.b32 %r654, %r214, 2139062143;
sub.s32 %r655, %r653, %r654;
xor.b32 %r217, %r509, %r655;

dp4a.s32.s32 %r216, %r217, %r169, %r317;

mov.u32 %r220, %r318;

prmt.b32 %r220, %r220, 0, 0xba98;

xor.b32 %r656, %r220, %r492;
or.b32 %r657, %r656, -2139062144;
and.b32 %r658, %r220, 2139062143;
sub.s32 %r659, %r657, %r658;
xor.b32 %r223, %r517, %r659;

dp4a.s32.s32 %r222, %r223, %r175, %r216;

mov.u32 %r226, %r324;

prmt.b32 %r226, %r226, 0, 0xba98;

xor.b32 %r660, %r520, %r226;
or.b32 %r661, %r660, -2139062144;
and.b32 %r662, %r226, 2139062143;
sub.s32 %r663, %r661, %r662;
xor.b32 %r229, %r539, %r663;

dp4a.s32.s32 %r228, %r229, %r181, %r222;

mov.u32 %r232, %r330;

prmt.b32 %r232, %r232, 0, 0xba98;

xor.b32 %r664, %r232, %r521;
or.b32 %r665, %r664, -2139062144;
and.b32 %r666, %r232, 2139062143;
sub.s32 %r667, %r665, %r666;
xor.b32 %r235, %r547, %r667;

dp4a.s32.s32 %r234, %r235, %r187, %r228;

mov.u32 %r238, %r336;

prmt.b32 %r238, %r238, 0, 0xba98;

xor.b32 %r668, %r549, %r238;
or.b32 %r669, %r668, -2139062144;
and.b32 %r670, %r238, 2139062143;
sub.s32 %r671, %r669, %r670;
xor.b32 %r241, %r568, %r671;

dp4a.s32.s32 %r240, %r241, %r193, %r234;

mov.u32 %r244, %r342;

prmt.b32 %r244, %r244, 0, 0xba98;

xor.b32 %r672, %r244, %r550;
or.b32 %r673, %r672, -2139062144;
and.b32 %r674, %r244, 2139062143;
sub.s32 %r675, %r673, %r674;
xor.b32 %r247, %r576, %r675;

dp4a.s32.s32 %r246, %r247, %r199, %r240;

mov.u32 %r250, %r348;

prmt.b32 %r250, %r250, 0, 0xba98;

xor.b32 %r676, %r579, %r250;
or.b32 %r677, %r676, -2139062144;
and.b32 %r678, %r250, 2139062143;
sub.s32 %r679, %r677, %r678;
xor.b32 %r253, %r598, %r679;

dp4a.s32.s32 %r252, %r253, %r205, %r246;

mov.u32 %r256, %r354;

prmt.b32 %r256, %r256, 0, 0xba98;

xor.b32 %r680, %r256, %r580;
or.b32 %r681, %r680, -2139062144;
and.b32 %r682, %r256, 2139062143;
sub.s32 %r683, %r681, %r682;
xor.b32 %r259, %r606, %r683;

dp4a.s32.s32 %r258, %r259, %r211, %r252;

mul.lo.s32 %r684, %r258, %r609;
shr.s32 %r685, %r684, 31;
shr.u32 %r686, %r685, 29;
add.s32 %r687, %r684, %r686;
shr.s32 %r688, %r687, 3;

{ cvt.f32.f16 %f43, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r213;
cvt.f32.f16 %f44, low;}


mul.ftz.f32 %f55, %f43, %f44;
cvt.rn.f32.s32 %f56, %r688;
fma.rn.ftz.f32 %f159, %f55, %f56, %f159;
cvt.u64.u32 %rd60, %r905;
add.s64 %rd61, %rd1, %rd60;
add.s64 %rd62, %rd61, %rd2;
mul.lo.s64 %rd63, %rd62, 36;

prmt.b32 %r263, %r263, 0, 0xba98;

xor.b32 %r689, %r364, %r263;
or.b32 %r690, %r689, -2139062144;
and.b32 %r691, %r263, 2139062143;
sub.s32 %r692, %r690, %r691;
xor.b32 %r266, %r382, %r692;
add.s64 %rd64, %rd7, %rd63;
ld.global.nc.u32 %r267, [%rd64+4];

dp4a.s32.s32 %r265, %r266, %r267, %r317;


prmt.b32 %r269, %r269, 0, 0xba98;

xor.b32 %r693, %r269, %r365;
or.b32 %r694, %r693, -2139062144;
and.b32 %r695, %r269, 2139062143;
sub.s32 %r696, %r694, %r695;
xor.b32 %r272, %r390, %r696;
ld.global.nc.u32 %r273, [%rd64+8];

dp4a.s32.s32 %r271, %r272, %r273, %r265;


prmt.b32 %r275, %r275, 0, 0xba98;

xor.b32 %r697, %r393, %r275;
or.b32 %r698, %r697, -2139062144;
and.b32 %r699, %r275, 2139062143;
sub.s32 %r700, %r698, %r699;
xor.b32 %r278, %r412, %r700;
ld.global.nc.u32 %r279, [%rd64+12];

dp4a.s32.s32 %r277, %r278, %r279, %r271;


prmt.b32 %r281, %r281, 0, 0xba98;

xor.b32 %r701, %r281, %r394;
or.b32 %r702, %r701, -2139062144;
and.b32 %r703, %r281, 2139062143;
sub.s32 %r704, %r702, %r703;
xor.b32 %r284, %r420, %r704;
ld.global.nc.u32 %r285, [%rd64+16];

dp4a.s32.s32 %r283, %r284, %r285, %r277;


prmt.b32 %r287, %r287, 0, 0xba98;

xor.b32 %r705, %r422, %r287;
or.b32 %r706, %r705, -2139062144;
and.b32 %r707, %r287, 2139062143;
sub.s32 %r708, %r706, %r707;
xor.b32 %r290, %r441, %r708;
ld.global.nc.u32 %r291, [%rd64+20];

dp4a.s32.s32 %r289, %r290, %r291, %r283;


prmt.b32 %r293, %r293, 0, 0xba98;

xor.b32 %r709, %r293, %r423;
or.b32 %r710, %r709, -2139062144;
and.b32 %r711, %r293, 2139062143;
sub.s32 %r712, %r710, %r711;
xor.b32 %r296, %r449, %r712;
ld.global.nc.u32 %r297, [%rd64+24];

dp4a.s32.s32 %r295, %r296, %r297, %r289;


prmt.b32 %r299, %r299, 0, 0xba98;

xor.b32 %r713, %r452, %r299;
or.b32 %r714, %r713, -2139062144;
and.b32 %r715, %r299, 2139062143;
sub.s32 %r716, %r714, %r715;
xor.b32 %r302, %r471, %r716;
ld.global.nc.u32 %r303, [%rd64+28];

dp4a.s32.s32 %r301, %r302, %r303, %r295;


prmt.b32 %r305, %r305, 0, 0xba98;

xor.b32 %r717, %r305, %r453;
or.b32 %r718, %r717, -2139062144;
and.b32 %r719, %r305, 2139062143;
sub.s32 %r720, %r718, %r719;
xor.b32 %r308, %r479, %r720;
ld.global.nc.u32 %r309, [%rd64+32];

dp4a.s32.s32 %r307, %r308, %r309, %r301;

mul.lo.s32 %r721, %r307, %r482;
shr.s32 %r722, %r721, 31;
shr.u32 %r723, %r722, 29;
add.s32 %r724, %r721, %r723;
shr.s32 %r725, %r724, 3;

{ cvt.f32.f16 %f45, %rs1;}


ld.global.nc.u32 %r311, [%rd64];

{.reg .f16 low,high;
mov.b32 {low,high},%r311;
cvt.f32.f16 %f46, low;}


mul.ftz.f32 %f57, %f45, %f46;
cvt.rn.f32.s32 %f58, %r725;
fma.rn.ftz.f32 %f158, %f57, %f58, %f158;

prmt.b32 %r312, %r312, 0, 0xba98;

xor.b32 %r726, %r491, %r312;
or.b32 %r727, %r726, -2139062144;
and.b32 %r728, %r312, 2139062143;
sub.s32 %r729, %r727, %r728;
xor.b32 %r315, %r509, %r729;

dp4a.s32.s32 %r314, %r315, %r267, %r317;


prmt.b32 %r318, %r318, 0, 0xba98;

xor.b32 %r730, %r318, %r492;
or.b32 %r731, %r730, -2139062144;
and.b32 %r732, %r318, 2139062143;
sub.s32 %r733, %r731, %r732;
xor.b32 %r321, %r517, %r733;

dp4a.s32.s32 %r320, %r321, %r273, %r314;


prmt.b32 %r324, %r324, 0, 0xba98;

xor.b32 %r734, %r520, %r324;
or.b32 %r735, %r734, -2139062144;
and.b32 %r736, %r324, 2139062143;
sub.s32 %r737, %r735, %r736;
xor.b32 %r327, %r539, %r737;

dp4a.s32.s32 %r326, %r327, %r279, %r320;


prmt.b32 %r330, %r330, 0, 0xba98;

xor.b32 %r738, %r330, %r521;
or.b32 %r739, %r738, -2139062144;
and.b32 %r740, %r330, 2139062143;
sub.s32 %r741, %r739, %r740;
xor.b32 %r333, %r547, %r741;

dp4a.s32.s32 %r332, %r333, %r285, %r326;


prmt.b32 %r336, %r336, 0, 0xba98;

xor.b32 %r742, %r549, %r336;
or.b32 %r743, %r742, -2139062144;
and.b32 %r744, %r336, 2139062143;
sub.s32 %r745, %r743, %r744;
xor.b32 %r339, %r568, %r745;

dp4a.s32.s32 %r338, %r339, %r291, %r332;


prmt.b32 %r342, %r342, 0, 0xba98;

xor.b32 %r746, %r342, %r550;
or.b32 %r747, %r746, -2139062144;
and.b32 %r748, %r342, 2139062143;
sub.s32 %r749, %r747, %r748;
xor.b32 %r345, %r576, %r749;

dp4a.s32.s32 %r344, %r345, %r297, %r338;


prmt.b32 %r348, %r348, 0, 0xba98;

xor.b32 %r750, %r579, %r348;
or.b32 %r751, %r750, -2139062144;
and.b32 %r752, %r348, 2139062143;
sub.s32 %r753, %r751, %r752;
xor.b32 %r351, %r598, %r753;

dp4a.s32.s32 %r350, %r351, %r303, %r344;


prmt.b32 %r354, %r354, 0, 0xba98;

xor.b32 %r754, %r354, %r580;
or.b32 %r755, %r754, -2139062144;
and.b32 %r756, %r354, 2139062143;
sub.s32 %r757, %r755, %r756;
xor.b32 %r357, %r606, %r757;

dp4a.s32.s32 %r356, %r357, %r309, %r350;

mul.lo.s32 %r758, %r356, %r609;
shr.s32 %r759, %r758, 31;
shr.u32 %r760, %r759, 29;
add.s32 %r761, %r758, %r760;
shr.s32 %r762, %r761, 3;

{ cvt.f32.f16 %f47, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r311;
cvt.f32.f16 %f48, low;}


mul.ftz.f32 %f59, %f47, %f48;
cvt.rn.f32.s32 %f60, %r762;
fma.rn.ftz.f32 %f157, %f59, %f60, %f157;
add.s32 %r905, %r905, 128;
add.s32 %r904, %r904, 128;
add.s32 %r903, %r903, 128;
add.s64 %rd87, %rd87, 1056;
add.s64 %rd86, %rd86, 1056;
add.s64 %rd85, %rd85, 1056;
add.s32 %r906, %r906, 16;
setp.lt.s32 %p3, %r906, %r4;
@%p3 bra $L__BB205_2;

$L__BB205_3:
setp.eq.s32 %p4, %r1, 0;
@%p4 bra $L__BB205_5;

mov.u32 %r764, _ZZ13mul_mat_vec_qIL9ggml_type16ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r765, %r1, 768, %r764;
shl.b32 %r766, %r35, 2;
add.s32 %r767, %r765, %r766;
st.shared.f32 [%r767+-768], %f162;
st.shared.f32 [%r767+-640], %f161;
st.shared.f32 [%r767+-512], %f160;
st.shared.f32 [%r767+-384], %f159;
st.shared.f32 [%r767+-256], %f158;
st.shared.f32 [%r767+-128], %f157;

$L__BB205_5:
setp.ne.s32 %p5, %r1, 0;
bar.sync 0;
@%p5 bra $L__BB205_18;

mov.u32 %r768, %ctaid.z;
mov.u32 %r769, %ctaid.y;
mov.u32 %r770, %ctaid.x;
shl.b32 %r771, %r770, 1;
mov.u32 %r772, 1;
mad.lo.s32 %r773, %r769, %r26, %r771;
mad.lo.s32 %r774, %r768, %r32, %r773;
cvt.u64.u32 %rd14, %r774;
shl.b32 %r776, %r35, 2;
mov.u32 %r777, _ZZ13mul_mat_vec_qIL9ggml_type16ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r16, %r777, %r776;
mov.u32 %r778, 2;
ld.shared.f32 %f61, [%r16];
add.ftz.f32 %f62, %f61, %f162;
ld.shared.f32 %f63, [%r16+768];
add.ftz.f32 %f64, %f63, %f62;
ld.shared.f32 %f65, [%r16+1536];
add.ftz.f32 %f66, %f65, %f64;
mov.b32 %r779, %f66;
mov.u32 %r780, 31;
mov.u32 %r781, 16;
mov.u32 %r782, -1;
shfl.sync.bfly.b32 %r783|%p6, %r779, %r781, %r780, %r782;
mov.b32 %f67, %r783;
add.ftz.f32 %f68, %f66, %f67;
mov.b32 %r784, %f68;
mov.u32 %r785, 8;
shfl.sync.bfly.b32 %r786|%p7, %r784, %r785, %r780, %r782;
mov.b32 %f69, %r786;
add.ftz.f32 %f70, %f68, %f69;
mov.b32 %r787, %f70;
mov.u32 %r788, 4;
shfl.sync.bfly.b32 %r789|%p8, %r787, %r788, %r780, %r782;
mov.b32 %f71, %r789;
add.ftz.f32 %f72, %f70, %f71;
mov.b32 %r790, %f72;
shfl.sync.bfly.b32 %r791|%p9, %r790, %r778, %r780, %r782;
mov.b32 %f73, %r791;
add.ftz.f32 %f74, %f72, %f73;
mov.b32 %r792, %f74;
shfl.sync.bfly.b32 %r793|%p10, %r792, %r772, %r780, %r782;
mov.b32 %f75, %r793;
add.ftz.f32 %f19, %f74, %f75;
setp.ne.s32 %p11, %r35, 0;
setp.ge.u32 %p12, %r771, %r20;
or.pred %p13, %p11, %p12;
@%p13 bra $L__BB205_8;

cvta.to.global.u64 %rd65, %rd18;
mul.wide.u32 %rd66, %r774, 4;
add.s64 %rd67, %rd65, %rd66;
st.global.f32 [%rd67], %f19;

$L__BB205_8:
ld.shared.f32 %f76, [%r16+128];
add.ftz.f32 %f77, %f76, %f161;
ld.shared.f32 %f78, [%r16+896];
add.ftz.f32 %f79, %f78, %f77;
ld.shared.f32 %f80, [%r16+1664];
add.ftz.f32 %f81, %f80, %f79;
mov.b32 %r800, %f81;
shfl.sync.bfly.b32 %r804|%p14, %r800, %r781, %r780, %r782;
mov.b32 %f82, %r804;
add.ftz.f32 %f83, %f81, %f82;
mov.b32 %r805, %f83;
shfl.sync.bfly.b32 %r807|%p15, %r805, %r785, %r780, %r782;
mov.b32 %f84, %r807;
add.ftz.f32 %f85, %f83, %f84;
mov.b32 %r808, %f85;
shfl.sync.bfly.b32 %r810|%p16, %r808, %r788, %r780, %r782;
mov.b32 %f86, %r810;
add.ftz.f32 %f87, %f85, %f86;
mov.b32 %r811, %f87;
shfl.sync.bfly.b32 %r813|%p17, %r811, %r778, %r780, %r782;
mov.b32 %f88, %r813;
add.ftz.f32 %f89, %f87, %f88;
mov.b32 %r814, %f89;
shfl.sync.bfly.b32 %r816|%p18, %r814, %r772, %r780, %r782;
mov.b32 %f90, %r816;
add.ftz.f32 %f20, %f89, %f90;
setp.eq.s32 %p19, %r35, 1;
or.b32 %r820, %r771, 1;
setp.lt.u32 %p20, %r820, %r20;
and.pred %p1, %p19, %p20;
not.pred %p21, %p1;
@%p21 bra $L__BB205_10;

cvta.to.global.u64 %rd68, %rd18;
mul.wide.u32 %rd69, %r774, 4;
add.s64 %rd70, %rd68, %rd69;
st.global.f32 [%rd70+4], %f20;

$L__BB205_10:
ld.shared.f32 %f91, [%r16+256];
add.ftz.f32 %f92, %f91, %f160;
ld.shared.f32 %f93, [%r16+1024];
add.ftz.f32 %f94, %f93, %f92;
ld.shared.f32 %f95, [%r16+1792];
add.ftz.f32 %f96, %f95, %f94;
mov.b32 %r827, %f96;
mov.u32 %r828, 31;
mov.u32 %r829, 16;
mov.u32 %r830, -1;
shfl.sync.bfly.b32 %r831|%p22, %r827, %r829, %r828, %r830;
mov.b32 %f97, %r831;
add.ftz.f32 %f98, %f96, %f97;
mov.b32 %r832, %f98;
mov.u32 %r833, 8;
shfl.sync.bfly.b32 %r834|%p23, %r832, %r833, %r828, %r830;
mov.b32 %f99, %r834;
add.ftz.f32 %f100, %f98, %f99;
mov.b32 %r835, %f100;
mov.u32 %r836, 4;
shfl.sync.bfly.b32 %r837|%p24, %r835, %r836, %r828, %r830;
mov.b32 %f101, %r837;
add.ftz.f32 %f102, %f100, %f101;
mov.b32 %r838, %f102;
mov.u32 %r839, 2;
shfl.sync.bfly.b32 %r840|%p25, %r838, %r839, %r828, %r830;
mov.b32 %f103, %r840;
add.ftz.f32 %f104, %f102, %f103;
mov.b32 %r841, %f104;
mov.u32 %r842, 1;
shfl.sync.bfly.b32 %r843|%p26, %r841, %r842, %r828, %r830;
mov.b32 %f105, %r843;
add.ftz.f32 %f21, %f104, %f105;
@%p13 bra $L__BB205_12;

cvt.u64.u32 %rd71, %r20;
add.s64 %rd72, %rd71, %rd14;
cvta.to.global.u64 %rd73, %rd18;
shl.b64 %rd74, %rd72, 2;
add.s64 %rd75, %rd73, %rd74;
st.global.f32 [%rd75], %f21;

$L__BB205_12:
ld.shared.f32 %f106, [%r16+384];
add.ftz.f32 %f107, %f106, %f159;
ld.shared.f32 %f108, [%r16+1152];
add.ftz.f32 %f109, %f108, %f107;
ld.shared.f32 %f110, [%r16+1920];
add.ftz.f32 %f111, %f110, %f109;
mov.b32 %r847, %f111;
shfl.sync.bfly.b32 %r851|%p30, %r847, %r829, %r828, %r830;
mov.b32 %f112, %r851;
add.ftz.f32 %f113, %f111, %f112;
mov.b32 %r852, %f113;
shfl.sync.bfly.b32 %r854|%p31, %r852, %r833, %r828, %r830;
mov.b32 %f114, %r854;
add.ftz.f32 %f115, %f113, %f114;
mov.b32 %r855, %f115;
shfl.sync.bfly.b32 %r857|%p32, %r855, %r836, %r828, %r830;
mov.b32 %f116, %r857;
add.ftz.f32 %f117, %f115, %f116;
mov.b32 %r858, %f117;
shfl.sync.bfly.b32 %r860|%p33, %r858, %r839, %r828, %r830;
mov.b32 %f118, %r860;
add.ftz.f32 %f119, %f117, %f118;
mov.b32 %r861, %f119;
shfl.sync.bfly.b32 %r863|%p34, %r861, %r842, %r828, %r830;
mov.b32 %f120, %r863;
add.ftz.f32 %f22, %f119, %f120;
@%p21 bra $L__BB205_14;

add.s32 %r864, %r20, 1;
cvt.u64.u32 %rd76, %r864;
add.s64 %rd77, %rd76, %rd14;
cvta.to.global.u64 %rd78, %rd18;
shl.b64 %rd79, %rd77, 2;
add.s64 %rd80, %rd78, %rd79;
st.global.f32 [%rd80], %f22;

$L__BB205_14:
shl.b32 %r865, %r20, 1;
mov.u32 %r866, 1;
ld.shared.f32 %f121, [%r16+512];
add.ftz.f32 %f122, %f121, %f158;
ld.shared.f32 %f123, [%r16+1280];
add.ftz.f32 %f124, %f123, %f122;
ld.shared.f32 %f125, [%r16+2048];
add.ftz.f32 %f126, %f125, %f124;
mov.b32 %r867, %f126;
mov.u32 %r868, 31;
mov.u32 %r869, 16;
mov.u32 %r870, -1;
shfl.sync.bfly.b32 %r871|%p36, %r867, %r869, %r868, %r870;
mov.b32 %f127, %r871;
add.ftz.f32 %f128, %f126, %f127;
mov.b32 %r872, %f128;
mov.u32 %r873, 8;
shfl.sync.bfly.b32 %r874|%p37, %r872, %r873, %r868, %r870;
mov.b32 %f129, %r874;
add.ftz.f32 %f130, %f128, %f129;
mov.b32 %r875, %f130;
mov.u32 %r876, 4;
shfl.sync.bfly.b32 %r877|%p38, %r875, %r876, %r868, %r870;
mov.b32 %f131, %r877;
add.ftz.f32 %f132, %f130, %f131;
mov.b32 %r878, %f132;
mov.u32 %r879, 2;
shfl.sync.bfly.b32 %r880|%p39, %r878, %r879, %r868, %r870;
mov.b32 %f133, %r880;
add.ftz.f32 %f134, %f132, %f133;
mov.b32 %r881, %f134;
shfl.sync.bfly.b32 %r882|%p40, %r881, %r866, %r868, %r870;
mov.b32 %f135, %r882;
add.ftz.f32 %f23, %f134, %f135;
cvt.u64.u32 %rd81, %r865;
add.s64 %rd82, %rd81, %rd14;
cvta.to.global.u64 %rd83, %rd18;
shl.b64 %rd84, %rd82, 2;
add.s64 %rd15, %rd83, %rd84;
@%p13 bra $L__BB205_16;

st.global.f32 [%rd15], %f23;

$L__BB205_16:
ld.shared.f32 %f136, [%r16+640];
add.ftz.f32 %f137, %f136, %f157;
ld.shared.f32 %f138, [%r16+1408];
add.ftz.f32 %f139, %f138, %f137;
ld.shared.f32 %f140, [%r16+2176];
add.ftz.f32 %f141, %f140, %f139;
mov.b32 %r886, %f141;
shfl.sync.bfly.b32 %r890|%p44, %r886, %r869, %r868, %r870;
mov.b32 %f142, %r890;
add.ftz.f32 %f143, %f141, %f142;
mov.b32 %r891, %f143;
shfl.sync.bfly.b32 %r893|%p45, %r891, %r873, %r868, %r870;
mov.b32 %f144, %r893;
add.ftz.f32 %f145, %f143, %f144;
mov.b32 %r894, %f145;
shfl.sync.bfly.b32 %r896|%p46, %r894, %r876, %r868, %r870;
mov.b32 %f146, %r896;
add.ftz.f32 %f147, %f145, %f146;
mov.b32 %r897, %f147;
shfl.sync.bfly.b32 %r899|%p47, %r897, %r879, %r868, %r870;
mov.b32 %f148, %r899;
add.ftz.f32 %f149, %f147, %f148;
mov.b32 %r900, %f149;
shfl.sync.bfly.b32 %r902|%p48, %r900, %r866, %r868, %r870;
mov.b32 %f150, %r902;
add.ftz.f32 %f24, %f149, %f150;
@%p21 bra $L__BB205_18;

st.global.f32 [%rd15+4], %f24;

$L__BB205_18:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type16ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type16ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type16ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type16ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type16ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type16ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type16ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type16ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type16ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<64>;
.reg .b16 %rs<25>;
.reg .f32 %f<217>;
.reg .b32 %r<1178>;
.reg .b64 %rd<107>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type16ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[3072];

ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type16ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type16ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type16ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r18, [_Z13mul_mat_vec_qIL9ggml_type16ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r19, [_Z13mul_mat_vec_qIL9ggml_type16ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type16ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type16ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type16ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type16ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type16ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type16ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type16ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type16ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type16ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type16ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type16ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type16ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r34, %tid.y;
shl.b32 %r35, %r34, 5;
mov.u32 %r36, %tid.x;
add.s32 %r37, %r35, %r36;
shr.s32 %r38, %r37, 31;
shr.u32 %r39, %r38, 29;
add.s32 %r40, %r37, %r39;
shr.s32 %r1177, %r40, 3;
shr.u32 %r41, %r18, 8;
setp.ge.s32 %p2, %r1177, %r41;
mov.f32 %f209, 0f00000000;
mov.f32 %f210, %f209;
mov.f32 %f211, %f209;
mov.f32 %f212, %f209;
mov.f32 %f213, %f209;
mov.f32 %f214, %f209;
mov.f32 %f215, %f209;
mov.f32 %f216, %f209;
@%p2 bra $L__BB206_3;

mov.u32 %r42, %ctaid.y;
mov.u32 %r43, %ctaid.z;
mul.lo.s32 %r44, %r42, %r26;
cvt.u64.u32 %rd18, %r44;
mul.lo.s32 %r45, %r43, %r32;
cvt.u64.u32 %rd19, %r45;
add.s64 %rd1, %rd19, %rd18;
mul.hi.u32 %r46, %r43, %r28;
add.s32 %r47, %r46, %r43;
shr.u32 %r48, %r47, %r29;
mul.lo.s32 %r49, %r48, %r31;
mul.hi.u32 %r50, %r42, %r22;
add.s32 %r51, %r50, %r42;
shr.u32 %r52, %r51, %r23;
mul.lo.s32 %r53, %r52, %r25;
cvta.to.global.u64 %rd20, %rd15;
and.b32 %r61, %r40, -8;
sub.s32 %r62, %r37, %r61;
shl.b32 %r63, %r62, 2;
cvt.s64.s32 %rd2, %r62;
mov.u32 %r64, %ctaid.x;
shl.b32 %r65, %r64, 1;
shl.b32 %r66, %r20, 1;
shl.b32 %r1173, %r1177, 3;
add.s32 %r1176, %r1173, %r66;
mad.lo.s32 %r1175, %r20, 3, %r1173;
add.s32 %r1174, %r20, %r1173;
add.s32 %r67, %r1177, %r49;
add.s32 %r68, %r67, %r53;
or.b32 %r69, %r65, 1;
mad.lo.s32 %r70, %r19, %r69, %r68;
mul.wide.s32 %rd21, %r70, 66;
add.s64 %rd106, %rd20, %rd21;
mad.lo.s32 %r71, %r65, %r19, %r1177;
add.s32 %r72, %r71, %r49;
add.s32 %r73, %r72, %r53;
mul.wide.s32 %rd22, %r73, 66;
add.s64 %rd23, %rd20, %rd22;
add.s64 %rd105, %rd23, 4;
mul.wide.s32 %rd24, %r63, 2;
add.s64 %rd25, %rd22, %rd24;
add.s64 %rd26, %rd20, %rd25;
add.s64 %rd104, %rd26, 8;
or.b64 %rd6, %rd24, 4;
cvta.to.global.u64 %rd7, %rd16;
mov.f32 %f209, 0f00000000;

$L__BB206_2:
cvt.u64.u32 %rd27, %r1173;
add.s64 %rd28, %rd1, %rd27;
add.s64 %rd29, %rd28, %rd2;
mul.lo.s64 %rd30, %rd29, 36;
ld.global.nc.u16 %rs9, [%rd104+-6];
shr.u16 %rs10, %rs9, 8;
ld.global.nc.u16 %rs11, [%rd104+-4];
shr.u16 %rs12, %rs11, 8;
ld.global.nc.u16 %rs13, [%rd104+-2];
cvt.u32.u16 %r466, %rs13;
ld.global.nc.u16 %rs14, [%rd104];
cvt.u32.u16 %r467, %rs14;
mov.b32 %r468, {%rs13, %rs14};
shl.b16 %rs15, %rs9, 3;
cvt.u64.u16 %rd31, %rs15;
and.b64 %rd32, %rd31, 2040;
mov.u64 %rd33, iq2xxs_grid;
add.s64 %rd34, %rd33, %rd32;
ld.global.nc.v2.u32 {%r469, %r470}, [%rd34];
not.b32 %r472, %r469;
mov.u32 %r422, 0;
not.b32 %r474, %r470;
and.b32 %r475, %r466, 255;
popc.b32 %r476, %r475;
shl.b32 %r477, %r476, 7;
and.b32 %r478, %r477, 128;
xor.b32 %r479, %r478, %r475;
mul.lo.s32 %r480, %r479, 16843009;
and.b32 %r481, %r480, 134480385;
or.b32 %r482, %r481, -2139062144;
add.s32 %r483, %r482, -16843009;
or.b32 %r368, %r483, %r481;
mov.u32 %r74, %r368;

prmt.b32 %r74, %r74, 0, 0xba98;

xor.b32 %r484, %r469, %r74;
or.b32 %r485, %r484, -2139062144;
and.b32 %r486, %r74, 2139062143;
and.b32 %r487, %r472, -2139062144;
sub.s32 %r488, %r485, %r486;
xor.b32 %r77, %r487, %r488;
add.s64 %rd35, %rd7, %rd30;
ld.global.nc.u32 %r78, [%rd35+4];

dp4a.s32.s32 %r76, %r77, %r78, %r422;

and.b32 %r489, %r480, -2143281136;
or.b32 %r490, %r489, -2139062144;
add.s32 %r491, %r490, -16843009;
or.b32 %r374, %r491, %r489;
mov.u32 %r80, %r374;

prmt.b32 %r80, %r80, 0, 0xba98;

xor.b32 %r492, %r80, %r470;
or.b32 %r493, %r492, -2139062144;
and.b32 %r494, %r80, 2139062143;
and.b32 %r495, %r474, -2139062144;
sub.s32 %r496, %r493, %r494;
xor.b32 %r83, %r495, %r496;
ld.global.nc.u32 %r84, [%rd35+8];

dp4a.s32.s32 %r82, %r83, %r84, %r76;

cvt.u32.u16 %r497, %rs10;
mul.wide.u32 %rd36, %r497, 8;
add.s64 %rd37, %rd33, %rd36;
ld.global.nc.v2.u32 {%r498, %r499}, [%rd37];
not.b32 %r501, %r498;
not.b32 %r503, %r499;
shr.u32 %r504, %r466, 7;
and.b32 %r505, %r504, 255;
popc.b32 %r506, %r505;
shl.b32 %r507, %r506, 7;
and.b32 %r508, %r507, 128;
xor.b32 %r509, %r508, %r505;
mul.lo.s32 %r510, %r509, 16843009;
and.b32 %r511, %r510, 134480385;
or.b32 %r512, %r511, -2139062144;
add.s32 %r513, %r512, -16843009;
or.b32 %r380, %r513, %r511;
mov.u32 %r86, %r380;

prmt.b32 %r86, %r86, 0, 0xba98;

xor.b32 %r514, %r498, %r86;
or.b32 %r515, %r514, -2139062144;
and.b32 %r516, %r86, 2139062143;
and.b32 %r517, %r501, -2139062144;
sub.s32 %r518, %r515, %r516;
xor.b32 %r89, %r517, %r518;
ld.global.nc.u32 %r90, [%rd35+12];

dp4a.s32.s32 %r88, %r89, %r90, %r82;

and.b32 %r519, %r510, -2143281136;
or.b32 %r520, %r519, -2139062144;
add.s32 %r521, %r520, -16843009;
or.b32 %r386, %r521, %r519;
mov.u32 %r92, %r386;

prmt.b32 %r92, %r92, 0, 0xba98;

xor.b32 %r522, %r92, %r499;
or.b32 %r523, %r522, -2139062144;
and.b32 %r524, %r92, 2139062143;
and.b32 %r525, %r503, -2139062144;
sub.s32 %r526, %r523, %r524;
xor.b32 %r95, %r525, %r526;
ld.global.nc.u32 %r96, [%rd35+16];

dp4a.s32.s32 %r94, %r95, %r96, %r88;

shl.b16 %rs16, %rs11, 3;
cvt.u64.u16 %rd38, %rs16;
and.b64 %rd39, %rd38, 2040;
add.s64 %rd40, %rd33, %rd39;
ld.global.nc.v2.u32 {%r527, %r528}, [%rd40];
not.b32 %r530, %r527;
not.b32 %r532, %r528;
shr.u32 %r533, %r468, 14;
and.b32 %r534, %r533, 255;
popc.b32 %r535, %r534;
shl.b32 %r536, %r535, 7;
and.b32 %r537, %r536, 128;
xor.b32 %r538, %r537, %r534;
mul.lo.s32 %r539, %r538, 16843009;
and.b32 %r540, %r539, 134480385;
or.b32 %r541, %r540, -2139062144;
add.s32 %r542, %r541, -16843009;
or.b32 %r392, %r542, %r540;
mov.u32 %r98, %r392;

prmt.b32 %r98, %r98, 0, 0xba98;

xor.b32 %r543, %r527, %r98;
or.b32 %r544, %r543, -2139062144;
and.b32 %r545, %r98, 2139062143;
and.b32 %r546, %r530, -2139062144;
sub.s32 %r547, %r544, %r545;
xor.b32 %r101, %r546, %r547;
ld.global.nc.u32 %r102, [%rd35+20];

dp4a.s32.s32 %r100, %r101, %r102, %r94;

and.b32 %r548, %r539, -2143281136;
or.b32 %r549, %r548, -2139062144;
add.s32 %r550, %r549, -16843009;
or.b32 %r398, %r550, %r548;
mov.u32 %r104, %r398;

prmt.b32 %r104, %r104, 0, 0xba98;

xor.b32 %r551, %r104, %r528;
or.b32 %r552, %r551, -2139062144;
and.b32 %r553, %r104, 2139062143;
and.b32 %r554, %r532, -2139062144;
sub.s32 %r555, %r552, %r553;
xor.b32 %r107, %r554, %r555;
ld.global.nc.u32 %r108, [%rd35+24];

dp4a.s32.s32 %r106, %r107, %r108, %r100;

cvt.u32.u16 %r556, %rs12;
mul.wide.u32 %rd41, %r556, 8;
add.s64 %rd42, %rd33, %rd41;
ld.global.nc.v2.u32 {%r557, %r558}, [%rd42];
not.b32 %r560, %r557;
not.b32 %r562, %r558;
shr.u32 %r563, %r467, 5;
and.b32 %r564, %r563, 255;
popc.b32 %r565, %r564;
shl.b32 %r566, %r565, 7;
and.b32 %r567, %r566, 128;
xor.b32 %r568, %r567, %r564;
mul.lo.s32 %r569, %r568, 16843009;
and.b32 %r570, %r569, 134480385;
or.b32 %r571, %r570, -2139062144;
add.s32 %r572, %r571, -16843009;
or.b32 %r404, %r572, %r570;
mov.u32 %r110, %r404;

prmt.b32 %r110, %r110, 0, 0xba98;

xor.b32 %r573, %r557, %r110;
or.b32 %r574, %r573, -2139062144;
and.b32 %r575, %r110, 2139062143;
and.b32 %r576, %r560, -2139062144;
sub.s32 %r577, %r574, %r575;
xor.b32 %r113, %r576, %r577;
ld.global.nc.u32 %r114, [%rd35+28];

dp4a.s32.s32 %r112, %r113, %r114, %r106;

and.b32 %r578, %r569, -2143281136;
or.b32 %r579, %r578, -2139062144;
add.s32 %r580, %r579, -16843009;
or.b32 %r410, %r580, %r578;
mov.u32 %r116, %r410;

prmt.b32 %r116, %r116, 0, 0xba98;

xor.b32 %r581, %r116, %r558;
or.b32 %r582, %r581, -2139062144;
and.b32 %r583, %r116, 2139062143;
and.b32 %r584, %r562, -2139062144;
sub.s32 %r585, %r582, %r583;
xor.b32 %r119, %r584, %r585;
ld.global.nc.u32 %r120, [%rd35+32];

dp4a.s32.s32 %r118, %r119, %r120, %r112;

shr.u32 %r586, %r467, 11;
or.b32 %r587, %r586, 1;
mul.lo.s32 %r588, %r118, %r587;
shr.s32 %r589, %r588, 31;
shr.u32 %r590, %r589, 29;
add.s32 %r591, %r588, %r590;
shr.s32 %r592, %r591, 3;
ld.global.nc.u16 %rs1, [%rd105+-4];

{ cvt.f32.f16 %f49, %rs1;}


ld.global.nc.u32 %r122, [%rd35];

{.reg .f16 low,high;
mov.b32 {low,high},%r122;
cvt.f32.f16 %f50, low;}


mul.ftz.f32 %f65, %f49, %f50;
cvt.rn.f32.s32 %f66, %r592;
fma.rn.ftz.f32 %f216, %f65, %f66, %f216;
add.s64 %rd43, %rd106, %rd6;
ld.global.nc.u16 %rs17, [%rd43+-2];
shr.u16 %rs18, %rs17, 8;
ld.global.nc.u16 %rs19, [%rd43];
shr.u16 %rs20, %rs19, 8;
ld.global.nc.u16 %rs21, [%rd43+2];
cvt.u32.u16 %r593, %rs21;
ld.global.nc.u16 %rs22, [%rd43+4];
cvt.u32.u16 %r594, %rs22;
mov.b32 %r595, {%rs21, %rs22};
shl.b16 %rs23, %rs17, 3;
cvt.u64.u16 %rd44, %rs23;
and.b64 %rd45, %rd44, 2040;
add.s64 %rd46, %rd33, %rd45;
ld.global.nc.v2.u32 {%r596, %r597}, [%rd46];
not.b32 %r599, %r596;
not.b32 %r601, %r597;
and.b32 %r602, %r593, 255;
popc.b32 %r603, %r602;
shl.b32 %r604, %r603, 7;
and.b32 %r605, %r604, 128;
xor.b32 %r606, %r605, %r602;
mul.lo.s32 %r607, %r606, 16843009;
and.b32 %r608, %r607, 134480385;
or.b32 %r609, %r608, -2139062144;
add.s32 %r610, %r609, -16843009;
or.b32 %r417, %r610, %r608;
mov.u32 %r123, %r417;

prmt.b32 %r123, %r123, 0, 0xba98;

xor.b32 %r611, %r596, %r123;
or.b32 %r612, %r611, -2139062144;
and.b32 %r613, %r123, 2139062143;
and.b32 %r614, %r599, -2139062144;
sub.s32 %r615, %r612, %r613;
xor.b32 %r126, %r614, %r615;

dp4a.s32.s32 %r125, %r126, %r78, %r422;

and.b32 %r616, %r607, -2143281136;
or.b32 %r617, %r616, -2139062144;
add.s32 %r618, %r617, -16843009;
or.b32 %r423, %r618, %r616;
mov.u32 %r129, %r423;

prmt.b32 %r129, %r129, 0, 0xba98;

xor.b32 %r619, %r129, %r597;
or.b32 %r620, %r619, -2139062144;
and.b32 %r621, %r129, 2139062143;
and.b32 %r622, %r601, -2139062144;
sub.s32 %r623, %r620, %r621;
xor.b32 %r132, %r622, %r623;

dp4a.s32.s32 %r131, %r132, %r84, %r125;

cvt.u32.u16 %r624, %rs18;
mul.wide.u32 %rd47, %r624, 8;
add.s64 %rd48, %rd33, %rd47;
ld.global.nc.v2.u32 {%r625, %r626}, [%rd48];
not.b32 %r628, %r625;
not.b32 %r630, %r626;
shr.u32 %r631, %r593, 7;
and.b32 %r632, %r631, 255;
popc.b32 %r633, %r632;
shl.b32 %r634, %r633, 7;
and.b32 %r635, %r634, 128;
xor.b32 %r636, %r635, %r632;
mul.lo.s32 %r637, %r636, 16843009;
and.b32 %r638, %r637, 134480385;
or.b32 %r639, %r638, -2139062144;
add.s32 %r640, %r639, -16843009;
or.b32 %r429, %r640, %r638;
mov.u32 %r135, %r429;

prmt.b32 %r135, %r135, 0, 0xba98;

xor.b32 %r641, %r625, %r135;
or.b32 %r642, %r641, -2139062144;
and.b32 %r643, %r135, 2139062143;
and.b32 %r644, %r628, -2139062144;
sub.s32 %r645, %r642, %r643;
xor.b32 %r138, %r644, %r645;

dp4a.s32.s32 %r137, %r138, %r90, %r131;

and.b32 %r646, %r637, -2143281136;
or.b32 %r647, %r646, -2139062144;
add.s32 %r648, %r647, -16843009;
or.b32 %r435, %r648, %r646;
mov.u32 %r141, %r435;

prmt.b32 %r141, %r141, 0, 0xba98;

xor.b32 %r649, %r141, %r626;
or.b32 %r650, %r649, -2139062144;
and.b32 %r651, %r141, 2139062143;
and.b32 %r652, %r630, -2139062144;
sub.s32 %r653, %r650, %r651;
xor.b32 %r144, %r652, %r653;

dp4a.s32.s32 %r143, %r144, %r96, %r137;

shl.b16 %rs24, %rs19, 3;
cvt.u64.u16 %rd49, %rs24;
and.b64 %rd50, %rd49, 2040;
add.s64 %rd51, %rd33, %rd50;
ld.global.nc.v2.u32 {%r654, %r655}, [%rd51];
not.b32 %r657, %r654;
not.b32 %r659, %r655;
shr.u32 %r660, %r595, 14;
and.b32 %r661, %r660, 255;
popc.b32 %r662, %r661;
shl.b32 %r663, %r662, 7;
and.b32 %r664, %r663, 128;
xor.b32 %r665, %r664, %r661;
mul.lo.s32 %r666, %r665, 16843009;
and.b32 %r667, %r666, 134480385;
or.b32 %r668, %r667, -2139062144;
add.s32 %r669, %r668, -16843009;
or.b32 %r441, %r669, %r667;
mov.u32 %r147, %r441;

prmt.b32 %r147, %r147, 0, 0xba98;

xor.b32 %r670, %r654, %r147;
or.b32 %r671, %r670, -2139062144;
and.b32 %r672, %r147, 2139062143;
and.b32 %r673, %r657, -2139062144;
sub.s32 %r674, %r671, %r672;
xor.b32 %r150, %r673, %r674;

dp4a.s32.s32 %r149, %r150, %r102, %r143;

and.b32 %r675, %r666, -2143281136;
or.b32 %r676, %r675, -2139062144;
add.s32 %r677, %r676, -16843009;
or.b32 %r447, %r677, %r675;
mov.u32 %r153, %r447;

prmt.b32 %r153, %r153, 0, 0xba98;

xor.b32 %r678, %r153, %r655;
or.b32 %r679, %r678, -2139062144;
and.b32 %r680, %r153, 2139062143;
and.b32 %r681, %r659, -2139062144;
sub.s32 %r682, %r679, %r680;
xor.b32 %r156, %r681, %r682;

dp4a.s32.s32 %r155, %r156, %r108, %r149;

cvt.u32.u16 %r683, %rs20;
mul.wide.u32 %rd52, %r683, 8;
add.s64 %rd53, %rd33, %rd52;
ld.global.nc.v2.u32 {%r684, %r685}, [%rd53];
not.b32 %r687, %r684;
not.b32 %r689, %r685;
shr.u32 %r690, %r594, 5;
and.b32 %r691, %r690, 255;
popc.b32 %r692, %r691;
shl.b32 %r693, %r692, 7;
and.b32 %r694, %r693, 128;
xor.b32 %r695, %r694, %r691;
mul.lo.s32 %r696, %r695, 16843009;
and.b32 %r697, %r696, 134480385;
or.b32 %r698, %r697, -2139062144;
add.s32 %r699, %r698, -16843009;
or.b32 %r453, %r699, %r697;
mov.u32 %r159, %r453;

prmt.b32 %r159, %r159, 0, 0xba98;

xor.b32 %r700, %r684, %r159;
or.b32 %r701, %r700, -2139062144;
and.b32 %r702, %r159, 2139062143;
and.b32 %r703, %r687, -2139062144;
sub.s32 %r704, %r701, %r702;
xor.b32 %r162, %r703, %r704;

dp4a.s32.s32 %r161, %r162, %r114, %r155;

and.b32 %r705, %r696, -2143281136;
or.b32 %r706, %r705, -2139062144;
add.s32 %r707, %r706, -16843009;
or.b32 %r459, %r707, %r705;
mov.u32 %r165, %r459;

prmt.b32 %r165, %r165, 0, 0xba98;

xor.b32 %r708, %r165, %r685;
or.b32 %r709, %r708, -2139062144;
and.b32 %r710, %r165, 2139062143;
and.b32 %r711, %r689, -2139062144;
sub.s32 %r712, %r709, %r710;
xor.b32 %r168, %r711, %r712;

dp4a.s32.s32 %r167, %r168, %r120, %r161;

shr.u32 %r713, %r594, 11;
or.b32 %r714, %r713, 1;
mul.lo.s32 %r715, %r167, %r714;
shr.s32 %r716, %r715, 31;
shr.u32 %r717, %r716, 29;
add.s32 %r718, %r715, %r717;
shr.s32 %r719, %r718, 3;
ld.global.nc.u16 %rs2, [%rd106];

{ cvt.f32.f16 %f51, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r122;
cvt.f32.f16 %f52, low;}


mul.ftz.f32 %f67, %f51, %f52;
cvt.rn.f32.s32 %f68, %r719;
fma.rn.ftz.f32 %f215, %f67, %f68, %f215;
cvt.u64.u32 %rd54, %r1174;
add.s64 %rd55, %rd1, %rd54;
add.s64 %rd56, %rd55, %rd2;
mul.lo.s64 %rd57, %rd56, 36;
mov.u32 %r172, %r368;

prmt.b32 %r172, %r172, 0, 0xba98;

xor.b32 %r720, %r469, %r172;
or.b32 %r721, %r720, -2139062144;
and.b32 %r722, %r172, 2139062143;
sub.s32 %r723, %r721, %r722;
xor.b32 %r175, %r487, %r723;
add.s64 %rd58, %rd7, %rd57;
ld.global.nc.u32 %r176, [%rd58+4];

dp4a.s32.s32 %r174, %r175, %r176, %r422;

mov.u32 %r178, %r374;

prmt.b32 %r178, %r178, 0, 0xba98;

xor.b32 %r724, %r178, %r470;
or.b32 %r725, %r724, -2139062144;
and.b32 %r726, %r178, 2139062143;
sub.s32 %r727, %r725, %r726;
xor.b32 %r181, %r495, %r727;
ld.global.nc.u32 %r182, [%rd58+8];

dp4a.s32.s32 %r180, %r181, %r182, %r174;

mov.u32 %r184, %r380;

prmt.b32 %r184, %r184, 0, 0xba98;

xor.b32 %r728, %r498, %r184;
or.b32 %r729, %r728, -2139062144;
and.b32 %r730, %r184, 2139062143;
sub.s32 %r731, %r729, %r730;
xor.b32 %r187, %r517, %r731;
ld.global.nc.u32 %r188, [%rd58+12];

dp4a.s32.s32 %r186, %r187, %r188, %r180;

mov.u32 %r190, %r386;

prmt.b32 %r190, %r190, 0, 0xba98;

xor.b32 %r732, %r190, %r499;
or.b32 %r733, %r732, -2139062144;
and.b32 %r734, %r190, 2139062143;
sub.s32 %r735, %r733, %r734;
xor.b32 %r193, %r525, %r735;
ld.global.nc.u32 %r194, [%rd58+16];

dp4a.s32.s32 %r192, %r193, %r194, %r186;

mov.u32 %r196, %r392;

prmt.b32 %r196, %r196, 0, 0xba98;

xor.b32 %r736, %r527, %r196;
or.b32 %r737, %r736, -2139062144;
and.b32 %r738, %r196, 2139062143;
sub.s32 %r739, %r737, %r738;
xor.b32 %r199, %r546, %r739;
ld.global.nc.u32 %r200, [%rd58+20];

dp4a.s32.s32 %r198, %r199, %r200, %r192;

mov.u32 %r202, %r398;

prmt.b32 %r202, %r202, 0, 0xba98;

xor.b32 %r740, %r202, %r528;
or.b32 %r741, %r740, -2139062144;
and.b32 %r742, %r202, 2139062143;
sub.s32 %r743, %r741, %r742;
xor.b32 %r205, %r554, %r743;
ld.global.nc.u32 %r206, [%rd58+24];

dp4a.s32.s32 %r204, %r205, %r206, %r198;

mov.u32 %r208, %r404;

prmt.b32 %r208, %r208, 0, 0xba98;

xor.b32 %r744, %r557, %r208;
or.b32 %r745, %r744, -2139062144;
and.b32 %r746, %r208, 2139062143;
sub.s32 %r747, %r745, %r746;
xor.b32 %r211, %r576, %r747;
ld.global.nc.u32 %r212, [%rd58+28];

dp4a.s32.s32 %r210, %r211, %r212, %r204;

mov.u32 %r214, %r410;

prmt.b32 %r214, %r214, 0, 0xba98;

xor.b32 %r748, %r214, %r558;
or.b32 %r749, %r748, -2139062144;
and.b32 %r750, %r214, 2139062143;
sub.s32 %r751, %r749, %r750;
xor.b32 %r217, %r584, %r751;
ld.global.nc.u32 %r218, [%rd58+32];

dp4a.s32.s32 %r216, %r217, %r218, %r210;

mul.lo.s32 %r752, %r216, %r587;
shr.s32 %r753, %r752, 31;
shr.u32 %r754, %r753, 29;
add.s32 %r755, %r752, %r754;
shr.s32 %r756, %r755, 3;

{ cvt.f32.f16 %f53, %rs1;}


ld.global.nc.u32 %r220, [%rd58];

{.reg .f16 low,high;
mov.b32 {low,high},%r220;
cvt.f32.f16 %f54, low;}


mul.ftz.f32 %f69, %f53, %f54;
cvt.rn.f32.s32 %f70, %r756;
fma.rn.ftz.f32 %f214, %f69, %f70, %f214;
mov.u32 %r221, %r417;

prmt.b32 %r221, %r221, 0, 0xba98;

xor.b32 %r757, %r596, %r221;
or.b32 %r758, %r757, -2139062144;
and.b32 %r759, %r221, 2139062143;
sub.s32 %r760, %r758, %r759;
xor.b32 %r224, %r614, %r760;

dp4a.s32.s32 %r223, %r224, %r176, %r422;

mov.u32 %r227, %r423;

prmt.b32 %r227, %r227, 0, 0xba98;

xor.b32 %r761, %r227, %r597;
or.b32 %r762, %r761, -2139062144;
and.b32 %r763, %r227, 2139062143;
sub.s32 %r764, %r762, %r763;
xor.b32 %r230, %r622, %r764;

dp4a.s32.s32 %r229, %r230, %r182, %r223;

mov.u32 %r233, %r429;

prmt.b32 %r233, %r233, 0, 0xba98;

xor.b32 %r765, %r625, %r233;
or.b32 %r766, %r765, -2139062144;
and.b32 %r767, %r233, 2139062143;
sub.s32 %r768, %r766, %r767;
xor.b32 %r236, %r644, %r768;

dp4a.s32.s32 %r235, %r236, %r188, %r229;

mov.u32 %r239, %r435;

prmt.b32 %r239, %r239, 0, 0xba98;

xor.b32 %r769, %r239, %r626;
or.b32 %r770, %r769, -2139062144;
and.b32 %r771, %r239, 2139062143;
sub.s32 %r772, %r770, %r771;
xor.b32 %r242, %r652, %r772;

dp4a.s32.s32 %r241, %r242, %r194, %r235;

mov.u32 %r245, %r441;

prmt.b32 %r245, %r245, 0, 0xba98;

xor.b32 %r773, %r654, %r245;
or.b32 %r774, %r773, -2139062144;
and.b32 %r775, %r245, 2139062143;
sub.s32 %r776, %r774, %r775;
xor.b32 %r248, %r673, %r776;

dp4a.s32.s32 %r247, %r248, %r200, %r241;

mov.u32 %r251, %r447;

prmt.b32 %r251, %r251, 0, 0xba98;

xor.b32 %r777, %r251, %r655;
or.b32 %r778, %r777, -2139062144;
and.b32 %r779, %r251, 2139062143;
sub.s32 %r780, %r778, %r779;
xor.b32 %r254, %r681, %r780;

dp4a.s32.s32 %r253, %r254, %r206, %r247;

mov.u32 %r257, %r453;

prmt.b32 %r257, %r257, 0, 0xba98;

xor.b32 %r781, %r684, %r257;
or.b32 %r782, %r781, -2139062144;
and.b32 %r783, %r257, 2139062143;
sub.s32 %r784, %r782, %r783;
xor.b32 %r260, %r703, %r784;

dp4a.s32.s32 %r259, %r260, %r212, %r253;

mov.u32 %r263, %r459;

prmt.b32 %r263, %r263, 0, 0xba98;

xor.b32 %r785, %r263, %r685;
or.b32 %r786, %r785, -2139062144;
and.b32 %r787, %r263, 2139062143;
sub.s32 %r788, %r786, %r787;
xor.b32 %r266, %r711, %r788;

dp4a.s32.s32 %r265, %r266, %r218, %r259;

mul.lo.s32 %r789, %r265, %r714;
shr.s32 %r790, %r789, 31;
shr.u32 %r791, %r790, 29;
add.s32 %r792, %r789, %r791;
shr.s32 %r793, %r792, 3;

{ cvt.f32.f16 %f55, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r220;
cvt.f32.f16 %f56, low;}


mul.ftz.f32 %f71, %f55, %f56;
cvt.rn.f32.s32 %f72, %r793;
fma.rn.ftz.f32 %f213, %f71, %f72, %f213;
cvt.u64.u32 %rd59, %r1176;
add.s64 %rd60, %rd1, %rd59;
add.s64 %rd61, %rd60, %rd2;
mul.lo.s64 %rd62, %rd61, 36;
mov.u32 %r270, %r368;

prmt.b32 %r270, %r270, 0, 0xba98;

xor.b32 %r794, %r469, %r270;
or.b32 %r795, %r794, -2139062144;
and.b32 %r796, %r270, 2139062143;
sub.s32 %r797, %r795, %r796;
xor.b32 %r273, %r487, %r797;
add.s64 %rd63, %rd7, %rd62;
ld.global.nc.u32 %r274, [%rd63+4];

dp4a.s32.s32 %r272, %r273, %r274, %r422;

mov.u32 %r276, %r374;

prmt.b32 %r276, %r276, 0, 0xba98;

xor.b32 %r798, %r276, %r470;
or.b32 %r799, %r798, -2139062144;
and.b32 %r800, %r276, 2139062143;
sub.s32 %r801, %r799, %r800;
xor.b32 %r279, %r495, %r801;
ld.global.nc.u32 %r280, [%rd63+8];

dp4a.s32.s32 %r278, %r279, %r280, %r272;

mov.u32 %r282, %r380;

prmt.b32 %r282, %r282, 0, 0xba98;

xor.b32 %r802, %r498, %r282;
or.b32 %r803, %r802, -2139062144;
and.b32 %r804, %r282, 2139062143;
sub.s32 %r805, %r803, %r804;
xor.b32 %r285, %r517, %r805;
ld.global.nc.u32 %r286, [%rd63+12];

dp4a.s32.s32 %r284, %r285, %r286, %r278;

mov.u32 %r288, %r386;

prmt.b32 %r288, %r288, 0, 0xba98;

xor.b32 %r806, %r288, %r499;
or.b32 %r807, %r806, -2139062144;
and.b32 %r808, %r288, 2139062143;
sub.s32 %r809, %r807, %r808;
xor.b32 %r291, %r525, %r809;
ld.global.nc.u32 %r292, [%rd63+16];

dp4a.s32.s32 %r290, %r291, %r292, %r284;

mov.u32 %r294, %r392;

prmt.b32 %r294, %r294, 0, 0xba98;

xor.b32 %r810, %r527, %r294;
or.b32 %r811, %r810, -2139062144;
and.b32 %r812, %r294, 2139062143;
sub.s32 %r813, %r811, %r812;
xor.b32 %r297, %r546, %r813;
ld.global.nc.u32 %r298, [%rd63+20];

dp4a.s32.s32 %r296, %r297, %r298, %r290;

mov.u32 %r300, %r398;

prmt.b32 %r300, %r300, 0, 0xba98;

xor.b32 %r814, %r300, %r528;
or.b32 %r815, %r814, -2139062144;
and.b32 %r816, %r300, 2139062143;
sub.s32 %r817, %r815, %r816;
xor.b32 %r303, %r554, %r817;
ld.global.nc.u32 %r304, [%rd63+24];

dp4a.s32.s32 %r302, %r303, %r304, %r296;

mov.u32 %r306, %r404;

prmt.b32 %r306, %r306, 0, 0xba98;

xor.b32 %r818, %r557, %r306;
or.b32 %r819, %r818, -2139062144;
and.b32 %r820, %r306, 2139062143;
sub.s32 %r821, %r819, %r820;
xor.b32 %r309, %r576, %r821;
ld.global.nc.u32 %r310, [%rd63+28];

dp4a.s32.s32 %r308, %r309, %r310, %r302;

mov.u32 %r312, %r410;

prmt.b32 %r312, %r312, 0, 0xba98;

xor.b32 %r822, %r312, %r558;
or.b32 %r823, %r822, -2139062144;
and.b32 %r824, %r312, 2139062143;
sub.s32 %r825, %r823, %r824;
xor.b32 %r315, %r584, %r825;
ld.global.nc.u32 %r316, [%rd63+32];

dp4a.s32.s32 %r314, %r315, %r316, %r308;

mul.lo.s32 %r826, %r314, %r587;
shr.s32 %r827, %r826, 31;
shr.u32 %r828, %r827, 29;
add.s32 %r829, %r826, %r828;
shr.s32 %r830, %r829, 3;

{ cvt.f32.f16 %f57, %rs1;}


ld.global.nc.u32 %r318, [%rd63];

{.reg .f16 low,high;
mov.b32 {low,high},%r318;
cvt.f32.f16 %f58, low;}


mul.ftz.f32 %f73, %f57, %f58;
cvt.rn.f32.s32 %f74, %r830;
fma.rn.ftz.f32 %f212, %f73, %f74, %f212;
mov.u32 %r319, %r417;

prmt.b32 %r319, %r319, 0, 0xba98;

xor.b32 %r831, %r596, %r319;
or.b32 %r832, %r831, -2139062144;
and.b32 %r833, %r319, 2139062143;
sub.s32 %r834, %r832, %r833;
xor.b32 %r322, %r614, %r834;

dp4a.s32.s32 %r321, %r322, %r274, %r422;

mov.u32 %r325, %r423;

prmt.b32 %r325, %r325, 0, 0xba98;

xor.b32 %r835, %r325, %r597;
or.b32 %r836, %r835, -2139062144;
and.b32 %r837, %r325, 2139062143;
sub.s32 %r838, %r836, %r837;
xor.b32 %r328, %r622, %r838;

dp4a.s32.s32 %r327, %r328, %r280, %r321;

mov.u32 %r331, %r429;

prmt.b32 %r331, %r331, 0, 0xba98;

xor.b32 %r839, %r625, %r331;
or.b32 %r840, %r839, -2139062144;
and.b32 %r841, %r331, 2139062143;
sub.s32 %r842, %r840, %r841;
xor.b32 %r334, %r644, %r842;

dp4a.s32.s32 %r333, %r334, %r286, %r327;

mov.u32 %r337, %r435;

prmt.b32 %r337, %r337, 0, 0xba98;

xor.b32 %r843, %r337, %r626;
or.b32 %r844, %r843, -2139062144;
and.b32 %r845, %r337, 2139062143;
sub.s32 %r846, %r844, %r845;
xor.b32 %r340, %r652, %r846;

dp4a.s32.s32 %r339, %r340, %r292, %r333;

mov.u32 %r343, %r441;

prmt.b32 %r343, %r343, 0, 0xba98;

xor.b32 %r847, %r654, %r343;
or.b32 %r848, %r847, -2139062144;
and.b32 %r849, %r343, 2139062143;
sub.s32 %r850, %r848, %r849;
xor.b32 %r346, %r673, %r850;

dp4a.s32.s32 %r345, %r346, %r298, %r339;

mov.u32 %r349, %r447;

prmt.b32 %r349, %r349, 0, 0xba98;

xor.b32 %r851, %r349, %r655;
or.b32 %r852, %r851, -2139062144;
and.b32 %r853, %r349, 2139062143;
sub.s32 %r854, %r852, %r853;
xor.b32 %r352, %r681, %r854;

dp4a.s32.s32 %r351, %r352, %r304, %r345;

mov.u32 %r355, %r453;

prmt.b32 %r355, %r355, 0, 0xba98;

xor.b32 %r855, %r684, %r355;
or.b32 %r856, %r855, -2139062144;
and.b32 %r857, %r355, 2139062143;
sub.s32 %r858, %r856, %r857;
xor.b32 %r358, %r703, %r858;

dp4a.s32.s32 %r357, %r358, %r310, %r351;

mov.u32 %r361, %r459;

prmt.b32 %r361, %r361, 0, 0xba98;

xor.b32 %r859, %r361, %r685;
or.b32 %r860, %r859, -2139062144;
and.b32 %r861, %r361, 2139062143;
sub.s32 %r862, %r860, %r861;
xor.b32 %r364, %r711, %r862;

dp4a.s32.s32 %r363, %r364, %r316, %r357;

mul.lo.s32 %r863, %r363, %r714;
shr.s32 %r864, %r863, 31;
shr.u32 %r865, %r864, 29;
add.s32 %r866, %r863, %r865;
shr.s32 %r867, %r866, 3;

{ cvt.f32.f16 %f59, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r318;
cvt.f32.f16 %f60, low;}


mul.ftz.f32 %f75, %f59, %f60;
cvt.rn.f32.s32 %f76, %r867;
fma.rn.ftz.f32 %f211, %f75, %f76, %f211;
cvt.u64.u32 %rd64, %r1175;
add.s64 %rd65, %rd1, %rd64;
add.s64 %rd66, %rd65, %rd2;
mul.lo.s64 %rd67, %rd66, 36;

prmt.b32 %r368, %r368, 0, 0xba98;

xor.b32 %r868, %r469, %r368;
or.b32 %r869, %r868, -2139062144;
and.b32 %r870, %r368, 2139062143;
sub.s32 %r871, %r869, %r870;
xor.b32 %r371, %r487, %r871;
add.s64 %rd68, %rd7, %rd67;
ld.global.nc.u32 %r372, [%rd68+4];

dp4a.s32.s32 %r370, %r371, %r372, %r422;


prmt.b32 %r374, %r374, 0, 0xba98;

xor.b32 %r872, %r374, %r470;
or.b32 %r873, %r872, -2139062144;
and.b32 %r874, %r374, 2139062143;
sub.s32 %r875, %r873, %r874;
xor.b32 %r377, %r495, %r875;
ld.global.nc.u32 %r378, [%rd68+8];

dp4a.s32.s32 %r376, %r377, %r378, %r370;


prmt.b32 %r380, %r380, 0, 0xba98;

xor.b32 %r876, %r498, %r380;
or.b32 %r877, %r876, -2139062144;
and.b32 %r878, %r380, 2139062143;
sub.s32 %r879, %r877, %r878;
xor.b32 %r383, %r517, %r879;
ld.global.nc.u32 %r384, [%rd68+12];

dp4a.s32.s32 %r382, %r383, %r384, %r376;


prmt.b32 %r386, %r386, 0, 0xba98;

xor.b32 %r880, %r386, %r499;
or.b32 %r881, %r880, -2139062144;
and.b32 %r882, %r386, 2139062143;
sub.s32 %r883, %r881, %r882;
xor.b32 %r389, %r525, %r883;
ld.global.nc.u32 %r390, [%rd68+16];

dp4a.s32.s32 %r388, %r389, %r390, %r382;


prmt.b32 %r392, %r392, 0, 0xba98;

xor.b32 %r884, %r527, %r392;
or.b32 %r885, %r884, -2139062144;
and.b32 %r886, %r392, 2139062143;
sub.s32 %r887, %r885, %r886;
xor.b32 %r395, %r546, %r887;
ld.global.nc.u32 %r396, [%rd68+20];

dp4a.s32.s32 %r394, %r395, %r396, %r388;


prmt.b32 %r398, %r398, 0, 0xba98;

xor.b32 %r888, %r398, %r528;
or.b32 %r889, %r888, -2139062144;
and.b32 %r890, %r398, 2139062143;
sub.s32 %r891, %r889, %r890;
xor.b32 %r401, %r554, %r891;
ld.global.nc.u32 %r402, [%rd68+24];

dp4a.s32.s32 %r400, %r401, %r402, %r394;


prmt.b32 %r404, %r404, 0, 0xba98;

xor.b32 %r892, %r557, %r404;
or.b32 %r893, %r892, -2139062144;
and.b32 %r894, %r404, 2139062143;
sub.s32 %r895, %r893, %r894;
xor.b32 %r407, %r576, %r895;
ld.global.nc.u32 %r408, [%rd68+28];

dp4a.s32.s32 %r406, %r407, %r408, %r400;


prmt.b32 %r410, %r410, 0, 0xba98;

xor.b32 %r896, %r410, %r558;
or.b32 %r897, %r896, -2139062144;
and.b32 %r898, %r410, 2139062143;
sub.s32 %r899, %r897, %r898;
xor.b32 %r413, %r584, %r899;
ld.global.nc.u32 %r414, [%rd68+32];

dp4a.s32.s32 %r412, %r413, %r414, %r406;

mul.lo.s32 %r900, %r412, %r587;
shr.s32 %r901, %r900, 31;
shr.u32 %r902, %r901, 29;
add.s32 %r903, %r900, %r902;
shr.s32 %r904, %r903, 3;

{ cvt.f32.f16 %f61, %rs1;}


ld.global.nc.u32 %r416, [%rd68];

{.reg .f16 low,high;
mov.b32 {low,high},%r416;
cvt.f32.f16 %f62, low;}


mul.ftz.f32 %f77, %f61, %f62;
cvt.rn.f32.s32 %f78, %r904;
fma.rn.ftz.f32 %f210, %f77, %f78, %f210;

prmt.b32 %r417, %r417, 0, 0xba98;

xor.b32 %r905, %r596, %r417;
or.b32 %r906, %r905, -2139062144;
and.b32 %r907, %r417, 2139062143;
sub.s32 %r908, %r906, %r907;
xor.b32 %r420, %r614, %r908;

dp4a.s32.s32 %r419, %r420, %r372, %r422;


prmt.b32 %r423, %r423, 0, 0xba98;

xor.b32 %r909, %r423, %r597;
or.b32 %r910, %r909, -2139062144;
and.b32 %r911, %r423, 2139062143;
sub.s32 %r912, %r910, %r911;
xor.b32 %r426, %r622, %r912;

dp4a.s32.s32 %r425, %r426, %r378, %r419;


prmt.b32 %r429, %r429, 0, 0xba98;

xor.b32 %r913, %r625, %r429;
or.b32 %r914, %r913, -2139062144;
and.b32 %r915, %r429, 2139062143;
sub.s32 %r916, %r914, %r915;
xor.b32 %r432, %r644, %r916;

dp4a.s32.s32 %r431, %r432, %r384, %r425;


prmt.b32 %r435, %r435, 0, 0xba98;

xor.b32 %r917, %r435, %r626;
or.b32 %r918, %r917, -2139062144;
and.b32 %r919, %r435, 2139062143;
sub.s32 %r920, %r918, %r919;
xor.b32 %r438, %r652, %r920;

dp4a.s32.s32 %r437, %r438, %r390, %r431;


prmt.b32 %r441, %r441, 0, 0xba98;

xor.b32 %r921, %r654, %r441;
or.b32 %r922, %r921, -2139062144;
and.b32 %r923, %r441, 2139062143;
sub.s32 %r924, %r922, %r923;
xor.b32 %r444, %r673, %r924;

dp4a.s32.s32 %r443, %r444, %r396, %r437;


prmt.b32 %r447, %r447, 0, 0xba98;

xor.b32 %r925, %r447, %r655;
or.b32 %r926, %r925, -2139062144;
and.b32 %r927, %r447, 2139062143;
sub.s32 %r928, %r926, %r927;
xor.b32 %r450, %r681, %r928;

dp4a.s32.s32 %r449, %r450, %r402, %r443;


prmt.b32 %r453, %r453, 0, 0xba98;

xor.b32 %r929, %r684, %r453;
or.b32 %r930, %r929, -2139062144;
and.b32 %r931, %r453, 2139062143;
sub.s32 %r932, %r930, %r931;
xor.b32 %r456, %r703, %r932;

dp4a.s32.s32 %r455, %r456, %r408, %r449;


prmt.b32 %r459, %r459, 0, 0xba98;

xor.b32 %r933, %r459, %r685;
or.b32 %r934, %r933, -2139062144;
and.b32 %r935, %r459, 2139062143;
sub.s32 %r936, %r934, %r935;
xor.b32 %r462, %r711, %r936;

dp4a.s32.s32 %r461, %r462, %r414, %r455;

mul.lo.s32 %r937, %r461, %r714;
shr.s32 %r938, %r937, 31;
shr.u32 %r939, %r938, 29;
add.s32 %r940, %r937, %r939;
shr.s32 %r941, %r940, 3;

{ cvt.f32.f16 %f63, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r416;
cvt.f32.f16 %f64, low;}


mul.ftz.f32 %f79, %f63, %f64;
cvt.rn.f32.s32 %f80, %r941;
fma.rn.ftz.f32 %f209, %f79, %f80, %f209;
add.s32 %r1176, %r1176, 128;
add.s32 %r1175, %r1175, 128;
add.s32 %r1174, %r1174, 128;
add.s32 %r1173, %r1173, 128;
add.s64 %rd106, %rd106, 1056;
add.s64 %rd105, %rd105, 1056;
add.s64 %rd104, %rd104, 1056;
add.s32 %r1177, %r1177, 16;
setp.lt.s32 %p3, %r1177, %r41;
@%p3 bra $L__BB206_2;

$L__BB206_3:
setp.eq.s32 %p4, %r34, 0;
@%p4 bra $L__BB206_5;

shl.b32 %r945, %r34, 10;
mov.u32 %r946, _ZZ13mul_mat_vec_qIL9ggml_type16ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r947, %r946, %r945;
shl.b32 %r948, %r36, 2;
add.s32 %r949, %r947, %r948;
st.shared.f32 [%r949+-1024], %f216;
st.shared.f32 [%r949+-896], %f215;
st.shared.f32 [%r949+-768], %f214;
st.shared.f32 [%r949+-640], %f213;
st.shared.f32 [%r949+-512], %f212;
st.shared.f32 [%r949+-384], %f211;
st.shared.f32 [%r949+-256], %f210;
st.shared.f32 [%r949+-128], %f209;

$L__BB206_5:
setp.ne.s32 %p5, %r34, 0;
bar.sync 0;
@%p5 bra $L__BB206_22;

mov.u32 %r951, %ctaid.x;
shl.b32 %r952, %r951, 1;
mov.u32 %r953, 1;
shl.b32 %r955, %r36, 2;
mov.u32 %r956, _ZZ13mul_mat_vec_qIL9ggml_type16ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r957, %r956, %r955;
mov.u32 %r958, 2;
ld.shared.f32 %f81, [%r957];
add.ftz.f32 %f82, %f81, %f216;
ld.shared.f32 %f83, [%r957+1024];
add.ftz.f32 %f84, %f83, %f82;
ld.shared.f32 %f85, [%r957+2048];
add.ftz.f32 %f86, %f85, %f84;
mov.b32 %r959, %f86;
mov.u32 %r960, 31;
mov.u32 %r961, 16;
mov.u32 %r962, -1;
shfl.sync.bfly.b32 %r963|%p6, %r959, %r961, %r960, %r962;
mov.b32 %f87, %r963;
add.ftz.f32 %f88, %f86, %f87;
mov.b32 %r964, %f88;
mov.u32 %r965, 8;
shfl.sync.bfly.b32 %r966|%p7, %r964, %r965, %r960, %r962;
mov.b32 %f89, %r966;
add.ftz.f32 %f90, %f88, %f89;
mov.b32 %r967, %f90;
mov.u32 %r968, 4;
shfl.sync.bfly.b32 %r969|%p8, %r967, %r968, %r960, %r962;
mov.b32 %f91, %r969;
add.ftz.f32 %f92, %f90, %f91;
mov.b32 %r970, %f92;
shfl.sync.bfly.b32 %r971|%p9, %r970, %r958, %r960, %r962;
mov.b32 %f93, %r971;
add.ftz.f32 %f94, %f92, %f93;
mov.b32 %r972, %f94;
shfl.sync.bfly.b32 %r973|%p10, %r972, %r953, %r960, %r962;
mov.b32 %f95, %r973;
add.ftz.f32 %f25, %f94, %f95;
setp.ne.s32 %p11, %r36, 0;
setp.ge.u32 %p12, %r952, %r21;
or.pred %p13, %p11, %p12;
@%p13 bra $L__BB206_8;

cvta.to.global.u64 %rd69, %rd17;
mov.u32 %r974, %ctaid.y;
mad.lo.s32 %r977, %r974, %r27, %r952;
mov.u32 %r978, %ctaid.z;
mad.lo.s32 %r979, %r978, %r33, %r977;
mul.wide.u32 %rd70, %r979, 4;
add.s64 %rd71, %rd69, %rd70;
st.global.f32 [%rd71], %f25;

$L__BB206_8:
ld.shared.f32 %f96, [%r957+128];
add.ftz.f32 %f97, %f96, %f215;
ld.shared.f32 %f98, [%r957+1152];
add.ftz.f32 %f99, %f98, %f97;
ld.shared.f32 %f100, [%r957+2176];
add.ftz.f32 %f101, %f100, %f99;
mov.b32 %r985, %f101;
shfl.sync.bfly.b32 %r989|%p14, %r985, %r961, %r960, %r962;
mov.b32 %f102, %r989;
add.ftz.f32 %f103, %f101, %f102;
mov.b32 %r990, %f103;
shfl.sync.bfly.b32 %r992|%p15, %r990, %r965, %r960, %r962;
mov.b32 %f104, %r992;
add.ftz.f32 %f105, %f103, %f104;
mov.b32 %r993, %f105;
shfl.sync.bfly.b32 %r995|%p16, %r993, %r968, %r960, %r962;
mov.b32 %f106, %r995;
add.ftz.f32 %f107, %f105, %f106;
mov.b32 %r996, %f107;
shfl.sync.bfly.b32 %r997|%p17, %r996, %r958, %r960, %r962;
mov.b32 %f108, %r997;
add.ftz.f32 %f109, %f107, %f108;
mov.b32 %r998, %f109;
shfl.sync.bfly.b32 %r1000|%p18, %r998, %r953, %r960, %r962;
mov.b32 %f110, %r1000;
add.ftz.f32 %f26, %f109, %f110;
setp.eq.s32 %p19, %r36, 1;
or.b32 %r1003, %r952, 1;
setp.lt.u32 %p20, %r1003, %r21;
and.pred %p1, %p19, %p20;
not.pred %p21, %p1;
@%p21 bra $L__BB206_10;

cvta.to.global.u64 %rd72, %rd17;
mov.u32 %r1004, %ctaid.y;
mad.lo.s32 %r1007, %r1004, %r27, %r952;
mov.u32 %r1008, %ctaid.z;
mad.lo.s32 %r1009, %r1008, %r33, %r1007;
mul.wide.u32 %rd73, %r1009, 4;
add.s64 %rd74, %rd72, %rd73;
st.global.f32 [%rd74+4], %f26;

$L__BB206_10:
mov.u32 %r1014, 2;
ld.shared.f32 %f111, [%r957+256];
add.ftz.f32 %f112, %f111, %f214;
ld.shared.f32 %f113, [%r957+1280];
add.ftz.f32 %f114, %f113, %f112;
ld.shared.f32 %f115, [%r957+2304];
add.ftz.f32 %f116, %f115, %f114;
mov.b32 %r1015, %f116;
mov.u32 %r1016, 31;
mov.u32 %r1017, 16;
mov.u32 %r1018, -1;
shfl.sync.bfly.b32 %r1019|%p22, %r1015, %r1017, %r1016, %r1018;
mov.b32 %f117, %r1019;
add.ftz.f32 %f118, %f116, %f117;
mov.b32 %r1020, %f118;
mov.u32 %r1021, 8;
shfl.sync.bfly.b32 %r1022|%p23, %r1020, %r1021, %r1016, %r1018;
mov.b32 %f119, %r1022;
add.ftz.f32 %f120, %f118, %f119;
mov.b32 %r1023, %f120;
mov.u32 %r1024, 4;
shfl.sync.bfly.b32 %r1025|%p24, %r1023, %r1024, %r1016, %r1018;
mov.b32 %f121, %r1025;
add.ftz.f32 %f122, %f120, %f121;
mov.b32 %r1026, %f122;
shfl.sync.bfly.b32 %r1027|%p25, %r1026, %r1014, %r1016, %r1018;
mov.b32 %f123, %r1027;
add.ftz.f32 %f124, %f122, %f123;
mov.b32 %r1028, %f124;
mov.u32 %r1029, 1;
shfl.sync.bfly.b32 %r1030|%p26, %r1028, %r1029, %r1016, %r1018;
mov.b32 %f125, %r1030;
add.ftz.f32 %f27, %f124, %f125;
@%p13 bra $L__BB206_12;

cvt.u64.u32 %rd75, %r21;
mov.u32 %r1033, %ctaid.y;
mad.lo.s32 %r1036, %r1033, %r27, %r952;
mov.u32 %r1037, %ctaid.z;
mad.lo.s32 %r1038, %r1037, %r33, %r1036;
cvt.u64.u32 %rd76, %r1038;
add.s64 %rd77, %rd75, %rd76;
cvta.to.global.u64 %rd78, %rd17;
shl.b64 %rd79, %rd77, 2;
add.s64 %rd80, %rd78, %rd79;
st.global.f32 [%rd80], %f27;

$L__BB206_12:
ld.shared.f32 %f126, [%r957+384];
add.ftz.f32 %f127, %f126, %f213;
ld.shared.f32 %f128, [%r957+1408];
add.ftz.f32 %f129, %f128, %f127;
ld.shared.f32 %f130, [%r957+2432];
add.ftz.f32 %f131, %f130, %f129;
mov.b32 %r1044, %f131;
shfl.sync.bfly.b32 %r1048|%p30, %r1044, %r1017, %r1016, %r1018;
mov.b32 %f132, %r1048;
add.ftz.f32 %f133, %f131, %f132;
mov.b32 %r1049, %f133;
shfl.sync.bfly.b32 %r1051|%p31, %r1049, %r1021, %r1016, %r1018;
mov.b32 %f134, %r1051;
add.ftz.f32 %f135, %f133, %f134;
mov.b32 %r1052, %f135;
shfl.sync.bfly.b32 %r1054|%p32, %r1052, %r1024, %r1016, %r1018;
mov.b32 %f136, %r1054;
add.ftz.f32 %f137, %f135, %f136;
mov.b32 %r1055, %f137;
shfl.sync.bfly.b32 %r1056|%p33, %r1055, %r1014, %r1016, %r1018;
mov.b32 %f138, %r1056;
add.ftz.f32 %f139, %f137, %f138;
mov.b32 %r1057, %f139;
shfl.sync.bfly.b32 %r1059|%p34, %r1057, %r1029, %r1016, %r1018;
mov.b32 %f140, %r1059;
add.ftz.f32 %f28, %f139, %f140;
@%p21 bra $L__BB206_14;

add.s32 %r1060, %r21, 1;
cvt.u64.u32 %rd81, %r1060;
mov.u32 %r1061, %ctaid.y;
mad.lo.s32 %r1064, %r1061, %r27, %r952;
mov.u32 %r1065, %ctaid.z;
mad.lo.s32 %r1066, %r1065, %r33, %r1064;
cvt.u64.u32 %rd82, %r1066;
add.s64 %rd83, %rd81, %rd82;
cvta.to.global.u64 %rd84, %rd17;
shl.b64 %rd85, %rd83, 2;
add.s64 %rd86, %rd84, %rd85;
st.global.f32 [%rd86], %f28;

$L__BB206_14:
shl.b32 %r1067, %r21, 1;
mov.u32 %r1068, 1;
mov.u32 %r1073, 2;
ld.shared.f32 %f141, [%r957+512];
add.ftz.f32 %f142, %f141, %f212;
ld.shared.f32 %f143, [%r957+1536];
add.ftz.f32 %f144, %f143, %f142;
ld.shared.f32 %f145, [%r957+2560];
add.ftz.f32 %f146, %f145, %f144;
mov.b32 %r1074, %f146;
mov.u32 %r1075, 31;
mov.u32 %r1076, 16;
mov.u32 %r1077, -1;
shfl.sync.bfly.b32 %r1078|%p36, %r1074, %r1076, %r1075, %r1077;
mov.b32 %f147, %r1078;
add.ftz.f32 %f148, %f146, %f147;
mov.b32 %r1079, %f148;
mov.u32 %r1080, 8;
shfl.sync.bfly.b32 %r1081|%p37, %r1079, %r1080, %r1075, %r1077;
mov.b32 %f149, %r1081;
add.ftz.f32 %f150, %f148, %f149;
mov.b32 %r1082, %f150;
mov.u32 %r1083, 4;
shfl.sync.bfly.b32 %r1084|%p38, %r1082, %r1083, %r1075, %r1077;
mov.b32 %f151, %r1084;
add.ftz.f32 %f152, %f150, %f151;
mov.b32 %r1085, %f152;
shfl.sync.bfly.b32 %r1086|%p39, %r1085, %r1073, %r1075, %r1077;
mov.b32 %f153, %r1086;
add.ftz.f32 %f154, %f152, %f153;
mov.b32 %r1087, %f154;
shfl.sync.bfly.b32 %r1088|%p40, %r1087, %r1068, %r1075, %r1077;
mov.b32 %f155, %r1088;
add.ftz.f32 %f29, %f154, %f155;
cvt.u64.u32 %rd87, %r1067;
mov.u32 %r1089, %ctaid.y;
mad.lo.s32 %r1092, %r1089, %r27, %r952;
mov.u32 %r1093, %ctaid.z;
mad.lo.s32 %r1094, %r1093, %r33, %r1092;
cvt.u64.u32 %rd88, %r1094;
add.s64 %rd89, %rd87, %rd88;
cvta.to.global.u64 %rd90, %rd17;
shl.b64 %rd91, %rd89, 2;
add.s64 %rd14, %rd90, %rd91;
@%p13 bra $L__BB206_16;

st.global.f32 [%rd14], %f29;

$L__BB206_16:
ld.shared.f32 %f156, [%r957+640];
add.ftz.f32 %f157, %f156, %f211;
ld.shared.f32 %f158, [%r957+1664];
add.ftz.f32 %f159, %f158, %f157;
ld.shared.f32 %f160, [%r957+2688];
add.ftz.f32 %f161, %f160, %f159;
mov.b32 %r1100, %f161;
shfl.sync.bfly.b32 %r1104|%p44, %r1100, %r1076, %r1075, %r1077;
mov.b32 %f162, %r1104;
add.ftz.f32 %f163, %f161, %f162;
mov.b32 %r1105, %f163;
shfl.sync.bfly.b32 %r1107|%p45, %r1105, %r1080, %r1075, %r1077;
mov.b32 %f164, %r1107;
add.ftz.f32 %f165, %f163, %f164;
mov.b32 %r1108, %f165;
shfl.sync.bfly.b32 %r1110|%p46, %r1108, %r1083, %r1075, %r1077;
mov.b32 %f166, %r1110;
add.ftz.f32 %f167, %f165, %f166;
mov.b32 %r1111, %f167;
shfl.sync.bfly.b32 %r1112|%p47, %r1111, %r1073, %r1075, %r1077;
mov.b32 %f168, %r1112;
add.ftz.f32 %f169, %f167, %f168;
mov.b32 %r1113, %f169;
shfl.sync.bfly.b32 %r1115|%p48, %r1113, %r1068, %r1075, %r1077;
mov.b32 %f170, %r1115;
add.ftz.f32 %f30, %f169, %f170;
@%p21 bra $L__BB206_18;

st.global.f32 [%rd14+4], %f30;

$L__BB206_18:
mul.lo.s32 %r17, %r21, 3;
mov.u32 %r1120, 2;
ld.shared.f32 %f171, [%r957+768];
add.ftz.f32 %f172, %f171, %f210;
ld.shared.f32 %f173, [%r957+1792];
add.ftz.f32 %f174, %f173, %f172;
ld.shared.f32 %f175, [%r957+2816];
add.ftz.f32 %f176, %f175, %f174;
mov.b32 %r1121, %f176;
mov.u32 %r1122, 31;
mov.u32 %r1123, 16;
mov.u32 %r1124, -1;
shfl.sync.bfly.b32 %r1125|%p50, %r1121, %r1123, %r1122, %r1124;
mov.b32 %f177, %r1125;
add.ftz.f32 %f178, %f176, %f177;
mov.b32 %r1126, %f178;
mov.u32 %r1127, 8;
shfl.sync.bfly.b32 %r1128|%p51, %r1126, %r1127, %r1122, %r1124;
mov.b32 %f179, %r1128;
add.ftz.f32 %f180, %f178, %f179;
mov.b32 %r1129, %f180;
mov.u32 %r1130, 4;
shfl.sync.bfly.b32 %r1131|%p52, %r1129, %r1130, %r1122, %r1124;
mov.b32 %f181, %r1131;
add.ftz.f32 %f182, %f180, %f181;
mov.b32 %r1132, %f182;
shfl.sync.bfly.b32 %r1133|%p53, %r1132, %r1120, %r1122, %r1124;
mov.b32 %f183, %r1133;
add.ftz.f32 %f184, %f182, %f183;
mov.b32 %r1134, %f184;
mov.u32 %r1135, 1;
shfl.sync.bfly.b32 %r1136|%p54, %r1134, %r1135, %r1122, %r1124;
mov.b32 %f185, %r1136;
add.ftz.f32 %f31, %f184, %f185;
@%p13 bra $L__BB206_20;

cvt.u64.u32 %rd92, %r17;
add.s64 %rd94, %rd92, %rd88;
shl.b64 %rd96, %rd94, 2;
add.s64 %rd97, %rd90, %rd96;
st.global.f32 [%rd97], %f31;

$L__BB206_20:
ld.shared.f32 %f186, [%r957+896];
add.ftz.f32 %f187, %f186, %f209;
ld.shared.f32 %f188, [%r957+1920];
add.ftz.f32 %f189, %f188, %f187;
ld.shared.f32 %f190, [%r957+2944];
add.ftz.f32 %f191, %f190, %f189;
mov.b32 %r1150, %f191;
shfl.sync.bfly.b32 %r1154|%p58, %r1150, %r1123, %r1122, %r1124;
mov.b32 %f192, %r1154;
add.ftz.f32 %f193, %f191, %f192;
mov.b32 %r1155, %f193;
shfl.sync.bfly.b32 %r1157|%p59, %r1155, %r1127, %r1122, %r1124;
mov.b32 %f194, %r1157;
add.ftz.f32 %f195, %f193, %f194;
mov.b32 %r1158, %f195;
shfl.sync.bfly.b32 %r1160|%p60, %r1158, %r1130, %r1122, %r1124;
mov.b32 %f196, %r1160;
add.ftz.f32 %f197, %f195, %f196;
mov.b32 %r1161, %f197;
shfl.sync.bfly.b32 %r1162|%p61, %r1161, %r1120, %r1122, %r1124;
mov.b32 %f198, %r1162;
add.ftz.f32 %f199, %f197, %f198;
mov.b32 %r1163, %f199;
shfl.sync.bfly.b32 %r1165|%p62, %r1163, %r1135, %r1122, %r1124;
mov.b32 %f200, %r1165;
add.ftz.f32 %f32, %f199, %f200;
@%p21 bra $L__BB206_22;

add.s32 %r1166, %r17, 1;
cvt.u64.u32 %rd98, %r1166;
add.s64 %rd100, %rd98, %rd88;
shl.b64 %rd102, %rd100, 2;
add.s64 %rd103, %rd90, %rd102;
st.global.f32 [%rd103], %f32;

$L__BB206_22:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type16ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type16ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type16ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type16ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type16ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type16ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type16ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type16ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type16ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<78>;
.reg .b16 %rs<27>;
.reg .f32 %f<231>;
.reg .b32 %r<1402>;
.reg .b64 %rd<118>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type16ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1280];

ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type16ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type16ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type16ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type16ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type16ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type16ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type16ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type16ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type16ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type16ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type16ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type16ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type16ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type16ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type16ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type16ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type16ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r37, %tid.y;
shl.b32 %r38, %r37, 5;
mov.u32 %r39, %tid.x;
add.s32 %r40, %r38, %r39;
shr.s32 %r41, %r40, 31;
shr.u32 %r42, %r41, 29;
add.s32 %r43, %r40, %r42;
shr.s32 %r1401, %r43, 3;
shr.u32 %r2, %r36, 8;
setp.ge.s32 %p2, %r1401, %r2;
mov.f32 %f221, 0f00000000;
mov.f32 %f222, %f221;
mov.f32 %f223, %f221;
mov.f32 %f224, %f221;
mov.f32 %f225, %f221;
mov.f32 %f226, %f221;
mov.f32 %f227, %f221;
mov.f32 %f228, %f221;
mov.f32 %f229, %f221;
mov.f32 %f230, %f221;
@%p2 bra $L__BB207_3;

mov.u32 %r44, %ctaid.y;
mov.u32 %r45, %ctaid.z;
mul.lo.s32 %r46, %r44, %r28;
cvt.u64.u32 %rd19, %r46;
mul.lo.s32 %r47, %r45, %r34;
cvt.u64.u32 %rd20, %r47;
add.s64 %rd1, %rd20, %rd19;
mul.hi.u32 %r48, %r45, %r30;
add.s32 %r49, %r48, %r45;
shr.u32 %r50, %r49, %r31;
mul.lo.s32 %r51, %r50, %r33;
mul.hi.u32 %r52, %r44, %r24;
add.s32 %r53, %r52, %r44;
shr.u32 %r54, %r53, %r25;
mul.lo.s32 %r55, %r54, %r27;
cvta.to.global.u64 %rd21, %rd16;
and.b32 %r63, %r43, -8;
sub.s32 %r64, %r40, %r63;
shl.b32 %r65, %r64, 2;
cvt.s64.s32 %rd2, %r64;
mov.u32 %r66, %ctaid.x;
shl.b32 %r67, %r66, 1;
shl.b32 %r68, %r22, 1;
shl.b32 %r1396, %r1401, 3;
add.s32 %r1400, %r1396, %r68;
mad.lo.s32 %r1399, %r22, 3, %r1396;
shl.b32 %r69, %r22, 2;
add.s32 %r1398, %r1396, %r69;
add.s32 %r1397, %r22, %r1396;
add.s32 %r70, %r1401, %r51;
add.s32 %r71, %r70, %r55;
or.b32 %r72, %r67, 1;
mad.lo.s32 %r73, %r21, %r72, %r71;
mul.wide.s32 %rd22, %r73, 66;
add.s64 %rd117, %rd21, %rd22;
mad.lo.s32 %r74, %r67, %r21, %r1401;
add.s32 %r75, %r74, %r51;
add.s32 %r76, %r75, %r55;
mul.wide.s32 %rd23, %r76, 66;
add.s64 %rd24, %rd21, %rd23;
add.s64 %rd116, %rd24, 4;
mul.wide.s32 %rd25, %r65, 2;
add.s64 %rd26, %rd23, %rd25;
add.s64 %rd27, %rd21, %rd26;
add.s64 %rd115, %rd27, 8;
or.b64 %rd6, %rd25, 4;
cvta.to.global.u64 %rd7, %rd17;
mov.f32 %f221, 0f00000000;

$L__BB207_2:
cvt.u64.u32 %rd28, %r1396;
add.s64 %rd29, %rd1, %rd28;
add.s64 %rd30, %rd29, %rd2;
mul.lo.s64 %rd31, %rd30, 36;
ld.global.nc.u16 %rs11, [%rd115+-6];
shr.u16 %rs12, %rs11, 8;
ld.global.nc.u16 %rs13, [%rd115+-4];
shr.u16 %rs14, %rs13, 8;
ld.global.nc.u16 %rs15, [%rd115+-2];
cvt.u32.u16 %r567, %rs15;
ld.global.nc.u16 %rs16, [%rd115];
cvt.u32.u16 %r568, %rs16;
mov.b32 %r569, {%rs15, %rs16};
shl.b16 %rs17, %rs11, 3;
cvt.u64.u16 %rd32, %rs17;
and.b64 %rd33, %rd32, 2040;
mov.u64 %rd34, iq2xxs_grid;
add.s64 %rd35, %rd34, %rd33;
ld.global.nc.v2.u32 {%r570, %r571}, [%rd35];
not.b32 %r573, %r570;
mov.u32 %r523, 0;
not.b32 %r575, %r571;
and.b32 %r576, %r567, 255;
popc.b32 %r577, %r576;
shl.b32 %r578, %r577, 7;
and.b32 %r579, %r578, 128;
xor.b32 %r580, %r579, %r576;
mul.lo.s32 %r581, %r580, 16843009;
and.b32 %r582, %r581, 134480385;
or.b32 %r583, %r582, -2139062144;
add.s32 %r584, %r583, -16843009;
or.b32 %r469, %r584, %r582;
mov.u32 %r77, %r469;

prmt.b32 %r77, %r77, 0, 0xba98;

xor.b32 %r585, %r570, %r77;
or.b32 %r586, %r585, -2139062144;
and.b32 %r587, %r77, 2139062143;
and.b32 %r588, %r573, -2139062144;
sub.s32 %r589, %r586, %r587;
xor.b32 %r80, %r588, %r589;
add.s64 %rd36, %rd7, %rd31;
ld.global.nc.u32 %r81, [%rd36+4];

dp4a.s32.s32 %r79, %r80, %r81, %r523;

and.b32 %r590, %r581, -2143281136;
or.b32 %r591, %r590, -2139062144;
add.s32 %r592, %r591, -16843009;
or.b32 %r475, %r592, %r590;
mov.u32 %r83, %r475;

prmt.b32 %r83, %r83, 0, 0xba98;

xor.b32 %r593, %r83, %r571;
or.b32 %r594, %r593, -2139062144;
and.b32 %r595, %r83, 2139062143;
and.b32 %r596, %r575, -2139062144;
sub.s32 %r597, %r594, %r595;
xor.b32 %r86, %r596, %r597;
ld.global.nc.u32 %r87, [%rd36+8];

dp4a.s32.s32 %r85, %r86, %r87, %r79;

cvt.u32.u16 %r598, %rs12;
mul.wide.u32 %rd37, %r598, 8;
add.s64 %rd38, %rd34, %rd37;
ld.global.nc.v2.u32 {%r599, %r600}, [%rd38];
not.b32 %r602, %r599;
not.b32 %r604, %r600;
shr.u32 %r605, %r567, 7;
and.b32 %r606, %r605, 255;
popc.b32 %r607, %r606;
shl.b32 %r608, %r607, 7;
and.b32 %r609, %r608, 128;
xor.b32 %r610, %r609, %r606;
mul.lo.s32 %r611, %r610, 16843009;
and.b32 %r612, %r611, 134480385;
or.b32 %r613, %r612, -2139062144;
add.s32 %r614, %r613, -16843009;
or.b32 %r481, %r614, %r612;
mov.u32 %r89, %r481;

prmt.b32 %r89, %r89, 0, 0xba98;

xor.b32 %r615, %r599, %r89;
or.b32 %r616, %r615, -2139062144;
and.b32 %r617, %r89, 2139062143;
and.b32 %r618, %r602, -2139062144;
sub.s32 %r619, %r616, %r617;
xor.b32 %r92, %r618, %r619;
ld.global.nc.u32 %r93, [%rd36+12];

dp4a.s32.s32 %r91, %r92, %r93, %r85;

and.b32 %r620, %r611, -2143281136;
or.b32 %r621, %r620, -2139062144;
add.s32 %r622, %r621, -16843009;
or.b32 %r487, %r622, %r620;
mov.u32 %r95, %r487;

prmt.b32 %r95, %r95, 0, 0xba98;

xor.b32 %r623, %r95, %r600;
or.b32 %r624, %r623, -2139062144;
and.b32 %r625, %r95, 2139062143;
and.b32 %r626, %r604, -2139062144;
sub.s32 %r627, %r624, %r625;
xor.b32 %r98, %r626, %r627;
ld.global.nc.u32 %r99, [%rd36+16];

dp4a.s32.s32 %r97, %r98, %r99, %r91;

shl.b16 %rs18, %rs13, 3;
cvt.u64.u16 %rd39, %rs18;
and.b64 %rd40, %rd39, 2040;
add.s64 %rd41, %rd34, %rd40;
ld.global.nc.v2.u32 {%r628, %r629}, [%rd41];
not.b32 %r631, %r628;
not.b32 %r633, %r629;
shr.u32 %r634, %r569, 14;
and.b32 %r635, %r634, 255;
popc.b32 %r636, %r635;
shl.b32 %r637, %r636, 7;
and.b32 %r638, %r637, 128;
xor.b32 %r639, %r638, %r635;
mul.lo.s32 %r640, %r639, 16843009;
and.b32 %r641, %r640, 134480385;
or.b32 %r642, %r641, -2139062144;
add.s32 %r643, %r642, -16843009;
or.b32 %r493, %r643, %r641;
mov.u32 %r101, %r493;

prmt.b32 %r101, %r101, 0, 0xba98;

xor.b32 %r644, %r628, %r101;
or.b32 %r645, %r644, -2139062144;
and.b32 %r646, %r101, 2139062143;
and.b32 %r647, %r631, -2139062144;
sub.s32 %r648, %r645, %r646;
xor.b32 %r104, %r647, %r648;
ld.global.nc.u32 %r105, [%rd36+20];

dp4a.s32.s32 %r103, %r104, %r105, %r97;

and.b32 %r649, %r640, -2143281136;
or.b32 %r650, %r649, -2139062144;
add.s32 %r651, %r650, -16843009;
or.b32 %r499, %r651, %r649;
mov.u32 %r107, %r499;

prmt.b32 %r107, %r107, 0, 0xba98;

xor.b32 %r652, %r107, %r629;
or.b32 %r653, %r652, -2139062144;
and.b32 %r654, %r107, 2139062143;
and.b32 %r655, %r633, -2139062144;
sub.s32 %r656, %r653, %r654;
xor.b32 %r110, %r655, %r656;
ld.global.nc.u32 %r111, [%rd36+24];

dp4a.s32.s32 %r109, %r110, %r111, %r103;

cvt.u32.u16 %r657, %rs14;
mul.wide.u32 %rd42, %r657, 8;
add.s64 %rd43, %rd34, %rd42;
ld.global.nc.v2.u32 {%r658, %r659}, [%rd43];
not.b32 %r661, %r658;
not.b32 %r663, %r659;
shr.u32 %r664, %r568, 5;
and.b32 %r665, %r664, 255;
popc.b32 %r666, %r665;
shl.b32 %r667, %r666, 7;
and.b32 %r668, %r667, 128;
xor.b32 %r669, %r668, %r665;
mul.lo.s32 %r670, %r669, 16843009;
and.b32 %r671, %r670, 134480385;
or.b32 %r672, %r671, -2139062144;
add.s32 %r673, %r672, -16843009;
or.b32 %r505, %r673, %r671;
mov.u32 %r113, %r505;

prmt.b32 %r113, %r113, 0, 0xba98;

xor.b32 %r674, %r658, %r113;
or.b32 %r675, %r674, -2139062144;
and.b32 %r676, %r113, 2139062143;
and.b32 %r677, %r661, -2139062144;
sub.s32 %r678, %r675, %r676;
xor.b32 %r116, %r677, %r678;
ld.global.nc.u32 %r117, [%rd36+28];

dp4a.s32.s32 %r115, %r116, %r117, %r109;

and.b32 %r679, %r670, -2143281136;
or.b32 %r680, %r679, -2139062144;
add.s32 %r681, %r680, -16843009;
or.b32 %r511, %r681, %r679;
mov.u32 %r119, %r511;

prmt.b32 %r119, %r119, 0, 0xba98;

xor.b32 %r682, %r119, %r659;
or.b32 %r683, %r682, -2139062144;
and.b32 %r684, %r119, 2139062143;
and.b32 %r685, %r663, -2139062144;
sub.s32 %r686, %r683, %r684;
xor.b32 %r122, %r685, %r686;
ld.global.nc.u32 %r123, [%rd36+32];

dp4a.s32.s32 %r121, %r122, %r123, %r115;

shr.u32 %r687, %r568, 11;
or.b32 %r688, %r687, 1;
mul.lo.s32 %r689, %r121, %r688;
shr.s32 %r690, %r689, 31;
shr.u32 %r691, %r690, 29;
add.s32 %r692, %r689, %r691;
shr.s32 %r693, %r692, 3;
ld.global.nc.u16 %rs1, [%rd116+-4];

{ cvt.f32.f16 %f61, %rs1;}


ld.global.nc.u32 %r125, [%rd36];

{.reg .f16 low,high;
mov.b32 {low,high},%r125;
cvt.f32.f16 %f62, low;}


mul.ftz.f32 %f81, %f61, %f62;
cvt.rn.f32.s32 %f82, %r693;
fma.rn.ftz.f32 %f230, %f81, %f82, %f230;
add.s64 %rd44, %rd117, %rd6;
ld.global.nc.u16 %rs19, [%rd44+-2];
shr.u16 %rs20, %rs19, 8;
ld.global.nc.u16 %rs21, [%rd44];
shr.u16 %rs22, %rs21, 8;
ld.global.nc.u16 %rs23, [%rd44+2];
cvt.u32.u16 %r694, %rs23;
ld.global.nc.u16 %rs24, [%rd44+4];
cvt.u32.u16 %r695, %rs24;
mov.b32 %r696, {%rs23, %rs24};
shl.b16 %rs25, %rs19, 3;
cvt.u64.u16 %rd45, %rs25;
and.b64 %rd46, %rd45, 2040;
add.s64 %rd47, %rd34, %rd46;
ld.global.nc.v2.u32 {%r697, %r698}, [%rd47];
not.b32 %r700, %r697;
not.b32 %r702, %r698;
and.b32 %r703, %r694, 255;
popc.b32 %r704, %r703;
shl.b32 %r705, %r704, 7;
and.b32 %r706, %r705, 128;
xor.b32 %r707, %r706, %r703;
mul.lo.s32 %r708, %r707, 16843009;
and.b32 %r709, %r708, 134480385;
or.b32 %r710, %r709, -2139062144;
add.s32 %r711, %r710, -16843009;
or.b32 %r518, %r711, %r709;
mov.u32 %r126, %r518;

prmt.b32 %r126, %r126, 0, 0xba98;

xor.b32 %r712, %r697, %r126;
or.b32 %r713, %r712, -2139062144;
and.b32 %r714, %r126, 2139062143;
and.b32 %r715, %r700, -2139062144;
sub.s32 %r716, %r713, %r714;
xor.b32 %r129, %r715, %r716;

dp4a.s32.s32 %r128, %r129, %r81, %r523;

and.b32 %r717, %r708, -2143281136;
or.b32 %r718, %r717, -2139062144;
add.s32 %r719, %r718, -16843009;
or.b32 %r524, %r719, %r717;
mov.u32 %r132, %r524;

prmt.b32 %r132, %r132, 0, 0xba98;

xor.b32 %r720, %r132, %r698;
or.b32 %r721, %r720, -2139062144;
and.b32 %r722, %r132, 2139062143;
and.b32 %r723, %r702, -2139062144;
sub.s32 %r724, %r721, %r722;
xor.b32 %r135, %r723, %r724;

dp4a.s32.s32 %r134, %r135, %r87, %r128;

cvt.u32.u16 %r725, %rs20;
mul.wide.u32 %rd48, %r725, 8;
add.s64 %rd49, %rd34, %rd48;
ld.global.nc.v2.u32 {%r726, %r727}, [%rd49];
not.b32 %r729, %r726;
not.b32 %r731, %r727;
shr.u32 %r732, %r694, 7;
and.b32 %r733, %r732, 255;
popc.b32 %r734, %r733;
shl.b32 %r735, %r734, 7;
and.b32 %r736, %r735, 128;
xor.b32 %r737, %r736, %r733;
mul.lo.s32 %r738, %r737, 16843009;
and.b32 %r739, %r738, 134480385;
or.b32 %r740, %r739, -2139062144;
add.s32 %r741, %r740, -16843009;
or.b32 %r530, %r741, %r739;
mov.u32 %r138, %r530;

prmt.b32 %r138, %r138, 0, 0xba98;

xor.b32 %r742, %r726, %r138;
or.b32 %r743, %r742, -2139062144;
and.b32 %r744, %r138, 2139062143;
and.b32 %r745, %r729, -2139062144;
sub.s32 %r746, %r743, %r744;
xor.b32 %r141, %r745, %r746;

dp4a.s32.s32 %r140, %r141, %r93, %r134;

and.b32 %r747, %r738, -2143281136;
or.b32 %r748, %r747, -2139062144;
add.s32 %r749, %r748, -16843009;
or.b32 %r536, %r749, %r747;
mov.u32 %r144, %r536;

prmt.b32 %r144, %r144, 0, 0xba98;

xor.b32 %r750, %r144, %r727;
or.b32 %r751, %r750, -2139062144;
and.b32 %r752, %r144, 2139062143;
and.b32 %r753, %r731, -2139062144;
sub.s32 %r754, %r751, %r752;
xor.b32 %r147, %r753, %r754;

dp4a.s32.s32 %r146, %r147, %r99, %r140;

shl.b16 %rs26, %rs21, 3;
cvt.u64.u16 %rd50, %rs26;
and.b64 %rd51, %rd50, 2040;
add.s64 %rd52, %rd34, %rd51;
ld.global.nc.v2.u32 {%r755, %r756}, [%rd52];
not.b32 %r758, %r755;
not.b32 %r760, %r756;
shr.u32 %r761, %r696, 14;
and.b32 %r762, %r761, 255;
popc.b32 %r763, %r762;
shl.b32 %r764, %r763, 7;
and.b32 %r765, %r764, 128;
xor.b32 %r766, %r765, %r762;
mul.lo.s32 %r767, %r766, 16843009;
and.b32 %r768, %r767, 134480385;
or.b32 %r769, %r768, -2139062144;
add.s32 %r770, %r769, -16843009;
or.b32 %r542, %r770, %r768;
mov.u32 %r150, %r542;

prmt.b32 %r150, %r150, 0, 0xba98;

xor.b32 %r771, %r755, %r150;
or.b32 %r772, %r771, -2139062144;
and.b32 %r773, %r150, 2139062143;
and.b32 %r774, %r758, -2139062144;
sub.s32 %r775, %r772, %r773;
xor.b32 %r153, %r774, %r775;

dp4a.s32.s32 %r152, %r153, %r105, %r146;

and.b32 %r776, %r767, -2143281136;
or.b32 %r777, %r776, -2139062144;
add.s32 %r778, %r777, -16843009;
or.b32 %r548, %r778, %r776;
mov.u32 %r156, %r548;

prmt.b32 %r156, %r156, 0, 0xba98;

xor.b32 %r779, %r156, %r756;
or.b32 %r780, %r779, -2139062144;
and.b32 %r781, %r156, 2139062143;
and.b32 %r782, %r760, -2139062144;
sub.s32 %r783, %r780, %r781;
xor.b32 %r159, %r782, %r783;

dp4a.s32.s32 %r158, %r159, %r111, %r152;

cvt.u32.u16 %r784, %rs22;
mul.wide.u32 %rd53, %r784, 8;
add.s64 %rd54, %rd34, %rd53;
ld.global.nc.v2.u32 {%r785, %r786}, [%rd54];
not.b32 %r788, %r785;
not.b32 %r790, %r786;
shr.u32 %r791, %r695, 5;
and.b32 %r792, %r791, 255;
popc.b32 %r793, %r792;
shl.b32 %r794, %r793, 7;
and.b32 %r795, %r794, 128;
xor.b32 %r796, %r795, %r792;
mul.lo.s32 %r797, %r796, 16843009;
and.b32 %r798, %r797, 134480385;
or.b32 %r799, %r798, -2139062144;
add.s32 %r800, %r799, -16843009;
or.b32 %r554, %r800, %r798;
mov.u32 %r162, %r554;

prmt.b32 %r162, %r162, 0, 0xba98;

xor.b32 %r801, %r785, %r162;
or.b32 %r802, %r801, -2139062144;
and.b32 %r803, %r162, 2139062143;
and.b32 %r804, %r788, -2139062144;
sub.s32 %r805, %r802, %r803;
xor.b32 %r165, %r804, %r805;

dp4a.s32.s32 %r164, %r165, %r117, %r158;

and.b32 %r806, %r797, -2143281136;
or.b32 %r807, %r806, -2139062144;
add.s32 %r808, %r807, -16843009;
or.b32 %r560, %r808, %r806;
mov.u32 %r168, %r560;

prmt.b32 %r168, %r168, 0, 0xba98;

xor.b32 %r809, %r168, %r786;
or.b32 %r810, %r809, -2139062144;
and.b32 %r811, %r168, 2139062143;
and.b32 %r812, %r790, -2139062144;
sub.s32 %r813, %r810, %r811;
xor.b32 %r171, %r812, %r813;

dp4a.s32.s32 %r170, %r171, %r123, %r164;

shr.u32 %r814, %r695, 11;
or.b32 %r815, %r814, 1;
mul.lo.s32 %r816, %r170, %r815;
shr.s32 %r817, %r816, 31;
shr.u32 %r818, %r817, 29;
add.s32 %r819, %r816, %r818;
shr.s32 %r820, %r819, 3;
ld.global.nc.u16 %rs2, [%rd117];

{ cvt.f32.f16 %f63, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r125;
cvt.f32.f16 %f64, low;}


mul.ftz.f32 %f83, %f63, %f64;
cvt.rn.f32.s32 %f84, %r820;
fma.rn.ftz.f32 %f229, %f83, %f84, %f229;
cvt.u64.u32 %rd55, %r1397;
add.s64 %rd56, %rd1, %rd55;
add.s64 %rd57, %rd56, %rd2;
mul.lo.s64 %rd58, %rd57, 36;
mov.u32 %r175, %r469;

prmt.b32 %r175, %r175, 0, 0xba98;

xor.b32 %r821, %r570, %r175;
or.b32 %r822, %r821, -2139062144;
and.b32 %r823, %r175, 2139062143;
sub.s32 %r824, %r822, %r823;
xor.b32 %r178, %r588, %r824;
add.s64 %rd59, %rd7, %rd58;
ld.global.nc.u32 %r179, [%rd59+4];

dp4a.s32.s32 %r177, %r178, %r179, %r523;

mov.u32 %r181, %r475;

prmt.b32 %r181, %r181, 0, 0xba98;

xor.b32 %r825, %r181, %r571;
or.b32 %r826, %r825, -2139062144;
and.b32 %r827, %r181, 2139062143;
sub.s32 %r828, %r826, %r827;
xor.b32 %r184, %r596, %r828;
ld.global.nc.u32 %r185, [%rd59+8];

dp4a.s32.s32 %r183, %r184, %r185, %r177;

mov.u32 %r187, %r481;

prmt.b32 %r187, %r187, 0, 0xba98;

xor.b32 %r829, %r599, %r187;
or.b32 %r830, %r829, -2139062144;
and.b32 %r831, %r187, 2139062143;
sub.s32 %r832, %r830, %r831;
xor.b32 %r190, %r618, %r832;
ld.global.nc.u32 %r191, [%rd59+12];

dp4a.s32.s32 %r189, %r190, %r191, %r183;

mov.u32 %r193, %r487;

prmt.b32 %r193, %r193, 0, 0xba98;

xor.b32 %r833, %r193, %r600;
or.b32 %r834, %r833, -2139062144;
and.b32 %r835, %r193, 2139062143;
sub.s32 %r836, %r834, %r835;
xor.b32 %r196, %r626, %r836;
ld.global.nc.u32 %r197, [%rd59+16];

dp4a.s32.s32 %r195, %r196, %r197, %r189;

mov.u32 %r199, %r493;

prmt.b32 %r199, %r199, 0, 0xba98;

xor.b32 %r837, %r628, %r199;
or.b32 %r838, %r837, -2139062144;
and.b32 %r839, %r199, 2139062143;
sub.s32 %r840, %r838, %r839;
xor.b32 %r202, %r647, %r840;
ld.global.nc.u32 %r203, [%rd59+20];

dp4a.s32.s32 %r201, %r202, %r203, %r195;

mov.u32 %r205, %r499;

prmt.b32 %r205, %r205, 0, 0xba98;

xor.b32 %r841, %r205, %r629;
or.b32 %r842, %r841, -2139062144;
and.b32 %r843, %r205, 2139062143;
sub.s32 %r844, %r842, %r843;
xor.b32 %r208, %r655, %r844;
ld.global.nc.u32 %r209, [%rd59+24];

dp4a.s32.s32 %r207, %r208, %r209, %r201;

mov.u32 %r211, %r505;

prmt.b32 %r211, %r211, 0, 0xba98;

xor.b32 %r845, %r658, %r211;
or.b32 %r846, %r845, -2139062144;
and.b32 %r847, %r211, 2139062143;
sub.s32 %r848, %r846, %r847;
xor.b32 %r214, %r677, %r848;
ld.global.nc.u32 %r215, [%rd59+28];

dp4a.s32.s32 %r213, %r214, %r215, %r207;

mov.u32 %r217, %r511;

prmt.b32 %r217, %r217, 0, 0xba98;

xor.b32 %r849, %r217, %r659;
or.b32 %r850, %r849, -2139062144;
and.b32 %r851, %r217, 2139062143;
sub.s32 %r852, %r850, %r851;
xor.b32 %r220, %r685, %r852;
ld.global.nc.u32 %r221, [%rd59+32];

dp4a.s32.s32 %r219, %r220, %r221, %r213;

mul.lo.s32 %r853, %r219, %r688;
shr.s32 %r854, %r853, 31;
shr.u32 %r855, %r854, 29;
add.s32 %r856, %r853, %r855;
shr.s32 %r857, %r856, 3;

{ cvt.f32.f16 %f65, %rs1;}


ld.global.nc.u32 %r223, [%rd59];

{.reg .f16 low,high;
mov.b32 {low,high},%r223;
cvt.f32.f16 %f66, low;}


mul.ftz.f32 %f85, %f65, %f66;
cvt.rn.f32.s32 %f86, %r857;
fma.rn.ftz.f32 %f228, %f85, %f86, %f228;
mov.u32 %r224, %r518;

prmt.b32 %r224, %r224, 0, 0xba98;

xor.b32 %r858, %r697, %r224;
or.b32 %r859, %r858, -2139062144;
and.b32 %r860, %r224, 2139062143;
sub.s32 %r861, %r859, %r860;
xor.b32 %r227, %r715, %r861;

dp4a.s32.s32 %r226, %r227, %r179, %r523;

mov.u32 %r230, %r524;

prmt.b32 %r230, %r230, 0, 0xba98;

xor.b32 %r862, %r230, %r698;
or.b32 %r863, %r862, -2139062144;
and.b32 %r864, %r230, 2139062143;
sub.s32 %r865, %r863, %r864;
xor.b32 %r233, %r723, %r865;

dp4a.s32.s32 %r232, %r233, %r185, %r226;

mov.u32 %r236, %r530;

prmt.b32 %r236, %r236, 0, 0xba98;

xor.b32 %r866, %r726, %r236;
or.b32 %r867, %r866, -2139062144;
and.b32 %r868, %r236, 2139062143;
sub.s32 %r869, %r867, %r868;
xor.b32 %r239, %r745, %r869;

dp4a.s32.s32 %r238, %r239, %r191, %r232;

mov.u32 %r242, %r536;

prmt.b32 %r242, %r242, 0, 0xba98;

xor.b32 %r870, %r242, %r727;
or.b32 %r871, %r870, -2139062144;
and.b32 %r872, %r242, 2139062143;
sub.s32 %r873, %r871, %r872;
xor.b32 %r245, %r753, %r873;

dp4a.s32.s32 %r244, %r245, %r197, %r238;

mov.u32 %r248, %r542;

prmt.b32 %r248, %r248, 0, 0xba98;

xor.b32 %r874, %r755, %r248;
or.b32 %r875, %r874, -2139062144;
and.b32 %r876, %r248, 2139062143;
sub.s32 %r877, %r875, %r876;
xor.b32 %r251, %r774, %r877;

dp4a.s32.s32 %r250, %r251, %r203, %r244;

mov.u32 %r254, %r548;

prmt.b32 %r254, %r254, 0, 0xba98;

xor.b32 %r878, %r254, %r756;
or.b32 %r879, %r878, -2139062144;
and.b32 %r880, %r254, 2139062143;
sub.s32 %r881, %r879, %r880;
xor.b32 %r257, %r782, %r881;

dp4a.s32.s32 %r256, %r257, %r209, %r250;

mov.u32 %r260, %r554;

prmt.b32 %r260, %r260, 0, 0xba98;

xor.b32 %r882, %r785, %r260;
or.b32 %r883, %r882, -2139062144;
and.b32 %r884, %r260, 2139062143;
sub.s32 %r885, %r883, %r884;
xor.b32 %r263, %r804, %r885;

dp4a.s32.s32 %r262, %r263, %r215, %r256;

mov.u32 %r266, %r560;

prmt.b32 %r266, %r266, 0, 0xba98;

xor.b32 %r886, %r266, %r786;
or.b32 %r887, %r886, -2139062144;
and.b32 %r888, %r266, 2139062143;
sub.s32 %r889, %r887, %r888;
xor.b32 %r269, %r812, %r889;

dp4a.s32.s32 %r268, %r269, %r221, %r262;

mul.lo.s32 %r890, %r268, %r815;
shr.s32 %r891, %r890, 31;
shr.u32 %r892, %r891, 29;
add.s32 %r893, %r890, %r892;
shr.s32 %r894, %r893, 3;

{ cvt.f32.f16 %f67, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r223;
cvt.f32.f16 %f68, low;}


mul.ftz.f32 %f87, %f67, %f68;
cvt.rn.f32.s32 %f88, %r894;
fma.rn.ftz.f32 %f227, %f87, %f88, %f227;
cvt.u64.u32 %rd60, %r1400;
add.s64 %rd61, %rd1, %rd60;
add.s64 %rd62, %rd61, %rd2;
mul.lo.s64 %rd63, %rd62, 36;
mov.u32 %r273, %r469;

prmt.b32 %r273, %r273, 0, 0xba98;

xor.b32 %r895, %r570, %r273;
or.b32 %r896, %r895, -2139062144;
and.b32 %r897, %r273, 2139062143;
sub.s32 %r898, %r896, %r897;
xor.b32 %r276, %r588, %r898;
add.s64 %rd64, %rd7, %rd63;
ld.global.nc.u32 %r277, [%rd64+4];

dp4a.s32.s32 %r275, %r276, %r277, %r523;

mov.u32 %r279, %r475;

prmt.b32 %r279, %r279, 0, 0xba98;

xor.b32 %r899, %r279, %r571;
or.b32 %r900, %r899, -2139062144;
and.b32 %r901, %r279, 2139062143;
sub.s32 %r902, %r900, %r901;
xor.b32 %r282, %r596, %r902;
ld.global.nc.u32 %r283, [%rd64+8];

dp4a.s32.s32 %r281, %r282, %r283, %r275;

mov.u32 %r285, %r481;

prmt.b32 %r285, %r285, 0, 0xba98;

xor.b32 %r903, %r599, %r285;
or.b32 %r904, %r903, -2139062144;
and.b32 %r905, %r285, 2139062143;
sub.s32 %r906, %r904, %r905;
xor.b32 %r288, %r618, %r906;
ld.global.nc.u32 %r289, [%rd64+12];

dp4a.s32.s32 %r287, %r288, %r289, %r281;

mov.u32 %r291, %r487;

prmt.b32 %r291, %r291, 0, 0xba98;

xor.b32 %r907, %r291, %r600;
or.b32 %r908, %r907, -2139062144;
and.b32 %r909, %r291, 2139062143;
sub.s32 %r910, %r908, %r909;
xor.b32 %r294, %r626, %r910;
ld.global.nc.u32 %r295, [%rd64+16];

dp4a.s32.s32 %r293, %r294, %r295, %r287;

mov.u32 %r297, %r493;

prmt.b32 %r297, %r297, 0, 0xba98;

xor.b32 %r911, %r628, %r297;
or.b32 %r912, %r911, -2139062144;
and.b32 %r913, %r297, 2139062143;
sub.s32 %r914, %r912, %r913;
xor.b32 %r300, %r647, %r914;
ld.global.nc.u32 %r301, [%rd64+20];

dp4a.s32.s32 %r299, %r300, %r301, %r293;

mov.u32 %r303, %r499;

prmt.b32 %r303, %r303, 0, 0xba98;

xor.b32 %r915, %r303, %r629;
or.b32 %r916, %r915, -2139062144;
and.b32 %r917, %r303, 2139062143;
sub.s32 %r918, %r916, %r917;
xor.b32 %r306, %r655, %r918;
ld.global.nc.u32 %r307, [%rd64+24];

dp4a.s32.s32 %r305, %r306, %r307, %r299;

mov.u32 %r309, %r505;

prmt.b32 %r309, %r309, 0, 0xba98;

xor.b32 %r919, %r658, %r309;
or.b32 %r920, %r919, -2139062144;
and.b32 %r921, %r309, 2139062143;
sub.s32 %r922, %r920, %r921;
xor.b32 %r312, %r677, %r922;
ld.global.nc.u32 %r313, [%rd64+28];

dp4a.s32.s32 %r311, %r312, %r313, %r305;

mov.u32 %r315, %r511;

prmt.b32 %r315, %r315, 0, 0xba98;

xor.b32 %r923, %r315, %r659;
or.b32 %r924, %r923, -2139062144;
and.b32 %r925, %r315, 2139062143;
sub.s32 %r926, %r924, %r925;
xor.b32 %r318, %r685, %r926;
ld.global.nc.u32 %r319, [%rd64+32];

dp4a.s32.s32 %r317, %r318, %r319, %r311;

mul.lo.s32 %r927, %r317, %r688;
shr.s32 %r928, %r927, 31;
shr.u32 %r929, %r928, 29;
add.s32 %r930, %r927, %r929;
shr.s32 %r931, %r930, 3;

{ cvt.f32.f16 %f69, %rs1;}


ld.global.nc.u32 %r321, [%rd64];

{.reg .f16 low,high;
mov.b32 {low,high},%r321;
cvt.f32.f16 %f70, low;}


mul.ftz.f32 %f89, %f69, %f70;
cvt.rn.f32.s32 %f90, %r931;
fma.rn.ftz.f32 %f226, %f89, %f90, %f226;
mov.u32 %r322, %r518;

prmt.b32 %r322, %r322, 0, 0xba98;

xor.b32 %r932, %r697, %r322;
or.b32 %r933, %r932, -2139062144;
and.b32 %r934, %r322, 2139062143;
sub.s32 %r935, %r933, %r934;
xor.b32 %r325, %r715, %r935;

dp4a.s32.s32 %r324, %r325, %r277, %r523;

mov.u32 %r328, %r524;

prmt.b32 %r328, %r328, 0, 0xba98;

xor.b32 %r936, %r328, %r698;
or.b32 %r937, %r936, -2139062144;
and.b32 %r938, %r328, 2139062143;
sub.s32 %r939, %r937, %r938;
xor.b32 %r331, %r723, %r939;

dp4a.s32.s32 %r330, %r331, %r283, %r324;

mov.u32 %r334, %r530;

prmt.b32 %r334, %r334, 0, 0xba98;

xor.b32 %r940, %r726, %r334;
or.b32 %r941, %r940, -2139062144;
and.b32 %r942, %r334, 2139062143;
sub.s32 %r943, %r941, %r942;
xor.b32 %r337, %r745, %r943;

dp4a.s32.s32 %r336, %r337, %r289, %r330;

mov.u32 %r340, %r536;

prmt.b32 %r340, %r340, 0, 0xba98;

xor.b32 %r944, %r340, %r727;
or.b32 %r945, %r944, -2139062144;
and.b32 %r946, %r340, 2139062143;
sub.s32 %r947, %r945, %r946;
xor.b32 %r343, %r753, %r947;

dp4a.s32.s32 %r342, %r343, %r295, %r336;

mov.u32 %r346, %r542;

prmt.b32 %r346, %r346, 0, 0xba98;

xor.b32 %r948, %r755, %r346;
or.b32 %r949, %r948, -2139062144;
and.b32 %r950, %r346, 2139062143;
sub.s32 %r951, %r949, %r950;
xor.b32 %r349, %r774, %r951;

dp4a.s32.s32 %r348, %r349, %r301, %r342;

mov.u32 %r352, %r548;

prmt.b32 %r352, %r352, 0, 0xba98;

xor.b32 %r952, %r352, %r756;
or.b32 %r953, %r952, -2139062144;
and.b32 %r954, %r352, 2139062143;
sub.s32 %r955, %r953, %r954;
xor.b32 %r355, %r782, %r955;

dp4a.s32.s32 %r354, %r355, %r307, %r348;

mov.u32 %r358, %r554;

prmt.b32 %r358, %r358, 0, 0xba98;

xor.b32 %r956, %r785, %r358;
or.b32 %r957, %r956, -2139062144;
and.b32 %r958, %r358, 2139062143;
sub.s32 %r959, %r957, %r958;
xor.b32 %r361, %r804, %r959;

dp4a.s32.s32 %r360, %r361, %r313, %r354;

mov.u32 %r364, %r560;

prmt.b32 %r364, %r364, 0, 0xba98;

xor.b32 %r960, %r364, %r786;
or.b32 %r961, %r960, -2139062144;
and.b32 %r962, %r364, 2139062143;
sub.s32 %r963, %r961, %r962;
xor.b32 %r367, %r812, %r963;

dp4a.s32.s32 %r366, %r367, %r319, %r360;

mul.lo.s32 %r964, %r366, %r815;
shr.s32 %r965, %r964, 31;
shr.u32 %r966, %r965, 29;
add.s32 %r967, %r964, %r966;
shr.s32 %r968, %r967, 3;

{ cvt.f32.f16 %f71, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r321;
cvt.f32.f16 %f72, low;}


mul.ftz.f32 %f91, %f71, %f72;
cvt.rn.f32.s32 %f92, %r968;
fma.rn.ftz.f32 %f225, %f91, %f92, %f225;
cvt.u64.u32 %rd65, %r1399;
add.s64 %rd66, %rd1, %rd65;
add.s64 %rd67, %rd66, %rd2;
mul.lo.s64 %rd68, %rd67, 36;
mov.u32 %r371, %r469;

prmt.b32 %r371, %r371, 0, 0xba98;

xor.b32 %r969, %r570, %r371;
or.b32 %r970, %r969, -2139062144;
and.b32 %r971, %r371, 2139062143;
sub.s32 %r972, %r970, %r971;
xor.b32 %r374, %r588, %r972;
add.s64 %rd69, %rd7, %rd68;
ld.global.nc.u32 %r375, [%rd69+4];

dp4a.s32.s32 %r373, %r374, %r375, %r523;

mov.u32 %r377, %r475;

prmt.b32 %r377, %r377, 0, 0xba98;

xor.b32 %r973, %r377, %r571;
or.b32 %r974, %r973, -2139062144;
and.b32 %r975, %r377, 2139062143;
sub.s32 %r976, %r974, %r975;
xor.b32 %r380, %r596, %r976;
ld.global.nc.u32 %r381, [%rd69+8];

dp4a.s32.s32 %r379, %r380, %r381, %r373;

mov.u32 %r383, %r481;

prmt.b32 %r383, %r383, 0, 0xba98;

xor.b32 %r977, %r599, %r383;
or.b32 %r978, %r977, -2139062144;
and.b32 %r979, %r383, 2139062143;
sub.s32 %r980, %r978, %r979;
xor.b32 %r386, %r618, %r980;
ld.global.nc.u32 %r387, [%rd69+12];

dp4a.s32.s32 %r385, %r386, %r387, %r379;

mov.u32 %r389, %r487;

prmt.b32 %r389, %r389, 0, 0xba98;

xor.b32 %r981, %r389, %r600;
or.b32 %r982, %r981, -2139062144;
and.b32 %r983, %r389, 2139062143;
sub.s32 %r984, %r982, %r983;
xor.b32 %r392, %r626, %r984;
ld.global.nc.u32 %r393, [%rd69+16];

dp4a.s32.s32 %r391, %r392, %r393, %r385;

mov.u32 %r395, %r493;

prmt.b32 %r395, %r395, 0, 0xba98;

xor.b32 %r985, %r628, %r395;
or.b32 %r986, %r985, -2139062144;
and.b32 %r987, %r395, 2139062143;
sub.s32 %r988, %r986, %r987;
xor.b32 %r398, %r647, %r988;
ld.global.nc.u32 %r399, [%rd69+20];

dp4a.s32.s32 %r397, %r398, %r399, %r391;

mov.u32 %r401, %r499;

prmt.b32 %r401, %r401, 0, 0xba98;

xor.b32 %r989, %r401, %r629;
or.b32 %r990, %r989, -2139062144;
and.b32 %r991, %r401, 2139062143;
sub.s32 %r992, %r990, %r991;
xor.b32 %r404, %r655, %r992;
ld.global.nc.u32 %r405, [%rd69+24];

dp4a.s32.s32 %r403, %r404, %r405, %r397;

mov.u32 %r407, %r505;

prmt.b32 %r407, %r407, 0, 0xba98;

xor.b32 %r993, %r658, %r407;
or.b32 %r994, %r993, -2139062144;
and.b32 %r995, %r407, 2139062143;
sub.s32 %r996, %r994, %r995;
xor.b32 %r410, %r677, %r996;
ld.global.nc.u32 %r411, [%rd69+28];

dp4a.s32.s32 %r409, %r410, %r411, %r403;

mov.u32 %r413, %r511;

prmt.b32 %r413, %r413, 0, 0xba98;

xor.b32 %r997, %r413, %r659;
or.b32 %r998, %r997, -2139062144;
and.b32 %r999, %r413, 2139062143;
sub.s32 %r1000, %r998, %r999;
xor.b32 %r416, %r685, %r1000;
ld.global.nc.u32 %r417, [%rd69+32];

dp4a.s32.s32 %r415, %r416, %r417, %r409;

mul.lo.s32 %r1001, %r415, %r688;
shr.s32 %r1002, %r1001, 31;
shr.u32 %r1003, %r1002, 29;
add.s32 %r1004, %r1001, %r1003;
shr.s32 %r1005, %r1004, 3;

{ cvt.f32.f16 %f73, %rs1;}


ld.global.nc.u32 %r419, [%rd69];

{.reg .f16 low,high;
mov.b32 {low,high},%r419;
cvt.f32.f16 %f74, low;}


mul.ftz.f32 %f93, %f73, %f74;
cvt.rn.f32.s32 %f94, %r1005;
fma.rn.ftz.f32 %f224, %f93, %f94, %f224;
mov.u32 %r420, %r518;

prmt.b32 %r420, %r420, 0, 0xba98;

xor.b32 %r1006, %r697, %r420;
or.b32 %r1007, %r1006, -2139062144;
and.b32 %r1008, %r420, 2139062143;
sub.s32 %r1009, %r1007, %r1008;
xor.b32 %r423, %r715, %r1009;

dp4a.s32.s32 %r422, %r423, %r375, %r523;

mov.u32 %r426, %r524;

prmt.b32 %r426, %r426, 0, 0xba98;

xor.b32 %r1010, %r426, %r698;
or.b32 %r1011, %r1010, -2139062144;
and.b32 %r1012, %r426, 2139062143;
sub.s32 %r1013, %r1011, %r1012;
xor.b32 %r429, %r723, %r1013;

dp4a.s32.s32 %r428, %r429, %r381, %r422;

mov.u32 %r432, %r530;

prmt.b32 %r432, %r432, 0, 0xba98;

xor.b32 %r1014, %r726, %r432;
or.b32 %r1015, %r1014, -2139062144;
and.b32 %r1016, %r432, 2139062143;
sub.s32 %r1017, %r1015, %r1016;
xor.b32 %r435, %r745, %r1017;

dp4a.s32.s32 %r434, %r435, %r387, %r428;

mov.u32 %r438, %r536;

prmt.b32 %r438, %r438, 0, 0xba98;

xor.b32 %r1018, %r438, %r727;
or.b32 %r1019, %r1018, -2139062144;
and.b32 %r1020, %r438, 2139062143;
sub.s32 %r1021, %r1019, %r1020;
xor.b32 %r441, %r753, %r1021;

dp4a.s32.s32 %r440, %r441, %r393, %r434;

mov.u32 %r444, %r542;

prmt.b32 %r444, %r444, 0, 0xba98;

xor.b32 %r1022, %r755, %r444;
or.b32 %r1023, %r1022, -2139062144;
and.b32 %r1024, %r444, 2139062143;
sub.s32 %r1025, %r1023, %r1024;
xor.b32 %r447, %r774, %r1025;

dp4a.s32.s32 %r446, %r447, %r399, %r440;

mov.u32 %r450, %r548;

prmt.b32 %r450, %r450, 0, 0xba98;

xor.b32 %r1026, %r450, %r756;
or.b32 %r1027, %r1026, -2139062144;
and.b32 %r1028, %r450, 2139062143;
sub.s32 %r1029, %r1027, %r1028;
xor.b32 %r453, %r782, %r1029;

dp4a.s32.s32 %r452, %r453, %r405, %r446;

mov.u32 %r456, %r554;

prmt.b32 %r456, %r456, 0, 0xba98;

xor.b32 %r1030, %r785, %r456;
or.b32 %r1031, %r1030, -2139062144;
and.b32 %r1032, %r456, 2139062143;
sub.s32 %r1033, %r1031, %r1032;
xor.b32 %r459, %r804, %r1033;

dp4a.s32.s32 %r458, %r459, %r411, %r452;

mov.u32 %r462, %r560;

prmt.b32 %r462, %r462, 0, 0xba98;

xor.b32 %r1034, %r462, %r786;
or.b32 %r1035, %r1034, -2139062144;
and.b32 %r1036, %r462, 2139062143;
sub.s32 %r1037, %r1035, %r1036;
xor.b32 %r465, %r812, %r1037;

dp4a.s32.s32 %r464, %r465, %r417, %r458;

mul.lo.s32 %r1038, %r464, %r815;
shr.s32 %r1039, %r1038, 31;
shr.u32 %r1040, %r1039, 29;
add.s32 %r1041, %r1038, %r1040;
shr.s32 %r1042, %r1041, 3;

{ cvt.f32.f16 %f75, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r419;
cvt.f32.f16 %f76, low;}


mul.ftz.f32 %f95, %f75, %f76;
cvt.rn.f32.s32 %f96, %r1042;
fma.rn.ftz.f32 %f223, %f95, %f96, %f223;
cvt.u64.u32 %rd70, %r1398;
add.s64 %rd71, %rd1, %rd70;
add.s64 %rd72, %rd71, %rd2;
mul.lo.s64 %rd73, %rd72, 36;

prmt.b32 %r469, %r469, 0, 0xba98;

xor.b32 %r1043, %r570, %r469;
or.b32 %r1044, %r1043, -2139062144;
and.b32 %r1045, %r469, 2139062143;
sub.s32 %r1046, %r1044, %r1045;
xor.b32 %r472, %r588, %r1046;
add.s64 %rd74, %rd7, %rd73;
ld.global.nc.u32 %r473, [%rd74+4];

dp4a.s32.s32 %r471, %r472, %r473, %r523;


prmt.b32 %r475, %r475, 0, 0xba98;

xor.b32 %r1047, %r475, %r571;
or.b32 %r1048, %r1047, -2139062144;
and.b32 %r1049, %r475, 2139062143;
sub.s32 %r1050, %r1048, %r1049;
xor.b32 %r478, %r596, %r1050;
ld.global.nc.u32 %r479, [%rd74+8];

dp4a.s32.s32 %r477, %r478, %r479, %r471;


prmt.b32 %r481, %r481, 0, 0xba98;

xor.b32 %r1051, %r599, %r481;
or.b32 %r1052, %r1051, -2139062144;
and.b32 %r1053, %r481, 2139062143;
sub.s32 %r1054, %r1052, %r1053;
xor.b32 %r484, %r618, %r1054;
ld.global.nc.u32 %r485, [%rd74+12];

dp4a.s32.s32 %r483, %r484, %r485, %r477;


prmt.b32 %r487, %r487, 0, 0xba98;

xor.b32 %r1055, %r487, %r600;
or.b32 %r1056, %r1055, -2139062144;
and.b32 %r1057, %r487, 2139062143;
sub.s32 %r1058, %r1056, %r1057;
xor.b32 %r490, %r626, %r1058;
ld.global.nc.u32 %r491, [%rd74+16];

dp4a.s32.s32 %r489, %r490, %r491, %r483;


prmt.b32 %r493, %r493, 0, 0xba98;

xor.b32 %r1059, %r628, %r493;
or.b32 %r1060, %r1059, -2139062144;
and.b32 %r1061, %r493, 2139062143;
sub.s32 %r1062, %r1060, %r1061;
xor.b32 %r496, %r647, %r1062;
ld.global.nc.u32 %r497, [%rd74+20];

dp4a.s32.s32 %r495, %r496, %r497, %r489;


prmt.b32 %r499, %r499, 0, 0xba98;

xor.b32 %r1063, %r499, %r629;
or.b32 %r1064, %r1063, -2139062144;
and.b32 %r1065, %r499, 2139062143;
sub.s32 %r1066, %r1064, %r1065;
xor.b32 %r502, %r655, %r1066;
ld.global.nc.u32 %r503, [%rd74+24];

dp4a.s32.s32 %r501, %r502, %r503, %r495;


prmt.b32 %r505, %r505, 0, 0xba98;

xor.b32 %r1067, %r658, %r505;
or.b32 %r1068, %r1067, -2139062144;
and.b32 %r1069, %r505, 2139062143;
sub.s32 %r1070, %r1068, %r1069;
xor.b32 %r508, %r677, %r1070;
ld.global.nc.u32 %r509, [%rd74+28];

dp4a.s32.s32 %r507, %r508, %r509, %r501;


prmt.b32 %r511, %r511, 0, 0xba98;

xor.b32 %r1071, %r511, %r659;
or.b32 %r1072, %r1071, -2139062144;
and.b32 %r1073, %r511, 2139062143;
sub.s32 %r1074, %r1072, %r1073;
xor.b32 %r514, %r685, %r1074;
ld.global.nc.u32 %r515, [%rd74+32];

dp4a.s32.s32 %r513, %r514, %r515, %r507;

mul.lo.s32 %r1075, %r513, %r688;
shr.s32 %r1076, %r1075, 31;
shr.u32 %r1077, %r1076, 29;
add.s32 %r1078, %r1075, %r1077;
shr.s32 %r1079, %r1078, 3;

{ cvt.f32.f16 %f77, %rs1;}


ld.global.nc.u32 %r517, [%rd74];

{.reg .f16 low,high;
mov.b32 {low,high},%r517;
cvt.f32.f16 %f78, low;}


mul.ftz.f32 %f97, %f77, %f78;
cvt.rn.f32.s32 %f98, %r1079;
fma.rn.ftz.f32 %f222, %f97, %f98, %f222;

prmt.b32 %r518, %r518, 0, 0xba98;

xor.b32 %r1080, %r697, %r518;
or.b32 %r1081, %r1080, -2139062144;
and.b32 %r1082, %r518, 2139062143;
sub.s32 %r1083, %r1081, %r1082;
xor.b32 %r521, %r715, %r1083;

dp4a.s32.s32 %r520, %r521, %r473, %r523;


prmt.b32 %r524, %r524, 0, 0xba98;

xor.b32 %r1084, %r524, %r698;
or.b32 %r1085, %r1084, -2139062144;
and.b32 %r1086, %r524, 2139062143;
sub.s32 %r1087, %r1085, %r1086;
xor.b32 %r527, %r723, %r1087;

dp4a.s32.s32 %r526, %r527, %r479, %r520;


prmt.b32 %r530, %r530, 0, 0xba98;

xor.b32 %r1088, %r726, %r530;
or.b32 %r1089, %r1088, -2139062144;
and.b32 %r1090, %r530, 2139062143;
sub.s32 %r1091, %r1089, %r1090;
xor.b32 %r533, %r745, %r1091;

dp4a.s32.s32 %r532, %r533, %r485, %r526;


prmt.b32 %r536, %r536, 0, 0xba98;

xor.b32 %r1092, %r536, %r727;
or.b32 %r1093, %r1092, -2139062144;
and.b32 %r1094, %r536, 2139062143;
sub.s32 %r1095, %r1093, %r1094;
xor.b32 %r539, %r753, %r1095;

dp4a.s32.s32 %r538, %r539, %r491, %r532;


prmt.b32 %r542, %r542, 0, 0xba98;

xor.b32 %r1096, %r755, %r542;
or.b32 %r1097, %r1096, -2139062144;
and.b32 %r1098, %r542, 2139062143;
sub.s32 %r1099, %r1097, %r1098;
xor.b32 %r545, %r774, %r1099;

dp4a.s32.s32 %r544, %r545, %r497, %r538;


prmt.b32 %r548, %r548, 0, 0xba98;

xor.b32 %r1100, %r548, %r756;
or.b32 %r1101, %r1100, -2139062144;
and.b32 %r1102, %r548, 2139062143;
sub.s32 %r1103, %r1101, %r1102;
xor.b32 %r551, %r782, %r1103;

dp4a.s32.s32 %r550, %r551, %r503, %r544;


prmt.b32 %r554, %r554, 0, 0xba98;

xor.b32 %r1104, %r785, %r554;
or.b32 %r1105, %r1104, -2139062144;
and.b32 %r1106, %r554, 2139062143;
sub.s32 %r1107, %r1105, %r1106;
xor.b32 %r557, %r804, %r1107;

dp4a.s32.s32 %r556, %r557, %r509, %r550;


prmt.b32 %r560, %r560, 0, 0xba98;

xor.b32 %r1108, %r560, %r786;
or.b32 %r1109, %r1108, -2139062144;
and.b32 %r1110, %r560, 2139062143;
sub.s32 %r1111, %r1109, %r1110;
xor.b32 %r563, %r812, %r1111;

dp4a.s32.s32 %r562, %r563, %r515, %r556;

mul.lo.s32 %r1112, %r562, %r815;
shr.s32 %r1113, %r1112, 31;
shr.u32 %r1114, %r1113, 29;
add.s32 %r1115, %r1112, %r1114;
shr.s32 %r1116, %r1115, 3;

{ cvt.f32.f16 %f79, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r517;
cvt.f32.f16 %f80, low;}


mul.ftz.f32 %f99, %f79, %f80;
cvt.rn.f32.s32 %f100, %r1116;
fma.rn.ftz.f32 %f221, %f99, %f100, %f221;
add.s32 %r1400, %r1400, 64;
add.s32 %r1399, %r1399, 64;
add.s32 %r1398, %r1398, 64;
add.s32 %r1397, %r1397, 64;
add.s32 %r1396, %r1396, 64;
add.s64 %rd117, %rd117, 528;
add.s64 %rd116, %rd116, 528;
add.s64 %rd115, %rd115, 528;
add.s32 %r1401, %r1401, 8;
setp.lt.s32 %p3, %r1401, %r2;
@%p3 bra $L__BB207_2;

$L__BB207_3:
setp.eq.s32 %p4, %r37, 0;
@%p4 bra $L__BB207_5;

mov.u32 %r1120, _ZZ13mul_mat_vec_qIL9ggml_type16ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r1121, %r37, 1280, %r1120;
shl.b32 %r1122, %r39, 2;
add.s32 %r1123, %r1121, %r1122;
st.shared.f32 [%r1123+-1280], %f230;
st.shared.f32 [%r1123+-1152], %f229;
st.shared.f32 [%r1123+-1024], %f228;
st.shared.f32 [%r1123+-896], %f227;
st.shared.f32 [%r1123+-768], %f226;
st.shared.f32 [%r1123+-640], %f225;
st.shared.f32 [%r1123+-512], %f224;
st.shared.f32 [%r1123+-384], %f223;
st.shared.f32 [%r1123+-256], %f222;
st.shared.f32 [%r1123+-128], %f221;

$L__BB207_5:
setp.ne.s32 %p5, %r37, 0;
bar.sync 0;
@%p5 bra $L__BB207_26;

mov.u32 %r1125, %ctaid.x;
shl.b32 %r1126, %r1125, 1;
mov.u32 %r1127, 1;
shl.b32 %r1129, %r39, 2;
mov.u32 %r1130, _ZZ13mul_mat_vec_qIL9ggml_type16ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r1131, %r1130, %r1129;
mov.u32 %r1132, 2;
ld.shared.f32 %f101, [%r1131];
add.ftz.f32 %f102, %f101, %f230;
mov.b32 %r1133, %f102;
mov.u32 %r1134, 31;
mov.u32 %r1135, 16;
mov.u32 %r1136, -1;
shfl.sync.bfly.b32 %r1137|%p6, %r1133, %r1135, %r1134, %r1136;
mov.b32 %f103, %r1137;
add.ftz.f32 %f104, %f102, %f103;
mov.b32 %r1138, %f104;
mov.u32 %r1139, 8;
shfl.sync.bfly.b32 %r1140|%p7, %r1138, %r1139, %r1134, %r1136;
mov.b32 %f105, %r1140;
add.ftz.f32 %f106, %f104, %f105;
mov.b32 %r1141, %f106;
mov.u32 %r1142, 4;
shfl.sync.bfly.b32 %r1143|%p8, %r1141, %r1142, %r1134, %r1136;
mov.b32 %f107, %r1143;
add.ftz.f32 %f108, %f106, %f107;
mov.b32 %r1144, %f108;
shfl.sync.bfly.b32 %r1145|%p9, %r1144, %r1132, %r1134, %r1136;
mov.b32 %f109, %r1145;
add.ftz.f32 %f110, %f108, %f109;
mov.b32 %r1146, %f110;
shfl.sync.bfly.b32 %r1147|%p10, %r1146, %r1127, %r1134, %r1136;
mov.b32 %f111, %r1147;
add.ftz.f32 %f31, %f110, %f111;
setp.ne.s32 %p11, %r39, 0;
setp.ge.u32 %p12, %r1126, %r23;
or.pred %p13, %p11, %p12;
@%p13 bra $L__BB207_8;

cvta.to.global.u64 %rd75, %rd18;
mov.u32 %r1148, %ctaid.y;
mad.lo.s32 %r1151, %r1148, %r29, %r1126;
mov.u32 %r1152, %ctaid.z;
mad.lo.s32 %r1153, %r1152, %r35, %r1151;
mul.wide.u32 %rd76, %r1153, 4;
add.s64 %rd77, %rd75, %rd76;
st.global.f32 [%rd77], %f31;

$L__BB207_8:
ld.shared.f32 %f112, [%r1131+128];
add.ftz.f32 %f113, %f112, %f229;
mov.b32 %r1159, %f113;
shfl.sync.bfly.b32 %r1163|%p14, %r1159, %r1135, %r1134, %r1136;
mov.b32 %f114, %r1163;
add.ftz.f32 %f115, %f113, %f114;
mov.b32 %r1164, %f115;
shfl.sync.bfly.b32 %r1166|%p15, %r1164, %r1139, %r1134, %r1136;
mov.b32 %f116, %r1166;
add.ftz.f32 %f117, %f115, %f116;
mov.b32 %r1167, %f117;
shfl.sync.bfly.b32 %r1169|%p16, %r1167, %r1142, %r1134, %r1136;
mov.b32 %f118, %r1169;
add.ftz.f32 %f119, %f117, %f118;
mov.b32 %r1170, %f119;
shfl.sync.bfly.b32 %r1171|%p17, %r1170, %r1132, %r1134, %r1136;
mov.b32 %f120, %r1171;
add.ftz.f32 %f121, %f119, %f120;
mov.b32 %r1172, %f121;
shfl.sync.bfly.b32 %r1174|%p18, %r1172, %r1127, %r1134, %r1136;
mov.b32 %f122, %r1174;
add.ftz.f32 %f32, %f121, %f122;
setp.eq.s32 %p19, %r39, 1;
or.b32 %r1177, %r1126, 1;
setp.lt.u32 %p20, %r1177, %r23;
and.pred %p1, %p19, %p20;
not.pred %p21, %p1;
@%p21 bra $L__BB207_10;

cvta.to.global.u64 %rd78, %rd18;
mov.u32 %r1178, %ctaid.y;
mad.lo.s32 %r1181, %r1178, %r29, %r1126;
mov.u32 %r1182, %ctaid.z;
mad.lo.s32 %r1183, %r1182, %r35, %r1181;
mul.wide.u32 %rd79, %r1183, 4;
add.s64 %rd80, %rd78, %rd79;
st.global.f32 [%rd80+4], %f32;

$L__BB207_10:
mov.u32 %r1188, 2;
ld.shared.f32 %f123, [%r1131+256];
add.ftz.f32 %f124, %f123, %f228;
mov.b32 %r1189, %f124;
mov.u32 %r1190, 31;
mov.u32 %r1191, 16;
mov.u32 %r1192, -1;
shfl.sync.bfly.b32 %r1193|%p22, %r1189, %r1191, %r1190, %r1192;
mov.b32 %f125, %r1193;
add.ftz.f32 %f126, %f124, %f125;
mov.b32 %r1194, %f126;
mov.u32 %r1195, 8;
shfl.sync.bfly.b32 %r1196|%p23, %r1194, %r1195, %r1190, %r1192;
mov.b32 %f127, %r1196;
add.ftz.f32 %f128, %f126, %f127;
mov.b32 %r1197, %f128;
mov.u32 %r1198, 4;
shfl.sync.bfly.b32 %r1199|%p24, %r1197, %r1198, %r1190, %r1192;
mov.b32 %f129, %r1199;
add.ftz.f32 %f130, %f128, %f129;
mov.b32 %r1200, %f130;
shfl.sync.bfly.b32 %r1201|%p25, %r1200, %r1188, %r1190, %r1192;
mov.b32 %f131, %r1201;
add.ftz.f32 %f132, %f130, %f131;
mov.b32 %r1202, %f132;
mov.u32 %r1203, 1;
shfl.sync.bfly.b32 %r1204|%p26, %r1202, %r1203, %r1190, %r1192;
mov.b32 %f133, %r1204;
add.ftz.f32 %f33, %f132, %f133;
@%p13 bra $L__BB207_12;

cvt.u64.u32 %rd81, %r23;
mov.u32 %r1207, %ctaid.y;
mad.lo.s32 %r1210, %r1207, %r29, %r1126;
mov.u32 %r1211, %ctaid.z;
mad.lo.s32 %r1212, %r1211, %r35, %r1210;
cvt.u64.u32 %rd82, %r1212;
add.s64 %rd83, %rd81, %rd82;
cvta.to.global.u64 %rd84, %rd18;
shl.b64 %rd85, %rd83, 2;
add.s64 %rd86, %rd84, %rd85;
st.global.f32 [%rd86], %f33;

$L__BB207_12:
ld.shared.f32 %f134, [%r1131+384];
add.ftz.f32 %f135, %f134, %f227;
mov.b32 %r1218, %f135;
shfl.sync.bfly.b32 %r1222|%p30, %r1218, %r1191, %r1190, %r1192;
mov.b32 %f136, %r1222;
add.ftz.f32 %f137, %f135, %f136;
mov.b32 %r1223, %f137;
shfl.sync.bfly.b32 %r1225|%p31, %r1223, %r1195, %r1190, %r1192;
mov.b32 %f138, %r1225;
add.ftz.f32 %f139, %f137, %f138;
mov.b32 %r1226, %f139;
shfl.sync.bfly.b32 %r1228|%p32, %r1226, %r1198, %r1190, %r1192;
mov.b32 %f140, %r1228;
add.ftz.f32 %f141, %f139, %f140;
mov.b32 %r1229, %f141;
shfl.sync.bfly.b32 %r1230|%p33, %r1229, %r1188, %r1190, %r1192;
mov.b32 %f142, %r1230;
add.ftz.f32 %f143, %f141, %f142;
mov.b32 %r1231, %f143;
shfl.sync.bfly.b32 %r1233|%p34, %r1231, %r1203, %r1190, %r1192;
mov.b32 %f144, %r1233;
add.ftz.f32 %f34, %f143, %f144;
@%p21 bra $L__BB207_14;

add.s32 %r1234, %r23, 1;
cvt.u64.u32 %rd87, %r1234;
mov.u32 %r1235, %ctaid.y;
mad.lo.s32 %r1238, %r1235, %r29, %r1126;
mov.u32 %r1239, %ctaid.z;
mad.lo.s32 %r1240, %r1239, %r35, %r1238;
cvt.u64.u32 %rd88, %r1240;
add.s64 %rd89, %rd87, %rd88;
cvta.to.global.u64 %rd90, %rd18;
shl.b64 %rd91, %rd89, 2;
add.s64 %rd92, %rd90, %rd91;
st.global.f32 [%rd92], %f34;

$L__BB207_14:
shl.b32 %r1241, %r23, 1;
mov.u32 %r1242, 1;
mov.u32 %r1247, 2;
ld.shared.f32 %f145, [%r1131+512];
add.ftz.f32 %f146, %f145, %f226;
mov.b32 %r1248, %f146;
mov.u32 %r1249, 31;
mov.u32 %r1250, 16;
mov.u32 %r1251, -1;
shfl.sync.bfly.b32 %r1252|%p36, %r1248, %r1250, %r1249, %r1251;
mov.b32 %f147, %r1252;
add.ftz.f32 %f148, %f146, %f147;
mov.b32 %r1253, %f148;
mov.u32 %r1254, 8;
shfl.sync.bfly.b32 %r1255|%p37, %r1253, %r1254, %r1249, %r1251;
mov.b32 %f149, %r1255;
add.ftz.f32 %f150, %f148, %f149;
mov.b32 %r1256, %f150;
mov.u32 %r1257, 4;
shfl.sync.bfly.b32 %r1258|%p38, %r1256, %r1257, %r1249, %r1251;
mov.b32 %f151, %r1258;
add.ftz.f32 %f152, %f150, %f151;
mov.b32 %r1259, %f152;
shfl.sync.bfly.b32 %r1260|%p39, %r1259, %r1247, %r1249, %r1251;
mov.b32 %f153, %r1260;
add.ftz.f32 %f154, %f152, %f153;
mov.b32 %r1261, %f154;
shfl.sync.bfly.b32 %r1262|%p40, %r1261, %r1242, %r1249, %r1251;
mov.b32 %f155, %r1262;
add.ftz.f32 %f35, %f154, %f155;
cvt.u64.u32 %rd93, %r1241;
mov.u32 %r1263, %ctaid.y;
mad.lo.s32 %r1266, %r1263, %r29, %r1126;
mov.u32 %r1267, %ctaid.z;
mad.lo.s32 %r1268, %r1267, %r35, %r1266;
cvt.u64.u32 %rd94, %r1268;
add.s64 %rd95, %rd93, %rd94;
cvta.to.global.u64 %rd96, %rd18;
shl.b64 %rd97, %rd95, 2;
add.s64 %rd14, %rd96, %rd97;
@%p13 bra $L__BB207_16;

st.global.f32 [%rd14], %f35;

$L__BB207_16:
ld.shared.f32 %f156, [%r1131+640];
add.ftz.f32 %f157, %f156, %f225;
mov.b32 %r1274, %f157;
shfl.sync.bfly.b32 %r1278|%p44, %r1274, %r1250, %r1249, %r1251;
mov.b32 %f158, %r1278;
add.ftz.f32 %f159, %f157, %f158;
mov.b32 %r1279, %f159;
shfl.sync.bfly.b32 %r1281|%p45, %r1279, %r1254, %r1249, %r1251;
mov.b32 %f160, %r1281;
add.ftz.f32 %f161, %f159, %f160;
mov.b32 %r1282, %f161;
shfl.sync.bfly.b32 %r1284|%p46, %r1282, %r1257, %r1249, %r1251;
mov.b32 %f162, %r1284;
add.ftz.f32 %f163, %f161, %f162;
mov.b32 %r1285, %f163;
shfl.sync.bfly.b32 %r1286|%p47, %r1285, %r1247, %r1249, %r1251;
mov.b32 %f164, %r1286;
add.ftz.f32 %f165, %f163, %f164;
mov.b32 %r1287, %f165;
shfl.sync.bfly.b32 %r1289|%p48, %r1287, %r1242, %r1249, %r1251;
mov.b32 %f166, %r1289;
add.ftz.f32 %f36, %f165, %f166;
@%p21 bra $L__BB207_18;

st.global.f32 [%rd14+4], %f36;

$L__BB207_18:
mul.lo.s32 %r20, %r23, 3;
mov.u32 %r1294, 2;
ld.shared.f32 %f167, [%r1131+768];
add.ftz.f32 %f168, %f167, %f224;
mov.b32 %r1295, %f168;
mov.u32 %r1296, 31;
mov.u32 %r1297, 16;
mov.u32 %r1298, -1;
shfl.sync.bfly.b32 %r1299|%p50, %r1295, %r1297, %r1296, %r1298;
mov.b32 %f169, %r1299;
add.ftz.f32 %f170, %f168, %f169;
mov.b32 %r1300, %f170;
mov.u32 %r1301, 8;
shfl.sync.bfly.b32 %r1302|%p51, %r1300, %r1301, %r1296, %r1298;
mov.b32 %f171, %r1302;
add.ftz.f32 %f172, %f170, %f171;
mov.b32 %r1303, %f172;
mov.u32 %r1304, 4;
shfl.sync.bfly.b32 %r1305|%p52, %r1303, %r1304, %r1296, %r1298;
mov.b32 %f173, %r1305;
add.ftz.f32 %f174, %f172, %f173;
mov.b32 %r1306, %f174;
shfl.sync.bfly.b32 %r1307|%p53, %r1306, %r1294, %r1296, %r1298;
mov.b32 %f175, %r1307;
add.ftz.f32 %f176, %f174, %f175;
mov.b32 %r1308, %f176;
mov.u32 %r1309, 1;
shfl.sync.bfly.b32 %r1310|%p54, %r1308, %r1309, %r1296, %r1298;
mov.b32 %f177, %r1310;
add.ftz.f32 %f37, %f176, %f177;
@%p13 bra $L__BB207_20;

cvt.u64.u32 %rd98, %r20;
add.s64 %rd100, %rd98, %rd94;
shl.b64 %rd102, %rd100, 2;
add.s64 %rd103, %rd96, %rd102;
st.global.f32 [%rd103], %f37;

$L__BB207_20:
ld.shared.f32 %f178, [%r1131+896];
add.ftz.f32 %f179, %f178, %f223;
mov.b32 %r1324, %f179;
shfl.sync.bfly.b32 %r1328|%p58, %r1324, %r1297, %r1296, %r1298;
mov.b32 %f180, %r1328;
add.ftz.f32 %f181, %f179, %f180;
mov.b32 %r1329, %f181;
shfl.sync.bfly.b32 %r1331|%p59, %r1329, %r1301, %r1296, %r1298;
mov.b32 %f182, %r1331;
add.ftz.f32 %f183, %f181, %f182;
mov.b32 %r1332, %f183;
shfl.sync.bfly.b32 %r1334|%p60, %r1332, %r1304, %r1296, %r1298;
mov.b32 %f184, %r1334;
add.ftz.f32 %f185, %f183, %f184;
mov.b32 %r1335, %f185;
shfl.sync.bfly.b32 %r1336|%p61, %r1335, %r1294, %r1296, %r1298;
mov.b32 %f186, %r1336;
add.ftz.f32 %f187, %f185, %f186;
mov.b32 %r1337, %f187;
shfl.sync.bfly.b32 %r1339|%p62, %r1337, %r1309, %r1296, %r1298;
mov.b32 %f188, %r1339;
add.ftz.f32 %f38, %f187, %f188;
@%p21 bra $L__BB207_22;

add.s32 %r1340, %r20, 1;
cvt.u64.u32 %rd104, %r1340;
add.s64 %rd106, %rd104, %rd94;
shl.b64 %rd108, %rd106, 2;
add.s64 %rd109, %rd96, %rd108;
st.global.f32 [%rd109], %f38;

$L__BB207_22:
shl.b32 %r1347, %r23, 2;
mov.u32 %r1348, 2;
ld.shared.f32 %f189, [%r1131+1024];
add.ftz.f32 %f190, %f189, %f222;
mov.b32 %r1353, %f190;
mov.u32 %r1354, 31;
mov.u32 %r1355, 16;
mov.u32 %r1356, -1;
shfl.sync.bfly.b32 %r1357|%p64, %r1353, %r1355, %r1354, %r1356;
mov.b32 %f191, %r1357;
add.ftz.f32 %f192, %f190, %f191;
mov.b32 %r1358, %f192;
mov.u32 %r1359, 8;
shfl.sync.bfly.b32 %r1360|%p65, %r1358, %r1359, %r1354, %r1356;
mov.b32 %f193, %r1360;
add.ftz.f32 %f194, %f192, %f193;
mov.b32 %r1361, %f194;
mov.u32 %r1362, 4;
shfl.sync.bfly.b32 %r1363|%p66, %r1361, %r1362, %r1354, %r1356;
mov.b32 %f195, %r1363;
add.ftz.f32 %f196, %f194, %f195;
mov.b32 %r1364, %f196;
shfl.sync.bfly.b32 %r1365|%p67, %r1364, %r1348, %r1354, %r1356;
mov.b32 %f197, %r1365;
add.ftz.f32 %f198, %f196, %f197;
mov.b32 %r1366, %f198;
mov.u32 %r1367, 1;
shfl.sync.bfly.b32 %r1368|%p68, %r1366, %r1367, %r1354, %r1356;
mov.b32 %f199, %r1368;
add.ftz.f32 %f39, %f198, %f199;
cvt.u64.u32 %rd110, %r1347;
add.s64 %rd112, %rd110, %rd94;
shl.b64 %rd114, %rd112, 2;
add.s64 %rd15, %rd96, %rd114;
@%p13 bra $L__BB207_24;

st.global.f32 [%rd15], %f39;

$L__BB207_24:
ld.shared.f32 %f200, [%r1131+1152];
add.ftz.f32 %f201, %f200, %f221;
mov.b32 %r1380, %f201;
shfl.sync.bfly.b32 %r1384|%p72, %r1380, %r1355, %r1354, %r1356;
mov.b32 %f202, %r1384;
add.ftz.f32 %f203, %f201, %f202;
mov.b32 %r1385, %f203;
shfl.sync.bfly.b32 %r1387|%p73, %r1385, %r1359, %r1354, %r1356;
mov.b32 %f204, %r1387;
add.ftz.f32 %f205, %f203, %f204;
mov.b32 %r1388, %f205;
shfl.sync.bfly.b32 %r1390|%p74, %r1388, %r1362, %r1354, %r1356;
mov.b32 %f206, %r1390;
add.ftz.f32 %f207, %f205, %f206;
mov.b32 %r1391, %f207;
shfl.sync.bfly.b32 %r1392|%p75, %r1391, %r1348, %r1354, %r1356;
mov.b32 %f208, %r1392;
add.ftz.f32 %f209, %f207, %f208;
mov.b32 %r1393, %f209;
shfl.sync.bfly.b32 %r1395|%p76, %r1393, %r1367, %r1354, %r1356;
mov.b32 %f210, %r1395;
add.ftz.f32 %f40, %f209, %f210;
@%p21 bra $L__BB207_26;

st.global.f32 [%rd15+4], %f40;

$L__BB207_26:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type16ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type16ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type16ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type16ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type16ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type16ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type16ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type16ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type16ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<92>;
.reg .b16 %rs<29>;
.reg .f32 %f<277>;
.reg .b32 %r<1636>;
.reg .b64 %rd<135>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type16ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type16ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type16ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type16ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type16ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type16ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type16ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type16ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type16ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type16ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type16ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type16ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type16ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type16ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type16ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type16ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type16ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type16ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r41, %tid.y;
shl.b32 %r42, %r41, 5;
mov.u32 %r43, %tid.x;
add.s32 %r44, %r42, %r43;
shr.s32 %r45, %r44, 31;
shr.u32 %r46, %r45, 29;
add.s32 %r47, %r44, %r46;
shr.s32 %r1635, %r47, 3;
shr.u32 %r2, %r40, 8;
setp.ge.s32 %p2, %r1635, %r2;
mov.f32 %f265, 0f00000000;
mov.f32 %f266, %f265;
mov.f32 %f267, %f265;
mov.f32 %f268, %f265;
mov.f32 %f269, %f265;
mov.f32 %f270, %f265;
mov.f32 %f271, %f265;
mov.f32 %f272, %f265;
mov.f32 %f273, %f265;
mov.f32 %f274, %f265;
mov.f32 %f275, %f265;
mov.f32 %f276, %f265;
@%p2 bra $L__BB208_3;

mov.u32 %r48, %ctaid.y;
mov.u32 %r49, %ctaid.z;
mul.lo.s32 %r50, %r48, %r32;
cvt.u64.u32 %rd19, %r50;
mul.lo.s32 %r51, %r49, %r38;
cvt.u64.u32 %rd20, %r51;
add.s64 %rd1, %rd20, %rd19;
mul.hi.u32 %r52, %r49, %r34;
add.s32 %r53, %r52, %r49;
shr.u32 %r54, %r53, %r35;
mul.lo.s32 %r55, %r54, %r37;
mul.hi.u32 %r56, %r48, %r28;
add.s32 %r57, %r56, %r48;
shr.u32 %r58, %r57, %r29;
mul.lo.s32 %r59, %r58, %r31;
cvta.to.global.u64 %rd21, %rd16;
and.b32 %r67, %r47, -8;
sub.s32 %r68, %r44, %r67;
shl.b32 %r69, %r68, 2;
cvt.s64.s32 %rd2, %r68;
mov.u32 %r70, %ctaid.x;
shl.b32 %r71, %r70, 1;
shl.b32 %r72, %r26, 1;
shl.b32 %r1629, %r1635, 3;
add.s32 %r1634, %r1629, %r72;
mad.lo.s32 %r1633, %r26, 3, %r1629;
shl.b32 %r73, %r26, 2;
add.s32 %r1632, %r1629, %r73;
mad.lo.s32 %r1631, %r26, 5, %r1629;
add.s32 %r1630, %r26, %r1629;
add.s32 %r74, %r1635, %r55;
add.s32 %r75, %r74, %r59;
or.b32 %r76, %r71, 1;
mad.lo.s32 %r77, %r25, %r76, %r75;
mul.wide.s32 %rd22, %r77, 66;
add.s64 %rd134, %rd21, %rd22;
mad.lo.s32 %r78, %r71, %r25, %r1635;
add.s32 %r79, %r78, %r55;
add.s32 %r80, %r79, %r59;
mul.wide.s32 %rd23, %r80, 66;
add.s64 %rd24, %rd21, %rd23;
add.s64 %rd133, %rd24, 4;
mul.wide.s32 %rd25, %r69, 2;
add.s64 %rd26, %rd23, %rd25;
add.s64 %rd27, %rd21, %rd26;
add.s64 %rd132, %rd27, 8;
or.b64 %rd6, %rd25, 4;
cvta.to.global.u64 %rd7, %rd17;
mov.f32 %f265, 0f00000000;

$L__BB208_2:
cvt.u64.u32 %rd28, %r1629;
add.s64 %rd29, %rd1, %rd28;
add.s64 %rd30, %rd29, %rd2;
mul.lo.s64 %rd31, %rd30, 36;
ld.global.nc.u16 %rs13, [%rd132+-6];
shr.u16 %rs14, %rs13, 8;
ld.global.nc.u16 %rs15, [%rd132+-4];
shr.u16 %rs16, %rs15, 8;
ld.global.nc.u16 %rs17, [%rd132+-2];
cvt.u32.u16 %r669, %rs17;
ld.global.nc.u16 %rs18, [%rd132];
cvt.u32.u16 %r670, %rs18;
mov.b32 %r671, {%rs17, %rs18};
shl.b16 %rs19, %rs13, 3;
cvt.u64.u16 %rd32, %rs19;
and.b64 %rd33, %rd32, 2040;
mov.u64 %rd34, iq2xxs_grid;
add.s64 %rd35, %rd34, %rd33;
ld.global.nc.v2.u32 {%r672, %r673}, [%rd35];
not.b32 %r675, %r672;
mov.u32 %r625, 0;
not.b32 %r677, %r673;
and.b32 %r678, %r669, 255;
popc.b32 %r679, %r678;
shl.b32 %r680, %r679, 7;
and.b32 %r681, %r680, 128;
xor.b32 %r682, %r681, %r678;
mul.lo.s32 %r683, %r682, 16843009;
and.b32 %r684, %r683, 134480385;
or.b32 %r685, %r684, -2139062144;
add.s32 %r686, %r685, -16843009;
or.b32 %r571, %r686, %r684;
mov.u32 %r81, %r571;

prmt.b32 %r81, %r81, 0, 0xba98;

xor.b32 %r687, %r672, %r81;
or.b32 %r688, %r687, -2139062144;
and.b32 %r689, %r81, 2139062143;
and.b32 %r690, %r675, -2139062144;
sub.s32 %r691, %r688, %r689;
xor.b32 %r84, %r690, %r691;
add.s64 %rd36, %rd7, %rd31;
ld.global.nc.u32 %r85, [%rd36+4];

dp4a.s32.s32 %r83, %r84, %r85, %r625;

and.b32 %r692, %r683, -2143281136;
or.b32 %r693, %r692, -2139062144;
add.s32 %r694, %r693, -16843009;
or.b32 %r577, %r694, %r692;
mov.u32 %r87, %r577;

prmt.b32 %r87, %r87, 0, 0xba98;

xor.b32 %r695, %r87, %r673;
or.b32 %r696, %r695, -2139062144;
and.b32 %r697, %r87, 2139062143;
and.b32 %r698, %r677, -2139062144;
sub.s32 %r699, %r696, %r697;
xor.b32 %r90, %r698, %r699;
ld.global.nc.u32 %r91, [%rd36+8];

dp4a.s32.s32 %r89, %r90, %r91, %r83;

cvt.u32.u16 %r700, %rs14;
mul.wide.u32 %rd37, %r700, 8;
add.s64 %rd38, %rd34, %rd37;
ld.global.nc.v2.u32 {%r701, %r702}, [%rd38];
not.b32 %r704, %r701;
not.b32 %r706, %r702;
shr.u32 %r707, %r669, 7;
and.b32 %r708, %r707, 255;
popc.b32 %r709, %r708;
shl.b32 %r710, %r709, 7;
and.b32 %r711, %r710, 128;
xor.b32 %r712, %r711, %r708;
mul.lo.s32 %r713, %r712, 16843009;
and.b32 %r714, %r713, 134480385;
or.b32 %r715, %r714, -2139062144;
add.s32 %r716, %r715, -16843009;
or.b32 %r583, %r716, %r714;
mov.u32 %r93, %r583;

prmt.b32 %r93, %r93, 0, 0xba98;

xor.b32 %r717, %r701, %r93;
or.b32 %r718, %r717, -2139062144;
and.b32 %r719, %r93, 2139062143;
and.b32 %r720, %r704, -2139062144;
sub.s32 %r721, %r718, %r719;
xor.b32 %r96, %r720, %r721;
ld.global.nc.u32 %r97, [%rd36+12];

dp4a.s32.s32 %r95, %r96, %r97, %r89;

and.b32 %r722, %r713, -2143281136;
or.b32 %r723, %r722, -2139062144;
add.s32 %r724, %r723, -16843009;
or.b32 %r589, %r724, %r722;
mov.u32 %r99, %r589;

prmt.b32 %r99, %r99, 0, 0xba98;

xor.b32 %r725, %r99, %r702;
or.b32 %r726, %r725, -2139062144;
and.b32 %r727, %r99, 2139062143;
and.b32 %r728, %r706, -2139062144;
sub.s32 %r729, %r726, %r727;
xor.b32 %r102, %r728, %r729;
ld.global.nc.u32 %r103, [%rd36+16];

dp4a.s32.s32 %r101, %r102, %r103, %r95;

shl.b16 %rs20, %rs15, 3;
cvt.u64.u16 %rd39, %rs20;
and.b64 %rd40, %rd39, 2040;
add.s64 %rd41, %rd34, %rd40;
ld.global.nc.v2.u32 {%r730, %r731}, [%rd41];
not.b32 %r733, %r730;
not.b32 %r735, %r731;
shr.u32 %r736, %r671, 14;
and.b32 %r737, %r736, 255;
popc.b32 %r738, %r737;
shl.b32 %r739, %r738, 7;
and.b32 %r740, %r739, 128;
xor.b32 %r741, %r740, %r737;
mul.lo.s32 %r742, %r741, 16843009;
and.b32 %r743, %r742, 134480385;
or.b32 %r744, %r743, -2139062144;
add.s32 %r745, %r744, -16843009;
or.b32 %r595, %r745, %r743;
mov.u32 %r105, %r595;

prmt.b32 %r105, %r105, 0, 0xba98;

xor.b32 %r746, %r730, %r105;
or.b32 %r747, %r746, -2139062144;
and.b32 %r748, %r105, 2139062143;
and.b32 %r749, %r733, -2139062144;
sub.s32 %r750, %r747, %r748;
xor.b32 %r108, %r749, %r750;
ld.global.nc.u32 %r109, [%rd36+20];

dp4a.s32.s32 %r107, %r108, %r109, %r101;

and.b32 %r751, %r742, -2143281136;
or.b32 %r752, %r751, -2139062144;
add.s32 %r753, %r752, -16843009;
or.b32 %r601, %r753, %r751;
mov.u32 %r111, %r601;

prmt.b32 %r111, %r111, 0, 0xba98;

xor.b32 %r754, %r111, %r731;
or.b32 %r755, %r754, -2139062144;
and.b32 %r756, %r111, 2139062143;
and.b32 %r757, %r735, -2139062144;
sub.s32 %r758, %r755, %r756;
xor.b32 %r114, %r757, %r758;
ld.global.nc.u32 %r115, [%rd36+24];

dp4a.s32.s32 %r113, %r114, %r115, %r107;

cvt.u32.u16 %r759, %rs16;
mul.wide.u32 %rd42, %r759, 8;
add.s64 %rd43, %rd34, %rd42;
ld.global.nc.v2.u32 {%r760, %r761}, [%rd43];
not.b32 %r763, %r760;
not.b32 %r765, %r761;
shr.u32 %r766, %r670, 5;
and.b32 %r767, %r766, 255;
popc.b32 %r768, %r767;
shl.b32 %r769, %r768, 7;
and.b32 %r770, %r769, 128;
xor.b32 %r771, %r770, %r767;
mul.lo.s32 %r772, %r771, 16843009;
and.b32 %r773, %r772, 134480385;
or.b32 %r774, %r773, -2139062144;
add.s32 %r775, %r774, -16843009;
or.b32 %r607, %r775, %r773;
mov.u32 %r117, %r607;

prmt.b32 %r117, %r117, 0, 0xba98;

xor.b32 %r776, %r760, %r117;
or.b32 %r777, %r776, -2139062144;
and.b32 %r778, %r117, 2139062143;
and.b32 %r779, %r763, -2139062144;
sub.s32 %r780, %r777, %r778;
xor.b32 %r120, %r779, %r780;
ld.global.nc.u32 %r121, [%rd36+28];

dp4a.s32.s32 %r119, %r120, %r121, %r113;

and.b32 %r781, %r772, -2143281136;
or.b32 %r782, %r781, -2139062144;
add.s32 %r783, %r782, -16843009;
or.b32 %r613, %r783, %r781;
mov.u32 %r123, %r613;

prmt.b32 %r123, %r123, 0, 0xba98;

xor.b32 %r784, %r123, %r761;
or.b32 %r785, %r784, -2139062144;
and.b32 %r786, %r123, 2139062143;
and.b32 %r787, %r765, -2139062144;
sub.s32 %r788, %r785, %r786;
xor.b32 %r126, %r787, %r788;
ld.global.nc.u32 %r127, [%rd36+32];

dp4a.s32.s32 %r125, %r126, %r127, %r119;

shr.u32 %r789, %r670, 11;
or.b32 %r790, %r789, 1;
mul.lo.s32 %r791, %r125, %r790;
shr.s32 %r792, %r791, 31;
shr.u32 %r793, %r792, 29;
add.s32 %r794, %r791, %r793;
shr.s32 %r795, %r794, 3;
ld.global.nc.u16 %rs1, [%rd133+-4];

{ cvt.f32.f16 %f73, %rs1;}


ld.global.nc.u32 %r129, [%rd36];

{.reg .f16 low,high;
mov.b32 {low,high},%r129;
cvt.f32.f16 %f74, low;}


mul.ftz.f32 %f97, %f73, %f74;
cvt.rn.f32.s32 %f98, %r795;
fma.rn.ftz.f32 %f276, %f97, %f98, %f276;
add.s64 %rd44, %rd134, %rd6;
ld.global.nc.u16 %rs21, [%rd44+-2];
shr.u16 %rs22, %rs21, 8;
ld.global.nc.u16 %rs23, [%rd44];
shr.u16 %rs24, %rs23, 8;
ld.global.nc.u16 %rs25, [%rd44+2];
cvt.u32.u16 %r796, %rs25;
ld.global.nc.u16 %rs26, [%rd44+4];
cvt.u32.u16 %r797, %rs26;
mov.b32 %r798, {%rs25, %rs26};
shl.b16 %rs27, %rs21, 3;
cvt.u64.u16 %rd45, %rs27;
and.b64 %rd46, %rd45, 2040;
add.s64 %rd47, %rd34, %rd46;
ld.global.nc.v2.u32 {%r799, %r800}, [%rd47];
not.b32 %r802, %r799;
not.b32 %r804, %r800;
and.b32 %r805, %r796, 255;
popc.b32 %r806, %r805;
shl.b32 %r807, %r806, 7;
and.b32 %r808, %r807, 128;
xor.b32 %r809, %r808, %r805;
mul.lo.s32 %r810, %r809, 16843009;
and.b32 %r811, %r810, 134480385;
or.b32 %r812, %r811, -2139062144;
add.s32 %r813, %r812, -16843009;
or.b32 %r620, %r813, %r811;
mov.u32 %r130, %r620;

prmt.b32 %r130, %r130, 0, 0xba98;

xor.b32 %r814, %r799, %r130;
or.b32 %r815, %r814, -2139062144;
and.b32 %r816, %r130, 2139062143;
and.b32 %r817, %r802, -2139062144;
sub.s32 %r818, %r815, %r816;
xor.b32 %r133, %r817, %r818;

dp4a.s32.s32 %r132, %r133, %r85, %r625;

and.b32 %r819, %r810, -2143281136;
or.b32 %r820, %r819, -2139062144;
add.s32 %r821, %r820, -16843009;
or.b32 %r626, %r821, %r819;
mov.u32 %r136, %r626;

prmt.b32 %r136, %r136, 0, 0xba98;

xor.b32 %r822, %r136, %r800;
or.b32 %r823, %r822, -2139062144;
and.b32 %r824, %r136, 2139062143;
and.b32 %r825, %r804, -2139062144;
sub.s32 %r826, %r823, %r824;
xor.b32 %r139, %r825, %r826;

dp4a.s32.s32 %r138, %r139, %r91, %r132;

cvt.u32.u16 %r827, %rs22;
mul.wide.u32 %rd48, %r827, 8;
add.s64 %rd49, %rd34, %rd48;
ld.global.nc.v2.u32 {%r828, %r829}, [%rd49];
not.b32 %r831, %r828;
not.b32 %r833, %r829;
shr.u32 %r834, %r796, 7;
and.b32 %r835, %r834, 255;
popc.b32 %r836, %r835;
shl.b32 %r837, %r836, 7;
and.b32 %r838, %r837, 128;
xor.b32 %r839, %r838, %r835;
mul.lo.s32 %r840, %r839, 16843009;
and.b32 %r841, %r840, 134480385;
or.b32 %r842, %r841, -2139062144;
add.s32 %r843, %r842, -16843009;
or.b32 %r632, %r843, %r841;
mov.u32 %r142, %r632;

prmt.b32 %r142, %r142, 0, 0xba98;

xor.b32 %r844, %r828, %r142;
or.b32 %r845, %r844, -2139062144;
and.b32 %r846, %r142, 2139062143;
and.b32 %r847, %r831, -2139062144;
sub.s32 %r848, %r845, %r846;
xor.b32 %r145, %r847, %r848;

dp4a.s32.s32 %r144, %r145, %r97, %r138;

and.b32 %r849, %r840, -2143281136;
or.b32 %r850, %r849, -2139062144;
add.s32 %r851, %r850, -16843009;
or.b32 %r638, %r851, %r849;
mov.u32 %r148, %r638;

prmt.b32 %r148, %r148, 0, 0xba98;

xor.b32 %r852, %r148, %r829;
or.b32 %r853, %r852, -2139062144;
and.b32 %r854, %r148, 2139062143;
and.b32 %r855, %r833, -2139062144;
sub.s32 %r856, %r853, %r854;
xor.b32 %r151, %r855, %r856;

dp4a.s32.s32 %r150, %r151, %r103, %r144;

shl.b16 %rs28, %rs23, 3;
cvt.u64.u16 %rd50, %rs28;
and.b64 %rd51, %rd50, 2040;
add.s64 %rd52, %rd34, %rd51;
ld.global.nc.v2.u32 {%r857, %r858}, [%rd52];
not.b32 %r860, %r857;
not.b32 %r862, %r858;
shr.u32 %r863, %r798, 14;
and.b32 %r864, %r863, 255;
popc.b32 %r865, %r864;
shl.b32 %r866, %r865, 7;
and.b32 %r867, %r866, 128;
xor.b32 %r868, %r867, %r864;
mul.lo.s32 %r869, %r868, 16843009;
and.b32 %r870, %r869, 134480385;
or.b32 %r871, %r870, -2139062144;
add.s32 %r872, %r871, -16843009;
or.b32 %r644, %r872, %r870;
mov.u32 %r154, %r644;

prmt.b32 %r154, %r154, 0, 0xba98;

xor.b32 %r873, %r857, %r154;
or.b32 %r874, %r873, -2139062144;
and.b32 %r875, %r154, 2139062143;
and.b32 %r876, %r860, -2139062144;
sub.s32 %r877, %r874, %r875;
xor.b32 %r157, %r876, %r877;

dp4a.s32.s32 %r156, %r157, %r109, %r150;

and.b32 %r878, %r869, -2143281136;
or.b32 %r879, %r878, -2139062144;
add.s32 %r880, %r879, -16843009;
or.b32 %r650, %r880, %r878;
mov.u32 %r160, %r650;

prmt.b32 %r160, %r160, 0, 0xba98;

xor.b32 %r881, %r160, %r858;
or.b32 %r882, %r881, -2139062144;
and.b32 %r883, %r160, 2139062143;
and.b32 %r884, %r862, -2139062144;
sub.s32 %r885, %r882, %r883;
xor.b32 %r163, %r884, %r885;

dp4a.s32.s32 %r162, %r163, %r115, %r156;

cvt.u32.u16 %r886, %rs24;
mul.wide.u32 %rd53, %r886, 8;
add.s64 %rd54, %rd34, %rd53;
ld.global.nc.v2.u32 {%r887, %r888}, [%rd54];
not.b32 %r890, %r887;
not.b32 %r892, %r888;
shr.u32 %r893, %r797, 5;
and.b32 %r894, %r893, 255;
popc.b32 %r895, %r894;
shl.b32 %r896, %r895, 7;
and.b32 %r897, %r896, 128;
xor.b32 %r898, %r897, %r894;
mul.lo.s32 %r899, %r898, 16843009;
and.b32 %r900, %r899, 134480385;
or.b32 %r901, %r900, -2139062144;
add.s32 %r902, %r901, -16843009;
or.b32 %r656, %r902, %r900;
mov.u32 %r166, %r656;

prmt.b32 %r166, %r166, 0, 0xba98;

xor.b32 %r903, %r887, %r166;
or.b32 %r904, %r903, -2139062144;
and.b32 %r905, %r166, 2139062143;
and.b32 %r906, %r890, -2139062144;
sub.s32 %r907, %r904, %r905;
xor.b32 %r169, %r906, %r907;

dp4a.s32.s32 %r168, %r169, %r121, %r162;

and.b32 %r908, %r899, -2143281136;
or.b32 %r909, %r908, -2139062144;
add.s32 %r910, %r909, -16843009;
or.b32 %r662, %r910, %r908;
mov.u32 %r172, %r662;

prmt.b32 %r172, %r172, 0, 0xba98;

xor.b32 %r911, %r172, %r888;
or.b32 %r912, %r911, -2139062144;
and.b32 %r913, %r172, 2139062143;
and.b32 %r914, %r892, -2139062144;
sub.s32 %r915, %r912, %r913;
xor.b32 %r175, %r914, %r915;

dp4a.s32.s32 %r174, %r175, %r127, %r168;

shr.u32 %r916, %r797, 11;
or.b32 %r917, %r916, 1;
mul.lo.s32 %r918, %r174, %r917;
shr.s32 %r919, %r918, 31;
shr.u32 %r920, %r919, 29;
add.s32 %r921, %r918, %r920;
shr.s32 %r922, %r921, 3;
ld.global.nc.u16 %rs2, [%rd134];

{ cvt.f32.f16 %f75, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r129;
cvt.f32.f16 %f76, low;}


mul.ftz.f32 %f99, %f75, %f76;
cvt.rn.f32.s32 %f100, %r922;
fma.rn.ftz.f32 %f275, %f99, %f100, %f275;
cvt.u64.u32 %rd55, %r1630;
add.s64 %rd56, %rd1, %rd55;
add.s64 %rd57, %rd56, %rd2;
mul.lo.s64 %rd58, %rd57, 36;
mov.u32 %r179, %r571;

prmt.b32 %r179, %r179, 0, 0xba98;

xor.b32 %r923, %r672, %r179;
or.b32 %r924, %r923, -2139062144;
and.b32 %r925, %r179, 2139062143;
sub.s32 %r926, %r924, %r925;
xor.b32 %r182, %r690, %r926;
add.s64 %rd59, %rd7, %rd58;
ld.global.nc.u32 %r183, [%rd59+4];

dp4a.s32.s32 %r181, %r182, %r183, %r625;

mov.u32 %r185, %r577;

prmt.b32 %r185, %r185, 0, 0xba98;

xor.b32 %r927, %r185, %r673;
or.b32 %r928, %r927, -2139062144;
and.b32 %r929, %r185, 2139062143;
sub.s32 %r930, %r928, %r929;
xor.b32 %r188, %r698, %r930;
ld.global.nc.u32 %r189, [%rd59+8];

dp4a.s32.s32 %r187, %r188, %r189, %r181;

mov.u32 %r191, %r583;

prmt.b32 %r191, %r191, 0, 0xba98;

xor.b32 %r931, %r701, %r191;
or.b32 %r932, %r931, -2139062144;
and.b32 %r933, %r191, 2139062143;
sub.s32 %r934, %r932, %r933;
xor.b32 %r194, %r720, %r934;
ld.global.nc.u32 %r195, [%rd59+12];

dp4a.s32.s32 %r193, %r194, %r195, %r187;

mov.u32 %r197, %r589;

prmt.b32 %r197, %r197, 0, 0xba98;

xor.b32 %r935, %r197, %r702;
or.b32 %r936, %r935, -2139062144;
and.b32 %r937, %r197, 2139062143;
sub.s32 %r938, %r936, %r937;
xor.b32 %r200, %r728, %r938;
ld.global.nc.u32 %r201, [%rd59+16];

dp4a.s32.s32 %r199, %r200, %r201, %r193;

mov.u32 %r203, %r595;

prmt.b32 %r203, %r203, 0, 0xba98;

xor.b32 %r939, %r730, %r203;
or.b32 %r940, %r939, -2139062144;
and.b32 %r941, %r203, 2139062143;
sub.s32 %r942, %r940, %r941;
xor.b32 %r206, %r749, %r942;
ld.global.nc.u32 %r207, [%rd59+20];

dp4a.s32.s32 %r205, %r206, %r207, %r199;

mov.u32 %r209, %r601;

prmt.b32 %r209, %r209, 0, 0xba98;

xor.b32 %r943, %r209, %r731;
or.b32 %r944, %r943, -2139062144;
and.b32 %r945, %r209, 2139062143;
sub.s32 %r946, %r944, %r945;
xor.b32 %r212, %r757, %r946;
ld.global.nc.u32 %r213, [%rd59+24];

dp4a.s32.s32 %r211, %r212, %r213, %r205;

mov.u32 %r215, %r607;

prmt.b32 %r215, %r215, 0, 0xba98;

xor.b32 %r947, %r760, %r215;
or.b32 %r948, %r947, -2139062144;
and.b32 %r949, %r215, 2139062143;
sub.s32 %r950, %r948, %r949;
xor.b32 %r218, %r779, %r950;
ld.global.nc.u32 %r219, [%rd59+28];

dp4a.s32.s32 %r217, %r218, %r219, %r211;

mov.u32 %r221, %r613;

prmt.b32 %r221, %r221, 0, 0xba98;

xor.b32 %r951, %r221, %r761;
or.b32 %r952, %r951, -2139062144;
and.b32 %r953, %r221, 2139062143;
sub.s32 %r954, %r952, %r953;
xor.b32 %r224, %r787, %r954;
ld.global.nc.u32 %r225, [%rd59+32];

dp4a.s32.s32 %r223, %r224, %r225, %r217;

mul.lo.s32 %r955, %r223, %r790;
shr.s32 %r956, %r955, 31;
shr.u32 %r957, %r956, 29;
add.s32 %r958, %r955, %r957;
shr.s32 %r959, %r958, 3;

{ cvt.f32.f16 %f77, %rs1;}


ld.global.nc.u32 %r227, [%rd59];

{.reg .f16 low,high;
mov.b32 {low,high},%r227;
cvt.f32.f16 %f78, low;}


mul.ftz.f32 %f101, %f77, %f78;
cvt.rn.f32.s32 %f102, %r959;
fma.rn.ftz.f32 %f274, %f101, %f102, %f274;
mov.u32 %r228, %r620;

prmt.b32 %r228, %r228, 0, 0xba98;

xor.b32 %r960, %r799, %r228;
or.b32 %r961, %r960, -2139062144;
and.b32 %r962, %r228, 2139062143;
sub.s32 %r963, %r961, %r962;
xor.b32 %r231, %r817, %r963;

dp4a.s32.s32 %r230, %r231, %r183, %r625;

mov.u32 %r234, %r626;

prmt.b32 %r234, %r234, 0, 0xba98;

xor.b32 %r964, %r234, %r800;
or.b32 %r965, %r964, -2139062144;
and.b32 %r966, %r234, 2139062143;
sub.s32 %r967, %r965, %r966;
xor.b32 %r237, %r825, %r967;

dp4a.s32.s32 %r236, %r237, %r189, %r230;

mov.u32 %r240, %r632;

prmt.b32 %r240, %r240, 0, 0xba98;

xor.b32 %r968, %r828, %r240;
or.b32 %r969, %r968, -2139062144;
and.b32 %r970, %r240, 2139062143;
sub.s32 %r971, %r969, %r970;
xor.b32 %r243, %r847, %r971;

dp4a.s32.s32 %r242, %r243, %r195, %r236;

mov.u32 %r246, %r638;

prmt.b32 %r246, %r246, 0, 0xba98;

xor.b32 %r972, %r246, %r829;
or.b32 %r973, %r972, -2139062144;
and.b32 %r974, %r246, 2139062143;
sub.s32 %r975, %r973, %r974;
xor.b32 %r249, %r855, %r975;

dp4a.s32.s32 %r248, %r249, %r201, %r242;

mov.u32 %r252, %r644;

prmt.b32 %r252, %r252, 0, 0xba98;

xor.b32 %r976, %r857, %r252;
or.b32 %r977, %r976, -2139062144;
and.b32 %r978, %r252, 2139062143;
sub.s32 %r979, %r977, %r978;
xor.b32 %r255, %r876, %r979;

dp4a.s32.s32 %r254, %r255, %r207, %r248;

mov.u32 %r258, %r650;

prmt.b32 %r258, %r258, 0, 0xba98;

xor.b32 %r980, %r258, %r858;
or.b32 %r981, %r980, -2139062144;
and.b32 %r982, %r258, 2139062143;
sub.s32 %r983, %r981, %r982;
xor.b32 %r261, %r884, %r983;

dp4a.s32.s32 %r260, %r261, %r213, %r254;

mov.u32 %r264, %r656;

prmt.b32 %r264, %r264, 0, 0xba98;

xor.b32 %r984, %r887, %r264;
or.b32 %r985, %r984, -2139062144;
and.b32 %r986, %r264, 2139062143;
sub.s32 %r987, %r985, %r986;
xor.b32 %r267, %r906, %r987;

dp4a.s32.s32 %r266, %r267, %r219, %r260;

mov.u32 %r270, %r662;

prmt.b32 %r270, %r270, 0, 0xba98;

xor.b32 %r988, %r270, %r888;
or.b32 %r989, %r988, -2139062144;
and.b32 %r990, %r270, 2139062143;
sub.s32 %r991, %r989, %r990;
xor.b32 %r273, %r914, %r991;

dp4a.s32.s32 %r272, %r273, %r225, %r266;

mul.lo.s32 %r992, %r272, %r917;
shr.s32 %r993, %r992, 31;
shr.u32 %r994, %r993, 29;
add.s32 %r995, %r992, %r994;
shr.s32 %r996, %r995, 3;

{ cvt.f32.f16 %f79, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r227;
cvt.f32.f16 %f80, low;}


mul.ftz.f32 %f103, %f79, %f80;
cvt.rn.f32.s32 %f104, %r996;
fma.rn.ftz.f32 %f273, %f103, %f104, %f273;
cvt.u64.u32 %rd60, %r1634;
add.s64 %rd61, %rd1, %rd60;
add.s64 %rd62, %rd61, %rd2;
mul.lo.s64 %rd63, %rd62, 36;
mov.u32 %r277, %r571;

prmt.b32 %r277, %r277, 0, 0xba98;

xor.b32 %r997, %r672, %r277;
or.b32 %r998, %r997, -2139062144;
and.b32 %r999, %r277, 2139062143;
sub.s32 %r1000, %r998, %r999;
xor.b32 %r280, %r690, %r1000;
add.s64 %rd64, %rd7, %rd63;
ld.global.nc.u32 %r281, [%rd64+4];

dp4a.s32.s32 %r279, %r280, %r281, %r625;

mov.u32 %r283, %r577;

prmt.b32 %r283, %r283, 0, 0xba98;

xor.b32 %r1001, %r283, %r673;
or.b32 %r1002, %r1001, -2139062144;
and.b32 %r1003, %r283, 2139062143;
sub.s32 %r1004, %r1002, %r1003;
xor.b32 %r286, %r698, %r1004;
ld.global.nc.u32 %r287, [%rd64+8];

dp4a.s32.s32 %r285, %r286, %r287, %r279;

mov.u32 %r289, %r583;

prmt.b32 %r289, %r289, 0, 0xba98;

xor.b32 %r1005, %r701, %r289;
or.b32 %r1006, %r1005, -2139062144;
and.b32 %r1007, %r289, 2139062143;
sub.s32 %r1008, %r1006, %r1007;
xor.b32 %r292, %r720, %r1008;
ld.global.nc.u32 %r293, [%rd64+12];

dp4a.s32.s32 %r291, %r292, %r293, %r285;

mov.u32 %r295, %r589;

prmt.b32 %r295, %r295, 0, 0xba98;

xor.b32 %r1009, %r295, %r702;
or.b32 %r1010, %r1009, -2139062144;
and.b32 %r1011, %r295, 2139062143;
sub.s32 %r1012, %r1010, %r1011;
xor.b32 %r298, %r728, %r1012;
ld.global.nc.u32 %r299, [%rd64+16];

dp4a.s32.s32 %r297, %r298, %r299, %r291;

mov.u32 %r301, %r595;

prmt.b32 %r301, %r301, 0, 0xba98;

xor.b32 %r1013, %r730, %r301;
or.b32 %r1014, %r1013, -2139062144;
and.b32 %r1015, %r301, 2139062143;
sub.s32 %r1016, %r1014, %r1015;
xor.b32 %r304, %r749, %r1016;
ld.global.nc.u32 %r305, [%rd64+20];

dp4a.s32.s32 %r303, %r304, %r305, %r297;

mov.u32 %r307, %r601;

prmt.b32 %r307, %r307, 0, 0xba98;

xor.b32 %r1017, %r307, %r731;
or.b32 %r1018, %r1017, -2139062144;
and.b32 %r1019, %r307, 2139062143;
sub.s32 %r1020, %r1018, %r1019;
xor.b32 %r310, %r757, %r1020;
ld.global.nc.u32 %r311, [%rd64+24];

dp4a.s32.s32 %r309, %r310, %r311, %r303;

mov.u32 %r313, %r607;

prmt.b32 %r313, %r313, 0, 0xba98;

xor.b32 %r1021, %r760, %r313;
or.b32 %r1022, %r1021, -2139062144;
and.b32 %r1023, %r313, 2139062143;
sub.s32 %r1024, %r1022, %r1023;
xor.b32 %r316, %r779, %r1024;
ld.global.nc.u32 %r317, [%rd64+28];

dp4a.s32.s32 %r315, %r316, %r317, %r309;

mov.u32 %r319, %r613;

prmt.b32 %r319, %r319, 0, 0xba98;

xor.b32 %r1025, %r319, %r761;
or.b32 %r1026, %r1025, -2139062144;
and.b32 %r1027, %r319, 2139062143;
sub.s32 %r1028, %r1026, %r1027;
xor.b32 %r322, %r787, %r1028;
ld.global.nc.u32 %r323, [%rd64+32];

dp4a.s32.s32 %r321, %r322, %r323, %r315;

mul.lo.s32 %r1029, %r321, %r790;
shr.s32 %r1030, %r1029, 31;
shr.u32 %r1031, %r1030, 29;
add.s32 %r1032, %r1029, %r1031;
shr.s32 %r1033, %r1032, 3;

{ cvt.f32.f16 %f81, %rs1;}


ld.global.nc.u32 %r325, [%rd64];

{.reg .f16 low,high;
mov.b32 {low,high},%r325;
cvt.f32.f16 %f82, low;}


mul.ftz.f32 %f105, %f81, %f82;
cvt.rn.f32.s32 %f106, %r1033;
fma.rn.ftz.f32 %f272, %f105, %f106, %f272;
mov.u32 %r326, %r620;

prmt.b32 %r326, %r326, 0, 0xba98;

xor.b32 %r1034, %r799, %r326;
or.b32 %r1035, %r1034, -2139062144;
and.b32 %r1036, %r326, 2139062143;
sub.s32 %r1037, %r1035, %r1036;
xor.b32 %r329, %r817, %r1037;

dp4a.s32.s32 %r328, %r329, %r281, %r625;

mov.u32 %r332, %r626;

prmt.b32 %r332, %r332, 0, 0xba98;

xor.b32 %r1038, %r332, %r800;
or.b32 %r1039, %r1038, -2139062144;
and.b32 %r1040, %r332, 2139062143;
sub.s32 %r1041, %r1039, %r1040;
xor.b32 %r335, %r825, %r1041;

dp4a.s32.s32 %r334, %r335, %r287, %r328;

mov.u32 %r338, %r632;

prmt.b32 %r338, %r338, 0, 0xba98;

xor.b32 %r1042, %r828, %r338;
or.b32 %r1043, %r1042, -2139062144;
and.b32 %r1044, %r338, 2139062143;
sub.s32 %r1045, %r1043, %r1044;
xor.b32 %r341, %r847, %r1045;

dp4a.s32.s32 %r340, %r341, %r293, %r334;

mov.u32 %r344, %r638;

prmt.b32 %r344, %r344, 0, 0xba98;

xor.b32 %r1046, %r344, %r829;
or.b32 %r1047, %r1046, -2139062144;
and.b32 %r1048, %r344, 2139062143;
sub.s32 %r1049, %r1047, %r1048;
xor.b32 %r347, %r855, %r1049;

dp4a.s32.s32 %r346, %r347, %r299, %r340;

mov.u32 %r350, %r644;

prmt.b32 %r350, %r350, 0, 0xba98;

xor.b32 %r1050, %r857, %r350;
or.b32 %r1051, %r1050, -2139062144;
and.b32 %r1052, %r350, 2139062143;
sub.s32 %r1053, %r1051, %r1052;
xor.b32 %r353, %r876, %r1053;

dp4a.s32.s32 %r352, %r353, %r305, %r346;

mov.u32 %r356, %r650;

prmt.b32 %r356, %r356, 0, 0xba98;

xor.b32 %r1054, %r356, %r858;
or.b32 %r1055, %r1054, -2139062144;
and.b32 %r1056, %r356, 2139062143;
sub.s32 %r1057, %r1055, %r1056;
xor.b32 %r359, %r884, %r1057;

dp4a.s32.s32 %r358, %r359, %r311, %r352;

mov.u32 %r362, %r656;

prmt.b32 %r362, %r362, 0, 0xba98;

xor.b32 %r1058, %r887, %r362;
or.b32 %r1059, %r1058, -2139062144;
and.b32 %r1060, %r362, 2139062143;
sub.s32 %r1061, %r1059, %r1060;
xor.b32 %r365, %r906, %r1061;

dp4a.s32.s32 %r364, %r365, %r317, %r358;

mov.u32 %r368, %r662;

prmt.b32 %r368, %r368, 0, 0xba98;

xor.b32 %r1062, %r368, %r888;
or.b32 %r1063, %r1062, -2139062144;
and.b32 %r1064, %r368, 2139062143;
sub.s32 %r1065, %r1063, %r1064;
xor.b32 %r371, %r914, %r1065;

dp4a.s32.s32 %r370, %r371, %r323, %r364;

mul.lo.s32 %r1066, %r370, %r917;
shr.s32 %r1067, %r1066, 31;
shr.u32 %r1068, %r1067, 29;
add.s32 %r1069, %r1066, %r1068;
shr.s32 %r1070, %r1069, 3;

{ cvt.f32.f16 %f83, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r325;
cvt.f32.f16 %f84, low;}


mul.ftz.f32 %f107, %f83, %f84;
cvt.rn.f32.s32 %f108, %r1070;
fma.rn.ftz.f32 %f271, %f107, %f108, %f271;
cvt.u64.u32 %rd65, %r1633;
add.s64 %rd66, %rd1, %rd65;
add.s64 %rd67, %rd66, %rd2;
mul.lo.s64 %rd68, %rd67, 36;
mov.u32 %r375, %r571;

prmt.b32 %r375, %r375, 0, 0xba98;

xor.b32 %r1071, %r672, %r375;
or.b32 %r1072, %r1071, -2139062144;
and.b32 %r1073, %r375, 2139062143;
sub.s32 %r1074, %r1072, %r1073;
xor.b32 %r378, %r690, %r1074;
add.s64 %rd69, %rd7, %rd68;
ld.global.nc.u32 %r379, [%rd69+4];

dp4a.s32.s32 %r377, %r378, %r379, %r625;

mov.u32 %r381, %r577;

prmt.b32 %r381, %r381, 0, 0xba98;

xor.b32 %r1075, %r381, %r673;
or.b32 %r1076, %r1075, -2139062144;
and.b32 %r1077, %r381, 2139062143;
sub.s32 %r1078, %r1076, %r1077;
xor.b32 %r384, %r698, %r1078;
ld.global.nc.u32 %r385, [%rd69+8];

dp4a.s32.s32 %r383, %r384, %r385, %r377;

mov.u32 %r387, %r583;

prmt.b32 %r387, %r387, 0, 0xba98;

xor.b32 %r1079, %r701, %r387;
or.b32 %r1080, %r1079, -2139062144;
and.b32 %r1081, %r387, 2139062143;
sub.s32 %r1082, %r1080, %r1081;
xor.b32 %r390, %r720, %r1082;
ld.global.nc.u32 %r391, [%rd69+12];

dp4a.s32.s32 %r389, %r390, %r391, %r383;

mov.u32 %r393, %r589;

prmt.b32 %r393, %r393, 0, 0xba98;

xor.b32 %r1083, %r393, %r702;
or.b32 %r1084, %r1083, -2139062144;
and.b32 %r1085, %r393, 2139062143;
sub.s32 %r1086, %r1084, %r1085;
xor.b32 %r396, %r728, %r1086;
ld.global.nc.u32 %r397, [%rd69+16];

dp4a.s32.s32 %r395, %r396, %r397, %r389;

mov.u32 %r399, %r595;

prmt.b32 %r399, %r399, 0, 0xba98;

xor.b32 %r1087, %r730, %r399;
or.b32 %r1088, %r1087, -2139062144;
and.b32 %r1089, %r399, 2139062143;
sub.s32 %r1090, %r1088, %r1089;
xor.b32 %r402, %r749, %r1090;
ld.global.nc.u32 %r403, [%rd69+20];

dp4a.s32.s32 %r401, %r402, %r403, %r395;

mov.u32 %r405, %r601;

prmt.b32 %r405, %r405, 0, 0xba98;

xor.b32 %r1091, %r405, %r731;
or.b32 %r1092, %r1091, -2139062144;
and.b32 %r1093, %r405, 2139062143;
sub.s32 %r1094, %r1092, %r1093;
xor.b32 %r408, %r757, %r1094;
ld.global.nc.u32 %r409, [%rd69+24];

dp4a.s32.s32 %r407, %r408, %r409, %r401;

mov.u32 %r411, %r607;

prmt.b32 %r411, %r411, 0, 0xba98;

xor.b32 %r1095, %r760, %r411;
or.b32 %r1096, %r1095, -2139062144;
and.b32 %r1097, %r411, 2139062143;
sub.s32 %r1098, %r1096, %r1097;
xor.b32 %r414, %r779, %r1098;
ld.global.nc.u32 %r415, [%rd69+28];

dp4a.s32.s32 %r413, %r414, %r415, %r407;

mov.u32 %r417, %r613;

prmt.b32 %r417, %r417, 0, 0xba98;

xor.b32 %r1099, %r417, %r761;
or.b32 %r1100, %r1099, -2139062144;
and.b32 %r1101, %r417, 2139062143;
sub.s32 %r1102, %r1100, %r1101;
xor.b32 %r420, %r787, %r1102;
ld.global.nc.u32 %r421, [%rd69+32];

dp4a.s32.s32 %r419, %r420, %r421, %r413;

mul.lo.s32 %r1103, %r419, %r790;
shr.s32 %r1104, %r1103, 31;
shr.u32 %r1105, %r1104, 29;
add.s32 %r1106, %r1103, %r1105;
shr.s32 %r1107, %r1106, 3;

{ cvt.f32.f16 %f85, %rs1;}


ld.global.nc.u32 %r423, [%rd69];

{.reg .f16 low,high;
mov.b32 {low,high},%r423;
cvt.f32.f16 %f86, low;}


mul.ftz.f32 %f109, %f85, %f86;
cvt.rn.f32.s32 %f110, %r1107;
fma.rn.ftz.f32 %f270, %f109, %f110, %f270;
mov.u32 %r424, %r620;

prmt.b32 %r424, %r424, 0, 0xba98;

xor.b32 %r1108, %r799, %r424;
or.b32 %r1109, %r1108, -2139062144;
and.b32 %r1110, %r424, 2139062143;
sub.s32 %r1111, %r1109, %r1110;
xor.b32 %r427, %r817, %r1111;

dp4a.s32.s32 %r426, %r427, %r379, %r625;

mov.u32 %r430, %r626;

prmt.b32 %r430, %r430, 0, 0xba98;

xor.b32 %r1112, %r430, %r800;
or.b32 %r1113, %r1112, -2139062144;
and.b32 %r1114, %r430, 2139062143;
sub.s32 %r1115, %r1113, %r1114;
xor.b32 %r433, %r825, %r1115;

dp4a.s32.s32 %r432, %r433, %r385, %r426;

mov.u32 %r436, %r632;

prmt.b32 %r436, %r436, 0, 0xba98;

xor.b32 %r1116, %r828, %r436;
or.b32 %r1117, %r1116, -2139062144;
and.b32 %r1118, %r436, 2139062143;
sub.s32 %r1119, %r1117, %r1118;
xor.b32 %r439, %r847, %r1119;

dp4a.s32.s32 %r438, %r439, %r391, %r432;

mov.u32 %r442, %r638;

prmt.b32 %r442, %r442, 0, 0xba98;

xor.b32 %r1120, %r442, %r829;
or.b32 %r1121, %r1120, -2139062144;
and.b32 %r1122, %r442, 2139062143;
sub.s32 %r1123, %r1121, %r1122;
xor.b32 %r445, %r855, %r1123;

dp4a.s32.s32 %r444, %r445, %r397, %r438;

mov.u32 %r448, %r644;

prmt.b32 %r448, %r448, 0, 0xba98;

xor.b32 %r1124, %r857, %r448;
or.b32 %r1125, %r1124, -2139062144;
and.b32 %r1126, %r448, 2139062143;
sub.s32 %r1127, %r1125, %r1126;
xor.b32 %r451, %r876, %r1127;

dp4a.s32.s32 %r450, %r451, %r403, %r444;

mov.u32 %r454, %r650;

prmt.b32 %r454, %r454, 0, 0xba98;

xor.b32 %r1128, %r454, %r858;
or.b32 %r1129, %r1128, -2139062144;
and.b32 %r1130, %r454, 2139062143;
sub.s32 %r1131, %r1129, %r1130;
xor.b32 %r457, %r884, %r1131;

dp4a.s32.s32 %r456, %r457, %r409, %r450;

mov.u32 %r460, %r656;

prmt.b32 %r460, %r460, 0, 0xba98;

xor.b32 %r1132, %r887, %r460;
or.b32 %r1133, %r1132, -2139062144;
and.b32 %r1134, %r460, 2139062143;
sub.s32 %r1135, %r1133, %r1134;
xor.b32 %r463, %r906, %r1135;

dp4a.s32.s32 %r462, %r463, %r415, %r456;

mov.u32 %r466, %r662;

prmt.b32 %r466, %r466, 0, 0xba98;

xor.b32 %r1136, %r466, %r888;
or.b32 %r1137, %r1136, -2139062144;
and.b32 %r1138, %r466, 2139062143;
sub.s32 %r1139, %r1137, %r1138;
xor.b32 %r469, %r914, %r1139;

dp4a.s32.s32 %r468, %r469, %r421, %r462;

mul.lo.s32 %r1140, %r468, %r917;
shr.s32 %r1141, %r1140, 31;
shr.u32 %r1142, %r1141, 29;
add.s32 %r1143, %r1140, %r1142;
shr.s32 %r1144, %r1143, 3;

{ cvt.f32.f16 %f87, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r423;
cvt.f32.f16 %f88, low;}


mul.ftz.f32 %f111, %f87, %f88;
cvt.rn.f32.s32 %f112, %r1144;
fma.rn.ftz.f32 %f269, %f111, %f112, %f269;
cvt.u64.u32 %rd70, %r1632;
add.s64 %rd71, %rd1, %rd70;
add.s64 %rd72, %rd71, %rd2;
mul.lo.s64 %rd73, %rd72, 36;
mov.u32 %r473, %r571;

prmt.b32 %r473, %r473, 0, 0xba98;

xor.b32 %r1145, %r672, %r473;
or.b32 %r1146, %r1145, -2139062144;
and.b32 %r1147, %r473, 2139062143;
sub.s32 %r1148, %r1146, %r1147;
xor.b32 %r476, %r690, %r1148;
add.s64 %rd74, %rd7, %rd73;
ld.global.nc.u32 %r477, [%rd74+4];

dp4a.s32.s32 %r475, %r476, %r477, %r625;

mov.u32 %r479, %r577;

prmt.b32 %r479, %r479, 0, 0xba98;

xor.b32 %r1149, %r479, %r673;
or.b32 %r1150, %r1149, -2139062144;
and.b32 %r1151, %r479, 2139062143;
sub.s32 %r1152, %r1150, %r1151;
xor.b32 %r482, %r698, %r1152;
ld.global.nc.u32 %r483, [%rd74+8];

dp4a.s32.s32 %r481, %r482, %r483, %r475;

mov.u32 %r485, %r583;

prmt.b32 %r485, %r485, 0, 0xba98;

xor.b32 %r1153, %r701, %r485;
or.b32 %r1154, %r1153, -2139062144;
and.b32 %r1155, %r485, 2139062143;
sub.s32 %r1156, %r1154, %r1155;
xor.b32 %r488, %r720, %r1156;
ld.global.nc.u32 %r489, [%rd74+12];

dp4a.s32.s32 %r487, %r488, %r489, %r481;

mov.u32 %r491, %r589;

prmt.b32 %r491, %r491, 0, 0xba98;

xor.b32 %r1157, %r491, %r702;
or.b32 %r1158, %r1157, -2139062144;
and.b32 %r1159, %r491, 2139062143;
sub.s32 %r1160, %r1158, %r1159;
xor.b32 %r494, %r728, %r1160;
ld.global.nc.u32 %r495, [%rd74+16];

dp4a.s32.s32 %r493, %r494, %r495, %r487;

mov.u32 %r497, %r595;

prmt.b32 %r497, %r497, 0, 0xba98;

xor.b32 %r1161, %r730, %r497;
or.b32 %r1162, %r1161, -2139062144;
and.b32 %r1163, %r497, 2139062143;
sub.s32 %r1164, %r1162, %r1163;
xor.b32 %r500, %r749, %r1164;
ld.global.nc.u32 %r501, [%rd74+20];

dp4a.s32.s32 %r499, %r500, %r501, %r493;

mov.u32 %r503, %r601;

prmt.b32 %r503, %r503, 0, 0xba98;

xor.b32 %r1165, %r503, %r731;
or.b32 %r1166, %r1165, -2139062144;
and.b32 %r1167, %r503, 2139062143;
sub.s32 %r1168, %r1166, %r1167;
xor.b32 %r506, %r757, %r1168;
ld.global.nc.u32 %r507, [%rd74+24];

dp4a.s32.s32 %r505, %r506, %r507, %r499;

mov.u32 %r509, %r607;

prmt.b32 %r509, %r509, 0, 0xba98;

xor.b32 %r1169, %r760, %r509;
or.b32 %r1170, %r1169, -2139062144;
and.b32 %r1171, %r509, 2139062143;
sub.s32 %r1172, %r1170, %r1171;
xor.b32 %r512, %r779, %r1172;
ld.global.nc.u32 %r513, [%rd74+28];

dp4a.s32.s32 %r511, %r512, %r513, %r505;

mov.u32 %r515, %r613;

prmt.b32 %r515, %r515, 0, 0xba98;

xor.b32 %r1173, %r515, %r761;
or.b32 %r1174, %r1173, -2139062144;
and.b32 %r1175, %r515, 2139062143;
sub.s32 %r1176, %r1174, %r1175;
xor.b32 %r518, %r787, %r1176;
ld.global.nc.u32 %r519, [%rd74+32];

dp4a.s32.s32 %r517, %r518, %r519, %r511;

mul.lo.s32 %r1177, %r517, %r790;
shr.s32 %r1178, %r1177, 31;
shr.u32 %r1179, %r1178, 29;
add.s32 %r1180, %r1177, %r1179;
shr.s32 %r1181, %r1180, 3;

{ cvt.f32.f16 %f89, %rs1;}


ld.global.nc.u32 %r521, [%rd74];

{.reg .f16 low,high;
mov.b32 {low,high},%r521;
cvt.f32.f16 %f90, low;}


mul.ftz.f32 %f113, %f89, %f90;
cvt.rn.f32.s32 %f114, %r1181;
fma.rn.ftz.f32 %f268, %f113, %f114, %f268;
mov.u32 %r522, %r620;

prmt.b32 %r522, %r522, 0, 0xba98;

xor.b32 %r1182, %r799, %r522;
or.b32 %r1183, %r1182, -2139062144;
and.b32 %r1184, %r522, 2139062143;
sub.s32 %r1185, %r1183, %r1184;
xor.b32 %r525, %r817, %r1185;

dp4a.s32.s32 %r524, %r525, %r477, %r625;

mov.u32 %r528, %r626;

prmt.b32 %r528, %r528, 0, 0xba98;

xor.b32 %r1186, %r528, %r800;
or.b32 %r1187, %r1186, -2139062144;
and.b32 %r1188, %r528, 2139062143;
sub.s32 %r1189, %r1187, %r1188;
xor.b32 %r531, %r825, %r1189;

dp4a.s32.s32 %r530, %r531, %r483, %r524;

mov.u32 %r534, %r632;

prmt.b32 %r534, %r534, 0, 0xba98;

xor.b32 %r1190, %r828, %r534;
or.b32 %r1191, %r1190, -2139062144;
and.b32 %r1192, %r534, 2139062143;
sub.s32 %r1193, %r1191, %r1192;
xor.b32 %r537, %r847, %r1193;

dp4a.s32.s32 %r536, %r537, %r489, %r530;

mov.u32 %r540, %r638;

prmt.b32 %r540, %r540, 0, 0xba98;

xor.b32 %r1194, %r540, %r829;
or.b32 %r1195, %r1194, -2139062144;
and.b32 %r1196, %r540, 2139062143;
sub.s32 %r1197, %r1195, %r1196;
xor.b32 %r543, %r855, %r1197;

dp4a.s32.s32 %r542, %r543, %r495, %r536;

mov.u32 %r546, %r644;

prmt.b32 %r546, %r546, 0, 0xba98;

xor.b32 %r1198, %r857, %r546;
or.b32 %r1199, %r1198, -2139062144;
and.b32 %r1200, %r546, 2139062143;
sub.s32 %r1201, %r1199, %r1200;
xor.b32 %r549, %r876, %r1201;

dp4a.s32.s32 %r548, %r549, %r501, %r542;

mov.u32 %r552, %r650;

prmt.b32 %r552, %r552, 0, 0xba98;

xor.b32 %r1202, %r552, %r858;
or.b32 %r1203, %r1202, -2139062144;
and.b32 %r1204, %r552, 2139062143;
sub.s32 %r1205, %r1203, %r1204;
xor.b32 %r555, %r884, %r1205;

dp4a.s32.s32 %r554, %r555, %r507, %r548;

mov.u32 %r558, %r656;

prmt.b32 %r558, %r558, 0, 0xba98;

xor.b32 %r1206, %r887, %r558;
or.b32 %r1207, %r1206, -2139062144;
and.b32 %r1208, %r558, 2139062143;
sub.s32 %r1209, %r1207, %r1208;
xor.b32 %r561, %r906, %r1209;

dp4a.s32.s32 %r560, %r561, %r513, %r554;

mov.u32 %r564, %r662;

prmt.b32 %r564, %r564, 0, 0xba98;

xor.b32 %r1210, %r564, %r888;
or.b32 %r1211, %r1210, -2139062144;
and.b32 %r1212, %r564, 2139062143;
sub.s32 %r1213, %r1211, %r1212;
xor.b32 %r567, %r914, %r1213;

dp4a.s32.s32 %r566, %r567, %r519, %r560;

mul.lo.s32 %r1214, %r566, %r917;
shr.s32 %r1215, %r1214, 31;
shr.u32 %r1216, %r1215, 29;
add.s32 %r1217, %r1214, %r1216;
shr.s32 %r1218, %r1217, 3;

{ cvt.f32.f16 %f91, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r521;
cvt.f32.f16 %f92, low;}


mul.ftz.f32 %f115, %f91, %f92;
cvt.rn.f32.s32 %f116, %r1218;
fma.rn.ftz.f32 %f267, %f115, %f116, %f267;
cvt.u64.u32 %rd75, %r1631;
add.s64 %rd76, %rd1, %rd75;
add.s64 %rd77, %rd76, %rd2;
mul.lo.s64 %rd78, %rd77, 36;

prmt.b32 %r571, %r571, 0, 0xba98;

xor.b32 %r1219, %r672, %r571;
or.b32 %r1220, %r1219, -2139062144;
and.b32 %r1221, %r571, 2139062143;
sub.s32 %r1222, %r1220, %r1221;
xor.b32 %r574, %r690, %r1222;
add.s64 %rd79, %rd7, %rd78;
ld.global.nc.u32 %r575, [%rd79+4];

dp4a.s32.s32 %r573, %r574, %r575, %r625;


prmt.b32 %r577, %r577, 0, 0xba98;

xor.b32 %r1223, %r577, %r673;
or.b32 %r1224, %r1223, -2139062144;
and.b32 %r1225, %r577, 2139062143;
sub.s32 %r1226, %r1224, %r1225;
xor.b32 %r580, %r698, %r1226;
ld.global.nc.u32 %r581, [%rd79+8];

dp4a.s32.s32 %r579, %r580, %r581, %r573;


prmt.b32 %r583, %r583, 0, 0xba98;

xor.b32 %r1227, %r701, %r583;
or.b32 %r1228, %r1227, -2139062144;
and.b32 %r1229, %r583, 2139062143;
sub.s32 %r1230, %r1228, %r1229;
xor.b32 %r586, %r720, %r1230;
ld.global.nc.u32 %r587, [%rd79+12];

dp4a.s32.s32 %r585, %r586, %r587, %r579;


prmt.b32 %r589, %r589, 0, 0xba98;

xor.b32 %r1231, %r589, %r702;
or.b32 %r1232, %r1231, -2139062144;
and.b32 %r1233, %r589, 2139062143;
sub.s32 %r1234, %r1232, %r1233;
xor.b32 %r592, %r728, %r1234;
ld.global.nc.u32 %r593, [%rd79+16];

dp4a.s32.s32 %r591, %r592, %r593, %r585;


prmt.b32 %r595, %r595, 0, 0xba98;

xor.b32 %r1235, %r730, %r595;
or.b32 %r1236, %r1235, -2139062144;
and.b32 %r1237, %r595, 2139062143;
sub.s32 %r1238, %r1236, %r1237;
xor.b32 %r598, %r749, %r1238;
ld.global.nc.u32 %r599, [%rd79+20];

dp4a.s32.s32 %r597, %r598, %r599, %r591;


prmt.b32 %r601, %r601, 0, 0xba98;

xor.b32 %r1239, %r601, %r731;
or.b32 %r1240, %r1239, -2139062144;
and.b32 %r1241, %r601, 2139062143;
sub.s32 %r1242, %r1240, %r1241;
xor.b32 %r604, %r757, %r1242;
ld.global.nc.u32 %r605, [%rd79+24];

dp4a.s32.s32 %r603, %r604, %r605, %r597;


prmt.b32 %r607, %r607, 0, 0xba98;

xor.b32 %r1243, %r760, %r607;
or.b32 %r1244, %r1243, -2139062144;
and.b32 %r1245, %r607, 2139062143;
sub.s32 %r1246, %r1244, %r1245;
xor.b32 %r610, %r779, %r1246;
ld.global.nc.u32 %r611, [%rd79+28];

dp4a.s32.s32 %r609, %r610, %r611, %r603;


prmt.b32 %r613, %r613, 0, 0xba98;

xor.b32 %r1247, %r613, %r761;
or.b32 %r1248, %r1247, -2139062144;
and.b32 %r1249, %r613, 2139062143;
sub.s32 %r1250, %r1248, %r1249;
xor.b32 %r616, %r787, %r1250;
ld.global.nc.u32 %r617, [%rd79+32];

dp4a.s32.s32 %r615, %r616, %r617, %r609;

mul.lo.s32 %r1251, %r615, %r790;
shr.s32 %r1252, %r1251, 31;
shr.u32 %r1253, %r1252, 29;
add.s32 %r1254, %r1251, %r1253;
shr.s32 %r1255, %r1254, 3;

{ cvt.f32.f16 %f93, %rs1;}


ld.global.nc.u32 %r619, [%rd79];

{.reg .f16 low,high;
mov.b32 {low,high},%r619;
cvt.f32.f16 %f94, low;}


mul.ftz.f32 %f117, %f93, %f94;
cvt.rn.f32.s32 %f118, %r1255;
fma.rn.ftz.f32 %f266, %f117, %f118, %f266;

prmt.b32 %r620, %r620, 0, 0xba98;

xor.b32 %r1256, %r799, %r620;
or.b32 %r1257, %r1256, -2139062144;
and.b32 %r1258, %r620, 2139062143;
sub.s32 %r1259, %r1257, %r1258;
xor.b32 %r623, %r817, %r1259;

dp4a.s32.s32 %r622, %r623, %r575, %r625;


prmt.b32 %r626, %r626, 0, 0xba98;

xor.b32 %r1260, %r626, %r800;
or.b32 %r1261, %r1260, -2139062144;
and.b32 %r1262, %r626, 2139062143;
sub.s32 %r1263, %r1261, %r1262;
xor.b32 %r629, %r825, %r1263;

dp4a.s32.s32 %r628, %r629, %r581, %r622;


prmt.b32 %r632, %r632, 0, 0xba98;

xor.b32 %r1264, %r828, %r632;
or.b32 %r1265, %r1264, -2139062144;
and.b32 %r1266, %r632, 2139062143;
sub.s32 %r1267, %r1265, %r1266;
xor.b32 %r635, %r847, %r1267;

dp4a.s32.s32 %r634, %r635, %r587, %r628;


prmt.b32 %r638, %r638, 0, 0xba98;

xor.b32 %r1268, %r638, %r829;
or.b32 %r1269, %r1268, -2139062144;
and.b32 %r1270, %r638, 2139062143;
sub.s32 %r1271, %r1269, %r1270;
xor.b32 %r641, %r855, %r1271;

dp4a.s32.s32 %r640, %r641, %r593, %r634;


prmt.b32 %r644, %r644, 0, 0xba98;

xor.b32 %r1272, %r857, %r644;
or.b32 %r1273, %r1272, -2139062144;
and.b32 %r1274, %r644, 2139062143;
sub.s32 %r1275, %r1273, %r1274;
xor.b32 %r647, %r876, %r1275;

dp4a.s32.s32 %r646, %r647, %r599, %r640;


prmt.b32 %r650, %r650, 0, 0xba98;

xor.b32 %r1276, %r650, %r858;
or.b32 %r1277, %r1276, -2139062144;
and.b32 %r1278, %r650, 2139062143;
sub.s32 %r1279, %r1277, %r1278;
xor.b32 %r653, %r884, %r1279;

dp4a.s32.s32 %r652, %r653, %r605, %r646;


prmt.b32 %r656, %r656, 0, 0xba98;

xor.b32 %r1280, %r887, %r656;
or.b32 %r1281, %r1280, -2139062144;
and.b32 %r1282, %r656, 2139062143;
sub.s32 %r1283, %r1281, %r1282;
xor.b32 %r659, %r906, %r1283;

dp4a.s32.s32 %r658, %r659, %r611, %r652;


prmt.b32 %r662, %r662, 0, 0xba98;

xor.b32 %r1284, %r662, %r888;
or.b32 %r1285, %r1284, -2139062144;
and.b32 %r1286, %r662, 2139062143;
sub.s32 %r1287, %r1285, %r1286;
xor.b32 %r665, %r914, %r1287;

dp4a.s32.s32 %r664, %r665, %r617, %r658;

mul.lo.s32 %r1288, %r664, %r917;
shr.s32 %r1289, %r1288, 31;
shr.u32 %r1290, %r1289, 29;
add.s32 %r1291, %r1288, %r1290;
shr.s32 %r1292, %r1291, 3;

{ cvt.f32.f16 %f95, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r619;
cvt.f32.f16 %f96, low;}


mul.ftz.f32 %f119, %f95, %f96;
cvt.rn.f32.s32 %f120, %r1292;
fma.rn.ftz.f32 %f265, %f119, %f120, %f265;
add.s32 %r1634, %r1634, 64;
add.s32 %r1633, %r1633, 64;
add.s32 %r1632, %r1632, 64;
add.s32 %r1631, %r1631, 64;
add.s32 %r1630, %r1630, 64;
add.s32 %r1629, %r1629, 64;
add.s64 %rd134, %rd134, 528;
add.s64 %rd133, %rd133, 528;
add.s64 %rd132, %rd132, 528;
add.s32 %r1635, %r1635, 8;
setp.lt.s32 %p3, %r1635, %r2;
@%p3 bra $L__BB208_2;

$L__BB208_3:
setp.eq.s32 %p4, %r41, 0;
@%p4 bra $L__BB208_5;

mov.u32 %r1296, _ZZ13mul_mat_vec_qIL9ggml_type16ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r1297, %r41, 1536, %r1296;
shl.b32 %r1298, %r43, 2;
add.s32 %r1299, %r1297, %r1298;
st.shared.f32 [%r1299+-1536], %f276;
st.shared.f32 [%r1299+-1408], %f275;
st.shared.f32 [%r1299+-1280], %f274;
st.shared.f32 [%r1299+-1152], %f273;
st.shared.f32 [%r1299+-1024], %f272;
st.shared.f32 [%r1299+-896], %f271;
st.shared.f32 [%r1299+-768], %f270;
st.shared.f32 [%r1299+-640], %f269;
st.shared.f32 [%r1299+-512], %f268;
st.shared.f32 [%r1299+-384], %f267;
st.shared.f32 [%r1299+-256], %f266;
st.shared.f32 [%r1299+-128], %f265;

$L__BB208_5:
setp.ne.s32 %p5, %r41, 0;
bar.sync 0;
@%p5 bra $L__BB208_30;

mov.u32 %r1301, %ctaid.x;
shl.b32 %r1302, %r1301, 1;
mov.u32 %r1303, 1;
shl.b32 %r1305, %r43, 2;
mov.u32 %r1306, _ZZ13mul_mat_vec_qIL9ggml_type16ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r1307, %r1306, %r1305;
mov.u32 %r1308, 2;
ld.shared.f32 %f121, [%r1307];
add.ftz.f32 %f122, %f121, %f276;
mov.b32 %r1309, %f122;
mov.u32 %r1310, 31;
mov.u32 %r1311, 16;
mov.u32 %r1312, -1;
shfl.sync.bfly.b32 %r1313|%p6, %r1309, %r1311, %r1310, %r1312;
mov.b32 %f123, %r1313;
add.ftz.f32 %f124, %f122, %f123;
mov.b32 %r1314, %f124;
mov.u32 %r1315, 8;
shfl.sync.bfly.b32 %r1316|%p7, %r1314, %r1315, %r1310, %r1312;
mov.b32 %f125, %r1316;
add.ftz.f32 %f126, %f124, %f125;
mov.b32 %r1317, %f126;
mov.u32 %r1318, 4;
shfl.sync.bfly.b32 %r1319|%p8, %r1317, %r1318, %r1310, %r1312;
mov.b32 %f127, %r1319;
add.ftz.f32 %f128, %f126, %f127;
mov.b32 %r1320, %f128;
shfl.sync.bfly.b32 %r1321|%p9, %r1320, %r1308, %r1310, %r1312;
mov.b32 %f129, %r1321;
add.ftz.f32 %f130, %f128, %f129;
mov.b32 %r1322, %f130;
shfl.sync.bfly.b32 %r1323|%p10, %r1322, %r1303, %r1310, %r1312;
mov.b32 %f131, %r1323;
add.ftz.f32 %f37, %f130, %f131;
setp.ne.s32 %p11, %r43, 0;
setp.ge.u32 %p12, %r1302, %r27;
or.pred %p13, %p11, %p12;
@%p13 bra $L__BB208_8;

cvta.to.global.u64 %rd80, %rd18;
mov.u32 %r1324, %ctaid.y;
mad.lo.s32 %r1327, %r1324, %r33, %r1302;
mov.u32 %r1328, %ctaid.z;
mad.lo.s32 %r1329, %r1328, %r39, %r1327;
mul.wide.u32 %rd81, %r1329, 4;
add.s64 %rd82, %rd80, %rd81;
st.global.f32 [%rd82], %f37;

$L__BB208_8:
ld.shared.f32 %f132, [%r1307+128];
add.ftz.f32 %f133, %f132, %f275;
mov.b32 %r1335, %f133;
shfl.sync.bfly.b32 %r1339|%p14, %r1335, %r1311, %r1310, %r1312;
mov.b32 %f134, %r1339;
add.ftz.f32 %f135, %f133, %f134;
mov.b32 %r1340, %f135;
shfl.sync.bfly.b32 %r1342|%p15, %r1340, %r1315, %r1310, %r1312;
mov.b32 %f136, %r1342;
add.ftz.f32 %f137, %f135, %f136;
mov.b32 %r1343, %f137;
shfl.sync.bfly.b32 %r1345|%p16, %r1343, %r1318, %r1310, %r1312;
mov.b32 %f138, %r1345;
add.ftz.f32 %f139, %f137, %f138;
mov.b32 %r1346, %f139;
shfl.sync.bfly.b32 %r1347|%p17, %r1346, %r1308, %r1310, %r1312;
mov.b32 %f140, %r1347;
add.ftz.f32 %f141, %f139, %f140;
mov.b32 %r1348, %f141;
shfl.sync.bfly.b32 %r1350|%p18, %r1348, %r1303, %r1310, %r1312;
mov.b32 %f142, %r1350;
add.ftz.f32 %f38, %f141, %f142;
setp.eq.s32 %p19, %r43, 1;
or.b32 %r1353, %r1302, 1;
setp.lt.u32 %p20, %r1353, %r27;
and.pred %p1, %p19, %p20;
not.pred %p21, %p1;
@%p21 bra $L__BB208_10;

cvta.to.global.u64 %rd83, %rd18;
mov.u32 %r1354, %ctaid.y;
mad.lo.s32 %r1357, %r1354, %r33, %r1302;
mov.u32 %r1358, %ctaid.z;
mad.lo.s32 %r1359, %r1358, %r39, %r1357;
mul.wide.u32 %rd84, %r1359, 4;
add.s64 %rd85, %rd83, %rd84;
st.global.f32 [%rd85+4], %f38;

$L__BB208_10:
mov.u32 %r1364, 2;
ld.shared.f32 %f143, [%r1307+256];
add.ftz.f32 %f144, %f143, %f274;
mov.b32 %r1365, %f144;
mov.u32 %r1366, 31;
mov.u32 %r1367, 16;
mov.u32 %r1368, -1;
shfl.sync.bfly.b32 %r1369|%p22, %r1365, %r1367, %r1366, %r1368;
mov.b32 %f145, %r1369;
add.ftz.f32 %f146, %f144, %f145;
mov.b32 %r1370, %f146;
mov.u32 %r1371, 8;
shfl.sync.bfly.b32 %r1372|%p23, %r1370, %r1371, %r1366, %r1368;
mov.b32 %f147, %r1372;
add.ftz.f32 %f148, %f146, %f147;
mov.b32 %r1373, %f148;
mov.u32 %r1374, 4;
shfl.sync.bfly.b32 %r1375|%p24, %r1373, %r1374, %r1366, %r1368;
mov.b32 %f149, %r1375;
add.ftz.f32 %f150, %f148, %f149;
mov.b32 %r1376, %f150;
shfl.sync.bfly.b32 %r1377|%p25, %r1376, %r1364, %r1366, %r1368;
mov.b32 %f151, %r1377;
add.ftz.f32 %f152, %f150, %f151;
mov.b32 %r1378, %f152;
mov.u32 %r1379, 1;
shfl.sync.bfly.b32 %r1380|%p26, %r1378, %r1379, %r1366, %r1368;
mov.b32 %f153, %r1380;
add.ftz.f32 %f39, %f152, %f153;
@%p13 bra $L__BB208_12;

cvt.u64.u32 %rd86, %r27;
mov.u32 %r1383, %ctaid.y;
mad.lo.s32 %r1386, %r1383, %r33, %r1302;
mov.u32 %r1387, %ctaid.z;
mad.lo.s32 %r1388, %r1387, %r39, %r1386;
cvt.u64.u32 %rd87, %r1388;
add.s64 %rd88, %rd86, %rd87;
cvta.to.global.u64 %rd89, %rd18;
shl.b64 %rd90, %rd88, 2;
add.s64 %rd91, %rd89, %rd90;
st.global.f32 [%rd91], %f39;

$L__BB208_12:
ld.shared.f32 %f154, [%r1307+384];
add.ftz.f32 %f155, %f154, %f273;
mov.b32 %r1394, %f155;
shfl.sync.bfly.b32 %r1398|%p30, %r1394, %r1367, %r1366, %r1368;
mov.b32 %f156, %r1398;
add.ftz.f32 %f157, %f155, %f156;
mov.b32 %r1399, %f157;
shfl.sync.bfly.b32 %r1401|%p31, %r1399, %r1371, %r1366, %r1368;
mov.b32 %f158, %r1401;
add.ftz.f32 %f159, %f157, %f158;
mov.b32 %r1402, %f159;
shfl.sync.bfly.b32 %r1404|%p32, %r1402, %r1374, %r1366, %r1368;
mov.b32 %f160, %r1404;
add.ftz.f32 %f161, %f159, %f160;
mov.b32 %r1405, %f161;
shfl.sync.bfly.b32 %r1406|%p33, %r1405, %r1364, %r1366, %r1368;
mov.b32 %f162, %r1406;
add.ftz.f32 %f163, %f161, %f162;
mov.b32 %r1407, %f163;
shfl.sync.bfly.b32 %r1409|%p34, %r1407, %r1379, %r1366, %r1368;
mov.b32 %f164, %r1409;
add.ftz.f32 %f40, %f163, %f164;
@%p21 bra $L__BB208_14;

add.s32 %r1410, %r27, 1;
cvt.u64.u32 %rd92, %r1410;
mov.u32 %r1411, %ctaid.y;
mad.lo.s32 %r1414, %r1411, %r33, %r1302;
mov.u32 %r1415, %ctaid.z;
mad.lo.s32 %r1416, %r1415, %r39, %r1414;
cvt.u64.u32 %rd93, %r1416;
add.s64 %rd94, %rd92, %rd93;
cvta.to.global.u64 %rd95, %rd18;
shl.b64 %rd96, %rd94, 2;
add.s64 %rd97, %rd95, %rd96;
st.global.f32 [%rd97], %f40;

$L__BB208_14:
shl.b32 %r1417, %r27, 1;
mov.u32 %r1418, 1;
mov.u32 %r1423, 2;
ld.shared.f32 %f165, [%r1307+512];
add.ftz.f32 %f166, %f165, %f272;
mov.b32 %r1424, %f166;
mov.u32 %r1425, 31;
mov.u32 %r1426, 16;
mov.u32 %r1427, -1;
shfl.sync.bfly.b32 %r1428|%p36, %r1424, %r1426, %r1425, %r1427;
mov.b32 %f167, %r1428;
add.ftz.f32 %f168, %f166, %f167;
mov.b32 %r1429, %f168;
mov.u32 %r1430, 8;
shfl.sync.bfly.b32 %r1431|%p37, %r1429, %r1430, %r1425, %r1427;
mov.b32 %f169, %r1431;
add.ftz.f32 %f170, %f168, %f169;
mov.b32 %r1432, %f170;
mov.u32 %r1433, 4;
shfl.sync.bfly.b32 %r1434|%p38, %r1432, %r1433, %r1425, %r1427;
mov.b32 %f171, %r1434;
add.ftz.f32 %f172, %f170, %f171;
mov.b32 %r1435, %f172;
shfl.sync.bfly.b32 %r1436|%p39, %r1435, %r1423, %r1425, %r1427;
mov.b32 %f173, %r1436;
add.ftz.f32 %f174, %f172, %f173;
mov.b32 %r1437, %f174;
shfl.sync.bfly.b32 %r1438|%p40, %r1437, %r1418, %r1425, %r1427;
mov.b32 %f175, %r1438;
add.ftz.f32 %f41, %f174, %f175;
cvt.u64.u32 %rd98, %r1417;
mov.u32 %r1439, %ctaid.y;
mad.lo.s32 %r1442, %r1439, %r33, %r1302;
mov.u32 %r1443, %ctaid.z;
mad.lo.s32 %r1444, %r1443, %r39, %r1442;
cvt.u64.u32 %rd99, %r1444;
add.s64 %rd100, %rd98, %rd99;
cvta.to.global.u64 %rd101, %rd18;
shl.b64 %rd102, %rd100, 2;
add.s64 %rd14, %rd101, %rd102;
@%p13 bra $L__BB208_16;

st.global.f32 [%rd14], %f41;

$L__BB208_16:
ld.shared.f32 %f176, [%r1307+640];
add.ftz.f32 %f177, %f176, %f271;
mov.b32 %r1450, %f177;
shfl.sync.bfly.b32 %r1454|%p44, %r1450, %r1426, %r1425, %r1427;
mov.b32 %f178, %r1454;
add.ftz.f32 %f179, %f177, %f178;
mov.b32 %r1455, %f179;
shfl.sync.bfly.b32 %r1457|%p45, %r1455, %r1430, %r1425, %r1427;
mov.b32 %f180, %r1457;
add.ftz.f32 %f181, %f179, %f180;
mov.b32 %r1458, %f181;
shfl.sync.bfly.b32 %r1460|%p46, %r1458, %r1433, %r1425, %r1427;
mov.b32 %f182, %r1460;
add.ftz.f32 %f183, %f181, %f182;
mov.b32 %r1461, %f183;
shfl.sync.bfly.b32 %r1462|%p47, %r1461, %r1423, %r1425, %r1427;
mov.b32 %f184, %r1462;
add.ftz.f32 %f185, %f183, %f184;
mov.b32 %r1463, %f185;
shfl.sync.bfly.b32 %r1465|%p48, %r1463, %r1418, %r1425, %r1427;
mov.b32 %f186, %r1465;
add.ftz.f32 %f42, %f185, %f186;
@%p21 bra $L__BB208_18;

st.global.f32 [%rd14+4], %f42;

$L__BB208_18:
mul.lo.s32 %r23, %r27, 3;
mov.u32 %r1470, 2;
ld.shared.f32 %f187, [%r1307+768];
add.ftz.f32 %f188, %f187, %f270;
mov.b32 %r1471, %f188;
mov.u32 %r1472, 31;
mov.u32 %r1473, 16;
mov.u32 %r1474, -1;
shfl.sync.bfly.b32 %r1475|%p50, %r1471, %r1473, %r1472, %r1474;
mov.b32 %f189, %r1475;
add.ftz.f32 %f190, %f188, %f189;
mov.b32 %r1476, %f190;
mov.u32 %r1477, 8;
shfl.sync.bfly.b32 %r1478|%p51, %r1476, %r1477, %r1472, %r1474;
mov.b32 %f191, %r1478;
add.ftz.f32 %f192, %f190, %f191;
mov.b32 %r1479, %f192;
mov.u32 %r1480, 4;
shfl.sync.bfly.b32 %r1481|%p52, %r1479, %r1480, %r1472, %r1474;
mov.b32 %f193, %r1481;
add.ftz.f32 %f194, %f192, %f193;
mov.b32 %r1482, %f194;
shfl.sync.bfly.b32 %r1483|%p53, %r1482, %r1470, %r1472, %r1474;
mov.b32 %f195, %r1483;
add.ftz.f32 %f196, %f194, %f195;
mov.b32 %r1484, %f196;
mov.u32 %r1485, 1;
shfl.sync.bfly.b32 %r1486|%p54, %r1484, %r1485, %r1472, %r1474;
mov.b32 %f197, %r1486;
add.ftz.f32 %f43, %f196, %f197;
@%p13 bra $L__BB208_20;

cvt.u64.u32 %rd103, %r23;
add.s64 %rd105, %rd103, %rd99;
shl.b64 %rd107, %rd105, 2;
add.s64 %rd108, %rd101, %rd107;
st.global.f32 [%rd108], %f43;

$L__BB208_20:
ld.shared.f32 %f198, [%r1307+896];
add.ftz.f32 %f199, %f198, %f269;
mov.b32 %r1500, %f199;
shfl.sync.bfly.b32 %r1504|%p58, %r1500, %r1473, %r1472, %r1474;
mov.b32 %f200, %r1504;
add.ftz.f32 %f201, %f199, %f200;
mov.b32 %r1505, %f201;
shfl.sync.bfly.b32 %r1507|%p59, %r1505, %r1477, %r1472, %r1474;
mov.b32 %f202, %r1507;
add.ftz.f32 %f203, %f201, %f202;
mov.b32 %r1508, %f203;
shfl.sync.bfly.b32 %r1510|%p60, %r1508, %r1480, %r1472, %r1474;
mov.b32 %f204, %r1510;
add.ftz.f32 %f205, %f203, %f204;
mov.b32 %r1511, %f205;
shfl.sync.bfly.b32 %r1512|%p61, %r1511, %r1470, %r1472, %r1474;
mov.b32 %f206, %r1512;
add.ftz.f32 %f207, %f205, %f206;
mov.b32 %r1513, %f207;
shfl.sync.bfly.b32 %r1515|%p62, %r1513, %r1485, %r1472, %r1474;
mov.b32 %f208, %r1515;
add.ftz.f32 %f44, %f207, %f208;
@%p21 bra $L__BB208_22;

add.s32 %r1516, %r23, 1;
cvt.u64.u32 %rd109, %r1516;
add.s64 %rd111, %rd109, %rd99;
shl.b64 %rd113, %rd111, 2;
add.s64 %rd114, %rd101, %rd113;
st.global.f32 [%rd114], %f44;

$L__BB208_22:
shl.b32 %r1523, %r27, 2;
mov.u32 %r1524, 2;
ld.shared.f32 %f209, [%r1307+1024];
add.ftz.f32 %f210, %f209, %f268;
mov.b32 %r1529, %f210;
mov.u32 %r1530, 31;
mov.u32 %r1531, 16;
mov.u32 %r1532, -1;
shfl.sync.bfly.b32 %r1533|%p64, %r1529, %r1531, %r1530, %r1532;
mov.b32 %f211, %r1533;
add.ftz.f32 %f212, %f210, %f211;
mov.b32 %r1534, %f212;
mov.u32 %r1535, 8;
shfl.sync.bfly.b32 %r1536|%p65, %r1534, %r1535, %r1530, %r1532;
mov.b32 %f213, %r1536;
add.ftz.f32 %f214, %f212, %f213;
mov.b32 %r1537, %f214;
mov.u32 %r1538, 4;
shfl.sync.bfly.b32 %r1539|%p66, %r1537, %r1538, %r1530, %r1532;
mov.b32 %f215, %r1539;
add.ftz.f32 %f216, %f214, %f215;
mov.b32 %r1540, %f216;
shfl.sync.bfly.b32 %r1541|%p67, %r1540, %r1524, %r1530, %r1532;
mov.b32 %f217, %r1541;
add.ftz.f32 %f218, %f216, %f217;
mov.b32 %r1542, %f218;
mov.u32 %r1543, 1;
shfl.sync.bfly.b32 %r1544|%p68, %r1542, %r1543, %r1530, %r1532;
mov.b32 %f219, %r1544;
add.ftz.f32 %f45, %f218, %f219;
cvt.u64.u32 %rd115, %r1523;
add.s64 %rd117, %rd115, %rd99;
shl.b64 %rd119, %rd117, 2;
add.s64 %rd15, %rd101, %rd119;
@%p13 bra $L__BB208_24;

st.global.f32 [%rd15], %f45;

$L__BB208_24:
ld.shared.f32 %f220, [%r1307+1152];
add.ftz.f32 %f221, %f220, %f267;
mov.b32 %r1556, %f221;
shfl.sync.bfly.b32 %r1560|%p72, %r1556, %r1531, %r1530, %r1532;
mov.b32 %f222, %r1560;
add.ftz.f32 %f223, %f221, %f222;
mov.b32 %r1561, %f223;
shfl.sync.bfly.b32 %r1563|%p73, %r1561, %r1535, %r1530, %r1532;
mov.b32 %f224, %r1563;
add.ftz.f32 %f225, %f223, %f224;
mov.b32 %r1564, %f225;
shfl.sync.bfly.b32 %r1566|%p74, %r1564, %r1538, %r1530, %r1532;
mov.b32 %f226, %r1566;
add.ftz.f32 %f227, %f225, %f226;
mov.b32 %r1567, %f227;
shfl.sync.bfly.b32 %r1568|%p75, %r1567, %r1524, %r1530, %r1532;
mov.b32 %f228, %r1568;
add.ftz.f32 %f229, %f227, %f228;
mov.b32 %r1569, %f229;
shfl.sync.bfly.b32 %r1571|%p76, %r1569, %r1543, %r1530, %r1532;
mov.b32 %f230, %r1571;
add.ftz.f32 %f46, %f229, %f230;
@%p21 bra $L__BB208_26;

st.global.f32 [%rd15+4], %f46;

$L__BB208_26:
mul.lo.s32 %r24, %r27, 5;
mov.u32 %r1576, 2;
ld.shared.f32 %f231, [%r1307+1280];
add.ftz.f32 %f232, %f231, %f266;
mov.b32 %r1577, %f232;
mov.u32 %r1578, 31;
mov.u32 %r1579, 16;
mov.u32 %r1580, -1;
shfl.sync.bfly.b32 %r1581|%p78, %r1577, %r1579, %r1578, %r1580;
mov.b32 %f233, %r1581;
add.ftz.f32 %f234, %f232, %f233;
mov.b32 %r1582, %f234;
mov.u32 %r1583, 8;
shfl.sync.bfly.b32 %r1584|%p79, %r1582, %r1583, %r1578, %r1580;
mov.b32 %f235, %r1584;
add.ftz.f32 %f236, %f234, %f235;
mov.b32 %r1585, %f236;
mov.u32 %r1586, 4;
shfl.sync.bfly.b32 %r1587|%p80, %r1585, %r1586, %r1578, %r1580;
mov.b32 %f237, %r1587;
add.ftz.f32 %f238, %f236, %f237;
mov.b32 %r1588, %f238;
shfl.sync.bfly.b32 %r1589|%p81, %r1588, %r1576, %r1578, %r1580;
mov.b32 %f239, %r1589;
add.ftz.f32 %f240, %f238, %f239;
mov.b32 %r1590, %f240;
mov.u32 %r1591, 1;
shfl.sync.bfly.b32 %r1592|%p82, %r1590, %r1591, %r1578, %r1580;
mov.b32 %f241, %r1592;
add.ftz.f32 %f47, %f240, %f241;
@%p13 bra $L__BB208_28;

cvt.u64.u32 %rd120, %r24;
add.s64 %rd122, %rd120, %rd99;
shl.b64 %rd124, %rd122, 2;
add.s64 %rd125, %rd101, %rd124;
st.global.f32 [%rd125], %f47;

$L__BB208_28:
ld.shared.f32 %f242, [%r1307+1408];
add.ftz.f32 %f243, %f242, %f265;
mov.b32 %r1606, %f243;
shfl.sync.bfly.b32 %r1610|%p86, %r1606, %r1579, %r1578, %r1580;
mov.b32 %f244, %r1610;
add.ftz.f32 %f245, %f243, %f244;
mov.b32 %r1611, %f245;
shfl.sync.bfly.b32 %r1613|%p87, %r1611, %r1583, %r1578, %r1580;
mov.b32 %f246, %r1613;
add.ftz.f32 %f247, %f245, %f246;
mov.b32 %r1614, %f247;
shfl.sync.bfly.b32 %r1616|%p88, %r1614, %r1586, %r1578, %r1580;
mov.b32 %f248, %r1616;
add.ftz.f32 %f249, %f247, %f248;
mov.b32 %r1617, %f249;
shfl.sync.bfly.b32 %r1618|%p89, %r1617, %r1576, %r1578, %r1580;
mov.b32 %f250, %r1618;
add.ftz.f32 %f251, %f249, %f250;
mov.b32 %r1619, %f251;
shfl.sync.bfly.b32 %r1621|%p90, %r1619, %r1591, %r1578, %r1580;
mov.b32 %f252, %r1621;
add.ftz.f32 %f48, %f251, %f252;
@%p21 bra $L__BB208_30;

add.s32 %r1622, %r24, 1;
cvt.u64.u32 %rd126, %r1622;
add.s64 %rd128, %rd126, %rd99;
shl.b64 %rd130, %rd128, 2;
add.s64 %rd131, %rd101, %rd130;
st.global.f32 [%rd131], %f48;

$L__BB208_30:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type16ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type16ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type16ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type16ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type16ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type16ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type16ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type16ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type16ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<106>;
.reg .b16 %rs<31>;
.reg .f32 %f<323>;
.reg .b32 %r<1856>;
.reg .b64 %rd<146>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type16ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1792];

ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type16ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type16ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type16ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r45, [_Z13mul_mat_vec_qIL9ggml_type16ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type16ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type16ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type16ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type16ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type16ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type16ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type16ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type16ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type16ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type16ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type16ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type16ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type16ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r46, %r1, 5;
mov.u32 %r47, %tid.x;
add.s32 %r2, %r46, %r47;
shr.s32 %r48, %r2, 31;
shr.u32 %r49, %r48, 29;
add.s32 %r50, %r2, %r49;
shr.s32 %r1855, %r50, 3;
shr.u32 %r4, %r45, 8;
setp.ge.s32 %p2, %r1855, %r4;
mov.f32 %f309, 0f00000000;
mov.f32 %f310, %f309;
mov.f32 %f311, %f309;
mov.f32 %f312, %f309;
mov.f32 %f313, %f309;
mov.f32 %f314, %f309;
mov.f32 %f315, %f309;
mov.f32 %f316, %f309;
mov.f32 %f317, %f309;
mov.f32 %f318, %f309;
mov.f32 %f319, %f309;
mov.f32 %f320, %f309;
mov.f32 %f321, %f309;
mov.f32 %f322, %f309;
@%p2 bra $L__BB209_3;

mov.u32 %r51, %ctaid.x;
shl.b32 %r52, %r51, 1;
mov.u32 %r53, %ctaid.y;
mov.u32 %r54, %ctaid.z;
mul.lo.s32 %r55, %r53, %r37;
cvt.u64.u32 %rd20, %r55;
mul.lo.s32 %r56, %r54, %r43;
cvt.u64.u32 %rd21, %r56;
add.s64 %rd1, %rd21, %rd20;
mul.hi.u32 %r57, %r54, %r39;
add.s32 %r58, %r57, %r54;
shr.u32 %r59, %r58, %r40;
mul.lo.s32 %r60, %r59, %r42;
mul.hi.u32 %r61, %r53, %r33;
add.s32 %r62, %r61, %r53;
shr.u32 %r63, %r62, %r34;
mul.lo.s32 %r64, %r63, %r36;
cvta.to.global.u64 %rd22, %rd17;
and.b32 %r68, %r50, -8;
sub.s32 %r69, %r2, %r68;
shl.b32 %r70, %r69, 2;
cvt.s64.s32 %rd2, %r69;
shl.b32 %r71, %r31, 1;
shl.b32 %r1848, %r1855, 3;
add.s32 %r1854, %r1848, %r71;
mad.lo.s32 %r1853, %r31, 3, %r1848;
shl.b32 %r72, %r31, 2;
add.s32 %r1852, %r1848, %r72;
mad.lo.s32 %r1851, %r31, 5, %r1848;
mad.lo.s32 %r1850, %r31, 6, %r1848;
add.s32 %r1849, %r31, %r1848;
add.s32 %r73, %r1855, %r60;
add.s32 %r74, %r73, %r64;
or.b32 %r75, %r52, 1;
mad.lo.s32 %r76, %r30, %r75, %r74;
mul.wide.s32 %rd23, %r76, 66;
add.s64 %rd145, %rd22, %rd23;
mad.lo.s32 %r77, %r52, %r30, %r1855;
add.s32 %r78, %r77, %r60;
add.s32 %r79, %r78, %r64;
mul.wide.s32 %rd24, %r79, 66;
add.s64 %rd25, %rd22, %rd24;
add.s64 %rd144, %rd25, 4;
mul.wide.s32 %rd26, %r70, 2;
add.s64 %rd27, %rd24, %rd26;
add.s64 %rd28, %rd22, %rd27;
add.s64 %rd143, %rd28, 8;
or.b64 %rd6, %rd26, 4;
cvta.to.global.u64 %rd7, %rd18;
mov.f32 %f309, 0f00000000;

$L__BB209_2:
cvt.u64.u32 %rd29, %r1848;
add.s64 %rd30, %rd1, %rd29;
add.s64 %rd31, %rd30, %rd2;
mul.lo.s64 %rd32, %rd31, 36;
ld.global.nc.u16 %rs15, [%rd143+-6];
shr.u16 %rs16, %rs15, 8;
ld.global.nc.u16 %rs17, [%rd143+-4];
shr.u16 %rs18, %rs17, 8;
ld.global.nc.u16 %rs19, [%rd143+-2];
cvt.u32.u16 %r766, %rs19;
ld.global.nc.u16 %rs20, [%rd143];
cvt.u32.u16 %r767, %rs20;
mov.b32 %r768, {%rs19, %rs20};
shl.b16 %rs21, %rs15, 3;
cvt.u64.u16 %rd33, %rs21;
and.b64 %rd34, %rd33, 2040;
mov.u64 %rd35, iq2xxs_grid;
add.s64 %rd36, %rd35, %rd34;
ld.global.nc.v2.u32 {%r769, %r770}, [%rd36];
not.b32 %r772, %r769;
mov.u32 %r722, 0;
not.b32 %r774, %r770;
and.b32 %r775, %r766, 255;
popc.b32 %r776, %r775;
shl.b32 %r777, %r776, 7;
and.b32 %r778, %r777, 128;
xor.b32 %r779, %r778, %r775;
mul.lo.s32 %r780, %r779, 16843009;
and.b32 %r781, %r780, 134480385;
or.b32 %r782, %r781, -2139062144;
add.s32 %r783, %r782, -16843009;
or.b32 %r668, %r783, %r781;
mov.u32 %r80, %r668;

prmt.b32 %r80, %r80, 0, 0xba98;

xor.b32 %r784, %r769, %r80;
or.b32 %r785, %r784, -2139062144;
and.b32 %r786, %r80, 2139062143;
and.b32 %r787, %r772, -2139062144;
sub.s32 %r788, %r785, %r786;
xor.b32 %r83, %r787, %r788;
add.s64 %rd37, %rd7, %rd32;
ld.global.nc.u32 %r84, [%rd37+4];

dp4a.s32.s32 %r82, %r83, %r84, %r722;

and.b32 %r789, %r780, -2143281136;
or.b32 %r790, %r789, -2139062144;
add.s32 %r791, %r790, -16843009;
or.b32 %r674, %r791, %r789;
mov.u32 %r86, %r674;

prmt.b32 %r86, %r86, 0, 0xba98;

xor.b32 %r792, %r86, %r770;
or.b32 %r793, %r792, -2139062144;
and.b32 %r794, %r86, 2139062143;
and.b32 %r795, %r774, -2139062144;
sub.s32 %r796, %r793, %r794;
xor.b32 %r89, %r795, %r796;
ld.global.nc.u32 %r90, [%rd37+8];

dp4a.s32.s32 %r88, %r89, %r90, %r82;

cvt.u32.u16 %r797, %rs16;
mul.wide.u32 %rd38, %r797, 8;
add.s64 %rd39, %rd35, %rd38;
ld.global.nc.v2.u32 {%r798, %r799}, [%rd39];
not.b32 %r801, %r798;
not.b32 %r803, %r799;
shr.u32 %r804, %r766, 7;
and.b32 %r805, %r804, 255;
popc.b32 %r806, %r805;
shl.b32 %r807, %r806, 7;
and.b32 %r808, %r807, 128;
xor.b32 %r809, %r808, %r805;
mul.lo.s32 %r810, %r809, 16843009;
and.b32 %r811, %r810, 134480385;
or.b32 %r812, %r811, -2139062144;
add.s32 %r813, %r812, -16843009;
or.b32 %r680, %r813, %r811;
mov.u32 %r92, %r680;

prmt.b32 %r92, %r92, 0, 0xba98;

xor.b32 %r814, %r798, %r92;
or.b32 %r815, %r814, -2139062144;
and.b32 %r816, %r92, 2139062143;
and.b32 %r817, %r801, -2139062144;
sub.s32 %r818, %r815, %r816;
xor.b32 %r95, %r817, %r818;
ld.global.nc.u32 %r96, [%rd37+12];

dp4a.s32.s32 %r94, %r95, %r96, %r88;

and.b32 %r819, %r810, -2143281136;
or.b32 %r820, %r819, -2139062144;
add.s32 %r821, %r820, -16843009;
or.b32 %r686, %r821, %r819;
mov.u32 %r98, %r686;

prmt.b32 %r98, %r98, 0, 0xba98;

xor.b32 %r822, %r98, %r799;
or.b32 %r823, %r822, -2139062144;
and.b32 %r824, %r98, 2139062143;
and.b32 %r825, %r803, -2139062144;
sub.s32 %r826, %r823, %r824;
xor.b32 %r101, %r825, %r826;
ld.global.nc.u32 %r102, [%rd37+16];

dp4a.s32.s32 %r100, %r101, %r102, %r94;

shl.b16 %rs22, %rs17, 3;
cvt.u64.u16 %rd40, %rs22;
and.b64 %rd41, %rd40, 2040;
add.s64 %rd42, %rd35, %rd41;
ld.global.nc.v2.u32 {%r827, %r828}, [%rd42];
not.b32 %r830, %r827;
not.b32 %r832, %r828;
shr.u32 %r833, %r768, 14;
and.b32 %r834, %r833, 255;
popc.b32 %r835, %r834;
shl.b32 %r836, %r835, 7;
and.b32 %r837, %r836, 128;
xor.b32 %r838, %r837, %r834;
mul.lo.s32 %r839, %r838, 16843009;
and.b32 %r840, %r839, 134480385;
or.b32 %r841, %r840, -2139062144;
add.s32 %r842, %r841, -16843009;
or.b32 %r692, %r842, %r840;
mov.u32 %r104, %r692;

prmt.b32 %r104, %r104, 0, 0xba98;

xor.b32 %r843, %r827, %r104;
or.b32 %r844, %r843, -2139062144;
and.b32 %r845, %r104, 2139062143;
and.b32 %r846, %r830, -2139062144;
sub.s32 %r847, %r844, %r845;
xor.b32 %r107, %r846, %r847;
ld.global.nc.u32 %r108, [%rd37+20];

dp4a.s32.s32 %r106, %r107, %r108, %r100;

and.b32 %r848, %r839, -2143281136;
or.b32 %r849, %r848, -2139062144;
add.s32 %r850, %r849, -16843009;
or.b32 %r698, %r850, %r848;
mov.u32 %r110, %r698;

prmt.b32 %r110, %r110, 0, 0xba98;

xor.b32 %r851, %r110, %r828;
or.b32 %r852, %r851, -2139062144;
and.b32 %r853, %r110, 2139062143;
and.b32 %r854, %r832, -2139062144;
sub.s32 %r855, %r852, %r853;
xor.b32 %r113, %r854, %r855;
ld.global.nc.u32 %r114, [%rd37+24];

dp4a.s32.s32 %r112, %r113, %r114, %r106;

cvt.u32.u16 %r856, %rs18;
mul.wide.u32 %rd43, %r856, 8;
add.s64 %rd44, %rd35, %rd43;
ld.global.nc.v2.u32 {%r857, %r858}, [%rd44];
not.b32 %r860, %r857;
not.b32 %r862, %r858;
shr.u32 %r863, %r767, 5;
and.b32 %r864, %r863, 255;
popc.b32 %r865, %r864;
shl.b32 %r866, %r865, 7;
and.b32 %r867, %r866, 128;
xor.b32 %r868, %r867, %r864;
mul.lo.s32 %r869, %r868, 16843009;
and.b32 %r870, %r869, 134480385;
or.b32 %r871, %r870, -2139062144;
add.s32 %r872, %r871, -16843009;
or.b32 %r704, %r872, %r870;
mov.u32 %r116, %r704;

prmt.b32 %r116, %r116, 0, 0xba98;

xor.b32 %r873, %r857, %r116;
or.b32 %r874, %r873, -2139062144;
and.b32 %r875, %r116, 2139062143;
and.b32 %r876, %r860, -2139062144;
sub.s32 %r877, %r874, %r875;
xor.b32 %r119, %r876, %r877;
ld.global.nc.u32 %r120, [%rd37+28];

dp4a.s32.s32 %r118, %r119, %r120, %r112;

and.b32 %r878, %r869, -2143281136;
or.b32 %r879, %r878, -2139062144;
add.s32 %r880, %r879, -16843009;
or.b32 %r710, %r880, %r878;
mov.u32 %r122, %r710;

prmt.b32 %r122, %r122, 0, 0xba98;

xor.b32 %r881, %r122, %r858;
or.b32 %r882, %r881, -2139062144;
and.b32 %r883, %r122, 2139062143;
and.b32 %r884, %r862, -2139062144;
sub.s32 %r885, %r882, %r883;
xor.b32 %r125, %r884, %r885;
ld.global.nc.u32 %r126, [%rd37+32];

dp4a.s32.s32 %r124, %r125, %r126, %r118;

shr.u32 %r886, %r767, 11;
or.b32 %r887, %r886, 1;
mul.lo.s32 %r888, %r124, %r887;
shr.s32 %r889, %r888, 31;
shr.u32 %r890, %r889, 29;
add.s32 %r891, %r888, %r890;
shr.s32 %r892, %r891, 3;
ld.global.nc.u16 %rs1, [%rd144+-4];

{ cvt.f32.f16 %f85, %rs1;}


ld.global.nc.u32 %r128, [%rd37];

{.reg .f16 low,high;
mov.b32 {low,high},%r128;
cvt.f32.f16 %f86, low;}


mul.ftz.f32 %f113, %f85, %f86;
cvt.rn.f32.s32 %f114, %r892;
fma.rn.ftz.f32 %f322, %f113, %f114, %f322;
add.s64 %rd45, %rd145, %rd6;
ld.global.nc.u16 %rs23, [%rd45+-2];
shr.u16 %rs24, %rs23, 8;
ld.global.nc.u16 %rs25, [%rd45];
shr.u16 %rs26, %rs25, 8;
ld.global.nc.u16 %rs27, [%rd45+2];
cvt.u32.u16 %r893, %rs27;
ld.global.nc.u16 %rs28, [%rd45+4];
cvt.u32.u16 %r894, %rs28;
mov.b32 %r895, {%rs27, %rs28};
shl.b16 %rs29, %rs23, 3;
cvt.u64.u16 %rd46, %rs29;
and.b64 %rd47, %rd46, 2040;
add.s64 %rd48, %rd35, %rd47;
ld.global.nc.v2.u32 {%r896, %r897}, [%rd48];
not.b32 %r899, %r896;
not.b32 %r901, %r897;
and.b32 %r902, %r893, 255;
popc.b32 %r903, %r902;
shl.b32 %r904, %r903, 7;
and.b32 %r905, %r904, 128;
xor.b32 %r906, %r905, %r902;
mul.lo.s32 %r907, %r906, 16843009;
and.b32 %r908, %r907, 134480385;
or.b32 %r909, %r908, -2139062144;
add.s32 %r910, %r909, -16843009;
or.b32 %r717, %r910, %r908;
mov.u32 %r129, %r717;

prmt.b32 %r129, %r129, 0, 0xba98;

xor.b32 %r911, %r896, %r129;
or.b32 %r912, %r911, -2139062144;
and.b32 %r913, %r129, 2139062143;
and.b32 %r914, %r899, -2139062144;
sub.s32 %r915, %r912, %r913;
xor.b32 %r132, %r914, %r915;

dp4a.s32.s32 %r131, %r132, %r84, %r722;

and.b32 %r916, %r907, -2143281136;
or.b32 %r917, %r916, -2139062144;
add.s32 %r918, %r917, -16843009;
or.b32 %r723, %r918, %r916;
mov.u32 %r135, %r723;

prmt.b32 %r135, %r135, 0, 0xba98;

xor.b32 %r919, %r135, %r897;
or.b32 %r920, %r919, -2139062144;
and.b32 %r921, %r135, 2139062143;
and.b32 %r922, %r901, -2139062144;
sub.s32 %r923, %r920, %r921;
xor.b32 %r138, %r922, %r923;

dp4a.s32.s32 %r137, %r138, %r90, %r131;

cvt.u32.u16 %r924, %rs24;
mul.wide.u32 %rd49, %r924, 8;
add.s64 %rd50, %rd35, %rd49;
ld.global.nc.v2.u32 {%r925, %r926}, [%rd50];
not.b32 %r928, %r925;
not.b32 %r930, %r926;
shr.u32 %r931, %r893, 7;
and.b32 %r932, %r931, 255;
popc.b32 %r933, %r932;
shl.b32 %r934, %r933, 7;
and.b32 %r935, %r934, 128;
xor.b32 %r936, %r935, %r932;
mul.lo.s32 %r937, %r936, 16843009;
and.b32 %r938, %r937, 134480385;
or.b32 %r939, %r938, -2139062144;
add.s32 %r940, %r939, -16843009;
or.b32 %r729, %r940, %r938;
mov.u32 %r141, %r729;

prmt.b32 %r141, %r141, 0, 0xba98;

xor.b32 %r941, %r925, %r141;
or.b32 %r942, %r941, -2139062144;
and.b32 %r943, %r141, 2139062143;
and.b32 %r944, %r928, -2139062144;
sub.s32 %r945, %r942, %r943;
xor.b32 %r144, %r944, %r945;

dp4a.s32.s32 %r143, %r144, %r96, %r137;

and.b32 %r946, %r937, -2143281136;
or.b32 %r947, %r946, -2139062144;
add.s32 %r948, %r947, -16843009;
or.b32 %r735, %r948, %r946;
mov.u32 %r147, %r735;

prmt.b32 %r147, %r147, 0, 0xba98;

xor.b32 %r949, %r147, %r926;
or.b32 %r950, %r949, -2139062144;
and.b32 %r951, %r147, 2139062143;
and.b32 %r952, %r930, -2139062144;
sub.s32 %r953, %r950, %r951;
xor.b32 %r150, %r952, %r953;

dp4a.s32.s32 %r149, %r150, %r102, %r143;

shl.b16 %rs30, %rs25, 3;
cvt.u64.u16 %rd51, %rs30;
and.b64 %rd52, %rd51, 2040;
add.s64 %rd53, %rd35, %rd52;
ld.global.nc.v2.u32 {%r954, %r955}, [%rd53];
not.b32 %r957, %r954;
not.b32 %r959, %r955;
shr.u32 %r960, %r895, 14;
and.b32 %r961, %r960, 255;
popc.b32 %r962, %r961;
shl.b32 %r963, %r962, 7;
and.b32 %r964, %r963, 128;
xor.b32 %r965, %r964, %r961;
mul.lo.s32 %r966, %r965, 16843009;
and.b32 %r967, %r966, 134480385;
or.b32 %r968, %r967, -2139062144;
add.s32 %r969, %r968, -16843009;
or.b32 %r741, %r969, %r967;
mov.u32 %r153, %r741;

prmt.b32 %r153, %r153, 0, 0xba98;

xor.b32 %r970, %r954, %r153;
or.b32 %r971, %r970, -2139062144;
and.b32 %r972, %r153, 2139062143;
and.b32 %r973, %r957, -2139062144;
sub.s32 %r974, %r971, %r972;
xor.b32 %r156, %r973, %r974;

dp4a.s32.s32 %r155, %r156, %r108, %r149;

and.b32 %r975, %r966, -2143281136;
or.b32 %r976, %r975, -2139062144;
add.s32 %r977, %r976, -16843009;
or.b32 %r747, %r977, %r975;
mov.u32 %r159, %r747;

prmt.b32 %r159, %r159, 0, 0xba98;

xor.b32 %r978, %r159, %r955;
or.b32 %r979, %r978, -2139062144;
and.b32 %r980, %r159, 2139062143;
and.b32 %r981, %r959, -2139062144;
sub.s32 %r982, %r979, %r980;
xor.b32 %r162, %r981, %r982;

dp4a.s32.s32 %r161, %r162, %r114, %r155;

cvt.u32.u16 %r983, %rs26;
mul.wide.u32 %rd54, %r983, 8;
add.s64 %rd55, %rd35, %rd54;
ld.global.nc.v2.u32 {%r984, %r985}, [%rd55];
not.b32 %r987, %r984;
not.b32 %r989, %r985;
shr.u32 %r990, %r894, 5;
and.b32 %r991, %r990, 255;
popc.b32 %r992, %r991;
shl.b32 %r993, %r992, 7;
and.b32 %r994, %r993, 128;
xor.b32 %r995, %r994, %r991;
mul.lo.s32 %r996, %r995, 16843009;
and.b32 %r997, %r996, 134480385;
or.b32 %r998, %r997, -2139062144;
add.s32 %r999, %r998, -16843009;
or.b32 %r753, %r999, %r997;
mov.u32 %r165, %r753;

prmt.b32 %r165, %r165, 0, 0xba98;

xor.b32 %r1000, %r984, %r165;
or.b32 %r1001, %r1000, -2139062144;
and.b32 %r1002, %r165, 2139062143;
and.b32 %r1003, %r987, -2139062144;
sub.s32 %r1004, %r1001, %r1002;
xor.b32 %r168, %r1003, %r1004;

dp4a.s32.s32 %r167, %r168, %r120, %r161;

and.b32 %r1005, %r996, -2143281136;
or.b32 %r1006, %r1005, -2139062144;
add.s32 %r1007, %r1006, -16843009;
or.b32 %r759, %r1007, %r1005;
mov.u32 %r171, %r759;

prmt.b32 %r171, %r171, 0, 0xba98;

xor.b32 %r1008, %r171, %r985;
or.b32 %r1009, %r1008, -2139062144;
and.b32 %r1010, %r171, 2139062143;
and.b32 %r1011, %r989, -2139062144;
sub.s32 %r1012, %r1009, %r1010;
xor.b32 %r174, %r1011, %r1012;

dp4a.s32.s32 %r173, %r174, %r126, %r167;

shr.u32 %r1013, %r894, 11;
or.b32 %r1014, %r1013, 1;
mul.lo.s32 %r1015, %r173, %r1014;
shr.s32 %r1016, %r1015, 31;
shr.u32 %r1017, %r1016, 29;
add.s32 %r1018, %r1015, %r1017;
shr.s32 %r1019, %r1018, 3;
ld.global.nc.u16 %rs2, [%rd145];

{ cvt.f32.f16 %f87, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r128;
cvt.f32.f16 %f88, low;}


mul.ftz.f32 %f115, %f87, %f88;
cvt.rn.f32.s32 %f116, %r1019;
fma.rn.ftz.f32 %f321, %f115, %f116, %f321;
cvt.u64.u32 %rd56, %r1849;
add.s64 %rd57, %rd1, %rd56;
add.s64 %rd58, %rd57, %rd2;
mul.lo.s64 %rd59, %rd58, 36;
mov.u32 %r178, %r668;

prmt.b32 %r178, %r178, 0, 0xba98;

xor.b32 %r1020, %r769, %r178;
or.b32 %r1021, %r1020, -2139062144;
and.b32 %r1022, %r178, 2139062143;
sub.s32 %r1023, %r1021, %r1022;
xor.b32 %r181, %r787, %r1023;
add.s64 %rd60, %rd7, %rd59;
ld.global.nc.u32 %r182, [%rd60+4];

dp4a.s32.s32 %r180, %r181, %r182, %r722;

mov.u32 %r184, %r674;

prmt.b32 %r184, %r184, 0, 0xba98;

xor.b32 %r1024, %r184, %r770;
or.b32 %r1025, %r1024, -2139062144;
and.b32 %r1026, %r184, 2139062143;
sub.s32 %r1027, %r1025, %r1026;
xor.b32 %r187, %r795, %r1027;
ld.global.nc.u32 %r188, [%rd60+8];

dp4a.s32.s32 %r186, %r187, %r188, %r180;

mov.u32 %r190, %r680;

prmt.b32 %r190, %r190, 0, 0xba98;

xor.b32 %r1028, %r798, %r190;
or.b32 %r1029, %r1028, -2139062144;
and.b32 %r1030, %r190, 2139062143;
sub.s32 %r1031, %r1029, %r1030;
xor.b32 %r193, %r817, %r1031;
ld.global.nc.u32 %r194, [%rd60+12];

dp4a.s32.s32 %r192, %r193, %r194, %r186;

mov.u32 %r196, %r686;

prmt.b32 %r196, %r196, 0, 0xba98;

xor.b32 %r1032, %r196, %r799;
or.b32 %r1033, %r1032, -2139062144;
and.b32 %r1034, %r196, 2139062143;
sub.s32 %r1035, %r1033, %r1034;
xor.b32 %r199, %r825, %r1035;
ld.global.nc.u32 %r200, [%rd60+16];

dp4a.s32.s32 %r198, %r199, %r200, %r192;

mov.u32 %r202, %r692;

prmt.b32 %r202, %r202, 0, 0xba98;

xor.b32 %r1036, %r827, %r202;
or.b32 %r1037, %r1036, -2139062144;
and.b32 %r1038, %r202, 2139062143;
sub.s32 %r1039, %r1037, %r1038;
xor.b32 %r205, %r846, %r1039;
ld.global.nc.u32 %r206, [%rd60+20];

dp4a.s32.s32 %r204, %r205, %r206, %r198;

mov.u32 %r208, %r698;

prmt.b32 %r208, %r208, 0, 0xba98;

xor.b32 %r1040, %r208, %r828;
or.b32 %r1041, %r1040, -2139062144;
and.b32 %r1042, %r208, 2139062143;
sub.s32 %r1043, %r1041, %r1042;
xor.b32 %r211, %r854, %r1043;
ld.global.nc.u32 %r212, [%rd60+24];

dp4a.s32.s32 %r210, %r211, %r212, %r204;

mov.u32 %r214, %r704;

prmt.b32 %r214, %r214, 0, 0xba98;

xor.b32 %r1044, %r857, %r214;
or.b32 %r1045, %r1044, -2139062144;
and.b32 %r1046, %r214, 2139062143;
sub.s32 %r1047, %r1045, %r1046;
xor.b32 %r217, %r876, %r1047;
ld.global.nc.u32 %r218, [%rd60+28];

dp4a.s32.s32 %r216, %r217, %r218, %r210;

mov.u32 %r220, %r710;

prmt.b32 %r220, %r220, 0, 0xba98;

xor.b32 %r1048, %r220, %r858;
or.b32 %r1049, %r1048, -2139062144;
and.b32 %r1050, %r220, 2139062143;
sub.s32 %r1051, %r1049, %r1050;
xor.b32 %r223, %r884, %r1051;
ld.global.nc.u32 %r224, [%rd60+32];

dp4a.s32.s32 %r222, %r223, %r224, %r216;

mul.lo.s32 %r1052, %r222, %r887;
shr.s32 %r1053, %r1052, 31;
shr.u32 %r1054, %r1053, 29;
add.s32 %r1055, %r1052, %r1054;
shr.s32 %r1056, %r1055, 3;

{ cvt.f32.f16 %f89, %rs1;}


ld.global.nc.u32 %r226, [%rd60];

{.reg .f16 low,high;
mov.b32 {low,high},%r226;
cvt.f32.f16 %f90, low;}


mul.ftz.f32 %f117, %f89, %f90;
cvt.rn.f32.s32 %f118, %r1056;
fma.rn.ftz.f32 %f320, %f117, %f118, %f320;
mov.u32 %r227, %r717;

prmt.b32 %r227, %r227, 0, 0xba98;

xor.b32 %r1057, %r896, %r227;
or.b32 %r1058, %r1057, -2139062144;
and.b32 %r1059, %r227, 2139062143;
sub.s32 %r1060, %r1058, %r1059;
xor.b32 %r230, %r914, %r1060;

dp4a.s32.s32 %r229, %r230, %r182, %r722;

mov.u32 %r233, %r723;

prmt.b32 %r233, %r233, 0, 0xba98;

xor.b32 %r1061, %r233, %r897;
or.b32 %r1062, %r1061, -2139062144;
and.b32 %r1063, %r233, 2139062143;
sub.s32 %r1064, %r1062, %r1063;
xor.b32 %r236, %r922, %r1064;

dp4a.s32.s32 %r235, %r236, %r188, %r229;

mov.u32 %r239, %r729;

prmt.b32 %r239, %r239, 0, 0xba98;

xor.b32 %r1065, %r925, %r239;
or.b32 %r1066, %r1065, -2139062144;
and.b32 %r1067, %r239, 2139062143;
sub.s32 %r1068, %r1066, %r1067;
xor.b32 %r242, %r944, %r1068;

dp4a.s32.s32 %r241, %r242, %r194, %r235;

mov.u32 %r245, %r735;

prmt.b32 %r245, %r245, 0, 0xba98;

xor.b32 %r1069, %r245, %r926;
or.b32 %r1070, %r1069, -2139062144;
and.b32 %r1071, %r245, 2139062143;
sub.s32 %r1072, %r1070, %r1071;
xor.b32 %r248, %r952, %r1072;

dp4a.s32.s32 %r247, %r248, %r200, %r241;

mov.u32 %r251, %r741;

prmt.b32 %r251, %r251, 0, 0xba98;

xor.b32 %r1073, %r954, %r251;
or.b32 %r1074, %r1073, -2139062144;
and.b32 %r1075, %r251, 2139062143;
sub.s32 %r1076, %r1074, %r1075;
xor.b32 %r254, %r973, %r1076;

dp4a.s32.s32 %r253, %r254, %r206, %r247;

mov.u32 %r257, %r747;

prmt.b32 %r257, %r257, 0, 0xba98;

xor.b32 %r1077, %r257, %r955;
or.b32 %r1078, %r1077, -2139062144;
and.b32 %r1079, %r257, 2139062143;
sub.s32 %r1080, %r1078, %r1079;
xor.b32 %r260, %r981, %r1080;

dp4a.s32.s32 %r259, %r260, %r212, %r253;

mov.u32 %r263, %r753;

prmt.b32 %r263, %r263, 0, 0xba98;

xor.b32 %r1081, %r984, %r263;
or.b32 %r1082, %r1081, -2139062144;
and.b32 %r1083, %r263, 2139062143;
sub.s32 %r1084, %r1082, %r1083;
xor.b32 %r266, %r1003, %r1084;

dp4a.s32.s32 %r265, %r266, %r218, %r259;

mov.u32 %r269, %r759;

prmt.b32 %r269, %r269, 0, 0xba98;

xor.b32 %r1085, %r269, %r985;
or.b32 %r1086, %r1085, -2139062144;
and.b32 %r1087, %r269, 2139062143;
sub.s32 %r1088, %r1086, %r1087;
xor.b32 %r272, %r1011, %r1088;

dp4a.s32.s32 %r271, %r272, %r224, %r265;

mul.lo.s32 %r1089, %r271, %r1014;
shr.s32 %r1090, %r1089, 31;
shr.u32 %r1091, %r1090, 29;
add.s32 %r1092, %r1089, %r1091;
shr.s32 %r1093, %r1092, 3;

{ cvt.f32.f16 %f91, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r226;
cvt.f32.f16 %f92, low;}


mul.ftz.f32 %f119, %f91, %f92;
cvt.rn.f32.s32 %f120, %r1093;
fma.rn.ftz.f32 %f319, %f119, %f120, %f319;
cvt.u64.u32 %rd61, %r1854;
add.s64 %rd62, %rd1, %rd61;
add.s64 %rd63, %rd62, %rd2;
mul.lo.s64 %rd64, %rd63, 36;
mov.u32 %r276, %r668;

prmt.b32 %r276, %r276, 0, 0xba98;

xor.b32 %r1094, %r769, %r276;
or.b32 %r1095, %r1094, -2139062144;
and.b32 %r1096, %r276, 2139062143;
sub.s32 %r1097, %r1095, %r1096;
xor.b32 %r279, %r787, %r1097;
add.s64 %rd65, %rd7, %rd64;
ld.global.nc.u32 %r280, [%rd65+4];

dp4a.s32.s32 %r278, %r279, %r280, %r722;

mov.u32 %r282, %r674;

prmt.b32 %r282, %r282, 0, 0xba98;

xor.b32 %r1098, %r282, %r770;
or.b32 %r1099, %r1098, -2139062144;
and.b32 %r1100, %r282, 2139062143;
sub.s32 %r1101, %r1099, %r1100;
xor.b32 %r285, %r795, %r1101;
ld.global.nc.u32 %r286, [%rd65+8];

dp4a.s32.s32 %r284, %r285, %r286, %r278;

mov.u32 %r288, %r680;

prmt.b32 %r288, %r288, 0, 0xba98;

xor.b32 %r1102, %r798, %r288;
or.b32 %r1103, %r1102, -2139062144;
and.b32 %r1104, %r288, 2139062143;
sub.s32 %r1105, %r1103, %r1104;
xor.b32 %r291, %r817, %r1105;
ld.global.nc.u32 %r292, [%rd65+12];

dp4a.s32.s32 %r290, %r291, %r292, %r284;

mov.u32 %r294, %r686;

prmt.b32 %r294, %r294, 0, 0xba98;

xor.b32 %r1106, %r294, %r799;
or.b32 %r1107, %r1106, -2139062144;
and.b32 %r1108, %r294, 2139062143;
sub.s32 %r1109, %r1107, %r1108;
xor.b32 %r297, %r825, %r1109;
ld.global.nc.u32 %r298, [%rd65+16];

dp4a.s32.s32 %r296, %r297, %r298, %r290;

mov.u32 %r300, %r692;

prmt.b32 %r300, %r300, 0, 0xba98;

xor.b32 %r1110, %r827, %r300;
or.b32 %r1111, %r1110, -2139062144;
and.b32 %r1112, %r300, 2139062143;
sub.s32 %r1113, %r1111, %r1112;
xor.b32 %r303, %r846, %r1113;
ld.global.nc.u32 %r304, [%rd65+20];

dp4a.s32.s32 %r302, %r303, %r304, %r296;

mov.u32 %r306, %r698;

prmt.b32 %r306, %r306, 0, 0xba98;

xor.b32 %r1114, %r306, %r828;
or.b32 %r1115, %r1114, -2139062144;
and.b32 %r1116, %r306, 2139062143;
sub.s32 %r1117, %r1115, %r1116;
xor.b32 %r309, %r854, %r1117;
ld.global.nc.u32 %r310, [%rd65+24];

dp4a.s32.s32 %r308, %r309, %r310, %r302;

mov.u32 %r312, %r704;

prmt.b32 %r312, %r312, 0, 0xba98;

xor.b32 %r1118, %r857, %r312;
or.b32 %r1119, %r1118, -2139062144;
and.b32 %r1120, %r312, 2139062143;
sub.s32 %r1121, %r1119, %r1120;
xor.b32 %r315, %r876, %r1121;
ld.global.nc.u32 %r316, [%rd65+28];

dp4a.s32.s32 %r314, %r315, %r316, %r308;

mov.u32 %r318, %r710;

prmt.b32 %r318, %r318, 0, 0xba98;

xor.b32 %r1122, %r318, %r858;
or.b32 %r1123, %r1122, -2139062144;
and.b32 %r1124, %r318, 2139062143;
sub.s32 %r1125, %r1123, %r1124;
xor.b32 %r321, %r884, %r1125;
ld.global.nc.u32 %r322, [%rd65+32];

dp4a.s32.s32 %r320, %r321, %r322, %r314;

mul.lo.s32 %r1126, %r320, %r887;
shr.s32 %r1127, %r1126, 31;
shr.u32 %r1128, %r1127, 29;
add.s32 %r1129, %r1126, %r1128;
shr.s32 %r1130, %r1129, 3;

{ cvt.f32.f16 %f93, %rs1;}


ld.global.nc.u32 %r324, [%rd65];

{.reg .f16 low,high;
mov.b32 {low,high},%r324;
cvt.f32.f16 %f94, low;}


mul.ftz.f32 %f121, %f93, %f94;
cvt.rn.f32.s32 %f122, %r1130;
fma.rn.ftz.f32 %f318, %f121, %f122, %f318;
mov.u32 %r325, %r717;

prmt.b32 %r325, %r325, 0, 0xba98;

xor.b32 %r1131, %r896, %r325;
or.b32 %r1132, %r1131, -2139062144;
and.b32 %r1133, %r325, 2139062143;
sub.s32 %r1134, %r1132, %r1133;
xor.b32 %r328, %r914, %r1134;

dp4a.s32.s32 %r327, %r328, %r280, %r722;

mov.u32 %r331, %r723;

prmt.b32 %r331, %r331, 0, 0xba98;

xor.b32 %r1135, %r331, %r897;
or.b32 %r1136, %r1135, -2139062144;
and.b32 %r1137, %r331, 2139062143;
sub.s32 %r1138, %r1136, %r1137;
xor.b32 %r334, %r922, %r1138;

dp4a.s32.s32 %r333, %r334, %r286, %r327;

mov.u32 %r337, %r729;

prmt.b32 %r337, %r337, 0, 0xba98;

xor.b32 %r1139, %r925, %r337;
or.b32 %r1140, %r1139, -2139062144;
and.b32 %r1141, %r337, 2139062143;
sub.s32 %r1142, %r1140, %r1141;
xor.b32 %r340, %r944, %r1142;

dp4a.s32.s32 %r339, %r340, %r292, %r333;

mov.u32 %r343, %r735;

prmt.b32 %r343, %r343, 0, 0xba98;

xor.b32 %r1143, %r343, %r926;
or.b32 %r1144, %r1143, -2139062144;
and.b32 %r1145, %r343, 2139062143;
sub.s32 %r1146, %r1144, %r1145;
xor.b32 %r346, %r952, %r1146;

dp4a.s32.s32 %r345, %r346, %r298, %r339;

mov.u32 %r349, %r741;

prmt.b32 %r349, %r349, 0, 0xba98;

xor.b32 %r1147, %r954, %r349;
or.b32 %r1148, %r1147, -2139062144;
and.b32 %r1149, %r349, 2139062143;
sub.s32 %r1150, %r1148, %r1149;
xor.b32 %r352, %r973, %r1150;

dp4a.s32.s32 %r351, %r352, %r304, %r345;

mov.u32 %r355, %r747;

prmt.b32 %r355, %r355, 0, 0xba98;

xor.b32 %r1151, %r355, %r955;
or.b32 %r1152, %r1151, -2139062144;
and.b32 %r1153, %r355, 2139062143;
sub.s32 %r1154, %r1152, %r1153;
xor.b32 %r358, %r981, %r1154;

dp4a.s32.s32 %r357, %r358, %r310, %r351;

mov.u32 %r361, %r753;

prmt.b32 %r361, %r361, 0, 0xba98;

xor.b32 %r1155, %r984, %r361;
or.b32 %r1156, %r1155, -2139062144;
and.b32 %r1157, %r361, 2139062143;
sub.s32 %r1158, %r1156, %r1157;
xor.b32 %r364, %r1003, %r1158;

dp4a.s32.s32 %r363, %r364, %r316, %r357;

mov.u32 %r367, %r759;

prmt.b32 %r367, %r367, 0, 0xba98;

xor.b32 %r1159, %r367, %r985;
or.b32 %r1160, %r1159, -2139062144;
and.b32 %r1161, %r367, 2139062143;
sub.s32 %r1162, %r1160, %r1161;
xor.b32 %r370, %r1011, %r1162;

dp4a.s32.s32 %r369, %r370, %r322, %r363;

mul.lo.s32 %r1163, %r369, %r1014;
shr.s32 %r1164, %r1163, 31;
shr.u32 %r1165, %r1164, 29;
add.s32 %r1166, %r1163, %r1165;
shr.s32 %r1167, %r1166, 3;

{ cvt.f32.f16 %f95, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r324;
cvt.f32.f16 %f96, low;}


mul.ftz.f32 %f123, %f95, %f96;
cvt.rn.f32.s32 %f124, %r1167;
fma.rn.ftz.f32 %f317, %f123, %f124, %f317;
cvt.u64.u32 %rd66, %r1853;
add.s64 %rd67, %rd1, %rd66;
add.s64 %rd68, %rd67, %rd2;
mul.lo.s64 %rd69, %rd68, 36;
mov.u32 %r374, %r668;

prmt.b32 %r374, %r374, 0, 0xba98;

xor.b32 %r1168, %r769, %r374;
or.b32 %r1169, %r1168, -2139062144;
and.b32 %r1170, %r374, 2139062143;
sub.s32 %r1171, %r1169, %r1170;
xor.b32 %r377, %r787, %r1171;
add.s64 %rd70, %rd7, %rd69;
ld.global.nc.u32 %r378, [%rd70+4];

dp4a.s32.s32 %r376, %r377, %r378, %r722;

mov.u32 %r380, %r674;

prmt.b32 %r380, %r380, 0, 0xba98;

xor.b32 %r1172, %r380, %r770;
or.b32 %r1173, %r1172, -2139062144;
and.b32 %r1174, %r380, 2139062143;
sub.s32 %r1175, %r1173, %r1174;
xor.b32 %r383, %r795, %r1175;
ld.global.nc.u32 %r384, [%rd70+8];

dp4a.s32.s32 %r382, %r383, %r384, %r376;

mov.u32 %r386, %r680;

prmt.b32 %r386, %r386, 0, 0xba98;

xor.b32 %r1176, %r798, %r386;
or.b32 %r1177, %r1176, -2139062144;
and.b32 %r1178, %r386, 2139062143;
sub.s32 %r1179, %r1177, %r1178;
xor.b32 %r389, %r817, %r1179;
ld.global.nc.u32 %r390, [%rd70+12];

dp4a.s32.s32 %r388, %r389, %r390, %r382;

mov.u32 %r392, %r686;

prmt.b32 %r392, %r392, 0, 0xba98;

xor.b32 %r1180, %r392, %r799;
or.b32 %r1181, %r1180, -2139062144;
and.b32 %r1182, %r392, 2139062143;
sub.s32 %r1183, %r1181, %r1182;
xor.b32 %r395, %r825, %r1183;
ld.global.nc.u32 %r396, [%rd70+16];

dp4a.s32.s32 %r394, %r395, %r396, %r388;

mov.u32 %r398, %r692;

prmt.b32 %r398, %r398, 0, 0xba98;

xor.b32 %r1184, %r827, %r398;
or.b32 %r1185, %r1184, -2139062144;
and.b32 %r1186, %r398, 2139062143;
sub.s32 %r1187, %r1185, %r1186;
xor.b32 %r401, %r846, %r1187;
ld.global.nc.u32 %r402, [%rd70+20];

dp4a.s32.s32 %r400, %r401, %r402, %r394;

mov.u32 %r404, %r698;

prmt.b32 %r404, %r404, 0, 0xba98;

xor.b32 %r1188, %r404, %r828;
or.b32 %r1189, %r1188, -2139062144;
and.b32 %r1190, %r404, 2139062143;
sub.s32 %r1191, %r1189, %r1190;
xor.b32 %r407, %r854, %r1191;
ld.global.nc.u32 %r408, [%rd70+24];

dp4a.s32.s32 %r406, %r407, %r408, %r400;

mov.u32 %r410, %r704;

prmt.b32 %r410, %r410, 0, 0xba98;

xor.b32 %r1192, %r857, %r410;
or.b32 %r1193, %r1192, -2139062144;
and.b32 %r1194, %r410, 2139062143;
sub.s32 %r1195, %r1193, %r1194;
xor.b32 %r413, %r876, %r1195;
ld.global.nc.u32 %r414, [%rd70+28];

dp4a.s32.s32 %r412, %r413, %r414, %r406;

mov.u32 %r416, %r710;

prmt.b32 %r416, %r416, 0, 0xba98;

xor.b32 %r1196, %r416, %r858;
or.b32 %r1197, %r1196, -2139062144;
and.b32 %r1198, %r416, 2139062143;
sub.s32 %r1199, %r1197, %r1198;
xor.b32 %r419, %r884, %r1199;
ld.global.nc.u32 %r420, [%rd70+32];

dp4a.s32.s32 %r418, %r419, %r420, %r412;

mul.lo.s32 %r1200, %r418, %r887;
shr.s32 %r1201, %r1200, 31;
shr.u32 %r1202, %r1201, 29;
add.s32 %r1203, %r1200, %r1202;
shr.s32 %r1204, %r1203, 3;

{ cvt.f32.f16 %f97, %rs1;}


ld.global.nc.u32 %r422, [%rd70];

{.reg .f16 low,high;
mov.b32 {low,high},%r422;
cvt.f32.f16 %f98, low;}


mul.ftz.f32 %f125, %f97, %f98;
cvt.rn.f32.s32 %f126, %r1204;
fma.rn.ftz.f32 %f316, %f125, %f126, %f316;
mov.u32 %r423, %r717;

prmt.b32 %r423, %r423, 0, 0xba98;

xor.b32 %r1205, %r896, %r423;
or.b32 %r1206, %r1205, -2139062144;
and.b32 %r1207, %r423, 2139062143;
sub.s32 %r1208, %r1206, %r1207;
xor.b32 %r426, %r914, %r1208;

dp4a.s32.s32 %r425, %r426, %r378, %r722;

mov.u32 %r429, %r723;

prmt.b32 %r429, %r429, 0, 0xba98;

xor.b32 %r1209, %r429, %r897;
or.b32 %r1210, %r1209, -2139062144;
and.b32 %r1211, %r429, 2139062143;
sub.s32 %r1212, %r1210, %r1211;
xor.b32 %r432, %r922, %r1212;

dp4a.s32.s32 %r431, %r432, %r384, %r425;

mov.u32 %r435, %r729;

prmt.b32 %r435, %r435, 0, 0xba98;

xor.b32 %r1213, %r925, %r435;
or.b32 %r1214, %r1213, -2139062144;
and.b32 %r1215, %r435, 2139062143;
sub.s32 %r1216, %r1214, %r1215;
xor.b32 %r438, %r944, %r1216;

dp4a.s32.s32 %r437, %r438, %r390, %r431;

mov.u32 %r441, %r735;

prmt.b32 %r441, %r441, 0, 0xba98;

xor.b32 %r1217, %r441, %r926;
or.b32 %r1218, %r1217, -2139062144;
and.b32 %r1219, %r441, 2139062143;
sub.s32 %r1220, %r1218, %r1219;
xor.b32 %r444, %r952, %r1220;

dp4a.s32.s32 %r443, %r444, %r396, %r437;

mov.u32 %r447, %r741;

prmt.b32 %r447, %r447, 0, 0xba98;

xor.b32 %r1221, %r954, %r447;
or.b32 %r1222, %r1221, -2139062144;
and.b32 %r1223, %r447, 2139062143;
sub.s32 %r1224, %r1222, %r1223;
xor.b32 %r450, %r973, %r1224;

dp4a.s32.s32 %r449, %r450, %r402, %r443;

mov.u32 %r453, %r747;

prmt.b32 %r453, %r453, 0, 0xba98;

xor.b32 %r1225, %r453, %r955;
or.b32 %r1226, %r1225, -2139062144;
and.b32 %r1227, %r453, 2139062143;
sub.s32 %r1228, %r1226, %r1227;
xor.b32 %r456, %r981, %r1228;

dp4a.s32.s32 %r455, %r456, %r408, %r449;

mov.u32 %r459, %r753;

prmt.b32 %r459, %r459, 0, 0xba98;

xor.b32 %r1229, %r984, %r459;
or.b32 %r1230, %r1229, -2139062144;
and.b32 %r1231, %r459, 2139062143;
sub.s32 %r1232, %r1230, %r1231;
xor.b32 %r462, %r1003, %r1232;

dp4a.s32.s32 %r461, %r462, %r414, %r455;

mov.u32 %r465, %r759;

prmt.b32 %r465, %r465, 0, 0xba98;

xor.b32 %r1233, %r465, %r985;
or.b32 %r1234, %r1233, -2139062144;
and.b32 %r1235, %r465, 2139062143;
sub.s32 %r1236, %r1234, %r1235;
xor.b32 %r468, %r1011, %r1236;

dp4a.s32.s32 %r467, %r468, %r420, %r461;

mul.lo.s32 %r1237, %r467, %r1014;
shr.s32 %r1238, %r1237, 31;
shr.u32 %r1239, %r1238, 29;
add.s32 %r1240, %r1237, %r1239;
shr.s32 %r1241, %r1240, 3;

{ cvt.f32.f16 %f99, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r422;
cvt.f32.f16 %f100, low;}


mul.ftz.f32 %f127, %f99, %f100;
cvt.rn.f32.s32 %f128, %r1241;
fma.rn.ftz.f32 %f315, %f127, %f128, %f315;
cvt.u64.u32 %rd71, %r1852;
add.s64 %rd72, %rd1, %rd71;
add.s64 %rd73, %rd72, %rd2;
mul.lo.s64 %rd74, %rd73, 36;
mov.u32 %r472, %r668;

prmt.b32 %r472, %r472, 0, 0xba98;

xor.b32 %r1242, %r769, %r472;
or.b32 %r1243, %r1242, -2139062144;
and.b32 %r1244, %r472, 2139062143;
sub.s32 %r1245, %r1243, %r1244;
xor.b32 %r475, %r787, %r1245;
add.s64 %rd75, %rd7, %rd74;
ld.global.nc.u32 %r476, [%rd75+4];

dp4a.s32.s32 %r474, %r475, %r476, %r722;

mov.u32 %r478, %r674;

prmt.b32 %r478, %r478, 0, 0xba98;

xor.b32 %r1246, %r478, %r770;
or.b32 %r1247, %r1246, -2139062144;
and.b32 %r1248, %r478, 2139062143;
sub.s32 %r1249, %r1247, %r1248;
xor.b32 %r481, %r795, %r1249;
ld.global.nc.u32 %r482, [%rd75+8];

dp4a.s32.s32 %r480, %r481, %r482, %r474;

mov.u32 %r484, %r680;

prmt.b32 %r484, %r484, 0, 0xba98;

xor.b32 %r1250, %r798, %r484;
or.b32 %r1251, %r1250, -2139062144;
and.b32 %r1252, %r484, 2139062143;
sub.s32 %r1253, %r1251, %r1252;
xor.b32 %r487, %r817, %r1253;
ld.global.nc.u32 %r488, [%rd75+12];

dp4a.s32.s32 %r486, %r487, %r488, %r480;

mov.u32 %r490, %r686;

prmt.b32 %r490, %r490, 0, 0xba98;

xor.b32 %r1254, %r490, %r799;
or.b32 %r1255, %r1254, -2139062144;
and.b32 %r1256, %r490, 2139062143;
sub.s32 %r1257, %r1255, %r1256;
xor.b32 %r493, %r825, %r1257;
ld.global.nc.u32 %r494, [%rd75+16];

dp4a.s32.s32 %r492, %r493, %r494, %r486;

mov.u32 %r496, %r692;

prmt.b32 %r496, %r496, 0, 0xba98;

xor.b32 %r1258, %r827, %r496;
or.b32 %r1259, %r1258, -2139062144;
and.b32 %r1260, %r496, 2139062143;
sub.s32 %r1261, %r1259, %r1260;
xor.b32 %r499, %r846, %r1261;
ld.global.nc.u32 %r500, [%rd75+20];

dp4a.s32.s32 %r498, %r499, %r500, %r492;

mov.u32 %r502, %r698;

prmt.b32 %r502, %r502, 0, 0xba98;

xor.b32 %r1262, %r502, %r828;
or.b32 %r1263, %r1262, -2139062144;
and.b32 %r1264, %r502, 2139062143;
sub.s32 %r1265, %r1263, %r1264;
xor.b32 %r505, %r854, %r1265;
ld.global.nc.u32 %r506, [%rd75+24];

dp4a.s32.s32 %r504, %r505, %r506, %r498;

mov.u32 %r508, %r704;

prmt.b32 %r508, %r508, 0, 0xba98;

xor.b32 %r1266, %r857, %r508;
or.b32 %r1267, %r1266, -2139062144;
and.b32 %r1268, %r508, 2139062143;
sub.s32 %r1269, %r1267, %r1268;
xor.b32 %r511, %r876, %r1269;
ld.global.nc.u32 %r512, [%rd75+28];

dp4a.s32.s32 %r510, %r511, %r512, %r504;

mov.u32 %r514, %r710;

prmt.b32 %r514, %r514, 0, 0xba98;

xor.b32 %r1270, %r514, %r858;
or.b32 %r1271, %r1270, -2139062144;
and.b32 %r1272, %r514, 2139062143;
sub.s32 %r1273, %r1271, %r1272;
xor.b32 %r517, %r884, %r1273;
ld.global.nc.u32 %r518, [%rd75+32];

dp4a.s32.s32 %r516, %r517, %r518, %r510;

mul.lo.s32 %r1274, %r516, %r887;
shr.s32 %r1275, %r1274, 31;
shr.u32 %r1276, %r1275, 29;
add.s32 %r1277, %r1274, %r1276;
shr.s32 %r1278, %r1277, 3;

{ cvt.f32.f16 %f101, %rs1;}


ld.global.nc.u32 %r520, [%rd75];

{.reg .f16 low,high;
mov.b32 {low,high},%r520;
cvt.f32.f16 %f102, low;}


mul.ftz.f32 %f129, %f101, %f102;
cvt.rn.f32.s32 %f130, %r1278;
fma.rn.ftz.f32 %f314, %f129, %f130, %f314;
mov.u32 %r521, %r717;

prmt.b32 %r521, %r521, 0, 0xba98;

xor.b32 %r1279, %r896, %r521;
or.b32 %r1280, %r1279, -2139062144;
and.b32 %r1281, %r521, 2139062143;
sub.s32 %r1282, %r1280, %r1281;
xor.b32 %r524, %r914, %r1282;

dp4a.s32.s32 %r523, %r524, %r476, %r722;

mov.u32 %r527, %r723;

prmt.b32 %r527, %r527, 0, 0xba98;

xor.b32 %r1283, %r527, %r897;
or.b32 %r1284, %r1283, -2139062144;
and.b32 %r1285, %r527, 2139062143;
sub.s32 %r1286, %r1284, %r1285;
xor.b32 %r530, %r922, %r1286;

dp4a.s32.s32 %r529, %r530, %r482, %r523;

mov.u32 %r533, %r729;

prmt.b32 %r533, %r533, 0, 0xba98;

xor.b32 %r1287, %r925, %r533;
or.b32 %r1288, %r1287, -2139062144;
and.b32 %r1289, %r533, 2139062143;
sub.s32 %r1290, %r1288, %r1289;
xor.b32 %r536, %r944, %r1290;

dp4a.s32.s32 %r535, %r536, %r488, %r529;

mov.u32 %r539, %r735;

prmt.b32 %r539, %r539, 0, 0xba98;

xor.b32 %r1291, %r539, %r926;
or.b32 %r1292, %r1291, -2139062144;
and.b32 %r1293, %r539, 2139062143;
sub.s32 %r1294, %r1292, %r1293;
xor.b32 %r542, %r952, %r1294;

dp4a.s32.s32 %r541, %r542, %r494, %r535;

mov.u32 %r545, %r741;

prmt.b32 %r545, %r545, 0, 0xba98;

xor.b32 %r1295, %r954, %r545;
or.b32 %r1296, %r1295, -2139062144;
and.b32 %r1297, %r545, 2139062143;
sub.s32 %r1298, %r1296, %r1297;
xor.b32 %r548, %r973, %r1298;

dp4a.s32.s32 %r547, %r548, %r500, %r541;

mov.u32 %r551, %r747;

prmt.b32 %r551, %r551, 0, 0xba98;

xor.b32 %r1299, %r551, %r955;
or.b32 %r1300, %r1299, -2139062144;
and.b32 %r1301, %r551, 2139062143;
sub.s32 %r1302, %r1300, %r1301;
xor.b32 %r554, %r981, %r1302;

dp4a.s32.s32 %r553, %r554, %r506, %r547;

mov.u32 %r557, %r753;

prmt.b32 %r557, %r557, 0, 0xba98;

xor.b32 %r1303, %r984, %r557;
or.b32 %r1304, %r1303, -2139062144;
and.b32 %r1305, %r557, 2139062143;
sub.s32 %r1306, %r1304, %r1305;
xor.b32 %r560, %r1003, %r1306;

dp4a.s32.s32 %r559, %r560, %r512, %r553;

mov.u32 %r563, %r759;

prmt.b32 %r563, %r563, 0, 0xba98;

xor.b32 %r1307, %r563, %r985;
or.b32 %r1308, %r1307, -2139062144;
and.b32 %r1309, %r563, 2139062143;
sub.s32 %r1310, %r1308, %r1309;
xor.b32 %r566, %r1011, %r1310;

dp4a.s32.s32 %r565, %r566, %r518, %r559;

mul.lo.s32 %r1311, %r565, %r1014;
shr.s32 %r1312, %r1311, 31;
shr.u32 %r1313, %r1312, 29;
add.s32 %r1314, %r1311, %r1313;
shr.s32 %r1315, %r1314, 3;

{ cvt.f32.f16 %f103, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r520;
cvt.f32.f16 %f104, low;}


mul.ftz.f32 %f131, %f103, %f104;
cvt.rn.f32.s32 %f132, %r1315;
fma.rn.ftz.f32 %f313, %f131, %f132, %f313;
cvt.u64.u32 %rd76, %r1851;
add.s64 %rd77, %rd1, %rd76;
add.s64 %rd78, %rd77, %rd2;
mul.lo.s64 %rd79, %rd78, 36;
mov.u32 %r570, %r668;

prmt.b32 %r570, %r570, 0, 0xba98;

xor.b32 %r1316, %r769, %r570;
or.b32 %r1317, %r1316, -2139062144;
and.b32 %r1318, %r570, 2139062143;
sub.s32 %r1319, %r1317, %r1318;
xor.b32 %r573, %r787, %r1319;
add.s64 %rd80, %rd7, %rd79;
ld.global.nc.u32 %r574, [%rd80+4];

dp4a.s32.s32 %r572, %r573, %r574, %r722;

mov.u32 %r576, %r674;

prmt.b32 %r576, %r576, 0, 0xba98;

xor.b32 %r1320, %r576, %r770;
or.b32 %r1321, %r1320, -2139062144;
and.b32 %r1322, %r576, 2139062143;
sub.s32 %r1323, %r1321, %r1322;
xor.b32 %r579, %r795, %r1323;
ld.global.nc.u32 %r580, [%rd80+8];

dp4a.s32.s32 %r578, %r579, %r580, %r572;

mov.u32 %r582, %r680;

prmt.b32 %r582, %r582, 0, 0xba98;

xor.b32 %r1324, %r798, %r582;
or.b32 %r1325, %r1324, -2139062144;
and.b32 %r1326, %r582, 2139062143;
sub.s32 %r1327, %r1325, %r1326;
xor.b32 %r585, %r817, %r1327;
ld.global.nc.u32 %r586, [%rd80+12];

dp4a.s32.s32 %r584, %r585, %r586, %r578;

mov.u32 %r588, %r686;

prmt.b32 %r588, %r588, 0, 0xba98;

xor.b32 %r1328, %r588, %r799;
or.b32 %r1329, %r1328, -2139062144;
and.b32 %r1330, %r588, 2139062143;
sub.s32 %r1331, %r1329, %r1330;
xor.b32 %r591, %r825, %r1331;
ld.global.nc.u32 %r592, [%rd80+16];

dp4a.s32.s32 %r590, %r591, %r592, %r584;

mov.u32 %r594, %r692;

prmt.b32 %r594, %r594, 0, 0xba98;

xor.b32 %r1332, %r827, %r594;
or.b32 %r1333, %r1332, -2139062144;
and.b32 %r1334, %r594, 2139062143;
sub.s32 %r1335, %r1333, %r1334;
xor.b32 %r597, %r846, %r1335;
ld.global.nc.u32 %r598, [%rd80+20];

dp4a.s32.s32 %r596, %r597, %r598, %r590;

mov.u32 %r600, %r698;

prmt.b32 %r600, %r600, 0, 0xba98;

xor.b32 %r1336, %r600, %r828;
or.b32 %r1337, %r1336, -2139062144;
and.b32 %r1338, %r600, 2139062143;
sub.s32 %r1339, %r1337, %r1338;
xor.b32 %r603, %r854, %r1339;
ld.global.nc.u32 %r604, [%rd80+24];

dp4a.s32.s32 %r602, %r603, %r604, %r596;

mov.u32 %r606, %r704;

prmt.b32 %r606, %r606, 0, 0xba98;

xor.b32 %r1340, %r857, %r606;
or.b32 %r1341, %r1340, -2139062144;
and.b32 %r1342, %r606, 2139062143;
sub.s32 %r1343, %r1341, %r1342;
xor.b32 %r609, %r876, %r1343;
ld.global.nc.u32 %r610, [%rd80+28];

dp4a.s32.s32 %r608, %r609, %r610, %r602;

mov.u32 %r612, %r710;

prmt.b32 %r612, %r612, 0, 0xba98;

xor.b32 %r1344, %r612, %r858;
or.b32 %r1345, %r1344, -2139062144;
and.b32 %r1346, %r612, 2139062143;
sub.s32 %r1347, %r1345, %r1346;
xor.b32 %r615, %r884, %r1347;
ld.global.nc.u32 %r616, [%rd80+32];

dp4a.s32.s32 %r614, %r615, %r616, %r608;

mul.lo.s32 %r1348, %r614, %r887;
shr.s32 %r1349, %r1348, 31;
shr.u32 %r1350, %r1349, 29;
add.s32 %r1351, %r1348, %r1350;
shr.s32 %r1352, %r1351, 3;

{ cvt.f32.f16 %f105, %rs1;}


ld.global.nc.u32 %r618, [%rd80];

{.reg .f16 low,high;
mov.b32 {low,high},%r618;
cvt.f32.f16 %f106, low;}


mul.ftz.f32 %f133, %f105, %f106;
cvt.rn.f32.s32 %f134, %r1352;
fma.rn.ftz.f32 %f312, %f133, %f134, %f312;
mov.u32 %r619, %r717;

prmt.b32 %r619, %r619, 0, 0xba98;

xor.b32 %r1353, %r896, %r619;
or.b32 %r1354, %r1353, -2139062144;
and.b32 %r1355, %r619, 2139062143;
sub.s32 %r1356, %r1354, %r1355;
xor.b32 %r622, %r914, %r1356;

dp4a.s32.s32 %r621, %r622, %r574, %r722;

mov.u32 %r625, %r723;

prmt.b32 %r625, %r625, 0, 0xba98;

xor.b32 %r1357, %r625, %r897;
or.b32 %r1358, %r1357, -2139062144;
and.b32 %r1359, %r625, 2139062143;
sub.s32 %r1360, %r1358, %r1359;
xor.b32 %r628, %r922, %r1360;

dp4a.s32.s32 %r627, %r628, %r580, %r621;

mov.u32 %r631, %r729;

prmt.b32 %r631, %r631, 0, 0xba98;

xor.b32 %r1361, %r925, %r631;
or.b32 %r1362, %r1361, -2139062144;
and.b32 %r1363, %r631, 2139062143;
sub.s32 %r1364, %r1362, %r1363;
xor.b32 %r634, %r944, %r1364;

dp4a.s32.s32 %r633, %r634, %r586, %r627;

mov.u32 %r637, %r735;

prmt.b32 %r637, %r637, 0, 0xba98;

xor.b32 %r1365, %r637, %r926;
or.b32 %r1366, %r1365, -2139062144;
and.b32 %r1367, %r637, 2139062143;
sub.s32 %r1368, %r1366, %r1367;
xor.b32 %r640, %r952, %r1368;

dp4a.s32.s32 %r639, %r640, %r592, %r633;

mov.u32 %r643, %r741;

prmt.b32 %r643, %r643, 0, 0xba98;

xor.b32 %r1369, %r954, %r643;
or.b32 %r1370, %r1369, -2139062144;
and.b32 %r1371, %r643, 2139062143;
sub.s32 %r1372, %r1370, %r1371;
xor.b32 %r646, %r973, %r1372;

dp4a.s32.s32 %r645, %r646, %r598, %r639;

mov.u32 %r649, %r747;

prmt.b32 %r649, %r649, 0, 0xba98;

xor.b32 %r1373, %r649, %r955;
or.b32 %r1374, %r1373, -2139062144;
and.b32 %r1375, %r649, 2139062143;
sub.s32 %r1376, %r1374, %r1375;
xor.b32 %r652, %r981, %r1376;

dp4a.s32.s32 %r651, %r652, %r604, %r645;

mov.u32 %r655, %r753;

prmt.b32 %r655, %r655, 0, 0xba98;

xor.b32 %r1377, %r984, %r655;
or.b32 %r1378, %r1377, -2139062144;
and.b32 %r1379, %r655, 2139062143;
sub.s32 %r1380, %r1378, %r1379;
xor.b32 %r658, %r1003, %r1380;

dp4a.s32.s32 %r657, %r658, %r610, %r651;

mov.u32 %r661, %r759;

prmt.b32 %r661, %r661, 0, 0xba98;

xor.b32 %r1381, %r661, %r985;
or.b32 %r1382, %r1381, -2139062144;
and.b32 %r1383, %r661, 2139062143;
sub.s32 %r1384, %r1382, %r1383;
xor.b32 %r664, %r1011, %r1384;

dp4a.s32.s32 %r663, %r664, %r616, %r657;

mul.lo.s32 %r1385, %r663, %r1014;
shr.s32 %r1386, %r1385, 31;
shr.u32 %r1387, %r1386, 29;
add.s32 %r1388, %r1385, %r1387;
shr.s32 %r1389, %r1388, 3;

{ cvt.f32.f16 %f107, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r618;
cvt.f32.f16 %f108, low;}


mul.ftz.f32 %f135, %f107, %f108;
cvt.rn.f32.s32 %f136, %r1389;
fma.rn.ftz.f32 %f311, %f135, %f136, %f311;
cvt.u64.u32 %rd81, %r1850;
add.s64 %rd82, %rd1, %rd81;
add.s64 %rd83, %rd82, %rd2;
mul.lo.s64 %rd84, %rd83, 36;

prmt.b32 %r668, %r668, 0, 0xba98;

xor.b32 %r1390, %r769, %r668;
or.b32 %r1391, %r1390, -2139062144;
and.b32 %r1392, %r668, 2139062143;
sub.s32 %r1393, %r1391, %r1392;
xor.b32 %r671, %r787, %r1393;
add.s64 %rd85, %rd7, %rd84;
ld.global.nc.u32 %r672, [%rd85+4];

dp4a.s32.s32 %r670, %r671, %r672, %r722;


prmt.b32 %r674, %r674, 0, 0xba98;

xor.b32 %r1394, %r674, %r770;
or.b32 %r1395, %r1394, -2139062144;
and.b32 %r1396, %r674, 2139062143;
sub.s32 %r1397, %r1395, %r1396;
xor.b32 %r677, %r795, %r1397;
ld.global.nc.u32 %r678, [%rd85+8];

dp4a.s32.s32 %r676, %r677, %r678, %r670;


prmt.b32 %r680, %r680, 0, 0xba98;

xor.b32 %r1398, %r798, %r680;
or.b32 %r1399, %r1398, -2139062144;
and.b32 %r1400, %r680, 2139062143;
sub.s32 %r1401, %r1399, %r1400;
xor.b32 %r683, %r817, %r1401;
ld.global.nc.u32 %r684, [%rd85+12];

dp4a.s32.s32 %r682, %r683, %r684, %r676;


prmt.b32 %r686, %r686, 0, 0xba98;

xor.b32 %r1402, %r686, %r799;
or.b32 %r1403, %r1402, -2139062144;
and.b32 %r1404, %r686, 2139062143;
sub.s32 %r1405, %r1403, %r1404;
xor.b32 %r689, %r825, %r1405;
ld.global.nc.u32 %r690, [%rd85+16];

dp4a.s32.s32 %r688, %r689, %r690, %r682;


prmt.b32 %r692, %r692, 0, 0xba98;

xor.b32 %r1406, %r827, %r692;
or.b32 %r1407, %r1406, -2139062144;
and.b32 %r1408, %r692, 2139062143;
sub.s32 %r1409, %r1407, %r1408;
xor.b32 %r695, %r846, %r1409;
ld.global.nc.u32 %r696, [%rd85+20];

dp4a.s32.s32 %r694, %r695, %r696, %r688;


prmt.b32 %r698, %r698, 0, 0xba98;

xor.b32 %r1410, %r698, %r828;
or.b32 %r1411, %r1410, -2139062144;
and.b32 %r1412, %r698, 2139062143;
sub.s32 %r1413, %r1411, %r1412;
xor.b32 %r701, %r854, %r1413;
ld.global.nc.u32 %r702, [%rd85+24];

dp4a.s32.s32 %r700, %r701, %r702, %r694;


prmt.b32 %r704, %r704, 0, 0xba98;

xor.b32 %r1414, %r857, %r704;
or.b32 %r1415, %r1414, -2139062144;
and.b32 %r1416, %r704, 2139062143;
sub.s32 %r1417, %r1415, %r1416;
xor.b32 %r707, %r876, %r1417;
ld.global.nc.u32 %r708, [%rd85+28];

dp4a.s32.s32 %r706, %r707, %r708, %r700;


prmt.b32 %r710, %r710, 0, 0xba98;

xor.b32 %r1418, %r710, %r858;
or.b32 %r1419, %r1418, -2139062144;
and.b32 %r1420, %r710, 2139062143;
sub.s32 %r1421, %r1419, %r1420;
xor.b32 %r713, %r884, %r1421;
ld.global.nc.u32 %r714, [%rd85+32];

dp4a.s32.s32 %r712, %r713, %r714, %r706;

mul.lo.s32 %r1422, %r712, %r887;
shr.s32 %r1423, %r1422, 31;
shr.u32 %r1424, %r1423, 29;
add.s32 %r1425, %r1422, %r1424;
shr.s32 %r1426, %r1425, 3;

{ cvt.f32.f16 %f109, %rs1;}


ld.global.nc.u32 %r716, [%rd85];

{.reg .f16 low,high;
mov.b32 {low,high},%r716;
cvt.f32.f16 %f110, low;}


mul.ftz.f32 %f137, %f109, %f110;
cvt.rn.f32.s32 %f138, %r1426;
fma.rn.ftz.f32 %f310, %f137, %f138, %f310;

prmt.b32 %r717, %r717, 0, 0xba98;

xor.b32 %r1427, %r896, %r717;
or.b32 %r1428, %r1427, -2139062144;
and.b32 %r1429, %r717, 2139062143;
sub.s32 %r1430, %r1428, %r1429;
xor.b32 %r720, %r914, %r1430;

dp4a.s32.s32 %r719, %r720, %r672, %r722;


prmt.b32 %r723, %r723, 0, 0xba98;

xor.b32 %r1431, %r723, %r897;
or.b32 %r1432, %r1431, -2139062144;
and.b32 %r1433, %r723, 2139062143;
sub.s32 %r1434, %r1432, %r1433;
xor.b32 %r726, %r922, %r1434;

dp4a.s32.s32 %r725, %r726, %r678, %r719;


prmt.b32 %r729, %r729, 0, 0xba98;

xor.b32 %r1435, %r925, %r729;
or.b32 %r1436, %r1435, -2139062144;
and.b32 %r1437, %r729, 2139062143;
sub.s32 %r1438, %r1436, %r1437;
xor.b32 %r732, %r944, %r1438;

dp4a.s32.s32 %r731, %r732, %r684, %r725;


prmt.b32 %r735, %r735, 0, 0xba98;

xor.b32 %r1439, %r735, %r926;
or.b32 %r1440, %r1439, -2139062144;
and.b32 %r1441, %r735, 2139062143;
sub.s32 %r1442, %r1440, %r1441;
xor.b32 %r738, %r952, %r1442;

dp4a.s32.s32 %r737, %r738, %r690, %r731;


prmt.b32 %r741, %r741, 0, 0xba98;

xor.b32 %r1443, %r954, %r741;
or.b32 %r1444, %r1443, -2139062144;
and.b32 %r1445, %r741, 2139062143;
sub.s32 %r1446, %r1444, %r1445;
xor.b32 %r744, %r973, %r1446;

dp4a.s32.s32 %r743, %r744, %r696, %r737;


prmt.b32 %r747, %r747, 0, 0xba98;

xor.b32 %r1447, %r747, %r955;
or.b32 %r1448, %r1447, -2139062144;
and.b32 %r1449, %r747, 2139062143;
sub.s32 %r1450, %r1448, %r1449;
xor.b32 %r750, %r981, %r1450;

dp4a.s32.s32 %r749, %r750, %r702, %r743;


prmt.b32 %r753, %r753, 0, 0xba98;

xor.b32 %r1451, %r984, %r753;
or.b32 %r1452, %r1451, -2139062144;
and.b32 %r1453, %r753, 2139062143;
sub.s32 %r1454, %r1452, %r1453;
xor.b32 %r756, %r1003, %r1454;

dp4a.s32.s32 %r755, %r756, %r708, %r749;


prmt.b32 %r759, %r759, 0, 0xba98;

xor.b32 %r1455, %r759, %r985;
or.b32 %r1456, %r1455, -2139062144;
and.b32 %r1457, %r759, 2139062143;
sub.s32 %r1458, %r1456, %r1457;
xor.b32 %r762, %r1011, %r1458;

dp4a.s32.s32 %r761, %r762, %r714, %r755;

mul.lo.s32 %r1459, %r761, %r1014;
shr.s32 %r1460, %r1459, 31;
shr.u32 %r1461, %r1460, 29;
add.s32 %r1462, %r1459, %r1461;
shr.s32 %r1463, %r1462, 3;

{ cvt.f32.f16 %f111, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r716;
cvt.f32.f16 %f112, low;}


mul.ftz.f32 %f139, %f111, %f112;
cvt.rn.f32.s32 %f140, %r1463;
fma.rn.ftz.f32 %f309, %f139, %f140, %f309;
add.s32 %r1854, %r1854, 64;
add.s32 %r1853, %r1853, 64;
add.s32 %r1852, %r1852, 64;
add.s32 %r1851, %r1851, 64;
add.s32 %r1850, %r1850, 64;
add.s32 %r1849, %r1849, 64;
add.s32 %r1848, %r1848, 64;
add.s64 %rd145, %rd145, 528;
add.s64 %rd144, %rd144, 528;
add.s64 %rd143, %rd143, 528;
add.s32 %r1855, %r1855, 8;
setp.lt.s32 %p3, %r1855, %r4;
@%p3 bra $L__BB209_2;

$L__BB209_3:
setp.eq.s32 %p4, %r1, 0;
@%p4 bra $L__BB209_5;

mov.u32 %r1466, _ZZ13mul_mat_vec_qIL9ggml_type16ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r1467, %r1, 1792, %r1466;
shl.b32 %r1468, %r47, 2;
add.s32 %r1469, %r1467, %r1468;
st.shared.f32 [%r1469+-1792], %f322;
st.shared.f32 [%r1469+-1664], %f321;
st.shared.f32 [%r1469+-1536], %f320;
st.shared.f32 [%r1469+-1408], %f319;
st.shared.f32 [%r1469+-1280], %f318;
st.shared.f32 [%r1469+-1152], %f317;
st.shared.f32 [%r1469+-1024], %f316;
st.shared.f32 [%r1469+-896], %f315;
st.shared.f32 [%r1469+-768], %f314;
st.shared.f32 [%r1469+-640], %f313;
st.shared.f32 [%r1469+-512], %f312;
st.shared.f32 [%r1469+-384], %f311;
st.shared.f32 [%r1469+-256], %f310;
st.shared.f32 [%r1469+-128], %f309;

$L__BB209_5:
setp.ne.s32 %p5, %r1, 0;
bar.sync 0;
@%p5 bra $L__BB209_34;

mov.u32 %r1471, %ctaid.x;
shl.b32 %r1472, %r1471, 1;
mov.u32 %r1473, 1;
shl.b32 %r1475, %r47, 2;
mov.u32 %r1476, _ZZ13mul_mat_vec_qIL9ggml_type16ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r1477, %r1476, %r1475;
mov.u32 %r1478, 2;
ld.shared.f32 %f141, [%r1477];
add.ftz.f32 %f142, %f141, %f322;
mov.b32 %r1479, %f142;
mov.u32 %r1480, 31;
mov.u32 %r1481, 16;
mov.u32 %r1482, -1;
shfl.sync.bfly.b32 %r1483|%p6, %r1479, %r1481, %r1480, %r1482;
mov.b32 %f143, %r1483;
add.ftz.f32 %f144, %f142, %f143;
mov.b32 %r1484, %f144;
mov.u32 %r1485, 8;
shfl.sync.bfly.b32 %r1486|%p7, %r1484, %r1485, %r1480, %r1482;
mov.b32 %f145, %r1486;
add.ftz.f32 %f146, %f144, %f145;
mov.b32 %r1487, %f146;
mov.u32 %r1488, 4;
shfl.sync.bfly.b32 %r1489|%p8, %r1487, %r1488, %r1480, %r1482;
mov.b32 %f147, %r1489;
add.ftz.f32 %f148, %f146, %f147;
mov.b32 %r1490, %f148;
shfl.sync.bfly.b32 %r1491|%p9, %r1490, %r1478, %r1480, %r1482;
mov.b32 %f149, %r1491;
add.ftz.f32 %f150, %f148, %f149;
mov.b32 %r1492, %f150;
shfl.sync.bfly.b32 %r1493|%p10, %r1492, %r1473, %r1480, %r1482;
mov.b32 %f151, %r1493;
add.ftz.f32 %f43, %f150, %f151;
setp.ne.s32 %p11, %r47, 0;
setp.ge.u32 %p12, %r1472, %r32;
or.pred %p13, %p11, %p12;
@%p13 bra $L__BB209_8;

cvta.to.global.u64 %rd86, %rd19;
mov.u32 %r1494, %ctaid.y;
mad.lo.s32 %r1497, %r1494, %r38, %r1472;
mov.u32 %r1498, %ctaid.z;
mad.lo.s32 %r1499, %r1498, %r44, %r1497;
mul.wide.u32 %rd87, %r1499, 4;
add.s64 %rd88, %rd86, %rd87;
st.global.f32 [%rd88], %f43;

$L__BB209_8:
ld.shared.f32 %f152, [%r1477+128];
add.ftz.f32 %f153, %f152, %f321;
mov.b32 %r1505, %f153;
shfl.sync.bfly.b32 %r1509|%p14, %r1505, %r1481, %r1480, %r1482;
mov.b32 %f154, %r1509;
add.ftz.f32 %f155, %f153, %f154;
mov.b32 %r1510, %f155;
shfl.sync.bfly.b32 %r1512|%p15, %r1510, %r1485, %r1480, %r1482;
mov.b32 %f156, %r1512;
add.ftz.f32 %f157, %f155, %f156;
mov.b32 %r1513, %f157;
shfl.sync.bfly.b32 %r1515|%p16, %r1513, %r1488, %r1480, %r1482;
mov.b32 %f158, %r1515;
add.ftz.f32 %f159, %f157, %f158;
mov.b32 %r1516, %f159;
shfl.sync.bfly.b32 %r1517|%p17, %r1516, %r1478, %r1480, %r1482;
mov.b32 %f160, %r1517;
add.ftz.f32 %f161, %f159, %f160;
mov.b32 %r1518, %f161;
shfl.sync.bfly.b32 %r1520|%p18, %r1518, %r1473, %r1480, %r1482;
mov.b32 %f162, %r1520;
add.ftz.f32 %f44, %f161, %f162;
setp.eq.s32 %p19, %r47, 1;
or.b32 %r1523, %r1472, 1;
setp.lt.u32 %p20, %r1523, %r32;
and.pred %p1, %p19, %p20;
not.pred %p21, %p1;
@%p21 bra $L__BB209_10;

cvta.to.global.u64 %rd89, %rd19;
mov.u32 %r1524, %ctaid.y;
mad.lo.s32 %r1527, %r1524, %r38, %r1472;
mov.u32 %r1528, %ctaid.z;
mad.lo.s32 %r1529, %r1528, %r44, %r1527;
mul.wide.u32 %rd90, %r1529, 4;
add.s64 %rd91, %rd89, %rd90;
st.global.f32 [%rd91+4], %f44;

$L__BB209_10:
mov.u32 %r1534, 2;
ld.shared.f32 %f163, [%r1477+256];
add.ftz.f32 %f164, %f163, %f320;
mov.b32 %r1535, %f164;
mov.u32 %r1536, 31;
mov.u32 %r1537, 16;
mov.u32 %r1538, -1;
shfl.sync.bfly.b32 %r1539|%p22, %r1535, %r1537, %r1536, %r1538;
mov.b32 %f165, %r1539;
add.ftz.f32 %f166, %f164, %f165;
mov.b32 %r1540, %f166;
mov.u32 %r1541, 8;
shfl.sync.bfly.b32 %r1542|%p23, %r1540, %r1541, %r1536, %r1538;
mov.b32 %f167, %r1542;
add.ftz.f32 %f168, %f166, %f167;
mov.b32 %r1543, %f168;
mov.u32 %r1544, 4;
shfl.sync.bfly.b32 %r1545|%p24, %r1543, %r1544, %r1536, %r1538;
mov.b32 %f169, %r1545;
add.ftz.f32 %f170, %f168, %f169;
mov.b32 %r1546, %f170;
shfl.sync.bfly.b32 %r1547|%p25, %r1546, %r1534, %r1536, %r1538;
mov.b32 %f171, %r1547;
add.ftz.f32 %f172, %f170, %f171;
mov.b32 %r1548, %f172;
mov.u32 %r1549, 1;
shfl.sync.bfly.b32 %r1550|%p26, %r1548, %r1549, %r1536, %r1538;
mov.b32 %f173, %r1550;
add.ftz.f32 %f45, %f172, %f173;
@%p13 bra $L__BB209_12;

cvt.u64.u32 %rd92, %r32;
mov.u32 %r1553, %ctaid.y;
mad.lo.s32 %r1556, %r1553, %r38, %r1472;
mov.u32 %r1557, %ctaid.z;
mad.lo.s32 %r1558, %r1557, %r44, %r1556;
cvt.u64.u32 %rd93, %r1558;
add.s64 %rd94, %rd92, %rd93;
cvta.to.global.u64 %rd95, %rd19;
shl.b64 %rd96, %rd94, 2;
add.s64 %rd97, %rd95, %rd96;
st.global.f32 [%rd97], %f45;

$L__BB209_12:
ld.shared.f32 %f174, [%r1477+384];
add.ftz.f32 %f175, %f174, %f319;
mov.b32 %r1564, %f175;
shfl.sync.bfly.b32 %r1568|%p30, %r1564, %r1537, %r1536, %r1538;
mov.b32 %f176, %r1568;
add.ftz.f32 %f177, %f175, %f176;
mov.b32 %r1569, %f177;
shfl.sync.bfly.b32 %r1571|%p31, %r1569, %r1541, %r1536, %r1538;
mov.b32 %f178, %r1571;
add.ftz.f32 %f179, %f177, %f178;
mov.b32 %r1572, %f179;
shfl.sync.bfly.b32 %r1574|%p32, %r1572, %r1544, %r1536, %r1538;
mov.b32 %f180, %r1574;
add.ftz.f32 %f181, %f179, %f180;
mov.b32 %r1575, %f181;
shfl.sync.bfly.b32 %r1576|%p33, %r1575, %r1534, %r1536, %r1538;
mov.b32 %f182, %r1576;
add.ftz.f32 %f183, %f181, %f182;
mov.b32 %r1577, %f183;
shfl.sync.bfly.b32 %r1579|%p34, %r1577, %r1549, %r1536, %r1538;
mov.b32 %f184, %r1579;
add.ftz.f32 %f46, %f183, %f184;
@%p21 bra $L__BB209_14;

add.s32 %r1580, %r32, 1;
cvt.u64.u32 %rd98, %r1580;
mov.u32 %r1581, %ctaid.y;
mad.lo.s32 %r1584, %r1581, %r38, %r1472;
mov.u32 %r1585, %ctaid.z;
mad.lo.s32 %r1586, %r1585, %r44, %r1584;
cvt.u64.u32 %rd99, %r1586;
add.s64 %rd100, %rd98, %rd99;
cvta.to.global.u64 %rd101, %rd19;
shl.b64 %rd102, %rd100, 2;
add.s64 %rd103, %rd101, %rd102;
st.global.f32 [%rd103], %f46;

$L__BB209_14:
shl.b32 %r1587, %r32, 1;
mov.u32 %r1588, 1;
mov.u32 %r1593, 2;
ld.shared.f32 %f185, [%r1477+512];
add.ftz.f32 %f186, %f185, %f318;
mov.b32 %r1594, %f186;
mov.u32 %r1595, 31;
mov.u32 %r1596, 16;
mov.u32 %r1597, -1;
shfl.sync.bfly.b32 %r1598|%p36, %r1594, %r1596, %r1595, %r1597;
mov.b32 %f187, %r1598;
add.ftz.f32 %f188, %f186, %f187;
mov.b32 %r1599, %f188;
mov.u32 %r1600, 8;
shfl.sync.bfly.b32 %r1601|%p37, %r1599, %r1600, %r1595, %r1597;
mov.b32 %f189, %r1601;
add.ftz.f32 %f190, %f188, %f189;
mov.b32 %r1602, %f190;
mov.u32 %r1603, 4;
shfl.sync.bfly.b32 %r1604|%p38, %r1602, %r1603, %r1595, %r1597;
mov.b32 %f191, %r1604;
add.ftz.f32 %f192, %f190, %f191;
mov.b32 %r1605, %f192;
shfl.sync.bfly.b32 %r1606|%p39, %r1605, %r1593, %r1595, %r1597;
mov.b32 %f193, %r1606;
add.ftz.f32 %f194, %f192, %f193;
mov.b32 %r1607, %f194;
shfl.sync.bfly.b32 %r1608|%p40, %r1607, %r1588, %r1595, %r1597;
mov.b32 %f195, %r1608;
add.ftz.f32 %f47, %f194, %f195;
cvt.u64.u32 %rd104, %r1587;
mov.u32 %r1609, %ctaid.y;
mad.lo.s32 %r1612, %r1609, %r38, %r1472;
mov.u32 %r1613, %ctaid.z;
mad.lo.s32 %r1614, %r1613, %r44, %r1612;
cvt.u64.u32 %rd105, %r1614;
add.s64 %rd106, %rd104, %rd105;
cvta.to.global.u64 %rd107, %rd19;
shl.b64 %rd108, %rd106, 2;
add.s64 %rd14, %rd107, %rd108;
@%p13 bra $L__BB209_16;

st.global.f32 [%rd14], %f47;

$L__BB209_16:
ld.shared.f32 %f196, [%r1477+640];
add.ftz.f32 %f197, %f196, %f317;
mov.b32 %r1620, %f197;
shfl.sync.bfly.b32 %r1624|%p44, %r1620, %r1596, %r1595, %r1597;
mov.b32 %f198, %r1624;
add.ftz.f32 %f199, %f197, %f198;
mov.b32 %r1625, %f199;
shfl.sync.bfly.b32 %r1627|%p45, %r1625, %r1600, %r1595, %r1597;
mov.b32 %f200, %r1627;
add.ftz.f32 %f201, %f199, %f200;
mov.b32 %r1628, %f201;
shfl.sync.bfly.b32 %r1630|%p46, %r1628, %r1603, %r1595, %r1597;
mov.b32 %f202, %r1630;
add.ftz.f32 %f203, %f201, %f202;
mov.b32 %r1631, %f203;
shfl.sync.bfly.b32 %r1632|%p47, %r1631, %r1593, %r1595, %r1597;
mov.b32 %f204, %r1632;
add.ftz.f32 %f205, %f203, %f204;
mov.b32 %r1633, %f205;
shfl.sync.bfly.b32 %r1635|%p48, %r1633, %r1588, %r1595, %r1597;
mov.b32 %f206, %r1635;
add.ftz.f32 %f48, %f205, %f206;
@%p21 bra $L__BB209_18;

st.global.f32 [%rd14+4], %f48;

$L__BB209_18:
mul.lo.s32 %r28, %r32, 3;
mov.u32 %r1640, 2;
ld.shared.f32 %f207, [%r1477+768];
add.ftz.f32 %f208, %f207, %f316;
mov.b32 %r1641, %f208;
mov.u32 %r1642, 31;
mov.u32 %r1643, 16;
mov.u32 %r1644, -1;
shfl.sync.bfly.b32 %r1645|%p50, %r1641, %r1643, %r1642, %r1644;
mov.b32 %f209, %r1645;
add.ftz.f32 %f210, %f208, %f209;
mov.b32 %r1646, %f210;
mov.u32 %r1647, 8;
shfl.sync.bfly.b32 %r1648|%p51, %r1646, %r1647, %r1642, %r1644;
mov.b32 %f211, %r1648;
add.ftz.f32 %f212, %f210, %f211;
mov.b32 %r1649, %f212;
mov.u32 %r1650, 4;
shfl.sync.bfly.b32 %r1651|%p52, %r1649, %r1650, %r1642, %r1644;
mov.b32 %f213, %r1651;
add.ftz.f32 %f214, %f212, %f213;
mov.b32 %r1652, %f214;
shfl.sync.bfly.b32 %r1653|%p53, %r1652, %r1640, %r1642, %r1644;
mov.b32 %f215, %r1653;
add.ftz.f32 %f216, %f214, %f215;
mov.b32 %r1654, %f216;
mov.u32 %r1655, 1;
shfl.sync.bfly.b32 %r1656|%p54, %r1654, %r1655, %r1642, %r1644;
mov.b32 %f217, %r1656;
add.ftz.f32 %f49, %f216, %f217;
@%p13 bra $L__BB209_20;

cvt.u64.u32 %rd109, %r28;
add.s64 %rd111, %rd109, %rd105;
shl.b64 %rd113, %rd111, 2;
add.s64 %rd114, %rd107, %rd113;
st.global.f32 [%rd114], %f49;

$L__BB209_20:
ld.shared.f32 %f218, [%r1477+896];
add.ftz.f32 %f219, %f218, %f315;
mov.b32 %r1670, %f219;
shfl.sync.bfly.b32 %r1674|%p58, %r1670, %r1643, %r1642, %r1644;
mov.b32 %f220, %r1674;
add.ftz.f32 %f221, %f219, %f220;
mov.b32 %r1675, %f221;
shfl.sync.bfly.b32 %r1677|%p59, %r1675, %r1647, %r1642, %r1644;
mov.b32 %f222, %r1677;
add.ftz.f32 %f223, %f221, %f222;
mov.b32 %r1678, %f223;
shfl.sync.bfly.b32 %r1680|%p60, %r1678, %r1650, %r1642, %r1644;
mov.b32 %f224, %r1680;
add.ftz.f32 %f225, %f223, %f224;
mov.b32 %r1681, %f225;
shfl.sync.bfly.b32 %r1682|%p61, %r1681, %r1640, %r1642, %r1644;
mov.b32 %f226, %r1682;
add.ftz.f32 %f227, %f225, %f226;
mov.b32 %r1683, %f227;
shfl.sync.bfly.b32 %r1685|%p62, %r1683, %r1655, %r1642, %r1644;
mov.b32 %f228, %r1685;
add.ftz.f32 %f50, %f227, %f228;
@%p21 bra $L__BB209_22;

add.s32 %r1686, %r28, 1;
cvt.u64.u32 %rd115, %r1686;
add.s64 %rd117, %rd115, %rd105;
shl.b64 %rd119, %rd117, 2;
add.s64 %rd120, %rd107, %rd119;
st.global.f32 [%rd120], %f50;

$L__BB209_22:
shl.b32 %r1693, %r32, 2;
mov.u32 %r1694, 2;
ld.shared.f32 %f229, [%r1477+1024];
add.ftz.f32 %f230, %f229, %f314;
mov.b32 %r1699, %f230;
mov.u32 %r1700, 31;
mov.u32 %r1701, 16;
mov.u32 %r1702, -1;
shfl.sync.bfly.b32 %r1703|%p64, %r1699, %r1701, %r1700, %r1702;
mov.b32 %f231, %r1703;
add.ftz.f32 %f232, %f230, %f231;
mov.b32 %r1704, %f232;
mov.u32 %r1705, 8;
shfl.sync.bfly.b32 %r1706|%p65, %r1704, %r1705, %r1700, %r1702;
mov.b32 %f233, %r1706;
add.ftz.f32 %f234, %f232, %f233;
mov.b32 %r1707, %f234;
mov.u32 %r1708, 4;
shfl.sync.bfly.b32 %r1709|%p66, %r1707, %r1708, %r1700, %r1702;
mov.b32 %f235, %r1709;
add.ftz.f32 %f236, %f234, %f235;
mov.b32 %r1710, %f236;
shfl.sync.bfly.b32 %r1711|%p67, %r1710, %r1694, %r1700, %r1702;
mov.b32 %f237, %r1711;
add.ftz.f32 %f238, %f236, %f237;
mov.b32 %r1712, %f238;
mov.u32 %r1713, 1;
shfl.sync.bfly.b32 %r1714|%p68, %r1712, %r1713, %r1700, %r1702;
mov.b32 %f239, %r1714;
add.ftz.f32 %f51, %f238, %f239;
cvt.u64.u32 %rd121, %r1693;
add.s64 %rd123, %rd121, %rd105;
shl.b64 %rd125, %rd123, 2;
add.s64 %rd15, %rd107, %rd125;
@%p13 bra $L__BB209_24;

st.global.f32 [%rd15], %f51;

$L__BB209_24:
ld.shared.f32 %f240, [%r1477+1152];
add.ftz.f32 %f241, %f240, %f313;
mov.b32 %r1726, %f241;
shfl.sync.bfly.b32 %r1730|%p72, %r1726, %r1701, %r1700, %r1702;
mov.b32 %f242, %r1730;
add.ftz.f32 %f243, %f241, %f242;
mov.b32 %r1731, %f243;
shfl.sync.bfly.b32 %r1733|%p73, %r1731, %r1705, %r1700, %r1702;
mov.b32 %f244, %r1733;
add.ftz.f32 %f245, %f243, %f244;
mov.b32 %r1734, %f245;
shfl.sync.bfly.b32 %r1736|%p74, %r1734, %r1708, %r1700, %r1702;
mov.b32 %f246, %r1736;
add.ftz.f32 %f247, %f245, %f246;
mov.b32 %r1737, %f247;
shfl.sync.bfly.b32 %r1738|%p75, %r1737, %r1694, %r1700, %r1702;
mov.b32 %f248, %r1738;
add.ftz.f32 %f249, %f247, %f248;
mov.b32 %r1739, %f249;
shfl.sync.bfly.b32 %r1741|%p76, %r1739, %r1713, %r1700, %r1702;
mov.b32 %f250, %r1741;
add.ftz.f32 %f52, %f249, %f250;
@%p21 bra $L__BB209_26;

st.global.f32 [%rd15+4], %f52;

$L__BB209_26:
mul.lo.s32 %r29, %r32, 5;
mov.u32 %r1746, 2;
ld.shared.f32 %f251, [%r1477+1280];
add.ftz.f32 %f252, %f251, %f312;
mov.b32 %r1747, %f252;
mov.u32 %r1748, 31;
mov.u32 %r1749, 16;
mov.u32 %r1750, -1;
shfl.sync.bfly.b32 %r1751|%p78, %r1747, %r1749, %r1748, %r1750;
mov.b32 %f253, %r1751;
add.ftz.f32 %f254, %f252, %f253;
mov.b32 %r1752, %f254;
mov.u32 %r1753, 8;
shfl.sync.bfly.b32 %r1754|%p79, %r1752, %r1753, %r1748, %r1750;
mov.b32 %f255, %r1754;
add.ftz.f32 %f256, %f254, %f255;
mov.b32 %r1755, %f256;
mov.u32 %r1756, 4;
shfl.sync.bfly.b32 %r1757|%p80, %r1755, %r1756, %r1748, %r1750;
mov.b32 %f257, %r1757;
add.ftz.f32 %f258, %f256, %f257;
mov.b32 %r1758, %f258;
shfl.sync.bfly.b32 %r1759|%p81, %r1758, %r1746, %r1748, %r1750;
mov.b32 %f259, %r1759;
add.ftz.f32 %f260, %f258, %f259;
mov.b32 %r1760, %f260;
mov.u32 %r1761, 1;
shfl.sync.bfly.b32 %r1762|%p82, %r1760, %r1761, %r1748, %r1750;
mov.b32 %f261, %r1762;
add.ftz.f32 %f53, %f260, %f261;
@%p13 bra $L__BB209_28;

cvt.u64.u32 %rd126, %r29;
add.s64 %rd128, %rd126, %rd105;
shl.b64 %rd130, %rd128, 2;
add.s64 %rd131, %rd107, %rd130;
st.global.f32 [%rd131], %f53;

$L__BB209_28:
ld.shared.f32 %f262, [%r1477+1408];
add.ftz.f32 %f263, %f262, %f311;
mov.b32 %r1776, %f263;
shfl.sync.bfly.b32 %r1780|%p86, %r1776, %r1749, %r1748, %r1750;
mov.b32 %f264, %r1780;
add.ftz.f32 %f265, %f263, %f264;
mov.b32 %r1781, %f265;
shfl.sync.bfly.b32 %r1783|%p87, %r1781, %r1753, %r1748, %r1750;
mov.b32 %f266, %r1783;
add.ftz.f32 %f267, %f265, %f266;
mov.b32 %r1784, %f267;
shfl.sync.bfly.b32 %r1786|%p88, %r1784, %r1756, %r1748, %r1750;
mov.b32 %f268, %r1786;
add.ftz.f32 %f269, %f267, %f268;
mov.b32 %r1787, %f269;
shfl.sync.bfly.b32 %r1788|%p89, %r1787, %r1746, %r1748, %r1750;
mov.b32 %f270, %r1788;
add.ftz.f32 %f271, %f269, %f270;
mov.b32 %r1789, %f271;
shfl.sync.bfly.b32 %r1791|%p90, %r1789, %r1761, %r1748, %r1750;
mov.b32 %f272, %r1791;
add.ftz.f32 %f54, %f271, %f272;
@%p21 bra $L__BB209_30;

add.s32 %r1792, %r29, 1;
cvt.u64.u32 %rd132, %r1792;
add.s64 %rd134, %rd132, %rd105;
shl.b64 %rd136, %rd134, 2;
add.s64 %rd137, %rd107, %rd136;
st.global.f32 [%rd137], %f54;

$L__BB209_30:
mul.lo.s32 %r1799, %r32, 6;
mov.u32 %r1804, 2;
ld.shared.f32 %f273, [%r1477+1536];
add.ftz.f32 %f274, %f273, %f310;
mov.b32 %r1805, %f274;
mov.u32 %r1806, 31;
mov.u32 %r1807, 16;
mov.u32 %r1808, -1;
shfl.sync.bfly.b32 %r1809|%p92, %r1805, %r1807, %r1806, %r1808;
mov.b32 %f275, %r1809;
add.ftz.f32 %f276, %f274, %f275;
mov.b32 %r1810, %f276;
mov.u32 %r1811, 8;
shfl.sync.bfly.b32 %r1812|%p93, %r1810, %r1811, %r1806, %r1808;
mov.b32 %f277, %r1812;
add.ftz.f32 %f278, %f276, %f277;
mov.b32 %r1813, %f278;
mov.u32 %r1814, 4;
shfl.sync.bfly.b32 %r1815|%p94, %r1813, %r1814, %r1806, %r1808;
mov.b32 %f279, %r1815;
add.ftz.f32 %f280, %f278, %f279;
mov.b32 %r1816, %f280;
shfl.sync.bfly.b32 %r1817|%p95, %r1816, %r1804, %r1806, %r1808;
mov.b32 %f281, %r1817;
add.ftz.f32 %f282, %f280, %f281;
mov.b32 %r1818, %f282;
mov.u32 %r1819, 1;
shfl.sync.bfly.b32 %r1820|%p96, %r1818, %r1819, %r1806, %r1808;
mov.b32 %f283, %r1820;
add.ftz.f32 %f55, %f282, %f283;
cvt.u64.u32 %rd138, %r1799;
add.s64 %rd140, %rd138, %rd105;
shl.b64 %rd142, %rd140, 2;
add.s64 %rd16, %rd107, %rd142;
@%p13 bra $L__BB209_32;

st.global.f32 [%rd16], %f55;

$L__BB209_32:
ld.shared.f32 %f284, [%r1477+1664];
add.ftz.f32 %f285, %f284, %f309;
mov.b32 %r1832, %f285;
shfl.sync.bfly.b32 %r1836|%p100, %r1832, %r1807, %r1806, %r1808;
mov.b32 %f286, %r1836;
add.ftz.f32 %f287, %f285, %f286;
mov.b32 %r1837, %f287;
shfl.sync.bfly.b32 %r1839|%p101, %r1837, %r1811, %r1806, %r1808;
mov.b32 %f288, %r1839;
add.ftz.f32 %f289, %f287, %f288;
mov.b32 %r1840, %f289;
shfl.sync.bfly.b32 %r1842|%p102, %r1840, %r1814, %r1806, %r1808;
mov.b32 %f290, %r1842;
add.ftz.f32 %f291, %f289, %f290;
mov.b32 %r1843, %f291;
shfl.sync.bfly.b32 %r1844|%p103, %r1843, %r1804, %r1806, %r1808;
mov.b32 %f292, %r1844;
add.ftz.f32 %f293, %f291, %f292;
mov.b32 %r1845, %f293;
shfl.sync.bfly.b32 %r1847|%p104, %r1845, %r1819, %r1806, %r1808;
mov.b32 %f294, %r1847;
add.ftz.f32 %f56, %f293, %f294;
@%p21 bra $L__BB209_34;

st.global.f32 [%rd16+4], %f56;

$L__BB209_34:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type16ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type16ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type16ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type16ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type16ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type16ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type16ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type16ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type16ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type16ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<107>;
.reg .b16 %rs<33>;
.reg .f32 %f<369>;
.reg .b32 %r<1944>;
.reg .b64 %rd<150>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type16ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[2048];

ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type16ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd20, [_Z13mul_mat_vec_qIL9ggml_type16ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd21, [_Z13mul_mat_vec_qIL9ggml_type16ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r52, [_Z13mul_mat_vec_qIL9ggml_type16ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type16ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type16ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type16ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type16ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type16ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r45, [_Z13mul_mat_vec_qIL9ggml_type16ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r49, [_Z13mul_mat_vec_qIL9ggml_type16ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r50, [_Z13mul_mat_vec_qIL9ggml_type16ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r51, [_Z13mul_mat_vec_qIL9ggml_type16ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r47, [_Z13mul_mat_vec_qIL9ggml_type16ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r46, [_Z13mul_mat_vec_qIL9ggml_type16ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type16ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type16ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r53, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r53, %r2;
mov.u32 %r54, %ctaid.x;
shl.b32 %r4, %r54, 1;
shr.s32 %r55, %r3, 31;
shr.u32 %r56, %r55, 29;
add.s32 %r57, %r3, %r56;
shr.s32 %r1943, %r57, 3;
shr.u32 %r6, %r52, 8;
setp.ge.s32 %p3, %r1943, %r6;
mov.f32 %f353, 0f00000000;
mov.f32 %f354, %f353;
mov.f32 %f355, %f353;
mov.f32 %f356, %f353;
mov.f32 %f357, %f353;
mov.f32 %f358, %f353;
mov.f32 %f359, %f353;
mov.f32 %f360, %f353;
mov.f32 %f361, %f353;
mov.f32 %f362, %f353;
mov.f32 %f363, %f353;
mov.f32 %f364, %f353;
mov.f32 %f365, %f353;
mov.f32 %f366, %f353;
mov.f32 %f367, %f353;
mov.f32 %f368, %f353;
@%p3 bra $L__BB210_3;

mov.u32 %r58, %ctaid.y;
mov.u32 %r59, %ctaid.z;
mul.lo.s32 %r60, %r58, %r44;
cvt.u64.u32 %rd22, %r60;
mul.lo.s32 %r61, %r59, %r50;
cvt.u64.u32 %rd23, %r61;
add.s64 %rd1, %rd23, %rd22;
mul.hi.u32 %r62, %r59, %r46;
add.s32 %r63, %r62, %r59;
shr.u32 %r64, %r63, %r47;
mul.lo.s32 %r65, %r64, %r49;
mul.hi.u32 %r66, %r58, %r40;
add.s32 %r67, %r66, %r58;
shr.u32 %r68, %r67, %r41;
mul.lo.s32 %r69, %r68, %r43;
cvta.to.global.u64 %rd24, %rd19;
and.b32 %r73, %r57, -8;
sub.s32 %r74, %r3, %r73;
shl.b32 %r75, %r74, 2;
cvt.s64.s32 %rd2, %r74;
shl.b32 %r76, %r38, 1;
shl.b32 %r1935, %r1943, 3;
add.s32 %r1942, %r1935, %r76;
mad.lo.s32 %r1941, %r38, 3, %r1935;
shl.b32 %r77, %r38, 2;
add.s32 %r1940, %r1935, %r77;
mad.lo.s32 %r1939, %r38, 5, %r1935;
mad.lo.s32 %r1938, %r38, 6, %r1935;
mad.lo.s32 %r1937, %r38, 7, %r1935;
add.s32 %r1936, %r38, %r1935;
add.s32 %r78, %r1943, %r65;
add.s32 %r79, %r78, %r69;
add.s32 %r80, %r4, 1;
mad.lo.s32 %r81, %r37, %r80, %r79;
mul.wide.s32 %rd25, %r81, 66;
add.s64 %rd149, %rd24, %rd25;
mad.lo.s32 %r82, %r4, %r37, %r1943;
add.s32 %r83, %r82, %r65;
add.s32 %r84, %r83, %r69;
mul.wide.s32 %rd26, %r84, 66;
add.s64 %rd27, %rd24, %rd26;
add.s64 %rd148, %rd27, 4;
mul.wide.s32 %rd28, %r75, 2;
add.s64 %rd29, %rd26, %rd28;
add.s64 %rd30, %rd24, %rd29;
add.s64 %rd147, %rd30, 8;
or.b64 %rd6, %rd28, 4;
cvta.to.global.u64 %rd7, %rd20;
mov.f32 %f353, 0f00000000;

$L__BB210_2:
cvt.u64.u32 %rd31, %r1935;
add.s64 %rd32, %rd1, %rd31;
add.s64 %rd33, %rd32, %rd2;
mul.lo.s64 %rd34, %rd33, 36;
ld.global.nc.u16 %rs17, [%rd147+-6];
shr.u16 %rs18, %rs17, 8;
ld.global.nc.u16 %rs19, [%rd147+-4];
shr.u16 %rs20, %rs19, 8;
ld.global.nc.u16 %rs21, [%rd147+-2];
cvt.u32.u16 %r869, %rs21;
ld.global.nc.u16 %rs22, [%rd147];
cvt.u32.u16 %r870, %rs22;
mov.b32 %r871, {%rs21, %rs22};
shl.b16 %rs23, %rs17, 3;
cvt.u64.u16 %rd35, %rs23;
and.b64 %rd36, %rd35, 2040;
mov.u64 %rd37, iq2xxs_grid;
add.s64 %rd38, %rd37, %rd36;
ld.global.nc.v2.u32 {%r872, %r873}, [%rd38];
not.b32 %r875, %r872;
mov.u32 %r825, 0;
not.b32 %r877, %r873;
and.b32 %r878, %r869, 255;
popc.b32 %r879, %r878;
shl.b32 %r880, %r879, 7;
and.b32 %r881, %r880, 128;
xor.b32 %r882, %r881, %r878;
mul.lo.s32 %r883, %r882, 16843009;
and.b32 %r884, %r883, 134480385;
or.b32 %r885, %r884, -2139062144;
add.s32 %r886, %r885, -16843009;
or.b32 %r771, %r886, %r884;
mov.u32 %r85, %r771;

prmt.b32 %r85, %r85, 0, 0xba98;

xor.b32 %r887, %r872, %r85;
or.b32 %r888, %r887, -2139062144;
and.b32 %r889, %r85, 2139062143;
and.b32 %r890, %r875, -2139062144;
sub.s32 %r891, %r888, %r889;
xor.b32 %r88, %r890, %r891;
add.s64 %rd39, %rd7, %rd34;
ld.global.nc.u32 %r89, [%rd39+4];

dp4a.s32.s32 %r87, %r88, %r89, %r825;

and.b32 %r892, %r883, -2143281136;
or.b32 %r893, %r892, -2139062144;
add.s32 %r894, %r893, -16843009;
or.b32 %r777, %r894, %r892;
mov.u32 %r91, %r777;

prmt.b32 %r91, %r91, 0, 0xba98;

xor.b32 %r895, %r91, %r873;
or.b32 %r896, %r895, -2139062144;
and.b32 %r897, %r91, 2139062143;
and.b32 %r898, %r877, -2139062144;
sub.s32 %r899, %r896, %r897;
xor.b32 %r94, %r898, %r899;
ld.global.nc.u32 %r95, [%rd39+8];

dp4a.s32.s32 %r93, %r94, %r95, %r87;

cvt.u32.u16 %r900, %rs18;
mul.wide.u32 %rd40, %r900, 8;
add.s64 %rd41, %rd37, %rd40;
ld.global.nc.v2.u32 {%r901, %r902}, [%rd41];
not.b32 %r904, %r901;
not.b32 %r906, %r902;
shr.u32 %r907, %r869, 7;
and.b32 %r908, %r907, 255;
popc.b32 %r909, %r908;
shl.b32 %r910, %r909, 7;
and.b32 %r911, %r910, 128;
xor.b32 %r912, %r911, %r908;
mul.lo.s32 %r913, %r912, 16843009;
and.b32 %r914, %r913, 134480385;
or.b32 %r915, %r914, -2139062144;
add.s32 %r916, %r915, -16843009;
or.b32 %r783, %r916, %r914;
mov.u32 %r97, %r783;

prmt.b32 %r97, %r97, 0, 0xba98;

xor.b32 %r917, %r901, %r97;
or.b32 %r918, %r917, -2139062144;
and.b32 %r919, %r97, 2139062143;
and.b32 %r920, %r904, -2139062144;
sub.s32 %r921, %r918, %r919;
xor.b32 %r100, %r920, %r921;
ld.global.nc.u32 %r101, [%rd39+12];

dp4a.s32.s32 %r99, %r100, %r101, %r93;

and.b32 %r922, %r913, -2143281136;
or.b32 %r923, %r922, -2139062144;
add.s32 %r924, %r923, -16843009;
or.b32 %r789, %r924, %r922;
mov.u32 %r103, %r789;

prmt.b32 %r103, %r103, 0, 0xba98;

xor.b32 %r925, %r103, %r902;
or.b32 %r926, %r925, -2139062144;
and.b32 %r927, %r103, 2139062143;
and.b32 %r928, %r906, -2139062144;
sub.s32 %r929, %r926, %r927;
xor.b32 %r106, %r928, %r929;
ld.global.nc.u32 %r107, [%rd39+16];

dp4a.s32.s32 %r105, %r106, %r107, %r99;

shl.b16 %rs24, %rs19, 3;
cvt.u64.u16 %rd42, %rs24;
and.b64 %rd43, %rd42, 2040;
add.s64 %rd44, %rd37, %rd43;
ld.global.nc.v2.u32 {%r930, %r931}, [%rd44];
not.b32 %r933, %r930;
not.b32 %r935, %r931;
shr.u32 %r936, %r871, 14;
and.b32 %r937, %r936, 255;
popc.b32 %r938, %r937;
shl.b32 %r939, %r938, 7;
and.b32 %r940, %r939, 128;
xor.b32 %r941, %r940, %r937;
mul.lo.s32 %r942, %r941, 16843009;
and.b32 %r943, %r942, 134480385;
or.b32 %r944, %r943, -2139062144;
add.s32 %r945, %r944, -16843009;
or.b32 %r795, %r945, %r943;
mov.u32 %r109, %r795;

prmt.b32 %r109, %r109, 0, 0xba98;

xor.b32 %r946, %r930, %r109;
or.b32 %r947, %r946, -2139062144;
and.b32 %r948, %r109, 2139062143;
and.b32 %r949, %r933, -2139062144;
sub.s32 %r950, %r947, %r948;
xor.b32 %r112, %r949, %r950;
ld.global.nc.u32 %r113, [%rd39+20];

dp4a.s32.s32 %r111, %r112, %r113, %r105;

and.b32 %r951, %r942, -2143281136;
or.b32 %r952, %r951, -2139062144;
add.s32 %r953, %r952, -16843009;
or.b32 %r801, %r953, %r951;
mov.u32 %r115, %r801;

prmt.b32 %r115, %r115, 0, 0xba98;

xor.b32 %r954, %r115, %r931;
or.b32 %r955, %r954, -2139062144;
and.b32 %r956, %r115, 2139062143;
and.b32 %r957, %r935, -2139062144;
sub.s32 %r958, %r955, %r956;
xor.b32 %r118, %r957, %r958;
ld.global.nc.u32 %r119, [%rd39+24];

dp4a.s32.s32 %r117, %r118, %r119, %r111;

cvt.u32.u16 %r959, %rs20;
mul.wide.u32 %rd45, %r959, 8;
add.s64 %rd46, %rd37, %rd45;
ld.global.nc.v2.u32 {%r960, %r961}, [%rd46];
not.b32 %r963, %r960;
not.b32 %r965, %r961;
shr.u32 %r966, %r870, 5;
and.b32 %r967, %r966, 255;
popc.b32 %r968, %r967;
shl.b32 %r969, %r968, 7;
and.b32 %r970, %r969, 128;
xor.b32 %r971, %r970, %r967;
mul.lo.s32 %r972, %r971, 16843009;
and.b32 %r973, %r972, 134480385;
or.b32 %r974, %r973, -2139062144;
add.s32 %r975, %r974, -16843009;
or.b32 %r807, %r975, %r973;
mov.u32 %r121, %r807;

prmt.b32 %r121, %r121, 0, 0xba98;

xor.b32 %r976, %r960, %r121;
or.b32 %r977, %r976, -2139062144;
and.b32 %r978, %r121, 2139062143;
and.b32 %r979, %r963, -2139062144;
sub.s32 %r980, %r977, %r978;
xor.b32 %r124, %r979, %r980;
ld.global.nc.u32 %r125, [%rd39+28];

dp4a.s32.s32 %r123, %r124, %r125, %r117;

and.b32 %r981, %r972, -2143281136;
or.b32 %r982, %r981, -2139062144;
add.s32 %r983, %r982, -16843009;
or.b32 %r813, %r983, %r981;
mov.u32 %r127, %r813;

prmt.b32 %r127, %r127, 0, 0xba98;

xor.b32 %r984, %r127, %r961;
or.b32 %r985, %r984, -2139062144;
and.b32 %r986, %r127, 2139062143;
and.b32 %r987, %r965, -2139062144;
sub.s32 %r988, %r985, %r986;
xor.b32 %r130, %r987, %r988;
ld.global.nc.u32 %r131, [%rd39+32];

dp4a.s32.s32 %r129, %r130, %r131, %r123;

shr.u32 %r989, %r870, 11;
or.b32 %r990, %r989, 1;
mul.lo.s32 %r991, %r129, %r990;
shr.s32 %r992, %r991, 31;
shr.u32 %r993, %r992, 29;
add.s32 %r994, %r991, %r993;
shr.s32 %r995, %r994, 3;
ld.global.nc.u16 %rs1, [%rd148+-4];

{ cvt.f32.f16 %f97, %rs1;}


ld.global.nc.u32 %r133, [%rd39];

{.reg .f16 low,high;
mov.b32 {low,high},%r133;
cvt.f32.f16 %f98, low;}


mul.ftz.f32 %f129, %f97, %f98;
cvt.rn.f32.s32 %f130, %r995;
fma.rn.ftz.f32 %f368, %f129, %f130, %f368;
add.s64 %rd47, %rd149, %rd6;
ld.global.nc.u16 %rs25, [%rd47+-2];
shr.u16 %rs26, %rs25, 8;
ld.global.nc.u16 %rs27, [%rd47];
shr.u16 %rs28, %rs27, 8;
ld.global.nc.u16 %rs29, [%rd47+2];
cvt.u32.u16 %r996, %rs29;
ld.global.nc.u16 %rs30, [%rd47+4];
cvt.u32.u16 %r997, %rs30;
mov.b32 %r998, {%rs29, %rs30};
shl.b16 %rs31, %rs25, 3;
cvt.u64.u16 %rd48, %rs31;
and.b64 %rd49, %rd48, 2040;
add.s64 %rd50, %rd37, %rd49;
ld.global.nc.v2.u32 {%r999, %r1000}, [%rd50];
not.b32 %r1002, %r999;
not.b32 %r1004, %r1000;
and.b32 %r1005, %r996, 255;
popc.b32 %r1006, %r1005;
shl.b32 %r1007, %r1006, 7;
and.b32 %r1008, %r1007, 128;
xor.b32 %r1009, %r1008, %r1005;
mul.lo.s32 %r1010, %r1009, 16843009;
and.b32 %r1011, %r1010, 134480385;
or.b32 %r1012, %r1011, -2139062144;
add.s32 %r1013, %r1012, -16843009;
or.b32 %r820, %r1013, %r1011;
mov.u32 %r134, %r820;

prmt.b32 %r134, %r134, 0, 0xba98;

xor.b32 %r1014, %r999, %r134;
or.b32 %r1015, %r1014, -2139062144;
and.b32 %r1016, %r134, 2139062143;
and.b32 %r1017, %r1002, -2139062144;
sub.s32 %r1018, %r1015, %r1016;
xor.b32 %r137, %r1017, %r1018;

dp4a.s32.s32 %r136, %r137, %r89, %r825;

and.b32 %r1019, %r1010, -2143281136;
or.b32 %r1020, %r1019, -2139062144;
add.s32 %r1021, %r1020, -16843009;
or.b32 %r826, %r1021, %r1019;
mov.u32 %r140, %r826;

prmt.b32 %r140, %r140, 0, 0xba98;

xor.b32 %r1022, %r140, %r1000;
or.b32 %r1023, %r1022, -2139062144;
and.b32 %r1024, %r140, 2139062143;
and.b32 %r1025, %r1004, -2139062144;
sub.s32 %r1026, %r1023, %r1024;
xor.b32 %r143, %r1025, %r1026;

dp4a.s32.s32 %r142, %r143, %r95, %r136;

cvt.u32.u16 %r1027, %rs26;
mul.wide.u32 %rd51, %r1027, 8;
add.s64 %rd52, %rd37, %rd51;
ld.global.nc.v2.u32 {%r1028, %r1029}, [%rd52];
not.b32 %r1031, %r1028;
not.b32 %r1033, %r1029;
shr.u32 %r1034, %r996, 7;
and.b32 %r1035, %r1034, 255;
popc.b32 %r1036, %r1035;
shl.b32 %r1037, %r1036, 7;
and.b32 %r1038, %r1037, 128;
xor.b32 %r1039, %r1038, %r1035;
mul.lo.s32 %r1040, %r1039, 16843009;
and.b32 %r1041, %r1040, 134480385;
or.b32 %r1042, %r1041, -2139062144;
add.s32 %r1043, %r1042, -16843009;
or.b32 %r832, %r1043, %r1041;
mov.u32 %r146, %r832;

prmt.b32 %r146, %r146, 0, 0xba98;

xor.b32 %r1044, %r1028, %r146;
or.b32 %r1045, %r1044, -2139062144;
and.b32 %r1046, %r146, 2139062143;
and.b32 %r1047, %r1031, -2139062144;
sub.s32 %r1048, %r1045, %r1046;
xor.b32 %r149, %r1047, %r1048;

dp4a.s32.s32 %r148, %r149, %r101, %r142;

and.b32 %r1049, %r1040, -2143281136;
or.b32 %r1050, %r1049, -2139062144;
add.s32 %r1051, %r1050, -16843009;
or.b32 %r838, %r1051, %r1049;
mov.u32 %r152, %r838;

prmt.b32 %r152, %r152, 0, 0xba98;

xor.b32 %r1052, %r152, %r1029;
or.b32 %r1053, %r1052, -2139062144;
and.b32 %r1054, %r152, 2139062143;
and.b32 %r1055, %r1033, -2139062144;
sub.s32 %r1056, %r1053, %r1054;
xor.b32 %r155, %r1055, %r1056;

dp4a.s32.s32 %r154, %r155, %r107, %r148;

shl.b16 %rs32, %rs27, 3;
cvt.u64.u16 %rd53, %rs32;
and.b64 %rd54, %rd53, 2040;
add.s64 %rd55, %rd37, %rd54;
ld.global.nc.v2.u32 {%r1057, %r1058}, [%rd55];
not.b32 %r1060, %r1057;
not.b32 %r1062, %r1058;
shr.u32 %r1063, %r998, 14;
and.b32 %r1064, %r1063, 255;
popc.b32 %r1065, %r1064;
shl.b32 %r1066, %r1065, 7;
and.b32 %r1067, %r1066, 128;
xor.b32 %r1068, %r1067, %r1064;
mul.lo.s32 %r1069, %r1068, 16843009;
and.b32 %r1070, %r1069, 134480385;
or.b32 %r1071, %r1070, -2139062144;
add.s32 %r1072, %r1071, -16843009;
or.b32 %r844, %r1072, %r1070;
mov.u32 %r158, %r844;

prmt.b32 %r158, %r158, 0, 0xba98;

xor.b32 %r1073, %r1057, %r158;
or.b32 %r1074, %r1073, -2139062144;
and.b32 %r1075, %r158, 2139062143;
and.b32 %r1076, %r1060, -2139062144;
sub.s32 %r1077, %r1074, %r1075;
xor.b32 %r161, %r1076, %r1077;

dp4a.s32.s32 %r160, %r161, %r113, %r154;

and.b32 %r1078, %r1069, -2143281136;
or.b32 %r1079, %r1078, -2139062144;
add.s32 %r1080, %r1079, -16843009;
or.b32 %r850, %r1080, %r1078;
mov.u32 %r164, %r850;

prmt.b32 %r164, %r164, 0, 0xba98;

xor.b32 %r1081, %r164, %r1058;
or.b32 %r1082, %r1081, -2139062144;
and.b32 %r1083, %r164, 2139062143;
and.b32 %r1084, %r1062, -2139062144;
sub.s32 %r1085, %r1082, %r1083;
xor.b32 %r167, %r1084, %r1085;

dp4a.s32.s32 %r166, %r167, %r119, %r160;

cvt.u32.u16 %r1086, %rs28;
mul.wide.u32 %rd56, %r1086, 8;
add.s64 %rd57, %rd37, %rd56;
ld.global.nc.v2.u32 {%r1087, %r1088}, [%rd57];
not.b32 %r1090, %r1087;
not.b32 %r1092, %r1088;
shr.u32 %r1093, %r997, 5;
and.b32 %r1094, %r1093, 255;
popc.b32 %r1095, %r1094;
shl.b32 %r1096, %r1095, 7;
and.b32 %r1097, %r1096, 128;
xor.b32 %r1098, %r1097, %r1094;
mul.lo.s32 %r1099, %r1098, 16843009;
and.b32 %r1100, %r1099, 134480385;
or.b32 %r1101, %r1100, -2139062144;
add.s32 %r1102, %r1101, -16843009;
or.b32 %r856, %r1102, %r1100;
mov.u32 %r170, %r856;

prmt.b32 %r170, %r170, 0, 0xba98;

xor.b32 %r1103, %r1087, %r170;
or.b32 %r1104, %r1103, -2139062144;
and.b32 %r1105, %r170, 2139062143;
and.b32 %r1106, %r1090, -2139062144;
sub.s32 %r1107, %r1104, %r1105;
xor.b32 %r173, %r1106, %r1107;

dp4a.s32.s32 %r172, %r173, %r125, %r166;

and.b32 %r1108, %r1099, -2143281136;
or.b32 %r1109, %r1108, -2139062144;
add.s32 %r1110, %r1109, -16843009;
or.b32 %r862, %r1110, %r1108;
mov.u32 %r176, %r862;

prmt.b32 %r176, %r176, 0, 0xba98;

xor.b32 %r1111, %r176, %r1088;
or.b32 %r1112, %r1111, -2139062144;
and.b32 %r1113, %r176, 2139062143;
and.b32 %r1114, %r1092, -2139062144;
sub.s32 %r1115, %r1112, %r1113;
xor.b32 %r179, %r1114, %r1115;

dp4a.s32.s32 %r178, %r179, %r131, %r172;

shr.u32 %r1116, %r997, 11;
or.b32 %r1117, %r1116, 1;
mul.lo.s32 %r1118, %r178, %r1117;
shr.s32 %r1119, %r1118, 31;
shr.u32 %r1120, %r1119, 29;
add.s32 %r1121, %r1118, %r1120;
shr.s32 %r1122, %r1121, 3;
ld.global.nc.u16 %rs2, [%rd149];

{ cvt.f32.f16 %f99, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r133;
cvt.f32.f16 %f100, low;}


mul.ftz.f32 %f131, %f99, %f100;
cvt.rn.f32.s32 %f132, %r1122;
fma.rn.ftz.f32 %f367, %f131, %f132, %f367;
cvt.u64.u32 %rd58, %r1936;
add.s64 %rd59, %rd1, %rd58;
add.s64 %rd60, %rd59, %rd2;
mul.lo.s64 %rd61, %rd60, 36;
mov.u32 %r183, %r771;

prmt.b32 %r183, %r183, 0, 0xba98;

xor.b32 %r1123, %r872, %r183;
or.b32 %r1124, %r1123, -2139062144;
and.b32 %r1125, %r183, 2139062143;
sub.s32 %r1126, %r1124, %r1125;
xor.b32 %r186, %r890, %r1126;
add.s64 %rd62, %rd7, %rd61;
ld.global.nc.u32 %r187, [%rd62+4];

dp4a.s32.s32 %r185, %r186, %r187, %r825;

mov.u32 %r189, %r777;

prmt.b32 %r189, %r189, 0, 0xba98;

xor.b32 %r1127, %r189, %r873;
or.b32 %r1128, %r1127, -2139062144;
and.b32 %r1129, %r189, 2139062143;
sub.s32 %r1130, %r1128, %r1129;
xor.b32 %r192, %r898, %r1130;
ld.global.nc.u32 %r193, [%rd62+8];

dp4a.s32.s32 %r191, %r192, %r193, %r185;

mov.u32 %r195, %r783;

prmt.b32 %r195, %r195, 0, 0xba98;

xor.b32 %r1131, %r901, %r195;
or.b32 %r1132, %r1131, -2139062144;
and.b32 %r1133, %r195, 2139062143;
sub.s32 %r1134, %r1132, %r1133;
xor.b32 %r198, %r920, %r1134;
ld.global.nc.u32 %r199, [%rd62+12];

dp4a.s32.s32 %r197, %r198, %r199, %r191;

mov.u32 %r201, %r789;

prmt.b32 %r201, %r201, 0, 0xba98;

xor.b32 %r1135, %r201, %r902;
or.b32 %r1136, %r1135, -2139062144;
and.b32 %r1137, %r201, 2139062143;
sub.s32 %r1138, %r1136, %r1137;
xor.b32 %r204, %r928, %r1138;
ld.global.nc.u32 %r205, [%rd62+16];

dp4a.s32.s32 %r203, %r204, %r205, %r197;

mov.u32 %r207, %r795;

prmt.b32 %r207, %r207, 0, 0xba98;

xor.b32 %r1139, %r930, %r207;
or.b32 %r1140, %r1139, -2139062144;
and.b32 %r1141, %r207, 2139062143;
sub.s32 %r1142, %r1140, %r1141;
xor.b32 %r210, %r949, %r1142;
ld.global.nc.u32 %r211, [%rd62+20];

dp4a.s32.s32 %r209, %r210, %r211, %r203;

mov.u32 %r213, %r801;

prmt.b32 %r213, %r213, 0, 0xba98;

xor.b32 %r1143, %r213, %r931;
or.b32 %r1144, %r1143, -2139062144;
and.b32 %r1145, %r213, 2139062143;
sub.s32 %r1146, %r1144, %r1145;
xor.b32 %r216, %r957, %r1146;
ld.global.nc.u32 %r217, [%rd62+24];

dp4a.s32.s32 %r215, %r216, %r217, %r209;

mov.u32 %r219, %r807;

prmt.b32 %r219, %r219, 0, 0xba98;

xor.b32 %r1147, %r960, %r219;
or.b32 %r1148, %r1147, -2139062144;
and.b32 %r1149, %r219, 2139062143;
sub.s32 %r1150, %r1148, %r1149;
xor.b32 %r222, %r979, %r1150;
ld.global.nc.u32 %r223, [%rd62+28];

dp4a.s32.s32 %r221, %r222, %r223, %r215;

mov.u32 %r225, %r813;

prmt.b32 %r225, %r225, 0, 0xba98;

xor.b32 %r1151, %r225, %r961;
or.b32 %r1152, %r1151, -2139062144;
and.b32 %r1153, %r225, 2139062143;
sub.s32 %r1154, %r1152, %r1153;
xor.b32 %r228, %r987, %r1154;
ld.global.nc.u32 %r229, [%rd62+32];

dp4a.s32.s32 %r227, %r228, %r229, %r221;

mul.lo.s32 %r1155, %r227, %r990;
shr.s32 %r1156, %r1155, 31;
shr.u32 %r1157, %r1156, 29;
add.s32 %r1158, %r1155, %r1157;
shr.s32 %r1159, %r1158, 3;

{ cvt.f32.f16 %f101, %rs1;}


ld.global.nc.u32 %r231, [%rd62];

{.reg .f16 low,high;
mov.b32 {low,high},%r231;
cvt.f32.f16 %f102, low;}


mul.ftz.f32 %f133, %f101, %f102;
cvt.rn.f32.s32 %f134, %r1159;
fma.rn.ftz.f32 %f366, %f133, %f134, %f366;
mov.u32 %r232, %r820;

prmt.b32 %r232, %r232, 0, 0xba98;

xor.b32 %r1160, %r999, %r232;
or.b32 %r1161, %r1160, -2139062144;
and.b32 %r1162, %r232, 2139062143;
sub.s32 %r1163, %r1161, %r1162;
xor.b32 %r235, %r1017, %r1163;

dp4a.s32.s32 %r234, %r235, %r187, %r825;

mov.u32 %r238, %r826;

prmt.b32 %r238, %r238, 0, 0xba98;

xor.b32 %r1164, %r238, %r1000;
or.b32 %r1165, %r1164, -2139062144;
and.b32 %r1166, %r238, 2139062143;
sub.s32 %r1167, %r1165, %r1166;
xor.b32 %r241, %r1025, %r1167;

dp4a.s32.s32 %r240, %r241, %r193, %r234;

mov.u32 %r244, %r832;

prmt.b32 %r244, %r244, 0, 0xba98;

xor.b32 %r1168, %r1028, %r244;
or.b32 %r1169, %r1168, -2139062144;
and.b32 %r1170, %r244, 2139062143;
sub.s32 %r1171, %r1169, %r1170;
xor.b32 %r247, %r1047, %r1171;

dp4a.s32.s32 %r246, %r247, %r199, %r240;

mov.u32 %r250, %r838;

prmt.b32 %r250, %r250, 0, 0xba98;

xor.b32 %r1172, %r250, %r1029;
or.b32 %r1173, %r1172, -2139062144;
and.b32 %r1174, %r250, 2139062143;
sub.s32 %r1175, %r1173, %r1174;
xor.b32 %r253, %r1055, %r1175;

dp4a.s32.s32 %r252, %r253, %r205, %r246;

mov.u32 %r256, %r844;

prmt.b32 %r256, %r256, 0, 0xba98;

xor.b32 %r1176, %r1057, %r256;
or.b32 %r1177, %r1176, -2139062144;
and.b32 %r1178, %r256, 2139062143;
sub.s32 %r1179, %r1177, %r1178;
xor.b32 %r259, %r1076, %r1179;

dp4a.s32.s32 %r258, %r259, %r211, %r252;

mov.u32 %r262, %r850;

prmt.b32 %r262, %r262, 0, 0xba98;

xor.b32 %r1180, %r262, %r1058;
or.b32 %r1181, %r1180, -2139062144;
and.b32 %r1182, %r262, 2139062143;
sub.s32 %r1183, %r1181, %r1182;
xor.b32 %r265, %r1084, %r1183;

dp4a.s32.s32 %r264, %r265, %r217, %r258;

mov.u32 %r268, %r856;

prmt.b32 %r268, %r268, 0, 0xba98;

xor.b32 %r1184, %r1087, %r268;
or.b32 %r1185, %r1184, -2139062144;
and.b32 %r1186, %r268, 2139062143;
sub.s32 %r1187, %r1185, %r1186;
xor.b32 %r271, %r1106, %r1187;

dp4a.s32.s32 %r270, %r271, %r223, %r264;

mov.u32 %r274, %r862;

prmt.b32 %r274, %r274, 0, 0xba98;

xor.b32 %r1188, %r274, %r1088;
or.b32 %r1189, %r1188, -2139062144;
and.b32 %r1190, %r274, 2139062143;
sub.s32 %r1191, %r1189, %r1190;
xor.b32 %r277, %r1114, %r1191;

dp4a.s32.s32 %r276, %r277, %r229, %r270;

mul.lo.s32 %r1192, %r276, %r1117;
shr.s32 %r1193, %r1192, 31;
shr.u32 %r1194, %r1193, 29;
add.s32 %r1195, %r1192, %r1194;
shr.s32 %r1196, %r1195, 3;

{ cvt.f32.f16 %f103, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r231;
cvt.f32.f16 %f104, low;}


mul.ftz.f32 %f135, %f103, %f104;
cvt.rn.f32.s32 %f136, %r1196;
fma.rn.ftz.f32 %f365, %f135, %f136, %f365;
cvt.u64.u32 %rd63, %r1942;
add.s64 %rd64, %rd1, %rd63;
add.s64 %rd65, %rd64, %rd2;
mul.lo.s64 %rd66, %rd65, 36;
mov.u32 %r281, %r771;

prmt.b32 %r281, %r281, 0, 0xba98;

xor.b32 %r1197, %r872, %r281;
or.b32 %r1198, %r1197, -2139062144;
and.b32 %r1199, %r281, 2139062143;
sub.s32 %r1200, %r1198, %r1199;
xor.b32 %r284, %r890, %r1200;
add.s64 %rd67, %rd7, %rd66;
ld.global.nc.u32 %r285, [%rd67+4];

dp4a.s32.s32 %r283, %r284, %r285, %r825;

mov.u32 %r287, %r777;

prmt.b32 %r287, %r287, 0, 0xba98;

xor.b32 %r1201, %r287, %r873;
or.b32 %r1202, %r1201, -2139062144;
and.b32 %r1203, %r287, 2139062143;
sub.s32 %r1204, %r1202, %r1203;
xor.b32 %r290, %r898, %r1204;
ld.global.nc.u32 %r291, [%rd67+8];

dp4a.s32.s32 %r289, %r290, %r291, %r283;

mov.u32 %r293, %r783;

prmt.b32 %r293, %r293, 0, 0xba98;

xor.b32 %r1205, %r901, %r293;
or.b32 %r1206, %r1205, -2139062144;
and.b32 %r1207, %r293, 2139062143;
sub.s32 %r1208, %r1206, %r1207;
xor.b32 %r296, %r920, %r1208;
ld.global.nc.u32 %r297, [%rd67+12];

dp4a.s32.s32 %r295, %r296, %r297, %r289;

mov.u32 %r299, %r789;

prmt.b32 %r299, %r299, 0, 0xba98;

xor.b32 %r1209, %r299, %r902;
or.b32 %r1210, %r1209, -2139062144;
and.b32 %r1211, %r299, 2139062143;
sub.s32 %r1212, %r1210, %r1211;
xor.b32 %r302, %r928, %r1212;
ld.global.nc.u32 %r303, [%rd67+16];

dp4a.s32.s32 %r301, %r302, %r303, %r295;

mov.u32 %r305, %r795;

prmt.b32 %r305, %r305, 0, 0xba98;

xor.b32 %r1213, %r930, %r305;
or.b32 %r1214, %r1213, -2139062144;
and.b32 %r1215, %r305, 2139062143;
sub.s32 %r1216, %r1214, %r1215;
xor.b32 %r308, %r949, %r1216;
ld.global.nc.u32 %r309, [%rd67+20];

dp4a.s32.s32 %r307, %r308, %r309, %r301;

mov.u32 %r311, %r801;

prmt.b32 %r311, %r311, 0, 0xba98;

xor.b32 %r1217, %r311, %r931;
or.b32 %r1218, %r1217, -2139062144;
and.b32 %r1219, %r311, 2139062143;
sub.s32 %r1220, %r1218, %r1219;
xor.b32 %r314, %r957, %r1220;
ld.global.nc.u32 %r315, [%rd67+24];

dp4a.s32.s32 %r313, %r314, %r315, %r307;

mov.u32 %r317, %r807;

prmt.b32 %r317, %r317, 0, 0xba98;

xor.b32 %r1221, %r960, %r317;
or.b32 %r1222, %r1221, -2139062144;
and.b32 %r1223, %r317, 2139062143;
sub.s32 %r1224, %r1222, %r1223;
xor.b32 %r320, %r979, %r1224;
ld.global.nc.u32 %r321, [%rd67+28];

dp4a.s32.s32 %r319, %r320, %r321, %r313;

mov.u32 %r323, %r813;

prmt.b32 %r323, %r323, 0, 0xba98;

xor.b32 %r1225, %r323, %r961;
or.b32 %r1226, %r1225, -2139062144;
and.b32 %r1227, %r323, 2139062143;
sub.s32 %r1228, %r1226, %r1227;
xor.b32 %r326, %r987, %r1228;
ld.global.nc.u32 %r327, [%rd67+32];

dp4a.s32.s32 %r325, %r326, %r327, %r319;

mul.lo.s32 %r1229, %r325, %r990;
shr.s32 %r1230, %r1229, 31;
shr.u32 %r1231, %r1230, 29;
add.s32 %r1232, %r1229, %r1231;
shr.s32 %r1233, %r1232, 3;

{ cvt.f32.f16 %f105, %rs1;}


ld.global.nc.u32 %r329, [%rd67];

{.reg .f16 low,high;
mov.b32 {low,high},%r329;
cvt.f32.f16 %f106, low;}


mul.ftz.f32 %f137, %f105, %f106;
cvt.rn.f32.s32 %f138, %r1233;
fma.rn.ftz.f32 %f364, %f137, %f138, %f364;
mov.u32 %r330, %r820;

prmt.b32 %r330, %r330, 0, 0xba98;

xor.b32 %r1234, %r999, %r330;
or.b32 %r1235, %r1234, -2139062144;
and.b32 %r1236, %r330, 2139062143;
sub.s32 %r1237, %r1235, %r1236;
xor.b32 %r333, %r1017, %r1237;

dp4a.s32.s32 %r332, %r333, %r285, %r825;

mov.u32 %r336, %r826;

prmt.b32 %r336, %r336, 0, 0xba98;

xor.b32 %r1238, %r336, %r1000;
or.b32 %r1239, %r1238, -2139062144;
and.b32 %r1240, %r336, 2139062143;
sub.s32 %r1241, %r1239, %r1240;
xor.b32 %r339, %r1025, %r1241;

dp4a.s32.s32 %r338, %r339, %r291, %r332;

mov.u32 %r342, %r832;

prmt.b32 %r342, %r342, 0, 0xba98;

xor.b32 %r1242, %r1028, %r342;
or.b32 %r1243, %r1242, -2139062144;
and.b32 %r1244, %r342, 2139062143;
sub.s32 %r1245, %r1243, %r1244;
xor.b32 %r345, %r1047, %r1245;

dp4a.s32.s32 %r344, %r345, %r297, %r338;

mov.u32 %r348, %r838;

prmt.b32 %r348, %r348, 0, 0xba98;

xor.b32 %r1246, %r348, %r1029;
or.b32 %r1247, %r1246, -2139062144;
and.b32 %r1248, %r348, 2139062143;
sub.s32 %r1249, %r1247, %r1248;
xor.b32 %r351, %r1055, %r1249;

dp4a.s32.s32 %r350, %r351, %r303, %r344;

mov.u32 %r354, %r844;

prmt.b32 %r354, %r354, 0, 0xba98;

xor.b32 %r1250, %r1057, %r354;
or.b32 %r1251, %r1250, -2139062144;
and.b32 %r1252, %r354, 2139062143;
sub.s32 %r1253, %r1251, %r1252;
xor.b32 %r357, %r1076, %r1253;

dp4a.s32.s32 %r356, %r357, %r309, %r350;

mov.u32 %r360, %r850;

prmt.b32 %r360, %r360, 0, 0xba98;

xor.b32 %r1254, %r360, %r1058;
or.b32 %r1255, %r1254, -2139062144;
and.b32 %r1256, %r360, 2139062143;
sub.s32 %r1257, %r1255, %r1256;
xor.b32 %r363, %r1084, %r1257;

dp4a.s32.s32 %r362, %r363, %r315, %r356;

mov.u32 %r366, %r856;

prmt.b32 %r366, %r366, 0, 0xba98;

xor.b32 %r1258, %r1087, %r366;
or.b32 %r1259, %r1258, -2139062144;
and.b32 %r1260, %r366, 2139062143;
sub.s32 %r1261, %r1259, %r1260;
xor.b32 %r369, %r1106, %r1261;

dp4a.s32.s32 %r368, %r369, %r321, %r362;

mov.u32 %r372, %r862;

prmt.b32 %r372, %r372, 0, 0xba98;

xor.b32 %r1262, %r372, %r1088;
or.b32 %r1263, %r1262, -2139062144;
and.b32 %r1264, %r372, 2139062143;
sub.s32 %r1265, %r1263, %r1264;
xor.b32 %r375, %r1114, %r1265;

dp4a.s32.s32 %r374, %r375, %r327, %r368;

mul.lo.s32 %r1266, %r374, %r1117;
shr.s32 %r1267, %r1266, 31;
shr.u32 %r1268, %r1267, 29;
add.s32 %r1269, %r1266, %r1268;
shr.s32 %r1270, %r1269, 3;

{ cvt.f32.f16 %f107, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r329;
cvt.f32.f16 %f108, low;}


mul.ftz.f32 %f139, %f107, %f108;
cvt.rn.f32.s32 %f140, %r1270;
fma.rn.ftz.f32 %f363, %f139, %f140, %f363;
cvt.u64.u32 %rd68, %r1941;
add.s64 %rd69, %rd1, %rd68;
add.s64 %rd70, %rd69, %rd2;
mul.lo.s64 %rd71, %rd70, 36;
mov.u32 %r379, %r771;

prmt.b32 %r379, %r379, 0, 0xba98;

xor.b32 %r1271, %r872, %r379;
or.b32 %r1272, %r1271, -2139062144;
and.b32 %r1273, %r379, 2139062143;
sub.s32 %r1274, %r1272, %r1273;
xor.b32 %r382, %r890, %r1274;
add.s64 %rd72, %rd7, %rd71;
ld.global.nc.u32 %r383, [%rd72+4];

dp4a.s32.s32 %r381, %r382, %r383, %r825;

mov.u32 %r385, %r777;

prmt.b32 %r385, %r385, 0, 0xba98;

xor.b32 %r1275, %r385, %r873;
or.b32 %r1276, %r1275, -2139062144;
and.b32 %r1277, %r385, 2139062143;
sub.s32 %r1278, %r1276, %r1277;
xor.b32 %r388, %r898, %r1278;
ld.global.nc.u32 %r389, [%rd72+8];

dp4a.s32.s32 %r387, %r388, %r389, %r381;

mov.u32 %r391, %r783;

prmt.b32 %r391, %r391, 0, 0xba98;

xor.b32 %r1279, %r901, %r391;
or.b32 %r1280, %r1279, -2139062144;
and.b32 %r1281, %r391, 2139062143;
sub.s32 %r1282, %r1280, %r1281;
xor.b32 %r394, %r920, %r1282;
ld.global.nc.u32 %r395, [%rd72+12];

dp4a.s32.s32 %r393, %r394, %r395, %r387;

mov.u32 %r397, %r789;

prmt.b32 %r397, %r397, 0, 0xba98;

xor.b32 %r1283, %r397, %r902;
or.b32 %r1284, %r1283, -2139062144;
and.b32 %r1285, %r397, 2139062143;
sub.s32 %r1286, %r1284, %r1285;
xor.b32 %r400, %r928, %r1286;
ld.global.nc.u32 %r401, [%rd72+16];

dp4a.s32.s32 %r399, %r400, %r401, %r393;

mov.u32 %r403, %r795;

prmt.b32 %r403, %r403, 0, 0xba98;

xor.b32 %r1287, %r930, %r403;
or.b32 %r1288, %r1287, -2139062144;
and.b32 %r1289, %r403, 2139062143;
sub.s32 %r1290, %r1288, %r1289;
xor.b32 %r406, %r949, %r1290;
ld.global.nc.u32 %r407, [%rd72+20];

dp4a.s32.s32 %r405, %r406, %r407, %r399;

mov.u32 %r409, %r801;

prmt.b32 %r409, %r409, 0, 0xba98;

xor.b32 %r1291, %r409, %r931;
or.b32 %r1292, %r1291, -2139062144;
and.b32 %r1293, %r409, 2139062143;
sub.s32 %r1294, %r1292, %r1293;
xor.b32 %r412, %r957, %r1294;
ld.global.nc.u32 %r413, [%rd72+24];

dp4a.s32.s32 %r411, %r412, %r413, %r405;

mov.u32 %r415, %r807;

prmt.b32 %r415, %r415, 0, 0xba98;

xor.b32 %r1295, %r960, %r415;
or.b32 %r1296, %r1295, -2139062144;
and.b32 %r1297, %r415, 2139062143;
sub.s32 %r1298, %r1296, %r1297;
xor.b32 %r418, %r979, %r1298;
ld.global.nc.u32 %r419, [%rd72+28];

dp4a.s32.s32 %r417, %r418, %r419, %r411;

mov.u32 %r421, %r813;

prmt.b32 %r421, %r421, 0, 0xba98;

xor.b32 %r1299, %r421, %r961;
or.b32 %r1300, %r1299, -2139062144;
and.b32 %r1301, %r421, 2139062143;
sub.s32 %r1302, %r1300, %r1301;
xor.b32 %r424, %r987, %r1302;
ld.global.nc.u32 %r425, [%rd72+32];

dp4a.s32.s32 %r423, %r424, %r425, %r417;

mul.lo.s32 %r1303, %r423, %r990;
shr.s32 %r1304, %r1303, 31;
shr.u32 %r1305, %r1304, 29;
add.s32 %r1306, %r1303, %r1305;
shr.s32 %r1307, %r1306, 3;

{ cvt.f32.f16 %f109, %rs1;}


ld.global.nc.u32 %r427, [%rd72];

{.reg .f16 low,high;
mov.b32 {low,high},%r427;
cvt.f32.f16 %f110, low;}


mul.ftz.f32 %f141, %f109, %f110;
cvt.rn.f32.s32 %f142, %r1307;
fma.rn.ftz.f32 %f362, %f141, %f142, %f362;
mov.u32 %r428, %r820;

prmt.b32 %r428, %r428, 0, 0xba98;

xor.b32 %r1308, %r999, %r428;
or.b32 %r1309, %r1308, -2139062144;
and.b32 %r1310, %r428, 2139062143;
sub.s32 %r1311, %r1309, %r1310;
xor.b32 %r431, %r1017, %r1311;

dp4a.s32.s32 %r430, %r431, %r383, %r825;

mov.u32 %r434, %r826;

prmt.b32 %r434, %r434, 0, 0xba98;

xor.b32 %r1312, %r434, %r1000;
or.b32 %r1313, %r1312, -2139062144;
and.b32 %r1314, %r434, 2139062143;
sub.s32 %r1315, %r1313, %r1314;
xor.b32 %r437, %r1025, %r1315;

dp4a.s32.s32 %r436, %r437, %r389, %r430;

mov.u32 %r440, %r832;

prmt.b32 %r440, %r440, 0, 0xba98;

xor.b32 %r1316, %r1028, %r440;
or.b32 %r1317, %r1316, -2139062144;
and.b32 %r1318, %r440, 2139062143;
sub.s32 %r1319, %r1317, %r1318;
xor.b32 %r443, %r1047, %r1319;

dp4a.s32.s32 %r442, %r443, %r395, %r436;

mov.u32 %r446, %r838;

prmt.b32 %r446, %r446, 0, 0xba98;

xor.b32 %r1320, %r446, %r1029;
or.b32 %r1321, %r1320, -2139062144;
and.b32 %r1322, %r446, 2139062143;
sub.s32 %r1323, %r1321, %r1322;
xor.b32 %r449, %r1055, %r1323;

dp4a.s32.s32 %r448, %r449, %r401, %r442;

mov.u32 %r452, %r844;

prmt.b32 %r452, %r452, 0, 0xba98;

xor.b32 %r1324, %r1057, %r452;
or.b32 %r1325, %r1324, -2139062144;
and.b32 %r1326, %r452, 2139062143;
sub.s32 %r1327, %r1325, %r1326;
xor.b32 %r455, %r1076, %r1327;

dp4a.s32.s32 %r454, %r455, %r407, %r448;

mov.u32 %r458, %r850;

prmt.b32 %r458, %r458, 0, 0xba98;

xor.b32 %r1328, %r458, %r1058;
or.b32 %r1329, %r1328, -2139062144;
and.b32 %r1330, %r458, 2139062143;
sub.s32 %r1331, %r1329, %r1330;
xor.b32 %r461, %r1084, %r1331;

dp4a.s32.s32 %r460, %r461, %r413, %r454;

mov.u32 %r464, %r856;

prmt.b32 %r464, %r464, 0, 0xba98;

xor.b32 %r1332, %r1087, %r464;
or.b32 %r1333, %r1332, -2139062144;
and.b32 %r1334, %r464, 2139062143;
sub.s32 %r1335, %r1333, %r1334;
xor.b32 %r467, %r1106, %r1335;

dp4a.s32.s32 %r466, %r467, %r419, %r460;

mov.u32 %r470, %r862;

prmt.b32 %r470, %r470, 0, 0xba98;

xor.b32 %r1336, %r470, %r1088;
or.b32 %r1337, %r1336, -2139062144;
and.b32 %r1338, %r470, 2139062143;
sub.s32 %r1339, %r1337, %r1338;
xor.b32 %r473, %r1114, %r1339;

dp4a.s32.s32 %r472, %r473, %r425, %r466;

mul.lo.s32 %r1340, %r472, %r1117;
shr.s32 %r1341, %r1340, 31;
shr.u32 %r1342, %r1341, 29;
add.s32 %r1343, %r1340, %r1342;
shr.s32 %r1344, %r1343, 3;

{ cvt.f32.f16 %f111, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r427;
cvt.f32.f16 %f112, low;}


mul.ftz.f32 %f143, %f111, %f112;
cvt.rn.f32.s32 %f144, %r1344;
fma.rn.ftz.f32 %f361, %f143, %f144, %f361;
cvt.u64.u32 %rd73, %r1940;
add.s64 %rd74, %rd1, %rd73;
add.s64 %rd75, %rd74, %rd2;
mul.lo.s64 %rd76, %rd75, 36;
mov.u32 %r477, %r771;

prmt.b32 %r477, %r477, 0, 0xba98;

xor.b32 %r1345, %r872, %r477;
or.b32 %r1346, %r1345, -2139062144;
and.b32 %r1347, %r477, 2139062143;
sub.s32 %r1348, %r1346, %r1347;
xor.b32 %r480, %r890, %r1348;
add.s64 %rd77, %rd7, %rd76;
ld.global.nc.u32 %r481, [%rd77+4];

dp4a.s32.s32 %r479, %r480, %r481, %r825;

mov.u32 %r483, %r777;

prmt.b32 %r483, %r483, 0, 0xba98;

xor.b32 %r1349, %r483, %r873;
or.b32 %r1350, %r1349, -2139062144;
and.b32 %r1351, %r483, 2139062143;
sub.s32 %r1352, %r1350, %r1351;
xor.b32 %r486, %r898, %r1352;
ld.global.nc.u32 %r487, [%rd77+8];

dp4a.s32.s32 %r485, %r486, %r487, %r479;

mov.u32 %r489, %r783;

prmt.b32 %r489, %r489, 0, 0xba98;

xor.b32 %r1353, %r901, %r489;
or.b32 %r1354, %r1353, -2139062144;
and.b32 %r1355, %r489, 2139062143;
sub.s32 %r1356, %r1354, %r1355;
xor.b32 %r492, %r920, %r1356;
ld.global.nc.u32 %r493, [%rd77+12];

dp4a.s32.s32 %r491, %r492, %r493, %r485;

mov.u32 %r495, %r789;

prmt.b32 %r495, %r495, 0, 0xba98;

xor.b32 %r1357, %r495, %r902;
or.b32 %r1358, %r1357, -2139062144;
and.b32 %r1359, %r495, 2139062143;
sub.s32 %r1360, %r1358, %r1359;
xor.b32 %r498, %r928, %r1360;
ld.global.nc.u32 %r499, [%rd77+16];

dp4a.s32.s32 %r497, %r498, %r499, %r491;

mov.u32 %r501, %r795;

prmt.b32 %r501, %r501, 0, 0xba98;

xor.b32 %r1361, %r930, %r501;
or.b32 %r1362, %r1361, -2139062144;
and.b32 %r1363, %r501, 2139062143;
sub.s32 %r1364, %r1362, %r1363;
xor.b32 %r504, %r949, %r1364;
ld.global.nc.u32 %r505, [%rd77+20];

dp4a.s32.s32 %r503, %r504, %r505, %r497;

mov.u32 %r507, %r801;

prmt.b32 %r507, %r507, 0, 0xba98;

xor.b32 %r1365, %r507, %r931;
or.b32 %r1366, %r1365, -2139062144;
and.b32 %r1367, %r507, 2139062143;
sub.s32 %r1368, %r1366, %r1367;
xor.b32 %r510, %r957, %r1368;
ld.global.nc.u32 %r511, [%rd77+24];

dp4a.s32.s32 %r509, %r510, %r511, %r503;

mov.u32 %r513, %r807;

prmt.b32 %r513, %r513, 0, 0xba98;

xor.b32 %r1369, %r960, %r513;
or.b32 %r1370, %r1369, -2139062144;
and.b32 %r1371, %r513, 2139062143;
sub.s32 %r1372, %r1370, %r1371;
xor.b32 %r516, %r979, %r1372;
ld.global.nc.u32 %r517, [%rd77+28];

dp4a.s32.s32 %r515, %r516, %r517, %r509;

mov.u32 %r519, %r813;

prmt.b32 %r519, %r519, 0, 0xba98;

xor.b32 %r1373, %r519, %r961;
or.b32 %r1374, %r1373, -2139062144;
and.b32 %r1375, %r519, 2139062143;
sub.s32 %r1376, %r1374, %r1375;
xor.b32 %r522, %r987, %r1376;
ld.global.nc.u32 %r523, [%rd77+32];

dp4a.s32.s32 %r521, %r522, %r523, %r515;

mul.lo.s32 %r1377, %r521, %r990;
shr.s32 %r1378, %r1377, 31;
shr.u32 %r1379, %r1378, 29;
add.s32 %r1380, %r1377, %r1379;
shr.s32 %r1381, %r1380, 3;

{ cvt.f32.f16 %f113, %rs1;}


ld.global.nc.u32 %r525, [%rd77];

{.reg .f16 low,high;
mov.b32 {low,high},%r525;
cvt.f32.f16 %f114, low;}


mul.ftz.f32 %f145, %f113, %f114;
cvt.rn.f32.s32 %f146, %r1381;
fma.rn.ftz.f32 %f360, %f145, %f146, %f360;
mov.u32 %r526, %r820;

prmt.b32 %r526, %r526, 0, 0xba98;

xor.b32 %r1382, %r999, %r526;
or.b32 %r1383, %r1382, -2139062144;
and.b32 %r1384, %r526, 2139062143;
sub.s32 %r1385, %r1383, %r1384;
xor.b32 %r529, %r1017, %r1385;

dp4a.s32.s32 %r528, %r529, %r481, %r825;

mov.u32 %r532, %r826;

prmt.b32 %r532, %r532, 0, 0xba98;

xor.b32 %r1386, %r532, %r1000;
or.b32 %r1387, %r1386, -2139062144;
and.b32 %r1388, %r532, 2139062143;
sub.s32 %r1389, %r1387, %r1388;
xor.b32 %r535, %r1025, %r1389;

dp4a.s32.s32 %r534, %r535, %r487, %r528;

mov.u32 %r538, %r832;

prmt.b32 %r538, %r538, 0, 0xba98;

xor.b32 %r1390, %r1028, %r538;
or.b32 %r1391, %r1390, -2139062144;
and.b32 %r1392, %r538, 2139062143;
sub.s32 %r1393, %r1391, %r1392;
xor.b32 %r541, %r1047, %r1393;

dp4a.s32.s32 %r540, %r541, %r493, %r534;

mov.u32 %r544, %r838;

prmt.b32 %r544, %r544, 0, 0xba98;

xor.b32 %r1394, %r544, %r1029;
or.b32 %r1395, %r1394, -2139062144;
and.b32 %r1396, %r544, 2139062143;
sub.s32 %r1397, %r1395, %r1396;
xor.b32 %r547, %r1055, %r1397;

dp4a.s32.s32 %r546, %r547, %r499, %r540;

mov.u32 %r550, %r844;

prmt.b32 %r550, %r550, 0, 0xba98;

xor.b32 %r1398, %r1057, %r550;
or.b32 %r1399, %r1398, -2139062144;
and.b32 %r1400, %r550, 2139062143;
sub.s32 %r1401, %r1399, %r1400;
xor.b32 %r553, %r1076, %r1401;

dp4a.s32.s32 %r552, %r553, %r505, %r546;

mov.u32 %r556, %r850;

prmt.b32 %r556, %r556, 0, 0xba98;

xor.b32 %r1402, %r556, %r1058;
or.b32 %r1403, %r1402, -2139062144;
and.b32 %r1404, %r556, 2139062143;
sub.s32 %r1405, %r1403, %r1404;
xor.b32 %r559, %r1084, %r1405;

dp4a.s32.s32 %r558, %r559, %r511, %r552;

mov.u32 %r562, %r856;

prmt.b32 %r562, %r562, 0, 0xba98;

xor.b32 %r1406, %r1087, %r562;
or.b32 %r1407, %r1406, -2139062144;
and.b32 %r1408, %r562, 2139062143;
sub.s32 %r1409, %r1407, %r1408;
xor.b32 %r565, %r1106, %r1409;

dp4a.s32.s32 %r564, %r565, %r517, %r558;

mov.u32 %r568, %r862;

prmt.b32 %r568, %r568, 0, 0xba98;

xor.b32 %r1410, %r568, %r1088;
or.b32 %r1411, %r1410, -2139062144;
and.b32 %r1412, %r568, 2139062143;
sub.s32 %r1413, %r1411, %r1412;
xor.b32 %r571, %r1114, %r1413;

dp4a.s32.s32 %r570, %r571, %r523, %r564;

mul.lo.s32 %r1414, %r570, %r1117;
shr.s32 %r1415, %r1414, 31;
shr.u32 %r1416, %r1415, 29;
add.s32 %r1417, %r1414, %r1416;
shr.s32 %r1418, %r1417, 3;

{ cvt.f32.f16 %f115, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r525;
cvt.f32.f16 %f116, low;}


mul.ftz.f32 %f147, %f115, %f116;
cvt.rn.f32.s32 %f148, %r1418;
fma.rn.ftz.f32 %f359, %f147, %f148, %f359;
cvt.u64.u32 %rd78, %r1939;
add.s64 %rd79, %rd1, %rd78;
add.s64 %rd80, %rd79, %rd2;
mul.lo.s64 %rd81, %rd80, 36;
mov.u32 %r575, %r771;

prmt.b32 %r575, %r575, 0, 0xba98;

xor.b32 %r1419, %r872, %r575;
or.b32 %r1420, %r1419, -2139062144;
and.b32 %r1421, %r575, 2139062143;
sub.s32 %r1422, %r1420, %r1421;
xor.b32 %r578, %r890, %r1422;
add.s64 %rd82, %rd7, %rd81;
ld.global.nc.u32 %r579, [%rd82+4];

dp4a.s32.s32 %r577, %r578, %r579, %r825;

mov.u32 %r581, %r777;

prmt.b32 %r581, %r581, 0, 0xba98;

xor.b32 %r1423, %r581, %r873;
or.b32 %r1424, %r1423, -2139062144;
and.b32 %r1425, %r581, 2139062143;
sub.s32 %r1426, %r1424, %r1425;
xor.b32 %r584, %r898, %r1426;
ld.global.nc.u32 %r585, [%rd82+8];

dp4a.s32.s32 %r583, %r584, %r585, %r577;

mov.u32 %r587, %r783;

prmt.b32 %r587, %r587, 0, 0xba98;

xor.b32 %r1427, %r901, %r587;
or.b32 %r1428, %r1427, -2139062144;
and.b32 %r1429, %r587, 2139062143;
sub.s32 %r1430, %r1428, %r1429;
xor.b32 %r590, %r920, %r1430;
ld.global.nc.u32 %r591, [%rd82+12];

dp4a.s32.s32 %r589, %r590, %r591, %r583;

mov.u32 %r593, %r789;

prmt.b32 %r593, %r593, 0, 0xba98;

xor.b32 %r1431, %r593, %r902;
or.b32 %r1432, %r1431, -2139062144;
and.b32 %r1433, %r593, 2139062143;
sub.s32 %r1434, %r1432, %r1433;
xor.b32 %r596, %r928, %r1434;
ld.global.nc.u32 %r597, [%rd82+16];

dp4a.s32.s32 %r595, %r596, %r597, %r589;

mov.u32 %r599, %r795;

prmt.b32 %r599, %r599, 0, 0xba98;

xor.b32 %r1435, %r930, %r599;
or.b32 %r1436, %r1435, -2139062144;
and.b32 %r1437, %r599, 2139062143;
sub.s32 %r1438, %r1436, %r1437;
xor.b32 %r602, %r949, %r1438;
ld.global.nc.u32 %r603, [%rd82+20];

dp4a.s32.s32 %r601, %r602, %r603, %r595;

mov.u32 %r605, %r801;

prmt.b32 %r605, %r605, 0, 0xba98;

xor.b32 %r1439, %r605, %r931;
or.b32 %r1440, %r1439, -2139062144;
and.b32 %r1441, %r605, 2139062143;
sub.s32 %r1442, %r1440, %r1441;
xor.b32 %r608, %r957, %r1442;
ld.global.nc.u32 %r609, [%rd82+24];

dp4a.s32.s32 %r607, %r608, %r609, %r601;

mov.u32 %r611, %r807;

prmt.b32 %r611, %r611, 0, 0xba98;

xor.b32 %r1443, %r960, %r611;
or.b32 %r1444, %r1443, -2139062144;
and.b32 %r1445, %r611, 2139062143;
sub.s32 %r1446, %r1444, %r1445;
xor.b32 %r614, %r979, %r1446;
ld.global.nc.u32 %r615, [%rd82+28];

dp4a.s32.s32 %r613, %r614, %r615, %r607;

mov.u32 %r617, %r813;

prmt.b32 %r617, %r617, 0, 0xba98;

xor.b32 %r1447, %r617, %r961;
or.b32 %r1448, %r1447, -2139062144;
and.b32 %r1449, %r617, 2139062143;
sub.s32 %r1450, %r1448, %r1449;
xor.b32 %r620, %r987, %r1450;
ld.global.nc.u32 %r621, [%rd82+32];

dp4a.s32.s32 %r619, %r620, %r621, %r613;

mul.lo.s32 %r1451, %r619, %r990;
shr.s32 %r1452, %r1451, 31;
shr.u32 %r1453, %r1452, 29;
add.s32 %r1454, %r1451, %r1453;
shr.s32 %r1455, %r1454, 3;

{ cvt.f32.f16 %f117, %rs1;}


ld.global.nc.u32 %r623, [%rd82];

{.reg .f16 low,high;
mov.b32 {low,high},%r623;
cvt.f32.f16 %f118, low;}


mul.ftz.f32 %f149, %f117, %f118;
cvt.rn.f32.s32 %f150, %r1455;
fma.rn.ftz.f32 %f358, %f149, %f150, %f358;
mov.u32 %r624, %r820;

prmt.b32 %r624, %r624, 0, 0xba98;

xor.b32 %r1456, %r999, %r624;
or.b32 %r1457, %r1456, -2139062144;
and.b32 %r1458, %r624, 2139062143;
sub.s32 %r1459, %r1457, %r1458;
xor.b32 %r627, %r1017, %r1459;

dp4a.s32.s32 %r626, %r627, %r579, %r825;

mov.u32 %r630, %r826;

prmt.b32 %r630, %r630, 0, 0xba98;

xor.b32 %r1460, %r630, %r1000;
or.b32 %r1461, %r1460, -2139062144;
and.b32 %r1462, %r630, 2139062143;
sub.s32 %r1463, %r1461, %r1462;
xor.b32 %r633, %r1025, %r1463;

dp4a.s32.s32 %r632, %r633, %r585, %r626;

mov.u32 %r636, %r832;

prmt.b32 %r636, %r636, 0, 0xba98;

xor.b32 %r1464, %r1028, %r636;
or.b32 %r1465, %r1464, -2139062144;
and.b32 %r1466, %r636, 2139062143;
sub.s32 %r1467, %r1465, %r1466;
xor.b32 %r639, %r1047, %r1467;

dp4a.s32.s32 %r638, %r639, %r591, %r632;

mov.u32 %r642, %r838;

prmt.b32 %r642, %r642, 0, 0xba98;

xor.b32 %r1468, %r642, %r1029;
or.b32 %r1469, %r1468, -2139062144;
and.b32 %r1470, %r642, 2139062143;
sub.s32 %r1471, %r1469, %r1470;
xor.b32 %r645, %r1055, %r1471;

dp4a.s32.s32 %r644, %r645, %r597, %r638;

mov.u32 %r648, %r844;

prmt.b32 %r648, %r648, 0, 0xba98;

xor.b32 %r1472, %r1057, %r648;
or.b32 %r1473, %r1472, -2139062144;
and.b32 %r1474, %r648, 2139062143;
sub.s32 %r1475, %r1473, %r1474;
xor.b32 %r651, %r1076, %r1475;

dp4a.s32.s32 %r650, %r651, %r603, %r644;

mov.u32 %r654, %r850;

prmt.b32 %r654, %r654, 0, 0xba98;

xor.b32 %r1476, %r654, %r1058;
or.b32 %r1477, %r1476, -2139062144;
and.b32 %r1478, %r654, 2139062143;
sub.s32 %r1479, %r1477, %r1478;
xor.b32 %r657, %r1084, %r1479;

dp4a.s32.s32 %r656, %r657, %r609, %r650;

mov.u32 %r660, %r856;

prmt.b32 %r660, %r660, 0, 0xba98;

xor.b32 %r1480, %r1087, %r660;
or.b32 %r1481, %r1480, -2139062144;
and.b32 %r1482, %r660, 2139062143;
sub.s32 %r1483, %r1481, %r1482;
xor.b32 %r663, %r1106, %r1483;

dp4a.s32.s32 %r662, %r663, %r615, %r656;

mov.u32 %r666, %r862;

prmt.b32 %r666, %r666, 0, 0xba98;

xor.b32 %r1484, %r666, %r1088;
or.b32 %r1485, %r1484, -2139062144;
and.b32 %r1486, %r666, 2139062143;
sub.s32 %r1487, %r1485, %r1486;
xor.b32 %r669, %r1114, %r1487;

dp4a.s32.s32 %r668, %r669, %r621, %r662;

mul.lo.s32 %r1488, %r668, %r1117;
shr.s32 %r1489, %r1488, 31;
shr.u32 %r1490, %r1489, 29;
add.s32 %r1491, %r1488, %r1490;
shr.s32 %r1492, %r1491, 3;

{ cvt.f32.f16 %f119, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r623;
cvt.f32.f16 %f120, low;}


mul.ftz.f32 %f151, %f119, %f120;
cvt.rn.f32.s32 %f152, %r1492;
fma.rn.ftz.f32 %f357, %f151, %f152, %f357;
cvt.u64.u32 %rd83, %r1938;
add.s64 %rd84, %rd1, %rd83;
add.s64 %rd85, %rd84, %rd2;
mul.lo.s64 %rd86, %rd85, 36;
mov.u32 %r673, %r771;

prmt.b32 %r673, %r673, 0, 0xba98;

xor.b32 %r1493, %r872, %r673;
or.b32 %r1494, %r1493, -2139062144;
and.b32 %r1495, %r673, 2139062143;
sub.s32 %r1496, %r1494, %r1495;
xor.b32 %r676, %r890, %r1496;
add.s64 %rd87, %rd7, %rd86;
ld.global.nc.u32 %r677, [%rd87+4];

dp4a.s32.s32 %r675, %r676, %r677, %r825;

mov.u32 %r679, %r777;

prmt.b32 %r679, %r679, 0, 0xba98;

xor.b32 %r1497, %r679, %r873;
or.b32 %r1498, %r1497, -2139062144;
and.b32 %r1499, %r679, 2139062143;
sub.s32 %r1500, %r1498, %r1499;
xor.b32 %r682, %r898, %r1500;
ld.global.nc.u32 %r683, [%rd87+8];

dp4a.s32.s32 %r681, %r682, %r683, %r675;

mov.u32 %r685, %r783;

prmt.b32 %r685, %r685, 0, 0xba98;

xor.b32 %r1501, %r901, %r685;
or.b32 %r1502, %r1501, -2139062144;
and.b32 %r1503, %r685, 2139062143;
sub.s32 %r1504, %r1502, %r1503;
xor.b32 %r688, %r920, %r1504;
ld.global.nc.u32 %r689, [%rd87+12];

dp4a.s32.s32 %r687, %r688, %r689, %r681;

mov.u32 %r691, %r789;

prmt.b32 %r691, %r691, 0, 0xba98;

xor.b32 %r1505, %r691, %r902;
or.b32 %r1506, %r1505, -2139062144;
and.b32 %r1507, %r691, 2139062143;
sub.s32 %r1508, %r1506, %r1507;
xor.b32 %r694, %r928, %r1508;
ld.global.nc.u32 %r695, [%rd87+16];

dp4a.s32.s32 %r693, %r694, %r695, %r687;

mov.u32 %r697, %r795;

prmt.b32 %r697, %r697, 0, 0xba98;

xor.b32 %r1509, %r930, %r697;
or.b32 %r1510, %r1509, -2139062144;
and.b32 %r1511, %r697, 2139062143;
sub.s32 %r1512, %r1510, %r1511;
xor.b32 %r700, %r949, %r1512;
ld.global.nc.u32 %r701, [%rd87+20];

dp4a.s32.s32 %r699, %r700, %r701, %r693;

mov.u32 %r703, %r801;

prmt.b32 %r703, %r703, 0, 0xba98;

xor.b32 %r1513, %r703, %r931;
or.b32 %r1514, %r1513, -2139062144;
and.b32 %r1515, %r703, 2139062143;
sub.s32 %r1516, %r1514, %r1515;
xor.b32 %r706, %r957, %r1516;
ld.global.nc.u32 %r707, [%rd87+24];

dp4a.s32.s32 %r705, %r706, %r707, %r699;

mov.u32 %r709, %r807;

prmt.b32 %r709, %r709, 0, 0xba98;

xor.b32 %r1517, %r960, %r709;
or.b32 %r1518, %r1517, -2139062144;
and.b32 %r1519, %r709, 2139062143;
sub.s32 %r1520, %r1518, %r1519;
xor.b32 %r712, %r979, %r1520;
ld.global.nc.u32 %r713, [%rd87+28];

dp4a.s32.s32 %r711, %r712, %r713, %r705;

mov.u32 %r715, %r813;

prmt.b32 %r715, %r715, 0, 0xba98;

xor.b32 %r1521, %r715, %r961;
or.b32 %r1522, %r1521, -2139062144;
and.b32 %r1523, %r715, 2139062143;
sub.s32 %r1524, %r1522, %r1523;
xor.b32 %r718, %r987, %r1524;
ld.global.nc.u32 %r719, [%rd87+32];

dp4a.s32.s32 %r717, %r718, %r719, %r711;

mul.lo.s32 %r1525, %r717, %r990;
shr.s32 %r1526, %r1525, 31;
shr.u32 %r1527, %r1526, 29;
add.s32 %r1528, %r1525, %r1527;
shr.s32 %r1529, %r1528, 3;

{ cvt.f32.f16 %f121, %rs1;}


ld.global.nc.u32 %r721, [%rd87];

{.reg .f16 low,high;
mov.b32 {low,high},%r721;
cvt.f32.f16 %f122, low;}


mul.ftz.f32 %f153, %f121, %f122;
cvt.rn.f32.s32 %f154, %r1529;
fma.rn.ftz.f32 %f356, %f153, %f154, %f356;
mov.u32 %r722, %r820;

prmt.b32 %r722, %r722, 0, 0xba98;

xor.b32 %r1530, %r999, %r722;
or.b32 %r1531, %r1530, -2139062144;
and.b32 %r1532, %r722, 2139062143;
sub.s32 %r1533, %r1531, %r1532;
xor.b32 %r725, %r1017, %r1533;

dp4a.s32.s32 %r724, %r725, %r677, %r825;

mov.u32 %r728, %r826;

prmt.b32 %r728, %r728, 0, 0xba98;

xor.b32 %r1534, %r728, %r1000;
or.b32 %r1535, %r1534, -2139062144;
and.b32 %r1536, %r728, 2139062143;
sub.s32 %r1537, %r1535, %r1536;
xor.b32 %r731, %r1025, %r1537;

dp4a.s32.s32 %r730, %r731, %r683, %r724;

mov.u32 %r734, %r832;

prmt.b32 %r734, %r734, 0, 0xba98;

xor.b32 %r1538, %r1028, %r734;
or.b32 %r1539, %r1538, -2139062144;
and.b32 %r1540, %r734, 2139062143;
sub.s32 %r1541, %r1539, %r1540;
xor.b32 %r737, %r1047, %r1541;

dp4a.s32.s32 %r736, %r737, %r689, %r730;

mov.u32 %r740, %r838;

prmt.b32 %r740, %r740, 0, 0xba98;

xor.b32 %r1542, %r740, %r1029;
or.b32 %r1543, %r1542, -2139062144;
and.b32 %r1544, %r740, 2139062143;
sub.s32 %r1545, %r1543, %r1544;
xor.b32 %r743, %r1055, %r1545;

dp4a.s32.s32 %r742, %r743, %r695, %r736;

mov.u32 %r746, %r844;

prmt.b32 %r746, %r746, 0, 0xba98;

xor.b32 %r1546, %r1057, %r746;
or.b32 %r1547, %r1546, -2139062144;
and.b32 %r1548, %r746, 2139062143;
sub.s32 %r1549, %r1547, %r1548;
xor.b32 %r749, %r1076, %r1549;

dp4a.s32.s32 %r748, %r749, %r701, %r742;

mov.u32 %r752, %r850;

prmt.b32 %r752, %r752, 0, 0xba98;

xor.b32 %r1550, %r752, %r1058;
or.b32 %r1551, %r1550, -2139062144;
and.b32 %r1552, %r752, 2139062143;
sub.s32 %r1553, %r1551, %r1552;
xor.b32 %r755, %r1084, %r1553;

dp4a.s32.s32 %r754, %r755, %r707, %r748;

mov.u32 %r758, %r856;

prmt.b32 %r758, %r758, 0, 0xba98;

xor.b32 %r1554, %r1087, %r758;
or.b32 %r1555, %r1554, -2139062144;
and.b32 %r1556, %r758, 2139062143;
sub.s32 %r1557, %r1555, %r1556;
xor.b32 %r761, %r1106, %r1557;

dp4a.s32.s32 %r760, %r761, %r713, %r754;

mov.u32 %r764, %r862;

prmt.b32 %r764, %r764, 0, 0xba98;

xor.b32 %r1558, %r764, %r1088;
or.b32 %r1559, %r1558, -2139062144;
and.b32 %r1560, %r764, 2139062143;
sub.s32 %r1561, %r1559, %r1560;
xor.b32 %r767, %r1114, %r1561;

dp4a.s32.s32 %r766, %r767, %r719, %r760;

mul.lo.s32 %r1562, %r766, %r1117;
shr.s32 %r1563, %r1562, 31;
shr.u32 %r1564, %r1563, 29;
add.s32 %r1565, %r1562, %r1564;
shr.s32 %r1566, %r1565, 3;

{ cvt.f32.f16 %f123, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r721;
cvt.f32.f16 %f124, low;}


mul.ftz.f32 %f155, %f123, %f124;
cvt.rn.f32.s32 %f156, %r1566;
fma.rn.ftz.f32 %f355, %f155, %f156, %f355;
cvt.u64.u32 %rd88, %r1937;
add.s64 %rd89, %rd1, %rd88;
add.s64 %rd90, %rd89, %rd2;
mul.lo.s64 %rd91, %rd90, 36;

prmt.b32 %r771, %r771, 0, 0xba98;

xor.b32 %r1567, %r872, %r771;
or.b32 %r1568, %r1567, -2139062144;
and.b32 %r1569, %r771, 2139062143;
sub.s32 %r1570, %r1568, %r1569;
xor.b32 %r774, %r890, %r1570;
add.s64 %rd92, %rd7, %rd91;
ld.global.nc.u32 %r775, [%rd92+4];

dp4a.s32.s32 %r773, %r774, %r775, %r825;


prmt.b32 %r777, %r777, 0, 0xba98;

xor.b32 %r1571, %r777, %r873;
or.b32 %r1572, %r1571, -2139062144;
and.b32 %r1573, %r777, 2139062143;
sub.s32 %r1574, %r1572, %r1573;
xor.b32 %r780, %r898, %r1574;
ld.global.nc.u32 %r781, [%rd92+8];

dp4a.s32.s32 %r779, %r780, %r781, %r773;


prmt.b32 %r783, %r783, 0, 0xba98;

xor.b32 %r1575, %r901, %r783;
or.b32 %r1576, %r1575, -2139062144;
and.b32 %r1577, %r783, 2139062143;
sub.s32 %r1578, %r1576, %r1577;
xor.b32 %r786, %r920, %r1578;
ld.global.nc.u32 %r787, [%rd92+12];

dp4a.s32.s32 %r785, %r786, %r787, %r779;


prmt.b32 %r789, %r789, 0, 0xba98;

xor.b32 %r1579, %r789, %r902;
or.b32 %r1580, %r1579, -2139062144;
and.b32 %r1581, %r789, 2139062143;
sub.s32 %r1582, %r1580, %r1581;
xor.b32 %r792, %r928, %r1582;
ld.global.nc.u32 %r793, [%rd92+16];

dp4a.s32.s32 %r791, %r792, %r793, %r785;


prmt.b32 %r795, %r795, 0, 0xba98;

xor.b32 %r1583, %r930, %r795;
or.b32 %r1584, %r1583, -2139062144;
and.b32 %r1585, %r795, 2139062143;
sub.s32 %r1586, %r1584, %r1585;
xor.b32 %r798, %r949, %r1586;
ld.global.nc.u32 %r799, [%rd92+20];

dp4a.s32.s32 %r797, %r798, %r799, %r791;


prmt.b32 %r801, %r801, 0, 0xba98;

xor.b32 %r1587, %r801, %r931;
or.b32 %r1588, %r1587, -2139062144;
and.b32 %r1589, %r801, 2139062143;
sub.s32 %r1590, %r1588, %r1589;
xor.b32 %r804, %r957, %r1590;
ld.global.nc.u32 %r805, [%rd92+24];

dp4a.s32.s32 %r803, %r804, %r805, %r797;


prmt.b32 %r807, %r807, 0, 0xba98;

xor.b32 %r1591, %r960, %r807;
or.b32 %r1592, %r1591, -2139062144;
and.b32 %r1593, %r807, 2139062143;
sub.s32 %r1594, %r1592, %r1593;
xor.b32 %r810, %r979, %r1594;
ld.global.nc.u32 %r811, [%rd92+28];

dp4a.s32.s32 %r809, %r810, %r811, %r803;


prmt.b32 %r813, %r813, 0, 0xba98;

xor.b32 %r1595, %r813, %r961;
or.b32 %r1596, %r1595, -2139062144;
and.b32 %r1597, %r813, 2139062143;
sub.s32 %r1598, %r1596, %r1597;
xor.b32 %r816, %r987, %r1598;
ld.global.nc.u32 %r817, [%rd92+32];

dp4a.s32.s32 %r815, %r816, %r817, %r809;

mul.lo.s32 %r1599, %r815, %r990;
shr.s32 %r1600, %r1599, 31;
shr.u32 %r1601, %r1600, 29;
add.s32 %r1602, %r1599, %r1601;
shr.s32 %r1603, %r1602, 3;

{ cvt.f32.f16 %f125, %rs1;}


ld.global.nc.u32 %r819, [%rd92];

{.reg .f16 low,high;
mov.b32 {low,high},%r819;
cvt.f32.f16 %f126, low;}


mul.ftz.f32 %f157, %f125, %f126;
cvt.rn.f32.s32 %f158, %r1603;
fma.rn.ftz.f32 %f354, %f157, %f158, %f354;

prmt.b32 %r820, %r820, 0, 0xba98;

xor.b32 %r1604, %r999, %r820;
or.b32 %r1605, %r1604, -2139062144;
and.b32 %r1606, %r820, 2139062143;
sub.s32 %r1607, %r1605, %r1606;
xor.b32 %r823, %r1017, %r1607;

dp4a.s32.s32 %r822, %r823, %r775, %r825;


prmt.b32 %r826, %r826, 0, 0xba98;

xor.b32 %r1608, %r826, %r1000;
or.b32 %r1609, %r1608, -2139062144;
and.b32 %r1610, %r826, 2139062143;
sub.s32 %r1611, %r1609, %r1610;
xor.b32 %r829, %r1025, %r1611;

dp4a.s32.s32 %r828, %r829, %r781, %r822;


prmt.b32 %r832, %r832, 0, 0xba98;

xor.b32 %r1612, %r1028, %r832;
or.b32 %r1613, %r1612, -2139062144;
and.b32 %r1614, %r832, 2139062143;
sub.s32 %r1615, %r1613, %r1614;
xor.b32 %r835, %r1047, %r1615;

dp4a.s32.s32 %r834, %r835, %r787, %r828;


prmt.b32 %r838, %r838, 0, 0xba98;

xor.b32 %r1616, %r838, %r1029;
or.b32 %r1617, %r1616, -2139062144;
and.b32 %r1618, %r838, 2139062143;
sub.s32 %r1619, %r1617, %r1618;
xor.b32 %r841, %r1055, %r1619;

dp4a.s32.s32 %r840, %r841, %r793, %r834;


prmt.b32 %r844, %r844, 0, 0xba98;

xor.b32 %r1620, %r1057, %r844;
or.b32 %r1621, %r1620, -2139062144;
and.b32 %r1622, %r844, 2139062143;
sub.s32 %r1623, %r1621, %r1622;
xor.b32 %r847, %r1076, %r1623;

dp4a.s32.s32 %r846, %r847, %r799, %r840;


prmt.b32 %r850, %r850, 0, 0xba98;

xor.b32 %r1624, %r850, %r1058;
or.b32 %r1625, %r1624, -2139062144;
and.b32 %r1626, %r850, 2139062143;
sub.s32 %r1627, %r1625, %r1626;
xor.b32 %r853, %r1084, %r1627;

dp4a.s32.s32 %r852, %r853, %r805, %r846;


prmt.b32 %r856, %r856, 0, 0xba98;

xor.b32 %r1628, %r1087, %r856;
or.b32 %r1629, %r1628, -2139062144;
and.b32 %r1630, %r856, 2139062143;
sub.s32 %r1631, %r1629, %r1630;
xor.b32 %r859, %r1106, %r1631;

dp4a.s32.s32 %r858, %r859, %r811, %r852;


prmt.b32 %r862, %r862, 0, 0xba98;

xor.b32 %r1632, %r862, %r1088;
or.b32 %r1633, %r1632, -2139062144;
and.b32 %r1634, %r862, 2139062143;
sub.s32 %r1635, %r1633, %r1634;
xor.b32 %r865, %r1114, %r1635;

dp4a.s32.s32 %r864, %r865, %r817, %r858;

mul.lo.s32 %r1636, %r864, %r1117;
shr.s32 %r1637, %r1636, 31;
shr.u32 %r1638, %r1637, 29;
add.s32 %r1639, %r1636, %r1638;
shr.s32 %r1640, %r1639, 3;

{ cvt.f32.f16 %f127, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r819;
cvt.f32.f16 %f128, low;}


mul.ftz.f32 %f159, %f127, %f128;
cvt.rn.f32.s32 %f160, %r1640;
fma.rn.ftz.f32 %f353, %f159, %f160, %f353;
add.s32 %r1942, %r1942, 64;
add.s32 %r1941, %r1941, 64;
add.s32 %r1940, %r1940, 64;
add.s32 %r1939, %r1939, 64;
add.s32 %r1938, %r1938, 64;
add.s32 %r1937, %r1937, 64;
add.s32 %r1936, %r1936, 64;
add.s32 %r1935, %r1935, 64;
add.s64 %rd149, %rd149, 528;
add.s64 %rd148, %rd148, 528;
add.s64 %rd147, %rd147, 528;
add.s32 %r1943, %r1943, 8;
setp.lt.s32 %p4, %r1943, %r6;
@%p4 bra $L__BB210_2;

$L__BB210_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB210_5;

shl.b32 %r1641, %r1, 11;
mov.u32 %r1642, _ZZ13mul_mat_vec_qIL9ggml_type16ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r1643, %r1642, %r1641;
shl.b32 %r1644, %r2, 2;
add.s32 %r1645, %r1643, %r1644;
st.shared.f32 [%r1645+-2048], %f368;
st.shared.f32 [%r1645+-1920], %f367;
st.shared.f32 [%r1645+-1792], %f366;
st.shared.f32 [%r1645+-1664], %f365;
st.shared.f32 [%r1645+-1536], %f364;
st.shared.f32 [%r1645+-1408], %f363;
st.shared.f32 [%r1645+-1280], %f362;
st.shared.f32 [%r1645+-1152], %f361;
st.shared.f32 [%r1645+-1024], %f360;
st.shared.f32 [%r1645+-896], %f359;
st.shared.f32 [%r1645+-768], %f358;
st.shared.f32 [%r1645+-640], %f357;
st.shared.f32 [%r1645+-512], %f356;
st.shared.f32 [%r1645+-384], %f355;
st.shared.f32 [%r1645+-256], %f354;
st.shared.f32 [%r1645+-128], %f353;

$L__BB210_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB210_38;

mov.u32 %r1646, %ctaid.z;
mov.u32 %r1647, %ctaid.y;
mad.lo.s32 %r1648, %r1647, %r45, %r4;
mad.lo.s32 %r1649, %r1646, %r51, %r1648;
cvt.u64.u32 %rd14, %r1649;
shl.b32 %r1650, %r2, 2;
mov.u32 %r1651, _ZZ13mul_mat_vec_qIL9ggml_type16ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r33, %r1651, %r1650;
mov.u32 %r1652, 2;
ld.shared.f32 %f161, [%r33];
add.ftz.f32 %f162, %f161, %f368;
mov.b32 %r1653, %f162;
mov.u32 %r1654, 31;
mov.u32 %r1655, 16;
mov.u32 %r1656, -1;
shfl.sync.bfly.b32 %r1657|%p7, %r1653, %r1655, %r1654, %r1656;
mov.b32 %f163, %r1657;
add.ftz.f32 %f164, %f162, %f163;
mov.b32 %r1658, %f164;
mov.u32 %r1659, 8;
shfl.sync.bfly.b32 %r1660|%p8, %r1658, %r1659, %r1654, %r1656;
mov.b32 %f165, %r1660;
add.ftz.f32 %f166, %f164, %f165;
mov.b32 %r1661, %f166;
mov.u32 %r1662, 4;
shfl.sync.bfly.b32 %r1663|%p9, %r1661, %r1662, %r1654, %r1656;
mov.b32 %f167, %r1663;
add.ftz.f32 %f168, %f166, %f167;
mov.b32 %r1664, %f168;
shfl.sync.bfly.b32 %r1665|%p10, %r1664, %r1652, %r1654, %r1656;
mov.b32 %f169, %r1665;
add.ftz.f32 %f170, %f168, %f169;
mov.b32 %r1666, %f170;
mov.u32 %r1667, 1;
shfl.sync.bfly.b32 %r1668|%p11, %r1666, %r1667, %r1654, %r1656;
mov.b32 %f171, %r1668;
add.ftz.f32 %f49, %f170, %f171;
setp.lt.u32 %p12, %r4, %r39;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
cvta.to.global.u64 %rd93, %rd21;
mul.wide.u32 %rd94, %r1649, 4;
add.s64 %rd15, %rd93, %rd94;
not.pred %p14, %p1;
@%p14 bra $L__BB210_8;

st.global.f32 [%rd15], %f49;

$L__BB210_8:
ld.shared.f32 %f172, [%r33+128];
add.ftz.f32 %f173, %f172, %f367;
mov.b32 %r1669, %f173;
shfl.sync.bfly.b32 %r1673|%p15, %r1669, %r1655, %r1654, %r1656;
mov.b32 %f174, %r1673;
add.ftz.f32 %f175, %f173, %f174;
mov.b32 %r1674, %f175;
shfl.sync.bfly.b32 %r1676|%p16, %r1674, %r1659, %r1654, %r1656;
mov.b32 %f176, %r1676;
add.ftz.f32 %f177, %f175, %f176;
mov.b32 %r1677, %f177;
shfl.sync.bfly.b32 %r1679|%p17, %r1677, %r1662, %r1654, %r1656;
mov.b32 %f178, %r1679;
add.ftz.f32 %f179, %f177, %f178;
mov.b32 %r1680, %f179;
shfl.sync.bfly.b32 %r1682|%p18, %r1680, %r1652, %r1654, %r1656;
mov.b32 %f180, %r1682;
add.ftz.f32 %f181, %f179, %f180;
mov.b32 %r1683, %f181;
shfl.sync.bfly.b32 %r1685|%p19, %r1683, %r1667, %r1654, %r1656;
mov.b32 %f182, %r1685;
add.ftz.f32 %f50, %f181, %f182;
setp.eq.s32 %p20, %r2, 1;
or.b32 %r1689, %r4, 1;
setp.lt.u32 %p21, %r1689, %r39;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB210_10;

st.global.f32 [%rd15+4], %f50;

$L__BB210_10:
ld.shared.f32 %f183, [%r33+256];
add.ftz.f32 %f184, %f183, %f366;
mov.b32 %r1690, %f184;
mov.u32 %r1691, 31;
mov.u32 %r1692, 16;
mov.u32 %r1693, -1;
shfl.sync.bfly.b32 %r1694|%p23, %r1690, %r1692, %r1691, %r1693;
mov.b32 %f185, %r1694;
add.ftz.f32 %f186, %f184, %f185;
mov.b32 %r1695, %f186;
mov.u32 %r1696, 8;
shfl.sync.bfly.b32 %r1697|%p24, %r1695, %r1696, %r1691, %r1693;
mov.b32 %f187, %r1697;
add.ftz.f32 %f188, %f186, %f187;
mov.b32 %r1698, %f188;
mov.u32 %r1699, 4;
shfl.sync.bfly.b32 %r1700|%p25, %r1698, %r1699, %r1691, %r1693;
mov.b32 %f189, %r1700;
add.ftz.f32 %f190, %f188, %f189;
mov.b32 %r1701, %f190;
mov.u32 %r1702, 2;
shfl.sync.bfly.b32 %r1703|%p26, %r1701, %r1702, %r1691, %r1693;
mov.b32 %f191, %r1703;
add.ftz.f32 %f192, %f190, %f191;
mov.b32 %r1704, %f192;
mov.u32 %r1705, 1;
shfl.sync.bfly.b32 %r1706|%p27, %r1704, %r1705, %r1691, %r1693;
mov.b32 %f193, %r1706;
add.ftz.f32 %f51, %f192, %f193;
@%p14 bra $L__BB210_12;

cvt.u64.u32 %rd95, %r39;
add.s64 %rd96, %rd95, %rd14;
shl.b64 %rd98, %rd96, 2;
add.s64 %rd99, %rd93, %rd98;
st.global.f32 [%rd99], %f51;

$L__BB210_12:
ld.shared.f32 %f194, [%r33+384];
add.ftz.f32 %f195, %f194, %f365;
mov.b32 %r1707, %f195;
shfl.sync.bfly.b32 %r1711|%p29, %r1707, %r1692, %r1691, %r1693;
mov.b32 %f196, %r1711;
add.ftz.f32 %f197, %f195, %f196;
mov.b32 %r1712, %f197;
shfl.sync.bfly.b32 %r1714|%p30, %r1712, %r1696, %r1691, %r1693;
mov.b32 %f198, %r1714;
add.ftz.f32 %f199, %f197, %f198;
mov.b32 %r1715, %f199;
shfl.sync.bfly.b32 %r1717|%p31, %r1715, %r1699, %r1691, %r1693;
mov.b32 %f200, %r1717;
add.ftz.f32 %f201, %f199, %f200;
mov.b32 %r1718, %f201;
shfl.sync.bfly.b32 %r1720|%p32, %r1718, %r1702, %r1691, %r1693;
mov.b32 %f202, %r1720;
add.ftz.f32 %f203, %f201, %f202;
mov.b32 %r1721, %f203;
shfl.sync.bfly.b32 %r1723|%p33, %r1721, %r1705, %r1691, %r1693;
mov.b32 %f204, %r1723;
add.ftz.f32 %f52, %f203, %f204;
@%p22 bra $L__BB210_14;

add.s32 %r1724, %r39, 1;
cvt.u64.u32 %rd100, %r1724;
add.s64 %rd101, %rd100, %rd14;
shl.b64 %rd103, %rd101, 2;
add.s64 %rd104, %rd93, %rd103;
st.global.f32 [%rd104], %f52;

$L__BB210_14:
shl.b32 %r1725, %r39, 1;
mov.u32 %r1726, 1;
ld.shared.f32 %f205, [%r33+512];
add.ftz.f32 %f206, %f205, %f364;
mov.b32 %r1727, %f206;
mov.u32 %r1728, 31;
mov.u32 %r1729, 16;
mov.u32 %r1730, -1;
shfl.sync.bfly.b32 %r1731|%p35, %r1727, %r1729, %r1728, %r1730;
mov.b32 %f207, %r1731;
add.ftz.f32 %f208, %f206, %f207;
mov.b32 %r1732, %f208;
mov.u32 %r1733, 8;
shfl.sync.bfly.b32 %r1734|%p36, %r1732, %r1733, %r1728, %r1730;
mov.b32 %f209, %r1734;
add.ftz.f32 %f210, %f208, %f209;
mov.b32 %r1735, %f210;
mov.u32 %r1736, 4;
shfl.sync.bfly.b32 %r1737|%p37, %r1735, %r1736, %r1728, %r1730;
mov.b32 %f211, %r1737;
add.ftz.f32 %f212, %f210, %f211;
mov.b32 %r1738, %f212;
mov.u32 %r1739, 2;
shfl.sync.bfly.b32 %r1740|%p38, %r1738, %r1739, %r1728, %r1730;
mov.b32 %f213, %r1740;
add.ftz.f32 %f214, %f212, %f213;
mov.b32 %r1741, %f214;
shfl.sync.bfly.b32 %r1742|%p39, %r1741, %r1726, %r1728, %r1730;
mov.b32 %f215, %r1742;
add.ftz.f32 %f53, %f214, %f215;
cvt.u64.u32 %rd105, %r1725;
add.s64 %rd106, %rd105, %rd14;
shl.b64 %rd108, %rd106, 2;
add.s64 %rd16, %rd93, %rd108;
@%p14 bra $L__BB210_16;

st.global.f32 [%rd16], %f53;

$L__BB210_16:
ld.shared.f32 %f216, [%r33+640];
add.ftz.f32 %f217, %f216, %f363;
mov.b32 %r1743, %f217;
shfl.sync.bfly.b32 %r1747|%p41, %r1743, %r1729, %r1728, %r1730;
mov.b32 %f218, %r1747;
add.ftz.f32 %f219, %f217, %f218;
mov.b32 %r1748, %f219;
shfl.sync.bfly.b32 %r1750|%p42, %r1748, %r1733, %r1728, %r1730;
mov.b32 %f220, %r1750;
add.ftz.f32 %f221, %f219, %f220;
mov.b32 %r1751, %f221;
shfl.sync.bfly.b32 %r1753|%p43, %r1751, %r1736, %r1728, %r1730;
mov.b32 %f222, %r1753;
add.ftz.f32 %f223, %f221, %f222;
mov.b32 %r1754, %f223;
shfl.sync.bfly.b32 %r1756|%p44, %r1754, %r1739, %r1728, %r1730;
mov.b32 %f224, %r1756;
add.ftz.f32 %f225, %f223, %f224;
mov.b32 %r1757, %f225;
shfl.sync.bfly.b32 %r1759|%p45, %r1757, %r1726, %r1728, %r1730;
mov.b32 %f226, %r1759;
add.ftz.f32 %f54, %f225, %f226;
@%p22 bra $L__BB210_18;

st.global.f32 [%rd16+4], %f54;

$L__BB210_18:
mul.lo.s32 %r34, %r39, 3;
ld.shared.f32 %f227, [%r33+768];
add.ftz.f32 %f228, %f227, %f362;
mov.b32 %r1760, %f228;
mov.u32 %r1761, 31;
mov.u32 %r1762, 16;
mov.u32 %r1763, -1;
shfl.sync.bfly.b32 %r1764|%p47, %r1760, %r1762, %r1761, %r1763;
mov.b32 %f229, %r1764;
add.ftz.f32 %f230, %f228, %f229;
mov.b32 %r1765, %f230;
mov.u32 %r1766, 8;
shfl.sync.bfly.b32 %r1767|%p48, %r1765, %r1766, %r1761, %r1763;
mov.b32 %f231, %r1767;
add.ftz.f32 %f232, %f230, %f231;
mov.b32 %r1768, %f232;
mov.u32 %r1769, 4;
shfl.sync.bfly.b32 %r1770|%p49, %r1768, %r1769, %r1761, %r1763;
mov.b32 %f233, %r1770;
add.ftz.f32 %f234, %f232, %f233;
mov.b32 %r1771, %f234;
mov.u32 %r1772, 2;
shfl.sync.bfly.b32 %r1773|%p50, %r1771, %r1772, %r1761, %r1763;
mov.b32 %f235, %r1773;
add.ftz.f32 %f236, %f234, %f235;
mov.b32 %r1774, %f236;
mov.u32 %r1775, 1;
shfl.sync.bfly.b32 %r1776|%p51, %r1774, %r1775, %r1761, %r1763;
mov.b32 %f237, %r1776;
add.ftz.f32 %f55, %f236, %f237;
@%p14 bra $L__BB210_20;

cvt.u64.u32 %rd109, %r34;
add.s64 %rd110, %rd109, %rd14;
shl.b64 %rd112, %rd110, 2;
add.s64 %rd113, %rd93, %rd112;
st.global.f32 [%rd113], %f55;

$L__BB210_20:
ld.shared.f32 %f238, [%r33+896];
add.ftz.f32 %f239, %f238, %f361;
mov.b32 %r1777, %f239;
shfl.sync.bfly.b32 %r1781|%p53, %r1777, %r1762, %r1761, %r1763;
mov.b32 %f240, %r1781;
add.ftz.f32 %f241, %f239, %f240;
mov.b32 %r1782, %f241;
shfl.sync.bfly.b32 %r1784|%p54, %r1782, %r1766, %r1761, %r1763;
mov.b32 %f242, %r1784;
add.ftz.f32 %f243, %f241, %f242;
mov.b32 %r1785, %f243;
shfl.sync.bfly.b32 %r1787|%p55, %r1785, %r1769, %r1761, %r1763;
mov.b32 %f244, %r1787;
add.ftz.f32 %f245, %f243, %f244;
mov.b32 %r1788, %f245;
shfl.sync.bfly.b32 %r1790|%p56, %r1788, %r1772, %r1761, %r1763;
mov.b32 %f246, %r1790;
add.ftz.f32 %f247, %f245, %f246;
mov.b32 %r1791, %f247;
shfl.sync.bfly.b32 %r1793|%p57, %r1791, %r1775, %r1761, %r1763;
mov.b32 %f248, %r1793;
add.ftz.f32 %f56, %f247, %f248;
@%p22 bra $L__BB210_22;

add.s32 %r1794, %r34, 1;
cvt.u64.u32 %rd114, %r1794;
add.s64 %rd115, %rd114, %rd14;
shl.b64 %rd117, %rd115, 2;
add.s64 %rd118, %rd93, %rd117;
st.global.f32 [%rd118], %f56;

$L__BB210_22:
shl.b32 %r1795, %r39, 2;
mov.u32 %r1796, 2;
ld.shared.f32 %f249, [%r33+1024];
add.ftz.f32 %f250, %f249, %f360;
mov.b32 %r1797, %f250;
mov.u32 %r1798, 31;
mov.u32 %r1799, 16;
mov.u32 %r1800, -1;
shfl.sync.bfly.b32 %r1801|%p59, %r1797, %r1799, %r1798, %r1800;
mov.b32 %f251, %r1801;
add.ftz.f32 %f252, %f250, %f251;
mov.b32 %r1802, %f252;
mov.u32 %r1803, 8;
shfl.sync.bfly.b32 %r1804|%p60, %r1802, %r1803, %r1798, %r1800;
mov.b32 %f253, %r1804;
add.ftz.f32 %f254, %f252, %f253;
mov.b32 %r1805, %f254;
mov.u32 %r1806, 4;
shfl.sync.bfly.b32 %r1807|%p61, %r1805, %r1806, %r1798, %r1800;
mov.b32 %f255, %r1807;
add.ftz.f32 %f256, %f254, %f255;
mov.b32 %r1808, %f256;
shfl.sync.bfly.b32 %r1809|%p62, %r1808, %r1796, %r1798, %r1800;
mov.b32 %f257, %r1809;
add.ftz.f32 %f258, %f256, %f257;
mov.b32 %r1810, %f258;
mov.u32 %r1811, 1;
shfl.sync.bfly.b32 %r1812|%p63, %r1810, %r1811, %r1798, %r1800;
mov.b32 %f259, %r1812;
add.ftz.f32 %f57, %f258, %f259;
cvt.u64.u32 %rd119, %r1795;
add.s64 %rd120, %rd119, %rd14;
shl.b64 %rd122, %rd120, 2;
add.s64 %rd17, %rd93, %rd122;
@%p14 bra $L__BB210_24;

st.global.f32 [%rd17], %f57;

$L__BB210_24:
ld.shared.f32 %f260, [%r33+1152];
add.ftz.f32 %f261, %f260, %f359;
mov.b32 %r1813, %f261;
shfl.sync.bfly.b32 %r1817|%p65, %r1813, %r1799, %r1798, %r1800;
mov.b32 %f262, %r1817;
add.ftz.f32 %f263, %f261, %f262;
mov.b32 %r1818, %f263;
shfl.sync.bfly.b32 %r1820|%p66, %r1818, %r1803, %r1798, %r1800;
mov.b32 %f264, %r1820;
add.ftz.f32 %f265, %f263, %f264;
mov.b32 %r1821, %f265;
shfl.sync.bfly.b32 %r1823|%p67, %r1821, %r1806, %r1798, %r1800;
mov.b32 %f266, %r1823;
add.ftz.f32 %f267, %f265, %f266;
mov.b32 %r1824, %f267;
shfl.sync.bfly.b32 %r1826|%p68, %r1824, %r1796, %r1798, %r1800;
mov.b32 %f268, %r1826;
add.ftz.f32 %f269, %f267, %f268;
mov.b32 %r1827, %f269;
shfl.sync.bfly.b32 %r1829|%p69, %r1827, %r1811, %r1798, %r1800;
mov.b32 %f270, %r1829;
add.ftz.f32 %f58, %f269, %f270;
@%p22 bra $L__BB210_26;

st.global.f32 [%rd17+4], %f58;

$L__BB210_26:
mul.lo.s32 %r35, %r39, 5;
ld.shared.f32 %f271, [%r33+1280];
add.ftz.f32 %f272, %f271, %f358;
mov.b32 %r1830, %f272;
mov.u32 %r1831, 31;
mov.u32 %r1832, 16;
mov.u32 %r1833, -1;
shfl.sync.bfly.b32 %r1834|%p71, %r1830, %r1832, %r1831, %r1833;
mov.b32 %f273, %r1834;
add.ftz.f32 %f274, %f272, %f273;
mov.b32 %r1835, %f274;
mov.u32 %r1836, 8;
shfl.sync.bfly.b32 %r1837|%p72, %r1835, %r1836, %r1831, %r1833;
mov.b32 %f275, %r1837;
add.ftz.f32 %f276, %f274, %f275;
mov.b32 %r1838, %f276;
mov.u32 %r1839, 4;
shfl.sync.bfly.b32 %r1840|%p73, %r1838, %r1839, %r1831, %r1833;
mov.b32 %f277, %r1840;
add.ftz.f32 %f278, %f276, %f277;
mov.b32 %r1841, %f278;
mov.u32 %r1842, 2;
shfl.sync.bfly.b32 %r1843|%p74, %r1841, %r1842, %r1831, %r1833;
mov.b32 %f279, %r1843;
add.ftz.f32 %f280, %f278, %f279;
mov.b32 %r1844, %f280;
mov.u32 %r1845, 1;
shfl.sync.bfly.b32 %r1846|%p75, %r1844, %r1845, %r1831, %r1833;
mov.b32 %f281, %r1846;
add.ftz.f32 %f59, %f280, %f281;
@%p14 bra $L__BB210_28;

cvt.u64.u32 %rd123, %r35;
add.s64 %rd124, %rd123, %rd14;
shl.b64 %rd126, %rd124, 2;
add.s64 %rd127, %rd93, %rd126;
st.global.f32 [%rd127], %f59;

$L__BB210_28:
ld.shared.f32 %f282, [%r33+1408];
add.ftz.f32 %f283, %f282, %f357;
mov.b32 %r1847, %f283;
shfl.sync.bfly.b32 %r1851|%p77, %r1847, %r1832, %r1831, %r1833;
mov.b32 %f284, %r1851;
add.ftz.f32 %f285, %f283, %f284;
mov.b32 %r1852, %f285;
shfl.sync.bfly.b32 %r1854|%p78, %r1852, %r1836, %r1831, %r1833;
mov.b32 %f286, %r1854;
add.ftz.f32 %f287, %f285, %f286;
mov.b32 %r1855, %f287;
shfl.sync.bfly.b32 %r1857|%p79, %r1855, %r1839, %r1831, %r1833;
mov.b32 %f288, %r1857;
add.ftz.f32 %f289, %f287, %f288;
mov.b32 %r1858, %f289;
shfl.sync.bfly.b32 %r1860|%p80, %r1858, %r1842, %r1831, %r1833;
mov.b32 %f290, %r1860;
add.ftz.f32 %f291, %f289, %f290;
mov.b32 %r1861, %f291;
shfl.sync.bfly.b32 %r1863|%p81, %r1861, %r1845, %r1831, %r1833;
mov.b32 %f292, %r1863;
add.ftz.f32 %f60, %f291, %f292;
@%p22 bra $L__BB210_30;

add.s32 %r1864, %r35, 1;
cvt.u64.u32 %rd128, %r1864;
add.s64 %rd129, %rd128, %rd14;
shl.b64 %rd131, %rd129, 2;
add.s64 %rd132, %rd93, %rd131;
st.global.f32 [%rd132], %f60;

$L__BB210_30:
mul.lo.s32 %r1865, %r39, 6;
ld.shared.f32 %f293, [%r33+1536];
add.ftz.f32 %f294, %f293, %f356;
mov.b32 %r1866, %f294;
mov.u32 %r1867, 31;
mov.u32 %r1868, 16;
mov.u32 %r1869, -1;
shfl.sync.bfly.b32 %r1870|%p83, %r1866, %r1868, %r1867, %r1869;
mov.b32 %f295, %r1870;
add.ftz.f32 %f296, %f294, %f295;
mov.b32 %r1871, %f296;
mov.u32 %r1872, 8;
shfl.sync.bfly.b32 %r1873|%p84, %r1871, %r1872, %r1867, %r1869;
mov.b32 %f297, %r1873;
add.ftz.f32 %f298, %f296, %f297;
mov.b32 %r1874, %f298;
mov.u32 %r1875, 4;
shfl.sync.bfly.b32 %r1876|%p85, %r1874, %r1875, %r1867, %r1869;
mov.b32 %f299, %r1876;
add.ftz.f32 %f300, %f298, %f299;
mov.b32 %r1877, %f300;
mov.u32 %r1878, 2;
shfl.sync.bfly.b32 %r1879|%p86, %r1877, %r1878, %r1867, %r1869;
mov.b32 %f301, %r1879;
add.ftz.f32 %f302, %f300, %f301;
mov.b32 %r1880, %f302;
mov.u32 %r1881, 1;
shfl.sync.bfly.b32 %r1882|%p87, %r1880, %r1881, %r1867, %r1869;
mov.b32 %f303, %r1882;
add.ftz.f32 %f61, %f302, %f303;
cvt.u64.u32 %rd133, %r1865;
add.s64 %rd134, %rd133, %rd14;
shl.b64 %rd136, %rd134, 2;
add.s64 %rd18, %rd93, %rd136;
@%p14 bra $L__BB210_32;

st.global.f32 [%rd18], %f61;

$L__BB210_32:
ld.shared.f32 %f304, [%r33+1664];
add.ftz.f32 %f305, %f304, %f355;
mov.b32 %r1883, %f305;
shfl.sync.bfly.b32 %r1887|%p89, %r1883, %r1868, %r1867, %r1869;
mov.b32 %f306, %r1887;
add.ftz.f32 %f307, %f305, %f306;
mov.b32 %r1888, %f307;
shfl.sync.bfly.b32 %r1890|%p90, %r1888, %r1872, %r1867, %r1869;
mov.b32 %f308, %r1890;
add.ftz.f32 %f309, %f307, %f308;
mov.b32 %r1891, %f309;
shfl.sync.bfly.b32 %r1893|%p91, %r1891, %r1875, %r1867, %r1869;
mov.b32 %f310, %r1893;
add.ftz.f32 %f311, %f309, %f310;
mov.b32 %r1894, %f311;
shfl.sync.bfly.b32 %r1896|%p92, %r1894, %r1878, %r1867, %r1869;
mov.b32 %f312, %r1896;
add.ftz.f32 %f313, %f311, %f312;
mov.b32 %r1897, %f313;
shfl.sync.bfly.b32 %r1899|%p93, %r1897, %r1881, %r1867, %r1869;
mov.b32 %f314, %r1899;
add.ftz.f32 %f62, %f313, %f314;
@%p22 bra $L__BB210_34;

st.global.f32 [%rd18+4], %f62;

$L__BB210_34:
mul.lo.s32 %r36, %r39, 7;
ld.shared.f32 %f315, [%r33+1792];
add.ftz.f32 %f316, %f315, %f354;
mov.b32 %r1900, %f316;
mov.u32 %r1901, 31;
mov.u32 %r1902, 16;
mov.u32 %r1903, -1;
shfl.sync.bfly.b32 %r1904|%p95, %r1900, %r1902, %r1901, %r1903;
mov.b32 %f317, %r1904;
add.ftz.f32 %f318, %f316, %f317;
mov.b32 %r1905, %f318;
mov.u32 %r1906, 8;
shfl.sync.bfly.b32 %r1907|%p96, %r1905, %r1906, %r1901, %r1903;
mov.b32 %f319, %r1907;
add.ftz.f32 %f320, %f318, %f319;
mov.b32 %r1908, %f320;
mov.u32 %r1909, 4;
shfl.sync.bfly.b32 %r1910|%p97, %r1908, %r1909, %r1901, %r1903;
mov.b32 %f321, %r1910;
add.ftz.f32 %f322, %f320, %f321;
mov.b32 %r1911, %f322;
mov.u32 %r1912, 2;
shfl.sync.bfly.b32 %r1913|%p98, %r1911, %r1912, %r1901, %r1903;
mov.b32 %f323, %r1913;
add.ftz.f32 %f324, %f322, %f323;
mov.b32 %r1914, %f324;
mov.u32 %r1915, 1;
shfl.sync.bfly.b32 %r1916|%p99, %r1914, %r1915, %r1901, %r1903;
mov.b32 %f325, %r1916;
add.ftz.f32 %f63, %f324, %f325;
@%p14 bra $L__BB210_36;

cvt.u64.u32 %rd137, %r36;
add.s64 %rd138, %rd137, %rd14;
shl.b64 %rd140, %rd138, 2;
add.s64 %rd141, %rd93, %rd140;
st.global.f32 [%rd141], %f63;

$L__BB210_36:
ld.shared.f32 %f326, [%r33+1920];
add.ftz.f32 %f327, %f326, %f353;
mov.b32 %r1917, %f327;
shfl.sync.bfly.b32 %r1921|%p101, %r1917, %r1902, %r1901, %r1903;
mov.b32 %f328, %r1921;
add.ftz.f32 %f329, %f327, %f328;
mov.b32 %r1922, %f329;
shfl.sync.bfly.b32 %r1924|%p102, %r1922, %r1906, %r1901, %r1903;
mov.b32 %f330, %r1924;
add.ftz.f32 %f331, %f329, %f330;
mov.b32 %r1925, %f331;
shfl.sync.bfly.b32 %r1927|%p103, %r1925, %r1909, %r1901, %r1903;
mov.b32 %f332, %r1927;
add.ftz.f32 %f333, %f331, %f332;
mov.b32 %r1928, %f333;
shfl.sync.bfly.b32 %r1930|%p104, %r1928, %r1912, %r1901, %r1903;
mov.b32 %f334, %r1930;
add.ftz.f32 %f335, %f333, %f334;
mov.b32 %r1931, %f335;
shfl.sync.bfly.b32 %r1933|%p105, %r1931, %r1915, %r1901, %r1903;
mov.b32 %f336, %r1933;
add.ftz.f32 %f64, %f335, %f336;
@%p22 bra $L__BB210_38;

add.s32 %r1934, %r36, 1;
cvt.u64.u32 %rd142, %r1934;
add.s64 %rd143, %rd142, %rd14;
shl.b64 %rd145, %rd143, 2;
add.s64 %rd146, %rd93, %rd145;
st.global.f32 [%rd146], %f64;

$L__BB210_38:
ret;

}
.entry _Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj(
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2,
.param .align 8 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3[48],
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5,
.param .align 4 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6[12],
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15
)
.maxntid 256, 1, 1
.minnctapersm 1
{
.local .align 8 .b8 __local_depot211[16];
.reg .b64 %SP;
.reg .b64 %SPL;
.reg .pred %p<48>;
.reg .b16 %rs<67>;
.reg .f32 %f<169>;
.reg .b32 %r<851>;
.reg .b64 %rd<151>;


mov.u64 %SPL, __local_depot211;
ld.param.u64 %rd20, [_Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0];
ld.param.u64 %rd21, [_Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1];
ld.param.u64 %rd22, [_Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2];
ld.param.u32 %r24, [_Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5];
ld.param.u32 %r29, [_Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8];
ld.param.u32 %r30, [_Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9];
ld.param.u32 %r32, [_Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11];
ld.param.u32 %r33, [_Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12];
ld.param.u32 %r36, [_Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14];
ld.param.u32 %r35, [_Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15];
ld.param.u32 %r27, [_Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+8];
ld.param.u32 %r26, [_Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+4];
ld.param.u32 %r25, [_Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6];
ld.param.u64 %rd24, [_Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+8];
add.u64 %rd1, %SPL, 0;
add.u64 %rd2, %SPL, 8;
mov.u32 %r1, %tid.y;
setp.ge.u32 %p1, %r1, %r36;
@%p1 bra $L__BB211_29;

mov.u32 %r2, %ctaid.y;
cvta.to.global.u64 %rd31, %rd22;
mad.lo.s32 %r37, %r1, %r35, %r2;
mul.wide.u32 %rd32, %r37, 4;
add.s64 %rd33, %rd31, %rd32;
ld.global.nc.u32 %r3, [%rd33];
mov.f32 %f162, 0f00000000;
st.local.v2.f32 [%rd1], {%f162, %f162};
st.local.v2.f32 [%rd2], {%f162, %f162};
mov.u32 %r4, %tid.x;
shr.u32 %r850, %r4, 3;
shr.u32 %r38, %r24, 8;
setp.ge.u32 %p2, %r850, %r38;
mov.f32 %f163, %f162;
mov.f32 %f164, %f162;
mov.f32 %f165, %f162;
@%p2 bra $L__BB211_9;

mov.u32 %r39, %ctaid.x;
shl.b32 %r40, %r39, 1;
mul.lo.s32 %r41, %r3, %r32;
and.b32 %r42, %r4, 7;
cvt.u64.u32 %rd34, %r42;
shl.b32 %r849, %r850, 3;
or.b32 %r43, %r40, 1;
add.s32 %r44, %r850, %r41;
mad.lo.s32 %r45, %r29, %r43, %r44;
mul.wide.s32 %rd4, %r45, 74;
add.s64 %rd35, %rd4, %rd34;
add.s64 %rd5, %rd35, 66;
mad.lo.s32 %r46, %r40, %r29, %r850;
add.s32 %r47, %r46, %r41;
mul.wide.s32 %rd6, %r47, 74;
add.s64 %rd36, %rd6, %rd34;
add.s64 %rd7, %rd36, 66;
shl.b32 %r48, %r4, 3;
cvt.u64.u32 %rd37, %r48;
and.b64 %rd38, %rd37, 56;
add.s64 %rd39, %rd38, %rd4;
add.s64 %rd8, %rd39, 4;
add.s64 %rd40, %rd38, %rd6;
add.s64 %rd9, %rd40, 8;
cvta.to.global.u64 %rd149, %rd20;
cvta.to.global.u64 %rd150, %rd24;
mul.lo.s32 %r49, %r1, %r30;
cvt.u64.u32 %rd41, %r49;
mul.hi.u32 %r50, %r2, %r25;
add.s32 %r51, %r50, %r2;
shr.u32 %r52, %r51, %r26;
mul.lo.s32 %r53, %r52, %r27;
sub.s32 %r54, %r2, %r53;
mul.lo.s32 %r55, %r54, %r33;
cvt.u64.u32 %rd42, %r55;
add.s64 %rd12, %rd42, %rd41;
mov.f32 %f162, 0f00000000;
cvta.to.global.u64 %rd43, %rd21;

$L__BB211_3:
cvt.s64.s32 %rd44, %r849;
add.s64 %rd45, %rd12, %rd44;
add.s64 %rd46, %rd149, %rd9;
ld.global.nc.u16 %rs2, [%rd46+-4];
ld.global.nc.u16 %rs3, [%rd46+-2];
ld.global.nc.u16 %rs4, [%rd46];
add.s64 %rd47, %rd149, %rd7;
ld.global.nc.u8 %rs5, [%rd47];
and.b16 %rs6, %rs5, 240;
and.b16 %rs7, %rs5, 15;
shr.u16 %rs8, %rs6, 4;
add.s64 %rd49, %rd45, %rd34;
mul.lo.s64 %rd50, %rd49, 36;
ld.global.nc.u16 %rs9, [%rd46+-6];
shl.b16 %rs10, %rs9, 3;
cvt.u64.u16 %rd51, %rs10;
and.b64 %rd52, %rd51, 4088;
mov.u64 %rd53, iq2xs_grid;
add.s64 %rd54, %rd53, %rd52;
ld.global.nc.v2.u32 {%r107, %r108}, [%rd54];
mov.u32 %r87, 0;
shr.u16 %rs11, %rs9, 9;
cvt.u32.u16 %r111, %rs11;
popc.b32 %r112, %r111;
shl.b32 %r113, %r112, 7;
and.b32 %r114, %r113, 128;
or.b32 %r115, %r114, %r111;
mul.lo.s32 %r116, %r115, 16843009;
and.b32 %r117, %r116, 134480385;
or.b32 %r118, %r117, -2139062144;
add.s32 %r119, %r118, -16843009;
or.b32 %r56, %r119, %r117;

prmt.b32 %r56, %r56, 0, 0xba98;

xor.b32 %r120, %r56, %r107;
or.b32 %r121, %r120, -2139062144;
and.b32 %r122, %r56, 2139062143;
and.b32 %r123, %r107, -2139062144;
sub.s32 %r124, %r121, %r122;
xor.b32 %r125, %r123, %r124;
xor.b32 %r61, %r125, -2139062144;
add.s64 %rd55, %rd43, %rd50;
ld.global.nc.u32 %r9, [%rd55+4];
and.b32 %r126, %r116, -2143281136;
or.b32 %r127, %r126, -2139062144;
add.s32 %r128, %r127, -16843009;
or.b32 %r58, %r128, %r126;

prmt.b32 %r58, %r58, 0, 0xba98;

xor.b32 %r129, %r58, %r108;
or.b32 %r130, %r129, -2139062144;
and.b32 %r131, %r58, 2139062143;
and.b32 %r132, %r108, -2139062144;
sub.s32 %r133, %r130, %r131;
xor.b32 %r134, %r132, %r133;
xor.b32 %r65, %r134, -2139062144;
ld.global.nc.u32 %r10, [%rd55+8];

dp4a.s32.s32 %r60, %r61, %r9, %r87;


dp4a.s32.s32 %r64, %r65, %r10, %r60;

shl.b16 %rs12, %rs2, 3;
cvt.u64.u16 %rd56, %rs12;
and.b64 %rd57, %rd56, 4088;
add.s64 %rd58, %rd53, %rd57;
ld.global.nc.v2.u32 {%r135, %r136}, [%rd58];
shr.u16 %rs13, %rs2, 9;
cvt.u32.u16 %r139, %rs13;
popc.b32 %r140, %r139;
shl.b32 %r141, %r140, 7;
and.b32 %r142, %r141, 128;
or.b32 %r143, %r142, %r139;
mul.lo.s32 %r144, %r143, 16843009;
and.b32 %r145, %r144, 134480385;
or.b32 %r146, %r145, -2139062144;
add.s32 %r147, %r146, -16843009;
or.b32 %r68, %r147, %r145;

prmt.b32 %r68, %r68, 0, 0xba98;

xor.b32 %r148, %r68, %r135;
or.b32 %r149, %r148, -2139062144;
and.b32 %r150, %r68, 2139062143;
and.b32 %r151, %r135, -2139062144;
sub.s32 %r152, %r149, %r150;
xor.b32 %r153, %r151, %r152;
xor.b32 %r73, %r153, -2139062144;
ld.global.nc.u32 %r11, [%rd55+12];
and.b32 %r154, %r144, -2143281136;
or.b32 %r155, %r154, -2139062144;
add.s32 %r156, %r155, -16843009;
or.b32 %r70, %r156, %r154;

prmt.b32 %r70, %r70, 0, 0xba98;

xor.b32 %r157, %r70, %r136;
or.b32 %r158, %r157, -2139062144;
and.b32 %r159, %r70, 2139062143;
and.b32 %r160, %r136, -2139062144;
sub.s32 %r161, %r158, %r159;
xor.b32 %r162, %r160, %r161;
xor.b32 %r77, %r162, -2139062144;
ld.global.nc.u32 %r12, [%rd55+16];

dp4a.s32.s32 %r72, %r73, %r11, %r64;


dp4a.s32.s32 %r76, %r77, %r12, %r72;

shl.b16 %rs14, %rs3, 3;
cvt.u64.u16 %rd59, %rs14;
and.b64 %rd60, %rd59, 4088;
add.s64 %rd61, %rd53, %rd60;
ld.global.nc.v2.u32 {%r163, %r164}, [%rd61];
shr.u16 %rs15, %rs3, 9;
cvt.u32.u16 %r167, %rs15;
popc.b32 %r168, %r167;
shl.b32 %r169, %r168, 7;
and.b32 %r170, %r169, 128;
or.b32 %r171, %r170, %r167;
mul.lo.s32 %r172, %r171, 16843009;
and.b32 %r173, %r172, 134480385;
or.b32 %r174, %r173, -2139062144;
add.s32 %r175, %r174, -16843009;
or.b32 %r80, %r175, %r173;

prmt.b32 %r80, %r80, 0, 0xba98;

xor.b32 %r176, %r80, %r163;
or.b32 %r177, %r176, -2139062144;
and.b32 %r178, %r80, 2139062143;
and.b32 %r179, %r163, -2139062144;
sub.s32 %r180, %r177, %r178;
xor.b32 %r181, %r179, %r180;
xor.b32 %r85, %r181, -2139062144;
ld.global.nc.u32 %r13, [%rd55+20];
and.b32 %r182, %r172, -2143281136;
or.b32 %r183, %r182, -2139062144;
add.s32 %r184, %r183, -16843009;
or.b32 %r82, %r184, %r182;

prmt.b32 %r82, %r82, 0, 0xba98;

xor.b32 %r185, %r82, %r164;
or.b32 %r186, %r185, -2139062144;
and.b32 %r187, %r82, 2139062143;
and.b32 %r188, %r164, -2139062144;
sub.s32 %r189, %r186, %r187;
xor.b32 %r190, %r188, %r189;
xor.b32 %r89, %r190, -2139062144;
ld.global.nc.u32 %r14, [%rd55+24];

dp4a.s32.s32 %r84, %r85, %r13, %r87;


dp4a.s32.s32 %r88, %r89, %r14, %r84;

shl.b16 %rs16, %rs4, 3;
cvt.u64.u16 %rd62, %rs16;
and.b64 %rd63, %rd62, 4088;
add.s64 %rd64, %rd53, %rd63;
ld.global.nc.v2.u32 {%r191, %r192}, [%rd64];
shr.u16 %rs17, %rs4, 9;
cvt.u32.u16 %r195, %rs17;
popc.b32 %r196, %r195;
shl.b32 %r197, %r196, 7;
and.b32 %r198, %r197, 128;
or.b32 %r199, %r198, %r195;
mul.lo.s32 %r200, %r199, 16843009;
and.b32 %r201, %r200, 134480385;
or.b32 %r202, %r201, -2139062144;
add.s32 %r203, %r202, -16843009;
or.b32 %r92, %r203, %r201;

prmt.b32 %r92, %r92, 0, 0xba98;

xor.b32 %r204, %r92, %r191;
or.b32 %r205, %r204, -2139062144;
and.b32 %r206, %r92, 2139062143;
and.b32 %r207, %r191, -2139062144;
sub.s32 %r208, %r205, %r206;
xor.b32 %r209, %r207, %r208;
xor.b32 %r97, %r209, -2139062144;
ld.global.nc.u32 %r15, [%rd55+28];
and.b32 %r210, %r200, -2143281136;
or.b32 %r211, %r210, -2139062144;
add.s32 %r212, %r211, -16843009;
or.b32 %r94, %r212, %r210;

prmt.b32 %r94, %r94, 0, 0xba98;

xor.b32 %r213, %r94, %r192;
or.b32 %r214, %r213, -2139062144;
and.b32 %r215, %r94, 2139062143;
and.b32 %r216, %r192, -2139062144;
sub.s32 %r217, %r214, %r215;
xor.b32 %r218, %r216, %r217;
xor.b32 %r101, %r218, -2139062144;
ld.global.nc.u32 %r16, [%rd55+32];

dp4a.s32.s32 %r96, %r97, %r15, %r88;


dp4a.s32.s32 %r100, %r101, %r16, %r96;

cvt.u32.u16 %r219, %rs7;
cvt.u32.u16 %r220, %rs8;
mul.lo.s32 %r221, %r76, %r219;
mad.lo.s32 %r222, %r100, %r220, %r221;
add.s32 %r223, %r100, %r76;
shr.u32 %r224, %r223, 31;
add.s32 %r225, %r223, %r224;
shr.s32 %r226, %r225, 1;
add.s32 %r227, %r222, %r226;
shr.s32 %r228, %r227, 31;
shr.u32 %r229, %r228, 30;
add.s32 %r230, %r227, %r229;
shr.s32 %r231, %r230, 2;
add.s64 %rd65, %rd149, %rd6;
ld.global.nc.u16 %rs1, [%rd65];

{ cvt.f32.f16 %f42, %rs1;}


ld.global.nc.u32 %r17, [%rd55];

{.reg .f16 low,high;
mov.b32 {low,high},%r17;
cvt.f32.f16 %f43, low;}


mul.ftz.f32 %f44, %f42, %f43;
cvt.rn.f32.s32 %f45, %r231;
fma.rn.ftz.f32 %f165, %f44, %f45, %f165;
setp.eq.s64 %p3, %rd24, 0;
@%p3 bra $L__BB211_5;

mov.u64 %rd133, iq2xs_grid;
add.s64 %rd66, %rd150, %rd9;
ld.global.u16 %rs19, [%rd66+-4];
ld.global.u16 %rs20, [%rd66+-2];
ld.global.u16 %rs21, [%rd66];
add.s64 %rd67, %rd150, %rd7;
ld.global.u8 %rs22, [%rd67];
and.b16 %rs23, %rs22, 15;
shr.u16 %rs24, %rs22, 4;
ld.global.u16 %rs25, [%rd66+-6];
shl.b16 %rs26, %rs25, 3;
cvt.u64.u16 %rd68, %rs26;
and.b64 %rd69, %rd68, 4088;
add.s64 %rd71, %rd133, %rd69;
ld.global.nc.v2.u32 {%r281, %r282}, [%rd71];
shr.u16 %rs27, %rs25, 9;
cvt.u32.u16 %r285, %rs27;
popc.b32 %r286, %r285;
shl.b32 %r287, %r286, 7;
and.b32 %r288, %r287, 128;
or.b32 %r289, %r288, %r285;
mul.lo.s32 %r290, %r289, 16843009;
and.b32 %r291, %r290, 134480385;
or.b32 %r292, %r291, -2139062144;
add.s32 %r293, %r292, -16843009;
or.b32 %r232, %r293, %r291;

prmt.b32 %r232, %r232, 0, 0xba98;

xor.b32 %r294, %r232, %r281;
or.b32 %r295, %r294, -2139062144;
and.b32 %r296, %r232, 2139062143;
and.b32 %r297, %r281, -2139062144;
sub.s32 %r298, %r295, %r296;
xor.b32 %r299, %r297, %r298;
xor.b32 %r237, %r299, -2139062144;
and.b32 %r300, %r290, -2143281136;
or.b32 %r301, %r300, -2139062144;
add.s32 %r302, %r301, -16843009;
or.b32 %r234, %r302, %r300;

prmt.b32 %r234, %r234, 0, 0xba98;

xor.b32 %r303, %r234, %r282;
or.b32 %r304, %r303, -2139062144;
and.b32 %r305, %r234, 2139062143;
and.b32 %r306, %r282, -2139062144;
sub.s32 %r307, %r304, %r305;
xor.b32 %r308, %r306, %r307;
xor.b32 %r241, %r308, -2139062144;

dp4a.s32.s32 %r236, %r237, %r9, %r87;


dp4a.s32.s32 %r240, %r241, %r10, %r236;

shl.b16 %rs28, %rs19, 3;
cvt.u64.u16 %rd72, %rs28;
and.b64 %rd73, %rd72, 4088;
add.s64 %rd74, %rd133, %rd73;
ld.global.nc.v2.u32 {%r309, %r310}, [%rd74];
shr.u16 %rs29, %rs19, 9;
cvt.u32.u16 %r313, %rs29;
popc.b32 %r314, %r313;
shl.b32 %r315, %r314, 7;
and.b32 %r316, %r315, 128;
or.b32 %r317, %r316, %r313;
mul.lo.s32 %r318, %r317, 16843009;
and.b32 %r319, %r318, 134480385;
or.b32 %r320, %r319, -2139062144;
add.s32 %r321, %r320, -16843009;
or.b32 %r244, %r321, %r319;

prmt.b32 %r244, %r244, 0, 0xba98;

xor.b32 %r322, %r244, %r309;
or.b32 %r323, %r322, -2139062144;
and.b32 %r324, %r244, 2139062143;
and.b32 %r325, %r309, -2139062144;
sub.s32 %r326, %r323, %r324;
xor.b32 %r327, %r325, %r326;
xor.b32 %r249, %r327, -2139062144;
and.b32 %r328, %r318, -2143281136;
or.b32 %r329, %r328, -2139062144;
add.s32 %r330, %r329, -16843009;
or.b32 %r246, %r330, %r328;

prmt.b32 %r246, %r246, 0, 0xba98;

xor.b32 %r331, %r246, %r310;
or.b32 %r332, %r331, -2139062144;
and.b32 %r333, %r246, 2139062143;
and.b32 %r334, %r310, -2139062144;
sub.s32 %r335, %r332, %r333;
xor.b32 %r336, %r334, %r335;
xor.b32 %r253, %r336, -2139062144;

dp4a.s32.s32 %r248, %r249, %r11, %r240;


dp4a.s32.s32 %r252, %r253, %r12, %r248;

shl.b16 %rs30, %rs20, 3;
cvt.u64.u16 %rd75, %rs30;
and.b64 %rd76, %rd75, 4088;
add.s64 %rd77, %rd133, %rd76;
ld.global.nc.v2.u32 {%r337, %r338}, [%rd77];
shr.u16 %rs31, %rs20, 9;
cvt.u32.u16 %r341, %rs31;
popc.b32 %r342, %r341;
shl.b32 %r343, %r342, 7;
and.b32 %r344, %r343, 128;
or.b32 %r345, %r344, %r341;
mul.lo.s32 %r346, %r345, 16843009;
and.b32 %r347, %r346, 134480385;
or.b32 %r348, %r347, -2139062144;
add.s32 %r349, %r348, -16843009;
or.b32 %r256, %r349, %r347;

prmt.b32 %r256, %r256, 0, 0xba98;

xor.b32 %r350, %r256, %r337;
or.b32 %r351, %r350, -2139062144;
and.b32 %r352, %r256, 2139062143;
and.b32 %r353, %r337, -2139062144;
sub.s32 %r354, %r351, %r352;
xor.b32 %r355, %r353, %r354;
xor.b32 %r261, %r355, -2139062144;
and.b32 %r356, %r346, -2143281136;
or.b32 %r357, %r356, -2139062144;
add.s32 %r358, %r357, -16843009;
or.b32 %r258, %r358, %r356;

prmt.b32 %r258, %r258, 0, 0xba98;

xor.b32 %r359, %r258, %r338;
or.b32 %r360, %r359, -2139062144;
and.b32 %r361, %r258, 2139062143;
and.b32 %r362, %r338, -2139062144;
sub.s32 %r363, %r360, %r361;
xor.b32 %r364, %r362, %r363;
xor.b32 %r265, %r364, -2139062144;

dp4a.s32.s32 %r260, %r261, %r13, %r87;


dp4a.s32.s32 %r264, %r265, %r14, %r260;

shl.b16 %rs32, %rs21, 3;
cvt.u64.u16 %rd78, %rs32;
and.b64 %rd79, %rd78, 4088;
add.s64 %rd80, %rd133, %rd79;
ld.global.nc.v2.u32 {%r365, %r366}, [%rd80];
shr.u16 %rs33, %rs21, 9;
cvt.u32.u16 %r369, %rs33;
popc.b32 %r370, %r369;
shl.b32 %r371, %r370, 7;
and.b32 %r372, %r371, 128;
or.b32 %r373, %r372, %r369;
mul.lo.s32 %r374, %r373, 16843009;
and.b32 %r375, %r374, 134480385;
or.b32 %r376, %r375, -2139062144;
add.s32 %r377, %r376, -16843009;
or.b32 %r268, %r377, %r375;

prmt.b32 %r268, %r268, 0, 0xba98;

xor.b32 %r378, %r268, %r365;
or.b32 %r379, %r378, -2139062144;
and.b32 %r380, %r268, 2139062143;
and.b32 %r381, %r365, -2139062144;
sub.s32 %r382, %r379, %r380;
xor.b32 %r383, %r381, %r382;
xor.b32 %r273, %r383, -2139062144;
and.b32 %r384, %r374, -2143281136;
or.b32 %r385, %r384, -2139062144;
add.s32 %r386, %r385, -16843009;
or.b32 %r270, %r386, %r384;

prmt.b32 %r270, %r270, 0, 0xba98;

xor.b32 %r387, %r270, %r366;
or.b32 %r388, %r387, -2139062144;
and.b32 %r389, %r270, 2139062143;
and.b32 %r390, %r366, -2139062144;
sub.s32 %r391, %r388, %r389;
xor.b32 %r392, %r390, %r391;
xor.b32 %r277, %r392, -2139062144;

dp4a.s32.s32 %r272, %r273, %r15, %r264;


dp4a.s32.s32 %r276, %r277, %r16, %r272;

cvt.u32.u16 %r393, %rs23;
cvt.u32.u16 %r394, %rs24;
mul.lo.s32 %r395, %r252, %r393;
mad.lo.s32 %r396, %r276, %r394, %r395;
add.s32 %r397, %r276, %r252;
shr.u32 %r398, %r397, 31;
add.s32 %r399, %r397, %r398;
shr.s32 %r400, %r399, 1;
add.s32 %r401, %r396, %r400;
shr.s32 %r402, %r401, 31;
shr.u32 %r403, %r402, 30;
add.s32 %r404, %r401, %r403;
shr.s32 %r405, %r404, 2;
add.s64 %rd81, %rd150, %rd6;
ld.global.u16 %rs18, [%rd81];

{ cvt.f32.f16 %f46, %rs18;}



{.reg .f16 low,high;
mov.b32 {low,high},%r17;
cvt.f32.f16 %f47, low;}


mul.ftz.f32 %f48, %f46, %f47;
cvt.rn.f32.s32 %f49, %r405;
fma.rn.ftz.f32 %f163, %f48, %f49, %f163;

$L__BB211_5:
mov.u64 %rd127, iq2xs_grid;
add.s64 %rd82, %rd149, %rd8;
ld.global.nc.u16 %rs35, [%rd82];
ld.global.nc.u16 %rs36, [%rd82+2];
ld.global.nc.u16 %rs37, [%rd82+4];
add.s64 %rd83, %rd149, %rd5;
ld.global.nc.u8 %rs38, [%rd83];
and.b16 %rs39, %rs38, 240;
and.b16 %rs40, %rs38, 15;
shr.u16 %rs41, %rs39, 4;
ld.global.nc.u16 %rs42, [%rd82+-2];
shl.b16 %rs43, %rs42, 3;
cvt.u64.u16 %rd84, %rs43;
and.b64 %rd85, %rd84, 4088;
add.s64 %rd87, %rd127, %rd85;
ld.global.nc.v2.u32 {%r455, %r456}, [%rd87];
shr.u16 %rs44, %rs42, 9;
cvt.u32.u16 %r459, %rs44;
popc.b32 %r460, %r459;
shl.b32 %r461, %r460, 7;
and.b32 %r462, %r461, 128;
or.b32 %r463, %r462, %r459;
mul.lo.s32 %r464, %r463, 16843009;
and.b32 %r465, %r464, 134480385;
or.b32 %r466, %r465, -2139062144;
add.s32 %r467, %r466, -16843009;
or.b32 %r406, %r467, %r465;

prmt.b32 %r406, %r406, 0, 0xba98;

xor.b32 %r468, %r406, %r455;
or.b32 %r469, %r468, -2139062144;
and.b32 %r470, %r406, 2139062143;
and.b32 %r471, %r455, -2139062144;
sub.s32 %r472, %r469, %r470;
xor.b32 %r473, %r471, %r472;
xor.b32 %r411, %r473, -2139062144;
and.b32 %r474, %r464, -2143281136;
or.b32 %r475, %r474, -2139062144;
add.s32 %r476, %r475, -16843009;
or.b32 %r408, %r476, %r474;

prmt.b32 %r408, %r408, 0, 0xba98;

xor.b32 %r477, %r408, %r456;
or.b32 %r478, %r477, -2139062144;
and.b32 %r479, %r408, 2139062143;
and.b32 %r480, %r456, -2139062144;
sub.s32 %r481, %r478, %r479;
xor.b32 %r482, %r480, %r481;
xor.b32 %r415, %r482, -2139062144;

dp4a.s32.s32 %r410, %r411, %r9, %r87;


dp4a.s32.s32 %r414, %r415, %r10, %r410;

shl.b16 %rs45, %rs35, 3;
cvt.u64.u16 %rd88, %rs45;
and.b64 %rd89, %rd88, 4088;
add.s64 %rd90, %rd127, %rd89;
ld.global.nc.v2.u32 {%r483, %r484}, [%rd90];
shr.u16 %rs46, %rs35, 9;
cvt.u32.u16 %r487, %rs46;
popc.b32 %r488, %r487;
shl.b32 %r489, %r488, 7;
and.b32 %r490, %r489, 128;
or.b32 %r491, %r490, %r487;
mul.lo.s32 %r492, %r491, 16843009;
and.b32 %r493, %r492, 134480385;
or.b32 %r494, %r493, -2139062144;
add.s32 %r495, %r494, -16843009;
or.b32 %r418, %r495, %r493;

prmt.b32 %r418, %r418, 0, 0xba98;

xor.b32 %r496, %r418, %r483;
or.b32 %r497, %r496, -2139062144;
and.b32 %r498, %r418, 2139062143;
and.b32 %r499, %r483, -2139062144;
sub.s32 %r500, %r497, %r498;
xor.b32 %r501, %r499, %r500;
xor.b32 %r423, %r501, -2139062144;
and.b32 %r502, %r492, -2143281136;
or.b32 %r503, %r502, -2139062144;
add.s32 %r504, %r503, -16843009;
or.b32 %r420, %r504, %r502;

prmt.b32 %r420, %r420, 0, 0xba98;

xor.b32 %r505, %r420, %r484;
or.b32 %r506, %r505, -2139062144;
and.b32 %r507, %r420, 2139062143;
and.b32 %r508, %r484, -2139062144;
sub.s32 %r509, %r506, %r507;
xor.b32 %r510, %r508, %r509;
xor.b32 %r427, %r510, -2139062144;

dp4a.s32.s32 %r422, %r423, %r11, %r414;


dp4a.s32.s32 %r426, %r427, %r12, %r422;

shl.b16 %rs47, %rs36, 3;
cvt.u64.u16 %rd91, %rs47;
and.b64 %rd92, %rd91, 4088;
add.s64 %rd93, %rd127, %rd92;
ld.global.nc.v2.u32 {%r511, %r512}, [%rd93];
shr.u16 %rs48, %rs36, 9;
cvt.u32.u16 %r515, %rs48;
popc.b32 %r516, %r515;
shl.b32 %r517, %r516, 7;
and.b32 %r518, %r517, 128;
or.b32 %r519, %r518, %r515;
mul.lo.s32 %r520, %r519, 16843009;
and.b32 %r521, %r520, 134480385;
or.b32 %r522, %r521, -2139062144;
add.s32 %r523, %r522, -16843009;
or.b32 %r430, %r523, %r521;

prmt.b32 %r430, %r430, 0, 0xba98;

xor.b32 %r524, %r430, %r511;
or.b32 %r525, %r524, -2139062144;
and.b32 %r526, %r430, 2139062143;
and.b32 %r527, %r511, -2139062144;
sub.s32 %r528, %r525, %r526;
xor.b32 %r529, %r527, %r528;
xor.b32 %r435, %r529, -2139062144;
and.b32 %r530, %r520, -2143281136;
or.b32 %r531, %r530, -2139062144;
add.s32 %r532, %r531, -16843009;
or.b32 %r432, %r532, %r530;

prmt.b32 %r432, %r432, 0, 0xba98;

xor.b32 %r533, %r432, %r512;
or.b32 %r534, %r533, -2139062144;
and.b32 %r535, %r432, 2139062143;
and.b32 %r536, %r512, -2139062144;
sub.s32 %r537, %r534, %r535;
xor.b32 %r538, %r536, %r537;
xor.b32 %r439, %r538, -2139062144;

dp4a.s32.s32 %r434, %r435, %r13, %r87;


dp4a.s32.s32 %r438, %r439, %r14, %r434;

shl.b16 %rs49, %rs37, 3;
cvt.u64.u16 %rd94, %rs49;
and.b64 %rd95, %rd94, 4088;
add.s64 %rd96, %rd127, %rd95;
ld.global.nc.v2.u32 {%r539, %r540}, [%rd96];
shr.u16 %rs50, %rs37, 9;
cvt.u32.u16 %r543, %rs50;
popc.b32 %r544, %r543;
shl.b32 %r545, %r544, 7;
and.b32 %r546, %r545, 128;
or.b32 %r547, %r546, %r543;
mul.lo.s32 %r548, %r547, 16843009;
and.b32 %r549, %r548, 134480385;
or.b32 %r550, %r549, -2139062144;
add.s32 %r551, %r550, -16843009;
or.b32 %r442, %r551, %r549;

prmt.b32 %r442, %r442, 0, 0xba98;

xor.b32 %r552, %r442, %r539;
or.b32 %r553, %r552, -2139062144;
and.b32 %r554, %r442, 2139062143;
and.b32 %r555, %r539, -2139062144;
sub.s32 %r556, %r553, %r554;
xor.b32 %r557, %r555, %r556;
xor.b32 %r447, %r557, -2139062144;
and.b32 %r558, %r548, -2143281136;
or.b32 %r559, %r558, -2139062144;
add.s32 %r560, %r559, -16843009;
or.b32 %r444, %r560, %r558;

prmt.b32 %r444, %r444, 0, 0xba98;

xor.b32 %r561, %r444, %r540;
or.b32 %r562, %r561, -2139062144;
and.b32 %r563, %r444, 2139062143;
and.b32 %r564, %r540, -2139062144;
sub.s32 %r565, %r562, %r563;
xor.b32 %r566, %r564, %r565;
xor.b32 %r451, %r566, -2139062144;

dp4a.s32.s32 %r446, %r447, %r15, %r438;


dp4a.s32.s32 %r450, %r451, %r16, %r446;

cvt.u32.u16 %r567, %rs40;
cvt.u32.u16 %r568, %rs41;
mul.lo.s32 %r569, %r426, %r567;
mad.lo.s32 %r570, %r450, %r568, %r569;
add.s32 %r571, %r450, %r426;
shr.u32 %r572, %r571, 31;
add.s32 %r573, %r571, %r572;
shr.s32 %r574, %r573, 1;
add.s32 %r575, %r570, %r574;
shr.s32 %r576, %r575, 31;
shr.u32 %r577, %r576, 30;
add.s32 %r578, %r575, %r577;
shr.s32 %r579, %r578, 2;
add.s64 %rd97, %rd149, %rd4;
ld.global.nc.u16 %rs34, [%rd97];

{ cvt.f32.f16 %f50, %rs34;}



{.reg .f16 low,high;
mov.b32 {low,high},%r17;
cvt.f32.f16 %f51, low;}


mul.ftz.f32 %f52, %f50, %f51;
cvt.rn.f32.s32 %f53, %r579;
fma.rn.ftz.f32 %f164, %f52, %f53, %f164;
@%p3 bra $L__BB211_7;

mov.u64 %rd134, iq2xs_grid;
add.s64 %rd98, %rd150, %rd8;
ld.global.u16 %rs52, [%rd98];
ld.global.u16 %rs53, [%rd98+2];
ld.global.u16 %rs54, [%rd98+4];
add.s64 %rd99, %rd150, %rd5;
ld.global.u8 %rs55, [%rd99];
and.b16 %rs56, %rs55, 15;
shr.u16 %rs57, %rs55, 4;
ld.global.u16 %rs58, [%rd98+-2];
shl.b16 %rs59, %rs58, 3;
cvt.u64.u16 %rd100, %rs59;
and.b64 %rd101, %rd100, 4088;
add.s64 %rd103, %rd134, %rd101;
ld.global.nc.v2.u32 {%r629, %r630}, [%rd103];
mov.u32 %r611, 0;
shr.u16 %rs60, %rs58, 9;
cvt.u32.u16 %r633, %rs60;
popc.b32 %r634, %r633;
shl.b32 %r635, %r634, 7;
and.b32 %r636, %r635, 128;
or.b32 %r637, %r636, %r633;
mul.lo.s32 %r638, %r637, 16843009;
and.b32 %r639, %r638, 134480385;
or.b32 %r640, %r639, -2139062144;
add.s32 %r641, %r640, -16843009;
or.b32 %r580, %r641, %r639;

prmt.b32 %r580, %r580, 0, 0xba98;

xor.b32 %r642, %r580, %r629;
or.b32 %r643, %r642, -2139062144;
and.b32 %r644, %r580, 2139062143;
and.b32 %r645, %r629, -2139062144;
sub.s32 %r646, %r643, %r644;
xor.b32 %r647, %r645, %r646;
xor.b32 %r585, %r647, -2139062144;
and.b32 %r648, %r638, -2143281136;
or.b32 %r649, %r648, -2139062144;
add.s32 %r650, %r649, -16843009;
or.b32 %r582, %r650, %r648;

prmt.b32 %r582, %r582, 0, 0xba98;

xor.b32 %r651, %r582, %r630;
or.b32 %r652, %r651, -2139062144;
and.b32 %r653, %r582, 2139062143;
and.b32 %r654, %r630, -2139062144;
sub.s32 %r655, %r652, %r653;
xor.b32 %r656, %r654, %r655;
xor.b32 %r589, %r656, -2139062144;

dp4a.s32.s32 %r584, %r585, %r9, %r611;


dp4a.s32.s32 %r588, %r589, %r10, %r584;

shl.b16 %rs61, %rs52, 3;
cvt.u64.u16 %rd104, %rs61;
and.b64 %rd105, %rd104, 4088;
add.s64 %rd106, %rd134, %rd105;
ld.global.nc.v2.u32 {%r657, %r658}, [%rd106];
shr.u16 %rs62, %rs52, 9;
cvt.u32.u16 %r661, %rs62;
popc.b32 %r662, %r661;
shl.b32 %r663, %r662, 7;
and.b32 %r664, %r663, 128;
or.b32 %r665, %r664, %r661;
mul.lo.s32 %r666, %r665, 16843009;
and.b32 %r667, %r666, 134480385;
or.b32 %r668, %r667, -2139062144;
add.s32 %r669, %r668, -16843009;
or.b32 %r592, %r669, %r667;

prmt.b32 %r592, %r592, 0, 0xba98;

xor.b32 %r670, %r592, %r657;
or.b32 %r671, %r670, -2139062144;
and.b32 %r672, %r592, 2139062143;
and.b32 %r673, %r657, -2139062144;
sub.s32 %r674, %r671, %r672;
xor.b32 %r675, %r673, %r674;
xor.b32 %r597, %r675, -2139062144;
and.b32 %r676, %r666, -2143281136;
or.b32 %r677, %r676, -2139062144;
add.s32 %r678, %r677, -16843009;
or.b32 %r594, %r678, %r676;

prmt.b32 %r594, %r594, 0, 0xba98;

xor.b32 %r679, %r594, %r658;
or.b32 %r680, %r679, -2139062144;
and.b32 %r681, %r594, 2139062143;
and.b32 %r682, %r658, -2139062144;
sub.s32 %r683, %r680, %r681;
xor.b32 %r684, %r682, %r683;
xor.b32 %r601, %r684, -2139062144;

dp4a.s32.s32 %r596, %r597, %r11, %r588;


dp4a.s32.s32 %r600, %r601, %r12, %r596;

shl.b16 %rs63, %rs53, 3;
cvt.u64.u16 %rd107, %rs63;
and.b64 %rd108, %rd107, 4088;
add.s64 %rd109, %rd134, %rd108;
ld.global.nc.v2.u32 {%r685, %r686}, [%rd109];
shr.u16 %rs64, %rs53, 9;
cvt.u32.u16 %r689, %rs64;
popc.b32 %r690, %r689;
shl.b32 %r691, %r690, 7;
and.b32 %r692, %r691, 128;
or.b32 %r693, %r692, %r689;
mul.lo.s32 %r694, %r693, 16843009;
and.b32 %r695, %r694, 134480385;
or.b32 %r696, %r695, -2139062144;
add.s32 %r697, %r696, -16843009;
or.b32 %r604, %r697, %r695;

prmt.b32 %r604, %r604, 0, 0xba98;

xor.b32 %r698, %r604, %r685;
or.b32 %r699, %r698, -2139062144;
and.b32 %r700, %r604, 2139062143;
and.b32 %r701, %r685, -2139062144;
sub.s32 %r702, %r699, %r700;
xor.b32 %r703, %r701, %r702;
xor.b32 %r609, %r703, -2139062144;
and.b32 %r704, %r694, -2143281136;
or.b32 %r705, %r704, -2139062144;
add.s32 %r706, %r705, -16843009;
or.b32 %r606, %r706, %r704;

prmt.b32 %r606, %r606, 0, 0xba98;

xor.b32 %r707, %r606, %r686;
or.b32 %r708, %r707, -2139062144;
and.b32 %r709, %r606, 2139062143;
and.b32 %r710, %r686, -2139062144;
sub.s32 %r711, %r708, %r709;
xor.b32 %r712, %r710, %r711;
xor.b32 %r613, %r712, -2139062144;

dp4a.s32.s32 %r608, %r609, %r13, %r611;


dp4a.s32.s32 %r612, %r613, %r14, %r608;

shl.b16 %rs65, %rs54, 3;
cvt.u64.u16 %rd110, %rs65;
and.b64 %rd111, %rd110, 4088;
add.s64 %rd112, %rd134, %rd111;
ld.global.nc.v2.u32 {%r713, %r714}, [%rd112];
shr.u16 %rs66, %rs54, 9;
cvt.u32.u16 %r717, %rs66;
popc.b32 %r718, %r717;
shl.b32 %r719, %r718, 7;
and.b32 %r720, %r719, 128;
or.b32 %r721, %r720, %r717;
mul.lo.s32 %r722, %r721, 16843009;
and.b32 %r723, %r722, 134480385;
or.b32 %r724, %r723, -2139062144;
add.s32 %r725, %r724, -16843009;
or.b32 %r616, %r725, %r723;

prmt.b32 %r616, %r616, 0, 0xba98;

xor.b32 %r726, %r616, %r713;
or.b32 %r727, %r726, -2139062144;
and.b32 %r728, %r616, 2139062143;
and.b32 %r729, %r713, -2139062144;
sub.s32 %r730, %r727, %r728;
xor.b32 %r731, %r729, %r730;
xor.b32 %r621, %r731, -2139062144;
and.b32 %r732, %r722, -2143281136;
or.b32 %r733, %r732, -2139062144;
add.s32 %r734, %r733, -16843009;
or.b32 %r618, %r734, %r732;

prmt.b32 %r618, %r618, 0, 0xba98;

xor.b32 %r735, %r618, %r714;
or.b32 %r736, %r735, -2139062144;
and.b32 %r737, %r618, 2139062143;
and.b32 %r738, %r714, -2139062144;
sub.s32 %r739, %r736, %r737;
xor.b32 %r740, %r738, %r739;
xor.b32 %r625, %r740, -2139062144;

dp4a.s32.s32 %r620, %r621, %r15, %r612;


dp4a.s32.s32 %r624, %r625, %r16, %r620;

cvt.u32.u16 %r741, %rs56;
cvt.u32.u16 %r742, %rs57;
mul.lo.s32 %r743, %r600, %r741;
mad.lo.s32 %r744, %r624, %r742, %r743;
add.s32 %r745, %r624, %r600;
shr.u32 %r746, %r745, 31;
add.s32 %r747, %r745, %r746;
shr.s32 %r748, %r747, 1;
add.s32 %r749, %r744, %r748;
shr.s32 %r750, %r749, 31;
shr.u32 %r751, %r750, 30;
add.s32 %r752, %r749, %r751;
shr.s32 %r753, %r752, 2;
add.s64 %rd113, %rd150, %rd4;
ld.global.u16 %rs51, [%rd113];

{ cvt.f32.f16 %f54, %rs51;}



{.reg .f16 low,high;
mov.b32 {low,high},%r17;
cvt.f32.f16 %f55, low;}


mul.ftz.f32 %f56, %f54, %f55;
cvt.rn.f32.s32 %f57, %r753;
fma.rn.ftz.f32 %f162, %f56, %f57, %f162;

$L__BB211_7:
add.s32 %r849, %r849, 32;
add.s64 %rd150, %rd150, 296;
add.s64 %rd149, %rd149, 296;
add.s32 %r850, %r850, 4;
setp.lt.u32 %p5, %r850, %r38;
@%p5 bra $L__BB211_3;

add.u64 %rd141, %SPL, 0;
add.u64 %rd135, %SPL, 8;
st.local.v2.f32 [%rd141], {%f165, %f164};
st.local.v2.f32 [%rd135], {%f163, %f162};

$L__BB211_9:
mov.b32 %r20, %f165;
setp.ne.s64 %p6, %rd24, 0;
@%p6 bra $L__BB211_11;
bra.uni $L__BB211_10;

$L__BB211_11:
add.u64 %rd147, %SPL, 0;
add.u64 %rd139, %SPL, 8;
mov.u32 %r781, 31;
mov.u32 %r782, 16;
mov.u32 %r783, -1;
shfl.sync.bfly.b32 %r784|%p17, %r20, %r782, %r781, %r783;
mov.b32 %f78, %r784;
add.ftz.f32 %f79, %f165, %f78;
mov.b32 %r785, %f79;
mov.u32 %r786, 8;
shfl.sync.bfly.b32 %r787|%p18, %r785, %r786, %r781, %r783;
mov.b32 %f80, %r787;
add.ftz.f32 %f81, %f79, %f80;
mov.b32 %r788, %f81;
mov.u32 %r789, 4;
shfl.sync.bfly.b32 %r790|%p19, %r788, %r789, %r781, %r783;
mov.b32 %f82, %r790;
add.ftz.f32 %f83, %f81, %f82;
mov.b32 %r791, %f83;
mov.u32 %r792, 2;
shfl.sync.bfly.b32 %r793|%p20, %r791, %r792, %r781, %r783;
mov.b32 %f84, %r793;
add.ftz.f32 %f85, %f83, %f84;
mov.b32 %r794, %f85;
mov.u32 %r795, 1;
shfl.sync.bfly.b32 %r796|%p21, %r794, %r795, %r781, %r783;
mov.b32 %f86, %r796;
add.ftz.f32 %f87, %f85, %f86;
st.local.f32 [%rd147], %f87;
mov.b32 %r797, %f163;
shfl.sync.bfly.b32 %r798|%p22, %r797, %r782, %r781, %r783;
mov.b32 %f88, %r798;
add.ftz.f32 %f89, %f163, %f88;
mov.b32 %r799, %f89;
shfl.sync.bfly.b32 %r800|%p23, %r799, %r786, %r781, %r783;
mov.b32 %f90, %r800;
add.ftz.f32 %f91, %f89, %f90;
mov.b32 %r801, %f91;
shfl.sync.bfly.b32 %r802|%p24, %r801, %r789, %r781, %r783;
mov.b32 %f92, %r802;
add.ftz.f32 %f93, %f91, %f92;
mov.b32 %r803, %f93;
shfl.sync.bfly.b32 %r804|%p25, %r803, %r792, %r781, %r783;
mov.b32 %f94, %r804;
add.ftz.f32 %f95, %f93, %f94;
mov.b32 %r805, %f95;
shfl.sync.bfly.b32 %r806|%p26, %r805, %r795, %r781, %r783;
mov.b32 %f96, %r806;
add.ftz.f32 %f97, %f95, %f96;
st.local.f32 [%rd139], %f97;
mov.b32 %r807, %f164;
shfl.sync.bfly.b32 %r808|%p27, %r807, %r782, %r781, %r783;
mov.b32 %f98, %r808;
add.ftz.f32 %f99, %f164, %f98;
mov.b32 %r809, %f99;
shfl.sync.bfly.b32 %r810|%p28, %r809, %r786, %r781, %r783;
mov.b32 %f100, %r810;
add.ftz.f32 %f101, %f99, %f100;
mov.b32 %r811, %f101;
shfl.sync.bfly.b32 %r812|%p29, %r811, %r789, %r781, %r783;
mov.b32 %f102, %r812;
add.ftz.f32 %f103, %f101, %f102;
mov.b32 %r813, %f103;
shfl.sync.bfly.b32 %r814|%p30, %r813, %r792, %r781, %r783;
mov.b32 %f104, %r814;
add.ftz.f32 %f105, %f103, %f104;
mov.b32 %r815, %f105;
shfl.sync.bfly.b32 %r816|%p31, %r815, %r795, %r781, %r783;
mov.b32 %f106, %r816;
add.ftz.f32 %f107, %f105, %f106;
st.local.f32 [%rd147+4], %f107;
mov.b32 %r817, %f162;
shfl.sync.bfly.b32 %r818|%p32, %r817, %r782, %r781, %r783;
mov.b32 %f108, %r818;
add.ftz.f32 %f109, %f162, %f108;
mov.b32 %r819, %f109;
shfl.sync.bfly.b32 %r820|%p33, %r819, %r786, %r781, %r783;
mov.b32 %f110, %r820;
add.ftz.f32 %f111, %f109, %f110;
mov.b32 %r821, %f111;
shfl.sync.bfly.b32 %r822|%p34, %r821, %r789, %r781, %r783;
mov.b32 %f112, %r822;
add.ftz.f32 %f113, %f111, %f112;
mov.b32 %r823, %f113;
shfl.sync.bfly.b32 %r824|%p35, %r823, %r792, %r781, %r783;
mov.b32 %f114, %r824;
add.ftz.f32 %f115, %f113, %f114;
mov.b32 %r825, %f115;
shfl.sync.bfly.b32 %r826|%p36, %r825, %r795, %r781, %r783;
mov.b32 %f116, %r826;
add.ftz.f32 %f117, %f115, %f116;
st.local.f32 [%rd139+4], %f117;
bra.uni $L__BB211_12;

$L__BB211_10:
add.u64 %rd143, %SPL, 0;
mov.u32 %r755, 31;
mov.u32 %r756, 16;
mov.u32 %r757, -1;
shfl.sync.bfly.b32 %r758|%p7, %r20, %r756, %r755, %r757;
mov.b32 %f58, %r758;
add.ftz.f32 %f59, %f165, %f58;
mov.b32 %r759, %f59;
mov.u32 %r760, 8;
shfl.sync.bfly.b32 %r761|%p8, %r759, %r760, %r755, %r757;
mov.b32 %f60, %r761;
add.ftz.f32 %f61, %f59, %f60;
mov.b32 %r762, %f61;
mov.u32 %r763, 4;
shfl.sync.bfly.b32 %r764|%p9, %r762, %r763, %r755, %r757;
mov.b32 %f62, %r764;
add.ftz.f32 %f63, %f61, %f62;
mov.b32 %r765, %f63;
mov.u32 %r766, 2;
shfl.sync.bfly.b32 %r767|%p10, %r765, %r766, %r755, %r757;
mov.b32 %f64, %r767;
add.ftz.f32 %f65, %f63, %f64;
mov.b32 %r768, %f65;
mov.u32 %r769, 1;
shfl.sync.bfly.b32 %r770|%p11, %r768, %r769, %r755, %r757;
mov.b32 %f66, %r770;
add.ftz.f32 %f67, %f65, %f66;
st.local.f32 [%rd143], %f67;
mov.b32 %r771, %f164;
shfl.sync.bfly.b32 %r772|%p12, %r771, %r756, %r755, %r757;
mov.b32 %f68, %r772;
add.ftz.f32 %f69, %f164, %f68;
mov.b32 %r773, %f69;
shfl.sync.bfly.b32 %r774|%p13, %r773, %r760, %r755, %r757;
mov.b32 %f70, %r774;
add.ftz.f32 %f71, %f69, %f70;
mov.b32 %r775, %f71;
shfl.sync.bfly.b32 %r776|%p14, %r775, %r763, %r755, %r757;
mov.b32 %f72, %r776;
add.ftz.f32 %f73, %f71, %f72;
mov.b32 %r777, %f73;
shfl.sync.bfly.b32 %r778|%p15, %r777, %r766, %r755, %r757;
mov.b32 %f74, %r778;
add.ftz.f32 %f75, %f73, %f74;
mov.b32 %r779, %f75;
shfl.sync.bfly.b32 %r780|%p16, %r779, %r769, %r755, %r757;
mov.b32 %f76, %r780;
add.ftz.f32 %f77, %f75, %f76;
st.local.f32 [%rd143+4], %f77;

$L__BB211_12:
ld.param.u32 %r837, [_Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7];
mov.u32 %r836, %tid.x;
setp.gt.u32 %p37, %r836, 1;
mov.u32 %r828, %ctaid.x;
shl.b32 %r829, %r828, 1;
add.s32 %r21, %r829, %r836;
setp.ge.u32 %p38, %r21, %r837;
or.pred %p39, %p37, %p38;
@%p39 bra $L__BB211_29;

add.u64 %rd145, %SPL, 0;
ld.param.u64 %rd128, [_Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3];
ld.param.u32 %r839, [_Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13];
mov.u32 %r838, %tid.x;
cvt.u64.u32 %rd18, %r838;
mul.wide.u32 %rd114, %r838, 4;
add.s64 %rd115, %rd145, %rd114;
ld.local.f32 %f168, [%rd115];
mad.lo.s32 %r22, %r3, %r839, %r21;
setp.eq.s64 %p40, %rd128, 0;
@%p40 bra $L__BB211_15;

ld.param.u64 %rd129, [_Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3];
cvta.to.global.u64 %rd116, %rd129;
mul.wide.u32 %rd117, %r22, 4;
add.s64 %rd118, %rd116, %rd117;
ld.global.f32 %f118, [%rd118];
add.ftz.f32 %f168, %f168, %f118;

$L__BB211_15:
setp.eq.s64 %p41, %rd24, 0;
@%p41 bra $L__BB211_28;

add.u64 %rd137, %SPL, 8;
ld.param.u64 %rd131, [_Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+16];
shl.b64 %rd119, %rd18, 2;
add.s64 %rd120, %rd137, %rd119;
ld.local.f32 %f24, [%rd120];
setp.eq.s64 %p42, %rd131, 0;
@%p42 bra $L__BB211_18;

ld.param.u64 %rd132, [_Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+16];
cvta.to.global.u64 %rd121, %rd132;
mul.wide.u32 %rd122, %r22, 4;
add.s64 %rd123, %rd121, %rd122;
ld.global.f32 %f119, [%rd123];
add.ftz.f32 %f24, %f24, %f119;

$L__BB211_18:
ld.param.u32 %r844, [_Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.gt.s32 %p43, %r844, 2;
@%p43 bra $L__BB211_22;

ld.param.u32 %r847, [_Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.eq.s32 %p46, %r847, 1;
@%p46 bra $L__BB211_26;

ld.param.u32 %r848, [_Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.eq.s32 %p47, %r848, 2;
@%p47 bra $L__BB211_21;
bra.uni $L__BB211_27;

$L__BB211_21:
mul.ftz.f32 %f139, %f24, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f140, %f139;
add.ftz.f32 %f141, %f140, 0f3F800000;
div.approx.ftz.f32 %f142, %f24, %f141;
mul.ftz.f32 %f168, %f168, %f142;
bra.uni $L__BB211_28;

$L__BB211_22:
ld.param.u32 %r845, [_Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.eq.s32 %p44, %r845, 3;
@%p44 bra $L__BB211_25;

ld.param.u32 %r846, [_Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.ne.s32 %p45, %r846, 6;
@%p45 bra $L__BB211_27;

ld.param.f32 %f151, [_Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+44];
min.ftz.f32 %f120, %f24, %f151;
neg.ftz.f32 %f121, %f151;
min.ftz.f32 %f122, %f168, %f151;
max.ftz.f32 %f123, %f122, %f121;
mul.ftz.f32 %f124, %f120, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f125, %f124;
add.ftz.f32 %f126, %f125, 0f3F800000;
div.approx.ftz.f32 %f127, %f120, %f126;
mul.ftz.f32 %f168, %f123, %f127;
bra.uni $L__BB211_28;

$L__BB211_27:
mul.ftz.f32 %f168, %f168, %f24;
bra.uni $L__BB211_28;

$L__BB211_26:
mul.ftz.f32 %f143, %f24, 0f3F000000;
mul.ftz.f32 %f144, %f24, 0f3D372713;
fma.rn.ftz.f32 %f145, %f24, %f144, 0f3F800000;
mul.ftz.f32 %f146, %f24, 0f3F4C422A;
mul.ftz.f32 %f147, %f146, %f145;
tanh.approx.f32 %f148, %f147;
add.ftz.f32 %f149, %f148, 0f3F800000;
mul.ftz.f32 %f150, %f143, %f149;
mul.ftz.f32 %f168, %f168, %f150;
bra.uni $L__BB211_28;

$L__BB211_25:
mov.f32 %f128, 0f40E00000;
min.ftz.f32 %f129, %f24, %f128;
min.ftz.f32 %f130, %f168, %f128;
mov.f32 %f131, 0fC0E00000;
max.ftz.f32 %f132, %f130, %f131;
mul.ftz.f32 %f133, %f129, 0f3FD9DB23;
mul.ftz.f32 %f134, %f133, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f135, %f134;
add.ftz.f32 %f136, %f135, 0f3F800000;
div.approx.ftz.f32 %f137, %f129, %f136;
add.ftz.f32 %f138, %f132, 0f3F800000;
mul.ftz.f32 %f168, %f138, %f137;

$L__BB211_28:
ld.param.u64 %rd130, [_Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4];
ld.param.u32 %r843, [_Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10];
ld.param.u32 %r842, [_Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13];
mov.u32 %r841, %ctaid.y;
mov.u32 %r840, %tid.y;
mad.lo.s32 %r834, %r840, %r843, %r21;
mad.lo.s32 %r835, %r841, %r842, %r834;
cvta.to.global.u64 %rd124, %rd130;
mul.wide.u32 %rd125, %r835, 4;
add.s64 %rd126, %rd124, %rd125;
st.global.f32 [%rd126], %f168;

$L__BB211_29:
ret;

}
.entry _Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj(
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2,
.param .align 8 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3[48],
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5,
.param .align 4 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6[12],
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15
)
.maxntid 256, 1, 1
.minnctapersm 1
{
.local .align 8 .b8 __local_depot212[8];
.reg .b64 %SP;
.reg .b64 %SPL;
.reg .pred %p<17>;
.reg .b16 %rs<35>;
.reg .f32 %f<44>;
.reg .b32 %r<425>;
.reg .b64 %rd<73>;


mov.u64 %SPL, __local_depot212;
ld.param.u64 %rd15, [_Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0];
ld.param.u64 %rd16, [_Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1];
ld.param.u64 %rd17, [_Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2];
ld.param.u64 %rd18, [_Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4];
ld.param.u32 %r23, [_Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5];
ld.param.u32 %r15, [_Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7];
ld.param.u32 %r16, [_Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8];
ld.param.u32 %r17, [_Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9];
ld.param.u32 %r18, [_Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10];
ld.param.u32 %r19, [_Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11];
ld.param.u32 %r20, [_Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12];
ld.param.u32 %r21, [_Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13];
ld.param.u32 %r24, [_Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14];
ld.param.u32 %r22, [_Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15];
ld.param.u32 %r14, [_Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+8];
ld.param.u32 %r13, [_Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+4];
ld.param.u32 %r12, [_Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6];
add.u64 %rd1, %SPL, 0;
mov.u32 %r25, %ctaid.x;
shl.b32 %r1, %r25, 1;
shr.u32 %r2, %r23, 8;
mov.u32 %r3, %tid.y;
setp.ge.u32 %p1, %r3, %r24;
@%p1 bra $L__BB212_7;

mov.f32 %f42, 0f00000000;
st.local.v2.f32 [%rd1], {%f42, %f42};
mov.u32 %r4, %tid.x;
shr.u32 %r424, %r4, 3;
setp.ge.u32 %p2, %r424, %r2;
mov.f32 %f43, %f42;
@%p2 bra $L__BB212_5;

mov.u32 %r26, %ctaid.y;
mad.lo.s32 %r27, %r3, %r22, %r26;
cvta.to.global.u64 %rd20, %rd17;
mul.wide.u32 %rd21, %r27, 4;
add.s64 %rd22, %rd20, %rd21;
ld.global.nc.u32 %r28, [%rd22];
mul.lo.s32 %r29, %r28, %r19;
and.b32 %r30, %r4, 7;
cvt.u64.u32 %rd3, %r30;
shl.b32 %r423, %r424, 3;
add.s32 %r31, %r1, 1;
add.s32 %r32, %r424, %r29;
mad.lo.s32 %r33, %r16, %r31, %r32;
mul.wide.s32 %rd4, %r33, 74;
add.s64 %rd23, %rd4, %rd3;
add.s64 %rd5, %rd23, 66;
mad.lo.s32 %r34, %r1, %r16, %r424;
add.s32 %r35, %r34, %r29;
mul.wide.s32 %rd6, %r35, 74;
add.s64 %rd24, %rd6, %rd3;
add.s64 %rd7, %rd24, 66;
shl.b32 %r36, %r4, 3;
cvt.u64.u32 %rd25, %r36;
and.b64 %rd26, %rd25, 56;
add.s64 %rd27, %rd26, %rd4;
add.s64 %rd8, %rd27, 4;
add.s64 %rd28, %rd26, %rd6;
add.s64 %rd9, %rd28, 8;
cvta.to.global.u64 %rd72, %rd15;
mul.lo.s32 %r37, %r3, %r17;
cvt.u64.u32 %rd29, %r37;
mul.hi.u32 %r38, %r26, %r12;
add.s32 %r39, %r38, %r26;
shr.u32 %r40, %r39, %r13;
mul.lo.s32 %r41, %r40, %r14;
sub.s32 %r42, %r26, %r41;
mul.lo.s32 %r43, %r42, %r20;
cvt.u64.u32 %rd30, %r43;
add.s64 %rd11, %rd30, %rd29;
cvta.to.global.u64 %rd12, %rd16;
mov.f32 %f42, 0f00000000;

$L__BB212_3:
cvt.s64.s32 %rd31, %r423;
add.s64 %rd32, %rd11, %rd31;
add.s64 %rd33, %rd32, %rd3;
mul.lo.s64 %rd34, %rd33, 36;
add.s64 %rd35, %rd72, %rd9;
ld.global.nc.u16 %rs3, [%rd35+-4];
ld.global.nc.u16 %rs4, [%rd35+-2];
ld.global.nc.u16 %rs5, [%rd35];
add.s64 %rd36, %rd72, %rd7;
ld.global.nc.u8 %rs6, [%rd36];
and.b16 %rs7, %rs6, 240;
and.b16 %rs8, %rs6, 15;
shr.u16 %rs9, %rs7, 4;
ld.global.nc.u16 %rs10, [%rd35+-6];
shl.b16 %rs11, %rs10, 3;
cvt.u64.u16 %rd37, %rs11;
and.b64 %rd38, %rd37, 4088;
mov.u64 %rd39, iq2xs_grid;
add.s64 %rd40, %rd39, %rd38;
ld.global.nc.v2.u32 {%r142, %r143}, [%rd40];
mov.u32 %r124, 0;
shr.u16 %rs12, %rs10, 9;
cvt.u32.u16 %r146, %rs12;
popc.b32 %r147, %r146;
shl.b32 %r148, %r147, 7;
and.b32 %r149, %r148, 128;
or.b32 %r150, %r149, %r146;
mul.lo.s32 %r151, %r150, 16843009;
and.b32 %r152, %r151, 134480385;
or.b32 %r153, %r152, -2139062144;
add.s32 %r154, %r153, -16843009;
or.b32 %r44, %r154, %r152;

prmt.b32 %r44, %r44, 0, 0xba98;

xor.b32 %r155, %r44, %r142;
or.b32 %r156, %r155, -2139062144;
and.b32 %r157, %r44, 2139062143;
and.b32 %r158, %r142, -2139062144;
sub.s32 %r159, %r156, %r157;
xor.b32 %r160, %r158, %r159;
xor.b32 %r49, %r160, -2139062144;
add.s64 %rd41, %rd12, %rd34;
ld.global.nc.u32 %r50, [%rd41+4];
and.b32 %r161, %r151, -2143281136;
or.b32 %r162, %r161, -2139062144;
add.s32 %r163, %r162, -16843009;
or.b32 %r46, %r163, %r161;

prmt.b32 %r46, %r46, 0, 0xba98;

xor.b32 %r164, %r46, %r143;
or.b32 %r165, %r164, -2139062144;
and.b32 %r166, %r46, 2139062143;
and.b32 %r167, %r143, -2139062144;
sub.s32 %r168, %r165, %r166;
xor.b32 %r169, %r167, %r168;
xor.b32 %r53, %r169, -2139062144;
ld.global.nc.u32 %r54, [%rd41+8];

dp4a.s32.s32 %r48, %r49, %r50, %r124;


dp4a.s32.s32 %r52, %r53, %r54, %r48;

shl.b16 %rs13, %rs3, 3;
cvt.u64.u16 %rd42, %rs13;
and.b64 %rd43, %rd42, 4088;
add.s64 %rd44, %rd39, %rd43;
ld.global.nc.v2.u32 {%r170, %r171}, [%rd44];
shr.u16 %rs14, %rs3, 9;
cvt.u32.u16 %r174, %rs14;
popc.b32 %r175, %r174;
shl.b32 %r176, %r175, 7;
and.b32 %r177, %r176, 128;
or.b32 %r178, %r177, %r174;
mul.lo.s32 %r179, %r178, 16843009;
and.b32 %r180, %r179, 134480385;
or.b32 %r181, %r180, -2139062144;
add.s32 %r182, %r181, -16843009;
or.b32 %r56, %r182, %r180;

prmt.b32 %r56, %r56, 0, 0xba98;

xor.b32 %r183, %r56, %r170;
or.b32 %r184, %r183, -2139062144;
and.b32 %r185, %r56, 2139062143;
and.b32 %r186, %r170, -2139062144;
sub.s32 %r187, %r184, %r185;
xor.b32 %r188, %r186, %r187;
xor.b32 %r61, %r188, -2139062144;
ld.global.nc.u32 %r62, [%rd41+12];
and.b32 %r189, %r179, -2143281136;
or.b32 %r190, %r189, -2139062144;
add.s32 %r191, %r190, -16843009;
or.b32 %r58, %r191, %r189;

prmt.b32 %r58, %r58, 0, 0xba98;

xor.b32 %r192, %r58, %r171;
or.b32 %r193, %r192, -2139062144;
and.b32 %r194, %r58, 2139062143;
and.b32 %r195, %r171, -2139062144;
sub.s32 %r196, %r193, %r194;
xor.b32 %r197, %r195, %r196;
xor.b32 %r65, %r197, -2139062144;
ld.global.nc.u32 %r66, [%rd41+16];

dp4a.s32.s32 %r60, %r61, %r62, %r52;


dp4a.s32.s32 %r64, %r65, %r66, %r60;

shl.b16 %rs15, %rs4, 3;
cvt.u64.u16 %rd45, %rs15;
and.b64 %rd46, %rd45, 4088;
add.s64 %rd47, %rd39, %rd46;
ld.global.nc.v2.u32 {%r198, %r199}, [%rd47];
shr.u16 %rs16, %rs4, 9;
cvt.u32.u16 %r202, %rs16;
popc.b32 %r203, %r202;
shl.b32 %r204, %r203, 7;
and.b32 %r205, %r204, 128;
or.b32 %r206, %r205, %r202;
mul.lo.s32 %r207, %r206, 16843009;
and.b32 %r208, %r207, 134480385;
or.b32 %r209, %r208, -2139062144;
add.s32 %r210, %r209, -16843009;
or.b32 %r68, %r210, %r208;

prmt.b32 %r68, %r68, 0, 0xba98;

xor.b32 %r211, %r68, %r198;
or.b32 %r212, %r211, -2139062144;
and.b32 %r213, %r68, 2139062143;
and.b32 %r214, %r198, -2139062144;
sub.s32 %r215, %r212, %r213;
xor.b32 %r216, %r214, %r215;
xor.b32 %r73, %r216, -2139062144;
ld.global.nc.u32 %r74, [%rd41+20];
and.b32 %r217, %r207, -2143281136;
or.b32 %r218, %r217, -2139062144;
add.s32 %r219, %r218, -16843009;
or.b32 %r70, %r219, %r217;

prmt.b32 %r70, %r70, 0, 0xba98;

xor.b32 %r220, %r70, %r199;
or.b32 %r221, %r220, -2139062144;
and.b32 %r222, %r70, 2139062143;
and.b32 %r223, %r199, -2139062144;
sub.s32 %r224, %r221, %r222;
xor.b32 %r225, %r223, %r224;
xor.b32 %r77, %r225, -2139062144;
ld.global.nc.u32 %r78, [%rd41+24];

dp4a.s32.s32 %r72, %r73, %r74, %r124;


dp4a.s32.s32 %r76, %r77, %r78, %r72;

shl.b16 %rs17, %rs5, 3;
cvt.u64.u16 %rd48, %rs17;
and.b64 %rd49, %rd48, 4088;
add.s64 %rd50, %rd39, %rd49;
ld.global.nc.v2.u32 {%r226, %r227}, [%rd50];
shr.u16 %rs18, %rs5, 9;
cvt.u32.u16 %r230, %rs18;
popc.b32 %r231, %r230;
shl.b32 %r232, %r231, 7;
and.b32 %r233, %r232, 128;
or.b32 %r234, %r233, %r230;
mul.lo.s32 %r235, %r234, 16843009;
and.b32 %r236, %r235, 134480385;
or.b32 %r237, %r236, -2139062144;
add.s32 %r238, %r237, -16843009;
or.b32 %r80, %r238, %r236;

prmt.b32 %r80, %r80, 0, 0xba98;

xor.b32 %r239, %r80, %r226;
or.b32 %r240, %r239, -2139062144;
and.b32 %r241, %r80, 2139062143;
and.b32 %r242, %r226, -2139062144;
sub.s32 %r243, %r240, %r241;
xor.b32 %r244, %r242, %r243;
xor.b32 %r85, %r244, -2139062144;
ld.global.nc.u32 %r86, [%rd41+28];
and.b32 %r245, %r235, -2143281136;
or.b32 %r246, %r245, -2139062144;
add.s32 %r247, %r246, -16843009;
or.b32 %r82, %r247, %r245;

prmt.b32 %r82, %r82, 0, 0xba98;

xor.b32 %r248, %r82, %r227;
or.b32 %r249, %r248, -2139062144;
and.b32 %r250, %r82, 2139062143;
and.b32 %r251, %r227, -2139062144;
sub.s32 %r252, %r249, %r250;
xor.b32 %r253, %r251, %r252;
xor.b32 %r89, %r253, -2139062144;
ld.global.nc.u32 %r90, [%rd41+32];

dp4a.s32.s32 %r84, %r85, %r86, %r76;


dp4a.s32.s32 %r88, %r89, %r90, %r84;

cvt.u32.u16 %r254, %rs8;
cvt.u32.u16 %r255, %rs9;
mul.lo.s32 %r256, %r64, %r254;
mad.lo.s32 %r257, %r88, %r255, %r256;
add.s32 %r258, %r88, %r64;
shr.u32 %r259, %r258, 31;
add.s32 %r260, %r258, %r259;
shr.s32 %r261, %r260, 1;
add.s32 %r262, %r257, %r261;
shr.s32 %r263, %r262, 31;
shr.u32 %r264, %r263, 30;
add.s32 %r265, %r262, %r264;
shr.s32 %r266, %r265, 2;
add.s64 %rd51, %rd72, %rd6;
ld.global.nc.u16 %rs1, [%rd51];

{ cvt.f32.f16 %f11, %rs1;}


ld.global.nc.u32 %r92, [%rd41];

{.reg .f16 low,high;
mov.b32 {low,high},%r92;
cvt.f32.f16 %f12, low;}


mul.ftz.f32 %f15, %f11, %f12;
cvt.rn.f32.s32 %f16, %r266;
fma.rn.ftz.f32 %f43, %f15, %f16, %f43;
add.s64 %rd52, %rd72, %rd8;
ld.global.nc.u16 %rs19, [%rd52];
ld.global.nc.u16 %rs20, [%rd52+2];
ld.global.nc.u16 %rs21, [%rd52+4];
add.s64 %rd53, %rd72, %rd5;
ld.global.nc.u8 %rs22, [%rd53];
and.b16 %rs23, %rs22, 240;
and.b16 %rs24, %rs22, 15;
shr.u16 %rs25, %rs23, 4;
ld.global.nc.u16 %rs26, [%rd52+-2];
shl.b16 %rs27, %rs26, 3;
cvt.u64.u16 %rd54, %rs27;
and.b64 %rd55, %rd54, 4088;
add.s64 %rd56, %rd39, %rd55;
ld.global.nc.v2.u32 {%r267, %r268}, [%rd56];
shr.u16 %rs28, %rs26, 9;
cvt.u32.u16 %r271, %rs28;
popc.b32 %r272, %r271;
shl.b32 %r273, %r272, 7;
and.b32 %r274, %r273, 128;
or.b32 %r275, %r274, %r271;
mul.lo.s32 %r276, %r275, 16843009;
and.b32 %r277, %r276, 134480385;
or.b32 %r278, %r277, -2139062144;
add.s32 %r279, %r278, -16843009;
or.b32 %r93, %r279, %r277;

prmt.b32 %r93, %r93, 0, 0xba98;

xor.b32 %r280, %r93, %r267;
or.b32 %r281, %r280, -2139062144;
and.b32 %r282, %r93, 2139062143;
and.b32 %r283, %r267, -2139062144;
sub.s32 %r284, %r281, %r282;
xor.b32 %r285, %r283, %r284;
xor.b32 %r98, %r285, -2139062144;
and.b32 %r286, %r276, -2143281136;
or.b32 %r287, %r286, -2139062144;
add.s32 %r288, %r287, -16843009;
or.b32 %r95, %r288, %r286;

prmt.b32 %r95, %r95, 0, 0xba98;

xor.b32 %r289, %r95, %r268;
or.b32 %r290, %r289, -2139062144;
and.b32 %r291, %r95, 2139062143;
and.b32 %r292, %r268, -2139062144;
sub.s32 %r293, %r290, %r291;
xor.b32 %r294, %r292, %r293;
xor.b32 %r102, %r294, -2139062144;

dp4a.s32.s32 %r97, %r98, %r50, %r124;


dp4a.s32.s32 %r101, %r102, %r54, %r97;

shl.b16 %rs29, %rs19, 3;
cvt.u64.u16 %rd57, %rs29;
and.b64 %rd58, %rd57, 4088;
add.s64 %rd59, %rd39, %rd58;
ld.global.nc.v2.u32 {%r295, %r296}, [%rd59];
shr.u16 %rs30, %rs19, 9;
cvt.u32.u16 %r299, %rs30;
popc.b32 %r300, %r299;
shl.b32 %r301, %r300, 7;
and.b32 %r302, %r301, 128;
or.b32 %r303, %r302, %r299;
mul.lo.s32 %r304, %r303, 16843009;
and.b32 %r305, %r304, 134480385;
or.b32 %r306, %r305, -2139062144;
add.s32 %r307, %r306, -16843009;
or.b32 %r105, %r307, %r305;

prmt.b32 %r105, %r105, 0, 0xba98;

xor.b32 %r308, %r105, %r295;
or.b32 %r309, %r308, -2139062144;
and.b32 %r310, %r105, 2139062143;
and.b32 %r311, %r295, -2139062144;
sub.s32 %r312, %r309, %r310;
xor.b32 %r313, %r311, %r312;
xor.b32 %r110, %r313, -2139062144;
and.b32 %r314, %r304, -2143281136;
or.b32 %r315, %r314, -2139062144;
add.s32 %r316, %r315, -16843009;
or.b32 %r107, %r316, %r314;

prmt.b32 %r107, %r107, 0, 0xba98;

xor.b32 %r317, %r107, %r296;
or.b32 %r318, %r317, -2139062144;
and.b32 %r319, %r107, 2139062143;
and.b32 %r320, %r296, -2139062144;
sub.s32 %r321, %r318, %r319;
xor.b32 %r322, %r320, %r321;
xor.b32 %r114, %r322, -2139062144;

dp4a.s32.s32 %r109, %r110, %r62, %r101;


dp4a.s32.s32 %r113, %r114, %r66, %r109;

shl.b16 %rs31, %rs20, 3;
cvt.u64.u16 %rd60, %rs31;
and.b64 %rd61, %rd60, 4088;
add.s64 %rd62, %rd39, %rd61;
ld.global.nc.v2.u32 {%r323, %r324}, [%rd62];
shr.u16 %rs32, %rs20, 9;
cvt.u32.u16 %r327, %rs32;
popc.b32 %r328, %r327;
shl.b32 %r329, %r328, 7;
and.b32 %r330, %r329, 128;
or.b32 %r331, %r330, %r327;
mul.lo.s32 %r332, %r331, 16843009;
and.b32 %r333, %r332, 134480385;
or.b32 %r334, %r333, -2139062144;
add.s32 %r335, %r334, -16843009;
or.b32 %r117, %r335, %r333;

prmt.b32 %r117, %r117, 0, 0xba98;

xor.b32 %r336, %r117, %r323;
or.b32 %r337, %r336, -2139062144;
and.b32 %r338, %r117, 2139062143;
and.b32 %r339, %r323, -2139062144;
sub.s32 %r340, %r337, %r338;
xor.b32 %r341, %r339, %r340;
xor.b32 %r122, %r341, -2139062144;
and.b32 %r342, %r332, -2143281136;
or.b32 %r343, %r342, -2139062144;
add.s32 %r344, %r343, -16843009;
or.b32 %r119, %r344, %r342;

prmt.b32 %r119, %r119, 0, 0xba98;

xor.b32 %r345, %r119, %r324;
or.b32 %r346, %r345, -2139062144;
and.b32 %r347, %r119, 2139062143;
and.b32 %r348, %r324, -2139062144;
sub.s32 %r349, %r346, %r347;
xor.b32 %r350, %r348, %r349;
xor.b32 %r126, %r350, -2139062144;

dp4a.s32.s32 %r121, %r122, %r74, %r124;


dp4a.s32.s32 %r125, %r126, %r78, %r121;

shl.b16 %rs33, %rs21, 3;
cvt.u64.u16 %rd63, %rs33;
and.b64 %rd64, %rd63, 4088;
add.s64 %rd65, %rd39, %rd64;
ld.global.nc.v2.u32 {%r351, %r352}, [%rd65];
shr.u16 %rs34, %rs21, 9;
cvt.u32.u16 %r355, %rs34;
popc.b32 %r356, %r355;
shl.b32 %r357, %r356, 7;
and.b32 %r358, %r357, 128;
or.b32 %r359, %r358, %r355;
mul.lo.s32 %r360, %r359, 16843009;
and.b32 %r361, %r360, 134480385;
or.b32 %r362, %r361, -2139062144;
add.s32 %r363, %r362, -16843009;
or.b32 %r129, %r363, %r361;

prmt.b32 %r129, %r129, 0, 0xba98;

xor.b32 %r364, %r129, %r351;
or.b32 %r365, %r364, -2139062144;
and.b32 %r366, %r129, 2139062143;
and.b32 %r367, %r351, -2139062144;
sub.s32 %r368, %r365, %r366;
xor.b32 %r369, %r367, %r368;
xor.b32 %r134, %r369, -2139062144;
and.b32 %r370, %r360, -2143281136;
or.b32 %r371, %r370, -2139062144;
add.s32 %r372, %r371, -16843009;
or.b32 %r131, %r372, %r370;

prmt.b32 %r131, %r131, 0, 0xba98;

xor.b32 %r373, %r131, %r352;
or.b32 %r374, %r373, -2139062144;
and.b32 %r375, %r131, 2139062143;
and.b32 %r376, %r352, -2139062144;
sub.s32 %r377, %r374, %r375;
xor.b32 %r378, %r376, %r377;
xor.b32 %r138, %r378, -2139062144;

dp4a.s32.s32 %r133, %r134, %r86, %r125;


dp4a.s32.s32 %r137, %r138, %r90, %r133;

cvt.u32.u16 %r379, %rs24;
cvt.u32.u16 %r380, %rs25;
mul.lo.s32 %r381, %r113, %r379;
mad.lo.s32 %r382, %r137, %r380, %r381;
add.s32 %r383, %r137, %r113;
shr.u32 %r384, %r383, 31;
add.s32 %r385, %r383, %r384;
shr.s32 %r386, %r385, 1;
add.s32 %r387, %r382, %r386;
shr.s32 %r388, %r387, 31;
shr.u32 %r389, %r388, 30;
add.s32 %r390, %r387, %r389;
shr.s32 %r391, %r390, 2;
add.s64 %rd66, %rd72, %rd4;
ld.global.nc.u16 %rs2, [%rd66];

{ cvt.f32.f16 %f13, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r92;
cvt.f32.f16 %f14, low;}


mul.ftz.f32 %f17, %f13, %f14;
cvt.rn.f32.s32 %f18, %r391;
fma.rn.ftz.f32 %f42, %f17, %f18, %f42;
add.s32 %r423, %r423, 32;
add.s64 %rd72, %rd72, 296;
add.s32 %r424, %r424, 4;
setp.lt.u32 %p3, %r424, %r2;
@%p3 bra $L__BB212_3;

st.local.v2.f32 [%rd1], {%f43, %f42};

$L__BB212_5:
mov.b32 %r392, %f43;
mov.u32 %r393, 31;
mov.u32 %r394, 16;
mov.u32 %r395, -1;
shfl.sync.bfly.b32 %r396|%p4, %r392, %r394, %r393, %r395;
mov.b32 %f19, %r396;
add.ftz.f32 %f20, %f43, %f19;
mov.b32 %r397, %f20;
mov.u32 %r398, 8;
shfl.sync.bfly.b32 %r399|%p5, %r397, %r398, %r393, %r395;
mov.b32 %f21, %r399;
add.ftz.f32 %f22, %f20, %f21;
mov.b32 %r400, %f22;
mov.u32 %r401, 4;
shfl.sync.bfly.b32 %r402|%p6, %r400, %r401, %r393, %r395;
mov.b32 %f23, %r402;
add.ftz.f32 %f24, %f22, %f23;
mov.b32 %r403, %f24;
mov.u32 %r404, 2;
shfl.sync.bfly.b32 %r405|%p7, %r403, %r404, %r393, %r395;
mov.b32 %f25, %r405;
add.ftz.f32 %f26, %f24, %f25;
mov.b32 %r406, %f26;
mov.u32 %r407, 1;
shfl.sync.bfly.b32 %r408|%p8, %r406, %r407, %r393, %r395;
mov.b32 %f27, %r408;
add.ftz.f32 %f28, %f26, %f27;
st.local.f32 [%rd1], %f28;
mov.b32 %r409, %f42;
shfl.sync.bfly.b32 %r410|%p9, %r409, %r394, %r393, %r395;
mov.b32 %f29, %r410;
add.ftz.f32 %f30, %f42, %f29;
mov.b32 %r411, %f30;
shfl.sync.bfly.b32 %r412|%p10, %r411, %r398, %r393, %r395;
mov.b32 %f31, %r412;
add.ftz.f32 %f32, %f30, %f31;
mov.b32 %r413, %f32;
shfl.sync.bfly.b32 %r414|%p11, %r413, %r401, %r393, %r395;
mov.b32 %f33, %r414;
add.ftz.f32 %f34, %f32, %f33;
mov.b32 %r415, %f34;
shfl.sync.bfly.b32 %r416|%p12, %r415, %r404, %r393, %r395;
mov.b32 %f35, %r416;
add.ftz.f32 %f36, %f34, %f35;
mov.b32 %r417, %f36;
shfl.sync.bfly.b32 %r418|%p13, %r417, %r407, %r393, %r395;
mov.b32 %f37, %r418;
add.ftz.f32 %f38, %f36, %f37;
st.local.f32 [%rd1+4], %f38;
add.s32 %r11, %r1, %r4;
setp.ge.u32 %p14, %r11, %r15;
setp.gt.u32 %p15, %r4, 1;
or.pred %p16, %p15, %p14;
@%p16 bra $L__BB212_7;

mul.wide.u32 %rd67, %r4, 4;
add.s64 %rd68, %rd1, %rd67;
ld.local.f32 %f39, [%rd68];
mov.u32 %r419, %ctaid.y;
mad.lo.s32 %r421, %r3, %r18, %r11;
mad.lo.s32 %r422, %r419, %r21, %r421;
cvta.to.global.u64 %rd69, %rd18;
mul.wide.u32 %rd70, %r422, 4;
add.s64 %rd71, %rd69, %rd70;
st.global.f32 [%rd71], %f39;

$L__BB212_7:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<320>;
.reg .b16 %rs<133>;
.reg .f32 %f<1005>;
.reg .b32 %r<2161>;
.reg .b64 %rd<218>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate[1536];

ld.param.u64 %rd36, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd37, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd38, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u32 %r63, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r49, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r50, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r54, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r55, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r56, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r60, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r61, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r62, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r58, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r57, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r52, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r51, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r48, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r47, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r46, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
ld.param.u64 %rd41, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+16];
ld.param.u64 %rd40, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+8];
ld.param.u64 %rd39, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3];
mov.u32 %r1, %tid.y;
shl.b32 %r64, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r64, %r2;
mov.u32 %r65, %ctaid.x;
shl.b32 %r4, %r65, 2;
shr.u32 %r5, %r63, 8;
mov.u32 %r6, %ctaid.y;
setp.eq.s64 %p1, %rd38, 0;
@%p1 bra $L__BB213_2;

cvta.to.global.u64 %rd45, %rd38;
mul.wide.u32 %rd46, %r6, 4;
add.s64 %rd47, %rd45, %rd46;
ld.global.nc.u32 %r2157, [%rd47];
mul.hi.u32 %r66, %r6, %r46;
add.s32 %r67, %r66, %r6;
shr.u32 %r68, %r67, %r47;
mul.lo.s32 %r69, %r68, %r48;
sub.s32 %r2158, %r6, %r69;
bra.uni $L__BB213_3;

$L__BB213_2:
mul.hi.u32 %r70, %r6, %r51;
add.s32 %r71, %r70, %r6;
shr.u32 %r2157, %r71, %r52;
mov.u32 %r2158, %r6;

$L__BB213_3:
mov.u32 %r12, %ctaid.z;
selp.b32 %r13, %r6, %r2157, %p1;
setp.ne.s32 %p3, %r1, 0;
setp.gt.u32 %p4, %r2, 3;
or.pred %p5, %p4, %p3;
add.s32 %r72, %r4, %r2;
setp.ge.u32 %p6, %r72, %r50;
mov.f32 %f27, 0f00000000;
or.pred %p7, %p5, %p6;
mov.f32 %f4, %f27;
mov.f32 %f5, %f27;
@%p7 bra $L__BB213_8;

setp.eq.s64 %p8, %rd39, 0;
mov.f32 %f4, 0f00000000;
mov.f32 %f5, %f4;
@%p8 bra $L__BB213_6;

cvta.to.global.u64 %rd48, %rd39;
mul.lo.s32 %r73, %r12, %r62;
cvt.u64.u32 %rd49, %r73;
mul.lo.s32 %r74, %r13, %r56;
cvt.u64.u32 %rd50, %r74;
cvt.u64.u32 %rd51, %r2;
cvt.s64.s32 %rd52, %r4;
add.s64 %rd53, %rd52, %rd51;
add.s64 %rd54, %rd53, %rd49;
add.s64 %rd55, %rd54, %rd50;
shl.b64 %rd56, %rd55, 2;
add.s64 %rd57, %rd48, %rd56;
ld.global.f32 %f5, [%rd57];

$L__BB213_6:
setp.eq.s64 %p9, %rd41, 0;
setp.eq.s64 %p10, %rd40, 0;
or.pred %p11, %p9, %p10;
@%p11 bra $L__BB213_8;

cvta.to.global.u64 %rd58, %rd41;
mul.lo.s32 %r75, %r12, %r62;
cvt.u64.u32 %rd59, %r75;
mul.lo.s32 %r76, %r13, %r56;
cvt.u64.u32 %rd60, %r76;
cvt.u64.u32 %rd61, %r2;
cvt.s64.s32 %rd62, %r4;
add.s64 %rd63, %rd62, %rd61;
add.s64 %rd64, %rd63, %rd59;
add.s64 %rd65, %rd64, %rd60;
shl.b64 %rd66, %rd65, 2;
add.s64 %rd67, %rd58, %rd66;
ld.global.f32 %f4, [%rd67];

$L__BB213_8:
mul.lo.s32 %r77, %r12, %r61;
cvt.u64.u32 %rd68, %r77;
mul.lo.s32 %r78, %r2158, %r55;
cvt.u64.u32 %rd69, %r78;
add.s64 %rd4, %rd69, %rd68;
mul.hi.u32 %r79, %r12, %r57;
add.s32 %r80, %r79, %r12;
shr.u32 %r14, %r80, %r58;
shr.s32 %r81, %r3, 31;
shr.u32 %r82, %r81, 29;
add.s32 %r83, %r3, %r82;
shr.s32 %r2160, %r83, 3;
setp.ge.s32 %p12, %r2160, %r5;
mov.f32 %f28, %f27;
mov.f32 %f29, %f27;
mov.f32 %f30, %f27;
mov.f32 %f31, %f27;
mov.f32 %f32, %f27;
mov.f32 %f33, %f27;
mov.f32 %f34, %f27;
@%p12 bra $L__BB213_19;

mul.lo.s32 %r84, %r14, %r60;
mul.lo.s32 %r85, %r2157, %r54;
and.b32 %r89, %r83, -8;
sub.s32 %r90, %r3, %r89;
shl.b32 %r91, %r90, 2;
cvt.s64.s32 %rd5, %r90;
shl.b32 %r2159, %r2160, 3;
add.s32 %r92, %r2160, %r84;
add.s32 %r93, %r92, %r85;
add.s32 %r94, %r4, 1;
mad.lo.s32 %r95, %r49, %r94, %r93;
mul.wide.s32 %rd6, %r95, 74;
add.s64 %rd70, %rd6, %rd5;
add.s64 %rd7, %rd70, 66;
mad.lo.s32 %r96, %r4, %r49, %r2160;
add.s32 %r97, %r96, %r84;
add.s32 %r98, %r97, %r85;
mul.wide.s32 %rd8, %r98, 74;
add.s64 %rd71, %rd8, %rd5;
add.s64 %rd9, %rd71, 66;
mul.wide.s32 %rd72, %r91, 2;
add.s64 %rd73, %rd6, %rd72;
add.s64 %rd10, %rd73, 4;
add.s64 %rd74, %rd8, %rd72;
add.s64 %rd11, %rd74, 8;
add.s32 %r99, %r4, 2;
mad.lo.s32 %r100, %r49, %r99, %r93;
mul.wide.s32 %rd12, %r100, 74;
add.s64 %rd75, %rd12, %rd5;
add.s64 %rd13, %rd75, 66;
add.s32 %r101, %r4, 3;
mad.lo.s32 %r102, %r49, %r101, %r93;
mul.wide.s32 %rd14, %r102, 74;
add.s64 %rd76, %rd14, %rd72;
add.s64 %rd15, %rd76, 8;
add.s64 %rd77, %rd14, %rd5;
add.s64 %rd16, %rd77, 66;
add.s64 %rd78, %rd12, %rd72;
add.s64 %rd17, %rd78, 4;
mul.wide.s32 %rd18, %r49, 74;
cvta.to.global.u64 %rd217, %rd40;
cvta.to.global.u64 %rd216, %rd36;
cvta.to.global.u64 %rd21, %rd37;
mov.f32 %f27, 0f00000000;
mov.f32 %f28, %f27;
mov.f32 %f29, %f27;
mov.f32 %f30, %f27;
mov.f32 %f31, %f27;
mov.f32 %f32, %f27;
mov.f32 %f33, %f27;
mov.f32 %f34, %f27;

$L__BB213_10:
cvt.u64.u32 %rd79, %r2159;
add.s64 %rd80, %rd4, %rd79;
add.s64 %rd81, %rd216, %rd11;
ld.global.nc.u16 %rs2, [%rd81+-4];
ld.global.nc.u16 %rs3, [%rd81+-2];
ld.global.nc.u16 %rs4, [%rd81];
add.s64 %rd25, %rd216, %rd9;
ld.global.nc.u8 %rs5, [%rd25];
and.b16 %rs6, %rs5, 240;
and.b16 %rs7, %rs5, 15;
shr.u16 %rs8, %rs6, 4;
add.s64 %rd82, %rd80, %rd5;
mul.lo.s64 %rd83, %rd82, 36;
ld.global.nc.u16 %rs9, [%rd81+-6];
shl.b16 %rs10, %rs9, 3;
cvt.u64.u16 %rd84, %rs10;
and.b64 %rd85, %rd84, 4088;
mov.u64 %rd86, iq2xs_grid;
add.s64 %rd87, %rd86, %rd85;
ld.global.nc.v2.u32 {%r152, %r153}, [%rd87];
mov.u32 %r134, 0;
shr.u16 %rs11, %rs9, 9;
cvt.u32.u16 %r156, %rs11;
popc.b32 %r157, %r156;
shl.b32 %r158, %r157, 7;
and.b32 %r159, %r158, 128;
or.b32 %r160, %r159, %r156;
mul.lo.s32 %r161, %r160, 16843009;
and.b32 %r162, %r161, 134480385;
or.b32 %r163, %r162, -2139062144;
add.s32 %r164, %r163, -16843009;
or.b32 %r103, %r164, %r162;

prmt.b32 %r103, %r103, 0, 0xba98;

xor.b32 %r165, %r103, %r152;
or.b32 %r166, %r165, -2139062144;
and.b32 %r167, %r103, 2139062143;
and.b32 %r168, %r152, -2139062144;
sub.s32 %r169, %r166, %r167;
xor.b32 %r170, %r168, %r169;
xor.b32 %r108, %r170, -2139062144;
add.s64 %rd88, %rd21, %rd83;
ld.global.nc.u32 %r19, [%rd88+4];
and.b32 %r171, %r161, -2143281136;
or.b32 %r172, %r171, -2139062144;
add.s32 %r173, %r172, -16843009;
or.b32 %r105, %r173, %r171;

prmt.b32 %r105, %r105, 0, 0xba98;

xor.b32 %r174, %r105, %r153;
or.b32 %r175, %r174, -2139062144;
and.b32 %r176, %r105, 2139062143;
and.b32 %r177, %r153, -2139062144;
sub.s32 %r178, %r175, %r176;
xor.b32 %r179, %r177, %r178;
xor.b32 %r112, %r179, -2139062144;
ld.global.nc.u32 %r20, [%rd88+8];

dp4a.s32.s32 %r107, %r108, %r19, %r134;


dp4a.s32.s32 %r111, %r112, %r20, %r107;

shl.b16 %rs12, %rs2, 3;
cvt.u64.u16 %rd89, %rs12;
and.b64 %rd90, %rd89, 4088;
add.s64 %rd91, %rd86, %rd90;
ld.global.nc.v2.u32 {%r180, %r181}, [%rd91];
shr.u16 %rs13, %rs2, 9;
cvt.u32.u16 %r184, %rs13;
popc.b32 %r185, %r184;
shl.b32 %r186, %r185, 7;
and.b32 %r187, %r186, 128;
or.b32 %r188, %r187, %r184;
mul.lo.s32 %r189, %r188, 16843009;
and.b32 %r190, %r189, 134480385;
or.b32 %r191, %r190, -2139062144;
add.s32 %r192, %r191, -16843009;
or.b32 %r115, %r192, %r190;

prmt.b32 %r115, %r115, 0, 0xba98;

xor.b32 %r193, %r115, %r180;
or.b32 %r194, %r193, -2139062144;
and.b32 %r195, %r115, 2139062143;
and.b32 %r196, %r180, -2139062144;
sub.s32 %r197, %r194, %r195;
xor.b32 %r198, %r196, %r197;
xor.b32 %r120, %r198, -2139062144;
ld.global.nc.u32 %r21, [%rd88+12];
and.b32 %r199, %r189, -2143281136;
or.b32 %r200, %r199, -2139062144;
add.s32 %r201, %r200, -16843009;
or.b32 %r117, %r201, %r199;

prmt.b32 %r117, %r117, 0, 0xba98;

xor.b32 %r202, %r117, %r181;
or.b32 %r203, %r202, -2139062144;
and.b32 %r204, %r117, 2139062143;
and.b32 %r205, %r181, -2139062144;
sub.s32 %r206, %r203, %r204;
xor.b32 %r207, %r205, %r206;
xor.b32 %r124, %r207, -2139062144;
ld.global.nc.u32 %r22, [%rd88+16];

dp4a.s32.s32 %r119, %r120, %r21, %r111;


dp4a.s32.s32 %r123, %r124, %r22, %r119;

shl.b16 %rs14, %rs3, 3;
cvt.u64.u16 %rd92, %rs14;
and.b64 %rd93, %rd92, 4088;
add.s64 %rd94, %rd86, %rd93;
ld.global.nc.v2.u32 {%r208, %r209}, [%rd94];
shr.u16 %rs15, %rs3, 9;
cvt.u32.u16 %r212, %rs15;
popc.b32 %r213, %r212;
shl.b32 %r214, %r213, 7;
and.b32 %r215, %r214, 128;
or.b32 %r216, %r215, %r212;
mul.lo.s32 %r217, %r216, 16843009;
and.b32 %r218, %r217, 134480385;
or.b32 %r219, %r218, -2139062144;
add.s32 %r220, %r219, -16843009;
or.b32 %r127, %r220, %r218;

prmt.b32 %r127, %r127, 0, 0xba98;

xor.b32 %r221, %r127, %r208;
or.b32 %r222, %r221, -2139062144;
and.b32 %r223, %r127, 2139062143;
and.b32 %r224, %r208, -2139062144;
sub.s32 %r225, %r222, %r223;
xor.b32 %r226, %r224, %r225;
xor.b32 %r132, %r226, -2139062144;
ld.global.nc.u32 %r23, [%rd88+20];
and.b32 %r227, %r217, -2143281136;
or.b32 %r228, %r227, -2139062144;
add.s32 %r229, %r228, -16843009;
or.b32 %r129, %r229, %r227;

prmt.b32 %r129, %r129, 0, 0xba98;

xor.b32 %r230, %r129, %r209;
or.b32 %r231, %r230, -2139062144;
and.b32 %r232, %r129, 2139062143;
and.b32 %r233, %r209, -2139062144;
sub.s32 %r234, %r231, %r232;
xor.b32 %r235, %r233, %r234;
xor.b32 %r136, %r235, -2139062144;
ld.global.nc.u32 %r24, [%rd88+24];

dp4a.s32.s32 %r131, %r132, %r23, %r134;


dp4a.s32.s32 %r135, %r136, %r24, %r131;

shl.b16 %rs16, %rs4, 3;
cvt.u64.u16 %rd95, %rs16;
and.b64 %rd96, %rd95, 4088;
add.s64 %rd97, %rd86, %rd96;
ld.global.nc.v2.u32 {%r236, %r237}, [%rd97];
shr.u16 %rs17, %rs4, 9;
cvt.u32.u16 %r240, %rs17;
popc.b32 %r241, %r240;
shl.b32 %r242, %r241, 7;
and.b32 %r243, %r242, 128;
or.b32 %r244, %r243, %r240;
mul.lo.s32 %r245, %r244, 16843009;
and.b32 %r246, %r245, 134480385;
or.b32 %r247, %r246, -2139062144;
add.s32 %r248, %r247, -16843009;
or.b32 %r139, %r248, %r246;

prmt.b32 %r139, %r139, 0, 0xba98;

xor.b32 %r249, %r139, %r236;
or.b32 %r250, %r249, -2139062144;
and.b32 %r251, %r139, 2139062143;
and.b32 %r252, %r236, -2139062144;
sub.s32 %r253, %r250, %r251;
xor.b32 %r254, %r252, %r253;
xor.b32 %r144, %r254, -2139062144;
ld.global.nc.u32 %r25, [%rd88+28];
and.b32 %r255, %r245, -2143281136;
or.b32 %r256, %r255, -2139062144;
add.s32 %r257, %r256, -16843009;
or.b32 %r141, %r257, %r255;

prmt.b32 %r141, %r141, 0, 0xba98;

xor.b32 %r258, %r141, %r237;
or.b32 %r259, %r258, -2139062144;
and.b32 %r260, %r141, 2139062143;
and.b32 %r261, %r237, -2139062144;
sub.s32 %r262, %r259, %r260;
xor.b32 %r263, %r261, %r262;
xor.b32 %r148, %r263, -2139062144;
ld.global.nc.u32 %r26, [%rd88+32];

dp4a.s32.s32 %r143, %r144, %r25, %r135;


dp4a.s32.s32 %r147, %r148, %r26, %r143;

cvt.u32.u16 %r264, %rs7;
cvt.u32.u16 %r265, %rs8;
mul.lo.s32 %r266, %r123, %r264;
mad.lo.s32 %r267, %r147, %r265, %r266;
add.s32 %r268, %r147, %r123;
shr.u32 %r269, %r268, 31;
add.s32 %r270, %r268, %r269;
shr.s32 %r271, %r270, 1;
add.s32 %r272, %r267, %r271;
shr.s32 %r273, %r272, 31;
shr.u32 %r274, %r273, 30;
add.s32 %r275, %r272, %r274;
shr.s32 %r276, %r275, 2;
add.s64 %rd26, %rd216, %rd8;
ld.global.nc.u16 %rs1, [%rd26];

{ cvt.f32.f16 %f101, %rs1;}


ld.global.nc.u32 %r27, [%rd88];

{.reg .f16 low,high;
mov.b32 {low,high},%r27;
cvt.f32.f16 %f102, low;}


mul.ftz.f32 %f103, %f101, %f102;
cvt.rn.f32.s32 %f104, %r276;
fma.rn.ftz.f32 %f30, %f103, %f104, %f30;
setp.eq.s64 %p13, %rd40, 0;
@%p13 bra $L__BB213_12;

mov.u64 %rd215, iq2xs_grid;
add.s64 %rd98, %rd217, %rd11;
ld.global.u16 %rs19, [%rd98+-4];
ld.global.u16 %rs20, [%rd98+-2];
ld.global.u16 %rs21, [%rd98];
add.s64 %rd99, %rd217, %rd9;
ld.global.u8 %rs22, [%rd99];
and.b16 %rs23, %rs22, 15;
shr.u16 %rs24, %rs22, 4;
ld.global.u16 %rs25, [%rd98+-6];
shl.b16 %rs26, %rs25, 3;
cvt.u64.u16 %rd100, %rs26;
and.b64 %rd101, %rd100, 4088;
add.s64 %rd103, %rd215, %rd101;
ld.global.nc.v2.u32 {%r326, %r327}, [%rd103];
shr.u16 %rs27, %rs25, 9;
cvt.u32.u16 %r330, %rs27;
popc.b32 %r331, %r330;
shl.b32 %r332, %r331, 7;
and.b32 %r333, %r332, 128;
or.b32 %r334, %r333, %r330;
mul.lo.s32 %r335, %r334, 16843009;
and.b32 %r336, %r335, 134480385;
or.b32 %r337, %r336, -2139062144;
add.s32 %r338, %r337, -16843009;
or.b32 %r277, %r338, %r336;

prmt.b32 %r277, %r277, 0, 0xba98;

xor.b32 %r339, %r277, %r326;
or.b32 %r340, %r339, -2139062144;
and.b32 %r341, %r277, 2139062143;
and.b32 %r342, %r326, -2139062144;
sub.s32 %r343, %r340, %r341;
xor.b32 %r344, %r342, %r343;
xor.b32 %r282, %r344, -2139062144;
and.b32 %r345, %r335, -2143281136;
or.b32 %r346, %r345, -2139062144;
add.s32 %r347, %r346, -16843009;
or.b32 %r279, %r347, %r345;

prmt.b32 %r279, %r279, 0, 0xba98;

xor.b32 %r348, %r279, %r327;
or.b32 %r349, %r348, -2139062144;
and.b32 %r350, %r279, 2139062143;
and.b32 %r351, %r327, -2139062144;
sub.s32 %r352, %r349, %r350;
xor.b32 %r353, %r351, %r352;
xor.b32 %r286, %r353, -2139062144;

dp4a.s32.s32 %r281, %r282, %r19, %r134;


dp4a.s32.s32 %r285, %r286, %r20, %r281;

shl.b16 %rs28, %rs19, 3;
cvt.u64.u16 %rd104, %rs28;
and.b64 %rd105, %rd104, 4088;
add.s64 %rd106, %rd215, %rd105;
ld.global.nc.v2.u32 {%r354, %r355}, [%rd106];
shr.u16 %rs29, %rs19, 9;
cvt.u32.u16 %r358, %rs29;
popc.b32 %r359, %r358;
shl.b32 %r360, %r359, 7;
and.b32 %r361, %r360, 128;
or.b32 %r362, %r361, %r358;
mul.lo.s32 %r363, %r362, 16843009;
and.b32 %r364, %r363, 134480385;
or.b32 %r365, %r364, -2139062144;
add.s32 %r366, %r365, -16843009;
or.b32 %r289, %r366, %r364;

prmt.b32 %r289, %r289, 0, 0xba98;

xor.b32 %r367, %r289, %r354;
or.b32 %r368, %r367, -2139062144;
and.b32 %r369, %r289, 2139062143;
and.b32 %r370, %r354, -2139062144;
sub.s32 %r371, %r368, %r369;
xor.b32 %r372, %r370, %r371;
xor.b32 %r294, %r372, -2139062144;
and.b32 %r373, %r363, -2143281136;
or.b32 %r374, %r373, -2139062144;
add.s32 %r375, %r374, -16843009;
or.b32 %r291, %r375, %r373;

prmt.b32 %r291, %r291, 0, 0xba98;

xor.b32 %r376, %r291, %r355;
or.b32 %r377, %r376, -2139062144;
and.b32 %r378, %r291, 2139062143;
and.b32 %r379, %r355, -2139062144;
sub.s32 %r380, %r377, %r378;
xor.b32 %r381, %r379, %r380;
xor.b32 %r298, %r381, -2139062144;

dp4a.s32.s32 %r293, %r294, %r21, %r285;


dp4a.s32.s32 %r297, %r298, %r22, %r293;

shl.b16 %rs30, %rs20, 3;
cvt.u64.u16 %rd107, %rs30;
and.b64 %rd108, %rd107, 4088;
add.s64 %rd109, %rd215, %rd108;
ld.global.nc.v2.u32 {%r382, %r383}, [%rd109];
shr.u16 %rs31, %rs20, 9;
cvt.u32.u16 %r386, %rs31;
popc.b32 %r387, %r386;
shl.b32 %r388, %r387, 7;
and.b32 %r389, %r388, 128;
or.b32 %r390, %r389, %r386;
mul.lo.s32 %r391, %r390, 16843009;
and.b32 %r392, %r391, 134480385;
or.b32 %r393, %r392, -2139062144;
add.s32 %r394, %r393, -16843009;
or.b32 %r301, %r394, %r392;

prmt.b32 %r301, %r301, 0, 0xba98;

xor.b32 %r395, %r301, %r382;
or.b32 %r396, %r395, -2139062144;
and.b32 %r397, %r301, 2139062143;
and.b32 %r398, %r382, -2139062144;
sub.s32 %r399, %r396, %r397;
xor.b32 %r400, %r398, %r399;
xor.b32 %r306, %r400, -2139062144;
and.b32 %r401, %r391, -2143281136;
or.b32 %r402, %r401, -2139062144;
add.s32 %r403, %r402, -16843009;
or.b32 %r303, %r403, %r401;

prmt.b32 %r303, %r303, 0, 0xba98;

xor.b32 %r404, %r303, %r383;
or.b32 %r405, %r404, -2139062144;
and.b32 %r406, %r303, 2139062143;
and.b32 %r407, %r383, -2139062144;
sub.s32 %r408, %r405, %r406;
xor.b32 %r409, %r407, %r408;
xor.b32 %r310, %r409, -2139062144;

dp4a.s32.s32 %r305, %r306, %r23, %r134;


dp4a.s32.s32 %r309, %r310, %r24, %r305;

shl.b16 %rs32, %rs21, 3;
cvt.u64.u16 %rd110, %rs32;
and.b64 %rd111, %rd110, 4088;
add.s64 %rd112, %rd215, %rd111;
ld.global.nc.v2.u32 {%r410, %r411}, [%rd112];
shr.u16 %rs33, %rs21, 9;
cvt.u32.u16 %r414, %rs33;
popc.b32 %r415, %r414;
shl.b32 %r416, %r415, 7;
and.b32 %r417, %r416, 128;
or.b32 %r418, %r417, %r414;
mul.lo.s32 %r419, %r418, 16843009;
and.b32 %r420, %r419, 134480385;
or.b32 %r421, %r420, -2139062144;
add.s32 %r422, %r421, -16843009;
or.b32 %r313, %r422, %r420;

prmt.b32 %r313, %r313, 0, 0xba98;

xor.b32 %r423, %r313, %r410;
or.b32 %r424, %r423, -2139062144;
and.b32 %r425, %r313, 2139062143;
and.b32 %r426, %r410, -2139062144;
sub.s32 %r427, %r424, %r425;
xor.b32 %r428, %r426, %r427;
xor.b32 %r318, %r428, -2139062144;
and.b32 %r429, %r419, -2143281136;
or.b32 %r430, %r429, -2139062144;
add.s32 %r431, %r430, -16843009;
or.b32 %r315, %r431, %r429;

prmt.b32 %r315, %r315, 0, 0xba98;

xor.b32 %r432, %r315, %r411;
or.b32 %r433, %r432, -2139062144;
and.b32 %r434, %r315, 2139062143;
and.b32 %r435, %r411, -2139062144;
sub.s32 %r436, %r433, %r434;
xor.b32 %r437, %r435, %r436;
xor.b32 %r322, %r437, -2139062144;

dp4a.s32.s32 %r317, %r318, %r25, %r309;


dp4a.s32.s32 %r321, %r322, %r26, %r317;

cvt.u32.u16 %r438, %rs23;
cvt.u32.u16 %r439, %rs24;
mul.lo.s32 %r440, %r297, %r438;
mad.lo.s32 %r441, %r321, %r439, %r440;
add.s32 %r442, %r321, %r297;
shr.u32 %r443, %r442, 31;
add.s32 %r444, %r442, %r443;
shr.s32 %r445, %r444, 1;
add.s32 %r446, %r441, %r445;
shr.s32 %r447, %r446, 31;
shr.u32 %r448, %r447, 30;
add.s32 %r449, %r446, %r448;
shr.s32 %r450, %r449, 2;
add.s64 %rd113, %rd217, %rd8;
ld.global.u16 %rs18, [%rd113];

{ cvt.f32.f16 %f105, %rs18;}



{.reg .f16 low,high;
mov.b32 {low,high},%r27;
cvt.f32.f16 %f106, low;}


mul.ftz.f32 %f107, %f105, %f106;
cvt.rn.f32.s32 %f108, %r450;
fma.rn.ftz.f32 %f34, %f107, %f108, %f34;

$L__BB213_12:
mov.u64 %rd212, iq2xs_grid;
add.s64 %rd207, %rd216, %rd11;
add.s64 %rd206, %rd207, -6;
add.s64 %rd27, %rd206, %rd18;
ld.global.nc.u16 %rs35, [%rd27+2];
ld.global.nc.u16 %rs36, [%rd27+4];
ld.global.nc.u16 %rs37, [%rd27+6];
add.s64 %rd28, %rd25, %rd18;
ld.global.nc.u8 %rs38, [%rd28];
and.b16 %rs39, %rs38, 240;
and.b16 %rs40, %rs38, 15;
shr.u16 %rs41, %rs39, 4;
ld.global.nc.u16 %rs42, [%rd27];
shl.b16 %rs43, %rs42, 3;
cvt.u64.u16 %rd114, %rs43;
and.b64 %rd115, %rd114, 4088;
add.s64 %rd117, %rd212, %rd115;
ld.global.nc.v2.u32 {%r500, %r501}, [%rd117];
shr.u16 %rs44, %rs42, 9;
cvt.u32.u16 %r504, %rs44;
popc.b32 %r505, %r504;
shl.b32 %r506, %r505, 7;
and.b32 %r507, %r506, 128;
or.b32 %r508, %r507, %r504;
mul.lo.s32 %r509, %r508, 16843009;
and.b32 %r510, %r509, 134480385;
or.b32 %r511, %r510, -2139062144;
add.s32 %r512, %r511, -16843009;
or.b32 %r451, %r512, %r510;

prmt.b32 %r451, %r451, 0, 0xba98;

xor.b32 %r513, %r451, %r500;
or.b32 %r514, %r513, -2139062144;
and.b32 %r515, %r451, 2139062143;
and.b32 %r516, %r500, -2139062144;
sub.s32 %r517, %r514, %r515;
xor.b32 %r518, %r516, %r517;
xor.b32 %r456, %r518, -2139062144;
and.b32 %r519, %r509, -2143281136;
or.b32 %r520, %r519, -2139062144;
add.s32 %r521, %r520, -16843009;
or.b32 %r453, %r521, %r519;

prmt.b32 %r453, %r453, 0, 0xba98;

xor.b32 %r522, %r453, %r501;
or.b32 %r523, %r522, -2139062144;
and.b32 %r524, %r453, 2139062143;
and.b32 %r525, %r501, -2139062144;
sub.s32 %r526, %r523, %r524;
xor.b32 %r527, %r525, %r526;
xor.b32 %r460, %r527, -2139062144;

dp4a.s32.s32 %r455, %r456, %r19, %r134;


dp4a.s32.s32 %r459, %r460, %r20, %r455;

shl.b16 %rs45, %rs35, 3;
cvt.u64.u16 %rd118, %rs45;
and.b64 %rd119, %rd118, 4088;
add.s64 %rd120, %rd212, %rd119;
ld.global.nc.v2.u32 {%r528, %r529}, [%rd120];
shr.u16 %rs46, %rs35, 9;
cvt.u32.u16 %r532, %rs46;
popc.b32 %r533, %r532;
shl.b32 %r534, %r533, 7;
and.b32 %r535, %r534, 128;
or.b32 %r536, %r535, %r532;
mul.lo.s32 %r537, %r536, 16843009;
and.b32 %r538, %r537, 134480385;
or.b32 %r539, %r538, -2139062144;
add.s32 %r540, %r539, -16843009;
or.b32 %r463, %r540, %r538;

prmt.b32 %r463, %r463, 0, 0xba98;

xor.b32 %r541, %r463, %r528;
or.b32 %r542, %r541, -2139062144;
and.b32 %r543, %r463, 2139062143;
and.b32 %r544, %r528, -2139062144;
sub.s32 %r545, %r542, %r543;
xor.b32 %r546, %r544, %r545;
xor.b32 %r468, %r546, -2139062144;
and.b32 %r547, %r537, -2143281136;
or.b32 %r548, %r547, -2139062144;
add.s32 %r549, %r548, -16843009;
or.b32 %r465, %r549, %r547;

prmt.b32 %r465, %r465, 0, 0xba98;

xor.b32 %r550, %r465, %r529;
or.b32 %r551, %r550, -2139062144;
and.b32 %r552, %r465, 2139062143;
and.b32 %r553, %r529, -2139062144;
sub.s32 %r554, %r551, %r552;
xor.b32 %r555, %r553, %r554;
xor.b32 %r472, %r555, -2139062144;

dp4a.s32.s32 %r467, %r468, %r21, %r459;


dp4a.s32.s32 %r471, %r472, %r22, %r467;

shl.b16 %rs47, %rs36, 3;
cvt.u64.u16 %rd121, %rs47;
and.b64 %rd122, %rd121, 4088;
add.s64 %rd123, %rd212, %rd122;
ld.global.nc.v2.u32 {%r556, %r557}, [%rd123];
shr.u16 %rs48, %rs36, 9;
cvt.u32.u16 %r560, %rs48;
popc.b32 %r561, %r560;
shl.b32 %r562, %r561, 7;
and.b32 %r563, %r562, 128;
or.b32 %r564, %r563, %r560;
mul.lo.s32 %r565, %r564, 16843009;
and.b32 %r566, %r565, 134480385;
or.b32 %r567, %r566, -2139062144;
add.s32 %r568, %r567, -16843009;
or.b32 %r475, %r568, %r566;

prmt.b32 %r475, %r475, 0, 0xba98;

xor.b32 %r569, %r475, %r556;
or.b32 %r570, %r569, -2139062144;
and.b32 %r571, %r475, 2139062143;
and.b32 %r572, %r556, -2139062144;
sub.s32 %r573, %r570, %r571;
xor.b32 %r574, %r572, %r573;
xor.b32 %r480, %r574, -2139062144;
and.b32 %r575, %r565, -2143281136;
or.b32 %r576, %r575, -2139062144;
add.s32 %r577, %r576, -16843009;
or.b32 %r477, %r577, %r575;

prmt.b32 %r477, %r477, 0, 0xba98;

xor.b32 %r578, %r477, %r557;
or.b32 %r579, %r578, -2139062144;
and.b32 %r580, %r477, 2139062143;
and.b32 %r581, %r557, -2139062144;
sub.s32 %r582, %r579, %r580;
xor.b32 %r583, %r581, %r582;
xor.b32 %r484, %r583, -2139062144;

dp4a.s32.s32 %r479, %r480, %r23, %r134;


dp4a.s32.s32 %r483, %r484, %r24, %r479;

shl.b16 %rs49, %rs37, 3;
cvt.u64.u16 %rd124, %rs49;
and.b64 %rd125, %rd124, 4088;
add.s64 %rd126, %rd212, %rd125;
ld.global.nc.v2.u32 {%r584, %r585}, [%rd126];
shr.u16 %rs50, %rs37, 9;
cvt.u32.u16 %r588, %rs50;
popc.b32 %r589, %r588;
shl.b32 %r590, %r589, 7;
and.b32 %r591, %r590, 128;
or.b32 %r592, %r591, %r588;
mul.lo.s32 %r593, %r592, 16843009;
and.b32 %r594, %r593, 134480385;
or.b32 %r595, %r594, -2139062144;
add.s32 %r596, %r595, -16843009;
or.b32 %r487, %r596, %r594;

prmt.b32 %r487, %r487, 0, 0xba98;

xor.b32 %r597, %r487, %r584;
or.b32 %r598, %r597, -2139062144;
and.b32 %r599, %r487, 2139062143;
and.b32 %r600, %r584, -2139062144;
sub.s32 %r601, %r598, %r599;
xor.b32 %r602, %r600, %r601;
xor.b32 %r492, %r602, -2139062144;
and.b32 %r603, %r593, -2143281136;
or.b32 %r604, %r603, -2139062144;
add.s32 %r605, %r604, -16843009;
or.b32 %r489, %r605, %r603;

prmt.b32 %r489, %r489, 0, 0xba98;

xor.b32 %r606, %r489, %r585;
or.b32 %r607, %r606, -2139062144;
and.b32 %r608, %r489, 2139062143;
and.b32 %r609, %r585, -2139062144;
sub.s32 %r610, %r607, %r608;
xor.b32 %r611, %r609, %r610;
xor.b32 %r496, %r611, -2139062144;

dp4a.s32.s32 %r491, %r492, %r25, %r483;


dp4a.s32.s32 %r495, %r496, %r26, %r491;

cvt.u32.u16 %r612, %rs40;
cvt.u32.u16 %r613, %rs41;
mul.lo.s32 %r614, %r471, %r612;
mad.lo.s32 %r615, %r495, %r613, %r614;
add.s32 %r616, %r495, %r471;
shr.u32 %r617, %r616, 31;
add.s32 %r618, %r616, %r617;
shr.s32 %r619, %r618, 1;
add.s32 %r620, %r615, %r619;
shr.s32 %r621, %r620, 31;
shr.u32 %r622, %r621, 30;
add.s32 %r623, %r620, %r622;
shr.s32 %r624, %r623, 2;
add.s64 %rd29, %rd26, %rd18;
ld.global.nc.u16 %rs34, [%rd29];

{ cvt.f32.f16 %f109, %rs34;}



{.reg .f16 low,high;
mov.b32 {low,high},%r27;
cvt.f32.f16 %f110, low;}


mul.ftz.f32 %f111, %f109, %f110;
cvt.rn.f32.s32 %f112, %r624;
fma.rn.ftz.f32 %f29, %f111, %f112, %f29;
@%p13 bra $L__BB213_14;

mov.u64 %rd214, iq2xs_grid;
add.s64 %rd127, %rd217, %rd10;
ld.global.u16 %rs52, [%rd127];
ld.global.u16 %rs53, [%rd127+2];
ld.global.u16 %rs54, [%rd127+4];
add.s64 %rd128, %rd217, %rd7;
ld.global.u8 %rs55, [%rd128];
and.b16 %rs56, %rs55, 15;
shr.u16 %rs57, %rs55, 4;
ld.global.u16 %rs58, [%rd127+-2];
shl.b16 %rs59, %rs58, 3;
cvt.u64.u16 %rd129, %rs59;
and.b64 %rd130, %rd129, 4088;
add.s64 %rd132, %rd214, %rd130;
ld.global.nc.v2.u32 {%r674, %r675}, [%rd132];
mov.u32 %r656, 0;
shr.u16 %rs60, %rs58, 9;
cvt.u32.u16 %r678, %rs60;
popc.b32 %r679, %r678;
shl.b32 %r680, %r679, 7;
and.b32 %r681, %r680, 128;
or.b32 %r682, %r681, %r678;
mul.lo.s32 %r683, %r682, 16843009;
and.b32 %r684, %r683, 134480385;
or.b32 %r685, %r684, -2139062144;
add.s32 %r686, %r685, -16843009;
or.b32 %r625, %r686, %r684;

prmt.b32 %r625, %r625, 0, 0xba98;

xor.b32 %r687, %r625, %r674;
or.b32 %r688, %r687, -2139062144;
and.b32 %r689, %r625, 2139062143;
and.b32 %r690, %r674, -2139062144;
sub.s32 %r691, %r688, %r689;
xor.b32 %r692, %r690, %r691;
xor.b32 %r630, %r692, -2139062144;
and.b32 %r693, %r683, -2143281136;
or.b32 %r694, %r693, -2139062144;
add.s32 %r695, %r694, -16843009;
or.b32 %r627, %r695, %r693;

prmt.b32 %r627, %r627, 0, 0xba98;

xor.b32 %r696, %r627, %r675;
or.b32 %r697, %r696, -2139062144;
and.b32 %r698, %r627, 2139062143;
and.b32 %r699, %r675, -2139062144;
sub.s32 %r700, %r697, %r698;
xor.b32 %r701, %r699, %r700;
xor.b32 %r634, %r701, -2139062144;

dp4a.s32.s32 %r629, %r630, %r19, %r656;


dp4a.s32.s32 %r633, %r634, %r20, %r629;

shl.b16 %rs61, %rs52, 3;
cvt.u64.u16 %rd133, %rs61;
and.b64 %rd134, %rd133, 4088;
add.s64 %rd135, %rd214, %rd134;
ld.global.nc.v2.u32 {%r702, %r703}, [%rd135];
shr.u16 %rs62, %rs52, 9;
cvt.u32.u16 %r706, %rs62;
popc.b32 %r707, %r706;
shl.b32 %r708, %r707, 7;
and.b32 %r709, %r708, 128;
or.b32 %r710, %r709, %r706;
mul.lo.s32 %r711, %r710, 16843009;
and.b32 %r712, %r711, 134480385;
or.b32 %r713, %r712, -2139062144;
add.s32 %r714, %r713, -16843009;
or.b32 %r637, %r714, %r712;

prmt.b32 %r637, %r637, 0, 0xba98;

xor.b32 %r715, %r637, %r702;
or.b32 %r716, %r715, -2139062144;
and.b32 %r717, %r637, 2139062143;
and.b32 %r718, %r702, -2139062144;
sub.s32 %r719, %r716, %r717;
xor.b32 %r720, %r718, %r719;
xor.b32 %r642, %r720, -2139062144;
and.b32 %r721, %r711, -2143281136;
or.b32 %r722, %r721, -2139062144;
add.s32 %r723, %r722, -16843009;
or.b32 %r639, %r723, %r721;

prmt.b32 %r639, %r639, 0, 0xba98;

xor.b32 %r724, %r639, %r703;
or.b32 %r725, %r724, -2139062144;
and.b32 %r726, %r639, 2139062143;
and.b32 %r727, %r703, -2139062144;
sub.s32 %r728, %r725, %r726;
xor.b32 %r729, %r727, %r728;
xor.b32 %r646, %r729, -2139062144;

dp4a.s32.s32 %r641, %r642, %r21, %r633;


dp4a.s32.s32 %r645, %r646, %r22, %r641;

shl.b16 %rs63, %rs53, 3;
cvt.u64.u16 %rd136, %rs63;
and.b64 %rd137, %rd136, 4088;
add.s64 %rd138, %rd214, %rd137;
ld.global.nc.v2.u32 {%r730, %r731}, [%rd138];
shr.u16 %rs64, %rs53, 9;
cvt.u32.u16 %r734, %rs64;
popc.b32 %r735, %r734;
shl.b32 %r736, %r735, 7;
and.b32 %r737, %r736, 128;
or.b32 %r738, %r737, %r734;
mul.lo.s32 %r739, %r738, 16843009;
and.b32 %r740, %r739, 134480385;
or.b32 %r741, %r740, -2139062144;
add.s32 %r742, %r741, -16843009;
or.b32 %r649, %r742, %r740;

prmt.b32 %r649, %r649, 0, 0xba98;

xor.b32 %r743, %r649, %r730;
or.b32 %r744, %r743, -2139062144;
and.b32 %r745, %r649, 2139062143;
and.b32 %r746, %r730, -2139062144;
sub.s32 %r747, %r744, %r745;
xor.b32 %r748, %r746, %r747;
xor.b32 %r654, %r748, -2139062144;
and.b32 %r749, %r739, -2143281136;
or.b32 %r750, %r749, -2139062144;
add.s32 %r751, %r750, -16843009;
or.b32 %r651, %r751, %r749;

prmt.b32 %r651, %r651, 0, 0xba98;

xor.b32 %r752, %r651, %r731;
or.b32 %r753, %r752, -2139062144;
and.b32 %r754, %r651, 2139062143;
and.b32 %r755, %r731, -2139062144;
sub.s32 %r756, %r753, %r754;
xor.b32 %r757, %r755, %r756;
xor.b32 %r658, %r757, -2139062144;

dp4a.s32.s32 %r653, %r654, %r23, %r656;


dp4a.s32.s32 %r657, %r658, %r24, %r653;

shl.b16 %rs65, %rs54, 3;
cvt.u64.u16 %rd139, %rs65;
and.b64 %rd140, %rd139, 4088;
add.s64 %rd141, %rd214, %rd140;
ld.global.nc.v2.u32 {%r758, %r759}, [%rd141];
shr.u16 %rs66, %rs54, 9;
cvt.u32.u16 %r762, %rs66;
popc.b32 %r763, %r762;
shl.b32 %r764, %r763, 7;
and.b32 %r765, %r764, 128;
or.b32 %r766, %r765, %r762;
mul.lo.s32 %r767, %r766, 16843009;
and.b32 %r768, %r767, 134480385;
or.b32 %r769, %r768, -2139062144;
add.s32 %r770, %r769, -16843009;
or.b32 %r661, %r770, %r768;

prmt.b32 %r661, %r661, 0, 0xba98;

xor.b32 %r771, %r661, %r758;
or.b32 %r772, %r771, -2139062144;
and.b32 %r773, %r661, 2139062143;
and.b32 %r774, %r758, -2139062144;
sub.s32 %r775, %r772, %r773;
xor.b32 %r776, %r774, %r775;
xor.b32 %r666, %r776, -2139062144;
and.b32 %r777, %r767, -2143281136;
or.b32 %r778, %r777, -2139062144;
add.s32 %r779, %r778, -16843009;
or.b32 %r663, %r779, %r777;

prmt.b32 %r663, %r663, 0, 0xba98;

xor.b32 %r780, %r663, %r759;
or.b32 %r781, %r780, -2139062144;
and.b32 %r782, %r663, 2139062143;
and.b32 %r783, %r759, -2139062144;
sub.s32 %r784, %r781, %r782;
xor.b32 %r785, %r783, %r784;
xor.b32 %r670, %r785, -2139062144;

dp4a.s32.s32 %r665, %r666, %r25, %r657;


dp4a.s32.s32 %r669, %r670, %r26, %r665;

cvt.u32.u16 %r786, %rs56;
cvt.u32.u16 %r787, %rs57;
mul.lo.s32 %r788, %r645, %r786;
mad.lo.s32 %r789, %r669, %r787, %r788;
add.s32 %r790, %r669, %r645;
shr.u32 %r791, %r790, 31;
add.s32 %r792, %r790, %r791;
shr.s32 %r793, %r792, 1;
add.s32 %r794, %r789, %r793;
shr.s32 %r795, %r794, 31;
shr.u32 %r796, %r795, 30;
add.s32 %r797, %r794, %r796;
shr.s32 %r798, %r797, 2;
add.s64 %rd142, %rd217, %rd6;
ld.global.u16 %rs51, [%rd142];

{ cvt.f32.f16 %f113, %rs51;}



{.reg .f16 low,high;
mov.b32 {low,high},%r27;
cvt.f32.f16 %f114, low;}


mul.ftz.f32 %f115, %f113, %f114;
cvt.rn.f32.s32 %f116, %r798;
fma.rn.ftz.f32 %f33, %f115, %f116, %f33;

$L__BB213_14:
mov.u64 %rd213, iq2xs_grid;
add.s64 %rd30, %rd27, %rd18;
ld.global.nc.u16 %rs68, [%rd30+2];
ld.global.nc.u16 %rs69, [%rd30+4];
ld.global.nc.u16 %rs70, [%rd30+6];
add.s64 %rd31, %rd28, %rd18;
ld.global.nc.u8 %rs71, [%rd31];
and.b16 %rs72, %rs71, 240;
and.b16 %rs73, %rs71, 15;
shr.u16 %rs74, %rs72, 4;
ld.global.nc.u16 %rs75, [%rd30];
shl.b16 %rs76, %rs75, 3;
cvt.u64.u16 %rd143, %rs76;
and.b64 %rd144, %rd143, 4088;
add.s64 %rd146, %rd213, %rd144;
ld.global.nc.v2.u32 {%r848, %r849}, [%rd146];
mov.u32 %r830, 0;
shr.u16 %rs77, %rs75, 9;
cvt.u32.u16 %r852, %rs77;
popc.b32 %r853, %r852;
shl.b32 %r854, %r853, 7;
and.b32 %r855, %r854, 128;
or.b32 %r856, %r855, %r852;
mul.lo.s32 %r857, %r856, 16843009;
and.b32 %r858, %r857, 134480385;
or.b32 %r859, %r858, -2139062144;
add.s32 %r860, %r859, -16843009;
or.b32 %r799, %r860, %r858;

prmt.b32 %r799, %r799, 0, 0xba98;

xor.b32 %r861, %r799, %r848;
or.b32 %r862, %r861, -2139062144;
and.b32 %r863, %r799, 2139062143;
and.b32 %r864, %r848, -2139062144;
sub.s32 %r865, %r862, %r863;
xor.b32 %r866, %r864, %r865;
xor.b32 %r804, %r866, -2139062144;
and.b32 %r867, %r857, -2143281136;
or.b32 %r868, %r867, -2139062144;
add.s32 %r869, %r868, -16843009;
or.b32 %r801, %r869, %r867;

prmt.b32 %r801, %r801, 0, 0xba98;

xor.b32 %r870, %r801, %r849;
or.b32 %r871, %r870, -2139062144;
and.b32 %r872, %r801, 2139062143;
and.b32 %r873, %r849, -2139062144;
sub.s32 %r874, %r871, %r872;
xor.b32 %r875, %r873, %r874;
xor.b32 %r808, %r875, -2139062144;

dp4a.s32.s32 %r803, %r804, %r19, %r830;


dp4a.s32.s32 %r807, %r808, %r20, %r803;

shl.b16 %rs78, %rs68, 3;
cvt.u64.u16 %rd147, %rs78;
and.b64 %rd148, %rd147, 4088;
add.s64 %rd149, %rd213, %rd148;
ld.global.nc.v2.u32 {%r876, %r877}, [%rd149];
shr.u16 %rs79, %rs68, 9;
cvt.u32.u16 %r880, %rs79;
popc.b32 %r881, %r880;
shl.b32 %r882, %r881, 7;
and.b32 %r883, %r882, 128;
or.b32 %r884, %r883, %r880;
mul.lo.s32 %r885, %r884, 16843009;
and.b32 %r886, %r885, 134480385;
or.b32 %r887, %r886, -2139062144;
add.s32 %r888, %r887, -16843009;
or.b32 %r811, %r888, %r886;

prmt.b32 %r811, %r811, 0, 0xba98;

xor.b32 %r889, %r811, %r876;
or.b32 %r890, %r889, -2139062144;
and.b32 %r891, %r811, 2139062143;
and.b32 %r892, %r876, -2139062144;
sub.s32 %r893, %r890, %r891;
xor.b32 %r894, %r892, %r893;
xor.b32 %r816, %r894, -2139062144;
and.b32 %r895, %r885, -2143281136;
or.b32 %r896, %r895, -2139062144;
add.s32 %r897, %r896, -16843009;
or.b32 %r813, %r897, %r895;

prmt.b32 %r813, %r813, 0, 0xba98;

xor.b32 %r898, %r813, %r877;
or.b32 %r899, %r898, -2139062144;
and.b32 %r900, %r813, 2139062143;
and.b32 %r901, %r877, -2139062144;
sub.s32 %r902, %r899, %r900;
xor.b32 %r903, %r901, %r902;
xor.b32 %r820, %r903, -2139062144;

dp4a.s32.s32 %r815, %r816, %r21, %r807;


dp4a.s32.s32 %r819, %r820, %r22, %r815;

shl.b16 %rs80, %rs69, 3;
cvt.u64.u16 %rd150, %rs80;
and.b64 %rd151, %rd150, 4088;
add.s64 %rd152, %rd213, %rd151;
ld.global.nc.v2.u32 {%r904, %r905}, [%rd152];
shr.u16 %rs81, %rs69, 9;
cvt.u32.u16 %r908, %rs81;
popc.b32 %r909, %r908;
shl.b32 %r910, %r909, 7;
and.b32 %r911, %r910, 128;
or.b32 %r912, %r911, %r908;
mul.lo.s32 %r913, %r912, 16843009;
and.b32 %r914, %r913, 134480385;
or.b32 %r915, %r914, -2139062144;
add.s32 %r916, %r915, -16843009;
or.b32 %r823, %r916, %r914;

prmt.b32 %r823, %r823, 0, 0xba98;

xor.b32 %r917, %r823, %r904;
or.b32 %r918, %r917, -2139062144;
and.b32 %r919, %r823, 2139062143;
and.b32 %r920, %r904, -2139062144;
sub.s32 %r921, %r918, %r919;
xor.b32 %r922, %r920, %r921;
xor.b32 %r828, %r922, -2139062144;
and.b32 %r923, %r913, -2143281136;
or.b32 %r924, %r923, -2139062144;
add.s32 %r925, %r924, -16843009;
or.b32 %r825, %r925, %r923;

prmt.b32 %r825, %r825, 0, 0xba98;

xor.b32 %r926, %r825, %r905;
or.b32 %r927, %r926, -2139062144;
and.b32 %r928, %r825, 2139062143;
and.b32 %r929, %r905, -2139062144;
sub.s32 %r930, %r927, %r928;
xor.b32 %r931, %r929, %r930;
xor.b32 %r832, %r931, -2139062144;

dp4a.s32.s32 %r827, %r828, %r23, %r830;


dp4a.s32.s32 %r831, %r832, %r24, %r827;

shl.b16 %rs82, %rs70, 3;
cvt.u64.u16 %rd153, %rs82;
and.b64 %rd154, %rd153, 4088;
add.s64 %rd155, %rd213, %rd154;
ld.global.nc.v2.u32 {%r932, %r933}, [%rd155];
shr.u16 %rs83, %rs70, 9;
cvt.u32.u16 %r936, %rs83;
popc.b32 %r937, %r936;
shl.b32 %r938, %r937, 7;
and.b32 %r939, %r938, 128;
or.b32 %r940, %r939, %r936;
mul.lo.s32 %r941, %r940, 16843009;
and.b32 %r942, %r941, 134480385;
or.b32 %r943, %r942, -2139062144;
add.s32 %r944, %r943, -16843009;
or.b32 %r835, %r944, %r942;

prmt.b32 %r835, %r835, 0, 0xba98;

xor.b32 %r945, %r835, %r932;
or.b32 %r946, %r945, -2139062144;
and.b32 %r947, %r835, 2139062143;
and.b32 %r948, %r932, -2139062144;
sub.s32 %r949, %r946, %r947;
xor.b32 %r950, %r948, %r949;
xor.b32 %r840, %r950, -2139062144;
and.b32 %r951, %r941, -2143281136;
or.b32 %r952, %r951, -2139062144;
add.s32 %r953, %r952, -16843009;
or.b32 %r837, %r953, %r951;

prmt.b32 %r837, %r837, 0, 0xba98;

xor.b32 %r954, %r837, %r933;
or.b32 %r955, %r954, -2139062144;
and.b32 %r956, %r837, 2139062143;
and.b32 %r957, %r933, -2139062144;
sub.s32 %r958, %r955, %r956;
xor.b32 %r959, %r957, %r958;
xor.b32 %r844, %r959, -2139062144;

dp4a.s32.s32 %r839, %r840, %r25, %r831;


dp4a.s32.s32 %r843, %r844, %r26, %r839;

cvt.u32.u16 %r960, %rs73;
cvt.u32.u16 %r961, %rs74;
mul.lo.s32 %r962, %r819, %r960;
mad.lo.s32 %r963, %r843, %r961, %r962;
add.s32 %r964, %r843, %r819;
shr.u32 %r965, %r964, 31;
add.s32 %r966, %r964, %r965;
shr.s32 %r967, %r966, 1;
add.s32 %r968, %r963, %r967;
shr.s32 %r969, %r968, 31;
shr.u32 %r970, %r969, 30;
add.s32 %r971, %r968, %r970;
shr.s32 %r972, %r971, 2;
add.s64 %rd32, %rd29, %rd18;
ld.global.nc.u16 %rs67, [%rd32];

{ cvt.f32.f16 %f117, %rs67;}



{.reg .f16 low,high;
mov.b32 {low,high},%r27;
cvt.f32.f16 %f118, low;}


mul.ftz.f32 %f119, %f117, %f118;
cvt.rn.f32.s32 %f120, %r972;
fma.rn.ftz.f32 %f28, %f119, %f120, %f28;
@%p13 bra $L__BB213_16;

mov.u64 %rd211, iq2xs_grid;
add.s64 %rd156, %rd217, %rd17;
ld.global.u16 %rs85, [%rd156];
ld.global.u16 %rs86, [%rd156+2];
ld.global.u16 %rs87, [%rd156+4];
add.s64 %rd157, %rd217, %rd13;
ld.global.u8 %rs88, [%rd157];
and.b16 %rs89, %rs88, 15;
shr.u16 %rs90, %rs88, 4;
ld.global.u16 %rs91, [%rd156+-2];
shl.b16 %rs92, %rs91, 3;
cvt.u64.u16 %rd158, %rs92;
and.b64 %rd159, %rd158, 4088;
add.s64 %rd161, %rd211, %rd159;
ld.global.nc.v2.u32 {%r1022, %r1023}, [%rd161];
shr.u16 %rs93, %rs91, 9;
cvt.u32.u16 %r1026, %rs93;
popc.b32 %r1027, %r1026;
shl.b32 %r1028, %r1027, 7;
and.b32 %r1029, %r1028, 128;
or.b32 %r1030, %r1029, %r1026;
mul.lo.s32 %r1031, %r1030, 16843009;
and.b32 %r1032, %r1031, 134480385;
or.b32 %r1033, %r1032, -2139062144;
add.s32 %r1034, %r1033, -16843009;
or.b32 %r973, %r1034, %r1032;

prmt.b32 %r973, %r973, 0, 0xba98;

xor.b32 %r1035, %r973, %r1022;
or.b32 %r1036, %r1035, -2139062144;
and.b32 %r1037, %r973, 2139062143;
and.b32 %r1038, %r1022, -2139062144;
sub.s32 %r1039, %r1036, %r1037;
xor.b32 %r1040, %r1038, %r1039;
xor.b32 %r978, %r1040, -2139062144;
and.b32 %r1041, %r1031, -2143281136;
or.b32 %r1042, %r1041, -2139062144;
add.s32 %r1043, %r1042, -16843009;
or.b32 %r975, %r1043, %r1041;

prmt.b32 %r975, %r975, 0, 0xba98;

xor.b32 %r1044, %r975, %r1023;
or.b32 %r1045, %r1044, -2139062144;
and.b32 %r1046, %r975, 2139062143;
and.b32 %r1047, %r1023, -2139062144;
sub.s32 %r1048, %r1045, %r1046;
xor.b32 %r1049, %r1047, %r1048;
xor.b32 %r982, %r1049, -2139062144;

dp4a.s32.s32 %r977, %r978, %r19, %r830;


dp4a.s32.s32 %r981, %r982, %r20, %r977;

shl.b16 %rs94, %rs85, 3;
cvt.u64.u16 %rd162, %rs94;
and.b64 %rd163, %rd162, 4088;
add.s64 %rd164, %rd211, %rd163;
ld.global.nc.v2.u32 {%r1050, %r1051}, [%rd164];
shr.u16 %rs95, %rs85, 9;
cvt.u32.u16 %r1054, %rs95;
popc.b32 %r1055, %r1054;
shl.b32 %r1056, %r1055, 7;
and.b32 %r1057, %r1056, 128;
or.b32 %r1058, %r1057, %r1054;
mul.lo.s32 %r1059, %r1058, 16843009;
and.b32 %r1060, %r1059, 134480385;
or.b32 %r1061, %r1060, -2139062144;
add.s32 %r1062, %r1061, -16843009;
or.b32 %r985, %r1062, %r1060;

prmt.b32 %r985, %r985, 0, 0xba98;

xor.b32 %r1063, %r985, %r1050;
or.b32 %r1064, %r1063, -2139062144;
and.b32 %r1065, %r985, 2139062143;
and.b32 %r1066, %r1050, -2139062144;
sub.s32 %r1067, %r1064, %r1065;
xor.b32 %r1068, %r1066, %r1067;
xor.b32 %r990, %r1068, -2139062144;
and.b32 %r1069, %r1059, -2143281136;
or.b32 %r1070, %r1069, -2139062144;
add.s32 %r1071, %r1070, -16843009;
or.b32 %r987, %r1071, %r1069;

prmt.b32 %r987, %r987, 0, 0xba98;

xor.b32 %r1072, %r987, %r1051;
or.b32 %r1073, %r1072, -2139062144;
and.b32 %r1074, %r987, 2139062143;
and.b32 %r1075, %r1051, -2139062144;
sub.s32 %r1076, %r1073, %r1074;
xor.b32 %r1077, %r1075, %r1076;
xor.b32 %r994, %r1077, -2139062144;

dp4a.s32.s32 %r989, %r990, %r21, %r981;


dp4a.s32.s32 %r993, %r994, %r22, %r989;

shl.b16 %rs96, %rs86, 3;
cvt.u64.u16 %rd165, %rs96;
and.b64 %rd166, %rd165, 4088;
add.s64 %rd167, %rd211, %rd166;
ld.global.nc.v2.u32 {%r1078, %r1079}, [%rd167];
shr.u16 %rs97, %rs86, 9;
cvt.u32.u16 %r1082, %rs97;
popc.b32 %r1083, %r1082;
shl.b32 %r1084, %r1083, 7;
and.b32 %r1085, %r1084, 128;
or.b32 %r1086, %r1085, %r1082;
mul.lo.s32 %r1087, %r1086, 16843009;
and.b32 %r1088, %r1087, 134480385;
or.b32 %r1089, %r1088, -2139062144;
add.s32 %r1090, %r1089, -16843009;
or.b32 %r997, %r1090, %r1088;

prmt.b32 %r997, %r997, 0, 0xba98;

xor.b32 %r1091, %r997, %r1078;
or.b32 %r1092, %r1091, -2139062144;
and.b32 %r1093, %r997, 2139062143;
and.b32 %r1094, %r1078, -2139062144;
sub.s32 %r1095, %r1092, %r1093;
xor.b32 %r1096, %r1094, %r1095;
xor.b32 %r1002, %r1096, -2139062144;
and.b32 %r1097, %r1087, -2143281136;
or.b32 %r1098, %r1097, -2139062144;
add.s32 %r1099, %r1098, -16843009;
or.b32 %r999, %r1099, %r1097;

prmt.b32 %r999, %r999, 0, 0xba98;

xor.b32 %r1100, %r999, %r1079;
or.b32 %r1101, %r1100, -2139062144;
and.b32 %r1102, %r999, 2139062143;
and.b32 %r1103, %r1079, -2139062144;
sub.s32 %r1104, %r1101, %r1102;
xor.b32 %r1105, %r1103, %r1104;
xor.b32 %r1006, %r1105, -2139062144;

dp4a.s32.s32 %r1001, %r1002, %r23, %r830;


dp4a.s32.s32 %r1005, %r1006, %r24, %r1001;

shl.b16 %rs98, %rs87, 3;
cvt.u64.u16 %rd168, %rs98;
and.b64 %rd169, %rd168, 4088;
add.s64 %rd170, %rd211, %rd169;
ld.global.nc.v2.u32 {%r1106, %r1107}, [%rd170];
shr.u16 %rs99, %rs87, 9;
cvt.u32.u16 %r1110, %rs99;
popc.b32 %r1111, %r1110;
shl.b32 %r1112, %r1111, 7;
and.b32 %r1113, %r1112, 128;
or.b32 %r1114, %r1113, %r1110;
mul.lo.s32 %r1115, %r1114, 16843009;
and.b32 %r1116, %r1115, 134480385;
or.b32 %r1117, %r1116, -2139062144;
add.s32 %r1118, %r1117, -16843009;
or.b32 %r1009, %r1118, %r1116;

prmt.b32 %r1009, %r1009, 0, 0xba98;

xor.b32 %r1119, %r1009, %r1106;
or.b32 %r1120, %r1119, -2139062144;
and.b32 %r1121, %r1009, 2139062143;
and.b32 %r1122, %r1106, -2139062144;
sub.s32 %r1123, %r1120, %r1121;
xor.b32 %r1124, %r1122, %r1123;
xor.b32 %r1014, %r1124, -2139062144;
and.b32 %r1125, %r1115, -2143281136;
or.b32 %r1126, %r1125, -2139062144;
add.s32 %r1127, %r1126, -16843009;
or.b32 %r1011, %r1127, %r1125;

prmt.b32 %r1011, %r1011, 0, 0xba98;

xor.b32 %r1128, %r1011, %r1107;
or.b32 %r1129, %r1128, -2139062144;
and.b32 %r1130, %r1011, 2139062143;
and.b32 %r1131, %r1107, -2139062144;
sub.s32 %r1132, %r1129, %r1130;
xor.b32 %r1133, %r1131, %r1132;
xor.b32 %r1018, %r1133, -2139062144;

dp4a.s32.s32 %r1013, %r1014, %r25, %r1005;


dp4a.s32.s32 %r1017, %r1018, %r26, %r1013;

cvt.u32.u16 %r1134, %rs89;
cvt.u32.u16 %r1135, %rs90;
mul.lo.s32 %r1136, %r993, %r1134;
mad.lo.s32 %r1137, %r1017, %r1135, %r1136;
add.s32 %r1138, %r1017, %r993;
shr.u32 %r1139, %r1138, 31;
add.s32 %r1140, %r1138, %r1139;
shr.s32 %r1141, %r1140, 1;
add.s32 %r1142, %r1137, %r1141;
shr.s32 %r1143, %r1142, 31;
shr.u32 %r1144, %r1143, 30;
add.s32 %r1145, %r1142, %r1144;
shr.s32 %r1146, %r1145, 2;
add.s64 %rd171, %rd217, %rd12;
ld.global.u16 %rs84, [%rd171];

{ cvt.f32.f16 %f121, %rs84;}



{.reg .f16 low,high;
mov.b32 {low,high},%r27;
cvt.f32.f16 %f122, low;}


mul.ftz.f32 %f123, %f121, %f122;
cvt.rn.f32.s32 %f124, %r1146;
fma.rn.ftz.f32 %f32, %f123, %f124, %f32;

$L__BB213_16:
mov.u64 %rd209, iq2xs_grid;
add.s64 %rd172, %rd30, %rd18;
ld.global.nc.u16 %rs101, [%rd172+2];
ld.global.nc.u16 %rs102, [%rd172+4];
ld.global.nc.u16 %rs103, [%rd172+6];
add.s64 %rd173, %rd31, %rd18;
ld.global.nc.u8 %rs104, [%rd173];
and.b16 %rs105, %rs104, 240;
and.b16 %rs106, %rs104, 15;
shr.u16 %rs107, %rs105, 4;
ld.global.nc.u16 %rs108, [%rd172];
shl.b16 %rs109, %rs108, 3;
cvt.u64.u16 %rd174, %rs109;
and.b64 %rd175, %rd174, 4088;
add.s64 %rd177, %rd209, %rd175;
ld.global.nc.v2.u32 {%r1196, %r1197}, [%rd177];
shr.u16 %rs110, %rs108, 9;
cvt.u32.u16 %r1200, %rs110;
popc.b32 %r1201, %r1200;
shl.b32 %r1202, %r1201, 7;
and.b32 %r1203, %r1202, 128;
or.b32 %r1204, %r1203, %r1200;
mul.lo.s32 %r1205, %r1204, 16843009;
and.b32 %r1206, %r1205, 134480385;
or.b32 %r1207, %r1206, -2139062144;
add.s32 %r1208, %r1207, -16843009;
or.b32 %r1147, %r1208, %r1206;

prmt.b32 %r1147, %r1147, 0, 0xba98;

xor.b32 %r1209, %r1147, %r1196;
or.b32 %r1210, %r1209, -2139062144;
and.b32 %r1211, %r1147, 2139062143;
and.b32 %r1212, %r1196, -2139062144;
sub.s32 %r1213, %r1210, %r1211;
xor.b32 %r1214, %r1212, %r1213;
xor.b32 %r1152, %r1214, -2139062144;
and.b32 %r1215, %r1205, -2143281136;
or.b32 %r1216, %r1215, -2139062144;
add.s32 %r1217, %r1216, -16843009;
or.b32 %r1149, %r1217, %r1215;

prmt.b32 %r1149, %r1149, 0, 0xba98;

xor.b32 %r1218, %r1149, %r1197;
or.b32 %r1219, %r1218, -2139062144;
and.b32 %r1220, %r1149, 2139062143;
and.b32 %r1221, %r1197, -2139062144;
sub.s32 %r1222, %r1219, %r1220;
xor.b32 %r1223, %r1221, %r1222;
xor.b32 %r1156, %r1223, -2139062144;

dp4a.s32.s32 %r1151, %r1152, %r19, %r830;


dp4a.s32.s32 %r1155, %r1156, %r20, %r1151;

shl.b16 %rs111, %rs101, 3;
cvt.u64.u16 %rd178, %rs111;
and.b64 %rd179, %rd178, 4088;
add.s64 %rd180, %rd209, %rd179;
ld.global.nc.v2.u32 {%r1224, %r1225}, [%rd180];
shr.u16 %rs112, %rs101, 9;
cvt.u32.u16 %r1228, %rs112;
popc.b32 %r1229, %r1228;
shl.b32 %r1230, %r1229, 7;
and.b32 %r1231, %r1230, 128;
or.b32 %r1232, %r1231, %r1228;
mul.lo.s32 %r1233, %r1232, 16843009;
and.b32 %r1234, %r1233, 134480385;
or.b32 %r1235, %r1234, -2139062144;
add.s32 %r1236, %r1235, -16843009;
or.b32 %r1159, %r1236, %r1234;

prmt.b32 %r1159, %r1159, 0, 0xba98;

xor.b32 %r1237, %r1159, %r1224;
or.b32 %r1238, %r1237, -2139062144;
and.b32 %r1239, %r1159, 2139062143;
and.b32 %r1240, %r1224, -2139062144;
sub.s32 %r1241, %r1238, %r1239;
xor.b32 %r1242, %r1240, %r1241;
xor.b32 %r1164, %r1242, -2139062144;
and.b32 %r1243, %r1233, -2143281136;
or.b32 %r1244, %r1243, -2139062144;
add.s32 %r1245, %r1244, -16843009;
or.b32 %r1161, %r1245, %r1243;

prmt.b32 %r1161, %r1161, 0, 0xba98;

xor.b32 %r1246, %r1161, %r1225;
or.b32 %r1247, %r1246, -2139062144;
and.b32 %r1248, %r1161, 2139062143;
and.b32 %r1249, %r1225, -2139062144;
sub.s32 %r1250, %r1247, %r1248;
xor.b32 %r1251, %r1249, %r1250;
xor.b32 %r1168, %r1251, -2139062144;

dp4a.s32.s32 %r1163, %r1164, %r21, %r1155;


dp4a.s32.s32 %r1167, %r1168, %r22, %r1163;

shl.b16 %rs113, %rs102, 3;
cvt.u64.u16 %rd181, %rs113;
and.b64 %rd182, %rd181, 4088;
add.s64 %rd183, %rd209, %rd182;
ld.global.nc.v2.u32 {%r1252, %r1253}, [%rd183];
shr.u16 %rs114, %rs102, 9;
cvt.u32.u16 %r1256, %rs114;
popc.b32 %r1257, %r1256;
shl.b32 %r1258, %r1257, 7;
and.b32 %r1259, %r1258, 128;
or.b32 %r1260, %r1259, %r1256;
mul.lo.s32 %r1261, %r1260, 16843009;
and.b32 %r1262, %r1261, 134480385;
or.b32 %r1263, %r1262, -2139062144;
add.s32 %r1264, %r1263, -16843009;
or.b32 %r1171, %r1264, %r1262;

prmt.b32 %r1171, %r1171, 0, 0xba98;

xor.b32 %r1265, %r1171, %r1252;
or.b32 %r1266, %r1265, -2139062144;
and.b32 %r1267, %r1171, 2139062143;
and.b32 %r1268, %r1252, -2139062144;
sub.s32 %r1269, %r1266, %r1267;
xor.b32 %r1270, %r1268, %r1269;
xor.b32 %r1176, %r1270, -2139062144;
and.b32 %r1271, %r1261, -2143281136;
or.b32 %r1272, %r1271, -2139062144;
add.s32 %r1273, %r1272, -16843009;
or.b32 %r1173, %r1273, %r1271;

prmt.b32 %r1173, %r1173, 0, 0xba98;

xor.b32 %r1274, %r1173, %r1253;
or.b32 %r1275, %r1274, -2139062144;
and.b32 %r1276, %r1173, 2139062143;
and.b32 %r1277, %r1253, -2139062144;
sub.s32 %r1278, %r1275, %r1276;
xor.b32 %r1279, %r1277, %r1278;
xor.b32 %r1180, %r1279, -2139062144;

dp4a.s32.s32 %r1175, %r1176, %r23, %r830;


dp4a.s32.s32 %r1179, %r1180, %r24, %r1175;

shl.b16 %rs115, %rs103, 3;
cvt.u64.u16 %rd184, %rs115;
and.b64 %rd185, %rd184, 4088;
add.s64 %rd186, %rd209, %rd185;
ld.global.nc.v2.u32 {%r1280, %r1281}, [%rd186];
shr.u16 %rs116, %rs103, 9;
cvt.u32.u16 %r1284, %rs116;
popc.b32 %r1285, %r1284;
shl.b32 %r1286, %r1285, 7;
and.b32 %r1287, %r1286, 128;
or.b32 %r1288, %r1287, %r1284;
mul.lo.s32 %r1289, %r1288, 16843009;
and.b32 %r1290, %r1289, 134480385;
or.b32 %r1291, %r1290, -2139062144;
add.s32 %r1292, %r1291, -16843009;
or.b32 %r1183, %r1292, %r1290;

prmt.b32 %r1183, %r1183, 0, 0xba98;

xor.b32 %r1293, %r1183, %r1280;
or.b32 %r1294, %r1293, -2139062144;
and.b32 %r1295, %r1183, 2139062143;
and.b32 %r1296, %r1280, -2139062144;
sub.s32 %r1297, %r1294, %r1295;
xor.b32 %r1298, %r1296, %r1297;
xor.b32 %r1188, %r1298, -2139062144;
and.b32 %r1299, %r1289, -2143281136;
or.b32 %r1300, %r1299, -2139062144;
add.s32 %r1301, %r1300, -16843009;
or.b32 %r1185, %r1301, %r1299;

prmt.b32 %r1185, %r1185, 0, 0xba98;

xor.b32 %r1302, %r1185, %r1281;
or.b32 %r1303, %r1302, -2139062144;
and.b32 %r1304, %r1185, 2139062143;
and.b32 %r1305, %r1281, -2139062144;
sub.s32 %r1306, %r1303, %r1304;
xor.b32 %r1307, %r1305, %r1306;
xor.b32 %r1192, %r1307, -2139062144;

dp4a.s32.s32 %r1187, %r1188, %r25, %r1179;


dp4a.s32.s32 %r1191, %r1192, %r26, %r1187;

cvt.u32.u16 %r1308, %rs106;
cvt.u32.u16 %r1309, %rs107;
mul.lo.s32 %r1310, %r1167, %r1308;
mad.lo.s32 %r1311, %r1191, %r1309, %r1310;
add.s32 %r1312, %r1191, %r1167;
shr.u32 %r1313, %r1312, 31;
add.s32 %r1314, %r1312, %r1313;
shr.s32 %r1315, %r1314, 1;
add.s32 %r1316, %r1311, %r1315;
shr.s32 %r1317, %r1316, 31;
shr.u32 %r1318, %r1317, 30;
add.s32 %r1319, %r1316, %r1318;
shr.s32 %r1320, %r1319, 2;
add.s64 %rd187, %rd32, %rd18;
ld.global.nc.u16 %rs100, [%rd187];

{ cvt.f32.f16 %f125, %rs100;}



{.reg .f16 low,high;
mov.b32 {low,high},%r27;
cvt.f32.f16 %f126, low;}


mul.ftz.f32 %f127, %f125, %f126;
cvt.rn.f32.s32 %f128, %r1320;
fma.rn.ftz.f32 %f27, %f127, %f128, %f27;
@%p13 bra $L__BB213_18;

mov.u64 %rd210, iq2xs_grid;
add.s64 %rd188, %rd217, %rd15;
ld.global.u16 %rs118, [%rd188+-4];
ld.global.u16 %rs119, [%rd188+-2];
ld.global.u16 %rs120, [%rd188];
add.s64 %rd189, %rd217, %rd16;
ld.global.u8 %rs121, [%rd189];
and.b16 %rs122, %rs121, 15;
shr.u16 %rs123, %rs121, 4;
ld.global.u16 %rs124, [%rd188+-6];
shl.b16 %rs125, %rs124, 3;
cvt.u64.u16 %rd190, %rs125;
and.b64 %rd191, %rd190, 4088;
add.s64 %rd193, %rd210, %rd191;
ld.global.nc.v2.u32 {%r1370, %r1371}, [%rd193];
mov.u32 %r1352, 0;
shr.u16 %rs126, %rs124, 9;
cvt.u32.u16 %r1374, %rs126;
popc.b32 %r1375, %r1374;
shl.b32 %r1376, %r1375, 7;
and.b32 %r1377, %r1376, 128;
or.b32 %r1378, %r1377, %r1374;
mul.lo.s32 %r1379, %r1378, 16843009;
and.b32 %r1380, %r1379, 134480385;
or.b32 %r1381, %r1380, -2139062144;
add.s32 %r1382, %r1381, -16843009;
or.b32 %r1321, %r1382, %r1380;

prmt.b32 %r1321, %r1321, 0, 0xba98;

xor.b32 %r1383, %r1321, %r1370;
or.b32 %r1384, %r1383, -2139062144;
and.b32 %r1385, %r1321, 2139062143;
and.b32 %r1386, %r1370, -2139062144;
sub.s32 %r1387, %r1384, %r1385;
xor.b32 %r1388, %r1386, %r1387;
xor.b32 %r1326, %r1388, -2139062144;
and.b32 %r1389, %r1379, -2143281136;
or.b32 %r1390, %r1389, -2139062144;
add.s32 %r1391, %r1390, -16843009;
or.b32 %r1323, %r1391, %r1389;

prmt.b32 %r1323, %r1323, 0, 0xba98;

xor.b32 %r1392, %r1323, %r1371;
or.b32 %r1393, %r1392, -2139062144;
and.b32 %r1394, %r1323, 2139062143;
and.b32 %r1395, %r1371, -2139062144;
sub.s32 %r1396, %r1393, %r1394;
xor.b32 %r1397, %r1395, %r1396;
xor.b32 %r1330, %r1397, -2139062144;

dp4a.s32.s32 %r1325, %r1326, %r19, %r1352;


dp4a.s32.s32 %r1329, %r1330, %r20, %r1325;

shl.b16 %rs127, %rs118, 3;
cvt.u64.u16 %rd194, %rs127;
and.b64 %rd195, %rd194, 4088;
add.s64 %rd196, %rd210, %rd195;
ld.global.nc.v2.u32 {%r1398, %r1399}, [%rd196];
shr.u16 %rs128, %rs118, 9;
cvt.u32.u16 %r1402, %rs128;
popc.b32 %r1403, %r1402;
shl.b32 %r1404, %r1403, 7;
and.b32 %r1405, %r1404, 128;
or.b32 %r1406, %r1405, %r1402;
mul.lo.s32 %r1407, %r1406, 16843009;
and.b32 %r1408, %r1407, 134480385;
or.b32 %r1409, %r1408, -2139062144;
add.s32 %r1410, %r1409, -16843009;
or.b32 %r1333, %r1410, %r1408;

prmt.b32 %r1333, %r1333, 0, 0xba98;

xor.b32 %r1411, %r1333, %r1398;
or.b32 %r1412, %r1411, -2139062144;
and.b32 %r1413, %r1333, 2139062143;
and.b32 %r1414, %r1398, -2139062144;
sub.s32 %r1415, %r1412, %r1413;
xor.b32 %r1416, %r1414, %r1415;
xor.b32 %r1338, %r1416, -2139062144;
and.b32 %r1417, %r1407, -2143281136;
or.b32 %r1418, %r1417, -2139062144;
add.s32 %r1419, %r1418, -16843009;
or.b32 %r1335, %r1419, %r1417;

prmt.b32 %r1335, %r1335, 0, 0xba98;

xor.b32 %r1420, %r1335, %r1399;
or.b32 %r1421, %r1420, -2139062144;
and.b32 %r1422, %r1335, 2139062143;
and.b32 %r1423, %r1399, -2139062144;
sub.s32 %r1424, %r1421, %r1422;
xor.b32 %r1425, %r1423, %r1424;
xor.b32 %r1342, %r1425, -2139062144;

dp4a.s32.s32 %r1337, %r1338, %r21, %r1329;


dp4a.s32.s32 %r1341, %r1342, %r22, %r1337;

shl.b16 %rs129, %rs119, 3;
cvt.u64.u16 %rd197, %rs129;
and.b64 %rd198, %rd197, 4088;
add.s64 %rd199, %rd210, %rd198;
ld.global.nc.v2.u32 {%r1426, %r1427}, [%rd199];
shr.u16 %rs130, %rs119, 9;
cvt.u32.u16 %r1430, %rs130;
popc.b32 %r1431, %r1430;
shl.b32 %r1432, %r1431, 7;
and.b32 %r1433, %r1432, 128;
or.b32 %r1434, %r1433, %r1430;
mul.lo.s32 %r1435, %r1434, 16843009;
and.b32 %r1436, %r1435, 134480385;
or.b32 %r1437, %r1436, -2139062144;
add.s32 %r1438, %r1437, -16843009;
or.b32 %r1345, %r1438, %r1436;

prmt.b32 %r1345, %r1345, 0, 0xba98;

xor.b32 %r1439, %r1345, %r1426;
or.b32 %r1440, %r1439, -2139062144;
and.b32 %r1441, %r1345, 2139062143;
and.b32 %r1442, %r1426, -2139062144;
sub.s32 %r1443, %r1440, %r1441;
xor.b32 %r1444, %r1442, %r1443;
xor.b32 %r1350, %r1444, -2139062144;
and.b32 %r1445, %r1435, -2143281136;
or.b32 %r1446, %r1445, -2139062144;
add.s32 %r1447, %r1446, -16843009;
or.b32 %r1347, %r1447, %r1445;

prmt.b32 %r1347, %r1347, 0, 0xba98;

xor.b32 %r1448, %r1347, %r1427;
or.b32 %r1449, %r1448, -2139062144;
and.b32 %r1450, %r1347, 2139062143;
and.b32 %r1451, %r1427, -2139062144;
sub.s32 %r1452, %r1449, %r1450;
xor.b32 %r1453, %r1451, %r1452;
xor.b32 %r1354, %r1453, -2139062144;

dp4a.s32.s32 %r1349, %r1350, %r23, %r1352;


dp4a.s32.s32 %r1353, %r1354, %r24, %r1349;

shl.b16 %rs131, %rs120, 3;
cvt.u64.u16 %rd200, %rs131;
and.b64 %rd201, %rd200, 4088;
add.s64 %rd202, %rd210, %rd201;
ld.global.nc.v2.u32 {%r1454, %r1455}, [%rd202];
shr.u16 %rs132, %rs120, 9;
cvt.u32.u16 %r1458, %rs132;
popc.b32 %r1459, %r1458;
shl.b32 %r1460, %r1459, 7;
and.b32 %r1461, %r1460, 128;
or.b32 %r1462, %r1461, %r1458;
mul.lo.s32 %r1463, %r1462, 16843009;
and.b32 %r1464, %r1463, 134480385;
or.b32 %r1465, %r1464, -2139062144;
add.s32 %r1466, %r1465, -16843009;
or.b32 %r1357, %r1466, %r1464;

prmt.b32 %r1357, %r1357, 0, 0xba98;

xor.b32 %r1467, %r1357, %r1454;
or.b32 %r1468, %r1467, -2139062144;
and.b32 %r1469, %r1357, 2139062143;
and.b32 %r1470, %r1454, -2139062144;
sub.s32 %r1471, %r1468, %r1469;
xor.b32 %r1472, %r1470, %r1471;
xor.b32 %r1362, %r1472, -2139062144;
and.b32 %r1473, %r1463, -2143281136;
or.b32 %r1474, %r1473, -2139062144;
add.s32 %r1475, %r1474, -16843009;
or.b32 %r1359, %r1475, %r1473;

prmt.b32 %r1359, %r1359, 0, 0xba98;

xor.b32 %r1476, %r1359, %r1455;
or.b32 %r1477, %r1476, -2139062144;
and.b32 %r1478, %r1359, 2139062143;
and.b32 %r1479, %r1455, -2139062144;
sub.s32 %r1480, %r1477, %r1478;
xor.b32 %r1481, %r1479, %r1480;
xor.b32 %r1366, %r1481, -2139062144;

dp4a.s32.s32 %r1361, %r1362, %r25, %r1353;


dp4a.s32.s32 %r1365, %r1366, %r26, %r1361;

cvt.u32.u16 %r1482, %rs122;
cvt.u32.u16 %r1483, %rs123;
mul.lo.s32 %r1484, %r1341, %r1482;
mad.lo.s32 %r1485, %r1365, %r1483, %r1484;
add.s32 %r1486, %r1365, %r1341;
shr.u32 %r1487, %r1486, 31;
add.s32 %r1488, %r1486, %r1487;
shr.s32 %r1489, %r1488, 1;
add.s32 %r1490, %r1485, %r1489;
shr.s32 %r1491, %r1490, 31;
shr.u32 %r1492, %r1491, 30;
add.s32 %r1493, %r1490, %r1492;
shr.s32 %r1494, %r1493, 2;
add.s64 %rd203, %rd217, %rd14;
ld.global.u16 %rs117, [%rd203];

{ cvt.f32.f16 %f129, %rs117;}



{.reg .f16 low,high;
mov.b32 {low,high},%r27;
cvt.f32.f16 %f130, low;}


mul.ftz.f32 %f131, %f129, %f130;
cvt.rn.f32.s32 %f132, %r1494;
fma.rn.ftz.f32 %f31, %f131, %f132, %f31;

$L__BB213_18:
add.s32 %r2159, %r2159, 128;
add.s64 %rd217, %rd217, 1184;
add.s64 %rd216, %rd216, 1184;
add.s32 %r2160, %r2160, 16;
setp.lt.s32 %p17, %r2160, %r5;
@%p17 bra $L__BB213_10;

$L__BB213_19:
mov.u32 %r2146, %tid.y;
setp.eq.s32 %p18, %r2146, 0;
@%p18 bra $L__BB213_23;

mov.u32 %r2147, %tid.y;
setp.ne.s64 %p19, %rd40, 0;
add.s32 %r30, %r2147, -1;
shl.b32 %r1495, %r2147, 9;
mov.u32 %r1496, _ZZ13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r1497, %r1496, %r1495;
shl.b32 %r1498, %r2, 2;
add.s32 %r1499, %r1497, %r1498;
st.shared.f32 [%r1499+-512], %f30;
@%p19 bra $L__BB213_22;
bra.uni $L__BB213_21;

$L__BB213_22:
shl.b32 %r1500, %r30, 9;
mov.u32 %r1501, _ZZ13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r1502, %r1501, %r1500;
add.s32 %r1504, %r1502, %r1498;
st.shared.f32 [%r1504], %f34;
add.s32 %r2143, %r1499, -512;
st.shared.f32 [%r2143+128], %f29;
st.shared.f32 [%r1504+128], %f33;
add.s32 %r2144, %r1499, -512;
st.shared.f32 [%r2144+256], %f28;
st.shared.f32 [%r1504+256], %f32;
add.s32 %r2145, %r1499, -512;
st.shared.f32 [%r2145+384], %f27;
st.shared.f32 [%r1504+384], %f31;
bra.uni $L__BB213_23;

$L__BB213_21:
add.s32 %r2140, %r1499, -512;
st.shared.f32 [%r2140+128], %f29;
add.s32 %r2141, %r1499, -512;
st.shared.f32 [%r2141+256], %f28;
add.s32 %r2142, %r1499, -512;
st.shared.f32 [%r2142+384], %f27;

$L__BB213_23:
mov.u32 %r2148, %tid.y;
setp.ne.s32 %p319, %r2148, 0;
bar.sync 0;
@%p319 bra $L__BB213_78;
bra.uni $L__BB213_24;

$L__BB213_78:
ret;

$L__BB213_24:
ld.param.u64 %rd208, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.f32 %f977, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
ld.param.u32 %r2151, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r2150, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
mov.u32 %r2149, %ctaid.z;
setp.ne.s64 %p21, %rd40, 0;
mov.u32 %r1506, %ctaid.y;
mad.lo.s32 %r1507, %r1506, %r2151, %r4;
mad.lo.s32 %r1508, %r2149, %r2150, %r1507;
neg.ftz.f32 %f35, %f977;
shl.b32 %r1509, %r2, 2;
mov.u32 %r1510, _ZZ13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r32, %r1510, %r1509;
cvta.to.global.u64 %rd204, %rd208;
mul.wide.u32 %rd205, %r1508, 4;
add.s64 %rd35, %rd204, %rd205;
@%p21 bra $L__BB213_33;
bra.uni $L__BB213_25;

$L__BB213_33:
ld.param.u32 %r2152, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
mov.u32 %r1584, _ZZ13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r34, %r1584, %r1509;
setp.gt.s32 %p54, %r2152, 2;
@%p54 bra $L__BB213_44;

ld.param.u32 %r2155, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p57, %r2155, 1;
@%p57 bra $L__BB213_62;

ld.param.u32 %r2156, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p58, %r2156, 2;
@%p58 bra $L__BB213_36;
bra.uni $L__BB213_70;

$L__BB213_36:
ld.shared.f32 %f533, [%r32];
add.ftz.f32 %f534, %f533, %f30;
ld.shared.f32 %f535, [%r34];
add.ftz.f32 %f536, %f535, %f34;
ld.shared.f32 %f537, [%r32+512];
add.ftz.f32 %f538, %f537, %f534;
ld.shared.f32 %f539, [%r34+512];
add.ftz.f32 %f540, %f539, %f536;
ld.shared.f32 %f541, [%r32+1024];
add.ftz.f32 %f542, %f541, %f538;
ld.shared.f32 %f543, [%r34+1024];
add.ftz.f32 %f544, %f543, %f540;
mov.b32 %r1807, %f542;
mov.u32 %r1808, 31;
mov.u32 %r1809, 16;
mov.u32 %r1810, -1;
shfl.sync.bfly.b32 %r1811|%p163, %r1807, %r1809, %r1808, %r1810;
mov.b32 %f545, %r1811;
add.ftz.f32 %f546, %f542, %f545;
mov.b32 %r1812, %f546;
mov.u32 %r1813, 8;
shfl.sync.bfly.b32 %r1814|%p164, %r1812, %r1813, %r1808, %r1810;
mov.b32 %f547, %r1814;
add.ftz.f32 %f548, %f546, %f547;
mov.b32 %r1815, %f548;
mov.u32 %r1816, 4;
shfl.sync.bfly.b32 %r1817|%p165, %r1815, %r1816, %r1808, %r1810;
mov.b32 %f549, %r1817;
add.ftz.f32 %f550, %f548, %f549;
mov.b32 %r1818, %f550;
mov.u32 %r1819, 2;
shfl.sync.bfly.b32 %r1820|%p166, %r1818, %r1819, %r1808, %r1810;
mov.b32 %f551, %r1820;
add.ftz.f32 %f552, %f550, %f551;
mov.b32 %r1821, %f552;
mov.u32 %r1822, 1;
shfl.sync.bfly.b32 %r1823|%p167, %r1821, %r1822, %r1808, %r1810;
mov.b32 %f553, %r1823;
add.ftz.f32 %f56, %f552, %f553;
mov.b32 %r1824, %f544;
shfl.sync.bfly.b32 %r1825|%p168, %r1824, %r1809, %r1808, %r1810;
mov.b32 %f554, %r1825;
add.ftz.f32 %f555, %f544, %f554;
mov.b32 %r1826, %f555;
shfl.sync.bfly.b32 %r1827|%p169, %r1826, %r1813, %r1808, %r1810;
mov.b32 %f556, %r1827;
add.ftz.f32 %f557, %f555, %f556;
mov.b32 %r1828, %f557;
shfl.sync.bfly.b32 %r1829|%p170, %r1828, %r1816, %r1808, %r1810;
mov.b32 %f558, %r1829;
add.ftz.f32 %f559, %f557, %f558;
mov.b32 %r1830, %f559;
shfl.sync.bfly.b32 %r1831|%p171, %r1830, %r1819, %r1808, %r1810;
mov.b32 %f560, %r1831;
add.ftz.f32 %f561, %f559, %f560;
mov.b32 %r1832, %f561;
shfl.sync.bfly.b32 %r1833|%p172, %r1832, %r1822, %r1808, %r1810;
mov.b32 %f562, %r1833;
add.ftz.f32 %f57, %f561, %f562;
setp.ge.u32 %p173, %r4, %r50;
setp.ne.s32 %p174, %r2, 0;
or.pred %p175, %p174, %p173;
@%p175 bra $L__BB213_38;

add.ftz.f32 %f563, %f5, %f56;
add.ftz.f32 %f564, %f4, %f57;
mul.ftz.f32 %f565, %f564, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f566, %f565;
add.ftz.f32 %f567, %f566, 0f3F800000;
div.approx.ftz.f32 %f568, %f564, %f567;
mul.ftz.f32 %f569, %f563, %f568;
st.global.f32 [%rd35], %f569;

$L__BB213_38:
ld.shared.f32 %f570, [%r32+128];
add.ftz.f32 %f571, %f570, %f29;
ld.shared.f32 %f572, [%r34+128];
add.ftz.f32 %f573, %f572, %f33;
ld.shared.f32 %f574, [%r32+640];
add.ftz.f32 %f575, %f574, %f571;
ld.shared.f32 %f576, [%r34+640];
add.ftz.f32 %f577, %f576, %f573;
ld.shared.f32 %f578, [%r32+1152];
add.ftz.f32 %f579, %f578, %f575;
ld.shared.f32 %f580, [%r34+1152];
add.ftz.f32 %f581, %f580, %f577;
mov.b32 %r1834, %f579;
shfl.sync.bfly.b32 %r1838|%p176, %r1834, %r1809, %r1808, %r1810;
mov.b32 %f582, %r1838;
add.ftz.f32 %f583, %f579, %f582;
mov.b32 %r1839, %f583;
shfl.sync.bfly.b32 %r1841|%p177, %r1839, %r1813, %r1808, %r1810;
mov.b32 %f584, %r1841;
add.ftz.f32 %f585, %f583, %f584;
mov.b32 %r1842, %f585;
shfl.sync.bfly.b32 %r1844|%p178, %r1842, %r1816, %r1808, %r1810;
mov.b32 %f586, %r1844;
add.ftz.f32 %f587, %f585, %f586;
mov.b32 %r1845, %f587;
shfl.sync.bfly.b32 %r1847|%p179, %r1845, %r1819, %r1808, %r1810;
mov.b32 %f588, %r1847;
add.ftz.f32 %f589, %f587, %f588;
mov.b32 %r1848, %f589;
shfl.sync.bfly.b32 %r1850|%p180, %r1848, %r1822, %r1808, %r1810;
mov.b32 %f590, %r1850;
add.ftz.f32 %f58, %f589, %f590;
mov.b32 %r1851, %f581;
shfl.sync.bfly.b32 %r1852|%p181, %r1851, %r1809, %r1808, %r1810;
mov.b32 %f591, %r1852;
add.ftz.f32 %f592, %f581, %f591;
mov.b32 %r1853, %f592;
shfl.sync.bfly.b32 %r1854|%p182, %r1853, %r1813, %r1808, %r1810;
mov.b32 %f593, %r1854;
add.ftz.f32 %f594, %f592, %f593;
mov.b32 %r1855, %f594;
shfl.sync.bfly.b32 %r1856|%p183, %r1855, %r1816, %r1808, %r1810;
mov.b32 %f595, %r1856;
add.ftz.f32 %f596, %f594, %f595;
mov.b32 %r1857, %f596;
shfl.sync.bfly.b32 %r1858|%p184, %r1857, %r1819, %r1808, %r1810;
mov.b32 %f597, %r1858;
add.ftz.f32 %f598, %f596, %f597;
mov.b32 %r1859, %f598;
shfl.sync.bfly.b32 %r1860|%p185, %r1859, %r1822, %r1808, %r1810;
mov.b32 %f599, %r1860;
add.ftz.f32 %f59, %f598, %f599;
add.s32 %r1861, %r4, 1;
setp.ge.u32 %p186, %r1861, %r50;
setp.ne.s32 %p187, %r2, 1;
or.pred %p188, %p187, %p186;
@%p188 bra $L__BB213_40;

add.ftz.f32 %f600, %f5, %f58;
add.ftz.f32 %f601, %f4, %f59;
mul.ftz.f32 %f602, %f601, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f603, %f602;
add.ftz.f32 %f604, %f603, 0f3F800000;
div.approx.ftz.f32 %f605, %f601, %f604;
mul.ftz.f32 %f606, %f600, %f605;
st.global.f32 [%rd35+4], %f606;

$L__BB213_40:
ld.shared.f32 %f607, [%r32+256];
add.ftz.f32 %f608, %f607, %f28;
ld.shared.f32 %f609, [%r34+256];
add.ftz.f32 %f610, %f609, %f32;
ld.shared.f32 %f611, [%r32+768];
add.ftz.f32 %f612, %f611, %f608;
ld.shared.f32 %f613, [%r34+768];
add.ftz.f32 %f614, %f613, %f610;
ld.shared.f32 %f615, [%r32+1280];
add.ftz.f32 %f616, %f615, %f612;
ld.shared.f32 %f617, [%r34+1280];
add.ftz.f32 %f618, %f617, %f614;
mov.b32 %r1862, %f616;
mov.u32 %r1863, 31;
mov.u32 %r1864, 16;
mov.u32 %r1865, -1;
shfl.sync.bfly.b32 %r1866|%p189, %r1862, %r1864, %r1863, %r1865;
mov.b32 %f619, %r1866;
add.ftz.f32 %f620, %f616, %f619;
mov.b32 %r1867, %f620;
mov.u32 %r1868, 8;
shfl.sync.bfly.b32 %r1869|%p190, %r1867, %r1868, %r1863, %r1865;
mov.b32 %f621, %r1869;
add.ftz.f32 %f622, %f620, %f621;
mov.b32 %r1870, %f622;
mov.u32 %r1871, 4;
shfl.sync.bfly.b32 %r1872|%p191, %r1870, %r1871, %r1863, %r1865;
mov.b32 %f623, %r1872;
add.ftz.f32 %f624, %f622, %f623;
mov.b32 %r1873, %f624;
mov.u32 %r1874, 2;
shfl.sync.bfly.b32 %r1875|%p192, %r1873, %r1874, %r1863, %r1865;
mov.b32 %f625, %r1875;
add.ftz.f32 %f626, %f624, %f625;
mov.b32 %r1876, %f626;
mov.u32 %r1877, 1;
shfl.sync.bfly.b32 %r1878|%p193, %r1876, %r1877, %r1863, %r1865;
mov.b32 %f627, %r1878;
add.ftz.f32 %f60, %f626, %f627;
mov.b32 %r1879, %f618;
shfl.sync.bfly.b32 %r1880|%p194, %r1879, %r1864, %r1863, %r1865;
mov.b32 %f628, %r1880;
add.ftz.f32 %f629, %f618, %f628;
mov.b32 %r1881, %f629;
shfl.sync.bfly.b32 %r1882|%p195, %r1881, %r1868, %r1863, %r1865;
mov.b32 %f630, %r1882;
add.ftz.f32 %f631, %f629, %f630;
mov.b32 %r1883, %f631;
shfl.sync.bfly.b32 %r1884|%p196, %r1883, %r1871, %r1863, %r1865;
mov.b32 %f632, %r1884;
add.ftz.f32 %f633, %f631, %f632;
mov.b32 %r1885, %f633;
shfl.sync.bfly.b32 %r1886|%p197, %r1885, %r1874, %r1863, %r1865;
mov.b32 %f634, %r1886;
add.ftz.f32 %f635, %f633, %f634;
mov.b32 %r1887, %f635;
shfl.sync.bfly.b32 %r1888|%p198, %r1887, %r1877, %r1863, %r1865;
mov.b32 %f636, %r1888;
add.ftz.f32 %f61, %f635, %f636;
add.s32 %r1889, %r4, 2;
setp.ge.u32 %p199, %r1889, %r50;
setp.ne.s32 %p200, %r2, 2;
or.pred %p201, %p200, %p199;
@%p201 bra $L__BB213_42;

add.ftz.f32 %f637, %f5, %f60;
add.ftz.f32 %f638, %f4, %f61;
mul.ftz.f32 %f639, %f638, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f640, %f639;
add.ftz.f32 %f641, %f640, 0f3F800000;
div.approx.ftz.f32 %f642, %f638, %f641;
mul.ftz.f32 %f643, %f637, %f642;
st.global.f32 [%rd35+8], %f643;

$L__BB213_42:
ld.shared.f32 %f644, [%r32+384];
add.ftz.f32 %f645, %f644, %f27;
ld.shared.f32 %f646, [%r34+384];
add.ftz.f32 %f647, %f646, %f31;
ld.shared.f32 %f648, [%r32+896];
add.ftz.f32 %f649, %f648, %f645;
ld.shared.f32 %f650, [%r34+896];
add.ftz.f32 %f651, %f650, %f647;
ld.shared.f32 %f652, [%r32+1408];
add.ftz.f32 %f653, %f652, %f649;
ld.shared.f32 %f654, [%r34+1408];
add.ftz.f32 %f655, %f654, %f651;
mov.b32 %r1890, %f653;
shfl.sync.bfly.b32 %r1894|%p202, %r1890, %r1864, %r1863, %r1865;
mov.b32 %f656, %r1894;
add.ftz.f32 %f657, %f653, %f656;
mov.b32 %r1895, %f657;
shfl.sync.bfly.b32 %r1897|%p203, %r1895, %r1868, %r1863, %r1865;
mov.b32 %f658, %r1897;
add.ftz.f32 %f659, %f657, %f658;
mov.b32 %r1898, %f659;
shfl.sync.bfly.b32 %r1900|%p204, %r1898, %r1871, %r1863, %r1865;
mov.b32 %f660, %r1900;
add.ftz.f32 %f661, %f659, %f660;
mov.b32 %r1901, %f661;
shfl.sync.bfly.b32 %r1903|%p205, %r1901, %r1874, %r1863, %r1865;
mov.b32 %f662, %r1903;
add.ftz.f32 %f663, %f661, %f662;
mov.b32 %r1904, %f663;
shfl.sync.bfly.b32 %r1906|%p206, %r1904, %r1877, %r1863, %r1865;
mov.b32 %f664, %r1906;
add.ftz.f32 %f62, %f663, %f664;
mov.b32 %r1907, %f655;
shfl.sync.bfly.b32 %r1908|%p207, %r1907, %r1864, %r1863, %r1865;
mov.b32 %f665, %r1908;
add.ftz.f32 %f666, %f655, %f665;
mov.b32 %r1909, %f666;
shfl.sync.bfly.b32 %r1910|%p208, %r1909, %r1868, %r1863, %r1865;
mov.b32 %f667, %r1910;
add.ftz.f32 %f668, %f666, %f667;
mov.b32 %r1911, %f668;
shfl.sync.bfly.b32 %r1912|%p209, %r1911, %r1871, %r1863, %r1865;
mov.b32 %f669, %r1912;
add.ftz.f32 %f670, %f668, %f669;
mov.b32 %r1913, %f670;
shfl.sync.bfly.b32 %r1914|%p210, %r1913, %r1874, %r1863, %r1865;
mov.b32 %f671, %r1914;
add.ftz.f32 %f672, %f670, %f671;
mov.b32 %r1915, %f672;
shfl.sync.bfly.b32 %r1916|%p211, %r1915, %r1877, %r1863, %r1865;
mov.b32 %f673, %r1916;
add.ftz.f32 %f63, %f672, %f673;
add.s32 %r1917, %r4, 3;
setp.ge.u32 %p212, %r1917, %r50;
setp.ne.s32 %p213, %r2, 3;
or.pred %p214, %p213, %p212;
@%p214 bra $L__BB213_78;

add.ftz.f32 %f674, %f5, %f62;
add.ftz.f32 %f675, %f4, %f63;
mul.ftz.f32 %f676, %f675, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f677, %f676;
add.ftz.f32 %f678, %f677, 0f3F800000;
div.approx.ftz.f32 %f679, %f675, %f678;
mul.ftz.f32 %f680, %f674, %f679;
st.global.f32 [%rd35+12], %f680;
bra.uni $L__BB213_78;

$L__BB213_25:
ld.shared.f32 %f133, [%r32];
add.ftz.f32 %f134, %f133, %f30;
ld.shared.f32 %f135, [%r32+512];
add.ftz.f32 %f136, %f135, %f134;
ld.shared.f32 %f137, [%r32+1024];
add.ftz.f32 %f138, %f137, %f136;
mov.b32 %r1511, %f138;
mov.u32 %r1512, 31;
mov.u32 %r1513, 16;
mov.u32 %r1514, -1;
shfl.sync.bfly.b32 %r1515|%p22, %r1511, %r1513, %r1512, %r1514;
mov.b32 %f139, %r1515;
add.ftz.f32 %f140, %f138, %f139;
mov.b32 %r1516, %f140;
mov.u32 %r1517, 8;
shfl.sync.bfly.b32 %r1518|%p23, %r1516, %r1517, %r1512, %r1514;
mov.b32 %f141, %r1518;
add.ftz.f32 %f142, %f140, %f141;
mov.b32 %r1519, %f142;
mov.u32 %r1520, 4;
shfl.sync.bfly.b32 %r1521|%p24, %r1519, %r1520, %r1512, %r1514;
mov.b32 %f143, %r1521;
add.ftz.f32 %f144, %f142, %f143;
mov.b32 %r1522, %f144;
mov.u32 %r1523, 2;
shfl.sync.bfly.b32 %r1524|%p25, %r1522, %r1523, %r1512, %r1514;
mov.b32 %f145, %r1524;
add.ftz.f32 %f146, %f144, %f145;
mov.b32 %r1525, %f146;
mov.u32 %r1526, 1;
shfl.sync.bfly.b32 %r1527|%p26, %r1525, %r1526, %r1512, %r1514;
mov.b32 %f147, %r1527;
add.ftz.f32 %f36, %f146, %f147;
setp.ge.u32 %p27, %r4, %r50;
setp.ne.s32 %p28, %r2, 0;
or.pred %p29, %p28, %p27;
@%p29 bra $L__BB213_27;

add.ftz.f32 %f148, %f5, %f36;
st.global.f32 [%rd35], %f148;

$L__BB213_27:
ld.shared.f32 %f149, [%r32+128];
add.ftz.f32 %f150, %f149, %f29;
ld.shared.f32 %f151, [%r32+640];
add.ftz.f32 %f152, %f151, %f150;
ld.shared.f32 %f153, [%r32+1152];
add.ftz.f32 %f154, %f153, %f152;
mov.b32 %r1528, %f154;
shfl.sync.bfly.b32 %r1532|%p30, %r1528, %r1513, %r1512, %r1514;
mov.b32 %f155, %r1532;
add.ftz.f32 %f156, %f154, %f155;
mov.b32 %r1533, %f156;
shfl.sync.bfly.b32 %r1535|%p31, %r1533, %r1517, %r1512, %r1514;
mov.b32 %f157, %r1535;
add.ftz.f32 %f158, %f156, %f157;
mov.b32 %r1536, %f158;
shfl.sync.bfly.b32 %r1538|%p32, %r1536, %r1520, %r1512, %r1514;
mov.b32 %f159, %r1538;
add.ftz.f32 %f160, %f158, %f159;
mov.b32 %r1539, %f160;
shfl.sync.bfly.b32 %r1541|%p33, %r1539, %r1523, %r1512, %r1514;
mov.b32 %f161, %r1541;
add.ftz.f32 %f162, %f160, %f161;
mov.b32 %r1542, %f162;
shfl.sync.bfly.b32 %r1544|%p34, %r1542, %r1526, %r1512, %r1514;
mov.b32 %f163, %r1544;
add.ftz.f32 %f37, %f162, %f163;
add.s32 %r1545, %r4, 1;
setp.ge.u32 %p35, %r1545, %r50;
setp.ne.s32 %p36, %r2, 1;
or.pred %p37, %p36, %p35;
@%p37 bra $L__BB213_29;

add.ftz.f32 %f164, %f5, %f37;
st.global.f32 [%rd35+4], %f164;

$L__BB213_29:
ld.shared.f32 %f165, [%r32+256];
add.ftz.f32 %f166, %f165, %f28;
ld.shared.f32 %f167, [%r32+768];
add.ftz.f32 %f168, %f167, %f166;
ld.shared.f32 %f169, [%r32+1280];
add.ftz.f32 %f170, %f169, %f168;
mov.b32 %r1546, %f170;
mov.u32 %r1547, 31;
mov.u32 %r1548, 16;
mov.u32 %r1549, -1;
shfl.sync.bfly.b32 %r1550|%p38, %r1546, %r1548, %r1547, %r1549;
mov.b32 %f171, %r1550;
add.ftz.f32 %f172, %f170, %f171;
mov.b32 %r1551, %f172;
mov.u32 %r1552, 8;
shfl.sync.bfly.b32 %r1553|%p39, %r1551, %r1552, %r1547, %r1549;
mov.b32 %f173, %r1553;
add.ftz.f32 %f174, %f172, %f173;
mov.b32 %r1554, %f174;
mov.u32 %r1555, 4;
shfl.sync.bfly.b32 %r1556|%p40, %r1554, %r1555, %r1547, %r1549;
mov.b32 %f175, %r1556;
add.ftz.f32 %f176, %f174, %f175;
mov.b32 %r1557, %f176;
mov.u32 %r1558, 2;
shfl.sync.bfly.b32 %r1559|%p41, %r1557, %r1558, %r1547, %r1549;
mov.b32 %f177, %r1559;
add.ftz.f32 %f178, %f176, %f177;
mov.b32 %r1560, %f178;
mov.u32 %r1561, 1;
shfl.sync.bfly.b32 %r1562|%p42, %r1560, %r1561, %r1547, %r1549;
mov.b32 %f179, %r1562;
add.ftz.f32 %f38, %f178, %f179;
add.s32 %r1563, %r4, 2;
setp.ge.u32 %p43, %r1563, %r50;
setp.ne.s32 %p44, %r2, 2;
or.pred %p45, %p44, %p43;
@%p45 bra $L__BB213_31;

add.ftz.f32 %f180, %f5, %f38;
st.global.f32 [%rd35+8], %f180;

$L__BB213_31:
ld.shared.f32 %f181, [%r32+384];
add.ftz.f32 %f182, %f181, %f27;
ld.shared.f32 %f183, [%r32+896];
add.ftz.f32 %f184, %f183, %f182;
ld.shared.f32 %f185, [%r32+1408];
add.ftz.f32 %f186, %f185, %f184;
mov.b32 %r1564, %f186;
shfl.sync.bfly.b32 %r1568|%p46, %r1564, %r1548, %r1547, %r1549;
mov.b32 %f187, %r1568;
add.ftz.f32 %f188, %f186, %f187;
mov.b32 %r1569, %f188;
shfl.sync.bfly.b32 %r1571|%p47, %r1569, %r1552, %r1547, %r1549;
mov.b32 %f189, %r1571;
add.ftz.f32 %f190, %f188, %f189;
mov.b32 %r1572, %f190;
shfl.sync.bfly.b32 %r1574|%p48, %r1572, %r1555, %r1547, %r1549;
mov.b32 %f191, %r1574;
add.ftz.f32 %f192, %f190, %f191;
mov.b32 %r1575, %f192;
shfl.sync.bfly.b32 %r1577|%p49, %r1575, %r1558, %r1547, %r1549;
mov.b32 %f193, %r1577;
add.ftz.f32 %f194, %f192, %f193;
mov.b32 %r1578, %f194;
shfl.sync.bfly.b32 %r1580|%p50, %r1578, %r1561, %r1547, %r1549;
mov.b32 %f195, %r1580;
add.ftz.f32 %f39, %f194, %f195;
add.s32 %r1581, %r4, 3;
setp.ge.u32 %p51, %r1581, %r50;
setp.ne.s32 %p52, %r2, 3;
or.pred %p53, %p52, %p51;
@%p53 bra $L__BB213_78;

add.ftz.f32 %f196, %f5, %f39;
st.global.f32 [%rd35+12], %f196;
bra.uni $L__BB213_78;

$L__BB213_44:
ld.param.u32 %r2153, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p55, %r2153, 3;
@%p55 bra $L__BB213_54;

ld.param.u32 %r2154, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.ne.s32 %p56, %r2154, 6;
@%p56 bra $L__BB213_70;

ld.shared.f32 %f197, [%r32];
add.ftz.f32 %f198, %f197, %f30;
ld.shared.f32 %f199, [%r34];
add.ftz.f32 %f200, %f199, %f34;
ld.shared.f32 %f201, [%r32+512];
add.ftz.f32 %f202, %f201, %f198;
ld.shared.f32 %f203, [%r34+512];
add.ftz.f32 %f204, %f203, %f200;
ld.shared.f32 %f205, [%r32+1024];
add.ftz.f32 %f206, %f205, %f202;
ld.shared.f32 %f207, [%r34+1024];
add.ftz.f32 %f208, %f207, %f204;
mov.b32 %r1585, %f206;
mov.u32 %r1586, 31;
mov.u32 %r1587, 16;
mov.u32 %r1588, -1;
shfl.sync.bfly.b32 %r1589|%p59, %r1585, %r1587, %r1586, %r1588;
mov.b32 %f209, %r1589;
add.ftz.f32 %f210, %f206, %f209;
mov.b32 %r1590, %f210;
mov.u32 %r1591, 8;
shfl.sync.bfly.b32 %r1592|%p60, %r1590, %r1591, %r1586, %r1588;
mov.b32 %f211, %r1592;
add.ftz.f32 %f212, %f210, %f211;
mov.b32 %r1593, %f212;
mov.u32 %r1594, 4;
shfl.sync.bfly.b32 %r1595|%p61, %r1593, %r1594, %r1586, %r1588;
mov.b32 %f213, %r1595;
add.ftz.f32 %f214, %f212, %f213;
mov.b32 %r1596, %f214;
mov.u32 %r1597, 2;
shfl.sync.bfly.b32 %r1598|%p62, %r1596, %r1597, %r1586, %r1588;
mov.b32 %f215, %r1598;
add.ftz.f32 %f216, %f214, %f215;
mov.b32 %r1599, %f216;
mov.u32 %r1600, 1;
shfl.sync.bfly.b32 %r1601|%p63, %r1599, %r1600, %r1586, %r1588;
mov.b32 %f217, %r1601;
add.ftz.f32 %f40, %f216, %f217;
mov.b32 %r1602, %f208;
shfl.sync.bfly.b32 %r1603|%p64, %r1602, %r1587, %r1586, %r1588;
mov.b32 %f218, %r1603;
add.ftz.f32 %f219, %f208, %f218;
mov.b32 %r1604, %f219;
shfl.sync.bfly.b32 %r1605|%p65, %r1604, %r1591, %r1586, %r1588;
mov.b32 %f220, %r1605;
add.ftz.f32 %f221, %f219, %f220;
mov.b32 %r1606, %f221;
shfl.sync.bfly.b32 %r1607|%p66, %r1606, %r1594, %r1586, %r1588;
mov.b32 %f222, %r1607;
add.ftz.f32 %f223, %f221, %f222;
mov.b32 %r1608, %f223;
shfl.sync.bfly.b32 %r1609|%p67, %r1608, %r1597, %r1586, %r1588;
mov.b32 %f224, %r1609;
add.ftz.f32 %f225, %f223, %f224;
mov.b32 %r1610, %f225;
shfl.sync.bfly.b32 %r1611|%p68, %r1610, %r1600, %r1586, %r1588;
mov.b32 %f226, %r1611;
add.ftz.f32 %f41, %f225, %f226;
setp.ge.u32 %p69, %r4, %r50;
setp.ne.s32 %p70, %r2, 0;
or.pred %p71, %p70, %p69;
@%p71 bra $L__BB213_48;

ld.param.f32 %f981, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f227, %f5, %f40;
add.ftz.f32 %f228, %f4, %f41;
min.ftz.f32 %f229, %f228, %f981;
min.ftz.f32 %f230, %f227, %f981;
max.ftz.f32 %f231, %f230, %f35;
mul.ftz.f32 %f232, %f229, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f233, %f232;
add.ftz.f32 %f234, %f233, 0f3F800000;
div.approx.ftz.f32 %f235, %f229, %f234;
mul.ftz.f32 %f236, %f231, %f235;
st.global.f32 [%rd35], %f236;

$L__BB213_48:
ld.shared.f32 %f237, [%r32+128];
add.ftz.f32 %f238, %f237, %f29;
ld.shared.f32 %f239, [%r34+128];
add.ftz.f32 %f240, %f239, %f33;
ld.shared.f32 %f241, [%r32+640];
add.ftz.f32 %f242, %f241, %f238;
ld.shared.f32 %f243, [%r34+640];
add.ftz.f32 %f244, %f243, %f240;
ld.shared.f32 %f245, [%r32+1152];
add.ftz.f32 %f246, %f245, %f242;
ld.shared.f32 %f247, [%r34+1152];
add.ftz.f32 %f248, %f247, %f244;
mov.b32 %r1612, %f246;
shfl.sync.bfly.b32 %r1616|%p72, %r1612, %r1587, %r1586, %r1588;
mov.b32 %f249, %r1616;
add.ftz.f32 %f250, %f246, %f249;
mov.b32 %r1617, %f250;
shfl.sync.bfly.b32 %r1619|%p73, %r1617, %r1591, %r1586, %r1588;
mov.b32 %f251, %r1619;
add.ftz.f32 %f252, %f250, %f251;
mov.b32 %r1620, %f252;
shfl.sync.bfly.b32 %r1622|%p74, %r1620, %r1594, %r1586, %r1588;
mov.b32 %f253, %r1622;
add.ftz.f32 %f254, %f252, %f253;
mov.b32 %r1623, %f254;
shfl.sync.bfly.b32 %r1625|%p75, %r1623, %r1597, %r1586, %r1588;
mov.b32 %f255, %r1625;
add.ftz.f32 %f256, %f254, %f255;
mov.b32 %r1626, %f256;
shfl.sync.bfly.b32 %r1628|%p76, %r1626, %r1600, %r1586, %r1588;
mov.b32 %f257, %r1628;
add.ftz.f32 %f42, %f256, %f257;
mov.b32 %r1629, %f248;
shfl.sync.bfly.b32 %r1630|%p77, %r1629, %r1587, %r1586, %r1588;
mov.b32 %f258, %r1630;
add.ftz.f32 %f259, %f248, %f258;
mov.b32 %r1631, %f259;
shfl.sync.bfly.b32 %r1632|%p78, %r1631, %r1591, %r1586, %r1588;
mov.b32 %f260, %r1632;
add.ftz.f32 %f261, %f259, %f260;
mov.b32 %r1633, %f261;
shfl.sync.bfly.b32 %r1634|%p79, %r1633, %r1594, %r1586, %r1588;
mov.b32 %f262, %r1634;
add.ftz.f32 %f263, %f261, %f262;
mov.b32 %r1635, %f263;
shfl.sync.bfly.b32 %r1636|%p80, %r1635, %r1597, %r1586, %r1588;
mov.b32 %f264, %r1636;
add.ftz.f32 %f265, %f263, %f264;
mov.b32 %r1637, %f265;
shfl.sync.bfly.b32 %r1638|%p81, %r1637, %r1600, %r1586, %r1588;
mov.b32 %f266, %r1638;
add.ftz.f32 %f43, %f265, %f266;
add.s32 %r1639, %r4, 1;
setp.ge.u32 %p82, %r1639, %r50;
setp.ne.s32 %p83, %r2, 1;
or.pred %p84, %p83, %p82;
@%p84 bra $L__BB213_50;

ld.param.f32 %f980, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f267, %f5, %f42;
add.ftz.f32 %f268, %f4, %f43;
min.ftz.f32 %f269, %f268, %f980;
min.ftz.f32 %f270, %f267, %f980;
max.ftz.f32 %f271, %f270, %f35;
mul.ftz.f32 %f272, %f269, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f273, %f272;
add.ftz.f32 %f274, %f273, 0f3F800000;
div.approx.ftz.f32 %f275, %f269, %f274;
mul.ftz.f32 %f276, %f271, %f275;
st.global.f32 [%rd35+4], %f276;

$L__BB213_50:
ld.shared.f32 %f277, [%r32+256];
add.ftz.f32 %f278, %f277, %f28;
ld.shared.f32 %f279, [%r34+256];
add.ftz.f32 %f280, %f279, %f32;
ld.shared.f32 %f281, [%r32+768];
add.ftz.f32 %f282, %f281, %f278;
ld.shared.f32 %f283, [%r34+768];
add.ftz.f32 %f284, %f283, %f280;
ld.shared.f32 %f285, [%r32+1280];
add.ftz.f32 %f286, %f285, %f282;
ld.shared.f32 %f287, [%r34+1280];
add.ftz.f32 %f288, %f287, %f284;
mov.b32 %r1640, %f286;
mov.u32 %r1641, 31;
mov.u32 %r1642, 16;
mov.u32 %r1643, -1;
shfl.sync.bfly.b32 %r1644|%p85, %r1640, %r1642, %r1641, %r1643;
mov.b32 %f289, %r1644;
add.ftz.f32 %f290, %f286, %f289;
mov.b32 %r1645, %f290;
mov.u32 %r1646, 8;
shfl.sync.bfly.b32 %r1647|%p86, %r1645, %r1646, %r1641, %r1643;
mov.b32 %f291, %r1647;
add.ftz.f32 %f292, %f290, %f291;
mov.b32 %r1648, %f292;
mov.u32 %r1649, 4;
shfl.sync.bfly.b32 %r1650|%p87, %r1648, %r1649, %r1641, %r1643;
mov.b32 %f293, %r1650;
add.ftz.f32 %f294, %f292, %f293;
mov.b32 %r1651, %f294;
mov.u32 %r1652, 2;
shfl.sync.bfly.b32 %r1653|%p88, %r1651, %r1652, %r1641, %r1643;
mov.b32 %f295, %r1653;
add.ftz.f32 %f296, %f294, %f295;
mov.b32 %r1654, %f296;
mov.u32 %r1655, 1;
shfl.sync.bfly.b32 %r1656|%p89, %r1654, %r1655, %r1641, %r1643;
mov.b32 %f297, %r1656;
add.ftz.f32 %f44, %f296, %f297;
mov.b32 %r1657, %f288;
shfl.sync.bfly.b32 %r1658|%p90, %r1657, %r1642, %r1641, %r1643;
mov.b32 %f298, %r1658;
add.ftz.f32 %f299, %f288, %f298;
mov.b32 %r1659, %f299;
shfl.sync.bfly.b32 %r1660|%p91, %r1659, %r1646, %r1641, %r1643;
mov.b32 %f300, %r1660;
add.ftz.f32 %f301, %f299, %f300;
mov.b32 %r1661, %f301;
shfl.sync.bfly.b32 %r1662|%p92, %r1661, %r1649, %r1641, %r1643;
mov.b32 %f302, %r1662;
add.ftz.f32 %f303, %f301, %f302;
mov.b32 %r1663, %f303;
shfl.sync.bfly.b32 %r1664|%p93, %r1663, %r1652, %r1641, %r1643;
mov.b32 %f304, %r1664;
add.ftz.f32 %f305, %f303, %f304;
mov.b32 %r1665, %f305;
shfl.sync.bfly.b32 %r1666|%p94, %r1665, %r1655, %r1641, %r1643;
mov.b32 %f306, %r1666;
add.ftz.f32 %f45, %f305, %f306;
add.s32 %r1667, %r4, 2;
setp.ge.u32 %p95, %r1667, %r50;
setp.ne.s32 %p96, %r2, 2;
or.pred %p97, %p96, %p95;
@%p97 bra $L__BB213_52;

ld.param.f32 %f979, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f307, %f5, %f44;
add.ftz.f32 %f308, %f4, %f45;
min.ftz.f32 %f309, %f308, %f979;
min.ftz.f32 %f310, %f307, %f979;
max.ftz.f32 %f311, %f310, %f35;
mul.ftz.f32 %f312, %f309, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f313, %f312;
add.ftz.f32 %f314, %f313, 0f3F800000;
div.approx.ftz.f32 %f315, %f309, %f314;
mul.ftz.f32 %f316, %f311, %f315;
st.global.f32 [%rd35+8], %f316;

$L__BB213_52:
ld.shared.f32 %f317, [%r32+384];
add.ftz.f32 %f318, %f317, %f27;
ld.shared.f32 %f319, [%r34+384];
add.ftz.f32 %f320, %f319, %f31;
ld.shared.f32 %f321, [%r32+896];
add.ftz.f32 %f322, %f321, %f318;
ld.shared.f32 %f323, [%r34+896];
add.ftz.f32 %f324, %f323, %f320;
ld.shared.f32 %f325, [%r32+1408];
add.ftz.f32 %f326, %f325, %f322;
ld.shared.f32 %f327, [%r34+1408];
add.ftz.f32 %f328, %f327, %f324;
mov.b32 %r1668, %f326;
shfl.sync.bfly.b32 %r1672|%p98, %r1668, %r1642, %r1641, %r1643;
mov.b32 %f329, %r1672;
add.ftz.f32 %f330, %f326, %f329;
mov.b32 %r1673, %f330;
shfl.sync.bfly.b32 %r1675|%p99, %r1673, %r1646, %r1641, %r1643;
mov.b32 %f331, %r1675;
add.ftz.f32 %f332, %f330, %f331;
mov.b32 %r1676, %f332;
shfl.sync.bfly.b32 %r1678|%p100, %r1676, %r1649, %r1641, %r1643;
mov.b32 %f333, %r1678;
add.ftz.f32 %f334, %f332, %f333;
mov.b32 %r1679, %f334;
shfl.sync.bfly.b32 %r1681|%p101, %r1679, %r1652, %r1641, %r1643;
mov.b32 %f335, %r1681;
add.ftz.f32 %f336, %f334, %f335;
mov.b32 %r1682, %f336;
shfl.sync.bfly.b32 %r1684|%p102, %r1682, %r1655, %r1641, %r1643;
mov.b32 %f337, %r1684;
add.ftz.f32 %f46, %f336, %f337;
mov.b32 %r1685, %f328;
shfl.sync.bfly.b32 %r1686|%p103, %r1685, %r1642, %r1641, %r1643;
mov.b32 %f338, %r1686;
add.ftz.f32 %f339, %f328, %f338;
mov.b32 %r1687, %f339;
shfl.sync.bfly.b32 %r1688|%p104, %r1687, %r1646, %r1641, %r1643;
mov.b32 %f340, %r1688;
add.ftz.f32 %f341, %f339, %f340;
mov.b32 %r1689, %f341;
shfl.sync.bfly.b32 %r1690|%p105, %r1689, %r1649, %r1641, %r1643;
mov.b32 %f342, %r1690;
add.ftz.f32 %f343, %f341, %f342;
mov.b32 %r1691, %f343;
shfl.sync.bfly.b32 %r1692|%p106, %r1691, %r1652, %r1641, %r1643;
mov.b32 %f344, %r1692;
add.ftz.f32 %f345, %f343, %f344;
mov.b32 %r1693, %f345;
shfl.sync.bfly.b32 %r1694|%p107, %r1693, %r1655, %r1641, %r1643;
mov.b32 %f346, %r1694;
add.ftz.f32 %f47, %f345, %f346;
add.s32 %r1695, %r4, 3;
setp.ge.u32 %p108, %r1695, %r50;
setp.ne.s32 %p109, %r2, 3;
or.pred %p110, %p109, %p108;
@%p110 bra $L__BB213_78;

ld.param.f32 %f978, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f347, %f5, %f46;
add.ftz.f32 %f348, %f4, %f47;
min.ftz.f32 %f349, %f348, %f978;
min.ftz.f32 %f350, %f347, %f978;
max.ftz.f32 %f351, %f350, %f35;
mul.ftz.f32 %f352, %f349, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f353, %f352;
add.ftz.f32 %f354, %f353, 0f3F800000;
div.approx.ftz.f32 %f355, %f349, %f354;
mul.ftz.f32 %f356, %f351, %f355;
st.global.f32 [%rd35+12], %f356;
bra.uni $L__BB213_78;

$L__BB213_70:
ld.shared.f32 %f845, [%r32];
add.ftz.f32 %f846, %f845, %f30;
ld.shared.f32 %f847, [%r34];
add.ftz.f32 %f848, %f847, %f34;
ld.shared.f32 %f849, [%r32+512];
add.ftz.f32 %f850, %f849, %f846;
ld.shared.f32 %f851, [%r34+512];
add.ftz.f32 %f852, %f851, %f848;
ld.shared.f32 %f853, [%r32+1024];
add.ftz.f32 %f854, %f853, %f850;
ld.shared.f32 %f855, [%r34+1024];
add.ftz.f32 %f856, %f855, %f852;
mov.b32 %r2029, %f854;
mov.u32 %r2030, 31;
mov.u32 %r2031, 16;
mov.u32 %r2032, -1;
shfl.sync.bfly.b32 %r2033|%p267, %r2029, %r2031, %r2030, %r2032;
mov.b32 %f857, %r2033;
add.ftz.f32 %f858, %f854, %f857;
mov.b32 %r2034, %f858;
mov.u32 %r2035, 8;
shfl.sync.bfly.b32 %r2036|%p268, %r2034, %r2035, %r2030, %r2032;
mov.b32 %f859, %r2036;
add.ftz.f32 %f860, %f858, %f859;
mov.b32 %r2037, %f860;
mov.u32 %r2038, 4;
shfl.sync.bfly.b32 %r2039|%p269, %r2037, %r2038, %r2030, %r2032;
mov.b32 %f861, %r2039;
add.ftz.f32 %f862, %f860, %f861;
mov.b32 %r2040, %f862;
mov.u32 %r2041, 2;
shfl.sync.bfly.b32 %r2042|%p270, %r2040, %r2041, %r2030, %r2032;
mov.b32 %f863, %r2042;
add.ftz.f32 %f864, %f862, %f863;
mov.b32 %r2043, %f864;
mov.u32 %r2044, 1;
shfl.sync.bfly.b32 %r2045|%p271, %r2043, %r2044, %r2030, %r2032;
mov.b32 %f865, %r2045;
add.ftz.f32 %f72, %f864, %f865;
mov.b32 %r2046, %f856;
shfl.sync.bfly.b32 %r2047|%p272, %r2046, %r2031, %r2030, %r2032;
mov.b32 %f866, %r2047;
add.ftz.f32 %f867, %f856, %f866;
mov.b32 %r2048, %f867;
shfl.sync.bfly.b32 %r2049|%p273, %r2048, %r2035, %r2030, %r2032;
mov.b32 %f868, %r2049;
add.ftz.f32 %f869, %f867, %f868;
mov.b32 %r2050, %f869;
shfl.sync.bfly.b32 %r2051|%p274, %r2050, %r2038, %r2030, %r2032;
mov.b32 %f870, %r2051;
add.ftz.f32 %f871, %f869, %f870;
mov.b32 %r2052, %f871;
shfl.sync.bfly.b32 %r2053|%p275, %r2052, %r2041, %r2030, %r2032;
mov.b32 %f872, %r2053;
add.ftz.f32 %f873, %f871, %f872;
mov.b32 %r2054, %f873;
shfl.sync.bfly.b32 %r2055|%p276, %r2054, %r2044, %r2030, %r2032;
mov.b32 %f874, %r2055;
add.ftz.f32 %f73, %f873, %f874;
setp.ge.u32 %p277, %r4, %r50;
setp.ne.s32 %p278, %r2, 0;
or.pred %p279, %p278, %p277;
@%p279 bra $L__BB213_72;

add.ftz.f32 %f875, %f5, %f72;
add.ftz.f32 %f876, %f4, %f73;
mul.ftz.f32 %f877, %f875, %f876;
st.global.f32 [%rd35], %f877;

$L__BB213_72:
ld.shared.f32 %f878, [%r32+128];
add.ftz.f32 %f879, %f878, %f29;
ld.shared.f32 %f880, [%r34+128];
add.ftz.f32 %f881, %f880, %f33;
ld.shared.f32 %f882, [%r32+640];
add.ftz.f32 %f883, %f882, %f879;
ld.shared.f32 %f884, [%r34+640];
add.ftz.f32 %f885, %f884, %f881;
ld.shared.f32 %f886, [%r32+1152];
add.ftz.f32 %f887, %f886, %f883;
ld.shared.f32 %f888, [%r34+1152];
add.ftz.f32 %f889, %f888, %f885;
mov.b32 %r2056, %f887;
shfl.sync.bfly.b32 %r2060|%p280, %r2056, %r2031, %r2030, %r2032;
mov.b32 %f890, %r2060;
add.ftz.f32 %f891, %f887, %f890;
mov.b32 %r2061, %f891;
shfl.sync.bfly.b32 %r2063|%p281, %r2061, %r2035, %r2030, %r2032;
mov.b32 %f892, %r2063;
add.ftz.f32 %f893, %f891, %f892;
mov.b32 %r2064, %f893;
shfl.sync.bfly.b32 %r2066|%p282, %r2064, %r2038, %r2030, %r2032;
mov.b32 %f894, %r2066;
add.ftz.f32 %f895, %f893, %f894;
mov.b32 %r2067, %f895;
shfl.sync.bfly.b32 %r2069|%p283, %r2067, %r2041, %r2030, %r2032;
mov.b32 %f896, %r2069;
add.ftz.f32 %f897, %f895, %f896;
mov.b32 %r2070, %f897;
shfl.sync.bfly.b32 %r2072|%p284, %r2070, %r2044, %r2030, %r2032;
mov.b32 %f898, %r2072;
add.ftz.f32 %f74, %f897, %f898;
mov.b32 %r2073, %f889;
shfl.sync.bfly.b32 %r2074|%p285, %r2073, %r2031, %r2030, %r2032;
mov.b32 %f899, %r2074;
add.ftz.f32 %f900, %f889, %f899;
mov.b32 %r2075, %f900;
shfl.sync.bfly.b32 %r2076|%p286, %r2075, %r2035, %r2030, %r2032;
mov.b32 %f901, %r2076;
add.ftz.f32 %f902, %f900, %f901;
mov.b32 %r2077, %f902;
shfl.sync.bfly.b32 %r2078|%p287, %r2077, %r2038, %r2030, %r2032;
mov.b32 %f903, %r2078;
add.ftz.f32 %f904, %f902, %f903;
mov.b32 %r2079, %f904;
shfl.sync.bfly.b32 %r2080|%p288, %r2079, %r2041, %r2030, %r2032;
mov.b32 %f905, %r2080;
add.ftz.f32 %f906, %f904, %f905;
mov.b32 %r2081, %f906;
shfl.sync.bfly.b32 %r2082|%p289, %r2081, %r2044, %r2030, %r2032;
mov.b32 %f907, %r2082;
add.ftz.f32 %f75, %f906, %f907;
add.s32 %r2083, %r4, 1;
setp.ge.u32 %p290, %r2083, %r50;
setp.ne.s32 %p291, %r2, 1;
or.pred %p292, %p291, %p290;
@%p292 bra $L__BB213_74;

add.ftz.f32 %f908, %f5, %f74;
add.ftz.f32 %f909, %f4, %f75;
mul.ftz.f32 %f910, %f908, %f909;
st.global.f32 [%rd35+4], %f910;

$L__BB213_74:
ld.shared.f32 %f911, [%r32+256];
add.ftz.f32 %f912, %f911, %f28;
ld.shared.f32 %f913, [%r34+256];
add.ftz.f32 %f914, %f913, %f32;
ld.shared.f32 %f915, [%r32+768];
add.ftz.f32 %f916, %f915, %f912;
ld.shared.f32 %f917, [%r34+768];
add.ftz.f32 %f918, %f917, %f914;
ld.shared.f32 %f919, [%r32+1280];
add.ftz.f32 %f920, %f919, %f916;
ld.shared.f32 %f921, [%r34+1280];
add.ftz.f32 %f922, %f921, %f918;
mov.b32 %r2084, %f920;
mov.u32 %r2085, 31;
mov.u32 %r2086, 16;
mov.u32 %r2087, -1;
shfl.sync.bfly.b32 %r2088|%p293, %r2084, %r2086, %r2085, %r2087;
mov.b32 %f923, %r2088;
add.ftz.f32 %f924, %f920, %f923;
mov.b32 %r2089, %f924;
mov.u32 %r2090, 8;
shfl.sync.bfly.b32 %r2091|%p294, %r2089, %r2090, %r2085, %r2087;
mov.b32 %f925, %r2091;
add.ftz.f32 %f926, %f924, %f925;
mov.b32 %r2092, %f926;
mov.u32 %r2093, 4;
shfl.sync.bfly.b32 %r2094|%p295, %r2092, %r2093, %r2085, %r2087;
mov.b32 %f927, %r2094;
add.ftz.f32 %f928, %f926, %f927;
mov.b32 %r2095, %f928;
mov.u32 %r2096, 2;
shfl.sync.bfly.b32 %r2097|%p296, %r2095, %r2096, %r2085, %r2087;
mov.b32 %f929, %r2097;
add.ftz.f32 %f930, %f928, %f929;
mov.b32 %r2098, %f930;
mov.u32 %r2099, 1;
shfl.sync.bfly.b32 %r2100|%p297, %r2098, %r2099, %r2085, %r2087;
mov.b32 %f931, %r2100;
add.ftz.f32 %f76, %f930, %f931;
mov.b32 %r2101, %f922;
shfl.sync.bfly.b32 %r2102|%p298, %r2101, %r2086, %r2085, %r2087;
mov.b32 %f932, %r2102;
add.ftz.f32 %f933, %f922, %f932;
mov.b32 %r2103, %f933;
shfl.sync.bfly.b32 %r2104|%p299, %r2103, %r2090, %r2085, %r2087;
mov.b32 %f934, %r2104;
add.ftz.f32 %f935, %f933, %f934;
mov.b32 %r2105, %f935;
shfl.sync.bfly.b32 %r2106|%p300, %r2105, %r2093, %r2085, %r2087;
mov.b32 %f936, %r2106;
add.ftz.f32 %f937, %f935, %f936;
mov.b32 %r2107, %f937;
shfl.sync.bfly.b32 %r2108|%p301, %r2107, %r2096, %r2085, %r2087;
mov.b32 %f938, %r2108;
add.ftz.f32 %f939, %f937, %f938;
mov.b32 %r2109, %f939;
shfl.sync.bfly.b32 %r2110|%p302, %r2109, %r2099, %r2085, %r2087;
mov.b32 %f940, %r2110;
add.ftz.f32 %f77, %f939, %f940;
add.s32 %r2111, %r4, 2;
setp.ge.u32 %p303, %r2111, %r50;
setp.ne.s32 %p304, %r2, 2;
or.pred %p305, %p304, %p303;
@%p305 bra $L__BB213_76;

add.ftz.f32 %f941, %f5, %f76;
add.ftz.f32 %f942, %f4, %f77;
mul.ftz.f32 %f943, %f941, %f942;
st.global.f32 [%rd35+8], %f943;

$L__BB213_76:
ld.shared.f32 %f944, [%r32+384];
add.ftz.f32 %f945, %f944, %f27;
ld.shared.f32 %f946, [%r34+384];
add.ftz.f32 %f947, %f946, %f31;
ld.shared.f32 %f948, [%r32+896];
add.ftz.f32 %f949, %f948, %f945;
ld.shared.f32 %f950, [%r34+896];
add.ftz.f32 %f951, %f950, %f947;
ld.shared.f32 %f952, [%r32+1408];
add.ftz.f32 %f953, %f952, %f949;
ld.shared.f32 %f954, [%r34+1408];
add.ftz.f32 %f955, %f954, %f951;
mov.b32 %r2112, %f953;
shfl.sync.bfly.b32 %r2116|%p306, %r2112, %r2086, %r2085, %r2087;
mov.b32 %f956, %r2116;
add.ftz.f32 %f957, %f953, %f956;
mov.b32 %r2117, %f957;
shfl.sync.bfly.b32 %r2119|%p307, %r2117, %r2090, %r2085, %r2087;
mov.b32 %f958, %r2119;
add.ftz.f32 %f959, %f957, %f958;
mov.b32 %r2120, %f959;
shfl.sync.bfly.b32 %r2122|%p308, %r2120, %r2093, %r2085, %r2087;
mov.b32 %f960, %r2122;
add.ftz.f32 %f961, %f959, %f960;
mov.b32 %r2123, %f961;
shfl.sync.bfly.b32 %r2125|%p309, %r2123, %r2096, %r2085, %r2087;
mov.b32 %f962, %r2125;
add.ftz.f32 %f963, %f961, %f962;
mov.b32 %r2126, %f963;
shfl.sync.bfly.b32 %r2128|%p310, %r2126, %r2099, %r2085, %r2087;
mov.b32 %f964, %r2128;
add.ftz.f32 %f78, %f963, %f964;
mov.b32 %r2129, %f955;
shfl.sync.bfly.b32 %r2130|%p311, %r2129, %r2086, %r2085, %r2087;
mov.b32 %f965, %r2130;
add.ftz.f32 %f966, %f955, %f965;
mov.b32 %r2131, %f966;
shfl.sync.bfly.b32 %r2132|%p312, %r2131, %r2090, %r2085, %r2087;
mov.b32 %f967, %r2132;
add.ftz.f32 %f968, %f966, %f967;
mov.b32 %r2133, %f968;
shfl.sync.bfly.b32 %r2134|%p313, %r2133, %r2093, %r2085, %r2087;
mov.b32 %f969, %r2134;
add.ftz.f32 %f970, %f968, %f969;
mov.b32 %r2135, %f970;
shfl.sync.bfly.b32 %r2136|%p314, %r2135, %r2096, %r2085, %r2087;
mov.b32 %f971, %r2136;
add.ftz.f32 %f972, %f970, %f971;
mov.b32 %r2137, %f972;
shfl.sync.bfly.b32 %r2138|%p315, %r2137, %r2099, %r2085, %r2087;
mov.b32 %f973, %r2138;
add.ftz.f32 %f79, %f972, %f973;
add.s32 %r2139, %r4, 3;
setp.ge.u32 %p316, %r2139, %r50;
setp.ne.s32 %p317, %r2, 3;
or.pred %p318, %p317, %p316;
@%p318 bra $L__BB213_78;

add.ftz.f32 %f974, %f5, %f78;
add.ftz.f32 %f975, %f4, %f79;
mul.ftz.f32 %f976, %f974, %f975;
st.global.f32 [%rd35+12], %f976;
bra.uni $L__BB213_78;

$L__BB213_62:
ld.shared.f32 %f681, [%r32];
add.ftz.f32 %f682, %f681, %f30;
ld.shared.f32 %f683, [%r34];
add.ftz.f32 %f684, %f683, %f34;
ld.shared.f32 %f685, [%r32+512];
add.ftz.f32 %f686, %f685, %f682;
ld.shared.f32 %f687, [%r34+512];
add.ftz.f32 %f688, %f687, %f684;
ld.shared.f32 %f689, [%r32+1024];
add.ftz.f32 %f690, %f689, %f686;
ld.shared.f32 %f691, [%r34+1024];
add.ftz.f32 %f692, %f691, %f688;
mov.b32 %r1918, %f690;
mov.u32 %r1919, 31;
mov.u32 %r1920, 16;
mov.u32 %r1921, -1;
shfl.sync.bfly.b32 %r1922|%p215, %r1918, %r1920, %r1919, %r1921;
mov.b32 %f693, %r1922;
add.ftz.f32 %f694, %f690, %f693;
mov.b32 %r1923, %f694;
mov.u32 %r1924, 8;
shfl.sync.bfly.b32 %r1925|%p216, %r1923, %r1924, %r1919, %r1921;
mov.b32 %f695, %r1925;
add.ftz.f32 %f696, %f694, %f695;
mov.b32 %r1926, %f696;
mov.u32 %r1927, 4;
shfl.sync.bfly.b32 %r1928|%p217, %r1926, %r1927, %r1919, %r1921;
mov.b32 %f697, %r1928;
add.ftz.f32 %f698, %f696, %f697;
mov.b32 %r1929, %f698;
mov.u32 %r1930, 2;
shfl.sync.bfly.b32 %r1931|%p218, %r1929, %r1930, %r1919, %r1921;
mov.b32 %f699, %r1931;
add.ftz.f32 %f700, %f698, %f699;
mov.b32 %r1932, %f700;
mov.u32 %r1933, 1;
shfl.sync.bfly.b32 %r1934|%p219, %r1932, %r1933, %r1919, %r1921;
mov.b32 %f701, %r1934;
add.ftz.f32 %f64, %f700, %f701;
mov.b32 %r1935, %f692;
shfl.sync.bfly.b32 %r1936|%p220, %r1935, %r1920, %r1919, %r1921;
mov.b32 %f702, %r1936;
add.ftz.f32 %f703, %f692, %f702;
mov.b32 %r1937, %f703;
shfl.sync.bfly.b32 %r1938|%p221, %r1937, %r1924, %r1919, %r1921;
mov.b32 %f704, %r1938;
add.ftz.f32 %f705, %f703, %f704;
mov.b32 %r1939, %f705;
shfl.sync.bfly.b32 %r1940|%p222, %r1939, %r1927, %r1919, %r1921;
mov.b32 %f706, %r1940;
add.ftz.f32 %f707, %f705, %f706;
mov.b32 %r1941, %f707;
shfl.sync.bfly.b32 %r1942|%p223, %r1941, %r1930, %r1919, %r1921;
mov.b32 %f708, %r1942;
add.ftz.f32 %f709, %f707, %f708;
mov.b32 %r1943, %f709;
shfl.sync.bfly.b32 %r1944|%p224, %r1943, %r1933, %r1919, %r1921;
mov.b32 %f710, %r1944;
add.ftz.f32 %f65, %f709, %f710;
setp.ge.u32 %p225, %r4, %r50;
setp.ne.s32 %p226, %r2, 0;
or.pred %p227, %p226, %p225;
@%p227 bra $L__BB213_64;

add.ftz.f32 %f711, %f5, %f64;
add.ftz.f32 %f712, %f4, %f65;
mul.ftz.f32 %f713, %f712, 0f3F000000;
mul.ftz.f32 %f714, %f712, 0f3F4C422A;
mul.ftz.f32 %f715, %f712, 0f3D372713;
fma.rn.ftz.f32 %f716, %f712, %f715, 0f3F800000;
mul.ftz.f32 %f717, %f714, %f716;
tanh.approx.f32 %f718, %f717;
add.ftz.f32 %f719, %f718, 0f3F800000;
mul.ftz.f32 %f720, %f713, %f719;
mul.ftz.f32 %f721, %f711, %f720;
st.global.f32 [%rd35], %f721;

$L__BB213_64:
ld.shared.f32 %f722, [%r32+128];
add.ftz.f32 %f723, %f722, %f29;
ld.shared.f32 %f724, [%r34+128];
add.ftz.f32 %f725, %f724, %f33;
ld.shared.f32 %f726, [%r32+640];
add.ftz.f32 %f727, %f726, %f723;
ld.shared.f32 %f728, [%r34+640];
add.ftz.f32 %f729, %f728, %f725;
ld.shared.f32 %f730, [%r32+1152];
add.ftz.f32 %f731, %f730, %f727;
ld.shared.f32 %f732, [%r34+1152];
add.ftz.f32 %f733, %f732, %f729;
mov.b32 %r1945, %f731;
shfl.sync.bfly.b32 %r1949|%p228, %r1945, %r1920, %r1919, %r1921;
mov.b32 %f734, %r1949;
add.ftz.f32 %f735, %f731, %f734;
mov.b32 %r1950, %f735;
shfl.sync.bfly.b32 %r1952|%p229, %r1950, %r1924, %r1919, %r1921;
mov.b32 %f736, %r1952;
add.ftz.f32 %f737, %f735, %f736;
mov.b32 %r1953, %f737;
shfl.sync.bfly.b32 %r1955|%p230, %r1953, %r1927, %r1919, %r1921;
mov.b32 %f738, %r1955;
add.ftz.f32 %f739, %f737, %f738;
mov.b32 %r1956, %f739;
shfl.sync.bfly.b32 %r1958|%p231, %r1956, %r1930, %r1919, %r1921;
mov.b32 %f740, %r1958;
add.ftz.f32 %f741, %f739, %f740;
mov.b32 %r1959, %f741;
shfl.sync.bfly.b32 %r1961|%p232, %r1959, %r1933, %r1919, %r1921;
mov.b32 %f742, %r1961;
add.ftz.f32 %f66, %f741, %f742;
mov.b32 %r1962, %f733;
shfl.sync.bfly.b32 %r1963|%p233, %r1962, %r1920, %r1919, %r1921;
mov.b32 %f743, %r1963;
add.ftz.f32 %f744, %f733, %f743;
mov.b32 %r1964, %f744;
shfl.sync.bfly.b32 %r1965|%p234, %r1964, %r1924, %r1919, %r1921;
mov.b32 %f745, %r1965;
add.ftz.f32 %f746, %f744, %f745;
mov.b32 %r1966, %f746;
shfl.sync.bfly.b32 %r1967|%p235, %r1966, %r1927, %r1919, %r1921;
mov.b32 %f747, %r1967;
add.ftz.f32 %f748, %f746, %f747;
mov.b32 %r1968, %f748;
shfl.sync.bfly.b32 %r1969|%p236, %r1968, %r1930, %r1919, %r1921;
mov.b32 %f749, %r1969;
add.ftz.f32 %f750, %f748, %f749;
mov.b32 %r1970, %f750;
shfl.sync.bfly.b32 %r1971|%p237, %r1970, %r1933, %r1919, %r1921;
mov.b32 %f751, %r1971;
add.ftz.f32 %f67, %f750, %f751;
add.s32 %r1972, %r4, 1;
setp.ge.u32 %p238, %r1972, %r50;
setp.ne.s32 %p239, %r2, 1;
or.pred %p240, %p239, %p238;
@%p240 bra $L__BB213_66;

add.ftz.f32 %f752, %f5, %f66;
add.ftz.f32 %f753, %f4, %f67;
mul.ftz.f32 %f754, %f753, 0f3F000000;
mul.ftz.f32 %f755, %f753, 0f3F4C422A;
mul.ftz.f32 %f756, %f753, 0f3D372713;
fma.rn.ftz.f32 %f757, %f753, %f756, 0f3F800000;
mul.ftz.f32 %f758, %f755, %f757;
tanh.approx.f32 %f759, %f758;
add.ftz.f32 %f760, %f759, 0f3F800000;
mul.ftz.f32 %f761, %f754, %f760;
mul.ftz.f32 %f762, %f752, %f761;
st.global.f32 [%rd35+4], %f762;

$L__BB213_66:
ld.shared.f32 %f763, [%r32+256];
add.ftz.f32 %f764, %f763, %f28;
ld.shared.f32 %f765, [%r34+256];
add.ftz.f32 %f766, %f765, %f32;
ld.shared.f32 %f767, [%r32+768];
add.ftz.f32 %f768, %f767, %f764;
ld.shared.f32 %f769, [%r34+768];
add.ftz.f32 %f770, %f769, %f766;
ld.shared.f32 %f771, [%r32+1280];
add.ftz.f32 %f772, %f771, %f768;
ld.shared.f32 %f773, [%r34+1280];
add.ftz.f32 %f774, %f773, %f770;
mov.b32 %r1973, %f772;
mov.u32 %r1974, 31;
mov.u32 %r1975, 16;
mov.u32 %r1976, -1;
shfl.sync.bfly.b32 %r1977|%p241, %r1973, %r1975, %r1974, %r1976;
mov.b32 %f775, %r1977;
add.ftz.f32 %f776, %f772, %f775;
mov.b32 %r1978, %f776;
mov.u32 %r1979, 8;
shfl.sync.bfly.b32 %r1980|%p242, %r1978, %r1979, %r1974, %r1976;
mov.b32 %f777, %r1980;
add.ftz.f32 %f778, %f776, %f777;
mov.b32 %r1981, %f778;
mov.u32 %r1982, 4;
shfl.sync.bfly.b32 %r1983|%p243, %r1981, %r1982, %r1974, %r1976;
mov.b32 %f779, %r1983;
add.ftz.f32 %f780, %f778, %f779;
mov.b32 %r1984, %f780;
mov.u32 %r1985, 2;
shfl.sync.bfly.b32 %r1986|%p244, %r1984, %r1985, %r1974, %r1976;
mov.b32 %f781, %r1986;
add.ftz.f32 %f782, %f780, %f781;
mov.b32 %r1987, %f782;
mov.u32 %r1988, 1;
shfl.sync.bfly.b32 %r1989|%p245, %r1987, %r1988, %r1974, %r1976;
mov.b32 %f783, %r1989;
add.ftz.f32 %f68, %f782, %f783;
mov.b32 %r1990, %f774;
shfl.sync.bfly.b32 %r1991|%p246, %r1990, %r1975, %r1974, %r1976;
mov.b32 %f784, %r1991;
add.ftz.f32 %f785, %f774, %f784;
mov.b32 %r1992, %f785;
shfl.sync.bfly.b32 %r1993|%p247, %r1992, %r1979, %r1974, %r1976;
mov.b32 %f786, %r1993;
add.ftz.f32 %f787, %f785, %f786;
mov.b32 %r1994, %f787;
shfl.sync.bfly.b32 %r1995|%p248, %r1994, %r1982, %r1974, %r1976;
mov.b32 %f788, %r1995;
add.ftz.f32 %f789, %f787, %f788;
mov.b32 %r1996, %f789;
shfl.sync.bfly.b32 %r1997|%p249, %r1996, %r1985, %r1974, %r1976;
mov.b32 %f790, %r1997;
add.ftz.f32 %f791, %f789, %f790;
mov.b32 %r1998, %f791;
shfl.sync.bfly.b32 %r1999|%p250, %r1998, %r1988, %r1974, %r1976;
mov.b32 %f792, %r1999;
add.ftz.f32 %f69, %f791, %f792;
add.s32 %r2000, %r4, 2;
setp.ge.u32 %p251, %r2000, %r50;
setp.ne.s32 %p252, %r2, 2;
or.pred %p253, %p252, %p251;
@%p253 bra $L__BB213_68;

add.ftz.f32 %f793, %f5, %f68;
add.ftz.f32 %f794, %f4, %f69;
mul.ftz.f32 %f795, %f794, 0f3F000000;
mul.ftz.f32 %f796, %f794, 0f3F4C422A;
mul.ftz.f32 %f797, %f794, 0f3D372713;
fma.rn.ftz.f32 %f798, %f794, %f797, 0f3F800000;
mul.ftz.f32 %f799, %f796, %f798;
tanh.approx.f32 %f800, %f799;
add.ftz.f32 %f801, %f800, 0f3F800000;
mul.ftz.f32 %f802, %f795, %f801;
mul.ftz.f32 %f803, %f793, %f802;
st.global.f32 [%rd35+8], %f803;

$L__BB213_68:
ld.shared.f32 %f804, [%r32+384];
add.ftz.f32 %f805, %f804, %f27;
ld.shared.f32 %f806, [%r34+384];
add.ftz.f32 %f807, %f806, %f31;
ld.shared.f32 %f808, [%r32+896];
add.ftz.f32 %f809, %f808, %f805;
ld.shared.f32 %f810, [%r34+896];
add.ftz.f32 %f811, %f810, %f807;
ld.shared.f32 %f812, [%r32+1408];
add.ftz.f32 %f813, %f812, %f809;
ld.shared.f32 %f814, [%r34+1408];
add.ftz.f32 %f815, %f814, %f811;
mov.b32 %r2001, %f813;
shfl.sync.bfly.b32 %r2005|%p254, %r2001, %r1975, %r1974, %r1976;
mov.b32 %f816, %r2005;
add.ftz.f32 %f817, %f813, %f816;
mov.b32 %r2006, %f817;
shfl.sync.bfly.b32 %r2008|%p255, %r2006, %r1979, %r1974, %r1976;
mov.b32 %f818, %r2008;
add.ftz.f32 %f819, %f817, %f818;
mov.b32 %r2009, %f819;
shfl.sync.bfly.b32 %r2011|%p256, %r2009, %r1982, %r1974, %r1976;
mov.b32 %f820, %r2011;
add.ftz.f32 %f821, %f819, %f820;
mov.b32 %r2012, %f821;
shfl.sync.bfly.b32 %r2014|%p257, %r2012, %r1985, %r1974, %r1976;
mov.b32 %f822, %r2014;
add.ftz.f32 %f823, %f821, %f822;
mov.b32 %r2015, %f823;
shfl.sync.bfly.b32 %r2017|%p258, %r2015, %r1988, %r1974, %r1976;
mov.b32 %f824, %r2017;
add.ftz.f32 %f70, %f823, %f824;
mov.b32 %r2018, %f815;
shfl.sync.bfly.b32 %r2019|%p259, %r2018, %r1975, %r1974, %r1976;
mov.b32 %f825, %r2019;
add.ftz.f32 %f826, %f815, %f825;
mov.b32 %r2020, %f826;
shfl.sync.bfly.b32 %r2021|%p260, %r2020, %r1979, %r1974, %r1976;
mov.b32 %f827, %r2021;
add.ftz.f32 %f828, %f826, %f827;
mov.b32 %r2022, %f828;
shfl.sync.bfly.b32 %r2023|%p261, %r2022, %r1982, %r1974, %r1976;
mov.b32 %f829, %r2023;
add.ftz.f32 %f830, %f828, %f829;
mov.b32 %r2024, %f830;
shfl.sync.bfly.b32 %r2025|%p262, %r2024, %r1985, %r1974, %r1976;
mov.b32 %f831, %r2025;
add.ftz.f32 %f832, %f830, %f831;
mov.b32 %r2026, %f832;
shfl.sync.bfly.b32 %r2027|%p263, %r2026, %r1988, %r1974, %r1976;
mov.b32 %f833, %r2027;
add.ftz.f32 %f71, %f832, %f833;
add.s32 %r2028, %r4, 3;
setp.ge.u32 %p264, %r2028, %r50;
setp.ne.s32 %p265, %r2, 3;
or.pred %p266, %p265, %p264;
@%p266 bra $L__BB213_78;

add.ftz.f32 %f834, %f5, %f70;
add.ftz.f32 %f835, %f4, %f71;
mul.ftz.f32 %f836, %f835, 0f3F000000;
mul.ftz.f32 %f837, %f835, 0f3F4C422A;
mul.ftz.f32 %f838, %f835, 0f3D372713;
fma.rn.ftz.f32 %f839, %f835, %f838, 0f3F800000;
mul.ftz.f32 %f840, %f837, %f839;
tanh.approx.f32 %f841, %f840;
add.ftz.f32 %f842, %f841, 0f3F800000;
mul.ftz.f32 %f843, %f836, %f842;
mul.ftz.f32 %f844, %f834, %f843;
st.global.f32 [%rd35+12], %f844;
bra.uni $L__BB213_78;

$L__BB213_54:
ld.shared.f32 %f357, [%r32];
add.ftz.f32 %f358, %f357, %f30;
ld.shared.f32 %f359, [%r34];
add.ftz.f32 %f360, %f359, %f34;
ld.shared.f32 %f361, [%r32+512];
add.ftz.f32 %f362, %f361, %f358;
ld.shared.f32 %f363, [%r34+512];
add.ftz.f32 %f364, %f363, %f360;
ld.shared.f32 %f365, [%r32+1024];
add.ftz.f32 %f366, %f365, %f362;
ld.shared.f32 %f367, [%r34+1024];
add.ftz.f32 %f368, %f367, %f364;
mov.b32 %r1696, %f366;
mov.u32 %r1697, 31;
mov.u32 %r1698, 16;
mov.u32 %r1699, -1;
shfl.sync.bfly.b32 %r1700|%p111, %r1696, %r1698, %r1697, %r1699;
mov.b32 %f369, %r1700;
add.ftz.f32 %f370, %f366, %f369;
mov.b32 %r1701, %f370;
mov.u32 %r1702, 8;
shfl.sync.bfly.b32 %r1703|%p112, %r1701, %r1702, %r1697, %r1699;
mov.b32 %f371, %r1703;
add.ftz.f32 %f372, %f370, %f371;
mov.b32 %r1704, %f372;
mov.u32 %r1705, 4;
shfl.sync.bfly.b32 %r1706|%p113, %r1704, %r1705, %r1697, %r1699;
mov.b32 %f373, %r1706;
add.ftz.f32 %f374, %f372, %f373;
mov.b32 %r1707, %f374;
mov.u32 %r1708, 2;
shfl.sync.bfly.b32 %r1709|%p114, %r1707, %r1708, %r1697, %r1699;
mov.b32 %f375, %r1709;
add.ftz.f32 %f376, %f374, %f375;
mov.b32 %r1710, %f376;
mov.u32 %r1711, 1;
shfl.sync.bfly.b32 %r1712|%p115, %r1710, %r1711, %r1697, %r1699;
mov.b32 %f377, %r1712;
add.ftz.f32 %f48, %f376, %f377;
mov.b32 %r1713, %f368;
shfl.sync.bfly.b32 %r1714|%p116, %r1713, %r1698, %r1697, %r1699;
mov.b32 %f378, %r1714;
add.ftz.f32 %f379, %f368, %f378;
mov.b32 %r1715, %f379;
shfl.sync.bfly.b32 %r1716|%p117, %r1715, %r1702, %r1697, %r1699;
mov.b32 %f380, %r1716;
add.ftz.f32 %f381, %f379, %f380;
mov.b32 %r1717, %f381;
shfl.sync.bfly.b32 %r1718|%p118, %r1717, %r1705, %r1697, %r1699;
mov.b32 %f382, %r1718;
add.ftz.f32 %f383, %f381, %f382;
mov.b32 %r1719, %f383;
shfl.sync.bfly.b32 %r1720|%p119, %r1719, %r1708, %r1697, %r1699;
mov.b32 %f384, %r1720;
add.ftz.f32 %f385, %f383, %f384;
mov.b32 %r1721, %f385;
shfl.sync.bfly.b32 %r1722|%p120, %r1721, %r1711, %r1697, %r1699;
mov.b32 %f386, %r1722;
add.ftz.f32 %f49, %f385, %f386;
setp.ge.u32 %p121, %r4, %r50;
setp.ne.s32 %p122, %r2, 0;
or.pred %p123, %p122, %p121;
@%p123 bra $L__BB213_56;

add.ftz.f32 %f387, %f5, %f48;
add.ftz.f32 %f388, %f4, %f49;
mov.f32 %f389, 0f40E00000;
min.ftz.f32 %f390, %f388, %f389;
min.ftz.f32 %f391, %f387, %f389;
mov.f32 %f392, 0fC0E00000;
max.ftz.f32 %f393, %f391, %f392;
mul.ftz.f32 %f394, %f390, 0f3FD9DB23;
mul.ftz.f32 %f395, %f394, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f396, %f395;
add.ftz.f32 %f397, %f396, 0f3F800000;
div.approx.ftz.f32 %f398, %f390, %f397;
add.ftz.f32 %f399, %f393, 0f3F800000;
mul.ftz.f32 %f400, %f399, %f398;
st.global.f32 [%rd35], %f400;

$L__BB213_56:
ld.shared.f32 %f401, [%r32+128];
add.ftz.f32 %f402, %f401, %f29;
ld.shared.f32 %f403, [%r34+128];
add.ftz.f32 %f404, %f403, %f33;
ld.shared.f32 %f405, [%r32+640];
add.ftz.f32 %f406, %f405, %f402;
ld.shared.f32 %f407, [%r34+640];
add.ftz.f32 %f408, %f407, %f404;
ld.shared.f32 %f409, [%r32+1152];
add.ftz.f32 %f410, %f409, %f406;
ld.shared.f32 %f411, [%r34+1152];
add.ftz.f32 %f412, %f411, %f408;
mov.b32 %r1723, %f410;
shfl.sync.bfly.b32 %r1727|%p124, %r1723, %r1698, %r1697, %r1699;
mov.b32 %f413, %r1727;
add.ftz.f32 %f414, %f410, %f413;
mov.b32 %r1728, %f414;
shfl.sync.bfly.b32 %r1730|%p125, %r1728, %r1702, %r1697, %r1699;
mov.b32 %f415, %r1730;
add.ftz.f32 %f416, %f414, %f415;
mov.b32 %r1731, %f416;
shfl.sync.bfly.b32 %r1733|%p126, %r1731, %r1705, %r1697, %r1699;
mov.b32 %f417, %r1733;
add.ftz.f32 %f418, %f416, %f417;
mov.b32 %r1734, %f418;
shfl.sync.bfly.b32 %r1736|%p127, %r1734, %r1708, %r1697, %r1699;
mov.b32 %f419, %r1736;
add.ftz.f32 %f420, %f418, %f419;
mov.b32 %r1737, %f420;
shfl.sync.bfly.b32 %r1739|%p128, %r1737, %r1711, %r1697, %r1699;
mov.b32 %f421, %r1739;
add.ftz.f32 %f50, %f420, %f421;
mov.b32 %r1740, %f412;
shfl.sync.bfly.b32 %r1741|%p129, %r1740, %r1698, %r1697, %r1699;
mov.b32 %f422, %r1741;
add.ftz.f32 %f423, %f412, %f422;
mov.b32 %r1742, %f423;
shfl.sync.bfly.b32 %r1743|%p130, %r1742, %r1702, %r1697, %r1699;
mov.b32 %f424, %r1743;
add.ftz.f32 %f425, %f423, %f424;
mov.b32 %r1744, %f425;
shfl.sync.bfly.b32 %r1745|%p131, %r1744, %r1705, %r1697, %r1699;
mov.b32 %f426, %r1745;
add.ftz.f32 %f427, %f425, %f426;
mov.b32 %r1746, %f427;
shfl.sync.bfly.b32 %r1747|%p132, %r1746, %r1708, %r1697, %r1699;
mov.b32 %f428, %r1747;
add.ftz.f32 %f429, %f427, %f428;
mov.b32 %r1748, %f429;
shfl.sync.bfly.b32 %r1749|%p133, %r1748, %r1711, %r1697, %r1699;
mov.b32 %f430, %r1749;
add.ftz.f32 %f51, %f429, %f430;
add.s32 %r1750, %r4, 1;
setp.ge.u32 %p134, %r1750, %r50;
setp.ne.s32 %p135, %r2, 1;
or.pred %p136, %p135, %p134;
@%p136 bra $L__BB213_58;

add.ftz.f32 %f431, %f5, %f50;
add.ftz.f32 %f432, %f4, %f51;
mov.f32 %f433, 0f40E00000;
min.ftz.f32 %f434, %f432, %f433;
min.ftz.f32 %f435, %f431, %f433;
mov.f32 %f436, 0fC0E00000;
max.ftz.f32 %f437, %f435, %f436;
mul.ftz.f32 %f438, %f434, 0f3FD9DB23;
mul.ftz.f32 %f439, %f438, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f440, %f439;
add.ftz.f32 %f441, %f440, 0f3F800000;
div.approx.ftz.f32 %f442, %f434, %f441;
add.ftz.f32 %f443, %f437, 0f3F800000;
mul.ftz.f32 %f444, %f443, %f442;
st.global.f32 [%rd35+4], %f444;

$L__BB213_58:
ld.shared.f32 %f445, [%r32+256];
add.ftz.f32 %f446, %f445, %f28;
ld.shared.f32 %f447, [%r34+256];
add.ftz.f32 %f448, %f447, %f32;
ld.shared.f32 %f449, [%r32+768];
add.ftz.f32 %f450, %f449, %f446;
ld.shared.f32 %f451, [%r34+768];
add.ftz.f32 %f452, %f451, %f448;
ld.shared.f32 %f453, [%r32+1280];
add.ftz.f32 %f454, %f453, %f450;
ld.shared.f32 %f455, [%r34+1280];
add.ftz.f32 %f456, %f455, %f452;
mov.b32 %r1751, %f454;
mov.u32 %r1752, 31;
mov.u32 %r1753, 16;
mov.u32 %r1754, -1;
shfl.sync.bfly.b32 %r1755|%p137, %r1751, %r1753, %r1752, %r1754;
mov.b32 %f457, %r1755;
add.ftz.f32 %f458, %f454, %f457;
mov.b32 %r1756, %f458;
mov.u32 %r1757, 8;
shfl.sync.bfly.b32 %r1758|%p138, %r1756, %r1757, %r1752, %r1754;
mov.b32 %f459, %r1758;
add.ftz.f32 %f460, %f458, %f459;
mov.b32 %r1759, %f460;
mov.u32 %r1760, 4;
shfl.sync.bfly.b32 %r1761|%p139, %r1759, %r1760, %r1752, %r1754;
mov.b32 %f461, %r1761;
add.ftz.f32 %f462, %f460, %f461;
mov.b32 %r1762, %f462;
mov.u32 %r1763, 2;
shfl.sync.bfly.b32 %r1764|%p140, %r1762, %r1763, %r1752, %r1754;
mov.b32 %f463, %r1764;
add.ftz.f32 %f464, %f462, %f463;
mov.b32 %r1765, %f464;
mov.u32 %r1766, 1;
shfl.sync.bfly.b32 %r1767|%p141, %r1765, %r1766, %r1752, %r1754;
mov.b32 %f465, %r1767;
add.ftz.f32 %f52, %f464, %f465;
mov.b32 %r1768, %f456;
shfl.sync.bfly.b32 %r1769|%p142, %r1768, %r1753, %r1752, %r1754;
mov.b32 %f466, %r1769;
add.ftz.f32 %f467, %f456, %f466;
mov.b32 %r1770, %f467;
shfl.sync.bfly.b32 %r1771|%p143, %r1770, %r1757, %r1752, %r1754;
mov.b32 %f468, %r1771;
add.ftz.f32 %f469, %f467, %f468;
mov.b32 %r1772, %f469;
shfl.sync.bfly.b32 %r1773|%p144, %r1772, %r1760, %r1752, %r1754;
mov.b32 %f470, %r1773;
add.ftz.f32 %f471, %f469, %f470;
mov.b32 %r1774, %f471;
shfl.sync.bfly.b32 %r1775|%p145, %r1774, %r1763, %r1752, %r1754;
mov.b32 %f472, %r1775;
add.ftz.f32 %f473, %f471, %f472;
mov.b32 %r1776, %f473;
shfl.sync.bfly.b32 %r1777|%p146, %r1776, %r1766, %r1752, %r1754;
mov.b32 %f474, %r1777;
add.ftz.f32 %f53, %f473, %f474;
add.s32 %r1778, %r4, 2;
setp.ge.u32 %p147, %r1778, %r50;
setp.ne.s32 %p148, %r2, 2;
or.pred %p149, %p148, %p147;
@%p149 bra $L__BB213_60;

add.ftz.f32 %f475, %f5, %f52;
add.ftz.f32 %f476, %f4, %f53;
mov.f32 %f477, 0f40E00000;
min.ftz.f32 %f478, %f476, %f477;
min.ftz.f32 %f479, %f475, %f477;
mov.f32 %f480, 0fC0E00000;
max.ftz.f32 %f481, %f479, %f480;
mul.ftz.f32 %f482, %f478, 0f3FD9DB23;
mul.ftz.f32 %f483, %f482, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f484, %f483;
add.ftz.f32 %f485, %f484, 0f3F800000;
div.approx.ftz.f32 %f486, %f478, %f485;
add.ftz.f32 %f487, %f481, 0f3F800000;
mul.ftz.f32 %f488, %f487, %f486;
st.global.f32 [%rd35+8], %f488;

$L__BB213_60:
ld.shared.f32 %f489, [%r32+384];
add.ftz.f32 %f490, %f489, %f27;
ld.shared.f32 %f491, [%r34+384];
add.ftz.f32 %f492, %f491, %f31;
ld.shared.f32 %f493, [%r32+896];
add.ftz.f32 %f494, %f493, %f490;
ld.shared.f32 %f495, [%r34+896];
add.ftz.f32 %f496, %f495, %f492;
ld.shared.f32 %f497, [%r32+1408];
add.ftz.f32 %f498, %f497, %f494;
ld.shared.f32 %f499, [%r34+1408];
add.ftz.f32 %f500, %f499, %f496;
mov.b32 %r1779, %f498;
shfl.sync.bfly.b32 %r1783|%p150, %r1779, %r1753, %r1752, %r1754;
mov.b32 %f501, %r1783;
add.ftz.f32 %f502, %f498, %f501;
mov.b32 %r1784, %f502;
shfl.sync.bfly.b32 %r1786|%p151, %r1784, %r1757, %r1752, %r1754;
mov.b32 %f503, %r1786;
add.ftz.f32 %f504, %f502, %f503;
mov.b32 %r1787, %f504;
shfl.sync.bfly.b32 %r1789|%p152, %r1787, %r1760, %r1752, %r1754;
mov.b32 %f505, %r1789;
add.ftz.f32 %f506, %f504, %f505;
mov.b32 %r1790, %f506;
shfl.sync.bfly.b32 %r1792|%p153, %r1790, %r1763, %r1752, %r1754;
mov.b32 %f507, %r1792;
add.ftz.f32 %f508, %f506, %f507;
mov.b32 %r1793, %f508;
shfl.sync.bfly.b32 %r1795|%p154, %r1793, %r1766, %r1752, %r1754;
mov.b32 %f509, %r1795;
add.ftz.f32 %f54, %f508, %f509;
mov.b32 %r1796, %f500;
shfl.sync.bfly.b32 %r1797|%p155, %r1796, %r1753, %r1752, %r1754;
mov.b32 %f510, %r1797;
add.ftz.f32 %f511, %f500, %f510;
mov.b32 %r1798, %f511;
shfl.sync.bfly.b32 %r1799|%p156, %r1798, %r1757, %r1752, %r1754;
mov.b32 %f512, %r1799;
add.ftz.f32 %f513, %f511, %f512;
mov.b32 %r1800, %f513;
shfl.sync.bfly.b32 %r1801|%p157, %r1800, %r1760, %r1752, %r1754;
mov.b32 %f514, %r1801;
add.ftz.f32 %f515, %f513, %f514;
mov.b32 %r1802, %f515;
shfl.sync.bfly.b32 %r1803|%p158, %r1802, %r1763, %r1752, %r1754;
mov.b32 %f516, %r1803;
add.ftz.f32 %f517, %f515, %f516;
mov.b32 %r1804, %f517;
shfl.sync.bfly.b32 %r1805|%p159, %r1804, %r1766, %r1752, %r1754;
mov.b32 %f518, %r1805;
add.ftz.f32 %f55, %f517, %f518;
add.s32 %r1806, %r4, 3;
setp.ge.u32 %p160, %r1806, %r50;
setp.ne.s32 %p161, %r2, 3;
or.pred %p162, %p161, %p160;
@%p162 bra $L__BB213_78;

add.ftz.f32 %f519, %f5, %f54;
add.ftz.f32 %f520, %f4, %f55;
mov.f32 %f521, 0f40E00000;
min.ftz.f32 %f522, %f520, %f521;
min.ftz.f32 %f523, %f519, %f521;
mov.f32 %f524, 0fC0E00000;
max.ftz.f32 %f525, %f523, %f524;
mul.ftz.f32 %f526, %f522, 0f3FD9DB23;
mul.ftz.f32 %f527, %f526, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f528, %f527;
add.ftz.f32 %f529, %f528, 0f3F800000;
div.approx.ftz.f32 %f530, %f522, %f529;
add.ftz.f32 %f531, %f525, 0f3F800000;
mul.ftz.f32 %f532, %f531, %f530;
st.global.f32 [%rd35+12], %f532;
bra.uni $L__BB213_78;

}
.entry _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<38>;
.reg .b16 %rs<69>;
.reg .f32 %f<109>;
.reg .b32 %r<845>;
.reg .b64 %rd<94>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

ld.param.u64 %rd12, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd13, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd14, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
mov.u32 %r1, %tid.y;
shl.b32 %r39, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r39, %r2;
mov.u32 %r40, %ctaid.x;
shl.b32 %r4, %r40, 2;
shr.u32 %r5, %r38, 8;
mov.u32 %r842, %ctaid.y;
setp.eq.s64 %p1, %rd14, 0;
@%p1 bra $L__BB214_2;

cvta.to.global.u64 %rd16, %rd14;
mul.wide.u32 %rd17, %r842, 4;
add.s64 %rd18, %rd16, %rd17;
ld.global.nc.u32 %r841, [%rd18];
mul.hi.u32 %r41, %r842, %r21;
add.s32 %r42, %r41, %r842;
shr.u32 %r43, %r42, %r22;
mul.lo.s32 %r44, %r43, %r23;
sub.s32 %r842, %r842, %r44;
bra.uni $L__BB214_3;

$L__BB214_2:
mul.hi.u32 %r45, %r842, %r26;
add.s32 %r46, %r45, %r842;
shr.u32 %r841, %r46, %r27;

$L__BB214_3:
mov.u32 %r12, %ctaid.z;
shr.s32 %r47, %r3, 31;
shr.u32 %r48, %r47, 29;
add.s32 %r49, %r3, %r48;
shr.s32 %r844, %r49, 3;
setp.ge.s32 %p2, %r844, %r5;
mov.f32 %f105, 0f00000000;
mov.f32 %f106, %f105;
mov.f32 %f107, %f105;
mov.f32 %f108, %f105;
@%p2 bra $L__BB214_6;

and.b32 %r53, %r49, -8;
sub.s32 %r54, %r3, %r53;
shl.b32 %r55, %r54, 2;
cvt.s64.s32 %rd1, %r54;
shl.b32 %r843, %r844, 3;
mad.lo.s32 %r56, %r4, %r24, %r844;
mul.hi.u32 %r57, %r12, %r32;
add.s32 %r58, %r57, %r12;
shr.u32 %r59, %r58, %r33;
mad.lo.s32 %r60, %r59, %r35, %r56;
mad.lo.s32 %r61, %r841, %r29, %r60;
mul.wide.s32 %rd2, %r61, 74;
add.s64 %rd19, %rd2, %rd1;
add.s64 %rd3, %rd19, 66;
mul.wide.s32 %rd20, %r55, 2;
add.s64 %rd21, %rd2, %rd20;
add.s64 %rd4, %rd21, 8;
mul.wide.s32 %rd5, %r24, 74;
cvta.to.global.u64 %rd6, %rd13;
cvta.to.global.u64 %rd93, %rd12;
mul.lo.s32 %r62, %r842, %r30;
cvt.u64.u32 %rd22, %r62;
mul.lo.s32 %r63, %r12, %r36;
cvt.u64.u32 %rd23, %r63;
add.s64 %rd8, %rd22, %rd23;
mov.f32 %f105, 0f00000000;

$L__BB214_5:
cvt.u64.u32 %rd24, %r843;
add.s64 %rd25, %rd8, %rd24;
add.s64 %rd26, %rd93, %rd4;
add.s64 %rd27, %rd26, -6;
ld.global.nc.u16 %rs5, [%rd26+-4];
ld.global.nc.u16 %rs6, [%rd26+-2];
ld.global.nc.u16 %rs7, [%rd26];
add.s64 %rd28, %rd93, %rd3;
ld.global.nc.u8 %rs8, [%rd28];
and.b16 %rs9, %rs8, 240;
and.b16 %rs10, %rs8, 15;
shr.u16 %rs11, %rs9, 4;
add.s64 %rd29, %rd25, %rd1;
mul.lo.s64 %rd30, %rd29, 36;
ld.global.nc.u16 %rs12, [%rd26+-6];
shl.b16 %rs13, %rs12, 3;
cvt.u64.u16 %rd31, %rs13;
and.b64 %rd32, %rd31, 4088;
mov.u64 %rd33, iq2xs_grid;
add.s64 %rd34, %rd33, %rd32;
ld.global.nc.v2.u32 {%r260, %r261}, [%rd34];
mov.u32 %r242, 0;
shr.u16 %rs14, %rs12, 9;
cvt.u32.u16 %r264, %rs14;
popc.b32 %r265, %r264;
shl.b32 %r266, %r265, 7;
and.b32 %r267, %r266, 128;
or.b32 %r268, %r267, %r264;
mul.lo.s32 %r269, %r268, 16843009;
and.b32 %r270, %r269, 134480385;
or.b32 %r271, %r270, -2139062144;
add.s32 %r272, %r271, -16843009;
or.b32 %r64, %r272, %r270;

prmt.b32 %r64, %r64, 0, 0xba98;

xor.b32 %r273, %r64, %r260;
or.b32 %r274, %r273, -2139062144;
and.b32 %r275, %r64, 2139062143;
and.b32 %r276, %r260, -2139062144;
sub.s32 %r277, %r274, %r275;
xor.b32 %r278, %r276, %r277;
xor.b32 %r69, %r278, -2139062144;
add.s64 %rd35, %rd6, %rd30;
ld.global.nc.u32 %r70, [%rd35+4];
and.b32 %r279, %r269, -2143281136;
or.b32 %r280, %r279, -2139062144;
add.s32 %r281, %r280, -16843009;
or.b32 %r66, %r281, %r279;

prmt.b32 %r66, %r66, 0, 0xba98;

xor.b32 %r282, %r66, %r261;
or.b32 %r283, %r282, -2139062144;
and.b32 %r284, %r66, 2139062143;
and.b32 %r285, %r261, -2139062144;
sub.s32 %r286, %r283, %r284;
xor.b32 %r287, %r285, %r286;
xor.b32 %r73, %r287, -2139062144;
ld.global.nc.u32 %r74, [%rd35+8];

dp4a.s32.s32 %r68, %r69, %r70, %r242;


dp4a.s32.s32 %r72, %r73, %r74, %r68;

shl.b16 %rs15, %rs5, 3;
cvt.u64.u16 %rd36, %rs15;
and.b64 %rd37, %rd36, 4088;
add.s64 %rd38, %rd33, %rd37;
ld.global.nc.v2.u32 {%r288, %r289}, [%rd38];
shr.u16 %rs16, %rs5, 9;
cvt.u32.u16 %r292, %rs16;
popc.b32 %r293, %r292;
shl.b32 %r294, %r293, 7;
and.b32 %r295, %r294, 128;
or.b32 %r296, %r295, %r292;
mul.lo.s32 %r297, %r296, 16843009;
and.b32 %r298, %r297, 134480385;
or.b32 %r299, %r298, -2139062144;
add.s32 %r300, %r299, -16843009;
or.b32 %r76, %r300, %r298;

prmt.b32 %r76, %r76, 0, 0xba98;

xor.b32 %r301, %r76, %r288;
or.b32 %r302, %r301, -2139062144;
and.b32 %r303, %r76, 2139062143;
and.b32 %r304, %r288, -2139062144;
sub.s32 %r305, %r302, %r303;
xor.b32 %r306, %r304, %r305;
xor.b32 %r81, %r306, -2139062144;
ld.global.nc.u32 %r82, [%rd35+12];
and.b32 %r307, %r297, -2143281136;
or.b32 %r308, %r307, -2139062144;
add.s32 %r309, %r308, -16843009;
or.b32 %r78, %r309, %r307;

prmt.b32 %r78, %r78, 0, 0xba98;

xor.b32 %r310, %r78, %r289;
or.b32 %r311, %r310, -2139062144;
and.b32 %r312, %r78, 2139062143;
and.b32 %r313, %r289, -2139062144;
sub.s32 %r314, %r311, %r312;
xor.b32 %r315, %r313, %r314;
xor.b32 %r85, %r315, -2139062144;
ld.global.nc.u32 %r86, [%rd35+16];

dp4a.s32.s32 %r80, %r81, %r82, %r72;


dp4a.s32.s32 %r84, %r85, %r86, %r80;

shl.b16 %rs17, %rs6, 3;
cvt.u64.u16 %rd39, %rs17;
and.b64 %rd40, %rd39, 4088;
add.s64 %rd41, %rd33, %rd40;
ld.global.nc.v2.u32 {%r316, %r317}, [%rd41];
shr.u16 %rs18, %rs6, 9;
cvt.u32.u16 %r320, %rs18;
popc.b32 %r321, %r320;
shl.b32 %r322, %r321, 7;
and.b32 %r323, %r322, 128;
or.b32 %r324, %r323, %r320;
mul.lo.s32 %r325, %r324, 16843009;
and.b32 %r326, %r325, 134480385;
or.b32 %r327, %r326, -2139062144;
add.s32 %r328, %r327, -16843009;
or.b32 %r88, %r328, %r326;

prmt.b32 %r88, %r88, 0, 0xba98;

xor.b32 %r329, %r88, %r316;
or.b32 %r330, %r329, -2139062144;
and.b32 %r331, %r88, 2139062143;
and.b32 %r332, %r316, -2139062144;
sub.s32 %r333, %r330, %r331;
xor.b32 %r334, %r332, %r333;
xor.b32 %r93, %r334, -2139062144;
ld.global.nc.u32 %r94, [%rd35+20];
and.b32 %r335, %r325, -2143281136;
or.b32 %r336, %r335, -2139062144;
add.s32 %r337, %r336, -16843009;
or.b32 %r90, %r337, %r335;

prmt.b32 %r90, %r90, 0, 0xba98;

xor.b32 %r338, %r90, %r317;
or.b32 %r339, %r338, -2139062144;
and.b32 %r340, %r90, 2139062143;
and.b32 %r341, %r317, -2139062144;
sub.s32 %r342, %r339, %r340;
xor.b32 %r343, %r341, %r342;
xor.b32 %r97, %r343, -2139062144;
ld.global.nc.u32 %r98, [%rd35+24];

dp4a.s32.s32 %r92, %r93, %r94, %r242;


dp4a.s32.s32 %r96, %r97, %r98, %r92;

shl.b16 %rs19, %rs7, 3;
cvt.u64.u16 %rd42, %rs19;
and.b64 %rd43, %rd42, 4088;
add.s64 %rd44, %rd33, %rd43;
ld.global.nc.v2.u32 {%r344, %r345}, [%rd44];
shr.u16 %rs20, %rs7, 9;
cvt.u32.u16 %r348, %rs20;
popc.b32 %r349, %r348;
shl.b32 %r350, %r349, 7;
and.b32 %r351, %r350, 128;
or.b32 %r352, %r351, %r348;
mul.lo.s32 %r353, %r352, 16843009;
and.b32 %r354, %r353, 134480385;
or.b32 %r355, %r354, -2139062144;
add.s32 %r356, %r355, -16843009;
or.b32 %r100, %r356, %r354;

prmt.b32 %r100, %r100, 0, 0xba98;

xor.b32 %r357, %r100, %r344;
or.b32 %r358, %r357, -2139062144;
and.b32 %r359, %r100, 2139062143;
and.b32 %r360, %r344, -2139062144;
sub.s32 %r361, %r358, %r359;
xor.b32 %r362, %r360, %r361;
xor.b32 %r105, %r362, -2139062144;
ld.global.nc.u32 %r106, [%rd35+28];
and.b32 %r363, %r353, -2143281136;
or.b32 %r364, %r363, -2139062144;
add.s32 %r365, %r364, -16843009;
or.b32 %r102, %r365, %r363;

prmt.b32 %r102, %r102, 0, 0xba98;

xor.b32 %r366, %r102, %r345;
or.b32 %r367, %r366, -2139062144;
and.b32 %r368, %r102, 2139062143;
and.b32 %r369, %r345, -2139062144;
sub.s32 %r370, %r367, %r368;
xor.b32 %r371, %r369, %r370;
xor.b32 %r109, %r371, -2139062144;
ld.global.nc.u32 %r110, [%rd35+32];

dp4a.s32.s32 %r104, %r105, %r106, %r96;


dp4a.s32.s32 %r108, %r109, %r110, %r104;

cvt.u32.u16 %r372, %rs10;
cvt.u32.u16 %r373, %rs11;
mul.lo.s32 %r374, %r84, %r372;
mad.lo.s32 %r375, %r108, %r373, %r374;
add.s32 %r376, %r108, %r84;
shr.u32 %r377, %r376, 31;
add.s32 %r378, %r376, %r377;
shr.s32 %r379, %r378, 1;
add.s32 %r380, %r375, %r379;
shr.s32 %r381, %r380, 31;
shr.u32 %r382, %r381, 30;
add.s32 %r383, %r380, %r382;
shr.s32 %r384, %r383, 2;
add.s64 %rd45, %rd93, %rd2;
ld.global.nc.u16 %rs1, [%rd45];

{ cvt.f32.f16 %f25, %rs1;}


ld.global.nc.u32 %r112, [%rd35];

{.reg .f16 low,high;
mov.b32 {low,high},%r112;
cvt.f32.f16 %f26, low;}


mul.ftz.f32 %f33, %f25, %f26;
cvt.rn.f32.s32 %f34, %r384;
fma.rn.ftz.f32 %f108, %f33, %f34, %f108;
add.s64 %rd46, %rd27, %rd5;
ld.global.nc.u16 %rs21, [%rd46+2];
ld.global.nc.u16 %rs22, [%rd46+4];
ld.global.nc.u16 %rs23, [%rd46+6];
add.s64 %rd47, %rd28, %rd5;
ld.global.nc.u8 %rs24, [%rd47];
and.b16 %rs25, %rs24, 240;
and.b16 %rs26, %rs24, 15;
shr.u16 %rs27, %rs25, 4;
ld.global.nc.u16 %rs28, [%rd46];
shl.b16 %rs29, %rs28, 3;
cvt.u64.u16 %rd48, %rs29;
and.b64 %rd49, %rd48, 4088;
add.s64 %rd50, %rd33, %rd49;
ld.global.nc.v2.u32 {%r385, %r386}, [%rd50];
shr.u16 %rs30, %rs28, 9;
cvt.u32.u16 %r389, %rs30;
popc.b32 %r390, %r389;
shl.b32 %r391, %r390, 7;
and.b32 %r392, %r391, 128;
or.b32 %r393, %r392, %r389;
mul.lo.s32 %r394, %r393, 16843009;
and.b32 %r395, %r394, 134480385;
or.b32 %r396, %r395, -2139062144;
add.s32 %r397, %r396, -16843009;
or.b32 %r113, %r397, %r395;

prmt.b32 %r113, %r113, 0, 0xba98;

xor.b32 %r398, %r113, %r385;
or.b32 %r399, %r398, -2139062144;
and.b32 %r400, %r113, 2139062143;
and.b32 %r401, %r385, -2139062144;
sub.s32 %r402, %r399, %r400;
xor.b32 %r403, %r401, %r402;
xor.b32 %r118, %r403, -2139062144;
and.b32 %r404, %r394, -2143281136;
or.b32 %r405, %r404, -2139062144;
add.s32 %r406, %r405, -16843009;
or.b32 %r115, %r406, %r404;

prmt.b32 %r115, %r115, 0, 0xba98;

xor.b32 %r407, %r115, %r386;
or.b32 %r408, %r407, -2139062144;
and.b32 %r409, %r115, 2139062143;
and.b32 %r410, %r386, -2139062144;
sub.s32 %r411, %r408, %r409;
xor.b32 %r412, %r410, %r411;
xor.b32 %r122, %r412, -2139062144;

dp4a.s32.s32 %r117, %r118, %r70, %r242;


dp4a.s32.s32 %r121, %r122, %r74, %r117;

shl.b16 %rs31, %rs21, 3;
cvt.u64.u16 %rd51, %rs31;
and.b64 %rd52, %rd51, 4088;
add.s64 %rd53, %rd33, %rd52;
ld.global.nc.v2.u32 {%r413, %r414}, [%rd53];
shr.u16 %rs32, %rs21, 9;
cvt.u32.u16 %r417, %rs32;
popc.b32 %r418, %r417;
shl.b32 %r419, %r418, 7;
and.b32 %r420, %r419, 128;
or.b32 %r421, %r420, %r417;
mul.lo.s32 %r422, %r421, 16843009;
and.b32 %r423, %r422, 134480385;
or.b32 %r424, %r423, -2139062144;
add.s32 %r425, %r424, -16843009;
or.b32 %r125, %r425, %r423;

prmt.b32 %r125, %r125, 0, 0xba98;

xor.b32 %r426, %r125, %r413;
or.b32 %r427, %r426, -2139062144;
and.b32 %r428, %r125, 2139062143;
and.b32 %r429, %r413, -2139062144;
sub.s32 %r430, %r427, %r428;
xor.b32 %r431, %r429, %r430;
xor.b32 %r130, %r431, -2139062144;
and.b32 %r432, %r422, -2143281136;
or.b32 %r433, %r432, -2139062144;
add.s32 %r434, %r433, -16843009;
or.b32 %r127, %r434, %r432;

prmt.b32 %r127, %r127, 0, 0xba98;

xor.b32 %r435, %r127, %r414;
or.b32 %r436, %r435, -2139062144;
and.b32 %r437, %r127, 2139062143;
and.b32 %r438, %r414, -2139062144;
sub.s32 %r439, %r436, %r437;
xor.b32 %r440, %r438, %r439;
xor.b32 %r134, %r440, -2139062144;

dp4a.s32.s32 %r129, %r130, %r82, %r121;


dp4a.s32.s32 %r133, %r134, %r86, %r129;

shl.b16 %rs33, %rs22, 3;
cvt.u64.u16 %rd54, %rs33;
and.b64 %rd55, %rd54, 4088;
add.s64 %rd56, %rd33, %rd55;
ld.global.nc.v2.u32 {%r441, %r442}, [%rd56];
shr.u16 %rs34, %rs22, 9;
cvt.u32.u16 %r445, %rs34;
popc.b32 %r446, %r445;
shl.b32 %r447, %r446, 7;
and.b32 %r448, %r447, 128;
or.b32 %r449, %r448, %r445;
mul.lo.s32 %r450, %r449, 16843009;
and.b32 %r451, %r450, 134480385;
or.b32 %r452, %r451, -2139062144;
add.s32 %r453, %r452, -16843009;
or.b32 %r137, %r453, %r451;

prmt.b32 %r137, %r137, 0, 0xba98;

xor.b32 %r454, %r137, %r441;
or.b32 %r455, %r454, -2139062144;
and.b32 %r456, %r137, 2139062143;
and.b32 %r457, %r441, -2139062144;
sub.s32 %r458, %r455, %r456;
xor.b32 %r459, %r457, %r458;
xor.b32 %r142, %r459, -2139062144;
and.b32 %r460, %r450, -2143281136;
or.b32 %r461, %r460, -2139062144;
add.s32 %r462, %r461, -16843009;
or.b32 %r139, %r462, %r460;

prmt.b32 %r139, %r139, 0, 0xba98;

xor.b32 %r463, %r139, %r442;
or.b32 %r464, %r463, -2139062144;
and.b32 %r465, %r139, 2139062143;
and.b32 %r466, %r442, -2139062144;
sub.s32 %r467, %r464, %r465;
xor.b32 %r468, %r466, %r467;
xor.b32 %r146, %r468, -2139062144;

dp4a.s32.s32 %r141, %r142, %r94, %r242;


dp4a.s32.s32 %r145, %r146, %r98, %r141;

shl.b16 %rs35, %rs23, 3;
cvt.u64.u16 %rd57, %rs35;
and.b64 %rd58, %rd57, 4088;
add.s64 %rd59, %rd33, %rd58;
ld.global.nc.v2.u32 {%r469, %r470}, [%rd59];
shr.u16 %rs36, %rs23, 9;
cvt.u32.u16 %r473, %rs36;
popc.b32 %r474, %r473;
shl.b32 %r475, %r474, 7;
and.b32 %r476, %r475, 128;
or.b32 %r477, %r476, %r473;
mul.lo.s32 %r478, %r477, 16843009;
and.b32 %r479, %r478, 134480385;
or.b32 %r480, %r479, -2139062144;
add.s32 %r481, %r480, -16843009;
or.b32 %r149, %r481, %r479;

prmt.b32 %r149, %r149, 0, 0xba98;

xor.b32 %r482, %r149, %r469;
or.b32 %r483, %r482, -2139062144;
and.b32 %r484, %r149, 2139062143;
and.b32 %r485, %r469, -2139062144;
sub.s32 %r486, %r483, %r484;
xor.b32 %r487, %r485, %r486;
xor.b32 %r154, %r487, -2139062144;
and.b32 %r488, %r478, -2143281136;
or.b32 %r489, %r488, -2139062144;
add.s32 %r490, %r489, -16843009;
or.b32 %r151, %r490, %r488;

prmt.b32 %r151, %r151, 0, 0xba98;

xor.b32 %r491, %r151, %r470;
or.b32 %r492, %r491, -2139062144;
and.b32 %r493, %r151, 2139062143;
and.b32 %r494, %r470, -2139062144;
sub.s32 %r495, %r492, %r493;
xor.b32 %r496, %r494, %r495;
xor.b32 %r158, %r496, -2139062144;

dp4a.s32.s32 %r153, %r154, %r106, %r145;


dp4a.s32.s32 %r157, %r158, %r110, %r153;

cvt.u32.u16 %r497, %rs26;
cvt.u32.u16 %r498, %rs27;
mul.lo.s32 %r499, %r133, %r497;
mad.lo.s32 %r500, %r157, %r498, %r499;
add.s32 %r501, %r157, %r133;
shr.u32 %r502, %r501, 31;
add.s32 %r503, %r501, %r502;
shr.s32 %r504, %r503, 1;
add.s32 %r505, %r500, %r504;
shr.s32 %r506, %r505, 31;
shr.u32 %r507, %r506, 30;
add.s32 %r508, %r505, %r507;
shr.s32 %r509, %r508, 2;
add.s64 %rd60, %rd45, %rd5;
ld.global.nc.u16 %rs2, [%rd60];

{ cvt.f32.f16 %f27, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r112;
cvt.f32.f16 %f28, low;}


mul.ftz.f32 %f35, %f27, %f28;
cvt.rn.f32.s32 %f36, %r509;
fma.rn.ftz.f32 %f107, %f35, %f36, %f107;
add.s64 %rd61, %rd46, %rd5;
ld.global.nc.u16 %rs37, [%rd61+2];
ld.global.nc.u16 %rs38, [%rd61+4];
ld.global.nc.u16 %rs39, [%rd61+6];
add.s64 %rd62, %rd47, %rd5;
ld.global.nc.u8 %rs40, [%rd62];
and.b16 %rs41, %rs40, 240;
and.b16 %rs42, %rs40, 15;
shr.u16 %rs43, %rs41, 4;
ld.global.nc.u16 %rs44, [%rd61];
shl.b16 %rs45, %rs44, 3;
cvt.u64.u16 %rd63, %rs45;
and.b64 %rd64, %rd63, 4088;
add.s64 %rd65, %rd33, %rd64;
ld.global.nc.v2.u32 {%r510, %r511}, [%rd65];
shr.u16 %rs46, %rs44, 9;
cvt.u32.u16 %r514, %rs46;
popc.b32 %r515, %r514;
shl.b32 %r516, %r515, 7;
and.b32 %r517, %r516, 128;
or.b32 %r518, %r517, %r514;
mul.lo.s32 %r519, %r518, 16843009;
and.b32 %r520, %r519, 134480385;
or.b32 %r521, %r520, -2139062144;
add.s32 %r522, %r521, -16843009;
or.b32 %r162, %r522, %r520;

prmt.b32 %r162, %r162, 0, 0xba98;

xor.b32 %r523, %r162, %r510;
or.b32 %r524, %r523, -2139062144;
and.b32 %r525, %r162, 2139062143;
and.b32 %r526, %r510, -2139062144;
sub.s32 %r527, %r524, %r525;
xor.b32 %r528, %r526, %r527;
xor.b32 %r167, %r528, -2139062144;
and.b32 %r529, %r519, -2143281136;
or.b32 %r530, %r529, -2139062144;
add.s32 %r531, %r530, -16843009;
or.b32 %r164, %r531, %r529;

prmt.b32 %r164, %r164, 0, 0xba98;

xor.b32 %r532, %r164, %r511;
or.b32 %r533, %r532, -2139062144;
and.b32 %r534, %r164, 2139062143;
and.b32 %r535, %r511, -2139062144;
sub.s32 %r536, %r533, %r534;
xor.b32 %r537, %r535, %r536;
xor.b32 %r171, %r537, -2139062144;

dp4a.s32.s32 %r166, %r167, %r70, %r242;


dp4a.s32.s32 %r170, %r171, %r74, %r166;

shl.b16 %rs47, %rs37, 3;
cvt.u64.u16 %rd66, %rs47;
and.b64 %rd67, %rd66, 4088;
add.s64 %rd68, %rd33, %rd67;
ld.global.nc.v2.u32 {%r538, %r539}, [%rd68];
shr.u16 %rs48, %rs37, 9;
cvt.u32.u16 %r542, %rs48;
popc.b32 %r543, %r542;
shl.b32 %r544, %r543, 7;
and.b32 %r545, %r544, 128;
or.b32 %r546, %r545, %r542;
mul.lo.s32 %r547, %r546, 16843009;
and.b32 %r548, %r547, 134480385;
or.b32 %r549, %r548, -2139062144;
add.s32 %r550, %r549, -16843009;
or.b32 %r174, %r550, %r548;

prmt.b32 %r174, %r174, 0, 0xba98;

xor.b32 %r551, %r174, %r538;
or.b32 %r552, %r551, -2139062144;
and.b32 %r553, %r174, 2139062143;
and.b32 %r554, %r538, -2139062144;
sub.s32 %r555, %r552, %r553;
xor.b32 %r556, %r554, %r555;
xor.b32 %r179, %r556, -2139062144;
and.b32 %r557, %r547, -2143281136;
or.b32 %r558, %r557, -2139062144;
add.s32 %r559, %r558, -16843009;
or.b32 %r176, %r559, %r557;

prmt.b32 %r176, %r176, 0, 0xba98;

xor.b32 %r560, %r176, %r539;
or.b32 %r561, %r560, -2139062144;
and.b32 %r562, %r176, 2139062143;
and.b32 %r563, %r539, -2139062144;
sub.s32 %r564, %r561, %r562;
xor.b32 %r565, %r563, %r564;
xor.b32 %r183, %r565, -2139062144;

dp4a.s32.s32 %r178, %r179, %r82, %r170;


dp4a.s32.s32 %r182, %r183, %r86, %r178;

shl.b16 %rs49, %rs38, 3;
cvt.u64.u16 %rd69, %rs49;
and.b64 %rd70, %rd69, 4088;
add.s64 %rd71, %rd33, %rd70;
ld.global.nc.v2.u32 {%r566, %r567}, [%rd71];
shr.u16 %rs50, %rs38, 9;
cvt.u32.u16 %r570, %rs50;
popc.b32 %r571, %r570;
shl.b32 %r572, %r571, 7;
and.b32 %r573, %r572, 128;
or.b32 %r574, %r573, %r570;
mul.lo.s32 %r575, %r574, 16843009;
and.b32 %r576, %r575, 134480385;
or.b32 %r577, %r576, -2139062144;
add.s32 %r578, %r577, -16843009;
or.b32 %r186, %r578, %r576;

prmt.b32 %r186, %r186, 0, 0xba98;

xor.b32 %r579, %r186, %r566;
or.b32 %r580, %r579, -2139062144;
and.b32 %r581, %r186, 2139062143;
and.b32 %r582, %r566, -2139062144;
sub.s32 %r583, %r580, %r581;
xor.b32 %r584, %r582, %r583;
xor.b32 %r191, %r584, -2139062144;
and.b32 %r585, %r575, -2143281136;
or.b32 %r586, %r585, -2139062144;
add.s32 %r587, %r586, -16843009;
or.b32 %r188, %r587, %r585;

prmt.b32 %r188, %r188, 0, 0xba98;

xor.b32 %r588, %r188, %r567;
or.b32 %r589, %r588, -2139062144;
and.b32 %r590, %r188, 2139062143;
and.b32 %r591, %r567, -2139062144;
sub.s32 %r592, %r589, %r590;
xor.b32 %r593, %r591, %r592;
xor.b32 %r195, %r593, -2139062144;

dp4a.s32.s32 %r190, %r191, %r94, %r242;


dp4a.s32.s32 %r194, %r195, %r98, %r190;

shl.b16 %rs51, %rs39, 3;
cvt.u64.u16 %rd72, %rs51;
and.b64 %rd73, %rd72, 4088;
add.s64 %rd74, %rd33, %rd73;
ld.global.nc.v2.u32 {%r594, %r595}, [%rd74];
shr.u16 %rs52, %rs39, 9;
cvt.u32.u16 %r598, %rs52;
popc.b32 %r599, %r598;
shl.b32 %r600, %r599, 7;
and.b32 %r601, %r600, 128;
or.b32 %r602, %r601, %r598;
mul.lo.s32 %r603, %r602, 16843009;
and.b32 %r604, %r603, 134480385;
or.b32 %r605, %r604, -2139062144;
add.s32 %r606, %r605, -16843009;
or.b32 %r198, %r606, %r604;

prmt.b32 %r198, %r198, 0, 0xba98;

xor.b32 %r607, %r198, %r594;
or.b32 %r608, %r607, -2139062144;
and.b32 %r609, %r198, 2139062143;
and.b32 %r610, %r594, -2139062144;
sub.s32 %r611, %r608, %r609;
xor.b32 %r612, %r610, %r611;
xor.b32 %r203, %r612, -2139062144;
and.b32 %r613, %r603, -2143281136;
or.b32 %r614, %r613, -2139062144;
add.s32 %r615, %r614, -16843009;
or.b32 %r200, %r615, %r613;

prmt.b32 %r200, %r200, 0, 0xba98;

xor.b32 %r616, %r200, %r595;
or.b32 %r617, %r616, -2139062144;
and.b32 %r618, %r200, 2139062143;
and.b32 %r619, %r595, -2139062144;
sub.s32 %r620, %r617, %r618;
xor.b32 %r621, %r619, %r620;
xor.b32 %r207, %r621, -2139062144;

dp4a.s32.s32 %r202, %r203, %r106, %r194;


dp4a.s32.s32 %r206, %r207, %r110, %r202;

cvt.u32.u16 %r622, %rs42;
cvt.u32.u16 %r623, %rs43;
mul.lo.s32 %r624, %r182, %r622;
mad.lo.s32 %r625, %r206, %r623, %r624;
add.s32 %r626, %r206, %r182;
shr.u32 %r627, %r626, 31;
add.s32 %r628, %r626, %r627;
shr.s32 %r629, %r628, 1;
add.s32 %r630, %r625, %r629;
shr.s32 %r631, %r630, 31;
shr.u32 %r632, %r631, 30;
add.s32 %r633, %r630, %r632;
shr.s32 %r634, %r633, 2;
add.s64 %rd75, %rd60, %rd5;
ld.global.nc.u16 %rs3, [%rd75];

{ cvt.f32.f16 %f29, %rs3;}



{.reg .f16 low,high;
mov.b32 {low,high},%r112;
cvt.f32.f16 %f30, low;}


mul.ftz.f32 %f37, %f29, %f30;
cvt.rn.f32.s32 %f38, %r634;
fma.rn.ftz.f32 %f106, %f37, %f38, %f106;
add.s64 %rd76, %rd61, %rd5;
ld.global.nc.u16 %rs53, [%rd76+2];
ld.global.nc.u16 %rs54, [%rd76+4];
ld.global.nc.u16 %rs55, [%rd76+6];
add.s64 %rd77, %rd62, %rd5;
ld.global.nc.u8 %rs56, [%rd77];
and.b16 %rs57, %rs56, 240;
and.b16 %rs58, %rs56, 15;
shr.u16 %rs59, %rs57, 4;
ld.global.nc.u16 %rs60, [%rd76];
shl.b16 %rs61, %rs60, 3;
cvt.u64.u16 %rd78, %rs61;
and.b64 %rd79, %rd78, 4088;
add.s64 %rd80, %rd33, %rd79;
ld.global.nc.v2.u32 {%r635, %r636}, [%rd80];
shr.u16 %rs62, %rs60, 9;
cvt.u32.u16 %r639, %rs62;
popc.b32 %r640, %r639;
shl.b32 %r641, %r640, 7;
and.b32 %r642, %r641, 128;
or.b32 %r643, %r642, %r639;
mul.lo.s32 %r644, %r643, 16843009;
and.b32 %r645, %r644, 134480385;
or.b32 %r646, %r645, -2139062144;
add.s32 %r647, %r646, -16843009;
or.b32 %r211, %r647, %r645;

prmt.b32 %r211, %r211, 0, 0xba98;

xor.b32 %r648, %r211, %r635;
or.b32 %r649, %r648, -2139062144;
and.b32 %r650, %r211, 2139062143;
and.b32 %r651, %r635, -2139062144;
sub.s32 %r652, %r649, %r650;
xor.b32 %r653, %r651, %r652;
xor.b32 %r216, %r653, -2139062144;
and.b32 %r654, %r644, -2143281136;
or.b32 %r655, %r654, -2139062144;
add.s32 %r656, %r655, -16843009;
or.b32 %r213, %r656, %r654;

prmt.b32 %r213, %r213, 0, 0xba98;

xor.b32 %r657, %r213, %r636;
or.b32 %r658, %r657, -2139062144;
and.b32 %r659, %r213, 2139062143;
and.b32 %r660, %r636, -2139062144;
sub.s32 %r661, %r658, %r659;
xor.b32 %r662, %r660, %r661;
xor.b32 %r220, %r662, -2139062144;

dp4a.s32.s32 %r215, %r216, %r70, %r242;


dp4a.s32.s32 %r219, %r220, %r74, %r215;

shl.b16 %rs63, %rs53, 3;
cvt.u64.u16 %rd81, %rs63;
and.b64 %rd82, %rd81, 4088;
add.s64 %rd83, %rd33, %rd82;
ld.global.nc.v2.u32 {%r663, %r664}, [%rd83];
shr.u16 %rs64, %rs53, 9;
cvt.u32.u16 %r667, %rs64;
popc.b32 %r668, %r667;
shl.b32 %r669, %r668, 7;
and.b32 %r670, %r669, 128;
or.b32 %r671, %r670, %r667;
mul.lo.s32 %r672, %r671, 16843009;
and.b32 %r673, %r672, 134480385;
or.b32 %r674, %r673, -2139062144;
add.s32 %r675, %r674, -16843009;
or.b32 %r223, %r675, %r673;

prmt.b32 %r223, %r223, 0, 0xba98;

xor.b32 %r676, %r223, %r663;
or.b32 %r677, %r676, -2139062144;
and.b32 %r678, %r223, 2139062143;
and.b32 %r679, %r663, -2139062144;
sub.s32 %r680, %r677, %r678;
xor.b32 %r681, %r679, %r680;
xor.b32 %r228, %r681, -2139062144;
and.b32 %r682, %r672, -2143281136;
or.b32 %r683, %r682, -2139062144;
add.s32 %r684, %r683, -16843009;
or.b32 %r225, %r684, %r682;

prmt.b32 %r225, %r225, 0, 0xba98;

xor.b32 %r685, %r225, %r664;
or.b32 %r686, %r685, -2139062144;
and.b32 %r687, %r225, 2139062143;
and.b32 %r688, %r664, -2139062144;
sub.s32 %r689, %r686, %r687;
xor.b32 %r690, %r688, %r689;
xor.b32 %r232, %r690, -2139062144;

dp4a.s32.s32 %r227, %r228, %r82, %r219;


dp4a.s32.s32 %r231, %r232, %r86, %r227;

shl.b16 %rs65, %rs54, 3;
cvt.u64.u16 %rd84, %rs65;
and.b64 %rd85, %rd84, 4088;
add.s64 %rd86, %rd33, %rd85;
ld.global.nc.v2.u32 {%r691, %r692}, [%rd86];
shr.u16 %rs66, %rs54, 9;
cvt.u32.u16 %r695, %rs66;
popc.b32 %r696, %r695;
shl.b32 %r697, %r696, 7;
and.b32 %r698, %r697, 128;
or.b32 %r699, %r698, %r695;
mul.lo.s32 %r700, %r699, 16843009;
and.b32 %r701, %r700, 134480385;
or.b32 %r702, %r701, -2139062144;
add.s32 %r703, %r702, -16843009;
or.b32 %r235, %r703, %r701;

prmt.b32 %r235, %r235, 0, 0xba98;

xor.b32 %r704, %r235, %r691;
or.b32 %r705, %r704, -2139062144;
and.b32 %r706, %r235, 2139062143;
and.b32 %r707, %r691, -2139062144;
sub.s32 %r708, %r705, %r706;
xor.b32 %r709, %r707, %r708;
xor.b32 %r240, %r709, -2139062144;
and.b32 %r710, %r700, -2143281136;
or.b32 %r711, %r710, -2139062144;
add.s32 %r712, %r711, -16843009;
or.b32 %r237, %r712, %r710;

prmt.b32 %r237, %r237, 0, 0xba98;

xor.b32 %r713, %r237, %r692;
or.b32 %r714, %r713, -2139062144;
and.b32 %r715, %r237, 2139062143;
and.b32 %r716, %r692, -2139062144;
sub.s32 %r717, %r714, %r715;
xor.b32 %r718, %r716, %r717;
xor.b32 %r244, %r718, -2139062144;

dp4a.s32.s32 %r239, %r240, %r94, %r242;


dp4a.s32.s32 %r243, %r244, %r98, %r239;

shl.b16 %rs67, %rs55, 3;
cvt.u64.u16 %rd87, %rs67;
and.b64 %rd88, %rd87, 4088;
add.s64 %rd89, %rd33, %rd88;
ld.global.nc.v2.u32 {%r719, %r720}, [%rd89];
shr.u16 %rs68, %rs55, 9;
cvt.u32.u16 %r723, %rs68;
popc.b32 %r724, %r723;
shl.b32 %r725, %r724, 7;
and.b32 %r726, %r725, 128;
or.b32 %r727, %r726, %r723;
mul.lo.s32 %r728, %r727, 16843009;
and.b32 %r729, %r728, 134480385;
or.b32 %r730, %r729, -2139062144;
add.s32 %r731, %r730, -16843009;
or.b32 %r247, %r731, %r729;

prmt.b32 %r247, %r247, 0, 0xba98;

xor.b32 %r732, %r247, %r719;
or.b32 %r733, %r732, -2139062144;
and.b32 %r734, %r247, 2139062143;
and.b32 %r735, %r719, -2139062144;
sub.s32 %r736, %r733, %r734;
xor.b32 %r737, %r735, %r736;
xor.b32 %r252, %r737, -2139062144;
and.b32 %r738, %r728, -2143281136;
or.b32 %r739, %r738, -2139062144;
add.s32 %r740, %r739, -16843009;
or.b32 %r249, %r740, %r738;

prmt.b32 %r249, %r249, 0, 0xba98;

xor.b32 %r741, %r249, %r720;
or.b32 %r742, %r741, -2139062144;
and.b32 %r743, %r249, 2139062143;
and.b32 %r744, %r720, -2139062144;
sub.s32 %r745, %r742, %r743;
xor.b32 %r746, %r744, %r745;
xor.b32 %r256, %r746, -2139062144;

dp4a.s32.s32 %r251, %r252, %r106, %r243;


dp4a.s32.s32 %r255, %r256, %r110, %r251;

cvt.u32.u16 %r747, %rs58;
cvt.u32.u16 %r748, %rs59;
mul.lo.s32 %r749, %r231, %r747;
mad.lo.s32 %r750, %r255, %r748, %r749;
add.s32 %r751, %r255, %r231;
shr.u32 %r752, %r751, 31;
add.s32 %r753, %r751, %r752;
shr.s32 %r754, %r753, 1;
add.s32 %r755, %r750, %r754;
shr.s32 %r756, %r755, 31;
shr.u32 %r757, %r756, 30;
add.s32 %r758, %r755, %r757;
shr.s32 %r759, %r758, 2;
add.s64 %rd90, %rd75, %rd5;
ld.global.nc.u16 %rs4, [%rd90];

{ cvt.f32.f16 %f31, %rs4;}



{.reg .f16 low,high;
mov.b32 {low,high},%r112;
cvt.f32.f16 %f32, low;}


mul.ftz.f32 %f39, %f31, %f32;
cvt.rn.f32.s32 %f40, %r759;
fma.rn.ftz.f32 %f105, %f39, %f40, %f105;
add.s32 %r843, %r843, 128;
add.s64 %rd93, %rd93, 1184;
add.s32 %r844, %r844, 16;
setp.lt.s32 %p3, %r844, %r5;
@%p3 bra $L__BB214_5;

$L__BB214_6:
setp.eq.s32 %p4, %r1, 0;
@%p4 bra $L__BB214_8;

shl.b32 %r760, %r1, 9;
mov.u32 %r761, _ZZ13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r762, %r761, %r760;
shl.b32 %r763, %r2, 2;
add.s32 %r764, %r762, %r763;
st.shared.f32 [%r764+-512], %f108;
st.shared.f32 [%r764+-384], %f107;
st.shared.f32 [%r764+-256], %f106;
st.shared.f32 [%r764+-128], %f105;

$L__BB214_8:
setp.ne.s32 %p5, %r1, 0;
bar.sync 0;
@%p5 bra $L__BB214_17;

mov.u32 %r765, %ctaid.y;
mad.lo.s32 %r766, %r765, %r31, %r4;
mad.lo.s32 %r767, %r12, %r37, %r766;
shl.b32 %r768, %r2, 2;
mov.u32 %r769, _ZZ13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r19, %r769, %r768;
mov.u32 %r770, 2;
ld.shared.f32 %f41, [%r19];
add.ftz.f32 %f42, %f41, %f108;
ld.shared.f32 %f43, [%r19+512];
add.ftz.f32 %f44, %f43, %f42;
ld.shared.f32 %f45, [%r19+1024];
add.ftz.f32 %f46, %f45, %f44;
mov.b32 %r771, %f46;
mov.u32 %r772, 31;
mov.u32 %r773, 16;
mov.u32 %r774, -1;
shfl.sync.bfly.b32 %r775|%p6, %r771, %r773, %r772, %r774;
mov.b32 %f47, %r775;
add.ftz.f32 %f48, %f46, %f47;
mov.b32 %r776, %f48;
mov.u32 %r777, 8;
shfl.sync.bfly.b32 %r778|%p7, %r776, %r777, %r772, %r774;
mov.b32 %f49, %r778;
add.ftz.f32 %f50, %f48, %f49;
mov.b32 %r779, %f50;
mov.u32 %r780, 4;
shfl.sync.bfly.b32 %r781|%p8, %r779, %r780, %r772, %r774;
mov.b32 %f51, %r781;
add.ftz.f32 %f52, %f50, %f51;
mov.b32 %r782, %f52;
shfl.sync.bfly.b32 %r783|%p9, %r782, %r770, %r772, %r774;
mov.b32 %f53, %r783;
add.ftz.f32 %f54, %f52, %f53;
mov.b32 %r784, %f54;
mov.u32 %r785, 1;
shfl.sync.bfly.b32 %r786|%p10, %r784, %r785, %r772, %r774;
mov.b32 %f55, %r786;
add.ftz.f32 %f13, %f54, %f55;
setp.ge.u32 %p11, %r4, %r25;
setp.ne.s32 %p12, %r2, 0;
cvta.to.global.u64 %rd91, %rd15;
mul.wide.u32 %rd92, %r767, 4;
add.s64 %rd11, %rd91, %rd92;
or.pred %p13, %p12, %p11;
@%p13 bra $L__BB214_11;

st.global.f32 [%rd11], %f13;

$L__BB214_11:
ld.shared.f32 %f56, [%r19+128];
add.ftz.f32 %f57, %f56, %f107;
ld.shared.f32 %f58, [%r19+640];
add.ftz.f32 %f59, %f58, %f57;
ld.shared.f32 %f60, [%r19+1152];
add.ftz.f32 %f61, %f60, %f59;
mov.b32 %r787, %f61;
shfl.sync.bfly.b32 %r791|%p14, %r787, %r773, %r772, %r774;
mov.b32 %f62, %r791;
add.ftz.f32 %f63, %f61, %f62;
mov.b32 %r792, %f63;
shfl.sync.bfly.b32 %r794|%p15, %r792, %r777, %r772, %r774;
mov.b32 %f64, %r794;
add.ftz.f32 %f65, %f63, %f64;
mov.b32 %r795, %f65;
shfl.sync.bfly.b32 %r797|%p16, %r795, %r780, %r772, %r774;
mov.b32 %f66, %r797;
add.ftz.f32 %f67, %f65, %f66;
mov.b32 %r798, %f67;
shfl.sync.bfly.b32 %r800|%p17, %r798, %r770, %r772, %r774;
mov.b32 %f68, %r800;
add.ftz.f32 %f69, %f67, %f68;
mov.b32 %r801, %f69;
shfl.sync.bfly.b32 %r803|%p18, %r801, %r785, %r772, %r774;
mov.b32 %f70, %r803;
add.ftz.f32 %f14, %f69, %f70;
add.s32 %r804, %r4, 1;
setp.ge.u32 %p19, %r804, %r25;
setp.ne.s32 %p20, %r2, 1;
or.pred %p21, %p20, %p19;
@%p21 bra $L__BB214_13;

st.global.f32 [%rd11+4], %f14;

$L__BB214_13:
ld.shared.f32 %f71, [%r19+256];
add.ftz.f32 %f72, %f71, %f106;
ld.shared.f32 %f73, [%r19+768];
add.ftz.f32 %f74, %f73, %f72;
ld.shared.f32 %f75, [%r19+1280];
add.ftz.f32 %f76, %f75, %f74;
mov.b32 %r805, %f76;
mov.u32 %r806, 31;
mov.u32 %r807, 16;
mov.u32 %r808, -1;
shfl.sync.bfly.b32 %r809|%p22, %r805, %r807, %r806, %r808;
mov.b32 %f77, %r809;
add.ftz.f32 %f78, %f76, %f77;
mov.b32 %r810, %f78;
mov.u32 %r811, 8;
shfl.sync.bfly.b32 %r812|%p23, %r810, %r811, %r806, %r808;
mov.b32 %f79, %r812;
add.ftz.f32 %f80, %f78, %f79;
mov.b32 %r813, %f80;
mov.u32 %r814, 4;
shfl.sync.bfly.b32 %r815|%p24, %r813, %r814, %r806, %r808;
mov.b32 %f81, %r815;
add.ftz.f32 %f82, %f80, %f81;
mov.b32 %r816, %f82;
mov.u32 %r817, 2;
shfl.sync.bfly.b32 %r818|%p25, %r816, %r817, %r806, %r808;
mov.b32 %f83, %r818;
add.ftz.f32 %f84, %f82, %f83;
mov.b32 %r819, %f84;
mov.u32 %r820, 1;
shfl.sync.bfly.b32 %r821|%p26, %r819, %r820, %r806, %r808;
mov.b32 %f85, %r821;
add.ftz.f32 %f15, %f84, %f85;
add.s32 %r822, %r4, 2;
setp.ge.u32 %p27, %r822, %r25;
setp.ne.s32 %p28, %r2, 2;
or.pred %p29, %p28, %p27;
@%p29 bra $L__BB214_15;

st.global.f32 [%rd11+8], %f15;

$L__BB214_15:
ld.shared.f32 %f86, [%r19+384];
add.ftz.f32 %f87, %f86, %f105;
ld.shared.f32 %f88, [%r19+896];
add.ftz.f32 %f89, %f88, %f87;
ld.shared.f32 %f90, [%r19+1408];
add.ftz.f32 %f91, %f90, %f89;
mov.b32 %r823, %f91;
shfl.sync.bfly.b32 %r827|%p30, %r823, %r807, %r806, %r808;
mov.b32 %f92, %r827;
add.ftz.f32 %f93, %f91, %f92;
mov.b32 %r828, %f93;
shfl.sync.bfly.b32 %r830|%p31, %r828, %r811, %r806, %r808;
mov.b32 %f94, %r830;
add.ftz.f32 %f95, %f93, %f94;
mov.b32 %r831, %f95;
shfl.sync.bfly.b32 %r833|%p32, %r831, %r814, %r806, %r808;
mov.b32 %f96, %r833;
add.ftz.f32 %f97, %f95, %f96;
mov.b32 %r834, %f97;
shfl.sync.bfly.b32 %r836|%p33, %r834, %r817, %r806, %r808;
mov.b32 %f98, %r836;
add.ftz.f32 %f99, %f97, %f98;
mov.b32 %r837, %f99;
shfl.sync.bfly.b32 %r839|%p34, %r837, %r820, %r806, %r808;
mov.b32 %f100, %r839;
add.ftz.f32 %f16, %f99, %f100;
add.s32 %r840, %r4, 3;
setp.ge.u32 %p35, %r840, %r25;
setp.ne.s32 %p36, %r2, 3;
or.pred %p37, %p36, %p35;
@%p37 bra $L__BB214_17;

st.global.f32 [%rd11+12], %f16;

$L__BB214_17:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<34>;
.reg .b16 %rs<34>;
.reg .f32 %f<121>;
.reg .b32 %r<514>;
.reg .b64 %rd<95>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[384];

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate[384];

ld.param.u64 %rd8, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd9, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd10, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r45, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r46, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r47, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
ld.param.u64 %rd13, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+16];
ld.param.u64 %rd12, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+8];
ld.param.u64 %rd11, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3];
mov.u32 %r1, %tid.y;
shl.b32 %r48, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r48, %r2;
mov.u32 %r4, %ctaid.x;
mov.u32 %r5, %ctaid.y;
setp.eq.s64 %p2, %rd10, 0;
@%p2 bra $L__BB215_2;

cvta.to.global.u64 %rd17, %rd10;
mul.wide.u32 %rd18, %r5, 4;
add.s64 %rd19, %rd17, %rd18;
ld.global.nc.u32 %r510, [%rd19];
mul.hi.u32 %r49, %r5, %r32;
add.s32 %r50, %r49, %r5;
shr.u32 %r51, %r50, %r33;
mul.lo.s32 %r52, %r51, %r34;
sub.s32 %r511, %r5, %r52;
bra.uni $L__BB215_3;

$L__BB215_2:
mul.hi.u32 %r53, %r5, %r36;
add.s32 %r54, %r53, %r5;
shr.u32 %r510, %r54, %r37;
mov.u32 %r511, %r5;

$L__BB215_3:
or.b32 %r55, %r2, %r1;
setp.ne.s32 %p3, %r55, 0;
mov.f32 %f114, 0f00000000;
mov.f32 %f110, %f114;
mov.f32 %f111, %f114;
@%p3 bra $L__BB215_8;

setp.eq.s64 %p4, %rd11, 0;
mov.f32 %f110, 0f00000000;
mov.f32 %f111, %f110;
@%p4 bra $L__BB215_6;

cvta.to.global.u64 %rd20, %rd11;
mov.u32 %r56, %ctaid.z;
mul.lo.s32 %r57, %r56, %r47;
cvt.u64.u32 %rd21, %r57;
selp.b32 %r58, %r5, %r510, %p2;
mul.lo.s32 %r59, %r58, %r41;
cvt.u64.u32 %rd22, %r59;
cvt.u64.u32 %rd23, %r2;
cvt.s64.s32 %rd24, %r4;
add.s64 %rd25, %rd23, %rd24;
add.s64 %rd26, %rd25, %rd21;
add.s64 %rd27, %rd26, %rd22;
shl.b64 %rd28, %rd27, 2;
add.s64 %rd29, %rd20, %rd28;
ld.global.f32 %f111, [%rd29];

$L__BB215_6:
setp.eq.s64 %p6, %rd13, 0;
setp.eq.s64 %p7, %rd12, 0;
or.pred %p8, %p6, %p7;
@%p8 bra $L__BB215_8;

cvta.to.global.u64 %rd30, %rd13;
mov.u32 %r61, %ctaid.z;
mul.lo.s32 %r62, %r61, %r47;
cvt.u64.u32 %rd31, %r62;
selp.b32 %r63, %r5, %r510, %p2;
mul.lo.s32 %r64, %r63, %r41;
cvt.u64.u32 %rd32, %r64;
cvt.u64.u32 %rd33, %r2;
cvt.s64.s32 %rd34, %r4;
add.s64 %rd35, %rd33, %rd34;
add.s64 %rd36, %rd35, %rd31;
add.s64 %rd37, %rd36, %rd32;
shl.b64 %rd38, %rd37, 2;
add.s64 %rd39, %rd30, %rd38;
ld.global.f32 %f110, [%rd39];

$L__BB215_8:
shr.u32 %r66, %r31, 8;
shr.s32 %r67, %r3, 31;
shr.u32 %r68, %r67, 29;
add.s32 %r69, %r3, %r68;
shr.s32 %r513, %r69, 3;
setp.ge.s32 %p10, %r513, %r66;
mov.f32 %f116, %f114;
@%p10 bra $L__BB215_13;

mov.u32 %r70, %ctaid.z;
and.b32 %r74, %r69, -8;
sub.s32 %r12, %r3, %r74;
mul.hi.u32 %r75, %r70, %r42;
add.s32 %r76, %r75, %r70;
shr.u32 %r77, %r76, %r43;
mad.lo.s32 %r78, %r77, %r45, %r513;
mad.lo.s32 %r79, %r510, %r39, %r78;
mad.lo.s32 %r80, %r4, %r35, %r79;
mul.wide.s32 %rd94, %r80, 74;
shl.b32 %r512, %r513, 3;
mul.lo.s32 %r81, %r511, %r40;
cvt.u64.u32 %rd40, %r81;
mul.lo.s32 %r82, %r70, %r46;
cvt.u64.u32 %rd41, %r82;
add.s64 %rd5, %rd40, %rd41;
mov.f32 %f116, 0f00000000;
cvta.to.global.u64 %rd42, %rd9;
cvta.to.global.u64 %rd46, %rd8;
cvt.s64.s32 %rd49, %r12;
add.s64 %rd50, %rd46, %rd49;
mov.f32 %f114, %f116;

$L__BB215_10:
cvt.u64.u32 %rd43, %r512;
add.s64 %rd44, %rd5, %rd43;
shl.b32 %r132, %r12, 2;
mul.wide.s32 %rd45, %r132, 2;
add.s64 %rd47, %rd46, %rd45;
add.s64 %rd48, %rd47, %rd94;
ld.global.nc.u16 %rs2, [%rd48+4];
ld.global.nc.u16 %rs3, [%rd48+6];
ld.global.nc.u16 %rs4, [%rd48+8];
add.s64 %rd51, %rd50, %rd94;
ld.global.nc.u8 %rs5, [%rd51+66];
and.b16 %rs6, %rs5, 240;
and.b16 %rs7, %rs5, 15;
shr.u16 %rs8, %rs6, 4;
add.s64 %rd52, %rd44, %rd49;
mul.lo.s64 %rd53, %rd52, 36;
ld.global.nc.u16 %rs9, [%rd48+2];
shl.b16 %rs10, %rs9, 3;
cvt.u64.u16 %rd54, %rs10;
and.b64 %rd55, %rd54, 4088;
mov.u64 %rd56, iq2xs_grid;
add.s64 %rd57, %rd56, %rd55;
ld.global.nc.v2.u32 {%r133, %r134}, [%rd57];
mov.u32 %r114, 0;
shr.u16 %rs11, %rs9, 9;
cvt.u32.u16 %r137, %rs11;
popc.b32 %r138, %r137;
shl.b32 %r139, %r138, 7;
and.b32 %r140, %r139, 128;
or.b32 %r141, %r140, %r137;
mul.lo.s32 %r142, %r141, 16843009;
and.b32 %r143, %r142, 134480385;
or.b32 %r144, %r143, -2139062144;
add.s32 %r145, %r144, -16843009;
or.b32 %r83, %r145, %r143;

prmt.b32 %r83, %r83, 0, 0xba98;

xor.b32 %r146, %r83, %r133;
or.b32 %r147, %r146, -2139062144;
and.b32 %r148, %r83, 2139062143;
and.b32 %r149, %r133, -2139062144;
sub.s32 %r150, %r147, %r148;
xor.b32 %r151, %r149, %r150;
xor.b32 %r88, %r151, -2139062144;
add.s64 %rd58, %rd42, %rd53;
ld.global.nc.u32 %r16, [%rd58+4];
and.b32 %r152, %r142, -2143281136;
or.b32 %r153, %r152, -2139062144;
add.s32 %r154, %r153, -16843009;
or.b32 %r85, %r154, %r152;

prmt.b32 %r85, %r85, 0, 0xba98;

xor.b32 %r155, %r85, %r134;
or.b32 %r156, %r155, -2139062144;
and.b32 %r157, %r85, 2139062143;
and.b32 %r158, %r134, -2139062144;
sub.s32 %r159, %r156, %r157;
xor.b32 %r160, %r158, %r159;
xor.b32 %r92, %r160, -2139062144;
ld.global.nc.u32 %r17, [%rd58+8];

dp4a.s32.s32 %r87, %r88, %r16, %r114;


dp4a.s32.s32 %r91, %r92, %r17, %r87;

shl.b16 %rs12, %rs2, 3;
cvt.u64.u16 %rd59, %rs12;
and.b64 %rd60, %rd59, 4088;
add.s64 %rd61, %rd56, %rd60;
ld.global.nc.v2.u32 {%r161, %r162}, [%rd61];
shr.u16 %rs13, %rs2, 9;
cvt.u32.u16 %r165, %rs13;
popc.b32 %r166, %r165;
shl.b32 %r167, %r166, 7;
and.b32 %r168, %r167, 128;
or.b32 %r169, %r168, %r165;
mul.lo.s32 %r170, %r169, 16843009;
and.b32 %r171, %r170, 134480385;
or.b32 %r172, %r171, -2139062144;
add.s32 %r173, %r172, -16843009;
or.b32 %r95, %r173, %r171;

prmt.b32 %r95, %r95, 0, 0xba98;

xor.b32 %r174, %r95, %r161;
or.b32 %r175, %r174, -2139062144;
and.b32 %r176, %r95, 2139062143;
and.b32 %r177, %r161, -2139062144;
sub.s32 %r178, %r175, %r176;
xor.b32 %r179, %r177, %r178;
xor.b32 %r100, %r179, -2139062144;
ld.global.nc.u32 %r18, [%rd58+12];
and.b32 %r180, %r170, -2143281136;
or.b32 %r181, %r180, -2139062144;
add.s32 %r182, %r181, -16843009;
or.b32 %r97, %r182, %r180;

prmt.b32 %r97, %r97, 0, 0xba98;

xor.b32 %r183, %r97, %r162;
or.b32 %r184, %r183, -2139062144;
and.b32 %r185, %r97, 2139062143;
and.b32 %r186, %r162, -2139062144;
sub.s32 %r187, %r184, %r185;
xor.b32 %r188, %r186, %r187;
xor.b32 %r104, %r188, -2139062144;
ld.global.nc.u32 %r19, [%rd58+16];

dp4a.s32.s32 %r99, %r100, %r18, %r91;


dp4a.s32.s32 %r103, %r104, %r19, %r99;

shl.b16 %rs14, %rs3, 3;
cvt.u64.u16 %rd62, %rs14;
and.b64 %rd63, %rd62, 4088;
add.s64 %rd64, %rd56, %rd63;
ld.global.nc.v2.u32 {%r189, %r190}, [%rd64];
shr.u16 %rs15, %rs3, 9;
cvt.u32.u16 %r193, %rs15;
popc.b32 %r194, %r193;
shl.b32 %r195, %r194, 7;
and.b32 %r196, %r195, 128;
or.b32 %r197, %r196, %r193;
mul.lo.s32 %r198, %r197, 16843009;
and.b32 %r199, %r198, 134480385;
or.b32 %r200, %r199, -2139062144;
add.s32 %r201, %r200, -16843009;
or.b32 %r107, %r201, %r199;

prmt.b32 %r107, %r107, 0, 0xba98;

xor.b32 %r202, %r107, %r189;
or.b32 %r203, %r202, -2139062144;
and.b32 %r204, %r107, 2139062143;
and.b32 %r205, %r189, -2139062144;
sub.s32 %r206, %r203, %r204;
xor.b32 %r207, %r205, %r206;
xor.b32 %r112, %r207, -2139062144;
ld.global.nc.u32 %r20, [%rd58+20];
and.b32 %r208, %r198, -2143281136;
or.b32 %r209, %r208, -2139062144;
add.s32 %r210, %r209, -16843009;
or.b32 %r109, %r210, %r208;

prmt.b32 %r109, %r109, 0, 0xba98;

xor.b32 %r211, %r109, %r190;
or.b32 %r212, %r211, -2139062144;
and.b32 %r213, %r109, 2139062143;
and.b32 %r214, %r190, -2139062144;
sub.s32 %r215, %r212, %r213;
xor.b32 %r216, %r214, %r215;
xor.b32 %r116, %r216, -2139062144;
ld.global.nc.u32 %r21, [%rd58+24];

dp4a.s32.s32 %r111, %r112, %r20, %r114;


dp4a.s32.s32 %r115, %r116, %r21, %r111;

shl.b16 %rs16, %rs4, 3;
cvt.u64.u16 %rd65, %rs16;
and.b64 %rd66, %rd65, 4088;
add.s64 %rd67, %rd56, %rd66;
ld.global.nc.v2.u32 {%r217, %r218}, [%rd67];
shr.u16 %rs17, %rs4, 9;
cvt.u32.u16 %r221, %rs17;
popc.b32 %r222, %r221;
shl.b32 %r223, %r222, 7;
and.b32 %r224, %r223, 128;
or.b32 %r225, %r224, %r221;
mul.lo.s32 %r226, %r225, 16843009;
and.b32 %r227, %r226, 134480385;
or.b32 %r228, %r227, -2139062144;
add.s32 %r229, %r228, -16843009;
or.b32 %r119, %r229, %r227;

prmt.b32 %r119, %r119, 0, 0xba98;

xor.b32 %r230, %r119, %r217;
or.b32 %r231, %r230, -2139062144;
and.b32 %r232, %r119, 2139062143;
and.b32 %r233, %r217, -2139062144;
sub.s32 %r234, %r231, %r232;
xor.b32 %r235, %r233, %r234;
xor.b32 %r124, %r235, -2139062144;
ld.global.nc.u32 %r22, [%rd58+28];
and.b32 %r236, %r226, -2143281136;
or.b32 %r237, %r236, -2139062144;
add.s32 %r238, %r237, -16843009;
or.b32 %r121, %r238, %r236;

prmt.b32 %r121, %r121, 0, 0xba98;

xor.b32 %r239, %r121, %r218;
or.b32 %r240, %r239, -2139062144;
and.b32 %r241, %r121, 2139062143;
and.b32 %r242, %r218, -2139062144;
sub.s32 %r243, %r240, %r241;
xor.b32 %r244, %r242, %r243;
xor.b32 %r128, %r244, -2139062144;
ld.global.nc.u32 %r23, [%rd58+32];

dp4a.s32.s32 %r123, %r124, %r22, %r115;


dp4a.s32.s32 %r127, %r128, %r23, %r123;

cvt.u32.u16 %r245, %rs7;
cvt.u32.u16 %r246, %rs8;
mul.lo.s32 %r247, %r103, %r245;
mad.lo.s32 %r248, %r127, %r246, %r247;
add.s32 %r249, %r127, %r103;
shr.u32 %r250, %r249, 31;
add.s32 %r251, %r249, %r250;
shr.s32 %r252, %r251, 1;
add.s32 %r253, %r248, %r252;
shr.s32 %r254, %r253, 31;
shr.u32 %r255, %r254, 30;
add.s32 %r256, %r253, %r255;
shr.s32 %r257, %r256, 2;
add.s64 %rd68, %rd46, %rd94;
ld.global.nc.u16 %rs1, [%rd68];

{ cvt.f32.f16 %f40, %rs1;}


ld.global.nc.u32 %r24, [%rd58];

{.reg .f16 low,high;
mov.b32 {low,high},%r24;
cvt.f32.f16 %f41, low;}


mul.ftz.f32 %f42, %f40, %f41;
cvt.rn.f32.s32 %f43, %r257;
fma.rn.ftz.f32 %f116, %f42, %f43, %f116;
setp.eq.s64 %p11, %rd12, 0;
@%p11 bra $L__BB215_12;

cvta.to.global.u64 %rd70, %rd12;
add.s64 %rd71, %rd70, %rd45;
add.s64 %rd72, %rd71, %rd94;
ld.global.u16 %rs19, [%rd72+4];
ld.global.u16 %rs20, [%rd72+6];
ld.global.u16 %rs21, [%rd72+8];
add.s64 %rd74, %rd70, %rd49;
add.s64 %rd75, %rd74, %rd94;
ld.global.u8 %rs22, [%rd75+66];
and.b16 %rs23, %rs22, 15;
shr.u16 %rs24, %rs22, 4;
ld.global.u16 %rs25, [%rd72+2];
shl.b16 %rs26, %rs25, 3;
cvt.u64.u16 %rd76, %rs26;
and.b64 %rd77, %rd76, 4088;
add.s64 %rd79, %rd56, %rd77;
ld.global.nc.v2.u32 {%r308, %r309}, [%rd79];
shr.u16 %rs27, %rs25, 9;
cvt.u32.u16 %r312, %rs27;
popc.b32 %r313, %r312;
shl.b32 %r314, %r313, 7;
and.b32 %r315, %r314, 128;
or.b32 %r316, %r315, %r312;
mul.lo.s32 %r317, %r316, 16843009;
and.b32 %r318, %r317, 134480385;
or.b32 %r319, %r318, -2139062144;
add.s32 %r320, %r319, -16843009;
or.b32 %r258, %r320, %r318;

prmt.b32 %r258, %r258, 0, 0xba98;

xor.b32 %r321, %r258, %r308;
or.b32 %r322, %r321, -2139062144;
and.b32 %r323, %r258, 2139062143;
and.b32 %r324, %r308, -2139062144;
sub.s32 %r325, %r322, %r323;
xor.b32 %r326, %r324, %r325;
xor.b32 %r263, %r326, -2139062144;
and.b32 %r327, %r317, -2143281136;
or.b32 %r328, %r327, -2139062144;
add.s32 %r329, %r328, -16843009;
or.b32 %r260, %r329, %r327;

prmt.b32 %r260, %r260, 0, 0xba98;

xor.b32 %r330, %r260, %r309;
or.b32 %r331, %r330, -2139062144;
and.b32 %r332, %r260, 2139062143;
and.b32 %r333, %r309, -2139062144;
sub.s32 %r334, %r331, %r332;
xor.b32 %r335, %r333, %r334;
xor.b32 %r267, %r335, -2139062144;

dp4a.s32.s32 %r262, %r263, %r16, %r114;


dp4a.s32.s32 %r266, %r267, %r17, %r262;

shl.b16 %rs28, %rs19, 3;
cvt.u64.u16 %rd80, %rs28;
and.b64 %rd81, %rd80, 4088;
add.s64 %rd82, %rd56, %rd81;
ld.global.nc.v2.u32 {%r336, %r337}, [%rd82];
shr.u16 %rs29, %rs19, 9;
cvt.u32.u16 %r340, %rs29;
popc.b32 %r341, %r340;
shl.b32 %r342, %r341, 7;
and.b32 %r343, %r342, 128;
or.b32 %r344, %r343, %r340;
mul.lo.s32 %r345, %r344, 16843009;
and.b32 %r346, %r345, 134480385;
or.b32 %r347, %r346, -2139062144;
add.s32 %r348, %r347, -16843009;
or.b32 %r270, %r348, %r346;

prmt.b32 %r270, %r270, 0, 0xba98;

xor.b32 %r349, %r270, %r336;
or.b32 %r350, %r349, -2139062144;
and.b32 %r351, %r270, 2139062143;
and.b32 %r352, %r336, -2139062144;
sub.s32 %r353, %r350, %r351;
xor.b32 %r354, %r352, %r353;
xor.b32 %r275, %r354, -2139062144;
and.b32 %r355, %r345, -2143281136;
or.b32 %r356, %r355, -2139062144;
add.s32 %r357, %r356, -16843009;
or.b32 %r272, %r357, %r355;

prmt.b32 %r272, %r272, 0, 0xba98;

xor.b32 %r358, %r272, %r337;
or.b32 %r359, %r358, -2139062144;
and.b32 %r360, %r272, 2139062143;
and.b32 %r361, %r337, -2139062144;
sub.s32 %r362, %r359, %r360;
xor.b32 %r363, %r361, %r362;
xor.b32 %r279, %r363, -2139062144;

dp4a.s32.s32 %r274, %r275, %r18, %r266;


dp4a.s32.s32 %r278, %r279, %r19, %r274;

shl.b16 %rs30, %rs20, 3;
cvt.u64.u16 %rd83, %rs30;
and.b64 %rd84, %rd83, 4088;
add.s64 %rd85, %rd56, %rd84;
ld.global.nc.v2.u32 {%r364, %r365}, [%rd85];
shr.u16 %rs31, %rs20, 9;
cvt.u32.u16 %r368, %rs31;
popc.b32 %r369, %r368;
shl.b32 %r370, %r369, 7;
and.b32 %r371, %r370, 128;
or.b32 %r372, %r371, %r368;
mul.lo.s32 %r373, %r372, 16843009;
and.b32 %r374, %r373, 134480385;
or.b32 %r375, %r374, -2139062144;
add.s32 %r376, %r375, -16843009;
or.b32 %r282, %r376, %r374;

prmt.b32 %r282, %r282, 0, 0xba98;

xor.b32 %r377, %r282, %r364;
or.b32 %r378, %r377, -2139062144;
and.b32 %r379, %r282, 2139062143;
and.b32 %r380, %r364, -2139062144;
sub.s32 %r381, %r378, %r379;
xor.b32 %r382, %r380, %r381;
xor.b32 %r287, %r382, -2139062144;
and.b32 %r383, %r373, -2143281136;
or.b32 %r384, %r383, -2139062144;
add.s32 %r385, %r384, -16843009;
or.b32 %r284, %r385, %r383;

prmt.b32 %r284, %r284, 0, 0xba98;

xor.b32 %r386, %r284, %r365;
or.b32 %r387, %r386, -2139062144;
and.b32 %r388, %r284, 2139062143;
and.b32 %r389, %r365, -2139062144;
sub.s32 %r390, %r387, %r388;
xor.b32 %r391, %r389, %r390;
xor.b32 %r291, %r391, -2139062144;

dp4a.s32.s32 %r286, %r287, %r20, %r114;


dp4a.s32.s32 %r290, %r291, %r21, %r286;

shl.b16 %rs32, %rs21, 3;
cvt.u64.u16 %rd86, %rs32;
and.b64 %rd87, %rd86, 4088;
add.s64 %rd88, %rd56, %rd87;
ld.global.nc.v2.u32 {%r392, %r393}, [%rd88];
shr.u16 %rs33, %rs21, 9;
cvt.u32.u16 %r396, %rs33;
popc.b32 %r397, %r396;
shl.b32 %r398, %r397, 7;
and.b32 %r399, %r398, 128;
or.b32 %r400, %r399, %r396;
mul.lo.s32 %r401, %r400, 16843009;
and.b32 %r402, %r401, 134480385;
or.b32 %r403, %r402, -2139062144;
add.s32 %r404, %r403, -16843009;
or.b32 %r294, %r404, %r402;

prmt.b32 %r294, %r294, 0, 0xba98;

xor.b32 %r405, %r294, %r392;
or.b32 %r406, %r405, -2139062144;
and.b32 %r407, %r294, 2139062143;
and.b32 %r408, %r392, -2139062144;
sub.s32 %r409, %r406, %r407;
xor.b32 %r410, %r408, %r409;
xor.b32 %r299, %r410, -2139062144;
and.b32 %r411, %r401, -2143281136;
or.b32 %r412, %r411, -2139062144;
add.s32 %r413, %r412, -16843009;
or.b32 %r296, %r413, %r411;

prmt.b32 %r296, %r296, 0, 0xba98;

xor.b32 %r414, %r296, %r393;
or.b32 %r415, %r414, -2139062144;
and.b32 %r416, %r296, 2139062143;
and.b32 %r417, %r393, -2139062144;
sub.s32 %r418, %r415, %r416;
xor.b32 %r419, %r417, %r418;
xor.b32 %r303, %r419, -2139062144;

dp4a.s32.s32 %r298, %r299, %r22, %r290;


dp4a.s32.s32 %r302, %r303, %r23, %r298;

cvt.u32.u16 %r420, %rs23;
cvt.u32.u16 %r421, %rs24;
mul.lo.s32 %r422, %r278, %r420;
mad.lo.s32 %r423, %r302, %r421, %r422;
add.s32 %r424, %r302, %r278;
shr.u32 %r425, %r424, 31;
add.s32 %r426, %r424, %r425;
shr.s32 %r427, %r426, 1;
add.s32 %r428, %r423, %r427;
shr.s32 %r429, %r428, 31;
shr.u32 %r430, %r429, 30;
add.s32 %r431, %r428, %r430;
shr.s32 %r432, %r431, 2;
add.s64 %rd89, %rd70, %rd94;
ld.global.u16 %rs18, [%rd89];

{ cvt.f32.f16 %f44, %rs18;}



{.reg .f16 low,high;
mov.b32 {low,high},%r24;
cvt.f32.f16 %f45, low;}


mul.ftz.f32 %f46, %f44, %f45;
cvt.rn.f32.s32 %f47, %r432;
fma.rn.ftz.f32 %f114, %f46, %f47, %f114;

$L__BB215_12:
add.s64 %rd94, %rd94, 1184;
add.s32 %r512, %r512, 128;
add.s32 %r513, %r513, 16;
setp.lt.s32 %p12, %r513, %r66;
@%p12 bra $L__BB215_10;

$L__BB215_13:
mov.u32 %r496, %tid.y;
setp.eq.s32 %p13, %r496, 0;
@%p13 bra $L__BB215_16;

mov.u32 %r498, %tid.y;
mov.u32 %r497, %tid.x;
setp.eq.s64 %p14, %rd12, 0;
add.s32 %r27, %r498, -1;
shl.b32 %r437, %r498, 7;
mov.u32 %r438, _ZZ13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r439, %r438, %r437;
shl.b32 %r440, %r497, 2;
add.s32 %r441, %r439, %r440;
st.shared.f32 [%r441+-128], %f116;
@%p14 bra $L__BB215_16;

shl.b32 %r443, %r27, 7;
mov.u32 %r444, _ZZ13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r445, %r444, %r443;
add.s32 %r447, %r445, %r440;
st.shared.f32 [%r447], %f114;

$L__BB215_16:
mov.u32 %r499, %tid.y;
setp.ne.s32 %p15, %r499, 0;
bar.sync 0;
@%p15 bra $L__BB215_35;

mov.u32 %r500, %tid.x;
setp.ne.s64 %p16, %rd12, 0;
shl.b32 %r450, %r500, 2;
mov.u32 %r451, _ZZ13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r28, %r451, %r450;
ld.shared.f32 %f48, [%r28];
add.ftz.f32 %f13, %f116, %f48;
@%p16 bra $L__BB215_19;
bra.uni $L__BB215_18;

$L__BB215_19:
mov.u32 %r454, _ZZ13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r455, %r454, %r450;
ld.shared.f32 %f52, [%r455];
add.ftz.f32 %f53, %f114, %f52;
ld.shared.f32 %f54, [%r28+128];
add.ftz.f32 %f55, %f13, %f54;
ld.shared.f32 %f56, [%r455+128];
add.ftz.f32 %f57, %f53, %f56;
ld.shared.f32 %f58, [%r28+256];
add.ftz.f32 %f118, %f55, %f58;
ld.shared.f32 %f59, [%r455+256];
add.ftz.f32 %f114, %f57, %f59;
bra.uni $L__BB215_20;

$L__BB215_18:
ld.shared.f32 %f49, [%r28+128];
add.ftz.f32 %f50, %f13, %f49;
ld.shared.f32 %f51, [%r28+256];
add.ftz.f32 %f118, %f50, %f51;

$L__BB215_20:
mov.b32 %r456, %f118;
mov.u32 %r457, 31;
mov.u32 %r458, 16;
mov.u32 %r459, -1;
shfl.sync.bfly.b32 %r460|%p17, %r456, %r458, %r457, %r459;
mov.b32 %f60, %r460;
add.ftz.f32 %f61, %f118, %f60;
mov.b32 %r461, %f61;
mov.u32 %r462, 8;
shfl.sync.bfly.b32 %r463|%p18, %r461, %r462, %r457, %r459;
mov.b32 %f62, %r463;
add.ftz.f32 %f63, %f61, %f62;
mov.b32 %r464, %f63;
mov.u32 %r465, 4;
shfl.sync.bfly.b32 %r466|%p19, %r464, %r465, %r457, %r459;
mov.b32 %f64, %r466;
add.ftz.f32 %f65, %f63, %f64;
mov.b32 %r467, %f65;
mov.u32 %r468, 2;
shfl.sync.bfly.b32 %r469|%p20, %r467, %r468, %r457, %r459;
mov.b32 %f66, %r469;
add.ftz.f32 %f19, %f65, %f66;
mov.b32 %r470, %f19;
mov.u32 %r471, 1;
shfl.sync.bfly.b32 %r29|%p1, %r470, %r471, %r457, %r459;
setp.eq.s64 %p21, %rd12, 0;
@%p21 bra $L__BB215_22;

mov.b32 %r472, %f114;
shfl.sync.bfly.b32 %r476|%p22, %r472, %r458, %r457, %r459;
mov.b32 %f67, %r476;
add.ftz.f32 %f68, %f114, %f67;
mov.b32 %r477, %f68;
shfl.sync.bfly.b32 %r479|%p23, %r477, %r462, %r457, %r459;
mov.b32 %f69, %r479;
add.ftz.f32 %f70, %f68, %f69;
mov.b32 %r480, %f70;
shfl.sync.bfly.b32 %r482|%p24, %r480, %r465, %r457, %r459;
mov.b32 %f71, %r482;
add.ftz.f32 %f72, %f70, %f71;
mov.b32 %r483, %f72;
shfl.sync.bfly.b32 %r485|%p25, %r483, %r468, %r457, %r459;
mov.b32 %f73, %r485;
add.ftz.f32 %f74, %f72, %f73;
mov.b32 %r486, %f74;
shfl.sync.bfly.b32 %r488|%p26, %r486, %r471, %r457, %r459;
mov.b32 %f75, %r488;
add.ftz.f32 %f114, %f74, %f75;

$L__BB215_22:
mov.u32 %r501, %tid.x;
setp.ne.s32 %p27, %r501, 0;
mov.b32 %f76, %r29;
add.ftz.f32 %f22, %f19, %f76;
@%p27 bra $L__BB215_35;

add.ftz.f32 %f120, %f111, %f22;
@%p21 bra $L__BB215_34;

ld.param.u32 %r505, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
add.ftz.f32 %f24, %f110, %f114;
setp.gt.s32 %p29, %r505, 2;
@%p29 bra $L__BB215_28;

ld.param.u32 %r508, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p32, %r508, 1;
@%p32 bra $L__BB215_32;

ld.param.u32 %r509, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p33, %r509, 2;
@%p33 bra $L__BB215_27;
bra.uni $L__BB215_33;

$L__BB215_27:
mul.ftz.f32 %f96, %f24, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f97, %f96;
add.ftz.f32 %f98, %f97, 0f3F800000;
div.approx.ftz.f32 %f99, %f24, %f98;
mul.ftz.f32 %f120, %f120, %f99;
bra.uni $L__BB215_34;

$L__BB215_28:
ld.param.u32 %r506, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p30, %r506, 3;
@%p30 bra $L__BB215_31;

ld.param.u32 %r507, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.ne.s32 %p31, %r507, 6;
@%p31 bra $L__BB215_33;

ld.param.f32 %f108, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
min.ftz.f32 %f77, %f24, %f108;
neg.ftz.f32 %f78, %f108;
min.ftz.f32 %f79, %f120, %f108;
max.ftz.f32 %f80, %f79, %f78;
mul.ftz.f32 %f81, %f77, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f82, %f81;
add.ftz.f32 %f83, %f82, 0f3F800000;
div.approx.ftz.f32 %f84, %f77, %f83;
mul.ftz.f32 %f120, %f80, %f84;
bra.uni $L__BB215_34;

$L__BB215_32:
mul.ftz.f32 %f100, %f24, 0f3F000000;
mul.ftz.f32 %f101, %f24, 0f3D372713;
fma.rn.ftz.f32 %f102, %f24, %f101, 0f3F800000;
mul.ftz.f32 %f103, %f24, 0f3F4C422A;
mul.ftz.f32 %f104, %f103, %f102;
tanh.approx.f32 %f105, %f104;
add.ftz.f32 %f106, %f105, 0f3F800000;
mul.ftz.f32 %f107, %f100, %f106;
mul.ftz.f32 %f120, %f120, %f107;
bra.uni $L__BB215_34;

$L__BB215_31:
mov.f32 %f85, 0f40E00000;
min.ftz.f32 %f86, %f24, %f85;
min.ftz.f32 %f87, %f120, %f85;
mov.f32 %f88, 0fC0E00000;
max.ftz.f32 %f89, %f87, %f88;
mul.ftz.f32 %f90, %f86, 0f3FD9DB23;
mul.ftz.f32 %f91, %f90, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f92, %f91;
add.ftz.f32 %f93, %f92, 0f3F800000;
div.approx.ftz.f32 %f94, %f86, %f93;
add.ftz.f32 %f95, %f89, 0f3F800000;
mul.ftz.f32 %f120, %f95, %f94;
bra.uni $L__BB215_34;

$L__BB215_33:
mul.ftz.f32 %f120, %f120, %f24;

$L__BB215_34:
ld.param.u64 %rd93, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
mov.u32 %r504, %ctaid.x;
ld.param.u32 %r503, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r502, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
cvta.to.global.u64 %rd90, %rd93;
mov.u32 %r491, %ctaid.z;
mov.u32 %r492, %ctaid.y;
mad.lo.s32 %r494, %r492, %r503, %r504;
mad.lo.s32 %r495, %r491, %r502, %r494;
mul.wide.u32 %rd91, %r495, 4;
add.s64 %rd92, %rd90, %rd91;
st.global.f32 [%rd92], %f120;

$L__BB215_35:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<12>;
.reg .b16 %rs<18>;
.reg .f32 %f<28>;
.reg .b32 %r<271>;
.reg .b64 %rd<53>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[384];

ld.param.u64 %rd13, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd14, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r19, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r18, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r17, [_Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
mov.u32 %r1, %tid.y;
shl.b32 %r34, %r1, 5;
mov.u32 %r35, %tid.x;
add.s32 %r2, %r34, %r35;
shr.u32 %r3, %r33, 8;
setp.eq.s64 %p2, %rd15, 0;
@%p2 bra $L__BB216_2;

cvta.to.global.u64 %rd17, %rd15;
mov.u32 %r36, %ctaid.y;
mul.wide.u32 %rd18, %r36, 4;
add.s64 %rd19, %rd17, %rd18;
ld.global.nc.u32 %r267, [%rd19];
mul.hi.u32 %r37, %r36, %r17;
add.s32 %r38, %r37, %r36;
shr.u32 %r39, %r38, %r18;
mul.lo.s32 %r40, %r39, %r19;
sub.s32 %r268, %r36, %r40;
bra.uni $L__BB216_3;

$L__BB216_2:
mov.u32 %r268, %ctaid.y;
mul.hi.u32 %r41, %r268, %r21;
add.s32 %r42, %r41, %r268;
shr.u32 %r267, %r42, %r22;

$L__BB216_3:
shr.s32 %r43, %r2, 31;
shr.u32 %r44, %r43, 29;
add.s32 %r45, %r2, %r44;
shr.s32 %r270, %r45, 3;
setp.ge.s32 %p3, %r270, %r3;
mov.f32 %f27, 0f00000000;
@%p3 bra $L__BB216_6;

mov.u32 %r46, %ctaid.z;
and.b32 %r50, %r45, -8;
sub.s32 %r51, %r2, %r50;
shl.b32 %r52, %r51, 2;
cvt.s64.s32 %rd1, %r51;
shl.b32 %r269, %r270, 3;
mul.hi.u32 %r53, %r46, %r27;
add.s32 %r54, %r53, %r46;
shr.u32 %r55, %r54, %r28;
mad.lo.s32 %r56, %r55, %r30, %r270;
mad.lo.s32 %r57, %r267, %r24, %r56;
mov.u32 %r58, %ctaid.x;
mad.lo.s32 %r59, %r58, %r20, %r57;
mul.wide.s32 %rd20, %r59, 74;
mul.wide.s32 %rd21, %r52, 2;
add.s64 %rd22, %rd20, %rd21;
cvta.to.global.u64 %rd23, %rd13;
add.s64 %rd24, %rd23, %rd22;
add.s64 %rd52, %rd24, 4;
add.s64 %rd25, %rd20, %rd1;
add.s64 %rd26, %rd23, %rd25;
add.s64 %rd51, %rd26, 66;
add.s64 %rd50, %rd23, %rd20;
cvta.to.global.u64 %rd5, %rd14;
mul.lo.s32 %r60, %r268, %r25;
cvt.u64.u32 %rd27, %r60;
mul.lo.s32 %r61, %r46, %r31;
cvt.u64.u32 %rd28, %r61;
add.s64 %rd6, %rd27, %rd28;
mov.f32 %f27, 0f00000000;

$L__BB216_5:
cvt.u64.u32 %rd29, %r269;
add.s64 %rd30, %rd6, %rd29;
ld.global.nc.u16 %rs2, [%rd52];
ld.global.nc.u16 %rs3, [%rd52+2];
ld.global.nc.u16 %rs4, [%rd52+4];
ld.global.nc.u8 %rs5, [%rd51];
and.b16 %rs6, %rs5, 240;
and.b16 %rs7, %rs5, 15;
shr.u16 %rs8, %rs6, 4;
add.s64 %rd31, %rd30, %rd1;
mul.lo.s64 %rd32, %rd31, 36;
ld.global.nc.u16 %rs9, [%rd52+-2];
shl.b16 %rs10, %rs9, 3;
cvt.u64.u16 %rd33, %rs10;
and.b64 %rd34, %rd33, 4088;
mov.u64 %rd35, iq2xs_grid;
add.s64 %rd36, %rd35, %rd34;
ld.global.nc.v2.u32 {%r111, %r112}, [%rd36];
mov.u32 %r93, 0;
shr.u16 %rs11, %rs9, 9;
cvt.u32.u16 %r115, %rs11;
popc.b32 %r116, %r115;
shl.b32 %r117, %r116, 7;
and.b32 %r118, %r117, 128;
or.b32 %r119, %r118, %r115;
mul.lo.s32 %r120, %r119, 16843009;
and.b32 %r121, %r120, 134480385;
or.b32 %r122, %r121, -2139062144;
add.s32 %r123, %r122, -16843009;
or.b32 %r62, %r123, %r121;

prmt.b32 %r62, %r62, 0, 0xba98;

xor.b32 %r124, %r62, %r111;
or.b32 %r125, %r124, -2139062144;
and.b32 %r126, %r62, 2139062143;
and.b32 %r127, %r111, -2139062144;
sub.s32 %r128, %r125, %r126;
xor.b32 %r129, %r127, %r128;
xor.b32 %r67, %r129, -2139062144;
add.s64 %rd37, %rd5, %rd32;
ld.global.nc.u32 %r68, [%rd37+4];
and.b32 %r130, %r120, -2143281136;
or.b32 %r131, %r130, -2139062144;
add.s32 %r132, %r131, -16843009;
or.b32 %r64, %r132, %r130;

prmt.b32 %r64, %r64, 0, 0xba98;

xor.b32 %r133, %r64, %r112;
or.b32 %r134, %r133, -2139062144;
and.b32 %r135, %r64, 2139062143;
and.b32 %r136, %r112, -2139062144;
sub.s32 %r137, %r134, %r135;
xor.b32 %r138, %r136, %r137;
xor.b32 %r71, %r138, -2139062144;
ld.global.nc.u32 %r72, [%rd37+8];

dp4a.s32.s32 %r66, %r67, %r68, %r93;


dp4a.s32.s32 %r70, %r71, %r72, %r66;

shl.b16 %rs12, %rs2, 3;
cvt.u64.u16 %rd38, %rs12;
and.b64 %rd39, %rd38, 4088;
add.s64 %rd40, %rd35, %rd39;
ld.global.nc.v2.u32 {%r139, %r140}, [%rd40];
shr.u16 %rs13, %rs2, 9;
cvt.u32.u16 %r143, %rs13;
popc.b32 %r144, %r143;
shl.b32 %r145, %r144, 7;
and.b32 %r146, %r145, 128;
or.b32 %r147, %r146, %r143;
mul.lo.s32 %r148, %r147, 16843009;
and.b32 %r149, %r148, 134480385;
or.b32 %r150, %r149, -2139062144;
add.s32 %r151, %r150, -16843009;
or.b32 %r74, %r151, %r149;

prmt.b32 %r74, %r74, 0, 0xba98;

xor.b32 %r152, %r74, %r139;
or.b32 %r153, %r152, -2139062144;
and.b32 %r154, %r74, 2139062143;
and.b32 %r155, %r139, -2139062144;
sub.s32 %r156, %r153, %r154;
xor.b32 %r157, %r155, %r156;
xor.b32 %r79, %r157, -2139062144;
ld.global.nc.u32 %r80, [%rd37+12];
and.b32 %r158, %r148, -2143281136;
or.b32 %r159, %r158, -2139062144;
add.s32 %r160, %r159, -16843009;
or.b32 %r76, %r160, %r158;

prmt.b32 %r76, %r76, 0, 0xba98;

xor.b32 %r161, %r76, %r140;
or.b32 %r162, %r161, -2139062144;
and.b32 %r163, %r76, 2139062143;
and.b32 %r164, %r140, -2139062144;
sub.s32 %r165, %r162, %r163;
xor.b32 %r166, %r164, %r165;
xor.b32 %r83, %r166, -2139062144;
ld.global.nc.u32 %r84, [%rd37+16];

dp4a.s32.s32 %r78, %r79, %r80, %r70;


dp4a.s32.s32 %r82, %r83, %r84, %r78;

shl.b16 %rs14, %rs3, 3;
cvt.u64.u16 %rd41, %rs14;
and.b64 %rd42, %rd41, 4088;
add.s64 %rd43, %rd35, %rd42;
ld.global.nc.v2.u32 {%r167, %r168}, [%rd43];
shr.u16 %rs15, %rs3, 9;
cvt.u32.u16 %r171, %rs15;
popc.b32 %r172, %r171;
shl.b32 %r173, %r172, 7;
and.b32 %r174, %r173, 128;
or.b32 %r175, %r174, %r171;
mul.lo.s32 %r176, %r175, 16843009;
and.b32 %r177, %r176, 134480385;
or.b32 %r178, %r177, -2139062144;
add.s32 %r179, %r178, -16843009;
or.b32 %r86, %r179, %r177;

prmt.b32 %r86, %r86, 0, 0xba98;

xor.b32 %r180, %r86, %r167;
or.b32 %r181, %r180, -2139062144;
and.b32 %r182, %r86, 2139062143;
and.b32 %r183, %r167, -2139062144;
sub.s32 %r184, %r181, %r182;
xor.b32 %r185, %r183, %r184;
xor.b32 %r91, %r185, -2139062144;
ld.global.nc.u32 %r92, [%rd37+20];
and.b32 %r186, %r176, -2143281136;
or.b32 %r187, %r186, -2139062144;
add.s32 %r188, %r187, -16843009;
or.b32 %r88, %r188, %r186;

prmt.b32 %r88, %r88, 0, 0xba98;

xor.b32 %r189, %r88, %r168;
or.b32 %r190, %r189, -2139062144;
and.b32 %r191, %r88, 2139062143;
and.b32 %r192, %r168, -2139062144;
sub.s32 %r193, %r190, %r191;
xor.b32 %r194, %r192, %r193;
xor.b32 %r95, %r194, -2139062144;
ld.global.nc.u32 %r96, [%rd37+24];

dp4a.s32.s32 %r90, %r91, %r92, %r93;


dp4a.s32.s32 %r94, %r95, %r96, %r90;

shl.b16 %rs16, %rs4, 3;
cvt.u64.u16 %rd44, %rs16;
and.b64 %rd45, %rd44, 4088;
add.s64 %rd46, %rd35, %rd45;
ld.global.nc.v2.u32 {%r195, %r196}, [%rd46];
shr.u16 %rs17, %rs4, 9;
cvt.u32.u16 %r199, %rs17;
popc.b32 %r200, %r199;
shl.b32 %r201, %r200, 7;
and.b32 %r202, %r201, 128;
or.b32 %r203, %r202, %r199;
mul.lo.s32 %r204, %r203, 16843009;
and.b32 %r205, %r204, 134480385;
or.b32 %r206, %r205, -2139062144;
add.s32 %r207, %r206, -16843009;
or.b32 %r98, %r207, %r205;

prmt.b32 %r98, %r98, 0, 0xba98;

xor.b32 %r208, %r98, %r195;
or.b32 %r209, %r208, -2139062144;
and.b32 %r210, %r98, 2139062143;
and.b32 %r211, %r195, -2139062144;
sub.s32 %r212, %r209, %r210;
xor.b32 %r213, %r211, %r212;
xor.b32 %r103, %r213, -2139062144;
ld.global.nc.u32 %r104, [%rd37+28];
and.b32 %r214, %r204, -2143281136;
or.b32 %r215, %r214, -2139062144;
add.s32 %r216, %r215, -16843009;
or.b32 %r100, %r216, %r214;

prmt.b32 %r100, %r100, 0, 0xba98;

xor.b32 %r217, %r100, %r196;
or.b32 %r218, %r217, -2139062144;
and.b32 %r219, %r100, 2139062143;
and.b32 %r220, %r196, -2139062144;
sub.s32 %r221, %r218, %r219;
xor.b32 %r222, %r220, %r221;
xor.b32 %r107, %r222, -2139062144;
ld.global.nc.u32 %r108, [%rd37+32];

dp4a.s32.s32 %r102, %r103, %r104, %r94;


dp4a.s32.s32 %r106, %r107, %r108, %r102;

cvt.u32.u16 %r223, %rs7;
cvt.u32.u16 %r224, %rs8;
mul.lo.s32 %r225, %r82, %r223;
mad.lo.s32 %r226, %r106, %r224, %r225;
add.s32 %r227, %r106, %r82;
shr.u32 %r228, %r227, 31;
add.s32 %r229, %r227, %r228;
shr.s32 %r230, %r229, 1;
add.s32 %r231, %r226, %r230;
shr.s32 %r232, %r231, 31;
shr.u32 %r233, %r232, 30;
add.s32 %r234, %r231, %r233;
shr.s32 %r235, %r234, 2;
ld.global.nc.u16 %rs1, [%rd50];

{ cvt.f32.f16 %f7, %rs1;}


ld.global.nc.u32 %r110, [%rd37];

{.reg .f16 low,high;
mov.b32 {low,high},%r110;
cvt.f32.f16 %f8, low;}


mul.ftz.f32 %f9, %f7, %f8;
cvt.rn.f32.s32 %f10, %r235;
fma.rn.ftz.f32 %f27, %f9, %f10, %f27;
add.s32 %r269, %r269, 128;
add.s64 %rd52, %rd52, 1184;
add.s64 %rd51, %rd51, 1184;
add.s64 %rd50, %rd50, 1184;
add.s32 %r270, %r270, 16;
setp.lt.s32 %p4, %r270, %r3;
@%p4 bra $L__BB216_5;

$L__BB216_6:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB216_8;

shl.b32 %r237, %r1, 7;
mov.u32 %r238, _ZZ13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r239, %r238, %r237;
shl.b32 %r240, %r35, 2;
add.s32 %r241, %r239, %r240;
st.shared.f32 [%r241+-128], %f27;

$L__BB216_8:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB216_11;

shl.b32 %r243, %r35, 2;
mov.u32 %r244, _ZZ13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r245, %r244, %r243;
mov.u32 %r246, 2;
ld.shared.f32 %f11, [%r245];
add.ftz.f32 %f12, %f27, %f11;
ld.shared.f32 %f13, [%r245+128];
add.ftz.f32 %f14, %f12, %f13;
ld.shared.f32 %f15, [%r245+256];
add.ftz.f32 %f16, %f14, %f15;
mov.b32 %r247, %f16;
mov.u32 %r248, 31;
mov.u32 %r249, 16;
mov.u32 %r250, -1;
shfl.sync.bfly.b32 %r251|%p7, %r247, %r249, %r248, %r250;
mov.b32 %f17, %r251;
add.ftz.f32 %f18, %f16, %f17;
mov.b32 %r252, %f18;
mov.u32 %r253, 8;
shfl.sync.bfly.b32 %r254|%p8, %r252, %r253, %r248, %r250;
mov.b32 %f19, %r254;
add.ftz.f32 %f20, %f18, %f19;
mov.b32 %r255, %f20;
mov.u32 %r256, 4;
shfl.sync.bfly.b32 %r257|%p9, %r255, %r256, %r248, %r250;
mov.b32 %f21, %r257;
add.ftz.f32 %f22, %f20, %f21;
mov.b32 %r258, %f22;
shfl.sync.bfly.b32 %r259|%p10, %r258, %r246, %r248, %r250;
mov.b32 %f23, %r259;
add.ftz.f32 %f4, %f22, %f23;
mov.b32 %r260, %f4;
mov.u32 %r261, 1;
shfl.sync.bfly.b32 %r16|%p1, %r260, %r261, %r248, %r250;
setp.ne.s32 %p11, %r35, 0;
@%p11 bra $L__BB216_11;

mov.u32 %r262, %ctaid.x;
mov.u32 %r263, %ctaid.y;
mad.lo.s32 %r264, %r263, %r26, %r262;
mov.u32 %r265, %ctaid.z;
mad.lo.s32 %r266, %r265, %r32, %r264;
cvta.to.global.u64 %rd47, %rd16;
mul.wide.u32 %rd48, %r266, 4;
add.s64 %rd49, %rd47, %rd48;
mov.b32 %f24, %r16;
add.ftz.f32 %f25, %f4, %f24;
st.global.f32 [%rd49], %f25;

$L__BB216_11:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type17ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type17ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type17ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type17ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type17ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type17ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type17ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type17ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type17ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<35>;
.reg .b16 %rs<37>;
.reg .f32 %f<109>;
.reg .b32 %r<682>;
.reg .b64 %rd<77>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type17ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type17ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type17ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type17ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type17ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r16, [_Z13mul_mat_vec_qIL9ggml_type17ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r17, [_Z13mul_mat_vec_qIL9ggml_type17ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r18, [_Z13mul_mat_vec_qIL9ggml_type17ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type17ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type17ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type17ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type17ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type17ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type17ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type17ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type17ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type17ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r19, [_Z13mul_mat_vec_qIL9ggml_type17ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
cvta.to.global.u64 %rd1, %rd18;
mov.u32 %r1, %tid.y;
shl.b32 %r32, %r1, 5;
mov.u32 %r33, %tid.x;
add.s32 %r2, %r32, %r33;
mov.u32 %r34, %ctaid.x;
shl.b32 %r3, %r34, 1;
shr.s32 %r35, %r2, 31;
shr.u32 %r36, %r35, 29;
add.s32 %r37, %r2, %r36;
shr.s32 %r681, %r37, 3;
shr.u32 %r5, %r31, 8;
setp.ge.s32 %p3, %r681, %r5;
mov.f32 %f105, 0f00000000;
mov.f32 %f106, %f105;
mov.f32 %f107, %f105;
mov.f32 %f108, %f105;
@%p3 bra $L__BB217_3;

mov.u32 %r38, %ctaid.y;
mov.u32 %r39, %ctaid.z;
mul.lo.s32 %r40, %r38, %r23;
cvt.u64.u32 %rd19, %r40;
mul.lo.s32 %r41, %r39, %r29;
cvt.u64.u32 %rd20, %r41;
add.s64 %rd2, %rd20, %rd19;
mul.hi.u32 %r42, %r39, %r25;
add.s32 %r43, %r42, %r39;
shr.u32 %r44, %r43, %r26;
mul.lo.s32 %r45, %r44, %r28;
mul.hi.u32 %r46, %r38, %r19;
add.s32 %r47, %r46, %r38;
shr.u32 %r48, %r47, %r20;
mul.lo.s32 %r49, %r48, %r22;
cvta.to.global.u64 %rd76, %rd16;
and.b32 %r53, %r37, -8;
sub.s32 %r54, %r2, %r53;
shl.b32 %r55, %r54, 2;
cvt.s64.s32 %rd4, %r54;
shl.b32 %r679, %r681, 3;
add.s32 %r680, %r17, %r679;
add.s32 %r56, %r681, %r45;
add.s32 %r57, %r56, %r49;
add.s32 %r58, %r3, 1;
mad.lo.s32 %r59, %r16, %r58, %r57;
mul.wide.s32 %rd5, %r59, 74;
add.s64 %rd21, %rd5, %rd4;
add.s64 %rd6, %rd21, 66;
mad.lo.s32 %r60, %r3, %r16, %r681;
add.s32 %r61, %r60, %r45;
add.s32 %r62, %r61, %r49;
mul.wide.s32 %rd7, %r62, 74;
add.s64 %rd22, %rd7, %rd4;
add.s64 %rd8, %rd22, 66;
mul.wide.s32 %rd23, %r55, 2;
add.s64 %rd24, %rd5, %rd23;
add.s64 %rd9, %rd24, 4;
add.s64 %rd25, %rd7, %rd23;
add.s64 %rd10, %rd25, 8;
cvta.to.global.u64 %rd11, %rd17;
mov.f32 %f105, 0f00000000;

$L__BB217_2:
cvt.u64.u32 %rd26, %r679;
add.s64 %rd27, %rd2, %rd26;
add.s64 %rd28, %rd27, %rd4;
mul.lo.s64 %rd29, %rd28, 36;
add.s64 %rd30, %rd76, %rd10;
ld.global.nc.u16 %rs5, [%rd30+-4];
ld.global.nc.u16 %rs6, [%rd30+-2];
ld.global.nc.u16 %rs7, [%rd30];
add.s64 %rd31, %rd76, %rd8;
ld.global.nc.u8 %rs8, [%rd31];
and.b16 %rs9, %rs8, 240;
and.b16 %rs10, %rs8, 15;
shr.u16 %rs11, %rs9, 4;
ld.global.nc.u16 %rs12, [%rd30+-6];
shl.b16 %rs13, %rs12, 3;
cvt.u64.u16 %rd32, %rs13;
and.b64 %rd33, %rd32, 4088;
mov.u64 %rd34, iq2xs_grid;
add.s64 %rd35, %rd34, %rd33;
ld.global.nc.v2.u32 {%r259, %r260}, [%rd35];
not.b32 %r262, %r259;
mov.u32 %r241, 0;
not.b32 %r264, %r260;
shr.u16 %rs14, %rs12, 9;
cvt.u32.u16 %r265, %rs14;
popc.b32 %r266, %r265;
shl.b32 %r267, %r266, 7;
and.b32 %r268, %r267, 128;
or.b32 %r269, %r268, %r265;
mul.lo.s32 %r270, %r269, 16843009;
and.b32 %r271, %r270, 134480385;
or.b32 %r272, %r271, -2139062144;
add.s32 %r273, %r272, -16843009;
or.b32 %r161, %r273, %r271;
mov.u32 %r63, %r161;

prmt.b32 %r63, %r63, 0, 0xba98;

xor.b32 %r274, %r63, %r259;
or.b32 %r275, %r274, -2139062144;
and.b32 %r276, %r63, 2139062143;
and.b32 %r277, %r262, -2139062144;
sub.s32 %r278, %r275, %r276;
xor.b32 %r68, %r277, %r278;
add.s64 %rd36, %rd11, %rd29;
ld.global.nc.u32 %r69, [%rd36+4];
and.b32 %r279, %r270, -2143281136;
or.b32 %r280, %r279, -2139062144;
add.s32 %r281, %r280, -16843009;
or.b32 %r163, %r281, %r279;
mov.u32 %r65, %r163;

prmt.b32 %r65, %r65, 0, 0xba98;

xor.b32 %r282, %r65, %r260;
or.b32 %r283, %r282, -2139062144;
and.b32 %r284, %r65, 2139062143;
and.b32 %r285, %r264, -2139062144;
sub.s32 %r286, %r283, %r284;
xor.b32 %r72, %r285, %r286;
ld.global.nc.u32 %r73, [%rd36+8];

dp4a.s32.s32 %r67, %r68, %r69, %r241;


dp4a.s32.s32 %r71, %r72, %r73, %r67;

shl.b16 %rs15, %rs5, 3;
cvt.u64.u16 %rd37, %rs15;
and.b64 %rd38, %rd37, 4088;
add.s64 %rd39, %rd34, %rd38;
ld.global.nc.v2.u32 {%r287, %r288}, [%rd39];
not.b32 %r290, %r287;
not.b32 %r292, %r288;
shr.u16 %rs16, %rs5, 9;
cvt.u32.u16 %r293, %rs16;
popc.b32 %r294, %r293;
shl.b32 %r295, %r294, 7;
and.b32 %r296, %r295, 128;
or.b32 %r297, %r296, %r293;
mul.lo.s32 %r298, %r297, 16843009;
and.b32 %r299, %r298, 134480385;
or.b32 %r300, %r299, -2139062144;
add.s32 %r301, %r300, -16843009;
or.b32 %r173, %r301, %r299;
mov.u32 %r75, %r173;

prmt.b32 %r75, %r75, 0, 0xba98;

xor.b32 %r302, %r75, %r287;
or.b32 %r303, %r302, -2139062144;
and.b32 %r304, %r75, 2139062143;
and.b32 %r305, %r290, -2139062144;
sub.s32 %r306, %r303, %r304;
xor.b32 %r80, %r305, %r306;
ld.global.nc.u32 %r81, [%rd36+12];
and.b32 %r307, %r298, -2143281136;
or.b32 %r308, %r307, -2139062144;
add.s32 %r309, %r308, -16843009;
or.b32 %r175, %r309, %r307;
mov.u32 %r77, %r175;

prmt.b32 %r77, %r77, 0, 0xba98;

xor.b32 %r310, %r77, %r288;
or.b32 %r311, %r310, -2139062144;
and.b32 %r312, %r77, 2139062143;
and.b32 %r313, %r292, -2139062144;
sub.s32 %r314, %r311, %r312;
xor.b32 %r84, %r313, %r314;
ld.global.nc.u32 %r85, [%rd36+16];

dp4a.s32.s32 %r79, %r80, %r81, %r71;


dp4a.s32.s32 %r83, %r84, %r85, %r79;

shl.b16 %rs17, %rs6, 3;
cvt.u64.u16 %rd40, %rs17;
and.b64 %rd41, %rd40, 4088;
add.s64 %rd42, %rd34, %rd41;
ld.global.nc.v2.u32 {%r315, %r316}, [%rd42];
not.b32 %r318, %r315;
not.b32 %r320, %r316;
shr.u16 %rs18, %rs6, 9;
cvt.u32.u16 %r321, %rs18;
popc.b32 %r322, %r321;
shl.b32 %r323, %r322, 7;
and.b32 %r324, %r323, 128;
or.b32 %r325, %r324, %r321;
mul.lo.s32 %r326, %r325, 16843009;
and.b32 %r327, %r326, 134480385;
or.b32 %r328, %r327, -2139062144;
add.s32 %r329, %r328, -16843009;
or.b32 %r185, %r329, %r327;
mov.u32 %r87, %r185;

prmt.b32 %r87, %r87, 0, 0xba98;

xor.b32 %r330, %r87, %r315;
or.b32 %r331, %r330, -2139062144;
and.b32 %r332, %r87, 2139062143;
and.b32 %r333, %r318, -2139062144;
sub.s32 %r334, %r331, %r332;
xor.b32 %r92, %r333, %r334;
ld.global.nc.u32 %r93, [%rd36+20];
and.b32 %r335, %r326, -2143281136;
or.b32 %r336, %r335, -2139062144;
add.s32 %r337, %r336, -16843009;
or.b32 %r187, %r337, %r335;
mov.u32 %r89, %r187;

prmt.b32 %r89, %r89, 0, 0xba98;

xor.b32 %r338, %r89, %r316;
or.b32 %r339, %r338, -2139062144;
and.b32 %r340, %r89, 2139062143;
and.b32 %r341, %r320, -2139062144;
sub.s32 %r342, %r339, %r340;
xor.b32 %r96, %r341, %r342;
ld.global.nc.u32 %r97, [%rd36+24];

dp4a.s32.s32 %r91, %r92, %r93, %r241;


dp4a.s32.s32 %r95, %r96, %r97, %r91;

shl.b16 %rs19, %rs7, 3;
cvt.u64.u16 %rd43, %rs19;
and.b64 %rd44, %rd43, 4088;
add.s64 %rd45, %rd34, %rd44;
ld.global.nc.v2.u32 {%r343, %r344}, [%rd45];
not.b32 %r346, %r343;
not.b32 %r348, %r344;
shr.u16 %rs20, %rs7, 9;
cvt.u32.u16 %r349, %rs20;
popc.b32 %r350, %r349;
shl.b32 %r351, %r350, 7;
and.b32 %r352, %r351, 128;
or.b32 %r353, %r352, %r349;
mul.lo.s32 %r354, %r353, 16843009;
and.b32 %r355, %r354, 134480385;
or.b32 %r356, %r355, -2139062144;
add.s32 %r357, %r356, -16843009;
or.b32 %r197, %r357, %r355;
mov.u32 %r99, %r197;

prmt.b32 %r99, %r99, 0, 0xba98;

xor.b32 %r358, %r99, %r343;
or.b32 %r359, %r358, -2139062144;
and.b32 %r360, %r99, 2139062143;
and.b32 %r361, %r346, -2139062144;
sub.s32 %r362, %r359, %r360;
xor.b32 %r104, %r361, %r362;
ld.global.nc.u32 %r105, [%rd36+28];
and.b32 %r363, %r354, -2143281136;
or.b32 %r364, %r363, -2139062144;
add.s32 %r365, %r364, -16843009;
or.b32 %r199, %r365, %r363;
mov.u32 %r101, %r199;

prmt.b32 %r101, %r101, 0, 0xba98;

xor.b32 %r366, %r101, %r344;
or.b32 %r367, %r366, -2139062144;
and.b32 %r368, %r101, 2139062143;
and.b32 %r369, %r348, -2139062144;
sub.s32 %r370, %r367, %r368;
xor.b32 %r108, %r369, %r370;
ld.global.nc.u32 %r109, [%rd36+32];

dp4a.s32.s32 %r103, %r104, %r105, %r95;


dp4a.s32.s32 %r107, %r108, %r109, %r103;

cvt.u32.u16 %r371, %rs10;
cvt.u32.u16 %r372, %rs11;
mul.lo.s32 %r373, %r83, %r371;
mad.lo.s32 %r374, %r107, %r372, %r373;
add.s32 %r375, %r107, %r83;
shr.u32 %r376, %r375, 31;
add.s32 %r377, %r375, %r376;
shr.s32 %r378, %r377, 1;
add.s32 %r379, %r374, %r378;
shr.s32 %r380, %r379, 31;
shr.u32 %r381, %r380, 30;
add.s32 %r382, %r379, %r381;
shr.s32 %r383, %r382, 2;
add.s64 %rd46, %rd76, %rd7;
ld.global.nc.u16 %rs1, [%rd46];

{ cvt.f32.f16 %f25, %rs1;}


ld.global.nc.u32 %r111, [%rd36];

{.reg .f16 low,high;
mov.b32 {low,high},%r111;
cvt.f32.f16 %f26, low;}


mul.ftz.f32 %f33, %f25, %f26;
cvt.rn.f32.s32 %f34, %r383;
fma.rn.ftz.f32 %f108, %f33, %f34, %f108;
add.s64 %rd47, %rd76, %rd9;
ld.global.nc.u16 %rs21, [%rd47];
ld.global.nc.u16 %rs22, [%rd47+2];
ld.global.nc.u16 %rs23, [%rd47+4];
add.s64 %rd48, %rd76, %rd6;
ld.global.nc.u8 %rs24, [%rd48];
and.b16 %rs25, %rs24, 240;
and.b16 %rs26, %rs24, 15;
shr.u16 %rs27, %rs25, 4;
ld.global.nc.u16 %rs28, [%rd47+-2];
shl.b16 %rs29, %rs28, 3;
cvt.u64.u16 %rd49, %rs29;
and.b64 %rd50, %rd49, 4088;
add.s64 %rd51, %rd34, %rd50;
ld.global.nc.v2.u32 {%r384, %r385}, [%rd51];
not.b32 %r387, %r384;
not.b32 %r389, %r385;
shr.u16 %rs30, %rs28, 9;
cvt.u32.u16 %r390, %rs30;
popc.b32 %r391, %r390;
shl.b32 %r392, %r391, 7;
and.b32 %r393, %r392, 128;
or.b32 %r394, %r393, %r390;
mul.lo.s32 %r395, %r394, 16843009;
and.b32 %r396, %r395, 134480385;
or.b32 %r397, %r396, -2139062144;
add.s32 %r398, %r397, -16843009;
or.b32 %r210, %r398, %r396;
mov.u32 %r112, %r210;

prmt.b32 %r112, %r112, 0, 0xba98;

xor.b32 %r399, %r112, %r384;
or.b32 %r400, %r399, -2139062144;
and.b32 %r401, %r112, 2139062143;
and.b32 %r402, %r387, -2139062144;
sub.s32 %r403, %r400, %r401;
xor.b32 %r117, %r402, %r403;
and.b32 %r404, %r395, -2143281136;
or.b32 %r405, %r404, -2139062144;
add.s32 %r406, %r405, -16843009;
or.b32 %r212, %r406, %r404;
mov.u32 %r114, %r212;

prmt.b32 %r114, %r114, 0, 0xba98;

xor.b32 %r407, %r114, %r385;
or.b32 %r408, %r407, -2139062144;
and.b32 %r409, %r114, 2139062143;
and.b32 %r410, %r389, -2139062144;
sub.s32 %r411, %r408, %r409;
xor.b32 %r121, %r410, %r411;

dp4a.s32.s32 %r116, %r117, %r69, %r241;


dp4a.s32.s32 %r120, %r121, %r73, %r116;

shl.b16 %rs31, %rs21, 3;
cvt.u64.u16 %rd52, %rs31;
and.b64 %rd53, %rd52, 4088;
add.s64 %rd54, %rd34, %rd53;
ld.global.nc.v2.u32 {%r412, %r413}, [%rd54];
not.b32 %r415, %r412;
not.b32 %r417, %r413;
shr.u16 %rs32, %rs21, 9;
cvt.u32.u16 %r418, %rs32;
popc.b32 %r419, %r418;
shl.b32 %r420, %r419, 7;
and.b32 %r421, %r420, 128;
or.b32 %r422, %r421, %r418;
mul.lo.s32 %r423, %r422, 16843009;
and.b32 %r424, %r423, 134480385;
or.b32 %r425, %r424, -2139062144;
add.s32 %r426, %r425, -16843009;
or.b32 %r222, %r426, %r424;
mov.u32 %r124, %r222;

prmt.b32 %r124, %r124, 0, 0xba98;

xor.b32 %r427, %r124, %r412;
or.b32 %r428, %r427, -2139062144;
and.b32 %r429, %r124, 2139062143;
and.b32 %r430, %r415, -2139062144;
sub.s32 %r431, %r428, %r429;
xor.b32 %r129, %r430, %r431;
and.b32 %r432, %r423, -2143281136;
or.b32 %r433, %r432, -2139062144;
add.s32 %r434, %r433, -16843009;
or.b32 %r224, %r434, %r432;
mov.u32 %r126, %r224;

prmt.b32 %r126, %r126, 0, 0xba98;

xor.b32 %r435, %r126, %r413;
or.b32 %r436, %r435, -2139062144;
and.b32 %r437, %r126, 2139062143;
and.b32 %r438, %r417, -2139062144;
sub.s32 %r439, %r436, %r437;
xor.b32 %r133, %r438, %r439;

dp4a.s32.s32 %r128, %r129, %r81, %r120;


dp4a.s32.s32 %r132, %r133, %r85, %r128;

shl.b16 %rs33, %rs22, 3;
cvt.u64.u16 %rd55, %rs33;
and.b64 %rd56, %rd55, 4088;
add.s64 %rd57, %rd34, %rd56;
ld.global.nc.v2.u32 {%r440, %r441}, [%rd57];
not.b32 %r443, %r440;
not.b32 %r445, %r441;
shr.u16 %rs34, %rs22, 9;
cvt.u32.u16 %r446, %rs34;
popc.b32 %r447, %r446;
shl.b32 %r448, %r447, 7;
and.b32 %r449, %r448, 128;
or.b32 %r450, %r449, %r446;
mul.lo.s32 %r451, %r450, 16843009;
and.b32 %r452, %r451, 134480385;
or.b32 %r453, %r452, -2139062144;
add.s32 %r454, %r453, -16843009;
or.b32 %r234, %r454, %r452;
mov.u32 %r136, %r234;

prmt.b32 %r136, %r136, 0, 0xba98;

xor.b32 %r455, %r136, %r440;
or.b32 %r456, %r455, -2139062144;
and.b32 %r457, %r136, 2139062143;
and.b32 %r458, %r443, -2139062144;
sub.s32 %r459, %r456, %r457;
xor.b32 %r141, %r458, %r459;
and.b32 %r460, %r451, -2143281136;
or.b32 %r461, %r460, -2139062144;
add.s32 %r462, %r461, -16843009;
or.b32 %r236, %r462, %r460;
mov.u32 %r138, %r236;

prmt.b32 %r138, %r138, 0, 0xba98;

xor.b32 %r463, %r138, %r441;
or.b32 %r464, %r463, -2139062144;
and.b32 %r465, %r138, 2139062143;
and.b32 %r466, %r445, -2139062144;
sub.s32 %r467, %r464, %r465;
xor.b32 %r145, %r466, %r467;

dp4a.s32.s32 %r140, %r141, %r93, %r241;


dp4a.s32.s32 %r144, %r145, %r97, %r140;

shl.b16 %rs35, %rs23, 3;
cvt.u64.u16 %rd58, %rs35;
and.b64 %rd59, %rd58, 4088;
add.s64 %rd60, %rd34, %rd59;
ld.global.nc.v2.u32 {%r468, %r469}, [%rd60];
not.b32 %r471, %r468;
not.b32 %r473, %r469;
shr.u16 %rs36, %rs23, 9;
cvt.u32.u16 %r474, %rs36;
popc.b32 %r475, %r474;
shl.b32 %r476, %r475, 7;
and.b32 %r477, %r476, 128;
or.b32 %r478, %r477, %r474;
mul.lo.s32 %r479, %r478, 16843009;
and.b32 %r480, %r479, 134480385;
or.b32 %r481, %r480, -2139062144;
add.s32 %r482, %r481, -16843009;
or.b32 %r246, %r482, %r480;
mov.u32 %r148, %r246;

prmt.b32 %r148, %r148, 0, 0xba98;

xor.b32 %r483, %r148, %r468;
or.b32 %r484, %r483, -2139062144;
and.b32 %r485, %r148, 2139062143;
and.b32 %r486, %r471, -2139062144;
sub.s32 %r487, %r484, %r485;
xor.b32 %r153, %r486, %r487;
and.b32 %r488, %r479, -2143281136;
or.b32 %r489, %r488, -2139062144;
add.s32 %r490, %r489, -16843009;
or.b32 %r248, %r490, %r488;
mov.u32 %r150, %r248;

prmt.b32 %r150, %r150, 0, 0xba98;

xor.b32 %r491, %r150, %r469;
or.b32 %r492, %r491, -2139062144;
and.b32 %r493, %r150, 2139062143;
and.b32 %r494, %r473, -2139062144;
sub.s32 %r495, %r492, %r493;
xor.b32 %r157, %r494, %r495;

dp4a.s32.s32 %r152, %r153, %r105, %r144;


dp4a.s32.s32 %r156, %r157, %r109, %r152;

cvt.u32.u16 %r496, %rs26;
cvt.u32.u16 %r497, %rs27;
mul.lo.s32 %r498, %r132, %r496;
mad.lo.s32 %r499, %r156, %r497, %r498;
add.s32 %r500, %r156, %r132;
shr.u32 %r501, %r500, 31;
add.s32 %r502, %r500, %r501;
shr.s32 %r503, %r502, 1;
add.s32 %r504, %r499, %r503;
shr.s32 %r505, %r504, 31;
shr.u32 %r506, %r505, 30;
add.s32 %r507, %r504, %r506;
shr.s32 %r508, %r507, 2;
add.s64 %rd61, %rd76, %rd5;
ld.global.nc.u16 %rs2, [%rd61];

{ cvt.f32.f16 %f27, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r111;
cvt.f32.f16 %f28, low;}


mul.ftz.f32 %f35, %f27, %f28;
cvt.rn.f32.s32 %f36, %r508;
fma.rn.ftz.f32 %f107, %f35, %f36, %f107;
cvt.u64.u32 %rd62, %r680;
add.s64 %rd63, %rd2, %rd62;
add.s64 %rd64, %rd63, %rd4;
mul.lo.s64 %rd65, %rd64, 36;

prmt.b32 %r161, %r161, 0, 0xba98;

xor.b32 %r509, %r161, %r259;
or.b32 %r510, %r509, -2139062144;
and.b32 %r511, %r161, 2139062143;
sub.s32 %r512, %r510, %r511;
xor.b32 %r166, %r277, %r512;
add.s64 %rd66, %rd11, %rd65;
ld.global.nc.u32 %r167, [%rd66+4];

prmt.b32 %r163, %r163, 0, 0xba98;

xor.b32 %r513, %r163, %r260;
or.b32 %r514, %r513, -2139062144;
and.b32 %r515, %r163, 2139062143;
sub.s32 %r516, %r514, %r515;
xor.b32 %r170, %r285, %r516;
ld.global.nc.u32 %r171, [%rd66+8];

dp4a.s32.s32 %r165, %r166, %r167, %r241;


dp4a.s32.s32 %r169, %r170, %r171, %r165;


prmt.b32 %r173, %r173, 0, 0xba98;

xor.b32 %r517, %r173, %r287;
or.b32 %r518, %r517, -2139062144;
and.b32 %r519, %r173, 2139062143;
sub.s32 %r520, %r518, %r519;
xor.b32 %r178, %r305, %r520;
ld.global.nc.u32 %r179, [%rd66+12];

prmt.b32 %r175, %r175, 0, 0xba98;

xor.b32 %r521, %r175, %r288;
or.b32 %r522, %r521, -2139062144;
and.b32 %r523, %r175, 2139062143;
sub.s32 %r524, %r522, %r523;
xor.b32 %r182, %r313, %r524;
ld.global.nc.u32 %r183, [%rd66+16];

dp4a.s32.s32 %r177, %r178, %r179, %r169;


dp4a.s32.s32 %r181, %r182, %r183, %r177;


prmt.b32 %r185, %r185, 0, 0xba98;

xor.b32 %r525, %r185, %r315;
or.b32 %r526, %r525, -2139062144;
and.b32 %r527, %r185, 2139062143;
sub.s32 %r528, %r526, %r527;
xor.b32 %r190, %r333, %r528;
ld.global.nc.u32 %r191, [%rd66+20];

prmt.b32 %r187, %r187, 0, 0xba98;

xor.b32 %r529, %r187, %r316;
or.b32 %r530, %r529, -2139062144;
and.b32 %r531, %r187, 2139062143;
sub.s32 %r532, %r530, %r531;
xor.b32 %r194, %r341, %r532;
ld.global.nc.u32 %r195, [%rd66+24];

dp4a.s32.s32 %r189, %r190, %r191, %r241;


dp4a.s32.s32 %r193, %r194, %r195, %r189;


prmt.b32 %r197, %r197, 0, 0xba98;

xor.b32 %r533, %r197, %r343;
or.b32 %r534, %r533, -2139062144;
and.b32 %r535, %r197, 2139062143;
sub.s32 %r536, %r534, %r535;
xor.b32 %r202, %r361, %r536;
ld.global.nc.u32 %r203, [%rd66+28];

prmt.b32 %r199, %r199, 0, 0xba98;

xor.b32 %r537, %r199, %r344;
or.b32 %r538, %r537, -2139062144;
and.b32 %r539, %r199, 2139062143;
sub.s32 %r540, %r538, %r539;
xor.b32 %r206, %r369, %r540;
ld.global.nc.u32 %r207, [%rd66+32];

dp4a.s32.s32 %r201, %r202, %r203, %r193;


dp4a.s32.s32 %r205, %r206, %r207, %r201;

mul.lo.s32 %r541, %r181, %r371;
mad.lo.s32 %r542, %r205, %r372, %r541;
add.s32 %r543, %r205, %r181;
shr.u32 %r544, %r543, 31;
add.s32 %r545, %r543, %r544;
shr.s32 %r546, %r545, 1;
add.s32 %r547, %r542, %r546;
shr.s32 %r548, %r547, 31;
shr.u32 %r549, %r548, 30;
add.s32 %r550, %r547, %r549;
shr.s32 %r551, %r550, 2;

{ cvt.f32.f16 %f29, %rs1;}


ld.global.nc.u32 %r209, [%rd66];

{.reg .f16 low,high;
mov.b32 {low,high},%r209;
cvt.f32.f16 %f30, low;}


mul.ftz.f32 %f37, %f29, %f30;
cvt.rn.f32.s32 %f38, %r551;
fma.rn.ftz.f32 %f106, %f37, %f38, %f106;

prmt.b32 %r210, %r210, 0, 0xba98;

xor.b32 %r552, %r210, %r384;
or.b32 %r553, %r552, -2139062144;
and.b32 %r554, %r210, 2139062143;
sub.s32 %r555, %r553, %r554;
xor.b32 %r215, %r402, %r555;

prmt.b32 %r212, %r212, 0, 0xba98;

xor.b32 %r556, %r212, %r385;
or.b32 %r557, %r556, -2139062144;
and.b32 %r558, %r212, 2139062143;
sub.s32 %r559, %r557, %r558;
xor.b32 %r219, %r410, %r559;

dp4a.s32.s32 %r214, %r215, %r167, %r241;


dp4a.s32.s32 %r218, %r219, %r171, %r214;


prmt.b32 %r222, %r222, 0, 0xba98;

xor.b32 %r560, %r222, %r412;
or.b32 %r561, %r560, -2139062144;
and.b32 %r562, %r222, 2139062143;
sub.s32 %r563, %r561, %r562;
xor.b32 %r227, %r430, %r563;

prmt.b32 %r224, %r224, 0, 0xba98;

xor.b32 %r564, %r224, %r413;
or.b32 %r565, %r564, -2139062144;
and.b32 %r566, %r224, 2139062143;
sub.s32 %r567, %r565, %r566;
xor.b32 %r231, %r438, %r567;

dp4a.s32.s32 %r226, %r227, %r179, %r218;


dp4a.s32.s32 %r230, %r231, %r183, %r226;


prmt.b32 %r234, %r234, 0, 0xba98;

xor.b32 %r568, %r234, %r440;
or.b32 %r569, %r568, -2139062144;
and.b32 %r570, %r234, 2139062143;
sub.s32 %r571, %r569, %r570;
xor.b32 %r239, %r458, %r571;

prmt.b32 %r236, %r236, 0, 0xba98;

xor.b32 %r572, %r236, %r441;
or.b32 %r573, %r572, -2139062144;
and.b32 %r574, %r236, 2139062143;
sub.s32 %r575, %r573, %r574;
xor.b32 %r243, %r466, %r575;

dp4a.s32.s32 %r238, %r239, %r191, %r241;


dp4a.s32.s32 %r242, %r243, %r195, %r238;


prmt.b32 %r246, %r246, 0, 0xba98;

xor.b32 %r576, %r246, %r468;
or.b32 %r577, %r576, -2139062144;
and.b32 %r578, %r246, 2139062143;
sub.s32 %r579, %r577, %r578;
xor.b32 %r251, %r486, %r579;

prmt.b32 %r248, %r248, 0, 0xba98;

xor.b32 %r580, %r248, %r469;
or.b32 %r581, %r580, -2139062144;
and.b32 %r582, %r248, 2139062143;
sub.s32 %r583, %r581, %r582;
xor.b32 %r255, %r494, %r583;

dp4a.s32.s32 %r250, %r251, %r203, %r242;


dp4a.s32.s32 %r254, %r255, %r207, %r250;

mul.lo.s32 %r584, %r230, %r496;
mad.lo.s32 %r585, %r254, %r497, %r584;
add.s32 %r586, %r254, %r230;
shr.u32 %r587, %r586, 31;
add.s32 %r588, %r586, %r587;
shr.s32 %r589, %r588, 1;
add.s32 %r590, %r585, %r589;
shr.s32 %r591, %r590, 31;
shr.u32 %r592, %r591, 30;
add.s32 %r593, %r590, %r592;
shr.s32 %r594, %r593, 2;

{ cvt.f32.f16 %f31, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r209;
cvt.f32.f16 %f32, low;}


mul.ftz.f32 %f39, %f31, %f32;
cvt.rn.f32.s32 %f40, %r594;
fma.rn.ftz.f32 %f105, %f39, %f40, %f105;
add.s32 %r680, %r680, 128;
add.s32 %r679, %r679, 128;
add.s64 %rd76, %rd76, 1184;
add.s32 %r681, %r681, 16;
setp.lt.s32 %p4, %r681, %r5;
@%p4 bra $L__BB217_2;

$L__BB217_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB217_5;

shl.b32 %r596, %r1, 9;
mov.u32 %r597, _ZZ13mul_mat_vec_qIL9ggml_type17ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r598, %r597, %r596;
shl.b32 %r599, %r33, 2;
add.s32 %r600, %r598, %r599;
st.shared.f32 [%r600+-512], %f108;
st.shared.f32 [%r600+-384], %f107;
st.shared.f32 [%r600+-256], %f106;
st.shared.f32 [%r600+-128], %f105;

$L__BB217_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB217_14;

mov.u32 %r601, %ctaid.z;
mov.u32 %r602, %ctaid.y;
mad.lo.s32 %r603, %r602, %r24, %r3;
mad.lo.s32 %r604, %r601, %r30, %r603;
cvt.u64.u32 %rd14, %r604;
shl.b32 %r606, %r33, 2;
mov.u32 %r607, _ZZ13mul_mat_vec_qIL9ggml_type17ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r14, %r607, %r606;
mov.u32 %r608, 2;
ld.shared.f32 %f41, [%r14];
add.ftz.f32 %f42, %f41, %f108;
ld.shared.f32 %f43, [%r14+512];
add.ftz.f32 %f44, %f43, %f42;
ld.shared.f32 %f45, [%r14+1024];
add.ftz.f32 %f46, %f45, %f44;
mov.b32 %r609, %f46;
mov.u32 %r610, 31;
mov.u32 %r611, 16;
mov.u32 %r612, -1;
shfl.sync.bfly.b32 %r613|%p7, %r609, %r611, %r610, %r612;
mov.b32 %f47, %r613;
add.ftz.f32 %f48, %f46, %f47;
mov.b32 %r614, %f48;
mov.u32 %r615, 8;
shfl.sync.bfly.b32 %r616|%p8, %r614, %r615, %r610, %r612;
mov.b32 %f49, %r616;
add.ftz.f32 %f50, %f48, %f49;
mov.b32 %r617, %f50;
mov.u32 %r618, 4;
shfl.sync.bfly.b32 %r619|%p9, %r617, %r618, %r610, %r612;
mov.b32 %f51, %r619;
add.ftz.f32 %f52, %f50, %f51;
mov.b32 %r620, %f52;
shfl.sync.bfly.b32 %r621|%p10, %r620, %r608, %r610, %r612;
mov.b32 %f53, %r621;
add.ftz.f32 %f54, %f52, %f53;
mov.b32 %r622, %f54;
mov.u32 %r623, 1;
shfl.sync.bfly.b32 %r624|%p11, %r622, %r623, %r610, %r612;
mov.b32 %f55, %r624;
add.ftz.f32 %f13, %f54, %f55;
setp.eq.s32 %p12, %r33, 0;
setp.lt.u32 %p13, %r3, %r18;
and.pred %p1, %p12, %p13;
mul.wide.u32 %rd67, %r604, 4;
add.s64 %rd15, %rd1, %rd67;
not.pred %p14, %p1;
@%p14 bra $L__BB217_8;

st.global.f32 [%rd15], %f13;

$L__BB217_8:
ld.shared.f32 %f56, [%r14+128];
add.ftz.f32 %f57, %f56, %f107;
ld.shared.f32 %f58, [%r14+640];
add.ftz.f32 %f59, %f58, %f57;
ld.shared.f32 %f60, [%r14+1152];
add.ftz.f32 %f61, %f60, %f59;
mov.b32 %r625, %f61;
shfl.sync.bfly.b32 %r629|%p15, %r625, %r611, %r610, %r612;
mov.b32 %f62, %r629;
add.ftz.f32 %f63, %f61, %f62;
mov.b32 %r630, %f63;
shfl.sync.bfly.b32 %r632|%p16, %r630, %r615, %r610, %r612;
mov.b32 %f64, %r632;
add.ftz.f32 %f65, %f63, %f64;
mov.b32 %r633, %f65;
shfl.sync.bfly.b32 %r635|%p17, %r633, %r618, %r610, %r612;
mov.b32 %f66, %r635;
add.ftz.f32 %f67, %f65, %f66;
mov.b32 %r636, %f67;
shfl.sync.bfly.b32 %r638|%p18, %r636, %r608, %r610, %r612;
mov.b32 %f68, %r638;
add.ftz.f32 %f69, %f67, %f68;
mov.b32 %r639, %f69;
shfl.sync.bfly.b32 %r641|%p19, %r639, %r623, %r610, %r612;
mov.b32 %f70, %r641;
add.ftz.f32 %f14, %f69, %f70;
setp.eq.s32 %p20, %r33, 1;
add.s32 %r643, %r3, 1;
setp.lt.u32 %p21, %r643, %r18;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB217_10;

st.global.f32 [%rd15+4], %f14;

$L__BB217_10:
ld.shared.f32 %f71, [%r14+256];
add.ftz.f32 %f72, %f71, %f106;
ld.shared.f32 %f73, [%r14+768];
add.ftz.f32 %f74, %f73, %f72;
ld.shared.f32 %f75, [%r14+1280];
add.ftz.f32 %f76, %f75, %f74;
mov.b32 %r644, %f76;
mov.u32 %r645, 31;
mov.u32 %r646, 16;
mov.u32 %r647, -1;
shfl.sync.bfly.b32 %r648|%p23, %r644, %r646, %r645, %r647;
mov.b32 %f77, %r648;
add.ftz.f32 %f78, %f76, %f77;
mov.b32 %r649, %f78;
mov.u32 %r650, 8;
shfl.sync.bfly.b32 %r651|%p24, %r649, %r650, %r645, %r647;
mov.b32 %f79, %r651;
add.ftz.f32 %f80, %f78, %f79;
mov.b32 %r652, %f80;
mov.u32 %r653, 4;
shfl.sync.bfly.b32 %r654|%p25, %r652, %r653, %r645, %r647;
mov.b32 %f81, %r654;
add.ftz.f32 %f82, %f80, %f81;
mov.b32 %r655, %f82;
mov.u32 %r656, 2;
shfl.sync.bfly.b32 %r657|%p26, %r655, %r656, %r645, %r647;
mov.b32 %f83, %r657;
add.ftz.f32 %f84, %f82, %f83;
mov.b32 %r658, %f84;
mov.u32 %r659, 1;
shfl.sync.bfly.b32 %r660|%p27, %r658, %r659, %r645, %r647;
mov.b32 %f85, %r660;
add.ftz.f32 %f15, %f84, %f85;
@%p14 bra $L__BB217_12;

cvt.u64.u32 %rd68, %r18;
add.s64 %rd69, %rd68, %rd14;
shl.b64 %rd70, %rd69, 2;
add.s64 %rd71, %rd1, %rd70;
st.global.f32 [%rd71], %f15;

$L__BB217_12:
ld.shared.f32 %f86, [%r14+384];
add.ftz.f32 %f87, %f86, %f105;
ld.shared.f32 %f88, [%r14+896];
add.ftz.f32 %f89, %f88, %f87;
ld.shared.f32 %f90, [%r14+1408];
add.ftz.f32 %f91, %f90, %f89;
mov.b32 %r661, %f91;
shfl.sync.bfly.b32 %r665|%p29, %r661, %r646, %r645, %r647;
mov.b32 %f92, %r665;
add.ftz.f32 %f93, %f91, %f92;
mov.b32 %r666, %f93;
shfl.sync.bfly.b32 %r668|%p30, %r666, %r650, %r645, %r647;
mov.b32 %f94, %r668;
add.ftz.f32 %f95, %f93, %f94;
mov.b32 %r669, %f95;
shfl.sync.bfly.b32 %r671|%p31, %r669, %r653, %r645, %r647;
mov.b32 %f96, %r671;
add.ftz.f32 %f97, %f95, %f96;
mov.b32 %r672, %f97;
shfl.sync.bfly.b32 %r674|%p32, %r672, %r656, %r645, %r647;
mov.b32 %f98, %r674;
add.ftz.f32 %f99, %f97, %f98;
mov.b32 %r675, %f99;
shfl.sync.bfly.b32 %r677|%p33, %r675, %r659, %r645, %r647;
mov.b32 %f100, %r677;
add.ftz.f32 %f16, %f99, %f100;
@%p22 bra $L__BB217_14;

add.s32 %r678, %r18, 1;
cvt.u64.u32 %rd72, %r678;
add.s64 %rd73, %rd72, %rd14;
shl.b64 %rd74, %rd73, 2;
add.s64 %rd75, %rd1, %rd74;
st.global.f32 [%rd75], %f16;

$L__BB217_14:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type17ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type17ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type17ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type17ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type17ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type17ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type17ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type17ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type17ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<47>;
.reg .b16 %rs<39>;
.reg .f32 %f<163>;
.reg .b32 %r<905>;
.reg .b64 %rd<86>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type17ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[2304];

ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type17ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type17ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type17ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type17ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r19, [_Z13mul_mat_vec_qIL9ggml_type17ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type17ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type17ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type17ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type17ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type17ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type17ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type17ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type17ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type17ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type17ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type17ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type17ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
cvta.to.global.u64 %rd1, %rd19;
mov.u32 %r1, %tid.y;
shl.b32 %r35, %r1, 5;
mov.u32 %r36, %tid.x;
add.s32 %r2, %r35, %r36;
mov.u32 %r37, %ctaid.x;
shl.b32 %r3, %r37, 1;
shr.s32 %r38, %r2, 31;
shr.u32 %r39, %r38, 29;
add.s32 %r40, %r2, %r39;
shr.s32 %r904, %r40, 3;
shr.u32 %r5, %r34, 8;
setp.ge.s32 %p3, %r904, %r5;
mov.f32 %f157, 0f00000000;
mov.f32 %f158, %f157;
mov.f32 %f159, %f157;
mov.f32 %f160, %f157;
mov.f32 %f161, %f157;
mov.f32 %f162, %f157;
@%p3 bra $L__BB218_3;

mov.u32 %r41, %ctaid.y;
mov.u32 %r42, %ctaid.z;
mul.lo.s32 %r43, %r41, %r26;
cvt.u64.u32 %rd20, %r43;
mul.lo.s32 %r44, %r42, %r32;
cvt.u64.u32 %rd21, %r44;
add.s64 %rd2, %rd21, %rd20;
mul.hi.u32 %r45, %r42, %r28;
add.s32 %r46, %r45, %r42;
shr.u32 %r47, %r46, %r29;
mul.lo.s32 %r48, %r47, %r31;
mul.hi.u32 %r49, %r41, %r22;
add.s32 %r50, %r49, %r41;
shr.u32 %r51, %r50, %r23;
mul.lo.s32 %r52, %r51, %r25;
cvta.to.global.u64 %rd85, %rd17;
and.b32 %r56, %r40, -8;
sub.s32 %r57, %r2, %r56;
shl.b32 %r58, %r57, 2;
cvt.s64.s32 %rd4, %r57;
shl.b32 %r59, %r20, 1;
shl.b32 %r901, %r904, 3;
add.s32 %r903, %r901, %r59;
add.s32 %r902, %r20, %r901;
add.s32 %r60, %r904, %r48;
add.s32 %r61, %r60, %r52;
add.s32 %r62, %r3, 1;
mad.lo.s32 %r63, %r19, %r62, %r61;
mul.wide.s32 %rd5, %r63, 74;
add.s64 %rd22, %rd5, %rd4;
add.s64 %rd6, %rd22, 66;
mad.lo.s32 %r64, %r3, %r19, %r904;
add.s32 %r65, %r64, %r48;
add.s32 %r66, %r65, %r52;
mul.wide.s32 %rd7, %r66, 74;
add.s64 %rd23, %rd7, %rd4;
add.s64 %rd8, %rd23, 66;
mul.wide.s32 %rd24, %r58, 2;
add.s64 %rd25, %rd5, %rd24;
add.s64 %rd9, %rd25, 4;
add.s64 %rd26, %rd7, %rd24;
add.s64 %rd10, %rd26, 8;
cvta.to.global.u64 %rd11, %rd18;
mov.f32 %f157, 0f00000000;

$L__BB218_2:
cvt.u64.u32 %rd27, %r901;
add.s64 %rd28, %rd2, %rd27;
add.s64 %rd29, %rd28, %rd4;
mul.lo.s64 %rd30, %rd29, 36;
add.s64 %rd31, %rd85, %rd10;
ld.global.nc.u16 %rs7, [%rd31+-4];
ld.global.nc.u16 %rs8, [%rd31+-2];
ld.global.nc.u16 %rs9, [%rd31];
add.s64 %rd32, %rd85, %rd8;
ld.global.nc.u8 %rs10, [%rd32];
and.b16 %rs11, %rs10, 240;
and.b16 %rs12, %rs10, 15;
shr.u16 %rs13, %rs11, 4;
ld.global.nc.u16 %rs14, [%rd31+-6];
shl.b16 %rs15, %rs14, 3;
cvt.u64.u16 %rd33, %rs15;
and.b64 %rd34, %rd33, 4088;
mov.u64 %rd35, iq2xs_grid;
add.s64 %rd36, %rd35, %rd34;
ld.global.nc.v2.u32 {%r361, %r362}, [%rd36];
not.b32 %r364, %r361;
mov.u32 %r343, 0;
not.b32 %r366, %r362;
shr.u16 %rs16, %rs14, 9;
cvt.u32.u16 %r367, %rs16;
popc.b32 %r368, %r367;
shl.b32 %r369, %r368, 7;
and.b32 %r370, %r369, 128;
or.b32 %r371, %r370, %r367;
mul.lo.s32 %r372, %r371, 16843009;
and.b32 %r373, %r372, 134480385;
or.b32 %r374, %r373, -2139062144;
add.s32 %r375, %r374, -16843009;
or.b32 %r263, %r375, %r373;
mov.u32 %r67, %r263;

prmt.b32 %r67, %r67, 0, 0xba98;

xor.b32 %r376, %r67, %r361;
or.b32 %r377, %r376, -2139062144;
and.b32 %r378, %r67, 2139062143;
and.b32 %r379, %r364, -2139062144;
sub.s32 %r380, %r377, %r378;
xor.b32 %r72, %r379, %r380;
add.s64 %rd37, %rd11, %rd30;
ld.global.nc.u32 %r73, [%rd37+4];
and.b32 %r381, %r372, -2143281136;
or.b32 %r382, %r381, -2139062144;
add.s32 %r383, %r382, -16843009;
or.b32 %r265, %r383, %r381;
mov.u32 %r69, %r265;

prmt.b32 %r69, %r69, 0, 0xba98;

xor.b32 %r384, %r69, %r362;
or.b32 %r385, %r384, -2139062144;
and.b32 %r386, %r69, 2139062143;
and.b32 %r387, %r366, -2139062144;
sub.s32 %r388, %r385, %r386;
xor.b32 %r76, %r387, %r388;
ld.global.nc.u32 %r77, [%rd37+8];

dp4a.s32.s32 %r71, %r72, %r73, %r343;


dp4a.s32.s32 %r75, %r76, %r77, %r71;

shl.b16 %rs17, %rs7, 3;
cvt.u64.u16 %rd38, %rs17;
and.b64 %rd39, %rd38, 4088;
add.s64 %rd40, %rd35, %rd39;
ld.global.nc.v2.u32 {%r389, %r390}, [%rd40];
not.b32 %r392, %r389;
not.b32 %r394, %r390;
shr.u16 %rs18, %rs7, 9;
cvt.u32.u16 %r395, %rs18;
popc.b32 %r396, %r395;
shl.b32 %r397, %r396, 7;
and.b32 %r398, %r397, 128;
or.b32 %r399, %r398, %r395;
mul.lo.s32 %r400, %r399, 16843009;
and.b32 %r401, %r400, 134480385;
or.b32 %r402, %r401, -2139062144;
add.s32 %r403, %r402, -16843009;
or.b32 %r275, %r403, %r401;
mov.u32 %r79, %r275;

prmt.b32 %r79, %r79, 0, 0xba98;

xor.b32 %r404, %r79, %r389;
or.b32 %r405, %r404, -2139062144;
and.b32 %r406, %r79, 2139062143;
and.b32 %r407, %r392, -2139062144;
sub.s32 %r408, %r405, %r406;
xor.b32 %r84, %r407, %r408;
ld.global.nc.u32 %r85, [%rd37+12];
and.b32 %r409, %r400, -2143281136;
or.b32 %r410, %r409, -2139062144;
add.s32 %r411, %r410, -16843009;
or.b32 %r277, %r411, %r409;
mov.u32 %r81, %r277;

prmt.b32 %r81, %r81, 0, 0xba98;

xor.b32 %r412, %r81, %r390;
or.b32 %r413, %r412, -2139062144;
and.b32 %r414, %r81, 2139062143;
and.b32 %r415, %r394, -2139062144;
sub.s32 %r416, %r413, %r414;
xor.b32 %r88, %r415, %r416;
ld.global.nc.u32 %r89, [%rd37+16];

dp4a.s32.s32 %r83, %r84, %r85, %r75;


dp4a.s32.s32 %r87, %r88, %r89, %r83;

shl.b16 %rs19, %rs8, 3;
cvt.u64.u16 %rd41, %rs19;
and.b64 %rd42, %rd41, 4088;
add.s64 %rd43, %rd35, %rd42;
ld.global.nc.v2.u32 {%r417, %r418}, [%rd43];
not.b32 %r420, %r417;
not.b32 %r422, %r418;
shr.u16 %rs20, %rs8, 9;
cvt.u32.u16 %r423, %rs20;
popc.b32 %r424, %r423;
shl.b32 %r425, %r424, 7;
and.b32 %r426, %r425, 128;
or.b32 %r427, %r426, %r423;
mul.lo.s32 %r428, %r427, 16843009;
and.b32 %r429, %r428, 134480385;
or.b32 %r430, %r429, -2139062144;
add.s32 %r431, %r430, -16843009;
or.b32 %r287, %r431, %r429;
mov.u32 %r91, %r287;

prmt.b32 %r91, %r91, 0, 0xba98;

xor.b32 %r432, %r91, %r417;
or.b32 %r433, %r432, -2139062144;
and.b32 %r434, %r91, 2139062143;
and.b32 %r435, %r420, -2139062144;
sub.s32 %r436, %r433, %r434;
xor.b32 %r96, %r435, %r436;
ld.global.nc.u32 %r97, [%rd37+20];
and.b32 %r437, %r428, -2143281136;
or.b32 %r438, %r437, -2139062144;
add.s32 %r439, %r438, -16843009;
or.b32 %r289, %r439, %r437;
mov.u32 %r93, %r289;

prmt.b32 %r93, %r93, 0, 0xba98;

xor.b32 %r440, %r93, %r418;
or.b32 %r441, %r440, -2139062144;
and.b32 %r442, %r93, 2139062143;
and.b32 %r443, %r422, -2139062144;
sub.s32 %r444, %r441, %r442;
xor.b32 %r100, %r443, %r444;
ld.global.nc.u32 %r101, [%rd37+24];

dp4a.s32.s32 %r95, %r96, %r97, %r343;


dp4a.s32.s32 %r99, %r100, %r101, %r95;

shl.b16 %rs21, %rs9, 3;
cvt.u64.u16 %rd44, %rs21;
and.b64 %rd45, %rd44, 4088;
add.s64 %rd46, %rd35, %rd45;
ld.global.nc.v2.u32 {%r445, %r446}, [%rd46];
not.b32 %r448, %r445;
not.b32 %r450, %r446;
shr.u16 %rs22, %rs9, 9;
cvt.u32.u16 %r451, %rs22;
popc.b32 %r452, %r451;
shl.b32 %r453, %r452, 7;
and.b32 %r454, %r453, 128;
or.b32 %r455, %r454, %r451;
mul.lo.s32 %r456, %r455, 16843009;
and.b32 %r457, %r456, 134480385;
or.b32 %r458, %r457, -2139062144;
add.s32 %r459, %r458, -16843009;
or.b32 %r299, %r459, %r457;
mov.u32 %r103, %r299;

prmt.b32 %r103, %r103, 0, 0xba98;

xor.b32 %r460, %r103, %r445;
or.b32 %r461, %r460, -2139062144;
and.b32 %r462, %r103, 2139062143;
and.b32 %r463, %r448, -2139062144;
sub.s32 %r464, %r461, %r462;
xor.b32 %r108, %r463, %r464;
ld.global.nc.u32 %r109, [%rd37+28];
and.b32 %r465, %r456, -2143281136;
or.b32 %r466, %r465, -2139062144;
add.s32 %r467, %r466, -16843009;
or.b32 %r301, %r467, %r465;
mov.u32 %r105, %r301;

prmt.b32 %r105, %r105, 0, 0xba98;

xor.b32 %r468, %r105, %r446;
or.b32 %r469, %r468, -2139062144;
and.b32 %r470, %r105, 2139062143;
and.b32 %r471, %r450, -2139062144;
sub.s32 %r472, %r469, %r470;
xor.b32 %r112, %r471, %r472;
ld.global.nc.u32 %r113, [%rd37+32];

dp4a.s32.s32 %r107, %r108, %r109, %r99;


dp4a.s32.s32 %r111, %r112, %r113, %r107;

cvt.u32.u16 %r473, %rs12;
cvt.u32.u16 %r474, %rs13;
mul.lo.s32 %r475, %r87, %r473;
mad.lo.s32 %r476, %r111, %r474, %r475;
add.s32 %r477, %r111, %r87;
shr.u32 %r478, %r477, 31;
add.s32 %r479, %r477, %r478;
shr.s32 %r480, %r479, 1;
add.s32 %r481, %r476, %r480;
shr.s32 %r482, %r481, 31;
shr.u32 %r483, %r482, 30;
add.s32 %r484, %r481, %r483;
shr.s32 %r485, %r484, 2;
add.s64 %rd47, %rd85, %rd7;
ld.global.nc.u16 %rs1, [%rd47];

{ cvt.f32.f16 %f37, %rs1;}


ld.global.nc.u32 %r115, [%rd37];

{.reg .f16 low,high;
mov.b32 {low,high},%r115;
cvt.f32.f16 %f38, low;}


mul.ftz.f32 %f49, %f37, %f38;
cvt.rn.f32.s32 %f50, %r485;
fma.rn.ftz.f32 %f162, %f49, %f50, %f162;
add.s64 %rd48, %rd85, %rd9;
ld.global.nc.u16 %rs23, [%rd48];
ld.global.nc.u16 %rs24, [%rd48+2];
ld.global.nc.u16 %rs25, [%rd48+4];
add.s64 %rd49, %rd85, %rd6;
ld.global.nc.u8 %rs26, [%rd49];
and.b16 %rs27, %rs26, 240;
and.b16 %rs28, %rs26, 15;
shr.u16 %rs29, %rs27, 4;
ld.global.nc.u16 %rs30, [%rd48+-2];
shl.b16 %rs31, %rs30, 3;
cvt.u64.u16 %rd50, %rs31;
and.b64 %rd51, %rd50, 4088;
add.s64 %rd52, %rd35, %rd51;
ld.global.nc.v2.u32 {%r486, %r487}, [%rd52];
not.b32 %r489, %r486;
not.b32 %r491, %r487;
shr.u16 %rs32, %rs30, 9;
cvt.u32.u16 %r492, %rs32;
popc.b32 %r493, %r492;
shl.b32 %r494, %r493, 7;
and.b32 %r495, %r494, 128;
or.b32 %r496, %r495, %r492;
mul.lo.s32 %r497, %r496, 16843009;
and.b32 %r498, %r497, 134480385;
or.b32 %r499, %r498, -2139062144;
add.s32 %r500, %r499, -16843009;
or.b32 %r312, %r500, %r498;
mov.u32 %r116, %r312;

prmt.b32 %r116, %r116, 0, 0xba98;

xor.b32 %r501, %r116, %r486;
or.b32 %r502, %r501, -2139062144;
and.b32 %r503, %r116, 2139062143;
and.b32 %r504, %r489, -2139062144;
sub.s32 %r505, %r502, %r503;
xor.b32 %r121, %r504, %r505;
and.b32 %r506, %r497, -2143281136;
or.b32 %r507, %r506, -2139062144;
add.s32 %r508, %r507, -16843009;
or.b32 %r314, %r508, %r506;
mov.u32 %r118, %r314;

prmt.b32 %r118, %r118, 0, 0xba98;

xor.b32 %r509, %r118, %r487;
or.b32 %r510, %r509, -2139062144;
and.b32 %r511, %r118, 2139062143;
and.b32 %r512, %r491, -2139062144;
sub.s32 %r513, %r510, %r511;
xor.b32 %r125, %r512, %r513;

dp4a.s32.s32 %r120, %r121, %r73, %r343;


dp4a.s32.s32 %r124, %r125, %r77, %r120;

shl.b16 %rs33, %rs23, 3;
cvt.u64.u16 %rd53, %rs33;
and.b64 %rd54, %rd53, 4088;
add.s64 %rd55, %rd35, %rd54;
ld.global.nc.v2.u32 {%r514, %r515}, [%rd55];
not.b32 %r517, %r514;
not.b32 %r519, %r515;
shr.u16 %rs34, %rs23, 9;
cvt.u32.u16 %r520, %rs34;
popc.b32 %r521, %r520;
shl.b32 %r522, %r521, 7;
and.b32 %r523, %r522, 128;
or.b32 %r524, %r523, %r520;
mul.lo.s32 %r525, %r524, 16843009;
and.b32 %r526, %r525, 134480385;
or.b32 %r527, %r526, -2139062144;
add.s32 %r528, %r527, -16843009;
or.b32 %r324, %r528, %r526;
mov.u32 %r128, %r324;

prmt.b32 %r128, %r128, 0, 0xba98;

xor.b32 %r529, %r128, %r514;
or.b32 %r530, %r529, -2139062144;
and.b32 %r531, %r128, 2139062143;
and.b32 %r532, %r517, -2139062144;
sub.s32 %r533, %r530, %r531;
xor.b32 %r133, %r532, %r533;
and.b32 %r534, %r525, -2143281136;
or.b32 %r535, %r534, -2139062144;
add.s32 %r536, %r535, -16843009;
or.b32 %r326, %r536, %r534;
mov.u32 %r130, %r326;

prmt.b32 %r130, %r130, 0, 0xba98;

xor.b32 %r537, %r130, %r515;
or.b32 %r538, %r537, -2139062144;
and.b32 %r539, %r130, 2139062143;
and.b32 %r540, %r519, -2139062144;
sub.s32 %r541, %r538, %r539;
xor.b32 %r137, %r540, %r541;

dp4a.s32.s32 %r132, %r133, %r85, %r124;


dp4a.s32.s32 %r136, %r137, %r89, %r132;

shl.b16 %rs35, %rs24, 3;
cvt.u64.u16 %rd56, %rs35;
and.b64 %rd57, %rd56, 4088;
add.s64 %rd58, %rd35, %rd57;
ld.global.nc.v2.u32 {%r542, %r543}, [%rd58];
not.b32 %r545, %r542;
not.b32 %r547, %r543;
shr.u16 %rs36, %rs24, 9;
cvt.u32.u16 %r548, %rs36;
popc.b32 %r549, %r548;
shl.b32 %r550, %r549, 7;
and.b32 %r551, %r550, 128;
or.b32 %r552, %r551, %r548;
mul.lo.s32 %r553, %r552, 16843009;
and.b32 %r554, %r553, 134480385;
or.b32 %r555, %r554, -2139062144;
add.s32 %r556, %r555, -16843009;
or.b32 %r336, %r556, %r554;
mov.u32 %r140, %r336;

prmt.b32 %r140, %r140, 0, 0xba98;

xor.b32 %r557, %r140, %r542;
or.b32 %r558, %r557, -2139062144;
and.b32 %r559, %r140, 2139062143;
and.b32 %r560, %r545, -2139062144;
sub.s32 %r561, %r558, %r559;
xor.b32 %r145, %r560, %r561;
and.b32 %r562, %r553, -2143281136;
or.b32 %r563, %r562, -2139062144;
add.s32 %r564, %r563, -16843009;
or.b32 %r338, %r564, %r562;
mov.u32 %r142, %r338;

prmt.b32 %r142, %r142, 0, 0xba98;

xor.b32 %r565, %r142, %r543;
or.b32 %r566, %r565, -2139062144;
and.b32 %r567, %r142, 2139062143;
and.b32 %r568, %r547, -2139062144;
sub.s32 %r569, %r566, %r567;
xor.b32 %r149, %r568, %r569;

dp4a.s32.s32 %r144, %r145, %r97, %r343;


dp4a.s32.s32 %r148, %r149, %r101, %r144;

shl.b16 %rs37, %rs25, 3;
cvt.u64.u16 %rd59, %rs37;
and.b64 %rd60, %rd59, 4088;
add.s64 %rd61, %rd35, %rd60;
ld.global.nc.v2.u32 {%r570, %r571}, [%rd61];
not.b32 %r573, %r570;
not.b32 %r575, %r571;
shr.u16 %rs38, %rs25, 9;
cvt.u32.u16 %r576, %rs38;
popc.b32 %r577, %r576;
shl.b32 %r578, %r577, 7;
and.b32 %r579, %r578, 128;
or.b32 %r580, %r579, %r576;
mul.lo.s32 %r581, %r580, 16843009;
and.b32 %r582, %r581, 134480385;
or.b32 %r583, %r582, -2139062144;
add.s32 %r584, %r583, -16843009;
or.b32 %r348, %r584, %r582;
mov.u32 %r152, %r348;

prmt.b32 %r152, %r152, 0, 0xba98;

xor.b32 %r585, %r152, %r570;
or.b32 %r586, %r585, -2139062144;
and.b32 %r587, %r152, 2139062143;
and.b32 %r588, %r573, -2139062144;
sub.s32 %r589, %r586, %r587;
xor.b32 %r157, %r588, %r589;
and.b32 %r590, %r581, -2143281136;
or.b32 %r591, %r590, -2139062144;
add.s32 %r592, %r591, -16843009;
or.b32 %r350, %r592, %r590;
mov.u32 %r154, %r350;

prmt.b32 %r154, %r154, 0, 0xba98;

xor.b32 %r593, %r154, %r571;
or.b32 %r594, %r593, -2139062144;
and.b32 %r595, %r154, 2139062143;
and.b32 %r596, %r575, -2139062144;
sub.s32 %r597, %r594, %r595;
xor.b32 %r161, %r596, %r597;

dp4a.s32.s32 %r156, %r157, %r109, %r148;


dp4a.s32.s32 %r160, %r161, %r113, %r156;

cvt.u32.u16 %r598, %rs28;
cvt.u32.u16 %r599, %rs29;
mul.lo.s32 %r600, %r136, %r598;
mad.lo.s32 %r601, %r160, %r599, %r600;
add.s32 %r602, %r160, %r136;
shr.u32 %r603, %r602, 31;
add.s32 %r604, %r602, %r603;
shr.s32 %r605, %r604, 1;
add.s32 %r606, %r601, %r605;
shr.s32 %r607, %r606, 31;
shr.u32 %r608, %r607, 30;
add.s32 %r609, %r606, %r608;
shr.s32 %r610, %r609, 2;
add.s64 %rd62, %rd85, %rd5;
ld.global.nc.u16 %rs2, [%rd62];

{ cvt.f32.f16 %f39, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r115;
cvt.f32.f16 %f40, low;}


mul.ftz.f32 %f51, %f39, %f40;
cvt.rn.f32.s32 %f52, %r610;
fma.rn.ftz.f32 %f161, %f51, %f52, %f161;
cvt.u64.u32 %rd63, %r902;
add.s64 %rd64, %rd2, %rd63;
add.s64 %rd65, %rd64, %rd4;
mul.lo.s64 %rd66, %rd65, 36;
mov.u32 %r165, %r263;

prmt.b32 %r165, %r165, 0, 0xba98;

xor.b32 %r611, %r165, %r361;
or.b32 %r612, %r611, -2139062144;
and.b32 %r613, %r165, 2139062143;
sub.s32 %r614, %r612, %r613;
xor.b32 %r170, %r379, %r614;
add.s64 %rd67, %rd11, %rd66;
ld.global.nc.u32 %r171, [%rd67+4];
mov.u32 %r167, %r265;

prmt.b32 %r167, %r167, 0, 0xba98;

xor.b32 %r615, %r167, %r362;
or.b32 %r616, %r615, -2139062144;
and.b32 %r617, %r167, 2139062143;
sub.s32 %r618, %r616, %r617;
xor.b32 %r174, %r387, %r618;
ld.global.nc.u32 %r175, [%rd67+8];

dp4a.s32.s32 %r169, %r170, %r171, %r343;


dp4a.s32.s32 %r173, %r174, %r175, %r169;

mov.u32 %r177, %r275;

prmt.b32 %r177, %r177, 0, 0xba98;

xor.b32 %r619, %r177, %r389;
or.b32 %r620, %r619, -2139062144;
and.b32 %r621, %r177, 2139062143;
sub.s32 %r622, %r620, %r621;
xor.b32 %r182, %r407, %r622;
ld.global.nc.u32 %r183, [%rd67+12];
mov.u32 %r179, %r277;

prmt.b32 %r179, %r179, 0, 0xba98;

xor.b32 %r623, %r179, %r390;
or.b32 %r624, %r623, -2139062144;
and.b32 %r625, %r179, 2139062143;
sub.s32 %r626, %r624, %r625;
xor.b32 %r186, %r415, %r626;
ld.global.nc.u32 %r187, [%rd67+16];

dp4a.s32.s32 %r181, %r182, %r183, %r173;


dp4a.s32.s32 %r185, %r186, %r187, %r181;

mov.u32 %r189, %r287;

prmt.b32 %r189, %r189, 0, 0xba98;

xor.b32 %r627, %r189, %r417;
or.b32 %r628, %r627, -2139062144;
and.b32 %r629, %r189, 2139062143;
sub.s32 %r630, %r628, %r629;
xor.b32 %r194, %r435, %r630;
ld.global.nc.u32 %r195, [%rd67+20];
mov.u32 %r191, %r289;

prmt.b32 %r191, %r191, 0, 0xba98;

xor.b32 %r631, %r191, %r418;
or.b32 %r632, %r631, -2139062144;
and.b32 %r633, %r191, 2139062143;
sub.s32 %r634, %r632, %r633;
xor.b32 %r198, %r443, %r634;
ld.global.nc.u32 %r199, [%rd67+24];

dp4a.s32.s32 %r193, %r194, %r195, %r343;


dp4a.s32.s32 %r197, %r198, %r199, %r193;

mov.u32 %r201, %r299;

prmt.b32 %r201, %r201, 0, 0xba98;

xor.b32 %r635, %r201, %r445;
or.b32 %r636, %r635, -2139062144;
and.b32 %r637, %r201, 2139062143;
sub.s32 %r638, %r636, %r637;
xor.b32 %r206, %r463, %r638;
ld.global.nc.u32 %r207, [%rd67+28];
mov.u32 %r203, %r301;

prmt.b32 %r203, %r203, 0, 0xba98;

xor.b32 %r639, %r203, %r446;
or.b32 %r640, %r639, -2139062144;
and.b32 %r641, %r203, 2139062143;
sub.s32 %r642, %r640, %r641;
xor.b32 %r210, %r471, %r642;
ld.global.nc.u32 %r211, [%rd67+32];

dp4a.s32.s32 %r205, %r206, %r207, %r197;


dp4a.s32.s32 %r209, %r210, %r211, %r205;

mul.lo.s32 %r643, %r185, %r473;
mad.lo.s32 %r644, %r209, %r474, %r643;
add.s32 %r645, %r209, %r185;
shr.u32 %r646, %r645, 31;
add.s32 %r647, %r645, %r646;
shr.s32 %r648, %r647, 1;
add.s32 %r649, %r644, %r648;
shr.s32 %r650, %r649, 31;
shr.u32 %r651, %r650, 30;
add.s32 %r652, %r649, %r651;
shr.s32 %r653, %r652, 2;

{ cvt.f32.f16 %f41, %rs1;}


ld.global.nc.u32 %r213, [%rd67];

{.reg .f16 low,high;
mov.b32 {low,high},%r213;
cvt.f32.f16 %f42, low;}


mul.ftz.f32 %f53, %f41, %f42;
cvt.rn.f32.s32 %f54, %r653;
fma.rn.ftz.f32 %f160, %f53, %f54, %f160;
mov.u32 %r214, %r312;

prmt.b32 %r214, %r214, 0, 0xba98;

xor.b32 %r654, %r214, %r486;
or.b32 %r655, %r654, -2139062144;
and.b32 %r656, %r214, 2139062143;
sub.s32 %r657, %r655, %r656;
xor.b32 %r219, %r504, %r657;
mov.u32 %r216, %r314;

prmt.b32 %r216, %r216, 0, 0xba98;

xor.b32 %r658, %r216, %r487;
or.b32 %r659, %r658, -2139062144;
and.b32 %r660, %r216, 2139062143;
sub.s32 %r661, %r659, %r660;
xor.b32 %r223, %r512, %r661;

dp4a.s32.s32 %r218, %r219, %r171, %r343;


dp4a.s32.s32 %r222, %r223, %r175, %r218;

mov.u32 %r226, %r324;

prmt.b32 %r226, %r226, 0, 0xba98;

xor.b32 %r662, %r226, %r514;
or.b32 %r663, %r662, -2139062144;
and.b32 %r664, %r226, 2139062143;
sub.s32 %r665, %r663, %r664;
xor.b32 %r231, %r532, %r665;
mov.u32 %r228, %r326;

prmt.b32 %r228, %r228, 0, 0xba98;

xor.b32 %r666, %r228, %r515;
or.b32 %r667, %r666, -2139062144;
and.b32 %r668, %r228, 2139062143;
sub.s32 %r669, %r667, %r668;
xor.b32 %r235, %r540, %r669;

dp4a.s32.s32 %r230, %r231, %r183, %r222;


dp4a.s32.s32 %r234, %r235, %r187, %r230;

mov.u32 %r238, %r336;

prmt.b32 %r238, %r238, 0, 0xba98;

xor.b32 %r670, %r238, %r542;
or.b32 %r671, %r670, -2139062144;
and.b32 %r672, %r238, 2139062143;
sub.s32 %r673, %r671, %r672;
xor.b32 %r243, %r560, %r673;
mov.u32 %r240, %r338;

prmt.b32 %r240, %r240, 0, 0xba98;

xor.b32 %r674, %r240, %r543;
or.b32 %r675, %r674, -2139062144;
and.b32 %r676, %r240, 2139062143;
sub.s32 %r677, %r675, %r676;
xor.b32 %r247, %r568, %r677;

dp4a.s32.s32 %r242, %r243, %r195, %r343;


dp4a.s32.s32 %r246, %r247, %r199, %r242;

mov.u32 %r250, %r348;

prmt.b32 %r250, %r250, 0, 0xba98;

xor.b32 %r678, %r250, %r570;
or.b32 %r679, %r678, -2139062144;
and.b32 %r680, %r250, 2139062143;
sub.s32 %r681, %r679, %r680;
xor.b32 %r255, %r588, %r681;
mov.u32 %r252, %r350;

prmt.b32 %r252, %r252, 0, 0xba98;

xor.b32 %r682, %r252, %r571;
or.b32 %r683, %r682, -2139062144;
and.b32 %r684, %r252, 2139062143;
sub.s32 %r685, %r683, %r684;
xor.b32 %r259, %r596, %r685;

dp4a.s32.s32 %r254, %r255, %r207, %r246;


dp4a.s32.s32 %r258, %r259, %r211, %r254;

mul.lo.s32 %r686, %r234, %r598;
mad.lo.s32 %r687, %r258, %r599, %r686;
add.s32 %r688, %r258, %r234;
shr.u32 %r689, %r688, 31;
add.s32 %r690, %r688, %r689;
shr.s32 %r691, %r690, 1;
add.s32 %r692, %r687, %r691;
shr.s32 %r693, %r692, 31;
shr.u32 %r694, %r693, 30;
add.s32 %r695, %r692, %r694;
shr.s32 %r696, %r695, 2;

{ cvt.f32.f16 %f43, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r213;
cvt.f32.f16 %f44, low;}


mul.ftz.f32 %f55, %f43, %f44;
cvt.rn.f32.s32 %f56, %r696;
fma.rn.ftz.f32 %f159, %f55, %f56, %f159;
cvt.u64.u32 %rd68, %r903;
add.s64 %rd69, %rd2, %rd68;
add.s64 %rd70, %rd69, %rd4;
mul.lo.s64 %rd71, %rd70, 36;

prmt.b32 %r263, %r263, 0, 0xba98;

xor.b32 %r697, %r263, %r361;
or.b32 %r698, %r697, -2139062144;
and.b32 %r699, %r263, 2139062143;
sub.s32 %r700, %r698, %r699;
xor.b32 %r268, %r379, %r700;
add.s64 %rd72, %rd11, %rd71;
ld.global.nc.u32 %r269, [%rd72+4];

prmt.b32 %r265, %r265, 0, 0xba98;

xor.b32 %r701, %r265, %r362;
or.b32 %r702, %r701, -2139062144;
and.b32 %r703, %r265, 2139062143;
sub.s32 %r704, %r702, %r703;
xor.b32 %r272, %r387, %r704;
ld.global.nc.u32 %r273, [%rd72+8];

dp4a.s32.s32 %r267, %r268, %r269, %r343;


dp4a.s32.s32 %r271, %r272, %r273, %r267;


prmt.b32 %r275, %r275, 0, 0xba98;

xor.b32 %r705, %r275, %r389;
or.b32 %r706, %r705, -2139062144;
and.b32 %r707, %r275, 2139062143;
sub.s32 %r708, %r706, %r707;
xor.b32 %r280, %r407, %r708;
ld.global.nc.u32 %r281, [%rd72+12];

prmt.b32 %r277, %r277, 0, 0xba98;

xor.b32 %r709, %r277, %r390;
or.b32 %r710, %r709, -2139062144;
and.b32 %r711, %r277, 2139062143;
sub.s32 %r712, %r710, %r711;
xor.b32 %r284, %r415, %r712;
ld.global.nc.u32 %r285, [%rd72+16];

dp4a.s32.s32 %r279, %r280, %r281, %r271;


dp4a.s32.s32 %r283, %r284, %r285, %r279;


prmt.b32 %r287, %r287, 0, 0xba98;

xor.b32 %r713, %r287, %r417;
or.b32 %r714, %r713, -2139062144;
and.b32 %r715, %r287, 2139062143;
sub.s32 %r716, %r714, %r715;
xor.b32 %r292, %r435, %r716;
ld.global.nc.u32 %r293, [%rd72+20];

prmt.b32 %r289, %r289, 0, 0xba98;

xor.b32 %r717, %r289, %r418;
or.b32 %r718, %r717, -2139062144;
and.b32 %r719, %r289, 2139062143;
sub.s32 %r720, %r718, %r719;
xor.b32 %r296, %r443, %r720;
ld.global.nc.u32 %r297, [%rd72+24];

dp4a.s32.s32 %r291, %r292, %r293, %r343;


dp4a.s32.s32 %r295, %r296, %r297, %r291;


prmt.b32 %r299, %r299, 0, 0xba98;

xor.b32 %r721, %r299, %r445;
or.b32 %r722, %r721, -2139062144;
and.b32 %r723, %r299, 2139062143;
sub.s32 %r724, %r722, %r723;
xor.b32 %r304, %r463, %r724;
ld.global.nc.u32 %r305, [%rd72+28];

prmt.b32 %r301, %r301, 0, 0xba98;

xor.b32 %r725, %r301, %r446;
or.b32 %r726, %r725, -2139062144;
and.b32 %r727, %r301, 2139062143;
sub.s32 %r728, %r726, %r727;
xor.b32 %r308, %r471, %r728;
ld.global.nc.u32 %r309, [%rd72+32];

dp4a.s32.s32 %r303, %r304, %r305, %r295;


dp4a.s32.s32 %r307, %r308, %r309, %r303;

mul.lo.s32 %r729, %r283, %r473;
mad.lo.s32 %r730, %r307, %r474, %r729;
add.s32 %r731, %r307, %r283;
shr.u32 %r732, %r731, 31;
add.s32 %r733, %r731, %r732;
shr.s32 %r734, %r733, 1;
add.s32 %r735, %r730, %r734;
shr.s32 %r736, %r735, 31;
shr.u32 %r737, %r736, 30;
add.s32 %r738, %r735, %r737;
shr.s32 %r739, %r738, 2;

{ cvt.f32.f16 %f45, %rs1;}


ld.global.nc.u32 %r311, [%rd72];

{.reg .f16 low,high;
mov.b32 {low,high},%r311;
cvt.f32.f16 %f46, low;}


mul.ftz.f32 %f57, %f45, %f46;
cvt.rn.f32.s32 %f58, %r739;
fma.rn.ftz.f32 %f158, %f57, %f58, %f158;

prmt.b32 %r312, %r312, 0, 0xba98;

xor.b32 %r740, %r312, %r486;
or.b32 %r741, %r740, -2139062144;
and.b32 %r742, %r312, 2139062143;
sub.s32 %r743, %r741, %r742;
xor.b32 %r317, %r504, %r743;

prmt.b32 %r314, %r314, 0, 0xba98;

xor.b32 %r744, %r314, %r487;
or.b32 %r745, %r744, -2139062144;
and.b32 %r746, %r314, 2139062143;
sub.s32 %r747, %r745, %r746;
xor.b32 %r321, %r512, %r747;

dp4a.s32.s32 %r316, %r317, %r269, %r343;


dp4a.s32.s32 %r320, %r321, %r273, %r316;


prmt.b32 %r324, %r324, 0, 0xba98;

xor.b32 %r748, %r324, %r514;
or.b32 %r749, %r748, -2139062144;
and.b32 %r750, %r324, 2139062143;
sub.s32 %r751, %r749, %r750;
xor.b32 %r329, %r532, %r751;

prmt.b32 %r326, %r326, 0, 0xba98;

xor.b32 %r752, %r326, %r515;
or.b32 %r753, %r752, -2139062144;
and.b32 %r754, %r326, 2139062143;
sub.s32 %r755, %r753, %r754;
xor.b32 %r333, %r540, %r755;

dp4a.s32.s32 %r328, %r329, %r281, %r320;


dp4a.s32.s32 %r332, %r333, %r285, %r328;


prmt.b32 %r336, %r336, 0, 0xba98;

xor.b32 %r756, %r336, %r542;
or.b32 %r757, %r756, -2139062144;
and.b32 %r758, %r336, 2139062143;
sub.s32 %r759, %r757, %r758;
xor.b32 %r341, %r560, %r759;

prmt.b32 %r338, %r338, 0, 0xba98;

xor.b32 %r760, %r338, %r543;
or.b32 %r761, %r760, -2139062144;
and.b32 %r762, %r338, 2139062143;
sub.s32 %r763, %r761, %r762;
xor.b32 %r345, %r568, %r763;

dp4a.s32.s32 %r340, %r341, %r293, %r343;


dp4a.s32.s32 %r344, %r345, %r297, %r340;


prmt.b32 %r348, %r348, 0, 0xba98;

xor.b32 %r764, %r348, %r570;
or.b32 %r765, %r764, -2139062144;
and.b32 %r766, %r348, 2139062143;
sub.s32 %r767, %r765, %r766;
xor.b32 %r353, %r588, %r767;

prmt.b32 %r350, %r350, 0, 0xba98;

xor.b32 %r768, %r350, %r571;
or.b32 %r769, %r768, -2139062144;
and.b32 %r770, %r350, 2139062143;
sub.s32 %r771, %r769, %r770;
xor.b32 %r357, %r596, %r771;

dp4a.s32.s32 %r352, %r353, %r305, %r344;


dp4a.s32.s32 %r356, %r357, %r309, %r352;

mul.lo.s32 %r772, %r332, %r598;
mad.lo.s32 %r773, %r356, %r599, %r772;
add.s32 %r774, %r356, %r332;
shr.u32 %r775, %r774, 31;
add.s32 %r776, %r774, %r775;
shr.s32 %r777, %r776, 1;
add.s32 %r778, %r773, %r777;
shr.s32 %r779, %r778, 31;
shr.u32 %r780, %r779, 30;
add.s32 %r781, %r778, %r780;
shr.s32 %r782, %r781, 2;

{ cvt.f32.f16 %f47, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r311;
cvt.f32.f16 %f48, low;}


mul.ftz.f32 %f59, %f47, %f48;
cvt.rn.f32.s32 %f60, %r782;
fma.rn.ftz.f32 %f157, %f59, %f60, %f157;
add.s32 %r903, %r903, 128;
add.s32 %r902, %r902, 128;
add.s32 %r901, %r901, 128;
add.s64 %rd85, %rd85, 1184;
add.s32 %r904, %r904, 16;
setp.lt.s32 %p4, %r904, %r5;
@%p4 bra $L__BB218_2;

$L__BB218_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB218_5;

mov.u32 %r784, _ZZ13mul_mat_vec_qIL9ggml_type17ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r785, %r1, 768, %r784;
shl.b32 %r786, %r36, 2;
add.s32 %r787, %r785, %r786;
st.shared.f32 [%r787+-768], %f162;
st.shared.f32 [%r787+-640], %f161;
st.shared.f32 [%r787+-512], %f160;
st.shared.f32 [%r787+-384], %f159;
st.shared.f32 [%r787+-256], %f158;
st.shared.f32 [%r787+-128], %f157;

$L__BB218_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB218_18;

mov.u32 %r788, %ctaid.z;
mov.u32 %r789, %ctaid.y;
mad.lo.s32 %r790, %r789, %r27, %r3;
mad.lo.s32 %r791, %r788, %r33, %r790;
cvt.u64.u32 %rd14, %r791;
shl.b32 %r793, %r36, 2;
mov.u32 %r794, _ZZ13mul_mat_vec_qIL9ggml_type17ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r17, %r794, %r793;
mov.u32 %r795, 2;
ld.shared.f32 %f61, [%r17];
add.ftz.f32 %f62, %f61, %f162;
ld.shared.f32 %f63, [%r17+768];
add.ftz.f32 %f64, %f63, %f62;
ld.shared.f32 %f65, [%r17+1536];
add.ftz.f32 %f66, %f65, %f64;
mov.b32 %r796, %f66;
mov.u32 %r797, 31;
mov.u32 %r798, 16;
mov.u32 %r799, -1;
shfl.sync.bfly.b32 %r800|%p7, %r796, %r798, %r797, %r799;
mov.b32 %f67, %r800;
add.ftz.f32 %f68, %f66, %f67;
mov.b32 %r801, %f68;
mov.u32 %r802, 8;
shfl.sync.bfly.b32 %r803|%p8, %r801, %r802, %r797, %r799;
mov.b32 %f69, %r803;
add.ftz.f32 %f70, %f68, %f69;
mov.b32 %r804, %f70;
mov.u32 %r805, 4;
shfl.sync.bfly.b32 %r806|%p9, %r804, %r805, %r797, %r799;
mov.b32 %f71, %r806;
add.ftz.f32 %f72, %f70, %f71;
mov.b32 %r807, %f72;
shfl.sync.bfly.b32 %r808|%p10, %r807, %r795, %r797, %r799;
mov.b32 %f73, %r808;
add.ftz.f32 %f74, %f72, %f73;
mov.b32 %r809, %f74;
mov.u32 %r810, 1;
shfl.sync.bfly.b32 %r811|%p11, %r809, %r810, %r797, %r799;
mov.b32 %f75, %r811;
add.ftz.f32 %f19, %f74, %f75;
setp.eq.s32 %p12, %r36, 0;
setp.lt.u32 %p13, %r3, %r21;
and.pred %p1, %p12, %p13;
mul.wide.u32 %rd73, %r791, 4;
add.s64 %rd15, %rd1, %rd73;
not.pred %p14, %p1;
@%p14 bra $L__BB218_8;

st.global.f32 [%rd15], %f19;

$L__BB218_8:
ld.shared.f32 %f76, [%r17+128];
add.ftz.f32 %f77, %f76, %f161;
ld.shared.f32 %f78, [%r17+896];
add.ftz.f32 %f79, %f78, %f77;
ld.shared.f32 %f80, [%r17+1664];
add.ftz.f32 %f81, %f80, %f79;
mov.b32 %r812, %f81;
shfl.sync.bfly.b32 %r816|%p15, %r812, %r798, %r797, %r799;
mov.b32 %f82, %r816;
add.ftz.f32 %f83, %f81, %f82;
mov.b32 %r817, %f83;
shfl.sync.bfly.b32 %r819|%p16, %r817, %r802, %r797, %r799;
mov.b32 %f84, %r819;
add.ftz.f32 %f85, %f83, %f84;
mov.b32 %r820, %f85;
shfl.sync.bfly.b32 %r822|%p17, %r820, %r805, %r797, %r799;
mov.b32 %f86, %r822;
add.ftz.f32 %f87, %f85, %f86;
mov.b32 %r823, %f87;
shfl.sync.bfly.b32 %r825|%p18, %r823, %r795, %r797, %r799;
mov.b32 %f88, %r825;
add.ftz.f32 %f89, %f87, %f88;
mov.b32 %r826, %f89;
shfl.sync.bfly.b32 %r828|%p19, %r826, %r810, %r797, %r799;
mov.b32 %f90, %r828;
add.ftz.f32 %f20, %f89, %f90;
setp.eq.s32 %p20, %r36, 1;
add.s32 %r830, %r3, 1;
setp.lt.u32 %p21, %r830, %r21;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB218_10;

st.global.f32 [%rd15+4], %f20;

$L__BB218_10:
ld.shared.f32 %f91, [%r17+256];
add.ftz.f32 %f92, %f91, %f160;
ld.shared.f32 %f93, [%r17+1024];
add.ftz.f32 %f94, %f93, %f92;
ld.shared.f32 %f95, [%r17+1792];
add.ftz.f32 %f96, %f95, %f94;
mov.b32 %r831, %f96;
mov.u32 %r832, 31;
mov.u32 %r833, 16;
mov.u32 %r834, -1;
shfl.sync.bfly.b32 %r835|%p23, %r831, %r833, %r832, %r834;
mov.b32 %f97, %r835;
add.ftz.f32 %f98, %f96, %f97;
mov.b32 %r836, %f98;
mov.u32 %r837, 8;
shfl.sync.bfly.b32 %r838|%p24, %r836, %r837, %r832, %r834;
mov.b32 %f99, %r838;
add.ftz.f32 %f100, %f98, %f99;
mov.b32 %r839, %f100;
mov.u32 %r840, 4;
shfl.sync.bfly.b32 %r841|%p25, %r839, %r840, %r832, %r834;
mov.b32 %f101, %r841;
add.ftz.f32 %f102, %f100, %f101;
mov.b32 %r842, %f102;
mov.u32 %r843, 2;
shfl.sync.bfly.b32 %r844|%p26, %r842, %r843, %r832, %r834;
mov.b32 %f103, %r844;
add.ftz.f32 %f104, %f102, %f103;
mov.b32 %r845, %f104;
mov.u32 %r846, 1;
shfl.sync.bfly.b32 %r847|%p27, %r845, %r846, %r832, %r834;
mov.b32 %f105, %r847;
add.ftz.f32 %f21, %f104, %f105;
@%p14 bra $L__BB218_12;

cvt.u64.u32 %rd74, %r21;
add.s64 %rd75, %rd74, %rd14;
shl.b64 %rd76, %rd75, 2;
add.s64 %rd77, %rd1, %rd76;
st.global.f32 [%rd77], %f21;

$L__BB218_12:
ld.shared.f32 %f106, [%r17+384];
add.ftz.f32 %f107, %f106, %f159;
ld.shared.f32 %f108, [%r17+1152];
add.ftz.f32 %f109, %f108, %f107;
ld.shared.f32 %f110, [%r17+1920];
add.ftz.f32 %f111, %f110, %f109;
mov.b32 %r848, %f111;
shfl.sync.bfly.b32 %r852|%p29, %r848, %r833, %r832, %r834;
mov.b32 %f112, %r852;
add.ftz.f32 %f113, %f111, %f112;
mov.b32 %r853, %f113;
shfl.sync.bfly.b32 %r855|%p30, %r853, %r837, %r832, %r834;
mov.b32 %f114, %r855;
add.ftz.f32 %f115, %f113, %f114;
mov.b32 %r856, %f115;
shfl.sync.bfly.b32 %r858|%p31, %r856, %r840, %r832, %r834;
mov.b32 %f116, %r858;
add.ftz.f32 %f117, %f115, %f116;
mov.b32 %r859, %f117;
shfl.sync.bfly.b32 %r861|%p32, %r859, %r843, %r832, %r834;
mov.b32 %f118, %r861;
add.ftz.f32 %f119, %f117, %f118;
mov.b32 %r862, %f119;
shfl.sync.bfly.b32 %r864|%p33, %r862, %r846, %r832, %r834;
mov.b32 %f120, %r864;
add.ftz.f32 %f22, %f119, %f120;
@%p22 bra $L__BB218_14;

add.s32 %r865, %r21, 1;
cvt.u64.u32 %rd78, %r865;
add.s64 %rd79, %rd78, %rd14;
shl.b64 %rd80, %rd79, 2;
add.s64 %rd81, %rd1, %rd80;
st.global.f32 [%rd81], %f22;

$L__BB218_14:
shl.b32 %r866, %r21, 1;
mov.u32 %r867, 1;
ld.shared.f32 %f121, [%r17+512];
add.ftz.f32 %f122, %f121, %f158;
ld.shared.f32 %f123, [%r17+1280];
add.ftz.f32 %f124, %f123, %f122;
ld.shared.f32 %f125, [%r17+2048];
add.ftz.f32 %f126, %f125, %f124;
mov.b32 %r868, %f126;
mov.u32 %r869, 31;
mov.u32 %r870, 16;
mov.u32 %r871, -1;
shfl.sync.bfly.b32 %r872|%p35, %r868, %r870, %r869, %r871;
mov.b32 %f127, %r872;
add.ftz.f32 %f128, %f126, %f127;
mov.b32 %r873, %f128;
mov.u32 %r874, 8;
shfl.sync.bfly.b32 %r875|%p36, %r873, %r874, %r869, %r871;
mov.b32 %f129, %r875;
add.ftz.f32 %f130, %f128, %f129;
mov.b32 %r876, %f130;
mov.u32 %r877, 4;
shfl.sync.bfly.b32 %r878|%p37, %r876, %r877, %r869, %r871;
mov.b32 %f131, %r878;
add.ftz.f32 %f132, %f130, %f131;
mov.b32 %r879, %f132;
mov.u32 %r880, 2;
shfl.sync.bfly.b32 %r881|%p38, %r879, %r880, %r869, %r871;
mov.b32 %f133, %r881;
add.ftz.f32 %f134, %f132, %f133;
mov.b32 %r882, %f134;
shfl.sync.bfly.b32 %r883|%p39, %r882, %r867, %r869, %r871;
mov.b32 %f135, %r883;
add.ftz.f32 %f23, %f134, %f135;
cvt.u64.u32 %rd82, %r866;
add.s64 %rd83, %rd82, %rd14;
shl.b64 %rd84, %rd83, 2;
add.s64 %rd16, %rd1, %rd84;
@%p14 bra $L__BB218_16;

st.global.f32 [%rd16], %f23;

$L__BB218_16:
ld.shared.f32 %f136, [%r17+640];
add.ftz.f32 %f137, %f136, %f157;
ld.shared.f32 %f138, [%r17+1408];
add.ftz.f32 %f139, %f138, %f137;
ld.shared.f32 %f140, [%r17+2176];
add.ftz.f32 %f141, %f140, %f139;
mov.b32 %r884, %f141;
shfl.sync.bfly.b32 %r888|%p41, %r884, %r870, %r869, %r871;
mov.b32 %f142, %r888;
add.ftz.f32 %f143, %f141, %f142;
mov.b32 %r889, %f143;
shfl.sync.bfly.b32 %r891|%p42, %r889, %r874, %r869, %r871;
mov.b32 %f144, %r891;
add.ftz.f32 %f145, %f143, %f144;
mov.b32 %r892, %f145;
shfl.sync.bfly.b32 %r894|%p43, %r892, %r877, %r869, %r871;
mov.b32 %f146, %r894;
add.ftz.f32 %f147, %f145, %f146;
mov.b32 %r895, %f147;
shfl.sync.bfly.b32 %r897|%p44, %r895, %r880, %r869, %r871;
mov.b32 %f148, %r897;
add.ftz.f32 %f149, %f147, %f148;
mov.b32 %r898, %f149;
shfl.sync.bfly.b32 %r900|%p45, %r898, %r867, %r869, %r871;
mov.b32 %f150, %r900;
add.ftz.f32 %f24, %f149, %f150;
@%p22 bra $L__BB218_18;

st.global.f32 [%rd16+4], %f24;

$L__BB218_18:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type17ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type17ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type17ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type17ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type17ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type17ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type17ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type17ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type17ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<59>;
.reg .b16 %rs<41>;
.reg .f32 %f<217>;
.reg .b32 %r<1132>;
.reg .b64 %rd<99>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type17ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[3072];

ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type17ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type17ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type17ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type17ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type17ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type17ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type17ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type17ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type17ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type17ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type17ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type17ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type17ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type17ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type17ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type17ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type17ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
cvta.to.global.u64 %rd1, %rd19;
mov.u32 %r1, %tid.y;
shl.b32 %r39, %r1, 5;
mov.u32 %r40, %tid.x;
add.s32 %r2, %r39, %r40;
mov.u32 %r41, %ctaid.x;
shl.b32 %r3, %r41, 1;
shr.s32 %r42, %r2, 31;
shr.u32 %r43, %r42, 29;
add.s32 %r44, %r2, %r43;
shr.s32 %r1131, %r44, 3;
shr.u32 %r5, %r38, 8;
setp.ge.s32 %p3, %r1131, %r5;
mov.f32 %f209, 0f00000000;
mov.f32 %f210, %f209;
mov.f32 %f211, %f209;
mov.f32 %f212, %f209;
mov.f32 %f213, %f209;
mov.f32 %f214, %f209;
mov.f32 %f215, %f209;
mov.f32 %f216, %f209;
@%p3 bra $L__BB219_3;

mov.u32 %r45, %ctaid.y;
mov.u32 %r46, %ctaid.z;
mul.lo.s32 %r47, %r45, %r30;
cvt.u64.u32 %rd20, %r47;
mul.lo.s32 %r48, %r46, %r36;
cvt.u64.u32 %rd21, %r48;
add.s64 %rd2, %rd21, %rd20;
mul.hi.u32 %r49, %r46, %r32;
add.s32 %r50, %r49, %r46;
shr.u32 %r51, %r50, %r33;
mul.lo.s32 %r52, %r51, %r35;
mul.hi.u32 %r53, %r45, %r26;
add.s32 %r54, %r53, %r45;
shr.u32 %r55, %r54, %r27;
mul.lo.s32 %r56, %r55, %r29;
cvta.to.global.u64 %rd98, %rd17;
and.b32 %r60, %r44, -8;
sub.s32 %r61, %r2, %r60;
shl.b32 %r62, %r61, 2;
cvt.s64.s32 %rd4, %r61;
shl.b32 %r63, %r24, 1;
shl.b32 %r1127, %r1131, 3;
add.s32 %r1130, %r1127, %r63;
mad.lo.s32 %r1129, %r24, 3, %r1127;
add.s32 %r1128, %r24, %r1127;
add.s32 %r64, %r1131, %r52;
add.s32 %r65, %r64, %r56;
add.s32 %r66, %r3, 1;
mad.lo.s32 %r67, %r23, %r66, %r65;
mul.wide.s32 %rd5, %r67, 74;
add.s64 %rd22, %rd5, %rd4;
add.s64 %rd6, %rd22, 66;
mad.lo.s32 %r68, %r3, %r23, %r1131;
add.s32 %r69, %r68, %r52;
add.s32 %r70, %r69, %r56;
mul.wide.s32 %rd7, %r70, 74;
add.s64 %rd23, %rd7, %rd4;
add.s64 %rd8, %rd23, 66;
mul.wide.s32 %rd24, %r62, 2;
add.s64 %rd25, %rd5, %rd24;
add.s64 %rd9, %rd25, 4;
add.s64 %rd26, %rd7, %rd24;
add.s64 %rd10, %rd26, 8;
cvta.to.global.u64 %rd11, %rd18;
mov.f32 %f209, 0f00000000;

$L__BB219_2:
cvt.u64.u32 %rd27, %r1127;
add.s64 %rd28, %rd2, %rd27;
add.s64 %rd29, %rd28, %rd4;
mul.lo.s64 %rd30, %rd29, 36;
add.s64 %rd31, %rd98, %rd10;
ld.global.nc.u16 %rs9, [%rd31+-4];
ld.global.nc.u16 %rs10, [%rd31+-2];
ld.global.nc.u16 %rs11, [%rd31];
add.s64 %rd32, %rd98, %rd8;
ld.global.nc.u8 %rs12, [%rd32];
and.b16 %rs13, %rs12, 240;
and.b16 %rs14, %rs12, 15;
shr.u16 %rs15, %rs13, 4;
ld.global.nc.u16 %rs16, [%rd31+-6];
shl.b16 %rs17, %rs16, 3;
cvt.u64.u16 %rd33, %rs17;
and.b64 %rd34, %rd33, 4088;
mov.u64 %rd35, iq2xs_grid;
add.s64 %rd36, %rd35, %rd34;
ld.global.nc.v2.u32 {%r463, %r464}, [%rd36];
not.b32 %r466, %r463;
mov.u32 %r445, 0;
not.b32 %r468, %r464;
shr.u16 %rs18, %rs16, 9;
cvt.u32.u16 %r469, %rs18;
popc.b32 %r470, %r469;
shl.b32 %r471, %r470, 7;
and.b32 %r472, %r471, 128;
or.b32 %r473, %r472, %r469;
mul.lo.s32 %r474, %r473, 16843009;
and.b32 %r475, %r474, 134480385;
or.b32 %r476, %r475, -2139062144;
add.s32 %r477, %r476, -16843009;
or.b32 %r365, %r477, %r475;
mov.u32 %r71, %r365;

prmt.b32 %r71, %r71, 0, 0xba98;

xor.b32 %r478, %r71, %r463;
or.b32 %r479, %r478, -2139062144;
and.b32 %r480, %r71, 2139062143;
and.b32 %r481, %r466, -2139062144;
sub.s32 %r482, %r479, %r480;
xor.b32 %r76, %r481, %r482;
add.s64 %rd37, %rd11, %rd30;
ld.global.nc.u32 %r77, [%rd37+4];
and.b32 %r483, %r474, -2143281136;
or.b32 %r484, %r483, -2139062144;
add.s32 %r485, %r484, -16843009;
or.b32 %r367, %r485, %r483;
mov.u32 %r73, %r367;

prmt.b32 %r73, %r73, 0, 0xba98;

xor.b32 %r486, %r73, %r464;
or.b32 %r487, %r486, -2139062144;
and.b32 %r488, %r73, 2139062143;
and.b32 %r489, %r468, -2139062144;
sub.s32 %r490, %r487, %r488;
xor.b32 %r80, %r489, %r490;
ld.global.nc.u32 %r81, [%rd37+8];

dp4a.s32.s32 %r75, %r76, %r77, %r445;


dp4a.s32.s32 %r79, %r80, %r81, %r75;

shl.b16 %rs19, %rs9, 3;
cvt.u64.u16 %rd38, %rs19;
and.b64 %rd39, %rd38, 4088;
add.s64 %rd40, %rd35, %rd39;
ld.global.nc.v2.u32 {%r491, %r492}, [%rd40];
not.b32 %r494, %r491;
not.b32 %r496, %r492;
shr.u16 %rs20, %rs9, 9;
cvt.u32.u16 %r497, %rs20;
popc.b32 %r498, %r497;
shl.b32 %r499, %r498, 7;
and.b32 %r500, %r499, 128;
or.b32 %r501, %r500, %r497;
mul.lo.s32 %r502, %r501, 16843009;
and.b32 %r503, %r502, 134480385;
or.b32 %r504, %r503, -2139062144;
add.s32 %r505, %r504, -16843009;
or.b32 %r377, %r505, %r503;
mov.u32 %r83, %r377;

prmt.b32 %r83, %r83, 0, 0xba98;

xor.b32 %r506, %r83, %r491;
or.b32 %r507, %r506, -2139062144;
and.b32 %r508, %r83, 2139062143;
and.b32 %r509, %r494, -2139062144;
sub.s32 %r510, %r507, %r508;
xor.b32 %r88, %r509, %r510;
ld.global.nc.u32 %r89, [%rd37+12];
and.b32 %r511, %r502, -2143281136;
or.b32 %r512, %r511, -2139062144;
add.s32 %r513, %r512, -16843009;
or.b32 %r379, %r513, %r511;
mov.u32 %r85, %r379;

prmt.b32 %r85, %r85, 0, 0xba98;

xor.b32 %r514, %r85, %r492;
or.b32 %r515, %r514, -2139062144;
and.b32 %r516, %r85, 2139062143;
and.b32 %r517, %r496, -2139062144;
sub.s32 %r518, %r515, %r516;
xor.b32 %r92, %r517, %r518;
ld.global.nc.u32 %r93, [%rd37+16];

dp4a.s32.s32 %r87, %r88, %r89, %r79;


dp4a.s32.s32 %r91, %r92, %r93, %r87;

shl.b16 %rs21, %rs10, 3;
cvt.u64.u16 %rd41, %rs21;
and.b64 %rd42, %rd41, 4088;
add.s64 %rd43, %rd35, %rd42;
ld.global.nc.v2.u32 {%r519, %r520}, [%rd43];
not.b32 %r522, %r519;
not.b32 %r524, %r520;
shr.u16 %rs22, %rs10, 9;
cvt.u32.u16 %r525, %rs22;
popc.b32 %r526, %r525;
shl.b32 %r527, %r526, 7;
and.b32 %r528, %r527, 128;
or.b32 %r529, %r528, %r525;
mul.lo.s32 %r530, %r529, 16843009;
and.b32 %r531, %r530, 134480385;
or.b32 %r532, %r531, -2139062144;
add.s32 %r533, %r532, -16843009;
or.b32 %r389, %r533, %r531;
mov.u32 %r95, %r389;

prmt.b32 %r95, %r95, 0, 0xba98;

xor.b32 %r534, %r95, %r519;
or.b32 %r535, %r534, -2139062144;
and.b32 %r536, %r95, 2139062143;
and.b32 %r537, %r522, -2139062144;
sub.s32 %r538, %r535, %r536;
xor.b32 %r100, %r537, %r538;
ld.global.nc.u32 %r101, [%rd37+20];
and.b32 %r539, %r530, -2143281136;
or.b32 %r540, %r539, -2139062144;
add.s32 %r541, %r540, -16843009;
or.b32 %r391, %r541, %r539;
mov.u32 %r97, %r391;

prmt.b32 %r97, %r97, 0, 0xba98;

xor.b32 %r542, %r97, %r520;
or.b32 %r543, %r542, -2139062144;
and.b32 %r544, %r97, 2139062143;
and.b32 %r545, %r524, -2139062144;
sub.s32 %r546, %r543, %r544;
xor.b32 %r104, %r545, %r546;
ld.global.nc.u32 %r105, [%rd37+24];

dp4a.s32.s32 %r99, %r100, %r101, %r445;


dp4a.s32.s32 %r103, %r104, %r105, %r99;

shl.b16 %rs23, %rs11, 3;
cvt.u64.u16 %rd44, %rs23;
and.b64 %rd45, %rd44, 4088;
add.s64 %rd46, %rd35, %rd45;
ld.global.nc.v2.u32 {%r547, %r548}, [%rd46];
not.b32 %r550, %r547;
not.b32 %r552, %r548;
shr.u16 %rs24, %rs11, 9;
cvt.u32.u16 %r553, %rs24;
popc.b32 %r554, %r553;
shl.b32 %r555, %r554, 7;
and.b32 %r556, %r555, 128;
or.b32 %r557, %r556, %r553;
mul.lo.s32 %r558, %r557, 16843009;
and.b32 %r559, %r558, 134480385;
or.b32 %r560, %r559, -2139062144;
add.s32 %r561, %r560, -16843009;
or.b32 %r401, %r561, %r559;
mov.u32 %r107, %r401;

prmt.b32 %r107, %r107, 0, 0xba98;

xor.b32 %r562, %r107, %r547;
or.b32 %r563, %r562, -2139062144;
and.b32 %r564, %r107, 2139062143;
and.b32 %r565, %r550, -2139062144;
sub.s32 %r566, %r563, %r564;
xor.b32 %r112, %r565, %r566;
ld.global.nc.u32 %r113, [%rd37+28];
and.b32 %r567, %r558, -2143281136;
or.b32 %r568, %r567, -2139062144;
add.s32 %r569, %r568, -16843009;
or.b32 %r403, %r569, %r567;
mov.u32 %r109, %r403;

prmt.b32 %r109, %r109, 0, 0xba98;

xor.b32 %r570, %r109, %r548;
or.b32 %r571, %r570, -2139062144;
and.b32 %r572, %r109, 2139062143;
and.b32 %r573, %r552, -2139062144;
sub.s32 %r574, %r571, %r572;
xor.b32 %r116, %r573, %r574;
ld.global.nc.u32 %r117, [%rd37+32];

dp4a.s32.s32 %r111, %r112, %r113, %r103;


dp4a.s32.s32 %r115, %r116, %r117, %r111;

cvt.u32.u16 %r575, %rs14;
cvt.u32.u16 %r576, %rs15;
mul.lo.s32 %r577, %r91, %r575;
mad.lo.s32 %r578, %r115, %r576, %r577;
add.s32 %r579, %r115, %r91;
shr.u32 %r580, %r579, 31;
add.s32 %r581, %r579, %r580;
shr.s32 %r582, %r581, 1;
add.s32 %r583, %r578, %r582;
shr.s32 %r584, %r583, 31;
shr.u32 %r585, %r584, 30;
add.s32 %r586, %r583, %r585;
shr.s32 %r587, %r586, 2;
add.s64 %rd47, %rd98, %rd7;
ld.global.nc.u16 %rs1, [%rd47];

{ cvt.f32.f16 %f49, %rs1;}


ld.global.nc.u32 %r119, [%rd37];

{.reg .f16 low,high;
mov.b32 {low,high},%r119;
cvt.f32.f16 %f50, low;}


mul.ftz.f32 %f65, %f49, %f50;
cvt.rn.f32.s32 %f66, %r587;
fma.rn.ftz.f32 %f216, %f65, %f66, %f216;
add.s64 %rd48, %rd98, %rd9;
ld.global.nc.u16 %rs25, [%rd48];
ld.global.nc.u16 %rs26, [%rd48+2];
ld.global.nc.u16 %rs27, [%rd48+4];
add.s64 %rd49, %rd98, %rd6;
ld.global.nc.u8 %rs28, [%rd49];
and.b16 %rs29, %rs28, 240;
and.b16 %rs30, %rs28, 15;
shr.u16 %rs31, %rs29, 4;
ld.global.nc.u16 %rs32, [%rd48+-2];
shl.b16 %rs33, %rs32, 3;
cvt.u64.u16 %rd50, %rs33;
and.b64 %rd51, %rd50, 4088;
add.s64 %rd52, %rd35, %rd51;
ld.global.nc.v2.u32 {%r588, %r589}, [%rd52];
not.b32 %r591, %r588;
not.b32 %r593, %r589;
shr.u16 %rs34, %rs32, 9;
cvt.u32.u16 %r594, %rs34;
popc.b32 %r595, %r594;
shl.b32 %r596, %r595, 7;
and.b32 %r597, %r596, 128;
or.b32 %r598, %r597, %r594;
mul.lo.s32 %r599, %r598, 16843009;
and.b32 %r600, %r599, 134480385;
or.b32 %r601, %r600, -2139062144;
add.s32 %r602, %r601, -16843009;
or.b32 %r414, %r602, %r600;
mov.u32 %r120, %r414;

prmt.b32 %r120, %r120, 0, 0xba98;

xor.b32 %r603, %r120, %r588;
or.b32 %r604, %r603, -2139062144;
and.b32 %r605, %r120, 2139062143;
and.b32 %r606, %r591, -2139062144;
sub.s32 %r607, %r604, %r605;
xor.b32 %r125, %r606, %r607;
and.b32 %r608, %r599, -2143281136;
or.b32 %r609, %r608, -2139062144;
add.s32 %r610, %r609, -16843009;
or.b32 %r416, %r610, %r608;
mov.u32 %r122, %r416;

prmt.b32 %r122, %r122, 0, 0xba98;

xor.b32 %r611, %r122, %r589;
or.b32 %r612, %r611, -2139062144;
and.b32 %r613, %r122, 2139062143;
and.b32 %r614, %r593, -2139062144;
sub.s32 %r615, %r612, %r613;
xor.b32 %r129, %r614, %r615;

dp4a.s32.s32 %r124, %r125, %r77, %r445;


dp4a.s32.s32 %r128, %r129, %r81, %r124;

shl.b16 %rs35, %rs25, 3;
cvt.u64.u16 %rd53, %rs35;
and.b64 %rd54, %rd53, 4088;
add.s64 %rd55, %rd35, %rd54;
ld.global.nc.v2.u32 {%r616, %r617}, [%rd55];
not.b32 %r619, %r616;
not.b32 %r621, %r617;
shr.u16 %rs36, %rs25, 9;
cvt.u32.u16 %r622, %rs36;
popc.b32 %r623, %r622;
shl.b32 %r624, %r623, 7;
and.b32 %r625, %r624, 128;
or.b32 %r626, %r625, %r622;
mul.lo.s32 %r627, %r626, 16843009;
and.b32 %r628, %r627, 134480385;
or.b32 %r629, %r628, -2139062144;
add.s32 %r630, %r629, -16843009;
or.b32 %r426, %r630, %r628;
mov.u32 %r132, %r426;

prmt.b32 %r132, %r132, 0, 0xba98;

xor.b32 %r631, %r132, %r616;
or.b32 %r632, %r631, -2139062144;
and.b32 %r633, %r132, 2139062143;
and.b32 %r634, %r619, -2139062144;
sub.s32 %r635, %r632, %r633;
xor.b32 %r137, %r634, %r635;
and.b32 %r636, %r627, -2143281136;
or.b32 %r637, %r636, -2139062144;
add.s32 %r638, %r637, -16843009;
or.b32 %r428, %r638, %r636;
mov.u32 %r134, %r428;

prmt.b32 %r134, %r134, 0, 0xba98;

xor.b32 %r639, %r134, %r617;
or.b32 %r640, %r639, -2139062144;
and.b32 %r641, %r134, 2139062143;
and.b32 %r642, %r621, -2139062144;
sub.s32 %r643, %r640, %r641;
xor.b32 %r141, %r642, %r643;

dp4a.s32.s32 %r136, %r137, %r89, %r128;


dp4a.s32.s32 %r140, %r141, %r93, %r136;

shl.b16 %rs37, %rs26, 3;
cvt.u64.u16 %rd56, %rs37;
and.b64 %rd57, %rd56, 4088;
add.s64 %rd58, %rd35, %rd57;
ld.global.nc.v2.u32 {%r644, %r645}, [%rd58];
not.b32 %r647, %r644;
not.b32 %r649, %r645;
shr.u16 %rs38, %rs26, 9;
cvt.u32.u16 %r650, %rs38;
popc.b32 %r651, %r650;
shl.b32 %r652, %r651, 7;
and.b32 %r653, %r652, 128;
or.b32 %r654, %r653, %r650;
mul.lo.s32 %r655, %r654, 16843009;
and.b32 %r656, %r655, 134480385;
or.b32 %r657, %r656, -2139062144;
add.s32 %r658, %r657, -16843009;
or.b32 %r438, %r658, %r656;
mov.u32 %r144, %r438;

prmt.b32 %r144, %r144, 0, 0xba98;

xor.b32 %r659, %r144, %r644;
or.b32 %r660, %r659, -2139062144;
and.b32 %r661, %r144, 2139062143;
and.b32 %r662, %r647, -2139062144;
sub.s32 %r663, %r660, %r661;
xor.b32 %r149, %r662, %r663;
and.b32 %r664, %r655, -2143281136;
or.b32 %r665, %r664, -2139062144;
add.s32 %r666, %r665, -16843009;
or.b32 %r440, %r666, %r664;
mov.u32 %r146, %r440;

prmt.b32 %r146, %r146, 0, 0xba98;

xor.b32 %r667, %r146, %r645;
or.b32 %r668, %r667, -2139062144;
and.b32 %r669, %r146, 2139062143;
and.b32 %r670, %r649, -2139062144;
sub.s32 %r671, %r668, %r669;
xor.b32 %r153, %r670, %r671;

dp4a.s32.s32 %r148, %r149, %r101, %r445;


dp4a.s32.s32 %r152, %r153, %r105, %r148;

shl.b16 %rs39, %rs27, 3;
cvt.u64.u16 %rd59, %rs39;
and.b64 %rd60, %rd59, 4088;
add.s64 %rd61, %rd35, %rd60;
ld.global.nc.v2.u32 {%r672, %r673}, [%rd61];
not.b32 %r675, %r672;
not.b32 %r677, %r673;
shr.u16 %rs40, %rs27, 9;
cvt.u32.u16 %r678, %rs40;
popc.b32 %r679, %r678;
shl.b32 %r680, %r679, 7;
and.b32 %r681, %r680, 128;
or.b32 %r682, %r681, %r678;
mul.lo.s32 %r683, %r682, 16843009;
and.b32 %r684, %r683, 134480385;
or.b32 %r685, %r684, -2139062144;
add.s32 %r686, %r685, -16843009;
or.b32 %r450, %r686, %r684;
mov.u32 %r156, %r450;

prmt.b32 %r156, %r156, 0, 0xba98;

xor.b32 %r687, %r156, %r672;
or.b32 %r688, %r687, -2139062144;
and.b32 %r689, %r156, 2139062143;
and.b32 %r690, %r675, -2139062144;
sub.s32 %r691, %r688, %r689;
xor.b32 %r161, %r690, %r691;
and.b32 %r692, %r683, -2143281136;
or.b32 %r693, %r692, -2139062144;
add.s32 %r694, %r693, -16843009;
or.b32 %r452, %r694, %r692;
mov.u32 %r158, %r452;

prmt.b32 %r158, %r158, 0, 0xba98;

xor.b32 %r695, %r158, %r673;
or.b32 %r696, %r695, -2139062144;
and.b32 %r697, %r158, 2139062143;
and.b32 %r698, %r677, -2139062144;
sub.s32 %r699, %r696, %r697;
xor.b32 %r165, %r698, %r699;

dp4a.s32.s32 %r160, %r161, %r113, %r152;


dp4a.s32.s32 %r164, %r165, %r117, %r160;

cvt.u32.u16 %r700, %rs30;
cvt.u32.u16 %r701, %rs31;
mul.lo.s32 %r702, %r140, %r700;
mad.lo.s32 %r703, %r164, %r701, %r702;
add.s32 %r704, %r164, %r140;
shr.u32 %r705, %r704, 31;
add.s32 %r706, %r704, %r705;
shr.s32 %r707, %r706, 1;
add.s32 %r708, %r703, %r707;
shr.s32 %r709, %r708, 31;
shr.u32 %r710, %r709, 30;
add.s32 %r711, %r708, %r710;
shr.s32 %r712, %r711, 2;
add.s64 %rd62, %rd98, %rd5;
ld.global.nc.u16 %rs2, [%rd62];

{ cvt.f32.f16 %f51, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r119;
cvt.f32.f16 %f52, low;}


mul.ftz.f32 %f67, %f51, %f52;
cvt.rn.f32.s32 %f68, %r712;
fma.rn.ftz.f32 %f215, %f67, %f68, %f215;
cvt.u64.u32 %rd63, %r1128;
add.s64 %rd64, %rd2, %rd63;
add.s64 %rd65, %rd64, %rd4;
mul.lo.s64 %rd66, %rd65, 36;
mov.u32 %r169, %r365;

prmt.b32 %r169, %r169, 0, 0xba98;

xor.b32 %r713, %r169, %r463;
or.b32 %r714, %r713, -2139062144;
and.b32 %r715, %r169, 2139062143;
sub.s32 %r716, %r714, %r715;
xor.b32 %r174, %r481, %r716;
add.s64 %rd67, %rd11, %rd66;
ld.global.nc.u32 %r175, [%rd67+4];
mov.u32 %r171, %r367;

prmt.b32 %r171, %r171, 0, 0xba98;

xor.b32 %r717, %r171, %r464;
or.b32 %r718, %r717, -2139062144;
and.b32 %r719, %r171, 2139062143;
sub.s32 %r720, %r718, %r719;
xor.b32 %r178, %r489, %r720;
ld.global.nc.u32 %r179, [%rd67+8];

dp4a.s32.s32 %r173, %r174, %r175, %r445;


dp4a.s32.s32 %r177, %r178, %r179, %r173;

mov.u32 %r181, %r377;

prmt.b32 %r181, %r181, 0, 0xba98;

xor.b32 %r721, %r181, %r491;
or.b32 %r722, %r721, -2139062144;
and.b32 %r723, %r181, 2139062143;
sub.s32 %r724, %r722, %r723;
xor.b32 %r186, %r509, %r724;
ld.global.nc.u32 %r187, [%rd67+12];
mov.u32 %r183, %r379;

prmt.b32 %r183, %r183, 0, 0xba98;

xor.b32 %r725, %r183, %r492;
or.b32 %r726, %r725, -2139062144;
and.b32 %r727, %r183, 2139062143;
sub.s32 %r728, %r726, %r727;
xor.b32 %r190, %r517, %r728;
ld.global.nc.u32 %r191, [%rd67+16];

dp4a.s32.s32 %r185, %r186, %r187, %r177;


dp4a.s32.s32 %r189, %r190, %r191, %r185;

mov.u32 %r193, %r389;

prmt.b32 %r193, %r193, 0, 0xba98;

xor.b32 %r729, %r193, %r519;
or.b32 %r730, %r729, -2139062144;
and.b32 %r731, %r193, 2139062143;
sub.s32 %r732, %r730, %r731;
xor.b32 %r198, %r537, %r732;
ld.global.nc.u32 %r199, [%rd67+20];
mov.u32 %r195, %r391;

prmt.b32 %r195, %r195, 0, 0xba98;

xor.b32 %r733, %r195, %r520;
or.b32 %r734, %r733, -2139062144;
and.b32 %r735, %r195, 2139062143;
sub.s32 %r736, %r734, %r735;
xor.b32 %r202, %r545, %r736;
ld.global.nc.u32 %r203, [%rd67+24];

dp4a.s32.s32 %r197, %r198, %r199, %r445;


dp4a.s32.s32 %r201, %r202, %r203, %r197;

mov.u32 %r205, %r401;

prmt.b32 %r205, %r205, 0, 0xba98;

xor.b32 %r737, %r205, %r547;
or.b32 %r738, %r737, -2139062144;
and.b32 %r739, %r205, 2139062143;
sub.s32 %r740, %r738, %r739;
xor.b32 %r210, %r565, %r740;
ld.global.nc.u32 %r211, [%rd67+28];
mov.u32 %r207, %r403;

prmt.b32 %r207, %r207, 0, 0xba98;

xor.b32 %r741, %r207, %r548;
or.b32 %r742, %r741, -2139062144;
and.b32 %r743, %r207, 2139062143;
sub.s32 %r744, %r742, %r743;
xor.b32 %r214, %r573, %r744;
ld.global.nc.u32 %r215, [%rd67+32];

dp4a.s32.s32 %r209, %r210, %r211, %r201;


dp4a.s32.s32 %r213, %r214, %r215, %r209;

mul.lo.s32 %r745, %r189, %r575;
mad.lo.s32 %r746, %r213, %r576, %r745;
add.s32 %r747, %r213, %r189;
shr.u32 %r748, %r747, 31;
add.s32 %r749, %r747, %r748;
shr.s32 %r750, %r749, 1;
add.s32 %r751, %r746, %r750;
shr.s32 %r752, %r751, 31;
shr.u32 %r753, %r752, 30;
add.s32 %r754, %r751, %r753;
shr.s32 %r755, %r754, 2;

{ cvt.f32.f16 %f53, %rs1;}


ld.global.nc.u32 %r217, [%rd67];

{.reg .f16 low,high;
mov.b32 {low,high},%r217;
cvt.f32.f16 %f54, low;}


mul.ftz.f32 %f69, %f53, %f54;
cvt.rn.f32.s32 %f70, %r755;
fma.rn.ftz.f32 %f214, %f69, %f70, %f214;
mov.u32 %r218, %r414;

prmt.b32 %r218, %r218, 0, 0xba98;

xor.b32 %r756, %r218, %r588;
or.b32 %r757, %r756, -2139062144;
and.b32 %r758, %r218, 2139062143;
sub.s32 %r759, %r757, %r758;
xor.b32 %r223, %r606, %r759;
mov.u32 %r220, %r416;

prmt.b32 %r220, %r220, 0, 0xba98;

xor.b32 %r760, %r220, %r589;
or.b32 %r761, %r760, -2139062144;
and.b32 %r762, %r220, 2139062143;
sub.s32 %r763, %r761, %r762;
xor.b32 %r227, %r614, %r763;

dp4a.s32.s32 %r222, %r223, %r175, %r445;


dp4a.s32.s32 %r226, %r227, %r179, %r222;

mov.u32 %r230, %r426;

prmt.b32 %r230, %r230, 0, 0xba98;

xor.b32 %r764, %r230, %r616;
or.b32 %r765, %r764, -2139062144;
and.b32 %r766, %r230, 2139062143;
sub.s32 %r767, %r765, %r766;
xor.b32 %r235, %r634, %r767;
mov.u32 %r232, %r428;

prmt.b32 %r232, %r232, 0, 0xba98;

xor.b32 %r768, %r232, %r617;
or.b32 %r769, %r768, -2139062144;
and.b32 %r770, %r232, 2139062143;
sub.s32 %r771, %r769, %r770;
xor.b32 %r239, %r642, %r771;

dp4a.s32.s32 %r234, %r235, %r187, %r226;


dp4a.s32.s32 %r238, %r239, %r191, %r234;

mov.u32 %r242, %r438;

prmt.b32 %r242, %r242, 0, 0xba98;

xor.b32 %r772, %r242, %r644;
or.b32 %r773, %r772, -2139062144;
and.b32 %r774, %r242, 2139062143;
sub.s32 %r775, %r773, %r774;
xor.b32 %r247, %r662, %r775;
mov.u32 %r244, %r440;

prmt.b32 %r244, %r244, 0, 0xba98;

xor.b32 %r776, %r244, %r645;
or.b32 %r777, %r776, -2139062144;
and.b32 %r778, %r244, 2139062143;
sub.s32 %r779, %r777, %r778;
xor.b32 %r251, %r670, %r779;

dp4a.s32.s32 %r246, %r247, %r199, %r445;


dp4a.s32.s32 %r250, %r251, %r203, %r246;

mov.u32 %r254, %r450;

prmt.b32 %r254, %r254, 0, 0xba98;

xor.b32 %r780, %r254, %r672;
or.b32 %r781, %r780, -2139062144;
and.b32 %r782, %r254, 2139062143;
sub.s32 %r783, %r781, %r782;
xor.b32 %r259, %r690, %r783;
mov.u32 %r256, %r452;

prmt.b32 %r256, %r256, 0, 0xba98;

xor.b32 %r784, %r256, %r673;
or.b32 %r785, %r784, -2139062144;
and.b32 %r786, %r256, 2139062143;
sub.s32 %r787, %r785, %r786;
xor.b32 %r263, %r698, %r787;

dp4a.s32.s32 %r258, %r259, %r211, %r250;


dp4a.s32.s32 %r262, %r263, %r215, %r258;

mul.lo.s32 %r788, %r238, %r700;
mad.lo.s32 %r789, %r262, %r701, %r788;
add.s32 %r790, %r262, %r238;
shr.u32 %r791, %r790, 31;
add.s32 %r792, %r790, %r791;
shr.s32 %r793, %r792, 1;
add.s32 %r794, %r789, %r793;
shr.s32 %r795, %r794, 31;
shr.u32 %r796, %r795, 30;
add.s32 %r797, %r794, %r796;
shr.s32 %r798, %r797, 2;

{ cvt.f32.f16 %f55, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r217;
cvt.f32.f16 %f56, low;}


mul.ftz.f32 %f71, %f55, %f56;
cvt.rn.f32.s32 %f72, %r798;
fma.rn.ftz.f32 %f213, %f71, %f72, %f213;
cvt.u64.u32 %rd68, %r1130;
add.s64 %rd69, %rd2, %rd68;
add.s64 %rd70, %rd69, %rd4;
mul.lo.s64 %rd71, %rd70, 36;
mov.u32 %r267, %r365;

prmt.b32 %r267, %r267, 0, 0xba98;

xor.b32 %r799, %r267, %r463;
or.b32 %r800, %r799, -2139062144;
and.b32 %r801, %r267, 2139062143;
sub.s32 %r802, %r800, %r801;
xor.b32 %r272, %r481, %r802;
add.s64 %rd72, %rd11, %rd71;
ld.global.nc.u32 %r273, [%rd72+4];
mov.u32 %r269, %r367;

prmt.b32 %r269, %r269, 0, 0xba98;

xor.b32 %r803, %r269, %r464;
or.b32 %r804, %r803, -2139062144;
and.b32 %r805, %r269, 2139062143;
sub.s32 %r806, %r804, %r805;
xor.b32 %r276, %r489, %r806;
ld.global.nc.u32 %r277, [%rd72+8];

dp4a.s32.s32 %r271, %r272, %r273, %r445;


dp4a.s32.s32 %r275, %r276, %r277, %r271;

mov.u32 %r279, %r377;

prmt.b32 %r279, %r279, 0, 0xba98;

xor.b32 %r807, %r279, %r491;
or.b32 %r808, %r807, -2139062144;
and.b32 %r809, %r279, 2139062143;
sub.s32 %r810, %r808, %r809;
xor.b32 %r284, %r509, %r810;
ld.global.nc.u32 %r285, [%rd72+12];
mov.u32 %r281, %r379;

prmt.b32 %r281, %r281, 0, 0xba98;

xor.b32 %r811, %r281, %r492;
or.b32 %r812, %r811, -2139062144;
and.b32 %r813, %r281, 2139062143;
sub.s32 %r814, %r812, %r813;
xor.b32 %r288, %r517, %r814;
ld.global.nc.u32 %r289, [%rd72+16];

dp4a.s32.s32 %r283, %r284, %r285, %r275;


dp4a.s32.s32 %r287, %r288, %r289, %r283;

mov.u32 %r291, %r389;

prmt.b32 %r291, %r291, 0, 0xba98;

xor.b32 %r815, %r291, %r519;
or.b32 %r816, %r815, -2139062144;
and.b32 %r817, %r291, 2139062143;
sub.s32 %r818, %r816, %r817;
xor.b32 %r296, %r537, %r818;
ld.global.nc.u32 %r297, [%rd72+20];
mov.u32 %r293, %r391;

prmt.b32 %r293, %r293, 0, 0xba98;

xor.b32 %r819, %r293, %r520;
or.b32 %r820, %r819, -2139062144;
and.b32 %r821, %r293, 2139062143;
sub.s32 %r822, %r820, %r821;
xor.b32 %r300, %r545, %r822;
ld.global.nc.u32 %r301, [%rd72+24];

dp4a.s32.s32 %r295, %r296, %r297, %r445;


dp4a.s32.s32 %r299, %r300, %r301, %r295;

mov.u32 %r303, %r401;

prmt.b32 %r303, %r303, 0, 0xba98;

xor.b32 %r823, %r303, %r547;
or.b32 %r824, %r823, -2139062144;
and.b32 %r825, %r303, 2139062143;
sub.s32 %r826, %r824, %r825;
xor.b32 %r308, %r565, %r826;
ld.global.nc.u32 %r309, [%rd72+28];
mov.u32 %r305, %r403;

prmt.b32 %r305, %r305, 0, 0xba98;

xor.b32 %r827, %r305, %r548;
or.b32 %r828, %r827, -2139062144;
and.b32 %r829, %r305, 2139062143;
sub.s32 %r830, %r828, %r829;
xor.b32 %r312, %r573, %r830;
ld.global.nc.u32 %r313, [%rd72+32];

dp4a.s32.s32 %r307, %r308, %r309, %r299;


dp4a.s32.s32 %r311, %r312, %r313, %r307;

mul.lo.s32 %r831, %r287, %r575;
mad.lo.s32 %r832, %r311, %r576, %r831;
add.s32 %r833, %r311, %r287;
shr.u32 %r834, %r833, 31;
add.s32 %r835, %r833, %r834;
shr.s32 %r836, %r835, 1;
add.s32 %r837, %r832, %r836;
shr.s32 %r838, %r837, 31;
shr.u32 %r839, %r838, 30;
add.s32 %r840, %r837, %r839;
shr.s32 %r841, %r840, 2;

{ cvt.f32.f16 %f57, %rs1;}


ld.global.nc.u32 %r315, [%rd72];

{.reg .f16 low,high;
mov.b32 {low,high},%r315;
cvt.f32.f16 %f58, low;}


mul.ftz.f32 %f73, %f57, %f58;
cvt.rn.f32.s32 %f74, %r841;
fma.rn.ftz.f32 %f212, %f73, %f74, %f212;
mov.u32 %r316, %r414;

prmt.b32 %r316, %r316, 0, 0xba98;

xor.b32 %r842, %r316, %r588;
or.b32 %r843, %r842, -2139062144;
and.b32 %r844, %r316, 2139062143;
sub.s32 %r845, %r843, %r844;
xor.b32 %r321, %r606, %r845;
mov.u32 %r318, %r416;

prmt.b32 %r318, %r318, 0, 0xba98;

xor.b32 %r846, %r318, %r589;
or.b32 %r847, %r846, -2139062144;
and.b32 %r848, %r318, 2139062143;
sub.s32 %r849, %r847, %r848;
xor.b32 %r325, %r614, %r849;

dp4a.s32.s32 %r320, %r321, %r273, %r445;


dp4a.s32.s32 %r324, %r325, %r277, %r320;

mov.u32 %r328, %r426;

prmt.b32 %r328, %r328, 0, 0xba98;

xor.b32 %r850, %r328, %r616;
or.b32 %r851, %r850, -2139062144;
and.b32 %r852, %r328, 2139062143;
sub.s32 %r853, %r851, %r852;
xor.b32 %r333, %r634, %r853;
mov.u32 %r330, %r428;

prmt.b32 %r330, %r330, 0, 0xba98;

xor.b32 %r854, %r330, %r617;
or.b32 %r855, %r854, -2139062144;
and.b32 %r856, %r330, 2139062143;
sub.s32 %r857, %r855, %r856;
xor.b32 %r337, %r642, %r857;

dp4a.s32.s32 %r332, %r333, %r285, %r324;


dp4a.s32.s32 %r336, %r337, %r289, %r332;

mov.u32 %r340, %r438;

prmt.b32 %r340, %r340, 0, 0xba98;

xor.b32 %r858, %r340, %r644;
or.b32 %r859, %r858, -2139062144;
and.b32 %r860, %r340, 2139062143;
sub.s32 %r861, %r859, %r860;
xor.b32 %r345, %r662, %r861;
mov.u32 %r342, %r440;

prmt.b32 %r342, %r342, 0, 0xba98;

xor.b32 %r862, %r342, %r645;
or.b32 %r863, %r862, -2139062144;
and.b32 %r864, %r342, 2139062143;
sub.s32 %r865, %r863, %r864;
xor.b32 %r349, %r670, %r865;

dp4a.s32.s32 %r344, %r345, %r297, %r445;


dp4a.s32.s32 %r348, %r349, %r301, %r344;

mov.u32 %r352, %r450;

prmt.b32 %r352, %r352, 0, 0xba98;

xor.b32 %r866, %r352, %r672;
or.b32 %r867, %r866, -2139062144;
and.b32 %r868, %r352, 2139062143;
sub.s32 %r869, %r867, %r868;
xor.b32 %r357, %r690, %r869;
mov.u32 %r354, %r452;

prmt.b32 %r354, %r354, 0, 0xba98;

xor.b32 %r870, %r354, %r673;
or.b32 %r871, %r870, -2139062144;
and.b32 %r872, %r354, 2139062143;
sub.s32 %r873, %r871, %r872;
xor.b32 %r361, %r698, %r873;

dp4a.s32.s32 %r356, %r357, %r309, %r348;


dp4a.s32.s32 %r360, %r361, %r313, %r356;

mul.lo.s32 %r874, %r336, %r700;
mad.lo.s32 %r875, %r360, %r701, %r874;
add.s32 %r876, %r360, %r336;
shr.u32 %r877, %r876, 31;
add.s32 %r878, %r876, %r877;
shr.s32 %r879, %r878, 1;
add.s32 %r880, %r875, %r879;
shr.s32 %r881, %r880, 31;
shr.u32 %r882, %r881, 30;
add.s32 %r883, %r880, %r882;
shr.s32 %r884, %r883, 2;

{ cvt.f32.f16 %f59, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r315;
cvt.f32.f16 %f60, low;}


mul.ftz.f32 %f75, %f59, %f60;
cvt.rn.f32.s32 %f76, %r884;
fma.rn.ftz.f32 %f211, %f75, %f76, %f211;
cvt.u64.u32 %rd73, %r1129;
add.s64 %rd74, %rd2, %rd73;
add.s64 %rd75, %rd74, %rd4;
mul.lo.s64 %rd76, %rd75, 36;

prmt.b32 %r365, %r365, 0, 0xba98;

xor.b32 %r885, %r365, %r463;
or.b32 %r886, %r885, -2139062144;
and.b32 %r887, %r365, 2139062143;
sub.s32 %r888, %r886, %r887;
xor.b32 %r370, %r481, %r888;
add.s64 %rd77, %rd11, %rd76;
ld.global.nc.u32 %r371, [%rd77+4];

prmt.b32 %r367, %r367, 0, 0xba98;

xor.b32 %r889, %r367, %r464;
or.b32 %r890, %r889, -2139062144;
and.b32 %r891, %r367, 2139062143;
sub.s32 %r892, %r890, %r891;
xor.b32 %r374, %r489, %r892;
ld.global.nc.u32 %r375, [%rd77+8];

dp4a.s32.s32 %r369, %r370, %r371, %r445;


dp4a.s32.s32 %r373, %r374, %r375, %r369;


prmt.b32 %r377, %r377, 0, 0xba98;

xor.b32 %r893, %r377, %r491;
or.b32 %r894, %r893, -2139062144;
and.b32 %r895, %r377, 2139062143;
sub.s32 %r896, %r894, %r895;
xor.b32 %r382, %r509, %r896;
ld.global.nc.u32 %r383, [%rd77+12];

prmt.b32 %r379, %r379, 0, 0xba98;

xor.b32 %r897, %r379, %r492;
or.b32 %r898, %r897, -2139062144;
and.b32 %r899, %r379, 2139062143;
sub.s32 %r900, %r898, %r899;
xor.b32 %r386, %r517, %r900;
ld.global.nc.u32 %r387, [%rd77+16];

dp4a.s32.s32 %r381, %r382, %r383, %r373;


dp4a.s32.s32 %r385, %r386, %r387, %r381;


prmt.b32 %r389, %r389, 0, 0xba98;

xor.b32 %r901, %r389, %r519;
or.b32 %r902, %r901, -2139062144;
and.b32 %r903, %r389, 2139062143;
sub.s32 %r904, %r902, %r903;
xor.b32 %r394, %r537, %r904;
ld.global.nc.u32 %r395, [%rd77+20];

prmt.b32 %r391, %r391, 0, 0xba98;

xor.b32 %r905, %r391, %r520;
or.b32 %r906, %r905, -2139062144;
and.b32 %r907, %r391, 2139062143;
sub.s32 %r908, %r906, %r907;
xor.b32 %r398, %r545, %r908;
ld.global.nc.u32 %r399, [%rd77+24];

dp4a.s32.s32 %r393, %r394, %r395, %r445;


dp4a.s32.s32 %r397, %r398, %r399, %r393;


prmt.b32 %r401, %r401, 0, 0xba98;

xor.b32 %r909, %r401, %r547;
or.b32 %r910, %r909, -2139062144;
and.b32 %r911, %r401, 2139062143;
sub.s32 %r912, %r910, %r911;
xor.b32 %r406, %r565, %r912;
ld.global.nc.u32 %r407, [%rd77+28];

prmt.b32 %r403, %r403, 0, 0xba98;

xor.b32 %r913, %r403, %r548;
or.b32 %r914, %r913, -2139062144;
and.b32 %r915, %r403, 2139062143;
sub.s32 %r916, %r914, %r915;
xor.b32 %r410, %r573, %r916;
ld.global.nc.u32 %r411, [%rd77+32];

dp4a.s32.s32 %r405, %r406, %r407, %r397;


dp4a.s32.s32 %r409, %r410, %r411, %r405;

mul.lo.s32 %r917, %r385, %r575;
mad.lo.s32 %r918, %r409, %r576, %r917;
add.s32 %r919, %r409, %r385;
shr.u32 %r920, %r919, 31;
add.s32 %r921, %r919, %r920;
shr.s32 %r922, %r921, 1;
add.s32 %r923, %r918, %r922;
shr.s32 %r924, %r923, 31;
shr.u32 %r925, %r924, 30;
add.s32 %r926, %r923, %r925;
shr.s32 %r927, %r926, 2;

{ cvt.f32.f16 %f61, %rs1;}


ld.global.nc.u32 %r413, [%rd77];

{.reg .f16 low,high;
mov.b32 {low,high},%r413;
cvt.f32.f16 %f62, low;}


mul.ftz.f32 %f77, %f61, %f62;
cvt.rn.f32.s32 %f78, %r927;
fma.rn.ftz.f32 %f210, %f77, %f78, %f210;

prmt.b32 %r414, %r414, 0, 0xba98;

xor.b32 %r928, %r414, %r588;
or.b32 %r929, %r928, -2139062144;
and.b32 %r930, %r414, 2139062143;
sub.s32 %r931, %r929, %r930;
xor.b32 %r419, %r606, %r931;

prmt.b32 %r416, %r416, 0, 0xba98;

xor.b32 %r932, %r416, %r589;
or.b32 %r933, %r932, -2139062144;
and.b32 %r934, %r416, 2139062143;
sub.s32 %r935, %r933, %r934;
xor.b32 %r423, %r614, %r935;

dp4a.s32.s32 %r418, %r419, %r371, %r445;


dp4a.s32.s32 %r422, %r423, %r375, %r418;


prmt.b32 %r426, %r426, 0, 0xba98;

xor.b32 %r936, %r426, %r616;
or.b32 %r937, %r936, -2139062144;
and.b32 %r938, %r426, 2139062143;
sub.s32 %r939, %r937, %r938;
xor.b32 %r431, %r634, %r939;

prmt.b32 %r428, %r428, 0, 0xba98;

xor.b32 %r940, %r428, %r617;
or.b32 %r941, %r940, -2139062144;
and.b32 %r942, %r428, 2139062143;
sub.s32 %r943, %r941, %r942;
xor.b32 %r435, %r642, %r943;

dp4a.s32.s32 %r430, %r431, %r383, %r422;


dp4a.s32.s32 %r434, %r435, %r387, %r430;


prmt.b32 %r438, %r438, 0, 0xba98;

xor.b32 %r944, %r438, %r644;
or.b32 %r945, %r944, -2139062144;
and.b32 %r946, %r438, 2139062143;
sub.s32 %r947, %r945, %r946;
xor.b32 %r443, %r662, %r947;

prmt.b32 %r440, %r440, 0, 0xba98;

xor.b32 %r948, %r440, %r645;
or.b32 %r949, %r948, -2139062144;
and.b32 %r950, %r440, 2139062143;
sub.s32 %r951, %r949, %r950;
xor.b32 %r447, %r670, %r951;

dp4a.s32.s32 %r442, %r443, %r395, %r445;


dp4a.s32.s32 %r446, %r447, %r399, %r442;


prmt.b32 %r450, %r450, 0, 0xba98;

xor.b32 %r952, %r450, %r672;
or.b32 %r953, %r952, -2139062144;
and.b32 %r954, %r450, 2139062143;
sub.s32 %r955, %r953, %r954;
xor.b32 %r455, %r690, %r955;

prmt.b32 %r452, %r452, 0, 0xba98;

xor.b32 %r956, %r452, %r673;
or.b32 %r957, %r956, -2139062144;
and.b32 %r958, %r452, 2139062143;
sub.s32 %r959, %r957, %r958;
xor.b32 %r459, %r698, %r959;

dp4a.s32.s32 %r454, %r455, %r407, %r446;


dp4a.s32.s32 %r458, %r459, %r411, %r454;

mul.lo.s32 %r960, %r434, %r700;
mad.lo.s32 %r961, %r458, %r701, %r960;
add.s32 %r962, %r458, %r434;
shr.u32 %r963, %r962, 31;
add.s32 %r964, %r962, %r963;
shr.s32 %r965, %r964, 1;
add.s32 %r966, %r961, %r965;
shr.s32 %r967, %r966, 31;
shr.u32 %r968, %r967, 30;
add.s32 %r969, %r966, %r968;
shr.s32 %r970, %r969, 2;

{ cvt.f32.f16 %f63, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r413;
cvt.f32.f16 %f64, low;}


mul.ftz.f32 %f79, %f63, %f64;
cvt.rn.f32.s32 %f80, %r970;
fma.rn.ftz.f32 %f209, %f79, %f80, %f209;
add.s32 %r1130, %r1130, 128;
add.s32 %r1129, %r1129, 128;
add.s32 %r1128, %r1128, 128;
add.s32 %r1127, %r1127, 128;
add.s64 %rd98, %rd98, 1184;
add.s32 %r1131, %r1131, 16;
setp.lt.s32 %p4, %r1131, %r5;
@%p4 bra $L__BB219_2;

$L__BB219_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB219_5;

shl.b32 %r972, %r1, 10;
mov.u32 %r973, _ZZ13mul_mat_vec_qIL9ggml_type17ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r974, %r973, %r972;
shl.b32 %r975, %r40, 2;
add.s32 %r976, %r974, %r975;
st.shared.f32 [%r976+-1024], %f216;
st.shared.f32 [%r976+-896], %f215;
st.shared.f32 [%r976+-768], %f214;
st.shared.f32 [%r976+-640], %f213;
st.shared.f32 [%r976+-512], %f212;
st.shared.f32 [%r976+-384], %f211;
st.shared.f32 [%r976+-256], %f210;
st.shared.f32 [%r976+-128], %f209;

$L__BB219_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB219_22;

mov.u32 %r977, %ctaid.z;
mov.u32 %r978, %ctaid.y;
mad.lo.s32 %r979, %r978, %r31, %r3;
mad.lo.s32 %r980, %r977, %r37, %r979;
cvt.u64.u32 %rd14, %r980;
shl.b32 %r982, %r40, 2;
mov.u32 %r983, _ZZ13mul_mat_vec_qIL9ggml_type17ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r20, %r983, %r982;
mov.u32 %r984, 2;
ld.shared.f32 %f81, [%r20];
add.ftz.f32 %f82, %f81, %f216;
ld.shared.f32 %f83, [%r20+1024];
add.ftz.f32 %f84, %f83, %f82;
ld.shared.f32 %f85, [%r20+2048];
add.ftz.f32 %f86, %f85, %f84;
mov.b32 %r985, %f86;
mov.u32 %r986, 31;
mov.u32 %r987, 16;
mov.u32 %r988, -1;
shfl.sync.bfly.b32 %r989|%p7, %r985, %r987, %r986, %r988;
mov.b32 %f87, %r989;
add.ftz.f32 %f88, %f86, %f87;
mov.b32 %r990, %f88;
mov.u32 %r991, 8;
shfl.sync.bfly.b32 %r992|%p8, %r990, %r991, %r986, %r988;
mov.b32 %f89, %r992;
add.ftz.f32 %f90, %f88, %f89;
mov.b32 %r993, %f90;
mov.u32 %r994, 4;
shfl.sync.bfly.b32 %r995|%p9, %r993, %r994, %r986, %r988;
mov.b32 %f91, %r995;
add.ftz.f32 %f92, %f90, %f91;
mov.b32 %r996, %f92;
shfl.sync.bfly.b32 %r997|%p10, %r996, %r984, %r986, %r988;
mov.b32 %f93, %r997;
add.ftz.f32 %f94, %f92, %f93;
mov.b32 %r998, %f94;
mov.u32 %r999, 1;
shfl.sync.bfly.b32 %r1000|%p11, %r998, %r999, %r986, %r988;
mov.b32 %f95, %r1000;
add.ftz.f32 %f25, %f94, %f95;
setp.eq.s32 %p12, %r40, 0;
setp.lt.u32 %p13, %r3, %r25;
and.pred %p1, %p12, %p13;
mul.wide.u32 %rd78, %r980, 4;
add.s64 %rd15, %rd1, %rd78;
not.pred %p14, %p1;
@%p14 bra $L__BB219_8;

st.global.f32 [%rd15], %f25;

$L__BB219_8:
ld.shared.f32 %f96, [%r20+128];
add.ftz.f32 %f97, %f96, %f215;
ld.shared.f32 %f98, [%r20+1152];
add.ftz.f32 %f99, %f98, %f97;
ld.shared.f32 %f100, [%r20+2176];
add.ftz.f32 %f101, %f100, %f99;
mov.b32 %r1001, %f101;
shfl.sync.bfly.b32 %r1005|%p15, %r1001, %r987, %r986, %r988;
mov.b32 %f102, %r1005;
add.ftz.f32 %f103, %f101, %f102;
mov.b32 %r1006, %f103;
shfl.sync.bfly.b32 %r1008|%p16, %r1006, %r991, %r986, %r988;
mov.b32 %f104, %r1008;
add.ftz.f32 %f105, %f103, %f104;
mov.b32 %r1009, %f105;
shfl.sync.bfly.b32 %r1011|%p17, %r1009, %r994, %r986, %r988;
mov.b32 %f106, %r1011;
add.ftz.f32 %f107, %f105, %f106;
mov.b32 %r1012, %f107;
shfl.sync.bfly.b32 %r1014|%p18, %r1012, %r984, %r986, %r988;
mov.b32 %f108, %r1014;
add.ftz.f32 %f109, %f107, %f108;
mov.b32 %r1015, %f109;
shfl.sync.bfly.b32 %r1017|%p19, %r1015, %r999, %r986, %r988;
mov.b32 %f110, %r1017;
add.ftz.f32 %f26, %f109, %f110;
setp.eq.s32 %p20, %r40, 1;
or.b32 %r1021, %r3, 1;
setp.lt.u32 %p21, %r1021, %r25;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB219_10;

st.global.f32 [%rd15+4], %f26;

$L__BB219_10:
ld.shared.f32 %f111, [%r20+256];
add.ftz.f32 %f112, %f111, %f214;
ld.shared.f32 %f113, [%r20+1280];
add.ftz.f32 %f114, %f113, %f112;
ld.shared.f32 %f115, [%r20+2304];
add.ftz.f32 %f116, %f115, %f114;
mov.b32 %r1022, %f116;
mov.u32 %r1023, 31;
mov.u32 %r1024, 16;
mov.u32 %r1025, -1;
shfl.sync.bfly.b32 %r1026|%p23, %r1022, %r1024, %r1023, %r1025;
mov.b32 %f117, %r1026;
add.ftz.f32 %f118, %f116, %f117;
mov.b32 %r1027, %f118;
mov.u32 %r1028, 8;
shfl.sync.bfly.b32 %r1029|%p24, %r1027, %r1028, %r1023, %r1025;
mov.b32 %f119, %r1029;
add.ftz.f32 %f120, %f118, %f119;
mov.b32 %r1030, %f120;
mov.u32 %r1031, 4;
shfl.sync.bfly.b32 %r1032|%p25, %r1030, %r1031, %r1023, %r1025;
mov.b32 %f121, %r1032;
add.ftz.f32 %f122, %f120, %f121;
mov.b32 %r1033, %f122;
mov.u32 %r1034, 2;
shfl.sync.bfly.b32 %r1035|%p26, %r1033, %r1034, %r1023, %r1025;
mov.b32 %f123, %r1035;
add.ftz.f32 %f124, %f122, %f123;
mov.b32 %r1036, %f124;
mov.u32 %r1037, 1;
shfl.sync.bfly.b32 %r1038|%p27, %r1036, %r1037, %r1023, %r1025;
mov.b32 %f125, %r1038;
add.ftz.f32 %f27, %f124, %f125;
@%p14 bra $L__BB219_12;

cvt.u64.u32 %rd79, %r25;
add.s64 %rd80, %rd79, %rd14;
shl.b64 %rd81, %rd80, 2;
add.s64 %rd82, %rd1, %rd81;
st.global.f32 [%rd82], %f27;

$L__BB219_12:
ld.shared.f32 %f126, [%r20+384];
add.ftz.f32 %f127, %f126, %f213;
ld.shared.f32 %f128, [%r20+1408];
add.ftz.f32 %f129, %f128, %f127;
ld.shared.f32 %f130, [%r20+2432];
add.ftz.f32 %f131, %f130, %f129;
mov.b32 %r1039, %f131;
shfl.sync.bfly.b32 %r1043|%p29, %r1039, %r1024, %r1023, %r1025;
mov.b32 %f132, %r1043;
add.ftz.f32 %f133, %f131, %f132;
mov.b32 %r1044, %f133;
shfl.sync.bfly.b32 %r1046|%p30, %r1044, %r1028, %r1023, %r1025;
mov.b32 %f134, %r1046;
add.ftz.f32 %f135, %f133, %f134;
mov.b32 %r1047, %f135;
shfl.sync.bfly.b32 %r1049|%p31, %r1047, %r1031, %r1023, %r1025;
mov.b32 %f136, %r1049;
add.ftz.f32 %f137, %f135, %f136;
mov.b32 %r1050, %f137;
shfl.sync.bfly.b32 %r1052|%p32, %r1050, %r1034, %r1023, %r1025;
mov.b32 %f138, %r1052;
add.ftz.f32 %f139, %f137, %f138;
mov.b32 %r1053, %f139;
shfl.sync.bfly.b32 %r1055|%p33, %r1053, %r1037, %r1023, %r1025;
mov.b32 %f140, %r1055;
add.ftz.f32 %f28, %f139, %f140;
@%p22 bra $L__BB219_14;

add.s32 %r1056, %r25, 1;
cvt.u64.u32 %rd83, %r1056;
add.s64 %rd84, %rd83, %rd14;
shl.b64 %rd85, %rd84, 2;
add.s64 %rd86, %rd1, %rd85;
st.global.f32 [%rd86], %f28;

$L__BB219_14:
shl.b32 %r1057, %r25, 1;
mov.u32 %r1058, 1;
ld.shared.f32 %f141, [%r20+512];
add.ftz.f32 %f142, %f141, %f212;
ld.shared.f32 %f143, [%r20+1536];
add.ftz.f32 %f144, %f143, %f142;
ld.shared.f32 %f145, [%r20+2560];
add.ftz.f32 %f146, %f145, %f144;
mov.b32 %r1059, %f146;
mov.u32 %r1060, 31;
mov.u32 %r1061, 16;
mov.u32 %r1062, -1;
shfl.sync.bfly.b32 %r1063|%p35, %r1059, %r1061, %r1060, %r1062;
mov.b32 %f147, %r1063;
add.ftz.f32 %f148, %f146, %f147;
mov.b32 %r1064, %f148;
mov.u32 %r1065, 8;
shfl.sync.bfly.b32 %r1066|%p36, %r1064, %r1065, %r1060, %r1062;
mov.b32 %f149, %r1066;
add.ftz.f32 %f150, %f148, %f149;
mov.b32 %r1067, %f150;
mov.u32 %r1068, 4;
shfl.sync.bfly.b32 %r1069|%p37, %r1067, %r1068, %r1060, %r1062;
mov.b32 %f151, %r1069;
add.ftz.f32 %f152, %f150, %f151;
mov.b32 %r1070, %f152;
mov.u32 %r1071, 2;
shfl.sync.bfly.b32 %r1072|%p38, %r1070, %r1071, %r1060, %r1062;
mov.b32 %f153, %r1072;
add.ftz.f32 %f154, %f152, %f153;
mov.b32 %r1073, %f154;
shfl.sync.bfly.b32 %r1074|%p39, %r1073, %r1058, %r1060, %r1062;
mov.b32 %f155, %r1074;
add.ftz.f32 %f29, %f154, %f155;
cvt.u64.u32 %rd87, %r1057;
add.s64 %rd88, %rd87, %rd14;
shl.b64 %rd89, %rd88, 2;
add.s64 %rd16, %rd1, %rd89;
@%p14 bra $L__BB219_16;

st.global.f32 [%rd16], %f29;

$L__BB219_16:
ld.shared.f32 %f156, [%r20+640];
add.ftz.f32 %f157, %f156, %f211;
ld.shared.f32 %f158, [%r20+1664];
add.ftz.f32 %f159, %f158, %f157;
ld.shared.f32 %f160, [%r20+2688];
add.ftz.f32 %f161, %f160, %f159;
mov.b32 %r1075, %f161;
shfl.sync.bfly.b32 %r1079|%p41, %r1075, %r1061, %r1060, %r1062;
mov.b32 %f162, %r1079;
add.ftz.f32 %f163, %f161, %f162;
mov.b32 %r1080, %f163;
shfl.sync.bfly.b32 %r1082|%p42, %r1080, %r1065, %r1060, %r1062;
mov.b32 %f164, %r1082;
add.ftz.f32 %f165, %f163, %f164;
mov.b32 %r1083, %f165;
shfl.sync.bfly.b32 %r1085|%p43, %r1083, %r1068, %r1060, %r1062;
mov.b32 %f166, %r1085;
add.ftz.f32 %f167, %f165, %f166;
mov.b32 %r1086, %f167;
shfl.sync.bfly.b32 %r1088|%p44, %r1086, %r1071, %r1060, %r1062;
mov.b32 %f168, %r1088;
add.ftz.f32 %f169, %f167, %f168;
mov.b32 %r1089, %f169;
shfl.sync.bfly.b32 %r1091|%p45, %r1089, %r1058, %r1060, %r1062;
mov.b32 %f170, %r1091;
add.ftz.f32 %f30, %f169, %f170;
@%p22 bra $L__BB219_18;

st.global.f32 [%rd16+4], %f30;

$L__BB219_18:
mul.lo.s32 %r22, %r25, 3;
ld.shared.f32 %f171, [%r20+768];
add.ftz.f32 %f172, %f171, %f210;
ld.shared.f32 %f173, [%r20+1792];
add.ftz.f32 %f174, %f173, %f172;
ld.shared.f32 %f175, [%r20+2816];
add.ftz.f32 %f176, %f175, %f174;
mov.b32 %r1092, %f176;
mov.u32 %r1093, 31;
mov.u32 %r1094, 16;
mov.u32 %r1095, -1;
shfl.sync.bfly.b32 %r1096|%p47, %r1092, %r1094, %r1093, %r1095;
mov.b32 %f177, %r1096;
add.ftz.f32 %f178, %f176, %f177;
mov.b32 %r1097, %f178;
mov.u32 %r1098, 8;
shfl.sync.bfly.b32 %r1099|%p48, %r1097, %r1098, %r1093, %r1095;
mov.b32 %f179, %r1099;
add.ftz.f32 %f180, %f178, %f179;
mov.b32 %r1100, %f180;
mov.u32 %r1101, 4;
shfl.sync.bfly.b32 %r1102|%p49, %r1100, %r1101, %r1093, %r1095;
mov.b32 %f181, %r1102;
add.ftz.f32 %f182, %f180, %f181;
mov.b32 %r1103, %f182;
mov.u32 %r1104, 2;
shfl.sync.bfly.b32 %r1105|%p50, %r1103, %r1104, %r1093, %r1095;
mov.b32 %f183, %r1105;
add.ftz.f32 %f184, %f182, %f183;
mov.b32 %r1106, %f184;
mov.u32 %r1107, 1;
shfl.sync.bfly.b32 %r1108|%p51, %r1106, %r1107, %r1093, %r1095;
mov.b32 %f185, %r1108;
add.ftz.f32 %f31, %f184, %f185;
@%p14 bra $L__BB219_20;

cvt.u64.u32 %rd90, %r22;
add.s64 %rd91, %rd90, %rd14;
shl.b64 %rd92, %rd91, 2;
add.s64 %rd93, %rd1, %rd92;
st.global.f32 [%rd93], %f31;

$L__BB219_20:
ld.shared.f32 %f186, [%r20+896];
add.ftz.f32 %f187, %f186, %f209;
ld.shared.f32 %f188, [%r20+1920];
add.ftz.f32 %f189, %f188, %f187;
ld.shared.f32 %f190, [%r20+2944];
add.ftz.f32 %f191, %f190, %f189;
mov.b32 %r1109, %f191;
shfl.sync.bfly.b32 %r1113|%p53, %r1109, %r1094, %r1093, %r1095;
mov.b32 %f192, %r1113;
add.ftz.f32 %f193, %f191, %f192;
mov.b32 %r1114, %f193;
shfl.sync.bfly.b32 %r1116|%p54, %r1114, %r1098, %r1093, %r1095;
mov.b32 %f194, %r1116;
add.ftz.f32 %f195, %f193, %f194;
mov.b32 %r1117, %f195;
shfl.sync.bfly.b32 %r1119|%p55, %r1117, %r1101, %r1093, %r1095;
mov.b32 %f196, %r1119;
add.ftz.f32 %f197, %f195, %f196;
mov.b32 %r1120, %f197;
shfl.sync.bfly.b32 %r1122|%p56, %r1120, %r1104, %r1093, %r1095;
mov.b32 %f198, %r1122;
add.ftz.f32 %f199, %f197, %f198;
mov.b32 %r1123, %f199;
shfl.sync.bfly.b32 %r1125|%p57, %r1123, %r1107, %r1093, %r1095;
mov.b32 %f200, %r1125;
add.ftz.f32 %f32, %f199, %f200;
@%p22 bra $L__BB219_22;

add.s32 %r1126, %r22, 1;
cvt.u64.u32 %rd94, %r1126;
add.s64 %rd95, %rd94, %rd14;
shl.b64 %rd96, %rd95, 2;
add.s64 %rd97, %rd1, %rd96;
st.global.f32 [%rd97], %f32;

$L__BB219_22:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type17ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type17ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type17ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type17ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type17ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type17ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type17ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type17ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type17ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<71>;
.reg .b16 %rs<43>;
.reg .f32 %f<231>;
.reg .b32 %r<1352>;
.reg .b64 %rd<114>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type17ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1280];

ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type17ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type17ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type17ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type17ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type17ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type17ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type17ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type17ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type17ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type17ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type17ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type17ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type17ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type17ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type17ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type17ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type17ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r42, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r42, %r2;
mov.u32 %r43, %ctaid.x;
shl.b32 %r4, %r43, 1;
shr.s32 %r44, %r3, 31;
shr.u32 %r45, %r44, 29;
add.s32 %r46, %r3, %r45;
shr.s32 %r1351, %r46, 3;
shr.u32 %r6, %r41, 8;
setp.ge.s32 %p3, %r1351, %r6;
mov.f32 %f221, 0f00000000;
mov.f32 %f222, %f221;
mov.f32 %f223, %f221;
mov.f32 %f224, %f221;
mov.f32 %f225, %f221;
mov.f32 %f226, %f221;
mov.f32 %f227, %f221;
mov.f32 %f228, %f221;
mov.f32 %f229, %f221;
mov.f32 %f230, %f221;
@%p3 bra $L__BB220_3;

mov.u32 %r47, %ctaid.y;
mov.u32 %r48, %ctaid.z;
mul.lo.s32 %r49, %r47, %r33;
cvt.u64.u32 %rd20, %r49;
mul.lo.s32 %r50, %r48, %r39;
cvt.u64.u32 %rd21, %r50;
add.s64 %rd1, %rd21, %rd20;
mul.hi.u32 %r51, %r48, %r35;
add.s32 %r52, %r51, %r48;
shr.u32 %r53, %r52, %r36;
mul.lo.s32 %r54, %r53, %r38;
mul.hi.u32 %r55, %r47, %r29;
add.s32 %r56, %r55, %r47;
shr.u32 %r57, %r56, %r30;
mul.lo.s32 %r58, %r57, %r32;
cvta.to.global.u64 %rd113, %rd17;
and.b32 %r62, %r46, -8;
sub.s32 %r63, %r3, %r62;
shl.b32 %r64, %r63, 2;
cvt.s64.s32 %rd3, %r63;
shl.b32 %r65, %r27, 1;
shl.b32 %r1346, %r1351, 3;
add.s32 %r1350, %r1346, %r65;
mad.lo.s32 %r1349, %r27, 3, %r1346;
shl.b32 %r66, %r27, 2;
add.s32 %r1348, %r1346, %r66;
add.s32 %r1347, %r27, %r1346;
add.s32 %r67, %r1351, %r54;
add.s32 %r68, %r67, %r58;
add.s32 %r69, %r4, 1;
mad.lo.s32 %r70, %r26, %r69, %r68;
mul.wide.s32 %rd4, %r70, 74;
add.s64 %rd22, %rd4, %rd3;
add.s64 %rd5, %rd22, 66;
mad.lo.s32 %r71, %r4, %r26, %r1351;
add.s32 %r72, %r71, %r54;
add.s32 %r73, %r72, %r58;
mul.wide.s32 %rd6, %r73, 74;
add.s64 %rd23, %rd6, %rd3;
add.s64 %rd7, %rd23, 66;
mul.wide.s32 %rd24, %r64, 2;
add.s64 %rd25, %rd4, %rd24;
add.s64 %rd8, %rd25, 4;
add.s64 %rd26, %rd6, %rd24;
add.s64 %rd9, %rd26, 8;
cvta.to.global.u64 %rd10, %rd18;
mov.f32 %f221, 0f00000000;

$L__BB220_2:
cvt.u64.u32 %rd27, %r1346;
add.s64 %rd28, %rd1, %rd27;
add.s64 %rd29, %rd28, %rd3;
mul.lo.s64 %rd30, %rd29, 36;
add.s64 %rd31, %rd113, %rd9;
ld.global.nc.u16 %rs11, [%rd31+-4];
ld.global.nc.u16 %rs12, [%rd31+-2];
ld.global.nc.u16 %rs13, [%rd31];
add.s64 %rd32, %rd113, %rd7;
ld.global.nc.u8 %rs14, [%rd32];
and.b16 %rs15, %rs14, 240;
and.b16 %rs16, %rs14, 15;
shr.u16 %rs17, %rs15, 4;
ld.global.nc.u16 %rs18, [%rd31+-6];
shl.b16 %rs19, %rs18, 3;
cvt.u64.u16 %rd33, %rs19;
and.b64 %rd34, %rd33, 4088;
mov.u64 %rd35, iq2xs_grid;
add.s64 %rd36, %rd35, %rd34;
ld.global.nc.v2.u32 {%r564, %r565}, [%rd36];
not.b32 %r567, %r564;
mov.u32 %r546, 0;
not.b32 %r569, %r565;
shr.u16 %rs20, %rs18, 9;
cvt.u32.u16 %r570, %rs20;
popc.b32 %r571, %r570;
shl.b32 %r572, %r571, 7;
and.b32 %r573, %r572, 128;
or.b32 %r574, %r573, %r570;
mul.lo.s32 %r575, %r574, 16843009;
and.b32 %r576, %r575, 134480385;
or.b32 %r577, %r576, -2139062144;
add.s32 %r578, %r577, -16843009;
or.b32 %r466, %r578, %r576;
mov.u32 %r74, %r466;

prmt.b32 %r74, %r74, 0, 0xba98;

xor.b32 %r579, %r74, %r564;
or.b32 %r580, %r579, -2139062144;
and.b32 %r581, %r74, 2139062143;
and.b32 %r582, %r567, -2139062144;
sub.s32 %r583, %r580, %r581;
xor.b32 %r79, %r582, %r583;
add.s64 %rd37, %rd10, %rd30;
ld.global.nc.u32 %r80, [%rd37+4];
and.b32 %r584, %r575, -2143281136;
or.b32 %r585, %r584, -2139062144;
add.s32 %r586, %r585, -16843009;
or.b32 %r468, %r586, %r584;
mov.u32 %r76, %r468;

prmt.b32 %r76, %r76, 0, 0xba98;

xor.b32 %r587, %r76, %r565;
or.b32 %r588, %r587, -2139062144;
and.b32 %r589, %r76, 2139062143;
and.b32 %r590, %r569, -2139062144;
sub.s32 %r591, %r588, %r589;
xor.b32 %r83, %r590, %r591;
ld.global.nc.u32 %r84, [%rd37+8];

dp4a.s32.s32 %r78, %r79, %r80, %r546;


dp4a.s32.s32 %r82, %r83, %r84, %r78;

shl.b16 %rs21, %rs11, 3;
cvt.u64.u16 %rd38, %rs21;
and.b64 %rd39, %rd38, 4088;
add.s64 %rd40, %rd35, %rd39;
ld.global.nc.v2.u32 {%r592, %r593}, [%rd40];
not.b32 %r595, %r592;
not.b32 %r597, %r593;
shr.u16 %rs22, %rs11, 9;
cvt.u32.u16 %r598, %rs22;
popc.b32 %r599, %r598;
shl.b32 %r600, %r599, 7;
and.b32 %r601, %r600, 128;
or.b32 %r602, %r601, %r598;
mul.lo.s32 %r603, %r602, 16843009;
and.b32 %r604, %r603, 134480385;
or.b32 %r605, %r604, -2139062144;
add.s32 %r606, %r605, -16843009;
or.b32 %r478, %r606, %r604;
mov.u32 %r86, %r478;

prmt.b32 %r86, %r86, 0, 0xba98;

xor.b32 %r607, %r86, %r592;
or.b32 %r608, %r607, -2139062144;
and.b32 %r609, %r86, 2139062143;
and.b32 %r610, %r595, -2139062144;
sub.s32 %r611, %r608, %r609;
xor.b32 %r91, %r610, %r611;
ld.global.nc.u32 %r92, [%rd37+12];
and.b32 %r612, %r603, -2143281136;
or.b32 %r613, %r612, -2139062144;
add.s32 %r614, %r613, -16843009;
or.b32 %r480, %r614, %r612;
mov.u32 %r88, %r480;

prmt.b32 %r88, %r88, 0, 0xba98;

xor.b32 %r615, %r88, %r593;
or.b32 %r616, %r615, -2139062144;
and.b32 %r617, %r88, 2139062143;
and.b32 %r618, %r597, -2139062144;
sub.s32 %r619, %r616, %r617;
xor.b32 %r95, %r618, %r619;
ld.global.nc.u32 %r96, [%rd37+16];

dp4a.s32.s32 %r90, %r91, %r92, %r82;


dp4a.s32.s32 %r94, %r95, %r96, %r90;

shl.b16 %rs23, %rs12, 3;
cvt.u64.u16 %rd41, %rs23;
and.b64 %rd42, %rd41, 4088;
add.s64 %rd43, %rd35, %rd42;
ld.global.nc.v2.u32 {%r620, %r621}, [%rd43];
not.b32 %r623, %r620;
not.b32 %r625, %r621;
shr.u16 %rs24, %rs12, 9;
cvt.u32.u16 %r626, %rs24;
popc.b32 %r627, %r626;
shl.b32 %r628, %r627, 7;
and.b32 %r629, %r628, 128;
or.b32 %r630, %r629, %r626;
mul.lo.s32 %r631, %r630, 16843009;
and.b32 %r632, %r631, 134480385;
or.b32 %r633, %r632, -2139062144;
add.s32 %r634, %r633, -16843009;
or.b32 %r490, %r634, %r632;
mov.u32 %r98, %r490;

prmt.b32 %r98, %r98, 0, 0xba98;

xor.b32 %r635, %r98, %r620;
or.b32 %r636, %r635, -2139062144;
and.b32 %r637, %r98, 2139062143;
and.b32 %r638, %r623, -2139062144;
sub.s32 %r639, %r636, %r637;
xor.b32 %r103, %r638, %r639;
ld.global.nc.u32 %r104, [%rd37+20];
and.b32 %r640, %r631, -2143281136;
or.b32 %r641, %r640, -2139062144;
add.s32 %r642, %r641, -16843009;
or.b32 %r492, %r642, %r640;
mov.u32 %r100, %r492;

prmt.b32 %r100, %r100, 0, 0xba98;

xor.b32 %r643, %r100, %r621;
or.b32 %r644, %r643, -2139062144;
and.b32 %r645, %r100, 2139062143;
and.b32 %r646, %r625, -2139062144;
sub.s32 %r647, %r644, %r645;
xor.b32 %r107, %r646, %r647;
ld.global.nc.u32 %r108, [%rd37+24];

dp4a.s32.s32 %r102, %r103, %r104, %r546;


dp4a.s32.s32 %r106, %r107, %r108, %r102;

shl.b16 %rs25, %rs13, 3;
cvt.u64.u16 %rd44, %rs25;
and.b64 %rd45, %rd44, 4088;
add.s64 %rd46, %rd35, %rd45;
ld.global.nc.v2.u32 {%r648, %r649}, [%rd46];
not.b32 %r651, %r648;
not.b32 %r653, %r649;
shr.u16 %rs26, %rs13, 9;
cvt.u32.u16 %r654, %rs26;
popc.b32 %r655, %r654;
shl.b32 %r656, %r655, 7;
and.b32 %r657, %r656, 128;
or.b32 %r658, %r657, %r654;
mul.lo.s32 %r659, %r658, 16843009;
and.b32 %r660, %r659, 134480385;
or.b32 %r661, %r660, -2139062144;
add.s32 %r662, %r661, -16843009;
or.b32 %r502, %r662, %r660;
mov.u32 %r110, %r502;

prmt.b32 %r110, %r110, 0, 0xba98;

xor.b32 %r663, %r110, %r648;
or.b32 %r664, %r663, -2139062144;
and.b32 %r665, %r110, 2139062143;
and.b32 %r666, %r651, -2139062144;
sub.s32 %r667, %r664, %r665;
xor.b32 %r115, %r666, %r667;
ld.global.nc.u32 %r116, [%rd37+28];
and.b32 %r668, %r659, -2143281136;
or.b32 %r669, %r668, -2139062144;
add.s32 %r670, %r669, -16843009;
or.b32 %r504, %r670, %r668;
mov.u32 %r112, %r504;

prmt.b32 %r112, %r112, 0, 0xba98;

xor.b32 %r671, %r112, %r649;
or.b32 %r672, %r671, -2139062144;
and.b32 %r673, %r112, 2139062143;
and.b32 %r674, %r653, -2139062144;
sub.s32 %r675, %r672, %r673;
xor.b32 %r119, %r674, %r675;
ld.global.nc.u32 %r120, [%rd37+32];

dp4a.s32.s32 %r114, %r115, %r116, %r106;


dp4a.s32.s32 %r118, %r119, %r120, %r114;

cvt.u32.u16 %r676, %rs16;
cvt.u32.u16 %r677, %rs17;
mul.lo.s32 %r678, %r94, %r676;
mad.lo.s32 %r679, %r118, %r677, %r678;
add.s32 %r680, %r118, %r94;
shr.u32 %r681, %r680, 31;
add.s32 %r682, %r680, %r681;
shr.s32 %r683, %r682, 1;
add.s32 %r684, %r679, %r683;
shr.s32 %r685, %r684, 31;
shr.u32 %r686, %r685, 30;
add.s32 %r687, %r684, %r686;
shr.s32 %r688, %r687, 2;
add.s64 %rd47, %rd113, %rd6;
ld.global.nc.u16 %rs1, [%rd47];

{ cvt.f32.f16 %f61, %rs1;}


ld.global.nc.u32 %r122, [%rd37];

{.reg .f16 low,high;
mov.b32 {low,high},%r122;
cvt.f32.f16 %f62, low;}


mul.ftz.f32 %f81, %f61, %f62;
cvt.rn.f32.s32 %f82, %r688;
fma.rn.ftz.f32 %f230, %f81, %f82, %f230;
add.s64 %rd48, %rd113, %rd8;
ld.global.nc.u16 %rs27, [%rd48];
ld.global.nc.u16 %rs28, [%rd48+2];
ld.global.nc.u16 %rs29, [%rd48+4];
add.s64 %rd49, %rd113, %rd5;
ld.global.nc.u8 %rs30, [%rd49];
and.b16 %rs31, %rs30, 240;
and.b16 %rs32, %rs30, 15;
shr.u16 %rs33, %rs31, 4;
ld.global.nc.u16 %rs34, [%rd48+-2];
shl.b16 %rs35, %rs34, 3;
cvt.u64.u16 %rd50, %rs35;
and.b64 %rd51, %rd50, 4088;
add.s64 %rd52, %rd35, %rd51;
ld.global.nc.v2.u32 {%r689, %r690}, [%rd52];
not.b32 %r692, %r689;
not.b32 %r694, %r690;
shr.u16 %rs36, %rs34, 9;
cvt.u32.u16 %r695, %rs36;
popc.b32 %r696, %r695;
shl.b32 %r697, %r696, 7;
and.b32 %r698, %r697, 128;
or.b32 %r699, %r698, %r695;
mul.lo.s32 %r700, %r699, 16843009;
and.b32 %r701, %r700, 134480385;
or.b32 %r702, %r701, -2139062144;
add.s32 %r703, %r702, -16843009;
or.b32 %r515, %r703, %r701;
mov.u32 %r123, %r515;

prmt.b32 %r123, %r123, 0, 0xba98;

xor.b32 %r704, %r123, %r689;
or.b32 %r705, %r704, -2139062144;
and.b32 %r706, %r123, 2139062143;
and.b32 %r707, %r692, -2139062144;
sub.s32 %r708, %r705, %r706;
xor.b32 %r128, %r707, %r708;
and.b32 %r709, %r700, -2143281136;
or.b32 %r710, %r709, -2139062144;
add.s32 %r711, %r710, -16843009;
or.b32 %r517, %r711, %r709;
mov.u32 %r125, %r517;

prmt.b32 %r125, %r125, 0, 0xba98;

xor.b32 %r712, %r125, %r690;
or.b32 %r713, %r712, -2139062144;
and.b32 %r714, %r125, 2139062143;
and.b32 %r715, %r694, -2139062144;
sub.s32 %r716, %r713, %r714;
xor.b32 %r132, %r715, %r716;

dp4a.s32.s32 %r127, %r128, %r80, %r546;


dp4a.s32.s32 %r131, %r132, %r84, %r127;

shl.b16 %rs37, %rs27, 3;
cvt.u64.u16 %rd53, %rs37;
and.b64 %rd54, %rd53, 4088;
add.s64 %rd55, %rd35, %rd54;
ld.global.nc.v2.u32 {%r717, %r718}, [%rd55];
not.b32 %r720, %r717;
not.b32 %r722, %r718;
shr.u16 %rs38, %rs27, 9;
cvt.u32.u16 %r723, %rs38;
popc.b32 %r724, %r723;
shl.b32 %r725, %r724, 7;
and.b32 %r726, %r725, 128;
or.b32 %r727, %r726, %r723;
mul.lo.s32 %r728, %r727, 16843009;
and.b32 %r729, %r728, 134480385;
or.b32 %r730, %r729, -2139062144;
add.s32 %r731, %r730, -16843009;
or.b32 %r527, %r731, %r729;
mov.u32 %r135, %r527;

prmt.b32 %r135, %r135, 0, 0xba98;

xor.b32 %r732, %r135, %r717;
or.b32 %r733, %r732, -2139062144;
and.b32 %r734, %r135, 2139062143;
and.b32 %r735, %r720, -2139062144;
sub.s32 %r736, %r733, %r734;
xor.b32 %r140, %r735, %r736;
and.b32 %r737, %r728, -2143281136;
or.b32 %r738, %r737, -2139062144;
add.s32 %r739, %r738, -16843009;
or.b32 %r529, %r739, %r737;
mov.u32 %r137, %r529;

prmt.b32 %r137, %r137, 0, 0xba98;

xor.b32 %r740, %r137, %r718;
or.b32 %r741, %r740, -2139062144;
and.b32 %r742, %r137, 2139062143;
and.b32 %r743, %r722, -2139062144;
sub.s32 %r744, %r741, %r742;
xor.b32 %r144, %r743, %r744;

dp4a.s32.s32 %r139, %r140, %r92, %r131;


dp4a.s32.s32 %r143, %r144, %r96, %r139;

shl.b16 %rs39, %rs28, 3;
cvt.u64.u16 %rd56, %rs39;
and.b64 %rd57, %rd56, 4088;
add.s64 %rd58, %rd35, %rd57;
ld.global.nc.v2.u32 {%r745, %r746}, [%rd58];
not.b32 %r748, %r745;
not.b32 %r750, %r746;
shr.u16 %rs40, %rs28, 9;
cvt.u32.u16 %r751, %rs40;
popc.b32 %r752, %r751;
shl.b32 %r753, %r752, 7;
and.b32 %r754, %r753, 128;
or.b32 %r755, %r754, %r751;
mul.lo.s32 %r756, %r755, 16843009;
and.b32 %r757, %r756, 134480385;
or.b32 %r758, %r757, -2139062144;
add.s32 %r759, %r758, -16843009;
or.b32 %r539, %r759, %r757;
mov.u32 %r147, %r539;

prmt.b32 %r147, %r147, 0, 0xba98;

xor.b32 %r760, %r147, %r745;
or.b32 %r761, %r760, -2139062144;
and.b32 %r762, %r147, 2139062143;
and.b32 %r763, %r748, -2139062144;
sub.s32 %r764, %r761, %r762;
xor.b32 %r152, %r763, %r764;
and.b32 %r765, %r756, -2143281136;
or.b32 %r766, %r765, -2139062144;
add.s32 %r767, %r766, -16843009;
or.b32 %r541, %r767, %r765;
mov.u32 %r149, %r541;

prmt.b32 %r149, %r149, 0, 0xba98;

xor.b32 %r768, %r149, %r746;
or.b32 %r769, %r768, -2139062144;
and.b32 %r770, %r149, 2139062143;
and.b32 %r771, %r750, -2139062144;
sub.s32 %r772, %r769, %r770;
xor.b32 %r156, %r771, %r772;

dp4a.s32.s32 %r151, %r152, %r104, %r546;


dp4a.s32.s32 %r155, %r156, %r108, %r151;

shl.b16 %rs41, %rs29, 3;
cvt.u64.u16 %rd59, %rs41;
and.b64 %rd60, %rd59, 4088;
add.s64 %rd61, %rd35, %rd60;
ld.global.nc.v2.u32 {%r773, %r774}, [%rd61];
not.b32 %r776, %r773;
not.b32 %r778, %r774;
shr.u16 %rs42, %rs29, 9;
cvt.u32.u16 %r779, %rs42;
popc.b32 %r780, %r779;
shl.b32 %r781, %r780, 7;
and.b32 %r782, %r781, 128;
or.b32 %r783, %r782, %r779;
mul.lo.s32 %r784, %r783, 16843009;
and.b32 %r785, %r784, 134480385;
or.b32 %r786, %r785, -2139062144;
add.s32 %r787, %r786, -16843009;
or.b32 %r551, %r787, %r785;
mov.u32 %r159, %r551;

prmt.b32 %r159, %r159, 0, 0xba98;

xor.b32 %r788, %r159, %r773;
or.b32 %r789, %r788, -2139062144;
and.b32 %r790, %r159, 2139062143;
and.b32 %r791, %r776, -2139062144;
sub.s32 %r792, %r789, %r790;
xor.b32 %r164, %r791, %r792;
and.b32 %r793, %r784, -2143281136;
or.b32 %r794, %r793, -2139062144;
add.s32 %r795, %r794, -16843009;
or.b32 %r553, %r795, %r793;
mov.u32 %r161, %r553;

prmt.b32 %r161, %r161, 0, 0xba98;

xor.b32 %r796, %r161, %r774;
or.b32 %r797, %r796, -2139062144;
and.b32 %r798, %r161, 2139062143;
and.b32 %r799, %r778, -2139062144;
sub.s32 %r800, %r797, %r798;
xor.b32 %r168, %r799, %r800;

dp4a.s32.s32 %r163, %r164, %r116, %r155;


dp4a.s32.s32 %r167, %r168, %r120, %r163;

cvt.u32.u16 %r801, %rs32;
cvt.u32.u16 %r802, %rs33;
mul.lo.s32 %r803, %r143, %r801;
mad.lo.s32 %r804, %r167, %r802, %r803;
add.s32 %r805, %r167, %r143;
shr.u32 %r806, %r805, 31;
add.s32 %r807, %r805, %r806;
shr.s32 %r808, %r807, 1;
add.s32 %r809, %r804, %r808;
shr.s32 %r810, %r809, 31;
shr.u32 %r811, %r810, 30;
add.s32 %r812, %r809, %r811;
shr.s32 %r813, %r812, 2;
add.s64 %rd62, %rd113, %rd4;
ld.global.nc.u16 %rs2, [%rd62];

{ cvt.f32.f16 %f63, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r122;
cvt.f32.f16 %f64, low;}


mul.ftz.f32 %f83, %f63, %f64;
cvt.rn.f32.s32 %f84, %r813;
fma.rn.ftz.f32 %f229, %f83, %f84, %f229;
cvt.u64.u32 %rd63, %r1347;
add.s64 %rd64, %rd1, %rd63;
add.s64 %rd65, %rd64, %rd3;
mul.lo.s64 %rd66, %rd65, 36;
mov.u32 %r172, %r466;

prmt.b32 %r172, %r172, 0, 0xba98;

xor.b32 %r814, %r172, %r564;
or.b32 %r815, %r814, -2139062144;
and.b32 %r816, %r172, 2139062143;
sub.s32 %r817, %r815, %r816;
xor.b32 %r177, %r582, %r817;
add.s64 %rd67, %rd10, %rd66;
ld.global.nc.u32 %r178, [%rd67+4];
mov.u32 %r174, %r468;

prmt.b32 %r174, %r174, 0, 0xba98;

xor.b32 %r818, %r174, %r565;
or.b32 %r819, %r818, -2139062144;
and.b32 %r820, %r174, 2139062143;
sub.s32 %r821, %r819, %r820;
xor.b32 %r181, %r590, %r821;
ld.global.nc.u32 %r182, [%rd67+8];

dp4a.s32.s32 %r176, %r177, %r178, %r546;


dp4a.s32.s32 %r180, %r181, %r182, %r176;

mov.u32 %r184, %r478;

prmt.b32 %r184, %r184, 0, 0xba98;

xor.b32 %r822, %r184, %r592;
or.b32 %r823, %r822, -2139062144;
and.b32 %r824, %r184, 2139062143;
sub.s32 %r825, %r823, %r824;
xor.b32 %r189, %r610, %r825;
ld.global.nc.u32 %r190, [%rd67+12];
mov.u32 %r186, %r480;

prmt.b32 %r186, %r186, 0, 0xba98;

xor.b32 %r826, %r186, %r593;
or.b32 %r827, %r826, -2139062144;
and.b32 %r828, %r186, 2139062143;
sub.s32 %r829, %r827, %r828;
xor.b32 %r193, %r618, %r829;
ld.global.nc.u32 %r194, [%rd67+16];

dp4a.s32.s32 %r188, %r189, %r190, %r180;


dp4a.s32.s32 %r192, %r193, %r194, %r188;

mov.u32 %r196, %r490;

prmt.b32 %r196, %r196, 0, 0xba98;

xor.b32 %r830, %r196, %r620;
or.b32 %r831, %r830, -2139062144;
and.b32 %r832, %r196, 2139062143;
sub.s32 %r833, %r831, %r832;
xor.b32 %r201, %r638, %r833;
ld.global.nc.u32 %r202, [%rd67+20];
mov.u32 %r198, %r492;

prmt.b32 %r198, %r198, 0, 0xba98;

xor.b32 %r834, %r198, %r621;
or.b32 %r835, %r834, -2139062144;
and.b32 %r836, %r198, 2139062143;
sub.s32 %r837, %r835, %r836;
xor.b32 %r205, %r646, %r837;
ld.global.nc.u32 %r206, [%rd67+24];

dp4a.s32.s32 %r200, %r201, %r202, %r546;


dp4a.s32.s32 %r204, %r205, %r206, %r200;

mov.u32 %r208, %r502;

prmt.b32 %r208, %r208, 0, 0xba98;

xor.b32 %r838, %r208, %r648;
or.b32 %r839, %r838, -2139062144;
and.b32 %r840, %r208, 2139062143;
sub.s32 %r841, %r839, %r840;
xor.b32 %r213, %r666, %r841;
ld.global.nc.u32 %r214, [%rd67+28];
mov.u32 %r210, %r504;

prmt.b32 %r210, %r210, 0, 0xba98;

xor.b32 %r842, %r210, %r649;
or.b32 %r843, %r842, -2139062144;
and.b32 %r844, %r210, 2139062143;
sub.s32 %r845, %r843, %r844;
xor.b32 %r217, %r674, %r845;
ld.global.nc.u32 %r218, [%rd67+32];

dp4a.s32.s32 %r212, %r213, %r214, %r204;


dp4a.s32.s32 %r216, %r217, %r218, %r212;

mul.lo.s32 %r846, %r192, %r676;
mad.lo.s32 %r847, %r216, %r677, %r846;
add.s32 %r848, %r216, %r192;
shr.u32 %r849, %r848, 31;
add.s32 %r850, %r848, %r849;
shr.s32 %r851, %r850, 1;
add.s32 %r852, %r847, %r851;
shr.s32 %r853, %r852, 31;
shr.u32 %r854, %r853, 30;
add.s32 %r855, %r852, %r854;
shr.s32 %r856, %r855, 2;

{ cvt.f32.f16 %f65, %rs1;}


ld.global.nc.u32 %r220, [%rd67];

{.reg .f16 low,high;
mov.b32 {low,high},%r220;
cvt.f32.f16 %f66, low;}


mul.ftz.f32 %f85, %f65, %f66;
cvt.rn.f32.s32 %f86, %r856;
fma.rn.ftz.f32 %f228, %f85, %f86, %f228;
mov.u32 %r221, %r515;

prmt.b32 %r221, %r221, 0, 0xba98;

xor.b32 %r857, %r221, %r689;
or.b32 %r858, %r857, -2139062144;
and.b32 %r859, %r221, 2139062143;
sub.s32 %r860, %r858, %r859;
xor.b32 %r226, %r707, %r860;
mov.u32 %r223, %r517;

prmt.b32 %r223, %r223, 0, 0xba98;

xor.b32 %r861, %r223, %r690;
or.b32 %r862, %r861, -2139062144;
and.b32 %r863, %r223, 2139062143;
sub.s32 %r864, %r862, %r863;
xor.b32 %r230, %r715, %r864;

dp4a.s32.s32 %r225, %r226, %r178, %r546;


dp4a.s32.s32 %r229, %r230, %r182, %r225;

mov.u32 %r233, %r527;

prmt.b32 %r233, %r233, 0, 0xba98;

xor.b32 %r865, %r233, %r717;
or.b32 %r866, %r865, -2139062144;
and.b32 %r867, %r233, 2139062143;
sub.s32 %r868, %r866, %r867;
xor.b32 %r238, %r735, %r868;
mov.u32 %r235, %r529;

prmt.b32 %r235, %r235, 0, 0xba98;

xor.b32 %r869, %r235, %r718;
or.b32 %r870, %r869, -2139062144;
and.b32 %r871, %r235, 2139062143;
sub.s32 %r872, %r870, %r871;
xor.b32 %r242, %r743, %r872;

dp4a.s32.s32 %r237, %r238, %r190, %r229;


dp4a.s32.s32 %r241, %r242, %r194, %r237;

mov.u32 %r245, %r539;

prmt.b32 %r245, %r245, 0, 0xba98;

xor.b32 %r873, %r245, %r745;
or.b32 %r874, %r873, -2139062144;
and.b32 %r875, %r245, 2139062143;
sub.s32 %r876, %r874, %r875;
xor.b32 %r250, %r763, %r876;
mov.u32 %r247, %r541;

prmt.b32 %r247, %r247, 0, 0xba98;

xor.b32 %r877, %r247, %r746;
or.b32 %r878, %r877, -2139062144;
and.b32 %r879, %r247, 2139062143;
sub.s32 %r880, %r878, %r879;
xor.b32 %r254, %r771, %r880;

dp4a.s32.s32 %r249, %r250, %r202, %r546;


dp4a.s32.s32 %r253, %r254, %r206, %r249;

mov.u32 %r257, %r551;

prmt.b32 %r257, %r257, 0, 0xba98;

xor.b32 %r881, %r257, %r773;
or.b32 %r882, %r881, -2139062144;
and.b32 %r883, %r257, 2139062143;
sub.s32 %r884, %r882, %r883;
xor.b32 %r262, %r791, %r884;
mov.u32 %r259, %r553;

prmt.b32 %r259, %r259, 0, 0xba98;

xor.b32 %r885, %r259, %r774;
or.b32 %r886, %r885, -2139062144;
and.b32 %r887, %r259, 2139062143;
sub.s32 %r888, %r886, %r887;
xor.b32 %r266, %r799, %r888;

dp4a.s32.s32 %r261, %r262, %r214, %r253;


dp4a.s32.s32 %r265, %r266, %r218, %r261;

mul.lo.s32 %r889, %r241, %r801;
mad.lo.s32 %r890, %r265, %r802, %r889;
add.s32 %r891, %r265, %r241;
shr.u32 %r892, %r891, 31;
add.s32 %r893, %r891, %r892;
shr.s32 %r894, %r893, 1;
add.s32 %r895, %r890, %r894;
shr.s32 %r896, %r895, 31;
shr.u32 %r897, %r896, 30;
add.s32 %r898, %r895, %r897;
shr.s32 %r899, %r898, 2;

{ cvt.f32.f16 %f67, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r220;
cvt.f32.f16 %f68, low;}


mul.ftz.f32 %f87, %f67, %f68;
cvt.rn.f32.s32 %f88, %r899;
fma.rn.ftz.f32 %f227, %f87, %f88, %f227;
cvt.u64.u32 %rd68, %r1350;
add.s64 %rd69, %rd1, %rd68;
add.s64 %rd70, %rd69, %rd3;
mul.lo.s64 %rd71, %rd70, 36;
mov.u32 %r270, %r466;

prmt.b32 %r270, %r270, 0, 0xba98;

xor.b32 %r900, %r270, %r564;
or.b32 %r901, %r900, -2139062144;
and.b32 %r902, %r270, 2139062143;
sub.s32 %r903, %r901, %r902;
xor.b32 %r275, %r582, %r903;
add.s64 %rd72, %rd10, %rd71;
ld.global.nc.u32 %r276, [%rd72+4];
mov.u32 %r272, %r468;

prmt.b32 %r272, %r272, 0, 0xba98;

xor.b32 %r904, %r272, %r565;
or.b32 %r905, %r904, -2139062144;
and.b32 %r906, %r272, 2139062143;
sub.s32 %r907, %r905, %r906;
xor.b32 %r279, %r590, %r907;
ld.global.nc.u32 %r280, [%rd72+8];

dp4a.s32.s32 %r274, %r275, %r276, %r546;


dp4a.s32.s32 %r278, %r279, %r280, %r274;

mov.u32 %r282, %r478;

prmt.b32 %r282, %r282, 0, 0xba98;

xor.b32 %r908, %r282, %r592;
or.b32 %r909, %r908, -2139062144;
and.b32 %r910, %r282, 2139062143;
sub.s32 %r911, %r909, %r910;
xor.b32 %r287, %r610, %r911;
ld.global.nc.u32 %r288, [%rd72+12];
mov.u32 %r284, %r480;

prmt.b32 %r284, %r284, 0, 0xba98;

xor.b32 %r912, %r284, %r593;
or.b32 %r913, %r912, -2139062144;
and.b32 %r914, %r284, 2139062143;
sub.s32 %r915, %r913, %r914;
xor.b32 %r291, %r618, %r915;
ld.global.nc.u32 %r292, [%rd72+16];

dp4a.s32.s32 %r286, %r287, %r288, %r278;


dp4a.s32.s32 %r290, %r291, %r292, %r286;

mov.u32 %r294, %r490;

prmt.b32 %r294, %r294, 0, 0xba98;

xor.b32 %r916, %r294, %r620;
or.b32 %r917, %r916, -2139062144;
and.b32 %r918, %r294, 2139062143;
sub.s32 %r919, %r917, %r918;
xor.b32 %r299, %r638, %r919;
ld.global.nc.u32 %r300, [%rd72+20];
mov.u32 %r296, %r492;

prmt.b32 %r296, %r296, 0, 0xba98;

xor.b32 %r920, %r296, %r621;
or.b32 %r921, %r920, -2139062144;
and.b32 %r922, %r296, 2139062143;
sub.s32 %r923, %r921, %r922;
xor.b32 %r303, %r646, %r923;
ld.global.nc.u32 %r304, [%rd72+24];

dp4a.s32.s32 %r298, %r299, %r300, %r546;


dp4a.s32.s32 %r302, %r303, %r304, %r298;

mov.u32 %r306, %r502;

prmt.b32 %r306, %r306, 0, 0xba98;

xor.b32 %r924, %r306, %r648;
or.b32 %r925, %r924, -2139062144;
and.b32 %r926, %r306, 2139062143;
sub.s32 %r927, %r925, %r926;
xor.b32 %r311, %r666, %r927;
ld.global.nc.u32 %r312, [%rd72+28];
mov.u32 %r308, %r504;

prmt.b32 %r308, %r308, 0, 0xba98;

xor.b32 %r928, %r308, %r649;
or.b32 %r929, %r928, -2139062144;
and.b32 %r930, %r308, 2139062143;
sub.s32 %r931, %r929, %r930;
xor.b32 %r315, %r674, %r931;
ld.global.nc.u32 %r316, [%rd72+32];

dp4a.s32.s32 %r310, %r311, %r312, %r302;


dp4a.s32.s32 %r314, %r315, %r316, %r310;

mul.lo.s32 %r932, %r290, %r676;
mad.lo.s32 %r933, %r314, %r677, %r932;
add.s32 %r934, %r314, %r290;
shr.u32 %r935, %r934, 31;
add.s32 %r936, %r934, %r935;
shr.s32 %r937, %r936, 1;
add.s32 %r938, %r933, %r937;
shr.s32 %r939, %r938, 31;
shr.u32 %r940, %r939, 30;
add.s32 %r941, %r938, %r940;
shr.s32 %r942, %r941, 2;

{ cvt.f32.f16 %f69, %rs1;}


ld.global.nc.u32 %r318, [%rd72];

{.reg .f16 low,high;
mov.b32 {low,high},%r318;
cvt.f32.f16 %f70, low;}


mul.ftz.f32 %f89, %f69, %f70;
cvt.rn.f32.s32 %f90, %r942;
fma.rn.ftz.f32 %f226, %f89, %f90, %f226;
mov.u32 %r319, %r515;

prmt.b32 %r319, %r319, 0, 0xba98;

xor.b32 %r943, %r319, %r689;
or.b32 %r944, %r943, -2139062144;
and.b32 %r945, %r319, 2139062143;
sub.s32 %r946, %r944, %r945;
xor.b32 %r324, %r707, %r946;
mov.u32 %r321, %r517;

prmt.b32 %r321, %r321, 0, 0xba98;

xor.b32 %r947, %r321, %r690;
or.b32 %r948, %r947, -2139062144;
and.b32 %r949, %r321, 2139062143;
sub.s32 %r950, %r948, %r949;
xor.b32 %r328, %r715, %r950;

dp4a.s32.s32 %r323, %r324, %r276, %r546;


dp4a.s32.s32 %r327, %r328, %r280, %r323;

mov.u32 %r331, %r527;

prmt.b32 %r331, %r331, 0, 0xba98;

xor.b32 %r951, %r331, %r717;
or.b32 %r952, %r951, -2139062144;
and.b32 %r953, %r331, 2139062143;
sub.s32 %r954, %r952, %r953;
xor.b32 %r336, %r735, %r954;
mov.u32 %r333, %r529;

prmt.b32 %r333, %r333, 0, 0xba98;

xor.b32 %r955, %r333, %r718;
or.b32 %r956, %r955, -2139062144;
and.b32 %r957, %r333, 2139062143;
sub.s32 %r958, %r956, %r957;
xor.b32 %r340, %r743, %r958;

dp4a.s32.s32 %r335, %r336, %r288, %r327;


dp4a.s32.s32 %r339, %r340, %r292, %r335;

mov.u32 %r343, %r539;

prmt.b32 %r343, %r343, 0, 0xba98;

xor.b32 %r959, %r343, %r745;
or.b32 %r960, %r959, -2139062144;
and.b32 %r961, %r343, 2139062143;
sub.s32 %r962, %r960, %r961;
xor.b32 %r348, %r763, %r962;
mov.u32 %r345, %r541;

prmt.b32 %r345, %r345, 0, 0xba98;

xor.b32 %r963, %r345, %r746;
or.b32 %r964, %r963, -2139062144;
and.b32 %r965, %r345, 2139062143;
sub.s32 %r966, %r964, %r965;
xor.b32 %r352, %r771, %r966;

dp4a.s32.s32 %r347, %r348, %r300, %r546;


dp4a.s32.s32 %r351, %r352, %r304, %r347;

mov.u32 %r355, %r551;

prmt.b32 %r355, %r355, 0, 0xba98;

xor.b32 %r967, %r355, %r773;
or.b32 %r968, %r967, -2139062144;
and.b32 %r969, %r355, 2139062143;
sub.s32 %r970, %r968, %r969;
xor.b32 %r360, %r791, %r970;
mov.u32 %r357, %r553;

prmt.b32 %r357, %r357, 0, 0xba98;

xor.b32 %r971, %r357, %r774;
or.b32 %r972, %r971, -2139062144;
and.b32 %r973, %r357, 2139062143;
sub.s32 %r974, %r972, %r973;
xor.b32 %r364, %r799, %r974;

dp4a.s32.s32 %r359, %r360, %r312, %r351;


dp4a.s32.s32 %r363, %r364, %r316, %r359;

mul.lo.s32 %r975, %r339, %r801;
mad.lo.s32 %r976, %r363, %r802, %r975;
add.s32 %r977, %r363, %r339;
shr.u32 %r978, %r977, 31;
add.s32 %r979, %r977, %r978;
shr.s32 %r980, %r979, 1;
add.s32 %r981, %r976, %r980;
shr.s32 %r982, %r981, 31;
shr.u32 %r983, %r982, 30;
add.s32 %r984, %r981, %r983;
shr.s32 %r985, %r984, 2;

{ cvt.f32.f16 %f71, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r318;
cvt.f32.f16 %f72, low;}


mul.ftz.f32 %f91, %f71, %f72;
cvt.rn.f32.s32 %f92, %r985;
fma.rn.ftz.f32 %f225, %f91, %f92, %f225;
cvt.u64.u32 %rd73, %r1349;
add.s64 %rd74, %rd1, %rd73;
add.s64 %rd75, %rd74, %rd3;
mul.lo.s64 %rd76, %rd75, 36;
mov.u32 %r368, %r466;

prmt.b32 %r368, %r368, 0, 0xba98;

xor.b32 %r986, %r368, %r564;
or.b32 %r987, %r986, -2139062144;
and.b32 %r988, %r368, 2139062143;
sub.s32 %r989, %r987, %r988;
xor.b32 %r373, %r582, %r989;
add.s64 %rd77, %rd10, %rd76;
ld.global.nc.u32 %r374, [%rd77+4];
mov.u32 %r370, %r468;

prmt.b32 %r370, %r370, 0, 0xba98;

xor.b32 %r990, %r370, %r565;
or.b32 %r991, %r990, -2139062144;
and.b32 %r992, %r370, 2139062143;
sub.s32 %r993, %r991, %r992;
xor.b32 %r377, %r590, %r993;
ld.global.nc.u32 %r378, [%rd77+8];

dp4a.s32.s32 %r372, %r373, %r374, %r546;


dp4a.s32.s32 %r376, %r377, %r378, %r372;

mov.u32 %r380, %r478;

prmt.b32 %r380, %r380, 0, 0xba98;

xor.b32 %r994, %r380, %r592;
or.b32 %r995, %r994, -2139062144;
and.b32 %r996, %r380, 2139062143;
sub.s32 %r997, %r995, %r996;
xor.b32 %r385, %r610, %r997;
ld.global.nc.u32 %r386, [%rd77+12];
mov.u32 %r382, %r480;

prmt.b32 %r382, %r382, 0, 0xba98;

xor.b32 %r998, %r382, %r593;
or.b32 %r999, %r998, -2139062144;
and.b32 %r1000, %r382, 2139062143;
sub.s32 %r1001, %r999, %r1000;
xor.b32 %r389, %r618, %r1001;
ld.global.nc.u32 %r390, [%rd77+16];

dp4a.s32.s32 %r384, %r385, %r386, %r376;


dp4a.s32.s32 %r388, %r389, %r390, %r384;

mov.u32 %r392, %r490;

prmt.b32 %r392, %r392, 0, 0xba98;

xor.b32 %r1002, %r392, %r620;
or.b32 %r1003, %r1002, -2139062144;
and.b32 %r1004, %r392, 2139062143;
sub.s32 %r1005, %r1003, %r1004;
xor.b32 %r397, %r638, %r1005;
ld.global.nc.u32 %r398, [%rd77+20];
mov.u32 %r394, %r492;

prmt.b32 %r394, %r394, 0, 0xba98;

xor.b32 %r1006, %r394, %r621;
or.b32 %r1007, %r1006, -2139062144;
and.b32 %r1008, %r394, 2139062143;
sub.s32 %r1009, %r1007, %r1008;
xor.b32 %r401, %r646, %r1009;
ld.global.nc.u32 %r402, [%rd77+24];

dp4a.s32.s32 %r396, %r397, %r398, %r546;


dp4a.s32.s32 %r400, %r401, %r402, %r396;

mov.u32 %r404, %r502;

prmt.b32 %r404, %r404, 0, 0xba98;

xor.b32 %r1010, %r404, %r648;
or.b32 %r1011, %r1010, -2139062144;
and.b32 %r1012, %r404, 2139062143;
sub.s32 %r1013, %r1011, %r1012;
xor.b32 %r409, %r666, %r1013;
ld.global.nc.u32 %r410, [%rd77+28];
mov.u32 %r406, %r504;

prmt.b32 %r406, %r406, 0, 0xba98;

xor.b32 %r1014, %r406, %r649;
or.b32 %r1015, %r1014, -2139062144;
and.b32 %r1016, %r406, 2139062143;
sub.s32 %r1017, %r1015, %r1016;
xor.b32 %r413, %r674, %r1017;
ld.global.nc.u32 %r414, [%rd77+32];

dp4a.s32.s32 %r408, %r409, %r410, %r400;


dp4a.s32.s32 %r412, %r413, %r414, %r408;

mul.lo.s32 %r1018, %r388, %r676;
mad.lo.s32 %r1019, %r412, %r677, %r1018;
add.s32 %r1020, %r412, %r388;
shr.u32 %r1021, %r1020, 31;
add.s32 %r1022, %r1020, %r1021;
shr.s32 %r1023, %r1022, 1;
add.s32 %r1024, %r1019, %r1023;
shr.s32 %r1025, %r1024, 31;
shr.u32 %r1026, %r1025, 30;
add.s32 %r1027, %r1024, %r1026;
shr.s32 %r1028, %r1027, 2;

{ cvt.f32.f16 %f73, %rs1;}


ld.global.nc.u32 %r416, [%rd77];

{.reg .f16 low,high;
mov.b32 {low,high},%r416;
cvt.f32.f16 %f74, low;}


mul.ftz.f32 %f93, %f73, %f74;
cvt.rn.f32.s32 %f94, %r1028;
fma.rn.ftz.f32 %f224, %f93, %f94, %f224;
mov.u32 %r417, %r515;

prmt.b32 %r417, %r417, 0, 0xba98;

xor.b32 %r1029, %r417, %r689;
or.b32 %r1030, %r1029, -2139062144;
and.b32 %r1031, %r417, 2139062143;
sub.s32 %r1032, %r1030, %r1031;
xor.b32 %r422, %r707, %r1032;
mov.u32 %r419, %r517;

prmt.b32 %r419, %r419, 0, 0xba98;

xor.b32 %r1033, %r419, %r690;
or.b32 %r1034, %r1033, -2139062144;
and.b32 %r1035, %r419, 2139062143;
sub.s32 %r1036, %r1034, %r1035;
xor.b32 %r426, %r715, %r1036;

dp4a.s32.s32 %r421, %r422, %r374, %r546;


dp4a.s32.s32 %r425, %r426, %r378, %r421;

mov.u32 %r429, %r527;

prmt.b32 %r429, %r429, 0, 0xba98;

xor.b32 %r1037, %r429, %r717;
or.b32 %r1038, %r1037, -2139062144;
and.b32 %r1039, %r429, 2139062143;
sub.s32 %r1040, %r1038, %r1039;
xor.b32 %r434, %r735, %r1040;
mov.u32 %r431, %r529;

prmt.b32 %r431, %r431, 0, 0xba98;

xor.b32 %r1041, %r431, %r718;
or.b32 %r1042, %r1041, -2139062144;
and.b32 %r1043, %r431, 2139062143;
sub.s32 %r1044, %r1042, %r1043;
xor.b32 %r438, %r743, %r1044;

dp4a.s32.s32 %r433, %r434, %r386, %r425;


dp4a.s32.s32 %r437, %r438, %r390, %r433;

mov.u32 %r441, %r539;

prmt.b32 %r441, %r441, 0, 0xba98;

xor.b32 %r1045, %r441, %r745;
or.b32 %r1046, %r1045, -2139062144;
and.b32 %r1047, %r441, 2139062143;
sub.s32 %r1048, %r1046, %r1047;
xor.b32 %r446, %r763, %r1048;
mov.u32 %r443, %r541;

prmt.b32 %r443, %r443, 0, 0xba98;

xor.b32 %r1049, %r443, %r746;
or.b32 %r1050, %r1049, -2139062144;
and.b32 %r1051, %r443, 2139062143;
sub.s32 %r1052, %r1050, %r1051;
xor.b32 %r450, %r771, %r1052;

dp4a.s32.s32 %r445, %r446, %r398, %r546;


dp4a.s32.s32 %r449, %r450, %r402, %r445;

mov.u32 %r453, %r551;

prmt.b32 %r453, %r453, 0, 0xba98;

xor.b32 %r1053, %r453, %r773;
or.b32 %r1054, %r1053, -2139062144;
and.b32 %r1055, %r453, 2139062143;
sub.s32 %r1056, %r1054, %r1055;
xor.b32 %r458, %r791, %r1056;
mov.u32 %r455, %r553;

prmt.b32 %r455, %r455, 0, 0xba98;

xor.b32 %r1057, %r455, %r774;
or.b32 %r1058, %r1057, -2139062144;
and.b32 %r1059, %r455, 2139062143;
sub.s32 %r1060, %r1058, %r1059;
xor.b32 %r462, %r799, %r1060;

dp4a.s32.s32 %r457, %r458, %r410, %r449;


dp4a.s32.s32 %r461, %r462, %r414, %r457;

mul.lo.s32 %r1061, %r437, %r801;
mad.lo.s32 %r1062, %r461, %r802, %r1061;
add.s32 %r1063, %r461, %r437;
shr.u32 %r1064, %r1063, 31;
add.s32 %r1065, %r1063, %r1064;
shr.s32 %r1066, %r1065, 1;
add.s32 %r1067, %r1062, %r1066;
shr.s32 %r1068, %r1067, 31;
shr.u32 %r1069, %r1068, 30;
add.s32 %r1070, %r1067, %r1069;
shr.s32 %r1071, %r1070, 2;

{ cvt.f32.f16 %f75, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r416;
cvt.f32.f16 %f76, low;}


mul.ftz.f32 %f95, %f75, %f76;
cvt.rn.f32.s32 %f96, %r1071;
fma.rn.ftz.f32 %f223, %f95, %f96, %f223;
cvt.u64.u32 %rd78, %r1348;
add.s64 %rd79, %rd1, %rd78;
add.s64 %rd80, %rd79, %rd3;
mul.lo.s64 %rd81, %rd80, 36;

prmt.b32 %r466, %r466, 0, 0xba98;

xor.b32 %r1072, %r466, %r564;
or.b32 %r1073, %r1072, -2139062144;
and.b32 %r1074, %r466, 2139062143;
sub.s32 %r1075, %r1073, %r1074;
xor.b32 %r471, %r582, %r1075;
add.s64 %rd82, %rd10, %rd81;
ld.global.nc.u32 %r472, [%rd82+4];

prmt.b32 %r468, %r468, 0, 0xba98;

xor.b32 %r1076, %r468, %r565;
or.b32 %r1077, %r1076, -2139062144;
and.b32 %r1078, %r468, 2139062143;
sub.s32 %r1079, %r1077, %r1078;
xor.b32 %r475, %r590, %r1079;
ld.global.nc.u32 %r476, [%rd82+8];

dp4a.s32.s32 %r470, %r471, %r472, %r546;


dp4a.s32.s32 %r474, %r475, %r476, %r470;


prmt.b32 %r478, %r478, 0, 0xba98;

xor.b32 %r1080, %r478, %r592;
or.b32 %r1081, %r1080, -2139062144;
and.b32 %r1082, %r478, 2139062143;
sub.s32 %r1083, %r1081, %r1082;
xor.b32 %r483, %r610, %r1083;
ld.global.nc.u32 %r484, [%rd82+12];

prmt.b32 %r480, %r480, 0, 0xba98;

xor.b32 %r1084, %r480, %r593;
or.b32 %r1085, %r1084, -2139062144;
and.b32 %r1086, %r480, 2139062143;
sub.s32 %r1087, %r1085, %r1086;
xor.b32 %r487, %r618, %r1087;
ld.global.nc.u32 %r488, [%rd82+16];

dp4a.s32.s32 %r482, %r483, %r484, %r474;


dp4a.s32.s32 %r486, %r487, %r488, %r482;


prmt.b32 %r490, %r490, 0, 0xba98;

xor.b32 %r1088, %r490, %r620;
or.b32 %r1089, %r1088, -2139062144;
and.b32 %r1090, %r490, 2139062143;
sub.s32 %r1091, %r1089, %r1090;
xor.b32 %r495, %r638, %r1091;
ld.global.nc.u32 %r496, [%rd82+20];

prmt.b32 %r492, %r492, 0, 0xba98;

xor.b32 %r1092, %r492, %r621;
or.b32 %r1093, %r1092, -2139062144;
and.b32 %r1094, %r492, 2139062143;
sub.s32 %r1095, %r1093, %r1094;
xor.b32 %r499, %r646, %r1095;
ld.global.nc.u32 %r500, [%rd82+24];

dp4a.s32.s32 %r494, %r495, %r496, %r546;


dp4a.s32.s32 %r498, %r499, %r500, %r494;


prmt.b32 %r502, %r502, 0, 0xba98;

xor.b32 %r1096, %r502, %r648;
or.b32 %r1097, %r1096, -2139062144;
and.b32 %r1098, %r502, 2139062143;
sub.s32 %r1099, %r1097, %r1098;
xor.b32 %r507, %r666, %r1099;
ld.global.nc.u32 %r508, [%rd82+28];

prmt.b32 %r504, %r504, 0, 0xba98;

xor.b32 %r1100, %r504, %r649;
or.b32 %r1101, %r1100, -2139062144;
and.b32 %r1102, %r504, 2139062143;
sub.s32 %r1103, %r1101, %r1102;
xor.b32 %r511, %r674, %r1103;
ld.global.nc.u32 %r512, [%rd82+32];

dp4a.s32.s32 %r506, %r507, %r508, %r498;


dp4a.s32.s32 %r510, %r511, %r512, %r506;

mul.lo.s32 %r1104, %r486, %r676;
mad.lo.s32 %r1105, %r510, %r677, %r1104;
add.s32 %r1106, %r510, %r486;
shr.u32 %r1107, %r1106, 31;
add.s32 %r1108, %r1106, %r1107;
shr.s32 %r1109, %r1108, 1;
add.s32 %r1110, %r1105, %r1109;
shr.s32 %r1111, %r1110, 31;
shr.u32 %r1112, %r1111, 30;
add.s32 %r1113, %r1110, %r1112;
shr.s32 %r1114, %r1113, 2;

{ cvt.f32.f16 %f77, %rs1;}


ld.global.nc.u32 %r514, [%rd82];

{.reg .f16 low,high;
mov.b32 {low,high},%r514;
cvt.f32.f16 %f78, low;}


mul.ftz.f32 %f97, %f77, %f78;
cvt.rn.f32.s32 %f98, %r1114;
fma.rn.ftz.f32 %f222, %f97, %f98, %f222;

prmt.b32 %r515, %r515, 0, 0xba98;

xor.b32 %r1115, %r515, %r689;
or.b32 %r1116, %r1115, -2139062144;
and.b32 %r1117, %r515, 2139062143;
sub.s32 %r1118, %r1116, %r1117;
xor.b32 %r520, %r707, %r1118;

prmt.b32 %r517, %r517, 0, 0xba98;

xor.b32 %r1119, %r517, %r690;
or.b32 %r1120, %r1119, -2139062144;
and.b32 %r1121, %r517, 2139062143;
sub.s32 %r1122, %r1120, %r1121;
xor.b32 %r524, %r715, %r1122;

dp4a.s32.s32 %r519, %r520, %r472, %r546;


dp4a.s32.s32 %r523, %r524, %r476, %r519;


prmt.b32 %r527, %r527, 0, 0xba98;

xor.b32 %r1123, %r527, %r717;
or.b32 %r1124, %r1123, -2139062144;
and.b32 %r1125, %r527, 2139062143;
sub.s32 %r1126, %r1124, %r1125;
xor.b32 %r532, %r735, %r1126;

prmt.b32 %r529, %r529, 0, 0xba98;

xor.b32 %r1127, %r529, %r718;
or.b32 %r1128, %r1127, -2139062144;
and.b32 %r1129, %r529, 2139062143;
sub.s32 %r1130, %r1128, %r1129;
xor.b32 %r536, %r743, %r1130;

dp4a.s32.s32 %r531, %r532, %r484, %r523;


dp4a.s32.s32 %r535, %r536, %r488, %r531;


prmt.b32 %r539, %r539, 0, 0xba98;

xor.b32 %r1131, %r539, %r745;
or.b32 %r1132, %r1131, -2139062144;
and.b32 %r1133, %r539, 2139062143;
sub.s32 %r1134, %r1132, %r1133;
xor.b32 %r544, %r763, %r1134;

prmt.b32 %r541, %r541, 0, 0xba98;

xor.b32 %r1135, %r541, %r746;
or.b32 %r1136, %r1135, -2139062144;
and.b32 %r1137, %r541, 2139062143;
sub.s32 %r1138, %r1136, %r1137;
xor.b32 %r548, %r771, %r1138;

dp4a.s32.s32 %r543, %r544, %r496, %r546;


dp4a.s32.s32 %r547, %r548, %r500, %r543;


prmt.b32 %r551, %r551, 0, 0xba98;

xor.b32 %r1139, %r551, %r773;
or.b32 %r1140, %r1139, -2139062144;
and.b32 %r1141, %r551, 2139062143;
sub.s32 %r1142, %r1140, %r1141;
xor.b32 %r556, %r791, %r1142;

prmt.b32 %r553, %r553, 0, 0xba98;

xor.b32 %r1143, %r553, %r774;
or.b32 %r1144, %r1143, -2139062144;
and.b32 %r1145, %r553, 2139062143;
sub.s32 %r1146, %r1144, %r1145;
xor.b32 %r560, %r799, %r1146;

dp4a.s32.s32 %r555, %r556, %r508, %r547;


dp4a.s32.s32 %r559, %r560, %r512, %r555;

mul.lo.s32 %r1147, %r535, %r801;
mad.lo.s32 %r1148, %r559, %r802, %r1147;
add.s32 %r1149, %r559, %r535;
shr.u32 %r1150, %r1149, 31;
add.s32 %r1151, %r1149, %r1150;
shr.s32 %r1152, %r1151, 1;
add.s32 %r1153, %r1148, %r1152;
shr.s32 %r1154, %r1153, 31;
shr.u32 %r1155, %r1154, 30;
add.s32 %r1156, %r1153, %r1155;
shr.s32 %r1157, %r1156, 2;

{ cvt.f32.f16 %f79, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r514;
cvt.f32.f16 %f80, low;}


mul.ftz.f32 %f99, %f79, %f80;
cvt.rn.f32.s32 %f100, %r1157;
fma.rn.ftz.f32 %f221, %f99, %f100, %f221;
add.s32 %r1350, %r1350, 64;
add.s32 %r1349, %r1349, 64;
add.s32 %r1348, %r1348, 64;
add.s32 %r1347, %r1347, 64;
add.s32 %r1346, %r1346, 64;
add.s64 %rd113, %rd113, 592;
add.s32 %r1351, %r1351, 8;
setp.lt.s32 %p4, %r1351, %r6;
@%p4 bra $L__BB220_2;

$L__BB220_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB220_5;

mov.u32 %r1158, _ZZ13mul_mat_vec_qIL9ggml_type17ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r1159, %r1, 1280, %r1158;
shl.b32 %r1160, %r2, 2;
add.s32 %r1161, %r1159, %r1160;
st.shared.f32 [%r1161+-1280], %f230;
st.shared.f32 [%r1161+-1152], %f229;
st.shared.f32 [%r1161+-1024], %f228;
st.shared.f32 [%r1161+-896], %f227;
st.shared.f32 [%r1161+-768], %f226;
st.shared.f32 [%r1161+-640], %f225;
st.shared.f32 [%r1161+-512], %f224;
st.shared.f32 [%r1161+-384], %f223;
st.shared.f32 [%r1161+-256], %f222;
st.shared.f32 [%r1161+-128], %f221;

$L__BB220_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB220_26;

mov.u32 %r1162, %ctaid.z;
mov.u32 %r1163, %ctaid.y;
mad.lo.s32 %r1164, %r1163, %r34, %r4;
mad.lo.s32 %r1165, %r1162, %r40, %r1164;
cvt.u64.u32 %rd13, %r1165;
shl.b32 %r1166, %r2, 2;
mov.u32 %r1167, _ZZ13mul_mat_vec_qIL9ggml_type17ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r24, %r1167, %r1166;
mov.u32 %r1168, 2;
ld.shared.f32 %f101, [%r24];
add.ftz.f32 %f102, %f101, %f230;
mov.b32 %r1169, %f102;
mov.u32 %r1170, 31;
mov.u32 %r1171, 16;
mov.u32 %r1172, -1;
shfl.sync.bfly.b32 %r1173|%p7, %r1169, %r1171, %r1170, %r1172;
mov.b32 %f103, %r1173;
add.ftz.f32 %f104, %f102, %f103;
mov.b32 %r1174, %f104;
mov.u32 %r1175, 8;
shfl.sync.bfly.b32 %r1176|%p8, %r1174, %r1175, %r1170, %r1172;
mov.b32 %f105, %r1176;
add.ftz.f32 %f106, %f104, %f105;
mov.b32 %r1177, %f106;
mov.u32 %r1178, 4;
shfl.sync.bfly.b32 %r1179|%p9, %r1177, %r1178, %r1170, %r1172;
mov.b32 %f107, %r1179;
add.ftz.f32 %f108, %f106, %f107;
mov.b32 %r1180, %f108;
shfl.sync.bfly.b32 %r1181|%p10, %r1180, %r1168, %r1170, %r1172;
mov.b32 %f109, %r1181;
add.ftz.f32 %f110, %f108, %f109;
mov.b32 %r1182, %f110;
mov.u32 %r1183, 1;
shfl.sync.bfly.b32 %r1184|%p11, %r1182, %r1183, %r1170, %r1172;
mov.b32 %f111, %r1184;
add.ftz.f32 %f31, %f110, %f111;
setp.lt.u32 %p12, %r4, %r28;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
cvta.to.global.u64 %rd83, %rd19;
mul.wide.u32 %rd84, %r1165, 4;
add.s64 %rd14, %rd83, %rd84;
not.pred %p14, %p1;
@%p14 bra $L__BB220_8;

st.global.f32 [%rd14], %f31;

$L__BB220_8:
ld.shared.f32 %f112, [%r24+128];
add.ftz.f32 %f113, %f112, %f229;
mov.b32 %r1185, %f113;
shfl.sync.bfly.b32 %r1189|%p15, %r1185, %r1171, %r1170, %r1172;
mov.b32 %f114, %r1189;
add.ftz.f32 %f115, %f113, %f114;
mov.b32 %r1190, %f115;
shfl.sync.bfly.b32 %r1192|%p16, %r1190, %r1175, %r1170, %r1172;
mov.b32 %f116, %r1192;
add.ftz.f32 %f117, %f115, %f116;
mov.b32 %r1193, %f117;
shfl.sync.bfly.b32 %r1195|%p17, %r1193, %r1178, %r1170, %r1172;
mov.b32 %f118, %r1195;
add.ftz.f32 %f119, %f117, %f118;
mov.b32 %r1196, %f119;
shfl.sync.bfly.b32 %r1198|%p18, %r1196, %r1168, %r1170, %r1172;
mov.b32 %f120, %r1198;
add.ftz.f32 %f121, %f119, %f120;
mov.b32 %r1199, %f121;
shfl.sync.bfly.b32 %r1201|%p19, %r1199, %r1183, %r1170, %r1172;
mov.b32 %f122, %r1201;
add.ftz.f32 %f32, %f121, %f122;
setp.eq.s32 %p20, %r2, 1;
or.b32 %r1205, %r4, 1;
setp.lt.u32 %p21, %r1205, %r28;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB220_10;

st.global.f32 [%rd14+4], %f32;

$L__BB220_10:
ld.shared.f32 %f123, [%r24+256];
add.ftz.f32 %f124, %f123, %f228;
mov.b32 %r1206, %f124;
mov.u32 %r1207, 31;
mov.u32 %r1208, 16;
mov.u32 %r1209, -1;
shfl.sync.bfly.b32 %r1210|%p23, %r1206, %r1208, %r1207, %r1209;
mov.b32 %f125, %r1210;
add.ftz.f32 %f126, %f124, %f125;
mov.b32 %r1211, %f126;
mov.u32 %r1212, 8;
shfl.sync.bfly.b32 %r1213|%p24, %r1211, %r1212, %r1207, %r1209;
mov.b32 %f127, %r1213;
add.ftz.f32 %f128, %f126, %f127;
mov.b32 %r1214, %f128;
mov.u32 %r1215, 4;
shfl.sync.bfly.b32 %r1216|%p25, %r1214, %r1215, %r1207, %r1209;
mov.b32 %f129, %r1216;
add.ftz.f32 %f130, %f128, %f129;
mov.b32 %r1217, %f130;
mov.u32 %r1218, 2;
shfl.sync.bfly.b32 %r1219|%p26, %r1217, %r1218, %r1207, %r1209;
mov.b32 %f131, %r1219;
add.ftz.f32 %f132, %f130, %f131;
mov.b32 %r1220, %f132;
mov.u32 %r1221, 1;
shfl.sync.bfly.b32 %r1222|%p27, %r1220, %r1221, %r1207, %r1209;
mov.b32 %f133, %r1222;
add.ftz.f32 %f33, %f132, %f133;
@%p14 bra $L__BB220_12;

cvt.u64.u32 %rd85, %r28;
add.s64 %rd86, %rd85, %rd13;
shl.b64 %rd88, %rd86, 2;
add.s64 %rd89, %rd83, %rd88;
st.global.f32 [%rd89], %f33;

$L__BB220_12:
ld.shared.f32 %f134, [%r24+384];
add.ftz.f32 %f135, %f134, %f227;
mov.b32 %r1223, %f135;
shfl.sync.bfly.b32 %r1227|%p29, %r1223, %r1208, %r1207, %r1209;
mov.b32 %f136, %r1227;
add.ftz.f32 %f137, %f135, %f136;
mov.b32 %r1228, %f137;
shfl.sync.bfly.b32 %r1230|%p30, %r1228, %r1212, %r1207, %r1209;
mov.b32 %f138, %r1230;
add.ftz.f32 %f139, %f137, %f138;
mov.b32 %r1231, %f139;
shfl.sync.bfly.b32 %r1233|%p31, %r1231, %r1215, %r1207, %r1209;
mov.b32 %f140, %r1233;
add.ftz.f32 %f141, %f139, %f140;
mov.b32 %r1234, %f141;
shfl.sync.bfly.b32 %r1236|%p32, %r1234, %r1218, %r1207, %r1209;
mov.b32 %f142, %r1236;
add.ftz.f32 %f143, %f141, %f142;
mov.b32 %r1237, %f143;
shfl.sync.bfly.b32 %r1239|%p33, %r1237, %r1221, %r1207, %r1209;
mov.b32 %f144, %r1239;
add.ftz.f32 %f34, %f143, %f144;
@%p22 bra $L__BB220_14;

add.s32 %r1240, %r28, 1;
cvt.u64.u32 %rd90, %r1240;
add.s64 %rd91, %rd90, %rd13;
shl.b64 %rd93, %rd91, 2;
add.s64 %rd94, %rd83, %rd93;
st.global.f32 [%rd94], %f34;

$L__BB220_14:
shl.b32 %r1241, %r28, 1;
mov.u32 %r1242, 1;
ld.shared.f32 %f145, [%r24+512];
add.ftz.f32 %f146, %f145, %f226;
mov.b32 %r1243, %f146;
mov.u32 %r1244, 31;
mov.u32 %r1245, 16;
mov.u32 %r1246, -1;
shfl.sync.bfly.b32 %r1247|%p35, %r1243, %r1245, %r1244, %r1246;
mov.b32 %f147, %r1247;
add.ftz.f32 %f148, %f146, %f147;
mov.b32 %r1248, %f148;
mov.u32 %r1249, 8;
shfl.sync.bfly.b32 %r1250|%p36, %r1248, %r1249, %r1244, %r1246;
mov.b32 %f149, %r1250;
add.ftz.f32 %f150, %f148, %f149;
mov.b32 %r1251, %f150;
mov.u32 %r1252, 4;
shfl.sync.bfly.b32 %r1253|%p37, %r1251, %r1252, %r1244, %r1246;
mov.b32 %f151, %r1253;
add.ftz.f32 %f152, %f150, %f151;
mov.b32 %r1254, %f152;
mov.u32 %r1255, 2;
shfl.sync.bfly.b32 %r1256|%p38, %r1254, %r1255, %r1244, %r1246;
mov.b32 %f153, %r1256;
add.ftz.f32 %f154, %f152, %f153;
mov.b32 %r1257, %f154;
shfl.sync.bfly.b32 %r1258|%p39, %r1257, %r1242, %r1244, %r1246;
mov.b32 %f155, %r1258;
add.ftz.f32 %f35, %f154, %f155;
cvt.u64.u32 %rd95, %r1241;
add.s64 %rd96, %rd95, %rd13;
shl.b64 %rd98, %rd96, 2;
add.s64 %rd15, %rd83, %rd98;
@%p14 bra $L__BB220_16;

st.global.f32 [%rd15], %f35;

$L__BB220_16:
ld.shared.f32 %f156, [%r24+640];
add.ftz.f32 %f157, %f156, %f225;
mov.b32 %r1259, %f157;
shfl.sync.bfly.b32 %r1263|%p41, %r1259, %r1245, %r1244, %r1246;
mov.b32 %f158, %r1263;
add.ftz.f32 %f159, %f157, %f158;
mov.b32 %r1264, %f159;
shfl.sync.bfly.b32 %r1266|%p42, %r1264, %r1249, %r1244, %r1246;
mov.b32 %f160, %r1266;
add.ftz.f32 %f161, %f159, %f160;
mov.b32 %r1267, %f161;
shfl.sync.bfly.b32 %r1269|%p43, %r1267, %r1252, %r1244, %r1246;
mov.b32 %f162, %r1269;
add.ftz.f32 %f163, %f161, %f162;
mov.b32 %r1270, %f163;
shfl.sync.bfly.b32 %r1272|%p44, %r1270, %r1255, %r1244, %r1246;
mov.b32 %f164, %r1272;
add.ftz.f32 %f165, %f163, %f164;
mov.b32 %r1273, %f165;
shfl.sync.bfly.b32 %r1275|%p45, %r1273, %r1242, %r1244, %r1246;
mov.b32 %f166, %r1275;
add.ftz.f32 %f36, %f165, %f166;
@%p22 bra $L__BB220_18;

st.global.f32 [%rd15+4], %f36;

$L__BB220_18:
mul.lo.s32 %r25, %r28, 3;
ld.shared.f32 %f167, [%r24+768];
add.ftz.f32 %f168, %f167, %f224;
mov.b32 %r1276, %f168;
mov.u32 %r1277, 31;
mov.u32 %r1278, 16;
mov.u32 %r1279, -1;
shfl.sync.bfly.b32 %r1280|%p47, %r1276, %r1278, %r1277, %r1279;
mov.b32 %f169, %r1280;
add.ftz.f32 %f170, %f168, %f169;
mov.b32 %r1281, %f170;
mov.u32 %r1282, 8;
shfl.sync.bfly.b32 %r1283|%p48, %r1281, %r1282, %r1277, %r1279;
mov.b32 %f171, %r1283;
add.ftz.f32 %f172, %f170, %f171;
mov.b32 %r1284, %f172;
mov.u32 %r1285, 4;
shfl.sync.bfly.b32 %r1286|%p49, %r1284, %r1285, %r1277, %r1279;
mov.b32 %f173, %r1286;
add.ftz.f32 %f174, %f172, %f173;
mov.b32 %r1287, %f174;
mov.u32 %r1288, 2;
shfl.sync.bfly.b32 %r1289|%p50, %r1287, %r1288, %r1277, %r1279;
mov.b32 %f175, %r1289;
add.ftz.f32 %f176, %f174, %f175;
mov.b32 %r1290, %f176;
mov.u32 %r1291, 1;
shfl.sync.bfly.b32 %r1292|%p51, %r1290, %r1291, %r1277, %r1279;
mov.b32 %f177, %r1292;
add.ftz.f32 %f37, %f176, %f177;
@%p14 bra $L__BB220_20;

cvt.u64.u32 %rd99, %r25;
add.s64 %rd100, %rd99, %rd13;
shl.b64 %rd102, %rd100, 2;
add.s64 %rd103, %rd83, %rd102;
st.global.f32 [%rd103], %f37;

$L__BB220_20:
ld.shared.f32 %f178, [%r24+896];
add.ftz.f32 %f179, %f178, %f223;
mov.b32 %r1293, %f179;
shfl.sync.bfly.b32 %r1297|%p53, %r1293, %r1278, %r1277, %r1279;
mov.b32 %f180, %r1297;
add.ftz.f32 %f181, %f179, %f180;
mov.b32 %r1298, %f181;
shfl.sync.bfly.b32 %r1300|%p54, %r1298, %r1282, %r1277, %r1279;
mov.b32 %f182, %r1300;
add.ftz.f32 %f183, %f181, %f182;
mov.b32 %r1301, %f183;
shfl.sync.bfly.b32 %r1303|%p55, %r1301, %r1285, %r1277, %r1279;
mov.b32 %f184, %r1303;
add.ftz.f32 %f185, %f183, %f184;
mov.b32 %r1304, %f185;
shfl.sync.bfly.b32 %r1306|%p56, %r1304, %r1288, %r1277, %r1279;
mov.b32 %f186, %r1306;
add.ftz.f32 %f187, %f185, %f186;
mov.b32 %r1307, %f187;
shfl.sync.bfly.b32 %r1309|%p57, %r1307, %r1291, %r1277, %r1279;
mov.b32 %f188, %r1309;
add.ftz.f32 %f38, %f187, %f188;
@%p22 bra $L__BB220_22;

add.s32 %r1310, %r25, 1;
cvt.u64.u32 %rd104, %r1310;
add.s64 %rd105, %rd104, %rd13;
shl.b64 %rd107, %rd105, 2;
add.s64 %rd108, %rd83, %rd107;
st.global.f32 [%rd108], %f38;

$L__BB220_22:
shl.b32 %r1311, %r28, 2;
mov.u32 %r1312, 2;
ld.shared.f32 %f189, [%r24+1024];
add.ftz.f32 %f190, %f189, %f222;
mov.b32 %r1313, %f190;
mov.u32 %r1314, 31;
mov.u32 %r1315, 16;
mov.u32 %r1316, -1;
shfl.sync.bfly.b32 %r1317|%p59, %r1313, %r1315, %r1314, %r1316;
mov.b32 %f191, %r1317;
add.ftz.f32 %f192, %f190, %f191;
mov.b32 %r1318, %f192;
mov.u32 %r1319, 8;
shfl.sync.bfly.b32 %r1320|%p60, %r1318, %r1319, %r1314, %r1316;
mov.b32 %f193, %r1320;
add.ftz.f32 %f194, %f192, %f193;
mov.b32 %r1321, %f194;
mov.u32 %r1322, 4;
shfl.sync.bfly.b32 %r1323|%p61, %r1321, %r1322, %r1314, %r1316;
mov.b32 %f195, %r1323;
add.ftz.f32 %f196, %f194, %f195;
mov.b32 %r1324, %f196;
shfl.sync.bfly.b32 %r1325|%p62, %r1324, %r1312, %r1314, %r1316;
mov.b32 %f197, %r1325;
add.ftz.f32 %f198, %f196, %f197;
mov.b32 %r1326, %f198;
mov.u32 %r1327, 1;
shfl.sync.bfly.b32 %r1328|%p63, %r1326, %r1327, %r1314, %r1316;
mov.b32 %f199, %r1328;
add.ftz.f32 %f39, %f198, %f199;
cvt.u64.u32 %rd109, %r1311;
add.s64 %rd110, %rd109, %rd13;
shl.b64 %rd112, %rd110, 2;
add.s64 %rd16, %rd83, %rd112;
@%p14 bra $L__BB220_24;

st.global.f32 [%rd16], %f39;

$L__BB220_24:
ld.shared.f32 %f200, [%r24+1152];
add.ftz.f32 %f201, %f200, %f221;
mov.b32 %r1329, %f201;
shfl.sync.bfly.b32 %r1333|%p65, %r1329, %r1315, %r1314, %r1316;
mov.b32 %f202, %r1333;
add.ftz.f32 %f203, %f201, %f202;
mov.b32 %r1334, %f203;
shfl.sync.bfly.b32 %r1336|%p66, %r1334, %r1319, %r1314, %r1316;
mov.b32 %f204, %r1336;
add.ftz.f32 %f205, %f203, %f204;
mov.b32 %r1337, %f205;
shfl.sync.bfly.b32 %r1339|%p67, %r1337, %r1322, %r1314, %r1316;
mov.b32 %f206, %r1339;
add.ftz.f32 %f207, %f205, %f206;
mov.b32 %r1340, %f207;
shfl.sync.bfly.b32 %r1342|%p68, %r1340, %r1312, %r1314, %r1316;
mov.b32 %f208, %r1342;
add.ftz.f32 %f209, %f207, %f208;
mov.b32 %r1343, %f209;
shfl.sync.bfly.b32 %r1345|%p69, %r1343, %r1327, %r1314, %r1316;
mov.b32 %f210, %r1345;
add.ftz.f32 %f40, %f209, %f210;
@%p22 bra $L__BB220_26;

st.global.f32 [%rd16+4], %f40;

$L__BB220_26:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type17ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type17ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type17ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type17ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type17ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type17ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type17ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type17ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type17ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<83>;
.reg .b16 %rs<45>;
.reg .f32 %f<277>;
.reg .b32 %r<1638>;
.reg .b64 %rd<132>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type17ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type17ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type17ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type17ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type17ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type17ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type17ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type17ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type17ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type17ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type17ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type17ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type17ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type17ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type17ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type17ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type17ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type17ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r42, %tid.y;
shl.b32 %r43, %r42, 5;
mov.u32 %r44, %tid.x;
add.s32 %r1, %r43, %r44;
shr.s32 %r45, %r1, 31;
shr.u32 %r46, %r45, 29;
add.s32 %r47, %r1, %r46;
shr.s32 %r1637, %r47, 3;
shr.u32 %r3, %r41, 8;
setp.ge.s32 %p3, %r1637, %r3;
mov.f32 %f265, 0f00000000;
mov.f32 %f266, %f265;
mov.f32 %f267, %f265;
mov.f32 %f268, %f265;
mov.f32 %f269, %f265;
mov.f32 %f270, %f265;
mov.f32 %f271, %f265;
mov.f32 %f272, %f265;
mov.f32 %f273, %f265;
mov.f32 %f274, %f265;
mov.f32 %f275, %f265;
mov.f32 %f276, %f265;
@%p3 bra $L__BB221_3;

mov.u32 %r48, %ctaid.x;
shl.b32 %r49, %r48, 1;
mov.u32 %r50, %ctaid.y;
mov.u32 %r51, %ctaid.z;
mul.lo.s32 %r52, %r50, %r33;
cvt.u64.u32 %rd19, %r52;
mul.lo.s32 %r53, %r51, %r39;
cvt.u64.u32 %rd20, %r53;
add.s64 %rd1, %rd20, %rd19;
mul.hi.u32 %r54, %r51, %r35;
add.s32 %r55, %r54, %r51;
shr.u32 %r56, %r55, %r36;
mul.lo.s32 %r57, %r56, %r38;
mul.hi.u32 %r58, %r50, %r29;
add.s32 %r59, %r58, %r50;
shr.u32 %r60, %r59, %r30;
mul.lo.s32 %r61, %r60, %r32;
cvta.to.global.u64 %rd131, %rd16;
and.b32 %r65, %r47, -8;
sub.s32 %r66, %r1, %r65;
shl.b32 %r67, %r66, 2;
cvt.s64.s32 %rd3, %r66;
shl.b32 %r68, %r27, 1;
shl.b32 %r1631, %r1637, 3;
add.s32 %r1636, %r1631, %r68;
mad.lo.s32 %r1635, %r27, 3, %r1631;
shl.b32 %r69, %r27, 2;
add.s32 %r1634, %r1631, %r69;
mad.lo.s32 %r1633, %r27, 5, %r1631;
add.s32 %r1632, %r27, %r1631;
add.s32 %r70, %r1637, %r57;
add.s32 %r71, %r70, %r61;
or.b32 %r72, %r49, 1;
mad.lo.s32 %r73, %r26, %r72, %r71;
mul.wide.s32 %rd4, %r73, 74;
add.s64 %rd21, %rd4, %rd3;
add.s64 %rd5, %rd21, 66;
mad.lo.s32 %r74, %r49, %r26, %r1637;
add.s32 %r75, %r74, %r57;
add.s32 %r76, %r75, %r61;
mul.wide.s32 %rd6, %r76, 74;
add.s64 %rd22, %rd6, %rd3;
add.s64 %rd7, %rd22, 66;
mul.wide.s32 %rd23, %r67, 2;
add.s64 %rd24, %rd4, %rd23;
add.s64 %rd8, %rd24, 4;
add.s64 %rd25, %rd6, %rd23;
add.s64 %rd9, %rd25, 8;
cvta.to.global.u64 %rd10, %rd17;
mov.f32 %f265, 0f00000000;

$L__BB221_2:
cvt.u64.u32 %rd26, %r1631;
add.s64 %rd27, %rd1, %rd26;
add.s64 %rd28, %rd27, %rd3;
mul.lo.s64 %rd29, %rd28, 36;
add.s64 %rd30, %rd131, %rd9;
ld.global.nc.u16 %rs13, [%rd30+-4];
ld.global.nc.u16 %rs14, [%rd30+-2];
ld.global.nc.u16 %rs15, [%rd30];
add.s64 %rd31, %rd131, %rd7;
ld.global.nc.u8 %rs16, [%rd31];
and.b16 %rs17, %rs16, 240;
and.b16 %rs18, %rs16, 15;
shr.u16 %rs19, %rs17, 4;
ld.global.nc.u16 %rs20, [%rd30+-6];
shl.b16 %rs21, %rs20, 3;
cvt.u64.u16 %rd32, %rs21;
and.b64 %rd33, %rd32, 4088;
mov.u64 %rd34, iq2xs_grid;
add.s64 %rd35, %rd34, %rd33;
ld.global.nc.v2.u32 {%r665, %r666}, [%rd35];
not.b32 %r668, %r665;
mov.u32 %r647, 0;
not.b32 %r670, %r666;
shr.u16 %rs22, %rs20, 9;
cvt.u32.u16 %r671, %rs22;
popc.b32 %r672, %r671;
shl.b32 %r673, %r672, 7;
and.b32 %r674, %r673, 128;
or.b32 %r675, %r674, %r671;
mul.lo.s32 %r676, %r675, 16843009;
and.b32 %r677, %r676, 134480385;
or.b32 %r678, %r677, -2139062144;
add.s32 %r679, %r678, -16843009;
or.b32 %r567, %r679, %r677;
mov.u32 %r77, %r567;

prmt.b32 %r77, %r77, 0, 0xba98;

xor.b32 %r680, %r77, %r665;
or.b32 %r681, %r680, -2139062144;
and.b32 %r682, %r77, 2139062143;
and.b32 %r683, %r668, -2139062144;
sub.s32 %r684, %r681, %r682;
xor.b32 %r82, %r683, %r684;
add.s64 %rd36, %rd10, %rd29;
ld.global.nc.u32 %r83, [%rd36+4];
and.b32 %r685, %r676, -2143281136;
or.b32 %r686, %r685, -2139062144;
add.s32 %r687, %r686, -16843009;
or.b32 %r569, %r687, %r685;
mov.u32 %r79, %r569;

prmt.b32 %r79, %r79, 0, 0xba98;

xor.b32 %r688, %r79, %r666;
or.b32 %r689, %r688, -2139062144;
and.b32 %r690, %r79, 2139062143;
and.b32 %r691, %r670, -2139062144;
sub.s32 %r692, %r689, %r690;
xor.b32 %r86, %r691, %r692;
ld.global.nc.u32 %r87, [%rd36+8];

dp4a.s32.s32 %r81, %r82, %r83, %r647;


dp4a.s32.s32 %r85, %r86, %r87, %r81;

shl.b16 %rs23, %rs13, 3;
cvt.u64.u16 %rd37, %rs23;
and.b64 %rd38, %rd37, 4088;
add.s64 %rd39, %rd34, %rd38;
ld.global.nc.v2.u32 {%r693, %r694}, [%rd39];
not.b32 %r696, %r693;
not.b32 %r698, %r694;
shr.u16 %rs24, %rs13, 9;
cvt.u32.u16 %r699, %rs24;
popc.b32 %r700, %r699;
shl.b32 %r701, %r700, 7;
and.b32 %r702, %r701, 128;
or.b32 %r703, %r702, %r699;
mul.lo.s32 %r704, %r703, 16843009;
and.b32 %r705, %r704, 134480385;
or.b32 %r706, %r705, -2139062144;
add.s32 %r707, %r706, -16843009;
or.b32 %r579, %r707, %r705;
mov.u32 %r89, %r579;

prmt.b32 %r89, %r89, 0, 0xba98;

xor.b32 %r708, %r89, %r693;
or.b32 %r709, %r708, -2139062144;
and.b32 %r710, %r89, 2139062143;
and.b32 %r711, %r696, -2139062144;
sub.s32 %r712, %r709, %r710;
xor.b32 %r94, %r711, %r712;
ld.global.nc.u32 %r95, [%rd36+12];
and.b32 %r713, %r704, -2143281136;
or.b32 %r714, %r713, -2139062144;
add.s32 %r715, %r714, -16843009;
or.b32 %r581, %r715, %r713;
mov.u32 %r91, %r581;

prmt.b32 %r91, %r91, 0, 0xba98;

xor.b32 %r716, %r91, %r694;
or.b32 %r717, %r716, -2139062144;
and.b32 %r718, %r91, 2139062143;
and.b32 %r719, %r698, -2139062144;
sub.s32 %r720, %r717, %r718;
xor.b32 %r98, %r719, %r720;
ld.global.nc.u32 %r99, [%rd36+16];

dp4a.s32.s32 %r93, %r94, %r95, %r85;


dp4a.s32.s32 %r97, %r98, %r99, %r93;

shl.b16 %rs25, %rs14, 3;
cvt.u64.u16 %rd40, %rs25;
and.b64 %rd41, %rd40, 4088;
add.s64 %rd42, %rd34, %rd41;
ld.global.nc.v2.u32 {%r721, %r722}, [%rd42];
not.b32 %r724, %r721;
not.b32 %r726, %r722;
shr.u16 %rs26, %rs14, 9;
cvt.u32.u16 %r727, %rs26;
popc.b32 %r728, %r727;
shl.b32 %r729, %r728, 7;
and.b32 %r730, %r729, 128;
or.b32 %r731, %r730, %r727;
mul.lo.s32 %r732, %r731, 16843009;
and.b32 %r733, %r732, 134480385;
or.b32 %r734, %r733, -2139062144;
add.s32 %r735, %r734, -16843009;
or.b32 %r591, %r735, %r733;
mov.u32 %r101, %r591;

prmt.b32 %r101, %r101, 0, 0xba98;

xor.b32 %r736, %r101, %r721;
or.b32 %r737, %r736, -2139062144;
and.b32 %r738, %r101, 2139062143;
and.b32 %r739, %r724, -2139062144;
sub.s32 %r740, %r737, %r738;
xor.b32 %r106, %r739, %r740;
ld.global.nc.u32 %r107, [%rd36+20];
and.b32 %r741, %r732, -2143281136;
or.b32 %r742, %r741, -2139062144;
add.s32 %r743, %r742, -16843009;
or.b32 %r593, %r743, %r741;
mov.u32 %r103, %r593;

prmt.b32 %r103, %r103, 0, 0xba98;

xor.b32 %r744, %r103, %r722;
or.b32 %r745, %r744, -2139062144;
and.b32 %r746, %r103, 2139062143;
and.b32 %r747, %r726, -2139062144;
sub.s32 %r748, %r745, %r746;
xor.b32 %r110, %r747, %r748;
ld.global.nc.u32 %r111, [%rd36+24];

dp4a.s32.s32 %r105, %r106, %r107, %r647;


dp4a.s32.s32 %r109, %r110, %r111, %r105;

shl.b16 %rs27, %rs15, 3;
cvt.u64.u16 %rd43, %rs27;
and.b64 %rd44, %rd43, 4088;
add.s64 %rd45, %rd34, %rd44;
ld.global.nc.v2.u32 {%r749, %r750}, [%rd45];
not.b32 %r752, %r749;
not.b32 %r754, %r750;
shr.u16 %rs28, %rs15, 9;
cvt.u32.u16 %r755, %rs28;
popc.b32 %r756, %r755;
shl.b32 %r757, %r756, 7;
and.b32 %r758, %r757, 128;
or.b32 %r759, %r758, %r755;
mul.lo.s32 %r760, %r759, 16843009;
and.b32 %r761, %r760, 134480385;
or.b32 %r762, %r761, -2139062144;
add.s32 %r763, %r762, -16843009;
or.b32 %r603, %r763, %r761;
mov.u32 %r113, %r603;

prmt.b32 %r113, %r113, 0, 0xba98;

xor.b32 %r764, %r113, %r749;
or.b32 %r765, %r764, -2139062144;
and.b32 %r766, %r113, 2139062143;
and.b32 %r767, %r752, -2139062144;
sub.s32 %r768, %r765, %r766;
xor.b32 %r118, %r767, %r768;
ld.global.nc.u32 %r119, [%rd36+28];
and.b32 %r769, %r760, -2143281136;
or.b32 %r770, %r769, -2139062144;
add.s32 %r771, %r770, -16843009;
or.b32 %r605, %r771, %r769;
mov.u32 %r115, %r605;

prmt.b32 %r115, %r115, 0, 0xba98;

xor.b32 %r772, %r115, %r750;
or.b32 %r773, %r772, -2139062144;
and.b32 %r774, %r115, 2139062143;
and.b32 %r775, %r754, -2139062144;
sub.s32 %r776, %r773, %r774;
xor.b32 %r122, %r775, %r776;
ld.global.nc.u32 %r123, [%rd36+32];

dp4a.s32.s32 %r117, %r118, %r119, %r109;


dp4a.s32.s32 %r121, %r122, %r123, %r117;

cvt.u32.u16 %r777, %rs18;
cvt.u32.u16 %r778, %rs19;
mul.lo.s32 %r779, %r97, %r777;
mad.lo.s32 %r780, %r121, %r778, %r779;
add.s32 %r781, %r121, %r97;
shr.u32 %r782, %r781, 31;
add.s32 %r783, %r781, %r782;
shr.s32 %r784, %r783, 1;
add.s32 %r785, %r780, %r784;
shr.s32 %r786, %r785, 31;
shr.u32 %r787, %r786, 30;
add.s32 %r788, %r785, %r787;
shr.s32 %r789, %r788, 2;
add.s64 %rd46, %rd131, %rd6;
ld.global.nc.u16 %rs1, [%rd46];

{ cvt.f32.f16 %f73, %rs1;}


ld.global.nc.u32 %r125, [%rd36];

{.reg .f16 low,high;
mov.b32 {low,high},%r125;
cvt.f32.f16 %f74, low;}


mul.ftz.f32 %f97, %f73, %f74;
cvt.rn.f32.s32 %f98, %r789;
fma.rn.ftz.f32 %f276, %f97, %f98, %f276;
add.s64 %rd47, %rd131, %rd8;
ld.global.nc.u16 %rs29, [%rd47];
ld.global.nc.u16 %rs30, [%rd47+2];
ld.global.nc.u16 %rs31, [%rd47+4];
add.s64 %rd48, %rd131, %rd5;
ld.global.nc.u8 %rs32, [%rd48];
and.b16 %rs33, %rs32, 240;
and.b16 %rs34, %rs32, 15;
shr.u16 %rs35, %rs33, 4;
ld.global.nc.u16 %rs36, [%rd47+-2];
shl.b16 %rs37, %rs36, 3;
cvt.u64.u16 %rd49, %rs37;
and.b64 %rd50, %rd49, 4088;
add.s64 %rd51, %rd34, %rd50;
ld.global.nc.v2.u32 {%r790, %r791}, [%rd51];
not.b32 %r793, %r790;
not.b32 %r795, %r791;
shr.u16 %rs38, %rs36, 9;
cvt.u32.u16 %r796, %rs38;
popc.b32 %r797, %r796;
shl.b32 %r798, %r797, 7;
and.b32 %r799, %r798, 128;
or.b32 %r800, %r799, %r796;
mul.lo.s32 %r801, %r800, 16843009;
and.b32 %r802, %r801, 134480385;
or.b32 %r803, %r802, -2139062144;
add.s32 %r804, %r803, -16843009;
or.b32 %r616, %r804, %r802;
mov.u32 %r126, %r616;

prmt.b32 %r126, %r126, 0, 0xba98;

xor.b32 %r805, %r126, %r790;
or.b32 %r806, %r805, -2139062144;
and.b32 %r807, %r126, 2139062143;
and.b32 %r808, %r793, -2139062144;
sub.s32 %r809, %r806, %r807;
xor.b32 %r131, %r808, %r809;
and.b32 %r810, %r801, -2143281136;
or.b32 %r811, %r810, -2139062144;
add.s32 %r812, %r811, -16843009;
or.b32 %r618, %r812, %r810;
mov.u32 %r128, %r618;

prmt.b32 %r128, %r128, 0, 0xba98;

xor.b32 %r813, %r128, %r791;
or.b32 %r814, %r813, -2139062144;
and.b32 %r815, %r128, 2139062143;
and.b32 %r816, %r795, -2139062144;
sub.s32 %r817, %r814, %r815;
xor.b32 %r135, %r816, %r817;

dp4a.s32.s32 %r130, %r131, %r83, %r647;


dp4a.s32.s32 %r134, %r135, %r87, %r130;

shl.b16 %rs39, %rs29, 3;
cvt.u64.u16 %rd52, %rs39;
and.b64 %rd53, %rd52, 4088;
add.s64 %rd54, %rd34, %rd53;
ld.global.nc.v2.u32 {%r818, %r819}, [%rd54];
not.b32 %r821, %r818;
not.b32 %r823, %r819;
shr.u16 %rs40, %rs29, 9;
cvt.u32.u16 %r824, %rs40;
popc.b32 %r825, %r824;
shl.b32 %r826, %r825, 7;
and.b32 %r827, %r826, 128;
or.b32 %r828, %r827, %r824;
mul.lo.s32 %r829, %r828, 16843009;
and.b32 %r830, %r829, 134480385;
or.b32 %r831, %r830, -2139062144;
add.s32 %r832, %r831, -16843009;
or.b32 %r628, %r832, %r830;
mov.u32 %r138, %r628;

prmt.b32 %r138, %r138, 0, 0xba98;

xor.b32 %r833, %r138, %r818;
or.b32 %r834, %r833, -2139062144;
and.b32 %r835, %r138, 2139062143;
and.b32 %r836, %r821, -2139062144;
sub.s32 %r837, %r834, %r835;
xor.b32 %r143, %r836, %r837;
and.b32 %r838, %r829, -2143281136;
or.b32 %r839, %r838, -2139062144;
add.s32 %r840, %r839, -16843009;
or.b32 %r630, %r840, %r838;
mov.u32 %r140, %r630;

prmt.b32 %r140, %r140, 0, 0xba98;

xor.b32 %r841, %r140, %r819;
or.b32 %r842, %r841, -2139062144;
and.b32 %r843, %r140, 2139062143;
and.b32 %r844, %r823, -2139062144;
sub.s32 %r845, %r842, %r843;
xor.b32 %r147, %r844, %r845;

dp4a.s32.s32 %r142, %r143, %r95, %r134;


dp4a.s32.s32 %r146, %r147, %r99, %r142;

shl.b16 %rs41, %rs30, 3;
cvt.u64.u16 %rd55, %rs41;
and.b64 %rd56, %rd55, 4088;
add.s64 %rd57, %rd34, %rd56;
ld.global.nc.v2.u32 {%r846, %r847}, [%rd57];
not.b32 %r849, %r846;
not.b32 %r851, %r847;
shr.u16 %rs42, %rs30, 9;
cvt.u32.u16 %r852, %rs42;
popc.b32 %r853, %r852;
shl.b32 %r854, %r853, 7;
and.b32 %r855, %r854, 128;
or.b32 %r856, %r855, %r852;
mul.lo.s32 %r857, %r856, 16843009;
and.b32 %r858, %r857, 134480385;
or.b32 %r859, %r858, -2139062144;
add.s32 %r860, %r859, -16843009;
or.b32 %r640, %r860, %r858;
mov.u32 %r150, %r640;

prmt.b32 %r150, %r150, 0, 0xba98;

xor.b32 %r861, %r150, %r846;
or.b32 %r862, %r861, -2139062144;
and.b32 %r863, %r150, 2139062143;
and.b32 %r864, %r849, -2139062144;
sub.s32 %r865, %r862, %r863;
xor.b32 %r155, %r864, %r865;
and.b32 %r866, %r857, -2143281136;
or.b32 %r867, %r866, -2139062144;
add.s32 %r868, %r867, -16843009;
or.b32 %r642, %r868, %r866;
mov.u32 %r152, %r642;

prmt.b32 %r152, %r152, 0, 0xba98;

xor.b32 %r869, %r152, %r847;
or.b32 %r870, %r869, -2139062144;
and.b32 %r871, %r152, 2139062143;
and.b32 %r872, %r851, -2139062144;
sub.s32 %r873, %r870, %r871;
xor.b32 %r159, %r872, %r873;

dp4a.s32.s32 %r154, %r155, %r107, %r647;


dp4a.s32.s32 %r158, %r159, %r111, %r154;

shl.b16 %rs43, %rs31, 3;
cvt.u64.u16 %rd58, %rs43;
and.b64 %rd59, %rd58, 4088;
add.s64 %rd60, %rd34, %rd59;
ld.global.nc.v2.u32 {%r874, %r875}, [%rd60];
not.b32 %r877, %r874;
not.b32 %r879, %r875;
shr.u16 %rs44, %rs31, 9;
cvt.u32.u16 %r880, %rs44;
popc.b32 %r881, %r880;
shl.b32 %r882, %r881, 7;
and.b32 %r883, %r882, 128;
or.b32 %r884, %r883, %r880;
mul.lo.s32 %r885, %r884, 16843009;
and.b32 %r886, %r885, 134480385;
or.b32 %r887, %r886, -2139062144;
add.s32 %r888, %r887, -16843009;
or.b32 %r652, %r888, %r886;
mov.u32 %r162, %r652;

prmt.b32 %r162, %r162, 0, 0xba98;

xor.b32 %r889, %r162, %r874;
or.b32 %r890, %r889, -2139062144;
and.b32 %r891, %r162, 2139062143;
and.b32 %r892, %r877, -2139062144;
sub.s32 %r893, %r890, %r891;
xor.b32 %r167, %r892, %r893;
and.b32 %r894, %r885, -2143281136;
or.b32 %r895, %r894, -2139062144;
add.s32 %r896, %r895, -16843009;
or.b32 %r654, %r896, %r894;
mov.u32 %r164, %r654;

prmt.b32 %r164, %r164, 0, 0xba98;

xor.b32 %r897, %r164, %r875;
or.b32 %r898, %r897, -2139062144;
and.b32 %r899, %r164, 2139062143;
and.b32 %r900, %r879, -2139062144;
sub.s32 %r901, %r898, %r899;
xor.b32 %r171, %r900, %r901;

dp4a.s32.s32 %r166, %r167, %r119, %r158;


dp4a.s32.s32 %r170, %r171, %r123, %r166;

cvt.u32.u16 %r902, %rs34;
cvt.u32.u16 %r903, %rs35;
mul.lo.s32 %r904, %r146, %r902;
mad.lo.s32 %r905, %r170, %r903, %r904;
add.s32 %r906, %r170, %r146;
shr.u32 %r907, %r906, 31;
add.s32 %r908, %r906, %r907;
shr.s32 %r909, %r908, 1;
add.s32 %r910, %r905, %r909;
shr.s32 %r911, %r910, 31;
shr.u32 %r912, %r911, 30;
add.s32 %r913, %r910, %r912;
shr.s32 %r914, %r913, 2;
add.s64 %rd61, %rd131, %rd4;
ld.global.nc.u16 %rs2, [%rd61];

{ cvt.f32.f16 %f75, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r125;
cvt.f32.f16 %f76, low;}


mul.ftz.f32 %f99, %f75, %f76;
cvt.rn.f32.s32 %f100, %r914;
fma.rn.ftz.f32 %f275, %f99, %f100, %f275;
cvt.u64.u32 %rd62, %r1632;
add.s64 %rd63, %rd1, %rd62;
add.s64 %rd64, %rd63, %rd3;
mul.lo.s64 %rd65, %rd64, 36;
mov.u32 %r175, %r567;

prmt.b32 %r175, %r175, 0, 0xba98;

xor.b32 %r915, %r175, %r665;
or.b32 %r916, %r915, -2139062144;
and.b32 %r917, %r175, 2139062143;
sub.s32 %r918, %r916, %r917;
xor.b32 %r180, %r683, %r918;
add.s64 %rd66, %rd10, %rd65;
ld.global.nc.u32 %r181, [%rd66+4];
mov.u32 %r177, %r569;

prmt.b32 %r177, %r177, 0, 0xba98;

xor.b32 %r919, %r177, %r666;
or.b32 %r920, %r919, -2139062144;
and.b32 %r921, %r177, 2139062143;
sub.s32 %r922, %r920, %r921;
xor.b32 %r184, %r691, %r922;
ld.global.nc.u32 %r185, [%rd66+8];

dp4a.s32.s32 %r179, %r180, %r181, %r647;


dp4a.s32.s32 %r183, %r184, %r185, %r179;

mov.u32 %r187, %r579;

prmt.b32 %r187, %r187, 0, 0xba98;

xor.b32 %r923, %r187, %r693;
or.b32 %r924, %r923, -2139062144;
and.b32 %r925, %r187, 2139062143;
sub.s32 %r926, %r924, %r925;
xor.b32 %r192, %r711, %r926;
ld.global.nc.u32 %r193, [%rd66+12];
mov.u32 %r189, %r581;

prmt.b32 %r189, %r189, 0, 0xba98;

xor.b32 %r927, %r189, %r694;
or.b32 %r928, %r927, -2139062144;
and.b32 %r929, %r189, 2139062143;
sub.s32 %r930, %r928, %r929;
xor.b32 %r196, %r719, %r930;
ld.global.nc.u32 %r197, [%rd66+16];

dp4a.s32.s32 %r191, %r192, %r193, %r183;


dp4a.s32.s32 %r195, %r196, %r197, %r191;

mov.u32 %r199, %r591;

prmt.b32 %r199, %r199, 0, 0xba98;

xor.b32 %r931, %r199, %r721;
or.b32 %r932, %r931, -2139062144;
and.b32 %r933, %r199, 2139062143;
sub.s32 %r934, %r932, %r933;
xor.b32 %r204, %r739, %r934;
ld.global.nc.u32 %r205, [%rd66+20];
mov.u32 %r201, %r593;

prmt.b32 %r201, %r201, 0, 0xba98;

xor.b32 %r935, %r201, %r722;
or.b32 %r936, %r935, -2139062144;
and.b32 %r937, %r201, 2139062143;
sub.s32 %r938, %r936, %r937;
xor.b32 %r208, %r747, %r938;
ld.global.nc.u32 %r209, [%rd66+24];

dp4a.s32.s32 %r203, %r204, %r205, %r647;


dp4a.s32.s32 %r207, %r208, %r209, %r203;

mov.u32 %r211, %r603;

prmt.b32 %r211, %r211, 0, 0xba98;

xor.b32 %r939, %r211, %r749;
or.b32 %r940, %r939, -2139062144;
and.b32 %r941, %r211, 2139062143;
sub.s32 %r942, %r940, %r941;
xor.b32 %r216, %r767, %r942;
ld.global.nc.u32 %r217, [%rd66+28];
mov.u32 %r213, %r605;

prmt.b32 %r213, %r213, 0, 0xba98;

xor.b32 %r943, %r213, %r750;
or.b32 %r944, %r943, -2139062144;
and.b32 %r945, %r213, 2139062143;
sub.s32 %r946, %r944, %r945;
xor.b32 %r220, %r775, %r946;
ld.global.nc.u32 %r221, [%rd66+32];

dp4a.s32.s32 %r215, %r216, %r217, %r207;


dp4a.s32.s32 %r219, %r220, %r221, %r215;

mul.lo.s32 %r947, %r195, %r777;
mad.lo.s32 %r948, %r219, %r778, %r947;
add.s32 %r949, %r219, %r195;
shr.u32 %r950, %r949, 31;
add.s32 %r951, %r949, %r950;
shr.s32 %r952, %r951, 1;
add.s32 %r953, %r948, %r952;
shr.s32 %r954, %r953, 31;
shr.u32 %r955, %r954, 30;
add.s32 %r956, %r953, %r955;
shr.s32 %r957, %r956, 2;

{ cvt.f32.f16 %f77, %rs1;}


ld.global.nc.u32 %r223, [%rd66];

{.reg .f16 low,high;
mov.b32 {low,high},%r223;
cvt.f32.f16 %f78, low;}


mul.ftz.f32 %f101, %f77, %f78;
cvt.rn.f32.s32 %f102, %r957;
fma.rn.ftz.f32 %f274, %f101, %f102, %f274;
mov.u32 %r224, %r616;

prmt.b32 %r224, %r224, 0, 0xba98;

xor.b32 %r958, %r224, %r790;
or.b32 %r959, %r958, -2139062144;
and.b32 %r960, %r224, 2139062143;
sub.s32 %r961, %r959, %r960;
xor.b32 %r229, %r808, %r961;
mov.u32 %r226, %r618;

prmt.b32 %r226, %r226, 0, 0xba98;

xor.b32 %r962, %r226, %r791;
or.b32 %r963, %r962, -2139062144;
and.b32 %r964, %r226, 2139062143;
sub.s32 %r965, %r963, %r964;
xor.b32 %r233, %r816, %r965;

dp4a.s32.s32 %r228, %r229, %r181, %r647;


dp4a.s32.s32 %r232, %r233, %r185, %r228;

mov.u32 %r236, %r628;

prmt.b32 %r236, %r236, 0, 0xba98;

xor.b32 %r966, %r236, %r818;
or.b32 %r967, %r966, -2139062144;
and.b32 %r968, %r236, 2139062143;
sub.s32 %r969, %r967, %r968;
xor.b32 %r241, %r836, %r969;
mov.u32 %r238, %r630;

prmt.b32 %r238, %r238, 0, 0xba98;

xor.b32 %r970, %r238, %r819;
or.b32 %r971, %r970, -2139062144;
and.b32 %r972, %r238, 2139062143;
sub.s32 %r973, %r971, %r972;
xor.b32 %r245, %r844, %r973;

dp4a.s32.s32 %r240, %r241, %r193, %r232;


dp4a.s32.s32 %r244, %r245, %r197, %r240;

mov.u32 %r248, %r640;

prmt.b32 %r248, %r248, 0, 0xba98;

xor.b32 %r974, %r248, %r846;
or.b32 %r975, %r974, -2139062144;
and.b32 %r976, %r248, 2139062143;
sub.s32 %r977, %r975, %r976;
xor.b32 %r253, %r864, %r977;
mov.u32 %r250, %r642;

prmt.b32 %r250, %r250, 0, 0xba98;

xor.b32 %r978, %r250, %r847;
or.b32 %r979, %r978, -2139062144;
and.b32 %r980, %r250, 2139062143;
sub.s32 %r981, %r979, %r980;
xor.b32 %r257, %r872, %r981;

dp4a.s32.s32 %r252, %r253, %r205, %r647;


dp4a.s32.s32 %r256, %r257, %r209, %r252;

mov.u32 %r260, %r652;

prmt.b32 %r260, %r260, 0, 0xba98;

xor.b32 %r982, %r260, %r874;
or.b32 %r983, %r982, -2139062144;
and.b32 %r984, %r260, 2139062143;
sub.s32 %r985, %r983, %r984;
xor.b32 %r265, %r892, %r985;
mov.u32 %r262, %r654;

prmt.b32 %r262, %r262, 0, 0xba98;

xor.b32 %r986, %r262, %r875;
or.b32 %r987, %r986, -2139062144;
and.b32 %r988, %r262, 2139062143;
sub.s32 %r989, %r987, %r988;
xor.b32 %r269, %r900, %r989;

dp4a.s32.s32 %r264, %r265, %r217, %r256;


dp4a.s32.s32 %r268, %r269, %r221, %r264;

mul.lo.s32 %r990, %r244, %r902;
mad.lo.s32 %r991, %r268, %r903, %r990;
add.s32 %r992, %r268, %r244;
shr.u32 %r993, %r992, 31;
add.s32 %r994, %r992, %r993;
shr.s32 %r995, %r994, 1;
add.s32 %r996, %r991, %r995;
shr.s32 %r997, %r996, 31;
shr.u32 %r998, %r997, 30;
add.s32 %r999, %r996, %r998;
shr.s32 %r1000, %r999, 2;

{ cvt.f32.f16 %f79, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r223;
cvt.f32.f16 %f80, low;}


mul.ftz.f32 %f103, %f79, %f80;
cvt.rn.f32.s32 %f104, %r1000;
fma.rn.ftz.f32 %f273, %f103, %f104, %f273;
cvt.u64.u32 %rd67, %r1636;
add.s64 %rd68, %rd1, %rd67;
add.s64 %rd69, %rd68, %rd3;
mul.lo.s64 %rd70, %rd69, 36;
mov.u32 %r273, %r567;

prmt.b32 %r273, %r273, 0, 0xba98;

xor.b32 %r1001, %r273, %r665;
or.b32 %r1002, %r1001, -2139062144;
and.b32 %r1003, %r273, 2139062143;
sub.s32 %r1004, %r1002, %r1003;
xor.b32 %r278, %r683, %r1004;
add.s64 %rd71, %rd10, %rd70;
ld.global.nc.u32 %r279, [%rd71+4];
mov.u32 %r275, %r569;

prmt.b32 %r275, %r275, 0, 0xba98;

xor.b32 %r1005, %r275, %r666;
or.b32 %r1006, %r1005, -2139062144;
and.b32 %r1007, %r275, 2139062143;
sub.s32 %r1008, %r1006, %r1007;
xor.b32 %r282, %r691, %r1008;
ld.global.nc.u32 %r283, [%rd71+8];

dp4a.s32.s32 %r277, %r278, %r279, %r647;


dp4a.s32.s32 %r281, %r282, %r283, %r277;

mov.u32 %r285, %r579;

prmt.b32 %r285, %r285, 0, 0xba98;

xor.b32 %r1009, %r285, %r693;
or.b32 %r1010, %r1009, -2139062144;
and.b32 %r1011, %r285, 2139062143;
sub.s32 %r1012, %r1010, %r1011;
xor.b32 %r290, %r711, %r1012;
ld.global.nc.u32 %r291, [%rd71+12];
mov.u32 %r287, %r581;

prmt.b32 %r287, %r287, 0, 0xba98;

xor.b32 %r1013, %r287, %r694;
or.b32 %r1014, %r1013, -2139062144;
and.b32 %r1015, %r287, 2139062143;
sub.s32 %r1016, %r1014, %r1015;
xor.b32 %r294, %r719, %r1016;
ld.global.nc.u32 %r295, [%rd71+16];

dp4a.s32.s32 %r289, %r290, %r291, %r281;


dp4a.s32.s32 %r293, %r294, %r295, %r289;

mov.u32 %r297, %r591;

prmt.b32 %r297, %r297, 0, 0xba98;

xor.b32 %r1017, %r297, %r721;
or.b32 %r1018, %r1017, -2139062144;
and.b32 %r1019, %r297, 2139062143;
sub.s32 %r1020, %r1018, %r1019;
xor.b32 %r302, %r739, %r1020;
ld.global.nc.u32 %r303, [%rd71+20];
mov.u32 %r299, %r593;

prmt.b32 %r299, %r299, 0, 0xba98;

xor.b32 %r1021, %r299, %r722;
or.b32 %r1022, %r1021, -2139062144;
and.b32 %r1023, %r299, 2139062143;
sub.s32 %r1024, %r1022, %r1023;
xor.b32 %r306, %r747, %r1024;
ld.global.nc.u32 %r307, [%rd71+24];

dp4a.s32.s32 %r301, %r302, %r303, %r647;


dp4a.s32.s32 %r305, %r306, %r307, %r301;

mov.u32 %r309, %r603;

prmt.b32 %r309, %r309, 0, 0xba98;

xor.b32 %r1025, %r309, %r749;
or.b32 %r1026, %r1025, -2139062144;
and.b32 %r1027, %r309, 2139062143;
sub.s32 %r1028, %r1026, %r1027;
xor.b32 %r314, %r767, %r1028;
ld.global.nc.u32 %r315, [%rd71+28];
mov.u32 %r311, %r605;

prmt.b32 %r311, %r311, 0, 0xba98;

xor.b32 %r1029, %r311, %r750;
or.b32 %r1030, %r1029, -2139062144;
and.b32 %r1031, %r311, 2139062143;
sub.s32 %r1032, %r1030, %r1031;
xor.b32 %r318, %r775, %r1032;
ld.global.nc.u32 %r319, [%rd71+32];

dp4a.s32.s32 %r313, %r314, %r315, %r305;


dp4a.s32.s32 %r317, %r318, %r319, %r313;

mul.lo.s32 %r1033, %r293, %r777;
mad.lo.s32 %r1034, %r317, %r778, %r1033;
add.s32 %r1035, %r317, %r293;
shr.u32 %r1036, %r1035, 31;
add.s32 %r1037, %r1035, %r1036;
shr.s32 %r1038, %r1037, 1;
add.s32 %r1039, %r1034, %r1038;
shr.s32 %r1040, %r1039, 31;
shr.u32 %r1041, %r1040, 30;
add.s32 %r1042, %r1039, %r1041;
shr.s32 %r1043, %r1042, 2;

{ cvt.f32.f16 %f81, %rs1;}


ld.global.nc.u32 %r321, [%rd71];

{.reg .f16 low,high;
mov.b32 {low,high},%r321;
cvt.f32.f16 %f82, low;}


mul.ftz.f32 %f105, %f81, %f82;
cvt.rn.f32.s32 %f106, %r1043;
fma.rn.ftz.f32 %f272, %f105, %f106, %f272;
mov.u32 %r322, %r616;

prmt.b32 %r322, %r322, 0, 0xba98;

xor.b32 %r1044, %r322, %r790;
or.b32 %r1045, %r1044, -2139062144;
and.b32 %r1046, %r322, 2139062143;
sub.s32 %r1047, %r1045, %r1046;
xor.b32 %r327, %r808, %r1047;
mov.u32 %r324, %r618;

prmt.b32 %r324, %r324, 0, 0xba98;

xor.b32 %r1048, %r324, %r791;
or.b32 %r1049, %r1048, -2139062144;
and.b32 %r1050, %r324, 2139062143;
sub.s32 %r1051, %r1049, %r1050;
xor.b32 %r331, %r816, %r1051;

dp4a.s32.s32 %r326, %r327, %r279, %r647;


dp4a.s32.s32 %r330, %r331, %r283, %r326;

mov.u32 %r334, %r628;

prmt.b32 %r334, %r334, 0, 0xba98;

xor.b32 %r1052, %r334, %r818;
or.b32 %r1053, %r1052, -2139062144;
and.b32 %r1054, %r334, 2139062143;
sub.s32 %r1055, %r1053, %r1054;
xor.b32 %r339, %r836, %r1055;
mov.u32 %r336, %r630;

prmt.b32 %r336, %r336, 0, 0xba98;

xor.b32 %r1056, %r336, %r819;
or.b32 %r1057, %r1056, -2139062144;
and.b32 %r1058, %r336, 2139062143;
sub.s32 %r1059, %r1057, %r1058;
xor.b32 %r343, %r844, %r1059;

dp4a.s32.s32 %r338, %r339, %r291, %r330;


dp4a.s32.s32 %r342, %r343, %r295, %r338;

mov.u32 %r346, %r640;

prmt.b32 %r346, %r346, 0, 0xba98;

xor.b32 %r1060, %r346, %r846;
or.b32 %r1061, %r1060, -2139062144;
and.b32 %r1062, %r346, 2139062143;
sub.s32 %r1063, %r1061, %r1062;
xor.b32 %r351, %r864, %r1063;
mov.u32 %r348, %r642;

prmt.b32 %r348, %r348, 0, 0xba98;

xor.b32 %r1064, %r348, %r847;
or.b32 %r1065, %r1064, -2139062144;
and.b32 %r1066, %r348, 2139062143;
sub.s32 %r1067, %r1065, %r1066;
xor.b32 %r355, %r872, %r1067;

dp4a.s32.s32 %r350, %r351, %r303, %r647;


dp4a.s32.s32 %r354, %r355, %r307, %r350;

mov.u32 %r358, %r652;

prmt.b32 %r358, %r358, 0, 0xba98;

xor.b32 %r1068, %r358, %r874;
or.b32 %r1069, %r1068, -2139062144;
and.b32 %r1070, %r358, 2139062143;
sub.s32 %r1071, %r1069, %r1070;
xor.b32 %r363, %r892, %r1071;
mov.u32 %r360, %r654;

prmt.b32 %r360, %r360, 0, 0xba98;

xor.b32 %r1072, %r360, %r875;
or.b32 %r1073, %r1072, -2139062144;
and.b32 %r1074, %r360, 2139062143;
sub.s32 %r1075, %r1073, %r1074;
xor.b32 %r367, %r900, %r1075;

dp4a.s32.s32 %r362, %r363, %r315, %r354;


dp4a.s32.s32 %r366, %r367, %r319, %r362;

mul.lo.s32 %r1076, %r342, %r902;
mad.lo.s32 %r1077, %r366, %r903, %r1076;
add.s32 %r1078, %r366, %r342;
shr.u32 %r1079, %r1078, 31;
add.s32 %r1080, %r1078, %r1079;
shr.s32 %r1081, %r1080, 1;
add.s32 %r1082, %r1077, %r1081;
shr.s32 %r1083, %r1082, 31;
shr.u32 %r1084, %r1083, 30;
add.s32 %r1085, %r1082, %r1084;
shr.s32 %r1086, %r1085, 2;

{ cvt.f32.f16 %f83, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r321;
cvt.f32.f16 %f84, low;}


mul.ftz.f32 %f107, %f83, %f84;
cvt.rn.f32.s32 %f108, %r1086;
fma.rn.ftz.f32 %f271, %f107, %f108, %f271;
cvt.u64.u32 %rd72, %r1635;
add.s64 %rd73, %rd1, %rd72;
add.s64 %rd74, %rd73, %rd3;
mul.lo.s64 %rd75, %rd74, 36;
mov.u32 %r371, %r567;

prmt.b32 %r371, %r371, 0, 0xba98;

xor.b32 %r1087, %r371, %r665;
or.b32 %r1088, %r1087, -2139062144;
and.b32 %r1089, %r371, 2139062143;
sub.s32 %r1090, %r1088, %r1089;
xor.b32 %r376, %r683, %r1090;
add.s64 %rd76, %rd10, %rd75;
ld.global.nc.u32 %r377, [%rd76+4];
mov.u32 %r373, %r569;

prmt.b32 %r373, %r373, 0, 0xba98;

xor.b32 %r1091, %r373, %r666;
or.b32 %r1092, %r1091, -2139062144;
and.b32 %r1093, %r373, 2139062143;
sub.s32 %r1094, %r1092, %r1093;
xor.b32 %r380, %r691, %r1094;
ld.global.nc.u32 %r381, [%rd76+8];

dp4a.s32.s32 %r375, %r376, %r377, %r647;


dp4a.s32.s32 %r379, %r380, %r381, %r375;

mov.u32 %r383, %r579;

prmt.b32 %r383, %r383, 0, 0xba98;

xor.b32 %r1095, %r383, %r693;
or.b32 %r1096, %r1095, -2139062144;
and.b32 %r1097, %r383, 2139062143;
sub.s32 %r1098, %r1096, %r1097;
xor.b32 %r388, %r711, %r1098;
ld.global.nc.u32 %r389, [%rd76+12];
mov.u32 %r385, %r581;

prmt.b32 %r385, %r385, 0, 0xba98;

xor.b32 %r1099, %r385, %r694;
or.b32 %r1100, %r1099, -2139062144;
and.b32 %r1101, %r385, 2139062143;
sub.s32 %r1102, %r1100, %r1101;
xor.b32 %r392, %r719, %r1102;
ld.global.nc.u32 %r393, [%rd76+16];

dp4a.s32.s32 %r387, %r388, %r389, %r379;


dp4a.s32.s32 %r391, %r392, %r393, %r387;

mov.u32 %r395, %r591;

prmt.b32 %r395, %r395, 0, 0xba98;

xor.b32 %r1103, %r395, %r721;
or.b32 %r1104, %r1103, -2139062144;
and.b32 %r1105, %r395, 2139062143;
sub.s32 %r1106, %r1104, %r1105;
xor.b32 %r400, %r739, %r1106;
ld.global.nc.u32 %r401, [%rd76+20];
mov.u32 %r397, %r593;

prmt.b32 %r397, %r397, 0, 0xba98;

xor.b32 %r1107, %r397, %r722;
or.b32 %r1108, %r1107, -2139062144;
and.b32 %r1109, %r397, 2139062143;
sub.s32 %r1110, %r1108, %r1109;
xor.b32 %r404, %r747, %r1110;
ld.global.nc.u32 %r405, [%rd76+24];

dp4a.s32.s32 %r399, %r400, %r401, %r647;


dp4a.s32.s32 %r403, %r404, %r405, %r399;

mov.u32 %r407, %r603;

prmt.b32 %r407, %r407, 0, 0xba98;

xor.b32 %r1111, %r407, %r749;
or.b32 %r1112, %r1111, -2139062144;
and.b32 %r1113, %r407, 2139062143;
sub.s32 %r1114, %r1112, %r1113;
xor.b32 %r412, %r767, %r1114;
ld.global.nc.u32 %r413, [%rd76+28];
mov.u32 %r409, %r605;

prmt.b32 %r409, %r409, 0, 0xba98;

xor.b32 %r1115, %r409, %r750;
or.b32 %r1116, %r1115, -2139062144;
and.b32 %r1117, %r409, 2139062143;
sub.s32 %r1118, %r1116, %r1117;
xor.b32 %r416, %r775, %r1118;
ld.global.nc.u32 %r417, [%rd76+32];

dp4a.s32.s32 %r411, %r412, %r413, %r403;


dp4a.s32.s32 %r415, %r416, %r417, %r411;

mul.lo.s32 %r1119, %r391, %r777;
mad.lo.s32 %r1120, %r415, %r778, %r1119;
add.s32 %r1121, %r415, %r391;
shr.u32 %r1122, %r1121, 31;
add.s32 %r1123, %r1121, %r1122;
shr.s32 %r1124, %r1123, 1;
add.s32 %r1125, %r1120, %r1124;
shr.s32 %r1126, %r1125, 31;
shr.u32 %r1127, %r1126, 30;
add.s32 %r1128, %r1125, %r1127;
shr.s32 %r1129, %r1128, 2;

{ cvt.f32.f16 %f85, %rs1;}


ld.global.nc.u32 %r419, [%rd76];

{.reg .f16 low,high;
mov.b32 {low,high},%r419;
cvt.f32.f16 %f86, low;}


mul.ftz.f32 %f109, %f85, %f86;
cvt.rn.f32.s32 %f110, %r1129;
fma.rn.ftz.f32 %f270, %f109, %f110, %f270;
mov.u32 %r420, %r616;

prmt.b32 %r420, %r420, 0, 0xba98;

xor.b32 %r1130, %r420, %r790;
or.b32 %r1131, %r1130, -2139062144;
and.b32 %r1132, %r420, 2139062143;
sub.s32 %r1133, %r1131, %r1132;
xor.b32 %r425, %r808, %r1133;
mov.u32 %r422, %r618;

prmt.b32 %r422, %r422, 0, 0xba98;

xor.b32 %r1134, %r422, %r791;
or.b32 %r1135, %r1134, -2139062144;
and.b32 %r1136, %r422, 2139062143;
sub.s32 %r1137, %r1135, %r1136;
xor.b32 %r429, %r816, %r1137;

dp4a.s32.s32 %r424, %r425, %r377, %r647;


dp4a.s32.s32 %r428, %r429, %r381, %r424;

mov.u32 %r432, %r628;

prmt.b32 %r432, %r432, 0, 0xba98;

xor.b32 %r1138, %r432, %r818;
or.b32 %r1139, %r1138, -2139062144;
and.b32 %r1140, %r432, 2139062143;
sub.s32 %r1141, %r1139, %r1140;
xor.b32 %r437, %r836, %r1141;
mov.u32 %r434, %r630;

prmt.b32 %r434, %r434, 0, 0xba98;

xor.b32 %r1142, %r434, %r819;
or.b32 %r1143, %r1142, -2139062144;
and.b32 %r1144, %r434, 2139062143;
sub.s32 %r1145, %r1143, %r1144;
xor.b32 %r441, %r844, %r1145;

dp4a.s32.s32 %r436, %r437, %r389, %r428;


dp4a.s32.s32 %r440, %r441, %r393, %r436;

mov.u32 %r444, %r640;

prmt.b32 %r444, %r444, 0, 0xba98;

xor.b32 %r1146, %r444, %r846;
or.b32 %r1147, %r1146, -2139062144;
and.b32 %r1148, %r444, 2139062143;
sub.s32 %r1149, %r1147, %r1148;
xor.b32 %r449, %r864, %r1149;
mov.u32 %r446, %r642;

prmt.b32 %r446, %r446, 0, 0xba98;

xor.b32 %r1150, %r446, %r847;
or.b32 %r1151, %r1150, -2139062144;
and.b32 %r1152, %r446, 2139062143;
sub.s32 %r1153, %r1151, %r1152;
xor.b32 %r453, %r872, %r1153;

dp4a.s32.s32 %r448, %r449, %r401, %r647;


dp4a.s32.s32 %r452, %r453, %r405, %r448;

mov.u32 %r456, %r652;

prmt.b32 %r456, %r456, 0, 0xba98;

xor.b32 %r1154, %r456, %r874;
or.b32 %r1155, %r1154, -2139062144;
and.b32 %r1156, %r456, 2139062143;
sub.s32 %r1157, %r1155, %r1156;
xor.b32 %r461, %r892, %r1157;
mov.u32 %r458, %r654;

prmt.b32 %r458, %r458, 0, 0xba98;

xor.b32 %r1158, %r458, %r875;
or.b32 %r1159, %r1158, -2139062144;
and.b32 %r1160, %r458, 2139062143;
sub.s32 %r1161, %r1159, %r1160;
xor.b32 %r465, %r900, %r1161;

dp4a.s32.s32 %r460, %r461, %r413, %r452;


dp4a.s32.s32 %r464, %r465, %r417, %r460;

mul.lo.s32 %r1162, %r440, %r902;
mad.lo.s32 %r1163, %r464, %r903, %r1162;
add.s32 %r1164, %r464, %r440;
shr.u32 %r1165, %r1164, 31;
add.s32 %r1166, %r1164, %r1165;
shr.s32 %r1167, %r1166, 1;
add.s32 %r1168, %r1163, %r1167;
shr.s32 %r1169, %r1168, 31;
shr.u32 %r1170, %r1169, 30;
add.s32 %r1171, %r1168, %r1170;
shr.s32 %r1172, %r1171, 2;

{ cvt.f32.f16 %f87, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r419;
cvt.f32.f16 %f88, low;}


mul.ftz.f32 %f111, %f87, %f88;
cvt.rn.f32.s32 %f112, %r1172;
fma.rn.ftz.f32 %f269, %f111, %f112, %f269;
cvt.u64.u32 %rd77, %r1634;
add.s64 %rd78, %rd1, %rd77;
add.s64 %rd79, %rd78, %rd3;
mul.lo.s64 %rd80, %rd79, 36;
mov.u32 %r469, %r567;

prmt.b32 %r469, %r469, 0, 0xba98;

xor.b32 %r1173, %r469, %r665;
or.b32 %r1174, %r1173, -2139062144;
and.b32 %r1175, %r469, 2139062143;
sub.s32 %r1176, %r1174, %r1175;
xor.b32 %r474, %r683, %r1176;
add.s64 %rd81, %rd10, %rd80;
ld.global.nc.u32 %r475, [%rd81+4];
mov.u32 %r471, %r569;

prmt.b32 %r471, %r471, 0, 0xba98;

xor.b32 %r1177, %r471, %r666;
or.b32 %r1178, %r1177, -2139062144;
and.b32 %r1179, %r471, 2139062143;
sub.s32 %r1180, %r1178, %r1179;
xor.b32 %r478, %r691, %r1180;
ld.global.nc.u32 %r479, [%rd81+8];

dp4a.s32.s32 %r473, %r474, %r475, %r647;


dp4a.s32.s32 %r477, %r478, %r479, %r473;

mov.u32 %r481, %r579;

prmt.b32 %r481, %r481, 0, 0xba98;

xor.b32 %r1181, %r481, %r693;
or.b32 %r1182, %r1181, -2139062144;
and.b32 %r1183, %r481, 2139062143;
sub.s32 %r1184, %r1182, %r1183;
xor.b32 %r486, %r711, %r1184;
ld.global.nc.u32 %r487, [%rd81+12];
mov.u32 %r483, %r581;

prmt.b32 %r483, %r483, 0, 0xba98;

xor.b32 %r1185, %r483, %r694;
or.b32 %r1186, %r1185, -2139062144;
and.b32 %r1187, %r483, 2139062143;
sub.s32 %r1188, %r1186, %r1187;
xor.b32 %r490, %r719, %r1188;
ld.global.nc.u32 %r491, [%rd81+16];

dp4a.s32.s32 %r485, %r486, %r487, %r477;


dp4a.s32.s32 %r489, %r490, %r491, %r485;

mov.u32 %r493, %r591;

prmt.b32 %r493, %r493, 0, 0xba98;

xor.b32 %r1189, %r493, %r721;
or.b32 %r1190, %r1189, -2139062144;
and.b32 %r1191, %r493, 2139062143;
sub.s32 %r1192, %r1190, %r1191;
xor.b32 %r498, %r739, %r1192;
ld.global.nc.u32 %r499, [%rd81+20];
mov.u32 %r495, %r593;

prmt.b32 %r495, %r495, 0, 0xba98;

xor.b32 %r1193, %r495, %r722;
or.b32 %r1194, %r1193, -2139062144;
and.b32 %r1195, %r495, 2139062143;
sub.s32 %r1196, %r1194, %r1195;
xor.b32 %r502, %r747, %r1196;
ld.global.nc.u32 %r503, [%rd81+24];

dp4a.s32.s32 %r497, %r498, %r499, %r647;


dp4a.s32.s32 %r501, %r502, %r503, %r497;

mov.u32 %r505, %r603;

prmt.b32 %r505, %r505, 0, 0xba98;

xor.b32 %r1197, %r505, %r749;
or.b32 %r1198, %r1197, -2139062144;
and.b32 %r1199, %r505, 2139062143;
sub.s32 %r1200, %r1198, %r1199;
xor.b32 %r510, %r767, %r1200;
ld.global.nc.u32 %r511, [%rd81+28];
mov.u32 %r507, %r605;

prmt.b32 %r507, %r507, 0, 0xba98;

xor.b32 %r1201, %r507, %r750;
or.b32 %r1202, %r1201, -2139062144;
and.b32 %r1203, %r507, 2139062143;
sub.s32 %r1204, %r1202, %r1203;
xor.b32 %r514, %r775, %r1204;
ld.global.nc.u32 %r515, [%rd81+32];

dp4a.s32.s32 %r509, %r510, %r511, %r501;


dp4a.s32.s32 %r513, %r514, %r515, %r509;

mul.lo.s32 %r1205, %r489, %r777;
mad.lo.s32 %r1206, %r513, %r778, %r1205;
add.s32 %r1207, %r513, %r489;
shr.u32 %r1208, %r1207, 31;
add.s32 %r1209, %r1207, %r1208;
shr.s32 %r1210, %r1209, 1;
add.s32 %r1211, %r1206, %r1210;
shr.s32 %r1212, %r1211, 31;
shr.u32 %r1213, %r1212, 30;
add.s32 %r1214, %r1211, %r1213;
shr.s32 %r1215, %r1214, 2;

{ cvt.f32.f16 %f89, %rs1;}


ld.global.nc.u32 %r517, [%rd81];

{.reg .f16 low,high;
mov.b32 {low,high},%r517;
cvt.f32.f16 %f90, low;}


mul.ftz.f32 %f113, %f89, %f90;
cvt.rn.f32.s32 %f114, %r1215;
fma.rn.ftz.f32 %f268, %f113, %f114, %f268;
mov.u32 %r518, %r616;

prmt.b32 %r518, %r518, 0, 0xba98;

xor.b32 %r1216, %r518, %r790;
or.b32 %r1217, %r1216, -2139062144;
and.b32 %r1218, %r518, 2139062143;
sub.s32 %r1219, %r1217, %r1218;
xor.b32 %r523, %r808, %r1219;
mov.u32 %r520, %r618;

prmt.b32 %r520, %r520, 0, 0xba98;

xor.b32 %r1220, %r520, %r791;
or.b32 %r1221, %r1220, -2139062144;
and.b32 %r1222, %r520, 2139062143;
sub.s32 %r1223, %r1221, %r1222;
xor.b32 %r527, %r816, %r1223;

dp4a.s32.s32 %r522, %r523, %r475, %r647;


dp4a.s32.s32 %r526, %r527, %r479, %r522;

mov.u32 %r530, %r628;

prmt.b32 %r530, %r530, 0, 0xba98;

xor.b32 %r1224, %r530, %r818;
or.b32 %r1225, %r1224, -2139062144;
and.b32 %r1226, %r530, 2139062143;
sub.s32 %r1227, %r1225, %r1226;
xor.b32 %r535, %r836, %r1227;
mov.u32 %r532, %r630;

prmt.b32 %r532, %r532, 0, 0xba98;

xor.b32 %r1228, %r532, %r819;
or.b32 %r1229, %r1228, -2139062144;
and.b32 %r1230, %r532, 2139062143;
sub.s32 %r1231, %r1229, %r1230;
xor.b32 %r539, %r844, %r1231;

dp4a.s32.s32 %r534, %r535, %r487, %r526;


dp4a.s32.s32 %r538, %r539, %r491, %r534;

mov.u32 %r542, %r640;

prmt.b32 %r542, %r542, 0, 0xba98;

xor.b32 %r1232, %r542, %r846;
or.b32 %r1233, %r1232, -2139062144;
and.b32 %r1234, %r542, 2139062143;
sub.s32 %r1235, %r1233, %r1234;
xor.b32 %r547, %r864, %r1235;
mov.u32 %r544, %r642;

prmt.b32 %r544, %r544, 0, 0xba98;

xor.b32 %r1236, %r544, %r847;
or.b32 %r1237, %r1236, -2139062144;
and.b32 %r1238, %r544, 2139062143;
sub.s32 %r1239, %r1237, %r1238;
xor.b32 %r551, %r872, %r1239;

dp4a.s32.s32 %r546, %r547, %r499, %r647;


dp4a.s32.s32 %r550, %r551, %r503, %r546;

mov.u32 %r554, %r652;

prmt.b32 %r554, %r554, 0, 0xba98;

xor.b32 %r1240, %r554, %r874;
or.b32 %r1241, %r1240, -2139062144;
and.b32 %r1242, %r554, 2139062143;
sub.s32 %r1243, %r1241, %r1242;
xor.b32 %r559, %r892, %r1243;
mov.u32 %r556, %r654;

prmt.b32 %r556, %r556, 0, 0xba98;

xor.b32 %r1244, %r556, %r875;
or.b32 %r1245, %r1244, -2139062144;
and.b32 %r1246, %r556, 2139062143;
sub.s32 %r1247, %r1245, %r1246;
xor.b32 %r563, %r900, %r1247;

dp4a.s32.s32 %r558, %r559, %r511, %r550;


dp4a.s32.s32 %r562, %r563, %r515, %r558;

mul.lo.s32 %r1248, %r538, %r902;
mad.lo.s32 %r1249, %r562, %r903, %r1248;
add.s32 %r1250, %r562, %r538;
shr.u32 %r1251, %r1250, 31;
add.s32 %r1252, %r1250, %r1251;
shr.s32 %r1253, %r1252, 1;
add.s32 %r1254, %r1249, %r1253;
shr.s32 %r1255, %r1254, 31;
shr.u32 %r1256, %r1255, 30;
add.s32 %r1257, %r1254, %r1256;
shr.s32 %r1258, %r1257, 2;

{ cvt.f32.f16 %f91, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r517;
cvt.f32.f16 %f92, low;}


mul.ftz.f32 %f115, %f91, %f92;
cvt.rn.f32.s32 %f116, %r1258;
fma.rn.ftz.f32 %f267, %f115, %f116, %f267;
cvt.u64.u32 %rd82, %r1633;
add.s64 %rd83, %rd1, %rd82;
add.s64 %rd84, %rd83, %rd3;
mul.lo.s64 %rd85, %rd84, 36;

prmt.b32 %r567, %r567, 0, 0xba98;

xor.b32 %r1259, %r567, %r665;
or.b32 %r1260, %r1259, -2139062144;
and.b32 %r1261, %r567, 2139062143;
sub.s32 %r1262, %r1260, %r1261;
xor.b32 %r572, %r683, %r1262;
add.s64 %rd86, %rd10, %rd85;
ld.global.nc.u32 %r573, [%rd86+4];

prmt.b32 %r569, %r569, 0, 0xba98;

xor.b32 %r1263, %r569, %r666;
or.b32 %r1264, %r1263, -2139062144;
and.b32 %r1265, %r569, 2139062143;
sub.s32 %r1266, %r1264, %r1265;
xor.b32 %r576, %r691, %r1266;
ld.global.nc.u32 %r577, [%rd86+8];

dp4a.s32.s32 %r571, %r572, %r573, %r647;


dp4a.s32.s32 %r575, %r576, %r577, %r571;


prmt.b32 %r579, %r579, 0, 0xba98;

xor.b32 %r1267, %r579, %r693;
or.b32 %r1268, %r1267, -2139062144;
and.b32 %r1269, %r579, 2139062143;
sub.s32 %r1270, %r1268, %r1269;
xor.b32 %r584, %r711, %r1270;
ld.global.nc.u32 %r585, [%rd86+12];

prmt.b32 %r581, %r581, 0, 0xba98;

xor.b32 %r1271, %r581, %r694;
or.b32 %r1272, %r1271, -2139062144;
and.b32 %r1273, %r581, 2139062143;
sub.s32 %r1274, %r1272, %r1273;
xor.b32 %r588, %r719, %r1274;
ld.global.nc.u32 %r589, [%rd86+16];

dp4a.s32.s32 %r583, %r584, %r585, %r575;


dp4a.s32.s32 %r587, %r588, %r589, %r583;


prmt.b32 %r591, %r591, 0, 0xba98;

xor.b32 %r1275, %r591, %r721;
or.b32 %r1276, %r1275, -2139062144;
and.b32 %r1277, %r591, 2139062143;
sub.s32 %r1278, %r1276, %r1277;
xor.b32 %r596, %r739, %r1278;
ld.global.nc.u32 %r597, [%rd86+20];

prmt.b32 %r593, %r593, 0, 0xba98;

xor.b32 %r1279, %r593, %r722;
or.b32 %r1280, %r1279, -2139062144;
and.b32 %r1281, %r593, 2139062143;
sub.s32 %r1282, %r1280, %r1281;
xor.b32 %r600, %r747, %r1282;
ld.global.nc.u32 %r601, [%rd86+24];

dp4a.s32.s32 %r595, %r596, %r597, %r647;


dp4a.s32.s32 %r599, %r600, %r601, %r595;


prmt.b32 %r603, %r603, 0, 0xba98;

xor.b32 %r1283, %r603, %r749;
or.b32 %r1284, %r1283, -2139062144;
and.b32 %r1285, %r603, 2139062143;
sub.s32 %r1286, %r1284, %r1285;
xor.b32 %r608, %r767, %r1286;
ld.global.nc.u32 %r609, [%rd86+28];

prmt.b32 %r605, %r605, 0, 0xba98;

xor.b32 %r1287, %r605, %r750;
or.b32 %r1288, %r1287, -2139062144;
and.b32 %r1289, %r605, 2139062143;
sub.s32 %r1290, %r1288, %r1289;
xor.b32 %r612, %r775, %r1290;
ld.global.nc.u32 %r613, [%rd86+32];

dp4a.s32.s32 %r607, %r608, %r609, %r599;


dp4a.s32.s32 %r611, %r612, %r613, %r607;

mul.lo.s32 %r1291, %r587, %r777;
mad.lo.s32 %r1292, %r611, %r778, %r1291;
add.s32 %r1293, %r611, %r587;
shr.u32 %r1294, %r1293, 31;
add.s32 %r1295, %r1293, %r1294;
shr.s32 %r1296, %r1295, 1;
add.s32 %r1297, %r1292, %r1296;
shr.s32 %r1298, %r1297, 31;
shr.u32 %r1299, %r1298, 30;
add.s32 %r1300, %r1297, %r1299;
shr.s32 %r1301, %r1300, 2;

{ cvt.f32.f16 %f93, %rs1;}


ld.global.nc.u32 %r615, [%rd86];

{.reg .f16 low,high;
mov.b32 {low,high},%r615;
cvt.f32.f16 %f94, low;}


mul.ftz.f32 %f117, %f93, %f94;
cvt.rn.f32.s32 %f118, %r1301;
fma.rn.ftz.f32 %f266, %f117, %f118, %f266;

prmt.b32 %r616, %r616, 0, 0xba98;

xor.b32 %r1302, %r616, %r790;
or.b32 %r1303, %r1302, -2139062144;
and.b32 %r1304, %r616, 2139062143;
sub.s32 %r1305, %r1303, %r1304;
xor.b32 %r621, %r808, %r1305;

prmt.b32 %r618, %r618, 0, 0xba98;

xor.b32 %r1306, %r618, %r791;
or.b32 %r1307, %r1306, -2139062144;
and.b32 %r1308, %r618, 2139062143;
sub.s32 %r1309, %r1307, %r1308;
xor.b32 %r625, %r816, %r1309;

dp4a.s32.s32 %r620, %r621, %r573, %r647;


dp4a.s32.s32 %r624, %r625, %r577, %r620;


prmt.b32 %r628, %r628, 0, 0xba98;

xor.b32 %r1310, %r628, %r818;
or.b32 %r1311, %r1310, -2139062144;
and.b32 %r1312, %r628, 2139062143;
sub.s32 %r1313, %r1311, %r1312;
xor.b32 %r633, %r836, %r1313;

prmt.b32 %r630, %r630, 0, 0xba98;

xor.b32 %r1314, %r630, %r819;
or.b32 %r1315, %r1314, -2139062144;
and.b32 %r1316, %r630, 2139062143;
sub.s32 %r1317, %r1315, %r1316;
xor.b32 %r637, %r844, %r1317;

dp4a.s32.s32 %r632, %r633, %r585, %r624;


dp4a.s32.s32 %r636, %r637, %r589, %r632;


prmt.b32 %r640, %r640, 0, 0xba98;

xor.b32 %r1318, %r640, %r846;
or.b32 %r1319, %r1318, -2139062144;
and.b32 %r1320, %r640, 2139062143;
sub.s32 %r1321, %r1319, %r1320;
xor.b32 %r645, %r864, %r1321;

prmt.b32 %r642, %r642, 0, 0xba98;

xor.b32 %r1322, %r642, %r847;
or.b32 %r1323, %r1322, -2139062144;
and.b32 %r1324, %r642, 2139062143;
sub.s32 %r1325, %r1323, %r1324;
xor.b32 %r649, %r872, %r1325;

dp4a.s32.s32 %r644, %r645, %r597, %r647;


dp4a.s32.s32 %r648, %r649, %r601, %r644;


prmt.b32 %r652, %r652, 0, 0xba98;

xor.b32 %r1326, %r652, %r874;
or.b32 %r1327, %r1326, -2139062144;
and.b32 %r1328, %r652, 2139062143;
sub.s32 %r1329, %r1327, %r1328;
xor.b32 %r657, %r892, %r1329;

prmt.b32 %r654, %r654, 0, 0xba98;

xor.b32 %r1330, %r654, %r875;
or.b32 %r1331, %r1330, -2139062144;
and.b32 %r1332, %r654, 2139062143;
sub.s32 %r1333, %r1331, %r1332;
xor.b32 %r661, %r900, %r1333;

dp4a.s32.s32 %r656, %r657, %r609, %r648;


dp4a.s32.s32 %r660, %r661, %r613, %r656;

mul.lo.s32 %r1334, %r636, %r902;
mad.lo.s32 %r1335, %r660, %r903, %r1334;
add.s32 %r1336, %r660, %r636;
shr.u32 %r1337, %r1336, 31;
add.s32 %r1338, %r1336, %r1337;
shr.s32 %r1339, %r1338, 1;
add.s32 %r1340, %r1335, %r1339;
shr.s32 %r1341, %r1340, 31;
shr.u32 %r1342, %r1341, 30;
add.s32 %r1343, %r1340, %r1342;
shr.s32 %r1344, %r1343, 2;

{ cvt.f32.f16 %f95, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r615;
cvt.f32.f16 %f96, low;}


mul.ftz.f32 %f119, %f95, %f96;
cvt.rn.f32.s32 %f120, %r1344;
fma.rn.ftz.f32 %f265, %f119, %f120, %f265;
add.s32 %r1636, %r1636, 64;
add.s32 %r1635, %r1635, 64;
add.s32 %r1634, %r1634, 64;
add.s32 %r1633, %r1633, 64;
add.s32 %r1632, %r1632, 64;
add.s32 %r1631, %r1631, 64;
add.s64 %rd131, %rd131, 592;
add.s32 %r1637, %r1637, 8;
setp.lt.s32 %p4, %r1637, %r3;
@%p4 bra $L__BB221_2;

$L__BB221_3:
setp.eq.s32 %p5, %r42, 0;
@%p5 bra $L__BB221_5;

mov.u32 %r1348, _ZZ13mul_mat_vec_qIL9ggml_type17ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r1349, %r42, 1536, %r1348;
shl.b32 %r1350, %r44, 2;
add.s32 %r1351, %r1349, %r1350;
st.shared.f32 [%r1351+-1536], %f276;
st.shared.f32 [%r1351+-1408], %f275;
st.shared.f32 [%r1351+-1280], %f274;
st.shared.f32 [%r1351+-1152], %f273;
st.shared.f32 [%r1351+-1024], %f272;
st.shared.f32 [%r1351+-896], %f271;
st.shared.f32 [%r1351+-768], %f270;
st.shared.f32 [%r1351+-640], %f269;
st.shared.f32 [%r1351+-512], %f268;
st.shared.f32 [%r1351+-384], %f267;
st.shared.f32 [%r1351+-256], %f266;
st.shared.f32 [%r1351+-128], %f265;

$L__BB221_5:
setp.ne.s32 %p6, %r42, 0;
bar.sync 0;
@%p6 bra $L__BB221_30;

mov.u32 %r1353, %ctaid.z;
mov.u32 %r1354, %ctaid.y;
mov.u32 %r1355, %ctaid.x;
shl.b32 %r1356, %r1355, 1;
mov.u32 %r1357, 1;
mad.lo.s32 %r1358, %r1354, %r34, %r1356;
mad.lo.s32 %r1359, %r1353, %r40, %r1358;
cvt.u64.u32 %rd13, %r1359;
shl.b32 %r1361, %r44, 2;
mov.u32 %r1362, _ZZ13mul_mat_vec_qIL9ggml_type17ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r1363, %r1362, %r1361;
mov.u32 %r1364, 2;
ld.shared.f32 %f121, [%r1363];
add.ftz.f32 %f122, %f121, %f276;
mov.b32 %r1365, %f122;
mov.u32 %r1366, 31;
mov.u32 %r1367, 16;
mov.u32 %r1368, -1;
shfl.sync.bfly.b32 %r1369|%p7, %r1365, %r1367, %r1366, %r1368;
mov.b32 %f123, %r1369;
add.ftz.f32 %f124, %f122, %f123;
mov.b32 %r1370, %f124;
mov.u32 %r1371, 8;
shfl.sync.bfly.b32 %r1372|%p8, %r1370, %r1371, %r1366, %r1368;
mov.b32 %f125, %r1372;
add.ftz.f32 %f126, %f124, %f125;
mov.b32 %r1373, %f126;
mov.u32 %r1374, 4;
shfl.sync.bfly.b32 %r1375|%p9, %r1373, %r1374, %r1366, %r1368;
mov.b32 %f127, %r1375;
add.ftz.f32 %f128, %f126, %f127;
mov.b32 %r1376, %f128;
shfl.sync.bfly.b32 %r1377|%p10, %r1376, %r1364, %r1366, %r1368;
mov.b32 %f129, %r1377;
add.ftz.f32 %f130, %f128, %f129;
mov.b32 %r1378, %f130;
shfl.sync.bfly.b32 %r1379|%p11, %r1378, %r1357, %r1366, %r1368;
mov.b32 %f131, %r1379;
add.ftz.f32 %f37, %f130, %f131;
setp.eq.s32 %p12, %r44, 0;
setp.lt.u32 %p13, %r1356, %r28;
and.pred %p1, %p12, %p13;
not.pred %p14, %p1;
@%p14 bra $L__BB221_8;

cvta.to.global.u64 %rd87, %rd18;
mul.wide.u32 %rd88, %r1359, 4;
add.s64 %rd89, %rd87, %rd88;
st.global.f32 [%rd89], %f37;

$L__BB221_8:
ld.shared.f32 %f132, [%r1363+128];
add.ftz.f32 %f133, %f132, %f275;
mov.b32 %r1391, %f133;
shfl.sync.bfly.b32 %r1395|%p15, %r1391, %r1367, %r1366, %r1368;
mov.b32 %f134, %r1395;
add.ftz.f32 %f135, %f133, %f134;
mov.b32 %r1396, %f135;
shfl.sync.bfly.b32 %r1398|%p16, %r1396, %r1371, %r1366, %r1368;
mov.b32 %f136, %r1398;
add.ftz.f32 %f137, %f135, %f136;
mov.b32 %r1399, %f137;
shfl.sync.bfly.b32 %r1401|%p17, %r1399, %r1374, %r1366, %r1368;
mov.b32 %f138, %r1401;
add.ftz.f32 %f139, %f137, %f138;
mov.b32 %r1402, %f139;
shfl.sync.bfly.b32 %r1403|%p18, %r1402, %r1364, %r1366, %r1368;
mov.b32 %f140, %r1403;
add.ftz.f32 %f141, %f139, %f140;
mov.b32 %r1404, %f141;
shfl.sync.bfly.b32 %r1406|%p19, %r1404, %r1357, %r1366, %r1368;
mov.b32 %f142, %r1406;
add.ftz.f32 %f38, %f141, %f142;
setp.eq.s32 %p20, %r44, 1;
or.b32 %r1409, %r1356, 1;
setp.lt.u32 %p21, %r1409, %r28;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB221_10;

cvta.to.global.u64 %rd90, %rd18;
mul.wide.u32 %rd91, %r1359, 4;
add.s64 %rd92, %rd90, %rd91;
st.global.f32 [%rd92+4], %f38;

$L__BB221_10:
mov.u32 %r1420, 2;
ld.shared.f32 %f143, [%r1363+256];
add.ftz.f32 %f144, %f143, %f274;
mov.b32 %r1421, %f144;
mov.u32 %r1422, 31;
mov.u32 %r1423, 16;
mov.u32 %r1424, -1;
shfl.sync.bfly.b32 %r1425|%p23, %r1421, %r1423, %r1422, %r1424;
mov.b32 %f145, %r1425;
add.ftz.f32 %f146, %f144, %f145;
mov.b32 %r1426, %f146;
mov.u32 %r1427, 8;
shfl.sync.bfly.b32 %r1428|%p24, %r1426, %r1427, %r1422, %r1424;
mov.b32 %f147, %r1428;
add.ftz.f32 %f148, %f146, %f147;
mov.b32 %r1429, %f148;
mov.u32 %r1430, 4;
shfl.sync.bfly.b32 %r1431|%p25, %r1429, %r1430, %r1422, %r1424;
mov.b32 %f149, %r1431;
add.ftz.f32 %f150, %f148, %f149;
mov.b32 %r1432, %f150;
shfl.sync.bfly.b32 %r1433|%p26, %r1432, %r1420, %r1422, %r1424;
mov.b32 %f151, %r1433;
add.ftz.f32 %f152, %f150, %f151;
mov.b32 %r1434, %f152;
mov.u32 %r1435, 1;
shfl.sync.bfly.b32 %r1436|%p27, %r1434, %r1435, %r1422, %r1424;
mov.b32 %f153, %r1436;
add.ftz.f32 %f39, %f152, %f153;
@%p14 bra $L__BB221_12;

cvt.u64.u32 %rd93, %r28;
add.s64 %rd94, %rd93, %rd13;
cvta.to.global.u64 %rd95, %rd18;
shl.b64 %rd96, %rd94, 2;
add.s64 %rd97, %rd95, %rd96;
st.global.f32 [%rd97], %f39;

$L__BB221_12:
ld.shared.f32 %f154, [%r1363+384];
add.ftz.f32 %f155, %f154, %f273;
mov.b32 %r1442, %f155;
shfl.sync.bfly.b32 %r1446|%p29, %r1442, %r1423, %r1422, %r1424;
mov.b32 %f156, %r1446;
add.ftz.f32 %f157, %f155, %f156;
mov.b32 %r1447, %f157;
shfl.sync.bfly.b32 %r1449|%p30, %r1447, %r1427, %r1422, %r1424;
mov.b32 %f158, %r1449;
add.ftz.f32 %f159, %f157, %f158;
mov.b32 %r1450, %f159;
shfl.sync.bfly.b32 %r1452|%p31, %r1450, %r1430, %r1422, %r1424;
mov.b32 %f160, %r1452;
add.ftz.f32 %f161, %f159, %f160;
mov.b32 %r1453, %f161;
shfl.sync.bfly.b32 %r1454|%p32, %r1453, %r1420, %r1422, %r1424;
mov.b32 %f162, %r1454;
add.ftz.f32 %f163, %f161, %f162;
mov.b32 %r1455, %f163;
shfl.sync.bfly.b32 %r1457|%p33, %r1455, %r1435, %r1422, %r1424;
mov.b32 %f164, %r1457;
add.ftz.f32 %f40, %f163, %f164;
@%p22 bra $L__BB221_14;

add.s32 %r1458, %r28, 1;
cvt.u64.u32 %rd98, %r1458;
add.s64 %rd99, %rd98, %rd13;
cvta.to.global.u64 %rd100, %rd18;
shl.b64 %rd101, %rd99, 2;
add.s64 %rd102, %rd100, %rd101;
st.global.f32 [%rd102], %f40;

$L__BB221_14:
shl.b32 %r1459, %r28, 1;
mov.u32 %r1460, 1;
mov.u32 %r1465, 2;
ld.shared.f32 %f165, [%r1363+512];
add.ftz.f32 %f166, %f165, %f272;
mov.b32 %r1466, %f166;
mov.u32 %r1467, 31;
mov.u32 %r1468, 16;
mov.u32 %r1469, -1;
shfl.sync.bfly.b32 %r1470|%p35, %r1466, %r1468, %r1467, %r1469;
mov.b32 %f167, %r1470;
add.ftz.f32 %f168, %f166, %f167;
mov.b32 %r1471, %f168;
mov.u32 %r1472, 8;
shfl.sync.bfly.b32 %r1473|%p36, %r1471, %r1472, %r1467, %r1469;
mov.b32 %f169, %r1473;
add.ftz.f32 %f170, %f168, %f169;
mov.b32 %r1474, %f170;
mov.u32 %r1475, 4;
shfl.sync.bfly.b32 %r1476|%p37, %r1474, %r1475, %r1467, %r1469;
mov.b32 %f171, %r1476;
add.ftz.f32 %f172, %f170, %f171;
mov.b32 %r1477, %f172;
shfl.sync.bfly.b32 %r1478|%p38, %r1477, %r1465, %r1467, %r1469;
mov.b32 %f173, %r1478;
add.ftz.f32 %f174, %f172, %f173;
mov.b32 %r1479, %f174;
shfl.sync.bfly.b32 %r1480|%p39, %r1479, %r1460, %r1467, %r1469;
mov.b32 %f175, %r1480;
add.ftz.f32 %f41, %f174, %f175;
cvt.u64.u32 %rd103, %r1459;
add.s64 %rd104, %rd103, %rd13;
cvta.to.global.u64 %rd105, %rd18;
shl.b64 %rd106, %rd104, 2;
add.s64 %rd14, %rd105, %rd106;
@%p14 bra $L__BB221_16;

st.global.f32 [%rd14], %f41;

$L__BB221_16:
ld.shared.f32 %f176, [%r1363+640];
add.ftz.f32 %f177, %f176, %f271;
mov.b32 %r1486, %f177;
shfl.sync.bfly.b32 %r1490|%p41, %r1486, %r1468, %r1467, %r1469;
mov.b32 %f178, %r1490;
add.ftz.f32 %f179, %f177, %f178;
mov.b32 %r1491, %f179;
shfl.sync.bfly.b32 %r1493|%p42, %r1491, %r1472, %r1467, %r1469;
mov.b32 %f180, %r1493;
add.ftz.f32 %f181, %f179, %f180;
mov.b32 %r1494, %f181;
shfl.sync.bfly.b32 %r1496|%p43, %r1494, %r1475, %r1467, %r1469;
mov.b32 %f182, %r1496;
add.ftz.f32 %f183, %f181, %f182;
mov.b32 %r1497, %f183;
shfl.sync.bfly.b32 %r1498|%p44, %r1497, %r1465, %r1467, %r1469;
mov.b32 %f184, %r1498;
add.ftz.f32 %f185, %f183, %f184;
mov.b32 %r1499, %f185;
shfl.sync.bfly.b32 %r1501|%p45, %r1499, %r1460, %r1467, %r1469;
mov.b32 %f186, %r1501;
add.ftz.f32 %f42, %f185, %f186;
@%p22 bra $L__BB221_18;

st.global.f32 [%rd14+4], %f42;

$L__BB221_18:
mul.lo.s32 %r24, %r28, 3;
mov.u32 %r1506, 2;
ld.shared.f32 %f187, [%r1363+768];
add.ftz.f32 %f188, %f187, %f270;
mov.b32 %r1507, %f188;
mov.u32 %r1508, 31;
mov.u32 %r1509, 16;
mov.u32 %r1510, -1;
shfl.sync.bfly.b32 %r1511|%p47, %r1507, %r1509, %r1508, %r1510;
mov.b32 %f189, %r1511;
add.ftz.f32 %f190, %f188, %f189;
mov.b32 %r1512, %f190;
mov.u32 %r1513, 8;
shfl.sync.bfly.b32 %r1514|%p48, %r1512, %r1513, %r1508, %r1510;
mov.b32 %f191, %r1514;
add.ftz.f32 %f192, %f190, %f191;
mov.b32 %r1515, %f192;
mov.u32 %r1516, 4;
shfl.sync.bfly.b32 %r1517|%p49, %r1515, %r1516, %r1508, %r1510;
mov.b32 %f193, %r1517;
add.ftz.f32 %f194, %f192, %f193;
mov.b32 %r1518, %f194;
shfl.sync.bfly.b32 %r1519|%p50, %r1518, %r1506, %r1508, %r1510;
mov.b32 %f195, %r1519;
add.ftz.f32 %f196, %f194, %f195;
mov.b32 %r1520, %f196;
mov.u32 %r1521, 1;
shfl.sync.bfly.b32 %r1522|%p51, %r1520, %r1521, %r1508, %r1510;
mov.b32 %f197, %r1522;
add.ftz.f32 %f43, %f196, %f197;
@%p14 bra $L__BB221_20;

cvt.u64.u32 %rd107, %r24;
add.s64 %rd108, %rd107, %rd13;
shl.b64 %rd110, %rd108, 2;
add.s64 %rd111, %rd105, %rd110;
st.global.f32 [%rd111], %f43;

$L__BB221_20:
ld.shared.f32 %f198, [%r1363+896];
add.ftz.f32 %f199, %f198, %f269;
mov.b32 %r1528, %f199;
shfl.sync.bfly.b32 %r1532|%p53, %r1528, %r1509, %r1508, %r1510;
mov.b32 %f200, %r1532;
add.ftz.f32 %f201, %f199, %f200;
mov.b32 %r1533, %f201;
shfl.sync.bfly.b32 %r1535|%p54, %r1533, %r1513, %r1508, %r1510;
mov.b32 %f202, %r1535;
add.ftz.f32 %f203, %f201, %f202;
mov.b32 %r1536, %f203;
shfl.sync.bfly.b32 %r1538|%p55, %r1536, %r1516, %r1508, %r1510;
mov.b32 %f204, %r1538;
add.ftz.f32 %f205, %f203, %f204;
mov.b32 %r1539, %f205;
shfl.sync.bfly.b32 %r1540|%p56, %r1539, %r1506, %r1508, %r1510;
mov.b32 %f206, %r1540;
add.ftz.f32 %f207, %f205, %f206;
mov.b32 %r1541, %f207;
shfl.sync.bfly.b32 %r1543|%p57, %r1541, %r1521, %r1508, %r1510;
mov.b32 %f208, %r1543;
add.ftz.f32 %f44, %f207, %f208;
@%p22 bra $L__BB221_22;

add.s32 %r1544, %r24, 1;
cvt.u64.u32 %rd112, %r1544;
add.s64 %rd113, %rd112, %rd13;
shl.b64 %rd115, %rd113, 2;
add.s64 %rd116, %rd105, %rd115;
st.global.f32 [%rd116], %f44;

$L__BB221_22:
shl.b32 %r1545, %r28, 2;
mov.u32 %r1546, 2;
ld.shared.f32 %f209, [%r1363+1024];
add.ftz.f32 %f210, %f209, %f268;
mov.b32 %r1551, %f210;
mov.u32 %r1552, 31;
mov.u32 %r1553, 16;
mov.u32 %r1554, -1;
shfl.sync.bfly.b32 %r1555|%p59, %r1551, %r1553, %r1552, %r1554;
mov.b32 %f211, %r1555;
add.ftz.f32 %f212, %f210, %f211;
mov.b32 %r1556, %f212;
mov.u32 %r1557, 8;
shfl.sync.bfly.b32 %r1558|%p60, %r1556, %r1557, %r1552, %r1554;
mov.b32 %f213, %r1558;
add.ftz.f32 %f214, %f212, %f213;
mov.b32 %r1559, %f214;
mov.u32 %r1560, 4;
shfl.sync.bfly.b32 %r1561|%p61, %r1559, %r1560, %r1552, %r1554;
mov.b32 %f215, %r1561;
add.ftz.f32 %f216, %f214, %f215;
mov.b32 %r1562, %f216;
shfl.sync.bfly.b32 %r1563|%p62, %r1562, %r1546, %r1552, %r1554;
mov.b32 %f217, %r1563;
add.ftz.f32 %f218, %f216, %f217;
mov.b32 %r1564, %f218;
mov.u32 %r1565, 1;
shfl.sync.bfly.b32 %r1566|%p63, %r1564, %r1565, %r1552, %r1554;
mov.b32 %f219, %r1566;
add.ftz.f32 %f45, %f218, %f219;
cvt.u64.u32 %rd117, %r1545;
add.s64 %rd118, %rd117, %rd13;
shl.b64 %rd120, %rd118, 2;
add.s64 %rd15, %rd105, %rd120;
@%p14 bra $L__BB221_24;

st.global.f32 [%rd15], %f45;

$L__BB221_24:
ld.shared.f32 %f220, [%r1363+1152];
add.ftz.f32 %f221, %f220, %f267;
mov.b32 %r1572, %f221;
shfl.sync.bfly.b32 %r1576|%p65, %r1572, %r1553, %r1552, %r1554;
mov.b32 %f222, %r1576;
add.ftz.f32 %f223, %f221, %f222;
mov.b32 %r1577, %f223;
shfl.sync.bfly.b32 %r1579|%p66, %r1577, %r1557, %r1552, %r1554;
mov.b32 %f224, %r1579;
add.ftz.f32 %f225, %f223, %f224;
mov.b32 %r1580, %f225;
shfl.sync.bfly.b32 %r1582|%p67, %r1580, %r1560, %r1552, %r1554;
mov.b32 %f226, %r1582;
add.ftz.f32 %f227, %f225, %f226;
mov.b32 %r1583, %f227;
shfl.sync.bfly.b32 %r1584|%p68, %r1583, %r1546, %r1552, %r1554;
mov.b32 %f228, %r1584;
add.ftz.f32 %f229, %f227, %f228;
mov.b32 %r1585, %f229;
shfl.sync.bfly.b32 %r1587|%p69, %r1585, %r1565, %r1552, %r1554;
mov.b32 %f230, %r1587;
add.ftz.f32 %f46, %f229, %f230;
@%p22 bra $L__BB221_26;

st.global.f32 [%rd15+4], %f46;

$L__BB221_26:
mul.lo.s32 %r25, %r28, 5;
mov.u32 %r1592, 2;
ld.shared.f32 %f231, [%r1363+1280];
add.ftz.f32 %f232, %f231, %f266;
mov.b32 %r1593, %f232;
mov.u32 %r1594, 31;
mov.u32 %r1595, 16;
mov.u32 %r1596, -1;
shfl.sync.bfly.b32 %r1597|%p71, %r1593, %r1595, %r1594, %r1596;
mov.b32 %f233, %r1597;
add.ftz.f32 %f234, %f232, %f233;
mov.b32 %r1598, %f234;
mov.u32 %r1599, 8;
shfl.sync.bfly.b32 %r1600|%p72, %r1598, %r1599, %r1594, %r1596;
mov.b32 %f235, %r1600;
add.ftz.f32 %f236, %f234, %f235;
mov.b32 %r1601, %f236;
mov.u32 %r1602, 4;
shfl.sync.bfly.b32 %r1603|%p73, %r1601, %r1602, %r1594, %r1596;
mov.b32 %f237, %r1603;
add.ftz.f32 %f238, %f236, %f237;
mov.b32 %r1604, %f238;
shfl.sync.bfly.b32 %r1605|%p74, %r1604, %r1592, %r1594, %r1596;
mov.b32 %f239, %r1605;
add.ftz.f32 %f240, %f238, %f239;
mov.b32 %r1606, %f240;
mov.u32 %r1607, 1;
shfl.sync.bfly.b32 %r1608|%p75, %r1606, %r1607, %r1594, %r1596;
mov.b32 %f241, %r1608;
add.ftz.f32 %f47, %f240, %f241;
@%p14 bra $L__BB221_28;

cvt.u64.u32 %rd121, %r25;
add.s64 %rd122, %rd121, %rd13;
shl.b64 %rd124, %rd122, 2;
add.s64 %rd125, %rd105, %rd124;
st.global.f32 [%rd125], %f47;

$L__BB221_28:
ld.shared.f32 %f242, [%r1363+1408];
add.ftz.f32 %f243, %f242, %f265;
mov.b32 %r1614, %f243;
shfl.sync.bfly.b32 %r1618|%p77, %r1614, %r1595, %r1594, %r1596;
mov.b32 %f244, %r1618;
add.ftz.f32 %f245, %f243, %f244;
mov.b32 %r1619, %f245;
shfl.sync.bfly.b32 %r1621|%p78, %r1619, %r1599, %r1594, %r1596;
mov.b32 %f246, %r1621;
add.ftz.f32 %f247, %f245, %f246;
mov.b32 %r1622, %f247;
shfl.sync.bfly.b32 %r1624|%p79, %r1622, %r1602, %r1594, %r1596;
mov.b32 %f248, %r1624;
add.ftz.f32 %f249, %f247, %f248;
mov.b32 %r1625, %f249;
shfl.sync.bfly.b32 %r1626|%p80, %r1625, %r1592, %r1594, %r1596;
mov.b32 %f250, %r1626;
add.ftz.f32 %f251, %f249, %f250;
mov.b32 %r1627, %f251;
shfl.sync.bfly.b32 %r1629|%p81, %r1627, %r1607, %r1594, %r1596;
mov.b32 %f252, %r1629;
add.ftz.f32 %f48, %f251, %f252;
@%p22 bra $L__BB221_30;

add.s32 %r1630, %r25, 1;
cvt.u64.u32 %rd126, %r1630;
add.s64 %rd127, %rd126, %rd13;
shl.b64 %rd129, %rd127, 2;
add.s64 %rd130, %rd105, %rd129;
st.global.f32 [%rd130], %f48;

$L__BB221_30:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type17ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type17ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type17ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type17ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type17ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type17ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type17ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type17ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type17ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<106>;
.reg .b16 %rs<47>;
.reg .f32 %f<323>;
.reg .b32 %r<1926>;
.reg .b64 %rd<150>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type17ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1792];

ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type17ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type17ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type17ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type17ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type17ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type17ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type17ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type17ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type17ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type17ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type17ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type17ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type17ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type17ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type17ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type17ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type17ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r44, %tid.y;
shl.b32 %r45, %r44, 5;
mov.u32 %r46, %tid.x;
add.s32 %r1, %r45, %r46;
shr.s32 %r47, %r1, 31;
shr.u32 %r48, %r47, 29;
add.s32 %r49, %r1, %r48;
shr.s32 %r1925, %r49, 3;
shr.u32 %r50, %r28, 8;
setp.ge.s32 %p2, %r1925, %r50;
mov.f32 %f309, 0f00000000;
mov.f32 %f310, %f309;
mov.f32 %f311, %f309;
mov.f32 %f312, %f309;
mov.f32 %f313, %f309;
mov.f32 %f314, %f309;
mov.f32 %f315, %f309;
mov.f32 %f316, %f309;
mov.f32 %f317, %f309;
mov.f32 %f318, %f309;
mov.f32 %f319, %f309;
mov.f32 %f320, %f309;
mov.f32 %f321, %f309;
mov.f32 %f322, %f309;
@%p2 bra $L__BB222_3;

mov.u32 %r51, %ctaid.x;
shl.b32 %r52, %r51, 1;
mov.u32 %r53, %ctaid.y;
mov.u32 %r54, %ctaid.z;
mul.lo.s32 %r55, %r53, %r36;
cvt.u64.u32 %rd19, %r55;
mul.lo.s32 %r56, %r54, %r42;
cvt.u64.u32 %rd20, %r56;
add.s64 %rd1, %rd20, %rd19;
mul.hi.u32 %r57, %r54, %r38;
add.s32 %r58, %r57, %r54;
shr.u32 %r59, %r58, %r39;
mul.lo.s32 %r60, %r59, %r41;
mul.hi.u32 %r61, %r53, %r32;
add.s32 %r62, %r61, %r53;
shr.u32 %r63, %r62, %r33;
mul.lo.s32 %r64, %r63, %r35;
cvta.to.global.u64 %rd149, %rd16;
and.b32 %r68, %r49, -8;
sub.s32 %r69, %r1, %r68;
shl.b32 %r70, %r69, 2;
cvt.s64.s32 %rd3, %r69;
shl.b32 %r71, %r30, 1;
shl.b32 %r1918, %r1925, 3;
add.s32 %r1924, %r1918, %r71;
mad.lo.s32 %r1923, %r30, 3, %r1918;
shl.b32 %r72, %r30, 2;
add.s32 %r1922, %r1918, %r72;
mad.lo.s32 %r1921, %r30, 5, %r1918;
mad.lo.s32 %r1920, %r30, 6, %r1918;
add.s32 %r1919, %r30, %r1918;
add.s32 %r73, %r1925, %r60;
add.s32 %r74, %r73, %r64;
or.b32 %r75, %r52, 1;
mad.lo.s32 %r76, %r29, %r75, %r74;
mul.wide.s32 %rd4, %r76, 74;
add.s64 %rd21, %rd4, %rd3;
add.s64 %rd5, %rd21, 66;
mad.lo.s32 %r77, %r52, %r29, %r1925;
add.s32 %r78, %r77, %r60;
add.s32 %r79, %r78, %r64;
mul.wide.s32 %rd6, %r79, 74;
add.s64 %rd22, %rd6, %rd3;
add.s64 %rd7, %rd22, 66;
mul.wide.s32 %rd23, %r70, 2;
add.s64 %rd24, %rd4, %rd23;
add.s64 %rd8, %rd24, 4;
add.s64 %rd25, %rd6, %rd23;
add.s64 %rd9, %rd25, 8;
cvta.to.global.u64 %rd10, %rd17;
mov.f32 %f309, 0f00000000;

$L__BB222_2:
cvt.u64.u32 %rd26, %r1918;
add.s64 %rd27, %rd1, %rd26;
add.s64 %rd28, %rd27, %rd3;
mul.lo.s64 %rd29, %rd28, 36;
add.s64 %rd30, %rd149, %rd9;
ld.global.nc.u16 %rs15, [%rd30+-4];
ld.global.nc.u16 %rs16, [%rd30+-2];
ld.global.nc.u16 %rs17, [%rd30];
add.s64 %rd31, %rd149, %rd7;
ld.global.nc.u8 %rs18, [%rd31];
and.b16 %rs19, %rs18, 240;
and.b16 %rs20, %rs18, 15;
shr.u16 %rs21, %rs19, 4;
ld.global.nc.u16 %rs22, [%rd30+-6];
shl.b16 %rs23, %rs22, 3;
cvt.u64.u16 %rd32, %rs23;
and.b64 %rd33, %rd32, 4088;
mov.u64 %rd34, iq2xs_grid;
add.s64 %rd35, %rd34, %rd33;
ld.global.nc.v2.u32 {%r766, %r767}, [%rd35];
not.b32 %r769, %r766;
mov.u32 %r748, 0;
not.b32 %r771, %r767;
shr.u16 %rs24, %rs22, 9;
cvt.u32.u16 %r772, %rs24;
popc.b32 %r773, %r772;
shl.b32 %r774, %r773, 7;
and.b32 %r775, %r774, 128;
or.b32 %r776, %r775, %r772;
mul.lo.s32 %r777, %r776, 16843009;
and.b32 %r778, %r777, 134480385;
or.b32 %r779, %r778, -2139062144;
add.s32 %r780, %r779, -16843009;
or.b32 %r668, %r780, %r778;
mov.u32 %r80, %r668;

prmt.b32 %r80, %r80, 0, 0xba98;

xor.b32 %r781, %r80, %r766;
or.b32 %r782, %r781, -2139062144;
and.b32 %r783, %r80, 2139062143;
and.b32 %r784, %r769, -2139062144;
sub.s32 %r785, %r782, %r783;
xor.b32 %r85, %r784, %r785;
add.s64 %rd36, %rd10, %rd29;
ld.global.nc.u32 %r86, [%rd36+4];
and.b32 %r786, %r777, -2143281136;
or.b32 %r787, %r786, -2139062144;
add.s32 %r788, %r787, -16843009;
or.b32 %r670, %r788, %r786;
mov.u32 %r82, %r670;

prmt.b32 %r82, %r82, 0, 0xba98;

xor.b32 %r789, %r82, %r767;
or.b32 %r790, %r789, -2139062144;
and.b32 %r791, %r82, 2139062143;
and.b32 %r792, %r771, -2139062144;
sub.s32 %r793, %r790, %r791;
xor.b32 %r89, %r792, %r793;
ld.global.nc.u32 %r90, [%rd36+8];

dp4a.s32.s32 %r84, %r85, %r86, %r748;


dp4a.s32.s32 %r88, %r89, %r90, %r84;

shl.b16 %rs25, %rs15, 3;
cvt.u64.u16 %rd37, %rs25;
and.b64 %rd38, %rd37, 4088;
add.s64 %rd39, %rd34, %rd38;
ld.global.nc.v2.u32 {%r794, %r795}, [%rd39];
not.b32 %r797, %r794;
not.b32 %r799, %r795;
shr.u16 %rs26, %rs15, 9;
cvt.u32.u16 %r800, %rs26;
popc.b32 %r801, %r800;
shl.b32 %r802, %r801, 7;
and.b32 %r803, %r802, 128;
or.b32 %r804, %r803, %r800;
mul.lo.s32 %r805, %r804, 16843009;
and.b32 %r806, %r805, 134480385;
or.b32 %r807, %r806, -2139062144;
add.s32 %r808, %r807, -16843009;
or.b32 %r680, %r808, %r806;
mov.u32 %r92, %r680;

prmt.b32 %r92, %r92, 0, 0xba98;

xor.b32 %r809, %r92, %r794;
or.b32 %r810, %r809, -2139062144;
and.b32 %r811, %r92, 2139062143;
and.b32 %r812, %r797, -2139062144;
sub.s32 %r813, %r810, %r811;
xor.b32 %r97, %r812, %r813;
ld.global.nc.u32 %r98, [%rd36+12];
and.b32 %r814, %r805, -2143281136;
or.b32 %r815, %r814, -2139062144;
add.s32 %r816, %r815, -16843009;
or.b32 %r682, %r816, %r814;
mov.u32 %r94, %r682;

prmt.b32 %r94, %r94, 0, 0xba98;

xor.b32 %r817, %r94, %r795;
or.b32 %r818, %r817, -2139062144;
and.b32 %r819, %r94, 2139062143;
and.b32 %r820, %r799, -2139062144;
sub.s32 %r821, %r818, %r819;
xor.b32 %r101, %r820, %r821;
ld.global.nc.u32 %r102, [%rd36+16];

dp4a.s32.s32 %r96, %r97, %r98, %r88;


dp4a.s32.s32 %r100, %r101, %r102, %r96;

shl.b16 %rs27, %rs16, 3;
cvt.u64.u16 %rd40, %rs27;
and.b64 %rd41, %rd40, 4088;
add.s64 %rd42, %rd34, %rd41;
ld.global.nc.v2.u32 {%r822, %r823}, [%rd42];
not.b32 %r825, %r822;
not.b32 %r827, %r823;
shr.u16 %rs28, %rs16, 9;
cvt.u32.u16 %r828, %rs28;
popc.b32 %r829, %r828;
shl.b32 %r830, %r829, 7;
and.b32 %r831, %r830, 128;
or.b32 %r832, %r831, %r828;
mul.lo.s32 %r833, %r832, 16843009;
and.b32 %r834, %r833, 134480385;
or.b32 %r835, %r834, -2139062144;
add.s32 %r836, %r835, -16843009;
or.b32 %r692, %r836, %r834;
mov.u32 %r104, %r692;

prmt.b32 %r104, %r104, 0, 0xba98;

xor.b32 %r837, %r104, %r822;
or.b32 %r838, %r837, -2139062144;
and.b32 %r839, %r104, 2139062143;
and.b32 %r840, %r825, -2139062144;
sub.s32 %r841, %r838, %r839;
xor.b32 %r109, %r840, %r841;
ld.global.nc.u32 %r110, [%rd36+20];
and.b32 %r842, %r833, -2143281136;
or.b32 %r843, %r842, -2139062144;
add.s32 %r844, %r843, -16843009;
or.b32 %r694, %r844, %r842;
mov.u32 %r106, %r694;

prmt.b32 %r106, %r106, 0, 0xba98;

xor.b32 %r845, %r106, %r823;
or.b32 %r846, %r845, -2139062144;
and.b32 %r847, %r106, 2139062143;
and.b32 %r848, %r827, -2139062144;
sub.s32 %r849, %r846, %r847;
xor.b32 %r113, %r848, %r849;
ld.global.nc.u32 %r114, [%rd36+24];

dp4a.s32.s32 %r108, %r109, %r110, %r748;


dp4a.s32.s32 %r112, %r113, %r114, %r108;

shl.b16 %rs29, %rs17, 3;
cvt.u64.u16 %rd43, %rs29;
and.b64 %rd44, %rd43, 4088;
add.s64 %rd45, %rd34, %rd44;
ld.global.nc.v2.u32 {%r850, %r851}, [%rd45];
not.b32 %r853, %r850;
not.b32 %r855, %r851;
shr.u16 %rs30, %rs17, 9;
cvt.u32.u16 %r856, %rs30;
popc.b32 %r857, %r856;
shl.b32 %r858, %r857, 7;
and.b32 %r859, %r858, 128;
or.b32 %r860, %r859, %r856;
mul.lo.s32 %r861, %r860, 16843009;
and.b32 %r862, %r861, 134480385;
or.b32 %r863, %r862, -2139062144;
add.s32 %r864, %r863, -16843009;
or.b32 %r704, %r864, %r862;
mov.u32 %r116, %r704;

prmt.b32 %r116, %r116, 0, 0xba98;

xor.b32 %r865, %r116, %r850;
or.b32 %r866, %r865, -2139062144;
and.b32 %r867, %r116, 2139062143;
and.b32 %r868, %r853, -2139062144;
sub.s32 %r869, %r866, %r867;
xor.b32 %r121, %r868, %r869;
ld.global.nc.u32 %r122, [%rd36+28];
and.b32 %r870, %r861, -2143281136;
or.b32 %r871, %r870, -2139062144;
add.s32 %r872, %r871, -16843009;
or.b32 %r706, %r872, %r870;
mov.u32 %r118, %r706;

prmt.b32 %r118, %r118, 0, 0xba98;

xor.b32 %r873, %r118, %r851;
or.b32 %r874, %r873, -2139062144;
and.b32 %r875, %r118, 2139062143;
and.b32 %r876, %r855, -2139062144;
sub.s32 %r877, %r874, %r875;
xor.b32 %r125, %r876, %r877;
ld.global.nc.u32 %r126, [%rd36+32];

dp4a.s32.s32 %r120, %r121, %r122, %r112;


dp4a.s32.s32 %r124, %r125, %r126, %r120;

cvt.u32.u16 %r878, %rs20;
cvt.u32.u16 %r879, %rs21;
mul.lo.s32 %r880, %r100, %r878;
mad.lo.s32 %r881, %r124, %r879, %r880;
add.s32 %r882, %r124, %r100;
shr.u32 %r883, %r882, 31;
add.s32 %r884, %r882, %r883;
shr.s32 %r885, %r884, 1;
add.s32 %r886, %r881, %r885;
shr.s32 %r887, %r886, 31;
shr.u32 %r888, %r887, 30;
add.s32 %r889, %r886, %r888;
shr.s32 %r890, %r889, 2;
add.s64 %rd46, %rd149, %rd6;
ld.global.nc.u16 %rs1, [%rd46];

{ cvt.f32.f16 %f85, %rs1;}


ld.global.nc.u32 %r128, [%rd36];

{.reg .f16 low,high;
mov.b32 {low,high},%r128;
cvt.f32.f16 %f86, low;}


mul.ftz.f32 %f113, %f85, %f86;
cvt.rn.f32.s32 %f114, %r890;
fma.rn.ftz.f32 %f322, %f113, %f114, %f322;
add.s64 %rd47, %rd149, %rd8;
ld.global.nc.u16 %rs31, [%rd47];
ld.global.nc.u16 %rs32, [%rd47+2];
ld.global.nc.u16 %rs33, [%rd47+4];
add.s64 %rd48, %rd149, %rd5;
ld.global.nc.u8 %rs34, [%rd48];
and.b16 %rs35, %rs34, 240;
and.b16 %rs36, %rs34, 15;
shr.u16 %rs37, %rs35, 4;
ld.global.nc.u16 %rs38, [%rd47+-2];
shl.b16 %rs39, %rs38, 3;
cvt.u64.u16 %rd49, %rs39;
and.b64 %rd50, %rd49, 4088;
add.s64 %rd51, %rd34, %rd50;
ld.global.nc.v2.u32 {%r891, %r892}, [%rd51];
not.b32 %r894, %r891;
not.b32 %r896, %r892;
shr.u16 %rs40, %rs38, 9;
cvt.u32.u16 %r897, %rs40;
popc.b32 %r898, %r897;
shl.b32 %r899, %r898, 7;
and.b32 %r900, %r899, 128;
or.b32 %r901, %r900, %r897;
mul.lo.s32 %r902, %r901, 16843009;
and.b32 %r903, %r902, 134480385;
or.b32 %r904, %r903, -2139062144;
add.s32 %r905, %r904, -16843009;
or.b32 %r717, %r905, %r903;
mov.u32 %r129, %r717;

prmt.b32 %r129, %r129, 0, 0xba98;

xor.b32 %r906, %r129, %r891;
or.b32 %r907, %r906, -2139062144;
and.b32 %r908, %r129, 2139062143;
and.b32 %r909, %r894, -2139062144;
sub.s32 %r910, %r907, %r908;
xor.b32 %r134, %r909, %r910;
and.b32 %r911, %r902, -2143281136;
or.b32 %r912, %r911, -2139062144;
add.s32 %r913, %r912, -16843009;
or.b32 %r719, %r913, %r911;
mov.u32 %r131, %r719;

prmt.b32 %r131, %r131, 0, 0xba98;

xor.b32 %r914, %r131, %r892;
or.b32 %r915, %r914, -2139062144;
and.b32 %r916, %r131, 2139062143;
and.b32 %r917, %r896, -2139062144;
sub.s32 %r918, %r915, %r916;
xor.b32 %r138, %r917, %r918;

dp4a.s32.s32 %r133, %r134, %r86, %r748;


dp4a.s32.s32 %r137, %r138, %r90, %r133;

shl.b16 %rs41, %rs31, 3;
cvt.u64.u16 %rd52, %rs41;
and.b64 %rd53, %rd52, 4088;
add.s64 %rd54, %rd34, %rd53;
ld.global.nc.v2.u32 {%r919, %r920}, [%rd54];
not.b32 %r922, %r919;
not.b32 %r924, %r920;
shr.u16 %rs42, %rs31, 9;
cvt.u32.u16 %r925, %rs42;
popc.b32 %r926, %r925;
shl.b32 %r927, %r926, 7;
and.b32 %r928, %r927, 128;
or.b32 %r929, %r928, %r925;
mul.lo.s32 %r930, %r929, 16843009;
and.b32 %r931, %r930, 134480385;
or.b32 %r932, %r931, -2139062144;
add.s32 %r933, %r932, -16843009;
or.b32 %r729, %r933, %r931;
mov.u32 %r141, %r729;

prmt.b32 %r141, %r141, 0, 0xba98;

xor.b32 %r934, %r141, %r919;
or.b32 %r935, %r934, -2139062144;
and.b32 %r936, %r141, 2139062143;
and.b32 %r937, %r922, -2139062144;
sub.s32 %r938, %r935, %r936;
xor.b32 %r146, %r937, %r938;
and.b32 %r939, %r930, -2143281136;
or.b32 %r940, %r939, -2139062144;
add.s32 %r941, %r940, -16843009;
or.b32 %r731, %r941, %r939;
mov.u32 %r143, %r731;

prmt.b32 %r143, %r143, 0, 0xba98;

xor.b32 %r942, %r143, %r920;
or.b32 %r943, %r942, -2139062144;
and.b32 %r944, %r143, 2139062143;
and.b32 %r945, %r924, -2139062144;
sub.s32 %r946, %r943, %r944;
xor.b32 %r150, %r945, %r946;

dp4a.s32.s32 %r145, %r146, %r98, %r137;


dp4a.s32.s32 %r149, %r150, %r102, %r145;

shl.b16 %rs43, %rs32, 3;
cvt.u64.u16 %rd55, %rs43;
and.b64 %rd56, %rd55, 4088;
add.s64 %rd57, %rd34, %rd56;
ld.global.nc.v2.u32 {%r947, %r948}, [%rd57];
not.b32 %r950, %r947;
not.b32 %r952, %r948;
shr.u16 %rs44, %rs32, 9;
cvt.u32.u16 %r953, %rs44;
popc.b32 %r954, %r953;
shl.b32 %r955, %r954, 7;
and.b32 %r956, %r955, 128;
or.b32 %r957, %r956, %r953;
mul.lo.s32 %r958, %r957, 16843009;
and.b32 %r959, %r958, 134480385;
or.b32 %r960, %r959, -2139062144;
add.s32 %r961, %r960, -16843009;
or.b32 %r741, %r961, %r959;
mov.u32 %r153, %r741;

prmt.b32 %r153, %r153, 0, 0xba98;

xor.b32 %r962, %r153, %r947;
or.b32 %r963, %r962, -2139062144;
and.b32 %r964, %r153, 2139062143;
and.b32 %r965, %r950, -2139062144;
sub.s32 %r966, %r963, %r964;
xor.b32 %r158, %r965, %r966;
and.b32 %r967, %r958, -2143281136;
or.b32 %r968, %r967, -2139062144;
add.s32 %r969, %r968, -16843009;
or.b32 %r743, %r969, %r967;
mov.u32 %r155, %r743;

prmt.b32 %r155, %r155, 0, 0xba98;

xor.b32 %r970, %r155, %r948;
or.b32 %r971, %r970, -2139062144;
and.b32 %r972, %r155, 2139062143;
and.b32 %r973, %r952, -2139062144;
sub.s32 %r974, %r971, %r972;
xor.b32 %r162, %r973, %r974;

dp4a.s32.s32 %r157, %r158, %r110, %r748;


dp4a.s32.s32 %r161, %r162, %r114, %r157;

shl.b16 %rs45, %rs33, 3;
cvt.u64.u16 %rd58, %rs45;
and.b64 %rd59, %rd58, 4088;
add.s64 %rd60, %rd34, %rd59;
ld.global.nc.v2.u32 {%r975, %r976}, [%rd60];
not.b32 %r978, %r975;
not.b32 %r980, %r976;
shr.u16 %rs46, %rs33, 9;
cvt.u32.u16 %r981, %rs46;
popc.b32 %r982, %r981;
shl.b32 %r983, %r982, 7;
and.b32 %r984, %r983, 128;
or.b32 %r985, %r984, %r981;
mul.lo.s32 %r986, %r985, 16843009;
and.b32 %r987, %r986, 134480385;
or.b32 %r988, %r987, -2139062144;
add.s32 %r989, %r988, -16843009;
or.b32 %r753, %r989, %r987;
mov.u32 %r165, %r753;

prmt.b32 %r165, %r165, 0, 0xba98;

xor.b32 %r990, %r165, %r975;
or.b32 %r991, %r990, -2139062144;
and.b32 %r992, %r165, 2139062143;
and.b32 %r993, %r978, -2139062144;
sub.s32 %r994, %r991, %r992;
xor.b32 %r170, %r993, %r994;
and.b32 %r995, %r986, -2143281136;
or.b32 %r996, %r995, -2139062144;
add.s32 %r997, %r996, -16843009;
or.b32 %r755, %r997, %r995;
mov.u32 %r167, %r755;

prmt.b32 %r167, %r167, 0, 0xba98;

xor.b32 %r998, %r167, %r976;
or.b32 %r999, %r998, -2139062144;
and.b32 %r1000, %r167, 2139062143;
and.b32 %r1001, %r980, -2139062144;
sub.s32 %r1002, %r999, %r1000;
xor.b32 %r174, %r1001, %r1002;

dp4a.s32.s32 %r169, %r170, %r122, %r161;


dp4a.s32.s32 %r173, %r174, %r126, %r169;

cvt.u32.u16 %r1003, %rs36;
cvt.u32.u16 %r1004, %rs37;
mul.lo.s32 %r1005, %r149, %r1003;
mad.lo.s32 %r1006, %r173, %r1004, %r1005;
add.s32 %r1007, %r173, %r149;
shr.u32 %r1008, %r1007, 31;
add.s32 %r1009, %r1007, %r1008;
shr.s32 %r1010, %r1009, 1;
add.s32 %r1011, %r1006, %r1010;
shr.s32 %r1012, %r1011, 31;
shr.u32 %r1013, %r1012, 30;
add.s32 %r1014, %r1011, %r1013;
shr.s32 %r1015, %r1014, 2;
add.s64 %rd61, %rd149, %rd4;
ld.global.nc.u16 %rs2, [%rd61];

{ cvt.f32.f16 %f87, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r128;
cvt.f32.f16 %f88, low;}


mul.ftz.f32 %f115, %f87, %f88;
cvt.rn.f32.s32 %f116, %r1015;
fma.rn.ftz.f32 %f321, %f115, %f116, %f321;
cvt.u64.u32 %rd62, %r1919;
add.s64 %rd63, %rd1, %rd62;
add.s64 %rd64, %rd63, %rd3;
mul.lo.s64 %rd65, %rd64, 36;
mov.u32 %r178, %r668;

prmt.b32 %r178, %r178, 0, 0xba98;

xor.b32 %r1016, %r178, %r766;
or.b32 %r1017, %r1016, -2139062144;
and.b32 %r1018, %r178, 2139062143;
sub.s32 %r1019, %r1017, %r1018;
xor.b32 %r183, %r784, %r1019;
add.s64 %rd66, %rd10, %rd65;
ld.global.nc.u32 %r184, [%rd66+4];
mov.u32 %r180, %r670;

prmt.b32 %r180, %r180, 0, 0xba98;

xor.b32 %r1020, %r180, %r767;
or.b32 %r1021, %r1020, -2139062144;
and.b32 %r1022, %r180, 2139062143;
sub.s32 %r1023, %r1021, %r1022;
xor.b32 %r187, %r792, %r1023;
ld.global.nc.u32 %r188, [%rd66+8];

dp4a.s32.s32 %r182, %r183, %r184, %r748;


dp4a.s32.s32 %r186, %r187, %r188, %r182;

mov.u32 %r190, %r680;

prmt.b32 %r190, %r190, 0, 0xba98;

xor.b32 %r1024, %r190, %r794;
or.b32 %r1025, %r1024, -2139062144;
and.b32 %r1026, %r190, 2139062143;
sub.s32 %r1027, %r1025, %r1026;
xor.b32 %r195, %r812, %r1027;
ld.global.nc.u32 %r196, [%rd66+12];
mov.u32 %r192, %r682;

prmt.b32 %r192, %r192, 0, 0xba98;

xor.b32 %r1028, %r192, %r795;
or.b32 %r1029, %r1028, -2139062144;
and.b32 %r1030, %r192, 2139062143;
sub.s32 %r1031, %r1029, %r1030;
xor.b32 %r199, %r820, %r1031;
ld.global.nc.u32 %r200, [%rd66+16];

dp4a.s32.s32 %r194, %r195, %r196, %r186;


dp4a.s32.s32 %r198, %r199, %r200, %r194;

mov.u32 %r202, %r692;

prmt.b32 %r202, %r202, 0, 0xba98;

xor.b32 %r1032, %r202, %r822;
or.b32 %r1033, %r1032, -2139062144;
and.b32 %r1034, %r202, 2139062143;
sub.s32 %r1035, %r1033, %r1034;
xor.b32 %r207, %r840, %r1035;
ld.global.nc.u32 %r208, [%rd66+20];
mov.u32 %r204, %r694;

prmt.b32 %r204, %r204, 0, 0xba98;

xor.b32 %r1036, %r204, %r823;
or.b32 %r1037, %r1036, -2139062144;
and.b32 %r1038, %r204, 2139062143;
sub.s32 %r1039, %r1037, %r1038;
xor.b32 %r211, %r848, %r1039;
ld.global.nc.u32 %r212, [%rd66+24];

dp4a.s32.s32 %r206, %r207, %r208, %r748;


dp4a.s32.s32 %r210, %r211, %r212, %r206;

mov.u32 %r214, %r704;

prmt.b32 %r214, %r214, 0, 0xba98;

xor.b32 %r1040, %r214, %r850;
or.b32 %r1041, %r1040, -2139062144;
and.b32 %r1042, %r214, 2139062143;
sub.s32 %r1043, %r1041, %r1042;
xor.b32 %r219, %r868, %r1043;
ld.global.nc.u32 %r220, [%rd66+28];
mov.u32 %r216, %r706;

prmt.b32 %r216, %r216, 0, 0xba98;

xor.b32 %r1044, %r216, %r851;
or.b32 %r1045, %r1044, -2139062144;
and.b32 %r1046, %r216, 2139062143;
sub.s32 %r1047, %r1045, %r1046;
xor.b32 %r223, %r876, %r1047;
ld.global.nc.u32 %r224, [%rd66+32];

dp4a.s32.s32 %r218, %r219, %r220, %r210;


dp4a.s32.s32 %r222, %r223, %r224, %r218;

mul.lo.s32 %r1048, %r198, %r878;
mad.lo.s32 %r1049, %r222, %r879, %r1048;
add.s32 %r1050, %r222, %r198;
shr.u32 %r1051, %r1050, 31;
add.s32 %r1052, %r1050, %r1051;
shr.s32 %r1053, %r1052, 1;
add.s32 %r1054, %r1049, %r1053;
shr.s32 %r1055, %r1054, 31;
shr.u32 %r1056, %r1055, 30;
add.s32 %r1057, %r1054, %r1056;
shr.s32 %r1058, %r1057, 2;

{ cvt.f32.f16 %f89, %rs1;}


ld.global.nc.u32 %r226, [%rd66];

{.reg .f16 low,high;
mov.b32 {low,high},%r226;
cvt.f32.f16 %f90, low;}


mul.ftz.f32 %f117, %f89, %f90;
cvt.rn.f32.s32 %f118, %r1058;
fma.rn.ftz.f32 %f320, %f117, %f118, %f320;
mov.u32 %r227, %r717;

prmt.b32 %r227, %r227, 0, 0xba98;

xor.b32 %r1059, %r227, %r891;
or.b32 %r1060, %r1059, -2139062144;
and.b32 %r1061, %r227, 2139062143;
sub.s32 %r1062, %r1060, %r1061;
xor.b32 %r232, %r909, %r1062;
mov.u32 %r229, %r719;

prmt.b32 %r229, %r229, 0, 0xba98;

xor.b32 %r1063, %r229, %r892;
or.b32 %r1064, %r1063, -2139062144;
and.b32 %r1065, %r229, 2139062143;
sub.s32 %r1066, %r1064, %r1065;
xor.b32 %r236, %r917, %r1066;

dp4a.s32.s32 %r231, %r232, %r184, %r748;


dp4a.s32.s32 %r235, %r236, %r188, %r231;

mov.u32 %r239, %r729;

prmt.b32 %r239, %r239, 0, 0xba98;

xor.b32 %r1067, %r239, %r919;
or.b32 %r1068, %r1067, -2139062144;
and.b32 %r1069, %r239, 2139062143;
sub.s32 %r1070, %r1068, %r1069;
xor.b32 %r244, %r937, %r1070;
mov.u32 %r241, %r731;

prmt.b32 %r241, %r241, 0, 0xba98;

xor.b32 %r1071, %r241, %r920;
or.b32 %r1072, %r1071, -2139062144;
and.b32 %r1073, %r241, 2139062143;
sub.s32 %r1074, %r1072, %r1073;
xor.b32 %r248, %r945, %r1074;

dp4a.s32.s32 %r243, %r244, %r196, %r235;


dp4a.s32.s32 %r247, %r248, %r200, %r243;

mov.u32 %r251, %r741;

prmt.b32 %r251, %r251, 0, 0xba98;

xor.b32 %r1075, %r251, %r947;
or.b32 %r1076, %r1075, -2139062144;
and.b32 %r1077, %r251, 2139062143;
sub.s32 %r1078, %r1076, %r1077;
xor.b32 %r256, %r965, %r1078;
mov.u32 %r253, %r743;

prmt.b32 %r253, %r253, 0, 0xba98;

xor.b32 %r1079, %r253, %r948;
or.b32 %r1080, %r1079, -2139062144;
and.b32 %r1081, %r253, 2139062143;
sub.s32 %r1082, %r1080, %r1081;
xor.b32 %r260, %r973, %r1082;

dp4a.s32.s32 %r255, %r256, %r208, %r748;


dp4a.s32.s32 %r259, %r260, %r212, %r255;

mov.u32 %r263, %r753;

prmt.b32 %r263, %r263, 0, 0xba98;

xor.b32 %r1083, %r263, %r975;
or.b32 %r1084, %r1083, -2139062144;
and.b32 %r1085, %r263, 2139062143;
sub.s32 %r1086, %r1084, %r1085;
xor.b32 %r268, %r993, %r1086;
mov.u32 %r265, %r755;

prmt.b32 %r265, %r265, 0, 0xba98;

xor.b32 %r1087, %r265, %r976;
or.b32 %r1088, %r1087, -2139062144;
and.b32 %r1089, %r265, 2139062143;
sub.s32 %r1090, %r1088, %r1089;
xor.b32 %r272, %r1001, %r1090;

dp4a.s32.s32 %r267, %r268, %r220, %r259;


dp4a.s32.s32 %r271, %r272, %r224, %r267;

mul.lo.s32 %r1091, %r247, %r1003;
mad.lo.s32 %r1092, %r271, %r1004, %r1091;
add.s32 %r1093, %r271, %r247;
shr.u32 %r1094, %r1093, 31;
add.s32 %r1095, %r1093, %r1094;
shr.s32 %r1096, %r1095, 1;
add.s32 %r1097, %r1092, %r1096;
shr.s32 %r1098, %r1097, 31;
shr.u32 %r1099, %r1098, 30;
add.s32 %r1100, %r1097, %r1099;
shr.s32 %r1101, %r1100, 2;

{ cvt.f32.f16 %f91, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r226;
cvt.f32.f16 %f92, low;}


mul.ftz.f32 %f119, %f91, %f92;
cvt.rn.f32.s32 %f120, %r1101;
fma.rn.ftz.f32 %f319, %f119, %f120, %f319;
cvt.u64.u32 %rd67, %r1924;
add.s64 %rd68, %rd1, %rd67;
add.s64 %rd69, %rd68, %rd3;
mul.lo.s64 %rd70, %rd69, 36;
mov.u32 %r276, %r668;

prmt.b32 %r276, %r276, 0, 0xba98;

xor.b32 %r1102, %r276, %r766;
or.b32 %r1103, %r1102, -2139062144;
and.b32 %r1104, %r276, 2139062143;
sub.s32 %r1105, %r1103, %r1104;
xor.b32 %r281, %r784, %r1105;
add.s64 %rd71, %rd10, %rd70;
ld.global.nc.u32 %r282, [%rd71+4];
mov.u32 %r278, %r670;

prmt.b32 %r278, %r278, 0, 0xba98;

xor.b32 %r1106, %r278, %r767;
or.b32 %r1107, %r1106, -2139062144;
and.b32 %r1108, %r278, 2139062143;
sub.s32 %r1109, %r1107, %r1108;
xor.b32 %r285, %r792, %r1109;
ld.global.nc.u32 %r286, [%rd71+8];

dp4a.s32.s32 %r280, %r281, %r282, %r748;


dp4a.s32.s32 %r284, %r285, %r286, %r280;

mov.u32 %r288, %r680;

prmt.b32 %r288, %r288, 0, 0xba98;

xor.b32 %r1110, %r288, %r794;
or.b32 %r1111, %r1110, -2139062144;
and.b32 %r1112, %r288, 2139062143;
sub.s32 %r1113, %r1111, %r1112;
xor.b32 %r293, %r812, %r1113;
ld.global.nc.u32 %r294, [%rd71+12];
mov.u32 %r290, %r682;

prmt.b32 %r290, %r290, 0, 0xba98;

xor.b32 %r1114, %r290, %r795;
or.b32 %r1115, %r1114, -2139062144;
and.b32 %r1116, %r290, 2139062143;
sub.s32 %r1117, %r1115, %r1116;
xor.b32 %r297, %r820, %r1117;
ld.global.nc.u32 %r298, [%rd71+16];

dp4a.s32.s32 %r292, %r293, %r294, %r284;


dp4a.s32.s32 %r296, %r297, %r298, %r292;

mov.u32 %r300, %r692;

prmt.b32 %r300, %r300, 0, 0xba98;

xor.b32 %r1118, %r300, %r822;
or.b32 %r1119, %r1118, -2139062144;
and.b32 %r1120, %r300, 2139062143;
sub.s32 %r1121, %r1119, %r1120;
xor.b32 %r305, %r840, %r1121;
ld.global.nc.u32 %r306, [%rd71+20];
mov.u32 %r302, %r694;

prmt.b32 %r302, %r302, 0, 0xba98;

xor.b32 %r1122, %r302, %r823;
or.b32 %r1123, %r1122, -2139062144;
and.b32 %r1124, %r302, 2139062143;
sub.s32 %r1125, %r1123, %r1124;
xor.b32 %r309, %r848, %r1125;
ld.global.nc.u32 %r310, [%rd71+24];

dp4a.s32.s32 %r304, %r305, %r306, %r748;


dp4a.s32.s32 %r308, %r309, %r310, %r304;

mov.u32 %r312, %r704;

prmt.b32 %r312, %r312, 0, 0xba98;

xor.b32 %r1126, %r312, %r850;
or.b32 %r1127, %r1126, -2139062144;
and.b32 %r1128, %r312, 2139062143;
sub.s32 %r1129, %r1127, %r1128;
xor.b32 %r317, %r868, %r1129;
ld.global.nc.u32 %r318, [%rd71+28];
mov.u32 %r314, %r706;

prmt.b32 %r314, %r314, 0, 0xba98;

xor.b32 %r1130, %r314, %r851;
or.b32 %r1131, %r1130, -2139062144;
and.b32 %r1132, %r314, 2139062143;
sub.s32 %r1133, %r1131, %r1132;
xor.b32 %r321, %r876, %r1133;
ld.global.nc.u32 %r322, [%rd71+32];

dp4a.s32.s32 %r316, %r317, %r318, %r308;


dp4a.s32.s32 %r320, %r321, %r322, %r316;

mul.lo.s32 %r1134, %r296, %r878;
mad.lo.s32 %r1135, %r320, %r879, %r1134;
add.s32 %r1136, %r320, %r296;
shr.u32 %r1137, %r1136, 31;
add.s32 %r1138, %r1136, %r1137;
shr.s32 %r1139, %r1138, 1;
add.s32 %r1140, %r1135, %r1139;
shr.s32 %r1141, %r1140, 31;
shr.u32 %r1142, %r1141, 30;
add.s32 %r1143, %r1140, %r1142;
shr.s32 %r1144, %r1143, 2;

{ cvt.f32.f16 %f93, %rs1;}


ld.global.nc.u32 %r324, [%rd71];

{.reg .f16 low,high;
mov.b32 {low,high},%r324;
cvt.f32.f16 %f94, low;}


mul.ftz.f32 %f121, %f93, %f94;
cvt.rn.f32.s32 %f122, %r1144;
fma.rn.ftz.f32 %f318, %f121, %f122, %f318;
mov.u32 %r325, %r717;

prmt.b32 %r325, %r325, 0, 0xba98;

xor.b32 %r1145, %r325, %r891;
or.b32 %r1146, %r1145, -2139062144;
and.b32 %r1147, %r325, 2139062143;
sub.s32 %r1148, %r1146, %r1147;
xor.b32 %r330, %r909, %r1148;
mov.u32 %r327, %r719;

prmt.b32 %r327, %r327, 0, 0xba98;

xor.b32 %r1149, %r327, %r892;
or.b32 %r1150, %r1149, -2139062144;
and.b32 %r1151, %r327, 2139062143;
sub.s32 %r1152, %r1150, %r1151;
xor.b32 %r334, %r917, %r1152;

dp4a.s32.s32 %r329, %r330, %r282, %r748;


dp4a.s32.s32 %r333, %r334, %r286, %r329;

mov.u32 %r337, %r729;

prmt.b32 %r337, %r337, 0, 0xba98;

xor.b32 %r1153, %r337, %r919;
or.b32 %r1154, %r1153, -2139062144;
and.b32 %r1155, %r337, 2139062143;
sub.s32 %r1156, %r1154, %r1155;
xor.b32 %r342, %r937, %r1156;
mov.u32 %r339, %r731;

prmt.b32 %r339, %r339, 0, 0xba98;

xor.b32 %r1157, %r339, %r920;
or.b32 %r1158, %r1157, -2139062144;
and.b32 %r1159, %r339, 2139062143;
sub.s32 %r1160, %r1158, %r1159;
xor.b32 %r346, %r945, %r1160;

dp4a.s32.s32 %r341, %r342, %r294, %r333;


dp4a.s32.s32 %r345, %r346, %r298, %r341;

mov.u32 %r349, %r741;

prmt.b32 %r349, %r349, 0, 0xba98;

xor.b32 %r1161, %r349, %r947;
or.b32 %r1162, %r1161, -2139062144;
and.b32 %r1163, %r349, 2139062143;
sub.s32 %r1164, %r1162, %r1163;
xor.b32 %r354, %r965, %r1164;
mov.u32 %r351, %r743;

prmt.b32 %r351, %r351, 0, 0xba98;

xor.b32 %r1165, %r351, %r948;
or.b32 %r1166, %r1165, -2139062144;
and.b32 %r1167, %r351, 2139062143;
sub.s32 %r1168, %r1166, %r1167;
xor.b32 %r358, %r973, %r1168;

dp4a.s32.s32 %r353, %r354, %r306, %r748;


dp4a.s32.s32 %r357, %r358, %r310, %r353;

mov.u32 %r361, %r753;

prmt.b32 %r361, %r361, 0, 0xba98;

xor.b32 %r1169, %r361, %r975;
or.b32 %r1170, %r1169, -2139062144;
and.b32 %r1171, %r361, 2139062143;
sub.s32 %r1172, %r1170, %r1171;
xor.b32 %r366, %r993, %r1172;
mov.u32 %r363, %r755;

prmt.b32 %r363, %r363, 0, 0xba98;

xor.b32 %r1173, %r363, %r976;
or.b32 %r1174, %r1173, -2139062144;
and.b32 %r1175, %r363, 2139062143;
sub.s32 %r1176, %r1174, %r1175;
xor.b32 %r370, %r1001, %r1176;

dp4a.s32.s32 %r365, %r366, %r318, %r357;


dp4a.s32.s32 %r369, %r370, %r322, %r365;

mul.lo.s32 %r1177, %r345, %r1003;
mad.lo.s32 %r1178, %r369, %r1004, %r1177;
add.s32 %r1179, %r369, %r345;
shr.u32 %r1180, %r1179, 31;
add.s32 %r1181, %r1179, %r1180;
shr.s32 %r1182, %r1181, 1;
add.s32 %r1183, %r1178, %r1182;
shr.s32 %r1184, %r1183, 31;
shr.u32 %r1185, %r1184, 30;
add.s32 %r1186, %r1183, %r1185;
shr.s32 %r1187, %r1186, 2;

{ cvt.f32.f16 %f95, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r324;
cvt.f32.f16 %f96, low;}


mul.ftz.f32 %f123, %f95, %f96;
cvt.rn.f32.s32 %f124, %r1187;
fma.rn.ftz.f32 %f317, %f123, %f124, %f317;
cvt.u64.u32 %rd72, %r1923;
add.s64 %rd73, %rd1, %rd72;
add.s64 %rd74, %rd73, %rd3;
mul.lo.s64 %rd75, %rd74, 36;
mov.u32 %r374, %r668;

prmt.b32 %r374, %r374, 0, 0xba98;

xor.b32 %r1188, %r374, %r766;
or.b32 %r1189, %r1188, -2139062144;
and.b32 %r1190, %r374, 2139062143;
sub.s32 %r1191, %r1189, %r1190;
xor.b32 %r379, %r784, %r1191;
add.s64 %rd76, %rd10, %rd75;
ld.global.nc.u32 %r380, [%rd76+4];
mov.u32 %r376, %r670;

prmt.b32 %r376, %r376, 0, 0xba98;

xor.b32 %r1192, %r376, %r767;
or.b32 %r1193, %r1192, -2139062144;
and.b32 %r1194, %r376, 2139062143;
sub.s32 %r1195, %r1193, %r1194;
xor.b32 %r383, %r792, %r1195;
ld.global.nc.u32 %r384, [%rd76+8];

dp4a.s32.s32 %r378, %r379, %r380, %r748;


dp4a.s32.s32 %r382, %r383, %r384, %r378;

mov.u32 %r386, %r680;

prmt.b32 %r386, %r386, 0, 0xba98;

xor.b32 %r1196, %r386, %r794;
or.b32 %r1197, %r1196, -2139062144;
and.b32 %r1198, %r386, 2139062143;
sub.s32 %r1199, %r1197, %r1198;
xor.b32 %r391, %r812, %r1199;
ld.global.nc.u32 %r392, [%rd76+12];
mov.u32 %r388, %r682;

prmt.b32 %r388, %r388, 0, 0xba98;

xor.b32 %r1200, %r388, %r795;
or.b32 %r1201, %r1200, -2139062144;
and.b32 %r1202, %r388, 2139062143;
sub.s32 %r1203, %r1201, %r1202;
xor.b32 %r395, %r820, %r1203;
ld.global.nc.u32 %r396, [%rd76+16];

dp4a.s32.s32 %r390, %r391, %r392, %r382;


dp4a.s32.s32 %r394, %r395, %r396, %r390;

mov.u32 %r398, %r692;

prmt.b32 %r398, %r398, 0, 0xba98;

xor.b32 %r1204, %r398, %r822;
or.b32 %r1205, %r1204, -2139062144;
and.b32 %r1206, %r398, 2139062143;
sub.s32 %r1207, %r1205, %r1206;
xor.b32 %r403, %r840, %r1207;
ld.global.nc.u32 %r404, [%rd76+20];
mov.u32 %r400, %r694;

prmt.b32 %r400, %r400, 0, 0xba98;

xor.b32 %r1208, %r400, %r823;
or.b32 %r1209, %r1208, -2139062144;
and.b32 %r1210, %r400, 2139062143;
sub.s32 %r1211, %r1209, %r1210;
xor.b32 %r407, %r848, %r1211;
ld.global.nc.u32 %r408, [%rd76+24];

dp4a.s32.s32 %r402, %r403, %r404, %r748;


dp4a.s32.s32 %r406, %r407, %r408, %r402;

mov.u32 %r410, %r704;

prmt.b32 %r410, %r410, 0, 0xba98;

xor.b32 %r1212, %r410, %r850;
or.b32 %r1213, %r1212, -2139062144;
and.b32 %r1214, %r410, 2139062143;
sub.s32 %r1215, %r1213, %r1214;
xor.b32 %r415, %r868, %r1215;
ld.global.nc.u32 %r416, [%rd76+28];
mov.u32 %r412, %r706;

prmt.b32 %r412, %r412, 0, 0xba98;

xor.b32 %r1216, %r412, %r851;
or.b32 %r1217, %r1216, -2139062144;
and.b32 %r1218, %r412, 2139062143;
sub.s32 %r1219, %r1217, %r1218;
xor.b32 %r419, %r876, %r1219;
ld.global.nc.u32 %r420, [%rd76+32];

dp4a.s32.s32 %r414, %r415, %r416, %r406;


dp4a.s32.s32 %r418, %r419, %r420, %r414;

mul.lo.s32 %r1220, %r394, %r878;
mad.lo.s32 %r1221, %r418, %r879, %r1220;
add.s32 %r1222, %r418, %r394;
shr.u32 %r1223, %r1222, 31;
add.s32 %r1224, %r1222, %r1223;
shr.s32 %r1225, %r1224, 1;
add.s32 %r1226, %r1221, %r1225;
shr.s32 %r1227, %r1226, 31;
shr.u32 %r1228, %r1227, 30;
add.s32 %r1229, %r1226, %r1228;
shr.s32 %r1230, %r1229, 2;

{ cvt.f32.f16 %f97, %rs1;}


ld.global.nc.u32 %r422, [%rd76];

{.reg .f16 low,high;
mov.b32 {low,high},%r422;
cvt.f32.f16 %f98, low;}


mul.ftz.f32 %f125, %f97, %f98;
cvt.rn.f32.s32 %f126, %r1230;
fma.rn.ftz.f32 %f316, %f125, %f126, %f316;
mov.u32 %r423, %r717;

prmt.b32 %r423, %r423, 0, 0xba98;

xor.b32 %r1231, %r423, %r891;
or.b32 %r1232, %r1231, -2139062144;
and.b32 %r1233, %r423, 2139062143;
sub.s32 %r1234, %r1232, %r1233;
xor.b32 %r428, %r909, %r1234;
mov.u32 %r425, %r719;

prmt.b32 %r425, %r425, 0, 0xba98;

xor.b32 %r1235, %r425, %r892;
or.b32 %r1236, %r1235, -2139062144;
and.b32 %r1237, %r425, 2139062143;
sub.s32 %r1238, %r1236, %r1237;
xor.b32 %r432, %r917, %r1238;

dp4a.s32.s32 %r427, %r428, %r380, %r748;


dp4a.s32.s32 %r431, %r432, %r384, %r427;

mov.u32 %r435, %r729;

prmt.b32 %r435, %r435, 0, 0xba98;

xor.b32 %r1239, %r435, %r919;
or.b32 %r1240, %r1239, -2139062144;
and.b32 %r1241, %r435, 2139062143;
sub.s32 %r1242, %r1240, %r1241;
xor.b32 %r440, %r937, %r1242;
mov.u32 %r437, %r731;

prmt.b32 %r437, %r437, 0, 0xba98;

xor.b32 %r1243, %r437, %r920;
or.b32 %r1244, %r1243, -2139062144;
and.b32 %r1245, %r437, 2139062143;
sub.s32 %r1246, %r1244, %r1245;
xor.b32 %r444, %r945, %r1246;

dp4a.s32.s32 %r439, %r440, %r392, %r431;


dp4a.s32.s32 %r443, %r444, %r396, %r439;

mov.u32 %r447, %r741;

prmt.b32 %r447, %r447, 0, 0xba98;

xor.b32 %r1247, %r447, %r947;
or.b32 %r1248, %r1247, -2139062144;
and.b32 %r1249, %r447, 2139062143;
sub.s32 %r1250, %r1248, %r1249;
xor.b32 %r452, %r965, %r1250;
mov.u32 %r449, %r743;

prmt.b32 %r449, %r449, 0, 0xba98;

xor.b32 %r1251, %r449, %r948;
or.b32 %r1252, %r1251, -2139062144;
and.b32 %r1253, %r449, 2139062143;
sub.s32 %r1254, %r1252, %r1253;
xor.b32 %r456, %r973, %r1254;

dp4a.s32.s32 %r451, %r452, %r404, %r748;


dp4a.s32.s32 %r455, %r456, %r408, %r451;

mov.u32 %r459, %r753;

prmt.b32 %r459, %r459, 0, 0xba98;

xor.b32 %r1255, %r459, %r975;
or.b32 %r1256, %r1255, -2139062144;
and.b32 %r1257, %r459, 2139062143;
sub.s32 %r1258, %r1256, %r1257;
xor.b32 %r464, %r993, %r1258;
mov.u32 %r461, %r755;

prmt.b32 %r461, %r461, 0, 0xba98;

xor.b32 %r1259, %r461, %r976;
or.b32 %r1260, %r1259, -2139062144;
and.b32 %r1261, %r461, 2139062143;
sub.s32 %r1262, %r1260, %r1261;
xor.b32 %r468, %r1001, %r1262;

dp4a.s32.s32 %r463, %r464, %r416, %r455;


dp4a.s32.s32 %r467, %r468, %r420, %r463;

mul.lo.s32 %r1263, %r443, %r1003;
mad.lo.s32 %r1264, %r467, %r1004, %r1263;
add.s32 %r1265, %r467, %r443;
shr.u32 %r1266, %r1265, 31;
add.s32 %r1267, %r1265, %r1266;
shr.s32 %r1268, %r1267, 1;
add.s32 %r1269, %r1264, %r1268;
shr.s32 %r1270, %r1269, 31;
shr.u32 %r1271, %r1270, 30;
add.s32 %r1272, %r1269, %r1271;
shr.s32 %r1273, %r1272, 2;

{ cvt.f32.f16 %f99, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r422;
cvt.f32.f16 %f100, low;}


mul.ftz.f32 %f127, %f99, %f100;
cvt.rn.f32.s32 %f128, %r1273;
fma.rn.ftz.f32 %f315, %f127, %f128, %f315;
cvt.u64.u32 %rd77, %r1922;
add.s64 %rd78, %rd1, %rd77;
add.s64 %rd79, %rd78, %rd3;
mul.lo.s64 %rd80, %rd79, 36;
mov.u32 %r472, %r668;

prmt.b32 %r472, %r472, 0, 0xba98;

xor.b32 %r1274, %r472, %r766;
or.b32 %r1275, %r1274, -2139062144;
and.b32 %r1276, %r472, 2139062143;
sub.s32 %r1277, %r1275, %r1276;
xor.b32 %r477, %r784, %r1277;
add.s64 %rd81, %rd10, %rd80;
ld.global.nc.u32 %r478, [%rd81+4];
mov.u32 %r474, %r670;

prmt.b32 %r474, %r474, 0, 0xba98;

xor.b32 %r1278, %r474, %r767;
or.b32 %r1279, %r1278, -2139062144;
and.b32 %r1280, %r474, 2139062143;
sub.s32 %r1281, %r1279, %r1280;
xor.b32 %r481, %r792, %r1281;
ld.global.nc.u32 %r482, [%rd81+8];

dp4a.s32.s32 %r476, %r477, %r478, %r748;


dp4a.s32.s32 %r480, %r481, %r482, %r476;

mov.u32 %r484, %r680;

prmt.b32 %r484, %r484, 0, 0xba98;

xor.b32 %r1282, %r484, %r794;
or.b32 %r1283, %r1282, -2139062144;
and.b32 %r1284, %r484, 2139062143;
sub.s32 %r1285, %r1283, %r1284;
xor.b32 %r489, %r812, %r1285;
ld.global.nc.u32 %r490, [%rd81+12];
mov.u32 %r486, %r682;

prmt.b32 %r486, %r486, 0, 0xba98;

xor.b32 %r1286, %r486, %r795;
or.b32 %r1287, %r1286, -2139062144;
and.b32 %r1288, %r486, 2139062143;
sub.s32 %r1289, %r1287, %r1288;
xor.b32 %r493, %r820, %r1289;
ld.global.nc.u32 %r494, [%rd81+16];

dp4a.s32.s32 %r488, %r489, %r490, %r480;


dp4a.s32.s32 %r492, %r493, %r494, %r488;

mov.u32 %r496, %r692;

prmt.b32 %r496, %r496, 0, 0xba98;

xor.b32 %r1290, %r496, %r822;
or.b32 %r1291, %r1290, -2139062144;
and.b32 %r1292, %r496, 2139062143;
sub.s32 %r1293, %r1291, %r1292;
xor.b32 %r501, %r840, %r1293;
ld.global.nc.u32 %r502, [%rd81+20];
mov.u32 %r498, %r694;

prmt.b32 %r498, %r498, 0, 0xba98;

xor.b32 %r1294, %r498, %r823;
or.b32 %r1295, %r1294, -2139062144;
and.b32 %r1296, %r498, 2139062143;
sub.s32 %r1297, %r1295, %r1296;
xor.b32 %r505, %r848, %r1297;
ld.global.nc.u32 %r506, [%rd81+24];

dp4a.s32.s32 %r500, %r501, %r502, %r748;


dp4a.s32.s32 %r504, %r505, %r506, %r500;

mov.u32 %r508, %r704;

prmt.b32 %r508, %r508, 0, 0xba98;

xor.b32 %r1298, %r508, %r850;
or.b32 %r1299, %r1298, -2139062144;
and.b32 %r1300, %r508, 2139062143;
sub.s32 %r1301, %r1299, %r1300;
xor.b32 %r513, %r868, %r1301;
ld.global.nc.u32 %r514, [%rd81+28];
mov.u32 %r510, %r706;

prmt.b32 %r510, %r510, 0, 0xba98;

xor.b32 %r1302, %r510, %r851;
or.b32 %r1303, %r1302, -2139062144;
and.b32 %r1304, %r510, 2139062143;
sub.s32 %r1305, %r1303, %r1304;
xor.b32 %r517, %r876, %r1305;
ld.global.nc.u32 %r518, [%rd81+32];

dp4a.s32.s32 %r512, %r513, %r514, %r504;


dp4a.s32.s32 %r516, %r517, %r518, %r512;

mul.lo.s32 %r1306, %r492, %r878;
mad.lo.s32 %r1307, %r516, %r879, %r1306;
add.s32 %r1308, %r516, %r492;
shr.u32 %r1309, %r1308, 31;
add.s32 %r1310, %r1308, %r1309;
shr.s32 %r1311, %r1310, 1;
add.s32 %r1312, %r1307, %r1311;
shr.s32 %r1313, %r1312, 31;
shr.u32 %r1314, %r1313, 30;
add.s32 %r1315, %r1312, %r1314;
shr.s32 %r1316, %r1315, 2;

{ cvt.f32.f16 %f101, %rs1;}


ld.global.nc.u32 %r520, [%rd81];

{.reg .f16 low,high;
mov.b32 {low,high},%r520;
cvt.f32.f16 %f102, low;}


mul.ftz.f32 %f129, %f101, %f102;
cvt.rn.f32.s32 %f130, %r1316;
fma.rn.ftz.f32 %f314, %f129, %f130, %f314;
mov.u32 %r521, %r717;

prmt.b32 %r521, %r521, 0, 0xba98;

xor.b32 %r1317, %r521, %r891;
or.b32 %r1318, %r1317, -2139062144;
and.b32 %r1319, %r521, 2139062143;
sub.s32 %r1320, %r1318, %r1319;
xor.b32 %r526, %r909, %r1320;
mov.u32 %r523, %r719;

prmt.b32 %r523, %r523, 0, 0xba98;

xor.b32 %r1321, %r523, %r892;
or.b32 %r1322, %r1321, -2139062144;
and.b32 %r1323, %r523, 2139062143;
sub.s32 %r1324, %r1322, %r1323;
xor.b32 %r530, %r917, %r1324;

dp4a.s32.s32 %r525, %r526, %r478, %r748;


dp4a.s32.s32 %r529, %r530, %r482, %r525;

mov.u32 %r533, %r729;

prmt.b32 %r533, %r533, 0, 0xba98;

xor.b32 %r1325, %r533, %r919;
or.b32 %r1326, %r1325, -2139062144;
and.b32 %r1327, %r533, 2139062143;
sub.s32 %r1328, %r1326, %r1327;
xor.b32 %r538, %r937, %r1328;
mov.u32 %r535, %r731;

prmt.b32 %r535, %r535, 0, 0xba98;

xor.b32 %r1329, %r535, %r920;
or.b32 %r1330, %r1329, -2139062144;
and.b32 %r1331, %r535, 2139062143;
sub.s32 %r1332, %r1330, %r1331;
xor.b32 %r542, %r945, %r1332;

dp4a.s32.s32 %r537, %r538, %r490, %r529;


dp4a.s32.s32 %r541, %r542, %r494, %r537;

mov.u32 %r545, %r741;

prmt.b32 %r545, %r545, 0, 0xba98;

xor.b32 %r1333, %r545, %r947;
or.b32 %r1334, %r1333, -2139062144;
and.b32 %r1335, %r545, 2139062143;
sub.s32 %r1336, %r1334, %r1335;
xor.b32 %r550, %r965, %r1336;
mov.u32 %r547, %r743;

prmt.b32 %r547, %r547, 0, 0xba98;

xor.b32 %r1337, %r547, %r948;
or.b32 %r1338, %r1337, -2139062144;
and.b32 %r1339, %r547, 2139062143;
sub.s32 %r1340, %r1338, %r1339;
xor.b32 %r554, %r973, %r1340;

dp4a.s32.s32 %r549, %r550, %r502, %r748;


dp4a.s32.s32 %r553, %r554, %r506, %r549;

mov.u32 %r557, %r753;

prmt.b32 %r557, %r557, 0, 0xba98;

xor.b32 %r1341, %r557, %r975;
or.b32 %r1342, %r1341, -2139062144;
and.b32 %r1343, %r557, 2139062143;
sub.s32 %r1344, %r1342, %r1343;
xor.b32 %r562, %r993, %r1344;
mov.u32 %r559, %r755;

prmt.b32 %r559, %r559, 0, 0xba98;

xor.b32 %r1345, %r559, %r976;
or.b32 %r1346, %r1345, -2139062144;
and.b32 %r1347, %r559, 2139062143;
sub.s32 %r1348, %r1346, %r1347;
xor.b32 %r566, %r1001, %r1348;

dp4a.s32.s32 %r561, %r562, %r514, %r553;


dp4a.s32.s32 %r565, %r566, %r518, %r561;

mul.lo.s32 %r1349, %r541, %r1003;
mad.lo.s32 %r1350, %r565, %r1004, %r1349;
add.s32 %r1351, %r565, %r541;
shr.u32 %r1352, %r1351, 31;
add.s32 %r1353, %r1351, %r1352;
shr.s32 %r1354, %r1353, 1;
add.s32 %r1355, %r1350, %r1354;
shr.s32 %r1356, %r1355, 31;
shr.u32 %r1357, %r1356, 30;
add.s32 %r1358, %r1355, %r1357;
shr.s32 %r1359, %r1358, 2;

{ cvt.f32.f16 %f103, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r520;
cvt.f32.f16 %f104, low;}


mul.ftz.f32 %f131, %f103, %f104;
cvt.rn.f32.s32 %f132, %r1359;
fma.rn.ftz.f32 %f313, %f131, %f132, %f313;
cvt.u64.u32 %rd82, %r1921;
add.s64 %rd83, %rd1, %rd82;
add.s64 %rd84, %rd83, %rd3;
mul.lo.s64 %rd85, %rd84, 36;
mov.u32 %r570, %r668;

prmt.b32 %r570, %r570, 0, 0xba98;

xor.b32 %r1360, %r570, %r766;
or.b32 %r1361, %r1360, -2139062144;
and.b32 %r1362, %r570, 2139062143;
sub.s32 %r1363, %r1361, %r1362;
xor.b32 %r575, %r784, %r1363;
add.s64 %rd86, %rd10, %rd85;
ld.global.nc.u32 %r576, [%rd86+4];
mov.u32 %r572, %r670;

prmt.b32 %r572, %r572, 0, 0xba98;

xor.b32 %r1364, %r572, %r767;
or.b32 %r1365, %r1364, -2139062144;
and.b32 %r1366, %r572, 2139062143;
sub.s32 %r1367, %r1365, %r1366;
xor.b32 %r579, %r792, %r1367;
ld.global.nc.u32 %r580, [%rd86+8];

dp4a.s32.s32 %r574, %r575, %r576, %r748;


dp4a.s32.s32 %r578, %r579, %r580, %r574;

mov.u32 %r582, %r680;

prmt.b32 %r582, %r582, 0, 0xba98;

xor.b32 %r1368, %r582, %r794;
or.b32 %r1369, %r1368, -2139062144;
and.b32 %r1370, %r582, 2139062143;
sub.s32 %r1371, %r1369, %r1370;
xor.b32 %r587, %r812, %r1371;
ld.global.nc.u32 %r588, [%rd86+12];
mov.u32 %r584, %r682;

prmt.b32 %r584, %r584, 0, 0xba98;

xor.b32 %r1372, %r584, %r795;
or.b32 %r1373, %r1372, -2139062144;
and.b32 %r1374, %r584, 2139062143;
sub.s32 %r1375, %r1373, %r1374;
xor.b32 %r591, %r820, %r1375;
ld.global.nc.u32 %r592, [%rd86+16];

dp4a.s32.s32 %r586, %r587, %r588, %r578;


dp4a.s32.s32 %r590, %r591, %r592, %r586;

mov.u32 %r594, %r692;

prmt.b32 %r594, %r594, 0, 0xba98;

xor.b32 %r1376, %r594, %r822;
or.b32 %r1377, %r1376, -2139062144;
and.b32 %r1378, %r594, 2139062143;
sub.s32 %r1379, %r1377, %r1378;
xor.b32 %r599, %r840, %r1379;
ld.global.nc.u32 %r600, [%rd86+20];
mov.u32 %r596, %r694;

prmt.b32 %r596, %r596, 0, 0xba98;

xor.b32 %r1380, %r596, %r823;
or.b32 %r1381, %r1380, -2139062144;
and.b32 %r1382, %r596, 2139062143;
sub.s32 %r1383, %r1381, %r1382;
xor.b32 %r603, %r848, %r1383;
ld.global.nc.u32 %r604, [%rd86+24];

dp4a.s32.s32 %r598, %r599, %r600, %r748;


dp4a.s32.s32 %r602, %r603, %r604, %r598;

mov.u32 %r606, %r704;

prmt.b32 %r606, %r606, 0, 0xba98;

xor.b32 %r1384, %r606, %r850;
or.b32 %r1385, %r1384, -2139062144;
and.b32 %r1386, %r606, 2139062143;
sub.s32 %r1387, %r1385, %r1386;
xor.b32 %r611, %r868, %r1387;
ld.global.nc.u32 %r612, [%rd86+28];
mov.u32 %r608, %r706;

prmt.b32 %r608, %r608, 0, 0xba98;

xor.b32 %r1388, %r608, %r851;
or.b32 %r1389, %r1388, -2139062144;
and.b32 %r1390, %r608, 2139062143;
sub.s32 %r1391, %r1389, %r1390;
xor.b32 %r615, %r876, %r1391;
ld.global.nc.u32 %r616, [%rd86+32];

dp4a.s32.s32 %r610, %r611, %r612, %r602;


dp4a.s32.s32 %r614, %r615, %r616, %r610;

mul.lo.s32 %r1392, %r590, %r878;
mad.lo.s32 %r1393, %r614, %r879, %r1392;
add.s32 %r1394, %r614, %r590;
shr.u32 %r1395, %r1394, 31;
add.s32 %r1396, %r1394, %r1395;
shr.s32 %r1397, %r1396, 1;
add.s32 %r1398, %r1393, %r1397;
shr.s32 %r1399, %r1398, 31;
shr.u32 %r1400, %r1399, 30;
add.s32 %r1401, %r1398, %r1400;
shr.s32 %r1402, %r1401, 2;

{ cvt.f32.f16 %f105, %rs1;}


ld.global.nc.u32 %r618, [%rd86];

{.reg .f16 low,high;
mov.b32 {low,high},%r618;
cvt.f32.f16 %f106, low;}


mul.ftz.f32 %f133, %f105, %f106;
cvt.rn.f32.s32 %f134, %r1402;
fma.rn.ftz.f32 %f312, %f133, %f134, %f312;
mov.u32 %r619, %r717;

prmt.b32 %r619, %r619, 0, 0xba98;

xor.b32 %r1403, %r619, %r891;
or.b32 %r1404, %r1403, -2139062144;
and.b32 %r1405, %r619, 2139062143;
sub.s32 %r1406, %r1404, %r1405;
xor.b32 %r624, %r909, %r1406;
mov.u32 %r621, %r719;

prmt.b32 %r621, %r621, 0, 0xba98;

xor.b32 %r1407, %r621, %r892;
or.b32 %r1408, %r1407, -2139062144;
and.b32 %r1409, %r621, 2139062143;
sub.s32 %r1410, %r1408, %r1409;
xor.b32 %r628, %r917, %r1410;

dp4a.s32.s32 %r623, %r624, %r576, %r748;


dp4a.s32.s32 %r627, %r628, %r580, %r623;

mov.u32 %r631, %r729;

prmt.b32 %r631, %r631, 0, 0xba98;

xor.b32 %r1411, %r631, %r919;
or.b32 %r1412, %r1411, -2139062144;
and.b32 %r1413, %r631, 2139062143;
sub.s32 %r1414, %r1412, %r1413;
xor.b32 %r636, %r937, %r1414;
mov.u32 %r633, %r731;

prmt.b32 %r633, %r633, 0, 0xba98;

xor.b32 %r1415, %r633, %r920;
or.b32 %r1416, %r1415, -2139062144;
and.b32 %r1417, %r633, 2139062143;
sub.s32 %r1418, %r1416, %r1417;
xor.b32 %r640, %r945, %r1418;

dp4a.s32.s32 %r635, %r636, %r588, %r627;


dp4a.s32.s32 %r639, %r640, %r592, %r635;

mov.u32 %r643, %r741;

prmt.b32 %r643, %r643, 0, 0xba98;

xor.b32 %r1419, %r643, %r947;
or.b32 %r1420, %r1419, -2139062144;
and.b32 %r1421, %r643, 2139062143;
sub.s32 %r1422, %r1420, %r1421;
xor.b32 %r648, %r965, %r1422;
mov.u32 %r645, %r743;

prmt.b32 %r645, %r645, 0, 0xba98;

xor.b32 %r1423, %r645, %r948;
or.b32 %r1424, %r1423, -2139062144;
and.b32 %r1425, %r645, 2139062143;
sub.s32 %r1426, %r1424, %r1425;
xor.b32 %r652, %r973, %r1426;

dp4a.s32.s32 %r647, %r648, %r600, %r748;


dp4a.s32.s32 %r651, %r652, %r604, %r647;

mov.u32 %r655, %r753;

prmt.b32 %r655, %r655, 0, 0xba98;

xor.b32 %r1427, %r655, %r975;
or.b32 %r1428, %r1427, -2139062144;
and.b32 %r1429, %r655, 2139062143;
sub.s32 %r1430, %r1428, %r1429;
xor.b32 %r660, %r993, %r1430;
mov.u32 %r657, %r755;

prmt.b32 %r657, %r657, 0, 0xba98;

xor.b32 %r1431, %r657, %r976;
or.b32 %r1432, %r1431, -2139062144;
and.b32 %r1433, %r657, 2139062143;
sub.s32 %r1434, %r1432, %r1433;
xor.b32 %r664, %r1001, %r1434;

dp4a.s32.s32 %r659, %r660, %r612, %r651;


dp4a.s32.s32 %r663, %r664, %r616, %r659;

mul.lo.s32 %r1435, %r639, %r1003;
mad.lo.s32 %r1436, %r663, %r1004, %r1435;
add.s32 %r1437, %r663, %r639;
shr.u32 %r1438, %r1437, 31;
add.s32 %r1439, %r1437, %r1438;
shr.s32 %r1440, %r1439, 1;
add.s32 %r1441, %r1436, %r1440;
shr.s32 %r1442, %r1441, 31;
shr.u32 %r1443, %r1442, 30;
add.s32 %r1444, %r1441, %r1443;
shr.s32 %r1445, %r1444, 2;

{ cvt.f32.f16 %f107, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r618;
cvt.f32.f16 %f108, low;}


mul.ftz.f32 %f135, %f107, %f108;
cvt.rn.f32.s32 %f136, %r1445;
fma.rn.ftz.f32 %f311, %f135, %f136, %f311;
cvt.u64.u32 %rd87, %r1920;
add.s64 %rd88, %rd1, %rd87;
add.s64 %rd89, %rd88, %rd3;
mul.lo.s64 %rd90, %rd89, 36;

prmt.b32 %r668, %r668, 0, 0xba98;

xor.b32 %r1446, %r668, %r766;
or.b32 %r1447, %r1446, -2139062144;
and.b32 %r1448, %r668, 2139062143;
sub.s32 %r1449, %r1447, %r1448;
xor.b32 %r673, %r784, %r1449;
add.s64 %rd91, %rd10, %rd90;
ld.global.nc.u32 %r674, [%rd91+4];

prmt.b32 %r670, %r670, 0, 0xba98;

xor.b32 %r1450, %r670, %r767;
or.b32 %r1451, %r1450, -2139062144;
and.b32 %r1452, %r670, 2139062143;
sub.s32 %r1453, %r1451, %r1452;
xor.b32 %r677, %r792, %r1453;
ld.global.nc.u32 %r678, [%rd91+8];

dp4a.s32.s32 %r672, %r673, %r674, %r748;


dp4a.s32.s32 %r676, %r677, %r678, %r672;


prmt.b32 %r680, %r680, 0, 0xba98;

xor.b32 %r1454, %r680, %r794;
or.b32 %r1455, %r1454, -2139062144;
and.b32 %r1456, %r680, 2139062143;
sub.s32 %r1457, %r1455, %r1456;
xor.b32 %r685, %r812, %r1457;
ld.global.nc.u32 %r686, [%rd91+12];

prmt.b32 %r682, %r682, 0, 0xba98;

xor.b32 %r1458, %r682, %r795;
or.b32 %r1459, %r1458, -2139062144;
and.b32 %r1460, %r682, 2139062143;
sub.s32 %r1461, %r1459, %r1460;
xor.b32 %r689, %r820, %r1461;
ld.global.nc.u32 %r690, [%rd91+16];

dp4a.s32.s32 %r684, %r685, %r686, %r676;


dp4a.s32.s32 %r688, %r689, %r690, %r684;


prmt.b32 %r692, %r692, 0, 0xba98;

xor.b32 %r1462, %r692, %r822;
or.b32 %r1463, %r1462, -2139062144;
and.b32 %r1464, %r692, 2139062143;
sub.s32 %r1465, %r1463, %r1464;
xor.b32 %r697, %r840, %r1465;
ld.global.nc.u32 %r698, [%rd91+20];

prmt.b32 %r694, %r694, 0, 0xba98;

xor.b32 %r1466, %r694, %r823;
or.b32 %r1467, %r1466, -2139062144;
and.b32 %r1468, %r694, 2139062143;
sub.s32 %r1469, %r1467, %r1468;
xor.b32 %r701, %r848, %r1469;
ld.global.nc.u32 %r702, [%rd91+24];

dp4a.s32.s32 %r696, %r697, %r698, %r748;


dp4a.s32.s32 %r700, %r701, %r702, %r696;


prmt.b32 %r704, %r704, 0, 0xba98;

xor.b32 %r1470, %r704, %r850;
or.b32 %r1471, %r1470, -2139062144;
and.b32 %r1472, %r704, 2139062143;
sub.s32 %r1473, %r1471, %r1472;
xor.b32 %r709, %r868, %r1473;
ld.global.nc.u32 %r710, [%rd91+28];

prmt.b32 %r706, %r706, 0, 0xba98;

xor.b32 %r1474, %r706, %r851;
or.b32 %r1475, %r1474, -2139062144;
and.b32 %r1476, %r706, 2139062143;
sub.s32 %r1477, %r1475, %r1476;
xor.b32 %r713, %r876, %r1477;
ld.global.nc.u32 %r714, [%rd91+32];

dp4a.s32.s32 %r708, %r709, %r710, %r700;


dp4a.s32.s32 %r712, %r713, %r714, %r708;

mul.lo.s32 %r1478, %r688, %r878;
mad.lo.s32 %r1479, %r712, %r879, %r1478;
add.s32 %r1480, %r712, %r688;
shr.u32 %r1481, %r1480, 31;
add.s32 %r1482, %r1480, %r1481;
shr.s32 %r1483, %r1482, 1;
add.s32 %r1484, %r1479, %r1483;
shr.s32 %r1485, %r1484, 31;
shr.u32 %r1486, %r1485, 30;
add.s32 %r1487, %r1484, %r1486;
shr.s32 %r1488, %r1487, 2;

{ cvt.f32.f16 %f109, %rs1;}


ld.global.nc.u32 %r716, [%rd91];

{.reg .f16 low,high;
mov.b32 {low,high},%r716;
cvt.f32.f16 %f110, low;}


mul.ftz.f32 %f137, %f109, %f110;
cvt.rn.f32.s32 %f138, %r1488;
fma.rn.ftz.f32 %f310, %f137, %f138, %f310;

prmt.b32 %r717, %r717, 0, 0xba98;

xor.b32 %r1489, %r717, %r891;
or.b32 %r1490, %r1489, -2139062144;
and.b32 %r1491, %r717, 2139062143;
sub.s32 %r1492, %r1490, %r1491;
xor.b32 %r722, %r909, %r1492;

prmt.b32 %r719, %r719, 0, 0xba98;

xor.b32 %r1493, %r719, %r892;
or.b32 %r1494, %r1493, -2139062144;
and.b32 %r1495, %r719, 2139062143;
sub.s32 %r1496, %r1494, %r1495;
xor.b32 %r726, %r917, %r1496;

dp4a.s32.s32 %r721, %r722, %r674, %r748;


dp4a.s32.s32 %r725, %r726, %r678, %r721;


prmt.b32 %r729, %r729, 0, 0xba98;

xor.b32 %r1497, %r729, %r919;
or.b32 %r1498, %r1497, -2139062144;
and.b32 %r1499, %r729, 2139062143;
sub.s32 %r1500, %r1498, %r1499;
xor.b32 %r734, %r937, %r1500;

prmt.b32 %r731, %r731, 0, 0xba98;

xor.b32 %r1501, %r731, %r920;
or.b32 %r1502, %r1501, -2139062144;
and.b32 %r1503, %r731, 2139062143;
sub.s32 %r1504, %r1502, %r1503;
xor.b32 %r738, %r945, %r1504;

dp4a.s32.s32 %r733, %r734, %r686, %r725;


dp4a.s32.s32 %r737, %r738, %r690, %r733;


prmt.b32 %r741, %r741, 0, 0xba98;

xor.b32 %r1505, %r741, %r947;
or.b32 %r1506, %r1505, -2139062144;
and.b32 %r1507, %r741, 2139062143;
sub.s32 %r1508, %r1506, %r1507;
xor.b32 %r746, %r965, %r1508;

prmt.b32 %r743, %r743, 0, 0xba98;

xor.b32 %r1509, %r743, %r948;
or.b32 %r1510, %r1509, -2139062144;
and.b32 %r1511, %r743, 2139062143;
sub.s32 %r1512, %r1510, %r1511;
xor.b32 %r750, %r973, %r1512;

dp4a.s32.s32 %r745, %r746, %r698, %r748;


dp4a.s32.s32 %r749, %r750, %r702, %r745;


prmt.b32 %r753, %r753, 0, 0xba98;

xor.b32 %r1513, %r753, %r975;
or.b32 %r1514, %r1513, -2139062144;
and.b32 %r1515, %r753, 2139062143;
sub.s32 %r1516, %r1514, %r1515;
xor.b32 %r758, %r993, %r1516;

prmt.b32 %r755, %r755, 0, 0xba98;

xor.b32 %r1517, %r755, %r976;
or.b32 %r1518, %r1517, -2139062144;
and.b32 %r1519, %r755, 2139062143;
sub.s32 %r1520, %r1518, %r1519;
xor.b32 %r762, %r1001, %r1520;

dp4a.s32.s32 %r757, %r758, %r710, %r749;


dp4a.s32.s32 %r761, %r762, %r714, %r757;

mul.lo.s32 %r1521, %r737, %r1003;
mad.lo.s32 %r1522, %r761, %r1004, %r1521;
add.s32 %r1523, %r761, %r737;
shr.u32 %r1524, %r1523, 31;
add.s32 %r1525, %r1523, %r1524;
shr.s32 %r1526, %r1525, 1;
add.s32 %r1527, %r1522, %r1526;
shr.s32 %r1528, %r1527, 31;
shr.u32 %r1529, %r1528, 30;
add.s32 %r1530, %r1527, %r1529;
shr.s32 %r1531, %r1530, 2;

{ cvt.f32.f16 %f111, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r716;
cvt.f32.f16 %f112, low;}


mul.ftz.f32 %f139, %f111, %f112;
cvt.rn.f32.s32 %f140, %r1531;
fma.rn.ftz.f32 %f309, %f139, %f140, %f309;
add.s32 %r1924, %r1924, 64;
add.s32 %r1923, %r1923, 64;
add.s32 %r1922, %r1922, 64;
add.s32 %r1921, %r1921, 64;
add.s32 %r1920, %r1920, 64;
add.s32 %r1919, %r1919, 64;
add.s32 %r1918, %r1918, 64;
add.s64 %rd149, %rd149, 592;
add.s32 %r1925, %r1925, 8;
setp.lt.s32 %p3, %r1925, %r50;
@%p3 bra $L__BB222_2;

$L__BB222_3:
setp.eq.s32 %p4, %r44, 0;
@%p4 bra $L__BB222_5;

mov.u32 %r1536, _ZZ13mul_mat_vec_qIL9ggml_type17ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r1537, %r44, 1792, %r1536;
shl.b32 %r1538, %r46, 2;
add.s32 %r1539, %r1537, %r1538;
st.shared.f32 [%r1539+-1792], %f322;
st.shared.f32 [%r1539+-1664], %f321;
st.shared.f32 [%r1539+-1536], %f320;
st.shared.f32 [%r1539+-1408], %f319;
st.shared.f32 [%r1539+-1280], %f318;
st.shared.f32 [%r1539+-1152], %f317;
st.shared.f32 [%r1539+-1024], %f316;
st.shared.f32 [%r1539+-896], %f315;
st.shared.f32 [%r1539+-768], %f314;
st.shared.f32 [%r1539+-640], %f313;
st.shared.f32 [%r1539+-512], %f312;
st.shared.f32 [%r1539+-384], %f311;
st.shared.f32 [%r1539+-256], %f310;
st.shared.f32 [%r1539+-128], %f309;

$L__BB222_5:
setp.ne.s32 %p5, %r44, 0;
bar.sync 0;
@%p5 bra $L__BB222_34;

mov.u32 %r1541, %ctaid.x;
shl.b32 %r1542, %r1541, 1;
mov.u32 %r1543, 1;
shl.b32 %r1545, %r46, 2;
mov.u32 %r1546, _ZZ13mul_mat_vec_qIL9ggml_type17ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r1547, %r1546, %r1545;
mov.u32 %r1548, 2;
ld.shared.f32 %f141, [%r1547];
add.ftz.f32 %f142, %f141, %f322;
mov.b32 %r1549, %f142;
mov.u32 %r1550, 31;
mov.u32 %r1551, 16;
mov.u32 %r1552, -1;
shfl.sync.bfly.b32 %r1553|%p6, %r1549, %r1551, %r1550, %r1552;
mov.b32 %f143, %r1553;
add.ftz.f32 %f144, %f142, %f143;
mov.b32 %r1554, %f144;
mov.u32 %r1555, 8;
shfl.sync.bfly.b32 %r1556|%p7, %r1554, %r1555, %r1550, %r1552;
mov.b32 %f145, %r1556;
add.ftz.f32 %f146, %f144, %f145;
mov.b32 %r1557, %f146;
mov.u32 %r1558, 4;
shfl.sync.bfly.b32 %r1559|%p8, %r1557, %r1558, %r1550, %r1552;
mov.b32 %f147, %r1559;
add.ftz.f32 %f148, %f146, %f147;
mov.b32 %r1560, %f148;
shfl.sync.bfly.b32 %r1561|%p9, %r1560, %r1548, %r1550, %r1552;
mov.b32 %f149, %r1561;
add.ftz.f32 %f150, %f148, %f149;
mov.b32 %r1562, %f150;
shfl.sync.bfly.b32 %r1563|%p10, %r1562, %r1543, %r1550, %r1552;
mov.b32 %f151, %r1563;
add.ftz.f32 %f43, %f150, %f151;
setp.ne.s32 %p11, %r46, 0;
setp.ge.u32 %p12, %r1542, %r31;
or.pred %p13, %p11, %p12;
@%p13 bra $L__BB222_8;

cvta.to.global.u64 %rd92, %rd18;
mov.u32 %r1564, %ctaid.y;
mad.lo.s32 %r1567, %r1564, %r37, %r1542;
mov.u32 %r1568, %ctaid.z;
mad.lo.s32 %r1569, %r1568, %r43, %r1567;
mul.wide.u32 %rd93, %r1569, 4;
add.s64 %rd94, %rd92, %rd93;
st.global.f32 [%rd94], %f43;

$L__BB222_8:
ld.shared.f32 %f152, [%r1547+128];
add.ftz.f32 %f153, %f152, %f321;
mov.b32 %r1575, %f153;
shfl.sync.bfly.b32 %r1579|%p14, %r1575, %r1551, %r1550, %r1552;
mov.b32 %f154, %r1579;
add.ftz.f32 %f155, %f153, %f154;
mov.b32 %r1580, %f155;
shfl.sync.bfly.b32 %r1582|%p15, %r1580, %r1555, %r1550, %r1552;
mov.b32 %f156, %r1582;
add.ftz.f32 %f157, %f155, %f156;
mov.b32 %r1583, %f157;
shfl.sync.bfly.b32 %r1585|%p16, %r1583, %r1558, %r1550, %r1552;
mov.b32 %f158, %r1585;
add.ftz.f32 %f159, %f157, %f158;
mov.b32 %r1586, %f159;
shfl.sync.bfly.b32 %r1587|%p17, %r1586, %r1548, %r1550, %r1552;
mov.b32 %f160, %r1587;
add.ftz.f32 %f161, %f159, %f160;
mov.b32 %r1588, %f161;
shfl.sync.bfly.b32 %r1590|%p18, %r1588, %r1543, %r1550, %r1552;
mov.b32 %f162, %r1590;
add.ftz.f32 %f44, %f161, %f162;
setp.eq.s32 %p19, %r46, 1;
or.b32 %r1593, %r1542, 1;
setp.lt.u32 %p20, %r1593, %r31;
and.pred %p1, %p19, %p20;
not.pred %p21, %p1;
@%p21 bra $L__BB222_10;

cvta.to.global.u64 %rd95, %rd18;
mov.u32 %r1594, %ctaid.y;
mad.lo.s32 %r1597, %r1594, %r37, %r1542;
mov.u32 %r1598, %ctaid.z;
mad.lo.s32 %r1599, %r1598, %r43, %r1597;
mul.wide.u32 %rd96, %r1599, 4;
add.s64 %rd97, %rd95, %rd96;
st.global.f32 [%rd97+4], %f44;

$L__BB222_10:
mov.u32 %r1604, 2;
ld.shared.f32 %f163, [%r1547+256];
add.ftz.f32 %f164, %f163, %f320;
mov.b32 %r1605, %f164;
mov.u32 %r1606, 31;
mov.u32 %r1607, 16;
mov.u32 %r1608, -1;
shfl.sync.bfly.b32 %r1609|%p22, %r1605, %r1607, %r1606, %r1608;
mov.b32 %f165, %r1609;
add.ftz.f32 %f166, %f164, %f165;
mov.b32 %r1610, %f166;
mov.u32 %r1611, 8;
shfl.sync.bfly.b32 %r1612|%p23, %r1610, %r1611, %r1606, %r1608;
mov.b32 %f167, %r1612;
add.ftz.f32 %f168, %f166, %f167;
mov.b32 %r1613, %f168;
mov.u32 %r1614, 4;
shfl.sync.bfly.b32 %r1615|%p24, %r1613, %r1614, %r1606, %r1608;
mov.b32 %f169, %r1615;
add.ftz.f32 %f170, %f168, %f169;
mov.b32 %r1616, %f170;
shfl.sync.bfly.b32 %r1617|%p25, %r1616, %r1604, %r1606, %r1608;
mov.b32 %f171, %r1617;
add.ftz.f32 %f172, %f170, %f171;
mov.b32 %r1618, %f172;
mov.u32 %r1619, 1;
shfl.sync.bfly.b32 %r1620|%p26, %r1618, %r1619, %r1606, %r1608;
mov.b32 %f173, %r1620;
add.ftz.f32 %f45, %f172, %f173;
@%p13 bra $L__BB222_12;

cvt.u64.u32 %rd98, %r31;
mov.u32 %r1623, %ctaid.y;
mad.lo.s32 %r1626, %r1623, %r37, %r1542;
mov.u32 %r1627, %ctaid.z;
mad.lo.s32 %r1628, %r1627, %r43, %r1626;
cvt.u64.u32 %rd99, %r1628;
add.s64 %rd100, %rd98, %rd99;
cvta.to.global.u64 %rd101, %rd18;
shl.b64 %rd102, %rd100, 2;
add.s64 %rd103, %rd101, %rd102;
st.global.f32 [%rd103], %f45;

$L__BB222_12:
ld.shared.f32 %f174, [%r1547+384];
add.ftz.f32 %f175, %f174, %f319;
mov.b32 %r1634, %f175;
shfl.sync.bfly.b32 %r1638|%p30, %r1634, %r1607, %r1606, %r1608;
mov.b32 %f176, %r1638;
add.ftz.f32 %f177, %f175, %f176;
mov.b32 %r1639, %f177;
shfl.sync.bfly.b32 %r1641|%p31, %r1639, %r1611, %r1606, %r1608;
mov.b32 %f178, %r1641;
add.ftz.f32 %f179, %f177, %f178;
mov.b32 %r1642, %f179;
shfl.sync.bfly.b32 %r1644|%p32, %r1642, %r1614, %r1606, %r1608;
mov.b32 %f180, %r1644;
add.ftz.f32 %f181, %f179, %f180;
mov.b32 %r1645, %f181;
shfl.sync.bfly.b32 %r1646|%p33, %r1645, %r1604, %r1606, %r1608;
mov.b32 %f182, %r1646;
add.ftz.f32 %f183, %f181, %f182;
mov.b32 %r1647, %f183;
shfl.sync.bfly.b32 %r1649|%p34, %r1647, %r1619, %r1606, %r1608;
mov.b32 %f184, %r1649;
add.ftz.f32 %f46, %f183, %f184;
@%p21 bra $L__BB222_14;

add.s32 %r1650, %r31, 1;
cvt.u64.u32 %rd104, %r1650;
mov.u32 %r1651, %ctaid.y;
mad.lo.s32 %r1654, %r1651, %r37, %r1542;
mov.u32 %r1655, %ctaid.z;
mad.lo.s32 %r1656, %r1655, %r43, %r1654;
cvt.u64.u32 %rd105, %r1656;
add.s64 %rd106, %rd104, %rd105;
cvta.to.global.u64 %rd107, %rd18;
shl.b64 %rd108, %rd106, 2;
add.s64 %rd109, %rd107, %rd108;
st.global.f32 [%rd109], %f46;

$L__BB222_14:
shl.b32 %r1657, %r31, 1;
mov.u32 %r1658, 1;
mov.u32 %r1663, 2;
ld.shared.f32 %f185, [%r1547+512];
add.ftz.f32 %f186, %f185, %f318;
mov.b32 %r1664, %f186;
mov.u32 %r1665, 31;
mov.u32 %r1666, 16;
mov.u32 %r1667, -1;
shfl.sync.bfly.b32 %r1668|%p36, %r1664, %r1666, %r1665, %r1667;
mov.b32 %f187, %r1668;
add.ftz.f32 %f188, %f186, %f187;
mov.b32 %r1669, %f188;
mov.u32 %r1670, 8;
shfl.sync.bfly.b32 %r1671|%p37, %r1669, %r1670, %r1665, %r1667;
mov.b32 %f189, %r1671;
add.ftz.f32 %f190, %f188, %f189;
mov.b32 %r1672, %f190;
mov.u32 %r1673, 4;
shfl.sync.bfly.b32 %r1674|%p38, %r1672, %r1673, %r1665, %r1667;
mov.b32 %f191, %r1674;
add.ftz.f32 %f192, %f190, %f191;
mov.b32 %r1675, %f192;
shfl.sync.bfly.b32 %r1676|%p39, %r1675, %r1663, %r1665, %r1667;
mov.b32 %f193, %r1676;
add.ftz.f32 %f194, %f192, %f193;
mov.b32 %r1677, %f194;
shfl.sync.bfly.b32 %r1678|%p40, %r1677, %r1658, %r1665, %r1667;
mov.b32 %f195, %r1678;
add.ftz.f32 %f47, %f194, %f195;
cvt.u64.u32 %rd110, %r1657;
mov.u32 %r1679, %ctaid.y;
mad.lo.s32 %r1682, %r1679, %r37, %r1542;
mov.u32 %r1683, %ctaid.z;
mad.lo.s32 %r1684, %r1683, %r43, %r1682;
cvt.u64.u32 %rd111, %r1684;
add.s64 %rd112, %rd110, %rd111;
cvta.to.global.u64 %rd113, %rd18;
shl.b64 %rd114, %rd112, 2;
add.s64 %rd13, %rd113, %rd114;
@%p13 bra $L__BB222_16;

st.global.f32 [%rd13], %f47;

$L__BB222_16:
ld.shared.f32 %f196, [%r1547+640];
add.ftz.f32 %f197, %f196, %f317;
mov.b32 %r1690, %f197;
shfl.sync.bfly.b32 %r1694|%p44, %r1690, %r1666, %r1665, %r1667;
mov.b32 %f198, %r1694;
add.ftz.f32 %f199, %f197, %f198;
mov.b32 %r1695, %f199;
shfl.sync.bfly.b32 %r1697|%p45, %r1695, %r1670, %r1665, %r1667;
mov.b32 %f200, %r1697;
add.ftz.f32 %f201, %f199, %f200;
mov.b32 %r1698, %f201;
shfl.sync.bfly.b32 %r1700|%p46, %r1698, %r1673, %r1665, %r1667;
mov.b32 %f202, %r1700;
add.ftz.f32 %f203, %f201, %f202;
mov.b32 %r1701, %f203;
shfl.sync.bfly.b32 %r1702|%p47, %r1701, %r1663, %r1665, %r1667;
mov.b32 %f204, %r1702;
add.ftz.f32 %f205, %f203, %f204;
mov.b32 %r1703, %f205;
shfl.sync.bfly.b32 %r1705|%p48, %r1703, %r1658, %r1665, %r1667;
mov.b32 %f206, %r1705;
add.ftz.f32 %f48, %f205, %f206;
@%p21 bra $L__BB222_18;

st.global.f32 [%rd13+4], %f48;

$L__BB222_18:
mul.lo.s32 %r26, %r31, 3;
mov.u32 %r1710, 2;
ld.shared.f32 %f207, [%r1547+768];
add.ftz.f32 %f208, %f207, %f316;
mov.b32 %r1711, %f208;
mov.u32 %r1712, 31;
mov.u32 %r1713, 16;
mov.u32 %r1714, -1;
shfl.sync.bfly.b32 %r1715|%p50, %r1711, %r1713, %r1712, %r1714;
mov.b32 %f209, %r1715;
add.ftz.f32 %f210, %f208, %f209;
mov.b32 %r1716, %f210;
mov.u32 %r1717, 8;
shfl.sync.bfly.b32 %r1718|%p51, %r1716, %r1717, %r1712, %r1714;
mov.b32 %f211, %r1718;
add.ftz.f32 %f212, %f210, %f211;
mov.b32 %r1719, %f212;
mov.u32 %r1720, 4;
shfl.sync.bfly.b32 %r1721|%p52, %r1719, %r1720, %r1712, %r1714;
mov.b32 %f213, %r1721;
add.ftz.f32 %f214, %f212, %f213;
mov.b32 %r1722, %f214;
shfl.sync.bfly.b32 %r1723|%p53, %r1722, %r1710, %r1712, %r1714;
mov.b32 %f215, %r1723;
add.ftz.f32 %f216, %f214, %f215;
mov.b32 %r1724, %f216;
mov.u32 %r1725, 1;
shfl.sync.bfly.b32 %r1726|%p54, %r1724, %r1725, %r1712, %r1714;
mov.b32 %f217, %r1726;
add.ftz.f32 %f49, %f216, %f217;
@%p13 bra $L__BB222_20;

cvt.u64.u32 %rd115, %r26;
add.s64 %rd117, %rd115, %rd111;
shl.b64 %rd119, %rd117, 2;
add.s64 %rd120, %rd113, %rd119;
st.global.f32 [%rd120], %f49;

$L__BB222_20:
ld.shared.f32 %f218, [%r1547+896];
add.ftz.f32 %f219, %f218, %f315;
mov.b32 %r1740, %f219;
shfl.sync.bfly.b32 %r1744|%p58, %r1740, %r1713, %r1712, %r1714;
mov.b32 %f220, %r1744;
add.ftz.f32 %f221, %f219, %f220;
mov.b32 %r1745, %f221;
shfl.sync.bfly.b32 %r1747|%p59, %r1745, %r1717, %r1712, %r1714;
mov.b32 %f222, %r1747;
add.ftz.f32 %f223, %f221, %f222;
mov.b32 %r1748, %f223;
shfl.sync.bfly.b32 %r1750|%p60, %r1748, %r1720, %r1712, %r1714;
mov.b32 %f224, %r1750;
add.ftz.f32 %f225, %f223, %f224;
mov.b32 %r1751, %f225;
shfl.sync.bfly.b32 %r1752|%p61, %r1751, %r1710, %r1712, %r1714;
mov.b32 %f226, %r1752;
add.ftz.f32 %f227, %f225, %f226;
mov.b32 %r1753, %f227;
shfl.sync.bfly.b32 %r1755|%p62, %r1753, %r1725, %r1712, %r1714;
mov.b32 %f228, %r1755;
add.ftz.f32 %f50, %f227, %f228;
@%p21 bra $L__BB222_22;

add.s32 %r1756, %r26, 1;
cvt.u64.u32 %rd121, %r1756;
add.s64 %rd123, %rd121, %rd111;
shl.b64 %rd125, %rd123, 2;
add.s64 %rd126, %rd113, %rd125;
st.global.f32 [%rd126], %f50;

$L__BB222_22:
shl.b32 %r1763, %r31, 2;
mov.u32 %r1764, 2;
ld.shared.f32 %f229, [%r1547+1024];
add.ftz.f32 %f230, %f229, %f314;
mov.b32 %r1769, %f230;
mov.u32 %r1770, 31;
mov.u32 %r1771, 16;
mov.u32 %r1772, -1;
shfl.sync.bfly.b32 %r1773|%p64, %r1769, %r1771, %r1770, %r1772;
mov.b32 %f231, %r1773;
add.ftz.f32 %f232, %f230, %f231;
mov.b32 %r1774, %f232;
mov.u32 %r1775, 8;
shfl.sync.bfly.b32 %r1776|%p65, %r1774, %r1775, %r1770, %r1772;
mov.b32 %f233, %r1776;
add.ftz.f32 %f234, %f232, %f233;
mov.b32 %r1777, %f234;
mov.u32 %r1778, 4;
shfl.sync.bfly.b32 %r1779|%p66, %r1777, %r1778, %r1770, %r1772;
mov.b32 %f235, %r1779;
add.ftz.f32 %f236, %f234, %f235;
mov.b32 %r1780, %f236;
shfl.sync.bfly.b32 %r1781|%p67, %r1780, %r1764, %r1770, %r1772;
mov.b32 %f237, %r1781;
add.ftz.f32 %f238, %f236, %f237;
mov.b32 %r1782, %f238;
mov.u32 %r1783, 1;
shfl.sync.bfly.b32 %r1784|%p68, %r1782, %r1783, %r1770, %r1772;
mov.b32 %f239, %r1784;
add.ftz.f32 %f51, %f238, %f239;
cvt.u64.u32 %rd127, %r1763;
add.s64 %rd129, %rd127, %rd111;
shl.b64 %rd131, %rd129, 2;
add.s64 %rd14, %rd113, %rd131;
@%p13 bra $L__BB222_24;

st.global.f32 [%rd14], %f51;

$L__BB222_24:
ld.shared.f32 %f240, [%r1547+1152];
add.ftz.f32 %f241, %f240, %f313;
mov.b32 %r1796, %f241;
shfl.sync.bfly.b32 %r1800|%p72, %r1796, %r1771, %r1770, %r1772;
mov.b32 %f242, %r1800;
add.ftz.f32 %f243, %f241, %f242;
mov.b32 %r1801, %f243;
shfl.sync.bfly.b32 %r1803|%p73, %r1801, %r1775, %r1770, %r1772;
mov.b32 %f244, %r1803;
add.ftz.f32 %f245, %f243, %f244;
mov.b32 %r1804, %f245;
shfl.sync.bfly.b32 %r1806|%p74, %r1804, %r1778, %r1770, %r1772;
mov.b32 %f246, %r1806;
add.ftz.f32 %f247, %f245, %f246;
mov.b32 %r1807, %f247;
shfl.sync.bfly.b32 %r1808|%p75, %r1807, %r1764, %r1770, %r1772;
mov.b32 %f248, %r1808;
add.ftz.f32 %f249, %f247, %f248;
mov.b32 %r1809, %f249;
shfl.sync.bfly.b32 %r1811|%p76, %r1809, %r1783, %r1770, %r1772;
mov.b32 %f250, %r1811;
add.ftz.f32 %f52, %f249, %f250;
@%p21 bra $L__BB222_26;

st.global.f32 [%rd14+4], %f52;

$L__BB222_26:
mul.lo.s32 %r27, %r31, 5;
mov.u32 %r1816, 2;
ld.shared.f32 %f251, [%r1547+1280];
add.ftz.f32 %f252, %f251, %f312;
mov.b32 %r1817, %f252;
mov.u32 %r1818, 31;
mov.u32 %r1819, 16;
mov.u32 %r1820, -1;
shfl.sync.bfly.b32 %r1821|%p78, %r1817, %r1819, %r1818, %r1820;
mov.b32 %f253, %r1821;
add.ftz.f32 %f254, %f252, %f253;
mov.b32 %r1822, %f254;
mov.u32 %r1823, 8;
shfl.sync.bfly.b32 %r1824|%p79, %r1822, %r1823, %r1818, %r1820;
mov.b32 %f255, %r1824;
add.ftz.f32 %f256, %f254, %f255;
mov.b32 %r1825, %f256;
mov.u32 %r1826, 4;
shfl.sync.bfly.b32 %r1827|%p80, %r1825, %r1826, %r1818, %r1820;
mov.b32 %f257, %r1827;
add.ftz.f32 %f258, %f256, %f257;
mov.b32 %r1828, %f258;
shfl.sync.bfly.b32 %r1829|%p81, %r1828, %r1816, %r1818, %r1820;
mov.b32 %f259, %r1829;
add.ftz.f32 %f260, %f258, %f259;
mov.b32 %r1830, %f260;
mov.u32 %r1831, 1;
shfl.sync.bfly.b32 %r1832|%p82, %r1830, %r1831, %r1818, %r1820;
mov.b32 %f261, %r1832;
add.ftz.f32 %f53, %f260, %f261;
@%p13 bra $L__BB222_28;

cvt.u64.u32 %rd132, %r27;
add.s64 %rd134, %rd132, %rd111;
shl.b64 %rd136, %rd134, 2;
add.s64 %rd137, %rd113, %rd136;
st.global.f32 [%rd137], %f53;

$L__BB222_28:
ld.shared.f32 %f262, [%r1547+1408];
add.ftz.f32 %f263, %f262, %f311;
mov.b32 %r1846, %f263;
shfl.sync.bfly.b32 %r1850|%p86, %r1846, %r1819, %r1818, %r1820;
mov.b32 %f264, %r1850;
add.ftz.f32 %f265, %f263, %f264;
mov.b32 %r1851, %f265;
shfl.sync.bfly.b32 %r1853|%p87, %r1851, %r1823, %r1818, %r1820;
mov.b32 %f266, %r1853;
add.ftz.f32 %f267, %f265, %f266;
mov.b32 %r1854, %f267;
shfl.sync.bfly.b32 %r1856|%p88, %r1854, %r1826, %r1818, %r1820;
mov.b32 %f268, %r1856;
add.ftz.f32 %f269, %f267, %f268;
mov.b32 %r1857, %f269;
shfl.sync.bfly.b32 %r1858|%p89, %r1857, %r1816, %r1818, %r1820;
mov.b32 %f270, %r1858;
add.ftz.f32 %f271, %f269, %f270;
mov.b32 %r1859, %f271;
shfl.sync.bfly.b32 %r1861|%p90, %r1859, %r1831, %r1818, %r1820;
mov.b32 %f272, %r1861;
add.ftz.f32 %f54, %f271, %f272;
@%p21 bra $L__BB222_30;

add.s32 %r1862, %r27, 1;
cvt.u64.u32 %rd138, %r1862;
add.s64 %rd140, %rd138, %rd111;
shl.b64 %rd142, %rd140, 2;
add.s64 %rd143, %rd113, %rd142;
st.global.f32 [%rd143], %f54;

$L__BB222_30:
mul.lo.s32 %r1869, %r31, 6;
mov.u32 %r1874, 2;
ld.shared.f32 %f273, [%r1547+1536];
add.ftz.f32 %f274, %f273, %f310;
mov.b32 %r1875, %f274;
mov.u32 %r1876, 31;
mov.u32 %r1877, 16;
mov.u32 %r1878, -1;
shfl.sync.bfly.b32 %r1879|%p92, %r1875, %r1877, %r1876, %r1878;
mov.b32 %f275, %r1879;
add.ftz.f32 %f276, %f274, %f275;
mov.b32 %r1880, %f276;
mov.u32 %r1881, 8;
shfl.sync.bfly.b32 %r1882|%p93, %r1880, %r1881, %r1876, %r1878;
mov.b32 %f277, %r1882;
add.ftz.f32 %f278, %f276, %f277;
mov.b32 %r1883, %f278;
mov.u32 %r1884, 4;
shfl.sync.bfly.b32 %r1885|%p94, %r1883, %r1884, %r1876, %r1878;
mov.b32 %f279, %r1885;
add.ftz.f32 %f280, %f278, %f279;
mov.b32 %r1886, %f280;
shfl.sync.bfly.b32 %r1887|%p95, %r1886, %r1874, %r1876, %r1878;
mov.b32 %f281, %r1887;
add.ftz.f32 %f282, %f280, %f281;
mov.b32 %r1888, %f282;
mov.u32 %r1889, 1;
shfl.sync.bfly.b32 %r1890|%p96, %r1888, %r1889, %r1876, %r1878;
mov.b32 %f283, %r1890;
add.ftz.f32 %f55, %f282, %f283;
cvt.u64.u32 %rd144, %r1869;
add.s64 %rd146, %rd144, %rd111;
shl.b64 %rd148, %rd146, 2;
add.s64 %rd15, %rd113, %rd148;
@%p13 bra $L__BB222_32;

st.global.f32 [%rd15], %f55;

$L__BB222_32:
ld.shared.f32 %f284, [%r1547+1664];
add.ftz.f32 %f285, %f284, %f309;
mov.b32 %r1902, %f285;
shfl.sync.bfly.b32 %r1906|%p100, %r1902, %r1877, %r1876, %r1878;
mov.b32 %f286, %r1906;
add.ftz.f32 %f287, %f285, %f286;
mov.b32 %r1907, %f287;
shfl.sync.bfly.b32 %r1909|%p101, %r1907, %r1881, %r1876, %r1878;
mov.b32 %f288, %r1909;
add.ftz.f32 %f289, %f287, %f288;
mov.b32 %r1910, %f289;
shfl.sync.bfly.b32 %r1912|%p102, %r1910, %r1884, %r1876, %r1878;
mov.b32 %f290, %r1912;
add.ftz.f32 %f291, %f289, %f290;
mov.b32 %r1913, %f291;
shfl.sync.bfly.b32 %r1914|%p103, %r1913, %r1874, %r1876, %r1878;
mov.b32 %f292, %r1914;
add.ftz.f32 %f293, %f291, %f292;
mov.b32 %r1915, %f293;
shfl.sync.bfly.b32 %r1917|%p104, %r1915, %r1889, %r1876, %r1878;
mov.b32 %f294, %r1917;
add.ftz.f32 %f56, %f293, %f294;
@%p21 bra $L__BB222_34;

st.global.f32 [%rd15+4], %f56;

$L__BB222_34:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type17ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type17ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type17ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type17ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type17ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type17ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type17ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type17ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type17ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type17ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<120>;
.reg .b16 %rs<49>;
.reg .f32 %f<369>;
.reg .b32 %r<2173>;
.reg .b64 %rd<167>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type17ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[2048];

ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type17ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type17ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type17ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type17ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type17ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type17ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type17ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type17ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type17ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type17ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r45, [_Z13mul_mat_vec_qIL9ggml_type17ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r46, [_Z13mul_mat_vec_qIL9ggml_type17ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r47, [_Z13mul_mat_vec_qIL9ggml_type17ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type17ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type17ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type17ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type17ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r48, %tid.y;
shl.b32 %r49, %r48, 5;
mov.u32 %r50, %tid.x;
add.s32 %r1, %r49, %r50;
shr.s32 %r51, %r1, 31;
shr.u32 %r52, %r51, 29;
add.s32 %r53, %r1, %r52;
shr.s32 %r2172, %r53, 3;
shr.u32 %r54, %r32, 8;
setp.ge.s32 %p2, %r2172, %r54;
mov.f32 %f353, 0f00000000;
mov.f32 %f354, %f353;
mov.f32 %f355, %f353;
mov.f32 %f356, %f353;
mov.f32 %f357, %f353;
mov.f32 %f358, %f353;
mov.f32 %f359, %f353;
mov.f32 %f360, %f353;
mov.f32 %f361, %f353;
mov.f32 %f362, %f353;
mov.f32 %f363, %f353;
mov.f32 %f364, %f353;
mov.f32 %f365, %f353;
mov.f32 %f366, %f353;
mov.f32 %f367, %f353;
mov.f32 %f368, %f353;
@%p2 bra $L__BB223_3;

mov.u32 %r55, %ctaid.x;
shl.b32 %r56, %r55, 1;
mov.u32 %r57, %ctaid.y;
mov.u32 %r58, %ctaid.z;
mul.lo.s32 %r59, %r57, %r40;
cvt.u64.u32 %rd19, %r59;
mul.lo.s32 %r60, %r58, %r46;
cvt.u64.u32 %rd20, %r60;
add.s64 %rd1, %rd20, %rd19;
mul.hi.u32 %r61, %r58, %r42;
add.s32 %r62, %r61, %r58;
shr.u32 %r63, %r62, %r43;
mul.lo.s32 %r64, %r63, %r45;
mul.hi.u32 %r65, %r57, %r36;
add.s32 %r66, %r65, %r57;
shr.u32 %r67, %r66, %r37;
mul.lo.s32 %r68, %r67, %r39;
cvta.to.global.u64 %rd166, %rd16;
and.b32 %r72, %r53, -8;
sub.s32 %r73, %r1, %r72;
shl.b32 %r74, %r73, 2;
cvt.s64.s32 %rd3, %r73;
shl.b32 %r75, %r34, 1;
shl.b32 %r2164, %r2172, 3;
add.s32 %r2171, %r2164, %r75;
mad.lo.s32 %r2170, %r34, 3, %r2164;
shl.b32 %r76, %r34, 2;
add.s32 %r2169, %r2164, %r76;
mad.lo.s32 %r2168, %r34, 5, %r2164;
mad.lo.s32 %r2167, %r34, 6, %r2164;
mad.lo.s32 %r2166, %r34, 7, %r2164;
add.s32 %r2165, %r34, %r2164;
add.s32 %r77, %r2172, %r64;
add.s32 %r78, %r77, %r68;
or.b32 %r79, %r56, 1;
mad.lo.s32 %r80, %r33, %r79, %r78;
mul.wide.s32 %rd4, %r80, 74;
add.s64 %rd21, %rd4, %rd3;
add.s64 %rd5, %rd21, 66;
mad.lo.s32 %r81, %r56, %r33, %r2172;
add.s32 %r82, %r81, %r64;
add.s32 %r83, %r82, %r68;
mul.wide.s32 %rd6, %r83, 74;
add.s64 %rd22, %rd6, %rd3;
add.s64 %rd7, %rd22, 66;
mul.wide.s32 %rd23, %r74, 2;
add.s64 %rd24, %rd4, %rd23;
add.s64 %rd8, %rd24, 4;
add.s64 %rd25, %rd6, %rd23;
add.s64 %rd9, %rd25, 8;
cvta.to.global.u64 %rd10, %rd17;
mov.f32 %f353, 0f00000000;

$L__BB223_2:
cvt.u64.u32 %rd26, %r2164;
add.s64 %rd27, %rd1, %rd26;
add.s64 %rd28, %rd27, %rd3;
mul.lo.s64 %rd29, %rd28, 36;
add.s64 %rd30, %rd166, %rd9;
ld.global.nc.u16 %rs17, [%rd30+-4];
ld.global.nc.u16 %rs18, [%rd30+-2];
ld.global.nc.u16 %rs19, [%rd30];
add.s64 %rd31, %rd166, %rd7;
ld.global.nc.u8 %rs20, [%rd31];
and.b16 %rs21, %rs20, 240;
and.b16 %rs22, %rs20, 15;
shr.u16 %rs23, %rs21, 4;
ld.global.nc.u16 %rs24, [%rd30+-6];
shl.b16 %rs25, %rs24, 3;
cvt.u64.u16 %rd32, %rs25;
and.b64 %rd33, %rd32, 4088;
mov.u64 %rd34, iq2xs_grid;
add.s64 %rd35, %rd34, %rd33;
ld.global.nc.v2.u32 {%r868, %r869}, [%rd35];
not.b32 %r871, %r868;
mov.u32 %r850, 0;
not.b32 %r873, %r869;
shr.u16 %rs26, %rs24, 9;
cvt.u32.u16 %r874, %rs26;
popc.b32 %r875, %r874;
shl.b32 %r876, %r875, 7;
and.b32 %r877, %r876, 128;
or.b32 %r878, %r877, %r874;
mul.lo.s32 %r879, %r878, 16843009;
and.b32 %r880, %r879, 134480385;
or.b32 %r881, %r880, -2139062144;
add.s32 %r882, %r881, -16843009;
or.b32 %r770, %r882, %r880;
mov.u32 %r84, %r770;

prmt.b32 %r84, %r84, 0, 0xba98;

xor.b32 %r883, %r84, %r868;
or.b32 %r884, %r883, -2139062144;
and.b32 %r885, %r84, 2139062143;
and.b32 %r886, %r871, -2139062144;
sub.s32 %r887, %r884, %r885;
xor.b32 %r89, %r886, %r887;
add.s64 %rd36, %rd10, %rd29;
ld.global.nc.u32 %r90, [%rd36+4];
and.b32 %r888, %r879, -2143281136;
or.b32 %r889, %r888, -2139062144;
add.s32 %r890, %r889, -16843009;
or.b32 %r772, %r890, %r888;
mov.u32 %r86, %r772;

prmt.b32 %r86, %r86, 0, 0xba98;

xor.b32 %r891, %r86, %r869;
or.b32 %r892, %r891, -2139062144;
and.b32 %r893, %r86, 2139062143;
and.b32 %r894, %r873, -2139062144;
sub.s32 %r895, %r892, %r893;
xor.b32 %r93, %r894, %r895;
ld.global.nc.u32 %r94, [%rd36+8];

dp4a.s32.s32 %r88, %r89, %r90, %r850;


dp4a.s32.s32 %r92, %r93, %r94, %r88;

shl.b16 %rs27, %rs17, 3;
cvt.u64.u16 %rd37, %rs27;
and.b64 %rd38, %rd37, 4088;
add.s64 %rd39, %rd34, %rd38;
ld.global.nc.v2.u32 {%r896, %r897}, [%rd39];
not.b32 %r899, %r896;
not.b32 %r901, %r897;
shr.u16 %rs28, %rs17, 9;
cvt.u32.u16 %r902, %rs28;
popc.b32 %r903, %r902;
shl.b32 %r904, %r903, 7;
and.b32 %r905, %r904, 128;
or.b32 %r906, %r905, %r902;
mul.lo.s32 %r907, %r906, 16843009;
and.b32 %r908, %r907, 134480385;
or.b32 %r909, %r908, -2139062144;
add.s32 %r910, %r909, -16843009;
or.b32 %r782, %r910, %r908;
mov.u32 %r96, %r782;

prmt.b32 %r96, %r96, 0, 0xba98;

xor.b32 %r911, %r96, %r896;
or.b32 %r912, %r911, -2139062144;
and.b32 %r913, %r96, 2139062143;
and.b32 %r914, %r899, -2139062144;
sub.s32 %r915, %r912, %r913;
xor.b32 %r101, %r914, %r915;
ld.global.nc.u32 %r102, [%rd36+12];
and.b32 %r916, %r907, -2143281136;
or.b32 %r917, %r916, -2139062144;
add.s32 %r918, %r917, -16843009;
or.b32 %r784, %r918, %r916;
mov.u32 %r98, %r784;

prmt.b32 %r98, %r98, 0, 0xba98;

xor.b32 %r919, %r98, %r897;
or.b32 %r920, %r919, -2139062144;
and.b32 %r921, %r98, 2139062143;
and.b32 %r922, %r901, -2139062144;
sub.s32 %r923, %r920, %r921;
xor.b32 %r105, %r922, %r923;
ld.global.nc.u32 %r106, [%rd36+16];

dp4a.s32.s32 %r100, %r101, %r102, %r92;


dp4a.s32.s32 %r104, %r105, %r106, %r100;

shl.b16 %rs29, %rs18, 3;
cvt.u64.u16 %rd40, %rs29;
and.b64 %rd41, %rd40, 4088;
add.s64 %rd42, %rd34, %rd41;
ld.global.nc.v2.u32 {%r924, %r925}, [%rd42];
not.b32 %r927, %r924;
not.b32 %r929, %r925;
shr.u16 %rs30, %rs18, 9;
cvt.u32.u16 %r930, %rs30;
popc.b32 %r931, %r930;
shl.b32 %r932, %r931, 7;
and.b32 %r933, %r932, 128;
or.b32 %r934, %r933, %r930;
mul.lo.s32 %r935, %r934, 16843009;
and.b32 %r936, %r935, 134480385;
or.b32 %r937, %r936, -2139062144;
add.s32 %r938, %r937, -16843009;
or.b32 %r794, %r938, %r936;
mov.u32 %r108, %r794;

prmt.b32 %r108, %r108, 0, 0xba98;

xor.b32 %r939, %r108, %r924;
or.b32 %r940, %r939, -2139062144;
and.b32 %r941, %r108, 2139062143;
and.b32 %r942, %r927, -2139062144;
sub.s32 %r943, %r940, %r941;
xor.b32 %r113, %r942, %r943;
ld.global.nc.u32 %r114, [%rd36+20];
and.b32 %r944, %r935, -2143281136;
or.b32 %r945, %r944, -2139062144;
add.s32 %r946, %r945, -16843009;
or.b32 %r796, %r946, %r944;
mov.u32 %r110, %r796;

prmt.b32 %r110, %r110, 0, 0xba98;

xor.b32 %r947, %r110, %r925;
or.b32 %r948, %r947, -2139062144;
and.b32 %r949, %r110, 2139062143;
and.b32 %r950, %r929, -2139062144;
sub.s32 %r951, %r948, %r949;
xor.b32 %r117, %r950, %r951;
ld.global.nc.u32 %r118, [%rd36+24];

dp4a.s32.s32 %r112, %r113, %r114, %r850;


dp4a.s32.s32 %r116, %r117, %r118, %r112;

shl.b16 %rs31, %rs19, 3;
cvt.u64.u16 %rd43, %rs31;
and.b64 %rd44, %rd43, 4088;
add.s64 %rd45, %rd34, %rd44;
ld.global.nc.v2.u32 {%r952, %r953}, [%rd45];
not.b32 %r955, %r952;
not.b32 %r957, %r953;
shr.u16 %rs32, %rs19, 9;
cvt.u32.u16 %r958, %rs32;
popc.b32 %r959, %r958;
shl.b32 %r960, %r959, 7;
and.b32 %r961, %r960, 128;
or.b32 %r962, %r961, %r958;
mul.lo.s32 %r963, %r962, 16843009;
and.b32 %r964, %r963, 134480385;
or.b32 %r965, %r964, -2139062144;
add.s32 %r966, %r965, -16843009;
or.b32 %r806, %r966, %r964;
mov.u32 %r120, %r806;

prmt.b32 %r120, %r120, 0, 0xba98;

xor.b32 %r967, %r120, %r952;
or.b32 %r968, %r967, -2139062144;
and.b32 %r969, %r120, 2139062143;
and.b32 %r970, %r955, -2139062144;
sub.s32 %r971, %r968, %r969;
xor.b32 %r125, %r970, %r971;
ld.global.nc.u32 %r126, [%rd36+28];
and.b32 %r972, %r963, -2143281136;
or.b32 %r973, %r972, -2139062144;
add.s32 %r974, %r973, -16843009;
or.b32 %r808, %r974, %r972;
mov.u32 %r122, %r808;

prmt.b32 %r122, %r122, 0, 0xba98;

xor.b32 %r975, %r122, %r953;
or.b32 %r976, %r975, -2139062144;
and.b32 %r977, %r122, 2139062143;
and.b32 %r978, %r957, -2139062144;
sub.s32 %r979, %r976, %r977;
xor.b32 %r129, %r978, %r979;
ld.global.nc.u32 %r130, [%rd36+32];

dp4a.s32.s32 %r124, %r125, %r126, %r116;


dp4a.s32.s32 %r128, %r129, %r130, %r124;

cvt.u32.u16 %r980, %rs22;
cvt.u32.u16 %r981, %rs23;
mul.lo.s32 %r982, %r104, %r980;
mad.lo.s32 %r983, %r128, %r981, %r982;
add.s32 %r984, %r128, %r104;
shr.u32 %r985, %r984, 31;
add.s32 %r986, %r984, %r985;
shr.s32 %r987, %r986, 1;
add.s32 %r988, %r983, %r987;
shr.s32 %r989, %r988, 31;
shr.u32 %r990, %r989, 30;
add.s32 %r991, %r988, %r990;
shr.s32 %r992, %r991, 2;
add.s64 %rd46, %rd166, %rd6;
ld.global.nc.u16 %rs1, [%rd46];

{ cvt.f32.f16 %f97, %rs1;}


ld.global.nc.u32 %r132, [%rd36];

{.reg .f16 low,high;
mov.b32 {low,high},%r132;
cvt.f32.f16 %f98, low;}


mul.ftz.f32 %f129, %f97, %f98;
cvt.rn.f32.s32 %f130, %r992;
fma.rn.ftz.f32 %f368, %f129, %f130, %f368;
add.s64 %rd47, %rd166, %rd8;
ld.global.nc.u16 %rs33, [%rd47];
ld.global.nc.u16 %rs34, [%rd47+2];
ld.global.nc.u16 %rs35, [%rd47+4];
add.s64 %rd48, %rd166, %rd5;
ld.global.nc.u8 %rs36, [%rd48];
and.b16 %rs37, %rs36, 240;
and.b16 %rs38, %rs36, 15;
shr.u16 %rs39, %rs37, 4;
ld.global.nc.u16 %rs40, [%rd47+-2];
shl.b16 %rs41, %rs40, 3;
cvt.u64.u16 %rd49, %rs41;
and.b64 %rd50, %rd49, 4088;
add.s64 %rd51, %rd34, %rd50;
ld.global.nc.v2.u32 {%r993, %r994}, [%rd51];
not.b32 %r996, %r993;
not.b32 %r998, %r994;
shr.u16 %rs42, %rs40, 9;
cvt.u32.u16 %r999, %rs42;
popc.b32 %r1000, %r999;
shl.b32 %r1001, %r1000, 7;
and.b32 %r1002, %r1001, 128;
or.b32 %r1003, %r1002, %r999;
mul.lo.s32 %r1004, %r1003, 16843009;
and.b32 %r1005, %r1004, 134480385;
or.b32 %r1006, %r1005, -2139062144;
add.s32 %r1007, %r1006, -16843009;
or.b32 %r819, %r1007, %r1005;
mov.u32 %r133, %r819;

prmt.b32 %r133, %r133, 0, 0xba98;

xor.b32 %r1008, %r133, %r993;
or.b32 %r1009, %r1008, -2139062144;
and.b32 %r1010, %r133, 2139062143;
and.b32 %r1011, %r996, -2139062144;
sub.s32 %r1012, %r1009, %r1010;
xor.b32 %r138, %r1011, %r1012;
and.b32 %r1013, %r1004, -2143281136;
or.b32 %r1014, %r1013, -2139062144;
add.s32 %r1015, %r1014, -16843009;
or.b32 %r821, %r1015, %r1013;
mov.u32 %r135, %r821;

prmt.b32 %r135, %r135, 0, 0xba98;

xor.b32 %r1016, %r135, %r994;
or.b32 %r1017, %r1016, -2139062144;
and.b32 %r1018, %r135, 2139062143;
and.b32 %r1019, %r998, -2139062144;
sub.s32 %r1020, %r1017, %r1018;
xor.b32 %r142, %r1019, %r1020;

dp4a.s32.s32 %r137, %r138, %r90, %r850;


dp4a.s32.s32 %r141, %r142, %r94, %r137;

shl.b16 %rs43, %rs33, 3;
cvt.u64.u16 %rd52, %rs43;
and.b64 %rd53, %rd52, 4088;
add.s64 %rd54, %rd34, %rd53;
ld.global.nc.v2.u32 {%r1021, %r1022}, [%rd54];
not.b32 %r1024, %r1021;
not.b32 %r1026, %r1022;
shr.u16 %rs44, %rs33, 9;
cvt.u32.u16 %r1027, %rs44;
popc.b32 %r1028, %r1027;
shl.b32 %r1029, %r1028, 7;
and.b32 %r1030, %r1029, 128;
or.b32 %r1031, %r1030, %r1027;
mul.lo.s32 %r1032, %r1031, 16843009;
and.b32 %r1033, %r1032, 134480385;
or.b32 %r1034, %r1033, -2139062144;
add.s32 %r1035, %r1034, -16843009;
or.b32 %r831, %r1035, %r1033;
mov.u32 %r145, %r831;

prmt.b32 %r145, %r145, 0, 0xba98;

xor.b32 %r1036, %r145, %r1021;
or.b32 %r1037, %r1036, -2139062144;
and.b32 %r1038, %r145, 2139062143;
and.b32 %r1039, %r1024, -2139062144;
sub.s32 %r1040, %r1037, %r1038;
xor.b32 %r150, %r1039, %r1040;
and.b32 %r1041, %r1032, -2143281136;
or.b32 %r1042, %r1041, -2139062144;
add.s32 %r1043, %r1042, -16843009;
or.b32 %r833, %r1043, %r1041;
mov.u32 %r147, %r833;

prmt.b32 %r147, %r147, 0, 0xba98;

xor.b32 %r1044, %r147, %r1022;
or.b32 %r1045, %r1044, -2139062144;
and.b32 %r1046, %r147, 2139062143;
and.b32 %r1047, %r1026, -2139062144;
sub.s32 %r1048, %r1045, %r1046;
xor.b32 %r154, %r1047, %r1048;

dp4a.s32.s32 %r149, %r150, %r102, %r141;


dp4a.s32.s32 %r153, %r154, %r106, %r149;

shl.b16 %rs45, %rs34, 3;
cvt.u64.u16 %rd55, %rs45;
and.b64 %rd56, %rd55, 4088;
add.s64 %rd57, %rd34, %rd56;
ld.global.nc.v2.u32 {%r1049, %r1050}, [%rd57];
not.b32 %r1052, %r1049;
not.b32 %r1054, %r1050;
shr.u16 %rs46, %rs34, 9;
cvt.u32.u16 %r1055, %rs46;
popc.b32 %r1056, %r1055;
shl.b32 %r1057, %r1056, 7;
and.b32 %r1058, %r1057, 128;
or.b32 %r1059, %r1058, %r1055;
mul.lo.s32 %r1060, %r1059, 16843009;
and.b32 %r1061, %r1060, 134480385;
or.b32 %r1062, %r1061, -2139062144;
add.s32 %r1063, %r1062, -16843009;
or.b32 %r843, %r1063, %r1061;
mov.u32 %r157, %r843;

prmt.b32 %r157, %r157, 0, 0xba98;

xor.b32 %r1064, %r157, %r1049;
or.b32 %r1065, %r1064, -2139062144;
and.b32 %r1066, %r157, 2139062143;
and.b32 %r1067, %r1052, -2139062144;
sub.s32 %r1068, %r1065, %r1066;
xor.b32 %r162, %r1067, %r1068;
and.b32 %r1069, %r1060, -2143281136;
or.b32 %r1070, %r1069, -2139062144;
add.s32 %r1071, %r1070, -16843009;
or.b32 %r845, %r1071, %r1069;
mov.u32 %r159, %r845;

prmt.b32 %r159, %r159, 0, 0xba98;

xor.b32 %r1072, %r159, %r1050;
or.b32 %r1073, %r1072, -2139062144;
and.b32 %r1074, %r159, 2139062143;
and.b32 %r1075, %r1054, -2139062144;
sub.s32 %r1076, %r1073, %r1074;
xor.b32 %r166, %r1075, %r1076;

dp4a.s32.s32 %r161, %r162, %r114, %r850;


dp4a.s32.s32 %r165, %r166, %r118, %r161;

shl.b16 %rs47, %rs35, 3;
cvt.u64.u16 %rd58, %rs47;
and.b64 %rd59, %rd58, 4088;
add.s64 %rd60, %rd34, %rd59;
ld.global.nc.v2.u32 {%r1077, %r1078}, [%rd60];
not.b32 %r1080, %r1077;
not.b32 %r1082, %r1078;
shr.u16 %rs48, %rs35, 9;
cvt.u32.u16 %r1083, %rs48;
popc.b32 %r1084, %r1083;
shl.b32 %r1085, %r1084, 7;
and.b32 %r1086, %r1085, 128;
or.b32 %r1087, %r1086, %r1083;
mul.lo.s32 %r1088, %r1087, 16843009;
and.b32 %r1089, %r1088, 134480385;
or.b32 %r1090, %r1089, -2139062144;
add.s32 %r1091, %r1090, -16843009;
or.b32 %r855, %r1091, %r1089;
mov.u32 %r169, %r855;

prmt.b32 %r169, %r169, 0, 0xba98;

xor.b32 %r1092, %r169, %r1077;
or.b32 %r1093, %r1092, -2139062144;
and.b32 %r1094, %r169, 2139062143;
and.b32 %r1095, %r1080, -2139062144;
sub.s32 %r1096, %r1093, %r1094;
xor.b32 %r174, %r1095, %r1096;
and.b32 %r1097, %r1088, -2143281136;
or.b32 %r1098, %r1097, -2139062144;
add.s32 %r1099, %r1098, -16843009;
or.b32 %r857, %r1099, %r1097;
mov.u32 %r171, %r857;

prmt.b32 %r171, %r171, 0, 0xba98;

xor.b32 %r1100, %r171, %r1078;
or.b32 %r1101, %r1100, -2139062144;
and.b32 %r1102, %r171, 2139062143;
and.b32 %r1103, %r1082, -2139062144;
sub.s32 %r1104, %r1101, %r1102;
xor.b32 %r178, %r1103, %r1104;

dp4a.s32.s32 %r173, %r174, %r126, %r165;


dp4a.s32.s32 %r177, %r178, %r130, %r173;

cvt.u32.u16 %r1105, %rs38;
cvt.u32.u16 %r1106, %rs39;
mul.lo.s32 %r1107, %r153, %r1105;
mad.lo.s32 %r1108, %r177, %r1106, %r1107;
add.s32 %r1109, %r177, %r153;
shr.u32 %r1110, %r1109, 31;
add.s32 %r1111, %r1109, %r1110;
shr.s32 %r1112, %r1111, 1;
add.s32 %r1113, %r1108, %r1112;
shr.s32 %r1114, %r1113, 31;
shr.u32 %r1115, %r1114, 30;
add.s32 %r1116, %r1113, %r1115;
shr.s32 %r1117, %r1116, 2;
add.s64 %rd61, %rd166, %rd4;
ld.global.nc.u16 %rs2, [%rd61];

{ cvt.f32.f16 %f99, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r132;
cvt.f32.f16 %f100, low;}


mul.ftz.f32 %f131, %f99, %f100;
cvt.rn.f32.s32 %f132, %r1117;
fma.rn.ftz.f32 %f367, %f131, %f132, %f367;
cvt.u64.u32 %rd62, %r2165;
add.s64 %rd63, %rd1, %rd62;
add.s64 %rd64, %rd63, %rd3;
mul.lo.s64 %rd65, %rd64, 36;
mov.u32 %r182, %r770;

prmt.b32 %r182, %r182, 0, 0xba98;

xor.b32 %r1118, %r182, %r868;
or.b32 %r1119, %r1118, -2139062144;
and.b32 %r1120, %r182, 2139062143;
sub.s32 %r1121, %r1119, %r1120;
xor.b32 %r187, %r886, %r1121;
add.s64 %rd66, %rd10, %rd65;
ld.global.nc.u32 %r188, [%rd66+4];
mov.u32 %r184, %r772;

prmt.b32 %r184, %r184, 0, 0xba98;

xor.b32 %r1122, %r184, %r869;
or.b32 %r1123, %r1122, -2139062144;
and.b32 %r1124, %r184, 2139062143;
sub.s32 %r1125, %r1123, %r1124;
xor.b32 %r191, %r894, %r1125;
ld.global.nc.u32 %r192, [%rd66+8];

dp4a.s32.s32 %r186, %r187, %r188, %r850;


dp4a.s32.s32 %r190, %r191, %r192, %r186;

mov.u32 %r194, %r782;

prmt.b32 %r194, %r194, 0, 0xba98;

xor.b32 %r1126, %r194, %r896;
or.b32 %r1127, %r1126, -2139062144;
and.b32 %r1128, %r194, 2139062143;
sub.s32 %r1129, %r1127, %r1128;
xor.b32 %r199, %r914, %r1129;
ld.global.nc.u32 %r200, [%rd66+12];
mov.u32 %r196, %r784;

prmt.b32 %r196, %r196, 0, 0xba98;

xor.b32 %r1130, %r196, %r897;
or.b32 %r1131, %r1130, -2139062144;
and.b32 %r1132, %r196, 2139062143;
sub.s32 %r1133, %r1131, %r1132;
xor.b32 %r203, %r922, %r1133;
ld.global.nc.u32 %r204, [%rd66+16];

dp4a.s32.s32 %r198, %r199, %r200, %r190;


dp4a.s32.s32 %r202, %r203, %r204, %r198;

mov.u32 %r206, %r794;

prmt.b32 %r206, %r206, 0, 0xba98;

xor.b32 %r1134, %r206, %r924;
or.b32 %r1135, %r1134, -2139062144;
and.b32 %r1136, %r206, 2139062143;
sub.s32 %r1137, %r1135, %r1136;
xor.b32 %r211, %r942, %r1137;
ld.global.nc.u32 %r212, [%rd66+20];
mov.u32 %r208, %r796;

prmt.b32 %r208, %r208, 0, 0xba98;

xor.b32 %r1138, %r208, %r925;
or.b32 %r1139, %r1138, -2139062144;
and.b32 %r1140, %r208, 2139062143;
sub.s32 %r1141, %r1139, %r1140;
xor.b32 %r215, %r950, %r1141;
ld.global.nc.u32 %r216, [%rd66+24];

dp4a.s32.s32 %r210, %r211, %r212, %r850;


dp4a.s32.s32 %r214, %r215, %r216, %r210;

mov.u32 %r218, %r806;

prmt.b32 %r218, %r218, 0, 0xba98;

xor.b32 %r1142, %r218, %r952;
or.b32 %r1143, %r1142, -2139062144;
and.b32 %r1144, %r218, 2139062143;
sub.s32 %r1145, %r1143, %r1144;
xor.b32 %r223, %r970, %r1145;
ld.global.nc.u32 %r224, [%rd66+28];
mov.u32 %r220, %r808;

prmt.b32 %r220, %r220, 0, 0xba98;

xor.b32 %r1146, %r220, %r953;
or.b32 %r1147, %r1146, -2139062144;
and.b32 %r1148, %r220, 2139062143;
sub.s32 %r1149, %r1147, %r1148;
xor.b32 %r227, %r978, %r1149;
ld.global.nc.u32 %r228, [%rd66+32];

dp4a.s32.s32 %r222, %r223, %r224, %r214;


dp4a.s32.s32 %r226, %r227, %r228, %r222;

mul.lo.s32 %r1150, %r202, %r980;
mad.lo.s32 %r1151, %r226, %r981, %r1150;
add.s32 %r1152, %r226, %r202;
shr.u32 %r1153, %r1152, 31;
add.s32 %r1154, %r1152, %r1153;
shr.s32 %r1155, %r1154, 1;
add.s32 %r1156, %r1151, %r1155;
shr.s32 %r1157, %r1156, 31;
shr.u32 %r1158, %r1157, 30;
add.s32 %r1159, %r1156, %r1158;
shr.s32 %r1160, %r1159, 2;

{ cvt.f32.f16 %f101, %rs1;}


ld.global.nc.u32 %r230, [%rd66];

{.reg .f16 low,high;
mov.b32 {low,high},%r230;
cvt.f32.f16 %f102, low;}


mul.ftz.f32 %f133, %f101, %f102;
cvt.rn.f32.s32 %f134, %r1160;
fma.rn.ftz.f32 %f366, %f133, %f134, %f366;
mov.u32 %r231, %r819;

prmt.b32 %r231, %r231, 0, 0xba98;

xor.b32 %r1161, %r231, %r993;
or.b32 %r1162, %r1161, -2139062144;
and.b32 %r1163, %r231, 2139062143;
sub.s32 %r1164, %r1162, %r1163;
xor.b32 %r236, %r1011, %r1164;
mov.u32 %r233, %r821;

prmt.b32 %r233, %r233, 0, 0xba98;

xor.b32 %r1165, %r233, %r994;
or.b32 %r1166, %r1165, -2139062144;
and.b32 %r1167, %r233, 2139062143;
sub.s32 %r1168, %r1166, %r1167;
xor.b32 %r240, %r1019, %r1168;

dp4a.s32.s32 %r235, %r236, %r188, %r850;


dp4a.s32.s32 %r239, %r240, %r192, %r235;

mov.u32 %r243, %r831;

prmt.b32 %r243, %r243, 0, 0xba98;

xor.b32 %r1169, %r243, %r1021;
or.b32 %r1170, %r1169, -2139062144;
and.b32 %r1171, %r243, 2139062143;
sub.s32 %r1172, %r1170, %r1171;
xor.b32 %r248, %r1039, %r1172;
mov.u32 %r245, %r833;

prmt.b32 %r245, %r245, 0, 0xba98;

xor.b32 %r1173, %r245, %r1022;
or.b32 %r1174, %r1173, -2139062144;
and.b32 %r1175, %r245, 2139062143;
sub.s32 %r1176, %r1174, %r1175;
xor.b32 %r252, %r1047, %r1176;

dp4a.s32.s32 %r247, %r248, %r200, %r239;


dp4a.s32.s32 %r251, %r252, %r204, %r247;

mov.u32 %r255, %r843;

prmt.b32 %r255, %r255, 0, 0xba98;

xor.b32 %r1177, %r255, %r1049;
or.b32 %r1178, %r1177, -2139062144;
and.b32 %r1179, %r255, 2139062143;
sub.s32 %r1180, %r1178, %r1179;
xor.b32 %r260, %r1067, %r1180;
mov.u32 %r257, %r845;

prmt.b32 %r257, %r257, 0, 0xba98;

xor.b32 %r1181, %r257, %r1050;
or.b32 %r1182, %r1181, -2139062144;
and.b32 %r1183, %r257, 2139062143;
sub.s32 %r1184, %r1182, %r1183;
xor.b32 %r264, %r1075, %r1184;

dp4a.s32.s32 %r259, %r260, %r212, %r850;


dp4a.s32.s32 %r263, %r264, %r216, %r259;

mov.u32 %r267, %r855;

prmt.b32 %r267, %r267, 0, 0xba98;

xor.b32 %r1185, %r267, %r1077;
or.b32 %r1186, %r1185, -2139062144;
and.b32 %r1187, %r267, 2139062143;
sub.s32 %r1188, %r1186, %r1187;
xor.b32 %r272, %r1095, %r1188;
mov.u32 %r269, %r857;

prmt.b32 %r269, %r269, 0, 0xba98;

xor.b32 %r1189, %r269, %r1078;
or.b32 %r1190, %r1189, -2139062144;
and.b32 %r1191, %r269, 2139062143;
sub.s32 %r1192, %r1190, %r1191;
xor.b32 %r276, %r1103, %r1192;

dp4a.s32.s32 %r271, %r272, %r224, %r263;


dp4a.s32.s32 %r275, %r276, %r228, %r271;

mul.lo.s32 %r1193, %r251, %r1105;
mad.lo.s32 %r1194, %r275, %r1106, %r1193;
add.s32 %r1195, %r275, %r251;
shr.u32 %r1196, %r1195, 31;
add.s32 %r1197, %r1195, %r1196;
shr.s32 %r1198, %r1197, 1;
add.s32 %r1199, %r1194, %r1198;
shr.s32 %r1200, %r1199, 31;
shr.u32 %r1201, %r1200, 30;
add.s32 %r1202, %r1199, %r1201;
shr.s32 %r1203, %r1202, 2;

{ cvt.f32.f16 %f103, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r230;
cvt.f32.f16 %f104, low;}


mul.ftz.f32 %f135, %f103, %f104;
cvt.rn.f32.s32 %f136, %r1203;
fma.rn.ftz.f32 %f365, %f135, %f136, %f365;
cvt.u64.u32 %rd67, %r2171;
add.s64 %rd68, %rd1, %rd67;
add.s64 %rd69, %rd68, %rd3;
mul.lo.s64 %rd70, %rd69, 36;
mov.u32 %r280, %r770;

prmt.b32 %r280, %r280, 0, 0xba98;

xor.b32 %r1204, %r280, %r868;
or.b32 %r1205, %r1204, -2139062144;
and.b32 %r1206, %r280, 2139062143;
sub.s32 %r1207, %r1205, %r1206;
xor.b32 %r285, %r886, %r1207;
add.s64 %rd71, %rd10, %rd70;
ld.global.nc.u32 %r286, [%rd71+4];
mov.u32 %r282, %r772;

prmt.b32 %r282, %r282, 0, 0xba98;

xor.b32 %r1208, %r282, %r869;
or.b32 %r1209, %r1208, -2139062144;
and.b32 %r1210, %r282, 2139062143;
sub.s32 %r1211, %r1209, %r1210;
xor.b32 %r289, %r894, %r1211;
ld.global.nc.u32 %r290, [%rd71+8];

dp4a.s32.s32 %r284, %r285, %r286, %r850;


dp4a.s32.s32 %r288, %r289, %r290, %r284;

mov.u32 %r292, %r782;

prmt.b32 %r292, %r292, 0, 0xba98;

xor.b32 %r1212, %r292, %r896;
or.b32 %r1213, %r1212, -2139062144;
and.b32 %r1214, %r292, 2139062143;
sub.s32 %r1215, %r1213, %r1214;
xor.b32 %r297, %r914, %r1215;
ld.global.nc.u32 %r298, [%rd71+12];
mov.u32 %r294, %r784;

prmt.b32 %r294, %r294, 0, 0xba98;

xor.b32 %r1216, %r294, %r897;
or.b32 %r1217, %r1216, -2139062144;
and.b32 %r1218, %r294, 2139062143;
sub.s32 %r1219, %r1217, %r1218;
xor.b32 %r301, %r922, %r1219;
ld.global.nc.u32 %r302, [%rd71+16];

dp4a.s32.s32 %r296, %r297, %r298, %r288;


dp4a.s32.s32 %r300, %r301, %r302, %r296;

mov.u32 %r304, %r794;

prmt.b32 %r304, %r304, 0, 0xba98;

xor.b32 %r1220, %r304, %r924;
or.b32 %r1221, %r1220, -2139062144;
and.b32 %r1222, %r304, 2139062143;
sub.s32 %r1223, %r1221, %r1222;
xor.b32 %r309, %r942, %r1223;
ld.global.nc.u32 %r310, [%rd71+20];
mov.u32 %r306, %r796;

prmt.b32 %r306, %r306, 0, 0xba98;

xor.b32 %r1224, %r306, %r925;
or.b32 %r1225, %r1224, -2139062144;
and.b32 %r1226, %r306, 2139062143;
sub.s32 %r1227, %r1225, %r1226;
xor.b32 %r313, %r950, %r1227;
ld.global.nc.u32 %r314, [%rd71+24];

dp4a.s32.s32 %r308, %r309, %r310, %r850;


dp4a.s32.s32 %r312, %r313, %r314, %r308;

mov.u32 %r316, %r806;

prmt.b32 %r316, %r316, 0, 0xba98;

xor.b32 %r1228, %r316, %r952;
or.b32 %r1229, %r1228, -2139062144;
and.b32 %r1230, %r316, 2139062143;
sub.s32 %r1231, %r1229, %r1230;
xor.b32 %r321, %r970, %r1231;
ld.global.nc.u32 %r322, [%rd71+28];
mov.u32 %r318, %r808;

prmt.b32 %r318, %r318, 0, 0xba98;

xor.b32 %r1232, %r318, %r953;
or.b32 %r1233, %r1232, -2139062144;
and.b32 %r1234, %r318, 2139062143;
sub.s32 %r1235, %r1233, %r1234;
xor.b32 %r325, %r978, %r1235;
ld.global.nc.u32 %r326, [%rd71+32];

dp4a.s32.s32 %r320, %r321, %r322, %r312;


dp4a.s32.s32 %r324, %r325, %r326, %r320;

mul.lo.s32 %r1236, %r300, %r980;
mad.lo.s32 %r1237, %r324, %r981, %r1236;
add.s32 %r1238, %r324, %r300;
shr.u32 %r1239, %r1238, 31;
add.s32 %r1240, %r1238, %r1239;
shr.s32 %r1241, %r1240, 1;
add.s32 %r1242, %r1237, %r1241;
shr.s32 %r1243, %r1242, 31;
shr.u32 %r1244, %r1243, 30;
add.s32 %r1245, %r1242, %r1244;
shr.s32 %r1246, %r1245, 2;

{ cvt.f32.f16 %f105, %rs1;}


ld.global.nc.u32 %r328, [%rd71];

{.reg .f16 low,high;
mov.b32 {low,high},%r328;
cvt.f32.f16 %f106, low;}


mul.ftz.f32 %f137, %f105, %f106;
cvt.rn.f32.s32 %f138, %r1246;
fma.rn.ftz.f32 %f364, %f137, %f138, %f364;
mov.u32 %r329, %r819;

prmt.b32 %r329, %r329, 0, 0xba98;

xor.b32 %r1247, %r329, %r993;
or.b32 %r1248, %r1247, -2139062144;
and.b32 %r1249, %r329, 2139062143;
sub.s32 %r1250, %r1248, %r1249;
xor.b32 %r334, %r1011, %r1250;
mov.u32 %r331, %r821;

prmt.b32 %r331, %r331, 0, 0xba98;

xor.b32 %r1251, %r331, %r994;
or.b32 %r1252, %r1251, -2139062144;
and.b32 %r1253, %r331, 2139062143;
sub.s32 %r1254, %r1252, %r1253;
xor.b32 %r338, %r1019, %r1254;

dp4a.s32.s32 %r333, %r334, %r286, %r850;


dp4a.s32.s32 %r337, %r338, %r290, %r333;

mov.u32 %r341, %r831;

prmt.b32 %r341, %r341, 0, 0xba98;

xor.b32 %r1255, %r341, %r1021;
or.b32 %r1256, %r1255, -2139062144;
and.b32 %r1257, %r341, 2139062143;
sub.s32 %r1258, %r1256, %r1257;
xor.b32 %r346, %r1039, %r1258;
mov.u32 %r343, %r833;

prmt.b32 %r343, %r343, 0, 0xba98;

xor.b32 %r1259, %r343, %r1022;
or.b32 %r1260, %r1259, -2139062144;
and.b32 %r1261, %r343, 2139062143;
sub.s32 %r1262, %r1260, %r1261;
xor.b32 %r350, %r1047, %r1262;

dp4a.s32.s32 %r345, %r346, %r298, %r337;


dp4a.s32.s32 %r349, %r350, %r302, %r345;

mov.u32 %r353, %r843;

prmt.b32 %r353, %r353, 0, 0xba98;

xor.b32 %r1263, %r353, %r1049;
or.b32 %r1264, %r1263, -2139062144;
and.b32 %r1265, %r353, 2139062143;
sub.s32 %r1266, %r1264, %r1265;
xor.b32 %r358, %r1067, %r1266;
mov.u32 %r355, %r845;

prmt.b32 %r355, %r355, 0, 0xba98;

xor.b32 %r1267, %r355, %r1050;
or.b32 %r1268, %r1267, -2139062144;
and.b32 %r1269, %r355, 2139062143;
sub.s32 %r1270, %r1268, %r1269;
xor.b32 %r362, %r1075, %r1270;

dp4a.s32.s32 %r357, %r358, %r310, %r850;


dp4a.s32.s32 %r361, %r362, %r314, %r357;

mov.u32 %r365, %r855;

prmt.b32 %r365, %r365, 0, 0xba98;

xor.b32 %r1271, %r365, %r1077;
or.b32 %r1272, %r1271, -2139062144;
and.b32 %r1273, %r365, 2139062143;
sub.s32 %r1274, %r1272, %r1273;
xor.b32 %r370, %r1095, %r1274;
mov.u32 %r367, %r857;

prmt.b32 %r367, %r367, 0, 0xba98;

xor.b32 %r1275, %r367, %r1078;
or.b32 %r1276, %r1275, -2139062144;
and.b32 %r1277, %r367, 2139062143;
sub.s32 %r1278, %r1276, %r1277;
xor.b32 %r374, %r1103, %r1278;

dp4a.s32.s32 %r369, %r370, %r322, %r361;


dp4a.s32.s32 %r373, %r374, %r326, %r369;

mul.lo.s32 %r1279, %r349, %r1105;
mad.lo.s32 %r1280, %r373, %r1106, %r1279;
add.s32 %r1281, %r373, %r349;
shr.u32 %r1282, %r1281, 31;
add.s32 %r1283, %r1281, %r1282;
shr.s32 %r1284, %r1283, 1;
add.s32 %r1285, %r1280, %r1284;
shr.s32 %r1286, %r1285, 31;
shr.u32 %r1287, %r1286, 30;
add.s32 %r1288, %r1285, %r1287;
shr.s32 %r1289, %r1288, 2;

{ cvt.f32.f16 %f107, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r328;
cvt.f32.f16 %f108, low;}


mul.ftz.f32 %f139, %f107, %f108;
cvt.rn.f32.s32 %f140, %r1289;
fma.rn.ftz.f32 %f363, %f139, %f140, %f363;
cvt.u64.u32 %rd72, %r2170;
add.s64 %rd73, %rd1, %rd72;
add.s64 %rd74, %rd73, %rd3;
mul.lo.s64 %rd75, %rd74, 36;
mov.u32 %r378, %r770;

prmt.b32 %r378, %r378, 0, 0xba98;

xor.b32 %r1290, %r378, %r868;
or.b32 %r1291, %r1290, -2139062144;
and.b32 %r1292, %r378, 2139062143;
sub.s32 %r1293, %r1291, %r1292;
xor.b32 %r383, %r886, %r1293;
add.s64 %rd76, %rd10, %rd75;
ld.global.nc.u32 %r384, [%rd76+4];
mov.u32 %r380, %r772;

prmt.b32 %r380, %r380, 0, 0xba98;

xor.b32 %r1294, %r380, %r869;
or.b32 %r1295, %r1294, -2139062144;
and.b32 %r1296, %r380, 2139062143;
sub.s32 %r1297, %r1295, %r1296;
xor.b32 %r387, %r894, %r1297;
ld.global.nc.u32 %r388, [%rd76+8];

dp4a.s32.s32 %r382, %r383, %r384, %r850;


dp4a.s32.s32 %r386, %r387, %r388, %r382;

mov.u32 %r390, %r782;

prmt.b32 %r390, %r390, 0, 0xba98;

xor.b32 %r1298, %r390, %r896;
or.b32 %r1299, %r1298, -2139062144;
and.b32 %r1300, %r390, 2139062143;
sub.s32 %r1301, %r1299, %r1300;
xor.b32 %r395, %r914, %r1301;
ld.global.nc.u32 %r396, [%rd76+12];
mov.u32 %r392, %r784;

prmt.b32 %r392, %r392, 0, 0xba98;

xor.b32 %r1302, %r392, %r897;
or.b32 %r1303, %r1302, -2139062144;
and.b32 %r1304, %r392, 2139062143;
sub.s32 %r1305, %r1303, %r1304;
xor.b32 %r399, %r922, %r1305;
ld.global.nc.u32 %r400, [%rd76+16];

dp4a.s32.s32 %r394, %r395, %r396, %r386;


dp4a.s32.s32 %r398, %r399, %r400, %r394;

mov.u32 %r402, %r794;

prmt.b32 %r402, %r402, 0, 0xba98;

xor.b32 %r1306, %r402, %r924;
or.b32 %r1307, %r1306, -2139062144;
and.b32 %r1308, %r402, 2139062143;
sub.s32 %r1309, %r1307, %r1308;
xor.b32 %r407, %r942, %r1309;
ld.global.nc.u32 %r408, [%rd76+20];
mov.u32 %r404, %r796;

prmt.b32 %r404, %r404, 0, 0xba98;

xor.b32 %r1310, %r404, %r925;
or.b32 %r1311, %r1310, -2139062144;
and.b32 %r1312, %r404, 2139062143;
sub.s32 %r1313, %r1311, %r1312;
xor.b32 %r411, %r950, %r1313;
ld.global.nc.u32 %r412, [%rd76+24];

dp4a.s32.s32 %r406, %r407, %r408, %r850;


dp4a.s32.s32 %r410, %r411, %r412, %r406;

mov.u32 %r414, %r806;

prmt.b32 %r414, %r414, 0, 0xba98;

xor.b32 %r1314, %r414, %r952;
or.b32 %r1315, %r1314, -2139062144;
and.b32 %r1316, %r414, 2139062143;
sub.s32 %r1317, %r1315, %r1316;
xor.b32 %r419, %r970, %r1317;
ld.global.nc.u32 %r420, [%rd76+28];
mov.u32 %r416, %r808;

prmt.b32 %r416, %r416, 0, 0xba98;

xor.b32 %r1318, %r416, %r953;
or.b32 %r1319, %r1318, -2139062144;
and.b32 %r1320, %r416, 2139062143;
sub.s32 %r1321, %r1319, %r1320;
xor.b32 %r423, %r978, %r1321;
ld.global.nc.u32 %r424, [%rd76+32];

dp4a.s32.s32 %r418, %r419, %r420, %r410;


dp4a.s32.s32 %r422, %r423, %r424, %r418;

mul.lo.s32 %r1322, %r398, %r980;
mad.lo.s32 %r1323, %r422, %r981, %r1322;
add.s32 %r1324, %r422, %r398;
shr.u32 %r1325, %r1324, 31;
add.s32 %r1326, %r1324, %r1325;
shr.s32 %r1327, %r1326, 1;
add.s32 %r1328, %r1323, %r1327;
shr.s32 %r1329, %r1328, 31;
shr.u32 %r1330, %r1329, 30;
add.s32 %r1331, %r1328, %r1330;
shr.s32 %r1332, %r1331, 2;

{ cvt.f32.f16 %f109, %rs1;}


ld.global.nc.u32 %r426, [%rd76];

{.reg .f16 low,high;
mov.b32 {low,high},%r426;
cvt.f32.f16 %f110, low;}


mul.ftz.f32 %f141, %f109, %f110;
cvt.rn.f32.s32 %f142, %r1332;
fma.rn.ftz.f32 %f362, %f141, %f142, %f362;
mov.u32 %r427, %r819;

prmt.b32 %r427, %r427, 0, 0xba98;

xor.b32 %r1333, %r427, %r993;
or.b32 %r1334, %r1333, -2139062144;
and.b32 %r1335, %r427, 2139062143;
sub.s32 %r1336, %r1334, %r1335;
xor.b32 %r432, %r1011, %r1336;
mov.u32 %r429, %r821;

prmt.b32 %r429, %r429, 0, 0xba98;

xor.b32 %r1337, %r429, %r994;
or.b32 %r1338, %r1337, -2139062144;
and.b32 %r1339, %r429, 2139062143;
sub.s32 %r1340, %r1338, %r1339;
xor.b32 %r436, %r1019, %r1340;

dp4a.s32.s32 %r431, %r432, %r384, %r850;


dp4a.s32.s32 %r435, %r436, %r388, %r431;

mov.u32 %r439, %r831;

prmt.b32 %r439, %r439, 0, 0xba98;

xor.b32 %r1341, %r439, %r1021;
or.b32 %r1342, %r1341, -2139062144;
and.b32 %r1343, %r439, 2139062143;
sub.s32 %r1344, %r1342, %r1343;
xor.b32 %r444, %r1039, %r1344;
mov.u32 %r441, %r833;

prmt.b32 %r441, %r441, 0, 0xba98;

xor.b32 %r1345, %r441, %r1022;
or.b32 %r1346, %r1345, -2139062144;
and.b32 %r1347, %r441, 2139062143;
sub.s32 %r1348, %r1346, %r1347;
xor.b32 %r448, %r1047, %r1348;

dp4a.s32.s32 %r443, %r444, %r396, %r435;


dp4a.s32.s32 %r447, %r448, %r400, %r443;

mov.u32 %r451, %r843;

prmt.b32 %r451, %r451, 0, 0xba98;

xor.b32 %r1349, %r451, %r1049;
or.b32 %r1350, %r1349, -2139062144;
and.b32 %r1351, %r451, 2139062143;
sub.s32 %r1352, %r1350, %r1351;
xor.b32 %r456, %r1067, %r1352;
mov.u32 %r453, %r845;

prmt.b32 %r453, %r453, 0, 0xba98;

xor.b32 %r1353, %r453, %r1050;
or.b32 %r1354, %r1353, -2139062144;
and.b32 %r1355, %r453, 2139062143;
sub.s32 %r1356, %r1354, %r1355;
xor.b32 %r460, %r1075, %r1356;

dp4a.s32.s32 %r455, %r456, %r408, %r850;


dp4a.s32.s32 %r459, %r460, %r412, %r455;

mov.u32 %r463, %r855;

prmt.b32 %r463, %r463, 0, 0xba98;

xor.b32 %r1357, %r463, %r1077;
or.b32 %r1358, %r1357, -2139062144;
and.b32 %r1359, %r463, 2139062143;
sub.s32 %r1360, %r1358, %r1359;
xor.b32 %r468, %r1095, %r1360;
mov.u32 %r465, %r857;

prmt.b32 %r465, %r465, 0, 0xba98;

xor.b32 %r1361, %r465, %r1078;
or.b32 %r1362, %r1361, -2139062144;
and.b32 %r1363, %r465, 2139062143;
sub.s32 %r1364, %r1362, %r1363;
xor.b32 %r472, %r1103, %r1364;

dp4a.s32.s32 %r467, %r468, %r420, %r459;


dp4a.s32.s32 %r471, %r472, %r424, %r467;

mul.lo.s32 %r1365, %r447, %r1105;
mad.lo.s32 %r1366, %r471, %r1106, %r1365;
add.s32 %r1367, %r471, %r447;
shr.u32 %r1368, %r1367, 31;
add.s32 %r1369, %r1367, %r1368;
shr.s32 %r1370, %r1369, 1;
add.s32 %r1371, %r1366, %r1370;
shr.s32 %r1372, %r1371, 31;
shr.u32 %r1373, %r1372, 30;
add.s32 %r1374, %r1371, %r1373;
shr.s32 %r1375, %r1374, 2;

{ cvt.f32.f16 %f111, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r426;
cvt.f32.f16 %f112, low;}


mul.ftz.f32 %f143, %f111, %f112;
cvt.rn.f32.s32 %f144, %r1375;
fma.rn.ftz.f32 %f361, %f143, %f144, %f361;
cvt.u64.u32 %rd77, %r2169;
add.s64 %rd78, %rd1, %rd77;
add.s64 %rd79, %rd78, %rd3;
mul.lo.s64 %rd80, %rd79, 36;
mov.u32 %r476, %r770;

prmt.b32 %r476, %r476, 0, 0xba98;

xor.b32 %r1376, %r476, %r868;
or.b32 %r1377, %r1376, -2139062144;
and.b32 %r1378, %r476, 2139062143;
sub.s32 %r1379, %r1377, %r1378;
xor.b32 %r481, %r886, %r1379;
add.s64 %rd81, %rd10, %rd80;
ld.global.nc.u32 %r482, [%rd81+4];
mov.u32 %r478, %r772;

prmt.b32 %r478, %r478, 0, 0xba98;

xor.b32 %r1380, %r478, %r869;
or.b32 %r1381, %r1380, -2139062144;
and.b32 %r1382, %r478, 2139062143;
sub.s32 %r1383, %r1381, %r1382;
xor.b32 %r485, %r894, %r1383;
ld.global.nc.u32 %r486, [%rd81+8];

dp4a.s32.s32 %r480, %r481, %r482, %r850;


dp4a.s32.s32 %r484, %r485, %r486, %r480;

mov.u32 %r488, %r782;

prmt.b32 %r488, %r488, 0, 0xba98;

xor.b32 %r1384, %r488, %r896;
or.b32 %r1385, %r1384, -2139062144;
and.b32 %r1386, %r488, 2139062143;
sub.s32 %r1387, %r1385, %r1386;
xor.b32 %r493, %r914, %r1387;
ld.global.nc.u32 %r494, [%rd81+12];
mov.u32 %r490, %r784;

prmt.b32 %r490, %r490, 0, 0xba98;

xor.b32 %r1388, %r490, %r897;
or.b32 %r1389, %r1388, -2139062144;
and.b32 %r1390, %r490, 2139062143;
sub.s32 %r1391, %r1389, %r1390;
xor.b32 %r497, %r922, %r1391;
ld.global.nc.u32 %r498, [%rd81+16];

dp4a.s32.s32 %r492, %r493, %r494, %r484;


dp4a.s32.s32 %r496, %r497, %r498, %r492;

mov.u32 %r500, %r794;

prmt.b32 %r500, %r500, 0, 0xba98;

xor.b32 %r1392, %r500, %r924;
or.b32 %r1393, %r1392, -2139062144;
and.b32 %r1394, %r500, 2139062143;
sub.s32 %r1395, %r1393, %r1394;
xor.b32 %r505, %r942, %r1395;
ld.global.nc.u32 %r506, [%rd81+20];
mov.u32 %r502, %r796;

prmt.b32 %r502, %r502, 0, 0xba98;

xor.b32 %r1396, %r502, %r925;
or.b32 %r1397, %r1396, -2139062144;
and.b32 %r1398, %r502, 2139062143;
sub.s32 %r1399, %r1397, %r1398;
xor.b32 %r509, %r950, %r1399;
ld.global.nc.u32 %r510, [%rd81+24];

dp4a.s32.s32 %r504, %r505, %r506, %r850;


dp4a.s32.s32 %r508, %r509, %r510, %r504;

mov.u32 %r512, %r806;

prmt.b32 %r512, %r512, 0, 0xba98;

xor.b32 %r1400, %r512, %r952;
or.b32 %r1401, %r1400, -2139062144;
and.b32 %r1402, %r512, 2139062143;
sub.s32 %r1403, %r1401, %r1402;
xor.b32 %r517, %r970, %r1403;
ld.global.nc.u32 %r518, [%rd81+28];
mov.u32 %r514, %r808;

prmt.b32 %r514, %r514, 0, 0xba98;

xor.b32 %r1404, %r514, %r953;
or.b32 %r1405, %r1404, -2139062144;
and.b32 %r1406, %r514, 2139062143;
sub.s32 %r1407, %r1405, %r1406;
xor.b32 %r521, %r978, %r1407;
ld.global.nc.u32 %r522, [%rd81+32];

dp4a.s32.s32 %r516, %r517, %r518, %r508;


dp4a.s32.s32 %r520, %r521, %r522, %r516;

mul.lo.s32 %r1408, %r496, %r980;
mad.lo.s32 %r1409, %r520, %r981, %r1408;
add.s32 %r1410, %r520, %r496;
shr.u32 %r1411, %r1410, 31;
add.s32 %r1412, %r1410, %r1411;
shr.s32 %r1413, %r1412, 1;
add.s32 %r1414, %r1409, %r1413;
shr.s32 %r1415, %r1414, 31;
shr.u32 %r1416, %r1415, 30;
add.s32 %r1417, %r1414, %r1416;
shr.s32 %r1418, %r1417, 2;

{ cvt.f32.f16 %f113, %rs1;}


ld.global.nc.u32 %r524, [%rd81];

{.reg .f16 low,high;
mov.b32 {low,high},%r524;
cvt.f32.f16 %f114, low;}


mul.ftz.f32 %f145, %f113, %f114;
cvt.rn.f32.s32 %f146, %r1418;
fma.rn.ftz.f32 %f360, %f145, %f146, %f360;
mov.u32 %r525, %r819;

prmt.b32 %r525, %r525, 0, 0xba98;

xor.b32 %r1419, %r525, %r993;
or.b32 %r1420, %r1419, -2139062144;
and.b32 %r1421, %r525, 2139062143;
sub.s32 %r1422, %r1420, %r1421;
xor.b32 %r530, %r1011, %r1422;
mov.u32 %r527, %r821;

prmt.b32 %r527, %r527, 0, 0xba98;

xor.b32 %r1423, %r527, %r994;
or.b32 %r1424, %r1423, -2139062144;
and.b32 %r1425, %r527, 2139062143;
sub.s32 %r1426, %r1424, %r1425;
xor.b32 %r534, %r1019, %r1426;

dp4a.s32.s32 %r529, %r530, %r482, %r850;


dp4a.s32.s32 %r533, %r534, %r486, %r529;

mov.u32 %r537, %r831;

prmt.b32 %r537, %r537, 0, 0xba98;

xor.b32 %r1427, %r537, %r1021;
or.b32 %r1428, %r1427, -2139062144;
and.b32 %r1429, %r537, 2139062143;
sub.s32 %r1430, %r1428, %r1429;
xor.b32 %r542, %r1039, %r1430;
mov.u32 %r539, %r833;

prmt.b32 %r539, %r539, 0, 0xba98;

xor.b32 %r1431, %r539, %r1022;
or.b32 %r1432, %r1431, -2139062144;
and.b32 %r1433, %r539, 2139062143;
sub.s32 %r1434, %r1432, %r1433;
xor.b32 %r546, %r1047, %r1434;

dp4a.s32.s32 %r541, %r542, %r494, %r533;


dp4a.s32.s32 %r545, %r546, %r498, %r541;

mov.u32 %r549, %r843;

prmt.b32 %r549, %r549, 0, 0xba98;

xor.b32 %r1435, %r549, %r1049;
or.b32 %r1436, %r1435, -2139062144;
and.b32 %r1437, %r549, 2139062143;
sub.s32 %r1438, %r1436, %r1437;
xor.b32 %r554, %r1067, %r1438;
mov.u32 %r551, %r845;

prmt.b32 %r551, %r551, 0, 0xba98;

xor.b32 %r1439, %r551, %r1050;
or.b32 %r1440, %r1439, -2139062144;
and.b32 %r1441, %r551, 2139062143;
sub.s32 %r1442, %r1440, %r1441;
xor.b32 %r558, %r1075, %r1442;

dp4a.s32.s32 %r553, %r554, %r506, %r850;


dp4a.s32.s32 %r557, %r558, %r510, %r553;

mov.u32 %r561, %r855;

prmt.b32 %r561, %r561, 0, 0xba98;

xor.b32 %r1443, %r561, %r1077;
or.b32 %r1444, %r1443, -2139062144;
and.b32 %r1445, %r561, 2139062143;
sub.s32 %r1446, %r1444, %r1445;
xor.b32 %r566, %r1095, %r1446;
mov.u32 %r563, %r857;

prmt.b32 %r563, %r563, 0, 0xba98;

xor.b32 %r1447, %r563, %r1078;
or.b32 %r1448, %r1447, -2139062144;
and.b32 %r1449, %r563, 2139062143;
sub.s32 %r1450, %r1448, %r1449;
xor.b32 %r570, %r1103, %r1450;

dp4a.s32.s32 %r565, %r566, %r518, %r557;


dp4a.s32.s32 %r569, %r570, %r522, %r565;

mul.lo.s32 %r1451, %r545, %r1105;
mad.lo.s32 %r1452, %r569, %r1106, %r1451;
add.s32 %r1453, %r569, %r545;
shr.u32 %r1454, %r1453, 31;
add.s32 %r1455, %r1453, %r1454;
shr.s32 %r1456, %r1455, 1;
add.s32 %r1457, %r1452, %r1456;
shr.s32 %r1458, %r1457, 31;
shr.u32 %r1459, %r1458, 30;
add.s32 %r1460, %r1457, %r1459;
shr.s32 %r1461, %r1460, 2;

{ cvt.f32.f16 %f115, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r524;
cvt.f32.f16 %f116, low;}


mul.ftz.f32 %f147, %f115, %f116;
cvt.rn.f32.s32 %f148, %r1461;
fma.rn.ftz.f32 %f359, %f147, %f148, %f359;
cvt.u64.u32 %rd82, %r2168;
add.s64 %rd83, %rd1, %rd82;
add.s64 %rd84, %rd83, %rd3;
mul.lo.s64 %rd85, %rd84, 36;
mov.u32 %r574, %r770;

prmt.b32 %r574, %r574, 0, 0xba98;

xor.b32 %r1462, %r574, %r868;
or.b32 %r1463, %r1462, -2139062144;
and.b32 %r1464, %r574, 2139062143;
sub.s32 %r1465, %r1463, %r1464;
xor.b32 %r579, %r886, %r1465;
add.s64 %rd86, %rd10, %rd85;
ld.global.nc.u32 %r580, [%rd86+4];
mov.u32 %r576, %r772;

prmt.b32 %r576, %r576, 0, 0xba98;

xor.b32 %r1466, %r576, %r869;
or.b32 %r1467, %r1466, -2139062144;
and.b32 %r1468, %r576, 2139062143;
sub.s32 %r1469, %r1467, %r1468;
xor.b32 %r583, %r894, %r1469;
ld.global.nc.u32 %r584, [%rd86+8];

dp4a.s32.s32 %r578, %r579, %r580, %r850;


dp4a.s32.s32 %r582, %r583, %r584, %r578;

mov.u32 %r586, %r782;

prmt.b32 %r586, %r586, 0, 0xba98;

xor.b32 %r1470, %r586, %r896;
or.b32 %r1471, %r1470, -2139062144;
and.b32 %r1472, %r586, 2139062143;
sub.s32 %r1473, %r1471, %r1472;
xor.b32 %r591, %r914, %r1473;
ld.global.nc.u32 %r592, [%rd86+12];
mov.u32 %r588, %r784;

prmt.b32 %r588, %r588, 0, 0xba98;

xor.b32 %r1474, %r588, %r897;
or.b32 %r1475, %r1474, -2139062144;
and.b32 %r1476, %r588, 2139062143;
sub.s32 %r1477, %r1475, %r1476;
xor.b32 %r595, %r922, %r1477;
ld.global.nc.u32 %r596, [%rd86+16];

dp4a.s32.s32 %r590, %r591, %r592, %r582;


dp4a.s32.s32 %r594, %r595, %r596, %r590;

mov.u32 %r598, %r794;

prmt.b32 %r598, %r598, 0, 0xba98;

xor.b32 %r1478, %r598, %r924;
or.b32 %r1479, %r1478, -2139062144;
and.b32 %r1480, %r598, 2139062143;
sub.s32 %r1481, %r1479, %r1480;
xor.b32 %r603, %r942, %r1481;
ld.global.nc.u32 %r604, [%rd86+20];
mov.u32 %r600, %r796;

prmt.b32 %r600, %r600, 0, 0xba98;

xor.b32 %r1482, %r600, %r925;
or.b32 %r1483, %r1482, -2139062144;
and.b32 %r1484, %r600, 2139062143;
sub.s32 %r1485, %r1483, %r1484;
xor.b32 %r607, %r950, %r1485;
ld.global.nc.u32 %r608, [%rd86+24];

dp4a.s32.s32 %r602, %r603, %r604, %r850;


dp4a.s32.s32 %r606, %r607, %r608, %r602;

mov.u32 %r610, %r806;

prmt.b32 %r610, %r610, 0, 0xba98;

xor.b32 %r1486, %r610, %r952;
or.b32 %r1487, %r1486, -2139062144;
and.b32 %r1488, %r610, 2139062143;
sub.s32 %r1489, %r1487, %r1488;
xor.b32 %r615, %r970, %r1489;
ld.global.nc.u32 %r616, [%rd86+28];
mov.u32 %r612, %r808;

prmt.b32 %r612, %r612, 0, 0xba98;

xor.b32 %r1490, %r612, %r953;
or.b32 %r1491, %r1490, -2139062144;
and.b32 %r1492, %r612, 2139062143;
sub.s32 %r1493, %r1491, %r1492;
xor.b32 %r619, %r978, %r1493;
ld.global.nc.u32 %r620, [%rd86+32];

dp4a.s32.s32 %r614, %r615, %r616, %r606;


dp4a.s32.s32 %r618, %r619, %r620, %r614;

mul.lo.s32 %r1494, %r594, %r980;
mad.lo.s32 %r1495, %r618, %r981, %r1494;
add.s32 %r1496, %r618, %r594;
shr.u32 %r1497, %r1496, 31;
add.s32 %r1498, %r1496, %r1497;
shr.s32 %r1499, %r1498, 1;
add.s32 %r1500, %r1495, %r1499;
shr.s32 %r1501, %r1500, 31;
shr.u32 %r1502, %r1501, 30;
add.s32 %r1503, %r1500, %r1502;
shr.s32 %r1504, %r1503, 2;

{ cvt.f32.f16 %f117, %rs1;}


ld.global.nc.u32 %r622, [%rd86];

{.reg .f16 low,high;
mov.b32 {low,high},%r622;
cvt.f32.f16 %f118, low;}


mul.ftz.f32 %f149, %f117, %f118;
cvt.rn.f32.s32 %f150, %r1504;
fma.rn.ftz.f32 %f358, %f149, %f150, %f358;
mov.u32 %r623, %r819;

prmt.b32 %r623, %r623, 0, 0xba98;

xor.b32 %r1505, %r623, %r993;
or.b32 %r1506, %r1505, -2139062144;
and.b32 %r1507, %r623, 2139062143;
sub.s32 %r1508, %r1506, %r1507;
xor.b32 %r628, %r1011, %r1508;
mov.u32 %r625, %r821;

prmt.b32 %r625, %r625, 0, 0xba98;

xor.b32 %r1509, %r625, %r994;
or.b32 %r1510, %r1509, -2139062144;
and.b32 %r1511, %r625, 2139062143;
sub.s32 %r1512, %r1510, %r1511;
xor.b32 %r632, %r1019, %r1512;

dp4a.s32.s32 %r627, %r628, %r580, %r850;


dp4a.s32.s32 %r631, %r632, %r584, %r627;

mov.u32 %r635, %r831;

prmt.b32 %r635, %r635, 0, 0xba98;

xor.b32 %r1513, %r635, %r1021;
or.b32 %r1514, %r1513, -2139062144;
and.b32 %r1515, %r635, 2139062143;
sub.s32 %r1516, %r1514, %r1515;
xor.b32 %r640, %r1039, %r1516;
mov.u32 %r637, %r833;

prmt.b32 %r637, %r637, 0, 0xba98;

xor.b32 %r1517, %r637, %r1022;
or.b32 %r1518, %r1517, -2139062144;
and.b32 %r1519, %r637, 2139062143;
sub.s32 %r1520, %r1518, %r1519;
xor.b32 %r644, %r1047, %r1520;

dp4a.s32.s32 %r639, %r640, %r592, %r631;


dp4a.s32.s32 %r643, %r644, %r596, %r639;

mov.u32 %r647, %r843;

prmt.b32 %r647, %r647, 0, 0xba98;

xor.b32 %r1521, %r647, %r1049;
or.b32 %r1522, %r1521, -2139062144;
and.b32 %r1523, %r647, 2139062143;
sub.s32 %r1524, %r1522, %r1523;
xor.b32 %r652, %r1067, %r1524;
mov.u32 %r649, %r845;

prmt.b32 %r649, %r649, 0, 0xba98;

xor.b32 %r1525, %r649, %r1050;
or.b32 %r1526, %r1525, -2139062144;
and.b32 %r1527, %r649, 2139062143;
sub.s32 %r1528, %r1526, %r1527;
xor.b32 %r656, %r1075, %r1528;

dp4a.s32.s32 %r651, %r652, %r604, %r850;


dp4a.s32.s32 %r655, %r656, %r608, %r651;

mov.u32 %r659, %r855;

prmt.b32 %r659, %r659, 0, 0xba98;

xor.b32 %r1529, %r659, %r1077;
or.b32 %r1530, %r1529, -2139062144;
and.b32 %r1531, %r659, 2139062143;
sub.s32 %r1532, %r1530, %r1531;
xor.b32 %r664, %r1095, %r1532;
mov.u32 %r661, %r857;

prmt.b32 %r661, %r661, 0, 0xba98;

xor.b32 %r1533, %r661, %r1078;
or.b32 %r1534, %r1533, -2139062144;
and.b32 %r1535, %r661, 2139062143;
sub.s32 %r1536, %r1534, %r1535;
xor.b32 %r668, %r1103, %r1536;

dp4a.s32.s32 %r663, %r664, %r616, %r655;


dp4a.s32.s32 %r667, %r668, %r620, %r663;

mul.lo.s32 %r1537, %r643, %r1105;
mad.lo.s32 %r1538, %r667, %r1106, %r1537;
add.s32 %r1539, %r667, %r643;
shr.u32 %r1540, %r1539, 31;
add.s32 %r1541, %r1539, %r1540;
shr.s32 %r1542, %r1541, 1;
add.s32 %r1543, %r1538, %r1542;
shr.s32 %r1544, %r1543, 31;
shr.u32 %r1545, %r1544, 30;
add.s32 %r1546, %r1543, %r1545;
shr.s32 %r1547, %r1546, 2;

{ cvt.f32.f16 %f119, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r622;
cvt.f32.f16 %f120, low;}


mul.ftz.f32 %f151, %f119, %f120;
cvt.rn.f32.s32 %f152, %r1547;
fma.rn.ftz.f32 %f357, %f151, %f152, %f357;
cvt.u64.u32 %rd87, %r2167;
add.s64 %rd88, %rd1, %rd87;
add.s64 %rd89, %rd88, %rd3;
mul.lo.s64 %rd90, %rd89, 36;
mov.u32 %r672, %r770;

prmt.b32 %r672, %r672, 0, 0xba98;

xor.b32 %r1548, %r672, %r868;
or.b32 %r1549, %r1548, -2139062144;
and.b32 %r1550, %r672, 2139062143;
sub.s32 %r1551, %r1549, %r1550;
xor.b32 %r677, %r886, %r1551;
add.s64 %rd91, %rd10, %rd90;
ld.global.nc.u32 %r678, [%rd91+4];
mov.u32 %r674, %r772;

prmt.b32 %r674, %r674, 0, 0xba98;

xor.b32 %r1552, %r674, %r869;
or.b32 %r1553, %r1552, -2139062144;
and.b32 %r1554, %r674, 2139062143;
sub.s32 %r1555, %r1553, %r1554;
xor.b32 %r681, %r894, %r1555;
ld.global.nc.u32 %r682, [%rd91+8];

dp4a.s32.s32 %r676, %r677, %r678, %r850;


dp4a.s32.s32 %r680, %r681, %r682, %r676;

mov.u32 %r684, %r782;

prmt.b32 %r684, %r684, 0, 0xba98;

xor.b32 %r1556, %r684, %r896;
or.b32 %r1557, %r1556, -2139062144;
and.b32 %r1558, %r684, 2139062143;
sub.s32 %r1559, %r1557, %r1558;
xor.b32 %r689, %r914, %r1559;
ld.global.nc.u32 %r690, [%rd91+12];
mov.u32 %r686, %r784;

prmt.b32 %r686, %r686, 0, 0xba98;

xor.b32 %r1560, %r686, %r897;
or.b32 %r1561, %r1560, -2139062144;
and.b32 %r1562, %r686, 2139062143;
sub.s32 %r1563, %r1561, %r1562;
xor.b32 %r693, %r922, %r1563;
ld.global.nc.u32 %r694, [%rd91+16];

dp4a.s32.s32 %r688, %r689, %r690, %r680;


dp4a.s32.s32 %r692, %r693, %r694, %r688;

mov.u32 %r696, %r794;

prmt.b32 %r696, %r696, 0, 0xba98;

xor.b32 %r1564, %r696, %r924;
or.b32 %r1565, %r1564, -2139062144;
and.b32 %r1566, %r696, 2139062143;
sub.s32 %r1567, %r1565, %r1566;
xor.b32 %r701, %r942, %r1567;
ld.global.nc.u32 %r702, [%rd91+20];
mov.u32 %r698, %r796;

prmt.b32 %r698, %r698, 0, 0xba98;

xor.b32 %r1568, %r698, %r925;
or.b32 %r1569, %r1568, -2139062144;
and.b32 %r1570, %r698, 2139062143;
sub.s32 %r1571, %r1569, %r1570;
xor.b32 %r705, %r950, %r1571;
ld.global.nc.u32 %r706, [%rd91+24];

dp4a.s32.s32 %r700, %r701, %r702, %r850;


dp4a.s32.s32 %r704, %r705, %r706, %r700;

mov.u32 %r708, %r806;

prmt.b32 %r708, %r708, 0, 0xba98;

xor.b32 %r1572, %r708, %r952;
or.b32 %r1573, %r1572, -2139062144;
and.b32 %r1574, %r708, 2139062143;
sub.s32 %r1575, %r1573, %r1574;
xor.b32 %r713, %r970, %r1575;
ld.global.nc.u32 %r714, [%rd91+28];
mov.u32 %r710, %r808;

prmt.b32 %r710, %r710, 0, 0xba98;

xor.b32 %r1576, %r710, %r953;
or.b32 %r1577, %r1576, -2139062144;
and.b32 %r1578, %r710, 2139062143;
sub.s32 %r1579, %r1577, %r1578;
xor.b32 %r717, %r978, %r1579;
ld.global.nc.u32 %r718, [%rd91+32];

dp4a.s32.s32 %r712, %r713, %r714, %r704;


dp4a.s32.s32 %r716, %r717, %r718, %r712;

mul.lo.s32 %r1580, %r692, %r980;
mad.lo.s32 %r1581, %r716, %r981, %r1580;
add.s32 %r1582, %r716, %r692;
shr.u32 %r1583, %r1582, 31;
add.s32 %r1584, %r1582, %r1583;
shr.s32 %r1585, %r1584, 1;
add.s32 %r1586, %r1581, %r1585;
shr.s32 %r1587, %r1586, 31;
shr.u32 %r1588, %r1587, 30;
add.s32 %r1589, %r1586, %r1588;
shr.s32 %r1590, %r1589, 2;

{ cvt.f32.f16 %f121, %rs1;}


ld.global.nc.u32 %r720, [%rd91];

{.reg .f16 low,high;
mov.b32 {low,high},%r720;
cvt.f32.f16 %f122, low;}


mul.ftz.f32 %f153, %f121, %f122;
cvt.rn.f32.s32 %f154, %r1590;
fma.rn.ftz.f32 %f356, %f153, %f154, %f356;
mov.u32 %r721, %r819;

prmt.b32 %r721, %r721, 0, 0xba98;

xor.b32 %r1591, %r721, %r993;
or.b32 %r1592, %r1591, -2139062144;
and.b32 %r1593, %r721, 2139062143;
sub.s32 %r1594, %r1592, %r1593;
xor.b32 %r726, %r1011, %r1594;
mov.u32 %r723, %r821;

prmt.b32 %r723, %r723, 0, 0xba98;

xor.b32 %r1595, %r723, %r994;
or.b32 %r1596, %r1595, -2139062144;
and.b32 %r1597, %r723, 2139062143;
sub.s32 %r1598, %r1596, %r1597;
xor.b32 %r730, %r1019, %r1598;

dp4a.s32.s32 %r725, %r726, %r678, %r850;


dp4a.s32.s32 %r729, %r730, %r682, %r725;

mov.u32 %r733, %r831;

prmt.b32 %r733, %r733, 0, 0xba98;

xor.b32 %r1599, %r733, %r1021;
or.b32 %r1600, %r1599, -2139062144;
and.b32 %r1601, %r733, 2139062143;
sub.s32 %r1602, %r1600, %r1601;
xor.b32 %r738, %r1039, %r1602;
mov.u32 %r735, %r833;

prmt.b32 %r735, %r735, 0, 0xba98;

xor.b32 %r1603, %r735, %r1022;
or.b32 %r1604, %r1603, -2139062144;
and.b32 %r1605, %r735, 2139062143;
sub.s32 %r1606, %r1604, %r1605;
xor.b32 %r742, %r1047, %r1606;

dp4a.s32.s32 %r737, %r738, %r690, %r729;


dp4a.s32.s32 %r741, %r742, %r694, %r737;

mov.u32 %r745, %r843;

prmt.b32 %r745, %r745, 0, 0xba98;

xor.b32 %r1607, %r745, %r1049;
or.b32 %r1608, %r1607, -2139062144;
and.b32 %r1609, %r745, 2139062143;
sub.s32 %r1610, %r1608, %r1609;
xor.b32 %r750, %r1067, %r1610;
mov.u32 %r747, %r845;

prmt.b32 %r747, %r747, 0, 0xba98;

xor.b32 %r1611, %r747, %r1050;
or.b32 %r1612, %r1611, -2139062144;
and.b32 %r1613, %r747, 2139062143;
sub.s32 %r1614, %r1612, %r1613;
xor.b32 %r754, %r1075, %r1614;

dp4a.s32.s32 %r749, %r750, %r702, %r850;


dp4a.s32.s32 %r753, %r754, %r706, %r749;

mov.u32 %r757, %r855;

prmt.b32 %r757, %r757, 0, 0xba98;

xor.b32 %r1615, %r757, %r1077;
or.b32 %r1616, %r1615, -2139062144;
and.b32 %r1617, %r757, 2139062143;
sub.s32 %r1618, %r1616, %r1617;
xor.b32 %r762, %r1095, %r1618;
mov.u32 %r759, %r857;

prmt.b32 %r759, %r759, 0, 0xba98;

xor.b32 %r1619, %r759, %r1078;
or.b32 %r1620, %r1619, -2139062144;
and.b32 %r1621, %r759, 2139062143;
sub.s32 %r1622, %r1620, %r1621;
xor.b32 %r766, %r1103, %r1622;

dp4a.s32.s32 %r761, %r762, %r714, %r753;


dp4a.s32.s32 %r765, %r766, %r718, %r761;

mul.lo.s32 %r1623, %r741, %r1105;
mad.lo.s32 %r1624, %r765, %r1106, %r1623;
add.s32 %r1625, %r765, %r741;
shr.u32 %r1626, %r1625, 31;
add.s32 %r1627, %r1625, %r1626;
shr.s32 %r1628, %r1627, 1;
add.s32 %r1629, %r1624, %r1628;
shr.s32 %r1630, %r1629, 31;
shr.u32 %r1631, %r1630, 30;
add.s32 %r1632, %r1629, %r1631;
shr.s32 %r1633, %r1632, 2;

{ cvt.f32.f16 %f123, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r720;
cvt.f32.f16 %f124, low;}


mul.ftz.f32 %f155, %f123, %f124;
cvt.rn.f32.s32 %f156, %r1633;
fma.rn.ftz.f32 %f355, %f155, %f156, %f355;
cvt.u64.u32 %rd92, %r2166;
add.s64 %rd93, %rd1, %rd92;
add.s64 %rd94, %rd93, %rd3;
mul.lo.s64 %rd95, %rd94, 36;

prmt.b32 %r770, %r770, 0, 0xba98;

xor.b32 %r1634, %r770, %r868;
or.b32 %r1635, %r1634, -2139062144;
and.b32 %r1636, %r770, 2139062143;
sub.s32 %r1637, %r1635, %r1636;
xor.b32 %r775, %r886, %r1637;
add.s64 %rd96, %rd10, %rd95;
ld.global.nc.u32 %r776, [%rd96+4];

prmt.b32 %r772, %r772, 0, 0xba98;

xor.b32 %r1638, %r772, %r869;
or.b32 %r1639, %r1638, -2139062144;
and.b32 %r1640, %r772, 2139062143;
sub.s32 %r1641, %r1639, %r1640;
xor.b32 %r779, %r894, %r1641;
ld.global.nc.u32 %r780, [%rd96+8];

dp4a.s32.s32 %r774, %r775, %r776, %r850;


dp4a.s32.s32 %r778, %r779, %r780, %r774;


prmt.b32 %r782, %r782, 0, 0xba98;

xor.b32 %r1642, %r782, %r896;
or.b32 %r1643, %r1642, -2139062144;
and.b32 %r1644, %r782, 2139062143;
sub.s32 %r1645, %r1643, %r1644;
xor.b32 %r787, %r914, %r1645;
ld.global.nc.u32 %r788, [%rd96+12];

prmt.b32 %r784, %r784, 0, 0xba98;

xor.b32 %r1646, %r784, %r897;
or.b32 %r1647, %r1646, -2139062144;
and.b32 %r1648, %r784, 2139062143;
sub.s32 %r1649, %r1647, %r1648;
xor.b32 %r791, %r922, %r1649;
ld.global.nc.u32 %r792, [%rd96+16];

dp4a.s32.s32 %r786, %r787, %r788, %r778;


dp4a.s32.s32 %r790, %r791, %r792, %r786;


prmt.b32 %r794, %r794, 0, 0xba98;

xor.b32 %r1650, %r794, %r924;
or.b32 %r1651, %r1650, -2139062144;
and.b32 %r1652, %r794, 2139062143;
sub.s32 %r1653, %r1651, %r1652;
xor.b32 %r799, %r942, %r1653;
ld.global.nc.u32 %r800, [%rd96+20];

prmt.b32 %r796, %r796, 0, 0xba98;

xor.b32 %r1654, %r796, %r925;
or.b32 %r1655, %r1654, -2139062144;
and.b32 %r1656, %r796, 2139062143;
sub.s32 %r1657, %r1655, %r1656;
xor.b32 %r803, %r950, %r1657;
ld.global.nc.u32 %r804, [%rd96+24];

dp4a.s32.s32 %r798, %r799, %r800, %r850;


dp4a.s32.s32 %r802, %r803, %r804, %r798;


prmt.b32 %r806, %r806, 0, 0xba98;

xor.b32 %r1658, %r806, %r952;
or.b32 %r1659, %r1658, -2139062144;
and.b32 %r1660, %r806, 2139062143;
sub.s32 %r1661, %r1659, %r1660;
xor.b32 %r811, %r970, %r1661;
ld.global.nc.u32 %r812, [%rd96+28];

prmt.b32 %r808, %r808, 0, 0xba98;

xor.b32 %r1662, %r808, %r953;
or.b32 %r1663, %r1662, -2139062144;
and.b32 %r1664, %r808, 2139062143;
sub.s32 %r1665, %r1663, %r1664;
xor.b32 %r815, %r978, %r1665;
ld.global.nc.u32 %r816, [%rd96+32];

dp4a.s32.s32 %r810, %r811, %r812, %r802;


dp4a.s32.s32 %r814, %r815, %r816, %r810;

mul.lo.s32 %r1666, %r790, %r980;
mad.lo.s32 %r1667, %r814, %r981, %r1666;
add.s32 %r1668, %r814, %r790;
shr.u32 %r1669, %r1668, 31;
add.s32 %r1670, %r1668, %r1669;
shr.s32 %r1671, %r1670, 1;
add.s32 %r1672, %r1667, %r1671;
shr.s32 %r1673, %r1672, 31;
shr.u32 %r1674, %r1673, 30;
add.s32 %r1675, %r1672, %r1674;
shr.s32 %r1676, %r1675, 2;

{ cvt.f32.f16 %f125, %rs1;}


ld.global.nc.u32 %r818, [%rd96];

{.reg .f16 low,high;
mov.b32 {low,high},%r818;
cvt.f32.f16 %f126, low;}


mul.ftz.f32 %f157, %f125, %f126;
cvt.rn.f32.s32 %f158, %r1676;
fma.rn.ftz.f32 %f354, %f157, %f158, %f354;

prmt.b32 %r819, %r819, 0, 0xba98;

xor.b32 %r1677, %r819, %r993;
or.b32 %r1678, %r1677, -2139062144;
and.b32 %r1679, %r819, 2139062143;
sub.s32 %r1680, %r1678, %r1679;
xor.b32 %r824, %r1011, %r1680;

prmt.b32 %r821, %r821, 0, 0xba98;

xor.b32 %r1681, %r821, %r994;
or.b32 %r1682, %r1681, -2139062144;
and.b32 %r1683, %r821, 2139062143;
sub.s32 %r1684, %r1682, %r1683;
xor.b32 %r828, %r1019, %r1684;

dp4a.s32.s32 %r823, %r824, %r776, %r850;


dp4a.s32.s32 %r827, %r828, %r780, %r823;


prmt.b32 %r831, %r831, 0, 0xba98;

xor.b32 %r1685, %r831, %r1021;
or.b32 %r1686, %r1685, -2139062144;
and.b32 %r1687, %r831, 2139062143;
sub.s32 %r1688, %r1686, %r1687;
xor.b32 %r836, %r1039, %r1688;

prmt.b32 %r833, %r833, 0, 0xba98;

xor.b32 %r1689, %r833, %r1022;
or.b32 %r1690, %r1689, -2139062144;
and.b32 %r1691, %r833, 2139062143;
sub.s32 %r1692, %r1690, %r1691;
xor.b32 %r840, %r1047, %r1692;

dp4a.s32.s32 %r835, %r836, %r788, %r827;


dp4a.s32.s32 %r839, %r840, %r792, %r835;


prmt.b32 %r843, %r843, 0, 0xba98;

xor.b32 %r1693, %r843, %r1049;
or.b32 %r1694, %r1693, -2139062144;
and.b32 %r1695, %r843, 2139062143;
sub.s32 %r1696, %r1694, %r1695;
xor.b32 %r848, %r1067, %r1696;

prmt.b32 %r845, %r845, 0, 0xba98;

xor.b32 %r1697, %r845, %r1050;
or.b32 %r1698, %r1697, -2139062144;
and.b32 %r1699, %r845, 2139062143;
sub.s32 %r1700, %r1698, %r1699;
xor.b32 %r852, %r1075, %r1700;

dp4a.s32.s32 %r847, %r848, %r800, %r850;


dp4a.s32.s32 %r851, %r852, %r804, %r847;


prmt.b32 %r855, %r855, 0, 0xba98;

xor.b32 %r1701, %r855, %r1077;
or.b32 %r1702, %r1701, -2139062144;
and.b32 %r1703, %r855, 2139062143;
sub.s32 %r1704, %r1702, %r1703;
xor.b32 %r860, %r1095, %r1704;

prmt.b32 %r857, %r857, 0, 0xba98;

xor.b32 %r1705, %r857, %r1078;
or.b32 %r1706, %r1705, -2139062144;
and.b32 %r1707, %r857, 2139062143;
sub.s32 %r1708, %r1706, %r1707;
xor.b32 %r864, %r1103, %r1708;

dp4a.s32.s32 %r859, %r860, %r812, %r851;


dp4a.s32.s32 %r863, %r864, %r816, %r859;

mul.lo.s32 %r1709, %r839, %r1105;
mad.lo.s32 %r1710, %r863, %r1106, %r1709;
add.s32 %r1711, %r863, %r839;
shr.u32 %r1712, %r1711, 31;
add.s32 %r1713, %r1711, %r1712;
shr.s32 %r1714, %r1713, 1;
add.s32 %r1715, %r1710, %r1714;
shr.s32 %r1716, %r1715, 31;
shr.u32 %r1717, %r1716, 30;
add.s32 %r1718, %r1715, %r1717;
shr.s32 %r1719, %r1718, 2;

{ cvt.f32.f16 %f127, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r818;
cvt.f32.f16 %f128, low;}


mul.ftz.f32 %f159, %f127, %f128;
cvt.rn.f32.s32 %f160, %r1719;
fma.rn.ftz.f32 %f353, %f159, %f160, %f353;
add.s32 %r2171, %r2171, 64;
add.s32 %r2170, %r2170, 64;
add.s32 %r2169, %r2169, 64;
add.s32 %r2168, %r2168, 64;
add.s32 %r2167, %r2167, 64;
add.s32 %r2166, %r2166, 64;
add.s32 %r2165, %r2165, 64;
add.s32 %r2164, %r2164, 64;
add.s64 %rd166, %rd166, 592;
add.s32 %r2172, %r2172, 8;
setp.lt.s32 %p3, %r2172, %r54;
@%p3 bra $L__BB223_2;

$L__BB223_3:
setp.eq.s32 %p4, %r48, 0;
@%p4 bra $L__BB223_5;

shl.b32 %r1724, %r48, 11;
mov.u32 %r1725, _ZZ13mul_mat_vec_qIL9ggml_type17ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r1726, %r1725, %r1724;
shl.b32 %r1727, %r50, 2;
add.s32 %r1728, %r1726, %r1727;
st.shared.f32 [%r1728+-2048], %f368;
st.shared.f32 [%r1728+-1920], %f367;
st.shared.f32 [%r1728+-1792], %f366;
st.shared.f32 [%r1728+-1664], %f365;
st.shared.f32 [%r1728+-1536], %f364;
st.shared.f32 [%r1728+-1408], %f363;
st.shared.f32 [%r1728+-1280], %f362;
st.shared.f32 [%r1728+-1152], %f361;
st.shared.f32 [%r1728+-1024], %f360;
st.shared.f32 [%r1728+-896], %f359;
st.shared.f32 [%r1728+-768], %f358;
st.shared.f32 [%r1728+-640], %f357;
st.shared.f32 [%r1728+-512], %f356;
st.shared.f32 [%r1728+-384], %f355;
st.shared.f32 [%r1728+-256], %f354;
st.shared.f32 [%r1728+-128], %f353;

$L__BB223_5:
setp.ne.s32 %p5, %r48, 0;
bar.sync 0;
@%p5 bra $L__BB223_38;

mov.u32 %r1730, %ctaid.x;
shl.b32 %r1731, %r1730, 1;
mov.u32 %r1732, 1;
shl.b32 %r1734, %r50, 2;
mov.u32 %r1735, _ZZ13mul_mat_vec_qIL9ggml_type17ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r1736, %r1735, %r1734;
mov.u32 %r1737, 2;
ld.shared.f32 %f161, [%r1736];
add.ftz.f32 %f162, %f161, %f368;
mov.b32 %r1738, %f162;
mov.u32 %r1739, 31;
mov.u32 %r1740, 16;
mov.u32 %r1741, -1;
shfl.sync.bfly.b32 %r1742|%p6, %r1738, %r1740, %r1739, %r1741;
mov.b32 %f163, %r1742;
add.ftz.f32 %f164, %f162, %f163;
mov.b32 %r1743, %f164;
mov.u32 %r1744, 8;
shfl.sync.bfly.b32 %r1745|%p7, %r1743, %r1744, %r1739, %r1741;
mov.b32 %f165, %r1745;
add.ftz.f32 %f166, %f164, %f165;
mov.b32 %r1746, %f166;
mov.u32 %r1747, 4;
shfl.sync.bfly.b32 %r1748|%p8, %r1746, %r1747, %r1739, %r1741;
mov.b32 %f167, %r1748;
add.ftz.f32 %f168, %f166, %f167;
mov.b32 %r1749, %f168;
shfl.sync.bfly.b32 %r1750|%p9, %r1749, %r1737, %r1739, %r1741;
mov.b32 %f169, %r1750;
add.ftz.f32 %f170, %f168, %f169;
mov.b32 %r1751, %f170;
shfl.sync.bfly.b32 %r1752|%p10, %r1751, %r1732, %r1739, %r1741;
mov.b32 %f171, %r1752;
add.ftz.f32 %f49, %f170, %f171;
setp.ne.s32 %p11, %r50, 0;
setp.ge.u32 %p12, %r1731, %r35;
or.pred %p13, %p11, %p12;
@%p13 bra $L__BB223_8;

cvta.to.global.u64 %rd97, %rd18;
mov.u32 %r1753, %ctaid.y;
mad.lo.s32 %r1756, %r1753, %r41, %r1731;
mov.u32 %r1757, %ctaid.z;
mad.lo.s32 %r1758, %r1757, %r47, %r1756;
mul.wide.u32 %rd98, %r1758, 4;
add.s64 %rd99, %rd97, %rd98;
st.global.f32 [%rd99], %f49;

$L__BB223_8:
ld.shared.f32 %f172, [%r1736+128];
add.ftz.f32 %f173, %f172, %f367;
mov.b32 %r1764, %f173;
shfl.sync.bfly.b32 %r1768|%p14, %r1764, %r1740, %r1739, %r1741;
mov.b32 %f174, %r1768;
add.ftz.f32 %f175, %f173, %f174;
mov.b32 %r1769, %f175;
shfl.sync.bfly.b32 %r1771|%p15, %r1769, %r1744, %r1739, %r1741;
mov.b32 %f176, %r1771;
add.ftz.f32 %f177, %f175, %f176;
mov.b32 %r1772, %f177;
shfl.sync.bfly.b32 %r1774|%p16, %r1772, %r1747, %r1739, %r1741;
mov.b32 %f178, %r1774;
add.ftz.f32 %f179, %f177, %f178;
mov.b32 %r1775, %f179;
shfl.sync.bfly.b32 %r1776|%p17, %r1775, %r1737, %r1739, %r1741;
mov.b32 %f180, %r1776;
add.ftz.f32 %f181, %f179, %f180;
mov.b32 %r1777, %f181;
shfl.sync.bfly.b32 %r1779|%p18, %r1777, %r1732, %r1739, %r1741;
mov.b32 %f182, %r1779;
add.ftz.f32 %f50, %f181, %f182;
setp.eq.s32 %p19, %r50, 1;
or.b32 %r1782, %r1731, 1;
setp.lt.u32 %p20, %r1782, %r35;
and.pred %p1, %p19, %p20;
not.pred %p21, %p1;
@%p21 bra $L__BB223_10;

cvta.to.global.u64 %rd100, %rd18;
mov.u32 %r1783, %ctaid.y;
mad.lo.s32 %r1786, %r1783, %r41, %r1731;
mov.u32 %r1787, %ctaid.z;
mad.lo.s32 %r1788, %r1787, %r47, %r1786;
mul.wide.u32 %rd101, %r1788, 4;
add.s64 %rd102, %rd100, %rd101;
st.global.f32 [%rd102+4], %f50;

$L__BB223_10:
mov.u32 %r1793, 2;
ld.shared.f32 %f183, [%r1736+256];
add.ftz.f32 %f184, %f183, %f366;
mov.b32 %r1794, %f184;
mov.u32 %r1795, 31;
mov.u32 %r1796, 16;
mov.u32 %r1797, -1;
shfl.sync.bfly.b32 %r1798|%p22, %r1794, %r1796, %r1795, %r1797;
mov.b32 %f185, %r1798;
add.ftz.f32 %f186, %f184, %f185;
mov.b32 %r1799, %f186;
mov.u32 %r1800, 8;
shfl.sync.bfly.b32 %r1801|%p23, %r1799, %r1800, %r1795, %r1797;
mov.b32 %f187, %r1801;
add.ftz.f32 %f188, %f186, %f187;
mov.b32 %r1802, %f188;
mov.u32 %r1803, 4;
shfl.sync.bfly.b32 %r1804|%p24, %r1802, %r1803, %r1795, %r1797;
mov.b32 %f189, %r1804;
add.ftz.f32 %f190, %f188, %f189;
mov.b32 %r1805, %f190;
shfl.sync.bfly.b32 %r1806|%p25, %r1805, %r1793, %r1795, %r1797;
mov.b32 %f191, %r1806;
add.ftz.f32 %f192, %f190, %f191;
mov.b32 %r1807, %f192;
mov.u32 %r1808, 1;
shfl.sync.bfly.b32 %r1809|%p26, %r1807, %r1808, %r1795, %r1797;
mov.b32 %f193, %r1809;
add.ftz.f32 %f51, %f192, %f193;
@%p13 bra $L__BB223_12;

cvt.u64.u32 %rd103, %r35;
mov.u32 %r1812, %ctaid.y;
mad.lo.s32 %r1815, %r1812, %r41, %r1731;
mov.u32 %r1816, %ctaid.z;
mad.lo.s32 %r1817, %r1816, %r47, %r1815;
cvt.u64.u32 %rd104, %r1817;
add.s64 %rd105, %rd103, %rd104;
cvta.to.global.u64 %rd106, %rd18;
shl.b64 %rd107, %rd105, 2;
add.s64 %rd108, %rd106, %rd107;
st.global.f32 [%rd108], %f51;

$L__BB223_12:
ld.shared.f32 %f194, [%r1736+384];
add.ftz.f32 %f195, %f194, %f365;
mov.b32 %r1823, %f195;
shfl.sync.bfly.b32 %r1827|%p30, %r1823, %r1796, %r1795, %r1797;
mov.b32 %f196, %r1827;
add.ftz.f32 %f197, %f195, %f196;
mov.b32 %r1828, %f197;
shfl.sync.bfly.b32 %r1830|%p31, %r1828, %r1800, %r1795, %r1797;
mov.b32 %f198, %r1830;
add.ftz.f32 %f199, %f197, %f198;
mov.b32 %r1831, %f199;
shfl.sync.bfly.b32 %r1833|%p32, %r1831, %r1803, %r1795, %r1797;
mov.b32 %f200, %r1833;
add.ftz.f32 %f201, %f199, %f200;
mov.b32 %r1834, %f201;
shfl.sync.bfly.b32 %r1835|%p33, %r1834, %r1793, %r1795, %r1797;
mov.b32 %f202, %r1835;
add.ftz.f32 %f203, %f201, %f202;
mov.b32 %r1836, %f203;
shfl.sync.bfly.b32 %r1838|%p34, %r1836, %r1808, %r1795, %r1797;
mov.b32 %f204, %r1838;
add.ftz.f32 %f52, %f203, %f204;
@%p21 bra $L__BB223_14;

add.s32 %r1839, %r35, 1;
cvt.u64.u32 %rd109, %r1839;
mov.u32 %r1840, %ctaid.y;
mad.lo.s32 %r1843, %r1840, %r41, %r1731;
mov.u32 %r1844, %ctaid.z;
mad.lo.s32 %r1845, %r1844, %r47, %r1843;
cvt.u64.u32 %rd110, %r1845;
add.s64 %rd111, %rd109, %rd110;
cvta.to.global.u64 %rd112, %rd18;
shl.b64 %rd113, %rd111, 2;
add.s64 %rd114, %rd112, %rd113;
st.global.f32 [%rd114], %f52;

$L__BB223_14:
shl.b32 %r1846, %r35, 1;
mov.u32 %r1847, 1;
mov.u32 %r1852, 2;
ld.shared.f32 %f205, [%r1736+512];
add.ftz.f32 %f206, %f205, %f364;
mov.b32 %r1853, %f206;
mov.u32 %r1854, 31;
mov.u32 %r1855, 16;
mov.u32 %r1856, -1;
shfl.sync.bfly.b32 %r1857|%p36, %r1853, %r1855, %r1854, %r1856;
mov.b32 %f207, %r1857;
add.ftz.f32 %f208, %f206, %f207;
mov.b32 %r1858, %f208;
mov.u32 %r1859, 8;
shfl.sync.bfly.b32 %r1860|%p37, %r1858, %r1859, %r1854, %r1856;
mov.b32 %f209, %r1860;
add.ftz.f32 %f210, %f208, %f209;
mov.b32 %r1861, %f210;
mov.u32 %r1862, 4;
shfl.sync.bfly.b32 %r1863|%p38, %r1861, %r1862, %r1854, %r1856;
mov.b32 %f211, %r1863;
add.ftz.f32 %f212, %f210, %f211;
mov.b32 %r1864, %f212;
shfl.sync.bfly.b32 %r1865|%p39, %r1864, %r1852, %r1854, %r1856;
mov.b32 %f213, %r1865;
add.ftz.f32 %f214, %f212, %f213;
mov.b32 %r1866, %f214;
shfl.sync.bfly.b32 %r1867|%p40, %r1866, %r1847, %r1854, %r1856;
mov.b32 %f215, %r1867;
add.ftz.f32 %f53, %f214, %f215;
cvt.u64.u32 %rd115, %r1846;
mov.u32 %r1868, %ctaid.y;
mad.lo.s32 %r1871, %r1868, %r41, %r1731;
mov.u32 %r1872, %ctaid.z;
mad.lo.s32 %r1873, %r1872, %r47, %r1871;
cvt.u64.u32 %rd116, %r1873;
add.s64 %rd117, %rd115, %rd116;
cvta.to.global.u64 %rd118, %rd18;
shl.b64 %rd119, %rd117, 2;
add.s64 %rd13, %rd118, %rd119;
@%p13 bra $L__BB223_16;

st.global.f32 [%rd13], %f53;

$L__BB223_16:
ld.shared.f32 %f216, [%r1736+640];
add.ftz.f32 %f217, %f216, %f363;
mov.b32 %r1879, %f217;
shfl.sync.bfly.b32 %r1883|%p44, %r1879, %r1855, %r1854, %r1856;
mov.b32 %f218, %r1883;
add.ftz.f32 %f219, %f217, %f218;
mov.b32 %r1884, %f219;
shfl.sync.bfly.b32 %r1886|%p45, %r1884, %r1859, %r1854, %r1856;
mov.b32 %f220, %r1886;
add.ftz.f32 %f221, %f219, %f220;
mov.b32 %r1887, %f221;
shfl.sync.bfly.b32 %r1889|%p46, %r1887, %r1862, %r1854, %r1856;
mov.b32 %f222, %r1889;
add.ftz.f32 %f223, %f221, %f222;
mov.b32 %r1890, %f223;
shfl.sync.bfly.b32 %r1891|%p47, %r1890, %r1852, %r1854, %r1856;
mov.b32 %f224, %r1891;
add.ftz.f32 %f225, %f223, %f224;
mov.b32 %r1892, %f225;
shfl.sync.bfly.b32 %r1894|%p48, %r1892, %r1847, %r1854, %r1856;
mov.b32 %f226, %r1894;
add.ftz.f32 %f54, %f225, %f226;
@%p21 bra $L__BB223_18;

st.global.f32 [%rd13+4], %f54;

$L__BB223_18:
mul.lo.s32 %r29, %r35, 3;
mov.u32 %r1899, 2;
ld.shared.f32 %f227, [%r1736+768];
add.ftz.f32 %f228, %f227, %f362;
mov.b32 %r1900, %f228;
mov.u32 %r1901, 31;
mov.u32 %r1902, 16;
mov.u32 %r1903, -1;
shfl.sync.bfly.b32 %r1904|%p50, %r1900, %r1902, %r1901, %r1903;
mov.b32 %f229, %r1904;
add.ftz.f32 %f230, %f228, %f229;
mov.b32 %r1905, %f230;
mov.u32 %r1906, 8;
shfl.sync.bfly.b32 %r1907|%p51, %r1905, %r1906, %r1901, %r1903;
mov.b32 %f231, %r1907;
add.ftz.f32 %f232, %f230, %f231;
mov.b32 %r1908, %f232;
mov.u32 %r1909, 4;
shfl.sync.bfly.b32 %r1910|%p52, %r1908, %r1909, %r1901, %r1903;
mov.b32 %f233, %r1910;
add.ftz.f32 %f234, %f232, %f233;
mov.b32 %r1911, %f234;
shfl.sync.bfly.b32 %r1912|%p53, %r1911, %r1899, %r1901, %r1903;
mov.b32 %f235, %r1912;
add.ftz.f32 %f236, %f234, %f235;
mov.b32 %r1913, %f236;
mov.u32 %r1914, 1;
shfl.sync.bfly.b32 %r1915|%p54, %r1913, %r1914, %r1901, %r1903;
mov.b32 %f237, %r1915;
add.ftz.f32 %f55, %f236, %f237;
@%p13 bra $L__BB223_20;

cvt.u64.u32 %rd120, %r29;
add.s64 %rd122, %rd120, %rd116;
shl.b64 %rd124, %rd122, 2;
add.s64 %rd125, %rd118, %rd124;
st.global.f32 [%rd125], %f55;

$L__BB223_20:
ld.shared.f32 %f238, [%r1736+896];
add.ftz.f32 %f239, %f238, %f361;
mov.b32 %r1929, %f239;
shfl.sync.bfly.b32 %r1933|%p58, %r1929, %r1902, %r1901, %r1903;
mov.b32 %f240, %r1933;
add.ftz.f32 %f241, %f239, %f240;
mov.b32 %r1934, %f241;
shfl.sync.bfly.b32 %r1936|%p59, %r1934, %r1906, %r1901, %r1903;
mov.b32 %f242, %r1936;
add.ftz.f32 %f243, %f241, %f242;
mov.b32 %r1937, %f243;
shfl.sync.bfly.b32 %r1939|%p60, %r1937, %r1909, %r1901, %r1903;
mov.b32 %f244, %r1939;
add.ftz.f32 %f245, %f243, %f244;
mov.b32 %r1940, %f245;
shfl.sync.bfly.b32 %r1941|%p61, %r1940, %r1899, %r1901, %r1903;
mov.b32 %f246, %r1941;
add.ftz.f32 %f247, %f245, %f246;
mov.b32 %r1942, %f247;
shfl.sync.bfly.b32 %r1944|%p62, %r1942, %r1914, %r1901, %r1903;
mov.b32 %f248, %r1944;
add.ftz.f32 %f56, %f247, %f248;
@%p21 bra $L__BB223_22;

add.s32 %r1945, %r29, 1;
cvt.u64.u32 %rd126, %r1945;
add.s64 %rd128, %rd126, %rd116;
shl.b64 %rd130, %rd128, 2;
add.s64 %rd131, %rd118, %rd130;
st.global.f32 [%rd131], %f56;

$L__BB223_22:
shl.b32 %r1952, %r35, 2;
mov.u32 %r1953, 2;
ld.shared.f32 %f249, [%r1736+1024];
add.ftz.f32 %f250, %f249, %f360;
mov.b32 %r1958, %f250;
mov.u32 %r1959, 31;
mov.u32 %r1960, 16;
mov.u32 %r1961, -1;
shfl.sync.bfly.b32 %r1962|%p64, %r1958, %r1960, %r1959, %r1961;
mov.b32 %f251, %r1962;
add.ftz.f32 %f252, %f250, %f251;
mov.b32 %r1963, %f252;
mov.u32 %r1964, 8;
shfl.sync.bfly.b32 %r1965|%p65, %r1963, %r1964, %r1959, %r1961;
mov.b32 %f253, %r1965;
add.ftz.f32 %f254, %f252, %f253;
mov.b32 %r1966, %f254;
mov.u32 %r1967, 4;
shfl.sync.bfly.b32 %r1968|%p66, %r1966, %r1967, %r1959, %r1961;
mov.b32 %f255, %r1968;
add.ftz.f32 %f256, %f254, %f255;
mov.b32 %r1969, %f256;
shfl.sync.bfly.b32 %r1970|%p67, %r1969, %r1953, %r1959, %r1961;
mov.b32 %f257, %r1970;
add.ftz.f32 %f258, %f256, %f257;
mov.b32 %r1971, %f258;
mov.u32 %r1972, 1;
shfl.sync.bfly.b32 %r1973|%p68, %r1971, %r1972, %r1959, %r1961;
mov.b32 %f259, %r1973;
add.ftz.f32 %f57, %f258, %f259;
cvt.u64.u32 %rd132, %r1952;
add.s64 %rd134, %rd132, %rd116;
shl.b64 %rd136, %rd134, 2;
add.s64 %rd14, %rd118, %rd136;
@%p13 bra $L__BB223_24;

st.global.f32 [%rd14], %f57;

$L__BB223_24:
ld.shared.f32 %f260, [%r1736+1152];
add.ftz.f32 %f261, %f260, %f359;
mov.b32 %r1985, %f261;
shfl.sync.bfly.b32 %r1989|%p72, %r1985, %r1960, %r1959, %r1961;
mov.b32 %f262, %r1989;
add.ftz.f32 %f263, %f261, %f262;
mov.b32 %r1990, %f263;
shfl.sync.bfly.b32 %r1992|%p73, %r1990, %r1964, %r1959, %r1961;
mov.b32 %f264, %r1992;
add.ftz.f32 %f265, %f263, %f264;
mov.b32 %r1993, %f265;
shfl.sync.bfly.b32 %r1995|%p74, %r1993, %r1967, %r1959, %r1961;
mov.b32 %f266, %r1995;
add.ftz.f32 %f267, %f265, %f266;
mov.b32 %r1996, %f267;
shfl.sync.bfly.b32 %r1997|%p75, %r1996, %r1953, %r1959, %r1961;
mov.b32 %f268, %r1997;
add.ftz.f32 %f269, %f267, %f268;
mov.b32 %r1998, %f269;
shfl.sync.bfly.b32 %r2000|%p76, %r1998, %r1972, %r1959, %r1961;
mov.b32 %f270, %r2000;
add.ftz.f32 %f58, %f269, %f270;
@%p21 bra $L__BB223_26;

st.global.f32 [%rd14+4], %f58;

$L__BB223_26:
mul.lo.s32 %r30, %r35, 5;
mov.u32 %r2005, 2;
ld.shared.f32 %f271, [%r1736+1280];
add.ftz.f32 %f272, %f271, %f358;
mov.b32 %r2006, %f272;
mov.u32 %r2007, 31;
mov.u32 %r2008, 16;
mov.u32 %r2009, -1;
shfl.sync.bfly.b32 %r2010|%p78, %r2006, %r2008, %r2007, %r2009;
mov.b32 %f273, %r2010;
add.ftz.f32 %f274, %f272, %f273;
mov.b32 %r2011, %f274;
mov.u32 %r2012, 8;
shfl.sync.bfly.b32 %r2013|%p79, %r2011, %r2012, %r2007, %r2009;
mov.b32 %f275, %r2013;
add.ftz.f32 %f276, %f274, %f275;
mov.b32 %r2014, %f276;
mov.u32 %r2015, 4;
shfl.sync.bfly.b32 %r2016|%p80, %r2014, %r2015, %r2007, %r2009;
mov.b32 %f277, %r2016;
add.ftz.f32 %f278, %f276, %f277;
mov.b32 %r2017, %f278;
shfl.sync.bfly.b32 %r2018|%p81, %r2017, %r2005, %r2007, %r2009;
mov.b32 %f279, %r2018;
add.ftz.f32 %f280, %f278, %f279;
mov.b32 %r2019, %f280;
mov.u32 %r2020, 1;
shfl.sync.bfly.b32 %r2021|%p82, %r2019, %r2020, %r2007, %r2009;
mov.b32 %f281, %r2021;
add.ftz.f32 %f59, %f280, %f281;
@%p13 bra $L__BB223_28;

cvt.u64.u32 %rd137, %r30;
add.s64 %rd139, %rd137, %rd116;
shl.b64 %rd141, %rd139, 2;
add.s64 %rd142, %rd118, %rd141;
st.global.f32 [%rd142], %f59;

$L__BB223_28:
ld.shared.f32 %f282, [%r1736+1408];
add.ftz.f32 %f283, %f282, %f357;
mov.b32 %r2035, %f283;
shfl.sync.bfly.b32 %r2039|%p86, %r2035, %r2008, %r2007, %r2009;
mov.b32 %f284, %r2039;
add.ftz.f32 %f285, %f283, %f284;
mov.b32 %r2040, %f285;
shfl.sync.bfly.b32 %r2042|%p87, %r2040, %r2012, %r2007, %r2009;
mov.b32 %f286, %r2042;
add.ftz.f32 %f287, %f285, %f286;
mov.b32 %r2043, %f287;
shfl.sync.bfly.b32 %r2045|%p88, %r2043, %r2015, %r2007, %r2009;
mov.b32 %f288, %r2045;
add.ftz.f32 %f289, %f287, %f288;
mov.b32 %r2046, %f289;
shfl.sync.bfly.b32 %r2047|%p89, %r2046, %r2005, %r2007, %r2009;
mov.b32 %f290, %r2047;
add.ftz.f32 %f291, %f289, %f290;
mov.b32 %r2048, %f291;
shfl.sync.bfly.b32 %r2050|%p90, %r2048, %r2020, %r2007, %r2009;
mov.b32 %f292, %r2050;
add.ftz.f32 %f60, %f291, %f292;
@%p21 bra $L__BB223_30;

add.s32 %r2051, %r30, 1;
cvt.u64.u32 %rd143, %r2051;
add.s64 %rd145, %rd143, %rd116;
shl.b64 %rd147, %rd145, 2;
add.s64 %rd148, %rd118, %rd147;
st.global.f32 [%rd148], %f60;

$L__BB223_30:
mul.lo.s32 %r2058, %r35, 6;
mov.u32 %r2063, 2;
ld.shared.f32 %f293, [%r1736+1536];
add.ftz.f32 %f294, %f293, %f356;
mov.b32 %r2064, %f294;
mov.u32 %r2065, 31;
mov.u32 %r2066, 16;
mov.u32 %r2067, -1;
shfl.sync.bfly.b32 %r2068|%p92, %r2064, %r2066, %r2065, %r2067;
mov.b32 %f295, %r2068;
add.ftz.f32 %f296, %f294, %f295;
mov.b32 %r2069, %f296;
mov.u32 %r2070, 8;
shfl.sync.bfly.b32 %r2071|%p93, %r2069, %r2070, %r2065, %r2067;
mov.b32 %f297, %r2071;
add.ftz.f32 %f298, %f296, %f297;
mov.b32 %r2072, %f298;
mov.u32 %r2073, 4;
shfl.sync.bfly.b32 %r2074|%p94, %r2072, %r2073, %r2065, %r2067;
mov.b32 %f299, %r2074;
add.ftz.f32 %f300, %f298, %f299;
mov.b32 %r2075, %f300;
shfl.sync.bfly.b32 %r2076|%p95, %r2075, %r2063, %r2065, %r2067;
mov.b32 %f301, %r2076;
add.ftz.f32 %f302, %f300, %f301;
mov.b32 %r2077, %f302;
mov.u32 %r2078, 1;
shfl.sync.bfly.b32 %r2079|%p96, %r2077, %r2078, %r2065, %r2067;
mov.b32 %f303, %r2079;
add.ftz.f32 %f61, %f302, %f303;
cvt.u64.u32 %rd149, %r2058;
add.s64 %rd151, %rd149, %rd116;
shl.b64 %rd153, %rd151, 2;
add.s64 %rd15, %rd118, %rd153;
@%p13 bra $L__BB223_32;

st.global.f32 [%rd15], %f61;

$L__BB223_32:
ld.shared.f32 %f304, [%r1736+1664];
add.ftz.f32 %f305, %f304, %f355;
mov.b32 %r2091, %f305;
shfl.sync.bfly.b32 %r2095|%p100, %r2091, %r2066, %r2065, %r2067;
mov.b32 %f306, %r2095;
add.ftz.f32 %f307, %f305, %f306;
mov.b32 %r2096, %f307;
shfl.sync.bfly.b32 %r2098|%p101, %r2096, %r2070, %r2065, %r2067;
mov.b32 %f308, %r2098;
add.ftz.f32 %f309, %f307, %f308;
mov.b32 %r2099, %f309;
shfl.sync.bfly.b32 %r2101|%p102, %r2099, %r2073, %r2065, %r2067;
mov.b32 %f310, %r2101;
add.ftz.f32 %f311, %f309, %f310;
mov.b32 %r2102, %f311;
shfl.sync.bfly.b32 %r2103|%p103, %r2102, %r2063, %r2065, %r2067;
mov.b32 %f312, %r2103;
add.ftz.f32 %f313, %f311, %f312;
mov.b32 %r2104, %f313;
shfl.sync.bfly.b32 %r2106|%p104, %r2104, %r2078, %r2065, %r2067;
mov.b32 %f314, %r2106;
add.ftz.f32 %f62, %f313, %f314;
@%p21 bra $L__BB223_34;

st.global.f32 [%rd15+4], %f62;

$L__BB223_34:
mul.lo.s32 %r31, %r35, 7;
mov.u32 %r2111, 2;
ld.shared.f32 %f315, [%r1736+1792];
add.ftz.f32 %f316, %f315, %f354;
mov.b32 %r2112, %f316;
mov.u32 %r2113, 31;
mov.u32 %r2114, 16;
mov.u32 %r2115, -1;
shfl.sync.bfly.b32 %r2116|%p106, %r2112, %r2114, %r2113, %r2115;
mov.b32 %f317, %r2116;
add.ftz.f32 %f318, %f316, %f317;
mov.b32 %r2117, %f318;
mov.u32 %r2118, 8;
shfl.sync.bfly.b32 %r2119|%p107, %r2117, %r2118, %r2113, %r2115;
mov.b32 %f319, %r2119;
add.ftz.f32 %f320, %f318, %f319;
mov.b32 %r2120, %f320;
mov.u32 %r2121, 4;
shfl.sync.bfly.b32 %r2122|%p108, %r2120, %r2121, %r2113, %r2115;
mov.b32 %f321, %r2122;
add.ftz.f32 %f322, %f320, %f321;
mov.b32 %r2123, %f322;
shfl.sync.bfly.b32 %r2124|%p109, %r2123, %r2111, %r2113, %r2115;
mov.b32 %f323, %r2124;
add.ftz.f32 %f324, %f322, %f323;
mov.b32 %r2125, %f324;
mov.u32 %r2126, 1;
shfl.sync.bfly.b32 %r2127|%p110, %r2125, %r2126, %r2113, %r2115;
mov.b32 %f325, %r2127;
add.ftz.f32 %f63, %f324, %f325;
@%p13 bra $L__BB223_36;

cvt.u64.u32 %rd154, %r31;
add.s64 %rd156, %rd154, %rd116;
shl.b64 %rd158, %rd156, 2;
add.s64 %rd159, %rd118, %rd158;
st.global.f32 [%rd159], %f63;

$L__BB223_36:
ld.shared.f32 %f326, [%r1736+1920];
add.ftz.f32 %f327, %f326, %f353;
mov.b32 %r2141, %f327;
shfl.sync.bfly.b32 %r2145|%p114, %r2141, %r2114, %r2113, %r2115;
mov.b32 %f328, %r2145;
add.ftz.f32 %f329, %f327, %f328;
mov.b32 %r2146, %f329;
shfl.sync.bfly.b32 %r2148|%p115, %r2146, %r2118, %r2113, %r2115;
mov.b32 %f330, %r2148;
add.ftz.f32 %f331, %f329, %f330;
mov.b32 %r2149, %f331;
shfl.sync.bfly.b32 %r2151|%p116, %r2149, %r2121, %r2113, %r2115;
mov.b32 %f332, %r2151;
add.ftz.f32 %f333, %f331, %f332;
mov.b32 %r2152, %f333;
shfl.sync.bfly.b32 %r2153|%p117, %r2152, %r2111, %r2113, %r2115;
mov.b32 %f334, %r2153;
add.ftz.f32 %f335, %f333, %f334;
mov.b32 %r2154, %f335;
shfl.sync.bfly.b32 %r2156|%p118, %r2154, %r2126, %r2113, %r2115;
mov.b32 %f336, %r2156;
add.ftz.f32 %f64, %f335, %f336;
@%p21 bra $L__BB223_38;

add.s32 %r2157, %r31, 1;
cvt.u64.u32 %rd160, %r2157;
add.s64 %rd162, %rd160, %rd116;
shl.b64 %rd164, %rd162, 2;
add.s64 %rd165, %rd118, %rd164;
st.global.f32 [%rd165], %f64;

$L__BB223_38:
ret;

}
.entry _Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj(
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2,
.param .align 8 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3[48],
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5,
.param .align 4 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6[12],
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15
)
.maxntid 256, 1, 1
.minnctapersm 1
{
.local .align 8 .b8 __local_depot224[16];
.reg .b64 %SP;
.reg .b64 %SPL;
.reg .pred %p<48>;
.reg .b16 %rs<77>;
.reg .f32 %f<169>;
.reg .b32 %r<929>;
.reg .b64 %rd<133>;


mov.u64 %SPL, __local_depot224;
ld.param.u64 %rd21, [_Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0];
ld.param.u64 %rd22, [_Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1];
ld.param.u64 %rd23, [_Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2];
ld.param.u32 %r24, [_Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5];
ld.param.u32 %r29, [_Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8];
ld.param.u32 %r30, [_Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9];
ld.param.u32 %r32, [_Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11];
ld.param.u32 %r33, [_Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12];
ld.param.u32 %r36, [_Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14];
ld.param.u32 %r35, [_Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15];
ld.param.u32 %r27, [_Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+8];
ld.param.u32 %r26, [_Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+4];
ld.param.u32 %r25, [_Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6];
ld.param.u64 %rd25, [_Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+8];
add.u64 %rd1, %SPL, 0;
add.u64 %rd2, %SPL, 8;
mov.u32 %r1, %tid.y;
setp.ge.u32 %p1, %r1, %r36;
@%p1 bra $L__BB224_29;

mov.u32 %r2, %ctaid.y;
cvta.to.global.u64 %rd32, %rd23;
mad.lo.s32 %r37, %r1, %r35, %r2;
mul.wide.u32 %rd33, %r37, 4;
add.s64 %rd34, %rd32, %rd33;
ld.global.nc.u32 %r3, [%rd34];
mov.f32 %f162, 0f00000000;
st.local.v2.f32 [%rd1], {%f162, %f162};
st.local.v2.f32 [%rd2], {%f162, %f162};
mov.u32 %r4, %tid.x;
shr.u32 %r928, %r4, 3;
shr.u32 %r38, %r24, 8;
setp.ge.u32 %p2, %r928, %r38;
mov.f32 %f163, %f162;
mov.f32 %f164, %f162;
mov.f32 %f165, %f162;
@%p2 bra $L__BB224_9;

mov.u32 %r39, %ctaid.x;
shl.b32 %r40, %r39, 1;
mul.lo.s32 %r41, %r3, %r32;
and.b32 %r42, %r4, 7;
cvt.u64.u32 %rd4, %r42;
shl.b32 %r927, %r928, 3;
or.b32 %r43, %r40, 1;
add.s32 %r44, %r928, %r41;
mad.lo.s32 %r45, %r29, %r43, %r44;
mul.wide.s32 %rd5, %r45, 82;
shl.b32 %r46, %r4, 2;
cvt.u64.u32 %rd35, %r46;
and.b64 %rd36, %rd35, 28;
mad.lo.s32 %r47, %r40, %r29, %r928;
add.s32 %r48, %r47, %r41;
mul.wide.s32 %rd6, %r48, 82;
add.s64 %rd37, %rd36, %rd6;
add.s64 %rd7, %rd37, 36;
add.s64 %rd38, %rd5, %rd4;
add.s64 %rd8, %rd38, 66;
add.s64 %rd9, %rd6, %rd4;
add.s64 %rd39, %rd36, %rd5;
add.s64 %rd10, %rd39, 34;
cvta.to.global.u64 %rd131, %rd21;
cvta.to.global.u64 %rd132, %rd25;
mul.lo.s32 %r49, %r1, %r30;
cvt.u64.u32 %rd40, %r49;
mul.hi.u32 %r50, %r2, %r25;
add.s32 %r51, %r50, %r2;
shr.u32 %r52, %r51, %r26;
mul.lo.s32 %r53, %r52, %r27;
sub.s32 %r54, %r2, %r53;
mul.lo.s32 %r55, %r54, %r33;
cvt.u64.u32 %rd41, %r55;
add.s64 %rd13, %rd41, %rd40;
mov.f32 %f162, 0f00000000;
cvta.to.global.u64 %rd42, %rd22;

$L__BB224_3:
cvt.s64.s32 %rd43, %r927;
add.s64 %rd44, %rd13, %rd43;
add.s64 %rd45, %rd131, %rd7;
ld.global.nc.u16 %rs2, [%rd45+-34];
shr.u16 %rs3, %rs2, 8;
cvt.u32.u16 %r105, %rs3;
ld.global.nc.u16 %rs4, [%rd45+-32];
shr.u16 %rs5, %rs4, 8;
cvt.u32.u16 %r106, %rs5;
add.s64 %rd46, %rd131, %rd9;
ld.global.nc.u8 %rs6, [%rd46+66];
ld.global.nc.u16 %rs7, [%rd45+-2];
ld.global.nc.u16 %rs8, [%rd45];
ld.global.nc.u8 %rs9, [%rd46+74];
and.b16 %rs10, %rs9, 240;
and.b16 %rs11, %rs9, 15;
shr.u16 %rs12, %rs10, 4;
add.s64 %rd47, %rd44, %rd4;
and.b16 %rs13, %rs2, 255;
cvt.u32.u16 %r107, %rs13;
shl.b16 %rs14, %rs6, 8;
cvt.u32.u16 %r108, %rs14;
and.b32 %r109, %r108, 768;
or.b32 %r110, %r109, %r107;
mul.wide.u32 %rd48, %r110, 8;
mov.u64 %rd49, iq2s_grid;
add.s64 %rd50, %rd49, %rd48;
and.b16 %rs15, %rs7, 255;
mul.wide.u16 %r111, %rs15, 128;
cvt.u32.u16 %r112, %rs7;
shl.b32 %r113, %r112, 21;
or.b32 %r114, %r111, %r113;
and.b32 %r115, %r114, 25166208;
or.b32 %r116, %r115, -2139062144;
add.s32 %r117, %r116, -16843009;
or.b32 %r56, %r117, %r115;

prmt.b32 %r56, %r56, 0, 0xba98;

mul.wide.u16 %r118, %rs15, 8;
shl.b32 %r119, %r112, 17;
or.b32 %r120, %r118, %r119;
and.b32 %r121, %r120, 25166208;
or.b32 %r122, %r121, -2139062144;
add.s32 %r123, %r122, -16843009;
or.b32 %r58, %r123, %r121;

prmt.b32 %r58, %r58, 0, 0xba98;

ld.global.nc.v2.u32 {%r124, %r125}, [%rd50];
mov.u32 %r87, 0;
xor.b32 %r128, %r124, %r56;
or.b32 %r129, %r128, -2139062144;
and.b32 %r130, %r56, 2139062143;
and.b32 %r131, %r124, -2139062144;
sub.s32 %r132, %r129, %r130;
xor.b32 %r133, %r131, %r132;
xor.b32 %r61, %r133, -2139062144;
xor.b32 %r134, %r125, %r58;
or.b32 %r135, %r134, -2139062144;
and.b32 %r136, %r58, 2139062143;
and.b32 %r137, %r125, -2139062144;
sub.s32 %r138, %r135, %r136;
xor.b32 %r139, %r137, %r138;
xor.b32 %r65, %r139, -2139062144;
mul.lo.s64 %rd51, %rd47, 36;
add.s64 %rd52, %rd42, %rd51;
ld.global.nc.u32 %r9, [%rd52+8];
ld.global.nc.u32 %r10, [%rd52+4];

dp4a.s32.s32 %r60, %r61, %r10, %r87;


dp4a.s32.s32 %r64, %r65, %r9, %r60;

mul.wide.u16 %r140, %rs6, 64;
and.b32 %r141, %r140, 768;
or.b32 %r142, %r141, %r105;
mul.wide.u32 %rd53, %r142, 8;
add.s64 %rd54, %rd49, %rd53;
shr.u16 %rs17, %rs7, 8;
mul.wide.u16 %r143, %rs17, 128;
cvt.u32.u16 %r144, %rs17;
shl.b32 %r145, %r144, 21;
or.b32 %r146, %r143, %r145;
and.b32 %r147, %r146, 25166208;
or.b32 %r148, %r147, -2139062144;
add.s32 %r149, %r148, -16843009;
or.b32 %r68, %r149, %r147;

prmt.b32 %r68, %r68, 0, 0xba98;

mul.wide.u16 %r150, %rs17, 8;
shl.b32 %r151, %r144, 17;
or.b32 %r152, %r150, %r151;
and.b32 %r153, %r152, 25166208;
or.b32 %r154, %r153, -2139062144;
add.s32 %r155, %r154, -16843009;
or.b32 %r70, %r155, %r153;

prmt.b32 %r70, %r70, 0, 0xba98;

ld.global.nc.v2.u32 {%r156, %r157}, [%rd54];
xor.b32 %r160, %r156, %r68;
or.b32 %r161, %r160, -2139062144;
and.b32 %r162, %r68, 2139062143;
and.b32 %r163, %r156, -2139062144;
sub.s32 %r164, %r161, %r162;
xor.b32 %r165, %r163, %r164;
xor.b32 %r73, %r165, -2139062144;
xor.b32 %r166, %r157, %r70;
or.b32 %r167, %r166, -2139062144;
and.b32 %r168, %r70, 2139062143;
and.b32 %r169, %r157, -2139062144;
sub.s32 %r170, %r167, %r168;
xor.b32 %r171, %r169, %r170;
xor.b32 %r77, %r171, -2139062144;
ld.global.nc.u32 %r11, [%rd52+16];
ld.global.nc.u32 %r12, [%rd52+12];

dp4a.s32.s32 %r72, %r73, %r12, %r64;


dp4a.s32.s32 %r76, %r77, %r11, %r72;

and.b16 %rs18, %rs4, 255;
cvt.u32.u16 %r172, %rs18;
mul.wide.u16 %r173, %rs6, 16;
and.b32 %r174, %r173, 768;
or.b32 %r175, %r174, %r172;
mul.wide.u32 %rd55, %r175, 8;
add.s64 %rd56, %rd49, %rd55;
and.b16 %rs19, %rs8, 255;
mul.wide.u16 %r176, %rs19, 128;
cvt.u32.u16 %r177, %rs8;
shl.b32 %r178, %r177, 21;
or.b32 %r179, %r176, %r178;
and.b32 %r180, %r179, 25166208;
or.b32 %r181, %r180, -2139062144;
add.s32 %r182, %r181, -16843009;
or.b32 %r80, %r182, %r180;

prmt.b32 %r80, %r80, 0, 0xba98;

mul.wide.u16 %r183, %rs19, 8;
shl.b32 %r184, %r177, 17;
or.b32 %r185, %r183, %r184;
and.b32 %r186, %r185, 25166208;
or.b32 %r187, %r186, -2139062144;
add.s32 %r188, %r187, -16843009;
or.b32 %r82, %r188, %r186;

prmt.b32 %r82, %r82, 0, 0xba98;

ld.global.nc.v2.u32 {%r189, %r190}, [%rd56];
xor.b32 %r193, %r189, %r80;
or.b32 %r194, %r193, -2139062144;
and.b32 %r195, %r80, 2139062143;
and.b32 %r196, %r189, -2139062144;
sub.s32 %r197, %r194, %r195;
xor.b32 %r198, %r196, %r197;
xor.b32 %r85, %r198, -2139062144;
xor.b32 %r199, %r190, %r82;
or.b32 %r200, %r199, -2139062144;
and.b32 %r201, %r82, 2139062143;
and.b32 %r202, %r190, -2139062144;
sub.s32 %r203, %r200, %r201;
xor.b32 %r204, %r202, %r203;
xor.b32 %r89, %r204, -2139062144;
ld.global.nc.u32 %r13, [%rd52+24];
ld.global.nc.u32 %r14, [%rd52+20];

dp4a.s32.s32 %r84, %r85, %r14, %r87;


dp4a.s32.s32 %r88, %r89, %r13, %r84;

mul.wide.u16 %r205, %rs6, 4;
and.b32 %r206, %r205, 768;
or.b32 %r207, %r206, %r106;
mul.wide.u32 %rd57, %r207, 8;
add.s64 %rd58, %rd49, %rd57;
shr.u16 %rs20, %rs8, 8;
mul.wide.u16 %r208, %rs20, 128;
cvt.u32.u16 %r209, %rs20;
shl.b32 %r210, %r209, 21;
or.b32 %r211, %r208, %r210;
and.b32 %r212, %r211, 25166208;
or.b32 %r213, %r212, -2139062144;
add.s32 %r214, %r213, -16843009;
or.b32 %r92, %r214, %r212;

prmt.b32 %r92, %r92, 0, 0xba98;

mul.wide.u16 %r215, %rs20, 8;
shl.b32 %r216, %r209, 17;
or.b32 %r217, %r215, %r216;
and.b32 %r218, %r217, 25166208;
or.b32 %r219, %r218, -2139062144;
add.s32 %r220, %r219, -16843009;
or.b32 %r94, %r220, %r218;

prmt.b32 %r94, %r94, 0, 0xba98;

ld.global.nc.v2.u32 {%r221, %r222}, [%rd58];
xor.b32 %r225, %r221, %r92;
or.b32 %r226, %r225, -2139062144;
and.b32 %r227, %r92, 2139062143;
and.b32 %r228, %r221, -2139062144;
sub.s32 %r229, %r226, %r227;
xor.b32 %r230, %r228, %r229;
xor.b32 %r97, %r230, -2139062144;
xor.b32 %r231, %r222, %r94;
or.b32 %r232, %r231, -2139062144;
and.b32 %r233, %r94, 2139062143;
and.b32 %r234, %r222, -2139062144;
sub.s32 %r235, %r232, %r233;
xor.b32 %r236, %r234, %r235;
xor.b32 %r101, %r236, -2139062144;
ld.global.nc.u32 %r15, [%rd52+32];
ld.global.nc.u32 %r16, [%rd52+28];

dp4a.s32.s32 %r96, %r97, %r16, %r88;


dp4a.s32.s32 %r100, %r101, %r15, %r96;

cvt.u32.u16 %r237, %rs11;
cvt.u32.u16 %r238, %rs12;
mul.lo.s32 %r239, %r76, %r237;
mad.lo.s32 %r240, %r100, %r238, %r239;
add.s32 %r241, %r100, %r76;
shr.u32 %r242, %r241, 31;
add.s32 %r243, %r241, %r242;
shr.s32 %r244, %r243, 1;
add.s32 %r245, %r240, %r244;
shr.s32 %r246, %r245, 31;
shr.u32 %r247, %r246, 30;
add.s32 %r248, %r245, %r247;
shr.s32 %r249, %r248, 2;
add.s64 %rd59, %rd131, %rd6;
ld.global.nc.u16 %rs1, [%rd59];

{ cvt.f32.f16 %f42, %rs1;}


ld.global.nc.u32 %r17, [%rd52];

{.reg .f16 low,high;
mov.b32 {low,high},%r17;
cvt.f32.f16 %f43, low;}


mul.ftz.f32 %f44, %f42, %f43;
cvt.rn.f32.s32 %f45, %r249;
fma.rn.ftz.f32 %f165, %f44, %f45, %f165;
setp.eq.s64 %p3, %rd25, 0;
@%p3 bra $L__BB224_5;

mov.u64 %rd115, iq2s_grid;
add.s64 %rd60, %rd132, %rd7;
ld.global.u16 %rs22, [%rd60+-34];
shr.u16 %rs23, %rs22, 8;
cvt.u32.u16 %r299, %rs23;
ld.global.u16 %rs24, [%rd60+-32];
shr.u16 %rs25, %rs24, 8;
cvt.u32.u16 %r300, %rs25;
add.s64 %rd61, %rd132, %rd9;
ld.global.u16 %rs26, [%rd60+-2];
ld.global.u16 %rs27, [%rd60];
ld.global.u8 %rs28, [%rd61+74];
and.b16 %rs29, %rs28, 15;
shr.u16 %rs30, %rs28, 4;
and.b16 %rs31, %rs22, 255;
cvt.u32.u16 %r301, %rs31;
ld.global.u8 %rs32, [%rd61+66];
shl.b16 %rs33, %rs32, 8;
cvt.u32.u16 %r302, %rs33;
and.b32 %r303, %r302, 768;
or.b32 %r304, %r303, %r301;
mul.wide.u32 %rd62, %r304, 8;
add.s64 %rd64, %rd115, %rd62;
and.b16 %rs34, %rs26, 255;
mul.wide.u16 %r305, %rs34, 128;
cvt.u32.u16 %r306, %rs26;
shl.b32 %r307, %r306, 21;
or.b32 %r308, %r305, %r307;
and.b32 %r309, %r308, 25166208;
or.b32 %r310, %r309, -2139062144;
add.s32 %r311, %r310, -16843009;
or.b32 %r250, %r311, %r309;

prmt.b32 %r250, %r250, 0, 0xba98;

mul.wide.u16 %r312, %rs34, 8;
shl.b32 %r313, %r306, 17;
or.b32 %r314, %r312, %r313;
and.b32 %r315, %r314, 25166208;
or.b32 %r316, %r315, -2139062144;
add.s32 %r317, %r316, -16843009;
or.b32 %r252, %r317, %r315;

prmt.b32 %r252, %r252, 0, 0xba98;

ld.global.nc.v2.u32 {%r318, %r319}, [%rd64];
xor.b32 %r322, %r318, %r250;
or.b32 %r323, %r322, -2139062144;
and.b32 %r324, %r250, 2139062143;
and.b32 %r325, %r318, -2139062144;
sub.s32 %r326, %r323, %r324;
xor.b32 %r327, %r325, %r326;
xor.b32 %r255, %r327, -2139062144;
xor.b32 %r328, %r319, %r252;
or.b32 %r329, %r328, -2139062144;
and.b32 %r330, %r252, 2139062143;
and.b32 %r331, %r319, -2139062144;
sub.s32 %r332, %r329, %r330;
xor.b32 %r333, %r331, %r332;
xor.b32 %r259, %r333, -2139062144;

dp4a.s32.s32 %r254, %r255, %r10, %r87;


dp4a.s32.s32 %r258, %r259, %r9, %r254;

mul.wide.u16 %r334, %rs32, 64;
and.b32 %r335, %r334, 768;
or.b32 %r336, %r335, %r299;
mul.wide.u32 %rd65, %r336, 8;
add.s64 %rd66, %rd115, %rd65;
shr.u16 %rs35, %rs26, 8;
mul.wide.u16 %r337, %rs35, 128;
cvt.u32.u16 %r338, %rs35;
shl.b32 %r339, %r338, 21;
or.b32 %r340, %r337, %r339;
and.b32 %r341, %r340, 25166208;
or.b32 %r342, %r341, -2139062144;
add.s32 %r343, %r342, -16843009;
or.b32 %r262, %r343, %r341;

prmt.b32 %r262, %r262, 0, 0xba98;

mul.wide.u16 %r344, %rs35, 8;
shl.b32 %r345, %r338, 17;
or.b32 %r346, %r344, %r345;
and.b32 %r347, %r346, 25166208;
or.b32 %r348, %r347, -2139062144;
add.s32 %r349, %r348, -16843009;
or.b32 %r264, %r349, %r347;

prmt.b32 %r264, %r264, 0, 0xba98;

ld.global.nc.v2.u32 {%r350, %r351}, [%rd66];
xor.b32 %r354, %r350, %r262;
or.b32 %r355, %r354, -2139062144;
and.b32 %r356, %r262, 2139062143;
and.b32 %r357, %r350, -2139062144;
sub.s32 %r358, %r355, %r356;
xor.b32 %r359, %r357, %r358;
xor.b32 %r267, %r359, -2139062144;
xor.b32 %r360, %r351, %r264;
or.b32 %r361, %r360, -2139062144;
and.b32 %r362, %r264, 2139062143;
and.b32 %r363, %r351, -2139062144;
sub.s32 %r364, %r361, %r362;
xor.b32 %r365, %r363, %r364;
xor.b32 %r271, %r365, -2139062144;

dp4a.s32.s32 %r266, %r267, %r12, %r258;


dp4a.s32.s32 %r270, %r271, %r11, %r266;

and.b16 %rs36, %rs24, 255;
cvt.u32.u16 %r366, %rs36;
mul.wide.u16 %r367, %rs32, 16;
and.b32 %r368, %r367, 768;
or.b32 %r369, %r368, %r366;
mul.wide.u32 %rd67, %r369, 8;
add.s64 %rd68, %rd115, %rd67;
and.b16 %rs37, %rs27, 255;
mul.wide.u16 %r370, %rs37, 128;
cvt.u32.u16 %r371, %rs27;
shl.b32 %r372, %r371, 21;
or.b32 %r373, %r370, %r372;
and.b32 %r374, %r373, 25166208;
or.b32 %r375, %r374, -2139062144;
add.s32 %r376, %r375, -16843009;
or.b32 %r274, %r376, %r374;

prmt.b32 %r274, %r274, 0, 0xba98;

mul.wide.u16 %r377, %rs37, 8;
shl.b32 %r378, %r371, 17;
or.b32 %r379, %r377, %r378;
and.b32 %r380, %r379, 25166208;
or.b32 %r381, %r380, -2139062144;
add.s32 %r382, %r381, -16843009;
or.b32 %r276, %r382, %r380;

prmt.b32 %r276, %r276, 0, 0xba98;

ld.global.nc.v2.u32 {%r383, %r384}, [%rd68];
xor.b32 %r387, %r383, %r274;
or.b32 %r388, %r387, -2139062144;
and.b32 %r389, %r274, 2139062143;
and.b32 %r390, %r383, -2139062144;
sub.s32 %r391, %r388, %r389;
xor.b32 %r392, %r390, %r391;
xor.b32 %r279, %r392, -2139062144;
xor.b32 %r393, %r384, %r276;
or.b32 %r394, %r393, -2139062144;
and.b32 %r395, %r276, 2139062143;
and.b32 %r396, %r384, -2139062144;
sub.s32 %r397, %r394, %r395;
xor.b32 %r398, %r396, %r397;
xor.b32 %r283, %r398, -2139062144;

dp4a.s32.s32 %r278, %r279, %r14, %r87;


dp4a.s32.s32 %r282, %r283, %r13, %r278;

mul.wide.u16 %r399, %rs32, 4;
and.b32 %r400, %r399, 768;
or.b32 %r401, %r400, %r300;
mul.wide.u32 %rd69, %r401, 8;
add.s64 %rd70, %rd115, %rd69;
shr.u16 %rs38, %rs27, 8;
mul.wide.u16 %r402, %rs38, 128;
cvt.u32.u16 %r403, %rs38;
shl.b32 %r404, %r403, 21;
or.b32 %r405, %r402, %r404;
and.b32 %r406, %r405, 25166208;
or.b32 %r407, %r406, -2139062144;
add.s32 %r408, %r407, -16843009;
or.b32 %r286, %r408, %r406;

prmt.b32 %r286, %r286, 0, 0xba98;

mul.wide.u16 %r409, %rs38, 8;
shl.b32 %r410, %r403, 17;
or.b32 %r411, %r409, %r410;
and.b32 %r412, %r411, 25166208;
or.b32 %r413, %r412, -2139062144;
add.s32 %r414, %r413, -16843009;
or.b32 %r288, %r414, %r412;

prmt.b32 %r288, %r288, 0, 0xba98;

ld.global.nc.v2.u32 {%r415, %r416}, [%rd70];
xor.b32 %r419, %r415, %r286;
or.b32 %r420, %r419, -2139062144;
and.b32 %r421, %r286, 2139062143;
and.b32 %r422, %r415, -2139062144;
sub.s32 %r423, %r420, %r421;
xor.b32 %r424, %r422, %r423;
xor.b32 %r291, %r424, -2139062144;
xor.b32 %r425, %r416, %r288;
or.b32 %r426, %r425, -2139062144;
and.b32 %r427, %r288, 2139062143;
and.b32 %r428, %r416, -2139062144;
sub.s32 %r429, %r426, %r427;
xor.b32 %r430, %r428, %r429;
xor.b32 %r295, %r430, -2139062144;

dp4a.s32.s32 %r290, %r291, %r16, %r282;


dp4a.s32.s32 %r294, %r295, %r15, %r290;

cvt.u32.u16 %r431, %rs29;
cvt.u32.u16 %r432, %rs30;
mul.lo.s32 %r433, %r270, %r431;
mad.lo.s32 %r434, %r294, %r432, %r433;
add.s32 %r435, %r294, %r270;
shr.u32 %r436, %r435, 31;
add.s32 %r437, %r435, %r436;
shr.s32 %r438, %r437, 1;
add.s32 %r439, %r434, %r438;
shr.s32 %r440, %r439, 31;
shr.u32 %r441, %r440, 30;
add.s32 %r442, %r439, %r441;
shr.s32 %r443, %r442, 2;
add.s64 %rd71, %rd132, %rd6;
ld.global.u16 %rs21, [%rd71];

{ cvt.f32.f16 %f46, %rs21;}



{.reg .f16 low,high;
mov.b32 {low,high},%r17;
cvt.f32.f16 %f47, low;}


mul.ftz.f32 %f48, %f46, %f47;
cvt.rn.f32.s32 %f49, %r443;
fma.rn.ftz.f32 %f163, %f48, %f49, %f163;

$L__BB224_5:
mov.u64 %rd109, iq2s_grid;
add.s64 %rd72, %rd131, %rd10;
ld.global.nc.u16 %rs40, [%rd72+-32];
shr.u16 %rs41, %rs40, 8;
cvt.u32.u16 %r493, %rs41;
ld.global.nc.u16 %rs42, [%rd72+-30];
shr.u16 %rs43, %rs42, 8;
cvt.u32.u16 %r494, %rs43;
add.s64 %rd73, %rd131, %rd8;
ld.global.nc.u8 %rs44, [%rd73];
ld.global.nc.u16 %rs45, [%rd72];
ld.global.nc.u16 %rs46, [%rd72+2];
ld.global.nc.u8 %rs47, [%rd73+8];
and.b16 %rs48, %rs47, 240;
and.b16 %rs49, %rs47, 15;
shr.u16 %rs50, %rs48, 4;
and.b16 %rs51, %rs40, 255;
cvt.u32.u16 %r495, %rs51;
shl.b16 %rs52, %rs44, 8;
cvt.u32.u16 %r496, %rs52;
and.b32 %r497, %r496, 768;
or.b32 %r498, %r497, %r495;
mul.wide.u32 %rd74, %r498, 8;
add.s64 %rd76, %rd109, %rd74;
and.b16 %rs53, %rs45, 255;
mul.wide.u16 %r499, %rs53, 128;
cvt.u32.u16 %r500, %rs45;
shl.b32 %r501, %r500, 21;
or.b32 %r502, %r499, %r501;
and.b32 %r503, %r502, 25166208;
or.b32 %r504, %r503, -2139062144;
add.s32 %r505, %r504, -16843009;
or.b32 %r444, %r505, %r503;

prmt.b32 %r444, %r444, 0, 0xba98;

mul.wide.u16 %r506, %rs53, 8;
shl.b32 %r507, %r500, 17;
or.b32 %r508, %r506, %r507;
and.b32 %r509, %r508, 25166208;
or.b32 %r510, %r509, -2139062144;
add.s32 %r511, %r510, -16843009;
or.b32 %r446, %r511, %r509;

prmt.b32 %r446, %r446, 0, 0xba98;

ld.global.nc.v2.u32 {%r512, %r513}, [%rd76];
xor.b32 %r516, %r512, %r444;
or.b32 %r517, %r516, -2139062144;
and.b32 %r518, %r444, 2139062143;
and.b32 %r519, %r512, -2139062144;
sub.s32 %r520, %r517, %r518;
xor.b32 %r521, %r519, %r520;
xor.b32 %r449, %r521, -2139062144;
xor.b32 %r522, %r513, %r446;
or.b32 %r523, %r522, -2139062144;
and.b32 %r524, %r446, 2139062143;
and.b32 %r525, %r513, -2139062144;
sub.s32 %r526, %r523, %r524;
xor.b32 %r527, %r525, %r526;
xor.b32 %r453, %r527, -2139062144;

dp4a.s32.s32 %r448, %r449, %r10, %r87;


dp4a.s32.s32 %r452, %r453, %r9, %r448;

mul.wide.u16 %r528, %rs44, 64;
and.b32 %r529, %r528, 768;
or.b32 %r530, %r529, %r493;
mul.wide.u32 %rd77, %r530, 8;
add.s64 %rd78, %rd109, %rd77;
shr.u16 %rs55, %rs45, 8;
mul.wide.u16 %r531, %rs55, 128;
cvt.u32.u16 %r532, %rs55;
shl.b32 %r533, %r532, 21;
or.b32 %r534, %r531, %r533;
and.b32 %r535, %r534, 25166208;
or.b32 %r536, %r535, -2139062144;
add.s32 %r537, %r536, -16843009;
or.b32 %r456, %r537, %r535;

prmt.b32 %r456, %r456, 0, 0xba98;

mul.wide.u16 %r538, %rs55, 8;
shl.b32 %r539, %r532, 17;
or.b32 %r540, %r538, %r539;
and.b32 %r541, %r540, 25166208;
or.b32 %r542, %r541, -2139062144;
add.s32 %r543, %r542, -16843009;
or.b32 %r458, %r543, %r541;

prmt.b32 %r458, %r458, 0, 0xba98;

ld.global.nc.v2.u32 {%r544, %r545}, [%rd78];
xor.b32 %r548, %r544, %r456;
or.b32 %r549, %r548, -2139062144;
and.b32 %r550, %r456, 2139062143;
and.b32 %r551, %r544, -2139062144;
sub.s32 %r552, %r549, %r550;
xor.b32 %r553, %r551, %r552;
xor.b32 %r461, %r553, -2139062144;
xor.b32 %r554, %r545, %r458;
or.b32 %r555, %r554, -2139062144;
and.b32 %r556, %r458, 2139062143;
and.b32 %r557, %r545, -2139062144;
sub.s32 %r558, %r555, %r556;
xor.b32 %r559, %r557, %r558;
xor.b32 %r465, %r559, -2139062144;

dp4a.s32.s32 %r460, %r461, %r12, %r452;


dp4a.s32.s32 %r464, %r465, %r11, %r460;

and.b16 %rs56, %rs42, 255;
cvt.u32.u16 %r560, %rs56;
mul.wide.u16 %r561, %rs44, 16;
and.b32 %r562, %r561, 768;
or.b32 %r563, %r562, %r560;
mul.wide.u32 %rd79, %r563, 8;
add.s64 %rd80, %rd109, %rd79;
and.b16 %rs57, %rs46, 255;
mul.wide.u16 %r564, %rs57, 128;
cvt.u32.u16 %r565, %rs46;
shl.b32 %r566, %r565, 21;
or.b32 %r567, %r564, %r566;
and.b32 %r568, %r567, 25166208;
or.b32 %r569, %r568, -2139062144;
add.s32 %r570, %r569, -16843009;
or.b32 %r468, %r570, %r568;

prmt.b32 %r468, %r468, 0, 0xba98;

mul.wide.u16 %r571, %rs57, 8;
shl.b32 %r572, %r565, 17;
or.b32 %r573, %r571, %r572;
and.b32 %r574, %r573, 25166208;
or.b32 %r575, %r574, -2139062144;
add.s32 %r576, %r575, -16843009;
or.b32 %r470, %r576, %r574;

prmt.b32 %r470, %r470, 0, 0xba98;

ld.global.nc.v2.u32 {%r577, %r578}, [%rd80];
xor.b32 %r581, %r577, %r468;
or.b32 %r582, %r581, -2139062144;
and.b32 %r583, %r468, 2139062143;
and.b32 %r584, %r577, -2139062144;
sub.s32 %r585, %r582, %r583;
xor.b32 %r586, %r584, %r585;
xor.b32 %r473, %r586, -2139062144;
xor.b32 %r587, %r578, %r470;
or.b32 %r588, %r587, -2139062144;
and.b32 %r589, %r470, 2139062143;
and.b32 %r590, %r578, -2139062144;
sub.s32 %r591, %r588, %r589;
xor.b32 %r592, %r590, %r591;
xor.b32 %r477, %r592, -2139062144;

dp4a.s32.s32 %r472, %r473, %r14, %r87;


dp4a.s32.s32 %r476, %r477, %r13, %r472;

mul.wide.u16 %r593, %rs44, 4;
and.b32 %r594, %r593, 768;
or.b32 %r595, %r594, %r494;
mul.wide.u32 %rd81, %r595, 8;
add.s64 %rd82, %rd109, %rd81;
shr.u16 %rs58, %rs46, 8;
mul.wide.u16 %r596, %rs58, 128;
cvt.u32.u16 %r597, %rs58;
shl.b32 %r598, %r597, 21;
or.b32 %r599, %r596, %r598;
and.b32 %r600, %r599, 25166208;
or.b32 %r601, %r600, -2139062144;
add.s32 %r602, %r601, -16843009;
or.b32 %r480, %r602, %r600;

prmt.b32 %r480, %r480, 0, 0xba98;

mul.wide.u16 %r603, %rs58, 8;
shl.b32 %r604, %r597, 17;
or.b32 %r605, %r603, %r604;
and.b32 %r606, %r605, 25166208;
or.b32 %r607, %r606, -2139062144;
add.s32 %r608, %r607, -16843009;
or.b32 %r482, %r608, %r606;

prmt.b32 %r482, %r482, 0, 0xba98;

ld.global.nc.v2.u32 {%r609, %r610}, [%rd82];
xor.b32 %r613, %r609, %r480;
or.b32 %r614, %r613, -2139062144;
and.b32 %r615, %r480, 2139062143;
and.b32 %r616, %r609, -2139062144;
sub.s32 %r617, %r614, %r615;
xor.b32 %r618, %r616, %r617;
xor.b32 %r485, %r618, -2139062144;
xor.b32 %r619, %r610, %r482;
or.b32 %r620, %r619, -2139062144;
and.b32 %r621, %r482, 2139062143;
and.b32 %r622, %r610, -2139062144;
sub.s32 %r623, %r620, %r621;
xor.b32 %r624, %r622, %r623;
xor.b32 %r489, %r624, -2139062144;

dp4a.s32.s32 %r484, %r485, %r16, %r476;


dp4a.s32.s32 %r488, %r489, %r15, %r484;

cvt.u32.u16 %r625, %rs49;
cvt.u32.u16 %r626, %rs50;
mul.lo.s32 %r627, %r464, %r625;
mad.lo.s32 %r628, %r488, %r626, %r627;
add.s32 %r629, %r488, %r464;
shr.u32 %r630, %r629, 31;
add.s32 %r631, %r629, %r630;
shr.s32 %r632, %r631, 1;
add.s32 %r633, %r628, %r632;
shr.s32 %r634, %r633, 31;
shr.u32 %r635, %r634, 30;
add.s32 %r636, %r633, %r635;
shr.s32 %r637, %r636, 2;
add.s64 %rd83, %rd131, %rd5;
ld.global.nc.u16 %rs39, [%rd83];

{ cvt.f32.f16 %f50, %rs39;}



{.reg .f16 low,high;
mov.b32 {low,high},%r17;
cvt.f32.f16 %f51, low;}


mul.ftz.f32 %f52, %f50, %f51;
cvt.rn.f32.s32 %f53, %r637;
fma.rn.ftz.f32 %f164, %f52, %f53, %f164;
@%p3 bra $L__BB224_7;

mov.u64 %rd116, iq2s_grid;
add.s64 %rd84, %rd132, %rd10;
ld.global.u16 %rs60, [%rd84+-32];
shr.u16 %rs61, %rs60, 8;
cvt.u32.u16 %r687, %rs61;
ld.global.u16 %rs62, [%rd84+-30];
shr.u16 %rs63, %rs62, 8;
cvt.u32.u16 %r688, %rs63;
add.s64 %rd85, %rd132, %rd8;
ld.global.u16 %rs64, [%rd84];
ld.global.u16 %rs65, [%rd84+2];
ld.global.u8 %rs66, [%rd85+8];
and.b16 %rs67, %rs66, 15;
shr.u16 %rs68, %rs66, 4;
and.b16 %rs69, %rs60, 255;
cvt.u32.u16 %r689, %rs69;
ld.global.u8 %rs70, [%rd85];
shl.b16 %rs71, %rs70, 8;
cvt.u32.u16 %r690, %rs71;
and.b32 %r691, %r690, 768;
or.b32 %r692, %r691, %r689;
mul.wide.u32 %rd86, %r692, 8;
add.s64 %rd88, %rd116, %rd86;
and.b16 %rs72, %rs64, 255;
mul.wide.u16 %r693, %rs72, 128;
cvt.u32.u16 %r694, %rs64;
shl.b32 %r695, %r694, 21;
or.b32 %r696, %r693, %r695;
and.b32 %r697, %r696, 25166208;
or.b32 %r698, %r697, -2139062144;
add.s32 %r699, %r698, -16843009;
or.b32 %r638, %r699, %r697;

prmt.b32 %r638, %r638, 0, 0xba98;

mul.wide.u16 %r700, %rs72, 8;
shl.b32 %r701, %r694, 17;
or.b32 %r702, %r700, %r701;
and.b32 %r703, %r702, 25166208;
or.b32 %r704, %r703, -2139062144;
add.s32 %r705, %r704, -16843009;
or.b32 %r640, %r705, %r703;

prmt.b32 %r640, %r640, 0, 0xba98;

ld.global.nc.v2.u32 {%r706, %r707}, [%rd88];
mov.u32 %r669, 0;
xor.b32 %r710, %r706, %r638;
or.b32 %r711, %r710, -2139062144;
and.b32 %r712, %r638, 2139062143;
and.b32 %r713, %r706, -2139062144;
sub.s32 %r714, %r711, %r712;
xor.b32 %r715, %r713, %r714;
xor.b32 %r643, %r715, -2139062144;
xor.b32 %r716, %r707, %r640;
or.b32 %r717, %r716, -2139062144;
and.b32 %r718, %r640, 2139062143;
and.b32 %r719, %r707, -2139062144;
sub.s32 %r720, %r717, %r718;
xor.b32 %r721, %r719, %r720;
xor.b32 %r647, %r721, -2139062144;

dp4a.s32.s32 %r642, %r643, %r10, %r669;


dp4a.s32.s32 %r646, %r647, %r9, %r642;

mul.wide.u16 %r722, %rs70, 64;
and.b32 %r723, %r722, 768;
or.b32 %r724, %r723, %r687;
mul.wide.u32 %rd89, %r724, 8;
add.s64 %rd90, %rd116, %rd89;
shr.u16 %rs73, %rs64, 8;
mul.wide.u16 %r725, %rs73, 128;
cvt.u32.u16 %r726, %rs73;
shl.b32 %r727, %r726, 21;
or.b32 %r728, %r725, %r727;
and.b32 %r729, %r728, 25166208;
or.b32 %r730, %r729, -2139062144;
add.s32 %r731, %r730, -16843009;
or.b32 %r650, %r731, %r729;

prmt.b32 %r650, %r650, 0, 0xba98;

mul.wide.u16 %r732, %rs73, 8;
shl.b32 %r733, %r726, 17;
or.b32 %r734, %r732, %r733;
and.b32 %r735, %r734, 25166208;
or.b32 %r736, %r735, -2139062144;
add.s32 %r737, %r736, -16843009;
or.b32 %r652, %r737, %r735;

prmt.b32 %r652, %r652, 0, 0xba98;

ld.global.nc.v2.u32 {%r738, %r739}, [%rd90];
xor.b32 %r742, %r738, %r650;
or.b32 %r743, %r742, -2139062144;
and.b32 %r744, %r650, 2139062143;
and.b32 %r745, %r738, -2139062144;
sub.s32 %r746, %r743, %r744;
xor.b32 %r747, %r745, %r746;
xor.b32 %r655, %r747, -2139062144;
xor.b32 %r748, %r739, %r652;
or.b32 %r749, %r748, -2139062144;
and.b32 %r750, %r652, 2139062143;
and.b32 %r751, %r739, -2139062144;
sub.s32 %r752, %r749, %r750;
xor.b32 %r753, %r751, %r752;
xor.b32 %r659, %r753, -2139062144;

dp4a.s32.s32 %r654, %r655, %r12, %r646;


dp4a.s32.s32 %r658, %r659, %r11, %r654;

and.b16 %rs74, %rs62, 255;
cvt.u32.u16 %r754, %rs74;
mul.wide.u16 %r755, %rs70, 16;
and.b32 %r756, %r755, 768;
or.b32 %r757, %r756, %r754;
mul.wide.u32 %rd91, %r757, 8;
add.s64 %rd92, %rd116, %rd91;
and.b16 %rs75, %rs65, 255;
mul.wide.u16 %r758, %rs75, 128;
cvt.u32.u16 %r759, %rs65;
shl.b32 %r760, %r759, 21;
or.b32 %r761, %r758, %r760;
and.b32 %r762, %r761, 25166208;
or.b32 %r763, %r762, -2139062144;
add.s32 %r764, %r763, -16843009;
or.b32 %r662, %r764, %r762;

prmt.b32 %r662, %r662, 0, 0xba98;

mul.wide.u16 %r765, %rs75, 8;
shl.b32 %r766, %r759, 17;
or.b32 %r767, %r765, %r766;
and.b32 %r768, %r767, 25166208;
or.b32 %r769, %r768, -2139062144;
add.s32 %r770, %r769, -16843009;
or.b32 %r664, %r770, %r768;

prmt.b32 %r664, %r664, 0, 0xba98;

ld.global.nc.v2.u32 {%r771, %r772}, [%rd92];
xor.b32 %r775, %r771, %r662;
or.b32 %r776, %r775, -2139062144;
and.b32 %r777, %r662, 2139062143;
and.b32 %r778, %r771, -2139062144;
sub.s32 %r779, %r776, %r777;
xor.b32 %r780, %r778, %r779;
xor.b32 %r667, %r780, -2139062144;
xor.b32 %r781, %r772, %r664;
or.b32 %r782, %r781, -2139062144;
and.b32 %r783, %r664, 2139062143;
and.b32 %r784, %r772, -2139062144;
sub.s32 %r785, %r782, %r783;
xor.b32 %r786, %r784, %r785;
xor.b32 %r671, %r786, -2139062144;

dp4a.s32.s32 %r666, %r667, %r14, %r669;


dp4a.s32.s32 %r670, %r671, %r13, %r666;

mul.wide.u16 %r787, %rs70, 4;
and.b32 %r788, %r787, 768;
or.b32 %r789, %r788, %r688;
mul.wide.u32 %rd93, %r789, 8;
add.s64 %rd94, %rd116, %rd93;
shr.u16 %rs76, %rs65, 8;
mul.wide.u16 %r790, %rs76, 128;
cvt.u32.u16 %r791, %rs76;
shl.b32 %r792, %r791, 21;
or.b32 %r793, %r790, %r792;
and.b32 %r794, %r793, 25166208;
or.b32 %r795, %r794, -2139062144;
add.s32 %r796, %r795, -16843009;
or.b32 %r674, %r796, %r794;

prmt.b32 %r674, %r674, 0, 0xba98;

mul.wide.u16 %r797, %rs76, 8;
shl.b32 %r798, %r791, 17;
or.b32 %r799, %r797, %r798;
and.b32 %r800, %r799, 25166208;
or.b32 %r801, %r800, -2139062144;
add.s32 %r802, %r801, -16843009;
or.b32 %r676, %r802, %r800;

prmt.b32 %r676, %r676, 0, 0xba98;

ld.global.nc.v2.u32 {%r803, %r804}, [%rd94];
xor.b32 %r807, %r803, %r674;
or.b32 %r808, %r807, -2139062144;
and.b32 %r809, %r674, 2139062143;
and.b32 %r810, %r803, -2139062144;
sub.s32 %r811, %r808, %r809;
xor.b32 %r812, %r810, %r811;
xor.b32 %r679, %r812, -2139062144;
xor.b32 %r813, %r804, %r676;
or.b32 %r814, %r813, -2139062144;
and.b32 %r815, %r676, 2139062143;
and.b32 %r816, %r804, -2139062144;
sub.s32 %r817, %r814, %r815;
xor.b32 %r818, %r816, %r817;
xor.b32 %r683, %r818, -2139062144;

dp4a.s32.s32 %r678, %r679, %r16, %r670;


dp4a.s32.s32 %r682, %r683, %r15, %r678;

cvt.u32.u16 %r819, %rs67;
cvt.u32.u16 %r820, %rs68;
mul.lo.s32 %r821, %r658, %r819;
mad.lo.s32 %r822, %r682, %r820, %r821;
add.s32 %r823, %r682, %r658;
shr.u32 %r824, %r823, 31;
add.s32 %r825, %r823, %r824;
shr.s32 %r826, %r825, 1;
add.s32 %r827, %r822, %r826;
shr.s32 %r828, %r827, 31;
shr.u32 %r829, %r828, 30;
add.s32 %r830, %r827, %r829;
shr.s32 %r831, %r830, 2;
add.s64 %rd95, %rd132, %rd5;
ld.global.u16 %rs59, [%rd95];

{ cvt.f32.f16 %f54, %rs59;}



{.reg .f16 low,high;
mov.b32 {low,high},%r17;
cvt.f32.f16 %f55, low;}


mul.ftz.f32 %f56, %f54, %f55;
cvt.rn.f32.s32 %f57, %r831;
fma.rn.ftz.f32 %f162, %f56, %f57, %f162;

$L__BB224_7:
add.s32 %r927, %r927, 32;
add.s64 %rd132, %rd132, 328;
add.s64 %rd131, %rd131, 328;
add.s32 %r928, %r928, 4;
setp.lt.u32 %p5, %r928, %r38;
@%p5 bra $L__BB224_3;

add.u64 %rd123, %SPL, 0;
add.u64 %rd117, %SPL, 8;
st.local.v2.f32 [%rd123], {%f165, %f164};
st.local.v2.f32 [%rd117], {%f163, %f162};

$L__BB224_9:
mov.b32 %r20, %f165;
setp.ne.s64 %p6, %rd25, 0;
@%p6 bra $L__BB224_11;
bra.uni $L__BB224_10;

$L__BB224_11:
add.u64 %rd129, %SPL, 0;
add.u64 %rd121, %SPL, 8;
mov.u32 %r859, 31;
mov.u32 %r860, 16;
mov.u32 %r861, -1;
shfl.sync.bfly.b32 %r862|%p17, %r20, %r860, %r859, %r861;
mov.b32 %f78, %r862;
add.ftz.f32 %f79, %f165, %f78;
mov.b32 %r863, %f79;
mov.u32 %r864, 8;
shfl.sync.bfly.b32 %r865|%p18, %r863, %r864, %r859, %r861;
mov.b32 %f80, %r865;
add.ftz.f32 %f81, %f79, %f80;
mov.b32 %r866, %f81;
mov.u32 %r867, 4;
shfl.sync.bfly.b32 %r868|%p19, %r866, %r867, %r859, %r861;
mov.b32 %f82, %r868;
add.ftz.f32 %f83, %f81, %f82;
mov.b32 %r869, %f83;
mov.u32 %r870, 2;
shfl.sync.bfly.b32 %r871|%p20, %r869, %r870, %r859, %r861;
mov.b32 %f84, %r871;
add.ftz.f32 %f85, %f83, %f84;
mov.b32 %r872, %f85;
mov.u32 %r873, 1;
shfl.sync.bfly.b32 %r874|%p21, %r872, %r873, %r859, %r861;
mov.b32 %f86, %r874;
add.ftz.f32 %f87, %f85, %f86;
st.local.f32 [%rd129], %f87;
mov.b32 %r875, %f163;
shfl.sync.bfly.b32 %r876|%p22, %r875, %r860, %r859, %r861;
mov.b32 %f88, %r876;
add.ftz.f32 %f89, %f163, %f88;
mov.b32 %r877, %f89;
shfl.sync.bfly.b32 %r878|%p23, %r877, %r864, %r859, %r861;
mov.b32 %f90, %r878;
add.ftz.f32 %f91, %f89, %f90;
mov.b32 %r879, %f91;
shfl.sync.bfly.b32 %r880|%p24, %r879, %r867, %r859, %r861;
mov.b32 %f92, %r880;
add.ftz.f32 %f93, %f91, %f92;
mov.b32 %r881, %f93;
shfl.sync.bfly.b32 %r882|%p25, %r881, %r870, %r859, %r861;
mov.b32 %f94, %r882;
add.ftz.f32 %f95, %f93, %f94;
mov.b32 %r883, %f95;
shfl.sync.bfly.b32 %r884|%p26, %r883, %r873, %r859, %r861;
mov.b32 %f96, %r884;
add.ftz.f32 %f97, %f95, %f96;
st.local.f32 [%rd121], %f97;
mov.b32 %r885, %f164;
shfl.sync.bfly.b32 %r886|%p27, %r885, %r860, %r859, %r861;
mov.b32 %f98, %r886;
add.ftz.f32 %f99, %f164, %f98;
mov.b32 %r887, %f99;
shfl.sync.bfly.b32 %r888|%p28, %r887, %r864, %r859, %r861;
mov.b32 %f100, %r888;
add.ftz.f32 %f101, %f99, %f100;
mov.b32 %r889, %f101;
shfl.sync.bfly.b32 %r890|%p29, %r889, %r867, %r859, %r861;
mov.b32 %f102, %r890;
add.ftz.f32 %f103, %f101, %f102;
mov.b32 %r891, %f103;
shfl.sync.bfly.b32 %r892|%p30, %r891, %r870, %r859, %r861;
mov.b32 %f104, %r892;
add.ftz.f32 %f105, %f103, %f104;
mov.b32 %r893, %f105;
shfl.sync.bfly.b32 %r894|%p31, %r893, %r873, %r859, %r861;
mov.b32 %f106, %r894;
add.ftz.f32 %f107, %f105, %f106;
st.local.f32 [%rd129+4], %f107;
mov.b32 %r895, %f162;
shfl.sync.bfly.b32 %r896|%p32, %r895, %r860, %r859, %r861;
mov.b32 %f108, %r896;
add.ftz.f32 %f109, %f162, %f108;
mov.b32 %r897, %f109;
shfl.sync.bfly.b32 %r898|%p33, %r897, %r864, %r859, %r861;
mov.b32 %f110, %r898;
add.ftz.f32 %f111, %f109, %f110;
mov.b32 %r899, %f111;
shfl.sync.bfly.b32 %r900|%p34, %r899, %r867, %r859, %r861;
mov.b32 %f112, %r900;
add.ftz.f32 %f113, %f111, %f112;
mov.b32 %r901, %f113;
shfl.sync.bfly.b32 %r902|%p35, %r901, %r870, %r859, %r861;
mov.b32 %f114, %r902;
add.ftz.f32 %f115, %f113, %f114;
mov.b32 %r903, %f115;
shfl.sync.bfly.b32 %r904|%p36, %r903, %r873, %r859, %r861;
mov.b32 %f116, %r904;
add.ftz.f32 %f117, %f115, %f116;
st.local.f32 [%rd121+4], %f117;
bra.uni $L__BB224_12;

$L__BB224_10:
add.u64 %rd125, %SPL, 0;
mov.u32 %r833, 31;
mov.u32 %r834, 16;
mov.u32 %r835, -1;
shfl.sync.bfly.b32 %r836|%p7, %r20, %r834, %r833, %r835;
mov.b32 %f58, %r836;
add.ftz.f32 %f59, %f165, %f58;
mov.b32 %r837, %f59;
mov.u32 %r838, 8;
shfl.sync.bfly.b32 %r839|%p8, %r837, %r838, %r833, %r835;
mov.b32 %f60, %r839;
add.ftz.f32 %f61, %f59, %f60;
mov.b32 %r840, %f61;
mov.u32 %r841, 4;
shfl.sync.bfly.b32 %r842|%p9, %r840, %r841, %r833, %r835;
mov.b32 %f62, %r842;
add.ftz.f32 %f63, %f61, %f62;
mov.b32 %r843, %f63;
mov.u32 %r844, 2;
shfl.sync.bfly.b32 %r845|%p10, %r843, %r844, %r833, %r835;
mov.b32 %f64, %r845;
add.ftz.f32 %f65, %f63, %f64;
mov.b32 %r846, %f65;
mov.u32 %r847, 1;
shfl.sync.bfly.b32 %r848|%p11, %r846, %r847, %r833, %r835;
mov.b32 %f66, %r848;
add.ftz.f32 %f67, %f65, %f66;
st.local.f32 [%rd125], %f67;
mov.b32 %r849, %f164;
shfl.sync.bfly.b32 %r850|%p12, %r849, %r834, %r833, %r835;
mov.b32 %f68, %r850;
add.ftz.f32 %f69, %f164, %f68;
mov.b32 %r851, %f69;
shfl.sync.bfly.b32 %r852|%p13, %r851, %r838, %r833, %r835;
mov.b32 %f70, %r852;
add.ftz.f32 %f71, %f69, %f70;
mov.b32 %r853, %f71;
shfl.sync.bfly.b32 %r854|%p14, %r853, %r841, %r833, %r835;
mov.b32 %f72, %r854;
add.ftz.f32 %f73, %f71, %f72;
mov.b32 %r855, %f73;
shfl.sync.bfly.b32 %r856|%p15, %r855, %r844, %r833, %r835;
mov.b32 %f74, %r856;
add.ftz.f32 %f75, %f73, %f74;
mov.b32 %r857, %f75;
shfl.sync.bfly.b32 %r858|%p16, %r857, %r847, %r833, %r835;
mov.b32 %f76, %r858;
add.ftz.f32 %f77, %f75, %f76;
st.local.f32 [%rd125+4], %f77;

$L__BB224_12:
ld.param.u32 %r915, [_Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7];
mov.u32 %r914, %tid.x;
setp.gt.u32 %p37, %r914, 1;
mov.u32 %r906, %ctaid.x;
shl.b32 %r907, %r906, 1;
add.s32 %r21, %r907, %r914;
setp.ge.u32 %p38, %r21, %r915;
or.pred %p39, %p37, %p38;
@%p39 bra $L__BB224_29;

add.u64 %rd127, %SPL, 0;
ld.param.u64 %rd110, [_Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3];
ld.param.u32 %r917, [_Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13];
mov.u32 %r916, %tid.x;
cvt.u64.u32 %rd19, %r916;
mul.wide.u32 %rd96, %r916, 4;
add.s64 %rd97, %rd127, %rd96;
ld.local.f32 %f168, [%rd97];
mad.lo.s32 %r22, %r3, %r917, %r21;
setp.eq.s64 %p40, %rd110, 0;
@%p40 bra $L__BB224_15;

ld.param.u64 %rd111, [_Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3];
cvta.to.global.u64 %rd98, %rd111;
mul.wide.u32 %rd99, %r22, 4;
add.s64 %rd100, %rd98, %rd99;
ld.global.f32 %f118, [%rd100];
add.ftz.f32 %f168, %f168, %f118;

$L__BB224_15:
setp.eq.s64 %p41, %rd25, 0;
@%p41 bra $L__BB224_28;

add.u64 %rd119, %SPL, 8;
ld.param.u64 %rd113, [_Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+16];
shl.b64 %rd101, %rd19, 2;
add.s64 %rd102, %rd119, %rd101;
ld.local.f32 %f24, [%rd102];
setp.eq.s64 %p42, %rd113, 0;
@%p42 bra $L__BB224_18;

ld.param.u64 %rd114, [_Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+16];
cvta.to.global.u64 %rd103, %rd114;
mul.wide.u32 %rd104, %r22, 4;
add.s64 %rd105, %rd103, %rd104;
ld.global.f32 %f119, [%rd105];
add.ftz.f32 %f24, %f24, %f119;

$L__BB224_18:
ld.param.u32 %r922, [_Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.gt.s32 %p43, %r922, 2;
@%p43 bra $L__BB224_22;

ld.param.u32 %r925, [_Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.eq.s32 %p46, %r925, 1;
@%p46 bra $L__BB224_26;

ld.param.u32 %r926, [_Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.eq.s32 %p47, %r926, 2;
@%p47 bra $L__BB224_21;
bra.uni $L__BB224_27;

$L__BB224_21:
mul.ftz.f32 %f139, %f24, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f140, %f139;
add.ftz.f32 %f141, %f140, 0f3F800000;
div.approx.ftz.f32 %f142, %f24, %f141;
mul.ftz.f32 %f168, %f168, %f142;
bra.uni $L__BB224_28;

$L__BB224_22:
ld.param.u32 %r923, [_Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.eq.s32 %p44, %r923, 3;
@%p44 bra $L__BB224_25;

ld.param.u32 %r924, [_Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.ne.s32 %p45, %r924, 6;
@%p45 bra $L__BB224_27;

ld.param.f32 %f151, [_Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+44];
min.ftz.f32 %f120, %f24, %f151;
neg.ftz.f32 %f121, %f151;
min.ftz.f32 %f122, %f168, %f151;
max.ftz.f32 %f123, %f122, %f121;
mul.ftz.f32 %f124, %f120, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f125, %f124;
add.ftz.f32 %f126, %f125, 0f3F800000;
div.approx.ftz.f32 %f127, %f120, %f126;
mul.ftz.f32 %f168, %f123, %f127;
bra.uni $L__BB224_28;

$L__BB224_27:
mul.ftz.f32 %f168, %f168, %f24;
bra.uni $L__BB224_28;

$L__BB224_26:
mul.ftz.f32 %f143, %f24, 0f3F000000;
mul.ftz.f32 %f144, %f24, 0f3D372713;
fma.rn.ftz.f32 %f145, %f24, %f144, 0f3F800000;
mul.ftz.f32 %f146, %f24, 0f3F4C422A;
mul.ftz.f32 %f147, %f146, %f145;
tanh.approx.f32 %f148, %f147;
add.ftz.f32 %f149, %f148, 0f3F800000;
mul.ftz.f32 %f150, %f143, %f149;
mul.ftz.f32 %f168, %f168, %f150;
bra.uni $L__BB224_28;

$L__BB224_25:
mov.f32 %f128, 0f40E00000;
min.ftz.f32 %f129, %f24, %f128;
min.ftz.f32 %f130, %f168, %f128;
mov.f32 %f131, 0fC0E00000;
max.ftz.f32 %f132, %f130, %f131;
mul.ftz.f32 %f133, %f129, 0f3FD9DB23;
mul.ftz.f32 %f134, %f133, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f135, %f134;
add.ftz.f32 %f136, %f135, 0f3F800000;
div.approx.ftz.f32 %f137, %f129, %f136;
add.ftz.f32 %f138, %f132, 0f3F800000;
mul.ftz.f32 %f168, %f138, %f137;

$L__BB224_28:
ld.param.u64 %rd112, [_Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4];
ld.param.u32 %r921, [_Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10];
ld.param.u32 %r920, [_Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13];
mov.u32 %r919, %ctaid.y;
mov.u32 %r918, %tid.y;
mad.lo.s32 %r912, %r918, %r921, %r21;
mad.lo.s32 %r913, %r919, %r920, %r912;
cvta.to.global.u64 %rd106, %rd112;
mul.wide.u32 %rd107, %r913, 4;
add.s64 %rd108, %rd106, %rd107;
st.global.f32 [%rd108], %f168;

$L__BB224_29:
ret;

}
.entry _Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj(
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2,
.param .align 8 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3[48],
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5,
.param .align 4 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6[12],
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15
)
.maxntid 256, 1, 1
.minnctapersm 1
{
.local .align 8 .b8 __local_depot225[8];
.reg .b64 %SP;
.reg .b64 %SPL;
.reg .pred %p<17>;
.reg .b16 %rs<41>;
.reg .f32 %f<44>;
.reg .b32 %r<465>;
.reg .b64 %rd<65>;


mov.u64 %SPL, __local_depot225;
ld.param.u64 %rd15, [_Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0];
ld.param.u64 %rd16, [_Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1];
ld.param.u64 %rd17, [_Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2];
ld.param.u64 %rd18, [_Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4];
ld.param.u32 %r23, [_Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5];
ld.param.u32 %r15, [_Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7];
ld.param.u32 %r16, [_Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8];
ld.param.u32 %r17, [_Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9];
ld.param.u32 %r18, [_Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10];
ld.param.u32 %r19, [_Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11];
ld.param.u32 %r20, [_Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12];
ld.param.u32 %r21, [_Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13];
ld.param.u32 %r24, [_Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14];
ld.param.u32 %r22, [_Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15];
ld.param.u32 %r14, [_Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+8];
ld.param.u32 %r13, [_Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+4];
ld.param.u32 %r12, [_Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6];
add.u64 %rd1, %SPL, 0;
mov.u32 %r25, %ctaid.x;
shl.b32 %r1, %r25, 1;
shr.u32 %r2, %r23, 8;
mov.u32 %r3, %tid.y;
setp.ge.u32 %p1, %r3, %r24;
@%p1 bra $L__BB225_7;

mov.f32 %f42, 0f00000000;
st.local.v2.f32 [%rd1], {%f42, %f42};
mov.u32 %r4, %tid.x;
shr.u32 %r464, %r4, 3;
setp.ge.u32 %p2, %r464, %r2;
mov.f32 %f43, %f42;
@%p2 bra $L__BB225_5;

mov.u32 %r26, %ctaid.y;
mad.lo.s32 %r27, %r3, %r22, %r26;
cvta.to.global.u64 %rd20, %rd17;
mul.wide.u32 %rd21, %r27, 4;
add.s64 %rd22, %rd20, %rd21;
ld.global.nc.u32 %r28, [%rd22];
mul.lo.s32 %r29, %r28, %r19;
and.b32 %r30, %r4, 7;
cvt.u64.u32 %rd3, %r30;
shl.b32 %r463, %r464, 3;
add.s32 %r31, %r1, 1;
add.s32 %r32, %r464, %r29;
mad.lo.s32 %r33, %r16, %r31, %r32;
mul.wide.s32 %rd4, %r33, 82;
shl.b32 %r34, %r4, 2;
cvt.u64.u32 %rd23, %r34;
and.b64 %rd24, %rd23, 28;
mad.lo.s32 %r35, %r1, %r16, %r464;
add.s32 %r36, %r35, %r29;
mul.wide.s32 %rd5, %r36, 82;
add.s64 %rd25, %rd24, %rd5;
add.s64 %rd6, %rd25, 36;
add.s64 %rd26, %rd4, %rd3;
add.s64 %rd7, %rd26, 66;
add.s64 %rd27, %rd5, %rd3;
add.s64 %rd8, %rd27, 74;
add.s64 %rd28, %rd24, %rd4;
add.s64 %rd9, %rd28, 34;
cvta.to.global.u64 %rd64, %rd15;
mul.lo.s32 %r37, %r3, %r17;
cvt.u64.u32 %rd29, %r37;
mul.hi.u32 %r38, %r26, %r12;
add.s32 %r39, %r38, %r26;
shr.u32 %r40, %r39, %r13;
mul.lo.s32 %r41, %r40, %r14;
sub.s32 %r42, %r26, %r41;
mul.lo.s32 %r43, %r42, %r20;
cvt.u64.u32 %rd30, %r43;
add.s64 %rd11, %rd30, %rd29;
cvta.to.global.u64 %rd12, %rd16;
mov.f32 %f42, 0f00000000;

$L__BB225_3:
cvt.s64.s32 %rd31, %r463;
add.s64 %rd32, %rd11, %rd31;
add.s64 %rd33, %rd32, %rd3;
add.s64 %rd34, %rd64, %rd6;
ld.global.nc.u16 %rs3, [%rd34+-34];
shr.u16 %rs4, %rs3, 8;
cvt.u32.u16 %r142, %rs4;
ld.global.nc.u16 %rs5, [%rd34+-32];
shr.u16 %rs6, %rs5, 8;
cvt.u32.u16 %r143, %rs6;
add.s64 %rd35, %rd64, %rd8;
ld.global.nc.u8 %rs7, [%rd35+-8];
ld.global.nc.u16 %rs8, [%rd34+-2];
ld.global.nc.u16 %rs9, [%rd34];
ld.global.nc.u8 %rs10, [%rd35];
and.b16 %rs11, %rs10, 240;
and.b16 %rs12, %rs10, 15;
shr.u16 %rs13, %rs11, 4;
and.b16 %rs14, %rs3, 255;
cvt.u32.u16 %r144, %rs14;
shl.b16 %rs15, %rs7, 8;
cvt.u32.u16 %r145, %rs15;
and.b32 %r146, %r145, 768;
or.b32 %r147, %r146, %r144;
mul.wide.u32 %rd36, %r147, 8;
mov.u64 %rd37, iq2s_grid;
add.s64 %rd38, %rd37, %rd36;
and.b16 %rs16, %rs8, 255;
mul.wide.u16 %r148, %rs16, 128;
cvt.u32.u16 %r149, %rs8;
shl.b32 %r150, %r149, 21;
or.b32 %r151, %r148, %r150;
and.b32 %r152, %r151, 25166208;
or.b32 %r153, %r152, -2139062144;
add.s32 %r154, %r153, -16843009;
or.b32 %r44, %r154, %r152;

prmt.b32 %r44, %r44, 0, 0xba98;

mul.wide.u16 %r155, %rs16, 8;
shl.b32 %r156, %r149, 17;
or.b32 %r157, %r155, %r156;
and.b32 %r158, %r157, 25166208;
or.b32 %r159, %r158, -2139062144;
add.s32 %r160, %r159, -16843009;
or.b32 %r46, %r160, %r158;

prmt.b32 %r46, %r46, 0, 0xba98;

ld.global.nc.v2.u32 {%r161, %r162}, [%rd38];
mov.u32 %r124, 0;
xor.b32 %r165, %r161, %r44;
or.b32 %r166, %r165, -2139062144;
and.b32 %r167, %r44, 2139062143;
and.b32 %r168, %r161, -2139062144;
sub.s32 %r169, %r166, %r167;
xor.b32 %r170, %r168, %r169;
xor.b32 %r49, %r170, -2139062144;
xor.b32 %r171, %r162, %r46;
or.b32 %r172, %r171, -2139062144;
and.b32 %r173, %r46, 2139062143;
and.b32 %r174, %r162, -2139062144;
sub.s32 %r175, %r172, %r173;
xor.b32 %r176, %r174, %r175;
xor.b32 %r53, %r176, -2139062144;
mul.lo.s64 %rd39, %rd33, 36;
add.s64 %rd40, %rd12, %rd39;
ld.global.nc.u32 %r54, [%rd40+8];
ld.global.nc.u32 %r50, [%rd40+4];

dp4a.s32.s32 %r48, %r49, %r50, %r124;


dp4a.s32.s32 %r52, %r53, %r54, %r48;

mul.wide.u16 %r177, %rs7, 64;
and.b32 %r178, %r177, 768;
or.b32 %r179, %r178, %r142;
mul.wide.u32 %rd41, %r179, 8;
add.s64 %rd42, %rd37, %rd41;
shr.u16 %rs18, %rs8, 8;
mul.wide.u16 %r180, %rs18, 128;
cvt.u32.u16 %r181, %rs18;
shl.b32 %r182, %r181, 21;
or.b32 %r183, %r180, %r182;
and.b32 %r184, %r183, 25166208;
or.b32 %r185, %r184, -2139062144;
add.s32 %r186, %r185, -16843009;
or.b32 %r56, %r186, %r184;

prmt.b32 %r56, %r56, 0, 0xba98;

mul.wide.u16 %r187, %rs18, 8;
shl.b32 %r188, %r181, 17;
or.b32 %r189, %r187, %r188;
and.b32 %r190, %r189, 25166208;
or.b32 %r191, %r190, -2139062144;
add.s32 %r192, %r191, -16843009;
or.b32 %r58, %r192, %r190;

prmt.b32 %r58, %r58, 0, 0xba98;

ld.global.nc.v2.u32 {%r193, %r194}, [%rd42];
xor.b32 %r197, %r193, %r56;
or.b32 %r198, %r197, -2139062144;
and.b32 %r199, %r56, 2139062143;
and.b32 %r200, %r193, -2139062144;
sub.s32 %r201, %r198, %r199;
xor.b32 %r202, %r200, %r201;
xor.b32 %r61, %r202, -2139062144;
xor.b32 %r203, %r194, %r58;
or.b32 %r204, %r203, -2139062144;
and.b32 %r205, %r58, 2139062143;
and.b32 %r206, %r194, -2139062144;
sub.s32 %r207, %r204, %r205;
xor.b32 %r208, %r206, %r207;
xor.b32 %r65, %r208, -2139062144;
ld.global.nc.u32 %r66, [%rd40+16];
ld.global.nc.u32 %r62, [%rd40+12];

dp4a.s32.s32 %r60, %r61, %r62, %r52;


dp4a.s32.s32 %r64, %r65, %r66, %r60;

and.b16 %rs19, %rs5, 255;
cvt.u32.u16 %r209, %rs19;
mul.wide.u16 %r210, %rs7, 16;
and.b32 %r211, %r210, 768;
or.b32 %r212, %r211, %r209;
mul.wide.u32 %rd43, %r212, 8;
add.s64 %rd44, %rd37, %rd43;
and.b16 %rs20, %rs9, 255;
mul.wide.u16 %r213, %rs20, 128;
cvt.u32.u16 %r214, %rs9;
shl.b32 %r215, %r214, 21;
or.b32 %r216, %r213, %r215;
and.b32 %r217, %r216, 25166208;
or.b32 %r218, %r217, -2139062144;
add.s32 %r219, %r218, -16843009;
or.b32 %r68, %r219, %r217;

prmt.b32 %r68, %r68, 0, 0xba98;

mul.wide.u16 %r220, %rs20, 8;
shl.b32 %r221, %r214, 17;
or.b32 %r222, %r220, %r221;
and.b32 %r223, %r222, 25166208;
or.b32 %r224, %r223, -2139062144;
add.s32 %r225, %r224, -16843009;
or.b32 %r70, %r225, %r223;

prmt.b32 %r70, %r70, 0, 0xba98;

ld.global.nc.v2.u32 {%r226, %r227}, [%rd44];
xor.b32 %r230, %r226, %r68;
or.b32 %r231, %r230, -2139062144;
and.b32 %r232, %r68, 2139062143;
and.b32 %r233, %r226, -2139062144;
sub.s32 %r234, %r231, %r232;
xor.b32 %r235, %r233, %r234;
xor.b32 %r73, %r235, -2139062144;
xor.b32 %r236, %r227, %r70;
or.b32 %r237, %r236, -2139062144;
and.b32 %r238, %r70, 2139062143;
and.b32 %r239, %r227, -2139062144;
sub.s32 %r240, %r237, %r238;
xor.b32 %r241, %r239, %r240;
xor.b32 %r77, %r241, -2139062144;
ld.global.nc.u32 %r78, [%rd40+24];
ld.global.nc.u32 %r74, [%rd40+20];

dp4a.s32.s32 %r72, %r73, %r74, %r124;


dp4a.s32.s32 %r76, %r77, %r78, %r72;

mul.wide.u16 %r242, %rs7, 4;
and.b32 %r243, %r242, 768;
or.b32 %r244, %r243, %r143;
mul.wide.u32 %rd45, %r244, 8;
add.s64 %rd46, %rd37, %rd45;
shr.u16 %rs21, %rs9, 8;
mul.wide.u16 %r245, %rs21, 128;
cvt.u32.u16 %r246, %rs21;
shl.b32 %r247, %r246, 21;
or.b32 %r248, %r245, %r247;
and.b32 %r249, %r248, 25166208;
or.b32 %r250, %r249, -2139062144;
add.s32 %r251, %r250, -16843009;
or.b32 %r80, %r251, %r249;

prmt.b32 %r80, %r80, 0, 0xba98;

mul.wide.u16 %r252, %rs21, 8;
shl.b32 %r253, %r246, 17;
or.b32 %r254, %r252, %r253;
and.b32 %r255, %r254, 25166208;
or.b32 %r256, %r255, -2139062144;
add.s32 %r257, %r256, -16843009;
or.b32 %r82, %r257, %r255;

prmt.b32 %r82, %r82, 0, 0xba98;

ld.global.nc.v2.u32 {%r258, %r259}, [%rd46];
xor.b32 %r262, %r258, %r80;
or.b32 %r263, %r262, -2139062144;
and.b32 %r264, %r80, 2139062143;
and.b32 %r265, %r258, -2139062144;
sub.s32 %r266, %r263, %r264;
xor.b32 %r267, %r265, %r266;
xor.b32 %r85, %r267, -2139062144;
xor.b32 %r268, %r259, %r82;
or.b32 %r269, %r268, -2139062144;
and.b32 %r270, %r82, 2139062143;
and.b32 %r271, %r259, -2139062144;
sub.s32 %r272, %r269, %r270;
xor.b32 %r273, %r271, %r272;
xor.b32 %r89, %r273, -2139062144;
ld.global.nc.u32 %r90, [%rd40+32];
ld.global.nc.u32 %r86, [%rd40+28];

dp4a.s32.s32 %r84, %r85, %r86, %r76;


dp4a.s32.s32 %r88, %r89, %r90, %r84;

cvt.u32.u16 %r274, %rs12;
cvt.u32.u16 %r275, %rs13;
mul.lo.s32 %r276, %r64, %r274;
mad.lo.s32 %r277, %r88, %r275, %r276;
add.s32 %r278, %r88, %r64;
shr.u32 %r279, %r278, 31;
add.s32 %r280, %r278, %r279;
shr.s32 %r281, %r280, 1;
add.s32 %r282, %r277, %r281;
shr.s32 %r283, %r282, 31;
shr.u32 %r284, %r283, 30;
add.s32 %r285, %r282, %r284;
shr.s32 %r286, %r285, 2;
add.s64 %rd47, %rd64, %rd5;
ld.global.nc.u16 %rs1, [%rd47];

{ cvt.f32.f16 %f11, %rs1;}


ld.global.nc.u32 %r92, [%rd40];

{.reg .f16 low,high;
mov.b32 {low,high},%r92;
cvt.f32.f16 %f12, low;}


mul.ftz.f32 %f15, %f11, %f12;
cvt.rn.f32.s32 %f16, %r286;
fma.rn.ftz.f32 %f43, %f15, %f16, %f43;
add.s64 %rd48, %rd64, %rd9;
ld.global.nc.u16 %rs22, [%rd48+-32];
shr.u16 %rs23, %rs22, 8;
cvt.u32.u16 %r287, %rs23;
ld.global.nc.u16 %rs24, [%rd48+-30];
shr.u16 %rs25, %rs24, 8;
cvt.u32.u16 %r288, %rs25;
add.s64 %rd49, %rd64, %rd7;
ld.global.nc.u8 %rs26, [%rd49];
ld.global.nc.u16 %rs27, [%rd48];
ld.global.nc.u16 %rs28, [%rd48+2];
ld.global.nc.u8 %rs29, [%rd49+8];
and.b16 %rs30, %rs29, 240;
and.b16 %rs31, %rs29, 15;
shr.u16 %rs32, %rs30, 4;
and.b16 %rs33, %rs22, 255;
cvt.u32.u16 %r289, %rs33;
shl.b16 %rs34, %rs26, 8;
cvt.u32.u16 %r290, %rs34;
and.b32 %r291, %r290, 768;
or.b32 %r292, %r291, %r289;
mul.wide.u32 %rd50, %r292, 8;
add.s64 %rd51, %rd37, %rd50;
and.b16 %rs35, %rs27, 255;
mul.wide.u16 %r293, %rs35, 128;
cvt.u32.u16 %r294, %rs27;
shl.b32 %r295, %r294, 21;
or.b32 %r296, %r293, %r295;
and.b32 %r297, %r296, 25166208;
or.b32 %r298, %r297, -2139062144;
add.s32 %r299, %r298, -16843009;
or.b32 %r93, %r299, %r297;

prmt.b32 %r93, %r93, 0, 0xba98;

mul.wide.u16 %r300, %rs35, 8;
shl.b32 %r301, %r294, 17;
or.b32 %r302, %r300, %r301;
and.b32 %r303, %r302, 25166208;
or.b32 %r304, %r303, -2139062144;
add.s32 %r305, %r304, -16843009;
or.b32 %r95, %r305, %r303;

prmt.b32 %r95, %r95, 0, 0xba98;

ld.global.nc.v2.u32 {%r306, %r307}, [%rd51];
xor.b32 %r310, %r306, %r93;
or.b32 %r311, %r310, -2139062144;
and.b32 %r312, %r93, 2139062143;
and.b32 %r313, %r306, -2139062144;
sub.s32 %r314, %r311, %r312;
xor.b32 %r315, %r313, %r314;
xor.b32 %r98, %r315, -2139062144;
xor.b32 %r316, %r307, %r95;
or.b32 %r317, %r316, -2139062144;
and.b32 %r318, %r95, 2139062143;
and.b32 %r319, %r307, -2139062144;
sub.s32 %r320, %r317, %r318;
xor.b32 %r321, %r319, %r320;
xor.b32 %r102, %r321, -2139062144;

dp4a.s32.s32 %r97, %r98, %r50, %r124;


dp4a.s32.s32 %r101, %r102, %r54, %r97;

mul.wide.u16 %r322, %rs26, 64;
and.b32 %r323, %r322, 768;
or.b32 %r324, %r323, %r287;
mul.wide.u32 %rd52, %r324, 8;
add.s64 %rd53, %rd37, %rd52;
shr.u16 %rs37, %rs27, 8;
mul.wide.u16 %r325, %rs37, 128;
cvt.u32.u16 %r326, %rs37;
shl.b32 %r327, %r326, 21;
or.b32 %r328, %r325, %r327;
and.b32 %r329, %r328, 25166208;
or.b32 %r330, %r329, -2139062144;
add.s32 %r331, %r330, -16843009;
or.b32 %r105, %r331, %r329;

prmt.b32 %r105, %r105, 0, 0xba98;

mul.wide.u16 %r332, %rs37, 8;
shl.b32 %r333, %r326, 17;
or.b32 %r334, %r332, %r333;
and.b32 %r335, %r334, 25166208;
or.b32 %r336, %r335, -2139062144;
add.s32 %r337, %r336, -16843009;
or.b32 %r107, %r337, %r335;

prmt.b32 %r107, %r107, 0, 0xba98;

ld.global.nc.v2.u32 {%r338, %r339}, [%rd53];
xor.b32 %r342, %r338, %r105;
or.b32 %r343, %r342, -2139062144;
and.b32 %r344, %r105, 2139062143;
and.b32 %r345, %r338, -2139062144;
sub.s32 %r346, %r343, %r344;
xor.b32 %r347, %r345, %r346;
xor.b32 %r110, %r347, -2139062144;
xor.b32 %r348, %r339, %r107;
or.b32 %r349, %r348, -2139062144;
and.b32 %r350, %r107, 2139062143;
and.b32 %r351, %r339, -2139062144;
sub.s32 %r352, %r349, %r350;
xor.b32 %r353, %r351, %r352;
xor.b32 %r114, %r353, -2139062144;

dp4a.s32.s32 %r109, %r110, %r62, %r101;


dp4a.s32.s32 %r113, %r114, %r66, %r109;

and.b16 %rs38, %rs24, 255;
cvt.u32.u16 %r354, %rs38;
mul.wide.u16 %r355, %rs26, 16;
and.b32 %r356, %r355, 768;
or.b32 %r357, %r356, %r354;
mul.wide.u32 %rd54, %r357, 8;
add.s64 %rd55, %rd37, %rd54;
and.b16 %rs39, %rs28, 255;
mul.wide.u16 %r358, %rs39, 128;
cvt.u32.u16 %r359, %rs28;
shl.b32 %r360, %r359, 21;
or.b32 %r361, %r358, %r360;
and.b32 %r362, %r361, 25166208;
or.b32 %r363, %r362, -2139062144;
add.s32 %r364, %r363, -16843009;
or.b32 %r117, %r364, %r362;

prmt.b32 %r117, %r117, 0, 0xba98;

mul.wide.u16 %r365, %rs39, 8;
shl.b32 %r366, %r359, 17;
or.b32 %r367, %r365, %r366;
and.b32 %r368, %r367, 25166208;
or.b32 %r369, %r368, -2139062144;
add.s32 %r370, %r369, -16843009;
or.b32 %r119, %r370, %r368;

prmt.b32 %r119, %r119, 0, 0xba98;

ld.global.nc.v2.u32 {%r371, %r372}, [%rd55];
xor.b32 %r375, %r371, %r117;
or.b32 %r376, %r375, -2139062144;
and.b32 %r377, %r117, 2139062143;
and.b32 %r378, %r371, -2139062144;
sub.s32 %r379, %r376, %r377;
xor.b32 %r380, %r378, %r379;
xor.b32 %r122, %r380, -2139062144;
xor.b32 %r381, %r372, %r119;
or.b32 %r382, %r381, -2139062144;
and.b32 %r383, %r119, 2139062143;
and.b32 %r384, %r372, -2139062144;
sub.s32 %r385, %r382, %r383;
xor.b32 %r386, %r384, %r385;
xor.b32 %r126, %r386, -2139062144;

dp4a.s32.s32 %r121, %r122, %r74, %r124;


dp4a.s32.s32 %r125, %r126, %r78, %r121;

mul.wide.u16 %r387, %rs26, 4;
and.b32 %r388, %r387, 768;
or.b32 %r389, %r388, %r288;
mul.wide.u32 %rd56, %r389, 8;
add.s64 %rd57, %rd37, %rd56;
shr.u16 %rs40, %rs28, 8;
mul.wide.u16 %r390, %rs40, 128;
cvt.u32.u16 %r391, %rs40;
shl.b32 %r392, %r391, 21;
or.b32 %r393, %r390, %r392;
and.b32 %r394, %r393, 25166208;
or.b32 %r395, %r394, -2139062144;
add.s32 %r396, %r395, -16843009;
or.b32 %r129, %r396, %r394;

prmt.b32 %r129, %r129, 0, 0xba98;

mul.wide.u16 %r397, %rs40, 8;
shl.b32 %r398, %r391, 17;
or.b32 %r399, %r397, %r398;
and.b32 %r400, %r399, 25166208;
or.b32 %r401, %r400, -2139062144;
add.s32 %r402, %r401, -16843009;
or.b32 %r131, %r402, %r400;

prmt.b32 %r131, %r131, 0, 0xba98;

ld.global.nc.v2.u32 {%r403, %r404}, [%rd57];
xor.b32 %r407, %r403, %r129;
or.b32 %r408, %r407, -2139062144;
and.b32 %r409, %r129, 2139062143;
and.b32 %r410, %r403, -2139062144;
sub.s32 %r411, %r408, %r409;
xor.b32 %r412, %r410, %r411;
xor.b32 %r134, %r412, -2139062144;
xor.b32 %r413, %r404, %r131;
or.b32 %r414, %r413, -2139062144;
and.b32 %r415, %r131, 2139062143;
and.b32 %r416, %r404, -2139062144;
sub.s32 %r417, %r414, %r415;
xor.b32 %r418, %r416, %r417;
xor.b32 %r138, %r418, -2139062144;

dp4a.s32.s32 %r133, %r134, %r86, %r125;


dp4a.s32.s32 %r137, %r138, %r90, %r133;

cvt.u32.u16 %r419, %rs31;
cvt.u32.u16 %r420, %rs32;
mul.lo.s32 %r421, %r113, %r419;
mad.lo.s32 %r422, %r137, %r420, %r421;
add.s32 %r423, %r137, %r113;
shr.u32 %r424, %r423, 31;
add.s32 %r425, %r423, %r424;
shr.s32 %r426, %r425, 1;
add.s32 %r427, %r422, %r426;
shr.s32 %r428, %r427, 31;
shr.u32 %r429, %r428, 30;
add.s32 %r430, %r427, %r429;
shr.s32 %r431, %r430, 2;
add.s64 %rd58, %rd64, %rd4;
ld.global.nc.u16 %rs2, [%rd58];

{ cvt.f32.f16 %f13, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r92;
cvt.f32.f16 %f14, low;}


mul.ftz.f32 %f17, %f13, %f14;
cvt.rn.f32.s32 %f18, %r431;
fma.rn.ftz.f32 %f42, %f17, %f18, %f42;
add.s32 %r463, %r463, 32;
add.s64 %rd64, %rd64, 328;
add.s32 %r464, %r464, 4;
setp.lt.u32 %p3, %r464, %r2;
@%p3 bra $L__BB225_3;

st.local.v2.f32 [%rd1], {%f43, %f42};

$L__BB225_5:
mov.b32 %r432, %f43;
mov.u32 %r433, 31;
mov.u32 %r434, 16;
mov.u32 %r435, -1;
shfl.sync.bfly.b32 %r436|%p4, %r432, %r434, %r433, %r435;
mov.b32 %f19, %r436;
add.ftz.f32 %f20, %f43, %f19;
mov.b32 %r437, %f20;
mov.u32 %r438, 8;
shfl.sync.bfly.b32 %r439|%p5, %r437, %r438, %r433, %r435;
mov.b32 %f21, %r439;
add.ftz.f32 %f22, %f20, %f21;
mov.b32 %r440, %f22;
mov.u32 %r441, 4;
shfl.sync.bfly.b32 %r442|%p6, %r440, %r441, %r433, %r435;
mov.b32 %f23, %r442;
add.ftz.f32 %f24, %f22, %f23;
mov.b32 %r443, %f24;
mov.u32 %r444, 2;
shfl.sync.bfly.b32 %r445|%p7, %r443, %r444, %r433, %r435;
mov.b32 %f25, %r445;
add.ftz.f32 %f26, %f24, %f25;
mov.b32 %r446, %f26;
mov.u32 %r447, 1;
shfl.sync.bfly.b32 %r448|%p8, %r446, %r447, %r433, %r435;
mov.b32 %f27, %r448;
add.ftz.f32 %f28, %f26, %f27;
st.local.f32 [%rd1], %f28;
mov.b32 %r449, %f42;
shfl.sync.bfly.b32 %r450|%p9, %r449, %r434, %r433, %r435;
mov.b32 %f29, %r450;
add.ftz.f32 %f30, %f42, %f29;
mov.b32 %r451, %f30;
shfl.sync.bfly.b32 %r452|%p10, %r451, %r438, %r433, %r435;
mov.b32 %f31, %r452;
add.ftz.f32 %f32, %f30, %f31;
mov.b32 %r453, %f32;
shfl.sync.bfly.b32 %r454|%p11, %r453, %r441, %r433, %r435;
mov.b32 %f33, %r454;
add.ftz.f32 %f34, %f32, %f33;
mov.b32 %r455, %f34;
shfl.sync.bfly.b32 %r456|%p12, %r455, %r444, %r433, %r435;
mov.b32 %f35, %r456;
add.ftz.f32 %f36, %f34, %f35;
mov.b32 %r457, %f36;
shfl.sync.bfly.b32 %r458|%p13, %r457, %r447, %r433, %r435;
mov.b32 %f37, %r458;
add.ftz.f32 %f38, %f36, %f37;
st.local.f32 [%rd1+4], %f38;
add.s32 %r11, %r1, %r4;
setp.ge.u32 %p14, %r11, %r15;
setp.gt.u32 %p15, %r4, 1;
or.pred %p16, %p15, %p14;
@%p16 bra $L__BB225_7;

mul.wide.u32 %rd59, %r4, 4;
add.s64 %rd60, %rd1, %rd59;
ld.local.f32 %f39, [%rd60];
mov.u32 %r459, %ctaid.y;
mad.lo.s32 %r461, %r3, %r18, %r11;
mad.lo.s32 %r462, %r459, %r21, %r461;
cvta.to.global.u64 %rd61, %rd18;
mul.wide.u32 %rd62, %r462, 4;
add.s64 %rd63, %rd61, %rd62;
st.global.f32 [%rd63], %f39;

$L__BB225_7:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<320>;
.reg .b16 %rs<153>;
.reg .f32 %f<1005>;
.reg .b32 %r<2415>;
.reg .b64 %rd<194>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate[1536];

ld.param.u64 %rd31, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd32, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd33, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u32 %r46, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r50, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r51, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r55, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r56, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r57, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r61, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r62, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r63, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r59, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r58, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r53, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r52, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r49, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r48, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r47, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
ld.param.u64 %rd36, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+16];
ld.param.u64 %rd35, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+8];
ld.param.u64 %rd34, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3];
mov.u32 %r1, %tid.y;
shl.b32 %r64, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r64, %r2;
mov.u32 %r65, %ctaid.x;
shl.b32 %r4, %r65, 2;
mov.u32 %r5, %ctaid.y;
setp.eq.s64 %p1, %rd33, 0;
@%p1 bra $L__BB226_2;

cvta.to.global.u64 %rd40, %rd33;
mul.wide.u32 %rd41, %r5, 4;
add.s64 %rd42, %rd40, %rd41;
ld.global.nc.u32 %r2411, [%rd42];
mul.hi.u32 %r66, %r5, %r47;
add.s32 %r67, %r66, %r5;
shr.u32 %r68, %r67, %r48;
mul.lo.s32 %r69, %r68, %r49;
sub.s32 %r2412, %r5, %r69;
bra.uni $L__BB226_3;

$L__BB226_2:
mul.hi.u32 %r70, %r5, %r52;
add.s32 %r71, %r70, %r5;
shr.u32 %r2411, %r71, %r53;
mov.u32 %r2412, %r5;

$L__BB226_3:
mov.u32 %r11, %ctaid.z;
selp.b32 %r12, %r5, %r2411, %p1;
setp.ne.s32 %p3, %r1, 0;
setp.gt.u32 %p4, %r2, 3;
or.pred %p5, %p4, %p3;
add.s32 %r72, %r4, %r2;
setp.ge.u32 %p6, %r72, %r51;
mov.f32 %f27, 0f00000000;
or.pred %p7, %p5, %p6;
mov.f32 %f4, %f27;
mov.f32 %f5, %f27;
@%p7 bra $L__BB226_8;

setp.eq.s64 %p8, %rd34, 0;
mov.f32 %f4, 0f00000000;
mov.f32 %f5, %f4;
@%p8 bra $L__BB226_6;

cvta.to.global.u64 %rd43, %rd34;
mul.lo.s32 %r73, %r11, %r63;
cvt.u64.u32 %rd44, %r73;
mul.lo.s32 %r74, %r12, %r57;
cvt.u64.u32 %rd45, %r74;
cvt.u64.u32 %rd46, %r2;
cvt.s64.s32 %rd47, %r4;
add.s64 %rd48, %rd47, %rd46;
add.s64 %rd49, %rd48, %rd44;
add.s64 %rd50, %rd49, %rd45;
shl.b64 %rd51, %rd50, 2;
add.s64 %rd52, %rd43, %rd51;
ld.global.f32 %f5, [%rd52];

$L__BB226_6:
setp.eq.s64 %p9, %rd36, 0;
setp.eq.s64 %p10, %rd35, 0;
or.pred %p11, %p9, %p10;
@%p11 bra $L__BB226_8;

cvta.to.global.u64 %rd53, %rd36;
mul.lo.s32 %r75, %r11, %r63;
cvt.u64.u32 %rd54, %r75;
mul.lo.s32 %r76, %r12, %r57;
cvt.u64.u32 %rd55, %r76;
cvt.u64.u32 %rd56, %r2;
cvt.s64.s32 %rd57, %r4;
add.s64 %rd58, %rd57, %rd56;
add.s64 %rd59, %rd58, %rd54;
add.s64 %rd60, %rd59, %rd55;
shl.b64 %rd61, %rd60, 2;
add.s64 %rd62, %rd53, %rd61;
ld.global.f32 %f4, [%rd62];

$L__BB226_8:
mul.lo.s32 %r77, %r11, %r62;
cvt.u64.u32 %rd63, %r77;
mul.lo.s32 %r78, %r2412, %r56;
cvt.u64.u32 %rd64, %r78;
add.s64 %rd4, %rd64, %rd63;
mul.hi.u32 %r79, %r11, %r58;
add.s32 %r80, %r79, %r11;
shr.u32 %r13, %r80, %r59;
shr.s32 %r81, %r3, 31;
shr.u32 %r82, %r81, 29;
add.s32 %r83, %r3, %r82;
shr.s32 %r2414, %r83, 3;
shr.u32 %r84, %r46, 8;
setp.ge.s32 %p12, %r2414, %r84;
mov.f32 %f28, %f27;
mov.f32 %f29, %f27;
mov.f32 %f30, %f27;
mov.f32 %f31, %f27;
mov.f32 %f32, %f27;
mov.f32 %f33, %f27;
mov.f32 %f34, %f27;
@%p12 bra $L__BB226_19;

mul.lo.s32 %r85, %r13, %r61;
mul.lo.s32 %r86, %r2411, %r55;
and.b32 %r90, %r83, -8;
sub.s32 %r91, %r3, %r90;
shl.b32 %r92, %r91, 1;
cvt.s64.s32 %rd5, %r91;
shl.b32 %r2413, %r2414, 3;
add.s32 %r93, %r2414, %r85;
add.s32 %r16, %r93, %r86;
add.s32 %r94, %r4, 1;
mad.lo.s32 %r95, %r50, %r94, %r16;
mul.wide.s32 %rd65, %r95, 82;
mad.lo.s32 %r96, %r4, %r50, %r2414;
add.s32 %r97, %r96, %r85;
add.s32 %r98, %r97, %r86;
mul.wide.s32 %rd6, %r98, 82;
mul.wide.s32 %rd66, %r92, 2;
add.s64 %rd67, %rd6, %rd66;
add.s64 %rd7, %rd67, 36;
add.s64 %rd68, %rd65, %rd5;
add.s64 %rd8, %rd68, 66;
add.s64 %rd9, %rd6, %rd5;
add.s64 %rd69, %rd65, %rd66;
add.s64 %rd10, %rd69, 34;
add.s32 %r99, %r4, 3;
mad.lo.s32 %r100, %r50, %r99, %r16;
mul.wide.s32 %rd70, %r100, 82;
add.s64 %rd71, %rd70, %rd66;
add.s64 %rd11, %rd71, 36;
add.s64 %rd72, %rd70, %rd5;
add.s64 %rd12, %rd72, 74;
add.s32 %r101, %r4, 2;
mad.lo.s32 %r102, %r50, %r101, %r16;
mul.wide.s32 %rd73, %r102, 82;
add.s64 %rd74, %rd73, %rd5;
add.s64 %rd13, %rd74, 66;
add.s64 %rd14, %rd73, %rd66;
cvta.to.global.u64 %rd193, %rd35;
cvta.to.global.u64 %rd192, %rd31;
mov.f32 %f27, 0f00000000;
cvta.to.global.u64 %rd75, %rd32;
mov.f32 %f28, %f27;
mov.f32 %f29, %f27;
mov.f32 %f30, %f27;
mov.f32 %f31, %f27;
mov.f32 %f32, %f27;
mov.f32 %f33, %f27;
mov.f32 %f34, %f27;

$L__BB226_10:
cvt.u64.u32 %rd76, %r2413;
add.s64 %rd77, %rd4, %rd76;
add.s64 %rd78, %rd192, %rd7;
ld.global.nc.u16 %rs2, [%rd78+-34];
shr.u16 %rs3, %rs2, 8;
cvt.u32.u16 %r152, %rs3;
ld.global.nc.u16 %rs4, [%rd78+-32];
shr.u16 %rs5, %rs4, 8;
cvt.u32.u16 %r153, %rs5;
add.s64 %rd79, %rd192, %rd9;
ld.global.nc.u8 %rs6, [%rd79+66];
ld.global.nc.u16 %rs7, [%rd78+-2];
ld.global.nc.u16 %rs8, [%rd78];
ld.global.nc.u8 %rs9, [%rd79+74];
and.b16 %rs10, %rs9, 240;
and.b16 %rs11, %rs9, 15;
shr.u16 %rs12, %rs10, 4;
add.s64 %rd80, %rd77, %rd5;
and.b16 %rs13, %rs2, 255;
cvt.u32.u16 %r154, %rs13;
shl.b16 %rs14, %rs6, 8;
cvt.u32.u16 %r155, %rs14;
and.b32 %r156, %r155, 768;
or.b32 %r157, %r156, %r154;
mul.wide.u32 %rd81, %r157, 8;
mov.u64 %rd82, iq2s_grid;
add.s64 %rd83, %rd82, %rd81;
and.b16 %rs15, %rs7, 255;
mul.wide.u16 %r158, %rs15, 128;
cvt.u32.u16 %r159, %rs7;
shl.b32 %r160, %r159, 21;
or.b32 %r161, %r158, %r160;
and.b32 %r162, %r161, 25166208;
or.b32 %r163, %r162, -2139062144;
add.s32 %r164, %r163, -16843009;
or.b32 %r103, %r164, %r162;

prmt.b32 %r103, %r103, 0, 0xba98;

mul.wide.u16 %r165, %rs15, 8;
shl.b32 %r166, %r159, 17;
or.b32 %r167, %r165, %r166;
and.b32 %r168, %r167, 25166208;
or.b32 %r169, %r168, -2139062144;
add.s32 %r170, %r169, -16843009;
or.b32 %r105, %r170, %r168;

prmt.b32 %r105, %r105, 0, 0xba98;

ld.global.nc.v2.u32 {%r171, %r172}, [%rd83];
mov.u32 %r134, 0;
xor.b32 %r175, %r171, %r103;
or.b32 %r176, %r175, -2139062144;
and.b32 %r177, %r103, 2139062143;
and.b32 %r178, %r171, -2139062144;
sub.s32 %r179, %r176, %r177;
xor.b32 %r180, %r178, %r179;
xor.b32 %r108, %r180, -2139062144;
xor.b32 %r181, %r172, %r105;
or.b32 %r182, %r181, -2139062144;
and.b32 %r183, %r105, 2139062143;
and.b32 %r184, %r172, -2139062144;
sub.s32 %r185, %r182, %r183;
xor.b32 %r186, %r184, %r185;
xor.b32 %r112, %r186, -2139062144;
mul.lo.s64 %rd84, %rd80, 36;
add.s64 %rd85, %rd75, %rd84;
ld.global.nc.u32 %r19, [%rd85+8];
ld.global.nc.u32 %r20, [%rd85+4];

dp4a.s32.s32 %r107, %r108, %r20, %r134;


dp4a.s32.s32 %r111, %r112, %r19, %r107;

mul.wide.u16 %r187, %rs6, 64;
and.b32 %r188, %r187, 768;
or.b32 %r189, %r188, %r152;
mul.wide.u32 %rd86, %r189, 8;
add.s64 %rd87, %rd82, %rd86;
shr.u16 %rs17, %rs7, 8;
mul.wide.u16 %r190, %rs17, 128;
cvt.u32.u16 %r191, %rs17;
shl.b32 %r192, %r191, 21;
or.b32 %r193, %r190, %r192;
and.b32 %r194, %r193, 25166208;
or.b32 %r195, %r194, -2139062144;
add.s32 %r196, %r195, -16843009;
or.b32 %r115, %r196, %r194;

prmt.b32 %r115, %r115, 0, 0xba98;

mul.wide.u16 %r197, %rs17, 8;
shl.b32 %r198, %r191, 17;
or.b32 %r199, %r197, %r198;
and.b32 %r200, %r199, 25166208;
or.b32 %r201, %r200, -2139062144;
add.s32 %r202, %r201, -16843009;
or.b32 %r117, %r202, %r200;

prmt.b32 %r117, %r117, 0, 0xba98;

ld.global.nc.v2.u32 {%r203, %r204}, [%rd87];
xor.b32 %r207, %r203, %r115;
or.b32 %r208, %r207, -2139062144;
and.b32 %r209, %r115, 2139062143;
and.b32 %r210, %r203, -2139062144;
sub.s32 %r211, %r208, %r209;
xor.b32 %r212, %r210, %r211;
xor.b32 %r120, %r212, -2139062144;
xor.b32 %r213, %r204, %r117;
or.b32 %r214, %r213, -2139062144;
and.b32 %r215, %r117, 2139062143;
and.b32 %r216, %r204, -2139062144;
sub.s32 %r217, %r214, %r215;
xor.b32 %r218, %r216, %r217;
xor.b32 %r124, %r218, -2139062144;
ld.global.nc.u32 %r21, [%rd85+16];
ld.global.nc.u32 %r22, [%rd85+12];

dp4a.s32.s32 %r119, %r120, %r22, %r111;


dp4a.s32.s32 %r123, %r124, %r21, %r119;

and.b16 %rs18, %rs4, 255;
cvt.u32.u16 %r219, %rs18;
mul.wide.u16 %r220, %rs6, 16;
and.b32 %r221, %r220, 768;
or.b32 %r222, %r221, %r219;
mul.wide.u32 %rd88, %r222, 8;
add.s64 %rd89, %rd82, %rd88;
and.b16 %rs19, %rs8, 255;
mul.wide.u16 %r223, %rs19, 128;
cvt.u32.u16 %r224, %rs8;
shl.b32 %r225, %r224, 21;
or.b32 %r226, %r223, %r225;
and.b32 %r227, %r226, 25166208;
or.b32 %r228, %r227, -2139062144;
add.s32 %r229, %r228, -16843009;
or.b32 %r127, %r229, %r227;

prmt.b32 %r127, %r127, 0, 0xba98;

mul.wide.u16 %r230, %rs19, 8;
shl.b32 %r231, %r224, 17;
or.b32 %r232, %r230, %r231;
and.b32 %r233, %r232, 25166208;
or.b32 %r234, %r233, -2139062144;
add.s32 %r235, %r234, -16843009;
or.b32 %r129, %r235, %r233;

prmt.b32 %r129, %r129, 0, 0xba98;

ld.global.nc.v2.u32 {%r236, %r237}, [%rd89];
xor.b32 %r240, %r236, %r127;
or.b32 %r241, %r240, -2139062144;
and.b32 %r242, %r127, 2139062143;
and.b32 %r243, %r236, -2139062144;
sub.s32 %r244, %r241, %r242;
xor.b32 %r245, %r243, %r244;
xor.b32 %r132, %r245, -2139062144;
xor.b32 %r246, %r237, %r129;
or.b32 %r247, %r246, -2139062144;
and.b32 %r248, %r129, 2139062143;
and.b32 %r249, %r237, -2139062144;
sub.s32 %r250, %r247, %r248;
xor.b32 %r251, %r249, %r250;
xor.b32 %r136, %r251, -2139062144;
ld.global.nc.u32 %r23, [%rd85+24];
ld.global.nc.u32 %r24, [%rd85+20];

dp4a.s32.s32 %r131, %r132, %r24, %r134;


dp4a.s32.s32 %r135, %r136, %r23, %r131;

mul.wide.u16 %r252, %rs6, 4;
and.b32 %r253, %r252, 768;
or.b32 %r254, %r253, %r153;
mul.wide.u32 %rd90, %r254, 8;
add.s64 %rd91, %rd82, %rd90;
shr.u16 %rs20, %rs8, 8;
mul.wide.u16 %r255, %rs20, 128;
cvt.u32.u16 %r256, %rs20;
shl.b32 %r257, %r256, 21;
or.b32 %r258, %r255, %r257;
and.b32 %r259, %r258, 25166208;
or.b32 %r260, %r259, -2139062144;
add.s32 %r261, %r260, -16843009;
or.b32 %r139, %r261, %r259;

prmt.b32 %r139, %r139, 0, 0xba98;

mul.wide.u16 %r262, %rs20, 8;
shl.b32 %r263, %r256, 17;
or.b32 %r264, %r262, %r263;
and.b32 %r265, %r264, 25166208;
or.b32 %r266, %r265, -2139062144;
add.s32 %r267, %r266, -16843009;
or.b32 %r141, %r267, %r265;

prmt.b32 %r141, %r141, 0, 0xba98;

ld.global.nc.v2.u32 {%r268, %r269}, [%rd91];
xor.b32 %r272, %r268, %r139;
or.b32 %r273, %r272, -2139062144;
and.b32 %r274, %r139, 2139062143;
and.b32 %r275, %r268, -2139062144;
sub.s32 %r276, %r273, %r274;
xor.b32 %r277, %r275, %r276;
xor.b32 %r144, %r277, -2139062144;
xor.b32 %r278, %r269, %r141;
or.b32 %r279, %r278, -2139062144;
and.b32 %r280, %r141, 2139062143;
and.b32 %r281, %r269, -2139062144;
sub.s32 %r282, %r279, %r280;
xor.b32 %r283, %r281, %r282;
xor.b32 %r148, %r283, -2139062144;
ld.global.nc.u32 %r25, [%rd85+32];
ld.global.nc.u32 %r26, [%rd85+28];

dp4a.s32.s32 %r143, %r144, %r26, %r135;


dp4a.s32.s32 %r147, %r148, %r25, %r143;

cvt.u32.u16 %r284, %rs11;
cvt.u32.u16 %r285, %rs12;
mul.lo.s32 %r286, %r123, %r284;
mad.lo.s32 %r287, %r147, %r285, %r286;
add.s32 %r288, %r147, %r123;
shr.u32 %r289, %r288, 31;
add.s32 %r290, %r288, %r289;
shr.s32 %r291, %r290, 1;
add.s32 %r292, %r287, %r291;
shr.s32 %r293, %r292, 31;
shr.u32 %r294, %r293, 30;
add.s32 %r295, %r292, %r294;
shr.s32 %r296, %r295, 2;
add.s64 %rd21, %rd192, %rd6;
ld.global.nc.u16 %rs1, [%rd21];

{ cvt.f32.f16 %f101, %rs1;}


ld.global.nc.u32 %r27, [%rd85];

{.reg .f16 low,high;
mov.b32 {low,high},%r27;
cvt.f32.f16 %f102, low;}


mul.ftz.f32 %f103, %f101, %f102;
cvt.rn.f32.s32 %f104, %r296;
fma.rn.ftz.f32 %f30, %f103, %f104, %f30;
setp.eq.s64 %p13, %rd35, 0;
@%p13 bra $L__BB226_12;

mov.u64 %rd183, iq2s_grid;
add.s64 %rd92, %rd193, %rd7;
ld.global.u16 %rs22, [%rd92+-34];
shr.u16 %rs23, %rs22, 8;
cvt.u32.u16 %r346, %rs23;
ld.global.u16 %rs24, [%rd92+-32];
shr.u16 %rs25, %rs24, 8;
cvt.u32.u16 %r347, %rs25;
add.s64 %rd93, %rd193, %rd9;
ld.global.u16 %rs26, [%rd92+-2];
ld.global.u16 %rs27, [%rd92];
ld.global.u8 %rs28, [%rd93+74];
and.b16 %rs29, %rs28, 15;
shr.u16 %rs30, %rs28, 4;
and.b16 %rs31, %rs22, 255;
cvt.u32.u16 %r348, %rs31;
ld.global.u8 %rs32, [%rd93+66];
shl.b16 %rs33, %rs32, 8;
cvt.u32.u16 %r349, %rs33;
and.b32 %r350, %r349, 768;
or.b32 %r351, %r350, %r348;
mul.wide.u32 %rd94, %r351, 8;
add.s64 %rd96, %rd183, %rd94;
and.b16 %rs34, %rs26, 255;
mul.wide.u16 %r352, %rs34, 128;
cvt.u32.u16 %r353, %rs26;
shl.b32 %r354, %r353, 21;
or.b32 %r355, %r352, %r354;
and.b32 %r356, %r355, 25166208;
or.b32 %r357, %r356, -2139062144;
add.s32 %r358, %r357, -16843009;
or.b32 %r297, %r358, %r356;

prmt.b32 %r297, %r297, 0, 0xba98;

mul.wide.u16 %r359, %rs34, 8;
shl.b32 %r360, %r353, 17;
or.b32 %r361, %r359, %r360;
and.b32 %r362, %r361, 25166208;
or.b32 %r363, %r362, -2139062144;
add.s32 %r364, %r363, -16843009;
or.b32 %r299, %r364, %r362;

prmt.b32 %r299, %r299, 0, 0xba98;

ld.global.nc.v2.u32 {%r365, %r366}, [%rd96];
xor.b32 %r369, %r365, %r297;
or.b32 %r370, %r369, -2139062144;
and.b32 %r371, %r297, 2139062143;
and.b32 %r372, %r365, -2139062144;
sub.s32 %r373, %r370, %r371;
xor.b32 %r374, %r372, %r373;
xor.b32 %r302, %r374, -2139062144;
xor.b32 %r375, %r366, %r299;
or.b32 %r376, %r375, -2139062144;
and.b32 %r377, %r299, 2139062143;
and.b32 %r378, %r366, -2139062144;
sub.s32 %r379, %r376, %r377;
xor.b32 %r380, %r378, %r379;
xor.b32 %r306, %r380, -2139062144;

dp4a.s32.s32 %r301, %r302, %r20, %r134;


dp4a.s32.s32 %r305, %r306, %r19, %r301;

mul.wide.u16 %r381, %rs32, 64;
and.b32 %r382, %r381, 768;
or.b32 %r383, %r382, %r346;
mul.wide.u32 %rd97, %r383, 8;
add.s64 %rd98, %rd183, %rd97;
shr.u16 %rs35, %rs26, 8;
mul.wide.u16 %r384, %rs35, 128;
cvt.u32.u16 %r385, %rs35;
shl.b32 %r386, %r385, 21;
or.b32 %r387, %r384, %r386;
and.b32 %r388, %r387, 25166208;
or.b32 %r389, %r388, -2139062144;
add.s32 %r390, %r389, -16843009;
or.b32 %r309, %r390, %r388;

prmt.b32 %r309, %r309, 0, 0xba98;

mul.wide.u16 %r391, %rs35, 8;
shl.b32 %r392, %r385, 17;
or.b32 %r393, %r391, %r392;
and.b32 %r394, %r393, 25166208;
or.b32 %r395, %r394, -2139062144;
add.s32 %r396, %r395, -16843009;
or.b32 %r311, %r396, %r394;

prmt.b32 %r311, %r311, 0, 0xba98;

ld.global.nc.v2.u32 {%r397, %r398}, [%rd98];
xor.b32 %r401, %r397, %r309;
or.b32 %r402, %r401, -2139062144;
and.b32 %r403, %r309, 2139062143;
and.b32 %r404, %r397, -2139062144;
sub.s32 %r405, %r402, %r403;
xor.b32 %r406, %r404, %r405;
xor.b32 %r314, %r406, -2139062144;
xor.b32 %r407, %r398, %r311;
or.b32 %r408, %r407, -2139062144;
and.b32 %r409, %r311, 2139062143;
and.b32 %r410, %r398, -2139062144;
sub.s32 %r411, %r408, %r409;
xor.b32 %r412, %r410, %r411;
xor.b32 %r318, %r412, -2139062144;

dp4a.s32.s32 %r313, %r314, %r22, %r305;


dp4a.s32.s32 %r317, %r318, %r21, %r313;

and.b16 %rs36, %rs24, 255;
cvt.u32.u16 %r413, %rs36;
mul.wide.u16 %r414, %rs32, 16;
and.b32 %r415, %r414, 768;
or.b32 %r416, %r415, %r413;
mul.wide.u32 %rd99, %r416, 8;
add.s64 %rd100, %rd183, %rd99;
and.b16 %rs37, %rs27, 255;
mul.wide.u16 %r417, %rs37, 128;
cvt.u32.u16 %r418, %rs27;
shl.b32 %r419, %r418, 21;
or.b32 %r420, %r417, %r419;
and.b32 %r421, %r420, 25166208;
or.b32 %r422, %r421, -2139062144;
add.s32 %r423, %r422, -16843009;
or.b32 %r321, %r423, %r421;

prmt.b32 %r321, %r321, 0, 0xba98;

mul.wide.u16 %r424, %rs37, 8;
shl.b32 %r425, %r418, 17;
or.b32 %r426, %r424, %r425;
and.b32 %r427, %r426, 25166208;
or.b32 %r428, %r427, -2139062144;
add.s32 %r429, %r428, -16843009;
or.b32 %r323, %r429, %r427;

prmt.b32 %r323, %r323, 0, 0xba98;

ld.global.nc.v2.u32 {%r430, %r431}, [%rd100];
xor.b32 %r434, %r430, %r321;
or.b32 %r435, %r434, -2139062144;
and.b32 %r436, %r321, 2139062143;
and.b32 %r437, %r430, -2139062144;
sub.s32 %r438, %r435, %r436;
xor.b32 %r439, %r437, %r438;
xor.b32 %r326, %r439, -2139062144;
xor.b32 %r440, %r431, %r323;
or.b32 %r441, %r440, -2139062144;
and.b32 %r442, %r323, 2139062143;
and.b32 %r443, %r431, -2139062144;
sub.s32 %r444, %r441, %r442;
xor.b32 %r445, %r443, %r444;
xor.b32 %r330, %r445, -2139062144;

dp4a.s32.s32 %r325, %r326, %r24, %r134;


dp4a.s32.s32 %r329, %r330, %r23, %r325;

mul.wide.u16 %r446, %rs32, 4;
and.b32 %r447, %r446, 768;
or.b32 %r448, %r447, %r347;
mul.wide.u32 %rd101, %r448, 8;
add.s64 %rd102, %rd183, %rd101;
shr.u16 %rs38, %rs27, 8;
mul.wide.u16 %r449, %rs38, 128;
cvt.u32.u16 %r450, %rs38;
shl.b32 %r451, %r450, 21;
or.b32 %r452, %r449, %r451;
and.b32 %r453, %r452, 25166208;
or.b32 %r454, %r453, -2139062144;
add.s32 %r455, %r454, -16843009;
or.b32 %r333, %r455, %r453;

prmt.b32 %r333, %r333, 0, 0xba98;

mul.wide.u16 %r456, %rs38, 8;
shl.b32 %r457, %r450, 17;
or.b32 %r458, %r456, %r457;
and.b32 %r459, %r458, 25166208;
or.b32 %r460, %r459, -2139062144;
add.s32 %r461, %r460, -16843009;
or.b32 %r335, %r461, %r459;

prmt.b32 %r335, %r335, 0, 0xba98;

ld.global.nc.v2.u32 {%r462, %r463}, [%rd102];
xor.b32 %r466, %r462, %r333;
or.b32 %r467, %r466, -2139062144;
and.b32 %r468, %r333, 2139062143;
and.b32 %r469, %r462, -2139062144;
sub.s32 %r470, %r467, %r468;
xor.b32 %r471, %r469, %r470;
xor.b32 %r338, %r471, -2139062144;
xor.b32 %r472, %r463, %r335;
or.b32 %r473, %r472, -2139062144;
and.b32 %r474, %r335, 2139062143;
and.b32 %r475, %r463, -2139062144;
sub.s32 %r476, %r473, %r474;
xor.b32 %r477, %r475, %r476;
xor.b32 %r342, %r477, -2139062144;

dp4a.s32.s32 %r337, %r338, %r26, %r329;


dp4a.s32.s32 %r341, %r342, %r25, %r337;

cvt.u32.u16 %r478, %rs29;
cvt.u32.u16 %r479, %rs30;
mul.lo.s32 %r480, %r317, %r478;
mad.lo.s32 %r481, %r341, %r479, %r480;
add.s32 %r482, %r341, %r317;
shr.u32 %r483, %r482, 31;
add.s32 %r484, %r482, %r483;
shr.s32 %r485, %r484, 1;
add.s32 %r486, %r481, %r485;
shr.s32 %r487, %r486, 31;
shr.u32 %r488, %r487, 30;
add.s32 %r489, %r486, %r488;
shr.s32 %r490, %r489, 2;
add.s64 %rd103, %rd193, %rd6;
ld.global.u16 %rs21, [%rd103];

{ cvt.f32.f16 %f105, %rs21;}



{.reg .f16 low,high;
mov.b32 {low,high},%r27;
cvt.f32.f16 %f106, low;}


mul.ftz.f32 %f107, %f105, %f106;
cvt.rn.f32.s32 %f108, %r490;
fma.rn.ftz.f32 %f34, %f107, %f108, %f34;

$L__BB226_12:
add.s64 %rd189, %rd192, %rd9;
add.s64 %rd188, %rd189, 66;
add.s64 %rd187, %rd192, %rd7;
add.s64 %rd186, %rd187, -34;
mov.u64 %rd180, iq2s_grid;
mul.wide.s32 %rd104, %r50, 82;
add.s64 %rd22, %rd186, %rd104;
ld.global.nc.u16 %rs40, [%rd22];
shr.u16 %rs41, %rs40, 8;
cvt.u32.u16 %r540, %rs41;
ld.global.nc.u16 %rs42, [%rd22+2];
shr.u16 %rs43, %rs42, 8;
cvt.u32.u16 %r541, %rs43;
add.s64 %rd23, %rd188, %rd104;
ld.global.nc.u8 %rs44, [%rd23];
ld.global.nc.u16 %rs45, [%rd22+32];
ld.global.nc.u16 %rs46, [%rd22+34];
ld.global.nc.u8 %rs47, [%rd23+8];
and.b16 %rs48, %rs47, 240;
and.b16 %rs49, %rs47, 15;
shr.u16 %rs50, %rs48, 4;
and.b16 %rs51, %rs40, 255;
cvt.u32.u16 %r542, %rs51;
shl.b16 %rs52, %rs44, 8;
cvt.u32.u16 %r543, %rs52;
and.b32 %r544, %r543, 768;
or.b32 %r545, %r544, %r542;
mul.wide.u32 %rd105, %r545, 8;
add.s64 %rd107, %rd180, %rd105;
and.b16 %rs53, %rs45, 255;
mul.wide.u16 %r546, %rs53, 128;
cvt.u32.u16 %r547, %rs45;
shl.b32 %r548, %r547, 21;
or.b32 %r549, %r546, %r548;
and.b32 %r550, %r549, 25166208;
or.b32 %r551, %r550, -2139062144;
add.s32 %r552, %r551, -16843009;
or.b32 %r491, %r552, %r550;

prmt.b32 %r491, %r491, 0, 0xba98;

mul.wide.u16 %r553, %rs53, 8;
shl.b32 %r554, %r547, 17;
or.b32 %r555, %r553, %r554;
and.b32 %r556, %r555, 25166208;
or.b32 %r557, %r556, -2139062144;
add.s32 %r558, %r557, -16843009;
or.b32 %r493, %r558, %r556;

prmt.b32 %r493, %r493, 0, 0xba98;

ld.global.nc.v2.u32 {%r559, %r560}, [%rd107];
xor.b32 %r563, %r559, %r491;
or.b32 %r564, %r563, -2139062144;
and.b32 %r565, %r491, 2139062143;
and.b32 %r566, %r559, -2139062144;
sub.s32 %r567, %r564, %r565;
xor.b32 %r568, %r566, %r567;
xor.b32 %r496, %r568, -2139062144;
xor.b32 %r569, %r560, %r493;
or.b32 %r570, %r569, -2139062144;
and.b32 %r571, %r493, 2139062143;
and.b32 %r572, %r560, -2139062144;
sub.s32 %r573, %r570, %r571;
xor.b32 %r574, %r572, %r573;
xor.b32 %r500, %r574, -2139062144;

dp4a.s32.s32 %r495, %r496, %r20, %r134;


dp4a.s32.s32 %r499, %r500, %r19, %r495;

mul.wide.u16 %r575, %rs44, 64;
and.b32 %r576, %r575, 768;
or.b32 %r577, %r576, %r540;
mul.wide.u32 %rd108, %r577, 8;
add.s64 %rd109, %rd180, %rd108;
shr.u16 %rs55, %rs45, 8;
mul.wide.u16 %r578, %rs55, 128;
cvt.u32.u16 %r579, %rs55;
shl.b32 %r580, %r579, 21;
or.b32 %r581, %r578, %r580;
and.b32 %r582, %r581, 25166208;
or.b32 %r583, %r582, -2139062144;
add.s32 %r584, %r583, -16843009;
or.b32 %r503, %r584, %r582;

prmt.b32 %r503, %r503, 0, 0xba98;

mul.wide.u16 %r585, %rs55, 8;
shl.b32 %r586, %r579, 17;
or.b32 %r587, %r585, %r586;
and.b32 %r588, %r587, 25166208;
or.b32 %r589, %r588, -2139062144;
add.s32 %r590, %r589, -16843009;
or.b32 %r505, %r590, %r588;

prmt.b32 %r505, %r505, 0, 0xba98;

ld.global.nc.v2.u32 {%r591, %r592}, [%rd109];
xor.b32 %r595, %r591, %r503;
or.b32 %r596, %r595, -2139062144;
and.b32 %r597, %r503, 2139062143;
and.b32 %r598, %r591, -2139062144;
sub.s32 %r599, %r596, %r597;
xor.b32 %r600, %r598, %r599;
xor.b32 %r508, %r600, -2139062144;
xor.b32 %r601, %r592, %r505;
or.b32 %r602, %r601, -2139062144;
and.b32 %r603, %r505, 2139062143;
and.b32 %r604, %r592, -2139062144;
sub.s32 %r605, %r602, %r603;
xor.b32 %r606, %r604, %r605;
xor.b32 %r512, %r606, -2139062144;

dp4a.s32.s32 %r507, %r508, %r22, %r499;


dp4a.s32.s32 %r511, %r512, %r21, %r507;

and.b16 %rs56, %rs42, 255;
cvt.u32.u16 %r607, %rs56;
mul.wide.u16 %r608, %rs44, 16;
and.b32 %r609, %r608, 768;
or.b32 %r610, %r609, %r607;
mul.wide.u32 %rd110, %r610, 8;
add.s64 %rd111, %rd180, %rd110;
and.b16 %rs57, %rs46, 255;
mul.wide.u16 %r611, %rs57, 128;
cvt.u32.u16 %r612, %rs46;
shl.b32 %r613, %r612, 21;
or.b32 %r614, %r611, %r613;
and.b32 %r615, %r614, 25166208;
or.b32 %r616, %r615, -2139062144;
add.s32 %r617, %r616, -16843009;
or.b32 %r515, %r617, %r615;

prmt.b32 %r515, %r515, 0, 0xba98;

mul.wide.u16 %r618, %rs57, 8;
shl.b32 %r619, %r612, 17;
or.b32 %r620, %r618, %r619;
and.b32 %r621, %r620, 25166208;
or.b32 %r622, %r621, -2139062144;
add.s32 %r623, %r622, -16843009;
or.b32 %r517, %r623, %r621;

prmt.b32 %r517, %r517, 0, 0xba98;

ld.global.nc.v2.u32 {%r624, %r625}, [%rd111];
xor.b32 %r628, %r624, %r515;
or.b32 %r629, %r628, -2139062144;
and.b32 %r630, %r515, 2139062143;
and.b32 %r631, %r624, -2139062144;
sub.s32 %r632, %r629, %r630;
xor.b32 %r633, %r631, %r632;
xor.b32 %r520, %r633, -2139062144;
xor.b32 %r634, %r625, %r517;
or.b32 %r635, %r634, -2139062144;
and.b32 %r636, %r517, 2139062143;
and.b32 %r637, %r625, -2139062144;
sub.s32 %r638, %r635, %r636;
xor.b32 %r639, %r637, %r638;
xor.b32 %r524, %r639, -2139062144;

dp4a.s32.s32 %r519, %r520, %r24, %r134;


dp4a.s32.s32 %r523, %r524, %r23, %r519;

mul.wide.u16 %r640, %rs44, 4;
and.b32 %r641, %r640, 768;
or.b32 %r642, %r641, %r541;
mul.wide.u32 %rd112, %r642, 8;
add.s64 %rd113, %rd180, %rd112;
shr.u16 %rs58, %rs46, 8;
mul.wide.u16 %r643, %rs58, 128;
cvt.u32.u16 %r644, %rs58;
shl.b32 %r645, %r644, 21;
or.b32 %r646, %r643, %r645;
and.b32 %r647, %r646, 25166208;
or.b32 %r648, %r647, -2139062144;
add.s32 %r649, %r648, -16843009;
or.b32 %r527, %r649, %r647;

prmt.b32 %r527, %r527, 0, 0xba98;

mul.wide.u16 %r650, %rs58, 8;
shl.b32 %r651, %r644, 17;
or.b32 %r652, %r650, %r651;
and.b32 %r653, %r652, 25166208;
or.b32 %r654, %r653, -2139062144;
add.s32 %r655, %r654, -16843009;
or.b32 %r529, %r655, %r653;

prmt.b32 %r529, %r529, 0, 0xba98;

ld.global.nc.v2.u32 {%r656, %r657}, [%rd113];
xor.b32 %r660, %r656, %r527;
or.b32 %r661, %r660, -2139062144;
and.b32 %r662, %r527, 2139062143;
and.b32 %r663, %r656, -2139062144;
sub.s32 %r664, %r661, %r662;
xor.b32 %r665, %r663, %r664;
xor.b32 %r532, %r665, -2139062144;
xor.b32 %r666, %r657, %r529;
or.b32 %r667, %r666, -2139062144;
and.b32 %r668, %r529, 2139062143;
and.b32 %r669, %r657, -2139062144;
sub.s32 %r670, %r667, %r668;
xor.b32 %r671, %r669, %r670;
xor.b32 %r536, %r671, -2139062144;

dp4a.s32.s32 %r531, %r532, %r26, %r523;


dp4a.s32.s32 %r535, %r536, %r25, %r531;

cvt.u32.u16 %r672, %rs49;
cvt.u32.u16 %r673, %rs50;
mul.lo.s32 %r674, %r511, %r672;
mad.lo.s32 %r675, %r535, %r673, %r674;
add.s32 %r676, %r535, %r511;
shr.u32 %r677, %r676, 31;
add.s32 %r678, %r676, %r677;
shr.s32 %r679, %r678, 1;
add.s32 %r680, %r675, %r679;
shr.s32 %r681, %r680, 31;
shr.u32 %r682, %r681, 30;
add.s32 %r683, %r680, %r682;
shr.s32 %r684, %r683, 2;
add.s64 %rd24, %rd21, %rd104;
ld.global.nc.u16 %rs39, [%rd24];

{ cvt.f32.f16 %f109, %rs39;}



{.reg .f16 low,high;
mov.b32 {low,high},%r27;
cvt.f32.f16 %f110, low;}


mul.ftz.f32 %f111, %f109, %f110;
cvt.rn.f32.s32 %f112, %r684;
fma.rn.ftz.f32 %f29, %f111, %f112, %f29;
@%p13 bra $L__BB226_14;

mov.u64 %rd182, iq2s_grid;
add.s64 %rd114, %rd193, %rd10;
ld.global.u16 %rs60, [%rd114+-32];
shr.u16 %rs61, %rs60, 8;
cvt.u32.u16 %r734, %rs61;
ld.global.u16 %rs62, [%rd114+-30];
shr.u16 %rs63, %rs62, 8;
cvt.u32.u16 %r735, %rs63;
add.s64 %rd115, %rd193, %rd8;
ld.global.u16 %rs64, [%rd114];
ld.global.u16 %rs65, [%rd114+2];
ld.global.u8 %rs66, [%rd115+8];
and.b16 %rs67, %rs66, 15;
shr.u16 %rs68, %rs66, 4;
and.b16 %rs69, %rs60, 255;
cvt.u32.u16 %r736, %rs69;
ld.global.u8 %rs70, [%rd115];
shl.b16 %rs71, %rs70, 8;
cvt.u32.u16 %r737, %rs71;
and.b32 %r738, %r737, 768;
or.b32 %r739, %r738, %r736;
mul.wide.u32 %rd116, %r739, 8;
add.s64 %rd118, %rd182, %rd116;
and.b16 %rs72, %rs64, 255;
mul.wide.u16 %r740, %rs72, 128;
cvt.u32.u16 %r741, %rs64;
shl.b32 %r742, %r741, 21;
or.b32 %r743, %r740, %r742;
and.b32 %r744, %r743, 25166208;
or.b32 %r745, %r744, -2139062144;
add.s32 %r746, %r745, -16843009;
or.b32 %r685, %r746, %r744;

prmt.b32 %r685, %r685, 0, 0xba98;

mul.wide.u16 %r747, %rs72, 8;
shl.b32 %r748, %r741, 17;
or.b32 %r749, %r747, %r748;
and.b32 %r750, %r749, 25166208;
or.b32 %r751, %r750, -2139062144;
add.s32 %r752, %r751, -16843009;
or.b32 %r687, %r752, %r750;

prmt.b32 %r687, %r687, 0, 0xba98;

ld.global.nc.v2.u32 {%r753, %r754}, [%rd118];
mov.u32 %r716, 0;
xor.b32 %r757, %r753, %r685;
or.b32 %r758, %r757, -2139062144;
and.b32 %r759, %r685, 2139062143;
and.b32 %r760, %r753, -2139062144;
sub.s32 %r761, %r758, %r759;
xor.b32 %r762, %r760, %r761;
xor.b32 %r690, %r762, -2139062144;
xor.b32 %r763, %r754, %r687;
or.b32 %r764, %r763, -2139062144;
and.b32 %r765, %r687, 2139062143;
and.b32 %r766, %r754, -2139062144;
sub.s32 %r767, %r764, %r765;
xor.b32 %r768, %r766, %r767;
xor.b32 %r694, %r768, -2139062144;

dp4a.s32.s32 %r689, %r690, %r20, %r716;


dp4a.s32.s32 %r693, %r694, %r19, %r689;

mul.wide.u16 %r769, %rs70, 64;
and.b32 %r770, %r769, 768;
or.b32 %r771, %r770, %r734;
mul.wide.u32 %rd119, %r771, 8;
add.s64 %rd120, %rd182, %rd119;
shr.u16 %rs73, %rs64, 8;
mul.wide.u16 %r772, %rs73, 128;
cvt.u32.u16 %r773, %rs73;
shl.b32 %r774, %r773, 21;
or.b32 %r775, %r772, %r774;
and.b32 %r776, %r775, 25166208;
or.b32 %r777, %r776, -2139062144;
add.s32 %r778, %r777, -16843009;
or.b32 %r697, %r778, %r776;

prmt.b32 %r697, %r697, 0, 0xba98;

mul.wide.u16 %r779, %rs73, 8;
shl.b32 %r780, %r773, 17;
or.b32 %r781, %r779, %r780;
and.b32 %r782, %r781, 25166208;
or.b32 %r783, %r782, -2139062144;
add.s32 %r784, %r783, -16843009;
or.b32 %r699, %r784, %r782;

prmt.b32 %r699, %r699, 0, 0xba98;

ld.global.nc.v2.u32 {%r785, %r786}, [%rd120];
xor.b32 %r789, %r785, %r697;
or.b32 %r790, %r789, -2139062144;
and.b32 %r791, %r697, 2139062143;
and.b32 %r792, %r785, -2139062144;
sub.s32 %r793, %r790, %r791;
xor.b32 %r794, %r792, %r793;
xor.b32 %r702, %r794, -2139062144;
xor.b32 %r795, %r786, %r699;
or.b32 %r796, %r795, -2139062144;
and.b32 %r797, %r699, 2139062143;
and.b32 %r798, %r786, -2139062144;
sub.s32 %r799, %r796, %r797;
xor.b32 %r800, %r798, %r799;
xor.b32 %r706, %r800, -2139062144;

dp4a.s32.s32 %r701, %r702, %r22, %r693;


dp4a.s32.s32 %r705, %r706, %r21, %r701;

and.b16 %rs74, %rs62, 255;
cvt.u32.u16 %r801, %rs74;
mul.wide.u16 %r802, %rs70, 16;
and.b32 %r803, %r802, 768;
or.b32 %r804, %r803, %r801;
mul.wide.u32 %rd121, %r804, 8;
add.s64 %rd122, %rd182, %rd121;
and.b16 %rs75, %rs65, 255;
mul.wide.u16 %r805, %rs75, 128;
cvt.u32.u16 %r806, %rs65;
shl.b32 %r807, %r806, 21;
or.b32 %r808, %r805, %r807;
and.b32 %r809, %r808, 25166208;
or.b32 %r810, %r809, -2139062144;
add.s32 %r811, %r810, -16843009;
or.b32 %r709, %r811, %r809;

prmt.b32 %r709, %r709, 0, 0xba98;

mul.wide.u16 %r812, %rs75, 8;
shl.b32 %r813, %r806, 17;
or.b32 %r814, %r812, %r813;
and.b32 %r815, %r814, 25166208;
or.b32 %r816, %r815, -2139062144;
add.s32 %r817, %r816, -16843009;
or.b32 %r711, %r817, %r815;

prmt.b32 %r711, %r711, 0, 0xba98;

ld.global.nc.v2.u32 {%r818, %r819}, [%rd122];
xor.b32 %r822, %r818, %r709;
or.b32 %r823, %r822, -2139062144;
and.b32 %r824, %r709, 2139062143;
and.b32 %r825, %r818, -2139062144;
sub.s32 %r826, %r823, %r824;
xor.b32 %r827, %r825, %r826;
xor.b32 %r714, %r827, -2139062144;
xor.b32 %r828, %r819, %r711;
or.b32 %r829, %r828, -2139062144;
and.b32 %r830, %r711, 2139062143;
and.b32 %r831, %r819, -2139062144;
sub.s32 %r832, %r829, %r830;
xor.b32 %r833, %r831, %r832;
xor.b32 %r718, %r833, -2139062144;

dp4a.s32.s32 %r713, %r714, %r24, %r716;


dp4a.s32.s32 %r717, %r718, %r23, %r713;

mul.wide.u16 %r834, %rs70, 4;
and.b32 %r835, %r834, 768;
or.b32 %r836, %r835, %r735;
mul.wide.u32 %rd123, %r836, 8;
add.s64 %rd124, %rd182, %rd123;
shr.u16 %rs76, %rs65, 8;
mul.wide.u16 %r837, %rs76, 128;
cvt.u32.u16 %r838, %rs76;
shl.b32 %r839, %r838, 21;
or.b32 %r840, %r837, %r839;
and.b32 %r841, %r840, 25166208;
or.b32 %r842, %r841, -2139062144;
add.s32 %r843, %r842, -16843009;
or.b32 %r721, %r843, %r841;

prmt.b32 %r721, %r721, 0, 0xba98;

mul.wide.u16 %r844, %rs76, 8;
shl.b32 %r845, %r838, 17;
or.b32 %r846, %r844, %r845;
and.b32 %r847, %r846, 25166208;
or.b32 %r848, %r847, -2139062144;
add.s32 %r849, %r848, -16843009;
or.b32 %r723, %r849, %r847;

prmt.b32 %r723, %r723, 0, 0xba98;

ld.global.nc.v2.u32 {%r850, %r851}, [%rd124];
xor.b32 %r854, %r850, %r721;
or.b32 %r855, %r854, -2139062144;
and.b32 %r856, %r721, 2139062143;
and.b32 %r857, %r850, -2139062144;
sub.s32 %r858, %r855, %r856;
xor.b32 %r859, %r857, %r858;
xor.b32 %r726, %r859, -2139062144;
xor.b32 %r860, %r851, %r723;
or.b32 %r861, %r860, -2139062144;
and.b32 %r862, %r723, 2139062143;
and.b32 %r863, %r851, -2139062144;
sub.s32 %r864, %r861, %r862;
xor.b32 %r865, %r863, %r864;
xor.b32 %r730, %r865, -2139062144;

dp4a.s32.s32 %r725, %r726, %r26, %r717;


dp4a.s32.s32 %r729, %r730, %r25, %r725;

cvt.u32.u16 %r866, %rs67;
cvt.u32.u16 %r867, %rs68;
mul.lo.s32 %r868, %r705, %r866;
mad.lo.s32 %r869, %r729, %r867, %r868;
add.s32 %r870, %r729, %r705;
shr.u32 %r871, %r870, 31;
add.s32 %r872, %r870, %r871;
shr.s32 %r873, %r872, 1;
add.s32 %r874, %r869, %r873;
shr.s32 %r875, %r874, 31;
shr.u32 %r876, %r875, 30;
add.s32 %r877, %r874, %r876;
shr.s32 %r878, %r877, 2;
or.b32 %r881, %r4, 1;
mad.lo.s32 %r882, %r50, %r881, %r16;
mul.wide.s32 %rd125, %r882, 82;
add.s64 %rd126, %rd193, %rd125;
ld.global.u16 %rs59, [%rd126];

{ cvt.f32.f16 %f113, %rs59;}



{.reg .f16 low,high;
mov.b32 {low,high},%r27;
cvt.f32.f16 %f114, low;}


mul.ftz.f32 %f115, %f113, %f114;
cvt.rn.f32.s32 %f116, %r878;
fma.rn.ftz.f32 %f33, %f115, %f116, %f33;

$L__BB226_14:
mul.wide.s32 %rd190, %r50, 82;
mov.u64 %rd181, iq2s_grid;
add.s64 %rd25, %rd22, %rd190;
ld.global.nc.u16 %rs78, [%rd25];
shr.u16 %rs79, %rs78, 8;
cvt.u32.u16 %r932, %rs79;
ld.global.nc.u16 %rs80, [%rd25+2];
shr.u16 %rs81, %rs80, 8;
cvt.u32.u16 %r933, %rs81;
add.s64 %rd26, %rd23, %rd190;
ld.global.nc.u8 %rs82, [%rd26];
ld.global.nc.u16 %rs83, [%rd25+32];
ld.global.nc.u16 %rs84, [%rd25+34];
ld.global.nc.u8 %rs85, [%rd26+8];
and.b16 %rs86, %rs85, 240;
and.b16 %rs87, %rs85, 15;
shr.u16 %rs88, %rs86, 4;
and.b16 %rs89, %rs78, 255;
cvt.u32.u16 %r934, %rs89;
shl.b16 %rs90, %rs82, 8;
cvt.u32.u16 %r935, %rs90;
and.b32 %r936, %r935, 768;
or.b32 %r937, %r936, %r934;
mul.wide.u32 %rd128, %r937, 8;
add.s64 %rd130, %rd181, %rd128;
and.b16 %rs91, %rs83, 255;
mul.wide.u16 %r938, %rs91, 128;
cvt.u32.u16 %r939, %rs83;
shl.b32 %r940, %r939, 21;
or.b32 %r941, %r938, %r940;
and.b32 %r942, %r941, 25166208;
or.b32 %r943, %r942, -2139062144;
add.s32 %r944, %r943, -16843009;
or.b32 %r883, %r944, %r942;

prmt.b32 %r883, %r883, 0, 0xba98;

mul.wide.u16 %r945, %rs91, 8;
shl.b32 %r946, %r939, 17;
or.b32 %r947, %r945, %r946;
and.b32 %r948, %r947, 25166208;
or.b32 %r949, %r948, -2139062144;
add.s32 %r950, %r949, -16843009;
or.b32 %r885, %r950, %r948;

prmt.b32 %r885, %r885, 0, 0xba98;

ld.global.nc.v2.u32 {%r951, %r952}, [%rd130];
mov.u32 %r914, 0;
xor.b32 %r955, %r951, %r883;
or.b32 %r956, %r955, -2139062144;
and.b32 %r957, %r883, 2139062143;
and.b32 %r958, %r951, -2139062144;
sub.s32 %r959, %r956, %r957;
xor.b32 %r960, %r958, %r959;
xor.b32 %r888, %r960, -2139062144;
xor.b32 %r961, %r952, %r885;
or.b32 %r962, %r961, -2139062144;
and.b32 %r963, %r885, 2139062143;
and.b32 %r964, %r952, -2139062144;
sub.s32 %r965, %r962, %r963;
xor.b32 %r966, %r964, %r965;
xor.b32 %r892, %r966, -2139062144;

dp4a.s32.s32 %r887, %r888, %r20, %r914;


dp4a.s32.s32 %r891, %r892, %r19, %r887;

mul.wide.u16 %r967, %rs82, 64;
and.b32 %r968, %r967, 768;
or.b32 %r969, %r968, %r932;
mul.wide.u32 %rd131, %r969, 8;
add.s64 %rd132, %rd181, %rd131;
shr.u16 %rs93, %rs83, 8;
mul.wide.u16 %r970, %rs93, 128;
cvt.u32.u16 %r971, %rs93;
shl.b32 %r972, %r971, 21;
or.b32 %r973, %r970, %r972;
and.b32 %r974, %r973, 25166208;
or.b32 %r975, %r974, -2139062144;
add.s32 %r976, %r975, -16843009;
or.b32 %r895, %r976, %r974;

prmt.b32 %r895, %r895, 0, 0xba98;

mul.wide.u16 %r977, %rs93, 8;
shl.b32 %r978, %r971, 17;
or.b32 %r979, %r977, %r978;
and.b32 %r980, %r979, 25166208;
or.b32 %r981, %r980, -2139062144;
add.s32 %r982, %r981, -16843009;
or.b32 %r897, %r982, %r980;

prmt.b32 %r897, %r897, 0, 0xba98;

ld.global.nc.v2.u32 {%r983, %r984}, [%rd132];
xor.b32 %r987, %r983, %r895;
or.b32 %r988, %r987, -2139062144;
and.b32 %r989, %r895, 2139062143;
and.b32 %r990, %r983, -2139062144;
sub.s32 %r991, %r988, %r989;
xor.b32 %r992, %r990, %r991;
xor.b32 %r900, %r992, -2139062144;
xor.b32 %r993, %r984, %r897;
or.b32 %r994, %r993, -2139062144;
and.b32 %r995, %r897, 2139062143;
and.b32 %r996, %r984, -2139062144;
sub.s32 %r997, %r994, %r995;
xor.b32 %r998, %r996, %r997;
xor.b32 %r904, %r998, -2139062144;

dp4a.s32.s32 %r899, %r900, %r22, %r891;


dp4a.s32.s32 %r903, %r904, %r21, %r899;

and.b16 %rs94, %rs80, 255;
cvt.u32.u16 %r999, %rs94;
mul.wide.u16 %r1000, %rs82, 16;
and.b32 %r1001, %r1000, 768;
or.b32 %r1002, %r1001, %r999;
mul.wide.u32 %rd133, %r1002, 8;
add.s64 %rd134, %rd181, %rd133;
and.b16 %rs95, %rs84, 255;
mul.wide.u16 %r1003, %rs95, 128;
cvt.u32.u16 %r1004, %rs84;
shl.b32 %r1005, %r1004, 21;
or.b32 %r1006, %r1003, %r1005;
and.b32 %r1007, %r1006, 25166208;
or.b32 %r1008, %r1007, -2139062144;
add.s32 %r1009, %r1008, -16843009;
or.b32 %r907, %r1009, %r1007;

prmt.b32 %r907, %r907, 0, 0xba98;

mul.wide.u16 %r1010, %rs95, 8;
shl.b32 %r1011, %r1004, 17;
or.b32 %r1012, %r1010, %r1011;
and.b32 %r1013, %r1012, 25166208;
or.b32 %r1014, %r1013, -2139062144;
add.s32 %r1015, %r1014, -16843009;
or.b32 %r909, %r1015, %r1013;

prmt.b32 %r909, %r909, 0, 0xba98;

ld.global.nc.v2.u32 {%r1016, %r1017}, [%rd134];
xor.b32 %r1020, %r1016, %r907;
or.b32 %r1021, %r1020, -2139062144;
and.b32 %r1022, %r907, 2139062143;
and.b32 %r1023, %r1016, -2139062144;
sub.s32 %r1024, %r1021, %r1022;
xor.b32 %r1025, %r1023, %r1024;
xor.b32 %r912, %r1025, -2139062144;
xor.b32 %r1026, %r1017, %r909;
or.b32 %r1027, %r1026, -2139062144;
and.b32 %r1028, %r909, 2139062143;
and.b32 %r1029, %r1017, -2139062144;
sub.s32 %r1030, %r1027, %r1028;
xor.b32 %r1031, %r1029, %r1030;
xor.b32 %r916, %r1031, -2139062144;

dp4a.s32.s32 %r911, %r912, %r24, %r914;


dp4a.s32.s32 %r915, %r916, %r23, %r911;

mul.wide.u16 %r1032, %rs82, 4;
and.b32 %r1033, %r1032, 768;
or.b32 %r1034, %r1033, %r933;
mul.wide.u32 %rd135, %r1034, 8;
add.s64 %rd136, %rd181, %rd135;
shr.u16 %rs96, %rs84, 8;
mul.wide.u16 %r1035, %rs96, 128;
cvt.u32.u16 %r1036, %rs96;
shl.b32 %r1037, %r1036, 21;
or.b32 %r1038, %r1035, %r1037;
and.b32 %r1039, %r1038, 25166208;
or.b32 %r1040, %r1039, -2139062144;
add.s32 %r1041, %r1040, -16843009;
or.b32 %r919, %r1041, %r1039;

prmt.b32 %r919, %r919, 0, 0xba98;

mul.wide.u16 %r1042, %rs96, 8;
shl.b32 %r1043, %r1036, 17;
or.b32 %r1044, %r1042, %r1043;
and.b32 %r1045, %r1044, 25166208;
or.b32 %r1046, %r1045, -2139062144;
add.s32 %r1047, %r1046, -16843009;
or.b32 %r921, %r1047, %r1045;

prmt.b32 %r921, %r921, 0, 0xba98;

ld.global.nc.v2.u32 {%r1048, %r1049}, [%rd136];
xor.b32 %r1052, %r1048, %r919;
or.b32 %r1053, %r1052, -2139062144;
and.b32 %r1054, %r919, 2139062143;
and.b32 %r1055, %r1048, -2139062144;
sub.s32 %r1056, %r1053, %r1054;
xor.b32 %r1057, %r1055, %r1056;
xor.b32 %r924, %r1057, -2139062144;
xor.b32 %r1058, %r1049, %r921;
or.b32 %r1059, %r1058, -2139062144;
and.b32 %r1060, %r921, 2139062143;
and.b32 %r1061, %r1049, -2139062144;
sub.s32 %r1062, %r1059, %r1060;
xor.b32 %r1063, %r1061, %r1062;
xor.b32 %r928, %r1063, -2139062144;

dp4a.s32.s32 %r923, %r924, %r26, %r915;


dp4a.s32.s32 %r927, %r928, %r25, %r923;

cvt.u32.u16 %r1064, %rs87;
cvt.u32.u16 %r1065, %rs88;
mul.lo.s32 %r1066, %r903, %r1064;
mad.lo.s32 %r1067, %r927, %r1065, %r1066;
add.s32 %r1068, %r927, %r903;
shr.u32 %r1069, %r1068, 31;
add.s32 %r1070, %r1068, %r1069;
shr.s32 %r1071, %r1070, 1;
add.s32 %r1072, %r1067, %r1071;
shr.s32 %r1073, %r1072, 31;
shr.u32 %r1074, %r1073, 30;
add.s32 %r1075, %r1072, %r1074;
shr.s32 %r1076, %r1075, 2;
add.s64 %rd27, %rd24, %rd190;
ld.global.nc.u16 %rs77, [%rd27];

{ cvt.f32.f16 %f117, %rs77;}



{.reg .f16 low,high;
mov.b32 {low,high},%r27;
cvt.f32.f16 %f118, low;}


mul.ftz.f32 %f119, %f117, %f118;
cvt.rn.f32.s32 %f120, %r1076;
fma.rn.ftz.f32 %f28, %f119, %f120, %f28;
@%p13 bra $L__BB226_16;

mov.u64 %rd185, iq2s_grid;
add.s64 %rd137, %rd193, %rd14;
ld.global.u16 %rs98, [%rd137+2];
shr.u16 %rs99, %rs98, 8;
cvt.u32.u16 %r1126, %rs99;
ld.global.u16 %rs100, [%rd137+4];
shr.u16 %rs101, %rs100, 8;
cvt.u32.u16 %r1127, %rs101;
add.s64 %rd138, %rd193, %rd13;
ld.global.u16 %rs102, [%rd137+34];
ld.global.u16 %rs103, [%rd137+36];
ld.global.u8 %rs104, [%rd138+8];
and.b16 %rs105, %rs104, 15;
shr.u16 %rs106, %rs104, 4;
and.b16 %rs107, %rs98, 255;
cvt.u32.u16 %r1128, %rs107;
ld.global.u8 %rs108, [%rd138];
shl.b16 %rs109, %rs108, 8;
cvt.u32.u16 %r1129, %rs109;
and.b32 %r1130, %r1129, 768;
or.b32 %r1131, %r1130, %r1128;
mul.wide.u32 %rd139, %r1131, 8;
add.s64 %rd141, %rd185, %rd139;
and.b16 %rs110, %rs102, 255;
mul.wide.u16 %r1132, %rs110, 128;
cvt.u32.u16 %r1133, %rs102;
shl.b32 %r1134, %r1133, 21;
or.b32 %r1135, %r1132, %r1134;
and.b32 %r1136, %r1135, 25166208;
or.b32 %r1137, %r1136, -2139062144;
add.s32 %r1138, %r1137, -16843009;
or.b32 %r1077, %r1138, %r1136;

prmt.b32 %r1077, %r1077, 0, 0xba98;

mul.wide.u16 %r1139, %rs110, 8;
shl.b32 %r1140, %r1133, 17;
or.b32 %r1141, %r1139, %r1140;
and.b32 %r1142, %r1141, 25166208;
or.b32 %r1143, %r1142, -2139062144;
add.s32 %r1144, %r1143, -16843009;
or.b32 %r1079, %r1144, %r1142;

prmt.b32 %r1079, %r1079, 0, 0xba98;

ld.global.nc.v2.u32 {%r1145, %r1146}, [%rd141];
xor.b32 %r1149, %r1145, %r1077;
or.b32 %r1150, %r1149, -2139062144;
and.b32 %r1151, %r1077, 2139062143;
and.b32 %r1152, %r1145, -2139062144;
sub.s32 %r1153, %r1150, %r1151;
xor.b32 %r1154, %r1152, %r1153;
xor.b32 %r1082, %r1154, -2139062144;
xor.b32 %r1155, %r1146, %r1079;
or.b32 %r1156, %r1155, -2139062144;
and.b32 %r1157, %r1079, 2139062143;
and.b32 %r1158, %r1146, -2139062144;
sub.s32 %r1159, %r1156, %r1157;
xor.b32 %r1160, %r1158, %r1159;
xor.b32 %r1086, %r1160, -2139062144;

dp4a.s32.s32 %r1081, %r1082, %r20, %r914;


dp4a.s32.s32 %r1085, %r1086, %r19, %r1081;

mul.wide.u16 %r1161, %rs108, 64;
and.b32 %r1162, %r1161, 768;
or.b32 %r1163, %r1162, %r1126;
mul.wide.u32 %rd142, %r1163, 8;
add.s64 %rd143, %rd185, %rd142;
shr.u16 %rs111, %rs102, 8;
mul.wide.u16 %r1164, %rs111, 128;
cvt.u32.u16 %r1165, %rs111;
shl.b32 %r1166, %r1165, 21;
or.b32 %r1167, %r1164, %r1166;
and.b32 %r1168, %r1167, 25166208;
or.b32 %r1169, %r1168, -2139062144;
add.s32 %r1170, %r1169, -16843009;
or.b32 %r1089, %r1170, %r1168;

prmt.b32 %r1089, %r1089, 0, 0xba98;

mul.wide.u16 %r1171, %rs111, 8;
shl.b32 %r1172, %r1165, 17;
or.b32 %r1173, %r1171, %r1172;
and.b32 %r1174, %r1173, 25166208;
or.b32 %r1175, %r1174, -2139062144;
add.s32 %r1176, %r1175, -16843009;
or.b32 %r1091, %r1176, %r1174;

prmt.b32 %r1091, %r1091, 0, 0xba98;

ld.global.nc.v2.u32 {%r1177, %r1178}, [%rd143];
xor.b32 %r1181, %r1177, %r1089;
or.b32 %r1182, %r1181, -2139062144;
and.b32 %r1183, %r1089, 2139062143;
and.b32 %r1184, %r1177, -2139062144;
sub.s32 %r1185, %r1182, %r1183;
xor.b32 %r1186, %r1184, %r1185;
xor.b32 %r1094, %r1186, -2139062144;
xor.b32 %r1187, %r1178, %r1091;
or.b32 %r1188, %r1187, -2139062144;
and.b32 %r1189, %r1091, 2139062143;
and.b32 %r1190, %r1178, -2139062144;
sub.s32 %r1191, %r1188, %r1189;
xor.b32 %r1192, %r1190, %r1191;
xor.b32 %r1098, %r1192, -2139062144;

dp4a.s32.s32 %r1093, %r1094, %r22, %r1085;


dp4a.s32.s32 %r1097, %r1098, %r21, %r1093;

and.b16 %rs112, %rs100, 255;
cvt.u32.u16 %r1193, %rs112;
mul.wide.u16 %r1194, %rs108, 16;
and.b32 %r1195, %r1194, 768;
or.b32 %r1196, %r1195, %r1193;
mul.wide.u32 %rd144, %r1196, 8;
add.s64 %rd145, %rd185, %rd144;
and.b16 %rs113, %rs103, 255;
mul.wide.u16 %r1197, %rs113, 128;
cvt.u32.u16 %r1198, %rs103;
shl.b32 %r1199, %r1198, 21;
or.b32 %r1200, %r1197, %r1199;
and.b32 %r1201, %r1200, 25166208;
or.b32 %r1202, %r1201, -2139062144;
add.s32 %r1203, %r1202, -16843009;
or.b32 %r1101, %r1203, %r1201;

prmt.b32 %r1101, %r1101, 0, 0xba98;

mul.wide.u16 %r1204, %rs113, 8;
shl.b32 %r1205, %r1198, 17;
or.b32 %r1206, %r1204, %r1205;
and.b32 %r1207, %r1206, 25166208;
or.b32 %r1208, %r1207, -2139062144;
add.s32 %r1209, %r1208, -16843009;
or.b32 %r1103, %r1209, %r1207;

prmt.b32 %r1103, %r1103, 0, 0xba98;

ld.global.nc.v2.u32 {%r1210, %r1211}, [%rd145];
xor.b32 %r1214, %r1210, %r1101;
or.b32 %r1215, %r1214, -2139062144;
and.b32 %r1216, %r1101, 2139062143;
and.b32 %r1217, %r1210, -2139062144;
sub.s32 %r1218, %r1215, %r1216;
xor.b32 %r1219, %r1217, %r1218;
xor.b32 %r1106, %r1219, -2139062144;
xor.b32 %r1220, %r1211, %r1103;
or.b32 %r1221, %r1220, -2139062144;
and.b32 %r1222, %r1103, 2139062143;
and.b32 %r1223, %r1211, -2139062144;
sub.s32 %r1224, %r1221, %r1222;
xor.b32 %r1225, %r1223, %r1224;
xor.b32 %r1110, %r1225, -2139062144;

dp4a.s32.s32 %r1105, %r1106, %r24, %r914;


dp4a.s32.s32 %r1109, %r1110, %r23, %r1105;

mul.wide.u16 %r1226, %rs108, 4;
and.b32 %r1227, %r1226, 768;
or.b32 %r1228, %r1227, %r1127;
mul.wide.u32 %rd146, %r1228, 8;
add.s64 %rd147, %rd185, %rd146;
shr.u16 %rs114, %rs103, 8;
mul.wide.u16 %r1229, %rs114, 128;
cvt.u32.u16 %r1230, %rs114;
shl.b32 %r1231, %r1230, 21;
or.b32 %r1232, %r1229, %r1231;
and.b32 %r1233, %r1232, 25166208;
or.b32 %r1234, %r1233, -2139062144;
add.s32 %r1235, %r1234, -16843009;
or.b32 %r1113, %r1235, %r1233;

prmt.b32 %r1113, %r1113, 0, 0xba98;

mul.wide.u16 %r1236, %rs114, 8;
shl.b32 %r1237, %r1230, 17;
or.b32 %r1238, %r1236, %r1237;
and.b32 %r1239, %r1238, 25166208;
or.b32 %r1240, %r1239, -2139062144;
add.s32 %r1241, %r1240, -16843009;
or.b32 %r1115, %r1241, %r1239;

prmt.b32 %r1115, %r1115, 0, 0xba98;

ld.global.nc.v2.u32 {%r1242, %r1243}, [%rd147];
xor.b32 %r1246, %r1242, %r1113;
or.b32 %r1247, %r1246, -2139062144;
and.b32 %r1248, %r1113, 2139062143;
and.b32 %r1249, %r1242, -2139062144;
sub.s32 %r1250, %r1247, %r1248;
xor.b32 %r1251, %r1249, %r1250;
xor.b32 %r1118, %r1251, -2139062144;
xor.b32 %r1252, %r1243, %r1115;
or.b32 %r1253, %r1252, -2139062144;
and.b32 %r1254, %r1115, 2139062143;
and.b32 %r1255, %r1243, -2139062144;
sub.s32 %r1256, %r1253, %r1254;
xor.b32 %r1257, %r1255, %r1256;
xor.b32 %r1122, %r1257, -2139062144;

dp4a.s32.s32 %r1117, %r1118, %r26, %r1109;


dp4a.s32.s32 %r1121, %r1122, %r25, %r1117;

cvt.u32.u16 %r1258, %rs105;
cvt.u32.u16 %r1259, %rs106;
mul.lo.s32 %r1260, %r1097, %r1258;
mad.lo.s32 %r1261, %r1121, %r1259, %r1260;
add.s32 %r1262, %r1121, %r1097;
shr.u32 %r1263, %r1262, 31;
add.s32 %r1264, %r1262, %r1263;
shr.s32 %r1265, %r1264, 1;
add.s32 %r1266, %r1261, %r1265;
shr.s32 %r1267, %r1266, 31;
shr.u32 %r1268, %r1267, 30;
add.s32 %r1269, %r1266, %r1268;
shr.s32 %r1270, %r1269, 2;
or.b32 %r1273, %r4, 2;
mad.lo.s32 %r1274, %r50, %r1273, %r16;
mul.wide.s32 %rd148, %r1274, 82;
add.s64 %rd149, %rd193, %rd148;
ld.global.u16 %rs97, [%rd149];

{ cvt.f32.f16 %f121, %rs97;}



{.reg .f16 low,high;
mov.b32 {low,high},%r27;
cvt.f32.f16 %f122, low;}


mul.ftz.f32 %f123, %f121, %f122;
cvt.rn.f32.s32 %f124, %r1270;
fma.rn.ftz.f32 %f32, %f123, %f124, %f32;

$L__BB226_16:
mul.wide.s32 %rd191, %r50, 82;
mov.u64 %rd178, iq2s_grid;
add.s64 %rd151, %rd25, %rd191;
ld.global.nc.u16 %rs116, [%rd151];
shr.u16 %rs117, %rs116, 8;
cvt.u32.u16 %r1324, %rs117;
ld.global.nc.u16 %rs118, [%rd151+2];
shr.u16 %rs119, %rs118, 8;
cvt.u32.u16 %r1325, %rs119;
add.s64 %rd152, %rd26, %rd191;
ld.global.nc.u8 %rs120, [%rd152];
ld.global.nc.u16 %rs121, [%rd151+32];
ld.global.nc.u16 %rs122, [%rd151+34];
ld.global.nc.u8 %rs123, [%rd152+8];
and.b16 %rs124, %rs123, 240;
and.b16 %rs125, %rs123, 15;
shr.u16 %rs126, %rs124, 4;
and.b16 %rs127, %rs116, 255;
cvt.u32.u16 %r1326, %rs127;
shl.b16 %rs128, %rs120, 8;
cvt.u32.u16 %r1327, %rs128;
and.b32 %r1328, %r1327, 768;
or.b32 %r1329, %r1328, %r1326;
mul.wide.u32 %rd153, %r1329, 8;
add.s64 %rd155, %rd178, %rd153;
and.b16 %rs129, %rs121, 255;
mul.wide.u16 %r1330, %rs129, 128;
cvt.u32.u16 %r1331, %rs121;
shl.b32 %r1332, %r1331, 21;
or.b32 %r1333, %r1330, %r1332;
and.b32 %r1334, %r1333, 25166208;
or.b32 %r1335, %r1334, -2139062144;
add.s32 %r1336, %r1335, -16843009;
or.b32 %r1275, %r1336, %r1334;

prmt.b32 %r1275, %r1275, 0, 0xba98;

mul.wide.u16 %r1337, %rs129, 8;
shl.b32 %r1338, %r1331, 17;
or.b32 %r1339, %r1337, %r1338;
and.b32 %r1340, %r1339, 25166208;
or.b32 %r1341, %r1340, -2139062144;
add.s32 %r1342, %r1341, -16843009;
or.b32 %r1277, %r1342, %r1340;

prmt.b32 %r1277, %r1277, 0, 0xba98;

ld.global.nc.v2.u32 {%r1343, %r1344}, [%rd155];
xor.b32 %r1347, %r1343, %r1275;
or.b32 %r1348, %r1347, -2139062144;
and.b32 %r1349, %r1275, 2139062143;
and.b32 %r1350, %r1343, -2139062144;
sub.s32 %r1351, %r1348, %r1349;
xor.b32 %r1352, %r1350, %r1351;
xor.b32 %r1280, %r1352, -2139062144;
xor.b32 %r1353, %r1344, %r1277;
or.b32 %r1354, %r1353, -2139062144;
and.b32 %r1355, %r1277, 2139062143;
and.b32 %r1356, %r1344, -2139062144;
sub.s32 %r1357, %r1354, %r1355;
xor.b32 %r1358, %r1356, %r1357;
xor.b32 %r1284, %r1358, -2139062144;

dp4a.s32.s32 %r1279, %r1280, %r20, %r914;


dp4a.s32.s32 %r1283, %r1284, %r19, %r1279;

mul.wide.u16 %r1359, %rs120, 64;
and.b32 %r1360, %r1359, 768;
or.b32 %r1361, %r1360, %r1324;
mul.wide.u32 %rd156, %r1361, 8;
add.s64 %rd157, %rd178, %rd156;
shr.u16 %rs131, %rs121, 8;
mul.wide.u16 %r1362, %rs131, 128;
cvt.u32.u16 %r1363, %rs131;
shl.b32 %r1364, %r1363, 21;
or.b32 %r1365, %r1362, %r1364;
and.b32 %r1366, %r1365, 25166208;
or.b32 %r1367, %r1366, -2139062144;
add.s32 %r1368, %r1367, -16843009;
or.b32 %r1287, %r1368, %r1366;

prmt.b32 %r1287, %r1287, 0, 0xba98;

mul.wide.u16 %r1369, %rs131, 8;
shl.b32 %r1370, %r1363, 17;
or.b32 %r1371, %r1369, %r1370;
and.b32 %r1372, %r1371, 25166208;
or.b32 %r1373, %r1372, -2139062144;
add.s32 %r1374, %r1373, -16843009;
or.b32 %r1289, %r1374, %r1372;

prmt.b32 %r1289, %r1289, 0, 0xba98;

ld.global.nc.v2.u32 {%r1375, %r1376}, [%rd157];
xor.b32 %r1379, %r1375, %r1287;
or.b32 %r1380, %r1379, -2139062144;
and.b32 %r1381, %r1287, 2139062143;
and.b32 %r1382, %r1375, -2139062144;
sub.s32 %r1383, %r1380, %r1381;
xor.b32 %r1384, %r1382, %r1383;
xor.b32 %r1292, %r1384, -2139062144;
xor.b32 %r1385, %r1376, %r1289;
or.b32 %r1386, %r1385, -2139062144;
and.b32 %r1387, %r1289, 2139062143;
and.b32 %r1388, %r1376, -2139062144;
sub.s32 %r1389, %r1386, %r1387;
xor.b32 %r1390, %r1388, %r1389;
xor.b32 %r1296, %r1390, -2139062144;

dp4a.s32.s32 %r1291, %r1292, %r22, %r1283;


dp4a.s32.s32 %r1295, %r1296, %r21, %r1291;

and.b16 %rs132, %rs118, 255;
cvt.u32.u16 %r1391, %rs132;
mul.wide.u16 %r1392, %rs120, 16;
and.b32 %r1393, %r1392, 768;
or.b32 %r1394, %r1393, %r1391;
mul.wide.u32 %rd158, %r1394, 8;
add.s64 %rd159, %rd178, %rd158;
and.b16 %rs133, %rs122, 255;
mul.wide.u16 %r1395, %rs133, 128;
cvt.u32.u16 %r1396, %rs122;
shl.b32 %r1397, %r1396, 21;
or.b32 %r1398, %r1395, %r1397;
and.b32 %r1399, %r1398, 25166208;
or.b32 %r1400, %r1399, -2139062144;
add.s32 %r1401, %r1400, -16843009;
or.b32 %r1299, %r1401, %r1399;

prmt.b32 %r1299, %r1299, 0, 0xba98;

mul.wide.u16 %r1402, %rs133, 8;
shl.b32 %r1403, %r1396, 17;
or.b32 %r1404, %r1402, %r1403;
and.b32 %r1405, %r1404, 25166208;
or.b32 %r1406, %r1405, -2139062144;
add.s32 %r1407, %r1406, -16843009;
or.b32 %r1301, %r1407, %r1405;

prmt.b32 %r1301, %r1301, 0, 0xba98;

ld.global.nc.v2.u32 {%r1408, %r1409}, [%rd159];
xor.b32 %r1412, %r1408, %r1299;
or.b32 %r1413, %r1412, -2139062144;
and.b32 %r1414, %r1299, 2139062143;
and.b32 %r1415, %r1408, -2139062144;
sub.s32 %r1416, %r1413, %r1414;
xor.b32 %r1417, %r1415, %r1416;
xor.b32 %r1304, %r1417, -2139062144;
xor.b32 %r1418, %r1409, %r1301;
or.b32 %r1419, %r1418, -2139062144;
and.b32 %r1420, %r1301, 2139062143;
and.b32 %r1421, %r1409, -2139062144;
sub.s32 %r1422, %r1419, %r1420;
xor.b32 %r1423, %r1421, %r1422;
xor.b32 %r1308, %r1423, -2139062144;

dp4a.s32.s32 %r1303, %r1304, %r24, %r914;


dp4a.s32.s32 %r1307, %r1308, %r23, %r1303;

mul.wide.u16 %r1424, %rs120, 4;
and.b32 %r1425, %r1424, 768;
or.b32 %r1426, %r1425, %r1325;
mul.wide.u32 %rd160, %r1426, 8;
add.s64 %rd161, %rd178, %rd160;
shr.u16 %rs134, %rs122, 8;
mul.wide.u16 %r1427, %rs134, 128;
cvt.u32.u16 %r1428, %rs134;
shl.b32 %r1429, %r1428, 21;
or.b32 %r1430, %r1427, %r1429;
and.b32 %r1431, %r1430, 25166208;
or.b32 %r1432, %r1431, -2139062144;
add.s32 %r1433, %r1432, -16843009;
or.b32 %r1311, %r1433, %r1431;

prmt.b32 %r1311, %r1311, 0, 0xba98;

mul.wide.u16 %r1434, %rs134, 8;
shl.b32 %r1435, %r1428, 17;
or.b32 %r1436, %r1434, %r1435;
and.b32 %r1437, %r1436, 25166208;
or.b32 %r1438, %r1437, -2139062144;
add.s32 %r1439, %r1438, -16843009;
or.b32 %r1313, %r1439, %r1437;

prmt.b32 %r1313, %r1313, 0, 0xba98;

ld.global.nc.v2.u32 {%r1440, %r1441}, [%rd161];
xor.b32 %r1444, %r1440, %r1311;
or.b32 %r1445, %r1444, -2139062144;
and.b32 %r1446, %r1311, 2139062143;
and.b32 %r1447, %r1440, -2139062144;
sub.s32 %r1448, %r1445, %r1446;
xor.b32 %r1449, %r1447, %r1448;
xor.b32 %r1316, %r1449, -2139062144;
xor.b32 %r1450, %r1441, %r1313;
or.b32 %r1451, %r1450, -2139062144;
and.b32 %r1452, %r1313, 2139062143;
and.b32 %r1453, %r1441, -2139062144;
sub.s32 %r1454, %r1451, %r1452;
xor.b32 %r1455, %r1453, %r1454;
xor.b32 %r1320, %r1455, -2139062144;

dp4a.s32.s32 %r1315, %r1316, %r26, %r1307;


dp4a.s32.s32 %r1319, %r1320, %r25, %r1315;

cvt.u32.u16 %r1456, %rs125;
cvt.u32.u16 %r1457, %rs126;
mul.lo.s32 %r1458, %r1295, %r1456;
mad.lo.s32 %r1459, %r1319, %r1457, %r1458;
add.s32 %r1460, %r1319, %r1295;
shr.u32 %r1461, %r1460, 31;
add.s32 %r1462, %r1460, %r1461;
shr.s32 %r1463, %r1462, 1;
add.s32 %r1464, %r1459, %r1463;
shr.s32 %r1465, %r1464, 31;
shr.u32 %r1466, %r1465, 30;
add.s32 %r1467, %r1464, %r1466;
shr.s32 %r1468, %r1467, 2;
add.s64 %rd162, %rd27, %rd191;
ld.global.nc.u16 %rs115, [%rd162];

{ cvt.f32.f16 %f125, %rs115;}



{.reg .f16 low,high;
mov.b32 {low,high},%r27;
cvt.f32.f16 %f126, low;}


mul.ftz.f32 %f127, %f125, %f126;
cvt.rn.f32.s32 %f128, %r1468;
fma.rn.ftz.f32 %f27, %f127, %f128, %f27;
@%p13 bra $L__BB226_18;

mov.u64 %rd184, iq2s_grid;
add.s64 %rd163, %rd193, %rd11;
ld.global.u16 %rs136, [%rd163+-34];
shr.u16 %rs137, %rs136, 8;
cvt.u32.u16 %r1518, %rs137;
ld.global.u16 %rs138, [%rd163+-32];
shr.u16 %rs139, %rs138, 8;
cvt.u32.u16 %r1519, %rs139;
ld.global.u16 %rs140, [%rd163+-2];
ld.global.u16 %rs141, [%rd163];
add.s64 %rd164, %rd193, %rd12;
ld.global.u8 %rs142, [%rd164];
and.b16 %rs143, %rs142, 15;
shr.u16 %rs144, %rs142, 4;
and.b16 %rs145, %rs136, 255;
cvt.u32.u16 %r1520, %rs145;
ld.global.u8 %rs146, [%rd164+-8];
shl.b16 %rs147, %rs146, 8;
cvt.u32.u16 %r1521, %rs147;
and.b32 %r1522, %r1521, 768;
or.b32 %r1523, %r1522, %r1520;
mul.wide.u32 %rd165, %r1523, 8;
add.s64 %rd167, %rd184, %rd165;
and.b16 %rs148, %rs140, 255;
mul.wide.u16 %r1524, %rs148, 128;
cvt.u32.u16 %r1525, %rs140;
shl.b32 %r1526, %r1525, 21;
or.b32 %r1527, %r1524, %r1526;
and.b32 %r1528, %r1527, 25166208;
or.b32 %r1529, %r1528, -2139062144;
add.s32 %r1530, %r1529, -16843009;
or.b32 %r1469, %r1530, %r1528;

prmt.b32 %r1469, %r1469, 0, 0xba98;

mul.wide.u16 %r1531, %rs148, 8;
shl.b32 %r1532, %r1525, 17;
or.b32 %r1533, %r1531, %r1532;
and.b32 %r1534, %r1533, 25166208;
or.b32 %r1535, %r1534, -2139062144;
add.s32 %r1536, %r1535, -16843009;
or.b32 %r1471, %r1536, %r1534;

prmt.b32 %r1471, %r1471, 0, 0xba98;

ld.global.nc.v2.u32 {%r1537, %r1538}, [%rd167];
mov.u32 %r1500, 0;
xor.b32 %r1541, %r1537, %r1469;
or.b32 %r1542, %r1541, -2139062144;
and.b32 %r1543, %r1469, 2139062143;
and.b32 %r1544, %r1537, -2139062144;
sub.s32 %r1545, %r1542, %r1543;
xor.b32 %r1546, %r1544, %r1545;
xor.b32 %r1474, %r1546, -2139062144;
xor.b32 %r1547, %r1538, %r1471;
or.b32 %r1548, %r1547, -2139062144;
and.b32 %r1549, %r1471, 2139062143;
and.b32 %r1550, %r1538, -2139062144;
sub.s32 %r1551, %r1548, %r1549;
xor.b32 %r1552, %r1550, %r1551;
xor.b32 %r1478, %r1552, -2139062144;

dp4a.s32.s32 %r1473, %r1474, %r20, %r1500;


dp4a.s32.s32 %r1477, %r1478, %r19, %r1473;

mul.wide.u16 %r1553, %rs146, 64;
and.b32 %r1554, %r1553, 768;
or.b32 %r1555, %r1554, %r1518;
mul.wide.u32 %rd168, %r1555, 8;
add.s64 %rd169, %rd184, %rd168;
shr.u16 %rs149, %rs140, 8;
mul.wide.u16 %r1556, %rs149, 128;
cvt.u32.u16 %r1557, %rs149;
shl.b32 %r1558, %r1557, 21;
or.b32 %r1559, %r1556, %r1558;
and.b32 %r1560, %r1559, 25166208;
or.b32 %r1561, %r1560, -2139062144;
add.s32 %r1562, %r1561, -16843009;
or.b32 %r1481, %r1562, %r1560;

prmt.b32 %r1481, %r1481, 0, 0xba98;

mul.wide.u16 %r1563, %rs149, 8;
shl.b32 %r1564, %r1557, 17;
or.b32 %r1565, %r1563, %r1564;
and.b32 %r1566, %r1565, 25166208;
or.b32 %r1567, %r1566, -2139062144;
add.s32 %r1568, %r1567, -16843009;
or.b32 %r1483, %r1568, %r1566;

prmt.b32 %r1483, %r1483, 0, 0xba98;

ld.global.nc.v2.u32 {%r1569, %r1570}, [%rd169];
xor.b32 %r1573, %r1569, %r1481;
or.b32 %r1574, %r1573, -2139062144;
and.b32 %r1575, %r1481, 2139062143;
and.b32 %r1576, %r1569, -2139062144;
sub.s32 %r1577, %r1574, %r1575;
xor.b32 %r1578, %r1576, %r1577;
xor.b32 %r1486, %r1578, -2139062144;
xor.b32 %r1579, %r1570, %r1483;
or.b32 %r1580, %r1579, -2139062144;
and.b32 %r1581, %r1483, 2139062143;
and.b32 %r1582, %r1570, -2139062144;
sub.s32 %r1583, %r1580, %r1581;
xor.b32 %r1584, %r1582, %r1583;
xor.b32 %r1490, %r1584, -2139062144;

dp4a.s32.s32 %r1485, %r1486, %r22, %r1477;


dp4a.s32.s32 %r1489, %r1490, %r21, %r1485;

and.b16 %rs150, %rs138, 255;
cvt.u32.u16 %r1585, %rs150;
mul.wide.u16 %r1586, %rs146, 16;
and.b32 %r1587, %r1586, 768;
or.b32 %r1588, %r1587, %r1585;
mul.wide.u32 %rd170, %r1588, 8;
add.s64 %rd171, %rd184, %rd170;
and.b16 %rs151, %rs141, 255;
mul.wide.u16 %r1589, %rs151, 128;
cvt.u32.u16 %r1590, %rs141;
shl.b32 %r1591, %r1590, 21;
or.b32 %r1592, %r1589, %r1591;
and.b32 %r1593, %r1592, 25166208;
or.b32 %r1594, %r1593, -2139062144;
add.s32 %r1595, %r1594, -16843009;
or.b32 %r1493, %r1595, %r1593;

prmt.b32 %r1493, %r1493, 0, 0xba98;

mul.wide.u16 %r1596, %rs151, 8;
shl.b32 %r1597, %r1590, 17;
or.b32 %r1598, %r1596, %r1597;
and.b32 %r1599, %r1598, 25166208;
or.b32 %r1600, %r1599, -2139062144;
add.s32 %r1601, %r1600, -16843009;
or.b32 %r1495, %r1601, %r1599;

prmt.b32 %r1495, %r1495, 0, 0xba98;

ld.global.nc.v2.u32 {%r1602, %r1603}, [%rd171];
xor.b32 %r1606, %r1602, %r1493;
or.b32 %r1607, %r1606, -2139062144;
and.b32 %r1608, %r1493, 2139062143;
and.b32 %r1609, %r1602, -2139062144;
sub.s32 %r1610, %r1607, %r1608;
xor.b32 %r1611, %r1609, %r1610;
xor.b32 %r1498, %r1611, -2139062144;
xor.b32 %r1612, %r1603, %r1495;
or.b32 %r1613, %r1612, -2139062144;
and.b32 %r1614, %r1495, 2139062143;
and.b32 %r1615, %r1603, -2139062144;
sub.s32 %r1616, %r1613, %r1614;
xor.b32 %r1617, %r1615, %r1616;
xor.b32 %r1502, %r1617, -2139062144;

dp4a.s32.s32 %r1497, %r1498, %r24, %r1500;


dp4a.s32.s32 %r1501, %r1502, %r23, %r1497;

mul.wide.u16 %r1618, %rs146, 4;
and.b32 %r1619, %r1618, 768;
or.b32 %r1620, %r1619, %r1519;
mul.wide.u32 %rd172, %r1620, 8;
add.s64 %rd173, %rd184, %rd172;
shr.u16 %rs152, %rs141, 8;
mul.wide.u16 %r1621, %rs152, 128;
cvt.u32.u16 %r1622, %rs152;
shl.b32 %r1623, %r1622, 21;
or.b32 %r1624, %r1621, %r1623;
and.b32 %r1625, %r1624, 25166208;
or.b32 %r1626, %r1625, -2139062144;
add.s32 %r1627, %r1626, -16843009;
or.b32 %r1505, %r1627, %r1625;

prmt.b32 %r1505, %r1505, 0, 0xba98;

mul.wide.u16 %r1628, %rs152, 8;
shl.b32 %r1629, %r1622, 17;
or.b32 %r1630, %r1628, %r1629;
and.b32 %r1631, %r1630, 25166208;
or.b32 %r1632, %r1631, -2139062144;
add.s32 %r1633, %r1632, -16843009;
or.b32 %r1507, %r1633, %r1631;

prmt.b32 %r1507, %r1507, 0, 0xba98;

ld.global.nc.v2.u32 {%r1634, %r1635}, [%rd173];
xor.b32 %r1638, %r1634, %r1505;
or.b32 %r1639, %r1638, -2139062144;
and.b32 %r1640, %r1505, 2139062143;
and.b32 %r1641, %r1634, -2139062144;
sub.s32 %r1642, %r1639, %r1640;
xor.b32 %r1643, %r1641, %r1642;
xor.b32 %r1510, %r1643, -2139062144;
xor.b32 %r1644, %r1635, %r1507;
or.b32 %r1645, %r1644, -2139062144;
and.b32 %r1646, %r1507, 2139062143;
and.b32 %r1647, %r1635, -2139062144;
sub.s32 %r1648, %r1645, %r1646;
xor.b32 %r1649, %r1647, %r1648;
xor.b32 %r1514, %r1649, -2139062144;

dp4a.s32.s32 %r1509, %r1510, %r26, %r1501;


dp4a.s32.s32 %r1513, %r1514, %r25, %r1509;

cvt.u32.u16 %r1650, %rs143;
cvt.u32.u16 %r1651, %rs144;
mul.lo.s32 %r1652, %r1489, %r1650;
mad.lo.s32 %r1653, %r1513, %r1651, %r1652;
add.s32 %r1654, %r1513, %r1489;
shr.u32 %r1655, %r1654, 31;
add.s32 %r1656, %r1654, %r1655;
shr.s32 %r1657, %r1656, 1;
add.s32 %r1658, %r1653, %r1657;
shr.s32 %r1659, %r1658, 31;
shr.u32 %r1660, %r1659, 30;
add.s32 %r1661, %r1658, %r1660;
shr.s32 %r1662, %r1661, 2;
or.b32 %r1665, %r4, 3;
mad.lo.s32 %r1666, %r50, %r1665, %r16;
mul.wide.s32 %rd174, %r1666, 82;
add.s64 %rd175, %rd193, %rd174;
ld.global.u16 %rs135, [%rd175];

{ cvt.f32.f16 %f129, %rs135;}



{.reg .f16 low,high;
mov.b32 {low,high},%r27;
cvt.f32.f16 %f130, low;}


mul.ftz.f32 %f131, %f129, %f130;
cvt.rn.f32.s32 %f132, %r1662;
fma.rn.ftz.f32 %f31, %f131, %f132, %f31;

$L__BB226_18:
add.s32 %r2413, %r2413, 128;
add.s64 %rd193, %rd193, 1312;
add.s64 %rd192, %rd192, 1312;
add.s32 %r2414, %r2414, 16;
setp.lt.s32 %p17, %r2414, %r84;
@%p17 bra $L__BB226_10;

$L__BB226_19:
mov.u32 %r2400, %tid.y;
setp.eq.s32 %p18, %r2400, 0;
@%p18 bra $L__BB226_23;

mov.u32 %r2401, %tid.y;
setp.ne.s64 %p19, %rd35, 0;
add.s32 %r30, %r2401, -1;
shl.b32 %r1671, %r2401, 9;
mov.u32 %r1672, _ZZ13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r1673, %r1672, %r1671;
shl.b32 %r1674, %r2, 2;
add.s32 %r1675, %r1673, %r1674;
st.shared.f32 [%r1675+-512], %f30;
@%p19 bra $L__BB226_22;
bra.uni $L__BB226_21;

$L__BB226_22:
shl.b32 %r1677, %r30, 9;
mov.u32 %r1678, _ZZ13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r1679, %r1678, %r1677;
add.s32 %r1681, %r1679, %r1674;
st.shared.f32 [%r1681], %f34;
add.s32 %r2397, %r1675, -512;
st.shared.f32 [%r2397+128], %f29;
st.shared.f32 [%r1681+128], %f33;
add.s32 %r2398, %r1675, -512;
st.shared.f32 [%r2398+256], %f28;
st.shared.f32 [%r1681+256], %f32;
add.s32 %r2399, %r1675, -512;
st.shared.f32 [%r2399+384], %f27;
st.shared.f32 [%r1681+384], %f31;
bra.uni $L__BB226_23;

$L__BB226_21:
add.s32 %r2394, %r1675, -512;
st.shared.f32 [%r2394+128], %f29;
add.s32 %r2395, %r1675, -512;
st.shared.f32 [%r2395+256], %f28;
add.s32 %r2396, %r1675, -512;
st.shared.f32 [%r2396+384], %f27;

$L__BB226_23:
mov.u32 %r2402, %tid.y;
setp.ne.s32 %p319, %r2402, 0;
bar.sync 0;
@%p319 bra $L__BB226_78;
bra.uni $L__BB226_24;

$L__BB226_78:
ret;

$L__BB226_24:
ld.param.u64 %rd179, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.f32 %f977, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
ld.param.u32 %r2405, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r2404, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
mov.u32 %r2403, %ctaid.z;
setp.ne.s64 %p21, %rd35, 0;
mov.u32 %r1684, %ctaid.y;
mad.lo.s32 %r1687, %r1684, %r2405, %r4;
mad.lo.s32 %r1688, %r2403, %r2404, %r1687;
neg.ftz.f32 %f35, %f977;
shl.b32 %r1690, %r2, 2;
mov.u32 %r1691, _ZZ13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r32, %r1691, %r1690;
cvta.to.global.u64 %rd176, %rd179;
mul.wide.u32 %rd177, %r1688, 4;
add.s64 %rd30, %rd176, %rd177;
@%p21 bra $L__BB226_33;
bra.uni $L__BB226_25;

$L__BB226_33:
ld.param.u32 %r2406, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
mov.u32 %r1778, _ZZ13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r34, %r1778, %r1690;
setp.gt.s32 %p54, %r2406, 2;
@%p54 bra $L__BB226_44;

ld.param.u32 %r2409, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p57, %r2409, 1;
@%p57 bra $L__BB226_62;

ld.param.u32 %r2410, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p58, %r2410, 2;
@%p58 bra $L__BB226_36;
bra.uni $L__BB226_70;

$L__BB226_36:
ld.shared.f32 %f533, [%r32];
add.ftz.f32 %f534, %f533, %f30;
ld.shared.f32 %f535, [%r34];
add.ftz.f32 %f536, %f535, %f34;
ld.shared.f32 %f537, [%r32+512];
add.ftz.f32 %f538, %f537, %f534;
ld.shared.f32 %f539, [%r34+512];
add.ftz.f32 %f540, %f539, %f536;
ld.shared.f32 %f541, [%r32+1024];
add.ftz.f32 %f542, %f541, %f538;
ld.shared.f32 %f543, [%r34+1024];
add.ftz.f32 %f544, %f543, %f540;
mov.b32 %r2025, %f542;
mov.u32 %r2026, 31;
mov.u32 %r2027, 16;
mov.u32 %r2028, -1;
shfl.sync.bfly.b32 %r2029|%p163, %r2025, %r2027, %r2026, %r2028;
mov.b32 %f545, %r2029;
add.ftz.f32 %f546, %f542, %f545;
mov.b32 %r2030, %f546;
mov.u32 %r2031, 8;
shfl.sync.bfly.b32 %r2032|%p164, %r2030, %r2031, %r2026, %r2028;
mov.b32 %f547, %r2032;
add.ftz.f32 %f548, %f546, %f547;
mov.b32 %r2033, %f548;
mov.u32 %r2034, 4;
shfl.sync.bfly.b32 %r2035|%p165, %r2033, %r2034, %r2026, %r2028;
mov.b32 %f549, %r2035;
add.ftz.f32 %f550, %f548, %f549;
mov.b32 %r2036, %f550;
mov.u32 %r2037, 2;
shfl.sync.bfly.b32 %r2038|%p166, %r2036, %r2037, %r2026, %r2028;
mov.b32 %f551, %r2038;
add.ftz.f32 %f552, %f550, %f551;
mov.b32 %r2039, %f552;
mov.u32 %r2040, 1;
shfl.sync.bfly.b32 %r2041|%p167, %r2039, %r2040, %r2026, %r2028;
mov.b32 %f553, %r2041;
add.ftz.f32 %f56, %f552, %f553;
mov.b32 %r2042, %f544;
shfl.sync.bfly.b32 %r2043|%p168, %r2042, %r2027, %r2026, %r2028;
mov.b32 %f554, %r2043;
add.ftz.f32 %f555, %f544, %f554;
mov.b32 %r2044, %f555;
shfl.sync.bfly.b32 %r2045|%p169, %r2044, %r2031, %r2026, %r2028;
mov.b32 %f556, %r2045;
add.ftz.f32 %f557, %f555, %f556;
mov.b32 %r2046, %f557;
shfl.sync.bfly.b32 %r2047|%p170, %r2046, %r2034, %r2026, %r2028;
mov.b32 %f558, %r2047;
add.ftz.f32 %f559, %f557, %f558;
mov.b32 %r2048, %f559;
shfl.sync.bfly.b32 %r2049|%p171, %r2048, %r2037, %r2026, %r2028;
mov.b32 %f560, %r2049;
add.ftz.f32 %f561, %f559, %f560;
mov.b32 %r2050, %f561;
shfl.sync.bfly.b32 %r2051|%p172, %r2050, %r2040, %r2026, %r2028;
mov.b32 %f562, %r2051;
add.ftz.f32 %f57, %f561, %f562;
setp.ne.s32 %p173, %r2, 0;
setp.ge.u32 %p174, %r4, %r51;
or.pred %p175, %p173, %p174;
@%p175 bra $L__BB226_38;

add.ftz.f32 %f563, %f5, %f56;
add.ftz.f32 %f564, %f4, %f57;
mul.ftz.f32 %f565, %f564, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f566, %f565;
add.ftz.f32 %f567, %f566, 0f3F800000;
div.approx.ftz.f32 %f568, %f564, %f567;
mul.ftz.f32 %f569, %f563, %f568;
st.global.f32 [%rd30], %f569;

$L__BB226_38:
ld.shared.f32 %f570, [%r32+128];
add.ftz.f32 %f571, %f570, %f29;
ld.shared.f32 %f572, [%r34+128];
add.ftz.f32 %f573, %f572, %f33;
ld.shared.f32 %f574, [%r32+640];
add.ftz.f32 %f575, %f574, %f571;
ld.shared.f32 %f576, [%r34+640];
add.ftz.f32 %f577, %f576, %f573;
ld.shared.f32 %f578, [%r32+1152];
add.ftz.f32 %f579, %f578, %f575;
ld.shared.f32 %f580, [%r34+1152];
add.ftz.f32 %f581, %f580, %f577;
mov.b32 %r2055, %f579;
shfl.sync.bfly.b32 %r2059|%p176, %r2055, %r2027, %r2026, %r2028;
mov.b32 %f582, %r2059;
add.ftz.f32 %f583, %f579, %f582;
mov.b32 %r2060, %f583;
shfl.sync.bfly.b32 %r2062|%p177, %r2060, %r2031, %r2026, %r2028;
mov.b32 %f584, %r2062;
add.ftz.f32 %f585, %f583, %f584;
mov.b32 %r2063, %f585;
shfl.sync.bfly.b32 %r2065|%p178, %r2063, %r2034, %r2026, %r2028;
mov.b32 %f586, %r2065;
add.ftz.f32 %f587, %f585, %f586;
mov.b32 %r2066, %f587;
shfl.sync.bfly.b32 %r2068|%p179, %r2066, %r2037, %r2026, %r2028;
mov.b32 %f588, %r2068;
add.ftz.f32 %f589, %f587, %f588;
mov.b32 %r2069, %f589;
shfl.sync.bfly.b32 %r2071|%p180, %r2069, %r2040, %r2026, %r2028;
mov.b32 %f590, %r2071;
add.ftz.f32 %f58, %f589, %f590;
mov.b32 %r2072, %f581;
shfl.sync.bfly.b32 %r2073|%p181, %r2072, %r2027, %r2026, %r2028;
mov.b32 %f591, %r2073;
add.ftz.f32 %f592, %f581, %f591;
mov.b32 %r2074, %f592;
shfl.sync.bfly.b32 %r2075|%p182, %r2074, %r2031, %r2026, %r2028;
mov.b32 %f593, %r2075;
add.ftz.f32 %f594, %f592, %f593;
mov.b32 %r2076, %f594;
shfl.sync.bfly.b32 %r2077|%p183, %r2076, %r2034, %r2026, %r2028;
mov.b32 %f595, %r2077;
add.ftz.f32 %f596, %f594, %f595;
mov.b32 %r2078, %f596;
shfl.sync.bfly.b32 %r2079|%p184, %r2078, %r2037, %r2026, %r2028;
mov.b32 %f597, %r2079;
add.ftz.f32 %f598, %f596, %f597;
mov.b32 %r2080, %f598;
shfl.sync.bfly.b32 %r2081|%p185, %r2080, %r2040, %r2026, %r2028;
mov.b32 %f599, %r2081;
add.ftz.f32 %f59, %f598, %f599;
setp.ne.s32 %p186, %r2, 1;
or.b32 %r2085, %r4, 1;
setp.ge.u32 %p187, %r2085, %r51;
or.pred %p188, %p186, %p187;
@%p188 bra $L__BB226_40;

add.ftz.f32 %f600, %f5, %f58;
add.ftz.f32 %f601, %f4, %f59;
mul.ftz.f32 %f602, %f601, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f603, %f602;
add.ftz.f32 %f604, %f603, 0f3F800000;
div.approx.ftz.f32 %f605, %f601, %f604;
mul.ftz.f32 %f606, %f600, %f605;
st.global.f32 [%rd30+4], %f606;

$L__BB226_40:
ld.shared.f32 %f607, [%r32+256];
add.ftz.f32 %f608, %f607, %f28;
ld.shared.f32 %f609, [%r34+256];
add.ftz.f32 %f610, %f609, %f32;
ld.shared.f32 %f611, [%r32+768];
add.ftz.f32 %f612, %f611, %f608;
ld.shared.f32 %f613, [%r34+768];
add.ftz.f32 %f614, %f613, %f610;
ld.shared.f32 %f615, [%r32+1280];
add.ftz.f32 %f616, %f615, %f612;
ld.shared.f32 %f617, [%r34+1280];
add.ftz.f32 %f618, %f617, %f614;
mov.b32 %r2086, %f616;
mov.u32 %r2087, 31;
mov.u32 %r2088, 16;
mov.u32 %r2089, -1;
shfl.sync.bfly.b32 %r2090|%p189, %r2086, %r2088, %r2087, %r2089;
mov.b32 %f619, %r2090;
add.ftz.f32 %f620, %f616, %f619;
mov.b32 %r2091, %f620;
mov.u32 %r2092, 8;
shfl.sync.bfly.b32 %r2093|%p190, %r2091, %r2092, %r2087, %r2089;
mov.b32 %f621, %r2093;
add.ftz.f32 %f622, %f620, %f621;
mov.b32 %r2094, %f622;
mov.u32 %r2095, 4;
shfl.sync.bfly.b32 %r2096|%p191, %r2094, %r2095, %r2087, %r2089;
mov.b32 %f623, %r2096;
add.ftz.f32 %f624, %f622, %f623;
mov.b32 %r2097, %f624;
mov.u32 %r2098, 2;
shfl.sync.bfly.b32 %r2099|%p192, %r2097, %r2098, %r2087, %r2089;
mov.b32 %f625, %r2099;
add.ftz.f32 %f626, %f624, %f625;
mov.b32 %r2100, %f626;
mov.u32 %r2101, 1;
shfl.sync.bfly.b32 %r2102|%p193, %r2100, %r2101, %r2087, %r2089;
mov.b32 %f627, %r2102;
add.ftz.f32 %f60, %f626, %f627;
mov.b32 %r2103, %f618;
shfl.sync.bfly.b32 %r2104|%p194, %r2103, %r2088, %r2087, %r2089;
mov.b32 %f628, %r2104;
add.ftz.f32 %f629, %f618, %f628;
mov.b32 %r2105, %f629;
shfl.sync.bfly.b32 %r2106|%p195, %r2105, %r2092, %r2087, %r2089;
mov.b32 %f630, %r2106;
add.ftz.f32 %f631, %f629, %f630;
mov.b32 %r2107, %f631;
shfl.sync.bfly.b32 %r2108|%p196, %r2107, %r2095, %r2087, %r2089;
mov.b32 %f632, %r2108;
add.ftz.f32 %f633, %f631, %f632;
mov.b32 %r2109, %f633;
shfl.sync.bfly.b32 %r2110|%p197, %r2109, %r2098, %r2087, %r2089;
mov.b32 %f634, %r2110;
add.ftz.f32 %f635, %f633, %f634;
mov.b32 %r2111, %f635;
shfl.sync.bfly.b32 %r2112|%p198, %r2111, %r2101, %r2087, %r2089;
mov.b32 %f636, %r2112;
add.ftz.f32 %f61, %f635, %f636;
setp.ne.s32 %p199, %r2, 2;
or.b32 %r2116, %r4, 2;
setp.ge.u32 %p200, %r2116, %r51;
or.pred %p201, %p199, %p200;
@%p201 bra $L__BB226_42;

add.ftz.f32 %f637, %f5, %f60;
add.ftz.f32 %f638, %f4, %f61;
mul.ftz.f32 %f639, %f638, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f640, %f639;
add.ftz.f32 %f641, %f640, 0f3F800000;
div.approx.ftz.f32 %f642, %f638, %f641;
mul.ftz.f32 %f643, %f637, %f642;
st.global.f32 [%rd30+8], %f643;

$L__BB226_42:
ld.shared.f32 %f644, [%r32+384];
add.ftz.f32 %f645, %f644, %f27;
ld.shared.f32 %f646, [%r34+384];
add.ftz.f32 %f647, %f646, %f31;
ld.shared.f32 %f648, [%r32+896];
add.ftz.f32 %f649, %f648, %f645;
ld.shared.f32 %f650, [%r34+896];
add.ftz.f32 %f651, %f650, %f647;
ld.shared.f32 %f652, [%r32+1408];
add.ftz.f32 %f653, %f652, %f649;
ld.shared.f32 %f654, [%r34+1408];
add.ftz.f32 %f655, %f654, %f651;
mov.b32 %r2117, %f653;
shfl.sync.bfly.b32 %r2121|%p202, %r2117, %r2088, %r2087, %r2089;
mov.b32 %f656, %r2121;
add.ftz.f32 %f657, %f653, %f656;
mov.b32 %r2122, %f657;
shfl.sync.bfly.b32 %r2124|%p203, %r2122, %r2092, %r2087, %r2089;
mov.b32 %f658, %r2124;
add.ftz.f32 %f659, %f657, %f658;
mov.b32 %r2125, %f659;
shfl.sync.bfly.b32 %r2127|%p204, %r2125, %r2095, %r2087, %r2089;
mov.b32 %f660, %r2127;
add.ftz.f32 %f661, %f659, %f660;
mov.b32 %r2128, %f661;
shfl.sync.bfly.b32 %r2130|%p205, %r2128, %r2098, %r2087, %r2089;
mov.b32 %f662, %r2130;
add.ftz.f32 %f663, %f661, %f662;
mov.b32 %r2131, %f663;
shfl.sync.bfly.b32 %r2133|%p206, %r2131, %r2101, %r2087, %r2089;
mov.b32 %f664, %r2133;
add.ftz.f32 %f62, %f663, %f664;
mov.b32 %r2134, %f655;
shfl.sync.bfly.b32 %r2135|%p207, %r2134, %r2088, %r2087, %r2089;
mov.b32 %f665, %r2135;
add.ftz.f32 %f666, %f655, %f665;
mov.b32 %r2136, %f666;
shfl.sync.bfly.b32 %r2137|%p208, %r2136, %r2092, %r2087, %r2089;
mov.b32 %f667, %r2137;
add.ftz.f32 %f668, %f666, %f667;
mov.b32 %r2138, %f668;
shfl.sync.bfly.b32 %r2139|%p209, %r2138, %r2095, %r2087, %r2089;
mov.b32 %f669, %r2139;
add.ftz.f32 %f670, %f668, %f669;
mov.b32 %r2140, %f670;
shfl.sync.bfly.b32 %r2141|%p210, %r2140, %r2098, %r2087, %r2089;
mov.b32 %f671, %r2141;
add.ftz.f32 %f672, %f670, %f671;
mov.b32 %r2142, %f672;
shfl.sync.bfly.b32 %r2143|%p211, %r2142, %r2101, %r2087, %r2089;
mov.b32 %f673, %r2143;
add.ftz.f32 %f63, %f672, %f673;
setp.ne.s32 %p212, %r2, 3;
or.b32 %r2147, %r4, 3;
setp.ge.u32 %p213, %r2147, %r51;
or.pred %p214, %p212, %p213;
@%p214 bra $L__BB226_78;

add.ftz.f32 %f674, %f5, %f62;
add.ftz.f32 %f675, %f4, %f63;
mul.ftz.f32 %f676, %f675, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f677, %f676;
add.ftz.f32 %f678, %f677, 0f3F800000;
div.approx.ftz.f32 %f679, %f675, %f678;
mul.ftz.f32 %f680, %f674, %f679;
st.global.f32 [%rd30+12], %f680;
bra.uni $L__BB226_78;

$L__BB226_25:
ld.shared.f32 %f133, [%r32];
add.ftz.f32 %f134, %f133, %f30;
ld.shared.f32 %f135, [%r32+512];
add.ftz.f32 %f136, %f135, %f134;
ld.shared.f32 %f137, [%r32+1024];
add.ftz.f32 %f138, %f137, %f136;
mov.b32 %r1692, %f138;
mov.u32 %r1693, 31;
mov.u32 %r1694, 16;
mov.u32 %r1695, -1;
shfl.sync.bfly.b32 %r1696|%p22, %r1692, %r1694, %r1693, %r1695;
mov.b32 %f139, %r1696;
add.ftz.f32 %f140, %f138, %f139;
mov.b32 %r1697, %f140;
mov.u32 %r1698, 8;
shfl.sync.bfly.b32 %r1699|%p23, %r1697, %r1698, %r1693, %r1695;
mov.b32 %f141, %r1699;
add.ftz.f32 %f142, %f140, %f141;
mov.b32 %r1700, %f142;
mov.u32 %r1701, 4;
shfl.sync.bfly.b32 %r1702|%p24, %r1700, %r1701, %r1693, %r1695;
mov.b32 %f143, %r1702;
add.ftz.f32 %f144, %f142, %f143;
mov.b32 %r1703, %f144;
mov.u32 %r1704, 2;
shfl.sync.bfly.b32 %r1705|%p25, %r1703, %r1704, %r1693, %r1695;
mov.b32 %f145, %r1705;
add.ftz.f32 %f146, %f144, %f145;
mov.b32 %r1706, %f146;
mov.u32 %r1707, 1;
shfl.sync.bfly.b32 %r1708|%p26, %r1706, %r1707, %r1693, %r1695;
mov.b32 %f147, %r1708;
add.ftz.f32 %f36, %f146, %f147;
setp.ne.s32 %p27, %r2, 0;
setp.ge.u32 %p28, %r4, %r51;
or.pred %p29, %p27, %p28;
@%p29 bra $L__BB226_27;

add.ftz.f32 %f148, %f5, %f36;
st.global.f32 [%rd30], %f148;

$L__BB226_27:
ld.shared.f32 %f149, [%r32+128];
add.ftz.f32 %f150, %f149, %f29;
ld.shared.f32 %f151, [%r32+640];
add.ftz.f32 %f152, %f151, %f150;
ld.shared.f32 %f153, [%r32+1152];
add.ftz.f32 %f154, %f153, %f152;
mov.b32 %r1712, %f154;
shfl.sync.bfly.b32 %r1716|%p30, %r1712, %r1694, %r1693, %r1695;
mov.b32 %f155, %r1716;
add.ftz.f32 %f156, %f154, %f155;
mov.b32 %r1717, %f156;
shfl.sync.bfly.b32 %r1719|%p31, %r1717, %r1698, %r1693, %r1695;
mov.b32 %f157, %r1719;
add.ftz.f32 %f158, %f156, %f157;
mov.b32 %r1720, %f158;
shfl.sync.bfly.b32 %r1722|%p32, %r1720, %r1701, %r1693, %r1695;
mov.b32 %f159, %r1722;
add.ftz.f32 %f160, %f158, %f159;
mov.b32 %r1723, %f160;
shfl.sync.bfly.b32 %r1725|%p33, %r1723, %r1704, %r1693, %r1695;
mov.b32 %f161, %r1725;
add.ftz.f32 %f162, %f160, %f161;
mov.b32 %r1726, %f162;
shfl.sync.bfly.b32 %r1728|%p34, %r1726, %r1707, %r1693, %r1695;
mov.b32 %f163, %r1728;
add.ftz.f32 %f37, %f162, %f163;
setp.ne.s32 %p35, %r2, 1;
or.b32 %r1732, %r4, 1;
setp.ge.u32 %p36, %r1732, %r51;
or.pred %p37, %p35, %p36;
@%p37 bra $L__BB226_29;

add.ftz.f32 %f164, %f5, %f37;
st.global.f32 [%rd30+4], %f164;

$L__BB226_29:
ld.shared.f32 %f165, [%r32+256];
add.ftz.f32 %f166, %f165, %f28;
ld.shared.f32 %f167, [%r32+768];
add.ftz.f32 %f168, %f167, %f166;
ld.shared.f32 %f169, [%r32+1280];
add.ftz.f32 %f170, %f169, %f168;
mov.b32 %r1733, %f170;
mov.u32 %r1734, 31;
mov.u32 %r1735, 16;
mov.u32 %r1736, -1;
shfl.sync.bfly.b32 %r1737|%p38, %r1733, %r1735, %r1734, %r1736;
mov.b32 %f171, %r1737;
add.ftz.f32 %f172, %f170, %f171;
mov.b32 %r1738, %f172;
mov.u32 %r1739, 8;
shfl.sync.bfly.b32 %r1740|%p39, %r1738, %r1739, %r1734, %r1736;
mov.b32 %f173, %r1740;
add.ftz.f32 %f174, %f172, %f173;
mov.b32 %r1741, %f174;
mov.u32 %r1742, 4;
shfl.sync.bfly.b32 %r1743|%p40, %r1741, %r1742, %r1734, %r1736;
mov.b32 %f175, %r1743;
add.ftz.f32 %f176, %f174, %f175;
mov.b32 %r1744, %f176;
mov.u32 %r1745, 2;
shfl.sync.bfly.b32 %r1746|%p41, %r1744, %r1745, %r1734, %r1736;
mov.b32 %f177, %r1746;
add.ftz.f32 %f178, %f176, %f177;
mov.b32 %r1747, %f178;
mov.u32 %r1748, 1;
shfl.sync.bfly.b32 %r1749|%p42, %r1747, %r1748, %r1734, %r1736;
mov.b32 %f179, %r1749;
add.ftz.f32 %f38, %f178, %f179;
setp.ne.s32 %p43, %r2, 2;
or.b32 %r1753, %r4, 2;
setp.ge.u32 %p44, %r1753, %r51;
or.pred %p45, %p43, %p44;
@%p45 bra $L__BB226_31;

add.ftz.f32 %f180, %f5, %f38;
st.global.f32 [%rd30+8], %f180;

$L__BB226_31:
ld.shared.f32 %f181, [%r32+384];
add.ftz.f32 %f182, %f181, %f27;
ld.shared.f32 %f183, [%r32+896];
add.ftz.f32 %f184, %f183, %f182;
ld.shared.f32 %f185, [%r32+1408];
add.ftz.f32 %f186, %f185, %f184;
mov.b32 %r1754, %f186;
shfl.sync.bfly.b32 %r1758|%p46, %r1754, %r1735, %r1734, %r1736;
mov.b32 %f187, %r1758;
add.ftz.f32 %f188, %f186, %f187;
mov.b32 %r1759, %f188;
shfl.sync.bfly.b32 %r1761|%p47, %r1759, %r1739, %r1734, %r1736;
mov.b32 %f189, %r1761;
add.ftz.f32 %f190, %f188, %f189;
mov.b32 %r1762, %f190;
shfl.sync.bfly.b32 %r1764|%p48, %r1762, %r1742, %r1734, %r1736;
mov.b32 %f191, %r1764;
add.ftz.f32 %f192, %f190, %f191;
mov.b32 %r1765, %f192;
shfl.sync.bfly.b32 %r1767|%p49, %r1765, %r1745, %r1734, %r1736;
mov.b32 %f193, %r1767;
add.ftz.f32 %f194, %f192, %f193;
mov.b32 %r1768, %f194;
shfl.sync.bfly.b32 %r1770|%p50, %r1768, %r1748, %r1734, %r1736;
mov.b32 %f195, %r1770;
add.ftz.f32 %f39, %f194, %f195;
setp.ne.s32 %p51, %r2, 3;
or.b32 %r1774, %r4, 3;
setp.ge.u32 %p52, %r1774, %r51;
or.pred %p53, %p51, %p52;
@%p53 bra $L__BB226_78;

add.ftz.f32 %f196, %f5, %f39;
st.global.f32 [%rd30+12], %f196;
bra.uni $L__BB226_78;

$L__BB226_44:
ld.param.u32 %r2407, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p55, %r2407, 3;
@%p55 bra $L__BB226_54;

ld.param.u32 %r2408, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.ne.s32 %p56, %r2408, 6;
@%p56 bra $L__BB226_70;

ld.shared.f32 %f197, [%r32];
add.ftz.f32 %f198, %f197, %f30;
ld.shared.f32 %f199, [%r34];
add.ftz.f32 %f200, %f199, %f34;
ld.shared.f32 %f201, [%r32+512];
add.ftz.f32 %f202, %f201, %f198;
ld.shared.f32 %f203, [%r34+512];
add.ftz.f32 %f204, %f203, %f200;
ld.shared.f32 %f205, [%r32+1024];
add.ftz.f32 %f206, %f205, %f202;
ld.shared.f32 %f207, [%r34+1024];
add.ftz.f32 %f208, %f207, %f204;
mov.b32 %r1779, %f206;
mov.u32 %r1780, 31;
mov.u32 %r1781, 16;
mov.u32 %r1782, -1;
shfl.sync.bfly.b32 %r1783|%p59, %r1779, %r1781, %r1780, %r1782;
mov.b32 %f209, %r1783;
add.ftz.f32 %f210, %f206, %f209;
mov.b32 %r1784, %f210;
mov.u32 %r1785, 8;
shfl.sync.bfly.b32 %r1786|%p60, %r1784, %r1785, %r1780, %r1782;
mov.b32 %f211, %r1786;
add.ftz.f32 %f212, %f210, %f211;
mov.b32 %r1787, %f212;
mov.u32 %r1788, 4;
shfl.sync.bfly.b32 %r1789|%p61, %r1787, %r1788, %r1780, %r1782;
mov.b32 %f213, %r1789;
add.ftz.f32 %f214, %f212, %f213;
mov.b32 %r1790, %f214;
mov.u32 %r1791, 2;
shfl.sync.bfly.b32 %r1792|%p62, %r1790, %r1791, %r1780, %r1782;
mov.b32 %f215, %r1792;
add.ftz.f32 %f216, %f214, %f215;
mov.b32 %r1793, %f216;
mov.u32 %r1794, 1;
shfl.sync.bfly.b32 %r1795|%p63, %r1793, %r1794, %r1780, %r1782;
mov.b32 %f217, %r1795;
add.ftz.f32 %f40, %f216, %f217;
mov.b32 %r1796, %f208;
shfl.sync.bfly.b32 %r1797|%p64, %r1796, %r1781, %r1780, %r1782;
mov.b32 %f218, %r1797;
add.ftz.f32 %f219, %f208, %f218;
mov.b32 %r1798, %f219;
shfl.sync.bfly.b32 %r1799|%p65, %r1798, %r1785, %r1780, %r1782;
mov.b32 %f220, %r1799;
add.ftz.f32 %f221, %f219, %f220;
mov.b32 %r1800, %f221;
shfl.sync.bfly.b32 %r1801|%p66, %r1800, %r1788, %r1780, %r1782;
mov.b32 %f222, %r1801;
add.ftz.f32 %f223, %f221, %f222;
mov.b32 %r1802, %f223;
shfl.sync.bfly.b32 %r1803|%p67, %r1802, %r1791, %r1780, %r1782;
mov.b32 %f224, %r1803;
add.ftz.f32 %f225, %f223, %f224;
mov.b32 %r1804, %f225;
shfl.sync.bfly.b32 %r1805|%p68, %r1804, %r1794, %r1780, %r1782;
mov.b32 %f226, %r1805;
add.ftz.f32 %f41, %f225, %f226;
setp.ne.s32 %p69, %r2, 0;
setp.ge.u32 %p70, %r4, %r51;
or.pred %p71, %p69, %p70;
@%p71 bra $L__BB226_48;

ld.param.f32 %f981, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f227, %f5, %f40;
add.ftz.f32 %f228, %f4, %f41;
min.ftz.f32 %f229, %f228, %f981;
min.ftz.f32 %f230, %f227, %f981;
max.ftz.f32 %f231, %f230, %f35;
mul.ftz.f32 %f232, %f229, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f233, %f232;
add.ftz.f32 %f234, %f233, 0f3F800000;
div.approx.ftz.f32 %f235, %f229, %f234;
mul.ftz.f32 %f236, %f231, %f235;
st.global.f32 [%rd30], %f236;

$L__BB226_48:
ld.shared.f32 %f237, [%r32+128];
add.ftz.f32 %f238, %f237, %f29;
ld.shared.f32 %f239, [%r34+128];
add.ftz.f32 %f240, %f239, %f33;
ld.shared.f32 %f241, [%r32+640];
add.ftz.f32 %f242, %f241, %f238;
ld.shared.f32 %f243, [%r34+640];
add.ftz.f32 %f244, %f243, %f240;
ld.shared.f32 %f245, [%r32+1152];
add.ftz.f32 %f246, %f245, %f242;
ld.shared.f32 %f247, [%r34+1152];
add.ftz.f32 %f248, %f247, %f244;
mov.b32 %r1809, %f246;
shfl.sync.bfly.b32 %r1813|%p72, %r1809, %r1781, %r1780, %r1782;
mov.b32 %f249, %r1813;
add.ftz.f32 %f250, %f246, %f249;
mov.b32 %r1814, %f250;
shfl.sync.bfly.b32 %r1816|%p73, %r1814, %r1785, %r1780, %r1782;
mov.b32 %f251, %r1816;
add.ftz.f32 %f252, %f250, %f251;
mov.b32 %r1817, %f252;
shfl.sync.bfly.b32 %r1819|%p74, %r1817, %r1788, %r1780, %r1782;
mov.b32 %f253, %r1819;
add.ftz.f32 %f254, %f252, %f253;
mov.b32 %r1820, %f254;
shfl.sync.bfly.b32 %r1822|%p75, %r1820, %r1791, %r1780, %r1782;
mov.b32 %f255, %r1822;
add.ftz.f32 %f256, %f254, %f255;
mov.b32 %r1823, %f256;
shfl.sync.bfly.b32 %r1825|%p76, %r1823, %r1794, %r1780, %r1782;
mov.b32 %f257, %r1825;
add.ftz.f32 %f42, %f256, %f257;
mov.b32 %r1826, %f248;
shfl.sync.bfly.b32 %r1827|%p77, %r1826, %r1781, %r1780, %r1782;
mov.b32 %f258, %r1827;
add.ftz.f32 %f259, %f248, %f258;
mov.b32 %r1828, %f259;
shfl.sync.bfly.b32 %r1829|%p78, %r1828, %r1785, %r1780, %r1782;
mov.b32 %f260, %r1829;
add.ftz.f32 %f261, %f259, %f260;
mov.b32 %r1830, %f261;
shfl.sync.bfly.b32 %r1831|%p79, %r1830, %r1788, %r1780, %r1782;
mov.b32 %f262, %r1831;
add.ftz.f32 %f263, %f261, %f262;
mov.b32 %r1832, %f263;
shfl.sync.bfly.b32 %r1833|%p80, %r1832, %r1791, %r1780, %r1782;
mov.b32 %f264, %r1833;
add.ftz.f32 %f265, %f263, %f264;
mov.b32 %r1834, %f265;
shfl.sync.bfly.b32 %r1835|%p81, %r1834, %r1794, %r1780, %r1782;
mov.b32 %f266, %r1835;
add.ftz.f32 %f43, %f265, %f266;
setp.ne.s32 %p82, %r2, 1;
or.b32 %r1839, %r4, 1;
setp.ge.u32 %p83, %r1839, %r51;
or.pred %p84, %p82, %p83;
@%p84 bra $L__BB226_50;

ld.param.f32 %f980, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f267, %f5, %f42;
add.ftz.f32 %f268, %f4, %f43;
min.ftz.f32 %f269, %f268, %f980;
min.ftz.f32 %f270, %f267, %f980;
max.ftz.f32 %f271, %f270, %f35;
mul.ftz.f32 %f272, %f269, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f273, %f272;
add.ftz.f32 %f274, %f273, 0f3F800000;
div.approx.ftz.f32 %f275, %f269, %f274;
mul.ftz.f32 %f276, %f271, %f275;
st.global.f32 [%rd30+4], %f276;

$L__BB226_50:
ld.shared.f32 %f277, [%r32+256];
add.ftz.f32 %f278, %f277, %f28;
ld.shared.f32 %f279, [%r34+256];
add.ftz.f32 %f280, %f279, %f32;
ld.shared.f32 %f281, [%r32+768];
add.ftz.f32 %f282, %f281, %f278;
ld.shared.f32 %f283, [%r34+768];
add.ftz.f32 %f284, %f283, %f280;
ld.shared.f32 %f285, [%r32+1280];
add.ftz.f32 %f286, %f285, %f282;
ld.shared.f32 %f287, [%r34+1280];
add.ftz.f32 %f288, %f287, %f284;
mov.b32 %r1840, %f286;
mov.u32 %r1841, 31;
mov.u32 %r1842, 16;
mov.u32 %r1843, -1;
shfl.sync.bfly.b32 %r1844|%p85, %r1840, %r1842, %r1841, %r1843;
mov.b32 %f289, %r1844;
add.ftz.f32 %f290, %f286, %f289;
mov.b32 %r1845, %f290;
mov.u32 %r1846, 8;
shfl.sync.bfly.b32 %r1847|%p86, %r1845, %r1846, %r1841, %r1843;
mov.b32 %f291, %r1847;
add.ftz.f32 %f292, %f290, %f291;
mov.b32 %r1848, %f292;
mov.u32 %r1849, 4;
shfl.sync.bfly.b32 %r1850|%p87, %r1848, %r1849, %r1841, %r1843;
mov.b32 %f293, %r1850;
add.ftz.f32 %f294, %f292, %f293;
mov.b32 %r1851, %f294;
mov.u32 %r1852, 2;
shfl.sync.bfly.b32 %r1853|%p88, %r1851, %r1852, %r1841, %r1843;
mov.b32 %f295, %r1853;
add.ftz.f32 %f296, %f294, %f295;
mov.b32 %r1854, %f296;
mov.u32 %r1855, 1;
shfl.sync.bfly.b32 %r1856|%p89, %r1854, %r1855, %r1841, %r1843;
mov.b32 %f297, %r1856;
add.ftz.f32 %f44, %f296, %f297;
mov.b32 %r1857, %f288;
shfl.sync.bfly.b32 %r1858|%p90, %r1857, %r1842, %r1841, %r1843;
mov.b32 %f298, %r1858;
add.ftz.f32 %f299, %f288, %f298;
mov.b32 %r1859, %f299;
shfl.sync.bfly.b32 %r1860|%p91, %r1859, %r1846, %r1841, %r1843;
mov.b32 %f300, %r1860;
add.ftz.f32 %f301, %f299, %f300;
mov.b32 %r1861, %f301;
shfl.sync.bfly.b32 %r1862|%p92, %r1861, %r1849, %r1841, %r1843;
mov.b32 %f302, %r1862;
add.ftz.f32 %f303, %f301, %f302;
mov.b32 %r1863, %f303;
shfl.sync.bfly.b32 %r1864|%p93, %r1863, %r1852, %r1841, %r1843;
mov.b32 %f304, %r1864;
add.ftz.f32 %f305, %f303, %f304;
mov.b32 %r1865, %f305;
shfl.sync.bfly.b32 %r1866|%p94, %r1865, %r1855, %r1841, %r1843;
mov.b32 %f306, %r1866;
add.ftz.f32 %f45, %f305, %f306;
setp.ne.s32 %p95, %r2, 2;
or.b32 %r1870, %r4, 2;
setp.ge.u32 %p96, %r1870, %r51;
or.pred %p97, %p95, %p96;
@%p97 bra $L__BB226_52;

ld.param.f32 %f979, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f307, %f5, %f44;
add.ftz.f32 %f308, %f4, %f45;
min.ftz.f32 %f309, %f308, %f979;
min.ftz.f32 %f310, %f307, %f979;
max.ftz.f32 %f311, %f310, %f35;
mul.ftz.f32 %f312, %f309, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f313, %f312;
add.ftz.f32 %f314, %f313, 0f3F800000;
div.approx.ftz.f32 %f315, %f309, %f314;
mul.ftz.f32 %f316, %f311, %f315;
st.global.f32 [%rd30+8], %f316;

$L__BB226_52:
ld.shared.f32 %f317, [%r32+384];
add.ftz.f32 %f318, %f317, %f27;
ld.shared.f32 %f319, [%r34+384];
add.ftz.f32 %f320, %f319, %f31;
ld.shared.f32 %f321, [%r32+896];
add.ftz.f32 %f322, %f321, %f318;
ld.shared.f32 %f323, [%r34+896];
add.ftz.f32 %f324, %f323, %f320;
ld.shared.f32 %f325, [%r32+1408];
add.ftz.f32 %f326, %f325, %f322;
ld.shared.f32 %f327, [%r34+1408];
add.ftz.f32 %f328, %f327, %f324;
mov.b32 %r1871, %f326;
shfl.sync.bfly.b32 %r1875|%p98, %r1871, %r1842, %r1841, %r1843;
mov.b32 %f329, %r1875;
add.ftz.f32 %f330, %f326, %f329;
mov.b32 %r1876, %f330;
shfl.sync.bfly.b32 %r1878|%p99, %r1876, %r1846, %r1841, %r1843;
mov.b32 %f331, %r1878;
add.ftz.f32 %f332, %f330, %f331;
mov.b32 %r1879, %f332;
shfl.sync.bfly.b32 %r1881|%p100, %r1879, %r1849, %r1841, %r1843;
mov.b32 %f333, %r1881;
add.ftz.f32 %f334, %f332, %f333;
mov.b32 %r1882, %f334;
shfl.sync.bfly.b32 %r1884|%p101, %r1882, %r1852, %r1841, %r1843;
mov.b32 %f335, %r1884;
add.ftz.f32 %f336, %f334, %f335;
mov.b32 %r1885, %f336;
shfl.sync.bfly.b32 %r1887|%p102, %r1885, %r1855, %r1841, %r1843;
mov.b32 %f337, %r1887;
add.ftz.f32 %f46, %f336, %f337;
mov.b32 %r1888, %f328;
shfl.sync.bfly.b32 %r1889|%p103, %r1888, %r1842, %r1841, %r1843;
mov.b32 %f338, %r1889;
add.ftz.f32 %f339, %f328, %f338;
mov.b32 %r1890, %f339;
shfl.sync.bfly.b32 %r1891|%p104, %r1890, %r1846, %r1841, %r1843;
mov.b32 %f340, %r1891;
add.ftz.f32 %f341, %f339, %f340;
mov.b32 %r1892, %f341;
shfl.sync.bfly.b32 %r1893|%p105, %r1892, %r1849, %r1841, %r1843;
mov.b32 %f342, %r1893;
add.ftz.f32 %f343, %f341, %f342;
mov.b32 %r1894, %f343;
shfl.sync.bfly.b32 %r1895|%p106, %r1894, %r1852, %r1841, %r1843;
mov.b32 %f344, %r1895;
add.ftz.f32 %f345, %f343, %f344;
mov.b32 %r1896, %f345;
shfl.sync.bfly.b32 %r1897|%p107, %r1896, %r1855, %r1841, %r1843;
mov.b32 %f346, %r1897;
add.ftz.f32 %f47, %f345, %f346;
setp.ne.s32 %p108, %r2, 3;
or.b32 %r1901, %r4, 3;
setp.ge.u32 %p109, %r1901, %r51;
or.pred %p110, %p108, %p109;
@%p110 bra $L__BB226_78;

ld.param.f32 %f978, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f347, %f5, %f46;
add.ftz.f32 %f348, %f4, %f47;
min.ftz.f32 %f349, %f348, %f978;
min.ftz.f32 %f350, %f347, %f978;
max.ftz.f32 %f351, %f350, %f35;
mul.ftz.f32 %f352, %f349, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f353, %f352;
add.ftz.f32 %f354, %f353, 0f3F800000;
div.approx.ftz.f32 %f355, %f349, %f354;
mul.ftz.f32 %f356, %f351, %f355;
st.global.f32 [%rd30+12], %f356;
bra.uni $L__BB226_78;

$L__BB226_70:
ld.shared.f32 %f845, [%r32];
add.ftz.f32 %f846, %f845, %f30;
ld.shared.f32 %f847, [%r34];
add.ftz.f32 %f848, %f847, %f34;
ld.shared.f32 %f849, [%r32+512];
add.ftz.f32 %f850, %f849, %f846;
ld.shared.f32 %f851, [%r34+512];
add.ftz.f32 %f852, %f851, %f848;
ld.shared.f32 %f853, [%r32+1024];
add.ftz.f32 %f854, %f853, %f850;
ld.shared.f32 %f855, [%r34+1024];
add.ftz.f32 %f856, %f855, %f852;
mov.b32 %r2271, %f854;
mov.u32 %r2272, 31;
mov.u32 %r2273, 16;
mov.u32 %r2274, -1;
shfl.sync.bfly.b32 %r2275|%p267, %r2271, %r2273, %r2272, %r2274;
mov.b32 %f857, %r2275;
add.ftz.f32 %f858, %f854, %f857;
mov.b32 %r2276, %f858;
mov.u32 %r2277, 8;
shfl.sync.bfly.b32 %r2278|%p268, %r2276, %r2277, %r2272, %r2274;
mov.b32 %f859, %r2278;
add.ftz.f32 %f860, %f858, %f859;
mov.b32 %r2279, %f860;
mov.u32 %r2280, 4;
shfl.sync.bfly.b32 %r2281|%p269, %r2279, %r2280, %r2272, %r2274;
mov.b32 %f861, %r2281;
add.ftz.f32 %f862, %f860, %f861;
mov.b32 %r2282, %f862;
mov.u32 %r2283, 2;
shfl.sync.bfly.b32 %r2284|%p270, %r2282, %r2283, %r2272, %r2274;
mov.b32 %f863, %r2284;
add.ftz.f32 %f864, %f862, %f863;
mov.b32 %r2285, %f864;
mov.u32 %r2286, 1;
shfl.sync.bfly.b32 %r2287|%p271, %r2285, %r2286, %r2272, %r2274;
mov.b32 %f865, %r2287;
add.ftz.f32 %f72, %f864, %f865;
mov.b32 %r2288, %f856;
shfl.sync.bfly.b32 %r2289|%p272, %r2288, %r2273, %r2272, %r2274;
mov.b32 %f866, %r2289;
add.ftz.f32 %f867, %f856, %f866;
mov.b32 %r2290, %f867;
shfl.sync.bfly.b32 %r2291|%p273, %r2290, %r2277, %r2272, %r2274;
mov.b32 %f868, %r2291;
add.ftz.f32 %f869, %f867, %f868;
mov.b32 %r2292, %f869;
shfl.sync.bfly.b32 %r2293|%p274, %r2292, %r2280, %r2272, %r2274;
mov.b32 %f870, %r2293;
add.ftz.f32 %f871, %f869, %f870;
mov.b32 %r2294, %f871;
shfl.sync.bfly.b32 %r2295|%p275, %r2294, %r2283, %r2272, %r2274;
mov.b32 %f872, %r2295;
add.ftz.f32 %f873, %f871, %f872;
mov.b32 %r2296, %f873;
shfl.sync.bfly.b32 %r2297|%p276, %r2296, %r2286, %r2272, %r2274;
mov.b32 %f874, %r2297;
add.ftz.f32 %f73, %f873, %f874;
setp.ne.s32 %p277, %r2, 0;
setp.ge.u32 %p278, %r4, %r51;
or.pred %p279, %p277, %p278;
@%p279 bra $L__BB226_72;

add.ftz.f32 %f875, %f5, %f72;
add.ftz.f32 %f876, %f4, %f73;
mul.ftz.f32 %f877, %f875, %f876;
st.global.f32 [%rd30], %f877;

$L__BB226_72:
ld.shared.f32 %f878, [%r32+128];
add.ftz.f32 %f879, %f878, %f29;
ld.shared.f32 %f880, [%r34+128];
add.ftz.f32 %f881, %f880, %f33;
ld.shared.f32 %f882, [%r32+640];
add.ftz.f32 %f883, %f882, %f879;
ld.shared.f32 %f884, [%r34+640];
add.ftz.f32 %f885, %f884, %f881;
ld.shared.f32 %f886, [%r32+1152];
add.ftz.f32 %f887, %f886, %f883;
ld.shared.f32 %f888, [%r34+1152];
add.ftz.f32 %f889, %f888, %f885;
mov.b32 %r2301, %f887;
shfl.sync.bfly.b32 %r2305|%p280, %r2301, %r2273, %r2272, %r2274;
mov.b32 %f890, %r2305;
add.ftz.f32 %f891, %f887, %f890;
mov.b32 %r2306, %f891;
shfl.sync.bfly.b32 %r2308|%p281, %r2306, %r2277, %r2272, %r2274;
mov.b32 %f892, %r2308;
add.ftz.f32 %f893, %f891, %f892;
mov.b32 %r2309, %f893;
shfl.sync.bfly.b32 %r2311|%p282, %r2309, %r2280, %r2272, %r2274;
mov.b32 %f894, %r2311;
add.ftz.f32 %f895, %f893, %f894;
mov.b32 %r2312, %f895;
shfl.sync.bfly.b32 %r2314|%p283, %r2312, %r2283, %r2272, %r2274;
mov.b32 %f896, %r2314;
add.ftz.f32 %f897, %f895, %f896;
mov.b32 %r2315, %f897;
shfl.sync.bfly.b32 %r2317|%p284, %r2315, %r2286, %r2272, %r2274;
mov.b32 %f898, %r2317;
add.ftz.f32 %f74, %f897, %f898;
mov.b32 %r2318, %f889;
shfl.sync.bfly.b32 %r2319|%p285, %r2318, %r2273, %r2272, %r2274;
mov.b32 %f899, %r2319;
add.ftz.f32 %f900, %f889, %f899;
mov.b32 %r2320, %f900;
shfl.sync.bfly.b32 %r2321|%p286, %r2320, %r2277, %r2272, %r2274;
mov.b32 %f901, %r2321;
add.ftz.f32 %f902, %f900, %f901;
mov.b32 %r2322, %f902;
shfl.sync.bfly.b32 %r2323|%p287, %r2322, %r2280, %r2272, %r2274;
mov.b32 %f903, %r2323;
add.ftz.f32 %f904, %f902, %f903;
mov.b32 %r2324, %f904;
shfl.sync.bfly.b32 %r2325|%p288, %r2324, %r2283, %r2272, %r2274;
mov.b32 %f905, %r2325;
add.ftz.f32 %f906, %f904, %f905;
mov.b32 %r2326, %f906;
shfl.sync.bfly.b32 %r2327|%p289, %r2326, %r2286, %r2272, %r2274;
mov.b32 %f907, %r2327;
add.ftz.f32 %f75, %f906, %f907;
setp.ne.s32 %p290, %r2, 1;
or.b32 %r2331, %r4, 1;
setp.ge.u32 %p291, %r2331, %r51;
or.pred %p292, %p290, %p291;
@%p292 bra $L__BB226_74;

add.ftz.f32 %f908, %f5, %f74;
add.ftz.f32 %f909, %f4, %f75;
mul.ftz.f32 %f910, %f908, %f909;
st.global.f32 [%rd30+4], %f910;

$L__BB226_74:
ld.shared.f32 %f911, [%r32+256];
add.ftz.f32 %f912, %f911, %f28;
ld.shared.f32 %f913, [%r34+256];
add.ftz.f32 %f914, %f913, %f32;
ld.shared.f32 %f915, [%r32+768];
add.ftz.f32 %f916, %f915, %f912;
ld.shared.f32 %f917, [%r34+768];
add.ftz.f32 %f918, %f917, %f914;
ld.shared.f32 %f919, [%r32+1280];
add.ftz.f32 %f920, %f919, %f916;
ld.shared.f32 %f921, [%r34+1280];
add.ftz.f32 %f922, %f921, %f918;
mov.b32 %r2332, %f920;
mov.u32 %r2333, 31;
mov.u32 %r2334, 16;
mov.u32 %r2335, -1;
shfl.sync.bfly.b32 %r2336|%p293, %r2332, %r2334, %r2333, %r2335;
mov.b32 %f923, %r2336;
add.ftz.f32 %f924, %f920, %f923;
mov.b32 %r2337, %f924;
mov.u32 %r2338, 8;
shfl.sync.bfly.b32 %r2339|%p294, %r2337, %r2338, %r2333, %r2335;
mov.b32 %f925, %r2339;
add.ftz.f32 %f926, %f924, %f925;
mov.b32 %r2340, %f926;
mov.u32 %r2341, 4;
shfl.sync.bfly.b32 %r2342|%p295, %r2340, %r2341, %r2333, %r2335;
mov.b32 %f927, %r2342;
add.ftz.f32 %f928, %f926, %f927;
mov.b32 %r2343, %f928;
mov.u32 %r2344, 2;
shfl.sync.bfly.b32 %r2345|%p296, %r2343, %r2344, %r2333, %r2335;
mov.b32 %f929, %r2345;
add.ftz.f32 %f930, %f928, %f929;
mov.b32 %r2346, %f930;
mov.u32 %r2347, 1;
shfl.sync.bfly.b32 %r2348|%p297, %r2346, %r2347, %r2333, %r2335;
mov.b32 %f931, %r2348;
add.ftz.f32 %f76, %f930, %f931;
mov.b32 %r2349, %f922;
shfl.sync.bfly.b32 %r2350|%p298, %r2349, %r2334, %r2333, %r2335;
mov.b32 %f932, %r2350;
add.ftz.f32 %f933, %f922, %f932;
mov.b32 %r2351, %f933;
shfl.sync.bfly.b32 %r2352|%p299, %r2351, %r2338, %r2333, %r2335;
mov.b32 %f934, %r2352;
add.ftz.f32 %f935, %f933, %f934;
mov.b32 %r2353, %f935;
shfl.sync.bfly.b32 %r2354|%p300, %r2353, %r2341, %r2333, %r2335;
mov.b32 %f936, %r2354;
add.ftz.f32 %f937, %f935, %f936;
mov.b32 %r2355, %f937;
shfl.sync.bfly.b32 %r2356|%p301, %r2355, %r2344, %r2333, %r2335;
mov.b32 %f938, %r2356;
add.ftz.f32 %f939, %f937, %f938;
mov.b32 %r2357, %f939;
shfl.sync.bfly.b32 %r2358|%p302, %r2357, %r2347, %r2333, %r2335;
mov.b32 %f940, %r2358;
add.ftz.f32 %f77, %f939, %f940;
setp.ne.s32 %p303, %r2, 2;
or.b32 %r2362, %r4, 2;
setp.ge.u32 %p304, %r2362, %r51;
or.pred %p305, %p303, %p304;
@%p305 bra $L__BB226_76;

add.ftz.f32 %f941, %f5, %f76;
add.ftz.f32 %f942, %f4, %f77;
mul.ftz.f32 %f943, %f941, %f942;
st.global.f32 [%rd30+8], %f943;

$L__BB226_76:
ld.shared.f32 %f944, [%r32+384];
add.ftz.f32 %f945, %f944, %f27;
ld.shared.f32 %f946, [%r34+384];
add.ftz.f32 %f947, %f946, %f31;
ld.shared.f32 %f948, [%r32+896];
add.ftz.f32 %f949, %f948, %f945;
ld.shared.f32 %f950, [%r34+896];
add.ftz.f32 %f951, %f950, %f947;
ld.shared.f32 %f952, [%r32+1408];
add.ftz.f32 %f953, %f952, %f949;
ld.shared.f32 %f954, [%r34+1408];
add.ftz.f32 %f955, %f954, %f951;
mov.b32 %r2363, %f953;
shfl.sync.bfly.b32 %r2367|%p306, %r2363, %r2334, %r2333, %r2335;
mov.b32 %f956, %r2367;
add.ftz.f32 %f957, %f953, %f956;
mov.b32 %r2368, %f957;
shfl.sync.bfly.b32 %r2370|%p307, %r2368, %r2338, %r2333, %r2335;
mov.b32 %f958, %r2370;
add.ftz.f32 %f959, %f957, %f958;
mov.b32 %r2371, %f959;
shfl.sync.bfly.b32 %r2373|%p308, %r2371, %r2341, %r2333, %r2335;
mov.b32 %f960, %r2373;
add.ftz.f32 %f961, %f959, %f960;
mov.b32 %r2374, %f961;
shfl.sync.bfly.b32 %r2376|%p309, %r2374, %r2344, %r2333, %r2335;
mov.b32 %f962, %r2376;
add.ftz.f32 %f963, %f961, %f962;
mov.b32 %r2377, %f963;
shfl.sync.bfly.b32 %r2379|%p310, %r2377, %r2347, %r2333, %r2335;
mov.b32 %f964, %r2379;
add.ftz.f32 %f78, %f963, %f964;
mov.b32 %r2380, %f955;
shfl.sync.bfly.b32 %r2381|%p311, %r2380, %r2334, %r2333, %r2335;
mov.b32 %f965, %r2381;
add.ftz.f32 %f966, %f955, %f965;
mov.b32 %r2382, %f966;
shfl.sync.bfly.b32 %r2383|%p312, %r2382, %r2338, %r2333, %r2335;
mov.b32 %f967, %r2383;
add.ftz.f32 %f968, %f966, %f967;
mov.b32 %r2384, %f968;
shfl.sync.bfly.b32 %r2385|%p313, %r2384, %r2341, %r2333, %r2335;
mov.b32 %f969, %r2385;
add.ftz.f32 %f970, %f968, %f969;
mov.b32 %r2386, %f970;
shfl.sync.bfly.b32 %r2387|%p314, %r2386, %r2344, %r2333, %r2335;
mov.b32 %f971, %r2387;
add.ftz.f32 %f972, %f970, %f971;
mov.b32 %r2388, %f972;
shfl.sync.bfly.b32 %r2389|%p315, %r2388, %r2347, %r2333, %r2335;
mov.b32 %f973, %r2389;
add.ftz.f32 %f79, %f972, %f973;
setp.ne.s32 %p316, %r2, 3;
or.b32 %r2393, %r4, 3;
setp.ge.u32 %p317, %r2393, %r51;
or.pred %p318, %p316, %p317;
@%p318 bra $L__BB226_78;

add.ftz.f32 %f974, %f5, %f78;
add.ftz.f32 %f975, %f4, %f79;
mul.ftz.f32 %f976, %f974, %f975;
st.global.f32 [%rd30+12], %f976;
bra.uni $L__BB226_78;

$L__BB226_62:
ld.shared.f32 %f681, [%r32];
add.ftz.f32 %f682, %f681, %f30;
ld.shared.f32 %f683, [%r34];
add.ftz.f32 %f684, %f683, %f34;
ld.shared.f32 %f685, [%r32+512];
add.ftz.f32 %f686, %f685, %f682;
ld.shared.f32 %f687, [%r34+512];
add.ftz.f32 %f688, %f687, %f684;
ld.shared.f32 %f689, [%r32+1024];
add.ftz.f32 %f690, %f689, %f686;
ld.shared.f32 %f691, [%r34+1024];
add.ftz.f32 %f692, %f691, %f688;
mov.b32 %r2148, %f690;
mov.u32 %r2149, 31;
mov.u32 %r2150, 16;
mov.u32 %r2151, -1;
shfl.sync.bfly.b32 %r2152|%p215, %r2148, %r2150, %r2149, %r2151;
mov.b32 %f693, %r2152;
add.ftz.f32 %f694, %f690, %f693;
mov.b32 %r2153, %f694;
mov.u32 %r2154, 8;
shfl.sync.bfly.b32 %r2155|%p216, %r2153, %r2154, %r2149, %r2151;
mov.b32 %f695, %r2155;
add.ftz.f32 %f696, %f694, %f695;
mov.b32 %r2156, %f696;
mov.u32 %r2157, 4;
shfl.sync.bfly.b32 %r2158|%p217, %r2156, %r2157, %r2149, %r2151;
mov.b32 %f697, %r2158;
add.ftz.f32 %f698, %f696, %f697;
mov.b32 %r2159, %f698;
mov.u32 %r2160, 2;
shfl.sync.bfly.b32 %r2161|%p218, %r2159, %r2160, %r2149, %r2151;
mov.b32 %f699, %r2161;
add.ftz.f32 %f700, %f698, %f699;
mov.b32 %r2162, %f700;
mov.u32 %r2163, 1;
shfl.sync.bfly.b32 %r2164|%p219, %r2162, %r2163, %r2149, %r2151;
mov.b32 %f701, %r2164;
add.ftz.f32 %f64, %f700, %f701;
mov.b32 %r2165, %f692;
shfl.sync.bfly.b32 %r2166|%p220, %r2165, %r2150, %r2149, %r2151;
mov.b32 %f702, %r2166;
add.ftz.f32 %f703, %f692, %f702;
mov.b32 %r2167, %f703;
shfl.sync.bfly.b32 %r2168|%p221, %r2167, %r2154, %r2149, %r2151;
mov.b32 %f704, %r2168;
add.ftz.f32 %f705, %f703, %f704;
mov.b32 %r2169, %f705;
shfl.sync.bfly.b32 %r2170|%p222, %r2169, %r2157, %r2149, %r2151;
mov.b32 %f706, %r2170;
add.ftz.f32 %f707, %f705, %f706;
mov.b32 %r2171, %f707;
shfl.sync.bfly.b32 %r2172|%p223, %r2171, %r2160, %r2149, %r2151;
mov.b32 %f708, %r2172;
add.ftz.f32 %f709, %f707, %f708;
mov.b32 %r2173, %f709;
shfl.sync.bfly.b32 %r2174|%p224, %r2173, %r2163, %r2149, %r2151;
mov.b32 %f710, %r2174;
add.ftz.f32 %f65, %f709, %f710;
setp.ne.s32 %p225, %r2, 0;
setp.ge.u32 %p226, %r4, %r51;
or.pred %p227, %p225, %p226;
@%p227 bra $L__BB226_64;

add.ftz.f32 %f711, %f5, %f64;
add.ftz.f32 %f712, %f4, %f65;
mul.ftz.f32 %f713, %f712, 0f3F000000;
mul.ftz.f32 %f714, %f712, 0f3F4C422A;
mul.ftz.f32 %f715, %f712, 0f3D372713;
fma.rn.ftz.f32 %f716, %f712, %f715, 0f3F800000;
mul.ftz.f32 %f717, %f714, %f716;
tanh.approx.f32 %f718, %f717;
add.ftz.f32 %f719, %f718, 0f3F800000;
mul.ftz.f32 %f720, %f713, %f719;
mul.ftz.f32 %f721, %f711, %f720;
st.global.f32 [%rd30], %f721;

$L__BB226_64:
ld.shared.f32 %f722, [%r32+128];
add.ftz.f32 %f723, %f722, %f29;
ld.shared.f32 %f724, [%r34+128];
add.ftz.f32 %f725, %f724, %f33;
ld.shared.f32 %f726, [%r32+640];
add.ftz.f32 %f727, %f726, %f723;
ld.shared.f32 %f728, [%r34+640];
add.ftz.f32 %f729, %f728, %f725;
ld.shared.f32 %f730, [%r32+1152];
add.ftz.f32 %f731, %f730, %f727;
ld.shared.f32 %f732, [%r34+1152];
add.ftz.f32 %f733, %f732, %f729;
mov.b32 %r2178, %f731;
shfl.sync.bfly.b32 %r2182|%p228, %r2178, %r2150, %r2149, %r2151;
mov.b32 %f734, %r2182;
add.ftz.f32 %f735, %f731, %f734;
mov.b32 %r2183, %f735;
shfl.sync.bfly.b32 %r2185|%p229, %r2183, %r2154, %r2149, %r2151;
mov.b32 %f736, %r2185;
add.ftz.f32 %f737, %f735, %f736;
mov.b32 %r2186, %f737;
shfl.sync.bfly.b32 %r2188|%p230, %r2186, %r2157, %r2149, %r2151;
mov.b32 %f738, %r2188;
add.ftz.f32 %f739, %f737, %f738;
mov.b32 %r2189, %f739;
shfl.sync.bfly.b32 %r2191|%p231, %r2189, %r2160, %r2149, %r2151;
mov.b32 %f740, %r2191;
add.ftz.f32 %f741, %f739, %f740;
mov.b32 %r2192, %f741;
shfl.sync.bfly.b32 %r2194|%p232, %r2192, %r2163, %r2149, %r2151;
mov.b32 %f742, %r2194;
add.ftz.f32 %f66, %f741, %f742;
mov.b32 %r2195, %f733;
shfl.sync.bfly.b32 %r2196|%p233, %r2195, %r2150, %r2149, %r2151;
mov.b32 %f743, %r2196;
add.ftz.f32 %f744, %f733, %f743;
mov.b32 %r2197, %f744;
shfl.sync.bfly.b32 %r2198|%p234, %r2197, %r2154, %r2149, %r2151;
mov.b32 %f745, %r2198;
add.ftz.f32 %f746, %f744, %f745;
mov.b32 %r2199, %f746;
shfl.sync.bfly.b32 %r2200|%p235, %r2199, %r2157, %r2149, %r2151;
mov.b32 %f747, %r2200;
add.ftz.f32 %f748, %f746, %f747;
mov.b32 %r2201, %f748;
shfl.sync.bfly.b32 %r2202|%p236, %r2201, %r2160, %r2149, %r2151;
mov.b32 %f749, %r2202;
add.ftz.f32 %f750, %f748, %f749;
mov.b32 %r2203, %f750;
shfl.sync.bfly.b32 %r2204|%p237, %r2203, %r2163, %r2149, %r2151;
mov.b32 %f751, %r2204;
add.ftz.f32 %f67, %f750, %f751;
setp.ne.s32 %p238, %r2, 1;
or.b32 %r2208, %r4, 1;
setp.ge.u32 %p239, %r2208, %r51;
or.pred %p240, %p238, %p239;
@%p240 bra $L__BB226_66;

add.ftz.f32 %f752, %f5, %f66;
add.ftz.f32 %f753, %f4, %f67;
mul.ftz.f32 %f754, %f753, 0f3F000000;
mul.ftz.f32 %f755, %f753, 0f3F4C422A;
mul.ftz.f32 %f756, %f753, 0f3D372713;
fma.rn.ftz.f32 %f757, %f753, %f756, 0f3F800000;
mul.ftz.f32 %f758, %f755, %f757;
tanh.approx.f32 %f759, %f758;
add.ftz.f32 %f760, %f759, 0f3F800000;
mul.ftz.f32 %f761, %f754, %f760;
mul.ftz.f32 %f762, %f752, %f761;
st.global.f32 [%rd30+4], %f762;

$L__BB226_66:
ld.shared.f32 %f763, [%r32+256];
add.ftz.f32 %f764, %f763, %f28;
ld.shared.f32 %f765, [%r34+256];
add.ftz.f32 %f766, %f765, %f32;
ld.shared.f32 %f767, [%r32+768];
add.ftz.f32 %f768, %f767, %f764;
ld.shared.f32 %f769, [%r34+768];
add.ftz.f32 %f770, %f769, %f766;
ld.shared.f32 %f771, [%r32+1280];
add.ftz.f32 %f772, %f771, %f768;
ld.shared.f32 %f773, [%r34+1280];
add.ftz.f32 %f774, %f773, %f770;
mov.b32 %r2209, %f772;
mov.u32 %r2210, 31;
mov.u32 %r2211, 16;
mov.u32 %r2212, -1;
shfl.sync.bfly.b32 %r2213|%p241, %r2209, %r2211, %r2210, %r2212;
mov.b32 %f775, %r2213;
add.ftz.f32 %f776, %f772, %f775;
mov.b32 %r2214, %f776;
mov.u32 %r2215, 8;
shfl.sync.bfly.b32 %r2216|%p242, %r2214, %r2215, %r2210, %r2212;
mov.b32 %f777, %r2216;
add.ftz.f32 %f778, %f776, %f777;
mov.b32 %r2217, %f778;
mov.u32 %r2218, 4;
shfl.sync.bfly.b32 %r2219|%p243, %r2217, %r2218, %r2210, %r2212;
mov.b32 %f779, %r2219;
add.ftz.f32 %f780, %f778, %f779;
mov.b32 %r2220, %f780;
mov.u32 %r2221, 2;
shfl.sync.bfly.b32 %r2222|%p244, %r2220, %r2221, %r2210, %r2212;
mov.b32 %f781, %r2222;
add.ftz.f32 %f782, %f780, %f781;
mov.b32 %r2223, %f782;
mov.u32 %r2224, 1;
shfl.sync.bfly.b32 %r2225|%p245, %r2223, %r2224, %r2210, %r2212;
mov.b32 %f783, %r2225;
add.ftz.f32 %f68, %f782, %f783;
mov.b32 %r2226, %f774;
shfl.sync.bfly.b32 %r2227|%p246, %r2226, %r2211, %r2210, %r2212;
mov.b32 %f784, %r2227;
add.ftz.f32 %f785, %f774, %f784;
mov.b32 %r2228, %f785;
shfl.sync.bfly.b32 %r2229|%p247, %r2228, %r2215, %r2210, %r2212;
mov.b32 %f786, %r2229;
add.ftz.f32 %f787, %f785, %f786;
mov.b32 %r2230, %f787;
shfl.sync.bfly.b32 %r2231|%p248, %r2230, %r2218, %r2210, %r2212;
mov.b32 %f788, %r2231;
add.ftz.f32 %f789, %f787, %f788;
mov.b32 %r2232, %f789;
shfl.sync.bfly.b32 %r2233|%p249, %r2232, %r2221, %r2210, %r2212;
mov.b32 %f790, %r2233;
add.ftz.f32 %f791, %f789, %f790;
mov.b32 %r2234, %f791;
shfl.sync.bfly.b32 %r2235|%p250, %r2234, %r2224, %r2210, %r2212;
mov.b32 %f792, %r2235;
add.ftz.f32 %f69, %f791, %f792;
setp.ne.s32 %p251, %r2, 2;
or.b32 %r2239, %r4, 2;
setp.ge.u32 %p252, %r2239, %r51;
or.pred %p253, %p251, %p252;
@%p253 bra $L__BB226_68;

add.ftz.f32 %f793, %f5, %f68;
add.ftz.f32 %f794, %f4, %f69;
mul.ftz.f32 %f795, %f794, 0f3F000000;
mul.ftz.f32 %f796, %f794, 0f3F4C422A;
mul.ftz.f32 %f797, %f794, 0f3D372713;
fma.rn.ftz.f32 %f798, %f794, %f797, 0f3F800000;
mul.ftz.f32 %f799, %f796, %f798;
tanh.approx.f32 %f800, %f799;
add.ftz.f32 %f801, %f800, 0f3F800000;
mul.ftz.f32 %f802, %f795, %f801;
mul.ftz.f32 %f803, %f793, %f802;
st.global.f32 [%rd30+8], %f803;

$L__BB226_68:
ld.shared.f32 %f804, [%r32+384];
add.ftz.f32 %f805, %f804, %f27;
ld.shared.f32 %f806, [%r34+384];
add.ftz.f32 %f807, %f806, %f31;
ld.shared.f32 %f808, [%r32+896];
add.ftz.f32 %f809, %f808, %f805;
ld.shared.f32 %f810, [%r34+896];
add.ftz.f32 %f811, %f810, %f807;
ld.shared.f32 %f812, [%r32+1408];
add.ftz.f32 %f813, %f812, %f809;
ld.shared.f32 %f814, [%r34+1408];
add.ftz.f32 %f815, %f814, %f811;
mov.b32 %r2240, %f813;
shfl.sync.bfly.b32 %r2244|%p254, %r2240, %r2211, %r2210, %r2212;
mov.b32 %f816, %r2244;
add.ftz.f32 %f817, %f813, %f816;
mov.b32 %r2245, %f817;
shfl.sync.bfly.b32 %r2247|%p255, %r2245, %r2215, %r2210, %r2212;
mov.b32 %f818, %r2247;
add.ftz.f32 %f819, %f817, %f818;
mov.b32 %r2248, %f819;
shfl.sync.bfly.b32 %r2250|%p256, %r2248, %r2218, %r2210, %r2212;
mov.b32 %f820, %r2250;
add.ftz.f32 %f821, %f819, %f820;
mov.b32 %r2251, %f821;
shfl.sync.bfly.b32 %r2253|%p257, %r2251, %r2221, %r2210, %r2212;
mov.b32 %f822, %r2253;
add.ftz.f32 %f823, %f821, %f822;
mov.b32 %r2254, %f823;
shfl.sync.bfly.b32 %r2256|%p258, %r2254, %r2224, %r2210, %r2212;
mov.b32 %f824, %r2256;
add.ftz.f32 %f70, %f823, %f824;
mov.b32 %r2257, %f815;
shfl.sync.bfly.b32 %r2258|%p259, %r2257, %r2211, %r2210, %r2212;
mov.b32 %f825, %r2258;
add.ftz.f32 %f826, %f815, %f825;
mov.b32 %r2259, %f826;
shfl.sync.bfly.b32 %r2260|%p260, %r2259, %r2215, %r2210, %r2212;
mov.b32 %f827, %r2260;
add.ftz.f32 %f828, %f826, %f827;
mov.b32 %r2261, %f828;
shfl.sync.bfly.b32 %r2262|%p261, %r2261, %r2218, %r2210, %r2212;
mov.b32 %f829, %r2262;
add.ftz.f32 %f830, %f828, %f829;
mov.b32 %r2263, %f830;
shfl.sync.bfly.b32 %r2264|%p262, %r2263, %r2221, %r2210, %r2212;
mov.b32 %f831, %r2264;
add.ftz.f32 %f832, %f830, %f831;
mov.b32 %r2265, %f832;
shfl.sync.bfly.b32 %r2266|%p263, %r2265, %r2224, %r2210, %r2212;
mov.b32 %f833, %r2266;
add.ftz.f32 %f71, %f832, %f833;
setp.ne.s32 %p264, %r2, 3;
or.b32 %r2270, %r4, 3;
setp.ge.u32 %p265, %r2270, %r51;
or.pred %p266, %p264, %p265;
@%p266 bra $L__BB226_78;

add.ftz.f32 %f834, %f5, %f70;
add.ftz.f32 %f835, %f4, %f71;
mul.ftz.f32 %f836, %f835, 0f3F000000;
mul.ftz.f32 %f837, %f835, 0f3F4C422A;
mul.ftz.f32 %f838, %f835, 0f3D372713;
fma.rn.ftz.f32 %f839, %f835, %f838, 0f3F800000;
mul.ftz.f32 %f840, %f837, %f839;
tanh.approx.f32 %f841, %f840;
add.ftz.f32 %f842, %f841, 0f3F800000;
mul.ftz.f32 %f843, %f836, %f842;
mul.ftz.f32 %f844, %f834, %f843;
st.global.f32 [%rd30+12], %f844;
bra.uni $L__BB226_78;

$L__BB226_54:
ld.shared.f32 %f357, [%r32];
add.ftz.f32 %f358, %f357, %f30;
ld.shared.f32 %f359, [%r34];
add.ftz.f32 %f360, %f359, %f34;
ld.shared.f32 %f361, [%r32+512];
add.ftz.f32 %f362, %f361, %f358;
ld.shared.f32 %f363, [%r34+512];
add.ftz.f32 %f364, %f363, %f360;
ld.shared.f32 %f365, [%r32+1024];
add.ftz.f32 %f366, %f365, %f362;
ld.shared.f32 %f367, [%r34+1024];
add.ftz.f32 %f368, %f367, %f364;
mov.b32 %r1902, %f366;
mov.u32 %r1903, 31;
mov.u32 %r1904, 16;
mov.u32 %r1905, -1;
shfl.sync.bfly.b32 %r1906|%p111, %r1902, %r1904, %r1903, %r1905;
mov.b32 %f369, %r1906;
add.ftz.f32 %f370, %f366, %f369;
mov.b32 %r1907, %f370;
mov.u32 %r1908, 8;
shfl.sync.bfly.b32 %r1909|%p112, %r1907, %r1908, %r1903, %r1905;
mov.b32 %f371, %r1909;
add.ftz.f32 %f372, %f370, %f371;
mov.b32 %r1910, %f372;
mov.u32 %r1911, 4;
shfl.sync.bfly.b32 %r1912|%p113, %r1910, %r1911, %r1903, %r1905;
mov.b32 %f373, %r1912;
add.ftz.f32 %f374, %f372, %f373;
mov.b32 %r1913, %f374;
mov.u32 %r1914, 2;
shfl.sync.bfly.b32 %r1915|%p114, %r1913, %r1914, %r1903, %r1905;
mov.b32 %f375, %r1915;
add.ftz.f32 %f376, %f374, %f375;
mov.b32 %r1916, %f376;
mov.u32 %r1917, 1;
shfl.sync.bfly.b32 %r1918|%p115, %r1916, %r1917, %r1903, %r1905;
mov.b32 %f377, %r1918;
add.ftz.f32 %f48, %f376, %f377;
mov.b32 %r1919, %f368;
shfl.sync.bfly.b32 %r1920|%p116, %r1919, %r1904, %r1903, %r1905;
mov.b32 %f378, %r1920;
add.ftz.f32 %f379, %f368, %f378;
mov.b32 %r1921, %f379;
shfl.sync.bfly.b32 %r1922|%p117, %r1921, %r1908, %r1903, %r1905;
mov.b32 %f380, %r1922;
add.ftz.f32 %f381, %f379, %f380;
mov.b32 %r1923, %f381;
shfl.sync.bfly.b32 %r1924|%p118, %r1923, %r1911, %r1903, %r1905;
mov.b32 %f382, %r1924;
add.ftz.f32 %f383, %f381, %f382;
mov.b32 %r1925, %f383;
shfl.sync.bfly.b32 %r1926|%p119, %r1925, %r1914, %r1903, %r1905;
mov.b32 %f384, %r1926;
add.ftz.f32 %f385, %f383, %f384;
mov.b32 %r1927, %f385;
shfl.sync.bfly.b32 %r1928|%p120, %r1927, %r1917, %r1903, %r1905;
mov.b32 %f386, %r1928;
add.ftz.f32 %f49, %f385, %f386;
setp.ne.s32 %p121, %r2, 0;
setp.ge.u32 %p122, %r4, %r51;
or.pred %p123, %p121, %p122;
@%p123 bra $L__BB226_56;

add.ftz.f32 %f387, %f5, %f48;
add.ftz.f32 %f388, %f4, %f49;
mov.f32 %f389, 0f40E00000;
min.ftz.f32 %f390, %f388, %f389;
min.ftz.f32 %f391, %f387, %f389;
mov.f32 %f392, 0fC0E00000;
max.ftz.f32 %f393, %f391, %f392;
mul.ftz.f32 %f394, %f390, 0f3FD9DB23;
mul.ftz.f32 %f395, %f394, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f396, %f395;
add.ftz.f32 %f397, %f396, 0f3F800000;
div.approx.ftz.f32 %f398, %f390, %f397;
add.ftz.f32 %f399, %f393, 0f3F800000;
mul.ftz.f32 %f400, %f399, %f398;
st.global.f32 [%rd30], %f400;

$L__BB226_56:
ld.shared.f32 %f401, [%r32+128];
add.ftz.f32 %f402, %f401, %f29;
ld.shared.f32 %f403, [%r34+128];
add.ftz.f32 %f404, %f403, %f33;
ld.shared.f32 %f405, [%r32+640];
add.ftz.f32 %f406, %f405, %f402;
ld.shared.f32 %f407, [%r34+640];
add.ftz.f32 %f408, %f407, %f404;
ld.shared.f32 %f409, [%r32+1152];
add.ftz.f32 %f410, %f409, %f406;
ld.shared.f32 %f411, [%r34+1152];
add.ftz.f32 %f412, %f411, %f408;
mov.b32 %r1932, %f410;
shfl.sync.bfly.b32 %r1936|%p124, %r1932, %r1904, %r1903, %r1905;
mov.b32 %f413, %r1936;
add.ftz.f32 %f414, %f410, %f413;
mov.b32 %r1937, %f414;
shfl.sync.bfly.b32 %r1939|%p125, %r1937, %r1908, %r1903, %r1905;
mov.b32 %f415, %r1939;
add.ftz.f32 %f416, %f414, %f415;
mov.b32 %r1940, %f416;
shfl.sync.bfly.b32 %r1942|%p126, %r1940, %r1911, %r1903, %r1905;
mov.b32 %f417, %r1942;
add.ftz.f32 %f418, %f416, %f417;
mov.b32 %r1943, %f418;
shfl.sync.bfly.b32 %r1945|%p127, %r1943, %r1914, %r1903, %r1905;
mov.b32 %f419, %r1945;
add.ftz.f32 %f420, %f418, %f419;
mov.b32 %r1946, %f420;
shfl.sync.bfly.b32 %r1948|%p128, %r1946, %r1917, %r1903, %r1905;
mov.b32 %f421, %r1948;
add.ftz.f32 %f50, %f420, %f421;
mov.b32 %r1949, %f412;
shfl.sync.bfly.b32 %r1950|%p129, %r1949, %r1904, %r1903, %r1905;
mov.b32 %f422, %r1950;
add.ftz.f32 %f423, %f412, %f422;
mov.b32 %r1951, %f423;
shfl.sync.bfly.b32 %r1952|%p130, %r1951, %r1908, %r1903, %r1905;
mov.b32 %f424, %r1952;
add.ftz.f32 %f425, %f423, %f424;
mov.b32 %r1953, %f425;
shfl.sync.bfly.b32 %r1954|%p131, %r1953, %r1911, %r1903, %r1905;
mov.b32 %f426, %r1954;
add.ftz.f32 %f427, %f425, %f426;
mov.b32 %r1955, %f427;
shfl.sync.bfly.b32 %r1956|%p132, %r1955, %r1914, %r1903, %r1905;
mov.b32 %f428, %r1956;
add.ftz.f32 %f429, %f427, %f428;
mov.b32 %r1957, %f429;
shfl.sync.bfly.b32 %r1958|%p133, %r1957, %r1917, %r1903, %r1905;
mov.b32 %f430, %r1958;
add.ftz.f32 %f51, %f429, %f430;
setp.ne.s32 %p134, %r2, 1;
or.b32 %r1962, %r4, 1;
setp.ge.u32 %p135, %r1962, %r51;
or.pred %p136, %p134, %p135;
@%p136 bra $L__BB226_58;

add.ftz.f32 %f431, %f5, %f50;
add.ftz.f32 %f432, %f4, %f51;
mov.f32 %f433, 0f40E00000;
min.ftz.f32 %f434, %f432, %f433;
min.ftz.f32 %f435, %f431, %f433;
mov.f32 %f436, 0fC0E00000;
max.ftz.f32 %f437, %f435, %f436;
mul.ftz.f32 %f438, %f434, 0f3FD9DB23;
mul.ftz.f32 %f439, %f438, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f440, %f439;
add.ftz.f32 %f441, %f440, 0f3F800000;
div.approx.ftz.f32 %f442, %f434, %f441;
add.ftz.f32 %f443, %f437, 0f3F800000;
mul.ftz.f32 %f444, %f443, %f442;
st.global.f32 [%rd30+4], %f444;

$L__BB226_58:
ld.shared.f32 %f445, [%r32+256];
add.ftz.f32 %f446, %f445, %f28;
ld.shared.f32 %f447, [%r34+256];
add.ftz.f32 %f448, %f447, %f32;
ld.shared.f32 %f449, [%r32+768];
add.ftz.f32 %f450, %f449, %f446;
ld.shared.f32 %f451, [%r34+768];
add.ftz.f32 %f452, %f451, %f448;
ld.shared.f32 %f453, [%r32+1280];
add.ftz.f32 %f454, %f453, %f450;
ld.shared.f32 %f455, [%r34+1280];
add.ftz.f32 %f456, %f455, %f452;
mov.b32 %r1963, %f454;
mov.u32 %r1964, 31;
mov.u32 %r1965, 16;
mov.u32 %r1966, -1;
shfl.sync.bfly.b32 %r1967|%p137, %r1963, %r1965, %r1964, %r1966;
mov.b32 %f457, %r1967;
add.ftz.f32 %f458, %f454, %f457;
mov.b32 %r1968, %f458;
mov.u32 %r1969, 8;
shfl.sync.bfly.b32 %r1970|%p138, %r1968, %r1969, %r1964, %r1966;
mov.b32 %f459, %r1970;
add.ftz.f32 %f460, %f458, %f459;
mov.b32 %r1971, %f460;
mov.u32 %r1972, 4;
shfl.sync.bfly.b32 %r1973|%p139, %r1971, %r1972, %r1964, %r1966;
mov.b32 %f461, %r1973;
add.ftz.f32 %f462, %f460, %f461;
mov.b32 %r1974, %f462;
mov.u32 %r1975, 2;
shfl.sync.bfly.b32 %r1976|%p140, %r1974, %r1975, %r1964, %r1966;
mov.b32 %f463, %r1976;
add.ftz.f32 %f464, %f462, %f463;
mov.b32 %r1977, %f464;
mov.u32 %r1978, 1;
shfl.sync.bfly.b32 %r1979|%p141, %r1977, %r1978, %r1964, %r1966;
mov.b32 %f465, %r1979;
add.ftz.f32 %f52, %f464, %f465;
mov.b32 %r1980, %f456;
shfl.sync.bfly.b32 %r1981|%p142, %r1980, %r1965, %r1964, %r1966;
mov.b32 %f466, %r1981;
add.ftz.f32 %f467, %f456, %f466;
mov.b32 %r1982, %f467;
shfl.sync.bfly.b32 %r1983|%p143, %r1982, %r1969, %r1964, %r1966;
mov.b32 %f468, %r1983;
add.ftz.f32 %f469, %f467, %f468;
mov.b32 %r1984, %f469;
shfl.sync.bfly.b32 %r1985|%p144, %r1984, %r1972, %r1964, %r1966;
mov.b32 %f470, %r1985;
add.ftz.f32 %f471, %f469, %f470;
mov.b32 %r1986, %f471;
shfl.sync.bfly.b32 %r1987|%p145, %r1986, %r1975, %r1964, %r1966;
mov.b32 %f472, %r1987;
add.ftz.f32 %f473, %f471, %f472;
mov.b32 %r1988, %f473;
shfl.sync.bfly.b32 %r1989|%p146, %r1988, %r1978, %r1964, %r1966;
mov.b32 %f474, %r1989;
add.ftz.f32 %f53, %f473, %f474;
setp.ne.s32 %p147, %r2, 2;
or.b32 %r1993, %r4, 2;
setp.ge.u32 %p148, %r1993, %r51;
or.pred %p149, %p147, %p148;
@%p149 bra $L__BB226_60;

add.ftz.f32 %f475, %f5, %f52;
add.ftz.f32 %f476, %f4, %f53;
mov.f32 %f477, 0f40E00000;
min.ftz.f32 %f478, %f476, %f477;
min.ftz.f32 %f479, %f475, %f477;
mov.f32 %f480, 0fC0E00000;
max.ftz.f32 %f481, %f479, %f480;
mul.ftz.f32 %f482, %f478, 0f3FD9DB23;
mul.ftz.f32 %f483, %f482, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f484, %f483;
add.ftz.f32 %f485, %f484, 0f3F800000;
div.approx.ftz.f32 %f486, %f478, %f485;
add.ftz.f32 %f487, %f481, 0f3F800000;
mul.ftz.f32 %f488, %f487, %f486;
st.global.f32 [%rd30+8], %f488;

$L__BB226_60:
ld.shared.f32 %f489, [%r32+384];
add.ftz.f32 %f490, %f489, %f27;
ld.shared.f32 %f491, [%r34+384];
add.ftz.f32 %f492, %f491, %f31;
ld.shared.f32 %f493, [%r32+896];
add.ftz.f32 %f494, %f493, %f490;
ld.shared.f32 %f495, [%r34+896];
add.ftz.f32 %f496, %f495, %f492;
ld.shared.f32 %f497, [%r32+1408];
add.ftz.f32 %f498, %f497, %f494;
ld.shared.f32 %f499, [%r34+1408];
add.ftz.f32 %f500, %f499, %f496;
mov.b32 %r1994, %f498;
shfl.sync.bfly.b32 %r1998|%p150, %r1994, %r1965, %r1964, %r1966;
mov.b32 %f501, %r1998;
add.ftz.f32 %f502, %f498, %f501;
mov.b32 %r1999, %f502;
shfl.sync.bfly.b32 %r2001|%p151, %r1999, %r1969, %r1964, %r1966;
mov.b32 %f503, %r2001;
add.ftz.f32 %f504, %f502, %f503;
mov.b32 %r2002, %f504;
shfl.sync.bfly.b32 %r2004|%p152, %r2002, %r1972, %r1964, %r1966;
mov.b32 %f505, %r2004;
add.ftz.f32 %f506, %f504, %f505;
mov.b32 %r2005, %f506;
shfl.sync.bfly.b32 %r2007|%p153, %r2005, %r1975, %r1964, %r1966;
mov.b32 %f507, %r2007;
add.ftz.f32 %f508, %f506, %f507;
mov.b32 %r2008, %f508;
shfl.sync.bfly.b32 %r2010|%p154, %r2008, %r1978, %r1964, %r1966;
mov.b32 %f509, %r2010;
add.ftz.f32 %f54, %f508, %f509;
mov.b32 %r2011, %f500;
shfl.sync.bfly.b32 %r2012|%p155, %r2011, %r1965, %r1964, %r1966;
mov.b32 %f510, %r2012;
add.ftz.f32 %f511, %f500, %f510;
mov.b32 %r2013, %f511;
shfl.sync.bfly.b32 %r2014|%p156, %r2013, %r1969, %r1964, %r1966;
mov.b32 %f512, %r2014;
add.ftz.f32 %f513, %f511, %f512;
mov.b32 %r2015, %f513;
shfl.sync.bfly.b32 %r2016|%p157, %r2015, %r1972, %r1964, %r1966;
mov.b32 %f514, %r2016;
add.ftz.f32 %f515, %f513, %f514;
mov.b32 %r2017, %f515;
shfl.sync.bfly.b32 %r2018|%p158, %r2017, %r1975, %r1964, %r1966;
mov.b32 %f516, %r2018;
add.ftz.f32 %f517, %f515, %f516;
mov.b32 %r2019, %f517;
shfl.sync.bfly.b32 %r2020|%p159, %r2019, %r1978, %r1964, %r1966;
mov.b32 %f518, %r2020;
add.ftz.f32 %f55, %f517, %f518;
setp.ne.s32 %p160, %r2, 3;
or.b32 %r2024, %r4, 3;
setp.ge.u32 %p161, %r2024, %r51;
or.pred %p162, %p160, %p161;
@%p162 bra $L__BB226_78;

add.ftz.f32 %f519, %f5, %f54;
add.ftz.f32 %f520, %f4, %f55;
mov.f32 %f521, 0f40E00000;
min.ftz.f32 %f522, %f520, %f521;
min.ftz.f32 %f523, %f519, %f521;
mov.f32 %f524, 0fC0E00000;
max.ftz.f32 %f525, %f523, %f524;
mul.ftz.f32 %f526, %f522, 0f3FD9DB23;
mul.ftz.f32 %f527, %f526, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f528, %f527;
add.ftz.f32 %f529, %f528, 0f3F800000;
div.approx.ftz.f32 %f530, %f522, %f529;
add.ftz.f32 %f531, %f525, 0f3F800000;
mul.ftz.f32 %f532, %f531, %f530;
st.global.f32 [%rd30+12], %f532;
bra.uni $L__BB226_78;

}
.entry _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<38>;
.reg .b16 %rs<81>;
.reg .f32 %f<109>;
.reg .b32 %r<931>;
.reg .b64 %rd<79>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

ld.param.u64 %rd12, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd13, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd14, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r19, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
mov.u32 %r1, %tid.y;
shl.b32 %r37, %r1, 5;
mov.u32 %r38, %tid.x;
add.s32 %r2, %r37, %r38;
mov.u32 %r39, %ctaid.x;
shl.b32 %r3, %r39, 2;
shr.u32 %r4, %r36, 8;
mov.u32 %r928, %ctaid.y;
setp.eq.s64 %p1, %rd14, 0;
@%p1 bra $L__BB227_2;

cvta.to.global.u64 %rd16, %rd14;
mul.wide.u32 %rd17, %r928, 4;
add.s64 %rd18, %rd16, %rd17;
ld.global.nc.u32 %r927, [%rd18];
mul.hi.u32 %r40, %r928, %r19;
add.s32 %r41, %r40, %r928;
shr.u32 %r42, %r41, %r20;
mul.lo.s32 %r43, %r42, %r21;
sub.s32 %r928, %r928, %r43;
bra.uni $L__BB227_3;

$L__BB227_2:
mul.hi.u32 %r44, %r928, %r24;
add.s32 %r45, %r44, %r928;
shr.u32 %r927, %r45, %r25;

$L__BB227_3:
shr.s32 %r46, %r2, 31;
shr.u32 %r47, %r46, 29;
add.s32 %r48, %r2, %r47;
shr.s32 %r930, %r48, 3;
setp.ge.s32 %p2, %r930, %r4;
mov.f32 %f105, 0f00000000;
mov.f32 %f106, %f105;
mov.f32 %f107, %f105;
mov.f32 %f108, %f105;
@%p2 bra $L__BB227_6;

mov.u32 %r49, %ctaid.z;
and.b32 %r53, %r48, -8;
sub.s32 %r54, %r2, %r53;
shl.b32 %r55, %r54, 1;
cvt.s64.s32 %rd1, %r54;
shl.b32 %r929, %r930, 3;
mad.lo.s32 %r56, %r3, %r22, %r930;
mul.hi.u32 %r57, %r49, %r30;
add.s32 %r58, %r57, %r49;
shr.u32 %r59, %r58, %r31;
mad.lo.s32 %r60, %r59, %r33, %r56;
mad.lo.s32 %r61, %r927, %r27, %r60;
mul.wide.s32 %rd2, %r61, 82;
mul.wide.s32 %rd19, %r55, 2;
add.s64 %rd20, %rd2, %rd19;
add.s64 %rd3, %rd20, 36;
add.s64 %rd21, %rd2, %rd1;
add.s64 %rd4, %rd21, 74;
mul.wide.s32 %rd5, %r22, 82;
cvta.to.global.u64 %rd6, %rd13;
cvta.to.global.u64 %rd78, %rd12;
mul.lo.s32 %r62, %r928, %r28;
cvt.u64.u32 %rd22, %r62;
mul.lo.s32 %r63, %r49, %r34;
cvt.u64.u32 %rd23, %r63;
add.s64 %rd8, %rd22, %rd23;
mov.f32 %f105, 0f00000000;

$L__BB227_5:
cvt.u64.u32 %rd24, %r929;
add.s64 %rd25, %rd8, %rd24;
add.s64 %rd26, %rd78, %rd3;
add.s64 %rd27, %rd26, -34;
ld.global.nc.u16 %rs5, [%rd26+-34];
shr.u16 %rs6, %rs5, 8;
cvt.u32.u16 %r260, %rs6;
ld.global.nc.u16 %rs7, [%rd26+-32];
shr.u16 %rs8, %rs7, 8;
cvt.u32.u16 %r261, %rs8;
add.s64 %rd28, %rd78, %rd4;
add.s64 %rd29, %rd28, -8;
ld.global.nc.u8 %rs9, [%rd28+-8];
ld.global.nc.u16 %rs10, [%rd26+-2];
ld.global.nc.u16 %rs11, [%rd26];
ld.global.nc.u8 %rs12, [%rd28];
and.b16 %rs13, %rs12, 240;
and.b16 %rs14, %rs12, 15;
shr.u16 %rs15, %rs13, 4;
add.s64 %rd30, %rd25, %rd1;
and.b16 %rs16, %rs5, 255;
cvt.u32.u16 %r262, %rs16;
shl.b16 %rs17, %rs9, 8;
cvt.u32.u16 %r263, %rs17;
and.b32 %r264, %r263, 768;
or.b32 %r265, %r264, %r262;
mul.wide.u32 %rd31, %r265, 8;
mov.u64 %rd32, iq2s_grid;
add.s64 %rd33, %rd32, %rd31;
and.b16 %rs18, %rs10, 255;
mul.wide.u16 %r266, %rs18, 128;
cvt.u32.u16 %r267, %rs10;
shl.b32 %r268, %r267, 21;
or.b32 %r269, %r266, %r268;
and.b32 %r270, %r269, 25166208;
or.b32 %r271, %r270, -2139062144;
add.s32 %r272, %r271, -16843009;
or.b32 %r64, %r272, %r270;

prmt.b32 %r64, %r64, 0, 0xba98;

mul.wide.u16 %r273, %rs18, 8;
shl.b32 %r274, %r267, 17;
or.b32 %r275, %r273, %r274;
and.b32 %r276, %r275, 25166208;
or.b32 %r277, %r276, -2139062144;
add.s32 %r278, %r277, -16843009;
or.b32 %r66, %r278, %r276;

prmt.b32 %r66, %r66, 0, 0xba98;

ld.global.nc.v2.u32 {%r279, %r280}, [%rd33];
mov.u32 %r242, 0;
xor.b32 %r283, %r279, %r64;
or.b32 %r284, %r283, -2139062144;
and.b32 %r285, %r64, 2139062143;
and.b32 %r286, %r279, -2139062144;
sub.s32 %r287, %r284, %r285;
xor.b32 %r288, %r286, %r287;
xor.b32 %r69, %r288, -2139062144;
xor.b32 %r289, %r280, %r66;
or.b32 %r290, %r289, -2139062144;
and.b32 %r291, %r66, 2139062143;
and.b32 %r292, %r280, -2139062144;
sub.s32 %r293, %r290, %r291;
xor.b32 %r294, %r292, %r293;
xor.b32 %r73, %r294, -2139062144;
mul.lo.s64 %rd34, %rd30, 36;
add.s64 %rd35, %rd6, %rd34;
ld.global.nc.u32 %r74, [%rd35+8];
ld.global.nc.u32 %r70, [%rd35+4];

dp4a.s32.s32 %r68, %r69, %r70, %r242;


dp4a.s32.s32 %r72, %r73, %r74, %r68;

mul.wide.u16 %r295, %rs9, 64;
and.b32 %r296, %r295, 768;
or.b32 %r297, %r296, %r260;
mul.wide.u32 %rd36, %r297, 8;
add.s64 %rd37, %rd32, %rd36;
shr.u16 %rs20, %rs10, 8;
mul.wide.u16 %r298, %rs20, 128;
cvt.u32.u16 %r299, %rs20;
shl.b32 %r300, %r299, 21;
or.b32 %r301, %r298, %r300;
and.b32 %r302, %r301, 25166208;
or.b32 %r303, %r302, -2139062144;
add.s32 %r304, %r303, -16843009;
or.b32 %r76, %r304, %r302;

prmt.b32 %r76, %r76, 0, 0xba98;

mul.wide.u16 %r305, %rs20, 8;
shl.b32 %r306, %r299, 17;
or.b32 %r307, %r305, %r306;
and.b32 %r308, %r307, 25166208;
or.b32 %r309, %r308, -2139062144;
add.s32 %r310, %r309, -16843009;
or.b32 %r78, %r310, %r308;

prmt.b32 %r78, %r78, 0, 0xba98;

ld.global.nc.v2.u32 {%r311, %r312}, [%rd37];
xor.b32 %r315, %r311, %r76;
or.b32 %r316, %r315, -2139062144;
and.b32 %r317, %r76, 2139062143;
and.b32 %r318, %r311, -2139062144;
sub.s32 %r319, %r316, %r317;
xor.b32 %r320, %r318, %r319;
xor.b32 %r81, %r320, -2139062144;
xor.b32 %r321, %r312, %r78;
or.b32 %r322, %r321, -2139062144;
and.b32 %r323, %r78, 2139062143;
and.b32 %r324, %r312, -2139062144;
sub.s32 %r325, %r322, %r323;
xor.b32 %r326, %r324, %r325;
xor.b32 %r85, %r326, -2139062144;
ld.global.nc.u32 %r86, [%rd35+16];
ld.global.nc.u32 %r82, [%rd35+12];

dp4a.s32.s32 %r80, %r81, %r82, %r72;


dp4a.s32.s32 %r84, %r85, %r86, %r80;

and.b16 %rs21, %rs7, 255;
cvt.u32.u16 %r327, %rs21;
mul.wide.u16 %r328, %rs9, 16;
and.b32 %r329, %r328, 768;
or.b32 %r330, %r329, %r327;
mul.wide.u32 %rd38, %r330, 8;
add.s64 %rd39, %rd32, %rd38;
and.b16 %rs22, %rs11, 255;
mul.wide.u16 %r331, %rs22, 128;
cvt.u32.u16 %r332, %rs11;
shl.b32 %r333, %r332, 21;
or.b32 %r334, %r331, %r333;
and.b32 %r335, %r334, 25166208;
or.b32 %r336, %r335, -2139062144;
add.s32 %r337, %r336, -16843009;
or.b32 %r88, %r337, %r335;

prmt.b32 %r88, %r88, 0, 0xba98;

mul.wide.u16 %r338, %rs22, 8;
shl.b32 %r339, %r332, 17;
or.b32 %r340, %r338, %r339;
and.b32 %r341, %r340, 25166208;
or.b32 %r342, %r341, -2139062144;
add.s32 %r343, %r342, -16843009;
or.b32 %r90, %r343, %r341;

prmt.b32 %r90, %r90, 0, 0xba98;

ld.global.nc.v2.u32 {%r344, %r345}, [%rd39];
xor.b32 %r348, %r344, %r88;
or.b32 %r349, %r348, -2139062144;
and.b32 %r350, %r88, 2139062143;
and.b32 %r351, %r344, -2139062144;
sub.s32 %r352, %r349, %r350;
xor.b32 %r353, %r351, %r352;
xor.b32 %r93, %r353, -2139062144;
xor.b32 %r354, %r345, %r90;
or.b32 %r355, %r354, -2139062144;
and.b32 %r356, %r90, 2139062143;
and.b32 %r357, %r345, -2139062144;
sub.s32 %r358, %r355, %r356;
xor.b32 %r359, %r357, %r358;
xor.b32 %r97, %r359, -2139062144;
ld.global.nc.u32 %r98, [%rd35+24];
ld.global.nc.u32 %r94, [%rd35+20];

dp4a.s32.s32 %r92, %r93, %r94, %r242;


dp4a.s32.s32 %r96, %r97, %r98, %r92;

mul.wide.u16 %r360, %rs9, 4;
and.b32 %r361, %r360, 768;
or.b32 %r362, %r361, %r261;
mul.wide.u32 %rd40, %r362, 8;
add.s64 %rd41, %rd32, %rd40;
shr.u16 %rs23, %rs11, 8;
mul.wide.u16 %r363, %rs23, 128;
cvt.u32.u16 %r364, %rs23;
shl.b32 %r365, %r364, 21;
or.b32 %r366, %r363, %r365;
and.b32 %r367, %r366, 25166208;
or.b32 %r368, %r367, -2139062144;
add.s32 %r369, %r368, -16843009;
or.b32 %r100, %r369, %r367;

prmt.b32 %r100, %r100, 0, 0xba98;

mul.wide.u16 %r370, %rs23, 8;
shl.b32 %r371, %r364, 17;
or.b32 %r372, %r370, %r371;
and.b32 %r373, %r372, 25166208;
or.b32 %r374, %r373, -2139062144;
add.s32 %r375, %r374, -16843009;
or.b32 %r102, %r375, %r373;

prmt.b32 %r102, %r102, 0, 0xba98;

ld.global.nc.v2.u32 {%r376, %r377}, [%rd41];
xor.b32 %r380, %r376, %r100;
or.b32 %r381, %r380, -2139062144;
and.b32 %r382, %r100, 2139062143;
and.b32 %r383, %r376, -2139062144;
sub.s32 %r384, %r381, %r382;
xor.b32 %r385, %r383, %r384;
xor.b32 %r105, %r385, -2139062144;
xor.b32 %r386, %r377, %r102;
or.b32 %r387, %r386, -2139062144;
and.b32 %r388, %r102, 2139062143;
and.b32 %r389, %r377, -2139062144;
sub.s32 %r390, %r387, %r388;
xor.b32 %r391, %r389, %r390;
xor.b32 %r109, %r391, -2139062144;
ld.global.nc.u32 %r110, [%rd35+32];
ld.global.nc.u32 %r106, [%rd35+28];

dp4a.s32.s32 %r104, %r105, %r106, %r96;


dp4a.s32.s32 %r108, %r109, %r110, %r104;

cvt.u32.u16 %r392, %rs14;
cvt.u32.u16 %r393, %rs15;
mul.lo.s32 %r394, %r84, %r392;
mad.lo.s32 %r395, %r108, %r393, %r394;
add.s32 %r396, %r108, %r84;
shr.u32 %r397, %r396, 31;
add.s32 %r398, %r396, %r397;
shr.s32 %r399, %r398, 1;
add.s32 %r400, %r395, %r399;
shr.s32 %r401, %r400, 31;
shr.u32 %r402, %r401, 30;
add.s32 %r403, %r400, %r402;
shr.s32 %r404, %r403, 2;
add.s64 %rd42, %rd78, %rd2;
ld.global.nc.u16 %rs1, [%rd42];

{ cvt.f32.f16 %f25, %rs1;}


ld.global.nc.u32 %r112, [%rd35];

{.reg .f16 low,high;
mov.b32 {low,high},%r112;
cvt.f32.f16 %f26, low;}


mul.ftz.f32 %f33, %f25, %f26;
cvt.rn.f32.s32 %f34, %r404;
fma.rn.ftz.f32 %f108, %f33, %f34, %f108;
add.s64 %rd43, %rd27, %rd5;
ld.global.nc.u16 %rs24, [%rd43];
shr.u16 %rs25, %rs24, 8;
cvt.u32.u16 %r405, %rs25;
ld.global.nc.u16 %rs26, [%rd43+2];
shr.u16 %rs27, %rs26, 8;
cvt.u32.u16 %r406, %rs27;
add.s64 %rd44, %rd29, %rd5;
ld.global.nc.u8 %rs28, [%rd44];
ld.global.nc.u16 %rs29, [%rd43+32];
ld.global.nc.u16 %rs30, [%rd43+34];
ld.global.nc.u8 %rs31, [%rd44+8];
and.b16 %rs32, %rs31, 240;
and.b16 %rs33, %rs31, 15;
shr.u16 %rs34, %rs32, 4;
and.b16 %rs35, %rs24, 255;
cvt.u32.u16 %r407, %rs35;
shl.b16 %rs36, %rs28, 8;
cvt.u32.u16 %r408, %rs36;
and.b32 %r409, %r408, 768;
or.b32 %r410, %r409, %r407;
mul.wide.u32 %rd45, %r410, 8;
add.s64 %rd46, %rd32, %rd45;
and.b16 %rs37, %rs29, 255;
mul.wide.u16 %r411, %rs37, 128;
cvt.u32.u16 %r412, %rs29;
shl.b32 %r413, %r412, 21;
or.b32 %r414, %r411, %r413;
and.b32 %r415, %r414, 25166208;
or.b32 %r416, %r415, -2139062144;
add.s32 %r417, %r416, -16843009;
or.b32 %r113, %r417, %r415;

prmt.b32 %r113, %r113, 0, 0xba98;

mul.wide.u16 %r418, %rs37, 8;
shl.b32 %r419, %r412, 17;
or.b32 %r420, %r418, %r419;
and.b32 %r421, %r420, 25166208;
or.b32 %r422, %r421, -2139062144;
add.s32 %r423, %r422, -16843009;
or.b32 %r115, %r423, %r421;

prmt.b32 %r115, %r115, 0, 0xba98;

ld.global.nc.v2.u32 {%r424, %r425}, [%rd46];
xor.b32 %r428, %r424, %r113;
or.b32 %r429, %r428, -2139062144;
and.b32 %r430, %r113, 2139062143;
and.b32 %r431, %r424, -2139062144;
sub.s32 %r432, %r429, %r430;
xor.b32 %r433, %r431, %r432;
xor.b32 %r118, %r433, -2139062144;
xor.b32 %r434, %r425, %r115;
or.b32 %r435, %r434, -2139062144;
and.b32 %r436, %r115, 2139062143;
and.b32 %r437, %r425, -2139062144;
sub.s32 %r438, %r435, %r436;
xor.b32 %r439, %r437, %r438;
xor.b32 %r122, %r439, -2139062144;

dp4a.s32.s32 %r117, %r118, %r70, %r242;


dp4a.s32.s32 %r121, %r122, %r74, %r117;

mul.wide.u16 %r440, %rs28, 64;
and.b32 %r441, %r440, 768;
or.b32 %r442, %r441, %r405;
mul.wide.u32 %rd47, %r442, 8;
add.s64 %rd48, %rd32, %rd47;
shr.u16 %rs39, %rs29, 8;
mul.wide.u16 %r443, %rs39, 128;
cvt.u32.u16 %r444, %rs39;
shl.b32 %r445, %r444, 21;
or.b32 %r446, %r443, %r445;
and.b32 %r447, %r446, 25166208;
or.b32 %r448, %r447, -2139062144;
add.s32 %r449, %r448, -16843009;
or.b32 %r125, %r449, %r447;

prmt.b32 %r125, %r125, 0, 0xba98;

mul.wide.u16 %r450, %rs39, 8;
shl.b32 %r451, %r444, 17;
or.b32 %r452, %r450, %r451;
and.b32 %r453, %r452, 25166208;
or.b32 %r454, %r453, -2139062144;
add.s32 %r455, %r454, -16843009;
or.b32 %r127, %r455, %r453;

prmt.b32 %r127, %r127, 0, 0xba98;

ld.global.nc.v2.u32 {%r456, %r457}, [%rd48];
xor.b32 %r460, %r456, %r125;
or.b32 %r461, %r460, -2139062144;
and.b32 %r462, %r125, 2139062143;
and.b32 %r463, %r456, -2139062144;
sub.s32 %r464, %r461, %r462;
xor.b32 %r465, %r463, %r464;
xor.b32 %r130, %r465, -2139062144;
xor.b32 %r466, %r457, %r127;
or.b32 %r467, %r466, -2139062144;
and.b32 %r468, %r127, 2139062143;
and.b32 %r469, %r457, -2139062144;
sub.s32 %r470, %r467, %r468;
xor.b32 %r471, %r469, %r470;
xor.b32 %r134, %r471, -2139062144;

dp4a.s32.s32 %r129, %r130, %r82, %r121;


dp4a.s32.s32 %r133, %r134, %r86, %r129;

and.b16 %rs40, %rs26, 255;
cvt.u32.u16 %r472, %rs40;
mul.wide.u16 %r473, %rs28, 16;
and.b32 %r474, %r473, 768;
or.b32 %r475, %r474, %r472;
mul.wide.u32 %rd49, %r475, 8;
add.s64 %rd50, %rd32, %rd49;
and.b16 %rs41, %rs30, 255;
mul.wide.u16 %r476, %rs41, 128;
cvt.u32.u16 %r477, %rs30;
shl.b32 %r478, %r477, 21;
or.b32 %r479, %r476, %r478;
and.b32 %r480, %r479, 25166208;
or.b32 %r481, %r480, -2139062144;
add.s32 %r482, %r481, -16843009;
or.b32 %r137, %r482, %r480;

prmt.b32 %r137, %r137, 0, 0xba98;

mul.wide.u16 %r483, %rs41, 8;
shl.b32 %r484, %r477, 17;
or.b32 %r485, %r483, %r484;
and.b32 %r486, %r485, 25166208;
or.b32 %r487, %r486, -2139062144;
add.s32 %r488, %r487, -16843009;
or.b32 %r139, %r488, %r486;

prmt.b32 %r139, %r139, 0, 0xba98;

ld.global.nc.v2.u32 {%r489, %r490}, [%rd50];
xor.b32 %r493, %r489, %r137;
or.b32 %r494, %r493, -2139062144;
and.b32 %r495, %r137, 2139062143;
and.b32 %r496, %r489, -2139062144;
sub.s32 %r497, %r494, %r495;
xor.b32 %r498, %r496, %r497;
xor.b32 %r142, %r498, -2139062144;
xor.b32 %r499, %r490, %r139;
or.b32 %r500, %r499, -2139062144;
and.b32 %r501, %r139, 2139062143;
and.b32 %r502, %r490, -2139062144;
sub.s32 %r503, %r500, %r501;
xor.b32 %r504, %r502, %r503;
xor.b32 %r146, %r504, -2139062144;

dp4a.s32.s32 %r141, %r142, %r94, %r242;


dp4a.s32.s32 %r145, %r146, %r98, %r141;

mul.wide.u16 %r505, %rs28, 4;
and.b32 %r506, %r505, 768;
or.b32 %r507, %r506, %r406;
mul.wide.u32 %rd51, %r507, 8;
add.s64 %rd52, %rd32, %rd51;
shr.u16 %rs42, %rs30, 8;
mul.wide.u16 %r508, %rs42, 128;
cvt.u32.u16 %r509, %rs42;
shl.b32 %r510, %r509, 21;
or.b32 %r511, %r508, %r510;
and.b32 %r512, %r511, 25166208;
or.b32 %r513, %r512, -2139062144;
add.s32 %r514, %r513, -16843009;
or.b32 %r149, %r514, %r512;

prmt.b32 %r149, %r149, 0, 0xba98;

mul.wide.u16 %r515, %rs42, 8;
shl.b32 %r516, %r509, 17;
or.b32 %r517, %r515, %r516;
and.b32 %r518, %r517, 25166208;
or.b32 %r519, %r518, -2139062144;
add.s32 %r520, %r519, -16843009;
or.b32 %r151, %r520, %r518;

prmt.b32 %r151, %r151, 0, 0xba98;

ld.global.nc.v2.u32 {%r521, %r522}, [%rd52];
xor.b32 %r525, %r521, %r149;
or.b32 %r526, %r525, -2139062144;
and.b32 %r527, %r149, 2139062143;
and.b32 %r528, %r521, -2139062144;
sub.s32 %r529, %r526, %r527;
xor.b32 %r530, %r528, %r529;
xor.b32 %r154, %r530, -2139062144;
xor.b32 %r531, %r522, %r151;
or.b32 %r532, %r531, -2139062144;
and.b32 %r533, %r151, 2139062143;
and.b32 %r534, %r522, -2139062144;
sub.s32 %r535, %r532, %r533;
xor.b32 %r536, %r534, %r535;
xor.b32 %r158, %r536, -2139062144;

dp4a.s32.s32 %r153, %r154, %r106, %r145;


dp4a.s32.s32 %r157, %r158, %r110, %r153;

cvt.u32.u16 %r537, %rs33;
cvt.u32.u16 %r538, %rs34;
mul.lo.s32 %r539, %r133, %r537;
mad.lo.s32 %r540, %r157, %r538, %r539;
add.s32 %r541, %r157, %r133;
shr.u32 %r542, %r541, 31;
add.s32 %r543, %r541, %r542;
shr.s32 %r544, %r543, 1;
add.s32 %r545, %r540, %r544;
shr.s32 %r546, %r545, 31;
shr.u32 %r547, %r546, 30;
add.s32 %r548, %r545, %r547;
shr.s32 %r549, %r548, 2;
add.s64 %rd53, %rd42, %rd5;
ld.global.nc.u16 %rs2, [%rd53];

{ cvt.f32.f16 %f27, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r112;
cvt.f32.f16 %f28, low;}


mul.ftz.f32 %f35, %f27, %f28;
cvt.rn.f32.s32 %f36, %r549;
fma.rn.ftz.f32 %f107, %f35, %f36, %f107;
add.s64 %rd54, %rd43, %rd5;
ld.global.nc.u16 %rs43, [%rd54];
shr.u16 %rs44, %rs43, 8;
cvt.u32.u16 %r550, %rs44;
ld.global.nc.u16 %rs45, [%rd54+2];
shr.u16 %rs46, %rs45, 8;
cvt.u32.u16 %r551, %rs46;
add.s64 %rd55, %rd44, %rd5;
ld.global.nc.u8 %rs47, [%rd55];
ld.global.nc.u16 %rs48, [%rd54+32];
ld.global.nc.u16 %rs49, [%rd54+34];
ld.global.nc.u8 %rs50, [%rd55+8];
and.b16 %rs51, %rs50, 240;
and.b16 %rs52, %rs50, 15;
shr.u16 %rs53, %rs51, 4;
and.b16 %rs54, %rs43, 255;
cvt.u32.u16 %r552, %rs54;
shl.b16 %rs55, %rs47, 8;
cvt.u32.u16 %r553, %rs55;
and.b32 %r554, %r553, 768;
or.b32 %r555, %r554, %r552;
mul.wide.u32 %rd56, %r555, 8;
add.s64 %rd57, %rd32, %rd56;
and.b16 %rs56, %rs48, 255;
mul.wide.u16 %r556, %rs56, 128;
cvt.u32.u16 %r557, %rs48;
shl.b32 %r558, %r557, 21;
or.b32 %r559, %r556, %r558;
and.b32 %r560, %r559, 25166208;
or.b32 %r561, %r560, -2139062144;
add.s32 %r562, %r561, -16843009;
or.b32 %r162, %r562, %r560;

prmt.b32 %r162, %r162, 0, 0xba98;

mul.wide.u16 %r563, %rs56, 8;
shl.b32 %r564, %r557, 17;
or.b32 %r565, %r563, %r564;
and.b32 %r566, %r565, 25166208;
or.b32 %r567, %r566, -2139062144;
add.s32 %r568, %r567, -16843009;
or.b32 %r164, %r568, %r566;

prmt.b32 %r164, %r164, 0, 0xba98;

ld.global.nc.v2.u32 {%r569, %r570}, [%rd57];
xor.b32 %r573, %r569, %r162;
or.b32 %r574, %r573, -2139062144;
and.b32 %r575, %r162, 2139062143;
and.b32 %r576, %r569, -2139062144;
sub.s32 %r577, %r574, %r575;
xor.b32 %r578, %r576, %r577;
xor.b32 %r167, %r578, -2139062144;
xor.b32 %r579, %r570, %r164;
or.b32 %r580, %r579, -2139062144;
and.b32 %r581, %r164, 2139062143;
and.b32 %r582, %r570, -2139062144;
sub.s32 %r583, %r580, %r581;
xor.b32 %r584, %r582, %r583;
xor.b32 %r171, %r584, -2139062144;

dp4a.s32.s32 %r166, %r167, %r70, %r242;


dp4a.s32.s32 %r170, %r171, %r74, %r166;

mul.wide.u16 %r585, %rs47, 64;
and.b32 %r586, %r585, 768;
or.b32 %r587, %r586, %r550;
mul.wide.u32 %rd58, %r587, 8;
add.s64 %rd59, %rd32, %rd58;
shr.u16 %rs58, %rs48, 8;
mul.wide.u16 %r588, %rs58, 128;
cvt.u32.u16 %r589, %rs58;
shl.b32 %r590, %r589, 21;
or.b32 %r591, %r588, %r590;
and.b32 %r592, %r591, 25166208;
or.b32 %r593, %r592, -2139062144;
add.s32 %r594, %r593, -16843009;
or.b32 %r174, %r594, %r592;

prmt.b32 %r174, %r174, 0, 0xba98;

mul.wide.u16 %r595, %rs58, 8;
shl.b32 %r596, %r589, 17;
or.b32 %r597, %r595, %r596;
and.b32 %r598, %r597, 25166208;
or.b32 %r599, %r598, -2139062144;
add.s32 %r600, %r599, -16843009;
or.b32 %r176, %r600, %r598;

prmt.b32 %r176, %r176, 0, 0xba98;

ld.global.nc.v2.u32 {%r601, %r602}, [%rd59];
xor.b32 %r605, %r601, %r174;
or.b32 %r606, %r605, -2139062144;
and.b32 %r607, %r174, 2139062143;
and.b32 %r608, %r601, -2139062144;
sub.s32 %r609, %r606, %r607;
xor.b32 %r610, %r608, %r609;
xor.b32 %r179, %r610, -2139062144;
xor.b32 %r611, %r602, %r176;
or.b32 %r612, %r611, -2139062144;
and.b32 %r613, %r176, 2139062143;
and.b32 %r614, %r602, -2139062144;
sub.s32 %r615, %r612, %r613;
xor.b32 %r616, %r614, %r615;
xor.b32 %r183, %r616, -2139062144;

dp4a.s32.s32 %r178, %r179, %r82, %r170;


dp4a.s32.s32 %r182, %r183, %r86, %r178;

and.b16 %rs59, %rs45, 255;
cvt.u32.u16 %r617, %rs59;
mul.wide.u16 %r618, %rs47, 16;
and.b32 %r619, %r618, 768;
or.b32 %r620, %r619, %r617;
mul.wide.u32 %rd60, %r620, 8;
add.s64 %rd61, %rd32, %rd60;
and.b16 %rs60, %rs49, 255;
mul.wide.u16 %r621, %rs60, 128;
cvt.u32.u16 %r622, %rs49;
shl.b32 %r623, %r622, 21;
or.b32 %r624, %r621, %r623;
and.b32 %r625, %r624, 25166208;
or.b32 %r626, %r625, -2139062144;
add.s32 %r627, %r626, -16843009;
or.b32 %r186, %r627, %r625;

prmt.b32 %r186, %r186, 0, 0xba98;

mul.wide.u16 %r628, %rs60, 8;
shl.b32 %r629, %r622, 17;
or.b32 %r630, %r628, %r629;
and.b32 %r631, %r630, 25166208;
or.b32 %r632, %r631, -2139062144;
add.s32 %r633, %r632, -16843009;
or.b32 %r188, %r633, %r631;

prmt.b32 %r188, %r188, 0, 0xba98;

ld.global.nc.v2.u32 {%r634, %r635}, [%rd61];
xor.b32 %r638, %r634, %r186;
or.b32 %r639, %r638, -2139062144;
and.b32 %r640, %r186, 2139062143;
and.b32 %r641, %r634, -2139062144;
sub.s32 %r642, %r639, %r640;
xor.b32 %r643, %r641, %r642;
xor.b32 %r191, %r643, -2139062144;
xor.b32 %r644, %r635, %r188;
or.b32 %r645, %r644, -2139062144;
and.b32 %r646, %r188, 2139062143;
and.b32 %r647, %r635, -2139062144;
sub.s32 %r648, %r645, %r646;
xor.b32 %r649, %r647, %r648;
xor.b32 %r195, %r649, -2139062144;

dp4a.s32.s32 %r190, %r191, %r94, %r242;


dp4a.s32.s32 %r194, %r195, %r98, %r190;

mul.wide.u16 %r650, %rs47, 4;
and.b32 %r651, %r650, 768;
or.b32 %r652, %r651, %r551;
mul.wide.u32 %rd62, %r652, 8;
add.s64 %rd63, %rd32, %rd62;
shr.u16 %rs61, %rs49, 8;
mul.wide.u16 %r653, %rs61, 128;
cvt.u32.u16 %r654, %rs61;
shl.b32 %r655, %r654, 21;
or.b32 %r656, %r653, %r655;
and.b32 %r657, %r656, 25166208;
or.b32 %r658, %r657, -2139062144;
add.s32 %r659, %r658, -16843009;
or.b32 %r198, %r659, %r657;

prmt.b32 %r198, %r198, 0, 0xba98;

mul.wide.u16 %r660, %rs61, 8;
shl.b32 %r661, %r654, 17;
or.b32 %r662, %r660, %r661;
and.b32 %r663, %r662, 25166208;
or.b32 %r664, %r663, -2139062144;
add.s32 %r665, %r664, -16843009;
or.b32 %r200, %r665, %r663;

prmt.b32 %r200, %r200, 0, 0xba98;

ld.global.nc.v2.u32 {%r666, %r667}, [%rd63];
xor.b32 %r670, %r666, %r198;
or.b32 %r671, %r670, -2139062144;
and.b32 %r672, %r198, 2139062143;
and.b32 %r673, %r666, -2139062144;
sub.s32 %r674, %r671, %r672;
xor.b32 %r675, %r673, %r674;
xor.b32 %r203, %r675, -2139062144;
xor.b32 %r676, %r667, %r200;
or.b32 %r677, %r676, -2139062144;
and.b32 %r678, %r200, 2139062143;
and.b32 %r679, %r667, -2139062144;
sub.s32 %r680, %r677, %r678;
xor.b32 %r681, %r679, %r680;
xor.b32 %r207, %r681, -2139062144;

dp4a.s32.s32 %r202, %r203, %r106, %r194;


dp4a.s32.s32 %r206, %r207, %r110, %r202;

cvt.u32.u16 %r682, %rs52;
cvt.u32.u16 %r683, %rs53;
mul.lo.s32 %r684, %r182, %r682;
mad.lo.s32 %r685, %r206, %r683, %r684;
add.s32 %r686, %r206, %r182;
shr.u32 %r687, %r686, 31;
add.s32 %r688, %r686, %r687;
shr.s32 %r689, %r688, 1;
add.s32 %r690, %r685, %r689;
shr.s32 %r691, %r690, 31;
shr.u32 %r692, %r691, 30;
add.s32 %r693, %r690, %r692;
shr.s32 %r694, %r693, 2;
add.s64 %rd64, %rd53, %rd5;
ld.global.nc.u16 %rs3, [%rd64];

{ cvt.f32.f16 %f29, %rs3;}



{.reg .f16 low,high;
mov.b32 {low,high},%r112;
cvt.f32.f16 %f30, low;}


mul.ftz.f32 %f37, %f29, %f30;
cvt.rn.f32.s32 %f38, %r694;
fma.rn.ftz.f32 %f106, %f37, %f38, %f106;
add.s64 %rd65, %rd54, %rd5;
ld.global.nc.u16 %rs62, [%rd65];
shr.u16 %rs63, %rs62, 8;
cvt.u32.u16 %r695, %rs63;
ld.global.nc.u16 %rs64, [%rd65+2];
shr.u16 %rs65, %rs64, 8;
cvt.u32.u16 %r696, %rs65;
add.s64 %rd66, %rd55, %rd5;
ld.global.nc.u8 %rs66, [%rd66];
ld.global.nc.u16 %rs67, [%rd65+32];
ld.global.nc.u16 %rs68, [%rd65+34];
ld.global.nc.u8 %rs69, [%rd66+8];
and.b16 %rs70, %rs69, 240;
and.b16 %rs71, %rs69, 15;
shr.u16 %rs72, %rs70, 4;
and.b16 %rs73, %rs62, 255;
cvt.u32.u16 %r697, %rs73;
shl.b16 %rs74, %rs66, 8;
cvt.u32.u16 %r698, %rs74;
and.b32 %r699, %r698, 768;
or.b32 %r700, %r699, %r697;
mul.wide.u32 %rd67, %r700, 8;
add.s64 %rd68, %rd32, %rd67;
and.b16 %rs75, %rs67, 255;
mul.wide.u16 %r701, %rs75, 128;
cvt.u32.u16 %r702, %rs67;
shl.b32 %r703, %r702, 21;
or.b32 %r704, %r701, %r703;
and.b32 %r705, %r704, 25166208;
or.b32 %r706, %r705, -2139062144;
add.s32 %r707, %r706, -16843009;
or.b32 %r211, %r707, %r705;

prmt.b32 %r211, %r211, 0, 0xba98;

mul.wide.u16 %r708, %rs75, 8;
shl.b32 %r709, %r702, 17;
or.b32 %r710, %r708, %r709;
and.b32 %r711, %r710, 25166208;
or.b32 %r712, %r711, -2139062144;
add.s32 %r713, %r712, -16843009;
or.b32 %r213, %r713, %r711;

prmt.b32 %r213, %r213, 0, 0xba98;

ld.global.nc.v2.u32 {%r714, %r715}, [%rd68];
xor.b32 %r718, %r714, %r211;
or.b32 %r719, %r718, -2139062144;
and.b32 %r720, %r211, 2139062143;
and.b32 %r721, %r714, -2139062144;
sub.s32 %r722, %r719, %r720;
xor.b32 %r723, %r721, %r722;
xor.b32 %r216, %r723, -2139062144;
xor.b32 %r724, %r715, %r213;
or.b32 %r725, %r724, -2139062144;
and.b32 %r726, %r213, 2139062143;
and.b32 %r727, %r715, -2139062144;
sub.s32 %r728, %r725, %r726;
xor.b32 %r729, %r727, %r728;
xor.b32 %r220, %r729, -2139062144;

dp4a.s32.s32 %r215, %r216, %r70, %r242;


dp4a.s32.s32 %r219, %r220, %r74, %r215;

mul.wide.u16 %r730, %rs66, 64;
and.b32 %r731, %r730, 768;
or.b32 %r732, %r731, %r695;
mul.wide.u32 %rd69, %r732, 8;
add.s64 %rd70, %rd32, %rd69;
shr.u16 %rs77, %rs67, 8;
mul.wide.u16 %r733, %rs77, 128;
cvt.u32.u16 %r734, %rs77;
shl.b32 %r735, %r734, 21;
or.b32 %r736, %r733, %r735;
and.b32 %r737, %r736, 25166208;
or.b32 %r738, %r737, -2139062144;
add.s32 %r739, %r738, -16843009;
or.b32 %r223, %r739, %r737;

prmt.b32 %r223, %r223, 0, 0xba98;

mul.wide.u16 %r740, %rs77, 8;
shl.b32 %r741, %r734, 17;
or.b32 %r742, %r740, %r741;
and.b32 %r743, %r742, 25166208;
or.b32 %r744, %r743, -2139062144;
add.s32 %r745, %r744, -16843009;
or.b32 %r225, %r745, %r743;

prmt.b32 %r225, %r225, 0, 0xba98;

ld.global.nc.v2.u32 {%r746, %r747}, [%rd70];
xor.b32 %r750, %r746, %r223;
or.b32 %r751, %r750, -2139062144;
and.b32 %r752, %r223, 2139062143;
and.b32 %r753, %r746, -2139062144;
sub.s32 %r754, %r751, %r752;
xor.b32 %r755, %r753, %r754;
xor.b32 %r228, %r755, -2139062144;
xor.b32 %r756, %r747, %r225;
or.b32 %r757, %r756, -2139062144;
and.b32 %r758, %r225, 2139062143;
and.b32 %r759, %r747, -2139062144;
sub.s32 %r760, %r757, %r758;
xor.b32 %r761, %r759, %r760;
xor.b32 %r232, %r761, -2139062144;

dp4a.s32.s32 %r227, %r228, %r82, %r219;


dp4a.s32.s32 %r231, %r232, %r86, %r227;

and.b16 %rs78, %rs64, 255;
cvt.u32.u16 %r762, %rs78;
mul.wide.u16 %r763, %rs66, 16;
and.b32 %r764, %r763, 768;
or.b32 %r765, %r764, %r762;
mul.wide.u32 %rd71, %r765, 8;
add.s64 %rd72, %rd32, %rd71;
and.b16 %rs79, %rs68, 255;
mul.wide.u16 %r766, %rs79, 128;
cvt.u32.u16 %r767, %rs68;
shl.b32 %r768, %r767, 21;
or.b32 %r769, %r766, %r768;
and.b32 %r770, %r769, 25166208;
or.b32 %r771, %r770, -2139062144;
add.s32 %r772, %r771, -16843009;
or.b32 %r235, %r772, %r770;

prmt.b32 %r235, %r235, 0, 0xba98;

mul.wide.u16 %r773, %rs79, 8;
shl.b32 %r774, %r767, 17;
or.b32 %r775, %r773, %r774;
and.b32 %r776, %r775, 25166208;
or.b32 %r777, %r776, -2139062144;
add.s32 %r778, %r777, -16843009;
or.b32 %r237, %r778, %r776;

prmt.b32 %r237, %r237, 0, 0xba98;

ld.global.nc.v2.u32 {%r779, %r780}, [%rd72];
xor.b32 %r783, %r779, %r235;
or.b32 %r784, %r783, -2139062144;
and.b32 %r785, %r235, 2139062143;
and.b32 %r786, %r779, -2139062144;
sub.s32 %r787, %r784, %r785;
xor.b32 %r788, %r786, %r787;
xor.b32 %r240, %r788, -2139062144;
xor.b32 %r789, %r780, %r237;
or.b32 %r790, %r789, -2139062144;
and.b32 %r791, %r237, 2139062143;
and.b32 %r792, %r780, -2139062144;
sub.s32 %r793, %r790, %r791;
xor.b32 %r794, %r792, %r793;
xor.b32 %r244, %r794, -2139062144;

dp4a.s32.s32 %r239, %r240, %r94, %r242;


dp4a.s32.s32 %r243, %r244, %r98, %r239;

mul.wide.u16 %r795, %rs66, 4;
and.b32 %r796, %r795, 768;
or.b32 %r797, %r796, %r696;
mul.wide.u32 %rd73, %r797, 8;
add.s64 %rd74, %rd32, %rd73;
shr.u16 %rs80, %rs68, 8;
mul.wide.u16 %r798, %rs80, 128;
cvt.u32.u16 %r799, %rs80;
shl.b32 %r800, %r799, 21;
or.b32 %r801, %r798, %r800;
and.b32 %r802, %r801, 25166208;
or.b32 %r803, %r802, -2139062144;
add.s32 %r804, %r803, -16843009;
or.b32 %r247, %r804, %r802;

prmt.b32 %r247, %r247, 0, 0xba98;

mul.wide.u16 %r805, %rs80, 8;
shl.b32 %r806, %r799, 17;
or.b32 %r807, %r805, %r806;
and.b32 %r808, %r807, 25166208;
or.b32 %r809, %r808, -2139062144;
add.s32 %r810, %r809, -16843009;
or.b32 %r249, %r810, %r808;

prmt.b32 %r249, %r249, 0, 0xba98;

ld.global.nc.v2.u32 {%r811, %r812}, [%rd74];
xor.b32 %r815, %r811, %r247;
or.b32 %r816, %r815, -2139062144;
and.b32 %r817, %r247, 2139062143;
and.b32 %r818, %r811, -2139062144;
sub.s32 %r819, %r816, %r817;
xor.b32 %r820, %r818, %r819;
xor.b32 %r252, %r820, -2139062144;
xor.b32 %r821, %r812, %r249;
or.b32 %r822, %r821, -2139062144;
and.b32 %r823, %r249, 2139062143;
and.b32 %r824, %r812, -2139062144;
sub.s32 %r825, %r822, %r823;
xor.b32 %r826, %r824, %r825;
xor.b32 %r256, %r826, -2139062144;

dp4a.s32.s32 %r251, %r252, %r106, %r243;


dp4a.s32.s32 %r255, %r256, %r110, %r251;

cvt.u32.u16 %r827, %rs71;
cvt.u32.u16 %r828, %rs72;
mul.lo.s32 %r829, %r231, %r827;
mad.lo.s32 %r830, %r255, %r828, %r829;
add.s32 %r831, %r255, %r231;
shr.u32 %r832, %r831, 31;
add.s32 %r833, %r831, %r832;
shr.s32 %r834, %r833, 1;
add.s32 %r835, %r830, %r834;
shr.s32 %r836, %r835, 31;
shr.u32 %r837, %r836, 30;
add.s32 %r838, %r835, %r837;
shr.s32 %r839, %r838, 2;
add.s64 %rd75, %rd64, %rd5;
ld.global.nc.u16 %rs4, [%rd75];

{ cvt.f32.f16 %f31, %rs4;}



{.reg .f16 low,high;
mov.b32 {low,high},%r112;
cvt.f32.f16 %f32, low;}


mul.ftz.f32 %f39, %f31, %f32;
cvt.rn.f32.s32 %f40, %r839;
fma.rn.ftz.f32 %f105, %f39, %f40, %f105;
add.s32 %r929, %r929, 128;
add.s64 %rd78, %rd78, 1312;
add.s32 %r930, %r930, 16;
setp.lt.s32 %p3, %r930, %r4;
@%p3 bra $L__BB227_5;

$L__BB227_6:
setp.eq.s32 %p4, %r1, 0;
@%p4 bra $L__BB227_8;

shl.b32 %r841, %r1, 9;
mov.u32 %r842, _ZZ13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r843, %r842, %r841;
shl.b32 %r844, %r38, 2;
add.s32 %r845, %r843, %r844;
st.shared.f32 [%r845+-512], %f108;
st.shared.f32 [%r845+-384], %f107;
st.shared.f32 [%r845+-256], %f106;
st.shared.f32 [%r845+-128], %f105;

$L__BB227_8:
setp.ne.s32 %p5, %r1, 0;
bar.sync 0;
@%p5 bra $L__BB227_17;

mov.u32 %r846, %ctaid.z;
mov.u32 %r847, %ctaid.y;
mad.lo.s32 %r848, %r847, %r29, %r3;
mad.lo.s32 %r849, %r846, %r35, %r848;
shl.b32 %r851, %r38, 2;
mov.u32 %r852, _ZZ13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r17, %r852, %r851;
mov.u32 %r853, 2;
ld.shared.f32 %f41, [%r17];
add.ftz.f32 %f42, %f41, %f108;
ld.shared.f32 %f43, [%r17+512];
add.ftz.f32 %f44, %f43, %f42;
ld.shared.f32 %f45, [%r17+1024];
add.ftz.f32 %f46, %f45, %f44;
mov.b32 %r854, %f46;
mov.u32 %r855, 31;
mov.u32 %r856, 16;
mov.u32 %r857, -1;
shfl.sync.bfly.b32 %r858|%p6, %r854, %r856, %r855, %r857;
mov.b32 %f47, %r858;
add.ftz.f32 %f48, %f46, %f47;
mov.b32 %r859, %f48;
mov.u32 %r860, 8;
shfl.sync.bfly.b32 %r861|%p7, %r859, %r860, %r855, %r857;
mov.b32 %f49, %r861;
add.ftz.f32 %f50, %f48, %f49;
mov.b32 %r862, %f50;
mov.u32 %r863, 4;
shfl.sync.bfly.b32 %r864|%p8, %r862, %r863, %r855, %r857;
mov.b32 %f51, %r864;
add.ftz.f32 %f52, %f50, %f51;
mov.b32 %r865, %f52;
shfl.sync.bfly.b32 %r866|%p9, %r865, %r853, %r855, %r857;
mov.b32 %f53, %r866;
add.ftz.f32 %f54, %f52, %f53;
mov.b32 %r867, %f54;
mov.u32 %r868, 1;
shfl.sync.bfly.b32 %r869|%p10, %r867, %r868, %r855, %r857;
mov.b32 %f55, %r869;
add.ftz.f32 %f13, %f54, %f55;
setp.ne.s32 %p11, %r38, 0;
setp.ge.u32 %p12, %r3, %r23;
cvta.to.global.u64 %rd76, %rd15;
mul.wide.u32 %rd77, %r849, 4;
add.s64 %rd11, %rd76, %rd77;
or.pred %p13, %p11, %p12;
@%p13 bra $L__BB227_11;

st.global.f32 [%rd11], %f13;

$L__BB227_11:
ld.shared.f32 %f56, [%r17+128];
add.ftz.f32 %f57, %f56, %f107;
ld.shared.f32 %f58, [%r17+640];
add.ftz.f32 %f59, %f58, %f57;
ld.shared.f32 %f60, [%r17+1152];
add.ftz.f32 %f61, %f60, %f59;
mov.b32 %r870, %f61;
shfl.sync.bfly.b32 %r874|%p14, %r870, %r856, %r855, %r857;
mov.b32 %f62, %r874;
add.ftz.f32 %f63, %f61, %f62;
mov.b32 %r875, %f63;
shfl.sync.bfly.b32 %r877|%p15, %r875, %r860, %r855, %r857;
mov.b32 %f64, %r877;
add.ftz.f32 %f65, %f63, %f64;
mov.b32 %r878, %f65;
shfl.sync.bfly.b32 %r880|%p16, %r878, %r863, %r855, %r857;
mov.b32 %f66, %r880;
add.ftz.f32 %f67, %f65, %f66;
mov.b32 %r881, %f67;
shfl.sync.bfly.b32 %r883|%p17, %r881, %r853, %r855, %r857;
mov.b32 %f68, %r883;
add.ftz.f32 %f69, %f67, %f68;
mov.b32 %r884, %f69;
shfl.sync.bfly.b32 %r886|%p18, %r884, %r868, %r855, %r857;
mov.b32 %f70, %r886;
add.ftz.f32 %f14, %f69, %f70;
setp.ne.s32 %p19, %r38, 1;
add.s32 %r888, %r3, 1;
setp.ge.u32 %p20, %r888, %r23;
or.pred %p21, %p19, %p20;
@%p21 bra $L__BB227_13;

st.global.f32 [%rd11+4], %f14;

$L__BB227_13:
ld.shared.f32 %f71, [%r17+256];
add.ftz.f32 %f72, %f71, %f106;
ld.shared.f32 %f73, [%r17+768];
add.ftz.f32 %f74, %f73, %f72;
ld.shared.f32 %f75, [%r17+1280];
add.ftz.f32 %f76, %f75, %f74;
mov.b32 %r889, %f76;
mov.u32 %r890, 31;
mov.u32 %r891, 16;
mov.u32 %r892, -1;
shfl.sync.bfly.b32 %r893|%p22, %r889, %r891, %r890, %r892;
mov.b32 %f77, %r893;
add.ftz.f32 %f78, %f76, %f77;
mov.b32 %r894, %f78;
mov.u32 %r895, 8;
shfl.sync.bfly.b32 %r896|%p23, %r894, %r895, %r890, %r892;
mov.b32 %f79, %r896;
add.ftz.f32 %f80, %f78, %f79;
mov.b32 %r897, %f80;
mov.u32 %r898, 4;
shfl.sync.bfly.b32 %r899|%p24, %r897, %r898, %r890, %r892;
mov.b32 %f81, %r899;
add.ftz.f32 %f82, %f80, %f81;
mov.b32 %r900, %f82;
mov.u32 %r901, 2;
shfl.sync.bfly.b32 %r902|%p25, %r900, %r901, %r890, %r892;
mov.b32 %f83, %r902;
add.ftz.f32 %f84, %f82, %f83;
mov.b32 %r903, %f84;
mov.u32 %r904, 1;
shfl.sync.bfly.b32 %r905|%p26, %r903, %r904, %r890, %r892;
mov.b32 %f85, %r905;
add.ftz.f32 %f15, %f84, %f85;
setp.ne.s32 %p27, %r38, 2;
add.s32 %r907, %r3, 2;
setp.ge.u32 %p28, %r907, %r23;
or.pred %p29, %p27, %p28;
@%p29 bra $L__BB227_15;

st.global.f32 [%rd11+8], %f15;

$L__BB227_15:
ld.shared.f32 %f86, [%r17+384];
add.ftz.f32 %f87, %f86, %f105;
ld.shared.f32 %f88, [%r17+896];
add.ftz.f32 %f89, %f88, %f87;
ld.shared.f32 %f90, [%r17+1408];
add.ftz.f32 %f91, %f90, %f89;
mov.b32 %r908, %f91;
shfl.sync.bfly.b32 %r912|%p30, %r908, %r891, %r890, %r892;
mov.b32 %f92, %r912;
add.ftz.f32 %f93, %f91, %f92;
mov.b32 %r913, %f93;
shfl.sync.bfly.b32 %r915|%p31, %r913, %r895, %r890, %r892;
mov.b32 %f94, %r915;
add.ftz.f32 %f95, %f93, %f94;
mov.b32 %r916, %f95;
shfl.sync.bfly.b32 %r918|%p32, %r916, %r898, %r890, %r892;
mov.b32 %f96, %r918;
add.ftz.f32 %f97, %f95, %f96;
mov.b32 %r919, %f97;
shfl.sync.bfly.b32 %r921|%p33, %r919, %r901, %r890, %r892;
mov.b32 %f98, %r921;
add.ftz.f32 %f99, %f97, %f98;
mov.b32 %r922, %f99;
shfl.sync.bfly.b32 %r924|%p34, %r922, %r904, %r890, %r892;
mov.b32 %f100, %r924;
add.ftz.f32 %f16, %f99, %f100;
setp.ne.s32 %p35, %r38, 3;
add.s32 %r926, %r3, 3;
setp.ge.u32 %p36, %r926, %r23;
or.pred %p37, %p35, %p36;
@%p37 bra $L__BB227_17;

st.global.f32 [%rd11+12], %f16;

$L__BB227_17:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<33>;
.reg .b16 %rs<39>;
.reg .f32 %f<121>;
.reg .b32 %r<548>;
.reg .b64 %rd<87>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[384];

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate[384];

ld.param.u64 %rd8, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd9, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd10, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r46, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r47, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r48, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
ld.param.u64 %rd13, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+16];
ld.param.u64 %rd12, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+8];
ld.param.u64 %rd11, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3];
mov.u32 %r1, %tid.y;
shl.b32 %r49, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r49, %r2;
mov.u32 %r4, %ctaid.x;
mov.u32 %r5, %ctaid.y;
setp.eq.s64 %p2, %rd10, 0;
@%p2 bra $L__BB228_2;

cvta.to.global.u64 %rd17, %rd10;
mul.wide.u32 %rd18, %r5, 4;
add.s64 %rd19, %rd17, %rd18;
ld.global.nc.u32 %r544, [%rd19];
mul.hi.u32 %r50, %r5, %r33;
add.s32 %r51, %r50, %r5;
shr.u32 %r52, %r51, %r34;
mul.lo.s32 %r53, %r52, %r35;
sub.s32 %r545, %r5, %r53;
bra.uni $L__BB228_3;

$L__BB228_2:
mul.hi.u32 %r54, %r5, %r37;
add.s32 %r55, %r54, %r5;
shr.u32 %r544, %r55, %r38;
mov.u32 %r545, %r5;

$L__BB228_3:
selp.b32 %r11, %r5, %r544, %p2;
or.b32 %r56, %r2, %r1;
setp.ne.s32 %p4, %r56, 0;
mov.f32 %f114, 0f00000000;
mov.f32 %f110, %f114;
mov.f32 %f111, %f114;
@%p4 bra $L__BB228_8;

setp.eq.s64 %p5, %rd11, 0;
mov.f32 %f110, 0f00000000;
mov.f32 %f111, %f110;
@%p5 bra $L__BB228_6;

cvta.to.global.u64 %rd20, %rd11;
mov.u32 %r57, %ctaid.z;
mul.lo.s32 %r58, %r57, %r48;
cvt.u64.u32 %rd21, %r58;
mul.lo.s32 %r59, %r11, %r42;
cvt.u64.u32 %rd22, %r59;
cvt.u64.u32 %rd23, %r2;
cvt.s64.s32 %rd24, %r4;
add.s64 %rd25, %rd23, %rd24;
add.s64 %rd26, %rd25, %rd21;
add.s64 %rd27, %rd26, %rd22;
shl.b64 %rd28, %rd27, 2;
add.s64 %rd29, %rd20, %rd28;
ld.global.f32 %f111, [%rd29];

$L__BB228_6:
setp.eq.s64 %p6, %rd13, 0;
setp.eq.s64 %p7, %rd12, 0;
or.pred %p8, %p6, %p7;
@%p8 bra $L__BB228_8;

cvta.to.global.u64 %rd30, %rd13;
mov.u32 %r61, %ctaid.z;
mul.lo.s32 %r62, %r61, %r48;
cvt.u64.u32 %rd31, %r62;
mul.lo.s32 %r63, %r11, %r42;
cvt.u64.u32 %rd32, %r63;
cvt.u64.u32 %rd33, %r2;
cvt.s64.s32 %rd34, %r4;
add.s64 %rd35, %rd33, %rd34;
add.s64 %rd36, %rd35, %rd31;
add.s64 %rd37, %rd36, %rd32;
shl.b64 %rd38, %rd37, 2;
add.s64 %rd39, %rd30, %rd38;
ld.global.f32 %f110, [%rd39];

$L__BB228_8:
shr.u32 %r65, %r32, 8;
shr.s32 %r66, %r3, 31;
shr.u32 %r67, %r66, 29;
add.s32 %r68, %r3, %r67;
shr.s32 %r547, %r68, 3;
setp.ge.s32 %p9, %r547, %r65;
mov.f32 %f116, %f114;
@%p9 bra $L__BB228_13;

mov.u32 %r69, %ctaid.z;
and.b32 %r73, %r68, -8;
sub.s32 %r13, %r3, %r73;
shl.b32 %r546, %r547, 3;
mul.hi.u32 %r74, %r69, %r43;
add.s32 %r75, %r74, %r69;
shr.u32 %r76, %r75, %r44;
mad.lo.s32 %r77, %r76, %r46, %r547;
mad.lo.s32 %r78, %r544, %r40, %r77;
mad.lo.s32 %r79, %r4, %r36, %r78;
mul.wide.s32 %rd86, %r79, 82;
mul.lo.s32 %r80, %r545, %r41;
cvt.u64.u32 %rd40, %r80;
mul.lo.s32 %r81, %r69, %r47;
cvt.u64.u32 %rd41, %r81;
add.s64 %rd5, %rd40, %rd41;
mov.f32 %f116, 0f00000000;
cvta.to.global.u64 %rd42, %rd9;
cvta.to.global.u64 %rd46, %rd8;
cvt.s64.s32 %rd49, %r13;
add.s64 %rd50, %rd46, %rd49;
mov.f32 %f114, %f116;

$L__BB228_10:
cvt.u64.u32 %rd43, %r546;
add.s64 %rd44, %rd5, %rd43;
shl.b32 %r131, %r13, 1;
mul.wide.s32 %rd45, %r131, 2;
add.s64 %rd47, %rd46, %rd45;
add.s64 %rd48, %rd47, %rd86;
ld.global.nc.u16 %rs2, [%rd48+2];
shr.u16 %rs3, %rs2, 8;
cvt.u32.u16 %r132, %rs3;
ld.global.nc.u16 %rs4, [%rd48+4];
shr.u16 %rs5, %rs4, 8;
cvt.u32.u16 %r133, %rs5;
add.s64 %rd51, %rd50, %rd86;
ld.global.nc.u8 %rs6, [%rd51+66];
ld.global.nc.u16 %rs7, [%rd48+34];
ld.global.nc.u16 %rs8, [%rd48+36];
ld.global.nc.u8 %rs9, [%rd51+74];
and.b16 %rs10, %rs9, 240;
and.b16 %rs11, %rs9, 15;
shr.u16 %rs12, %rs10, 4;
add.s64 %rd52, %rd44, %rd49;
and.b16 %rs13, %rs2, 255;
cvt.u32.u16 %r134, %rs13;
shl.b16 %rs14, %rs6, 8;
cvt.u32.u16 %r135, %rs14;
and.b32 %r136, %r135, 768;
or.b32 %r137, %r136, %r134;
mul.wide.u32 %rd53, %r137, 8;
mov.u64 %rd54, iq2s_grid;
add.s64 %rd55, %rd54, %rd53;
and.b16 %rs15, %rs7, 255;
mul.wide.u16 %r138, %rs15, 128;
cvt.u32.u16 %r139, %rs7;
shl.b32 %r140, %r139, 21;
or.b32 %r141, %r138, %r140;
and.b32 %r142, %r141, 25166208;
or.b32 %r143, %r142, -2139062144;
add.s32 %r144, %r143, -16843009;
or.b32 %r82, %r144, %r142;

prmt.b32 %r82, %r82, 0, 0xba98;

mul.wide.u16 %r145, %rs15, 8;
shl.b32 %r146, %r139, 17;
or.b32 %r147, %r145, %r146;
and.b32 %r148, %r147, 25166208;
or.b32 %r149, %r148, -2139062144;
add.s32 %r150, %r149, -16843009;
or.b32 %r84, %r150, %r148;

prmt.b32 %r84, %r84, 0, 0xba98;

ld.global.nc.v2.u32 {%r151, %r152}, [%rd55];
mov.u32 %r113, 0;
xor.b32 %r155, %r151, %r82;
or.b32 %r156, %r155, -2139062144;
and.b32 %r157, %r82, 2139062143;
and.b32 %r158, %r151, -2139062144;
sub.s32 %r159, %r156, %r157;
xor.b32 %r160, %r158, %r159;
xor.b32 %r87, %r160, -2139062144;
xor.b32 %r161, %r152, %r84;
or.b32 %r162, %r161, -2139062144;
and.b32 %r163, %r84, 2139062143;
and.b32 %r164, %r152, -2139062144;
sub.s32 %r165, %r162, %r163;
xor.b32 %r166, %r164, %r165;
xor.b32 %r91, %r166, -2139062144;
mul.lo.s64 %rd56, %rd52, 36;
add.s64 %rd57, %rd42, %rd56;
ld.global.nc.u32 %r17, [%rd57+8];
ld.global.nc.u32 %r18, [%rd57+4];

dp4a.s32.s32 %r86, %r87, %r18, %r113;


dp4a.s32.s32 %r90, %r91, %r17, %r86;

mul.wide.u16 %r167, %rs6, 64;
and.b32 %r168, %r167, 768;
or.b32 %r169, %r168, %r132;
mul.wide.u32 %rd58, %r169, 8;
add.s64 %rd59, %rd54, %rd58;
shr.u16 %rs17, %rs7, 8;
mul.wide.u16 %r170, %rs17, 128;
cvt.u32.u16 %r171, %rs17;
shl.b32 %r172, %r171, 21;
or.b32 %r173, %r170, %r172;
and.b32 %r174, %r173, 25166208;
or.b32 %r175, %r174, -2139062144;
add.s32 %r176, %r175, -16843009;
or.b32 %r94, %r176, %r174;

prmt.b32 %r94, %r94, 0, 0xba98;

mul.wide.u16 %r177, %rs17, 8;
shl.b32 %r178, %r171, 17;
or.b32 %r179, %r177, %r178;
and.b32 %r180, %r179, 25166208;
or.b32 %r181, %r180, -2139062144;
add.s32 %r182, %r181, -16843009;
or.b32 %r96, %r182, %r180;

prmt.b32 %r96, %r96, 0, 0xba98;

ld.global.nc.v2.u32 {%r183, %r184}, [%rd59];
xor.b32 %r187, %r183, %r94;
or.b32 %r188, %r187, -2139062144;
and.b32 %r189, %r94, 2139062143;
and.b32 %r190, %r183, -2139062144;
sub.s32 %r191, %r188, %r189;
xor.b32 %r192, %r190, %r191;
xor.b32 %r99, %r192, -2139062144;
xor.b32 %r193, %r184, %r96;
or.b32 %r194, %r193, -2139062144;
and.b32 %r195, %r96, 2139062143;
and.b32 %r196, %r184, -2139062144;
sub.s32 %r197, %r194, %r195;
xor.b32 %r198, %r196, %r197;
xor.b32 %r103, %r198, -2139062144;
ld.global.nc.u32 %r19, [%rd57+16];
ld.global.nc.u32 %r20, [%rd57+12];

dp4a.s32.s32 %r98, %r99, %r20, %r90;


dp4a.s32.s32 %r102, %r103, %r19, %r98;

and.b16 %rs18, %rs4, 255;
cvt.u32.u16 %r199, %rs18;
mul.wide.u16 %r200, %rs6, 16;
and.b32 %r201, %r200, 768;
or.b32 %r202, %r201, %r199;
mul.wide.u32 %rd60, %r202, 8;
add.s64 %rd61, %rd54, %rd60;
and.b16 %rs19, %rs8, 255;
mul.wide.u16 %r203, %rs19, 128;
cvt.u32.u16 %r204, %rs8;
shl.b32 %r205, %r204, 21;
or.b32 %r206, %r203, %r205;
and.b32 %r207, %r206, 25166208;
or.b32 %r208, %r207, -2139062144;
add.s32 %r209, %r208, -16843009;
or.b32 %r106, %r209, %r207;

prmt.b32 %r106, %r106, 0, 0xba98;

mul.wide.u16 %r210, %rs19, 8;
shl.b32 %r211, %r204, 17;
or.b32 %r212, %r210, %r211;
and.b32 %r213, %r212, 25166208;
or.b32 %r214, %r213, -2139062144;
add.s32 %r215, %r214, -16843009;
or.b32 %r108, %r215, %r213;

prmt.b32 %r108, %r108, 0, 0xba98;

ld.global.nc.v2.u32 {%r216, %r217}, [%rd61];
xor.b32 %r220, %r216, %r106;
or.b32 %r221, %r220, -2139062144;
and.b32 %r222, %r106, 2139062143;
and.b32 %r223, %r216, -2139062144;
sub.s32 %r224, %r221, %r222;
xor.b32 %r225, %r223, %r224;
xor.b32 %r111, %r225, -2139062144;
xor.b32 %r226, %r217, %r108;
or.b32 %r227, %r226, -2139062144;
and.b32 %r228, %r108, 2139062143;
and.b32 %r229, %r217, -2139062144;
sub.s32 %r230, %r227, %r228;
xor.b32 %r231, %r229, %r230;
xor.b32 %r115, %r231, -2139062144;
ld.global.nc.u32 %r21, [%rd57+24];
ld.global.nc.u32 %r22, [%rd57+20];

dp4a.s32.s32 %r110, %r111, %r22, %r113;


dp4a.s32.s32 %r114, %r115, %r21, %r110;

mul.wide.u16 %r232, %rs6, 4;
and.b32 %r233, %r232, 768;
or.b32 %r234, %r233, %r133;
mul.wide.u32 %rd62, %r234, 8;
add.s64 %rd63, %rd54, %rd62;
shr.u16 %rs20, %rs8, 8;
mul.wide.u16 %r235, %rs20, 128;
cvt.u32.u16 %r236, %rs20;
shl.b32 %r237, %r236, 21;
or.b32 %r238, %r235, %r237;
and.b32 %r239, %r238, 25166208;
or.b32 %r240, %r239, -2139062144;
add.s32 %r241, %r240, -16843009;
or.b32 %r118, %r241, %r239;

prmt.b32 %r118, %r118, 0, 0xba98;

mul.wide.u16 %r242, %rs20, 8;
shl.b32 %r243, %r236, 17;
or.b32 %r244, %r242, %r243;
and.b32 %r245, %r244, 25166208;
or.b32 %r246, %r245, -2139062144;
add.s32 %r247, %r246, -16843009;
or.b32 %r120, %r247, %r245;

prmt.b32 %r120, %r120, 0, 0xba98;

ld.global.nc.v2.u32 {%r248, %r249}, [%rd63];
xor.b32 %r252, %r248, %r118;
or.b32 %r253, %r252, -2139062144;
and.b32 %r254, %r118, 2139062143;
and.b32 %r255, %r248, -2139062144;
sub.s32 %r256, %r253, %r254;
xor.b32 %r257, %r255, %r256;
xor.b32 %r123, %r257, -2139062144;
xor.b32 %r258, %r249, %r120;
or.b32 %r259, %r258, -2139062144;
and.b32 %r260, %r120, 2139062143;
and.b32 %r261, %r249, -2139062144;
sub.s32 %r262, %r259, %r260;
xor.b32 %r263, %r261, %r262;
xor.b32 %r127, %r263, -2139062144;
ld.global.nc.u32 %r23, [%rd57+32];
ld.global.nc.u32 %r24, [%rd57+28];

dp4a.s32.s32 %r122, %r123, %r24, %r114;


dp4a.s32.s32 %r126, %r127, %r23, %r122;

cvt.u32.u16 %r264, %rs11;
cvt.u32.u16 %r265, %rs12;
mul.lo.s32 %r266, %r102, %r264;
mad.lo.s32 %r267, %r126, %r265, %r266;
add.s32 %r268, %r126, %r102;
shr.u32 %r269, %r268, 31;
add.s32 %r270, %r268, %r269;
shr.s32 %r271, %r270, 1;
add.s32 %r272, %r267, %r271;
shr.s32 %r273, %r272, 31;
shr.u32 %r274, %r273, 30;
add.s32 %r275, %r272, %r274;
shr.s32 %r276, %r275, 2;
add.s64 %rd64, %rd46, %rd86;
ld.global.nc.u16 %rs1, [%rd64];

{ cvt.f32.f16 %f40, %rs1;}


ld.global.nc.u32 %r25, [%rd57];

{.reg .f16 low,high;
mov.b32 {low,high},%r25;
cvt.f32.f16 %f41, low;}


mul.ftz.f32 %f42, %f40, %f41;
cvt.rn.f32.s32 %f43, %r276;
fma.rn.ftz.f32 %f116, %f42, %f43, %f116;
setp.eq.s64 %p10, %rd12, 0;
@%p10 bra $L__BB228_12;

cvta.to.global.u64 %rd66, %rd12;
add.s64 %rd67, %rd66, %rd45;
add.s64 %rd68, %rd67, %rd86;
ld.global.u16 %rs22, [%rd68+2];
shr.u16 %rs23, %rs22, 8;
cvt.u32.u16 %r327, %rs23;
ld.global.u16 %rs24, [%rd68+4];
shr.u16 %rs25, %rs24, 8;
cvt.u32.u16 %r328, %rs25;
add.s64 %rd70, %rd66, %rd49;
add.s64 %rd71, %rd70, %rd86;
ld.global.u16 %rs26, [%rd68+34];
ld.global.u16 %rs27, [%rd68+36];
ld.global.u8 %rs28, [%rd71+74];
and.b16 %rs29, %rs28, 15;
shr.u16 %rs30, %rs28, 4;
and.b16 %rs31, %rs22, 255;
cvt.u32.u16 %r329, %rs31;
ld.global.u8 %rs32, [%rd71+66];
shl.b16 %rs33, %rs32, 8;
cvt.u32.u16 %r330, %rs33;
and.b32 %r331, %r330, 768;
or.b32 %r332, %r331, %r329;
mul.wide.u32 %rd72, %r332, 8;
add.s64 %rd74, %rd54, %rd72;
and.b16 %rs34, %rs26, 255;
mul.wide.u16 %r333, %rs34, 128;
cvt.u32.u16 %r334, %rs26;
shl.b32 %r335, %r334, 21;
or.b32 %r336, %r333, %r335;
and.b32 %r337, %r336, 25166208;
or.b32 %r338, %r337, -2139062144;
add.s32 %r339, %r338, -16843009;
or.b32 %r277, %r339, %r337;

prmt.b32 %r277, %r277, 0, 0xba98;

mul.wide.u16 %r340, %rs34, 8;
shl.b32 %r341, %r334, 17;
or.b32 %r342, %r340, %r341;
and.b32 %r343, %r342, 25166208;
or.b32 %r344, %r343, -2139062144;
add.s32 %r345, %r344, -16843009;
or.b32 %r279, %r345, %r343;

prmt.b32 %r279, %r279, 0, 0xba98;

ld.global.nc.v2.u32 {%r346, %r347}, [%rd74];
xor.b32 %r350, %r346, %r277;
or.b32 %r351, %r350, -2139062144;
and.b32 %r352, %r277, 2139062143;
and.b32 %r353, %r346, -2139062144;
sub.s32 %r354, %r351, %r352;
xor.b32 %r355, %r353, %r354;
xor.b32 %r282, %r355, -2139062144;
xor.b32 %r356, %r347, %r279;
or.b32 %r357, %r356, -2139062144;
and.b32 %r358, %r279, 2139062143;
and.b32 %r359, %r347, -2139062144;
sub.s32 %r360, %r357, %r358;
xor.b32 %r361, %r359, %r360;
xor.b32 %r286, %r361, -2139062144;

dp4a.s32.s32 %r281, %r282, %r18, %r113;


dp4a.s32.s32 %r285, %r286, %r17, %r281;

mul.wide.u16 %r362, %rs32, 64;
and.b32 %r363, %r362, 768;
or.b32 %r364, %r363, %r327;
mul.wide.u32 %rd75, %r364, 8;
add.s64 %rd76, %rd54, %rd75;
shr.u16 %rs35, %rs26, 8;
mul.wide.u16 %r365, %rs35, 128;
cvt.u32.u16 %r366, %rs35;
shl.b32 %r367, %r366, 21;
or.b32 %r368, %r365, %r367;
and.b32 %r369, %r368, 25166208;
or.b32 %r370, %r369, -2139062144;
add.s32 %r371, %r370, -16843009;
or.b32 %r289, %r371, %r369;

prmt.b32 %r289, %r289, 0, 0xba98;

mul.wide.u16 %r372, %rs35, 8;
shl.b32 %r373, %r366, 17;
or.b32 %r374, %r372, %r373;
and.b32 %r375, %r374, 25166208;
or.b32 %r376, %r375, -2139062144;
add.s32 %r377, %r376, -16843009;
or.b32 %r291, %r377, %r375;

prmt.b32 %r291, %r291, 0, 0xba98;

ld.global.nc.v2.u32 {%r378, %r379}, [%rd76];
xor.b32 %r382, %r378, %r289;
or.b32 %r383, %r382, -2139062144;
and.b32 %r384, %r289, 2139062143;
and.b32 %r385, %r378, -2139062144;
sub.s32 %r386, %r383, %r384;
xor.b32 %r387, %r385, %r386;
xor.b32 %r294, %r387, -2139062144;
xor.b32 %r388, %r379, %r291;
or.b32 %r389, %r388, -2139062144;
and.b32 %r390, %r291, 2139062143;
and.b32 %r391, %r379, -2139062144;
sub.s32 %r392, %r389, %r390;
xor.b32 %r393, %r391, %r392;
xor.b32 %r298, %r393, -2139062144;

dp4a.s32.s32 %r293, %r294, %r20, %r285;


dp4a.s32.s32 %r297, %r298, %r19, %r293;

and.b16 %rs36, %rs24, 255;
cvt.u32.u16 %r394, %rs36;
mul.wide.u16 %r395, %rs32, 16;
and.b32 %r396, %r395, 768;
or.b32 %r397, %r396, %r394;
mul.wide.u32 %rd77, %r397, 8;
add.s64 %rd78, %rd54, %rd77;
and.b16 %rs37, %rs27, 255;
mul.wide.u16 %r398, %rs37, 128;
cvt.u32.u16 %r399, %rs27;
shl.b32 %r400, %r399, 21;
or.b32 %r401, %r398, %r400;
and.b32 %r402, %r401, 25166208;
or.b32 %r403, %r402, -2139062144;
add.s32 %r404, %r403, -16843009;
or.b32 %r301, %r404, %r402;

prmt.b32 %r301, %r301, 0, 0xba98;

mul.wide.u16 %r405, %rs37, 8;
shl.b32 %r406, %r399, 17;
or.b32 %r407, %r405, %r406;
and.b32 %r408, %r407, 25166208;
or.b32 %r409, %r408, -2139062144;
add.s32 %r410, %r409, -16843009;
or.b32 %r303, %r410, %r408;

prmt.b32 %r303, %r303, 0, 0xba98;

ld.global.nc.v2.u32 {%r411, %r412}, [%rd78];
xor.b32 %r415, %r411, %r301;
or.b32 %r416, %r415, -2139062144;
and.b32 %r417, %r301, 2139062143;
and.b32 %r418, %r411, -2139062144;
sub.s32 %r419, %r416, %r417;
xor.b32 %r420, %r418, %r419;
xor.b32 %r306, %r420, -2139062144;
xor.b32 %r421, %r412, %r303;
or.b32 %r422, %r421, -2139062144;
and.b32 %r423, %r303, 2139062143;
and.b32 %r424, %r412, -2139062144;
sub.s32 %r425, %r422, %r423;
xor.b32 %r426, %r424, %r425;
xor.b32 %r310, %r426, -2139062144;

dp4a.s32.s32 %r305, %r306, %r22, %r113;


dp4a.s32.s32 %r309, %r310, %r21, %r305;

mul.wide.u16 %r427, %rs32, 4;
and.b32 %r428, %r427, 768;
or.b32 %r429, %r428, %r328;
mul.wide.u32 %rd79, %r429, 8;
add.s64 %rd80, %rd54, %rd79;
shr.u16 %rs38, %rs27, 8;
mul.wide.u16 %r430, %rs38, 128;
cvt.u32.u16 %r431, %rs38;
shl.b32 %r432, %r431, 21;
or.b32 %r433, %r430, %r432;
and.b32 %r434, %r433, 25166208;
or.b32 %r435, %r434, -2139062144;
add.s32 %r436, %r435, -16843009;
or.b32 %r313, %r436, %r434;

prmt.b32 %r313, %r313, 0, 0xba98;

mul.wide.u16 %r437, %rs38, 8;
shl.b32 %r438, %r431, 17;
or.b32 %r439, %r437, %r438;
and.b32 %r440, %r439, 25166208;
or.b32 %r441, %r440, -2139062144;
add.s32 %r442, %r441, -16843009;
or.b32 %r315, %r442, %r440;

prmt.b32 %r315, %r315, 0, 0xba98;

ld.global.nc.v2.u32 {%r443, %r444}, [%rd80];
xor.b32 %r447, %r443, %r313;
or.b32 %r448, %r447, -2139062144;
and.b32 %r449, %r313, 2139062143;
and.b32 %r450, %r443, -2139062144;
sub.s32 %r451, %r448, %r449;
xor.b32 %r452, %r450, %r451;
xor.b32 %r318, %r452, -2139062144;
xor.b32 %r453, %r444, %r315;
or.b32 %r454, %r453, -2139062144;
and.b32 %r455, %r315, 2139062143;
and.b32 %r456, %r444, -2139062144;
sub.s32 %r457, %r454, %r455;
xor.b32 %r458, %r456, %r457;
xor.b32 %r322, %r458, -2139062144;

dp4a.s32.s32 %r317, %r318, %r24, %r309;


dp4a.s32.s32 %r321, %r322, %r23, %r317;

cvt.u32.u16 %r459, %rs29;
cvt.u32.u16 %r460, %rs30;
mul.lo.s32 %r461, %r297, %r459;
mad.lo.s32 %r462, %r321, %r460, %r461;
add.s32 %r463, %r321, %r297;
shr.u32 %r464, %r463, 31;
add.s32 %r465, %r463, %r464;
shr.s32 %r466, %r465, 1;
add.s32 %r467, %r462, %r466;
shr.s32 %r468, %r467, 31;
shr.u32 %r469, %r468, 30;
add.s32 %r470, %r467, %r469;
shr.s32 %r471, %r470, 2;
add.s64 %rd81, %rd66, %rd86;
ld.global.u16 %rs21, [%rd81];

{ cvt.f32.f16 %f44, %rs21;}



{.reg .f16 low,high;
mov.b32 {low,high},%r25;
cvt.f32.f16 %f45, low;}


mul.ftz.f32 %f46, %f44, %f45;
cvt.rn.f32.s32 %f47, %r471;
fma.rn.ftz.f32 %f114, %f46, %f47, %f114;

$L__BB228_12:
add.s32 %r546, %r546, 128;
add.s64 %rd86, %rd86, 1312;
add.s32 %r547, %r547, 16;
setp.lt.s32 %p11, %r547, %r65;
@%p11 bra $L__BB228_10;

$L__BB228_13:
mov.u32 %r535, %tid.y;
setp.eq.s32 %p12, %r535, 0;
@%p12 bra $L__BB228_16;

mov.u32 %r537, %tid.y;
mov.u32 %r536, %tid.x;
setp.eq.s64 %p13, %rd12, 0;
add.s32 %r28, %r537, -1;
shl.b32 %r476, %r537, 7;
mov.u32 %r477, _ZZ13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r478, %r477, %r476;
shl.b32 %r479, %r536, 2;
add.s32 %r480, %r478, %r479;
st.shared.f32 [%r480+-128], %f116;
@%p13 bra $L__BB228_16;

shl.b32 %r482, %r28, 7;
mov.u32 %r483, _ZZ13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r484, %r483, %r482;
add.s32 %r486, %r484, %r479;
st.shared.f32 [%r486], %f114;

$L__BB228_16:
mov.u32 %r538, %tid.y;
setp.ne.s32 %p14, %r538, 0;
bar.sync 0;
@%p14 bra $L__BB228_35;

mov.u32 %r539, %tid.x;
setp.ne.s64 %p15, %rd12, 0;
shl.b32 %r489, %r539, 2;
mov.u32 %r490, _ZZ13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r29, %r490, %r489;
ld.shared.f32 %f48, [%r29];
add.ftz.f32 %f13, %f116, %f48;
@%p15 bra $L__BB228_19;
bra.uni $L__BB228_18;

$L__BB228_19:
mov.u32 %r493, _ZZ13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r494, %r493, %r489;
ld.shared.f32 %f52, [%r494];
add.ftz.f32 %f53, %f114, %f52;
ld.shared.f32 %f54, [%r29+128];
add.ftz.f32 %f55, %f13, %f54;
ld.shared.f32 %f56, [%r494+128];
add.ftz.f32 %f57, %f53, %f56;
ld.shared.f32 %f58, [%r29+256];
add.ftz.f32 %f118, %f55, %f58;
ld.shared.f32 %f59, [%r494+256];
add.ftz.f32 %f114, %f57, %f59;
bra.uni $L__BB228_20;

$L__BB228_18:
ld.shared.f32 %f49, [%r29+128];
add.ftz.f32 %f50, %f13, %f49;
ld.shared.f32 %f51, [%r29+256];
add.ftz.f32 %f118, %f50, %f51;

$L__BB228_20:
mov.b32 %r495, %f118;
mov.u32 %r496, 31;
mov.u32 %r497, 16;
mov.u32 %r498, -1;
shfl.sync.bfly.b32 %r499|%p16, %r495, %r497, %r496, %r498;
mov.b32 %f60, %r499;
add.ftz.f32 %f61, %f118, %f60;
mov.b32 %r500, %f61;
mov.u32 %r501, 8;
shfl.sync.bfly.b32 %r502|%p17, %r500, %r501, %r496, %r498;
mov.b32 %f62, %r502;
add.ftz.f32 %f63, %f61, %f62;
mov.b32 %r503, %f63;
mov.u32 %r504, 4;
shfl.sync.bfly.b32 %r505|%p18, %r503, %r504, %r496, %r498;
mov.b32 %f64, %r505;
add.ftz.f32 %f65, %f63, %f64;
mov.b32 %r506, %f65;
mov.u32 %r507, 2;
shfl.sync.bfly.b32 %r508|%p19, %r506, %r507, %r496, %r498;
mov.b32 %f66, %r508;
add.ftz.f32 %f19, %f65, %f66;
mov.b32 %r509, %f19;
mov.u32 %r510, 1;
shfl.sync.bfly.b32 %r30|%p1, %r509, %r510, %r496, %r498;
setp.eq.s64 %p20, %rd12, 0;
@%p20 bra $L__BB228_22;

mov.b32 %r511, %f114;
shfl.sync.bfly.b32 %r515|%p21, %r511, %r497, %r496, %r498;
mov.b32 %f67, %r515;
add.ftz.f32 %f68, %f114, %f67;
mov.b32 %r516, %f68;
shfl.sync.bfly.b32 %r518|%p22, %r516, %r501, %r496, %r498;
mov.b32 %f69, %r518;
add.ftz.f32 %f70, %f68, %f69;
mov.b32 %r519, %f70;
shfl.sync.bfly.b32 %r521|%p23, %r519, %r504, %r496, %r498;
mov.b32 %f71, %r521;
add.ftz.f32 %f72, %f70, %f71;
mov.b32 %r522, %f72;
shfl.sync.bfly.b32 %r524|%p24, %r522, %r507, %r496, %r498;
mov.b32 %f73, %r524;
add.ftz.f32 %f74, %f72, %f73;
mov.b32 %r525, %f74;
shfl.sync.bfly.b32 %r527|%p25, %r525, %r510, %r496, %r498;
mov.b32 %f75, %r527;
add.ftz.f32 %f114, %f74, %f75;

$L__BB228_22:
mov.u32 %r540, %tid.x;
setp.ne.s32 %p26, %r540, 0;
mov.b32 %f76, %r30;
add.ftz.f32 %f22, %f19, %f76;
@%p26 bra $L__BB228_35;

add.ftz.f32 %f120, %f111, %f22;
@%p20 bra $L__BB228_34;

add.ftz.f32 %f24, %f110, %f114;
setp.gt.s32 %p28, %r31, 2;
@%p28 bra $L__BB228_28;

setp.eq.s32 %p31, %r31, 1;
@%p31 bra $L__BB228_32;

setp.eq.s32 %p32, %r31, 2;
@%p32 bra $L__BB228_27;
bra.uni $L__BB228_33;

$L__BB228_27:
mul.ftz.f32 %f96, %f24, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f97, %f96;
add.ftz.f32 %f98, %f97, 0f3F800000;
div.approx.ftz.f32 %f99, %f24, %f98;
mul.ftz.f32 %f120, %f120, %f99;
bra.uni $L__BB228_34;

$L__BB228_28:
setp.eq.s32 %p29, %r31, 3;
@%p29 bra $L__BB228_31;

setp.ne.s32 %p30, %r31, 6;
@%p30 bra $L__BB228_33;

ld.param.f32 %f108, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
min.ftz.f32 %f77, %f24, %f108;
neg.ftz.f32 %f78, %f108;
min.ftz.f32 %f79, %f120, %f108;
max.ftz.f32 %f80, %f79, %f78;
mul.ftz.f32 %f81, %f77, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f82, %f81;
add.ftz.f32 %f83, %f82, 0f3F800000;
div.approx.ftz.f32 %f84, %f77, %f83;
mul.ftz.f32 %f120, %f80, %f84;
bra.uni $L__BB228_34;

$L__BB228_32:
mul.ftz.f32 %f100, %f24, 0f3F000000;
mul.ftz.f32 %f101, %f24, 0f3D372713;
fma.rn.ftz.f32 %f102, %f24, %f101, 0f3F800000;
mul.ftz.f32 %f103, %f24, 0f3F4C422A;
mul.ftz.f32 %f104, %f103, %f102;
tanh.approx.f32 %f105, %f104;
add.ftz.f32 %f106, %f105, 0f3F800000;
mul.ftz.f32 %f107, %f100, %f106;
mul.ftz.f32 %f120, %f120, %f107;
bra.uni $L__BB228_34;

$L__BB228_31:
mov.f32 %f85, 0f40E00000;
min.ftz.f32 %f86, %f24, %f85;
min.ftz.f32 %f87, %f120, %f85;
mov.f32 %f88, 0fC0E00000;
max.ftz.f32 %f89, %f87, %f88;
mul.ftz.f32 %f90, %f86, 0f3FD9DB23;
mul.ftz.f32 %f91, %f90, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f92, %f91;
add.ftz.f32 %f93, %f92, 0f3F800000;
div.approx.ftz.f32 %f94, %f86, %f93;
add.ftz.f32 %f95, %f89, 0f3F800000;
mul.ftz.f32 %f120, %f95, %f94;
bra.uni $L__BB228_34;

$L__BB228_33:
mul.ftz.f32 %f120, %f120, %f24;

$L__BB228_34:
ld.param.u64 %rd85, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
mov.u32 %r543, %ctaid.x;
ld.param.u32 %r542, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r541, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
cvta.to.global.u64 %rd82, %rd85;
mov.u32 %r530, %ctaid.z;
mov.u32 %r531, %ctaid.y;
mad.lo.s32 %r533, %r531, %r542, %r543;
mad.lo.s32 %r534, %r530, %r541, %r533;
mul.wide.u32 %rd83, %r534, 4;
add.s64 %rd84, %rd82, %rd83;
st.global.f32 [%rd84], %f120;

$L__BB228_35:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<12>;
.reg .b16 %rs<21>;
.reg .f32 %f<28>;
.reg .b32 %r<288>;
.reg .b64 %rd<41>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[384];

ld.param.u64 %rd9, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd10, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd11, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd12, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r19, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r18, [_Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
mov.u32 %r1, %tid.y;
shl.b32 %r35, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r35, %r2;
shr.u32 %r4, %r34, 8;
mov.u32 %r285, %ctaid.y;
setp.eq.s64 %p2, %rd11, 0;
@%p2 bra $L__BB229_2;

cvta.to.global.u64 %rd13, %rd11;
mul.wide.u32 %rd14, %r285, 4;
add.s64 %rd15, %rd13, %rd14;
ld.global.nc.u32 %r284, [%rd15];
mul.hi.u32 %r36, %r285, %r18;
add.s32 %r37, %r36, %r285;
shr.u32 %r38, %r37, %r19;
mul.lo.s32 %r39, %r38, %r20;
sub.s32 %r285, %r285, %r39;
bra.uni $L__BB229_3;

$L__BB229_2:
mul.hi.u32 %r40, %r285, %r22;
add.s32 %r41, %r40, %r285;
shr.u32 %r284, %r41, %r23;

$L__BB229_3:
shr.s32 %r42, %r3, 31;
shr.u32 %r43, %r42, 29;
add.s32 %r44, %r3, %r43;
shr.s32 %r287, %r44, 3;
setp.ge.s32 %p3, %r287, %r4;
mov.f32 %f27, 0f00000000;
@%p3 bra $L__BB229_6;

mov.u32 %r45, %ctaid.z;
and.b32 %r49, %r44, -8;
sub.s32 %r50, %r3, %r49;
shl.b32 %r51, %r50, 1;
cvt.s64.s32 %rd1, %r50;
mul.hi.u32 %r52, %r45, %r28;
add.s32 %r53, %r52, %r45;
shr.u32 %r54, %r53, %r29;
mad.lo.s32 %r55, %r54, %r31, %r287;
mad.lo.s32 %r56, %r284, %r25, %r55;
mov.u32 %r57, %ctaid.x;
mad.lo.s32 %r58, %r57, %r21, %r56;
mul.wide.s32 %rd16, %r58, 82;
cvta.to.global.u64 %rd17, %rd9;
add.s64 %rd40, %rd17, %rd16;
shl.b32 %r286, %r287, 3;
add.s64 %rd3, %rd1, 66;
mul.wide.s32 %rd18, %r51, 2;
add.s64 %rd4, %rd18, 34;
cvta.to.global.u64 %rd5, %rd10;
mul.lo.s32 %r59, %r285, %r26;
cvt.u64.u32 %rd19, %r59;
mul.lo.s32 %r60, %r45, %r32;
cvt.u64.u32 %rd20, %r60;
add.s64 %rd6, %rd19, %rd20;
mov.f32 %f27, 0f00000000;

$L__BB229_5:
cvt.u64.u32 %rd21, %r286;
add.s64 %rd22, %rd6, %rd21;
add.s64 %rd23, %rd40, %rd4;
ld.global.nc.u16 %rs2, [%rd23+-32];
shr.u16 %rs3, %rs2, 8;
cvt.u32.u16 %r110, %rs3;
ld.global.nc.u16 %rs4, [%rd23+-30];
shr.u16 %rs5, %rs4, 8;
cvt.u32.u16 %r111, %rs5;
add.s64 %rd24, %rd40, %rd3;
ld.global.nc.u8 %rs6, [%rd24];
ld.global.nc.u16 %rs7, [%rd23];
ld.global.nc.u16 %rs8, [%rd23+2];
ld.global.nc.u8 %rs9, [%rd24+8];
and.b16 %rs10, %rs9, 240;
and.b16 %rs11, %rs9, 15;
shr.u16 %rs12, %rs10, 4;
add.s64 %rd25, %rd22, %rd1;
and.b16 %rs13, %rs2, 255;
cvt.u32.u16 %r112, %rs13;
shl.b16 %rs14, %rs6, 8;
cvt.u32.u16 %r113, %rs14;
and.b32 %r114, %r113, 768;
or.b32 %r115, %r114, %r112;
mul.wide.u32 %rd26, %r115, 8;
mov.u64 %rd27, iq2s_grid;
add.s64 %rd28, %rd27, %rd26;
and.b16 %rs15, %rs7, 255;
mul.wide.u16 %r116, %rs15, 128;
cvt.u32.u16 %r117, %rs7;
shl.b32 %r118, %r117, 21;
or.b32 %r119, %r116, %r118;
and.b32 %r120, %r119, 25166208;
or.b32 %r121, %r120, -2139062144;
add.s32 %r122, %r121, -16843009;
or.b32 %r61, %r122, %r120;

prmt.b32 %r61, %r61, 0, 0xba98;

mul.wide.u16 %r123, %rs15, 8;
shl.b32 %r124, %r117, 17;
or.b32 %r125, %r123, %r124;
and.b32 %r126, %r125, 25166208;
or.b32 %r127, %r126, -2139062144;
add.s32 %r128, %r127, -16843009;
or.b32 %r63, %r128, %r126;

prmt.b32 %r63, %r63, 0, 0xba98;

ld.global.nc.v2.u32 {%r129, %r130}, [%rd28];
mov.u32 %r92, 0;
xor.b32 %r133, %r129, %r61;
or.b32 %r134, %r133, -2139062144;
and.b32 %r135, %r61, 2139062143;
and.b32 %r136, %r129, -2139062144;
sub.s32 %r137, %r134, %r135;
xor.b32 %r138, %r136, %r137;
xor.b32 %r66, %r138, -2139062144;
xor.b32 %r139, %r130, %r63;
or.b32 %r140, %r139, -2139062144;
and.b32 %r141, %r63, 2139062143;
and.b32 %r142, %r130, -2139062144;
sub.s32 %r143, %r140, %r141;
xor.b32 %r144, %r142, %r143;
xor.b32 %r70, %r144, -2139062144;
mul.lo.s64 %rd29, %rd25, 36;
add.s64 %rd30, %rd5, %rd29;
ld.global.nc.u32 %r71, [%rd30+8];
ld.global.nc.u32 %r67, [%rd30+4];

dp4a.s32.s32 %r65, %r66, %r67, %r92;


dp4a.s32.s32 %r69, %r70, %r71, %r65;

mul.wide.u16 %r145, %rs6, 64;
and.b32 %r146, %r145, 768;
or.b32 %r147, %r146, %r110;
mul.wide.u32 %rd31, %r147, 8;
add.s64 %rd32, %rd27, %rd31;
shr.u16 %rs17, %rs7, 8;
mul.wide.u16 %r148, %rs17, 128;
cvt.u32.u16 %r149, %rs17;
shl.b32 %r150, %r149, 21;
or.b32 %r151, %r148, %r150;
and.b32 %r152, %r151, 25166208;
or.b32 %r153, %r152, -2139062144;
add.s32 %r154, %r153, -16843009;
or.b32 %r73, %r154, %r152;

prmt.b32 %r73, %r73, 0, 0xba98;

mul.wide.u16 %r155, %rs17, 8;
shl.b32 %r156, %r149, 17;
or.b32 %r157, %r155, %r156;
and.b32 %r158, %r157, 25166208;
or.b32 %r159, %r158, -2139062144;
add.s32 %r160, %r159, -16843009;
or.b32 %r75, %r160, %r158;

prmt.b32 %r75, %r75, 0, 0xba98;

ld.global.nc.v2.u32 {%r161, %r162}, [%rd32];
xor.b32 %r165, %r161, %r73;
or.b32 %r166, %r165, -2139062144;
and.b32 %r167, %r73, 2139062143;
and.b32 %r168, %r161, -2139062144;
sub.s32 %r169, %r166, %r167;
xor.b32 %r170, %r168, %r169;
xor.b32 %r78, %r170, -2139062144;
xor.b32 %r171, %r162, %r75;
or.b32 %r172, %r171, -2139062144;
and.b32 %r173, %r75, 2139062143;
and.b32 %r174, %r162, -2139062144;
sub.s32 %r175, %r172, %r173;
xor.b32 %r176, %r174, %r175;
xor.b32 %r82, %r176, -2139062144;
ld.global.nc.u32 %r83, [%rd30+16];
ld.global.nc.u32 %r79, [%rd30+12];

dp4a.s32.s32 %r77, %r78, %r79, %r69;


dp4a.s32.s32 %r81, %r82, %r83, %r77;

and.b16 %rs18, %rs4, 255;
cvt.u32.u16 %r177, %rs18;
mul.wide.u16 %r178, %rs6, 16;
and.b32 %r179, %r178, 768;
or.b32 %r180, %r179, %r177;
mul.wide.u32 %rd33, %r180, 8;
add.s64 %rd34, %rd27, %rd33;
and.b16 %rs19, %rs8, 255;
mul.wide.u16 %r181, %rs19, 128;
cvt.u32.u16 %r182, %rs8;
shl.b32 %r183, %r182, 21;
or.b32 %r184, %r181, %r183;
and.b32 %r185, %r184, 25166208;
or.b32 %r186, %r185, -2139062144;
add.s32 %r187, %r186, -16843009;
or.b32 %r85, %r187, %r185;

prmt.b32 %r85, %r85, 0, 0xba98;

mul.wide.u16 %r188, %rs19, 8;
shl.b32 %r189, %r182, 17;
or.b32 %r190, %r188, %r189;
and.b32 %r191, %r190, 25166208;
or.b32 %r192, %r191, -2139062144;
add.s32 %r193, %r192, -16843009;
or.b32 %r87, %r193, %r191;

prmt.b32 %r87, %r87, 0, 0xba98;

ld.global.nc.v2.u32 {%r194, %r195}, [%rd34];
xor.b32 %r198, %r194, %r85;
or.b32 %r199, %r198, -2139062144;
and.b32 %r200, %r85, 2139062143;
and.b32 %r201, %r194, -2139062144;
sub.s32 %r202, %r199, %r200;
xor.b32 %r203, %r201, %r202;
xor.b32 %r90, %r203, -2139062144;
xor.b32 %r204, %r195, %r87;
or.b32 %r205, %r204, -2139062144;
and.b32 %r206, %r87, 2139062143;
and.b32 %r207, %r195, -2139062144;
sub.s32 %r208, %r205, %r206;
xor.b32 %r209, %r207, %r208;
xor.b32 %r94, %r209, -2139062144;
ld.global.nc.u32 %r95, [%rd30+24];
ld.global.nc.u32 %r91, [%rd30+20];

dp4a.s32.s32 %r89, %r90, %r91, %r92;


dp4a.s32.s32 %r93, %r94, %r95, %r89;

mul.wide.u16 %r210, %rs6, 4;
and.b32 %r211, %r210, 768;
or.b32 %r212, %r211, %r111;
mul.wide.u32 %rd35, %r212, 8;
add.s64 %rd36, %rd27, %rd35;
shr.u16 %rs20, %rs8, 8;
mul.wide.u16 %r213, %rs20, 128;
cvt.u32.u16 %r214, %rs20;
shl.b32 %r215, %r214, 21;
or.b32 %r216, %r213, %r215;
and.b32 %r217, %r216, 25166208;
or.b32 %r218, %r217, -2139062144;
add.s32 %r219, %r218, -16843009;
or.b32 %r97, %r219, %r217;

prmt.b32 %r97, %r97, 0, 0xba98;

mul.wide.u16 %r220, %rs20, 8;
shl.b32 %r221, %r214, 17;
or.b32 %r222, %r220, %r221;
and.b32 %r223, %r222, 25166208;
or.b32 %r224, %r223, -2139062144;
add.s32 %r225, %r224, -16843009;
or.b32 %r99, %r225, %r223;

prmt.b32 %r99, %r99, 0, 0xba98;

ld.global.nc.v2.u32 {%r226, %r227}, [%rd36];
xor.b32 %r230, %r226, %r97;
or.b32 %r231, %r230, -2139062144;
and.b32 %r232, %r97, 2139062143;
and.b32 %r233, %r226, -2139062144;
sub.s32 %r234, %r231, %r232;
xor.b32 %r235, %r233, %r234;
xor.b32 %r102, %r235, -2139062144;
xor.b32 %r236, %r227, %r99;
or.b32 %r237, %r236, -2139062144;
and.b32 %r238, %r99, 2139062143;
and.b32 %r239, %r227, -2139062144;
sub.s32 %r240, %r237, %r238;
xor.b32 %r241, %r239, %r240;
xor.b32 %r106, %r241, -2139062144;
ld.global.nc.u32 %r107, [%rd30+32];
ld.global.nc.u32 %r103, [%rd30+28];

dp4a.s32.s32 %r101, %r102, %r103, %r93;


dp4a.s32.s32 %r105, %r106, %r107, %r101;

cvt.u32.u16 %r242, %rs11;
cvt.u32.u16 %r243, %rs12;
mul.lo.s32 %r244, %r81, %r242;
mad.lo.s32 %r245, %r105, %r243, %r244;
add.s32 %r246, %r105, %r81;
shr.u32 %r247, %r246, 31;
add.s32 %r248, %r246, %r247;
shr.s32 %r249, %r248, 1;
add.s32 %r250, %r245, %r249;
shr.s32 %r251, %r250, 31;
shr.u32 %r252, %r251, 30;
add.s32 %r253, %r250, %r252;
shr.s32 %r254, %r253, 2;
ld.global.nc.u16 %rs1, [%rd40];

{ cvt.f32.f16 %f7, %rs1;}


ld.global.nc.u32 %r109, [%rd30];

{.reg .f16 low,high;
mov.b32 {low,high},%r109;
cvt.f32.f16 %f8, low;}


mul.ftz.f32 %f9, %f7, %f8;
cvt.rn.f32.s32 %f10, %r254;
fma.rn.ftz.f32 %f27, %f9, %f10, %f27;
add.s64 %rd40, %rd40, 1312;
add.s32 %r286, %r286, 128;
add.s32 %r287, %r287, 16;
setp.lt.s32 %p4, %r287, %r4;
@%p4 bra $L__BB229_5;

$L__BB229_6:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB229_8;

shl.b32 %r255, %r1, 7;
mov.u32 %r256, _ZZ13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r257, %r256, %r255;
shl.b32 %r258, %r2, 2;
add.s32 %r259, %r257, %r258;
st.shared.f32 [%r259+-128], %f27;

$L__BB229_8:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB229_11;

shl.b32 %r260, %r2, 2;
mov.u32 %r261, _ZZ13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r262, %r261, %r260;
mov.u32 %r263, 2;
ld.shared.f32 %f11, [%r262];
add.ftz.f32 %f12, %f27, %f11;
ld.shared.f32 %f13, [%r262+128];
add.ftz.f32 %f14, %f12, %f13;
ld.shared.f32 %f15, [%r262+256];
add.ftz.f32 %f16, %f14, %f15;
mov.b32 %r264, %f16;
mov.u32 %r265, 31;
mov.u32 %r266, 16;
mov.u32 %r267, -1;
shfl.sync.bfly.b32 %r268|%p7, %r264, %r266, %r265, %r267;
mov.b32 %f17, %r268;
add.ftz.f32 %f18, %f16, %f17;
mov.b32 %r269, %f18;
mov.u32 %r270, 8;
shfl.sync.bfly.b32 %r271|%p8, %r269, %r270, %r265, %r267;
mov.b32 %f19, %r271;
add.ftz.f32 %f20, %f18, %f19;
mov.b32 %r272, %f20;
mov.u32 %r273, 4;
shfl.sync.bfly.b32 %r274|%p9, %r272, %r273, %r265, %r267;
mov.b32 %f21, %r274;
add.ftz.f32 %f22, %f20, %f21;
mov.b32 %r275, %f22;
shfl.sync.bfly.b32 %r276|%p10, %r275, %r263, %r265, %r267;
mov.b32 %f23, %r276;
add.ftz.f32 %f4, %f22, %f23;
mov.b32 %r277, %f4;
mov.u32 %r278, 1;
shfl.sync.bfly.b32 %r17|%p1, %r277, %r278, %r265, %r267;
setp.ne.s32 %p11, %r2, 0;
@%p11 bra $L__BB229_11;

mov.u32 %r279, %ctaid.x;
mov.u32 %r280, %ctaid.y;
mad.lo.s32 %r281, %r280, %r27, %r279;
mov.u32 %r282, %ctaid.z;
mad.lo.s32 %r283, %r282, %r33, %r281;
cvta.to.global.u64 %rd37, %rd12;
mul.wide.u32 %rd38, %r283, 4;
add.s64 %rd39, %rd37, %rd38;
mov.b32 %f24, %r17;
add.ftz.f32 %f25, %f4, %f24;
st.global.f32 [%rd39], %f25;

$L__BB229_11:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type22ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type22ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type22ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type22ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type22ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type22ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type22ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type22ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type22ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<35>;
.reg .b16 %rs<43>;
.reg .f32 %f<109>;
.reg .b32 %r<719>;
.reg .b64 %rd<69>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type22ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type22ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type22ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type22ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type22ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r17, [_Z13mul_mat_vec_qIL9ggml_type22ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r18, [_Z13mul_mat_vec_qIL9ggml_type22ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r19, [_Z13mul_mat_vec_qIL9ggml_type22ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type22ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type22ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type22ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type22ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type22ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type22ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type22ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type22ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type22ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type22ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
cvta.to.global.u64 %rd1, %rd18;
mov.u32 %r1, %tid.y;
shl.b32 %r33, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r33, %r2;
mov.u32 %r34, %ctaid.x;
shl.b32 %r4, %r34, 1;
shr.s32 %r35, %r3, 31;
shr.u32 %r36, %r35, 29;
add.s32 %r37, %r3, %r36;
shr.s32 %r718, %r37, 3;
shr.u32 %r6, %r32, 8;
setp.ge.s32 %p3, %r718, %r6;
mov.f32 %f105, 0f00000000;
mov.f32 %f106, %f105;
mov.f32 %f107, %f105;
mov.f32 %f108, %f105;
@%p3 bra $L__BB230_3;

mov.u32 %r38, %ctaid.y;
mov.u32 %r39, %ctaid.z;
mul.lo.s32 %r40, %r38, %r24;
cvt.u64.u32 %rd19, %r40;
mul.lo.s32 %r41, %r39, %r30;
cvt.u64.u32 %rd20, %r41;
add.s64 %rd2, %rd20, %rd19;
mul.hi.u32 %r42, %r39, %r26;
add.s32 %r43, %r42, %r39;
shr.u32 %r44, %r43, %r27;
mul.lo.s32 %r45, %r44, %r29;
mul.hi.u32 %r46, %r38, %r20;
add.s32 %r47, %r46, %r38;
shr.u32 %r48, %r47, %r21;
mul.lo.s32 %r49, %r48, %r23;
cvta.to.global.u64 %rd68, %rd16;
and.b32 %r53, %r37, -8;
sub.s32 %r54, %r3, %r53;
shl.b32 %r55, %r54, 1;
cvt.s64.s32 %rd4, %r54;
shl.b32 %r716, %r718, 3;
add.s32 %r717, %r18, %r716;
add.s32 %r56, %r718, %r45;
add.s32 %r57, %r56, %r49;
add.s32 %r58, %r4, 1;
mad.lo.s32 %r59, %r17, %r58, %r57;
mul.wide.s32 %rd5, %r59, 82;
mad.lo.s32 %r60, %r4, %r17, %r718;
add.s32 %r61, %r60, %r45;
add.s32 %r62, %r61, %r49;
mul.wide.s32 %rd6, %r62, 82;
mul.wide.s32 %rd21, %r55, 2;
add.s64 %rd22, %rd6, %rd21;
add.s64 %rd7, %rd22, 36;
add.s64 %rd23, %rd5, %rd4;
add.s64 %rd8, %rd23, 66;
add.s64 %rd24, %rd6, %rd4;
add.s64 %rd9, %rd24, 74;
add.s64 %rd25, %rd5, %rd21;
add.s64 %rd10, %rd25, 34;
cvta.to.global.u64 %rd11, %rd17;
mov.f32 %f105, 0f00000000;

$L__BB230_2:
cvt.u64.u32 %rd26, %r716;
add.s64 %rd27, %rd2, %rd26;
add.s64 %rd28, %rd27, %rd4;
add.s64 %rd29, %rd68, %rd7;
ld.global.nc.u16 %rs5, [%rd29+-34];
shr.u16 %rs6, %rs5, 8;
cvt.u32.u16 %r259, %rs6;
ld.global.nc.u16 %rs7, [%rd29+-32];
shr.u16 %rs8, %rs7, 8;
cvt.u32.u16 %r260, %rs8;
add.s64 %rd30, %rd68, %rd9;
ld.global.nc.u8 %rs9, [%rd30+-8];
ld.global.nc.u16 %rs10, [%rd29+-2];
ld.global.nc.u16 %rs11, [%rd29];
ld.global.nc.u8 %rs12, [%rd30];
and.b16 %rs13, %rs12, 240;
and.b16 %rs14, %rs12, 15;
shr.u16 %rs15, %rs13, 4;
and.b16 %rs16, %rs5, 255;
cvt.u32.u16 %r261, %rs16;
shl.b16 %rs17, %rs9, 8;
cvt.u32.u16 %r262, %rs17;
and.b32 %r263, %r262, 768;
or.b32 %r264, %r263, %r261;
mul.wide.u32 %rd31, %r264, 8;
mov.u64 %rd32, iq2s_grid;
add.s64 %rd33, %rd32, %rd31;
and.b16 %rs18, %rs10, 255;
mul.wide.u16 %r265, %rs18, 128;
cvt.u32.u16 %r266, %rs10;
shl.b32 %r267, %r266, 21;
or.b32 %r268, %r265, %r267;
and.b32 %r269, %r268, 25166208;
or.b32 %r270, %r269, -2139062144;
add.s32 %r271, %r270, -16843009;
or.b32 %r161, %r271, %r269;
mov.u32 %r63, %r161;

prmt.b32 %r63, %r63, 0, 0xba98;

mul.wide.u16 %r272, %rs18, 8;
shl.b32 %r273, %r266, 17;
or.b32 %r274, %r272, %r273;
and.b32 %r275, %r274, 25166208;
or.b32 %r276, %r275, -2139062144;
add.s32 %r277, %r276, -16843009;
or.b32 %r163, %r277, %r275;
mov.u32 %r65, %r163;

prmt.b32 %r65, %r65, 0, 0xba98;

ld.global.nc.v2.u32 {%r278, %r279}, [%rd33];
not.b32 %r281, %r278;
mov.u32 %r241, 0;
not.b32 %r283, %r279;
xor.b32 %r284, %r278, %r63;
or.b32 %r285, %r284, -2139062144;
and.b32 %r286, %r63, 2139062143;
and.b32 %r287, %r281, -2139062144;
sub.s32 %r288, %r285, %r286;
xor.b32 %r68, %r287, %r288;
xor.b32 %r289, %r279, %r65;
or.b32 %r290, %r289, -2139062144;
and.b32 %r291, %r65, 2139062143;
and.b32 %r292, %r283, -2139062144;
sub.s32 %r293, %r290, %r291;
xor.b32 %r72, %r292, %r293;
mul.lo.s64 %rd34, %rd28, 36;
add.s64 %rd35, %rd11, %rd34;
ld.global.nc.u32 %r73, [%rd35+8];
ld.global.nc.u32 %r69, [%rd35+4];

dp4a.s32.s32 %r67, %r68, %r69, %r241;


dp4a.s32.s32 %r71, %r72, %r73, %r67;

mul.wide.u16 %r294, %rs9, 64;
and.b32 %r295, %r294, 768;
or.b32 %r296, %r295, %r259;
mul.wide.u32 %rd36, %r296, 8;
add.s64 %rd37, %rd32, %rd36;
shr.u16 %rs20, %rs10, 8;
mul.wide.u16 %r297, %rs20, 128;
cvt.u32.u16 %r298, %rs20;
shl.b32 %r299, %r298, 21;
or.b32 %r300, %r297, %r299;
and.b32 %r301, %r300, 25166208;
or.b32 %r302, %r301, -2139062144;
add.s32 %r303, %r302, -16843009;
or.b32 %r173, %r303, %r301;
mov.u32 %r75, %r173;

prmt.b32 %r75, %r75, 0, 0xba98;

mul.wide.u16 %r304, %rs20, 8;
shl.b32 %r305, %r298, 17;
or.b32 %r306, %r304, %r305;
and.b32 %r307, %r306, 25166208;
or.b32 %r308, %r307, -2139062144;
add.s32 %r309, %r308, -16843009;
or.b32 %r175, %r309, %r307;
mov.u32 %r77, %r175;

prmt.b32 %r77, %r77, 0, 0xba98;

ld.global.nc.v2.u32 {%r310, %r311}, [%rd37];
not.b32 %r313, %r310;
not.b32 %r315, %r311;
xor.b32 %r316, %r310, %r75;
or.b32 %r317, %r316, -2139062144;
and.b32 %r318, %r75, 2139062143;
and.b32 %r319, %r313, -2139062144;
sub.s32 %r320, %r317, %r318;
xor.b32 %r80, %r319, %r320;
xor.b32 %r321, %r311, %r77;
or.b32 %r322, %r321, -2139062144;
and.b32 %r323, %r77, 2139062143;
and.b32 %r324, %r315, -2139062144;
sub.s32 %r325, %r322, %r323;
xor.b32 %r84, %r324, %r325;
ld.global.nc.u32 %r85, [%rd35+16];
ld.global.nc.u32 %r81, [%rd35+12];

dp4a.s32.s32 %r79, %r80, %r81, %r71;


dp4a.s32.s32 %r83, %r84, %r85, %r79;

and.b16 %rs21, %rs7, 255;
cvt.u32.u16 %r326, %rs21;
mul.wide.u16 %r327, %rs9, 16;
and.b32 %r328, %r327, 768;
or.b32 %r329, %r328, %r326;
mul.wide.u32 %rd38, %r329, 8;
add.s64 %rd39, %rd32, %rd38;
and.b16 %rs22, %rs11, 255;
mul.wide.u16 %r330, %rs22, 128;
cvt.u32.u16 %r331, %rs11;
shl.b32 %r332, %r331, 21;
or.b32 %r333, %r330, %r332;
and.b32 %r334, %r333, 25166208;
or.b32 %r335, %r334, -2139062144;
add.s32 %r336, %r335, -16843009;
or.b32 %r185, %r336, %r334;
mov.u32 %r87, %r185;

prmt.b32 %r87, %r87, 0, 0xba98;

mul.wide.u16 %r337, %rs22, 8;
shl.b32 %r338, %r331, 17;
or.b32 %r339, %r337, %r338;
and.b32 %r340, %r339, 25166208;
or.b32 %r341, %r340, -2139062144;
add.s32 %r342, %r341, -16843009;
or.b32 %r187, %r342, %r340;
mov.u32 %r89, %r187;

prmt.b32 %r89, %r89, 0, 0xba98;

ld.global.nc.v2.u32 {%r343, %r344}, [%rd39];
not.b32 %r346, %r343;
not.b32 %r348, %r344;
xor.b32 %r349, %r343, %r87;
or.b32 %r350, %r349, -2139062144;
and.b32 %r351, %r87, 2139062143;
and.b32 %r352, %r346, -2139062144;
sub.s32 %r353, %r350, %r351;
xor.b32 %r92, %r352, %r353;
xor.b32 %r354, %r344, %r89;
or.b32 %r355, %r354, -2139062144;
and.b32 %r356, %r89, 2139062143;
and.b32 %r357, %r348, -2139062144;
sub.s32 %r358, %r355, %r356;
xor.b32 %r96, %r357, %r358;
ld.global.nc.u32 %r97, [%rd35+24];
ld.global.nc.u32 %r93, [%rd35+20];

dp4a.s32.s32 %r91, %r92, %r93, %r241;


dp4a.s32.s32 %r95, %r96, %r97, %r91;

mul.wide.u16 %r359, %rs9, 4;
and.b32 %r360, %r359, 768;
or.b32 %r361, %r360, %r260;
mul.wide.u32 %rd40, %r361, 8;
add.s64 %rd41, %rd32, %rd40;
shr.u16 %rs23, %rs11, 8;
mul.wide.u16 %r362, %rs23, 128;
cvt.u32.u16 %r363, %rs23;
shl.b32 %r364, %r363, 21;
or.b32 %r365, %r362, %r364;
and.b32 %r366, %r365, 25166208;
or.b32 %r367, %r366, -2139062144;
add.s32 %r368, %r367, -16843009;
or.b32 %r197, %r368, %r366;
mov.u32 %r99, %r197;

prmt.b32 %r99, %r99, 0, 0xba98;

mul.wide.u16 %r369, %rs23, 8;
shl.b32 %r370, %r363, 17;
or.b32 %r371, %r369, %r370;
and.b32 %r372, %r371, 25166208;
or.b32 %r373, %r372, -2139062144;
add.s32 %r374, %r373, -16843009;
or.b32 %r199, %r374, %r372;
mov.u32 %r101, %r199;

prmt.b32 %r101, %r101, 0, 0xba98;

ld.global.nc.v2.u32 {%r375, %r376}, [%rd41];
not.b32 %r378, %r375;
not.b32 %r380, %r376;
xor.b32 %r381, %r375, %r99;
or.b32 %r382, %r381, -2139062144;
and.b32 %r383, %r99, 2139062143;
and.b32 %r384, %r378, -2139062144;
sub.s32 %r385, %r382, %r383;
xor.b32 %r104, %r384, %r385;
xor.b32 %r386, %r376, %r101;
or.b32 %r387, %r386, -2139062144;
and.b32 %r388, %r101, 2139062143;
and.b32 %r389, %r380, -2139062144;
sub.s32 %r390, %r387, %r388;
xor.b32 %r108, %r389, %r390;
ld.global.nc.u32 %r109, [%rd35+32];
ld.global.nc.u32 %r105, [%rd35+28];

dp4a.s32.s32 %r103, %r104, %r105, %r95;


dp4a.s32.s32 %r107, %r108, %r109, %r103;

cvt.u32.u16 %r391, %rs14;
cvt.u32.u16 %r392, %rs15;
mul.lo.s32 %r393, %r83, %r391;
mad.lo.s32 %r394, %r107, %r392, %r393;
add.s32 %r395, %r107, %r83;
shr.u32 %r396, %r395, 31;
add.s32 %r397, %r395, %r396;
shr.s32 %r398, %r397, 1;
add.s32 %r399, %r394, %r398;
shr.s32 %r400, %r399, 31;
shr.u32 %r401, %r400, 30;
add.s32 %r402, %r399, %r401;
shr.s32 %r403, %r402, 2;
add.s64 %rd42, %rd68, %rd6;
ld.global.nc.u16 %rs1, [%rd42];

{ cvt.f32.f16 %f25, %rs1;}


ld.global.nc.u32 %r111, [%rd35];

{.reg .f16 low,high;
mov.b32 {low,high},%r111;
cvt.f32.f16 %f26, low;}


mul.ftz.f32 %f33, %f25, %f26;
cvt.rn.f32.s32 %f34, %r403;
fma.rn.ftz.f32 %f108, %f33, %f34, %f108;
add.s64 %rd43, %rd68, %rd10;
ld.global.nc.u16 %rs24, [%rd43+-32];
shr.u16 %rs25, %rs24, 8;
cvt.u32.u16 %r404, %rs25;
ld.global.nc.u16 %rs26, [%rd43+-30];
shr.u16 %rs27, %rs26, 8;
cvt.u32.u16 %r405, %rs27;
add.s64 %rd44, %rd68, %rd8;
ld.global.nc.u8 %rs28, [%rd44];
ld.global.nc.u16 %rs29, [%rd43];
ld.global.nc.u16 %rs30, [%rd43+2];
ld.global.nc.u8 %rs31, [%rd44+8];
and.b16 %rs32, %rs31, 240;
and.b16 %rs33, %rs31, 15;
shr.u16 %rs34, %rs32, 4;
and.b16 %rs35, %rs24, 255;
cvt.u32.u16 %r406, %rs35;
shl.b16 %rs36, %rs28, 8;
cvt.u32.u16 %r407, %rs36;
and.b32 %r408, %r407, 768;
or.b32 %r409, %r408, %r406;
mul.wide.u32 %rd45, %r409, 8;
add.s64 %rd46, %rd32, %rd45;
and.b16 %rs37, %rs29, 255;
mul.wide.u16 %r410, %rs37, 128;
cvt.u32.u16 %r411, %rs29;
shl.b32 %r412, %r411, 21;
or.b32 %r413, %r410, %r412;
and.b32 %r414, %r413, 25166208;
or.b32 %r415, %r414, -2139062144;
add.s32 %r416, %r415, -16843009;
or.b32 %r210, %r416, %r414;
mov.u32 %r112, %r210;

prmt.b32 %r112, %r112, 0, 0xba98;

mul.wide.u16 %r417, %rs37, 8;
shl.b32 %r418, %r411, 17;
or.b32 %r419, %r417, %r418;
and.b32 %r420, %r419, 25166208;
or.b32 %r421, %r420, -2139062144;
add.s32 %r422, %r421, -16843009;
or.b32 %r212, %r422, %r420;
mov.u32 %r114, %r212;

prmt.b32 %r114, %r114, 0, 0xba98;

ld.global.nc.v2.u32 {%r423, %r424}, [%rd46];
not.b32 %r426, %r423;
not.b32 %r428, %r424;
xor.b32 %r429, %r423, %r112;
or.b32 %r430, %r429, -2139062144;
and.b32 %r431, %r112, 2139062143;
and.b32 %r432, %r426, -2139062144;
sub.s32 %r433, %r430, %r431;
xor.b32 %r117, %r432, %r433;
xor.b32 %r434, %r424, %r114;
or.b32 %r435, %r434, -2139062144;
and.b32 %r436, %r114, 2139062143;
and.b32 %r437, %r428, -2139062144;
sub.s32 %r438, %r435, %r436;
xor.b32 %r121, %r437, %r438;

dp4a.s32.s32 %r116, %r117, %r69, %r241;


dp4a.s32.s32 %r120, %r121, %r73, %r116;

mul.wide.u16 %r439, %rs28, 64;
and.b32 %r440, %r439, 768;
or.b32 %r441, %r440, %r404;
mul.wide.u32 %rd47, %r441, 8;
add.s64 %rd48, %rd32, %rd47;
shr.u16 %rs39, %rs29, 8;
mul.wide.u16 %r442, %rs39, 128;
cvt.u32.u16 %r443, %rs39;
shl.b32 %r444, %r443, 21;
or.b32 %r445, %r442, %r444;
and.b32 %r446, %r445, 25166208;
or.b32 %r447, %r446, -2139062144;
add.s32 %r448, %r447, -16843009;
or.b32 %r222, %r448, %r446;
mov.u32 %r124, %r222;

prmt.b32 %r124, %r124, 0, 0xba98;

mul.wide.u16 %r449, %rs39, 8;
shl.b32 %r450, %r443, 17;
or.b32 %r451, %r449, %r450;
and.b32 %r452, %r451, 25166208;
or.b32 %r453, %r452, -2139062144;
add.s32 %r454, %r453, -16843009;
or.b32 %r224, %r454, %r452;
mov.u32 %r126, %r224;

prmt.b32 %r126, %r126, 0, 0xba98;

ld.global.nc.v2.u32 {%r455, %r456}, [%rd48];
not.b32 %r458, %r455;
not.b32 %r460, %r456;
xor.b32 %r461, %r455, %r124;
or.b32 %r462, %r461, -2139062144;
and.b32 %r463, %r124, 2139062143;
and.b32 %r464, %r458, -2139062144;
sub.s32 %r465, %r462, %r463;
xor.b32 %r129, %r464, %r465;
xor.b32 %r466, %r456, %r126;
or.b32 %r467, %r466, -2139062144;
and.b32 %r468, %r126, 2139062143;
and.b32 %r469, %r460, -2139062144;
sub.s32 %r470, %r467, %r468;
xor.b32 %r133, %r469, %r470;

dp4a.s32.s32 %r128, %r129, %r81, %r120;


dp4a.s32.s32 %r132, %r133, %r85, %r128;

and.b16 %rs40, %rs26, 255;
cvt.u32.u16 %r471, %rs40;
mul.wide.u16 %r472, %rs28, 16;
and.b32 %r473, %r472, 768;
or.b32 %r474, %r473, %r471;
mul.wide.u32 %rd49, %r474, 8;
add.s64 %rd50, %rd32, %rd49;
and.b16 %rs41, %rs30, 255;
mul.wide.u16 %r475, %rs41, 128;
cvt.u32.u16 %r476, %rs30;
shl.b32 %r477, %r476, 21;
or.b32 %r478, %r475, %r477;
and.b32 %r479, %r478, 25166208;
or.b32 %r480, %r479, -2139062144;
add.s32 %r481, %r480, -16843009;
or.b32 %r234, %r481, %r479;
mov.u32 %r136, %r234;

prmt.b32 %r136, %r136, 0, 0xba98;

mul.wide.u16 %r482, %rs41, 8;
shl.b32 %r483, %r476, 17;
or.b32 %r484, %r482, %r483;
and.b32 %r485, %r484, 25166208;
or.b32 %r486, %r485, -2139062144;
add.s32 %r487, %r486, -16843009;
or.b32 %r236, %r487, %r485;
mov.u32 %r138, %r236;

prmt.b32 %r138, %r138, 0, 0xba98;

ld.global.nc.v2.u32 {%r488, %r489}, [%rd50];
not.b32 %r491, %r488;
not.b32 %r493, %r489;
xor.b32 %r494, %r488, %r136;
or.b32 %r495, %r494, -2139062144;
and.b32 %r496, %r136, 2139062143;
and.b32 %r497, %r491, -2139062144;
sub.s32 %r498, %r495, %r496;
xor.b32 %r141, %r497, %r498;
xor.b32 %r499, %r489, %r138;
or.b32 %r500, %r499, -2139062144;
and.b32 %r501, %r138, 2139062143;
and.b32 %r502, %r493, -2139062144;
sub.s32 %r503, %r500, %r501;
xor.b32 %r145, %r502, %r503;

dp4a.s32.s32 %r140, %r141, %r93, %r241;


dp4a.s32.s32 %r144, %r145, %r97, %r140;

mul.wide.u16 %r504, %rs28, 4;
and.b32 %r505, %r504, 768;
or.b32 %r506, %r505, %r405;
mul.wide.u32 %rd51, %r506, 8;
add.s64 %rd52, %rd32, %rd51;
shr.u16 %rs42, %rs30, 8;
mul.wide.u16 %r507, %rs42, 128;
cvt.u32.u16 %r508, %rs42;
shl.b32 %r509, %r508, 21;
or.b32 %r510, %r507, %r509;
and.b32 %r511, %r510, 25166208;
or.b32 %r512, %r511, -2139062144;
add.s32 %r513, %r512, -16843009;
or.b32 %r246, %r513, %r511;
mov.u32 %r148, %r246;

prmt.b32 %r148, %r148, 0, 0xba98;

mul.wide.u16 %r514, %rs42, 8;
shl.b32 %r515, %r508, 17;
or.b32 %r516, %r514, %r515;
and.b32 %r517, %r516, 25166208;
or.b32 %r518, %r517, -2139062144;
add.s32 %r519, %r518, -16843009;
or.b32 %r248, %r519, %r517;
mov.u32 %r150, %r248;

prmt.b32 %r150, %r150, 0, 0xba98;

ld.global.nc.v2.u32 {%r520, %r521}, [%rd52];
not.b32 %r523, %r520;
not.b32 %r525, %r521;
xor.b32 %r526, %r520, %r148;
or.b32 %r527, %r526, -2139062144;
and.b32 %r528, %r148, 2139062143;
and.b32 %r529, %r523, -2139062144;
sub.s32 %r530, %r527, %r528;
xor.b32 %r153, %r529, %r530;
xor.b32 %r531, %r521, %r150;
or.b32 %r532, %r531, -2139062144;
and.b32 %r533, %r150, 2139062143;
and.b32 %r534, %r525, -2139062144;
sub.s32 %r535, %r532, %r533;
xor.b32 %r157, %r534, %r535;

dp4a.s32.s32 %r152, %r153, %r105, %r144;


dp4a.s32.s32 %r156, %r157, %r109, %r152;

cvt.u32.u16 %r536, %rs33;
cvt.u32.u16 %r537, %rs34;
mul.lo.s32 %r538, %r132, %r536;
mad.lo.s32 %r539, %r156, %r537, %r538;
add.s32 %r540, %r156, %r132;
shr.u32 %r541, %r540, 31;
add.s32 %r542, %r540, %r541;
shr.s32 %r543, %r542, 1;
add.s32 %r544, %r539, %r543;
shr.s32 %r545, %r544, 31;
shr.u32 %r546, %r545, 30;
add.s32 %r547, %r544, %r546;
shr.s32 %r548, %r547, 2;
add.s64 %rd53, %rd68, %rd5;
ld.global.nc.u16 %rs2, [%rd53];

{ cvt.f32.f16 %f27, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r111;
cvt.f32.f16 %f28, low;}


mul.ftz.f32 %f35, %f27, %f28;
cvt.rn.f32.s32 %f36, %r548;
fma.rn.ftz.f32 %f107, %f35, %f36, %f107;
cvt.u64.u32 %rd54, %r717;
add.s64 %rd55, %rd2, %rd54;
add.s64 %rd56, %rd55, %rd4;

prmt.b32 %r161, %r161, 0, 0xba98;


prmt.b32 %r163, %r163, 0, 0xba98;

xor.b32 %r549, %r278, %r161;
or.b32 %r550, %r549, -2139062144;
and.b32 %r551, %r161, 2139062143;
sub.s32 %r552, %r550, %r551;
xor.b32 %r166, %r287, %r552;
xor.b32 %r553, %r279, %r163;
or.b32 %r554, %r553, -2139062144;
and.b32 %r555, %r163, 2139062143;
sub.s32 %r556, %r554, %r555;
xor.b32 %r170, %r292, %r556;
mul.lo.s64 %rd57, %rd56, 36;
add.s64 %rd58, %rd11, %rd57;
ld.global.nc.u32 %r171, [%rd58+8];
ld.global.nc.u32 %r167, [%rd58+4];

dp4a.s32.s32 %r165, %r166, %r167, %r241;


dp4a.s32.s32 %r169, %r170, %r171, %r165;


prmt.b32 %r173, %r173, 0, 0xba98;


prmt.b32 %r175, %r175, 0, 0xba98;

xor.b32 %r557, %r310, %r173;
or.b32 %r558, %r557, -2139062144;
and.b32 %r559, %r173, 2139062143;
sub.s32 %r560, %r558, %r559;
xor.b32 %r178, %r319, %r560;
xor.b32 %r561, %r311, %r175;
or.b32 %r562, %r561, -2139062144;
and.b32 %r563, %r175, 2139062143;
sub.s32 %r564, %r562, %r563;
xor.b32 %r182, %r324, %r564;
ld.global.nc.u32 %r183, [%rd58+16];
ld.global.nc.u32 %r179, [%rd58+12];

dp4a.s32.s32 %r177, %r178, %r179, %r169;


dp4a.s32.s32 %r181, %r182, %r183, %r177;


prmt.b32 %r185, %r185, 0, 0xba98;


prmt.b32 %r187, %r187, 0, 0xba98;

xor.b32 %r565, %r343, %r185;
or.b32 %r566, %r565, -2139062144;
and.b32 %r567, %r185, 2139062143;
sub.s32 %r568, %r566, %r567;
xor.b32 %r190, %r352, %r568;
xor.b32 %r569, %r344, %r187;
or.b32 %r570, %r569, -2139062144;
and.b32 %r571, %r187, 2139062143;
sub.s32 %r572, %r570, %r571;
xor.b32 %r194, %r357, %r572;
ld.global.nc.u32 %r195, [%rd58+24];
ld.global.nc.u32 %r191, [%rd58+20];

dp4a.s32.s32 %r189, %r190, %r191, %r241;


dp4a.s32.s32 %r193, %r194, %r195, %r189;


prmt.b32 %r197, %r197, 0, 0xba98;


prmt.b32 %r199, %r199, 0, 0xba98;

xor.b32 %r573, %r375, %r197;
or.b32 %r574, %r573, -2139062144;
and.b32 %r575, %r197, 2139062143;
sub.s32 %r576, %r574, %r575;
xor.b32 %r202, %r384, %r576;
xor.b32 %r577, %r376, %r199;
or.b32 %r578, %r577, -2139062144;
and.b32 %r579, %r199, 2139062143;
sub.s32 %r580, %r578, %r579;
xor.b32 %r206, %r389, %r580;
ld.global.nc.u32 %r207, [%rd58+32];
ld.global.nc.u32 %r203, [%rd58+28];

dp4a.s32.s32 %r201, %r202, %r203, %r193;


dp4a.s32.s32 %r205, %r206, %r207, %r201;

mul.lo.s32 %r581, %r181, %r391;
mad.lo.s32 %r582, %r205, %r392, %r581;
add.s32 %r583, %r205, %r181;
shr.u32 %r584, %r583, 31;
add.s32 %r585, %r583, %r584;
shr.s32 %r586, %r585, 1;
add.s32 %r587, %r582, %r586;
shr.s32 %r588, %r587, 31;
shr.u32 %r589, %r588, 30;
add.s32 %r590, %r587, %r589;
shr.s32 %r591, %r590, 2;

{ cvt.f32.f16 %f29, %rs1;}


ld.global.nc.u32 %r209, [%rd58];

{.reg .f16 low,high;
mov.b32 {low,high},%r209;
cvt.f32.f16 %f30, low;}


mul.ftz.f32 %f37, %f29, %f30;
cvt.rn.f32.s32 %f38, %r591;
fma.rn.ftz.f32 %f106, %f37, %f38, %f106;

prmt.b32 %r210, %r210, 0, 0xba98;


prmt.b32 %r212, %r212, 0, 0xba98;

xor.b32 %r592, %r423, %r210;
or.b32 %r593, %r592, -2139062144;
and.b32 %r594, %r210, 2139062143;
sub.s32 %r595, %r593, %r594;
xor.b32 %r215, %r432, %r595;
xor.b32 %r596, %r424, %r212;
or.b32 %r597, %r596, -2139062144;
and.b32 %r598, %r212, 2139062143;
sub.s32 %r599, %r597, %r598;
xor.b32 %r219, %r437, %r599;

dp4a.s32.s32 %r214, %r215, %r167, %r241;


dp4a.s32.s32 %r218, %r219, %r171, %r214;


prmt.b32 %r222, %r222, 0, 0xba98;


prmt.b32 %r224, %r224, 0, 0xba98;

xor.b32 %r600, %r455, %r222;
or.b32 %r601, %r600, -2139062144;
and.b32 %r602, %r222, 2139062143;
sub.s32 %r603, %r601, %r602;
xor.b32 %r227, %r464, %r603;
xor.b32 %r604, %r456, %r224;
or.b32 %r605, %r604, -2139062144;
and.b32 %r606, %r224, 2139062143;
sub.s32 %r607, %r605, %r606;
xor.b32 %r231, %r469, %r607;

dp4a.s32.s32 %r226, %r227, %r179, %r218;


dp4a.s32.s32 %r230, %r231, %r183, %r226;


prmt.b32 %r234, %r234, 0, 0xba98;


prmt.b32 %r236, %r236, 0, 0xba98;

xor.b32 %r608, %r488, %r234;
or.b32 %r609, %r608, -2139062144;
and.b32 %r610, %r234, 2139062143;
sub.s32 %r611, %r609, %r610;
xor.b32 %r239, %r497, %r611;
xor.b32 %r612, %r489, %r236;
or.b32 %r613, %r612, -2139062144;
and.b32 %r614, %r236, 2139062143;
sub.s32 %r615, %r613, %r614;
xor.b32 %r243, %r502, %r615;

dp4a.s32.s32 %r238, %r239, %r191, %r241;


dp4a.s32.s32 %r242, %r243, %r195, %r238;


prmt.b32 %r246, %r246, 0, 0xba98;


prmt.b32 %r248, %r248, 0, 0xba98;

xor.b32 %r616, %r520, %r246;
or.b32 %r617, %r616, -2139062144;
and.b32 %r618, %r246, 2139062143;
sub.s32 %r619, %r617, %r618;
xor.b32 %r251, %r529, %r619;
xor.b32 %r620, %r521, %r248;
or.b32 %r621, %r620, -2139062144;
and.b32 %r622, %r248, 2139062143;
sub.s32 %r623, %r621, %r622;
xor.b32 %r255, %r534, %r623;

dp4a.s32.s32 %r250, %r251, %r203, %r242;


dp4a.s32.s32 %r254, %r255, %r207, %r250;

mul.lo.s32 %r624, %r230, %r536;
mad.lo.s32 %r625, %r254, %r537, %r624;
add.s32 %r626, %r254, %r230;
shr.u32 %r627, %r626, 31;
add.s32 %r628, %r626, %r627;
shr.s32 %r629, %r628, 1;
add.s32 %r630, %r625, %r629;
shr.s32 %r631, %r630, 31;
shr.u32 %r632, %r631, 30;
add.s32 %r633, %r630, %r632;
shr.s32 %r634, %r633, 2;

{ cvt.f32.f16 %f31, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r209;
cvt.f32.f16 %f32, low;}


mul.ftz.f32 %f39, %f31, %f32;
cvt.rn.f32.s32 %f40, %r634;
fma.rn.ftz.f32 %f105, %f39, %f40, %f105;
add.s32 %r717, %r717, 128;
add.s32 %r716, %r716, 128;
add.s64 %rd68, %rd68, 1312;
add.s32 %r718, %r718, 16;
setp.lt.s32 %p4, %r718, %r6;
@%p4 bra $L__BB230_2;

$L__BB230_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB230_5;

shl.b32 %r635, %r1, 9;
mov.u32 %r636, _ZZ13mul_mat_vec_qIL9ggml_type22ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r637, %r636, %r635;
shl.b32 %r638, %r2, 2;
add.s32 %r639, %r637, %r638;
st.shared.f32 [%r639+-512], %f108;
st.shared.f32 [%r639+-384], %f107;
st.shared.f32 [%r639+-256], %f106;
st.shared.f32 [%r639+-128], %f105;

$L__BB230_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB230_14;

mov.u32 %r640, %ctaid.z;
mov.u32 %r641, %ctaid.y;
mad.lo.s32 %r642, %r641, %r25, %r4;
mad.lo.s32 %r643, %r640, %r31, %r642;
cvt.u64.u32 %rd14, %r643;
shl.b32 %r644, %r2, 2;
mov.u32 %r645, _ZZ13mul_mat_vec_qIL9ggml_type22ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r15, %r645, %r644;
mov.u32 %r646, 2;
ld.shared.f32 %f41, [%r15];
add.ftz.f32 %f42, %f41, %f108;
ld.shared.f32 %f43, [%r15+512];
add.ftz.f32 %f44, %f43, %f42;
ld.shared.f32 %f45, [%r15+1024];
add.ftz.f32 %f46, %f45, %f44;
mov.b32 %r647, %f46;
mov.u32 %r648, 31;
mov.u32 %r649, 16;
mov.u32 %r650, -1;
shfl.sync.bfly.b32 %r651|%p7, %r647, %r649, %r648, %r650;
mov.b32 %f47, %r651;
add.ftz.f32 %f48, %f46, %f47;
mov.b32 %r652, %f48;
mov.u32 %r653, 8;
shfl.sync.bfly.b32 %r654|%p8, %r652, %r653, %r648, %r650;
mov.b32 %f49, %r654;
add.ftz.f32 %f50, %f48, %f49;
mov.b32 %r655, %f50;
mov.u32 %r656, 4;
shfl.sync.bfly.b32 %r657|%p9, %r655, %r656, %r648, %r650;
mov.b32 %f51, %r657;
add.ftz.f32 %f52, %f50, %f51;
mov.b32 %r658, %f52;
shfl.sync.bfly.b32 %r659|%p10, %r658, %r646, %r648, %r650;
mov.b32 %f53, %r659;
add.ftz.f32 %f54, %f52, %f53;
mov.b32 %r660, %f54;
mov.u32 %r661, 1;
shfl.sync.bfly.b32 %r662|%p11, %r660, %r661, %r648, %r650;
mov.b32 %f55, %r662;
add.ftz.f32 %f13, %f54, %f55;
setp.lt.u32 %p12, %r4, %r19;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
mul.wide.u32 %rd59, %r643, 4;
add.s64 %rd15, %rd1, %rd59;
not.pred %p14, %p1;
@%p14 bra $L__BB230_8;

st.global.f32 [%rd15], %f13;

$L__BB230_8:
ld.shared.f32 %f56, [%r15+128];
add.ftz.f32 %f57, %f56, %f107;
ld.shared.f32 %f58, [%r15+640];
add.ftz.f32 %f59, %f58, %f57;
ld.shared.f32 %f60, [%r15+1152];
add.ftz.f32 %f61, %f60, %f59;
mov.b32 %r663, %f61;
shfl.sync.bfly.b32 %r667|%p15, %r663, %r649, %r648, %r650;
mov.b32 %f62, %r667;
add.ftz.f32 %f63, %f61, %f62;
mov.b32 %r668, %f63;
shfl.sync.bfly.b32 %r670|%p16, %r668, %r653, %r648, %r650;
mov.b32 %f64, %r670;
add.ftz.f32 %f65, %f63, %f64;
mov.b32 %r671, %f65;
shfl.sync.bfly.b32 %r673|%p17, %r671, %r656, %r648, %r650;
mov.b32 %f66, %r673;
add.ftz.f32 %f67, %f65, %f66;
mov.b32 %r674, %f67;
shfl.sync.bfly.b32 %r676|%p18, %r674, %r646, %r648, %r650;
mov.b32 %f68, %r676;
add.ftz.f32 %f69, %f67, %f68;
mov.b32 %r677, %f69;
shfl.sync.bfly.b32 %r679|%p19, %r677, %r661, %r648, %r650;
mov.b32 %f70, %r679;
add.ftz.f32 %f14, %f69, %f70;
add.s32 %r680, %r4, 1;
setp.lt.u32 %p20, %r680, %r19;
setp.eq.s32 %p21, %r2, 1;
and.pred %p2, %p21, %p20;
not.pred %p22, %p2;
@%p22 bra $L__BB230_10;

st.global.f32 [%rd15+4], %f14;

$L__BB230_10:
ld.shared.f32 %f71, [%r15+256];
add.ftz.f32 %f72, %f71, %f106;
ld.shared.f32 %f73, [%r15+768];
add.ftz.f32 %f74, %f73, %f72;
ld.shared.f32 %f75, [%r15+1280];
add.ftz.f32 %f76, %f75, %f74;
mov.b32 %r681, %f76;
mov.u32 %r682, 31;
mov.u32 %r683, 16;
mov.u32 %r684, -1;
shfl.sync.bfly.b32 %r685|%p23, %r681, %r683, %r682, %r684;
mov.b32 %f77, %r685;
add.ftz.f32 %f78, %f76, %f77;
mov.b32 %r686, %f78;
mov.u32 %r687, 8;
shfl.sync.bfly.b32 %r688|%p24, %r686, %r687, %r682, %r684;
mov.b32 %f79, %r688;
add.ftz.f32 %f80, %f78, %f79;
mov.b32 %r689, %f80;
mov.u32 %r690, 4;
shfl.sync.bfly.b32 %r691|%p25, %r689, %r690, %r682, %r684;
mov.b32 %f81, %r691;
add.ftz.f32 %f82, %f80, %f81;
mov.b32 %r692, %f82;
mov.u32 %r693, 2;
shfl.sync.bfly.b32 %r694|%p26, %r692, %r693, %r682, %r684;
mov.b32 %f83, %r694;
add.ftz.f32 %f84, %f82, %f83;
mov.b32 %r695, %f84;
mov.u32 %r696, 1;
shfl.sync.bfly.b32 %r697|%p27, %r695, %r696, %r682, %r684;
mov.b32 %f85, %r697;
add.ftz.f32 %f15, %f84, %f85;
@%p14 bra $L__BB230_12;

cvt.u64.u32 %rd60, %r19;
add.s64 %rd61, %rd60, %rd14;
shl.b64 %rd62, %rd61, 2;
add.s64 %rd63, %rd1, %rd62;
st.global.f32 [%rd63], %f15;

$L__BB230_12:
ld.shared.f32 %f86, [%r15+384];
add.ftz.f32 %f87, %f86, %f105;
ld.shared.f32 %f88, [%r15+896];
add.ftz.f32 %f89, %f88, %f87;
ld.shared.f32 %f90, [%r15+1408];
add.ftz.f32 %f91, %f90, %f89;
mov.b32 %r698, %f91;
shfl.sync.bfly.b32 %r702|%p29, %r698, %r683, %r682, %r684;
mov.b32 %f92, %r702;
add.ftz.f32 %f93, %f91, %f92;
mov.b32 %r703, %f93;
shfl.sync.bfly.b32 %r705|%p30, %r703, %r687, %r682, %r684;
mov.b32 %f94, %r705;
add.ftz.f32 %f95, %f93, %f94;
mov.b32 %r706, %f95;
shfl.sync.bfly.b32 %r708|%p31, %r706, %r690, %r682, %r684;
mov.b32 %f96, %r708;
add.ftz.f32 %f97, %f95, %f96;
mov.b32 %r709, %f97;
shfl.sync.bfly.b32 %r711|%p32, %r709, %r693, %r682, %r684;
mov.b32 %f98, %r711;
add.ftz.f32 %f99, %f97, %f98;
mov.b32 %r712, %f99;
shfl.sync.bfly.b32 %r714|%p33, %r712, %r696, %r682, %r684;
mov.b32 %f100, %r714;
add.ftz.f32 %f16, %f99, %f100;
@%p22 bra $L__BB230_14;

add.s32 %r715, %r19, 1;
cvt.u64.u32 %rd64, %r715;
add.s64 %rd65, %rd64, %rd14;
shl.b64 %rd66, %rd65, 2;
add.s64 %rd67, %rd1, %rd66;
st.global.f32 [%rd67], %f16;

$L__BB230_14:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type22ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type22ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type22ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type22ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type22ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type22ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type22ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type22ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type22ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<47>;
.reg .b16 %rs<45>;
.reg .f32 %f<163>;
.reg .b32 %r<942>;
.reg .b64 %rd<78>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type22ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[2304];

ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type22ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type22ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type22ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type22ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type22ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type22ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type22ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type22ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type22ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type22ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type22ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type22ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type22ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type22ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type22ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type22ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type22ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
cvta.to.global.u64 %rd1, %rd19;
mov.u32 %r1, %tid.y;
shl.b32 %r36, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r36, %r2;
mov.u32 %r37, %ctaid.x;
shl.b32 %r4, %r37, 1;
shr.s32 %r38, %r3, 31;
shr.u32 %r39, %r38, 29;
add.s32 %r40, %r3, %r39;
shr.s32 %r941, %r40, 3;
shr.u32 %r6, %r35, 8;
setp.ge.s32 %p3, %r941, %r6;
mov.f32 %f157, 0f00000000;
mov.f32 %f158, %f157;
mov.f32 %f159, %f157;
mov.f32 %f160, %f157;
mov.f32 %f161, %f157;
mov.f32 %f162, %f157;
@%p3 bra $L__BB231_3;

mov.u32 %r41, %ctaid.y;
mov.u32 %r42, %ctaid.z;
mul.lo.s32 %r43, %r41, %r27;
cvt.u64.u32 %rd20, %r43;
mul.lo.s32 %r44, %r42, %r33;
cvt.u64.u32 %rd21, %r44;
add.s64 %rd2, %rd21, %rd20;
mul.hi.u32 %r45, %r42, %r29;
add.s32 %r46, %r45, %r42;
shr.u32 %r47, %r46, %r30;
mul.lo.s32 %r48, %r47, %r32;
mul.hi.u32 %r49, %r41, %r23;
add.s32 %r50, %r49, %r41;
shr.u32 %r51, %r50, %r24;
mul.lo.s32 %r52, %r51, %r26;
cvta.to.global.u64 %rd77, %rd17;
and.b32 %r56, %r40, -8;
sub.s32 %r57, %r3, %r56;
shl.b32 %r58, %r57, 1;
cvt.s64.s32 %rd4, %r57;
shl.b32 %r59, %r21, 1;
shl.b32 %r938, %r941, 3;
add.s32 %r940, %r938, %r59;
add.s32 %r939, %r21, %r938;
add.s32 %r60, %r941, %r48;
add.s32 %r61, %r60, %r52;
add.s32 %r62, %r4, 1;
mad.lo.s32 %r63, %r20, %r62, %r61;
mul.wide.s32 %rd5, %r63, 82;
mad.lo.s32 %r64, %r4, %r20, %r941;
add.s32 %r65, %r64, %r48;
add.s32 %r66, %r65, %r52;
mul.wide.s32 %rd6, %r66, 82;
mul.wide.s32 %rd22, %r58, 2;
add.s64 %rd23, %rd6, %rd22;
add.s64 %rd7, %rd23, 36;
add.s64 %rd24, %rd5, %rd4;
add.s64 %rd8, %rd24, 66;
add.s64 %rd25, %rd6, %rd4;
add.s64 %rd9, %rd25, 74;
add.s64 %rd26, %rd5, %rd22;
add.s64 %rd10, %rd26, 34;
cvta.to.global.u64 %rd11, %rd18;
mov.f32 %f157, 0f00000000;

$L__BB231_2:
cvt.u64.u32 %rd27, %r938;
add.s64 %rd28, %rd2, %rd27;
add.s64 %rd29, %rd28, %rd4;
add.s64 %rd30, %rd77, %rd7;
ld.global.nc.u16 %rs7, [%rd30+-34];
shr.u16 %rs8, %rs7, 8;
cvt.u32.u16 %r361, %rs8;
ld.global.nc.u16 %rs9, [%rd30+-32];
shr.u16 %rs10, %rs9, 8;
cvt.u32.u16 %r362, %rs10;
add.s64 %rd31, %rd77, %rd9;
ld.global.nc.u8 %rs11, [%rd31+-8];
ld.global.nc.u16 %rs12, [%rd30+-2];
ld.global.nc.u16 %rs13, [%rd30];
ld.global.nc.u8 %rs14, [%rd31];
and.b16 %rs15, %rs14, 240;
and.b16 %rs16, %rs14, 15;
shr.u16 %rs17, %rs15, 4;
and.b16 %rs18, %rs7, 255;
cvt.u32.u16 %r363, %rs18;
shl.b16 %rs19, %rs11, 8;
cvt.u32.u16 %r364, %rs19;
and.b32 %r365, %r364, 768;
or.b32 %r366, %r365, %r363;
mul.wide.u32 %rd32, %r366, 8;
mov.u64 %rd33, iq2s_grid;
add.s64 %rd34, %rd33, %rd32;
and.b16 %rs20, %rs12, 255;
mul.wide.u16 %r367, %rs20, 128;
cvt.u32.u16 %r368, %rs12;
shl.b32 %r369, %r368, 21;
or.b32 %r370, %r367, %r369;
and.b32 %r371, %r370, 25166208;
or.b32 %r372, %r371, -2139062144;
add.s32 %r373, %r372, -16843009;
or.b32 %r263, %r373, %r371;
mov.u32 %r67, %r263;

prmt.b32 %r67, %r67, 0, 0xba98;

mul.wide.u16 %r374, %rs20, 8;
shl.b32 %r375, %r368, 17;
or.b32 %r376, %r374, %r375;
and.b32 %r377, %r376, 25166208;
or.b32 %r378, %r377, -2139062144;
add.s32 %r379, %r378, -16843009;
or.b32 %r265, %r379, %r377;
mov.u32 %r69, %r265;

prmt.b32 %r69, %r69, 0, 0xba98;

ld.global.nc.v2.u32 {%r380, %r381}, [%rd34];
not.b32 %r383, %r380;
mov.u32 %r343, 0;
not.b32 %r385, %r381;
xor.b32 %r386, %r380, %r67;
or.b32 %r387, %r386, -2139062144;
and.b32 %r388, %r67, 2139062143;
and.b32 %r389, %r383, -2139062144;
sub.s32 %r390, %r387, %r388;
xor.b32 %r72, %r389, %r390;
xor.b32 %r391, %r381, %r69;
or.b32 %r392, %r391, -2139062144;
and.b32 %r393, %r69, 2139062143;
and.b32 %r394, %r385, -2139062144;
sub.s32 %r395, %r392, %r393;
xor.b32 %r76, %r394, %r395;
mul.lo.s64 %rd35, %rd29, 36;
add.s64 %rd36, %rd11, %rd35;
ld.global.nc.u32 %r77, [%rd36+8];
ld.global.nc.u32 %r73, [%rd36+4];

dp4a.s32.s32 %r71, %r72, %r73, %r343;


dp4a.s32.s32 %r75, %r76, %r77, %r71;

mul.wide.u16 %r396, %rs11, 64;
and.b32 %r397, %r396, 768;
or.b32 %r398, %r397, %r361;
mul.wide.u32 %rd37, %r398, 8;
add.s64 %rd38, %rd33, %rd37;
shr.u16 %rs22, %rs12, 8;
mul.wide.u16 %r399, %rs22, 128;
cvt.u32.u16 %r400, %rs22;
shl.b32 %r401, %r400, 21;
or.b32 %r402, %r399, %r401;
and.b32 %r403, %r402, 25166208;
or.b32 %r404, %r403, -2139062144;
add.s32 %r405, %r404, -16843009;
or.b32 %r275, %r405, %r403;
mov.u32 %r79, %r275;

prmt.b32 %r79, %r79, 0, 0xba98;

mul.wide.u16 %r406, %rs22, 8;
shl.b32 %r407, %r400, 17;
or.b32 %r408, %r406, %r407;
and.b32 %r409, %r408, 25166208;
or.b32 %r410, %r409, -2139062144;
add.s32 %r411, %r410, -16843009;
or.b32 %r277, %r411, %r409;
mov.u32 %r81, %r277;

prmt.b32 %r81, %r81, 0, 0xba98;

ld.global.nc.v2.u32 {%r412, %r413}, [%rd38];
not.b32 %r415, %r412;
not.b32 %r417, %r413;
xor.b32 %r418, %r412, %r79;
or.b32 %r419, %r418, -2139062144;
and.b32 %r420, %r79, 2139062143;
and.b32 %r421, %r415, -2139062144;
sub.s32 %r422, %r419, %r420;
xor.b32 %r84, %r421, %r422;
xor.b32 %r423, %r413, %r81;
or.b32 %r424, %r423, -2139062144;
and.b32 %r425, %r81, 2139062143;
and.b32 %r426, %r417, -2139062144;
sub.s32 %r427, %r424, %r425;
xor.b32 %r88, %r426, %r427;
ld.global.nc.u32 %r89, [%rd36+16];
ld.global.nc.u32 %r85, [%rd36+12];

dp4a.s32.s32 %r83, %r84, %r85, %r75;


dp4a.s32.s32 %r87, %r88, %r89, %r83;

and.b16 %rs23, %rs9, 255;
cvt.u32.u16 %r428, %rs23;
mul.wide.u16 %r429, %rs11, 16;
and.b32 %r430, %r429, 768;
or.b32 %r431, %r430, %r428;
mul.wide.u32 %rd39, %r431, 8;
add.s64 %rd40, %rd33, %rd39;
and.b16 %rs24, %rs13, 255;
mul.wide.u16 %r432, %rs24, 128;
cvt.u32.u16 %r433, %rs13;
shl.b32 %r434, %r433, 21;
or.b32 %r435, %r432, %r434;
and.b32 %r436, %r435, 25166208;
or.b32 %r437, %r436, -2139062144;
add.s32 %r438, %r437, -16843009;
or.b32 %r287, %r438, %r436;
mov.u32 %r91, %r287;

prmt.b32 %r91, %r91, 0, 0xba98;

mul.wide.u16 %r439, %rs24, 8;
shl.b32 %r440, %r433, 17;
or.b32 %r441, %r439, %r440;
and.b32 %r442, %r441, 25166208;
or.b32 %r443, %r442, -2139062144;
add.s32 %r444, %r443, -16843009;
or.b32 %r289, %r444, %r442;
mov.u32 %r93, %r289;

prmt.b32 %r93, %r93, 0, 0xba98;

ld.global.nc.v2.u32 {%r445, %r446}, [%rd40];
not.b32 %r448, %r445;
not.b32 %r450, %r446;
xor.b32 %r451, %r445, %r91;
or.b32 %r452, %r451, -2139062144;
and.b32 %r453, %r91, 2139062143;
and.b32 %r454, %r448, -2139062144;
sub.s32 %r455, %r452, %r453;
xor.b32 %r96, %r454, %r455;
xor.b32 %r456, %r446, %r93;
or.b32 %r457, %r456, -2139062144;
and.b32 %r458, %r93, 2139062143;
and.b32 %r459, %r450, -2139062144;
sub.s32 %r460, %r457, %r458;
xor.b32 %r100, %r459, %r460;
ld.global.nc.u32 %r101, [%rd36+24];
ld.global.nc.u32 %r97, [%rd36+20];

dp4a.s32.s32 %r95, %r96, %r97, %r343;


dp4a.s32.s32 %r99, %r100, %r101, %r95;

mul.wide.u16 %r461, %rs11, 4;
and.b32 %r462, %r461, 768;
or.b32 %r463, %r462, %r362;
mul.wide.u32 %rd41, %r463, 8;
add.s64 %rd42, %rd33, %rd41;
shr.u16 %rs25, %rs13, 8;
mul.wide.u16 %r464, %rs25, 128;
cvt.u32.u16 %r465, %rs25;
shl.b32 %r466, %r465, 21;
or.b32 %r467, %r464, %r466;
and.b32 %r468, %r467, 25166208;
or.b32 %r469, %r468, -2139062144;
add.s32 %r470, %r469, -16843009;
or.b32 %r299, %r470, %r468;
mov.u32 %r103, %r299;

prmt.b32 %r103, %r103, 0, 0xba98;

mul.wide.u16 %r471, %rs25, 8;
shl.b32 %r472, %r465, 17;
or.b32 %r473, %r471, %r472;
and.b32 %r474, %r473, 25166208;
or.b32 %r475, %r474, -2139062144;
add.s32 %r476, %r475, -16843009;
or.b32 %r301, %r476, %r474;
mov.u32 %r105, %r301;

prmt.b32 %r105, %r105, 0, 0xba98;

ld.global.nc.v2.u32 {%r477, %r478}, [%rd42];
not.b32 %r480, %r477;
not.b32 %r482, %r478;
xor.b32 %r483, %r477, %r103;
or.b32 %r484, %r483, -2139062144;
and.b32 %r485, %r103, 2139062143;
and.b32 %r486, %r480, -2139062144;
sub.s32 %r487, %r484, %r485;
xor.b32 %r108, %r486, %r487;
xor.b32 %r488, %r478, %r105;
or.b32 %r489, %r488, -2139062144;
and.b32 %r490, %r105, 2139062143;
and.b32 %r491, %r482, -2139062144;
sub.s32 %r492, %r489, %r490;
xor.b32 %r112, %r491, %r492;
ld.global.nc.u32 %r113, [%rd36+32];
ld.global.nc.u32 %r109, [%rd36+28];

dp4a.s32.s32 %r107, %r108, %r109, %r99;


dp4a.s32.s32 %r111, %r112, %r113, %r107;

cvt.u32.u16 %r493, %rs16;
cvt.u32.u16 %r494, %rs17;
mul.lo.s32 %r495, %r87, %r493;
mad.lo.s32 %r496, %r111, %r494, %r495;
add.s32 %r497, %r111, %r87;
shr.u32 %r498, %r497, 31;
add.s32 %r499, %r497, %r498;
shr.s32 %r500, %r499, 1;
add.s32 %r501, %r496, %r500;
shr.s32 %r502, %r501, 31;
shr.u32 %r503, %r502, 30;
add.s32 %r504, %r501, %r503;
shr.s32 %r505, %r504, 2;
add.s64 %rd43, %rd77, %rd6;
ld.global.nc.u16 %rs1, [%rd43];

{ cvt.f32.f16 %f37, %rs1;}


ld.global.nc.u32 %r115, [%rd36];

{.reg .f16 low,high;
mov.b32 {low,high},%r115;
cvt.f32.f16 %f38, low;}


mul.ftz.f32 %f49, %f37, %f38;
cvt.rn.f32.s32 %f50, %r505;
fma.rn.ftz.f32 %f162, %f49, %f50, %f162;
add.s64 %rd44, %rd77, %rd10;
ld.global.nc.u16 %rs26, [%rd44+-32];
shr.u16 %rs27, %rs26, 8;
cvt.u32.u16 %r506, %rs27;
ld.global.nc.u16 %rs28, [%rd44+-30];
shr.u16 %rs29, %rs28, 8;
cvt.u32.u16 %r507, %rs29;
add.s64 %rd45, %rd77, %rd8;
ld.global.nc.u8 %rs30, [%rd45];
ld.global.nc.u16 %rs31, [%rd44];
ld.global.nc.u16 %rs32, [%rd44+2];
ld.global.nc.u8 %rs33, [%rd45+8];
and.b16 %rs34, %rs33, 240;
and.b16 %rs35, %rs33, 15;
shr.u16 %rs36, %rs34, 4;
and.b16 %rs37, %rs26, 255;
cvt.u32.u16 %r508, %rs37;
shl.b16 %rs38, %rs30, 8;
cvt.u32.u16 %r509, %rs38;
and.b32 %r510, %r509, 768;
or.b32 %r511, %r510, %r508;
mul.wide.u32 %rd46, %r511, 8;
add.s64 %rd47, %rd33, %rd46;
and.b16 %rs39, %rs31, 255;
mul.wide.u16 %r512, %rs39, 128;
cvt.u32.u16 %r513, %rs31;
shl.b32 %r514, %r513, 21;
or.b32 %r515, %r512, %r514;
and.b32 %r516, %r515, 25166208;
or.b32 %r517, %r516, -2139062144;
add.s32 %r518, %r517, -16843009;
or.b32 %r312, %r518, %r516;
mov.u32 %r116, %r312;

prmt.b32 %r116, %r116, 0, 0xba98;

mul.wide.u16 %r519, %rs39, 8;
shl.b32 %r520, %r513, 17;
or.b32 %r521, %r519, %r520;
and.b32 %r522, %r521, 25166208;
or.b32 %r523, %r522, -2139062144;
add.s32 %r524, %r523, -16843009;
or.b32 %r314, %r524, %r522;
mov.u32 %r118, %r314;

prmt.b32 %r118, %r118, 0, 0xba98;

ld.global.nc.v2.u32 {%r525, %r526}, [%rd47];
not.b32 %r528, %r525;
not.b32 %r530, %r526;
xor.b32 %r531, %r525, %r116;
or.b32 %r532, %r531, -2139062144;
and.b32 %r533, %r116, 2139062143;
and.b32 %r534, %r528, -2139062144;
sub.s32 %r535, %r532, %r533;
xor.b32 %r121, %r534, %r535;
xor.b32 %r536, %r526, %r118;
or.b32 %r537, %r536, -2139062144;
and.b32 %r538, %r118, 2139062143;
and.b32 %r539, %r530, -2139062144;
sub.s32 %r540, %r537, %r538;
xor.b32 %r125, %r539, %r540;

dp4a.s32.s32 %r120, %r121, %r73, %r343;


dp4a.s32.s32 %r124, %r125, %r77, %r120;

mul.wide.u16 %r541, %rs30, 64;
and.b32 %r542, %r541, 768;
or.b32 %r543, %r542, %r506;
mul.wide.u32 %rd48, %r543, 8;
add.s64 %rd49, %rd33, %rd48;
shr.u16 %rs41, %rs31, 8;
mul.wide.u16 %r544, %rs41, 128;
cvt.u32.u16 %r545, %rs41;
shl.b32 %r546, %r545, 21;
or.b32 %r547, %r544, %r546;
and.b32 %r548, %r547, 25166208;
or.b32 %r549, %r548, -2139062144;
add.s32 %r550, %r549, -16843009;
or.b32 %r324, %r550, %r548;
mov.u32 %r128, %r324;

prmt.b32 %r128, %r128, 0, 0xba98;

mul.wide.u16 %r551, %rs41, 8;
shl.b32 %r552, %r545, 17;
or.b32 %r553, %r551, %r552;
and.b32 %r554, %r553, 25166208;
or.b32 %r555, %r554, -2139062144;
add.s32 %r556, %r555, -16843009;
or.b32 %r326, %r556, %r554;
mov.u32 %r130, %r326;

prmt.b32 %r130, %r130, 0, 0xba98;

ld.global.nc.v2.u32 {%r557, %r558}, [%rd49];
not.b32 %r560, %r557;
not.b32 %r562, %r558;
xor.b32 %r563, %r557, %r128;
or.b32 %r564, %r563, -2139062144;
and.b32 %r565, %r128, 2139062143;
and.b32 %r566, %r560, -2139062144;
sub.s32 %r567, %r564, %r565;
xor.b32 %r133, %r566, %r567;
xor.b32 %r568, %r558, %r130;
or.b32 %r569, %r568, -2139062144;
and.b32 %r570, %r130, 2139062143;
and.b32 %r571, %r562, -2139062144;
sub.s32 %r572, %r569, %r570;
xor.b32 %r137, %r571, %r572;

dp4a.s32.s32 %r132, %r133, %r85, %r124;


dp4a.s32.s32 %r136, %r137, %r89, %r132;

and.b16 %rs42, %rs28, 255;
cvt.u32.u16 %r573, %rs42;
mul.wide.u16 %r574, %rs30, 16;
and.b32 %r575, %r574, 768;
or.b32 %r576, %r575, %r573;
mul.wide.u32 %rd50, %r576, 8;
add.s64 %rd51, %rd33, %rd50;
and.b16 %rs43, %rs32, 255;
mul.wide.u16 %r577, %rs43, 128;
cvt.u32.u16 %r578, %rs32;
shl.b32 %r579, %r578, 21;
or.b32 %r580, %r577, %r579;
and.b32 %r581, %r580, 25166208;
or.b32 %r582, %r581, -2139062144;
add.s32 %r583, %r582, -16843009;
or.b32 %r336, %r583, %r581;
mov.u32 %r140, %r336;

prmt.b32 %r140, %r140, 0, 0xba98;

mul.wide.u16 %r584, %rs43, 8;
shl.b32 %r585, %r578, 17;
or.b32 %r586, %r584, %r585;
and.b32 %r587, %r586, 25166208;
or.b32 %r588, %r587, -2139062144;
add.s32 %r589, %r588, -16843009;
or.b32 %r338, %r589, %r587;
mov.u32 %r142, %r338;

prmt.b32 %r142, %r142, 0, 0xba98;

ld.global.nc.v2.u32 {%r590, %r591}, [%rd51];
not.b32 %r593, %r590;
not.b32 %r595, %r591;
xor.b32 %r596, %r590, %r140;
or.b32 %r597, %r596, -2139062144;
and.b32 %r598, %r140, 2139062143;
and.b32 %r599, %r593, -2139062144;
sub.s32 %r600, %r597, %r598;
xor.b32 %r145, %r599, %r600;
xor.b32 %r601, %r591, %r142;
or.b32 %r602, %r601, -2139062144;
and.b32 %r603, %r142, 2139062143;
and.b32 %r604, %r595, -2139062144;
sub.s32 %r605, %r602, %r603;
xor.b32 %r149, %r604, %r605;

dp4a.s32.s32 %r144, %r145, %r97, %r343;


dp4a.s32.s32 %r148, %r149, %r101, %r144;

mul.wide.u16 %r606, %rs30, 4;
and.b32 %r607, %r606, 768;
or.b32 %r608, %r607, %r507;
mul.wide.u32 %rd52, %r608, 8;
add.s64 %rd53, %rd33, %rd52;
shr.u16 %rs44, %rs32, 8;
mul.wide.u16 %r609, %rs44, 128;
cvt.u32.u16 %r610, %rs44;
shl.b32 %r611, %r610, 21;
or.b32 %r612, %r609, %r611;
and.b32 %r613, %r612, 25166208;
or.b32 %r614, %r613, -2139062144;
add.s32 %r615, %r614, -16843009;
or.b32 %r348, %r615, %r613;
mov.u32 %r152, %r348;

prmt.b32 %r152, %r152, 0, 0xba98;

mul.wide.u16 %r616, %rs44, 8;
shl.b32 %r617, %r610, 17;
or.b32 %r618, %r616, %r617;
and.b32 %r619, %r618, 25166208;
or.b32 %r620, %r619, -2139062144;
add.s32 %r621, %r620, -16843009;
or.b32 %r350, %r621, %r619;
mov.u32 %r154, %r350;

prmt.b32 %r154, %r154, 0, 0xba98;

ld.global.nc.v2.u32 {%r622, %r623}, [%rd53];
not.b32 %r625, %r622;
not.b32 %r627, %r623;
xor.b32 %r628, %r622, %r152;
or.b32 %r629, %r628, -2139062144;
and.b32 %r630, %r152, 2139062143;
and.b32 %r631, %r625, -2139062144;
sub.s32 %r632, %r629, %r630;
xor.b32 %r157, %r631, %r632;
xor.b32 %r633, %r623, %r154;
or.b32 %r634, %r633, -2139062144;
and.b32 %r635, %r154, 2139062143;
and.b32 %r636, %r627, -2139062144;
sub.s32 %r637, %r634, %r635;
xor.b32 %r161, %r636, %r637;

dp4a.s32.s32 %r156, %r157, %r109, %r148;


dp4a.s32.s32 %r160, %r161, %r113, %r156;

cvt.u32.u16 %r638, %rs35;
cvt.u32.u16 %r639, %rs36;
mul.lo.s32 %r640, %r136, %r638;
mad.lo.s32 %r641, %r160, %r639, %r640;
add.s32 %r642, %r160, %r136;
shr.u32 %r643, %r642, 31;
add.s32 %r644, %r642, %r643;
shr.s32 %r645, %r644, 1;
add.s32 %r646, %r641, %r645;
shr.s32 %r647, %r646, 31;
shr.u32 %r648, %r647, 30;
add.s32 %r649, %r646, %r648;
shr.s32 %r650, %r649, 2;
add.s64 %rd54, %rd77, %rd5;
ld.global.nc.u16 %rs2, [%rd54];

{ cvt.f32.f16 %f39, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r115;
cvt.f32.f16 %f40, low;}


mul.ftz.f32 %f51, %f39, %f40;
cvt.rn.f32.s32 %f52, %r650;
fma.rn.ftz.f32 %f161, %f51, %f52, %f161;
cvt.u64.u32 %rd55, %r939;
add.s64 %rd56, %rd2, %rd55;
add.s64 %rd57, %rd56, %rd4;
mov.u32 %r165, %r263;

prmt.b32 %r165, %r165, 0, 0xba98;

mov.u32 %r167, %r265;

prmt.b32 %r167, %r167, 0, 0xba98;

xor.b32 %r651, %r380, %r165;
or.b32 %r652, %r651, -2139062144;
and.b32 %r653, %r165, 2139062143;
sub.s32 %r654, %r652, %r653;
xor.b32 %r170, %r389, %r654;
xor.b32 %r655, %r381, %r167;
or.b32 %r656, %r655, -2139062144;
and.b32 %r657, %r167, 2139062143;
sub.s32 %r658, %r656, %r657;
xor.b32 %r174, %r394, %r658;
mul.lo.s64 %rd58, %rd57, 36;
add.s64 %rd59, %rd11, %rd58;
ld.global.nc.u32 %r175, [%rd59+8];
ld.global.nc.u32 %r171, [%rd59+4];

dp4a.s32.s32 %r169, %r170, %r171, %r343;


dp4a.s32.s32 %r173, %r174, %r175, %r169;

mov.u32 %r177, %r275;

prmt.b32 %r177, %r177, 0, 0xba98;

mov.u32 %r179, %r277;

prmt.b32 %r179, %r179, 0, 0xba98;

xor.b32 %r659, %r412, %r177;
or.b32 %r660, %r659, -2139062144;
and.b32 %r661, %r177, 2139062143;
sub.s32 %r662, %r660, %r661;
xor.b32 %r182, %r421, %r662;
xor.b32 %r663, %r413, %r179;
or.b32 %r664, %r663, -2139062144;
and.b32 %r665, %r179, 2139062143;
sub.s32 %r666, %r664, %r665;
xor.b32 %r186, %r426, %r666;
ld.global.nc.u32 %r187, [%rd59+16];
ld.global.nc.u32 %r183, [%rd59+12];

dp4a.s32.s32 %r181, %r182, %r183, %r173;


dp4a.s32.s32 %r185, %r186, %r187, %r181;

mov.u32 %r189, %r287;

prmt.b32 %r189, %r189, 0, 0xba98;

mov.u32 %r191, %r289;

prmt.b32 %r191, %r191, 0, 0xba98;

xor.b32 %r667, %r445, %r189;
or.b32 %r668, %r667, -2139062144;
and.b32 %r669, %r189, 2139062143;
sub.s32 %r670, %r668, %r669;
xor.b32 %r194, %r454, %r670;
xor.b32 %r671, %r446, %r191;
or.b32 %r672, %r671, -2139062144;
and.b32 %r673, %r191, 2139062143;
sub.s32 %r674, %r672, %r673;
xor.b32 %r198, %r459, %r674;
ld.global.nc.u32 %r199, [%rd59+24];
ld.global.nc.u32 %r195, [%rd59+20];

dp4a.s32.s32 %r193, %r194, %r195, %r343;


dp4a.s32.s32 %r197, %r198, %r199, %r193;

mov.u32 %r201, %r299;

prmt.b32 %r201, %r201, 0, 0xba98;

mov.u32 %r203, %r301;

prmt.b32 %r203, %r203, 0, 0xba98;

xor.b32 %r675, %r477, %r201;
or.b32 %r676, %r675, -2139062144;
and.b32 %r677, %r201, 2139062143;
sub.s32 %r678, %r676, %r677;
xor.b32 %r206, %r486, %r678;
xor.b32 %r679, %r478, %r203;
or.b32 %r680, %r679, -2139062144;
and.b32 %r681, %r203, 2139062143;
sub.s32 %r682, %r680, %r681;
xor.b32 %r210, %r491, %r682;
ld.global.nc.u32 %r211, [%rd59+32];
ld.global.nc.u32 %r207, [%rd59+28];

dp4a.s32.s32 %r205, %r206, %r207, %r197;


dp4a.s32.s32 %r209, %r210, %r211, %r205;

mul.lo.s32 %r683, %r185, %r493;
mad.lo.s32 %r684, %r209, %r494, %r683;
add.s32 %r685, %r209, %r185;
shr.u32 %r686, %r685, 31;
add.s32 %r687, %r685, %r686;
shr.s32 %r688, %r687, 1;
add.s32 %r689, %r684, %r688;
shr.s32 %r690, %r689, 31;
shr.u32 %r691, %r690, 30;
add.s32 %r692, %r689, %r691;
shr.s32 %r693, %r692, 2;

{ cvt.f32.f16 %f41, %rs1;}


ld.global.nc.u32 %r213, [%rd59];

{.reg .f16 low,high;
mov.b32 {low,high},%r213;
cvt.f32.f16 %f42, low;}


mul.ftz.f32 %f53, %f41, %f42;
cvt.rn.f32.s32 %f54, %r693;
fma.rn.ftz.f32 %f160, %f53, %f54, %f160;
mov.u32 %r214, %r312;

prmt.b32 %r214, %r214, 0, 0xba98;

mov.u32 %r216, %r314;

prmt.b32 %r216, %r216, 0, 0xba98;

xor.b32 %r694, %r525, %r214;
or.b32 %r695, %r694, -2139062144;
and.b32 %r696, %r214, 2139062143;
sub.s32 %r697, %r695, %r696;
xor.b32 %r219, %r534, %r697;
xor.b32 %r698, %r526, %r216;
or.b32 %r699, %r698, -2139062144;
and.b32 %r700, %r216, 2139062143;
sub.s32 %r701, %r699, %r700;
xor.b32 %r223, %r539, %r701;

dp4a.s32.s32 %r218, %r219, %r171, %r343;


dp4a.s32.s32 %r222, %r223, %r175, %r218;

mov.u32 %r226, %r324;

prmt.b32 %r226, %r226, 0, 0xba98;

mov.u32 %r228, %r326;

prmt.b32 %r228, %r228, 0, 0xba98;

xor.b32 %r702, %r557, %r226;
or.b32 %r703, %r702, -2139062144;
and.b32 %r704, %r226, 2139062143;
sub.s32 %r705, %r703, %r704;
xor.b32 %r231, %r566, %r705;
xor.b32 %r706, %r558, %r228;
or.b32 %r707, %r706, -2139062144;
and.b32 %r708, %r228, 2139062143;
sub.s32 %r709, %r707, %r708;
xor.b32 %r235, %r571, %r709;

dp4a.s32.s32 %r230, %r231, %r183, %r222;


dp4a.s32.s32 %r234, %r235, %r187, %r230;

mov.u32 %r238, %r336;

prmt.b32 %r238, %r238, 0, 0xba98;

mov.u32 %r240, %r338;

prmt.b32 %r240, %r240, 0, 0xba98;

xor.b32 %r710, %r590, %r238;
or.b32 %r711, %r710, -2139062144;
and.b32 %r712, %r238, 2139062143;
sub.s32 %r713, %r711, %r712;
xor.b32 %r243, %r599, %r713;
xor.b32 %r714, %r591, %r240;
or.b32 %r715, %r714, -2139062144;
and.b32 %r716, %r240, 2139062143;
sub.s32 %r717, %r715, %r716;
xor.b32 %r247, %r604, %r717;

dp4a.s32.s32 %r242, %r243, %r195, %r343;


dp4a.s32.s32 %r246, %r247, %r199, %r242;

mov.u32 %r250, %r348;

prmt.b32 %r250, %r250, 0, 0xba98;

mov.u32 %r252, %r350;

prmt.b32 %r252, %r252, 0, 0xba98;

xor.b32 %r718, %r622, %r250;
or.b32 %r719, %r718, -2139062144;
and.b32 %r720, %r250, 2139062143;
sub.s32 %r721, %r719, %r720;
xor.b32 %r255, %r631, %r721;
xor.b32 %r722, %r623, %r252;
or.b32 %r723, %r722, -2139062144;
and.b32 %r724, %r252, 2139062143;
sub.s32 %r725, %r723, %r724;
xor.b32 %r259, %r636, %r725;

dp4a.s32.s32 %r254, %r255, %r207, %r246;


dp4a.s32.s32 %r258, %r259, %r211, %r254;

mul.lo.s32 %r726, %r234, %r638;
mad.lo.s32 %r727, %r258, %r639, %r726;
add.s32 %r728, %r258, %r234;
shr.u32 %r729, %r728, 31;
add.s32 %r730, %r728, %r729;
shr.s32 %r731, %r730, 1;
add.s32 %r732, %r727, %r731;
shr.s32 %r733, %r732, 31;
shr.u32 %r734, %r733, 30;
add.s32 %r735, %r732, %r734;
shr.s32 %r736, %r735, 2;

{ cvt.f32.f16 %f43, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r213;
cvt.f32.f16 %f44, low;}


mul.ftz.f32 %f55, %f43, %f44;
cvt.rn.f32.s32 %f56, %r736;
fma.rn.ftz.f32 %f159, %f55, %f56, %f159;
cvt.u64.u32 %rd60, %r940;
add.s64 %rd61, %rd2, %rd60;
add.s64 %rd62, %rd61, %rd4;

prmt.b32 %r263, %r263, 0, 0xba98;


prmt.b32 %r265, %r265, 0, 0xba98;

xor.b32 %r737, %r380, %r263;
or.b32 %r738, %r737, -2139062144;
and.b32 %r739, %r263, 2139062143;
sub.s32 %r740, %r738, %r739;
xor.b32 %r268, %r389, %r740;
xor.b32 %r741, %r381, %r265;
or.b32 %r742, %r741, -2139062144;
and.b32 %r743, %r265, 2139062143;
sub.s32 %r744, %r742, %r743;
xor.b32 %r272, %r394, %r744;
mul.lo.s64 %rd63, %rd62, 36;
add.s64 %rd64, %rd11, %rd63;
ld.global.nc.u32 %r273, [%rd64+8];
ld.global.nc.u32 %r269, [%rd64+4];

dp4a.s32.s32 %r267, %r268, %r269, %r343;


dp4a.s32.s32 %r271, %r272, %r273, %r267;


prmt.b32 %r275, %r275, 0, 0xba98;


prmt.b32 %r277, %r277, 0, 0xba98;

xor.b32 %r745, %r412, %r275;
or.b32 %r746, %r745, -2139062144;
and.b32 %r747, %r275, 2139062143;
sub.s32 %r748, %r746, %r747;
xor.b32 %r280, %r421, %r748;
xor.b32 %r749, %r413, %r277;
or.b32 %r750, %r749, -2139062144;
and.b32 %r751, %r277, 2139062143;
sub.s32 %r752, %r750, %r751;
xor.b32 %r284, %r426, %r752;
ld.global.nc.u32 %r285, [%rd64+16];
ld.global.nc.u32 %r281, [%rd64+12];

dp4a.s32.s32 %r279, %r280, %r281, %r271;


dp4a.s32.s32 %r283, %r284, %r285, %r279;


prmt.b32 %r287, %r287, 0, 0xba98;


prmt.b32 %r289, %r289, 0, 0xba98;

xor.b32 %r753, %r445, %r287;
or.b32 %r754, %r753, -2139062144;
and.b32 %r755, %r287, 2139062143;
sub.s32 %r756, %r754, %r755;
xor.b32 %r292, %r454, %r756;
xor.b32 %r757, %r446, %r289;
or.b32 %r758, %r757, -2139062144;
and.b32 %r759, %r289, 2139062143;
sub.s32 %r760, %r758, %r759;
xor.b32 %r296, %r459, %r760;
ld.global.nc.u32 %r297, [%rd64+24];
ld.global.nc.u32 %r293, [%rd64+20];

dp4a.s32.s32 %r291, %r292, %r293, %r343;


dp4a.s32.s32 %r295, %r296, %r297, %r291;


prmt.b32 %r299, %r299, 0, 0xba98;


prmt.b32 %r301, %r301, 0, 0xba98;

xor.b32 %r761, %r477, %r299;
or.b32 %r762, %r761, -2139062144;
and.b32 %r763, %r299, 2139062143;
sub.s32 %r764, %r762, %r763;
xor.b32 %r304, %r486, %r764;
xor.b32 %r765, %r478, %r301;
or.b32 %r766, %r765, -2139062144;
and.b32 %r767, %r301, 2139062143;
sub.s32 %r768, %r766, %r767;
xor.b32 %r308, %r491, %r768;
ld.global.nc.u32 %r309, [%rd64+32];
ld.global.nc.u32 %r305, [%rd64+28];

dp4a.s32.s32 %r303, %r304, %r305, %r295;


dp4a.s32.s32 %r307, %r308, %r309, %r303;

mul.lo.s32 %r769, %r283, %r493;
mad.lo.s32 %r770, %r307, %r494, %r769;
add.s32 %r771, %r307, %r283;
shr.u32 %r772, %r771, 31;
add.s32 %r773, %r771, %r772;
shr.s32 %r774, %r773, 1;
add.s32 %r775, %r770, %r774;
shr.s32 %r776, %r775, 31;
shr.u32 %r777, %r776, 30;
add.s32 %r778, %r775, %r777;
shr.s32 %r779, %r778, 2;

{ cvt.f32.f16 %f45, %rs1;}


ld.global.nc.u32 %r311, [%rd64];

{.reg .f16 low,high;
mov.b32 {low,high},%r311;
cvt.f32.f16 %f46, low;}


mul.ftz.f32 %f57, %f45, %f46;
cvt.rn.f32.s32 %f58, %r779;
fma.rn.ftz.f32 %f158, %f57, %f58, %f158;

prmt.b32 %r312, %r312, 0, 0xba98;


prmt.b32 %r314, %r314, 0, 0xba98;

xor.b32 %r780, %r525, %r312;
or.b32 %r781, %r780, -2139062144;
and.b32 %r782, %r312, 2139062143;
sub.s32 %r783, %r781, %r782;
xor.b32 %r317, %r534, %r783;
xor.b32 %r784, %r526, %r314;
or.b32 %r785, %r784, -2139062144;
and.b32 %r786, %r314, 2139062143;
sub.s32 %r787, %r785, %r786;
xor.b32 %r321, %r539, %r787;

dp4a.s32.s32 %r316, %r317, %r269, %r343;


dp4a.s32.s32 %r320, %r321, %r273, %r316;


prmt.b32 %r324, %r324, 0, 0xba98;


prmt.b32 %r326, %r326, 0, 0xba98;

xor.b32 %r788, %r557, %r324;
or.b32 %r789, %r788, -2139062144;
and.b32 %r790, %r324, 2139062143;
sub.s32 %r791, %r789, %r790;
xor.b32 %r329, %r566, %r791;
xor.b32 %r792, %r558, %r326;
or.b32 %r793, %r792, -2139062144;
and.b32 %r794, %r326, 2139062143;
sub.s32 %r795, %r793, %r794;
xor.b32 %r333, %r571, %r795;

dp4a.s32.s32 %r328, %r329, %r281, %r320;


dp4a.s32.s32 %r332, %r333, %r285, %r328;


prmt.b32 %r336, %r336, 0, 0xba98;


prmt.b32 %r338, %r338, 0, 0xba98;

xor.b32 %r796, %r590, %r336;
or.b32 %r797, %r796, -2139062144;
and.b32 %r798, %r336, 2139062143;
sub.s32 %r799, %r797, %r798;
xor.b32 %r341, %r599, %r799;
xor.b32 %r800, %r591, %r338;
or.b32 %r801, %r800, -2139062144;
and.b32 %r802, %r338, 2139062143;
sub.s32 %r803, %r801, %r802;
xor.b32 %r345, %r604, %r803;

dp4a.s32.s32 %r340, %r341, %r293, %r343;


dp4a.s32.s32 %r344, %r345, %r297, %r340;


prmt.b32 %r348, %r348, 0, 0xba98;


prmt.b32 %r350, %r350, 0, 0xba98;

xor.b32 %r804, %r622, %r348;
or.b32 %r805, %r804, -2139062144;
and.b32 %r806, %r348, 2139062143;
sub.s32 %r807, %r805, %r806;
xor.b32 %r353, %r631, %r807;
xor.b32 %r808, %r623, %r350;
or.b32 %r809, %r808, -2139062144;
and.b32 %r810, %r350, 2139062143;
sub.s32 %r811, %r809, %r810;
xor.b32 %r357, %r636, %r811;

dp4a.s32.s32 %r352, %r353, %r305, %r344;


dp4a.s32.s32 %r356, %r357, %r309, %r352;

mul.lo.s32 %r812, %r332, %r638;
mad.lo.s32 %r813, %r356, %r639, %r812;
add.s32 %r814, %r356, %r332;
shr.u32 %r815, %r814, 31;
add.s32 %r816, %r814, %r815;
shr.s32 %r817, %r816, 1;
add.s32 %r818, %r813, %r817;
shr.s32 %r819, %r818, 31;
shr.u32 %r820, %r819, 30;
add.s32 %r821, %r818, %r820;
shr.s32 %r822, %r821, 2;

{ cvt.f32.f16 %f47, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r311;
cvt.f32.f16 %f48, low;}


mul.ftz.f32 %f59, %f47, %f48;
cvt.rn.f32.s32 %f60, %r822;
fma.rn.ftz.f32 %f157, %f59, %f60, %f157;
add.s32 %r940, %r940, 128;
add.s32 %r939, %r939, 128;
add.s32 %r938, %r938, 128;
add.s64 %rd77, %rd77, 1312;
add.s32 %r941, %r941, 16;
setp.lt.s32 %p4, %r941, %r6;
@%p4 bra $L__BB231_2;

$L__BB231_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB231_5;

mov.u32 %r823, _ZZ13mul_mat_vec_qIL9ggml_type22ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r824, %r1, 768, %r823;
shl.b32 %r825, %r2, 2;
add.s32 %r826, %r824, %r825;
st.shared.f32 [%r826+-768], %f162;
st.shared.f32 [%r826+-640], %f161;
st.shared.f32 [%r826+-512], %f160;
st.shared.f32 [%r826+-384], %f159;
st.shared.f32 [%r826+-256], %f158;
st.shared.f32 [%r826+-128], %f157;

$L__BB231_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB231_18;

mov.u32 %r827, %ctaid.z;
mov.u32 %r828, %ctaid.y;
mad.lo.s32 %r829, %r828, %r28, %r4;
mad.lo.s32 %r830, %r827, %r34, %r829;
cvt.u64.u32 %rd14, %r830;
shl.b32 %r831, %r2, 2;
mov.u32 %r832, _ZZ13mul_mat_vec_qIL9ggml_type22ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r18, %r832, %r831;
mov.u32 %r833, 2;
ld.shared.f32 %f61, [%r18];
add.ftz.f32 %f62, %f61, %f162;
ld.shared.f32 %f63, [%r18+768];
add.ftz.f32 %f64, %f63, %f62;
ld.shared.f32 %f65, [%r18+1536];
add.ftz.f32 %f66, %f65, %f64;
mov.b32 %r834, %f66;
mov.u32 %r835, 31;
mov.u32 %r836, 16;
mov.u32 %r837, -1;
shfl.sync.bfly.b32 %r838|%p7, %r834, %r836, %r835, %r837;
mov.b32 %f67, %r838;
add.ftz.f32 %f68, %f66, %f67;
mov.b32 %r839, %f68;
mov.u32 %r840, 8;
shfl.sync.bfly.b32 %r841|%p8, %r839, %r840, %r835, %r837;
mov.b32 %f69, %r841;
add.ftz.f32 %f70, %f68, %f69;
mov.b32 %r842, %f70;
mov.u32 %r843, 4;
shfl.sync.bfly.b32 %r844|%p9, %r842, %r843, %r835, %r837;
mov.b32 %f71, %r844;
add.ftz.f32 %f72, %f70, %f71;
mov.b32 %r845, %f72;
shfl.sync.bfly.b32 %r846|%p10, %r845, %r833, %r835, %r837;
mov.b32 %f73, %r846;
add.ftz.f32 %f74, %f72, %f73;
mov.b32 %r847, %f74;
mov.u32 %r848, 1;
shfl.sync.bfly.b32 %r849|%p11, %r847, %r848, %r835, %r837;
mov.b32 %f75, %r849;
add.ftz.f32 %f19, %f74, %f75;
setp.lt.u32 %p12, %r4, %r22;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
mul.wide.u32 %rd65, %r830, 4;
add.s64 %rd15, %rd1, %rd65;
not.pred %p14, %p1;
@%p14 bra $L__BB231_8;

st.global.f32 [%rd15], %f19;

$L__BB231_8:
ld.shared.f32 %f76, [%r18+128];
add.ftz.f32 %f77, %f76, %f161;
ld.shared.f32 %f78, [%r18+896];
add.ftz.f32 %f79, %f78, %f77;
ld.shared.f32 %f80, [%r18+1664];
add.ftz.f32 %f81, %f80, %f79;
mov.b32 %r850, %f81;
shfl.sync.bfly.b32 %r854|%p15, %r850, %r836, %r835, %r837;
mov.b32 %f82, %r854;
add.ftz.f32 %f83, %f81, %f82;
mov.b32 %r855, %f83;
shfl.sync.bfly.b32 %r857|%p16, %r855, %r840, %r835, %r837;
mov.b32 %f84, %r857;
add.ftz.f32 %f85, %f83, %f84;
mov.b32 %r858, %f85;
shfl.sync.bfly.b32 %r860|%p17, %r858, %r843, %r835, %r837;
mov.b32 %f86, %r860;
add.ftz.f32 %f87, %f85, %f86;
mov.b32 %r861, %f87;
shfl.sync.bfly.b32 %r863|%p18, %r861, %r833, %r835, %r837;
mov.b32 %f88, %r863;
add.ftz.f32 %f89, %f87, %f88;
mov.b32 %r864, %f89;
shfl.sync.bfly.b32 %r866|%p19, %r864, %r848, %r835, %r837;
mov.b32 %f90, %r866;
add.ftz.f32 %f20, %f89, %f90;
add.s32 %r867, %r4, 1;
setp.lt.u32 %p20, %r867, %r22;
setp.eq.s32 %p21, %r2, 1;
and.pred %p2, %p21, %p20;
not.pred %p22, %p2;
@%p22 bra $L__BB231_10;

st.global.f32 [%rd15+4], %f20;

$L__BB231_10:
ld.shared.f32 %f91, [%r18+256];
add.ftz.f32 %f92, %f91, %f160;
ld.shared.f32 %f93, [%r18+1024];
add.ftz.f32 %f94, %f93, %f92;
ld.shared.f32 %f95, [%r18+1792];
add.ftz.f32 %f96, %f95, %f94;
mov.b32 %r868, %f96;
mov.u32 %r869, 31;
mov.u32 %r870, 16;
mov.u32 %r871, -1;
shfl.sync.bfly.b32 %r872|%p23, %r868, %r870, %r869, %r871;
mov.b32 %f97, %r872;
add.ftz.f32 %f98, %f96, %f97;
mov.b32 %r873, %f98;
mov.u32 %r874, 8;
shfl.sync.bfly.b32 %r875|%p24, %r873, %r874, %r869, %r871;
mov.b32 %f99, %r875;
add.ftz.f32 %f100, %f98, %f99;
mov.b32 %r876, %f100;
mov.u32 %r877, 4;
shfl.sync.bfly.b32 %r878|%p25, %r876, %r877, %r869, %r871;
mov.b32 %f101, %r878;
add.ftz.f32 %f102, %f100, %f101;
mov.b32 %r879, %f102;
mov.u32 %r880, 2;
shfl.sync.bfly.b32 %r881|%p26, %r879, %r880, %r869, %r871;
mov.b32 %f103, %r881;
add.ftz.f32 %f104, %f102, %f103;
mov.b32 %r882, %f104;
mov.u32 %r883, 1;
shfl.sync.bfly.b32 %r884|%p27, %r882, %r883, %r869, %r871;
mov.b32 %f105, %r884;
add.ftz.f32 %f21, %f104, %f105;
@%p14 bra $L__BB231_12;

cvt.u64.u32 %rd66, %r22;
add.s64 %rd67, %rd66, %rd14;
shl.b64 %rd68, %rd67, 2;
add.s64 %rd69, %rd1, %rd68;
st.global.f32 [%rd69], %f21;

$L__BB231_12:
ld.shared.f32 %f106, [%r18+384];
add.ftz.f32 %f107, %f106, %f159;
ld.shared.f32 %f108, [%r18+1152];
add.ftz.f32 %f109, %f108, %f107;
ld.shared.f32 %f110, [%r18+1920];
add.ftz.f32 %f111, %f110, %f109;
mov.b32 %r885, %f111;
shfl.sync.bfly.b32 %r889|%p29, %r885, %r870, %r869, %r871;
mov.b32 %f112, %r889;
add.ftz.f32 %f113, %f111, %f112;
mov.b32 %r890, %f113;
shfl.sync.bfly.b32 %r892|%p30, %r890, %r874, %r869, %r871;
mov.b32 %f114, %r892;
add.ftz.f32 %f115, %f113, %f114;
mov.b32 %r893, %f115;
shfl.sync.bfly.b32 %r895|%p31, %r893, %r877, %r869, %r871;
mov.b32 %f116, %r895;
add.ftz.f32 %f117, %f115, %f116;
mov.b32 %r896, %f117;
shfl.sync.bfly.b32 %r898|%p32, %r896, %r880, %r869, %r871;
mov.b32 %f118, %r898;
add.ftz.f32 %f119, %f117, %f118;
mov.b32 %r899, %f119;
shfl.sync.bfly.b32 %r901|%p33, %r899, %r883, %r869, %r871;
mov.b32 %f120, %r901;
add.ftz.f32 %f22, %f119, %f120;
@%p22 bra $L__BB231_14;

add.s32 %r902, %r22, 1;
cvt.u64.u32 %rd70, %r902;
add.s64 %rd71, %rd70, %rd14;
shl.b64 %rd72, %rd71, 2;
add.s64 %rd73, %rd1, %rd72;
st.global.f32 [%rd73], %f22;

$L__BB231_14:
shl.b32 %r903, %r22, 1;
mov.u32 %r904, 1;
ld.shared.f32 %f121, [%r18+512];
add.ftz.f32 %f122, %f121, %f158;
ld.shared.f32 %f123, [%r18+1280];
add.ftz.f32 %f124, %f123, %f122;
ld.shared.f32 %f125, [%r18+2048];
add.ftz.f32 %f126, %f125, %f124;
mov.b32 %r905, %f126;
mov.u32 %r906, 31;
mov.u32 %r907, 16;
mov.u32 %r908, -1;
shfl.sync.bfly.b32 %r909|%p35, %r905, %r907, %r906, %r908;
mov.b32 %f127, %r909;
add.ftz.f32 %f128, %f126, %f127;
mov.b32 %r910, %f128;
mov.u32 %r911, 8;
shfl.sync.bfly.b32 %r912|%p36, %r910, %r911, %r906, %r908;
mov.b32 %f129, %r912;
add.ftz.f32 %f130, %f128, %f129;
mov.b32 %r913, %f130;
mov.u32 %r914, 4;
shfl.sync.bfly.b32 %r915|%p37, %r913, %r914, %r906, %r908;
mov.b32 %f131, %r915;
add.ftz.f32 %f132, %f130, %f131;
mov.b32 %r916, %f132;
mov.u32 %r917, 2;
shfl.sync.bfly.b32 %r918|%p38, %r916, %r917, %r906, %r908;
mov.b32 %f133, %r918;
add.ftz.f32 %f134, %f132, %f133;
mov.b32 %r919, %f134;
shfl.sync.bfly.b32 %r920|%p39, %r919, %r904, %r906, %r908;
mov.b32 %f135, %r920;
add.ftz.f32 %f23, %f134, %f135;
cvt.u64.u32 %rd74, %r903;
add.s64 %rd75, %rd74, %rd14;
shl.b64 %rd76, %rd75, 2;
add.s64 %rd16, %rd1, %rd76;
@%p14 bra $L__BB231_16;

st.global.f32 [%rd16], %f23;

$L__BB231_16:
ld.shared.f32 %f136, [%r18+640];
add.ftz.f32 %f137, %f136, %f157;
ld.shared.f32 %f138, [%r18+1408];
add.ftz.f32 %f139, %f138, %f137;
ld.shared.f32 %f140, [%r18+2176];
add.ftz.f32 %f141, %f140, %f139;
mov.b32 %r921, %f141;
shfl.sync.bfly.b32 %r925|%p41, %r921, %r907, %r906, %r908;
mov.b32 %f142, %r925;
add.ftz.f32 %f143, %f141, %f142;
mov.b32 %r926, %f143;
shfl.sync.bfly.b32 %r928|%p42, %r926, %r911, %r906, %r908;
mov.b32 %f144, %r928;
add.ftz.f32 %f145, %f143, %f144;
mov.b32 %r929, %f145;
shfl.sync.bfly.b32 %r931|%p43, %r929, %r914, %r906, %r908;
mov.b32 %f146, %r931;
add.ftz.f32 %f147, %f145, %f146;
mov.b32 %r932, %f147;
shfl.sync.bfly.b32 %r934|%p44, %r932, %r917, %r906, %r908;
mov.b32 %f148, %r934;
add.ftz.f32 %f149, %f147, %f148;
mov.b32 %r935, %f149;
shfl.sync.bfly.b32 %r937|%p45, %r935, %r904, %r906, %r908;
mov.b32 %f150, %r937;
add.ftz.f32 %f24, %f149, %f150;
@%p22 bra $L__BB231_18;

st.global.f32 [%rd16+4], %f24;

$L__BB231_18:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type22ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type22ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type22ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type22ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type22ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type22ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type22ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type22ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type22ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<59>;
.reg .b16 %rs<47>;
.reg .f32 %f<217>;
.reg .b32 %r<1168>;
.reg .b64 %rd<91>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type22ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[3072];

ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type22ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type22ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type22ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type22ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type22ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type22ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type22ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type22ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type22ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type22ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type22ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type22ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type22ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type22ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type22ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type22ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type22ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
cvta.to.global.u64 %rd1, %rd19;
mov.u32 %r1, %tid.y;
shl.b32 %r40, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r40, %r2;
mov.u32 %r41, %ctaid.x;
shl.b32 %r4, %r41, 1;
shr.s32 %r42, %r3, 31;
shr.u32 %r43, %r42, 29;
add.s32 %r44, %r3, %r43;
shr.s32 %r1167, %r44, 3;
shr.u32 %r6, %r39, 8;
setp.ge.s32 %p3, %r1167, %r6;
mov.f32 %f209, 0f00000000;
mov.f32 %f210, %f209;
mov.f32 %f211, %f209;
mov.f32 %f212, %f209;
mov.f32 %f213, %f209;
mov.f32 %f214, %f209;
mov.f32 %f215, %f209;
mov.f32 %f216, %f209;
@%p3 bra $L__BB232_3;

mov.u32 %r45, %ctaid.y;
mov.u32 %r46, %ctaid.z;
mul.lo.s32 %r47, %r45, %r31;
cvt.u64.u32 %rd20, %r47;
mul.lo.s32 %r48, %r46, %r37;
cvt.u64.u32 %rd21, %r48;
add.s64 %rd2, %rd21, %rd20;
mul.hi.u32 %r49, %r46, %r33;
add.s32 %r50, %r49, %r46;
shr.u32 %r51, %r50, %r34;
mul.lo.s32 %r52, %r51, %r36;
mul.hi.u32 %r53, %r45, %r27;
add.s32 %r54, %r53, %r45;
shr.u32 %r55, %r54, %r28;
mul.lo.s32 %r56, %r55, %r30;
cvta.to.global.u64 %rd90, %rd17;
and.b32 %r60, %r44, -8;
sub.s32 %r61, %r3, %r60;
shl.b32 %r62, %r61, 1;
cvt.s64.s32 %rd4, %r61;
shl.b32 %r63, %r25, 1;
shl.b32 %r1163, %r1167, 3;
add.s32 %r1166, %r1163, %r63;
mad.lo.s32 %r1165, %r25, 3, %r1163;
add.s32 %r1164, %r25, %r1163;
add.s32 %r64, %r1167, %r52;
add.s32 %r65, %r64, %r56;
add.s32 %r66, %r4, 1;
mad.lo.s32 %r67, %r24, %r66, %r65;
mul.wide.s32 %rd5, %r67, 82;
mad.lo.s32 %r68, %r4, %r24, %r1167;
add.s32 %r69, %r68, %r52;
add.s32 %r70, %r69, %r56;
mul.wide.s32 %rd6, %r70, 82;
mul.wide.s32 %rd22, %r62, 2;
add.s64 %rd23, %rd6, %rd22;
add.s64 %rd7, %rd23, 36;
add.s64 %rd24, %rd5, %rd4;
add.s64 %rd8, %rd24, 66;
add.s64 %rd25, %rd6, %rd4;
add.s64 %rd9, %rd25, 74;
add.s64 %rd26, %rd5, %rd22;
add.s64 %rd10, %rd26, 34;
cvta.to.global.u64 %rd11, %rd18;
mov.f32 %f209, 0f00000000;

$L__BB232_2:
cvt.u64.u32 %rd27, %r1163;
add.s64 %rd28, %rd2, %rd27;
add.s64 %rd29, %rd28, %rd4;
add.s64 %rd30, %rd90, %rd7;
ld.global.nc.u16 %rs9, [%rd30+-34];
shr.u16 %rs10, %rs9, 8;
cvt.u32.u16 %r463, %rs10;
ld.global.nc.u16 %rs11, [%rd30+-32];
shr.u16 %rs12, %rs11, 8;
cvt.u32.u16 %r464, %rs12;
add.s64 %rd31, %rd90, %rd9;
ld.global.nc.u8 %rs13, [%rd31+-8];
ld.global.nc.u16 %rs14, [%rd30+-2];
ld.global.nc.u16 %rs15, [%rd30];
ld.global.nc.u8 %rs16, [%rd31];
and.b16 %rs17, %rs16, 240;
and.b16 %rs18, %rs16, 15;
shr.u16 %rs19, %rs17, 4;
and.b16 %rs20, %rs9, 255;
cvt.u32.u16 %r465, %rs20;
shl.b16 %rs21, %rs13, 8;
cvt.u32.u16 %r466, %rs21;
and.b32 %r467, %r466, 768;
or.b32 %r468, %r467, %r465;
mul.wide.u32 %rd32, %r468, 8;
mov.u64 %rd33, iq2s_grid;
add.s64 %rd34, %rd33, %rd32;
and.b16 %rs22, %rs14, 255;
mul.wide.u16 %r469, %rs22, 128;
cvt.u32.u16 %r470, %rs14;
shl.b32 %r471, %r470, 21;
or.b32 %r472, %r469, %r471;
and.b32 %r473, %r472, 25166208;
or.b32 %r474, %r473, -2139062144;
add.s32 %r475, %r474, -16843009;
or.b32 %r365, %r475, %r473;
mov.u32 %r71, %r365;

prmt.b32 %r71, %r71, 0, 0xba98;

mul.wide.u16 %r476, %rs22, 8;
shl.b32 %r477, %r470, 17;
or.b32 %r478, %r476, %r477;
and.b32 %r479, %r478, 25166208;
or.b32 %r480, %r479, -2139062144;
add.s32 %r481, %r480, -16843009;
or.b32 %r367, %r481, %r479;
mov.u32 %r73, %r367;

prmt.b32 %r73, %r73, 0, 0xba98;

ld.global.nc.v2.u32 {%r482, %r483}, [%rd34];
not.b32 %r485, %r482;
mov.u32 %r445, 0;
not.b32 %r487, %r483;
xor.b32 %r488, %r482, %r71;
or.b32 %r489, %r488, -2139062144;
and.b32 %r490, %r71, 2139062143;
and.b32 %r491, %r485, -2139062144;
sub.s32 %r492, %r489, %r490;
xor.b32 %r76, %r491, %r492;
xor.b32 %r493, %r483, %r73;
or.b32 %r494, %r493, -2139062144;
and.b32 %r495, %r73, 2139062143;
and.b32 %r496, %r487, -2139062144;
sub.s32 %r497, %r494, %r495;
xor.b32 %r80, %r496, %r497;
mul.lo.s64 %rd35, %rd29, 36;
add.s64 %rd36, %rd11, %rd35;
ld.global.nc.u32 %r81, [%rd36+8];
ld.global.nc.u32 %r77, [%rd36+4];

dp4a.s32.s32 %r75, %r76, %r77, %r445;


dp4a.s32.s32 %r79, %r80, %r81, %r75;

mul.wide.u16 %r498, %rs13, 64;
and.b32 %r499, %r498, 768;
or.b32 %r500, %r499, %r463;
mul.wide.u32 %rd37, %r500, 8;
add.s64 %rd38, %rd33, %rd37;
shr.u16 %rs24, %rs14, 8;
mul.wide.u16 %r501, %rs24, 128;
cvt.u32.u16 %r502, %rs24;
shl.b32 %r503, %r502, 21;
or.b32 %r504, %r501, %r503;
and.b32 %r505, %r504, 25166208;
or.b32 %r506, %r505, -2139062144;
add.s32 %r507, %r506, -16843009;
or.b32 %r377, %r507, %r505;
mov.u32 %r83, %r377;

prmt.b32 %r83, %r83, 0, 0xba98;

mul.wide.u16 %r508, %rs24, 8;
shl.b32 %r509, %r502, 17;
or.b32 %r510, %r508, %r509;
and.b32 %r511, %r510, 25166208;
or.b32 %r512, %r511, -2139062144;
add.s32 %r513, %r512, -16843009;
or.b32 %r379, %r513, %r511;
mov.u32 %r85, %r379;

prmt.b32 %r85, %r85, 0, 0xba98;

ld.global.nc.v2.u32 {%r514, %r515}, [%rd38];
not.b32 %r517, %r514;
not.b32 %r519, %r515;
xor.b32 %r520, %r514, %r83;
or.b32 %r521, %r520, -2139062144;
and.b32 %r522, %r83, 2139062143;
and.b32 %r523, %r517, -2139062144;
sub.s32 %r524, %r521, %r522;
xor.b32 %r88, %r523, %r524;
xor.b32 %r525, %r515, %r85;
or.b32 %r526, %r525, -2139062144;
and.b32 %r527, %r85, 2139062143;
and.b32 %r528, %r519, -2139062144;
sub.s32 %r529, %r526, %r527;
xor.b32 %r92, %r528, %r529;
ld.global.nc.u32 %r93, [%rd36+16];
ld.global.nc.u32 %r89, [%rd36+12];

dp4a.s32.s32 %r87, %r88, %r89, %r79;


dp4a.s32.s32 %r91, %r92, %r93, %r87;

and.b16 %rs25, %rs11, 255;
cvt.u32.u16 %r530, %rs25;
mul.wide.u16 %r531, %rs13, 16;
and.b32 %r532, %r531, 768;
or.b32 %r533, %r532, %r530;
mul.wide.u32 %rd39, %r533, 8;
add.s64 %rd40, %rd33, %rd39;
and.b16 %rs26, %rs15, 255;
mul.wide.u16 %r534, %rs26, 128;
cvt.u32.u16 %r535, %rs15;
shl.b32 %r536, %r535, 21;
or.b32 %r537, %r534, %r536;
and.b32 %r538, %r537, 25166208;
or.b32 %r539, %r538, -2139062144;
add.s32 %r540, %r539, -16843009;
or.b32 %r389, %r540, %r538;
mov.u32 %r95, %r389;

prmt.b32 %r95, %r95, 0, 0xba98;

mul.wide.u16 %r541, %rs26, 8;
shl.b32 %r542, %r535, 17;
or.b32 %r543, %r541, %r542;
and.b32 %r544, %r543, 25166208;
or.b32 %r545, %r544, -2139062144;
add.s32 %r546, %r545, -16843009;
or.b32 %r391, %r546, %r544;
mov.u32 %r97, %r391;

prmt.b32 %r97, %r97, 0, 0xba98;

ld.global.nc.v2.u32 {%r547, %r548}, [%rd40];
not.b32 %r550, %r547;
not.b32 %r552, %r548;
xor.b32 %r553, %r547, %r95;
or.b32 %r554, %r553, -2139062144;
and.b32 %r555, %r95, 2139062143;
and.b32 %r556, %r550, -2139062144;
sub.s32 %r557, %r554, %r555;
xor.b32 %r100, %r556, %r557;
xor.b32 %r558, %r548, %r97;
or.b32 %r559, %r558, -2139062144;
and.b32 %r560, %r97, 2139062143;
and.b32 %r561, %r552, -2139062144;
sub.s32 %r562, %r559, %r560;
xor.b32 %r104, %r561, %r562;
ld.global.nc.u32 %r105, [%rd36+24];
ld.global.nc.u32 %r101, [%rd36+20];

dp4a.s32.s32 %r99, %r100, %r101, %r445;


dp4a.s32.s32 %r103, %r104, %r105, %r99;

mul.wide.u16 %r563, %rs13, 4;
and.b32 %r564, %r563, 768;
or.b32 %r565, %r564, %r464;
mul.wide.u32 %rd41, %r565, 8;
add.s64 %rd42, %rd33, %rd41;
shr.u16 %rs27, %rs15, 8;
mul.wide.u16 %r566, %rs27, 128;
cvt.u32.u16 %r567, %rs27;
shl.b32 %r568, %r567, 21;
or.b32 %r569, %r566, %r568;
and.b32 %r570, %r569, 25166208;
or.b32 %r571, %r570, -2139062144;
add.s32 %r572, %r571, -16843009;
or.b32 %r401, %r572, %r570;
mov.u32 %r107, %r401;

prmt.b32 %r107, %r107, 0, 0xba98;

mul.wide.u16 %r573, %rs27, 8;
shl.b32 %r574, %r567, 17;
or.b32 %r575, %r573, %r574;
and.b32 %r576, %r575, 25166208;
or.b32 %r577, %r576, -2139062144;
add.s32 %r578, %r577, -16843009;
or.b32 %r403, %r578, %r576;
mov.u32 %r109, %r403;

prmt.b32 %r109, %r109, 0, 0xba98;

ld.global.nc.v2.u32 {%r579, %r580}, [%rd42];
not.b32 %r582, %r579;
not.b32 %r584, %r580;
xor.b32 %r585, %r579, %r107;
or.b32 %r586, %r585, -2139062144;
and.b32 %r587, %r107, 2139062143;
and.b32 %r588, %r582, -2139062144;
sub.s32 %r589, %r586, %r587;
xor.b32 %r112, %r588, %r589;
xor.b32 %r590, %r580, %r109;
or.b32 %r591, %r590, -2139062144;
and.b32 %r592, %r109, 2139062143;
and.b32 %r593, %r584, -2139062144;
sub.s32 %r594, %r591, %r592;
xor.b32 %r116, %r593, %r594;
ld.global.nc.u32 %r117, [%rd36+32];
ld.global.nc.u32 %r113, [%rd36+28];

dp4a.s32.s32 %r111, %r112, %r113, %r103;


dp4a.s32.s32 %r115, %r116, %r117, %r111;

cvt.u32.u16 %r595, %rs18;
cvt.u32.u16 %r596, %rs19;
mul.lo.s32 %r597, %r91, %r595;
mad.lo.s32 %r598, %r115, %r596, %r597;
add.s32 %r599, %r115, %r91;
shr.u32 %r600, %r599, 31;
add.s32 %r601, %r599, %r600;
shr.s32 %r602, %r601, 1;
add.s32 %r603, %r598, %r602;
shr.s32 %r604, %r603, 31;
shr.u32 %r605, %r604, 30;
add.s32 %r606, %r603, %r605;
shr.s32 %r607, %r606, 2;
add.s64 %rd43, %rd90, %rd6;
ld.global.nc.u16 %rs1, [%rd43];

{ cvt.f32.f16 %f49, %rs1;}


ld.global.nc.u32 %r119, [%rd36];

{.reg .f16 low,high;
mov.b32 {low,high},%r119;
cvt.f32.f16 %f50, low;}


mul.ftz.f32 %f65, %f49, %f50;
cvt.rn.f32.s32 %f66, %r607;
fma.rn.ftz.f32 %f216, %f65, %f66, %f216;
add.s64 %rd44, %rd90, %rd10;
ld.global.nc.u16 %rs28, [%rd44+-32];
shr.u16 %rs29, %rs28, 8;
cvt.u32.u16 %r608, %rs29;
ld.global.nc.u16 %rs30, [%rd44+-30];
shr.u16 %rs31, %rs30, 8;
cvt.u32.u16 %r609, %rs31;
add.s64 %rd45, %rd90, %rd8;
ld.global.nc.u8 %rs32, [%rd45];
ld.global.nc.u16 %rs33, [%rd44];
ld.global.nc.u16 %rs34, [%rd44+2];
ld.global.nc.u8 %rs35, [%rd45+8];
and.b16 %rs36, %rs35, 240;
and.b16 %rs37, %rs35, 15;
shr.u16 %rs38, %rs36, 4;
and.b16 %rs39, %rs28, 255;
cvt.u32.u16 %r610, %rs39;
shl.b16 %rs40, %rs32, 8;
cvt.u32.u16 %r611, %rs40;
and.b32 %r612, %r611, 768;
or.b32 %r613, %r612, %r610;
mul.wide.u32 %rd46, %r613, 8;
add.s64 %rd47, %rd33, %rd46;
and.b16 %rs41, %rs33, 255;
mul.wide.u16 %r614, %rs41, 128;
cvt.u32.u16 %r615, %rs33;
shl.b32 %r616, %r615, 21;
or.b32 %r617, %r614, %r616;
and.b32 %r618, %r617, 25166208;
or.b32 %r619, %r618, -2139062144;
add.s32 %r620, %r619, -16843009;
or.b32 %r414, %r620, %r618;
mov.u32 %r120, %r414;

prmt.b32 %r120, %r120, 0, 0xba98;

mul.wide.u16 %r621, %rs41, 8;
shl.b32 %r622, %r615, 17;
or.b32 %r623, %r621, %r622;
and.b32 %r624, %r623, 25166208;
or.b32 %r625, %r624, -2139062144;
add.s32 %r626, %r625, -16843009;
or.b32 %r416, %r626, %r624;
mov.u32 %r122, %r416;

prmt.b32 %r122, %r122, 0, 0xba98;

ld.global.nc.v2.u32 {%r627, %r628}, [%rd47];
not.b32 %r630, %r627;
not.b32 %r632, %r628;
xor.b32 %r633, %r627, %r120;
or.b32 %r634, %r633, -2139062144;
and.b32 %r635, %r120, 2139062143;
and.b32 %r636, %r630, -2139062144;
sub.s32 %r637, %r634, %r635;
xor.b32 %r125, %r636, %r637;
xor.b32 %r638, %r628, %r122;
or.b32 %r639, %r638, -2139062144;
and.b32 %r640, %r122, 2139062143;
and.b32 %r641, %r632, -2139062144;
sub.s32 %r642, %r639, %r640;
xor.b32 %r129, %r641, %r642;

dp4a.s32.s32 %r124, %r125, %r77, %r445;


dp4a.s32.s32 %r128, %r129, %r81, %r124;

mul.wide.u16 %r643, %rs32, 64;
and.b32 %r644, %r643, 768;
or.b32 %r645, %r644, %r608;
mul.wide.u32 %rd48, %r645, 8;
add.s64 %rd49, %rd33, %rd48;
shr.u16 %rs43, %rs33, 8;
mul.wide.u16 %r646, %rs43, 128;
cvt.u32.u16 %r647, %rs43;
shl.b32 %r648, %r647, 21;
or.b32 %r649, %r646, %r648;
and.b32 %r650, %r649, 25166208;
or.b32 %r651, %r650, -2139062144;
add.s32 %r652, %r651, -16843009;
or.b32 %r426, %r652, %r650;
mov.u32 %r132, %r426;

prmt.b32 %r132, %r132, 0, 0xba98;

mul.wide.u16 %r653, %rs43, 8;
shl.b32 %r654, %r647, 17;
or.b32 %r655, %r653, %r654;
and.b32 %r656, %r655, 25166208;
or.b32 %r657, %r656, -2139062144;
add.s32 %r658, %r657, -16843009;
or.b32 %r428, %r658, %r656;
mov.u32 %r134, %r428;

prmt.b32 %r134, %r134, 0, 0xba98;

ld.global.nc.v2.u32 {%r659, %r660}, [%rd49];
not.b32 %r662, %r659;
not.b32 %r664, %r660;
xor.b32 %r665, %r659, %r132;
or.b32 %r666, %r665, -2139062144;
and.b32 %r667, %r132, 2139062143;
and.b32 %r668, %r662, -2139062144;
sub.s32 %r669, %r666, %r667;
xor.b32 %r137, %r668, %r669;
xor.b32 %r670, %r660, %r134;
or.b32 %r671, %r670, -2139062144;
and.b32 %r672, %r134, 2139062143;
and.b32 %r673, %r664, -2139062144;
sub.s32 %r674, %r671, %r672;
xor.b32 %r141, %r673, %r674;

dp4a.s32.s32 %r136, %r137, %r89, %r128;


dp4a.s32.s32 %r140, %r141, %r93, %r136;

and.b16 %rs44, %rs30, 255;
cvt.u32.u16 %r675, %rs44;
mul.wide.u16 %r676, %rs32, 16;
and.b32 %r677, %r676, 768;
or.b32 %r678, %r677, %r675;
mul.wide.u32 %rd50, %r678, 8;
add.s64 %rd51, %rd33, %rd50;
and.b16 %rs45, %rs34, 255;
mul.wide.u16 %r679, %rs45, 128;
cvt.u32.u16 %r680, %rs34;
shl.b32 %r681, %r680, 21;
or.b32 %r682, %r679, %r681;
and.b32 %r683, %r682, 25166208;
or.b32 %r684, %r683, -2139062144;
add.s32 %r685, %r684, -16843009;
or.b32 %r438, %r685, %r683;
mov.u32 %r144, %r438;

prmt.b32 %r144, %r144, 0, 0xba98;

mul.wide.u16 %r686, %rs45, 8;
shl.b32 %r687, %r680, 17;
or.b32 %r688, %r686, %r687;
and.b32 %r689, %r688, 25166208;
or.b32 %r690, %r689, -2139062144;
add.s32 %r691, %r690, -16843009;
or.b32 %r440, %r691, %r689;
mov.u32 %r146, %r440;

prmt.b32 %r146, %r146, 0, 0xba98;

ld.global.nc.v2.u32 {%r692, %r693}, [%rd51];
not.b32 %r695, %r692;
not.b32 %r697, %r693;
xor.b32 %r698, %r692, %r144;
or.b32 %r699, %r698, -2139062144;
and.b32 %r700, %r144, 2139062143;
and.b32 %r701, %r695, -2139062144;
sub.s32 %r702, %r699, %r700;
xor.b32 %r149, %r701, %r702;
xor.b32 %r703, %r693, %r146;
or.b32 %r704, %r703, -2139062144;
and.b32 %r705, %r146, 2139062143;
and.b32 %r706, %r697, -2139062144;
sub.s32 %r707, %r704, %r705;
xor.b32 %r153, %r706, %r707;

dp4a.s32.s32 %r148, %r149, %r101, %r445;


dp4a.s32.s32 %r152, %r153, %r105, %r148;

mul.wide.u16 %r708, %rs32, 4;
and.b32 %r709, %r708, 768;
or.b32 %r710, %r709, %r609;
mul.wide.u32 %rd52, %r710, 8;
add.s64 %rd53, %rd33, %rd52;
shr.u16 %rs46, %rs34, 8;
mul.wide.u16 %r711, %rs46, 128;
cvt.u32.u16 %r712, %rs46;
shl.b32 %r713, %r712, 21;
or.b32 %r714, %r711, %r713;
and.b32 %r715, %r714, 25166208;
or.b32 %r716, %r715, -2139062144;
add.s32 %r717, %r716, -16843009;
or.b32 %r450, %r717, %r715;
mov.u32 %r156, %r450;

prmt.b32 %r156, %r156, 0, 0xba98;

mul.wide.u16 %r718, %rs46, 8;
shl.b32 %r719, %r712, 17;
or.b32 %r720, %r718, %r719;
and.b32 %r721, %r720, 25166208;
or.b32 %r722, %r721, -2139062144;
add.s32 %r723, %r722, -16843009;
or.b32 %r452, %r723, %r721;
mov.u32 %r158, %r452;

prmt.b32 %r158, %r158, 0, 0xba98;

ld.global.nc.v2.u32 {%r724, %r725}, [%rd53];
not.b32 %r727, %r724;
not.b32 %r729, %r725;
xor.b32 %r730, %r724, %r156;
or.b32 %r731, %r730, -2139062144;
and.b32 %r732, %r156, 2139062143;
and.b32 %r733, %r727, -2139062144;
sub.s32 %r734, %r731, %r732;
xor.b32 %r161, %r733, %r734;
xor.b32 %r735, %r725, %r158;
or.b32 %r736, %r735, -2139062144;
and.b32 %r737, %r158, 2139062143;
and.b32 %r738, %r729, -2139062144;
sub.s32 %r739, %r736, %r737;
xor.b32 %r165, %r738, %r739;

dp4a.s32.s32 %r160, %r161, %r113, %r152;


dp4a.s32.s32 %r164, %r165, %r117, %r160;

cvt.u32.u16 %r740, %rs37;
cvt.u32.u16 %r741, %rs38;
mul.lo.s32 %r742, %r140, %r740;
mad.lo.s32 %r743, %r164, %r741, %r742;
add.s32 %r744, %r164, %r140;
shr.u32 %r745, %r744, 31;
add.s32 %r746, %r744, %r745;
shr.s32 %r747, %r746, 1;
add.s32 %r748, %r743, %r747;
shr.s32 %r749, %r748, 31;
shr.u32 %r750, %r749, 30;
add.s32 %r751, %r748, %r750;
shr.s32 %r752, %r751, 2;
add.s64 %rd54, %rd90, %rd5;
ld.global.nc.u16 %rs2, [%rd54];

{ cvt.f32.f16 %f51, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r119;
cvt.f32.f16 %f52, low;}


mul.ftz.f32 %f67, %f51, %f52;
cvt.rn.f32.s32 %f68, %r752;
fma.rn.ftz.f32 %f215, %f67, %f68, %f215;
cvt.u64.u32 %rd55, %r1164;
add.s64 %rd56, %rd2, %rd55;
add.s64 %rd57, %rd56, %rd4;
mov.u32 %r169, %r365;

prmt.b32 %r169, %r169, 0, 0xba98;

mov.u32 %r171, %r367;

prmt.b32 %r171, %r171, 0, 0xba98;

xor.b32 %r753, %r482, %r169;
or.b32 %r754, %r753, -2139062144;
and.b32 %r755, %r169, 2139062143;
sub.s32 %r756, %r754, %r755;
xor.b32 %r174, %r491, %r756;
xor.b32 %r757, %r483, %r171;
or.b32 %r758, %r757, -2139062144;
and.b32 %r759, %r171, 2139062143;
sub.s32 %r760, %r758, %r759;
xor.b32 %r178, %r496, %r760;
mul.lo.s64 %rd58, %rd57, 36;
add.s64 %rd59, %rd11, %rd58;
ld.global.nc.u32 %r179, [%rd59+8];
ld.global.nc.u32 %r175, [%rd59+4];

dp4a.s32.s32 %r173, %r174, %r175, %r445;


dp4a.s32.s32 %r177, %r178, %r179, %r173;

mov.u32 %r181, %r377;

prmt.b32 %r181, %r181, 0, 0xba98;

mov.u32 %r183, %r379;

prmt.b32 %r183, %r183, 0, 0xba98;

xor.b32 %r761, %r514, %r181;
or.b32 %r762, %r761, -2139062144;
and.b32 %r763, %r181, 2139062143;
sub.s32 %r764, %r762, %r763;
xor.b32 %r186, %r523, %r764;
xor.b32 %r765, %r515, %r183;
or.b32 %r766, %r765, -2139062144;
and.b32 %r767, %r183, 2139062143;
sub.s32 %r768, %r766, %r767;
xor.b32 %r190, %r528, %r768;
ld.global.nc.u32 %r191, [%rd59+16];
ld.global.nc.u32 %r187, [%rd59+12];

dp4a.s32.s32 %r185, %r186, %r187, %r177;


dp4a.s32.s32 %r189, %r190, %r191, %r185;

mov.u32 %r193, %r389;

prmt.b32 %r193, %r193, 0, 0xba98;

mov.u32 %r195, %r391;

prmt.b32 %r195, %r195, 0, 0xba98;

xor.b32 %r769, %r547, %r193;
or.b32 %r770, %r769, -2139062144;
and.b32 %r771, %r193, 2139062143;
sub.s32 %r772, %r770, %r771;
xor.b32 %r198, %r556, %r772;
xor.b32 %r773, %r548, %r195;
or.b32 %r774, %r773, -2139062144;
and.b32 %r775, %r195, 2139062143;
sub.s32 %r776, %r774, %r775;
xor.b32 %r202, %r561, %r776;
ld.global.nc.u32 %r203, [%rd59+24];
ld.global.nc.u32 %r199, [%rd59+20];

dp4a.s32.s32 %r197, %r198, %r199, %r445;


dp4a.s32.s32 %r201, %r202, %r203, %r197;

mov.u32 %r205, %r401;

prmt.b32 %r205, %r205, 0, 0xba98;

mov.u32 %r207, %r403;

prmt.b32 %r207, %r207, 0, 0xba98;

xor.b32 %r777, %r579, %r205;
or.b32 %r778, %r777, -2139062144;
and.b32 %r779, %r205, 2139062143;
sub.s32 %r780, %r778, %r779;
xor.b32 %r210, %r588, %r780;
xor.b32 %r781, %r580, %r207;
or.b32 %r782, %r781, -2139062144;
and.b32 %r783, %r207, 2139062143;
sub.s32 %r784, %r782, %r783;
xor.b32 %r214, %r593, %r784;
ld.global.nc.u32 %r215, [%rd59+32];
ld.global.nc.u32 %r211, [%rd59+28];

dp4a.s32.s32 %r209, %r210, %r211, %r201;


dp4a.s32.s32 %r213, %r214, %r215, %r209;

mul.lo.s32 %r785, %r189, %r595;
mad.lo.s32 %r786, %r213, %r596, %r785;
add.s32 %r787, %r213, %r189;
shr.u32 %r788, %r787, 31;
add.s32 %r789, %r787, %r788;
shr.s32 %r790, %r789, 1;
add.s32 %r791, %r786, %r790;
shr.s32 %r792, %r791, 31;
shr.u32 %r793, %r792, 30;
add.s32 %r794, %r791, %r793;
shr.s32 %r795, %r794, 2;

{ cvt.f32.f16 %f53, %rs1;}


ld.global.nc.u32 %r217, [%rd59];

{.reg .f16 low,high;
mov.b32 {low,high},%r217;
cvt.f32.f16 %f54, low;}


mul.ftz.f32 %f69, %f53, %f54;
cvt.rn.f32.s32 %f70, %r795;
fma.rn.ftz.f32 %f214, %f69, %f70, %f214;
mov.u32 %r218, %r414;

prmt.b32 %r218, %r218, 0, 0xba98;

mov.u32 %r220, %r416;

prmt.b32 %r220, %r220, 0, 0xba98;

xor.b32 %r796, %r627, %r218;
or.b32 %r797, %r796, -2139062144;
and.b32 %r798, %r218, 2139062143;
sub.s32 %r799, %r797, %r798;
xor.b32 %r223, %r636, %r799;
xor.b32 %r800, %r628, %r220;
or.b32 %r801, %r800, -2139062144;
and.b32 %r802, %r220, 2139062143;
sub.s32 %r803, %r801, %r802;
xor.b32 %r227, %r641, %r803;

dp4a.s32.s32 %r222, %r223, %r175, %r445;


dp4a.s32.s32 %r226, %r227, %r179, %r222;

mov.u32 %r230, %r426;

prmt.b32 %r230, %r230, 0, 0xba98;

mov.u32 %r232, %r428;

prmt.b32 %r232, %r232, 0, 0xba98;

xor.b32 %r804, %r659, %r230;
or.b32 %r805, %r804, -2139062144;
and.b32 %r806, %r230, 2139062143;
sub.s32 %r807, %r805, %r806;
xor.b32 %r235, %r668, %r807;
xor.b32 %r808, %r660, %r232;
or.b32 %r809, %r808, -2139062144;
and.b32 %r810, %r232, 2139062143;
sub.s32 %r811, %r809, %r810;
xor.b32 %r239, %r673, %r811;

dp4a.s32.s32 %r234, %r235, %r187, %r226;


dp4a.s32.s32 %r238, %r239, %r191, %r234;

mov.u32 %r242, %r438;

prmt.b32 %r242, %r242, 0, 0xba98;

mov.u32 %r244, %r440;

prmt.b32 %r244, %r244, 0, 0xba98;

xor.b32 %r812, %r692, %r242;
or.b32 %r813, %r812, -2139062144;
and.b32 %r814, %r242, 2139062143;
sub.s32 %r815, %r813, %r814;
xor.b32 %r247, %r701, %r815;
xor.b32 %r816, %r693, %r244;
or.b32 %r817, %r816, -2139062144;
and.b32 %r818, %r244, 2139062143;
sub.s32 %r819, %r817, %r818;
xor.b32 %r251, %r706, %r819;

dp4a.s32.s32 %r246, %r247, %r199, %r445;


dp4a.s32.s32 %r250, %r251, %r203, %r246;

mov.u32 %r254, %r450;

prmt.b32 %r254, %r254, 0, 0xba98;

mov.u32 %r256, %r452;

prmt.b32 %r256, %r256, 0, 0xba98;

xor.b32 %r820, %r724, %r254;
or.b32 %r821, %r820, -2139062144;
and.b32 %r822, %r254, 2139062143;
sub.s32 %r823, %r821, %r822;
xor.b32 %r259, %r733, %r823;
xor.b32 %r824, %r725, %r256;
or.b32 %r825, %r824, -2139062144;
and.b32 %r826, %r256, 2139062143;
sub.s32 %r827, %r825, %r826;
xor.b32 %r263, %r738, %r827;

dp4a.s32.s32 %r258, %r259, %r211, %r250;


dp4a.s32.s32 %r262, %r263, %r215, %r258;

mul.lo.s32 %r828, %r238, %r740;
mad.lo.s32 %r829, %r262, %r741, %r828;
add.s32 %r830, %r262, %r238;
shr.u32 %r831, %r830, 31;
add.s32 %r832, %r830, %r831;
shr.s32 %r833, %r832, 1;
add.s32 %r834, %r829, %r833;
shr.s32 %r835, %r834, 31;
shr.u32 %r836, %r835, 30;
add.s32 %r837, %r834, %r836;
shr.s32 %r838, %r837, 2;

{ cvt.f32.f16 %f55, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r217;
cvt.f32.f16 %f56, low;}


mul.ftz.f32 %f71, %f55, %f56;
cvt.rn.f32.s32 %f72, %r838;
fma.rn.ftz.f32 %f213, %f71, %f72, %f213;
cvt.u64.u32 %rd60, %r1166;
add.s64 %rd61, %rd2, %rd60;
add.s64 %rd62, %rd61, %rd4;
mov.u32 %r267, %r365;

prmt.b32 %r267, %r267, 0, 0xba98;

mov.u32 %r269, %r367;

prmt.b32 %r269, %r269, 0, 0xba98;

xor.b32 %r839, %r482, %r267;
or.b32 %r840, %r839, -2139062144;
and.b32 %r841, %r267, 2139062143;
sub.s32 %r842, %r840, %r841;
xor.b32 %r272, %r491, %r842;
xor.b32 %r843, %r483, %r269;
or.b32 %r844, %r843, -2139062144;
and.b32 %r845, %r269, 2139062143;
sub.s32 %r846, %r844, %r845;
xor.b32 %r276, %r496, %r846;
mul.lo.s64 %rd63, %rd62, 36;
add.s64 %rd64, %rd11, %rd63;
ld.global.nc.u32 %r277, [%rd64+8];
ld.global.nc.u32 %r273, [%rd64+4];

dp4a.s32.s32 %r271, %r272, %r273, %r445;


dp4a.s32.s32 %r275, %r276, %r277, %r271;

mov.u32 %r279, %r377;

prmt.b32 %r279, %r279, 0, 0xba98;

mov.u32 %r281, %r379;

prmt.b32 %r281, %r281, 0, 0xba98;

xor.b32 %r847, %r514, %r279;
or.b32 %r848, %r847, -2139062144;
and.b32 %r849, %r279, 2139062143;
sub.s32 %r850, %r848, %r849;
xor.b32 %r284, %r523, %r850;
xor.b32 %r851, %r515, %r281;
or.b32 %r852, %r851, -2139062144;
and.b32 %r853, %r281, 2139062143;
sub.s32 %r854, %r852, %r853;
xor.b32 %r288, %r528, %r854;
ld.global.nc.u32 %r289, [%rd64+16];
ld.global.nc.u32 %r285, [%rd64+12];

dp4a.s32.s32 %r283, %r284, %r285, %r275;


dp4a.s32.s32 %r287, %r288, %r289, %r283;

mov.u32 %r291, %r389;

prmt.b32 %r291, %r291, 0, 0xba98;

mov.u32 %r293, %r391;

prmt.b32 %r293, %r293, 0, 0xba98;

xor.b32 %r855, %r547, %r291;
or.b32 %r856, %r855, -2139062144;
and.b32 %r857, %r291, 2139062143;
sub.s32 %r858, %r856, %r857;
xor.b32 %r296, %r556, %r858;
xor.b32 %r859, %r548, %r293;
or.b32 %r860, %r859, -2139062144;
and.b32 %r861, %r293, 2139062143;
sub.s32 %r862, %r860, %r861;
xor.b32 %r300, %r561, %r862;
ld.global.nc.u32 %r301, [%rd64+24];
ld.global.nc.u32 %r297, [%rd64+20];

dp4a.s32.s32 %r295, %r296, %r297, %r445;


dp4a.s32.s32 %r299, %r300, %r301, %r295;

mov.u32 %r303, %r401;

prmt.b32 %r303, %r303, 0, 0xba98;

mov.u32 %r305, %r403;

prmt.b32 %r305, %r305, 0, 0xba98;

xor.b32 %r863, %r579, %r303;
or.b32 %r864, %r863, -2139062144;
and.b32 %r865, %r303, 2139062143;
sub.s32 %r866, %r864, %r865;
xor.b32 %r308, %r588, %r866;
xor.b32 %r867, %r580, %r305;
or.b32 %r868, %r867, -2139062144;
and.b32 %r869, %r305, 2139062143;
sub.s32 %r870, %r868, %r869;
xor.b32 %r312, %r593, %r870;
ld.global.nc.u32 %r313, [%rd64+32];
ld.global.nc.u32 %r309, [%rd64+28];

dp4a.s32.s32 %r307, %r308, %r309, %r299;


dp4a.s32.s32 %r311, %r312, %r313, %r307;

mul.lo.s32 %r871, %r287, %r595;
mad.lo.s32 %r872, %r311, %r596, %r871;
add.s32 %r873, %r311, %r287;
shr.u32 %r874, %r873, 31;
add.s32 %r875, %r873, %r874;
shr.s32 %r876, %r875, 1;
add.s32 %r877, %r872, %r876;
shr.s32 %r878, %r877, 31;
shr.u32 %r879, %r878, 30;
add.s32 %r880, %r877, %r879;
shr.s32 %r881, %r880, 2;

{ cvt.f32.f16 %f57, %rs1;}


ld.global.nc.u32 %r315, [%rd64];

{.reg .f16 low,high;
mov.b32 {low,high},%r315;
cvt.f32.f16 %f58, low;}


mul.ftz.f32 %f73, %f57, %f58;
cvt.rn.f32.s32 %f74, %r881;
fma.rn.ftz.f32 %f212, %f73, %f74, %f212;
mov.u32 %r316, %r414;

prmt.b32 %r316, %r316, 0, 0xba98;

mov.u32 %r318, %r416;

prmt.b32 %r318, %r318, 0, 0xba98;

xor.b32 %r882, %r627, %r316;
or.b32 %r883, %r882, -2139062144;
and.b32 %r884, %r316, 2139062143;
sub.s32 %r885, %r883, %r884;
xor.b32 %r321, %r636, %r885;
xor.b32 %r886, %r628, %r318;
or.b32 %r887, %r886, -2139062144;
and.b32 %r888, %r318, 2139062143;
sub.s32 %r889, %r887, %r888;
xor.b32 %r325, %r641, %r889;

dp4a.s32.s32 %r320, %r321, %r273, %r445;


dp4a.s32.s32 %r324, %r325, %r277, %r320;

mov.u32 %r328, %r426;

prmt.b32 %r328, %r328, 0, 0xba98;

mov.u32 %r330, %r428;

prmt.b32 %r330, %r330, 0, 0xba98;

xor.b32 %r890, %r659, %r328;
or.b32 %r891, %r890, -2139062144;
and.b32 %r892, %r328, 2139062143;
sub.s32 %r893, %r891, %r892;
xor.b32 %r333, %r668, %r893;
xor.b32 %r894, %r660, %r330;
or.b32 %r895, %r894, -2139062144;
and.b32 %r896, %r330, 2139062143;
sub.s32 %r897, %r895, %r896;
xor.b32 %r337, %r673, %r897;

dp4a.s32.s32 %r332, %r333, %r285, %r324;


dp4a.s32.s32 %r336, %r337, %r289, %r332;

mov.u32 %r340, %r438;

prmt.b32 %r340, %r340, 0, 0xba98;

mov.u32 %r342, %r440;

prmt.b32 %r342, %r342, 0, 0xba98;

xor.b32 %r898, %r692, %r340;
or.b32 %r899, %r898, -2139062144;
and.b32 %r900, %r340, 2139062143;
sub.s32 %r901, %r899, %r900;
xor.b32 %r345, %r701, %r901;
xor.b32 %r902, %r693, %r342;
or.b32 %r903, %r902, -2139062144;
and.b32 %r904, %r342, 2139062143;
sub.s32 %r905, %r903, %r904;
xor.b32 %r349, %r706, %r905;

dp4a.s32.s32 %r344, %r345, %r297, %r445;


dp4a.s32.s32 %r348, %r349, %r301, %r344;

mov.u32 %r352, %r450;

prmt.b32 %r352, %r352, 0, 0xba98;

mov.u32 %r354, %r452;

prmt.b32 %r354, %r354, 0, 0xba98;

xor.b32 %r906, %r724, %r352;
or.b32 %r907, %r906, -2139062144;
and.b32 %r908, %r352, 2139062143;
sub.s32 %r909, %r907, %r908;
xor.b32 %r357, %r733, %r909;
xor.b32 %r910, %r725, %r354;
or.b32 %r911, %r910, -2139062144;
and.b32 %r912, %r354, 2139062143;
sub.s32 %r913, %r911, %r912;
xor.b32 %r361, %r738, %r913;

dp4a.s32.s32 %r356, %r357, %r309, %r348;


dp4a.s32.s32 %r360, %r361, %r313, %r356;

mul.lo.s32 %r914, %r336, %r740;
mad.lo.s32 %r915, %r360, %r741, %r914;
add.s32 %r916, %r360, %r336;
shr.u32 %r917, %r916, 31;
add.s32 %r918, %r916, %r917;
shr.s32 %r919, %r918, 1;
add.s32 %r920, %r915, %r919;
shr.s32 %r921, %r920, 31;
shr.u32 %r922, %r921, 30;
add.s32 %r923, %r920, %r922;
shr.s32 %r924, %r923, 2;

{ cvt.f32.f16 %f59, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r315;
cvt.f32.f16 %f60, low;}


mul.ftz.f32 %f75, %f59, %f60;
cvt.rn.f32.s32 %f76, %r924;
fma.rn.ftz.f32 %f211, %f75, %f76, %f211;
cvt.u64.u32 %rd65, %r1165;
add.s64 %rd66, %rd2, %rd65;
add.s64 %rd67, %rd66, %rd4;

prmt.b32 %r365, %r365, 0, 0xba98;


prmt.b32 %r367, %r367, 0, 0xba98;

xor.b32 %r925, %r482, %r365;
or.b32 %r926, %r925, -2139062144;
and.b32 %r927, %r365, 2139062143;
sub.s32 %r928, %r926, %r927;
xor.b32 %r370, %r491, %r928;
xor.b32 %r929, %r483, %r367;
or.b32 %r930, %r929, -2139062144;
and.b32 %r931, %r367, 2139062143;
sub.s32 %r932, %r930, %r931;
xor.b32 %r374, %r496, %r932;
mul.lo.s64 %rd68, %rd67, 36;
add.s64 %rd69, %rd11, %rd68;
ld.global.nc.u32 %r375, [%rd69+8];
ld.global.nc.u32 %r371, [%rd69+4];

dp4a.s32.s32 %r369, %r370, %r371, %r445;


dp4a.s32.s32 %r373, %r374, %r375, %r369;


prmt.b32 %r377, %r377, 0, 0xba98;


prmt.b32 %r379, %r379, 0, 0xba98;

xor.b32 %r933, %r514, %r377;
or.b32 %r934, %r933, -2139062144;
and.b32 %r935, %r377, 2139062143;
sub.s32 %r936, %r934, %r935;
xor.b32 %r382, %r523, %r936;
xor.b32 %r937, %r515, %r379;
or.b32 %r938, %r937, -2139062144;
and.b32 %r939, %r379, 2139062143;
sub.s32 %r940, %r938, %r939;
xor.b32 %r386, %r528, %r940;
ld.global.nc.u32 %r387, [%rd69+16];
ld.global.nc.u32 %r383, [%rd69+12];

dp4a.s32.s32 %r381, %r382, %r383, %r373;


dp4a.s32.s32 %r385, %r386, %r387, %r381;


prmt.b32 %r389, %r389, 0, 0xba98;


prmt.b32 %r391, %r391, 0, 0xba98;

xor.b32 %r941, %r547, %r389;
or.b32 %r942, %r941, -2139062144;
and.b32 %r943, %r389, 2139062143;
sub.s32 %r944, %r942, %r943;
xor.b32 %r394, %r556, %r944;
xor.b32 %r945, %r548, %r391;
or.b32 %r946, %r945, -2139062144;
and.b32 %r947, %r391, 2139062143;
sub.s32 %r948, %r946, %r947;
xor.b32 %r398, %r561, %r948;
ld.global.nc.u32 %r399, [%rd69+24];
ld.global.nc.u32 %r395, [%rd69+20];

dp4a.s32.s32 %r393, %r394, %r395, %r445;


dp4a.s32.s32 %r397, %r398, %r399, %r393;


prmt.b32 %r401, %r401, 0, 0xba98;


prmt.b32 %r403, %r403, 0, 0xba98;

xor.b32 %r949, %r579, %r401;
or.b32 %r950, %r949, -2139062144;
and.b32 %r951, %r401, 2139062143;
sub.s32 %r952, %r950, %r951;
xor.b32 %r406, %r588, %r952;
xor.b32 %r953, %r580, %r403;
or.b32 %r954, %r953, -2139062144;
and.b32 %r955, %r403, 2139062143;
sub.s32 %r956, %r954, %r955;
xor.b32 %r410, %r593, %r956;
ld.global.nc.u32 %r411, [%rd69+32];
ld.global.nc.u32 %r407, [%rd69+28];

dp4a.s32.s32 %r405, %r406, %r407, %r397;


dp4a.s32.s32 %r409, %r410, %r411, %r405;

mul.lo.s32 %r957, %r385, %r595;
mad.lo.s32 %r958, %r409, %r596, %r957;
add.s32 %r959, %r409, %r385;
shr.u32 %r960, %r959, 31;
add.s32 %r961, %r959, %r960;
shr.s32 %r962, %r961, 1;
add.s32 %r963, %r958, %r962;
shr.s32 %r964, %r963, 31;
shr.u32 %r965, %r964, 30;
add.s32 %r966, %r963, %r965;
shr.s32 %r967, %r966, 2;

{ cvt.f32.f16 %f61, %rs1;}


ld.global.nc.u32 %r413, [%rd69];

{.reg .f16 low,high;
mov.b32 {low,high},%r413;
cvt.f32.f16 %f62, low;}


mul.ftz.f32 %f77, %f61, %f62;
cvt.rn.f32.s32 %f78, %r967;
fma.rn.ftz.f32 %f210, %f77, %f78, %f210;

prmt.b32 %r414, %r414, 0, 0xba98;


prmt.b32 %r416, %r416, 0, 0xba98;

xor.b32 %r968, %r627, %r414;
or.b32 %r969, %r968, -2139062144;
and.b32 %r970, %r414, 2139062143;
sub.s32 %r971, %r969, %r970;
xor.b32 %r419, %r636, %r971;
xor.b32 %r972, %r628, %r416;
or.b32 %r973, %r972, -2139062144;
and.b32 %r974, %r416, 2139062143;
sub.s32 %r975, %r973, %r974;
xor.b32 %r423, %r641, %r975;

dp4a.s32.s32 %r418, %r419, %r371, %r445;


dp4a.s32.s32 %r422, %r423, %r375, %r418;


prmt.b32 %r426, %r426, 0, 0xba98;


prmt.b32 %r428, %r428, 0, 0xba98;

xor.b32 %r976, %r659, %r426;
or.b32 %r977, %r976, -2139062144;
and.b32 %r978, %r426, 2139062143;
sub.s32 %r979, %r977, %r978;
xor.b32 %r431, %r668, %r979;
xor.b32 %r980, %r660, %r428;
or.b32 %r981, %r980, -2139062144;
and.b32 %r982, %r428, 2139062143;
sub.s32 %r983, %r981, %r982;
xor.b32 %r435, %r673, %r983;

dp4a.s32.s32 %r430, %r431, %r383, %r422;


dp4a.s32.s32 %r434, %r435, %r387, %r430;


prmt.b32 %r438, %r438, 0, 0xba98;


prmt.b32 %r440, %r440, 0, 0xba98;

xor.b32 %r984, %r692, %r438;
or.b32 %r985, %r984, -2139062144;
and.b32 %r986, %r438, 2139062143;
sub.s32 %r987, %r985, %r986;
xor.b32 %r443, %r701, %r987;
xor.b32 %r988, %r693, %r440;
or.b32 %r989, %r988, -2139062144;
and.b32 %r990, %r440, 2139062143;
sub.s32 %r991, %r989, %r990;
xor.b32 %r447, %r706, %r991;

dp4a.s32.s32 %r442, %r443, %r395, %r445;


dp4a.s32.s32 %r446, %r447, %r399, %r442;


prmt.b32 %r450, %r450, 0, 0xba98;


prmt.b32 %r452, %r452, 0, 0xba98;

xor.b32 %r992, %r724, %r450;
or.b32 %r993, %r992, -2139062144;
and.b32 %r994, %r450, 2139062143;
sub.s32 %r995, %r993, %r994;
xor.b32 %r455, %r733, %r995;
xor.b32 %r996, %r725, %r452;
or.b32 %r997, %r996, -2139062144;
and.b32 %r998, %r452, 2139062143;
sub.s32 %r999, %r997, %r998;
xor.b32 %r459, %r738, %r999;

dp4a.s32.s32 %r454, %r455, %r407, %r446;


dp4a.s32.s32 %r458, %r459, %r411, %r454;

mul.lo.s32 %r1000, %r434, %r740;
mad.lo.s32 %r1001, %r458, %r741, %r1000;
add.s32 %r1002, %r458, %r434;
shr.u32 %r1003, %r1002, 31;
add.s32 %r1004, %r1002, %r1003;
shr.s32 %r1005, %r1004, 1;
add.s32 %r1006, %r1001, %r1005;
shr.s32 %r1007, %r1006, 31;
shr.u32 %r1008, %r1007, 30;
add.s32 %r1009, %r1006, %r1008;
shr.s32 %r1010, %r1009, 2;

{ cvt.f32.f16 %f63, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r413;
cvt.f32.f16 %f64, low;}


mul.ftz.f32 %f79, %f63, %f64;
cvt.rn.f32.s32 %f80, %r1010;
fma.rn.ftz.f32 %f209, %f79, %f80, %f209;
add.s32 %r1166, %r1166, 128;
add.s32 %r1165, %r1165, 128;
add.s32 %r1164, %r1164, 128;
add.s32 %r1163, %r1163, 128;
add.s64 %rd90, %rd90, 1312;
add.s32 %r1167, %r1167, 16;
setp.lt.s32 %p4, %r1167, %r6;
@%p4 bra $L__BB232_2;

$L__BB232_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB232_5;

shl.b32 %r1011, %r1, 10;
mov.u32 %r1012, _ZZ13mul_mat_vec_qIL9ggml_type22ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r1013, %r1012, %r1011;
shl.b32 %r1014, %r2, 2;
add.s32 %r1015, %r1013, %r1014;
st.shared.f32 [%r1015+-1024], %f216;
st.shared.f32 [%r1015+-896], %f215;
st.shared.f32 [%r1015+-768], %f214;
st.shared.f32 [%r1015+-640], %f213;
st.shared.f32 [%r1015+-512], %f212;
st.shared.f32 [%r1015+-384], %f211;
st.shared.f32 [%r1015+-256], %f210;
st.shared.f32 [%r1015+-128], %f209;

$L__BB232_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB232_22;

mov.u32 %r1016, %ctaid.z;
mov.u32 %r1017, %ctaid.y;
mad.lo.s32 %r1018, %r1017, %r32, %r4;
mad.lo.s32 %r1019, %r1016, %r38, %r1018;
cvt.u64.u32 %rd14, %r1019;
shl.b32 %r1020, %r2, 2;
mov.u32 %r1021, _ZZ13mul_mat_vec_qIL9ggml_type22ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r21, %r1021, %r1020;
mov.u32 %r1022, 2;
ld.shared.f32 %f81, [%r21];
add.ftz.f32 %f82, %f81, %f216;
ld.shared.f32 %f83, [%r21+1024];
add.ftz.f32 %f84, %f83, %f82;
ld.shared.f32 %f85, [%r21+2048];
add.ftz.f32 %f86, %f85, %f84;
mov.b32 %r1023, %f86;
mov.u32 %r1024, 31;
mov.u32 %r1025, 16;
mov.u32 %r1026, -1;
shfl.sync.bfly.b32 %r1027|%p7, %r1023, %r1025, %r1024, %r1026;
mov.b32 %f87, %r1027;
add.ftz.f32 %f88, %f86, %f87;
mov.b32 %r1028, %f88;
mov.u32 %r1029, 8;
shfl.sync.bfly.b32 %r1030|%p8, %r1028, %r1029, %r1024, %r1026;
mov.b32 %f89, %r1030;
add.ftz.f32 %f90, %f88, %f89;
mov.b32 %r1031, %f90;
mov.u32 %r1032, 4;
shfl.sync.bfly.b32 %r1033|%p9, %r1031, %r1032, %r1024, %r1026;
mov.b32 %f91, %r1033;
add.ftz.f32 %f92, %f90, %f91;
mov.b32 %r1034, %f92;
shfl.sync.bfly.b32 %r1035|%p10, %r1034, %r1022, %r1024, %r1026;
mov.b32 %f93, %r1035;
add.ftz.f32 %f94, %f92, %f93;
mov.b32 %r1036, %f94;
mov.u32 %r1037, 1;
shfl.sync.bfly.b32 %r1038|%p11, %r1036, %r1037, %r1024, %r1026;
mov.b32 %f95, %r1038;
add.ftz.f32 %f25, %f94, %f95;
setp.lt.u32 %p12, %r4, %r26;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
mul.wide.u32 %rd70, %r1019, 4;
add.s64 %rd15, %rd1, %rd70;
not.pred %p14, %p1;
@%p14 bra $L__BB232_8;

st.global.f32 [%rd15], %f25;

$L__BB232_8:
ld.shared.f32 %f96, [%r21+128];
add.ftz.f32 %f97, %f96, %f215;
ld.shared.f32 %f98, [%r21+1152];
add.ftz.f32 %f99, %f98, %f97;
ld.shared.f32 %f100, [%r21+2176];
add.ftz.f32 %f101, %f100, %f99;
mov.b32 %r1039, %f101;
shfl.sync.bfly.b32 %r1043|%p15, %r1039, %r1025, %r1024, %r1026;
mov.b32 %f102, %r1043;
add.ftz.f32 %f103, %f101, %f102;
mov.b32 %r1044, %f103;
shfl.sync.bfly.b32 %r1046|%p16, %r1044, %r1029, %r1024, %r1026;
mov.b32 %f104, %r1046;
add.ftz.f32 %f105, %f103, %f104;
mov.b32 %r1047, %f105;
shfl.sync.bfly.b32 %r1049|%p17, %r1047, %r1032, %r1024, %r1026;
mov.b32 %f106, %r1049;
add.ftz.f32 %f107, %f105, %f106;
mov.b32 %r1050, %f107;
shfl.sync.bfly.b32 %r1052|%p18, %r1050, %r1022, %r1024, %r1026;
mov.b32 %f108, %r1052;
add.ftz.f32 %f109, %f107, %f108;
mov.b32 %r1053, %f109;
shfl.sync.bfly.b32 %r1055|%p19, %r1053, %r1037, %r1024, %r1026;
mov.b32 %f110, %r1055;
add.ftz.f32 %f26, %f109, %f110;
setp.eq.s32 %p20, %r2, 1;
add.s32 %r1057, %r4, 1;
setp.lt.u32 %p21, %r1057, %r26;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB232_10;

st.global.f32 [%rd15+4], %f26;

$L__BB232_10:
ld.shared.f32 %f111, [%r21+256];
add.ftz.f32 %f112, %f111, %f214;
ld.shared.f32 %f113, [%r21+1280];
add.ftz.f32 %f114, %f113, %f112;
ld.shared.f32 %f115, [%r21+2304];
add.ftz.f32 %f116, %f115, %f114;
mov.b32 %r1058, %f116;
mov.u32 %r1059, 31;
mov.u32 %r1060, 16;
mov.u32 %r1061, -1;
shfl.sync.bfly.b32 %r1062|%p23, %r1058, %r1060, %r1059, %r1061;
mov.b32 %f117, %r1062;
add.ftz.f32 %f118, %f116, %f117;
mov.b32 %r1063, %f118;
mov.u32 %r1064, 8;
shfl.sync.bfly.b32 %r1065|%p24, %r1063, %r1064, %r1059, %r1061;
mov.b32 %f119, %r1065;
add.ftz.f32 %f120, %f118, %f119;
mov.b32 %r1066, %f120;
mov.u32 %r1067, 4;
shfl.sync.bfly.b32 %r1068|%p25, %r1066, %r1067, %r1059, %r1061;
mov.b32 %f121, %r1068;
add.ftz.f32 %f122, %f120, %f121;
mov.b32 %r1069, %f122;
mov.u32 %r1070, 2;
shfl.sync.bfly.b32 %r1071|%p26, %r1069, %r1070, %r1059, %r1061;
mov.b32 %f123, %r1071;
add.ftz.f32 %f124, %f122, %f123;
mov.b32 %r1072, %f124;
mov.u32 %r1073, 1;
shfl.sync.bfly.b32 %r1074|%p27, %r1072, %r1073, %r1059, %r1061;
mov.b32 %f125, %r1074;
add.ftz.f32 %f27, %f124, %f125;
@%p14 bra $L__BB232_12;

cvt.u64.u32 %rd71, %r26;
add.s64 %rd72, %rd71, %rd14;
shl.b64 %rd73, %rd72, 2;
add.s64 %rd74, %rd1, %rd73;
st.global.f32 [%rd74], %f27;

$L__BB232_12:
ld.shared.f32 %f126, [%r21+384];
add.ftz.f32 %f127, %f126, %f213;
ld.shared.f32 %f128, [%r21+1408];
add.ftz.f32 %f129, %f128, %f127;
ld.shared.f32 %f130, [%r21+2432];
add.ftz.f32 %f131, %f130, %f129;
mov.b32 %r1075, %f131;
shfl.sync.bfly.b32 %r1079|%p29, %r1075, %r1060, %r1059, %r1061;
mov.b32 %f132, %r1079;
add.ftz.f32 %f133, %f131, %f132;
mov.b32 %r1080, %f133;
shfl.sync.bfly.b32 %r1082|%p30, %r1080, %r1064, %r1059, %r1061;
mov.b32 %f134, %r1082;
add.ftz.f32 %f135, %f133, %f134;
mov.b32 %r1083, %f135;
shfl.sync.bfly.b32 %r1085|%p31, %r1083, %r1067, %r1059, %r1061;
mov.b32 %f136, %r1085;
add.ftz.f32 %f137, %f135, %f136;
mov.b32 %r1086, %f137;
shfl.sync.bfly.b32 %r1088|%p32, %r1086, %r1070, %r1059, %r1061;
mov.b32 %f138, %r1088;
add.ftz.f32 %f139, %f137, %f138;
mov.b32 %r1089, %f139;
shfl.sync.bfly.b32 %r1091|%p33, %r1089, %r1073, %r1059, %r1061;
mov.b32 %f140, %r1091;
add.ftz.f32 %f28, %f139, %f140;
@%p22 bra $L__BB232_14;

add.s32 %r1092, %r26, 1;
cvt.u64.u32 %rd75, %r1092;
add.s64 %rd76, %rd75, %rd14;
shl.b64 %rd77, %rd76, 2;
add.s64 %rd78, %rd1, %rd77;
st.global.f32 [%rd78], %f28;

$L__BB232_14:
shl.b32 %r1093, %r26, 1;
mov.u32 %r1094, 1;
ld.shared.f32 %f141, [%r21+512];
add.ftz.f32 %f142, %f141, %f212;
ld.shared.f32 %f143, [%r21+1536];
add.ftz.f32 %f144, %f143, %f142;
ld.shared.f32 %f145, [%r21+2560];
add.ftz.f32 %f146, %f145, %f144;
mov.b32 %r1095, %f146;
mov.u32 %r1096, 31;
mov.u32 %r1097, 16;
mov.u32 %r1098, -1;
shfl.sync.bfly.b32 %r1099|%p35, %r1095, %r1097, %r1096, %r1098;
mov.b32 %f147, %r1099;
add.ftz.f32 %f148, %f146, %f147;
mov.b32 %r1100, %f148;
mov.u32 %r1101, 8;
shfl.sync.bfly.b32 %r1102|%p36, %r1100, %r1101, %r1096, %r1098;
mov.b32 %f149, %r1102;
add.ftz.f32 %f150, %f148, %f149;
mov.b32 %r1103, %f150;
mov.u32 %r1104, 4;
shfl.sync.bfly.b32 %r1105|%p37, %r1103, %r1104, %r1096, %r1098;
mov.b32 %f151, %r1105;
add.ftz.f32 %f152, %f150, %f151;
mov.b32 %r1106, %f152;
mov.u32 %r1107, 2;
shfl.sync.bfly.b32 %r1108|%p38, %r1106, %r1107, %r1096, %r1098;
mov.b32 %f153, %r1108;
add.ftz.f32 %f154, %f152, %f153;
mov.b32 %r1109, %f154;
shfl.sync.bfly.b32 %r1110|%p39, %r1109, %r1094, %r1096, %r1098;
mov.b32 %f155, %r1110;
add.ftz.f32 %f29, %f154, %f155;
cvt.u64.u32 %rd79, %r1093;
add.s64 %rd80, %rd79, %rd14;
shl.b64 %rd81, %rd80, 2;
add.s64 %rd16, %rd1, %rd81;
@%p14 bra $L__BB232_16;

st.global.f32 [%rd16], %f29;

$L__BB232_16:
ld.shared.f32 %f156, [%r21+640];
add.ftz.f32 %f157, %f156, %f211;
ld.shared.f32 %f158, [%r21+1664];
add.ftz.f32 %f159, %f158, %f157;
ld.shared.f32 %f160, [%r21+2688];
add.ftz.f32 %f161, %f160, %f159;
mov.b32 %r1111, %f161;
shfl.sync.bfly.b32 %r1115|%p41, %r1111, %r1097, %r1096, %r1098;
mov.b32 %f162, %r1115;
add.ftz.f32 %f163, %f161, %f162;
mov.b32 %r1116, %f163;
shfl.sync.bfly.b32 %r1118|%p42, %r1116, %r1101, %r1096, %r1098;
mov.b32 %f164, %r1118;
add.ftz.f32 %f165, %f163, %f164;
mov.b32 %r1119, %f165;
shfl.sync.bfly.b32 %r1121|%p43, %r1119, %r1104, %r1096, %r1098;
mov.b32 %f166, %r1121;
add.ftz.f32 %f167, %f165, %f166;
mov.b32 %r1122, %f167;
shfl.sync.bfly.b32 %r1124|%p44, %r1122, %r1107, %r1096, %r1098;
mov.b32 %f168, %r1124;
add.ftz.f32 %f169, %f167, %f168;
mov.b32 %r1125, %f169;
shfl.sync.bfly.b32 %r1127|%p45, %r1125, %r1094, %r1096, %r1098;
mov.b32 %f170, %r1127;
add.ftz.f32 %f30, %f169, %f170;
@%p22 bra $L__BB232_18;

st.global.f32 [%rd16+4], %f30;

$L__BB232_18:
mul.lo.s32 %r23, %r26, 3;
ld.shared.f32 %f171, [%r21+768];
add.ftz.f32 %f172, %f171, %f210;
ld.shared.f32 %f173, [%r21+1792];
add.ftz.f32 %f174, %f173, %f172;
ld.shared.f32 %f175, [%r21+2816];
add.ftz.f32 %f176, %f175, %f174;
mov.b32 %r1128, %f176;
mov.u32 %r1129, 31;
mov.u32 %r1130, 16;
mov.u32 %r1131, -1;
shfl.sync.bfly.b32 %r1132|%p47, %r1128, %r1130, %r1129, %r1131;
mov.b32 %f177, %r1132;
add.ftz.f32 %f178, %f176, %f177;
mov.b32 %r1133, %f178;
mov.u32 %r1134, 8;
shfl.sync.bfly.b32 %r1135|%p48, %r1133, %r1134, %r1129, %r1131;
mov.b32 %f179, %r1135;
add.ftz.f32 %f180, %f178, %f179;
mov.b32 %r1136, %f180;
mov.u32 %r1137, 4;
shfl.sync.bfly.b32 %r1138|%p49, %r1136, %r1137, %r1129, %r1131;
mov.b32 %f181, %r1138;
add.ftz.f32 %f182, %f180, %f181;
mov.b32 %r1139, %f182;
mov.u32 %r1140, 2;
shfl.sync.bfly.b32 %r1141|%p50, %r1139, %r1140, %r1129, %r1131;
mov.b32 %f183, %r1141;
add.ftz.f32 %f184, %f182, %f183;
mov.b32 %r1142, %f184;
mov.u32 %r1143, 1;
shfl.sync.bfly.b32 %r1144|%p51, %r1142, %r1143, %r1129, %r1131;
mov.b32 %f185, %r1144;
add.ftz.f32 %f31, %f184, %f185;
@%p14 bra $L__BB232_20;

cvt.u64.u32 %rd82, %r23;
add.s64 %rd83, %rd82, %rd14;
shl.b64 %rd84, %rd83, 2;
add.s64 %rd85, %rd1, %rd84;
st.global.f32 [%rd85], %f31;

$L__BB232_20:
ld.shared.f32 %f186, [%r21+896];
add.ftz.f32 %f187, %f186, %f209;
ld.shared.f32 %f188, [%r21+1920];
add.ftz.f32 %f189, %f188, %f187;
ld.shared.f32 %f190, [%r21+2944];
add.ftz.f32 %f191, %f190, %f189;
mov.b32 %r1145, %f191;
shfl.sync.bfly.b32 %r1149|%p53, %r1145, %r1130, %r1129, %r1131;
mov.b32 %f192, %r1149;
add.ftz.f32 %f193, %f191, %f192;
mov.b32 %r1150, %f193;
shfl.sync.bfly.b32 %r1152|%p54, %r1150, %r1134, %r1129, %r1131;
mov.b32 %f194, %r1152;
add.ftz.f32 %f195, %f193, %f194;
mov.b32 %r1153, %f195;
shfl.sync.bfly.b32 %r1155|%p55, %r1153, %r1137, %r1129, %r1131;
mov.b32 %f196, %r1155;
add.ftz.f32 %f197, %f195, %f196;
mov.b32 %r1156, %f197;
shfl.sync.bfly.b32 %r1158|%p56, %r1156, %r1140, %r1129, %r1131;
mov.b32 %f198, %r1158;
add.ftz.f32 %f199, %f197, %f198;
mov.b32 %r1159, %f199;
shfl.sync.bfly.b32 %r1161|%p57, %r1159, %r1143, %r1129, %r1131;
mov.b32 %f200, %r1161;
add.ftz.f32 %f32, %f199, %f200;
@%p22 bra $L__BB232_22;

add.s32 %r1162, %r23, 1;
cvt.u64.u32 %rd86, %r1162;
add.s64 %rd87, %rd86, %rd14;
shl.b64 %rd88, %rd87, 2;
add.s64 %rd89, %rd1, %rd88;
st.global.f32 [%rd89], %f32;

$L__BB232_22:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type22ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type22ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type22ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type22ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type22ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type22ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type22ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type22ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type22ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<71>;
.reg .b16 %rs<49>;
.reg .f32 %f<231>;
.reg .b32 %r<1391>;
.reg .b64 %rd<100>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type22ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1280];

ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type22ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type22ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd20, [_Z13mul_mat_vec_qIL9ggml_type22ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type22ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type22ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type22ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type22ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type22ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type22ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type22ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type22ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type22ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type22ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type22ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type22ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type22ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type22ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
cvta.to.global.u64 %rd1, %rd20;
mov.u32 %r1, %tid.y;
shl.b32 %r43, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r43, %r2;
mov.u32 %r44, %ctaid.x;
shl.b32 %r4, %r44, 1;
mov.u32 %r5, %ctaid.y;
shr.s32 %r45, %r3, 31;
shr.u32 %r46, %r45, 29;
add.s32 %r47, %r3, %r46;
shr.s32 %r1390, %r47, 3;
shr.u32 %r7, %r42, 8;
setp.ge.s32 %p3, %r1390, %r7;
mov.f32 %f221, 0f00000000;
mov.f32 %f222, %f221;
mov.f32 %f223, %f221;
mov.f32 %f224, %f221;
mov.f32 %f225, %f221;
mov.f32 %f226, %f221;
mov.f32 %f227, %f221;
mov.f32 %f228, %f221;
mov.f32 %f229, %f221;
mov.f32 %f230, %f221;
@%p3 bra $L__BB233_3;

mov.u32 %r48, %ctaid.z;
mul.lo.s32 %r49, %r5, %r34;
cvt.u64.u32 %rd21, %r49;
mul.lo.s32 %r50, %r48, %r40;
cvt.u64.u32 %rd22, %r50;
add.s64 %rd2, %rd22, %rd21;
mul.hi.u32 %r51, %r48, %r36;
add.s32 %r52, %r51, %r48;
shr.u32 %r53, %r52, %r37;
mul.lo.s32 %r54, %r53, %r39;
mul.hi.u32 %r55, %r5, %r30;
add.s32 %r56, %r55, %r5;
shr.u32 %r57, %r56, %r31;
mul.lo.s32 %r58, %r57, %r33;
cvta.to.global.u64 %rd99, %rd18;
and.b32 %r62, %r47, -8;
sub.s32 %r63, %r3, %r62;
shl.b32 %r64, %r63, 1;
cvt.s64.s32 %rd4, %r63;
shl.b32 %r65, %r28, 1;
shl.b32 %r1385, %r1390, 3;
add.s32 %r1389, %r1385, %r65;
mad.lo.s32 %r1388, %r28, 3, %r1385;
shl.b32 %r66, %r28, 2;
add.s32 %r1387, %r1385, %r66;
add.s32 %r1386, %r28, %r1385;
add.s32 %r67, %r1390, %r54;
add.s32 %r68, %r67, %r58;
add.s32 %r69, %r4, 1;
mad.lo.s32 %r70, %r27, %r69, %r68;
mul.wide.s32 %rd5, %r70, 82;
mad.lo.s32 %r71, %r4, %r27, %r1390;
add.s32 %r72, %r71, %r54;
add.s32 %r73, %r72, %r58;
mul.wide.s32 %rd6, %r73, 82;
mul.wide.s32 %rd23, %r64, 2;
add.s64 %rd24, %rd6, %rd23;
add.s64 %rd7, %rd24, 36;
add.s64 %rd25, %rd5, %rd4;
add.s64 %rd8, %rd25, 66;
add.s64 %rd26, %rd6, %rd4;
add.s64 %rd9, %rd26, 74;
add.s64 %rd27, %rd5, %rd23;
add.s64 %rd10, %rd27, 34;
cvta.to.global.u64 %rd11, %rd19;
mov.f32 %f221, 0f00000000;

$L__BB233_2:
cvt.u64.u32 %rd28, %r1385;
add.s64 %rd29, %rd2, %rd28;
add.s64 %rd30, %rd29, %rd4;
add.s64 %rd31, %rd99, %rd7;
ld.global.nc.u16 %rs11, [%rd31+-34];
shr.u16 %rs12, %rs11, 8;
cvt.u32.u16 %r564, %rs12;
ld.global.nc.u16 %rs13, [%rd31+-32];
shr.u16 %rs14, %rs13, 8;
cvt.u32.u16 %r565, %rs14;
add.s64 %rd32, %rd99, %rd9;
ld.global.nc.u8 %rs15, [%rd32+-8];
ld.global.nc.u16 %rs16, [%rd31+-2];
ld.global.nc.u16 %rs17, [%rd31];
ld.global.nc.u8 %rs18, [%rd32];
and.b16 %rs19, %rs18, 240;
and.b16 %rs20, %rs18, 15;
shr.u16 %rs21, %rs19, 4;
and.b16 %rs22, %rs11, 255;
cvt.u32.u16 %r566, %rs22;
shl.b16 %rs23, %rs15, 8;
cvt.u32.u16 %r567, %rs23;
and.b32 %r568, %r567, 768;
or.b32 %r569, %r568, %r566;
mul.wide.u32 %rd33, %r569, 8;
mov.u64 %rd34, iq2s_grid;
add.s64 %rd35, %rd34, %rd33;
and.b16 %rs24, %rs16, 255;
mul.wide.u16 %r570, %rs24, 128;
cvt.u32.u16 %r571, %rs16;
shl.b32 %r572, %r571, 21;
or.b32 %r573, %r570, %r572;
and.b32 %r574, %r573, 25166208;
or.b32 %r575, %r574, -2139062144;
add.s32 %r576, %r575, -16843009;
or.b32 %r466, %r576, %r574;
mov.u32 %r74, %r466;

prmt.b32 %r74, %r74, 0, 0xba98;

mul.wide.u16 %r577, %rs24, 8;
shl.b32 %r578, %r571, 17;
or.b32 %r579, %r577, %r578;
and.b32 %r580, %r579, 25166208;
or.b32 %r581, %r580, -2139062144;
add.s32 %r582, %r581, -16843009;
or.b32 %r468, %r582, %r580;
mov.u32 %r76, %r468;

prmt.b32 %r76, %r76, 0, 0xba98;

ld.global.nc.v2.u32 {%r583, %r584}, [%rd35];
not.b32 %r586, %r583;
mov.u32 %r546, 0;
not.b32 %r588, %r584;
xor.b32 %r589, %r583, %r74;
or.b32 %r590, %r589, -2139062144;
and.b32 %r591, %r74, 2139062143;
and.b32 %r592, %r586, -2139062144;
sub.s32 %r593, %r590, %r591;
xor.b32 %r79, %r592, %r593;
xor.b32 %r594, %r584, %r76;
or.b32 %r595, %r594, -2139062144;
and.b32 %r596, %r76, 2139062143;
and.b32 %r597, %r588, -2139062144;
sub.s32 %r598, %r595, %r596;
xor.b32 %r83, %r597, %r598;
mul.lo.s64 %rd36, %rd30, 36;
add.s64 %rd37, %rd11, %rd36;
ld.global.nc.u32 %r84, [%rd37+8];
ld.global.nc.u32 %r80, [%rd37+4];

dp4a.s32.s32 %r78, %r79, %r80, %r546;


dp4a.s32.s32 %r82, %r83, %r84, %r78;

mul.wide.u16 %r599, %rs15, 64;
and.b32 %r600, %r599, 768;
or.b32 %r601, %r600, %r564;
mul.wide.u32 %rd38, %r601, 8;
add.s64 %rd39, %rd34, %rd38;
shr.u16 %rs26, %rs16, 8;
mul.wide.u16 %r602, %rs26, 128;
cvt.u32.u16 %r603, %rs26;
shl.b32 %r604, %r603, 21;
or.b32 %r605, %r602, %r604;
and.b32 %r606, %r605, 25166208;
or.b32 %r607, %r606, -2139062144;
add.s32 %r608, %r607, -16843009;
or.b32 %r478, %r608, %r606;
mov.u32 %r86, %r478;

prmt.b32 %r86, %r86, 0, 0xba98;

mul.wide.u16 %r609, %rs26, 8;
shl.b32 %r610, %r603, 17;
or.b32 %r611, %r609, %r610;
and.b32 %r612, %r611, 25166208;
or.b32 %r613, %r612, -2139062144;
add.s32 %r614, %r613, -16843009;
or.b32 %r480, %r614, %r612;
mov.u32 %r88, %r480;

prmt.b32 %r88, %r88, 0, 0xba98;

ld.global.nc.v2.u32 {%r615, %r616}, [%rd39];
not.b32 %r618, %r615;
not.b32 %r620, %r616;
xor.b32 %r621, %r615, %r86;
or.b32 %r622, %r621, -2139062144;
and.b32 %r623, %r86, 2139062143;
and.b32 %r624, %r618, -2139062144;
sub.s32 %r625, %r622, %r623;
xor.b32 %r91, %r624, %r625;
xor.b32 %r626, %r616, %r88;
or.b32 %r627, %r626, -2139062144;
and.b32 %r628, %r88, 2139062143;
and.b32 %r629, %r620, -2139062144;
sub.s32 %r630, %r627, %r628;
xor.b32 %r95, %r629, %r630;
ld.global.nc.u32 %r96, [%rd37+16];
ld.global.nc.u32 %r92, [%rd37+12];

dp4a.s32.s32 %r90, %r91, %r92, %r82;


dp4a.s32.s32 %r94, %r95, %r96, %r90;

and.b16 %rs27, %rs13, 255;
cvt.u32.u16 %r631, %rs27;
mul.wide.u16 %r632, %rs15, 16;
and.b32 %r633, %r632, 768;
or.b32 %r634, %r633, %r631;
mul.wide.u32 %rd40, %r634, 8;
add.s64 %rd41, %rd34, %rd40;
and.b16 %rs28, %rs17, 255;
mul.wide.u16 %r635, %rs28, 128;
cvt.u32.u16 %r636, %rs17;
shl.b32 %r637, %r636, 21;
or.b32 %r638, %r635, %r637;
and.b32 %r639, %r638, 25166208;
or.b32 %r640, %r639, -2139062144;
add.s32 %r641, %r640, -16843009;
or.b32 %r490, %r641, %r639;
mov.u32 %r98, %r490;

prmt.b32 %r98, %r98, 0, 0xba98;

mul.wide.u16 %r642, %rs28, 8;
shl.b32 %r643, %r636, 17;
or.b32 %r644, %r642, %r643;
and.b32 %r645, %r644, 25166208;
or.b32 %r646, %r645, -2139062144;
add.s32 %r647, %r646, -16843009;
or.b32 %r492, %r647, %r645;
mov.u32 %r100, %r492;

prmt.b32 %r100, %r100, 0, 0xba98;

ld.global.nc.v2.u32 {%r648, %r649}, [%rd41];
not.b32 %r651, %r648;
not.b32 %r653, %r649;
xor.b32 %r654, %r648, %r98;
or.b32 %r655, %r654, -2139062144;
and.b32 %r656, %r98, 2139062143;
and.b32 %r657, %r651, -2139062144;
sub.s32 %r658, %r655, %r656;
xor.b32 %r103, %r657, %r658;
xor.b32 %r659, %r649, %r100;
or.b32 %r660, %r659, -2139062144;
and.b32 %r661, %r100, 2139062143;
and.b32 %r662, %r653, -2139062144;
sub.s32 %r663, %r660, %r661;
xor.b32 %r107, %r662, %r663;
ld.global.nc.u32 %r108, [%rd37+24];
ld.global.nc.u32 %r104, [%rd37+20];

dp4a.s32.s32 %r102, %r103, %r104, %r546;


dp4a.s32.s32 %r106, %r107, %r108, %r102;

mul.wide.u16 %r664, %rs15, 4;
and.b32 %r665, %r664, 768;
or.b32 %r666, %r665, %r565;
mul.wide.u32 %rd42, %r666, 8;
add.s64 %rd43, %rd34, %rd42;
shr.u16 %rs29, %rs17, 8;
mul.wide.u16 %r667, %rs29, 128;
cvt.u32.u16 %r668, %rs29;
shl.b32 %r669, %r668, 21;
or.b32 %r670, %r667, %r669;
and.b32 %r671, %r670, 25166208;
or.b32 %r672, %r671, -2139062144;
add.s32 %r673, %r672, -16843009;
or.b32 %r502, %r673, %r671;
mov.u32 %r110, %r502;

prmt.b32 %r110, %r110, 0, 0xba98;

mul.wide.u16 %r674, %rs29, 8;
shl.b32 %r675, %r668, 17;
or.b32 %r676, %r674, %r675;
and.b32 %r677, %r676, 25166208;
or.b32 %r678, %r677, -2139062144;
add.s32 %r679, %r678, -16843009;
or.b32 %r504, %r679, %r677;
mov.u32 %r112, %r504;

prmt.b32 %r112, %r112, 0, 0xba98;

ld.global.nc.v2.u32 {%r680, %r681}, [%rd43];
not.b32 %r683, %r680;
not.b32 %r685, %r681;
xor.b32 %r686, %r680, %r110;
or.b32 %r687, %r686, -2139062144;
and.b32 %r688, %r110, 2139062143;
and.b32 %r689, %r683, -2139062144;
sub.s32 %r690, %r687, %r688;
xor.b32 %r115, %r689, %r690;
xor.b32 %r691, %r681, %r112;
or.b32 %r692, %r691, -2139062144;
and.b32 %r693, %r112, 2139062143;
and.b32 %r694, %r685, -2139062144;
sub.s32 %r695, %r692, %r693;
xor.b32 %r119, %r694, %r695;
ld.global.nc.u32 %r120, [%rd37+32];
ld.global.nc.u32 %r116, [%rd37+28];

dp4a.s32.s32 %r114, %r115, %r116, %r106;


dp4a.s32.s32 %r118, %r119, %r120, %r114;

cvt.u32.u16 %r696, %rs20;
cvt.u32.u16 %r697, %rs21;
mul.lo.s32 %r698, %r94, %r696;
mad.lo.s32 %r699, %r118, %r697, %r698;
add.s32 %r700, %r118, %r94;
shr.u32 %r701, %r700, 31;
add.s32 %r702, %r700, %r701;
shr.s32 %r703, %r702, 1;
add.s32 %r704, %r699, %r703;
shr.s32 %r705, %r704, 31;
shr.u32 %r706, %r705, 30;
add.s32 %r707, %r704, %r706;
shr.s32 %r708, %r707, 2;
add.s64 %rd44, %rd99, %rd6;
ld.global.nc.u16 %rs1, [%rd44];

{ cvt.f32.f16 %f61, %rs1;}


ld.global.nc.u32 %r122, [%rd37];

{.reg .f16 low,high;
mov.b32 {low,high},%r122;
cvt.f32.f16 %f62, low;}


mul.ftz.f32 %f81, %f61, %f62;
cvt.rn.f32.s32 %f82, %r708;
fma.rn.ftz.f32 %f230, %f81, %f82, %f230;
add.s64 %rd45, %rd99, %rd10;
ld.global.nc.u16 %rs30, [%rd45+-32];
shr.u16 %rs31, %rs30, 8;
cvt.u32.u16 %r709, %rs31;
ld.global.nc.u16 %rs32, [%rd45+-30];
shr.u16 %rs33, %rs32, 8;
cvt.u32.u16 %r710, %rs33;
add.s64 %rd46, %rd99, %rd8;
ld.global.nc.u8 %rs34, [%rd46];
ld.global.nc.u16 %rs35, [%rd45];
ld.global.nc.u16 %rs36, [%rd45+2];
ld.global.nc.u8 %rs37, [%rd46+8];
and.b16 %rs38, %rs37, 240;
and.b16 %rs39, %rs37, 15;
shr.u16 %rs40, %rs38, 4;
and.b16 %rs41, %rs30, 255;
cvt.u32.u16 %r711, %rs41;
shl.b16 %rs42, %rs34, 8;
cvt.u32.u16 %r712, %rs42;
and.b32 %r713, %r712, 768;
or.b32 %r714, %r713, %r711;
mul.wide.u32 %rd47, %r714, 8;
add.s64 %rd48, %rd34, %rd47;
and.b16 %rs43, %rs35, 255;
mul.wide.u16 %r715, %rs43, 128;
cvt.u32.u16 %r716, %rs35;
shl.b32 %r717, %r716, 21;
or.b32 %r718, %r715, %r717;
and.b32 %r719, %r718, 25166208;
or.b32 %r720, %r719, -2139062144;
add.s32 %r721, %r720, -16843009;
or.b32 %r515, %r721, %r719;
mov.u32 %r123, %r515;

prmt.b32 %r123, %r123, 0, 0xba98;

mul.wide.u16 %r722, %rs43, 8;
shl.b32 %r723, %r716, 17;
or.b32 %r724, %r722, %r723;
and.b32 %r725, %r724, 25166208;
or.b32 %r726, %r725, -2139062144;
add.s32 %r727, %r726, -16843009;
or.b32 %r517, %r727, %r725;
mov.u32 %r125, %r517;

prmt.b32 %r125, %r125, 0, 0xba98;

ld.global.nc.v2.u32 {%r728, %r729}, [%rd48];
not.b32 %r731, %r728;
not.b32 %r733, %r729;
xor.b32 %r734, %r728, %r123;
or.b32 %r735, %r734, -2139062144;
and.b32 %r736, %r123, 2139062143;
and.b32 %r737, %r731, -2139062144;
sub.s32 %r738, %r735, %r736;
xor.b32 %r128, %r737, %r738;
xor.b32 %r739, %r729, %r125;
or.b32 %r740, %r739, -2139062144;
and.b32 %r741, %r125, 2139062143;
and.b32 %r742, %r733, -2139062144;
sub.s32 %r743, %r740, %r741;
xor.b32 %r132, %r742, %r743;

dp4a.s32.s32 %r127, %r128, %r80, %r546;


dp4a.s32.s32 %r131, %r132, %r84, %r127;

mul.wide.u16 %r744, %rs34, 64;
and.b32 %r745, %r744, 768;
or.b32 %r746, %r745, %r709;
mul.wide.u32 %rd49, %r746, 8;
add.s64 %rd50, %rd34, %rd49;
shr.u16 %rs45, %rs35, 8;
mul.wide.u16 %r747, %rs45, 128;
cvt.u32.u16 %r748, %rs45;
shl.b32 %r749, %r748, 21;
or.b32 %r750, %r747, %r749;
and.b32 %r751, %r750, 25166208;
or.b32 %r752, %r751, -2139062144;
add.s32 %r753, %r752, -16843009;
or.b32 %r527, %r753, %r751;
mov.u32 %r135, %r527;

prmt.b32 %r135, %r135, 0, 0xba98;

mul.wide.u16 %r754, %rs45, 8;
shl.b32 %r755, %r748, 17;
or.b32 %r756, %r754, %r755;
and.b32 %r757, %r756, 25166208;
or.b32 %r758, %r757, -2139062144;
add.s32 %r759, %r758, -16843009;
or.b32 %r529, %r759, %r757;
mov.u32 %r137, %r529;

prmt.b32 %r137, %r137, 0, 0xba98;

ld.global.nc.v2.u32 {%r760, %r761}, [%rd50];
not.b32 %r763, %r760;
not.b32 %r765, %r761;
xor.b32 %r766, %r760, %r135;
or.b32 %r767, %r766, -2139062144;
and.b32 %r768, %r135, 2139062143;
and.b32 %r769, %r763, -2139062144;
sub.s32 %r770, %r767, %r768;
xor.b32 %r140, %r769, %r770;
xor.b32 %r771, %r761, %r137;
or.b32 %r772, %r771, -2139062144;
and.b32 %r773, %r137, 2139062143;
and.b32 %r774, %r765, -2139062144;
sub.s32 %r775, %r772, %r773;
xor.b32 %r144, %r774, %r775;

dp4a.s32.s32 %r139, %r140, %r92, %r131;


dp4a.s32.s32 %r143, %r144, %r96, %r139;

and.b16 %rs46, %rs32, 255;
cvt.u32.u16 %r776, %rs46;
mul.wide.u16 %r777, %rs34, 16;
and.b32 %r778, %r777, 768;
or.b32 %r779, %r778, %r776;
mul.wide.u32 %rd51, %r779, 8;
add.s64 %rd52, %rd34, %rd51;
and.b16 %rs47, %rs36, 255;
mul.wide.u16 %r780, %rs47, 128;
cvt.u32.u16 %r781, %rs36;
shl.b32 %r782, %r781, 21;
or.b32 %r783, %r780, %r782;
and.b32 %r784, %r783, 25166208;
or.b32 %r785, %r784, -2139062144;
add.s32 %r786, %r785, -16843009;
or.b32 %r539, %r786, %r784;
mov.u32 %r147, %r539;

prmt.b32 %r147, %r147, 0, 0xba98;

mul.wide.u16 %r787, %rs47, 8;
shl.b32 %r788, %r781, 17;
or.b32 %r789, %r787, %r788;
and.b32 %r790, %r789, 25166208;
or.b32 %r791, %r790, -2139062144;
add.s32 %r792, %r791, -16843009;
or.b32 %r541, %r792, %r790;
mov.u32 %r149, %r541;

prmt.b32 %r149, %r149, 0, 0xba98;

ld.global.nc.v2.u32 {%r793, %r794}, [%rd52];
not.b32 %r796, %r793;
not.b32 %r798, %r794;
xor.b32 %r799, %r793, %r147;
or.b32 %r800, %r799, -2139062144;
and.b32 %r801, %r147, 2139062143;
and.b32 %r802, %r796, -2139062144;
sub.s32 %r803, %r800, %r801;
xor.b32 %r152, %r802, %r803;
xor.b32 %r804, %r794, %r149;
or.b32 %r805, %r804, -2139062144;
and.b32 %r806, %r149, 2139062143;
and.b32 %r807, %r798, -2139062144;
sub.s32 %r808, %r805, %r806;
xor.b32 %r156, %r807, %r808;

dp4a.s32.s32 %r151, %r152, %r104, %r546;


dp4a.s32.s32 %r155, %r156, %r108, %r151;

mul.wide.u16 %r809, %rs34, 4;
and.b32 %r810, %r809, 768;
or.b32 %r811, %r810, %r710;
mul.wide.u32 %rd53, %r811, 8;
add.s64 %rd54, %rd34, %rd53;
shr.u16 %rs48, %rs36, 8;
mul.wide.u16 %r812, %rs48, 128;
cvt.u32.u16 %r813, %rs48;
shl.b32 %r814, %r813, 21;
or.b32 %r815, %r812, %r814;
and.b32 %r816, %r815, 25166208;
or.b32 %r817, %r816, -2139062144;
add.s32 %r818, %r817, -16843009;
or.b32 %r551, %r818, %r816;
mov.u32 %r159, %r551;

prmt.b32 %r159, %r159, 0, 0xba98;

mul.wide.u16 %r819, %rs48, 8;
shl.b32 %r820, %r813, 17;
or.b32 %r821, %r819, %r820;
and.b32 %r822, %r821, 25166208;
or.b32 %r823, %r822, -2139062144;
add.s32 %r824, %r823, -16843009;
or.b32 %r553, %r824, %r822;
mov.u32 %r161, %r553;

prmt.b32 %r161, %r161, 0, 0xba98;

ld.global.nc.v2.u32 {%r825, %r826}, [%rd54];
not.b32 %r828, %r825;
not.b32 %r830, %r826;
xor.b32 %r831, %r825, %r159;
or.b32 %r832, %r831, -2139062144;
and.b32 %r833, %r159, 2139062143;
and.b32 %r834, %r828, -2139062144;
sub.s32 %r835, %r832, %r833;
xor.b32 %r164, %r834, %r835;
xor.b32 %r836, %r826, %r161;
or.b32 %r837, %r836, -2139062144;
and.b32 %r838, %r161, 2139062143;
and.b32 %r839, %r830, -2139062144;
sub.s32 %r840, %r837, %r838;
xor.b32 %r168, %r839, %r840;

dp4a.s32.s32 %r163, %r164, %r116, %r155;


dp4a.s32.s32 %r167, %r168, %r120, %r163;

cvt.u32.u16 %r841, %rs39;
cvt.u32.u16 %r842, %rs40;
mul.lo.s32 %r843, %r143, %r841;
mad.lo.s32 %r844, %r167, %r842, %r843;
add.s32 %r845, %r167, %r143;
shr.u32 %r846, %r845, 31;
add.s32 %r847, %r845, %r846;
shr.s32 %r848, %r847, 1;
add.s32 %r849, %r844, %r848;
shr.s32 %r850, %r849, 31;
shr.u32 %r851, %r850, 30;
add.s32 %r852, %r849, %r851;
shr.s32 %r853, %r852, 2;
add.s64 %rd55, %rd99, %rd5;
ld.global.nc.u16 %rs2, [%rd55];

{ cvt.f32.f16 %f63, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r122;
cvt.f32.f16 %f64, low;}


mul.ftz.f32 %f83, %f63, %f64;
cvt.rn.f32.s32 %f84, %r853;
fma.rn.ftz.f32 %f229, %f83, %f84, %f229;
cvt.u64.u32 %rd56, %r1386;
add.s64 %rd57, %rd2, %rd56;
add.s64 %rd58, %rd57, %rd4;
mov.u32 %r172, %r466;

prmt.b32 %r172, %r172, 0, 0xba98;

mov.u32 %r174, %r468;

prmt.b32 %r174, %r174, 0, 0xba98;

xor.b32 %r854, %r583, %r172;
or.b32 %r855, %r854, -2139062144;
and.b32 %r856, %r172, 2139062143;
sub.s32 %r857, %r855, %r856;
xor.b32 %r177, %r592, %r857;
xor.b32 %r858, %r584, %r174;
or.b32 %r859, %r858, -2139062144;
and.b32 %r860, %r174, 2139062143;
sub.s32 %r861, %r859, %r860;
xor.b32 %r181, %r597, %r861;
mul.lo.s64 %rd59, %rd58, 36;
add.s64 %rd60, %rd11, %rd59;
ld.global.nc.u32 %r182, [%rd60+8];
ld.global.nc.u32 %r178, [%rd60+4];

dp4a.s32.s32 %r176, %r177, %r178, %r546;


dp4a.s32.s32 %r180, %r181, %r182, %r176;

mov.u32 %r184, %r478;

prmt.b32 %r184, %r184, 0, 0xba98;

mov.u32 %r186, %r480;

prmt.b32 %r186, %r186, 0, 0xba98;

xor.b32 %r862, %r615, %r184;
or.b32 %r863, %r862, -2139062144;
and.b32 %r864, %r184, 2139062143;
sub.s32 %r865, %r863, %r864;
xor.b32 %r189, %r624, %r865;
xor.b32 %r866, %r616, %r186;
or.b32 %r867, %r866, -2139062144;
and.b32 %r868, %r186, 2139062143;
sub.s32 %r869, %r867, %r868;
xor.b32 %r193, %r629, %r869;
ld.global.nc.u32 %r194, [%rd60+16];
ld.global.nc.u32 %r190, [%rd60+12];

dp4a.s32.s32 %r188, %r189, %r190, %r180;


dp4a.s32.s32 %r192, %r193, %r194, %r188;

mov.u32 %r196, %r490;

prmt.b32 %r196, %r196, 0, 0xba98;

mov.u32 %r198, %r492;

prmt.b32 %r198, %r198, 0, 0xba98;

xor.b32 %r870, %r648, %r196;
or.b32 %r871, %r870, -2139062144;
and.b32 %r872, %r196, 2139062143;
sub.s32 %r873, %r871, %r872;
xor.b32 %r201, %r657, %r873;
xor.b32 %r874, %r649, %r198;
or.b32 %r875, %r874, -2139062144;
and.b32 %r876, %r198, 2139062143;
sub.s32 %r877, %r875, %r876;
xor.b32 %r205, %r662, %r877;
ld.global.nc.u32 %r206, [%rd60+24];
ld.global.nc.u32 %r202, [%rd60+20];

dp4a.s32.s32 %r200, %r201, %r202, %r546;


dp4a.s32.s32 %r204, %r205, %r206, %r200;

mov.u32 %r208, %r502;

prmt.b32 %r208, %r208, 0, 0xba98;

mov.u32 %r210, %r504;

prmt.b32 %r210, %r210, 0, 0xba98;

xor.b32 %r878, %r680, %r208;
or.b32 %r879, %r878, -2139062144;
and.b32 %r880, %r208, 2139062143;
sub.s32 %r881, %r879, %r880;
xor.b32 %r213, %r689, %r881;
xor.b32 %r882, %r681, %r210;
or.b32 %r883, %r882, -2139062144;
and.b32 %r884, %r210, 2139062143;
sub.s32 %r885, %r883, %r884;
xor.b32 %r217, %r694, %r885;
ld.global.nc.u32 %r218, [%rd60+32];
ld.global.nc.u32 %r214, [%rd60+28];

dp4a.s32.s32 %r212, %r213, %r214, %r204;


dp4a.s32.s32 %r216, %r217, %r218, %r212;

mul.lo.s32 %r886, %r192, %r696;
mad.lo.s32 %r887, %r216, %r697, %r886;
add.s32 %r888, %r216, %r192;
shr.u32 %r889, %r888, 31;
add.s32 %r890, %r888, %r889;
shr.s32 %r891, %r890, 1;
add.s32 %r892, %r887, %r891;
shr.s32 %r893, %r892, 31;
shr.u32 %r894, %r893, 30;
add.s32 %r895, %r892, %r894;
shr.s32 %r896, %r895, 2;

{ cvt.f32.f16 %f65, %rs1;}


ld.global.nc.u32 %r220, [%rd60];

{.reg .f16 low,high;
mov.b32 {low,high},%r220;
cvt.f32.f16 %f66, low;}


mul.ftz.f32 %f85, %f65, %f66;
cvt.rn.f32.s32 %f86, %r896;
fma.rn.ftz.f32 %f228, %f85, %f86, %f228;
mov.u32 %r221, %r515;

prmt.b32 %r221, %r221, 0, 0xba98;

mov.u32 %r223, %r517;

prmt.b32 %r223, %r223, 0, 0xba98;

xor.b32 %r897, %r728, %r221;
or.b32 %r898, %r897, -2139062144;
and.b32 %r899, %r221, 2139062143;
sub.s32 %r900, %r898, %r899;
xor.b32 %r226, %r737, %r900;
xor.b32 %r901, %r729, %r223;
or.b32 %r902, %r901, -2139062144;
and.b32 %r903, %r223, 2139062143;
sub.s32 %r904, %r902, %r903;
xor.b32 %r230, %r742, %r904;

dp4a.s32.s32 %r225, %r226, %r178, %r546;


dp4a.s32.s32 %r229, %r230, %r182, %r225;

mov.u32 %r233, %r527;

prmt.b32 %r233, %r233, 0, 0xba98;

mov.u32 %r235, %r529;

prmt.b32 %r235, %r235, 0, 0xba98;

xor.b32 %r905, %r760, %r233;
or.b32 %r906, %r905, -2139062144;
and.b32 %r907, %r233, 2139062143;
sub.s32 %r908, %r906, %r907;
xor.b32 %r238, %r769, %r908;
xor.b32 %r909, %r761, %r235;
or.b32 %r910, %r909, -2139062144;
and.b32 %r911, %r235, 2139062143;
sub.s32 %r912, %r910, %r911;
xor.b32 %r242, %r774, %r912;

dp4a.s32.s32 %r237, %r238, %r190, %r229;


dp4a.s32.s32 %r241, %r242, %r194, %r237;

mov.u32 %r245, %r539;

prmt.b32 %r245, %r245, 0, 0xba98;

mov.u32 %r247, %r541;

prmt.b32 %r247, %r247, 0, 0xba98;

xor.b32 %r913, %r793, %r245;
or.b32 %r914, %r913, -2139062144;
and.b32 %r915, %r245, 2139062143;
sub.s32 %r916, %r914, %r915;
xor.b32 %r250, %r802, %r916;
xor.b32 %r917, %r794, %r247;
or.b32 %r918, %r917, -2139062144;
and.b32 %r919, %r247, 2139062143;
sub.s32 %r920, %r918, %r919;
xor.b32 %r254, %r807, %r920;

dp4a.s32.s32 %r249, %r250, %r202, %r546;


dp4a.s32.s32 %r253, %r254, %r206, %r249;

mov.u32 %r257, %r551;

prmt.b32 %r257, %r257, 0, 0xba98;

mov.u32 %r259, %r553;

prmt.b32 %r259, %r259, 0, 0xba98;

xor.b32 %r921, %r825, %r257;
or.b32 %r922, %r921, -2139062144;
and.b32 %r923, %r257, 2139062143;
sub.s32 %r924, %r922, %r923;
xor.b32 %r262, %r834, %r924;
xor.b32 %r925, %r826, %r259;
or.b32 %r926, %r925, -2139062144;
and.b32 %r927, %r259, 2139062143;
sub.s32 %r928, %r926, %r927;
xor.b32 %r266, %r839, %r928;

dp4a.s32.s32 %r261, %r262, %r214, %r253;


dp4a.s32.s32 %r265, %r266, %r218, %r261;

mul.lo.s32 %r929, %r241, %r841;
mad.lo.s32 %r930, %r265, %r842, %r929;
add.s32 %r931, %r265, %r241;
shr.u32 %r932, %r931, 31;
add.s32 %r933, %r931, %r932;
shr.s32 %r934, %r933, 1;
add.s32 %r935, %r930, %r934;
shr.s32 %r936, %r935, 31;
shr.u32 %r937, %r936, 30;
add.s32 %r938, %r935, %r937;
shr.s32 %r939, %r938, 2;

{ cvt.f32.f16 %f67, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r220;
cvt.f32.f16 %f68, low;}


mul.ftz.f32 %f87, %f67, %f68;
cvt.rn.f32.s32 %f88, %r939;
fma.rn.ftz.f32 %f227, %f87, %f88, %f227;
cvt.u64.u32 %rd61, %r1389;
add.s64 %rd62, %rd2, %rd61;
add.s64 %rd63, %rd62, %rd4;
mov.u32 %r270, %r466;

prmt.b32 %r270, %r270, 0, 0xba98;

mov.u32 %r272, %r468;

prmt.b32 %r272, %r272, 0, 0xba98;

xor.b32 %r940, %r583, %r270;
or.b32 %r941, %r940, -2139062144;
and.b32 %r942, %r270, 2139062143;
sub.s32 %r943, %r941, %r942;
xor.b32 %r275, %r592, %r943;
xor.b32 %r944, %r584, %r272;
or.b32 %r945, %r944, -2139062144;
and.b32 %r946, %r272, 2139062143;
sub.s32 %r947, %r945, %r946;
xor.b32 %r279, %r597, %r947;
mul.lo.s64 %rd64, %rd63, 36;
add.s64 %rd65, %rd11, %rd64;
ld.global.nc.u32 %r280, [%rd65+8];
ld.global.nc.u32 %r276, [%rd65+4];

dp4a.s32.s32 %r274, %r275, %r276, %r546;


dp4a.s32.s32 %r278, %r279, %r280, %r274;

mov.u32 %r282, %r478;

prmt.b32 %r282, %r282, 0, 0xba98;

mov.u32 %r284, %r480;

prmt.b32 %r284, %r284, 0, 0xba98;

xor.b32 %r948, %r615, %r282;
or.b32 %r949, %r948, -2139062144;
and.b32 %r950, %r282, 2139062143;
sub.s32 %r951, %r949, %r950;
xor.b32 %r287, %r624, %r951;
xor.b32 %r952, %r616, %r284;
or.b32 %r953, %r952, -2139062144;
and.b32 %r954, %r284, 2139062143;
sub.s32 %r955, %r953, %r954;
xor.b32 %r291, %r629, %r955;
ld.global.nc.u32 %r292, [%rd65+16];
ld.global.nc.u32 %r288, [%rd65+12];

dp4a.s32.s32 %r286, %r287, %r288, %r278;


dp4a.s32.s32 %r290, %r291, %r292, %r286;

mov.u32 %r294, %r490;

prmt.b32 %r294, %r294, 0, 0xba98;

mov.u32 %r296, %r492;

prmt.b32 %r296, %r296, 0, 0xba98;

xor.b32 %r956, %r648, %r294;
or.b32 %r957, %r956, -2139062144;
and.b32 %r958, %r294, 2139062143;
sub.s32 %r959, %r957, %r958;
xor.b32 %r299, %r657, %r959;
xor.b32 %r960, %r649, %r296;
or.b32 %r961, %r960, -2139062144;
and.b32 %r962, %r296, 2139062143;
sub.s32 %r963, %r961, %r962;
xor.b32 %r303, %r662, %r963;
ld.global.nc.u32 %r304, [%rd65+24];
ld.global.nc.u32 %r300, [%rd65+20];

dp4a.s32.s32 %r298, %r299, %r300, %r546;


dp4a.s32.s32 %r302, %r303, %r304, %r298;

mov.u32 %r306, %r502;

prmt.b32 %r306, %r306, 0, 0xba98;

mov.u32 %r308, %r504;

prmt.b32 %r308, %r308, 0, 0xba98;

xor.b32 %r964, %r680, %r306;
or.b32 %r965, %r964, -2139062144;
and.b32 %r966, %r306, 2139062143;
sub.s32 %r967, %r965, %r966;
xor.b32 %r311, %r689, %r967;
xor.b32 %r968, %r681, %r308;
or.b32 %r969, %r968, -2139062144;
and.b32 %r970, %r308, 2139062143;
sub.s32 %r971, %r969, %r970;
xor.b32 %r315, %r694, %r971;
ld.global.nc.u32 %r316, [%rd65+32];
ld.global.nc.u32 %r312, [%rd65+28];

dp4a.s32.s32 %r310, %r311, %r312, %r302;


dp4a.s32.s32 %r314, %r315, %r316, %r310;

mul.lo.s32 %r972, %r290, %r696;
mad.lo.s32 %r973, %r314, %r697, %r972;
add.s32 %r974, %r314, %r290;
shr.u32 %r975, %r974, 31;
add.s32 %r976, %r974, %r975;
shr.s32 %r977, %r976, 1;
add.s32 %r978, %r973, %r977;
shr.s32 %r979, %r978, 31;
shr.u32 %r980, %r979, 30;
add.s32 %r981, %r978, %r980;
shr.s32 %r982, %r981, 2;

{ cvt.f32.f16 %f69, %rs1;}


ld.global.nc.u32 %r318, [%rd65];

{.reg .f16 low,high;
mov.b32 {low,high},%r318;
cvt.f32.f16 %f70, low;}


mul.ftz.f32 %f89, %f69, %f70;
cvt.rn.f32.s32 %f90, %r982;
fma.rn.ftz.f32 %f226, %f89, %f90, %f226;
mov.u32 %r319, %r515;

prmt.b32 %r319, %r319, 0, 0xba98;

mov.u32 %r321, %r517;

prmt.b32 %r321, %r321, 0, 0xba98;

xor.b32 %r983, %r728, %r319;
or.b32 %r984, %r983, -2139062144;
and.b32 %r985, %r319, 2139062143;
sub.s32 %r986, %r984, %r985;
xor.b32 %r324, %r737, %r986;
xor.b32 %r987, %r729, %r321;
or.b32 %r988, %r987, -2139062144;
and.b32 %r989, %r321, 2139062143;
sub.s32 %r990, %r988, %r989;
xor.b32 %r328, %r742, %r990;

dp4a.s32.s32 %r323, %r324, %r276, %r546;


dp4a.s32.s32 %r327, %r328, %r280, %r323;

mov.u32 %r331, %r527;

prmt.b32 %r331, %r331, 0, 0xba98;

mov.u32 %r333, %r529;

prmt.b32 %r333, %r333, 0, 0xba98;

xor.b32 %r991, %r760, %r331;
or.b32 %r992, %r991, -2139062144;
and.b32 %r993, %r331, 2139062143;
sub.s32 %r994, %r992, %r993;
xor.b32 %r336, %r769, %r994;
xor.b32 %r995, %r761, %r333;
or.b32 %r996, %r995, -2139062144;
and.b32 %r997, %r333, 2139062143;
sub.s32 %r998, %r996, %r997;
xor.b32 %r340, %r774, %r998;

dp4a.s32.s32 %r335, %r336, %r288, %r327;


dp4a.s32.s32 %r339, %r340, %r292, %r335;

mov.u32 %r343, %r539;

prmt.b32 %r343, %r343, 0, 0xba98;

mov.u32 %r345, %r541;

prmt.b32 %r345, %r345, 0, 0xba98;

xor.b32 %r999, %r793, %r343;
or.b32 %r1000, %r999, -2139062144;
and.b32 %r1001, %r343, 2139062143;
sub.s32 %r1002, %r1000, %r1001;
xor.b32 %r348, %r802, %r1002;
xor.b32 %r1003, %r794, %r345;
or.b32 %r1004, %r1003, -2139062144;
and.b32 %r1005, %r345, 2139062143;
sub.s32 %r1006, %r1004, %r1005;
xor.b32 %r352, %r807, %r1006;

dp4a.s32.s32 %r347, %r348, %r300, %r546;


dp4a.s32.s32 %r351, %r352, %r304, %r347;

mov.u32 %r355, %r551;

prmt.b32 %r355, %r355, 0, 0xba98;

mov.u32 %r357, %r553;

prmt.b32 %r357, %r357, 0, 0xba98;

xor.b32 %r1007, %r825, %r355;
or.b32 %r1008, %r1007, -2139062144;
and.b32 %r1009, %r355, 2139062143;
sub.s32 %r1010, %r1008, %r1009;
xor.b32 %r360, %r834, %r1010;
xor.b32 %r1011, %r826, %r357;
or.b32 %r1012, %r1011, -2139062144;
and.b32 %r1013, %r357, 2139062143;
sub.s32 %r1014, %r1012, %r1013;
xor.b32 %r364, %r839, %r1014;

dp4a.s32.s32 %r359, %r360, %r312, %r351;


dp4a.s32.s32 %r363, %r364, %r316, %r359;

mul.lo.s32 %r1015, %r339, %r841;
mad.lo.s32 %r1016, %r363, %r842, %r1015;
add.s32 %r1017, %r363, %r339;
shr.u32 %r1018, %r1017, 31;
add.s32 %r1019, %r1017, %r1018;
shr.s32 %r1020, %r1019, 1;
add.s32 %r1021, %r1016, %r1020;
shr.s32 %r1022, %r1021, 31;
shr.u32 %r1023, %r1022, 30;
add.s32 %r1024, %r1021, %r1023;
shr.s32 %r1025, %r1024, 2;

{ cvt.f32.f16 %f71, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r318;
cvt.f32.f16 %f72, low;}


mul.ftz.f32 %f91, %f71, %f72;
cvt.rn.f32.s32 %f92, %r1025;
fma.rn.ftz.f32 %f225, %f91, %f92, %f225;
cvt.u64.u32 %rd66, %r1388;
add.s64 %rd67, %rd2, %rd66;
add.s64 %rd68, %rd67, %rd4;
mov.u32 %r368, %r466;

prmt.b32 %r368, %r368, 0, 0xba98;

mov.u32 %r370, %r468;

prmt.b32 %r370, %r370, 0, 0xba98;

xor.b32 %r1026, %r583, %r368;
or.b32 %r1027, %r1026, -2139062144;
and.b32 %r1028, %r368, 2139062143;
sub.s32 %r1029, %r1027, %r1028;
xor.b32 %r373, %r592, %r1029;
xor.b32 %r1030, %r584, %r370;
or.b32 %r1031, %r1030, -2139062144;
and.b32 %r1032, %r370, 2139062143;
sub.s32 %r1033, %r1031, %r1032;
xor.b32 %r377, %r597, %r1033;
mul.lo.s64 %rd69, %rd68, 36;
add.s64 %rd70, %rd11, %rd69;
ld.global.nc.u32 %r378, [%rd70+8];
ld.global.nc.u32 %r374, [%rd70+4];

dp4a.s32.s32 %r372, %r373, %r374, %r546;


dp4a.s32.s32 %r376, %r377, %r378, %r372;

mov.u32 %r380, %r478;

prmt.b32 %r380, %r380, 0, 0xba98;

mov.u32 %r382, %r480;

prmt.b32 %r382, %r382, 0, 0xba98;

xor.b32 %r1034, %r615, %r380;
or.b32 %r1035, %r1034, -2139062144;
and.b32 %r1036, %r380, 2139062143;
sub.s32 %r1037, %r1035, %r1036;
xor.b32 %r385, %r624, %r1037;
xor.b32 %r1038, %r616, %r382;
or.b32 %r1039, %r1038, -2139062144;
and.b32 %r1040, %r382, 2139062143;
sub.s32 %r1041, %r1039, %r1040;
xor.b32 %r389, %r629, %r1041;
ld.global.nc.u32 %r390, [%rd70+16];
ld.global.nc.u32 %r386, [%rd70+12];

dp4a.s32.s32 %r384, %r385, %r386, %r376;


dp4a.s32.s32 %r388, %r389, %r390, %r384;

mov.u32 %r392, %r490;

prmt.b32 %r392, %r392, 0, 0xba98;

mov.u32 %r394, %r492;

prmt.b32 %r394, %r394, 0, 0xba98;

xor.b32 %r1042, %r648, %r392;
or.b32 %r1043, %r1042, -2139062144;
and.b32 %r1044, %r392, 2139062143;
sub.s32 %r1045, %r1043, %r1044;
xor.b32 %r397, %r657, %r1045;
xor.b32 %r1046, %r649, %r394;
or.b32 %r1047, %r1046, -2139062144;
and.b32 %r1048, %r394, 2139062143;
sub.s32 %r1049, %r1047, %r1048;
xor.b32 %r401, %r662, %r1049;
ld.global.nc.u32 %r402, [%rd70+24];
ld.global.nc.u32 %r398, [%rd70+20];

dp4a.s32.s32 %r396, %r397, %r398, %r546;


dp4a.s32.s32 %r400, %r401, %r402, %r396;

mov.u32 %r404, %r502;

prmt.b32 %r404, %r404, 0, 0xba98;

mov.u32 %r406, %r504;

prmt.b32 %r406, %r406, 0, 0xba98;

xor.b32 %r1050, %r680, %r404;
or.b32 %r1051, %r1050, -2139062144;
and.b32 %r1052, %r404, 2139062143;
sub.s32 %r1053, %r1051, %r1052;
xor.b32 %r409, %r689, %r1053;
xor.b32 %r1054, %r681, %r406;
or.b32 %r1055, %r1054, -2139062144;
and.b32 %r1056, %r406, 2139062143;
sub.s32 %r1057, %r1055, %r1056;
xor.b32 %r413, %r694, %r1057;
ld.global.nc.u32 %r414, [%rd70+32];
ld.global.nc.u32 %r410, [%rd70+28];

dp4a.s32.s32 %r408, %r409, %r410, %r400;


dp4a.s32.s32 %r412, %r413, %r414, %r408;

mul.lo.s32 %r1058, %r388, %r696;
mad.lo.s32 %r1059, %r412, %r697, %r1058;
add.s32 %r1060, %r412, %r388;
shr.u32 %r1061, %r1060, 31;
add.s32 %r1062, %r1060, %r1061;
shr.s32 %r1063, %r1062, 1;
add.s32 %r1064, %r1059, %r1063;
shr.s32 %r1065, %r1064, 31;
shr.u32 %r1066, %r1065, 30;
add.s32 %r1067, %r1064, %r1066;
shr.s32 %r1068, %r1067, 2;

{ cvt.f32.f16 %f73, %rs1;}


ld.global.nc.u32 %r416, [%rd70];

{.reg .f16 low,high;
mov.b32 {low,high},%r416;
cvt.f32.f16 %f74, low;}


mul.ftz.f32 %f93, %f73, %f74;
cvt.rn.f32.s32 %f94, %r1068;
fma.rn.ftz.f32 %f224, %f93, %f94, %f224;
mov.u32 %r417, %r515;

prmt.b32 %r417, %r417, 0, 0xba98;

mov.u32 %r419, %r517;

prmt.b32 %r419, %r419, 0, 0xba98;

xor.b32 %r1069, %r728, %r417;
or.b32 %r1070, %r1069, -2139062144;
and.b32 %r1071, %r417, 2139062143;
sub.s32 %r1072, %r1070, %r1071;
xor.b32 %r422, %r737, %r1072;
xor.b32 %r1073, %r729, %r419;
or.b32 %r1074, %r1073, -2139062144;
and.b32 %r1075, %r419, 2139062143;
sub.s32 %r1076, %r1074, %r1075;
xor.b32 %r426, %r742, %r1076;

dp4a.s32.s32 %r421, %r422, %r374, %r546;


dp4a.s32.s32 %r425, %r426, %r378, %r421;

mov.u32 %r429, %r527;

prmt.b32 %r429, %r429, 0, 0xba98;

mov.u32 %r431, %r529;

prmt.b32 %r431, %r431, 0, 0xba98;

xor.b32 %r1077, %r760, %r429;
or.b32 %r1078, %r1077, -2139062144;
and.b32 %r1079, %r429, 2139062143;
sub.s32 %r1080, %r1078, %r1079;
xor.b32 %r434, %r769, %r1080;
xor.b32 %r1081, %r761, %r431;
or.b32 %r1082, %r1081, -2139062144;
and.b32 %r1083, %r431, 2139062143;
sub.s32 %r1084, %r1082, %r1083;
xor.b32 %r438, %r774, %r1084;

dp4a.s32.s32 %r433, %r434, %r386, %r425;


dp4a.s32.s32 %r437, %r438, %r390, %r433;

mov.u32 %r441, %r539;

prmt.b32 %r441, %r441, 0, 0xba98;

mov.u32 %r443, %r541;

prmt.b32 %r443, %r443, 0, 0xba98;

xor.b32 %r1085, %r793, %r441;
or.b32 %r1086, %r1085, -2139062144;
and.b32 %r1087, %r441, 2139062143;
sub.s32 %r1088, %r1086, %r1087;
xor.b32 %r446, %r802, %r1088;
xor.b32 %r1089, %r794, %r443;
or.b32 %r1090, %r1089, -2139062144;
and.b32 %r1091, %r443, 2139062143;
sub.s32 %r1092, %r1090, %r1091;
xor.b32 %r450, %r807, %r1092;

dp4a.s32.s32 %r445, %r446, %r398, %r546;


dp4a.s32.s32 %r449, %r450, %r402, %r445;

mov.u32 %r453, %r551;

prmt.b32 %r453, %r453, 0, 0xba98;

mov.u32 %r455, %r553;

prmt.b32 %r455, %r455, 0, 0xba98;

xor.b32 %r1093, %r825, %r453;
or.b32 %r1094, %r1093, -2139062144;
and.b32 %r1095, %r453, 2139062143;
sub.s32 %r1096, %r1094, %r1095;
xor.b32 %r458, %r834, %r1096;
xor.b32 %r1097, %r826, %r455;
or.b32 %r1098, %r1097, -2139062144;
and.b32 %r1099, %r455, 2139062143;
sub.s32 %r1100, %r1098, %r1099;
xor.b32 %r462, %r839, %r1100;

dp4a.s32.s32 %r457, %r458, %r410, %r449;


dp4a.s32.s32 %r461, %r462, %r414, %r457;

mul.lo.s32 %r1101, %r437, %r841;
mad.lo.s32 %r1102, %r461, %r842, %r1101;
add.s32 %r1103, %r461, %r437;
shr.u32 %r1104, %r1103, 31;
add.s32 %r1105, %r1103, %r1104;
shr.s32 %r1106, %r1105, 1;
add.s32 %r1107, %r1102, %r1106;
shr.s32 %r1108, %r1107, 31;
shr.u32 %r1109, %r1108, 30;
add.s32 %r1110, %r1107, %r1109;
shr.s32 %r1111, %r1110, 2;

{ cvt.f32.f16 %f75, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r416;
cvt.f32.f16 %f76, low;}


mul.ftz.f32 %f95, %f75, %f76;
cvt.rn.f32.s32 %f96, %r1111;
fma.rn.ftz.f32 %f223, %f95, %f96, %f223;
cvt.u64.u32 %rd71, %r1387;
add.s64 %rd72, %rd2, %rd71;
add.s64 %rd73, %rd72, %rd4;

prmt.b32 %r466, %r466, 0, 0xba98;


prmt.b32 %r468, %r468, 0, 0xba98;

xor.b32 %r1112, %r583, %r466;
or.b32 %r1113, %r1112, -2139062144;
and.b32 %r1114, %r466, 2139062143;
sub.s32 %r1115, %r1113, %r1114;
xor.b32 %r471, %r592, %r1115;
xor.b32 %r1116, %r584, %r468;
or.b32 %r1117, %r1116, -2139062144;
and.b32 %r1118, %r468, 2139062143;
sub.s32 %r1119, %r1117, %r1118;
xor.b32 %r475, %r597, %r1119;
mul.lo.s64 %rd74, %rd73, 36;
add.s64 %rd75, %rd11, %rd74;
ld.global.nc.u32 %r476, [%rd75+8];
ld.global.nc.u32 %r472, [%rd75+4];

dp4a.s32.s32 %r470, %r471, %r472, %r546;


dp4a.s32.s32 %r474, %r475, %r476, %r470;


prmt.b32 %r478, %r478, 0, 0xba98;


prmt.b32 %r480, %r480, 0, 0xba98;

xor.b32 %r1120, %r615, %r478;
or.b32 %r1121, %r1120, -2139062144;
and.b32 %r1122, %r478, 2139062143;
sub.s32 %r1123, %r1121, %r1122;
xor.b32 %r483, %r624, %r1123;
xor.b32 %r1124, %r616, %r480;
or.b32 %r1125, %r1124, -2139062144;
and.b32 %r1126, %r480, 2139062143;
sub.s32 %r1127, %r1125, %r1126;
xor.b32 %r487, %r629, %r1127;
ld.global.nc.u32 %r488, [%rd75+16];
ld.global.nc.u32 %r484, [%rd75+12];

dp4a.s32.s32 %r482, %r483, %r484, %r474;


dp4a.s32.s32 %r486, %r487, %r488, %r482;


prmt.b32 %r490, %r490, 0, 0xba98;


prmt.b32 %r492, %r492, 0, 0xba98;

xor.b32 %r1128, %r648, %r490;
or.b32 %r1129, %r1128, -2139062144;
and.b32 %r1130, %r490, 2139062143;
sub.s32 %r1131, %r1129, %r1130;
xor.b32 %r495, %r657, %r1131;
xor.b32 %r1132, %r649, %r492;
or.b32 %r1133, %r1132, -2139062144;
and.b32 %r1134, %r492, 2139062143;
sub.s32 %r1135, %r1133, %r1134;
xor.b32 %r499, %r662, %r1135;
ld.global.nc.u32 %r500, [%rd75+24];
ld.global.nc.u32 %r496, [%rd75+20];

dp4a.s32.s32 %r494, %r495, %r496, %r546;


dp4a.s32.s32 %r498, %r499, %r500, %r494;


prmt.b32 %r502, %r502, 0, 0xba98;


prmt.b32 %r504, %r504, 0, 0xba98;

xor.b32 %r1136, %r680, %r502;
or.b32 %r1137, %r1136, -2139062144;
and.b32 %r1138, %r502, 2139062143;
sub.s32 %r1139, %r1137, %r1138;
xor.b32 %r507, %r689, %r1139;
xor.b32 %r1140, %r681, %r504;
or.b32 %r1141, %r1140, -2139062144;
and.b32 %r1142, %r504, 2139062143;
sub.s32 %r1143, %r1141, %r1142;
xor.b32 %r511, %r694, %r1143;
ld.global.nc.u32 %r512, [%rd75+32];
ld.global.nc.u32 %r508, [%rd75+28];

dp4a.s32.s32 %r506, %r507, %r508, %r498;


dp4a.s32.s32 %r510, %r511, %r512, %r506;

mul.lo.s32 %r1144, %r486, %r696;
mad.lo.s32 %r1145, %r510, %r697, %r1144;
add.s32 %r1146, %r510, %r486;
shr.u32 %r1147, %r1146, 31;
add.s32 %r1148, %r1146, %r1147;
shr.s32 %r1149, %r1148, 1;
add.s32 %r1150, %r1145, %r1149;
shr.s32 %r1151, %r1150, 31;
shr.u32 %r1152, %r1151, 30;
add.s32 %r1153, %r1150, %r1152;
shr.s32 %r1154, %r1153, 2;

{ cvt.f32.f16 %f77, %rs1;}


ld.global.nc.u32 %r514, [%rd75];

{.reg .f16 low,high;
mov.b32 {low,high},%r514;
cvt.f32.f16 %f78, low;}


mul.ftz.f32 %f97, %f77, %f78;
cvt.rn.f32.s32 %f98, %r1154;
fma.rn.ftz.f32 %f222, %f97, %f98, %f222;

prmt.b32 %r515, %r515, 0, 0xba98;


prmt.b32 %r517, %r517, 0, 0xba98;

xor.b32 %r1155, %r728, %r515;
or.b32 %r1156, %r1155, -2139062144;
and.b32 %r1157, %r515, 2139062143;
sub.s32 %r1158, %r1156, %r1157;
xor.b32 %r520, %r737, %r1158;
xor.b32 %r1159, %r729, %r517;
or.b32 %r1160, %r1159, -2139062144;
and.b32 %r1161, %r517, 2139062143;
sub.s32 %r1162, %r1160, %r1161;
xor.b32 %r524, %r742, %r1162;

dp4a.s32.s32 %r519, %r520, %r472, %r546;


dp4a.s32.s32 %r523, %r524, %r476, %r519;


prmt.b32 %r527, %r527, 0, 0xba98;


prmt.b32 %r529, %r529, 0, 0xba98;

xor.b32 %r1163, %r760, %r527;
or.b32 %r1164, %r1163, -2139062144;
and.b32 %r1165, %r527, 2139062143;
sub.s32 %r1166, %r1164, %r1165;
xor.b32 %r532, %r769, %r1166;
xor.b32 %r1167, %r761, %r529;
or.b32 %r1168, %r1167, -2139062144;
and.b32 %r1169, %r529, 2139062143;
sub.s32 %r1170, %r1168, %r1169;
xor.b32 %r536, %r774, %r1170;

dp4a.s32.s32 %r531, %r532, %r484, %r523;


dp4a.s32.s32 %r535, %r536, %r488, %r531;


prmt.b32 %r539, %r539, 0, 0xba98;


prmt.b32 %r541, %r541, 0, 0xba98;

xor.b32 %r1171, %r793, %r539;
or.b32 %r1172, %r1171, -2139062144;
and.b32 %r1173, %r539, 2139062143;
sub.s32 %r1174, %r1172, %r1173;
xor.b32 %r544, %r802, %r1174;
xor.b32 %r1175, %r794, %r541;
or.b32 %r1176, %r1175, -2139062144;
and.b32 %r1177, %r541, 2139062143;
sub.s32 %r1178, %r1176, %r1177;
xor.b32 %r548, %r807, %r1178;

dp4a.s32.s32 %r543, %r544, %r496, %r546;


dp4a.s32.s32 %r547, %r548, %r500, %r543;


prmt.b32 %r551, %r551, 0, 0xba98;


prmt.b32 %r553, %r553, 0, 0xba98;

xor.b32 %r1179, %r825, %r551;
or.b32 %r1180, %r1179, -2139062144;
and.b32 %r1181, %r551, 2139062143;
sub.s32 %r1182, %r1180, %r1181;
xor.b32 %r556, %r834, %r1182;
xor.b32 %r1183, %r826, %r553;
or.b32 %r1184, %r1183, -2139062144;
and.b32 %r1185, %r553, 2139062143;
sub.s32 %r1186, %r1184, %r1185;
xor.b32 %r560, %r839, %r1186;

dp4a.s32.s32 %r555, %r556, %r508, %r547;


dp4a.s32.s32 %r559, %r560, %r512, %r555;

mul.lo.s32 %r1187, %r535, %r841;
mad.lo.s32 %r1188, %r559, %r842, %r1187;
add.s32 %r1189, %r559, %r535;
shr.u32 %r1190, %r1189, 31;
add.s32 %r1191, %r1189, %r1190;
shr.s32 %r1192, %r1191, 1;
add.s32 %r1193, %r1188, %r1192;
shr.s32 %r1194, %r1193, 31;
shr.u32 %r1195, %r1194, 30;
add.s32 %r1196, %r1193, %r1195;
shr.s32 %r1197, %r1196, 2;

{ cvt.f32.f16 %f79, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r514;
cvt.f32.f16 %f80, low;}


mul.ftz.f32 %f99, %f79, %f80;
cvt.rn.f32.s32 %f100, %r1197;
fma.rn.ftz.f32 %f221, %f99, %f100, %f221;
add.s32 %r1389, %r1389, 64;
add.s32 %r1388, %r1388, 64;
add.s32 %r1387, %r1387, 64;
add.s32 %r1386, %r1386, 64;
add.s32 %r1385, %r1385, 64;
add.s64 %rd99, %rd99, 656;
add.s32 %r1390, %r1390, 8;
setp.lt.s32 %p4, %r1390, %r7;
@%p4 bra $L__BB233_2;

$L__BB233_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB233_5;

mov.u32 %r1198, _ZZ13mul_mat_vec_qIL9ggml_type22ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r1199, %r1, 1280, %r1198;
shl.b32 %r1200, %r2, 2;
add.s32 %r1201, %r1199, %r1200;
st.shared.f32 [%r1201+-1280], %f230;
st.shared.f32 [%r1201+-1152], %f229;
st.shared.f32 [%r1201+-1024], %f228;
st.shared.f32 [%r1201+-896], %f227;
st.shared.f32 [%r1201+-768], %f226;
st.shared.f32 [%r1201+-640], %f225;
st.shared.f32 [%r1201+-512], %f224;
st.shared.f32 [%r1201+-384], %f223;
st.shared.f32 [%r1201+-256], %f222;
st.shared.f32 [%r1201+-128], %f221;

$L__BB233_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB233_26;

mov.u32 %r1202, %ctaid.z;
mad.lo.s32 %r1203, %r5, %r35, %r4;
mad.lo.s32 %r1204, %r1202, %r41, %r1203;
cvt.u64.u32 %rd14, %r1204;
shl.b32 %r1205, %r2, 2;
mov.u32 %r1206, _ZZ13mul_mat_vec_qIL9ggml_type22ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r25, %r1206, %r1205;
mov.u32 %r1207, 2;
ld.shared.f32 %f101, [%r25];
add.ftz.f32 %f102, %f101, %f230;
mov.b32 %r1208, %f102;
mov.u32 %r1209, 31;
mov.u32 %r1210, 16;
mov.u32 %r1211, -1;
shfl.sync.bfly.b32 %r1212|%p7, %r1208, %r1210, %r1209, %r1211;
mov.b32 %f103, %r1212;
add.ftz.f32 %f104, %f102, %f103;
mov.b32 %r1213, %f104;
mov.u32 %r1214, 8;
shfl.sync.bfly.b32 %r1215|%p8, %r1213, %r1214, %r1209, %r1211;
mov.b32 %f105, %r1215;
add.ftz.f32 %f106, %f104, %f105;
mov.b32 %r1216, %f106;
mov.u32 %r1217, 4;
shfl.sync.bfly.b32 %r1218|%p9, %r1216, %r1217, %r1209, %r1211;
mov.b32 %f107, %r1218;
add.ftz.f32 %f108, %f106, %f107;
mov.b32 %r1219, %f108;
shfl.sync.bfly.b32 %r1220|%p10, %r1219, %r1207, %r1209, %r1211;
mov.b32 %f109, %r1220;
add.ftz.f32 %f110, %f108, %f109;
mov.b32 %r1221, %f110;
mov.u32 %r1222, 1;
shfl.sync.bfly.b32 %r1223|%p11, %r1221, %r1222, %r1209, %r1211;
mov.b32 %f111, %r1223;
add.ftz.f32 %f31, %f110, %f111;
setp.lt.u32 %p12, %r4, %r29;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
mul.wide.u32 %rd76, %r1204, 4;
add.s64 %rd15, %rd1, %rd76;
not.pred %p14, %p1;
@%p14 bra $L__BB233_8;

st.global.f32 [%rd15], %f31;

$L__BB233_8:
ld.shared.f32 %f112, [%r25+128];
add.ftz.f32 %f113, %f112, %f229;
mov.b32 %r1224, %f113;
shfl.sync.bfly.b32 %r1228|%p15, %r1224, %r1210, %r1209, %r1211;
mov.b32 %f114, %r1228;
add.ftz.f32 %f115, %f113, %f114;
mov.b32 %r1229, %f115;
shfl.sync.bfly.b32 %r1231|%p16, %r1229, %r1214, %r1209, %r1211;
mov.b32 %f116, %r1231;
add.ftz.f32 %f117, %f115, %f116;
mov.b32 %r1232, %f117;
shfl.sync.bfly.b32 %r1234|%p17, %r1232, %r1217, %r1209, %r1211;
mov.b32 %f118, %r1234;
add.ftz.f32 %f119, %f117, %f118;
mov.b32 %r1235, %f119;
shfl.sync.bfly.b32 %r1237|%p18, %r1235, %r1207, %r1209, %r1211;
mov.b32 %f120, %r1237;
add.ftz.f32 %f121, %f119, %f120;
mov.b32 %r1238, %f121;
shfl.sync.bfly.b32 %r1240|%p19, %r1238, %r1222, %r1209, %r1211;
mov.b32 %f122, %r1240;
add.ftz.f32 %f32, %f121, %f122;
setp.eq.s32 %p20, %r2, 1;
or.b32 %r1244, %r4, 1;
setp.lt.u32 %p21, %r1244, %r29;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB233_10;

st.global.f32 [%rd15+4], %f32;

$L__BB233_10:
ld.shared.f32 %f123, [%r25+256];
add.ftz.f32 %f124, %f123, %f228;
mov.b32 %r1245, %f124;
mov.u32 %r1246, 31;
mov.u32 %r1247, 16;
mov.u32 %r1248, -1;
shfl.sync.bfly.b32 %r1249|%p23, %r1245, %r1247, %r1246, %r1248;
mov.b32 %f125, %r1249;
add.ftz.f32 %f126, %f124, %f125;
mov.b32 %r1250, %f126;
mov.u32 %r1251, 8;
shfl.sync.bfly.b32 %r1252|%p24, %r1250, %r1251, %r1246, %r1248;
mov.b32 %f127, %r1252;
add.ftz.f32 %f128, %f126, %f127;
mov.b32 %r1253, %f128;
mov.u32 %r1254, 4;
shfl.sync.bfly.b32 %r1255|%p25, %r1253, %r1254, %r1246, %r1248;
mov.b32 %f129, %r1255;
add.ftz.f32 %f130, %f128, %f129;
mov.b32 %r1256, %f130;
mov.u32 %r1257, 2;
shfl.sync.bfly.b32 %r1258|%p26, %r1256, %r1257, %r1246, %r1248;
mov.b32 %f131, %r1258;
add.ftz.f32 %f132, %f130, %f131;
mov.b32 %r1259, %f132;
mov.u32 %r1260, 1;
shfl.sync.bfly.b32 %r1261|%p27, %r1259, %r1260, %r1246, %r1248;
mov.b32 %f133, %r1261;
add.ftz.f32 %f33, %f132, %f133;
@%p14 bra $L__BB233_12;

cvt.u64.u32 %rd77, %r29;
add.s64 %rd78, %rd77, %rd14;
shl.b64 %rd79, %rd78, 2;
add.s64 %rd80, %rd1, %rd79;
st.global.f32 [%rd80], %f33;

$L__BB233_12:
ld.shared.f32 %f134, [%r25+384];
add.ftz.f32 %f135, %f134, %f227;
mov.b32 %r1262, %f135;
shfl.sync.bfly.b32 %r1266|%p29, %r1262, %r1247, %r1246, %r1248;
mov.b32 %f136, %r1266;
add.ftz.f32 %f137, %f135, %f136;
mov.b32 %r1267, %f137;
shfl.sync.bfly.b32 %r1269|%p30, %r1267, %r1251, %r1246, %r1248;
mov.b32 %f138, %r1269;
add.ftz.f32 %f139, %f137, %f138;
mov.b32 %r1270, %f139;
shfl.sync.bfly.b32 %r1272|%p31, %r1270, %r1254, %r1246, %r1248;
mov.b32 %f140, %r1272;
add.ftz.f32 %f141, %f139, %f140;
mov.b32 %r1273, %f141;
shfl.sync.bfly.b32 %r1275|%p32, %r1273, %r1257, %r1246, %r1248;
mov.b32 %f142, %r1275;
add.ftz.f32 %f143, %f141, %f142;
mov.b32 %r1276, %f143;
shfl.sync.bfly.b32 %r1278|%p33, %r1276, %r1260, %r1246, %r1248;
mov.b32 %f144, %r1278;
add.ftz.f32 %f34, %f143, %f144;
@%p22 bra $L__BB233_14;

add.s32 %r1279, %r29, 1;
cvt.u64.u32 %rd81, %r1279;
add.s64 %rd82, %rd81, %rd14;
shl.b64 %rd83, %rd82, 2;
add.s64 %rd84, %rd1, %rd83;
st.global.f32 [%rd84], %f34;

$L__BB233_14:
shl.b32 %r1280, %r29, 1;
mov.u32 %r1281, 1;
ld.shared.f32 %f145, [%r25+512];
add.ftz.f32 %f146, %f145, %f226;
mov.b32 %r1282, %f146;
mov.u32 %r1283, 31;
mov.u32 %r1284, 16;
mov.u32 %r1285, -1;
shfl.sync.bfly.b32 %r1286|%p35, %r1282, %r1284, %r1283, %r1285;
mov.b32 %f147, %r1286;
add.ftz.f32 %f148, %f146, %f147;
mov.b32 %r1287, %f148;
mov.u32 %r1288, 8;
shfl.sync.bfly.b32 %r1289|%p36, %r1287, %r1288, %r1283, %r1285;
mov.b32 %f149, %r1289;
add.ftz.f32 %f150, %f148, %f149;
mov.b32 %r1290, %f150;
mov.u32 %r1291, 4;
shfl.sync.bfly.b32 %r1292|%p37, %r1290, %r1291, %r1283, %r1285;
mov.b32 %f151, %r1292;
add.ftz.f32 %f152, %f150, %f151;
mov.b32 %r1293, %f152;
mov.u32 %r1294, 2;
shfl.sync.bfly.b32 %r1295|%p38, %r1293, %r1294, %r1283, %r1285;
mov.b32 %f153, %r1295;
add.ftz.f32 %f154, %f152, %f153;
mov.b32 %r1296, %f154;
shfl.sync.bfly.b32 %r1297|%p39, %r1296, %r1281, %r1283, %r1285;
mov.b32 %f155, %r1297;
add.ftz.f32 %f35, %f154, %f155;
cvt.u64.u32 %rd85, %r1280;
add.s64 %rd86, %rd85, %rd14;
shl.b64 %rd87, %rd86, 2;
add.s64 %rd16, %rd1, %rd87;
@%p14 bra $L__BB233_16;

st.global.f32 [%rd16], %f35;

$L__BB233_16:
ld.shared.f32 %f156, [%r25+640];
add.ftz.f32 %f157, %f156, %f225;
mov.b32 %r1298, %f157;
shfl.sync.bfly.b32 %r1302|%p41, %r1298, %r1284, %r1283, %r1285;
mov.b32 %f158, %r1302;
add.ftz.f32 %f159, %f157, %f158;
mov.b32 %r1303, %f159;
shfl.sync.bfly.b32 %r1305|%p42, %r1303, %r1288, %r1283, %r1285;
mov.b32 %f160, %r1305;
add.ftz.f32 %f161, %f159, %f160;
mov.b32 %r1306, %f161;
shfl.sync.bfly.b32 %r1308|%p43, %r1306, %r1291, %r1283, %r1285;
mov.b32 %f162, %r1308;
add.ftz.f32 %f163, %f161, %f162;
mov.b32 %r1309, %f163;
shfl.sync.bfly.b32 %r1311|%p44, %r1309, %r1294, %r1283, %r1285;
mov.b32 %f164, %r1311;
add.ftz.f32 %f165, %f163, %f164;
mov.b32 %r1312, %f165;
shfl.sync.bfly.b32 %r1314|%p45, %r1312, %r1281, %r1283, %r1285;
mov.b32 %f166, %r1314;
add.ftz.f32 %f36, %f165, %f166;
@%p22 bra $L__BB233_18;

st.global.f32 [%rd16+4], %f36;

$L__BB233_18:
mul.lo.s32 %r26, %r29, 3;
ld.shared.f32 %f167, [%r25+768];
add.ftz.f32 %f168, %f167, %f224;
mov.b32 %r1315, %f168;
mov.u32 %r1316, 31;
mov.u32 %r1317, 16;
mov.u32 %r1318, -1;
shfl.sync.bfly.b32 %r1319|%p47, %r1315, %r1317, %r1316, %r1318;
mov.b32 %f169, %r1319;
add.ftz.f32 %f170, %f168, %f169;
mov.b32 %r1320, %f170;
mov.u32 %r1321, 8;
shfl.sync.bfly.b32 %r1322|%p48, %r1320, %r1321, %r1316, %r1318;
mov.b32 %f171, %r1322;
add.ftz.f32 %f172, %f170, %f171;
mov.b32 %r1323, %f172;
mov.u32 %r1324, 4;
shfl.sync.bfly.b32 %r1325|%p49, %r1323, %r1324, %r1316, %r1318;
mov.b32 %f173, %r1325;
add.ftz.f32 %f174, %f172, %f173;
mov.b32 %r1326, %f174;
mov.u32 %r1327, 2;
shfl.sync.bfly.b32 %r1328|%p50, %r1326, %r1327, %r1316, %r1318;
mov.b32 %f175, %r1328;
add.ftz.f32 %f176, %f174, %f175;
mov.b32 %r1329, %f176;
mov.u32 %r1330, 1;
shfl.sync.bfly.b32 %r1331|%p51, %r1329, %r1330, %r1316, %r1318;
mov.b32 %f177, %r1331;
add.ftz.f32 %f37, %f176, %f177;
@%p14 bra $L__BB233_20;

cvt.u64.u32 %rd88, %r26;
add.s64 %rd89, %rd88, %rd14;
shl.b64 %rd90, %rd89, 2;
add.s64 %rd91, %rd1, %rd90;
st.global.f32 [%rd91], %f37;

$L__BB233_20:
ld.shared.f32 %f178, [%r25+896];
add.ftz.f32 %f179, %f178, %f223;
mov.b32 %r1332, %f179;
shfl.sync.bfly.b32 %r1336|%p53, %r1332, %r1317, %r1316, %r1318;
mov.b32 %f180, %r1336;
add.ftz.f32 %f181, %f179, %f180;
mov.b32 %r1337, %f181;
shfl.sync.bfly.b32 %r1339|%p54, %r1337, %r1321, %r1316, %r1318;
mov.b32 %f182, %r1339;
add.ftz.f32 %f183, %f181, %f182;
mov.b32 %r1340, %f183;
shfl.sync.bfly.b32 %r1342|%p55, %r1340, %r1324, %r1316, %r1318;
mov.b32 %f184, %r1342;
add.ftz.f32 %f185, %f183, %f184;
mov.b32 %r1343, %f185;
shfl.sync.bfly.b32 %r1345|%p56, %r1343, %r1327, %r1316, %r1318;
mov.b32 %f186, %r1345;
add.ftz.f32 %f187, %f185, %f186;
mov.b32 %r1346, %f187;
shfl.sync.bfly.b32 %r1348|%p57, %r1346, %r1330, %r1316, %r1318;
mov.b32 %f188, %r1348;
add.ftz.f32 %f38, %f187, %f188;
@%p22 bra $L__BB233_22;

add.s32 %r1349, %r26, 1;
cvt.u64.u32 %rd92, %r1349;
add.s64 %rd93, %rd92, %rd14;
shl.b64 %rd94, %rd93, 2;
add.s64 %rd95, %rd1, %rd94;
st.global.f32 [%rd95], %f38;

$L__BB233_22:
shl.b32 %r1350, %r29, 2;
mov.u32 %r1351, 2;
ld.shared.f32 %f189, [%r25+1024];
add.ftz.f32 %f190, %f189, %f222;
mov.b32 %r1352, %f190;
mov.u32 %r1353, 31;
mov.u32 %r1354, 16;
mov.u32 %r1355, -1;
shfl.sync.bfly.b32 %r1356|%p59, %r1352, %r1354, %r1353, %r1355;
mov.b32 %f191, %r1356;
add.ftz.f32 %f192, %f190, %f191;
mov.b32 %r1357, %f192;
mov.u32 %r1358, 8;
shfl.sync.bfly.b32 %r1359|%p60, %r1357, %r1358, %r1353, %r1355;
mov.b32 %f193, %r1359;
add.ftz.f32 %f194, %f192, %f193;
mov.b32 %r1360, %f194;
mov.u32 %r1361, 4;
shfl.sync.bfly.b32 %r1362|%p61, %r1360, %r1361, %r1353, %r1355;
mov.b32 %f195, %r1362;
add.ftz.f32 %f196, %f194, %f195;
mov.b32 %r1363, %f196;
shfl.sync.bfly.b32 %r1364|%p62, %r1363, %r1351, %r1353, %r1355;
mov.b32 %f197, %r1364;
add.ftz.f32 %f198, %f196, %f197;
mov.b32 %r1365, %f198;
mov.u32 %r1366, 1;
shfl.sync.bfly.b32 %r1367|%p63, %r1365, %r1366, %r1353, %r1355;
mov.b32 %f199, %r1367;
add.ftz.f32 %f39, %f198, %f199;
cvt.u64.u32 %rd96, %r1350;
add.s64 %rd97, %rd96, %rd14;
shl.b64 %rd98, %rd97, 2;
add.s64 %rd17, %rd1, %rd98;
@%p14 bra $L__BB233_24;

st.global.f32 [%rd17], %f39;

$L__BB233_24:
ld.shared.f32 %f200, [%r25+1152];
add.ftz.f32 %f201, %f200, %f221;
mov.b32 %r1368, %f201;
shfl.sync.bfly.b32 %r1372|%p65, %r1368, %r1354, %r1353, %r1355;
mov.b32 %f202, %r1372;
add.ftz.f32 %f203, %f201, %f202;
mov.b32 %r1373, %f203;
shfl.sync.bfly.b32 %r1375|%p66, %r1373, %r1358, %r1353, %r1355;
mov.b32 %f204, %r1375;
add.ftz.f32 %f205, %f203, %f204;
mov.b32 %r1376, %f205;
shfl.sync.bfly.b32 %r1378|%p67, %r1376, %r1361, %r1353, %r1355;
mov.b32 %f206, %r1378;
add.ftz.f32 %f207, %f205, %f206;
mov.b32 %r1379, %f207;
shfl.sync.bfly.b32 %r1381|%p68, %r1379, %r1351, %r1353, %r1355;
mov.b32 %f208, %r1381;
add.ftz.f32 %f209, %f207, %f208;
mov.b32 %r1382, %f209;
shfl.sync.bfly.b32 %r1384|%p69, %r1382, %r1366, %r1353, %r1355;
mov.b32 %f210, %r1384;
add.ftz.f32 %f40, %f209, %f210;
@%p22 bra $L__BB233_26;

st.global.f32 [%rd17+4], %f40;

$L__BB233_26:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type22ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type22ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type22ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type22ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type22ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type22ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type22ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type22ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type22ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<83>;
.reg .b16 %rs<51>;
.reg .f32 %f<277>;
.reg .b32 %r<1630>;
.reg .b64 %rd<124>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type22ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type22ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type22ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type22ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type22ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type22ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type22ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type22ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type22ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type22ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type22ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type22ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type22ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type22ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type22ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type22ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type22ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type22ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r45, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r45, %r2;
shr.s32 %r46, %r3, 31;
shr.u32 %r47, %r46, 29;
add.s32 %r48, %r3, %r47;
shr.s32 %r1629, %r48, 3;
shr.u32 %r5, %r44, 8;
setp.ge.s32 %p3, %r1629, %r5;
mov.f32 %f265, 0f00000000;
mov.f32 %f266, %f265;
mov.f32 %f267, %f265;
mov.f32 %f268, %f265;
mov.f32 %f269, %f265;
mov.f32 %f270, %f265;
mov.f32 %f271, %f265;
mov.f32 %f272, %f265;
mov.f32 %f273, %f265;
mov.f32 %f274, %f265;
mov.f32 %f275, %f265;
mov.f32 %f276, %f265;
@%p3 bra $L__BB234_3;

mov.u32 %r49, %ctaid.x;
shl.b32 %r50, %r49, 1;
mov.u32 %r51, %ctaid.y;
mov.u32 %r52, %ctaid.z;
mul.lo.s32 %r53, %r51, %r36;
cvt.u64.u32 %rd19, %r53;
mul.lo.s32 %r54, %r52, %r42;
cvt.u64.u32 %rd20, %r54;
add.s64 %rd1, %rd20, %rd19;
mul.hi.u32 %r55, %r52, %r38;
add.s32 %r56, %r55, %r52;
shr.u32 %r57, %r56, %r39;
mul.lo.s32 %r58, %r57, %r41;
mul.hi.u32 %r59, %r51, %r32;
add.s32 %r60, %r59, %r51;
shr.u32 %r61, %r60, %r33;
mul.lo.s32 %r62, %r61, %r35;
cvta.to.global.u64 %rd123, %rd16;
and.b32 %r66, %r48, -8;
sub.s32 %r67, %r3, %r66;
shl.b32 %r68, %r67, 1;
cvt.s64.s32 %rd3, %r67;
shl.b32 %r69, %r30, 1;
shl.b32 %r1623, %r1629, 3;
add.s32 %r1628, %r1623, %r69;
mad.lo.s32 %r1627, %r30, 3, %r1623;
shl.b32 %r70, %r30, 2;
add.s32 %r1626, %r1623, %r70;
mad.lo.s32 %r1625, %r30, 5, %r1623;
add.s32 %r1624, %r30, %r1623;
add.s32 %r71, %r1629, %r58;
add.s32 %r72, %r71, %r62;
or.b32 %r73, %r50, 1;
mad.lo.s32 %r74, %r29, %r73, %r72;
mul.wide.s32 %rd4, %r74, 82;
mad.lo.s32 %r75, %r50, %r29, %r1629;
add.s32 %r76, %r75, %r58;
add.s32 %r77, %r76, %r62;
mul.wide.s32 %rd5, %r77, 82;
mul.wide.s32 %rd21, %r68, 2;
add.s64 %rd22, %rd5, %rd21;
add.s64 %rd6, %rd22, 36;
add.s64 %rd23, %rd4, %rd3;
add.s64 %rd7, %rd23, 66;
add.s64 %rd24, %rd5, %rd3;
add.s64 %rd8, %rd24, 74;
add.s64 %rd25, %rd4, %rd21;
add.s64 %rd9, %rd25, 34;
cvta.to.global.u64 %rd10, %rd17;
mov.f32 %f265, 0f00000000;

$L__BB234_2:
cvt.u64.u32 %rd26, %r1623;
add.s64 %rd27, %rd1, %rd26;
add.s64 %rd28, %rd27, %rd3;
add.s64 %rd29, %rd123, %rd6;
ld.global.nc.u16 %rs13, [%rd29+-34];
shr.u16 %rs14, %rs13, 8;
cvt.u32.u16 %r666, %rs14;
ld.global.nc.u16 %rs15, [%rd29+-32];
shr.u16 %rs16, %rs15, 8;
cvt.u32.u16 %r667, %rs16;
add.s64 %rd30, %rd123, %rd8;
ld.global.nc.u8 %rs17, [%rd30+-8];
ld.global.nc.u16 %rs18, [%rd29+-2];
ld.global.nc.u16 %rs19, [%rd29];
ld.global.nc.u8 %rs20, [%rd30];
and.b16 %rs21, %rs20, 240;
and.b16 %rs22, %rs20, 15;
shr.u16 %rs23, %rs21, 4;
and.b16 %rs24, %rs13, 255;
cvt.u32.u16 %r668, %rs24;
shl.b16 %rs25, %rs17, 8;
cvt.u32.u16 %r669, %rs25;
and.b32 %r670, %r669, 768;
or.b32 %r671, %r670, %r668;
mul.wide.u32 %rd31, %r671, 8;
mov.u64 %rd32, iq2s_grid;
add.s64 %rd33, %rd32, %rd31;
and.b16 %rs26, %rs18, 255;
mul.wide.u16 %r672, %rs26, 128;
cvt.u32.u16 %r673, %rs18;
shl.b32 %r674, %r673, 21;
or.b32 %r675, %r672, %r674;
and.b32 %r676, %r675, 25166208;
or.b32 %r677, %r676, -2139062144;
add.s32 %r678, %r677, -16843009;
or.b32 %r568, %r678, %r676;
mov.u32 %r78, %r568;

prmt.b32 %r78, %r78, 0, 0xba98;

mul.wide.u16 %r679, %rs26, 8;
shl.b32 %r680, %r673, 17;
or.b32 %r681, %r679, %r680;
and.b32 %r682, %r681, 25166208;
or.b32 %r683, %r682, -2139062144;
add.s32 %r684, %r683, -16843009;
or.b32 %r570, %r684, %r682;
mov.u32 %r80, %r570;

prmt.b32 %r80, %r80, 0, 0xba98;

ld.global.nc.v2.u32 {%r685, %r686}, [%rd33];
not.b32 %r688, %r685;
mov.u32 %r648, 0;
not.b32 %r690, %r686;
xor.b32 %r691, %r685, %r78;
or.b32 %r692, %r691, -2139062144;
and.b32 %r693, %r78, 2139062143;
and.b32 %r694, %r688, -2139062144;
sub.s32 %r695, %r692, %r693;
xor.b32 %r83, %r694, %r695;
xor.b32 %r696, %r686, %r80;
or.b32 %r697, %r696, -2139062144;
and.b32 %r698, %r80, 2139062143;
and.b32 %r699, %r690, -2139062144;
sub.s32 %r700, %r697, %r698;
xor.b32 %r87, %r699, %r700;
mul.lo.s64 %rd34, %rd28, 36;
add.s64 %rd35, %rd10, %rd34;
ld.global.nc.u32 %r88, [%rd35+8];
ld.global.nc.u32 %r84, [%rd35+4];

dp4a.s32.s32 %r82, %r83, %r84, %r648;


dp4a.s32.s32 %r86, %r87, %r88, %r82;

mul.wide.u16 %r701, %rs17, 64;
and.b32 %r702, %r701, 768;
or.b32 %r703, %r702, %r666;
mul.wide.u32 %rd36, %r703, 8;
add.s64 %rd37, %rd32, %rd36;
shr.u16 %rs28, %rs18, 8;
mul.wide.u16 %r704, %rs28, 128;
cvt.u32.u16 %r705, %rs28;
shl.b32 %r706, %r705, 21;
or.b32 %r707, %r704, %r706;
and.b32 %r708, %r707, 25166208;
or.b32 %r709, %r708, -2139062144;
add.s32 %r710, %r709, -16843009;
or.b32 %r580, %r710, %r708;
mov.u32 %r90, %r580;

prmt.b32 %r90, %r90, 0, 0xba98;

mul.wide.u16 %r711, %rs28, 8;
shl.b32 %r712, %r705, 17;
or.b32 %r713, %r711, %r712;
and.b32 %r714, %r713, 25166208;
or.b32 %r715, %r714, -2139062144;
add.s32 %r716, %r715, -16843009;
or.b32 %r582, %r716, %r714;
mov.u32 %r92, %r582;

prmt.b32 %r92, %r92, 0, 0xba98;

ld.global.nc.v2.u32 {%r717, %r718}, [%rd37];
not.b32 %r720, %r717;
not.b32 %r722, %r718;
xor.b32 %r723, %r717, %r90;
or.b32 %r724, %r723, -2139062144;
and.b32 %r725, %r90, 2139062143;
and.b32 %r726, %r720, -2139062144;
sub.s32 %r727, %r724, %r725;
xor.b32 %r95, %r726, %r727;
xor.b32 %r728, %r718, %r92;
or.b32 %r729, %r728, -2139062144;
and.b32 %r730, %r92, 2139062143;
and.b32 %r731, %r722, -2139062144;
sub.s32 %r732, %r729, %r730;
xor.b32 %r99, %r731, %r732;
ld.global.nc.u32 %r100, [%rd35+16];
ld.global.nc.u32 %r96, [%rd35+12];

dp4a.s32.s32 %r94, %r95, %r96, %r86;


dp4a.s32.s32 %r98, %r99, %r100, %r94;

and.b16 %rs29, %rs15, 255;
cvt.u32.u16 %r733, %rs29;
mul.wide.u16 %r734, %rs17, 16;
and.b32 %r735, %r734, 768;
or.b32 %r736, %r735, %r733;
mul.wide.u32 %rd38, %r736, 8;
add.s64 %rd39, %rd32, %rd38;
and.b16 %rs30, %rs19, 255;
mul.wide.u16 %r737, %rs30, 128;
cvt.u32.u16 %r738, %rs19;
shl.b32 %r739, %r738, 21;
or.b32 %r740, %r737, %r739;
and.b32 %r741, %r740, 25166208;
or.b32 %r742, %r741, -2139062144;
add.s32 %r743, %r742, -16843009;
or.b32 %r592, %r743, %r741;
mov.u32 %r102, %r592;

prmt.b32 %r102, %r102, 0, 0xba98;

mul.wide.u16 %r744, %rs30, 8;
shl.b32 %r745, %r738, 17;
or.b32 %r746, %r744, %r745;
and.b32 %r747, %r746, 25166208;
or.b32 %r748, %r747, -2139062144;
add.s32 %r749, %r748, -16843009;
or.b32 %r594, %r749, %r747;
mov.u32 %r104, %r594;

prmt.b32 %r104, %r104, 0, 0xba98;

ld.global.nc.v2.u32 {%r750, %r751}, [%rd39];
not.b32 %r753, %r750;
not.b32 %r755, %r751;
xor.b32 %r756, %r750, %r102;
or.b32 %r757, %r756, -2139062144;
and.b32 %r758, %r102, 2139062143;
and.b32 %r759, %r753, -2139062144;
sub.s32 %r760, %r757, %r758;
xor.b32 %r107, %r759, %r760;
xor.b32 %r761, %r751, %r104;
or.b32 %r762, %r761, -2139062144;
and.b32 %r763, %r104, 2139062143;
and.b32 %r764, %r755, -2139062144;
sub.s32 %r765, %r762, %r763;
xor.b32 %r111, %r764, %r765;
ld.global.nc.u32 %r112, [%rd35+24];
ld.global.nc.u32 %r108, [%rd35+20];

dp4a.s32.s32 %r106, %r107, %r108, %r648;


dp4a.s32.s32 %r110, %r111, %r112, %r106;

mul.wide.u16 %r766, %rs17, 4;
and.b32 %r767, %r766, 768;
or.b32 %r768, %r767, %r667;
mul.wide.u32 %rd40, %r768, 8;
add.s64 %rd41, %rd32, %rd40;
shr.u16 %rs31, %rs19, 8;
mul.wide.u16 %r769, %rs31, 128;
cvt.u32.u16 %r770, %rs31;
shl.b32 %r771, %r770, 21;
or.b32 %r772, %r769, %r771;
and.b32 %r773, %r772, 25166208;
or.b32 %r774, %r773, -2139062144;
add.s32 %r775, %r774, -16843009;
or.b32 %r604, %r775, %r773;
mov.u32 %r114, %r604;

prmt.b32 %r114, %r114, 0, 0xba98;

mul.wide.u16 %r776, %rs31, 8;
shl.b32 %r777, %r770, 17;
or.b32 %r778, %r776, %r777;
and.b32 %r779, %r778, 25166208;
or.b32 %r780, %r779, -2139062144;
add.s32 %r781, %r780, -16843009;
or.b32 %r606, %r781, %r779;
mov.u32 %r116, %r606;

prmt.b32 %r116, %r116, 0, 0xba98;

ld.global.nc.v2.u32 {%r782, %r783}, [%rd41];
not.b32 %r785, %r782;
not.b32 %r787, %r783;
xor.b32 %r788, %r782, %r114;
or.b32 %r789, %r788, -2139062144;
and.b32 %r790, %r114, 2139062143;
and.b32 %r791, %r785, -2139062144;
sub.s32 %r792, %r789, %r790;
xor.b32 %r119, %r791, %r792;
xor.b32 %r793, %r783, %r116;
or.b32 %r794, %r793, -2139062144;
and.b32 %r795, %r116, 2139062143;
and.b32 %r796, %r787, -2139062144;
sub.s32 %r797, %r794, %r795;
xor.b32 %r123, %r796, %r797;
ld.global.nc.u32 %r124, [%rd35+32];
ld.global.nc.u32 %r120, [%rd35+28];

dp4a.s32.s32 %r118, %r119, %r120, %r110;


dp4a.s32.s32 %r122, %r123, %r124, %r118;

cvt.u32.u16 %r798, %rs22;
cvt.u32.u16 %r799, %rs23;
mul.lo.s32 %r800, %r98, %r798;
mad.lo.s32 %r801, %r122, %r799, %r800;
add.s32 %r802, %r122, %r98;
shr.u32 %r803, %r802, 31;
add.s32 %r804, %r802, %r803;
shr.s32 %r805, %r804, 1;
add.s32 %r806, %r801, %r805;
shr.s32 %r807, %r806, 31;
shr.u32 %r808, %r807, 30;
add.s32 %r809, %r806, %r808;
shr.s32 %r810, %r809, 2;
add.s64 %rd42, %rd123, %rd5;
ld.global.nc.u16 %rs1, [%rd42];

{ cvt.f32.f16 %f73, %rs1;}


ld.global.nc.u32 %r126, [%rd35];

{.reg .f16 low,high;
mov.b32 {low,high},%r126;
cvt.f32.f16 %f74, low;}


mul.ftz.f32 %f97, %f73, %f74;
cvt.rn.f32.s32 %f98, %r810;
fma.rn.ftz.f32 %f276, %f97, %f98, %f276;
add.s64 %rd43, %rd123, %rd9;
ld.global.nc.u16 %rs32, [%rd43+-32];
shr.u16 %rs33, %rs32, 8;
cvt.u32.u16 %r811, %rs33;
ld.global.nc.u16 %rs34, [%rd43+-30];
shr.u16 %rs35, %rs34, 8;
cvt.u32.u16 %r812, %rs35;
add.s64 %rd44, %rd123, %rd7;
ld.global.nc.u8 %rs36, [%rd44];
ld.global.nc.u16 %rs37, [%rd43];
ld.global.nc.u16 %rs38, [%rd43+2];
ld.global.nc.u8 %rs39, [%rd44+8];
and.b16 %rs40, %rs39, 240;
and.b16 %rs41, %rs39, 15;
shr.u16 %rs42, %rs40, 4;
and.b16 %rs43, %rs32, 255;
cvt.u32.u16 %r813, %rs43;
shl.b16 %rs44, %rs36, 8;
cvt.u32.u16 %r814, %rs44;
and.b32 %r815, %r814, 768;
or.b32 %r816, %r815, %r813;
mul.wide.u32 %rd45, %r816, 8;
add.s64 %rd46, %rd32, %rd45;
and.b16 %rs45, %rs37, 255;
mul.wide.u16 %r817, %rs45, 128;
cvt.u32.u16 %r818, %rs37;
shl.b32 %r819, %r818, 21;
or.b32 %r820, %r817, %r819;
and.b32 %r821, %r820, 25166208;
or.b32 %r822, %r821, -2139062144;
add.s32 %r823, %r822, -16843009;
or.b32 %r617, %r823, %r821;
mov.u32 %r127, %r617;

prmt.b32 %r127, %r127, 0, 0xba98;

mul.wide.u16 %r824, %rs45, 8;
shl.b32 %r825, %r818, 17;
or.b32 %r826, %r824, %r825;
and.b32 %r827, %r826, 25166208;
or.b32 %r828, %r827, -2139062144;
add.s32 %r829, %r828, -16843009;
or.b32 %r619, %r829, %r827;
mov.u32 %r129, %r619;

prmt.b32 %r129, %r129, 0, 0xba98;

ld.global.nc.v2.u32 {%r830, %r831}, [%rd46];
not.b32 %r833, %r830;
not.b32 %r835, %r831;
xor.b32 %r836, %r830, %r127;
or.b32 %r837, %r836, -2139062144;
and.b32 %r838, %r127, 2139062143;
and.b32 %r839, %r833, -2139062144;
sub.s32 %r840, %r837, %r838;
xor.b32 %r132, %r839, %r840;
xor.b32 %r841, %r831, %r129;
or.b32 %r842, %r841, -2139062144;
and.b32 %r843, %r129, 2139062143;
and.b32 %r844, %r835, -2139062144;
sub.s32 %r845, %r842, %r843;
xor.b32 %r136, %r844, %r845;

dp4a.s32.s32 %r131, %r132, %r84, %r648;


dp4a.s32.s32 %r135, %r136, %r88, %r131;

mul.wide.u16 %r846, %rs36, 64;
and.b32 %r847, %r846, 768;
or.b32 %r848, %r847, %r811;
mul.wide.u32 %rd47, %r848, 8;
add.s64 %rd48, %rd32, %rd47;
shr.u16 %rs47, %rs37, 8;
mul.wide.u16 %r849, %rs47, 128;
cvt.u32.u16 %r850, %rs47;
shl.b32 %r851, %r850, 21;
or.b32 %r852, %r849, %r851;
and.b32 %r853, %r852, 25166208;
or.b32 %r854, %r853, -2139062144;
add.s32 %r855, %r854, -16843009;
or.b32 %r629, %r855, %r853;
mov.u32 %r139, %r629;

prmt.b32 %r139, %r139, 0, 0xba98;

mul.wide.u16 %r856, %rs47, 8;
shl.b32 %r857, %r850, 17;
or.b32 %r858, %r856, %r857;
and.b32 %r859, %r858, 25166208;
or.b32 %r860, %r859, -2139062144;
add.s32 %r861, %r860, -16843009;
or.b32 %r631, %r861, %r859;
mov.u32 %r141, %r631;

prmt.b32 %r141, %r141, 0, 0xba98;

ld.global.nc.v2.u32 {%r862, %r863}, [%rd48];
not.b32 %r865, %r862;
not.b32 %r867, %r863;
xor.b32 %r868, %r862, %r139;
or.b32 %r869, %r868, -2139062144;
and.b32 %r870, %r139, 2139062143;
and.b32 %r871, %r865, -2139062144;
sub.s32 %r872, %r869, %r870;
xor.b32 %r144, %r871, %r872;
xor.b32 %r873, %r863, %r141;
or.b32 %r874, %r873, -2139062144;
and.b32 %r875, %r141, 2139062143;
and.b32 %r876, %r867, -2139062144;
sub.s32 %r877, %r874, %r875;
xor.b32 %r148, %r876, %r877;

dp4a.s32.s32 %r143, %r144, %r96, %r135;


dp4a.s32.s32 %r147, %r148, %r100, %r143;

and.b16 %rs48, %rs34, 255;
cvt.u32.u16 %r878, %rs48;
mul.wide.u16 %r879, %rs36, 16;
and.b32 %r880, %r879, 768;
or.b32 %r881, %r880, %r878;
mul.wide.u32 %rd49, %r881, 8;
add.s64 %rd50, %rd32, %rd49;
and.b16 %rs49, %rs38, 255;
mul.wide.u16 %r882, %rs49, 128;
cvt.u32.u16 %r883, %rs38;
shl.b32 %r884, %r883, 21;
or.b32 %r885, %r882, %r884;
and.b32 %r886, %r885, 25166208;
or.b32 %r887, %r886, -2139062144;
add.s32 %r888, %r887, -16843009;
or.b32 %r641, %r888, %r886;
mov.u32 %r151, %r641;

prmt.b32 %r151, %r151, 0, 0xba98;

mul.wide.u16 %r889, %rs49, 8;
shl.b32 %r890, %r883, 17;
or.b32 %r891, %r889, %r890;
and.b32 %r892, %r891, 25166208;
or.b32 %r893, %r892, -2139062144;
add.s32 %r894, %r893, -16843009;
or.b32 %r643, %r894, %r892;
mov.u32 %r153, %r643;

prmt.b32 %r153, %r153, 0, 0xba98;

ld.global.nc.v2.u32 {%r895, %r896}, [%rd50];
not.b32 %r898, %r895;
not.b32 %r900, %r896;
xor.b32 %r901, %r895, %r151;
or.b32 %r902, %r901, -2139062144;
and.b32 %r903, %r151, 2139062143;
and.b32 %r904, %r898, -2139062144;
sub.s32 %r905, %r902, %r903;
xor.b32 %r156, %r904, %r905;
xor.b32 %r906, %r896, %r153;
or.b32 %r907, %r906, -2139062144;
and.b32 %r908, %r153, 2139062143;
and.b32 %r909, %r900, -2139062144;
sub.s32 %r910, %r907, %r908;
xor.b32 %r160, %r909, %r910;

dp4a.s32.s32 %r155, %r156, %r108, %r648;


dp4a.s32.s32 %r159, %r160, %r112, %r155;

mul.wide.u16 %r911, %rs36, 4;
and.b32 %r912, %r911, 768;
or.b32 %r913, %r912, %r812;
mul.wide.u32 %rd51, %r913, 8;
add.s64 %rd52, %rd32, %rd51;
shr.u16 %rs50, %rs38, 8;
mul.wide.u16 %r914, %rs50, 128;
cvt.u32.u16 %r915, %rs50;
shl.b32 %r916, %r915, 21;
or.b32 %r917, %r914, %r916;
and.b32 %r918, %r917, 25166208;
or.b32 %r919, %r918, -2139062144;
add.s32 %r920, %r919, -16843009;
or.b32 %r653, %r920, %r918;
mov.u32 %r163, %r653;

prmt.b32 %r163, %r163, 0, 0xba98;

mul.wide.u16 %r921, %rs50, 8;
shl.b32 %r922, %r915, 17;
or.b32 %r923, %r921, %r922;
and.b32 %r924, %r923, 25166208;
or.b32 %r925, %r924, -2139062144;
add.s32 %r926, %r925, -16843009;
or.b32 %r655, %r926, %r924;
mov.u32 %r165, %r655;

prmt.b32 %r165, %r165, 0, 0xba98;

ld.global.nc.v2.u32 {%r927, %r928}, [%rd52];
not.b32 %r930, %r927;
not.b32 %r932, %r928;
xor.b32 %r933, %r927, %r163;
or.b32 %r934, %r933, -2139062144;
and.b32 %r935, %r163, 2139062143;
and.b32 %r936, %r930, -2139062144;
sub.s32 %r937, %r934, %r935;
xor.b32 %r168, %r936, %r937;
xor.b32 %r938, %r928, %r165;
or.b32 %r939, %r938, -2139062144;
and.b32 %r940, %r165, 2139062143;
and.b32 %r941, %r932, -2139062144;
sub.s32 %r942, %r939, %r940;
xor.b32 %r172, %r941, %r942;

dp4a.s32.s32 %r167, %r168, %r120, %r159;


dp4a.s32.s32 %r171, %r172, %r124, %r167;

cvt.u32.u16 %r943, %rs41;
cvt.u32.u16 %r944, %rs42;
mul.lo.s32 %r945, %r147, %r943;
mad.lo.s32 %r946, %r171, %r944, %r945;
add.s32 %r947, %r171, %r147;
shr.u32 %r948, %r947, 31;
add.s32 %r949, %r947, %r948;
shr.s32 %r950, %r949, 1;
add.s32 %r951, %r946, %r950;
shr.s32 %r952, %r951, 31;
shr.u32 %r953, %r952, 30;
add.s32 %r954, %r951, %r953;
shr.s32 %r955, %r954, 2;
add.s64 %rd53, %rd123, %rd4;
ld.global.nc.u16 %rs2, [%rd53];

{ cvt.f32.f16 %f75, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r126;
cvt.f32.f16 %f76, low;}


mul.ftz.f32 %f99, %f75, %f76;
cvt.rn.f32.s32 %f100, %r955;
fma.rn.ftz.f32 %f275, %f99, %f100, %f275;
cvt.u64.u32 %rd54, %r1624;
add.s64 %rd55, %rd1, %rd54;
add.s64 %rd56, %rd55, %rd3;
mov.u32 %r176, %r568;

prmt.b32 %r176, %r176, 0, 0xba98;

mov.u32 %r178, %r570;

prmt.b32 %r178, %r178, 0, 0xba98;

xor.b32 %r956, %r685, %r176;
or.b32 %r957, %r956, -2139062144;
and.b32 %r958, %r176, 2139062143;
sub.s32 %r959, %r957, %r958;
xor.b32 %r181, %r694, %r959;
xor.b32 %r960, %r686, %r178;
or.b32 %r961, %r960, -2139062144;
and.b32 %r962, %r178, 2139062143;
sub.s32 %r963, %r961, %r962;
xor.b32 %r185, %r699, %r963;
mul.lo.s64 %rd57, %rd56, 36;
add.s64 %rd58, %rd10, %rd57;
ld.global.nc.u32 %r186, [%rd58+8];
ld.global.nc.u32 %r182, [%rd58+4];

dp4a.s32.s32 %r180, %r181, %r182, %r648;


dp4a.s32.s32 %r184, %r185, %r186, %r180;

mov.u32 %r188, %r580;

prmt.b32 %r188, %r188, 0, 0xba98;

mov.u32 %r190, %r582;

prmt.b32 %r190, %r190, 0, 0xba98;

xor.b32 %r964, %r717, %r188;
or.b32 %r965, %r964, -2139062144;
and.b32 %r966, %r188, 2139062143;
sub.s32 %r967, %r965, %r966;
xor.b32 %r193, %r726, %r967;
xor.b32 %r968, %r718, %r190;
or.b32 %r969, %r968, -2139062144;
and.b32 %r970, %r190, 2139062143;
sub.s32 %r971, %r969, %r970;
xor.b32 %r197, %r731, %r971;
ld.global.nc.u32 %r198, [%rd58+16];
ld.global.nc.u32 %r194, [%rd58+12];

dp4a.s32.s32 %r192, %r193, %r194, %r184;


dp4a.s32.s32 %r196, %r197, %r198, %r192;

mov.u32 %r200, %r592;

prmt.b32 %r200, %r200, 0, 0xba98;

mov.u32 %r202, %r594;

prmt.b32 %r202, %r202, 0, 0xba98;

xor.b32 %r972, %r750, %r200;
or.b32 %r973, %r972, -2139062144;
and.b32 %r974, %r200, 2139062143;
sub.s32 %r975, %r973, %r974;
xor.b32 %r205, %r759, %r975;
xor.b32 %r976, %r751, %r202;
or.b32 %r977, %r976, -2139062144;
and.b32 %r978, %r202, 2139062143;
sub.s32 %r979, %r977, %r978;
xor.b32 %r209, %r764, %r979;
ld.global.nc.u32 %r210, [%rd58+24];
ld.global.nc.u32 %r206, [%rd58+20];

dp4a.s32.s32 %r204, %r205, %r206, %r648;


dp4a.s32.s32 %r208, %r209, %r210, %r204;

mov.u32 %r212, %r604;

prmt.b32 %r212, %r212, 0, 0xba98;

mov.u32 %r214, %r606;

prmt.b32 %r214, %r214, 0, 0xba98;

xor.b32 %r980, %r782, %r212;
or.b32 %r981, %r980, -2139062144;
and.b32 %r982, %r212, 2139062143;
sub.s32 %r983, %r981, %r982;
xor.b32 %r217, %r791, %r983;
xor.b32 %r984, %r783, %r214;
or.b32 %r985, %r984, -2139062144;
and.b32 %r986, %r214, 2139062143;
sub.s32 %r987, %r985, %r986;
xor.b32 %r221, %r796, %r987;
ld.global.nc.u32 %r222, [%rd58+32];
ld.global.nc.u32 %r218, [%rd58+28];

dp4a.s32.s32 %r216, %r217, %r218, %r208;


dp4a.s32.s32 %r220, %r221, %r222, %r216;

mul.lo.s32 %r988, %r196, %r798;
mad.lo.s32 %r989, %r220, %r799, %r988;
add.s32 %r990, %r220, %r196;
shr.u32 %r991, %r990, 31;
add.s32 %r992, %r990, %r991;
shr.s32 %r993, %r992, 1;
add.s32 %r994, %r989, %r993;
shr.s32 %r995, %r994, 31;
shr.u32 %r996, %r995, 30;
add.s32 %r997, %r994, %r996;
shr.s32 %r998, %r997, 2;

{ cvt.f32.f16 %f77, %rs1;}


ld.global.nc.u32 %r224, [%rd58];

{.reg .f16 low,high;
mov.b32 {low,high},%r224;
cvt.f32.f16 %f78, low;}


mul.ftz.f32 %f101, %f77, %f78;
cvt.rn.f32.s32 %f102, %r998;
fma.rn.ftz.f32 %f274, %f101, %f102, %f274;
mov.u32 %r225, %r617;

prmt.b32 %r225, %r225, 0, 0xba98;

mov.u32 %r227, %r619;

prmt.b32 %r227, %r227, 0, 0xba98;

xor.b32 %r999, %r830, %r225;
or.b32 %r1000, %r999, -2139062144;
and.b32 %r1001, %r225, 2139062143;
sub.s32 %r1002, %r1000, %r1001;
xor.b32 %r230, %r839, %r1002;
xor.b32 %r1003, %r831, %r227;
or.b32 %r1004, %r1003, -2139062144;
and.b32 %r1005, %r227, 2139062143;
sub.s32 %r1006, %r1004, %r1005;
xor.b32 %r234, %r844, %r1006;

dp4a.s32.s32 %r229, %r230, %r182, %r648;


dp4a.s32.s32 %r233, %r234, %r186, %r229;

mov.u32 %r237, %r629;

prmt.b32 %r237, %r237, 0, 0xba98;

mov.u32 %r239, %r631;

prmt.b32 %r239, %r239, 0, 0xba98;

xor.b32 %r1007, %r862, %r237;
or.b32 %r1008, %r1007, -2139062144;
and.b32 %r1009, %r237, 2139062143;
sub.s32 %r1010, %r1008, %r1009;
xor.b32 %r242, %r871, %r1010;
xor.b32 %r1011, %r863, %r239;
or.b32 %r1012, %r1011, -2139062144;
and.b32 %r1013, %r239, 2139062143;
sub.s32 %r1014, %r1012, %r1013;
xor.b32 %r246, %r876, %r1014;

dp4a.s32.s32 %r241, %r242, %r194, %r233;


dp4a.s32.s32 %r245, %r246, %r198, %r241;

mov.u32 %r249, %r641;

prmt.b32 %r249, %r249, 0, 0xba98;

mov.u32 %r251, %r643;

prmt.b32 %r251, %r251, 0, 0xba98;

xor.b32 %r1015, %r895, %r249;
or.b32 %r1016, %r1015, -2139062144;
and.b32 %r1017, %r249, 2139062143;
sub.s32 %r1018, %r1016, %r1017;
xor.b32 %r254, %r904, %r1018;
xor.b32 %r1019, %r896, %r251;
or.b32 %r1020, %r1019, -2139062144;
and.b32 %r1021, %r251, 2139062143;
sub.s32 %r1022, %r1020, %r1021;
xor.b32 %r258, %r909, %r1022;

dp4a.s32.s32 %r253, %r254, %r206, %r648;


dp4a.s32.s32 %r257, %r258, %r210, %r253;

mov.u32 %r261, %r653;

prmt.b32 %r261, %r261, 0, 0xba98;

mov.u32 %r263, %r655;

prmt.b32 %r263, %r263, 0, 0xba98;

xor.b32 %r1023, %r927, %r261;
or.b32 %r1024, %r1023, -2139062144;
and.b32 %r1025, %r261, 2139062143;
sub.s32 %r1026, %r1024, %r1025;
xor.b32 %r266, %r936, %r1026;
xor.b32 %r1027, %r928, %r263;
or.b32 %r1028, %r1027, -2139062144;
and.b32 %r1029, %r263, 2139062143;
sub.s32 %r1030, %r1028, %r1029;
xor.b32 %r270, %r941, %r1030;

dp4a.s32.s32 %r265, %r266, %r218, %r257;


dp4a.s32.s32 %r269, %r270, %r222, %r265;

mul.lo.s32 %r1031, %r245, %r943;
mad.lo.s32 %r1032, %r269, %r944, %r1031;
add.s32 %r1033, %r269, %r245;
shr.u32 %r1034, %r1033, 31;
add.s32 %r1035, %r1033, %r1034;
shr.s32 %r1036, %r1035, 1;
add.s32 %r1037, %r1032, %r1036;
shr.s32 %r1038, %r1037, 31;
shr.u32 %r1039, %r1038, 30;
add.s32 %r1040, %r1037, %r1039;
shr.s32 %r1041, %r1040, 2;

{ cvt.f32.f16 %f79, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r224;
cvt.f32.f16 %f80, low;}


mul.ftz.f32 %f103, %f79, %f80;
cvt.rn.f32.s32 %f104, %r1041;
fma.rn.ftz.f32 %f273, %f103, %f104, %f273;
cvt.u64.u32 %rd59, %r1628;
add.s64 %rd60, %rd1, %rd59;
add.s64 %rd61, %rd60, %rd3;
mov.u32 %r274, %r568;

prmt.b32 %r274, %r274, 0, 0xba98;

mov.u32 %r276, %r570;

prmt.b32 %r276, %r276, 0, 0xba98;

xor.b32 %r1042, %r685, %r274;
or.b32 %r1043, %r1042, -2139062144;
and.b32 %r1044, %r274, 2139062143;
sub.s32 %r1045, %r1043, %r1044;
xor.b32 %r279, %r694, %r1045;
xor.b32 %r1046, %r686, %r276;
or.b32 %r1047, %r1046, -2139062144;
and.b32 %r1048, %r276, 2139062143;
sub.s32 %r1049, %r1047, %r1048;
xor.b32 %r283, %r699, %r1049;
mul.lo.s64 %rd62, %rd61, 36;
add.s64 %rd63, %rd10, %rd62;
ld.global.nc.u32 %r284, [%rd63+8];
ld.global.nc.u32 %r280, [%rd63+4];

dp4a.s32.s32 %r278, %r279, %r280, %r648;


dp4a.s32.s32 %r282, %r283, %r284, %r278;

mov.u32 %r286, %r580;

prmt.b32 %r286, %r286, 0, 0xba98;

mov.u32 %r288, %r582;

prmt.b32 %r288, %r288, 0, 0xba98;

xor.b32 %r1050, %r717, %r286;
or.b32 %r1051, %r1050, -2139062144;
and.b32 %r1052, %r286, 2139062143;
sub.s32 %r1053, %r1051, %r1052;
xor.b32 %r291, %r726, %r1053;
xor.b32 %r1054, %r718, %r288;
or.b32 %r1055, %r1054, -2139062144;
and.b32 %r1056, %r288, 2139062143;
sub.s32 %r1057, %r1055, %r1056;
xor.b32 %r295, %r731, %r1057;
ld.global.nc.u32 %r296, [%rd63+16];
ld.global.nc.u32 %r292, [%rd63+12];

dp4a.s32.s32 %r290, %r291, %r292, %r282;


dp4a.s32.s32 %r294, %r295, %r296, %r290;

mov.u32 %r298, %r592;

prmt.b32 %r298, %r298, 0, 0xba98;

mov.u32 %r300, %r594;

prmt.b32 %r300, %r300, 0, 0xba98;

xor.b32 %r1058, %r750, %r298;
or.b32 %r1059, %r1058, -2139062144;
and.b32 %r1060, %r298, 2139062143;
sub.s32 %r1061, %r1059, %r1060;
xor.b32 %r303, %r759, %r1061;
xor.b32 %r1062, %r751, %r300;
or.b32 %r1063, %r1062, -2139062144;
and.b32 %r1064, %r300, 2139062143;
sub.s32 %r1065, %r1063, %r1064;
xor.b32 %r307, %r764, %r1065;
ld.global.nc.u32 %r308, [%rd63+24];
ld.global.nc.u32 %r304, [%rd63+20];

dp4a.s32.s32 %r302, %r303, %r304, %r648;


dp4a.s32.s32 %r306, %r307, %r308, %r302;

mov.u32 %r310, %r604;

prmt.b32 %r310, %r310, 0, 0xba98;

mov.u32 %r312, %r606;

prmt.b32 %r312, %r312, 0, 0xba98;

xor.b32 %r1066, %r782, %r310;
or.b32 %r1067, %r1066, -2139062144;
and.b32 %r1068, %r310, 2139062143;
sub.s32 %r1069, %r1067, %r1068;
xor.b32 %r315, %r791, %r1069;
xor.b32 %r1070, %r783, %r312;
or.b32 %r1071, %r1070, -2139062144;
and.b32 %r1072, %r312, 2139062143;
sub.s32 %r1073, %r1071, %r1072;
xor.b32 %r319, %r796, %r1073;
ld.global.nc.u32 %r320, [%rd63+32];
ld.global.nc.u32 %r316, [%rd63+28];

dp4a.s32.s32 %r314, %r315, %r316, %r306;


dp4a.s32.s32 %r318, %r319, %r320, %r314;

mul.lo.s32 %r1074, %r294, %r798;
mad.lo.s32 %r1075, %r318, %r799, %r1074;
add.s32 %r1076, %r318, %r294;
shr.u32 %r1077, %r1076, 31;
add.s32 %r1078, %r1076, %r1077;
shr.s32 %r1079, %r1078, 1;
add.s32 %r1080, %r1075, %r1079;
shr.s32 %r1081, %r1080, 31;
shr.u32 %r1082, %r1081, 30;
add.s32 %r1083, %r1080, %r1082;
shr.s32 %r1084, %r1083, 2;

{ cvt.f32.f16 %f81, %rs1;}


ld.global.nc.u32 %r322, [%rd63];

{.reg .f16 low,high;
mov.b32 {low,high},%r322;
cvt.f32.f16 %f82, low;}


mul.ftz.f32 %f105, %f81, %f82;
cvt.rn.f32.s32 %f106, %r1084;
fma.rn.ftz.f32 %f272, %f105, %f106, %f272;
mov.u32 %r323, %r617;

prmt.b32 %r323, %r323, 0, 0xba98;

mov.u32 %r325, %r619;

prmt.b32 %r325, %r325, 0, 0xba98;

xor.b32 %r1085, %r830, %r323;
or.b32 %r1086, %r1085, -2139062144;
and.b32 %r1087, %r323, 2139062143;
sub.s32 %r1088, %r1086, %r1087;
xor.b32 %r328, %r839, %r1088;
xor.b32 %r1089, %r831, %r325;
or.b32 %r1090, %r1089, -2139062144;
and.b32 %r1091, %r325, 2139062143;
sub.s32 %r1092, %r1090, %r1091;
xor.b32 %r332, %r844, %r1092;

dp4a.s32.s32 %r327, %r328, %r280, %r648;


dp4a.s32.s32 %r331, %r332, %r284, %r327;

mov.u32 %r335, %r629;

prmt.b32 %r335, %r335, 0, 0xba98;

mov.u32 %r337, %r631;

prmt.b32 %r337, %r337, 0, 0xba98;

xor.b32 %r1093, %r862, %r335;
or.b32 %r1094, %r1093, -2139062144;
and.b32 %r1095, %r335, 2139062143;
sub.s32 %r1096, %r1094, %r1095;
xor.b32 %r340, %r871, %r1096;
xor.b32 %r1097, %r863, %r337;
or.b32 %r1098, %r1097, -2139062144;
and.b32 %r1099, %r337, 2139062143;
sub.s32 %r1100, %r1098, %r1099;
xor.b32 %r344, %r876, %r1100;

dp4a.s32.s32 %r339, %r340, %r292, %r331;


dp4a.s32.s32 %r343, %r344, %r296, %r339;

mov.u32 %r347, %r641;

prmt.b32 %r347, %r347, 0, 0xba98;

mov.u32 %r349, %r643;

prmt.b32 %r349, %r349, 0, 0xba98;

xor.b32 %r1101, %r895, %r347;
or.b32 %r1102, %r1101, -2139062144;
and.b32 %r1103, %r347, 2139062143;
sub.s32 %r1104, %r1102, %r1103;
xor.b32 %r352, %r904, %r1104;
xor.b32 %r1105, %r896, %r349;
or.b32 %r1106, %r1105, -2139062144;
and.b32 %r1107, %r349, 2139062143;
sub.s32 %r1108, %r1106, %r1107;
xor.b32 %r356, %r909, %r1108;

dp4a.s32.s32 %r351, %r352, %r304, %r648;


dp4a.s32.s32 %r355, %r356, %r308, %r351;

mov.u32 %r359, %r653;

prmt.b32 %r359, %r359, 0, 0xba98;

mov.u32 %r361, %r655;

prmt.b32 %r361, %r361, 0, 0xba98;

xor.b32 %r1109, %r927, %r359;
or.b32 %r1110, %r1109, -2139062144;
and.b32 %r1111, %r359, 2139062143;
sub.s32 %r1112, %r1110, %r1111;
xor.b32 %r364, %r936, %r1112;
xor.b32 %r1113, %r928, %r361;
or.b32 %r1114, %r1113, -2139062144;
and.b32 %r1115, %r361, 2139062143;
sub.s32 %r1116, %r1114, %r1115;
xor.b32 %r368, %r941, %r1116;

dp4a.s32.s32 %r363, %r364, %r316, %r355;


dp4a.s32.s32 %r367, %r368, %r320, %r363;

mul.lo.s32 %r1117, %r343, %r943;
mad.lo.s32 %r1118, %r367, %r944, %r1117;
add.s32 %r1119, %r367, %r343;
shr.u32 %r1120, %r1119, 31;
add.s32 %r1121, %r1119, %r1120;
shr.s32 %r1122, %r1121, 1;
add.s32 %r1123, %r1118, %r1122;
shr.s32 %r1124, %r1123, 31;
shr.u32 %r1125, %r1124, 30;
add.s32 %r1126, %r1123, %r1125;
shr.s32 %r1127, %r1126, 2;

{ cvt.f32.f16 %f83, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r322;
cvt.f32.f16 %f84, low;}


mul.ftz.f32 %f107, %f83, %f84;
cvt.rn.f32.s32 %f108, %r1127;
fma.rn.ftz.f32 %f271, %f107, %f108, %f271;
cvt.u64.u32 %rd64, %r1627;
add.s64 %rd65, %rd1, %rd64;
add.s64 %rd66, %rd65, %rd3;
mov.u32 %r372, %r568;

prmt.b32 %r372, %r372, 0, 0xba98;

mov.u32 %r374, %r570;

prmt.b32 %r374, %r374, 0, 0xba98;

xor.b32 %r1128, %r685, %r372;
or.b32 %r1129, %r1128, -2139062144;
and.b32 %r1130, %r372, 2139062143;
sub.s32 %r1131, %r1129, %r1130;
xor.b32 %r377, %r694, %r1131;
xor.b32 %r1132, %r686, %r374;
or.b32 %r1133, %r1132, -2139062144;
and.b32 %r1134, %r374, 2139062143;
sub.s32 %r1135, %r1133, %r1134;
xor.b32 %r381, %r699, %r1135;
mul.lo.s64 %rd67, %rd66, 36;
add.s64 %rd68, %rd10, %rd67;
ld.global.nc.u32 %r382, [%rd68+8];
ld.global.nc.u32 %r378, [%rd68+4];

dp4a.s32.s32 %r376, %r377, %r378, %r648;


dp4a.s32.s32 %r380, %r381, %r382, %r376;

mov.u32 %r384, %r580;

prmt.b32 %r384, %r384, 0, 0xba98;

mov.u32 %r386, %r582;

prmt.b32 %r386, %r386, 0, 0xba98;

xor.b32 %r1136, %r717, %r384;
or.b32 %r1137, %r1136, -2139062144;
and.b32 %r1138, %r384, 2139062143;
sub.s32 %r1139, %r1137, %r1138;
xor.b32 %r389, %r726, %r1139;
xor.b32 %r1140, %r718, %r386;
or.b32 %r1141, %r1140, -2139062144;
and.b32 %r1142, %r386, 2139062143;
sub.s32 %r1143, %r1141, %r1142;
xor.b32 %r393, %r731, %r1143;
ld.global.nc.u32 %r394, [%rd68+16];
ld.global.nc.u32 %r390, [%rd68+12];

dp4a.s32.s32 %r388, %r389, %r390, %r380;


dp4a.s32.s32 %r392, %r393, %r394, %r388;

mov.u32 %r396, %r592;

prmt.b32 %r396, %r396, 0, 0xba98;

mov.u32 %r398, %r594;

prmt.b32 %r398, %r398, 0, 0xba98;

xor.b32 %r1144, %r750, %r396;
or.b32 %r1145, %r1144, -2139062144;
and.b32 %r1146, %r396, 2139062143;
sub.s32 %r1147, %r1145, %r1146;
xor.b32 %r401, %r759, %r1147;
xor.b32 %r1148, %r751, %r398;
or.b32 %r1149, %r1148, -2139062144;
and.b32 %r1150, %r398, 2139062143;
sub.s32 %r1151, %r1149, %r1150;
xor.b32 %r405, %r764, %r1151;
ld.global.nc.u32 %r406, [%rd68+24];
ld.global.nc.u32 %r402, [%rd68+20];

dp4a.s32.s32 %r400, %r401, %r402, %r648;


dp4a.s32.s32 %r404, %r405, %r406, %r400;

mov.u32 %r408, %r604;

prmt.b32 %r408, %r408, 0, 0xba98;

mov.u32 %r410, %r606;

prmt.b32 %r410, %r410, 0, 0xba98;

xor.b32 %r1152, %r782, %r408;
or.b32 %r1153, %r1152, -2139062144;
and.b32 %r1154, %r408, 2139062143;
sub.s32 %r1155, %r1153, %r1154;
xor.b32 %r413, %r791, %r1155;
xor.b32 %r1156, %r783, %r410;
or.b32 %r1157, %r1156, -2139062144;
and.b32 %r1158, %r410, 2139062143;
sub.s32 %r1159, %r1157, %r1158;
xor.b32 %r417, %r796, %r1159;
ld.global.nc.u32 %r418, [%rd68+32];
ld.global.nc.u32 %r414, [%rd68+28];

dp4a.s32.s32 %r412, %r413, %r414, %r404;


dp4a.s32.s32 %r416, %r417, %r418, %r412;

mul.lo.s32 %r1160, %r392, %r798;
mad.lo.s32 %r1161, %r416, %r799, %r1160;
add.s32 %r1162, %r416, %r392;
shr.u32 %r1163, %r1162, 31;
add.s32 %r1164, %r1162, %r1163;
shr.s32 %r1165, %r1164, 1;
add.s32 %r1166, %r1161, %r1165;
shr.s32 %r1167, %r1166, 31;
shr.u32 %r1168, %r1167, 30;
add.s32 %r1169, %r1166, %r1168;
shr.s32 %r1170, %r1169, 2;

{ cvt.f32.f16 %f85, %rs1;}


ld.global.nc.u32 %r420, [%rd68];

{.reg .f16 low,high;
mov.b32 {low,high},%r420;
cvt.f32.f16 %f86, low;}


mul.ftz.f32 %f109, %f85, %f86;
cvt.rn.f32.s32 %f110, %r1170;
fma.rn.ftz.f32 %f270, %f109, %f110, %f270;
mov.u32 %r421, %r617;

prmt.b32 %r421, %r421, 0, 0xba98;

mov.u32 %r423, %r619;

prmt.b32 %r423, %r423, 0, 0xba98;

xor.b32 %r1171, %r830, %r421;
or.b32 %r1172, %r1171, -2139062144;
and.b32 %r1173, %r421, 2139062143;
sub.s32 %r1174, %r1172, %r1173;
xor.b32 %r426, %r839, %r1174;
xor.b32 %r1175, %r831, %r423;
or.b32 %r1176, %r1175, -2139062144;
and.b32 %r1177, %r423, 2139062143;
sub.s32 %r1178, %r1176, %r1177;
xor.b32 %r430, %r844, %r1178;

dp4a.s32.s32 %r425, %r426, %r378, %r648;


dp4a.s32.s32 %r429, %r430, %r382, %r425;

mov.u32 %r433, %r629;

prmt.b32 %r433, %r433, 0, 0xba98;

mov.u32 %r435, %r631;

prmt.b32 %r435, %r435, 0, 0xba98;

xor.b32 %r1179, %r862, %r433;
or.b32 %r1180, %r1179, -2139062144;
and.b32 %r1181, %r433, 2139062143;
sub.s32 %r1182, %r1180, %r1181;
xor.b32 %r438, %r871, %r1182;
xor.b32 %r1183, %r863, %r435;
or.b32 %r1184, %r1183, -2139062144;
and.b32 %r1185, %r435, 2139062143;
sub.s32 %r1186, %r1184, %r1185;
xor.b32 %r442, %r876, %r1186;

dp4a.s32.s32 %r437, %r438, %r390, %r429;


dp4a.s32.s32 %r441, %r442, %r394, %r437;

mov.u32 %r445, %r641;

prmt.b32 %r445, %r445, 0, 0xba98;

mov.u32 %r447, %r643;

prmt.b32 %r447, %r447, 0, 0xba98;

xor.b32 %r1187, %r895, %r445;
or.b32 %r1188, %r1187, -2139062144;
and.b32 %r1189, %r445, 2139062143;
sub.s32 %r1190, %r1188, %r1189;
xor.b32 %r450, %r904, %r1190;
xor.b32 %r1191, %r896, %r447;
or.b32 %r1192, %r1191, -2139062144;
and.b32 %r1193, %r447, 2139062143;
sub.s32 %r1194, %r1192, %r1193;
xor.b32 %r454, %r909, %r1194;

dp4a.s32.s32 %r449, %r450, %r402, %r648;


dp4a.s32.s32 %r453, %r454, %r406, %r449;

mov.u32 %r457, %r653;

prmt.b32 %r457, %r457, 0, 0xba98;

mov.u32 %r459, %r655;

prmt.b32 %r459, %r459, 0, 0xba98;

xor.b32 %r1195, %r927, %r457;
or.b32 %r1196, %r1195, -2139062144;
and.b32 %r1197, %r457, 2139062143;
sub.s32 %r1198, %r1196, %r1197;
xor.b32 %r462, %r936, %r1198;
xor.b32 %r1199, %r928, %r459;
or.b32 %r1200, %r1199, -2139062144;
and.b32 %r1201, %r459, 2139062143;
sub.s32 %r1202, %r1200, %r1201;
xor.b32 %r466, %r941, %r1202;

dp4a.s32.s32 %r461, %r462, %r414, %r453;


dp4a.s32.s32 %r465, %r466, %r418, %r461;

mul.lo.s32 %r1203, %r441, %r943;
mad.lo.s32 %r1204, %r465, %r944, %r1203;
add.s32 %r1205, %r465, %r441;
shr.u32 %r1206, %r1205, 31;
add.s32 %r1207, %r1205, %r1206;
shr.s32 %r1208, %r1207, 1;
add.s32 %r1209, %r1204, %r1208;
shr.s32 %r1210, %r1209, 31;
shr.u32 %r1211, %r1210, 30;
add.s32 %r1212, %r1209, %r1211;
shr.s32 %r1213, %r1212, 2;

{ cvt.f32.f16 %f87, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r420;
cvt.f32.f16 %f88, low;}


mul.ftz.f32 %f111, %f87, %f88;
cvt.rn.f32.s32 %f112, %r1213;
fma.rn.ftz.f32 %f269, %f111, %f112, %f269;
cvt.u64.u32 %rd69, %r1626;
add.s64 %rd70, %rd1, %rd69;
add.s64 %rd71, %rd70, %rd3;
mov.u32 %r470, %r568;

prmt.b32 %r470, %r470, 0, 0xba98;

mov.u32 %r472, %r570;

prmt.b32 %r472, %r472, 0, 0xba98;

xor.b32 %r1214, %r685, %r470;
or.b32 %r1215, %r1214, -2139062144;
and.b32 %r1216, %r470, 2139062143;
sub.s32 %r1217, %r1215, %r1216;
xor.b32 %r475, %r694, %r1217;
xor.b32 %r1218, %r686, %r472;
or.b32 %r1219, %r1218, -2139062144;
and.b32 %r1220, %r472, 2139062143;
sub.s32 %r1221, %r1219, %r1220;
xor.b32 %r479, %r699, %r1221;
mul.lo.s64 %rd72, %rd71, 36;
add.s64 %rd73, %rd10, %rd72;
ld.global.nc.u32 %r480, [%rd73+8];
ld.global.nc.u32 %r476, [%rd73+4];

dp4a.s32.s32 %r474, %r475, %r476, %r648;


dp4a.s32.s32 %r478, %r479, %r480, %r474;

mov.u32 %r482, %r580;

prmt.b32 %r482, %r482, 0, 0xba98;

mov.u32 %r484, %r582;

prmt.b32 %r484, %r484, 0, 0xba98;

xor.b32 %r1222, %r717, %r482;
or.b32 %r1223, %r1222, -2139062144;
and.b32 %r1224, %r482, 2139062143;
sub.s32 %r1225, %r1223, %r1224;
xor.b32 %r487, %r726, %r1225;
xor.b32 %r1226, %r718, %r484;
or.b32 %r1227, %r1226, -2139062144;
and.b32 %r1228, %r484, 2139062143;
sub.s32 %r1229, %r1227, %r1228;
xor.b32 %r491, %r731, %r1229;
ld.global.nc.u32 %r492, [%rd73+16];
ld.global.nc.u32 %r488, [%rd73+12];

dp4a.s32.s32 %r486, %r487, %r488, %r478;


dp4a.s32.s32 %r490, %r491, %r492, %r486;

mov.u32 %r494, %r592;

prmt.b32 %r494, %r494, 0, 0xba98;

mov.u32 %r496, %r594;

prmt.b32 %r496, %r496, 0, 0xba98;

xor.b32 %r1230, %r750, %r494;
or.b32 %r1231, %r1230, -2139062144;
and.b32 %r1232, %r494, 2139062143;
sub.s32 %r1233, %r1231, %r1232;
xor.b32 %r499, %r759, %r1233;
xor.b32 %r1234, %r751, %r496;
or.b32 %r1235, %r1234, -2139062144;
and.b32 %r1236, %r496, 2139062143;
sub.s32 %r1237, %r1235, %r1236;
xor.b32 %r503, %r764, %r1237;
ld.global.nc.u32 %r504, [%rd73+24];
ld.global.nc.u32 %r500, [%rd73+20];

dp4a.s32.s32 %r498, %r499, %r500, %r648;


dp4a.s32.s32 %r502, %r503, %r504, %r498;

mov.u32 %r506, %r604;

prmt.b32 %r506, %r506, 0, 0xba98;

mov.u32 %r508, %r606;

prmt.b32 %r508, %r508, 0, 0xba98;

xor.b32 %r1238, %r782, %r506;
or.b32 %r1239, %r1238, -2139062144;
and.b32 %r1240, %r506, 2139062143;
sub.s32 %r1241, %r1239, %r1240;
xor.b32 %r511, %r791, %r1241;
xor.b32 %r1242, %r783, %r508;
or.b32 %r1243, %r1242, -2139062144;
and.b32 %r1244, %r508, 2139062143;
sub.s32 %r1245, %r1243, %r1244;
xor.b32 %r515, %r796, %r1245;
ld.global.nc.u32 %r516, [%rd73+32];
ld.global.nc.u32 %r512, [%rd73+28];

dp4a.s32.s32 %r510, %r511, %r512, %r502;


dp4a.s32.s32 %r514, %r515, %r516, %r510;

mul.lo.s32 %r1246, %r490, %r798;
mad.lo.s32 %r1247, %r514, %r799, %r1246;
add.s32 %r1248, %r514, %r490;
shr.u32 %r1249, %r1248, 31;
add.s32 %r1250, %r1248, %r1249;
shr.s32 %r1251, %r1250, 1;
add.s32 %r1252, %r1247, %r1251;
shr.s32 %r1253, %r1252, 31;
shr.u32 %r1254, %r1253, 30;
add.s32 %r1255, %r1252, %r1254;
shr.s32 %r1256, %r1255, 2;

{ cvt.f32.f16 %f89, %rs1;}


ld.global.nc.u32 %r518, [%rd73];

{.reg .f16 low,high;
mov.b32 {low,high},%r518;
cvt.f32.f16 %f90, low;}


mul.ftz.f32 %f113, %f89, %f90;
cvt.rn.f32.s32 %f114, %r1256;
fma.rn.ftz.f32 %f268, %f113, %f114, %f268;
mov.u32 %r519, %r617;

prmt.b32 %r519, %r519, 0, 0xba98;

mov.u32 %r521, %r619;

prmt.b32 %r521, %r521, 0, 0xba98;

xor.b32 %r1257, %r830, %r519;
or.b32 %r1258, %r1257, -2139062144;
and.b32 %r1259, %r519, 2139062143;
sub.s32 %r1260, %r1258, %r1259;
xor.b32 %r524, %r839, %r1260;
xor.b32 %r1261, %r831, %r521;
or.b32 %r1262, %r1261, -2139062144;
and.b32 %r1263, %r521, 2139062143;
sub.s32 %r1264, %r1262, %r1263;
xor.b32 %r528, %r844, %r1264;

dp4a.s32.s32 %r523, %r524, %r476, %r648;


dp4a.s32.s32 %r527, %r528, %r480, %r523;

mov.u32 %r531, %r629;

prmt.b32 %r531, %r531, 0, 0xba98;

mov.u32 %r533, %r631;

prmt.b32 %r533, %r533, 0, 0xba98;

xor.b32 %r1265, %r862, %r531;
or.b32 %r1266, %r1265, -2139062144;
and.b32 %r1267, %r531, 2139062143;
sub.s32 %r1268, %r1266, %r1267;
xor.b32 %r536, %r871, %r1268;
xor.b32 %r1269, %r863, %r533;
or.b32 %r1270, %r1269, -2139062144;
and.b32 %r1271, %r533, 2139062143;
sub.s32 %r1272, %r1270, %r1271;
xor.b32 %r540, %r876, %r1272;

dp4a.s32.s32 %r535, %r536, %r488, %r527;


dp4a.s32.s32 %r539, %r540, %r492, %r535;

mov.u32 %r543, %r641;

prmt.b32 %r543, %r543, 0, 0xba98;

mov.u32 %r545, %r643;

prmt.b32 %r545, %r545, 0, 0xba98;

xor.b32 %r1273, %r895, %r543;
or.b32 %r1274, %r1273, -2139062144;
and.b32 %r1275, %r543, 2139062143;
sub.s32 %r1276, %r1274, %r1275;
xor.b32 %r548, %r904, %r1276;
xor.b32 %r1277, %r896, %r545;
or.b32 %r1278, %r1277, -2139062144;
and.b32 %r1279, %r545, 2139062143;
sub.s32 %r1280, %r1278, %r1279;
xor.b32 %r552, %r909, %r1280;

dp4a.s32.s32 %r547, %r548, %r500, %r648;


dp4a.s32.s32 %r551, %r552, %r504, %r547;

mov.u32 %r555, %r653;

prmt.b32 %r555, %r555, 0, 0xba98;

mov.u32 %r557, %r655;

prmt.b32 %r557, %r557, 0, 0xba98;

xor.b32 %r1281, %r927, %r555;
or.b32 %r1282, %r1281, -2139062144;
and.b32 %r1283, %r555, 2139062143;
sub.s32 %r1284, %r1282, %r1283;
xor.b32 %r560, %r936, %r1284;
xor.b32 %r1285, %r928, %r557;
or.b32 %r1286, %r1285, -2139062144;
and.b32 %r1287, %r557, 2139062143;
sub.s32 %r1288, %r1286, %r1287;
xor.b32 %r564, %r941, %r1288;

dp4a.s32.s32 %r559, %r560, %r512, %r551;


dp4a.s32.s32 %r563, %r564, %r516, %r559;

mul.lo.s32 %r1289, %r539, %r943;
mad.lo.s32 %r1290, %r563, %r944, %r1289;
add.s32 %r1291, %r563, %r539;
shr.u32 %r1292, %r1291, 31;
add.s32 %r1293, %r1291, %r1292;
shr.s32 %r1294, %r1293, 1;
add.s32 %r1295, %r1290, %r1294;
shr.s32 %r1296, %r1295, 31;
shr.u32 %r1297, %r1296, 30;
add.s32 %r1298, %r1295, %r1297;
shr.s32 %r1299, %r1298, 2;

{ cvt.f32.f16 %f91, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r518;
cvt.f32.f16 %f92, low;}


mul.ftz.f32 %f115, %f91, %f92;
cvt.rn.f32.s32 %f116, %r1299;
fma.rn.ftz.f32 %f267, %f115, %f116, %f267;
cvt.u64.u32 %rd74, %r1625;
add.s64 %rd75, %rd1, %rd74;
add.s64 %rd76, %rd75, %rd3;

prmt.b32 %r568, %r568, 0, 0xba98;


prmt.b32 %r570, %r570, 0, 0xba98;

xor.b32 %r1300, %r685, %r568;
or.b32 %r1301, %r1300, -2139062144;
and.b32 %r1302, %r568, 2139062143;
sub.s32 %r1303, %r1301, %r1302;
xor.b32 %r573, %r694, %r1303;
xor.b32 %r1304, %r686, %r570;
or.b32 %r1305, %r1304, -2139062144;
and.b32 %r1306, %r570, 2139062143;
sub.s32 %r1307, %r1305, %r1306;
xor.b32 %r577, %r699, %r1307;
mul.lo.s64 %rd77, %rd76, 36;
add.s64 %rd78, %rd10, %rd77;
ld.global.nc.u32 %r578, [%rd78+8];
ld.global.nc.u32 %r574, [%rd78+4];

dp4a.s32.s32 %r572, %r573, %r574, %r648;


dp4a.s32.s32 %r576, %r577, %r578, %r572;


prmt.b32 %r580, %r580, 0, 0xba98;


prmt.b32 %r582, %r582, 0, 0xba98;

xor.b32 %r1308, %r717, %r580;
or.b32 %r1309, %r1308, -2139062144;
and.b32 %r1310, %r580, 2139062143;
sub.s32 %r1311, %r1309, %r1310;
xor.b32 %r585, %r726, %r1311;
xor.b32 %r1312, %r718, %r582;
or.b32 %r1313, %r1312, -2139062144;
and.b32 %r1314, %r582, 2139062143;
sub.s32 %r1315, %r1313, %r1314;
xor.b32 %r589, %r731, %r1315;
ld.global.nc.u32 %r590, [%rd78+16];
ld.global.nc.u32 %r586, [%rd78+12];

dp4a.s32.s32 %r584, %r585, %r586, %r576;


dp4a.s32.s32 %r588, %r589, %r590, %r584;


prmt.b32 %r592, %r592, 0, 0xba98;


prmt.b32 %r594, %r594, 0, 0xba98;

xor.b32 %r1316, %r750, %r592;
or.b32 %r1317, %r1316, -2139062144;
and.b32 %r1318, %r592, 2139062143;
sub.s32 %r1319, %r1317, %r1318;
xor.b32 %r597, %r759, %r1319;
xor.b32 %r1320, %r751, %r594;
or.b32 %r1321, %r1320, -2139062144;
and.b32 %r1322, %r594, 2139062143;
sub.s32 %r1323, %r1321, %r1322;
xor.b32 %r601, %r764, %r1323;
ld.global.nc.u32 %r602, [%rd78+24];
ld.global.nc.u32 %r598, [%rd78+20];

dp4a.s32.s32 %r596, %r597, %r598, %r648;


dp4a.s32.s32 %r600, %r601, %r602, %r596;


prmt.b32 %r604, %r604, 0, 0xba98;


prmt.b32 %r606, %r606, 0, 0xba98;

xor.b32 %r1324, %r782, %r604;
or.b32 %r1325, %r1324, -2139062144;
and.b32 %r1326, %r604, 2139062143;
sub.s32 %r1327, %r1325, %r1326;
xor.b32 %r609, %r791, %r1327;
xor.b32 %r1328, %r783, %r606;
or.b32 %r1329, %r1328, -2139062144;
and.b32 %r1330, %r606, 2139062143;
sub.s32 %r1331, %r1329, %r1330;
xor.b32 %r613, %r796, %r1331;
ld.global.nc.u32 %r614, [%rd78+32];
ld.global.nc.u32 %r610, [%rd78+28];

dp4a.s32.s32 %r608, %r609, %r610, %r600;


dp4a.s32.s32 %r612, %r613, %r614, %r608;

mul.lo.s32 %r1332, %r588, %r798;
mad.lo.s32 %r1333, %r612, %r799, %r1332;
add.s32 %r1334, %r612, %r588;
shr.u32 %r1335, %r1334, 31;
add.s32 %r1336, %r1334, %r1335;
shr.s32 %r1337, %r1336, 1;
add.s32 %r1338, %r1333, %r1337;
shr.s32 %r1339, %r1338, 31;
shr.u32 %r1340, %r1339, 30;
add.s32 %r1341, %r1338, %r1340;
shr.s32 %r1342, %r1341, 2;

{ cvt.f32.f16 %f93, %rs1;}


ld.global.nc.u32 %r616, [%rd78];

{.reg .f16 low,high;
mov.b32 {low,high},%r616;
cvt.f32.f16 %f94, low;}


mul.ftz.f32 %f117, %f93, %f94;
cvt.rn.f32.s32 %f118, %r1342;
fma.rn.ftz.f32 %f266, %f117, %f118, %f266;

prmt.b32 %r617, %r617, 0, 0xba98;


prmt.b32 %r619, %r619, 0, 0xba98;

xor.b32 %r1343, %r830, %r617;
or.b32 %r1344, %r1343, -2139062144;
and.b32 %r1345, %r617, 2139062143;
sub.s32 %r1346, %r1344, %r1345;
xor.b32 %r622, %r839, %r1346;
xor.b32 %r1347, %r831, %r619;
or.b32 %r1348, %r1347, -2139062144;
and.b32 %r1349, %r619, 2139062143;
sub.s32 %r1350, %r1348, %r1349;
xor.b32 %r626, %r844, %r1350;

dp4a.s32.s32 %r621, %r622, %r574, %r648;


dp4a.s32.s32 %r625, %r626, %r578, %r621;


prmt.b32 %r629, %r629, 0, 0xba98;


prmt.b32 %r631, %r631, 0, 0xba98;

xor.b32 %r1351, %r862, %r629;
or.b32 %r1352, %r1351, -2139062144;
and.b32 %r1353, %r629, 2139062143;
sub.s32 %r1354, %r1352, %r1353;
xor.b32 %r634, %r871, %r1354;
xor.b32 %r1355, %r863, %r631;
or.b32 %r1356, %r1355, -2139062144;
and.b32 %r1357, %r631, 2139062143;
sub.s32 %r1358, %r1356, %r1357;
xor.b32 %r638, %r876, %r1358;

dp4a.s32.s32 %r633, %r634, %r586, %r625;


dp4a.s32.s32 %r637, %r638, %r590, %r633;


prmt.b32 %r641, %r641, 0, 0xba98;


prmt.b32 %r643, %r643, 0, 0xba98;

xor.b32 %r1359, %r895, %r641;
or.b32 %r1360, %r1359, -2139062144;
and.b32 %r1361, %r641, 2139062143;
sub.s32 %r1362, %r1360, %r1361;
xor.b32 %r646, %r904, %r1362;
xor.b32 %r1363, %r896, %r643;
or.b32 %r1364, %r1363, -2139062144;
and.b32 %r1365, %r643, 2139062143;
sub.s32 %r1366, %r1364, %r1365;
xor.b32 %r650, %r909, %r1366;

dp4a.s32.s32 %r645, %r646, %r598, %r648;


dp4a.s32.s32 %r649, %r650, %r602, %r645;


prmt.b32 %r653, %r653, 0, 0xba98;


prmt.b32 %r655, %r655, 0, 0xba98;

xor.b32 %r1367, %r927, %r653;
or.b32 %r1368, %r1367, -2139062144;
and.b32 %r1369, %r653, 2139062143;
sub.s32 %r1370, %r1368, %r1369;
xor.b32 %r658, %r936, %r1370;
xor.b32 %r1371, %r928, %r655;
or.b32 %r1372, %r1371, -2139062144;
and.b32 %r1373, %r655, 2139062143;
sub.s32 %r1374, %r1372, %r1373;
xor.b32 %r662, %r941, %r1374;

dp4a.s32.s32 %r657, %r658, %r610, %r649;


dp4a.s32.s32 %r661, %r662, %r614, %r657;

mul.lo.s32 %r1375, %r637, %r943;
mad.lo.s32 %r1376, %r661, %r944, %r1375;
add.s32 %r1377, %r661, %r637;
shr.u32 %r1378, %r1377, 31;
add.s32 %r1379, %r1377, %r1378;
shr.s32 %r1380, %r1379, 1;
add.s32 %r1381, %r1376, %r1380;
shr.s32 %r1382, %r1381, 31;
shr.u32 %r1383, %r1382, 30;
add.s32 %r1384, %r1381, %r1383;
shr.s32 %r1385, %r1384, 2;

{ cvt.f32.f16 %f95, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r616;
cvt.f32.f16 %f96, low;}


mul.ftz.f32 %f119, %f95, %f96;
cvt.rn.f32.s32 %f120, %r1385;
fma.rn.ftz.f32 %f265, %f119, %f120, %f265;
add.s32 %r1628, %r1628, 64;
add.s32 %r1627, %r1627, 64;
add.s32 %r1626, %r1626, 64;
add.s32 %r1625, %r1625, 64;
add.s32 %r1624, %r1624, 64;
add.s32 %r1623, %r1623, 64;
add.s64 %rd123, %rd123, 656;
add.s32 %r1629, %r1629, 8;
setp.lt.s32 %p4, %r1629, %r5;
@%p4 bra $L__BB234_2;

$L__BB234_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB234_5;

mov.u32 %r1386, _ZZ13mul_mat_vec_qIL9ggml_type22ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r1387, %r1, 1536, %r1386;
shl.b32 %r1388, %r2, 2;
add.s32 %r1389, %r1387, %r1388;
st.shared.f32 [%r1389+-1536], %f276;
st.shared.f32 [%r1389+-1408], %f275;
st.shared.f32 [%r1389+-1280], %f274;
st.shared.f32 [%r1389+-1152], %f273;
st.shared.f32 [%r1389+-1024], %f272;
st.shared.f32 [%r1389+-896], %f271;
st.shared.f32 [%r1389+-768], %f270;
st.shared.f32 [%r1389+-640], %f269;
st.shared.f32 [%r1389+-512], %f268;
st.shared.f32 [%r1389+-384], %f267;
st.shared.f32 [%r1389+-256], %f266;
st.shared.f32 [%r1389+-128], %f265;

$L__BB234_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB234_30;

mov.u32 %r1390, %ctaid.z;
mov.u32 %r1391, %ctaid.y;
mov.u32 %r1392, %ctaid.x;
shl.b32 %r1393, %r1392, 1;
mov.u32 %r1394, 1;
mad.lo.s32 %r1395, %r1391, %r37, %r1393;
mad.lo.s32 %r1396, %r1390, %r43, %r1395;
cvt.u64.u32 %rd13, %r1396;
shl.b32 %r1397, %r2, 2;
mov.u32 %r1398, _ZZ13mul_mat_vec_qIL9ggml_type22ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r26, %r1398, %r1397;
mov.u32 %r1399, 2;
ld.shared.f32 %f121, [%r26];
add.ftz.f32 %f122, %f121, %f276;
mov.b32 %r1400, %f122;
mov.u32 %r1401, 31;
mov.u32 %r1402, 16;
mov.u32 %r1403, -1;
shfl.sync.bfly.b32 %r1404|%p7, %r1400, %r1402, %r1401, %r1403;
mov.b32 %f123, %r1404;
add.ftz.f32 %f124, %f122, %f123;
mov.b32 %r1405, %f124;
mov.u32 %r1406, 8;
shfl.sync.bfly.b32 %r1407|%p8, %r1405, %r1406, %r1401, %r1403;
mov.b32 %f125, %r1407;
add.ftz.f32 %f126, %f124, %f125;
mov.b32 %r1408, %f126;
mov.u32 %r1409, 4;
shfl.sync.bfly.b32 %r1410|%p9, %r1408, %r1409, %r1401, %r1403;
mov.b32 %f127, %r1410;
add.ftz.f32 %f128, %f126, %f127;
mov.b32 %r1411, %f128;
shfl.sync.bfly.b32 %r1412|%p10, %r1411, %r1399, %r1401, %r1403;
mov.b32 %f129, %r1412;
add.ftz.f32 %f130, %f128, %f129;
mov.b32 %r1413, %f130;
shfl.sync.bfly.b32 %r1414|%p11, %r1413, %r1394, %r1401, %r1403;
mov.b32 %f131, %r1414;
add.ftz.f32 %f37, %f130, %f131;
setp.lt.u32 %p12, %r1393, %r31;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
not.pred %p14, %p1;
@%p14 bra $L__BB234_8;

cvta.to.global.u64 %rd79, %rd18;
mul.wide.u32 %rd80, %r1396, 4;
add.s64 %rd81, %rd79, %rd80;
st.global.f32 [%rd81], %f37;

$L__BB234_8:
ld.shared.f32 %f132, [%r26+128];
add.ftz.f32 %f133, %f132, %f275;
mov.b32 %r1421, %f133;
shfl.sync.bfly.b32 %r1425|%p15, %r1421, %r1402, %r1401, %r1403;
mov.b32 %f134, %r1425;
add.ftz.f32 %f135, %f133, %f134;
mov.b32 %r1426, %f135;
shfl.sync.bfly.b32 %r1428|%p16, %r1426, %r1406, %r1401, %r1403;
mov.b32 %f136, %r1428;
add.ftz.f32 %f137, %f135, %f136;
mov.b32 %r1429, %f137;
shfl.sync.bfly.b32 %r1431|%p17, %r1429, %r1409, %r1401, %r1403;
mov.b32 %f138, %r1431;
add.ftz.f32 %f139, %f137, %f138;
mov.b32 %r1432, %f139;
shfl.sync.bfly.b32 %r1434|%p18, %r1432, %r1399, %r1401, %r1403;
mov.b32 %f140, %r1434;
add.ftz.f32 %f141, %f139, %f140;
mov.b32 %r1435, %f141;
shfl.sync.bfly.b32 %r1437|%p19, %r1435, %r1394, %r1401, %r1403;
mov.b32 %f142, %r1437;
add.ftz.f32 %f38, %f141, %f142;
setp.eq.s32 %p20, %r2, 1;
or.b32 %r1441, %r1393, 1;
setp.lt.u32 %p21, %r1441, %r31;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB234_10;

cvta.to.global.u64 %rd82, %rd18;
mul.wide.u32 %rd83, %r1396, 4;
add.s64 %rd84, %rd82, %rd83;
st.global.f32 [%rd84+4], %f38;

$L__BB234_10:
ld.shared.f32 %f143, [%r26+256];
add.ftz.f32 %f144, %f143, %f274;
mov.b32 %r1448, %f144;
mov.u32 %r1449, 31;
mov.u32 %r1450, 16;
mov.u32 %r1451, -1;
shfl.sync.bfly.b32 %r1452|%p23, %r1448, %r1450, %r1449, %r1451;
mov.b32 %f145, %r1452;
add.ftz.f32 %f146, %f144, %f145;
mov.b32 %r1453, %f146;
mov.u32 %r1454, 8;
shfl.sync.bfly.b32 %r1455|%p24, %r1453, %r1454, %r1449, %r1451;
mov.b32 %f147, %r1455;
add.ftz.f32 %f148, %f146, %f147;
mov.b32 %r1456, %f148;
mov.u32 %r1457, 4;
shfl.sync.bfly.b32 %r1458|%p25, %r1456, %r1457, %r1449, %r1451;
mov.b32 %f149, %r1458;
add.ftz.f32 %f150, %f148, %f149;
mov.b32 %r1459, %f150;
mov.u32 %r1460, 2;
shfl.sync.bfly.b32 %r1461|%p26, %r1459, %r1460, %r1449, %r1451;
mov.b32 %f151, %r1461;
add.ftz.f32 %f152, %f150, %f151;
mov.b32 %r1462, %f152;
mov.u32 %r1463, 1;
shfl.sync.bfly.b32 %r1464|%p27, %r1462, %r1463, %r1449, %r1451;
mov.b32 %f153, %r1464;
add.ftz.f32 %f39, %f152, %f153;
@%p14 bra $L__BB234_12;

cvt.u64.u32 %rd85, %r31;
add.s64 %rd86, %rd85, %rd13;
cvta.to.global.u64 %rd87, %rd18;
shl.b64 %rd88, %rd86, 2;
add.s64 %rd89, %rd87, %rd88;
st.global.f32 [%rd89], %f39;

$L__BB234_12:
ld.shared.f32 %f154, [%r26+384];
add.ftz.f32 %f155, %f154, %f273;
mov.b32 %r1465, %f155;
shfl.sync.bfly.b32 %r1469|%p29, %r1465, %r1450, %r1449, %r1451;
mov.b32 %f156, %r1469;
add.ftz.f32 %f157, %f155, %f156;
mov.b32 %r1470, %f157;
shfl.sync.bfly.b32 %r1472|%p30, %r1470, %r1454, %r1449, %r1451;
mov.b32 %f158, %r1472;
add.ftz.f32 %f159, %f157, %f158;
mov.b32 %r1473, %f159;
shfl.sync.bfly.b32 %r1475|%p31, %r1473, %r1457, %r1449, %r1451;
mov.b32 %f160, %r1475;
add.ftz.f32 %f161, %f159, %f160;
mov.b32 %r1476, %f161;
shfl.sync.bfly.b32 %r1478|%p32, %r1476, %r1460, %r1449, %r1451;
mov.b32 %f162, %r1478;
add.ftz.f32 %f163, %f161, %f162;
mov.b32 %r1479, %f163;
shfl.sync.bfly.b32 %r1481|%p33, %r1479, %r1463, %r1449, %r1451;
mov.b32 %f164, %r1481;
add.ftz.f32 %f40, %f163, %f164;
@%p22 bra $L__BB234_14;

add.s32 %r1482, %r31, 1;
cvt.u64.u32 %rd90, %r1482;
add.s64 %rd91, %rd90, %rd13;
cvta.to.global.u64 %rd92, %rd18;
shl.b64 %rd93, %rd91, 2;
add.s64 %rd94, %rd92, %rd93;
st.global.f32 [%rd94], %f40;

$L__BB234_14:
shl.b32 %r1483, %r31, 1;
mov.u32 %r1484, 1;
ld.shared.f32 %f165, [%r26+512];
add.ftz.f32 %f166, %f165, %f272;
mov.b32 %r1485, %f166;
mov.u32 %r1486, 31;
mov.u32 %r1487, 16;
mov.u32 %r1488, -1;
shfl.sync.bfly.b32 %r1489|%p35, %r1485, %r1487, %r1486, %r1488;
mov.b32 %f167, %r1489;
add.ftz.f32 %f168, %f166, %f167;
mov.b32 %r1490, %f168;
mov.u32 %r1491, 8;
shfl.sync.bfly.b32 %r1492|%p36, %r1490, %r1491, %r1486, %r1488;
mov.b32 %f169, %r1492;
add.ftz.f32 %f170, %f168, %f169;
mov.b32 %r1493, %f170;
mov.u32 %r1494, 4;
shfl.sync.bfly.b32 %r1495|%p37, %r1493, %r1494, %r1486, %r1488;
mov.b32 %f171, %r1495;
add.ftz.f32 %f172, %f170, %f171;
mov.b32 %r1496, %f172;
mov.u32 %r1497, 2;
shfl.sync.bfly.b32 %r1498|%p38, %r1496, %r1497, %r1486, %r1488;
mov.b32 %f173, %r1498;
add.ftz.f32 %f174, %f172, %f173;
mov.b32 %r1499, %f174;
shfl.sync.bfly.b32 %r1500|%p39, %r1499, %r1484, %r1486, %r1488;
mov.b32 %f175, %r1500;
add.ftz.f32 %f41, %f174, %f175;
cvt.u64.u32 %rd95, %r1483;
add.s64 %rd96, %rd95, %rd13;
cvta.to.global.u64 %rd97, %rd18;
shl.b64 %rd98, %rd96, 2;
add.s64 %rd14, %rd97, %rd98;
@%p14 bra $L__BB234_16;

st.global.f32 [%rd14], %f41;

$L__BB234_16:
ld.shared.f32 %f176, [%r26+640];
add.ftz.f32 %f177, %f176, %f271;
mov.b32 %r1501, %f177;
shfl.sync.bfly.b32 %r1505|%p41, %r1501, %r1487, %r1486, %r1488;
mov.b32 %f178, %r1505;
add.ftz.f32 %f179, %f177, %f178;
mov.b32 %r1506, %f179;
shfl.sync.bfly.b32 %r1508|%p42, %r1506, %r1491, %r1486, %r1488;
mov.b32 %f180, %r1508;
add.ftz.f32 %f181, %f179, %f180;
mov.b32 %r1509, %f181;
shfl.sync.bfly.b32 %r1511|%p43, %r1509, %r1494, %r1486, %r1488;
mov.b32 %f182, %r1511;
add.ftz.f32 %f183, %f181, %f182;
mov.b32 %r1512, %f183;
shfl.sync.bfly.b32 %r1514|%p44, %r1512, %r1497, %r1486, %r1488;
mov.b32 %f184, %r1514;
add.ftz.f32 %f185, %f183, %f184;
mov.b32 %r1515, %f185;
shfl.sync.bfly.b32 %r1517|%p45, %r1515, %r1484, %r1486, %r1488;
mov.b32 %f186, %r1517;
add.ftz.f32 %f42, %f185, %f186;
@%p22 bra $L__BB234_18;

st.global.f32 [%rd14+4], %f42;

$L__BB234_18:
mul.lo.s32 %r27, %r31, 3;
ld.shared.f32 %f187, [%r26+768];
add.ftz.f32 %f188, %f187, %f270;
mov.b32 %r1518, %f188;
mov.u32 %r1519, 31;
mov.u32 %r1520, 16;
mov.u32 %r1521, -1;
shfl.sync.bfly.b32 %r1522|%p47, %r1518, %r1520, %r1519, %r1521;
mov.b32 %f189, %r1522;
add.ftz.f32 %f190, %f188, %f189;
mov.b32 %r1523, %f190;
mov.u32 %r1524, 8;
shfl.sync.bfly.b32 %r1525|%p48, %r1523, %r1524, %r1519, %r1521;
mov.b32 %f191, %r1525;
add.ftz.f32 %f192, %f190, %f191;
mov.b32 %r1526, %f192;
mov.u32 %r1527, 4;
shfl.sync.bfly.b32 %r1528|%p49, %r1526, %r1527, %r1519, %r1521;
mov.b32 %f193, %r1528;
add.ftz.f32 %f194, %f192, %f193;
mov.b32 %r1529, %f194;
mov.u32 %r1530, 2;
shfl.sync.bfly.b32 %r1531|%p50, %r1529, %r1530, %r1519, %r1521;
mov.b32 %f195, %r1531;
add.ftz.f32 %f196, %f194, %f195;
mov.b32 %r1532, %f196;
mov.u32 %r1533, 1;
shfl.sync.bfly.b32 %r1534|%p51, %r1532, %r1533, %r1519, %r1521;
mov.b32 %f197, %r1534;
add.ftz.f32 %f43, %f196, %f197;
@%p14 bra $L__BB234_20;

cvt.u64.u32 %rd99, %r27;
add.s64 %rd100, %rd99, %rd13;
shl.b64 %rd102, %rd100, 2;
add.s64 %rd103, %rd97, %rd102;
st.global.f32 [%rd103], %f43;

$L__BB234_20:
ld.shared.f32 %f198, [%r26+896];
add.ftz.f32 %f199, %f198, %f269;
mov.b32 %r1535, %f199;
shfl.sync.bfly.b32 %r1539|%p53, %r1535, %r1520, %r1519, %r1521;
mov.b32 %f200, %r1539;
add.ftz.f32 %f201, %f199, %f200;
mov.b32 %r1540, %f201;
shfl.sync.bfly.b32 %r1542|%p54, %r1540, %r1524, %r1519, %r1521;
mov.b32 %f202, %r1542;
add.ftz.f32 %f203, %f201, %f202;
mov.b32 %r1543, %f203;
shfl.sync.bfly.b32 %r1545|%p55, %r1543, %r1527, %r1519, %r1521;
mov.b32 %f204, %r1545;
add.ftz.f32 %f205, %f203, %f204;
mov.b32 %r1546, %f205;
shfl.sync.bfly.b32 %r1548|%p56, %r1546, %r1530, %r1519, %r1521;
mov.b32 %f206, %r1548;
add.ftz.f32 %f207, %f205, %f206;
mov.b32 %r1549, %f207;
shfl.sync.bfly.b32 %r1551|%p57, %r1549, %r1533, %r1519, %r1521;
mov.b32 %f208, %r1551;
add.ftz.f32 %f44, %f207, %f208;
@%p22 bra $L__BB234_22;

add.s32 %r1552, %r27, 1;
cvt.u64.u32 %rd104, %r1552;
add.s64 %rd105, %rd104, %rd13;
shl.b64 %rd107, %rd105, 2;
add.s64 %rd108, %rd97, %rd107;
st.global.f32 [%rd108], %f44;

$L__BB234_22:
shl.b32 %r1553, %r31, 2;
mov.u32 %r1554, 2;
ld.shared.f32 %f209, [%r26+1024];
add.ftz.f32 %f210, %f209, %f268;
mov.b32 %r1555, %f210;
mov.u32 %r1556, 31;
mov.u32 %r1557, 16;
mov.u32 %r1558, -1;
shfl.sync.bfly.b32 %r1559|%p59, %r1555, %r1557, %r1556, %r1558;
mov.b32 %f211, %r1559;
add.ftz.f32 %f212, %f210, %f211;
mov.b32 %r1560, %f212;
mov.u32 %r1561, 8;
shfl.sync.bfly.b32 %r1562|%p60, %r1560, %r1561, %r1556, %r1558;
mov.b32 %f213, %r1562;
add.ftz.f32 %f214, %f212, %f213;
mov.b32 %r1563, %f214;
mov.u32 %r1564, 4;
shfl.sync.bfly.b32 %r1565|%p61, %r1563, %r1564, %r1556, %r1558;
mov.b32 %f215, %r1565;
add.ftz.f32 %f216, %f214, %f215;
mov.b32 %r1566, %f216;
shfl.sync.bfly.b32 %r1567|%p62, %r1566, %r1554, %r1556, %r1558;
mov.b32 %f217, %r1567;
add.ftz.f32 %f218, %f216, %f217;
mov.b32 %r1568, %f218;
mov.u32 %r1569, 1;
shfl.sync.bfly.b32 %r1570|%p63, %r1568, %r1569, %r1556, %r1558;
mov.b32 %f219, %r1570;
add.ftz.f32 %f45, %f218, %f219;
cvt.u64.u32 %rd109, %r1553;
add.s64 %rd110, %rd109, %rd13;
shl.b64 %rd112, %rd110, 2;
add.s64 %rd15, %rd97, %rd112;
@%p14 bra $L__BB234_24;

st.global.f32 [%rd15], %f45;

$L__BB234_24:
ld.shared.f32 %f220, [%r26+1152];
add.ftz.f32 %f221, %f220, %f267;
mov.b32 %r1571, %f221;
shfl.sync.bfly.b32 %r1575|%p65, %r1571, %r1557, %r1556, %r1558;
mov.b32 %f222, %r1575;
add.ftz.f32 %f223, %f221, %f222;
mov.b32 %r1576, %f223;
shfl.sync.bfly.b32 %r1578|%p66, %r1576, %r1561, %r1556, %r1558;
mov.b32 %f224, %r1578;
add.ftz.f32 %f225, %f223, %f224;
mov.b32 %r1579, %f225;
shfl.sync.bfly.b32 %r1581|%p67, %r1579, %r1564, %r1556, %r1558;
mov.b32 %f226, %r1581;
add.ftz.f32 %f227, %f225, %f226;
mov.b32 %r1582, %f227;
shfl.sync.bfly.b32 %r1584|%p68, %r1582, %r1554, %r1556, %r1558;
mov.b32 %f228, %r1584;
add.ftz.f32 %f229, %f227, %f228;
mov.b32 %r1585, %f229;
shfl.sync.bfly.b32 %r1587|%p69, %r1585, %r1569, %r1556, %r1558;
mov.b32 %f230, %r1587;
add.ftz.f32 %f46, %f229, %f230;
@%p22 bra $L__BB234_26;

st.global.f32 [%rd15+4], %f46;

$L__BB234_26:
mul.lo.s32 %r28, %r31, 5;
ld.shared.f32 %f231, [%r26+1280];
add.ftz.f32 %f232, %f231, %f266;
mov.b32 %r1588, %f232;
mov.u32 %r1589, 31;
mov.u32 %r1590, 16;
mov.u32 %r1591, -1;
shfl.sync.bfly.b32 %r1592|%p71, %r1588, %r1590, %r1589, %r1591;
mov.b32 %f233, %r1592;
add.ftz.f32 %f234, %f232, %f233;
mov.b32 %r1593, %f234;
mov.u32 %r1594, 8;
shfl.sync.bfly.b32 %r1595|%p72, %r1593, %r1594, %r1589, %r1591;
mov.b32 %f235, %r1595;
add.ftz.f32 %f236, %f234, %f235;
mov.b32 %r1596, %f236;
mov.u32 %r1597, 4;
shfl.sync.bfly.b32 %r1598|%p73, %r1596, %r1597, %r1589, %r1591;
mov.b32 %f237, %r1598;
add.ftz.f32 %f238, %f236, %f237;
mov.b32 %r1599, %f238;
mov.u32 %r1600, 2;
shfl.sync.bfly.b32 %r1601|%p74, %r1599, %r1600, %r1589, %r1591;
mov.b32 %f239, %r1601;
add.ftz.f32 %f240, %f238, %f239;
mov.b32 %r1602, %f240;
mov.u32 %r1603, 1;
shfl.sync.bfly.b32 %r1604|%p75, %r1602, %r1603, %r1589, %r1591;
mov.b32 %f241, %r1604;
add.ftz.f32 %f47, %f240, %f241;
@%p14 bra $L__BB234_28;

cvt.u64.u32 %rd113, %r28;
add.s64 %rd114, %rd113, %rd13;
shl.b64 %rd116, %rd114, 2;
add.s64 %rd117, %rd97, %rd116;
st.global.f32 [%rd117], %f47;

$L__BB234_28:
ld.shared.f32 %f242, [%r26+1408];
add.ftz.f32 %f243, %f242, %f265;
mov.b32 %r1605, %f243;
shfl.sync.bfly.b32 %r1609|%p77, %r1605, %r1590, %r1589, %r1591;
mov.b32 %f244, %r1609;
add.ftz.f32 %f245, %f243, %f244;
mov.b32 %r1610, %f245;
shfl.sync.bfly.b32 %r1612|%p78, %r1610, %r1594, %r1589, %r1591;
mov.b32 %f246, %r1612;
add.ftz.f32 %f247, %f245, %f246;
mov.b32 %r1613, %f247;
shfl.sync.bfly.b32 %r1615|%p79, %r1613, %r1597, %r1589, %r1591;
mov.b32 %f248, %r1615;
add.ftz.f32 %f249, %f247, %f248;
mov.b32 %r1616, %f249;
shfl.sync.bfly.b32 %r1618|%p80, %r1616, %r1600, %r1589, %r1591;
mov.b32 %f250, %r1618;
add.ftz.f32 %f251, %f249, %f250;
mov.b32 %r1619, %f251;
shfl.sync.bfly.b32 %r1621|%p81, %r1619, %r1603, %r1589, %r1591;
mov.b32 %f252, %r1621;
add.ftz.f32 %f48, %f251, %f252;
@%p22 bra $L__BB234_30;

add.s32 %r1622, %r28, 1;
cvt.u64.u32 %rd118, %r1622;
add.s64 %rd119, %rd118, %rd13;
shl.b64 %rd121, %rd119, 2;
add.s64 %rd122, %rd97, %rd121;
st.global.f32 [%rd122], %f48;

$L__BB234_30:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type22ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type22ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type22ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type22ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type22ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type22ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type22ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type22ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type22ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<106>;
.reg .b16 %rs<53>;
.reg .f32 %f<323>;
.reg .b32 %r<1921>;
.reg .b64 %rd<134>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type22ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1792];

ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type22ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type22ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type22ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type22ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type22ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type22ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type22ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type22ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type22ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type22ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type22ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type22ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type22ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type22ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type22ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type22ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type22ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r45, %tid.y;
shl.b32 %r46, %r45, 5;
mov.u32 %r47, %tid.x;
add.s32 %r1, %r46, %r47;
shr.s32 %r48, %r1, 31;
shr.u32 %r49, %r48, 29;
add.s32 %r50, %r1, %r49;
shr.s32 %r1920, %r50, 3;
shr.u32 %r3, %r44, 8;
setp.ge.s32 %p2, %r1920, %r3;
mov.f32 %f309, 0f00000000;
mov.f32 %f310, %f309;
mov.f32 %f311, %f309;
mov.f32 %f312, %f309;
mov.f32 %f313, %f309;
mov.f32 %f314, %f309;
mov.f32 %f315, %f309;
mov.f32 %f316, %f309;
mov.f32 %f317, %f309;
mov.f32 %f318, %f309;
mov.f32 %f319, %f309;
mov.f32 %f320, %f309;
mov.f32 %f321, %f309;
mov.f32 %f322, %f309;
@%p2 bra $L__BB235_3;

mov.u32 %r51, %ctaid.x;
shl.b32 %r52, %r51, 1;
mov.u32 %r53, %ctaid.y;
mov.u32 %r54, %ctaid.z;
mul.lo.s32 %r55, %r53, %r36;
cvt.u64.u32 %rd20, %r55;
mul.lo.s32 %r56, %r54, %r42;
cvt.u64.u32 %rd21, %r56;
add.s64 %rd1, %rd21, %rd20;
mul.hi.u32 %r57, %r54, %r38;
add.s32 %r58, %r57, %r54;
shr.u32 %r59, %r58, %r39;
mul.lo.s32 %r60, %r59, %r41;
mul.hi.u32 %r61, %r53, %r32;
add.s32 %r62, %r61, %r53;
shr.u32 %r63, %r62, %r33;
mul.lo.s32 %r64, %r63, %r35;
cvta.to.global.u64 %rd133, %rd17;
and.b32 %r68, %r50, -8;
sub.s32 %r69, %r1, %r68;
shl.b32 %r70, %r69, 1;
cvt.s64.s32 %rd3, %r69;
shl.b32 %r71, %r30, 1;
shl.b32 %r1913, %r1920, 3;
add.s32 %r1919, %r1913, %r71;
mad.lo.s32 %r1918, %r30, 3, %r1913;
shl.b32 %r72, %r30, 2;
add.s32 %r1917, %r1913, %r72;
mad.lo.s32 %r1916, %r30, 5, %r1913;
mad.lo.s32 %r1915, %r30, 6, %r1913;
add.s32 %r1914, %r30, %r1913;
add.s32 %r73, %r1920, %r60;
add.s32 %r74, %r73, %r64;
or.b32 %r75, %r52, 1;
mad.lo.s32 %r76, %r29, %r75, %r74;
mul.wide.s32 %rd4, %r76, 82;
mad.lo.s32 %r77, %r52, %r29, %r1920;
add.s32 %r78, %r77, %r60;
add.s32 %r79, %r78, %r64;
mul.wide.s32 %rd5, %r79, 82;
mul.wide.s32 %rd22, %r70, 2;
add.s64 %rd23, %rd5, %rd22;
add.s64 %rd6, %rd23, 36;
add.s64 %rd24, %rd4, %rd3;
add.s64 %rd7, %rd24, 66;
add.s64 %rd25, %rd5, %rd3;
add.s64 %rd8, %rd25, 74;
add.s64 %rd26, %rd4, %rd22;
add.s64 %rd9, %rd26, 34;
cvta.to.global.u64 %rd10, %rd18;
mov.f32 %f309, 0f00000000;

$L__BB235_2:
cvt.u64.u32 %rd27, %r1913;
add.s64 %rd28, %rd1, %rd27;
add.s64 %rd29, %rd28, %rd3;
add.s64 %rd30, %rd133, %rd6;
ld.global.nc.u16 %rs15, [%rd30+-34];
shr.u16 %rs16, %rs15, 8;
cvt.u32.u16 %r766, %rs16;
ld.global.nc.u16 %rs17, [%rd30+-32];
shr.u16 %rs18, %rs17, 8;
cvt.u32.u16 %r767, %rs18;
add.s64 %rd31, %rd133, %rd8;
ld.global.nc.u8 %rs19, [%rd31+-8];
ld.global.nc.u16 %rs20, [%rd30+-2];
ld.global.nc.u16 %rs21, [%rd30];
ld.global.nc.u8 %rs22, [%rd31];
and.b16 %rs23, %rs22, 240;
and.b16 %rs24, %rs22, 15;
shr.u16 %rs25, %rs23, 4;
and.b16 %rs26, %rs15, 255;
cvt.u32.u16 %r768, %rs26;
shl.b16 %rs27, %rs19, 8;
cvt.u32.u16 %r769, %rs27;
and.b32 %r770, %r769, 768;
or.b32 %r771, %r770, %r768;
mul.wide.u32 %rd32, %r771, 8;
mov.u64 %rd33, iq2s_grid;
add.s64 %rd34, %rd33, %rd32;
and.b16 %rs28, %rs20, 255;
mul.wide.u16 %r772, %rs28, 128;
cvt.u32.u16 %r773, %rs20;
shl.b32 %r774, %r773, 21;
or.b32 %r775, %r772, %r774;
and.b32 %r776, %r775, 25166208;
or.b32 %r777, %r776, -2139062144;
add.s32 %r778, %r777, -16843009;
or.b32 %r668, %r778, %r776;
mov.u32 %r80, %r668;

prmt.b32 %r80, %r80, 0, 0xba98;

mul.wide.u16 %r779, %rs28, 8;
shl.b32 %r780, %r773, 17;
or.b32 %r781, %r779, %r780;
and.b32 %r782, %r781, 25166208;
or.b32 %r783, %r782, -2139062144;
add.s32 %r784, %r783, -16843009;
or.b32 %r670, %r784, %r782;
mov.u32 %r82, %r670;

prmt.b32 %r82, %r82, 0, 0xba98;

ld.global.nc.v2.u32 {%r785, %r786}, [%rd34];
not.b32 %r788, %r785;
mov.u32 %r748, 0;
not.b32 %r790, %r786;
xor.b32 %r791, %r785, %r80;
or.b32 %r792, %r791, -2139062144;
and.b32 %r793, %r80, 2139062143;
and.b32 %r794, %r788, -2139062144;
sub.s32 %r795, %r792, %r793;
xor.b32 %r85, %r794, %r795;
xor.b32 %r796, %r786, %r82;
or.b32 %r797, %r796, -2139062144;
and.b32 %r798, %r82, 2139062143;
and.b32 %r799, %r790, -2139062144;
sub.s32 %r800, %r797, %r798;
xor.b32 %r89, %r799, %r800;
mul.lo.s64 %rd35, %rd29, 36;
add.s64 %rd36, %rd10, %rd35;
ld.global.nc.u32 %r90, [%rd36+8];
ld.global.nc.u32 %r86, [%rd36+4];

dp4a.s32.s32 %r84, %r85, %r86, %r748;


dp4a.s32.s32 %r88, %r89, %r90, %r84;

mul.wide.u16 %r801, %rs19, 64;
and.b32 %r802, %r801, 768;
or.b32 %r803, %r802, %r766;
mul.wide.u32 %rd37, %r803, 8;
add.s64 %rd38, %rd33, %rd37;
shr.u16 %rs30, %rs20, 8;
mul.wide.u16 %r804, %rs30, 128;
cvt.u32.u16 %r805, %rs30;
shl.b32 %r806, %r805, 21;
or.b32 %r807, %r804, %r806;
and.b32 %r808, %r807, 25166208;
or.b32 %r809, %r808, -2139062144;
add.s32 %r810, %r809, -16843009;
or.b32 %r680, %r810, %r808;
mov.u32 %r92, %r680;

prmt.b32 %r92, %r92, 0, 0xba98;

mul.wide.u16 %r811, %rs30, 8;
shl.b32 %r812, %r805, 17;
or.b32 %r813, %r811, %r812;
and.b32 %r814, %r813, 25166208;
or.b32 %r815, %r814, -2139062144;
add.s32 %r816, %r815, -16843009;
or.b32 %r682, %r816, %r814;
mov.u32 %r94, %r682;

prmt.b32 %r94, %r94, 0, 0xba98;

ld.global.nc.v2.u32 {%r817, %r818}, [%rd38];
not.b32 %r820, %r817;
not.b32 %r822, %r818;
xor.b32 %r823, %r817, %r92;
or.b32 %r824, %r823, -2139062144;
and.b32 %r825, %r92, 2139062143;
and.b32 %r826, %r820, -2139062144;
sub.s32 %r827, %r824, %r825;
xor.b32 %r97, %r826, %r827;
xor.b32 %r828, %r818, %r94;
or.b32 %r829, %r828, -2139062144;
and.b32 %r830, %r94, 2139062143;
and.b32 %r831, %r822, -2139062144;
sub.s32 %r832, %r829, %r830;
xor.b32 %r101, %r831, %r832;
ld.global.nc.u32 %r102, [%rd36+16];
ld.global.nc.u32 %r98, [%rd36+12];

dp4a.s32.s32 %r96, %r97, %r98, %r88;


dp4a.s32.s32 %r100, %r101, %r102, %r96;

and.b16 %rs31, %rs17, 255;
cvt.u32.u16 %r833, %rs31;
mul.wide.u16 %r834, %rs19, 16;
and.b32 %r835, %r834, 768;
or.b32 %r836, %r835, %r833;
mul.wide.u32 %rd39, %r836, 8;
add.s64 %rd40, %rd33, %rd39;
and.b16 %rs32, %rs21, 255;
mul.wide.u16 %r837, %rs32, 128;
cvt.u32.u16 %r838, %rs21;
shl.b32 %r839, %r838, 21;
or.b32 %r840, %r837, %r839;
and.b32 %r841, %r840, 25166208;
or.b32 %r842, %r841, -2139062144;
add.s32 %r843, %r842, -16843009;
or.b32 %r692, %r843, %r841;
mov.u32 %r104, %r692;

prmt.b32 %r104, %r104, 0, 0xba98;

mul.wide.u16 %r844, %rs32, 8;
shl.b32 %r845, %r838, 17;
or.b32 %r846, %r844, %r845;
and.b32 %r847, %r846, 25166208;
or.b32 %r848, %r847, -2139062144;
add.s32 %r849, %r848, -16843009;
or.b32 %r694, %r849, %r847;
mov.u32 %r106, %r694;

prmt.b32 %r106, %r106, 0, 0xba98;

ld.global.nc.v2.u32 {%r850, %r851}, [%rd40];
not.b32 %r853, %r850;
not.b32 %r855, %r851;
xor.b32 %r856, %r850, %r104;
or.b32 %r857, %r856, -2139062144;
and.b32 %r858, %r104, 2139062143;
and.b32 %r859, %r853, -2139062144;
sub.s32 %r860, %r857, %r858;
xor.b32 %r109, %r859, %r860;
xor.b32 %r861, %r851, %r106;
or.b32 %r862, %r861, -2139062144;
and.b32 %r863, %r106, 2139062143;
and.b32 %r864, %r855, -2139062144;
sub.s32 %r865, %r862, %r863;
xor.b32 %r113, %r864, %r865;
ld.global.nc.u32 %r114, [%rd36+24];
ld.global.nc.u32 %r110, [%rd36+20];

dp4a.s32.s32 %r108, %r109, %r110, %r748;


dp4a.s32.s32 %r112, %r113, %r114, %r108;

mul.wide.u16 %r866, %rs19, 4;
and.b32 %r867, %r866, 768;
or.b32 %r868, %r867, %r767;
mul.wide.u32 %rd41, %r868, 8;
add.s64 %rd42, %rd33, %rd41;
shr.u16 %rs33, %rs21, 8;
mul.wide.u16 %r869, %rs33, 128;
cvt.u32.u16 %r870, %rs33;
shl.b32 %r871, %r870, 21;
or.b32 %r872, %r869, %r871;
and.b32 %r873, %r872, 25166208;
or.b32 %r874, %r873, -2139062144;
add.s32 %r875, %r874, -16843009;
or.b32 %r704, %r875, %r873;
mov.u32 %r116, %r704;

prmt.b32 %r116, %r116, 0, 0xba98;

mul.wide.u16 %r876, %rs33, 8;
shl.b32 %r877, %r870, 17;
or.b32 %r878, %r876, %r877;
and.b32 %r879, %r878, 25166208;
or.b32 %r880, %r879, -2139062144;
add.s32 %r881, %r880, -16843009;
or.b32 %r706, %r881, %r879;
mov.u32 %r118, %r706;

prmt.b32 %r118, %r118, 0, 0xba98;

ld.global.nc.v2.u32 {%r882, %r883}, [%rd42];
not.b32 %r885, %r882;
not.b32 %r887, %r883;
xor.b32 %r888, %r882, %r116;
or.b32 %r889, %r888, -2139062144;
and.b32 %r890, %r116, 2139062143;
and.b32 %r891, %r885, -2139062144;
sub.s32 %r892, %r889, %r890;
xor.b32 %r121, %r891, %r892;
xor.b32 %r893, %r883, %r118;
or.b32 %r894, %r893, -2139062144;
and.b32 %r895, %r118, 2139062143;
and.b32 %r896, %r887, -2139062144;
sub.s32 %r897, %r894, %r895;
xor.b32 %r125, %r896, %r897;
ld.global.nc.u32 %r126, [%rd36+32];
ld.global.nc.u32 %r122, [%rd36+28];

dp4a.s32.s32 %r120, %r121, %r122, %r112;


dp4a.s32.s32 %r124, %r125, %r126, %r120;

cvt.u32.u16 %r898, %rs24;
cvt.u32.u16 %r899, %rs25;
mul.lo.s32 %r900, %r100, %r898;
mad.lo.s32 %r901, %r124, %r899, %r900;
add.s32 %r902, %r124, %r100;
shr.u32 %r903, %r902, 31;
add.s32 %r904, %r902, %r903;
shr.s32 %r905, %r904, 1;
add.s32 %r906, %r901, %r905;
shr.s32 %r907, %r906, 31;
shr.u32 %r908, %r907, 30;
add.s32 %r909, %r906, %r908;
shr.s32 %r910, %r909, 2;
add.s64 %rd43, %rd133, %rd5;
ld.global.nc.u16 %rs1, [%rd43];

{ cvt.f32.f16 %f85, %rs1;}


ld.global.nc.u32 %r128, [%rd36];

{.reg .f16 low,high;
mov.b32 {low,high},%r128;
cvt.f32.f16 %f86, low;}


mul.ftz.f32 %f113, %f85, %f86;
cvt.rn.f32.s32 %f114, %r910;
fma.rn.ftz.f32 %f322, %f113, %f114, %f322;
add.s64 %rd44, %rd133, %rd9;
ld.global.nc.u16 %rs34, [%rd44+-32];
shr.u16 %rs35, %rs34, 8;
cvt.u32.u16 %r911, %rs35;
ld.global.nc.u16 %rs36, [%rd44+-30];
shr.u16 %rs37, %rs36, 8;
cvt.u32.u16 %r912, %rs37;
add.s64 %rd45, %rd133, %rd7;
ld.global.nc.u8 %rs38, [%rd45];
ld.global.nc.u16 %rs39, [%rd44];
ld.global.nc.u16 %rs40, [%rd44+2];
ld.global.nc.u8 %rs41, [%rd45+8];
and.b16 %rs42, %rs41, 240;
and.b16 %rs43, %rs41, 15;
shr.u16 %rs44, %rs42, 4;
and.b16 %rs45, %rs34, 255;
cvt.u32.u16 %r913, %rs45;
shl.b16 %rs46, %rs38, 8;
cvt.u32.u16 %r914, %rs46;
and.b32 %r915, %r914, 768;
or.b32 %r916, %r915, %r913;
mul.wide.u32 %rd46, %r916, 8;
add.s64 %rd47, %rd33, %rd46;
and.b16 %rs47, %rs39, 255;
mul.wide.u16 %r917, %rs47, 128;
cvt.u32.u16 %r918, %rs39;
shl.b32 %r919, %r918, 21;
or.b32 %r920, %r917, %r919;
and.b32 %r921, %r920, 25166208;
or.b32 %r922, %r921, -2139062144;
add.s32 %r923, %r922, -16843009;
or.b32 %r717, %r923, %r921;
mov.u32 %r129, %r717;

prmt.b32 %r129, %r129, 0, 0xba98;

mul.wide.u16 %r924, %rs47, 8;
shl.b32 %r925, %r918, 17;
or.b32 %r926, %r924, %r925;
and.b32 %r927, %r926, 25166208;
or.b32 %r928, %r927, -2139062144;
add.s32 %r929, %r928, -16843009;
or.b32 %r719, %r929, %r927;
mov.u32 %r131, %r719;

prmt.b32 %r131, %r131, 0, 0xba98;

ld.global.nc.v2.u32 {%r930, %r931}, [%rd47];
not.b32 %r933, %r930;
not.b32 %r935, %r931;
xor.b32 %r936, %r930, %r129;
or.b32 %r937, %r936, -2139062144;
and.b32 %r938, %r129, 2139062143;
and.b32 %r939, %r933, -2139062144;
sub.s32 %r940, %r937, %r938;
xor.b32 %r134, %r939, %r940;
xor.b32 %r941, %r931, %r131;
or.b32 %r942, %r941, -2139062144;
and.b32 %r943, %r131, 2139062143;
and.b32 %r944, %r935, -2139062144;
sub.s32 %r945, %r942, %r943;
xor.b32 %r138, %r944, %r945;

dp4a.s32.s32 %r133, %r134, %r86, %r748;


dp4a.s32.s32 %r137, %r138, %r90, %r133;

mul.wide.u16 %r946, %rs38, 64;
and.b32 %r947, %r946, 768;
or.b32 %r948, %r947, %r911;
mul.wide.u32 %rd48, %r948, 8;
add.s64 %rd49, %rd33, %rd48;
shr.u16 %rs49, %rs39, 8;
mul.wide.u16 %r949, %rs49, 128;
cvt.u32.u16 %r950, %rs49;
shl.b32 %r951, %r950, 21;
or.b32 %r952, %r949, %r951;
and.b32 %r953, %r952, 25166208;
or.b32 %r954, %r953, -2139062144;
add.s32 %r955, %r954, -16843009;
or.b32 %r729, %r955, %r953;
mov.u32 %r141, %r729;

prmt.b32 %r141, %r141, 0, 0xba98;

mul.wide.u16 %r956, %rs49, 8;
shl.b32 %r957, %r950, 17;
or.b32 %r958, %r956, %r957;
and.b32 %r959, %r958, 25166208;
or.b32 %r960, %r959, -2139062144;
add.s32 %r961, %r960, -16843009;
or.b32 %r731, %r961, %r959;
mov.u32 %r143, %r731;

prmt.b32 %r143, %r143, 0, 0xba98;

ld.global.nc.v2.u32 {%r962, %r963}, [%rd49];
not.b32 %r965, %r962;
not.b32 %r967, %r963;
xor.b32 %r968, %r962, %r141;
or.b32 %r969, %r968, -2139062144;
and.b32 %r970, %r141, 2139062143;
and.b32 %r971, %r965, -2139062144;
sub.s32 %r972, %r969, %r970;
xor.b32 %r146, %r971, %r972;
xor.b32 %r973, %r963, %r143;
or.b32 %r974, %r973, -2139062144;
and.b32 %r975, %r143, 2139062143;
and.b32 %r976, %r967, -2139062144;
sub.s32 %r977, %r974, %r975;
xor.b32 %r150, %r976, %r977;

dp4a.s32.s32 %r145, %r146, %r98, %r137;


dp4a.s32.s32 %r149, %r150, %r102, %r145;

and.b16 %rs50, %rs36, 255;
cvt.u32.u16 %r978, %rs50;
mul.wide.u16 %r979, %rs38, 16;
and.b32 %r980, %r979, 768;
or.b32 %r981, %r980, %r978;
mul.wide.u32 %rd50, %r981, 8;
add.s64 %rd51, %rd33, %rd50;
and.b16 %rs51, %rs40, 255;
mul.wide.u16 %r982, %rs51, 128;
cvt.u32.u16 %r983, %rs40;
shl.b32 %r984, %r983, 21;
or.b32 %r985, %r982, %r984;
and.b32 %r986, %r985, 25166208;
or.b32 %r987, %r986, -2139062144;
add.s32 %r988, %r987, -16843009;
or.b32 %r741, %r988, %r986;
mov.u32 %r153, %r741;

prmt.b32 %r153, %r153, 0, 0xba98;

mul.wide.u16 %r989, %rs51, 8;
shl.b32 %r990, %r983, 17;
or.b32 %r991, %r989, %r990;
and.b32 %r992, %r991, 25166208;
or.b32 %r993, %r992, -2139062144;
add.s32 %r994, %r993, -16843009;
or.b32 %r743, %r994, %r992;
mov.u32 %r155, %r743;

prmt.b32 %r155, %r155, 0, 0xba98;

ld.global.nc.v2.u32 {%r995, %r996}, [%rd51];
not.b32 %r998, %r995;
not.b32 %r1000, %r996;
xor.b32 %r1001, %r995, %r153;
or.b32 %r1002, %r1001, -2139062144;
and.b32 %r1003, %r153, 2139062143;
and.b32 %r1004, %r998, -2139062144;
sub.s32 %r1005, %r1002, %r1003;
xor.b32 %r158, %r1004, %r1005;
xor.b32 %r1006, %r996, %r155;
or.b32 %r1007, %r1006, -2139062144;
and.b32 %r1008, %r155, 2139062143;
and.b32 %r1009, %r1000, -2139062144;
sub.s32 %r1010, %r1007, %r1008;
xor.b32 %r162, %r1009, %r1010;

dp4a.s32.s32 %r157, %r158, %r110, %r748;


dp4a.s32.s32 %r161, %r162, %r114, %r157;

mul.wide.u16 %r1011, %rs38, 4;
and.b32 %r1012, %r1011, 768;
or.b32 %r1013, %r1012, %r912;
mul.wide.u32 %rd52, %r1013, 8;
add.s64 %rd53, %rd33, %rd52;
shr.u16 %rs52, %rs40, 8;
mul.wide.u16 %r1014, %rs52, 128;
cvt.u32.u16 %r1015, %rs52;
shl.b32 %r1016, %r1015, 21;
or.b32 %r1017, %r1014, %r1016;
and.b32 %r1018, %r1017, 25166208;
or.b32 %r1019, %r1018, -2139062144;
add.s32 %r1020, %r1019, -16843009;
or.b32 %r753, %r1020, %r1018;
mov.u32 %r165, %r753;

prmt.b32 %r165, %r165, 0, 0xba98;

mul.wide.u16 %r1021, %rs52, 8;
shl.b32 %r1022, %r1015, 17;
or.b32 %r1023, %r1021, %r1022;
and.b32 %r1024, %r1023, 25166208;
or.b32 %r1025, %r1024, -2139062144;
add.s32 %r1026, %r1025, -16843009;
or.b32 %r755, %r1026, %r1024;
mov.u32 %r167, %r755;

prmt.b32 %r167, %r167, 0, 0xba98;

ld.global.nc.v2.u32 {%r1027, %r1028}, [%rd53];
not.b32 %r1030, %r1027;
not.b32 %r1032, %r1028;
xor.b32 %r1033, %r1027, %r165;
or.b32 %r1034, %r1033, -2139062144;
and.b32 %r1035, %r165, 2139062143;
and.b32 %r1036, %r1030, -2139062144;
sub.s32 %r1037, %r1034, %r1035;
xor.b32 %r170, %r1036, %r1037;
xor.b32 %r1038, %r1028, %r167;
or.b32 %r1039, %r1038, -2139062144;
and.b32 %r1040, %r167, 2139062143;
and.b32 %r1041, %r1032, -2139062144;
sub.s32 %r1042, %r1039, %r1040;
xor.b32 %r174, %r1041, %r1042;

dp4a.s32.s32 %r169, %r170, %r122, %r161;


dp4a.s32.s32 %r173, %r174, %r126, %r169;

cvt.u32.u16 %r1043, %rs43;
cvt.u32.u16 %r1044, %rs44;
mul.lo.s32 %r1045, %r149, %r1043;
mad.lo.s32 %r1046, %r173, %r1044, %r1045;
add.s32 %r1047, %r173, %r149;
shr.u32 %r1048, %r1047, 31;
add.s32 %r1049, %r1047, %r1048;
shr.s32 %r1050, %r1049, 1;
add.s32 %r1051, %r1046, %r1050;
shr.s32 %r1052, %r1051, 31;
shr.u32 %r1053, %r1052, 30;
add.s32 %r1054, %r1051, %r1053;
shr.s32 %r1055, %r1054, 2;
add.s64 %rd54, %rd133, %rd4;
ld.global.nc.u16 %rs2, [%rd54];

{ cvt.f32.f16 %f87, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r128;
cvt.f32.f16 %f88, low;}


mul.ftz.f32 %f115, %f87, %f88;
cvt.rn.f32.s32 %f116, %r1055;
fma.rn.ftz.f32 %f321, %f115, %f116, %f321;
cvt.u64.u32 %rd55, %r1914;
add.s64 %rd56, %rd1, %rd55;
add.s64 %rd57, %rd56, %rd3;
mov.u32 %r178, %r668;

prmt.b32 %r178, %r178, 0, 0xba98;

mov.u32 %r180, %r670;

prmt.b32 %r180, %r180, 0, 0xba98;

xor.b32 %r1056, %r785, %r178;
or.b32 %r1057, %r1056, -2139062144;
and.b32 %r1058, %r178, 2139062143;
sub.s32 %r1059, %r1057, %r1058;
xor.b32 %r183, %r794, %r1059;
xor.b32 %r1060, %r786, %r180;
or.b32 %r1061, %r1060, -2139062144;
and.b32 %r1062, %r180, 2139062143;
sub.s32 %r1063, %r1061, %r1062;
xor.b32 %r187, %r799, %r1063;
mul.lo.s64 %rd58, %rd57, 36;
add.s64 %rd59, %rd10, %rd58;
ld.global.nc.u32 %r188, [%rd59+8];
ld.global.nc.u32 %r184, [%rd59+4];

dp4a.s32.s32 %r182, %r183, %r184, %r748;


dp4a.s32.s32 %r186, %r187, %r188, %r182;

mov.u32 %r190, %r680;

prmt.b32 %r190, %r190, 0, 0xba98;

mov.u32 %r192, %r682;

prmt.b32 %r192, %r192, 0, 0xba98;

xor.b32 %r1064, %r817, %r190;
or.b32 %r1065, %r1064, -2139062144;
and.b32 %r1066, %r190, 2139062143;
sub.s32 %r1067, %r1065, %r1066;
xor.b32 %r195, %r826, %r1067;
xor.b32 %r1068, %r818, %r192;
or.b32 %r1069, %r1068, -2139062144;
and.b32 %r1070, %r192, 2139062143;
sub.s32 %r1071, %r1069, %r1070;
xor.b32 %r199, %r831, %r1071;
ld.global.nc.u32 %r200, [%rd59+16];
ld.global.nc.u32 %r196, [%rd59+12];

dp4a.s32.s32 %r194, %r195, %r196, %r186;


dp4a.s32.s32 %r198, %r199, %r200, %r194;

mov.u32 %r202, %r692;

prmt.b32 %r202, %r202, 0, 0xba98;

mov.u32 %r204, %r694;

prmt.b32 %r204, %r204, 0, 0xba98;

xor.b32 %r1072, %r850, %r202;
or.b32 %r1073, %r1072, -2139062144;
and.b32 %r1074, %r202, 2139062143;
sub.s32 %r1075, %r1073, %r1074;
xor.b32 %r207, %r859, %r1075;
xor.b32 %r1076, %r851, %r204;
or.b32 %r1077, %r1076, -2139062144;
and.b32 %r1078, %r204, 2139062143;
sub.s32 %r1079, %r1077, %r1078;
xor.b32 %r211, %r864, %r1079;
ld.global.nc.u32 %r212, [%rd59+24];
ld.global.nc.u32 %r208, [%rd59+20];

dp4a.s32.s32 %r206, %r207, %r208, %r748;


dp4a.s32.s32 %r210, %r211, %r212, %r206;

mov.u32 %r214, %r704;

prmt.b32 %r214, %r214, 0, 0xba98;

mov.u32 %r216, %r706;

prmt.b32 %r216, %r216, 0, 0xba98;

xor.b32 %r1080, %r882, %r214;
or.b32 %r1081, %r1080, -2139062144;
and.b32 %r1082, %r214, 2139062143;
sub.s32 %r1083, %r1081, %r1082;
xor.b32 %r219, %r891, %r1083;
xor.b32 %r1084, %r883, %r216;
or.b32 %r1085, %r1084, -2139062144;
and.b32 %r1086, %r216, 2139062143;
sub.s32 %r1087, %r1085, %r1086;
xor.b32 %r223, %r896, %r1087;
ld.global.nc.u32 %r224, [%rd59+32];
ld.global.nc.u32 %r220, [%rd59+28];

dp4a.s32.s32 %r218, %r219, %r220, %r210;


dp4a.s32.s32 %r222, %r223, %r224, %r218;

mul.lo.s32 %r1088, %r198, %r898;
mad.lo.s32 %r1089, %r222, %r899, %r1088;
add.s32 %r1090, %r222, %r198;
shr.u32 %r1091, %r1090, 31;
add.s32 %r1092, %r1090, %r1091;
shr.s32 %r1093, %r1092, 1;
add.s32 %r1094, %r1089, %r1093;
shr.s32 %r1095, %r1094, 31;
shr.u32 %r1096, %r1095, 30;
add.s32 %r1097, %r1094, %r1096;
shr.s32 %r1098, %r1097, 2;

{ cvt.f32.f16 %f89, %rs1;}


ld.global.nc.u32 %r226, [%rd59];

{.reg .f16 low,high;
mov.b32 {low,high},%r226;
cvt.f32.f16 %f90, low;}


mul.ftz.f32 %f117, %f89, %f90;
cvt.rn.f32.s32 %f118, %r1098;
fma.rn.ftz.f32 %f320, %f117, %f118, %f320;
mov.u32 %r227, %r717;

prmt.b32 %r227, %r227, 0, 0xba98;

mov.u32 %r229, %r719;

prmt.b32 %r229, %r229, 0, 0xba98;

xor.b32 %r1099, %r930, %r227;
or.b32 %r1100, %r1099, -2139062144;
and.b32 %r1101, %r227, 2139062143;
sub.s32 %r1102, %r1100, %r1101;
xor.b32 %r232, %r939, %r1102;
xor.b32 %r1103, %r931, %r229;
or.b32 %r1104, %r1103, -2139062144;
and.b32 %r1105, %r229, 2139062143;
sub.s32 %r1106, %r1104, %r1105;
xor.b32 %r236, %r944, %r1106;

dp4a.s32.s32 %r231, %r232, %r184, %r748;


dp4a.s32.s32 %r235, %r236, %r188, %r231;

mov.u32 %r239, %r729;

prmt.b32 %r239, %r239, 0, 0xba98;

mov.u32 %r241, %r731;

prmt.b32 %r241, %r241, 0, 0xba98;

xor.b32 %r1107, %r962, %r239;
or.b32 %r1108, %r1107, -2139062144;
and.b32 %r1109, %r239, 2139062143;
sub.s32 %r1110, %r1108, %r1109;
xor.b32 %r244, %r971, %r1110;
xor.b32 %r1111, %r963, %r241;
or.b32 %r1112, %r1111, -2139062144;
and.b32 %r1113, %r241, 2139062143;
sub.s32 %r1114, %r1112, %r1113;
xor.b32 %r248, %r976, %r1114;

dp4a.s32.s32 %r243, %r244, %r196, %r235;


dp4a.s32.s32 %r247, %r248, %r200, %r243;

mov.u32 %r251, %r741;

prmt.b32 %r251, %r251, 0, 0xba98;

mov.u32 %r253, %r743;

prmt.b32 %r253, %r253, 0, 0xba98;

xor.b32 %r1115, %r995, %r251;
or.b32 %r1116, %r1115, -2139062144;
and.b32 %r1117, %r251, 2139062143;
sub.s32 %r1118, %r1116, %r1117;
xor.b32 %r256, %r1004, %r1118;
xor.b32 %r1119, %r996, %r253;
or.b32 %r1120, %r1119, -2139062144;
and.b32 %r1121, %r253, 2139062143;
sub.s32 %r1122, %r1120, %r1121;
xor.b32 %r260, %r1009, %r1122;

dp4a.s32.s32 %r255, %r256, %r208, %r748;


dp4a.s32.s32 %r259, %r260, %r212, %r255;

mov.u32 %r263, %r753;

prmt.b32 %r263, %r263, 0, 0xba98;

mov.u32 %r265, %r755;

prmt.b32 %r265, %r265, 0, 0xba98;

xor.b32 %r1123, %r1027, %r263;
or.b32 %r1124, %r1123, -2139062144;
and.b32 %r1125, %r263, 2139062143;
sub.s32 %r1126, %r1124, %r1125;
xor.b32 %r268, %r1036, %r1126;
xor.b32 %r1127, %r1028, %r265;
or.b32 %r1128, %r1127, -2139062144;
and.b32 %r1129, %r265, 2139062143;
sub.s32 %r1130, %r1128, %r1129;
xor.b32 %r272, %r1041, %r1130;

dp4a.s32.s32 %r267, %r268, %r220, %r259;


dp4a.s32.s32 %r271, %r272, %r224, %r267;

mul.lo.s32 %r1131, %r247, %r1043;
mad.lo.s32 %r1132, %r271, %r1044, %r1131;
add.s32 %r1133, %r271, %r247;
shr.u32 %r1134, %r1133, 31;
add.s32 %r1135, %r1133, %r1134;
shr.s32 %r1136, %r1135, 1;
add.s32 %r1137, %r1132, %r1136;
shr.s32 %r1138, %r1137, 31;
shr.u32 %r1139, %r1138, 30;
add.s32 %r1140, %r1137, %r1139;
shr.s32 %r1141, %r1140, 2;

{ cvt.f32.f16 %f91, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r226;
cvt.f32.f16 %f92, low;}


mul.ftz.f32 %f119, %f91, %f92;
cvt.rn.f32.s32 %f120, %r1141;
fma.rn.ftz.f32 %f319, %f119, %f120, %f319;
cvt.u64.u32 %rd60, %r1919;
add.s64 %rd61, %rd1, %rd60;
add.s64 %rd62, %rd61, %rd3;
mov.u32 %r276, %r668;

prmt.b32 %r276, %r276, 0, 0xba98;

mov.u32 %r278, %r670;

prmt.b32 %r278, %r278, 0, 0xba98;

xor.b32 %r1142, %r785, %r276;
or.b32 %r1143, %r1142, -2139062144;
and.b32 %r1144, %r276, 2139062143;
sub.s32 %r1145, %r1143, %r1144;
xor.b32 %r281, %r794, %r1145;
xor.b32 %r1146, %r786, %r278;
or.b32 %r1147, %r1146, -2139062144;
and.b32 %r1148, %r278, 2139062143;
sub.s32 %r1149, %r1147, %r1148;
xor.b32 %r285, %r799, %r1149;
mul.lo.s64 %rd63, %rd62, 36;
add.s64 %rd64, %rd10, %rd63;
ld.global.nc.u32 %r286, [%rd64+8];
ld.global.nc.u32 %r282, [%rd64+4];

dp4a.s32.s32 %r280, %r281, %r282, %r748;


dp4a.s32.s32 %r284, %r285, %r286, %r280;

mov.u32 %r288, %r680;

prmt.b32 %r288, %r288, 0, 0xba98;

mov.u32 %r290, %r682;

prmt.b32 %r290, %r290, 0, 0xba98;

xor.b32 %r1150, %r817, %r288;
or.b32 %r1151, %r1150, -2139062144;
and.b32 %r1152, %r288, 2139062143;
sub.s32 %r1153, %r1151, %r1152;
xor.b32 %r293, %r826, %r1153;
xor.b32 %r1154, %r818, %r290;
or.b32 %r1155, %r1154, -2139062144;
and.b32 %r1156, %r290, 2139062143;
sub.s32 %r1157, %r1155, %r1156;
xor.b32 %r297, %r831, %r1157;
ld.global.nc.u32 %r298, [%rd64+16];
ld.global.nc.u32 %r294, [%rd64+12];

dp4a.s32.s32 %r292, %r293, %r294, %r284;


dp4a.s32.s32 %r296, %r297, %r298, %r292;

mov.u32 %r300, %r692;

prmt.b32 %r300, %r300, 0, 0xba98;

mov.u32 %r302, %r694;

prmt.b32 %r302, %r302, 0, 0xba98;

xor.b32 %r1158, %r850, %r300;
or.b32 %r1159, %r1158, -2139062144;
and.b32 %r1160, %r300, 2139062143;
sub.s32 %r1161, %r1159, %r1160;
xor.b32 %r305, %r859, %r1161;
xor.b32 %r1162, %r851, %r302;
or.b32 %r1163, %r1162, -2139062144;
and.b32 %r1164, %r302, 2139062143;
sub.s32 %r1165, %r1163, %r1164;
xor.b32 %r309, %r864, %r1165;
ld.global.nc.u32 %r310, [%rd64+24];
ld.global.nc.u32 %r306, [%rd64+20];

dp4a.s32.s32 %r304, %r305, %r306, %r748;


dp4a.s32.s32 %r308, %r309, %r310, %r304;

mov.u32 %r312, %r704;

prmt.b32 %r312, %r312, 0, 0xba98;

mov.u32 %r314, %r706;

prmt.b32 %r314, %r314, 0, 0xba98;

xor.b32 %r1166, %r882, %r312;
or.b32 %r1167, %r1166, -2139062144;
and.b32 %r1168, %r312, 2139062143;
sub.s32 %r1169, %r1167, %r1168;
xor.b32 %r317, %r891, %r1169;
xor.b32 %r1170, %r883, %r314;
or.b32 %r1171, %r1170, -2139062144;
and.b32 %r1172, %r314, 2139062143;
sub.s32 %r1173, %r1171, %r1172;
xor.b32 %r321, %r896, %r1173;
ld.global.nc.u32 %r322, [%rd64+32];
ld.global.nc.u32 %r318, [%rd64+28];

dp4a.s32.s32 %r316, %r317, %r318, %r308;


dp4a.s32.s32 %r320, %r321, %r322, %r316;

mul.lo.s32 %r1174, %r296, %r898;
mad.lo.s32 %r1175, %r320, %r899, %r1174;
add.s32 %r1176, %r320, %r296;
shr.u32 %r1177, %r1176, 31;
add.s32 %r1178, %r1176, %r1177;
shr.s32 %r1179, %r1178, 1;
add.s32 %r1180, %r1175, %r1179;
shr.s32 %r1181, %r1180, 31;
shr.u32 %r1182, %r1181, 30;
add.s32 %r1183, %r1180, %r1182;
shr.s32 %r1184, %r1183, 2;

{ cvt.f32.f16 %f93, %rs1;}


ld.global.nc.u32 %r324, [%rd64];

{.reg .f16 low,high;
mov.b32 {low,high},%r324;
cvt.f32.f16 %f94, low;}


mul.ftz.f32 %f121, %f93, %f94;
cvt.rn.f32.s32 %f122, %r1184;
fma.rn.ftz.f32 %f318, %f121, %f122, %f318;
mov.u32 %r325, %r717;

prmt.b32 %r325, %r325, 0, 0xba98;

mov.u32 %r327, %r719;

prmt.b32 %r327, %r327, 0, 0xba98;

xor.b32 %r1185, %r930, %r325;
or.b32 %r1186, %r1185, -2139062144;
and.b32 %r1187, %r325, 2139062143;
sub.s32 %r1188, %r1186, %r1187;
xor.b32 %r330, %r939, %r1188;
xor.b32 %r1189, %r931, %r327;
or.b32 %r1190, %r1189, -2139062144;
and.b32 %r1191, %r327, 2139062143;
sub.s32 %r1192, %r1190, %r1191;
xor.b32 %r334, %r944, %r1192;

dp4a.s32.s32 %r329, %r330, %r282, %r748;


dp4a.s32.s32 %r333, %r334, %r286, %r329;

mov.u32 %r337, %r729;

prmt.b32 %r337, %r337, 0, 0xba98;

mov.u32 %r339, %r731;

prmt.b32 %r339, %r339, 0, 0xba98;

xor.b32 %r1193, %r962, %r337;
or.b32 %r1194, %r1193, -2139062144;
and.b32 %r1195, %r337, 2139062143;
sub.s32 %r1196, %r1194, %r1195;
xor.b32 %r342, %r971, %r1196;
xor.b32 %r1197, %r963, %r339;
or.b32 %r1198, %r1197, -2139062144;
and.b32 %r1199, %r339, 2139062143;
sub.s32 %r1200, %r1198, %r1199;
xor.b32 %r346, %r976, %r1200;

dp4a.s32.s32 %r341, %r342, %r294, %r333;


dp4a.s32.s32 %r345, %r346, %r298, %r341;

mov.u32 %r349, %r741;

prmt.b32 %r349, %r349, 0, 0xba98;

mov.u32 %r351, %r743;

prmt.b32 %r351, %r351, 0, 0xba98;

xor.b32 %r1201, %r995, %r349;
or.b32 %r1202, %r1201, -2139062144;
and.b32 %r1203, %r349, 2139062143;
sub.s32 %r1204, %r1202, %r1203;
xor.b32 %r354, %r1004, %r1204;
xor.b32 %r1205, %r996, %r351;
or.b32 %r1206, %r1205, -2139062144;
and.b32 %r1207, %r351, 2139062143;
sub.s32 %r1208, %r1206, %r1207;
xor.b32 %r358, %r1009, %r1208;

dp4a.s32.s32 %r353, %r354, %r306, %r748;


dp4a.s32.s32 %r357, %r358, %r310, %r353;

mov.u32 %r361, %r753;

prmt.b32 %r361, %r361, 0, 0xba98;

mov.u32 %r363, %r755;

prmt.b32 %r363, %r363, 0, 0xba98;

xor.b32 %r1209, %r1027, %r361;
or.b32 %r1210, %r1209, -2139062144;
and.b32 %r1211, %r361, 2139062143;
sub.s32 %r1212, %r1210, %r1211;
xor.b32 %r366, %r1036, %r1212;
xor.b32 %r1213, %r1028, %r363;
or.b32 %r1214, %r1213, -2139062144;
and.b32 %r1215, %r363, 2139062143;
sub.s32 %r1216, %r1214, %r1215;
xor.b32 %r370, %r1041, %r1216;

dp4a.s32.s32 %r365, %r366, %r318, %r357;


dp4a.s32.s32 %r369, %r370, %r322, %r365;

mul.lo.s32 %r1217, %r345, %r1043;
mad.lo.s32 %r1218, %r369, %r1044, %r1217;
add.s32 %r1219, %r369, %r345;
shr.u32 %r1220, %r1219, 31;
add.s32 %r1221, %r1219, %r1220;
shr.s32 %r1222, %r1221, 1;
add.s32 %r1223, %r1218, %r1222;
shr.s32 %r1224, %r1223, 31;
shr.u32 %r1225, %r1224, 30;
add.s32 %r1226, %r1223, %r1225;
shr.s32 %r1227, %r1226, 2;

{ cvt.f32.f16 %f95, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r324;
cvt.f32.f16 %f96, low;}


mul.ftz.f32 %f123, %f95, %f96;
cvt.rn.f32.s32 %f124, %r1227;
fma.rn.ftz.f32 %f317, %f123, %f124, %f317;
cvt.u64.u32 %rd65, %r1918;
add.s64 %rd66, %rd1, %rd65;
add.s64 %rd67, %rd66, %rd3;
mov.u32 %r374, %r668;

prmt.b32 %r374, %r374, 0, 0xba98;

mov.u32 %r376, %r670;

prmt.b32 %r376, %r376, 0, 0xba98;

xor.b32 %r1228, %r785, %r374;
or.b32 %r1229, %r1228, -2139062144;
and.b32 %r1230, %r374, 2139062143;
sub.s32 %r1231, %r1229, %r1230;
xor.b32 %r379, %r794, %r1231;
xor.b32 %r1232, %r786, %r376;
or.b32 %r1233, %r1232, -2139062144;
and.b32 %r1234, %r376, 2139062143;
sub.s32 %r1235, %r1233, %r1234;
xor.b32 %r383, %r799, %r1235;
mul.lo.s64 %rd68, %rd67, 36;
add.s64 %rd69, %rd10, %rd68;
ld.global.nc.u32 %r384, [%rd69+8];
ld.global.nc.u32 %r380, [%rd69+4];

dp4a.s32.s32 %r378, %r379, %r380, %r748;


dp4a.s32.s32 %r382, %r383, %r384, %r378;

mov.u32 %r386, %r680;

prmt.b32 %r386, %r386, 0, 0xba98;

mov.u32 %r388, %r682;

prmt.b32 %r388, %r388, 0, 0xba98;

xor.b32 %r1236, %r817, %r386;
or.b32 %r1237, %r1236, -2139062144;
and.b32 %r1238, %r386, 2139062143;
sub.s32 %r1239, %r1237, %r1238;
xor.b32 %r391, %r826, %r1239;
xor.b32 %r1240, %r818, %r388;
or.b32 %r1241, %r1240, -2139062144;
and.b32 %r1242, %r388, 2139062143;
sub.s32 %r1243, %r1241, %r1242;
xor.b32 %r395, %r831, %r1243;
ld.global.nc.u32 %r396, [%rd69+16];
ld.global.nc.u32 %r392, [%rd69+12];

dp4a.s32.s32 %r390, %r391, %r392, %r382;


dp4a.s32.s32 %r394, %r395, %r396, %r390;

mov.u32 %r398, %r692;

prmt.b32 %r398, %r398, 0, 0xba98;

mov.u32 %r400, %r694;

prmt.b32 %r400, %r400, 0, 0xba98;

xor.b32 %r1244, %r850, %r398;
or.b32 %r1245, %r1244, -2139062144;
and.b32 %r1246, %r398, 2139062143;
sub.s32 %r1247, %r1245, %r1246;
xor.b32 %r403, %r859, %r1247;
xor.b32 %r1248, %r851, %r400;
or.b32 %r1249, %r1248, -2139062144;
and.b32 %r1250, %r400, 2139062143;
sub.s32 %r1251, %r1249, %r1250;
xor.b32 %r407, %r864, %r1251;
ld.global.nc.u32 %r408, [%rd69+24];
ld.global.nc.u32 %r404, [%rd69+20];

dp4a.s32.s32 %r402, %r403, %r404, %r748;


dp4a.s32.s32 %r406, %r407, %r408, %r402;

mov.u32 %r410, %r704;

prmt.b32 %r410, %r410, 0, 0xba98;

mov.u32 %r412, %r706;

prmt.b32 %r412, %r412, 0, 0xba98;

xor.b32 %r1252, %r882, %r410;
or.b32 %r1253, %r1252, -2139062144;
and.b32 %r1254, %r410, 2139062143;
sub.s32 %r1255, %r1253, %r1254;
xor.b32 %r415, %r891, %r1255;
xor.b32 %r1256, %r883, %r412;
or.b32 %r1257, %r1256, -2139062144;
and.b32 %r1258, %r412, 2139062143;
sub.s32 %r1259, %r1257, %r1258;
xor.b32 %r419, %r896, %r1259;
ld.global.nc.u32 %r420, [%rd69+32];
ld.global.nc.u32 %r416, [%rd69+28];

dp4a.s32.s32 %r414, %r415, %r416, %r406;


dp4a.s32.s32 %r418, %r419, %r420, %r414;

mul.lo.s32 %r1260, %r394, %r898;
mad.lo.s32 %r1261, %r418, %r899, %r1260;
add.s32 %r1262, %r418, %r394;
shr.u32 %r1263, %r1262, 31;
add.s32 %r1264, %r1262, %r1263;
shr.s32 %r1265, %r1264, 1;
add.s32 %r1266, %r1261, %r1265;
shr.s32 %r1267, %r1266, 31;
shr.u32 %r1268, %r1267, 30;
add.s32 %r1269, %r1266, %r1268;
shr.s32 %r1270, %r1269, 2;

{ cvt.f32.f16 %f97, %rs1;}


ld.global.nc.u32 %r422, [%rd69];

{.reg .f16 low,high;
mov.b32 {low,high},%r422;
cvt.f32.f16 %f98, low;}


mul.ftz.f32 %f125, %f97, %f98;
cvt.rn.f32.s32 %f126, %r1270;
fma.rn.ftz.f32 %f316, %f125, %f126, %f316;
mov.u32 %r423, %r717;

prmt.b32 %r423, %r423, 0, 0xba98;

mov.u32 %r425, %r719;

prmt.b32 %r425, %r425, 0, 0xba98;

xor.b32 %r1271, %r930, %r423;
or.b32 %r1272, %r1271, -2139062144;
and.b32 %r1273, %r423, 2139062143;
sub.s32 %r1274, %r1272, %r1273;
xor.b32 %r428, %r939, %r1274;
xor.b32 %r1275, %r931, %r425;
or.b32 %r1276, %r1275, -2139062144;
and.b32 %r1277, %r425, 2139062143;
sub.s32 %r1278, %r1276, %r1277;
xor.b32 %r432, %r944, %r1278;

dp4a.s32.s32 %r427, %r428, %r380, %r748;


dp4a.s32.s32 %r431, %r432, %r384, %r427;

mov.u32 %r435, %r729;

prmt.b32 %r435, %r435, 0, 0xba98;

mov.u32 %r437, %r731;

prmt.b32 %r437, %r437, 0, 0xba98;

xor.b32 %r1279, %r962, %r435;
or.b32 %r1280, %r1279, -2139062144;
and.b32 %r1281, %r435, 2139062143;
sub.s32 %r1282, %r1280, %r1281;
xor.b32 %r440, %r971, %r1282;
xor.b32 %r1283, %r963, %r437;
or.b32 %r1284, %r1283, -2139062144;
and.b32 %r1285, %r437, 2139062143;
sub.s32 %r1286, %r1284, %r1285;
xor.b32 %r444, %r976, %r1286;

dp4a.s32.s32 %r439, %r440, %r392, %r431;


dp4a.s32.s32 %r443, %r444, %r396, %r439;

mov.u32 %r447, %r741;

prmt.b32 %r447, %r447, 0, 0xba98;

mov.u32 %r449, %r743;

prmt.b32 %r449, %r449, 0, 0xba98;

xor.b32 %r1287, %r995, %r447;
or.b32 %r1288, %r1287, -2139062144;
and.b32 %r1289, %r447, 2139062143;
sub.s32 %r1290, %r1288, %r1289;
xor.b32 %r452, %r1004, %r1290;
xor.b32 %r1291, %r996, %r449;
or.b32 %r1292, %r1291, -2139062144;
and.b32 %r1293, %r449, 2139062143;
sub.s32 %r1294, %r1292, %r1293;
xor.b32 %r456, %r1009, %r1294;

dp4a.s32.s32 %r451, %r452, %r404, %r748;


dp4a.s32.s32 %r455, %r456, %r408, %r451;

mov.u32 %r459, %r753;

prmt.b32 %r459, %r459, 0, 0xba98;

mov.u32 %r461, %r755;

prmt.b32 %r461, %r461, 0, 0xba98;

xor.b32 %r1295, %r1027, %r459;
or.b32 %r1296, %r1295, -2139062144;
and.b32 %r1297, %r459, 2139062143;
sub.s32 %r1298, %r1296, %r1297;
xor.b32 %r464, %r1036, %r1298;
xor.b32 %r1299, %r1028, %r461;
or.b32 %r1300, %r1299, -2139062144;
and.b32 %r1301, %r461, 2139062143;
sub.s32 %r1302, %r1300, %r1301;
xor.b32 %r468, %r1041, %r1302;

dp4a.s32.s32 %r463, %r464, %r416, %r455;


dp4a.s32.s32 %r467, %r468, %r420, %r463;

mul.lo.s32 %r1303, %r443, %r1043;
mad.lo.s32 %r1304, %r467, %r1044, %r1303;
add.s32 %r1305, %r467, %r443;
shr.u32 %r1306, %r1305, 31;
add.s32 %r1307, %r1305, %r1306;
shr.s32 %r1308, %r1307, 1;
add.s32 %r1309, %r1304, %r1308;
shr.s32 %r1310, %r1309, 31;
shr.u32 %r1311, %r1310, 30;
add.s32 %r1312, %r1309, %r1311;
shr.s32 %r1313, %r1312, 2;

{ cvt.f32.f16 %f99, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r422;
cvt.f32.f16 %f100, low;}


mul.ftz.f32 %f127, %f99, %f100;
cvt.rn.f32.s32 %f128, %r1313;
fma.rn.ftz.f32 %f315, %f127, %f128, %f315;
cvt.u64.u32 %rd70, %r1917;
add.s64 %rd71, %rd1, %rd70;
add.s64 %rd72, %rd71, %rd3;
mov.u32 %r472, %r668;

prmt.b32 %r472, %r472, 0, 0xba98;

mov.u32 %r474, %r670;

prmt.b32 %r474, %r474, 0, 0xba98;

xor.b32 %r1314, %r785, %r472;
or.b32 %r1315, %r1314, -2139062144;
and.b32 %r1316, %r472, 2139062143;
sub.s32 %r1317, %r1315, %r1316;
xor.b32 %r477, %r794, %r1317;
xor.b32 %r1318, %r786, %r474;
or.b32 %r1319, %r1318, -2139062144;
and.b32 %r1320, %r474, 2139062143;
sub.s32 %r1321, %r1319, %r1320;
xor.b32 %r481, %r799, %r1321;
mul.lo.s64 %rd73, %rd72, 36;
add.s64 %rd74, %rd10, %rd73;
ld.global.nc.u32 %r482, [%rd74+8];
ld.global.nc.u32 %r478, [%rd74+4];

dp4a.s32.s32 %r476, %r477, %r478, %r748;


dp4a.s32.s32 %r480, %r481, %r482, %r476;

mov.u32 %r484, %r680;

prmt.b32 %r484, %r484, 0, 0xba98;

mov.u32 %r486, %r682;

prmt.b32 %r486, %r486, 0, 0xba98;

xor.b32 %r1322, %r817, %r484;
or.b32 %r1323, %r1322, -2139062144;
and.b32 %r1324, %r484, 2139062143;
sub.s32 %r1325, %r1323, %r1324;
xor.b32 %r489, %r826, %r1325;
xor.b32 %r1326, %r818, %r486;
or.b32 %r1327, %r1326, -2139062144;
and.b32 %r1328, %r486, 2139062143;
sub.s32 %r1329, %r1327, %r1328;
xor.b32 %r493, %r831, %r1329;
ld.global.nc.u32 %r494, [%rd74+16];
ld.global.nc.u32 %r490, [%rd74+12];

dp4a.s32.s32 %r488, %r489, %r490, %r480;


dp4a.s32.s32 %r492, %r493, %r494, %r488;

mov.u32 %r496, %r692;

prmt.b32 %r496, %r496, 0, 0xba98;

mov.u32 %r498, %r694;

prmt.b32 %r498, %r498, 0, 0xba98;

xor.b32 %r1330, %r850, %r496;
or.b32 %r1331, %r1330, -2139062144;
and.b32 %r1332, %r496, 2139062143;
sub.s32 %r1333, %r1331, %r1332;
xor.b32 %r501, %r859, %r1333;
xor.b32 %r1334, %r851, %r498;
or.b32 %r1335, %r1334, -2139062144;
and.b32 %r1336, %r498, 2139062143;
sub.s32 %r1337, %r1335, %r1336;
xor.b32 %r505, %r864, %r1337;
ld.global.nc.u32 %r506, [%rd74+24];
ld.global.nc.u32 %r502, [%rd74+20];

dp4a.s32.s32 %r500, %r501, %r502, %r748;


dp4a.s32.s32 %r504, %r505, %r506, %r500;

mov.u32 %r508, %r704;

prmt.b32 %r508, %r508, 0, 0xba98;

mov.u32 %r510, %r706;

prmt.b32 %r510, %r510, 0, 0xba98;

xor.b32 %r1338, %r882, %r508;
or.b32 %r1339, %r1338, -2139062144;
and.b32 %r1340, %r508, 2139062143;
sub.s32 %r1341, %r1339, %r1340;
xor.b32 %r513, %r891, %r1341;
xor.b32 %r1342, %r883, %r510;
or.b32 %r1343, %r1342, -2139062144;
and.b32 %r1344, %r510, 2139062143;
sub.s32 %r1345, %r1343, %r1344;
xor.b32 %r517, %r896, %r1345;
ld.global.nc.u32 %r518, [%rd74+32];
ld.global.nc.u32 %r514, [%rd74+28];

dp4a.s32.s32 %r512, %r513, %r514, %r504;


dp4a.s32.s32 %r516, %r517, %r518, %r512;

mul.lo.s32 %r1346, %r492, %r898;
mad.lo.s32 %r1347, %r516, %r899, %r1346;
add.s32 %r1348, %r516, %r492;
shr.u32 %r1349, %r1348, 31;
add.s32 %r1350, %r1348, %r1349;
shr.s32 %r1351, %r1350, 1;
add.s32 %r1352, %r1347, %r1351;
shr.s32 %r1353, %r1352, 31;
shr.u32 %r1354, %r1353, 30;
add.s32 %r1355, %r1352, %r1354;
shr.s32 %r1356, %r1355, 2;

{ cvt.f32.f16 %f101, %rs1;}


ld.global.nc.u32 %r520, [%rd74];

{.reg .f16 low,high;
mov.b32 {low,high},%r520;
cvt.f32.f16 %f102, low;}


mul.ftz.f32 %f129, %f101, %f102;
cvt.rn.f32.s32 %f130, %r1356;
fma.rn.ftz.f32 %f314, %f129, %f130, %f314;
mov.u32 %r521, %r717;

prmt.b32 %r521, %r521, 0, 0xba98;

mov.u32 %r523, %r719;

prmt.b32 %r523, %r523, 0, 0xba98;

xor.b32 %r1357, %r930, %r521;
or.b32 %r1358, %r1357, -2139062144;
and.b32 %r1359, %r521, 2139062143;
sub.s32 %r1360, %r1358, %r1359;
xor.b32 %r526, %r939, %r1360;
xor.b32 %r1361, %r931, %r523;
or.b32 %r1362, %r1361, -2139062144;
and.b32 %r1363, %r523, 2139062143;
sub.s32 %r1364, %r1362, %r1363;
xor.b32 %r530, %r944, %r1364;

dp4a.s32.s32 %r525, %r526, %r478, %r748;


dp4a.s32.s32 %r529, %r530, %r482, %r525;

mov.u32 %r533, %r729;

prmt.b32 %r533, %r533, 0, 0xba98;

mov.u32 %r535, %r731;

prmt.b32 %r535, %r535, 0, 0xba98;

xor.b32 %r1365, %r962, %r533;
or.b32 %r1366, %r1365, -2139062144;
and.b32 %r1367, %r533, 2139062143;
sub.s32 %r1368, %r1366, %r1367;
xor.b32 %r538, %r971, %r1368;
xor.b32 %r1369, %r963, %r535;
or.b32 %r1370, %r1369, -2139062144;
and.b32 %r1371, %r535, 2139062143;
sub.s32 %r1372, %r1370, %r1371;
xor.b32 %r542, %r976, %r1372;

dp4a.s32.s32 %r537, %r538, %r490, %r529;


dp4a.s32.s32 %r541, %r542, %r494, %r537;

mov.u32 %r545, %r741;

prmt.b32 %r545, %r545, 0, 0xba98;

mov.u32 %r547, %r743;

prmt.b32 %r547, %r547, 0, 0xba98;

xor.b32 %r1373, %r995, %r545;
or.b32 %r1374, %r1373, -2139062144;
and.b32 %r1375, %r545, 2139062143;
sub.s32 %r1376, %r1374, %r1375;
xor.b32 %r550, %r1004, %r1376;
xor.b32 %r1377, %r996, %r547;
or.b32 %r1378, %r1377, -2139062144;
and.b32 %r1379, %r547, 2139062143;
sub.s32 %r1380, %r1378, %r1379;
xor.b32 %r554, %r1009, %r1380;

dp4a.s32.s32 %r549, %r550, %r502, %r748;


dp4a.s32.s32 %r553, %r554, %r506, %r549;

mov.u32 %r557, %r753;

prmt.b32 %r557, %r557, 0, 0xba98;

mov.u32 %r559, %r755;

prmt.b32 %r559, %r559, 0, 0xba98;

xor.b32 %r1381, %r1027, %r557;
or.b32 %r1382, %r1381, -2139062144;
and.b32 %r1383, %r557, 2139062143;
sub.s32 %r1384, %r1382, %r1383;
xor.b32 %r562, %r1036, %r1384;
xor.b32 %r1385, %r1028, %r559;
or.b32 %r1386, %r1385, -2139062144;
and.b32 %r1387, %r559, 2139062143;
sub.s32 %r1388, %r1386, %r1387;
xor.b32 %r566, %r1041, %r1388;

dp4a.s32.s32 %r561, %r562, %r514, %r553;


dp4a.s32.s32 %r565, %r566, %r518, %r561;

mul.lo.s32 %r1389, %r541, %r1043;
mad.lo.s32 %r1390, %r565, %r1044, %r1389;
add.s32 %r1391, %r565, %r541;
shr.u32 %r1392, %r1391, 31;
add.s32 %r1393, %r1391, %r1392;
shr.s32 %r1394, %r1393, 1;
add.s32 %r1395, %r1390, %r1394;
shr.s32 %r1396, %r1395, 31;
shr.u32 %r1397, %r1396, 30;
add.s32 %r1398, %r1395, %r1397;
shr.s32 %r1399, %r1398, 2;

{ cvt.f32.f16 %f103, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r520;
cvt.f32.f16 %f104, low;}


mul.ftz.f32 %f131, %f103, %f104;
cvt.rn.f32.s32 %f132, %r1399;
fma.rn.ftz.f32 %f313, %f131, %f132, %f313;
cvt.u64.u32 %rd75, %r1916;
add.s64 %rd76, %rd1, %rd75;
add.s64 %rd77, %rd76, %rd3;
mov.u32 %r570, %r668;

prmt.b32 %r570, %r570, 0, 0xba98;

mov.u32 %r572, %r670;

prmt.b32 %r572, %r572, 0, 0xba98;

xor.b32 %r1400, %r785, %r570;
or.b32 %r1401, %r1400, -2139062144;
and.b32 %r1402, %r570, 2139062143;
sub.s32 %r1403, %r1401, %r1402;
xor.b32 %r575, %r794, %r1403;
xor.b32 %r1404, %r786, %r572;
or.b32 %r1405, %r1404, -2139062144;
and.b32 %r1406, %r572, 2139062143;
sub.s32 %r1407, %r1405, %r1406;
xor.b32 %r579, %r799, %r1407;
mul.lo.s64 %rd78, %rd77, 36;
add.s64 %rd79, %rd10, %rd78;
ld.global.nc.u32 %r580, [%rd79+8];
ld.global.nc.u32 %r576, [%rd79+4];

dp4a.s32.s32 %r574, %r575, %r576, %r748;


dp4a.s32.s32 %r578, %r579, %r580, %r574;

mov.u32 %r582, %r680;

prmt.b32 %r582, %r582, 0, 0xba98;

mov.u32 %r584, %r682;

prmt.b32 %r584, %r584, 0, 0xba98;

xor.b32 %r1408, %r817, %r582;
or.b32 %r1409, %r1408, -2139062144;
and.b32 %r1410, %r582, 2139062143;
sub.s32 %r1411, %r1409, %r1410;
xor.b32 %r587, %r826, %r1411;
xor.b32 %r1412, %r818, %r584;
or.b32 %r1413, %r1412, -2139062144;
and.b32 %r1414, %r584, 2139062143;
sub.s32 %r1415, %r1413, %r1414;
xor.b32 %r591, %r831, %r1415;
ld.global.nc.u32 %r592, [%rd79+16];
ld.global.nc.u32 %r588, [%rd79+12];

dp4a.s32.s32 %r586, %r587, %r588, %r578;


dp4a.s32.s32 %r590, %r591, %r592, %r586;

mov.u32 %r594, %r692;

prmt.b32 %r594, %r594, 0, 0xba98;

mov.u32 %r596, %r694;

prmt.b32 %r596, %r596, 0, 0xba98;

xor.b32 %r1416, %r850, %r594;
or.b32 %r1417, %r1416, -2139062144;
and.b32 %r1418, %r594, 2139062143;
sub.s32 %r1419, %r1417, %r1418;
xor.b32 %r599, %r859, %r1419;
xor.b32 %r1420, %r851, %r596;
or.b32 %r1421, %r1420, -2139062144;
and.b32 %r1422, %r596, 2139062143;
sub.s32 %r1423, %r1421, %r1422;
xor.b32 %r603, %r864, %r1423;
ld.global.nc.u32 %r604, [%rd79+24];
ld.global.nc.u32 %r600, [%rd79+20];

dp4a.s32.s32 %r598, %r599, %r600, %r748;


dp4a.s32.s32 %r602, %r603, %r604, %r598;

mov.u32 %r606, %r704;

prmt.b32 %r606, %r606, 0, 0xba98;

mov.u32 %r608, %r706;

prmt.b32 %r608, %r608, 0, 0xba98;

xor.b32 %r1424, %r882, %r606;
or.b32 %r1425, %r1424, -2139062144;
and.b32 %r1426, %r606, 2139062143;
sub.s32 %r1427, %r1425, %r1426;
xor.b32 %r611, %r891, %r1427;
xor.b32 %r1428, %r883, %r608;
or.b32 %r1429, %r1428, -2139062144;
and.b32 %r1430, %r608, 2139062143;
sub.s32 %r1431, %r1429, %r1430;
xor.b32 %r615, %r896, %r1431;
ld.global.nc.u32 %r616, [%rd79+32];
ld.global.nc.u32 %r612, [%rd79+28];

dp4a.s32.s32 %r610, %r611, %r612, %r602;


dp4a.s32.s32 %r614, %r615, %r616, %r610;

mul.lo.s32 %r1432, %r590, %r898;
mad.lo.s32 %r1433, %r614, %r899, %r1432;
add.s32 %r1434, %r614, %r590;
shr.u32 %r1435, %r1434, 31;
add.s32 %r1436, %r1434, %r1435;
shr.s32 %r1437, %r1436, 1;
add.s32 %r1438, %r1433, %r1437;
shr.s32 %r1439, %r1438, 31;
shr.u32 %r1440, %r1439, 30;
add.s32 %r1441, %r1438, %r1440;
shr.s32 %r1442, %r1441, 2;

{ cvt.f32.f16 %f105, %rs1;}


ld.global.nc.u32 %r618, [%rd79];

{.reg .f16 low,high;
mov.b32 {low,high},%r618;
cvt.f32.f16 %f106, low;}


mul.ftz.f32 %f133, %f105, %f106;
cvt.rn.f32.s32 %f134, %r1442;
fma.rn.ftz.f32 %f312, %f133, %f134, %f312;
mov.u32 %r619, %r717;

prmt.b32 %r619, %r619, 0, 0xba98;

mov.u32 %r621, %r719;

prmt.b32 %r621, %r621, 0, 0xba98;

xor.b32 %r1443, %r930, %r619;
or.b32 %r1444, %r1443, -2139062144;
and.b32 %r1445, %r619, 2139062143;
sub.s32 %r1446, %r1444, %r1445;
xor.b32 %r624, %r939, %r1446;
xor.b32 %r1447, %r931, %r621;
or.b32 %r1448, %r1447, -2139062144;
and.b32 %r1449, %r621, 2139062143;
sub.s32 %r1450, %r1448, %r1449;
xor.b32 %r628, %r944, %r1450;

dp4a.s32.s32 %r623, %r624, %r576, %r748;


dp4a.s32.s32 %r627, %r628, %r580, %r623;

mov.u32 %r631, %r729;

prmt.b32 %r631, %r631, 0, 0xba98;

mov.u32 %r633, %r731;

prmt.b32 %r633, %r633, 0, 0xba98;

xor.b32 %r1451, %r962, %r631;
or.b32 %r1452, %r1451, -2139062144;
and.b32 %r1453, %r631, 2139062143;
sub.s32 %r1454, %r1452, %r1453;
xor.b32 %r636, %r971, %r1454;
xor.b32 %r1455, %r963, %r633;
or.b32 %r1456, %r1455, -2139062144;
and.b32 %r1457, %r633, 2139062143;
sub.s32 %r1458, %r1456, %r1457;
xor.b32 %r640, %r976, %r1458;

dp4a.s32.s32 %r635, %r636, %r588, %r627;


dp4a.s32.s32 %r639, %r640, %r592, %r635;

mov.u32 %r643, %r741;

prmt.b32 %r643, %r643, 0, 0xba98;

mov.u32 %r645, %r743;

prmt.b32 %r645, %r645, 0, 0xba98;

xor.b32 %r1459, %r995, %r643;
or.b32 %r1460, %r1459, -2139062144;
and.b32 %r1461, %r643, 2139062143;
sub.s32 %r1462, %r1460, %r1461;
xor.b32 %r648, %r1004, %r1462;
xor.b32 %r1463, %r996, %r645;
or.b32 %r1464, %r1463, -2139062144;
and.b32 %r1465, %r645, 2139062143;
sub.s32 %r1466, %r1464, %r1465;
xor.b32 %r652, %r1009, %r1466;

dp4a.s32.s32 %r647, %r648, %r600, %r748;


dp4a.s32.s32 %r651, %r652, %r604, %r647;

mov.u32 %r655, %r753;

prmt.b32 %r655, %r655, 0, 0xba98;

mov.u32 %r657, %r755;

prmt.b32 %r657, %r657, 0, 0xba98;

xor.b32 %r1467, %r1027, %r655;
or.b32 %r1468, %r1467, -2139062144;
and.b32 %r1469, %r655, 2139062143;
sub.s32 %r1470, %r1468, %r1469;
xor.b32 %r660, %r1036, %r1470;
xor.b32 %r1471, %r1028, %r657;
or.b32 %r1472, %r1471, -2139062144;
and.b32 %r1473, %r657, 2139062143;
sub.s32 %r1474, %r1472, %r1473;
xor.b32 %r664, %r1041, %r1474;

dp4a.s32.s32 %r659, %r660, %r612, %r651;


dp4a.s32.s32 %r663, %r664, %r616, %r659;

mul.lo.s32 %r1475, %r639, %r1043;
mad.lo.s32 %r1476, %r663, %r1044, %r1475;
add.s32 %r1477, %r663, %r639;
shr.u32 %r1478, %r1477, 31;
add.s32 %r1479, %r1477, %r1478;
shr.s32 %r1480, %r1479, 1;
add.s32 %r1481, %r1476, %r1480;
shr.s32 %r1482, %r1481, 31;
shr.u32 %r1483, %r1482, 30;
add.s32 %r1484, %r1481, %r1483;
shr.s32 %r1485, %r1484, 2;

{ cvt.f32.f16 %f107, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r618;
cvt.f32.f16 %f108, low;}


mul.ftz.f32 %f135, %f107, %f108;
cvt.rn.f32.s32 %f136, %r1485;
fma.rn.ftz.f32 %f311, %f135, %f136, %f311;
cvt.u64.u32 %rd80, %r1915;
add.s64 %rd81, %rd1, %rd80;
add.s64 %rd82, %rd81, %rd3;

prmt.b32 %r668, %r668, 0, 0xba98;


prmt.b32 %r670, %r670, 0, 0xba98;

xor.b32 %r1486, %r785, %r668;
or.b32 %r1487, %r1486, -2139062144;
and.b32 %r1488, %r668, 2139062143;
sub.s32 %r1489, %r1487, %r1488;
xor.b32 %r673, %r794, %r1489;
xor.b32 %r1490, %r786, %r670;
or.b32 %r1491, %r1490, -2139062144;
and.b32 %r1492, %r670, 2139062143;
sub.s32 %r1493, %r1491, %r1492;
xor.b32 %r677, %r799, %r1493;
mul.lo.s64 %rd83, %rd82, 36;
add.s64 %rd84, %rd10, %rd83;
ld.global.nc.u32 %r678, [%rd84+8];
ld.global.nc.u32 %r674, [%rd84+4];

dp4a.s32.s32 %r672, %r673, %r674, %r748;


dp4a.s32.s32 %r676, %r677, %r678, %r672;


prmt.b32 %r680, %r680, 0, 0xba98;


prmt.b32 %r682, %r682, 0, 0xba98;

xor.b32 %r1494, %r817, %r680;
or.b32 %r1495, %r1494, -2139062144;
and.b32 %r1496, %r680, 2139062143;
sub.s32 %r1497, %r1495, %r1496;
xor.b32 %r685, %r826, %r1497;
xor.b32 %r1498, %r818, %r682;
or.b32 %r1499, %r1498, -2139062144;
and.b32 %r1500, %r682, 2139062143;
sub.s32 %r1501, %r1499, %r1500;
xor.b32 %r689, %r831, %r1501;
ld.global.nc.u32 %r690, [%rd84+16];
ld.global.nc.u32 %r686, [%rd84+12];

dp4a.s32.s32 %r684, %r685, %r686, %r676;


dp4a.s32.s32 %r688, %r689, %r690, %r684;


prmt.b32 %r692, %r692, 0, 0xba98;


prmt.b32 %r694, %r694, 0, 0xba98;

xor.b32 %r1502, %r850, %r692;
or.b32 %r1503, %r1502, -2139062144;
and.b32 %r1504, %r692, 2139062143;
sub.s32 %r1505, %r1503, %r1504;
xor.b32 %r697, %r859, %r1505;
xor.b32 %r1506, %r851, %r694;
or.b32 %r1507, %r1506, -2139062144;
and.b32 %r1508, %r694, 2139062143;
sub.s32 %r1509, %r1507, %r1508;
xor.b32 %r701, %r864, %r1509;
ld.global.nc.u32 %r702, [%rd84+24];
ld.global.nc.u32 %r698, [%rd84+20];

dp4a.s32.s32 %r696, %r697, %r698, %r748;


dp4a.s32.s32 %r700, %r701, %r702, %r696;


prmt.b32 %r704, %r704, 0, 0xba98;


prmt.b32 %r706, %r706, 0, 0xba98;

xor.b32 %r1510, %r882, %r704;
or.b32 %r1511, %r1510, -2139062144;
and.b32 %r1512, %r704, 2139062143;
sub.s32 %r1513, %r1511, %r1512;
xor.b32 %r709, %r891, %r1513;
xor.b32 %r1514, %r883, %r706;
or.b32 %r1515, %r1514, -2139062144;
and.b32 %r1516, %r706, 2139062143;
sub.s32 %r1517, %r1515, %r1516;
xor.b32 %r713, %r896, %r1517;
ld.global.nc.u32 %r714, [%rd84+32];
ld.global.nc.u32 %r710, [%rd84+28];

dp4a.s32.s32 %r708, %r709, %r710, %r700;


dp4a.s32.s32 %r712, %r713, %r714, %r708;

mul.lo.s32 %r1518, %r688, %r898;
mad.lo.s32 %r1519, %r712, %r899, %r1518;
add.s32 %r1520, %r712, %r688;
shr.u32 %r1521, %r1520, 31;
add.s32 %r1522, %r1520, %r1521;
shr.s32 %r1523, %r1522, 1;
add.s32 %r1524, %r1519, %r1523;
shr.s32 %r1525, %r1524, 31;
shr.u32 %r1526, %r1525, 30;
add.s32 %r1527, %r1524, %r1526;
shr.s32 %r1528, %r1527, 2;

{ cvt.f32.f16 %f109, %rs1;}


ld.global.nc.u32 %r716, [%rd84];

{.reg .f16 low,high;
mov.b32 {low,high},%r716;
cvt.f32.f16 %f110, low;}


mul.ftz.f32 %f137, %f109, %f110;
cvt.rn.f32.s32 %f138, %r1528;
fma.rn.ftz.f32 %f310, %f137, %f138, %f310;

prmt.b32 %r717, %r717, 0, 0xba98;


prmt.b32 %r719, %r719, 0, 0xba98;

xor.b32 %r1529, %r930, %r717;
or.b32 %r1530, %r1529, -2139062144;
and.b32 %r1531, %r717, 2139062143;
sub.s32 %r1532, %r1530, %r1531;
xor.b32 %r722, %r939, %r1532;
xor.b32 %r1533, %r931, %r719;
or.b32 %r1534, %r1533, -2139062144;
and.b32 %r1535, %r719, 2139062143;
sub.s32 %r1536, %r1534, %r1535;
xor.b32 %r726, %r944, %r1536;

dp4a.s32.s32 %r721, %r722, %r674, %r748;


dp4a.s32.s32 %r725, %r726, %r678, %r721;


prmt.b32 %r729, %r729, 0, 0xba98;


prmt.b32 %r731, %r731, 0, 0xba98;

xor.b32 %r1537, %r962, %r729;
or.b32 %r1538, %r1537, -2139062144;
and.b32 %r1539, %r729, 2139062143;
sub.s32 %r1540, %r1538, %r1539;
xor.b32 %r734, %r971, %r1540;
xor.b32 %r1541, %r963, %r731;
or.b32 %r1542, %r1541, -2139062144;
and.b32 %r1543, %r731, 2139062143;
sub.s32 %r1544, %r1542, %r1543;
xor.b32 %r738, %r976, %r1544;

dp4a.s32.s32 %r733, %r734, %r686, %r725;


dp4a.s32.s32 %r737, %r738, %r690, %r733;


prmt.b32 %r741, %r741, 0, 0xba98;


prmt.b32 %r743, %r743, 0, 0xba98;

xor.b32 %r1545, %r995, %r741;
or.b32 %r1546, %r1545, -2139062144;
and.b32 %r1547, %r741, 2139062143;
sub.s32 %r1548, %r1546, %r1547;
xor.b32 %r746, %r1004, %r1548;
xor.b32 %r1549, %r996, %r743;
or.b32 %r1550, %r1549, -2139062144;
and.b32 %r1551, %r743, 2139062143;
sub.s32 %r1552, %r1550, %r1551;
xor.b32 %r750, %r1009, %r1552;

dp4a.s32.s32 %r745, %r746, %r698, %r748;


dp4a.s32.s32 %r749, %r750, %r702, %r745;


prmt.b32 %r753, %r753, 0, 0xba98;


prmt.b32 %r755, %r755, 0, 0xba98;

xor.b32 %r1553, %r1027, %r753;
or.b32 %r1554, %r1553, -2139062144;
and.b32 %r1555, %r753, 2139062143;
sub.s32 %r1556, %r1554, %r1555;
xor.b32 %r758, %r1036, %r1556;
xor.b32 %r1557, %r1028, %r755;
or.b32 %r1558, %r1557, -2139062144;
and.b32 %r1559, %r755, 2139062143;
sub.s32 %r1560, %r1558, %r1559;
xor.b32 %r762, %r1041, %r1560;

dp4a.s32.s32 %r757, %r758, %r710, %r749;


dp4a.s32.s32 %r761, %r762, %r714, %r757;

mul.lo.s32 %r1561, %r737, %r1043;
mad.lo.s32 %r1562, %r761, %r1044, %r1561;
add.s32 %r1563, %r761, %r737;
shr.u32 %r1564, %r1563, 31;
add.s32 %r1565, %r1563, %r1564;
shr.s32 %r1566, %r1565, 1;
add.s32 %r1567, %r1562, %r1566;
shr.s32 %r1568, %r1567, 31;
shr.u32 %r1569, %r1568, 30;
add.s32 %r1570, %r1567, %r1569;
shr.s32 %r1571, %r1570, 2;

{ cvt.f32.f16 %f111, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r716;
cvt.f32.f16 %f112, low;}


mul.ftz.f32 %f139, %f111, %f112;
cvt.rn.f32.s32 %f140, %r1571;
fma.rn.ftz.f32 %f309, %f139, %f140, %f309;
add.s32 %r1919, %r1919, 64;
add.s32 %r1918, %r1918, 64;
add.s32 %r1917, %r1917, 64;
add.s32 %r1916, %r1916, 64;
add.s32 %r1915, %r1915, 64;
add.s32 %r1914, %r1914, 64;
add.s32 %r1913, %r1913, 64;
add.s64 %rd133, %rd133, 656;
add.s32 %r1920, %r1920, 8;
setp.lt.s32 %p3, %r1920, %r3;
@%p3 bra $L__BB235_2;

$L__BB235_3:
setp.eq.s32 %p4, %r45, 0;
@%p4 bra $L__BB235_5;

mov.u32 %r1575, _ZZ13mul_mat_vec_qIL9ggml_type22ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r1576, %r45, 1792, %r1575;
shl.b32 %r1577, %r47, 2;
add.s32 %r1578, %r1576, %r1577;
st.shared.f32 [%r1578+-1792], %f322;
st.shared.f32 [%r1578+-1664], %f321;
st.shared.f32 [%r1578+-1536], %f320;
st.shared.f32 [%r1578+-1408], %f319;
st.shared.f32 [%r1578+-1280], %f318;
st.shared.f32 [%r1578+-1152], %f317;
st.shared.f32 [%r1578+-1024], %f316;
st.shared.f32 [%r1578+-896], %f315;
st.shared.f32 [%r1578+-768], %f314;
st.shared.f32 [%r1578+-640], %f313;
st.shared.f32 [%r1578+-512], %f312;
st.shared.f32 [%r1578+-384], %f311;
st.shared.f32 [%r1578+-256], %f310;
st.shared.f32 [%r1578+-128], %f309;

$L__BB235_5:
setp.ne.s32 %p5, %r45, 0;
bar.sync 0;
@%p5 bra $L__BB235_34;

mov.u32 %r1580, %ctaid.z;
mov.u32 %r1581, %ctaid.y;
mov.u32 %r1582, %ctaid.x;
shl.b32 %r1583, %r1582, 1;
mov.u32 %r1584, 1;
mad.lo.s32 %r1585, %r1581, %r37, %r1583;
mad.lo.s32 %r1586, %r1580, %r43, %r1585;
cvt.u64.u32 %rd13, %r1586;
shl.b32 %r1588, %r47, 2;
mov.u32 %r1589, _ZZ13mul_mat_vec_qIL9ggml_type22ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r1590, %r1589, %r1588;
mov.u32 %r1591, 2;
ld.shared.f32 %f141, [%r1590];
add.ftz.f32 %f142, %f141, %f322;
mov.b32 %r1592, %f142;
mov.u32 %r1593, 31;
mov.u32 %r1594, 16;
mov.u32 %r1595, -1;
shfl.sync.bfly.b32 %r1596|%p6, %r1592, %r1594, %r1593, %r1595;
mov.b32 %f143, %r1596;
add.ftz.f32 %f144, %f142, %f143;
mov.b32 %r1597, %f144;
mov.u32 %r1598, 8;
shfl.sync.bfly.b32 %r1599|%p7, %r1597, %r1598, %r1593, %r1595;
mov.b32 %f145, %r1599;
add.ftz.f32 %f146, %f144, %f145;
mov.b32 %r1600, %f146;
mov.u32 %r1601, 4;
shfl.sync.bfly.b32 %r1602|%p8, %r1600, %r1601, %r1593, %r1595;
mov.b32 %f147, %r1602;
add.ftz.f32 %f148, %f146, %f147;
mov.b32 %r1603, %f148;
shfl.sync.bfly.b32 %r1604|%p9, %r1603, %r1591, %r1593, %r1595;
mov.b32 %f149, %r1604;
add.ftz.f32 %f150, %f148, %f149;
mov.b32 %r1605, %f150;
shfl.sync.bfly.b32 %r1606|%p10, %r1605, %r1584, %r1593, %r1595;
mov.b32 %f151, %r1606;
add.ftz.f32 %f43, %f150, %f151;
setp.ne.s32 %p11, %r47, 0;
setp.ge.u32 %p12, %r1583, %r31;
or.pred %p13, %p11, %p12;
@%p13 bra $L__BB235_8;

cvta.to.global.u64 %rd85, %rd19;
mul.wide.u32 %rd86, %r1586, 4;
add.s64 %rd87, %rd85, %rd86;
st.global.f32 [%rd87], %f43;

$L__BB235_8:
ld.shared.f32 %f152, [%r1590+128];
add.ftz.f32 %f153, %f152, %f321;
mov.b32 %r1618, %f153;
shfl.sync.bfly.b32 %r1622|%p14, %r1618, %r1594, %r1593, %r1595;
mov.b32 %f154, %r1622;
add.ftz.f32 %f155, %f153, %f154;
mov.b32 %r1623, %f155;
shfl.sync.bfly.b32 %r1625|%p15, %r1623, %r1598, %r1593, %r1595;
mov.b32 %f156, %r1625;
add.ftz.f32 %f157, %f155, %f156;
mov.b32 %r1626, %f157;
shfl.sync.bfly.b32 %r1628|%p16, %r1626, %r1601, %r1593, %r1595;
mov.b32 %f158, %r1628;
add.ftz.f32 %f159, %f157, %f158;
mov.b32 %r1629, %f159;
shfl.sync.bfly.b32 %r1630|%p17, %r1629, %r1591, %r1593, %r1595;
mov.b32 %f160, %r1630;
add.ftz.f32 %f161, %f159, %f160;
mov.b32 %r1631, %f161;
shfl.sync.bfly.b32 %r1633|%p18, %r1631, %r1584, %r1593, %r1595;
mov.b32 %f162, %r1633;
add.ftz.f32 %f44, %f161, %f162;
setp.eq.s32 %p19, %r47, 1;
or.b32 %r1636, %r1583, 1;
setp.lt.u32 %p20, %r1636, %r31;
and.pred %p1, %p19, %p20;
not.pred %p21, %p1;
@%p21 bra $L__BB235_10;

cvta.to.global.u64 %rd88, %rd19;
mul.wide.u32 %rd89, %r1586, 4;
add.s64 %rd90, %rd88, %rd89;
st.global.f32 [%rd90+4], %f44;

$L__BB235_10:
mov.u32 %r1647, 2;
ld.shared.f32 %f163, [%r1590+256];
add.ftz.f32 %f164, %f163, %f320;
mov.b32 %r1648, %f164;
mov.u32 %r1649, 31;
mov.u32 %r1650, 16;
mov.u32 %r1651, -1;
shfl.sync.bfly.b32 %r1652|%p22, %r1648, %r1650, %r1649, %r1651;
mov.b32 %f165, %r1652;
add.ftz.f32 %f166, %f164, %f165;
mov.b32 %r1653, %f166;
mov.u32 %r1654, 8;
shfl.sync.bfly.b32 %r1655|%p23, %r1653, %r1654, %r1649, %r1651;
mov.b32 %f167, %r1655;
add.ftz.f32 %f168, %f166, %f167;
mov.b32 %r1656, %f168;
mov.u32 %r1657, 4;
shfl.sync.bfly.b32 %r1658|%p24, %r1656, %r1657, %r1649, %r1651;
mov.b32 %f169, %r1658;
add.ftz.f32 %f170, %f168, %f169;
mov.b32 %r1659, %f170;
shfl.sync.bfly.b32 %r1660|%p25, %r1659, %r1647, %r1649, %r1651;
mov.b32 %f171, %r1660;
add.ftz.f32 %f172, %f170, %f171;
mov.b32 %r1661, %f172;
mov.u32 %r1662, 1;
shfl.sync.bfly.b32 %r1663|%p26, %r1661, %r1662, %r1649, %r1651;
mov.b32 %f173, %r1663;
add.ftz.f32 %f45, %f172, %f173;
@%p13 bra $L__BB235_12;

cvt.u64.u32 %rd91, %r31;
add.s64 %rd92, %rd91, %rd13;
cvta.to.global.u64 %rd93, %rd19;
shl.b64 %rd94, %rd92, 2;
add.s64 %rd95, %rd93, %rd94;
st.global.f32 [%rd95], %f45;

$L__BB235_12:
ld.shared.f32 %f174, [%r1590+384];
add.ftz.f32 %f175, %f174, %f319;
mov.b32 %r1671, %f175;
shfl.sync.bfly.b32 %r1675|%p30, %r1671, %r1650, %r1649, %r1651;
mov.b32 %f176, %r1675;
add.ftz.f32 %f177, %f175, %f176;
mov.b32 %r1676, %f177;
shfl.sync.bfly.b32 %r1678|%p31, %r1676, %r1654, %r1649, %r1651;
mov.b32 %f178, %r1678;
add.ftz.f32 %f179, %f177, %f178;
mov.b32 %r1679, %f179;
shfl.sync.bfly.b32 %r1681|%p32, %r1679, %r1657, %r1649, %r1651;
mov.b32 %f180, %r1681;
add.ftz.f32 %f181, %f179, %f180;
mov.b32 %r1682, %f181;
shfl.sync.bfly.b32 %r1683|%p33, %r1682, %r1647, %r1649, %r1651;
mov.b32 %f182, %r1683;
add.ftz.f32 %f183, %f181, %f182;
mov.b32 %r1684, %f183;
shfl.sync.bfly.b32 %r1686|%p34, %r1684, %r1662, %r1649, %r1651;
mov.b32 %f184, %r1686;
add.ftz.f32 %f46, %f183, %f184;
@%p21 bra $L__BB235_14;

add.s32 %r1687, %r31, 1;
cvt.u64.u32 %rd96, %r1687;
add.s64 %rd97, %rd96, %rd13;
cvta.to.global.u64 %rd98, %rd19;
shl.b64 %rd99, %rd97, 2;
add.s64 %rd100, %rd98, %rd99;
st.global.f32 [%rd100], %f46;

$L__BB235_14:
shl.b32 %r1688, %r31, 1;
mov.u32 %r1689, 1;
mov.u32 %r1694, 2;
ld.shared.f32 %f185, [%r1590+512];
add.ftz.f32 %f186, %f185, %f318;
mov.b32 %r1695, %f186;
mov.u32 %r1696, 31;
mov.u32 %r1697, 16;
mov.u32 %r1698, -1;
shfl.sync.bfly.b32 %r1699|%p36, %r1695, %r1697, %r1696, %r1698;
mov.b32 %f187, %r1699;
add.ftz.f32 %f188, %f186, %f187;
mov.b32 %r1700, %f188;
mov.u32 %r1701, 8;
shfl.sync.bfly.b32 %r1702|%p37, %r1700, %r1701, %r1696, %r1698;
mov.b32 %f189, %r1702;
add.ftz.f32 %f190, %f188, %f189;
mov.b32 %r1703, %f190;
mov.u32 %r1704, 4;
shfl.sync.bfly.b32 %r1705|%p38, %r1703, %r1704, %r1696, %r1698;
mov.b32 %f191, %r1705;
add.ftz.f32 %f192, %f190, %f191;
mov.b32 %r1706, %f192;
shfl.sync.bfly.b32 %r1707|%p39, %r1706, %r1694, %r1696, %r1698;
mov.b32 %f193, %r1707;
add.ftz.f32 %f194, %f192, %f193;
mov.b32 %r1708, %f194;
shfl.sync.bfly.b32 %r1709|%p40, %r1708, %r1689, %r1696, %r1698;
mov.b32 %f195, %r1709;
add.ftz.f32 %f47, %f194, %f195;
cvt.u64.u32 %rd101, %r1688;
add.s64 %rd102, %rd101, %rd13;
cvta.to.global.u64 %rd103, %rd19;
shl.b64 %rd104, %rd102, 2;
add.s64 %rd14, %rd103, %rd104;
@%p13 bra $L__BB235_16;

st.global.f32 [%rd14], %f47;

$L__BB235_16:
ld.shared.f32 %f196, [%r1590+640];
add.ftz.f32 %f197, %f196, %f317;
mov.b32 %r1717, %f197;
shfl.sync.bfly.b32 %r1721|%p44, %r1717, %r1697, %r1696, %r1698;
mov.b32 %f198, %r1721;
add.ftz.f32 %f199, %f197, %f198;
mov.b32 %r1722, %f199;
shfl.sync.bfly.b32 %r1724|%p45, %r1722, %r1701, %r1696, %r1698;
mov.b32 %f200, %r1724;
add.ftz.f32 %f201, %f199, %f200;
mov.b32 %r1725, %f201;
shfl.sync.bfly.b32 %r1727|%p46, %r1725, %r1704, %r1696, %r1698;
mov.b32 %f202, %r1727;
add.ftz.f32 %f203, %f201, %f202;
mov.b32 %r1728, %f203;
shfl.sync.bfly.b32 %r1729|%p47, %r1728, %r1694, %r1696, %r1698;
mov.b32 %f204, %r1729;
add.ftz.f32 %f205, %f203, %f204;
mov.b32 %r1730, %f205;
shfl.sync.bfly.b32 %r1732|%p48, %r1730, %r1689, %r1696, %r1698;
mov.b32 %f206, %r1732;
add.ftz.f32 %f48, %f205, %f206;
@%p21 bra $L__BB235_18;

st.global.f32 [%rd14+4], %f48;

$L__BB235_18:
mul.lo.s32 %r27, %r31, 3;
mov.u32 %r1737, 2;
ld.shared.f32 %f207, [%r1590+768];
add.ftz.f32 %f208, %f207, %f316;
mov.b32 %r1738, %f208;
mov.u32 %r1739, 31;
mov.u32 %r1740, 16;
mov.u32 %r1741, -1;
shfl.sync.bfly.b32 %r1742|%p50, %r1738, %r1740, %r1739, %r1741;
mov.b32 %f209, %r1742;
add.ftz.f32 %f210, %f208, %f209;
mov.b32 %r1743, %f210;
mov.u32 %r1744, 8;
shfl.sync.bfly.b32 %r1745|%p51, %r1743, %r1744, %r1739, %r1741;
mov.b32 %f211, %r1745;
add.ftz.f32 %f212, %f210, %f211;
mov.b32 %r1746, %f212;
mov.u32 %r1747, 4;
shfl.sync.bfly.b32 %r1748|%p52, %r1746, %r1747, %r1739, %r1741;
mov.b32 %f213, %r1748;
add.ftz.f32 %f214, %f212, %f213;
mov.b32 %r1749, %f214;
shfl.sync.bfly.b32 %r1750|%p53, %r1749, %r1737, %r1739, %r1741;
mov.b32 %f215, %r1750;
add.ftz.f32 %f216, %f214, %f215;
mov.b32 %r1751, %f216;
mov.u32 %r1752, 1;
shfl.sync.bfly.b32 %r1753|%p54, %r1751, %r1752, %r1739, %r1741;
mov.b32 %f217, %r1753;
add.ftz.f32 %f49, %f216, %f217;
@%p13 bra $L__BB235_20;

cvt.u64.u32 %rd105, %r27;
add.s64 %rd106, %rd105, %rd13;
shl.b64 %rd108, %rd106, 2;
add.s64 %rd109, %rd103, %rd108;
st.global.f32 [%rd109], %f49;

$L__BB235_20:
ld.shared.f32 %f218, [%r1590+896];
add.ftz.f32 %f219, %f218, %f315;
mov.b32 %r1761, %f219;
shfl.sync.bfly.b32 %r1765|%p58, %r1761, %r1740, %r1739, %r1741;
mov.b32 %f220, %r1765;
add.ftz.f32 %f221, %f219, %f220;
mov.b32 %r1766, %f221;
shfl.sync.bfly.b32 %r1768|%p59, %r1766, %r1744, %r1739, %r1741;
mov.b32 %f222, %r1768;
add.ftz.f32 %f223, %f221, %f222;
mov.b32 %r1769, %f223;
shfl.sync.bfly.b32 %r1771|%p60, %r1769, %r1747, %r1739, %r1741;
mov.b32 %f224, %r1771;
add.ftz.f32 %f225, %f223, %f224;
mov.b32 %r1772, %f225;
shfl.sync.bfly.b32 %r1773|%p61, %r1772, %r1737, %r1739, %r1741;
mov.b32 %f226, %r1773;
add.ftz.f32 %f227, %f225, %f226;
mov.b32 %r1774, %f227;
shfl.sync.bfly.b32 %r1776|%p62, %r1774, %r1752, %r1739, %r1741;
mov.b32 %f228, %r1776;
add.ftz.f32 %f50, %f227, %f228;
@%p21 bra $L__BB235_22;

add.s32 %r1777, %r27, 1;
cvt.u64.u32 %rd110, %r1777;
add.s64 %rd111, %rd110, %rd13;
shl.b64 %rd113, %rd111, 2;
add.s64 %rd114, %rd103, %rd113;
st.global.f32 [%rd114], %f50;

$L__BB235_22:
shl.b32 %r1778, %r31, 2;
mov.u32 %r1779, 2;
ld.shared.f32 %f229, [%r1590+1024];
add.ftz.f32 %f230, %f229, %f314;
mov.b32 %r1784, %f230;
mov.u32 %r1785, 31;
mov.u32 %r1786, 16;
mov.u32 %r1787, -1;
shfl.sync.bfly.b32 %r1788|%p64, %r1784, %r1786, %r1785, %r1787;
mov.b32 %f231, %r1788;
add.ftz.f32 %f232, %f230, %f231;
mov.b32 %r1789, %f232;
mov.u32 %r1790, 8;
shfl.sync.bfly.b32 %r1791|%p65, %r1789, %r1790, %r1785, %r1787;
mov.b32 %f233, %r1791;
add.ftz.f32 %f234, %f232, %f233;
mov.b32 %r1792, %f234;
mov.u32 %r1793, 4;
shfl.sync.bfly.b32 %r1794|%p66, %r1792, %r1793, %r1785, %r1787;
mov.b32 %f235, %r1794;
add.ftz.f32 %f236, %f234, %f235;
mov.b32 %r1795, %f236;
shfl.sync.bfly.b32 %r1796|%p67, %r1795, %r1779, %r1785, %r1787;
mov.b32 %f237, %r1796;
add.ftz.f32 %f238, %f236, %f237;
mov.b32 %r1797, %f238;
mov.u32 %r1798, 1;
shfl.sync.bfly.b32 %r1799|%p68, %r1797, %r1798, %r1785, %r1787;
mov.b32 %f239, %r1799;
add.ftz.f32 %f51, %f238, %f239;
cvt.u64.u32 %rd115, %r1778;
add.s64 %rd116, %rd115, %rd13;
shl.b64 %rd118, %rd116, 2;
add.s64 %rd15, %rd103, %rd118;
@%p13 bra $L__BB235_24;

st.global.f32 [%rd15], %f51;

$L__BB235_24:
ld.shared.f32 %f240, [%r1590+1152];
add.ftz.f32 %f241, %f240, %f313;
mov.b32 %r1807, %f241;
shfl.sync.bfly.b32 %r1811|%p72, %r1807, %r1786, %r1785, %r1787;
mov.b32 %f242, %r1811;
add.ftz.f32 %f243, %f241, %f242;
mov.b32 %r1812, %f243;
shfl.sync.bfly.b32 %r1814|%p73, %r1812, %r1790, %r1785, %r1787;
mov.b32 %f244, %r1814;
add.ftz.f32 %f245, %f243, %f244;
mov.b32 %r1815, %f245;
shfl.sync.bfly.b32 %r1817|%p74, %r1815, %r1793, %r1785, %r1787;
mov.b32 %f246, %r1817;
add.ftz.f32 %f247, %f245, %f246;
mov.b32 %r1818, %f247;
shfl.sync.bfly.b32 %r1819|%p75, %r1818, %r1779, %r1785, %r1787;
mov.b32 %f248, %r1819;
add.ftz.f32 %f249, %f247, %f248;
mov.b32 %r1820, %f249;
shfl.sync.bfly.b32 %r1822|%p76, %r1820, %r1798, %r1785, %r1787;
mov.b32 %f250, %r1822;
add.ftz.f32 %f52, %f249, %f250;
@%p21 bra $L__BB235_26;

st.global.f32 [%rd15+4], %f52;

$L__BB235_26:
mul.lo.s32 %r28, %r31, 5;
mov.u32 %r1827, 2;
ld.shared.f32 %f251, [%r1590+1280];
add.ftz.f32 %f252, %f251, %f312;
mov.b32 %r1828, %f252;
mov.u32 %r1829, 31;
mov.u32 %r1830, 16;
mov.u32 %r1831, -1;
shfl.sync.bfly.b32 %r1832|%p78, %r1828, %r1830, %r1829, %r1831;
mov.b32 %f253, %r1832;
add.ftz.f32 %f254, %f252, %f253;
mov.b32 %r1833, %f254;
mov.u32 %r1834, 8;
shfl.sync.bfly.b32 %r1835|%p79, %r1833, %r1834, %r1829, %r1831;
mov.b32 %f255, %r1835;
add.ftz.f32 %f256, %f254, %f255;
mov.b32 %r1836, %f256;
mov.u32 %r1837, 4;
shfl.sync.bfly.b32 %r1838|%p80, %r1836, %r1837, %r1829, %r1831;
mov.b32 %f257, %r1838;
add.ftz.f32 %f258, %f256, %f257;
mov.b32 %r1839, %f258;
shfl.sync.bfly.b32 %r1840|%p81, %r1839, %r1827, %r1829, %r1831;
mov.b32 %f259, %r1840;
add.ftz.f32 %f260, %f258, %f259;
mov.b32 %r1841, %f260;
mov.u32 %r1842, 1;
shfl.sync.bfly.b32 %r1843|%p82, %r1841, %r1842, %r1829, %r1831;
mov.b32 %f261, %r1843;
add.ftz.f32 %f53, %f260, %f261;
@%p13 bra $L__BB235_28;

cvt.u64.u32 %rd119, %r28;
add.s64 %rd120, %rd119, %rd13;
shl.b64 %rd122, %rd120, 2;
add.s64 %rd123, %rd103, %rd122;
st.global.f32 [%rd123], %f53;

$L__BB235_28:
ld.shared.f32 %f262, [%r1590+1408];
add.ftz.f32 %f263, %f262, %f311;
mov.b32 %r1851, %f263;
shfl.sync.bfly.b32 %r1855|%p86, %r1851, %r1830, %r1829, %r1831;
mov.b32 %f264, %r1855;
add.ftz.f32 %f265, %f263, %f264;
mov.b32 %r1856, %f265;
shfl.sync.bfly.b32 %r1858|%p87, %r1856, %r1834, %r1829, %r1831;
mov.b32 %f266, %r1858;
add.ftz.f32 %f267, %f265, %f266;
mov.b32 %r1859, %f267;
shfl.sync.bfly.b32 %r1861|%p88, %r1859, %r1837, %r1829, %r1831;
mov.b32 %f268, %r1861;
add.ftz.f32 %f269, %f267, %f268;
mov.b32 %r1862, %f269;
shfl.sync.bfly.b32 %r1863|%p89, %r1862, %r1827, %r1829, %r1831;
mov.b32 %f270, %r1863;
add.ftz.f32 %f271, %f269, %f270;
mov.b32 %r1864, %f271;
shfl.sync.bfly.b32 %r1866|%p90, %r1864, %r1842, %r1829, %r1831;
mov.b32 %f272, %r1866;
add.ftz.f32 %f54, %f271, %f272;
@%p21 bra $L__BB235_30;

add.s32 %r1867, %r28, 1;
cvt.u64.u32 %rd124, %r1867;
add.s64 %rd125, %rd124, %rd13;
shl.b64 %rd127, %rd125, 2;
add.s64 %rd128, %rd103, %rd127;
st.global.f32 [%rd128], %f54;

$L__BB235_30:
mul.lo.s32 %r1868, %r31, 6;
mov.u32 %r1873, 2;
ld.shared.f32 %f273, [%r1590+1536];
add.ftz.f32 %f274, %f273, %f310;
mov.b32 %r1874, %f274;
mov.u32 %r1875, 31;
mov.u32 %r1876, 16;
mov.u32 %r1877, -1;
shfl.sync.bfly.b32 %r1878|%p92, %r1874, %r1876, %r1875, %r1877;
mov.b32 %f275, %r1878;
add.ftz.f32 %f276, %f274, %f275;
mov.b32 %r1879, %f276;
mov.u32 %r1880, 8;
shfl.sync.bfly.b32 %r1881|%p93, %r1879, %r1880, %r1875, %r1877;
mov.b32 %f277, %r1881;
add.ftz.f32 %f278, %f276, %f277;
mov.b32 %r1882, %f278;
mov.u32 %r1883, 4;
shfl.sync.bfly.b32 %r1884|%p94, %r1882, %r1883, %r1875, %r1877;
mov.b32 %f279, %r1884;
add.ftz.f32 %f280, %f278, %f279;
mov.b32 %r1885, %f280;
shfl.sync.bfly.b32 %r1886|%p95, %r1885, %r1873, %r1875, %r1877;
mov.b32 %f281, %r1886;
add.ftz.f32 %f282, %f280, %f281;
mov.b32 %r1887, %f282;
mov.u32 %r1888, 1;
shfl.sync.bfly.b32 %r1889|%p96, %r1887, %r1888, %r1875, %r1877;
mov.b32 %f283, %r1889;
add.ftz.f32 %f55, %f282, %f283;
cvt.u64.u32 %rd129, %r1868;
add.s64 %rd130, %rd129, %rd13;
shl.b64 %rd132, %rd130, 2;
add.s64 %rd16, %rd103, %rd132;
@%p13 bra $L__BB235_32;

st.global.f32 [%rd16], %f55;

$L__BB235_32:
ld.shared.f32 %f284, [%r1590+1664];
add.ftz.f32 %f285, %f284, %f309;
mov.b32 %r1897, %f285;
shfl.sync.bfly.b32 %r1901|%p100, %r1897, %r1876, %r1875, %r1877;
mov.b32 %f286, %r1901;
add.ftz.f32 %f287, %f285, %f286;
mov.b32 %r1902, %f287;
shfl.sync.bfly.b32 %r1904|%p101, %r1902, %r1880, %r1875, %r1877;
mov.b32 %f288, %r1904;
add.ftz.f32 %f289, %f287, %f288;
mov.b32 %r1905, %f289;
shfl.sync.bfly.b32 %r1907|%p102, %r1905, %r1883, %r1875, %r1877;
mov.b32 %f290, %r1907;
add.ftz.f32 %f291, %f289, %f290;
mov.b32 %r1908, %f291;
shfl.sync.bfly.b32 %r1909|%p103, %r1908, %r1873, %r1875, %r1877;
mov.b32 %f292, %r1909;
add.ftz.f32 %f293, %f291, %f292;
mov.b32 %r1910, %f293;
shfl.sync.bfly.b32 %r1912|%p104, %r1910, %r1888, %r1875, %r1877;
mov.b32 %f294, %r1912;
add.ftz.f32 %f56, %f293, %f294;
@%p21 bra $L__BB235_34;

st.global.f32 [%rd16+4], %f56;

$L__BB235_34:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type22ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type22ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type22ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type22ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type22ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type22ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type22ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type22ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type22ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type22ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<120>;
.reg .b16 %rs<55>;
.reg .f32 %f<369>;
.reg .b32 %r<2212>;
.reg .b64 %rd<159>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type22ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[2048];

ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type22ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type22ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type22ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r48, [_Z13mul_mat_vec_qIL9ggml_type22ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type22ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type22ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type22ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type22ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type22ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type22ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r45, [_Z13mul_mat_vec_qIL9ggml_type22ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r46, [_Z13mul_mat_vec_qIL9ggml_type22ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r47, [_Z13mul_mat_vec_qIL9ggml_type22ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type22ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type22ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type22ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type22ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r49, %tid.y;
shl.b32 %r50, %r49, 5;
mov.u32 %r51, %tid.x;
add.s32 %r1, %r50, %r51;
shr.s32 %r52, %r1, 31;
shr.u32 %r53, %r52, 29;
add.s32 %r54, %r1, %r53;
shr.s32 %r2211, %r54, 3;
shr.u32 %r3, %r48, 8;
setp.ge.s32 %p2, %r2211, %r3;
mov.f32 %f353, 0f00000000;
mov.f32 %f354, %f353;
mov.f32 %f355, %f353;
mov.f32 %f356, %f353;
mov.f32 %f357, %f353;
mov.f32 %f358, %f353;
mov.f32 %f359, %f353;
mov.f32 %f360, %f353;
mov.f32 %f361, %f353;
mov.f32 %f362, %f353;
mov.f32 %f363, %f353;
mov.f32 %f364, %f353;
mov.f32 %f365, %f353;
mov.f32 %f366, %f353;
mov.f32 %f367, %f353;
mov.f32 %f368, %f353;
@%p2 bra $L__BB236_3;

mov.u32 %r55, %ctaid.x;
shl.b32 %r56, %r55, 1;
mov.u32 %r57, %ctaid.y;
mov.u32 %r58, %ctaid.z;
mul.lo.s32 %r59, %r57, %r40;
cvt.u64.u32 %rd19, %r59;
mul.lo.s32 %r60, %r58, %r46;
cvt.u64.u32 %rd20, %r60;
add.s64 %rd1, %rd20, %rd19;
mul.hi.u32 %r61, %r58, %r42;
add.s32 %r62, %r61, %r58;
shr.u32 %r63, %r62, %r43;
mul.lo.s32 %r64, %r63, %r45;
mul.hi.u32 %r65, %r57, %r36;
add.s32 %r66, %r65, %r57;
shr.u32 %r67, %r66, %r37;
mul.lo.s32 %r68, %r67, %r39;
cvta.to.global.u64 %rd158, %rd16;
and.b32 %r72, %r54, -8;
sub.s32 %r73, %r1, %r72;
shl.b32 %r74, %r73, 1;
cvt.s64.s32 %rd3, %r73;
shl.b32 %r75, %r34, 1;
shl.b32 %r2203, %r2211, 3;
add.s32 %r2210, %r2203, %r75;
mad.lo.s32 %r2209, %r34, 3, %r2203;
shl.b32 %r76, %r34, 2;
add.s32 %r2208, %r2203, %r76;
mad.lo.s32 %r2207, %r34, 5, %r2203;
mad.lo.s32 %r2206, %r34, 6, %r2203;
mad.lo.s32 %r2205, %r34, 7, %r2203;
add.s32 %r2204, %r34, %r2203;
add.s32 %r77, %r2211, %r64;
add.s32 %r78, %r77, %r68;
or.b32 %r79, %r56, 1;
mad.lo.s32 %r80, %r33, %r79, %r78;
mul.wide.s32 %rd4, %r80, 82;
mad.lo.s32 %r81, %r56, %r33, %r2211;
add.s32 %r82, %r81, %r64;
add.s32 %r83, %r82, %r68;
mul.wide.s32 %rd5, %r83, 82;
mul.wide.s32 %rd21, %r74, 2;
add.s64 %rd22, %rd5, %rd21;
add.s64 %rd6, %rd22, 36;
add.s64 %rd23, %rd4, %rd3;
add.s64 %rd7, %rd23, 66;
add.s64 %rd24, %rd5, %rd3;
add.s64 %rd8, %rd24, 74;
add.s64 %rd25, %rd4, %rd21;
add.s64 %rd9, %rd25, 34;
cvta.to.global.u64 %rd10, %rd17;
mov.f32 %f353, 0f00000000;

$L__BB236_2:
cvt.u64.u32 %rd26, %r2203;
add.s64 %rd27, %rd1, %rd26;
add.s64 %rd28, %rd27, %rd3;
add.s64 %rd29, %rd158, %rd6;
ld.global.nc.u16 %rs17, [%rd29+-34];
shr.u16 %rs18, %rs17, 8;
cvt.u32.u16 %r868, %rs18;
ld.global.nc.u16 %rs19, [%rd29+-32];
shr.u16 %rs20, %rs19, 8;
cvt.u32.u16 %r869, %rs20;
add.s64 %rd30, %rd158, %rd8;
ld.global.nc.u8 %rs21, [%rd30+-8];
ld.global.nc.u16 %rs22, [%rd29+-2];
ld.global.nc.u16 %rs23, [%rd29];
ld.global.nc.u8 %rs24, [%rd30];
and.b16 %rs25, %rs24, 240;
and.b16 %rs26, %rs24, 15;
shr.u16 %rs27, %rs25, 4;
and.b16 %rs28, %rs17, 255;
cvt.u32.u16 %r870, %rs28;
shl.b16 %rs29, %rs21, 8;
cvt.u32.u16 %r871, %rs29;
and.b32 %r872, %r871, 768;
or.b32 %r873, %r872, %r870;
mul.wide.u32 %rd31, %r873, 8;
mov.u64 %rd32, iq2s_grid;
add.s64 %rd33, %rd32, %rd31;
and.b16 %rs30, %rs22, 255;
mul.wide.u16 %r874, %rs30, 128;
cvt.u32.u16 %r875, %rs22;
shl.b32 %r876, %r875, 21;
or.b32 %r877, %r874, %r876;
and.b32 %r878, %r877, 25166208;
or.b32 %r879, %r878, -2139062144;
add.s32 %r880, %r879, -16843009;
or.b32 %r770, %r880, %r878;
mov.u32 %r84, %r770;

prmt.b32 %r84, %r84, 0, 0xba98;

mul.wide.u16 %r881, %rs30, 8;
shl.b32 %r882, %r875, 17;
or.b32 %r883, %r881, %r882;
and.b32 %r884, %r883, 25166208;
or.b32 %r885, %r884, -2139062144;
add.s32 %r886, %r885, -16843009;
or.b32 %r772, %r886, %r884;
mov.u32 %r86, %r772;

prmt.b32 %r86, %r86, 0, 0xba98;

ld.global.nc.v2.u32 {%r887, %r888}, [%rd33];
not.b32 %r890, %r887;
mov.u32 %r850, 0;
not.b32 %r892, %r888;
xor.b32 %r893, %r887, %r84;
or.b32 %r894, %r893, -2139062144;
and.b32 %r895, %r84, 2139062143;
and.b32 %r896, %r890, -2139062144;
sub.s32 %r897, %r894, %r895;
xor.b32 %r89, %r896, %r897;
xor.b32 %r898, %r888, %r86;
or.b32 %r899, %r898, -2139062144;
and.b32 %r900, %r86, 2139062143;
and.b32 %r901, %r892, -2139062144;
sub.s32 %r902, %r899, %r900;
xor.b32 %r93, %r901, %r902;
mul.lo.s64 %rd34, %rd28, 36;
add.s64 %rd35, %rd10, %rd34;
ld.global.nc.u32 %r94, [%rd35+8];
ld.global.nc.u32 %r90, [%rd35+4];

dp4a.s32.s32 %r88, %r89, %r90, %r850;


dp4a.s32.s32 %r92, %r93, %r94, %r88;

mul.wide.u16 %r903, %rs21, 64;
and.b32 %r904, %r903, 768;
or.b32 %r905, %r904, %r868;
mul.wide.u32 %rd36, %r905, 8;
add.s64 %rd37, %rd32, %rd36;
shr.u16 %rs32, %rs22, 8;
mul.wide.u16 %r906, %rs32, 128;
cvt.u32.u16 %r907, %rs32;
shl.b32 %r908, %r907, 21;
or.b32 %r909, %r906, %r908;
and.b32 %r910, %r909, 25166208;
or.b32 %r911, %r910, -2139062144;
add.s32 %r912, %r911, -16843009;
or.b32 %r782, %r912, %r910;
mov.u32 %r96, %r782;

prmt.b32 %r96, %r96, 0, 0xba98;

mul.wide.u16 %r913, %rs32, 8;
shl.b32 %r914, %r907, 17;
or.b32 %r915, %r913, %r914;
and.b32 %r916, %r915, 25166208;
or.b32 %r917, %r916, -2139062144;
add.s32 %r918, %r917, -16843009;
or.b32 %r784, %r918, %r916;
mov.u32 %r98, %r784;

prmt.b32 %r98, %r98, 0, 0xba98;

ld.global.nc.v2.u32 {%r919, %r920}, [%rd37];
not.b32 %r922, %r919;
not.b32 %r924, %r920;
xor.b32 %r925, %r919, %r96;
or.b32 %r926, %r925, -2139062144;
and.b32 %r927, %r96, 2139062143;
and.b32 %r928, %r922, -2139062144;
sub.s32 %r929, %r926, %r927;
xor.b32 %r101, %r928, %r929;
xor.b32 %r930, %r920, %r98;
or.b32 %r931, %r930, -2139062144;
and.b32 %r932, %r98, 2139062143;
and.b32 %r933, %r924, -2139062144;
sub.s32 %r934, %r931, %r932;
xor.b32 %r105, %r933, %r934;
ld.global.nc.u32 %r106, [%rd35+16];
ld.global.nc.u32 %r102, [%rd35+12];

dp4a.s32.s32 %r100, %r101, %r102, %r92;


dp4a.s32.s32 %r104, %r105, %r106, %r100;

and.b16 %rs33, %rs19, 255;
cvt.u32.u16 %r935, %rs33;
mul.wide.u16 %r936, %rs21, 16;
and.b32 %r937, %r936, 768;
or.b32 %r938, %r937, %r935;
mul.wide.u32 %rd38, %r938, 8;
add.s64 %rd39, %rd32, %rd38;
and.b16 %rs34, %rs23, 255;
mul.wide.u16 %r939, %rs34, 128;
cvt.u32.u16 %r940, %rs23;
shl.b32 %r941, %r940, 21;
or.b32 %r942, %r939, %r941;
and.b32 %r943, %r942, 25166208;
or.b32 %r944, %r943, -2139062144;
add.s32 %r945, %r944, -16843009;
or.b32 %r794, %r945, %r943;
mov.u32 %r108, %r794;

prmt.b32 %r108, %r108, 0, 0xba98;

mul.wide.u16 %r946, %rs34, 8;
shl.b32 %r947, %r940, 17;
or.b32 %r948, %r946, %r947;
and.b32 %r949, %r948, 25166208;
or.b32 %r950, %r949, -2139062144;
add.s32 %r951, %r950, -16843009;
or.b32 %r796, %r951, %r949;
mov.u32 %r110, %r796;

prmt.b32 %r110, %r110, 0, 0xba98;

ld.global.nc.v2.u32 {%r952, %r953}, [%rd39];
not.b32 %r955, %r952;
not.b32 %r957, %r953;
xor.b32 %r958, %r952, %r108;
or.b32 %r959, %r958, -2139062144;
and.b32 %r960, %r108, 2139062143;
and.b32 %r961, %r955, -2139062144;
sub.s32 %r962, %r959, %r960;
xor.b32 %r113, %r961, %r962;
xor.b32 %r963, %r953, %r110;
or.b32 %r964, %r963, -2139062144;
and.b32 %r965, %r110, 2139062143;
and.b32 %r966, %r957, -2139062144;
sub.s32 %r967, %r964, %r965;
xor.b32 %r117, %r966, %r967;
ld.global.nc.u32 %r118, [%rd35+24];
ld.global.nc.u32 %r114, [%rd35+20];

dp4a.s32.s32 %r112, %r113, %r114, %r850;


dp4a.s32.s32 %r116, %r117, %r118, %r112;

mul.wide.u16 %r968, %rs21, 4;
and.b32 %r969, %r968, 768;
or.b32 %r970, %r969, %r869;
mul.wide.u32 %rd40, %r970, 8;
add.s64 %rd41, %rd32, %rd40;
shr.u16 %rs35, %rs23, 8;
mul.wide.u16 %r971, %rs35, 128;
cvt.u32.u16 %r972, %rs35;
shl.b32 %r973, %r972, 21;
or.b32 %r974, %r971, %r973;
and.b32 %r975, %r974, 25166208;
or.b32 %r976, %r975, -2139062144;
add.s32 %r977, %r976, -16843009;
or.b32 %r806, %r977, %r975;
mov.u32 %r120, %r806;

prmt.b32 %r120, %r120, 0, 0xba98;

mul.wide.u16 %r978, %rs35, 8;
shl.b32 %r979, %r972, 17;
or.b32 %r980, %r978, %r979;
and.b32 %r981, %r980, 25166208;
or.b32 %r982, %r981, -2139062144;
add.s32 %r983, %r982, -16843009;
or.b32 %r808, %r983, %r981;
mov.u32 %r122, %r808;

prmt.b32 %r122, %r122, 0, 0xba98;

ld.global.nc.v2.u32 {%r984, %r985}, [%rd41];
not.b32 %r987, %r984;
not.b32 %r989, %r985;
xor.b32 %r990, %r984, %r120;
or.b32 %r991, %r990, -2139062144;
and.b32 %r992, %r120, 2139062143;
and.b32 %r993, %r987, -2139062144;
sub.s32 %r994, %r991, %r992;
xor.b32 %r125, %r993, %r994;
xor.b32 %r995, %r985, %r122;
or.b32 %r996, %r995, -2139062144;
and.b32 %r997, %r122, 2139062143;
and.b32 %r998, %r989, -2139062144;
sub.s32 %r999, %r996, %r997;
xor.b32 %r129, %r998, %r999;
ld.global.nc.u32 %r130, [%rd35+32];
ld.global.nc.u32 %r126, [%rd35+28];

dp4a.s32.s32 %r124, %r125, %r126, %r116;


dp4a.s32.s32 %r128, %r129, %r130, %r124;

cvt.u32.u16 %r1000, %rs26;
cvt.u32.u16 %r1001, %rs27;
mul.lo.s32 %r1002, %r104, %r1000;
mad.lo.s32 %r1003, %r128, %r1001, %r1002;
add.s32 %r1004, %r128, %r104;
shr.u32 %r1005, %r1004, 31;
add.s32 %r1006, %r1004, %r1005;
shr.s32 %r1007, %r1006, 1;
add.s32 %r1008, %r1003, %r1007;
shr.s32 %r1009, %r1008, 31;
shr.u32 %r1010, %r1009, 30;
add.s32 %r1011, %r1008, %r1010;
shr.s32 %r1012, %r1011, 2;
add.s64 %rd42, %rd158, %rd5;
ld.global.nc.u16 %rs1, [%rd42];

{ cvt.f32.f16 %f97, %rs1;}


ld.global.nc.u32 %r132, [%rd35];

{.reg .f16 low,high;
mov.b32 {low,high},%r132;
cvt.f32.f16 %f98, low;}


mul.ftz.f32 %f129, %f97, %f98;
cvt.rn.f32.s32 %f130, %r1012;
fma.rn.ftz.f32 %f368, %f129, %f130, %f368;
add.s64 %rd43, %rd158, %rd9;
ld.global.nc.u16 %rs36, [%rd43+-32];
shr.u16 %rs37, %rs36, 8;
cvt.u32.u16 %r1013, %rs37;
ld.global.nc.u16 %rs38, [%rd43+-30];
shr.u16 %rs39, %rs38, 8;
cvt.u32.u16 %r1014, %rs39;
add.s64 %rd44, %rd158, %rd7;
ld.global.nc.u8 %rs40, [%rd44];
ld.global.nc.u16 %rs41, [%rd43];
ld.global.nc.u16 %rs42, [%rd43+2];
ld.global.nc.u8 %rs43, [%rd44+8];
and.b16 %rs44, %rs43, 240;
and.b16 %rs45, %rs43, 15;
shr.u16 %rs46, %rs44, 4;
and.b16 %rs47, %rs36, 255;
cvt.u32.u16 %r1015, %rs47;
shl.b16 %rs48, %rs40, 8;
cvt.u32.u16 %r1016, %rs48;
and.b32 %r1017, %r1016, 768;
or.b32 %r1018, %r1017, %r1015;
mul.wide.u32 %rd45, %r1018, 8;
add.s64 %rd46, %rd32, %rd45;
and.b16 %rs49, %rs41, 255;
mul.wide.u16 %r1019, %rs49, 128;
cvt.u32.u16 %r1020, %rs41;
shl.b32 %r1021, %r1020, 21;
or.b32 %r1022, %r1019, %r1021;
and.b32 %r1023, %r1022, 25166208;
or.b32 %r1024, %r1023, -2139062144;
add.s32 %r1025, %r1024, -16843009;
or.b32 %r819, %r1025, %r1023;
mov.u32 %r133, %r819;

prmt.b32 %r133, %r133, 0, 0xba98;

mul.wide.u16 %r1026, %rs49, 8;
shl.b32 %r1027, %r1020, 17;
or.b32 %r1028, %r1026, %r1027;
and.b32 %r1029, %r1028, 25166208;
or.b32 %r1030, %r1029, -2139062144;
add.s32 %r1031, %r1030, -16843009;
or.b32 %r821, %r1031, %r1029;
mov.u32 %r135, %r821;

prmt.b32 %r135, %r135, 0, 0xba98;

ld.global.nc.v2.u32 {%r1032, %r1033}, [%rd46];
not.b32 %r1035, %r1032;
not.b32 %r1037, %r1033;
xor.b32 %r1038, %r1032, %r133;
or.b32 %r1039, %r1038, -2139062144;
and.b32 %r1040, %r133, 2139062143;
and.b32 %r1041, %r1035, -2139062144;
sub.s32 %r1042, %r1039, %r1040;
xor.b32 %r138, %r1041, %r1042;
xor.b32 %r1043, %r1033, %r135;
or.b32 %r1044, %r1043, -2139062144;
and.b32 %r1045, %r135, 2139062143;
and.b32 %r1046, %r1037, -2139062144;
sub.s32 %r1047, %r1044, %r1045;
xor.b32 %r142, %r1046, %r1047;

dp4a.s32.s32 %r137, %r138, %r90, %r850;


dp4a.s32.s32 %r141, %r142, %r94, %r137;

mul.wide.u16 %r1048, %rs40, 64;
and.b32 %r1049, %r1048, 768;
or.b32 %r1050, %r1049, %r1013;
mul.wide.u32 %rd47, %r1050, 8;
add.s64 %rd48, %rd32, %rd47;
shr.u16 %rs51, %rs41, 8;
mul.wide.u16 %r1051, %rs51, 128;
cvt.u32.u16 %r1052, %rs51;
shl.b32 %r1053, %r1052, 21;
or.b32 %r1054, %r1051, %r1053;
and.b32 %r1055, %r1054, 25166208;
or.b32 %r1056, %r1055, -2139062144;
add.s32 %r1057, %r1056, -16843009;
or.b32 %r831, %r1057, %r1055;
mov.u32 %r145, %r831;

prmt.b32 %r145, %r145, 0, 0xba98;

mul.wide.u16 %r1058, %rs51, 8;
shl.b32 %r1059, %r1052, 17;
or.b32 %r1060, %r1058, %r1059;
and.b32 %r1061, %r1060, 25166208;
or.b32 %r1062, %r1061, -2139062144;
add.s32 %r1063, %r1062, -16843009;
or.b32 %r833, %r1063, %r1061;
mov.u32 %r147, %r833;

prmt.b32 %r147, %r147, 0, 0xba98;

ld.global.nc.v2.u32 {%r1064, %r1065}, [%rd48];
not.b32 %r1067, %r1064;
not.b32 %r1069, %r1065;
xor.b32 %r1070, %r1064, %r145;
or.b32 %r1071, %r1070, -2139062144;
and.b32 %r1072, %r145, 2139062143;
and.b32 %r1073, %r1067, -2139062144;
sub.s32 %r1074, %r1071, %r1072;
xor.b32 %r150, %r1073, %r1074;
xor.b32 %r1075, %r1065, %r147;
or.b32 %r1076, %r1075, -2139062144;
and.b32 %r1077, %r147, 2139062143;
and.b32 %r1078, %r1069, -2139062144;
sub.s32 %r1079, %r1076, %r1077;
xor.b32 %r154, %r1078, %r1079;

dp4a.s32.s32 %r149, %r150, %r102, %r141;


dp4a.s32.s32 %r153, %r154, %r106, %r149;

and.b16 %rs52, %rs38, 255;
cvt.u32.u16 %r1080, %rs52;
mul.wide.u16 %r1081, %rs40, 16;
and.b32 %r1082, %r1081, 768;
or.b32 %r1083, %r1082, %r1080;
mul.wide.u32 %rd49, %r1083, 8;
add.s64 %rd50, %rd32, %rd49;
and.b16 %rs53, %rs42, 255;
mul.wide.u16 %r1084, %rs53, 128;
cvt.u32.u16 %r1085, %rs42;
shl.b32 %r1086, %r1085, 21;
or.b32 %r1087, %r1084, %r1086;
and.b32 %r1088, %r1087, 25166208;
or.b32 %r1089, %r1088, -2139062144;
add.s32 %r1090, %r1089, -16843009;
or.b32 %r843, %r1090, %r1088;
mov.u32 %r157, %r843;

prmt.b32 %r157, %r157, 0, 0xba98;

mul.wide.u16 %r1091, %rs53, 8;
shl.b32 %r1092, %r1085, 17;
or.b32 %r1093, %r1091, %r1092;
and.b32 %r1094, %r1093, 25166208;
or.b32 %r1095, %r1094, -2139062144;
add.s32 %r1096, %r1095, -16843009;
or.b32 %r845, %r1096, %r1094;
mov.u32 %r159, %r845;

prmt.b32 %r159, %r159, 0, 0xba98;

ld.global.nc.v2.u32 {%r1097, %r1098}, [%rd50];
not.b32 %r1100, %r1097;
not.b32 %r1102, %r1098;
xor.b32 %r1103, %r1097, %r157;
or.b32 %r1104, %r1103, -2139062144;
and.b32 %r1105, %r157, 2139062143;
and.b32 %r1106, %r1100, -2139062144;
sub.s32 %r1107, %r1104, %r1105;
xor.b32 %r162, %r1106, %r1107;
xor.b32 %r1108, %r1098, %r159;
or.b32 %r1109, %r1108, -2139062144;
and.b32 %r1110, %r159, 2139062143;
and.b32 %r1111, %r1102, -2139062144;
sub.s32 %r1112, %r1109, %r1110;
xor.b32 %r166, %r1111, %r1112;

dp4a.s32.s32 %r161, %r162, %r114, %r850;


dp4a.s32.s32 %r165, %r166, %r118, %r161;

mul.wide.u16 %r1113, %rs40, 4;
and.b32 %r1114, %r1113, 768;
or.b32 %r1115, %r1114, %r1014;
mul.wide.u32 %rd51, %r1115, 8;
add.s64 %rd52, %rd32, %rd51;
shr.u16 %rs54, %rs42, 8;
mul.wide.u16 %r1116, %rs54, 128;
cvt.u32.u16 %r1117, %rs54;
shl.b32 %r1118, %r1117, 21;
or.b32 %r1119, %r1116, %r1118;
and.b32 %r1120, %r1119, 25166208;
or.b32 %r1121, %r1120, -2139062144;
add.s32 %r1122, %r1121, -16843009;
or.b32 %r855, %r1122, %r1120;
mov.u32 %r169, %r855;

prmt.b32 %r169, %r169, 0, 0xba98;

mul.wide.u16 %r1123, %rs54, 8;
shl.b32 %r1124, %r1117, 17;
or.b32 %r1125, %r1123, %r1124;
and.b32 %r1126, %r1125, 25166208;
or.b32 %r1127, %r1126, -2139062144;
add.s32 %r1128, %r1127, -16843009;
or.b32 %r857, %r1128, %r1126;
mov.u32 %r171, %r857;

prmt.b32 %r171, %r171, 0, 0xba98;

ld.global.nc.v2.u32 {%r1129, %r1130}, [%rd52];
not.b32 %r1132, %r1129;
not.b32 %r1134, %r1130;
xor.b32 %r1135, %r1129, %r169;
or.b32 %r1136, %r1135, -2139062144;
and.b32 %r1137, %r169, 2139062143;
and.b32 %r1138, %r1132, -2139062144;
sub.s32 %r1139, %r1136, %r1137;
xor.b32 %r174, %r1138, %r1139;
xor.b32 %r1140, %r1130, %r171;
or.b32 %r1141, %r1140, -2139062144;
and.b32 %r1142, %r171, 2139062143;
and.b32 %r1143, %r1134, -2139062144;
sub.s32 %r1144, %r1141, %r1142;
xor.b32 %r178, %r1143, %r1144;

dp4a.s32.s32 %r173, %r174, %r126, %r165;


dp4a.s32.s32 %r177, %r178, %r130, %r173;

cvt.u32.u16 %r1145, %rs45;
cvt.u32.u16 %r1146, %rs46;
mul.lo.s32 %r1147, %r153, %r1145;
mad.lo.s32 %r1148, %r177, %r1146, %r1147;
add.s32 %r1149, %r177, %r153;
shr.u32 %r1150, %r1149, 31;
add.s32 %r1151, %r1149, %r1150;
shr.s32 %r1152, %r1151, 1;
add.s32 %r1153, %r1148, %r1152;
shr.s32 %r1154, %r1153, 31;
shr.u32 %r1155, %r1154, 30;
add.s32 %r1156, %r1153, %r1155;
shr.s32 %r1157, %r1156, 2;
add.s64 %rd53, %rd158, %rd4;
ld.global.nc.u16 %rs2, [%rd53];

{ cvt.f32.f16 %f99, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r132;
cvt.f32.f16 %f100, low;}


mul.ftz.f32 %f131, %f99, %f100;
cvt.rn.f32.s32 %f132, %r1157;
fma.rn.ftz.f32 %f367, %f131, %f132, %f367;
cvt.u64.u32 %rd54, %r2204;
add.s64 %rd55, %rd1, %rd54;
add.s64 %rd56, %rd55, %rd3;
mov.u32 %r182, %r770;

prmt.b32 %r182, %r182, 0, 0xba98;

mov.u32 %r184, %r772;

prmt.b32 %r184, %r184, 0, 0xba98;

xor.b32 %r1158, %r887, %r182;
or.b32 %r1159, %r1158, -2139062144;
and.b32 %r1160, %r182, 2139062143;
sub.s32 %r1161, %r1159, %r1160;
xor.b32 %r187, %r896, %r1161;
xor.b32 %r1162, %r888, %r184;
or.b32 %r1163, %r1162, -2139062144;
and.b32 %r1164, %r184, 2139062143;
sub.s32 %r1165, %r1163, %r1164;
xor.b32 %r191, %r901, %r1165;
mul.lo.s64 %rd57, %rd56, 36;
add.s64 %rd58, %rd10, %rd57;
ld.global.nc.u32 %r192, [%rd58+8];
ld.global.nc.u32 %r188, [%rd58+4];

dp4a.s32.s32 %r186, %r187, %r188, %r850;


dp4a.s32.s32 %r190, %r191, %r192, %r186;

mov.u32 %r194, %r782;

prmt.b32 %r194, %r194, 0, 0xba98;

mov.u32 %r196, %r784;

prmt.b32 %r196, %r196, 0, 0xba98;

xor.b32 %r1166, %r919, %r194;
or.b32 %r1167, %r1166, -2139062144;
and.b32 %r1168, %r194, 2139062143;
sub.s32 %r1169, %r1167, %r1168;
xor.b32 %r199, %r928, %r1169;
xor.b32 %r1170, %r920, %r196;
or.b32 %r1171, %r1170, -2139062144;
and.b32 %r1172, %r196, 2139062143;
sub.s32 %r1173, %r1171, %r1172;
xor.b32 %r203, %r933, %r1173;
ld.global.nc.u32 %r204, [%rd58+16];
ld.global.nc.u32 %r200, [%rd58+12];

dp4a.s32.s32 %r198, %r199, %r200, %r190;


dp4a.s32.s32 %r202, %r203, %r204, %r198;

mov.u32 %r206, %r794;

prmt.b32 %r206, %r206, 0, 0xba98;

mov.u32 %r208, %r796;

prmt.b32 %r208, %r208, 0, 0xba98;

xor.b32 %r1174, %r952, %r206;
or.b32 %r1175, %r1174, -2139062144;
and.b32 %r1176, %r206, 2139062143;
sub.s32 %r1177, %r1175, %r1176;
xor.b32 %r211, %r961, %r1177;
xor.b32 %r1178, %r953, %r208;
or.b32 %r1179, %r1178, -2139062144;
and.b32 %r1180, %r208, 2139062143;
sub.s32 %r1181, %r1179, %r1180;
xor.b32 %r215, %r966, %r1181;
ld.global.nc.u32 %r216, [%rd58+24];
ld.global.nc.u32 %r212, [%rd58+20];

dp4a.s32.s32 %r210, %r211, %r212, %r850;


dp4a.s32.s32 %r214, %r215, %r216, %r210;

mov.u32 %r218, %r806;

prmt.b32 %r218, %r218, 0, 0xba98;

mov.u32 %r220, %r808;

prmt.b32 %r220, %r220, 0, 0xba98;

xor.b32 %r1182, %r984, %r218;
or.b32 %r1183, %r1182, -2139062144;
and.b32 %r1184, %r218, 2139062143;
sub.s32 %r1185, %r1183, %r1184;
xor.b32 %r223, %r993, %r1185;
xor.b32 %r1186, %r985, %r220;
or.b32 %r1187, %r1186, -2139062144;
and.b32 %r1188, %r220, 2139062143;
sub.s32 %r1189, %r1187, %r1188;
xor.b32 %r227, %r998, %r1189;
ld.global.nc.u32 %r228, [%rd58+32];
ld.global.nc.u32 %r224, [%rd58+28];

dp4a.s32.s32 %r222, %r223, %r224, %r214;


dp4a.s32.s32 %r226, %r227, %r228, %r222;

mul.lo.s32 %r1190, %r202, %r1000;
mad.lo.s32 %r1191, %r226, %r1001, %r1190;
add.s32 %r1192, %r226, %r202;
shr.u32 %r1193, %r1192, 31;
add.s32 %r1194, %r1192, %r1193;
shr.s32 %r1195, %r1194, 1;
add.s32 %r1196, %r1191, %r1195;
shr.s32 %r1197, %r1196, 31;
shr.u32 %r1198, %r1197, 30;
add.s32 %r1199, %r1196, %r1198;
shr.s32 %r1200, %r1199, 2;

{ cvt.f32.f16 %f101, %rs1;}


ld.global.nc.u32 %r230, [%rd58];

{.reg .f16 low,high;
mov.b32 {low,high},%r230;
cvt.f32.f16 %f102, low;}


mul.ftz.f32 %f133, %f101, %f102;
cvt.rn.f32.s32 %f134, %r1200;
fma.rn.ftz.f32 %f366, %f133, %f134, %f366;
mov.u32 %r231, %r819;

prmt.b32 %r231, %r231, 0, 0xba98;

mov.u32 %r233, %r821;

prmt.b32 %r233, %r233, 0, 0xba98;

xor.b32 %r1201, %r1032, %r231;
or.b32 %r1202, %r1201, -2139062144;
and.b32 %r1203, %r231, 2139062143;
sub.s32 %r1204, %r1202, %r1203;
xor.b32 %r236, %r1041, %r1204;
xor.b32 %r1205, %r1033, %r233;
or.b32 %r1206, %r1205, -2139062144;
and.b32 %r1207, %r233, 2139062143;
sub.s32 %r1208, %r1206, %r1207;
xor.b32 %r240, %r1046, %r1208;

dp4a.s32.s32 %r235, %r236, %r188, %r850;


dp4a.s32.s32 %r239, %r240, %r192, %r235;

mov.u32 %r243, %r831;

prmt.b32 %r243, %r243, 0, 0xba98;

mov.u32 %r245, %r833;

prmt.b32 %r245, %r245, 0, 0xba98;

xor.b32 %r1209, %r1064, %r243;
or.b32 %r1210, %r1209, -2139062144;
and.b32 %r1211, %r243, 2139062143;
sub.s32 %r1212, %r1210, %r1211;
xor.b32 %r248, %r1073, %r1212;
xor.b32 %r1213, %r1065, %r245;
or.b32 %r1214, %r1213, -2139062144;
and.b32 %r1215, %r245, 2139062143;
sub.s32 %r1216, %r1214, %r1215;
xor.b32 %r252, %r1078, %r1216;

dp4a.s32.s32 %r247, %r248, %r200, %r239;


dp4a.s32.s32 %r251, %r252, %r204, %r247;

mov.u32 %r255, %r843;

prmt.b32 %r255, %r255, 0, 0xba98;

mov.u32 %r257, %r845;

prmt.b32 %r257, %r257, 0, 0xba98;

xor.b32 %r1217, %r1097, %r255;
or.b32 %r1218, %r1217, -2139062144;
and.b32 %r1219, %r255, 2139062143;
sub.s32 %r1220, %r1218, %r1219;
xor.b32 %r260, %r1106, %r1220;
xor.b32 %r1221, %r1098, %r257;
or.b32 %r1222, %r1221, -2139062144;
and.b32 %r1223, %r257, 2139062143;
sub.s32 %r1224, %r1222, %r1223;
xor.b32 %r264, %r1111, %r1224;

dp4a.s32.s32 %r259, %r260, %r212, %r850;


dp4a.s32.s32 %r263, %r264, %r216, %r259;

mov.u32 %r267, %r855;

prmt.b32 %r267, %r267, 0, 0xba98;

mov.u32 %r269, %r857;

prmt.b32 %r269, %r269, 0, 0xba98;

xor.b32 %r1225, %r1129, %r267;
or.b32 %r1226, %r1225, -2139062144;
and.b32 %r1227, %r267, 2139062143;
sub.s32 %r1228, %r1226, %r1227;
xor.b32 %r272, %r1138, %r1228;
xor.b32 %r1229, %r1130, %r269;
or.b32 %r1230, %r1229, -2139062144;
and.b32 %r1231, %r269, 2139062143;
sub.s32 %r1232, %r1230, %r1231;
xor.b32 %r276, %r1143, %r1232;

dp4a.s32.s32 %r271, %r272, %r224, %r263;


dp4a.s32.s32 %r275, %r276, %r228, %r271;

mul.lo.s32 %r1233, %r251, %r1145;
mad.lo.s32 %r1234, %r275, %r1146, %r1233;
add.s32 %r1235, %r275, %r251;
shr.u32 %r1236, %r1235, 31;
add.s32 %r1237, %r1235, %r1236;
shr.s32 %r1238, %r1237, 1;
add.s32 %r1239, %r1234, %r1238;
shr.s32 %r1240, %r1239, 31;
shr.u32 %r1241, %r1240, 30;
add.s32 %r1242, %r1239, %r1241;
shr.s32 %r1243, %r1242, 2;

{ cvt.f32.f16 %f103, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r230;
cvt.f32.f16 %f104, low;}


mul.ftz.f32 %f135, %f103, %f104;
cvt.rn.f32.s32 %f136, %r1243;
fma.rn.ftz.f32 %f365, %f135, %f136, %f365;
cvt.u64.u32 %rd59, %r2210;
add.s64 %rd60, %rd1, %rd59;
add.s64 %rd61, %rd60, %rd3;
mov.u32 %r280, %r770;

prmt.b32 %r280, %r280, 0, 0xba98;

mov.u32 %r282, %r772;

prmt.b32 %r282, %r282, 0, 0xba98;

xor.b32 %r1244, %r887, %r280;
or.b32 %r1245, %r1244, -2139062144;
and.b32 %r1246, %r280, 2139062143;
sub.s32 %r1247, %r1245, %r1246;
xor.b32 %r285, %r896, %r1247;
xor.b32 %r1248, %r888, %r282;
or.b32 %r1249, %r1248, -2139062144;
and.b32 %r1250, %r282, 2139062143;
sub.s32 %r1251, %r1249, %r1250;
xor.b32 %r289, %r901, %r1251;
mul.lo.s64 %rd62, %rd61, 36;
add.s64 %rd63, %rd10, %rd62;
ld.global.nc.u32 %r290, [%rd63+8];
ld.global.nc.u32 %r286, [%rd63+4];

dp4a.s32.s32 %r284, %r285, %r286, %r850;


dp4a.s32.s32 %r288, %r289, %r290, %r284;

mov.u32 %r292, %r782;

prmt.b32 %r292, %r292, 0, 0xba98;

mov.u32 %r294, %r784;

prmt.b32 %r294, %r294, 0, 0xba98;

xor.b32 %r1252, %r919, %r292;
or.b32 %r1253, %r1252, -2139062144;
and.b32 %r1254, %r292, 2139062143;
sub.s32 %r1255, %r1253, %r1254;
xor.b32 %r297, %r928, %r1255;
xor.b32 %r1256, %r920, %r294;
or.b32 %r1257, %r1256, -2139062144;
and.b32 %r1258, %r294, 2139062143;
sub.s32 %r1259, %r1257, %r1258;
xor.b32 %r301, %r933, %r1259;
ld.global.nc.u32 %r302, [%rd63+16];
ld.global.nc.u32 %r298, [%rd63+12];

dp4a.s32.s32 %r296, %r297, %r298, %r288;


dp4a.s32.s32 %r300, %r301, %r302, %r296;

mov.u32 %r304, %r794;

prmt.b32 %r304, %r304, 0, 0xba98;

mov.u32 %r306, %r796;

prmt.b32 %r306, %r306, 0, 0xba98;

xor.b32 %r1260, %r952, %r304;
or.b32 %r1261, %r1260, -2139062144;
and.b32 %r1262, %r304, 2139062143;
sub.s32 %r1263, %r1261, %r1262;
xor.b32 %r309, %r961, %r1263;
xor.b32 %r1264, %r953, %r306;
or.b32 %r1265, %r1264, -2139062144;
and.b32 %r1266, %r306, 2139062143;
sub.s32 %r1267, %r1265, %r1266;
xor.b32 %r313, %r966, %r1267;
ld.global.nc.u32 %r314, [%rd63+24];
ld.global.nc.u32 %r310, [%rd63+20];

dp4a.s32.s32 %r308, %r309, %r310, %r850;


dp4a.s32.s32 %r312, %r313, %r314, %r308;

mov.u32 %r316, %r806;

prmt.b32 %r316, %r316, 0, 0xba98;

mov.u32 %r318, %r808;

prmt.b32 %r318, %r318, 0, 0xba98;

xor.b32 %r1268, %r984, %r316;
or.b32 %r1269, %r1268, -2139062144;
and.b32 %r1270, %r316, 2139062143;
sub.s32 %r1271, %r1269, %r1270;
xor.b32 %r321, %r993, %r1271;
xor.b32 %r1272, %r985, %r318;
or.b32 %r1273, %r1272, -2139062144;
and.b32 %r1274, %r318, 2139062143;
sub.s32 %r1275, %r1273, %r1274;
xor.b32 %r325, %r998, %r1275;
ld.global.nc.u32 %r326, [%rd63+32];
ld.global.nc.u32 %r322, [%rd63+28];

dp4a.s32.s32 %r320, %r321, %r322, %r312;


dp4a.s32.s32 %r324, %r325, %r326, %r320;

mul.lo.s32 %r1276, %r300, %r1000;
mad.lo.s32 %r1277, %r324, %r1001, %r1276;
add.s32 %r1278, %r324, %r300;
shr.u32 %r1279, %r1278, 31;
add.s32 %r1280, %r1278, %r1279;
shr.s32 %r1281, %r1280, 1;
add.s32 %r1282, %r1277, %r1281;
shr.s32 %r1283, %r1282, 31;
shr.u32 %r1284, %r1283, 30;
add.s32 %r1285, %r1282, %r1284;
shr.s32 %r1286, %r1285, 2;

{ cvt.f32.f16 %f105, %rs1;}


ld.global.nc.u32 %r328, [%rd63];

{.reg .f16 low,high;
mov.b32 {low,high},%r328;
cvt.f32.f16 %f106, low;}


mul.ftz.f32 %f137, %f105, %f106;
cvt.rn.f32.s32 %f138, %r1286;
fma.rn.ftz.f32 %f364, %f137, %f138, %f364;
mov.u32 %r329, %r819;

prmt.b32 %r329, %r329, 0, 0xba98;

mov.u32 %r331, %r821;

prmt.b32 %r331, %r331, 0, 0xba98;

xor.b32 %r1287, %r1032, %r329;
or.b32 %r1288, %r1287, -2139062144;
and.b32 %r1289, %r329, 2139062143;
sub.s32 %r1290, %r1288, %r1289;
xor.b32 %r334, %r1041, %r1290;
xor.b32 %r1291, %r1033, %r331;
or.b32 %r1292, %r1291, -2139062144;
and.b32 %r1293, %r331, 2139062143;
sub.s32 %r1294, %r1292, %r1293;
xor.b32 %r338, %r1046, %r1294;

dp4a.s32.s32 %r333, %r334, %r286, %r850;


dp4a.s32.s32 %r337, %r338, %r290, %r333;

mov.u32 %r341, %r831;

prmt.b32 %r341, %r341, 0, 0xba98;

mov.u32 %r343, %r833;

prmt.b32 %r343, %r343, 0, 0xba98;

xor.b32 %r1295, %r1064, %r341;
or.b32 %r1296, %r1295, -2139062144;
and.b32 %r1297, %r341, 2139062143;
sub.s32 %r1298, %r1296, %r1297;
xor.b32 %r346, %r1073, %r1298;
xor.b32 %r1299, %r1065, %r343;
or.b32 %r1300, %r1299, -2139062144;
and.b32 %r1301, %r343, 2139062143;
sub.s32 %r1302, %r1300, %r1301;
xor.b32 %r350, %r1078, %r1302;

dp4a.s32.s32 %r345, %r346, %r298, %r337;


dp4a.s32.s32 %r349, %r350, %r302, %r345;

mov.u32 %r353, %r843;

prmt.b32 %r353, %r353, 0, 0xba98;

mov.u32 %r355, %r845;

prmt.b32 %r355, %r355, 0, 0xba98;

xor.b32 %r1303, %r1097, %r353;
or.b32 %r1304, %r1303, -2139062144;
and.b32 %r1305, %r353, 2139062143;
sub.s32 %r1306, %r1304, %r1305;
xor.b32 %r358, %r1106, %r1306;
xor.b32 %r1307, %r1098, %r355;
or.b32 %r1308, %r1307, -2139062144;
and.b32 %r1309, %r355, 2139062143;
sub.s32 %r1310, %r1308, %r1309;
xor.b32 %r362, %r1111, %r1310;

dp4a.s32.s32 %r357, %r358, %r310, %r850;


dp4a.s32.s32 %r361, %r362, %r314, %r357;

mov.u32 %r365, %r855;

prmt.b32 %r365, %r365, 0, 0xba98;

mov.u32 %r367, %r857;

prmt.b32 %r367, %r367, 0, 0xba98;

xor.b32 %r1311, %r1129, %r365;
or.b32 %r1312, %r1311, -2139062144;
and.b32 %r1313, %r365, 2139062143;
sub.s32 %r1314, %r1312, %r1313;
xor.b32 %r370, %r1138, %r1314;
xor.b32 %r1315, %r1130, %r367;
or.b32 %r1316, %r1315, -2139062144;
and.b32 %r1317, %r367, 2139062143;
sub.s32 %r1318, %r1316, %r1317;
xor.b32 %r374, %r1143, %r1318;

dp4a.s32.s32 %r369, %r370, %r322, %r361;


dp4a.s32.s32 %r373, %r374, %r326, %r369;

mul.lo.s32 %r1319, %r349, %r1145;
mad.lo.s32 %r1320, %r373, %r1146, %r1319;
add.s32 %r1321, %r373, %r349;
shr.u32 %r1322, %r1321, 31;
add.s32 %r1323, %r1321, %r1322;
shr.s32 %r1324, %r1323, 1;
add.s32 %r1325, %r1320, %r1324;
shr.s32 %r1326, %r1325, 31;
shr.u32 %r1327, %r1326, 30;
add.s32 %r1328, %r1325, %r1327;
shr.s32 %r1329, %r1328, 2;

{ cvt.f32.f16 %f107, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r328;
cvt.f32.f16 %f108, low;}


mul.ftz.f32 %f139, %f107, %f108;
cvt.rn.f32.s32 %f140, %r1329;
fma.rn.ftz.f32 %f363, %f139, %f140, %f363;
cvt.u64.u32 %rd64, %r2209;
add.s64 %rd65, %rd1, %rd64;
add.s64 %rd66, %rd65, %rd3;
mov.u32 %r378, %r770;

prmt.b32 %r378, %r378, 0, 0xba98;

mov.u32 %r380, %r772;

prmt.b32 %r380, %r380, 0, 0xba98;

xor.b32 %r1330, %r887, %r378;
or.b32 %r1331, %r1330, -2139062144;
and.b32 %r1332, %r378, 2139062143;
sub.s32 %r1333, %r1331, %r1332;
xor.b32 %r383, %r896, %r1333;
xor.b32 %r1334, %r888, %r380;
or.b32 %r1335, %r1334, -2139062144;
and.b32 %r1336, %r380, 2139062143;
sub.s32 %r1337, %r1335, %r1336;
xor.b32 %r387, %r901, %r1337;
mul.lo.s64 %rd67, %rd66, 36;
add.s64 %rd68, %rd10, %rd67;
ld.global.nc.u32 %r388, [%rd68+8];
ld.global.nc.u32 %r384, [%rd68+4];

dp4a.s32.s32 %r382, %r383, %r384, %r850;


dp4a.s32.s32 %r386, %r387, %r388, %r382;

mov.u32 %r390, %r782;

prmt.b32 %r390, %r390, 0, 0xba98;

mov.u32 %r392, %r784;

prmt.b32 %r392, %r392, 0, 0xba98;

xor.b32 %r1338, %r919, %r390;
or.b32 %r1339, %r1338, -2139062144;
and.b32 %r1340, %r390, 2139062143;
sub.s32 %r1341, %r1339, %r1340;
xor.b32 %r395, %r928, %r1341;
xor.b32 %r1342, %r920, %r392;
or.b32 %r1343, %r1342, -2139062144;
and.b32 %r1344, %r392, 2139062143;
sub.s32 %r1345, %r1343, %r1344;
xor.b32 %r399, %r933, %r1345;
ld.global.nc.u32 %r400, [%rd68+16];
ld.global.nc.u32 %r396, [%rd68+12];

dp4a.s32.s32 %r394, %r395, %r396, %r386;


dp4a.s32.s32 %r398, %r399, %r400, %r394;

mov.u32 %r402, %r794;

prmt.b32 %r402, %r402, 0, 0xba98;

mov.u32 %r404, %r796;

prmt.b32 %r404, %r404, 0, 0xba98;

xor.b32 %r1346, %r952, %r402;
or.b32 %r1347, %r1346, -2139062144;
and.b32 %r1348, %r402, 2139062143;
sub.s32 %r1349, %r1347, %r1348;
xor.b32 %r407, %r961, %r1349;
xor.b32 %r1350, %r953, %r404;
or.b32 %r1351, %r1350, -2139062144;
and.b32 %r1352, %r404, 2139062143;
sub.s32 %r1353, %r1351, %r1352;
xor.b32 %r411, %r966, %r1353;
ld.global.nc.u32 %r412, [%rd68+24];
ld.global.nc.u32 %r408, [%rd68+20];

dp4a.s32.s32 %r406, %r407, %r408, %r850;


dp4a.s32.s32 %r410, %r411, %r412, %r406;

mov.u32 %r414, %r806;

prmt.b32 %r414, %r414, 0, 0xba98;

mov.u32 %r416, %r808;

prmt.b32 %r416, %r416, 0, 0xba98;

xor.b32 %r1354, %r984, %r414;
or.b32 %r1355, %r1354, -2139062144;
and.b32 %r1356, %r414, 2139062143;
sub.s32 %r1357, %r1355, %r1356;
xor.b32 %r419, %r993, %r1357;
xor.b32 %r1358, %r985, %r416;
or.b32 %r1359, %r1358, -2139062144;
and.b32 %r1360, %r416, 2139062143;
sub.s32 %r1361, %r1359, %r1360;
xor.b32 %r423, %r998, %r1361;
ld.global.nc.u32 %r424, [%rd68+32];
ld.global.nc.u32 %r420, [%rd68+28];

dp4a.s32.s32 %r418, %r419, %r420, %r410;


dp4a.s32.s32 %r422, %r423, %r424, %r418;

mul.lo.s32 %r1362, %r398, %r1000;
mad.lo.s32 %r1363, %r422, %r1001, %r1362;
add.s32 %r1364, %r422, %r398;
shr.u32 %r1365, %r1364, 31;
add.s32 %r1366, %r1364, %r1365;
shr.s32 %r1367, %r1366, 1;
add.s32 %r1368, %r1363, %r1367;
shr.s32 %r1369, %r1368, 31;
shr.u32 %r1370, %r1369, 30;
add.s32 %r1371, %r1368, %r1370;
shr.s32 %r1372, %r1371, 2;

{ cvt.f32.f16 %f109, %rs1;}


ld.global.nc.u32 %r426, [%rd68];

{.reg .f16 low,high;
mov.b32 {low,high},%r426;
cvt.f32.f16 %f110, low;}


mul.ftz.f32 %f141, %f109, %f110;
cvt.rn.f32.s32 %f142, %r1372;
fma.rn.ftz.f32 %f362, %f141, %f142, %f362;
mov.u32 %r427, %r819;

prmt.b32 %r427, %r427, 0, 0xba98;

mov.u32 %r429, %r821;

prmt.b32 %r429, %r429, 0, 0xba98;

xor.b32 %r1373, %r1032, %r427;
or.b32 %r1374, %r1373, -2139062144;
and.b32 %r1375, %r427, 2139062143;
sub.s32 %r1376, %r1374, %r1375;
xor.b32 %r432, %r1041, %r1376;
xor.b32 %r1377, %r1033, %r429;
or.b32 %r1378, %r1377, -2139062144;
and.b32 %r1379, %r429, 2139062143;
sub.s32 %r1380, %r1378, %r1379;
xor.b32 %r436, %r1046, %r1380;

dp4a.s32.s32 %r431, %r432, %r384, %r850;


dp4a.s32.s32 %r435, %r436, %r388, %r431;

mov.u32 %r439, %r831;

prmt.b32 %r439, %r439, 0, 0xba98;

mov.u32 %r441, %r833;

prmt.b32 %r441, %r441, 0, 0xba98;

xor.b32 %r1381, %r1064, %r439;
or.b32 %r1382, %r1381, -2139062144;
and.b32 %r1383, %r439, 2139062143;
sub.s32 %r1384, %r1382, %r1383;
xor.b32 %r444, %r1073, %r1384;
xor.b32 %r1385, %r1065, %r441;
or.b32 %r1386, %r1385, -2139062144;
and.b32 %r1387, %r441, 2139062143;
sub.s32 %r1388, %r1386, %r1387;
xor.b32 %r448, %r1078, %r1388;

dp4a.s32.s32 %r443, %r444, %r396, %r435;


dp4a.s32.s32 %r447, %r448, %r400, %r443;

mov.u32 %r451, %r843;

prmt.b32 %r451, %r451, 0, 0xba98;

mov.u32 %r453, %r845;

prmt.b32 %r453, %r453, 0, 0xba98;

xor.b32 %r1389, %r1097, %r451;
or.b32 %r1390, %r1389, -2139062144;
and.b32 %r1391, %r451, 2139062143;
sub.s32 %r1392, %r1390, %r1391;
xor.b32 %r456, %r1106, %r1392;
xor.b32 %r1393, %r1098, %r453;
or.b32 %r1394, %r1393, -2139062144;
and.b32 %r1395, %r453, 2139062143;
sub.s32 %r1396, %r1394, %r1395;
xor.b32 %r460, %r1111, %r1396;

dp4a.s32.s32 %r455, %r456, %r408, %r850;


dp4a.s32.s32 %r459, %r460, %r412, %r455;

mov.u32 %r463, %r855;

prmt.b32 %r463, %r463, 0, 0xba98;

mov.u32 %r465, %r857;

prmt.b32 %r465, %r465, 0, 0xba98;

xor.b32 %r1397, %r1129, %r463;
or.b32 %r1398, %r1397, -2139062144;
and.b32 %r1399, %r463, 2139062143;
sub.s32 %r1400, %r1398, %r1399;
xor.b32 %r468, %r1138, %r1400;
xor.b32 %r1401, %r1130, %r465;
or.b32 %r1402, %r1401, -2139062144;
and.b32 %r1403, %r465, 2139062143;
sub.s32 %r1404, %r1402, %r1403;
xor.b32 %r472, %r1143, %r1404;

dp4a.s32.s32 %r467, %r468, %r420, %r459;


dp4a.s32.s32 %r471, %r472, %r424, %r467;

mul.lo.s32 %r1405, %r447, %r1145;
mad.lo.s32 %r1406, %r471, %r1146, %r1405;
add.s32 %r1407, %r471, %r447;
shr.u32 %r1408, %r1407, 31;
add.s32 %r1409, %r1407, %r1408;
shr.s32 %r1410, %r1409, 1;
add.s32 %r1411, %r1406, %r1410;
shr.s32 %r1412, %r1411, 31;
shr.u32 %r1413, %r1412, 30;
add.s32 %r1414, %r1411, %r1413;
shr.s32 %r1415, %r1414, 2;

{ cvt.f32.f16 %f111, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r426;
cvt.f32.f16 %f112, low;}


mul.ftz.f32 %f143, %f111, %f112;
cvt.rn.f32.s32 %f144, %r1415;
fma.rn.ftz.f32 %f361, %f143, %f144, %f361;
cvt.u64.u32 %rd69, %r2208;
add.s64 %rd70, %rd1, %rd69;
add.s64 %rd71, %rd70, %rd3;
mov.u32 %r476, %r770;

prmt.b32 %r476, %r476, 0, 0xba98;

mov.u32 %r478, %r772;

prmt.b32 %r478, %r478, 0, 0xba98;

xor.b32 %r1416, %r887, %r476;
or.b32 %r1417, %r1416, -2139062144;
and.b32 %r1418, %r476, 2139062143;
sub.s32 %r1419, %r1417, %r1418;
xor.b32 %r481, %r896, %r1419;
xor.b32 %r1420, %r888, %r478;
or.b32 %r1421, %r1420, -2139062144;
and.b32 %r1422, %r478, 2139062143;
sub.s32 %r1423, %r1421, %r1422;
xor.b32 %r485, %r901, %r1423;
mul.lo.s64 %rd72, %rd71, 36;
add.s64 %rd73, %rd10, %rd72;
ld.global.nc.u32 %r486, [%rd73+8];
ld.global.nc.u32 %r482, [%rd73+4];

dp4a.s32.s32 %r480, %r481, %r482, %r850;


dp4a.s32.s32 %r484, %r485, %r486, %r480;

mov.u32 %r488, %r782;

prmt.b32 %r488, %r488, 0, 0xba98;

mov.u32 %r490, %r784;

prmt.b32 %r490, %r490, 0, 0xba98;

xor.b32 %r1424, %r919, %r488;
or.b32 %r1425, %r1424, -2139062144;
and.b32 %r1426, %r488, 2139062143;
sub.s32 %r1427, %r1425, %r1426;
xor.b32 %r493, %r928, %r1427;
xor.b32 %r1428, %r920, %r490;
or.b32 %r1429, %r1428, -2139062144;
and.b32 %r1430, %r490, 2139062143;
sub.s32 %r1431, %r1429, %r1430;
xor.b32 %r497, %r933, %r1431;
ld.global.nc.u32 %r498, [%rd73+16];
ld.global.nc.u32 %r494, [%rd73+12];

dp4a.s32.s32 %r492, %r493, %r494, %r484;


dp4a.s32.s32 %r496, %r497, %r498, %r492;

mov.u32 %r500, %r794;

prmt.b32 %r500, %r500, 0, 0xba98;

mov.u32 %r502, %r796;

prmt.b32 %r502, %r502, 0, 0xba98;

xor.b32 %r1432, %r952, %r500;
or.b32 %r1433, %r1432, -2139062144;
and.b32 %r1434, %r500, 2139062143;
sub.s32 %r1435, %r1433, %r1434;
xor.b32 %r505, %r961, %r1435;
xor.b32 %r1436, %r953, %r502;
or.b32 %r1437, %r1436, -2139062144;
and.b32 %r1438, %r502, 2139062143;
sub.s32 %r1439, %r1437, %r1438;
xor.b32 %r509, %r966, %r1439;
ld.global.nc.u32 %r510, [%rd73+24];
ld.global.nc.u32 %r506, [%rd73+20];

dp4a.s32.s32 %r504, %r505, %r506, %r850;


dp4a.s32.s32 %r508, %r509, %r510, %r504;

mov.u32 %r512, %r806;

prmt.b32 %r512, %r512, 0, 0xba98;

mov.u32 %r514, %r808;

prmt.b32 %r514, %r514, 0, 0xba98;

xor.b32 %r1440, %r984, %r512;
or.b32 %r1441, %r1440, -2139062144;
and.b32 %r1442, %r512, 2139062143;
sub.s32 %r1443, %r1441, %r1442;
xor.b32 %r517, %r993, %r1443;
xor.b32 %r1444, %r985, %r514;
or.b32 %r1445, %r1444, -2139062144;
and.b32 %r1446, %r514, 2139062143;
sub.s32 %r1447, %r1445, %r1446;
xor.b32 %r521, %r998, %r1447;
ld.global.nc.u32 %r522, [%rd73+32];
ld.global.nc.u32 %r518, [%rd73+28];

dp4a.s32.s32 %r516, %r517, %r518, %r508;


dp4a.s32.s32 %r520, %r521, %r522, %r516;

mul.lo.s32 %r1448, %r496, %r1000;
mad.lo.s32 %r1449, %r520, %r1001, %r1448;
add.s32 %r1450, %r520, %r496;
shr.u32 %r1451, %r1450, 31;
add.s32 %r1452, %r1450, %r1451;
shr.s32 %r1453, %r1452, 1;
add.s32 %r1454, %r1449, %r1453;
shr.s32 %r1455, %r1454, 31;
shr.u32 %r1456, %r1455, 30;
add.s32 %r1457, %r1454, %r1456;
shr.s32 %r1458, %r1457, 2;

{ cvt.f32.f16 %f113, %rs1;}


ld.global.nc.u32 %r524, [%rd73];

{.reg .f16 low,high;
mov.b32 {low,high},%r524;
cvt.f32.f16 %f114, low;}


mul.ftz.f32 %f145, %f113, %f114;
cvt.rn.f32.s32 %f146, %r1458;
fma.rn.ftz.f32 %f360, %f145, %f146, %f360;
mov.u32 %r525, %r819;

prmt.b32 %r525, %r525, 0, 0xba98;

mov.u32 %r527, %r821;

prmt.b32 %r527, %r527, 0, 0xba98;

xor.b32 %r1459, %r1032, %r525;
or.b32 %r1460, %r1459, -2139062144;
and.b32 %r1461, %r525, 2139062143;
sub.s32 %r1462, %r1460, %r1461;
xor.b32 %r530, %r1041, %r1462;
xor.b32 %r1463, %r1033, %r527;
or.b32 %r1464, %r1463, -2139062144;
and.b32 %r1465, %r527, 2139062143;
sub.s32 %r1466, %r1464, %r1465;
xor.b32 %r534, %r1046, %r1466;

dp4a.s32.s32 %r529, %r530, %r482, %r850;


dp4a.s32.s32 %r533, %r534, %r486, %r529;

mov.u32 %r537, %r831;

prmt.b32 %r537, %r537, 0, 0xba98;

mov.u32 %r539, %r833;

prmt.b32 %r539, %r539, 0, 0xba98;

xor.b32 %r1467, %r1064, %r537;
or.b32 %r1468, %r1467, -2139062144;
and.b32 %r1469, %r537, 2139062143;
sub.s32 %r1470, %r1468, %r1469;
xor.b32 %r542, %r1073, %r1470;
xor.b32 %r1471, %r1065, %r539;
or.b32 %r1472, %r1471, -2139062144;
and.b32 %r1473, %r539, 2139062143;
sub.s32 %r1474, %r1472, %r1473;
xor.b32 %r546, %r1078, %r1474;

dp4a.s32.s32 %r541, %r542, %r494, %r533;


dp4a.s32.s32 %r545, %r546, %r498, %r541;

mov.u32 %r549, %r843;

prmt.b32 %r549, %r549, 0, 0xba98;

mov.u32 %r551, %r845;

prmt.b32 %r551, %r551, 0, 0xba98;

xor.b32 %r1475, %r1097, %r549;
or.b32 %r1476, %r1475, -2139062144;
and.b32 %r1477, %r549, 2139062143;
sub.s32 %r1478, %r1476, %r1477;
xor.b32 %r554, %r1106, %r1478;
xor.b32 %r1479, %r1098, %r551;
or.b32 %r1480, %r1479, -2139062144;
and.b32 %r1481, %r551, 2139062143;
sub.s32 %r1482, %r1480, %r1481;
xor.b32 %r558, %r1111, %r1482;

dp4a.s32.s32 %r553, %r554, %r506, %r850;


dp4a.s32.s32 %r557, %r558, %r510, %r553;

mov.u32 %r561, %r855;

prmt.b32 %r561, %r561, 0, 0xba98;

mov.u32 %r563, %r857;

prmt.b32 %r563, %r563, 0, 0xba98;

xor.b32 %r1483, %r1129, %r561;
or.b32 %r1484, %r1483, -2139062144;
and.b32 %r1485, %r561, 2139062143;
sub.s32 %r1486, %r1484, %r1485;
xor.b32 %r566, %r1138, %r1486;
xor.b32 %r1487, %r1130, %r563;
or.b32 %r1488, %r1487, -2139062144;
and.b32 %r1489, %r563, 2139062143;
sub.s32 %r1490, %r1488, %r1489;
xor.b32 %r570, %r1143, %r1490;

dp4a.s32.s32 %r565, %r566, %r518, %r557;


dp4a.s32.s32 %r569, %r570, %r522, %r565;

mul.lo.s32 %r1491, %r545, %r1145;
mad.lo.s32 %r1492, %r569, %r1146, %r1491;
add.s32 %r1493, %r569, %r545;
shr.u32 %r1494, %r1493, 31;
add.s32 %r1495, %r1493, %r1494;
shr.s32 %r1496, %r1495, 1;
add.s32 %r1497, %r1492, %r1496;
shr.s32 %r1498, %r1497, 31;
shr.u32 %r1499, %r1498, 30;
add.s32 %r1500, %r1497, %r1499;
shr.s32 %r1501, %r1500, 2;

{ cvt.f32.f16 %f115, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r524;
cvt.f32.f16 %f116, low;}


mul.ftz.f32 %f147, %f115, %f116;
cvt.rn.f32.s32 %f148, %r1501;
fma.rn.ftz.f32 %f359, %f147, %f148, %f359;
cvt.u64.u32 %rd74, %r2207;
add.s64 %rd75, %rd1, %rd74;
add.s64 %rd76, %rd75, %rd3;
mov.u32 %r574, %r770;

prmt.b32 %r574, %r574, 0, 0xba98;

mov.u32 %r576, %r772;

prmt.b32 %r576, %r576, 0, 0xba98;

xor.b32 %r1502, %r887, %r574;
or.b32 %r1503, %r1502, -2139062144;
and.b32 %r1504, %r574, 2139062143;
sub.s32 %r1505, %r1503, %r1504;
xor.b32 %r579, %r896, %r1505;
xor.b32 %r1506, %r888, %r576;
or.b32 %r1507, %r1506, -2139062144;
and.b32 %r1508, %r576, 2139062143;
sub.s32 %r1509, %r1507, %r1508;
xor.b32 %r583, %r901, %r1509;
mul.lo.s64 %rd77, %rd76, 36;
add.s64 %rd78, %rd10, %rd77;
ld.global.nc.u32 %r584, [%rd78+8];
ld.global.nc.u32 %r580, [%rd78+4];

dp4a.s32.s32 %r578, %r579, %r580, %r850;


dp4a.s32.s32 %r582, %r583, %r584, %r578;

mov.u32 %r586, %r782;

prmt.b32 %r586, %r586, 0, 0xba98;

mov.u32 %r588, %r784;

prmt.b32 %r588, %r588, 0, 0xba98;

xor.b32 %r1510, %r919, %r586;
or.b32 %r1511, %r1510, -2139062144;
and.b32 %r1512, %r586, 2139062143;
sub.s32 %r1513, %r1511, %r1512;
xor.b32 %r591, %r928, %r1513;
xor.b32 %r1514, %r920, %r588;
or.b32 %r1515, %r1514, -2139062144;
and.b32 %r1516, %r588, 2139062143;
sub.s32 %r1517, %r1515, %r1516;
xor.b32 %r595, %r933, %r1517;
ld.global.nc.u32 %r596, [%rd78+16];
ld.global.nc.u32 %r592, [%rd78+12];

dp4a.s32.s32 %r590, %r591, %r592, %r582;


dp4a.s32.s32 %r594, %r595, %r596, %r590;

mov.u32 %r598, %r794;

prmt.b32 %r598, %r598, 0, 0xba98;

mov.u32 %r600, %r796;

prmt.b32 %r600, %r600, 0, 0xba98;

xor.b32 %r1518, %r952, %r598;
or.b32 %r1519, %r1518, -2139062144;
and.b32 %r1520, %r598, 2139062143;
sub.s32 %r1521, %r1519, %r1520;
xor.b32 %r603, %r961, %r1521;
xor.b32 %r1522, %r953, %r600;
or.b32 %r1523, %r1522, -2139062144;
and.b32 %r1524, %r600, 2139062143;
sub.s32 %r1525, %r1523, %r1524;
xor.b32 %r607, %r966, %r1525;
ld.global.nc.u32 %r608, [%rd78+24];
ld.global.nc.u32 %r604, [%rd78+20];

dp4a.s32.s32 %r602, %r603, %r604, %r850;


dp4a.s32.s32 %r606, %r607, %r608, %r602;

mov.u32 %r610, %r806;

prmt.b32 %r610, %r610, 0, 0xba98;

mov.u32 %r612, %r808;

prmt.b32 %r612, %r612, 0, 0xba98;

xor.b32 %r1526, %r984, %r610;
or.b32 %r1527, %r1526, -2139062144;
and.b32 %r1528, %r610, 2139062143;
sub.s32 %r1529, %r1527, %r1528;
xor.b32 %r615, %r993, %r1529;
xor.b32 %r1530, %r985, %r612;
or.b32 %r1531, %r1530, -2139062144;
and.b32 %r1532, %r612, 2139062143;
sub.s32 %r1533, %r1531, %r1532;
xor.b32 %r619, %r998, %r1533;
ld.global.nc.u32 %r620, [%rd78+32];
ld.global.nc.u32 %r616, [%rd78+28];

dp4a.s32.s32 %r614, %r615, %r616, %r606;


dp4a.s32.s32 %r618, %r619, %r620, %r614;

mul.lo.s32 %r1534, %r594, %r1000;
mad.lo.s32 %r1535, %r618, %r1001, %r1534;
add.s32 %r1536, %r618, %r594;
shr.u32 %r1537, %r1536, 31;
add.s32 %r1538, %r1536, %r1537;
shr.s32 %r1539, %r1538, 1;
add.s32 %r1540, %r1535, %r1539;
shr.s32 %r1541, %r1540, 31;
shr.u32 %r1542, %r1541, 30;
add.s32 %r1543, %r1540, %r1542;
shr.s32 %r1544, %r1543, 2;

{ cvt.f32.f16 %f117, %rs1;}


ld.global.nc.u32 %r622, [%rd78];

{.reg .f16 low,high;
mov.b32 {low,high},%r622;
cvt.f32.f16 %f118, low;}


mul.ftz.f32 %f149, %f117, %f118;
cvt.rn.f32.s32 %f150, %r1544;
fma.rn.ftz.f32 %f358, %f149, %f150, %f358;
mov.u32 %r623, %r819;

prmt.b32 %r623, %r623, 0, 0xba98;

mov.u32 %r625, %r821;

prmt.b32 %r625, %r625, 0, 0xba98;

xor.b32 %r1545, %r1032, %r623;
or.b32 %r1546, %r1545, -2139062144;
and.b32 %r1547, %r623, 2139062143;
sub.s32 %r1548, %r1546, %r1547;
xor.b32 %r628, %r1041, %r1548;
xor.b32 %r1549, %r1033, %r625;
or.b32 %r1550, %r1549, -2139062144;
and.b32 %r1551, %r625, 2139062143;
sub.s32 %r1552, %r1550, %r1551;
xor.b32 %r632, %r1046, %r1552;

dp4a.s32.s32 %r627, %r628, %r580, %r850;


dp4a.s32.s32 %r631, %r632, %r584, %r627;

mov.u32 %r635, %r831;

prmt.b32 %r635, %r635, 0, 0xba98;

mov.u32 %r637, %r833;

prmt.b32 %r637, %r637, 0, 0xba98;

xor.b32 %r1553, %r1064, %r635;
or.b32 %r1554, %r1553, -2139062144;
and.b32 %r1555, %r635, 2139062143;
sub.s32 %r1556, %r1554, %r1555;
xor.b32 %r640, %r1073, %r1556;
xor.b32 %r1557, %r1065, %r637;
or.b32 %r1558, %r1557, -2139062144;
and.b32 %r1559, %r637, 2139062143;
sub.s32 %r1560, %r1558, %r1559;
xor.b32 %r644, %r1078, %r1560;

dp4a.s32.s32 %r639, %r640, %r592, %r631;


dp4a.s32.s32 %r643, %r644, %r596, %r639;

mov.u32 %r647, %r843;

prmt.b32 %r647, %r647, 0, 0xba98;

mov.u32 %r649, %r845;

prmt.b32 %r649, %r649, 0, 0xba98;

xor.b32 %r1561, %r1097, %r647;
or.b32 %r1562, %r1561, -2139062144;
and.b32 %r1563, %r647, 2139062143;
sub.s32 %r1564, %r1562, %r1563;
xor.b32 %r652, %r1106, %r1564;
xor.b32 %r1565, %r1098, %r649;
or.b32 %r1566, %r1565, -2139062144;
and.b32 %r1567, %r649, 2139062143;
sub.s32 %r1568, %r1566, %r1567;
xor.b32 %r656, %r1111, %r1568;

dp4a.s32.s32 %r651, %r652, %r604, %r850;


dp4a.s32.s32 %r655, %r656, %r608, %r651;

mov.u32 %r659, %r855;

prmt.b32 %r659, %r659, 0, 0xba98;

mov.u32 %r661, %r857;

prmt.b32 %r661, %r661, 0, 0xba98;

xor.b32 %r1569, %r1129, %r659;
or.b32 %r1570, %r1569, -2139062144;
and.b32 %r1571, %r659, 2139062143;
sub.s32 %r1572, %r1570, %r1571;
xor.b32 %r664, %r1138, %r1572;
xor.b32 %r1573, %r1130, %r661;
or.b32 %r1574, %r1573, -2139062144;
and.b32 %r1575, %r661, 2139062143;
sub.s32 %r1576, %r1574, %r1575;
xor.b32 %r668, %r1143, %r1576;

dp4a.s32.s32 %r663, %r664, %r616, %r655;


dp4a.s32.s32 %r667, %r668, %r620, %r663;

mul.lo.s32 %r1577, %r643, %r1145;
mad.lo.s32 %r1578, %r667, %r1146, %r1577;
add.s32 %r1579, %r667, %r643;
shr.u32 %r1580, %r1579, 31;
add.s32 %r1581, %r1579, %r1580;
shr.s32 %r1582, %r1581, 1;
add.s32 %r1583, %r1578, %r1582;
shr.s32 %r1584, %r1583, 31;
shr.u32 %r1585, %r1584, 30;
add.s32 %r1586, %r1583, %r1585;
shr.s32 %r1587, %r1586, 2;

{ cvt.f32.f16 %f119, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r622;
cvt.f32.f16 %f120, low;}


mul.ftz.f32 %f151, %f119, %f120;
cvt.rn.f32.s32 %f152, %r1587;
fma.rn.ftz.f32 %f357, %f151, %f152, %f357;
cvt.u64.u32 %rd79, %r2206;
add.s64 %rd80, %rd1, %rd79;
add.s64 %rd81, %rd80, %rd3;
mov.u32 %r672, %r770;

prmt.b32 %r672, %r672, 0, 0xba98;

mov.u32 %r674, %r772;

prmt.b32 %r674, %r674, 0, 0xba98;

xor.b32 %r1588, %r887, %r672;
or.b32 %r1589, %r1588, -2139062144;
and.b32 %r1590, %r672, 2139062143;
sub.s32 %r1591, %r1589, %r1590;
xor.b32 %r677, %r896, %r1591;
xor.b32 %r1592, %r888, %r674;
or.b32 %r1593, %r1592, -2139062144;
and.b32 %r1594, %r674, 2139062143;
sub.s32 %r1595, %r1593, %r1594;
xor.b32 %r681, %r901, %r1595;
mul.lo.s64 %rd82, %rd81, 36;
add.s64 %rd83, %rd10, %rd82;
ld.global.nc.u32 %r682, [%rd83+8];
ld.global.nc.u32 %r678, [%rd83+4];

dp4a.s32.s32 %r676, %r677, %r678, %r850;


dp4a.s32.s32 %r680, %r681, %r682, %r676;

mov.u32 %r684, %r782;

prmt.b32 %r684, %r684, 0, 0xba98;

mov.u32 %r686, %r784;

prmt.b32 %r686, %r686, 0, 0xba98;

xor.b32 %r1596, %r919, %r684;
or.b32 %r1597, %r1596, -2139062144;
and.b32 %r1598, %r684, 2139062143;
sub.s32 %r1599, %r1597, %r1598;
xor.b32 %r689, %r928, %r1599;
xor.b32 %r1600, %r920, %r686;
or.b32 %r1601, %r1600, -2139062144;
and.b32 %r1602, %r686, 2139062143;
sub.s32 %r1603, %r1601, %r1602;
xor.b32 %r693, %r933, %r1603;
ld.global.nc.u32 %r694, [%rd83+16];
ld.global.nc.u32 %r690, [%rd83+12];

dp4a.s32.s32 %r688, %r689, %r690, %r680;


dp4a.s32.s32 %r692, %r693, %r694, %r688;

mov.u32 %r696, %r794;

prmt.b32 %r696, %r696, 0, 0xba98;

mov.u32 %r698, %r796;

prmt.b32 %r698, %r698, 0, 0xba98;

xor.b32 %r1604, %r952, %r696;
or.b32 %r1605, %r1604, -2139062144;
and.b32 %r1606, %r696, 2139062143;
sub.s32 %r1607, %r1605, %r1606;
xor.b32 %r701, %r961, %r1607;
xor.b32 %r1608, %r953, %r698;
or.b32 %r1609, %r1608, -2139062144;
and.b32 %r1610, %r698, 2139062143;
sub.s32 %r1611, %r1609, %r1610;
xor.b32 %r705, %r966, %r1611;
ld.global.nc.u32 %r706, [%rd83+24];
ld.global.nc.u32 %r702, [%rd83+20];

dp4a.s32.s32 %r700, %r701, %r702, %r850;


dp4a.s32.s32 %r704, %r705, %r706, %r700;

mov.u32 %r708, %r806;

prmt.b32 %r708, %r708, 0, 0xba98;

mov.u32 %r710, %r808;

prmt.b32 %r710, %r710, 0, 0xba98;

xor.b32 %r1612, %r984, %r708;
or.b32 %r1613, %r1612, -2139062144;
and.b32 %r1614, %r708, 2139062143;
sub.s32 %r1615, %r1613, %r1614;
xor.b32 %r713, %r993, %r1615;
xor.b32 %r1616, %r985, %r710;
or.b32 %r1617, %r1616, -2139062144;
and.b32 %r1618, %r710, 2139062143;
sub.s32 %r1619, %r1617, %r1618;
xor.b32 %r717, %r998, %r1619;
ld.global.nc.u32 %r718, [%rd83+32];
ld.global.nc.u32 %r714, [%rd83+28];

dp4a.s32.s32 %r712, %r713, %r714, %r704;


dp4a.s32.s32 %r716, %r717, %r718, %r712;

mul.lo.s32 %r1620, %r692, %r1000;
mad.lo.s32 %r1621, %r716, %r1001, %r1620;
add.s32 %r1622, %r716, %r692;
shr.u32 %r1623, %r1622, 31;
add.s32 %r1624, %r1622, %r1623;
shr.s32 %r1625, %r1624, 1;
add.s32 %r1626, %r1621, %r1625;
shr.s32 %r1627, %r1626, 31;
shr.u32 %r1628, %r1627, 30;
add.s32 %r1629, %r1626, %r1628;
shr.s32 %r1630, %r1629, 2;

{ cvt.f32.f16 %f121, %rs1;}


ld.global.nc.u32 %r720, [%rd83];

{.reg .f16 low,high;
mov.b32 {low,high},%r720;
cvt.f32.f16 %f122, low;}


mul.ftz.f32 %f153, %f121, %f122;
cvt.rn.f32.s32 %f154, %r1630;
fma.rn.ftz.f32 %f356, %f153, %f154, %f356;
mov.u32 %r721, %r819;

prmt.b32 %r721, %r721, 0, 0xba98;

mov.u32 %r723, %r821;

prmt.b32 %r723, %r723, 0, 0xba98;

xor.b32 %r1631, %r1032, %r721;
or.b32 %r1632, %r1631, -2139062144;
and.b32 %r1633, %r721, 2139062143;
sub.s32 %r1634, %r1632, %r1633;
xor.b32 %r726, %r1041, %r1634;
xor.b32 %r1635, %r1033, %r723;
or.b32 %r1636, %r1635, -2139062144;
and.b32 %r1637, %r723, 2139062143;
sub.s32 %r1638, %r1636, %r1637;
xor.b32 %r730, %r1046, %r1638;

dp4a.s32.s32 %r725, %r726, %r678, %r850;


dp4a.s32.s32 %r729, %r730, %r682, %r725;

mov.u32 %r733, %r831;

prmt.b32 %r733, %r733, 0, 0xba98;

mov.u32 %r735, %r833;

prmt.b32 %r735, %r735, 0, 0xba98;

xor.b32 %r1639, %r1064, %r733;
or.b32 %r1640, %r1639, -2139062144;
and.b32 %r1641, %r733, 2139062143;
sub.s32 %r1642, %r1640, %r1641;
xor.b32 %r738, %r1073, %r1642;
xor.b32 %r1643, %r1065, %r735;
or.b32 %r1644, %r1643, -2139062144;
and.b32 %r1645, %r735, 2139062143;
sub.s32 %r1646, %r1644, %r1645;
xor.b32 %r742, %r1078, %r1646;

dp4a.s32.s32 %r737, %r738, %r690, %r729;


dp4a.s32.s32 %r741, %r742, %r694, %r737;

mov.u32 %r745, %r843;

prmt.b32 %r745, %r745, 0, 0xba98;

mov.u32 %r747, %r845;

prmt.b32 %r747, %r747, 0, 0xba98;

xor.b32 %r1647, %r1097, %r745;
or.b32 %r1648, %r1647, -2139062144;
and.b32 %r1649, %r745, 2139062143;
sub.s32 %r1650, %r1648, %r1649;
xor.b32 %r750, %r1106, %r1650;
xor.b32 %r1651, %r1098, %r747;
or.b32 %r1652, %r1651, -2139062144;
and.b32 %r1653, %r747, 2139062143;
sub.s32 %r1654, %r1652, %r1653;
xor.b32 %r754, %r1111, %r1654;

dp4a.s32.s32 %r749, %r750, %r702, %r850;


dp4a.s32.s32 %r753, %r754, %r706, %r749;

mov.u32 %r757, %r855;

prmt.b32 %r757, %r757, 0, 0xba98;

mov.u32 %r759, %r857;

prmt.b32 %r759, %r759, 0, 0xba98;

xor.b32 %r1655, %r1129, %r757;
or.b32 %r1656, %r1655, -2139062144;
and.b32 %r1657, %r757, 2139062143;
sub.s32 %r1658, %r1656, %r1657;
xor.b32 %r762, %r1138, %r1658;
xor.b32 %r1659, %r1130, %r759;
or.b32 %r1660, %r1659, -2139062144;
and.b32 %r1661, %r759, 2139062143;
sub.s32 %r1662, %r1660, %r1661;
xor.b32 %r766, %r1143, %r1662;

dp4a.s32.s32 %r761, %r762, %r714, %r753;


dp4a.s32.s32 %r765, %r766, %r718, %r761;

mul.lo.s32 %r1663, %r741, %r1145;
mad.lo.s32 %r1664, %r765, %r1146, %r1663;
add.s32 %r1665, %r765, %r741;
shr.u32 %r1666, %r1665, 31;
add.s32 %r1667, %r1665, %r1666;
shr.s32 %r1668, %r1667, 1;
add.s32 %r1669, %r1664, %r1668;
shr.s32 %r1670, %r1669, 31;
shr.u32 %r1671, %r1670, 30;
add.s32 %r1672, %r1669, %r1671;
shr.s32 %r1673, %r1672, 2;

{ cvt.f32.f16 %f123, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r720;
cvt.f32.f16 %f124, low;}


mul.ftz.f32 %f155, %f123, %f124;
cvt.rn.f32.s32 %f156, %r1673;
fma.rn.ftz.f32 %f355, %f155, %f156, %f355;
cvt.u64.u32 %rd84, %r2205;
add.s64 %rd85, %rd1, %rd84;
add.s64 %rd86, %rd85, %rd3;

prmt.b32 %r770, %r770, 0, 0xba98;


prmt.b32 %r772, %r772, 0, 0xba98;

xor.b32 %r1674, %r887, %r770;
or.b32 %r1675, %r1674, -2139062144;
and.b32 %r1676, %r770, 2139062143;
sub.s32 %r1677, %r1675, %r1676;
xor.b32 %r775, %r896, %r1677;
xor.b32 %r1678, %r888, %r772;
or.b32 %r1679, %r1678, -2139062144;
and.b32 %r1680, %r772, 2139062143;
sub.s32 %r1681, %r1679, %r1680;
xor.b32 %r779, %r901, %r1681;
mul.lo.s64 %rd87, %rd86, 36;
add.s64 %rd88, %rd10, %rd87;
ld.global.nc.u32 %r780, [%rd88+8];
ld.global.nc.u32 %r776, [%rd88+4];

dp4a.s32.s32 %r774, %r775, %r776, %r850;


dp4a.s32.s32 %r778, %r779, %r780, %r774;


prmt.b32 %r782, %r782, 0, 0xba98;


prmt.b32 %r784, %r784, 0, 0xba98;

xor.b32 %r1682, %r919, %r782;
or.b32 %r1683, %r1682, -2139062144;
and.b32 %r1684, %r782, 2139062143;
sub.s32 %r1685, %r1683, %r1684;
xor.b32 %r787, %r928, %r1685;
xor.b32 %r1686, %r920, %r784;
or.b32 %r1687, %r1686, -2139062144;
and.b32 %r1688, %r784, 2139062143;
sub.s32 %r1689, %r1687, %r1688;
xor.b32 %r791, %r933, %r1689;
ld.global.nc.u32 %r792, [%rd88+16];
ld.global.nc.u32 %r788, [%rd88+12];

dp4a.s32.s32 %r786, %r787, %r788, %r778;


dp4a.s32.s32 %r790, %r791, %r792, %r786;


prmt.b32 %r794, %r794, 0, 0xba98;


prmt.b32 %r796, %r796, 0, 0xba98;

xor.b32 %r1690, %r952, %r794;
or.b32 %r1691, %r1690, -2139062144;
and.b32 %r1692, %r794, 2139062143;
sub.s32 %r1693, %r1691, %r1692;
xor.b32 %r799, %r961, %r1693;
xor.b32 %r1694, %r953, %r796;
or.b32 %r1695, %r1694, -2139062144;
and.b32 %r1696, %r796, 2139062143;
sub.s32 %r1697, %r1695, %r1696;
xor.b32 %r803, %r966, %r1697;
ld.global.nc.u32 %r804, [%rd88+24];
ld.global.nc.u32 %r800, [%rd88+20];

dp4a.s32.s32 %r798, %r799, %r800, %r850;


dp4a.s32.s32 %r802, %r803, %r804, %r798;


prmt.b32 %r806, %r806, 0, 0xba98;


prmt.b32 %r808, %r808, 0, 0xba98;

xor.b32 %r1698, %r984, %r806;
or.b32 %r1699, %r1698, -2139062144;
and.b32 %r1700, %r806, 2139062143;
sub.s32 %r1701, %r1699, %r1700;
xor.b32 %r811, %r993, %r1701;
xor.b32 %r1702, %r985, %r808;
or.b32 %r1703, %r1702, -2139062144;
and.b32 %r1704, %r808, 2139062143;
sub.s32 %r1705, %r1703, %r1704;
xor.b32 %r815, %r998, %r1705;
ld.global.nc.u32 %r816, [%rd88+32];
ld.global.nc.u32 %r812, [%rd88+28];

dp4a.s32.s32 %r810, %r811, %r812, %r802;


dp4a.s32.s32 %r814, %r815, %r816, %r810;

mul.lo.s32 %r1706, %r790, %r1000;
mad.lo.s32 %r1707, %r814, %r1001, %r1706;
add.s32 %r1708, %r814, %r790;
shr.u32 %r1709, %r1708, 31;
add.s32 %r1710, %r1708, %r1709;
shr.s32 %r1711, %r1710, 1;
add.s32 %r1712, %r1707, %r1711;
shr.s32 %r1713, %r1712, 31;
shr.u32 %r1714, %r1713, 30;
add.s32 %r1715, %r1712, %r1714;
shr.s32 %r1716, %r1715, 2;

{ cvt.f32.f16 %f125, %rs1;}


ld.global.nc.u32 %r818, [%rd88];

{.reg .f16 low,high;
mov.b32 {low,high},%r818;
cvt.f32.f16 %f126, low;}


mul.ftz.f32 %f157, %f125, %f126;
cvt.rn.f32.s32 %f158, %r1716;
fma.rn.ftz.f32 %f354, %f157, %f158, %f354;

prmt.b32 %r819, %r819, 0, 0xba98;


prmt.b32 %r821, %r821, 0, 0xba98;

xor.b32 %r1717, %r1032, %r819;
or.b32 %r1718, %r1717, -2139062144;
and.b32 %r1719, %r819, 2139062143;
sub.s32 %r1720, %r1718, %r1719;
xor.b32 %r824, %r1041, %r1720;
xor.b32 %r1721, %r1033, %r821;
or.b32 %r1722, %r1721, -2139062144;
and.b32 %r1723, %r821, 2139062143;
sub.s32 %r1724, %r1722, %r1723;
xor.b32 %r828, %r1046, %r1724;

dp4a.s32.s32 %r823, %r824, %r776, %r850;


dp4a.s32.s32 %r827, %r828, %r780, %r823;


prmt.b32 %r831, %r831, 0, 0xba98;


prmt.b32 %r833, %r833, 0, 0xba98;

xor.b32 %r1725, %r1064, %r831;
or.b32 %r1726, %r1725, -2139062144;
and.b32 %r1727, %r831, 2139062143;
sub.s32 %r1728, %r1726, %r1727;
xor.b32 %r836, %r1073, %r1728;
xor.b32 %r1729, %r1065, %r833;
or.b32 %r1730, %r1729, -2139062144;
and.b32 %r1731, %r833, 2139062143;
sub.s32 %r1732, %r1730, %r1731;
xor.b32 %r840, %r1078, %r1732;

dp4a.s32.s32 %r835, %r836, %r788, %r827;


dp4a.s32.s32 %r839, %r840, %r792, %r835;


prmt.b32 %r843, %r843, 0, 0xba98;


prmt.b32 %r845, %r845, 0, 0xba98;

xor.b32 %r1733, %r1097, %r843;
or.b32 %r1734, %r1733, -2139062144;
and.b32 %r1735, %r843, 2139062143;
sub.s32 %r1736, %r1734, %r1735;
xor.b32 %r848, %r1106, %r1736;
xor.b32 %r1737, %r1098, %r845;
or.b32 %r1738, %r1737, -2139062144;
and.b32 %r1739, %r845, 2139062143;
sub.s32 %r1740, %r1738, %r1739;
xor.b32 %r852, %r1111, %r1740;

dp4a.s32.s32 %r847, %r848, %r800, %r850;


dp4a.s32.s32 %r851, %r852, %r804, %r847;


prmt.b32 %r855, %r855, 0, 0xba98;


prmt.b32 %r857, %r857, 0, 0xba98;

xor.b32 %r1741, %r1129, %r855;
or.b32 %r1742, %r1741, -2139062144;
and.b32 %r1743, %r855, 2139062143;
sub.s32 %r1744, %r1742, %r1743;
xor.b32 %r860, %r1138, %r1744;
xor.b32 %r1745, %r1130, %r857;
or.b32 %r1746, %r1745, -2139062144;
and.b32 %r1747, %r857, 2139062143;
sub.s32 %r1748, %r1746, %r1747;
xor.b32 %r864, %r1143, %r1748;

dp4a.s32.s32 %r859, %r860, %r812, %r851;


dp4a.s32.s32 %r863, %r864, %r816, %r859;

mul.lo.s32 %r1749, %r839, %r1145;
mad.lo.s32 %r1750, %r863, %r1146, %r1749;
add.s32 %r1751, %r863, %r839;
shr.u32 %r1752, %r1751, 31;
add.s32 %r1753, %r1751, %r1752;
shr.s32 %r1754, %r1753, 1;
add.s32 %r1755, %r1750, %r1754;
shr.s32 %r1756, %r1755, 31;
shr.u32 %r1757, %r1756, 30;
add.s32 %r1758, %r1755, %r1757;
shr.s32 %r1759, %r1758, 2;

{ cvt.f32.f16 %f127, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r818;
cvt.f32.f16 %f128, low;}


mul.ftz.f32 %f159, %f127, %f128;
cvt.rn.f32.s32 %f160, %r1759;
fma.rn.ftz.f32 %f353, %f159, %f160, %f353;
add.s32 %r2210, %r2210, 64;
add.s32 %r2209, %r2209, 64;
add.s32 %r2208, %r2208, 64;
add.s32 %r2207, %r2207, 64;
add.s32 %r2206, %r2206, 64;
add.s32 %r2205, %r2205, 64;
add.s32 %r2204, %r2204, 64;
add.s32 %r2203, %r2203, 64;
add.s64 %rd158, %rd158, 656;
add.s32 %r2211, %r2211, 8;
setp.lt.s32 %p3, %r2211, %r3;
@%p3 bra $L__BB236_2;

$L__BB236_3:
setp.eq.s32 %p4, %r49, 0;
@%p4 bra $L__BB236_5;

shl.b32 %r1763, %r49, 11;
mov.u32 %r1764, _ZZ13mul_mat_vec_qIL9ggml_type22ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r1765, %r1764, %r1763;
shl.b32 %r1766, %r51, 2;
add.s32 %r1767, %r1765, %r1766;
st.shared.f32 [%r1767+-2048], %f368;
st.shared.f32 [%r1767+-1920], %f367;
st.shared.f32 [%r1767+-1792], %f366;
st.shared.f32 [%r1767+-1664], %f365;
st.shared.f32 [%r1767+-1536], %f364;
st.shared.f32 [%r1767+-1408], %f363;
st.shared.f32 [%r1767+-1280], %f362;
st.shared.f32 [%r1767+-1152], %f361;
st.shared.f32 [%r1767+-1024], %f360;
st.shared.f32 [%r1767+-896], %f359;
st.shared.f32 [%r1767+-768], %f358;
st.shared.f32 [%r1767+-640], %f357;
st.shared.f32 [%r1767+-512], %f356;
st.shared.f32 [%r1767+-384], %f355;
st.shared.f32 [%r1767+-256], %f354;
st.shared.f32 [%r1767+-128], %f353;

$L__BB236_5:
setp.ne.s32 %p5, %r49, 0;
bar.sync 0;
@%p5 bra $L__BB236_38;

mov.u32 %r1769, %ctaid.x;
shl.b32 %r1770, %r1769, 1;
mov.u32 %r1771, 1;
shl.b32 %r1773, %r51, 2;
mov.u32 %r1774, _ZZ13mul_mat_vec_qIL9ggml_type22ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r1775, %r1774, %r1773;
mov.u32 %r1776, 2;
ld.shared.f32 %f161, [%r1775];
add.ftz.f32 %f162, %f161, %f368;
mov.b32 %r1777, %f162;
mov.u32 %r1778, 31;
mov.u32 %r1779, 16;
mov.u32 %r1780, -1;
shfl.sync.bfly.b32 %r1781|%p6, %r1777, %r1779, %r1778, %r1780;
mov.b32 %f163, %r1781;
add.ftz.f32 %f164, %f162, %f163;
mov.b32 %r1782, %f164;
mov.u32 %r1783, 8;
shfl.sync.bfly.b32 %r1784|%p7, %r1782, %r1783, %r1778, %r1780;
mov.b32 %f165, %r1784;
add.ftz.f32 %f166, %f164, %f165;
mov.b32 %r1785, %f166;
mov.u32 %r1786, 4;
shfl.sync.bfly.b32 %r1787|%p8, %r1785, %r1786, %r1778, %r1780;
mov.b32 %f167, %r1787;
add.ftz.f32 %f168, %f166, %f167;
mov.b32 %r1788, %f168;
shfl.sync.bfly.b32 %r1789|%p9, %r1788, %r1776, %r1778, %r1780;
mov.b32 %f169, %r1789;
add.ftz.f32 %f170, %f168, %f169;
mov.b32 %r1790, %f170;
shfl.sync.bfly.b32 %r1791|%p10, %r1790, %r1771, %r1778, %r1780;
mov.b32 %f171, %r1791;
add.ftz.f32 %f49, %f170, %f171;
setp.ne.s32 %p11, %r51, 0;
setp.ge.u32 %p12, %r1770, %r35;
or.pred %p13, %p11, %p12;
@%p13 bra $L__BB236_8;

cvta.to.global.u64 %rd89, %rd18;
mov.u32 %r1792, %ctaid.y;
mad.lo.s32 %r1795, %r1792, %r41, %r1770;
mov.u32 %r1796, %ctaid.z;
mad.lo.s32 %r1797, %r1796, %r47, %r1795;
mul.wide.u32 %rd90, %r1797, 4;
add.s64 %rd91, %rd89, %rd90;
st.global.f32 [%rd91], %f49;

$L__BB236_8:
ld.shared.f32 %f172, [%r1775+128];
add.ftz.f32 %f173, %f172, %f367;
mov.b32 %r1803, %f173;
shfl.sync.bfly.b32 %r1807|%p14, %r1803, %r1779, %r1778, %r1780;
mov.b32 %f174, %r1807;
add.ftz.f32 %f175, %f173, %f174;
mov.b32 %r1808, %f175;
shfl.sync.bfly.b32 %r1810|%p15, %r1808, %r1783, %r1778, %r1780;
mov.b32 %f176, %r1810;
add.ftz.f32 %f177, %f175, %f176;
mov.b32 %r1811, %f177;
shfl.sync.bfly.b32 %r1813|%p16, %r1811, %r1786, %r1778, %r1780;
mov.b32 %f178, %r1813;
add.ftz.f32 %f179, %f177, %f178;
mov.b32 %r1814, %f179;
shfl.sync.bfly.b32 %r1815|%p17, %r1814, %r1776, %r1778, %r1780;
mov.b32 %f180, %r1815;
add.ftz.f32 %f181, %f179, %f180;
mov.b32 %r1816, %f181;
shfl.sync.bfly.b32 %r1818|%p18, %r1816, %r1771, %r1778, %r1780;
mov.b32 %f182, %r1818;
add.ftz.f32 %f50, %f181, %f182;
setp.eq.s32 %p19, %r51, 1;
or.b32 %r1821, %r1770, 1;
setp.lt.u32 %p20, %r1821, %r35;
and.pred %p1, %p19, %p20;
not.pred %p21, %p1;
@%p21 bra $L__BB236_10;

cvta.to.global.u64 %rd92, %rd18;
mov.u32 %r1822, %ctaid.y;
mad.lo.s32 %r1825, %r1822, %r41, %r1770;
mov.u32 %r1826, %ctaid.z;
mad.lo.s32 %r1827, %r1826, %r47, %r1825;
mul.wide.u32 %rd93, %r1827, 4;
add.s64 %rd94, %rd92, %rd93;
st.global.f32 [%rd94+4], %f50;

$L__BB236_10:
mov.u32 %r1832, 2;
ld.shared.f32 %f183, [%r1775+256];
add.ftz.f32 %f184, %f183, %f366;
mov.b32 %r1833, %f184;
mov.u32 %r1834, 31;
mov.u32 %r1835, 16;
mov.u32 %r1836, -1;
shfl.sync.bfly.b32 %r1837|%p22, %r1833, %r1835, %r1834, %r1836;
mov.b32 %f185, %r1837;
add.ftz.f32 %f186, %f184, %f185;
mov.b32 %r1838, %f186;
mov.u32 %r1839, 8;
shfl.sync.bfly.b32 %r1840|%p23, %r1838, %r1839, %r1834, %r1836;
mov.b32 %f187, %r1840;
add.ftz.f32 %f188, %f186, %f187;
mov.b32 %r1841, %f188;
mov.u32 %r1842, 4;
shfl.sync.bfly.b32 %r1843|%p24, %r1841, %r1842, %r1834, %r1836;
mov.b32 %f189, %r1843;
add.ftz.f32 %f190, %f188, %f189;
mov.b32 %r1844, %f190;
shfl.sync.bfly.b32 %r1845|%p25, %r1844, %r1832, %r1834, %r1836;
mov.b32 %f191, %r1845;
add.ftz.f32 %f192, %f190, %f191;
mov.b32 %r1846, %f192;
mov.u32 %r1847, 1;
shfl.sync.bfly.b32 %r1848|%p26, %r1846, %r1847, %r1834, %r1836;
mov.b32 %f193, %r1848;
add.ftz.f32 %f51, %f192, %f193;
@%p13 bra $L__BB236_12;

cvt.u64.u32 %rd95, %r35;
mov.u32 %r1851, %ctaid.y;
mad.lo.s32 %r1854, %r1851, %r41, %r1770;
mov.u32 %r1855, %ctaid.z;
mad.lo.s32 %r1856, %r1855, %r47, %r1854;
cvt.u64.u32 %rd96, %r1856;
add.s64 %rd97, %rd95, %rd96;
cvta.to.global.u64 %rd98, %rd18;
shl.b64 %rd99, %rd97, 2;
add.s64 %rd100, %rd98, %rd99;
st.global.f32 [%rd100], %f51;

$L__BB236_12:
ld.shared.f32 %f194, [%r1775+384];
add.ftz.f32 %f195, %f194, %f365;
mov.b32 %r1862, %f195;
shfl.sync.bfly.b32 %r1866|%p30, %r1862, %r1835, %r1834, %r1836;
mov.b32 %f196, %r1866;
add.ftz.f32 %f197, %f195, %f196;
mov.b32 %r1867, %f197;
shfl.sync.bfly.b32 %r1869|%p31, %r1867, %r1839, %r1834, %r1836;
mov.b32 %f198, %r1869;
add.ftz.f32 %f199, %f197, %f198;
mov.b32 %r1870, %f199;
shfl.sync.bfly.b32 %r1872|%p32, %r1870, %r1842, %r1834, %r1836;
mov.b32 %f200, %r1872;
add.ftz.f32 %f201, %f199, %f200;
mov.b32 %r1873, %f201;
shfl.sync.bfly.b32 %r1874|%p33, %r1873, %r1832, %r1834, %r1836;
mov.b32 %f202, %r1874;
add.ftz.f32 %f203, %f201, %f202;
mov.b32 %r1875, %f203;
shfl.sync.bfly.b32 %r1877|%p34, %r1875, %r1847, %r1834, %r1836;
mov.b32 %f204, %r1877;
add.ftz.f32 %f52, %f203, %f204;
@%p21 bra $L__BB236_14;

add.s32 %r1878, %r35, 1;
cvt.u64.u32 %rd101, %r1878;
mov.u32 %r1879, %ctaid.y;
mad.lo.s32 %r1882, %r1879, %r41, %r1770;
mov.u32 %r1883, %ctaid.z;
mad.lo.s32 %r1884, %r1883, %r47, %r1882;
cvt.u64.u32 %rd102, %r1884;
add.s64 %rd103, %rd101, %rd102;
cvta.to.global.u64 %rd104, %rd18;
shl.b64 %rd105, %rd103, 2;
add.s64 %rd106, %rd104, %rd105;
st.global.f32 [%rd106], %f52;

$L__BB236_14:
shl.b32 %r1885, %r35, 1;
mov.u32 %r1886, 1;
mov.u32 %r1891, 2;
ld.shared.f32 %f205, [%r1775+512];
add.ftz.f32 %f206, %f205, %f364;
mov.b32 %r1892, %f206;
mov.u32 %r1893, 31;
mov.u32 %r1894, 16;
mov.u32 %r1895, -1;
shfl.sync.bfly.b32 %r1896|%p36, %r1892, %r1894, %r1893, %r1895;
mov.b32 %f207, %r1896;
add.ftz.f32 %f208, %f206, %f207;
mov.b32 %r1897, %f208;
mov.u32 %r1898, 8;
shfl.sync.bfly.b32 %r1899|%p37, %r1897, %r1898, %r1893, %r1895;
mov.b32 %f209, %r1899;
add.ftz.f32 %f210, %f208, %f209;
mov.b32 %r1900, %f210;
mov.u32 %r1901, 4;
shfl.sync.bfly.b32 %r1902|%p38, %r1900, %r1901, %r1893, %r1895;
mov.b32 %f211, %r1902;
add.ftz.f32 %f212, %f210, %f211;
mov.b32 %r1903, %f212;
shfl.sync.bfly.b32 %r1904|%p39, %r1903, %r1891, %r1893, %r1895;
mov.b32 %f213, %r1904;
add.ftz.f32 %f214, %f212, %f213;
mov.b32 %r1905, %f214;
shfl.sync.bfly.b32 %r1906|%p40, %r1905, %r1886, %r1893, %r1895;
mov.b32 %f215, %r1906;
add.ftz.f32 %f53, %f214, %f215;
cvt.u64.u32 %rd107, %r1885;
mov.u32 %r1907, %ctaid.y;
mad.lo.s32 %r1910, %r1907, %r41, %r1770;
mov.u32 %r1911, %ctaid.z;
mad.lo.s32 %r1912, %r1911, %r47, %r1910;
cvt.u64.u32 %rd108, %r1912;
add.s64 %rd109, %rd107, %rd108;
cvta.to.global.u64 %rd110, %rd18;
shl.b64 %rd111, %rd109, 2;
add.s64 %rd13, %rd110, %rd111;
@%p13 bra $L__BB236_16;

st.global.f32 [%rd13], %f53;

$L__BB236_16:
ld.shared.f32 %f216, [%r1775+640];
add.ftz.f32 %f217, %f216, %f363;
mov.b32 %r1918, %f217;
shfl.sync.bfly.b32 %r1922|%p44, %r1918, %r1894, %r1893, %r1895;
mov.b32 %f218, %r1922;
add.ftz.f32 %f219, %f217, %f218;
mov.b32 %r1923, %f219;
shfl.sync.bfly.b32 %r1925|%p45, %r1923, %r1898, %r1893, %r1895;
mov.b32 %f220, %r1925;
add.ftz.f32 %f221, %f219, %f220;
mov.b32 %r1926, %f221;
shfl.sync.bfly.b32 %r1928|%p46, %r1926, %r1901, %r1893, %r1895;
mov.b32 %f222, %r1928;
add.ftz.f32 %f223, %f221, %f222;
mov.b32 %r1929, %f223;
shfl.sync.bfly.b32 %r1930|%p47, %r1929, %r1891, %r1893, %r1895;
mov.b32 %f224, %r1930;
add.ftz.f32 %f225, %f223, %f224;
mov.b32 %r1931, %f225;
shfl.sync.bfly.b32 %r1933|%p48, %r1931, %r1886, %r1893, %r1895;
mov.b32 %f226, %r1933;
add.ftz.f32 %f54, %f225, %f226;
@%p21 bra $L__BB236_18;

st.global.f32 [%rd13+4], %f54;

$L__BB236_18:
mul.lo.s32 %r30, %r35, 3;
mov.u32 %r1938, 2;
ld.shared.f32 %f227, [%r1775+768];
add.ftz.f32 %f228, %f227, %f362;
mov.b32 %r1939, %f228;
mov.u32 %r1940, 31;
mov.u32 %r1941, 16;
mov.u32 %r1942, -1;
shfl.sync.bfly.b32 %r1943|%p50, %r1939, %r1941, %r1940, %r1942;
mov.b32 %f229, %r1943;
add.ftz.f32 %f230, %f228, %f229;
mov.b32 %r1944, %f230;
mov.u32 %r1945, 8;
shfl.sync.bfly.b32 %r1946|%p51, %r1944, %r1945, %r1940, %r1942;
mov.b32 %f231, %r1946;
add.ftz.f32 %f232, %f230, %f231;
mov.b32 %r1947, %f232;
mov.u32 %r1948, 4;
shfl.sync.bfly.b32 %r1949|%p52, %r1947, %r1948, %r1940, %r1942;
mov.b32 %f233, %r1949;
add.ftz.f32 %f234, %f232, %f233;
mov.b32 %r1950, %f234;
shfl.sync.bfly.b32 %r1951|%p53, %r1950, %r1938, %r1940, %r1942;
mov.b32 %f235, %r1951;
add.ftz.f32 %f236, %f234, %f235;
mov.b32 %r1952, %f236;
mov.u32 %r1953, 1;
shfl.sync.bfly.b32 %r1954|%p54, %r1952, %r1953, %r1940, %r1942;
mov.b32 %f237, %r1954;
add.ftz.f32 %f55, %f236, %f237;
@%p13 bra $L__BB236_20;

cvt.u64.u32 %rd112, %r30;
add.s64 %rd114, %rd112, %rd108;
shl.b64 %rd116, %rd114, 2;
add.s64 %rd117, %rd110, %rd116;
st.global.f32 [%rd117], %f55;

$L__BB236_20:
ld.shared.f32 %f238, [%r1775+896];
add.ftz.f32 %f239, %f238, %f361;
mov.b32 %r1968, %f239;
shfl.sync.bfly.b32 %r1972|%p58, %r1968, %r1941, %r1940, %r1942;
mov.b32 %f240, %r1972;
add.ftz.f32 %f241, %f239, %f240;
mov.b32 %r1973, %f241;
shfl.sync.bfly.b32 %r1975|%p59, %r1973, %r1945, %r1940, %r1942;
mov.b32 %f242, %r1975;
add.ftz.f32 %f243, %f241, %f242;
mov.b32 %r1976, %f243;
shfl.sync.bfly.b32 %r1978|%p60, %r1976, %r1948, %r1940, %r1942;
mov.b32 %f244, %r1978;
add.ftz.f32 %f245, %f243, %f244;
mov.b32 %r1979, %f245;
shfl.sync.bfly.b32 %r1980|%p61, %r1979, %r1938, %r1940, %r1942;
mov.b32 %f246, %r1980;
add.ftz.f32 %f247, %f245, %f246;
mov.b32 %r1981, %f247;
shfl.sync.bfly.b32 %r1983|%p62, %r1981, %r1953, %r1940, %r1942;
mov.b32 %f248, %r1983;
add.ftz.f32 %f56, %f247, %f248;
@%p21 bra $L__BB236_22;

add.s32 %r1984, %r30, 1;
cvt.u64.u32 %rd118, %r1984;
add.s64 %rd120, %rd118, %rd108;
shl.b64 %rd122, %rd120, 2;
add.s64 %rd123, %rd110, %rd122;
st.global.f32 [%rd123], %f56;

$L__BB236_22:
shl.b32 %r1991, %r35, 2;
mov.u32 %r1992, 2;
ld.shared.f32 %f249, [%r1775+1024];
add.ftz.f32 %f250, %f249, %f360;
mov.b32 %r1997, %f250;
mov.u32 %r1998, 31;
mov.u32 %r1999, 16;
mov.u32 %r2000, -1;
shfl.sync.bfly.b32 %r2001|%p64, %r1997, %r1999, %r1998, %r2000;
mov.b32 %f251, %r2001;
add.ftz.f32 %f252, %f250, %f251;
mov.b32 %r2002, %f252;
mov.u32 %r2003, 8;
shfl.sync.bfly.b32 %r2004|%p65, %r2002, %r2003, %r1998, %r2000;
mov.b32 %f253, %r2004;
add.ftz.f32 %f254, %f252, %f253;
mov.b32 %r2005, %f254;
mov.u32 %r2006, 4;
shfl.sync.bfly.b32 %r2007|%p66, %r2005, %r2006, %r1998, %r2000;
mov.b32 %f255, %r2007;
add.ftz.f32 %f256, %f254, %f255;
mov.b32 %r2008, %f256;
shfl.sync.bfly.b32 %r2009|%p67, %r2008, %r1992, %r1998, %r2000;
mov.b32 %f257, %r2009;
add.ftz.f32 %f258, %f256, %f257;
mov.b32 %r2010, %f258;
mov.u32 %r2011, 1;
shfl.sync.bfly.b32 %r2012|%p68, %r2010, %r2011, %r1998, %r2000;
mov.b32 %f259, %r2012;
add.ftz.f32 %f57, %f258, %f259;
cvt.u64.u32 %rd124, %r1991;
add.s64 %rd126, %rd124, %rd108;
shl.b64 %rd128, %rd126, 2;
add.s64 %rd14, %rd110, %rd128;
@%p13 bra $L__BB236_24;

st.global.f32 [%rd14], %f57;

$L__BB236_24:
ld.shared.f32 %f260, [%r1775+1152];
add.ftz.f32 %f261, %f260, %f359;
mov.b32 %r2024, %f261;
shfl.sync.bfly.b32 %r2028|%p72, %r2024, %r1999, %r1998, %r2000;
mov.b32 %f262, %r2028;
add.ftz.f32 %f263, %f261, %f262;
mov.b32 %r2029, %f263;
shfl.sync.bfly.b32 %r2031|%p73, %r2029, %r2003, %r1998, %r2000;
mov.b32 %f264, %r2031;
add.ftz.f32 %f265, %f263, %f264;
mov.b32 %r2032, %f265;
shfl.sync.bfly.b32 %r2034|%p74, %r2032, %r2006, %r1998, %r2000;
mov.b32 %f266, %r2034;
add.ftz.f32 %f267, %f265, %f266;
mov.b32 %r2035, %f267;
shfl.sync.bfly.b32 %r2036|%p75, %r2035, %r1992, %r1998, %r2000;
mov.b32 %f268, %r2036;
add.ftz.f32 %f269, %f267, %f268;
mov.b32 %r2037, %f269;
shfl.sync.bfly.b32 %r2039|%p76, %r2037, %r2011, %r1998, %r2000;
mov.b32 %f270, %r2039;
add.ftz.f32 %f58, %f269, %f270;
@%p21 bra $L__BB236_26;

st.global.f32 [%rd14+4], %f58;

$L__BB236_26:
mul.lo.s32 %r31, %r35, 5;
mov.u32 %r2044, 2;
ld.shared.f32 %f271, [%r1775+1280];
add.ftz.f32 %f272, %f271, %f358;
mov.b32 %r2045, %f272;
mov.u32 %r2046, 31;
mov.u32 %r2047, 16;
mov.u32 %r2048, -1;
shfl.sync.bfly.b32 %r2049|%p78, %r2045, %r2047, %r2046, %r2048;
mov.b32 %f273, %r2049;
add.ftz.f32 %f274, %f272, %f273;
mov.b32 %r2050, %f274;
mov.u32 %r2051, 8;
shfl.sync.bfly.b32 %r2052|%p79, %r2050, %r2051, %r2046, %r2048;
mov.b32 %f275, %r2052;
add.ftz.f32 %f276, %f274, %f275;
mov.b32 %r2053, %f276;
mov.u32 %r2054, 4;
shfl.sync.bfly.b32 %r2055|%p80, %r2053, %r2054, %r2046, %r2048;
mov.b32 %f277, %r2055;
add.ftz.f32 %f278, %f276, %f277;
mov.b32 %r2056, %f278;
shfl.sync.bfly.b32 %r2057|%p81, %r2056, %r2044, %r2046, %r2048;
mov.b32 %f279, %r2057;
add.ftz.f32 %f280, %f278, %f279;
mov.b32 %r2058, %f280;
mov.u32 %r2059, 1;
shfl.sync.bfly.b32 %r2060|%p82, %r2058, %r2059, %r2046, %r2048;
mov.b32 %f281, %r2060;
add.ftz.f32 %f59, %f280, %f281;
@%p13 bra $L__BB236_28;

cvt.u64.u32 %rd129, %r31;
add.s64 %rd131, %rd129, %rd108;
shl.b64 %rd133, %rd131, 2;
add.s64 %rd134, %rd110, %rd133;
st.global.f32 [%rd134], %f59;

$L__BB236_28:
ld.shared.f32 %f282, [%r1775+1408];
add.ftz.f32 %f283, %f282, %f357;
mov.b32 %r2074, %f283;
shfl.sync.bfly.b32 %r2078|%p86, %r2074, %r2047, %r2046, %r2048;
mov.b32 %f284, %r2078;
add.ftz.f32 %f285, %f283, %f284;
mov.b32 %r2079, %f285;
shfl.sync.bfly.b32 %r2081|%p87, %r2079, %r2051, %r2046, %r2048;
mov.b32 %f286, %r2081;
add.ftz.f32 %f287, %f285, %f286;
mov.b32 %r2082, %f287;
shfl.sync.bfly.b32 %r2084|%p88, %r2082, %r2054, %r2046, %r2048;
mov.b32 %f288, %r2084;
add.ftz.f32 %f289, %f287, %f288;
mov.b32 %r2085, %f289;
shfl.sync.bfly.b32 %r2086|%p89, %r2085, %r2044, %r2046, %r2048;
mov.b32 %f290, %r2086;
add.ftz.f32 %f291, %f289, %f290;
mov.b32 %r2087, %f291;
shfl.sync.bfly.b32 %r2089|%p90, %r2087, %r2059, %r2046, %r2048;
mov.b32 %f292, %r2089;
add.ftz.f32 %f60, %f291, %f292;
@%p21 bra $L__BB236_30;

add.s32 %r2090, %r31, 1;
cvt.u64.u32 %rd135, %r2090;
add.s64 %rd137, %rd135, %rd108;
shl.b64 %rd139, %rd137, 2;
add.s64 %rd140, %rd110, %rd139;
st.global.f32 [%rd140], %f60;

$L__BB236_30:
mul.lo.s32 %r2097, %r35, 6;
mov.u32 %r2102, 2;
ld.shared.f32 %f293, [%r1775+1536];
add.ftz.f32 %f294, %f293, %f356;
mov.b32 %r2103, %f294;
mov.u32 %r2104, 31;
mov.u32 %r2105, 16;
mov.u32 %r2106, -1;
shfl.sync.bfly.b32 %r2107|%p92, %r2103, %r2105, %r2104, %r2106;
mov.b32 %f295, %r2107;
add.ftz.f32 %f296, %f294, %f295;
mov.b32 %r2108, %f296;
mov.u32 %r2109, 8;
shfl.sync.bfly.b32 %r2110|%p93, %r2108, %r2109, %r2104, %r2106;
mov.b32 %f297, %r2110;
add.ftz.f32 %f298, %f296, %f297;
mov.b32 %r2111, %f298;
mov.u32 %r2112, 4;
shfl.sync.bfly.b32 %r2113|%p94, %r2111, %r2112, %r2104, %r2106;
mov.b32 %f299, %r2113;
add.ftz.f32 %f300, %f298, %f299;
mov.b32 %r2114, %f300;
shfl.sync.bfly.b32 %r2115|%p95, %r2114, %r2102, %r2104, %r2106;
mov.b32 %f301, %r2115;
add.ftz.f32 %f302, %f300, %f301;
mov.b32 %r2116, %f302;
mov.u32 %r2117, 1;
shfl.sync.bfly.b32 %r2118|%p96, %r2116, %r2117, %r2104, %r2106;
mov.b32 %f303, %r2118;
add.ftz.f32 %f61, %f302, %f303;
cvt.u64.u32 %rd141, %r2097;
add.s64 %rd143, %rd141, %rd108;
shl.b64 %rd145, %rd143, 2;
add.s64 %rd15, %rd110, %rd145;
@%p13 bra $L__BB236_32;

st.global.f32 [%rd15], %f61;

$L__BB236_32:
ld.shared.f32 %f304, [%r1775+1664];
add.ftz.f32 %f305, %f304, %f355;
mov.b32 %r2130, %f305;
shfl.sync.bfly.b32 %r2134|%p100, %r2130, %r2105, %r2104, %r2106;
mov.b32 %f306, %r2134;
add.ftz.f32 %f307, %f305, %f306;
mov.b32 %r2135, %f307;
shfl.sync.bfly.b32 %r2137|%p101, %r2135, %r2109, %r2104, %r2106;
mov.b32 %f308, %r2137;
add.ftz.f32 %f309, %f307, %f308;
mov.b32 %r2138, %f309;
shfl.sync.bfly.b32 %r2140|%p102, %r2138, %r2112, %r2104, %r2106;
mov.b32 %f310, %r2140;
add.ftz.f32 %f311, %f309, %f310;
mov.b32 %r2141, %f311;
shfl.sync.bfly.b32 %r2142|%p103, %r2141, %r2102, %r2104, %r2106;
mov.b32 %f312, %r2142;
add.ftz.f32 %f313, %f311, %f312;
mov.b32 %r2143, %f313;
shfl.sync.bfly.b32 %r2145|%p104, %r2143, %r2117, %r2104, %r2106;
mov.b32 %f314, %r2145;
add.ftz.f32 %f62, %f313, %f314;
@%p21 bra $L__BB236_34;

st.global.f32 [%rd15+4], %f62;

$L__BB236_34:
mul.lo.s32 %r32, %r35, 7;
mov.u32 %r2150, 2;
ld.shared.f32 %f315, [%r1775+1792];
add.ftz.f32 %f316, %f315, %f354;
mov.b32 %r2151, %f316;
mov.u32 %r2152, 31;
mov.u32 %r2153, 16;
mov.u32 %r2154, -1;
shfl.sync.bfly.b32 %r2155|%p106, %r2151, %r2153, %r2152, %r2154;
mov.b32 %f317, %r2155;
add.ftz.f32 %f318, %f316, %f317;
mov.b32 %r2156, %f318;
mov.u32 %r2157, 8;
shfl.sync.bfly.b32 %r2158|%p107, %r2156, %r2157, %r2152, %r2154;
mov.b32 %f319, %r2158;
add.ftz.f32 %f320, %f318, %f319;
mov.b32 %r2159, %f320;
mov.u32 %r2160, 4;
shfl.sync.bfly.b32 %r2161|%p108, %r2159, %r2160, %r2152, %r2154;
mov.b32 %f321, %r2161;
add.ftz.f32 %f322, %f320, %f321;
mov.b32 %r2162, %f322;
shfl.sync.bfly.b32 %r2163|%p109, %r2162, %r2150, %r2152, %r2154;
mov.b32 %f323, %r2163;
add.ftz.f32 %f324, %f322, %f323;
mov.b32 %r2164, %f324;
mov.u32 %r2165, 1;
shfl.sync.bfly.b32 %r2166|%p110, %r2164, %r2165, %r2152, %r2154;
mov.b32 %f325, %r2166;
add.ftz.f32 %f63, %f324, %f325;
@%p13 bra $L__BB236_36;

cvt.u64.u32 %rd146, %r32;
add.s64 %rd148, %rd146, %rd108;
shl.b64 %rd150, %rd148, 2;
add.s64 %rd151, %rd110, %rd150;
st.global.f32 [%rd151], %f63;

$L__BB236_36:
ld.shared.f32 %f326, [%r1775+1920];
add.ftz.f32 %f327, %f326, %f353;
mov.b32 %r2180, %f327;
shfl.sync.bfly.b32 %r2184|%p114, %r2180, %r2153, %r2152, %r2154;
mov.b32 %f328, %r2184;
add.ftz.f32 %f329, %f327, %f328;
mov.b32 %r2185, %f329;
shfl.sync.bfly.b32 %r2187|%p115, %r2185, %r2157, %r2152, %r2154;
mov.b32 %f330, %r2187;
add.ftz.f32 %f331, %f329, %f330;
mov.b32 %r2188, %f331;
shfl.sync.bfly.b32 %r2190|%p116, %r2188, %r2160, %r2152, %r2154;
mov.b32 %f332, %r2190;
add.ftz.f32 %f333, %f331, %f332;
mov.b32 %r2191, %f333;
shfl.sync.bfly.b32 %r2192|%p117, %r2191, %r2150, %r2152, %r2154;
mov.b32 %f334, %r2192;
add.ftz.f32 %f335, %f333, %f334;
mov.b32 %r2193, %f335;
shfl.sync.bfly.b32 %r2195|%p118, %r2193, %r2165, %r2152, %r2154;
mov.b32 %f336, %r2195;
add.ftz.f32 %f64, %f335, %f336;
@%p21 bra $L__BB236_38;

add.s32 %r2196, %r32, 1;
cvt.u64.u32 %rd152, %r2196;
add.s64 %rd154, %rd152, %rd108;
shl.b64 %rd156, %rd154, 2;
add.s64 %rd157, %rd110, %rd156;
st.global.f32 [%rd157], %f64;

$L__BB236_38:
ret;

}
.entry _Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj(
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2,
.param .align 8 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3[48],
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5,
.param .align 4 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6[12],
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15
)
.maxntid 256, 1, 1
.minnctapersm 1
{
.local .align 8 .b8 __local_depot237[16];
.reg .b64 %SP;
.reg .b64 %SPL;
.reg .pred %p<48>;
.reg .b16 %rs<61>;
.reg .f32 %f<169>;
.reg .b32 %r<840>;
.reg .b64 %rd<182>;


mov.u64 %SPL, __local_depot237;
ld.param.u64 %rd20, [_Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0];
ld.param.u64 %rd21, [_Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1];
ld.param.u64 %rd22, [_Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2];
ld.param.u32 %r24, [_Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5];
ld.param.u32 %r29, [_Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8];
ld.param.u32 %r30, [_Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9];
ld.param.u32 %r32, [_Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11];
ld.param.u32 %r33, [_Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12];
ld.param.u32 %r36, [_Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14];
ld.param.u32 %r35, [_Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15];
ld.param.u32 %r27, [_Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+8];
ld.param.u32 %r26, [_Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+4];
ld.param.u32 %r25, [_Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6];
ld.param.u64 %rd24, [_Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+8];
add.u64 %rd1, %SPL, 0;
add.u64 %rd2, %SPL, 8;
mov.u32 %r1, %tid.y;
setp.ge.u32 %p1, %r1, %r36;
@%p1 bra $L__BB237_29;

mov.u32 %r2, %ctaid.y;
cvta.to.global.u64 %rd31, %rd22;
mad.lo.s32 %r37, %r1, %r35, %r2;
mul.wide.u32 %rd32, %r37, 4;
add.s64 %rd33, %rd31, %rd32;
ld.global.nc.u32 %r3, [%rd33];
mov.f32 %f162, 0f00000000;
st.local.v2.f32 [%rd1], {%f162, %f162};
st.local.v2.f32 [%rd2], {%f162, %f162};
mov.u32 %r4, %tid.x;
shr.u32 %r839, %r4, 3;
shr.u32 %r38, %r24, 8;
setp.ge.u32 %p2, %r839, %r38;
mov.f32 %f163, %f162;
mov.f32 %f164, %f162;
mov.f32 %f165, %f162;
@%p2 bra $L__BB237_9;

mov.u32 %r39, %ctaid.x;
shl.b32 %r40, %r39, 1;
mul.lo.s32 %r41, %r3, %r32;
shl.b32 %r42, %r4, 2;
and.b32 %r43, %r42, 28;
cvt.u64.u32 %rd34, %r43;
shl.b32 %r838, %r839, 3;
or.b32 %r44, %r40, 1;
add.s32 %r45, %r839, %r41;
mad.lo.s32 %r46, %r29, %r44, %r45;
mul.wide.s32 %rd4, %r46, 98;
and.b32 %r47, %r4, 7;
mul.wide.u32 %rd35, %r47, 8;
mad.lo.s32 %r48, %r40, %r29, %r839;
add.s32 %r49, %r48, %r41;
mul.wide.s32 %rd5, %r49, 98;
add.s64 %rd36, %rd35, %rd5;
add.s64 %rd6, %rd36, 8;
add.s64 %rd37, %rd34, %rd4;
add.s64 %rd7, %rd37, 66;
add.s64 %rd38, %rd34, %rd5;
add.s64 %rd8, %rd38, 66;
add.s64 %rd39, %rd35, %rd4;
add.s64 %rd9, %rd39, 4;
cvta.to.global.u64 %rd180, %rd20;
cvta.to.global.u64 %rd181, %rd24;
mul.lo.s32 %r50, %r1, %r30;
cvt.u64.u32 %rd40, %r50;
mul.hi.u32 %r51, %r2, %r25;
add.s32 %r52, %r51, %r2;
shr.u32 %r53, %r52, %r26;
mul.lo.s32 %r54, %r53, %r27;
sub.s32 %r55, %r2, %r54;
mul.lo.s32 %r56, %r55, %r33;
cvt.u64.u32 %rd41, %r56;
add.s64 %rd12, %rd41, %rd40;
mov.f32 %f162, 0f00000000;
cvta.to.global.u64 %rd42, %rd21;

$L__BB237_3:
cvt.s64.s32 %rd43, %r838;
add.s64 %rd44, %rd12, %rd43;
add.s64 %rd45, %rd180, %rd6;
ld.global.nc.u16 %rs2, [%rd45+-4];
ld.global.nc.u16 %rs3, [%rd45+-2];
ld.global.nc.u16 %rs4, [%rd45];
add.s64 %rd46, %rd180, %rd8;
ld.global.nc.u16 %rs5, [%rd46];
cvt.u32.u16 %r106, %rs5;
ld.global.nc.u16 %rs6, [%rd46+2];
cvt.u32.u16 %r107, %rs6;
mov.b32 %r108, {%rs5, %rs6};
cvt.u64.u32 %rd47, %r47;
add.s64 %rd48, %rd44, %rd47;
mul.lo.s64 %rd49, %rd48, 36;
ld.global.nc.u16 %rs7, [%rd45+-6];
shl.b16 %rs8, %rs7, 2;
cvt.u64.u16 %rd50, %rs8;
and.b64 %rd51, %rd50, 1020;
mov.u64 %rd52, iq3xxs_grid;
add.s64 %rd53, %rd52, %rd51;
ld.global.nc.u32 %r111, [%rd53];
shr.u16 %rs9, %rs7, 8;
cvt.u32.u16 %r112, %rs9;
mul.wide.u32 %rd54, %r112, 4;
add.s64 %rd55, %rd52, %rd54;
ld.global.nc.u32 %r113, [%rd55];
and.b32 %r114, %r106, 255;
popc.b32 %r115, %r114;
shl.b32 %r116, %r115, 7;
and.b32 %r117, %r116, 128;
xor.b32 %r118, %r117, %r114;
mul.lo.s32 %r119, %r118, 16843009;
and.b32 %r120, %r119, 134480385;
or.b32 %r121, %r120, -2139062144;
add.s32 %r122, %r121, -16843009;
or.b32 %r57, %r122, %r120;

prmt.b32 %r57, %r57, 0, 0xba98;

xor.b32 %r123, %r57, %r111;
or.b32 %r124, %r123, -2139062144;
and.b32 %r125, %r57, 2139062143;
and.b32 %r126, %r111, -2139062144;
sub.s32 %r127, %r124, %r125;
xor.b32 %r128, %r126, %r127;
xor.b32 %r62, %r128, -2139062144;
add.s64 %rd56, %rd42, %rd49;
ld.global.nc.u32 %r9, [%rd56+4];
and.b32 %r129, %r119, -2143281136;
or.b32 %r130, %r129, -2139062144;
add.s32 %r131, %r130, -16843009;
or.b32 %r59, %r131, %r129;

prmt.b32 %r59, %r59, 0, 0xba98;

xor.b32 %r132, %r59, %r113;
or.b32 %r133, %r132, -2139062144;
and.b32 %r134, %r59, 2139062143;
and.b32 %r135, %r113, -2139062144;
sub.s32 %r136, %r133, %r134;
xor.b32 %r137, %r135, %r136;
xor.b32 %r66, %r137, -2139062144;
ld.global.nc.u32 %r10, [%rd56+8];
mov.u32 %r64, 0;

dp4a.s32.s32 %r61, %r62, %r9, %r64;


dp4a.s32.s32 %r65, %r66, %r10, %r61;

shl.b16 %rs10, %rs2, 2;
cvt.u64.u16 %rd57, %rs10;
and.b64 %rd58, %rd57, 1020;
add.s64 %rd59, %rd52, %rd58;
ld.global.nc.u32 %r138, [%rd59];
shr.u16 %rs11, %rs2, 8;
cvt.u32.u16 %r139, %rs11;
mul.wide.u32 %rd60, %r139, 4;
add.s64 %rd61, %rd52, %rd60;
ld.global.nc.u32 %r140, [%rd61];
shr.u32 %r141, %r106, 7;
and.b32 %r142, %r141, 255;
popc.b32 %r143, %r142;
shl.b32 %r144, %r143, 7;
and.b32 %r145, %r144, 128;
xor.b32 %r146, %r145, %r142;
mul.lo.s32 %r147, %r146, 16843009;
and.b32 %r148, %r147, 134480385;
or.b32 %r149, %r148, -2139062144;
add.s32 %r150, %r149, -16843009;
or.b32 %r69, %r150, %r148;

prmt.b32 %r69, %r69, 0, 0xba98;

xor.b32 %r151, %r69, %r138;
or.b32 %r152, %r151, -2139062144;
and.b32 %r153, %r69, 2139062143;
and.b32 %r154, %r138, -2139062144;
sub.s32 %r155, %r152, %r153;
xor.b32 %r156, %r154, %r155;
xor.b32 %r74, %r156, -2139062144;
ld.global.nc.u32 %r11, [%rd56+12];
and.b32 %r157, %r147, -2143281136;
or.b32 %r158, %r157, -2139062144;
add.s32 %r159, %r158, -16843009;
or.b32 %r71, %r159, %r157;

prmt.b32 %r71, %r71, 0, 0xba98;

xor.b32 %r160, %r71, %r140;
or.b32 %r161, %r160, -2139062144;
and.b32 %r162, %r71, 2139062143;
and.b32 %r163, %r140, -2139062144;
sub.s32 %r164, %r161, %r162;
xor.b32 %r165, %r163, %r164;
xor.b32 %r78, %r165, -2139062144;
ld.global.nc.u32 %r12, [%rd56+16];

dp4a.s32.s32 %r73, %r74, %r11, %r65;


dp4a.s32.s32 %r77, %r78, %r12, %r73;

shl.b16 %rs12, %rs3, 2;
cvt.u64.u16 %rd62, %rs12;
and.b64 %rd63, %rd62, 1020;
add.s64 %rd64, %rd52, %rd63;
ld.global.nc.u32 %r166, [%rd64];
shr.u16 %rs13, %rs3, 8;
cvt.u32.u16 %r167, %rs13;
mul.wide.u32 %rd65, %r167, 4;
add.s64 %rd66, %rd52, %rd65;
ld.global.nc.u32 %r168, [%rd66];
shr.u32 %r169, %r108, 14;
and.b32 %r170, %r169, 255;
popc.b32 %r171, %r170;
shl.b32 %r172, %r171, 7;
and.b32 %r173, %r172, 128;
xor.b32 %r174, %r173, %r170;
mul.lo.s32 %r175, %r174, 16843009;
and.b32 %r176, %r175, 134480385;
or.b32 %r177, %r176, -2139062144;
add.s32 %r178, %r177, -16843009;
or.b32 %r81, %r178, %r176;

prmt.b32 %r81, %r81, 0, 0xba98;

xor.b32 %r179, %r81, %r166;
or.b32 %r180, %r179, -2139062144;
and.b32 %r181, %r81, 2139062143;
and.b32 %r182, %r166, -2139062144;
sub.s32 %r183, %r180, %r181;
xor.b32 %r184, %r182, %r183;
xor.b32 %r86, %r184, -2139062144;
ld.global.nc.u32 %r13, [%rd56+20];
and.b32 %r185, %r175, -2143281136;
or.b32 %r186, %r185, -2139062144;
add.s32 %r187, %r186, -16843009;
or.b32 %r83, %r187, %r185;

prmt.b32 %r83, %r83, 0, 0xba98;

xor.b32 %r188, %r83, %r168;
or.b32 %r189, %r188, -2139062144;
and.b32 %r190, %r83, 2139062143;
and.b32 %r191, %r168, -2139062144;
sub.s32 %r192, %r189, %r190;
xor.b32 %r193, %r191, %r192;
xor.b32 %r90, %r193, -2139062144;
ld.global.nc.u32 %r14, [%rd56+24];

dp4a.s32.s32 %r85, %r86, %r13, %r77;


dp4a.s32.s32 %r89, %r90, %r14, %r85;

shl.b16 %rs14, %rs4, 2;
cvt.u64.u16 %rd67, %rs14;
and.b64 %rd68, %rd67, 1020;
add.s64 %rd69, %rd52, %rd68;
ld.global.nc.u32 %r194, [%rd69];
shr.u16 %rs15, %rs4, 8;
cvt.u32.u16 %r195, %rs15;
mul.wide.u32 %rd70, %r195, 4;
add.s64 %rd71, %rd52, %rd70;
ld.global.nc.u32 %r196, [%rd71];
shr.u32 %r197, %r107, 5;
and.b32 %r198, %r197, 255;
popc.b32 %r199, %r198;
shl.b32 %r200, %r199, 7;
and.b32 %r201, %r200, 128;
xor.b32 %r202, %r201, %r198;
mul.lo.s32 %r203, %r202, 16843009;
and.b32 %r204, %r203, 134480385;
or.b32 %r205, %r204, -2139062144;
add.s32 %r206, %r205, -16843009;
or.b32 %r93, %r206, %r204;

prmt.b32 %r93, %r93, 0, 0xba98;

xor.b32 %r207, %r93, %r194;
or.b32 %r208, %r207, -2139062144;
and.b32 %r209, %r93, 2139062143;
and.b32 %r210, %r194, -2139062144;
sub.s32 %r211, %r208, %r209;
xor.b32 %r212, %r210, %r211;
xor.b32 %r98, %r212, -2139062144;
ld.global.nc.u32 %r15, [%rd56+28];
and.b32 %r213, %r203, -2143281136;
or.b32 %r214, %r213, -2139062144;
add.s32 %r215, %r214, -16843009;
or.b32 %r95, %r215, %r213;

prmt.b32 %r95, %r95, 0, 0xba98;

xor.b32 %r216, %r95, %r196;
or.b32 %r217, %r216, -2139062144;
and.b32 %r218, %r95, 2139062143;
and.b32 %r219, %r196, -2139062144;
sub.s32 %r220, %r217, %r218;
xor.b32 %r221, %r219, %r220;
xor.b32 %r102, %r221, -2139062144;
ld.global.nc.u32 %r16, [%rd56+32];

dp4a.s32.s32 %r97, %r98, %r15, %r89;

shr.u32 %r222, %r107, 12;

dp4a.s32.s32 %r101, %r102, %r16, %r97;

shr.u32 %r223, %r101, 31;
add.s32 %r224, %r101, %r223;
shr.s32 %r225, %r224, 1;
mad.lo.s32 %r226, %r101, %r222, %r225;
shr.u32 %r227, %r226, 31;
add.s32 %r228, %r226, %r227;
shr.s32 %r229, %r228, 1;
add.s64 %rd72, %rd180, %rd5;
ld.global.nc.u16 %rs1, [%rd72];

{ cvt.f32.f16 %f42, %rs1;}


ld.global.nc.u32 %r17, [%rd56];

{.reg .f16 low,high;
mov.b32 {low,high},%r17;
cvt.f32.f16 %f43, low;}


mul.ftz.f32 %f44, %f42, %f43;
cvt.rn.f32.s32 %f45, %r229;
fma.rn.ftz.f32 %f165, %f44, %f45, %f165;
setp.eq.s64 %p3, %rd24, 0;
@%p3 bra $L__BB237_5;

mov.u64 %rd164, iq3xxs_grid;
add.s64 %rd73, %rd181, %rd6;
ld.global.u16 %rs17, [%rd73+-4];
ld.global.u16 %rs18, [%rd73+-2];
ld.global.u16 %rs19, [%rd73];
add.s64 %rd74, %rd181, %rd8;
ld.global.u16 %rs20, [%rd74];
cvt.u32.u16 %r279, %rs20;
ld.global.u16 %rs21, [%rd74+2];
cvt.u32.u16 %r280, %rs21;
mov.b32 %r281, {%rs20, %rs21};
ld.global.u16 %rs22, [%rd73+-6];
shl.b16 %rs23, %rs22, 2;
cvt.u64.u16 %rd75, %rs23;
and.b64 %rd76, %rd75, 1020;
add.s64 %rd78, %rd164, %rd76;
ld.global.nc.u32 %r282, [%rd78];
shr.u16 %rs24, %rs22, 8;
cvt.u32.u16 %r283, %rs24;
mul.wide.u32 %rd79, %r283, 4;
add.s64 %rd80, %rd164, %rd79;
ld.global.nc.u32 %r284, [%rd80];
and.b32 %r285, %r279, 255;
popc.b32 %r286, %r285;
shl.b32 %r287, %r286, 7;
and.b32 %r288, %r287, 128;
xor.b32 %r289, %r288, %r285;
mul.lo.s32 %r290, %r289, 16843009;
and.b32 %r291, %r290, 134480385;
or.b32 %r292, %r291, -2139062144;
add.s32 %r293, %r292, -16843009;
or.b32 %r230, %r293, %r291;

prmt.b32 %r230, %r230, 0, 0xba98;

xor.b32 %r294, %r230, %r282;
or.b32 %r295, %r294, -2139062144;
and.b32 %r296, %r230, 2139062143;
and.b32 %r297, %r282, -2139062144;
sub.s32 %r298, %r295, %r296;
xor.b32 %r299, %r297, %r298;
xor.b32 %r235, %r299, -2139062144;
and.b32 %r300, %r290, -2143281136;
or.b32 %r301, %r300, -2139062144;
add.s32 %r302, %r301, -16843009;
or.b32 %r232, %r302, %r300;

prmt.b32 %r232, %r232, 0, 0xba98;

xor.b32 %r303, %r232, %r284;
or.b32 %r304, %r303, -2139062144;
and.b32 %r305, %r232, 2139062143;
and.b32 %r306, %r284, -2139062144;
sub.s32 %r307, %r304, %r305;
xor.b32 %r308, %r306, %r307;
xor.b32 %r239, %r308, -2139062144;

dp4a.s32.s32 %r234, %r235, %r9, %r64;


dp4a.s32.s32 %r238, %r239, %r10, %r234;

shl.b16 %rs25, %rs17, 2;
cvt.u64.u16 %rd81, %rs25;
and.b64 %rd82, %rd81, 1020;
add.s64 %rd83, %rd164, %rd82;
ld.global.nc.u32 %r309, [%rd83];
shr.u16 %rs26, %rs17, 8;
cvt.u32.u16 %r310, %rs26;
mul.wide.u32 %rd84, %r310, 4;
add.s64 %rd85, %rd164, %rd84;
ld.global.nc.u32 %r311, [%rd85];
shr.u32 %r312, %r279, 7;
and.b32 %r313, %r312, 255;
popc.b32 %r314, %r313;
shl.b32 %r315, %r314, 7;
and.b32 %r316, %r315, 128;
xor.b32 %r317, %r316, %r313;
mul.lo.s32 %r318, %r317, 16843009;
and.b32 %r319, %r318, 134480385;
or.b32 %r320, %r319, -2139062144;
add.s32 %r321, %r320, -16843009;
or.b32 %r242, %r321, %r319;

prmt.b32 %r242, %r242, 0, 0xba98;

xor.b32 %r322, %r242, %r309;
or.b32 %r323, %r322, -2139062144;
and.b32 %r324, %r242, 2139062143;
and.b32 %r325, %r309, -2139062144;
sub.s32 %r326, %r323, %r324;
xor.b32 %r327, %r325, %r326;
xor.b32 %r247, %r327, -2139062144;
and.b32 %r328, %r318, -2143281136;
or.b32 %r329, %r328, -2139062144;
add.s32 %r330, %r329, -16843009;
or.b32 %r244, %r330, %r328;

prmt.b32 %r244, %r244, 0, 0xba98;

xor.b32 %r331, %r244, %r311;
or.b32 %r332, %r331, -2139062144;
and.b32 %r333, %r244, 2139062143;
and.b32 %r334, %r311, -2139062144;
sub.s32 %r335, %r332, %r333;
xor.b32 %r336, %r334, %r335;
xor.b32 %r251, %r336, -2139062144;

dp4a.s32.s32 %r246, %r247, %r11, %r238;


dp4a.s32.s32 %r250, %r251, %r12, %r246;

shl.b16 %rs27, %rs18, 2;
cvt.u64.u16 %rd86, %rs27;
and.b64 %rd87, %rd86, 1020;
add.s64 %rd88, %rd164, %rd87;
ld.global.nc.u32 %r337, [%rd88];
shr.u16 %rs28, %rs18, 8;
cvt.u32.u16 %r338, %rs28;
mul.wide.u32 %rd89, %r338, 4;
add.s64 %rd90, %rd164, %rd89;
ld.global.nc.u32 %r339, [%rd90];
shr.u32 %r340, %r281, 14;
and.b32 %r341, %r340, 255;
popc.b32 %r342, %r341;
shl.b32 %r343, %r342, 7;
and.b32 %r344, %r343, 128;
xor.b32 %r345, %r344, %r341;
mul.lo.s32 %r346, %r345, 16843009;
and.b32 %r347, %r346, 134480385;
or.b32 %r348, %r347, -2139062144;
add.s32 %r349, %r348, -16843009;
or.b32 %r254, %r349, %r347;

prmt.b32 %r254, %r254, 0, 0xba98;

xor.b32 %r350, %r254, %r337;
or.b32 %r351, %r350, -2139062144;
and.b32 %r352, %r254, 2139062143;
and.b32 %r353, %r337, -2139062144;
sub.s32 %r354, %r351, %r352;
xor.b32 %r355, %r353, %r354;
xor.b32 %r259, %r355, -2139062144;
and.b32 %r356, %r346, -2143281136;
or.b32 %r357, %r356, -2139062144;
add.s32 %r358, %r357, -16843009;
or.b32 %r256, %r358, %r356;

prmt.b32 %r256, %r256, 0, 0xba98;

xor.b32 %r359, %r256, %r339;
or.b32 %r360, %r359, -2139062144;
and.b32 %r361, %r256, 2139062143;
and.b32 %r362, %r339, -2139062144;
sub.s32 %r363, %r360, %r361;
xor.b32 %r364, %r362, %r363;
xor.b32 %r263, %r364, -2139062144;

dp4a.s32.s32 %r258, %r259, %r13, %r250;


dp4a.s32.s32 %r262, %r263, %r14, %r258;

shl.b16 %rs29, %rs19, 2;
cvt.u64.u16 %rd91, %rs29;
and.b64 %rd92, %rd91, 1020;
add.s64 %rd93, %rd164, %rd92;
ld.global.nc.u32 %r365, [%rd93];
shr.u16 %rs30, %rs19, 8;
cvt.u32.u16 %r366, %rs30;
mul.wide.u32 %rd94, %r366, 4;
add.s64 %rd95, %rd164, %rd94;
ld.global.nc.u32 %r367, [%rd95];
shr.u32 %r368, %r280, 5;
and.b32 %r369, %r368, 255;
popc.b32 %r370, %r369;
shl.b32 %r371, %r370, 7;
and.b32 %r372, %r371, 128;
xor.b32 %r373, %r372, %r369;
mul.lo.s32 %r374, %r373, 16843009;
and.b32 %r375, %r374, 134480385;
or.b32 %r376, %r375, -2139062144;
add.s32 %r377, %r376, -16843009;
or.b32 %r266, %r377, %r375;

prmt.b32 %r266, %r266, 0, 0xba98;

xor.b32 %r378, %r266, %r365;
or.b32 %r379, %r378, -2139062144;
and.b32 %r380, %r266, 2139062143;
and.b32 %r381, %r365, -2139062144;
sub.s32 %r382, %r379, %r380;
xor.b32 %r383, %r381, %r382;
xor.b32 %r271, %r383, -2139062144;
and.b32 %r384, %r374, -2143281136;
or.b32 %r385, %r384, -2139062144;
add.s32 %r386, %r385, -16843009;
or.b32 %r268, %r386, %r384;

prmt.b32 %r268, %r268, 0, 0xba98;

xor.b32 %r387, %r268, %r367;
or.b32 %r388, %r387, -2139062144;
and.b32 %r389, %r268, 2139062143;
and.b32 %r390, %r367, -2139062144;
sub.s32 %r391, %r388, %r389;
xor.b32 %r392, %r390, %r391;
xor.b32 %r275, %r392, -2139062144;

dp4a.s32.s32 %r270, %r271, %r15, %r262;

shr.u32 %r393, %r280, 12;

dp4a.s32.s32 %r274, %r275, %r16, %r270;

shr.u32 %r394, %r274, 31;
add.s32 %r395, %r274, %r394;
shr.s32 %r396, %r395, 1;
mad.lo.s32 %r397, %r274, %r393, %r396;
shr.u32 %r398, %r397, 31;
add.s32 %r399, %r397, %r398;
shr.s32 %r400, %r399, 1;
add.s64 %rd96, %rd181, %rd5;
ld.global.u16 %rs16, [%rd96];

{ cvt.f32.f16 %f46, %rs16;}



{.reg .f16 low,high;
mov.b32 {low,high},%r17;
cvt.f32.f16 %f47, low;}


mul.ftz.f32 %f48, %f46, %f47;
cvt.rn.f32.s32 %f49, %r400;
fma.rn.ftz.f32 %f163, %f48, %f49, %f163;

$L__BB237_5:
mov.u64 %rd158, iq3xxs_grid;
add.s64 %rd97, %rd180, %rd9;
ld.global.nc.u16 %rs32, [%rd97];
ld.global.nc.u16 %rs33, [%rd97+2];
ld.global.nc.u16 %rs34, [%rd97+4];
add.s64 %rd98, %rd180, %rd7;
ld.global.nc.u16 %rs35, [%rd98];
cvt.u32.u16 %r450, %rs35;
ld.global.nc.u16 %rs36, [%rd98+2];
cvt.u32.u16 %r451, %rs36;
mov.b32 %r452, {%rs35, %rs36};
ld.global.nc.u16 %rs37, [%rd97+-2];
shl.b16 %rs38, %rs37, 2;
cvt.u64.u16 %rd99, %rs38;
and.b64 %rd100, %rd99, 1020;
add.s64 %rd102, %rd158, %rd100;
ld.global.nc.u32 %r453, [%rd102];
shr.u16 %rs39, %rs37, 8;
cvt.u32.u16 %r454, %rs39;
mul.wide.u32 %rd103, %r454, 4;
add.s64 %rd104, %rd158, %rd103;
ld.global.nc.u32 %r455, [%rd104];
and.b32 %r456, %r450, 255;
popc.b32 %r457, %r456;
shl.b32 %r458, %r457, 7;
and.b32 %r459, %r458, 128;
xor.b32 %r460, %r459, %r456;
mul.lo.s32 %r461, %r460, 16843009;
and.b32 %r462, %r461, 134480385;
or.b32 %r463, %r462, -2139062144;
add.s32 %r464, %r463, -16843009;
or.b32 %r401, %r464, %r462;

prmt.b32 %r401, %r401, 0, 0xba98;

xor.b32 %r465, %r401, %r453;
or.b32 %r466, %r465, -2139062144;
and.b32 %r467, %r401, 2139062143;
and.b32 %r468, %r453, -2139062144;
sub.s32 %r469, %r466, %r467;
xor.b32 %r470, %r468, %r469;
xor.b32 %r406, %r470, -2139062144;
and.b32 %r471, %r461, -2143281136;
or.b32 %r472, %r471, -2139062144;
add.s32 %r473, %r472, -16843009;
or.b32 %r403, %r473, %r471;

prmt.b32 %r403, %r403, 0, 0xba98;

xor.b32 %r474, %r403, %r455;
or.b32 %r475, %r474, -2139062144;
and.b32 %r476, %r403, 2139062143;
and.b32 %r477, %r455, -2139062144;
sub.s32 %r478, %r475, %r476;
xor.b32 %r479, %r477, %r478;
xor.b32 %r410, %r479, -2139062144;

dp4a.s32.s32 %r405, %r406, %r9, %r64;


dp4a.s32.s32 %r409, %r410, %r10, %r405;

shl.b16 %rs40, %rs32, 2;
cvt.u64.u16 %rd105, %rs40;
and.b64 %rd106, %rd105, 1020;
add.s64 %rd107, %rd158, %rd106;
ld.global.nc.u32 %r480, [%rd107];
shr.u16 %rs41, %rs32, 8;
cvt.u32.u16 %r481, %rs41;
mul.wide.u32 %rd108, %r481, 4;
add.s64 %rd109, %rd158, %rd108;
ld.global.nc.u32 %r482, [%rd109];
shr.u32 %r483, %r450, 7;
and.b32 %r484, %r483, 255;
popc.b32 %r485, %r484;
shl.b32 %r486, %r485, 7;
and.b32 %r487, %r486, 128;
xor.b32 %r488, %r487, %r484;
mul.lo.s32 %r489, %r488, 16843009;
and.b32 %r490, %r489, 134480385;
or.b32 %r491, %r490, -2139062144;
add.s32 %r492, %r491, -16843009;
or.b32 %r413, %r492, %r490;

prmt.b32 %r413, %r413, 0, 0xba98;

xor.b32 %r493, %r413, %r480;
or.b32 %r494, %r493, -2139062144;
and.b32 %r495, %r413, 2139062143;
and.b32 %r496, %r480, -2139062144;
sub.s32 %r497, %r494, %r495;
xor.b32 %r498, %r496, %r497;
xor.b32 %r418, %r498, -2139062144;
and.b32 %r499, %r489, -2143281136;
or.b32 %r500, %r499, -2139062144;
add.s32 %r501, %r500, -16843009;
or.b32 %r415, %r501, %r499;

prmt.b32 %r415, %r415, 0, 0xba98;

xor.b32 %r502, %r415, %r482;
or.b32 %r503, %r502, -2139062144;
and.b32 %r504, %r415, 2139062143;
and.b32 %r505, %r482, -2139062144;
sub.s32 %r506, %r503, %r504;
xor.b32 %r507, %r505, %r506;
xor.b32 %r422, %r507, -2139062144;

dp4a.s32.s32 %r417, %r418, %r11, %r409;


dp4a.s32.s32 %r421, %r422, %r12, %r417;

shl.b16 %rs42, %rs33, 2;
cvt.u64.u16 %rd110, %rs42;
and.b64 %rd111, %rd110, 1020;
add.s64 %rd112, %rd158, %rd111;
ld.global.nc.u32 %r508, [%rd112];
shr.u16 %rs43, %rs33, 8;
cvt.u32.u16 %r509, %rs43;
mul.wide.u32 %rd113, %r509, 4;
add.s64 %rd114, %rd158, %rd113;
ld.global.nc.u32 %r510, [%rd114];
shr.u32 %r511, %r452, 14;
and.b32 %r512, %r511, 255;
popc.b32 %r513, %r512;
shl.b32 %r514, %r513, 7;
and.b32 %r515, %r514, 128;
xor.b32 %r516, %r515, %r512;
mul.lo.s32 %r517, %r516, 16843009;
and.b32 %r518, %r517, 134480385;
or.b32 %r519, %r518, -2139062144;
add.s32 %r520, %r519, -16843009;
or.b32 %r425, %r520, %r518;

prmt.b32 %r425, %r425, 0, 0xba98;

xor.b32 %r521, %r425, %r508;
or.b32 %r522, %r521, -2139062144;
and.b32 %r523, %r425, 2139062143;
and.b32 %r524, %r508, -2139062144;
sub.s32 %r525, %r522, %r523;
xor.b32 %r526, %r524, %r525;
xor.b32 %r430, %r526, -2139062144;
and.b32 %r527, %r517, -2143281136;
or.b32 %r528, %r527, -2139062144;
add.s32 %r529, %r528, -16843009;
or.b32 %r427, %r529, %r527;

prmt.b32 %r427, %r427, 0, 0xba98;

xor.b32 %r530, %r427, %r510;
or.b32 %r531, %r530, -2139062144;
and.b32 %r532, %r427, 2139062143;
and.b32 %r533, %r510, -2139062144;
sub.s32 %r534, %r531, %r532;
xor.b32 %r535, %r533, %r534;
xor.b32 %r434, %r535, -2139062144;

dp4a.s32.s32 %r429, %r430, %r13, %r421;


dp4a.s32.s32 %r433, %r434, %r14, %r429;

shl.b16 %rs44, %rs34, 2;
cvt.u64.u16 %rd115, %rs44;
and.b64 %rd116, %rd115, 1020;
add.s64 %rd117, %rd158, %rd116;
ld.global.nc.u32 %r536, [%rd117];
shr.u16 %rs45, %rs34, 8;
cvt.u32.u16 %r537, %rs45;
mul.wide.u32 %rd118, %r537, 4;
add.s64 %rd119, %rd158, %rd118;
ld.global.nc.u32 %r538, [%rd119];
shr.u32 %r539, %r451, 5;
and.b32 %r540, %r539, 255;
popc.b32 %r541, %r540;
shl.b32 %r542, %r541, 7;
and.b32 %r543, %r542, 128;
xor.b32 %r544, %r543, %r540;
mul.lo.s32 %r545, %r544, 16843009;
and.b32 %r546, %r545, 134480385;
or.b32 %r547, %r546, -2139062144;
add.s32 %r548, %r547, -16843009;
or.b32 %r437, %r548, %r546;

prmt.b32 %r437, %r437, 0, 0xba98;

xor.b32 %r549, %r437, %r536;
or.b32 %r550, %r549, -2139062144;
and.b32 %r551, %r437, 2139062143;
and.b32 %r552, %r536, -2139062144;
sub.s32 %r553, %r550, %r551;
xor.b32 %r554, %r552, %r553;
xor.b32 %r442, %r554, -2139062144;
and.b32 %r555, %r545, -2143281136;
or.b32 %r556, %r555, -2139062144;
add.s32 %r557, %r556, -16843009;
or.b32 %r439, %r557, %r555;

prmt.b32 %r439, %r439, 0, 0xba98;

xor.b32 %r558, %r439, %r538;
or.b32 %r559, %r558, -2139062144;
and.b32 %r560, %r439, 2139062143;
and.b32 %r561, %r538, -2139062144;
sub.s32 %r562, %r559, %r560;
xor.b32 %r563, %r561, %r562;
xor.b32 %r446, %r563, -2139062144;

dp4a.s32.s32 %r441, %r442, %r15, %r433;

shr.u32 %r564, %r451, 12;

dp4a.s32.s32 %r445, %r446, %r16, %r441;

shr.u32 %r565, %r445, 31;
add.s32 %r566, %r445, %r565;
shr.s32 %r567, %r566, 1;
mad.lo.s32 %r568, %r445, %r564, %r567;
shr.u32 %r569, %r568, 31;
add.s32 %r570, %r568, %r569;
shr.s32 %r571, %r570, 1;
add.s64 %rd120, %rd180, %rd4;
ld.global.nc.u16 %rs31, [%rd120];

{ cvt.f32.f16 %f50, %rs31;}



{.reg .f16 low,high;
mov.b32 {low,high},%r17;
cvt.f32.f16 %f51, low;}


mul.ftz.f32 %f52, %f50, %f51;
cvt.rn.f32.s32 %f53, %r571;
fma.rn.ftz.f32 %f164, %f52, %f53, %f164;
@%p3 bra $L__BB237_7;

mov.u64 %rd165, iq3xxs_grid;
add.s64 %rd121, %rd181, %rd9;
ld.global.u16 %rs47, [%rd121];
ld.global.u16 %rs48, [%rd121+2];
ld.global.u16 %rs49, [%rd121+4];
add.s64 %rd122, %rd181, %rd7;
ld.global.u16 %rs50, [%rd122];
cvt.u32.u16 %r621, %rs50;
ld.global.u16 %rs51, [%rd122+2];
cvt.u32.u16 %r622, %rs51;
mov.b32 %r623, {%rs50, %rs51};
ld.global.u16 %rs52, [%rd121+-2];
shl.b16 %rs53, %rs52, 2;
cvt.u64.u16 %rd123, %rs53;
and.b64 %rd124, %rd123, 1020;
add.s64 %rd126, %rd165, %rd124;
ld.global.nc.u32 %r624, [%rd126];
shr.u16 %rs54, %rs52, 8;
cvt.u32.u16 %r625, %rs54;
mul.wide.u32 %rd127, %r625, 4;
add.s64 %rd128, %rd165, %rd127;
ld.global.nc.u32 %r626, [%rd128];
and.b32 %r627, %r621, 255;
popc.b32 %r628, %r627;
shl.b32 %r629, %r628, 7;
and.b32 %r630, %r629, 128;
xor.b32 %r631, %r630, %r627;
mul.lo.s32 %r632, %r631, 16843009;
and.b32 %r633, %r632, 134480385;
or.b32 %r634, %r633, -2139062144;
add.s32 %r635, %r634, -16843009;
or.b32 %r572, %r635, %r633;

prmt.b32 %r572, %r572, 0, 0xba98;

xor.b32 %r636, %r572, %r624;
or.b32 %r637, %r636, -2139062144;
and.b32 %r638, %r572, 2139062143;
and.b32 %r639, %r624, -2139062144;
sub.s32 %r640, %r637, %r638;
xor.b32 %r641, %r639, %r640;
xor.b32 %r577, %r641, -2139062144;
and.b32 %r642, %r632, -2143281136;
or.b32 %r643, %r642, -2139062144;
add.s32 %r644, %r643, -16843009;
or.b32 %r574, %r644, %r642;

prmt.b32 %r574, %r574, 0, 0xba98;

xor.b32 %r645, %r574, %r626;
or.b32 %r646, %r645, -2139062144;
and.b32 %r647, %r574, 2139062143;
and.b32 %r648, %r626, -2139062144;
sub.s32 %r649, %r646, %r647;
xor.b32 %r650, %r648, %r649;
xor.b32 %r581, %r650, -2139062144;
mov.u32 %r579, 0;

dp4a.s32.s32 %r576, %r577, %r9, %r579;


dp4a.s32.s32 %r580, %r581, %r10, %r576;

shl.b16 %rs55, %rs47, 2;
cvt.u64.u16 %rd129, %rs55;
and.b64 %rd130, %rd129, 1020;
add.s64 %rd131, %rd165, %rd130;
ld.global.nc.u32 %r651, [%rd131];
shr.u16 %rs56, %rs47, 8;
cvt.u32.u16 %r652, %rs56;
mul.wide.u32 %rd132, %r652, 4;
add.s64 %rd133, %rd165, %rd132;
ld.global.nc.u32 %r653, [%rd133];
shr.u32 %r654, %r621, 7;
and.b32 %r655, %r654, 255;
popc.b32 %r656, %r655;
shl.b32 %r657, %r656, 7;
and.b32 %r658, %r657, 128;
xor.b32 %r659, %r658, %r655;
mul.lo.s32 %r660, %r659, 16843009;
and.b32 %r661, %r660, 134480385;
or.b32 %r662, %r661, -2139062144;
add.s32 %r663, %r662, -16843009;
or.b32 %r584, %r663, %r661;

prmt.b32 %r584, %r584, 0, 0xba98;

xor.b32 %r664, %r584, %r651;
or.b32 %r665, %r664, -2139062144;
and.b32 %r666, %r584, 2139062143;
and.b32 %r667, %r651, -2139062144;
sub.s32 %r668, %r665, %r666;
xor.b32 %r669, %r667, %r668;
xor.b32 %r589, %r669, -2139062144;
and.b32 %r670, %r660, -2143281136;
or.b32 %r671, %r670, -2139062144;
add.s32 %r672, %r671, -16843009;
or.b32 %r586, %r672, %r670;

prmt.b32 %r586, %r586, 0, 0xba98;

xor.b32 %r673, %r586, %r653;
or.b32 %r674, %r673, -2139062144;
and.b32 %r675, %r586, 2139062143;
and.b32 %r676, %r653, -2139062144;
sub.s32 %r677, %r674, %r675;
xor.b32 %r678, %r676, %r677;
xor.b32 %r593, %r678, -2139062144;

dp4a.s32.s32 %r588, %r589, %r11, %r580;


dp4a.s32.s32 %r592, %r593, %r12, %r588;

shl.b16 %rs57, %rs48, 2;
cvt.u64.u16 %rd134, %rs57;
and.b64 %rd135, %rd134, 1020;
add.s64 %rd136, %rd165, %rd135;
ld.global.nc.u32 %r679, [%rd136];
shr.u16 %rs58, %rs48, 8;
cvt.u32.u16 %r680, %rs58;
mul.wide.u32 %rd137, %r680, 4;
add.s64 %rd138, %rd165, %rd137;
ld.global.nc.u32 %r681, [%rd138];
shr.u32 %r682, %r623, 14;
and.b32 %r683, %r682, 255;
popc.b32 %r684, %r683;
shl.b32 %r685, %r684, 7;
and.b32 %r686, %r685, 128;
xor.b32 %r687, %r686, %r683;
mul.lo.s32 %r688, %r687, 16843009;
and.b32 %r689, %r688, 134480385;
or.b32 %r690, %r689, -2139062144;
add.s32 %r691, %r690, -16843009;
or.b32 %r596, %r691, %r689;

prmt.b32 %r596, %r596, 0, 0xba98;

xor.b32 %r692, %r596, %r679;
or.b32 %r693, %r692, -2139062144;
and.b32 %r694, %r596, 2139062143;
and.b32 %r695, %r679, -2139062144;
sub.s32 %r696, %r693, %r694;
xor.b32 %r697, %r695, %r696;
xor.b32 %r601, %r697, -2139062144;
and.b32 %r698, %r688, -2143281136;
or.b32 %r699, %r698, -2139062144;
add.s32 %r700, %r699, -16843009;
or.b32 %r598, %r700, %r698;

prmt.b32 %r598, %r598, 0, 0xba98;

xor.b32 %r701, %r598, %r681;
or.b32 %r702, %r701, -2139062144;
and.b32 %r703, %r598, 2139062143;
and.b32 %r704, %r681, -2139062144;
sub.s32 %r705, %r702, %r703;
xor.b32 %r706, %r704, %r705;
xor.b32 %r605, %r706, -2139062144;

dp4a.s32.s32 %r600, %r601, %r13, %r592;


dp4a.s32.s32 %r604, %r605, %r14, %r600;

shl.b16 %rs59, %rs49, 2;
cvt.u64.u16 %rd139, %rs59;
and.b64 %rd140, %rd139, 1020;
add.s64 %rd141, %rd165, %rd140;
ld.global.nc.u32 %r707, [%rd141];
shr.u16 %rs60, %rs49, 8;
cvt.u32.u16 %r708, %rs60;
mul.wide.u32 %rd142, %r708, 4;
add.s64 %rd143, %rd165, %rd142;
ld.global.nc.u32 %r709, [%rd143];
shr.u32 %r710, %r622, 5;
and.b32 %r711, %r710, 255;
popc.b32 %r712, %r711;
shl.b32 %r713, %r712, 7;
and.b32 %r714, %r713, 128;
xor.b32 %r715, %r714, %r711;
mul.lo.s32 %r716, %r715, 16843009;
and.b32 %r717, %r716, 134480385;
or.b32 %r718, %r717, -2139062144;
add.s32 %r719, %r718, -16843009;
or.b32 %r608, %r719, %r717;

prmt.b32 %r608, %r608, 0, 0xba98;

xor.b32 %r720, %r608, %r707;
or.b32 %r721, %r720, -2139062144;
and.b32 %r722, %r608, 2139062143;
and.b32 %r723, %r707, -2139062144;
sub.s32 %r724, %r721, %r722;
xor.b32 %r725, %r723, %r724;
xor.b32 %r613, %r725, -2139062144;
and.b32 %r726, %r716, -2143281136;
or.b32 %r727, %r726, -2139062144;
add.s32 %r728, %r727, -16843009;
or.b32 %r610, %r728, %r726;

prmt.b32 %r610, %r610, 0, 0xba98;

xor.b32 %r729, %r610, %r709;
or.b32 %r730, %r729, -2139062144;
and.b32 %r731, %r610, 2139062143;
and.b32 %r732, %r709, -2139062144;
sub.s32 %r733, %r730, %r731;
xor.b32 %r734, %r732, %r733;
xor.b32 %r617, %r734, -2139062144;

dp4a.s32.s32 %r612, %r613, %r15, %r604;

shr.u32 %r735, %r622, 12;

dp4a.s32.s32 %r616, %r617, %r16, %r612;

shr.u32 %r736, %r616, 31;
add.s32 %r737, %r616, %r736;
shr.s32 %r738, %r737, 1;
mad.lo.s32 %r739, %r616, %r735, %r738;
shr.u32 %r740, %r739, 31;
add.s32 %r741, %r739, %r740;
shr.s32 %r742, %r741, 1;
add.s64 %rd144, %rd181, %rd4;
ld.global.u16 %rs46, [%rd144];

{ cvt.f32.f16 %f54, %rs46;}



{.reg .f16 low,high;
mov.b32 {low,high},%r17;
cvt.f32.f16 %f55, low;}


mul.ftz.f32 %f56, %f54, %f55;
cvt.rn.f32.s32 %f57, %r742;
fma.rn.ftz.f32 %f162, %f56, %f57, %f162;

$L__BB237_7:
add.s32 %r838, %r838, 32;
add.s64 %rd181, %rd181, 392;
add.s64 %rd180, %rd180, 392;
add.s32 %r839, %r839, 4;
setp.lt.u32 %p5, %r839, %r38;
@%p5 bra $L__BB237_3;

add.u64 %rd172, %SPL, 0;
add.u64 %rd166, %SPL, 8;
st.local.v2.f32 [%rd172], {%f165, %f164};
st.local.v2.f32 [%rd166], {%f163, %f162};

$L__BB237_9:
mov.b32 %r20, %f165;
setp.ne.s64 %p6, %rd24, 0;
@%p6 bra $L__BB237_11;
bra.uni $L__BB237_10;

$L__BB237_11:
add.u64 %rd178, %SPL, 0;
add.u64 %rd170, %SPL, 8;
mov.u32 %r770, 31;
mov.u32 %r771, 16;
mov.u32 %r772, -1;
shfl.sync.bfly.b32 %r773|%p17, %r20, %r771, %r770, %r772;
mov.b32 %f78, %r773;
add.ftz.f32 %f79, %f165, %f78;
mov.b32 %r774, %f79;
mov.u32 %r775, 8;
shfl.sync.bfly.b32 %r776|%p18, %r774, %r775, %r770, %r772;
mov.b32 %f80, %r776;
add.ftz.f32 %f81, %f79, %f80;
mov.b32 %r777, %f81;
mov.u32 %r778, 4;
shfl.sync.bfly.b32 %r779|%p19, %r777, %r778, %r770, %r772;
mov.b32 %f82, %r779;
add.ftz.f32 %f83, %f81, %f82;
mov.b32 %r780, %f83;
mov.u32 %r781, 2;
shfl.sync.bfly.b32 %r782|%p20, %r780, %r781, %r770, %r772;
mov.b32 %f84, %r782;
add.ftz.f32 %f85, %f83, %f84;
mov.b32 %r783, %f85;
mov.u32 %r784, 1;
shfl.sync.bfly.b32 %r785|%p21, %r783, %r784, %r770, %r772;
mov.b32 %f86, %r785;
add.ftz.f32 %f87, %f85, %f86;
st.local.f32 [%rd178], %f87;
mov.b32 %r786, %f163;
shfl.sync.bfly.b32 %r787|%p22, %r786, %r771, %r770, %r772;
mov.b32 %f88, %r787;
add.ftz.f32 %f89, %f163, %f88;
mov.b32 %r788, %f89;
shfl.sync.bfly.b32 %r789|%p23, %r788, %r775, %r770, %r772;
mov.b32 %f90, %r789;
add.ftz.f32 %f91, %f89, %f90;
mov.b32 %r790, %f91;
shfl.sync.bfly.b32 %r791|%p24, %r790, %r778, %r770, %r772;
mov.b32 %f92, %r791;
add.ftz.f32 %f93, %f91, %f92;
mov.b32 %r792, %f93;
shfl.sync.bfly.b32 %r793|%p25, %r792, %r781, %r770, %r772;
mov.b32 %f94, %r793;
add.ftz.f32 %f95, %f93, %f94;
mov.b32 %r794, %f95;
shfl.sync.bfly.b32 %r795|%p26, %r794, %r784, %r770, %r772;
mov.b32 %f96, %r795;
add.ftz.f32 %f97, %f95, %f96;
st.local.f32 [%rd170], %f97;
mov.b32 %r796, %f164;
shfl.sync.bfly.b32 %r797|%p27, %r796, %r771, %r770, %r772;
mov.b32 %f98, %r797;
add.ftz.f32 %f99, %f164, %f98;
mov.b32 %r798, %f99;
shfl.sync.bfly.b32 %r799|%p28, %r798, %r775, %r770, %r772;
mov.b32 %f100, %r799;
add.ftz.f32 %f101, %f99, %f100;
mov.b32 %r800, %f101;
shfl.sync.bfly.b32 %r801|%p29, %r800, %r778, %r770, %r772;
mov.b32 %f102, %r801;
add.ftz.f32 %f103, %f101, %f102;
mov.b32 %r802, %f103;
shfl.sync.bfly.b32 %r803|%p30, %r802, %r781, %r770, %r772;
mov.b32 %f104, %r803;
add.ftz.f32 %f105, %f103, %f104;
mov.b32 %r804, %f105;
shfl.sync.bfly.b32 %r805|%p31, %r804, %r784, %r770, %r772;
mov.b32 %f106, %r805;
add.ftz.f32 %f107, %f105, %f106;
st.local.f32 [%rd178+4], %f107;
mov.b32 %r806, %f162;
shfl.sync.bfly.b32 %r807|%p32, %r806, %r771, %r770, %r772;
mov.b32 %f108, %r807;
add.ftz.f32 %f109, %f162, %f108;
mov.b32 %r808, %f109;
shfl.sync.bfly.b32 %r809|%p33, %r808, %r775, %r770, %r772;
mov.b32 %f110, %r809;
add.ftz.f32 %f111, %f109, %f110;
mov.b32 %r810, %f111;
shfl.sync.bfly.b32 %r811|%p34, %r810, %r778, %r770, %r772;
mov.b32 %f112, %r811;
add.ftz.f32 %f113, %f111, %f112;
mov.b32 %r812, %f113;
shfl.sync.bfly.b32 %r813|%p35, %r812, %r781, %r770, %r772;
mov.b32 %f114, %r813;
add.ftz.f32 %f115, %f113, %f114;
mov.b32 %r814, %f115;
shfl.sync.bfly.b32 %r815|%p36, %r814, %r784, %r770, %r772;
mov.b32 %f116, %r815;
add.ftz.f32 %f117, %f115, %f116;
st.local.f32 [%rd170+4], %f117;
bra.uni $L__BB237_12;

$L__BB237_10:
add.u64 %rd174, %SPL, 0;
mov.u32 %r744, 31;
mov.u32 %r745, 16;
mov.u32 %r746, -1;
shfl.sync.bfly.b32 %r747|%p7, %r20, %r745, %r744, %r746;
mov.b32 %f58, %r747;
add.ftz.f32 %f59, %f165, %f58;
mov.b32 %r748, %f59;
mov.u32 %r749, 8;
shfl.sync.bfly.b32 %r750|%p8, %r748, %r749, %r744, %r746;
mov.b32 %f60, %r750;
add.ftz.f32 %f61, %f59, %f60;
mov.b32 %r751, %f61;
mov.u32 %r752, 4;
shfl.sync.bfly.b32 %r753|%p9, %r751, %r752, %r744, %r746;
mov.b32 %f62, %r753;
add.ftz.f32 %f63, %f61, %f62;
mov.b32 %r754, %f63;
mov.u32 %r755, 2;
shfl.sync.bfly.b32 %r756|%p10, %r754, %r755, %r744, %r746;
mov.b32 %f64, %r756;
add.ftz.f32 %f65, %f63, %f64;
mov.b32 %r757, %f65;
mov.u32 %r758, 1;
shfl.sync.bfly.b32 %r759|%p11, %r757, %r758, %r744, %r746;
mov.b32 %f66, %r759;
add.ftz.f32 %f67, %f65, %f66;
st.local.f32 [%rd174], %f67;
mov.b32 %r760, %f164;
shfl.sync.bfly.b32 %r761|%p12, %r760, %r745, %r744, %r746;
mov.b32 %f68, %r761;
add.ftz.f32 %f69, %f164, %f68;
mov.b32 %r762, %f69;
shfl.sync.bfly.b32 %r763|%p13, %r762, %r749, %r744, %r746;
mov.b32 %f70, %r763;
add.ftz.f32 %f71, %f69, %f70;
mov.b32 %r764, %f71;
shfl.sync.bfly.b32 %r765|%p14, %r764, %r752, %r744, %r746;
mov.b32 %f72, %r765;
add.ftz.f32 %f73, %f71, %f72;
mov.b32 %r766, %f73;
shfl.sync.bfly.b32 %r767|%p15, %r766, %r755, %r744, %r746;
mov.b32 %f74, %r767;
add.ftz.f32 %f75, %f73, %f74;
mov.b32 %r768, %f75;
shfl.sync.bfly.b32 %r769|%p16, %r768, %r758, %r744, %r746;
mov.b32 %f76, %r769;
add.ftz.f32 %f77, %f75, %f76;
st.local.f32 [%rd174+4], %f77;

$L__BB237_12:
ld.param.u32 %r826, [_Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7];
mov.u32 %r825, %tid.x;
setp.gt.u32 %p37, %r825, 1;
mov.u32 %r817, %ctaid.x;
shl.b32 %r818, %r817, 1;
add.s32 %r21, %r818, %r825;
setp.ge.u32 %p38, %r21, %r826;
or.pred %p39, %p37, %p38;
@%p39 bra $L__BB237_29;

add.u64 %rd176, %SPL, 0;
ld.param.u64 %rd159, [_Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3];
ld.param.u32 %r828, [_Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13];
mov.u32 %r827, %tid.x;
cvt.u64.u32 %rd18, %r827;
mul.wide.u32 %rd145, %r827, 4;
add.s64 %rd146, %rd176, %rd145;
ld.local.f32 %f168, [%rd146];
mad.lo.s32 %r22, %r3, %r828, %r21;
setp.eq.s64 %p40, %rd159, 0;
@%p40 bra $L__BB237_15;

ld.param.u64 %rd160, [_Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3];
cvta.to.global.u64 %rd147, %rd160;
mul.wide.u32 %rd148, %r22, 4;
add.s64 %rd149, %rd147, %rd148;
ld.global.f32 %f118, [%rd149];
add.ftz.f32 %f168, %f168, %f118;

$L__BB237_15:
setp.eq.s64 %p41, %rd24, 0;
@%p41 bra $L__BB237_28;

add.u64 %rd168, %SPL, 8;
ld.param.u64 %rd162, [_Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+16];
shl.b64 %rd150, %rd18, 2;
add.s64 %rd151, %rd168, %rd150;
ld.local.f32 %f24, [%rd151];
setp.eq.s64 %p42, %rd162, 0;
@%p42 bra $L__BB237_18;

ld.param.u64 %rd163, [_Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+16];
cvta.to.global.u64 %rd152, %rd163;
mul.wide.u32 %rd153, %r22, 4;
add.s64 %rd154, %rd152, %rd153;
ld.global.f32 %f119, [%rd154];
add.ftz.f32 %f24, %f24, %f119;

$L__BB237_18:
ld.param.u32 %r833, [_Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.gt.s32 %p43, %r833, 2;
@%p43 bra $L__BB237_22;

ld.param.u32 %r836, [_Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.eq.s32 %p46, %r836, 1;
@%p46 bra $L__BB237_26;

ld.param.u32 %r837, [_Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.eq.s32 %p47, %r837, 2;
@%p47 bra $L__BB237_21;
bra.uni $L__BB237_27;

$L__BB237_21:
mul.ftz.f32 %f139, %f24, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f140, %f139;
add.ftz.f32 %f141, %f140, 0f3F800000;
div.approx.ftz.f32 %f142, %f24, %f141;
mul.ftz.f32 %f168, %f168, %f142;
bra.uni $L__BB237_28;

$L__BB237_22:
ld.param.u32 %r834, [_Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.eq.s32 %p44, %r834, 3;
@%p44 bra $L__BB237_25;

ld.param.u32 %r835, [_Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.ne.s32 %p45, %r835, 6;
@%p45 bra $L__BB237_27;

ld.param.f32 %f151, [_Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+44];
min.ftz.f32 %f120, %f24, %f151;
neg.ftz.f32 %f121, %f151;
min.ftz.f32 %f122, %f168, %f151;
max.ftz.f32 %f123, %f122, %f121;
mul.ftz.f32 %f124, %f120, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f125, %f124;
add.ftz.f32 %f126, %f125, 0f3F800000;
div.approx.ftz.f32 %f127, %f120, %f126;
mul.ftz.f32 %f168, %f123, %f127;
bra.uni $L__BB237_28;

$L__BB237_27:
mul.ftz.f32 %f168, %f168, %f24;
bra.uni $L__BB237_28;

$L__BB237_26:
mul.ftz.f32 %f143, %f24, 0f3F000000;
mul.ftz.f32 %f144, %f24, 0f3D372713;
fma.rn.ftz.f32 %f145, %f24, %f144, 0f3F800000;
mul.ftz.f32 %f146, %f24, 0f3F4C422A;
mul.ftz.f32 %f147, %f146, %f145;
tanh.approx.f32 %f148, %f147;
add.ftz.f32 %f149, %f148, 0f3F800000;
mul.ftz.f32 %f150, %f143, %f149;
mul.ftz.f32 %f168, %f168, %f150;
bra.uni $L__BB237_28;

$L__BB237_25:
mov.f32 %f128, 0f40E00000;
min.ftz.f32 %f129, %f24, %f128;
min.ftz.f32 %f130, %f168, %f128;
mov.f32 %f131, 0fC0E00000;
max.ftz.f32 %f132, %f130, %f131;
mul.ftz.f32 %f133, %f129, 0f3FD9DB23;
mul.ftz.f32 %f134, %f133, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f135, %f134;
add.ftz.f32 %f136, %f135, 0f3F800000;
div.approx.ftz.f32 %f137, %f129, %f136;
add.ftz.f32 %f138, %f132, 0f3F800000;
mul.ftz.f32 %f168, %f138, %f137;

$L__BB237_28:
ld.param.u64 %rd161, [_Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4];
ld.param.u32 %r832, [_Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10];
ld.param.u32 %r831, [_Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13];
mov.u32 %r830, %ctaid.y;
mov.u32 %r829, %tid.y;
mad.lo.s32 %r823, %r829, %r832, %r21;
mad.lo.s32 %r824, %r830, %r831, %r823;
cvta.to.global.u64 %rd155, %rd161;
mul.wide.u32 %rd156, %r824, 4;
add.s64 %rd157, %rd155, %rd156;
st.global.f32 [%rd157], %f168;

$L__BB237_29:
ret;

}
.entry _Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj(
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2,
.param .align 8 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3[48],
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5,
.param .align 4 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6[12],
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15
)
.maxntid 256, 1, 1
.minnctapersm 1
{
.local .align 8 .b8 __local_depot238[8];
.reg .b64 %SP;
.reg .b64 %SPL;
.reg .pred %p<17>;
.reg .b16 %rs<31>;
.reg .f32 %f<44>;
.reg .b32 %r<425>;
.reg .b64 %rd<88>;


mov.u64 %SPL, __local_depot238;
ld.param.u64 %rd14, [_Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0];
ld.param.u64 %rd15, [_Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1];
ld.param.u64 %rd16, [_Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2];
ld.param.u64 %rd17, [_Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4];
ld.param.u32 %r21, [_Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5];
ld.param.u32 %r13, [_Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7];
ld.param.u32 %r14, [_Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8];
ld.param.u32 %r15, [_Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9];
ld.param.u32 %r16, [_Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10];
ld.param.u32 %r17, [_Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11];
ld.param.u32 %r18, [_Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12];
ld.param.u32 %r19, [_Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13];
ld.param.u32 %r22, [_Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14];
ld.param.u32 %r20, [_Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15];
ld.param.u32 %r12, [_Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+8];
ld.param.u32 %r11, [_Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+4];
ld.param.u32 %r10, [_Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6];
add.u64 %rd1, %SPL, 0;
shr.u32 %r1, %r21, 8;
mov.u32 %r2, %tid.y;
setp.ge.u32 %p1, %r2, %r22;
@%p1 bra $L__BB238_7;

mov.f32 %f42, 0f00000000;
st.local.v2.f32 [%rd1], {%f42, %f42};
mov.u32 %r23, %tid.x;
shr.u32 %r424, %r23, 3;
setp.ge.u32 %p2, %r424, %r1;
mov.f32 %f43, %f42;
@%p2 bra $L__BB238_5;

mov.u32 %r24, %ctaid.x;
shl.b32 %r25, %r24, 1;
mov.u32 %r26, %ctaid.y;
mad.lo.s32 %r27, %r2, %r20, %r26;
cvta.to.global.u64 %rd19, %rd16;
mul.wide.u32 %rd20, %r27, 4;
add.s64 %rd21, %rd19, %rd20;
ld.global.nc.u32 %r28, [%rd21];
mul.lo.s32 %r29, %r28, %r17;
shl.b32 %r31, %r23, 2;
and.b32 %r32, %r31, 28;
cvt.u64.u32 %rd22, %r32;
and.b32 %r33, %r23, 7;
cvt.u64.u32 %rd2, %r33;
shl.b32 %r423, %r424, 3;
or.b32 %r34, %r25, 1;
add.s32 %r35, %r424, %r29;
mad.lo.s32 %r36, %r14, %r34, %r35;
mul.wide.s32 %rd3, %r36, 98;
mul.wide.u32 %rd23, %r33, 8;
mad.lo.s32 %r37, %r25, %r14, %r424;
add.s32 %r38, %r37, %r29;
mul.wide.s32 %rd4, %r38, 98;
add.s64 %rd24, %rd23, %rd4;
add.s64 %rd5, %rd24, 8;
add.s64 %rd25, %rd22, %rd3;
add.s64 %rd6, %rd25, 66;
add.s64 %rd26, %rd22, %rd4;
add.s64 %rd7, %rd26, 66;
add.s64 %rd27, %rd23, %rd3;
add.s64 %rd8, %rd27, 4;
cvta.to.global.u64 %rd87, %rd14;
mul.lo.s32 %r39, %r2, %r15;
cvt.u64.u32 %rd28, %r39;
mul.hi.u32 %r40, %r26, %r10;
add.s32 %r41, %r40, %r26;
shr.u32 %r42, %r41, %r11;
mul.lo.s32 %r43, %r42, %r12;
sub.s32 %r44, %r26, %r43;
mul.lo.s32 %r45, %r44, %r18;
cvt.u64.u32 %rd29, %r45;
add.s64 %rd10, %rd29, %rd28;
cvta.to.global.u64 %rd11, %rd15;
mov.f32 %f42, 0f00000000;

$L__BB238_3:
cvt.s64.s32 %rd30, %r423;
add.s64 %rd31, %rd10, %rd30;
add.s64 %rd32, %rd31, %rd2;
mul.lo.s64 %rd33, %rd32, 36;
add.s64 %rd34, %rd87, %rd5;
ld.global.nc.u16 %rs3, [%rd34+-4];
ld.global.nc.u16 %rs4, [%rd34+-2];
ld.global.nc.u16 %rs5, [%rd34];
add.s64 %rd35, %rd87, %rd7;
ld.global.nc.u16 %rs6, [%rd35];
cvt.u32.u16 %r144, %rs6;
ld.global.nc.u16 %rs7, [%rd35+2];
cvt.u32.u16 %r145, %rs7;
mov.b32 %r146, {%rs6, %rs7};
ld.global.nc.u16 %rs8, [%rd34+-6];
shl.b16 %rs9, %rs8, 2;
cvt.u64.u16 %rd36, %rs9;
and.b64 %rd37, %rd36, 1020;
mov.u64 %rd38, iq3xxs_grid;
add.s64 %rd39, %rd38, %rd37;
ld.global.nc.u32 %r147, [%rd39];
shr.u16 %rs10, %rs8, 8;
cvt.u32.u16 %r148, %rs10;
mul.wide.u32 %rd40, %r148, 4;
add.s64 %rd41, %rd38, %rd40;
ld.global.nc.u32 %r149, [%rd41];
and.b32 %r150, %r144, 255;
popc.b32 %r151, %r150;
shl.b32 %r152, %r151, 7;
and.b32 %r153, %r152, 128;
xor.b32 %r154, %r153, %r150;
mul.lo.s32 %r155, %r154, 16843009;
and.b32 %r156, %r155, 134480385;
or.b32 %r157, %r156, -2139062144;
add.s32 %r158, %r157, -16843009;
or.b32 %r46, %r158, %r156;

prmt.b32 %r46, %r46, 0, 0xba98;

xor.b32 %r159, %r46, %r147;
or.b32 %r160, %r159, -2139062144;
and.b32 %r161, %r46, 2139062143;
and.b32 %r162, %r147, -2139062144;
sub.s32 %r163, %r160, %r161;
xor.b32 %r164, %r162, %r163;
xor.b32 %r51, %r164, -2139062144;
add.s64 %rd42, %rd11, %rd33;
ld.global.nc.u32 %r52, [%rd42+4];
and.b32 %r165, %r155, -2143281136;
or.b32 %r166, %r165, -2139062144;
add.s32 %r167, %r166, -16843009;
or.b32 %r48, %r167, %r165;

prmt.b32 %r48, %r48, 0, 0xba98;

xor.b32 %r168, %r48, %r149;
or.b32 %r169, %r168, -2139062144;
and.b32 %r170, %r48, 2139062143;
and.b32 %r171, %r149, -2139062144;
sub.s32 %r172, %r169, %r170;
xor.b32 %r173, %r171, %r172;
xor.b32 %r55, %r173, -2139062144;
ld.global.nc.u32 %r56, [%rd42+8];
mov.u32 %r102, 0;

dp4a.s32.s32 %r50, %r51, %r52, %r102;


dp4a.s32.s32 %r54, %r55, %r56, %r50;

shl.b16 %rs11, %rs3, 2;
cvt.u64.u16 %rd43, %rs11;
and.b64 %rd44, %rd43, 1020;
add.s64 %rd45, %rd38, %rd44;
ld.global.nc.u32 %r174, [%rd45];
shr.u16 %rs12, %rs3, 8;
cvt.u32.u16 %r175, %rs12;
mul.wide.u32 %rd46, %r175, 4;
add.s64 %rd47, %rd38, %rd46;
ld.global.nc.u32 %r176, [%rd47];
shr.u32 %r177, %r144, 7;
and.b32 %r178, %r177, 255;
popc.b32 %r179, %r178;
shl.b32 %r180, %r179, 7;
and.b32 %r181, %r180, 128;
xor.b32 %r182, %r181, %r178;
mul.lo.s32 %r183, %r182, 16843009;
and.b32 %r184, %r183, 134480385;
or.b32 %r185, %r184, -2139062144;
add.s32 %r186, %r185, -16843009;
or.b32 %r58, %r186, %r184;

prmt.b32 %r58, %r58, 0, 0xba98;

xor.b32 %r187, %r58, %r174;
or.b32 %r188, %r187, -2139062144;
and.b32 %r189, %r58, 2139062143;
and.b32 %r190, %r174, -2139062144;
sub.s32 %r191, %r188, %r189;
xor.b32 %r192, %r190, %r191;
xor.b32 %r63, %r192, -2139062144;
ld.global.nc.u32 %r64, [%rd42+12];
and.b32 %r193, %r183, -2143281136;
or.b32 %r194, %r193, -2139062144;
add.s32 %r195, %r194, -16843009;
or.b32 %r60, %r195, %r193;

prmt.b32 %r60, %r60, 0, 0xba98;

xor.b32 %r196, %r60, %r176;
or.b32 %r197, %r196, -2139062144;
and.b32 %r198, %r60, 2139062143;
and.b32 %r199, %r176, -2139062144;
sub.s32 %r200, %r197, %r198;
xor.b32 %r201, %r199, %r200;
xor.b32 %r67, %r201, -2139062144;
ld.global.nc.u32 %r68, [%rd42+16];

dp4a.s32.s32 %r62, %r63, %r64, %r54;


dp4a.s32.s32 %r66, %r67, %r68, %r62;

shl.b16 %rs13, %rs4, 2;
cvt.u64.u16 %rd48, %rs13;
and.b64 %rd49, %rd48, 1020;
add.s64 %rd50, %rd38, %rd49;
ld.global.nc.u32 %r202, [%rd50];
shr.u16 %rs14, %rs4, 8;
cvt.u32.u16 %r203, %rs14;
mul.wide.u32 %rd51, %r203, 4;
add.s64 %rd52, %rd38, %rd51;
ld.global.nc.u32 %r204, [%rd52];
shr.u32 %r205, %r146, 14;
and.b32 %r206, %r205, 255;
popc.b32 %r207, %r206;
shl.b32 %r208, %r207, 7;
and.b32 %r209, %r208, 128;
xor.b32 %r210, %r209, %r206;
mul.lo.s32 %r211, %r210, 16843009;
and.b32 %r212, %r211, 134480385;
or.b32 %r213, %r212, -2139062144;
add.s32 %r214, %r213, -16843009;
or.b32 %r70, %r214, %r212;

prmt.b32 %r70, %r70, 0, 0xba98;

xor.b32 %r215, %r70, %r202;
or.b32 %r216, %r215, -2139062144;
and.b32 %r217, %r70, 2139062143;
and.b32 %r218, %r202, -2139062144;
sub.s32 %r219, %r216, %r217;
xor.b32 %r220, %r218, %r219;
xor.b32 %r75, %r220, -2139062144;
ld.global.nc.u32 %r76, [%rd42+20];
and.b32 %r221, %r211, -2143281136;
or.b32 %r222, %r221, -2139062144;
add.s32 %r223, %r222, -16843009;
or.b32 %r72, %r223, %r221;

prmt.b32 %r72, %r72, 0, 0xba98;

xor.b32 %r224, %r72, %r204;
or.b32 %r225, %r224, -2139062144;
and.b32 %r226, %r72, 2139062143;
and.b32 %r227, %r204, -2139062144;
sub.s32 %r228, %r225, %r226;
xor.b32 %r229, %r227, %r228;
xor.b32 %r79, %r229, -2139062144;
ld.global.nc.u32 %r80, [%rd42+24];

dp4a.s32.s32 %r74, %r75, %r76, %r66;


dp4a.s32.s32 %r78, %r79, %r80, %r74;

shl.b16 %rs15, %rs5, 2;
cvt.u64.u16 %rd53, %rs15;
and.b64 %rd54, %rd53, 1020;
add.s64 %rd55, %rd38, %rd54;
ld.global.nc.u32 %r230, [%rd55];
shr.u16 %rs16, %rs5, 8;
cvt.u32.u16 %r231, %rs16;
mul.wide.u32 %rd56, %r231, 4;
add.s64 %rd57, %rd38, %rd56;
ld.global.nc.u32 %r232, [%rd57];
shr.u32 %r233, %r145, 5;
and.b32 %r234, %r233, 255;
popc.b32 %r235, %r234;
shl.b32 %r236, %r235, 7;
and.b32 %r237, %r236, 128;
xor.b32 %r238, %r237, %r234;
mul.lo.s32 %r239, %r238, 16843009;
and.b32 %r240, %r239, 134480385;
or.b32 %r241, %r240, -2139062144;
add.s32 %r242, %r241, -16843009;
or.b32 %r82, %r242, %r240;

prmt.b32 %r82, %r82, 0, 0xba98;

xor.b32 %r243, %r82, %r230;
or.b32 %r244, %r243, -2139062144;
and.b32 %r245, %r82, 2139062143;
and.b32 %r246, %r230, -2139062144;
sub.s32 %r247, %r244, %r245;
xor.b32 %r248, %r246, %r247;
xor.b32 %r87, %r248, -2139062144;
ld.global.nc.u32 %r88, [%rd42+28];
and.b32 %r249, %r239, -2143281136;
or.b32 %r250, %r249, -2139062144;
add.s32 %r251, %r250, -16843009;
or.b32 %r84, %r251, %r249;

prmt.b32 %r84, %r84, 0, 0xba98;

xor.b32 %r252, %r84, %r232;
or.b32 %r253, %r252, -2139062144;
and.b32 %r254, %r84, 2139062143;
and.b32 %r255, %r232, -2139062144;
sub.s32 %r256, %r253, %r254;
xor.b32 %r257, %r255, %r256;
xor.b32 %r91, %r257, -2139062144;
ld.global.nc.u32 %r92, [%rd42+32];

dp4a.s32.s32 %r86, %r87, %r88, %r78;

shr.u32 %r258, %r145, 12;

dp4a.s32.s32 %r90, %r91, %r92, %r86;

shr.u32 %r259, %r90, 31;
add.s32 %r260, %r90, %r259;
shr.s32 %r261, %r260, 1;
mad.lo.s32 %r262, %r90, %r258, %r261;
shr.u32 %r263, %r262, 31;
add.s32 %r264, %r262, %r263;
shr.s32 %r265, %r264, 1;
add.s64 %rd58, %rd87, %rd4;
ld.global.nc.u16 %rs1, [%rd58];

{ cvt.f32.f16 %f11, %rs1;}


ld.global.nc.u32 %r94, [%rd42];

{.reg .f16 low,high;
mov.b32 {low,high},%r94;
cvt.f32.f16 %f12, low;}


mul.ftz.f32 %f15, %f11, %f12;
cvt.rn.f32.s32 %f16, %r265;
fma.rn.ftz.f32 %f43, %f15, %f16, %f43;
add.s64 %rd59, %rd87, %rd8;
ld.global.nc.u16 %rs17, [%rd59];
ld.global.nc.u16 %rs18, [%rd59+2];
ld.global.nc.u16 %rs19, [%rd59+4];
add.s64 %rd60, %rd87, %rd6;
ld.global.nc.u16 %rs20, [%rd60];
cvt.u32.u16 %r266, %rs20;
ld.global.nc.u16 %rs21, [%rd60+2];
cvt.u32.u16 %r267, %rs21;
mov.b32 %r268, {%rs20, %rs21};
ld.global.nc.u16 %rs22, [%rd59+-2];
shl.b16 %rs23, %rs22, 2;
cvt.u64.u16 %rd61, %rs23;
and.b64 %rd62, %rd61, 1020;
add.s64 %rd63, %rd38, %rd62;
ld.global.nc.u32 %r269, [%rd63];
shr.u16 %rs24, %rs22, 8;
cvt.u32.u16 %r270, %rs24;
mul.wide.u32 %rd64, %r270, 4;
add.s64 %rd65, %rd38, %rd64;
ld.global.nc.u32 %r271, [%rd65];
and.b32 %r272, %r266, 255;
popc.b32 %r273, %r272;
shl.b32 %r274, %r273, 7;
and.b32 %r275, %r274, 128;
xor.b32 %r276, %r275, %r272;
mul.lo.s32 %r277, %r276, 16843009;
and.b32 %r278, %r277, 134480385;
or.b32 %r279, %r278, -2139062144;
add.s32 %r280, %r279, -16843009;
or.b32 %r95, %r280, %r278;

prmt.b32 %r95, %r95, 0, 0xba98;

xor.b32 %r281, %r95, %r269;
or.b32 %r282, %r281, -2139062144;
and.b32 %r283, %r95, 2139062143;
and.b32 %r284, %r269, -2139062144;
sub.s32 %r285, %r282, %r283;
xor.b32 %r286, %r284, %r285;
xor.b32 %r100, %r286, -2139062144;
and.b32 %r287, %r277, -2143281136;
or.b32 %r288, %r287, -2139062144;
add.s32 %r289, %r288, -16843009;
or.b32 %r97, %r289, %r287;

prmt.b32 %r97, %r97, 0, 0xba98;

xor.b32 %r290, %r97, %r271;
or.b32 %r291, %r290, -2139062144;
and.b32 %r292, %r97, 2139062143;
and.b32 %r293, %r271, -2139062144;
sub.s32 %r294, %r291, %r292;
xor.b32 %r295, %r293, %r294;
xor.b32 %r104, %r295, -2139062144;

dp4a.s32.s32 %r99, %r100, %r52, %r102;


dp4a.s32.s32 %r103, %r104, %r56, %r99;

shl.b16 %rs25, %rs17, 2;
cvt.u64.u16 %rd66, %rs25;
and.b64 %rd67, %rd66, 1020;
add.s64 %rd68, %rd38, %rd67;
ld.global.nc.u32 %r296, [%rd68];
shr.u16 %rs26, %rs17, 8;
cvt.u32.u16 %r297, %rs26;
mul.wide.u32 %rd69, %r297, 4;
add.s64 %rd70, %rd38, %rd69;
ld.global.nc.u32 %r298, [%rd70];
shr.u32 %r299, %r266, 7;
and.b32 %r300, %r299, 255;
popc.b32 %r301, %r300;
shl.b32 %r302, %r301, 7;
and.b32 %r303, %r302, 128;
xor.b32 %r304, %r303, %r300;
mul.lo.s32 %r305, %r304, 16843009;
and.b32 %r306, %r305, 134480385;
or.b32 %r307, %r306, -2139062144;
add.s32 %r308, %r307, -16843009;
or.b32 %r107, %r308, %r306;

prmt.b32 %r107, %r107, 0, 0xba98;

xor.b32 %r309, %r107, %r296;
or.b32 %r310, %r309, -2139062144;
and.b32 %r311, %r107, 2139062143;
and.b32 %r312, %r296, -2139062144;
sub.s32 %r313, %r310, %r311;
xor.b32 %r314, %r312, %r313;
xor.b32 %r112, %r314, -2139062144;
and.b32 %r315, %r305, -2143281136;
or.b32 %r316, %r315, -2139062144;
add.s32 %r317, %r316, -16843009;
or.b32 %r109, %r317, %r315;

prmt.b32 %r109, %r109, 0, 0xba98;

xor.b32 %r318, %r109, %r298;
or.b32 %r319, %r318, -2139062144;
and.b32 %r320, %r109, 2139062143;
and.b32 %r321, %r298, -2139062144;
sub.s32 %r322, %r319, %r320;
xor.b32 %r323, %r321, %r322;
xor.b32 %r116, %r323, -2139062144;

dp4a.s32.s32 %r111, %r112, %r64, %r103;


dp4a.s32.s32 %r115, %r116, %r68, %r111;

shl.b16 %rs27, %rs18, 2;
cvt.u64.u16 %rd71, %rs27;
and.b64 %rd72, %rd71, 1020;
add.s64 %rd73, %rd38, %rd72;
ld.global.nc.u32 %r324, [%rd73];
shr.u16 %rs28, %rs18, 8;
cvt.u32.u16 %r325, %rs28;
mul.wide.u32 %rd74, %r325, 4;
add.s64 %rd75, %rd38, %rd74;
ld.global.nc.u32 %r326, [%rd75];
shr.u32 %r327, %r268, 14;
and.b32 %r328, %r327, 255;
popc.b32 %r329, %r328;
shl.b32 %r330, %r329, 7;
and.b32 %r331, %r330, 128;
xor.b32 %r332, %r331, %r328;
mul.lo.s32 %r333, %r332, 16843009;
and.b32 %r334, %r333, 134480385;
or.b32 %r335, %r334, -2139062144;
add.s32 %r336, %r335, -16843009;
or.b32 %r119, %r336, %r334;

prmt.b32 %r119, %r119, 0, 0xba98;

xor.b32 %r337, %r119, %r324;
or.b32 %r338, %r337, -2139062144;
and.b32 %r339, %r119, 2139062143;
and.b32 %r340, %r324, -2139062144;
sub.s32 %r341, %r338, %r339;
xor.b32 %r342, %r340, %r341;
xor.b32 %r124, %r342, -2139062144;
and.b32 %r343, %r333, -2143281136;
or.b32 %r344, %r343, -2139062144;
add.s32 %r345, %r344, -16843009;
or.b32 %r121, %r345, %r343;

prmt.b32 %r121, %r121, 0, 0xba98;

xor.b32 %r346, %r121, %r326;
or.b32 %r347, %r346, -2139062144;
and.b32 %r348, %r121, 2139062143;
and.b32 %r349, %r326, -2139062144;
sub.s32 %r350, %r347, %r348;
xor.b32 %r351, %r349, %r350;
xor.b32 %r128, %r351, -2139062144;

dp4a.s32.s32 %r123, %r124, %r76, %r115;


dp4a.s32.s32 %r127, %r128, %r80, %r123;

shl.b16 %rs29, %rs19, 2;
cvt.u64.u16 %rd76, %rs29;
and.b64 %rd77, %rd76, 1020;
add.s64 %rd78, %rd38, %rd77;
ld.global.nc.u32 %r352, [%rd78];
shr.u16 %rs30, %rs19, 8;
cvt.u32.u16 %r353, %rs30;
mul.wide.u32 %rd79, %r353, 4;
add.s64 %rd80, %rd38, %rd79;
ld.global.nc.u32 %r354, [%rd80];
shr.u32 %r355, %r267, 5;
and.b32 %r356, %r355, 255;
popc.b32 %r357, %r356;
shl.b32 %r358, %r357, 7;
and.b32 %r359, %r358, 128;
xor.b32 %r360, %r359, %r356;
mul.lo.s32 %r361, %r360, 16843009;
and.b32 %r362, %r361, 134480385;
or.b32 %r363, %r362, -2139062144;
add.s32 %r364, %r363, -16843009;
or.b32 %r131, %r364, %r362;

prmt.b32 %r131, %r131, 0, 0xba98;

xor.b32 %r365, %r131, %r352;
or.b32 %r366, %r365, -2139062144;
and.b32 %r367, %r131, 2139062143;
and.b32 %r368, %r352, -2139062144;
sub.s32 %r369, %r366, %r367;
xor.b32 %r370, %r368, %r369;
xor.b32 %r136, %r370, -2139062144;
and.b32 %r371, %r361, -2143281136;
or.b32 %r372, %r371, -2139062144;
add.s32 %r373, %r372, -16843009;
or.b32 %r133, %r373, %r371;

prmt.b32 %r133, %r133, 0, 0xba98;

xor.b32 %r374, %r133, %r354;
or.b32 %r375, %r374, -2139062144;
and.b32 %r376, %r133, 2139062143;
and.b32 %r377, %r354, -2139062144;
sub.s32 %r378, %r375, %r376;
xor.b32 %r379, %r377, %r378;
xor.b32 %r140, %r379, -2139062144;

dp4a.s32.s32 %r135, %r136, %r88, %r127;

shr.u32 %r380, %r267, 12;

dp4a.s32.s32 %r139, %r140, %r92, %r135;

shr.u32 %r381, %r139, 31;
add.s32 %r382, %r139, %r381;
shr.s32 %r383, %r382, 1;
mad.lo.s32 %r384, %r139, %r380, %r383;
shr.u32 %r385, %r384, 31;
add.s32 %r386, %r384, %r385;
shr.s32 %r387, %r386, 1;
add.s64 %rd81, %rd87, %rd3;
ld.global.nc.u16 %rs2, [%rd81];

{ cvt.f32.f16 %f13, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r94;
cvt.f32.f16 %f14, low;}


mul.ftz.f32 %f17, %f13, %f14;
cvt.rn.f32.s32 %f18, %r387;
fma.rn.ftz.f32 %f42, %f17, %f18, %f42;
add.s32 %r423, %r423, 32;
add.s64 %rd87, %rd87, 392;
add.s32 %r424, %r424, 4;
setp.lt.u32 %p3, %r424, %r1;
@%p3 bra $L__BB238_3;

st.local.v2.f32 [%rd1], {%f43, %f42};

$L__BB238_5:
mov.b32 %r388, %f43;
mov.u32 %r389, 31;
mov.u32 %r390, 16;
mov.u32 %r391, -1;
shfl.sync.bfly.b32 %r392|%p4, %r388, %r390, %r389, %r391;
mov.b32 %f19, %r392;
add.ftz.f32 %f20, %f43, %f19;
mov.b32 %r393, %f20;
mov.u32 %r394, 8;
shfl.sync.bfly.b32 %r395|%p5, %r393, %r394, %r389, %r391;
mov.b32 %f21, %r395;
add.ftz.f32 %f22, %f20, %f21;
mov.b32 %r396, %f22;
mov.u32 %r397, 4;
shfl.sync.bfly.b32 %r398|%p6, %r396, %r397, %r389, %r391;
mov.b32 %f23, %r398;
add.ftz.f32 %f24, %f22, %f23;
mov.b32 %r399, %f24;
mov.u32 %r400, 2;
shfl.sync.bfly.b32 %r401|%p7, %r399, %r400, %r389, %r391;
mov.b32 %f25, %r401;
add.ftz.f32 %f26, %f24, %f25;
mov.b32 %r402, %f26;
mov.u32 %r403, 1;
shfl.sync.bfly.b32 %r404|%p8, %r402, %r403, %r389, %r391;
mov.b32 %f27, %r404;
add.ftz.f32 %f28, %f26, %f27;
st.local.f32 [%rd1], %f28;
mov.b32 %r405, %f42;
shfl.sync.bfly.b32 %r406|%p9, %r405, %r390, %r389, %r391;
mov.b32 %f29, %r406;
add.ftz.f32 %f30, %f42, %f29;
mov.b32 %r407, %f30;
shfl.sync.bfly.b32 %r408|%p10, %r407, %r394, %r389, %r391;
mov.b32 %f31, %r408;
add.ftz.f32 %f32, %f30, %f31;
mov.b32 %r409, %f32;
shfl.sync.bfly.b32 %r410|%p11, %r409, %r397, %r389, %r391;
mov.b32 %f33, %r410;
add.ftz.f32 %f34, %f32, %f33;
mov.b32 %r411, %f34;
shfl.sync.bfly.b32 %r412|%p12, %r411, %r400, %r389, %r391;
mov.b32 %f35, %r412;
add.ftz.f32 %f36, %f34, %f35;
mov.b32 %r413, %f36;
shfl.sync.bfly.b32 %r414|%p13, %r413, %r403, %r389, %r391;
mov.b32 %f37, %r414;
add.ftz.f32 %f38, %f36, %f37;
st.local.f32 [%rd1+4], %f38;
setp.gt.u32 %p14, %r23, 1;
mov.u32 %r416, %ctaid.x;
shl.b32 %r417, %r416, 1;
add.s32 %r9, %r417, %r23;
setp.ge.u32 %p15, %r9, %r13;
or.pred %p16, %p14, %p15;
@%p16 bra $L__BB238_7;

mul.wide.u32 %rd82, %r23, 4;
add.s64 %rd83, %rd1, %rd82;
ld.local.f32 %f39, [%rd83];
mov.u32 %r419, %ctaid.y;
mad.lo.s32 %r421, %r2, %r16, %r9;
mad.lo.s32 %r422, %r419, %r19, %r421;
cvta.to.global.u64 %rd84, %rd17;
mul.wide.u32 %rd85, %r422, 4;
add.s64 %rd86, %rd84, %rd85;
st.global.f32 [%rd86], %f39;

$L__BB238_7:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<320>;
.reg .b16 %rs<121>;
.reg .f32 %f<1005>;
.reg .b32 %r<2139>;
.reg .b64 %rd<283>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate[1536];

ld.param.u64 %rd36, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd37, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd38, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u32 %r63, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r49, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r50, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r54, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r55, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r56, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r60, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r61, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r62, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r58, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r57, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r52, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r51, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r48, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r47, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r46, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
ld.param.u64 %rd41, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+16];
ld.param.u64 %rd40, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+8];
ld.param.u64 %rd39, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3];
mov.u32 %r1, %tid.y;
shl.b32 %r64, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r64, %r2;
mov.u32 %r65, %ctaid.x;
shl.b32 %r4, %r65, 2;
shr.u32 %r5, %r63, 8;
mov.u32 %r6, %ctaid.y;
setp.eq.s64 %p1, %rd38, 0;
@%p1 bra $L__BB239_2;

cvta.to.global.u64 %rd45, %rd38;
mul.wide.u32 %rd46, %r6, 4;
add.s64 %rd47, %rd45, %rd46;
ld.global.nc.u32 %r2135, [%rd47];
mul.hi.u32 %r66, %r6, %r46;
add.s32 %r67, %r66, %r6;
shr.u32 %r68, %r67, %r47;
mul.lo.s32 %r69, %r68, %r48;
sub.s32 %r2136, %r6, %r69;
bra.uni $L__BB239_3;

$L__BB239_2:
mul.hi.u32 %r70, %r6, %r51;
add.s32 %r71, %r70, %r6;
shr.u32 %r2135, %r71, %r52;
mov.u32 %r2136, %r6;

$L__BB239_3:
mov.u32 %r12, %ctaid.z;
selp.b32 %r13, %r6, %r2135, %p1;
setp.ne.s32 %p3, %r1, 0;
setp.gt.u32 %p4, %r2, 3;
or.pred %p5, %p4, %p3;
add.s32 %r72, %r4, %r2;
setp.ge.u32 %p6, %r72, %r50;
mov.f32 %f27, 0f00000000;
or.pred %p7, %p5, %p6;
mov.f32 %f4, %f27;
mov.f32 %f5, %f27;
@%p7 bra $L__BB239_8;

setp.eq.s64 %p8, %rd39, 0;
mov.f32 %f4, 0f00000000;
mov.f32 %f5, %f4;
@%p8 bra $L__BB239_6;

cvta.to.global.u64 %rd48, %rd39;
mul.lo.s32 %r73, %r12, %r62;
cvt.u64.u32 %rd49, %r73;
mul.lo.s32 %r74, %r13, %r56;
cvt.u64.u32 %rd50, %r74;
cvt.u64.u32 %rd51, %r2;
cvt.s64.s32 %rd52, %r4;
add.s64 %rd53, %rd52, %rd51;
add.s64 %rd54, %rd53, %rd49;
add.s64 %rd55, %rd54, %rd50;
shl.b64 %rd56, %rd55, 2;
add.s64 %rd57, %rd48, %rd56;
ld.global.f32 %f5, [%rd57];

$L__BB239_6:
setp.eq.s64 %p9, %rd41, 0;
setp.eq.s64 %p10, %rd40, 0;
or.pred %p11, %p9, %p10;
@%p11 bra $L__BB239_8;

cvta.to.global.u64 %rd58, %rd41;
mul.lo.s32 %r75, %r12, %r62;
cvt.u64.u32 %rd59, %r75;
mul.lo.s32 %r76, %r13, %r56;
cvt.u64.u32 %rd60, %r76;
cvt.u64.u32 %rd61, %r2;
cvt.s64.s32 %rd62, %r4;
add.s64 %rd63, %rd62, %rd61;
add.s64 %rd64, %rd63, %rd59;
add.s64 %rd65, %rd64, %rd60;
shl.b64 %rd66, %rd65, 2;
add.s64 %rd67, %rd58, %rd66;
ld.global.f32 %f4, [%rd67];

$L__BB239_8:
mul.lo.s32 %r77, %r12, %r61;
cvt.u64.u32 %rd68, %r77;
mul.lo.s32 %r78, %r2136, %r55;
cvt.u64.u32 %rd69, %r78;
add.s64 %rd4, %rd69, %rd68;
mul.hi.u32 %r79, %r12, %r57;
add.s32 %r80, %r79, %r12;
shr.u32 %r14, %r80, %r58;
shr.s32 %r81, %r3, 31;
shr.u32 %r82, %r81, 29;
add.s32 %r83, %r3, %r82;
shr.s32 %r2138, %r83, 3;
setp.ge.s32 %p12, %r2138, %r5;
mov.f32 %f28, %f27;
mov.f32 %f29, %f27;
mov.f32 %f30, %f27;
mov.f32 %f31, %f27;
mov.f32 %f32, %f27;
mov.f32 %f33, %f27;
mov.f32 %f34, %f27;
@%p12 bra $L__BB239_19;

mul.lo.s32 %r84, %r14, %r60;
mul.lo.s32 %r85, %r2135, %r54;
and.b32 %r89, %r83, -8;
sub.s32 %r90, %r3, %r89;
shl.b32 %r91, %r90, 2;
shl.b32 %r92, %r90, 1;
add.s32 %r93, %r92, 33;
cvt.s64.s32 %rd5, %r90;
shl.b32 %r2137, %r2138, 3;
add.s32 %r94, %r2138, %r84;
add.s32 %r95, %r94, %r85;
add.s32 %r96, %r4, 1;
mad.lo.s32 %r97, %r49, %r96, %r95;
mul.wide.s32 %rd6, %r97, 98;
mad.lo.s32 %r98, %r4, %r49, %r2138;
add.s32 %r99, %r98, %r84;
add.s32 %r100, %r99, %r85;
mul.wide.s32 %rd7, %r100, 98;
mul.wide.s32 %rd70, %r91, 2;
add.s64 %rd71, %rd7, %rd70;
add.s64 %rd8, %rd71, 8;
mul.wide.s32 %rd72, %r93, 2;
add.s64 %rd73, %rd72, -2;
add.s64 %rd74, %rd73, %rd6;
add.s64 %rd9, %rd74, 2;
add.s64 %rd75, %rd73, %rd7;
add.s64 %rd10, %rd75, 2;
add.s64 %rd76, %rd6, %rd70;
add.s64 %rd11, %rd76, 4;
add.s32 %r101, %r4, 3;
mad.lo.s32 %r102, %r49, %r101, %r95;
mul.wide.s32 %rd12, %r102, 98;
add.s64 %rd77, %rd12, %rd70;
add.s64 %rd13, %rd77, 8;
add.s64 %rd78, %rd73, %rd12;
add.s64 %rd14, %rd78, 4;
add.s32 %r103, %r4, 2;
mad.lo.s32 %r104, %r49, %r103, %r95;
mul.wide.s32 %rd15, %r104, 98;
add.s64 %rd79, %rd73, %rd15;
add.s64 %rd16, %rd79, 2;
add.s64 %rd17, %rd15, %rd70;
mul.wide.s32 %rd18, %r49, 98;
cvta.to.global.u64 %rd282, %rd40;
cvta.to.global.u64 %rd281, %rd36;
cvta.to.global.u64 %rd21, %rd37;
mov.f32 %f27, 0f00000000;
mov.f32 %f28, %f27;
mov.f32 %f29, %f27;
mov.f32 %f30, %f27;
mov.f32 %f31, %f27;
mov.f32 %f32, %f27;
mov.f32 %f33, %f27;
mov.f32 %f34, %f27;

$L__BB239_10:
cvt.u64.u32 %rd80, %r2137;
add.s64 %rd81, %rd4, %rd80;
add.s64 %rd82, %rd281, %rd8;
ld.global.nc.u16 %rs2, [%rd82+-4];
ld.global.nc.u16 %rs3, [%rd82+-2];
ld.global.nc.u16 %rs4, [%rd82];
add.s64 %rd25, %rd281, %rd10;
ld.global.nc.u16 %rs5, [%rd25];
cvt.u32.u16 %r154, %rs5;
ld.global.nc.u16 %rs6, [%rd25+2];
cvt.u32.u16 %r155, %rs6;
mov.b32 %r156, {%rs5, %rs6};
add.s64 %rd83, %rd81, %rd5;
mul.lo.s64 %rd84, %rd83, 36;
ld.global.nc.u16 %rs7, [%rd82+-6];
shl.b16 %rs8, %rs7, 2;
cvt.u64.u16 %rd85, %rs8;
and.b64 %rd86, %rd85, 1020;
mov.u64 %rd87, iq3xxs_grid;
add.s64 %rd88, %rd87, %rd86;
ld.global.nc.u32 %r157, [%rd88];
shr.u16 %rs9, %rs7, 8;
cvt.u32.u16 %r158, %rs9;
mul.wide.u32 %rd89, %r158, 4;
add.s64 %rd90, %rd87, %rd89;
ld.global.nc.u32 %r159, [%rd90];
and.b32 %r160, %r154, 255;
popc.b32 %r161, %r160;
shl.b32 %r162, %r161, 7;
and.b32 %r163, %r162, 128;
xor.b32 %r164, %r163, %r160;
mul.lo.s32 %r165, %r164, 16843009;
and.b32 %r166, %r165, 134480385;
or.b32 %r167, %r166, -2139062144;
add.s32 %r168, %r167, -16843009;
or.b32 %r105, %r168, %r166;

prmt.b32 %r105, %r105, 0, 0xba98;

xor.b32 %r169, %r105, %r157;
or.b32 %r170, %r169, -2139062144;
and.b32 %r171, %r105, 2139062143;
and.b32 %r172, %r157, -2139062144;
sub.s32 %r173, %r170, %r171;
xor.b32 %r174, %r172, %r173;
xor.b32 %r110, %r174, -2139062144;
add.s64 %rd91, %rd21, %rd84;
ld.global.nc.u32 %r19, [%rd91+4];
and.b32 %r175, %r165, -2143281136;
or.b32 %r176, %r175, -2139062144;
add.s32 %r177, %r176, -16843009;
or.b32 %r107, %r177, %r175;

prmt.b32 %r107, %r107, 0, 0xba98;

xor.b32 %r178, %r107, %r159;
or.b32 %r179, %r178, -2139062144;
and.b32 %r180, %r107, 2139062143;
and.b32 %r181, %r159, -2139062144;
sub.s32 %r182, %r179, %r180;
xor.b32 %r183, %r181, %r182;
xor.b32 %r114, %r183, -2139062144;
ld.global.nc.u32 %r20, [%rd91+8];
mov.u32 %r112, 0;

dp4a.s32.s32 %r109, %r110, %r19, %r112;


dp4a.s32.s32 %r113, %r114, %r20, %r109;

shl.b16 %rs10, %rs2, 2;
cvt.u64.u16 %rd92, %rs10;
and.b64 %rd93, %rd92, 1020;
add.s64 %rd94, %rd87, %rd93;
ld.global.nc.u32 %r184, [%rd94];
shr.u16 %rs11, %rs2, 8;
cvt.u32.u16 %r185, %rs11;
mul.wide.u32 %rd95, %r185, 4;
add.s64 %rd96, %rd87, %rd95;
ld.global.nc.u32 %r186, [%rd96];
shr.u32 %r187, %r154, 7;
and.b32 %r188, %r187, 255;
popc.b32 %r189, %r188;
shl.b32 %r190, %r189, 7;
and.b32 %r191, %r190, 128;
xor.b32 %r192, %r191, %r188;
mul.lo.s32 %r193, %r192, 16843009;
and.b32 %r194, %r193, 134480385;
or.b32 %r195, %r194, -2139062144;
add.s32 %r196, %r195, -16843009;
or.b32 %r117, %r196, %r194;

prmt.b32 %r117, %r117, 0, 0xba98;

xor.b32 %r197, %r117, %r184;
or.b32 %r198, %r197, -2139062144;
and.b32 %r199, %r117, 2139062143;
and.b32 %r200, %r184, -2139062144;
sub.s32 %r201, %r198, %r199;
xor.b32 %r202, %r200, %r201;
xor.b32 %r122, %r202, -2139062144;
ld.global.nc.u32 %r21, [%rd91+12];
and.b32 %r203, %r193, -2143281136;
or.b32 %r204, %r203, -2139062144;
add.s32 %r205, %r204, -16843009;
or.b32 %r119, %r205, %r203;

prmt.b32 %r119, %r119, 0, 0xba98;

xor.b32 %r206, %r119, %r186;
or.b32 %r207, %r206, -2139062144;
and.b32 %r208, %r119, 2139062143;
and.b32 %r209, %r186, -2139062144;
sub.s32 %r210, %r207, %r208;
xor.b32 %r211, %r209, %r210;
xor.b32 %r126, %r211, -2139062144;
ld.global.nc.u32 %r22, [%rd91+16];

dp4a.s32.s32 %r121, %r122, %r21, %r113;


dp4a.s32.s32 %r125, %r126, %r22, %r121;

shl.b16 %rs12, %rs3, 2;
cvt.u64.u16 %rd97, %rs12;
and.b64 %rd98, %rd97, 1020;
add.s64 %rd99, %rd87, %rd98;
ld.global.nc.u32 %r212, [%rd99];
shr.u16 %rs13, %rs3, 8;
cvt.u32.u16 %r213, %rs13;
mul.wide.u32 %rd100, %r213, 4;
add.s64 %rd101, %rd87, %rd100;
ld.global.nc.u32 %r214, [%rd101];
shr.u32 %r215, %r156, 14;
and.b32 %r216, %r215, 255;
popc.b32 %r217, %r216;
shl.b32 %r218, %r217, 7;
and.b32 %r219, %r218, 128;
xor.b32 %r220, %r219, %r216;
mul.lo.s32 %r221, %r220, 16843009;
and.b32 %r222, %r221, 134480385;
or.b32 %r223, %r222, -2139062144;
add.s32 %r224, %r223, -16843009;
or.b32 %r129, %r224, %r222;

prmt.b32 %r129, %r129, 0, 0xba98;

xor.b32 %r225, %r129, %r212;
or.b32 %r226, %r225, -2139062144;
and.b32 %r227, %r129, 2139062143;
and.b32 %r228, %r212, -2139062144;
sub.s32 %r229, %r226, %r227;
xor.b32 %r230, %r228, %r229;
xor.b32 %r134, %r230, -2139062144;
ld.global.nc.u32 %r23, [%rd91+20];
and.b32 %r231, %r221, -2143281136;
or.b32 %r232, %r231, -2139062144;
add.s32 %r233, %r232, -16843009;
or.b32 %r131, %r233, %r231;

prmt.b32 %r131, %r131, 0, 0xba98;

xor.b32 %r234, %r131, %r214;
or.b32 %r235, %r234, -2139062144;
and.b32 %r236, %r131, 2139062143;
and.b32 %r237, %r214, -2139062144;
sub.s32 %r238, %r235, %r236;
xor.b32 %r239, %r237, %r238;
xor.b32 %r138, %r239, -2139062144;
ld.global.nc.u32 %r24, [%rd91+24];

dp4a.s32.s32 %r133, %r134, %r23, %r125;


dp4a.s32.s32 %r137, %r138, %r24, %r133;

shl.b16 %rs14, %rs4, 2;
cvt.u64.u16 %rd102, %rs14;
and.b64 %rd103, %rd102, 1020;
add.s64 %rd104, %rd87, %rd103;
ld.global.nc.u32 %r240, [%rd104];
shr.u16 %rs15, %rs4, 8;
cvt.u32.u16 %r241, %rs15;
mul.wide.u32 %rd105, %r241, 4;
add.s64 %rd106, %rd87, %rd105;
ld.global.nc.u32 %r242, [%rd106];
shr.u32 %r243, %r155, 5;
and.b32 %r244, %r243, 255;
popc.b32 %r245, %r244;
shl.b32 %r246, %r245, 7;
and.b32 %r247, %r246, 128;
xor.b32 %r248, %r247, %r244;
mul.lo.s32 %r249, %r248, 16843009;
and.b32 %r250, %r249, 134480385;
or.b32 %r251, %r250, -2139062144;
add.s32 %r252, %r251, -16843009;
or.b32 %r141, %r252, %r250;

prmt.b32 %r141, %r141, 0, 0xba98;

xor.b32 %r253, %r141, %r240;
or.b32 %r254, %r253, -2139062144;
and.b32 %r255, %r141, 2139062143;
and.b32 %r256, %r240, -2139062144;
sub.s32 %r257, %r254, %r255;
xor.b32 %r258, %r256, %r257;
xor.b32 %r146, %r258, -2139062144;
ld.global.nc.u32 %r25, [%rd91+28];
and.b32 %r259, %r249, -2143281136;
or.b32 %r260, %r259, -2139062144;
add.s32 %r261, %r260, -16843009;
or.b32 %r143, %r261, %r259;

prmt.b32 %r143, %r143, 0, 0xba98;

xor.b32 %r262, %r143, %r242;
or.b32 %r263, %r262, -2139062144;
and.b32 %r264, %r143, 2139062143;
and.b32 %r265, %r242, -2139062144;
sub.s32 %r266, %r263, %r264;
xor.b32 %r267, %r265, %r266;
xor.b32 %r150, %r267, -2139062144;
ld.global.nc.u32 %r26, [%rd91+32];

dp4a.s32.s32 %r145, %r146, %r25, %r137;

shr.u32 %r268, %r155, 12;

dp4a.s32.s32 %r149, %r150, %r26, %r145;

shr.u32 %r269, %r149, 31;
add.s32 %r270, %r149, %r269;
shr.s32 %r271, %r270, 1;
mad.lo.s32 %r272, %r149, %r268, %r271;
shr.u32 %r273, %r272, 31;
add.s32 %r274, %r272, %r273;
shr.s32 %r275, %r274, 1;
add.s64 %rd26, %rd281, %rd7;
ld.global.nc.u16 %rs1, [%rd26];

{ cvt.f32.f16 %f101, %rs1;}


ld.global.nc.u32 %r27, [%rd91];

{.reg .f16 low,high;
mov.b32 {low,high},%r27;
cvt.f32.f16 %f102, low;}


mul.ftz.f32 %f103, %f101, %f102;
cvt.rn.f32.s32 %f104, %r275;
fma.rn.ftz.f32 %f30, %f103, %f104, %f30;
setp.eq.s64 %p13, %rd40, 0;
@%p13 bra $L__BB239_12;

mov.u64 %rd280, iq3xxs_grid;
add.s64 %rd107, %rd282, %rd8;
ld.global.u16 %rs17, [%rd107+-4];
ld.global.u16 %rs18, [%rd107+-2];
ld.global.u16 %rs19, [%rd107];
add.s64 %rd108, %rd282, %rd10;
ld.global.u16 %rs20, [%rd108];
cvt.u32.u16 %r325, %rs20;
ld.global.u16 %rs21, [%rd108+2];
cvt.u32.u16 %r326, %rs21;
mov.b32 %r327, {%rs20, %rs21};
ld.global.u16 %rs22, [%rd107+-6];
shl.b16 %rs23, %rs22, 2;
cvt.u64.u16 %rd109, %rs23;
and.b64 %rd110, %rd109, 1020;
add.s64 %rd112, %rd280, %rd110;
ld.global.nc.u32 %r328, [%rd112];
shr.u16 %rs24, %rs22, 8;
cvt.u32.u16 %r329, %rs24;
mul.wide.u32 %rd113, %r329, 4;
add.s64 %rd114, %rd280, %rd113;
ld.global.nc.u32 %r330, [%rd114];
and.b32 %r331, %r325, 255;
popc.b32 %r332, %r331;
shl.b32 %r333, %r332, 7;
and.b32 %r334, %r333, 128;
xor.b32 %r335, %r334, %r331;
mul.lo.s32 %r336, %r335, 16843009;
and.b32 %r337, %r336, 134480385;
or.b32 %r338, %r337, -2139062144;
add.s32 %r339, %r338, -16843009;
or.b32 %r276, %r339, %r337;

prmt.b32 %r276, %r276, 0, 0xba98;

xor.b32 %r340, %r276, %r328;
or.b32 %r341, %r340, -2139062144;
and.b32 %r342, %r276, 2139062143;
and.b32 %r343, %r328, -2139062144;
sub.s32 %r344, %r341, %r342;
xor.b32 %r345, %r343, %r344;
xor.b32 %r281, %r345, -2139062144;
and.b32 %r346, %r336, -2143281136;
or.b32 %r347, %r346, -2139062144;
add.s32 %r348, %r347, -16843009;
or.b32 %r278, %r348, %r346;

prmt.b32 %r278, %r278, 0, 0xba98;

xor.b32 %r349, %r278, %r330;
or.b32 %r350, %r349, -2139062144;
and.b32 %r351, %r278, 2139062143;
and.b32 %r352, %r330, -2139062144;
sub.s32 %r353, %r350, %r351;
xor.b32 %r354, %r352, %r353;
xor.b32 %r285, %r354, -2139062144;

dp4a.s32.s32 %r280, %r281, %r19, %r112;


dp4a.s32.s32 %r284, %r285, %r20, %r280;

shl.b16 %rs25, %rs17, 2;
cvt.u64.u16 %rd115, %rs25;
and.b64 %rd116, %rd115, 1020;
add.s64 %rd117, %rd280, %rd116;
ld.global.nc.u32 %r355, [%rd117];
shr.u16 %rs26, %rs17, 8;
cvt.u32.u16 %r356, %rs26;
mul.wide.u32 %rd118, %r356, 4;
add.s64 %rd119, %rd280, %rd118;
ld.global.nc.u32 %r357, [%rd119];
shr.u32 %r358, %r325, 7;
and.b32 %r359, %r358, 255;
popc.b32 %r360, %r359;
shl.b32 %r361, %r360, 7;
and.b32 %r362, %r361, 128;
xor.b32 %r363, %r362, %r359;
mul.lo.s32 %r364, %r363, 16843009;
and.b32 %r365, %r364, 134480385;
or.b32 %r366, %r365, -2139062144;
add.s32 %r367, %r366, -16843009;
or.b32 %r288, %r367, %r365;

prmt.b32 %r288, %r288, 0, 0xba98;

xor.b32 %r368, %r288, %r355;
or.b32 %r369, %r368, -2139062144;
and.b32 %r370, %r288, 2139062143;
and.b32 %r371, %r355, -2139062144;
sub.s32 %r372, %r369, %r370;
xor.b32 %r373, %r371, %r372;
xor.b32 %r293, %r373, -2139062144;
and.b32 %r374, %r364, -2143281136;
or.b32 %r375, %r374, -2139062144;
add.s32 %r376, %r375, -16843009;
or.b32 %r290, %r376, %r374;

prmt.b32 %r290, %r290, 0, 0xba98;

xor.b32 %r377, %r290, %r357;
or.b32 %r378, %r377, -2139062144;
and.b32 %r379, %r290, 2139062143;
and.b32 %r380, %r357, -2139062144;
sub.s32 %r381, %r378, %r379;
xor.b32 %r382, %r380, %r381;
xor.b32 %r297, %r382, -2139062144;

dp4a.s32.s32 %r292, %r293, %r21, %r284;


dp4a.s32.s32 %r296, %r297, %r22, %r292;

shl.b16 %rs27, %rs18, 2;
cvt.u64.u16 %rd120, %rs27;
and.b64 %rd121, %rd120, 1020;
add.s64 %rd122, %rd280, %rd121;
ld.global.nc.u32 %r383, [%rd122];
shr.u16 %rs28, %rs18, 8;
cvt.u32.u16 %r384, %rs28;
mul.wide.u32 %rd123, %r384, 4;
add.s64 %rd124, %rd280, %rd123;
ld.global.nc.u32 %r385, [%rd124];
shr.u32 %r386, %r327, 14;
and.b32 %r387, %r386, 255;
popc.b32 %r388, %r387;
shl.b32 %r389, %r388, 7;
and.b32 %r390, %r389, 128;
xor.b32 %r391, %r390, %r387;
mul.lo.s32 %r392, %r391, 16843009;
and.b32 %r393, %r392, 134480385;
or.b32 %r394, %r393, -2139062144;
add.s32 %r395, %r394, -16843009;
or.b32 %r300, %r395, %r393;

prmt.b32 %r300, %r300, 0, 0xba98;

xor.b32 %r396, %r300, %r383;
or.b32 %r397, %r396, -2139062144;
and.b32 %r398, %r300, 2139062143;
and.b32 %r399, %r383, -2139062144;
sub.s32 %r400, %r397, %r398;
xor.b32 %r401, %r399, %r400;
xor.b32 %r305, %r401, -2139062144;
and.b32 %r402, %r392, -2143281136;
or.b32 %r403, %r402, -2139062144;
add.s32 %r404, %r403, -16843009;
or.b32 %r302, %r404, %r402;

prmt.b32 %r302, %r302, 0, 0xba98;

xor.b32 %r405, %r302, %r385;
or.b32 %r406, %r405, -2139062144;
and.b32 %r407, %r302, 2139062143;
and.b32 %r408, %r385, -2139062144;
sub.s32 %r409, %r406, %r407;
xor.b32 %r410, %r408, %r409;
xor.b32 %r309, %r410, -2139062144;

dp4a.s32.s32 %r304, %r305, %r23, %r296;


dp4a.s32.s32 %r308, %r309, %r24, %r304;

shl.b16 %rs29, %rs19, 2;
cvt.u64.u16 %rd125, %rs29;
and.b64 %rd126, %rd125, 1020;
add.s64 %rd127, %rd280, %rd126;
ld.global.nc.u32 %r411, [%rd127];
shr.u16 %rs30, %rs19, 8;
cvt.u32.u16 %r412, %rs30;
mul.wide.u32 %rd128, %r412, 4;
add.s64 %rd129, %rd280, %rd128;
ld.global.nc.u32 %r413, [%rd129];
shr.u32 %r414, %r326, 5;
and.b32 %r415, %r414, 255;
popc.b32 %r416, %r415;
shl.b32 %r417, %r416, 7;
and.b32 %r418, %r417, 128;
xor.b32 %r419, %r418, %r415;
mul.lo.s32 %r420, %r419, 16843009;
and.b32 %r421, %r420, 134480385;
or.b32 %r422, %r421, -2139062144;
add.s32 %r423, %r422, -16843009;
or.b32 %r312, %r423, %r421;

prmt.b32 %r312, %r312, 0, 0xba98;

xor.b32 %r424, %r312, %r411;
or.b32 %r425, %r424, -2139062144;
and.b32 %r426, %r312, 2139062143;
and.b32 %r427, %r411, -2139062144;
sub.s32 %r428, %r425, %r426;
xor.b32 %r429, %r427, %r428;
xor.b32 %r317, %r429, -2139062144;
and.b32 %r430, %r420, -2143281136;
or.b32 %r431, %r430, -2139062144;
add.s32 %r432, %r431, -16843009;
or.b32 %r314, %r432, %r430;

prmt.b32 %r314, %r314, 0, 0xba98;

xor.b32 %r433, %r314, %r413;
or.b32 %r434, %r433, -2139062144;
and.b32 %r435, %r314, 2139062143;
and.b32 %r436, %r413, -2139062144;
sub.s32 %r437, %r434, %r435;
xor.b32 %r438, %r436, %r437;
xor.b32 %r321, %r438, -2139062144;

dp4a.s32.s32 %r316, %r317, %r25, %r308;

shr.u32 %r439, %r326, 12;

dp4a.s32.s32 %r320, %r321, %r26, %r316;

shr.u32 %r440, %r320, 31;
add.s32 %r441, %r320, %r440;
shr.s32 %r442, %r441, 1;
mad.lo.s32 %r443, %r320, %r439, %r442;
shr.u32 %r444, %r443, 31;
add.s32 %r445, %r443, %r444;
shr.s32 %r446, %r445, 1;
add.s64 %rd130, %rd282, %rd7;
ld.global.u16 %rs16, [%rd130];

{ cvt.f32.f16 %f105, %rs16;}



{.reg .f16 low,high;
mov.b32 {low,high},%r27;
cvt.f32.f16 %f106, low;}


mul.ftz.f32 %f107, %f105, %f106;
cvt.rn.f32.s32 %f108, %r446;
fma.rn.ftz.f32 %f34, %f107, %f108, %f34;

$L__BB239_12:
mov.u64 %rd277, iq3xxs_grid;
add.s64 %rd272, %rd281, %rd8;
add.s64 %rd271, %rd272, -6;
add.s64 %rd27, %rd271, %rd18;
ld.global.nc.u16 %rs32, [%rd27+2];
ld.global.nc.u16 %rs33, [%rd27+4];
ld.global.nc.u16 %rs34, [%rd27+6];
add.s64 %rd28, %rd25, %rd18;
ld.global.nc.u16 %rs35, [%rd28];
cvt.u32.u16 %r496, %rs35;
ld.global.nc.u16 %rs36, [%rd28+2];
cvt.u32.u16 %r497, %rs36;
mov.b32 %r498, {%rs35, %rs36};
ld.global.nc.u16 %rs37, [%rd27];
shl.b16 %rs38, %rs37, 2;
cvt.u64.u16 %rd131, %rs38;
and.b64 %rd132, %rd131, 1020;
add.s64 %rd134, %rd277, %rd132;
ld.global.nc.u32 %r499, [%rd134];
shr.u16 %rs39, %rs37, 8;
cvt.u32.u16 %r500, %rs39;
mul.wide.u32 %rd135, %r500, 4;
add.s64 %rd136, %rd277, %rd135;
ld.global.nc.u32 %r501, [%rd136];
and.b32 %r502, %r496, 255;
popc.b32 %r503, %r502;
shl.b32 %r504, %r503, 7;
and.b32 %r505, %r504, 128;
xor.b32 %r506, %r505, %r502;
mul.lo.s32 %r507, %r506, 16843009;
and.b32 %r508, %r507, 134480385;
or.b32 %r509, %r508, -2139062144;
add.s32 %r510, %r509, -16843009;
or.b32 %r447, %r510, %r508;

prmt.b32 %r447, %r447, 0, 0xba98;

xor.b32 %r511, %r447, %r499;
or.b32 %r512, %r511, -2139062144;
and.b32 %r513, %r447, 2139062143;
and.b32 %r514, %r499, -2139062144;
sub.s32 %r515, %r512, %r513;
xor.b32 %r516, %r514, %r515;
xor.b32 %r452, %r516, -2139062144;
and.b32 %r517, %r507, -2143281136;
or.b32 %r518, %r517, -2139062144;
add.s32 %r519, %r518, -16843009;
or.b32 %r449, %r519, %r517;

prmt.b32 %r449, %r449, 0, 0xba98;

xor.b32 %r520, %r449, %r501;
or.b32 %r521, %r520, -2139062144;
and.b32 %r522, %r449, 2139062143;
and.b32 %r523, %r501, -2139062144;
sub.s32 %r524, %r521, %r522;
xor.b32 %r525, %r523, %r524;
xor.b32 %r456, %r525, -2139062144;

dp4a.s32.s32 %r451, %r452, %r19, %r112;


dp4a.s32.s32 %r455, %r456, %r20, %r451;

shl.b16 %rs40, %rs32, 2;
cvt.u64.u16 %rd137, %rs40;
and.b64 %rd138, %rd137, 1020;
add.s64 %rd139, %rd277, %rd138;
ld.global.nc.u32 %r526, [%rd139];
shr.u16 %rs41, %rs32, 8;
cvt.u32.u16 %r527, %rs41;
mul.wide.u32 %rd140, %r527, 4;
add.s64 %rd141, %rd277, %rd140;
ld.global.nc.u32 %r528, [%rd141];
shr.u32 %r529, %r496, 7;
and.b32 %r530, %r529, 255;
popc.b32 %r531, %r530;
shl.b32 %r532, %r531, 7;
and.b32 %r533, %r532, 128;
xor.b32 %r534, %r533, %r530;
mul.lo.s32 %r535, %r534, 16843009;
and.b32 %r536, %r535, 134480385;
or.b32 %r537, %r536, -2139062144;
add.s32 %r538, %r537, -16843009;
or.b32 %r459, %r538, %r536;

prmt.b32 %r459, %r459, 0, 0xba98;

xor.b32 %r539, %r459, %r526;
or.b32 %r540, %r539, -2139062144;
and.b32 %r541, %r459, 2139062143;
and.b32 %r542, %r526, -2139062144;
sub.s32 %r543, %r540, %r541;
xor.b32 %r544, %r542, %r543;
xor.b32 %r464, %r544, -2139062144;
and.b32 %r545, %r535, -2143281136;
or.b32 %r546, %r545, -2139062144;
add.s32 %r547, %r546, -16843009;
or.b32 %r461, %r547, %r545;

prmt.b32 %r461, %r461, 0, 0xba98;

xor.b32 %r548, %r461, %r528;
or.b32 %r549, %r548, -2139062144;
and.b32 %r550, %r461, 2139062143;
and.b32 %r551, %r528, -2139062144;
sub.s32 %r552, %r549, %r550;
xor.b32 %r553, %r551, %r552;
xor.b32 %r468, %r553, -2139062144;

dp4a.s32.s32 %r463, %r464, %r21, %r455;


dp4a.s32.s32 %r467, %r468, %r22, %r463;

shl.b16 %rs42, %rs33, 2;
cvt.u64.u16 %rd142, %rs42;
and.b64 %rd143, %rd142, 1020;
add.s64 %rd144, %rd277, %rd143;
ld.global.nc.u32 %r554, [%rd144];
shr.u16 %rs43, %rs33, 8;
cvt.u32.u16 %r555, %rs43;
mul.wide.u32 %rd145, %r555, 4;
add.s64 %rd146, %rd277, %rd145;
ld.global.nc.u32 %r556, [%rd146];
shr.u32 %r557, %r498, 14;
and.b32 %r558, %r557, 255;
popc.b32 %r559, %r558;
shl.b32 %r560, %r559, 7;
and.b32 %r561, %r560, 128;
xor.b32 %r562, %r561, %r558;
mul.lo.s32 %r563, %r562, 16843009;
and.b32 %r564, %r563, 134480385;
or.b32 %r565, %r564, -2139062144;
add.s32 %r566, %r565, -16843009;
or.b32 %r471, %r566, %r564;

prmt.b32 %r471, %r471, 0, 0xba98;

xor.b32 %r567, %r471, %r554;
or.b32 %r568, %r567, -2139062144;
and.b32 %r569, %r471, 2139062143;
and.b32 %r570, %r554, -2139062144;
sub.s32 %r571, %r568, %r569;
xor.b32 %r572, %r570, %r571;
xor.b32 %r476, %r572, -2139062144;
and.b32 %r573, %r563, -2143281136;
or.b32 %r574, %r573, -2139062144;
add.s32 %r575, %r574, -16843009;
or.b32 %r473, %r575, %r573;

prmt.b32 %r473, %r473, 0, 0xba98;

xor.b32 %r576, %r473, %r556;
or.b32 %r577, %r576, -2139062144;
and.b32 %r578, %r473, 2139062143;
and.b32 %r579, %r556, -2139062144;
sub.s32 %r580, %r577, %r578;
xor.b32 %r581, %r579, %r580;
xor.b32 %r480, %r581, -2139062144;

dp4a.s32.s32 %r475, %r476, %r23, %r467;


dp4a.s32.s32 %r479, %r480, %r24, %r475;

shl.b16 %rs44, %rs34, 2;
cvt.u64.u16 %rd147, %rs44;
and.b64 %rd148, %rd147, 1020;
add.s64 %rd149, %rd277, %rd148;
ld.global.nc.u32 %r582, [%rd149];
shr.u16 %rs45, %rs34, 8;
cvt.u32.u16 %r583, %rs45;
mul.wide.u32 %rd150, %r583, 4;
add.s64 %rd151, %rd277, %rd150;
ld.global.nc.u32 %r584, [%rd151];
shr.u32 %r585, %r497, 5;
and.b32 %r586, %r585, 255;
popc.b32 %r587, %r586;
shl.b32 %r588, %r587, 7;
and.b32 %r589, %r588, 128;
xor.b32 %r590, %r589, %r586;
mul.lo.s32 %r591, %r590, 16843009;
and.b32 %r592, %r591, 134480385;
or.b32 %r593, %r592, -2139062144;
add.s32 %r594, %r593, -16843009;
or.b32 %r483, %r594, %r592;

prmt.b32 %r483, %r483, 0, 0xba98;

xor.b32 %r595, %r483, %r582;
or.b32 %r596, %r595, -2139062144;
and.b32 %r597, %r483, 2139062143;
and.b32 %r598, %r582, -2139062144;
sub.s32 %r599, %r596, %r597;
xor.b32 %r600, %r598, %r599;
xor.b32 %r488, %r600, -2139062144;
and.b32 %r601, %r591, -2143281136;
or.b32 %r602, %r601, -2139062144;
add.s32 %r603, %r602, -16843009;
or.b32 %r485, %r603, %r601;

prmt.b32 %r485, %r485, 0, 0xba98;

xor.b32 %r604, %r485, %r584;
or.b32 %r605, %r604, -2139062144;
and.b32 %r606, %r485, 2139062143;
and.b32 %r607, %r584, -2139062144;
sub.s32 %r608, %r605, %r606;
xor.b32 %r609, %r607, %r608;
xor.b32 %r492, %r609, -2139062144;

dp4a.s32.s32 %r487, %r488, %r25, %r479;

shr.u32 %r610, %r497, 12;

dp4a.s32.s32 %r491, %r492, %r26, %r487;

shr.u32 %r611, %r491, 31;
add.s32 %r612, %r491, %r611;
shr.s32 %r613, %r612, 1;
mad.lo.s32 %r614, %r491, %r610, %r613;
shr.u32 %r615, %r614, 31;
add.s32 %r616, %r614, %r615;
shr.s32 %r617, %r616, 1;
add.s64 %rd29, %rd26, %rd18;
ld.global.nc.u16 %rs31, [%rd29];

{ cvt.f32.f16 %f109, %rs31;}



{.reg .f16 low,high;
mov.b32 {low,high},%r27;
cvt.f32.f16 %f110, low;}


mul.ftz.f32 %f111, %f109, %f110;
cvt.rn.f32.s32 %f112, %r617;
fma.rn.ftz.f32 %f29, %f111, %f112, %f29;
@%p13 bra $L__BB239_14;

mov.u64 %rd279, iq3xxs_grid;
add.s64 %rd152, %rd282, %rd11;
ld.global.u16 %rs47, [%rd152];
ld.global.u16 %rs48, [%rd152+2];
ld.global.u16 %rs49, [%rd152+4];
add.s64 %rd153, %rd282, %rd9;
ld.global.u16 %rs50, [%rd153];
cvt.u32.u16 %r667, %rs50;
ld.global.u16 %rs51, [%rd153+2];
cvt.u32.u16 %r668, %rs51;
mov.b32 %r669, {%rs50, %rs51};
ld.global.u16 %rs52, [%rd152+-2];
shl.b16 %rs53, %rs52, 2;
cvt.u64.u16 %rd154, %rs53;
and.b64 %rd155, %rd154, 1020;
add.s64 %rd157, %rd279, %rd155;
ld.global.nc.u32 %r670, [%rd157];
shr.u16 %rs54, %rs52, 8;
cvt.u32.u16 %r671, %rs54;
mul.wide.u32 %rd158, %r671, 4;
add.s64 %rd159, %rd279, %rd158;
ld.global.nc.u32 %r672, [%rd159];
and.b32 %r673, %r667, 255;
popc.b32 %r674, %r673;
shl.b32 %r675, %r674, 7;
and.b32 %r676, %r675, 128;
xor.b32 %r677, %r676, %r673;
mul.lo.s32 %r678, %r677, 16843009;
and.b32 %r679, %r678, 134480385;
or.b32 %r680, %r679, -2139062144;
add.s32 %r681, %r680, -16843009;
or.b32 %r618, %r681, %r679;

prmt.b32 %r618, %r618, 0, 0xba98;

xor.b32 %r682, %r618, %r670;
or.b32 %r683, %r682, -2139062144;
and.b32 %r684, %r618, 2139062143;
and.b32 %r685, %r670, -2139062144;
sub.s32 %r686, %r683, %r684;
xor.b32 %r687, %r685, %r686;
xor.b32 %r623, %r687, -2139062144;
and.b32 %r688, %r678, -2143281136;
or.b32 %r689, %r688, -2139062144;
add.s32 %r690, %r689, -16843009;
or.b32 %r620, %r690, %r688;

prmt.b32 %r620, %r620, 0, 0xba98;

xor.b32 %r691, %r620, %r672;
or.b32 %r692, %r691, -2139062144;
and.b32 %r693, %r620, 2139062143;
and.b32 %r694, %r672, -2139062144;
sub.s32 %r695, %r692, %r693;
xor.b32 %r696, %r694, %r695;
xor.b32 %r627, %r696, -2139062144;
mov.u32 %r625, 0;

dp4a.s32.s32 %r622, %r623, %r19, %r625;


dp4a.s32.s32 %r626, %r627, %r20, %r622;

shl.b16 %rs55, %rs47, 2;
cvt.u64.u16 %rd160, %rs55;
and.b64 %rd161, %rd160, 1020;
add.s64 %rd162, %rd279, %rd161;
ld.global.nc.u32 %r697, [%rd162];
shr.u16 %rs56, %rs47, 8;
cvt.u32.u16 %r698, %rs56;
mul.wide.u32 %rd163, %r698, 4;
add.s64 %rd164, %rd279, %rd163;
ld.global.nc.u32 %r699, [%rd164];
shr.u32 %r700, %r667, 7;
and.b32 %r701, %r700, 255;
popc.b32 %r702, %r701;
shl.b32 %r703, %r702, 7;
and.b32 %r704, %r703, 128;
xor.b32 %r705, %r704, %r701;
mul.lo.s32 %r706, %r705, 16843009;
and.b32 %r707, %r706, 134480385;
or.b32 %r708, %r707, -2139062144;
add.s32 %r709, %r708, -16843009;
or.b32 %r630, %r709, %r707;

prmt.b32 %r630, %r630, 0, 0xba98;

xor.b32 %r710, %r630, %r697;
or.b32 %r711, %r710, -2139062144;
and.b32 %r712, %r630, 2139062143;
and.b32 %r713, %r697, -2139062144;
sub.s32 %r714, %r711, %r712;
xor.b32 %r715, %r713, %r714;
xor.b32 %r635, %r715, -2139062144;
and.b32 %r716, %r706, -2143281136;
or.b32 %r717, %r716, -2139062144;
add.s32 %r718, %r717, -16843009;
or.b32 %r632, %r718, %r716;

prmt.b32 %r632, %r632, 0, 0xba98;

xor.b32 %r719, %r632, %r699;
or.b32 %r720, %r719, -2139062144;
and.b32 %r721, %r632, 2139062143;
and.b32 %r722, %r699, -2139062144;
sub.s32 %r723, %r720, %r721;
xor.b32 %r724, %r722, %r723;
xor.b32 %r639, %r724, -2139062144;

dp4a.s32.s32 %r634, %r635, %r21, %r626;


dp4a.s32.s32 %r638, %r639, %r22, %r634;

shl.b16 %rs57, %rs48, 2;
cvt.u64.u16 %rd165, %rs57;
and.b64 %rd166, %rd165, 1020;
add.s64 %rd167, %rd279, %rd166;
ld.global.nc.u32 %r725, [%rd167];
shr.u16 %rs58, %rs48, 8;
cvt.u32.u16 %r726, %rs58;
mul.wide.u32 %rd168, %r726, 4;
add.s64 %rd169, %rd279, %rd168;
ld.global.nc.u32 %r727, [%rd169];
shr.u32 %r728, %r669, 14;
and.b32 %r729, %r728, 255;
popc.b32 %r730, %r729;
shl.b32 %r731, %r730, 7;
and.b32 %r732, %r731, 128;
xor.b32 %r733, %r732, %r729;
mul.lo.s32 %r734, %r733, 16843009;
and.b32 %r735, %r734, 134480385;
or.b32 %r736, %r735, -2139062144;
add.s32 %r737, %r736, -16843009;
or.b32 %r642, %r737, %r735;

prmt.b32 %r642, %r642, 0, 0xba98;

xor.b32 %r738, %r642, %r725;
or.b32 %r739, %r738, -2139062144;
and.b32 %r740, %r642, 2139062143;
and.b32 %r741, %r725, -2139062144;
sub.s32 %r742, %r739, %r740;
xor.b32 %r743, %r741, %r742;
xor.b32 %r647, %r743, -2139062144;
and.b32 %r744, %r734, -2143281136;
or.b32 %r745, %r744, -2139062144;
add.s32 %r746, %r745, -16843009;
or.b32 %r644, %r746, %r744;

prmt.b32 %r644, %r644, 0, 0xba98;

xor.b32 %r747, %r644, %r727;
or.b32 %r748, %r747, -2139062144;
and.b32 %r749, %r644, 2139062143;
and.b32 %r750, %r727, -2139062144;
sub.s32 %r751, %r748, %r749;
xor.b32 %r752, %r750, %r751;
xor.b32 %r651, %r752, -2139062144;

dp4a.s32.s32 %r646, %r647, %r23, %r638;


dp4a.s32.s32 %r650, %r651, %r24, %r646;

shl.b16 %rs59, %rs49, 2;
cvt.u64.u16 %rd170, %rs59;
and.b64 %rd171, %rd170, 1020;
add.s64 %rd172, %rd279, %rd171;
ld.global.nc.u32 %r753, [%rd172];
shr.u16 %rs60, %rs49, 8;
cvt.u32.u16 %r754, %rs60;
mul.wide.u32 %rd173, %r754, 4;
add.s64 %rd174, %rd279, %rd173;
ld.global.nc.u32 %r755, [%rd174];
shr.u32 %r756, %r668, 5;
and.b32 %r757, %r756, 255;
popc.b32 %r758, %r757;
shl.b32 %r759, %r758, 7;
and.b32 %r760, %r759, 128;
xor.b32 %r761, %r760, %r757;
mul.lo.s32 %r762, %r761, 16843009;
and.b32 %r763, %r762, 134480385;
or.b32 %r764, %r763, -2139062144;
add.s32 %r765, %r764, -16843009;
or.b32 %r654, %r765, %r763;

prmt.b32 %r654, %r654, 0, 0xba98;

xor.b32 %r766, %r654, %r753;
or.b32 %r767, %r766, -2139062144;
and.b32 %r768, %r654, 2139062143;
and.b32 %r769, %r753, -2139062144;
sub.s32 %r770, %r767, %r768;
xor.b32 %r771, %r769, %r770;
xor.b32 %r659, %r771, -2139062144;
and.b32 %r772, %r762, -2143281136;
or.b32 %r773, %r772, -2139062144;
add.s32 %r774, %r773, -16843009;
or.b32 %r656, %r774, %r772;

prmt.b32 %r656, %r656, 0, 0xba98;

xor.b32 %r775, %r656, %r755;
or.b32 %r776, %r775, -2139062144;
and.b32 %r777, %r656, 2139062143;
and.b32 %r778, %r755, -2139062144;
sub.s32 %r779, %r776, %r777;
xor.b32 %r780, %r778, %r779;
xor.b32 %r663, %r780, -2139062144;

dp4a.s32.s32 %r658, %r659, %r25, %r650;

shr.u32 %r781, %r668, 12;

dp4a.s32.s32 %r662, %r663, %r26, %r658;

shr.u32 %r782, %r662, 31;
add.s32 %r783, %r662, %r782;
shr.s32 %r784, %r783, 1;
mad.lo.s32 %r785, %r662, %r781, %r784;
shr.u32 %r786, %r785, 31;
add.s32 %r787, %r785, %r786;
shr.s32 %r788, %r787, 1;
add.s64 %rd175, %rd282, %rd6;
ld.global.u16 %rs46, [%rd175];

{ cvt.f32.f16 %f113, %rs46;}



{.reg .f16 low,high;
mov.b32 {low,high},%r27;
cvt.f32.f16 %f114, low;}


mul.ftz.f32 %f115, %f113, %f114;
cvt.rn.f32.s32 %f116, %r788;
fma.rn.ftz.f32 %f33, %f115, %f116, %f33;

$L__BB239_14:
mov.u64 %rd278, iq3xxs_grid;
add.s64 %rd30, %rd27, %rd18;
ld.global.nc.u16 %rs62, [%rd30+2];
ld.global.nc.u16 %rs63, [%rd30+4];
ld.global.nc.u16 %rs64, [%rd30+6];
add.s64 %rd31, %rd28, %rd18;
ld.global.nc.u16 %rs65, [%rd31];
cvt.u32.u16 %r838, %rs65;
ld.global.nc.u16 %rs66, [%rd31+2];
cvt.u32.u16 %r839, %rs66;
mov.b32 %r840, {%rs65, %rs66};
ld.global.nc.u16 %rs67, [%rd30];
shl.b16 %rs68, %rs67, 2;
cvt.u64.u16 %rd176, %rs68;
and.b64 %rd177, %rd176, 1020;
add.s64 %rd179, %rd278, %rd177;
ld.global.nc.u32 %r841, [%rd179];
shr.u16 %rs69, %rs67, 8;
cvt.u32.u16 %r842, %rs69;
mul.wide.u32 %rd180, %r842, 4;
add.s64 %rd181, %rd278, %rd180;
ld.global.nc.u32 %r843, [%rd181];
and.b32 %r844, %r838, 255;
popc.b32 %r845, %r844;
shl.b32 %r846, %r845, 7;
and.b32 %r847, %r846, 128;
xor.b32 %r848, %r847, %r844;
mul.lo.s32 %r849, %r848, 16843009;
and.b32 %r850, %r849, 134480385;
or.b32 %r851, %r850, -2139062144;
add.s32 %r852, %r851, -16843009;
or.b32 %r789, %r852, %r850;

prmt.b32 %r789, %r789, 0, 0xba98;

xor.b32 %r853, %r789, %r841;
or.b32 %r854, %r853, -2139062144;
and.b32 %r855, %r789, 2139062143;
and.b32 %r856, %r841, -2139062144;
sub.s32 %r857, %r854, %r855;
xor.b32 %r858, %r856, %r857;
xor.b32 %r794, %r858, -2139062144;
and.b32 %r859, %r849, -2143281136;
or.b32 %r860, %r859, -2139062144;
add.s32 %r861, %r860, -16843009;
or.b32 %r791, %r861, %r859;

prmt.b32 %r791, %r791, 0, 0xba98;

xor.b32 %r862, %r791, %r843;
or.b32 %r863, %r862, -2139062144;
and.b32 %r864, %r791, 2139062143;
and.b32 %r865, %r843, -2139062144;
sub.s32 %r866, %r863, %r864;
xor.b32 %r867, %r865, %r866;
xor.b32 %r798, %r867, -2139062144;
mov.u32 %r796, 0;

dp4a.s32.s32 %r793, %r794, %r19, %r796;


dp4a.s32.s32 %r797, %r798, %r20, %r793;

shl.b16 %rs70, %rs62, 2;
cvt.u64.u16 %rd182, %rs70;
and.b64 %rd183, %rd182, 1020;
add.s64 %rd184, %rd278, %rd183;
ld.global.nc.u32 %r868, [%rd184];
shr.u16 %rs71, %rs62, 8;
cvt.u32.u16 %r869, %rs71;
mul.wide.u32 %rd185, %r869, 4;
add.s64 %rd186, %rd278, %rd185;
ld.global.nc.u32 %r870, [%rd186];
shr.u32 %r871, %r838, 7;
and.b32 %r872, %r871, 255;
popc.b32 %r873, %r872;
shl.b32 %r874, %r873, 7;
and.b32 %r875, %r874, 128;
xor.b32 %r876, %r875, %r872;
mul.lo.s32 %r877, %r876, 16843009;
and.b32 %r878, %r877, 134480385;
or.b32 %r879, %r878, -2139062144;
add.s32 %r880, %r879, -16843009;
or.b32 %r801, %r880, %r878;

prmt.b32 %r801, %r801, 0, 0xba98;

xor.b32 %r881, %r801, %r868;
or.b32 %r882, %r881, -2139062144;
and.b32 %r883, %r801, 2139062143;
and.b32 %r884, %r868, -2139062144;
sub.s32 %r885, %r882, %r883;
xor.b32 %r886, %r884, %r885;
xor.b32 %r806, %r886, -2139062144;
and.b32 %r887, %r877, -2143281136;
or.b32 %r888, %r887, -2139062144;
add.s32 %r889, %r888, -16843009;
or.b32 %r803, %r889, %r887;

prmt.b32 %r803, %r803, 0, 0xba98;

xor.b32 %r890, %r803, %r870;
or.b32 %r891, %r890, -2139062144;
and.b32 %r892, %r803, 2139062143;
and.b32 %r893, %r870, -2139062144;
sub.s32 %r894, %r891, %r892;
xor.b32 %r895, %r893, %r894;
xor.b32 %r810, %r895, -2139062144;

dp4a.s32.s32 %r805, %r806, %r21, %r797;


dp4a.s32.s32 %r809, %r810, %r22, %r805;

shl.b16 %rs72, %rs63, 2;
cvt.u64.u16 %rd187, %rs72;
and.b64 %rd188, %rd187, 1020;
add.s64 %rd189, %rd278, %rd188;
ld.global.nc.u32 %r896, [%rd189];
shr.u16 %rs73, %rs63, 8;
cvt.u32.u16 %r897, %rs73;
mul.wide.u32 %rd190, %r897, 4;
add.s64 %rd191, %rd278, %rd190;
ld.global.nc.u32 %r898, [%rd191];
shr.u32 %r899, %r840, 14;
and.b32 %r900, %r899, 255;
popc.b32 %r901, %r900;
shl.b32 %r902, %r901, 7;
and.b32 %r903, %r902, 128;
xor.b32 %r904, %r903, %r900;
mul.lo.s32 %r905, %r904, 16843009;
and.b32 %r906, %r905, 134480385;
or.b32 %r907, %r906, -2139062144;
add.s32 %r908, %r907, -16843009;
or.b32 %r813, %r908, %r906;

prmt.b32 %r813, %r813, 0, 0xba98;

xor.b32 %r909, %r813, %r896;
or.b32 %r910, %r909, -2139062144;
and.b32 %r911, %r813, 2139062143;
and.b32 %r912, %r896, -2139062144;
sub.s32 %r913, %r910, %r911;
xor.b32 %r914, %r912, %r913;
xor.b32 %r818, %r914, -2139062144;
and.b32 %r915, %r905, -2143281136;
or.b32 %r916, %r915, -2139062144;
add.s32 %r917, %r916, -16843009;
or.b32 %r815, %r917, %r915;

prmt.b32 %r815, %r815, 0, 0xba98;

xor.b32 %r918, %r815, %r898;
or.b32 %r919, %r918, -2139062144;
and.b32 %r920, %r815, 2139062143;
and.b32 %r921, %r898, -2139062144;
sub.s32 %r922, %r919, %r920;
xor.b32 %r923, %r921, %r922;
xor.b32 %r822, %r923, -2139062144;

dp4a.s32.s32 %r817, %r818, %r23, %r809;


dp4a.s32.s32 %r821, %r822, %r24, %r817;

shl.b16 %rs74, %rs64, 2;
cvt.u64.u16 %rd192, %rs74;
and.b64 %rd193, %rd192, 1020;
add.s64 %rd194, %rd278, %rd193;
ld.global.nc.u32 %r924, [%rd194];
shr.u16 %rs75, %rs64, 8;
cvt.u32.u16 %r925, %rs75;
mul.wide.u32 %rd195, %r925, 4;
add.s64 %rd196, %rd278, %rd195;
ld.global.nc.u32 %r926, [%rd196];
shr.u32 %r927, %r839, 5;
and.b32 %r928, %r927, 255;
popc.b32 %r929, %r928;
shl.b32 %r930, %r929, 7;
and.b32 %r931, %r930, 128;
xor.b32 %r932, %r931, %r928;
mul.lo.s32 %r933, %r932, 16843009;
and.b32 %r934, %r933, 134480385;
or.b32 %r935, %r934, -2139062144;
add.s32 %r936, %r935, -16843009;
or.b32 %r825, %r936, %r934;

prmt.b32 %r825, %r825, 0, 0xba98;

xor.b32 %r937, %r825, %r924;
or.b32 %r938, %r937, -2139062144;
and.b32 %r939, %r825, 2139062143;
and.b32 %r940, %r924, -2139062144;
sub.s32 %r941, %r938, %r939;
xor.b32 %r942, %r940, %r941;
xor.b32 %r830, %r942, -2139062144;
and.b32 %r943, %r933, -2143281136;
or.b32 %r944, %r943, -2139062144;
add.s32 %r945, %r944, -16843009;
or.b32 %r827, %r945, %r943;

prmt.b32 %r827, %r827, 0, 0xba98;

xor.b32 %r946, %r827, %r926;
or.b32 %r947, %r946, -2139062144;
and.b32 %r948, %r827, 2139062143;
and.b32 %r949, %r926, -2139062144;
sub.s32 %r950, %r947, %r948;
xor.b32 %r951, %r949, %r950;
xor.b32 %r834, %r951, -2139062144;

dp4a.s32.s32 %r829, %r830, %r25, %r821;

shr.u32 %r952, %r839, 12;

dp4a.s32.s32 %r833, %r834, %r26, %r829;

shr.u32 %r953, %r833, 31;
add.s32 %r954, %r833, %r953;
shr.s32 %r955, %r954, 1;
mad.lo.s32 %r956, %r833, %r952, %r955;
shr.u32 %r957, %r956, 31;
add.s32 %r958, %r956, %r957;
shr.s32 %r959, %r958, 1;
add.s64 %rd32, %rd29, %rd18;
ld.global.nc.u16 %rs61, [%rd32];

{ cvt.f32.f16 %f117, %rs61;}



{.reg .f16 low,high;
mov.b32 {low,high},%r27;
cvt.f32.f16 %f118, low;}


mul.ftz.f32 %f119, %f117, %f118;
cvt.rn.f32.s32 %f120, %r959;
fma.rn.ftz.f32 %f28, %f119, %f120, %f28;
@%p13 bra $L__BB239_16;

mov.u64 %rd276, iq3xxs_grid;
add.s64 %rd197, %rd282, %rd17;
ld.global.u16 %rs77, [%rd197+4];
ld.global.u16 %rs78, [%rd197+6];
ld.global.u16 %rs79, [%rd197+8];
add.s64 %rd198, %rd282, %rd16;
ld.global.u16 %rs80, [%rd198];
cvt.u32.u16 %r1009, %rs80;
ld.global.u16 %rs81, [%rd198+2];
cvt.u32.u16 %r1010, %rs81;
mov.b32 %r1011, {%rs80, %rs81};
ld.global.u16 %rs82, [%rd197+2];
shl.b16 %rs83, %rs82, 2;
cvt.u64.u16 %rd199, %rs83;
and.b64 %rd200, %rd199, 1020;
add.s64 %rd202, %rd276, %rd200;
ld.global.nc.u32 %r1012, [%rd202];
shr.u16 %rs84, %rs82, 8;
cvt.u32.u16 %r1013, %rs84;
mul.wide.u32 %rd203, %r1013, 4;
add.s64 %rd204, %rd276, %rd203;
ld.global.nc.u32 %r1014, [%rd204];
and.b32 %r1015, %r1009, 255;
popc.b32 %r1016, %r1015;
shl.b32 %r1017, %r1016, 7;
and.b32 %r1018, %r1017, 128;
xor.b32 %r1019, %r1018, %r1015;
mul.lo.s32 %r1020, %r1019, 16843009;
and.b32 %r1021, %r1020, 134480385;
or.b32 %r1022, %r1021, -2139062144;
add.s32 %r1023, %r1022, -16843009;
or.b32 %r960, %r1023, %r1021;

prmt.b32 %r960, %r960, 0, 0xba98;

xor.b32 %r1024, %r960, %r1012;
or.b32 %r1025, %r1024, -2139062144;
and.b32 %r1026, %r960, 2139062143;
and.b32 %r1027, %r1012, -2139062144;
sub.s32 %r1028, %r1025, %r1026;
xor.b32 %r1029, %r1027, %r1028;
xor.b32 %r965, %r1029, -2139062144;
and.b32 %r1030, %r1020, -2143281136;
or.b32 %r1031, %r1030, -2139062144;
add.s32 %r1032, %r1031, -16843009;
or.b32 %r962, %r1032, %r1030;

prmt.b32 %r962, %r962, 0, 0xba98;

xor.b32 %r1033, %r962, %r1014;
or.b32 %r1034, %r1033, -2139062144;
and.b32 %r1035, %r962, 2139062143;
and.b32 %r1036, %r1014, -2139062144;
sub.s32 %r1037, %r1034, %r1035;
xor.b32 %r1038, %r1036, %r1037;
xor.b32 %r969, %r1038, -2139062144;

dp4a.s32.s32 %r964, %r965, %r19, %r796;


dp4a.s32.s32 %r968, %r969, %r20, %r964;

shl.b16 %rs85, %rs77, 2;
cvt.u64.u16 %rd205, %rs85;
and.b64 %rd206, %rd205, 1020;
add.s64 %rd207, %rd276, %rd206;
ld.global.nc.u32 %r1039, [%rd207];
shr.u16 %rs86, %rs77, 8;
cvt.u32.u16 %r1040, %rs86;
mul.wide.u32 %rd208, %r1040, 4;
add.s64 %rd209, %rd276, %rd208;
ld.global.nc.u32 %r1041, [%rd209];
shr.u32 %r1042, %r1009, 7;
and.b32 %r1043, %r1042, 255;
popc.b32 %r1044, %r1043;
shl.b32 %r1045, %r1044, 7;
and.b32 %r1046, %r1045, 128;
xor.b32 %r1047, %r1046, %r1043;
mul.lo.s32 %r1048, %r1047, 16843009;
and.b32 %r1049, %r1048, 134480385;
or.b32 %r1050, %r1049, -2139062144;
add.s32 %r1051, %r1050, -16843009;
or.b32 %r972, %r1051, %r1049;

prmt.b32 %r972, %r972, 0, 0xba98;

xor.b32 %r1052, %r972, %r1039;
or.b32 %r1053, %r1052, -2139062144;
and.b32 %r1054, %r972, 2139062143;
and.b32 %r1055, %r1039, -2139062144;
sub.s32 %r1056, %r1053, %r1054;
xor.b32 %r1057, %r1055, %r1056;
xor.b32 %r977, %r1057, -2139062144;
and.b32 %r1058, %r1048, -2143281136;
or.b32 %r1059, %r1058, -2139062144;
add.s32 %r1060, %r1059, -16843009;
or.b32 %r974, %r1060, %r1058;

prmt.b32 %r974, %r974, 0, 0xba98;

xor.b32 %r1061, %r974, %r1041;
or.b32 %r1062, %r1061, -2139062144;
and.b32 %r1063, %r974, 2139062143;
and.b32 %r1064, %r1041, -2139062144;
sub.s32 %r1065, %r1062, %r1063;
xor.b32 %r1066, %r1064, %r1065;
xor.b32 %r981, %r1066, -2139062144;

dp4a.s32.s32 %r976, %r977, %r21, %r968;


dp4a.s32.s32 %r980, %r981, %r22, %r976;

shl.b16 %rs87, %rs78, 2;
cvt.u64.u16 %rd210, %rs87;
and.b64 %rd211, %rd210, 1020;
add.s64 %rd212, %rd276, %rd211;
ld.global.nc.u32 %r1067, [%rd212];
shr.u16 %rs88, %rs78, 8;
cvt.u32.u16 %r1068, %rs88;
mul.wide.u32 %rd213, %r1068, 4;
add.s64 %rd214, %rd276, %rd213;
ld.global.nc.u32 %r1069, [%rd214];
shr.u32 %r1070, %r1011, 14;
and.b32 %r1071, %r1070, 255;
popc.b32 %r1072, %r1071;
shl.b32 %r1073, %r1072, 7;
and.b32 %r1074, %r1073, 128;
xor.b32 %r1075, %r1074, %r1071;
mul.lo.s32 %r1076, %r1075, 16843009;
and.b32 %r1077, %r1076, 134480385;
or.b32 %r1078, %r1077, -2139062144;
add.s32 %r1079, %r1078, -16843009;
or.b32 %r984, %r1079, %r1077;

prmt.b32 %r984, %r984, 0, 0xba98;

xor.b32 %r1080, %r984, %r1067;
or.b32 %r1081, %r1080, -2139062144;
and.b32 %r1082, %r984, 2139062143;
and.b32 %r1083, %r1067, -2139062144;
sub.s32 %r1084, %r1081, %r1082;
xor.b32 %r1085, %r1083, %r1084;
xor.b32 %r989, %r1085, -2139062144;
and.b32 %r1086, %r1076, -2143281136;
or.b32 %r1087, %r1086, -2139062144;
add.s32 %r1088, %r1087, -16843009;
or.b32 %r986, %r1088, %r1086;

prmt.b32 %r986, %r986, 0, 0xba98;

xor.b32 %r1089, %r986, %r1069;
or.b32 %r1090, %r1089, -2139062144;
and.b32 %r1091, %r986, 2139062143;
and.b32 %r1092, %r1069, -2139062144;
sub.s32 %r1093, %r1090, %r1091;
xor.b32 %r1094, %r1092, %r1093;
xor.b32 %r993, %r1094, -2139062144;

dp4a.s32.s32 %r988, %r989, %r23, %r980;


dp4a.s32.s32 %r992, %r993, %r24, %r988;

shl.b16 %rs89, %rs79, 2;
cvt.u64.u16 %rd215, %rs89;
and.b64 %rd216, %rd215, 1020;
add.s64 %rd217, %rd276, %rd216;
ld.global.nc.u32 %r1095, [%rd217];
shr.u16 %rs90, %rs79, 8;
cvt.u32.u16 %r1096, %rs90;
mul.wide.u32 %rd218, %r1096, 4;
add.s64 %rd219, %rd276, %rd218;
ld.global.nc.u32 %r1097, [%rd219];
shr.u32 %r1098, %r1010, 5;
and.b32 %r1099, %r1098, 255;
popc.b32 %r1100, %r1099;
shl.b32 %r1101, %r1100, 7;
and.b32 %r1102, %r1101, 128;
xor.b32 %r1103, %r1102, %r1099;
mul.lo.s32 %r1104, %r1103, 16843009;
and.b32 %r1105, %r1104, 134480385;
or.b32 %r1106, %r1105, -2139062144;
add.s32 %r1107, %r1106, -16843009;
or.b32 %r996, %r1107, %r1105;

prmt.b32 %r996, %r996, 0, 0xba98;

xor.b32 %r1108, %r996, %r1095;
or.b32 %r1109, %r1108, -2139062144;
and.b32 %r1110, %r996, 2139062143;
and.b32 %r1111, %r1095, -2139062144;
sub.s32 %r1112, %r1109, %r1110;
xor.b32 %r1113, %r1111, %r1112;
xor.b32 %r1001, %r1113, -2139062144;
and.b32 %r1114, %r1104, -2143281136;
or.b32 %r1115, %r1114, -2139062144;
add.s32 %r1116, %r1115, -16843009;
or.b32 %r998, %r1116, %r1114;

prmt.b32 %r998, %r998, 0, 0xba98;

xor.b32 %r1117, %r998, %r1097;
or.b32 %r1118, %r1117, -2139062144;
and.b32 %r1119, %r998, 2139062143;
and.b32 %r1120, %r1097, -2139062144;
sub.s32 %r1121, %r1118, %r1119;
xor.b32 %r1122, %r1120, %r1121;
xor.b32 %r1005, %r1122, -2139062144;

dp4a.s32.s32 %r1000, %r1001, %r25, %r992;

shr.u32 %r1123, %r1010, 12;

dp4a.s32.s32 %r1004, %r1005, %r26, %r1000;

shr.u32 %r1124, %r1004, 31;
add.s32 %r1125, %r1004, %r1124;
shr.s32 %r1126, %r1125, 1;
mad.lo.s32 %r1127, %r1004, %r1123, %r1126;
shr.u32 %r1128, %r1127, 31;
add.s32 %r1129, %r1127, %r1128;
shr.s32 %r1130, %r1129, 1;
add.s64 %rd220, %rd282, %rd15;
ld.global.u16 %rs76, [%rd220];

{ cvt.f32.f16 %f121, %rs76;}



{.reg .f16 low,high;
mov.b32 {low,high},%r27;
cvt.f32.f16 %f122, low;}


mul.ftz.f32 %f123, %f121, %f122;
cvt.rn.f32.s32 %f124, %r1130;
fma.rn.ftz.f32 %f32, %f123, %f124, %f32;

$L__BB239_16:
mov.u64 %rd274, iq3xxs_grid;
add.s64 %rd221, %rd30, %rd18;
ld.global.nc.u16 %rs92, [%rd221+2];
ld.global.nc.u16 %rs93, [%rd221+4];
ld.global.nc.u16 %rs94, [%rd221+6];
add.s64 %rd222, %rd31, %rd18;
ld.global.nc.u16 %rs95, [%rd222];
cvt.u32.u16 %r1180, %rs95;
ld.global.nc.u16 %rs96, [%rd222+2];
cvt.u32.u16 %r1181, %rs96;
mov.b32 %r1182, {%rs95, %rs96};
ld.global.nc.u16 %rs97, [%rd221];
shl.b16 %rs98, %rs97, 2;
cvt.u64.u16 %rd223, %rs98;
and.b64 %rd224, %rd223, 1020;
add.s64 %rd226, %rd274, %rd224;
ld.global.nc.u32 %r1183, [%rd226];
shr.u16 %rs99, %rs97, 8;
cvt.u32.u16 %r1184, %rs99;
mul.wide.u32 %rd227, %r1184, 4;
add.s64 %rd228, %rd274, %rd227;
ld.global.nc.u32 %r1185, [%rd228];
and.b32 %r1186, %r1180, 255;
popc.b32 %r1187, %r1186;
shl.b32 %r1188, %r1187, 7;
and.b32 %r1189, %r1188, 128;
xor.b32 %r1190, %r1189, %r1186;
mul.lo.s32 %r1191, %r1190, 16843009;
and.b32 %r1192, %r1191, 134480385;
or.b32 %r1193, %r1192, -2139062144;
add.s32 %r1194, %r1193, -16843009;
or.b32 %r1131, %r1194, %r1192;

prmt.b32 %r1131, %r1131, 0, 0xba98;

xor.b32 %r1195, %r1131, %r1183;
or.b32 %r1196, %r1195, -2139062144;
and.b32 %r1197, %r1131, 2139062143;
and.b32 %r1198, %r1183, -2139062144;
sub.s32 %r1199, %r1196, %r1197;
xor.b32 %r1200, %r1198, %r1199;
xor.b32 %r1136, %r1200, -2139062144;
and.b32 %r1201, %r1191, -2143281136;
or.b32 %r1202, %r1201, -2139062144;
add.s32 %r1203, %r1202, -16843009;
or.b32 %r1133, %r1203, %r1201;

prmt.b32 %r1133, %r1133, 0, 0xba98;

xor.b32 %r1204, %r1133, %r1185;
or.b32 %r1205, %r1204, -2139062144;
and.b32 %r1206, %r1133, 2139062143;
and.b32 %r1207, %r1185, -2139062144;
sub.s32 %r1208, %r1205, %r1206;
xor.b32 %r1209, %r1207, %r1208;
xor.b32 %r1140, %r1209, -2139062144;

dp4a.s32.s32 %r1135, %r1136, %r19, %r796;


dp4a.s32.s32 %r1139, %r1140, %r20, %r1135;

shl.b16 %rs100, %rs92, 2;
cvt.u64.u16 %rd229, %rs100;
and.b64 %rd230, %rd229, 1020;
add.s64 %rd231, %rd274, %rd230;
ld.global.nc.u32 %r1210, [%rd231];
shr.u16 %rs101, %rs92, 8;
cvt.u32.u16 %r1211, %rs101;
mul.wide.u32 %rd232, %r1211, 4;
add.s64 %rd233, %rd274, %rd232;
ld.global.nc.u32 %r1212, [%rd233];
shr.u32 %r1213, %r1180, 7;
and.b32 %r1214, %r1213, 255;
popc.b32 %r1215, %r1214;
shl.b32 %r1216, %r1215, 7;
and.b32 %r1217, %r1216, 128;
xor.b32 %r1218, %r1217, %r1214;
mul.lo.s32 %r1219, %r1218, 16843009;
and.b32 %r1220, %r1219, 134480385;
or.b32 %r1221, %r1220, -2139062144;
add.s32 %r1222, %r1221, -16843009;
or.b32 %r1143, %r1222, %r1220;

prmt.b32 %r1143, %r1143, 0, 0xba98;

xor.b32 %r1223, %r1143, %r1210;
or.b32 %r1224, %r1223, -2139062144;
and.b32 %r1225, %r1143, 2139062143;
and.b32 %r1226, %r1210, -2139062144;
sub.s32 %r1227, %r1224, %r1225;
xor.b32 %r1228, %r1226, %r1227;
xor.b32 %r1148, %r1228, -2139062144;
and.b32 %r1229, %r1219, -2143281136;
or.b32 %r1230, %r1229, -2139062144;
add.s32 %r1231, %r1230, -16843009;
or.b32 %r1145, %r1231, %r1229;

prmt.b32 %r1145, %r1145, 0, 0xba98;

xor.b32 %r1232, %r1145, %r1212;
or.b32 %r1233, %r1232, -2139062144;
and.b32 %r1234, %r1145, 2139062143;
and.b32 %r1235, %r1212, -2139062144;
sub.s32 %r1236, %r1233, %r1234;
xor.b32 %r1237, %r1235, %r1236;
xor.b32 %r1152, %r1237, -2139062144;

dp4a.s32.s32 %r1147, %r1148, %r21, %r1139;


dp4a.s32.s32 %r1151, %r1152, %r22, %r1147;

shl.b16 %rs102, %rs93, 2;
cvt.u64.u16 %rd234, %rs102;
and.b64 %rd235, %rd234, 1020;
add.s64 %rd236, %rd274, %rd235;
ld.global.nc.u32 %r1238, [%rd236];
shr.u16 %rs103, %rs93, 8;
cvt.u32.u16 %r1239, %rs103;
mul.wide.u32 %rd237, %r1239, 4;
add.s64 %rd238, %rd274, %rd237;
ld.global.nc.u32 %r1240, [%rd238];
shr.u32 %r1241, %r1182, 14;
and.b32 %r1242, %r1241, 255;
popc.b32 %r1243, %r1242;
shl.b32 %r1244, %r1243, 7;
and.b32 %r1245, %r1244, 128;
xor.b32 %r1246, %r1245, %r1242;
mul.lo.s32 %r1247, %r1246, 16843009;
and.b32 %r1248, %r1247, 134480385;
or.b32 %r1249, %r1248, -2139062144;
add.s32 %r1250, %r1249, -16843009;
or.b32 %r1155, %r1250, %r1248;

prmt.b32 %r1155, %r1155, 0, 0xba98;

xor.b32 %r1251, %r1155, %r1238;
or.b32 %r1252, %r1251, -2139062144;
and.b32 %r1253, %r1155, 2139062143;
and.b32 %r1254, %r1238, -2139062144;
sub.s32 %r1255, %r1252, %r1253;
xor.b32 %r1256, %r1254, %r1255;
xor.b32 %r1160, %r1256, -2139062144;
and.b32 %r1257, %r1247, -2143281136;
or.b32 %r1258, %r1257, -2139062144;
add.s32 %r1259, %r1258, -16843009;
or.b32 %r1157, %r1259, %r1257;

prmt.b32 %r1157, %r1157, 0, 0xba98;

xor.b32 %r1260, %r1157, %r1240;
or.b32 %r1261, %r1260, -2139062144;
and.b32 %r1262, %r1157, 2139062143;
and.b32 %r1263, %r1240, -2139062144;
sub.s32 %r1264, %r1261, %r1262;
xor.b32 %r1265, %r1263, %r1264;
xor.b32 %r1164, %r1265, -2139062144;

dp4a.s32.s32 %r1159, %r1160, %r23, %r1151;


dp4a.s32.s32 %r1163, %r1164, %r24, %r1159;

shl.b16 %rs104, %rs94, 2;
cvt.u64.u16 %rd239, %rs104;
and.b64 %rd240, %rd239, 1020;
add.s64 %rd241, %rd274, %rd240;
ld.global.nc.u32 %r1266, [%rd241];
shr.u16 %rs105, %rs94, 8;
cvt.u32.u16 %r1267, %rs105;
mul.wide.u32 %rd242, %r1267, 4;
add.s64 %rd243, %rd274, %rd242;
ld.global.nc.u32 %r1268, [%rd243];
shr.u32 %r1269, %r1181, 5;
and.b32 %r1270, %r1269, 255;
popc.b32 %r1271, %r1270;
shl.b32 %r1272, %r1271, 7;
and.b32 %r1273, %r1272, 128;
xor.b32 %r1274, %r1273, %r1270;
mul.lo.s32 %r1275, %r1274, 16843009;
and.b32 %r1276, %r1275, 134480385;
or.b32 %r1277, %r1276, -2139062144;
add.s32 %r1278, %r1277, -16843009;
or.b32 %r1167, %r1278, %r1276;

prmt.b32 %r1167, %r1167, 0, 0xba98;

xor.b32 %r1279, %r1167, %r1266;
or.b32 %r1280, %r1279, -2139062144;
and.b32 %r1281, %r1167, 2139062143;
and.b32 %r1282, %r1266, -2139062144;
sub.s32 %r1283, %r1280, %r1281;
xor.b32 %r1284, %r1282, %r1283;
xor.b32 %r1172, %r1284, -2139062144;
and.b32 %r1285, %r1275, -2143281136;
or.b32 %r1286, %r1285, -2139062144;
add.s32 %r1287, %r1286, -16843009;
or.b32 %r1169, %r1287, %r1285;

prmt.b32 %r1169, %r1169, 0, 0xba98;

xor.b32 %r1288, %r1169, %r1268;
or.b32 %r1289, %r1288, -2139062144;
and.b32 %r1290, %r1169, 2139062143;
and.b32 %r1291, %r1268, -2139062144;
sub.s32 %r1292, %r1289, %r1290;
xor.b32 %r1293, %r1291, %r1292;
xor.b32 %r1176, %r1293, -2139062144;

dp4a.s32.s32 %r1171, %r1172, %r25, %r1163;

shr.u32 %r1294, %r1181, 12;

dp4a.s32.s32 %r1175, %r1176, %r26, %r1171;

shr.u32 %r1295, %r1175, 31;
add.s32 %r1296, %r1175, %r1295;
shr.s32 %r1297, %r1296, 1;
mad.lo.s32 %r1298, %r1175, %r1294, %r1297;
shr.u32 %r1299, %r1298, 31;
add.s32 %r1300, %r1298, %r1299;
shr.s32 %r1301, %r1300, 1;
add.s64 %rd244, %rd32, %rd18;
ld.global.nc.u16 %rs91, [%rd244];

{ cvt.f32.f16 %f125, %rs91;}



{.reg .f16 low,high;
mov.b32 {low,high},%r27;
cvt.f32.f16 %f126, low;}


mul.ftz.f32 %f127, %f125, %f126;
cvt.rn.f32.s32 %f128, %r1301;
fma.rn.ftz.f32 %f27, %f127, %f128, %f27;
@%p13 bra $L__BB239_18;

mov.u64 %rd275, iq3xxs_grid;
add.s64 %rd245, %rd282, %rd13;
ld.global.u16 %rs107, [%rd245+-4];
ld.global.u16 %rs108, [%rd245+-2];
ld.global.u16 %rs109, [%rd245];
add.s64 %rd246, %rd282, %rd14;
ld.global.u16 %rs110, [%rd246+-2];
cvt.u32.u16 %r1351, %rs110;
ld.global.u16 %rs111, [%rd246];
cvt.u32.u16 %r1352, %rs111;
mov.b32 %r1353, {%rs110, %rs111};
ld.global.u16 %rs112, [%rd245+-6];
shl.b16 %rs113, %rs112, 2;
cvt.u64.u16 %rd247, %rs113;
and.b64 %rd248, %rd247, 1020;
add.s64 %rd250, %rd275, %rd248;
ld.global.nc.u32 %r1354, [%rd250];
shr.u16 %rs114, %rs112, 8;
cvt.u32.u16 %r1355, %rs114;
mul.wide.u32 %rd251, %r1355, 4;
add.s64 %rd252, %rd275, %rd251;
ld.global.nc.u32 %r1356, [%rd252];
and.b32 %r1357, %r1351, 255;
popc.b32 %r1358, %r1357;
shl.b32 %r1359, %r1358, 7;
and.b32 %r1360, %r1359, 128;
xor.b32 %r1361, %r1360, %r1357;
mul.lo.s32 %r1362, %r1361, 16843009;
and.b32 %r1363, %r1362, 134480385;
or.b32 %r1364, %r1363, -2139062144;
add.s32 %r1365, %r1364, -16843009;
or.b32 %r1302, %r1365, %r1363;

prmt.b32 %r1302, %r1302, 0, 0xba98;

xor.b32 %r1366, %r1302, %r1354;
or.b32 %r1367, %r1366, -2139062144;
and.b32 %r1368, %r1302, 2139062143;
and.b32 %r1369, %r1354, -2139062144;
sub.s32 %r1370, %r1367, %r1368;
xor.b32 %r1371, %r1369, %r1370;
xor.b32 %r1307, %r1371, -2139062144;
and.b32 %r1372, %r1362, -2143281136;
or.b32 %r1373, %r1372, -2139062144;
add.s32 %r1374, %r1373, -16843009;
or.b32 %r1304, %r1374, %r1372;

prmt.b32 %r1304, %r1304, 0, 0xba98;

xor.b32 %r1375, %r1304, %r1356;
or.b32 %r1376, %r1375, -2139062144;
and.b32 %r1377, %r1304, 2139062143;
and.b32 %r1378, %r1356, -2139062144;
sub.s32 %r1379, %r1376, %r1377;
xor.b32 %r1380, %r1378, %r1379;
xor.b32 %r1311, %r1380, -2139062144;
mov.u32 %r1309, 0;

dp4a.s32.s32 %r1306, %r1307, %r19, %r1309;


dp4a.s32.s32 %r1310, %r1311, %r20, %r1306;

shl.b16 %rs115, %rs107, 2;
cvt.u64.u16 %rd253, %rs115;
and.b64 %rd254, %rd253, 1020;
add.s64 %rd255, %rd275, %rd254;
ld.global.nc.u32 %r1381, [%rd255];
shr.u16 %rs116, %rs107, 8;
cvt.u32.u16 %r1382, %rs116;
mul.wide.u32 %rd256, %r1382, 4;
add.s64 %rd257, %rd275, %rd256;
ld.global.nc.u32 %r1383, [%rd257];
shr.u32 %r1384, %r1351, 7;
and.b32 %r1385, %r1384, 255;
popc.b32 %r1386, %r1385;
shl.b32 %r1387, %r1386, 7;
and.b32 %r1388, %r1387, 128;
xor.b32 %r1389, %r1388, %r1385;
mul.lo.s32 %r1390, %r1389, 16843009;
and.b32 %r1391, %r1390, 134480385;
or.b32 %r1392, %r1391, -2139062144;
add.s32 %r1393, %r1392, -16843009;
or.b32 %r1314, %r1393, %r1391;

prmt.b32 %r1314, %r1314, 0, 0xba98;

xor.b32 %r1394, %r1314, %r1381;
or.b32 %r1395, %r1394, -2139062144;
and.b32 %r1396, %r1314, 2139062143;
and.b32 %r1397, %r1381, -2139062144;
sub.s32 %r1398, %r1395, %r1396;
xor.b32 %r1399, %r1397, %r1398;
xor.b32 %r1319, %r1399, -2139062144;
and.b32 %r1400, %r1390, -2143281136;
or.b32 %r1401, %r1400, -2139062144;
add.s32 %r1402, %r1401, -16843009;
or.b32 %r1316, %r1402, %r1400;

prmt.b32 %r1316, %r1316, 0, 0xba98;

xor.b32 %r1403, %r1316, %r1383;
or.b32 %r1404, %r1403, -2139062144;
and.b32 %r1405, %r1316, 2139062143;
and.b32 %r1406, %r1383, -2139062144;
sub.s32 %r1407, %r1404, %r1405;
xor.b32 %r1408, %r1406, %r1407;
xor.b32 %r1323, %r1408, -2139062144;

dp4a.s32.s32 %r1318, %r1319, %r21, %r1310;


dp4a.s32.s32 %r1322, %r1323, %r22, %r1318;

shl.b16 %rs117, %rs108, 2;
cvt.u64.u16 %rd258, %rs117;
and.b64 %rd259, %rd258, 1020;
add.s64 %rd260, %rd275, %rd259;
ld.global.nc.u32 %r1409, [%rd260];
shr.u16 %rs118, %rs108, 8;
cvt.u32.u16 %r1410, %rs118;
mul.wide.u32 %rd261, %r1410, 4;
add.s64 %rd262, %rd275, %rd261;
ld.global.nc.u32 %r1411, [%rd262];
shr.u32 %r1412, %r1353, 14;
and.b32 %r1413, %r1412, 255;
popc.b32 %r1414, %r1413;
shl.b32 %r1415, %r1414, 7;
and.b32 %r1416, %r1415, 128;
xor.b32 %r1417, %r1416, %r1413;
mul.lo.s32 %r1418, %r1417, 16843009;
and.b32 %r1419, %r1418, 134480385;
or.b32 %r1420, %r1419, -2139062144;
add.s32 %r1421, %r1420, -16843009;
or.b32 %r1326, %r1421, %r1419;

prmt.b32 %r1326, %r1326, 0, 0xba98;

xor.b32 %r1422, %r1326, %r1409;
or.b32 %r1423, %r1422, -2139062144;
and.b32 %r1424, %r1326, 2139062143;
and.b32 %r1425, %r1409, -2139062144;
sub.s32 %r1426, %r1423, %r1424;
xor.b32 %r1427, %r1425, %r1426;
xor.b32 %r1331, %r1427, -2139062144;
and.b32 %r1428, %r1418, -2143281136;
or.b32 %r1429, %r1428, -2139062144;
add.s32 %r1430, %r1429, -16843009;
or.b32 %r1328, %r1430, %r1428;

prmt.b32 %r1328, %r1328, 0, 0xba98;

xor.b32 %r1431, %r1328, %r1411;
or.b32 %r1432, %r1431, -2139062144;
and.b32 %r1433, %r1328, 2139062143;
and.b32 %r1434, %r1411, -2139062144;
sub.s32 %r1435, %r1432, %r1433;
xor.b32 %r1436, %r1434, %r1435;
xor.b32 %r1335, %r1436, -2139062144;

dp4a.s32.s32 %r1330, %r1331, %r23, %r1322;


dp4a.s32.s32 %r1334, %r1335, %r24, %r1330;

shl.b16 %rs119, %rs109, 2;
cvt.u64.u16 %rd263, %rs119;
and.b64 %rd264, %rd263, 1020;
add.s64 %rd265, %rd275, %rd264;
ld.global.nc.u32 %r1437, [%rd265];
shr.u16 %rs120, %rs109, 8;
cvt.u32.u16 %r1438, %rs120;
mul.wide.u32 %rd266, %r1438, 4;
add.s64 %rd267, %rd275, %rd266;
ld.global.nc.u32 %r1439, [%rd267];
shr.u32 %r1440, %r1352, 5;
and.b32 %r1441, %r1440, 255;
popc.b32 %r1442, %r1441;
shl.b32 %r1443, %r1442, 7;
and.b32 %r1444, %r1443, 128;
xor.b32 %r1445, %r1444, %r1441;
mul.lo.s32 %r1446, %r1445, 16843009;
and.b32 %r1447, %r1446, 134480385;
or.b32 %r1448, %r1447, -2139062144;
add.s32 %r1449, %r1448, -16843009;
or.b32 %r1338, %r1449, %r1447;

prmt.b32 %r1338, %r1338, 0, 0xba98;

xor.b32 %r1450, %r1338, %r1437;
or.b32 %r1451, %r1450, -2139062144;
and.b32 %r1452, %r1338, 2139062143;
and.b32 %r1453, %r1437, -2139062144;
sub.s32 %r1454, %r1451, %r1452;
xor.b32 %r1455, %r1453, %r1454;
xor.b32 %r1343, %r1455, -2139062144;
and.b32 %r1456, %r1446, -2143281136;
or.b32 %r1457, %r1456, -2139062144;
add.s32 %r1458, %r1457, -16843009;
or.b32 %r1340, %r1458, %r1456;

prmt.b32 %r1340, %r1340, 0, 0xba98;

xor.b32 %r1459, %r1340, %r1439;
or.b32 %r1460, %r1459, -2139062144;
and.b32 %r1461, %r1340, 2139062143;
and.b32 %r1462, %r1439, -2139062144;
sub.s32 %r1463, %r1460, %r1461;
xor.b32 %r1464, %r1462, %r1463;
xor.b32 %r1347, %r1464, -2139062144;

dp4a.s32.s32 %r1342, %r1343, %r25, %r1334;

shr.u32 %r1465, %r1352, 12;

dp4a.s32.s32 %r1346, %r1347, %r26, %r1342;

shr.u32 %r1466, %r1346, 31;
add.s32 %r1467, %r1346, %r1466;
shr.s32 %r1468, %r1467, 1;
mad.lo.s32 %r1469, %r1346, %r1465, %r1468;
shr.u32 %r1470, %r1469, 31;
add.s32 %r1471, %r1469, %r1470;
shr.s32 %r1472, %r1471, 1;
add.s64 %rd268, %rd282, %rd12;
ld.global.u16 %rs106, [%rd268];

{ cvt.f32.f16 %f129, %rs106;}



{.reg .f16 low,high;
mov.b32 {low,high},%r27;
cvt.f32.f16 %f130, low;}


mul.ftz.f32 %f131, %f129, %f130;
cvt.rn.f32.s32 %f132, %r1472;
fma.rn.ftz.f32 %f31, %f131, %f132, %f31;

$L__BB239_18:
add.s32 %r2137, %r2137, 128;
add.s64 %rd282, %rd282, 1568;
add.s64 %rd281, %rd281, 1568;
add.s32 %r2138, %r2138, 16;
setp.lt.s32 %p17, %r2138, %r5;
@%p17 bra $L__BB239_10;

$L__BB239_19:
mov.u32 %r2124, %tid.y;
setp.eq.s32 %p18, %r2124, 0;
@%p18 bra $L__BB239_23;

mov.u32 %r2125, %tid.y;
setp.ne.s64 %p19, %rd40, 0;
add.s32 %r30, %r2125, -1;
shl.b32 %r1473, %r2125, 9;
mov.u32 %r1474, _ZZ13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r1475, %r1474, %r1473;
shl.b32 %r1476, %r2, 2;
add.s32 %r1477, %r1475, %r1476;
st.shared.f32 [%r1477+-512], %f30;
@%p19 bra $L__BB239_22;
bra.uni $L__BB239_21;

$L__BB239_22:
shl.b32 %r1478, %r30, 9;
mov.u32 %r1479, _ZZ13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r1480, %r1479, %r1478;
add.s32 %r1482, %r1480, %r1476;
st.shared.f32 [%r1482], %f34;
add.s32 %r2121, %r1477, -512;
st.shared.f32 [%r2121+128], %f29;
st.shared.f32 [%r1482+128], %f33;
add.s32 %r2122, %r1477, -512;
st.shared.f32 [%r2122+256], %f28;
st.shared.f32 [%r1482+256], %f32;
add.s32 %r2123, %r1477, -512;
st.shared.f32 [%r2123+384], %f27;
st.shared.f32 [%r1482+384], %f31;
bra.uni $L__BB239_23;

$L__BB239_21:
add.s32 %r2118, %r1477, -512;
st.shared.f32 [%r2118+128], %f29;
add.s32 %r2119, %r1477, -512;
st.shared.f32 [%r2119+256], %f28;
add.s32 %r2120, %r1477, -512;
st.shared.f32 [%r2120+384], %f27;

$L__BB239_23:
mov.u32 %r2126, %tid.y;
setp.ne.s32 %p319, %r2126, 0;
bar.sync 0;
@%p319 bra $L__BB239_78;
bra.uni $L__BB239_24;

$L__BB239_78:
ret;

$L__BB239_24:
ld.param.u64 %rd273, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.f32 %f977, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
ld.param.u32 %r2129, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r2128, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
mov.u32 %r2127, %ctaid.z;
setp.ne.s64 %p21, %rd40, 0;
mov.u32 %r1484, %ctaid.y;
mad.lo.s32 %r1485, %r1484, %r2129, %r4;
mad.lo.s32 %r1486, %r2127, %r2128, %r1485;
neg.ftz.f32 %f35, %f977;
shl.b32 %r1487, %r2, 2;
mov.u32 %r1488, _ZZ13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r32, %r1488, %r1487;
cvta.to.global.u64 %rd269, %rd273;
mul.wide.u32 %rd270, %r1486, 4;
add.s64 %rd35, %rd269, %rd270;
@%p21 bra $L__BB239_33;
bra.uni $L__BB239_25;

$L__BB239_33:
ld.param.u32 %r2130, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
mov.u32 %r1562, _ZZ13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r34, %r1562, %r1487;
setp.gt.s32 %p54, %r2130, 2;
@%p54 bra $L__BB239_44;

ld.param.u32 %r2133, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p57, %r2133, 1;
@%p57 bra $L__BB239_62;

ld.param.u32 %r2134, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p58, %r2134, 2;
@%p58 bra $L__BB239_36;
bra.uni $L__BB239_70;

$L__BB239_36:
ld.shared.f32 %f533, [%r32];
add.ftz.f32 %f534, %f533, %f30;
ld.shared.f32 %f535, [%r34];
add.ftz.f32 %f536, %f535, %f34;
ld.shared.f32 %f537, [%r32+512];
add.ftz.f32 %f538, %f537, %f534;
ld.shared.f32 %f539, [%r34+512];
add.ftz.f32 %f540, %f539, %f536;
ld.shared.f32 %f541, [%r32+1024];
add.ftz.f32 %f542, %f541, %f538;
ld.shared.f32 %f543, [%r34+1024];
add.ftz.f32 %f544, %f543, %f540;
mov.b32 %r1785, %f542;
mov.u32 %r1786, 31;
mov.u32 %r1787, 16;
mov.u32 %r1788, -1;
shfl.sync.bfly.b32 %r1789|%p163, %r1785, %r1787, %r1786, %r1788;
mov.b32 %f545, %r1789;
add.ftz.f32 %f546, %f542, %f545;
mov.b32 %r1790, %f546;
mov.u32 %r1791, 8;
shfl.sync.bfly.b32 %r1792|%p164, %r1790, %r1791, %r1786, %r1788;
mov.b32 %f547, %r1792;
add.ftz.f32 %f548, %f546, %f547;
mov.b32 %r1793, %f548;
mov.u32 %r1794, 4;
shfl.sync.bfly.b32 %r1795|%p165, %r1793, %r1794, %r1786, %r1788;
mov.b32 %f549, %r1795;
add.ftz.f32 %f550, %f548, %f549;
mov.b32 %r1796, %f550;
mov.u32 %r1797, 2;
shfl.sync.bfly.b32 %r1798|%p166, %r1796, %r1797, %r1786, %r1788;
mov.b32 %f551, %r1798;
add.ftz.f32 %f552, %f550, %f551;
mov.b32 %r1799, %f552;
mov.u32 %r1800, 1;
shfl.sync.bfly.b32 %r1801|%p167, %r1799, %r1800, %r1786, %r1788;
mov.b32 %f553, %r1801;
add.ftz.f32 %f56, %f552, %f553;
mov.b32 %r1802, %f544;
shfl.sync.bfly.b32 %r1803|%p168, %r1802, %r1787, %r1786, %r1788;
mov.b32 %f554, %r1803;
add.ftz.f32 %f555, %f544, %f554;
mov.b32 %r1804, %f555;
shfl.sync.bfly.b32 %r1805|%p169, %r1804, %r1791, %r1786, %r1788;
mov.b32 %f556, %r1805;
add.ftz.f32 %f557, %f555, %f556;
mov.b32 %r1806, %f557;
shfl.sync.bfly.b32 %r1807|%p170, %r1806, %r1794, %r1786, %r1788;
mov.b32 %f558, %r1807;
add.ftz.f32 %f559, %f557, %f558;
mov.b32 %r1808, %f559;
shfl.sync.bfly.b32 %r1809|%p171, %r1808, %r1797, %r1786, %r1788;
mov.b32 %f560, %r1809;
add.ftz.f32 %f561, %f559, %f560;
mov.b32 %r1810, %f561;
shfl.sync.bfly.b32 %r1811|%p172, %r1810, %r1800, %r1786, %r1788;
mov.b32 %f562, %r1811;
add.ftz.f32 %f57, %f561, %f562;
setp.ge.u32 %p173, %r4, %r50;
setp.ne.s32 %p174, %r2, 0;
or.pred %p175, %p174, %p173;
@%p175 bra $L__BB239_38;

add.ftz.f32 %f563, %f5, %f56;
add.ftz.f32 %f564, %f4, %f57;
mul.ftz.f32 %f565, %f564, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f566, %f565;
add.ftz.f32 %f567, %f566, 0f3F800000;
div.approx.ftz.f32 %f568, %f564, %f567;
mul.ftz.f32 %f569, %f563, %f568;
st.global.f32 [%rd35], %f569;

$L__BB239_38:
ld.shared.f32 %f570, [%r32+128];
add.ftz.f32 %f571, %f570, %f29;
ld.shared.f32 %f572, [%r34+128];
add.ftz.f32 %f573, %f572, %f33;
ld.shared.f32 %f574, [%r32+640];
add.ftz.f32 %f575, %f574, %f571;
ld.shared.f32 %f576, [%r34+640];
add.ftz.f32 %f577, %f576, %f573;
ld.shared.f32 %f578, [%r32+1152];
add.ftz.f32 %f579, %f578, %f575;
ld.shared.f32 %f580, [%r34+1152];
add.ftz.f32 %f581, %f580, %f577;
mov.b32 %r1812, %f579;
shfl.sync.bfly.b32 %r1816|%p176, %r1812, %r1787, %r1786, %r1788;
mov.b32 %f582, %r1816;
add.ftz.f32 %f583, %f579, %f582;
mov.b32 %r1817, %f583;
shfl.sync.bfly.b32 %r1819|%p177, %r1817, %r1791, %r1786, %r1788;
mov.b32 %f584, %r1819;
add.ftz.f32 %f585, %f583, %f584;
mov.b32 %r1820, %f585;
shfl.sync.bfly.b32 %r1822|%p178, %r1820, %r1794, %r1786, %r1788;
mov.b32 %f586, %r1822;
add.ftz.f32 %f587, %f585, %f586;
mov.b32 %r1823, %f587;
shfl.sync.bfly.b32 %r1825|%p179, %r1823, %r1797, %r1786, %r1788;
mov.b32 %f588, %r1825;
add.ftz.f32 %f589, %f587, %f588;
mov.b32 %r1826, %f589;
shfl.sync.bfly.b32 %r1828|%p180, %r1826, %r1800, %r1786, %r1788;
mov.b32 %f590, %r1828;
add.ftz.f32 %f58, %f589, %f590;
mov.b32 %r1829, %f581;
shfl.sync.bfly.b32 %r1830|%p181, %r1829, %r1787, %r1786, %r1788;
mov.b32 %f591, %r1830;
add.ftz.f32 %f592, %f581, %f591;
mov.b32 %r1831, %f592;
shfl.sync.bfly.b32 %r1832|%p182, %r1831, %r1791, %r1786, %r1788;
mov.b32 %f593, %r1832;
add.ftz.f32 %f594, %f592, %f593;
mov.b32 %r1833, %f594;
shfl.sync.bfly.b32 %r1834|%p183, %r1833, %r1794, %r1786, %r1788;
mov.b32 %f595, %r1834;
add.ftz.f32 %f596, %f594, %f595;
mov.b32 %r1835, %f596;
shfl.sync.bfly.b32 %r1836|%p184, %r1835, %r1797, %r1786, %r1788;
mov.b32 %f597, %r1836;
add.ftz.f32 %f598, %f596, %f597;
mov.b32 %r1837, %f598;
shfl.sync.bfly.b32 %r1838|%p185, %r1837, %r1800, %r1786, %r1788;
mov.b32 %f599, %r1838;
add.ftz.f32 %f59, %f598, %f599;
add.s32 %r1839, %r4, 1;
setp.ge.u32 %p186, %r1839, %r50;
setp.ne.s32 %p187, %r2, 1;
or.pred %p188, %p187, %p186;
@%p188 bra $L__BB239_40;

add.ftz.f32 %f600, %f5, %f58;
add.ftz.f32 %f601, %f4, %f59;
mul.ftz.f32 %f602, %f601, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f603, %f602;
add.ftz.f32 %f604, %f603, 0f3F800000;
div.approx.ftz.f32 %f605, %f601, %f604;
mul.ftz.f32 %f606, %f600, %f605;
st.global.f32 [%rd35+4], %f606;

$L__BB239_40:
ld.shared.f32 %f607, [%r32+256];
add.ftz.f32 %f608, %f607, %f28;
ld.shared.f32 %f609, [%r34+256];
add.ftz.f32 %f610, %f609, %f32;
ld.shared.f32 %f611, [%r32+768];
add.ftz.f32 %f612, %f611, %f608;
ld.shared.f32 %f613, [%r34+768];
add.ftz.f32 %f614, %f613, %f610;
ld.shared.f32 %f615, [%r32+1280];
add.ftz.f32 %f616, %f615, %f612;
ld.shared.f32 %f617, [%r34+1280];
add.ftz.f32 %f618, %f617, %f614;
mov.b32 %r1840, %f616;
mov.u32 %r1841, 31;
mov.u32 %r1842, 16;
mov.u32 %r1843, -1;
shfl.sync.bfly.b32 %r1844|%p189, %r1840, %r1842, %r1841, %r1843;
mov.b32 %f619, %r1844;
add.ftz.f32 %f620, %f616, %f619;
mov.b32 %r1845, %f620;
mov.u32 %r1846, 8;
shfl.sync.bfly.b32 %r1847|%p190, %r1845, %r1846, %r1841, %r1843;
mov.b32 %f621, %r1847;
add.ftz.f32 %f622, %f620, %f621;
mov.b32 %r1848, %f622;
mov.u32 %r1849, 4;
shfl.sync.bfly.b32 %r1850|%p191, %r1848, %r1849, %r1841, %r1843;
mov.b32 %f623, %r1850;
add.ftz.f32 %f624, %f622, %f623;
mov.b32 %r1851, %f624;
mov.u32 %r1852, 2;
shfl.sync.bfly.b32 %r1853|%p192, %r1851, %r1852, %r1841, %r1843;
mov.b32 %f625, %r1853;
add.ftz.f32 %f626, %f624, %f625;
mov.b32 %r1854, %f626;
mov.u32 %r1855, 1;
shfl.sync.bfly.b32 %r1856|%p193, %r1854, %r1855, %r1841, %r1843;
mov.b32 %f627, %r1856;
add.ftz.f32 %f60, %f626, %f627;
mov.b32 %r1857, %f618;
shfl.sync.bfly.b32 %r1858|%p194, %r1857, %r1842, %r1841, %r1843;
mov.b32 %f628, %r1858;
add.ftz.f32 %f629, %f618, %f628;
mov.b32 %r1859, %f629;
shfl.sync.bfly.b32 %r1860|%p195, %r1859, %r1846, %r1841, %r1843;
mov.b32 %f630, %r1860;
add.ftz.f32 %f631, %f629, %f630;
mov.b32 %r1861, %f631;
shfl.sync.bfly.b32 %r1862|%p196, %r1861, %r1849, %r1841, %r1843;
mov.b32 %f632, %r1862;
add.ftz.f32 %f633, %f631, %f632;
mov.b32 %r1863, %f633;
shfl.sync.bfly.b32 %r1864|%p197, %r1863, %r1852, %r1841, %r1843;
mov.b32 %f634, %r1864;
add.ftz.f32 %f635, %f633, %f634;
mov.b32 %r1865, %f635;
shfl.sync.bfly.b32 %r1866|%p198, %r1865, %r1855, %r1841, %r1843;
mov.b32 %f636, %r1866;
add.ftz.f32 %f61, %f635, %f636;
add.s32 %r1867, %r4, 2;
setp.ge.u32 %p199, %r1867, %r50;
setp.ne.s32 %p200, %r2, 2;
or.pred %p201, %p200, %p199;
@%p201 bra $L__BB239_42;

add.ftz.f32 %f637, %f5, %f60;
add.ftz.f32 %f638, %f4, %f61;
mul.ftz.f32 %f639, %f638, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f640, %f639;
add.ftz.f32 %f641, %f640, 0f3F800000;
div.approx.ftz.f32 %f642, %f638, %f641;
mul.ftz.f32 %f643, %f637, %f642;
st.global.f32 [%rd35+8], %f643;

$L__BB239_42:
ld.shared.f32 %f644, [%r32+384];
add.ftz.f32 %f645, %f644, %f27;
ld.shared.f32 %f646, [%r34+384];
add.ftz.f32 %f647, %f646, %f31;
ld.shared.f32 %f648, [%r32+896];
add.ftz.f32 %f649, %f648, %f645;
ld.shared.f32 %f650, [%r34+896];
add.ftz.f32 %f651, %f650, %f647;
ld.shared.f32 %f652, [%r32+1408];
add.ftz.f32 %f653, %f652, %f649;
ld.shared.f32 %f654, [%r34+1408];
add.ftz.f32 %f655, %f654, %f651;
mov.b32 %r1868, %f653;
shfl.sync.bfly.b32 %r1872|%p202, %r1868, %r1842, %r1841, %r1843;
mov.b32 %f656, %r1872;
add.ftz.f32 %f657, %f653, %f656;
mov.b32 %r1873, %f657;
shfl.sync.bfly.b32 %r1875|%p203, %r1873, %r1846, %r1841, %r1843;
mov.b32 %f658, %r1875;
add.ftz.f32 %f659, %f657, %f658;
mov.b32 %r1876, %f659;
shfl.sync.bfly.b32 %r1878|%p204, %r1876, %r1849, %r1841, %r1843;
mov.b32 %f660, %r1878;
add.ftz.f32 %f661, %f659, %f660;
mov.b32 %r1879, %f661;
shfl.sync.bfly.b32 %r1881|%p205, %r1879, %r1852, %r1841, %r1843;
mov.b32 %f662, %r1881;
add.ftz.f32 %f663, %f661, %f662;
mov.b32 %r1882, %f663;
shfl.sync.bfly.b32 %r1884|%p206, %r1882, %r1855, %r1841, %r1843;
mov.b32 %f664, %r1884;
add.ftz.f32 %f62, %f663, %f664;
mov.b32 %r1885, %f655;
shfl.sync.bfly.b32 %r1886|%p207, %r1885, %r1842, %r1841, %r1843;
mov.b32 %f665, %r1886;
add.ftz.f32 %f666, %f655, %f665;
mov.b32 %r1887, %f666;
shfl.sync.bfly.b32 %r1888|%p208, %r1887, %r1846, %r1841, %r1843;
mov.b32 %f667, %r1888;
add.ftz.f32 %f668, %f666, %f667;
mov.b32 %r1889, %f668;
shfl.sync.bfly.b32 %r1890|%p209, %r1889, %r1849, %r1841, %r1843;
mov.b32 %f669, %r1890;
add.ftz.f32 %f670, %f668, %f669;
mov.b32 %r1891, %f670;
shfl.sync.bfly.b32 %r1892|%p210, %r1891, %r1852, %r1841, %r1843;
mov.b32 %f671, %r1892;
add.ftz.f32 %f672, %f670, %f671;
mov.b32 %r1893, %f672;
shfl.sync.bfly.b32 %r1894|%p211, %r1893, %r1855, %r1841, %r1843;
mov.b32 %f673, %r1894;
add.ftz.f32 %f63, %f672, %f673;
add.s32 %r1895, %r4, 3;
setp.ge.u32 %p212, %r1895, %r50;
setp.ne.s32 %p213, %r2, 3;
or.pred %p214, %p213, %p212;
@%p214 bra $L__BB239_78;

add.ftz.f32 %f674, %f5, %f62;
add.ftz.f32 %f675, %f4, %f63;
mul.ftz.f32 %f676, %f675, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f677, %f676;
add.ftz.f32 %f678, %f677, 0f3F800000;
div.approx.ftz.f32 %f679, %f675, %f678;
mul.ftz.f32 %f680, %f674, %f679;
st.global.f32 [%rd35+12], %f680;
bra.uni $L__BB239_78;

$L__BB239_25:
ld.shared.f32 %f133, [%r32];
add.ftz.f32 %f134, %f133, %f30;
ld.shared.f32 %f135, [%r32+512];
add.ftz.f32 %f136, %f135, %f134;
ld.shared.f32 %f137, [%r32+1024];
add.ftz.f32 %f138, %f137, %f136;
mov.b32 %r1489, %f138;
mov.u32 %r1490, 31;
mov.u32 %r1491, 16;
mov.u32 %r1492, -1;
shfl.sync.bfly.b32 %r1493|%p22, %r1489, %r1491, %r1490, %r1492;
mov.b32 %f139, %r1493;
add.ftz.f32 %f140, %f138, %f139;
mov.b32 %r1494, %f140;
mov.u32 %r1495, 8;
shfl.sync.bfly.b32 %r1496|%p23, %r1494, %r1495, %r1490, %r1492;
mov.b32 %f141, %r1496;
add.ftz.f32 %f142, %f140, %f141;
mov.b32 %r1497, %f142;
mov.u32 %r1498, 4;
shfl.sync.bfly.b32 %r1499|%p24, %r1497, %r1498, %r1490, %r1492;
mov.b32 %f143, %r1499;
add.ftz.f32 %f144, %f142, %f143;
mov.b32 %r1500, %f144;
mov.u32 %r1501, 2;
shfl.sync.bfly.b32 %r1502|%p25, %r1500, %r1501, %r1490, %r1492;
mov.b32 %f145, %r1502;
add.ftz.f32 %f146, %f144, %f145;
mov.b32 %r1503, %f146;
mov.u32 %r1504, 1;
shfl.sync.bfly.b32 %r1505|%p26, %r1503, %r1504, %r1490, %r1492;
mov.b32 %f147, %r1505;
add.ftz.f32 %f36, %f146, %f147;
setp.ge.u32 %p27, %r4, %r50;
setp.ne.s32 %p28, %r2, 0;
or.pred %p29, %p28, %p27;
@%p29 bra $L__BB239_27;

add.ftz.f32 %f148, %f5, %f36;
st.global.f32 [%rd35], %f148;

$L__BB239_27:
ld.shared.f32 %f149, [%r32+128];
add.ftz.f32 %f150, %f149, %f29;
ld.shared.f32 %f151, [%r32+640];
add.ftz.f32 %f152, %f151, %f150;
ld.shared.f32 %f153, [%r32+1152];
add.ftz.f32 %f154, %f153, %f152;
mov.b32 %r1506, %f154;
shfl.sync.bfly.b32 %r1510|%p30, %r1506, %r1491, %r1490, %r1492;
mov.b32 %f155, %r1510;
add.ftz.f32 %f156, %f154, %f155;
mov.b32 %r1511, %f156;
shfl.sync.bfly.b32 %r1513|%p31, %r1511, %r1495, %r1490, %r1492;
mov.b32 %f157, %r1513;
add.ftz.f32 %f158, %f156, %f157;
mov.b32 %r1514, %f158;
shfl.sync.bfly.b32 %r1516|%p32, %r1514, %r1498, %r1490, %r1492;
mov.b32 %f159, %r1516;
add.ftz.f32 %f160, %f158, %f159;
mov.b32 %r1517, %f160;
shfl.sync.bfly.b32 %r1519|%p33, %r1517, %r1501, %r1490, %r1492;
mov.b32 %f161, %r1519;
add.ftz.f32 %f162, %f160, %f161;
mov.b32 %r1520, %f162;
shfl.sync.bfly.b32 %r1522|%p34, %r1520, %r1504, %r1490, %r1492;
mov.b32 %f163, %r1522;
add.ftz.f32 %f37, %f162, %f163;
add.s32 %r1523, %r4, 1;
setp.ge.u32 %p35, %r1523, %r50;
setp.ne.s32 %p36, %r2, 1;
or.pred %p37, %p36, %p35;
@%p37 bra $L__BB239_29;

add.ftz.f32 %f164, %f5, %f37;
st.global.f32 [%rd35+4], %f164;

$L__BB239_29:
ld.shared.f32 %f165, [%r32+256];
add.ftz.f32 %f166, %f165, %f28;
ld.shared.f32 %f167, [%r32+768];
add.ftz.f32 %f168, %f167, %f166;
ld.shared.f32 %f169, [%r32+1280];
add.ftz.f32 %f170, %f169, %f168;
mov.b32 %r1524, %f170;
mov.u32 %r1525, 31;
mov.u32 %r1526, 16;
mov.u32 %r1527, -1;
shfl.sync.bfly.b32 %r1528|%p38, %r1524, %r1526, %r1525, %r1527;
mov.b32 %f171, %r1528;
add.ftz.f32 %f172, %f170, %f171;
mov.b32 %r1529, %f172;
mov.u32 %r1530, 8;
shfl.sync.bfly.b32 %r1531|%p39, %r1529, %r1530, %r1525, %r1527;
mov.b32 %f173, %r1531;
add.ftz.f32 %f174, %f172, %f173;
mov.b32 %r1532, %f174;
mov.u32 %r1533, 4;
shfl.sync.bfly.b32 %r1534|%p40, %r1532, %r1533, %r1525, %r1527;
mov.b32 %f175, %r1534;
add.ftz.f32 %f176, %f174, %f175;
mov.b32 %r1535, %f176;
mov.u32 %r1536, 2;
shfl.sync.bfly.b32 %r1537|%p41, %r1535, %r1536, %r1525, %r1527;
mov.b32 %f177, %r1537;
add.ftz.f32 %f178, %f176, %f177;
mov.b32 %r1538, %f178;
mov.u32 %r1539, 1;
shfl.sync.bfly.b32 %r1540|%p42, %r1538, %r1539, %r1525, %r1527;
mov.b32 %f179, %r1540;
add.ftz.f32 %f38, %f178, %f179;
add.s32 %r1541, %r4, 2;
setp.ge.u32 %p43, %r1541, %r50;
setp.ne.s32 %p44, %r2, 2;
or.pred %p45, %p44, %p43;
@%p45 bra $L__BB239_31;

add.ftz.f32 %f180, %f5, %f38;
st.global.f32 [%rd35+8], %f180;

$L__BB239_31:
ld.shared.f32 %f181, [%r32+384];
add.ftz.f32 %f182, %f181, %f27;
ld.shared.f32 %f183, [%r32+896];
add.ftz.f32 %f184, %f183, %f182;
ld.shared.f32 %f185, [%r32+1408];
add.ftz.f32 %f186, %f185, %f184;
mov.b32 %r1542, %f186;
shfl.sync.bfly.b32 %r1546|%p46, %r1542, %r1526, %r1525, %r1527;
mov.b32 %f187, %r1546;
add.ftz.f32 %f188, %f186, %f187;
mov.b32 %r1547, %f188;
shfl.sync.bfly.b32 %r1549|%p47, %r1547, %r1530, %r1525, %r1527;
mov.b32 %f189, %r1549;
add.ftz.f32 %f190, %f188, %f189;
mov.b32 %r1550, %f190;
shfl.sync.bfly.b32 %r1552|%p48, %r1550, %r1533, %r1525, %r1527;
mov.b32 %f191, %r1552;
add.ftz.f32 %f192, %f190, %f191;
mov.b32 %r1553, %f192;
shfl.sync.bfly.b32 %r1555|%p49, %r1553, %r1536, %r1525, %r1527;
mov.b32 %f193, %r1555;
add.ftz.f32 %f194, %f192, %f193;
mov.b32 %r1556, %f194;
shfl.sync.bfly.b32 %r1558|%p50, %r1556, %r1539, %r1525, %r1527;
mov.b32 %f195, %r1558;
add.ftz.f32 %f39, %f194, %f195;
add.s32 %r1559, %r4, 3;
setp.ge.u32 %p51, %r1559, %r50;
setp.ne.s32 %p52, %r2, 3;
or.pred %p53, %p52, %p51;
@%p53 bra $L__BB239_78;

add.ftz.f32 %f196, %f5, %f39;
st.global.f32 [%rd35+12], %f196;
bra.uni $L__BB239_78;

$L__BB239_44:
ld.param.u32 %r2131, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p55, %r2131, 3;
@%p55 bra $L__BB239_54;

ld.param.u32 %r2132, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.ne.s32 %p56, %r2132, 6;
@%p56 bra $L__BB239_70;

ld.shared.f32 %f197, [%r32];
add.ftz.f32 %f198, %f197, %f30;
ld.shared.f32 %f199, [%r34];
add.ftz.f32 %f200, %f199, %f34;
ld.shared.f32 %f201, [%r32+512];
add.ftz.f32 %f202, %f201, %f198;
ld.shared.f32 %f203, [%r34+512];
add.ftz.f32 %f204, %f203, %f200;
ld.shared.f32 %f205, [%r32+1024];
add.ftz.f32 %f206, %f205, %f202;
ld.shared.f32 %f207, [%r34+1024];
add.ftz.f32 %f208, %f207, %f204;
mov.b32 %r1563, %f206;
mov.u32 %r1564, 31;
mov.u32 %r1565, 16;
mov.u32 %r1566, -1;
shfl.sync.bfly.b32 %r1567|%p59, %r1563, %r1565, %r1564, %r1566;
mov.b32 %f209, %r1567;
add.ftz.f32 %f210, %f206, %f209;
mov.b32 %r1568, %f210;
mov.u32 %r1569, 8;
shfl.sync.bfly.b32 %r1570|%p60, %r1568, %r1569, %r1564, %r1566;
mov.b32 %f211, %r1570;
add.ftz.f32 %f212, %f210, %f211;
mov.b32 %r1571, %f212;
mov.u32 %r1572, 4;
shfl.sync.bfly.b32 %r1573|%p61, %r1571, %r1572, %r1564, %r1566;
mov.b32 %f213, %r1573;
add.ftz.f32 %f214, %f212, %f213;
mov.b32 %r1574, %f214;
mov.u32 %r1575, 2;
shfl.sync.bfly.b32 %r1576|%p62, %r1574, %r1575, %r1564, %r1566;
mov.b32 %f215, %r1576;
add.ftz.f32 %f216, %f214, %f215;
mov.b32 %r1577, %f216;
mov.u32 %r1578, 1;
shfl.sync.bfly.b32 %r1579|%p63, %r1577, %r1578, %r1564, %r1566;
mov.b32 %f217, %r1579;
add.ftz.f32 %f40, %f216, %f217;
mov.b32 %r1580, %f208;
shfl.sync.bfly.b32 %r1581|%p64, %r1580, %r1565, %r1564, %r1566;
mov.b32 %f218, %r1581;
add.ftz.f32 %f219, %f208, %f218;
mov.b32 %r1582, %f219;
shfl.sync.bfly.b32 %r1583|%p65, %r1582, %r1569, %r1564, %r1566;
mov.b32 %f220, %r1583;
add.ftz.f32 %f221, %f219, %f220;
mov.b32 %r1584, %f221;
shfl.sync.bfly.b32 %r1585|%p66, %r1584, %r1572, %r1564, %r1566;
mov.b32 %f222, %r1585;
add.ftz.f32 %f223, %f221, %f222;
mov.b32 %r1586, %f223;
shfl.sync.bfly.b32 %r1587|%p67, %r1586, %r1575, %r1564, %r1566;
mov.b32 %f224, %r1587;
add.ftz.f32 %f225, %f223, %f224;
mov.b32 %r1588, %f225;
shfl.sync.bfly.b32 %r1589|%p68, %r1588, %r1578, %r1564, %r1566;
mov.b32 %f226, %r1589;
add.ftz.f32 %f41, %f225, %f226;
setp.ge.u32 %p69, %r4, %r50;
setp.ne.s32 %p70, %r2, 0;
or.pred %p71, %p70, %p69;
@%p71 bra $L__BB239_48;

ld.param.f32 %f981, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f227, %f5, %f40;
add.ftz.f32 %f228, %f4, %f41;
min.ftz.f32 %f229, %f228, %f981;
min.ftz.f32 %f230, %f227, %f981;
max.ftz.f32 %f231, %f230, %f35;
mul.ftz.f32 %f232, %f229, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f233, %f232;
add.ftz.f32 %f234, %f233, 0f3F800000;
div.approx.ftz.f32 %f235, %f229, %f234;
mul.ftz.f32 %f236, %f231, %f235;
st.global.f32 [%rd35], %f236;

$L__BB239_48:
ld.shared.f32 %f237, [%r32+128];
add.ftz.f32 %f238, %f237, %f29;
ld.shared.f32 %f239, [%r34+128];
add.ftz.f32 %f240, %f239, %f33;
ld.shared.f32 %f241, [%r32+640];
add.ftz.f32 %f242, %f241, %f238;
ld.shared.f32 %f243, [%r34+640];
add.ftz.f32 %f244, %f243, %f240;
ld.shared.f32 %f245, [%r32+1152];
add.ftz.f32 %f246, %f245, %f242;
ld.shared.f32 %f247, [%r34+1152];
add.ftz.f32 %f248, %f247, %f244;
mov.b32 %r1590, %f246;
shfl.sync.bfly.b32 %r1594|%p72, %r1590, %r1565, %r1564, %r1566;
mov.b32 %f249, %r1594;
add.ftz.f32 %f250, %f246, %f249;
mov.b32 %r1595, %f250;
shfl.sync.bfly.b32 %r1597|%p73, %r1595, %r1569, %r1564, %r1566;
mov.b32 %f251, %r1597;
add.ftz.f32 %f252, %f250, %f251;
mov.b32 %r1598, %f252;
shfl.sync.bfly.b32 %r1600|%p74, %r1598, %r1572, %r1564, %r1566;
mov.b32 %f253, %r1600;
add.ftz.f32 %f254, %f252, %f253;
mov.b32 %r1601, %f254;
shfl.sync.bfly.b32 %r1603|%p75, %r1601, %r1575, %r1564, %r1566;
mov.b32 %f255, %r1603;
add.ftz.f32 %f256, %f254, %f255;
mov.b32 %r1604, %f256;
shfl.sync.bfly.b32 %r1606|%p76, %r1604, %r1578, %r1564, %r1566;
mov.b32 %f257, %r1606;
add.ftz.f32 %f42, %f256, %f257;
mov.b32 %r1607, %f248;
shfl.sync.bfly.b32 %r1608|%p77, %r1607, %r1565, %r1564, %r1566;
mov.b32 %f258, %r1608;
add.ftz.f32 %f259, %f248, %f258;
mov.b32 %r1609, %f259;
shfl.sync.bfly.b32 %r1610|%p78, %r1609, %r1569, %r1564, %r1566;
mov.b32 %f260, %r1610;
add.ftz.f32 %f261, %f259, %f260;
mov.b32 %r1611, %f261;
shfl.sync.bfly.b32 %r1612|%p79, %r1611, %r1572, %r1564, %r1566;
mov.b32 %f262, %r1612;
add.ftz.f32 %f263, %f261, %f262;
mov.b32 %r1613, %f263;
shfl.sync.bfly.b32 %r1614|%p80, %r1613, %r1575, %r1564, %r1566;
mov.b32 %f264, %r1614;
add.ftz.f32 %f265, %f263, %f264;
mov.b32 %r1615, %f265;
shfl.sync.bfly.b32 %r1616|%p81, %r1615, %r1578, %r1564, %r1566;
mov.b32 %f266, %r1616;
add.ftz.f32 %f43, %f265, %f266;
add.s32 %r1617, %r4, 1;
setp.ge.u32 %p82, %r1617, %r50;
setp.ne.s32 %p83, %r2, 1;
or.pred %p84, %p83, %p82;
@%p84 bra $L__BB239_50;

ld.param.f32 %f980, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f267, %f5, %f42;
add.ftz.f32 %f268, %f4, %f43;
min.ftz.f32 %f269, %f268, %f980;
min.ftz.f32 %f270, %f267, %f980;
max.ftz.f32 %f271, %f270, %f35;
mul.ftz.f32 %f272, %f269, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f273, %f272;
add.ftz.f32 %f274, %f273, 0f3F800000;
div.approx.ftz.f32 %f275, %f269, %f274;
mul.ftz.f32 %f276, %f271, %f275;
st.global.f32 [%rd35+4], %f276;

$L__BB239_50:
ld.shared.f32 %f277, [%r32+256];
add.ftz.f32 %f278, %f277, %f28;
ld.shared.f32 %f279, [%r34+256];
add.ftz.f32 %f280, %f279, %f32;
ld.shared.f32 %f281, [%r32+768];
add.ftz.f32 %f282, %f281, %f278;
ld.shared.f32 %f283, [%r34+768];
add.ftz.f32 %f284, %f283, %f280;
ld.shared.f32 %f285, [%r32+1280];
add.ftz.f32 %f286, %f285, %f282;
ld.shared.f32 %f287, [%r34+1280];
add.ftz.f32 %f288, %f287, %f284;
mov.b32 %r1618, %f286;
mov.u32 %r1619, 31;
mov.u32 %r1620, 16;
mov.u32 %r1621, -1;
shfl.sync.bfly.b32 %r1622|%p85, %r1618, %r1620, %r1619, %r1621;
mov.b32 %f289, %r1622;
add.ftz.f32 %f290, %f286, %f289;
mov.b32 %r1623, %f290;
mov.u32 %r1624, 8;
shfl.sync.bfly.b32 %r1625|%p86, %r1623, %r1624, %r1619, %r1621;
mov.b32 %f291, %r1625;
add.ftz.f32 %f292, %f290, %f291;
mov.b32 %r1626, %f292;
mov.u32 %r1627, 4;
shfl.sync.bfly.b32 %r1628|%p87, %r1626, %r1627, %r1619, %r1621;
mov.b32 %f293, %r1628;
add.ftz.f32 %f294, %f292, %f293;
mov.b32 %r1629, %f294;
mov.u32 %r1630, 2;
shfl.sync.bfly.b32 %r1631|%p88, %r1629, %r1630, %r1619, %r1621;
mov.b32 %f295, %r1631;
add.ftz.f32 %f296, %f294, %f295;
mov.b32 %r1632, %f296;
mov.u32 %r1633, 1;
shfl.sync.bfly.b32 %r1634|%p89, %r1632, %r1633, %r1619, %r1621;
mov.b32 %f297, %r1634;
add.ftz.f32 %f44, %f296, %f297;
mov.b32 %r1635, %f288;
shfl.sync.bfly.b32 %r1636|%p90, %r1635, %r1620, %r1619, %r1621;
mov.b32 %f298, %r1636;
add.ftz.f32 %f299, %f288, %f298;
mov.b32 %r1637, %f299;
shfl.sync.bfly.b32 %r1638|%p91, %r1637, %r1624, %r1619, %r1621;
mov.b32 %f300, %r1638;
add.ftz.f32 %f301, %f299, %f300;
mov.b32 %r1639, %f301;
shfl.sync.bfly.b32 %r1640|%p92, %r1639, %r1627, %r1619, %r1621;
mov.b32 %f302, %r1640;
add.ftz.f32 %f303, %f301, %f302;
mov.b32 %r1641, %f303;
shfl.sync.bfly.b32 %r1642|%p93, %r1641, %r1630, %r1619, %r1621;
mov.b32 %f304, %r1642;
add.ftz.f32 %f305, %f303, %f304;
mov.b32 %r1643, %f305;
shfl.sync.bfly.b32 %r1644|%p94, %r1643, %r1633, %r1619, %r1621;
mov.b32 %f306, %r1644;
add.ftz.f32 %f45, %f305, %f306;
add.s32 %r1645, %r4, 2;
setp.ge.u32 %p95, %r1645, %r50;
setp.ne.s32 %p96, %r2, 2;
or.pred %p97, %p96, %p95;
@%p97 bra $L__BB239_52;

ld.param.f32 %f979, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f307, %f5, %f44;
add.ftz.f32 %f308, %f4, %f45;
min.ftz.f32 %f309, %f308, %f979;
min.ftz.f32 %f310, %f307, %f979;
max.ftz.f32 %f311, %f310, %f35;
mul.ftz.f32 %f312, %f309, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f313, %f312;
add.ftz.f32 %f314, %f313, 0f3F800000;
div.approx.ftz.f32 %f315, %f309, %f314;
mul.ftz.f32 %f316, %f311, %f315;
st.global.f32 [%rd35+8], %f316;

$L__BB239_52:
ld.shared.f32 %f317, [%r32+384];
add.ftz.f32 %f318, %f317, %f27;
ld.shared.f32 %f319, [%r34+384];
add.ftz.f32 %f320, %f319, %f31;
ld.shared.f32 %f321, [%r32+896];
add.ftz.f32 %f322, %f321, %f318;
ld.shared.f32 %f323, [%r34+896];
add.ftz.f32 %f324, %f323, %f320;
ld.shared.f32 %f325, [%r32+1408];
add.ftz.f32 %f326, %f325, %f322;
ld.shared.f32 %f327, [%r34+1408];
add.ftz.f32 %f328, %f327, %f324;
mov.b32 %r1646, %f326;
shfl.sync.bfly.b32 %r1650|%p98, %r1646, %r1620, %r1619, %r1621;
mov.b32 %f329, %r1650;
add.ftz.f32 %f330, %f326, %f329;
mov.b32 %r1651, %f330;
shfl.sync.bfly.b32 %r1653|%p99, %r1651, %r1624, %r1619, %r1621;
mov.b32 %f331, %r1653;
add.ftz.f32 %f332, %f330, %f331;
mov.b32 %r1654, %f332;
shfl.sync.bfly.b32 %r1656|%p100, %r1654, %r1627, %r1619, %r1621;
mov.b32 %f333, %r1656;
add.ftz.f32 %f334, %f332, %f333;
mov.b32 %r1657, %f334;
shfl.sync.bfly.b32 %r1659|%p101, %r1657, %r1630, %r1619, %r1621;
mov.b32 %f335, %r1659;
add.ftz.f32 %f336, %f334, %f335;
mov.b32 %r1660, %f336;
shfl.sync.bfly.b32 %r1662|%p102, %r1660, %r1633, %r1619, %r1621;
mov.b32 %f337, %r1662;
add.ftz.f32 %f46, %f336, %f337;
mov.b32 %r1663, %f328;
shfl.sync.bfly.b32 %r1664|%p103, %r1663, %r1620, %r1619, %r1621;
mov.b32 %f338, %r1664;
add.ftz.f32 %f339, %f328, %f338;
mov.b32 %r1665, %f339;
shfl.sync.bfly.b32 %r1666|%p104, %r1665, %r1624, %r1619, %r1621;
mov.b32 %f340, %r1666;
add.ftz.f32 %f341, %f339, %f340;
mov.b32 %r1667, %f341;
shfl.sync.bfly.b32 %r1668|%p105, %r1667, %r1627, %r1619, %r1621;
mov.b32 %f342, %r1668;
add.ftz.f32 %f343, %f341, %f342;
mov.b32 %r1669, %f343;
shfl.sync.bfly.b32 %r1670|%p106, %r1669, %r1630, %r1619, %r1621;
mov.b32 %f344, %r1670;
add.ftz.f32 %f345, %f343, %f344;
mov.b32 %r1671, %f345;
shfl.sync.bfly.b32 %r1672|%p107, %r1671, %r1633, %r1619, %r1621;
mov.b32 %f346, %r1672;
add.ftz.f32 %f47, %f345, %f346;
add.s32 %r1673, %r4, 3;
setp.ge.u32 %p108, %r1673, %r50;
setp.ne.s32 %p109, %r2, 3;
or.pred %p110, %p109, %p108;
@%p110 bra $L__BB239_78;

ld.param.f32 %f978, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f347, %f5, %f46;
add.ftz.f32 %f348, %f4, %f47;
min.ftz.f32 %f349, %f348, %f978;
min.ftz.f32 %f350, %f347, %f978;
max.ftz.f32 %f351, %f350, %f35;
mul.ftz.f32 %f352, %f349, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f353, %f352;
add.ftz.f32 %f354, %f353, 0f3F800000;
div.approx.ftz.f32 %f355, %f349, %f354;
mul.ftz.f32 %f356, %f351, %f355;
st.global.f32 [%rd35+12], %f356;
bra.uni $L__BB239_78;

$L__BB239_70:
ld.shared.f32 %f845, [%r32];
add.ftz.f32 %f846, %f845, %f30;
ld.shared.f32 %f847, [%r34];
add.ftz.f32 %f848, %f847, %f34;
ld.shared.f32 %f849, [%r32+512];
add.ftz.f32 %f850, %f849, %f846;
ld.shared.f32 %f851, [%r34+512];
add.ftz.f32 %f852, %f851, %f848;
ld.shared.f32 %f853, [%r32+1024];
add.ftz.f32 %f854, %f853, %f850;
ld.shared.f32 %f855, [%r34+1024];
add.ftz.f32 %f856, %f855, %f852;
mov.b32 %r2007, %f854;
mov.u32 %r2008, 31;
mov.u32 %r2009, 16;
mov.u32 %r2010, -1;
shfl.sync.bfly.b32 %r2011|%p267, %r2007, %r2009, %r2008, %r2010;
mov.b32 %f857, %r2011;
add.ftz.f32 %f858, %f854, %f857;
mov.b32 %r2012, %f858;
mov.u32 %r2013, 8;
shfl.sync.bfly.b32 %r2014|%p268, %r2012, %r2013, %r2008, %r2010;
mov.b32 %f859, %r2014;
add.ftz.f32 %f860, %f858, %f859;
mov.b32 %r2015, %f860;
mov.u32 %r2016, 4;
shfl.sync.bfly.b32 %r2017|%p269, %r2015, %r2016, %r2008, %r2010;
mov.b32 %f861, %r2017;
add.ftz.f32 %f862, %f860, %f861;
mov.b32 %r2018, %f862;
mov.u32 %r2019, 2;
shfl.sync.bfly.b32 %r2020|%p270, %r2018, %r2019, %r2008, %r2010;
mov.b32 %f863, %r2020;
add.ftz.f32 %f864, %f862, %f863;
mov.b32 %r2021, %f864;
mov.u32 %r2022, 1;
shfl.sync.bfly.b32 %r2023|%p271, %r2021, %r2022, %r2008, %r2010;
mov.b32 %f865, %r2023;
add.ftz.f32 %f72, %f864, %f865;
mov.b32 %r2024, %f856;
shfl.sync.bfly.b32 %r2025|%p272, %r2024, %r2009, %r2008, %r2010;
mov.b32 %f866, %r2025;
add.ftz.f32 %f867, %f856, %f866;
mov.b32 %r2026, %f867;
shfl.sync.bfly.b32 %r2027|%p273, %r2026, %r2013, %r2008, %r2010;
mov.b32 %f868, %r2027;
add.ftz.f32 %f869, %f867, %f868;
mov.b32 %r2028, %f869;
shfl.sync.bfly.b32 %r2029|%p274, %r2028, %r2016, %r2008, %r2010;
mov.b32 %f870, %r2029;
add.ftz.f32 %f871, %f869, %f870;
mov.b32 %r2030, %f871;
shfl.sync.bfly.b32 %r2031|%p275, %r2030, %r2019, %r2008, %r2010;
mov.b32 %f872, %r2031;
add.ftz.f32 %f873, %f871, %f872;
mov.b32 %r2032, %f873;
shfl.sync.bfly.b32 %r2033|%p276, %r2032, %r2022, %r2008, %r2010;
mov.b32 %f874, %r2033;
add.ftz.f32 %f73, %f873, %f874;
setp.ge.u32 %p277, %r4, %r50;
setp.ne.s32 %p278, %r2, 0;
or.pred %p279, %p278, %p277;
@%p279 bra $L__BB239_72;

add.ftz.f32 %f875, %f5, %f72;
add.ftz.f32 %f876, %f4, %f73;
mul.ftz.f32 %f877, %f875, %f876;
st.global.f32 [%rd35], %f877;

$L__BB239_72:
ld.shared.f32 %f878, [%r32+128];
add.ftz.f32 %f879, %f878, %f29;
ld.shared.f32 %f880, [%r34+128];
add.ftz.f32 %f881, %f880, %f33;
ld.shared.f32 %f882, [%r32+640];
add.ftz.f32 %f883, %f882, %f879;
ld.shared.f32 %f884, [%r34+640];
add.ftz.f32 %f885, %f884, %f881;
ld.shared.f32 %f886, [%r32+1152];
add.ftz.f32 %f887, %f886, %f883;
ld.shared.f32 %f888, [%r34+1152];
add.ftz.f32 %f889, %f888, %f885;
mov.b32 %r2034, %f887;
shfl.sync.bfly.b32 %r2038|%p280, %r2034, %r2009, %r2008, %r2010;
mov.b32 %f890, %r2038;
add.ftz.f32 %f891, %f887, %f890;
mov.b32 %r2039, %f891;
shfl.sync.bfly.b32 %r2041|%p281, %r2039, %r2013, %r2008, %r2010;
mov.b32 %f892, %r2041;
add.ftz.f32 %f893, %f891, %f892;
mov.b32 %r2042, %f893;
shfl.sync.bfly.b32 %r2044|%p282, %r2042, %r2016, %r2008, %r2010;
mov.b32 %f894, %r2044;
add.ftz.f32 %f895, %f893, %f894;
mov.b32 %r2045, %f895;
shfl.sync.bfly.b32 %r2047|%p283, %r2045, %r2019, %r2008, %r2010;
mov.b32 %f896, %r2047;
add.ftz.f32 %f897, %f895, %f896;
mov.b32 %r2048, %f897;
shfl.sync.bfly.b32 %r2050|%p284, %r2048, %r2022, %r2008, %r2010;
mov.b32 %f898, %r2050;
add.ftz.f32 %f74, %f897, %f898;
mov.b32 %r2051, %f889;
shfl.sync.bfly.b32 %r2052|%p285, %r2051, %r2009, %r2008, %r2010;
mov.b32 %f899, %r2052;
add.ftz.f32 %f900, %f889, %f899;
mov.b32 %r2053, %f900;
shfl.sync.bfly.b32 %r2054|%p286, %r2053, %r2013, %r2008, %r2010;
mov.b32 %f901, %r2054;
add.ftz.f32 %f902, %f900, %f901;
mov.b32 %r2055, %f902;
shfl.sync.bfly.b32 %r2056|%p287, %r2055, %r2016, %r2008, %r2010;
mov.b32 %f903, %r2056;
add.ftz.f32 %f904, %f902, %f903;
mov.b32 %r2057, %f904;
shfl.sync.bfly.b32 %r2058|%p288, %r2057, %r2019, %r2008, %r2010;
mov.b32 %f905, %r2058;
add.ftz.f32 %f906, %f904, %f905;
mov.b32 %r2059, %f906;
shfl.sync.bfly.b32 %r2060|%p289, %r2059, %r2022, %r2008, %r2010;
mov.b32 %f907, %r2060;
add.ftz.f32 %f75, %f906, %f907;
add.s32 %r2061, %r4, 1;
setp.ge.u32 %p290, %r2061, %r50;
setp.ne.s32 %p291, %r2, 1;
or.pred %p292, %p291, %p290;
@%p292 bra $L__BB239_74;

add.ftz.f32 %f908, %f5, %f74;
add.ftz.f32 %f909, %f4, %f75;
mul.ftz.f32 %f910, %f908, %f909;
st.global.f32 [%rd35+4], %f910;

$L__BB239_74:
ld.shared.f32 %f911, [%r32+256];
add.ftz.f32 %f912, %f911, %f28;
ld.shared.f32 %f913, [%r34+256];
add.ftz.f32 %f914, %f913, %f32;
ld.shared.f32 %f915, [%r32+768];
add.ftz.f32 %f916, %f915, %f912;
ld.shared.f32 %f917, [%r34+768];
add.ftz.f32 %f918, %f917, %f914;
ld.shared.f32 %f919, [%r32+1280];
add.ftz.f32 %f920, %f919, %f916;
ld.shared.f32 %f921, [%r34+1280];
add.ftz.f32 %f922, %f921, %f918;
mov.b32 %r2062, %f920;
mov.u32 %r2063, 31;
mov.u32 %r2064, 16;
mov.u32 %r2065, -1;
shfl.sync.bfly.b32 %r2066|%p293, %r2062, %r2064, %r2063, %r2065;
mov.b32 %f923, %r2066;
add.ftz.f32 %f924, %f920, %f923;
mov.b32 %r2067, %f924;
mov.u32 %r2068, 8;
shfl.sync.bfly.b32 %r2069|%p294, %r2067, %r2068, %r2063, %r2065;
mov.b32 %f925, %r2069;
add.ftz.f32 %f926, %f924, %f925;
mov.b32 %r2070, %f926;
mov.u32 %r2071, 4;
shfl.sync.bfly.b32 %r2072|%p295, %r2070, %r2071, %r2063, %r2065;
mov.b32 %f927, %r2072;
add.ftz.f32 %f928, %f926, %f927;
mov.b32 %r2073, %f928;
mov.u32 %r2074, 2;
shfl.sync.bfly.b32 %r2075|%p296, %r2073, %r2074, %r2063, %r2065;
mov.b32 %f929, %r2075;
add.ftz.f32 %f930, %f928, %f929;
mov.b32 %r2076, %f930;
mov.u32 %r2077, 1;
shfl.sync.bfly.b32 %r2078|%p297, %r2076, %r2077, %r2063, %r2065;
mov.b32 %f931, %r2078;
add.ftz.f32 %f76, %f930, %f931;
mov.b32 %r2079, %f922;
shfl.sync.bfly.b32 %r2080|%p298, %r2079, %r2064, %r2063, %r2065;
mov.b32 %f932, %r2080;
add.ftz.f32 %f933, %f922, %f932;
mov.b32 %r2081, %f933;
shfl.sync.bfly.b32 %r2082|%p299, %r2081, %r2068, %r2063, %r2065;
mov.b32 %f934, %r2082;
add.ftz.f32 %f935, %f933, %f934;
mov.b32 %r2083, %f935;
shfl.sync.bfly.b32 %r2084|%p300, %r2083, %r2071, %r2063, %r2065;
mov.b32 %f936, %r2084;
add.ftz.f32 %f937, %f935, %f936;
mov.b32 %r2085, %f937;
shfl.sync.bfly.b32 %r2086|%p301, %r2085, %r2074, %r2063, %r2065;
mov.b32 %f938, %r2086;
add.ftz.f32 %f939, %f937, %f938;
mov.b32 %r2087, %f939;
shfl.sync.bfly.b32 %r2088|%p302, %r2087, %r2077, %r2063, %r2065;
mov.b32 %f940, %r2088;
add.ftz.f32 %f77, %f939, %f940;
add.s32 %r2089, %r4, 2;
setp.ge.u32 %p303, %r2089, %r50;
setp.ne.s32 %p304, %r2, 2;
or.pred %p305, %p304, %p303;
@%p305 bra $L__BB239_76;

add.ftz.f32 %f941, %f5, %f76;
add.ftz.f32 %f942, %f4, %f77;
mul.ftz.f32 %f943, %f941, %f942;
st.global.f32 [%rd35+8], %f943;

$L__BB239_76:
ld.shared.f32 %f944, [%r32+384];
add.ftz.f32 %f945, %f944, %f27;
ld.shared.f32 %f946, [%r34+384];
add.ftz.f32 %f947, %f946, %f31;
ld.shared.f32 %f948, [%r32+896];
add.ftz.f32 %f949, %f948, %f945;
ld.shared.f32 %f950, [%r34+896];
add.ftz.f32 %f951, %f950, %f947;
ld.shared.f32 %f952, [%r32+1408];
add.ftz.f32 %f953, %f952, %f949;
ld.shared.f32 %f954, [%r34+1408];
add.ftz.f32 %f955, %f954, %f951;
mov.b32 %r2090, %f953;
shfl.sync.bfly.b32 %r2094|%p306, %r2090, %r2064, %r2063, %r2065;
mov.b32 %f956, %r2094;
add.ftz.f32 %f957, %f953, %f956;
mov.b32 %r2095, %f957;
shfl.sync.bfly.b32 %r2097|%p307, %r2095, %r2068, %r2063, %r2065;
mov.b32 %f958, %r2097;
add.ftz.f32 %f959, %f957, %f958;
mov.b32 %r2098, %f959;
shfl.sync.bfly.b32 %r2100|%p308, %r2098, %r2071, %r2063, %r2065;
mov.b32 %f960, %r2100;
add.ftz.f32 %f961, %f959, %f960;
mov.b32 %r2101, %f961;
shfl.sync.bfly.b32 %r2103|%p309, %r2101, %r2074, %r2063, %r2065;
mov.b32 %f962, %r2103;
add.ftz.f32 %f963, %f961, %f962;
mov.b32 %r2104, %f963;
shfl.sync.bfly.b32 %r2106|%p310, %r2104, %r2077, %r2063, %r2065;
mov.b32 %f964, %r2106;
add.ftz.f32 %f78, %f963, %f964;
mov.b32 %r2107, %f955;
shfl.sync.bfly.b32 %r2108|%p311, %r2107, %r2064, %r2063, %r2065;
mov.b32 %f965, %r2108;
add.ftz.f32 %f966, %f955, %f965;
mov.b32 %r2109, %f966;
shfl.sync.bfly.b32 %r2110|%p312, %r2109, %r2068, %r2063, %r2065;
mov.b32 %f967, %r2110;
add.ftz.f32 %f968, %f966, %f967;
mov.b32 %r2111, %f968;
shfl.sync.bfly.b32 %r2112|%p313, %r2111, %r2071, %r2063, %r2065;
mov.b32 %f969, %r2112;
add.ftz.f32 %f970, %f968, %f969;
mov.b32 %r2113, %f970;
shfl.sync.bfly.b32 %r2114|%p314, %r2113, %r2074, %r2063, %r2065;
mov.b32 %f971, %r2114;
add.ftz.f32 %f972, %f970, %f971;
mov.b32 %r2115, %f972;
shfl.sync.bfly.b32 %r2116|%p315, %r2115, %r2077, %r2063, %r2065;
mov.b32 %f973, %r2116;
add.ftz.f32 %f79, %f972, %f973;
add.s32 %r2117, %r4, 3;
setp.ge.u32 %p316, %r2117, %r50;
setp.ne.s32 %p317, %r2, 3;
or.pred %p318, %p317, %p316;
@%p318 bra $L__BB239_78;

add.ftz.f32 %f974, %f5, %f78;
add.ftz.f32 %f975, %f4, %f79;
mul.ftz.f32 %f976, %f974, %f975;
st.global.f32 [%rd35+12], %f976;
bra.uni $L__BB239_78;

$L__BB239_62:
ld.shared.f32 %f681, [%r32];
add.ftz.f32 %f682, %f681, %f30;
ld.shared.f32 %f683, [%r34];
add.ftz.f32 %f684, %f683, %f34;
ld.shared.f32 %f685, [%r32+512];
add.ftz.f32 %f686, %f685, %f682;
ld.shared.f32 %f687, [%r34+512];
add.ftz.f32 %f688, %f687, %f684;
ld.shared.f32 %f689, [%r32+1024];
add.ftz.f32 %f690, %f689, %f686;
ld.shared.f32 %f691, [%r34+1024];
add.ftz.f32 %f692, %f691, %f688;
mov.b32 %r1896, %f690;
mov.u32 %r1897, 31;
mov.u32 %r1898, 16;
mov.u32 %r1899, -1;
shfl.sync.bfly.b32 %r1900|%p215, %r1896, %r1898, %r1897, %r1899;
mov.b32 %f693, %r1900;
add.ftz.f32 %f694, %f690, %f693;
mov.b32 %r1901, %f694;
mov.u32 %r1902, 8;
shfl.sync.bfly.b32 %r1903|%p216, %r1901, %r1902, %r1897, %r1899;
mov.b32 %f695, %r1903;
add.ftz.f32 %f696, %f694, %f695;
mov.b32 %r1904, %f696;
mov.u32 %r1905, 4;
shfl.sync.bfly.b32 %r1906|%p217, %r1904, %r1905, %r1897, %r1899;
mov.b32 %f697, %r1906;
add.ftz.f32 %f698, %f696, %f697;
mov.b32 %r1907, %f698;
mov.u32 %r1908, 2;
shfl.sync.bfly.b32 %r1909|%p218, %r1907, %r1908, %r1897, %r1899;
mov.b32 %f699, %r1909;
add.ftz.f32 %f700, %f698, %f699;
mov.b32 %r1910, %f700;
mov.u32 %r1911, 1;
shfl.sync.bfly.b32 %r1912|%p219, %r1910, %r1911, %r1897, %r1899;
mov.b32 %f701, %r1912;
add.ftz.f32 %f64, %f700, %f701;
mov.b32 %r1913, %f692;
shfl.sync.bfly.b32 %r1914|%p220, %r1913, %r1898, %r1897, %r1899;
mov.b32 %f702, %r1914;
add.ftz.f32 %f703, %f692, %f702;
mov.b32 %r1915, %f703;
shfl.sync.bfly.b32 %r1916|%p221, %r1915, %r1902, %r1897, %r1899;
mov.b32 %f704, %r1916;
add.ftz.f32 %f705, %f703, %f704;
mov.b32 %r1917, %f705;
shfl.sync.bfly.b32 %r1918|%p222, %r1917, %r1905, %r1897, %r1899;
mov.b32 %f706, %r1918;
add.ftz.f32 %f707, %f705, %f706;
mov.b32 %r1919, %f707;
shfl.sync.bfly.b32 %r1920|%p223, %r1919, %r1908, %r1897, %r1899;
mov.b32 %f708, %r1920;
add.ftz.f32 %f709, %f707, %f708;
mov.b32 %r1921, %f709;
shfl.sync.bfly.b32 %r1922|%p224, %r1921, %r1911, %r1897, %r1899;
mov.b32 %f710, %r1922;
add.ftz.f32 %f65, %f709, %f710;
setp.ge.u32 %p225, %r4, %r50;
setp.ne.s32 %p226, %r2, 0;
or.pred %p227, %p226, %p225;
@%p227 bra $L__BB239_64;

add.ftz.f32 %f711, %f5, %f64;
add.ftz.f32 %f712, %f4, %f65;
mul.ftz.f32 %f713, %f712, 0f3F000000;
mul.ftz.f32 %f714, %f712, 0f3F4C422A;
mul.ftz.f32 %f715, %f712, 0f3D372713;
fma.rn.ftz.f32 %f716, %f712, %f715, 0f3F800000;
mul.ftz.f32 %f717, %f714, %f716;
tanh.approx.f32 %f718, %f717;
add.ftz.f32 %f719, %f718, 0f3F800000;
mul.ftz.f32 %f720, %f713, %f719;
mul.ftz.f32 %f721, %f711, %f720;
st.global.f32 [%rd35], %f721;

$L__BB239_64:
ld.shared.f32 %f722, [%r32+128];
add.ftz.f32 %f723, %f722, %f29;
ld.shared.f32 %f724, [%r34+128];
add.ftz.f32 %f725, %f724, %f33;
ld.shared.f32 %f726, [%r32+640];
add.ftz.f32 %f727, %f726, %f723;
ld.shared.f32 %f728, [%r34+640];
add.ftz.f32 %f729, %f728, %f725;
ld.shared.f32 %f730, [%r32+1152];
add.ftz.f32 %f731, %f730, %f727;
ld.shared.f32 %f732, [%r34+1152];
add.ftz.f32 %f733, %f732, %f729;
mov.b32 %r1923, %f731;
shfl.sync.bfly.b32 %r1927|%p228, %r1923, %r1898, %r1897, %r1899;
mov.b32 %f734, %r1927;
add.ftz.f32 %f735, %f731, %f734;
mov.b32 %r1928, %f735;
shfl.sync.bfly.b32 %r1930|%p229, %r1928, %r1902, %r1897, %r1899;
mov.b32 %f736, %r1930;
add.ftz.f32 %f737, %f735, %f736;
mov.b32 %r1931, %f737;
shfl.sync.bfly.b32 %r1933|%p230, %r1931, %r1905, %r1897, %r1899;
mov.b32 %f738, %r1933;
add.ftz.f32 %f739, %f737, %f738;
mov.b32 %r1934, %f739;
shfl.sync.bfly.b32 %r1936|%p231, %r1934, %r1908, %r1897, %r1899;
mov.b32 %f740, %r1936;
add.ftz.f32 %f741, %f739, %f740;
mov.b32 %r1937, %f741;
shfl.sync.bfly.b32 %r1939|%p232, %r1937, %r1911, %r1897, %r1899;
mov.b32 %f742, %r1939;
add.ftz.f32 %f66, %f741, %f742;
mov.b32 %r1940, %f733;
shfl.sync.bfly.b32 %r1941|%p233, %r1940, %r1898, %r1897, %r1899;
mov.b32 %f743, %r1941;
add.ftz.f32 %f744, %f733, %f743;
mov.b32 %r1942, %f744;
shfl.sync.bfly.b32 %r1943|%p234, %r1942, %r1902, %r1897, %r1899;
mov.b32 %f745, %r1943;
add.ftz.f32 %f746, %f744, %f745;
mov.b32 %r1944, %f746;
shfl.sync.bfly.b32 %r1945|%p235, %r1944, %r1905, %r1897, %r1899;
mov.b32 %f747, %r1945;
add.ftz.f32 %f748, %f746, %f747;
mov.b32 %r1946, %f748;
shfl.sync.bfly.b32 %r1947|%p236, %r1946, %r1908, %r1897, %r1899;
mov.b32 %f749, %r1947;
add.ftz.f32 %f750, %f748, %f749;
mov.b32 %r1948, %f750;
shfl.sync.bfly.b32 %r1949|%p237, %r1948, %r1911, %r1897, %r1899;
mov.b32 %f751, %r1949;
add.ftz.f32 %f67, %f750, %f751;
add.s32 %r1950, %r4, 1;
setp.ge.u32 %p238, %r1950, %r50;
setp.ne.s32 %p239, %r2, 1;
or.pred %p240, %p239, %p238;
@%p240 bra $L__BB239_66;

add.ftz.f32 %f752, %f5, %f66;
add.ftz.f32 %f753, %f4, %f67;
mul.ftz.f32 %f754, %f753, 0f3F000000;
mul.ftz.f32 %f755, %f753, 0f3F4C422A;
mul.ftz.f32 %f756, %f753, 0f3D372713;
fma.rn.ftz.f32 %f757, %f753, %f756, 0f3F800000;
mul.ftz.f32 %f758, %f755, %f757;
tanh.approx.f32 %f759, %f758;
add.ftz.f32 %f760, %f759, 0f3F800000;
mul.ftz.f32 %f761, %f754, %f760;
mul.ftz.f32 %f762, %f752, %f761;
st.global.f32 [%rd35+4], %f762;

$L__BB239_66:
ld.shared.f32 %f763, [%r32+256];
add.ftz.f32 %f764, %f763, %f28;
ld.shared.f32 %f765, [%r34+256];
add.ftz.f32 %f766, %f765, %f32;
ld.shared.f32 %f767, [%r32+768];
add.ftz.f32 %f768, %f767, %f764;
ld.shared.f32 %f769, [%r34+768];
add.ftz.f32 %f770, %f769, %f766;
ld.shared.f32 %f771, [%r32+1280];
add.ftz.f32 %f772, %f771, %f768;
ld.shared.f32 %f773, [%r34+1280];
add.ftz.f32 %f774, %f773, %f770;
mov.b32 %r1951, %f772;
mov.u32 %r1952, 31;
mov.u32 %r1953, 16;
mov.u32 %r1954, -1;
shfl.sync.bfly.b32 %r1955|%p241, %r1951, %r1953, %r1952, %r1954;
mov.b32 %f775, %r1955;
add.ftz.f32 %f776, %f772, %f775;
mov.b32 %r1956, %f776;
mov.u32 %r1957, 8;
shfl.sync.bfly.b32 %r1958|%p242, %r1956, %r1957, %r1952, %r1954;
mov.b32 %f777, %r1958;
add.ftz.f32 %f778, %f776, %f777;
mov.b32 %r1959, %f778;
mov.u32 %r1960, 4;
shfl.sync.bfly.b32 %r1961|%p243, %r1959, %r1960, %r1952, %r1954;
mov.b32 %f779, %r1961;
add.ftz.f32 %f780, %f778, %f779;
mov.b32 %r1962, %f780;
mov.u32 %r1963, 2;
shfl.sync.bfly.b32 %r1964|%p244, %r1962, %r1963, %r1952, %r1954;
mov.b32 %f781, %r1964;
add.ftz.f32 %f782, %f780, %f781;
mov.b32 %r1965, %f782;
mov.u32 %r1966, 1;
shfl.sync.bfly.b32 %r1967|%p245, %r1965, %r1966, %r1952, %r1954;
mov.b32 %f783, %r1967;
add.ftz.f32 %f68, %f782, %f783;
mov.b32 %r1968, %f774;
shfl.sync.bfly.b32 %r1969|%p246, %r1968, %r1953, %r1952, %r1954;
mov.b32 %f784, %r1969;
add.ftz.f32 %f785, %f774, %f784;
mov.b32 %r1970, %f785;
shfl.sync.bfly.b32 %r1971|%p247, %r1970, %r1957, %r1952, %r1954;
mov.b32 %f786, %r1971;
add.ftz.f32 %f787, %f785, %f786;
mov.b32 %r1972, %f787;
shfl.sync.bfly.b32 %r1973|%p248, %r1972, %r1960, %r1952, %r1954;
mov.b32 %f788, %r1973;
add.ftz.f32 %f789, %f787, %f788;
mov.b32 %r1974, %f789;
shfl.sync.bfly.b32 %r1975|%p249, %r1974, %r1963, %r1952, %r1954;
mov.b32 %f790, %r1975;
add.ftz.f32 %f791, %f789, %f790;
mov.b32 %r1976, %f791;
shfl.sync.bfly.b32 %r1977|%p250, %r1976, %r1966, %r1952, %r1954;
mov.b32 %f792, %r1977;
add.ftz.f32 %f69, %f791, %f792;
add.s32 %r1978, %r4, 2;
setp.ge.u32 %p251, %r1978, %r50;
setp.ne.s32 %p252, %r2, 2;
or.pred %p253, %p252, %p251;
@%p253 bra $L__BB239_68;

add.ftz.f32 %f793, %f5, %f68;
add.ftz.f32 %f794, %f4, %f69;
mul.ftz.f32 %f795, %f794, 0f3F000000;
mul.ftz.f32 %f796, %f794, 0f3F4C422A;
mul.ftz.f32 %f797, %f794, 0f3D372713;
fma.rn.ftz.f32 %f798, %f794, %f797, 0f3F800000;
mul.ftz.f32 %f799, %f796, %f798;
tanh.approx.f32 %f800, %f799;
add.ftz.f32 %f801, %f800, 0f3F800000;
mul.ftz.f32 %f802, %f795, %f801;
mul.ftz.f32 %f803, %f793, %f802;
st.global.f32 [%rd35+8], %f803;

$L__BB239_68:
ld.shared.f32 %f804, [%r32+384];
add.ftz.f32 %f805, %f804, %f27;
ld.shared.f32 %f806, [%r34+384];
add.ftz.f32 %f807, %f806, %f31;
ld.shared.f32 %f808, [%r32+896];
add.ftz.f32 %f809, %f808, %f805;
ld.shared.f32 %f810, [%r34+896];
add.ftz.f32 %f811, %f810, %f807;
ld.shared.f32 %f812, [%r32+1408];
add.ftz.f32 %f813, %f812, %f809;
ld.shared.f32 %f814, [%r34+1408];
add.ftz.f32 %f815, %f814, %f811;
mov.b32 %r1979, %f813;
shfl.sync.bfly.b32 %r1983|%p254, %r1979, %r1953, %r1952, %r1954;
mov.b32 %f816, %r1983;
add.ftz.f32 %f817, %f813, %f816;
mov.b32 %r1984, %f817;
shfl.sync.bfly.b32 %r1986|%p255, %r1984, %r1957, %r1952, %r1954;
mov.b32 %f818, %r1986;
add.ftz.f32 %f819, %f817, %f818;
mov.b32 %r1987, %f819;
shfl.sync.bfly.b32 %r1989|%p256, %r1987, %r1960, %r1952, %r1954;
mov.b32 %f820, %r1989;
add.ftz.f32 %f821, %f819, %f820;
mov.b32 %r1990, %f821;
shfl.sync.bfly.b32 %r1992|%p257, %r1990, %r1963, %r1952, %r1954;
mov.b32 %f822, %r1992;
add.ftz.f32 %f823, %f821, %f822;
mov.b32 %r1993, %f823;
shfl.sync.bfly.b32 %r1995|%p258, %r1993, %r1966, %r1952, %r1954;
mov.b32 %f824, %r1995;
add.ftz.f32 %f70, %f823, %f824;
mov.b32 %r1996, %f815;
shfl.sync.bfly.b32 %r1997|%p259, %r1996, %r1953, %r1952, %r1954;
mov.b32 %f825, %r1997;
add.ftz.f32 %f826, %f815, %f825;
mov.b32 %r1998, %f826;
shfl.sync.bfly.b32 %r1999|%p260, %r1998, %r1957, %r1952, %r1954;
mov.b32 %f827, %r1999;
add.ftz.f32 %f828, %f826, %f827;
mov.b32 %r2000, %f828;
shfl.sync.bfly.b32 %r2001|%p261, %r2000, %r1960, %r1952, %r1954;
mov.b32 %f829, %r2001;
add.ftz.f32 %f830, %f828, %f829;
mov.b32 %r2002, %f830;
shfl.sync.bfly.b32 %r2003|%p262, %r2002, %r1963, %r1952, %r1954;
mov.b32 %f831, %r2003;
add.ftz.f32 %f832, %f830, %f831;
mov.b32 %r2004, %f832;
shfl.sync.bfly.b32 %r2005|%p263, %r2004, %r1966, %r1952, %r1954;
mov.b32 %f833, %r2005;
add.ftz.f32 %f71, %f832, %f833;
add.s32 %r2006, %r4, 3;
setp.ge.u32 %p264, %r2006, %r50;
setp.ne.s32 %p265, %r2, 3;
or.pred %p266, %p265, %p264;
@%p266 bra $L__BB239_78;

add.ftz.f32 %f834, %f5, %f70;
add.ftz.f32 %f835, %f4, %f71;
mul.ftz.f32 %f836, %f835, 0f3F000000;
mul.ftz.f32 %f837, %f835, 0f3F4C422A;
mul.ftz.f32 %f838, %f835, 0f3D372713;
fma.rn.ftz.f32 %f839, %f835, %f838, 0f3F800000;
mul.ftz.f32 %f840, %f837, %f839;
tanh.approx.f32 %f841, %f840;
add.ftz.f32 %f842, %f841, 0f3F800000;
mul.ftz.f32 %f843, %f836, %f842;
mul.ftz.f32 %f844, %f834, %f843;
st.global.f32 [%rd35+12], %f844;
bra.uni $L__BB239_78;

$L__BB239_54:
ld.shared.f32 %f357, [%r32];
add.ftz.f32 %f358, %f357, %f30;
ld.shared.f32 %f359, [%r34];
add.ftz.f32 %f360, %f359, %f34;
ld.shared.f32 %f361, [%r32+512];
add.ftz.f32 %f362, %f361, %f358;
ld.shared.f32 %f363, [%r34+512];
add.ftz.f32 %f364, %f363, %f360;
ld.shared.f32 %f365, [%r32+1024];
add.ftz.f32 %f366, %f365, %f362;
ld.shared.f32 %f367, [%r34+1024];
add.ftz.f32 %f368, %f367, %f364;
mov.b32 %r1674, %f366;
mov.u32 %r1675, 31;
mov.u32 %r1676, 16;
mov.u32 %r1677, -1;
shfl.sync.bfly.b32 %r1678|%p111, %r1674, %r1676, %r1675, %r1677;
mov.b32 %f369, %r1678;
add.ftz.f32 %f370, %f366, %f369;
mov.b32 %r1679, %f370;
mov.u32 %r1680, 8;
shfl.sync.bfly.b32 %r1681|%p112, %r1679, %r1680, %r1675, %r1677;
mov.b32 %f371, %r1681;
add.ftz.f32 %f372, %f370, %f371;
mov.b32 %r1682, %f372;
mov.u32 %r1683, 4;
shfl.sync.bfly.b32 %r1684|%p113, %r1682, %r1683, %r1675, %r1677;
mov.b32 %f373, %r1684;
add.ftz.f32 %f374, %f372, %f373;
mov.b32 %r1685, %f374;
mov.u32 %r1686, 2;
shfl.sync.bfly.b32 %r1687|%p114, %r1685, %r1686, %r1675, %r1677;
mov.b32 %f375, %r1687;
add.ftz.f32 %f376, %f374, %f375;
mov.b32 %r1688, %f376;
mov.u32 %r1689, 1;
shfl.sync.bfly.b32 %r1690|%p115, %r1688, %r1689, %r1675, %r1677;
mov.b32 %f377, %r1690;
add.ftz.f32 %f48, %f376, %f377;
mov.b32 %r1691, %f368;
shfl.sync.bfly.b32 %r1692|%p116, %r1691, %r1676, %r1675, %r1677;
mov.b32 %f378, %r1692;
add.ftz.f32 %f379, %f368, %f378;
mov.b32 %r1693, %f379;
shfl.sync.bfly.b32 %r1694|%p117, %r1693, %r1680, %r1675, %r1677;
mov.b32 %f380, %r1694;
add.ftz.f32 %f381, %f379, %f380;
mov.b32 %r1695, %f381;
shfl.sync.bfly.b32 %r1696|%p118, %r1695, %r1683, %r1675, %r1677;
mov.b32 %f382, %r1696;
add.ftz.f32 %f383, %f381, %f382;
mov.b32 %r1697, %f383;
shfl.sync.bfly.b32 %r1698|%p119, %r1697, %r1686, %r1675, %r1677;
mov.b32 %f384, %r1698;
add.ftz.f32 %f385, %f383, %f384;
mov.b32 %r1699, %f385;
shfl.sync.bfly.b32 %r1700|%p120, %r1699, %r1689, %r1675, %r1677;
mov.b32 %f386, %r1700;
add.ftz.f32 %f49, %f385, %f386;
setp.ge.u32 %p121, %r4, %r50;
setp.ne.s32 %p122, %r2, 0;
or.pred %p123, %p122, %p121;
@%p123 bra $L__BB239_56;

add.ftz.f32 %f387, %f5, %f48;
add.ftz.f32 %f388, %f4, %f49;
mov.f32 %f389, 0f40E00000;
min.ftz.f32 %f390, %f388, %f389;
min.ftz.f32 %f391, %f387, %f389;
mov.f32 %f392, 0fC0E00000;
max.ftz.f32 %f393, %f391, %f392;
mul.ftz.f32 %f394, %f390, 0f3FD9DB23;
mul.ftz.f32 %f395, %f394, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f396, %f395;
add.ftz.f32 %f397, %f396, 0f3F800000;
div.approx.ftz.f32 %f398, %f390, %f397;
add.ftz.f32 %f399, %f393, 0f3F800000;
mul.ftz.f32 %f400, %f399, %f398;
st.global.f32 [%rd35], %f400;

$L__BB239_56:
ld.shared.f32 %f401, [%r32+128];
add.ftz.f32 %f402, %f401, %f29;
ld.shared.f32 %f403, [%r34+128];
add.ftz.f32 %f404, %f403, %f33;
ld.shared.f32 %f405, [%r32+640];
add.ftz.f32 %f406, %f405, %f402;
ld.shared.f32 %f407, [%r34+640];
add.ftz.f32 %f408, %f407, %f404;
ld.shared.f32 %f409, [%r32+1152];
add.ftz.f32 %f410, %f409, %f406;
ld.shared.f32 %f411, [%r34+1152];
add.ftz.f32 %f412, %f411, %f408;
mov.b32 %r1701, %f410;
shfl.sync.bfly.b32 %r1705|%p124, %r1701, %r1676, %r1675, %r1677;
mov.b32 %f413, %r1705;
add.ftz.f32 %f414, %f410, %f413;
mov.b32 %r1706, %f414;
shfl.sync.bfly.b32 %r1708|%p125, %r1706, %r1680, %r1675, %r1677;
mov.b32 %f415, %r1708;
add.ftz.f32 %f416, %f414, %f415;
mov.b32 %r1709, %f416;
shfl.sync.bfly.b32 %r1711|%p126, %r1709, %r1683, %r1675, %r1677;
mov.b32 %f417, %r1711;
add.ftz.f32 %f418, %f416, %f417;
mov.b32 %r1712, %f418;
shfl.sync.bfly.b32 %r1714|%p127, %r1712, %r1686, %r1675, %r1677;
mov.b32 %f419, %r1714;
add.ftz.f32 %f420, %f418, %f419;
mov.b32 %r1715, %f420;
shfl.sync.bfly.b32 %r1717|%p128, %r1715, %r1689, %r1675, %r1677;
mov.b32 %f421, %r1717;
add.ftz.f32 %f50, %f420, %f421;
mov.b32 %r1718, %f412;
shfl.sync.bfly.b32 %r1719|%p129, %r1718, %r1676, %r1675, %r1677;
mov.b32 %f422, %r1719;
add.ftz.f32 %f423, %f412, %f422;
mov.b32 %r1720, %f423;
shfl.sync.bfly.b32 %r1721|%p130, %r1720, %r1680, %r1675, %r1677;
mov.b32 %f424, %r1721;
add.ftz.f32 %f425, %f423, %f424;
mov.b32 %r1722, %f425;
shfl.sync.bfly.b32 %r1723|%p131, %r1722, %r1683, %r1675, %r1677;
mov.b32 %f426, %r1723;
add.ftz.f32 %f427, %f425, %f426;
mov.b32 %r1724, %f427;
shfl.sync.bfly.b32 %r1725|%p132, %r1724, %r1686, %r1675, %r1677;
mov.b32 %f428, %r1725;
add.ftz.f32 %f429, %f427, %f428;
mov.b32 %r1726, %f429;
shfl.sync.bfly.b32 %r1727|%p133, %r1726, %r1689, %r1675, %r1677;
mov.b32 %f430, %r1727;
add.ftz.f32 %f51, %f429, %f430;
add.s32 %r1728, %r4, 1;
setp.ge.u32 %p134, %r1728, %r50;
setp.ne.s32 %p135, %r2, 1;
or.pred %p136, %p135, %p134;
@%p136 bra $L__BB239_58;

add.ftz.f32 %f431, %f5, %f50;
add.ftz.f32 %f432, %f4, %f51;
mov.f32 %f433, 0f40E00000;
min.ftz.f32 %f434, %f432, %f433;
min.ftz.f32 %f435, %f431, %f433;
mov.f32 %f436, 0fC0E00000;
max.ftz.f32 %f437, %f435, %f436;
mul.ftz.f32 %f438, %f434, 0f3FD9DB23;
mul.ftz.f32 %f439, %f438, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f440, %f439;
add.ftz.f32 %f441, %f440, 0f3F800000;
div.approx.ftz.f32 %f442, %f434, %f441;
add.ftz.f32 %f443, %f437, 0f3F800000;
mul.ftz.f32 %f444, %f443, %f442;
st.global.f32 [%rd35+4], %f444;

$L__BB239_58:
ld.shared.f32 %f445, [%r32+256];
add.ftz.f32 %f446, %f445, %f28;
ld.shared.f32 %f447, [%r34+256];
add.ftz.f32 %f448, %f447, %f32;
ld.shared.f32 %f449, [%r32+768];
add.ftz.f32 %f450, %f449, %f446;
ld.shared.f32 %f451, [%r34+768];
add.ftz.f32 %f452, %f451, %f448;
ld.shared.f32 %f453, [%r32+1280];
add.ftz.f32 %f454, %f453, %f450;
ld.shared.f32 %f455, [%r34+1280];
add.ftz.f32 %f456, %f455, %f452;
mov.b32 %r1729, %f454;
mov.u32 %r1730, 31;
mov.u32 %r1731, 16;
mov.u32 %r1732, -1;
shfl.sync.bfly.b32 %r1733|%p137, %r1729, %r1731, %r1730, %r1732;
mov.b32 %f457, %r1733;
add.ftz.f32 %f458, %f454, %f457;
mov.b32 %r1734, %f458;
mov.u32 %r1735, 8;
shfl.sync.bfly.b32 %r1736|%p138, %r1734, %r1735, %r1730, %r1732;
mov.b32 %f459, %r1736;
add.ftz.f32 %f460, %f458, %f459;
mov.b32 %r1737, %f460;
mov.u32 %r1738, 4;
shfl.sync.bfly.b32 %r1739|%p139, %r1737, %r1738, %r1730, %r1732;
mov.b32 %f461, %r1739;
add.ftz.f32 %f462, %f460, %f461;
mov.b32 %r1740, %f462;
mov.u32 %r1741, 2;
shfl.sync.bfly.b32 %r1742|%p140, %r1740, %r1741, %r1730, %r1732;
mov.b32 %f463, %r1742;
add.ftz.f32 %f464, %f462, %f463;
mov.b32 %r1743, %f464;
mov.u32 %r1744, 1;
shfl.sync.bfly.b32 %r1745|%p141, %r1743, %r1744, %r1730, %r1732;
mov.b32 %f465, %r1745;
add.ftz.f32 %f52, %f464, %f465;
mov.b32 %r1746, %f456;
shfl.sync.bfly.b32 %r1747|%p142, %r1746, %r1731, %r1730, %r1732;
mov.b32 %f466, %r1747;
add.ftz.f32 %f467, %f456, %f466;
mov.b32 %r1748, %f467;
shfl.sync.bfly.b32 %r1749|%p143, %r1748, %r1735, %r1730, %r1732;
mov.b32 %f468, %r1749;
add.ftz.f32 %f469, %f467, %f468;
mov.b32 %r1750, %f469;
shfl.sync.bfly.b32 %r1751|%p144, %r1750, %r1738, %r1730, %r1732;
mov.b32 %f470, %r1751;
add.ftz.f32 %f471, %f469, %f470;
mov.b32 %r1752, %f471;
shfl.sync.bfly.b32 %r1753|%p145, %r1752, %r1741, %r1730, %r1732;
mov.b32 %f472, %r1753;
add.ftz.f32 %f473, %f471, %f472;
mov.b32 %r1754, %f473;
shfl.sync.bfly.b32 %r1755|%p146, %r1754, %r1744, %r1730, %r1732;
mov.b32 %f474, %r1755;
add.ftz.f32 %f53, %f473, %f474;
add.s32 %r1756, %r4, 2;
setp.ge.u32 %p147, %r1756, %r50;
setp.ne.s32 %p148, %r2, 2;
or.pred %p149, %p148, %p147;
@%p149 bra $L__BB239_60;

add.ftz.f32 %f475, %f5, %f52;
add.ftz.f32 %f476, %f4, %f53;
mov.f32 %f477, 0f40E00000;
min.ftz.f32 %f478, %f476, %f477;
min.ftz.f32 %f479, %f475, %f477;
mov.f32 %f480, 0fC0E00000;
max.ftz.f32 %f481, %f479, %f480;
mul.ftz.f32 %f482, %f478, 0f3FD9DB23;
mul.ftz.f32 %f483, %f482, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f484, %f483;
add.ftz.f32 %f485, %f484, 0f3F800000;
div.approx.ftz.f32 %f486, %f478, %f485;
add.ftz.f32 %f487, %f481, 0f3F800000;
mul.ftz.f32 %f488, %f487, %f486;
st.global.f32 [%rd35+8], %f488;

$L__BB239_60:
ld.shared.f32 %f489, [%r32+384];
add.ftz.f32 %f490, %f489, %f27;
ld.shared.f32 %f491, [%r34+384];
add.ftz.f32 %f492, %f491, %f31;
ld.shared.f32 %f493, [%r32+896];
add.ftz.f32 %f494, %f493, %f490;
ld.shared.f32 %f495, [%r34+896];
add.ftz.f32 %f496, %f495, %f492;
ld.shared.f32 %f497, [%r32+1408];
add.ftz.f32 %f498, %f497, %f494;
ld.shared.f32 %f499, [%r34+1408];
add.ftz.f32 %f500, %f499, %f496;
mov.b32 %r1757, %f498;
shfl.sync.bfly.b32 %r1761|%p150, %r1757, %r1731, %r1730, %r1732;
mov.b32 %f501, %r1761;
add.ftz.f32 %f502, %f498, %f501;
mov.b32 %r1762, %f502;
shfl.sync.bfly.b32 %r1764|%p151, %r1762, %r1735, %r1730, %r1732;
mov.b32 %f503, %r1764;
add.ftz.f32 %f504, %f502, %f503;
mov.b32 %r1765, %f504;
shfl.sync.bfly.b32 %r1767|%p152, %r1765, %r1738, %r1730, %r1732;
mov.b32 %f505, %r1767;
add.ftz.f32 %f506, %f504, %f505;
mov.b32 %r1768, %f506;
shfl.sync.bfly.b32 %r1770|%p153, %r1768, %r1741, %r1730, %r1732;
mov.b32 %f507, %r1770;
add.ftz.f32 %f508, %f506, %f507;
mov.b32 %r1771, %f508;
shfl.sync.bfly.b32 %r1773|%p154, %r1771, %r1744, %r1730, %r1732;
mov.b32 %f509, %r1773;
add.ftz.f32 %f54, %f508, %f509;
mov.b32 %r1774, %f500;
shfl.sync.bfly.b32 %r1775|%p155, %r1774, %r1731, %r1730, %r1732;
mov.b32 %f510, %r1775;
add.ftz.f32 %f511, %f500, %f510;
mov.b32 %r1776, %f511;
shfl.sync.bfly.b32 %r1777|%p156, %r1776, %r1735, %r1730, %r1732;
mov.b32 %f512, %r1777;
add.ftz.f32 %f513, %f511, %f512;
mov.b32 %r1778, %f513;
shfl.sync.bfly.b32 %r1779|%p157, %r1778, %r1738, %r1730, %r1732;
mov.b32 %f514, %r1779;
add.ftz.f32 %f515, %f513, %f514;
mov.b32 %r1780, %f515;
shfl.sync.bfly.b32 %r1781|%p158, %r1780, %r1741, %r1730, %r1732;
mov.b32 %f516, %r1781;
add.ftz.f32 %f517, %f515, %f516;
mov.b32 %r1782, %f517;
shfl.sync.bfly.b32 %r1783|%p159, %r1782, %r1744, %r1730, %r1732;
mov.b32 %f518, %r1783;
add.ftz.f32 %f55, %f517, %f518;
add.s32 %r1784, %r4, 3;
setp.ge.u32 %p160, %r1784, %r50;
setp.ne.s32 %p161, %r2, 3;
or.pred %p162, %p161, %p160;
@%p162 bra $L__BB239_78;

add.ftz.f32 %f519, %f5, %f54;
add.ftz.f32 %f520, %f4, %f55;
mov.f32 %f521, 0f40E00000;
min.ftz.f32 %f522, %f520, %f521;
min.ftz.f32 %f523, %f519, %f521;
mov.f32 %f524, 0fC0E00000;
max.ftz.f32 %f525, %f523, %f524;
mul.ftz.f32 %f526, %f522, 0f3FD9DB23;
mul.ftz.f32 %f527, %f526, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f528, %f527;
add.ftz.f32 %f529, %f528, 0f3F800000;
div.approx.ftz.f32 %f530, %f522, %f529;
add.ftz.f32 %f531, %f525, 0f3F800000;
mul.ftz.f32 %f532, %f531, %f530;
st.global.f32 [%rd35+12], %f532;
bra.uni $L__BB239_78;

}
.entry _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<38>;
.reg .b16 %rs<61>;
.reg .f32 %f<109>;
.reg .b32 %r<835>;
.reg .b64 %rd<126>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

ld.param.u64 %rd12, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd13, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd14, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
mov.u32 %r1, %tid.y;
shl.b32 %r39, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r39, %r2;
mov.u32 %r40, %ctaid.x;
shl.b32 %r4, %r40, 2;
shr.u32 %r5, %r38, 8;
mov.u32 %r832, %ctaid.y;
setp.eq.s64 %p1, %rd14, 0;
@%p1 bra $L__BB240_2;

cvta.to.global.u64 %rd16, %rd14;
mul.wide.u32 %rd17, %r832, 4;
add.s64 %rd18, %rd16, %rd17;
ld.global.nc.u32 %r831, [%rd18];
mul.hi.u32 %r41, %r832, %r21;
add.s32 %r42, %r41, %r832;
shr.u32 %r43, %r42, %r22;
mul.lo.s32 %r44, %r43, %r23;
sub.s32 %r832, %r832, %r44;
bra.uni $L__BB240_3;

$L__BB240_2:
mul.hi.u32 %r45, %r832, %r26;
add.s32 %r46, %r45, %r832;
shr.u32 %r831, %r46, %r27;

$L__BB240_3:
mov.u32 %r12, %ctaid.z;
shr.s32 %r47, %r3, 31;
shr.u32 %r48, %r47, 29;
add.s32 %r49, %r3, %r48;
shr.s32 %r834, %r49, 3;
setp.ge.s32 %p2, %r834, %r5;
mov.f32 %f105, 0f00000000;
mov.f32 %f106, %f105;
mov.f32 %f107, %f105;
mov.f32 %f108, %f105;
@%p2 bra $L__BB240_6;

and.b32 %r53, %r49, -8;
sub.s32 %r54, %r3, %r53;
shl.b32 %r55, %r54, 2;
shl.b32 %r56, %r54, 1;
add.s32 %r57, %r56, 33;
cvt.s64.s32 %rd1, %r54;
shl.b32 %r833, %r834, 3;
mad.lo.s32 %r58, %r4, %r24, %r834;
mul.hi.u32 %r59, %r12, %r32;
add.s32 %r60, %r59, %r12;
shr.u32 %r61, %r60, %r33;
mad.lo.s32 %r62, %r61, %r35, %r58;
mad.lo.s32 %r63, %r831, %r29, %r62;
mul.wide.s32 %rd2, %r63, 98;
mul.wide.s32 %rd19, %r55, 2;
add.s64 %rd20, %rd2, %rd19;
add.s64 %rd3, %rd20, 8;
mul.wide.s32 %rd21, %r57, 2;
add.s64 %rd4, %rd21, %rd2;
mul.wide.s32 %rd5, %r24, 98;
cvta.to.global.u64 %rd6, %rd13;
cvta.to.global.u64 %rd125, %rd12;
mul.lo.s32 %r64, %r832, %r30;
cvt.u64.u32 %rd22, %r64;
mul.lo.s32 %r65, %r12, %r36;
cvt.u64.u32 %rd23, %r65;
add.s64 %rd8, %rd22, %rd23;
mov.f32 %f105, 0f00000000;

$L__BB240_5:
cvt.u64.u32 %rd24, %r833;
add.s64 %rd25, %rd8, %rd24;
add.s64 %rd26, %rd125, %rd3;
add.s64 %rd27, %rd26, -6;
ld.global.nc.u16 %rs5, [%rd26+-4];
ld.global.nc.u16 %rs6, [%rd26+-2];
ld.global.nc.u16 %rs7, [%rd26];
add.s64 %rd28, %rd125, %rd4;
ld.global.nc.u16 %rs8, [%rd28];
cvt.u32.u16 %r262, %rs8;
ld.global.nc.u16 %rs9, [%rd28+2];
cvt.u32.u16 %r263, %rs9;
mov.b32 %r264, {%rs8, %rs9};
add.s64 %rd29, %rd25, %rd1;
mul.lo.s64 %rd30, %rd29, 36;
ld.global.nc.u16 %rs10, [%rd26+-6];
shl.b16 %rs11, %rs10, 2;
cvt.u64.u16 %rd31, %rs11;
and.b64 %rd32, %rd31, 1020;
mov.u64 %rd33, iq3xxs_grid;
add.s64 %rd34, %rd33, %rd32;
ld.global.nc.u32 %r265, [%rd34];
shr.u16 %rs12, %rs10, 8;
cvt.u32.u16 %r266, %rs12;
mul.wide.u32 %rd35, %r266, 4;
add.s64 %rd36, %rd33, %rd35;
ld.global.nc.u32 %r267, [%rd36];
and.b32 %r268, %r262, 255;
popc.b32 %r269, %r268;
shl.b32 %r270, %r269, 7;
and.b32 %r271, %r270, 128;
xor.b32 %r272, %r271, %r268;
mul.lo.s32 %r273, %r272, 16843009;
and.b32 %r274, %r273, 134480385;
or.b32 %r275, %r274, -2139062144;
add.s32 %r276, %r275, -16843009;
or.b32 %r66, %r276, %r274;

prmt.b32 %r66, %r66, 0, 0xba98;

xor.b32 %r277, %r66, %r265;
or.b32 %r278, %r277, -2139062144;
and.b32 %r279, %r66, 2139062143;
and.b32 %r280, %r265, -2139062144;
sub.s32 %r281, %r278, %r279;
xor.b32 %r282, %r280, %r281;
xor.b32 %r71, %r282, -2139062144;
add.s64 %rd37, %rd6, %rd30;
ld.global.nc.u32 %r72, [%rd37+4];
and.b32 %r283, %r273, -2143281136;
or.b32 %r284, %r283, -2139062144;
add.s32 %r285, %r284, -16843009;
or.b32 %r68, %r285, %r283;

prmt.b32 %r68, %r68, 0, 0xba98;

xor.b32 %r286, %r68, %r267;
or.b32 %r287, %r286, -2139062144;
and.b32 %r288, %r68, 2139062143;
and.b32 %r289, %r267, -2139062144;
sub.s32 %r290, %r287, %r288;
xor.b32 %r291, %r289, %r290;
xor.b32 %r75, %r291, -2139062144;
ld.global.nc.u32 %r76, [%rd37+8];
mov.u32 %r220, 0;

dp4a.s32.s32 %r70, %r71, %r72, %r220;


dp4a.s32.s32 %r74, %r75, %r76, %r70;

shl.b16 %rs13, %rs5, 2;
cvt.u64.u16 %rd38, %rs13;
and.b64 %rd39, %rd38, 1020;
add.s64 %rd40, %rd33, %rd39;
ld.global.nc.u32 %r292, [%rd40];
shr.u16 %rs14, %rs5, 8;
cvt.u32.u16 %r293, %rs14;
mul.wide.u32 %rd41, %r293, 4;
add.s64 %rd42, %rd33, %rd41;
ld.global.nc.u32 %r294, [%rd42];
shr.u32 %r295, %r262, 7;
and.b32 %r296, %r295, 255;
popc.b32 %r297, %r296;
shl.b32 %r298, %r297, 7;
and.b32 %r299, %r298, 128;
xor.b32 %r300, %r299, %r296;
mul.lo.s32 %r301, %r300, 16843009;
and.b32 %r302, %r301, 134480385;
or.b32 %r303, %r302, -2139062144;
add.s32 %r304, %r303, -16843009;
or.b32 %r78, %r304, %r302;

prmt.b32 %r78, %r78, 0, 0xba98;

xor.b32 %r305, %r78, %r292;
or.b32 %r306, %r305, -2139062144;
and.b32 %r307, %r78, 2139062143;
and.b32 %r308, %r292, -2139062144;
sub.s32 %r309, %r306, %r307;
xor.b32 %r310, %r308, %r309;
xor.b32 %r83, %r310, -2139062144;
ld.global.nc.u32 %r84, [%rd37+12];
and.b32 %r311, %r301, -2143281136;
or.b32 %r312, %r311, -2139062144;
add.s32 %r313, %r312, -16843009;
or.b32 %r80, %r313, %r311;

prmt.b32 %r80, %r80, 0, 0xba98;

xor.b32 %r314, %r80, %r294;
or.b32 %r315, %r314, -2139062144;
and.b32 %r316, %r80, 2139062143;
and.b32 %r317, %r294, -2139062144;
sub.s32 %r318, %r315, %r316;
xor.b32 %r319, %r317, %r318;
xor.b32 %r87, %r319, -2139062144;
ld.global.nc.u32 %r88, [%rd37+16];

dp4a.s32.s32 %r82, %r83, %r84, %r74;


dp4a.s32.s32 %r86, %r87, %r88, %r82;

shl.b16 %rs15, %rs6, 2;
cvt.u64.u16 %rd43, %rs15;
and.b64 %rd44, %rd43, 1020;
add.s64 %rd45, %rd33, %rd44;
ld.global.nc.u32 %r320, [%rd45];
shr.u16 %rs16, %rs6, 8;
cvt.u32.u16 %r321, %rs16;
mul.wide.u32 %rd46, %r321, 4;
add.s64 %rd47, %rd33, %rd46;
ld.global.nc.u32 %r322, [%rd47];
shr.u32 %r323, %r264, 14;
and.b32 %r324, %r323, 255;
popc.b32 %r325, %r324;
shl.b32 %r326, %r325, 7;
and.b32 %r327, %r326, 128;
xor.b32 %r328, %r327, %r324;
mul.lo.s32 %r329, %r328, 16843009;
and.b32 %r330, %r329, 134480385;
or.b32 %r331, %r330, -2139062144;
add.s32 %r332, %r331, -16843009;
or.b32 %r90, %r332, %r330;

prmt.b32 %r90, %r90, 0, 0xba98;

xor.b32 %r333, %r90, %r320;
or.b32 %r334, %r333, -2139062144;
and.b32 %r335, %r90, 2139062143;
and.b32 %r336, %r320, -2139062144;
sub.s32 %r337, %r334, %r335;
xor.b32 %r338, %r336, %r337;
xor.b32 %r95, %r338, -2139062144;
ld.global.nc.u32 %r96, [%rd37+20];
and.b32 %r339, %r329, -2143281136;
or.b32 %r340, %r339, -2139062144;
add.s32 %r341, %r340, -16843009;
or.b32 %r92, %r341, %r339;

prmt.b32 %r92, %r92, 0, 0xba98;

xor.b32 %r342, %r92, %r322;
or.b32 %r343, %r342, -2139062144;
and.b32 %r344, %r92, 2139062143;
and.b32 %r345, %r322, -2139062144;
sub.s32 %r346, %r343, %r344;
xor.b32 %r347, %r345, %r346;
xor.b32 %r99, %r347, -2139062144;
ld.global.nc.u32 %r100, [%rd37+24];

dp4a.s32.s32 %r94, %r95, %r96, %r86;


dp4a.s32.s32 %r98, %r99, %r100, %r94;

shl.b16 %rs17, %rs7, 2;
cvt.u64.u16 %rd48, %rs17;
and.b64 %rd49, %rd48, 1020;
add.s64 %rd50, %rd33, %rd49;
ld.global.nc.u32 %r348, [%rd50];
shr.u16 %rs18, %rs7, 8;
cvt.u32.u16 %r349, %rs18;
mul.wide.u32 %rd51, %r349, 4;
add.s64 %rd52, %rd33, %rd51;
ld.global.nc.u32 %r350, [%rd52];
shr.u32 %r351, %r263, 5;
and.b32 %r352, %r351, 255;
popc.b32 %r353, %r352;
shl.b32 %r354, %r353, 7;
and.b32 %r355, %r354, 128;
xor.b32 %r356, %r355, %r352;
mul.lo.s32 %r357, %r356, 16843009;
and.b32 %r358, %r357, 134480385;
or.b32 %r359, %r358, -2139062144;
add.s32 %r360, %r359, -16843009;
or.b32 %r102, %r360, %r358;

prmt.b32 %r102, %r102, 0, 0xba98;

xor.b32 %r361, %r102, %r348;
or.b32 %r362, %r361, -2139062144;
and.b32 %r363, %r102, 2139062143;
and.b32 %r364, %r348, -2139062144;
sub.s32 %r365, %r362, %r363;
xor.b32 %r366, %r364, %r365;
xor.b32 %r107, %r366, -2139062144;
ld.global.nc.u32 %r108, [%rd37+28];
and.b32 %r367, %r357, -2143281136;
or.b32 %r368, %r367, -2139062144;
add.s32 %r369, %r368, -16843009;
or.b32 %r104, %r369, %r367;

prmt.b32 %r104, %r104, 0, 0xba98;

xor.b32 %r370, %r104, %r350;
or.b32 %r371, %r370, -2139062144;
and.b32 %r372, %r104, 2139062143;
and.b32 %r373, %r350, -2139062144;
sub.s32 %r374, %r371, %r372;
xor.b32 %r375, %r373, %r374;
xor.b32 %r111, %r375, -2139062144;
ld.global.nc.u32 %r112, [%rd37+32];

dp4a.s32.s32 %r106, %r107, %r108, %r98;

shr.u32 %r376, %r263, 12;

dp4a.s32.s32 %r110, %r111, %r112, %r106;

shr.u32 %r377, %r110, 31;
add.s32 %r378, %r110, %r377;
shr.s32 %r379, %r378, 1;
mad.lo.s32 %r380, %r110, %r376, %r379;
shr.u32 %r381, %r380, 31;
add.s32 %r382, %r380, %r381;
shr.s32 %r383, %r382, 1;
add.s64 %rd53, %rd125, %rd2;
ld.global.nc.u16 %rs1, [%rd53];

{ cvt.f32.f16 %f25, %rs1;}


ld.global.nc.u32 %r114, [%rd37];

{.reg .f16 low,high;
mov.b32 {low,high},%r114;
cvt.f32.f16 %f26, low;}


mul.ftz.f32 %f33, %f25, %f26;
cvt.rn.f32.s32 %f34, %r383;
fma.rn.ftz.f32 %f108, %f33, %f34, %f108;
add.s64 %rd54, %rd27, %rd5;
ld.global.nc.u16 %rs19, [%rd54+2];
ld.global.nc.u16 %rs20, [%rd54+4];
ld.global.nc.u16 %rs21, [%rd54+6];
add.s64 %rd55, %rd28, %rd5;
ld.global.nc.u16 %rs22, [%rd55];
cvt.u32.u16 %r384, %rs22;
ld.global.nc.u16 %rs23, [%rd55+2];
cvt.u32.u16 %r385, %rs23;
mov.b32 %r386, {%rs22, %rs23};
ld.global.nc.u16 %rs24, [%rd54];
shl.b16 %rs25, %rs24, 2;
cvt.u64.u16 %rd56, %rs25;
and.b64 %rd57, %rd56, 1020;
add.s64 %rd58, %rd33, %rd57;
ld.global.nc.u32 %r387, [%rd58];
shr.u16 %rs26, %rs24, 8;
cvt.u32.u16 %r388, %rs26;
mul.wide.u32 %rd59, %r388, 4;
add.s64 %rd60, %rd33, %rd59;
ld.global.nc.u32 %r389, [%rd60];
and.b32 %r390, %r384, 255;
popc.b32 %r391, %r390;
shl.b32 %r392, %r391, 7;
and.b32 %r393, %r392, 128;
xor.b32 %r394, %r393, %r390;
mul.lo.s32 %r395, %r394, 16843009;
and.b32 %r396, %r395, 134480385;
or.b32 %r397, %r396, -2139062144;
add.s32 %r398, %r397, -16843009;
or.b32 %r115, %r398, %r396;

prmt.b32 %r115, %r115, 0, 0xba98;

xor.b32 %r399, %r115, %r387;
or.b32 %r400, %r399, -2139062144;
and.b32 %r401, %r115, 2139062143;
and.b32 %r402, %r387, -2139062144;
sub.s32 %r403, %r400, %r401;
xor.b32 %r404, %r402, %r403;
xor.b32 %r120, %r404, -2139062144;
and.b32 %r405, %r395, -2143281136;
or.b32 %r406, %r405, -2139062144;
add.s32 %r407, %r406, -16843009;
or.b32 %r117, %r407, %r405;

prmt.b32 %r117, %r117, 0, 0xba98;

xor.b32 %r408, %r117, %r389;
or.b32 %r409, %r408, -2139062144;
and.b32 %r410, %r117, 2139062143;
and.b32 %r411, %r389, -2139062144;
sub.s32 %r412, %r409, %r410;
xor.b32 %r413, %r411, %r412;
xor.b32 %r124, %r413, -2139062144;

dp4a.s32.s32 %r119, %r120, %r72, %r220;


dp4a.s32.s32 %r123, %r124, %r76, %r119;

shl.b16 %rs27, %rs19, 2;
cvt.u64.u16 %rd61, %rs27;
and.b64 %rd62, %rd61, 1020;
add.s64 %rd63, %rd33, %rd62;
ld.global.nc.u32 %r414, [%rd63];
shr.u16 %rs28, %rs19, 8;
cvt.u32.u16 %r415, %rs28;
mul.wide.u32 %rd64, %r415, 4;
add.s64 %rd65, %rd33, %rd64;
ld.global.nc.u32 %r416, [%rd65];
shr.u32 %r417, %r384, 7;
and.b32 %r418, %r417, 255;
popc.b32 %r419, %r418;
shl.b32 %r420, %r419, 7;
and.b32 %r421, %r420, 128;
xor.b32 %r422, %r421, %r418;
mul.lo.s32 %r423, %r422, 16843009;
and.b32 %r424, %r423, 134480385;
or.b32 %r425, %r424, -2139062144;
add.s32 %r426, %r425, -16843009;
or.b32 %r127, %r426, %r424;

prmt.b32 %r127, %r127, 0, 0xba98;

xor.b32 %r427, %r127, %r414;
or.b32 %r428, %r427, -2139062144;
and.b32 %r429, %r127, 2139062143;
and.b32 %r430, %r414, -2139062144;
sub.s32 %r431, %r428, %r429;
xor.b32 %r432, %r430, %r431;
xor.b32 %r132, %r432, -2139062144;
and.b32 %r433, %r423, -2143281136;
or.b32 %r434, %r433, -2139062144;
add.s32 %r435, %r434, -16843009;
or.b32 %r129, %r435, %r433;

prmt.b32 %r129, %r129, 0, 0xba98;

xor.b32 %r436, %r129, %r416;
or.b32 %r437, %r436, -2139062144;
and.b32 %r438, %r129, 2139062143;
and.b32 %r439, %r416, -2139062144;
sub.s32 %r440, %r437, %r438;
xor.b32 %r441, %r439, %r440;
xor.b32 %r136, %r441, -2139062144;

dp4a.s32.s32 %r131, %r132, %r84, %r123;


dp4a.s32.s32 %r135, %r136, %r88, %r131;

shl.b16 %rs29, %rs20, 2;
cvt.u64.u16 %rd66, %rs29;
and.b64 %rd67, %rd66, 1020;
add.s64 %rd68, %rd33, %rd67;
ld.global.nc.u32 %r442, [%rd68];
shr.u16 %rs30, %rs20, 8;
cvt.u32.u16 %r443, %rs30;
mul.wide.u32 %rd69, %r443, 4;
add.s64 %rd70, %rd33, %rd69;
ld.global.nc.u32 %r444, [%rd70];
shr.u32 %r445, %r386, 14;
and.b32 %r446, %r445, 255;
popc.b32 %r447, %r446;
shl.b32 %r448, %r447, 7;
and.b32 %r449, %r448, 128;
xor.b32 %r450, %r449, %r446;
mul.lo.s32 %r451, %r450, 16843009;
and.b32 %r452, %r451, 134480385;
or.b32 %r453, %r452, -2139062144;
add.s32 %r454, %r453, -16843009;
or.b32 %r139, %r454, %r452;

prmt.b32 %r139, %r139, 0, 0xba98;

xor.b32 %r455, %r139, %r442;
or.b32 %r456, %r455, -2139062144;
and.b32 %r457, %r139, 2139062143;
and.b32 %r458, %r442, -2139062144;
sub.s32 %r459, %r456, %r457;
xor.b32 %r460, %r458, %r459;
xor.b32 %r144, %r460, -2139062144;
and.b32 %r461, %r451, -2143281136;
or.b32 %r462, %r461, -2139062144;
add.s32 %r463, %r462, -16843009;
or.b32 %r141, %r463, %r461;

prmt.b32 %r141, %r141, 0, 0xba98;

xor.b32 %r464, %r141, %r444;
or.b32 %r465, %r464, -2139062144;
and.b32 %r466, %r141, 2139062143;
and.b32 %r467, %r444, -2139062144;
sub.s32 %r468, %r465, %r466;
xor.b32 %r469, %r467, %r468;
xor.b32 %r148, %r469, -2139062144;

dp4a.s32.s32 %r143, %r144, %r96, %r135;


dp4a.s32.s32 %r147, %r148, %r100, %r143;

shl.b16 %rs31, %rs21, 2;
cvt.u64.u16 %rd71, %rs31;
and.b64 %rd72, %rd71, 1020;
add.s64 %rd73, %rd33, %rd72;
ld.global.nc.u32 %r470, [%rd73];
shr.u16 %rs32, %rs21, 8;
cvt.u32.u16 %r471, %rs32;
mul.wide.u32 %rd74, %r471, 4;
add.s64 %rd75, %rd33, %rd74;
ld.global.nc.u32 %r472, [%rd75];
shr.u32 %r473, %r385, 5;
and.b32 %r474, %r473, 255;
popc.b32 %r475, %r474;
shl.b32 %r476, %r475, 7;
and.b32 %r477, %r476, 128;
xor.b32 %r478, %r477, %r474;
mul.lo.s32 %r479, %r478, 16843009;
and.b32 %r480, %r479, 134480385;
or.b32 %r481, %r480, -2139062144;
add.s32 %r482, %r481, -16843009;
or.b32 %r151, %r482, %r480;

prmt.b32 %r151, %r151, 0, 0xba98;

xor.b32 %r483, %r151, %r470;
or.b32 %r484, %r483, -2139062144;
and.b32 %r485, %r151, 2139062143;
and.b32 %r486, %r470, -2139062144;
sub.s32 %r487, %r484, %r485;
xor.b32 %r488, %r486, %r487;
xor.b32 %r156, %r488, -2139062144;
and.b32 %r489, %r479, -2143281136;
or.b32 %r490, %r489, -2139062144;
add.s32 %r491, %r490, -16843009;
or.b32 %r153, %r491, %r489;

prmt.b32 %r153, %r153, 0, 0xba98;

xor.b32 %r492, %r153, %r472;
or.b32 %r493, %r492, -2139062144;
and.b32 %r494, %r153, 2139062143;
and.b32 %r495, %r472, -2139062144;
sub.s32 %r496, %r493, %r494;
xor.b32 %r497, %r495, %r496;
xor.b32 %r160, %r497, -2139062144;

dp4a.s32.s32 %r155, %r156, %r108, %r147;

shr.u32 %r498, %r385, 12;

dp4a.s32.s32 %r159, %r160, %r112, %r155;

shr.u32 %r499, %r159, 31;
add.s32 %r500, %r159, %r499;
shr.s32 %r501, %r500, 1;
mad.lo.s32 %r502, %r159, %r498, %r501;
shr.u32 %r503, %r502, 31;
add.s32 %r504, %r502, %r503;
shr.s32 %r505, %r504, 1;
add.s64 %rd76, %rd53, %rd5;
ld.global.nc.u16 %rs2, [%rd76];

{ cvt.f32.f16 %f27, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r114;
cvt.f32.f16 %f28, low;}


mul.ftz.f32 %f35, %f27, %f28;
cvt.rn.f32.s32 %f36, %r505;
fma.rn.ftz.f32 %f107, %f35, %f36, %f107;
add.s64 %rd77, %rd54, %rd5;
ld.global.nc.u16 %rs33, [%rd77+2];
ld.global.nc.u16 %rs34, [%rd77+4];
ld.global.nc.u16 %rs35, [%rd77+6];
add.s64 %rd78, %rd55, %rd5;
ld.global.nc.u16 %rs36, [%rd78];
cvt.u32.u16 %r506, %rs36;
ld.global.nc.u16 %rs37, [%rd78+2];
cvt.u32.u16 %r507, %rs37;
mov.b32 %r508, {%rs36, %rs37};
ld.global.nc.u16 %rs38, [%rd77];
shl.b16 %rs39, %rs38, 2;
cvt.u64.u16 %rd79, %rs39;
and.b64 %rd80, %rd79, 1020;
add.s64 %rd81, %rd33, %rd80;
ld.global.nc.u32 %r509, [%rd81];
shr.u16 %rs40, %rs38, 8;
cvt.u32.u16 %r510, %rs40;
mul.wide.u32 %rd82, %r510, 4;
add.s64 %rd83, %rd33, %rd82;
ld.global.nc.u32 %r511, [%rd83];
and.b32 %r512, %r506, 255;
popc.b32 %r513, %r512;
shl.b32 %r514, %r513, 7;
and.b32 %r515, %r514, 128;
xor.b32 %r516, %r515, %r512;
mul.lo.s32 %r517, %r516, 16843009;
and.b32 %r518, %r517, 134480385;
or.b32 %r519, %r518, -2139062144;
add.s32 %r520, %r519, -16843009;
or.b32 %r164, %r520, %r518;

prmt.b32 %r164, %r164, 0, 0xba98;

xor.b32 %r521, %r164, %r509;
or.b32 %r522, %r521, -2139062144;
and.b32 %r523, %r164, 2139062143;
and.b32 %r524, %r509, -2139062144;
sub.s32 %r525, %r522, %r523;
xor.b32 %r526, %r524, %r525;
xor.b32 %r169, %r526, -2139062144;
and.b32 %r527, %r517, -2143281136;
or.b32 %r528, %r527, -2139062144;
add.s32 %r529, %r528, -16843009;
or.b32 %r166, %r529, %r527;

prmt.b32 %r166, %r166, 0, 0xba98;

xor.b32 %r530, %r166, %r511;
or.b32 %r531, %r530, -2139062144;
and.b32 %r532, %r166, 2139062143;
and.b32 %r533, %r511, -2139062144;
sub.s32 %r534, %r531, %r532;
xor.b32 %r535, %r533, %r534;
xor.b32 %r173, %r535, -2139062144;

dp4a.s32.s32 %r168, %r169, %r72, %r220;


dp4a.s32.s32 %r172, %r173, %r76, %r168;

shl.b16 %rs41, %rs33, 2;
cvt.u64.u16 %rd84, %rs41;
and.b64 %rd85, %rd84, 1020;
add.s64 %rd86, %rd33, %rd85;
ld.global.nc.u32 %r536, [%rd86];
shr.u16 %rs42, %rs33, 8;
cvt.u32.u16 %r537, %rs42;
mul.wide.u32 %rd87, %r537, 4;
add.s64 %rd88, %rd33, %rd87;
ld.global.nc.u32 %r538, [%rd88];
shr.u32 %r539, %r506, 7;
and.b32 %r540, %r539, 255;
popc.b32 %r541, %r540;
shl.b32 %r542, %r541, 7;
and.b32 %r543, %r542, 128;
xor.b32 %r544, %r543, %r540;
mul.lo.s32 %r545, %r544, 16843009;
and.b32 %r546, %r545, 134480385;
or.b32 %r547, %r546, -2139062144;
add.s32 %r548, %r547, -16843009;
or.b32 %r176, %r548, %r546;

prmt.b32 %r176, %r176, 0, 0xba98;

xor.b32 %r549, %r176, %r536;
or.b32 %r550, %r549, -2139062144;
and.b32 %r551, %r176, 2139062143;
and.b32 %r552, %r536, -2139062144;
sub.s32 %r553, %r550, %r551;
xor.b32 %r554, %r552, %r553;
xor.b32 %r181, %r554, -2139062144;
and.b32 %r555, %r545, -2143281136;
or.b32 %r556, %r555, -2139062144;
add.s32 %r557, %r556, -16843009;
or.b32 %r178, %r557, %r555;

prmt.b32 %r178, %r178, 0, 0xba98;

xor.b32 %r558, %r178, %r538;
or.b32 %r559, %r558, -2139062144;
and.b32 %r560, %r178, 2139062143;
and.b32 %r561, %r538, -2139062144;
sub.s32 %r562, %r559, %r560;
xor.b32 %r563, %r561, %r562;
xor.b32 %r185, %r563, -2139062144;

dp4a.s32.s32 %r180, %r181, %r84, %r172;


dp4a.s32.s32 %r184, %r185, %r88, %r180;

shl.b16 %rs43, %rs34, 2;
cvt.u64.u16 %rd89, %rs43;
and.b64 %rd90, %rd89, 1020;
add.s64 %rd91, %rd33, %rd90;
ld.global.nc.u32 %r564, [%rd91];
shr.u16 %rs44, %rs34, 8;
cvt.u32.u16 %r565, %rs44;
mul.wide.u32 %rd92, %r565, 4;
add.s64 %rd93, %rd33, %rd92;
ld.global.nc.u32 %r566, [%rd93];
shr.u32 %r567, %r508, 14;
and.b32 %r568, %r567, 255;
popc.b32 %r569, %r568;
shl.b32 %r570, %r569, 7;
and.b32 %r571, %r570, 128;
xor.b32 %r572, %r571, %r568;
mul.lo.s32 %r573, %r572, 16843009;
and.b32 %r574, %r573, 134480385;
or.b32 %r575, %r574, -2139062144;
add.s32 %r576, %r575, -16843009;
or.b32 %r188, %r576, %r574;

prmt.b32 %r188, %r188, 0, 0xba98;

xor.b32 %r577, %r188, %r564;
or.b32 %r578, %r577, -2139062144;
and.b32 %r579, %r188, 2139062143;
and.b32 %r580, %r564, -2139062144;
sub.s32 %r581, %r578, %r579;
xor.b32 %r582, %r580, %r581;
xor.b32 %r193, %r582, -2139062144;
and.b32 %r583, %r573, -2143281136;
or.b32 %r584, %r583, -2139062144;
add.s32 %r585, %r584, -16843009;
or.b32 %r190, %r585, %r583;

prmt.b32 %r190, %r190, 0, 0xba98;

xor.b32 %r586, %r190, %r566;
or.b32 %r587, %r586, -2139062144;
and.b32 %r588, %r190, 2139062143;
and.b32 %r589, %r566, -2139062144;
sub.s32 %r590, %r587, %r588;
xor.b32 %r591, %r589, %r590;
xor.b32 %r197, %r591, -2139062144;

dp4a.s32.s32 %r192, %r193, %r96, %r184;


dp4a.s32.s32 %r196, %r197, %r100, %r192;

shl.b16 %rs45, %rs35, 2;
cvt.u64.u16 %rd94, %rs45;
and.b64 %rd95, %rd94, 1020;
add.s64 %rd96, %rd33, %rd95;
ld.global.nc.u32 %r592, [%rd96];
shr.u16 %rs46, %rs35, 8;
cvt.u32.u16 %r593, %rs46;
mul.wide.u32 %rd97, %r593, 4;
add.s64 %rd98, %rd33, %rd97;
ld.global.nc.u32 %r594, [%rd98];
shr.u32 %r595, %r507, 5;
and.b32 %r596, %r595, 255;
popc.b32 %r597, %r596;
shl.b32 %r598, %r597, 7;
and.b32 %r599, %r598, 128;
xor.b32 %r600, %r599, %r596;
mul.lo.s32 %r601, %r600, 16843009;
and.b32 %r602, %r601, 134480385;
or.b32 %r603, %r602, -2139062144;
add.s32 %r604, %r603, -16843009;
or.b32 %r200, %r604, %r602;

prmt.b32 %r200, %r200, 0, 0xba98;

xor.b32 %r605, %r200, %r592;
or.b32 %r606, %r605, -2139062144;
and.b32 %r607, %r200, 2139062143;
and.b32 %r608, %r592, -2139062144;
sub.s32 %r609, %r606, %r607;
xor.b32 %r610, %r608, %r609;
xor.b32 %r205, %r610, -2139062144;
and.b32 %r611, %r601, -2143281136;
or.b32 %r612, %r611, -2139062144;
add.s32 %r613, %r612, -16843009;
or.b32 %r202, %r613, %r611;

prmt.b32 %r202, %r202, 0, 0xba98;

xor.b32 %r614, %r202, %r594;
or.b32 %r615, %r614, -2139062144;
and.b32 %r616, %r202, 2139062143;
and.b32 %r617, %r594, -2139062144;
sub.s32 %r618, %r615, %r616;
xor.b32 %r619, %r617, %r618;
xor.b32 %r209, %r619, -2139062144;

dp4a.s32.s32 %r204, %r205, %r108, %r196;

shr.u32 %r620, %r507, 12;

dp4a.s32.s32 %r208, %r209, %r112, %r204;

shr.u32 %r621, %r208, 31;
add.s32 %r622, %r208, %r621;
shr.s32 %r623, %r622, 1;
mad.lo.s32 %r624, %r208, %r620, %r623;
shr.u32 %r625, %r624, 31;
add.s32 %r626, %r624, %r625;
shr.s32 %r627, %r626, 1;
add.s64 %rd99, %rd76, %rd5;
ld.global.nc.u16 %rs3, [%rd99];

{ cvt.f32.f16 %f29, %rs3;}



{.reg .f16 low,high;
mov.b32 {low,high},%r114;
cvt.f32.f16 %f30, low;}


mul.ftz.f32 %f37, %f29, %f30;
cvt.rn.f32.s32 %f38, %r627;
fma.rn.ftz.f32 %f106, %f37, %f38, %f106;
add.s64 %rd100, %rd77, %rd5;
ld.global.nc.u16 %rs47, [%rd100+2];
ld.global.nc.u16 %rs48, [%rd100+4];
ld.global.nc.u16 %rs49, [%rd100+6];
add.s64 %rd101, %rd78, %rd5;
ld.global.nc.u16 %rs50, [%rd101];
cvt.u32.u16 %r628, %rs50;
ld.global.nc.u16 %rs51, [%rd101+2];
cvt.u32.u16 %r629, %rs51;
mov.b32 %r630, {%rs50, %rs51};
ld.global.nc.u16 %rs52, [%rd100];
shl.b16 %rs53, %rs52, 2;
cvt.u64.u16 %rd102, %rs53;
and.b64 %rd103, %rd102, 1020;
add.s64 %rd104, %rd33, %rd103;
ld.global.nc.u32 %r631, [%rd104];
shr.u16 %rs54, %rs52, 8;
cvt.u32.u16 %r632, %rs54;
mul.wide.u32 %rd105, %r632, 4;
add.s64 %rd106, %rd33, %rd105;
ld.global.nc.u32 %r633, [%rd106];
and.b32 %r634, %r628, 255;
popc.b32 %r635, %r634;
shl.b32 %r636, %r635, 7;
and.b32 %r637, %r636, 128;
xor.b32 %r638, %r637, %r634;
mul.lo.s32 %r639, %r638, 16843009;
and.b32 %r640, %r639, 134480385;
or.b32 %r641, %r640, -2139062144;
add.s32 %r642, %r641, -16843009;
or.b32 %r213, %r642, %r640;

prmt.b32 %r213, %r213, 0, 0xba98;

xor.b32 %r643, %r213, %r631;
or.b32 %r644, %r643, -2139062144;
and.b32 %r645, %r213, 2139062143;
and.b32 %r646, %r631, -2139062144;
sub.s32 %r647, %r644, %r645;
xor.b32 %r648, %r646, %r647;
xor.b32 %r218, %r648, -2139062144;
and.b32 %r649, %r639, -2143281136;
or.b32 %r650, %r649, -2139062144;
add.s32 %r651, %r650, -16843009;
or.b32 %r215, %r651, %r649;

prmt.b32 %r215, %r215, 0, 0xba98;

xor.b32 %r652, %r215, %r633;
or.b32 %r653, %r652, -2139062144;
and.b32 %r654, %r215, 2139062143;
and.b32 %r655, %r633, -2139062144;
sub.s32 %r656, %r653, %r654;
xor.b32 %r657, %r655, %r656;
xor.b32 %r222, %r657, -2139062144;

dp4a.s32.s32 %r217, %r218, %r72, %r220;


dp4a.s32.s32 %r221, %r222, %r76, %r217;

shl.b16 %rs55, %rs47, 2;
cvt.u64.u16 %rd107, %rs55;
and.b64 %rd108, %rd107, 1020;
add.s64 %rd109, %rd33, %rd108;
ld.global.nc.u32 %r658, [%rd109];
shr.u16 %rs56, %rs47, 8;
cvt.u32.u16 %r659, %rs56;
mul.wide.u32 %rd110, %r659, 4;
add.s64 %rd111, %rd33, %rd110;
ld.global.nc.u32 %r660, [%rd111];
shr.u32 %r661, %r628, 7;
and.b32 %r662, %r661, 255;
popc.b32 %r663, %r662;
shl.b32 %r664, %r663, 7;
and.b32 %r665, %r664, 128;
xor.b32 %r666, %r665, %r662;
mul.lo.s32 %r667, %r666, 16843009;
and.b32 %r668, %r667, 134480385;
or.b32 %r669, %r668, -2139062144;
add.s32 %r670, %r669, -16843009;
or.b32 %r225, %r670, %r668;

prmt.b32 %r225, %r225, 0, 0xba98;

xor.b32 %r671, %r225, %r658;
or.b32 %r672, %r671, -2139062144;
and.b32 %r673, %r225, 2139062143;
and.b32 %r674, %r658, -2139062144;
sub.s32 %r675, %r672, %r673;
xor.b32 %r676, %r674, %r675;
xor.b32 %r230, %r676, -2139062144;
and.b32 %r677, %r667, -2143281136;
or.b32 %r678, %r677, -2139062144;
add.s32 %r679, %r678, -16843009;
or.b32 %r227, %r679, %r677;

prmt.b32 %r227, %r227, 0, 0xba98;

xor.b32 %r680, %r227, %r660;
or.b32 %r681, %r680, -2139062144;
and.b32 %r682, %r227, 2139062143;
and.b32 %r683, %r660, -2139062144;
sub.s32 %r684, %r681, %r682;
xor.b32 %r685, %r683, %r684;
xor.b32 %r234, %r685, -2139062144;

dp4a.s32.s32 %r229, %r230, %r84, %r221;


dp4a.s32.s32 %r233, %r234, %r88, %r229;

shl.b16 %rs57, %rs48, 2;
cvt.u64.u16 %rd112, %rs57;
and.b64 %rd113, %rd112, 1020;
add.s64 %rd114, %rd33, %rd113;
ld.global.nc.u32 %r686, [%rd114];
shr.u16 %rs58, %rs48, 8;
cvt.u32.u16 %r687, %rs58;
mul.wide.u32 %rd115, %r687, 4;
add.s64 %rd116, %rd33, %rd115;
ld.global.nc.u32 %r688, [%rd116];
shr.u32 %r689, %r630, 14;
and.b32 %r690, %r689, 255;
popc.b32 %r691, %r690;
shl.b32 %r692, %r691, 7;
and.b32 %r693, %r692, 128;
xor.b32 %r694, %r693, %r690;
mul.lo.s32 %r695, %r694, 16843009;
and.b32 %r696, %r695, 134480385;
or.b32 %r697, %r696, -2139062144;
add.s32 %r698, %r697, -16843009;
or.b32 %r237, %r698, %r696;

prmt.b32 %r237, %r237, 0, 0xba98;

xor.b32 %r699, %r237, %r686;
or.b32 %r700, %r699, -2139062144;
and.b32 %r701, %r237, 2139062143;
and.b32 %r702, %r686, -2139062144;
sub.s32 %r703, %r700, %r701;
xor.b32 %r704, %r702, %r703;
xor.b32 %r242, %r704, -2139062144;
and.b32 %r705, %r695, -2143281136;
or.b32 %r706, %r705, -2139062144;
add.s32 %r707, %r706, -16843009;
or.b32 %r239, %r707, %r705;

prmt.b32 %r239, %r239, 0, 0xba98;

xor.b32 %r708, %r239, %r688;
or.b32 %r709, %r708, -2139062144;
and.b32 %r710, %r239, 2139062143;
and.b32 %r711, %r688, -2139062144;
sub.s32 %r712, %r709, %r710;
xor.b32 %r713, %r711, %r712;
xor.b32 %r246, %r713, -2139062144;

dp4a.s32.s32 %r241, %r242, %r96, %r233;


dp4a.s32.s32 %r245, %r246, %r100, %r241;

shl.b16 %rs59, %rs49, 2;
cvt.u64.u16 %rd117, %rs59;
and.b64 %rd118, %rd117, 1020;
add.s64 %rd119, %rd33, %rd118;
ld.global.nc.u32 %r714, [%rd119];
shr.u16 %rs60, %rs49, 8;
cvt.u32.u16 %r715, %rs60;
mul.wide.u32 %rd120, %r715, 4;
add.s64 %rd121, %rd33, %rd120;
ld.global.nc.u32 %r716, [%rd121];
shr.u32 %r717, %r629, 5;
and.b32 %r718, %r717, 255;
popc.b32 %r719, %r718;
shl.b32 %r720, %r719, 7;
and.b32 %r721, %r720, 128;
xor.b32 %r722, %r721, %r718;
mul.lo.s32 %r723, %r722, 16843009;
and.b32 %r724, %r723, 134480385;
or.b32 %r725, %r724, -2139062144;
add.s32 %r726, %r725, -16843009;
or.b32 %r249, %r726, %r724;

prmt.b32 %r249, %r249, 0, 0xba98;

xor.b32 %r727, %r249, %r714;
or.b32 %r728, %r727, -2139062144;
and.b32 %r729, %r249, 2139062143;
and.b32 %r730, %r714, -2139062144;
sub.s32 %r731, %r728, %r729;
xor.b32 %r732, %r730, %r731;
xor.b32 %r254, %r732, -2139062144;
and.b32 %r733, %r723, -2143281136;
or.b32 %r734, %r733, -2139062144;
add.s32 %r735, %r734, -16843009;
or.b32 %r251, %r735, %r733;

prmt.b32 %r251, %r251, 0, 0xba98;

xor.b32 %r736, %r251, %r716;
or.b32 %r737, %r736, -2139062144;
and.b32 %r738, %r251, 2139062143;
and.b32 %r739, %r716, -2139062144;
sub.s32 %r740, %r737, %r738;
xor.b32 %r741, %r739, %r740;
xor.b32 %r258, %r741, -2139062144;

dp4a.s32.s32 %r253, %r254, %r108, %r245;

shr.u32 %r742, %r629, 12;

dp4a.s32.s32 %r257, %r258, %r112, %r253;

shr.u32 %r743, %r257, 31;
add.s32 %r744, %r257, %r743;
shr.s32 %r745, %r744, 1;
mad.lo.s32 %r746, %r257, %r742, %r745;
shr.u32 %r747, %r746, 31;
add.s32 %r748, %r746, %r747;
shr.s32 %r749, %r748, 1;
add.s64 %rd122, %rd99, %rd5;
ld.global.nc.u16 %rs4, [%rd122];

{ cvt.f32.f16 %f31, %rs4;}



{.reg .f16 low,high;
mov.b32 {low,high},%r114;
cvt.f32.f16 %f32, low;}


mul.ftz.f32 %f39, %f31, %f32;
cvt.rn.f32.s32 %f40, %r749;
fma.rn.ftz.f32 %f105, %f39, %f40, %f105;
add.s32 %r833, %r833, 128;
add.s64 %rd125, %rd125, 1568;
add.s32 %r834, %r834, 16;
setp.lt.s32 %p3, %r834, %r5;
@%p3 bra $L__BB240_5;

$L__BB240_6:
setp.eq.s32 %p4, %r1, 0;
@%p4 bra $L__BB240_8;

shl.b32 %r750, %r1, 9;
mov.u32 %r751, _ZZ13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r752, %r751, %r750;
shl.b32 %r753, %r2, 2;
add.s32 %r754, %r752, %r753;
st.shared.f32 [%r754+-512], %f108;
st.shared.f32 [%r754+-384], %f107;
st.shared.f32 [%r754+-256], %f106;
st.shared.f32 [%r754+-128], %f105;

$L__BB240_8:
setp.ne.s32 %p5, %r1, 0;
bar.sync 0;
@%p5 bra $L__BB240_17;

mov.u32 %r755, %ctaid.y;
mad.lo.s32 %r756, %r755, %r31, %r4;
mad.lo.s32 %r757, %r12, %r37, %r756;
shl.b32 %r758, %r2, 2;
mov.u32 %r759, _ZZ13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r19, %r759, %r758;
mov.u32 %r760, 2;
ld.shared.f32 %f41, [%r19];
add.ftz.f32 %f42, %f41, %f108;
ld.shared.f32 %f43, [%r19+512];
add.ftz.f32 %f44, %f43, %f42;
ld.shared.f32 %f45, [%r19+1024];
add.ftz.f32 %f46, %f45, %f44;
mov.b32 %r761, %f46;
mov.u32 %r762, 31;
mov.u32 %r763, 16;
mov.u32 %r764, -1;
shfl.sync.bfly.b32 %r765|%p6, %r761, %r763, %r762, %r764;
mov.b32 %f47, %r765;
add.ftz.f32 %f48, %f46, %f47;
mov.b32 %r766, %f48;
mov.u32 %r767, 8;
shfl.sync.bfly.b32 %r768|%p7, %r766, %r767, %r762, %r764;
mov.b32 %f49, %r768;
add.ftz.f32 %f50, %f48, %f49;
mov.b32 %r769, %f50;
mov.u32 %r770, 4;
shfl.sync.bfly.b32 %r771|%p8, %r769, %r770, %r762, %r764;
mov.b32 %f51, %r771;
add.ftz.f32 %f52, %f50, %f51;
mov.b32 %r772, %f52;
shfl.sync.bfly.b32 %r773|%p9, %r772, %r760, %r762, %r764;
mov.b32 %f53, %r773;
add.ftz.f32 %f54, %f52, %f53;
mov.b32 %r774, %f54;
mov.u32 %r775, 1;
shfl.sync.bfly.b32 %r776|%p10, %r774, %r775, %r762, %r764;
mov.b32 %f55, %r776;
add.ftz.f32 %f13, %f54, %f55;
setp.ge.u32 %p11, %r4, %r25;
setp.ne.s32 %p12, %r2, 0;
cvta.to.global.u64 %rd123, %rd15;
mul.wide.u32 %rd124, %r757, 4;
add.s64 %rd11, %rd123, %rd124;
or.pred %p13, %p12, %p11;
@%p13 bra $L__BB240_11;

st.global.f32 [%rd11], %f13;

$L__BB240_11:
ld.shared.f32 %f56, [%r19+128];
add.ftz.f32 %f57, %f56, %f107;
ld.shared.f32 %f58, [%r19+640];
add.ftz.f32 %f59, %f58, %f57;
ld.shared.f32 %f60, [%r19+1152];
add.ftz.f32 %f61, %f60, %f59;
mov.b32 %r777, %f61;
shfl.sync.bfly.b32 %r781|%p14, %r777, %r763, %r762, %r764;
mov.b32 %f62, %r781;
add.ftz.f32 %f63, %f61, %f62;
mov.b32 %r782, %f63;
shfl.sync.bfly.b32 %r784|%p15, %r782, %r767, %r762, %r764;
mov.b32 %f64, %r784;
add.ftz.f32 %f65, %f63, %f64;
mov.b32 %r785, %f65;
shfl.sync.bfly.b32 %r787|%p16, %r785, %r770, %r762, %r764;
mov.b32 %f66, %r787;
add.ftz.f32 %f67, %f65, %f66;
mov.b32 %r788, %f67;
shfl.sync.bfly.b32 %r790|%p17, %r788, %r760, %r762, %r764;
mov.b32 %f68, %r790;
add.ftz.f32 %f69, %f67, %f68;
mov.b32 %r791, %f69;
shfl.sync.bfly.b32 %r793|%p18, %r791, %r775, %r762, %r764;
mov.b32 %f70, %r793;
add.ftz.f32 %f14, %f69, %f70;
add.s32 %r794, %r4, 1;
setp.ge.u32 %p19, %r794, %r25;
setp.ne.s32 %p20, %r2, 1;
or.pred %p21, %p20, %p19;
@%p21 bra $L__BB240_13;

st.global.f32 [%rd11+4], %f14;

$L__BB240_13:
ld.shared.f32 %f71, [%r19+256];
add.ftz.f32 %f72, %f71, %f106;
ld.shared.f32 %f73, [%r19+768];
add.ftz.f32 %f74, %f73, %f72;
ld.shared.f32 %f75, [%r19+1280];
add.ftz.f32 %f76, %f75, %f74;
mov.b32 %r795, %f76;
mov.u32 %r796, 31;
mov.u32 %r797, 16;
mov.u32 %r798, -1;
shfl.sync.bfly.b32 %r799|%p22, %r795, %r797, %r796, %r798;
mov.b32 %f77, %r799;
add.ftz.f32 %f78, %f76, %f77;
mov.b32 %r800, %f78;
mov.u32 %r801, 8;
shfl.sync.bfly.b32 %r802|%p23, %r800, %r801, %r796, %r798;
mov.b32 %f79, %r802;
add.ftz.f32 %f80, %f78, %f79;
mov.b32 %r803, %f80;
mov.u32 %r804, 4;
shfl.sync.bfly.b32 %r805|%p24, %r803, %r804, %r796, %r798;
mov.b32 %f81, %r805;
add.ftz.f32 %f82, %f80, %f81;
mov.b32 %r806, %f82;
mov.u32 %r807, 2;
shfl.sync.bfly.b32 %r808|%p25, %r806, %r807, %r796, %r798;
mov.b32 %f83, %r808;
add.ftz.f32 %f84, %f82, %f83;
mov.b32 %r809, %f84;
mov.u32 %r810, 1;
shfl.sync.bfly.b32 %r811|%p26, %r809, %r810, %r796, %r798;
mov.b32 %f85, %r811;
add.ftz.f32 %f15, %f84, %f85;
add.s32 %r812, %r4, 2;
setp.ge.u32 %p27, %r812, %r25;
setp.ne.s32 %p28, %r2, 2;
or.pred %p29, %p28, %p27;
@%p29 bra $L__BB240_15;

st.global.f32 [%rd11+8], %f15;

$L__BB240_15:
ld.shared.f32 %f86, [%r19+384];
add.ftz.f32 %f87, %f86, %f105;
ld.shared.f32 %f88, [%r19+896];
add.ftz.f32 %f89, %f88, %f87;
ld.shared.f32 %f90, [%r19+1408];
add.ftz.f32 %f91, %f90, %f89;
mov.b32 %r813, %f91;
shfl.sync.bfly.b32 %r817|%p30, %r813, %r797, %r796, %r798;
mov.b32 %f92, %r817;
add.ftz.f32 %f93, %f91, %f92;
mov.b32 %r818, %f93;
shfl.sync.bfly.b32 %r820|%p31, %r818, %r801, %r796, %r798;
mov.b32 %f94, %r820;
add.ftz.f32 %f95, %f93, %f94;
mov.b32 %r821, %f95;
shfl.sync.bfly.b32 %r823|%p32, %r821, %r804, %r796, %r798;
mov.b32 %f96, %r823;
add.ftz.f32 %f97, %f95, %f96;
mov.b32 %r824, %f97;
shfl.sync.bfly.b32 %r826|%p33, %r824, %r807, %r796, %r798;
mov.b32 %f98, %r826;
add.ftz.f32 %f99, %f97, %f98;
mov.b32 %r827, %f99;
shfl.sync.bfly.b32 %r829|%p34, %r827, %r810, %r796, %r798;
mov.b32 %f100, %r829;
add.ftz.f32 %f16, %f99, %f100;
add.s32 %r830, %r4, 3;
setp.ge.u32 %p35, %r830, %r25;
setp.ne.s32 %p36, %r2, 3;
or.pred %p37, %p36, %p35;
@%p37 bra $L__BB240_17;

st.global.f32 [%rd11+12], %f16;

$L__BB240_17:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<34>;
.reg .b16 %rs<31>;
.reg .f32 %f<121>;
.reg .b32 %r<512>;
.reg .b64 %rd<112>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[384];

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate[384];

ld.param.u64 %rd8, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd9, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd10, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r45, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r46, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r47, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
ld.param.u64 %rd13, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+16];
ld.param.u64 %rd12, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+8];
ld.param.u64 %rd11, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3];
mov.u32 %r1, %tid.y;
shl.b32 %r48, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r48, %r2;
mov.u32 %r4, %ctaid.x;
mov.u32 %r5, %ctaid.y;
setp.eq.s64 %p2, %rd10, 0;
@%p2 bra $L__BB241_2;

cvta.to.global.u64 %rd17, %rd10;
mul.wide.u32 %rd18, %r5, 4;
add.s64 %rd19, %rd17, %rd18;
ld.global.nc.u32 %r508, [%rd19];
mul.hi.u32 %r49, %r5, %r32;
add.s32 %r50, %r49, %r5;
shr.u32 %r51, %r50, %r33;
mul.lo.s32 %r52, %r51, %r34;
sub.s32 %r509, %r5, %r52;
bra.uni $L__BB241_3;

$L__BB241_2:
mul.hi.u32 %r53, %r5, %r36;
add.s32 %r54, %r53, %r5;
shr.u32 %r508, %r54, %r37;
mov.u32 %r509, %r5;

$L__BB241_3:
or.b32 %r55, %r2, %r1;
setp.ne.s32 %p3, %r55, 0;
mov.f32 %f114, 0f00000000;
mov.f32 %f110, %f114;
mov.f32 %f111, %f114;
@%p3 bra $L__BB241_8;

setp.eq.s64 %p4, %rd11, 0;
mov.f32 %f110, 0f00000000;
mov.f32 %f111, %f110;
@%p4 bra $L__BB241_6;

cvta.to.global.u64 %rd20, %rd11;
mov.u32 %r56, %ctaid.z;
mul.lo.s32 %r57, %r56, %r47;
cvt.u64.u32 %rd21, %r57;
selp.b32 %r58, %r5, %r508, %p2;
mul.lo.s32 %r59, %r58, %r41;
cvt.u64.u32 %rd22, %r59;
cvt.u64.u32 %rd23, %r2;
cvt.s64.s32 %rd24, %r4;
add.s64 %rd25, %rd23, %rd24;
add.s64 %rd26, %rd25, %rd21;
add.s64 %rd27, %rd26, %rd22;
shl.b64 %rd28, %rd27, 2;
add.s64 %rd29, %rd20, %rd28;
ld.global.f32 %f111, [%rd29];

$L__BB241_6:
setp.eq.s64 %p6, %rd13, 0;
setp.eq.s64 %p7, %rd12, 0;
or.pred %p8, %p6, %p7;
@%p8 bra $L__BB241_8;

cvta.to.global.u64 %rd30, %rd13;
mov.u32 %r61, %ctaid.z;
mul.lo.s32 %r62, %r61, %r47;
cvt.u64.u32 %rd31, %r62;
selp.b32 %r63, %r5, %r508, %p2;
mul.lo.s32 %r64, %r63, %r41;
cvt.u64.u32 %rd32, %r64;
cvt.u64.u32 %rd33, %r2;
cvt.s64.s32 %rd34, %r4;
add.s64 %rd35, %rd33, %rd34;
add.s64 %rd36, %rd35, %rd31;
add.s64 %rd37, %rd36, %rd32;
shl.b64 %rd38, %rd37, 2;
add.s64 %rd39, %rd30, %rd38;
ld.global.f32 %f110, [%rd39];

$L__BB241_8:
shr.u32 %r66, %r31, 8;
shr.s32 %r67, %r3, 31;
shr.u32 %r68, %r67, 29;
add.s32 %r69, %r3, %r68;
shr.s32 %r511, %r69, 3;
setp.ge.s32 %p10, %r511, %r66;
mov.f32 %f116, %f114;
@%p10 bra $L__BB241_13;

mov.u32 %r70, %ctaid.z;
and.b32 %r74, %r69, -8;
sub.s32 %r12, %r3, %r74;
shl.b32 %r510, %r511, 3;
mul.hi.u32 %r75, %r70, %r42;
add.s32 %r76, %r75, %r70;
shr.u32 %r77, %r76, %r43;
mad.lo.s32 %r78, %r77, %r45, %r511;
mad.lo.s32 %r79, %r508, %r39, %r78;
mad.lo.s32 %r80, %r4, %r35, %r79;
mul.wide.s32 %rd111, %r80, 98;
mul.lo.s32 %r81, %r509, %r40;
cvt.u64.u32 %rd40, %r81;
mul.lo.s32 %r82, %r70, %r46;
cvt.u64.u32 %rd41, %r82;
add.s64 %rd5, %rd40, %rd41;
mov.f32 %f116, 0f00000000;
cvta.to.global.u64 %rd42, %rd9;
cvta.to.global.u64 %rd46, %rd8;
cvt.s64.s32 %rd52, %r12;
mov.u64 %rd57, iq3xxs_grid;
mov.f32 %f114, %f116;

$L__BB241_10:
cvt.u64.u32 %rd43, %r510;
add.s64 %rd44, %rd5, %rd43;
shl.b32 %r132, %r12, 2;
mul.wide.s32 %rd45, %r132, 2;
add.s64 %rd47, %rd46, %rd45;
add.s64 %rd48, %rd47, %rd111;
ld.global.nc.u16 %rs2, [%rd48+4];
ld.global.nc.u16 %rs3, [%rd48+6];
ld.global.nc.u16 %rs4, [%rd48+8];
shl.b32 %r133, %r12, 1;
add.s32 %r134, %r133, 32;
mul.wide.s32 %rd49, %r134, 2;
add.s64 %rd50, %rd46, %rd49;
add.s64 %rd51, %rd50, %rd111;
ld.global.nc.u16 %rs5, [%rd51+2];
cvt.u32.u16 %r135, %rs5;
ld.global.nc.u16 %rs6, [%rd51+4];
cvt.u32.u16 %r136, %rs6;
mov.b32 %r137, {%rs5, %rs6};
add.s64 %rd53, %rd44, %rd52;
mul.lo.s64 %rd54, %rd53, 36;
ld.global.nc.u16 %rs7, [%rd48+2];
shl.b16 %rs8, %rs7, 2;
cvt.u64.u16 %rd55, %rs8;
and.b64 %rd56, %rd55, 1020;
add.s64 %rd58, %rd57, %rd56;
ld.global.nc.u32 %r138, [%rd58];
shr.u16 %rs9, %rs7, 8;
cvt.u32.u16 %r139, %rs9;
mul.wide.u32 %rd59, %r139, 4;
add.s64 %rd60, %rd57, %rd59;
ld.global.nc.u32 %r140, [%rd60];
and.b32 %r141, %r135, 255;
popc.b32 %r142, %r141;
shl.b32 %r143, %r142, 7;
and.b32 %r144, %r143, 128;
xor.b32 %r145, %r144, %r141;
mul.lo.s32 %r146, %r145, 16843009;
and.b32 %r147, %r146, 134480385;
or.b32 %r148, %r147, -2139062144;
add.s32 %r149, %r148, -16843009;
or.b32 %r83, %r149, %r147;

prmt.b32 %r83, %r83, 0, 0xba98;

xor.b32 %r150, %r83, %r138;
or.b32 %r151, %r150, -2139062144;
and.b32 %r152, %r83, 2139062143;
and.b32 %r153, %r138, -2139062144;
sub.s32 %r154, %r151, %r152;
xor.b32 %r155, %r153, %r154;
xor.b32 %r88, %r155, -2139062144;
add.s64 %rd61, %rd42, %rd54;
ld.global.nc.u32 %r16, [%rd61+4];
and.b32 %r156, %r146, -2143281136;
or.b32 %r157, %r156, -2139062144;
add.s32 %r158, %r157, -16843009;
or.b32 %r85, %r158, %r156;

prmt.b32 %r85, %r85, 0, 0xba98;

xor.b32 %r159, %r85, %r140;
or.b32 %r160, %r159, -2139062144;
and.b32 %r161, %r85, 2139062143;
and.b32 %r162, %r140, -2139062144;
sub.s32 %r163, %r160, %r161;
xor.b32 %r164, %r162, %r163;
xor.b32 %r92, %r164, -2139062144;
ld.global.nc.u32 %r17, [%rd61+8];
mov.u32 %r90, 0;

dp4a.s32.s32 %r87, %r88, %r16, %r90;


dp4a.s32.s32 %r91, %r92, %r17, %r87;

shl.b16 %rs10, %rs2, 2;
cvt.u64.u16 %rd62, %rs10;
and.b64 %rd63, %rd62, 1020;
add.s64 %rd64, %rd57, %rd63;
ld.global.nc.u32 %r165, [%rd64];
shr.u16 %rs11, %rs2, 8;
cvt.u32.u16 %r166, %rs11;
mul.wide.u32 %rd65, %r166, 4;
add.s64 %rd66, %rd57, %rd65;
ld.global.nc.u32 %r167, [%rd66];
shr.u32 %r168, %r135, 7;
and.b32 %r169, %r168, 255;
popc.b32 %r170, %r169;
shl.b32 %r171, %r170, 7;
and.b32 %r172, %r171, 128;
xor.b32 %r173, %r172, %r169;
mul.lo.s32 %r174, %r173, 16843009;
and.b32 %r175, %r174, 134480385;
or.b32 %r176, %r175, -2139062144;
add.s32 %r177, %r176, -16843009;
or.b32 %r95, %r177, %r175;

prmt.b32 %r95, %r95, 0, 0xba98;

xor.b32 %r178, %r95, %r165;
or.b32 %r179, %r178, -2139062144;
and.b32 %r180, %r95, 2139062143;
and.b32 %r181, %r165, -2139062144;
sub.s32 %r182, %r179, %r180;
xor.b32 %r183, %r181, %r182;
xor.b32 %r100, %r183, -2139062144;
ld.global.nc.u32 %r18, [%rd61+12];
and.b32 %r184, %r174, -2143281136;
or.b32 %r185, %r184, -2139062144;
add.s32 %r186, %r185, -16843009;
or.b32 %r97, %r186, %r184;

prmt.b32 %r97, %r97, 0, 0xba98;

xor.b32 %r187, %r97, %r167;
or.b32 %r188, %r187, -2139062144;
and.b32 %r189, %r97, 2139062143;
and.b32 %r190, %r167, -2139062144;
sub.s32 %r191, %r188, %r189;
xor.b32 %r192, %r190, %r191;
xor.b32 %r104, %r192, -2139062144;
ld.global.nc.u32 %r19, [%rd61+16];

dp4a.s32.s32 %r99, %r100, %r18, %r91;


dp4a.s32.s32 %r103, %r104, %r19, %r99;

shl.b16 %rs12, %rs3, 2;
cvt.u64.u16 %rd67, %rs12;
and.b64 %rd68, %rd67, 1020;
add.s64 %rd69, %rd57, %rd68;
ld.global.nc.u32 %r193, [%rd69];
shr.u16 %rs13, %rs3, 8;
cvt.u32.u16 %r194, %rs13;
mul.wide.u32 %rd70, %r194, 4;
add.s64 %rd71, %rd57, %rd70;
ld.global.nc.u32 %r195, [%rd71];
shr.u32 %r196, %r137, 14;
and.b32 %r197, %r196, 255;
popc.b32 %r198, %r197;
shl.b32 %r199, %r198, 7;
and.b32 %r200, %r199, 128;
xor.b32 %r201, %r200, %r197;
mul.lo.s32 %r202, %r201, 16843009;
and.b32 %r203, %r202, 134480385;
or.b32 %r204, %r203, -2139062144;
add.s32 %r205, %r204, -16843009;
or.b32 %r107, %r205, %r203;

prmt.b32 %r107, %r107, 0, 0xba98;

xor.b32 %r206, %r107, %r193;
or.b32 %r207, %r206, -2139062144;
and.b32 %r208, %r107, 2139062143;
and.b32 %r209, %r193, -2139062144;
sub.s32 %r210, %r207, %r208;
xor.b32 %r211, %r209, %r210;
xor.b32 %r112, %r211, -2139062144;
ld.global.nc.u32 %r20, [%rd61+20];
and.b32 %r212, %r202, -2143281136;
or.b32 %r213, %r212, -2139062144;
add.s32 %r214, %r213, -16843009;
or.b32 %r109, %r214, %r212;

prmt.b32 %r109, %r109, 0, 0xba98;

xor.b32 %r215, %r109, %r195;
or.b32 %r216, %r215, -2139062144;
and.b32 %r217, %r109, 2139062143;
and.b32 %r218, %r195, -2139062144;
sub.s32 %r219, %r216, %r217;
xor.b32 %r220, %r218, %r219;
xor.b32 %r116, %r220, -2139062144;
ld.global.nc.u32 %r21, [%rd61+24];

dp4a.s32.s32 %r111, %r112, %r20, %r103;


dp4a.s32.s32 %r115, %r116, %r21, %r111;

shl.b16 %rs14, %rs4, 2;
cvt.u64.u16 %rd72, %rs14;
and.b64 %rd73, %rd72, 1020;
add.s64 %rd74, %rd57, %rd73;
ld.global.nc.u32 %r221, [%rd74];
shr.u16 %rs15, %rs4, 8;
cvt.u32.u16 %r222, %rs15;
mul.wide.u32 %rd75, %r222, 4;
add.s64 %rd76, %rd57, %rd75;
ld.global.nc.u32 %r223, [%rd76];
shr.u32 %r224, %r136, 5;
and.b32 %r225, %r224, 255;
popc.b32 %r226, %r225;
shl.b32 %r227, %r226, 7;
and.b32 %r228, %r227, 128;
xor.b32 %r229, %r228, %r225;
mul.lo.s32 %r230, %r229, 16843009;
and.b32 %r231, %r230, 134480385;
or.b32 %r232, %r231, -2139062144;
add.s32 %r233, %r232, -16843009;
or.b32 %r119, %r233, %r231;

prmt.b32 %r119, %r119, 0, 0xba98;

xor.b32 %r234, %r119, %r221;
or.b32 %r235, %r234, -2139062144;
and.b32 %r236, %r119, 2139062143;
and.b32 %r237, %r221, -2139062144;
sub.s32 %r238, %r235, %r236;
xor.b32 %r239, %r237, %r238;
xor.b32 %r124, %r239, -2139062144;
ld.global.nc.u32 %r22, [%rd61+28];
and.b32 %r240, %r230, -2143281136;
or.b32 %r241, %r240, -2139062144;
add.s32 %r242, %r241, -16843009;
or.b32 %r121, %r242, %r240;

prmt.b32 %r121, %r121, 0, 0xba98;

xor.b32 %r243, %r121, %r223;
or.b32 %r244, %r243, -2139062144;
and.b32 %r245, %r121, 2139062143;
and.b32 %r246, %r223, -2139062144;
sub.s32 %r247, %r244, %r245;
xor.b32 %r248, %r246, %r247;
xor.b32 %r128, %r248, -2139062144;
ld.global.nc.u32 %r23, [%rd61+32];

dp4a.s32.s32 %r123, %r124, %r22, %r115;

shr.u32 %r249, %r136, 12;

dp4a.s32.s32 %r127, %r128, %r23, %r123;

shr.u32 %r250, %r127, 31;
add.s32 %r251, %r127, %r250;
shr.s32 %r252, %r251, 1;
mad.lo.s32 %r253, %r127, %r249, %r252;
shr.u32 %r254, %r253, 31;
add.s32 %r255, %r253, %r254;
shr.s32 %r256, %r255, 1;
add.s64 %rd77, %rd46, %rd111;
ld.global.nc.u16 %rs1, [%rd77];

{ cvt.f32.f16 %f40, %rs1;}


ld.global.nc.u32 %r24, [%rd61];

{.reg .f16 low,high;
mov.b32 {low,high},%r24;
cvt.f32.f16 %f41, low;}


mul.ftz.f32 %f42, %f40, %f41;
cvt.rn.f32.s32 %f43, %r256;
fma.rn.ftz.f32 %f116, %f42, %f43, %f116;
setp.eq.s64 %p11, %rd12, 0;
@%p11 bra $L__BB241_12;

cvta.to.global.u64 %rd79, %rd12;
add.s64 %rd80, %rd79, %rd45;
add.s64 %rd81, %rd80, %rd111;
ld.global.u16 %rs17, [%rd81+4];
ld.global.u16 %rs18, [%rd81+6];
ld.global.u16 %rs19, [%rd81+8];
add.s64 %rd83, %rd79, %rd49;
add.s64 %rd84, %rd83, %rd111;
ld.global.u16 %rs20, [%rd84+2];
cvt.u32.u16 %r309, %rs20;
ld.global.u16 %rs21, [%rd84+4];
cvt.u32.u16 %r310, %rs21;
mov.b32 %r311, {%rs20, %rs21};
ld.global.u16 %rs22, [%rd81+2];
shl.b16 %rs23, %rs22, 2;
cvt.u64.u16 %rd85, %rs23;
and.b64 %rd86, %rd85, 1020;
add.s64 %rd88, %rd57, %rd86;
ld.global.nc.u32 %r312, [%rd88];
shr.u16 %rs24, %rs22, 8;
cvt.u32.u16 %r313, %rs24;
mul.wide.u32 %rd89, %r313, 4;
add.s64 %rd90, %rd57, %rd89;
ld.global.nc.u32 %r314, [%rd90];
and.b32 %r315, %r309, 255;
popc.b32 %r316, %r315;
shl.b32 %r317, %r316, 7;
and.b32 %r318, %r317, 128;
xor.b32 %r319, %r318, %r315;
mul.lo.s32 %r320, %r319, 16843009;
and.b32 %r321, %r320, 134480385;
or.b32 %r322, %r321, -2139062144;
add.s32 %r323, %r322, -16843009;
or.b32 %r257, %r323, %r321;

prmt.b32 %r257, %r257, 0, 0xba98;

xor.b32 %r324, %r257, %r312;
or.b32 %r325, %r324, -2139062144;
and.b32 %r326, %r257, 2139062143;
and.b32 %r327, %r312, -2139062144;
sub.s32 %r328, %r325, %r326;
xor.b32 %r329, %r327, %r328;
xor.b32 %r262, %r329, -2139062144;
and.b32 %r330, %r320, -2143281136;
or.b32 %r331, %r330, -2139062144;
add.s32 %r332, %r331, -16843009;
or.b32 %r259, %r332, %r330;

prmt.b32 %r259, %r259, 0, 0xba98;

xor.b32 %r333, %r259, %r314;
or.b32 %r334, %r333, -2139062144;
and.b32 %r335, %r259, 2139062143;
and.b32 %r336, %r314, -2139062144;
sub.s32 %r337, %r334, %r335;
xor.b32 %r338, %r336, %r337;
xor.b32 %r266, %r338, -2139062144;

dp4a.s32.s32 %r261, %r262, %r16, %r90;


dp4a.s32.s32 %r265, %r266, %r17, %r261;

shl.b16 %rs25, %rs17, 2;
cvt.u64.u16 %rd91, %rs25;
and.b64 %rd92, %rd91, 1020;
add.s64 %rd93, %rd57, %rd92;
ld.global.nc.u32 %r339, [%rd93];
shr.u16 %rs26, %rs17, 8;
cvt.u32.u16 %r340, %rs26;
mul.wide.u32 %rd94, %r340, 4;
add.s64 %rd95, %rd57, %rd94;
ld.global.nc.u32 %r341, [%rd95];
shr.u32 %r342, %r309, 7;
and.b32 %r343, %r342, 255;
popc.b32 %r344, %r343;
shl.b32 %r345, %r344, 7;
and.b32 %r346, %r345, 128;
xor.b32 %r347, %r346, %r343;
mul.lo.s32 %r348, %r347, 16843009;
and.b32 %r349, %r348, 134480385;
or.b32 %r350, %r349, -2139062144;
add.s32 %r351, %r350, -16843009;
or.b32 %r269, %r351, %r349;

prmt.b32 %r269, %r269, 0, 0xba98;

xor.b32 %r352, %r269, %r339;
or.b32 %r353, %r352, -2139062144;
and.b32 %r354, %r269, 2139062143;
and.b32 %r355, %r339, -2139062144;
sub.s32 %r356, %r353, %r354;
xor.b32 %r357, %r355, %r356;
xor.b32 %r274, %r357, -2139062144;
and.b32 %r358, %r348, -2143281136;
or.b32 %r359, %r358, -2139062144;
add.s32 %r360, %r359, -16843009;
or.b32 %r271, %r360, %r358;

prmt.b32 %r271, %r271, 0, 0xba98;

xor.b32 %r361, %r271, %r341;
or.b32 %r362, %r361, -2139062144;
and.b32 %r363, %r271, 2139062143;
and.b32 %r364, %r341, -2139062144;
sub.s32 %r365, %r362, %r363;
xor.b32 %r366, %r364, %r365;
xor.b32 %r278, %r366, -2139062144;

dp4a.s32.s32 %r273, %r274, %r18, %r265;


dp4a.s32.s32 %r277, %r278, %r19, %r273;

shl.b16 %rs27, %rs18, 2;
cvt.u64.u16 %rd96, %rs27;
and.b64 %rd97, %rd96, 1020;
add.s64 %rd98, %rd57, %rd97;
ld.global.nc.u32 %r367, [%rd98];
shr.u16 %rs28, %rs18, 8;
cvt.u32.u16 %r368, %rs28;
mul.wide.u32 %rd99, %r368, 4;
add.s64 %rd100, %rd57, %rd99;
ld.global.nc.u32 %r369, [%rd100];
shr.u32 %r370, %r311, 14;
and.b32 %r371, %r370, 255;
popc.b32 %r372, %r371;
shl.b32 %r373, %r372, 7;
and.b32 %r374, %r373, 128;
xor.b32 %r375, %r374, %r371;
mul.lo.s32 %r376, %r375, 16843009;
and.b32 %r377, %r376, 134480385;
or.b32 %r378, %r377, -2139062144;
add.s32 %r379, %r378, -16843009;
or.b32 %r281, %r379, %r377;

prmt.b32 %r281, %r281, 0, 0xba98;

xor.b32 %r380, %r281, %r367;
or.b32 %r381, %r380, -2139062144;
and.b32 %r382, %r281, 2139062143;
and.b32 %r383, %r367, -2139062144;
sub.s32 %r384, %r381, %r382;
xor.b32 %r385, %r383, %r384;
xor.b32 %r286, %r385, -2139062144;
and.b32 %r386, %r376, -2143281136;
or.b32 %r387, %r386, -2139062144;
add.s32 %r388, %r387, -16843009;
or.b32 %r283, %r388, %r386;

prmt.b32 %r283, %r283, 0, 0xba98;

xor.b32 %r389, %r283, %r369;
or.b32 %r390, %r389, -2139062144;
and.b32 %r391, %r283, 2139062143;
and.b32 %r392, %r369, -2139062144;
sub.s32 %r393, %r390, %r391;
xor.b32 %r394, %r392, %r393;
xor.b32 %r290, %r394, -2139062144;

dp4a.s32.s32 %r285, %r286, %r20, %r277;


dp4a.s32.s32 %r289, %r290, %r21, %r285;

shl.b16 %rs29, %rs19, 2;
cvt.u64.u16 %rd101, %rs29;
and.b64 %rd102, %rd101, 1020;
add.s64 %rd103, %rd57, %rd102;
ld.global.nc.u32 %r395, [%rd103];
shr.u16 %rs30, %rs19, 8;
cvt.u32.u16 %r396, %rs30;
mul.wide.u32 %rd104, %r396, 4;
add.s64 %rd105, %rd57, %rd104;
ld.global.nc.u32 %r397, [%rd105];
shr.u32 %r398, %r310, 5;
and.b32 %r399, %r398, 255;
popc.b32 %r400, %r399;
shl.b32 %r401, %r400, 7;
and.b32 %r402, %r401, 128;
xor.b32 %r403, %r402, %r399;
mul.lo.s32 %r404, %r403, 16843009;
and.b32 %r405, %r404, 134480385;
or.b32 %r406, %r405, -2139062144;
add.s32 %r407, %r406, -16843009;
or.b32 %r293, %r407, %r405;

prmt.b32 %r293, %r293, 0, 0xba98;

xor.b32 %r408, %r293, %r395;
or.b32 %r409, %r408, -2139062144;
and.b32 %r410, %r293, 2139062143;
and.b32 %r411, %r395, -2139062144;
sub.s32 %r412, %r409, %r410;
xor.b32 %r413, %r411, %r412;
xor.b32 %r298, %r413, -2139062144;
and.b32 %r414, %r404, -2143281136;
or.b32 %r415, %r414, -2139062144;
add.s32 %r416, %r415, -16843009;
or.b32 %r295, %r416, %r414;

prmt.b32 %r295, %r295, 0, 0xba98;

xor.b32 %r417, %r295, %r397;
or.b32 %r418, %r417, -2139062144;
and.b32 %r419, %r295, 2139062143;
and.b32 %r420, %r397, -2139062144;
sub.s32 %r421, %r418, %r419;
xor.b32 %r422, %r420, %r421;
xor.b32 %r302, %r422, -2139062144;

dp4a.s32.s32 %r297, %r298, %r22, %r289;

shr.u32 %r423, %r310, 12;

dp4a.s32.s32 %r301, %r302, %r23, %r297;

shr.u32 %r424, %r301, 31;
add.s32 %r425, %r301, %r424;
shr.s32 %r426, %r425, 1;
mad.lo.s32 %r427, %r301, %r423, %r426;
shr.u32 %r428, %r427, 31;
add.s32 %r429, %r427, %r428;
shr.s32 %r430, %r429, 1;
add.s64 %rd106, %rd79, %rd111;
ld.global.u16 %rs16, [%rd106];

{ cvt.f32.f16 %f44, %rs16;}



{.reg .f16 low,high;
mov.b32 {low,high},%r24;
cvt.f32.f16 %f45, low;}


mul.ftz.f32 %f46, %f44, %f45;
cvt.rn.f32.s32 %f47, %r430;
fma.rn.ftz.f32 %f114, %f46, %f47, %f114;

$L__BB241_12:
add.s32 %r510, %r510, 128;
add.s64 %rd111, %rd111, 1568;
add.s32 %r511, %r511, 16;
setp.lt.s32 %p12, %r511, %r66;
@%p12 bra $L__BB241_10;

$L__BB241_13:
mov.u32 %r494, %tid.y;
setp.eq.s32 %p13, %r494, 0;
@%p13 bra $L__BB241_16;

mov.u32 %r496, %tid.y;
mov.u32 %r495, %tid.x;
setp.eq.s64 %p14, %rd12, 0;
add.s32 %r27, %r496, -1;
shl.b32 %r435, %r496, 7;
mov.u32 %r436, _ZZ13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r437, %r436, %r435;
shl.b32 %r438, %r495, 2;
add.s32 %r439, %r437, %r438;
st.shared.f32 [%r439+-128], %f116;
@%p14 bra $L__BB241_16;

shl.b32 %r441, %r27, 7;
mov.u32 %r442, _ZZ13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r443, %r442, %r441;
add.s32 %r445, %r443, %r438;
st.shared.f32 [%r445], %f114;

$L__BB241_16:
mov.u32 %r497, %tid.y;
setp.ne.s32 %p15, %r497, 0;
bar.sync 0;
@%p15 bra $L__BB241_35;

mov.u32 %r498, %tid.x;
setp.ne.s64 %p16, %rd12, 0;
shl.b32 %r448, %r498, 2;
mov.u32 %r449, _ZZ13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r28, %r449, %r448;
ld.shared.f32 %f48, [%r28];
add.ftz.f32 %f13, %f116, %f48;
@%p16 bra $L__BB241_19;
bra.uni $L__BB241_18;

$L__BB241_19:
mov.u32 %r452, _ZZ13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r453, %r452, %r448;
ld.shared.f32 %f52, [%r453];
add.ftz.f32 %f53, %f114, %f52;
ld.shared.f32 %f54, [%r28+128];
add.ftz.f32 %f55, %f13, %f54;
ld.shared.f32 %f56, [%r453+128];
add.ftz.f32 %f57, %f53, %f56;
ld.shared.f32 %f58, [%r28+256];
add.ftz.f32 %f118, %f55, %f58;
ld.shared.f32 %f59, [%r453+256];
add.ftz.f32 %f114, %f57, %f59;
bra.uni $L__BB241_20;

$L__BB241_18:
ld.shared.f32 %f49, [%r28+128];
add.ftz.f32 %f50, %f13, %f49;
ld.shared.f32 %f51, [%r28+256];
add.ftz.f32 %f118, %f50, %f51;

$L__BB241_20:
mov.b32 %r454, %f118;
mov.u32 %r455, 31;
mov.u32 %r456, 16;
mov.u32 %r457, -1;
shfl.sync.bfly.b32 %r458|%p17, %r454, %r456, %r455, %r457;
mov.b32 %f60, %r458;
add.ftz.f32 %f61, %f118, %f60;
mov.b32 %r459, %f61;
mov.u32 %r460, 8;
shfl.sync.bfly.b32 %r461|%p18, %r459, %r460, %r455, %r457;
mov.b32 %f62, %r461;
add.ftz.f32 %f63, %f61, %f62;
mov.b32 %r462, %f63;
mov.u32 %r463, 4;
shfl.sync.bfly.b32 %r464|%p19, %r462, %r463, %r455, %r457;
mov.b32 %f64, %r464;
add.ftz.f32 %f65, %f63, %f64;
mov.b32 %r465, %f65;
mov.u32 %r466, 2;
shfl.sync.bfly.b32 %r467|%p20, %r465, %r466, %r455, %r457;
mov.b32 %f66, %r467;
add.ftz.f32 %f19, %f65, %f66;
mov.b32 %r468, %f19;
mov.u32 %r469, 1;
shfl.sync.bfly.b32 %r29|%p1, %r468, %r469, %r455, %r457;
setp.eq.s64 %p21, %rd12, 0;
@%p21 bra $L__BB241_22;

mov.b32 %r470, %f114;
shfl.sync.bfly.b32 %r474|%p22, %r470, %r456, %r455, %r457;
mov.b32 %f67, %r474;
add.ftz.f32 %f68, %f114, %f67;
mov.b32 %r475, %f68;
shfl.sync.bfly.b32 %r477|%p23, %r475, %r460, %r455, %r457;
mov.b32 %f69, %r477;
add.ftz.f32 %f70, %f68, %f69;
mov.b32 %r478, %f70;
shfl.sync.bfly.b32 %r480|%p24, %r478, %r463, %r455, %r457;
mov.b32 %f71, %r480;
add.ftz.f32 %f72, %f70, %f71;
mov.b32 %r481, %f72;
shfl.sync.bfly.b32 %r483|%p25, %r481, %r466, %r455, %r457;
mov.b32 %f73, %r483;
add.ftz.f32 %f74, %f72, %f73;
mov.b32 %r484, %f74;
shfl.sync.bfly.b32 %r486|%p26, %r484, %r469, %r455, %r457;
mov.b32 %f75, %r486;
add.ftz.f32 %f114, %f74, %f75;

$L__BB241_22:
mov.u32 %r499, %tid.x;
setp.ne.s32 %p27, %r499, 0;
mov.b32 %f76, %r29;
add.ftz.f32 %f22, %f19, %f76;
@%p27 bra $L__BB241_35;

add.ftz.f32 %f120, %f111, %f22;
@%p21 bra $L__BB241_34;

ld.param.u32 %r503, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
add.ftz.f32 %f24, %f110, %f114;
setp.gt.s32 %p29, %r503, 2;
@%p29 bra $L__BB241_28;

ld.param.u32 %r506, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p32, %r506, 1;
@%p32 bra $L__BB241_32;

ld.param.u32 %r507, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p33, %r507, 2;
@%p33 bra $L__BB241_27;
bra.uni $L__BB241_33;

$L__BB241_27:
mul.ftz.f32 %f96, %f24, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f97, %f96;
add.ftz.f32 %f98, %f97, 0f3F800000;
div.approx.ftz.f32 %f99, %f24, %f98;
mul.ftz.f32 %f120, %f120, %f99;
bra.uni $L__BB241_34;

$L__BB241_28:
ld.param.u32 %r504, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p30, %r504, 3;
@%p30 bra $L__BB241_31;

ld.param.u32 %r505, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.ne.s32 %p31, %r505, 6;
@%p31 bra $L__BB241_33;

ld.param.f32 %f108, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
min.ftz.f32 %f77, %f24, %f108;
neg.ftz.f32 %f78, %f108;
min.ftz.f32 %f79, %f120, %f108;
max.ftz.f32 %f80, %f79, %f78;
mul.ftz.f32 %f81, %f77, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f82, %f81;
add.ftz.f32 %f83, %f82, 0f3F800000;
div.approx.ftz.f32 %f84, %f77, %f83;
mul.ftz.f32 %f120, %f80, %f84;
bra.uni $L__BB241_34;

$L__BB241_32:
mul.ftz.f32 %f100, %f24, 0f3F000000;
mul.ftz.f32 %f101, %f24, 0f3D372713;
fma.rn.ftz.f32 %f102, %f24, %f101, 0f3F800000;
mul.ftz.f32 %f103, %f24, 0f3F4C422A;
mul.ftz.f32 %f104, %f103, %f102;
tanh.approx.f32 %f105, %f104;
add.ftz.f32 %f106, %f105, 0f3F800000;
mul.ftz.f32 %f107, %f100, %f106;
mul.ftz.f32 %f120, %f120, %f107;
bra.uni $L__BB241_34;

$L__BB241_31:
mov.f32 %f85, 0f40E00000;
min.ftz.f32 %f86, %f24, %f85;
min.ftz.f32 %f87, %f120, %f85;
mov.f32 %f88, 0fC0E00000;
max.ftz.f32 %f89, %f87, %f88;
mul.ftz.f32 %f90, %f86, 0f3FD9DB23;
mul.ftz.f32 %f91, %f90, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f92, %f91;
add.ftz.f32 %f93, %f92, 0f3F800000;
div.approx.ftz.f32 %f94, %f86, %f93;
add.ftz.f32 %f95, %f89, 0f3F800000;
mul.ftz.f32 %f120, %f95, %f94;
bra.uni $L__BB241_34;

$L__BB241_33:
mul.ftz.f32 %f120, %f120, %f24;

$L__BB241_34:
ld.param.u64 %rd110, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
mov.u32 %r502, %ctaid.x;
ld.param.u32 %r501, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r500, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
cvta.to.global.u64 %rd107, %rd110;
mov.u32 %r489, %ctaid.z;
mov.u32 %r490, %ctaid.y;
mad.lo.s32 %r492, %r490, %r501, %r502;
mad.lo.s32 %r493, %r489, %r500, %r492;
mul.wide.u32 %rd108, %r493, 4;
add.s64 %rd109, %rd107, %rd108;
st.global.f32 [%rd109], %f120;

$L__BB241_35:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<12>;
.reg .b16 %rs<16>;
.reg .f32 %f<28>;
.reg .b32 %r<267>;
.reg .b64 %rd<54>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[384];

ld.param.u64 %rd9, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd10, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd11, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd12, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r19, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
mov.u32 %r1, %tid.y;
shl.b32 %r36, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r36, %r2;
mov.u32 %r264, %ctaid.y;
setp.eq.s64 %p1, %rd11, 0;
@%p1 bra $L__BB242_2;

cvta.to.global.u64 %rd13, %rd11;
mul.wide.u32 %rd14, %r264, 4;
add.s64 %rd15, %rd13, %rd14;
ld.global.nc.u32 %r263, [%rd15];
mul.hi.u32 %r37, %r264, %r20;
add.s32 %r38, %r37, %r264;
shr.u32 %r39, %r38, %r21;
mul.lo.s32 %r40, %r39, %r22;
sub.s32 %r264, %r264, %r40;
bra.uni $L__BB242_3;

$L__BB242_2:
mul.hi.u32 %r41, %r264, %r24;
add.s32 %r42, %r41, %r264;
shr.u32 %r263, %r42, %r25;

$L__BB242_3:
mov.u32 %r10, %ctaid.z;
shr.s32 %r43, %r3, 31;
shr.u32 %r44, %r43, 29;
add.s32 %r45, %r3, %r44;
shr.s32 %r266, %r45, 3;
shr.u32 %r46, %r19, 8;
setp.ge.s32 %p2, %r266, %r46;
mov.f32 %f27, 0f00000000;
@%p2 bra $L__BB242_6;

and.b32 %r50, %r45, -8;
sub.s32 %r51, %r3, %r50;
shl.b32 %r52, %r51, 2;
shl.b32 %r53, %r51, 1;
add.s32 %r54, %r53, 32;
cvt.s64.s32 %rd1, %r51;
mul.hi.u32 %r55, %r10, %r30;
add.s32 %r56, %r55, %r10;
shr.u32 %r57, %r56, %r31;
mad.lo.s32 %r58, %r57, %r33, %r266;
mad.lo.s32 %r59, %r263, %r27, %r58;
mov.u32 %r60, %ctaid.x;
mad.lo.s32 %r61, %r60, %r23, %r59;
mul.wide.s32 %rd16, %r61, 98;
cvta.to.global.u64 %rd17, %rd9;
add.s64 %rd53, %rd17, %rd16;
shl.b32 %r265, %r266, 3;
mul.wide.s32 %rd18, %r54, 2;
add.s64 %rd3, %rd18, 4;
mul.wide.s32 %rd19, %r52, 2;
or.b64 %rd4, %rd19, 4;
cvta.to.global.u64 %rd5, %rd10;
mul.lo.s32 %r62, %r264, %r28;
cvt.u64.u32 %rd20, %r62;
mul.lo.s32 %r63, %r10, %r34;
cvt.u64.u32 %rd21, %r63;
add.s64 %rd6, %rd20, %rd21;
mov.f32 %f27, 0f00000000;

$L__BB242_5:
cvt.u64.u32 %rd22, %r265;
add.s64 %rd23, %rd6, %rd22;
add.s64 %rd24, %rd53, %rd4;
ld.global.nc.u16 %rs2, [%rd24];
ld.global.nc.u16 %rs3, [%rd24+2];
ld.global.nc.u16 %rs4, [%rd24+4];
add.s64 %rd25, %rd53, %rd3;
ld.global.nc.u16 %rs5, [%rd25+-2];
cvt.u32.u16 %r113, %rs5;
ld.global.nc.u16 %rs6, [%rd25];
cvt.u32.u16 %r114, %rs6;
mov.b32 %r115, {%rs5, %rs6};
add.s64 %rd26, %rd23, %rd1;
mul.lo.s64 %rd27, %rd26, 36;
ld.global.nc.u16 %rs7, [%rd24+-2];
shl.b16 %rs8, %rs7, 2;
cvt.u64.u16 %rd28, %rs8;
and.b64 %rd29, %rd28, 1020;
mov.u64 %rd30, iq3xxs_grid;
add.s64 %rd31, %rd30, %rd29;
ld.global.nc.u32 %r116, [%rd31];
shr.u16 %rs9, %rs7, 8;
cvt.u32.u16 %r117, %rs9;
mul.wide.u32 %rd32, %r117, 4;
add.s64 %rd33, %rd30, %rd32;
ld.global.nc.u32 %r118, [%rd33];
and.b32 %r119, %r113, 255;
popc.b32 %r120, %r119;
shl.b32 %r121, %r120, 7;
and.b32 %r122, %r121, 128;
xor.b32 %r123, %r122, %r119;
mul.lo.s32 %r124, %r123, 16843009;
and.b32 %r125, %r124, 134480385;
or.b32 %r126, %r125, -2139062144;
add.s32 %r127, %r126, -16843009;
or.b32 %r64, %r127, %r125;

prmt.b32 %r64, %r64, 0, 0xba98;

xor.b32 %r128, %r64, %r116;
or.b32 %r129, %r128, -2139062144;
and.b32 %r130, %r64, 2139062143;
and.b32 %r131, %r116, -2139062144;
sub.s32 %r132, %r129, %r130;
xor.b32 %r133, %r131, %r132;
xor.b32 %r69, %r133, -2139062144;
add.s64 %rd34, %rd5, %rd27;
ld.global.nc.u32 %r70, [%rd34+4];
and.b32 %r134, %r124, -2143281136;
or.b32 %r135, %r134, -2139062144;
add.s32 %r136, %r135, -16843009;
or.b32 %r66, %r136, %r134;

prmt.b32 %r66, %r66, 0, 0xba98;

xor.b32 %r137, %r66, %r118;
or.b32 %r138, %r137, -2139062144;
and.b32 %r139, %r66, 2139062143;
and.b32 %r140, %r118, -2139062144;
sub.s32 %r141, %r138, %r139;
xor.b32 %r142, %r140, %r141;
xor.b32 %r73, %r142, -2139062144;
ld.global.nc.u32 %r74, [%rd34+8];
mov.u32 %r71, 0;

dp4a.s32.s32 %r68, %r69, %r70, %r71;


dp4a.s32.s32 %r72, %r73, %r74, %r68;

shl.b16 %rs10, %rs2, 2;
cvt.u64.u16 %rd35, %rs10;
and.b64 %rd36, %rd35, 1020;
add.s64 %rd37, %rd30, %rd36;
ld.global.nc.u32 %r143, [%rd37];
shr.u16 %rs11, %rs2, 8;
cvt.u32.u16 %r144, %rs11;
mul.wide.u32 %rd38, %r144, 4;
add.s64 %rd39, %rd30, %rd38;
ld.global.nc.u32 %r145, [%rd39];
shr.u32 %r146, %r113, 7;
and.b32 %r147, %r146, 255;
popc.b32 %r148, %r147;
shl.b32 %r149, %r148, 7;
and.b32 %r150, %r149, 128;
xor.b32 %r151, %r150, %r147;
mul.lo.s32 %r152, %r151, 16843009;
and.b32 %r153, %r152, 134480385;
or.b32 %r154, %r153, -2139062144;
add.s32 %r155, %r154, -16843009;
or.b32 %r76, %r155, %r153;

prmt.b32 %r76, %r76, 0, 0xba98;

xor.b32 %r156, %r76, %r143;
or.b32 %r157, %r156, -2139062144;
and.b32 %r158, %r76, 2139062143;
and.b32 %r159, %r143, -2139062144;
sub.s32 %r160, %r157, %r158;
xor.b32 %r161, %r159, %r160;
xor.b32 %r81, %r161, -2139062144;
ld.global.nc.u32 %r82, [%rd34+12];
and.b32 %r162, %r152, -2143281136;
or.b32 %r163, %r162, -2139062144;
add.s32 %r164, %r163, -16843009;
or.b32 %r78, %r164, %r162;

prmt.b32 %r78, %r78, 0, 0xba98;

xor.b32 %r165, %r78, %r145;
or.b32 %r166, %r165, -2139062144;
and.b32 %r167, %r78, 2139062143;
and.b32 %r168, %r145, -2139062144;
sub.s32 %r169, %r166, %r167;
xor.b32 %r170, %r168, %r169;
xor.b32 %r85, %r170, -2139062144;
ld.global.nc.u32 %r86, [%rd34+16];

dp4a.s32.s32 %r80, %r81, %r82, %r72;


dp4a.s32.s32 %r84, %r85, %r86, %r80;

shl.b16 %rs12, %rs3, 2;
cvt.u64.u16 %rd40, %rs12;
and.b64 %rd41, %rd40, 1020;
add.s64 %rd42, %rd30, %rd41;
ld.global.nc.u32 %r171, [%rd42];
shr.u16 %rs13, %rs3, 8;
cvt.u32.u16 %r172, %rs13;
mul.wide.u32 %rd43, %r172, 4;
add.s64 %rd44, %rd30, %rd43;
ld.global.nc.u32 %r173, [%rd44];
shr.u32 %r174, %r115, 14;
and.b32 %r175, %r174, 255;
popc.b32 %r176, %r175;
shl.b32 %r177, %r176, 7;
and.b32 %r178, %r177, 128;
xor.b32 %r179, %r178, %r175;
mul.lo.s32 %r180, %r179, 16843009;
and.b32 %r181, %r180, 134480385;
or.b32 %r182, %r181, -2139062144;
add.s32 %r183, %r182, -16843009;
or.b32 %r88, %r183, %r181;

prmt.b32 %r88, %r88, 0, 0xba98;

xor.b32 %r184, %r88, %r171;
or.b32 %r185, %r184, -2139062144;
and.b32 %r186, %r88, 2139062143;
and.b32 %r187, %r171, -2139062144;
sub.s32 %r188, %r185, %r186;
xor.b32 %r189, %r187, %r188;
xor.b32 %r93, %r189, -2139062144;
ld.global.nc.u32 %r94, [%rd34+20];
and.b32 %r190, %r180, -2143281136;
or.b32 %r191, %r190, -2139062144;
add.s32 %r192, %r191, -16843009;
or.b32 %r90, %r192, %r190;

prmt.b32 %r90, %r90, 0, 0xba98;

xor.b32 %r193, %r90, %r173;
or.b32 %r194, %r193, -2139062144;
and.b32 %r195, %r90, 2139062143;
and.b32 %r196, %r173, -2139062144;
sub.s32 %r197, %r194, %r195;
xor.b32 %r198, %r196, %r197;
xor.b32 %r97, %r198, -2139062144;
ld.global.nc.u32 %r98, [%rd34+24];

dp4a.s32.s32 %r92, %r93, %r94, %r84;


dp4a.s32.s32 %r96, %r97, %r98, %r92;

shl.b16 %rs14, %rs4, 2;
cvt.u64.u16 %rd45, %rs14;
and.b64 %rd46, %rd45, 1020;
add.s64 %rd47, %rd30, %rd46;
ld.global.nc.u32 %r199, [%rd47];
shr.u16 %rs15, %rs4, 8;
cvt.u32.u16 %r200, %rs15;
mul.wide.u32 %rd48, %r200, 4;
add.s64 %rd49, %rd30, %rd48;
ld.global.nc.u32 %r201, [%rd49];
shr.u32 %r202, %r114, 5;
and.b32 %r203, %r202, 255;
popc.b32 %r204, %r203;
shl.b32 %r205, %r204, 7;
and.b32 %r206, %r205, 128;
xor.b32 %r207, %r206, %r203;
mul.lo.s32 %r208, %r207, 16843009;
and.b32 %r209, %r208, 134480385;
or.b32 %r210, %r209, -2139062144;
add.s32 %r211, %r210, -16843009;
or.b32 %r100, %r211, %r209;

prmt.b32 %r100, %r100, 0, 0xba98;

xor.b32 %r212, %r100, %r199;
or.b32 %r213, %r212, -2139062144;
and.b32 %r214, %r100, 2139062143;
and.b32 %r215, %r199, -2139062144;
sub.s32 %r216, %r213, %r214;
xor.b32 %r217, %r215, %r216;
xor.b32 %r105, %r217, -2139062144;
ld.global.nc.u32 %r106, [%rd34+28];
and.b32 %r218, %r208, -2143281136;
or.b32 %r219, %r218, -2139062144;
add.s32 %r220, %r219, -16843009;
or.b32 %r102, %r220, %r218;

prmt.b32 %r102, %r102, 0, 0xba98;

xor.b32 %r221, %r102, %r201;
or.b32 %r222, %r221, -2139062144;
and.b32 %r223, %r102, 2139062143;
and.b32 %r224, %r201, -2139062144;
sub.s32 %r225, %r222, %r223;
xor.b32 %r226, %r224, %r225;
xor.b32 %r109, %r226, -2139062144;
ld.global.nc.u32 %r110, [%rd34+32];

dp4a.s32.s32 %r104, %r105, %r106, %r96;

shr.u32 %r227, %r114, 12;

dp4a.s32.s32 %r108, %r109, %r110, %r104;

shr.u32 %r228, %r108, 31;
add.s32 %r229, %r108, %r228;
shr.s32 %r230, %r229, 1;
mad.lo.s32 %r231, %r108, %r227, %r230;
shr.u32 %r232, %r231, 31;
add.s32 %r233, %r231, %r232;
shr.s32 %r234, %r233, 1;
ld.global.nc.u16 %rs1, [%rd53];

{ cvt.f32.f16 %f7, %rs1;}


ld.global.nc.u32 %r112, [%rd34];

{.reg .f16 low,high;
mov.b32 {low,high},%r112;
cvt.f32.f16 %f8, low;}


mul.ftz.f32 %f9, %f7, %f8;
cvt.rn.f32.s32 %f10, %r234;
fma.rn.ftz.f32 %f27, %f9, %f10, %f27;
add.s64 %rd53, %rd53, 1568;
add.s32 %r265, %r265, 128;
add.s32 %r266, %r266, 16;
setp.lt.s32 %p3, %r266, %r46;
@%p3 bra $L__BB242_5;

$L__BB242_6:
setp.eq.s32 %p4, %r1, 0;
@%p4 bra $L__BB242_8;

shl.b32 %r235, %r1, 7;
mov.u32 %r236, _ZZ13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r237, %r236, %r235;
shl.b32 %r238, %r2, 2;
add.s32 %r239, %r237, %r238;
st.shared.f32 [%r239+-128], %f27;

$L__BB242_8:
setp.ne.s32 %p5, %r1, 0;
bar.sync 0;
@%p5 bra $L__BB242_11;

mov.u32 %r240, %ctaid.y;
mov.u32 %r241, %ctaid.x;
mad.lo.s32 %r242, %r240, %r29, %r241;
mad.lo.s32 %r18, %r10, %r35, %r242;
shl.b32 %r243, %r2, 2;
mov.u32 %r244, _ZZ13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r245, %r244, %r243;
mov.u32 %r246, 2;
ld.shared.f32 %f11, [%r245];
add.ftz.f32 %f12, %f27, %f11;
ld.shared.f32 %f13, [%r245+128];
add.ftz.f32 %f14, %f12, %f13;
ld.shared.f32 %f15, [%r245+256];
add.ftz.f32 %f16, %f14, %f15;
mov.b32 %r247, %f16;
mov.u32 %r248, 31;
mov.u32 %r249, 16;
mov.u32 %r250, -1;
shfl.sync.bfly.b32 %r251|%p6, %r247, %r249, %r248, %r250;
mov.b32 %f17, %r251;
add.ftz.f32 %f18, %f16, %f17;
mov.b32 %r252, %f18;
mov.u32 %r253, 8;
shfl.sync.bfly.b32 %r254|%p7, %r252, %r253, %r248, %r250;
mov.b32 %f19, %r254;
add.ftz.f32 %f20, %f18, %f19;
mov.b32 %r255, %f20;
mov.u32 %r256, 4;
shfl.sync.bfly.b32 %r257|%p8, %r255, %r256, %r248, %r250;
mov.b32 %f21, %r257;
add.ftz.f32 %f22, %f20, %f21;
mov.b32 %r258, %f22;
shfl.sync.bfly.b32 %r259|%p9, %r258, %r246, %r248, %r250;
mov.b32 %f23, %r259;
add.ftz.f32 %f24, %f22, %f23;
mov.b32 %r260, %f24;
mov.u32 %r261, 1;
shfl.sync.bfly.b32 %r262|%p10, %r260, %r261, %r248, %r250;
mov.b32 %f25, %r262;
add.ftz.f32 %f4, %f24, %f25;
setp.ne.s32 %p11, %r2, 0;
@%p11 bra $L__BB242_11;

cvta.to.global.u64 %rd50, %rd12;
mul.wide.u32 %rd51, %r18, 4;
add.s64 %rd52, %rd50, %rd51;
st.global.f32 [%rd52], %f4;

$L__BB242_11:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type18ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type18ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type18ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type18ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type18ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type18ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type18ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type18ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type18ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<35>;
.reg .b16 %rs<33>;
.reg .f32 %f<109>;
.reg .b32 %r<667>;
.reg .b64 %rd<95>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type18ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type18ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type18ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type18ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type18ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r17, [_Z13mul_mat_vec_qIL9ggml_type18ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r18, [_Z13mul_mat_vec_qIL9ggml_type18ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r19, [_Z13mul_mat_vec_qIL9ggml_type18ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type18ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type18ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type18ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type18ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type18ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type18ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type18ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type18ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type18ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type18ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
cvta.to.global.u64 %rd1, %rd18;
mov.u32 %r1, %tid.y;
shl.b32 %r33, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r33, %r2;
mov.u32 %r34, %ctaid.x;
shl.b32 %r4, %r34, 1;
shr.s32 %r35, %r3, 31;
shr.u32 %r36, %r35, 29;
add.s32 %r37, %r3, %r36;
shr.s32 %r666, %r37, 3;
shr.u32 %r6, %r32, 8;
setp.ge.s32 %p3, %r666, %r6;
mov.f32 %f105, 0f00000000;
mov.f32 %f106, %f105;
mov.f32 %f107, %f105;
mov.f32 %f108, %f105;
@%p3 bra $L__BB243_3;

mov.u32 %r38, %ctaid.y;
mov.u32 %r39, %ctaid.z;
mul.lo.s32 %r40, %r38, %r24;
cvt.u64.u32 %rd19, %r40;
mul.lo.s32 %r41, %r39, %r30;
cvt.u64.u32 %rd20, %r41;
add.s64 %rd2, %rd20, %rd19;
mul.hi.u32 %r42, %r39, %r26;
add.s32 %r43, %r42, %r39;
shr.u32 %r44, %r43, %r27;
mul.lo.s32 %r45, %r44, %r29;
mul.hi.u32 %r46, %r38, %r20;
add.s32 %r47, %r46, %r38;
shr.u32 %r48, %r47, %r21;
mul.lo.s32 %r49, %r48, %r23;
cvta.to.global.u64 %rd94, %rd16;
and.b32 %r53, %r37, -8;
sub.s32 %r54, %r3, %r53;
shl.b32 %r55, %r54, 2;
shl.b32 %r56, %r54, 1;
add.s32 %r57, %r56, 33;
cvt.s64.s32 %rd4, %r54;
shl.b32 %r664, %r666, 3;
add.s32 %r665, %r18, %r664;
add.s32 %r58, %r666, %r45;
add.s32 %r59, %r58, %r49;
add.s32 %r60, %r4, 1;
mad.lo.s32 %r61, %r17, %r60, %r59;
mul.wide.s32 %rd5, %r61, 98;
mad.lo.s32 %r62, %r4, %r17, %r666;
add.s32 %r63, %r62, %r45;
add.s32 %r64, %r63, %r49;
mul.wide.s32 %rd6, %r64, 98;
mul.wide.s32 %rd21, %r55, 2;
add.s64 %rd22, %rd6, %rd21;
add.s64 %rd7, %rd22, 8;
mul.wide.s32 %rd23, %r57, 2;
add.s64 %rd24, %rd23, -2;
add.s64 %rd25, %rd24, %rd5;
add.s64 %rd8, %rd25, 2;
add.s64 %rd26, %rd24, %rd6;
add.s64 %rd9, %rd26, 2;
add.s64 %rd27, %rd5, %rd21;
add.s64 %rd10, %rd27, 4;
cvta.to.global.u64 %rd11, %rd17;
mov.f32 %f105, 0f00000000;

$L__BB243_2:
cvt.u64.u32 %rd28, %r664;
add.s64 %rd29, %rd2, %rd28;
add.s64 %rd30, %rd29, %rd4;
mul.lo.s64 %rd31, %rd30, 36;
add.s64 %rd32, %rd94, %rd7;
ld.global.nc.u16 %rs5, [%rd32+-4];
ld.global.nc.u16 %rs6, [%rd32+-2];
ld.global.nc.u16 %rs7, [%rd32];
add.s64 %rd33, %rd94, %rd9;
ld.global.nc.u16 %rs8, [%rd33];
cvt.u32.u16 %r261, %rs8;
ld.global.nc.u16 %rs9, [%rd33+2];
cvt.u32.u16 %r262, %rs9;
mov.b32 %r263, {%rs8, %rs9};
ld.global.nc.u16 %rs10, [%rd32+-6];
shl.b16 %rs11, %rs10, 2;
cvt.u64.u16 %rd34, %rs11;
and.b64 %rd35, %rd34, 1020;
mov.u64 %rd36, iq3xxs_grid;
add.s64 %rd37, %rd36, %rd35;
ld.global.nc.u32 %r264, [%rd37];
not.b32 %r265, %r264;
shr.u16 %rs12, %rs10, 8;
cvt.u32.u16 %r266, %rs12;
mul.wide.u32 %rd38, %r266, 4;
add.s64 %rd39, %rd36, %rd38;
ld.global.nc.u32 %r267, [%rd39];
not.b32 %r268, %r267;
and.b32 %r269, %r261, 255;
popc.b32 %r270, %r269;
shl.b32 %r271, %r270, 7;
and.b32 %r272, %r271, 128;
xor.b32 %r273, %r272, %r269;
mul.lo.s32 %r274, %r273, 16843009;
and.b32 %r275, %r274, 134480385;
or.b32 %r276, %r275, -2139062144;
add.s32 %r277, %r276, -16843009;
or.b32 %r163, %r277, %r275;
mov.u32 %r65, %r163;

prmt.b32 %r65, %r65, 0, 0xba98;

xor.b32 %r278, %r65, %r264;
or.b32 %r279, %r278, -2139062144;
and.b32 %r280, %r65, 2139062143;
and.b32 %r281, %r265, -2139062144;
sub.s32 %r282, %r279, %r280;
xor.b32 %r70, %r281, %r282;
add.s64 %rd40, %rd11, %rd31;
ld.global.nc.u32 %r71, [%rd40+4];
and.b32 %r283, %r274, -2143281136;
or.b32 %r284, %r283, -2139062144;
add.s32 %r285, %r284, -16843009;
or.b32 %r165, %r285, %r283;
mov.u32 %r67, %r165;

prmt.b32 %r67, %r67, 0, 0xba98;

xor.b32 %r286, %r67, %r267;
or.b32 %r287, %r286, -2139062144;
and.b32 %r288, %r67, 2139062143;
and.b32 %r289, %r268, -2139062144;
sub.s32 %r290, %r287, %r288;
xor.b32 %r74, %r289, %r290;
ld.global.nc.u32 %r75, [%rd40+8];
mov.u32 %r219, 0;

dp4a.s32.s32 %r69, %r70, %r71, %r219;


dp4a.s32.s32 %r73, %r74, %r75, %r69;

shl.b16 %rs13, %rs5, 2;
cvt.u64.u16 %rd41, %rs13;
and.b64 %rd42, %rd41, 1020;
add.s64 %rd43, %rd36, %rd42;
ld.global.nc.u32 %r291, [%rd43];
not.b32 %r292, %r291;
shr.u16 %rs14, %rs5, 8;
cvt.u32.u16 %r293, %rs14;
mul.wide.u32 %rd44, %r293, 4;
add.s64 %rd45, %rd36, %rd44;
ld.global.nc.u32 %r294, [%rd45];
not.b32 %r295, %r294;
shr.u32 %r296, %r261, 7;
and.b32 %r297, %r296, 255;
popc.b32 %r298, %r297;
shl.b32 %r299, %r298, 7;
and.b32 %r300, %r299, 128;
xor.b32 %r301, %r300, %r297;
mul.lo.s32 %r302, %r301, 16843009;
and.b32 %r303, %r302, 134480385;
or.b32 %r304, %r303, -2139062144;
add.s32 %r305, %r304, -16843009;
or.b32 %r175, %r305, %r303;
mov.u32 %r77, %r175;

prmt.b32 %r77, %r77, 0, 0xba98;

xor.b32 %r306, %r77, %r291;
or.b32 %r307, %r306, -2139062144;
and.b32 %r308, %r77, 2139062143;
and.b32 %r309, %r292, -2139062144;
sub.s32 %r310, %r307, %r308;
xor.b32 %r82, %r309, %r310;
ld.global.nc.u32 %r83, [%rd40+12];
and.b32 %r311, %r302, -2143281136;
or.b32 %r312, %r311, -2139062144;
add.s32 %r313, %r312, -16843009;
or.b32 %r177, %r313, %r311;
mov.u32 %r79, %r177;

prmt.b32 %r79, %r79, 0, 0xba98;

xor.b32 %r314, %r79, %r294;
or.b32 %r315, %r314, -2139062144;
and.b32 %r316, %r79, 2139062143;
and.b32 %r317, %r295, -2139062144;
sub.s32 %r318, %r315, %r316;
xor.b32 %r86, %r317, %r318;
ld.global.nc.u32 %r87, [%rd40+16];

dp4a.s32.s32 %r81, %r82, %r83, %r73;


dp4a.s32.s32 %r85, %r86, %r87, %r81;

shl.b16 %rs15, %rs6, 2;
cvt.u64.u16 %rd46, %rs15;
and.b64 %rd47, %rd46, 1020;
add.s64 %rd48, %rd36, %rd47;
ld.global.nc.u32 %r319, [%rd48];
not.b32 %r320, %r319;
shr.u16 %rs16, %rs6, 8;
cvt.u32.u16 %r321, %rs16;
mul.wide.u32 %rd49, %r321, 4;
add.s64 %rd50, %rd36, %rd49;
ld.global.nc.u32 %r322, [%rd50];
not.b32 %r323, %r322;
shr.u32 %r324, %r263, 14;
and.b32 %r325, %r324, 255;
popc.b32 %r326, %r325;
shl.b32 %r327, %r326, 7;
and.b32 %r328, %r327, 128;
xor.b32 %r329, %r328, %r325;
mul.lo.s32 %r330, %r329, 16843009;
and.b32 %r331, %r330, 134480385;
or.b32 %r332, %r331, -2139062144;
add.s32 %r333, %r332, -16843009;
or.b32 %r187, %r333, %r331;
mov.u32 %r89, %r187;

prmt.b32 %r89, %r89, 0, 0xba98;

xor.b32 %r334, %r89, %r319;
or.b32 %r335, %r334, -2139062144;
and.b32 %r336, %r89, 2139062143;
and.b32 %r337, %r320, -2139062144;
sub.s32 %r338, %r335, %r336;
xor.b32 %r94, %r337, %r338;
ld.global.nc.u32 %r95, [%rd40+20];
and.b32 %r339, %r330, -2143281136;
or.b32 %r340, %r339, -2139062144;
add.s32 %r341, %r340, -16843009;
or.b32 %r189, %r341, %r339;
mov.u32 %r91, %r189;

prmt.b32 %r91, %r91, 0, 0xba98;

xor.b32 %r342, %r91, %r322;
or.b32 %r343, %r342, -2139062144;
and.b32 %r344, %r91, 2139062143;
and.b32 %r345, %r323, -2139062144;
sub.s32 %r346, %r343, %r344;
xor.b32 %r98, %r345, %r346;
ld.global.nc.u32 %r99, [%rd40+24];

dp4a.s32.s32 %r93, %r94, %r95, %r85;


dp4a.s32.s32 %r97, %r98, %r99, %r93;

shl.b16 %rs17, %rs7, 2;
cvt.u64.u16 %rd51, %rs17;
and.b64 %rd52, %rd51, 1020;
add.s64 %rd53, %rd36, %rd52;
ld.global.nc.u32 %r347, [%rd53];
not.b32 %r348, %r347;
shr.u16 %rs18, %rs7, 8;
cvt.u32.u16 %r349, %rs18;
mul.wide.u32 %rd54, %r349, 4;
add.s64 %rd55, %rd36, %rd54;
ld.global.nc.u32 %r350, [%rd55];
not.b32 %r351, %r350;
shr.u32 %r352, %r262, 5;
and.b32 %r353, %r352, 255;
popc.b32 %r354, %r353;
shl.b32 %r355, %r354, 7;
and.b32 %r356, %r355, 128;
xor.b32 %r357, %r356, %r353;
mul.lo.s32 %r358, %r357, 16843009;
and.b32 %r359, %r358, 134480385;
or.b32 %r360, %r359, -2139062144;
add.s32 %r361, %r360, -16843009;
or.b32 %r199, %r361, %r359;
mov.u32 %r101, %r199;

prmt.b32 %r101, %r101, 0, 0xba98;

xor.b32 %r362, %r101, %r347;
or.b32 %r363, %r362, -2139062144;
and.b32 %r364, %r101, 2139062143;
and.b32 %r365, %r348, -2139062144;
sub.s32 %r366, %r363, %r364;
xor.b32 %r106, %r365, %r366;
ld.global.nc.u32 %r107, [%rd40+28];
and.b32 %r367, %r358, -2143281136;
or.b32 %r368, %r367, -2139062144;
add.s32 %r369, %r368, -16843009;
or.b32 %r201, %r369, %r367;
mov.u32 %r103, %r201;

prmt.b32 %r103, %r103, 0, 0xba98;

xor.b32 %r370, %r103, %r350;
or.b32 %r371, %r370, -2139062144;
and.b32 %r372, %r103, 2139062143;
and.b32 %r373, %r351, -2139062144;
sub.s32 %r374, %r371, %r372;
xor.b32 %r110, %r373, %r374;
ld.global.nc.u32 %r111, [%rd40+32];

dp4a.s32.s32 %r105, %r106, %r107, %r97;

shr.u32 %r375, %r262, 12;

dp4a.s32.s32 %r109, %r110, %r111, %r105;

shr.u32 %r376, %r109, 31;
add.s32 %r377, %r109, %r376;
shr.s32 %r378, %r377, 1;
mad.lo.s32 %r379, %r109, %r375, %r378;
shr.u32 %r380, %r379, 31;
add.s32 %r381, %r379, %r380;
shr.s32 %r382, %r381, 1;
add.s64 %rd56, %rd94, %rd6;
ld.global.nc.u16 %rs1, [%rd56];

{ cvt.f32.f16 %f25, %rs1;}


ld.global.nc.u32 %r113, [%rd40];

{.reg .f16 low,high;
mov.b32 {low,high},%r113;
cvt.f32.f16 %f26, low;}


mul.ftz.f32 %f33, %f25, %f26;
cvt.rn.f32.s32 %f34, %r382;
fma.rn.ftz.f32 %f108, %f33, %f34, %f108;
add.s64 %rd57, %rd94, %rd10;
ld.global.nc.u16 %rs19, [%rd57];
ld.global.nc.u16 %rs20, [%rd57+2];
ld.global.nc.u16 %rs21, [%rd57+4];
add.s64 %rd58, %rd94, %rd8;
ld.global.nc.u16 %rs22, [%rd58];
cvt.u32.u16 %r383, %rs22;
ld.global.nc.u16 %rs23, [%rd58+2];
cvt.u32.u16 %r384, %rs23;
mov.b32 %r385, {%rs22, %rs23};
ld.global.nc.u16 %rs24, [%rd57+-2];
shl.b16 %rs25, %rs24, 2;
cvt.u64.u16 %rd59, %rs25;
and.b64 %rd60, %rd59, 1020;
add.s64 %rd61, %rd36, %rd60;
ld.global.nc.u32 %r386, [%rd61];
not.b32 %r387, %r386;
shr.u16 %rs26, %rs24, 8;
cvt.u32.u16 %r388, %rs26;
mul.wide.u32 %rd62, %r388, 4;
add.s64 %rd63, %rd36, %rd62;
ld.global.nc.u32 %r389, [%rd63];
not.b32 %r390, %r389;
and.b32 %r391, %r383, 255;
popc.b32 %r392, %r391;
shl.b32 %r393, %r392, 7;
and.b32 %r394, %r393, 128;
xor.b32 %r395, %r394, %r391;
mul.lo.s32 %r396, %r395, 16843009;
and.b32 %r397, %r396, 134480385;
or.b32 %r398, %r397, -2139062144;
add.s32 %r399, %r398, -16843009;
or.b32 %r212, %r399, %r397;
mov.u32 %r114, %r212;

prmt.b32 %r114, %r114, 0, 0xba98;

xor.b32 %r400, %r114, %r386;
or.b32 %r401, %r400, -2139062144;
and.b32 %r402, %r114, 2139062143;
and.b32 %r403, %r387, -2139062144;
sub.s32 %r404, %r401, %r402;
xor.b32 %r119, %r403, %r404;
and.b32 %r405, %r396, -2143281136;
or.b32 %r406, %r405, -2139062144;
add.s32 %r407, %r406, -16843009;
or.b32 %r214, %r407, %r405;
mov.u32 %r116, %r214;

prmt.b32 %r116, %r116, 0, 0xba98;

xor.b32 %r408, %r116, %r389;
or.b32 %r409, %r408, -2139062144;
and.b32 %r410, %r116, 2139062143;
and.b32 %r411, %r390, -2139062144;
sub.s32 %r412, %r409, %r410;
xor.b32 %r123, %r411, %r412;

dp4a.s32.s32 %r118, %r119, %r71, %r219;


dp4a.s32.s32 %r122, %r123, %r75, %r118;

shl.b16 %rs27, %rs19, 2;
cvt.u64.u16 %rd64, %rs27;
and.b64 %rd65, %rd64, 1020;
add.s64 %rd66, %rd36, %rd65;
ld.global.nc.u32 %r413, [%rd66];
not.b32 %r414, %r413;
shr.u16 %rs28, %rs19, 8;
cvt.u32.u16 %r415, %rs28;
mul.wide.u32 %rd67, %r415, 4;
add.s64 %rd68, %rd36, %rd67;
ld.global.nc.u32 %r416, [%rd68];
not.b32 %r417, %r416;
shr.u32 %r418, %r383, 7;
and.b32 %r419, %r418, 255;
popc.b32 %r420, %r419;
shl.b32 %r421, %r420, 7;
and.b32 %r422, %r421, 128;
xor.b32 %r423, %r422, %r419;
mul.lo.s32 %r424, %r423, 16843009;
and.b32 %r425, %r424, 134480385;
or.b32 %r426, %r425, -2139062144;
add.s32 %r427, %r426, -16843009;
or.b32 %r224, %r427, %r425;
mov.u32 %r126, %r224;

prmt.b32 %r126, %r126, 0, 0xba98;

xor.b32 %r428, %r126, %r413;
or.b32 %r429, %r428, -2139062144;
and.b32 %r430, %r126, 2139062143;
and.b32 %r431, %r414, -2139062144;
sub.s32 %r432, %r429, %r430;
xor.b32 %r131, %r431, %r432;
and.b32 %r433, %r424, -2143281136;
or.b32 %r434, %r433, -2139062144;
add.s32 %r435, %r434, -16843009;
or.b32 %r226, %r435, %r433;
mov.u32 %r128, %r226;

prmt.b32 %r128, %r128, 0, 0xba98;

xor.b32 %r436, %r128, %r416;
or.b32 %r437, %r436, -2139062144;
and.b32 %r438, %r128, 2139062143;
and.b32 %r439, %r417, -2139062144;
sub.s32 %r440, %r437, %r438;
xor.b32 %r135, %r439, %r440;

dp4a.s32.s32 %r130, %r131, %r83, %r122;


dp4a.s32.s32 %r134, %r135, %r87, %r130;

shl.b16 %rs29, %rs20, 2;
cvt.u64.u16 %rd69, %rs29;
and.b64 %rd70, %rd69, 1020;
add.s64 %rd71, %rd36, %rd70;
ld.global.nc.u32 %r441, [%rd71];
not.b32 %r442, %r441;
shr.u16 %rs30, %rs20, 8;
cvt.u32.u16 %r443, %rs30;
mul.wide.u32 %rd72, %r443, 4;
add.s64 %rd73, %rd36, %rd72;
ld.global.nc.u32 %r444, [%rd73];
not.b32 %r445, %r444;
shr.u32 %r446, %r385, 14;
and.b32 %r447, %r446, 255;
popc.b32 %r448, %r447;
shl.b32 %r449, %r448, 7;
and.b32 %r450, %r449, 128;
xor.b32 %r451, %r450, %r447;
mul.lo.s32 %r452, %r451, 16843009;
and.b32 %r453, %r452, 134480385;
or.b32 %r454, %r453, -2139062144;
add.s32 %r455, %r454, -16843009;
or.b32 %r236, %r455, %r453;
mov.u32 %r138, %r236;

prmt.b32 %r138, %r138, 0, 0xba98;

xor.b32 %r456, %r138, %r441;
or.b32 %r457, %r456, -2139062144;
and.b32 %r458, %r138, 2139062143;
and.b32 %r459, %r442, -2139062144;
sub.s32 %r460, %r457, %r458;
xor.b32 %r143, %r459, %r460;
and.b32 %r461, %r452, -2143281136;
or.b32 %r462, %r461, -2139062144;
add.s32 %r463, %r462, -16843009;
or.b32 %r238, %r463, %r461;
mov.u32 %r140, %r238;

prmt.b32 %r140, %r140, 0, 0xba98;

xor.b32 %r464, %r140, %r444;
or.b32 %r465, %r464, -2139062144;
and.b32 %r466, %r140, 2139062143;
and.b32 %r467, %r445, -2139062144;
sub.s32 %r468, %r465, %r466;
xor.b32 %r147, %r467, %r468;

dp4a.s32.s32 %r142, %r143, %r95, %r134;


dp4a.s32.s32 %r146, %r147, %r99, %r142;

shl.b16 %rs31, %rs21, 2;
cvt.u64.u16 %rd74, %rs31;
and.b64 %rd75, %rd74, 1020;
add.s64 %rd76, %rd36, %rd75;
ld.global.nc.u32 %r469, [%rd76];
not.b32 %r470, %r469;
shr.u16 %rs32, %rs21, 8;
cvt.u32.u16 %r471, %rs32;
mul.wide.u32 %rd77, %r471, 4;
add.s64 %rd78, %rd36, %rd77;
ld.global.nc.u32 %r472, [%rd78];
not.b32 %r473, %r472;
shr.u32 %r474, %r384, 5;
and.b32 %r475, %r474, 255;
popc.b32 %r476, %r475;
shl.b32 %r477, %r476, 7;
and.b32 %r478, %r477, 128;
xor.b32 %r479, %r478, %r475;
mul.lo.s32 %r480, %r479, 16843009;
and.b32 %r481, %r480, 134480385;
or.b32 %r482, %r481, -2139062144;
add.s32 %r483, %r482, -16843009;
or.b32 %r248, %r483, %r481;
mov.u32 %r150, %r248;

prmt.b32 %r150, %r150, 0, 0xba98;

xor.b32 %r484, %r150, %r469;
or.b32 %r485, %r484, -2139062144;
and.b32 %r486, %r150, 2139062143;
and.b32 %r487, %r470, -2139062144;
sub.s32 %r488, %r485, %r486;
xor.b32 %r155, %r487, %r488;
and.b32 %r489, %r480, -2143281136;
or.b32 %r490, %r489, -2139062144;
add.s32 %r491, %r490, -16843009;
or.b32 %r250, %r491, %r489;
mov.u32 %r152, %r250;

prmt.b32 %r152, %r152, 0, 0xba98;

xor.b32 %r492, %r152, %r472;
or.b32 %r493, %r492, -2139062144;
and.b32 %r494, %r152, 2139062143;
and.b32 %r495, %r473, -2139062144;
sub.s32 %r496, %r493, %r494;
xor.b32 %r159, %r495, %r496;

dp4a.s32.s32 %r154, %r155, %r107, %r146;

shr.u32 %r497, %r384, 12;

dp4a.s32.s32 %r158, %r159, %r111, %r154;

shr.u32 %r498, %r158, 31;
add.s32 %r499, %r158, %r498;
shr.s32 %r500, %r499, 1;
mad.lo.s32 %r501, %r158, %r497, %r500;
shr.u32 %r502, %r501, 31;
add.s32 %r503, %r501, %r502;
shr.s32 %r504, %r503, 1;
add.s64 %rd79, %rd94, %rd5;
ld.global.nc.u16 %rs2, [%rd79];

{ cvt.f32.f16 %f27, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r113;
cvt.f32.f16 %f28, low;}


mul.ftz.f32 %f35, %f27, %f28;
cvt.rn.f32.s32 %f36, %r504;
fma.rn.ftz.f32 %f107, %f35, %f36, %f107;
cvt.u64.u32 %rd80, %r665;
add.s64 %rd81, %rd2, %rd80;
add.s64 %rd82, %rd81, %rd4;
mul.lo.s64 %rd83, %rd82, 36;

prmt.b32 %r163, %r163, 0, 0xba98;

xor.b32 %r505, %r163, %r264;
or.b32 %r506, %r505, -2139062144;
and.b32 %r507, %r163, 2139062143;
sub.s32 %r508, %r506, %r507;
xor.b32 %r168, %r281, %r508;
add.s64 %rd84, %rd11, %rd83;
ld.global.nc.u32 %r169, [%rd84+4];

prmt.b32 %r165, %r165, 0, 0xba98;

xor.b32 %r509, %r165, %r267;
or.b32 %r510, %r509, -2139062144;
and.b32 %r511, %r165, 2139062143;
sub.s32 %r512, %r510, %r511;
xor.b32 %r172, %r289, %r512;
ld.global.nc.u32 %r173, [%rd84+8];

dp4a.s32.s32 %r167, %r168, %r169, %r219;


dp4a.s32.s32 %r171, %r172, %r173, %r167;


prmt.b32 %r175, %r175, 0, 0xba98;

xor.b32 %r513, %r175, %r291;
or.b32 %r514, %r513, -2139062144;
and.b32 %r515, %r175, 2139062143;
sub.s32 %r516, %r514, %r515;
xor.b32 %r180, %r309, %r516;
ld.global.nc.u32 %r181, [%rd84+12];

prmt.b32 %r177, %r177, 0, 0xba98;

xor.b32 %r517, %r177, %r294;
or.b32 %r518, %r517, -2139062144;
and.b32 %r519, %r177, 2139062143;
sub.s32 %r520, %r518, %r519;
xor.b32 %r184, %r317, %r520;
ld.global.nc.u32 %r185, [%rd84+16];

dp4a.s32.s32 %r179, %r180, %r181, %r171;


dp4a.s32.s32 %r183, %r184, %r185, %r179;


prmt.b32 %r187, %r187, 0, 0xba98;

xor.b32 %r521, %r187, %r319;
or.b32 %r522, %r521, -2139062144;
and.b32 %r523, %r187, 2139062143;
sub.s32 %r524, %r522, %r523;
xor.b32 %r192, %r337, %r524;
ld.global.nc.u32 %r193, [%rd84+20];

prmt.b32 %r189, %r189, 0, 0xba98;

xor.b32 %r525, %r189, %r322;
or.b32 %r526, %r525, -2139062144;
and.b32 %r527, %r189, 2139062143;
sub.s32 %r528, %r526, %r527;
xor.b32 %r196, %r345, %r528;
ld.global.nc.u32 %r197, [%rd84+24];

dp4a.s32.s32 %r191, %r192, %r193, %r183;


dp4a.s32.s32 %r195, %r196, %r197, %r191;


prmt.b32 %r199, %r199, 0, 0xba98;

xor.b32 %r529, %r199, %r347;
or.b32 %r530, %r529, -2139062144;
and.b32 %r531, %r199, 2139062143;
sub.s32 %r532, %r530, %r531;
xor.b32 %r204, %r365, %r532;
ld.global.nc.u32 %r205, [%rd84+28];

prmt.b32 %r201, %r201, 0, 0xba98;

xor.b32 %r533, %r201, %r350;
or.b32 %r534, %r533, -2139062144;
and.b32 %r535, %r201, 2139062143;
sub.s32 %r536, %r534, %r535;
xor.b32 %r208, %r373, %r536;
ld.global.nc.u32 %r209, [%rd84+32];

dp4a.s32.s32 %r203, %r204, %r205, %r195;


dp4a.s32.s32 %r207, %r208, %r209, %r203;

shr.u32 %r537, %r207, 31;
add.s32 %r538, %r207, %r537;
shr.s32 %r539, %r538, 1;
mad.lo.s32 %r540, %r207, %r375, %r539;
shr.u32 %r541, %r540, 31;
add.s32 %r542, %r540, %r541;
shr.s32 %r543, %r542, 1;

{ cvt.f32.f16 %f29, %rs1;}


ld.global.nc.u32 %r211, [%rd84];

{.reg .f16 low,high;
mov.b32 {low,high},%r211;
cvt.f32.f16 %f30, low;}


mul.ftz.f32 %f37, %f29, %f30;
cvt.rn.f32.s32 %f38, %r543;
fma.rn.ftz.f32 %f106, %f37, %f38, %f106;

prmt.b32 %r212, %r212, 0, 0xba98;

xor.b32 %r544, %r212, %r386;
or.b32 %r545, %r544, -2139062144;
and.b32 %r546, %r212, 2139062143;
sub.s32 %r547, %r545, %r546;
xor.b32 %r217, %r403, %r547;

prmt.b32 %r214, %r214, 0, 0xba98;

xor.b32 %r548, %r214, %r389;
or.b32 %r549, %r548, -2139062144;
and.b32 %r550, %r214, 2139062143;
sub.s32 %r551, %r549, %r550;
xor.b32 %r221, %r411, %r551;

dp4a.s32.s32 %r216, %r217, %r169, %r219;


dp4a.s32.s32 %r220, %r221, %r173, %r216;


prmt.b32 %r224, %r224, 0, 0xba98;

xor.b32 %r552, %r224, %r413;
or.b32 %r553, %r552, -2139062144;
and.b32 %r554, %r224, 2139062143;
sub.s32 %r555, %r553, %r554;
xor.b32 %r229, %r431, %r555;

prmt.b32 %r226, %r226, 0, 0xba98;

xor.b32 %r556, %r226, %r416;
or.b32 %r557, %r556, -2139062144;
and.b32 %r558, %r226, 2139062143;
sub.s32 %r559, %r557, %r558;
xor.b32 %r233, %r439, %r559;

dp4a.s32.s32 %r228, %r229, %r181, %r220;


dp4a.s32.s32 %r232, %r233, %r185, %r228;


prmt.b32 %r236, %r236, 0, 0xba98;

xor.b32 %r560, %r236, %r441;
or.b32 %r561, %r560, -2139062144;
and.b32 %r562, %r236, 2139062143;
sub.s32 %r563, %r561, %r562;
xor.b32 %r241, %r459, %r563;

prmt.b32 %r238, %r238, 0, 0xba98;

xor.b32 %r564, %r238, %r444;
or.b32 %r565, %r564, -2139062144;
and.b32 %r566, %r238, 2139062143;
sub.s32 %r567, %r565, %r566;
xor.b32 %r245, %r467, %r567;

dp4a.s32.s32 %r240, %r241, %r193, %r232;


dp4a.s32.s32 %r244, %r245, %r197, %r240;


prmt.b32 %r248, %r248, 0, 0xba98;

xor.b32 %r568, %r248, %r469;
or.b32 %r569, %r568, -2139062144;
and.b32 %r570, %r248, 2139062143;
sub.s32 %r571, %r569, %r570;
xor.b32 %r253, %r487, %r571;

prmt.b32 %r250, %r250, 0, 0xba98;

xor.b32 %r572, %r250, %r472;
or.b32 %r573, %r572, -2139062144;
and.b32 %r574, %r250, 2139062143;
sub.s32 %r575, %r573, %r574;
xor.b32 %r257, %r495, %r575;

dp4a.s32.s32 %r252, %r253, %r205, %r244;


dp4a.s32.s32 %r256, %r257, %r209, %r252;

shr.u32 %r576, %r256, 31;
add.s32 %r577, %r256, %r576;
shr.s32 %r578, %r577, 1;
mad.lo.s32 %r579, %r256, %r497, %r578;
shr.u32 %r580, %r579, 31;
add.s32 %r581, %r579, %r580;
shr.s32 %r582, %r581, 1;

{ cvt.f32.f16 %f31, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r211;
cvt.f32.f16 %f32, low;}


mul.ftz.f32 %f39, %f31, %f32;
cvt.rn.f32.s32 %f40, %r582;
fma.rn.ftz.f32 %f105, %f39, %f40, %f105;
add.s32 %r665, %r665, 128;
add.s32 %r664, %r664, 128;
add.s64 %rd94, %rd94, 1568;
add.s32 %r666, %r666, 16;
setp.lt.s32 %p4, %r666, %r6;
@%p4 bra $L__BB243_2;

$L__BB243_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB243_5;

shl.b32 %r583, %r1, 9;
mov.u32 %r584, _ZZ13mul_mat_vec_qIL9ggml_type18ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r585, %r584, %r583;
shl.b32 %r586, %r2, 2;
add.s32 %r587, %r585, %r586;
st.shared.f32 [%r587+-512], %f108;
st.shared.f32 [%r587+-384], %f107;
st.shared.f32 [%r587+-256], %f106;
st.shared.f32 [%r587+-128], %f105;

$L__BB243_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB243_14;

mov.u32 %r588, %ctaid.z;
mov.u32 %r589, %ctaid.y;
mad.lo.s32 %r590, %r589, %r25, %r4;
mad.lo.s32 %r591, %r588, %r31, %r590;
cvt.u64.u32 %rd14, %r591;
shl.b32 %r592, %r2, 2;
mov.u32 %r593, _ZZ13mul_mat_vec_qIL9ggml_type18ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r15, %r593, %r592;
mov.u32 %r594, 2;
ld.shared.f32 %f41, [%r15];
add.ftz.f32 %f42, %f41, %f108;
ld.shared.f32 %f43, [%r15+512];
add.ftz.f32 %f44, %f43, %f42;
ld.shared.f32 %f45, [%r15+1024];
add.ftz.f32 %f46, %f45, %f44;
mov.b32 %r595, %f46;
mov.u32 %r596, 31;
mov.u32 %r597, 16;
mov.u32 %r598, -1;
shfl.sync.bfly.b32 %r599|%p7, %r595, %r597, %r596, %r598;
mov.b32 %f47, %r599;
add.ftz.f32 %f48, %f46, %f47;
mov.b32 %r600, %f48;
mov.u32 %r601, 8;
shfl.sync.bfly.b32 %r602|%p8, %r600, %r601, %r596, %r598;
mov.b32 %f49, %r602;
add.ftz.f32 %f50, %f48, %f49;
mov.b32 %r603, %f50;
mov.u32 %r604, 4;
shfl.sync.bfly.b32 %r605|%p9, %r603, %r604, %r596, %r598;
mov.b32 %f51, %r605;
add.ftz.f32 %f52, %f50, %f51;
mov.b32 %r606, %f52;
shfl.sync.bfly.b32 %r607|%p10, %r606, %r594, %r596, %r598;
mov.b32 %f53, %r607;
add.ftz.f32 %f54, %f52, %f53;
mov.b32 %r608, %f54;
mov.u32 %r609, 1;
shfl.sync.bfly.b32 %r610|%p11, %r608, %r609, %r596, %r598;
mov.b32 %f55, %r610;
add.ftz.f32 %f13, %f54, %f55;
setp.lt.u32 %p12, %r4, %r19;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
mul.wide.u32 %rd85, %r591, 4;
add.s64 %rd15, %rd1, %rd85;
not.pred %p14, %p1;
@%p14 bra $L__BB243_8;

st.global.f32 [%rd15], %f13;

$L__BB243_8:
ld.shared.f32 %f56, [%r15+128];
add.ftz.f32 %f57, %f56, %f107;
ld.shared.f32 %f58, [%r15+640];
add.ftz.f32 %f59, %f58, %f57;
ld.shared.f32 %f60, [%r15+1152];
add.ftz.f32 %f61, %f60, %f59;
mov.b32 %r611, %f61;
shfl.sync.bfly.b32 %r615|%p15, %r611, %r597, %r596, %r598;
mov.b32 %f62, %r615;
add.ftz.f32 %f63, %f61, %f62;
mov.b32 %r616, %f63;
shfl.sync.bfly.b32 %r618|%p16, %r616, %r601, %r596, %r598;
mov.b32 %f64, %r618;
add.ftz.f32 %f65, %f63, %f64;
mov.b32 %r619, %f65;
shfl.sync.bfly.b32 %r621|%p17, %r619, %r604, %r596, %r598;
mov.b32 %f66, %r621;
add.ftz.f32 %f67, %f65, %f66;
mov.b32 %r622, %f67;
shfl.sync.bfly.b32 %r624|%p18, %r622, %r594, %r596, %r598;
mov.b32 %f68, %r624;
add.ftz.f32 %f69, %f67, %f68;
mov.b32 %r625, %f69;
shfl.sync.bfly.b32 %r627|%p19, %r625, %r609, %r596, %r598;
mov.b32 %f70, %r627;
add.ftz.f32 %f14, %f69, %f70;
add.s32 %r628, %r4, 1;
setp.lt.u32 %p20, %r628, %r19;
setp.eq.s32 %p21, %r2, 1;
and.pred %p2, %p21, %p20;
not.pred %p22, %p2;
@%p22 bra $L__BB243_10;

st.global.f32 [%rd15+4], %f14;

$L__BB243_10:
ld.shared.f32 %f71, [%r15+256];
add.ftz.f32 %f72, %f71, %f106;
ld.shared.f32 %f73, [%r15+768];
add.ftz.f32 %f74, %f73, %f72;
ld.shared.f32 %f75, [%r15+1280];
add.ftz.f32 %f76, %f75, %f74;
mov.b32 %r629, %f76;
mov.u32 %r630, 31;
mov.u32 %r631, 16;
mov.u32 %r632, -1;
shfl.sync.bfly.b32 %r633|%p23, %r629, %r631, %r630, %r632;
mov.b32 %f77, %r633;
add.ftz.f32 %f78, %f76, %f77;
mov.b32 %r634, %f78;
mov.u32 %r635, 8;
shfl.sync.bfly.b32 %r636|%p24, %r634, %r635, %r630, %r632;
mov.b32 %f79, %r636;
add.ftz.f32 %f80, %f78, %f79;
mov.b32 %r637, %f80;
mov.u32 %r638, 4;
shfl.sync.bfly.b32 %r639|%p25, %r637, %r638, %r630, %r632;
mov.b32 %f81, %r639;
add.ftz.f32 %f82, %f80, %f81;
mov.b32 %r640, %f82;
mov.u32 %r641, 2;
shfl.sync.bfly.b32 %r642|%p26, %r640, %r641, %r630, %r632;
mov.b32 %f83, %r642;
add.ftz.f32 %f84, %f82, %f83;
mov.b32 %r643, %f84;
mov.u32 %r644, 1;
shfl.sync.bfly.b32 %r645|%p27, %r643, %r644, %r630, %r632;
mov.b32 %f85, %r645;
add.ftz.f32 %f15, %f84, %f85;
@%p14 bra $L__BB243_12;

cvt.u64.u32 %rd86, %r19;
add.s64 %rd87, %rd86, %rd14;
shl.b64 %rd88, %rd87, 2;
add.s64 %rd89, %rd1, %rd88;
st.global.f32 [%rd89], %f15;

$L__BB243_12:
ld.shared.f32 %f86, [%r15+384];
add.ftz.f32 %f87, %f86, %f105;
ld.shared.f32 %f88, [%r15+896];
add.ftz.f32 %f89, %f88, %f87;
ld.shared.f32 %f90, [%r15+1408];
add.ftz.f32 %f91, %f90, %f89;
mov.b32 %r646, %f91;
shfl.sync.bfly.b32 %r650|%p29, %r646, %r631, %r630, %r632;
mov.b32 %f92, %r650;
add.ftz.f32 %f93, %f91, %f92;
mov.b32 %r651, %f93;
shfl.sync.bfly.b32 %r653|%p30, %r651, %r635, %r630, %r632;
mov.b32 %f94, %r653;
add.ftz.f32 %f95, %f93, %f94;
mov.b32 %r654, %f95;
shfl.sync.bfly.b32 %r656|%p31, %r654, %r638, %r630, %r632;
mov.b32 %f96, %r656;
add.ftz.f32 %f97, %f95, %f96;
mov.b32 %r657, %f97;
shfl.sync.bfly.b32 %r659|%p32, %r657, %r641, %r630, %r632;
mov.b32 %f98, %r659;
add.ftz.f32 %f99, %f97, %f98;
mov.b32 %r660, %f99;
shfl.sync.bfly.b32 %r662|%p33, %r660, %r644, %r630, %r632;
mov.b32 %f100, %r662;
add.ftz.f32 %f16, %f99, %f100;
@%p22 bra $L__BB243_14;

add.s32 %r663, %r19, 1;
cvt.u64.u32 %rd90, %r663;
add.s64 %rd91, %rd90, %rd14;
shl.b64 %rd92, %rd91, 2;
add.s64 %rd93, %rd1, %rd92;
st.global.f32 [%rd93], %f16;

$L__BB243_14:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type18ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type18ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type18ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type18ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type18ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type18ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type18ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type18ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type18ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<47>;
.reg .b16 %rs<35>;
.reg .f32 %f<163>;
.reg .b32 %r<882>;
.reg .b64 %rd<104>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type18ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[2304];

ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type18ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type18ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type18ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type18ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type18ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type18ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type18ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type18ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type18ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type18ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type18ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type18ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type18ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type18ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type18ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type18ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type18ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
cvta.to.global.u64 %rd1, %rd19;
mov.u32 %r1, %tid.y;
shl.b32 %r36, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r36, %r2;
mov.u32 %r37, %ctaid.x;
shl.b32 %r4, %r37, 1;
shr.s32 %r38, %r3, 31;
shr.u32 %r39, %r38, 29;
add.s32 %r40, %r3, %r39;
shr.s32 %r881, %r40, 3;
shr.u32 %r6, %r35, 8;
setp.ge.s32 %p3, %r881, %r6;
mov.f32 %f157, 0f00000000;
mov.f32 %f158, %f157;
mov.f32 %f159, %f157;
mov.f32 %f160, %f157;
mov.f32 %f161, %f157;
mov.f32 %f162, %f157;
@%p3 bra $L__BB244_3;

mov.u32 %r41, %ctaid.y;
mov.u32 %r42, %ctaid.z;
mul.lo.s32 %r43, %r41, %r27;
cvt.u64.u32 %rd20, %r43;
mul.lo.s32 %r44, %r42, %r33;
cvt.u64.u32 %rd21, %r44;
add.s64 %rd2, %rd21, %rd20;
mul.hi.u32 %r45, %r42, %r29;
add.s32 %r46, %r45, %r42;
shr.u32 %r47, %r46, %r30;
mul.lo.s32 %r48, %r47, %r32;
mul.hi.u32 %r49, %r41, %r23;
add.s32 %r50, %r49, %r41;
shr.u32 %r51, %r50, %r24;
mul.lo.s32 %r52, %r51, %r26;
cvta.to.global.u64 %rd103, %rd17;
and.b32 %r56, %r40, -8;
sub.s32 %r57, %r3, %r56;
shl.b32 %r58, %r57, 2;
shl.b32 %r59, %r57, 1;
add.s32 %r60, %r59, 33;
cvt.s64.s32 %rd4, %r57;
shl.b32 %r61, %r21, 1;
shl.b32 %r878, %r881, 3;
add.s32 %r880, %r878, %r61;
add.s32 %r879, %r21, %r878;
add.s32 %r62, %r881, %r48;
add.s32 %r63, %r62, %r52;
add.s32 %r64, %r4, 1;
mad.lo.s32 %r65, %r20, %r64, %r63;
mul.wide.s32 %rd5, %r65, 98;
mad.lo.s32 %r66, %r4, %r20, %r881;
add.s32 %r67, %r66, %r48;
add.s32 %r68, %r67, %r52;
mul.wide.s32 %rd6, %r68, 98;
mul.wide.s32 %rd22, %r58, 2;
add.s64 %rd23, %rd6, %rd22;
add.s64 %rd7, %rd23, 8;
mul.wide.s32 %rd24, %r60, 2;
add.s64 %rd25, %rd24, -2;
add.s64 %rd26, %rd25, %rd5;
add.s64 %rd8, %rd26, 2;
add.s64 %rd27, %rd25, %rd6;
add.s64 %rd9, %rd27, 2;
add.s64 %rd28, %rd5, %rd22;
add.s64 %rd10, %rd28, 4;
cvta.to.global.u64 %rd11, %rd18;
mov.f32 %f157, 0f00000000;

$L__BB244_2:
cvt.u64.u32 %rd29, %r878;
add.s64 %rd30, %rd2, %rd29;
add.s64 %rd31, %rd30, %rd4;
mul.lo.s64 %rd32, %rd31, 36;
add.s64 %rd33, %rd103, %rd7;
ld.global.nc.u16 %rs7, [%rd33+-4];
ld.global.nc.u16 %rs8, [%rd33+-2];
ld.global.nc.u16 %rs9, [%rd33];
add.s64 %rd34, %rd103, %rd9;
ld.global.nc.u16 %rs10, [%rd34];
cvt.u32.u16 %r363, %rs10;
ld.global.nc.u16 %rs11, [%rd34+2];
cvt.u32.u16 %r364, %rs11;
mov.b32 %r365, {%rs10, %rs11};
ld.global.nc.u16 %rs12, [%rd33+-6];
shl.b16 %rs13, %rs12, 2;
cvt.u64.u16 %rd35, %rs13;
and.b64 %rd36, %rd35, 1020;
mov.u64 %rd37, iq3xxs_grid;
add.s64 %rd38, %rd37, %rd36;
ld.global.nc.u32 %r366, [%rd38];
not.b32 %r367, %r366;
shr.u16 %rs14, %rs12, 8;
cvt.u32.u16 %r368, %rs14;
mul.wide.u32 %rd39, %r368, 4;
add.s64 %rd40, %rd37, %rd39;
ld.global.nc.u32 %r369, [%rd40];
not.b32 %r370, %r369;
and.b32 %r371, %r363, 255;
popc.b32 %r372, %r371;
shl.b32 %r373, %r372, 7;
and.b32 %r374, %r373, 128;
xor.b32 %r375, %r374, %r371;
mul.lo.s32 %r376, %r375, 16843009;
and.b32 %r377, %r376, 134480385;
or.b32 %r378, %r377, -2139062144;
add.s32 %r379, %r378, -16843009;
or.b32 %r265, %r379, %r377;
mov.u32 %r69, %r265;

prmt.b32 %r69, %r69, 0, 0xba98;

xor.b32 %r380, %r69, %r366;
or.b32 %r381, %r380, -2139062144;
and.b32 %r382, %r69, 2139062143;
and.b32 %r383, %r367, -2139062144;
sub.s32 %r384, %r381, %r382;
xor.b32 %r74, %r383, %r384;
add.s64 %rd41, %rd11, %rd32;
ld.global.nc.u32 %r75, [%rd41+4];
and.b32 %r385, %r376, -2143281136;
or.b32 %r386, %r385, -2139062144;
add.s32 %r387, %r386, -16843009;
or.b32 %r267, %r387, %r385;
mov.u32 %r71, %r267;

prmt.b32 %r71, %r71, 0, 0xba98;

xor.b32 %r388, %r71, %r369;
or.b32 %r389, %r388, -2139062144;
and.b32 %r390, %r71, 2139062143;
and.b32 %r391, %r370, -2139062144;
sub.s32 %r392, %r389, %r390;
xor.b32 %r78, %r391, %r392;
ld.global.nc.u32 %r79, [%rd41+8];
mov.u32 %r321, 0;

dp4a.s32.s32 %r73, %r74, %r75, %r321;


dp4a.s32.s32 %r77, %r78, %r79, %r73;

shl.b16 %rs15, %rs7, 2;
cvt.u64.u16 %rd42, %rs15;
and.b64 %rd43, %rd42, 1020;
add.s64 %rd44, %rd37, %rd43;
ld.global.nc.u32 %r393, [%rd44];
not.b32 %r394, %r393;
shr.u16 %rs16, %rs7, 8;
cvt.u32.u16 %r395, %rs16;
mul.wide.u32 %rd45, %r395, 4;
add.s64 %rd46, %rd37, %rd45;
ld.global.nc.u32 %r396, [%rd46];
not.b32 %r397, %r396;
shr.u32 %r398, %r363, 7;
and.b32 %r399, %r398, 255;
popc.b32 %r400, %r399;
shl.b32 %r401, %r400, 7;
and.b32 %r402, %r401, 128;
xor.b32 %r403, %r402, %r399;
mul.lo.s32 %r404, %r403, 16843009;
and.b32 %r405, %r404, 134480385;
or.b32 %r406, %r405, -2139062144;
add.s32 %r407, %r406, -16843009;
or.b32 %r277, %r407, %r405;
mov.u32 %r81, %r277;

prmt.b32 %r81, %r81, 0, 0xba98;

xor.b32 %r408, %r81, %r393;
or.b32 %r409, %r408, -2139062144;
and.b32 %r410, %r81, 2139062143;
and.b32 %r411, %r394, -2139062144;
sub.s32 %r412, %r409, %r410;
xor.b32 %r86, %r411, %r412;
ld.global.nc.u32 %r87, [%rd41+12];
and.b32 %r413, %r404, -2143281136;
or.b32 %r414, %r413, -2139062144;
add.s32 %r415, %r414, -16843009;
or.b32 %r279, %r415, %r413;
mov.u32 %r83, %r279;

prmt.b32 %r83, %r83, 0, 0xba98;

xor.b32 %r416, %r83, %r396;
or.b32 %r417, %r416, -2139062144;
and.b32 %r418, %r83, 2139062143;
and.b32 %r419, %r397, -2139062144;
sub.s32 %r420, %r417, %r418;
xor.b32 %r90, %r419, %r420;
ld.global.nc.u32 %r91, [%rd41+16];

dp4a.s32.s32 %r85, %r86, %r87, %r77;


dp4a.s32.s32 %r89, %r90, %r91, %r85;

shl.b16 %rs17, %rs8, 2;
cvt.u64.u16 %rd47, %rs17;
and.b64 %rd48, %rd47, 1020;
add.s64 %rd49, %rd37, %rd48;
ld.global.nc.u32 %r421, [%rd49];
not.b32 %r422, %r421;
shr.u16 %rs18, %rs8, 8;
cvt.u32.u16 %r423, %rs18;
mul.wide.u32 %rd50, %r423, 4;
add.s64 %rd51, %rd37, %rd50;
ld.global.nc.u32 %r424, [%rd51];
not.b32 %r425, %r424;
shr.u32 %r426, %r365, 14;
and.b32 %r427, %r426, 255;
popc.b32 %r428, %r427;
shl.b32 %r429, %r428, 7;
and.b32 %r430, %r429, 128;
xor.b32 %r431, %r430, %r427;
mul.lo.s32 %r432, %r431, 16843009;
and.b32 %r433, %r432, 134480385;
or.b32 %r434, %r433, -2139062144;
add.s32 %r435, %r434, -16843009;
or.b32 %r289, %r435, %r433;
mov.u32 %r93, %r289;

prmt.b32 %r93, %r93, 0, 0xba98;

xor.b32 %r436, %r93, %r421;
or.b32 %r437, %r436, -2139062144;
and.b32 %r438, %r93, 2139062143;
and.b32 %r439, %r422, -2139062144;
sub.s32 %r440, %r437, %r438;
xor.b32 %r98, %r439, %r440;
ld.global.nc.u32 %r99, [%rd41+20];
and.b32 %r441, %r432, -2143281136;
or.b32 %r442, %r441, -2139062144;
add.s32 %r443, %r442, -16843009;
or.b32 %r291, %r443, %r441;
mov.u32 %r95, %r291;

prmt.b32 %r95, %r95, 0, 0xba98;

xor.b32 %r444, %r95, %r424;
or.b32 %r445, %r444, -2139062144;
and.b32 %r446, %r95, 2139062143;
and.b32 %r447, %r425, -2139062144;
sub.s32 %r448, %r445, %r446;
xor.b32 %r102, %r447, %r448;
ld.global.nc.u32 %r103, [%rd41+24];

dp4a.s32.s32 %r97, %r98, %r99, %r89;


dp4a.s32.s32 %r101, %r102, %r103, %r97;

shl.b16 %rs19, %rs9, 2;
cvt.u64.u16 %rd52, %rs19;
and.b64 %rd53, %rd52, 1020;
add.s64 %rd54, %rd37, %rd53;
ld.global.nc.u32 %r449, [%rd54];
not.b32 %r450, %r449;
shr.u16 %rs20, %rs9, 8;
cvt.u32.u16 %r451, %rs20;
mul.wide.u32 %rd55, %r451, 4;
add.s64 %rd56, %rd37, %rd55;
ld.global.nc.u32 %r452, [%rd56];
not.b32 %r453, %r452;
shr.u32 %r454, %r364, 5;
and.b32 %r455, %r454, 255;
popc.b32 %r456, %r455;
shl.b32 %r457, %r456, 7;
and.b32 %r458, %r457, 128;
xor.b32 %r459, %r458, %r455;
mul.lo.s32 %r460, %r459, 16843009;
and.b32 %r461, %r460, 134480385;
or.b32 %r462, %r461, -2139062144;
add.s32 %r463, %r462, -16843009;
or.b32 %r301, %r463, %r461;
mov.u32 %r105, %r301;

prmt.b32 %r105, %r105, 0, 0xba98;

xor.b32 %r464, %r105, %r449;
or.b32 %r465, %r464, -2139062144;
and.b32 %r466, %r105, 2139062143;
and.b32 %r467, %r450, -2139062144;
sub.s32 %r468, %r465, %r466;
xor.b32 %r110, %r467, %r468;
ld.global.nc.u32 %r111, [%rd41+28];
and.b32 %r469, %r460, -2143281136;
or.b32 %r470, %r469, -2139062144;
add.s32 %r471, %r470, -16843009;
or.b32 %r303, %r471, %r469;
mov.u32 %r107, %r303;

prmt.b32 %r107, %r107, 0, 0xba98;

xor.b32 %r472, %r107, %r452;
or.b32 %r473, %r472, -2139062144;
and.b32 %r474, %r107, 2139062143;
and.b32 %r475, %r453, -2139062144;
sub.s32 %r476, %r473, %r474;
xor.b32 %r114, %r475, %r476;
ld.global.nc.u32 %r115, [%rd41+32];

dp4a.s32.s32 %r109, %r110, %r111, %r101;

shr.u32 %r477, %r364, 12;

dp4a.s32.s32 %r113, %r114, %r115, %r109;

shr.u32 %r478, %r113, 31;
add.s32 %r479, %r113, %r478;
shr.s32 %r480, %r479, 1;
mad.lo.s32 %r481, %r113, %r477, %r480;
shr.u32 %r482, %r481, 31;
add.s32 %r483, %r481, %r482;
shr.s32 %r484, %r483, 1;
add.s64 %rd57, %rd103, %rd6;
ld.global.nc.u16 %rs1, [%rd57];

{ cvt.f32.f16 %f37, %rs1;}


ld.global.nc.u32 %r117, [%rd41];

{.reg .f16 low,high;
mov.b32 {low,high},%r117;
cvt.f32.f16 %f38, low;}


mul.ftz.f32 %f49, %f37, %f38;
cvt.rn.f32.s32 %f50, %r484;
fma.rn.ftz.f32 %f162, %f49, %f50, %f162;
add.s64 %rd58, %rd103, %rd10;
ld.global.nc.u16 %rs21, [%rd58];
ld.global.nc.u16 %rs22, [%rd58+2];
ld.global.nc.u16 %rs23, [%rd58+4];
add.s64 %rd59, %rd103, %rd8;
ld.global.nc.u16 %rs24, [%rd59];
cvt.u32.u16 %r485, %rs24;
ld.global.nc.u16 %rs25, [%rd59+2];
cvt.u32.u16 %r486, %rs25;
mov.b32 %r487, {%rs24, %rs25};
ld.global.nc.u16 %rs26, [%rd58+-2];
shl.b16 %rs27, %rs26, 2;
cvt.u64.u16 %rd60, %rs27;
and.b64 %rd61, %rd60, 1020;
add.s64 %rd62, %rd37, %rd61;
ld.global.nc.u32 %r488, [%rd62];
not.b32 %r489, %r488;
shr.u16 %rs28, %rs26, 8;
cvt.u32.u16 %r490, %rs28;
mul.wide.u32 %rd63, %r490, 4;
add.s64 %rd64, %rd37, %rd63;
ld.global.nc.u32 %r491, [%rd64];
not.b32 %r492, %r491;
and.b32 %r493, %r485, 255;
popc.b32 %r494, %r493;
shl.b32 %r495, %r494, 7;
and.b32 %r496, %r495, 128;
xor.b32 %r497, %r496, %r493;
mul.lo.s32 %r498, %r497, 16843009;
and.b32 %r499, %r498, 134480385;
or.b32 %r500, %r499, -2139062144;
add.s32 %r501, %r500, -16843009;
or.b32 %r314, %r501, %r499;
mov.u32 %r118, %r314;

prmt.b32 %r118, %r118, 0, 0xba98;

xor.b32 %r502, %r118, %r488;
or.b32 %r503, %r502, -2139062144;
and.b32 %r504, %r118, 2139062143;
and.b32 %r505, %r489, -2139062144;
sub.s32 %r506, %r503, %r504;
xor.b32 %r123, %r505, %r506;
and.b32 %r507, %r498, -2143281136;
or.b32 %r508, %r507, -2139062144;
add.s32 %r509, %r508, -16843009;
or.b32 %r316, %r509, %r507;
mov.u32 %r120, %r316;

prmt.b32 %r120, %r120, 0, 0xba98;

xor.b32 %r510, %r120, %r491;
or.b32 %r511, %r510, -2139062144;
and.b32 %r512, %r120, 2139062143;
and.b32 %r513, %r492, -2139062144;
sub.s32 %r514, %r511, %r512;
xor.b32 %r127, %r513, %r514;

dp4a.s32.s32 %r122, %r123, %r75, %r321;


dp4a.s32.s32 %r126, %r127, %r79, %r122;

shl.b16 %rs29, %rs21, 2;
cvt.u64.u16 %rd65, %rs29;
and.b64 %rd66, %rd65, 1020;
add.s64 %rd67, %rd37, %rd66;
ld.global.nc.u32 %r515, [%rd67];
not.b32 %r516, %r515;
shr.u16 %rs30, %rs21, 8;
cvt.u32.u16 %r517, %rs30;
mul.wide.u32 %rd68, %r517, 4;
add.s64 %rd69, %rd37, %rd68;
ld.global.nc.u32 %r518, [%rd69];
not.b32 %r519, %r518;
shr.u32 %r520, %r485, 7;
and.b32 %r521, %r520, 255;
popc.b32 %r522, %r521;
shl.b32 %r523, %r522, 7;
and.b32 %r524, %r523, 128;
xor.b32 %r525, %r524, %r521;
mul.lo.s32 %r526, %r525, 16843009;
and.b32 %r527, %r526, 134480385;
or.b32 %r528, %r527, -2139062144;
add.s32 %r529, %r528, -16843009;
or.b32 %r326, %r529, %r527;
mov.u32 %r130, %r326;

prmt.b32 %r130, %r130, 0, 0xba98;

xor.b32 %r530, %r130, %r515;
or.b32 %r531, %r530, -2139062144;
and.b32 %r532, %r130, 2139062143;
and.b32 %r533, %r516, -2139062144;
sub.s32 %r534, %r531, %r532;
xor.b32 %r135, %r533, %r534;
and.b32 %r535, %r526, -2143281136;
or.b32 %r536, %r535, -2139062144;
add.s32 %r537, %r536, -16843009;
or.b32 %r328, %r537, %r535;
mov.u32 %r132, %r328;

prmt.b32 %r132, %r132, 0, 0xba98;

xor.b32 %r538, %r132, %r518;
or.b32 %r539, %r538, -2139062144;
and.b32 %r540, %r132, 2139062143;
and.b32 %r541, %r519, -2139062144;
sub.s32 %r542, %r539, %r540;
xor.b32 %r139, %r541, %r542;

dp4a.s32.s32 %r134, %r135, %r87, %r126;


dp4a.s32.s32 %r138, %r139, %r91, %r134;

shl.b16 %rs31, %rs22, 2;
cvt.u64.u16 %rd70, %rs31;
and.b64 %rd71, %rd70, 1020;
add.s64 %rd72, %rd37, %rd71;
ld.global.nc.u32 %r543, [%rd72];
not.b32 %r544, %r543;
shr.u16 %rs32, %rs22, 8;
cvt.u32.u16 %r545, %rs32;
mul.wide.u32 %rd73, %r545, 4;
add.s64 %rd74, %rd37, %rd73;
ld.global.nc.u32 %r546, [%rd74];
not.b32 %r547, %r546;
shr.u32 %r548, %r487, 14;
and.b32 %r549, %r548, 255;
popc.b32 %r550, %r549;
shl.b32 %r551, %r550, 7;
and.b32 %r552, %r551, 128;
xor.b32 %r553, %r552, %r549;
mul.lo.s32 %r554, %r553, 16843009;
and.b32 %r555, %r554, 134480385;
or.b32 %r556, %r555, -2139062144;
add.s32 %r557, %r556, -16843009;
or.b32 %r338, %r557, %r555;
mov.u32 %r142, %r338;

prmt.b32 %r142, %r142, 0, 0xba98;

xor.b32 %r558, %r142, %r543;
or.b32 %r559, %r558, -2139062144;
and.b32 %r560, %r142, 2139062143;
and.b32 %r561, %r544, -2139062144;
sub.s32 %r562, %r559, %r560;
xor.b32 %r147, %r561, %r562;
and.b32 %r563, %r554, -2143281136;
or.b32 %r564, %r563, -2139062144;
add.s32 %r565, %r564, -16843009;
or.b32 %r340, %r565, %r563;
mov.u32 %r144, %r340;

prmt.b32 %r144, %r144, 0, 0xba98;

xor.b32 %r566, %r144, %r546;
or.b32 %r567, %r566, -2139062144;
and.b32 %r568, %r144, 2139062143;
and.b32 %r569, %r547, -2139062144;
sub.s32 %r570, %r567, %r568;
xor.b32 %r151, %r569, %r570;

dp4a.s32.s32 %r146, %r147, %r99, %r138;


dp4a.s32.s32 %r150, %r151, %r103, %r146;

shl.b16 %rs33, %rs23, 2;
cvt.u64.u16 %rd75, %rs33;
and.b64 %rd76, %rd75, 1020;
add.s64 %rd77, %rd37, %rd76;
ld.global.nc.u32 %r571, [%rd77];
not.b32 %r572, %r571;
shr.u16 %rs34, %rs23, 8;
cvt.u32.u16 %r573, %rs34;
mul.wide.u32 %rd78, %r573, 4;
add.s64 %rd79, %rd37, %rd78;
ld.global.nc.u32 %r574, [%rd79];
not.b32 %r575, %r574;
shr.u32 %r576, %r486, 5;
and.b32 %r577, %r576, 255;
popc.b32 %r578, %r577;
shl.b32 %r579, %r578, 7;
and.b32 %r580, %r579, 128;
xor.b32 %r581, %r580, %r577;
mul.lo.s32 %r582, %r581, 16843009;
and.b32 %r583, %r582, 134480385;
or.b32 %r584, %r583, -2139062144;
add.s32 %r585, %r584, -16843009;
or.b32 %r350, %r585, %r583;
mov.u32 %r154, %r350;

prmt.b32 %r154, %r154, 0, 0xba98;

xor.b32 %r586, %r154, %r571;
or.b32 %r587, %r586, -2139062144;
and.b32 %r588, %r154, 2139062143;
and.b32 %r589, %r572, -2139062144;
sub.s32 %r590, %r587, %r588;
xor.b32 %r159, %r589, %r590;
and.b32 %r591, %r582, -2143281136;
or.b32 %r592, %r591, -2139062144;
add.s32 %r593, %r592, -16843009;
or.b32 %r352, %r593, %r591;
mov.u32 %r156, %r352;

prmt.b32 %r156, %r156, 0, 0xba98;

xor.b32 %r594, %r156, %r574;
or.b32 %r595, %r594, -2139062144;
and.b32 %r596, %r156, 2139062143;
and.b32 %r597, %r575, -2139062144;
sub.s32 %r598, %r595, %r596;
xor.b32 %r163, %r597, %r598;

dp4a.s32.s32 %r158, %r159, %r111, %r150;

shr.u32 %r599, %r486, 12;

dp4a.s32.s32 %r162, %r163, %r115, %r158;

shr.u32 %r600, %r162, 31;
add.s32 %r601, %r162, %r600;
shr.s32 %r602, %r601, 1;
mad.lo.s32 %r603, %r162, %r599, %r602;
shr.u32 %r604, %r603, 31;
add.s32 %r605, %r603, %r604;
shr.s32 %r606, %r605, 1;
add.s64 %rd80, %rd103, %rd5;
ld.global.nc.u16 %rs2, [%rd80];

{ cvt.f32.f16 %f39, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r117;
cvt.f32.f16 %f40, low;}


mul.ftz.f32 %f51, %f39, %f40;
cvt.rn.f32.s32 %f52, %r606;
fma.rn.ftz.f32 %f161, %f51, %f52, %f161;
cvt.u64.u32 %rd81, %r879;
add.s64 %rd82, %rd2, %rd81;
add.s64 %rd83, %rd82, %rd4;
mul.lo.s64 %rd84, %rd83, 36;
mov.u32 %r167, %r265;

prmt.b32 %r167, %r167, 0, 0xba98;

xor.b32 %r607, %r167, %r366;
or.b32 %r608, %r607, -2139062144;
and.b32 %r609, %r167, 2139062143;
sub.s32 %r610, %r608, %r609;
xor.b32 %r172, %r383, %r610;
add.s64 %rd85, %rd11, %rd84;
ld.global.nc.u32 %r173, [%rd85+4];
mov.u32 %r169, %r267;

prmt.b32 %r169, %r169, 0, 0xba98;

xor.b32 %r611, %r169, %r369;
or.b32 %r612, %r611, -2139062144;
and.b32 %r613, %r169, 2139062143;
sub.s32 %r614, %r612, %r613;
xor.b32 %r176, %r391, %r614;
ld.global.nc.u32 %r177, [%rd85+8];

dp4a.s32.s32 %r171, %r172, %r173, %r321;


dp4a.s32.s32 %r175, %r176, %r177, %r171;

mov.u32 %r179, %r277;

prmt.b32 %r179, %r179, 0, 0xba98;

xor.b32 %r615, %r179, %r393;
or.b32 %r616, %r615, -2139062144;
and.b32 %r617, %r179, 2139062143;
sub.s32 %r618, %r616, %r617;
xor.b32 %r184, %r411, %r618;
ld.global.nc.u32 %r185, [%rd85+12];
mov.u32 %r181, %r279;

prmt.b32 %r181, %r181, 0, 0xba98;

xor.b32 %r619, %r181, %r396;
or.b32 %r620, %r619, -2139062144;
and.b32 %r621, %r181, 2139062143;
sub.s32 %r622, %r620, %r621;
xor.b32 %r188, %r419, %r622;
ld.global.nc.u32 %r189, [%rd85+16];

dp4a.s32.s32 %r183, %r184, %r185, %r175;


dp4a.s32.s32 %r187, %r188, %r189, %r183;

mov.u32 %r191, %r289;

prmt.b32 %r191, %r191, 0, 0xba98;

xor.b32 %r623, %r191, %r421;
or.b32 %r624, %r623, -2139062144;
and.b32 %r625, %r191, 2139062143;
sub.s32 %r626, %r624, %r625;
xor.b32 %r196, %r439, %r626;
ld.global.nc.u32 %r197, [%rd85+20];
mov.u32 %r193, %r291;

prmt.b32 %r193, %r193, 0, 0xba98;

xor.b32 %r627, %r193, %r424;
or.b32 %r628, %r627, -2139062144;
and.b32 %r629, %r193, 2139062143;
sub.s32 %r630, %r628, %r629;
xor.b32 %r200, %r447, %r630;
ld.global.nc.u32 %r201, [%rd85+24];

dp4a.s32.s32 %r195, %r196, %r197, %r187;


dp4a.s32.s32 %r199, %r200, %r201, %r195;

mov.u32 %r203, %r301;

prmt.b32 %r203, %r203, 0, 0xba98;

xor.b32 %r631, %r203, %r449;
or.b32 %r632, %r631, -2139062144;
and.b32 %r633, %r203, 2139062143;
sub.s32 %r634, %r632, %r633;
xor.b32 %r208, %r467, %r634;
ld.global.nc.u32 %r209, [%rd85+28];
mov.u32 %r205, %r303;

prmt.b32 %r205, %r205, 0, 0xba98;

xor.b32 %r635, %r205, %r452;
or.b32 %r636, %r635, -2139062144;
and.b32 %r637, %r205, 2139062143;
sub.s32 %r638, %r636, %r637;
xor.b32 %r212, %r475, %r638;
ld.global.nc.u32 %r213, [%rd85+32];

dp4a.s32.s32 %r207, %r208, %r209, %r199;


dp4a.s32.s32 %r211, %r212, %r213, %r207;

shr.u32 %r639, %r211, 31;
add.s32 %r640, %r211, %r639;
shr.s32 %r641, %r640, 1;
mad.lo.s32 %r642, %r211, %r477, %r641;
shr.u32 %r643, %r642, 31;
add.s32 %r644, %r642, %r643;
shr.s32 %r645, %r644, 1;

{ cvt.f32.f16 %f41, %rs1;}


ld.global.nc.u32 %r215, [%rd85];

{.reg .f16 low,high;
mov.b32 {low,high},%r215;
cvt.f32.f16 %f42, low;}


mul.ftz.f32 %f53, %f41, %f42;
cvt.rn.f32.s32 %f54, %r645;
fma.rn.ftz.f32 %f160, %f53, %f54, %f160;
mov.u32 %r216, %r314;

prmt.b32 %r216, %r216, 0, 0xba98;

xor.b32 %r646, %r216, %r488;
or.b32 %r647, %r646, -2139062144;
and.b32 %r648, %r216, 2139062143;
sub.s32 %r649, %r647, %r648;
xor.b32 %r221, %r505, %r649;
mov.u32 %r218, %r316;

prmt.b32 %r218, %r218, 0, 0xba98;

xor.b32 %r650, %r218, %r491;
or.b32 %r651, %r650, -2139062144;
and.b32 %r652, %r218, 2139062143;
sub.s32 %r653, %r651, %r652;
xor.b32 %r225, %r513, %r653;

dp4a.s32.s32 %r220, %r221, %r173, %r321;


dp4a.s32.s32 %r224, %r225, %r177, %r220;

mov.u32 %r228, %r326;

prmt.b32 %r228, %r228, 0, 0xba98;

xor.b32 %r654, %r228, %r515;
or.b32 %r655, %r654, -2139062144;
and.b32 %r656, %r228, 2139062143;
sub.s32 %r657, %r655, %r656;
xor.b32 %r233, %r533, %r657;
mov.u32 %r230, %r328;

prmt.b32 %r230, %r230, 0, 0xba98;

xor.b32 %r658, %r230, %r518;
or.b32 %r659, %r658, -2139062144;
and.b32 %r660, %r230, 2139062143;
sub.s32 %r661, %r659, %r660;
xor.b32 %r237, %r541, %r661;

dp4a.s32.s32 %r232, %r233, %r185, %r224;


dp4a.s32.s32 %r236, %r237, %r189, %r232;

mov.u32 %r240, %r338;

prmt.b32 %r240, %r240, 0, 0xba98;

xor.b32 %r662, %r240, %r543;
or.b32 %r663, %r662, -2139062144;
and.b32 %r664, %r240, 2139062143;
sub.s32 %r665, %r663, %r664;
xor.b32 %r245, %r561, %r665;
mov.u32 %r242, %r340;

prmt.b32 %r242, %r242, 0, 0xba98;

xor.b32 %r666, %r242, %r546;
or.b32 %r667, %r666, -2139062144;
and.b32 %r668, %r242, 2139062143;
sub.s32 %r669, %r667, %r668;
xor.b32 %r249, %r569, %r669;

dp4a.s32.s32 %r244, %r245, %r197, %r236;


dp4a.s32.s32 %r248, %r249, %r201, %r244;

mov.u32 %r252, %r350;

prmt.b32 %r252, %r252, 0, 0xba98;

xor.b32 %r670, %r252, %r571;
or.b32 %r671, %r670, -2139062144;
and.b32 %r672, %r252, 2139062143;
sub.s32 %r673, %r671, %r672;
xor.b32 %r257, %r589, %r673;
mov.u32 %r254, %r352;

prmt.b32 %r254, %r254, 0, 0xba98;

xor.b32 %r674, %r254, %r574;
or.b32 %r675, %r674, -2139062144;
and.b32 %r676, %r254, 2139062143;
sub.s32 %r677, %r675, %r676;
xor.b32 %r261, %r597, %r677;

dp4a.s32.s32 %r256, %r257, %r209, %r248;


dp4a.s32.s32 %r260, %r261, %r213, %r256;

shr.u32 %r678, %r260, 31;
add.s32 %r679, %r260, %r678;
shr.s32 %r680, %r679, 1;
mad.lo.s32 %r681, %r260, %r599, %r680;
shr.u32 %r682, %r681, 31;
add.s32 %r683, %r681, %r682;
shr.s32 %r684, %r683, 1;

{ cvt.f32.f16 %f43, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r215;
cvt.f32.f16 %f44, low;}


mul.ftz.f32 %f55, %f43, %f44;
cvt.rn.f32.s32 %f56, %r684;
fma.rn.ftz.f32 %f159, %f55, %f56, %f159;
cvt.u64.u32 %rd86, %r880;
add.s64 %rd87, %rd2, %rd86;
add.s64 %rd88, %rd87, %rd4;
mul.lo.s64 %rd89, %rd88, 36;

prmt.b32 %r265, %r265, 0, 0xba98;

xor.b32 %r685, %r265, %r366;
or.b32 %r686, %r685, -2139062144;
and.b32 %r687, %r265, 2139062143;
sub.s32 %r688, %r686, %r687;
xor.b32 %r270, %r383, %r688;
add.s64 %rd90, %rd11, %rd89;
ld.global.nc.u32 %r271, [%rd90+4];

prmt.b32 %r267, %r267, 0, 0xba98;

xor.b32 %r689, %r267, %r369;
or.b32 %r690, %r689, -2139062144;
and.b32 %r691, %r267, 2139062143;
sub.s32 %r692, %r690, %r691;
xor.b32 %r274, %r391, %r692;
ld.global.nc.u32 %r275, [%rd90+8];

dp4a.s32.s32 %r269, %r270, %r271, %r321;


dp4a.s32.s32 %r273, %r274, %r275, %r269;


prmt.b32 %r277, %r277, 0, 0xba98;

xor.b32 %r693, %r277, %r393;
or.b32 %r694, %r693, -2139062144;
and.b32 %r695, %r277, 2139062143;
sub.s32 %r696, %r694, %r695;
xor.b32 %r282, %r411, %r696;
ld.global.nc.u32 %r283, [%rd90+12];

prmt.b32 %r279, %r279, 0, 0xba98;

xor.b32 %r697, %r279, %r396;
or.b32 %r698, %r697, -2139062144;
and.b32 %r699, %r279, 2139062143;
sub.s32 %r700, %r698, %r699;
xor.b32 %r286, %r419, %r700;
ld.global.nc.u32 %r287, [%rd90+16];

dp4a.s32.s32 %r281, %r282, %r283, %r273;


dp4a.s32.s32 %r285, %r286, %r287, %r281;


prmt.b32 %r289, %r289, 0, 0xba98;

xor.b32 %r701, %r289, %r421;
or.b32 %r702, %r701, -2139062144;
and.b32 %r703, %r289, 2139062143;
sub.s32 %r704, %r702, %r703;
xor.b32 %r294, %r439, %r704;
ld.global.nc.u32 %r295, [%rd90+20];

prmt.b32 %r291, %r291, 0, 0xba98;

xor.b32 %r705, %r291, %r424;
or.b32 %r706, %r705, -2139062144;
and.b32 %r707, %r291, 2139062143;
sub.s32 %r708, %r706, %r707;
xor.b32 %r298, %r447, %r708;
ld.global.nc.u32 %r299, [%rd90+24];

dp4a.s32.s32 %r293, %r294, %r295, %r285;


dp4a.s32.s32 %r297, %r298, %r299, %r293;


prmt.b32 %r301, %r301, 0, 0xba98;

xor.b32 %r709, %r301, %r449;
or.b32 %r710, %r709, -2139062144;
and.b32 %r711, %r301, 2139062143;
sub.s32 %r712, %r710, %r711;
xor.b32 %r306, %r467, %r712;
ld.global.nc.u32 %r307, [%rd90+28];

prmt.b32 %r303, %r303, 0, 0xba98;

xor.b32 %r713, %r303, %r452;
or.b32 %r714, %r713, -2139062144;
and.b32 %r715, %r303, 2139062143;
sub.s32 %r716, %r714, %r715;
xor.b32 %r310, %r475, %r716;
ld.global.nc.u32 %r311, [%rd90+32];

dp4a.s32.s32 %r305, %r306, %r307, %r297;


dp4a.s32.s32 %r309, %r310, %r311, %r305;

shr.u32 %r717, %r309, 31;
add.s32 %r718, %r309, %r717;
shr.s32 %r719, %r718, 1;
mad.lo.s32 %r720, %r309, %r477, %r719;
shr.u32 %r721, %r720, 31;
add.s32 %r722, %r720, %r721;
shr.s32 %r723, %r722, 1;

{ cvt.f32.f16 %f45, %rs1;}


ld.global.nc.u32 %r313, [%rd90];

{.reg .f16 low,high;
mov.b32 {low,high},%r313;
cvt.f32.f16 %f46, low;}


mul.ftz.f32 %f57, %f45, %f46;
cvt.rn.f32.s32 %f58, %r723;
fma.rn.ftz.f32 %f158, %f57, %f58, %f158;

prmt.b32 %r314, %r314, 0, 0xba98;

xor.b32 %r724, %r314, %r488;
or.b32 %r725, %r724, -2139062144;
and.b32 %r726, %r314, 2139062143;
sub.s32 %r727, %r725, %r726;
xor.b32 %r319, %r505, %r727;

prmt.b32 %r316, %r316, 0, 0xba98;

xor.b32 %r728, %r316, %r491;
or.b32 %r729, %r728, -2139062144;
and.b32 %r730, %r316, 2139062143;
sub.s32 %r731, %r729, %r730;
xor.b32 %r323, %r513, %r731;

dp4a.s32.s32 %r318, %r319, %r271, %r321;


dp4a.s32.s32 %r322, %r323, %r275, %r318;


prmt.b32 %r326, %r326, 0, 0xba98;

xor.b32 %r732, %r326, %r515;
or.b32 %r733, %r732, -2139062144;
and.b32 %r734, %r326, 2139062143;
sub.s32 %r735, %r733, %r734;
xor.b32 %r331, %r533, %r735;

prmt.b32 %r328, %r328, 0, 0xba98;

xor.b32 %r736, %r328, %r518;
or.b32 %r737, %r736, -2139062144;
and.b32 %r738, %r328, 2139062143;
sub.s32 %r739, %r737, %r738;
xor.b32 %r335, %r541, %r739;

dp4a.s32.s32 %r330, %r331, %r283, %r322;


dp4a.s32.s32 %r334, %r335, %r287, %r330;


prmt.b32 %r338, %r338, 0, 0xba98;

xor.b32 %r740, %r338, %r543;
or.b32 %r741, %r740, -2139062144;
and.b32 %r742, %r338, 2139062143;
sub.s32 %r743, %r741, %r742;
xor.b32 %r343, %r561, %r743;

prmt.b32 %r340, %r340, 0, 0xba98;

xor.b32 %r744, %r340, %r546;
or.b32 %r745, %r744, -2139062144;
and.b32 %r746, %r340, 2139062143;
sub.s32 %r747, %r745, %r746;
xor.b32 %r347, %r569, %r747;

dp4a.s32.s32 %r342, %r343, %r295, %r334;


dp4a.s32.s32 %r346, %r347, %r299, %r342;


prmt.b32 %r350, %r350, 0, 0xba98;

xor.b32 %r748, %r350, %r571;
or.b32 %r749, %r748, -2139062144;
and.b32 %r750, %r350, 2139062143;
sub.s32 %r751, %r749, %r750;
xor.b32 %r355, %r589, %r751;

prmt.b32 %r352, %r352, 0, 0xba98;

xor.b32 %r752, %r352, %r574;
or.b32 %r753, %r752, -2139062144;
and.b32 %r754, %r352, 2139062143;
sub.s32 %r755, %r753, %r754;
xor.b32 %r359, %r597, %r755;

dp4a.s32.s32 %r354, %r355, %r307, %r346;


dp4a.s32.s32 %r358, %r359, %r311, %r354;

shr.u32 %r756, %r358, 31;
add.s32 %r757, %r358, %r756;
shr.s32 %r758, %r757, 1;
mad.lo.s32 %r759, %r358, %r599, %r758;
shr.u32 %r760, %r759, 31;
add.s32 %r761, %r759, %r760;
shr.s32 %r762, %r761, 1;

{ cvt.f32.f16 %f47, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r313;
cvt.f32.f16 %f48, low;}


mul.ftz.f32 %f59, %f47, %f48;
cvt.rn.f32.s32 %f60, %r762;
fma.rn.ftz.f32 %f157, %f59, %f60, %f157;
add.s32 %r880, %r880, 128;
add.s32 %r879, %r879, 128;
add.s32 %r878, %r878, 128;
add.s64 %rd103, %rd103, 1568;
add.s32 %r881, %r881, 16;
setp.lt.s32 %p4, %r881, %r6;
@%p4 bra $L__BB244_2;

$L__BB244_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB244_5;

mov.u32 %r763, _ZZ13mul_mat_vec_qIL9ggml_type18ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r764, %r1, 768, %r763;
shl.b32 %r765, %r2, 2;
add.s32 %r766, %r764, %r765;
st.shared.f32 [%r766+-768], %f162;
st.shared.f32 [%r766+-640], %f161;
st.shared.f32 [%r766+-512], %f160;
st.shared.f32 [%r766+-384], %f159;
st.shared.f32 [%r766+-256], %f158;
st.shared.f32 [%r766+-128], %f157;

$L__BB244_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB244_18;

mov.u32 %r767, %ctaid.z;
mov.u32 %r768, %ctaid.y;
mad.lo.s32 %r769, %r768, %r28, %r4;
mad.lo.s32 %r770, %r767, %r34, %r769;
cvt.u64.u32 %rd14, %r770;
shl.b32 %r771, %r2, 2;
mov.u32 %r772, _ZZ13mul_mat_vec_qIL9ggml_type18ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r18, %r772, %r771;
mov.u32 %r773, 2;
ld.shared.f32 %f61, [%r18];
add.ftz.f32 %f62, %f61, %f162;
ld.shared.f32 %f63, [%r18+768];
add.ftz.f32 %f64, %f63, %f62;
ld.shared.f32 %f65, [%r18+1536];
add.ftz.f32 %f66, %f65, %f64;
mov.b32 %r774, %f66;
mov.u32 %r775, 31;
mov.u32 %r776, 16;
mov.u32 %r777, -1;
shfl.sync.bfly.b32 %r778|%p7, %r774, %r776, %r775, %r777;
mov.b32 %f67, %r778;
add.ftz.f32 %f68, %f66, %f67;
mov.b32 %r779, %f68;
mov.u32 %r780, 8;
shfl.sync.bfly.b32 %r781|%p8, %r779, %r780, %r775, %r777;
mov.b32 %f69, %r781;
add.ftz.f32 %f70, %f68, %f69;
mov.b32 %r782, %f70;
mov.u32 %r783, 4;
shfl.sync.bfly.b32 %r784|%p9, %r782, %r783, %r775, %r777;
mov.b32 %f71, %r784;
add.ftz.f32 %f72, %f70, %f71;
mov.b32 %r785, %f72;
shfl.sync.bfly.b32 %r786|%p10, %r785, %r773, %r775, %r777;
mov.b32 %f73, %r786;
add.ftz.f32 %f74, %f72, %f73;
mov.b32 %r787, %f74;
mov.u32 %r788, 1;
shfl.sync.bfly.b32 %r789|%p11, %r787, %r788, %r775, %r777;
mov.b32 %f75, %r789;
add.ftz.f32 %f19, %f74, %f75;
setp.lt.u32 %p12, %r4, %r22;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
mul.wide.u32 %rd91, %r770, 4;
add.s64 %rd15, %rd1, %rd91;
not.pred %p14, %p1;
@%p14 bra $L__BB244_8;

st.global.f32 [%rd15], %f19;

$L__BB244_8:
ld.shared.f32 %f76, [%r18+128];
add.ftz.f32 %f77, %f76, %f161;
ld.shared.f32 %f78, [%r18+896];
add.ftz.f32 %f79, %f78, %f77;
ld.shared.f32 %f80, [%r18+1664];
add.ftz.f32 %f81, %f80, %f79;
mov.b32 %r790, %f81;
shfl.sync.bfly.b32 %r794|%p15, %r790, %r776, %r775, %r777;
mov.b32 %f82, %r794;
add.ftz.f32 %f83, %f81, %f82;
mov.b32 %r795, %f83;
shfl.sync.bfly.b32 %r797|%p16, %r795, %r780, %r775, %r777;
mov.b32 %f84, %r797;
add.ftz.f32 %f85, %f83, %f84;
mov.b32 %r798, %f85;
shfl.sync.bfly.b32 %r800|%p17, %r798, %r783, %r775, %r777;
mov.b32 %f86, %r800;
add.ftz.f32 %f87, %f85, %f86;
mov.b32 %r801, %f87;
shfl.sync.bfly.b32 %r803|%p18, %r801, %r773, %r775, %r777;
mov.b32 %f88, %r803;
add.ftz.f32 %f89, %f87, %f88;
mov.b32 %r804, %f89;
shfl.sync.bfly.b32 %r806|%p19, %r804, %r788, %r775, %r777;
mov.b32 %f90, %r806;
add.ftz.f32 %f20, %f89, %f90;
add.s32 %r807, %r4, 1;
setp.lt.u32 %p20, %r807, %r22;
setp.eq.s32 %p21, %r2, 1;
and.pred %p2, %p21, %p20;
not.pred %p22, %p2;
@%p22 bra $L__BB244_10;

st.global.f32 [%rd15+4], %f20;

$L__BB244_10:
ld.shared.f32 %f91, [%r18+256];
add.ftz.f32 %f92, %f91, %f160;
ld.shared.f32 %f93, [%r18+1024];
add.ftz.f32 %f94, %f93, %f92;
ld.shared.f32 %f95, [%r18+1792];
add.ftz.f32 %f96, %f95, %f94;
mov.b32 %r808, %f96;
mov.u32 %r809, 31;
mov.u32 %r810, 16;
mov.u32 %r811, -1;
shfl.sync.bfly.b32 %r812|%p23, %r808, %r810, %r809, %r811;
mov.b32 %f97, %r812;
add.ftz.f32 %f98, %f96, %f97;
mov.b32 %r813, %f98;
mov.u32 %r814, 8;
shfl.sync.bfly.b32 %r815|%p24, %r813, %r814, %r809, %r811;
mov.b32 %f99, %r815;
add.ftz.f32 %f100, %f98, %f99;
mov.b32 %r816, %f100;
mov.u32 %r817, 4;
shfl.sync.bfly.b32 %r818|%p25, %r816, %r817, %r809, %r811;
mov.b32 %f101, %r818;
add.ftz.f32 %f102, %f100, %f101;
mov.b32 %r819, %f102;
mov.u32 %r820, 2;
shfl.sync.bfly.b32 %r821|%p26, %r819, %r820, %r809, %r811;
mov.b32 %f103, %r821;
add.ftz.f32 %f104, %f102, %f103;
mov.b32 %r822, %f104;
mov.u32 %r823, 1;
shfl.sync.bfly.b32 %r824|%p27, %r822, %r823, %r809, %r811;
mov.b32 %f105, %r824;
add.ftz.f32 %f21, %f104, %f105;
@%p14 bra $L__BB244_12;

cvt.u64.u32 %rd92, %r22;
add.s64 %rd93, %rd92, %rd14;
shl.b64 %rd94, %rd93, 2;
add.s64 %rd95, %rd1, %rd94;
st.global.f32 [%rd95], %f21;

$L__BB244_12:
ld.shared.f32 %f106, [%r18+384];
add.ftz.f32 %f107, %f106, %f159;
ld.shared.f32 %f108, [%r18+1152];
add.ftz.f32 %f109, %f108, %f107;
ld.shared.f32 %f110, [%r18+1920];
add.ftz.f32 %f111, %f110, %f109;
mov.b32 %r825, %f111;
shfl.sync.bfly.b32 %r829|%p29, %r825, %r810, %r809, %r811;
mov.b32 %f112, %r829;
add.ftz.f32 %f113, %f111, %f112;
mov.b32 %r830, %f113;
shfl.sync.bfly.b32 %r832|%p30, %r830, %r814, %r809, %r811;
mov.b32 %f114, %r832;
add.ftz.f32 %f115, %f113, %f114;
mov.b32 %r833, %f115;
shfl.sync.bfly.b32 %r835|%p31, %r833, %r817, %r809, %r811;
mov.b32 %f116, %r835;
add.ftz.f32 %f117, %f115, %f116;
mov.b32 %r836, %f117;
shfl.sync.bfly.b32 %r838|%p32, %r836, %r820, %r809, %r811;
mov.b32 %f118, %r838;
add.ftz.f32 %f119, %f117, %f118;
mov.b32 %r839, %f119;
shfl.sync.bfly.b32 %r841|%p33, %r839, %r823, %r809, %r811;
mov.b32 %f120, %r841;
add.ftz.f32 %f22, %f119, %f120;
@%p22 bra $L__BB244_14;

add.s32 %r842, %r22, 1;
cvt.u64.u32 %rd96, %r842;
add.s64 %rd97, %rd96, %rd14;
shl.b64 %rd98, %rd97, 2;
add.s64 %rd99, %rd1, %rd98;
st.global.f32 [%rd99], %f22;

$L__BB244_14:
shl.b32 %r843, %r22, 1;
mov.u32 %r844, 1;
ld.shared.f32 %f121, [%r18+512];
add.ftz.f32 %f122, %f121, %f158;
ld.shared.f32 %f123, [%r18+1280];
add.ftz.f32 %f124, %f123, %f122;
ld.shared.f32 %f125, [%r18+2048];
add.ftz.f32 %f126, %f125, %f124;
mov.b32 %r845, %f126;
mov.u32 %r846, 31;
mov.u32 %r847, 16;
mov.u32 %r848, -1;
shfl.sync.bfly.b32 %r849|%p35, %r845, %r847, %r846, %r848;
mov.b32 %f127, %r849;
add.ftz.f32 %f128, %f126, %f127;
mov.b32 %r850, %f128;
mov.u32 %r851, 8;
shfl.sync.bfly.b32 %r852|%p36, %r850, %r851, %r846, %r848;
mov.b32 %f129, %r852;
add.ftz.f32 %f130, %f128, %f129;
mov.b32 %r853, %f130;
mov.u32 %r854, 4;
shfl.sync.bfly.b32 %r855|%p37, %r853, %r854, %r846, %r848;
mov.b32 %f131, %r855;
add.ftz.f32 %f132, %f130, %f131;
mov.b32 %r856, %f132;
mov.u32 %r857, 2;
shfl.sync.bfly.b32 %r858|%p38, %r856, %r857, %r846, %r848;
mov.b32 %f133, %r858;
add.ftz.f32 %f134, %f132, %f133;
mov.b32 %r859, %f134;
shfl.sync.bfly.b32 %r860|%p39, %r859, %r844, %r846, %r848;
mov.b32 %f135, %r860;
add.ftz.f32 %f23, %f134, %f135;
cvt.u64.u32 %rd100, %r843;
add.s64 %rd101, %rd100, %rd14;
shl.b64 %rd102, %rd101, 2;
add.s64 %rd16, %rd1, %rd102;
@%p14 bra $L__BB244_16;

st.global.f32 [%rd16], %f23;

$L__BB244_16:
ld.shared.f32 %f136, [%r18+640];
add.ftz.f32 %f137, %f136, %f157;
ld.shared.f32 %f138, [%r18+1408];
add.ftz.f32 %f139, %f138, %f137;
ld.shared.f32 %f140, [%r18+2176];
add.ftz.f32 %f141, %f140, %f139;
mov.b32 %r861, %f141;
shfl.sync.bfly.b32 %r865|%p41, %r861, %r847, %r846, %r848;
mov.b32 %f142, %r865;
add.ftz.f32 %f143, %f141, %f142;
mov.b32 %r866, %f143;
shfl.sync.bfly.b32 %r868|%p42, %r866, %r851, %r846, %r848;
mov.b32 %f144, %r868;
add.ftz.f32 %f145, %f143, %f144;
mov.b32 %r869, %f145;
shfl.sync.bfly.b32 %r871|%p43, %r869, %r854, %r846, %r848;
mov.b32 %f146, %r871;
add.ftz.f32 %f147, %f145, %f146;
mov.b32 %r872, %f147;
shfl.sync.bfly.b32 %r874|%p44, %r872, %r857, %r846, %r848;
mov.b32 %f148, %r874;
add.ftz.f32 %f149, %f147, %f148;
mov.b32 %r875, %f149;
shfl.sync.bfly.b32 %r877|%p45, %r875, %r844, %r846, %r848;
mov.b32 %f150, %r877;
add.ftz.f32 %f24, %f149, %f150;
@%p22 bra $L__BB244_18;

st.global.f32 [%rd16+4], %f24;

$L__BB244_18:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type18ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type18ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type18ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type18ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type18ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type18ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type18ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type18ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type18ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<59>;
.reg .b16 %rs<37>;
.reg .f32 %f<217>;
.reg .b32 %r<1100>;
.reg .b64 %rd<117>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type18ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[3072];

ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type18ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type18ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type18ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type18ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type18ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type18ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type18ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type18ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type18ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type18ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type18ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type18ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type18ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type18ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type18ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type18ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type18ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
cvta.to.global.u64 %rd1, %rd19;
mov.u32 %r1, %tid.y;
shl.b32 %r40, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r40, %r2;
mov.u32 %r41, %ctaid.x;
shl.b32 %r4, %r41, 1;
shr.s32 %r42, %r3, 31;
shr.u32 %r43, %r42, 29;
add.s32 %r44, %r3, %r43;
shr.s32 %r1099, %r44, 3;
shr.u32 %r6, %r39, 8;
setp.ge.s32 %p3, %r1099, %r6;
mov.f32 %f209, 0f00000000;
mov.f32 %f210, %f209;
mov.f32 %f211, %f209;
mov.f32 %f212, %f209;
mov.f32 %f213, %f209;
mov.f32 %f214, %f209;
mov.f32 %f215, %f209;
mov.f32 %f216, %f209;
@%p3 bra $L__BB245_3;

mov.u32 %r45, %ctaid.y;
mov.u32 %r46, %ctaid.z;
mul.lo.s32 %r47, %r45, %r31;
cvt.u64.u32 %rd20, %r47;
mul.lo.s32 %r48, %r46, %r37;
cvt.u64.u32 %rd21, %r48;
add.s64 %rd2, %rd21, %rd20;
mul.hi.u32 %r49, %r46, %r33;
add.s32 %r50, %r49, %r46;
shr.u32 %r51, %r50, %r34;
mul.lo.s32 %r52, %r51, %r36;
mul.hi.u32 %r53, %r45, %r27;
add.s32 %r54, %r53, %r45;
shr.u32 %r55, %r54, %r28;
mul.lo.s32 %r56, %r55, %r30;
cvta.to.global.u64 %rd116, %rd17;
and.b32 %r60, %r44, -8;
sub.s32 %r61, %r3, %r60;
shl.b32 %r62, %r61, 2;
shl.b32 %r63, %r61, 1;
add.s32 %r64, %r63, 33;
cvt.s64.s32 %rd4, %r61;
shl.b32 %r65, %r25, 1;
shl.b32 %r1095, %r1099, 3;
add.s32 %r1098, %r1095, %r65;
mad.lo.s32 %r1097, %r25, 3, %r1095;
add.s32 %r1096, %r25, %r1095;
add.s32 %r66, %r1099, %r52;
add.s32 %r67, %r66, %r56;
add.s32 %r68, %r4, 1;
mad.lo.s32 %r69, %r24, %r68, %r67;
mul.wide.s32 %rd5, %r69, 98;
mad.lo.s32 %r70, %r4, %r24, %r1099;
add.s32 %r71, %r70, %r52;
add.s32 %r72, %r71, %r56;
mul.wide.s32 %rd6, %r72, 98;
mul.wide.s32 %rd22, %r62, 2;
add.s64 %rd23, %rd6, %rd22;
add.s64 %rd7, %rd23, 8;
mul.wide.s32 %rd24, %r64, 2;
add.s64 %rd25, %rd24, -2;
add.s64 %rd26, %rd25, %rd5;
add.s64 %rd8, %rd26, 2;
add.s64 %rd27, %rd25, %rd6;
add.s64 %rd9, %rd27, 2;
add.s64 %rd28, %rd5, %rd22;
add.s64 %rd10, %rd28, 4;
cvta.to.global.u64 %rd11, %rd18;
mov.f32 %f209, 0f00000000;

$L__BB245_2:
cvt.u64.u32 %rd29, %r1095;
add.s64 %rd30, %rd2, %rd29;
add.s64 %rd31, %rd30, %rd4;
mul.lo.s64 %rd32, %rd31, 36;
add.s64 %rd33, %rd116, %rd7;
ld.global.nc.u16 %rs9, [%rd33+-4];
ld.global.nc.u16 %rs10, [%rd33+-2];
ld.global.nc.u16 %rs11, [%rd33];
add.s64 %rd34, %rd116, %rd9;
ld.global.nc.u16 %rs12, [%rd34];
cvt.u32.u16 %r465, %rs12;
ld.global.nc.u16 %rs13, [%rd34+2];
cvt.u32.u16 %r466, %rs13;
mov.b32 %r467, {%rs12, %rs13};
ld.global.nc.u16 %rs14, [%rd33+-6];
shl.b16 %rs15, %rs14, 2;
cvt.u64.u16 %rd35, %rs15;
and.b64 %rd36, %rd35, 1020;
mov.u64 %rd37, iq3xxs_grid;
add.s64 %rd38, %rd37, %rd36;
ld.global.nc.u32 %r468, [%rd38];
not.b32 %r469, %r468;
shr.u16 %rs16, %rs14, 8;
cvt.u32.u16 %r470, %rs16;
mul.wide.u32 %rd39, %r470, 4;
add.s64 %rd40, %rd37, %rd39;
ld.global.nc.u32 %r471, [%rd40];
not.b32 %r472, %r471;
and.b32 %r473, %r465, 255;
popc.b32 %r474, %r473;
shl.b32 %r475, %r474, 7;
and.b32 %r476, %r475, 128;
xor.b32 %r477, %r476, %r473;
mul.lo.s32 %r478, %r477, 16843009;
and.b32 %r479, %r478, 134480385;
or.b32 %r480, %r479, -2139062144;
add.s32 %r481, %r480, -16843009;
or.b32 %r367, %r481, %r479;
mov.u32 %r73, %r367;

prmt.b32 %r73, %r73, 0, 0xba98;

xor.b32 %r482, %r73, %r468;
or.b32 %r483, %r482, -2139062144;
and.b32 %r484, %r73, 2139062143;
and.b32 %r485, %r469, -2139062144;
sub.s32 %r486, %r483, %r484;
xor.b32 %r78, %r485, %r486;
add.s64 %rd41, %rd11, %rd32;
ld.global.nc.u32 %r79, [%rd41+4];
and.b32 %r487, %r478, -2143281136;
or.b32 %r488, %r487, -2139062144;
add.s32 %r489, %r488, -16843009;
or.b32 %r369, %r489, %r487;
mov.u32 %r75, %r369;

prmt.b32 %r75, %r75, 0, 0xba98;

xor.b32 %r490, %r75, %r471;
or.b32 %r491, %r490, -2139062144;
and.b32 %r492, %r75, 2139062143;
and.b32 %r493, %r472, -2139062144;
sub.s32 %r494, %r491, %r492;
xor.b32 %r82, %r493, %r494;
ld.global.nc.u32 %r83, [%rd41+8];
mov.u32 %r423, 0;

dp4a.s32.s32 %r77, %r78, %r79, %r423;


dp4a.s32.s32 %r81, %r82, %r83, %r77;

shl.b16 %rs17, %rs9, 2;
cvt.u64.u16 %rd42, %rs17;
and.b64 %rd43, %rd42, 1020;
add.s64 %rd44, %rd37, %rd43;
ld.global.nc.u32 %r495, [%rd44];
not.b32 %r496, %r495;
shr.u16 %rs18, %rs9, 8;
cvt.u32.u16 %r497, %rs18;
mul.wide.u32 %rd45, %r497, 4;
add.s64 %rd46, %rd37, %rd45;
ld.global.nc.u32 %r498, [%rd46];
not.b32 %r499, %r498;
shr.u32 %r500, %r465, 7;
and.b32 %r501, %r500, 255;
popc.b32 %r502, %r501;
shl.b32 %r503, %r502, 7;
and.b32 %r504, %r503, 128;
xor.b32 %r505, %r504, %r501;
mul.lo.s32 %r506, %r505, 16843009;
and.b32 %r507, %r506, 134480385;
or.b32 %r508, %r507, -2139062144;
add.s32 %r509, %r508, -16843009;
or.b32 %r379, %r509, %r507;
mov.u32 %r85, %r379;

prmt.b32 %r85, %r85, 0, 0xba98;

xor.b32 %r510, %r85, %r495;
or.b32 %r511, %r510, -2139062144;
and.b32 %r512, %r85, 2139062143;
and.b32 %r513, %r496, -2139062144;
sub.s32 %r514, %r511, %r512;
xor.b32 %r90, %r513, %r514;
ld.global.nc.u32 %r91, [%rd41+12];
and.b32 %r515, %r506, -2143281136;
or.b32 %r516, %r515, -2139062144;
add.s32 %r517, %r516, -16843009;
or.b32 %r381, %r517, %r515;
mov.u32 %r87, %r381;

prmt.b32 %r87, %r87, 0, 0xba98;

xor.b32 %r518, %r87, %r498;
or.b32 %r519, %r518, -2139062144;
and.b32 %r520, %r87, 2139062143;
and.b32 %r521, %r499, -2139062144;
sub.s32 %r522, %r519, %r520;
xor.b32 %r94, %r521, %r522;
ld.global.nc.u32 %r95, [%rd41+16];

dp4a.s32.s32 %r89, %r90, %r91, %r81;


dp4a.s32.s32 %r93, %r94, %r95, %r89;

shl.b16 %rs19, %rs10, 2;
cvt.u64.u16 %rd47, %rs19;
and.b64 %rd48, %rd47, 1020;
add.s64 %rd49, %rd37, %rd48;
ld.global.nc.u32 %r523, [%rd49];
not.b32 %r524, %r523;
shr.u16 %rs20, %rs10, 8;
cvt.u32.u16 %r525, %rs20;
mul.wide.u32 %rd50, %r525, 4;
add.s64 %rd51, %rd37, %rd50;
ld.global.nc.u32 %r526, [%rd51];
not.b32 %r527, %r526;
shr.u32 %r528, %r467, 14;
and.b32 %r529, %r528, 255;
popc.b32 %r530, %r529;
shl.b32 %r531, %r530, 7;
and.b32 %r532, %r531, 128;
xor.b32 %r533, %r532, %r529;
mul.lo.s32 %r534, %r533, 16843009;
and.b32 %r535, %r534, 134480385;
or.b32 %r536, %r535, -2139062144;
add.s32 %r537, %r536, -16843009;
or.b32 %r391, %r537, %r535;
mov.u32 %r97, %r391;

prmt.b32 %r97, %r97, 0, 0xba98;

xor.b32 %r538, %r97, %r523;
or.b32 %r539, %r538, -2139062144;
and.b32 %r540, %r97, 2139062143;
and.b32 %r541, %r524, -2139062144;
sub.s32 %r542, %r539, %r540;
xor.b32 %r102, %r541, %r542;
ld.global.nc.u32 %r103, [%rd41+20];
and.b32 %r543, %r534, -2143281136;
or.b32 %r544, %r543, -2139062144;
add.s32 %r545, %r544, -16843009;
or.b32 %r393, %r545, %r543;
mov.u32 %r99, %r393;

prmt.b32 %r99, %r99, 0, 0xba98;

xor.b32 %r546, %r99, %r526;
or.b32 %r547, %r546, -2139062144;
and.b32 %r548, %r99, 2139062143;
and.b32 %r549, %r527, -2139062144;
sub.s32 %r550, %r547, %r548;
xor.b32 %r106, %r549, %r550;
ld.global.nc.u32 %r107, [%rd41+24];

dp4a.s32.s32 %r101, %r102, %r103, %r93;


dp4a.s32.s32 %r105, %r106, %r107, %r101;

shl.b16 %rs21, %rs11, 2;
cvt.u64.u16 %rd52, %rs21;
and.b64 %rd53, %rd52, 1020;
add.s64 %rd54, %rd37, %rd53;
ld.global.nc.u32 %r551, [%rd54];
not.b32 %r552, %r551;
shr.u16 %rs22, %rs11, 8;
cvt.u32.u16 %r553, %rs22;
mul.wide.u32 %rd55, %r553, 4;
add.s64 %rd56, %rd37, %rd55;
ld.global.nc.u32 %r554, [%rd56];
not.b32 %r555, %r554;
shr.u32 %r556, %r466, 5;
and.b32 %r557, %r556, 255;
popc.b32 %r558, %r557;
shl.b32 %r559, %r558, 7;
and.b32 %r560, %r559, 128;
xor.b32 %r561, %r560, %r557;
mul.lo.s32 %r562, %r561, 16843009;
and.b32 %r563, %r562, 134480385;
or.b32 %r564, %r563, -2139062144;
add.s32 %r565, %r564, -16843009;
or.b32 %r403, %r565, %r563;
mov.u32 %r109, %r403;

prmt.b32 %r109, %r109, 0, 0xba98;

xor.b32 %r566, %r109, %r551;
or.b32 %r567, %r566, -2139062144;
and.b32 %r568, %r109, 2139062143;
and.b32 %r569, %r552, -2139062144;
sub.s32 %r570, %r567, %r568;
xor.b32 %r114, %r569, %r570;
ld.global.nc.u32 %r115, [%rd41+28];
and.b32 %r571, %r562, -2143281136;
or.b32 %r572, %r571, -2139062144;
add.s32 %r573, %r572, -16843009;
or.b32 %r405, %r573, %r571;
mov.u32 %r111, %r405;

prmt.b32 %r111, %r111, 0, 0xba98;

xor.b32 %r574, %r111, %r554;
or.b32 %r575, %r574, -2139062144;
and.b32 %r576, %r111, 2139062143;
and.b32 %r577, %r555, -2139062144;
sub.s32 %r578, %r575, %r576;
xor.b32 %r118, %r577, %r578;
ld.global.nc.u32 %r119, [%rd41+32];

dp4a.s32.s32 %r113, %r114, %r115, %r105;

shr.u32 %r579, %r466, 12;

dp4a.s32.s32 %r117, %r118, %r119, %r113;

shr.u32 %r580, %r117, 31;
add.s32 %r581, %r117, %r580;
shr.s32 %r582, %r581, 1;
mad.lo.s32 %r583, %r117, %r579, %r582;
shr.u32 %r584, %r583, 31;
add.s32 %r585, %r583, %r584;
shr.s32 %r586, %r585, 1;
add.s64 %rd57, %rd116, %rd6;
ld.global.nc.u16 %rs1, [%rd57];

{ cvt.f32.f16 %f49, %rs1;}


ld.global.nc.u32 %r121, [%rd41];

{.reg .f16 low,high;
mov.b32 {low,high},%r121;
cvt.f32.f16 %f50, low;}


mul.ftz.f32 %f65, %f49, %f50;
cvt.rn.f32.s32 %f66, %r586;
fma.rn.ftz.f32 %f216, %f65, %f66, %f216;
add.s64 %rd58, %rd116, %rd10;
ld.global.nc.u16 %rs23, [%rd58];
ld.global.nc.u16 %rs24, [%rd58+2];
ld.global.nc.u16 %rs25, [%rd58+4];
add.s64 %rd59, %rd116, %rd8;
ld.global.nc.u16 %rs26, [%rd59];
cvt.u32.u16 %r587, %rs26;
ld.global.nc.u16 %rs27, [%rd59+2];
cvt.u32.u16 %r588, %rs27;
mov.b32 %r589, {%rs26, %rs27};
ld.global.nc.u16 %rs28, [%rd58+-2];
shl.b16 %rs29, %rs28, 2;
cvt.u64.u16 %rd60, %rs29;
and.b64 %rd61, %rd60, 1020;
add.s64 %rd62, %rd37, %rd61;
ld.global.nc.u32 %r590, [%rd62];
not.b32 %r591, %r590;
shr.u16 %rs30, %rs28, 8;
cvt.u32.u16 %r592, %rs30;
mul.wide.u32 %rd63, %r592, 4;
add.s64 %rd64, %rd37, %rd63;
ld.global.nc.u32 %r593, [%rd64];
not.b32 %r594, %r593;
and.b32 %r595, %r587, 255;
popc.b32 %r596, %r595;
shl.b32 %r597, %r596, 7;
and.b32 %r598, %r597, 128;
xor.b32 %r599, %r598, %r595;
mul.lo.s32 %r600, %r599, 16843009;
and.b32 %r601, %r600, 134480385;
or.b32 %r602, %r601, -2139062144;
add.s32 %r603, %r602, -16843009;
or.b32 %r416, %r603, %r601;
mov.u32 %r122, %r416;

prmt.b32 %r122, %r122, 0, 0xba98;

xor.b32 %r604, %r122, %r590;
or.b32 %r605, %r604, -2139062144;
and.b32 %r606, %r122, 2139062143;
and.b32 %r607, %r591, -2139062144;
sub.s32 %r608, %r605, %r606;
xor.b32 %r127, %r607, %r608;
and.b32 %r609, %r600, -2143281136;
or.b32 %r610, %r609, -2139062144;
add.s32 %r611, %r610, -16843009;
or.b32 %r418, %r611, %r609;
mov.u32 %r124, %r418;

prmt.b32 %r124, %r124, 0, 0xba98;

xor.b32 %r612, %r124, %r593;
or.b32 %r613, %r612, -2139062144;
and.b32 %r614, %r124, 2139062143;
and.b32 %r615, %r594, -2139062144;
sub.s32 %r616, %r613, %r614;
xor.b32 %r131, %r615, %r616;

dp4a.s32.s32 %r126, %r127, %r79, %r423;


dp4a.s32.s32 %r130, %r131, %r83, %r126;

shl.b16 %rs31, %rs23, 2;
cvt.u64.u16 %rd65, %rs31;
and.b64 %rd66, %rd65, 1020;
add.s64 %rd67, %rd37, %rd66;
ld.global.nc.u32 %r617, [%rd67];
not.b32 %r618, %r617;
shr.u16 %rs32, %rs23, 8;
cvt.u32.u16 %r619, %rs32;
mul.wide.u32 %rd68, %r619, 4;
add.s64 %rd69, %rd37, %rd68;
ld.global.nc.u32 %r620, [%rd69];
not.b32 %r621, %r620;
shr.u32 %r622, %r587, 7;
and.b32 %r623, %r622, 255;
popc.b32 %r624, %r623;
shl.b32 %r625, %r624, 7;
and.b32 %r626, %r625, 128;
xor.b32 %r627, %r626, %r623;
mul.lo.s32 %r628, %r627, 16843009;
and.b32 %r629, %r628, 134480385;
or.b32 %r630, %r629, -2139062144;
add.s32 %r631, %r630, -16843009;
or.b32 %r428, %r631, %r629;
mov.u32 %r134, %r428;

prmt.b32 %r134, %r134, 0, 0xba98;

xor.b32 %r632, %r134, %r617;
or.b32 %r633, %r632, -2139062144;
and.b32 %r634, %r134, 2139062143;
and.b32 %r635, %r618, -2139062144;
sub.s32 %r636, %r633, %r634;
xor.b32 %r139, %r635, %r636;
and.b32 %r637, %r628, -2143281136;
or.b32 %r638, %r637, -2139062144;
add.s32 %r639, %r638, -16843009;
or.b32 %r430, %r639, %r637;
mov.u32 %r136, %r430;

prmt.b32 %r136, %r136, 0, 0xba98;

xor.b32 %r640, %r136, %r620;
or.b32 %r641, %r640, -2139062144;
and.b32 %r642, %r136, 2139062143;
and.b32 %r643, %r621, -2139062144;
sub.s32 %r644, %r641, %r642;
xor.b32 %r143, %r643, %r644;

dp4a.s32.s32 %r138, %r139, %r91, %r130;


dp4a.s32.s32 %r142, %r143, %r95, %r138;

shl.b16 %rs33, %rs24, 2;
cvt.u64.u16 %rd70, %rs33;
and.b64 %rd71, %rd70, 1020;
add.s64 %rd72, %rd37, %rd71;
ld.global.nc.u32 %r645, [%rd72];
not.b32 %r646, %r645;
shr.u16 %rs34, %rs24, 8;
cvt.u32.u16 %r647, %rs34;
mul.wide.u32 %rd73, %r647, 4;
add.s64 %rd74, %rd37, %rd73;
ld.global.nc.u32 %r648, [%rd74];
not.b32 %r649, %r648;
shr.u32 %r650, %r589, 14;
and.b32 %r651, %r650, 255;
popc.b32 %r652, %r651;
shl.b32 %r653, %r652, 7;
and.b32 %r654, %r653, 128;
xor.b32 %r655, %r654, %r651;
mul.lo.s32 %r656, %r655, 16843009;
and.b32 %r657, %r656, 134480385;
or.b32 %r658, %r657, -2139062144;
add.s32 %r659, %r658, -16843009;
or.b32 %r440, %r659, %r657;
mov.u32 %r146, %r440;

prmt.b32 %r146, %r146, 0, 0xba98;

xor.b32 %r660, %r146, %r645;
or.b32 %r661, %r660, -2139062144;
and.b32 %r662, %r146, 2139062143;
and.b32 %r663, %r646, -2139062144;
sub.s32 %r664, %r661, %r662;
xor.b32 %r151, %r663, %r664;
and.b32 %r665, %r656, -2143281136;
or.b32 %r666, %r665, -2139062144;
add.s32 %r667, %r666, -16843009;
or.b32 %r442, %r667, %r665;
mov.u32 %r148, %r442;

prmt.b32 %r148, %r148, 0, 0xba98;

xor.b32 %r668, %r148, %r648;
or.b32 %r669, %r668, -2139062144;
and.b32 %r670, %r148, 2139062143;
and.b32 %r671, %r649, -2139062144;
sub.s32 %r672, %r669, %r670;
xor.b32 %r155, %r671, %r672;

dp4a.s32.s32 %r150, %r151, %r103, %r142;


dp4a.s32.s32 %r154, %r155, %r107, %r150;

shl.b16 %rs35, %rs25, 2;
cvt.u64.u16 %rd75, %rs35;
and.b64 %rd76, %rd75, 1020;
add.s64 %rd77, %rd37, %rd76;
ld.global.nc.u32 %r673, [%rd77];
not.b32 %r674, %r673;
shr.u16 %rs36, %rs25, 8;
cvt.u32.u16 %r675, %rs36;
mul.wide.u32 %rd78, %r675, 4;
add.s64 %rd79, %rd37, %rd78;
ld.global.nc.u32 %r676, [%rd79];
not.b32 %r677, %r676;
shr.u32 %r678, %r588, 5;
and.b32 %r679, %r678, 255;
popc.b32 %r680, %r679;
shl.b32 %r681, %r680, 7;
and.b32 %r682, %r681, 128;
xor.b32 %r683, %r682, %r679;
mul.lo.s32 %r684, %r683, 16843009;
and.b32 %r685, %r684, 134480385;
or.b32 %r686, %r685, -2139062144;
add.s32 %r687, %r686, -16843009;
or.b32 %r452, %r687, %r685;
mov.u32 %r158, %r452;

prmt.b32 %r158, %r158, 0, 0xba98;

xor.b32 %r688, %r158, %r673;
or.b32 %r689, %r688, -2139062144;
and.b32 %r690, %r158, 2139062143;
and.b32 %r691, %r674, -2139062144;
sub.s32 %r692, %r689, %r690;
xor.b32 %r163, %r691, %r692;
and.b32 %r693, %r684, -2143281136;
or.b32 %r694, %r693, -2139062144;
add.s32 %r695, %r694, -16843009;
or.b32 %r454, %r695, %r693;
mov.u32 %r160, %r454;

prmt.b32 %r160, %r160, 0, 0xba98;

xor.b32 %r696, %r160, %r676;
or.b32 %r697, %r696, -2139062144;
and.b32 %r698, %r160, 2139062143;
and.b32 %r699, %r677, -2139062144;
sub.s32 %r700, %r697, %r698;
xor.b32 %r167, %r699, %r700;

dp4a.s32.s32 %r162, %r163, %r115, %r154;

shr.u32 %r701, %r588, 12;

dp4a.s32.s32 %r166, %r167, %r119, %r162;

shr.u32 %r702, %r166, 31;
add.s32 %r703, %r166, %r702;
shr.s32 %r704, %r703, 1;
mad.lo.s32 %r705, %r166, %r701, %r704;
shr.u32 %r706, %r705, 31;
add.s32 %r707, %r705, %r706;
shr.s32 %r708, %r707, 1;
add.s64 %rd80, %rd116, %rd5;
ld.global.nc.u16 %rs2, [%rd80];

{ cvt.f32.f16 %f51, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r121;
cvt.f32.f16 %f52, low;}


mul.ftz.f32 %f67, %f51, %f52;
cvt.rn.f32.s32 %f68, %r708;
fma.rn.ftz.f32 %f215, %f67, %f68, %f215;
cvt.u64.u32 %rd81, %r1096;
add.s64 %rd82, %rd2, %rd81;
add.s64 %rd83, %rd82, %rd4;
mul.lo.s64 %rd84, %rd83, 36;
mov.u32 %r171, %r367;

prmt.b32 %r171, %r171, 0, 0xba98;

xor.b32 %r709, %r171, %r468;
or.b32 %r710, %r709, -2139062144;
and.b32 %r711, %r171, 2139062143;
sub.s32 %r712, %r710, %r711;
xor.b32 %r176, %r485, %r712;
add.s64 %rd85, %rd11, %rd84;
ld.global.nc.u32 %r177, [%rd85+4];
mov.u32 %r173, %r369;

prmt.b32 %r173, %r173, 0, 0xba98;

xor.b32 %r713, %r173, %r471;
or.b32 %r714, %r713, -2139062144;
and.b32 %r715, %r173, 2139062143;
sub.s32 %r716, %r714, %r715;
xor.b32 %r180, %r493, %r716;
ld.global.nc.u32 %r181, [%rd85+8];

dp4a.s32.s32 %r175, %r176, %r177, %r423;


dp4a.s32.s32 %r179, %r180, %r181, %r175;

mov.u32 %r183, %r379;

prmt.b32 %r183, %r183, 0, 0xba98;

xor.b32 %r717, %r183, %r495;
or.b32 %r718, %r717, -2139062144;
and.b32 %r719, %r183, 2139062143;
sub.s32 %r720, %r718, %r719;
xor.b32 %r188, %r513, %r720;
ld.global.nc.u32 %r189, [%rd85+12];
mov.u32 %r185, %r381;

prmt.b32 %r185, %r185, 0, 0xba98;

xor.b32 %r721, %r185, %r498;
or.b32 %r722, %r721, -2139062144;
and.b32 %r723, %r185, 2139062143;
sub.s32 %r724, %r722, %r723;
xor.b32 %r192, %r521, %r724;
ld.global.nc.u32 %r193, [%rd85+16];

dp4a.s32.s32 %r187, %r188, %r189, %r179;


dp4a.s32.s32 %r191, %r192, %r193, %r187;

mov.u32 %r195, %r391;

prmt.b32 %r195, %r195, 0, 0xba98;

xor.b32 %r725, %r195, %r523;
or.b32 %r726, %r725, -2139062144;
and.b32 %r727, %r195, 2139062143;
sub.s32 %r728, %r726, %r727;
xor.b32 %r200, %r541, %r728;
ld.global.nc.u32 %r201, [%rd85+20];
mov.u32 %r197, %r393;

prmt.b32 %r197, %r197, 0, 0xba98;

xor.b32 %r729, %r197, %r526;
or.b32 %r730, %r729, -2139062144;
and.b32 %r731, %r197, 2139062143;
sub.s32 %r732, %r730, %r731;
xor.b32 %r204, %r549, %r732;
ld.global.nc.u32 %r205, [%rd85+24];

dp4a.s32.s32 %r199, %r200, %r201, %r191;


dp4a.s32.s32 %r203, %r204, %r205, %r199;

mov.u32 %r207, %r403;

prmt.b32 %r207, %r207, 0, 0xba98;

xor.b32 %r733, %r207, %r551;
or.b32 %r734, %r733, -2139062144;
and.b32 %r735, %r207, 2139062143;
sub.s32 %r736, %r734, %r735;
xor.b32 %r212, %r569, %r736;
ld.global.nc.u32 %r213, [%rd85+28];
mov.u32 %r209, %r405;

prmt.b32 %r209, %r209, 0, 0xba98;

xor.b32 %r737, %r209, %r554;
or.b32 %r738, %r737, -2139062144;
and.b32 %r739, %r209, 2139062143;
sub.s32 %r740, %r738, %r739;
xor.b32 %r216, %r577, %r740;
ld.global.nc.u32 %r217, [%rd85+32];

dp4a.s32.s32 %r211, %r212, %r213, %r203;


dp4a.s32.s32 %r215, %r216, %r217, %r211;

shr.u32 %r741, %r215, 31;
add.s32 %r742, %r215, %r741;
shr.s32 %r743, %r742, 1;
mad.lo.s32 %r744, %r215, %r579, %r743;
shr.u32 %r745, %r744, 31;
add.s32 %r746, %r744, %r745;
shr.s32 %r747, %r746, 1;

{ cvt.f32.f16 %f53, %rs1;}


ld.global.nc.u32 %r219, [%rd85];

{.reg .f16 low,high;
mov.b32 {low,high},%r219;
cvt.f32.f16 %f54, low;}


mul.ftz.f32 %f69, %f53, %f54;
cvt.rn.f32.s32 %f70, %r747;
fma.rn.ftz.f32 %f214, %f69, %f70, %f214;
mov.u32 %r220, %r416;

prmt.b32 %r220, %r220, 0, 0xba98;

xor.b32 %r748, %r220, %r590;
or.b32 %r749, %r748, -2139062144;
and.b32 %r750, %r220, 2139062143;
sub.s32 %r751, %r749, %r750;
xor.b32 %r225, %r607, %r751;
mov.u32 %r222, %r418;

prmt.b32 %r222, %r222, 0, 0xba98;

xor.b32 %r752, %r222, %r593;
or.b32 %r753, %r752, -2139062144;
and.b32 %r754, %r222, 2139062143;
sub.s32 %r755, %r753, %r754;
xor.b32 %r229, %r615, %r755;

dp4a.s32.s32 %r224, %r225, %r177, %r423;


dp4a.s32.s32 %r228, %r229, %r181, %r224;

mov.u32 %r232, %r428;

prmt.b32 %r232, %r232, 0, 0xba98;

xor.b32 %r756, %r232, %r617;
or.b32 %r757, %r756, -2139062144;
and.b32 %r758, %r232, 2139062143;
sub.s32 %r759, %r757, %r758;
xor.b32 %r237, %r635, %r759;
mov.u32 %r234, %r430;

prmt.b32 %r234, %r234, 0, 0xba98;

xor.b32 %r760, %r234, %r620;
or.b32 %r761, %r760, -2139062144;
and.b32 %r762, %r234, 2139062143;
sub.s32 %r763, %r761, %r762;
xor.b32 %r241, %r643, %r763;

dp4a.s32.s32 %r236, %r237, %r189, %r228;


dp4a.s32.s32 %r240, %r241, %r193, %r236;

mov.u32 %r244, %r440;

prmt.b32 %r244, %r244, 0, 0xba98;

xor.b32 %r764, %r244, %r645;
or.b32 %r765, %r764, -2139062144;
and.b32 %r766, %r244, 2139062143;
sub.s32 %r767, %r765, %r766;
xor.b32 %r249, %r663, %r767;
mov.u32 %r246, %r442;

prmt.b32 %r246, %r246, 0, 0xba98;

xor.b32 %r768, %r246, %r648;
or.b32 %r769, %r768, -2139062144;
and.b32 %r770, %r246, 2139062143;
sub.s32 %r771, %r769, %r770;
xor.b32 %r253, %r671, %r771;

dp4a.s32.s32 %r248, %r249, %r201, %r240;


dp4a.s32.s32 %r252, %r253, %r205, %r248;

mov.u32 %r256, %r452;

prmt.b32 %r256, %r256, 0, 0xba98;

xor.b32 %r772, %r256, %r673;
or.b32 %r773, %r772, -2139062144;
and.b32 %r774, %r256, 2139062143;
sub.s32 %r775, %r773, %r774;
xor.b32 %r261, %r691, %r775;
mov.u32 %r258, %r454;

prmt.b32 %r258, %r258, 0, 0xba98;

xor.b32 %r776, %r258, %r676;
or.b32 %r777, %r776, -2139062144;
and.b32 %r778, %r258, 2139062143;
sub.s32 %r779, %r777, %r778;
xor.b32 %r265, %r699, %r779;

dp4a.s32.s32 %r260, %r261, %r213, %r252;


dp4a.s32.s32 %r264, %r265, %r217, %r260;

shr.u32 %r780, %r264, 31;
add.s32 %r781, %r264, %r780;
shr.s32 %r782, %r781, 1;
mad.lo.s32 %r783, %r264, %r701, %r782;
shr.u32 %r784, %r783, 31;
add.s32 %r785, %r783, %r784;
shr.s32 %r786, %r785, 1;

{ cvt.f32.f16 %f55, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r219;
cvt.f32.f16 %f56, low;}


mul.ftz.f32 %f71, %f55, %f56;
cvt.rn.f32.s32 %f72, %r786;
fma.rn.ftz.f32 %f213, %f71, %f72, %f213;
cvt.u64.u32 %rd86, %r1098;
add.s64 %rd87, %rd2, %rd86;
add.s64 %rd88, %rd87, %rd4;
mul.lo.s64 %rd89, %rd88, 36;
mov.u32 %r269, %r367;

prmt.b32 %r269, %r269, 0, 0xba98;

xor.b32 %r787, %r269, %r468;
or.b32 %r788, %r787, -2139062144;
and.b32 %r789, %r269, 2139062143;
sub.s32 %r790, %r788, %r789;
xor.b32 %r274, %r485, %r790;
add.s64 %rd90, %rd11, %rd89;
ld.global.nc.u32 %r275, [%rd90+4];
mov.u32 %r271, %r369;

prmt.b32 %r271, %r271, 0, 0xba98;

xor.b32 %r791, %r271, %r471;
or.b32 %r792, %r791, -2139062144;
and.b32 %r793, %r271, 2139062143;
sub.s32 %r794, %r792, %r793;
xor.b32 %r278, %r493, %r794;
ld.global.nc.u32 %r279, [%rd90+8];

dp4a.s32.s32 %r273, %r274, %r275, %r423;


dp4a.s32.s32 %r277, %r278, %r279, %r273;

mov.u32 %r281, %r379;

prmt.b32 %r281, %r281, 0, 0xba98;

xor.b32 %r795, %r281, %r495;
or.b32 %r796, %r795, -2139062144;
and.b32 %r797, %r281, 2139062143;
sub.s32 %r798, %r796, %r797;
xor.b32 %r286, %r513, %r798;
ld.global.nc.u32 %r287, [%rd90+12];
mov.u32 %r283, %r381;

prmt.b32 %r283, %r283, 0, 0xba98;

xor.b32 %r799, %r283, %r498;
or.b32 %r800, %r799, -2139062144;
and.b32 %r801, %r283, 2139062143;
sub.s32 %r802, %r800, %r801;
xor.b32 %r290, %r521, %r802;
ld.global.nc.u32 %r291, [%rd90+16];

dp4a.s32.s32 %r285, %r286, %r287, %r277;


dp4a.s32.s32 %r289, %r290, %r291, %r285;

mov.u32 %r293, %r391;

prmt.b32 %r293, %r293, 0, 0xba98;

xor.b32 %r803, %r293, %r523;
or.b32 %r804, %r803, -2139062144;
and.b32 %r805, %r293, 2139062143;
sub.s32 %r806, %r804, %r805;
xor.b32 %r298, %r541, %r806;
ld.global.nc.u32 %r299, [%rd90+20];
mov.u32 %r295, %r393;

prmt.b32 %r295, %r295, 0, 0xba98;

xor.b32 %r807, %r295, %r526;
or.b32 %r808, %r807, -2139062144;
and.b32 %r809, %r295, 2139062143;
sub.s32 %r810, %r808, %r809;
xor.b32 %r302, %r549, %r810;
ld.global.nc.u32 %r303, [%rd90+24];

dp4a.s32.s32 %r297, %r298, %r299, %r289;


dp4a.s32.s32 %r301, %r302, %r303, %r297;

mov.u32 %r305, %r403;

prmt.b32 %r305, %r305, 0, 0xba98;

xor.b32 %r811, %r305, %r551;
or.b32 %r812, %r811, -2139062144;
and.b32 %r813, %r305, 2139062143;
sub.s32 %r814, %r812, %r813;
xor.b32 %r310, %r569, %r814;
ld.global.nc.u32 %r311, [%rd90+28];
mov.u32 %r307, %r405;

prmt.b32 %r307, %r307, 0, 0xba98;

xor.b32 %r815, %r307, %r554;
or.b32 %r816, %r815, -2139062144;
and.b32 %r817, %r307, 2139062143;
sub.s32 %r818, %r816, %r817;
xor.b32 %r314, %r577, %r818;
ld.global.nc.u32 %r315, [%rd90+32];

dp4a.s32.s32 %r309, %r310, %r311, %r301;


dp4a.s32.s32 %r313, %r314, %r315, %r309;

shr.u32 %r819, %r313, 31;
add.s32 %r820, %r313, %r819;
shr.s32 %r821, %r820, 1;
mad.lo.s32 %r822, %r313, %r579, %r821;
shr.u32 %r823, %r822, 31;
add.s32 %r824, %r822, %r823;
shr.s32 %r825, %r824, 1;

{ cvt.f32.f16 %f57, %rs1;}


ld.global.nc.u32 %r317, [%rd90];

{.reg .f16 low,high;
mov.b32 {low,high},%r317;
cvt.f32.f16 %f58, low;}


mul.ftz.f32 %f73, %f57, %f58;
cvt.rn.f32.s32 %f74, %r825;
fma.rn.ftz.f32 %f212, %f73, %f74, %f212;
mov.u32 %r318, %r416;

prmt.b32 %r318, %r318, 0, 0xba98;

xor.b32 %r826, %r318, %r590;
or.b32 %r827, %r826, -2139062144;
and.b32 %r828, %r318, 2139062143;
sub.s32 %r829, %r827, %r828;
xor.b32 %r323, %r607, %r829;
mov.u32 %r320, %r418;

prmt.b32 %r320, %r320, 0, 0xba98;

xor.b32 %r830, %r320, %r593;
or.b32 %r831, %r830, -2139062144;
and.b32 %r832, %r320, 2139062143;
sub.s32 %r833, %r831, %r832;
xor.b32 %r327, %r615, %r833;

dp4a.s32.s32 %r322, %r323, %r275, %r423;


dp4a.s32.s32 %r326, %r327, %r279, %r322;

mov.u32 %r330, %r428;

prmt.b32 %r330, %r330, 0, 0xba98;

xor.b32 %r834, %r330, %r617;
or.b32 %r835, %r834, -2139062144;
and.b32 %r836, %r330, 2139062143;
sub.s32 %r837, %r835, %r836;
xor.b32 %r335, %r635, %r837;
mov.u32 %r332, %r430;

prmt.b32 %r332, %r332, 0, 0xba98;

xor.b32 %r838, %r332, %r620;
or.b32 %r839, %r838, -2139062144;
and.b32 %r840, %r332, 2139062143;
sub.s32 %r841, %r839, %r840;
xor.b32 %r339, %r643, %r841;

dp4a.s32.s32 %r334, %r335, %r287, %r326;


dp4a.s32.s32 %r338, %r339, %r291, %r334;

mov.u32 %r342, %r440;

prmt.b32 %r342, %r342, 0, 0xba98;

xor.b32 %r842, %r342, %r645;
or.b32 %r843, %r842, -2139062144;
and.b32 %r844, %r342, 2139062143;
sub.s32 %r845, %r843, %r844;
xor.b32 %r347, %r663, %r845;
mov.u32 %r344, %r442;

prmt.b32 %r344, %r344, 0, 0xba98;

xor.b32 %r846, %r344, %r648;
or.b32 %r847, %r846, -2139062144;
and.b32 %r848, %r344, 2139062143;
sub.s32 %r849, %r847, %r848;
xor.b32 %r351, %r671, %r849;

dp4a.s32.s32 %r346, %r347, %r299, %r338;


dp4a.s32.s32 %r350, %r351, %r303, %r346;

mov.u32 %r354, %r452;

prmt.b32 %r354, %r354, 0, 0xba98;

xor.b32 %r850, %r354, %r673;
or.b32 %r851, %r850, -2139062144;
and.b32 %r852, %r354, 2139062143;
sub.s32 %r853, %r851, %r852;
xor.b32 %r359, %r691, %r853;
mov.u32 %r356, %r454;

prmt.b32 %r356, %r356, 0, 0xba98;

xor.b32 %r854, %r356, %r676;
or.b32 %r855, %r854, -2139062144;
and.b32 %r856, %r356, 2139062143;
sub.s32 %r857, %r855, %r856;
xor.b32 %r363, %r699, %r857;

dp4a.s32.s32 %r358, %r359, %r311, %r350;


dp4a.s32.s32 %r362, %r363, %r315, %r358;

shr.u32 %r858, %r362, 31;
add.s32 %r859, %r362, %r858;
shr.s32 %r860, %r859, 1;
mad.lo.s32 %r861, %r362, %r701, %r860;
shr.u32 %r862, %r861, 31;
add.s32 %r863, %r861, %r862;
shr.s32 %r864, %r863, 1;

{ cvt.f32.f16 %f59, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r317;
cvt.f32.f16 %f60, low;}


mul.ftz.f32 %f75, %f59, %f60;
cvt.rn.f32.s32 %f76, %r864;
fma.rn.ftz.f32 %f211, %f75, %f76, %f211;
cvt.u64.u32 %rd91, %r1097;
add.s64 %rd92, %rd2, %rd91;
add.s64 %rd93, %rd92, %rd4;
mul.lo.s64 %rd94, %rd93, 36;

prmt.b32 %r367, %r367, 0, 0xba98;

xor.b32 %r865, %r367, %r468;
or.b32 %r866, %r865, -2139062144;
and.b32 %r867, %r367, 2139062143;
sub.s32 %r868, %r866, %r867;
xor.b32 %r372, %r485, %r868;
add.s64 %rd95, %rd11, %rd94;
ld.global.nc.u32 %r373, [%rd95+4];

prmt.b32 %r369, %r369, 0, 0xba98;

xor.b32 %r869, %r369, %r471;
or.b32 %r870, %r869, -2139062144;
and.b32 %r871, %r369, 2139062143;
sub.s32 %r872, %r870, %r871;
xor.b32 %r376, %r493, %r872;
ld.global.nc.u32 %r377, [%rd95+8];

dp4a.s32.s32 %r371, %r372, %r373, %r423;


dp4a.s32.s32 %r375, %r376, %r377, %r371;


prmt.b32 %r379, %r379, 0, 0xba98;

xor.b32 %r873, %r379, %r495;
or.b32 %r874, %r873, -2139062144;
and.b32 %r875, %r379, 2139062143;
sub.s32 %r876, %r874, %r875;
xor.b32 %r384, %r513, %r876;
ld.global.nc.u32 %r385, [%rd95+12];

prmt.b32 %r381, %r381, 0, 0xba98;

xor.b32 %r877, %r381, %r498;
or.b32 %r878, %r877, -2139062144;
and.b32 %r879, %r381, 2139062143;
sub.s32 %r880, %r878, %r879;
xor.b32 %r388, %r521, %r880;
ld.global.nc.u32 %r389, [%rd95+16];

dp4a.s32.s32 %r383, %r384, %r385, %r375;


dp4a.s32.s32 %r387, %r388, %r389, %r383;


prmt.b32 %r391, %r391, 0, 0xba98;

xor.b32 %r881, %r391, %r523;
or.b32 %r882, %r881, -2139062144;
and.b32 %r883, %r391, 2139062143;
sub.s32 %r884, %r882, %r883;
xor.b32 %r396, %r541, %r884;
ld.global.nc.u32 %r397, [%rd95+20];

prmt.b32 %r393, %r393, 0, 0xba98;

xor.b32 %r885, %r393, %r526;
or.b32 %r886, %r885, -2139062144;
and.b32 %r887, %r393, 2139062143;
sub.s32 %r888, %r886, %r887;
xor.b32 %r400, %r549, %r888;
ld.global.nc.u32 %r401, [%rd95+24];

dp4a.s32.s32 %r395, %r396, %r397, %r387;


dp4a.s32.s32 %r399, %r400, %r401, %r395;


prmt.b32 %r403, %r403, 0, 0xba98;

xor.b32 %r889, %r403, %r551;
or.b32 %r890, %r889, -2139062144;
and.b32 %r891, %r403, 2139062143;
sub.s32 %r892, %r890, %r891;
xor.b32 %r408, %r569, %r892;
ld.global.nc.u32 %r409, [%rd95+28];

prmt.b32 %r405, %r405, 0, 0xba98;

xor.b32 %r893, %r405, %r554;
or.b32 %r894, %r893, -2139062144;
and.b32 %r895, %r405, 2139062143;
sub.s32 %r896, %r894, %r895;
xor.b32 %r412, %r577, %r896;
ld.global.nc.u32 %r413, [%rd95+32];

dp4a.s32.s32 %r407, %r408, %r409, %r399;


dp4a.s32.s32 %r411, %r412, %r413, %r407;

shr.u32 %r897, %r411, 31;
add.s32 %r898, %r411, %r897;
shr.s32 %r899, %r898, 1;
mad.lo.s32 %r900, %r411, %r579, %r899;
shr.u32 %r901, %r900, 31;
add.s32 %r902, %r900, %r901;
shr.s32 %r903, %r902, 1;

{ cvt.f32.f16 %f61, %rs1;}


ld.global.nc.u32 %r415, [%rd95];

{.reg .f16 low,high;
mov.b32 {low,high},%r415;
cvt.f32.f16 %f62, low;}


mul.ftz.f32 %f77, %f61, %f62;
cvt.rn.f32.s32 %f78, %r903;
fma.rn.ftz.f32 %f210, %f77, %f78, %f210;

prmt.b32 %r416, %r416, 0, 0xba98;

xor.b32 %r904, %r416, %r590;
or.b32 %r905, %r904, -2139062144;
and.b32 %r906, %r416, 2139062143;
sub.s32 %r907, %r905, %r906;
xor.b32 %r421, %r607, %r907;

prmt.b32 %r418, %r418, 0, 0xba98;

xor.b32 %r908, %r418, %r593;
or.b32 %r909, %r908, -2139062144;
and.b32 %r910, %r418, 2139062143;
sub.s32 %r911, %r909, %r910;
xor.b32 %r425, %r615, %r911;

dp4a.s32.s32 %r420, %r421, %r373, %r423;


dp4a.s32.s32 %r424, %r425, %r377, %r420;


prmt.b32 %r428, %r428, 0, 0xba98;

xor.b32 %r912, %r428, %r617;
or.b32 %r913, %r912, -2139062144;
and.b32 %r914, %r428, 2139062143;
sub.s32 %r915, %r913, %r914;
xor.b32 %r433, %r635, %r915;

prmt.b32 %r430, %r430, 0, 0xba98;

xor.b32 %r916, %r430, %r620;
or.b32 %r917, %r916, -2139062144;
and.b32 %r918, %r430, 2139062143;
sub.s32 %r919, %r917, %r918;
xor.b32 %r437, %r643, %r919;

dp4a.s32.s32 %r432, %r433, %r385, %r424;


dp4a.s32.s32 %r436, %r437, %r389, %r432;


prmt.b32 %r440, %r440, 0, 0xba98;

xor.b32 %r920, %r440, %r645;
or.b32 %r921, %r920, -2139062144;
and.b32 %r922, %r440, 2139062143;
sub.s32 %r923, %r921, %r922;
xor.b32 %r445, %r663, %r923;

prmt.b32 %r442, %r442, 0, 0xba98;

xor.b32 %r924, %r442, %r648;
or.b32 %r925, %r924, -2139062144;
and.b32 %r926, %r442, 2139062143;
sub.s32 %r927, %r925, %r926;
xor.b32 %r449, %r671, %r927;

dp4a.s32.s32 %r444, %r445, %r397, %r436;


dp4a.s32.s32 %r448, %r449, %r401, %r444;


prmt.b32 %r452, %r452, 0, 0xba98;

xor.b32 %r928, %r452, %r673;
or.b32 %r929, %r928, -2139062144;
and.b32 %r930, %r452, 2139062143;
sub.s32 %r931, %r929, %r930;
xor.b32 %r457, %r691, %r931;

prmt.b32 %r454, %r454, 0, 0xba98;

xor.b32 %r932, %r454, %r676;
or.b32 %r933, %r932, -2139062144;
and.b32 %r934, %r454, 2139062143;
sub.s32 %r935, %r933, %r934;
xor.b32 %r461, %r699, %r935;

dp4a.s32.s32 %r456, %r457, %r409, %r448;


dp4a.s32.s32 %r460, %r461, %r413, %r456;

shr.u32 %r936, %r460, 31;
add.s32 %r937, %r460, %r936;
shr.s32 %r938, %r937, 1;
mad.lo.s32 %r939, %r460, %r701, %r938;
shr.u32 %r940, %r939, 31;
add.s32 %r941, %r939, %r940;
shr.s32 %r942, %r941, 1;

{ cvt.f32.f16 %f63, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r415;
cvt.f32.f16 %f64, low;}


mul.ftz.f32 %f79, %f63, %f64;
cvt.rn.f32.s32 %f80, %r942;
fma.rn.ftz.f32 %f209, %f79, %f80, %f209;
add.s32 %r1098, %r1098, 128;
add.s32 %r1097, %r1097, 128;
add.s32 %r1096, %r1096, 128;
add.s32 %r1095, %r1095, 128;
add.s64 %rd116, %rd116, 1568;
add.s32 %r1099, %r1099, 16;
setp.lt.s32 %p4, %r1099, %r6;
@%p4 bra $L__BB245_2;

$L__BB245_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB245_5;

shl.b32 %r943, %r1, 10;
mov.u32 %r944, _ZZ13mul_mat_vec_qIL9ggml_type18ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r945, %r944, %r943;
shl.b32 %r946, %r2, 2;
add.s32 %r947, %r945, %r946;
st.shared.f32 [%r947+-1024], %f216;
st.shared.f32 [%r947+-896], %f215;
st.shared.f32 [%r947+-768], %f214;
st.shared.f32 [%r947+-640], %f213;
st.shared.f32 [%r947+-512], %f212;
st.shared.f32 [%r947+-384], %f211;
st.shared.f32 [%r947+-256], %f210;
st.shared.f32 [%r947+-128], %f209;

$L__BB245_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB245_22;

mov.u32 %r948, %ctaid.z;
mov.u32 %r949, %ctaid.y;
mad.lo.s32 %r950, %r949, %r32, %r4;
mad.lo.s32 %r951, %r948, %r38, %r950;
cvt.u64.u32 %rd14, %r951;
shl.b32 %r952, %r2, 2;
mov.u32 %r953, _ZZ13mul_mat_vec_qIL9ggml_type18ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r21, %r953, %r952;
mov.u32 %r954, 2;
ld.shared.f32 %f81, [%r21];
add.ftz.f32 %f82, %f81, %f216;
ld.shared.f32 %f83, [%r21+1024];
add.ftz.f32 %f84, %f83, %f82;
ld.shared.f32 %f85, [%r21+2048];
add.ftz.f32 %f86, %f85, %f84;
mov.b32 %r955, %f86;
mov.u32 %r956, 31;
mov.u32 %r957, 16;
mov.u32 %r958, -1;
shfl.sync.bfly.b32 %r959|%p7, %r955, %r957, %r956, %r958;
mov.b32 %f87, %r959;
add.ftz.f32 %f88, %f86, %f87;
mov.b32 %r960, %f88;
mov.u32 %r961, 8;
shfl.sync.bfly.b32 %r962|%p8, %r960, %r961, %r956, %r958;
mov.b32 %f89, %r962;
add.ftz.f32 %f90, %f88, %f89;
mov.b32 %r963, %f90;
mov.u32 %r964, 4;
shfl.sync.bfly.b32 %r965|%p9, %r963, %r964, %r956, %r958;
mov.b32 %f91, %r965;
add.ftz.f32 %f92, %f90, %f91;
mov.b32 %r966, %f92;
shfl.sync.bfly.b32 %r967|%p10, %r966, %r954, %r956, %r958;
mov.b32 %f93, %r967;
add.ftz.f32 %f94, %f92, %f93;
mov.b32 %r968, %f94;
mov.u32 %r969, 1;
shfl.sync.bfly.b32 %r970|%p11, %r968, %r969, %r956, %r958;
mov.b32 %f95, %r970;
add.ftz.f32 %f25, %f94, %f95;
setp.lt.u32 %p12, %r4, %r26;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
mul.wide.u32 %rd96, %r951, 4;
add.s64 %rd15, %rd1, %rd96;
not.pred %p14, %p1;
@%p14 bra $L__BB245_8;

st.global.f32 [%rd15], %f25;

$L__BB245_8:
ld.shared.f32 %f96, [%r21+128];
add.ftz.f32 %f97, %f96, %f215;
ld.shared.f32 %f98, [%r21+1152];
add.ftz.f32 %f99, %f98, %f97;
ld.shared.f32 %f100, [%r21+2176];
add.ftz.f32 %f101, %f100, %f99;
mov.b32 %r971, %f101;
shfl.sync.bfly.b32 %r975|%p15, %r971, %r957, %r956, %r958;
mov.b32 %f102, %r975;
add.ftz.f32 %f103, %f101, %f102;
mov.b32 %r976, %f103;
shfl.sync.bfly.b32 %r978|%p16, %r976, %r961, %r956, %r958;
mov.b32 %f104, %r978;
add.ftz.f32 %f105, %f103, %f104;
mov.b32 %r979, %f105;
shfl.sync.bfly.b32 %r981|%p17, %r979, %r964, %r956, %r958;
mov.b32 %f106, %r981;
add.ftz.f32 %f107, %f105, %f106;
mov.b32 %r982, %f107;
shfl.sync.bfly.b32 %r984|%p18, %r982, %r954, %r956, %r958;
mov.b32 %f108, %r984;
add.ftz.f32 %f109, %f107, %f108;
mov.b32 %r985, %f109;
shfl.sync.bfly.b32 %r987|%p19, %r985, %r969, %r956, %r958;
mov.b32 %f110, %r987;
add.ftz.f32 %f26, %f109, %f110;
setp.eq.s32 %p20, %r2, 1;
add.s32 %r989, %r4, 1;
setp.lt.u32 %p21, %r989, %r26;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB245_10;

st.global.f32 [%rd15+4], %f26;

$L__BB245_10:
ld.shared.f32 %f111, [%r21+256];
add.ftz.f32 %f112, %f111, %f214;
ld.shared.f32 %f113, [%r21+1280];
add.ftz.f32 %f114, %f113, %f112;
ld.shared.f32 %f115, [%r21+2304];
add.ftz.f32 %f116, %f115, %f114;
mov.b32 %r990, %f116;
mov.u32 %r991, 31;
mov.u32 %r992, 16;
mov.u32 %r993, -1;
shfl.sync.bfly.b32 %r994|%p23, %r990, %r992, %r991, %r993;
mov.b32 %f117, %r994;
add.ftz.f32 %f118, %f116, %f117;
mov.b32 %r995, %f118;
mov.u32 %r996, 8;
shfl.sync.bfly.b32 %r997|%p24, %r995, %r996, %r991, %r993;
mov.b32 %f119, %r997;
add.ftz.f32 %f120, %f118, %f119;
mov.b32 %r998, %f120;
mov.u32 %r999, 4;
shfl.sync.bfly.b32 %r1000|%p25, %r998, %r999, %r991, %r993;
mov.b32 %f121, %r1000;
add.ftz.f32 %f122, %f120, %f121;
mov.b32 %r1001, %f122;
mov.u32 %r1002, 2;
shfl.sync.bfly.b32 %r1003|%p26, %r1001, %r1002, %r991, %r993;
mov.b32 %f123, %r1003;
add.ftz.f32 %f124, %f122, %f123;
mov.b32 %r1004, %f124;
mov.u32 %r1005, 1;
shfl.sync.bfly.b32 %r1006|%p27, %r1004, %r1005, %r991, %r993;
mov.b32 %f125, %r1006;
add.ftz.f32 %f27, %f124, %f125;
@%p14 bra $L__BB245_12;

cvt.u64.u32 %rd97, %r26;
add.s64 %rd98, %rd97, %rd14;
shl.b64 %rd99, %rd98, 2;
add.s64 %rd100, %rd1, %rd99;
st.global.f32 [%rd100], %f27;

$L__BB245_12:
ld.shared.f32 %f126, [%r21+384];
add.ftz.f32 %f127, %f126, %f213;
ld.shared.f32 %f128, [%r21+1408];
add.ftz.f32 %f129, %f128, %f127;
ld.shared.f32 %f130, [%r21+2432];
add.ftz.f32 %f131, %f130, %f129;
mov.b32 %r1007, %f131;
shfl.sync.bfly.b32 %r1011|%p29, %r1007, %r992, %r991, %r993;
mov.b32 %f132, %r1011;
add.ftz.f32 %f133, %f131, %f132;
mov.b32 %r1012, %f133;
shfl.sync.bfly.b32 %r1014|%p30, %r1012, %r996, %r991, %r993;
mov.b32 %f134, %r1014;
add.ftz.f32 %f135, %f133, %f134;
mov.b32 %r1015, %f135;
shfl.sync.bfly.b32 %r1017|%p31, %r1015, %r999, %r991, %r993;
mov.b32 %f136, %r1017;
add.ftz.f32 %f137, %f135, %f136;
mov.b32 %r1018, %f137;
shfl.sync.bfly.b32 %r1020|%p32, %r1018, %r1002, %r991, %r993;
mov.b32 %f138, %r1020;
add.ftz.f32 %f139, %f137, %f138;
mov.b32 %r1021, %f139;
shfl.sync.bfly.b32 %r1023|%p33, %r1021, %r1005, %r991, %r993;
mov.b32 %f140, %r1023;
add.ftz.f32 %f28, %f139, %f140;
@%p22 bra $L__BB245_14;

add.s32 %r1024, %r26, 1;
cvt.u64.u32 %rd101, %r1024;
add.s64 %rd102, %rd101, %rd14;
shl.b64 %rd103, %rd102, 2;
add.s64 %rd104, %rd1, %rd103;
st.global.f32 [%rd104], %f28;

$L__BB245_14:
shl.b32 %r1025, %r26, 1;
mov.u32 %r1026, 1;
ld.shared.f32 %f141, [%r21+512];
add.ftz.f32 %f142, %f141, %f212;
ld.shared.f32 %f143, [%r21+1536];
add.ftz.f32 %f144, %f143, %f142;
ld.shared.f32 %f145, [%r21+2560];
add.ftz.f32 %f146, %f145, %f144;
mov.b32 %r1027, %f146;
mov.u32 %r1028, 31;
mov.u32 %r1029, 16;
mov.u32 %r1030, -1;
shfl.sync.bfly.b32 %r1031|%p35, %r1027, %r1029, %r1028, %r1030;
mov.b32 %f147, %r1031;
add.ftz.f32 %f148, %f146, %f147;
mov.b32 %r1032, %f148;
mov.u32 %r1033, 8;
shfl.sync.bfly.b32 %r1034|%p36, %r1032, %r1033, %r1028, %r1030;
mov.b32 %f149, %r1034;
add.ftz.f32 %f150, %f148, %f149;
mov.b32 %r1035, %f150;
mov.u32 %r1036, 4;
shfl.sync.bfly.b32 %r1037|%p37, %r1035, %r1036, %r1028, %r1030;
mov.b32 %f151, %r1037;
add.ftz.f32 %f152, %f150, %f151;
mov.b32 %r1038, %f152;
mov.u32 %r1039, 2;
shfl.sync.bfly.b32 %r1040|%p38, %r1038, %r1039, %r1028, %r1030;
mov.b32 %f153, %r1040;
add.ftz.f32 %f154, %f152, %f153;
mov.b32 %r1041, %f154;
shfl.sync.bfly.b32 %r1042|%p39, %r1041, %r1026, %r1028, %r1030;
mov.b32 %f155, %r1042;
add.ftz.f32 %f29, %f154, %f155;
cvt.u64.u32 %rd105, %r1025;
add.s64 %rd106, %rd105, %rd14;
shl.b64 %rd107, %rd106, 2;
add.s64 %rd16, %rd1, %rd107;
@%p14 bra $L__BB245_16;

st.global.f32 [%rd16], %f29;

$L__BB245_16:
ld.shared.f32 %f156, [%r21+640];
add.ftz.f32 %f157, %f156, %f211;
ld.shared.f32 %f158, [%r21+1664];
add.ftz.f32 %f159, %f158, %f157;
ld.shared.f32 %f160, [%r21+2688];
add.ftz.f32 %f161, %f160, %f159;
mov.b32 %r1043, %f161;
shfl.sync.bfly.b32 %r1047|%p41, %r1043, %r1029, %r1028, %r1030;
mov.b32 %f162, %r1047;
add.ftz.f32 %f163, %f161, %f162;
mov.b32 %r1048, %f163;
shfl.sync.bfly.b32 %r1050|%p42, %r1048, %r1033, %r1028, %r1030;
mov.b32 %f164, %r1050;
add.ftz.f32 %f165, %f163, %f164;
mov.b32 %r1051, %f165;
shfl.sync.bfly.b32 %r1053|%p43, %r1051, %r1036, %r1028, %r1030;
mov.b32 %f166, %r1053;
add.ftz.f32 %f167, %f165, %f166;
mov.b32 %r1054, %f167;
shfl.sync.bfly.b32 %r1056|%p44, %r1054, %r1039, %r1028, %r1030;
mov.b32 %f168, %r1056;
add.ftz.f32 %f169, %f167, %f168;
mov.b32 %r1057, %f169;
shfl.sync.bfly.b32 %r1059|%p45, %r1057, %r1026, %r1028, %r1030;
mov.b32 %f170, %r1059;
add.ftz.f32 %f30, %f169, %f170;
@%p22 bra $L__BB245_18;

st.global.f32 [%rd16+4], %f30;

$L__BB245_18:
mul.lo.s32 %r23, %r26, 3;
ld.shared.f32 %f171, [%r21+768];
add.ftz.f32 %f172, %f171, %f210;
ld.shared.f32 %f173, [%r21+1792];
add.ftz.f32 %f174, %f173, %f172;
ld.shared.f32 %f175, [%r21+2816];
add.ftz.f32 %f176, %f175, %f174;
mov.b32 %r1060, %f176;
mov.u32 %r1061, 31;
mov.u32 %r1062, 16;
mov.u32 %r1063, -1;
shfl.sync.bfly.b32 %r1064|%p47, %r1060, %r1062, %r1061, %r1063;
mov.b32 %f177, %r1064;
add.ftz.f32 %f178, %f176, %f177;
mov.b32 %r1065, %f178;
mov.u32 %r1066, 8;
shfl.sync.bfly.b32 %r1067|%p48, %r1065, %r1066, %r1061, %r1063;
mov.b32 %f179, %r1067;
add.ftz.f32 %f180, %f178, %f179;
mov.b32 %r1068, %f180;
mov.u32 %r1069, 4;
shfl.sync.bfly.b32 %r1070|%p49, %r1068, %r1069, %r1061, %r1063;
mov.b32 %f181, %r1070;
add.ftz.f32 %f182, %f180, %f181;
mov.b32 %r1071, %f182;
mov.u32 %r1072, 2;
shfl.sync.bfly.b32 %r1073|%p50, %r1071, %r1072, %r1061, %r1063;
mov.b32 %f183, %r1073;
add.ftz.f32 %f184, %f182, %f183;
mov.b32 %r1074, %f184;
mov.u32 %r1075, 1;
shfl.sync.bfly.b32 %r1076|%p51, %r1074, %r1075, %r1061, %r1063;
mov.b32 %f185, %r1076;
add.ftz.f32 %f31, %f184, %f185;
@%p14 bra $L__BB245_20;

cvt.u64.u32 %rd108, %r23;
add.s64 %rd109, %rd108, %rd14;
shl.b64 %rd110, %rd109, 2;
add.s64 %rd111, %rd1, %rd110;
st.global.f32 [%rd111], %f31;

$L__BB245_20:
ld.shared.f32 %f186, [%r21+896];
add.ftz.f32 %f187, %f186, %f209;
ld.shared.f32 %f188, [%r21+1920];
add.ftz.f32 %f189, %f188, %f187;
ld.shared.f32 %f190, [%r21+2944];
add.ftz.f32 %f191, %f190, %f189;
mov.b32 %r1077, %f191;
shfl.sync.bfly.b32 %r1081|%p53, %r1077, %r1062, %r1061, %r1063;
mov.b32 %f192, %r1081;
add.ftz.f32 %f193, %f191, %f192;
mov.b32 %r1082, %f193;
shfl.sync.bfly.b32 %r1084|%p54, %r1082, %r1066, %r1061, %r1063;
mov.b32 %f194, %r1084;
add.ftz.f32 %f195, %f193, %f194;
mov.b32 %r1085, %f195;
shfl.sync.bfly.b32 %r1087|%p55, %r1085, %r1069, %r1061, %r1063;
mov.b32 %f196, %r1087;
add.ftz.f32 %f197, %f195, %f196;
mov.b32 %r1088, %f197;
shfl.sync.bfly.b32 %r1090|%p56, %r1088, %r1072, %r1061, %r1063;
mov.b32 %f198, %r1090;
add.ftz.f32 %f199, %f197, %f198;
mov.b32 %r1091, %f199;
shfl.sync.bfly.b32 %r1093|%p57, %r1091, %r1075, %r1061, %r1063;
mov.b32 %f200, %r1093;
add.ftz.f32 %f32, %f199, %f200;
@%p22 bra $L__BB245_22;

add.s32 %r1094, %r23, 1;
cvt.u64.u32 %rd112, %r1094;
add.s64 %rd113, %rd112, %rd14;
shl.b64 %rd114, %rd113, 2;
add.s64 %rd115, %rd1, %rd114;
st.global.f32 [%rd115], %f32;

$L__BB245_22:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type18ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type18ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type18ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type18ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type18ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type18ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type18ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type18ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type18ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<71>;
.reg .b16 %rs<39>;
.reg .f32 %f<231>;
.reg .b32 %r<1316>;
.reg .b64 %rd<132>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type18ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1280];

ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type18ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type18ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type18ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type18ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type18ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type18ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type18ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type18ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type18ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type18ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type18ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type18ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type18ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type18ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type18ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type18ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type18ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r42, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r42, %r2;
mov.u32 %r43, %ctaid.x;
shl.b32 %r4, %r43, 1;
shr.s32 %r44, %r3, 31;
shr.u32 %r45, %r44, 29;
add.s32 %r46, %r3, %r45;
shr.s32 %r1315, %r46, 3;
shr.u32 %r6, %r41, 8;
setp.ge.s32 %p3, %r1315, %r6;
mov.f32 %f221, 0f00000000;
mov.f32 %f222, %f221;
mov.f32 %f223, %f221;
mov.f32 %f224, %f221;
mov.f32 %f225, %f221;
mov.f32 %f226, %f221;
mov.f32 %f227, %f221;
mov.f32 %f228, %f221;
mov.f32 %f229, %f221;
mov.f32 %f230, %f221;
@%p3 bra $L__BB246_3;

mov.u32 %r47, %ctaid.y;
mov.u32 %r48, %ctaid.z;
mul.lo.s32 %r49, %r47, %r33;
cvt.u64.u32 %rd20, %r49;
mul.lo.s32 %r50, %r48, %r39;
cvt.u64.u32 %rd21, %r50;
add.s64 %rd1, %rd21, %rd20;
mul.hi.u32 %r51, %r48, %r35;
add.s32 %r52, %r51, %r48;
shr.u32 %r53, %r52, %r36;
mul.lo.s32 %r54, %r53, %r38;
mul.hi.u32 %r55, %r47, %r29;
add.s32 %r56, %r55, %r47;
shr.u32 %r57, %r56, %r30;
mul.lo.s32 %r58, %r57, %r32;
cvta.to.global.u64 %rd131, %rd17;
and.b32 %r62, %r46, -8;
sub.s32 %r63, %r3, %r62;
shl.b32 %r64, %r63, 2;
shl.b32 %r65, %r63, 1;
add.s32 %r66, %r65, 33;
cvt.s64.s32 %rd3, %r63;
shl.b32 %r67, %r27, 1;
shl.b32 %r1310, %r1315, 3;
add.s32 %r1314, %r1310, %r67;
mad.lo.s32 %r1313, %r27, 3, %r1310;
shl.b32 %r68, %r27, 2;
add.s32 %r1312, %r1310, %r68;
add.s32 %r1311, %r27, %r1310;
add.s32 %r69, %r1315, %r54;
add.s32 %r70, %r69, %r58;
add.s32 %r71, %r4, 1;
mad.lo.s32 %r72, %r26, %r71, %r70;
mul.wide.s32 %rd4, %r72, 98;
mad.lo.s32 %r73, %r4, %r26, %r1315;
add.s32 %r74, %r73, %r54;
add.s32 %r75, %r74, %r58;
mul.wide.s32 %rd5, %r75, 98;
mul.wide.s32 %rd22, %r64, 2;
add.s64 %rd23, %rd5, %rd22;
add.s64 %rd6, %rd23, 8;
mul.wide.s32 %rd24, %r66, 2;
add.s64 %rd25, %rd24, -2;
add.s64 %rd26, %rd25, %rd4;
add.s64 %rd7, %rd26, 2;
add.s64 %rd27, %rd25, %rd5;
add.s64 %rd8, %rd27, 2;
add.s64 %rd28, %rd4, %rd22;
add.s64 %rd9, %rd28, 4;
cvta.to.global.u64 %rd10, %rd18;
mov.f32 %f221, 0f00000000;

$L__BB246_2:
cvt.u64.u32 %rd29, %r1310;
add.s64 %rd30, %rd1, %rd29;
add.s64 %rd31, %rd30, %rd3;
mul.lo.s64 %rd32, %rd31, 36;
add.s64 %rd33, %rd131, %rd6;
ld.global.nc.u16 %rs11, [%rd33+-4];
ld.global.nc.u16 %rs12, [%rd33+-2];
ld.global.nc.u16 %rs13, [%rd33];
add.s64 %rd34, %rd131, %rd8;
ld.global.nc.u16 %rs14, [%rd34];
cvt.u32.u16 %r566, %rs14;
ld.global.nc.u16 %rs15, [%rd34+2];
cvt.u32.u16 %r567, %rs15;
mov.b32 %r568, {%rs14, %rs15};
ld.global.nc.u16 %rs16, [%rd33+-6];
shl.b16 %rs17, %rs16, 2;
cvt.u64.u16 %rd35, %rs17;
and.b64 %rd36, %rd35, 1020;
mov.u64 %rd37, iq3xxs_grid;
add.s64 %rd38, %rd37, %rd36;
ld.global.nc.u32 %r569, [%rd38];
not.b32 %r570, %r569;
shr.u16 %rs18, %rs16, 8;
cvt.u32.u16 %r571, %rs18;
mul.wide.u32 %rd39, %r571, 4;
add.s64 %rd40, %rd37, %rd39;
ld.global.nc.u32 %r572, [%rd40];
not.b32 %r573, %r572;
and.b32 %r574, %r566, 255;
popc.b32 %r575, %r574;
shl.b32 %r576, %r575, 7;
and.b32 %r577, %r576, 128;
xor.b32 %r578, %r577, %r574;
mul.lo.s32 %r579, %r578, 16843009;
and.b32 %r580, %r579, 134480385;
or.b32 %r581, %r580, -2139062144;
add.s32 %r582, %r581, -16843009;
or.b32 %r468, %r582, %r580;
mov.u32 %r76, %r468;

prmt.b32 %r76, %r76, 0, 0xba98;

xor.b32 %r583, %r76, %r569;
or.b32 %r584, %r583, -2139062144;
and.b32 %r585, %r76, 2139062143;
and.b32 %r586, %r570, -2139062144;
sub.s32 %r587, %r584, %r585;
xor.b32 %r81, %r586, %r587;
add.s64 %rd41, %rd10, %rd32;
ld.global.nc.u32 %r82, [%rd41+4];
and.b32 %r588, %r579, -2143281136;
or.b32 %r589, %r588, -2139062144;
add.s32 %r590, %r589, -16843009;
or.b32 %r470, %r590, %r588;
mov.u32 %r78, %r470;

prmt.b32 %r78, %r78, 0, 0xba98;

xor.b32 %r591, %r78, %r572;
or.b32 %r592, %r591, -2139062144;
and.b32 %r593, %r78, 2139062143;
and.b32 %r594, %r573, -2139062144;
sub.s32 %r595, %r592, %r593;
xor.b32 %r85, %r594, %r595;
ld.global.nc.u32 %r86, [%rd41+8];
mov.u32 %r524, 0;

dp4a.s32.s32 %r80, %r81, %r82, %r524;


dp4a.s32.s32 %r84, %r85, %r86, %r80;

shl.b16 %rs19, %rs11, 2;
cvt.u64.u16 %rd42, %rs19;
and.b64 %rd43, %rd42, 1020;
add.s64 %rd44, %rd37, %rd43;
ld.global.nc.u32 %r596, [%rd44];
not.b32 %r597, %r596;
shr.u16 %rs20, %rs11, 8;
cvt.u32.u16 %r598, %rs20;
mul.wide.u32 %rd45, %r598, 4;
add.s64 %rd46, %rd37, %rd45;
ld.global.nc.u32 %r599, [%rd46];
not.b32 %r600, %r599;
shr.u32 %r601, %r566, 7;
and.b32 %r602, %r601, 255;
popc.b32 %r603, %r602;
shl.b32 %r604, %r603, 7;
and.b32 %r605, %r604, 128;
xor.b32 %r606, %r605, %r602;
mul.lo.s32 %r607, %r606, 16843009;
and.b32 %r608, %r607, 134480385;
or.b32 %r609, %r608, -2139062144;
add.s32 %r610, %r609, -16843009;
or.b32 %r480, %r610, %r608;
mov.u32 %r88, %r480;

prmt.b32 %r88, %r88, 0, 0xba98;

xor.b32 %r611, %r88, %r596;
or.b32 %r612, %r611, -2139062144;
and.b32 %r613, %r88, 2139062143;
and.b32 %r614, %r597, -2139062144;
sub.s32 %r615, %r612, %r613;
xor.b32 %r93, %r614, %r615;
ld.global.nc.u32 %r94, [%rd41+12];
and.b32 %r616, %r607, -2143281136;
or.b32 %r617, %r616, -2139062144;
add.s32 %r618, %r617, -16843009;
or.b32 %r482, %r618, %r616;
mov.u32 %r90, %r482;

prmt.b32 %r90, %r90, 0, 0xba98;

xor.b32 %r619, %r90, %r599;
or.b32 %r620, %r619, -2139062144;
and.b32 %r621, %r90, 2139062143;
and.b32 %r622, %r600, -2139062144;
sub.s32 %r623, %r620, %r621;
xor.b32 %r97, %r622, %r623;
ld.global.nc.u32 %r98, [%rd41+16];

dp4a.s32.s32 %r92, %r93, %r94, %r84;


dp4a.s32.s32 %r96, %r97, %r98, %r92;

shl.b16 %rs21, %rs12, 2;
cvt.u64.u16 %rd47, %rs21;
and.b64 %rd48, %rd47, 1020;
add.s64 %rd49, %rd37, %rd48;
ld.global.nc.u32 %r624, [%rd49];
not.b32 %r625, %r624;
shr.u16 %rs22, %rs12, 8;
cvt.u32.u16 %r626, %rs22;
mul.wide.u32 %rd50, %r626, 4;
add.s64 %rd51, %rd37, %rd50;
ld.global.nc.u32 %r627, [%rd51];
not.b32 %r628, %r627;
shr.u32 %r629, %r568, 14;
and.b32 %r630, %r629, 255;
popc.b32 %r631, %r630;
shl.b32 %r632, %r631, 7;
and.b32 %r633, %r632, 128;
xor.b32 %r634, %r633, %r630;
mul.lo.s32 %r635, %r634, 16843009;
and.b32 %r636, %r635, 134480385;
or.b32 %r637, %r636, -2139062144;
add.s32 %r638, %r637, -16843009;
or.b32 %r492, %r638, %r636;
mov.u32 %r100, %r492;

prmt.b32 %r100, %r100, 0, 0xba98;

xor.b32 %r639, %r100, %r624;
or.b32 %r640, %r639, -2139062144;
and.b32 %r641, %r100, 2139062143;
and.b32 %r642, %r625, -2139062144;
sub.s32 %r643, %r640, %r641;
xor.b32 %r105, %r642, %r643;
ld.global.nc.u32 %r106, [%rd41+20];
and.b32 %r644, %r635, -2143281136;
or.b32 %r645, %r644, -2139062144;
add.s32 %r646, %r645, -16843009;
or.b32 %r494, %r646, %r644;
mov.u32 %r102, %r494;

prmt.b32 %r102, %r102, 0, 0xba98;

xor.b32 %r647, %r102, %r627;
or.b32 %r648, %r647, -2139062144;
and.b32 %r649, %r102, 2139062143;
and.b32 %r650, %r628, -2139062144;
sub.s32 %r651, %r648, %r649;
xor.b32 %r109, %r650, %r651;
ld.global.nc.u32 %r110, [%rd41+24];

dp4a.s32.s32 %r104, %r105, %r106, %r96;


dp4a.s32.s32 %r108, %r109, %r110, %r104;

shl.b16 %rs23, %rs13, 2;
cvt.u64.u16 %rd52, %rs23;
and.b64 %rd53, %rd52, 1020;
add.s64 %rd54, %rd37, %rd53;
ld.global.nc.u32 %r652, [%rd54];
not.b32 %r653, %r652;
shr.u16 %rs24, %rs13, 8;
cvt.u32.u16 %r654, %rs24;
mul.wide.u32 %rd55, %r654, 4;
add.s64 %rd56, %rd37, %rd55;
ld.global.nc.u32 %r655, [%rd56];
not.b32 %r656, %r655;
shr.u32 %r657, %r567, 5;
and.b32 %r658, %r657, 255;
popc.b32 %r659, %r658;
shl.b32 %r660, %r659, 7;
and.b32 %r661, %r660, 128;
xor.b32 %r662, %r661, %r658;
mul.lo.s32 %r663, %r662, 16843009;
and.b32 %r664, %r663, 134480385;
or.b32 %r665, %r664, -2139062144;
add.s32 %r666, %r665, -16843009;
or.b32 %r504, %r666, %r664;
mov.u32 %r112, %r504;

prmt.b32 %r112, %r112, 0, 0xba98;

xor.b32 %r667, %r112, %r652;
or.b32 %r668, %r667, -2139062144;
and.b32 %r669, %r112, 2139062143;
and.b32 %r670, %r653, -2139062144;
sub.s32 %r671, %r668, %r669;
xor.b32 %r117, %r670, %r671;
ld.global.nc.u32 %r118, [%rd41+28];
and.b32 %r672, %r663, -2143281136;
or.b32 %r673, %r672, -2139062144;
add.s32 %r674, %r673, -16843009;
or.b32 %r506, %r674, %r672;
mov.u32 %r114, %r506;

prmt.b32 %r114, %r114, 0, 0xba98;

xor.b32 %r675, %r114, %r655;
or.b32 %r676, %r675, -2139062144;
and.b32 %r677, %r114, 2139062143;
and.b32 %r678, %r656, -2139062144;
sub.s32 %r679, %r676, %r677;
xor.b32 %r121, %r678, %r679;
ld.global.nc.u32 %r122, [%rd41+32];

dp4a.s32.s32 %r116, %r117, %r118, %r108;

shr.u32 %r680, %r567, 12;

dp4a.s32.s32 %r120, %r121, %r122, %r116;

shr.u32 %r681, %r120, 31;
add.s32 %r682, %r120, %r681;
shr.s32 %r683, %r682, 1;
mad.lo.s32 %r684, %r120, %r680, %r683;
shr.u32 %r685, %r684, 31;
add.s32 %r686, %r684, %r685;
shr.s32 %r687, %r686, 1;
add.s64 %rd57, %rd131, %rd5;
ld.global.nc.u16 %rs1, [%rd57];

{ cvt.f32.f16 %f61, %rs1;}


ld.global.nc.u32 %r124, [%rd41];

{.reg .f16 low,high;
mov.b32 {low,high},%r124;
cvt.f32.f16 %f62, low;}


mul.ftz.f32 %f81, %f61, %f62;
cvt.rn.f32.s32 %f82, %r687;
fma.rn.ftz.f32 %f230, %f81, %f82, %f230;
add.s64 %rd58, %rd131, %rd9;
ld.global.nc.u16 %rs25, [%rd58];
ld.global.nc.u16 %rs26, [%rd58+2];
ld.global.nc.u16 %rs27, [%rd58+4];
add.s64 %rd59, %rd131, %rd7;
ld.global.nc.u16 %rs28, [%rd59];
cvt.u32.u16 %r688, %rs28;
ld.global.nc.u16 %rs29, [%rd59+2];
cvt.u32.u16 %r689, %rs29;
mov.b32 %r690, {%rs28, %rs29};
ld.global.nc.u16 %rs30, [%rd58+-2];
shl.b16 %rs31, %rs30, 2;
cvt.u64.u16 %rd60, %rs31;
and.b64 %rd61, %rd60, 1020;
add.s64 %rd62, %rd37, %rd61;
ld.global.nc.u32 %r691, [%rd62];
not.b32 %r692, %r691;
shr.u16 %rs32, %rs30, 8;
cvt.u32.u16 %r693, %rs32;
mul.wide.u32 %rd63, %r693, 4;
add.s64 %rd64, %rd37, %rd63;
ld.global.nc.u32 %r694, [%rd64];
not.b32 %r695, %r694;
and.b32 %r696, %r688, 255;
popc.b32 %r697, %r696;
shl.b32 %r698, %r697, 7;
and.b32 %r699, %r698, 128;
xor.b32 %r700, %r699, %r696;
mul.lo.s32 %r701, %r700, 16843009;
and.b32 %r702, %r701, 134480385;
or.b32 %r703, %r702, -2139062144;
add.s32 %r704, %r703, -16843009;
or.b32 %r517, %r704, %r702;
mov.u32 %r125, %r517;

prmt.b32 %r125, %r125, 0, 0xba98;

xor.b32 %r705, %r125, %r691;
or.b32 %r706, %r705, -2139062144;
and.b32 %r707, %r125, 2139062143;
and.b32 %r708, %r692, -2139062144;
sub.s32 %r709, %r706, %r707;
xor.b32 %r130, %r708, %r709;
and.b32 %r710, %r701, -2143281136;
or.b32 %r711, %r710, -2139062144;
add.s32 %r712, %r711, -16843009;
or.b32 %r519, %r712, %r710;
mov.u32 %r127, %r519;

prmt.b32 %r127, %r127, 0, 0xba98;

xor.b32 %r713, %r127, %r694;
or.b32 %r714, %r713, -2139062144;
and.b32 %r715, %r127, 2139062143;
and.b32 %r716, %r695, -2139062144;
sub.s32 %r717, %r714, %r715;
xor.b32 %r134, %r716, %r717;

dp4a.s32.s32 %r129, %r130, %r82, %r524;


dp4a.s32.s32 %r133, %r134, %r86, %r129;

shl.b16 %rs33, %rs25, 2;
cvt.u64.u16 %rd65, %rs33;
and.b64 %rd66, %rd65, 1020;
add.s64 %rd67, %rd37, %rd66;
ld.global.nc.u32 %r718, [%rd67];
not.b32 %r719, %r718;
shr.u16 %rs34, %rs25, 8;
cvt.u32.u16 %r720, %rs34;
mul.wide.u32 %rd68, %r720, 4;
add.s64 %rd69, %rd37, %rd68;
ld.global.nc.u32 %r721, [%rd69];
not.b32 %r722, %r721;
shr.u32 %r723, %r688, 7;
and.b32 %r724, %r723, 255;
popc.b32 %r725, %r724;
shl.b32 %r726, %r725, 7;
and.b32 %r727, %r726, 128;
xor.b32 %r728, %r727, %r724;
mul.lo.s32 %r729, %r728, 16843009;
and.b32 %r730, %r729, 134480385;
or.b32 %r731, %r730, -2139062144;
add.s32 %r732, %r731, -16843009;
or.b32 %r529, %r732, %r730;
mov.u32 %r137, %r529;

prmt.b32 %r137, %r137, 0, 0xba98;

xor.b32 %r733, %r137, %r718;
or.b32 %r734, %r733, -2139062144;
and.b32 %r735, %r137, 2139062143;
and.b32 %r736, %r719, -2139062144;
sub.s32 %r737, %r734, %r735;
xor.b32 %r142, %r736, %r737;
and.b32 %r738, %r729, -2143281136;
or.b32 %r739, %r738, -2139062144;
add.s32 %r740, %r739, -16843009;
or.b32 %r531, %r740, %r738;
mov.u32 %r139, %r531;

prmt.b32 %r139, %r139, 0, 0xba98;

xor.b32 %r741, %r139, %r721;
or.b32 %r742, %r741, -2139062144;
and.b32 %r743, %r139, 2139062143;
and.b32 %r744, %r722, -2139062144;
sub.s32 %r745, %r742, %r743;
xor.b32 %r146, %r744, %r745;

dp4a.s32.s32 %r141, %r142, %r94, %r133;


dp4a.s32.s32 %r145, %r146, %r98, %r141;

shl.b16 %rs35, %rs26, 2;
cvt.u64.u16 %rd70, %rs35;
and.b64 %rd71, %rd70, 1020;
add.s64 %rd72, %rd37, %rd71;
ld.global.nc.u32 %r746, [%rd72];
not.b32 %r747, %r746;
shr.u16 %rs36, %rs26, 8;
cvt.u32.u16 %r748, %rs36;
mul.wide.u32 %rd73, %r748, 4;
add.s64 %rd74, %rd37, %rd73;
ld.global.nc.u32 %r749, [%rd74];
not.b32 %r750, %r749;
shr.u32 %r751, %r690, 14;
and.b32 %r752, %r751, 255;
popc.b32 %r753, %r752;
shl.b32 %r754, %r753, 7;
and.b32 %r755, %r754, 128;
xor.b32 %r756, %r755, %r752;
mul.lo.s32 %r757, %r756, 16843009;
and.b32 %r758, %r757, 134480385;
or.b32 %r759, %r758, -2139062144;
add.s32 %r760, %r759, -16843009;
or.b32 %r541, %r760, %r758;
mov.u32 %r149, %r541;

prmt.b32 %r149, %r149, 0, 0xba98;

xor.b32 %r761, %r149, %r746;
or.b32 %r762, %r761, -2139062144;
and.b32 %r763, %r149, 2139062143;
and.b32 %r764, %r747, -2139062144;
sub.s32 %r765, %r762, %r763;
xor.b32 %r154, %r764, %r765;
and.b32 %r766, %r757, -2143281136;
or.b32 %r767, %r766, -2139062144;
add.s32 %r768, %r767, -16843009;
or.b32 %r543, %r768, %r766;
mov.u32 %r151, %r543;

prmt.b32 %r151, %r151, 0, 0xba98;

xor.b32 %r769, %r151, %r749;
or.b32 %r770, %r769, -2139062144;
and.b32 %r771, %r151, 2139062143;
and.b32 %r772, %r750, -2139062144;
sub.s32 %r773, %r770, %r771;
xor.b32 %r158, %r772, %r773;

dp4a.s32.s32 %r153, %r154, %r106, %r145;


dp4a.s32.s32 %r157, %r158, %r110, %r153;

shl.b16 %rs37, %rs27, 2;
cvt.u64.u16 %rd75, %rs37;
and.b64 %rd76, %rd75, 1020;
add.s64 %rd77, %rd37, %rd76;
ld.global.nc.u32 %r774, [%rd77];
not.b32 %r775, %r774;
shr.u16 %rs38, %rs27, 8;
cvt.u32.u16 %r776, %rs38;
mul.wide.u32 %rd78, %r776, 4;
add.s64 %rd79, %rd37, %rd78;
ld.global.nc.u32 %r777, [%rd79];
not.b32 %r778, %r777;
shr.u32 %r779, %r689, 5;
and.b32 %r780, %r779, 255;
popc.b32 %r781, %r780;
shl.b32 %r782, %r781, 7;
and.b32 %r783, %r782, 128;
xor.b32 %r784, %r783, %r780;
mul.lo.s32 %r785, %r784, 16843009;
and.b32 %r786, %r785, 134480385;
or.b32 %r787, %r786, -2139062144;
add.s32 %r788, %r787, -16843009;
or.b32 %r553, %r788, %r786;
mov.u32 %r161, %r553;

prmt.b32 %r161, %r161, 0, 0xba98;

xor.b32 %r789, %r161, %r774;
or.b32 %r790, %r789, -2139062144;
and.b32 %r791, %r161, 2139062143;
and.b32 %r792, %r775, -2139062144;
sub.s32 %r793, %r790, %r791;
xor.b32 %r166, %r792, %r793;
and.b32 %r794, %r785, -2143281136;
or.b32 %r795, %r794, -2139062144;
add.s32 %r796, %r795, -16843009;
or.b32 %r555, %r796, %r794;
mov.u32 %r163, %r555;

prmt.b32 %r163, %r163, 0, 0xba98;

xor.b32 %r797, %r163, %r777;
or.b32 %r798, %r797, -2139062144;
and.b32 %r799, %r163, 2139062143;
and.b32 %r800, %r778, -2139062144;
sub.s32 %r801, %r798, %r799;
xor.b32 %r170, %r800, %r801;

dp4a.s32.s32 %r165, %r166, %r118, %r157;

shr.u32 %r802, %r689, 12;

dp4a.s32.s32 %r169, %r170, %r122, %r165;

shr.u32 %r803, %r169, 31;
add.s32 %r804, %r169, %r803;
shr.s32 %r805, %r804, 1;
mad.lo.s32 %r806, %r169, %r802, %r805;
shr.u32 %r807, %r806, 31;
add.s32 %r808, %r806, %r807;
shr.s32 %r809, %r808, 1;
add.s64 %rd80, %rd131, %rd4;
ld.global.nc.u16 %rs2, [%rd80];

{ cvt.f32.f16 %f63, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r124;
cvt.f32.f16 %f64, low;}


mul.ftz.f32 %f83, %f63, %f64;
cvt.rn.f32.s32 %f84, %r809;
fma.rn.ftz.f32 %f229, %f83, %f84, %f229;
cvt.u64.u32 %rd81, %r1311;
add.s64 %rd82, %rd1, %rd81;
add.s64 %rd83, %rd82, %rd3;
mul.lo.s64 %rd84, %rd83, 36;
mov.u32 %r174, %r468;

prmt.b32 %r174, %r174, 0, 0xba98;

xor.b32 %r810, %r174, %r569;
or.b32 %r811, %r810, -2139062144;
and.b32 %r812, %r174, 2139062143;
sub.s32 %r813, %r811, %r812;
xor.b32 %r179, %r586, %r813;
add.s64 %rd85, %rd10, %rd84;
ld.global.nc.u32 %r180, [%rd85+4];
mov.u32 %r176, %r470;

prmt.b32 %r176, %r176, 0, 0xba98;

xor.b32 %r814, %r176, %r572;
or.b32 %r815, %r814, -2139062144;
and.b32 %r816, %r176, 2139062143;
sub.s32 %r817, %r815, %r816;
xor.b32 %r183, %r594, %r817;
ld.global.nc.u32 %r184, [%rd85+8];

dp4a.s32.s32 %r178, %r179, %r180, %r524;


dp4a.s32.s32 %r182, %r183, %r184, %r178;

mov.u32 %r186, %r480;

prmt.b32 %r186, %r186, 0, 0xba98;

xor.b32 %r818, %r186, %r596;
or.b32 %r819, %r818, -2139062144;
and.b32 %r820, %r186, 2139062143;
sub.s32 %r821, %r819, %r820;
xor.b32 %r191, %r614, %r821;
ld.global.nc.u32 %r192, [%rd85+12];
mov.u32 %r188, %r482;

prmt.b32 %r188, %r188, 0, 0xba98;

xor.b32 %r822, %r188, %r599;
or.b32 %r823, %r822, -2139062144;
and.b32 %r824, %r188, 2139062143;
sub.s32 %r825, %r823, %r824;
xor.b32 %r195, %r622, %r825;
ld.global.nc.u32 %r196, [%rd85+16];

dp4a.s32.s32 %r190, %r191, %r192, %r182;


dp4a.s32.s32 %r194, %r195, %r196, %r190;

mov.u32 %r198, %r492;

prmt.b32 %r198, %r198, 0, 0xba98;

xor.b32 %r826, %r198, %r624;
or.b32 %r827, %r826, -2139062144;
and.b32 %r828, %r198, 2139062143;
sub.s32 %r829, %r827, %r828;
xor.b32 %r203, %r642, %r829;
ld.global.nc.u32 %r204, [%rd85+20];
mov.u32 %r200, %r494;

prmt.b32 %r200, %r200, 0, 0xba98;

xor.b32 %r830, %r200, %r627;
or.b32 %r831, %r830, -2139062144;
and.b32 %r832, %r200, 2139062143;
sub.s32 %r833, %r831, %r832;
xor.b32 %r207, %r650, %r833;
ld.global.nc.u32 %r208, [%rd85+24];

dp4a.s32.s32 %r202, %r203, %r204, %r194;


dp4a.s32.s32 %r206, %r207, %r208, %r202;

mov.u32 %r210, %r504;

prmt.b32 %r210, %r210, 0, 0xba98;

xor.b32 %r834, %r210, %r652;
or.b32 %r835, %r834, -2139062144;
and.b32 %r836, %r210, 2139062143;
sub.s32 %r837, %r835, %r836;
xor.b32 %r215, %r670, %r837;
ld.global.nc.u32 %r216, [%rd85+28];
mov.u32 %r212, %r506;

prmt.b32 %r212, %r212, 0, 0xba98;

xor.b32 %r838, %r212, %r655;
or.b32 %r839, %r838, -2139062144;
and.b32 %r840, %r212, 2139062143;
sub.s32 %r841, %r839, %r840;
xor.b32 %r219, %r678, %r841;
ld.global.nc.u32 %r220, [%rd85+32];

dp4a.s32.s32 %r214, %r215, %r216, %r206;


dp4a.s32.s32 %r218, %r219, %r220, %r214;

shr.u32 %r842, %r218, 31;
add.s32 %r843, %r218, %r842;
shr.s32 %r844, %r843, 1;
mad.lo.s32 %r845, %r218, %r680, %r844;
shr.u32 %r846, %r845, 31;
add.s32 %r847, %r845, %r846;
shr.s32 %r848, %r847, 1;

{ cvt.f32.f16 %f65, %rs1;}


ld.global.nc.u32 %r222, [%rd85];

{.reg .f16 low,high;
mov.b32 {low,high},%r222;
cvt.f32.f16 %f66, low;}


mul.ftz.f32 %f85, %f65, %f66;
cvt.rn.f32.s32 %f86, %r848;
fma.rn.ftz.f32 %f228, %f85, %f86, %f228;
mov.u32 %r223, %r517;

prmt.b32 %r223, %r223, 0, 0xba98;

xor.b32 %r849, %r223, %r691;
or.b32 %r850, %r849, -2139062144;
and.b32 %r851, %r223, 2139062143;
sub.s32 %r852, %r850, %r851;
xor.b32 %r228, %r708, %r852;
mov.u32 %r225, %r519;

prmt.b32 %r225, %r225, 0, 0xba98;

xor.b32 %r853, %r225, %r694;
or.b32 %r854, %r853, -2139062144;
and.b32 %r855, %r225, 2139062143;
sub.s32 %r856, %r854, %r855;
xor.b32 %r232, %r716, %r856;

dp4a.s32.s32 %r227, %r228, %r180, %r524;


dp4a.s32.s32 %r231, %r232, %r184, %r227;

mov.u32 %r235, %r529;

prmt.b32 %r235, %r235, 0, 0xba98;

xor.b32 %r857, %r235, %r718;
or.b32 %r858, %r857, -2139062144;
and.b32 %r859, %r235, 2139062143;
sub.s32 %r860, %r858, %r859;
xor.b32 %r240, %r736, %r860;
mov.u32 %r237, %r531;

prmt.b32 %r237, %r237, 0, 0xba98;

xor.b32 %r861, %r237, %r721;
or.b32 %r862, %r861, -2139062144;
and.b32 %r863, %r237, 2139062143;
sub.s32 %r864, %r862, %r863;
xor.b32 %r244, %r744, %r864;

dp4a.s32.s32 %r239, %r240, %r192, %r231;


dp4a.s32.s32 %r243, %r244, %r196, %r239;

mov.u32 %r247, %r541;

prmt.b32 %r247, %r247, 0, 0xba98;

xor.b32 %r865, %r247, %r746;
or.b32 %r866, %r865, -2139062144;
and.b32 %r867, %r247, 2139062143;
sub.s32 %r868, %r866, %r867;
xor.b32 %r252, %r764, %r868;
mov.u32 %r249, %r543;

prmt.b32 %r249, %r249, 0, 0xba98;

xor.b32 %r869, %r249, %r749;
or.b32 %r870, %r869, -2139062144;
and.b32 %r871, %r249, 2139062143;
sub.s32 %r872, %r870, %r871;
xor.b32 %r256, %r772, %r872;

dp4a.s32.s32 %r251, %r252, %r204, %r243;


dp4a.s32.s32 %r255, %r256, %r208, %r251;

mov.u32 %r259, %r553;

prmt.b32 %r259, %r259, 0, 0xba98;

xor.b32 %r873, %r259, %r774;
or.b32 %r874, %r873, -2139062144;
and.b32 %r875, %r259, 2139062143;
sub.s32 %r876, %r874, %r875;
xor.b32 %r264, %r792, %r876;
mov.u32 %r261, %r555;

prmt.b32 %r261, %r261, 0, 0xba98;

xor.b32 %r877, %r261, %r777;
or.b32 %r878, %r877, -2139062144;
and.b32 %r879, %r261, 2139062143;
sub.s32 %r880, %r878, %r879;
xor.b32 %r268, %r800, %r880;

dp4a.s32.s32 %r263, %r264, %r216, %r255;


dp4a.s32.s32 %r267, %r268, %r220, %r263;

shr.u32 %r881, %r267, 31;
add.s32 %r882, %r267, %r881;
shr.s32 %r883, %r882, 1;
mad.lo.s32 %r884, %r267, %r802, %r883;
shr.u32 %r885, %r884, 31;
add.s32 %r886, %r884, %r885;
shr.s32 %r887, %r886, 1;

{ cvt.f32.f16 %f67, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r222;
cvt.f32.f16 %f68, low;}


mul.ftz.f32 %f87, %f67, %f68;
cvt.rn.f32.s32 %f88, %r887;
fma.rn.ftz.f32 %f227, %f87, %f88, %f227;
cvt.u64.u32 %rd86, %r1314;
add.s64 %rd87, %rd1, %rd86;
add.s64 %rd88, %rd87, %rd3;
mul.lo.s64 %rd89, %rd88, 36;
mov.u32 %r272, %r468;

prmt.b32 %r272, %r272, 0, 0xba98;

xor.b32 %r888, %r272, %r569;
or.b32 %r889, %r888, -2139062144;
and.b32 %r890, %r272, 2139062143;
sub.s32 %r891, %r889, %r890;
xor.b32 %r277, %r586, %r891;
add.s64 %rd90, %rd10, %rd89;
ld.global.nc.u32 %r278, [%rd90+4];
mov.u32 %r274, %r470;

prmt.b32 %r274, %r274, 0, 0xba98;

xor.b32 %r892, %r274, %r572;
or.b32 %r893, %r892, -2139062144;
and.b32 %r894, %r274, 2139062143;
sub.s32 %r895, %r893, %r894;
xor.b32 %r281, %r594, %r895;
ld.global.nc.u32 %r282, [%rd90+8];

dp4a.s32.s32 %r276, %r277, %r278, %r524;


dp4a.s32.s32 %r280, %r281, %r282, %r276;

mov.u32 %r284, %r480;

prmt.b32 %r284, %r284, 0, 0xba98;

xor.b32 %r896, %r284, %r596;
or.b32 %r897, %r896, -2139062144;
and.b32 %r898, %r284, 2139062143;
sub.s32 %r899, %r897, %r898;
xor.b32 %r289, %r614, %r899;
ld.global.nc.u32 %r290, [%rd90+12];
mov.u32 %r286, %r482;

prmt.b32 %r286, %r286, 0, 0xba98;

xor.b32 %r900, %r286, %r599;
or.b32 %r901, %r900, -2139062144;
and.b32 %r902, %r286, 2139062143;
sub.s32 %r903, %r901, %r902;
xor.b32 %r293, %r622, %r903;
ld.global.nc.u32 %r294, [%rd90+16];

dp4a.s32.s32 %r288, %r289, %r290, %r280;


dp4a.s32.s32 %r292, %r293, %r294, %r288;

mov.u32 %r296, %r492;

prmt.b32 %r296, %r296, 0, 0xba98;

xor.b32 %r904, %r296, %r624;
or.b32 %r905, %r904, -2139062144;
and.b32 %r906, %r296, 2139062143;
sub.s32 %r907, %r905, %r906;
xor.b32 %r301, %r642, %r907;
ld.global.nc.u32 %r302, [%rd90+20];
mov.u32 %r298, %r494;

prmt.b32 %r298, %r298, 0, 0xba98;

xor.b32 %r908, %r298, %r627;
or.b32 %r909, %r908, -2139062144;
and.b32 %r910, %r298, 2139062143;
sub.s32 %r911, %r909, %r910;
xor.b32 %r305, %r650, %r911;
ld.global.nc.u32 %r306, [%rd90+24];

dp4a.s32.s32 %r300, %r301, %r302, %r292;


dp4a.s32.s32 %r304, %r305, %r306, %r300;

mov.u32 %r308, %r504;

prmt.b32 %r308, %r308, 0, 0xba98;

xor.b32 %r912, %r308, %r652;
or.b32 %r913, %r912, -2139062144;
and.b32 %r914, %r308, 2139062143;
sub.s32 %r915, %r913, %r914;
xor.b32 %r313, %r670, %r915;
ld.global.nc.u32 %r314, [%rd90+28];
mov.u32 %r310, %r506;

prmt.b32 %r310, %r310, 0, 0xba98;

xor.b32 %r916, %r310, %r655;
or.b32 %r917, %r916, -2139062144;
and.b32 %r918, %r310, 2139062143;
sub.s32 %r919, %r917, %r918;
xor.b32 %r317, %r678, %r919;
ld.global.nc.u32 %r318, [%rd90+32];

dp4a.s32.s32 %r312, %r313, %r314, %r304;


dp4a.s32.s32 %r316, %r317, %r318, %r312;

shr.u32 %r920, %r316, 31;
add.s32 %r921, %r316, %r920;
shr.s32 %r922, %r921, 1;
mad.lo.s32 %r923, %r316, %r680, %r922;
shr.u32 %r924, %r923, 31;
add.s32 %r925, %r923, %r924;
shr.s32 %r926, %r925, 1;

{ cvt.f32.f16 %f69, %rs1;}


ld.global.nc.u32 %r320, [%rd90];

{.reg .f16 low,high;
mov.b32 {low,high},%r320;
cvt.f32.f16 %f70, low;}


mul.ftz.f32 %f89, %f69, %f70;
cvt.rn.f32.s32 %f90, %r926;
fma.rn.ftz.f32 %f226, %f89, %f90, %f226;
mov.u32 %r321, %r517;

prmt.b32 %r321, %r321, 0, 0xba98;

xor.b32 %r927, %r321, %r691;
or.b32 %r928, %r927, -2139062144;
and.b32 %r929, %r321, 2139062143;
sub.s32 %r930, %r928, %r929;
xor.b32 %r326, %r708, %r930;
mov.u32 %r323, %r519;

prmt.b32 %r323, %r323, 0, 0xba98;

xor.b32 %r931, %r323, %r694;
or.b32 %r932, %r931, -2139062144;
and.b32 %r933, %r323, 2139062143;
sub.s32 %r934, %r932, %r933;
xor.b32 %r330, %r716, %r934;

dp4a.s32.s32 %r325, %r326, %r278, %r524;


dp4a.s32.s32 %r329, %r330, %r282, %r325;

mov.u32 %r333, %r529;

prmt.b32 %r333, %r333, 0, 0xba98;

xor.b32 %r935, %r333, %r718;
or.b32 %r936, %r935, -2139062144;
and.b32 %r937, %r333, 2139062143;
sub.s32 %r938, %r936, %r937;
xor.b32 %r338, %r736, %r938;
mov.u32 %r335, %r531;

prmt.b32 %r335, %r335, 0, 0xba98;

xor.b32 %r939, %r335, %r721;
or.b32 %r940, %r939, -2139062144;
and.b32 %r941, %r335, 2139062143;
sub.s32 %r942, %r940, %r941;
xor.b32 %r342, %r744, %r942;

dp4a.s32.s32 %r337, %r338, %r290, %r329;


dp4a.s32.s32 %r341, %r342, %r294, %r337;

mov.u32 %r345, %r541;

prmt.b32 %r345, %r345, 0, 0xba98;

xor.b32 %r943, %r345, %r746;
or.b32 %r944, %r943, -2139062144;
and.b32 %r945, %r345, 2139062143;
sub.s32 %r946, %r944, %r945;
xor.b32 %r350, %r764, %r946;
mov.u32 %r347, %r543;

prmt.b32 %r347, %r347, 0, 0xba98;

xor.b32 %r947, %r347, %r749;
or.b32 %r948, %r947, -2139062144;
and.b32 %r949, %r347, 2139062143;
sub.s32 %r950, %r948, %r949;
xor.b32 %r354, %r772, %r950;

dp4a.s32.s32 %r349, %r350, %r302, %r341;


dp4a.s32.s32 %r353, %r354, %r306, %r349;

mov.u32 %r357, %r553;

prmt.b32 %r357, %r357, 0, 0xba98;

xor.b32 %r951, %r357, %r774;
or.b32 %r952, %r951, -2139062144;
and.b32 %r953, %r357, 2139062143;
sub.s32 %r954, %r952, %r953;
xor.b32 %r362, %r792, %r954;
mov.u32 %r359, %r555;

prmt.b32 %r359, %r359, 0, 0xba98;

xor.b32 %r955, %r359, %r777;
or.b32 %r956, %r955, -2139062144;
and.b32 %r957, %r359, 2139062143;
sub.s32 %r958, %r956, %r957;
xor.b32 %r366, %r800, %r958;

dp4a.s32.s32 %r361, %r362, %r314, %r353;


dp4a.s32.s32 %r365, %r366, %r318, %r361;

shr.u32 %r959, %r365, 31;
add.s32 %r960, %r365, %r959;
shr.s32 %r961, %r960, 1;
mad.lo.s32 %r962, %r365, %r802, %r961;
shr.u32 %r963, %r962, 31;
add.s32 %r964, %r962, %r963;
shr.s32 %r965, %r964, 1;

{ cvt.f32.f16 %f71, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r320;
cvt.f32.f16 %f72, low;}


mul.ftz.f32 %f91, %f71, %f72;
cvt.rn.f32.s32 %f92, %r965;
fma.rn.ftz.f32 %f225, %f91, %f92, %f225;
cvt.u64.u32 %rd91, %r1313;
add.s64 %rd92, %rd1, %rd91;
add.s64 %rd93, %rd92, %rd3;
mul.lo.s64 %rd94, %rd93, 36;
mov.u32 %r370, %r468;

prmt.b32 %r370, %r370, 0, 0xba98;

xor.b32 %r966, %r370, %r569;
or.b32 %r967, %r966, -2139062144;
and.b32 %r968, %r370, 2139062143;
sub.s32 %r969, %r967, %r968;
xor.b32 %r375, %r586, %r969;
add.s64 %rd95, %rd10, %rd94;
ld.global.nc.u32 %r376, [%rd95+4];
mov.u32 %r372, %r470;

prmt.b32 %r372, %r372, 0, 0xba98;

xor.b32 %r970, %r372, %r572;
or.b32 %r971, %r970, -2139062144;
and.b32 %r972, %r372, 2139062143;
sub.s32 %r973, %r971, %r972;
xor.b32 %r379, %r594, %r973;
ld.global.nc.u32 %r380, [%rd95+8];

dp4a.s32.s32 %r374, %r375, %r376, %r524;


dp4a.s32.s32 %r378, %r379, %r380, %r374;

mov.u32 %r382, %r480;

prmt.b32 %r382, %r382, 0, 0xba98;

xor.b32 %r974, %r382, %r596;
or.b32 %r975, %r974, -2139062144;
and.b32 %r976, %r382, 2139062143;
sub.s32 %r977, %r975, %r976;
xor.b32 %r387, %r614, %r977;
ld.global.nc.u32 %r388, [%rd95+12];
mov.u32 %r384, %r482;

prmt.b32 %r384, %r384, 0, 0xba98;

xor.b32 %r978, %r384, %r599;
or.b32 %r979, %r978, -2139062144;
and.b32 %r980, %r384, 2139062143;
sub.s32 %r981, %r979, %r980;
xor.b32 %r391, %r622, %r981;
ld.global.nc.u32 %r392, [%rd95+16];

dp4a.s32.s32 %r386, %r387, %r388, %r378;


dp4a.s32.s32 %r390, %r391, %r392, %r386;

mov.u32 %r394, %r492;

prmt.b32 %r394, %r394, 0, 0xba98;

xor.b32 %r982, %r394, %r624;
or.b32 %r983, %r982, -2139062144;
and.b32 %r984, %r394, 2139062143;
sub.s32 %r985, %r983, %r984;
xor.b32 %r399, %r642, %r985;
ld.global.nc.u32 %r400, [%rd95+20];
mov.u32 %r396, %r494;

prmt.b32 %r396, %r396, 0, 0xba98;

xor.b32 %r986, %r396, %r627;
or.b32 %r987, %r986, -2139062144;
and.b32 %r988, %r396, 2139062143;
sub.s32 %r989, %r987, %r988;
xor.b32 %r403, %r650, %r989;
ld.global.nc.u32 %r404, [%rd95+24];

dp4a.s32.s32 %r398, %r399, %r400, %r390;


dp4a.s32.s32 %r402, %r403, %r404, %r398;

mov.u32 %r406, %r504;

prmt.b32 %r406, %r406, 0, 0xba98;

xor.b32 %r990, %r406, %r652;
or.b32 %r991, %r990, -2139062144;
and.b32 %r992, %r406, 2139062143;
sub.s32 %r993, %r991, %r992;
xor.b32 %r411, %r670, %r993;
ld.global.nc.u32 %r412, [%rd95+28];
mov.u32 %r408, %r506;

prmt.b32 %r408, %r408, 0, 0xba98;

xor.b32 %r994, %r408, %r655;
or.b32 %r995, %r994, -2139062144;
and.b32 %r996, %r408, 2139062143;
sub.s32 %r997, %r995, %r996;
xor.b32 %r415, %r678, %r997;
ld.global.nc.u32 %r416, [%rd95+32];

dp4a.s32.s32 %r410, %r411, %r412, %r402;


dp4a.s32.s32 %r414, %r415, %r416, %r410;

shr.u32 %r998, %r414, 31;
add.s32 %r999, %r414, %r998;
shr.s32 %r1000, %r999, 1;
mad.lo.s32 %r1001, %r414, %r680, %r1000;
shr.u32 %r1002, %r1001, 31;
add.s32 %r1003, %r1001, %r1002;
shr.s32 %r1004, %r1003, 1;

{ cvt.f32.f16 %f73, %rs1;}


ld.global.nc.u32 %r418, [%rd95];

{.reg .f16 low,high;
mov.b32 {low,high},%r418;
cvt.f32.f16 %f74, low;}


mul.ftz.f32 %f93, %f73, %f74;
cvt.rn.f32.s32 %f94, %r1004;
fma.rn.ftz.f32 %f224, %f93, %f94, %f224;
mov.u32 %r419, %r517;

prmt.b32 %r419, %r419, 0, 0xba98;

xor.b32 %r1005, %r419, %r691;
or.b32 %r1006, %r1005, -2139062144;
and.b32 %r1007, %r419, 2139062143;
sub.s32 %r1008, %r1006, %r1007;
xor.b32 %r424, %r708, %r1008;
mov.u32 %r421, %r519;

prmt.b32 %r421, %r421, 0, 0xba98;

xor.b32 %r1009, %r421, %r694;
or.b32 %r1010, %r1009, -2139062144;
and.b32 %r1011, %r421, 2139062143;
sub.s32 %r1012, %r1010, %r1011;
xor.b32 %r428, %r716, %r1012;

dp4a.s32.s32 %r423, %r424, %r376, %r524;


dp4a.s32.s32 %r427, %r428, %r380, %r423;

mov.u32 %r431, %r529;

prmt.b32 %r431, %r431, 0, 0xba98;

xor.b32 %r1013, %r431, %r718;
or.b32 %r1014, %r1013, -2139062144;
and.b32 %r1015, %r431, 2139062143;
sub.s32 %r1016, %r1014, %r1015;
xor.b32 %r436, %r736, %r1016;
mov.u32 %r433, %r531;

prmt.b32 %r433, %r433, 0, 0xba98;

xor.b32 %r1017, %r433, %r721;
or.b32 %r1018, %r1017, -2139062144;
and.b32 %r1019, %r433, 2139062143;
sub.s32 %r1020, %r1018, %r1019;
xor.b32 %r440, %r744, %r1020;

dp4a.s32.s32 %r435, %r436, %r388, %r427;


dp4a.s32.s32 %r439, %r440, %r392, %r435;

mov.u32 %r443, %r541;

prmt.b32 %r443, %r443, 0, 0xba98;

xor.b32 %r1021, %r443, %r746;
or.b32 %r1022, %r1021, -2139062144;
and.b32 %r1023, %r443, 2139062143;
sub.s32 %r1024, %r1022, %r1023;
xor.b32 %r448, %r764, %r1024;
mov.u32 %r445, %r543;

prmt.b32 %r445, %r445, 0, 0xba98;

xor.b32 %r1025, %r445, %r749;
or.b32 %r1026, %r1025, -2139062144;
and.b32 %r1027, %r445, 2139062143;
sub.s32 %r1028, %r1026, %r1027;
xor.b32 %r452, %r772, %r1028;

dp4a.s32.s32 %r447, %r448, %r400, %r439;


dp4a.s32.s32 %r451, %r452, %r404, %r447;

mov.u32 %r455, %r553;

prmt.b32 %r455, %r455, 0, 0xba98;

xor.b32 %r1029, %r455, %r774;
or.b32 %r1030, %r1029, -2139062144;
and.b32 %r1031, %r455, 2139062143;
sub.s32 %r1032, %r1030, %r1031;
xor.b32 %r460, %r792, %r1032;
mov.u32 %r457, %r555;

prmt.b32 %r457, %r457, 0, 0xba98;

xor.b32 %r1033, %r457, %r777;
or.b32 %r1034, %r1033, -2139062144;
and.b32 %r1035, %r457, 2139062143;
sub.s32 %r1036, %r1034, %r1035;
xor.b32 %r464, %r800, %r1036;

dp4a.s32.s32 %r459, %r460, %r412, %r451;


dp4a.s32.s32 %r463, %r464, %r416, %r459;

shr.u32 %r1037, %r463, 31;
add.s32 %r1038, %r463, %r1037;
shr.s32 %r1039, %r1038, 1;
mad.lo.s32 %r1040, %r463, %r802, %r1039;
shr.u32 %r1041, %r1040, 31;
add.s32 %r1042, %r1040, %r1041;
shr.s32 %r1043, %r1042, 1;

{ cvt.f32.f16 %f75, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r418;
cvt.f32.f16 %f76, low;}


mul.ftz.f32 %f95, %f75, %f76;
cvt.rn.f32.s32 %f96, %r1043;
fma.rn.ftz.f32 %f223, %f95, %f96, %f223;
cvt.u64.u32 %rd96, %r1312;
add.s64 %rd97, %rd1, %rd96;
add.s64 %rd98, %rd97, %rd3;
mul.lo.s64 %rd99, %rd98, 36;

prmt.b32 %r468, %r468, 0, 0xba98;

xor.b32 %r1044, %r468, %r569;
or.b32 %r1045, %r1044, -2139062144;
and.b32 %r1046, %r468, 2139062143;
sub.s32 %r1047, %r1045, %r1046;
xor.b32 %r473, %r586, %r1047;
add.s64 %rd100, %rd10, %rd99;
ld.global.nc.u32 %r474, [%rd100+4];

prmt.b32 %r470, %r470, 0, 0xba98;

xor.b32 %r1048, %r470, %r572;
or.b32 %r1049, %r1048, -2139062144;
and.b32 %r1050, %r470, 2139062143;
sub.s32 %r1051, %r1049, %r1050;
xor.b32 %r477, %r594, %r1051;
ld.global.nc.u32 %r478, [%rd100+8];

dp4a.s32.s32 %r472, %r473, %r474, %r524;


dp4a.s32.s32 %r476, %r477, %r478, %r472;


prmt.b32 %r480, %r480, 0, 0xba98;

xor.b32 %r1052, %r480, %r596;
or.b32 %r1053, %r1052, -2139062144;
and.b32 %r1054, %r480, 2139062143;
sub.s32 %r1055, %r1053, %r1054;
xor.b32 %r485, %r614, %r1055;
ld.global.nc.u32 %r486, [%rd100+12];

prmt.b32 %r482, %r482, 0, 0xba98;

xor.b32 %r1056, %r482, %r599;
or.b32 %r1057, %r1056, -2139062144;
and.b32 %r1058, %r482, 2139062143;
sub.s32 %r1059, %r1057, %r1058;
xor.b32 %r489, %r622, %r1059;
ld.global.nc.u32 %r490, [%rd100+16];

dp4a.s32.s32 %r484, %r485, %r486, %r476;


dp4a.s32.s32 %r488, %r489, %r490, %r484;


prmt.b32 %r492, %r492, 0, 0xba98;

xor.b32 %r1060, %r492, %r624;
or.b32 %r1061, %r1060, -2139062144;
and.b32 %r1062, %r492, 2139062143;
sub.s32 %r1063, %r1061, %r1062;
xor.b32 %r497, %r642, %r1063;
ld.global.nc.u32 %r498, [%rd100+20];

prmt.b32 %r494, %r494, 0, 0xba98;

xor.b32 %r1064, %r494, %r627;
or.b32 %r1065, %r1064, -2139062144;
and.b32 %r1066, %r494, 2139062143;
sub.s32 %r1067, %r1065, %r1066;
xor.b32 %r501, %r650, %r1067;
ld.global.nc.u32 %r502, [%rd100+24];

dp4a.s32.s32 %r496, %r497, %r498, %r488;


dp4a.s32.s32 %r500, %r501, %r502, %r496;


prmt.b32 %r504, %r504, 0, 0xba98;

xor.b32 %r1068, %r504, %r652;
or.b32 %r1069, %r1068, -2139062144;
and.b32 %r1070, %r504, 2139062143;
sub.s32 %r1071, %r1069, %r1070;
xor.b32 %r509, %r670, %r1071;
ld.global.nc.u32 %r510, [%rd100+28];

prmt.b32 %r506, %r506, 0, 0xba98;

xor.b32 %r1072, %r506, %r655;
or.b32 %r1073, %r1072, -2139062144;
and.b32 %r1074, %r506, 2139062143;
sub.s32 %r1075, %r1073, %r1074;
xor.b32 %r513, %r678, %r1075;
ld.global.nc.u32 %r514, [%rd100+32];

dp4a.s32.s32 %r508, %r509, %r510, %r500;


dp4a.s32.s32 %r512, %r513, %r514, %r508;

shr.u32 %r1076, %r512, 31;
add.s32 %r1077, %r512, %r1076;
shr.s32 %r1078, %r1077, 1;
mad.lo.s32 %r1079, %r512, %r680, %r1078;
shr.u32 %r1080, %r1079, 31;
add.s32 %r1081, %r1079, %r1080;
shr.s32 %r1082, %r1081, 1;

{ cvt.f32.f16 %f77, %rs1;}


ld.global.nc.u32 %r516, [%rd100];

{.reg .f16 low,high;
mov.b32 {low,high},%r516;
cvt.f32.f16 %f78, low;}


mul.ftz.f32 %f97, %f77, %f78;
cvt.rn.f32.s32 %f98, %r1082;
fma.rn.ftz.f32 %f222, %f97, %f98, %f222;

prmt.b32 %r517, %r517, 0, 0xba98;

xor.b32 %r1083, %r517, %r691;
or.b32 %r1084, %r1083, -2139062144;
and.b32 %r1085, %r517, 2139062143;
sub.s32 %r1086, %r1084, %r1085;
xor.b32 %r522, %r708, %r1086;

prmt.b32 %r519, %r519, 0, 0xba98;

xor.b32 %r1087, %r519, %r694;
or.b32 %r1088, %r1087, -2139062144;
and.b32 %r1089, %r519, 2139062143;
sub.s32 %r1090, %r1088, %r1089;
xor.b32 %r526, %r716, %r1090;

dp4a.s32.s32 %r521, %r522, %r474, %r524;


dp4a.s32.s32 %r525, %r526, %r478, %r521;


prmt.b32 %r529, %r529, 0, 0xba98;

xor.b32 %r1091, %r529, %r718;
or.b32 %r1092, %r1091, -2139062144;
and.b32 %r1093, %r529, 2139062143;
sub.s32 %r1094, %r1092, %r1093;
xor.b32 %r534, %r736, %r1094;

prmt.b32 %r531, %r531, 0, 0xba98;

xor.b32 %r1095, %r531, %r721;
or.b32 %r1096, %r1095, -2139062144;
and.b32 %r1097, %r531, 2139062143;
sub.s32 %r1098, %r1096, %r1097;
xor.b32 %r538, %r744, %r1098;

dp4a.s32.s32 %r533, %r534, %r486, %r525;


dp4a.s32.s32 %r537, %r538, %r490, %r533;


prmt.b32 %r541, %r541, 0, 0xba98;

xor.b32 %r1099, %r541, %r746;
or.b32 %r1100, %r1099, -2139062144;
and.b32 %r1101, %r541, 2139062143;
sub.s32 %r1102, %r1100, %r1101;
xor.b32 %r546, %r764, %r1102;

prmt.b32 %r543, %r543, 0, 0xba98;

xor.b32 %r1103, %r543, %r749;
or.b32 %r1104, %r1103, -2139062144;
and.b32 %r1105, %r543, 2139062143;
sub.s32 %r1106, %r1104, %r1105;
xor.b32 %r550, %r772, %r1106;

dp4a.s32.s32 %r545, %r546, %r498, %r537;


dp4a.s32.s32 %r549, %r550, %r502, %r545;


prmt.b32 %r553, %r553, 0, 0xba98;

xor.b32 %r1107, %r553, %r774;
or.b32 %r1108, %r1107, -2139062144;
and.b32 %r1109, %r553, 2139062143;
sub.s32 %r1110, %r1108, %r1109;
xor.b32 %r558, %r792, %r1110;

prmt.b32 %r555, %r555, 0, 0xba98;

xor.b32 %r1111, %r555, %r777;
or.b32 %r1112, %r1111, -2139062144;
and.b32 %r1113, %r555, 2139062143;
sub.s32 %r1114, %r1112, %r1113;
xor.b32 %r562, %r800, %r1114;

dp4a.s32.s32 %r557, %r558, %r510, %r549;


dp4a.s32.s32 %r561, %r562, %r514, %r557;

shr.u32 %r1115, %r561, 31;
add.s32 %r1116, %r561, %r1115;
shr.s32 %r1117, %r1116, 1;
mad.lo.s32 %r1118, %r561, %r802, %r1117;
shr.u32 %r1119, %r1118, 31;
add.s32 %r1120, %r1118, %r1119;
shr.s32 %r1121, %r1120, 1;

{ cvt.f32.f16 %f79, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r516;
cvt.f32.f16 %f80, low;}


mul.ftz.f32 %f99, %f79, %f80;
cvt.rn.f32.s32 %f100, %r1121;
fma.rn.ftz.f32 %f221, %f99, %f100, %f221;
add.s32 %r1314, %r1314, 64;
add.s32 %r1313, %r1313, 64;
add.s32 %r1312, %r1312, 64;
add.s32 %r1311, %r1311, 64;
add.s32 %r1310, %r1310, 64;
add.s64 %rd131, %rd131, 784;
add.s32 %r1315, %r1315, 8;
setp.lt.s32 %p4, %r1315, %r6;
@%p4 bra $L__BB246_2;

$L__BB246_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB246_5;

mov.u32 %r1122, _ZZ13mul_mat_vec_qIL9ggml_type18ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r1123, %r1, 1280, %r1122;
shl.b32 %r1124, %r2, 2;
add.s32 %r1125, %r1123, %r1124;
st.shared.f32 [%r1125+-1280], %f230;
st.shared.f32 [%r1125+-1152], %f229;
st.shared.f32 [%r1125+-1024], %f228;
st.shared.f32 [%r1125+-896], %f227;
st.shared.f32 [%r1125+-768], %f226;
st.shared.f32 [%r1125+-640], %f225;
st.shared.f32 [%r1125+-512], %f224;
st.shared.f32 [%r1125+-384], %f223;
st.shared.f32 [%r1125+-256], %f222;
st.shared.f32 [%r1125+-128], %f221;

$L__BB246_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB246_26;

mov.u32 %r1126, %ctaid.z;
mov.u32 %r1127, %ctaid.y;
mad.lo.s32 %r1128, %r1127, %r34, %r4;
mad.lo.s32 %r1129, %r1126, %r40, %r1128;
cvt.u64.u32 %rd13, %r1129;
shl.b32 %r1130, %r2, 2;
mov.u32 %r1131, _ZZ13mul_mat_vec_qIL9ggml_type18ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r24, %r1131, %r1130;
mov.u32 %r1132, 2;
ld.shared.f32 %f101, [%r24];
add.ftz.f32 %f102, %f101, %f230;
mov.b32 %r1133, %f102;
mov.u32 %r1134, 31;
mov.u32 %r1135, 16;
mov.u32 %r1136, -1;
shfl.sync.bfly.b32 %r1137|%p7, %r1133, %r1135, %r1134, %r1136;
mov.b32 %f103, %r1137;
add.ftz.f32 %f104, %f102, %f103;
mov.b32 %r1138, %f104;
mov.u32 %r1139, 8;
shfl.sync.bfly.b32 %r1140|%p8, %r1138, %r1139, %r1134, %r1136;
mov.b32 %f105, %r1140;
add.ftz.f32 %f106, %f104, %f105;
mov.b32 %r1141, %f106;
mov.u32 %r1142, 4;
shfl.sync.bfly.b32 %r1143|%p9, %r1141, %r1142, %r1134, %r1136;
mov.b32 %f107, %r1143;
add.ftz.f32 %f108, %f106, %f107;
mov.b32 %r1144, %f108;
shfl.sync.bfly.b32 %r1145|%p10, %r1144, %r1132, %r1134, %r1136;
mov.b32 %f109, %r1145;
add.ftz.f32 %f110, %f108, %f109;
mov.b32 %r1146, %f110;
mov.u32 %r1147, 1;
shfl.sync.bfly.b32 %r1148|%p11, %r1146, %r1147, %r1134, %r1136;
mov.b32 %f111, %r1148;
add.ftz.f32 %f31, %f110, %f111;
setp.lt.u32 %p12, %r4, %r28;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
cvta.to.global.u64 %rd101, %rd19;
mul.wide.u32 %rd102, %r1129, 4;
add.s64 %rd14, %rd101, %rd102;
not.pred %p14, %p1;
@%p14 bra $L__BB246_8;

st.global.f32 [%rd14], %f31;

$L__BB246_8:
ld.shared.f32 %f112, [%r24+128];
add.ftz.f32 %f113, %f112, %f229;
mov.b32 %r1149, %f113;
shfl.sync.bfly.b32 %r1153|%p15, %r1149, %r1135, %r1134, %r1136;
mov.b32 %f114, %r1153;
add.ftz.f32 %f115, %f113, %f114;
mov.b32 %r1154, %f115;
shfl.sync.bfly.b32 %r1156|%p16, %r1154, %r1139, %r1134, %r1136;
mov.b32 %f116, %r1156;
add.ftz.f32 %f117, %f115, %f116;
mov.b32 %r1157, %f117;
shfl.sync.bfly.b32 %r1159|%p17, %r1157, %r1142, %r1134, %r1136;
mov.b32 %f118, %r1159;
add.ftz.f32 %f119, %f117, %f118;
mov.b32 %r1160, %f119;
shfl.sync.bfly.b32 %r1162|%p18, %r1160, %r1132, %r1134, %r1136;
mov.b32 %f120, %r1162;
add.ftz.f32 %f121, %f119, %f120;
mov.b32 %r1163, %f121;
shfl.sync.bfly.b32 %r1165|%p19, %r1163, %r1147, %r1134, %r1136;
mov.b32 %f122, %r1165;
add.ftz.f32 %f32, %f121, %f122;
setp.eq.s32 %p20, %r2, 1;
or.b32 %r1169, %r4, 1;
setp.lt.u32 %p21, %r1169, %r28;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB246_10;

st.global.f32 [%rd14+4], %f32;

$L__BB246_10:
ld.shared.f32 %f123, [%r24+256];
add.ftz.f32 %f124, %f123, %f228;
mov.b32 %r1170, %f124;
mov.u32 %r1171, 31;
mov.u32 %r1172, 16;
mov.u32 %r1173, -1;
shfl.sync.bfly.b32 %r1174|%p23, %r1170, %r1172, %r1171, %r1173;
mov.b32 %f125, %r1174;
add.ftz.f32 %f126, %f124, %f125;
mov.b32 %r1175, %f126;
mov.u32 %r1176, 8;
shfl.sync.bfly.b32 %r1177|%p24, %r1175, %r1176, %r1171, %r1173;
mov.b32 %f127, %r1177;
add.ftz.f32 %f128, %f126, %f127;
mov.b32 %r1178, %f128;
mov.u32 %r1179, 4;
shfl.sync.bfly.b32 %r1180|%p25, %r1178, %r1179, %r1171, %r1173;
mov.b32 %f129, %r1180;
add.ftz.f32 %f130, %f128, %f129;
mov.b32 %r1181, %f130;
mov.u32 %r1182, 2;
shfl.sync.bfly.b32 %r1183|%p26, %r1181, %r1182, %r1171, %r1173;
mov.b32 %f131, %r1183;
add.ftz.f32 %f132, %f130, %f131;
mov.b32 %r1184, %f132;
mov.u32 %r1185, 1;
shfl.sync.bfly.b32 %r1186|%p27, %r1184, %r1185, %r1171, %r1173;
mov.b32 %f133, %r1186;
add.ftz.f32 %f33, %f132, %f133;
@%p14 bra $L__BB246_12;

cvt.u64.u32 %rd103, %r28;
add.s64 %rd104, %rd103, %rd13;
shl.b64 %rd106, %rd104, 2;
add.s64 %rd107, %rd101, %rd106;
st.global.f32 [%rd107], %f33;

$L__BB246_12:
ld.shared.f32 %f134, [%r24+384];
add.ftz.f32 %f135, %f134, %f227;
mov.b32 %r1187, %f135;
shfl.sync.bfly.b32 %r1191|%p29, %r1187, %r1172, %r1171, %r1173;
mov.b32 %f136, %r1191;
add.ftz.f32 %f137, %f135, %f136;
mov.b32 %r1192, %f137;
shfl.sync.bfly.b32 %r1194|%p30, %r1192, %r1176, %r1171, %r1173;
mov.b32 %f138, %r1194;
add.ftz.f32 %f139, %f137, %f138;
mov.b32 %r1195, %f139;
shfl.sync.bfly.b32 %r1197|%p31, %r1195, %r1179, %r1171, %r1173;
mov.b32 %f140, %r1197;
add.ftz.f32 %f141, %f139, %f140;
mov.b32 %r1198, %f141;
shfl.sync.bfly.b32 %r1200|%p32, %r1198, %r1182, %r1171, %r1173;
mov.b32 %f142, %r1200;
add.ftz.f32 %f143, %f141, %f142;
mov.b32 %r1201, %f143;
shfl.sync.bfly.b32 %r1203|%p33, %r1201, %r1185, %r1171, %r1173;
mov.b32 %f144, %r1203;
add.ftz.f32 %f34, %f143, %f144;
@%p22 bra $L__BB246_14;

add.s32 %r1204, %r28, 1;
cvt.u64.u32 %rd108, %r1204;
add.s64 %rd109, %rd108, %rd13;
shl.b64 %rd111, %rd109, 2;
add.s64 %rd112, %rd101, %rd111;
st.global.f32 [%rd112], %f34;

$L__BB246_14:
shl.b32 %r1205, %r28, 1;
mov.u32 %r1206, 1;
ld.shared.f32 %f145, [%r24+512];
add.ftz.f32 %f146, %f145, %f226;
mov.b32 %r1207, %f146;
mov.u32 %r1208, 31;
mov.u32 %r1209, 16;
mov.u32 %r1210, -1;
shfl.sync.bfly.b32 %r1211|%p35, %r1207, %r1209, %r1208, %r1210;
mov.b32 %f147, %r1211;
add.ftz.f32 %f148, %f146, %f147;
mov.b32 %r1212, %f148;
mov.u32 %r1213, 8;
shfl.sync.bfly.b32 %r1214|%p36, %r1212, %r1213, %r1208, %r1210;
mov.b32 %f149, %r1214;
add.ftz.f32 %f150, %f148, %f149;
mov.b32 %r1215, %f150;
mov.u32 %r1216, 4;
shfl.sync.bfly.b32 %r1217|%p37, %r1215, %r1216, %r1208, %r1210;
mov.b32 %f151, %r1217;
add.ftz.f32 %f152, %f150, %f151;
mov.b32 %r1218, %f152;
mov.u32 %r1219, 2;
shfl.sync.bfly.b32 %r1220|%p38, %r1218, %r1219, %r1208, %r1210;
mov.b32 %f153, %r1220;
add.ftz.f32 %f154, %f152, %f153;
mov.b32 %r1221, %f154;
shfl.sync.bfly.b32 %r1222|%p39, %r1221, %r1206, %r1208, %r1210;
mov.b32 %f155, %r1222;
add.ftz.f32 %f35, %f154, %f155;
cvt.u64.u32 %rd113, %r1205;
add.s64 %rd114, %rd113, %rd13;
shl.b64 %rd116, %rd114, 2;
add.s64 %rd15, %rd101, %rd116;
@%p14 bra $L__BB246_16;

st.global.f32 [%rd15], %f35;

$L__BB246_16:
ld.shared.f32 %f156, [%r24+640];
add.ftz.f32 %f157, %f156, %f225;
mov.b32 %r1223, %f157;
shfl.sync.bfly.b32 %r1227|%p41, %r1223, %r1209, %r1208, %r1210;
mov.b32 %f158, %r1227;
add.ftz.f32 %f159, %f157, %f158;
mov.b32 %r1228, %f159;
shfl.sync.bfly.b32 %r1230|%p42, %r1228, %r1213, %r1208, %r1210;
mov.b32 %f160, %r1230;
add.ftz.f32 %f161, %f159, %f160;
mov.b32 %r1231, %f161;
shfl.sync.bfly.b32 %r1233|%p43, %r1231, %r1216, %r1208, %r1210;
mov.b32 %f162, %r1233;
add.ftz.f32 %f163, %f161, %f162;
mov.b32 %r1234, %f163;
shfl.sync.bfly.b32 %r1236|%p44, %r1234, %r1219, %r1208, %r1210;
mov.b32 %f164, %r1236;
add.ftz.f32 %f165, %f163, %f164;
mov.b32 %r1237, %f165;
shfl.sync.bfly.b32 %r1239|%p45, %r1237, %r1206, %r1208, %r1210;
mov.b32 %f166, %r1239;
add.ftz.f32 %f36, %f165, %f166;
@%p22 bra $L__BB246_18;

st.global.f32 [%rd15+4], %f36;

$L__BB246_18:
mul.lo.s32 %r25, %r28, 3;
ld.shared.f32 %f167, [%r24+768];
add.ftz.f32 %f168, %f167, %f224;
mov.b32 %r1240, %f168;
mov.u32 %r1241, 31;
mov.u32 %r1242, 16;
mov.u32 %r1243, -1;
shfl.sync.bfly.b32 %r1244|%p47, %r1240, %r1242, %r1241, %r1243;
mov.b32 %f169, %r1244;
add.ftz.f32 %f170, %f168, %f169;
mov.b32 %r1245, %f170;
mov.u32 %r1246, 8;
shfl.sync.bfly.b32 %r1247|%p48, %r1245, %r1246, %r1241, %r1243;
mov.b32 %f171, %r1247;
add.ftz.f32 %f172, %f170, %f171;
mov.b32 %r1248, %f172;
mov.u32 %r1249, 4;
shfl.sync.bfly.b32 %r1250|%p49, %r1248, %r1249, %r1241, %r1243;
mov.b32 %f173, %r1250;
add.ftz.f32 %f174, %f172, %f173;
mov.b32 %r1251, %f174;
mov.u32 %r1252, 2;
shfl.sync.bfly.b32 %r1253|%p50, %r1251, %r1252, %r1241, %r1243;
mov.b32 %f175, %r1253;
add.ftz.f32 %f176, %f174, %f175;
mov.b32 %r1254, %f176;
mov.u32 %r1255, 1;
shfl.sync.bfly.b32 %r1256|%p51, %r1254, %r1255, %r1241, %r1243;
mov.b32 %f177, %r1256;
add.ftz.f32 %f37, %f176, %f177;
@%p14 bra $L__BB246_20;

cvt.u64.u32 %rd117, %r25;
add.s64 %rd118, %rd117, %rd13;
shl.b64 %rd120, %rd118, 2;
add.s64 %rd121, %rd101, %rd120;
st.global.f32 [%rd121], %f37;

$L__BB246_20:
ld.shared.f32 %f178, [%r24+896];
add.ftz.f32 %f179, %f178, %f223;
mov.b32 %r1257, %f179;
shfl.sync.bfly.b32 %r1261|%p53, %r1257, %r1242, %r1241, %r1243;
mov.b32 %f180, %r1261;
add.ftz.f32 %f181, %f179, %f180;
mov.b32 %r1262, %f181;
shfl.sync.bfly.b32 %r1264|%p54, %r1262, %r1246, %r1241, %r1243;
mov.b32 %f182, %r1264;
add.ftz.f32 %f183, %f181, %f182;
mov.b32 %r1265, %f183;
shfl.sync.bfly.b32 %r1267|%p55, %r1265, %r1249, %r1241, %r1243;
mov.b32 %f184, %r1267;
add.ftz.f32 %f185, %f183, %f184;
mov.b32 %r1268, %f185;
shfl.sync.bfly.b32 %r1270|%p56, %r1268, %r1252, %r1241, %r1243;
mov.b32 %f186, %r1270;
add.ftz.f32 %f187, %f185, %f186;
mov.b32 %r1271, %f187;
shfl.sync.bfly.b32 %r1273|%p57, %r1271, %r1255, %r1241, %r1243;
mov.b32 %f188, %r1273;
add.ftz.f32 %f38, %f187, %f188;
@%p22 bra $L__BB246_22;

add.s32 %r1274, %r25, 1;
cvt.u64.u32 %rd122, %r1274;
add.s64 %rd123, %rd122, %rd13;
shl.b64 %rd125, %rd123, 2;
add.s64 %rd126, %rd101, %rd125;
st.global.f32 [%rd126], %f38;

$L__BB246_22:
shl.b32 %r1275, %r28, 2;
mov.u32 %r1276, 2;
ld.shared.f32 %f189, [%r24+1024];
add.ftz.f32 %f190, %f189, %f222;
mov.b32 %r1277, %f190;
mov.u32 %r1278, 31;
mov.u32 %r1279, 16;
mov.u32 %r1280, -1;
shfl.sync.bfly.b32 %r1281|%p59, %r1277, %r1279, %r1278, %r1280;
mov.b32 %f191, %r1281;
add.ftz.f32 %f192, %f190, %f191;
mov.b32 %r1282, %f192;
mov.u32 %r1283, 8;
shfl.sync.bfly.b32 %r1284|%p60, %r1282, %r1283, %r1278, %r1280;
mov.b32 %f193, %r1284;
add.ftz.f32 %f194, %f192, %f193;
mov.b32 %r1285, %f194;
mov.u32 %r1286, 4;
shfl.sync.bfly.b32 %r1287|%p61, %r1285, %r1286, %r1278, %r1280;
mov.b32 %f195, %r1287;
add.ftz.f32 %f196, %f194, %f195;
mov.b32 %r1288, %f196;
shfl.sync.bfly.b32 %r1289|%p62, %r1288, %r1276, %r1278, %r1280;
mov.b32 %f197, %r1289;
add.ftz.f32 %f198, %f196, %f197;
mov.b32 %r1290, %f198;
mov.u32 %r1291, 1;
shfl.sync.bfly.b32 %r1292|%p63, %r1290, %r1291, %r1278, %r1280;
mov.b32 %f199, %r1292;
add.ftz.f32 %f39, %f198, %f199;
cvt.u64.u32 %rd127, %r1275;
add.s64 %rd128, %rd127, %rd13;
shl.b64 %rd130, %rd128, 2;
add.s64 %rd16, %rd101, %rd130;
@%p14 bra $L__BB246_24;

st.global.f32 [%rd16], %f39;

$L__BB246_24:
ld.shared.f32 %f200, [%r24+1152];
add.ftz.f32 %f201, %f200, %f221;
mov.b32 %r1293, %f201;
shfl.sync.bfly.b32 %r1297|%p65, %r1293, %r1279, %r1278, %r1280;
mov.b32 %f202, %r1297;
add.ftz.f32 %f203, %f201, %f202;
mov.b32 %r1298, %f203;
shfl.sync.bfly.b32 %r1300|%p66, %r1298, %r1283, %r1278, %r1280;
mov.b32 %f204, %r1300;
add.ftz.f32 %f205, %f203, %f204;
mov.b32 %r1301, %f205;
shfl.sync.bfly.b32 %r1303|%p67, %r1301, %r1286, %r1278, %r1280;
mov.b32 %f206, %r1303;
add.ftz.f32 %f207, %f205, %f206;
mov.b32 %r1304, %f207;
shfl.sync.bfly.b32 %r1306|%p68, %r1304, %r1276, %r1278, %r1280;
mov.b32 %f208, %r1306;
add.ftz.f32 %f209, %f207, %f208;
mov.b32 %r1307, %f209;
shfl.sync.bfly.b32 %r1309|%p69, %r1307, %r1291, %r1278, %r1280;
mov.b32 %f210, %r1309;
add.ftz.f32 %f40, %f209, %f210;
@%p22 bra $L__BB246_26;

st.global.f32 [%rd16+4], %f40;

$L__BB246_26:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type18ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type18ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type18ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type18ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type18ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type18ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type18ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type18ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type18ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<83>;
.reg .b16 %rs<41>;
.reg .f32 %f<277>;
.reg .b32 %r<1594>;
.reg .b64 %rd<150>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type18ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type18ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type18ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type18ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type18ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type18ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type18ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type18ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type18ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type18ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type18ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type18ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type18ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type18ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type18ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type18ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type18ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type18ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r42, %tid.y;
shl.b32 %r43, %r42, 5;
mov.u32 %r44, %tid.x;
add.s32 %r1, %r43, %r44;
shr.s32 %r45, %r1, 31;
shr.u32 %r46, %r45, 29;
add.s32 %r47, %r1, %r46;
shr.s32 %r1593, %r47, 3;
shr.u32 %r3, %r41, 8;
setp.ge.s32 %p3, %r1593, %r3;
mov.f32 %f265, 0f00000000;
mov.f32 %f266, %f265;
mov.f32 %f267, %f265;
mov.f32 %f268, %f265;
mov.f32 %f269, %f265;
mov.f32 %f270, %f265;
mov.f32 %f271, %f265;
mov.f32 %f272, %f265;
mov.f32 %f273, %f265;
mov.f32 %f274, %f265;
mov.f32 %f275, %f265;
mov.f32 %f276, %f265;
@%p3 bra $L__BB247_3;

mov.u32 %r48, %ctaid.x;
shl.b32 %r49, %r48, 1;
mov.u32 %r50, %ctaid.y;
mov.u32 %r51, %ctaid.z;
mul.lo.s32 %r52, %r50, %r33;
cvt.u64.u32 %rd19, %r52;
mul.lo.s32 %r53, %r51, %r39;
cvt.u64.u32 %rd20, %r53;
add.s64 %rd1, %rd20, %rd19;
mul.hi.u32 %r54, %r51, %r35;
add.s32 %r55, %r54, %r51;
shr.u32 %r56, %r55, %r36;
mul.lo.s32 %r57, %r56, %r38;
mul.hi.u32 %r58, %r50, %r29;
add.s32 %r59, %r58, %r50;
shr.u32 %r60, %r59, %r30;
mul.lo.s32 %r61, %r60, %r32;
cvta.to.global.u64 %rd149, %rd16;
and.b32 %r65, %r47, -8;
sub.s32 %r66, %r1, %r65;
shl.b32 %r67, %r66, 2;
shl.b32 %r68, %r66, 1;
add.s32 %r69, %r68, 33;
cvt.s64.s32 %rd3, %r66;
shl.b32 %r70, %r27, 1;
shl.b32 %r1587, %r1593, 3;
add.s32 %r1592, %r1587, %r70;
mad.lo.s32 %r1591, %r27, 3, %r1587;
shl.b32 %r71, %r27, 2;
add.s32 %r1590, %r1587, %r71;
mad.lo.s32 %r1589, %r27, 5, %r1587;
add.s32 %r1588, %r27, %r1587;
add.s32 %r72, %r1593, %r57;
add.s32 %r73, %r72, %r61;
or.b32 %r74, %r49, 1;
mad.lo.s32 %r75, %r26, %r74, %r73;
mul.wide.s32 %rd4, %r75, 98;
mad.lo.s32 %r76, %r49, %r26, %r1593;
add.s32 %r77, %r76, %r57;
add.s32 %r78, %r77, %r61;
mul.wide.s32 %rd5, %r78, 98;
mul.wide.s32 %rd21, %r67, 2;
add.s64 %rd22, %rd5, %rd21;
add.s64 %rd6, %rd22, 8;
mul.wide.s32 %rd23, %r69, 2;
add.s64 %rd24, %rd23, -2;
add.s64 %rd25, %rd24, %rd4;
add.s64 %rd7, %rd25, 2;
add.s64 %rd26, %rd24, %rd5;
add.s64 %rd8, %rd26, 2;
add.s64 %rd27, %rd4, %rd21;
add.s64 %rd9, %rd27, 4;
cvta.to.global.u64 %rd10, %rd17;
mov.f32 %f265, 0f00000000;

$L__BB247_2:
cvt.u64.u32 %rd28, %r1587;
add.s64 %rd29, %rd1, %rd28;
add.s64 %rd30, %rd29, %rd3;
mul.lo.s64 %rd31, %rd30, 36;
add.s64 %rd32, %rd149, %rd6;
ld.global.nc.u16 %rs13, [%rd32+-4];
ld.global.nc.u16 %rs14, [%rd32+-2];
ld.global.nc.u16 %rs15, [%rd32];
add.s64 %rd33, %rd149, %rd8;
ld.global.nc.u16 %rs16, [%rd33];
cvt.u32.u16 %r667, %rs16;
ld.global.nc.u16 %rs17, [%rd33+2];
cvt.u32.u16 %r668, %rs17;
mov.b32 %r669, {%rs16, %rs17};
ld.global.nc.u16 %rs18, [%rd32+-6];
shl.b16 %rs19, %rs18, 2;
cvt.u64.u16 %rd34, %rs19;
and.b64 %rd35, %rd34, 1020;
mov.u64 %rd36, iq3xxs_grid;
add.s64 %rd37, %rd36, %rd35;
ld.global.nc.u32 %r670, [%rd37];
not.b32 %r671, %r670;
shr.u16 %rs20, %rs18, 8;
cvt.u32.u16 %r672, %rs20;
mul.wide.u32 %rd38, %r672, 4;
add.s64 %rd39, %rd36, %rd38;
ld.global.nc.u32 %r673, [%rd39];
not.b32 %r674, %r673;
and.b32 %r675, %r667, 255;
popc.b32 %r676, %r675;
shl.b32 %r677, %r676, 7;
and.b32 %r678, %r677, 128;
xor.b32 %r679, %r678, %r675;
mul.lo.s32 %r680, %r679, 16843009;
and.b32 %r681, %r680, 134480385;
or.b32 %r682, %r681, -2139062144;
add.s32 %r683, %r682, -16843009;
or.b32 %r569, %r683, %r681;
mov.u32 %r79, %r569;

prmt.b32 %r79, %r79, 0, 0xba98;

xor.b32 %r684, %r79, %r670;
or.b32 %r685, %r684, -2139062144;
and.b32 %r686, %r79, 2139062143;
and.b32 %r687, %r671, -2139062144;
sub.s32 %r688, %r685, %r686;
xor.b32 %r84, %r687, %r688;
add.s64 %rd40, %rd10, %rd31;
ld.global.nc.u32 %r85, [%rd40+4];
and.b32 %r689, %r680, -2143281136;
or.b32 %r690, %r689, -2139062144;
add.s32 %r691, %r690, -16843009;
or.b32 %r571, %r691, %r689;
mov.u32 %r81, %r571;

prmt.b32 %r81, %r81, 0, 0xba98;

xor.b32 %r692, %r81, %r673;
or.b32 %r693, %r692, -2139062144;
and.b32 %r694, %r81, 2139062143;
and.b32 %r695, %r674, -2139062144;
sub.s32 %r696, %r693, %r694;
xor.b32 %r88, %r695, %r696;
ld.global.nc.u32 %r89, [%rd40+8];
mov.u32 %r625, 0;

dp4a.s32.s32 %r83, %r84, %r85, %r625;


dp4a.s32.s32 %r87, %r88, %r89, %r83;

shl.b16 %rs21, %rs13, 2;
cvt.u64.u16 %rd41, %rs21;
and.b64 %rd42, %rd41, 1020;
add.s64 %rd43, %rd36, %rd42;
ld.global.nc.u32 %r697, [%rd43];
not.b32 %r698, %r697;
shr.u16 %rs22, %rs13, 8;
cvt.u32.u16 %r699, %rs22;
mul.wide.u32 %rd44, %r699, 4;
add.s64 %rd45, %rd36, %rd44;
ld.global.nc.u32 %r700, [%rd45];
not.b32 %r701, %r700;
shr.u32 %r702, %r667, 7;
and.b32 %r703, %r702, 255;
popc.b32 %r704, %r703;
shl.b32 %r705, %r704, 7;
and.b32 %r706, %r705, 128;
xor.b32 %r707, %r706, %r703;
mul.lo.s32 %r708, %r707, 16843009;
and.b32 %r709, %r708, 134480385;
or.b32 %r710, %r709, -2139062144;
add.s32 %r711, %r710, -16843009;
or.b32 %r581, %r711, %r709;
mov.u32 %r91, %r581;

prmt.b32 %r91, %r91, 0, 0xba98;

xor.b32 %r712, %r91, %r697;
or.b32 %r713, %r712, -2139062144;
and.b32 %r714, %r91, 2139062143;
and.b32 %r715, %r698, -2139062144;
sub.s32 %r716, %r713, %r714;
xor.b32 %r96, %r715, %r716;
ld.global.nc.u32 %r97, [%rd40+12];
and.b32 %r717, %r708, -2143281136;
or.b32 %r718, %r717, -2139062144;
add.s32 %r719, %r718, -16843009;
or.b32 %r583, %r719, %r717;
mov.u32 %r93, %r583;

prmt.b32 %r93, %r93, 0, 0xba98;

xor.b32 %r720, %r93, %r700;
or.b32 %r721, %r720, -2139062144;
and.b32 %r722, %r93, 2139062143;
and.b32 %r723, %r701, -2139062144;
sub.s32 %r724, %r721, %r722;
xor.b32 %r100, %r723, %r724;
ld.global.nc.u32 %r101, [%rd40+16];

dp4a.s32.s32 %r95, %r96, %r97, %r87;


dp4a.s32.s32 %r99, %r100, %r101, %r95;

shl.b16 %rs23, %rs14, 2;
cvt.u64.u16 %rd46, %rs23;
and.b64 %rd47, %rd46, 1020;
add.s64 %rd48, %rd36, %rd47;
ld.global.nc.u32 %r725, [%rd48];
not.b32 %r726, %r725;
shr.u16 %rs24, %rs14, 8;
cvt.u32.u16 %r727, %rs24;
mul.wide.u32 %rd49, %r727, 4;
add.s64 %rd50, %rd36, %rd49;
ld.global.nc.u32 %r728, [%rd50];
not.b32 %r729, %r728;
shr.u32 %r730, %r669, 14;
and.b32 %r731, %r730, 255;
popc.b32 %r732, %r731;
shl.b32 %r733, %r732, 7;
and.b32 %r734, %r733, 128;
xor.b32 %r735, %r734, %r731;
mul.lo.s32 %r736, %r735, 16843009;
and.b32 %r737, %r736, 134480385;
or.b32 %r738, %r737, -2139062144;
add.s32 %r739, %r738, -16843009;
or.b32 %r593, %r739, %r737;
mov.u32 %r103, %r593;

prmt.b32 %r103, %r103, 0, 0xba98;

xor.b32 %r740, %r103, %r725;
or.b32 %r741, %r740, -2139062144;
and.b32 %r742, %r103, 2139062143;
and.b32 %r743, %r726, -2139062144;
sub.s32 %r744, %r741, %r742;
xor.b32 %r108, %r743, %r744;
ld.global.nc.u32 %r109, [%rd40+20];
and.b32 %r745, %r736, -2143281136;
or.b32 %r746, %r745, -2139062144;
add.s32 %r747, %r746, -16843009;
or.b32 %r595, %r747, %r745;
mov.u32 %r105, %r595;

prmt.b32 %r105, %r105, 0, 0xba98;

xor.b32 %r748, %r105, %r728;
or.b32 %r749, %r748, -2139062144;
and.b32 %r750, %r105, 2139062143;
and.b32 %r751, %r729, -2139062144;
sub.s32 %r752, %r749, %r750;
xor.b32 %r112, %r751, %r752;
ld.global.nc.u32 %r113, [%rd40+24];

dp4a.s32.s32 %r107, %r108, %r109, %r99;


dp4a.s32.s32 %r111, %r112, %r113, %r107;

shl.b16 %rs25, %rs15, 2;
cvt.u64.u16 %rd51, %rs25;
and.b64 %rd52, %rd51, 1020;
add.s64 %rd53, %rd36, %rd52;
ld.global.nc.u32 %r753, [%rd53];
not.b32 %r754, %r753;
shr.u16 %rs26, %rs15, 8;
cvt.u32.u16 %r755, %rs26;
mul.wide.u32 %rd54, %r755, 4;
add.s64 %rd55, %rd36, %rd54;
ld.global.nc.u32 %r756, [%rd55];
not.b32 %r757, %r756;
shr.u32 %r758, %r668, 5;
and.b32 %r759, %r758, 255;
popc.b32 %r760, %r759;
shl.b32 %r761, %r760, 7;
and.b32 %r762, %r761, 128;
xor.b32 %r763, %r762, %r759;
mul.lo.s32 %r764, %r763, 16843009;
and.b32 %r765, %r764, 134480385;
or.b32 %r766, %r765, -2139062144;
add.s32 %r767, %r766, -16843009;
or.b32 %r605, %r767, %r765;
mov.u32 %r115, %r605;

prmt.b32 %r115, %r115, 0, 0xba98;

xor.b32 %r768, %r115, %r753;
or.b32 %r769, %r768, -2139062144;
and.b32 %r770, %r115, 2139062143;
and.b32 %r771, %r754, -2139062144;
sub.s32 %r772, %r769, %r770;
xor.b32 %r120, %r771, %r772;
ld.global.nc.u32 %r121, [%rd40+28];
and.b32 %r773, %r764, -2143281136;
or.b32 %r774, %r773, -2139062144;
add.s32 %r775, %r774, -16843009;
or.b32 %r607, %r775, %r773;
mov.u32 %r117, %r607;

prmt.b32 %r117, %r117, 0, 0xba98;

xor.b32 %r776, %r117, %r756;
or.b32 %r777, %r776, -2139062144;
and.b32 %r778, %r117, 2139062143;
and.b32 %r779, %r757, -2139062144;
sub.s32 %r780, %r777, %r778;
xor.b32 %r124, %r779, %r780;
ld.global.nc.u32 %r125, [%rd40+32];

dp4a.s32.s32 %r119, %r120, %r121, %r111;

shr.u32 %r781, %r668, 12;

dp4a.s32.s32 %r123, %r124, %r125, %r119;

shr.u32 %r782, %r123, 31;
add.s32 %r783, %r123, %r782;
shr.s32 %r784, %r783, 1;
mad.lo.s32 %r785, %r123, %r781, %r784;
shr.u32 %r786, %r785, 31;
add.s32 %r787, %r785, %r786;
shr.s32 %r788, %r787, 1;
add.s64 %rd56, %rd149, %rd5;
ld.global.nc.u16 %rs1, [%rd56];

{ cvt.f32.f16 %f73, %rs1;}


ld.global.nc.u32 %r127, [%rd40];

{.reg .f16 low,high;
mov.b32 {low,high},%r127;
cvt.f32.f16 %f74, low;}


mul.ftz.f32 %f97, %f73, %f74;
cvt.rn.f32.s32 %f98, %r788;
fma.rn.ftz.f32 %f276, %f97, %f98, %f276;
add.s64 %rd57, %rd149, %rd9;
ld.global.nc.u16 %rs27, [%rd57];
ld.global.nc.u16 %rs28, [%rd57+2];
ld.global.nc.u16 %rs29, [%rd57+4];
add.s64 %rd58, %rd149, %rd7;
ld.global.nc.u16 %rs30, [%rd58];
cvt.u32.u16 %r789, %rs30;
ld.global.nc.u16 %rs31, [%rd58+2];
cvt.u32.u16 %r790, %rs31;
mov.b32 %r791, {%rs30, %rs31};
ld.global.nc.u16 %rs32, [%rd57+-2];
shl.b16 %rs33, %rs32, 2;
cvt.u64.u16 %rd59, %rs33;
and.b64 %rd60, %rd59, 1020;
add.s64 %rd61, %rd36, %rd60;
ld.global.nc.u32 %r792, [%rd61];
not.b32 %r793, %r792;
shr.u16 %rs34, %rs32, 8;
cvt.u32.u16 %r794, %rs34;
mul.wide.u32 %rd62, %r794, 4;
add.s64 %rd63, %rd36, %rd62;
ld.global.nc.u32 %r795, [%rd63];
not.b32 %r796, %r795;
and.b32 %r797, %r789, 255;
popc.b32 %r798, %r797;
shl.b32 %r799, %r798, 7;
and.b32 %r800, %r799, 128;
xor.b32 %r801, %r800, %r797;
mul.lo.s32 %r802, %r801, 16843009;
and.b32 %r803, %r802, 134480385;
or.b32 %r804, %r803, -2139062144;
add.s32 %r805, %r804, -16843009;
or.b32 %r618, %r805, %r803;
mov.u32 %r128, %r618;

prmt.b32 %r128, %r128, 0, 0xba98;

xor.b32 %r806, %r128, %r792;
or.b32 %r807, %r806, -2139062144;
and.b32 %r808, %r128, 2139062143;
and.b32 %r809, %r793, -2139062144;
sub.s32 %r810, %r807, %r808;
xor.b32 %r133, %r809, %r810;
and.b32 %r811, %r802, -2143281136;
or.b32 %r812, %r811, -2139062144;
add.s32 %r813, %r812, -16843009;
or.b32 %r620, %r813, %r811;
mov.u32 %r130, %r620;

prmt.b32 %r130, %r130, 0, 0xba98;

xor.b32 %r814, %r130, %r795;
or.b32 %r815, %r814, -2139062144;
and.b32 %r816, %r130, 2139062143;
and.b32 %r817, %r796, -2139062144;
sub.s32 %r818, %r815, %r816;
xor.b32 %r137, %r817, %r818;

dp4a.s32.s32 %r132, %r133, %r85, %r625;


dp4a.s32.s32 %r136, %r137, %r89, %r132;

shl.b16 %rs35, %rs27, 2;
cvt.u64.u16 %rd64, %rs35;
and.b64 %rd65, %rd64, 1020;
add.s64 %rd66, %rd36, %rd65;
ld.global.nc.u32 %r819, [%rd66];
not.b32 %r820, %r819;
shr.u16 %rs36, %rs27, 8;
cvt.u32.u16 %r821, %rs36;
mul.wide.u32 %rd67, %r821, 4;
add.s64 %rd68, %rd36, %rd67;
ld.global.nc.u32 %r822, [%rd68];
not.b32 %r823, %r822;
shr.u32 %r824, %r789, 7;
and.b32 %r825, %r824, 255;
popc.b32 %r826, %r825;
shl.b32 %r827, %r826, 7;
and.b32 %r828, %r827, 128;
xor.b32 %r829, %r828, %r825;
mul.lo.s32 %r830, %r829, 16843009;
and.b32 %r831, %r830, 134480385;
or.b32 %r832, %r831, -2139062144;
add.s32 %r833, %r832, -16843009;
or.b32 %r630, %r833, %r831;
mov.u32 %r140, %r630;

prmt.b32 %r140, %r140, 0, 0xba98;

xor.b32 %r834, %r140, %r819;
or.b32 %r835, %r834, -2139062144;
and.b32 %r836, %r140, 2139062143;
and.b32 %r837, %r820, -2139062144;
sub.s32 %r838, %r835, %r836;
xor.b32 %r145, %r837, %r838;
and.b32 %r839, %r830, -2143281136;
or.b32 %r840, %r839, -2139062144;
add.s32 %r841, %r840, -16843009;
or.b32 %r632, %r841, %r839;
mov.u32 %r142, %r632;

prmt.b32 %r142, %r142, 0, 0xba98;

xor.b32 %r842, %r142, %r822;
or.b32 %r843, %r842, -2139062144;
and.b32 %r844, %r142, 2139062143;
and.b32 %r845, %r823, -2139062144;
sub.s32 %r846, %r843, %r844;
xor.b32 %r149, %r845, %r846;

dp4a.s32.s32 %r144, %r145, %r97, %r136;


dp4a.s32.s32 %r148, %r149, %r101, %r144;

shl.b16 %rs37, %rs28, 2;
cvt.u64.u16 %rd69, %rs37;
and.b64 %rd70, %rd69, 1020;
add.s64 %rd71, %rd36, %rd70;
ld.global.nc.u32 %r847, [%rd71];
not.b32 %r848, %r847;
shr.u16 %rs38, %rs28, 8;
cvt.u32.u16 %r849, %rs38;
mul.wide.u32 %rd72, %r849, 4;
add.s64 %rd73, %rd36, %rd72;
ld.global.nc.u32 %r850, [%rd73];
not.b32 %r851, %r850;
shr.u32 %r852, %r791, 14;
and.b32 %r853, %r852, 255;
popc.b32 %r854, %r853;
shl.b32 %r855, %r854, 7;
and.b32 %r856, %r855, 128;
xor.b32 %r857, %r856, %r853;
mul.lo.s32 %r858, %r857, 16843009;
and.b32 %r859, %r858, 134480385;
or.b32 %r860, %r859, -2139062144;
add.s32 %r861, %r860, -16843009;
or.b32 %r642, %r861, %r859;
mov.u32 %r152, %r642;

prmt.b32 %r152, %r152, 0, 0xba98;

xor.b32 %r862, %r152, %r847;
or.b32 %r863, %r862, -2139062144;
and.b32 %r864, %r152, 2139062143;
and.b32 %r865, %r848, -2139062144;
sub.s32 %r866, %r863, %r864;
xor.b32 %r157, %r865, %r866;
and.b32 %r867, %r858, -2143281136;
or.b32 %r868, %r867, -2139062144;
add.s32 %r869, %r868, -16843009;
or.b32 %r644, %r869, %r867;
mov.u32 %r154, %r644;

prmt.b32 %r154, %r154, 0, 0xba98;

xor.b32 %r870, %r154, %r850;
or.b32 %r871, %r870, -2139062144;
and.b32 %r872, %r154, 2139062143;
and.b32 %r873, %r851, -2139062144;
sub.s32 %r874, %r871, %r872;
xor.b32 %r161, %r873, %r874;

dp4a.s32.s32 %r156, %r157, %r109, %r148;


dp4a.s32.s32 %r160, %r161, %r113, %r156;

shl.b16 %rs39, %rs29, 2;
cvt.u64.u16 %rd74, %rs39;
and.b64 %rd75, %rd74, 1020;
add.s64 %rd76, %rd36, %rd75;
ld.global.nc.u32 %r875, [%rd76];
not.b32 %r876, %r875;
shr.u16 %rs40, %rs29, 8;
cvt.u32.u16 %r877, %rs40;
mul.wide.u32 %rd77, %r877, 4;
add.s64 %rd78, %rd36, %rd77;
ld.global.nc.u32 %r878, [%rd78];
not.b32 %r879, %r878;
shr.u32 %r880, %r790, 5;
and.b32 %r881, %r880, 255;
popc.b32 %r882, %r881;
shl.b32 %r883, %r882, 7;
and.b32 %r884, %r883, 128;
xor.b32 %r885, %r884, %r881;
mul.lo.s32 %r886, %r885, 16843009;
and.b32 %r887, %r886, 134480385;
or.b32 %r888, %r887, -2139062144;
add.s32 %r889, %r888, -16843009;
or.b32 %r654, %r889, %r887;
mov.u32 %r164, %r654;

prmt.b32 %r164, %r164, 0, 0xba98;

xor.b32 %r890, %r164, %r875;
or.b32 %r891, %r890, -2139062144;
and.b32 %r892, %r164, 2139062143;
and.b32 %r893, %r876, -2139062144;
sub.s32 %r894, %r891, %r892;
xor.b32 %r169, %r893, %r894;
and.b32 %r895, %r886, -2143281136;
or.b32 %r896, %r895, -2139062144;
add.s32 %r897, %r896, -16843009;
or.b32 %r656, %r897, %r895;
mov.u32 %r166, %r656;

prmt.b32 %r166, %r166, 0, 0xba98;

xor.b32 %r898, %r166, %r878;
or.b32 %r899, %r898, -2139062144;
and.b32 %r900, %r166, 2139062143;
and.b32 %r901, %r879, -2139062144;
sub.s32 %r902, %r899, %r900;
xor.b32 %r173, %r901, %r902;

dp4a.s32.s32 %r168, %r169, %r121, %r160;

shr.u32 %r903, %r790, 12;

dp4a.s32.s32 %r172, %r173, %r125, %r168;

shr.u32 %r904, %r172, 31;
add.s32 %r905, %r172, %r904;
shr.s32 %r906, %r905, 1;
mad.lo.s32 %r907, %r172, %r903, %r906;
shr.u32 %r908, %r907, 31;
add.s32 %r909, %r907, %r908;
shr.s32 %r910, %r909, 1;
add.s64 %rd79, %rd149, %rd4;
ld.global.nc.u16 %rs2, [%rd79];

{ cvt.f32.f16 %f75, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r127;
cvt.f32.f16 %f76, low;}


mul.ftz.f32 %f99, %f75, %f76;
cvt.rn.f32.s32 %f100, %r910;
fma.rn.ftz.f32 %f275, %f99, %f100, %f275;
cvt.u64.u32 %rd80, %r1588;
add.s64 %rd81, %rd1, %rd80;
add.s64 %rd82, %rd81, %rd3;
mul.lo.s64 %rd83, %rd82, 36;
mov.u32 %r177, %r569;

prmt.b32 %r177, %r177, 0, 0xba98;

xor.b32 %r911, %r177, %r670;
or.b32 %r912, %r911, -2139062144;
and.b32 %r913, %r177, 2139062143;
sub.s32 %r914, %r912, %r913;
xor.b32 %r182, %r687, %r914;
add.s64 %rd84, %rd10, %rd83;
ld.global.nc.u32 %r183, [%rd84+4];
mov.u32 %r179, %r571;

prmt.b32 %r179, %r179, 0, 0xba98;

xor.b32 %r915, %r179, %r673;
or.b32 %r916, %r915, -2139062144;
and.b32 %r917, %r179, 2139062143;
sub.s32 %r918, %r916, %r917;
xor.b32 %r186, %r695, %r918;
ld.global.nc.u32 %r187, [%rd84+8];

dp4a.s32.s32 %r181, %r182, %r183, %r625;


dp4a.s32.s32 %r185, %r186, %r187, %r181;

mov.u32 %r189, %r581;

prmt.b32 %r189, %r189, 0, 0xba98;

xor.b32 %r919, %r189, %r697;
or.b32 %r920, %r919, -2139062144;
and.b32 %r921, %r189, 2139062143;
sub.s32 %r922, %r920, %r921;
xor.b32 %r194, %r715, %r922;
ld.global.nc.u32 %r195, [%rd84+12];
mov.u32 %r191, %r583;

prmt.b32 %r191, %r191, 0, 0xba98;

xor.b32 %r923, %r191, %r700;
or.b32 %r924, %r923, -2139062144;
and.b32 %r925, %r191, 2139062143;
sub.s32 %r926, %r924, %r925;
xor.b32 %r198, %r723, %r926;
ld.global.nc.u32 %r199, [%rd84+16];

dp4a.s32.s32 %r193, %r194, %r195, %r185;


dp4a.s32.s32 %r197, %r198, %r199, %r193;

mov.u32 %r201, %r593;

prmt.b32 %r201, %r201, 0, 0xba98;

xor.b32 %r927, %r201, %r725;
or.b32 %r928, %r927, -2139062144;
and.b32 %r929, %r201, 2139062143;
sub.s32 %r930, %r928, %r929;
xor.b32 %r206, %r743, %r930;
ld.global.nc.u32 %r207, [%rd84+20];
mov.u32 %r203, %r595;

prmt.b32 %r203, %r203, 0, 0xba98;

xor.b32 %r931, %r203, %r728;
or.b32 %r932, %r931, -2139062144;
and.b32 %r933, %r203, 2139062143;
sub.s32 %r934, %r932, %r933;
xor.b32 %r210, %r751, %r934;
ld.global.nc.u32 %r211, [%rd84+24];

dp4a.s32.s32 %r205, %r206, %r207, %r197;


dp4a.s32.s32 %r209, %r210, %r211, %r205;

mov.u32 %r213, %r605;

prmt.b32 %r213, %r213, 0, 0xba98;

xor.b32 %r935, %r213, %r753;
or.b32 %r936, %r935, -2139062144;
and.b32 %r937, %r213, 2139062143;
sub.s32 %r938, %r936, %r937;
xor.b32 %r218, %r771, %r938;
ld.global.nc.u32 %r219, [%rd84+28];
mov.u32 %r215, %r607;

prmt.b32 %r215, %r215, 0, 0xba98;

xor.b32 %r939, %r215, %r756;
or.b32 %r940, %r939, -2139062144;
and.b32 %r941, %r215, 2139062143;
sub.s32 %r942, %r940, %r941;
xor.b32 %r222, %r779, %r942;
ld.global.nc.u32 %r223, [%rd84+32];

dp4a.s32.s32 %r217, %r218, %r219, %r209;


dp4a.s32.s32 %r221, %r222, %r223, %r217;

shr.u32 %r943, %r221, 31;
add.s32 %r944, %r221, %r943;
shr.s32 %r945, %r944, 1;
mad.lo.s32 %r946, %r221, %r781, %r945;
shr.u32 %r947, %r946, 31;
add.s32 %r948, %r946, %r947;
shr.s32 %r949, %r948, 1;

{ cvt.f32.f16 %f77, %rs1;}


ld.global.nc.u32 %r225, [%rd84];

{.reg .f16 low,high;
mov.b32 {low,high},%r225;
cvt.f32.f16 %f78, low;}


mul.ftz.f32 %f101, %f77, %f78;
cvt.rn.f32.s32 %f102, %r949;
fma.rn.ftz.f32 %f274, %f101, %f102, %f274;
mov.u32 %r226, %r618;

prmt.b32 %r226, %r226, 0, 0xba98;

xor.b32 %r950, %r226, %r792;
or.b32 %r951, %r950, -2139062144;
and.b32 %r952, %r226, 2139062143;
sub.s32 %r953, %r951, %r952;
xor.b32 %r231, %r809, %r953;
mov.u32 %r228, %r620;

prmt.b32 %r228, %r228, 0, 0xba98;

xor.b32 %r954, %r228, %r795;
or.b32 %r955, %r954, -2139062144;
and.b32 %r956, %r228, 2139062143;
sub.s32 %r957, %r955, %r956;
xor.b32 %r235, %r817, %r957;

dp4a.s32.s32 %r230, %r231, %r183, %r625;


dp4a.s32.s32 %r234, %r235, %r187, %r230;

mov.u32 %r238, %r630;

prmt.b32 %r238, %r238, 0, 0xba98;

xor.b32 %r958, %r238, %r819;
or.b32 %r959, %r958, -2139062144;
and.b32 %r960, %r238, 2139062143;
sub.s32 %r961, %r959, %r960;
xor.b32 %r243, %r837, %r961;
mov.u32 %r240, %r632;

prmt.b32 %r240, %r240, 0, 0xba98;

xor.b32 %r962, %r240, %r822;
or.b32 %r963, %r962, -2139062144;
and.b32 %r964, %r240, 2139062143;
sub.s32 %r965, %r963, %r964;
xor.b32 %r247, %r845, %r965;

dp4a.s32.s32 %r242, %r243, %r195, %r234;


dp4a.s32.s32 %r246, %r247, %r199, %r242;

mov.u32 %r250, %r642;

prmt.b32 %r250, %r250, 0, 0xba98;

xor.b32 %r966, %r250, %r847;
or.b32 %r967, %r966, -2139062144;
and.b32 %r968, %r250, 2139062143;
sub.s32 %r969, %r967, %r968;
xor.b32 %r255, %r865, %r969;
mov.u32 %r252, %r644;

prmt.b32 %r252, %r252, 0, 0xba98;

xor.b32 %r970, %r252, %r850;
or.b32 %r971, %r970, -2139062144;
and.b32 %r972, %r252, 2139062143;
sub.s32 %r973, %r971, %r972;
xor.b32 %r259, %r873, %r973;

dp4a.s32.s32 %r254, %r255, %r207, %r246;


dp4a.s32.s32 %r258, %r259, %r211, %r254;

mov.u32 %r262, %r654;

prmt.b32 %r262, %r262, 0, 0xba98;

xor.b32 %r974, %r262, %r875;
or.b32 %r975, %r974, -2139062144;
and.b32 %r976, %r262, 2139062143;
sub.s32 %r977, %r975, %r976;
xor.b32 %r267, %r893, %r977;
mov.u32 %r264, %r656;

prmt.b32 %r264, %r264, 0, 0xba98;

xor.b32 %r978, %r264, %r878;
or.b32 %r979, %r978, -2139062144;
and.b32 %r980, %r264, 2139062143;
sub.s32 %r981, %r979, %r980;
xor.b32 %r271, %r901, %r981;

dp4a.s32.s32 %r266, %r267, %r219, %r258;


dp4a.s32.s32 %r270, %r271, %r223, %r266;

shr.u32 %r982, %r270, 31;
add.s32 %r983, %r270, %r982;
shr.s32 %r984, %r983, 1;
mad.lo.s32 %r985, %r270, %r903, %r984;
shr.u32 %r986, %r985, 31;
add.s32 %r987, %r985, %r986;
shr.s32 %r988, %r987, 1;

{ cvt.f32.f16 %f79, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r225;
cvt.f32.f16 %f80, low;}


mul.ftz.f32 %f103, %f79, %f80;
cvt.rn.f32.s32 %f104, %r988;
fma.rn.ftz.f32 %f273, %f103, %f104, %f273;
cvt.u64.u32 %rd85, %r1592;
add.s64 %rd86, %rd1, %rd85;
add.s64 %rd87, %rd86, %rd3;
mul.lo.s64 %rd88, %rd87, 36;
mov.u32 %r275, %r569;

prmt.b32 %r275, %r275, 0, 0xba98;

xor.b32 %r989, %r275, %r670;
or.b32 %r990, %r989, -2139062144;
and.b32 %r991, %r275, 2139062143;
sub.s32 %r992, %r990, %r991;
xor.b32 %r280, %r687, %r992;
add.s64 %rd89, %rd10, %rd88;
ld.global.nc.u32 %r281, [%rd89+4];
mov.u32 %r277, %r571;

prmt.b32 %r277, %r277, 0, 0xba98;

xor.b32 %r993, %r277, %r673;
or.b32 %r994, %r993, -2139062144;
and.b32 %r995, %r277, 2139062143;
sub.s32 %r996, %r994, %r995;
xor.b32 %r284, %r695, %r996;
ld.global.nc.u32 %r285, [%rd89+8];

dp4a.s32.s32 %r279, %r280, %r281, %r625;


dp4a.s32.s32 %r283, %r284, %r285, %r279;

mov.u32 %r287, %r581;

prmt.b32 %r287, %r287, 0, 0xba98;

xor.b32 %r997, %r287, %r697;
or.b32 %r998, %r997, -2139062144;
and.b32 %r999, %r287, 2139062143;
sub.s32 %r1000, %r998, %r999;
xor.b32 %r292, %r715, %r1000;
ld.global.nc.u32 %r293, [%rd89+12];
mov.u32 %r289, %r583;

prmt.b32 %r289, %r289, 0, 0xba98;

xor.b32 %r1001, %r289, %r700;
or.b32 %r1002, %r1001, -2139062144;
and.b32 %r1003, %r289, 2139062143;
sub.s32 %r1004, %r1002, %r1003;
xor.b32 %r296, %r723, %r1004;
ld.global.nc.u32 %r297, [%rd89+16];

dp4a.s32.s32 %r291, %r292, %r293, %r283;


dp4a.s32.s32 %r295, %r296, %r297, %r291;

mov.u32 %r299, %r593;

prmt.b32 %r299, %r299, 0, 0xba98;

xor.b32 %r1005, %r299, %r725;
or.b32 %r1006, %r1005, -2139062144;
and.b32 %r1007, %r299, 2139062143;
sub.s32 %r1008, %r1006, %r1007;
xor.b32 %r304, %r743, %r1008;
ld.global.nc.u32 %r305, [%rd89+20];
mov.u32 %r301, %r595;

prmt.b32 %r301, %r301, 0, 0xba98;

xor.b32 %r1009, %r301, %r728;
or.b32 %r1010, %r1009, -2139062144;
and.b32 %r1011, %r301, 2139062143;
sub.s32 %r1012, %r1010, %r1011;
xor.b32 %r308, %r751, %r1012;
ld.global.nc.u32 %r309, [%rd89+24];

dp4a.s32.s32 %r303, %r304, %r305, %r295;


dp4a.s32.s32 %r307, %r308, %r309, %r303;

mov.u32 %r311, %r605;

prmt.b32 %r311, %r311, 0, 0xba98;

xor.b32 %r1013, %r311, %r753;
or.b32 %r1014, %r1013, -2139062144;
and.b32 %r1015, %r311, 2139062143;
sub.s32 %r1016, %r1014, %r1015;
xor.b32 %r316, %r771, %r1016;
ld.global.nc.u32 %r317, [%rd89+28];
mov.u32 %r313, %r607;

prmt.b32 %r313, %r313, 0, 0xba98;

xor.b32 %r1017, %r313, %r756;
or.b32 %r1018, %r1017, -2139062144;
and.b32 %r1019, %r313, 2139062143;
sub.s32 %r1020, %r1018, %r1019;
xor.b32 %r320, %r779, %r1020;
ld.global.nc.u32 %r321, [%rd89+32];

dp4a.s32.s32 %r315, %r316, %r317, %r307;


dp4a.s32.s32 %r319, %r320, %r321, %r315;

shr.u32 %r1021, %r319, 31;
add.s32 %r1022, %r319, %r1021;
shr.s32 %r1023, %r1022, 1;
mad.lo.s32 %r1024, %r319, %r781, %r1023;
shr.u32 %r1025, %r1024, 31;
add.s32 %r1026, %r1024, %r1025;
shr.s32 %r1027, %r1026, 1;

{ cvt.f32.f16 %f81, %rs1;}


ld.global.nc.u32 %r323, [%rd89];

{.reg .f16 low,high;
mov.b32 {low,high},%r323;
cvt.f32.f16 %f82, low;}


mul.ftz.f32 %f105, %f81, %f82;
cvt.rn.f32.s32 %f106, %r1027;
fma.rn.ftz.f32 %f272, %f105, %f106, %f272;
mov.u32 %r324, %r618;

prmt.b32 %r324, %r324, 0, 0xba98;

xor.b32 %r1028, %r324, %r792;
or.b32 %r1029, %r1028, -2139062144;
and.b32 %r1030, %r324, 2139062143;
sub.s32 %r1031, %r1029, %r1030;
xor.b32 %r329, %r809, %r1031;
mov.u32 %r326, %r620;

prmt.b32 %r326, %r326, 0, 0xba98;

xor.b32 %r1032, %r326, %r795;
or.b32 %r1033, %r1032, -2139062144;
and.b32 %r1034, %r326, 2139062143;
sub.s32 %r1035, %r1033, %r1034;
xor.b32 %r333, %r817, %r1035;

dp4a.s32.s32 %r328, %r329, %r281, %r625;


dp4a.s32.s32 %r332, %r333, %r285, %r328;

mov.u32 %r336, %r630;

prmt.b32 %r336, %r336, 0, 0xba98;

xor.b32 %r1036, %r336, %r819;
or.b32 %r1037, %r1036, -2139062144;
and.b32 %r1038, %r336, 2139062143;
sub.s32 %r1039, %r1037, %r1038;
xor.b32 %r341, %r837, %r1039;
mov.u32 %r338, %r632;

prmt.b32 %r338, %r338, 0, 0xba98;

xor.b32 %r1040, %r338, %r822;
or.b32 %r1041, %r1040, -2139062144;
and.b32 %r1042, %r338, 2139062143;
sub.s32 %r1043, %r1041, %r1042;
xor.b32 %r345, %r845, %r1043;

dp4a.s32.s32 %r340, %r341, %r293, %r332;


dp4a.s32.s32 %r344, %r345, %r297, %r340;

mov.u32 %r348, %r642;

prmt.b32 %r348, %r348, 0, 0xba98;

xor.b32 %r1044, %r348, %r847;
or.b32 %r1045, %r1044, -2139062144;
and.b32 %r1046, %r348, 2139062143;
sub.s32 %r1047, %r1045, %r1046;
xor.b32 %r353, %r865, %r1047;
mov.u32 %r350, %r644;

prmt.b32 %r350, %r350, 0, 0xba98;

xor.b32 %r1048, %r350, %r850;
or.b32 %r1049, %r1048, -2139062144;
and.b32 %r1050, %r350, 2139062143;
sub.s32 %r1051, %r1049, %r1050;
xor.b32 %r357, %r873, %r1051;

dp4a.s32.s32 %r352, %r353, %r305, %r344;


dp4a.s32.s32 %r356, %r357, %r309, %r352;

mov.u32 %r360, %r654;

prmt.b32 %r360, %r360, 0, 0xba98;

xor.b32 %r1052, %r360, %r875;
or.b32 %r1053, %r1052, -2139062144;
and.b32 %r1054, %r360, 2139062143;
sub.s32 %r1055, %r1053, %r1054;
xor.b32 %r365, %r893, %r1055;
mov.u32 %r362, %r656;

prmt.b32 %r362, %r362, 0, 0xba98;

xor.b32 %r1056, %r362, %r878;
or.b32 %r1057, %r1056, -2139062144;
and.b32 %r1058, %r362, 2139062143;
sub.s32 %r1059, %r1057, %r1058;
xor.b32 %r369, %r901, %r1059;

dp4a.s32.s32 %r364, %r365, %r317, %r356;


dp4a.s32.s32 %r368, %r369, %r321, %r364;

shr.u32 %r1060, %r368, 31;
add.s32 %r1061, %r368, %r1060;
shr.s32 %r1062, %r1061, 1;
mad.lo.s32 %r1063, %r368, %r903, %r1062;
shr.u32 %r1064, %r1063, 31;
add.s32 %r1065, %r1063, %r1064;
shr.s32 %r1066, %r1065, 1;

{ cvt.f32.f16 %f83, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r323;
cvt.f32.f16 %f84, low;}


mul.ftz.f32 %f107, %f83, %f84;
cvt.rn.f32.s32 %f108, %r1066;
fma.rn.ftz.f32 %f271, %f107, %f108, %f271;
cvt.u64.u32 %rd90, %r1591;
add.s64 %rd91, %rd1, %rd90;
add.s64 %rd92, %rd91, %rd3;
mul.lo.s64 %rd93, %rd92, 36;
mov.u32 %r373, %r569;

prmt.b32 %r373, %r373, 0, 0xba98;

xor.b32 %r1067, %r373, %r670;
or.b32 %r1068, %r1067, -2139062144;
and.b32 %r1069, %r373, 2139062143;
sub.s32 %r1070, %r1068, %r1069;
xor.b32 %r378, %r687, %r1070;
add.s64 %rd94, %rd10, %rd93;
ld.global.nc.u32 %r379, [%rd94+4];
mov.u32 %r375, %r571;

prmt.b32 %r375, %r375, 0, 0xba98;

xor.b32 %r1071, %r375, %r673;
or.b32 %r1072, %r1071, -2139062144;
and.b32 %r1073, %r375, 2139062143;
sub.s32 %r1074, %r1072, %r1073;
xor.b32 %r382, %r695, %r1074;
ld.global.nc.u32 %r383, [%rd94+8];

dp4a.s32.s32 %r377, %r378, %r379, %r625;


dp4a.s32.s32 %r381, %r382, %r383, %r377;

mov.u32 %r385, %r581;

prmt.b32 %r385, %r385, 0, 0xba98;

xor.b32 %r1075, %r385, %r697;
or.b32 %r1076, %r1075, -2139062144;
and.b32 %r1077, %r385, 2139062143;
sub.s32 %r1078, %r1076, %r1077;
xor.b32 %r390, %r715, %r1078;
ld.global.nc.u32 %r391, [%rd94+12];
mov.u32 %r387, %r583;

prmt.b32 %r387, %r387, 0, 0xba98;

xor.b32 %r1079, %r387, %r700;
or.b32 %r1080, %r1079, -2139062144;
and.b32 %r1081, %r387, 2139062143;
sub.s32 %r1082, %r1080, %r1081;
xor.b32 %r394, %r723, %r1082;
ld.global.nc.u32 %r395, [%rd94+16];

dp4a.s32.s32 %r389, %r390, %r391, %r381;


dp4a.s32.s32 %r393, %r394, %r395, %r389;

mov.u32 %r397, %r593;

prmt.b32 %r397, %r397, 0, 0xba98;

xor.b32 %r1083, %r397, %r725;
or.b32 %r1084, %r1083, -2139062144;
and.b32 %r1085, %r397, 2139062143;
sub.s32 %r1086, %r1084, %r1085;
xor.b32 %r402, %r743, %r1086;
ld.global.nc.u32 %r403, [%rd94+20];
mov.u32 %r399, %r595;

prmt.b32 %r399, %r399, 0, 0xba98;

xor.b32 %r1087, %r399, %r728;
or.b32 %r1088, %r1087, -2139062144;
and.b32 %r1089, %r399, 2139062143;
sub.s32 %r1090, %r1088, %r1089;
xor.b32 %r406, %r751, %r1090;
ld.global.nc.u32 %r407, [%rd94+24];

dp4a.s32.s32 %r401, %r402, %r403, %r393;


dp4a.s32.s32 %r405, %r406, %r407, %r401;

mov.u32 %r409, %r605;

prmt.b32 %r409, %r409, 0, 0xba98;

xor.b32 %r1091, %r409, %r753;
or.b32 %r1092, %r1091, -2139062144;
and.b32 %r1093, %r409, 2139062143;
sub.s32 %r1094, %r1092, %r1093;
xor.b32 %r414, %r771, %r1094;
ld.global.nc.u32 %r415, [%rd94+28];
mov.u32 %r411, %r607;

prmt.b32 %r411, %r411, 0, 0xba98;

xor.b32 %r1095, %r411, %r756;
or.b32 %r1096, %r1095, -2139062144;
and.b32 %r1097, %r411, 2139062143;
sub.s32 %r1098, %r1096, %r1097;
xor.b32 %r418, %r779, %r1098;
ld.global.nc.u32 %r419, [%rd94+32];

dp4a.s32.s32 %r413, %r414, %r415, %r405;


dp4a.s32.s32 %r417, %r418, %r419, %r413;

shr.u32 %r1099, %r417, 31;
add.s32 %r1100, %r417, %r1099;
shr.s32 %r1101, %r1100, 1;
mad.lo.s32 %r1102, %r417, %r781, %r1101;
shr.u32 %r1103, %r1102, 31;
add.s32 %r1104, %r1102, %r1103;
shr.s32 %r1105, %r1104, 1;

{ cvt.f32.f16 %f85, %rs1;}


ld.global.nc.u32 %r421, [%rd94];

{.reg .f16 low,high;
mov.b32 {low,high},%r421;
cvt.f32.f16 %f86, low;}


mul.ftz.f32 %f109, %f85, %f86;
cvt.rn.f32.s32 %f110, %r1105;
fma.rn.ftz.f32 %f270, %f109, %f110, %f270;
mov.u32 %r422, %r618;

prmt.b32 %r422, %r422, 0, 0xba98;

xor.b32 %r1106, %r422, %r792;
or.b32 %r1107, %r1106, -2139062144;
and.b32 %r1108, %r422, 2139062143;
sub.s32 %r1109, %r1107, %r1108;
xor.b32 %r427, %r809, %r1109;
mov.u32 %r424, %r620;

prmt.b32 %r424, %r424, 0, 0xba98;

xor.b32 %r1110, %r424, %r795;
or.b32 %r1111, %r1110, -2139062144;
and.b32 %r1112, %r424, 2139062143;
sub.s32 %r1113, %r1111, %r1112;
xor.b32 %r431, %r817, %r1113;

dp4a.s32.s32 %r426, %r427, %r379, %r625;


dp4a.s32.s32 %r430, %r431, %r383, %r426;

mov.u32 %r434, %r630;

prmt.b32 %r434, %r434, 0, 0xba98;

xor.b32 %r1114, %r434, %r819;
or.b32 %r1115, %r1114, -2139062144;
and.b32 %r1116, %r434, 2139062143;
sub.s32 %r1117, %r1115, %r1116;
xor.b32 %r439, %r837, %r1117;
mov.u32 %r436, %r632;

prmt.b32 %r436, %r436, 0, 0xba98;

xor.b32 %r1118, %r436, %r822;
or.b32 %r1119, %r1118, -2139062144;
and.b32 %r1120, %r436, 2139062143;
sub.s32 %r1121, %r1119, %r1120;
xor.b32 %r443, %r845, %r1121;

dp4a.s32.s32 %r438, %r439, %r391, %r430;


dp4a.s32.s32 %r442, %r443, %r395, %r438;

mov.u32 %r446, %r642;

prmt.b32 %r446, %r446, 0, 0xba98;

xor.b32 %r1122, %r446, %r847;
or.b32 %r1123, %r1122, -2139062144;
and.b32 %r1124, %r446, 2139062143;
sub.s32 %r1125, %r1123, %r1124;
xor.b32 %r451, %r865, %r1125;
mov.u32 %r448, %r644;

prmt.b32 %r448, %r448, 0, 0xba98;

xor.b32 %r1126, %r448, %r850;
or.b32 %r1127, %r1126, -2139062144;
and.b32 %r1128, %r448, 2139062143;
sub.s32 %r1129, %r1127, %r1128;
xor.b32 %r455, %r873, %r1129;

dp4a.s32.s32 %r450, %r451, %r403, %r442;


dp4a.s32.s32 %r454, %r455, %r407, %r450;

mov.u32 %r458, %r654;

prmt.b32 %r458, %r458, 0, 0xba98;

xor.b32 %r1130, %r458, %r875;
or.b32 %r1131, %r1130, -2139062144;
and.b32 %r1132, %r458, 2139062143;
sub.s32 %r1133, %r1131, %r1132;
xor.b32 %r463, %r893, %r1133;
mov.u32 %r460, %r656;

prmt.b32 %r460, %r460, 0, 0xba98;

xor.b32 %r1134, %r460, %r878;
or.b32 %r1135, %r1134, -2139062144;
and.b32 %r1136, %r460, 2139062143;
sub.s32 %r1137, %r1135, %r1136;
xor.b32 %r467, %r901, %r1137;

dp4a.s32.s32 %r462, %r463, %r415, %r454;


dp4a.s32.s32 %r466, %r467, %r419, %r462;

shr.u32 %r1138, %r466, 31;
add.s32 %r1139, %r466, %r1138;
shr.s32 %r1140, %r1139, 1;
mad.lo.s32 %r1141, %r466, %r903, %r1140;
shr.u32 %r1142, %r1141, 31;
add.s32 %r1143, %r1141, %r1142;
shr.s32 %r1144, %r1143, 1;

{ cvt.f32.f16 %f87, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r421;
cvt.f32.f16 %f88, low;}


mul.ftz.f32 %f111, %f87, %f88;
cvt.rn.f32.s32 %f112, %r1144;
fma.rn.ftz.f32 %f269, %f111, %f112, %f269;
cvt.u64.u32 %rd95, %r1590;
add.s64 %rd96, %rd1, %rd95;
add.s64 %rd97, %rd96, %rd3;
mul.lo.s64 %rd98, %rd97, 36;
mov.u32 %r471, %r569;

prmt.b32 %r471, %r471, 0, 0xba98;

xor.b32 %r1145, %r471, %r670;
or.b32 %r1146, %r1145, -2139062144;
and.b32 %r1147, %r471, 2139062143;
sub.s32 %r1148, %r1146, %r1147;
xor.b32 %r476, %r687, %r1148;
add.s64 %rd99, %rd10, %rd98;
ld.global.nc.u32 %r477, [%rd99+4];
mov.u32 %r473, %r571;

prmt.b32 %r473, %r473, 0, 0xba98;

xor.b32 %r1149, %r473, %r673;
or.b32 %r1150, %r1149, -2139062144;
and.b32 %r1151, %r473, 2139062143;
sub.s32 %r1152, %r1150, %r1151;
xor.b32 %r480, %r695, %r1152;
ld.global.nc.u32 %r481, [%rd99+8];

dp4a.s32.s32 %r475, %r476, %r477, %r625;


dp4a.s32.s32 %r479, %r480, %r481, %r475;

mov.u32 %r483, %r581;

prmt.b32 %r483, %r483, 0, 0xba98;

xor.b32 %r1153, %r483, %r697;
or.b32 %r1154, %r1153, -2139062144;
and.b32 %r1155, %r483, 2139062143;
sub.s32 %r1156, %r1154, %r1155;
xor.b32 %r488, %r715, %r1156;
ld.global.nc.u32 %r489, [%rd99+12];
mov.u32 %r485, %r583;

prmt.b32 %r485, %r485, 0, 0xba98;

xor.b32 %r1157, %r485, %r700;
or.b32 %r1158, %r1157, -2139062144;
and.b32 %r1159, %r485, 2139062143;
sub.s32 %r1160, %r1158, %r1159;
xor.b32 %r492, %r723, %r1160;
ld.global.nc.u32 %r493, [%rd99+16];

dp4a.s32.s32 %r487, %r488, %r489, %r479;


dp4a.s32.s32 %r491, %r492, %r493, %r487;

mov.u32 %r495, %r593;

prmt.b32 %r495, %r495, 0, 0xba98;

xor.b32 %r1161, %r495, %r725;
or.b32 %r1162, %r1161, -2139062144;
and.b32 %r1163, %r495, 2139062143;
sub.s32 %r1164, %r1162, %r1163;
xor.b32 %r500, %r743, %r1164;
ld.global.nc.u32 %r501, [%rd99+20];
mov.u32 %r497, %r595;

prmt.b32 %r497, %r497, 0, 0xba98;

xor.b32 %r1165, %r497, %r728;
or.b32 %r1166, %r1165, -2139062144;
and.b32 %r1167, %r497, 2139062143;
sub.s32 %r1168, %r1166, %r1167;
xor.b32 %r504, %r751, %r1168;
ld.global.nc.u32 %r505, [%rd99+24];

dp4a.s32.s32 %r499, %r500, %r501, %r491;


dp4a.s32.s32 %r503, %r504, %r505, %r499;

mov.u32 %r507, %r605;

prmt.b32 %r507, %r507, 0, 0xba98;

xor.b32 %r1169, %r507, %r753;
or.b32 %r1170, %r1169, -2139062144;
and.b32 %r1171, %r507, 2139062143;
sub.s32 %r1172, %r1170, %r1171;
xor.b32 %r512, %r771, %r1172;
ld.global.nc.u32 %r513, [%rd99+28];
mov.u32 %r509, %r607;

prmt.b32 %r509, %r509, 0, 0xba98;

xor.b32 %r1173, %r509, %r756;
or.b32 %r1174, %r1173, -2139062144;
and.b32 %r1175, %r509, 2139062143;
sub.s32 %r1176, %r1174, %r1175;
xor.b32 %r516, %r779, %r1176;
ld.global.nc.u32 %r517, [%rd99+32];

dp4a.s32.s32 %r511, %r512, %r513, %r503;


dp4a.s32.s32 %r515, %r516, %r517, %r511;

shr.u32 %r1177, %r515, 31;
add.s32 %r1178, %r515, %r1177;
shr.s32 %r1179, %r1178, 1;
mad.lo.s32 %r1180, %r515, %r781, %r1179;
shr.u32 %r1181, %r1180, 31;
add.s32 %r1182, %r1180, %r1181;
shr.s32 %r1183, %r1182, 1;

{ cvt.f32.f16 %f89, %rs1;}


ld.global.nc.u32 %r519, [%rd99];

{.reg .f16 low,high;
mov.b32 {low,high},%r519;
cvt.f32.f16 %f90, low;}


mul.ftz.f32 %f113, %f89, %f90;
cvt.rn.f32.s32 %f114, %r1183;
fma.rn.ftz.f32 %f268, %f113, %f114, %f268;
mov.u32 %r520, %r618;

prmt.b32 %r520, %r520, 0, 0xba98;

xor.b32 %r1184, %r520, %r792;
or.b32 %r1185, %r1184, -2139062144;
and.b32 %r1186, %r520, 2139062143;
sub.s32 %r1187, %r1185, %r1186;
xor.b32 %r525, %r809, %r1187;
mov.u32 %r522, %r620;

prmt.b32 %r522, %r522, 0, 0xba98;

xor.b32 %r1188, %r522, %r795;
or.b32 %r1189, %r1188, -2139062144;
and.b32 %r1190, %r522, 2139062143;
sub.s32 %r1191, %r1189, %r1190;
xor.b32 %r529, %r817, %r1191;

dp4a.s32.s32 %r524, %r525, %r477, %r625;


dp4a.s32.s32 %r528, %r529, %r481, %r524;

mov.u32 %r532, %r630;

prmt.b32 %r532, %r532, 0, 0xba98;

xor.b32 %r1192, %r532, %r819;
or.b32 %r1193, %r1192, -2139062144;
and.b32 %r1194, %r532, 2139062143;
sub.s32 %r1195, %r1193, %r1194;
xor.b32 %r537, %r837, %r1195;
mov.u32 %r534, %r632;

prmt.b32 %r534, %r534, 0, 0xba98;

xor.b32 %r1196, %r534, %r822;
or.b32 %r1197, %r1196, -2139062144;
and.b32 %r1198, %r534, 2139062143;
sub.s32 %r1199, %r1197, %r1198;
xor.b32 %r541, %r845, %r1199;

dp4a.s32.s32 %r536, %r537, %r489, %r528;


dp4a.s32.s32 %r540, %r541, %r493, %r536;

mov.u32 %r544, %r642;

prmt.b32 %r544, %r544, 0, 0xba98;

xor.b32 %r1200, %r544, %r847;
or.b32 %r1201, %r1200, -2139062144;
and.b32 %r1202, %r544, 2139062143;
sub.s32 %r1203, %r1201, %r1202;
xor.b32 %r549, %r865, %r1203;
mov.u32 %r546, %r644;

prmt.b32 %r546, %r546, 0, 0xba98;

xor.b32 %r1204, %r546, %r850;
or.b32 %r1205, %r1204, -2139062144;
and.b32 %r1206, %r546, 2139062143;
sub.s32 %r1207, %r1205, %r1206;
xor.b32 %r553, %r873, %r1207;

dp4a.s32.s32 %r548, %r549, %r501, %r540;


dp4a.s32.s32 %r552, %r553, %r505, %r548;

mov.u32 %r556, %r654;

prmt.b32 %r556, %r556, 0, 0xba98;

xor.b32 %r1208, %r556, %r875;
or.b32 %r1209, %r1208, -2139062144;
and.b32 %r1210, %r556, 2139062143;
sub.s32 %r1211, %r1209, %r1210;
xor.b32 %r561, %r893, %r1211;
mov.u32 %r558, %r656;

prmt.b32 %r558, %r558, 0, 0xba98;

xor.b32 %r1212, %r558, %r878;
or.b32 %r1213, %r1212, -2139062144;
and.b32 %r1214, %r558, 2139062143;
sub.s32 %r1215, %r1213, %r1214;
xor.b32 %r565, %r901, %r1215;

dp4a.s32.s32 %r560, %r561, %r513, %r552;


dp4a.s32.s32 %r564, %r565, %r517, %r560;

shr.u32 %r1216, %r564, 31;
add.s32 %r1217, %r564, %r1216;
shr.s32 %r1218, %r1217, 1;
mad.lo.s32 %r1219, %r564, %r903, %r1218;
shr.u32 %r1220, %r1219, 31;
add.s32 %r1221, %r1219, %r1220;
shr.s32 %r1222, %r1221, 1;

{ cvt.f32.f16 %f91, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r519;
cvt.f32.f16 %f92, low;}


mul.ftz.f32 %f115, %f91, %f92;
cvt.rn.f32.s32 %f116, %r1222;
fma.rn.ftz.f32 %f267, %f115, %f116, %f267;
cvt.u64.u32 %rd100, %r1589;
add.s64 %rd101, %rd1, %rd100;
add.s64 %rd102, %rd101, %rd3;
mul.lo.s64 %rd103, %rd102, 36;

prmt.b32 %r569, %r569, 0, 0xba98;

xor.b32 %r1223, %r569, %r670;
or.b32 %r1224, %r1223, -2139062144;
and.b32 %r1225, %r569, 2139062143;
sub.s32 %r1226, %r1224, %r1225;
xor.b32 %r574, %r687, %r1226;
add.s64 %rd104, %rd10, %rd103;
ld.global.nc.u32 %r575, [%rd104+4];

prmt.b32 %r571, %r571, 0, 0xba98;

xor.b32 %r1227, %r571, %r673;
or.b32 %r1228, %r1227, -2139062144;
and.b32 %r1229, %r571, 2139062143;
sub.s32 %r1230, %r1228, %r1229;
xor.b32 %r578, %r695, %r1230;
ld.global.nc.u32 %r579, [%rd104+8];

dp4a.s32.s32 %r573, %r574, %r575, %r625;


dp4a.s32.s32 %r577, %r578, %r579, %r573;


prmt.b32 %r581, %r581, 0, 0xba98;

xor.b32 %r1231, %r581, %r697;
or.b32 %r1232, %r1231, -2139062144;
and.b32 %r1233, %r581, 2139062143;
sub.s32 %r1234, %r1232, %r1233;
xor.b32 %r586, %r715, %r1234;
ld.global.nc.u32 %r587, [%rd104+12];

prmt.b32 %r583, %r583, 0, 0xba98;

xor.b32 %r1235, %r583, %r700;
or.b32 %r1236, %r1235, -2139062144;
and.b32 %r1237, %r583, 2139062143;
sub.s32 %r1238, %r1236, %r1237;
xor.b32 %r590, %r723, %r1238;
ld.global.nc.u32 %r591, [%rd104+16];

dp4a.s32.s32 %r585, %r586, %r587, %r577;


dp4a.s32.s32 %r589, %r590, %r591, %r585;


prmt.b32 %r593, %r593, 0, 0xba98;

xor.b32 %r1239, %r593, %r725;
or.b32 %r1240, %r1239, -2139062144;
and.b32 %r1241, %r593, 2139062143;
sub.s32 %r1242, %r1240, %r1241;
xor.b32 %r598, %r743, %r1242;
ld.global.nc.u32 %r599, [%rd104+20];

prmt.b32 %r595, %r595, 0, 0xba98;

xor.b32 %r1243, %r595, %r728;
or.b32 %r1244, %r1243, -2139062144;
and.b32 %r1245, %r595, 2139062143;
sub.s32 %r1246, %r1244, %r1245;
xor.b32 %r602, %r751, %r1246;
ld.global.nc.u32 %r603, [%rd104+24];

dp4a.s32.s32 %r597, %r598, %r599, %r589;


dp4a.s32.s32 %r601, %r602, %r603, %r597;


prmt.b32 %r605, %r605, 0, 0xba98;

xor.b32 %r1247, %r605, %r753;
or.b32 %r1248, %r1247, -2139062144;
and.b32 %r1249, %r605, 2139062143;
sub.s32 %r1250, %r1248, %r1249;
xor.b32 %r610, %r771, %r1250;
ld.global.nc.u32 %r611, [%rd104+28];

prmt.b32 %r607, %r607, 0, 0xba98;

xor.b32 %r1251, %r607, %r756;
or.b32 %r1252, %r1251, -2139062144;
and.b32 %r1253, %r607, 2139062143;
sub.s32 %r1254, %r1252, %r1253;
xor.b32 %r614, %r779, %r1254;
ld.global.nc.u32 %r615, [%rd104+32];

dp4a.s32.s32 %r609, %r610, %r611, %r601;


dp4a.s32.s32 %r613, %r614, %r615, %r609;

shr.u32 %r1255, %r613, 31;
add.s32 %r1256, %r613, %r1255;
shr.s32 %r1257, %r1256, 1;
mad.lo.s32 %r1258, %r613, %r781, %r1257;
shr.u32 %r1259, %r1258, 31;
add.s32 %r1260, %r1258, %r1259;
shr.s32 %r1261, %r1260, 1;

{ cvt.f32.f16 %f93, %rs1;}


ld.global.nc.u32 %r617, [%rd104];

{.reg .f16 low,high;
mov.b32 {low,high},%r617;
cvt.f32.f16 %f94, low;}


mul.ftz.f32 %f117, %f93, %f94;
cvt.rn.f32.s32 %f118, %r1261;
fma.rn.ftz.f32 %f266, %f117, %f118, %f266;

prmt.b32 %r618, %r618, 0, 0xba98;

xor.b32 %r1262, %r618, %r792;
or.b32 %r1263, %r1262, -2139062144;
and.b32 %r1264, %r618, 2139062143;
sub.s32 %r1265, %r1263, %r1264;
xor.b32 %r623, %r809, %r1265;

prmt.b32 %r620, %r620, 0, 0xba98;

xor.b32 %r1266, %r620, %r795;
or.b32 %r1267, %r1266, -2139062144;
and.b32 %r1268, %r620, 2139062143;
sub.s32 %r1269, %r1267, %r1268;
xor.b32 %r627, %r817, %r1269;

dp4a.s32.s32 %r622, %r623, %r575, %r625;


dp4a.s32.s32 %r626, %r627, %r579, %r622;


prmt.b32 %r630, %r630, 0, 0xba98;

xor.b32 %r1270, %r630, %r819;
or.b32 %r1271, %r1270, -2139062144;
and.b32 %r1272, %r630, 2139062143;
sub.s32 %r1273, %r1271, %r1272;
xor.b32 %r635, %r837, %r1273;

prmt.b32 %r632, %r632, 0, 0xba98;

xor.b32 %r1274, %r632, %r822;
or.b32 %r1275, %r1274, -2139062144;
and.b32 %r1276, %r632, 2139062143;
sub.s32 %r1277, %r1275, %r1276;
xor.b32 %r639, %r845, %r1277;

dp4a.s32.s32 %r634, %r635, %r587, %r626;


dp4a.s32.s32 %r638, %r639, %r591, %r634;


prmt.b32 %r642, %r642, 0, 0xba98;

xor.b32 %r1278, %r642, %r847;
or.b32 %r1279, %r1278, -2139062144;
and.b32 %r1280, %r642, 2139062143;
sub.s32 %r1281, %r1279, %r1280;
xor.b32 %r647, %r865, %r1281;

prmt.b32 %r644, %r644, 0, 0xba98;

xor.b32 %r1282, %r644, %r850;
or.b32 %r1283, %r1282, -2139062144;
and.b32 %r1284, %r644, 2139062143;
sub.s32 %r1285, %r1283, %r1284;
xor.b32 %r651, %r873, %r1285;

dp4a.s32.s32 %r646, %r647, %r599, %r638;


dp4a.s32.s32 %r650, %r651, %r603, %r646;


prmt.b32 %r654, %r654, 0, 0xba98;

xor.b32 %r1286, %r654, %r875;
or.b32 %r1287, %r1286, -2139062144;
and.b32 %r1288, %r654, 2139062143;
sub.s32 %r1289, %r1287, %r1288;
xor.b32 %r659, %r893, %r1289;

prmt.b32 %r656, %r656, 0, 0xba98;

xor.b32 %r1290, %r656, %r878;
or.b32 %r1291, %r1290, -2139062144;
and.b32 %r1292, %r656, 2139062143;
sub.s32 %r1293, %r1291, %r1292;
xor.b32 %r663, %r901, %r1293;

dp4a.s32.s32 %r658, %r659, %r611, %r650;


dp4a.s32.s32 %r662, %r663, %r615, %r658;

shr.u32 %r1294, %r662, 31;
add.s32 %r1295, %r662, %r1294;
shr.s32 %r1296, %r1295, 1;
mad.lo.s32 %r1297, %r662, %r903, %r1296;
shr.u32 %r1298, %r1297, 31;
add.s32 %r1299, %r1297, %r1298;
shr.s32 %r1300, %r1299, 1;

{ cvt.f32.f16 %f95, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r617;
cvt.f32.f16 %f96, low;}


mul.ftz.f32 %f119, %f95, %f96;
cvt.rn.f32.s32 %f120, %r1300;
fma.rn.ftz.f32 %f265, %f119, %f120, %f265;
add.s32 %r1592, %r1592, 64;
add.s32 %r1591, %r1591, 64;
add.s32 %r1590, %r1590, 64;
add.s32 %r1589, %r1589, 64;
add.s32 %r1588, %r1588, 64;
add.s32 %r1587, %r1587, 64;
add.s64 %rd149, %rd149, 784;
add.s32 %r1593, %r1593, 8;
setp.lt.s32 %p4, %r1593, %r3;
@%p4 bra $L__BB247_2;

$L__BB247_3:
setp.eq.s32 %p5, %r42, 0;
@%p5 bra $L__BB247_5;

mov.u32 %r1304, _ZZ13mul_mat_vec_qIL9ggml_type18ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r1305, %r42, 1536, %r1304;
shl.b32 %r1306, %r44, 2;
add.s32 %r1307, %r1305, %r1306;
st.shared.f32 [%r1307+-1536], %f276;
st.shared.f32 [%r1307+-1408], %f275;
st.shared.f32 [%r1307+-1280], %f274;
st.shared.f32 [%r1307+-1152], %f273;
st.shared.f32 [%r1307+-1024], %f272;
st.shared.f32 [%r1307+-896], %f271;
st.shared.f32 [%r1307+-768], %f270;
st.shared.f32 [%r1307+-640], %f269;
st.shared.f32 [%r1307+-512], %f268;
st.shared.f32 [%r1307+-384], %f267;
st.shared.f32 [%r1307+-256], %f266;
st.shared.f32 [%r1307+-128], %f265;

$L__BB247_5:
setp.ne.s32 %p6, %r42, 0;
bar.sync 0;
@%p6 bra $L__BB247_30;

mov.u32 %r1309, %ctaid.z;
mov.u32 %r1310, %ctaid.y;
mov.u32 %r1311, %ctaid.x;
shl.b32 %r1312, %r1311, 1;
mov.u32 %r1313, 1;
mad.lo.s32 %r1314, %r1310, %r34, %r1312;
mad.lo.s32 %r1315, %r1309, %r40, %r1314;
cvt.u64.u32 %rd13, %r1315;
shl.b32 %r1317, %r44, 2;
mov.u32 %r1318, _ZZ13mul_mat_vec_qIL9ggml_type18ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r1319, %r1318, %r1317;
mov.u32 %r1320, 2;
ld.shared.f32 %f121, [%r1319];
add.ftz.f32 %f122, %f121, %f276;
mov.b32 %r1321, %f122;
mov.u32 %r1322, 31;
mov.u32 %r1323, 16;
mov.u32 %r1324, -1;
shfl.sync.bfly.b32 %r1325|%p7, %r1321, %r1323, %r1322, %r1324;
mov.b32 %f123, %r1325;
add.ftz.f32 %f124, %f122, %f123;
mov.b32 %r1326, %f124;
mov.u32 %r1327, 8;
shfl.sync.bfly.b32 %r1328|%p8, %r1326, %r1327, %r1322, %r1324;
mov.b32 %f125, %r1328;
add.ftz.f32 %f126, %f124, %f125;
mov.b32 %r1329, %f126;
mov.u32 %r1330, 4;
shfl.sync.bfly.b32 %r1331|%p9, %r1329, %r1330, %r1322, %r1324;
mov.b32 %f127, %r1331;
add.ftz.f32 %f128, %f126, %f127;
mov.b32 %r1332, %f128;
shfl.sync.bfly.b32 %r1333|%p10, %r1332, %r1320, %r1322, %r1324;
mov.b32 %f129, %r1333;
add.ftz.f32 %f130, %f128, %f129;
mov.b32 %r1334, %f130;
shfl.sync.bfly.b32 %r1335|%p11, %r1334, %r1313, %r1322, %r1324;
mov.b32 %f131, %r1335;
add.ftz.f32 %f37, %f130, %f131;
setp.eq.s32 %p12, %r44, 0;
setp.lt.u32 %p13, %r1312, %r28;
and.pred %p1, %p12, %p13;
not.pred %p14, %p1;
@%p14 bra $L__BB247_8;

cvta.to.global.u64 %rd105, %rd18;
mul.wide.u32 %rd106, %r1315, 4;
add.s64 %rd107, %rd105, %rd106;
st.global.f32 [%rd107], %f37;

$L__BB247_8:
ld.shared.f32 %f132, [%r1319+128];
add.ftz.f32 %f133, %f132, %f275;
mov.b32 %r1347, %f133;
shfl.sync.bfly.b32 %r1351|%p15, %r1347, %r1323, %r1322, %r1324;
mov.b32 %f134, %r1351;
add.ftz.f32 %f135, %f133, %f134;
mov.b32 %r1352, %f135;
shfl.sync.bfly.b32 %r1354|%p16, %r1352, %r1327, %r1322, %r1324;
mov.b32 %f136, %r1354;
add.ftz.f32 %f137, %f135, %f136;
mov.b32 %r1355, %f137;
shfl.sync.bfly.b32 %r1357|%p17, %r1355, %r1330, %r1322, %r1324;
mov.b32 %f138, %r1357;
add.ftz.f32 %f139, %f137, %f138;
mov.b32 %r1358, %f139;
shfl.sync.bfly.b32 %r1359|%p18, %r1358, %r1320, %r1322, %r1324;
mov.b32 %f140, %r1359;
add.ftz.f32 %f141, %f139, %f140;
mov.b32 %r1360, %f141;
shfl.sync.bfly.b32 %r1362|%p19, %r1360, %r1313, %r1322, %r1324;
mov.b32 %f142, %r1362;
add.ftz.f32 %f38, %f141, %f142;
setp.eq.s32 %p20, %r44, 1;
or.b32 %r1365, %r1312, 1;
setp.lt.u32 %p21, %r1365, %r28;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB247_10;

cvta.to.global.u64 %rd108, %rd18;
mul.wide.u32 %rd109, %r1315, 4;
add.s64 %rd110, %rd108, %rd109;
st.global.f32 [%rd110+4], %f38;

$L__BB247_10:
mov.u32 %r1376, 2;
ld.shared.f32 %f143, [%r1319+256];
add.ftz.f32 %f144, %f143, %f274;
mov.b32 %r1377, %f144;
mov.u32 %r1378, 31;
mov.u32 %r1379, 16;
mov.u32 %r1380, -1;
shfl.sync.bfly.b32 %r1381|%p23, %r1377, %r1379, %r1378, %r1380;
mov.b32 %f145, %r1381;
add.ftz.f32 %f146, %f144, %f145;
mov.b32 %r1382, %f146;
mov.u32 %r1383, 8;
shfl.sync.bfly.b32 %r1384|%p24, %r1382, %r1383, %r1378, %r1380;
mov.b32 %f147, %r1384;
add.ftz.f32 %f148, %f146, %f147;
mov.b32 %r1385, %f148;
mov.u32 %r1386, 4;
shfl.sync.bfly.b32 %r1387|%p25, %r1385, %r1386, %r1378, %r1380;
mov.b32 %f149, %r1387;
add.ftz.f32 %f150, %f148, %f149;
mov.b32 %r1388, %f150;
shfl.sync.bfly.b32 %r1389|%p26, %r1388, %r1376, %r1378, %r1380;
mov.b32 %f151, %r1389;
add.ftz.f32 %f152, %f150, %f151;
mov.b32 %r1390, %f152;
mov.u32 %r1391, 1;
shfl.sync.bfly.b32 %r1392|%p27, %r1390, %r1391, %r1378, %r1380;
mov.b32 %f153, %r1392;
add.ftz.f32 %f39, %f152, %f153;
@%p14 bra $L__BB247_12;

cvt.u64.u32 %rd111, %r28;
add.s64 %rd112, %rd111, %rd13;
cvta.to.global.u64 %rd113, %rd18;
shl.b64 %rd114, %rd112, 2;
add.s64 %rd115, %rd113, %rd114;
st.global.f32 [%rd115], %f39;

$L__BB247_12:
ld.shared.f32 %f154, [%r1319+384];
add.ftz.f32 %f155, %f154, %f273;
mov.b32 %r1398, %f155;
shfl.sync.bfly.b32 %r1402|%p29, %r1398, %r1379, %r1378, %r1380;
mov.b32 %f156, %r1402;
add.ftz.f32 %f157, %f155, %f156;
mov.b32 %r1403, %f157;
shfl.sync.bfly.b32 %r1405|%p30, %r1403, %r1383, %r1378, %r1380;
mov.b32 %f158, %r1405;
add.ftz.f32 %f159, %f157, %f158;
mov.b32 %r1406, %f159;
shfl.sync.bfly.b32 %r1408|%p31, %r1406, %r1386, %r1378, %r1380;
mov.b32 %f160, %r1408;
add.ftz.f32 %f161, %f159, %f160;
mov.b32 %r1409, %f161;
shfl.sync.bfly.b32 %r1410|%p32, %r1409, %r1376, %r1378, %r1380;
mov.b32 %f162, %r1410;
add.ftz.f32 %f163, %f161, %f162;
mov.b32 %r1411, %f163;
shfl.sync.bfly.b32 %r1413|%p33, %r1411, %r1391, %r1378, %r1380;
mov.b32 %f164, %r1413;
add.ftz.f32 %f40, %f163, %f164;
@%p22 bra $L__BB247_14;

add.s32 %r1414, %r28, 1;
cvt.u64.u32 %rd116, %r1414;
add.s64 %rd117, %rd116, %rd13;
cvta.to.global.u64 %rd118, %rd18;
shl.b64 %rd119, %rd117, 2;
add.s64 %rd120, %rd118, %rd119;
st.global.f32 [%rd120], %f40;

$L__BB247_14:
shl.b32 %r1415, %r28, 1;
mov.u32 %r1416, 1;
mov.u32 %r1421, 2;
ld.shared.f32 %f165, [%r1319+512];
add.ftz.f32 %f166, %f165, %f272;
mov.b32 %r1422, %f166;
mov.u32 %r1423, 31;
mov.u32 %r1424, 16;
mov.u32 %r1425, -1;
shfl.sync.bfly.b32 %r1426|%p35, %r1422, %r1424, %r1423, %r1425;
mov.b32 %f167, %r1426;
add.ftz.f32 %f168, %f166, %f167;
mov.b32 %r1427, %f168;
mov.u32 %r1428, 8;
shfl.sync.bfly.b32 %r1429|%p36, %r1427, %r1428, %r1423, %r1425;
mov.b32 %f169, %r1429;
add.ftz.f32 %f170, %f168, %f169;
mov.b32 %r1430, %f170;
mov.u32 %r1431, 4;
shfl.sync.bfly.b32 %r1432|%p37, %r1430, %r1431, %r1423, %r1425;
mov.b32 %f171, %r1432;
add.ftz.f32 %f172, %f170, %f171;
mov.b32 %r1433, %f172;
shfl.sync.bfly.b32 %r1434|%p38, %r1433, %r1421, %r1423, %r1425;
mov.b32 %f173, %r1434;
add.ftz.f32 %f174, %f172, %f173;
mov.b32 %r1435, %f174;
shfl.sync.bfly.b32 %r1436|%p39, %r1435, %r1416, %r1423, %r1425;
mov.b32 %f175, %r1436;
add.ftz.f32 %f41, %f174, %f175;
cvt.u64.u32 %rd121, %r1415;
add.s64 %rd122, %rd121, %rd13;
cvta.to.global.u64 %rd123, %rd18;
shl.b64 %rd124, %rd122, 2;
add.s64 %rd14, %rd123, %rd124;
@%p14 bra $L__BB247_16;

st.global.f32 [%rd14], %f41;

$L__BB247_16:
ld.shared.f32 %f176, [%r1319+640];
add.ftz.f32 %f177, %f176, %f271;
mov.b32 %r1442, %f177;
shfl.sync.bfly.b32 %r1446|%p41, %r1442, %r1424, %r1423, %r1425;
mov.b32 %f178, %r1446;
add.ftz.f32 %f179, %f177, %f178;
mov.b32 %r1447, %f179;
shfl.sync.bfly.b32 %r1449|%p42, %r1447, %r1428, %r1423, %r1425;
mov.b32 %f180, %r1449;
add.ftz.f32 %f181, %f179, %f180;
mov.b32 %r1450, %f181;
shfl.sync.bfly.b32 %r1452|%p43, %r1450, %r1431, %r1423, %r1425;
mov.b32 %f182, %r1452;
add.ftz.f32 %f183, %f181, %f182;
mov.b32 %r1453, %f183;
shfl.sync.bfly.b32 %r1454|%p44, %r1453, %r1421, %r1423, %r1425;
mov.b32 %f184, %r1454;
add.ftz.f32 %f185, %f183, %f184;
mov.b32 %r1455, %f185;
shfl.sync.bfly.b32 %r1457|%p45, %r1455, %r1416, %r1423, %r1425;
mov.b32 %f186, %r1457;
add.ftz.f32 %f42, %f185, %f186;
@%p22 bra $L__BB247_18;

st.global.f32 [%rd14+4], %f42;

$L__BB247_18:
mul.lo.s32 %r24, %r28, 3;
mov.u32 %r1462, 2;
ld.shared.f32 %f187, [%r1319+768];
add.ftz.f32 %f188, %f187, %f270;
mov.b32 %r1463, %f188;
mov.u32 %r1464, 31;
mov.u32 %r1465, 16;
mov.u32 %r1466, -1;
shfl.sync.bfly.b32 %r1467|%p47, %r1463, %r1465, %r1464, %r1466;
mov.b32 %f189, %r1467;
add.ftz.f32 %f190, %f188, %f189;
mov.b32 %r1468, %f190;
mov.u32 %r1469, 8;
shfl.sync.bfly.b32 %r1470|%p48, %r1468, %r1469, %r1464, %r1466;
mov.b32 %f191, %r1470;
add.ftz.f32 %f192, %f190, %f191;
mov.b32 %r1471, %f192;
mov.u32 %r1472, 4;
shfl.sync.bfly.b32 %r1473|%p49, %r1471, %r1472, %r1464, %r1466;
mov.b32 %f193, %r1473;
add.ftz.f32 %f194, %f192, %f193;
mov.b32 %r1474, %f194;
shfl.sync.bfly.b32 %r1475|%p50, %r1474, %r1462, %r1464, %r1466;
mov.b32 %f195, %r1475;
add.ftz.f32 %f196, %f194, %f195;
mov.b32 %r1476, %f196;
mov.u32 %r1477, 1;
shfl.sync.bfly.b32 %r1478|%p51, %r1476, %r1477, %r1464, %r1466;
mov.b32 %f197, %r1478;
add.ftz.f32 %f43, %f196, %f197;
@%p14 bra $L__BB247_20;

cvt.u64.u32 %rd125, %r24;
add.s64 %rd126, %rd125, %rd13;
shl.b64 %rd128, %rd126, 2;
add.s64 %rd129, %rd123, %rd128;
st.global.f32 [%rd129], %f43;

$L__BB247_20:
ld.shared.f32 %f198, [%r1319+896];
add.ftz.f32 %f199, %f198, %f269;
mov.b32 %r1484, %f199;
shfl.sync.bfly.b32 %r1488|%p53, %r1484, %r1465, %r1464, %r1466;
mov.b32 %f200, %r1488;
add.ftz.f32 %f201, %f199, %f200;
mov.b32 %r1489, %f201;
shfl.sync.bfly.b32 %r1491|%p54, %r1489, %r1469, %r1464, %r1466;
mov.b32 %f202, %r1491;
add.ftz.f32 %f203, %f201, %f202;
mov.b32 %r1492, %f203;
shfl.sync.bfly.b32 %r1494|%p55, %r1492, %r1472, %r1464, %r1466;
mov.b32 %f204, %r1494;
add.ftz.f32 %f205, %f203, %f204;
mov.b32 %r1495, %f205;
shfl.sync.bfly.b32 %r1496|%p56, %r1495, %r1462, %r1464, %r1466;
mov.b32 %f206, %r1496;
add.ftz.f32 %f207, %f205, %f206;
mov.b32 %r1497, %f207;
shfl.sync.bfly.b32 %r1499|%p57, %r1497, %r1477, %r1464, %r1466;
mov.b32 %f208, %r1499;
add.ftz.f32 %f44, %f207, %f208;
@%p22 bra $L__BB247_22;

add.s32 %r1500, %r24, 1;
cvt.u64.u32 %rd130, %r1500;
add.s64 %rd131, %rd130, %rd13;
shl.b64 %rd133, %rd131, 2;
add.s64 %rd134, %rd123, %rd133;
st.global.f32 [%rd134], %f44;

$L__BB247_22:
shl.b32 %r1501, %r28, 2;
mov.u32 %r1502, 2;
ld.shared.f32 %f209, [%r1319+1024];
add.ftz.f32 %f210, %f209, %f268;
mov.b32 %r1507, %f210;
mov.u32 %r1508, 31;
mov.u32 %r1509, 16;
mov.u32 %r1510, -1;
shfl.sync.bfly.b32 %r1511|%p59, %r1507, %r1509, %r1508, %r1510;
mov.b32 %f211, %r1511;
add.ftz.f32 %f212, %f210, %f211;
mov.b32 %r1512, %f212;
mov.u32 %r1513, 8;
shfl.sync.bfly.b32 %r1514|%p60, %r1512, %r1513, %r1508, %r1510;
mov.b32 %f213, %r1514;
add.ftz.f32 %f214, %f212, %f213;
mov.b32 %r1515, %f214;
mov.u32 %r1516, 4;
shfl.sync.bfly.b32 %r1517|%p61, %r1515, %r1516, %r1508, %r1510;
mov.b32 %f215, %r1517;
add.ftz.f32 %f216, %f214, %f215;
mov.b32 %r1518, %f216;
shfl.sync.bfly.b32 %r1519|%p62, %r1518, %r1502, %r1508, %r1510;
mov.b32 %f217, %r1519;
add.ftz.f32 %f218, %f216, %f217;
mov.b32 %r1520, %f218;
mov.u32 %r1521, 1;
shfl.sync.bfly.b32 %r1522|%p63, %r1520, %r1521, %r1508, %r1510;
mov.b32 %f219, %r1522;
add.ftz.f32 %f45, %f218, %f219;
cvt.u64.u32 %rd135, %r1501;
add.s64 %rd136, %rd135, %rd13;
shl.b64 %rd138, %rd136, 2;
add.s64 %rd15, %rd123, %rd138;
@%p14 bra $L__BB247_24;

st.global.f32 [%rd15], %f45;

$L__BB247_24:
ld.shared.f32 %f220, [%r1319+1152];
add.ftz.f32 %f221, %f220, %f267;
mov.b32 %r1528, %f221;
shfl.sync.bfly.b32 %r1532|%p65, %r1528, %r1509, %r1508, %r1510;
mov.b32 %f222, %r1532;
add.ftz.f32 %f223, %f221, %f222;
mov.b32 %r1533, %f223;
shfl.sync.bfly.b32 %r1535|%p66, %r1533, %r1513, %r1508, %r1510;
mov.b32 %f224, %r1535;
add.ftz.f32 %f225, %f223, %f224;
mov.b32 %r1536, %f225;
shfl.sync.bfly.b32 %r1538|%p67, %r1536, %r1516, %r1508, %r1510;
mov.b32 %f226, %r1538;
add.ftz.f32 %f227, %f225, %f226;
mov.b32 %r1539, %f227;
shfl.sync.bfly.b32 %r1540|%p68, %r1539, %r1502, %r1508, %r1510;
mov.b32 %f228, %r1540;
add.ftz.f32 %f229, %f227, %f228;
mov.b32 %r1541, %f229;
shfl.sync.bfly.b32 %r1543|%p69, %r1541, %r1521, %r1508, %r1510;
mov.b32 %f230, %r1543;
add.ftz.f32 %f46, %f229, %f230;
@%p22 bra $L__BB247_26;

st.global.f32 [%rd15+4], %f46;

$L__BB247_26:
mul.lo.s32 %r25, %r28, 5;
mov.u32 %r1548, 2;
ld.shared.f32 %f231, [%r1319+1280];
add.ftz.f32 %f232, %f231, %f266;
mov.b32 %r1549, %f232;
mov.u32 %r1550, 31;
mov.u32 %r1551, 16;
mov.u32 %r1552, -1;
shfl.sync.bfly.b32 %r1553|%p71, %r1549, %r1551, %r1550, %r1552;
mov.b32 %f233, %r1553;
add.ftz.f32 %f234, %f232, %f233;
mov.b32 %r1554, %f234;
mov.u32 %r1555, 8;
shfl.sync.bfly.b32 %r1556|%p72, %r1554, %r1555, %r1550, %r1552;
mov.b32 %f235, %r1556;
add.ftz.f32 %f236, %f234, %f235;
mov.b32 %r1557, %f236;
mov.u32 %r1558, 4;
shfl.sync.bfly.b32 %r1559|%p73, %r1557, %r1558, %r1550, %r1552;
mov.b32 %f237, %r1559;
add.ftz.f32 %f238, %f236, %f237;
mov.b32 %r1560, %f238;
shfl.sync.bfly.b32 %r1561|%p74, %r1560, %r1548, %r1550, %r1552;
mov.b32 %f239, %r1561;
add.ftz.f32 %f240, %f238, %f239;
mov.b32 %r1562, %f240;
mov.u32 %r1563, 1;
shfl.sync.bfly.b32 %r1564|%p75, %r1562, %r1563, %r1550, %r1552;
mov.b32 %f241, %r1564;
add.ftz.f32 %f47, %f240, %f241;
@%p14 bra $L__BB247_28;

cvt.u64.u32 %rd139, %r25;
add.s64 %rd140, %rd139, %rd13;
shl.b64 %rd142, %rd140, 2;
add.s64 %rd143, %rd123, %rd142;
st.global.f32 [%rd143], %f47;

$L__BB247_28:
ld.shared.f32 %f242, [%r1319+1408];
add.ftz.f32 %f243, %f242, %f265;
mov.b32 %r1570, %f243;
shfl.sync.bfly.b32 %r1574|%p77, %r1570, %r1551, %r1550, %r1552;
mov.b32 %f244, %r1574;
add.ftz.f32 %f245, %f243, %f244;
mov.b32 %r1575, %f245;
shfl.sync.bfly.b32 %r1577|%p78, %r1575, %r1555, %r1550, %r1552;
mov.b32 %f246, %r1577;
add.ftz.f32 %f247, %f245, %f246;
mov.b32 %r1578, %f247;
shfl.sync.bfly.b32 %r1580|%p79, %r1578, %r1558, %r1550, %r1552;
mov.b32 %f248, %r1580;
add.ftz.f32 %f249, %f247, %f248;
mov.b32 %r1581, %f249;
shfl.sync.bfly.b32 %r1582|%p80, %r1581, %r1548, %r1550, %r1552;
mov.b32 %f250, %r1582;
add.ftz.f32 %f251, %f249, %f250;
mov.b32 %r1583, %f251;
shfl.sync.bfly.b32 %r1585|%p81, %r1583, %r1563, %r1550, %r1552;
mov.b32 %f252, %r1585;
add.ftz.f32 %f48, %f251, %f252;
@%p22 bra $L__BB247_30;

add.s32 %r1586, %r25, 1;
cvt.u64.u32 %rd144, %r1586;
add.s64 %rd145, %rd144, %rd13;
shl.b64 %rd147, %rd145, 2;
add.s64 %rd148, %rd123, %rd147;
st.global.f32 [%rd148], %f48;

$L__BB247_30:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type18ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type18ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type18ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type18ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type18ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type18ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type18ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type18ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type18ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<106>;
.reg .b16 %rs<43>;
.reg .f32 %f<323>;
.reg .b32 %r<1874>;
.reg .b64 %rd<168>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type18ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1792];

ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type18ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type18ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type18ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type18ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type18ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type18ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type18ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type18ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type18ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type18ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type18ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type18ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type18ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type18ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type18ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type18ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type18ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r44, %tid.y;
shl.b32 %r45, %r44, 5;
mov.u32 %r46, %tid.x;
add.s32 %r1, %r45, %r46;
shr.s32 %r47, %r1, 31;
shr.u32 %r48, %r47, 29;
add.s32 %r49, %r1, %r48;
shr.s32 %r1873, %r49, 3;
shr.u32 %r50, %r28, 8;
setp.ge.s32 %p2, %r1873, %r50;
mov.f32 %f309, 0f00000000;
mov.f32 %f310, %f309;
mov.f32 %f311, %f309;
mov.f32 %f312, %f309;
mov.f32 %f313, %f309;
mov.f32 %f314, %f309;
mov.f32 %f315, %f309;
mov.f32 %f316, %f309;
mov.f32 %f317, %f309;
mov.f32 %f318, %f309;
mov.f32 %f319, %f309;
mov.f32 %f320, %f309;
mov.f32 %f321, %f309;
mov.f32 %f322, %f309;
@%p2 bra $L__BB248_3;

mov.u32 %r51, %ctaid.x;
shl.b32 %r52, %r51, 1;
mov.u32 %r53, %ctaid.y;
mov.u32 %r54, %ctaid.z;
mul.lo.s32 %r55, %r53, %r36;
cvt.u64.u32 %rd19, %r55;
mul.lo.s32 %r56, %r54, %r42;
cvt.u64.u32 %rd20, %r56;
add.s64 %rd1, %rd20, %rd19;
mul.hi.u32 %r57, %r54, %r38;
add.s32 %r58, %r57, %r54;
shr.u32 %r59, %r58, %r39;
mul.lo.s32 %r60, %r59, %r41;
mul.hi.u32 %r61, %r53, %r32;
add.s32 %r62, %r61, %r53;
shr.u32 %r63, %r62, %r33;
mul.lo.s32 %r64, %r63, %r35;
cvta.to.global.u64 %rd167, %rd16;
and.b32 %r68, %r49, -8;
sub.s32 %r69, %r1, %r68;
shl.b32 %r70, %r69, 2;
shl.b32 %r71, %r69, 1;
add.s32 %r72, %r71, 33;
cvt.s64.s32 %rd3, %r69;
shl.b32 %r73, %r30, 1;
shl.b32 %r1866, %r1873, 3;
add.s32 %r1872, %r1866, %r73;
mad.lo.s32 %r1871, %r30, 3, %r1866;
shl.b32 %r74, %r30, 2;
add.s32 %r1870, %r1866, %r74;
mad.lo.s32 %r1869, %r30, 5, %r1866;
mad.lo.s32 %r1868, %r30, 6, %r1866;
add.s32 %r1867, %r30, %r1866;
add.s32 %r75, %r1873, %r60;
add.s32 %r76, %r75, %r64;
or.b32 %r77, %r52, 1;
mad.lo.s32 %r78, %r29, %r77, %r76;
mul.wide.s32 %rd4, %r78, 98;
mad.lo.s32 %r79, %r52, %r29, %r1873;
add.s32 %r80, %r79, %r60;
add.s32 %r81, %r80, %r64;
mul.wide.s32 %rd5, %r81, 98;
mul.wide.s32 %rd21, %r70, 2;
add.s64 %rd22, %rd5, %rd21;
add.s64 %rd6, %rd22, 8;
mul.wide.s32 %rd23, %r72, 2;
add.s64 %rd24, %rd23, -2;
add.s64 %rd25, %rd24, %rd4;
add.s64 %rd7, %rd25, 2;
add.s64 %rd26, %rd24, %rd5;
add.s64 %rd8, %rd26, 2;
add.s64 %rd27, %rd4, %rd21;
add.s64 %rd9, %rd27, 4;
cvta.to.global.u64 %rd10, %rd17;
mov.f32 %f309, 0f00000000;

$L__BB248_2:
cvt.u64.u32 %rd28, %r1866;
add.s64 %rd29, %rd1, %rd28;
add.s64 %rd30, %rd29, %rd3;
mul.lo.s64 %rd31, %rd30, 36;
add.s64 %rd32, %rd167, %rd6;
ld.global.nc.u16 %rs15, [%rd32+-4];
ld.global.nc.u16 %rs16, [%rd32+-2];
ld.global.nc.u16 %rs17, [%rd32];
add.s64 %rd33, %rd167, %rd8;
ld.global.nc.u16 %rs18, [%rd33];
cvt.u32.u16 %r768, %rs18;
ld.global.nc.u16 %rs19, [%rd33+2];
cvt.u32.u16 %r769, %rs19;
mov.b32 %r770, {%rs18, %rs19};
ld.global.nc.u16 %rs20, [%rd32+-6];
shl.b16 %rs21, %rs20, 2;
cvt.u64.u16 %rd34, %rs21;
and.b64 %rd35, %rd34, 1020;
mov.u64 %rd36, iq3xxs_grid;
add.s64 %rd37, %rd36, %rd35;
ld.global.nc.u32 %r771, [%rd37];
not.b32 %r772, %r771;
shr.u16 %rs22, %rs20, 8;
cvt.u32.u16 %r773, %rs22;
mul.wide.u32 %rd38, %r773, 4;
add.s64 %rd39, %rd36, %rd38;
ld.global.nc.u32 %r774, [%rd39];
not.b32 %r775, %r774;
and.b32 %r776, %r768, 255;
popc.b32 %r777, %r776;
shl.b32 %r778, %r777, 7;
and.b32 %r779, %r778, 128;
xor.b32 %r780, %r779, %r776;
mul.lo.s32 %r781, %r780, 16843009;
and.b32 %r782, %r781, 134480385;
or.b32 %r783, %r782, -2139062144;
add.s32 %r784, %r783, -16843009;
or.b32 %r670, %r784, %r782;
mov.u32 %r82, %r670;

prmt.b32 %r82, %r82, 0, 0xba98;

xor.b32 %r785, %r82, %r771;
or.b32 %r786, %r785, -2139062144;
and.b32 %r787, %r82, 2139062143;
and.b32 %r788, %r772, -2139062144;
sub.s32 %r789, %r786, %r787;
xor.b32 %r87, %r788, %r789;
add.s64 %rd40, %rd10, %rd31;
ld.global.nc.u32 %r88, [%rd40+4];
and.b32 %r790, %r781, -2143281136;
or.b32 %r791, %r790, -2139062144;
add.s32 %r792, %r791, -16843009;
or.b32 %r672, %r792, %r790;
mov.u32 %r84, %r672;

prmt.b32 %r84, %r84, 0, 0xba98;

xor.b32 %r793, %r84, %r774;
or.b32 %r794, %r793, -2139062144;
and.b32 %r795, %r84, 2139062143;
and.b32 %r796, %r775, -2139062144;
sub.s32 %r797, %r794, %r795;
xor.b32 %r91, %r796, %r797;
ld.global.nc.u32 %r92, [%rd40+8];
mov.u32 %r726, 0;

dp4a.s32.s32 %r86, %r87, %r88, %r726;


dp4a.s32.s32 %r90, %r91, %r92, %r86;

shl.b16 %rs23, %rs15, 2;
cvt.u64.u16 %rd41, %rs23;
and.b64 %rd42, %rd41, 1020;
add.s64 %rd43, %rd36, %rd42;
ld.global.nc.u32 %r798, [%rd43];
not.b32 %r799, %r798;
shr.u16 %rs24, %rs15, 8;
cvt.u32.u16 %r800, %rs24;
mul.wide.u32 %rd44, %r800, 4;
add.s64 %rd45, %rd36, %rd44;
ld.global.nc.u32 %r801, [%rd45];
not.b32 %r802, %r801;
shr.u32 %r803, %r768, 7;
and.b32 %r804, %r803, 255;
popc.b32 %r805, %r804;
shl.b32 %r806, %r805, 7;
and.b32 %r807, %r806, 128;
xor.b32 %r808, %r807, %r804;
mul.lo.s32 %r809, %r808, 16843009;
and.b32 %r810, %r809, 134480385;
or.b32 %r811, %r810, -2139062144;
add.s32 %r812, %r811, -16843009;
or.b32 %r682, %r812, %r810;
mov.u32 %r94, %r682;

prmt.b32 %r94, %r94, 0, 0xba98;

xor.b32 %r813, %r94, %r798;
or.b32 %r814, %r813, -2139062144;
and.b32 %r815, %r94, 2139062143;
and.b32 %r816, %r799, -2139062144;
sub.s32 %r817, %r814, %r815;
xor.b32 %r99, %r816, %r817;
ld.global.nc.u32 %r100, [%rd40+12];
and.b32 %r818, %r809, -2143281136;
or.b32 %r819, %r818, -2139062144;
add.s32 %r820, %r819, -16843009;
or.b32 %r684, %r820, %r818;
mov.u32 %r96, %r684;

prmt.b32 %r96, %r96, 0, 0xba98;

xor.b32 %r821, %r96, %r801;
or.b32 %r822, %r821, -2139062144;
and.b32 %r823, %r96, 2139062143;
and.b32 %r824, %r802, -2139062144;
sub.s32 %r825, %r822, %r823;
xor.b32 %r103, %r824, %r825;
ld.global.nc.u32 %r104, [%rd40+16];

dp4a.s32.s32 %r98, %r99, %r100, %r90;


dp4a.s32.s32 %r102, %r103, %r104, %r98;

shl.b16 %rs25, %rs16, 2;
cvt.u64.u16 %rd46, %rs25;
and.b64 %rd47, %rd46, 1020;
add.s64 %rd48, %rd36, %rd47;
ld.global.nc.u32 %r826, [%rd48];
not.b32 %r827, %r826;
shr.u16 %rs26, %rs16, 8;
cvt.u32.u16 %r828, %rs26;
mul.wide.u32 %rd49, %r828, 4;
add.s64 %rd50, %rd36, %rd49;
ld.global.nc.u32 %r829, [%rd50];
not.b32 %r830, %r829;
shr.u32 %r831, %r770, 14;
and.b32 %r832, %r831, 255;
popc.b32 %r833, %r832;
shl.b32 %r834, %r833, 7;
and.b32 %r835, %r834, 128;
xor.b32 %r836, %r835, %r832;
mul.lo.s32 %r837, %r836, 16843009;
and.b32 %r838, %r837, 134480385;
or.b32 %r839, %r838, -2139062144;
add.s32 %r840, %r839, -16843009;
or.b32 %r694, %r840, %r838;
mov.u32 %r106, %r694;

prmt.b32 %r106, %r106, 0, 0xba98;

xor.b32 %r841, %r106, %r826;
or.b32 %r842, %r841, -2139062144;
and.b32 %r843, %r106, 2139062143;
and.b32 %r844, %r827, -2139062144;
sub.s32 %r845, %r842, %r843;
xor.b32 %r111, %r844, %r845;
ld.global.nc.u32 %r112, [%rd40+20];
and.b32 %r846, %r837, -2143281136;
or.b32 %r847, %r846, -2139062144;
add.s32 %r848, %r847, -16843009;
or.b32 %r696, %r848, %r846;
mov.u32 %r108, %r696;

prmt.b32 %r108, %r108, 0, 0xba98;

xor.b32 %r849, %r108, %r829;
or.b32 %r850, %r849, -2139062144;
and.b32 %r851, %r108, 2139062143;
and.b32 %r852, %r830, -2139062144;
sub.s32 %r853, %r850, %r851;
xor.b32 %r115, %r852, %r853;
ld.global.nc.u32 %r116, [%rd40+24];

dp4a.s32.s32 %r110, %r111, %r112, %r102;


dp4a.s32.s32 %r114, %r115, %r116, %r110;

shl.b16 %rs27, %rs17, 2;
cvt.u64.u16 %rd51, %rs27;
and.b64 %rd52, %rd51, 1020;
add.s64 %rd53, %rd36, %rd52;
ld.global.nc.u32 %r854, [%rd53];
not.b32 %r855, %r854;
shr.u16 %rs28, %rs17, 8;
cvt.u32.u16 %r856, %rs28;
mul.wide.u32 %rd54, %r856, 4;
add.s64 %rd55, %rd36, %rd54;
ld.global.nc.u32 %r857, [%rd55];
not.b32 %r858, %r857;
shr.u32 %r859, %r769, 5;
and.b32 %r860, %r859, 255;
popc.b32 %r861, %r860;
shl.b32 %r862, %r861, 7;
and.b32 %r863, %r862, 128;
xor.b32 %r864, %r863, %r860;
mul.lo.s32 %r865, %r864, 16843009;
and.b32 %r866, %r865, 134480385;
or.b32 %r867, %r866, -2139062144;
add.s32 %r868, %r867, -16843009;
or.b32 %r706, %r868, %r866;
mov.u32 %r118, %r706;

prmt.b32 %r118, %r118, 0, 0xba98;

xor.b32 %r869, %r118, %r854;
or.b32 %r870, %r869, -2139062144;
and.b32 %r871, %r118, 2139062143;
and.b32 %r872, %r855, -2139062144;
sub.s32 %r873, %r870, %r871;
xor.b32 %r123, %r872, %r873;
ld.global.nc.u32 %r124, [%rd40+28];
and.b32 %r874, %r865, -2143281136;
or.b32 %r875, %r874, -2139062144;
add.s32 %r876, %r875, -16843009;
or.b32 %r708, %r876, %r874;
mov.u32 %r120, %r708;

prmt.b32 %r120, %r120, 0, 0xba98;

xor.b32 %r877, %r120, %r857;
or.b32 %r878, %r877, -2139062144;
and.b32 %r879, %r120, 2139062143;
and.b32 %r880, %r858, -2139062144;
sub.s32 %r881, %r878, %r879;
xor.b32 %r127, %r880, %r881;
ld.global.nc.u32 %r128, [%rd40+32];

dp4a.s32.s32 %r122, %r123, %r124, %r114;

shr.u32 %r882, %r769, 12;

dp4a.s32.s32 %r126, %r127, %r128, %r122;

shr.u32 %r883, %r126, 31;
add.s32 %r884, %r126, %r883;
shr.s32 %r885, %r884, 1;
mad.lo.s32 %r886, %r126, %r882, %r885;
shr.u32 %r887, %r886, 31;
add.s32 %r888, %r886, %r887;
shr.s32 %r889, %r888, 1;
add.s64 %rd56, %rd167, %rd5;
ld.global.nc.u16 %rs1, [%rd56];

{ cvt.f32.f16 %f85, %rs1;}


ld.global.nc.u32 %r130, [%rd40];

{.reg .f16 low,high;
mov.b32 {low,high},%r130;
cvt.f32.f16 %f86, low;}


mul.ftz.f32 %f113, %f85, %f86;
cvt.rn.f32.s32 %f114, %r889;
fma.rn.ftz.f32 %f322, %f113, %f114, %f322;
add.s64 %rd57, %rd167, %rd9;
ld.global.nc.u16 %rs29, [%rd57];
ld.global.nc.u16 %rs30, [%rd57+2];
ld.global.nc.u16 %rs31, [%rd57+4];
add.s64 %rd58, %rd167, %rd7;
ld.global.nc.u16 %rs32, [%rd58];
cvt.u32.u16 %r890, %rs32;
ld.global.nc.u16 %rs33, [%rd58+2];
cvt.u32.u16 %r891, %rs33;
mov.b32 %r892, {%rs32, %rs33};
ld.global.nc.u16 %rs34, [%rd57+-2];
shl.b16 %rs35, %rs34, 2;
cvt.u64.u16 %rd59, %rs35;
and.b64 %rd60, %rd59, 1020;
add.s64 %rd61, %rd36, %rd60;
ld.global.nc.u32 %r893, [%rd61];
not.b32 %r894, %r893;
shr.u16 %rs36, %rs34, 8;
cvt.u32.u16 %r895, %rs36;
mul.wide.u32 %rd62, %r895, 4;
add.s64 %rd63, %rd36, %rd62;
ld.global.nc.u32 %r896, [%rd63];
not.b32 %r897, %r896;
and.b32 %r898, %r890, 255;
popc.b32 %r899, %r898;
shl.b32 %r900, %r899, 7;
and.b32 %r901, %r900, 128;
xor.b32 %r902, %r901, %r898;
mul.lo.s32 %r903, %r902, 16843009;
and.b32 %r904, %r903, 134480385;
or.b32 %r905, %r904, -2139062144;
add.s32 %r906, %r905, -16843009;
or.b32 %r719, %r906, %r904;
mov.u32 %r131, %r719;

prmt.b32 %r131, %r131, 0, 0xba98;

xor.b32 %r907, %r131, %r893;
or.b32 %r908, %r907, -2139062144;
and.b32 %r909, %r131, 2139062143;
and.b32 %r910, %r894, -2139062144;
sub.s32 %r911, %r908, %r909;
xor.b32 %r136, %r910, %r911;
and.b32 %r912, %r903, -2143281136;
or.b32 %r913, %r912, -2139062144;
add.s32 %r914, %r913, -16843009;
or.b32 %r721, %r914, %r912;
mov.u32 %r133, %r721;

prmt.b32 %r133, %r133, 0, 0xba98;

xor.b32 %r915, %r133, %r896;
or.b32 %r916, %r915, -2139062144;
and.b32 %r917, %r133, 2139062143;
and.b32 %r918, %r897, -2139062144;
sub.s32 %r919, %r916, %r917;
xor.b32 %r140, %r918, %r919;

dp4a.s32.s32 %r135, %r136, %r88, %r726;


dp4a.s32.s32 %r139, %r140, %r92, %r135;

shl.b16 %rs37, %rs29, 2;
cvt.u64.u16 %rd64, %rs37;
and.b64 %rd65, %rd64, 1020;
add.s64 %rd66, %rd36, %rd65;
ld.global.nc.u32 %r920, [%rd66];
not.b32 %r921, %r920;
shr.u16 %rs38, %rs29, 8;
cvt.u32.u16 %r922, %rs38;
mul.wide.u32 %rd67, %r922, 4;
add.s64 %rd68, %rd36, %rd67;
ld.global.nc.u32 %r923, [%rd68];
not.b32 %r924, %r923;
shr.u32 %r925, %r890, 7;
and.b32 %r926, %r925, 255;
popc.b32 %r927, %r926;
shl.b32 %r928, %r927, 7;
and.b32 %r929, %r928, 128;
xor.b32 %r930, %r929, %r926;
mul.lo.s32 %r931, %r930, 16843009;
and.b32 %r932, %r931, 134480385;
or.b32 %r933, %r932, -2139062144;
add.s32 %r934, %r933, -16843009;
or.b32 %r731, %r934, %r932;
mov.u32 %r143, %r731;

prmt.b32 %r143, %r143, 0, 0xba98;

xor.b32 %r935, %r143, %r920;
or.b32 %r936, %r935, -2139062144;
and.b32 %r937, %r143, 2139062143;
and.b32 %r938, %r921, -2139062144;
sub.s32 %r939, %r936, %r937;
xor.b32 %r148, %r938, %r939;
and.b32 %r940, %r931, -2143281136;
or.b32 %r941, %r940, -2139062144;
add.s32 %r942, %r941, -16843009;
or.b32 %r733, %r942, %r940;
mov.u32 %r145, %r733;

prmt.b32 %r145, %r145, 0, 0xba98;

xor.b32 %r943, %r145, %r923;
or.b32 %r944, %r943, -2139062144;
and.b32 %r945, %r145, 2139062143;
and.b32 %r946, %r924, -2139062144;
sub.s32 %r947, %r944, %r945;
xor.b32 %r152, %r946, %r947;

dp4a.s32.s32 %r147, %r148, %r100, %r139;


dp4a.s32.s32 %r151, %r152, %r104, %r147;

shl.b16 %rs39, %rs30, 2;
cvt.u64.u16 %rd69, %rs39;
and.b64 %rd70, %rd69, 1020;
add.s64 %rd71, %rd36, %rd70;
ld.global.nc.u32 %r948, [%rd71];
not.b32 %r949, %r948;
shr.u16 %rs40, %rs30, 8;
cvt.u32.u16 %r950, %rs40;
mul.wide.u32 %rd72, %r950, 4;
add.s64 %rd73, %rd36, %rd72;
ld.global.nc.u32 %r951, [%rd73];
not.b32 %r952, %r951;
shr.u32 %r953, %r892, 14;
and.b32 %r954, %r953, 255;
popc.b32 %r955, %r954;
shl.b32 %r956, %r955, 7;
and.b32 %r957, %r956, 128;
xor.b32 %r958, %r957, %r954;
mul.lo.s32 %r959, %r958, 16843009;
and.b32 %r960, %r959, 134480385;
or.b32 %r961, %r960, -2139062144;
add.s32 %r962, %r961, -16843009;
or.b32 %r743, %r962, %r960;
mov.u32 %r155, %r743;

prmt.b32 %r155, %r155, 0, 0xba98;

xor.b32 %r963, %r155, %r948;
or.b32 %r964, %r963, -2139062144;
and.b32 %r965, %r155, 2139062143;
and.b32 %r966, %r949, -2139062144;
sub.s32 %r967, %r964, %r965;
xor.b32 %r160, %r966, %r967;
and.b32 %r968, %r959, -2143281136;
or.b32 %r969, %r968, -2139062144;
add.s32 %r970, %r969, -16843009;
or.b32 %r745, %r970, %r968;
mov.u32 %r157, %r745;

prmt.b32 %r157, %r157, 0, 0xba98;

xor.b32 %r971, %r157, %r951;
or.b32 %r972, %r971, -2139062144;
and.b32 %r973, %r157, 2139062143;
and.b32 %r974, %r952, -2139062144;
sub.s32 %r975, %r972, %r973;
xor.b32 %r164, %r974, %r975;

dp4a.s32.s32 %r159, %r160, %r112, %r151;


dp4a.s32.s32 %r163, %r164, %r116, %r159;

shl.b16 %rs41, %rs31, 2;
cvt.u64.u16 %rd74, %rs41;
and.b64 %rd75, %rd74, 1020;
add.s64 %rd76, %rd36, %rd75;
ld.global.nc.u32 %r976, [%rd76];
not.b32 %r977, %r976;
shr.u16 %rs42, %rs31, 8;
cvt.u32.u16 %r978, %rs42;
mul.wide.u32 %rd77, %r978, 4;
add.s64 %rd78, %rd36, %rd77;
ld.global.nc.u32 %r979, [%rd78];
not.b32 %r980, %r979;
shr.u32 %r981, %r891, 5;
and.b32 %r982, %r981, 255;
popc.b32 %r983, %r982;
shl.b32 %r984, %r983, 7;
and.b32 %r985, %r984, 128;
xor.b32 %r986, %r985, %r982;
mul.lo.s32 %r987, %r986, 16843009;
and.b32 %r988, %r987, 134480385;
or.b32 %r989, %r988, -2139062144;
add.s32 %r990, %r989, -16843009;
or.b32 %r755, %r990, %r988;
mov.u32 %r167, %r755;

prmt.b32 %r167, %r167, 0, 0xba98;

xor.b32 %r991, %r167, %r976;
or.b32 %r992, %r991, -2139062144;
and.b32 %r993, %r167, 2139062143;
and.b32 %r994, %r977, -2139062144;
sub.s32 %r995, %r992, %r993;
xor.b32 %r172, %r994, %r995;
and.b32 %r996, %r987, -2143281136;
or.b32 %r997, %r996, -2139062144;
add.s32 %r998, %r997, -16843009;
or.b32 %r757, %r998, %r996;
mov.u32 %r169, %r757;

prmt.b32 %r169, %r169, 0, 0xba98;

xor.b32 %r999, %r169, %r979;
or.b32 %r1000, %r999, -2139062144;
and.b32 %r1001, %r169, 2139062143;
and.b32 %r1002, %r980, -2139062144;
sub.s32 %r1003, %r1000, %r1001;
xor.b32 %r176, %r1002, %r1003;

dp4a.s32.s32 %r171, %r172, %r124, %r163;

shr.u32 %r1004, %r891, 12;

dp4a.s32.s32 %r175, %r176, %r128, %r171;

shr.u32 %r1005, %r175, 31;
add.s32 %r1006, %r175, %r1005;
shr.s32 %r1007, %r1006, 1;
mad.lo.s32 %r1008, %r175, %r1004, %r1007;
shr.u32 %r1009, %r1008, 31;
add.s32 %r1010, %r1008, %r1009;
shr.s32 %r1011, %r1010, 1;
add.s64 %rd79, %rd167, %rd4;
ld.global.nc.u16 %rs2, [%rd79];

{ cvt.f32.f16 %f87, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r130;
cvt.f32.f16 %f88, low;}


mul.ftz.f32 %f115, %f87, %f88;
cvt.rn.f32.s32 %f116, %r1011;
fma.rn.ftz.f32 %f321, %f115, %f116, %f321;
cvt.u64.u32 %rd80, %r1867;
add.s64 %rd81, %rd1, %rd80;
add.s64 %rd82, %rd81, %rd3;
mul.lo.s64 %rd83, %rd82, 36;
mov.u32 %r180, %r670;

prmt.b32 %r180, %r180, 0, 0xba98;

xor.b32 %r1012, %r180, %r771;
or.b32 %r1013, %r1012, -2139062144;
and.b32 %r1014, %r180, 2139062143;
sub.s32 %r1015, %r1013, %r1014;
xor.b32 %r185, %r788, %r1015;
add.s64 %rd84, %rd10, %rd83;
ld.global.nc.u32 %r186, [%rd84+4];
mov.u32 %r182, %r672;

prmt.b32 %r182, %r182, 0, 0xba98;

xor.b32 %r1016, %r182, %r774;
or.b32 %r1017, %r1016, -2139062144;
and.b32 %r1018, %r182, 2139062143;
sub.s32 %r1019, %r1017, %r1018;
xor.b32 %r189, %r796, %r1019;
ld.global.nc.u32 %r190, [%rd84+8];

dp4a.s32.s32 %r184, %r185, %r186, %r726;


dp4a.s32.s32 %r188, %r189, %r190, %r184;

mov.u32 %r192, %r682;

prmt.b32 %r192, %r192, 0, 0xba98;

xor.b32 %r1020, %r192, %r798;
or.b32 %r1021, %r1020, -2139062144;
and.b32 %r1022, %r192, 2139062143;
sub.s32 %r1023, %r1021, %r1022;
xor.b32 %r197, %r816, %r1023;
ld.global.nc.u32 %r198, [%rd84+12];
mov.u32 %r194, %r684;

prmt.b32 %r194, %r194, 0, 0xba98;

xor.b32 %r1024, %r194, %r801;
or.b32 %r1025, %r1024, -2139062144;
and.b32 %r1026, %r194, 2139062143;
sub.s32 %r1027, %r1025, %r1026;
xor.b32 %r201, %r824, %r1027;
ld.global.nc.u32 %r202, [%rd84+16];

dp4a.s32.s32 %r196, %r197, %r198, %r188;


dp4a.s32.s32 %r200, %r201, %r202, %r196;

mov.u32 %r204, %r694;

prmt.b32 %r204, %r204, 0, 0xba98;

xor.b32 %r1028, %r204, %r826;
or.b32 %r1029, %r1028, -2139062144;
and.b32 %r1030, %r204, 2139062143;
sub.s32 %r1031, %r1029, %r1030;
xor.b32 %r209, %r844, %r1031;
ld.global.nc.u32 %r210, [%rd84+20];
mov.u32 %r206, %r696;

prmt.b32 %r206, %r206, 0, 0xba98;

xor.b32 %r1032, %r206, %r829;
or.b32 %r1033, %r1032, -2139062144;
and.b32 %r1034, %r206, 2139062143;
sub.s32 %r1035, %r1033, %r1034;
xor.b32 %r213, %r852, %r1035;
ld.global.nc.u32 %r214, [%rd84+24];

dp4a.s32.s32 %r208, %r209, %r210, %r200;


dp4a.s32.s32 %r212, %r213, %r214, %r208;

mov.u32 %r216, %r706;

prmt.b32 %r216, %r216, 0, 0xba98;

xor.b32 %r1036, %r216, %r854;
or.b32 %r1037, %r1036, -2139062144;
and.b32 %r1038, %r216, 2139062143;
sub.s32 %r1039, %r1037, %r1038;
xor.b32 %r221, %r872, %r1039;
ld.global.nc.u32 %r222, [%rd84+28];
mov.u32 %r218, %r708;

prmt.b32 %r218, %r218, 0, 0xba98;

xor.b32 %r1040, %r218, %r857;
or.b32 %r1041, %r1040, -2139062144;
and.b32 %r1042, %r218, 2139062143;
sub.s32 %r1043, %r1041, %r1042;
xor.b32 %r225, %r880, %r1043;
ld.global.nc.u32 %r226, [%rd84+32];

dp4a.s32.s32 %r220, %r221, %r222, %r212;


dp4a.s32.s32 %r224, %r225, %r226, %r220;

shr.u32 %r1044, %r224, 31;
add.s32 %r1045, %r224, %r1044;
shr.s32 %r1046, %r1045, 1;
mad.lo.s32 %r1047, %r224, %r882, %r1046;
shr.u32 %r1048, %r1047, 31;
add.s32 %r1049, %r1047, %r1048;
shr.s32 %r1050, %r1049, 1;

{ cvt.f32.f16 %f89, %rs1;}


ld.global.nc.u32 %r228, [%rd84];

{.reg .f16 low,high;
mov.b32 {low,high},%r228;
cvt.f32.f16 %f90, low;}


mul.ftz.f32 %f117, %f89, %f90;
cvt.rn.f32.s32 %f118, %r1050;
fma.rn.ftz.f32 %f320, %f117, %f118, %f320;
mov.u32 %r229, %r719;

prmt.b32 %r229, %r229, 0, 0xba98;

xor.b32 %r1051, %r229, %r893;
or.b32 %r1052, %r1051, -2139062144;
and.b32 %r1053, %r229, 2139062143;
sub.s32 %r1054, %r1052, %r1053;
xor.b32 %r234, %r910, %r1054;
mov.u32 %r231, %r721;

prmt.b32 %r231, %r231, 0, 0xba98;

xor.b32 %r1055, %r231, %r896;
or.b32 %r1056, %r1055, -2139062144;
and.b32 %r1057, %r231, 2139062143;
sub.s32 %r1058, %r1056, %r1057;
xor.b32 %r238, %r918, %r1058;

dp4a.s32.s32 %r233, %r234, %r186, %r726;


dp4a.s32.s32 %r237, %r238, %r190, %r233;

mov.u32 %r241, %r731;

prmt.b32 %r241, %r241, 0, 0xba98;

xor.b32 %r1059, %r241, %r920;
or.b32 %r1060, %r1059, -2139062144;
and.b32 %r1061, %r241, 2139062143;
sub.s32 %r1062, %r1060, %r1061;
xor.b32 %r246, %r938, %r1062;
mov.u32 %r243, %r733;

prmt.b32 %r243, %r243, 0, 0xba98;

xor.b32 %r1063, %r243, %r923;
or.b32 %r1064, %r1063, -2139062144;
and.b32 %r1065, %r243, 2139062143;
sub.s32 %r1066, %r1064, %r1065;
xor.b32 %r250, %r946, %r1066;

dp4a.s32.s32 %r245, %r246, %r198, %r237;


dp4a.s32.s32 %r249, %r250, %r202, %r245;

mov.u32 %r253, %r743;

prmt.b32 %r253, %r253, 0, 0xba98;

xor.b32 %r1067, %r253, %r948;
or.b32 %r1068, %r1067, -2139062144;
and.b32 %r1069, %r253, 2139062143;
sub.s32 %r1070, %r1068, %r1069;
xor.b32 %r258, %r966, %r1070;
mov.u32 %r255, %r745;

prmt.b32 %r255, %r255, 0, 0xba98;

xor.b32 %r1071, %r255, %r951;
or.b32 %r1072, %r1071, -2139062144;
and.b32 %r1073, %r255, 2139062143;
sub.s32 %r1074, %r1072, %r1073;
xor.b32 %r262, %r974, %r1074;

dp4a.s32.s32 %r257, %r258, %r210, %r249;


dp4a.s32.s32 %r261, %r262, %r214, %r257;

mov.u32 %r265, %r755;

prmt.b32 %r265, %r265, 0, 0xba98;

xor.b32 %r1075, %r265, %r976;
or.b32 %r1076, %r1075, -2139062144;
and.b32 %r1077, %r265, 2139062143;
sub.s32 %r1078, %r1076, %r1077;
xor.b32 %r270, %r994, %r1078;
mov.u32 %r267, %r757;

prmt.b32 %r267, %r267, 0, 0xba98;

xor.b32 %r1079, %r267, %r979;
or.b32 %r1080, %r1079, -2139062144;
and.b32 %r1081, %r267, 2139062143;
sub.s32 %r1082, %r1080, %r1081;
xor.b32 %r274, %r1002, %r1082;

dp4a.s32.s32 %r269, %r270, %r222, %r261;


dp4a.s32.s32 %r273, %r274, %r226, %r269;

shr.u32 %r1083, %r273, 31;
add.s32 %r1084, %r273, %r1083;
shr.s32 %r1085, %r1084, 1;
mad.lo.s32 %r1086, %r273, %r1004, %r1085;
shr.u32 %r1087, %r1086, 31;
add.s32 %r1088, %r1086, %r1087;
shr.s32 %r1089, %r1088, 1;

{ cvt.f32.f16 %f91, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r228;
cvt.f32.f16 %f92, low;}


mul.ftz.f32 %f119, %f91, %f92;
cvt.rn.f32.s32 %f120, %r1089;
fma.rn.ftz.f32 %f319, %f119, %f120, %f319;
cvt.u64.u32 %rd85, %r1872;
add.s64 %rd86, %rd1, %rd85;
add.s64 %rd87, %rd86, %rd3;
mul.lo.s64 %rd88, %rd87, 36;
mov.u32 %r278, %r670;

prmt.b32 %r278, %r278, 0, 0xba98;

xor.b32 %r1090, %r278, %r771;
or.b32 %r1091, %r1090, -2139062144;
and.b32 %r1092, %r278, 2139062143;
sub.s32 %r1093, %r1091, %r1092;
xor.b32 %r283, %r788, %r1093;
add.s64 %rd89, %rd10, %rd88;
ld.global.nc.u32 %r284, [%rd89+4];
mov.u32 %r280, %r672;

prmt.b32 %r280, %r280, 0, 0xba98;

xor.b32 %r1094, %r280, %r774;
or.b32 %r1095, %r1094, -2139062144;
and.b32 %r1096, %r280, 2139062143;
sub.s32 %r1097, %r1095, %r1096;
xor.b32 %r287, %r796, %r1097;
ld.global.nc.u32 %r288, [%rd89+8];

dp4a.s32.s32 %r282, %r283, %r284, %r726;


dp4a.s32.s32 %r286, %r287, %r288, %r282;

mov.u32 %r290, %r682;

prmt.b32 %r290, %r290, 0, 0xba98;

xor.b32 %r1098, %r290, %r798;
or.b32 %r1099, %r1098, -2139062144;
and.b32 %r1100, %r290, 2139062143;
sub.s32 %r1101, %r1099, %r1100;
xor.b32 %r295, %r816, %r1101;
ld.global.nc.u32 %r296, [%rd89+12];
mov.u32 %r292, %r684;

prmt.b32 %r292, %r292, 0, 0xba98;

xor.b32 %r1102, %r292, %r801;
or.b32 %r1103, %r1102, -2139062144;
and.b32 %r1104, %r292, 2139062143;
sub.s32 %r1105, %r1103, %r1104;
xor.b32 %r299, %r824, %r1105;
ld.global.nc.u32 %r300, [%rd89+16];

dp4a.s32.s32 %r294, %r295, %r296, %r286;


dp4a.s32.s32 %r298, %r299, %r300, %r294;

mov.u32 %r302, %r694;

prmt.b32 %r302, %r302, 0, 0xba98;

xor.b32 %r1106, %r302, %r826;
or.b32 %r1107, %r1106, -2139062144;
and.b32 %r1108, %r302, 2139062143;
sub.s32 %r1109, %r1107, %r1108;
xor.b32 %r307, %r844, %r1109;
ld.global.nc.u32 %r308, [%rd89+20];
mov.u32 %r304, %r696;

prmt.b32 %r304, %r304, 0, 0xba98;

xor.b32 %r1110, %r304, %r829;
or.b32 %r1111, %r1110, -2139062144;
and.b32 %r1112, %r304, 2139062143;
sub.s32 %r1113, %r1111, %r1112;
xor.b32 %r311, %r852, %r1113;
ld.global.nc.u32 %r312, [%rd89+24];

dp4a.s32.s32 %r306, %r307, %r308, %r298;


dp4a.s32.s32 %r310, %r311, %r312, %r306;

mov.u32 %r314, %r706;

prmt.b32 %r314, %r314, 0, 0xba98;

xor.b32 %r1114, %r314, %r854;
or.b32 %r1115, %r1114, -2139062144;
and.b32 %r1116, %r314, 2139062143;
sub.s32 %r1117, %r1115, %r1116;
xor.b32 %r319, %r872, %r1117;
ld.global.nc.u32 %r320, [%rd89+28];
mov.u32 %r316, %r708;

prmt.b32 %r316, %r316, 0, 0xba98;

xor.b32 %r1118, %r316, %r857;
or.b32 %r1119, %r1118, -2139062144;
and.b32 %r1120, %r316, 2139062143;
sub.s32 %r1121, %r1119, %r1120;
xor.b32 %r323, %r880, %r1121;
ld.global.nc.u32 %r324, [%rd89+32];

dp4a.s32.s32 %r318, %r319, %r320, %r310;


dp4a.s32.s32 %r322, %r323, %r324, %r318;

shr.u32 %r1122, %r322, 31;
add.s32 %r1123, %r322, %r1122;
shr.s32 %r1124, %r1123, 1;
mad.lo.s32 %r1125, %r322, %r882, %r1124;
shr.u32 %r1126, %r1125, 31;
add.s32 %r1127, %r1125, %r1126;
shr.s32 %r1128, %r1127, 1;

{ cvt.f32.f16 %f93, %rs1;}


ld.global.nc.u32 %r326, [%rd89];

{.reg .f16 low,high;
mov.b32 {low,high},%r326;
cvt.f32.f16 %f94, low;}


mul.ftz.f32 %f121, %f93, %f94;
cvt.rn.f32.s32 %f122, %r1128;
fma.rn.ftz.f32 %f318, %f121, %f122, %f318;
mov.u32 %r327, %r719;

prmt.b32 %r327, %r327, 0, 0xba98;

xor.b32 %r1129, %r327, %r893;
or.b32 %r1130, %r1129, -2139062144;
and.b32 %r1131, %r327, 2139062143;
sub.s32 %r1132, %r1130, %r1131;
xor.b32 %r332, %r910, %r1132;
mov.u32 %r329, %r721;

prmt.b32 %r329, %r329, 0, 0xba98;

xor.b32 %r1133, %r329, %r896;
or.b32 %r1134, %r1133, -2139062144;
and.b32 %r1135, %r329, 2139062143;
sub.s32 %r1136, %r1134, %r1135;
xor.b32 %r336, %r918, %r1136;

dp4a.s32.s32 %r331, %r332, %r284, %r726;


dp4a.s32.s32 %r335, %r336, %r288, %r331;

mov.u32 %r339, %r731;

prmt.b32 %r339, %r339, 0, 0xba98;

xor.b32 %r1137, %r339, %r920;
or.b32 %r1138, %r1137, -2139062144;
and.b32 %r1139, %r339, 2139062143;
sub.s32 %r1140, %r1138, %r1139;
xor.b32 %r344, %r938, %r1140;
mov.u32 %r341, %r733;

prmt.b32 %r341, %r341, 0, 0xba98;

xor.b32 %r1141, %r341, %r923;
or.b32 %r1142, %r1141, -2139062144;
and.b32 %r1143, %r341, 2139062143;
sub.s32 %r1144, %r1142, %r1143;
xor.b32 %r348, %r946, %r1144;

dp4a.s32.s32 %r343, %r344, %r296, %r335;


dp4a.s32.s32 %r347, %r348, %r300, %r343;

mov.u32 %r351, %r743;

prmt.b32 %r351, %r351, 0, 0xba98;

xor.b32 %r1145, %r351, %r948;
or.b32 %r1146, %r1145, -2139062144;
and.b32 %r1147, %r351, 2139062143;
sub.s32 %r1148, %r1146, %r1147;
xor.b32 %r356, %r966, %r1148;
mov.u32 %r353, %r745;

prmt.b32 %r353, %r353, 0, 0xba98;

xor.b32 %r1149, %r353, %r951;
or.b32 %r1150, %r1149, -2139062144;
and.b32 %r1151, %r353, 2139062143;
sub.s32 %r1152, %r1150, %r1151;
xor.b32 %r360, %r974, %r1152;

dp4a.s32.s32 %r355, %r356, %r308, %r347;


dp4a.s32.s32 %r359, %r360, %r312, %r355;

mov.u32 %r363, %r755;

prmt.b32 %r363, %r363, 0, 0xba98;

xor.b32 %r1153, %r363, %r976;
or.b32 %r1154, %r1153, -2139062144;
and.b32 %r1155, %r363, 2139062143;
sub.s32 %r1156, %r1154, %r1155;
xor.b32 %r368, %r994, %r1156;
mov.u32 %r365, %r757;

prmt.b32 %r365, %r365, 0, 0xba98;

xor.b32 %r1157, %r365, %r979;
or.b32 %r1158, %r1157, -2139062144;
and.b32 %r1159, %r365, 2139062143;
sub.s32 %r1160, %r1158, %r1159;
xor.b32 %r372, %r1002, %r1160;

dp4a.s32.s32 %r367, %r368, %r320, %r359;


dp4a.s32.s32 %r371, %r372, %r324, %r367;

shr.u32 %r1161, %r371, 31;
add.s32 %r1162, %r371, %r1161;
shr.s32 %r1163, %r1162, 1;
mad.lo.s32 %r1164, %r371, %r1004, %r1163;
shr.u32 %r1165, %r1164, 31;
add.s32 %r1166, %r1164, %r1165;
shr.s32 %r1167, %r1166, 1;

{ cvt.f32.f16 %f95, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r326;
cvt.f32.f16 %f96, low;}


mul.ftz.f32 %f123, %f95, %f96;
cvt.rn.f32.s32 %f124, %r1167;
fma.rn.ftz.f32 %f317, %f123, %f124, %f317;
cvt.u64.u32 %rd90, %r1871;
add.s64 %rd91, %rd1, %rd90;
add.s64 %rd92, %rd91, %rd3;
mul.lo.s64 %rd93, %rd92, 36;
mov.u32 %r376, %r670;

prmt.b32 %r376, %r376, 0, 0xba98;

xor.b32 %r1168, %r376, %r771;
or.b32 %r1169, %r1168, -2139062144;
and.b32 %r1170, %r376, 2139062143;
sub.s32 %r1171, %r1169, %r1170;
xor.b32 %r381, %r788, %r1171;
add.s64 %rd94, %rd10, %rd93;
ld.global.nc.u32 %r382, [%rd94+4];
mov.u32 %r378, %r672;

prmt.b32 %r378, %r378, 0, 0xba98;

xor.b32 %r1172, %r378, %r774;
or.b32 %r1173, %r1172, -2139062144;
and.b32 %r1174, %r378, 2139062143;
sub.s32 %r1175, %r1173, %r1174;
xor.b32 %r385, %r796, %r1175;
ld.global.nc.u32 %r386, [%rd94+8];

dp4a.s32.s32 %r380, %r381, %r382, %r726;


dp4a.s32.s32 %r384, %r385, %r386, %r380;

mov.u32 %r388, %r682;

prmt.b32 %r388, %r388, 0, 0xba98;

xor.b32 %r1176, %r388, %r798;
or.b32 %r1177, %r1176, -2139062144;
and.b32 %r1178, %r388, 2139062143;
sub.s32 %r1179, %r1177, %r1178;
xor.b32 %r393, %r816, %r1179;
ld.global.nc.u32 %r394, [%rd94+12];
mov.u32 %r390, %r684;

prmt.b32 %r390, %r390, 0, 0xba98;

xor.b32 %r1180, %r390, %r801;
or.b32 %r1181, %r1180, -2139062144;
and.b32 %r1182, %r390, 2139062143;
sub.s32 %r1183, %r1181, %r1182;
xor.b32 %r397, %r824, %r1183;
ld.global.nc.u32 %r398, [%rd94+16];

dp4a.s32.s32 %r392, %r393, %r394, %r384;


dp4a.s32.s32 %r396, %r397, %r398, %r392;

mov.u32 %r400, %r694;

prmt.b32 %r400, %r400, 0, 0xba98;

xor.b32 %r1184, %r400, %r826;
or.b32 %r1185, %r1184, -2139062144;
and.b32 %r1186, %r400, 2139062143;
sub.s32 %r1187, %r1185, %r1186;
xor.b32 %r405, %r844, %r1187;
ld.global.nc.u32 %r406, [%rd94+20];
mov.u32 %r402, %r696;

prmt.b32 %r402, %r402, 0, 0xba98;

xor.b32 %r1188, %r402, %r829;
or.b32 %r1189, %r1188, -2139062144;
and.b32 %r1190, %r402, 2139062143;
sub.s32 %r1191, %r1189, %r1190;
xor.b32 %r409, %r852, %r1191;
ld.global.nc.u32 %r410, [%rd94+24];

dp4a.s32.s32 %r404, %r405, %r406, %r396;


dp4a.s32.s32 %r408, %r409, %r410, %r404;

mov.u32 %r412, %r706;

prmt.b32 %r412, %r412, 0, 0xba98;

xor.b32 %r1192, %r412, %r854;
or.b32 %r1193, %r1192, -2139062144;
and.b32 %r1194, %r412, 2139062143;
sub.s32 %r1195, %r1193, %r1194;
xor.b32 %r417, %r872, %r1195;
ld.global.nc.u32 %r418, [%rd94+28];
mov.u32 %r414, %r708;

prmt.b32 %r414, %r414, 0, 0xba98;

xor.b32 %r1196, %r414, %r857;
or.b32 %r1197, %r1196, -2139062144;
and.b32 %r1198, %r414, 2139062143;
sub.s32 %r1199, %r1197, %r1198;
xor.b32 %r421, %r880, %r1199;
ld.global.nc.u32 %r422, [%rd94+32];

dp4a.s32.s32 %r416, %r417, %r418, %r408;


dp4a.s32.s32 %r420, %r421, %r422, %r416;

shr.u32 %r1200, %r420, 31;
add.s32 %r1201, %r420, %r1200;
shr.s32 %r1202, %r1201, 1;
mad.lo.s32 %r1203, %r420, %r882, %r1202;
shr.u32 %r1204, %r1203, 31;
add.s32 %r1205, %r1203, %r1204;
shr.s32 %r1206, %r1205, 1;

{ cvt.f32.f16 %f97, %rs1;}


ld.global.nc.u32 %r424, [%rd94];

{.reg .f16 low,high;
mov.b32 {low,high},%r424;
cvt.f32.f16 %f98, low;}


mul.ftz.f32 %f125, %f97, %f98;
cvt.rn.f32.s32 %f126, %r1206;
fma.rn.ftz.f32 %f316, %f125, %f126, %f316;
mov.u32 %r425, %r719;

prmt.b32 %r425, %r425, 0, 0xba98;

xor.b32 %r1207, %r425, %r893;
or.b32 %r1208, %r1207, -2139062144;
and.b32 %r1209, %r425, 2139062143;
sub.s32 %r1210, %r1208, %r1209;
xor.b32 %r430, %r910, %r1210;
mov.u32 %r427, %r721;

prmt.b32 %r427, %r427, 0, 0xba98;

xor.b32 %r1211, %r427, %r896;
or.b32 %r1212, %r1211, -2139062144;
and.b32 %r1213, %r427, 2139062143;
sub.s32 %r1214, %r1212, %r1213;
xor.b32 %r434, %r918, %r1214;

dp4a.s32.s32 %r429, %r430, %r382, %r726;


dp4a.s32.s32 %r433, %r434, %r386, %r429;

mov.u32 %r437, %r731;

prmt.b32 %r437, %r437, 0, 0xba98;

xor.b32 %r1215, %r437, %r920;
or.b32 %r1216, %r1215, -2139062144;
and.b32 %r1217, %r437, 2139062143;
sub.s32 %r1218, %r1216, %r1217;
xor.b32 %r442, %r938, %r1218;
mov.u32 %r439, %r733;

prmt.b32 %r439, %r439, 0, 0xba98;

xor.b32 %r1219, %r439, %r923;
or.b32 %r1220, %r1219, -2139062144;
and.b32 %r1221, %r439, 2139062143;
sub.s32 %r1222, %r1220, %r1221;
xor.b32 %r446, %r946, %r1222;

dp4a.s32.s32 %r441, %r442, %r394, %r433;


dp4a.s32.s32 %r445, %r446, %r398, %r441;

mov.u32 %r449, %r743;

prmt.b32 %r449, %r449, 0, 0xba98;

xor.b32 %r1223, %r449, %r948;
or.b32 %r1224, %r1223, -2139062144;
and.b32 %r1225, %r449, 2139062143;
sub.s32 %r1226, %r1224, %r1225;
xor.b32 %r454, %r966, %r1226;
mov.u32 %r451, %r745;

prmt.b32 %r451, %r451, 0, 0xba98;

xor.b32 %r1227, %r451, %r951;
or.b32 %r1228, %r1227, -2139062144;
and.b32 %r1229, %r451, 2139062143;
sub.s32 %r1230, %r1228, %r1229;
xor.b32 %r458, %r974, %r1230;

dp4a.s32.s32 %r453, %r454, %r406, %r445;


dp4a.s32.s32 %r457, %r458, %r410, %r453;

mov.u32 %r461, %r755;

prmt.b32 %r461, %r461, 0, 0xba98;

xor.b32 %r1231, %r461, %r976;
or.b32 %r1232, %r1231, -2139062144;
and.b32 %r1233, %r461, 2139062143;
sub.s32 %r1234, %r1232, %r1233;
xor.b32 %r466, %r994, %r1234;
mov.u32 %r463, %r757;

prmt.b32 %r463, %r463, 0, 0xba98;

xor.b32 %r1235, %r463, %r979;
or.b32 %r1236, %r1235, -2139062144;
and.b32 %r1237, %r463, 2139062143;
sub.s32 %r1238, %r1236, %r1237;
xor.b32 %r470, %r1002, %r1238;

dp4a.s32.s32 %r465, %r466, %r418, %r457;


dp4a.s32.s32 %r469, %r470, %r422, %r465;

shr.u32 %r1239, %r469, 31;
add.s32 %r1240, %r469, %r1239;
shr.s32 %r1241, %r1240, 1;
mad.lo.s32 %r1242, %r469, %r1004, %r1241;
shr.u32 %r1243, %r1242, 31;
add.s32 %r1244, %r1242, %r1243;
shr.s32 %r1245, %r1244, 1;

{ cvt.f32.f16 %f99, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r424;
cvt.f32.f16 %f100, low;}


mul.ftz.f32 %f127, %f99, %f100;
cvt.rn.f32.s32 %f128, %r1245;
fma.rn.ftz.f32 %f315, %f127, %f128, %f315;
cvt.u64.u32 %rd95, %r1870;
add.s64 %rd96, %rd1, %rd95;
add.s64 %rd97, %rd96, %rd3;
mul.lo.s64 %rd98, %rd97, 36;
mov.u32 %r474, %r670;

prmt.b32 %r474, %r474, 0, 0xba98;

xor.b32 %r1246, %r474, %r771;
or.b32 %r1247, %r1246, -2139062144;
and.b32 %r1248, %r474, 2139062143;
sub.s32 %r1249, %r1247, %r1248;
xor.b32 %r479, %r788, %r1249;
add.s64 %rd99, %rd10, %rd98;
ld.global.nc.u32 %r480, [%rd99+4];
mov.u32 %r476, %r672;

prmt.b32 %r476, %r476, 0, 0xba98;

xor.b32 %r1250, %r476, %r774;
or.b32 %r1251, %r1250, -2139062144;
and.b32 %r1252, %r476, 2139062143;
sub.s32 %r1253, %r1251, %r1252;
xor.b32 %r483, %r796, %r1253;
ld.global.nc.u32 %r484, [%rd99+8];

dp4a.s32.s32 %r478, %r479, %r480, %r726;


dp4a.s32.s32 %r482, %r483, %r484, %r478;

mov.u32 %r486, %r682;

prmt.b32 %r486, %r486, 0, 0xba98;

xor.b32 %r1254, %r486, %r798;
or.b32 %r1255, %r1254, -2139062144;
and.b32 %r1256, %r486, 2139062143;
sub.s32 %r1257, %r1255, %r1256;
xor.b32 %r491, %r816, %r1257;
ld.global.nc.u32 %r492, [%rd99+12];
mov.u32 %r488, %r684;

prmt.b32 %r488, %r488, 0, 0xba98;

xor.b32 %r1258, %r488, %r801;
or.b32 %r1259, %r1258, -2139062144;
and.b32 %r1260, %r488, 2139062143;
sub.s32 %r1261, %r1259, %r1260;
xor.b32 %r495, %r824, %r1261;
ld.global.nc.u32 %r496, [%rd99+16];

dp4a.s32.s32 %r490, %r491, %r492, %r482;


dp4a.s32.s32 %r494, %r495, %r496, %r490;

mov.u32 %r498, %r694;

prmt.b32 %r498, %r498, 0, 0xba98;

xor.b32 %r1262, %r498, %r826;
or.b32 %r1263, %r1262, -2139062144;
and.b32 %r1264, %r498, 2139062143;
sub.s32 %r1265, %r1263, %r1264;
xor.b32 %r503, %r844, %r1265;
ld.global.nc.u32 %r504, [%rd99+20];
mov.u32 %r500, %r696;

prmt.b32 %r500, %r500, 0, 0xba98;

xor.b32 %r1266, %r500, %r829;
or.b32 %r1267, %r1266, -2139062144;
and.b32 %r1268, %r500, 2139062143;
sub.s32 %r1269, %r1267, %r1268;
xor.b32 %r507, %r852, %r1269;
ld.global.nc.u32 %r508, [%rd99+24];

dp4a.s32.s32 %r502, %r503, %r504, %r494;


dp4a.s32.s32 %r506, %r507, %r508, %r502;

mov.u32 %r510, %r706;

prmt.b32 %r510, %r510, 0, 0xba98;

xor.b32 %r1270, %r510, %r854;
or.b32 %r1271, %r1270, -2139062144;
and.b32 %r1272, %r510, 2139062143;
sub.s32 %r1273, %r1271, %r1272;
xor.b32 %r515, %r872, %r1273;
ld.global.nc.u32 %r516, [%rd99+28];
mov.u32 %r512, %r708;

prmt.b32 %r512, %r512, 0, 0xba98;

xor.b32 %r1274, %r512, %r857;
or.b32 %r1275, %r1274, -2139062144;
and.b32 %r1276, %r512, 2139062143;
sub.s32 %r1277, %r1275, %r1276;
xor.b32 %r519, %r880, %r1277;
ld.global.nc.u32 %r520, [%rd99+32];

dp4a.s32.s32 %r514, %r515, %r516, %r506;


dp4a.s32.s32 %r518, %r519, %r520, %r514;

shr.u32 %r1278, %r518, 31;
add.s32 %r1279, %r518, %r1278;
shr.s32 %r1280, %r1279, 1;
mad.lo.s32 %r1281, %r518, %r882, %r1280;
shr.u32 %r1282, %r1281, 31;
add.s32 %r1283, %r1281, %r1282;
shr.s32 %r1284, %r1283, 1;

{ cvt.f32.f16 %f101, %rs1;}


ld.global.nc.u32 %r522, [%rd99];

{.reg .f16 low,high;
mov.b32 {low,high},%r522;
cvt.f32.f16 %f102, low;}


mul.ftz.f32 %f129, %f101, %f102;
cvt.rn.f32.s32 %f130, %r1284;
fma.rn.ftz.f32 %f314, %f129, %f130, %f314;
mov.u32 %r523, %r719;

prmt.b32 %r523, %r523, 0, 0xba98;

xor.b32 %r1285, %r523, %r893;
or.b32 %r1286, %r1285, -2139062144;
and.b32 %r1287, %r523, 2139062143;
sub.s32 %r1288, %r1286, %r1287;
xor.b32 %r528, %r910, %r1288;
mov.u32 %r525, %r721;

prmt.b32 %r525, %r525, 0, 0xba98;

xor.b32 %r1289, %r525, %r896;
or.b32 %r1290, %r1289, -2139062144;
and.b32 %r1291, %r525, 2139062143;
sub.s32 %r1292, %r1290, %r1291;
xor.b32 %r532, %r918, %r1292;

dp4a.s32.s32 %r527, %r528, %r480, %r726;


dp4a.s32.s32 %r531, %r532, %r484, %r527;

mov.u32 %r535, %r731;

prmt.b32 %r535, %r535, 0, 0xba98;

xor.b32 %r1293, %r535, %r920;
or.b32 %r1294, %r1293, -2139062144;
and.b32 %r1295, %r535, 2139062143;
sub.s32 %r1296, %r1294, %r1295;
xor.b32 %r540, %r938, %r1296;
mov.u32 %r537, %r733;

prmt.b32 %r537, %r537, 0, 0xba98;

xor.b32 %r1297, %r537, %r923;
or.b32 %r1298, %r1297, -2139062144;
and.b32 %r1299, %r537, 2139062143;
sub.s32 %r1300, %r1298, %r1299;
xor.b32 %r544, %r946, %r1300;

dp4a.s32.s32 %r539, %r540, %r492, %r531;


dp4a.s32.s32 %r543, %r544, %r496, %r539;

mov.u32 %r547, %r743;

prmt.b32 %r547, %r547, 0, 0xba98;

xor.b32 %r1301, %r547, %r948;
or.b32 %r1302, %r1301, -2139062144;
and.b32 %r1303, %r547, 2139062143;
sub.s32 %r1304, %r1302, %r1303;
xor.b32 %r552, %r966, %r1304;
mov.u32 %r549, %r745;

prmt.b32 %r549, %r549, 0, 0xba98;

xor.b32 %r1305, %r549, %r951;
or.b32 %r1306, %r1305, -2139062144;
and.b32 %r1307, %r549, 2139062143;
sub.s32 %r1308, %r1306, %r1307;
xor.b32 %r556, %r974, %r1308;

dp4a.s32.s32 %r551, %r552, %r504, %r543;


dp4a.s32.s32 %r555, %r556, %r508, %r551;

mov.u32 %r559, %r755;

prmt.b32 %r559, %r559, 0, 0xba98;

xor.b32 %r1309, %r559, %r976;
or.b32 %r1310, %r1309, -2139062144;
and.b32 %r1311, %r559, 2139062143;
sub.s32 %r1312, %r1310, %r1311;
xor.b32 %r564, %r994, %r1312;
mov.u32 %r561, %r757;

prmt.b32 %r561, %r561, 0, 0xba98;

xor.b32 %r1313, %r561, %r979;
or.b32 %r1314, %r1313, -2139062144;
and.b32 %r1315, %r561, 2139062143;
sub.s32 %r1316, %r1314, %r1315;
xor.b32 %r568, %r1002, %r1316;

dp4a.s32.s32 %r563, %r564, %r516, %r555;


dp4a.s32.s32 %r567, %r568, %r520, %r563;

shr.u32 %r1317, %r567, 31;
add.s32 %r1318, %r567, %r1317;
shr.s32 %r1319, %r1318, 1;
mad.lo.s32 %r1320, %r567, %r1004, %r1319;
shr.u32 %r1321, %r1320, 31;
add.s32 %r1322, %r1320, %r1321;
shr.s32 %r1323, %r1322, 1;

{ cvt.f32.f16 %f103, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r522;
cvt.f32.f16 %f104, low;}


mul.ftz.f32 %f131, %f103, %f104;
cvt.rn.f32.s32 %f132, %r1323;
fma.rn.ftz.f32 %f313, %f131, %f132, %f313;
cvt.u64.u32 %rd100, %r1869;
add.s64 %rd101, %rd1, %rd100;
add.s64 %rd102, %rd101, %rd3;
mul.lo.s64 %rd103, %rd102, 36;
mov.u32 %r572, %r670;

prmt.b32 %r572, %r572, 0, 0xba98;

xor.b32 %r1324, %r572, %r771;
or.b32 %r1325, %r1324, -2139062144;
and.b32 %r1326, %r572, 2139062143;
sub.s32 %r1327, %r1325, %r1326;
xor.b32 %r577, %r788, %r1327;
add.s64 %rd104, %rd10, %rd103;
ld.global.nc.u32 %r578, [%rd104+4];
mov.u32 %r574, %r672;

prmt.b32 %r574, %r574, 0, 0xba98;

xor.b32 %r1328, %r574, %r774;
or.b32 %r1329, %r1328, -2139062144;
and.b32 %r1330, %r574, 2139062143;
sub.s32 %r1331, %r1329, %r1330;
xor.b32 %r581, %r796, %r1331;
ld.global.nc.u32 %r582, [%rd104+8];

dp4a.s32.s32 %r576, %r577, %r578, %r726;


dp4a.s32.s32 %r580, %r581, %r582, %r576;

mov.u32 %r584, %r682;

prmt.b32 %r584, %r584, 0, 0xba98;

xor.b32 %r1332, %r584, %r798;
or.b32 %r1333, %r1332, -2139062144;
and.b32 %r1334, %r584, 2139062143;
sub.s32 %r1335, %r1333, %r1334;
xor.b32 %r589, %r816, %r1335;
ld.global.nc.u32 %r590, [%rd104+12];
mov.u32 %r586, %r684;

prmt.b32 %r586, %r586, 0, 0xba98;

xor.b32 %r1336, %r586, %r801;
or.b32 %r1337, %r1336, -2139062144;
and.b32 %r1338, %r586, 2139062143;
sub.s32 %r1339, %r1337, %r1338;
xor.b32 %r593, %r824, %r1339;
ld.global.nc.u32 %r594, [%rd104+16];

dp4a.s32.s32 %r588, %r589, %r590, %r580;


dp4a.s32.s32 %r592, %r593, %r594, %r588;

mov.u32 %r596, %r694;

prmt.b32 %r596, %r596, 0, 0xba98;

xor.b32 %r1340, %r596, %r826;
or.b32 %r1341, %r1340, -2139062144;
and.b32 %r1342, %r596, 2139062143;
sub.s32 %r1343, %r1341, %r1342;
xor.b32 %r601, %r844, %r1343;
ld.global.nc.u32 %r602, [%rd104+20];
mov.u32 %r598, %r696;

prmt.b32 %r598, %r598, 0, 0xba98;

xor.b32 %r1344, %r598, %r829;
or.b32 %r1345, %r1344, -2139062144;
and.b32 %r1346, %r598, 2139062143;
sub.s32 %r1347, %r1345, %r1346;
xor.b32 %r605, %r852, %r1347;
ld.global.nc.u32 %r606, [%rd104+24];

dp4a.s32.s32 %r600, %r601, %r602, %r592;


dp4a.s32.s32 %r604, %r605, %r606, %r600;

mov.u32 %r608, %r706;

prmt.b32 %r608, %r608, 0, 0xba98;

xor.b32 %r1348, %r608, %r854;
or.b32 %r1349, %r1348, -2139062144;
and.b32 %r1350, %r608, 2139062143;
sub.s32 %r1351, %r1349, %r1350;
xor.b32 %r613, %r872, %r1351;
ld.global.nc.u32 %r614, [%rd104+28];
mov.u32 %r610, %r708;

prmt.b32 %r610, %r610, 0, 0xba98;

xor.b32 %r1352, %r610, %r857;
or.b32 %r1353, %r1352, -2139062144;
and.b32 %r1354, %r610, 2139062143;
sub.s32 %r1355, %r1353, %r1354;
xor.b32 %r617, %r880, %r1355;
ld.global.nc.u32 %r618, [%rd104+32];

dp4a.s32.s32 %r612, %r613, %r614, %r604;


dp4a.s32.s32 %r616, %r617, %r618, %r612;

shr.u32 %r1356, %r616, 31;
add.s32 %r1357, %r616, %r1356;
shr.s32 %r1358, %r1357, 1;
mad.lo.s32 %r1359, %r616, %r882, %r1358;
shr.u32 %r1360, %r1359, 31;
add.s32 %r1361, %r1359, %r1360;
shr.s32 %r1362, %r1361, 1;

{ cvt.f32.f16 %f105, %rs1;}


ld.global.nc.u32 %r620, [%rd104];

{.reg .f16 low,high;
mov.b32 {low,high},%r620;
cvt.f32.f16 %f106, low;}


mul.ftz.f32 %f133, %f105, %f106;
cvt.rn.f32.s32 %f134, %r1362;
fma.rn.ftz.f32 %f312, %f133, %f134, %f312;
mov.u32 %r621, %r719;

prmt.b32 %r621, %r621, 0, 0xba98;

xor.b32 %r1363, %r621, %r893;
or.b32 %r1364, %r1363, -2139062144;
and.b32 %r1365, %r621, 2139062143;
sub.s32 %r1366, %r1364, %r1365;
xor.b32 %r626, %r910, %r1366;
mov.u32 %r623, %r721;

prmt.b32 %r623, %r623, 0, 0xba98;

xor.b32 %r1367, %r623, %r896;
or.b32 %r1368, %r1367, -2139062144;
and.b32 %r1369, %r623, 2139062143;
sub.s32 %r1370, %r1368, %r1369;
xor.b32 %r630, %r918, %r1370;

dp4a.s32.s32 %r625, %r626, %r578, %r726;


dp4a.s32.s32 %r629, %r630, %r582, %r625;

mov.u32 %r633, %r731;

prmt.b32 %r633, %r633, 0, 0xba98;

xor.b32 %r1371, %r633, %r920;
or.b32 %r1372, %r1371, -2139062144;
and.b32 %r1373, %r633, 2139062143;
sub.s32 %r1374, %r1372, %r1373;
xor.b32 %r638, %r938, %r1374;
mov.u32 %r635, %r733;

prmt.b32 %r635, %r635, 0, 0xba98;

xor.b32 %r1375, %r635, %r923;
or.b32 %r1376, %r1375, -2139062144;
and.b32 %r1377, %r635, 2139062143;
sub.s32 %r1378, %r1376, %r1377;
xor.b32 %r642, %r946, %r1378;

dp4a.s32.s32 %r637, %r638, %r590, %r629;


dp4a.s32.s32 %r641, %r642, %r594, %r637;

mov.u32 %r645, %r743;

prmt.b32 %r645, %r645, 0, 0xba98;

xor.b32 %r1379, %r645, %r948;
or.b32 %r1380, %r1379, -2139062144;
and.b32 %r1381, %r645, 2139062143;
sub.s32 %r1382, %r1380, %r1381;
xor.b32 %r650, %r966, %r1382;
mov.u32 %r647, %r745;

prmt.b32 %r647, %r647, 0, 0xba98;

xor.b32 %r1383, %r647, %r951;
or.b32 %r1384, %r1383, -2139062144;
and.b32 %r1385, %r647, 2139062143;
sub.s32 %r1386, %r1384, %r1385;
xor.b32 %r654, %r974, %r1386;

dp4a.s32.s32 %r649, %r650, %r602, %r641;


dp4a.s32.s32 %r653, %r654, %r606, %r649;

mov.u32 %r657, %r755;

prmt.b32 %r657, %r657, 0, 0xba98;

xor.b32 %r1387, %r657, %r976;
or.b32 %r1388, %r1387, -2139062144;
and.b32 %r1389, %r657, 2139062143;
sub.s32 %r1390, %r1388, %r1389;
xor.b32 %r662, %r994, %r1390;
mov.u32 %r659, %r757;

prmt.b32 %r659, %r659, 0, 0xba98;

xor.b32 %r1391, %r659, %r979;
or.b32 %r1392, %r1391, -2139062144;
and.b32 %r1393, %r659, 2139062143;
sub.s32 %r1394, %r1392, %r1393;
xor.b32 %r666, %r1002, %r1394;

dp4a.s32.s32 %r661, %r662, %r614, %r653;


dp4a.s32.s32 %r665, %r666, %r618, %r661;

shr.u32 %r1395, %r665, 31;
add.s32 %r1396, %r665, %r1395;
shr.s32 %r1397, %r1396, 1;
mad.lo.s32 %r1398, %r665, %r1004, %r1397;
shr.u32 %r1399, %r1398, 31;
add.s32 %r1400, %r1398, %r1399;
shr.s32 %r1401, %r1400, 1;

{ cvt.f32.f16 %f107, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r620;
cvt.f32.f16 %f108, low;}


mul.ftz.f32 %f135, %f107, %f108;
cvt.rn.f32.s32 %f136, %r1401;
fma.rn.ftz.f32 %f311, %f135, %f136, %f311;
cvt.u64.u32 %rd105, %r1868;
add.s64 %rd106, %rd1, %rd105;
add.s64 %rd107, %rd106, %rd3;
mul.lo.s64 %rd108, %rd107, 36;

prmt.b32 %r670, %r670, 0, 0xba98;

xor.b32 %r1402, %r670, %r771;
or.b32 %r1403, %r1402, -2139062144;
and.b32 %r1404, %r670, 2139062143;
sub.s32 %r1405, %r1403, %r1404;
xor.b32 %r675, %r788, %r1405;
add.s64 %rd109, %rd10, %rd108;
ld.global.nc.u32 %r676, [%rd109+4];

prmt.b32 %r672, %r672, 0, 0xba98;

xor.b32 %r1406, %r672, %r774;
or.b32 %r1407, %r1406, -2139062144;
and.b32 %r1408, %r672, 2139062143;
sub.s32 %r1409, %r1407, %r1408;
xor.b32 %r679, %r796, %r1409;
ld.global.nc.u32 %r680, [%rd109+8];

dp4a.s32.s32 %r674, %r675, %r676, %r726;


dp4a.s32.s32 %r678, %r679, %r680, %r674;


prmt.b32 %r682, %r682, 0, 0xba98;

xor.b32 %r1410, %r682, %r798;
or.b32 %r1411, %r1410, -2139062144;
and.b32 %r1412, %r682, 2139062143;
sub.s32 %r1413, %r1411, %r1412;
xor.b32 %r687, %r816, %r1413;
ld.global.nc.u32 %r688, [%rd109+12];

prmt.b32 %r684, %r684, 0, 0xba98;

xor.b32 %r1414, %r684, %r801;
or.b32 %r1415, %r1414, -2139062144;
and.b32 %r1416, %r684, 2139062143;
sub.s32 %r1417, %r1415, %r1416;
xor.b32 %r691, %r824, %r1417;
ld.global.nc.u32 %r692, [%rd109+16];

dp4a.s32.s32 %r686, %r687, %r688, %r678;


dp4a.s32.s32 %r690, %r691, %r692, %r686;


prmt.b32 %r694, %r694, 0, 0xba98;

xor.b32 %r1418, %r694, %r826;
or.b32 %r1419, %r1418, -2139062144;
and.b32 %r1420, %r694, 2139062143;
sub.s32 %r1421, %r1419, %r1420;
xor.b32 %r699, %r844, %r1421;
ld.global.nc.u32 %r700, [%rd109+20];

prmt.b32 %r696, %r696, 0, 0xba98;

xor.b32 %r1422, %r696, %r829;
or.b32 %r1423, %r1422, -2139062144;
and.b32 %r1424, %r696, 2139062143;
sub.s32 %r1425, %r1423, %r1424;
xor.b32 %r703, %r852, %r1425;
ld.global.nc.u32 %r704, [%rd109+24];

dp4a.s32.s32 %r698, %r699, %r700, %r690;


dp4a.s32.s32 %r702, %r703, %r704, %r698;


prmt.b32 %r706, %r706, 0, 0xba98;

xor.b32 %r1426, %r706, %r854;
or.b32 %r1427, %r1426, -2139062144;
and.b32 %r1428, %r706, 2139062143;
sub.s32 %r1429, %r1427, %r1428;
xor.b32 %r711, %r872, %r1429;
ld.global.nc.u32 %r712, [%rd109+28];

prmt.b32 %r708, %r708, 0, 0xba98;

xor.b32 %r1430, %r708, %r857;
or.b32 %r1431, %r1430, -2139062144;
and.b32 %r1432, %r708, 2139062143;
sub.s32 %r1433, %r1431, %r1432;
xor.b32 %r715, %r880, %r1433;
ld.global.nc.u32 %r716, [%rd109+32];

dp4a.s32.s32 %r710, %r711, %r712, %r702;


dp4a.s32.s32 %r714, %r715, %r716, %r710;

shr.u32 %r1434, %r714, 31;
add.s32 %r1435, %r714, %r1434;
shr.s32 %r1436, %r1435, 1;
mad.lo.s32 %r1437, %r714, %r882, %r1436;
shr.u32 %r1438, %r1437, 31;
add.s32 %r1439, %r1437, %r1438;
shr.s32 %r1440, %r1439, 1;

{ cvt.f32.f16 %f109, %rs1;}


ld.global.nc.u32 %r718, [%rd109];

{.reg .f16 low,high;
mov.b32 {low,high},%r718;
cvt.f32.f16 %f110, low;}


mul.ftz.f32 %f137, %f109, %f110;
cvt.rn.f32.s32 %f138, %r1440;
fma.rn.ftz.f32 %f310, %f137, %f138, %f310;

prmt.b32 %r719, %r719, 0, 0xba98;

xor.b32 %r1441, %r719, %r893;
or.b32 %r1442, %r1441, -2139062144;
and.b32 %r1443, %r719, 2139062143;
sub.s32 %r1444, %r1442, %r1443;
xor.b32 %r724, %r910, %r1444;

prmt.b32 %r721, %r721, 0, 0xba98;

xor.b32 %r1445, %r721, %r896;
or.b32 %r1446, %r1445, -2139062144;
and.b32 %r1447, %r721, 2139062143;
sub.s32 %r1448, %r1446, %r1447;
xor.b32 %r728, %r918, %r1448;

dp4a.s32.s32 %r723, %r724, %r676, %r726;


dp4a.s32.s32 %r727, %r728, %r680, %r723;


prmt.b32 %r731, %r731, 0, 0xba98;

xor.b32 %r1449, %r731, %r920;
or.b32 %r1450, %r1449, -2139062144;
and.b32 %r1451, %r731, 2139062143;
sub.s32 %r1452, %r1450, %r1451;
xor.b32 %r736, %r938, %r1452;

prmt.b32 %r733, %r733, 0, 0xba98;

xor.b32 %r1453, %r733, %r923;
or.b32 %r1454, %r1453, -2139062144;
and.b32 %r1455, %r733, 2139062143;
sub.s32 %r1456, %r1454, %r1455;
xor.b32 %r740, %r946, %r1456;

dp4a.s32.s32 %r735, %r736, %r688, %r727;


dp4a.s32.s32 %r739, %r740, %r692, %r735;


prmt.b32 %r743, %r743, 0, 0xba98;

xor.b32 %r1457, %r743, %r948;
or.b32 %r1458, %r1457, -2139062144;
and.b32 %r1459, %r743, 2139062143;
sub.s32 %r1460, %r1458, %r1459;
xor.b32 %r748, %r966, %r1460;

prmt.b32 %r745, %r745, 0, 0xba98;

xor.b32 %r1461, %r745, %r951;
or.b32 %r1462, %r1461, -2139062144;
and.b32 %r1463, %r745, 2139062143;
sub.s32 %r1464, %r1462, %r1463;
xor.b32 %r752, %r974, %r1464;

dp4a.s32.s32 %r747, %r748, %r700, %r739;


dp4a.s32.s32 %r751, %r752, %r704, %r747;


prmt.b32 %r755, %r755, 0, 0xba98;

xor.b32 %r1465, %r755, %r976;
or.b32 %r1466, %r1465, -2139062144;
and.b32 %r1467, %r755, 2139062143;
sub.s32 %r1468, %r1466, %r1467;
xor.b32 %r760, %r994, %r1468;

prmt.b32 %r757, %r757, 0, 0xba98;

xor.b32 %r1469, %r757, %r979;
or.b32 %r1470, %r1469, -2139062144;
and.b32 %r1471, %r757, 2139062143;
sub.s32 %r1472, %r1470, %r1471;
xor.b32 %r764, %r1002, %r1472;

dp4a.s32.s32 %r759, %r760, %r712, %r751;


dp4a.s32.s32 %r763, %r764, %r716, %r759;

shr.u32 %r1473, %r763, 31;
add.s32 %r1474, %r763, %r1473;
shr.s32 %r1475, %r1474, 1;
mad.lo.s32 %r1476, %r763, %r1004, %r1475;
shr.u32 %r1477, %r1476, 31;
add.s32 %r1478, %r1476, %r1477;
shr.s32 %r1479, %r1478, 1;

{ cvt.f32.f16 %f111, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r718;
cvt.f32.f16 %f112, low;}


mul.ftz.f32 %f139, %f111, %f112;
cvt.rn.f32.s32 %f140, %r1479;
fma.rn.ftz.f32 %f309, %f139, %f140, %f309;
add.s32 %r1872, %r1872, 64;
add.s32 %r1871, %r1871, 64;
add.s32 %r1870, %r1870, 64;
add.s32 %r1869, %r1869, 64;
add.s32 %r1868, %r1868, 64;
add.s32 %r1867, %r1867, 64;
add.s32 %r1866, %r1866, 64;
add.s64 %rd167, %rd167, 784;
add.s32 %r1873, %r1873, 8;
setp.lt.s32 %p3, %r1873, %r50;
@%p3 bra $L__BB248_2;

$L__BB248_3:
setp.eq.s32 %p4, %r44, 0;
@%p4 bra $L__BB248_5;

mov.u32 %r1484, _ZZ13mul_mat_vec_qIL9ggml_type18ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r1485, %r44, 1792, %r1484;
shl.b32 %r1486, %r46, 2;
add.s32 %r1487, %r1485, %r1486;
st.shared.f32 [%r1487+-1792], %f322;
st.shared.f32 [%r1487+-1664], %f321;
st.shared.f32 [%r1487+-1536], %f320;
st.shared.f32 [%r1487+-1408], %f319;
st.shared.f32 [%r1487+-1280], %f318;
st.shared.f32 [%r1487+-1152], %f317;
st.shared.f32 [%r1487+-1024], %f316;
st.shared.f32 [%r1487+-896], %f315;
st.shared.f32 [%r1487+-768], %f314;
st.shared.f32 [%r1487+-640], %f313;
st.shared.f32 [%r1487+-512], %f312;
st.shared.f32 [%r1487+-384], %f311;
st.shared.f32 [%r1487+-256], %f310;
st.shared.f32 [%r1487+-128], %f309;

$L__BB248_5:
setp.ne.s32 %p5, %r44, 0;
bar.sync 0;
@%p5 bra $L__BB248_34;

mov.u32 %r1489, %ctaid.x;
shl.b32 %r1490, %r1489, 1;
mov.u32 %r1491, 1;
shl.b32 %r1493, %r46, 2;
mov.u32 %r1494, _ZZ13mul_mat_vec_qIL9ggml_type18ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r1495, %r1494, %r1493;
mov.u32 %r1496, 2;
ld.shared.f32 %f141, [%r1495];
add.ftz.f32 %f142, %f141, %f322;
mov.b32 %r1497, %f142;
mov.u32 %r1498, 31;
mov.u32 %r1499, 16;
mov.u32 %r1500, -1;
shfl.sync.bfly.b32 %r1501|%p6, %r1497, %r1499, %r1498, %r1500;
mov.b32 %f143, %r1501;
add.ftz.f32 %f144, %f142, %f143;
mov.b32 %r1502, %f144;
mov.u32 %r1503, 8;
shfl.sync.bfly.b32 %r1504|%p7, %r1502, %r1503, %r1498, %r1500;
mov.b32 %f145, %r1504;
add.ftz.f32 %f146, %f144, %f145;
mov.b32 %r1505, %f146;
mov.u32 %r1506, 4;
shfl.sync.bfly.b32 %r1507|%p8, %r1505, %r1506, %r1498, %r1500;
mov.b32 %f147, %r1507;
add.ftz.f32 %f148, %f146, %f147;
mov.b32 %r1508, %f148;
shfl.sync.bfly.b32 %r1509|%p9, %r1508, %r1496, %r1498, %r1500;
mov.b32 %f149, %r1509;
add.ftz.f32 %f150, %f148, %f149;
mov.b32 %r1510, %f150;
shfl.sync.bfly.b32 %r1511|%p10, %r1510, %r1491, %r1498, %r1500;
mov.b32 %f151, %r1511;
add.ftz.f32 %f43, %f150, %f151;
setp.ne.s32 %p11, %r46, 0;
setp.ge.u32 %p12, %r1490, %r31;
or.pred %p13, %p11, %p12;
@%p13 bra $L__BB248_8;

cvta.to.global.u64 %rd110, %rd18;
mov.u32 %r1512, %ctaid.y;
mad.lo.s32 %r1515, %r1512, %r37, %r1490;
mov.u32 %r1516, %ctaid.z;
mad.lo.s32 %r1517, %r1516, %r43, %r1515;
mul.wide.u32 %rd111, %r1517, 4;
add.s64 %rd112, %rd110, %rd111;
st.global.f32 [%rd112], %f43;

$L__BB248_8:
ld.shared.f32 %f152, [%r1495+128];
add.ftz.f32 %f153, %f152, %f321;
mov.b32 %r1523, %f153;
shfl.sync.bfly.b32 %r1527|%p14, %r1523, %r1499, %r1498, %r1500;
mov.b32 %f154, %r1527;
add.ftz.f32 %f155, %f153, %f154;
mov.b32 %r1528, %f155;
shfl.sync.bfly.b32 %r1530|%p15, %r1528, %r1503, %r1498, %r1500;
mov.b32 %f156, %r1530;
add.ftz.f32 %f157, %f155, %f156;
mov.b32 %r1531, %f157;
shfl.sync.bfly.b32 %r1533|%p16, %r1531, %r1506, %r1498, %r1500;
mov.b32 %f158, %r1533;
add.ftz.f32 %f159, %f157, %f158;
mov.b32 %r1534, %f159;
shfl.sync.bfly.b32 %r1535|%p17, %r1534, %r1496, %r1498, %r1500;
mov.b32 %f160, %r1535;
add.ftz.f32 %f161, %f159, %f160;
mov.b32 %r1536, %f161;
shfl.sync.bfly.b32 %r1538|%p18, %r1536, %r1491, %r1498, %r1500;
mov.b32 %f162, %r1538;
add.ftz.f32 %f44, %f161, %f162;
setp.eq.s32 %p19, %r46, 1;
or.b32 %r1541, %r1490, 1;
setp.lt.u32 %p20, %r1541, %r31;
and.pred %p1, %p19, %p20;
not.pred %p21, %p1;
@%p21 bra $L__BB248_10;

cvta.to.global.u64 %rd113, %rd18;
mov.u32 %r1542, %ctaid.y;
mad.lo.s32 %r1545, %r1542, %r37, %r1490;
mov.u32 %r1546, %ctaid.z;
mad.lo.s32 %r1547, %r1546, %r43, %r1545;
mul.wide.u32 %rd114, %r1547, 4;
add.s64 %rd115, %rd113, %rd114;
st.global.f32 [%rd115+4], %f44;

$L__BB248_10:
mov.u32 %r1552, 2;
ld.shared.f32 %f163, [%r1495+256];
add.ftz.f32 %f164, %f163, %f320;
mov.b32 %r1553, %f164;
mov.u32 %r1554, 31;
mov.u32 %r1555, 16;
mov.u32 %r1556, -1;
shfl.sync.bfly.b32 %r1557|%p22, %r1553, %r1555, %r1554, %r1556;
mov.b32 %f165, %r1557;
add.ftz.f32 %f166, %f164, %f165;
mov.b32 %r1558, %f166;
mov.u32 %r1559, 8;
shfl.sync.bfly.b32 %r1560|%p23, %r1558, %r1559, %r1554, %r1556;
mov.b32 %f167, %r1560;
add.ftz.f32 %f168, %f166, %f167;
mov.b32 %r1561, %f168;
mov.u32 %r1562, 4;
shfl.sync.bfly.b32 %r1563|%p24, %r1561, %r1562, %r1554, %r1556;
mov.b32 %f169, %r1563;
add.ftz.f32 %f170, %f168, %f169;
mov.b32 %r1564, %f170;
shfl.sync.bfly.b32 %r1565|%p25, %r1564, %r1552, %r1554, %r1556;
mov.b32 %f171, %r1565;
add.ftz.f32 %f172, %f170, %f171;
mov.b32 %r1566, %f172;
mov.u32 %r1567, 1;
shfl.sync.bfly.b32 %r1568|%p26, %r1566, %r1567, %r1554, %r1556;
mov.b32 %f173, %r1568;
add.ftz.f32 %f45, %f172, %f173;
@%p13 bra $L__BB248_12;

cvt.u64.u32 %rd116, %r31;
mov.u32 %r1571, %ctaid.y;
mad.lo.s32 %r1574, %r1571, %r37, %r1490;
mov.u32 %r1575, %ctaid.z;
mad.lo.s32 %r1576, %r1575, %r43, %r1574;
cvt.u64.u32 %rd117, %r1576;
add.s64 %rd118, %rd116, %rd117;
cvta.to.global.u64 %rd119, %rd18;
shl.b64 %rd120, %rd118, 2;
add.s64 %rd121, %rd119, %rd120;
st.global.f32 [%rd121], %f45;

$L__BB248_12:
ld.shared.f32 %f174, [%r1495+384];
add.ftz.f32 %f175, %f174, %f319;
mov.b32 %r1582, %f175;
shfl.sync.bfly.b32 %r1586|%p30, %r1582, %r1555, %r1554, %r1556;
mov.b32 %f176, %r1586;
add.ftz.f32 %f177, %f175, %f176;
mov.b32 %r1587, %f177;
shfl.sync.bfly.b32 %r1589|%p31, %r1587, %r1559, %r1554, %r1556;
mov.b32 %f178, %r1589;
add.ftz.f32 %f179, %f177, %f178;
mov.b32 %r1590, %f179;
shfl.sync.bfly.b32 %r1592|%p32, %r1590, %r1562, %r1554, %r1556;
mov.b32 %f180, %r1592;
add.ftz.f32 %f181, %f179, %f180;
mov.b32 %r1593, %f181;
shfl.sync.bfly.b32 %r1594|%p33, %r1593, %r1552, %r1554, %r1556;
mov.b32 %f182, %r1594;
add.ftz.f32 %f183, %f181, %f182;
mov.b32 %r1595, %f183;
shfl.sync.bfly.b32 %r1597|%p34, %r1595, %r1567, %r1554, %r1556;
mov.b32 %f184, %r1597;
add.ftz.f32 %f46, %f183, %f184;
@%p21 bra $L__BB248_14;

add.s32 %r1598, %r31, 1;
cvt.u64.u32 %rd122, %r1598;
mov.u32 %r1599, %ctaid.y;
mad.lo.s32 %r1602, %r1599, %r37, %r1490;
mov.u32 %r1603, %ctaid.z;
mad.lo.s32 %r1604, %r1603, %r43, %r1602;
cvt.u64.u32 %rd123, %r1604;
add.s64 %rd124, %rd122, %rd123;
cvta.to.global.u64 %rd125, %rd18;
shl.b64 %rd126, %rd124, 2;
add.s64 %rd127, %rd125, %rd126;
st.global.f32 [%rd127], %f46;

$L__BB248_14:
shl.b32 %r1605, %r31, 1;
mov.u32 %r1606, 1;
mov.u32 %r1611, 2;
ld.shared.f32 %f185, [%r1495+512];
add.ftz.f32 %f186, %f185, %f318;
mov.b32 %r1612, %f186;
mov.u32 %r1613, 31;
mov.u32 %r1614, 16;
mov.u32 %r1615, -1;
shfl.sync.bfly.b32 %r1616|%p36, %r1612, %r1614, %r1613, %r1615;
mov.b32 %f187, %r1616;
add.ftz.f32 %f188, %f186, %f187;
mov.b32 %r1617, %f188;
mov.u32 %r1618, 8;
shfl.sync.bfly.b32 %r1619|%p37, %r1617, %r1618, %r1613, %r1615;
mov.b32 %f189, %r1619;
add.ftz.f32 %f190, %f188, %f189;
mov.b32 %r1620, %f190;
mov.u32 %r1621, 4;
shfl.sync.bfly.b32 %r1622|%p38, %r1620, %r1621, %r1613, %r1615;
mov.b32 %f191, %r1622;
add.ftz.f32 %f192, %f190, %f191;
mov.b32 %r1623, %f192;
shfl.sync.bfly.b32 %r1624|%p39, %r1623, %r1611, %r1613, %r1615;
mov.b32 %f193, %r1624;
add.ftz.f32 %f194, %f192, %f193;
mov.b32 %r1625, %f194;
shfl.sync.bfly.b32 %r1626|%p40, %r1625, %r1606, %r1613, %r1615;
mov.b32 %f195, %r1626;
add.ftz.f32 %f47, %f194, %f195;
cvt.u64.u32 %rd128, %r1605;
mov.u32 %r1627, %ctaid.y;
mad.lo.s32 %r1630, %r1627, %r37, %r1490;
mov.u32 %r1631, %ctaid.z;
mad.lo.s32 %r1632, %r1631, %r43, %r1630;
cvt.u64.u32 %rd129, %r1632;
add.s64 %rd130, %rd128, %rd129;
cvta.to.global.u64 %rd131, %rd18;
shl.b64 %rd132, %rd130, 2;
add.s64 %rd13, %rd131, %rd132;
@%p13 bra $L__BB248_16;

st.global.f32 [%rd13], %f47;

$L__BB248_16:
ld.shared.f32 %f196, [%r1495+640];
add.ftz.f32 %f197, %f196, %f317;
mov.b32 %r1638, %f197;
shfl.sync.bfly.b32 %r1642|%p44, %r1638, %r1614, %r1613, %r1615;
mov.b32 %f198, %r1642;
add.ftz.f32 %f199, %f197, %f198;
mov.b32 %r1643, %f199;
shfl.sync.bfly.b32 %r1645|%p45, %r1643, %r1618, %r1613, %r1615;
mov.b32 %f200, %r1645;
add.ftz.f32 %f201, %f199, %f200;
mov.b32 %r1646, %f201;
shfl.sync.bfly.b32 %r1648|%p46, %r1646, %r1621, %r1613, %r1615;
mov.b32 %f202, %r1648;
add.ftz.f32 %f203, %f201, %f202;
mov.b32 %r1649, %f203;
shfl.sync.bfly.b32 %r1650|%p47, %r1649, %r1611, %r1613, %r1615;
mov.b32 %f204, %r1650;
add.ftz.f32 %f205, %f203, %f204;
mov.b32 %r1651, %f205;
shfl.sync.bfly.b32 %r1653|%p48, %r1651, %r1606, %r1613, %r1615;
mov.b32 %f206, %r1653;
add.ftz.f32 %f48, %f205, %f206;
@%p21 bra $L__BB248_18;

st.global.f32 [%rd13+4], %f48;

$L__BB248_18:
mul.lo.s32 %r26, %r31, 3;
mov.u32 %r1658, 2;
ld.shared.f32 %f207, [%r1495+768];
add.ftz.f32 %f208, %f207, %f316;
mov.b32 %r1659, %f208;
mov.u32 %r1660, 31;
mov.u32 %r1661, 16;
mov.u32 %r1662, -1;
shfl.sync.bfly.b32 %r1663|%p50, %r1659, %r1661, %r1660, %r1662;
mov.b32 %f209, %r1663;
add.ftz.f32 %f210, %f208, %f209;
mov.b32 %r1664, %f210;
mov.u32 %r1665, 8;
shfl.sync.bfly.b32 %r1666|%p51, %r1664, %r1665, %r1660, %r1662;
mov.b32 %f211, %r1666;
add.ftz.f32 %f212, %f210, %f211;
mov.b32 %r1667, %f212;
mov.u32 %r1668, 4;
shfl.sync.bfly.b32 %r1669|%p52, %r1667, %r1668, %r1660, %r1662;
mov.b32 %f213, %r1669;
add.ftz.f32 %f214, %f212, %f213;
mov.b32 %r1670, %f214;
shfl.sync.bfly.b32 %r1671|%p53, %r1670, %r1658, %r1660, %r1662;
mov.b32 %f215, %r1671;
add.ftz.f32 %f216, %f214, %f215;
mov.b32 %r1672, %f216;
mov.u32 %r1673, 1;
shfl.sync.bfly.b32 %r1674|%p54, %r1672, %r1673, %r1660, %r1662;
mov.b32 %f217, %r1674;
add.ftz.f32 %f49, %f216, %f217;
@%p13 bra $L__BB248_20;

cvt.u64.u32 %rd133, %r26;
add.s64 %rd135, %rd133, %rd129;
shl.b64 %rd137, %rd135, 2;
add.s64 %rd138, %rd131, %rd137;
st.global.f32 [%rd138], %f49;

$L__BB248_20:
ld.shared.f32 %f218, [%r1495+896];
add.ftz.f32 %f219, %f218, %f315;
mov.b32 %r1688, %f219;
shfl.sync.bfly.b32 %r1692|%p58, %r1688, %r1661, %r1660, %r1662;
mov.b32 %f220, %r1692;
add.ftz.f32 %f221, %f219, %f220;
mov.b32 %r1693, %f221;
shfl.sync.bfly.b32 %r1695|%p59, %r1693, %r1665, %r1660, %r1662;
mov.b32 %f222, %r1695;
add.ftz.f32 %f223, %f221, %f222;
mov.b32 %r1696, %f223;
shfl.sync.bfly.b32 %r1698|%p60, %r1696, %r1668, %r1660, %r1662;
mov.b32 %f224, %r1698;
add.ftz.f32 %f225, %f223, %f224;
mov.b32 %r1699, %f225;
shfl.sync.bfly.b32 %r1700|%p61, %r1699, %r1658, %r1660, %r1662;
mov.b32 %f226, %r1700;
add.ftz.f32 %f227, %f225, %f226;
mov.b32 %r1701, %f227;
shfl.sync.bfly.b32 %r1703|%p62, %r1701, %r1673, %r1660, %r1662;
mov.b32 %f228, %r1703;
add.ftz.f32 %f50, %f227, %f228;
@%p21 bra $L__BB248_22;

add.s32 %r1704, %r26, 1;
cvt.u64.u32 %rd139, %r1704;
add.s64 %rd141, %rd139, %rd129;
shl.b64 %rd143, %rd141, 2;
add.s64 %rd144, %rd131, %rd143;
st.global.f32 [%rd144], %f50;

$L__BB248_22:
shl.b32 %r1711, %r31, 2;
mov.u32 %r1712, 2;
ld.shared.f32 %f229, [%r1495+1024];
add.ftz.f32 %f230, %f229, %f314;
mov.b32 %r1717, %f230;
mov.u32 %r1718, 31;
mov.u32 %r1719, 16;
mov.u32 %r1720, -1;
shfl.sync.bfly.b32 %r1721|%p64, %r1717, %r1719, %r1718, %r1720;
mov.b32 %f231, %r1721;
add.ftz.f32 %f232, %f230, %f231;
mov.b32 %r1722, %f232;
mov.u32 %r1723, 8;
shfl.sync.bfly.b32 %r1724|%p65, %r1722, %r1723, %r1718, %r1720;
mov.b32 %f233, %r1724;
add.ftz.f32 %f234, %f232, %f233;
mov.b32 %r1725, %f234;
mov.u32 %r1726, 4;
shfl.sync.bfly.b32 %r1727|%p66, %r1725, %r1726, %r1718, %r1720;
mov.b32 %f235, %r1727;
add.ftz.f32 %f236, %f234, %f235;
mov.b32 %r1728, %f236;
shfl.sync.bfly.b32 %r1729|%p67, %r1728, %r1712, %r1718, %r1720;
mov.b32 %f237, %r1729;
add.ftz.f32 %f238, %f236, %f237;
mov.b32 %r1730, %f238;
mov.u32 %r1731, 1;
shfl.sync.bfly.b32 %r1732|%p68, %r1730, %r1731, %r1718, %r1720;
mov.b32 %f239, %r1732;
add.ftz.f32 %f51, %f238, %f239;
cvt.u64.u32 %rd145, %r1711;
add.s64 %rd147, %rd145, %rd129;
shl.b64 %rd149, %rd147, 2;
add.s64 %rd14, %rd131, %rd149;
@%p13 bra $L__BB248_24;

st.global.f32 [%rd14], %f51;

$L__BB248_24:
ld.shared.f32 %f240, [%r1495+1152];
add.ftz.f32 %f241, %f240, %f313;
mov.b32 %r1744, %f241;
shfl.sync.bfly.b32 %r1748|%p72, %r1744, %r1719, %r1718, %r1720;
mov.b32 %f242, %r1748;
add.ftz.f32 %f243, %f241, %f242;
mov.b32 %r1749, %f243;
shfl.sync.bfly.b32 %r1751|%p73, %r1749, %r1723, %r1718, %r1720;
mov.b32 %f244, %r1751;
add.ftz.f32 %f245, %f243, %f244;
mov.b32 %r1752, %f245;
shfl.sync.bfly.b32 %r1754|%p74, %r1752, %r1726, %r1718, %r1720;
mov.b32 %f246, %r1754;
add.ftz.f32 %f247, %f245, %f246;
mov.b32 %r1755, %f247;
shfl.sync.bfly.b32 %r1756|%p75, %r1755, %r1712, %r1718, %r1720;
mov.b32 %f248, %r1756;
add.ftz.f32 %f249, %f247, %f248;
mov.b32 %r1757, %f249;
shfl.sync.bfly.b32 %r1759|%p76, %r1757, %r1731, %r1718, %r1720;
mov.b32 %f250, %r1759;
add.ftz.f32 %f52, %f249, %f250;
@%p21 bra $L__BB248_26;

st.global.f32 [%rd14+4], %f52;

$L__BB248_26:
mul.lo.s32 %r27, %r31, 5;
mov.u32 %r1764, 2;
ld.shared.f32 %f251, [%r1495+1280];
add.ftz.f32 %f252, %f251, %f312;
mov.b32 %r1765, %f252;
mov.u32 %r1766, 31;
mov.u32 %r1767, 16;
mov.u32 %r1768, -1;
shfl.sync.bfly.b32 %r1769|%p78, %r1765, %r1767, %r1766, %r1768;
mov.b32 %f253, %r1769;
add.ftz.f32 %f254, %f252, %f253;
mov.b32 %r1770, %f254;
mov.u32 %r1771, 8;
shfl.sync.bfly.b32 %r1772|%p79, %r1770, %r1771, %r1766, %r1768;
mov.b32 %f255, %r1772;
add.ftz.f32 %f256, %f254, %f255;
mov.b32 %r1773, %f256;
mov.u32 %r1774, 4;
shfl.sync.bfly.b32 %r1775|%p80, %r1773, %r1774, %r1766, %r1768;
mov.b32 %f257, %r1775;
add.ftz.f32 %f258, %f256, %f257;
mov.b32 %r1776, %f258;
shfl.sync.bfly.b32 %r1777|%p81, %r1776, %r1764, %r1766, %r1768;
mov.b32 %f259, %r1777;
add.ftz.f32 %f260, %f258, %f259;
mov.b32 %r1778, %f260;
mov.u32 %r1779, 1;
shfl.sync.bfly.b32 %r1780|%p82, %r1778, %r1779, %r1766, %r1768;
mov.b32 %f261, %r1780;
add.ftz.f32 %f53, %f260, %f261;
@%p13 bra $L__BB248_28;

cvt.u64.u32 %rd150, %r27;
add.s64 %rd152, %rd150, %rd129;
shl.b64 %rd154, %rd152, 2;
add.s64 %rd155, %rd131, %rd154;
st.global.f32 [%rd155], %f53;

$L__BB248_28:
ld.shared.f32 %f262, [%r1495+1408];
add.ftz.f32 %f263, %f262, %f311;
mov.b32 %r1794, %f263;
shfl.sync.bfly.b32 %r1798|%p86, %r1794, %r1767, %r1766, %r1768;
mov.b32 %f264, %r1798;
add.ftz.f32 %f265, %f263, %f264;
mov.b32 %r1799, %f265;
shfl.sync.bfly.b32 %r1801|%p87, %r1799, %r1771, %r1766, %r1768;
mov.b32 %f266, %r1801;
add.ftz.f32 %f267, %f265, %f266;
mov.b32 %r1802, %f267;
shfl.sync.bfly.b32 %r1804|%p88, %r1802, %r1774, %r1766, %r1768;
mov.b32 %f268, %r1804;
add.ftz.f32 %f269, %f267, %f268;
mov.b32 %r1805, %f269;
shfl.sync.bfly.b32 %r1806|%p89, %r1805, %r1764, %r1766, %r1768;
mov.b32 %f270, %r1806;
add.ftz.f32 %f271, %f269, %f270;
mov.b32 %r1807, %f271;
shfl.sync.bfly.b32 %r1809|%p90, %r1807, %r1779, %r1766, %r1768;
mov.b32 %f272, %r1809;
add.ftz.f32 %f54, %f271, %f272;
@%p21 bra $L__BB248_30;

add.s32 %r1810, %r27, 1;
cvt.u64.u32 %rd156, %r1810;
add.s64 %rd158, %rd156, %rd129;
shl.b64 %rd160, %rd158, 2;
add.s64 %rd161, %rd131, %rd160;
st.global.f32 [%rd161], %f54;

$L__BB248_30:
mul.lo.s32 %r1817, %r31, 6;
mov.u32 %r1822, 2;
ld.shared.f32 %f273, [%r1495+1536];
add.ftz.f32 %f274, %f273, %f310;
mov.b32 %r1823, %f274;
mov.u32 %r1824, 31;
mov.u32 %r1825, 16;
mov.u32 %r1826, -1;
shfl.sync.bfly.b32 %r1827|%p92, %r1823, %r1825, %r1824, %r1826;
mov.b32 %f275, %r1827;
add.ftz.f32 %f276, %f274, %f275;
mov.b32 %r1828, %f276;
mov.u32 %r1829, 8;
shfl.sync.bfly.b32 %r1830|%p93, %r1828, %r1829, %r1824, %r1826;
mov.b32 %f277, %r1830;
add.ftz.f32 %f278, %f276, %f277;
mov.b32 %r1831, %f278;
mov.u32 %r1832, 4;
shfl.sync.bfly.b32 %r1833|%p94, %r1831, %r1832, %r1824, %r1826;
mov.b32 %f279, %r1833;
add.ftz.f32 %f280, %f278, %f279;
mov.b32 %r1834, %f280;
shfl.sync.bfly.b32 %r1835|%p95, %r1834, %r1822, %r1824, %r1826;
mov.b32 %f281, %r1835;
add.ftz.f32 %f282, %f280, %f281;
mov.b32 %r1836, %f282;
mov.u32 %r1837, 1;
shfl.sync.bfly.b32 %r1838|%p96, %r1836, %r1837, %r1824, %r1826;
mov.b32 %f283, %r1838;
add.ftz.f32 %f55, %f282, %f283;
cvt.u64.u32 %rd162, %r1817;
add.s64 %rd164, %rd162, %rd129;
shl.b64 %rd166, %rd164, 2;
add.s64 %rd15, %rd131, %rd166;
@%p13 bra $L__BB248_32;

st.global.f32 [%rd15], %f55;

$L__BB248_32:
ld.shared.f32 %f284, [%r1495+1664];
add.ftz.f32 %f285, %f284, %f309;
mov.b32 %r1850, %f285;
shfl.sync.bfly.b32 %r1854|%p100, %r1850, %r1825, %r1824, %r1826;
mov.b32 %f286, %r1854;
add.ftz.f32 %f287, %f285, %f286;
mov.b32 %r1855, %f287;
shfl.sync.bfly.b32 %r1857|%p101, %r1855, %r1829, %r1824, %r1826;
mov.b32 %f288, %r1857;
add.ftz.f32 %f289, %f287, %f288;
mov.b32 %r1858, %f289;
shfl.sync.bfly.b32 %r1860|%p102, %r1858, %r1832, %r1824, %r1826;
mov.b32 %f290, %r1860;
add.ftz.f32 %f291, %f289, %f290;
mov.b32 %r1861, %f291;
shfl.sync.bfly.b32 %r1862|%p103, %r1861, %r1822, %r1824, %r1826;
mov.b32 %f292, %r1862;
add.ftz.f32 %f293, %f291, %f292;
mov.b32 %r1863, %f293;
shfl.sync.bfly.b32 %r1865|%p104, %r1863, %r1837, %r1824, %r1826;
mov.b32 %f294, %r1865;
add.ftz.f32 %f56, %f293, %f294;
@%p21 bra $L__BB248_34;

st.global.f32 [%rd15+4], %f56;

$L__BB248_34:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type18ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type18ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type18ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type18ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type18ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type18ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type18ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type18ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type18ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type18ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<120>;
.reg .b16 %rs<45>;
.reg .f32 %f<369>;
.reg .b32 %r<2113>;
.reg .b64 %rd<185>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type18ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[2048];

ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type18ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type18ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type18ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type18ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type18ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type18ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type18ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type18ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type18ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type18ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r45, [_Z13mul_mat_vec_qIL9ggml_type18ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r46, [_Z13mul_mat_vec_qIL9ggml_type18ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r47, [_Z13mul_mat_vec_qIL9ggml_type18ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type18ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type18ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type18ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type18ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r48, %tid.y;
shl.b32 %r49, %r48, 5;
mov.u32 %r50, %tid.x;
add.s32 %r1, %r49, %r50;
shr.s32 %r51, %r1, 31;
shr.u32 %r52, %r51, 29;
add.s32 %r53, %r1, %r52;
shr.s32 %r2112, %r53, 3;
shr.u32 %r54, %r32, 8;
setp.ge.s32 %p2, %r2112, %r54;
mov.f32 %f353, 0f00000000;
mov.f32 %f354, %f353;
mov.f32 %f355, %f353;
mov.f32 %f356, %f353;
mov.f32 %f357, %f353;
mov.f32 %f358, %f353;
mov.f32 %f359, %f353;
mov.f32 %f360, %f353;
mov.f32 %f361, %f353;
mov.f32 %f362, %f353;
mov.f32 %f363, %f353;
mov.f32 %f364, %f353;
mov.f32 %f365, %f353;
mov.f32 %f366, %f353;
mov.f32 %f367, %f353;
mov.f32 %f368, %f353;
@%p2 bra $L__BB249_3;

mov.u32 %r55, %ctaid.x;
shl.b32 %r56, %r55, 1;
mov.u32 %r57, %ctaid.y;
mov.u32 %r58, %ctaid.z;
mul.lo.s32 %r59, %r57, %r40;
cvt.u64.u32 %rd19, %r59;
mul.lo.s32 %r60, %r58, %r46;
cvt.u64.u32 %rd20, %r60;
add.s64 %rd1, %rd20, %rd19;
mul.hi.u32 %r61, %r58, %r42;
add.s32 %r62, %r61, %r58;
shr.u32 %r63, %r62, %r43;
mul.lo.s32 %r64, %r63, %r45;
mul.hi.u32 %r65, %r57, %r36;
add.s32 %r66, %r65, %r57;
shr.u32 %r67, %r66, %r37;
mul.lo.s32 %r68, %r67, %r39;
cvta.to.global.u64 %rd184, %rd16;
and.b32 %r72, %r53, -8;
sub.s32 %r73, %r1, %r72;
shl.b32 %r74, %r73, 2;
shl.b32 %r75, %r73, 1;
add.s32 %r76, %r75, 33;
cvt.s64.s32 %rd3, %r73;
shl.b32 %r77, %r34, 1;
shl.b32 %r2104, %r2112, 3;
add.s32 %r2111, %r2104, %r77;
mad.lo.s32 %r2110, %r34, 3, %r2104;
shl.b32 %r78, %r34, 2;
add.s32 %r2109, %r2104, %r78;
mad.lo.s32 %r2108, %r34, 5, %r2104;
mad.lo.s32 %r2107, %r34, 6, %r2104;
mad.lo.s32 %r2106, %r34, 7, %r2104;
add.s32 %r2105, %r34, %r2104;
add.s32 %r79, %r2112, %r64;
add.s32 %r80, %r79, %r68;
or.b32 %r81, %r56, 1;
mad.lo.s32 %r82, %r33, %r81, %r80;
mul.wide.s32 %rd4, %r82, 98;
mad.lo.s32 %r83, %r56, %r33, %r2112;
add.s32 %r84, %r83, %r64;
add.s32 %r85, %r84, %r68;
mul.wide.s32 %rd5, %r85, 98;
mul.wide.s32 %rd21, %r74, 2;
add.s64 %rd22, %rd5, %rd21;
add.s64 %rd6, %rd22, 8;
mul.wide.s32 %rd23, %r76, 2;
add.s64 %rd24, %rd23, -2;
add.s64 %rd25, %rd24, %rd4;
add.s64 %rd7, %rd25, 2;
add.s64 %rd26, %rd24, %rd5;
add.s64 %rd8, %rd26, 2;
add.s64 %rd27, %rd4, %rd21;
add.s64 %rd9, %rd27, 4;
cvta.to.global.u64 %rd10, %rd17;
mov.f32 %f353, 0f00000000;

$L__BB249_2:
cvt.u64.u32 %rd28, %r2104;
add.s64 %rd29, %rd1, %rd28;
add.s64 %rd30, %rd29, %rd3;
mul.lo.s64 %rd31, %rd30, 36;
add.s64 %rd32, %rd184, %rd6;
ld.global.nc.u16 %rs17, [%rd32+-4];
ld.global.nc.u16 %rs18, [%rd32+-2];
ld.global.nc.u16 %rs19, [%rd32];
add.s64 %rd33, %rd184, %rd8;
ld.global.nc.u16 %rs20, [%rd33];
cvt.u32.u16 %r870, %rs20;
ld.global.nc.u16 %rs21, [%rd33+2];
cvt.u32.u16 %r871, %rs21;
mov.b32 %r872, {%rs20, %rs21};
ld.global.nc.u16 %rs22, [%rd32+-6];
shl.b16 %rs23, %rs22, 2;
cvt.u64.u16 %rd34, %rs23;
and.b64 %rd35, %rd34, 1020;
mov.u64 %rd36, iq3xxs_grid;
add.s64 %rd37, %rd36, %rd35;
ld.global.nc.u32 %r873, [%rd37];
not.b32 %r874, %r873;
shr.u16 %rs24, %rs22, 8;
cvt.u32.u16 %r875, %rs24;
mul.wide.u32 %rd38, %r875, 4;
add.s64 %rd39, %rd36, %rd38;
ld.global.nc.u32 %r876, [%rd39];
not.b32 %r877, %r876;
and.b32 %r878, %r870, 255;
popc.b32 %r879, %r878;
shl.b32 %r880, %r879, 7;
and.b32 %r881, %r880, 128;
xor.b32 %r882, %r881, %r878;
mul.lo.s32 %r883, %r882, 16843009;
and.b32 %r884, %r883, 134480385;
or.b32 %r885, %r884, -2139062144;
add.s32 %r886, %r885, -16843009;
or.b32 %r772, %r886, %r884;
mov.u32 %r86, %r772;

prmt.b32 %r86, %r86, 0, 0xba98;

xor.b32 %r887, %r86, %r873;
or.b32 %r888, %r887, -2139062144;
and.b32 %r889, %r86, 2139062143;
and.b32 %r890, %r874, -2139062144;
sub.s32 %r891, %r888, %r889;
xor.b32 %r91, %r890, %r891;
add.s64 %rd40, %rd10, %rd31;
ld.global.nc.u32 %r92, [%rd40+4];
and.b32 %r892, %r883, -2143281136;
or.b32 %r893, %r892, -2139062144;
add.s32 %r894, %r893, -16843009;
or.b32 %r774, %r894, %r892;
mov.u32 %r88, %r774;

prmt.b32 %r88, %r88, 0, 0xba98;

xor.b32 %r895, %r88, %r876;
or.b32 %r896, %r895, -2139062144;
and.b32 %r897, %r88, 2139062143;
and.b32 %r898, %r877, -2139062144;
sub.s32 %r899, %r896, %r897;
xor.b32 %r95, %r898, %r899;
ld.global.nc.u32 %r96, [%rd40+8];
mov.u32 %r828, 0;

dp4a.s32.s32 %r90, %r91, %r92, %r828;


dp4a.s32.s32 %r94, %r95, %r96, %r90;

shl.b16 %rs25, %rs17, 2;
cvt.u64.u16 %rd41, %rs25;
and.b64 %rd42, %rd41, 1020;
add.s64 %rd43, %rd36, %rd42;
ld.global.nc.u32 %r900, [%rd43];
not.b32 %r901, %r900;
shr.u16 %rs26, %rs17, 8;
cvt.u32.u16 %r902, %rs26;
mul.wide.u32 %rd44, %r902, 4;
add.s64 %rd45, %rd36, %rd44;
ld.global.nc.u32 %r903, [%rd45];
not.b32 %r904, %r903;
shr.u32 %r905, %r870, 7;
and.b32 %r906, %r905, 255;
popc.b32 %r907, %r906;
shl.b32 %r908, %r907, 7;
and.b32 %r909, %r908, 128;
xor.b32 %r910, %r909, %r906;
mul.lo.s32 %r911, %r910, 16843009;
and.b32 %r912, %r911, 134480385;
or.b32 %r913, %r912, -2139062144;
add.s32 %r914, %r913, -16843009;
or.b32 %r784, %r914, %r912;
mov.u32 %r98, %r784;

prmt.b32 %r98, %r98, 0, 0xba98;

xor.b32 %r915, %r98, %r900;
or.b32 %r916, %r915, -2139062144;
and.b32 %r917, %r98, 2139062143;
and.b32 %r918, %r901, -2139062144;
sub.s32 %r919, %r916, %r917;
xor.b32 %r103, %r918, %r919;
ld.global.nc.u32 %r104, [%rd40+12];
and.b32 %r920, %r911, -2143281136;
or.b32 %r921, %r920, -2139062144;
add.s32 %r922, %r921, -16843009;
or.b32 %r786, %r922, %r920;
mov.u32 %r100, %r786;

prmt.b32 %r100, %r100, 0, 0xba98;

xor.b32 %r923, %r100, %r903;
or.b32 %r924, %r923, -2139062144;
and.b32 %r925, %r100, 2139062143;
and.b32 %r926, %r904, -2139062144;
sub.s32 %r927, %r924, %r925;
xor.b32 %r107, %r926, %r927;
ld.global.nc.u32 %r108, [%rd40+16];

dp4a.s32.s32 %r102, %r103, %r104, %r94;


dp4a.s32.s32 %r106, %r107, %r108, %r102;

shl.b16 %rs27, %rs18, 2;
cvt.u64.u16 %rd46, %rs27;
and.b64 %rd47, %rd46, 1020;
add.s64 %rd48, %rd36, %rd47;
ld.global.nc.u32 %r928, [%rd48];
not.b32 %r929, %r928;
shr.u16 %rs28, %rs18, 8;
cvt.u32.u16 %r930, %rs28;
mul.wide.u32 %rd49, %r930, 4;
add.s64 %rd50, %rd36, %rd49;
ld.global.nc.u32 %r931, [%rd50];
not.b32 %r932, %r931;
shr.u32 %r933, %r872, 14;
and.b32 %r934, %r933, 255;
popc.b32 %r935, %r934;
shl.b32 %r936, %r935, 7;
and.b32 %r937, %r936, 128;
xor.b32 %r938, %r937, %r934;
mul.lo.s32 %r939, %r938, 16843009;
and.b32 %r940, %r939, 134480385;
or.b32 %r941, %r940, -2139062144;
add.s32 %r942, %r941, -16843009;
or.b32 %r796, %r942, %r940;
mov.u32 %r110, %r796;

prmt.b32 %r110, %r110, 0, 0xba98;

xor.b32 %r943, %r110, %r928;
or.b32 %r944, %r943, -2139062144;
and.b32 %r945, %r110, 2139062143;
and.b32 %r946, %r929, -2139062144;
sub.s32 %r947, %r944, %r945;
xor.b32 %r115, %r946, %r947;
ld.global.nc.u32 %r116, [%rd40+20];
and.b32 %r948, %r939, -2143281136;
or.b32 %r949, %r948, -2139062144;
add.s32 %r950, %r949, -16843009;
or.b32 %r798, %r950, %r948;
mov.u32 %r112, %r798;

prmt.b32 %r112, %r112, 0, 0xba98;

xor.b32 %r951, %r112, %r931;
or.b32 %r952, %r951, -2139062144;
and.b32 %r953, %r112, 2139062143;
and.b32 %r954, %r932, -2139062144;
sub.s32 %r955, %r952, %r953;
xor.b32 %r119, %r954, %r955;
ld.global.nc.u32 %r120, [%rd40+24];

dp4a.s32.s32 %r114, %r115, %r116, %r106;


dp4a.s32.s32 %r118, %r119, %r120, %r114;

shl.b16 %rs29, %rs19, 2;
cvt.u64.u16 %rd51, %rs29;
and.b64 %rd52, %rd51, 1020;
add.s64 %rd53, %rd36, %rd52;
ld.global.nc.u32 %r956, [%rd53];
not.b32 %r957, %r956;
shr.u16 %rs30, %rs19, 8;
cvt.u32.u16 %r958, %rs30;
mul.wide.u32 %rd54, %r958, 4;
add.s64 %rd55, %rd36, %rd54;
ld.global.nc.u32 %r959, [%rd55];
not.b32 %r960, %r959;
shr.u32 %r961, %r871, 5;
and.b32 %r962, %r961, 255;
popc.b32 %r963, %r962;
shl.b32 %r964, %r963, 7;
and.b32 %r965, %r964, 128;
xor.b32 %r966, %r965, %r962;
mul.lo.s32 %r967, %r966, 16843009;
and.b32 %r968, %r967, 134480385;
or.b32 %r969, %r968, -2139062144;
add.s32 %r970, %r969, -16843009;
or.b32 %r808, %r970, %r968;
mov.u32 %r122, %r808;

prmt.b32 %r122, %r122, 0, 0xba98;

xor.b32 %r971, %r122, %r956;
or.b32 %r972, %r971, -2139062144;
and.b32 %r973, %r122, 2139062143;
and.b32 %r974, %r957, -2139062144;
sub.s32 %r975, %r972, %r973;
xor.b32 %r127, %r974, %r975;
ld.global.nc.u32 %r128, [%rd40+28];
and.b32 %r976, %r967, -2143281136;
or.b32 %r977, %r976, -2139062144;
add.s32 %r978, %r977, -16843009;
or.b32 %r810, %r978, %r976;
mov.u32 %r124, %r810;

prmt.b32 %r124, %r124, 0, 0xba98;

xor.b32 %r979, %r124, %r959;
or.b32 %r980, %r979, -2139062144;
and.b32 %r981, %r124, 2139062143;
and.b32 %r982, %r960, -2139062144;
sub.s32 %r983, %r980, %r981;
xor.b32 %r131, %r982, %r983;
ld.global.nc.u32 %r132, [%rd40+32];

dp4a.s32.s32 %r126, %r127, %r128, %r118;

shr.u32 %r984, %r871, 12;

dp4a.s32.s32 %r130, %r131, %r132, %r126;

shr.u32 %r985, %r130, 31;
add.s32 %r986, %r130, %r985;
shr.s32 %r987, %r986, 1;
mad.lo.s32 %r988, %r130, %r984, %r987;
shr.u32 %r989, %r988, 31;
add.s32 %r990, %r988, %r989;
shr.s32 %r991, %r990, 1;
add.s64 %rd56, %rd184, %rd5;
ld.global.nc.u16 %rs1, [%rd56];

{ cvt.f32.f16 %f97, %rs1;}


ld.global.nc.u32 %r134, [%rd40];

{.reg .f16 low,high;
mov.b32 {low,high},%r134;
cvt.f32.f16 %f98, low;}


mul.ftz.f32 %f129, %f97, %f98;
cvt.rn.f32.s32 %f130, %r991;
fma.rn.ftz.f32 %f368, %f129, %f130, %f368;
add.s64 %rd57, %rd184, %rd9;
ld.global.nc.u16 %rs31, [%rd57];
ld.global.nc.u16 %rs32, [%rd57+2];
ld.global.nc.u16 %rs33, [%rd57+4];
add.s64 %rd58, %rd184, %rd7;
ld.global.nc.u16 %rs34, [%rd58];
cvt.u32.u16 %r992, %rs34;
ld.global.nc.u16 %rs35, [%rd58+2];
cvt.u32.u16 %r993, %rs35;
mov.b32 %r994, {%rs34, %rs35};
ld.global.nc.u16 %rs36, [%rd57+-2];
shl.b16 %rs37, %rs36, 2;
cvt.u64.u16 %rd59, %rs37;
and.b64 %rd60, %rd59, 1020;
add.s64 %rd61, %rd36, %rd60;
ld.global.nc.u32 %r995, [%rd61];
not.b32 %r996, %r995;
shr.u16 %rs38, %rs36, 8;
cvt.u32.u16 %r997, %rs38;
mul.wide.u32 %rd62, %r997, 4;
add.s64 %rd63, %rd36, %rd62;
ld.global.nc.u32 %r998, [%rd63];
not.b32 %r999, %r998;
and.b32 %r1000, %r992, 255;
popc.b32 %r1001, %r1000;
shl.b32 %r1002, %r1001, 7;
and.b32 %r1003, %r1002, 128;
xor.b32 %r1004, %r1003, %r1000;
mul.lo.s32 %r1005, %r1004, 16843009;
and.b32 %r1006, %r1005, 134480385;
or.b32 %r1007, %r1006, -2139062144;
add.s32 %r1008, %r1007, -16843009;
or.b32 %r821, %r1008, %r1006;
mov.u32 %r135, %r821;

prmt.b32 %r135, %r135, 0, 0xba98;

xor.b32 %r1009, %r135, %r995;
or.b32 %r1010, %r1009, -2139062144;
and.b32 %r1011, %r135, 2139062143;
and.b32 %r1012, %r996, -2139062144;
sub.s32 %r1013, %r1010, %r1011;
xor.b32 %r140, %r1012, %r1013;
and.b32 %r1014, %r1005, -2143281136;
or.b32 %r1015, %r1014, -2139062144;
add.s32 %r1016, %r1015, -16843009;
or.b32 %r823, %r1016, %r1014;
mov.u32 %r137, %r823;

prmt.b32 %r137, %r137, 0, 0xba98;

xor.b32 %r1017, %r137, %r998;
or.b32 %r1018, %r1017, -2139062144;
and.b32 %r1019, %r137, 2139062143;
and.b32 %r1020, %r999, -2139062144;
sub.s32 %r1021, %r1018, %r1019;
xor.b32 %r144, %r1020, %r1021;

dp4a.s32.s32 %r139, %r140, %r92, %r828;


dp4a.s32.s32 %r143, %r144, %r96, %r139;

shl.b16 %rs39, %rs31, 2;
cvt.u64.u16 %rd64, %rs39;
and.b64 %rd65, %rd64, 1020;
add.s64 %rd66, %rd36, %rd65;
ld.global.nc.u32 %r1022, [%rd66];
not.b32 %r1023, %r1022;
shr.u16 %rs40, %rs31, 8;
cvt.u32.u16 %r1024, %rs40;
mul.wide.u32 %rd67, %r1024, 4;
add.s64 %rd68, %rd36, %rd67;
ld.global.nc.u32 %r1025, [%rd68];
not.b32 %r1026, %r1025;
shr.u32 %r1027, %r992, 7;
and.b32 %r1028, %r1027, 255;
popc.b32 %r1029, %r1028;
shl.b32 %r1030, %r1029, 7;
and.b32 %r1031, %r1030, 128;
xor.b32 %r1032, %r1031, %r1028;
mul.lo.s32 %r1033, %r1032, 16843009;
and.b32 %r1034, %r1033, 134480385;
or.b32 %r1035, %r1034, -2139062144;
add.s32 %r1036, %r1035, -16843009;
or.b32 %r833, %r1036, %r1034;
mov.u32 %r147, %r833;

prmt.b32 %r147, %r147, 0, 0xba98;

xor.b32 %r1037, %r147, %r1022;
or.b32 %r1038, %r1037, -2139062144;
and.b32 %r1039, %r147, 2139062143;
and.b32 %r1040, %r1023, -2139062144;
sub.s32 %r1041, %r1038, %r1039;
xor.b32 %r152, %r1040, %r1041;
and.b32 %r1042, %r1033, -2143281136;
or.b32 %r1043, %r1042, -2139062144;
add.s32 %r1044, %r1043, -16843009;
or.b32 %r835, %r1044, %r1042;
mov.u32 %r149, %r835;

prmt.b32 %r149, %r149, 0, 0xba98;

xor.b32 %r1045, %r149, %r1025;
or.b32 %r1046, %r1045, -2139062144;
and.b32 %r1047, %r149, 2139062143;
and.b32 %r1048, %r1026, -2139062144;
sub.s32 %r1049, %r1046, %r1047;
xor.b32 %r156, %r1048, %r1049;

dp4a.s32.s32 %r151, %r152, %r104, %r143;


dp4a.s32.s32 %r155, %r156, %r108, %r151;

shl.b16 %rs41, %rs32, 2;
cvt.u64.u16 %rd69, %rs41;
and.b64 %rd70, %rd69, 1020;
add.s64 %rd71, %rd36, %rd70;
ld.global.nc.u32 %r1050, [%rd71];
not.b32 %r1051, %r1050;
shr.u16 %rs42, %rs32, 8;
cvt.u32.u16 %r1052, %rs42;
mul.wide.u32 %rd72, %r1052, 4;
add.s64 %rd73, %rd36, %rd72;
ld.global.nc.u32 %r1053, [%rd73];
not.b32 %r1054, %r1053;
shr.u32 %r1055, %r994, 14;
and.b32 %r1056, %r1055, 255;
popc.b32 %r1057, %r1056;
shl.b32 %r1058, %r1057, 7;
and.b32 %r1059, %r1058, 128;
xor.b32 %r1060, %r1059, %r1056;
mul.lo.s32 %r1061, %r1060, 16843009;
and.b32 %r1062, %r1061, 134480385;
or.b32 %r1063, %r1062, -2139062144;
add.s32 %r1064, %r1063, -16843009;
or.b32 %r845, %r1064, %r1062;
mov.u32 %r159, %r845;

prmt.b32 %r159, %r159, 0, 0xba98;

xor.b32 %r1065, %r159, %r1050;
or.b32 %r1066, %r1065, -2139062144;
and.b32 %r1067, %r159, 2139062143;
and.b32 %r1068, %r1051, -2139062144;
sub.s32 %r1069, %r1066, %r1067;
xor.b32 %r164, %r1068, %r1069;
and.b32 %r1070, %r1061, -2143281136;
or.b32 %r1071, %r1070, -2139062144;
add.s32 %r1072, %r1071, -16843009;
or.b32 %r847, %r1072, %r1070;
mov.u32 %r161, %r847;

prmt.b32 %r161, %r161, 0, 0xba98;

xor.b32 %r1073, %r161, %r1053;
or.b32 %r1074, %r1073, -2139062144;
and.b32 %r1075, %r161, 2139062143;
and.b32 %r1076, %r1054, -2139062144;
sub.s32 %r1077, %r1074, %r1075;
xor.b32 %r168, %r1076, %r1077;

dp4a.s32.s32 %r163, %r164, %r116, %r155;


dp4a.s32.s32 %r167, %r168, %r120, %r163;

shl.b16 %rs43, %rs33, 2;
cvt.u64.u16 %rd74, %rs43;
and.b64 %rd75, %rd74, 1020;
add.s64 %rd76, %rd36, %rd75;
ld.global.nc.u32 %r1078, [%rd76];
not.b32 %r1079, %r1078;
shr.u16 %rs44, %rs33, 8;
cvt.u32.u16 %r1080, %rs44;
mul.wide.u32 %rd77, %r1080, 4;
add.s64 %rd78, %rd36, %rd77;
ld.global.nc.u32 %r1081, [%rd78];
not.b32 %r1082, %r1081;
shr.u32 %r1083, %r993, 5;
and.b32 %r1084, %r1083, 255;
popc.b32 %r1085, %r1084;
shl.b32 %r1086, %r1085, 7;
and.b32 %r1087, %r1086, 128;
xor.b32 %r1088, %r1087, %r1084;
mul.lo.s32 %r1089, %r1088, 16843009;
and.b32 %r1090, %r1089, 134480385;
or.b32 %r1091, %r1090, -2139062144;
add.s32 %r1092, %r1091, -16843009;
or.b32 %r857, %r1092, %r1090;
mov.u32 %r171, %r857;

prmt.b32 %r171, %r171, 0, 0xba98;

xor.b32 %r1093, %r171, %r1078;
or.b32 %r1094, %r1093, -2139062144;
and.b32 %r1095, %r171, 2139062143;
and.b32 %r1096, %r1079, -2139062144;
sub.s32 %r1097, %r1094, %r1095;
xor.b32 %r176, %r1096, %r1097;
and.b32 %r1098, %r1089, -2143281136;
or.b32 %r1099, %r1098, -2139062144;
add.s32 %r1100, %r1099, -16843009;
or.b32 %r859, %r1100, %r1098;
mov.u32 %r173, %r859;

prmt.b32 %r173, %r173, 0, 0xba98;

xor.b32 %r1101, %r173, %r1081;
or.b32 %r1102, %r1101, -2139062144;
and.b32 %r1103, %r173, 2139062143;
and.b32 %r1104, %r1082, -2139062144;
sub.s32 %r1105, %r1102, %r1103;
xor.b32 %r180, %r1104, %r1105;

dp4a.s32.s32 %r175, %r176, %r128, %r167;

shr.u32 %r1106, %r993, 12;

dp4a.s32.s32 %r179, %r180, %r132, %r175;

shr.u32 %r1107, %r179, 31;
add.s32 %r1108, %r179, %r1107;
shr.s32 %r1109, %r1108, 1;
mad.lo.s32 %r1110, %r179, %r1106, %r1109;
shr.u32 %r1111, %r1110, 31;
add.s32 %r1112, %r1110, %r1111;
shr.s32 %r1113, %r1112, 1;
add.s64 %rd79, %rd184, %rd4;
ld.global.nc.u16 %rs2, [%rd79];

{ cvt.f32.f16 %f99, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r134;
cvt.f32.f16 %f100, low;}


mul.ftz.f32 %f131, %f99, %f100;
cvt.rn.f32.s32 %f132, %r1113;
fma.rn.ftz.f32 %f367, %f131, %f132, %f367;
cvt.u64.u32 %rd80, %r2105;
add.s64 %rd81, %rd1, %rd80;
add.s64 %rd82, %rd81, %rd3;
mul.lo.s64 %rd83, %rd82, 36;
mov.u32 %r184, %r772;

prmt.b32 %r184, %r184, 0, 0xba98;

xor.b32 %r1114, %r184, %r873;
or.b32 %r1115, %r1114, -2139062144;
and.b32 %r1116, %r184, 2139062143;
sub.s32 %r1117, %r1115, %r1116;
xor.b32 %r189, %r890, %r1117;
add.s64 %rd84, %rd10, %rd83;
ld.global.nc.u32 %r190, [%rd84+4];
mov.u32 %r186, %r774;

prmt.b32 %r186, %r186, 0, 0xba98;

xor.b32 %r1118, %r186, %r876;
or.b32 %r1119, %r1118, -2139062144;
and.b32 %r1120, %r186, 2139062143;
sub.s32 %r1121, %r1119, %r1120;
xor.b32 %r193, %r898, %r1121;
ld.global.nc.u32 %r194, [%rd84+8];

dp4a.s32.s32 %r188, %r189, %r190, %r828;


dp4a.s32.s32 %r192, %r193, %r194, %r188;

mov.u32 %r196, %r784;

prmt.b32 %r196, %r196, 0, 0xba98;

xor.b32 %r1122, %r196, %r900;
or.b32 %r1123, %r1122, -2139062144;
and.b32 %r1124, %r196, 2139062143;
sub.s32 %r1125, %r1123, %r1124;
xor.b32 %r201, %r918, %r1125;
ld.global.nc.u32 %r202, [%rd84+12];
mov.u32 %r198, %r786;

prmt.b32 %r198, %r198, 0, 0xba98;

xor.b32 %r1126, %r198, %r903;
or.b32 %r1127, %r1126, -2139062144;
and.b32 %r1128, %r198, 2139062143;
sub.s32 %r1129, %r1127, %r1128;
xor.b32 %r205, %r926, %r1129;
ld.global.nc.u32 %r206, [%rd84+16];

dp4a.s32.s32 %r200, %r201, %r202, %r192;


dp4a.s32.s32 %r204, %r205, %r206, %r200;

mov.u32 %r208, %r796;

prmt.b32 %r208, %r208, 0, 0xba98;

xor.b32 %r1130, %r208, %r928;
or.b32 %r1131, %r1130, -2139062144;
and.b32 %r1132, %r208, 2139062143;
sub.s32 %r1133, %r1131, %r1132;
xor.b32 %r213, %r946, %r1133;
ld.global.nc.u32 %r214, [%rd84+20];
mov.u32 %r210, %r798;

prmt.b32 %r210, %r210, 0, 0xba98;

xor.b32 %r1134, %r210, %r931;
or.b32 %r1135, %r1134, -2139062144;
and.b32 %r1136, %r210, 2139062143;
sub.s32 %r1137, %r1135, %r1136;
xor.b32 %r217, %r954, %r1137;
ld.global.nc.u32 %r218, [%rd84+24];

dp4a.s32.s32 %r212, %r213, %r214, %r204;


dp4a.s32.s32 %r216, %r217, %r218, %r212;

mov.u32 %r220, %r808;

prmt.b32 %r220, %r220, 0, 0xba98;

xor.b32 %r1138, %r220, %r956;
or.b32 %r1139, %r1138, -2139062144;
and.b32 %r1140, %r220, 2139062143;
sub.s32 %r1141, %r1139, %r1140;
xor.b32 %r225, %r974, %r1141;
ld.global.nc.u32 %r226, [%rd84+28];
mov.u32 %r222, %r810;

prmt.b32 %r222, %r222, 0, 0xba98;

xor.b32 %r1142, %r222, %r959;
or.b32 %r1143, %r1142, -2139062144;
and.b32 %r1144, %r222, 2139062143;
sub.s32 %r1145, %r1143, %r1144;
xor.b32 %r229, %r982, %r1145;
ld.global.nc.u32 %r230, [%rd84+32];

dp4a.s32.s32 %r224, %r225, %r226, %r216;


dp4a.s32.s32 %r228, %r229, %r230, %r224;

shr.u32 %r1146, %r228, 31;
add.s32 %r1147, %r228, %r1146;
shr.s32 %r1148, %r1147, 1;
mad.lo.s32 %r1149, %r228, %r984, %r1148;
shr.u32 %r1150, %r1149, 31;
add.s32 %r1151, %r1149, %r1150;
shr.s32 %r1152, %r1151, 1;

{ cvt.f32.f16 %f101, %rs1;}


ld.global.nc.u32 %r232, [%rd84];

{.reg .f16 low,high;
mov.b32 {low,high},%r232;
cvt.f32.f16 %f102, low;}


mul.ftz.f32 %f133, %f101, %f102;
cvt.rn.f32.s32 %f134, %r1152;
fma.rn.ftz.f32 %f366, %f133, %f134, %f366;
mov.u32 %r233, %r821;

prmt.b32 %r233, %r233, 0, 0xba98;

xor.b32 %r1153, %r233, %r995;
or.b32 %r1154, %r1153, -2139062144;
and.b32 %r1155, %r233, 2139062143;
sub.s32 %r1156, %r1154, %r1155;
xor.b32 %r238, %r1012, %r1156;
mov.u32 %r235, %r823;

prmt.b32 %r235, %r235, 0, 0xba98;

xor.b32 %r1157, %r235, %r998;
or.b32 %r1158, %r1157, -2139062144;
and.b32 %r1159, %r235, 2139062143;
sub.s32 %r1160, %r1158, %r1159;
xor.b32 %r242, %r1020, %r1160;

dp4a.s32.s32 %r237, %r238, %r190, %r828;


dp4a.s32.s32 %r241, %r242, %r194, %r237;

mov.u32 %r245, %r833;

prmt.b32 %r245, %r245, 0, 0xba98;

xor.b32 %r1161, %r245, %r1022;
or.b32 %r1162, %r1161, -2139062144;
and.b32 %r1163, %r245, 2139062143;
sub.s32 %r1164, %r1162, %r1163;
xor.b32 %r250, %r1040, %r1164;
mov.u32 %r247, %r835;

prmt.b32 %r247, %r247, 0, 0xba98;

xor.b32 %r1165, %r247, %r1025;
or.b32 %r1166, %r1165, -2139062144;
and.b32 %r1167, %r247, 2139062143;
sub.s32 %r1168, %r1166, %r1167;
xor.b32 %r254, %r1048, %r1168;

dp4a.s32.s32 %r249, %r250, %r202, %r241;


dp4a.s32.s32 %r253, %r254, %r206, %r249;

mov.u32 %r257, %r845;

prmt.b32 %r257, %r257, 0, 0xba98;

xor.b32 %r1169, %r257, %r1050;
or.b32 %r1170, %r1169, -2139062144;
and.b32 %r1171, %r257, 2139062143;
sub.s32 %r1172, %r1170, %r1171;
xor.b32 %r262, %r1068, %r1172;
mov.u32 %r259, %r847;

prmt.b32 %r259, %r259, 0, 0xba98;

xor.b32 %r1173, %r259, %r1053;
or.b32 %r1174, %r1173, -2139062144;
and.b32 %r1175, %r259, 2139062143;
sub.s32 %r1176, %r1174, %r1175;
xor.b32 %r266, %r1076, %r1176;

dp4a.s32.s32 %r261, %r262, %r214, %r253;


dp4a.s32.s32 %r265, %r266, %r218, %r261;

mov.u32 %r269, %r857;

prmt.b32 %r269, %r269, 0, 0xba98;

xor.b32 %r1177, %r269, %r1078;
or.b32 %r1178, %r1177, -2139062144;
and.b32 %r1179, %r269, 2139062143;
sub.s32 %r1180, %r1178, %r1179;
xor.b32 %r274, %r1096, %r1180;
mov.u32 %r271, %r859;

prmt.b32 %r271, %r271, 0, 0xba98;

xor.b32 %r1181, %r271, %r1081;
or.b32 %r1182, %r1181, -2139062144;
and.b32 %r1183, %r271, 2139062143;
sub.s32 %r1184, %r1182, %r1183;
xor.b32 %r278, %r1104, %r1184;

dp4a.s32.s32 %r273, %r274, %r226, %r265;


dp4a.s32.s32 %r277, %r278, %r230, %r273;

shr.u32 %r1185, %r277, 31;
add.s32 %r1186, %r277, %r1185;
shr.s32 %r1187, %r1186, 1;
mad.lo.s32 %r1188, %r277, %r1106, %r1187;
shr.u32 %r1189, %r1188, 31;
add.s32 %r1190, %r1188, %r1189;
shr.s32 %r1191, %r1190, 1;

{ cvt.f32.f16 %f103, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r232;
cvt.f32.f16 %f104, low;}


mul.ftz.f32 %f135, %f103, %f104;
cvt.rn.f32.s32 %f136, %r1191;
fma.rn.ftz.f32 %f365, %f135, %f136, %f365;
cvt.u64.u32 %rd85, %r2111;
add.s64 %rd86, %rd1, %rd85;
add.s64 %rd87, %rd86, %rd3;
mul.lo.s64 %rd88, %rd87, 36;
mov.u32 %r282, %r772;

prmt.b32 %r282, %r282, 0, 0xba98;

xor.b32 %r1192, %r282, %r873;
or.b32 %r1193, %r1192, -2139062144;
and.b32 %r1194, %r282, 2139062143;
sub.s32 %r1195, %r1193, %r1194;
xor.b32 %r287, %r890, %r1195;
add.s64 %rd89, %rd10, %rd88;
ld.global.nc.u32 %r288, [%rd89+4];
mov.u32 %r284, %r774;

prmt.b32 %r284, %r284, 0, 0xba98;

xor.b32 %r1196, %r284, %r876;
or.b32 %r1197, %r1196, -2139062144;
and.b32 %r1198, %r284, 2139062143;
sub.s32 %r1199, %r1197, %r1198;
xor.b32 %r291, %r898, %r1199;
ld.global.nc.u32 %r292, [%rd89+8];

dp4a.s32.s32 %r286, %r287, %r288, %r828;


dp4a.s32.s32 %r290, %r291, %r292, %r286;

mov.u32 %r294, %r784;

prmt.b32 %r294, %r294, 0, 0xba98;

xor.b32 %r1200, %r294, %r900;
or.b32 %r1201, %r1200, -2139062144;
and.b32 %r1202, %r294, 2139062143;
sub.s32 %r1203, %r1201, %r1202;
xor.b32 %r299, %r918, %r1203;
ld.global.nc.u32 %r300, [%rd89+12];
mov.u32 %r296, %r786;

prmt.b32 %r296, %r296, 0, 0xba98;

xor.b32 %r1204, %r296, %r903;
or.b32 %r1205, %r1204, -2139062144;
and.b32 %r1206, %r296, 2139062143;
sub.s32 %r1207, %r1205, %r1206;
xor.b32 %r303, %r926, %r1207;
ld.global.nc.u32 %r304, [%rd89+16];

dp4a.s32.s32 %r298, %r299, %r300, %r290;


dp4a.s32.s32 %r302, %r303, %r304, %r298;

mov.u32 %r306, %r796;

prmt.b32 %r306, %r306, 0, 0xba98;

xor.b32 %r1208, %r306, %r928;
or.b32 %r1209, %r1208, -2139062144;
and.b32 %r1210, %r306, 2139062143;
sub.s32 %r1211, %r1209, %r1210;
xor.b32 %r311, %r946, %r1211;
ld.global.nc.u32 %r312, [%rd89+20];
mov.u32 %r308, %r798;

prmt.b32 %r308, %r308, 0, 0xba98;

xor.b32 %r1212, %r308, %r931;
or.b32 %r1213, %r1212, -2139062144;
and.b32 %r1214, %r308, 2139062143;
sub.s32 %r1215, %r1213, %r1214;
xor.b32 %r315, %r954, %r1215;
ld.global.nc.u32 %r316, [%rd89+24];

dp4a.s32.s32 %r310, %r311, %r312, %r302;


dp4a.s32.s32 %r314, %r315, %r316, %r310;

mov.u32 %r318, %r808;

prmt.b32 %r318, %r318, 0, 0xba98;

xor.b32 %r1216, %r318, %r956;
or.b32 %r1217, %r1216, -2139062144;
and.b32 %r1218, %r318, 2139062143;
sub.s32 %r1219, %r1217, %r1218;
xor.b32 %r323, %r974, %r1219;
ld.global.nc.u32 %r324, [%rd89+28];
mov.u32 %r320, %r810;

prmt.b32 %r320, %r320, 0, 0xba98;

xor.b32 %r1220, %r320, %r959;
or.b32 %r1221, %r1220, -2139062144;
and.b32 %r1222, %r320, 2139062143;
sub.s32 %r1223, %r1221, %r1222;
xor.b32 %r327, %r982, %r1223;
ld.global.nc.u32 %r328, [%rd89+32];

dp4a.s32.s32 %r322, %r323, %r324, %r314;


dp4a.s32.s32 %r326, %r327, %r328, %r322;

shr.u32 %r1224, %r326, 31;
add.s32 %r1225, %r326, %r1224;
shr.s32 %r1226, %r1225, 1;
mad.lo.s32 %r1227, %r326, %r984, %r1226;
shr.u32 %r1228, %r1227, 31;
add.s32 %r1229, %r1227, %r1228;
shr.s32 %r1230, %r1229, 1;

{ cvt.f32.f16 %f105, %rs1;}


ld.global.nc.u32 %r330, [%rd89];

{.reg .f16 low,high;
mov.b32 {low,high},%r330;
cvt.f32.f16 %f106, low;}


mul.ftz.f32 %f137, %f105, %f106;
cvt.rn.f32.s32 %f138, %r1230;
fma.rn.ftz.f32 %f364, %f137, %f138, %f364;
mov.u32 %r331, %r821;

prmt.b32 %r331, %r331, 0, 0xba98;

xor.b32 %r1231, %r331, %r995;
or.b32 %r1232, %r1231, -2139062144;
and.b32 %r1233, %r331, 2139062143;
sub.s32 %r1234, %r1232, %r1233;
xor.b32 %r336, %r1012, %r1234;
mov.u32 %r333, %r823;

prmt.b32 %r333, %r333, 0, 0xba98;

xor.b32 %r1235, %r333, %r998;
or.b32 %r1236, %r1235, -2139062144;
and.b32 %r1237, %r333, 2139062143;
sub.s32 %r1238, %r1236, %r1237;
xor.b32 %r340, %r1020, %r1238;

dp4a.s32.s32 %r335, %r336, %r288, %r828;


dp4a.s32.s32 %r339, %r340, %r292, %r335;

mov.u32 %r343, %r833;

prmt.b32 %r343, %r343, 0, 0xba98;

xor.b32 %r1239, %r343, %r1022;
or.b32 %r1240, %r1239, -2139062144;
and.b32 %r1241, %r343, 2139062143;
sub.s32 %r1242, %r1240, %r1241;
xor.b32 %r348, %r1040, %r1242;
mov.u32 %r345, %r835;

prmt.b32 %r345, %r345, 0, 0xba98;

xor.b32 %r1243, %r345, %r1025;
or.b32 %r1244, %r1243, -2139062144;
and.b32 %r1245, %r345, 2139062143;
sub.s32 %r1246, %r1244, %r1245;
xor.b32 %r352, %r1048, %r1246;

dp4a.s32.s32 %r347, %r348, %r300, %r339;


dp4a.s32.s32 %r351, %r352, %r304, %r347;

mov.u32 %r355, %r845;

prmt.b32 %r355, %r355, 0, 0xba98;

xor.b32 %r1247, %r355, %r1050;
or.b32 %r1248, %r1247, -2139062144;
and.b32 %r1249, %r355, 2139062143;
sub.s32 %r1250, %r1248, %r1249;
xor.b32 %r360, %r1068, %r1250;
mov.u32 %r357, %r847;

prmt.b32 %r357, %r357, 0, 0xba98;

xor.b32 %r1251, %r357, %r1053;
or.b32 %r1252, %r1251, -2139062144;
and.b32 %r1253, %r357, 2139062143;
sub.s32 %r1254, %r1252, %r1253;
xor.b32 %r364, %r1076, %r1254;

dp4a.s32.s32 %r359, %r360, %r312, %r351;


dp4a.s32.s32 %r363, %r364, %r316, %r359;

mov.u32 %r367, %r857;

prmt.b32 %r367, %r367, 0, 0xba98;

xor.b32 %r1255, %r367, %r1078;
or.b32 %r1256, %r1255, -2139062144;
and.b32 %r1257, %r367, 2139062143;
sub.s32 %r1258, %r1256, %r1257;
xor.b32 %r372, %r1096, %r1258;
mov.u32 %r369, %r859;

prmt.b32 %r369, %r369, 0, 0xba98;

xor.b32 %r1259, %r369, %r1081;
or.b32 %r1260, %r1259, -2139062144;
and.b32 %r1261, %r369, 2139062143;
sub.s32 %r1262, %r1260, %r1261;
xor.b32 %r376, %r1104, %r1262;

dp4a.s32.s32 %r371, %r372, %r324, %r363;


dp4a.s32.s32 %r375, %r376, %r328, %r371;

shr.u32 %r1263, %r375, 31;
add.s32 %r1264, %r375, %r1263;
shr.s32 %r1265, %r1264, 1;
mad.lo.s32 %r1266, %r375, %r1106, %r1265;
shr.u32 %r1267, %r1266, 31;
add.s32 %r1268, %r1266, %r1267;
shr.s32 %r1269, %r1268, 1;

{ cvt.f32.f16 %f107, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r330;
cvt.f32.f16 %f108, low;}


mul.ftz.f32 %f139, %f107, %f108;
cvt.rn.f32.s32 %f140, %r1269;
fma.rn.ftz.f32 %f363, %f139, %f140, %f363;
cvt.u64.u32 %rd90, %r2110;
add.s64 %rd91, %rd1, %rd90;
add.s64 %rd92, %rd91, %rd3;
mul.lo.s64 %rd93, %rd92, 36;
mov.u32 %r380, %r772;

prmt.b32 %r380, %r380, 0, 0xba98;

xor.b32 %r1270, %r380, %r873;
or.b32 %r1271, %r1270, -2139062144;
and.b32 %r1272, %r380, 2139062143;
sub.s32 %r1273, %r1271, %r1272;
xor.b32 %r385, %r890, %r1273;
add.s64 %rd94, %rd10, %rd93;
ld.global.nc.u32 %r386, [%rd94+4];
mov.u32 %r382, %r774;

prmt.b32 %r382, %r382, 0, 0xba98;

xor.b32 %r1274, %r382, %r876;
or.b32 %r1275, %r1274, -2139062144;
and.b32 %r1276, %r382, 2139062143;
sub.s32 %r1277, %r1275, %r1276;
xor.b32 %r389, %r898, %r1277;
ld.global.nc.u32 %r390, [%rd94+8];

dp4a.s32.s32 %r384, %r385, %r386, %r828;


dp4a.s32.s32 %r388, %r389, %r390, %r384;

mov.u32 %r392, %r784;

prmt.b32 %r392, %r392, 0, 0xba98;

xor.b32 %r1278, %r392, %r900;
or.b32 %r1279, %r1278, -2139062144;
and.b32 %r1280, %r392, 2139062143;
sub.s32 %r1281, %r1279, %r1280;
xor.b32 %r397, %r918, %r1281;
ld.global.nc.u32 %r398, [%rd94+12];
mov.u32 %r394, %r786;

prmt.b32 %r394, %r394, 0, 0xba98;

xor.b32 %r1282, %r394, %r903;
or.b32 %r1283, %r1282, -2139062144;
and.b32 %r1284, %r394, 2139062143;
sub.s32 %r1285, %r1283, %r1284;
xor.b32 %r401, %r926, %r1285;
ld.global.nc.u32 %r402, [%rd94+16];

dp4a.s32.s32 %r396, %r397, %r398, %r388;


dp4a.s32.s32 %r400, %r401, %r402, %r396;

mov.u32 %r404, %r796;

prmt.b32 %r404, %r404, 0, 0xba98;

xor.b32 %r1286, %r404, %r928;
or.b32 %r1287, %r1286, -2139062144;
and.b32 %r1288, %r404, 2139062143;
sub.s32 %r1289, %r1287, %r1288;
xor.b32 %r409, %r946, %r1289;
ld.global.nc.u32 %r410, [%rd94+20];
mov.u32 %r406, %r798;

prmt.b32 %r406, %r406, 0, 0xba98;

xor.b32 %r1290, %r406, %r931;
or.b32 %r1291, %r1290, -2139062144;
and.b32 %r1292, %r406, 2139062143;
sub.s32 %r1293, %r1291, %r1292;
xor.b32 %r413, %r954, %r1293;
ld.global.nc.u32 %r414, [%rd94+24];

dp4a.s32.s32 %r408, %r409, %r410, %r400;


dp4a.s32.s32 %r412, %r413, %r414, %r408;

mov.u32 %r416, %r808;

prmt.b32 %r416, %r416, 0, 0xba98;

xor.b32 %r1294, %r416, %r956;
or.b32 %r1295, %r1294, -2139062144;
and.b32 %r1296, %r416, 2139062143;
sub.s32 %r1297, %r1295, %r1296;
xor.b32 %r421, %r974, %r1297;
ld.global.nc.u32 %r422, [%rd94+28];
mov.u32 %r418, %r810;

prmt.b32 %r418, %r418, 0, 0xba98;

xor.b32 %r1298, %r418, %r959;
or.b32 %r1299, %r1298, -2139062144;
and.b32 %r1300, %r418, 2139062143;
sub.s32 %r1301, %r1299, %r1300;
xor.b32 %r425, %r982, %r1301;
ld.global.nc.u32 %r426, [%rd94+32];

dp4a.s32.s32 %r420, %r421, %r422, %r412;


dp4a.s32.s32 %r424, %r425, %r426, %r420;

shr.u32 %r1302, %r424, 31;
add.s32 %r1303, %r424, %r1302;
shr.s32 %r1304, %r1303, 1;
mad.lo.s32 %r1305, %r424, %r984, %r1304;
shr.u32 %r1306, %r1305, 31;
add.s32 %r1307, %r1305, %r1306;
shr.s32 %r1308, %r1307, 1;

{ cvt.f32.f16 %f109, %rs1;}


ld.global.nc.u32 %r428, [%rd94];

{.reg .f16 low,high;
mov.b32 {low,high},%r428;
cvt.f32.f16 %f110, low;}


mul.ftz.f32 %f141, %f109, %f110;
cvt.rn.f32.s32 %f142, %r1308;
fma.rn.ftz.f32 %f362, %f141, %f142, %f362;
mov.u32 %r429, %r821;

prmt.b32 %r429, %r429, 0, 0xba98;

xor.b32 %r1309, %r429, %r995;
or.b32 %r1310, %r1309, -2139062144;
and.b32 %r1311, %r429, 2139062143;
sub.s32 %r1312, %r1310, %r1311;
xor.b32 %r434, %r1012, %r1312;
mov.u32 %r431, %r823;

prmt.b32 %r431, %r431, 0, 0xba98;

xor.b32 %r1313, %r431, %r998;
or.b32 %r1314, %r1313, -2139062144;
and.b32 %r1315, %r431, 2139062143;
sub.s32 %r1316, %r1314, %r1315;
xor.b32 %r438, %r1020, %r1316;

dp4a.s32.s32 %r433, %r434, %r386, %r828;


dp4a.s32.s32 %r437, %r438, %r390, %r433;

mov.u32 %r441, %r833;

prmt.b32 %r441, %r441, 0, 0xba98;

xor.b32 %r1317, %r441, %r1022;
or.b32 %r1318, %r1317, -2139062144;
and.b32 %r1319, %r441, 2139062143;
sub.s32 %r1320, %r1318, %r1319;
xor.b32 %r446, %r1040, %r1320;
mov.u32 %r443, %r835;

prmt.b32 %r443, %r443, 0, 0xba98;

xor.b32 %r1321, %r443, %r1025;
or.b32 %r1322, %r1321, -2139062144;
and.b32 %r1323, %r443, 2139062143;
sub.s32 %r1324, %r1322, %r1323;
xor.b32 %r450, %r1048, %r1324;

dp4a.s32.s32 %r445, %r446, %r398, %r437;


dp4a.s32.s32 %r449, %r450, %r402, %r445;

mov.u32 %r453, %r845;

prmt.b32 %r453, %r453, 0, 0xba98;

xor.b32 %r1325, %r453, %r1050;
or.b32 %r1326, %r1325, -2139062144;
and.b32 %r1327, %r453, 2139062143;
sub.s32 %r1328, %r1326, %r1327;
xor.b32 %r458, %r1068, %r1328;
mov.u32 %r455, %r847;

prmt.b32 %r455, %r455, 0, 0xba98;

xor.b32 %r1329, %r455, %r1053;
or.b32 %r1330, %r1329, -2139062144;
and.b32 %r1331, %r455, 2139062143;
sub.s32 %r1332, %r1330, %r1331;
xor.b32 %r462, %r1076, %r1332;

dp4a.s32.s32 %r457, %r458, %r410, %r449;


dp4a.s32.s32 %r461, %r462, %r414, %r457;

mov.u32 %r465, %r857;

prmt.b32 %r465, %r465, 0, 0xba98;

xor.b32 %r1333, %r465, %r1078;
or.b32 %r1334, %r1333, -2139062144;
and.b32 %r1335, %r465, 2139062143;
sub.s32 %r1336, %r1334, %r1335;
xor.b32 %r470, %r1096, %r1336;
mov.u32 %r467, %r859;

prmt.b32 %r467, %r467, 0, 0xba98;

xor.b32 %r1337, %r467, %r1081;
or.b32 %r1338, %r1337, -2139062144;
and.b32 %r1339, %r467, 2139062143;
sub.s32 %r1340, %r1338, %r1339;
xor.b32 %r474, %r1104, %r1340;

dp4a.s32.s32 %r469, %r470, %r422, %r461;


dp4a.s32.s32 %r473, %r474, %r426, %r469;

shr.u32 %r1341, %r473, 31;
add.s32 %r1342, %r473, %r1341;
shr.s32 %r1343, %r1342, 1;
mad.lo.s32 %r1344, %r473, %r1106, %r1343;
shr.u32 %r1345, %r1344, 31;
add.s32 %r1346, %r1344, %r1345;
shr.s32 %r1347, %r1346, 1;

{ cvt.f32.f16 %f111, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r428;
cvt.f32.f16 %f112, low;}


mul.ftz.f32 %f143, %f111, %f112;
cvt.rn.f32.s32 %f144, %r1347;
fma.rn.ftz.f32 %f361, %f143, %f144, %f361;
cvt.u64.u32 %rd95, %r2109;
add.s64 %rd96, %rd1, %rd95;
add.s64 %rd97, %rd96, %rd3;
mul.lo.s64 %rd98, %rd97, 36;
mov.u32 %r478, %r772;

prmt.b32 %r478, %r478, 0, 0xba98;

xor.b32 %r1348, %r478, %r873;
or.b32 %r1349, %r1348, -2139062144;
and.b32 %r1350, %r478, 2139062143;
sub.s32 %r1351, %r1349, %r1350;
xor.b32 %r483, %r890, %r1351;
add.s64 %rd99, %rd10, %rd98;
ld.global.nc.u32 %r484, [%rd99+4];
mov.u32 %r480, %r774;

prmt.b32 %r480, %r480, 0, 0xba98;

xor.b32 %r1352, %r480, %r876;
or.b32 %r1353, %r1352, -2139062144;
and.b32 %r1354, %r480, 2139062143;
sub.s32 %r1355, %r1353, %r1354;
xor.b32 %r487, %r898, %r1355;
ld.global.nc.u32 %r488, [%rd99+8];

dp4a.s32.s32 %r482, %r483, %r484, %r828;


dp4a.s32.s32 %r486, %r487, %r488, %r482;

mov.u32 %r490, %r784;

prmt.b32 %r490, %r490, 0, 0xba98;

xor.b32 %r1356, %r490, %r900;
or.b32 %r1357, %r1356, -2139062144;
and.b32 %r1358, %r490, 2139062143;
sub.s32 %r1359, %r1357, %r1358;
xor.b32 %r495, %r918, %r1359;
ld.global.nc.u32 %r496, [%rd99+12];
mov.u32 %r492, %r786;

prmt.b32 %r492, %r492, 0, 0xba98;

xor.b32 %r1360, %r492, %r903;
or.b32 %r1361, %r1360, -2139062144;
and.b32 %r1362, %r492, 2139062143;
sub.s32 %r1363, %r1361, %r1362;
xor.b32 %r499, %r926, %r1363;
ld.global.nc.u32 %r500, [%rd99+16];

dp4a.s32.s32 %r494, %r495, %r496, %r486;


dp4a.s32.s32 %r498, %r499, %r500, %r494;

mov.u32 %r502, %r796;

prmt.b32 %r502, %r502, 0, 0xba98;

xor.b32 %r1364, %r502, %r928;
or.b32 %r1365, %r1364, -2139062144;
and.b32 %r1366, %r502, 2139062143;
sub.s32 %r1367, %r1365, %r1366;
xor.b32 %r507, %r946, %r1367;
ld.global.nc.u32 %r508, [%rd99+20];
mov.u32 %r504, %r798;

prmt.b32 %r504, %r504, 0, 0xba98;

xor.b32 %r1368, %r504, %r931;
or.b32 %r1369, %r1368, -2139062144;
and.b32 %r1370, %r504, 2139062143;
sub.s32 %r1371, %r1369, %r1370;
xor.b32 %r511, %r954, %r1371;
ld.global.nc.u32 %r512, [%rd99+24];

dp4a.s32.s32 %r506, %r507, %r508, %r498;


dp4a.s32.s32 %r510, %r511, %r512, %r506;

mov.u32 %r514, %r808;

prmt.b32 %r514, %r514, 0, 0xba98;

xor.b32 %r1372, %r514, %r956;
or.b32 %r1373, %r1372, -2139062144;
and.b32 %r1374, %r514, 2139062143;
sub.s32 %r1375, %r1373, %r1374;
xor.b32 %r519, %r974, %r1375;
ld.global.nc.u32 %r520, [%rd99+28];
mov.u32 %r516, %r810;

prmt.b32 %r516, %r516, 0, 0xba98;

xor.b32 %r1376, %r516, %r959;
or.b32 %r1377, %r1376, -2139062144;
and.b32 %r1378, %r516, 2139062143;
sub.s32 %r1379, %r1377, %r1378;
xor.b32 %r523, %r982, %r1379;
ld.global.nc.u32 %r524, [%rd99+32];

dp4a.s32.s32 %r518, %r519, %r520, %r510;


dp4a.s32.s32 %r522, %r523, %r524, %r518;

shr.u32 %r1380, %r522, 31;
add.s32 %r1381, %r522, %r1380;
shr.s32 %r1382, %r1381, 1;
mad.lo.s32 %r1383, %r522, %r984, %r1382;
shr.u32 %r1384, %r1383, 31;
add.s32 %r1385, %r1383, %r1384;
shr.s32 %r1386, %r1385, 1;

{ cvt.f32.f16 %f113, %rs1;}


ld.global.nc.u32 %r526, [%rd99];

{.reg .f16 low,high;
mov.b32 {low,high},%r526;
cvt.f32.f16 %f114, low;}


mul.ftz.f32 %f145, %f113, %f114;
cvt.rn.f32.s32 %f146, %r1386;
fma.rn.ftz.f32 %f360, %f145, %f146, %f360;
mov.u32 %r527, %r821;

prmt.b32 %r527, %r527, 0, 0xba98;

xor.b32 %r1387, %r527, %r995;
or.b32 %r1388, %r1387, -2139062144;
and.b32 %r1389, %r527, 2139062143;
sub.s32 %r1390, %r1388, %r1389;
xor.b32 %r532, %r1012, %r1390;
mov.u32 %r529, %r823;

prmt.b32 %r529, %r529, 0, 0xba98;

xor.b32 %r1391, %r529, %r998;
or.b32 %r1392, %r1391, -2139062144;
and.b32 %r1393, %r529, 2139062143;
sub.s32 %r1394, %r1392, %r1393;
xor.b32 %r536, %r1020, %r1394;

dp4a.s32.s32 %r531, %r532, %r484, %r828;


dp4a.s32.s32 %r535, %r536, %r488, %r531;

mov.u32 %r539, %r833;

prmt.b32 %r539, %r539, 0, 0xba98;

xor.b32 %r1395, %r539, %r1022;
or.b32 %r1396, %r1395, -2139062144;
and.b32 %r1397, %r539, 2139062143;
sub.s32 %r1398, %r1396, %r1397;
xor.b32 %r544, %r1040, %r1398;
mov.u32 %r541, %r835;

prmt.b32 %r541, %r541, 0, 0xba98;

xor.b32 %r1399, %r541, %r1025;
or.b32 %r1400, %r1399, -2139062144;
and.b32 %r1401, %r541, 2139062143;
sub.s32 %r1402, %r1400, %r1401;
xor.b32 %r548, %r1048, %r1402;

dp4a.s32.s32 %r543, %r544, %r496, %r535;


dp4a.s32.s32 %r547, %r548, %r500, %r543;

mov.u32 %r551, %r845;

prmt.b32 %r551, %r551, 0, 0xba98;

xor.b32 %r1403, %r551, %r1050;
or.b32 %r1404, %r1403, -2139062144;
and.b32 %r1405, %r551, 2139062143;
sub.s32 %r1406, %r1404, %r1405;
xor.b32 %r556, %r1068, %r1406;
mov.u32 %r553, %r847;

prmt.b32 %r553, %r553, 0, 0xba98;

xor.b32 %r1407, %r553, %r1053;
or.b32 %r1408, %r1407, -2139062144;
and.b32 %r1409, %r553, 2139062143;
sub.s32 %r1410, %r1408, %r1409;
xor.b32 %r560, %r1076, %r1410;

dp4a.s32.s32 %r555, %r556, %r508, %r547;


dp4a.s32.s32 %r559, %r560, %r512, %r555;

mov.u32 %r563, %r857;

prmt.b32 %r563, %r563, 0, 0xba98;

xor.b32 %r1411, %r563, %r1078;
or.b32 %r1412, %r1411, -2139062144;
and.b32 %r1413, %r563, 2139062143;
sub.s32 %r1414, %r1412, %r1413;
xor.b32 %r568, %r1096, %r1414;
mov.u32 %r565, %r859;

prmt.b32 %r565, %r565, 0, 0xba98;

xor.b32 %r1415, %r565, %r1081;
or.b32 %r1416, %r1415, -2139062144;
and.b32 %r1417, %r565, 2139062143;
sub.s32 %r1418, %r1416, %r1417;
xor.b32 %r572, %r1104, %r1418;

dp4a.s32.s32 %r567, %r568, %r520, %r559;


dp4a.s32.s32 %r571, %r572, %r524, %r567;

shr.u32 %r1419, %r571, 31;
add.s32 %r1420, %r571, %r1419;
shr.s32 %r1421, %r1420, 1;
mad.lo.s32 %r1422, %r571, %r1106, %r1421;
shr.u32 %r1423, %r1422, 31;
add.s32 %r1424, %r1422, %r1423;
shr.s32 %r1425, %r1424, 1;

{ cvt.f32.f16 %f115, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r526;
cvt.f32.f16 %f116, low;}


mul.ftz.f32 %f147, %f115, %f116;
cvt.rn.f32.s32 %f148, %r1425;
fma.rn.ftz.f32 %f359, %f147, %f148, %f359;
cvt.u64.u32 %rd100, %r2108;
add.s64 %rd101, %rd1, %rd100;
add.s64 %rd102, %rd101, %rd3;
mul.lo.s64 %rd103, %rd102, 36;
mov.u32 %r576, %r772;

prmt.b32 %r576, %r576, 0, 0xba98;

xor.b32 %r1426, %r576, %r873;
or.b32 %r1427, %r1426, -2139062144;
and.b32 %r1428, %r576, 2139062143;
sub.s32 %r1429, %r1427, %r1428;
xor.b32 %r581, %r890, %r1429;
add.s64 %rd104, %rd10, %rd103;
ld.global.nc.u32 %r582, [%rd104+4];
mov.u32 %r578, %r774;

prmt.b32 %r578, %r578, 0, 0xba98;

xor.b32 %r1430, %r578, %r876;
or.b32 %r1431, %r1430, -2139062144;
and.b32 %r1432, %r578, 2139062143;
sub.s32 %r1433, %r1431, %r1432;
xor.b32 %r585, %r898, %r1433;
ld.global.nc.u32 %r586, [%rd104+8];

dp4a.s32.s32 %r580, %r581, %r582, %r828;


dp4a.s32.s32 %r584, %r585, %r586, %r580;

mov.u32 %r588, %r784;

prmt.b32 %r588, %r588, 0, 0xba98;

xor.b32 %r1434, %r588, %r900;
or.b32 %r1435, %r1434, -2139062144;
and.b32 %r1436, %r588, 2139062143;
sub.s32 %r1437, %r1435, %r1436;
xor.b32 %r593, %r918, %r1437;
ld.global.nc.u32 %r594, [%rd104+12];
mov.u32 %r590, %r786;

prmt.b32 %r590, %r590, 0, 0xba98;

xor.b32 %r1438, %r590, %r903;
or.b32 %r1439, %r1438, -2139062144;
and.b32 %r1440, %r590, 2139062143;
sub.s32 %r1441, %r1439, %r1440;
xor.b32 %r597, %r926, %r1441;
ld.global.nc.u32 %r598, [%rd104+16];

dp4a.s32.s32 %r592, %r593, %r594, %r584;


dp4a.s32.s32 %r596, %r597, %r598, %r592;

mov.u32 %r600, %r796;

prmt.b32 %r600, %r600, 0, 0xba98;

xor.b32 %r1442, %r600, %r928;
or.b32 %r1443, %r1442, -2139062144;
and.b32 %r1444, %r600, 2139062143;
sub.s32 %r1445, %r1443, %r1444;
xor.b32 %r605, %r946, %r1445;
ld.global.nc.u32 %r606, [%rd104+20];
mov.u32 %r602, %r798;

prmt.b32 %r602, %r602, 0, 0xba98;

xor.b32 %r1446, %r602, %r931;
or.b32 %r1447, %r1446, -2139062144;
and.b32 %r1448, %r602, 2139062143;
sub.s32 %r1449, %r1447, %r1448;
xor.b32 %r609, %r954, %r1449;
ld.global.nc.u32 %r610, [%rd104+24];

dp4a.s32.s32 %r604, %r605, %r606, %r596;


dp4a.s32.s32 %r608, %r609, %r610, %r604;

mov.u32 %r612, %r808;

prmt.b32 %r612, %r612, 0, 0xba98;

xor.b32 %r1450, %r612, %r956;
or.b32 %r1451, %r1450, -2139062144;
and.b32 %r1452, %r612, 2139062143;
sub.s32 %r1453, %r1451, %r1452;
xor.b32 %r617, %r974, %r1453;
ld.global.nc.u32 %r618, [%rd104+28];
mov.u32 %r614, %r810;

prmt.b32 %r614, %r614, 0, 0xba98;

xor.b32 %r1454, %r614, %r959;
or.b32 %r1455, %r1454, -2139062144;
and.b32 %r1456, %r614, 2139062143;
sub.s32 %r1457, %r1455, %r1456;
xor.b32 %r621, %r982, %r1457;
ld.global.nc.u32 %r622, [%rd104+32];

dp4a.s32.s32 %r616, %r617, %r618, %r608;


dp4a.s32.s32 %r620, %r621, %r622, %r616;

shr.u32 %r1458, %r620, 31;
add.s32 %r1459, %r620, %r1458;
shr.s32 %r1460, %r1459, 1;
mad.lo.s32 %r1461, %r620, %r984, %r1460;
shr.u32 %r1462, %r1461, 31;
add.s32 %r1463, %r1461, %r1462;
shr.s32 %r1464, %r1463, 1;

{ cvt.f32.f16 %f117, %rs1;}


ld.global.nc.u32 %r624, [%rd104];

{.reg .f16 low,high;
mov.b32 {low,high},%r624;
cvt.f32.f16 %f118, low;}


mul.ftz.f32 %f149, %f117, %f118;
cvt.rn.f32.s32 %f150, %r1464;
fma.rn.ftz.f32 %f358, %f149, %f150, %f358;
mov.u32 %r625, %r821;

prmt.b32 %r625, %r625, 0, 0xba98;

xor.b32 %r1465, %r625, %r995;
or.b32 %r1466, %r1465, -2139062144;
and.b32 %r1467, %r625, 2139062143;
sub.s32 %r1468, %r1466, %r1467;
xor.b32 %r630, %r1012, %r1468;
mov.u32 %r627, %r823;

prmt.b32 %r627, %r627, 0, 0xba98;

xor.b32 %r1469, %r627, %r998;
or.b32 %r1470, %r1469, -2139062144;
and.b32 %r1471, %r627, 2139062143;
sub.s32 %r1472, %r1470, %r1471;
xor.b32 %r634, %r1020, %r1472;

dp4a.s32.s32 %r629, %r630, %r582, %r828;


dp4a.s32.s32 %r633, %r634, %r586, %r629;

mov.u32 %r637, %r833;

prmt.b32 %r637, %r637, 0, 0xba98;

xor.b32 %r1473, %r637, %r1022;
or.b32 %r1474, %r1473, -2139062144;
and.b32 %r1475, %r637, 2139062143;
sub.s32 %r1476, %r1474, %r1475;
xor.b32 %r642, %r1040, %r1476;
mov.u32 %r639, %r835;

prmt.b32 %r639, %r639, 0, 0xba98;

xor.b32 %r1477, %r639, %r1025;
or.b32 %r1478, %r1477, -2139062144;
and.b32 %r1479, %r639, 2139062143;
sub.s32 %r1480, %r1478, %r1479;
xor.b32 %r646, %r1048, %r1480;

dp4a.s32.s32 %r641, %r642, %r594, %r633;


dp4a.s32.s32 %r645, %r646, %r598, %r641;

mov.u32 %r649, %r845;

prmt.b32 %r649, %r649, 0, 0xba98;

xor.b32 %r1481, %r649, %r1050;
or.b32 %r1482, %r1481, -2139062144;
and.b32 %r1483, %r649, 2139062143;
sub.s32 %r1484, %r1482, %r1483;
xor.b32 %r654, %r1068, %r1484;
mov.u32 %r651, %r847;

prmt.b32 %r651, %r651, 0, 0xba98;

xor.b32 %r1485, %r651, %r1053;
or.b32 %r1486, %r1485, -2139062144;
and.b32 %r1487, %r651, 2139062143;
sub.s32 %r1488, %r1486, %r1487;
xor.b32 %r658, %r1076, %r1488;

dp4a.s32.s32 %r653, %r654, %r606, %r645;


dp4a.s32.s32 %r657, %r658, %r610, %r653;

mov.u32 %r661, %r857;

prmt.b32 %r661, %r661, 0, 0xba98;

xor.b32 %r1489, %r661, %r1078;
or.b32 %r1490, %r1489, -2139062144;
and.b32 %r1491, %r661, 2139062143;
sub.s32 %r1492, %r1490, %r1491;
xor.b32 %r666, %r1096, %r1492;
mov.u32 %r663, %r859;

prmt.b32 %r663, %r663, 0, 0xba98;

xor.b32 %r1493, %r663, %r1081;
or.b32 %r1494, %r1493, -2139062144;
and.b32 %r1495, %r663, 2139062143;
sub.s32 %r1496, %r1494, %r1495;
xor.b32 %r670, %r1104, %r1496;

dp4a.s32.s32 %r665, %r666, %r618, %r657;


dp4a.s32.s32 %r669, %r670, %r622, %r665;

shr.u32 %r1497, %r669, 31;
add.s32 %r1498, %r669, %r1497;
shr.s32 %r1499, %r1498, 1;
mad.lo.s32 %r1500, %r669, %r1106, %r1499;
shr.u32 %r1501, %r1500, 31;
add.s32 %r1502, %r1500, %r1501;
shr.s32 %r1503, %r1502, 1;

{ cvt.f32.f16 %f119, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r624;
cvt.f32.f16 %f120, low;}


mul.ftz.f32 %f151, %f119, %f120;
cvt.rn.f32.s32 %f152, %r1503;
fma.rn.ftz.f32 %f357, %f151, %f152, %f357;
cvt.u64.u32 %rd105, %r2107;
add.s64 %rd106, %rd1, %rd105;
add.s64 %rd107, %rd106, %rd3;
mul.lo.s64 %rd108, %rd107, 36;
mov.u32 %r674, %r772;

prmt.b32 %r674, %r674, 0, 0xba98;

xor.b32 %r1504, %r674, %r873;
or.b32 %r1505, %r1504, -2139062144;
and.b32 %r1506, %r674, 2139062143;
sub.s32 %r1507, %r1505, %r1506;
xor.b32 %r679, %r890, %r1507;
add.s64 %rd109, %rd10, %rd108;
ld.global.nc.u32 %r680, [%rd109+4];
mov.u32 %r676, %r774;

prmt.b32 %r676, %r676, 0, 0xba98;

xor.b32 %r1508, %r676, %r876;
or.b32 %r1509, %r1508, -2139062144;
and.b32 %r1510, %r676, 2139062143;
sub.s32 %r1511, %r1509, %r1510;
xor.b32 %r683, %r898, %r1511;
ld.global.nc.u32 %r684, [%rd109+8];

dp4a.s32.s32 %r678, %r679, %r680, %r828;


dp4a.s32.s32 %r682, %r683, %r684, %r678;

mov.u32 %r686, %r784;

prmt.b32 %r686, %r686, 0, 0xba98;

xor.b32 %r1512, %r686, %r900;
or.b32 %r1513, %r1512, -2139062144;
and.b32 %r1514, %r686, 2139062143;
sub.s32 %r1515, %r1513, %r1514;
xor.b32 %r691, %r918, %r1515;
ld.global.nc.u32 %r692, [%rd109+12];
mov.u32 %r688, %r786;

prmt.b32 %r688, %r688, 0, 0xba98;

xor.b32 %r1516, %r688, %r903;
or.b32 %r1517, %r1516, -2139062144;
and.b32 %r1518, %r688, 2139062143;
sub.s32 %r1519, %r1517, %r1518;
xor.b32 %r695, %r926, %r1519;
ld.global.nc.u32 %r696, [%rd109+16];

dp4a.s32.s32 %r690, %r691, %r692, %r682;


dp4a.s32.s32 %r694, %r695, %r696, %r690;

mov.u32 %r698, %r796;

prmt.b32 %r698, %r698, 0, 0xba98;

xor.b32 %r1520, %r698, %r928;
or.b32 %r1521, %r1520, -2139062144;
and.b32 %r1522, %r698, 2139062143;
sub.s32 %r1523, %r1521, %r1522;
xor.b32 %r703, %r946, %r1523;
ld.global.nc.u32 %r704, [%rd109+20];
mov.u32 %r700, %r798;

prmt.b32 %r700, %r700, 0, 0xba98;

xor.b32 %r1524, %r700, %r931;
or.b32 %r1525, %r1524, -2139062144;
and.b32 %r1526, %r700, 2139062143;
sub.s32 %r1527, %r1525, %r1526;
xor.b32 %r707, %r954, %r1527;
ld.global.nc.u32 %r708, [%rd109+24];

dp4a.s32.s32 %r702, %r703, %r704, %r694;


dp4a.s32.s32 %r706, %r707, %r708, %r702;

mov.u32 %r710, %r808;

prmt.b32 %r710, %r710, 0, 0xba98;

xor.b32 %r1528, %r710, %r956;
or.b32 %r1529, %r1528, -2139062144;
and.b32 %r1530, %r710, 2139062143;
sub.s32 %r1531, %r1529, %r1530;
xor.b32 %r715, %r974, %r1531;
ld.global.nc.u32 %r716, [%rd109+28];
mov.u32 %r712, %r810;

prmt.b32 %r712, %r712, 0, 0xba98;

xor.b32 %r1532, %r712, %r959;
or.b32 %r1533, %r1532, -2139062144;
and.b32 %r1534, %r712, 2139062143;
sub.s32 %r1535, %r1533, %r1534;
xor.b32 %r719, %r982, %r1535;
ld.global.nc.u32 %r720, [%rd109+32];

dp4a.s32.s32 %r714, %r715, %r716, %r706;


dp4a.s32.s32 %r718, %r719, %r720, %r714;

shr.u32 %r1536, %r718, 31;
add.s32 %r1537, %r718, %r1536;
shr.s32 %r1538, %r1537, 1;
mad.lo.s32 %r1539, %r718, %r984, %r1538;
shr.u32 %r1540, %r1539, 31;
add.s32 %r1541, %r1539, %r1540;
shr.s32 %r1542, %r1541, 1;

{ cvt.f32.f16 %f121, %rs1;}


ld.global.nc.u32 %r722, [%rd109];

{.reg .f16 low,high;
mov.b32 {low,high},%r722;
cvt.f32.f16 %f122, low;}


mul.ftz.f32 %f153, %f121, %f122;
cvt.rn.f32.s32 %f154, %r1542;
fma.rn.ftz.f32 %f356, %f153, %f154, %f356;
mov.u32 %r723, %r821;

prmt.b32 %r723, %r723, 0, 0xba98;

xor.b32 %r1543, %r723, %r995;
or.b32 %r1544, %r1543, -2139062144;
and.b32 %r1545, %r723, 2139062143;
sub.s32 %r1546, %r1544, %r1545;
xor.b32 %r728, %r1012, %r1546;
mov.u32 %r725, %r823;

prmt.b32 %r725, %r725, 0, 0xba98;

xor.b32 %r1547, %r725, %r998;
or.b32 %r1548, %r1547, -2139062144;
and.b32 %r1549, %r725, 2139062143;
sub.s32 %r1550, %r1548, %r1549;
xor.b32 %r732, %r1020, %r1550;

dp4a.s32.s32 %r727, %r728, %r680, %r828;


dp4a.s32.s32 %r731, %r732, %r684, %r727;

mov.u32 %r735, %r833;

prmt.b32 %r735, %r735, 0, 0xba98;

xor.b32 %r1551, %r735, %r1022;
or.b32 %r1552, %r1551, -2139062144;
and.b32 %r1553, %r735, 2139062143;
sub.s32 %r1554, %r1552, %r1553;
xor.b32 %r740, %r1040, %r1554;
mov.u32 %r737, %r835;

prmt.b32 %r737, %r737, 0, 0xba98;

xor.b32 %r1555, %r737, %r1025;
or.b32 %r1556, %r1555, -2139062144;
and.b32 %r1557, %r737, 2139062143;
sub.s32 %r1558, %r1556, %r1557;
xor.b32 %r744, %r1048, %r1558;

dp4a.s32.s32 %r739, %r740, %r692, %r731;


dp4a.s32.s32 %r743, %r744, %r696, %r739;

mov.u32 %r747, %r845;

prmt.b32 %r747, %r747, 0, 0xba98;

xor.b32 %r1559, %r747, %r1050;
or.b32 %r1560, %r1559, -2139062144;
and.b32 %r1561, %r747, 2139062143;
sub.s32 %r1562, %r1560, %r1561;
xor.b32 %r752, %r1068, %r1562;
mov.u32 %r749, %r847;

prmt.b32 %r749, %r749, 0, 0xba98;

xor.b32 %r1563, %r749, %r1053;
or.b32 %r1564, %r1563, -2139062144;
and.b32 %r1565, %r749, 2139062143;
sub.s32 %r1566, %r1564, %r1565;
xor.b32 %r756, %r1076, %r1566;

dp4a.s32.s32 %r751, %r752, %r704, %r743;


dp4a.s32.s32 %r755, %r756, %r708, %r751;

mov.u32 %r759, %r857;

prmt.b32 %r759, %r759, 0, 0xba98;

xor.b32 %r1567, %r759, %r1078;
or.b32 %r1568, %r1567, -2139062144;
and.b32 %r1569, %r759, 2139062143;
sub.s32 %r1570, %r1568, %r1569;
xor.b32 %r764, %r1096, %r1570;
mov.u32 %r761, %r859;

prmt.b32 %r761, %r761, 0, 0xba98;

xor.b32 %r1571, %r761, %r1081;
or.b32 %r1572, %r1571, -2139062144;
and.b32 %r1573, %r761, 2139062143;
sub.s32 %r1574, %r1572, %r1573;
xor.b32 %r768, %r1104, %r1574;

dp4a.s32.s32 %r763, %r764, %r716, %r755;


dp4a.s32.s32 %r767, %r768, %r720, %r763;

shr.u32 %r1575, %r767, 31;
add.s32 %r1576, %r767, %r1575;
shr.s32 %r1577, %r1576, 1;
mad.lo.s32 %r1578, %r767, %r1106, %r1577;
shr.u32 %r1579, %r1578, 31;
add.s32 %r1580, %r1578, %r1579;
shr.s32 %r1581, %r1580, 1;

{ cvt.f32.f16 %f123, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r722;
cvt.f32.f16 %f124, low;}


mul.ftz.f32 %f155, %f123, %f124;
cvt.rn.f32.s32 %f156, %r1581;
fma.rn.ftz.f32 %f355, %f155, %f156, %f355;
cvt.u64.u32 %rd110, %r2106;
add.s64 %rd111, %rd1, %rd110;
add.s64 %rd112, %rd111, %rd3;
mul.lo.s64 %rd113, %rd112, 36;

prmt.b32 %r772, %r772, 0, 0xba98;

xor.b32 %r1582, %r772, %r873;
or.b32 %r1583, %r1582, -2139062144;
and.b32 %r1584, %r772, 2139062143;
sub.s32 %r1585, %r1583, %r1584;
xor.b32 %r777, %r890, %r1585;
add.s64 %rd114, %rd10, %rd113;
ld.global.nc.u32 %r778, [%rd114+4];

prmt.b32 %r774, %r774, 0, 0xba98;

xor.b32 %r1586, %r774, %r876;
or.b32 %r1587, %r1586, -2139062144;
and.b32 %r1588, %r774, 2139062143;
sub.s32 %r1589, %r1587, %r1588;
xor.b32 %r781, %r898, %r1589;
ld.global.nc.u32 %r782, [%rd114+8];

dp4a.s32.s32 %r776, %r777, %r778, %r828;


dp4a.s32.s32 %r780, %r781, %r782, %r776;


prmt.b32 %r784, %r784, 0, 0xba98;

xor.b32 %r1590, %r784, %r900;
or.b32 %r1591, %r1590, -2139062144;
and.b32 %r1592, %r784, 2139062143;
sub.s32 %r1593, %r1591, %r1592;
xor.b32 %r789, %r918, %r1593;
ld.global.nc.u32 %r790, [%rd114+12];

prmt.b32 %r786, %r786, 0, 0xba98;

xor.b32 %r1594, %r786, %r903;
or.b32 %r1595, %r1594, -2139062144;
and.b32 %r1596, %r786, 2139062143;
sub.s32 %r1597, %r1595, %r1596;
xor.b32 %r793, %r926, %r1597;
ld.global.nc.u32 %r794, [%rd114+16];

dp4a.s32.s32 %r788, %r789, %r790, %r780;


dp4a.s32.s32 %r792, %r793, %r794, %r788;


prmt.b32 %r796, %r796, 0, 0xba98;

xor.b32 %r1598, %r796, %r928;
or.b32 %r1599, %r1598, -2139062144;
and.b32 %r1600, %r796, 2139062143;
sub.s32 %r1601, %r1599, %r1600;
xor.b32 %r801, %r946, %r1601;
ld.global.nc.u32 %r802, [%rd114+20];

prmt.b32 %r798, %r798, 0, 0xba98;

xor.b32 %r1602, %r798, %r931;
or.b32 %r1603, %r1602, -2139062144;
and.b32 %r1604, %r798, 2139062143;
sub.s32 %r1605, %r1603, %r1604;
xor.b32 %r805, %r954, %r1605;
ld.global.nc.u32 %r806, [%rd114+24];

dp4a.s32.s32 %r800, %r801, %r802, %r792;


dp4a.s32.s32 %r804, %r805, %r806, %r800;


prmt.b32 %r808, %r808, 0, 0xba98;

xor.b32 %r1606, %r808, %r956;
or.b32 %r1607, %r1606, -2139062144;
and.b32 %r1608, %r808, 2139062143;
sub.s32 %r1609, %r1607, %r1608;
xor.b32 %r813, %r974, %r1609;
ld.global.nc.u32 %r814, [%rd114+28];

prmt.b32 %r810, %r810, 0, 0xba98;

xor.b32 %r1610, %r810, %r959;
or.b32 %r1611, %r1610, -2139062144;
and.b32 %r1612, %r810, 2139062143;
sub.s32 %r1613, %r1611, %r1612;
xor.b32 %r817, %r982, %r1613;
ld.global.nc.u32 %r818, [%rd114+32];

dp4a.s32.s32 %r812, %r813, %r814, %r804;


dp4a.s32.s32 %r816, %r817, %r818, %r812;

shr.u32 %r1614, %r816, 31;
add.s32 %r1615, %r816, %r1614;
shr.s32 %r1616, %r1615, 1;
mad.lo.s32 %r1617, %r816, %r984, %r1616;
shr.u32 %r1618, %r1617, 31;
add.s32 %r1619, %r1617, %r1618;
shr.s32 %r1620, %r1619, 1;

{ cvt.f32.f16 %f125, %rs1;}


ld.global.nc.u32 %r820, [%rd114];

{.reg .f16 low,high;
mov.b32 {low,high},%r820;
cvt.f32.f16 %f126, low;}


mul.ftz.f32 %f157, %f125, %f126;
cvt.rn.f32.s32 %f158, %r1620;
fma.rn.ftz.f32 %f354, %f157, %f158, %f354;

prmt.b32 %r821, %r821, 0, 0xba98;

xor.b32 %r1621, %r821, %r995;
or.b32 %r1622, %r1621, -2139062144;
and.b32 %r1623, %r821, 2139062143;
sub.s32 %r1624, %r1622, %r1623;
xor.b32 %r826, %r1012, %r1624;

prmt.b32 %r823, %r823, 0, 0xba98;

xor.b32 %r1625, %r823, %r998;
or.b32 %r1626, %r1625, -2139062144;
and.b32 %r1627, %r823, 2139062143;
sub.s32 %r1628, %r1626, %r1627;
xor.b32 %r830, %r1020, %r1628;

dp4a.s32.s32 %r825, %r826, %r778, %r828;


dp4a.s32.s32 %r829, %r830, %r782, %r825;


prmt.b32 %r833, %r833, 0, 0xba98;

xor.b32 %r1629, %r833, %r1022;
or.b32 %r1630, %r1629, -2139062144;
and.b32 %r1631, %r833, 2139062143;
sub.s32 %r1632, %r1630, %r1631;
xor.b32 %r838, %r1040, %r1632;

prmt.b32 %r835, %r835, 0, 0xba98;

xor.b32 %r1633, %r835, %r1025;
or.b32 %r1634, %r1633, -2139062144;
and.b32 %r1635, %r835, 2139062143;
sub.s32 %r1636, %r1634, %r1635;
xor.b32 %r842, %r1048, %r1636;

dp4a.s32.s32 %r837, %r838, %r790, %r829;


dp4a.s32.s32 %r841, %r842, %r794, %r837;


prmt.b32 %r845, %r845, 0, 0xba98;

xor.b32 %r1637, %r845, %r1050;
or.b32 %r1638, %r1637, -2139062144;
and.b32 %r1639, %r845, 2139062143;
sub.s32 %r1640, %r1638, %r1639;
xor.b32 %r850, %r1068, %r1640;

prmt.b32 %r847, %r847, 0, 0xba98;

xor.b32 %r1641, %r847, %r1053;
or.b32 %r1642, %r1641, -2139062144;
and.b32 %r1643, %r847, 2139062143;
sub.s32 %r1644, %r1642, %r1643;
xor.b32 %r854, %r1076, %r1644;

dp4a.s32.s32 %r849, %r850, %r802, %r841;


dp4a.s32.s32 %r853, %r854, %r806, %r849;


prmt.b32 %r857, %r857, 0, 0xba98;

xor.b32 %r1645, %r857, %r1078;
or.b32 %r1646, %r1645, -2139062144;
and.b32 %r1647, %r857, 2139062143;
sub.s32 %r1648, %r1646, %r1647;
xor.b32 %r862, %r1096, %r1648;

prmt.b32 %r859, %r859, 0, 0xba98;

xor.b32 %r1649, %r859, %r1081;
or.b32 %r1650, %r1649, -2139062144;
and.b32 %r1651, %r859, 2139062143;
sub.s32 %r1652, %r1650, %r1651;
xor.b32 %r866, %r1104, %r1652;

dp4a.s32.s32 %r861, %r862, %r814, %r853;


dp4a.s32.s32 %r865, %r866, %r818, %r861;

shr.u32 %r1653, %r865, 31;
add.s32 %r1654, %r865, %r1653;
shr.s32 %r1655, %r1654, 1;
mad.lo.s32 %r1656, %r865, %r1106, %r1655;
shr.u32 %r1657, %r1656, 31;
add.s32 %r1658, %r1656, %r1657;
shr.s32 %r1659, %r1658, 1;

{ cvt.f32.f16 %f127, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r820;
cvt.f32.f16 %f128, low;}


mul.ftz.f32 %f159, %f127, %f128;
cvt.rn.f32.s32 %f160, %r1659;
fma.rn.ftz.f32 %f353, %f159, %f160, %f353;
add.s32 %r2111, %r2111, 64;
add.s32 %r2110, %r2110, 64;
add.s32 %r2109, %r2109, 64;
add.s32 %r2108, %r2108, 64;
add.s32 %r2107, %r2107, 64;
add.s32 %r2106, %r2106, 64;
add.s32 %r2105, %r2105, 64;
add.s32 %r2104, %r2104, 64;
add.s64 %rd184, %rd184, 784;
add.s32 %r2112, %r2112, 8;
setp.lt.s32 %p3, %r2112, %r54;
@%p3 bra $L__BB249_2;

$L__BB249_3:
setp.eq.s32 %p4, %r48, 0;
@%p4 bra $L__BB249_5;

shl.b32 %r1664, %r48, 11;
mov.u32 %r1665, _ZZ13mul_mat_vec_qIL9ggml_type18ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r1666, %r1665, %r1664;
shl.b32 %r1667, %r50, 2;
add.s32 %r1668, %r1666, %r1667;
st.shared.f32 [%r1668+-2048], %f368;
st.shared.f32 [%r1668+-1920], %f367;
st.shared.f32 [%r1668+-1792], %f366;
st.shared.f32 [%r1668+-1664], %f365;
st.shared.f32 [%r1668+-1536], %f364;
st.shared.f32 [%r1668+-1408], %f363;
st.shared.f32 [%r1668+-1280], %f362;
st.shared.f32 [%r1668+-1152], %f361;
st.shared.f32 [%r1668+-1024], %f360;
st.shared.f32 [%r1668+-896], %f359;
st.shared.f32 [%r1668+-768], %f358;
st.shared.f32 [%r1668+-640], %f357;
st.shared.f32 [%r1668+-512], %f356;
st.shared.f32 [%r1668+-384], %f355;
st.shared.f32 [%r1668+-256], %f354;
st.shared.f32 [%r1668+-128], %f353;

$L__BB249_5:
setp.ne.s32 %p5, %r48, 0;
bar.sync 0;
@%p5 bra $L__BB249_38;

mov.u32 %r1670, %ctaid.x;
shl.b32 %r1671, %r1670, 1;
mov.u32 %r1672, 1;
shl.b32 %r1674, %r50, 2;
mov.u32 %r1675, _ZZ13mul_mat_vec_qIL9ggml_type18ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r1676, %r1675, %r1674;
mov.u32 %r1677, 2;
ld.shared.f32 %f161, [%r1676];
add.ftz.f32 %f162, %f161, %f368;
mov.b32 %r1678, %f162;
mov.u32 %r1679, 31;
mov.u32 %r1680, 16;
mov.u32 %r1681, -1;
shfl.sync.bfly.b32 %r1682|%p6, %r1678, %r1680, %r1679, %r1681;
mov.b32 %f163, %r1682;
add.ftz.f32 %f164, %f162, %f163;
mov.b32 %r1683, %f164;
mov.u32 %r1684, 8;
shfl.sync.bfly.b32 %r1685|%p7, %r1683, %r1684, %r1679, %r1681;
mov.b32 %f165, %r1685;
add.ftz.f32 %f166, %f164, %f165;
mov.b32 %r1686, %f166;
mov.u32 %r1687, 4;
shfl.sync.bfly.b32 %r1688|%p8, %r1686, %r1687, %r1679, %r1681;
mov.b32 %f167, %r1688;
add.ftz.f32 %f168, %f166, %f167;
mov.b32 %r1689, %f168;
shfl.sync.bfly.b32 %r1690|%p9, %r1689, %r1677, %r1679, %r1681;
mov.b32 %f169, %r1690;
add.ftz.f32 %f170, %f168, %f169;
mov.b32 %r1691, %f170;
shfl.sync.bfly.b32 %r1692|%p10, %r1691, %r1672, %r1679, %r1681;
mov.b32 %f171, %r1692;
add.ftz.f32 %f49, %f170, %f171;
setp.ne.s32 %p11, %r50, 0;
setp.ge.u32 %p12, %r1671, %r35;
or.pred %p13, %p11, %p12;
@%p13 bra $L__BB249_8;

cvta.to.global.u64 %rd115, %rd18;
mov.u32 %r1693, %ctaid.y;
mad.lo.s32 %r1696, %r1693, %r41, %r1671;
mov.u32 %r1697, %ctaid.z;
mad.lo.s32 %r1698, %r1697, %r47, %r1696;
mul.wide.u32 %rd116, %r1698, 4;
add.s64 %rd117, %rd115, %rd116;
st.global.f32 [%rd117], %f49;

$L__BB249_8:
ld.shared.f32 %f172, [%r1676+128];
add.ftz.f32 %f173, %f172, %f367;
mov.b32 %r1704, %f173;
shfl.sync.bfly.b32 %r1708|%p14, %r1704, %r1680, %r1679, %r1681;
mov.b32 %f174, %r1708;
add.ftz.f32 %f175, %f173, %f174;
mov.b32 %r1709, %f175;
shfl.sync.bfly.b32 %r1711|%p15, %r1709, %r1684, %r1679, %r1681;
mov.b32 %f176, %r1711;
add.ftz.f32 %f177, %f175, %f176;
mov.b32 %r1712, %f177;
shfl.sync.bfly.b32 %r1714|%p16, %r1712, %r1687, %r1679, %r1681;
mov.b32 %f178, %r1714;
add.ftz.f32 %f179, %f177, %f178;
mov.b32 %r1715, %f179;
shfl.sync.bfly.b32 %r1716|%p17, %r1715, %r1677, %r1679, %r1681;
mov.b32 %f180, %r1716;
add.ftz.f32 %f181, %f179, %f180;
mov.b32 %r1717, %f181;
shfl.sync.bfly.b32 %r1719|%p18, %r1717, %r1672, %r1679, %r1681;
mov.b32 %f182, %r1719;
add.ftz.f32 %f50, %f181, %f182;
setp.eq.s32 %p19, %r50, 1;
or.b32 %r1722, %r1671, 1;
setp.lt.u32 %p20, %r1722, %r35;
and.pred %p1, %p19, %p20;
not.pred %p21, %p1;
@%p21 bra $L__BB249_10;

cvta.to.global.u64 %rd118, %rd18;
mov.u32 %r1723, %ctaid.y;
mad.lo.s32 %r1726, %r1723, %r41, %r1671;
mov.u32 %r1727, %ctaid.z;
mad.lo.s32 %r1728, %r1727, %r47, %r1726;
mul.wide.u32 %rd119, %r1728, 4;
add.s64 %rd120, %rd118, %rd119;
st.global.f32 [%rd120+4], %f50;

$L__BB249_10:
mov.u32 %r1733, 2;
ld.shared.f32 %f183, [%r1676+256];
add.ftz.f32 %f184, %f183, %f366;
mov.b32 %r1734, %f184;
mov.u32 %r1735, 31;
mov.u32 %r1736, 16;
mov.u32 %r1737, -1;
shfl.sync.bfly.b32 %r1738|%p22, %r1734, %r1736, %r1735, %r1737;
mov.b32 %f185, %r1738;
add.ftz.f32 %f186, %f184, %f185;
mov.b32 %r1739, %f186;
mov.u32 %r1740, 8;
shfl.sync.bfly.b32 %r1741|%p23, %r1739, %r1740, %r1735, %r1737;
mov.b32 %f187, %r1741;
add.ftz.f32 %f188, %f186, %f187;
mov.b32 %r1742, %f188;
mov.u32 %r1743, 4;
shfl.sync.bfly.b32 %r1744|%p24, %r1742, %r1743, %r1735, %r1737;
mov.b32 %f189, %r1744;
add.ftz.f32 %f190, %f188, %f189;
mov.b32 %r1745, %f190;
shfl.sync.bfly.b32 %r1746|%p25, %r1745, %r1733, %r1735, %r1737;
mov.b32 %f191, %r1746;
add.ftz.f32 %f192, %f190, %f191;
mov.b32 %r1747, %f192;
mov.u32 %r1748, 1;
shfl.sync.bfly.b32 %r1749|%p26, %r1747, %r1748, %r1735, %r1737;
mov.b32 %f193, %r1749;
add.ftz.f32 %f51, %f192, %f193;
@%p13 bra $L__BB249_12;

cvt.u64.u32 %rd121, %r35;
mov.u32 %r1752, %ctaid.y;
mad.lo.s32 %r1755, %r1752, %r41, %r1671;
mov.u32 %r1756, %ctaid.z;
mad.lo.s32 %r1757, %r1756, %r47, %r1755;
cvt.u64.u32 %rd122, %r1757;
add.s64 %rd123, %rd121, %rd122;
cvta.to.global.u64 %rd124, %rd18;
shl.b64 %rd125, %rd123, 2;
add.s64 %rd126, %rd124, %rd125;
st.global.f32 [%rd126], %f51;

$L__BB249_12:
ld.shared.f32 %f194, [%r1676+384];
add.ftz.f32 %f195, %f194, %f365;
mov.b32 %r1763, %f195;
shfl.sync.bfly.b32 %r1767|%p30, %r1763, %r1736, %r1735, %r1737;
mov.b32 %f196, %r1767;
add.ftz.f32 %f197, %f195, %f196;
mov.b32 %r1768, %f197;
shfl.sync.bfly.b32 %r1770|%p31, %r1768, %r1740, %r1735, %r1737;
mov.b32 %f198, %r1770;
add.ftz.f32 %f199, %f197, %f198;
mov.b32 %r1771, %f199;
shfl.sync.bfly.b32 %r1773|%p32, %r1771, %r1743, %r1735, %r1737;
mov.b32 %f200, %r1773;
add.ftz.f32 %f201, %f199, %f200;
mov.b32 %r1774, %f201;
shfl.sync.bfly.b32 %r1775|%p33, %r1774, %r1733, %r1735, %r1737;
mov.b32 %f202, %r1775;
add.ftz.f32 %f203, %f201, %f202;
mov.b32 %r1776, %f203;
shfl.sync.bfly.b32 %r1778|%p34, %r1776, %r1748, %r1735, %r1737;
mov.b32 %f204, %r1778;
add.ftz.f32 %f52, %f203, %f204;
@%p21 bra $L__BB249_14;

add.s32 %r1779, %r35, 1;
cvt.u64.u32 %rd127, %r1779;
mov.u32 %r1780, %ctaid.y;
mad.lo.s32 %r1783, %r1780, %r41, %r1671;
mov.u32 %r1784, %ctaid.z;
mad.lo.s32 %r1785, %r1784, %r47, %r1783;
cvt.u64.u32 %rd128, %r1785;
add.s64 %rd129, %rd127, %rd128;
cvta.to.global.u64 %rd130, %rd18;
shl.b64 %rd131, %rd129, 2;
add.s64 %rd132, %rd130, %rd131;
st.global.f32 [%rd132], %f52;

$L__BB249_14:
shl.b32 %r1786, %r35, 1;
mov.u32 %r1787, 1;
mov.u32 %r1792, 2;
ld.shared.f32 %f205, [%r1676+512];
add.ftz.f32 %f206, %f205, %f364;
mov.b32 %r1793, %f206;
mov.u32 %r1794, 31;
mov.u32 %r1795, 16;
mov.u32 %r1796, -1;
shfl.sync.bfly.b32 %r1797|%p36, %r1793, %r1795, %r1794, %r1796;
mov.b32 %f207, %r1797;
add.ftz.f32 %f208, %f206, %f207;
mov.b32 %r1798, %f208;
mov.u32 %r1799, 8;
shfl.sync.bfly.b32 %r1800|%p37, %r1798, %r1799, %r1794, %r1796;
mov.b32 %f209, %r1800;
add.ftz.f32 %f210, %f208, %f209;
mov.b32 %r1801, %f210;
mov.u32 %r1802, 4;
shfl.sync.bfly.b32 %r1803|%p38, %r1801, %r1802, %r1794, %r1796;
mov.b32 %f211, %r1803;
add.ftz.f32 %f212, %f210, %f211;
mov.b32 %r1804, %f212;
shfl.sync.bfly.b32 %r1805|%p39, %r1804, %r1792, %r1794, %r1796;
mov.b32 %f213, %r1805;
add.ftz.f32 %f214, %f212, %f213;
mov.b32 %r1806, %f214;
shfl.sync.bfly.b32 %r1807|%p40, %r1806, %r1787, %r1794, %r1796;
mov.b32 %f215, %r1807;
add.ftz.f32 %f53, %f214, %f215;
cvt.u64.u32 %rd133, %r1786;
mov.u32 %r1808, %ctaid.y;
mad.lo.s32 %r1811, %r1808, %r41, %r1671;
mov.u32 %r1812, %ctaid.z;
mad.lo.s32 %r1813, %r1812, %r47, %r1811;
cvt.u64.u32 %rd134, %r1813;
add.s64 %rd135, %rd133, %rd134;
cvta.to.global.u64 %rd136, %rd18;
shl.b64 %rd137, %rd135, 2;
add.s64 %rd13, %rd136, %rd137;
@%p13 bra $L__BB249_16;

st.global.f32 [%rd13], %f53;

$L__BB249_16:
ld.shared.f32 %f216, [%r1676+640];
add.ftz.f32 %f217, %f216, %f363;
mov.b32 %r1819, %f217;
shfl.sync.bfly.b32 %r1823|%p44, %r1819, %r1795, %r1794, %r1796;
mov.b32 %f218, %r1823;
add.ftz.f32 %f219, %f217, %f218;
mov.b32 %r1824, %f219;
shfl.sync.bfly.b32 %r1826|%p45, %r1824, %r1799, %r1794, %r1796;
mov.b32 %f220, %r1826;
add.ftz.f32 %f221, %f219, %f220;
mov.b32 %r1827, %f221;
shfl.sync.bfly.b32 %r1829|%p46, %r1827, %r1802, %r1794, %r1796;
mov.b32 %f222, %r1829;
add.ftz.f32 %f223, %f221, %f222;
mov.b32 %r1830, %f223;
shfl.sync.bfly.b32 %r1831|%p47, %r1830, %r1792, %r1794, %r1796;
mov.b32 %f224, %r1831;
add.ftz.f32 %f225, %f223, %f224;
mov.b32 %r1832, %f225;
shfl.sync.bfly.b32 %r1834|%p48, %r1832, %r1787, %r1794, %r1796;
mov.b32 %f226, %r1834;
add.ftz.f32 %f54, %f225, %f226;
@%p21 bra $L__BB249_18;

st.global.f32 [%rd13+4], %f54;

$L__BB249_18:
mul.lo.s32 %r29, %r35, 3;
mov.u32 %r1839, 2;
ld.shared.f32 %f227, [%r1676+768];
add.ftz.f32 %f228, %f227, %f362;
mov.b32 %r1840, %f228;
mov.u32 %r1841, 31;
mov.u32 %r1842, 16;
mov.u32 %r1843, -1;
shfl.sync.bfly.b32 %r1844|%p50, %r1840, %r1842, %r1841, %r1843;
mov.b32 %f229, %r1844;
add.ftz.f32 %f230, %f228, %f229;
mov.b32 %r1845, %f230;
mov.u32 %r1846, 8;
shfl.sync.bfly.b32 %r1847|%p51, %r1845, %r1846, %r1841, %r1843;
mov.b32 %f231, %r1847;
add.ftz.f32 %f232, %f230, %f231;
mov.b32 %r1848, %f232;
mov.u32 %r1849, 4;
shfl.sync.bfly.b32 %r1850|%p52, %r1848, %r1849, %r1841, %r1843;
mov.b32 %f233, %r1850;
add.ftz.f32 %f234, %f232, %f233;
mov.b32 %r1851, %f234;
shfl.sync.bfly.b32 %r1852|%p53, %r1851, %r1839, %r1841, %r1843;
mov.b32 %f235, %r1852;
add.ftz.f32 %f236, %f234, %f235;
mov.b32 %r1853, %f236;
mov.u32 %r1854, 1;
shfl.sync.bfly.b32 %r1855|%p54, %r1853, %r1854, %r1841, %r1843;
mov.b32 %f237, %r1855;
add.ftz.f32 %f55, %f236, %f237;
@%p13 bra $L__BB249_20;

cvt.u64.u32 %rd138, %r29;
add.s64 %rd140, %rd138, %rd134;
shl.b64 %rd142, %rd140, 2;
add.s64 %rd143, %rd136, %rd142;
st.global.f32 [%rd143], %f55;

$L__BB249_20:
ld.shared.f32 %f238, [%r1676+896];
add.ftz.f32 %f239, %f238, %f361;
mov.b32 %r1869, %f239;
shfl.sync.bfly.b32 %r1873|%p58, %r1869, %r1842, %r1841, %r1843;
mov.b32 %f240, %r1873;
add.ftz.f32 %f241, %f239, %f240;
mov.b32 %r1874, %f241;
shfl.sync.bfly.b32 %r1876|%p59, %r1874, %r1846, %r1841, %r1843;
mov.b32 %f242, %r1876;
add.ftz.f32 %f243, %f241, %f242;
mov.b32 %r1877, %f243;
shfl.sync.bfly.b32 %r1879|%p60, %r1877, %r1849, %r1841, %r1843;
mov.b32 %f244, %r1879;
add.ftz.f32 %f245, %f243, %f244;
mov.b32 %r1880, %f245;
shfl.sync.bfly.b32 %r1881|%p61, %r1880, %r1839, %r1841, %r1843;
mov.b32 %f246, %r1881;
add.ftz.f32 %f247, %f245, %f246;
mov.b32 %r1882, %f247;
shfl.sync.bfly.b32 %r1884|%p62, %r1882, %r1854, %r1841, %r1843;
mov.b32 %f248, %r1884;
add.ftz.f32 %f56, %f247, %f248;
@%p21 bra $L__BB249_22;

add.s32 %r1885, %r29, 1;
cvt.u64.u32 %rd144, %r1885;
add.s64 %rd146, %rd144, %rd134;
shl.b64 %rd148, %rd146, 2;
add.s64 %rd149, %rd136, %rd148;
st.global.f32 [%rd149], %f56;

$L__BB249_22:
shl.b32 %r1892, %r35, 2;
mov.u32 %r1893, 2;
ld.shared.f32 %f249, [%r1676+1024];
add.ftz.f32 %f250, %f249, %f360;
mov.b32 %r1898, %f250;
mov.u32 %r1899, 31;
mov.u32 %r1900, 16;
mov.u32 %r1901, -1;
shfl.sync.bfly.b32 %r1902|%p64, %r1898, %r1900, %r1899, %r1901;
mov.b32 %f251, %r1902;
add.ftz.f32 %f252, %f250, %f251;
mov.b32 %r1903, %f252;
mov.u32 %r1904, 8;
shfl.sync.bfly.b32 %r1905|%p65, %r1903, %r1904, %r1899, %r1901;
mov.b32 %f253, %r1905;
add.ftz.f32 %f254, %f252, %f253;
mov.b32 %r1906, %f254;
mov.u32 %r1907, 4;
shfl.sync.bfly.b32 %r1908|%p66, %r1906, %r1907, %r1899, %r1901;
mov.b32 %f255, %r1908;
add.ftz.f32 %f256, %f254, %f255;
mov.b32 %r1909, %f256;
shfl.sync.bfly.b32 %r1910|%p67, %r1909, %r1893, %r1899, %r1901;
mov.b32 %f257, %r1910;
add.ftz.f32 %f258, %f256, %f257;
mov.b32 %r1911, %f258;
mov.u32 %r1912, 1;
shfl.sync.bfly.b32 %r1913|%p68, %r1911, %r1912, %r1899, %r1901;
mov.b32 %f259, %r1913;
add.ftz.f32 %f57, %f258, %f259;
cvt.u64.u32 %rd150, %r1892;
add.s64 %rd152, %rd150, %rd134;
shl.b64 %rd154, %rd152, 2;
add.s64 %rd14, %rd136, %rd154;
@%p13 bra $L__BB249_24;

st.global.f32 [%rd14], %f57;

$L__BB249_24:
ld.shared.f32 %f260, [%r1676+1152];
add.ftz.f32 %f261, %f260, %f359;
mov.b32 %r1925, %f261;
shfl.sync.bfly.b32 %r1929|%p72, %r1925, %r1900, %r1899, %r1901;
mov.b32 %f262, %r1929;
add.ftz.f32 %f263, %f261, %f262;
mov.b32 %r1930, %f263;
shfl.sync.bfly.b32 %r1932|%p73, %r1930, %r1904, %r1899, %r1901;
mov.b32 %f264, %r1932;
add.ftz.f32 %f265, %f263, %f264;
mov.b32 %r1933, %f265;
shfl.sync.bfly.b32 %r1935|%p74, %r1933, %r1907, %r1899, %r1901;
mov.b32 %f266, %r1935;
add.ftz.f32 %f267, %f265, %f266;
mov.b32 %r1936, %f267;
shfl.sync.bfly.b32 %r1937|%p75, %r1936, %r1893, %r1899, %r1901;
mov.b32 %f268, %r1937;
add.ftz.f32 %f269, %f267, %f268;
mov.b32 %r1938, %f269;
shfl.sync.bfly.b32 %r1940|%p76, %r1938, %r1912, %r1899, %r1901;
mov.b32 %f270, %r1940;
add.ftz.f32 %f58, %f269, %f270;
@%p21 bra $L__BB249_26;

st.global.f32 [%rd14+4], %f58;

$L__BB249_26:
mul.lo.s32 %r30, %r35, 5;
mov.u32 %r1945, 2;
ld.shared.f32 %f271, [%r1676+1280];
add.ftz.f32 %f272, %f271, %f358;
mov.b32 %r1946, %f272;
mov.u32 %r1947, 31;
mov.u32 %r1948, 16;
mov.u32 %r1949, -1;
shfl.sync.bfly.b32 %r1950|%p78, %r1946, %r1948, %r1947, %r1949;
mov.b32 %f273, %r1950;
add.ftz.f32 %f274, %f272, %f273;
mov.b32 %r1951, %f274;
mov.u32 %r1952, 8;
shfl.sync.bfly.b32 %r1953|%p79, %r1951, %r1952, %r1947, %r1949;
mov.b32 %f275, %r1953;
add.ftz.f32 %f276, %f274, %f275;
mov.b32 %r1954, %f276;
mov.u32 %r1955, 4;
shfl.sync.bfly.b32 %r1956|%p80, %r1954, %r1955, %r1947, %r1949;
mov.b32 %f277, %r1956;
add.ftz.f32 %f278, %f276, %f277;
mov.b32 %r1957, %f278;
shfl.sync.bfly.b32 %r1958|%p81, %r1957, %r1945, %r1947, %r1949;
mov.b32 %f279, %r1958;
add.ftz.f32 %f280, %f278, %f279;
mov.b32 %r1959, %f280;
mov.u32 %r1960, 1;
shfl.sync.bfly.b32 %r1961|%p82, %r1959, %r1960, %r1947, %r1949;
mov.b32 %f281, %r1961;
add.ftz.f32 %f59, %f280, %f281;
@%p13 bra $L__BB249_28;

cvt.u64.u32 %rd155, %r30;
add.s64 %rd157, %rd155, %rd134;
shl.b64 %rd159, %rd157, 2;
add.s64 %rd160, %rd136, %rd159;
st.global.f32 [%rd160], %f59;

$L__BB249_28:
ld.shared.f32 %f282, [%r1676+1408];
add.ftz.f32 %f283, %f282, %f357;
mov.b32 %r1975, %f283;
shfl.sync.bfly.b32 %r1979|%p86, %r1975, %r1948, %r1947, %r1949;
mov.b32 %f284, %r1979;
add.ftz.f32 %f285, %f283, %f284;
mov.b32 %r1980, %f285;
shfl.sync.bfly.b32 %r1982|%p87, %r1980, %r1952, %r1947, %r1949;
mov.b32 %f286, %r1982;
add.ftz.f32 %f287, %f285, %f286;
mov.b32 %r1983, %f287;
shfl.sync.bfly.b32 %r1985|%p88, %r1983, %r1955, %r1947, %r1949;
mov.b32 %f288, %r1985;
add.ftz.f32 %f289, %f287, %f288;
mov.b32 %r1986, %f289;
shfl.sync.bfly.b32 %r1987|%p89, %r1986, %r1945, %r1947, %r1949;
mov.b32 %f290, %r1987;
add.ftz.f32 %f291, %f289, %f290;
mov.b32 %r1988, %f291;
shfl.sync.bfly.b32 %r1990|%p90, %r1988, %r1960, %r1947, %r1949;
mov.b32 %f292, %r1990;
add.ftz.f32 %f60, %f291, %f292;
@%p21 bra $L__BB249_30;

add.s32 %r1991, %r30, 1;
cvt.u64.u32 %rd161, %r1991;
add.s64 %rd163, %rd161, %rd134;
shl.b64 %rd165, %rd163, 2;
add.s64 %rd166, %rd136, %rd165;
st.global.f32 [%rd166], %f60;

$L__BB249_30:
mul.lo.s32 %r1998, %r35, 6;
mov.u32 %r2003, 2;
ld.shared.f32 %f293, [%r1676+1536];
add.ftz.f32 %f294, %f293, %f356;
mov.b32 %r2004, %f294;
mov.u32 %r2005, 31;
mov.u32 %r2006, 16;
mov.u32 %r2007, -1;
shfl.sync.bfly.b32 %r2008|%p92, %r2004, %r2006, %r2005, %r2007;
mov.b32 %f295, %r2008;
add.ftz.f32 %f296, %f294, %f295;
mov.b32 %r2009, %f296;
mov.u32 %r2010, 8;
shfl.sync.bfly.b32 %r2011|%p93, %r2009, %r2010, %r2005, %r2007;
mov.b32 %f297, %r2011;
add.ftz.f32 %f298, %f296, %f297;
mov.b32 %r2012, %f298;
mov.u32 %r2013, 4;
shfl.sync.bfly.b32 %r2014|%p94, %r2012, %r2013, %r2005, %r2007;
mov.b32 %f299, %r2014;
add.ftz.f32 %f300, %f298, %f299;
mov.b32 %r2015, %f300;
shfl.sync.bfly.b32 %r2016|%p95, %r2015, %r2003, %r2005, %r2007;
mov.b32 %f301, %r2016;
add.ftz.f32 %f302, %f300, %f301;
mov.b32 %r2017, %f302;
mov.u32 %r2018, 1;
shfl.sync.bfly.b32 %r2019|%p96, %r2017, %r2018, %r2005, %r2007;
mov.b32 %f303, %r2019;
add.ftz.f32 %f61, %f302, %f303;
cvt.u64.u32 %rd167, %r1998;
add.s64 %rd169, %rd167, %rd134;
shl.b64 %rd171, %rd169, 2;
add.s64 %rd15, %rd136, %rd171;
@%p13 bra $L__BB249_32;

st.global.f32 [%rd15], %f61;

$L__BB249_32:
ld.shared.f32 %f304, [%r1676+1664];
add.ftz.f32 %f305, %f304, %f355;
mov.b32 %r2031, %f305;
shfl.sync.bfly.b32 %r2035|%p100, %r2031, %r2006, %r2005, %r2007;
mov.b32 %f306, %r2035;
add.ftz.f32 %f307, %f305, %f306;
mov.b32 %r2036, %f307;
shfl.sync.bfly.b32 %r2038|%p101, %r2036, %r2010, %r2005, %r2007;
mov.b32 %f308, %r2038;
add.ftz.f32 %f309, %f307, %f308;
mov.b32 %r2039, %f309;
shfl.sync.bfly.b32 %r2041|%p102, %r2039, %r2013, %r2005, %r2007;
mov.b32 %f310, %r2041;
add.ftz.f32 %f311, %f309, %f310;
mov.b32 %r2042, %f311;
shfl.sync.bfly.b32 %r2043|%p103, %r2042, %r2003, %r2005, %r2007;
mov.b32 %f312, %r2043;
add.ftz.f32 %f313, %f311, %f312;
mov.b32 %r2044, %f313;
shfl.sync.bfly.b32 %r2046|%p104, %r2044, %r2018, %r2005, %r2007;
mov.b32 %f314, %r2046;
add.ftz.f32 %f62, %f313, %f314;
@%p21 bra $L__BB249_34;

st.global.f32 [%rd15+4], %f62;

$L__BB249_34:
mul.lo.s32 %r31, %r35, 7;
mov.u32 %r2051, 2;
ld.shared.f32 %f315, [%r1676+1792];
add.ftz.f32 %f316, %f315, %f354;
mov.b32 %r2052, %f316;
mov.u32 %r2053, 31;
mov.u32 %r2054, 16;
mov.u32 %r2055, -1;
shfl.sync.bfly.b32 %r2056|%p106, %r2052, %r2054, %r2053, %r2055;
mov.b32 %f317, %r2056;
add.ftz.f32 %f318, %f316, %f317;
mov.b32 %r2057, %f318;
mov.u32 %r2058, 8;
shfl.sync.bfly.b32 %r2059|%p107, %r2057, %r2058, %r2053, %r2055;
mov.b32 %f319, %r2059;
add.ftz.f32 %f320, %f318, %f319;
mov.b32 %r2060, %f320;
mov.u32 %r2061, 4;
shfl.sync.bfly.b32 %r2062|%p108, %r2060, %r2061, %r2053, %r2055;
mov.b32 %f321, %r2062;
add.ftz.f32 %f322, %f320, %f321;
mov.b32 %r2063, %f322;
shfl.sync.bfly.b32 %r2064|%p109, %r2063, %r2051, %r2053, %r2055;
mov.b32 %f323, %r2064;
add.ftz.f32 %f324, %f322, %f323;
mov.b32 %r2065, %f324;
mov.u32 %r2066, 1;
shfl.sync.bfly.b32 %r2067|%p110, %r2065, %r2066, %r2053, %r2055;
mov.b32 %f325, %r2067;
add.ftz.f32 %f63, %f324, %f325;
@%p13 bra $L__BB249_36;

cvt.u64.u32 %rd172, %r31;
add.s64 %rd174, %rd172, %rd134;
shl.b64 %rd176, %rd174, 2;
add.s64 %rd177, %rd136, %rd176;
st.global.f32 [%rd177], %f63;

$L__BB249_36:
ld.shared.f32 %f326, [%r1676+1920];
add.ftz.f32 %f327, %f326, %f353;
mov.b32 %r2081, %f327;
shfl.sync.bfly.b32 %r2085|%p114, %r2081, %r2054, %r2053, %r2055;
mov.b32 %f328, %r2085;
add.ftz.f32 %f329, %f327, %f328;
mov.b32 %r2086, %f329;
shfl.sync.bfly.b32 %r2088|%p115, %r2086, %r2058, %r2053, %r2055;
mov.b32 %f330, %r2088;
add.ftz.f32 %f331, %f329, %f330;
mov.b32 %r2089, %f331;
shfl.sync.bfly.b32 %r2091|%p116, %r2089, %r2061, %r2053, %r2055;
mov.b32 %f332, %r2091;
add.ftz.f32 %f333, %f331, %f332;
mov.b32 %r2092, %f333;
shfl.sync.bfly.b32 %r2093|%p117, %r2092, %r2051, %r2053, %r2055;
mov.b32 %f334, %r2093;
add.ftz.f32 %f335, %f333, %f334;
mov.b32 %r2094, %f335;
shfl.sync.bfly.b32 %r2096|%p118, %r2094, %r2066, %r2053, %r2055;
mov.b32 %f336, %r2096;
add.ftz.f32 %f64, %f335, %f336;
@%p21 bra $L__BB249_38;

add.s32 %r2097, %r31, 1;
cvt.u64.u32 %rd178, %r2097;
add.s64 %rd180, %rd178, %rd134;
shl.b64 %rd182, %rd180, 2;
add.s64 %rd183, %rd136, %rd182;
st.global.f32 [%rd183], %f64;

$L__BB249_38:
ret;

}
.entry _Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj(
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2,
.param .align 8 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3[48],
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5,
.param .align 4 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6[12],
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15
)
.maxntid 256, 1, 1
.minnctapersm 1
{
.local .align 8 .b8 __local_depot250[16];
.reg .b64 %SP;
.reg .b64 %SPL;
.reg .pred %p<48>;
.reg .b16 %rs<33>;
.reg .f32 %f<193>;
.reg .b32 %r<431>;
.reg .b64 %rd<146>;


mov.u64 %SPL, __local_depot250;
ld.param.u64 %rd18, [_Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0];
ld.param.u64 %rd19, [_Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1];
ld.param.u64 %rd20, [_Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2];
ld.param.u32 %r21, [_Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5];
ld.param.u32 %r26, [_Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8];
ld.param.u32 %r27, [_Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9];
ld.param.u32 %r29, [_Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11];
ld.param.u32 %r30, [_Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12];
ld.param.u32 %r33, [_Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14];
ld.param.u32 %r32, [_Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15];
ld.param.u32 %r24, [_Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+8];
ld.param.u32 %r23, [_Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+4];
ld.param.u32 %r22, [_Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6];
ld.param.u64 %rd22, [_Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+8];
add.u64 %rd1, %SPL, 0;
add.u64 %rd2, %SPL, 8;
mov.u32 %r1, %tid.y;
setp.ge.u32 %p1, %r1, %r33;
@%p1 bra $L__BB250_29;

mov.u32 %r2, %ctaid.y;
cvta.to.global.u64 %rd29, %rd20;
mad.lo.s32 %r34, %r1, %r32, %r2;
mul.wide.u32 %rd30, %r34, 4;
add.s64 %rd31, %rd29, %rd30;
ld.global.nc.u32 %r3, [%rd31];
mov.f32 %f186, 0f00000000;
st.local.v2.f32 [%rd1], {%f186, %f186};
st.local.v2.f32 [%rd2], {%f186, %f186};
mov.u32 %r4, %tid.x;
shr.u32 %r5, %r4, 3;
shr.u32 %r35, %r21, 8;
setp.ge.u32 %p2, %r5, %r35;
mov.f32 %f187, %f186;
mov.f32 %f188, %f186;
mov.f32 %f189, %f186;
@%p2 bra $L__BB250_9;

and.b32 %r36, %r4, 7;
shl.b32 %r37, %r4, 1;
and.b32 %r38, %r37, 14;
cvt.u64.u32 %rd32, %r38;
shl.b32 %r39, %r4, 2;
cvt.u64.u32 %rd33, %r39;
and.b64 %rd34, %rd33, 28;
mul.hi.u32 %r40, %r2, %r22;
add.s32 %r41, %r40, %r2;
shr.u32 %r42, %r41, %r23;
mul.lo.s32 %r43, %r42, %r24;
sub.s32 %r44, %r2, %r43;
mul.lo.s32 %r45, %r44, %r30;
mul.wide.u32 %rd35, %r45, 36;
mul.lo.s32 %r46, %r1, %r27;
mul.wide.u32 %rd36, %r46, 36;
add.s64 %rd37, %rd36, %rd35;
mul.wide.s32 %rd38, %r36, 36;
add.s64 %rd39, %rd37, %rd38;
shl.b32 %r47, %r5, 3;
mul.wide.s32 %rd40, %r47, 36;
add.s64 %rd41, %rd39, %rd40;
cvta.to.global.u64 %rd42, %rd19;
add.s64 %rd43, %rd42, %rd41;
add.s64 %rd143, %rd43, 16;
mov.u32 %r48, %ctaid.x;
shl.b32 %r49, %r48, 1;
or.b32 %r50, %r49, 1;
mad.lo.s32 %r51, %r3, %r29, %r5;
mad.lo.s32 %r52, %r26, %r50, %r51;
mul.wide.s32 %rd44, %r52, 50;
add.s64 %rd5, %rd34, %rd44;
add.s64 %rd45, %rd32, %rd44;
add.s64 %rd6, %rd45, 34;
cvta.to.global.u64 %rd145, %rd18;
cvta.to.global.u64 %rd144, %rd22;
mov.f32 %f186, 0f00000000;
mov.u32 %r430, %r5;

$L__BB250_3:
mad.lo.s32 %r92, %r49, %r26, %r5;
mad.lo.s32 %r93, %r3, %r29, %r92;
mul.wide.s32 %rd48, %r93, 50;
add.s64 %rd49, %rd34, %rd48;
add.s64 %rd50, %rd145, %rd49;
ld.global.nc.u16 %rs2, [%rd50+2];
shr.u16 %rs3, %rs2, 8;
cvt.u32.u16 %r94, %rs3;
ld.global.nc.u16 %rs4, [%rd50+4];
shr.u16 %rs5, %rs4, 8;
cvt.u32.u16 %r95, %rs5;
add.s64 %rd52, %rd32, %rd48;
add.s64 %rd53, %rd145, %rd52;
ld.global.nc.u16 %rs6, [%rd53+34];
cvt.u32.u16 %r98, %rs6;
and.b16 %rs7, %rs2, 255;
cvt.u32.u16 %r99, %rs7;
mul.wide.u16 %r100, %rs6, 256;
and.b32 %r101, %r100, 1792;
or.b32 %r102, %r101, %r99;
mul.wide.u32 %rd54, %r102, 4;
mov.u64 %rd55, iq1s_grid_gpu;
add.s64 %rd56, %rd55, %rd54;
ld.global.nc.u32 %r103, [%rd56];
and.b32 %r54, %r103, 252645135;
shr.u32 %r104, %r103, 4;
and.b32 %r58, %r104, 252645135;
ld.global.nc.u32 %r7, [%rd143+-12];
ld.global.nc.u32 %r8, [%rd143+-8];
mov.u32 %r56, 0;

dp4a.s32.s32 %r53, %r54, %r7, %r56;


dp4a.s32.s32 %r57, %r58, %r8, %r53;

mul.wide.u16 %r105, %rs6, 32;
and.b32 %r106, %r105, 1792;
or.b32 %r107, %r106, %r94;
mul.wide.u32 %rd57, %r107, 4;
add.s64 %rd58, %rd55, %rd57;
ld.global.nc.u32 %r108, [%rd58];
and.b32 %r62, %r108, 252645135;
shr.u32 %r109, %r108, 4;
and.b32 %r66, %r109, 252645135;
ld.global.nc.u32 %r9, [%rd143+-4];
ld.global.nc.u32 %r10, [%rd143];

dp4a.s32.s32 %r61, %r62, %r9, %r57;


dp4a.s32.s32 %r65, %r66, %r10, %r61;

and.b16 %rs8, %rs4, 255;
cvt.u32.u16 %r110, %rs8;
mul.wide.u16 %r111, %rs6, 4;
and.b32 %r112, %r111, 1792;
or.b32 %r113, %r112, %r110;
mul.wide.u32 %rd59, %r113, 4;
add.s64 %rd60, %rd55, %rd59;
ld.global.nc.u32 %r114, [%rd60];
and.b32 %r70, %r114, 252645135;
shr.u32 %r115, %r114, 4;
and.b32 %r74, %r115, 252645135;
ld.global.nc.u32 %r11, [%rd143+4];
ld.global.nc.u32 %r12, [%rd143+8];

dp4a.s32.s32 %r69, %r70, %r11, %r65;


dp4a.s32.s32 %r73, %r74, %r12, %r69;

shr.u32 %r116, %r98, 1;
and.b32 %r117, %r116, 1792;
or.b32 %r118, %r117, %r95;
mul.wide.u32 %rd61, %r118, 4;
add.s64 %rd62, %rd55, %rd61;
ld.global.nc.u32 %r119, [%rd62];
and.b32 %r78, %r119, 252645135;
shr.u32 %r120, %r119, 4;
and.b32 %r82, %r120, 252645135;
ld.global.nc.u32 %r13, [%rd143+12];
ld.global.nc.u32 %r14, [%rd143+16];

dp4a.s32.s32 %r77, %r78, %r13, %r73;


dp4a.s32.s32 %r81, %r82, %r14, %r77;

add.s64 %rd63, %rd145, %rd48;
ld.global.nc.u16 %rs1, [%rd63];

{ cvt.f32.f16 %f42, %rs1;}


shr.u32 %r121, %r98, 11;
and.b32 %r122, %r121, 14;
or.b32 %r123, %r122, 1;
cvt.rn.f32.s32 %f45, %r123;
mul.ftz.f32 %f46, %f42, %f45;
and.b32 %r124, %r98, 32768;
cvt.rn.f32.s32 %f47, %r124;
fma.rn.ftz.f32 %f48, %f47, 0fB7000000, 0fBF600000;
ld.global.nc.u32 %r15, [%rd143+-16];

{.reg .f16 low,high;
mov.b32 {low,high},%r15;
cvt.f32.f16 %f43, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r15;
cvt.f32.f16 %f44, high;}


cvt.rn.f32.s32 %f49, %r81;
mul.ftz.f32 %f50, %f48, %f44;
fma.rn.ftz.f32 %f51, %f43, %f49, %f50;
fma.rn.ftz.f32 %f189, %f46, %f51, %f189;
setp.eq.s64 %p3, %rd22, 0;
@%p3 bra $L__BB250_5;

mov.u64 %rd128, iq1s_grid_gpu;
shr.u32 %r161, %r4, 3;
mad.lo.s32 %r164, %r49, %r26, %r161;
mad.lo.s32 %r165, %r3, %r29, %r164;
mul.wide.s32 %rd66, %r165, 50;
add.s64 %rd67, %rd34, %rd66;
add.s64 %rd68, %rd144, %rd67;
ld.global.u16 %rs10, [%rd68+2];
shr.u16 %rs11, %rs10, 8;
cvt.u32.u16 %r166, %rs11;
ld.global.u16 %rs12, [%rd68+4];
shr.u16 %rs13, %rs12, 8;
cvt.u32.u16 %r167, %rs13;
add.s64 %rd70, %rd32, %rd66;
add.s64 %rd71, %rd144, %rd70;
ld.global.u16 %rs14, [%rd71+34];
cvt.u32.u16 %r170, %rs14;
and.b16 %rs15, %rs10, 255;
cvt.u32.u16 %r171, %rs15;
mul.wide.u16 %r172, %rs14, 256;
and.b32 %r173, %r172, 1792;
or.b32 %r174, %r173, %r171;
mul.wide.u32 %rd72, %r174, 4;
add.s64 %rd74, %rd128, %rd72;
ld.global.nc.u32 %r175, [%rd74];
and.b32 %r126, %r175, 252645135;
shr.u32 %r176, %r175, 4;
and.b32 %r130, %r176, 252645135;

dp4a.s32.s32 %r125, %r126, %r7, %r56;


dp4a.s32.s32 %r129, %r130, %r8, %r125;

mul.wide.u16 %r177, %rs14, 32;
and.b32 %r178, %r177, 1792;
or.b32 %r179, %r178, %r166;
mul.wide.u32 %rd75, %r179, 4;
add.s64 %rd76, %rd128, %rd75;
ld.global.nc.u32 %r180, [%rd76];
and.b32 %r134, %r180, 252645135;
shr.u32 %r181, %r180, 4;
and.b32 %r138, %r181, 252645135;

dp4a.s32.s32 %r133, %r134, %r9, %r129;


dp4a.s32.s32 %r137, %r138, %r10, %r133;

and.b16 %rs16, %rs12, 255;
cvt.u32.u16 %r182, %rs16;
mul.wide.u16 %r183, %rs14, 4;
and.b32 %r184, %r183, 1792;
or.b32 %r185, %r184, %r182;
mul.wide.u32 %rd77, %r185, 4;
add.s64 %rd78, %rd128, %rd77;
ld.global.nc.u32 %r186, [%rd78];
and.b32 %r142, %r186, 252645135;
shr.u32 %r187, %r186, 4;
and.b32 %r146, %r187, 252645135;

dp4a.s32.s32 %r141, %r142, %r11, %r137;


dp4a.s32.s32 %r145, %r146, %r12, %r141;

shr.u32 %r188, %r170, 1;
and.b32 %r189, %r188, 1792;
or.b32 %r190, %r189, %r167;
mul.wide.u32 %rd79, %r190, 4;
add.s64 %rd80, %rd128, %rd79;
ld.global.nc.u32 %r191, [%rd80];
and.b32 %r150, %r191, 252645135;
shr.u32 %r192, %r191, 4;
and.b32 %r154, %r192, 252645135;

dp4a.s32.s32 %r149, %r150, %r13, %r145;


dp4a.s32.s32 %r153, %r154, %r14, %r149;

add.s64 %rd81, %rd144, %rd66;
ld.global.u16 %rs9, [%rd81];

{ cvt.f32.f16 %f52, %rs9;}


shr.u32 %r193, %r170, 11;
and.b32 %r194, %r193, 14;
or.b32 %r195, %r194, 1;
cvt.rn.f32.s32 %f55, %r195;
mul.ftz.f32 %f56, %f52, %f55;
and.b32 %r196, %r170, 32768;
cvt.rn.f32.s32 %f57, %r196;
fma.rn.ftz.f32 %f58, %f57, 0fB7000000, 0fBF600000;

{.reg .f16 low,high;
mov.b32 {low,high},%r15;
cvt.f32.f16 %f53, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r15;
cvt.f32.f16 %f54, high;}


cvt.rn.f32.s32 %f59, %r153;
mul.ftz.f32 %f60, %f58, %f54;
fma.rn.ftz.f32 %f61, %f53, %f59, %f60;
fma.rn.ftz.f32 %f187, %f56, %f61, %f187;

$L__BB250_5:
mov.u64 %rd121, iq1s_grid_gpu;
add.s64 %rd82, %rd145, %rd5;
ld.global.nc.u16 %rs18, [%rd82+2];
shr.u16 %rs19, %rs18, 8;
cvt.u32.u16 %r231, %rs19;
ld.global.nc.u16 %rs20, [%rd82+4];
shr.u16 %rs21, %rs20, 8;
cvt.u32.u16 %r232, %rs21;
add.s64 %rd83, %rd145, %rd6;
ld.global.nc.u16 %rs22, [%rd83];
cvt.u32.u16 %r233, %rs22;
and.b16 %rs23, %rs18, 255;
cvt.u32.u16 %r234, %rs23;
mul.wide.u16 %r235, %rs22, 256;
and.b32 %r236, %r235, 1792;
or.b32 %r237, %r236, %r234;
mul.wide.u32 %rd84, %r237, 4;
add.s64 %rd86, %rd121, %rd84;
ld.global.nc.u32 %r238, [%rd86];
and.b32 %r198, %r238, 252645135;
shr.u32 %r239, %r238, 4;
and.b32 %r202, %r239, 252645135;

dp4a.s32.s32 %r197, %r198, %r7, %r56;


dp4a.s32.s32 %r201, %r202, %r8, %r197;

mul.wide.u16 %r240, %rs22, 32;
and.b32 %r241, %r240, 1792;
or.b32 %r242, %r241, %r231;
mul.wide.u32 %rd87, %r242, 4;
add.s64 %rd88, %rd121, %rd87;
ld.global.nc.u32 %r243, [%rd88];
and.b32 %r206, %r243, 252645135;
shr.u32 %r244, %r243, 4;
and.b32 %r210, %r244, 252645135;

dp4a.s32.s32 %r205, %r206, %r9, %r201;


dp4a.s32.s32 %r209, %r210, %r10, %r205;

and.b16 %rs24, %rs20, 255;
cvt.u32.u16 %r245, %rs24;
mul.wide.u16 %r246, %rs22, 4;
and.b32 %r247, %r246, 1792;
or.b32 %r248, %r247, %r245;
mul.wide.u32 %rd89, %r248, 4;
add.s64 %rd90, %rd121, %rd89;
ld.global.nc.u32 %r249, [%rd90];
and.b32 %r214, %r249, 252645135;
shr.u32 %r250, %r249, 4;
and.b32 %r218, %r250, 252645135;

dp4a.s32.s32 %r213, %r214, %r11, %r209;


dp4a.s32.s32 %r217, %r218, %r12, %r213;

shr.u32 %r251, %r233, 1;
and.b32 %r252, %r251, 1792;
or.b32 %r253, %r252, %r232;
mul.wide.u32 %rd91, %r253, 4;
add.s64 %rd92, %rd121, %rd91;
ld.global.nc.u32 %r254, [%rd92];
and.b32 %r222, %r254, 252645135;
shr.u32 %r255, %r254, 4;
and.b32 %r226, %r255, 252645135;

dp4a.s32.s32 %r221, %r222, %r13, %r217;


dp4a.s32.s32 %r225, %r226, %r14, %r221;

shr.u32 %r257, %r4, 3;
mad.lo.s32 %r258, %r3, %r29, %r257;
mad.lo.s32 %r262, %r26, %r50, %r258;
mul.wide.s32 %rd93, %r262, 50;
add.s64 %rd94, %rd145, %rd93;
ld.global.nc.u16 %rs17, [%rd94];

{ cvt.f32.f16 %f62, %rs17;}


shr.u32 %r263, %r233, 11;
and.b32 %r264, %r263, 14;
or.b32 %r265, %r264, 1;
cvt.rn.f32.s32 %f65, %r265;
mul.ftz.f32 %f66, %f62, %f65;
and.b32 %r266, %r233, 32768;
cvt.rn.f32.s32 %f67, %r266;
fma.rn.ftz.f32 %f68, %f67, 0fB7000000, 0fBF600000;

{.reg .f16 low,high;
mov.b32 {low,high},%r15;
cvt.f32.f16 %f63, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r15;
cvt.f32.f16 %f64, high;}


cvt.rn.f32.s32 %f69, %r225;
mul.ftz.f32 %f70, %f68, %f64;
fma.rn.ftz.f32 %f71, %f63, %f69, %f70;
fma.rn.ftz.f32 %f188, %f66, %f71, %f188;
@%p3 bra $L__BB250_7;

mov.u64 %rd127, iq1s_grid_gpu;
add.s64 %rd95, %rd144, %rd5;
ld.global.u16 %rs26, [%rd95+2];
shr.u16 %rs27, %rs26, 8;
cvt.u32.u16 %r301, %rs27;
ld.global.u16 %rs28, [%rd95+4];
shr.u16 %rs29, %rs28, 8;
cvt.u32.u16 %r302, %rs29;
add.s64 %rd96, %rd144, %rd6;
ld.global.u16 %rs30, [%rd96];
cvt.u32.u16 %r303, %rs30;
and.b16 %rs31, %rs26, 255;
cvt.u32.u16 %r304, %rs31;
mul.wide.u16 %r305, %rs30, 256;
and.b32 %r306, %r305, 1792;
or.b32 %r307, %r306, %r304;
mul.wide.u32 %rd97, %r307, 4;
add.s64 %rd99, %rd127, %rd97;
ld.global.nc.u32 %r308, [%rd99];
and.b32 %r268, %r308, 252645135;
shr.u32 %r309, %r308, 4;
and.b32 %r272, %r309, 252645135;
mov.u32 %r270, 0;

dp4a.s32.s32 %r267, %r268, %r7, %r270;


dp4a.s32.s32 %r271, %r272, %r8, %r267;

mul.wide.u16 %r310, %rs30, 32;
and.b32 %r311, %r310, 1792;
or.b32 %r312, %r311, %r301;
mul.wide.u32 %rd100, %r312, 4;
add.s64 %rd101, %rd127, %rd100;
ld.global.nc.u32 %r313, [%rd101];
and.b32 %r276, %r313, 252645135;
shr.u32 %r314, %r313, 4;
and.b32 %r280, %r314, 252645135;

dp4a.s32.s32 %r275, %r276, %r9, %r271;


dp4a.s32.s32 %r279, %r280, %r10, %r275;

and.b16 %rs32, %rs28, 255;
cvt.u32.u16 %r315, %rs32;
mul.wide.u16 %r316, %rs30, 4;
and.b32 %r317, %r316, 1792;
or.b32 %r318, %r317, %r315;
mul.wide.u32 %rd102, %r318, 4;
add.s64 %rd103, %rd127, %rd102;
ld.global.nc.u32 %r319, [%rd103];
and.b32 %r284, %r319, 252645135;
shr.u32 %r320, %r319, 4;
and.b32 %r288, %r320, 252645135;

dp4a.s32.s32 %r283, %r284, %r11, %r279;


dp4a.s32.s32 %r287, %r288, %r12, %r283;

shr.u32 %r321, %r303, 1;
and.b32 %r322, %r321, 1792;
or.b32 %r323, %r322, %r302;
mul.wide.u32 %rd104, %r323, 4;
add.s64 %rd105, %rd127, %rd104;
ld.global.nc.u32 %r324, [%rd105];
and.b32 %r292, %r324, 252645135;
shr.u32 %r325, %r324, 4;
and.b32 %r296, %r325, 252645135;

dp4a.s32.s32 %r291, %r292, %r13, %r287;


dp4a.s32.s32 %r295, %r296, %r14, %r291;

add.s64 %rd107, %rd144, %rd93;
ld.global.u16 %rs25, [%rd107];

{ cvt.f32.f16 %f72, %rs25;}


shr.u32 %r333, %r303, 11;
and.b32 %r334, %r333, 14;
or.b32 %r335, %r334, 1;
cvt.rn.f32.s32 %f75, %r335;
mul.ftz.f32 %f76, %f72, %f75;
and.b32 %r336, %r303, 32768;
cvt.rn.f32.s32 %f77, %r336;
fma.rn.ftz.f32 %f78, %f77, 0fB7000000, 0fBF600000;

{.reg .f16 low,high;
mov.b32 {low,high},%r15;
cvt.f32.f16 %f73, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r15;
cvt.f32.f16 %f74, high;}


cvt.rn.f32.s32 %f79, %r295;
mul.ftz.f32 %f80, %f78, %f74;
fma.rn.ftz.f32 %f81, %f73, %f79, %f80;
fma.rn.ftz.f32 %f186, %f76, %f81, %f186;

$L__BB250_7:
add.s64 %rd145, %rd145, 200;
add.s64 %rd144, %rd144, 200;
add.s64 %rd143, %rd143, 1152;
add.s32 %r430, %r430, 4;
setp.lt.u32 %p5, %r430, %r35;
@%p5 bra $L__BB250_3;

add.u64 %rd135, %SPL, 0;
add.u64 %rd129, %SPL, 8;
st.local.v2.f32 [%rd135], {%f189, %f188};
st.local.v2.f32 [%rd129], {%f187, %f186};

$L__BB250_9:
mov.b32 %r17, %f189;
setp.ne.s64 %p6, %rd22, 0;
@%p6 bra $L__BB250_11;
bra.uni $L__BB250_10;

$L__BB250_11:
add.u64 %rd141, %SPL, 0;
add.u64 %rd133, %SPL, 8;
mov.u32 %r364, 31;
mov.u32 %r365, 16;
mov.u32 %r366, -1;
shfl.sync.bfly.b32 %r367|%p17, %r17, %r365, %r364, %r366;
mov.b32 %f102, %r367;
add.ftz.f32 %f103, %f189, %f102;
mov.b32 %r368, %f103;
mov.u32 %r369, 8;
shfl.sync.bfly.b32 %r370|%p18, %r368, %r369, %r364, %r366;
mov.b32 %f104, %r370;
add.ftz.f32 %f105, %f103, %f104;
mov.b32 %r371, %f105;
mov.u32 %r372, 4;
shfl.sync.bfly.b32 %r373|%p19, %r371, %r372, %r364, %r366;
mov.b32 %f106, %r373;
add.ftz.f32 %f107, %f105, %f106;
mov.b32 %r374, %f107;
mov.u32 %r375, 2;
shfl.sync.bfly.b32 %r376|%p20, %r374, %r375, %r364, %r366;
mov.b32 %f108, %r376;
add.ftz.f32 %f109, %f107, %f108;
mov.b32 %r377, %f109;
mov.u32 %r378, 1;
shfl.sync.bfly.b32 %r379|%p21, %r377, %r378, %r364, %r366;
mov.b32 %f110, %r379;
add.ftz.f32 %f111, %f109, %f110;
st.local.f32 [%rd141], %f111;
mov.b32 %r380, %f187;
shfl.sync.bfly.b32 %r381|%p22, %r380, %r365, %r364, %r366;
mov.b32 %f112, %r381;
add.ftz.f32 %f113, %f187, %f112;
mov.b32 %r382, %f113;
shfl.sync.bfly.b32 %r383|%p23, %r382, %r369, %r364, %r366;
mov.b32 %f114, %r383;
add.ftz.f32 %f115, %f113, %f114;
mov.b32 %r384, %f115;
shfl.sync.bfly.b32 %r385|%p24, %r384, %r372, %r364, %r366;
mov.b32 %f116, %r385;
add.ftz.f32 %f117, %f115, %f116;
mov.b32 %r386, %f117;
shfl.sync.bfly.b32 %r387|%p25, %r386, %r375, %r364, %r366;
mov.b32 %f118, %r387;
add.ftz.f32 %f119, %f117, %f118;
mov.b32 %r388, %f119;
shfl.sync.bfly.b32 %r389|%p26, %r388, %r378, %r364, %r366;
mov.b32 %f120, %r389;
add.ftz.f32 %f121, %f119, %f120;
st.local.f32 [%rd133], %f121;
mov.b32 %r390, %f188;
shfl.sync.bfly.b32 %r391|%p27, %r390, %r365, %r364, %r366;
mov.b32 %f122, %r391;
add.ftz.f32 %f123, %f188, %f122;
mov.b32 %r392, %f123;
shfl.sync.bfly.b32 %r393|%p28, %r392, %r369, %r364, %r366;
mov.b32 %f124, %r393;
add.ftz.f32 %f125, %f123, %f124;
mov.b32 %r394, %f125;
shfl.sync.bfly.b32 %r395|%p29, %r394, %r372, %r364, %r366;
mov.b32 %f126, %r395;
add.ftz.f32 %f127, %f125, %f126;
mov.b32 %r396, %f127;
shfl.sync.bfly.b32 %r397|%p30, %r396, %r375, %r364, %r366;
mov.b32 %f128, %r397;
add.ftz.f32 %f129, %f127, %f128;
mov.b32 %r398, %f129;
shfl.sync.bfly.b32 %r399|%p31, %r398, %r378, %r364, %r366;
mov.b32 %f130, %r399;
add.ftz.f32 %f131, %f129, %f130;
st.local.f32 [%rd141+4], %f131;
mov.b32 %r400, %f186;
shfl.sync.bfly.b32 %r401|%p32, %r400, %r365, %r364, %r366;
mov.b32 %f132, %r401;
add.ftz.f32 %f133, %f186, %f132;
mov.b32 %r402, %f133;
shfl.sync.bfly.b32 %r403|%p33, %r402, %r369, %r364, %r366;
mov.b32 %f134, %r403;
add.ftz.f32 %f135, %f133, %f134;
mov.b32 %r404, %f135;
shfl.sync.bfly.b32 %r405|%p34, %r404, %r372, %r364, %r366;
mov.b32 %f136, %r405;
add.ftz.f32 %f137, %f135, %f136;
mov.b32 %r406, %f137;
shfl.sync.bfly.b32 %r407|%p35, %r406, %r375, %r364, %r366;
mov.b32 %f138, %r407;
add.ftz.f32 %f139, %f137, %f138;
mov.b32 %r408, %f139;
shfl.sync.bfly.b32 %r409|%p36, %r408, %r378, %r364, %r366;
mov.b32 %f140, %r409;
add.ftz.f32 %f141, %f139, %f140;
st.local.f32 [%rd133+4], %f141;
bra.uni $L__BB250_12;

$L__BB250_10:
add.u64 %rd137, %SPL, 0;
mov.u32 %r338, 31;
mov.u32 %r339, 16;
mov.u32 %r340, -1;
shfl.sync.bfly.b32 %r341|%p7, %r17, %r339, %r338, %r340;
mov.b32 %f82, %r341;
add.ftz.f32 %f83, %f189, %f82;
mov.b32 %r342, %f83;
mov.u32 %r343, 8;
shfl.sync.bfly.b32 %r344|%p8, %r342, %r343, %r338, %r340;
mov.b32 %f84, %r344;
add.ftz.f32 %f85, %f83, %f84;
mov.b32 %r345, %f85;
mov.u32 %r346, 4;
shfl.sync.bfly.b32 %r347|%p9, %r345, %r346, %r338, %r340;
mov.b32 %f86, %r347;
add.ftz.f32 %f87, %f85, %f86;
mov.b32 %r348, %f87;
mov.u32 %r349, 2;
shfl.sync.bfly.b32 %r350|%p10, %r348, %r349, %r338, %r340;
mov.b32 %f88, %r350;
add.ftz.f32 %f89, %f87, %f88;
mov.b32 %r351, %f89;
mov.u32 %r352, 1;
shfl.sync.bfly.b32 %r353|%p11, %r351, %r352, %r338, %r340;
mov.b32 %f90, %r353;
add.ftz.f32 %f91, %f89, %f90;
st.local.f32 [%rd137], %f91;
mov.b32 %r354, %f188;
shfl.sync.bfly.b32 %r355|%p12, %r354, %r339, %r338, %r340;
mov.b32 %f92, %r355;
add.ftz.f32 %f93, %f188, %f92;
mov.b32 %r356, %f93;
shfl.sync.bfly.b32 %r357|%p13, %r356, %r343, %r338, %r340;
mov.b32 %f94, %r357;
add.ftz.f32 %f95, %f93, %f94;
mov.b32 %r358, %f95;
shfl.sync.bfly.b32 %r359|%p14, %r358, %r346, %r338, %r340;
mov.b32 %f96, %r359;
add.ftz.f32 %f97, %f95, %f96;
mov.b32 %r360, %f97;
shfl.sync.bfly.b32 %r361|%p15, %r360, %r349, %r338, %r340;
mov.b32 %f98, %r361;
add.ftz.f32 %f99, %f97, %f98;
mov.b32 %r362, %f99;
shfl.sync.bfly.b32 %r363|%p16, %r362, %r352, %r338, %r340;
mov.b32 %f100, %r363;
add.ftz.f32 %f101, %f99, %f100;
st.local.f32 [%rd137+4], %f101;

$L__BB250_12:
ld.param.u32 %r419, [_Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7];
setp.gt.u32 %p37, %r4, 1;
mov.u32 %r411, %ctaid.x;
shl.b32 %r412, %r411, 1;
add.s32 %r18, %r412, %r4;
setp.ge.u32 %p38, %r18, %r419;
or.pred %p39, %p37, %p38;
@%p39 bra $L__BB250_29;

add.u64 %rd139, %SPL, 0;
ld.param.u64 %rd122, [_Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3];
ld.param.u32 %r420, [_Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13];
cvt.u64.u32 %rd16, %r4;
mul.wide.u32 %rd108, %r4, 4;
add.s64 %rd109, %rd139, %rd108;
ld.local.f32 %f192, [%rd109];
mad.lo.s32 %r19, %r3, %r420, %r18;
setp.eq.s64 %p40, %rd122, 0;
@%p40 bra $L__BB250_15;

ld.param.u64 %rd123, [_Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3];
cvta.to.global.u64 %rd110, %rd123;
mul.wide.u32 %rd111, %r19, 4;
add.s64 %rd112, %rd110, %rd111;
ld.global.f32 %f142, [%rd112];
add.ftz.f32 %f192, %f192, %f142;

$L__BB250_15:
setp.eq.s64 %p41, %rd22, 0;
@%p41 bra $L__BB250_28;

add.u64 %rd131, %SPL, 8;
ld.param.u64 %rd125, [_Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+16];
shl.b64 %rd113, %rd16, 2;
add.s64 %rd114, %rd131, %rd113;
ld.local.f32 %f24, [%rd114];
setp.eq.s64 %p42, %rd125, 0;
@%p42 bra $L__BB250_18;

ld.param.u64 %rd126, [_Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+16];
cvta.to.global.u64 %rd115, %rd126;
mul.wide.u32 %rd116, %r19, 4;
add.s64 %rd117, %rd115, %rd116;
ld.global.f32 %f143, [%rd117];
add.ftz.f32 %f24, %f24, %f143;

$L__BB250_18:
ld.param.u32 %r425, [_Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.gt.s32 %p43, %r425, 2;
@%p43 bra $L__BB250_22;

ld.param.u32 %r428, [_Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.eq.s32 %p46, %r428, 1;
@%p46 bra $L__BB250_26;

ld.param.u32 %r429, [_Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.eq.s32 %p47, %r429, 2;
@%p47 bra $L__BB250_21;
bra.uni $L__BB250_27;

$L__BB250_21:
mul.ftz.f32 %f163, %f24, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f164, %f163;
add.ftz.f32 %f165, %f164, 0f3F800000;
div.approx.ftz.f32 %f166, %f24, %f165;
mul.ftz.f32 %f192, %f192, %f166;
bra.uni $L__BB250_28;

$L__BB250_22:
ld.param.u32 %r426, [_Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.eq.s32 %p44, %r426, 3;
@%p44 bra $L__BB250_25;

ld.param.u32 %r427, [_Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.ne.s32 %p45, %r427, 6;
@%p45 bra $L__BB250_27;

ld.param.f32 %f175, [_Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+44];
min.ftz.f32 %f144, %f24, %f175;
neg.ftz.f32 %f145, %f175;
min.ftz.f32 %f146, %f192, %f175;
max.ftz.f32 %f147, %f146, %f145;
mul.ftz.f32 %f148, %f144, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f149, %f148;
add.ftz.f32 %f150, %f149, 0f3F800000;
div.approx.ftz.f32 %f151, %f144, %f150;
mul.ftz.f32 %f192, %f147, %f151;
bra.uni $L__BB250_28;

$L__BB250_27:
mul.ftz.f32 %f192, %f192, %f24;
bra.uni $L__BB250_28;

$L__BB250_26:
mul.ftz.f32 %f167, %f24, 0f3F000000;
mul.ftz.f32 %f168, %f24, 0f3D372713;
fma.rn.ftz.f32 %f169, %f24, %f168, 0f3F800000;
mul.ftz.f32 %f170, %f24, 0f3F4C422A;
mul.ftz.f32 %f171, %f170, %f169;
tanh.approx.f32 %f172, %f171;
add.ftz.f32 %f173, %f172, 0f3F800000;
mul.ftz.f32 %f174, %f167, %f173;
mul.ftz.f32 %f192, %f192, %f174;
bra.uni $L__BB250_28;

$L__BB250_25:
mov.f32 %f152, 0f40E00000;
min.ftz.f32 %f153, %f24, %f152;
min.ftz.f32 %f154, %f192, %f152;
mov.f32 %f155, 0fC0E00000;
max.ftz.f32 %f156, %f154, %f155;
mul.ftz.f32 %f157, %f153, 0f3FD9DB23;
mul.ftz.f32 %f158, %f157, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f159, %f158;
add.ftz.f32 %f160, %f159, 0f3F800000;
div.approx.ftz.f32 %f161, %f153, %f160;
add.ftz.f32 %f162, %f156, 0f3F800000;
mul.ftz.f32 %f192, %f162, %f161;

$L__BB250_28:
ld.param.u64 %rd124, [_Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4];
ld.param.u32 %r424, [_Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10];
ld.param.u32 %r423, [_Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13];
mov.u32 %r422, %ctaid.y;
mov.u32 %r421, %tid.y;
mad.lo.s32 %r417, %r421, %r424, %r18;
mad.lo.s32 %r418, %r422, %r423, %r417;
cvta.to.global.u64 %rd118, %rd124;
mul.wide.u32 %rd119, %r418, 4;
add.s64 %rd120, %rd118, %rd119;
st.global.f32 [%rd120], %f192;

$L__BB250_29:
ret;

}
.entry _Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj(
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2,
.param .align 8 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3[48],
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5,
.param .align 4 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6[12],
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15
)
.maxntid 256, 1, 1
.minnctapersm 1
{
.local .align 8 .b8 __local_depot251[8];
.reg .b64 %SP;
.reg .b64 %SPL;
.reg .pred %p<17>;
.reg .b16 %rs<17>;
.reg .f32 %f<59>;
.reg .b32 %r<201>;
.reg .b64 %rd<69>;


mov.u64 %SPL, __local_depot251;
ld.param.u64 %rd15, [_Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0];
ld.param.u64 %rd16, [_Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1];
ld.param.u64 %rd17, [_Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2];
ld.param.u64 %rd18, [_Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4];
ld.param.u32 %r21, [_Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5];
ld.param.u32 %r13, [_Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7];
ld.param.u32 %r14, [_Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8];
ld.param.u32 %r15, [_Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9];
ld.param.u32 %r16, [_Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10];
ld.param.u32 %r17, [_Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11];
ld.param.u32 %r18, [_Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12];
ld.param.u32 %r19, [_Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13];
ld.param.u32 %r22, [_Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14];
ld.param.u32 %r20, [_Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15];
ld.param.u32 %r12, [_Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+8];
ld.param.u32 %r11, [_Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+4];
ld.param.u32 %r10, [_Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6];
add.u64 %rd1, %SPL, 0;
mov.u32 %r23, %ctaid.x;
shl.b32 %r1, %r23, 1;
shr.u32 %r2, %r21, 8;
mov.u32 %r3, %ctaid.y;
mov.u32 %r4, %tid.y;
setp.ge.u32 %p1, %r4, %r22;
@%p1 bra $L__BB251_7;

mov.f32 %f57, 0f00000000;
st.local.v2.f32 [%rd1], {%f57, %f57};
mov.u32 %r5, %tid.x;
shr.u32 %r200, %r5, 3;
setp.ge.u32 %p2, %r200, %r2;
mov.f32 %f58, %f57;
@%p2 bra $L__BB251_5;

mad.lo.s32 %r24, %r4, %r20, %r3;
cvta.to.global.u64 %rd20, %rd17;
mul.wide.u32 %rd21, %r24, 4;
add.s64 %rd22, %rd20, %rd21;
ld.global.nc.u32 %r25, [%rd22];
mul.lo.s32 %r26, %r25, %r17;
and.b32 %r27, %r5, 7;
shl.b32 %r28, %r5, 1;
and.b32 %r29, %r28, 14;
cvt.u64.u32 %rd23, %r29;
shl.b32 %r30, %r5, 2;
cvt.u64.u32 %rd24, %r30;
and.b64 %rd25, %rd24, 28;
mad.lo.s32 %r31, %r1, %r14, %r200;
add.s32 %r32, %r31, %r26;
mul.wide.s32 %rd3, %r32, 50;
add.s64 %rd26, %rd25, %rd3;
add.s64 %rd4, %rd26, 4;
add.s64 %rd27, %rd23, %rd3;
add.s64 %rd5, %rd27, 34;
add.s32 %r33, %r1, 1;
add.s32 %r34, %r200, %r26;
mad.lo.s32 %r35, %r14, %r33, %r34;
mul.wide.s32 %rd6, %r35, 50;
add.s64 %rd28, %rd25, %rd6;
add.s64 %rd7, %rd28, 4;
add.s64 %rd29, %rd23, %rd6;
add.s64 %rd8, %rd29, 34;
mul.hi.u32 %r36, %r3, %r10;
add.s32 %r37, %r36, %r3;
shr.u32 %r38, %r37, %r11;
mul.lo.s32 %r39, %r38, %r12;
sub.s32 %r40, %r3, %r39;
mul.lo.s32 %r41, %r40, %r18;
mul.wide.u32 %rd30, %r41, 36;
mul.lo.s32 %r42, %r4, %r15;
mul.wide.u32 %rd31, %r42, 36;
add.s64 %rd32, %rd31, %rd30;
mul.wide.s32 %rd33, %r27, 36;
add.s64 %rd34, %rd32, %rd33;
shl.b32 %r43, %r200, 3;
mul.wide.s32 %rd35, %r43, 36;
add.s64 %rd36, %rd34, %rd35;
cvta.to.global.u64 %rd37, %rd16;
add.s64 %rd38, %rd37, %rd36;
add.s64 %rd67, %rd38, 16;
cvta.to.global.u64 %rd68, %rd15;
mov.f32 %f57, 0f00000000;

$L__BB251_3:
add.s64 %rd39, %rd68, %rd4;
ld.global.nc.u16 %rs3, [%rd39+-2];
shr.u16 %rs4, %rs3, 8;
cvt.u32.u16 %r112, %rs4;
ld.global.nc.u16 %rs5, [%rd39];
shr.u16 %rs6, %rs5, 8;
cvt.u32.u16 %r113, %rs6;
add.s64 %rd40, %rd68, %rd5;
ld.global.nc.u16 %rs7, [%rd40];
cvt.u32.u16 %r114, %rs7;
and.b16 %rs8, %rs3, 255;
cvt.u32.u16 %r115, %rs8;
mul.wide.u16 %r116, %rs7, 256;
and.b32 %r117, %r116, 1792;
or.b32 %r118, %r117, %r115;
mul.wide.u32 %rd41, %r118, 4;
mov.u64 %rd42, iq1s_grid_gpu;
add.s64 %rd43, %rd42, %rd41;
ld.global.nc.u32 %r119, [%rd43];
and.b32 %r45, %r119, 252645135;
shr.u32 %r120, %r119, 4;
and.b32 %r49, %r120, 252645135;
ld.global.nc.u32 %r46, [%rd67+-12];
ld.global.nc.u32 %r50, [%rd67+-8];
mov.u32 %r81, 0;

dp4a.s32.s32 %r44, %r45, %r46, %r81;


dp4a.s32.s32 %r48, %r49, %r50, %r44;

mul.wide.u16 %r121, %rs7, 32;
and.b32 %r122, %r121, 1792;
or.b32 %r123, %r122, %r112;
mul.wide.u32 %rd44, %r123, 4;
add.s64 %rd45, %rd42, %rd44;
ld.global.nc.u32 %r124, [%rd45];
and.b32 %r53, %r124, 252645135;
shr.u32 %r125, %r124, 4;
and.b32 %r57, %r125, 252645135;
ld.global.nc.u32 %r54, [%rd67+-4];
ld.global.nc.u32 %r58, [%rd67];

dp4a.s32.s32 %r52, %r53, %r54, %r48;


dp4a.s32.s32 %r56, %r57, %r58, %r52;

and.b16 %rs9, %rs5, 255;
cvt.u32.u16 %r126, %rs9;
mul.wide.u16 %r127, %rs7, 4;
and.b32 %r128, %r127, 1792;
or.b32 %r129, %r128, %r126;
mul.wide.u32 %rd46, %r129, 4;
add.s64 %rd47, %rd42, %rd46;
ld.global.nc.u32 %r130, [%rd47];
and.b32 %r61, %r130, 252645135;
shr.u32 %r131, %r130, 4;
and.b32 %r65, %r131, 252645135;
ld.global.nc.u32 %r62, [%rd67+4];
ld.global.nc.u32 %r66, [%rd67+8];

dp4a.s32.s32 %r60, %r61, %r62, %r56;


dp4a.s32.s32 %r64, %r65, %r66, %r60;

shr.u32 %r132, %r114, 1;
and.b32 %r133, %r132, 1792;
or.b32 %r134, %r133, %r113;
mul.wide.u32 %rd48, %r134, 4;
add.s64 %rd49, %rd42, %rd48;
ld.global.nc.u32 %r135, [%rd49];
and.b32 %r69, %r135, 252645135;
shr.u32 %r136, %r135, 4;
and.b32 %r73, %r136, 252645135;
ld.global.nc.u32 %r70, [%rd67+12];
ld.global.nc.u32 %r74, [%rd67+16];

dp4a.s32.s32 %r68, %r69, %r70, %r64;


dp4a.s32.s32 %r72, %r73, %r74, %r68;

add.s64 %rd50, %rd68, %rd3;
ld.global.nc.u16 %rs1, [%rd50];

{ cvt.f32.f16 %f11, %rs1;}


shr.u32 %r137, %r114, 11;
and.b32 %r138, %r137, 14;
or.b32 %r139, %r138, 1;
cvt.rn.f32.s32 %f17, %r139;
mul.ftz.f32 %f18, %f11, %f17;
and.b32 %r140, %r114, 32768;
cvt.rn.f32.s32 %f19, %r140;
mul.ftz.f32 %f20, %f19, 0f37000000;
mov.f32 %f21, 0fBF600000;
sub.ftz.f32 %f22, %f21, %f20;
ld.global.nc.u32 %r76, [%rd67+-16];

{.reg .f16 low,high;
mov.b32 {low,high},%r76;
cvt.f32.f16 %f12, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r76;
cvt.f32.f16 %f13, high;}


cvt.rn.f32.s32 %f23, %r72;
mul.ftz.f32 %f24, %f22, %f13;
fma.rn.ftz.f32 %f25, %f12, %f23, %f24;
fma.rn.ftz.f32 %f58, %f18, %f25, %f58;
add.s64 %rd51, %rd68, %rd7;
ld.global.nc.u16 %rs10, [%rd51+-2];
shr.u16 %rs11, %rs10, 8;
cvt.u32.u16 %r141, %rs11;
ld.global.nc.u16 %rs12, [%rd51];
shr.u16 %rs13, %rs12, 8;
cvt.u32.u16 %r142, %rs13;
add.s64 %rd52, %rd68, %rd8;
ld.global.nc.u16 %rs14, [%rd52];
cvt.u32.u16 %r143, %rs14;
and.b16 %rs15, %rs10, 255;
cvt.u32.u16 %r144, %rs15;
mul.wide.u16 %r145, %rs14, 256;
and.b32 %r146, %r145, 1792;
or.b32 %r147, %r146, %r144;
mul.wide.u32 %rd53, %r147, 4;
add.s64 %rd54, %rd42, %rd53;
ld.global.nc.u32 %r148, [%rd54];
and.b32 %r79, %r148, 252645135;
shr.u32 %r149, %r148, 4;
and.b32 %r83, %r149, 252645135;

dp4a.s32.s32 %r78, %r79, %r46, %r81;


dp4a.s32.s32 %r82, %r83, %r50, %r78;

mul.wide.u16 %r150, %rs14, 32;
and.b32 %r151, %r150, 1792;
or.b32 %r152, %r151, %r141;
mul.wide.u32 %rd55, %r152, 4;
add.s64 %rd56, %rd42, %rd55;
ld.global.nc.u32 %r153, [%rd56];
and.b32 %r87, %r153, 252645135;
shr.u32 %r154, %r153, 4;
and.b32 %r91, %r154, 252645135;

dp4a.s32.s32 %r86, %r87, %r54, %r82;


dp4a.s32.s32 %r90, %r91, %r58, %r86;

and.b16 %rs16, %rs12, 255;
cvt.u32.u16 %r155, %rs16;
mul.wide.u16 %r156, %rs14, 4;
and.b32 %r157, %r156, 1792;
or.b32 %r158, %r157, %r155;
mul.wide.u32 %rd57, %r158, 4;
add.s64 %rd58, %rd42, %rd57;
ld.global.nc.u32 %r159, [%rd58];
and.b32 %r95, %r159, 252645135;
shr.u32 %r160, %r159, 4;
and.b32 %r99, %r160, 252645135;

dp4a.s32.s32 %r94, %r95, %r62, %r90;


dp4a.s32.s32 %r98, %r99, %r66, %r94;

shr.u32 %r161, %r143, 1;
and.b32 %r162, %r161, 1792;
or.b32 %r163, %r162, %r142;
mul.wide.u32 %rd59, %r163, 4;
add.s64 %rd60, %rd42, %rd59;
ld.global.nc.u32 %r164, [%rd60];
and.b32 %r103, %r164, 252645135;
shr.u32 %r165, %r164, 4;
and.b32 %r107, %r165, 252645135;

dp4a.s32.s32 %r102, %r103, %r70, %r98;


dp4a.s32.s32 %r106, %r107, %r74, %r102;

add.s64 %rd61, %rd68, %rd6;
ld.global.nc.u16 %rs2, [%rd61];

{ cvt.f32.f16 %f14, %rs2;}


shr.u32 %r166, %r143, 11;
and.b32 %r167, %r166, 14;
or.b32 %r168, %r167, 1;
cvt.rn.f32.s32 %f26, %r168;
mul.ftz.f32 %f27, %f14, %f26;
and.b32 %r169, %r143, 32768;
cvt.rn.f32.s32 %f28, %r169;
mul.ftz.f32 %f29, %f28, 0f37000000;
sub.ftz.f32 %f30, %f21, %f29;

{.reg .f16 low,high;
mov.b32 {low,high},%r76;
cvt.f32.f16 %f15, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r76;
cvt.f32.f16 %f16, high;}


cvt.rn.f32.s32 %f31, %r106;
mul.ftz.f32 %f32, %f30, %f16;
fma.rn.ftz.f32 %f33, %f15, %f31, %f32;
fma.rn.ftz.f32 %f57, %f27, %f33, %f57;
add.s64 %rd68, %rd68, 200;
add.s64 %rd67, %rd67, 1152;
add.s32 %r200, %r200, 4;
setp.lt.u32 %p3, %r200, %r2;
@%p3 bra $L__BB251_3;

st.local.v2.f32 [%rd1], {%f58, %f57};

$L__BB251_5:
mov.b32 %r170, %f58;
mov.u32 %r171, 31;
mov.u32 %r172, 16;
mov.u32 %r173, -1;
shfl.sync.bfly.b32 %r174|%p4, %r170, %r172, %r171, %r173;
mov.b32 %f34, %r174;
add.ftz.f32 %f35, %f58, %f34;
mov.b32 %r175, %f35;
mov.u32 %r176, 8;
shfl.sync.bfly.b32 %r177|%p5, %r175, %r176, %r171, %r173;
mov.b32 %f36, %r177;
add.ftz.f32 %f37, %f35, %f36;
mov.b32 %r178, %f37;
mov.u32 %r179, 4;
shfl.sync.bfly.b32 %r180|%p6, %r178, %r179, %r171, %r173;
mov.b32 %f38, %r180;
add.ftz.f32 %f39, %f37, %f38;
mov.b32 %r181, %f39;
mov.u32 %r182, 2;
shfl.sync.bfly.b32 %r183|%p7, %r181, %r182, %r171, %r173;
mov.b32 %f40, %r183;
add.ftz.f32 %f41, %f39, %f40;
mov.b32 %r184, %f41;
mov.u32 %r185, 1;
shfl.sync.bfly.b32 %r186|%p8, %r184, %r185, %r171, %r173;
mov.b32 %f42, %r186;
add.ftz.f32 %f43, %f41, %f42;
st.local.f32 [%rd1], %f43;
mov.b32 %r187, %f57;
shfl.sync.bfly.b32 %r188|%p9, %r187, %r172, %r171, %r173;
mov.b32 %f44, %r188;
add.ftz.f32 %f45, %f57, %f44;
mov.b32 %r189, %f45;
shfl.sync.bfly.b32 %r190|%p10, %r189, %r176, %r171, %r173;
mov.b32 %f46, %r190;
add.ftz.f32 %f47, %f45, %f46;
mov.b32 %r191, %f47;
shfl.sync.bfly.b32 %r192|%p11, %r191, %r179, %r171, %r173;
mov.b32 %f48, %r192;
add.ftz.f32 %f49, %f47, %f48;
mov.b32 %r193, %f49;
shfl.sync.bfly.b32 %r194|%p12, %r193, %r182, %r171, %r173;
mov.b32 %f50, %r194;
add.ftz.f32 %f51, %f49, %f50;
mov.b32 %r195, %f51;
shfl.sync.bfly.b32 %r196|%p13, %r195, %r185, %r171, %r173;
mov.b32 %f52, %r196;
add.ftz.f32 %f53, %f51, %f52;
st.local.f32 [%rd1+4], %f53;
add.s32 %r9, %r1, %r5;
setp.ge.u32 %p14, %r9, %r13;
setp.gt.u32 %p15, %r5, 1;
or.pred %p16, %p15, %p14;
@%p16 bra $L__BB251_7;

mul.wide.u32 %rd62, %r5, 4;
add.s64 %rd63, %rd1, %rd62;
ld.local.f32 %f54, [%rd63];
mad.lo.s32 %r198, %r4, %r16, %r9;
mad.lo.s32 %r199, %r3, %r19, %r198;
cvta.to.global.u64 %rd64, %rd18;
mul.wide.u32 %rd65, %r199, 4;
add.s64 %rd66, %rd64, %rd65;
st.global.f32 [%rd66], %f54;

$L__BB251_7:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<320>;
.reg .b16 %rs<65>;
.reg .f32 %f<1053>;
.reg .b32 %r<1276>;
.reg .b64 %rd<187>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate[1536];

ld.param.u64 %rd35, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd36, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd37, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u32 %r63, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r49, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r50, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r54, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r55, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r56, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r60, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r61, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r62, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r58, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r57, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r52, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r51, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r48, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r47, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r46, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
ld.param.u64 %rd40, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+16];
ld.param.u64 %rd39, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+8];
ld.param.u64 %rd38, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3];
mov.u32 %r1, %tid.y;
shl.b32 %r64, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r64, %r2;
mov.u32 %r65, %ctaid.x;
shl.b32 %r4, %r65, 2;
shr.u32 %r5, %r63, 8;
mov.u32 %r6, %ctaid.y;
setp.eq.s64 %p1, %rd37, 0;
@%p1 bra $L__BB252_2;

cvta.to.global.u64 %rd44, %rd37;
mul.wide.u32 %rd45, %r6, 4;
add.s64 %rd46, %rd44, %rd45;
ld.global.nc.u32 %r1272, [%rd46];
mul.hi.u32 %r66, %r6, %r46;
add.s32 %r67, %r66, %r6;
shr.u32 %r68, %r67, %r47;
mul.lo.s32 %r69, %r68, %r48;
sub.s32 %r1273, %r6, %r69;
bra.uni $L__BB252_3;

$L__BB252_2:
mul.hi.u32 %r70, %r6, %r51;
add.s32 %r71, %r70, %r6;
shr.u32 %r1272, %r71, %r52;
mov.u32 %r1273, %r6;

$L__BB252_3:
mov.u32 %r12, %ctaid.z;
selp.b32 %r13, %r6, %r1272, %p1;
setp.ne.s32 %p3, %r1, 0;
setp.gt.u32 %p4, %r2, 3;
or.pred %p5, %p4, %p3;
add.s32 %r72, %r4, %r2;
setp.ge.u32 %p6, %r72, %r50;
mov.f32 %f27, 0f00000000;
or.pred %p7, %p5, %p6;
mov.f32 %f4, %f27;
mov.f32 %f5, %f27;
@%p7 bra $L__BB252_8;

setp.eq.s64 %p8, %rd38, 0;
mov.f32 %f4, 0f00000000;
mov.f32 %f5, %f4;
@%p8 bra $L__BB252_6;

cvta.to.global.u64 %rd47, %rd38;
mul.lo.s32 %r73, %r12, %r62;
cvt.u64.u32 %rd48, %r73;
mul.lo.s32 %r74, %r13, %r56;
cvt.u64.u32 %rd49, %r74;
cvt.u64.u32 %rd50, %r2;
cvt.s64.s32 %rd51, %r4;
add.s64 %rd52, %rd51, %rd50;
add.s64 %rd53, %rd52, %rd48;
add.s64 %rd54, %rd53, %rd49;
shl.b64 %rd55, %rd54, 2;
add.s64 %rd56, %rd47, %rd55;
ld.global.f32 %f5, [%rd56];

$L__BB252_6:
setp.eq.s64 %p9, %rd40, 0;
setp.eq.s64 %p10, %rd39, 0;
or.pred %p11, %p9, %p10;
@%p11 bra $L__BB252_8;

cvta.to.global.u64 %rd57, %rd40;
mul.lo.s32 %r75, %r12, %r62;
cvt.u64.u32 %rd58, %r75;
mul.lo.s32 %r76, %r13, %r56;
cvt.u64.u32 %rd59, %r76;
cvt.u64.u32 %rd60, %r2;
cvt.s64.s32 %rd61, %r4;
add.s64 %rd62, %rd61, %rd60;
add.s64 %rd63, %rd62, %rd58;
add.s64 %rd64, %rd63, %rd59;
shl.b64 %rd65, %rd64, 2;
add.s64 %rd66, %rd57, %rd65;
ld.global.f32 %f4, [%rd66];

$L__BB252_8:
mul.lo.s32 %r77, %r12, %r61;
cvt.u64.u32 %rd67, %r77;
mul.lo.s32 %r78, %r1273, %r55;
cvt.u64.u32 %rd68, %r78;
add.s64 %rd4, %rd68, %rd67;
mul.hi.u32 %r79, %r12, %r57;
add.s32 %r80, %r79, %r12;
shr.u32 %r14, %r80, %r58;
shr.s32 %r81, %r3, 31;
shr.u32 %r82, %r81, 29;
add.s32 %r83, %r3, %r82;
shr.s32 %r1275, %r83, 3;
setp.ge.s32 %p12, %r1275, %r5;
mov.f32 %f28, %f27;
mov.f32 %f29, %f27;
mov.f32 %f30, %f27;
mov.f32 %f31, %f27;
mov.f32 %f32, %f27;
mov.f32 %f33, %f27;
mov.f32 %f34, %f27;
@%p12 bra $L__BB252_19;

mul.lo.s32 %r84, %r14, %r60;
mul.lo.s32 %r85, %r1272, %r54;
and.b32 %r89, %r83, -8;
sub.s32 %r90, %r3, %r89;
shl.b32 %r91, %r90, 1;
cvt.s64.s32 %rd5, %r90;
shl.b32 %r1274, %r1275, 3;
add.s32 %r92, %r1275, %r84;
add.s32 %r93, %r92, %r85;
add.s32 %r94, %r4, 1;
mad.lo.s32 %r95, %r49, %r94, %r93;
mul.wide.s32 %rd6, %r95, 50;
mul.wide.s32 %rd69, %r90, 2;
add.s64 %rd70, %rd6, %rd69;
add.s64 %rd7, %rd70, 34;
mad.lo.s32 %r96, %r4, %r49, %r1275;
add.s32 %r97, %r96, %r84;
add.s32 %r98, %r97, %r85;
mul.wide.s32 %rd8, %r98, 50;
mul.wide.s32 %rd71, %r91, 2;
add.s64 %rd72, %rd8, %rd71;
add.s64 %rd9, %rd72, 4;
add.s64 %rd73, %rd6, %rd71;
add.s64 %rd10, %rd73, 2;
add.s64 %rd74, %rd8, %rd69;
add.s64 %rd11, %rd74, 34;
add.s32 %r99, %r4, 2;
mad.lo.s32 %r100, %r49, %r99, %r93;
mul.wide.s32 %rd12, %r100, 50;
add.s64 %rd75, %rd12, %rd69;
add.s64 %rd13, %rd75, 34;
add.s64 %rd76, %rd12, %rd71;
add.s64 %rd14, %rd76, 2;
add.s32 %r101, %r4, 3;
mad.lo.s32 %r102, %r49, %r101, %r93;
mul.wide.s32 %rd15, %r102, 50;
add.s64 %rd77, %rd15, %rd69;
add.s64 %rd16, %rd77, 34;
add.s64 %rd78, %rd15, %rd71;
add.s64 %rd17, %rd78, 2;
mul.wide.s32 %rd18, %r49, 50;
cvta.to.global.u64 %rd186, %rd39;
cvta.to.global.u64 %rd185, %rd35;
mov.f32 %f27, 0f00000000;
cvta.to.global.u64 %rd79, %rd36;
mov.f32 %f28, %f27;
mov.f32 %f29, %f27;
mov.f32 %f30, %f27;
mov.f32 %f31, %f27;
mov.f32 %f32, %f27;
mov.f32 %f33, %f27;
mov.f32 %f34, %f27;

$L__BB252_10:
cvt.u64.u32 %rd80, %r1274;
add.s64 %rd81, %rd4, %rd80;
add.s64 %rd82, %rd185, %rd9;
ld.global.nc.u16 %rs2, [%rd82+-2];
shr.u16 %rs3, %rs2, 8;
cvt.u32.u16 %r137, %rs3;
ld.global.nc.u16 %rs4, [%rd82];
shr.u16 %rs5, %rs4, 8;
cvt.u32.u16 %r138, %rs5;
add.s64 %rd24, %rd185, %rd11;
ld.global.nc.u16 %rs6, [%rd24];
cvt.u32.u16 %r139, %rs6;
add.s64 %rd83, %rd81, %rd5;
mul.lo.s64 %rd84, %rd83, 36;
and.b16 %rs7, %rs2, 255;
cvt.u32.u16 %r140, %rs7;
mul.wide.u16 %r141, %rs6, 256;
and.b32 %r142, %r141, 1792;
or.b32 %r143, %r142, %r140;
mul.wide.u32 %rd85, %r143, 4;
mov.u64 %rd86, iq1s_grid_gpu;
add.s64 %rd87, %rd86, %rd85;
ld.global.nc.u32 %r144, [%rd87];
and.b32 %r104, %r144, 252645135;
shr.u32 %r145, %r144, 4;
and.b32 %r108, %r145, 252645135;
add.s64 %rd88, %rd79, %rd84;
ld.global.nc.u32 %r19, [%rd88+4];
ld.global.nc.u32 %r20, [%rd88+8];
mov.u32 %r106, 0;

dp4a.s32.s32 %r103, %r104, %r19, %r106;


dp4a.s32.s32 %r107, %r108, %r20, %r103;

mul.wide.u16 %r146, %rs6, 32;
and.b32 %r147, %r146, 1792;
or.b32 %r148, %r147, %r137;
mul.wide.u32 %rd89, %r148, 4;
add.s64 %rd90, %rd86, %rd89;
ld.global.nc.u32 %r149, [%rd90];
and.b32 %r112, %r149, 252645135;
shr.u32 %r150, %r149, 4;
and.b32 %r116, %r150, 252645135;
ld.global.nc.u32 %r21, [%rd88+12];
ld.global.nc.u32 %r22, [%rd88+16];

dp4a.s32.s32 %r111, %r112, %r21, %r107;


dp4a.s32.s32 %r115, %r116, %r22, %r111;

and.b16 %rs8, %rs4, 255;
cvt.u32.u16 %r151, %rs8;
mul.wide.u16 %r152, %rs6, 4;
and.b32 %r153, %r152, 1792;
or.b32 %r154, %r153, %r151;
mul.wide.u32 %rd91, %r154, 4;
add.s64 %rd92, %rd86, %rd91;
ld.global.nc.u32 %r155, [%rd92];
and.b32 %r120, %r155, 252645135;
shr.u32 %r156, %r155, 4;
and.b32 %r124, %r156, 252645135;
ld.global.nc.u32 %r23, [%rd88+20];
ld.global.nc.u32 %r24, [%rd88+24];

dp4a.s32.s32 %r119, %r120, %r23, %r115;


dp4a.s32.s32 %r123, %r124, %r24, %r119;

shr.u32 %r157, %r139, 1;
and.b32 %r158, %r157, 1792;
or.b32 %r159, %r158, %r138;
mul.wide.u32 %rd93, %r159, 4;
add.s64 %rd94, %rd86, %rd93;
ld.global.nc.u32 %r160, [%rd94];
and.b32 %r128, %r160, 252645135;
shr.u32 %r161, %r160, 4;
and.b32 %r132, %r161, 252645135;
ld.global.nc.u32 %r25, [%rd88+28];
ld.global.nc.u32 %r26, [%rd88+32];

dp4a.s32.s32 %r127, %r128, %r25, %r123;


dp4a.s32.s32 %r131, %r132, %r26, %r127;

add.s64 %rd25, %rd185, %rd8;
ld.global.nc.u16 %rs1, [%rd25];

{ cvt.f32.f16 %f101, %rs1;}


shr.u32 %r162, %r139, 11;
and.b32 %r163, %r162, 14;
or.b32 %r164, %r163, 1;
cvt.rn.f32.s32 %f104, %r164;
mul.ftz.f32 %f105, %f101, %f104;
and.b32 %r165, %r139, 32768;
cvt.rn.f32.s32 %f106, %r165;
fma.rn.ftz.f32 %f107, %f106, 0fB7000000, 0fBF600000;
ld.global.nc.u32 %r27, [%rd88];

{.reg .f16 low,high;
mov.b32 {low,high},%r27;
cvt.f32.f16 %f102, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r27;
cvt.f32.f16 %f103, high;}


cvt.rn.f32.s32 %f108, %r131;
mul.ftz.f32 %f109, %f107, %f103;
fma.rn.ftz.f32 %f110, %f102, %f108, %f109;
fma.rn.ftz.f32 %f30, %f105, %f110, %f30;
setp.eq.s64 %p13, %rd39, 0;
@%p13 bra $L__BB252_12;

mov.u64 %rd180, iq1s_grid_gpu;
add.s64 %rd95, %rd186, %rd9;
ld.global.u16 %rs10, [%rd95+-2];
shr.u16 %rs11, %rs10, 8;
cvt.u32.u16 %r200, %rs11;
ld.global.u16 %rs12, [%rd95];
shr.u16 %rs13, %rs12, 8;
cvt.u32.u16 %r201, %rs13;
add.s64 %rd96, %rd186, %rd11;
ld.global.u16 %rs14, [%rd96];
cvt.u32.u16 %r202, %rs14;
and.b16 %rs15, %rs10, 255;
cvt.u32.u16 %r203, %rs15;
mul.wide.u16 %r204, %rs14, 256;
and.b32 %r205, %r204, 1792;
or.b32 %r206, %r205, %r203;
mul.wide.u32 %rd97, %r206, 4;
add.s64 %rd99, %rd180, %rd97;
ld.global.nc.u32 %r207, [%rd99];
and.b32 %r167, %r207, 252645135;
shr.u32 %r208, %r207, 4;
and.b32 %r171, %r208, 252645135;

dp4a.s32.s32 %r166, %r167, %r19, %r106;


dp4a.s32.s32 %r170, %r171, %r20, %r166;

mul.wide.u16 %r209, %rs14, 32;
and.b32 %r210, %r209, 1792;
or.b32 %r211, %r210, %r200;
mul.wide.u32 %rd100, %r211, 4;
add.s64 %rd101, %rd180, %rd100;
ld.global.nc.u32 %r212, [%rd101];
and.b32 %r175, %r212, 252645135;
shr.u32 %r213, %r212, 4;
and.b32 %r179, %r213, 252645135;

dp4a.s32.s32 %r174, %r175, %r21, %r170;


dp4a.s32.s32 %r178, %r179, %r22, %r174;

and.b16 %rs16, %rs12, 255;
cvt.u32.u16 %r214, %rs16;
mul.wide.u16 %r215, %rs14, 4;
and.b32 %r216, %r215, 1792;
or.b32 %r217, %r216, %r214;
mul.wide.u32 %rd102, %r217, 4;
add.s64 %rd103, %rd180, %rd102;
ld.global.nc.u32 %r218, [%rd103];
and.b32 %r183, %r218, 252645135;
shr.u32 %r219, %r218, 4;
and.b32 %r187, %r219, 252645135;

dp4a.s32.s32 %r182, %r183, %r23, %r178;


dp4a.s32.s32 %r186, %r187, %r24, %r182;

shr.u32 %r220, %r202, 1;
and.b32 %r221, %r220, 1792;
or.b32 %r222, %r221, %r201;
mul.wide.u32 %rd104, %r222, 4;
add.s64 %rd105, %rd180, %rd104;
ld.global.nc.u32 %r223, [%rd105];
and.b32 %r191, %r223, 252645135;
shr.u32 %r224, %r223, 4;
and.b32 %r195, %r224, 252645135;

dp4a.s32.s32 %r190, %r191, %r25, %r186;


dp4a.s32.s32 %r194, %r195, %r26, %r190;

add.s64 %rd106, %rd186, %rd8;
ld.global.u16 %rs9, [%rd106];

{ cvt.f32.f16 %f111, %rs9;}


shr.u32 %r225, %r202, 11;
and.b32 %r226, %r225, 14;
or.b32 %r227, %r226, 1;
cvt.rn.f32.s32 %f114, %r227;
mul.ftz.f32 %f115, %f111, %f114;
and.b32 %r228, %r202, 32768;
cvt.rn.f32.s32 %f116, %r228;
fma.rn.ftz.f32 %f117, %f116, 0fB7000000, 0fBF600000;

{.reg .f16 low,high;
mov.b32 {low,high},%r27;
cvt.f32.f16 %f112, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r27;
cvt.f32.f16 %f113, high;}


cvt.rn.f32.s32 %f118, %r194;
mul.ftz.f32 %f119, %f117, %f113;
fma.rn.ftz.f32 %f120, %f112, %f118, %f119;
fma.rn.ftz.f32 %f34, %f115, %f120, %f34;

$L__BB252_12:
add.s64 %rd184, %rd185, %rd9;
add.s64 %rd183, %rd184, -2;
mov.u64 %rd177, iq1s_grid_gpu;
add.s64 %rd26, %rd183, %rd18;
ld.global.nc.u16 %rs18, [%rd26];
shr.u16 %rs19, %rs18, 8;
cvt.u32.u16 %r263, %rs19;
ld.global.nc.u16 %rs20, [%rd26+2];
shr.u16 %rs21, %rs20, 8;
cvt.u32.u16 %r264, %rs21;
add.s64 %rd27, %rd24, %rd18;
ld.global.nc.u16 %rs22, [%rd27];
cvt.u32.u16 %r265, %rs22;
and.b16 %rs23, %rs18, 255;
cvt.u32.u16 %r266, %rs23;
mul.wide.u16 %r267, %rs22, 256;
and.b32 %r268, %r267, 1792;
or.b32 %r269, %r268, %r266;
mul.wide.u32 %rd107, %r269, 4;
add.s64 %rd109, %rd177, %rd107;
ld.global.nc.u32 %r270, [%rd109];
and.b32 %r230, %r270, 252645135;
shr.u32 %r271, %r270, 4;
and.b32 %r234, %r271, 252645135;

dp4a.s32.s32 %r229, %r230, %r19, %r106;


dp4a.s32.s32 %r233, %r234, %r20, %r229;

mul.wide.u16 %r272, %rs22, 32;
and.b32 %r273, %r272, 1792;
or.b32 %r274, %r273, %r263;
mul.wide.u32 %rd110, %r274, 4;
add.s64 %rd111, %rd177, %rd110;
ld.global.nc.u32 %r275, [%rd111];
and.b32 %r238, %r275, 252645135;
shr.u32 %r276, %r275, 4;
and.b32 %r242, %r276, 252645135;

dp4a.s32.s32 %r237, %r238, %r21, %r233;


dp4a.s32.s32 %r241, %r242, %r22, %r237;

and.b16 %rs24, %rs20, 255;
cvt.u32.u16 %r277, %rs24;
mul.wide.u16 %r278, %rs22, 4;
and.b32 %r279, %r278, 1792;
or.b32 %r280, %r279, %r277;
mul.wide.u32 %rd112, %r280, 4;
add.s64 %rd113, %rd177, %rd112;
ld.global.nc.u32 %r281, [%rd113];
and.b32 %r246, %r281, 252645135;
shr.u32 %r282, %r281, 4;
and.b32 %r250, %r282, 252645135;

dp4a.s32.s32 %r245, %r246, %r23, %r241;


dp4a.s32.s32 %r249, %r250, %r24, %r245;

shr.u32 %r283, %r265, 1;
and.b32 %r284, %r283, 1792;
or.b32 %r285, %r284, %r264;
mul.wide.u32 %rd114, %r285, 4;
add.s64 %rd115, %rd177, %rd114;
ld.global.nc.u32 %r286, [%rd115];
and.b32 %r254, %r286, 252645135;
shr.u32 %r287, %r286, 4;
and.b32 %r258, %r287, 252645135;

dp4a.s32.s32 %r253, %r254, %r25, %r249;


dp4a.s32.s32 %r257, %r258, %r26, %r253;

add.s64 %rd28, %rd25, %rd18;
ld.global.nc.u16 %rs17, [%rd28];

{ cvt.f32.f16 %f121, %rs17;}


shr.u32 %r288, %r265, 11;
and.b32 %r289, %r288, 14;
or.b32 %r290, %r289, 1;
cvt.rn.f32.s32 %f124, %r290;
mul.ftz.f32 %f125, %f121, %f124;
and.b32 %r291, %r265, 32768;
cvt.rn.f32.s32 %f126, %r291;
fma.rn.ftz.f32 %f127, %f126, 0fB7000000, 0fBF600000;

{.reg .f16 low,high;
mov.b32 {low,high},%r27;
cvt.f32.f16 %f122, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r27;
cvt.f32.f16 %f123, high;}


cvt.rn.f32.s32 %f128, %r257;
mul.ftz.f32 %f129, %f127, %f123;
fma.rn.ftz.f32 %f130, %f122, %f128, %f129;
fma.rn.ftz.f32 %f29, %f125, %f130, %f29;
@%p13 bra $L__BB252_14;

mov.u64 %rd179, iq1s_grid_gpu;
add.s64 %rd116, %rd186, %rd10;
ld.global.u16 %rs26, [%rd116];
shr.u16 %rs27, %rs26, 8;
cvt.u32.u16 %r326, %rs27;
ld.global.u16 %rs28, [%rd116+2];
shr.u16 %rs29, %rs28, 8;
cvt.u32.u16 %r327, %rs29;
add.s64 %rd117, %rd186, %rd7;
ld.global.u16 %rs30, [%rd117];
cvt.u32.u16 %r328, %rs30;
and.b16 %rs31, %rs26, 255;
cvt.u32.u16 %r329, %rs31;
mul.wide.u16 %r330, %rs30, 256;
and.b32 %r331, %r330, 1792;
or.b32 %r332, %r331, %r329;
mul.wide.u32 %rd118, %r332, 4;
add.s64 %rd120, %rd179, %rd118;
ld.global.nc.u32 %r333, [%rd120];
and.b32 %r293, %r333, 252645135;
shr.u32 %r334, %r333, 4;
and.b32 %r297, %r334, 252645135;
mov.u32 %r295, 0;

dp4a.s32.s32 %r292, %r293, %r19, %r295;


dp4a.s32.s32 %r296, %r297, %r20, %r292;

mul.wide.u16 %r335, %rs30, 32;
and.b32 %r336, %r335, 1792;
or.b32 %r337, %r336, %r326;
mul.wide.u32 %rd121, %r337, 4;
add.s64 %rd122, %rd179, %rd121;
ld.global.nc.u32 %r338, [%rd122];
and.b32 %r301, %r338, 252645135;
shr.u32 %r339, %r338, 4;
and.b32 %r305, %r339, 252645135;

dp4a.s32.s32 %r300, %r301, %r21, %r296;


dp4a.s32.s32 %r304, %r305, %r22, %r300;

and.b16 %rs32, %rs28, 255;
cvt.u32.u16 %r340, %rs32;
mul.wide.u16 %r341, %rs30, 4;
and.b32 %r342, %r341, 1792;
or.b32 %r343, %r342, %r340;
mul.wide.u32 %rd123, %r343, 4;
add.s64 %rd124, %rd179, %rd123;
ld.global.nc.u32 %r344, [%rd124];
and.b32 %r309, %r344, 252645135;
shr.u32 %r345, %r344, 4;
and.b32 %r313, %r345, 252645135;

dp4a.s32.s32 %r308, %r309, %r23, %r304;


dp4a.s32.s32 %r312, %r313, %r24, %r308;

shr.u32 %r346, %r328, 1;
and.b32 %r347, %r346, 1792;
or.b32 %r348, %r347, %r327;
mul.wide.u32 %rd125, %r348, 4;
add.s64 %rd126, %rd179, %rd125;
ld.global.nc.u32 %r349, [%rd126];
and.b32 %r317, %r349, 252645135;
shr.u32 %r350, %r349, 4;
and.b32 %r321, %r350, 252645135;

dp4a.s32.s32 %r316, %r317, %r25, %r312;


dp4a.s32.s32 %r320, %r321, %r26, %r316;

add.s64 %rd127, %rd186, %rd6;
ld.global.u16 %rs25, [%rd127];

{ cvt.f32.f16 %f131, %rs25;}


shr.u32 %r351, %r328, 11;
and.b32 %r352, %r351, 14;
or.b32 %r353, %r352, 1;
cvt.rn.f32.s32 %f134, %r353;
mul.ftz.f32 %f135, %f131, %f134;
and.b32 %r354, %r328, 32768;
cvt.rn.f32.s32 %f136, %r354;
fma.rn.ftz.f32 %f137, %f136, 0fB7000000, 0fBF600000;

{.reg .f16 low,high;
mov.b32 {low,high},%r27;
cvt.f32.f16 %f132, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r27;
cvt.f32.f16 %f133, high;}


cvt.rn.f32.s32 %f138, %r320;
mul.ftz.f32 %f139, %f137, %f133;
fma.rn.ftz.f32 %f140, %f132, %f138, %f139;
fma.rn.ftz.f32 %f33, %f135, %f140, %f33;

$L__BB252_14:
mov.u64 %rd178, iq1s_grid_gpu;
add.s64 %rd29, %rd26, %rd18;
ld.global.nc.u16 %rs34, [%rd29];
shr.u16 %rs35, %rs34, 8;
cvt.u32.u16 %r389, %rs35;
ld.global.nc.u16 %rs36, [%rd29+2];
shr.u16 %rs37, %rs36, 8;
cvt.u32.u16 %r390, %rs37;
add.s64 %rd30, %rd27, %rd18;
ld.global.nc.u16 %rs38, [%rd30];
cvt.u32.u16 %r391, %rs38;
and.b16 %rs39, %rs34, 255;
cvt.u32.u16 %r392, %rs39;
mul.wide.u16 %r393, %rs38, 256;
and.b32 %r394, %r393, 1792;
or.b32 %r395, %r394, %r392;
mul.wide.u32 %rd128, %r395, 4;
add.s64 %rd130, %rd178, %rd128;
ld.global.nc.u32 %r396, [%rd130];
and.b32 %r356, %r396, 252645135;
shr.u32 %r397, %r396, 4;
and.b32 %r360, %r397, 252645135;
mov.u32 %r358, 0;

dp4a.s32.s32 %r355, %r356, %r19, %r358;


dp4a.s32.s32 %r359, %r360, %r20, %r355;

mul.wide.u16 %r398, %rs38, 32;
and.b32 %r399, %r398, 1792;
or.b32 %r400, %r399, %r389;
mul.wide.u32 %rd131, %r400, 4;
add.s64 %rd132, %rd178, %rd131;
ld.global.nc.u32 %r401, [%rd132];
and.b32 %r364, %r401, 252645135;
shr.u32 %r402, %r401, 4;
and.b32 %r368, %r402, 252645135;

dp4a.s32.s32 %r363, %r364, %r21, %r359;


dp4a.s32.s32 %r367, %r368, %r22, %r363;

and.b16 %rs40, %rs36, 255;
cvt.u32.u16 %r403, %rs40;
mul.wide.u16 %r404, %rs38, 4;
and.b32 %r405, %r404, 1792;
or.b32 %r406, %r405, %r403;
mul.wide.u32 %rd133, %r406, 4;
add.s64 %rd134, %rd178, %rd133;
ld.global.nc.u32 %r407, [%rd134];
and.b32 %r372, %r407, 252645135;
shr.u32 %r408, %r407, 4;
and.b32 %r376, %r408, 252645135;

dp4a.s32.s32 %r371, %r372, %r23, %r367;


dp4a.s32.s32 %r375, %r376, %r24, %r371;

shr.u32 %r409, %r391, 1;
and.b32 %r410, %r409, 1792;
or.b32 %r411, %r410, %r390;
mul.wide.u32 %rd135, %r411, 4;
add.s64 %rd136, %rd178, %rd135;
ld.global.nc.u32 %r412, [%rd136];
and.b32 %r380, %r412, 252645135;
shr.u32 %r413, %r412, 4;
and.b32 %r384, %r413, 252645135;

dp4a.s32.s32 %r379, %r380, %r25, %r375;


dp4a.s32.s32 %r383, %r384, %r26, %r379;

add.s64 %rd31, %rd28, %rd18;
ld.global.nc.u16 %rs33, [%rd31];

{ cvt.f32.f16 %f141, %rs33;}


shr.u32 %r414, %r391, 11;
and.b32 %r415, %r414, 14;
or.b32 %r416, %r415, 1;
cvt.rn.f32.s32 %f144, %r416;
mul.ftz.f32 %f145, %f141, %f144;
and.b32 %r417, %r391, 32768;
cvt.rn.f32.s32 %f146, %r417;
fma.rn.ftz.f32 %f147, %f146, 0fB7000000, 0fBF600000;

{.reg .f16 low,high;
mov.b32 {low,high},%r27;
cvt.f32.f16 %f142, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r27;
cvt.f32.f16 %f143, high;}


cvt.rn.f32.s32 %f148, %r383;
mul.ftz.f32 %f149, %f147, %f143;
fma.rn.ftz.f32 %f150, %f142, %f148, %f149;
fma.rn.ftz.f32 %f28, %f145, %f150, %f28;
@%p13 bra $L__BB252_16;

mov.u64 %rd182, iq1s_grid_gpu;
add.s64 %rd137, %rd186, %rd14;
ld.global.u16 %rs42, [%rd137];
shr.u16 %rs43, %rs42, 8;
cvt.u32.u16 %r452, %rs43;
ld.global.u16 %rs44, [%rd137+2];
shr.u16 %rs45, %rs44, 8;
cvt.u32.u16 %r453, %rs45;
add.s64 %rd138, %rd186, %rd13;
ld.global.u16 %rs46, [%rd138];
cvt.u32.u16 %r454, %rs46;
and.b16 %rs47, %rs42, 255;
cvt.u32.u16 %r455, %rs47;
mul.wide.u16 %r456, %rs46, 256;
and.b32 %r457, %r456, 1792;
or.b32 %r458, %r457, %r455;
mul.wide.u32 %rd139, %r458, 4;
add.s64 %rd141, %rd182, %rd139;
ld.global.nc.u32 %r459, [%rd141];
and.b32 %r419, %r459, 252645135;
shr.u32 %r460, %r459, 4;
and.b32 %r423, %r460, 252645135;

dp4a.s32.s32 %r418, %r419, %r19, %r358;


dp4a.s32.s32 %r422, %r423, %r20, %r418;

mul.wide.u16 %r461, %rs46, 32;
and.b32 %r462, %r461, 1792;
or.b32 %r463, %r462, %r452;
mul.wide.u32 %rd142, %r463, 4;
add.s64 %rd143, %rd182, %rd142;
ld.global.nc.u32 %r464, [%rd143];
and.b32 %r427, %r464, 252645135;
shr.u32 %r465, %r464, 4;
and.b32 %r431, %r465, 252645135;

dp4a.s32.s32 %r426, %r427, %r21, %r422;


dp4a.s32.s32 %r430, %r431, %r22, %r426;

and.b16 %rs48, %rs44, 255;
cvt.u32.u16 %r466, %rs48;
mul.wide.u16 %r467, %rs46, 4;
and.b32 %r468, %r467, 1792;
or.b32 %r469, %r468, %r466;
mul.wide.u32 %rd144, %r469, 4;
add.s64 %rd145, %rd182, %rd144;
ld.global.nc.u32 %r470, [%rd145];
and.b32 %r435, %r470, 252645135;
shr.u32 %r471, %r470, 4;
and.b32 %r439, %r471, 252645135;

dp4a.s32.s32 %r434, %r435, %r23, %r430;


dp4a.s32.s32 %r438, %r439, %r24, %r434;

shr.u32 %r472, %r454, 1;
and.b32 %r473, %r472, 1792;
or.b32 %r474, %r473, %r453;
mul.wide.u32 %rd146, %r474, 4;
add.s64 %rd147, %rd182, %rd146;
ld.global.nc.u32 %r475, [%rd147];
and.b32 %r443, %r475, 252645135;
shr.u32 %r476, %r475, 4;
and.b32 %r447, %r476, 252645135;

dp4a.s32.s32 %r442, %r443, %r25, %r438;


dp4a.s32.s32 %r446, %r447, %r26, %r442;

add.s64 %rd148, %rd186, %rd12;
ld.global.u16 %rs41, [%rd148];

{ cvt.f32.f16 %f151, %rs41;}


shr.u32 %r477, %r454, 11;
and.b32 %r478, %r477, 14;
or.b32 %r479, %r478, 1;
cvt.rn.f32.s32 %f154, %r479;
mul.ftz.f32 %f155, %f151, %f154;
and.b32 %r480, %r454, 32768;
cvt.rn.f32.s32 %f156, %r480;
fma.rn.ftz.f32 %f157, %f156, 0fB7000000, 0fBF600000;

{.reg .f16 low,high;
mov.b32 {low,high},%r27;
cvt.f32.f16 %f152, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r27;
cvt.f32.f16 %f153, high;}


cvt.rn.f32.s32 %f158, %r446;
mul.ftz.f32 %f159, %f157, %f153;
fma.rn.ftz.f32 %f160, %f152, %f158, %f159;
fma.rn.ftz.f32 %f32, %f155, %f160, %f32;

$L__BB252_16:
mov.u64 %rd175, iq1s_grid_gpu;
add.s64 %rd149, %rd29, %rd18;
ld.global.nc.u16 %rs50, [%rd149];
shr.u16 %rs51, %rs50, 8;
cvt.u32.u16 %r515, %rs51;
ld.global.nc.u16 %rs52, [%rd149+2];
shr.u16 %rs53, %rs52, 8;
cvt.u32.u16 %r516, %rs53;
add.s64 %rd150, %rd30, %rd18;
ld.global.nc.u16 %rs54, [%rd150];
cvt.u32.u16 %r517, %rs54;
and.b16 %rs55, %rs50, 255;
cvt.u32.u16 %r518, %rs55;
mul.wide.u16 %r519, %rs54, 256;
and.b32 %r520, %r519, 1792;
or.b32 %r521, %r520, %r518;
mul.wide.u32 %rd151, %r521, 4;
add.s64 %rd153, %rd175, %rd151;
ld.global.nc.u32 %r522, [%rd153];
and.b32 %r482, %r522, 252645135;
shr.u32 %r523, %r522, 4;
and.b32 %r486, %r523, 252645135;

dp4a.s32.s32 %r481, %r482, %r19, %r358;


dp4a.s32.s32 %r485, %r486, %r20, %r481;

mul.wide.u16 %r524, %rs54, 32;
and.b32 %r525, %r524, 1792;
or.b32 %r526, %r525, %r515;
mul.wide.u32 %rd154, %r526, 4;
add.s64 %rd155, %rd175, %rd154;
ld.global.nc.u32 %r527, [%rd155];
and.b32 %r490, %r527, 252645135;
shr.u32 %r528, %r527, 4;
and.b32 %r494, %r528, 252645135;

dp4a.s32.s32 %r489, %r490, %r21, %r485;


dp4a.s32.s32 %r493, %r494, %r22, %r489;

and.b16 %rs56, %rs52, 255;
cvt.u32.u16 %r529, %rs56;
mul.wide.u16 %r530, %rs54, 4;
and.b32 %r531, %r530, 1792;
or.b32 %r532, %r531, %r529;
mul.wide.u32 %rd156, %r532, 4;
add.s64 %rd157, %rd175, %rd156;
ld.global.nc.u32 %r533, [%rd157];
and.b32 %r498, %r533, 252645135;
shr.u32 %r534, %r533, 4;
and.b32 %r502, %r534, 252645135;

dp4a.s32.s32 %r497, %r498, %r23, %r493;


dp4a.s32.s32 %r501, %r502, %r24, %r497;

shr.u32 %r535, %r517, 1;
and.b32 %r536, %r535, 1792;
or.b32 %r537, %r536, %r516;
mul.wide.u32 %rd158, %r537, 4;
add.s64 %rd159, %rd175, %rd158;
ld.global.nc.u32 %r538, [%rd159];
and.b32 %r506, %r538, 252645135;
shr.u32 %r539, %r538, 4;
and.b32 %r510, %r539, 252645135;

dp4a.s32.s32 %r505, %r506, %r25, %r501;


dp4a.s32.s32 %r509, %r510, %r26, %r505;

add.s64 %rd160, %rd31, %rd18;
ld.global.nc.u16 %rs49, [%rd160];

{ cvt.f32.f16 %f161, %rs49;}


shr.u32 %r540, %r517, 11;
and.b32 %r541, %r540, 14;
or.b32 %r542, %r541, 1;
cvt.rn.f32.s32 %f164, %r542;
mul.ftz.f32 %f165, %f161, %f164;
and.b32 %r543, %r517, 32768;
cvt.rn.f32.s32 %f166, %r543;
fma.rn.ftz.f32 %f167, %f166, 0fB7000000, 0fBF600000;

{.reg .f16 low,high;
mov.b32 {low,high},%r27;
cvt.f32.f16 %f162, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r27;
cvt.f32.f16 %f163, high;}


cvt.rn.f32.s32 %f168, %r509;
mul.ftz.f32 %f169, %f167, %f163;
fma.rn.ftz.f32 %f170, %f162, %f168, %f169;
fma.rn.ftz.f32 %f27, %f165, %f170, %f27;
@%p13 bra $L__BB252_18;

mov.u64 %rd181, iq1s_grid_gpu;
add.s64 %rd161, %rd186, %rd17;
ld.global.u16 %rs58, [%rd161];
shr.u16 %rs59, %rs58, 8;
cvt.u32.u16 %r578, %rs59;
ld.global.u16 %rs60, [%rd161+2];
shr.u16 %rs61, %rs60, 8;
cvt.u32.u16 %r579, %rs61;
add.s64 %rd162, %rd186, %rd16;
ld.global.u16 %rs62, [%rd162];
cvt.u32.u16 %r580, %rs62;
and.b16 %rs63, %rs58, 255;
cvt.u32.u16 %r581, %rs63;
mul.wide.u16 %r582, %rs62, 256;
and.b32 %r583, %r582, 1792;
or.b32 %r584, %r583, %r581;
mul.wide.u32 %rd163, %r584, 4;
add.s64 %rd165, %rd181, %rd163;
ld.global.nc.u32 %r585, [%rd165];
and.b32 %r545, %r585, 252645135;
shr.u32 %r586, %r585, 4;
and.b32 %r549, %r586, 252645135;
mov.u32 %r547, 0;

dp4a.s32.s32 %r544, %r545, %r19, %r547;


dp4a.s32.s32 %r548, %r549, %r20, %r544;

mul.wide.u16 %r587, %rs62, 32;
and.b32 %r588, %r587, 1792;
or.b32 %r589, %r588, %r578;
mul.wide.u32 %rd166, %r589, 4;
add.s64 %rd167, %rd181, %rd166;
ld.global.nc.u32 %r590, [%rd167];
and.b32 %r553, %r590, 252645135;
shr.u32 %r591, %r590, 4;
and.b32 %r557, %r591, 252645135;

dp4a.s32.s32 %r552, %r553, %r21, %r548;


dp4a.s32.s32 %r556, %r557, %r22, %r552;

and.b16 %rs64, %rs60, 255;
cvt.u32.u16 %r592, %rs64;
mul.wide.u16 %r593, %rs62, 4;
and.b32 %r594, %r593, 1792;
or.b32 %r595, %r594, %r592;
mul.wide.u32 %rd168, %r595, 4;
add.s64 %rd169, %rd181, %rd168;
ld.global.nc.u32 %r596, [%rd169];
and.b32 %r561, %r596, 252645135;
shr.u32 %r597, %r596, 4;
and.b32 %r565, %r597, 252645135;

dp4a.s32.s32 %r560, %r561, %r23, %r556;


dp4a.s32.s32 %r564, %r565, %r24, %r560;

shr.u32 %r598, %r580, 1;
and.b32 %r599, %r598, 1792;
or.b32 %r600, %r599, %r579;
mul.wide.u32 %rd170, %r600, 4;
add.s64 %rd171, %rd181, %rd170;
ld.global.nc.u32 %r601, [%rd171];
and.b32 %r569, %r601, 252645135;
shr.u32 %r602, %r601, 4;
and.b32 %r573, %r602, 252645135;

dp4a.s32.s32 %r568, %r569, %r25, %r564;


dp4a.s32.s32 %r572, %r573, %r26, %r568;

add.s64 %rd172, %rd186, %rd15;
ld.global.u16 %rs57, [%rd172];

{ cvt.f32.f16 %f171, %rs57;}


shr.u32 %r603, %r580, 11;
and.b32 %r604, %r603, 14;
or.b32 %r605, %r604, 1;
cvt.rn.f32.s32 %f174, %r605;
mul.ftz.f32 %f175, %f171, %f174;
and.b32 %r606, %r580, 32768;
cvt.rn.f32.s32 %f176, %r606;
fma.rn.ftz.f32 %f177, %f176, 0fB7000000, 0fBF600000;

{.reg .f16 low,high;
mov.b32 {low,high},%r27;
cvt.f32.f16 %f172, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r27;
cvt.f32.f16 %f173, high;}


cvt.rn.f32.s32 %f178, %r572;
mul.ftz.f32 %f179, %f177, %f173;
fma.rn.ftz.f32 %f180, %f172, %f178, %f179;
fma.rn.ftz.f32 %f31, %f175, %f180, %f31;

$L__BB252_18:
add.s32 %r1274, %r1274, 128;
add.s64 %rd186, %rd186, 800;
add.s64 %rd185, %rd185, 800;
add.s32 %r1275, %r1275, 16;
setp.lt.s32 %p17, %r1275, %r5;
@%p17 bra $L__BB252_10;

$L__BB252_19:
mov.u32 %r1261, %tid.y;
setp.eq.s32 %p18, %r1261, 0;
@%p18 bra $L__BB252_23;

mov.u32 %r1262, %tid.y;
setp.ne.s64 %p19, %rd39, 0;
add.s32 %r30, %r1262, -1;
shl.b32 %r609, %r1262, 9;
mov.u32 %r610, _ZZ13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r611, %r610, %r609;
shl.b32 %r612, %r2, 2;
add.s32 %r613, %r611, %r612;
st.shared.f32 [%r613+-512], %f30;
@%p19 bra $L__BB252_22;
bra.uni $L__BB252_21;

$L__BB252_22:
shl.b32 %r614, %r30, 9;
mov.u32 %r615, _ZZ13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r616, %r615, %r614;
add.s32 %r618, %r616, %r612;
st.shared.f32 [%r618], %f34;
add.s32 %r1258, %r613, -512;
st.shared.f32 [%r1258+128], %f29;
st.shared.f32 [%r618+128], %f33;
add.s32 %r1259, %r613, -512;
st.shared.f32 [%r1259+256], %f28;
st.shared.f32 [%r618+256], %f32;
add.s32 %r1260, %r613, -512;
st.shared.f32 [%r1260+384], %f27;
st.shared.f32 [%r618+384], %f31;
bra.uni $L__BB252_23;

$L__BB252_21:
add.s32 %r1255, %r613, -512;
st.shared.f32 [%r1255+128], %f29;
add.s32 %r1256, %r613, -512;
st.shared.f32 [%r1256+256], %f28;
add.s32 %r1257, %r613, -512;
st.shared.f32 [%r1257+384], %f27;

$L__BB252_23:
mov.u32 %r1263, %tid.y;
setp.ne.s32 %p319, %r1263, 0;
bar.sync 0;
@%p319 bra $L__BB252_78;
bra.uni $L__BB252_24;

$L__BB252_78:
ret;

$L__BB252_24:
ld.param.u64 %rd176, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.f32 %f1025, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
ld.param.u32 %r1266, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r1265, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
mov.u32 %r1264, %ctaid.z;
setp.ne.s64 %p21, %rd39, 0;
mov.u32 %r621, %ctaid.y;
mad.lo.s32 %r622, %r621, %r1266, %r4;
mad.lo.s32 %r623, %r1264, %r1265, %r622;
neg.ftz.f32 %f35, %f1025;
shl.b32 %r624, %r2, 2;
mov.u32 %r625, _ZZ13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r32, %r625, %r624;
cvta.to.global.u64 %rd173, %rd176;
mul.wide.u32 %rd174, %r623, 4;
add.s64 %rd34, %rd173, %rd174;
@%p21 bra $L__BB252_33;
bra.uni $L__BB252_25;

$L__BB252_33:
ld.param.u32 %r1267, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
mov.u32 %r699, _ZZ13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r34, %r699, %r624;
setp.gt.s32 %p54, %r1267, 2;
@%p54 bra $L__BB252_44;

ld.param.u32 %r1270, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p57, %r1270, 1;
@%p57 bra $L__BB252_62;

ld.param.u32 %r1271, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p58, %r1271, 2;
@%p58 bra $L__BB252_36;
bra.uni $L__BB252_70;

$L__BB252_36:
ld.shared.f32 %f581, [%r32];
add.ftz.f32 %f582, %f581, %f30;
ld.shared.f32 %f583, [%r34];
add.ftz.f32 %f584, %f583, %f34;
ld.shared.f32 %f585, [%r32+512];
add.ftz.f32 %f586, %f585, %f582;
ld.shared.f32 %f587, [%r34+512];
add.ftz.f32 %f588, %f587, %f584;
ld.shared.f32 %f589, [%r32+1024];
add.ftz.f32 %f590, %f589, %f586;
ld.shared.f32 %f591, [%r34+1024];
add.ftz.f32 %f592, %f591, %f588;
mov.b32 %r922, %f590;
mov.u32 %r923, 31;
mov.u32 %r924, 16;
mov.u32 %r925, -1;
shfl.sync.bfly.b32 %r926|%p163, %r922, %r924, %r923, %r925;
mov.b32 %f593, %r926;
add.ftz.f32 %f594, %f590, %f593;
mov.b32 %r927, %f594;
mov.u32 %r928, 8;
shfl.sync.bfly.b32 %r929|%p164, %r927, %r928, %r923, %r925;
mov.b32 %f595, %r929;
add.ftz.f32 %f596, %f594, %f595;
mov.b32 %r930, %f596;
mov.u32 %r931, 4;
shfl.sync.bfly.b32 %r932|%p165, %r930, %r931, %r923, %r925;
mov.b32 %f597, %r932;
add.ftz.f32 %f598, %f596, %f597;
mov.b32 %r933, %f598;
mov.u32 %r934, 2;
shfl.sync.bfly.b32 %r935|%p166, %r933, %r934, %r923, %r925;
mov.b32 %f599, %r935;
add.ftz.f32 %f600, %f598, %f599;
mov.b32 %r936, %f600;
mov.u32 %r937, 1;
shfl.sync.bfly.b32 %r938|%p167, %r936, %r937, %r923, %r925;
mov.b32 %f601, %r938;
add.ftz.f32 %f56, %f600, %f601;
mov.b32 %r939, %f592;
shfl.sync.bfly.b32 %r940|%p168, %r939, %r924, %r923, %r925;
mov.b32 %f602, %r940;
add.ftz.f32 %f603, %f592, %f602;
mov.b32 %r941, %f603;
shfl.sync.bfly.b32 %r942|%p169, %r941, %r928, %r923, %r925;
mov.b32 %f604, %r942;
add.ftz.f32 %f605, %f603, %f604;
mov.b32 %r943, %f605;
shfl.sync.bfly.b32 %r944|%p170, %r943, %r931, %r923, %r925;
mov.b32 %f606, %r944;
add.ftz.f32 %f607, %f605, %f606;
mov.b32 %r945, %f607;
shfl.sync.bfly.b32 %r946|%p171, %r945, %r934, %r923, %r925;
mov.b32 %f608, %r946;
add.ftz.f32 %f609, %f607, %f608;
mov.b32 %r947, %f609;
shfl.sync.bfly.b32 %r948|%p172, %r947, %r937, %r923, %r925;
mov.b32 %f610, %r948;
add.ftz.f32 %f57, %f609, %f610;
setp.ge.u32 %p173, %r4, %r50;
setp.ne.s32 %p174, %r2, 0;
or.pred %p175, %p174, %p173;
@%p175 bra $L__BB252_38;

add.ftz.f32 %f611, %f5, %f56;
add.ftz.f32 %f612, %f4, %f57;
mul.ftz.f32 %f613, %f612, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f614, %f613;
add.ftz.f32 %f615, %f614, 0f3F800000;
div.approx.ftz.f32 %f616, %f612, %f615;
mul.ftz.f32 %f617, %f611, %f616;
st.global.f32 [%rd34], %f617;

$L__BB252_38:
ld.shared.f32 %f618, [%r32+128];
add.ftz.f32 %f619, %f618, %f29;
ld.shared.f32 %f620, [%r34+128];
add.ftz.f32 %f621, %f620, %f33;
ld.shared.f32 %f622, [%r32+640];
add.ftz.f32 %f623, %f622, %f619;
ld.shared.f32 %f624, [%r34+640];
add.ftz.f32 %f625, %f624, %f621;
ld.shared.f32 %f626, [%r32+1152];
add.ftz.f32 %f627, %f626, %f623;
ld.shared.f32 %f628, [%r34+1152];
add.ftz.f32 %f629, %f628, %f625;
mov.b32 %r949, %f627;
shfl.sync.bfly.b32 %r953|%p176, %r949, %r924, %r923, %r925;
mov.b32 %f630, %r953;
add.ftz.f32 %f631, %f627, %f630;
mov.b32 %r954, %f631;
shfl.sync.bfly.b32 %r956|%p177, %r954, %r928, %r923, %r925;
mov.b32 %f632, %r956;
add.ftz.f32 %f633, %f631, %f632;
mov.b32 %r957, %f633;
shfl.sync.bfly.b32 %r959|%p178, %r957, %r931, %r923, %r925;
mov.b32 %f634, %r959;
add.ftz.f32 %f635, %f633, %f634;
mov.b32 %r960, %f635;
shfl.sync.bfly.b32 %r962|%p179, %r960, %r934, %r923, %r925;
mov.b32 %f636, %r962;
add.ftz.f32 %f637, %f635, %f636;
mov.b32 %r963, %f637;
shfl.sync.bfly.b32 %r965|%p180, %r963, %r937, %r923, %r925;
mov.b32 %f638, %r965;
add.ftz.f32 %f58, %f637, %f638;
mov.b32 %r966, %f629;
shfl.sync.bfly.b32 %r967|%p181, %r966, %r924, %r923, %r925;
mov.b32 %f639, %r967;
add.ftz.f32 %f640, %f629, %f639;
mov.b32 %r968, %f640;
shfl.sync.bfly.b32 %r969|%p182, %r968, %r928, %r923, %r925;
mov.b32 %f641, %r969;
add.ftz.f32 %f642, %f640, %f641;
mov.b32 %r970, %f642;
shfl.sync.bfly.b32 %r971|%p183, %r970, %r931, %r923, %r925;
mov.b32 %f643, %r971;
add.ftz.f32 %f644, %f642, %f643;
mov.b32 %r972, %f644;
shfl.sync.bfly.b32 %r973|%p184, %r972, %r934, %r923, %r925;
mov.b32 %f645, %r973;
add.ftz.f32 %f646, %f644, %f645;
mov.b32 %r974, %f646;
shfl.sync.bfly.b32 %r975|%p185, %r974, %r937, %r923, %r925;
mov.b32 %f647, %r975;
add.ftz.f32 %f59, %f646, %f647;
add.s32 %r976, %r4, 1;
setp.ge.u32 %p186, %r976, %r50;
setp.ne.s32 %p187, %r2, 1;
or.pred %p188, %p187, %p186;
@%p188 bra $L__BB252_40;

add.ftz.f32 %f648, %f5, %f58;
add.ftz.f32 %f649, %f4, %f59;
mul.ftz.f32 %f650, %f649, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f651, %f650;
add.ftz.f32 %f652, %f651, 0f3F800000;
div.approx.ftz.f32 %f653, %f649, %f652;
mul.ftz.f32 %f654, %f648, %f653;
st.global.f32 [%rd34+4], %f654;

$L__BB252_40:
ld.shared.f32 %f655, [%r32+256];
add.ftz.f32 %f656, %f655, %f28;
ld.shared.f32 %f657, [%r34+256];
add.ftz.f32 %f658, %f657, %f32;
ld.shared.f32 %f659, [%r32+768];
add.ftz.f32 %f660, %f659, %f656;
ld.shared.f32 %f661, [%r34+768];
add.ftz.f32 %f662, %f661, %f658;
ld.shared.f32 %f663, [%r32+1280];
add.ftz.f32 %f664, %f663, %f660;
ld.shared.f32 %f665, [%r34+1280];
add.ftz.f32 %f666, %f665, %f662;
mov.b32 %r977, %f664;
mov.u32 %r978, 31;
mov.u32 %r979, 16;
mov.u32 %r980, -1;
shfl.sync.bfly.b32 %r981|%p189, %r977, %r979, %r978, %r980;
mov.b32 %f667, %r981;
add.ftz.f32 %f668, %f664, %f667;
mov.b32 %r982, %f668;
mov.u32 %r983, 8;
shfl.sync.bfly.b32 %r984|%p190, %r982, %r983, %r978, %r980;
mov.b32 %f669, %r984;
add.ftz.f32 %f670, %f668, %f669;
mov.b32 %r985, %f670;
mov.u32 %r986, 4;
shfl.sync.bfly.b32 %r987|%p191, %r985, %r986, %r978, %r980;
mov.b32 %f671, %r987;
add.ftz.f32 %f672, %f670, %f671;
mov.b32 %r988, %f672;
mov.u32 %r989, 2;
shfl.sync.bfly.b32 %r990|%p192, %r988, %r989, %r978, %r980;
mov.b32 %f673, %r990;
add.ftz.f32 %f674, %f672, %f673;
mov.b32 %r991, %f674;
mov.u32 %r992, 1;
shfl.sync.bfly.b32 %r993|%p193, %r991, %r992, %r978, %r980;
mov.b32 %f675, %r993;
add.ftz.f32 %f60, %f674, %f675;
mov.b32 %r994, %f666;
shfl.sync.bfly.b32 %r995|%p194, %r994, %r979, %r978, %r980;
mov.b32 %f676, %r995;
add.ftz.f32 %f677, %f666, %f676;
mov.b32 %r996, %f677;
shfl.sync.bfly.b32 %r997|%p195, %r996, %r983, %r978, %r980;
mov.b32 %f678, %r997;
add.ftz.f32 %f679, %f677, %f678;
mov.b32 %r998, %f679;
shfl.sync.bfly.b32 %r999|%p196, %r998, %r986, %r978, %r980;
mov.b32 %f680, %r999;
add.ftz.f32 %f681, %f679, %f680;
mov.b32 %r1000, %f681;
shfl.sync.bfly.b32 %r1001|%p197, %r1000, %r989, %r978, %r980;
mov.b32 %f682, %r1001;
add.ftz.f32 %f683, %f681, %f682;
mov.b32 %r1002, %f683;
shfl.sync.bfly.b32 %r1003|%p198, %r1002, %r992, %r978, %r980;
mov.b32 %f684, %r1003;
add.ftz.f32 %f61, %f683, %f684;
add.s32 %r1004, %r4, 2;
setp.ge.u32 %p199, %r1004, %r50;
setp.ne.s32 %p200, %r2, 2;
or.pred %p201, %p200, %p199;
@%p201 bra $L__BB252_42;

add.ftz.f32 %f685, %f5, %f60;
add.ftz.f32 %f686, %f4, %f61;
mul.ftz.f32 %f687, %f686, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f688, %f687;
add.ftz.f32 %f689, %f688, 0f3F800000;
div.approx.ftz.f32 %f690, %f686, %f689;
mul.ftz.f32 %f691, %f685, %f690;
st.global.f32 [%rd34+8], %f691;

$L__BB252_42:
ld.shared.f32 %f692, [%r32+384];
add.ftz.f32 %f693, %f692, %f27;
ld.shared.f32 %f694, [%r34+384];
add.ftz.f32 %f695, %f694, %f31;
ld.shared.f32 %f696, [%r32+896];
add.ftz.f32 %f697, %f696, %f693;
ld.shared.f32 %f698, [%r34+896];
add.ftz.f32 %f699, %f698, %f695;
ld.shared.f32 %f700, [%r32+1408];
add.ftz.f32 %f701, %f700, %f697;
ld.shared.f32 %f702, [%r34+1408];
add.ftz.f32 %f703, %f702, %f699;
mov.b32 %r1005, %f701;
shfl.sync.bfly.b32 %r1009|%p202, %r1005, %r979, %r978, %r980;
mov.b32 %f704, %r1009;
add.ftz.f32 %f705, %f701, %f704;
mov.b32 %r1010, %f705;
shfl.sync.bfly.b32 %r1012|%p203, %r1010, %r983, %r978, %r980;
mov.b32 %f706, %r1012;
add.ftz.f32 %f707, %f705, %f706;
mov.b32 %r1013, %f707;
shfl.sync.bfly.b32 %r1015|%p204, %r1013, %r986, %r978, %r980;
mov.b32 %f708, %r1015;
add.ftz.f32 %f709, %f707, %f708;
mov.b32 %r1016, %f709;
shfl.sync.bfly.b32 %r1018|%p205, %r1016, %r989, %r978, %r980;
mov.b32 %f710, %r1018;
add.ftz.f32 %f711, %f709, %f710;
mov.b32 %r1019, %f711;
shfl.sync.bfly.b32 %r1021|%p206, %r1019, %r992, %r978, %r980;
mov.b32 %f712, %r1021;
add.ftz.f32 %f62, %f711, %f712;
mov.b32 %r1022, %f703;
shfl.sync.bfly.b32 %r1023|%p207, %r1022, %r979, %r978, %r980;
mov.b32 %f713, %r1023;
add.ftz.f32 %f714, %f703, %f713;
mov.b32 %r1024, %f714;
shfl.sync.bfly.b32 %r1025|%p208, %r1024, %r983, %r978, %r980;
mov.b32 %f715, %r1025;
add.ftz.f32 %f716, %f714, %f715;
mov.b32 %r1026, %f716;
shfl.sync.bfly.b32 %r1027|%p209, %r1026, %r986, %r978, %r980;
mov.b32 %f717, %r1027;
add.ftz.f32 %f718, %f716, %f717;
mov.b32 %r1028, %f718;
shfl.sync.bfly.b32 %r1029|%p210, %r1028, %r989, %r978, %r980;
mov.b32 %f719, %r1029;
add.ftz.f32 %f720, %f718, %f719;
mov.b32 %r1030, %f720;
shfl.sync.bfly.b32 %r1031|%p211, %r1030, %r992, %r978, %r980;
mov.b32 %f721, %r1031;
add.ftz.f32 %f63, %f720, %f721;
add.s32 %r1032, %r4, 3;
setp.ge.u32 %p212, %r1032, %r50;
setp.ne.s32 %p213, %r2, 3;
or.pred %p214, %p213, %p212;
@%p214 bra $L__BB252_78;

add.ftz.f32 %f722, %f5, %f62;
add.ftz.f32 %f723, %f4, %f63;
mul.ftz.f32 %f724, %f723, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f725, %f724;
add.ftz.f32 %f726, %f725, 0f3F800000;
div.approx.ftz.f32 %f727, %f723, %f726;
mul.ftz.f32 %f728, %f722, %f727;
st.global.f32 [%rd34+12], %f728;
bra.uni $L__BB252_78;

$L__BB252_25:
ld.shared.f32 %f181, [%r32];
add.ftz.f32 %f182, %f181, %f30;
ld.shared.f32 %f183, [%r32+512];
add.ftz.f32 %f184, %f183, %f182;
ld.shared.f32 %f185, [%r32+1024];
add.ftz.f32 %f186, %f185, %f184;
mov.b32 %r626, %f186;
mov.u32 %r627, 31;
mov.u32 %r628, 16;
mov.u32 %r629, -1;
shfl.sync.bfly.b32 %r630|%p22, %r626, %r628, %r627, %r629;
mov.b32 %f187, %r630;
add.ftz.f32 %f188, %f186, %f187;
mov.b32 %r631, %f188;
mov.u32 %r632, 8;
shfl.sync.bfly.b32 %r633|%p23, %r631, %r632, %r627, %r629;
mov.b32 %f189, %r633;
add.ftz.f32 %f190, %f188, %f189;
mov.b32 %r634, %f190;
mov.u32 %r635, 4;
shfl.sync.bfly.b32 %r636|%p24, %r634, %r635, %r627, %r629;
mov.b32 %f191, %r636;
add.ftz.f32 %f192, %f190, %f191;
mov.b32 %r637, %f192;
mov.u32 %r638, 2;
shfl.sync.bfly.b32 %r639|%p25, %r637, %r638, %r627, %r629;
mov.b32 %f193, %r639;
add.ftz.f32 %f194, %f192, %f193;
mov.b32 %r640, %f194;
mov.u32 %r641, 1;
shfl.sync.bfly.b32 %r642|%p26, %r640, %r641, %r627, %r629;
mov.b32 %f195, %r642;
add.ftz.f32 %f36, %f194, %f195;
setp.ge.u32 %p27, %r4, %r50;
setp.ne.s32 %p28, %r2, 0;
or.pred %p29, %p28, %p27;
@%p29 bra $L__BB252_27;

add.ftz.f32 %f196, %f5, %f36;
st.global.f32 [%rd34], %f196;

$L__BB252_27:
ld.shared.f32 %f197, [%r32+128];
add.ftz.f32 %f198, %f197, %f29;
ld.shared.f32 %f199, [%r32+640];
add.ftz.f32 %f200, %f199, %f198;
ld.shared.f32 %f201, [%r32+1152];
add.ftz.f32 %f202, %f201, %f200;
mov.b32 %r643, %f202;
shfl.sync.bfly.b32 %r647|%p30, %r643, %r628, %r627, %r629;
mov.b32 %f203, %r647;
add.ftz.f32 %f204, %f202, %f203;
mov.b32 %r648, %f204;
shfl.sync.bfly.b32 %r650|%p31, %r648, %r632, %r627, %r629;
mov.b32 %f205, %r650;
add.ftz.f32 %f206, %f204, %f205;
mov.b32 %r651, %f206;
shfl.sync.bfly.b32 %r653|%p32, %r651, %r635, %r627, %r629;
mov.b32 %f207, %r653;
add.ftz.f32 %f208, %f206, %f207;
mov.b32 %r654, %f208;
shfl.sync.bfly.b32 %r656|%p33, %r654, %r638, %r627, %r629;
mov.b32 %f209, %r656;
add.ftz.f32 %f210, %f208, %f209;
mov.b32 %r657, %f210;
shfl.sync.bfly.b32 %r659|%p34, %r657, %r641, %r627, %r629;
mov.b32 %f211, %r659;
add.ftz.f32 %f37, %f210, %f211;
add.s32 %r660, %r4, 1;
setp.ge.u32 %p35, %r660, %r50;
setp.ne.s32 %p36, %r2, 1;
or.pred %p37, %p36, %p35;
@%p37 bra $L__BB252_29;

add.ftz.f32 %f212, %f5, %f37;
st.global.f32 [%rd34+4], %f212;

$L__BB252_29:
ld.shared.f32 %f213, [%r32+256];
add.ftz.f32 %f214, %f213, %f28;
ld.shared.f32 %f215, [%r32+768];
add.ftz.f32 %f216, %f215, %f214;
ld.shared.f32 %f217, [%r32+1280];
add.ftz.f32 %f218, %f217, %f216;
mov.b32 %r661, %f218;
mov.u32 %r662, 31;
mov.u32 %r663, 16;
mov.u32 %r664, -1;
shfl.sync.bfly.b32 %r665|%p38, %r661, %r663, %r662, %r664;
mov.b32 %f219, %r665;
add.ftz.f32 %f220, %f218, %f219;
mov.b32 %r666, %f220;
mov.u32 %r667, 8;
shfl.sync.bfly.b32 %r668|%p39, %r666, %r667, %r662, %r664;
mov.b32 %f221, %r668;
add.ftz.f32 %f222, %f220, %f221;
mov.b32 %r669, %f222;
mov.u32 %r670, 4;
shfl.sync.bfly.b32 %r671|%p40, %r669, %r670, %r662, %r664;
mov.b32 %f223, %r671;
add.ftz.f32 %f224, %f222, %f223;
mov.b32 %r672, %f224;
mov.u32 %r673, 2;
shfl.sync.bfly.b32 %r674|%p41, %r672, %r673, %r662, %r664;
mov.b32 %f225, %r674;
add.ftz.f32 %f226, %f224, %f225;
mov.b32 %r675, %f226;
mov.u32 %r676, 1;
shfl.sync.bfly.b32 %r677|%p42, %r675, %r676, %r662, %r664;
mov.b32 %f227, %r677;
add.ftz.f32 %f38, %f226, %f227;
add.s32 %r678, %r4, 2;
setp.ge.u32 %p43, %r678, %r50;
setp.ne.s32 %p44, %r2, 2;
or.pred %p45, %p44, %p43;
@%p45 bra $L__BB252_31;

add.ftz.f32 %f228, %f5, %f38;
st.global.f32 [%rd34+8], %f228;

$L__BB252_31:
ld.shared.f32 %f229, [%r32+384];
add.ftz.f32 %f230, %f229, %f27;
ld.shared.f32 %f231, [%r32+896];
add.ftz.f32 %f232, %f231, %f230;
ld.shared.f32 %f233, [%r32+1408];
add.ftz.f32 %f234, %f233, %f232;
mov.b32 %r679, %f234;
shfl.sync.bfly.b32 %r683|%p46, %r679, %r663, %r662, %r664;
mov.b32 %f235, %r683;
add.ftz.f32 %f236, %f234, %f235;
mov.b32 %r684, %f236;
shfl.sync.bfly.b32 %r686|%p47, %r684, %r667, %r662, %r664;
mov.b32 %f237, %r686;
add.ftz.f32 %f238, %f236, %f237;
mov.b32 %r687, %f238;
shfl.sync.bfly.b32 %r689|%p48, %r687, %r670, %r662, %r664;
mov.b32 %f239, %r689;
add.ftz.f32 %f240, %f238, %f239;
mov.b32 %r690, %f240;
shfl.sync.bfly.b32 %r692|%p49, %r690, %r673, %r662, %r664;
mov.b32 %f241, %r692;
add.ftz.f32 %f242, %f240, %f241;
mov.b32 %r693, %f242;
shfl.sync.bfly.b32 %r695|%p50, %r693, %r676, %r662, %r664;
mov.b32 %f243, %r695;
add.ftz.f32 %f39, %f242, %f243;
add.s32 %r696, %r4, 3;
setp.ge.u32 %p51, %r696, %r50;
setp.ne.s32 %p52, %r2, 3;
or.pred %p53, %p52, %p51;
@%p53 bra $L__BB252_78;

add.ftz.f32 %f244, %f5, %f39;
st.global.f32 [%rd34+12], %f244;
bra.uni $L__BB252_78;

$L__BB252_44:
ld.param.u32 %r1268, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p55, %r1268, 3;
@%p55 bra $L__BB252_54;

ld.param.u32 %r1269, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.ne.s32 %p56, %r1269, 6;
@%p56 bra $L__BB252_70;

ld.shared.f32 %f245, [%r32];
add.ftz.f32 %f246, %f245, %f30;
ld.shared.f32 %f247, [%r34];
add.ftz.f32 %f248, %f247, %f34;
ld.shared.f32 %f249, [%r32+512];
add.ftz.f32 %f250, %f249, %f246;
ld.shared.f32 %f251, [%r34+512];
add.ftz.f32 %f252, %f251, %f248;
ld.shared.f32 %f253, [%r32+1024];
add.ftz.f32 %f254, %f253, %f250;
ld.shared.f32 %f255, [%r34+1024];
add.ftz.f32 %f256, %f255, %f252;
mov.b32 %r700, %f254;
mov.u32 %r701, 31;
mov.u32 %r702, 16;
mov.u32 %r703, -1;
shfl.sync.bfly.b32 %r704|%p59, %r700, %r702, %r701, %r703;
mov.b32 %f257, %r704;
add.ftz.f32 %f258, %f254, %f257;
mov.b32 %r705, %f258;
mov.u32 %r706, 8;
shfl.sync.bfly.b32 %r707|%p60, %r705, %r706, %r701, %r703;
mov.b32 %f259, %r707;
add.ftz.f32 %f260, %f258, %f259;
mov.b32 %r708, %f260;
mov.u32 %r709, 4;
shfl.sync.bfly.b32 %r710|%p61, %r708, %r709, %r701, %r703;
mov.b32 %f261, %r710;
add.ftz.f32 %f262, %f260, %f261;
mov.b32 %r711, %f262;
mov.u32 %r712, 2;
shfl.sync.bfly.b32 %r713|%p62, %r711, %r712, %r701, %r703;
mov.b32 %f263, %r713;
add.ftz.f32 %f264, %f262, %f263;
mov.b32 %r714, %f264;
mov.u32 %r715, 1;
shfl.sync.bfly.b32 %r716|%p63, %r714, %r715, %r701, %r703;
mov.b32 %f265, %r716;
add.ftz.f32 %f40, %f264, %f265;
mov.b32 %r717, %f256;
shfl.sync.bfly.b32 %r718|%p64, %r717, %r702, %r701, %r703;
mov.b32 %f266, %r718;
add.ftz.f32 %f267, %f256, %f266;
mov.b32 %r719, %f267;
shfl.sync.bfly.b32 %r720|%p65, %r719, %r706, %r701, %r703;
mov.b32 %f268, %r720;
add.ftz.f32 %f269, %f267, %f268;
mov.b32 %r721, %f269;
shfl.sync.bfly.b32 %r722|%p66, %r721, %r709, %r701, %r703;
mov.b32 %f270, %r722;
add.ftz.f32 %f271, %f269, %f270;
mov.b32 %r723, %f271;
shfl.sync.bfly.b32 %r724|%p67, %r723, %r712, %r701, %r703;
mov.b32 %f272, %r724;
add.ftz.f32 %f273, %f271, %f272;
mov.b32 %r725, %f273;
shfl.sync.bfly.b32 %r726|%p68, %r725, %r715, %r701, %r703;
mov.b32 %f274, %r726;
add.ftz.f32 %f41, %f273, %f274;
setp.ge.u32 %p69, %r4, %r50;
setp.ne.s32 %p70, %r2, 0;
or.pred %p71, %p70, %p69;
@%p71 bra $L__BB252_48;

ld.param.f32 %f1029, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f275, %f5, %f40;
add.ftz.f32 %f276, %f4, %f41;
min.ftz.f32 %f277, %f276, %f1029;
min.ftz.f32 %f278, %f275, %f1029;
max.ftz.f32 %f279, %f278, %f35;
mul.ftz.f32 %f280, %f277, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f281, %f280;
add.ftz.f32 %f282, %f281, 0f3F800000;
div.approx.ftz.f32 %f283, %f277, %f282;
mul.ftz.f32 %f284, %f279, %f283;
st.global.f32 [%rd34], %f284;

$L__BB252_48:
ld.shared.f32 %f285, [%r32+128];
add.ftz.f32 %f286, %f285, %f29;
ld.shared.f32 %f287, [%r34+128];
add.ftz.f32 %f288, %f287, %f33;
ld.shared.f32 %f289, [%r32+640];
add.ftz.f32 %f290, %f289, %f286;
ld.shared.f32 %f291, [%r34+640];
add.ftz.f32 %f292, %f291, %f288;
ld.shared.f32 %f293, [%r32+1152];
add.ftz.f32 %f294, %f293, %f290;
ld.shared.f32 %f295, [%r34+1152];
add.ftz.f32 %f296, %f295, %f292;
mov.b32 %r727, %f294;
shfl.sync.bfly.b32 %r731|%p72, %r727, %r702, %r701, %r703;
mov.b32 %f297, %r731;
add.ftz.f32 %f298, %f294, %f297;
mov.b32 %r732, %f298;
shfl.sync.bfly.b32 %r734|%p73, %r732, %r706, %r701, %r703;
mov.b32 %f299, %r734;
add.ftz.f32 %f300, %f298, %f299;
mov.b32 %r735, %f300;
shfl.sync.bfly.b32 %r737|%p74, %r735, %r709, %r701, %r703;
mov.b32 %f301, %r737;
add.ftz.f32 %f302, %f300, %f301;
mov.b32 %r738, %f302;
shfl.sync.bfly.b32 %r740|%p75, %r738, %r712, %r701, %r703;
mov.b32 %f303, %r740;
add.ftz.f32 %f304, %f302, %f303;
mov.b32 %r741, %f304;
shfl.sync.bfly.b32 %r743|%p76, %r741, %r715, %r701, %r703;
mov.b32 %f305, %r743;
add.ftz.f32 %f42, %f304, %f305;
mov.b32 %r744, %f296;
shfl.sync.bfly.b32 %r745|%p77, %r744, %r702, %r701, %r703;
mov.b32 %f306, %r745;
add.ftz.f32 %f307, %f296, %f306;
mov.b32 %r746, %f307;
shfl.sync.bfly.b32 %r747|%p78, %r746, %r706, %r701, %r703;
mov.b32 %f308, %r747;
add.ftz.f32 %f309, %f307, %f308;
mov.b32 %r748, %f309;
shfl.sync.bfly.b32 %r749|%p79, %r748, %r709, %r701, %r703;
mov.b32 %f310, %r749;
add.ftz.f32 %f311, %f309, %f310;
mov.b32 %r750, %f311;
shfl.sync.bfly.b32 %r751|%p80, %r750, %r712, %r701, %r703;
mov.b32 %f312, %r751;
add.ftz.f32 %f313, %f311, %f312;
mov.b32 %r752, %f313;
shfl.sync.bfly.b32 %r753|%p81, %r752, %r715, %r701, %r703;
mov.b32 %f314, %r753;
add.ftz.f32 %f43, %f313, %f314;
add.s32 %r754, %r4, 1;
setp.ge.u32 %p82, %r754, %r50;
setp.ne.s32 %p83, %r2, 1;
or.pred %p84, %p83, %p82;
@%p84 bra $L__BB252_50;

ld.param.f32 %f1028, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f315, %f5, %f42;
add.ftz.f32 %f316, %f4, %f43;
min.ftz.f32 %f317, %f316, %f1028;
min.ftz.f32 %f318, %f315, %f1028;
max.ftz.f32 %f319, %f318, %f35;
mul.ftz.f32 %f320, %f317, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f321, %f320;
add.ftz.f32 %f322, %f321, 0f3F800000;
div.approx.ftz.f32 %f323, %f317, %f322;
mul.ftz.f32 %f324, %f319, %f323;
st.global.f32 [%rd34+4], %f324;

$L__BB252_50:
ld.shared.f32 %f325, [%r32+256];
add.ftz.f32 %f326, %f325, %f28;
ld.shared.f32 %f327, [%r34+256];
add.ftz.f32 %f328, %f327, %f32;
ld.shared.f32 %f329, [%r32+768];
add.ftz.f32 %f330, %f329, %f326;
ld.shared.f32 %f331, [%r34+768];
add.ftz.f32 %f332, %f331, %f328;
ld.shared.f32 %f333, [%r32+1280];
add.ftz.f32 %f334, %f333, %f330;
ld.shared.f32 %f335, [%r34+1280];
add.ftz.f32 %f336, %f335, %f332;
mov.b32 %r755, %f334;
mov.u32 %r756, 31;
mov.u32 %r757, 16;
mov.u32 %r758, -1;
shfl.sync.bfly.b32 %r759|%p85, %r755, %r757, %r756, %r758;
mov.b32 %f337, %r759;
add.ftz.f32 %f338, %f334, %f337;
mov.b32 %r760, %f338;
mov.u32 %r761, 8;
shfl.sync.bfly.b32 %r762|%p86, %r760, %r761, %r756, %r758;
mov.b32 %f339, %r762;
add.ftz.f32 %f340, %f338, %f339;
mov.b32 %r763, %f340;
mov.u32 %r764, 4;
shfl.sync.bfly.b32 %r765|%p87, %r763, %r764, %r756, %r758;
mov.b32 %f341, %r765;
add.ftz.f32 %f342, %f340, %f341;
mov.b32 %r766, %f342;
mov.u32 %r767, 2;
shfl.sync.bfly.b32 %r768|%p88, %r766, %r767, %r756, %r758;
mov.b32 %f343, %r768;
add.ftz.f32 %f344, %f342, %f343;
mov.b32 %r769, %f344;
mov.u32 %r770, 1;
shfl.sync.bfly.b32 %r771|%p89, %r769, %r770, %r756, %r758;
mov.b32 %f345, %r771;
add.ftz.f32 %f44, %f344, %f345;
mov.b32 %r772, %f336;
shfl.sync.bfly.b32 %r773|%p90, %r772, %r757, %r756, %r758;
mov.b32 %f346, %r773;
add.ftz.f32 %f347, %f336, %f346;
mov.b32 %r774, %f347;
shfl.sync.bfly.b32 %r775|%p91, %r774, %r761, %r756, %r758;
mov.b32 %f348, %r775;
add.ftz.f32 %f349, %f347, %f348;
mov.b32 %r776, %f349;
shfl.sync.bfly.b32 %r777|%p92, %r776, %r764, %r756, %r758;
mov.b32 %f350, %r777;
add.ftz.f32 %f351, %f349, %f350;
mov.b32 %r778, %f351;
shfl.sync.bfly.b32 %r779|%p93, %r778, %r767, %r756, %r758;
mov.b32 %f352, %r779;
add.ftz.f32 %f353, %f351, %f352;
mov.b32 %r780, %f353;
shfl.sync.bfly.b32 %r781|%p94, %r780, %r770, %r756, %r758;
mov.b32 %f354, %r781;
add.ftz.f32 %f45, %f353, %f354;
add.s32 %r782, %r4, 2;
setp.ge.u32 %p95, %r782, %r50;
setp.ne.s32 %p96, %r2, 2;
or.pred %p97, %p96, %p95;
@%p97 bra $L__BB252_52;

ld.param.f32 %f1027, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f355, %f5, %f44;
add.ftz.f32 %f356, %f4, %f45;
min.ftz.f32 %f357, %f356, %f1027;
min.ftz.f32 %f358, %f355, %f1027;
max.ftz.f32 %f359, %f358, %f35;
mul.ftz.f32 %f360, %f357, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f361, %f360;
add.ftz.f32 %f362, %f361, 0f3F800000;
div.approx.ftz.f32 %f363, %f357, %f362;
mul.ftz.f32 %f364, %f359, %f363;
st.global.f32 [%rd34+8], %f364;

$L__BB252_52:
ld.shared.f32 %f365, [%r32+384];
add.ftz.f32 %f366, %f365, %f27;
ld.shared.f32 %f367, [%r34+384];
add.ftz.f32 %f368, %f367, %f31;
ld.shared.f32 %f369, [%r32+896];
add.ftz.f32 %f370, %f369, %f366;
ld.shared.f32 %f371, [%r34+896];
add.ftz.f32 %f372, %f371, %f368;
ld.shared.f32 %f373, [%r32+1408];
add.ftz.f32 %f374, %f373, %f370;
ld.shared.f32 %f375, [%r34+1408];
add.ftz.f32 %f376, %f375, %f372;
mov.b32 %r783, %f374;
shfl.sync.bfly.b32 %r787|%p98, %r783, %r757, %r756, %r758;
mov.b32 %f377, %r787;
add.ftz.f32 %f378, %f374, %f377;
mov.b32 %r788, %f378;
shfl.sync.bfly.b32 %r790|%p99, %r788, %r761, %r756, %r758;
mov.b32 %f379, %r790;
add.ftz.f32 %f380, %f378, %f379;
mov.b32 %r791, %f380;
shfl.sync.bfly.b32 %r793|%p100, %r791, %r764, %r756, %r758;
mov.b32 %f381, %r793;
add.ftz.f32 %f382, %f380, %f381;
mov.b32 %r794, %f382;
shfl.sync.bfly.b32 %r796|%p101, %r794, %r767, %r756, %r758;
mov.b32 %f383, %r796;
add.ftz.f32 %f384, %f382, %f383;
mov.b32 %r797, %f384;
shfl.sync.bfly.b32 %r799|%p102, %r797, %r770, %r756, %r758;
mov.b32 %f385, %r799;
add.ftz.f32 %f46, %f384, %f385;
mov.b32 %r800, %f376;
shfl.sync.bfly.b32 %r801|%p103, %r800, %r757, %r756, %r758;
mov.b32 %f386, %r801;
add.ftz.f32 %f387, %f376, %f386;
mov.b32 %r802, %f387;
shfl.sync.bfly.b32 %r803|%p104, %r802, %r761, %r756, %r758;
mov.b32 %f388, %r803;
add.ftz.f32 %f389, %f387, %f388;
mov.b32 %r804, %f389;
shfl.sync.bfly.b32 %r805|%p105, %r804, %r764, %r756, %r758;
mov.b32 %f390, %r805;
add.ftz.f32 %f391, %f389, %f390;
mov.b32 %r806, %f391;
shfl.sync.bfly.b32 %r807|%p106, %r806, %r767, %r756, %r758;
mov.b32 %f392, %r807;
add.ftz.f32 %f393, %f391, %f392;
mov.b32 %r808, %f393;
shfl.sync.bfly.b32 %r809|%p107, %r808, %r770, %r756, %r758;
mov.b32 %f394, %r809;
add.ftz.f32 %f47, %f393, %f394;
add.s32 %r810, %r4, 3;
setp.ge.u32 %p108, %r810, %r50;
setp.ne.s32 %p109, %r2, 3;
or.pred %p110, %p109, %p108;
@%p110 bra $L__BB252_78;

ld.param.f32 %f1026, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f395, %f5, %f46;
add.ftz.f32 %f396, %f4, %f47;
min.ftz.f32 %f397, %f396, %f1026;
min.ftz.f32 %f398, %f395, %f1026;
max.ftz.f32 %f399, %f398, %f35;
mul.ftz.f32 %f400, %f397, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f401, %f400;
add.ftz.f32 %f402, %f401, 0f3F800000;
div.approx.ftz.f32 %f403, %f397, %f402;
mul.ftz.f32 %f404, %f399, %f403;
st.global.f32 [%rd34+12], %f404;
bra.uni $L__BB252_78;

$L__BB252_70:
ld.shared.f32 %f893, [%r32];
add.ftz.f32 %f894, %f893, %f30;
ld.shared.f32 %f895, [%r34];
add.ftz.f32 %f896, %f895, %f34;
ld.shared.f32 %f897, [%r32+512];
add.ftz.f32 %f898, %f897, %f894;
ld.shared.f32 %f899, [%r34+512];
add.ftz.f32 %f900, %f899, %f896;
ld.shared.f32 %f901, [%r32+1024];
add.ftz.f32 %f902, %f901, %f898;
ld.shared.f32 %f903, [%r34+1024];
add.ftz.f32 %f904, %f903, %f900;
mov.b32 %r1144, %f902;
mov.u32 %r1145, 31;
mov.u32 %r1146, 16;
mov.u32 %r1147, -1;
shfl.sync.bfly.b32 %r1148|%p267, %r1144, %r1146, %r1145, %r1147;
mov.b32 %f905, %r1148;
add.ftz.f32 %f906, %f902, %f905;
mov.b32 %r1149, %f906;
mov.u32 %r1150, 8;
shfl.sync.bfly.b32 %r1151|%p268, %r1149, %r1150, %r1145, %r1147;
mov.b32 %f907, %r1151;
add.ftz.f32 %f908, %f906, %f907;
mov.b32 %r1152, %f908;
mov.u32 %r1153, 4;
shfl.sync.bfly.b32 %r1154|%p269, %r1152, %r1153, %r1145, %r1147;
mov.b32 %f909, %r1154;
add.ftz.f32 %f910, %f908, %f909;
mov.b32 %r1155, %f910;
mov.u32 %r1156, 2;
shfl.sync.bfly.b32 %r1157|%p270, %r1155, %r1156, %r1145, %r1147;
mov.b32 %f911, %r1157;
add.ftz.f32 %f912, %f910, %f911;
mov.b32 %r1158, %f912;
mov.u32 %r1159, 1;
shfl.sync.bfly.b32 %r1160|%p271, %r1158, %r1159, %r1145, %r1147;
mov.b32 %f913, %r1160;
add.ftz.f32 %f72, %f912, %f913;
mov.b32 %r1161, %f904;
shfl.sync.bfly.b32 %r1162|%p272, %r1161, %r1146, %r1145, %r1147;
mov.b32 %f914, %r1162;
add.ftz.f32 %f915, %f904, %f914;
mov.b32 %r1163, %f915;
shfl.sync.bfly.b32 %r1164|%p273, %r1163, %r1150, %r1145, %r1147;
mov.b32 %f916, %r1164;
add.ftz.f32 %f917, %f915, %f916;
mov.b32 %r1165, %f917;
shfl.sync.bfly.b32 %r1166|%p274, %r1165, %r1153, %r1145, %r1147;
mov.b32 %f918, %r1166;
add.ftz.f32 %f919, %f917, %f918;
mov.b32 %r1167, %f919;
shfl.sync.bfly.b32 %r1168|%p275, %r1167, %r1156, %r1145, %r1147;
mov.b32 %f920, %r1168;
add.ftz.f32 %f921, %f919, %f920;
mov.b32 %r1169, %f921;
shfl.sync.bfly.b32 %r1170|%p276, %r1169, %r1159, %r1145, %r1147;
mov.b32 %f922, %r1170;
add.ftz.f32 %f73, %f921, %f922;
setp.ge.u32 %p277, %r4, %r50;
setp.ne.s32 %p278, %r2, 0;
or.pred %p279, %p278, %p277;
@%p279 bra $L__BB252_72;

add.ftz.f32 %f923, %f5, %f72;
add.ftz.f32 %f924, %f4, %f73;
mul.ftz.f32 %f925, %f923, %f924;
st.global.f32 [%rd34], %f925;

$L__BB252_72:
ld.shared.f32 %f926, [%r32+128];
add.ftz.f32 %f927, %f926, %f29;
ld.shared.f32 %f928, [%r34+128];
add.ftz.f32 %f929, %f928, %f33;
ld.shared.f32 %f930, [%r32+640];
add.ftz.f32 %f931, %f930, %f927;
ld.shared.f32 %f932, [%r34+640];
add.ftz.f32 %f933, %f932, %f929;
ld.shared.f32 %f934, [%r32+1152];
add.ftz.f32 %f935, %f934, %f931;
ld.shared.f32 %f936, [%r34+1152];
add.ftz.f32 %f937, %f936, %f933;
mov.b32 %r1171, %f935;
shfl.sync.bfly.b32 %r1175|%p280, %r1171, %r1146, %r1145, %r1147;
mov.b32 %f938, %r1175;
add.ftz.f32 %f939, %f935, %f938;
mov.b32 %r1176, %f939;
shfl.sync.bfly.b32 %r1178|%p281, %r1176, %r1150, %r1145, %r1147;
mov.b32 %f940, %r1178;
add.ftz.f32 %f941, %f939, %f940;
mov.b32 %r1179, %f941;
shfl.sync.bfly.b32 %r1181|%p282, %r1179, %r1153, %r1145, %r1147;
mov.b32 %f942, %r1181;
add.ftz.f32 %f943, %f941, %f942;
mov.b32 %r1182, %f943;
shfl.sync.bfly.b32 %r1184|%p283, %r1182, %r1156, %r1145, %r1147;
mov.b32 %f944, %r1184;
add.ftz.f32 %f945, %f943, %f944;
mov.b32 %r1185, %f945;
shfl.sync.bfly.b32 %r1187|%p284, %r1185, %r1159, %r1145, %r1147;
mov.b32 %f946, %r1187;
add.ftz.f32 %f74, %f945, %f946;
mov.b32 %r1188, %f937;
shfl.sync.bfly.b32 %r1189|%p285, %r1188, %r1146, %r1145, %r1147;
mov.b32 %f947, %r1189;
add.ftz.f32 %f948, %f937, %f947;
mov.b32 %r1190, %f948;
shfl.sync.bfly.b32 %r1191|%p286, %r1190, %r1150, %r1145, %r1147;
mov.b32 %f949, %r1191;
add.ftz.f32 %f950, %f948, %f949;
mov.b32 %r1192, %f950;
shfl.sync.bfly.b32 %r1193|%p287, %r1192, %r1153, %r1145, %r1147;
mov.b32 %f951, %r1193;
add.ftz.f32 %f952, %f950, %f951;
mov.b32 %r1194, %f952;
shfl.sync.bfly.b32 %r1195|%p288, %r1194, %r1156, %r1145, %r1147;
mov.b32 %f953, %r1195;
add.ftz.f32 %f954, %f952, %f953;
mov.b32 %r1196, %f954;
shfl.sync.bfly.b32 %r1197|%p289, %r1196, %r1159, %r1145, %r1147;
mov.b32 %f955, %r1197;
add.ftz.f32 %f75, %f954, %f955;
add.s32 %r1198, %r4, 1;
setp.ge.u32 %p290, %r1198, %r50;
setp.ne.s32 %p291, %r2, 1;
or.pred %p292, %p291, %p290;
@%p292 bra $L__BB252_74;

add.ftz.f32 %f956, %f5, %f74;
add.ftz.f32 %f957, %f4, %f75;
mul.ftz.f32 %f958, %f956, %f957;
st.global.f32 [%rd34+4], %f958;

$L__BB252_74:
ld.shared.f32 %f959, [%r32+256];
add.ftz.f32 %f960, %f959, %f28;
ld.shared.f32 %f961, [%r34+256];
add.ftz.f32 %f962, %f961, %f32;
ld.shared.f32 %f963, [%r32+768];
add.ftz.f32 %f964, %f963, %f960;
ld.shared.f32 %f965, [%r34+768];
add.ftz.f32 %f966, %f965, %f962;
ld.shared.f32 %f967, [%r32+1280];
add.ftz.f32 %f968, %f967, %f964;
ld.shared.f32 %f969, [%r34+1280];
add.ftz.f32 %f970, %f969, %f966;
mov.b32 %r1199, %f968;
mov.u32 %r1200, 31;
mov.u32 %r1201, 16;
mov.u32 %r1202, -1;
shfl.sync.bfly.b32 %r1203|%p293, %r1199, %r1201, %r1200, %r1202;
mov.b32 %f971, %r1203;
add.ftz.f32 %f972, %f968, %f971;
mov.b32 %r1204, %f972;
mov.u32 %r1205, 8;
shfl.sync.bfly.b32 %r1206|%p294, %r1204, %r1205, %r1200, %r1202;
mov.b32 %f973, %r1206;
add.ftz.f32 %f974, %f972, %f973;
mov.b32 %r1207, %f974;
mov.u32 %r1208, 4;
shfl.sync.bfly.b32 %r1209|%p295, %r1207, %r1208, %r1200, %r1202;
mov.b32 %f975, %r1209;
add.ftz.f32 %f976, %f974, %f975;
mov.b32 %r1210, %f976;
mov.u32 %r1211, 2;
shfl.sync.bfly.b32 %r1212|%p296, %r1210, %r1211, %r1200, %r1202;
mov.b32 %f977, %r1212;
add.ftz.f32 %f978, %f976, %f977;
mov.b32 %r1213, %f978;
mov.u32 %r1214, 1;
shfl.sync.bfly.b32 %r1215|%p297, %r1213, %r1214, %r1200, %r1202;
mov.b32 %f979, %r1215;
add.ftz.f32 %f76, %f978, %f979;
mov.b32 %r1216, %f970;
shfl.sync.bfly.b32 %r1217|%p298, %r1216, %r1201, %r1200, %r1202;
mov.b32 %f980, %r1217;
add.ftz.f32 %f981, %f970, %f980;
mov.b32 %r1218, %f981;
shfl.sync.bfly.b32 %r1219|%p299, %r1218, %r1205, %r1200, %r1202;
mov.b32 %f982, %r1219;
add.ftz.f32 %f983, %f981, %f982;
mov.b32 %r1220, %f983;
shfl.sync.bfly.b32 %r1221|%p300, %r1220, %r1208, %r1200, %r1202;
mov.b32 %f984, %r1221;
add.ftz.f32 %f985, %f983, %f984;
mov.b32 %r1222, %f985;
shfl.sync.bfly.b32 %r1223|%p301, %r1222, %r1211, %r1200, %r1202;
mov.b32 %f986, %r1223;
add.ftz.f32 %f987, %f985, %f986;
mov.b32 %r1224, %f987;
shfl.sync.bfly.b32 %r1225|%p302, %r1224, %r1214, %r1200, %r1202;
mov.b32 %f988, %r1225;
add.ftz.f32 %f77, %f987, %f988;
add.s32 %r1226, %r4, 2;
setp.ge.u32 %p303, %r1226, %r50;
setp.ne.s32 %p304, %r2, 2;
or.pred %p305, %p304, %p303;
@%p305 bra $L__BB252_76;

add.ftz.f32 %f989, %f5, %f76;
add.ftz.f32 %f990, %f4, %f77;
mul.ftz.f32 %f991, %f989, %f990;
st.global.f32 [%rd34+8], %f991;

$L__BB252_76:
ld.shared.f32 %f992, [%r32+384];
add.ftz.f32 %f993, %f992, %f27;
ld.shared.f32 %f994, [%r34+384];
add.ftz.f32 %f995, %f994, %f31;
ld.shared.f32 %f996, [%r32+896];
add.ftz.f32 %f997, %f996, %f993;
ld.shared.f32 %f998, [%r34+896];
add.ftz.f32 %f999, %f998, %f995;
ld.shared.f32 %f1000, [%r32+1408];
add.ftz.f32 %f1001, %f1000, %f997;
ld.shared.f32 %f1002, [%r34+1408];
add.ftz.f32 %f1003, %f1002, %f999;
mov.b32 %r1227, %f1001;
shfl.sync.bfly.b32 %r1231|%p306, %r1227, %r1201, %r1200, %r1202;
mov.b32 %f1004, %r1231;
add.ftz.f32 %f1005, %f1001, %f1004;
mov.b32 %r1232, %f1005;
shfl.sync.bfly.b32 %r1234|%p307, %r1232, %r1205, %r1200, %r1202;
mov.b32 %f1006, %r1234;
add.ftz.f32 %f1007, %f1005, %f1006;
mov.b32 %r1235, %f1007;
shfl.sync.bfly.b32 %r1237|%p308, %r1235, %r1208, %r1200, %r1202;
mov.b32 %f1008, %r1237;
add.ftz.f32 %f1009, %f1007, %f1008;
mov.b32 %r1238, %f1009;
shfl.sync.bfly.b32 %r1240|%p309, %r1238, %r1211, %r1200, %r1202;
mov.b32 %f1010, %r1240;
add.ftz.f32 %f1011, %f1009, %f1010;
mov.b32 %r1241, %f1011;
shfl.sync.bfly.b32 %r1243|%p310, %r1241, %r1214, %r1200, %r1202;
mov.b32 %f1012, %r1243;
add.ftz.f32 %f78, %f1011, %f1012;
mov.b32 %r1244, %f1003;
shfl.sync.bfly.b32 %r1245|%p311, %r1244, %r1201, %r1200, %r1202;
mov.b32 %f1013, %r1245;
add.ftz.f32 %f1014, %f1003, %f1013;
mov.b32 %r1246, %f1014;
shfl.sync.bfly.b32 %r1247|%p312, %r1246, %r1205, %r1200, %r1202;
mov.b32 %f1015, %r1247;
add.ftz.f32 %f1016, %f1014, %f1015;
mov.b32 %r1248, %f1016;
shfl.sync.bfly.b32 %r1249|%p313, %r1248, %r1208, %r1200, %r1202;
mov.b32 %f1017, %r1249;
add.ftz.f32 %f1018, %f1016, %f1017;
mov.b32 %r1250, %f1018;
shfl.sync.bfly.b32 %r1251|%p314, %r1250, %r1211, %r1200, %r1202;
mov.b32 %f1019, %r1251;
add.ftz.f32 %f1020, %f1018, %f1019;
mov.b32 %r1252, %f1020;
shfl.sync.bfly.b32 %r1253|%p315, %r1252, %r1214, %r1200, %r1202;
mov.b32 %f1021, %r1253;
add.ftz.f32 %f79, %f1020, %f1021;
add.s32 %r1254, %r4, 3;
setp.ge.u32 %p316, %r1254, %r50;
setp.ne.s32 %p317, %r2, 3;
or.pred %p318, %p317, %p316;
@%p318 bra $L__BB252_78;

add.ftz.f32 %f1022, %f5, %f78;
add.ftz.f32 %f1023, %f4, %f79;
mul.ftz.f32 %f1024, %f1022, %f1023;
st.global.f32 [%rd34+12], %f1024;
bra.uni $L__BB252_78;

$L__BB252_62:
ld.shared.f32 %f729, [%r32];
add.ftz.f32 %f730, %f729, %f30;
ld.shared.f32 %f731, [%r34];
add.ftz.f32 %f732, %f731, %f34;
ld.shared.f32 %f733, [%r32+512];
add.ftz.f32 %f734, %f733, %f730;
ld.shared.f32 %f735, [%r34+512];
add.ftz.f32 %f736, %f735, %f732;
ld.shared.f32 %f737, [%r32+1024];
add.ftz.f32 %f738, %f737, %f734;
ld.shared.f32 %f739, [%r34+1024];
add.ftz.f32 %f740, %f739, %f736;
mov.b32 %r1033, %f738;
mov.u32 %r1034, 31;
mov.u32 %r1035, 16;
mov.u32 %r1036, -1;
shfl.sync.bfly.b32 %r1037|%p215, %r1033, %r1035, %r1034, %r1036;
mov.b32 %f741, %r1037;
add.ftz.f32 %f742, %f738, %f741;
mov.b32 %r1038, %f742;
mov.u32 %r1039, 8;
shfl.sync.bfly.b32 %r1040|%p216, %r1038, %r1039, %r1034, %r1036;
mov.b32 %f743, %r1040;
add.ftz.f32 %f744, %f742, %f743;
mov.b32 %r1041, %f744;
mov.u32 %r1042, 4;
shfl.sync.bfly.b32 %r1043|%p217, %r1041, %r1042, %r1034, %r1036;
mov.b32 %f745, %r1043;
add.ftz.f32 %f746, %f744, %f745;
mov.b32 %r1044, %f746;
mov.u32 %r1045, 2;
shfl.sync.bfly.b32 %r1046|%p218, %r1044, %r1045, %r1034, %r1036;
mov.b32 %f747, %r1046;
add.ftz.f32 %f748, %f746, %f747;
mov.b32 %r1047, %f748;
mov.u32 %r1048, 1;
shfl.sync.bfly.b32 %r1049|%p219, %r1047, %r1048, %r1034, %r1036;
mov.b32 %f749, %r1049;
add.ftz.f32 %f64, %f748, %f749;
mov.b32 %r1050, %f740;
shfl.sync.bfly.b32 %r1051|%p220, %r1050, %r1035, %r1034, %r1036;
mov.b32 %f750, %r1051;
add.ftz.f32 %f751, %f740, %f750;
mov.b32 %r1052, %f751;
shfl.sync.bfly.b32 %r1053|%p221, %r1052, %r1039, %r1034, %r1036;
mov.b32 %f752, %r1053;
add.ftz.f32 %f753, %f751, %f752;
mov.b32 %r1054, %f753;
shfl.sync.bfly.b32 %r1055|%p222, %r1054, %r1042, %r1034, %r1036;
mov.b32 %f754, %r1055;
add.ftz.f32 %f755, %f753, %f754;
mov.b32 %r1056, %f755;
shfl.sync.bfly.b32 %r1057|%p223, %r1056, %r1045, %r1034, %r1036;
mov.b32 %f756, %r1057;
add.ftz.f32 %f757, %f755, %f756;
mov.b32 %r1058, %f757;
shfl.sync.bfly.b32 %r1059|%p224, %r1058, %r1048, %r1034, %r1036;
mov.b32 %f758, %r1059;
add.ftz.f32 %f65, %f757, %f758;
setp.ge.u32 %p225, %r4, %r50;
setp.ne.s32 %p226, %r2, 0;
or.pred %p227, %p226, %p225;
@%p227 bra $L__BB252_64;

add.ftz.f32 %f759, %f5, %f64;
add.ftz.f32 %f760, %f4, %f65;
mul.ftz.f32 %f761, %f760, 0f3F000000;
mul.ftz.f32 %f762, %f760, 0f3F4C422A;
mul.ftz.f32 %f763, %f760, 0f3D372713;
fma.rn.ftz.f32 %f764, %f760, %f763, 0f3F800000;
mul.ftz.f32 %f765, %f762, %f764;
tanh.approx.f32 %f766, %f765;
add.ftz.f32 %f767, %f766, 0f3F800000;
mul.ftz.f32 %f768, %f761, %f767;
mul.ftz.f32 %f769, %f759, %f768;
st.global.f32 [%rd34], %f769;

$L__BB252_64:
ld.shared.f32 %f770, [%r32+128];
add.ftz.f32 %f771, %f770, %f29;
ld.shared.f32 %f772, [%r34+128];
add.ftz.f32 %f773, %f772, %f33;
ld.shared.f32 %f774, [%r32+640];
add.ftz.f32 %f775, %f774, %f771;
ld.shared.f32 %f776, [%r34+640];
add.ftz.f32 %f777, %f776, %f773;
ld.shared.f32 %f778, [%r32+1152];
add.ftz.f32 %f779, %f778, %f775;
ld.shared.f32 %f780, [%r34+1152];
add.ftz.f32 %f781, %f780, %f777;
mov.b32 %r1060, %f779;
shfl.sync.bfly.b32 %r1064|%p228, %r1060, %r1035, %r1034, %r1036;
mov.b32 %f782, %r1064;
add.ftz.f32 %f783, %f779, %f782;
mov.b32 %r1065, %f783;
shfl.sync.bfly.b32 %r1067|%p229, %r1065, %r1039, %r1034, %r1036;
mov.b32 %f784, %r1067;
add.ftz.f32 %f785, %f783, %f784;
mov.b32 %r1068, %f785;
shfl.sync.bfly.b32 %r1070|%p230, %r1068, %r1042, %r1034, %r1036;
mov.b32 %f786, %r1070;
add.ftz.f32 %f787, %f785, %f786;
mov.b32 %r1071, %f787;
shfl.sync.bfly.b32 %r1073|%p231, %r1071, %r1045, %r1034, %r1036;
mov.b32 %f788, %r1073;
add.ftz.f32 %f789, %f787, %f788;
mov.b32 %r1074, %f789;
shfl.sync.bfly.b32 %r1076|%p232, %r1074, %r1048, %r1034, %r1036;
mov.b32 %f790, %r1076;
add.ftz.f32 %f66, %f789, %f790;
mov.b32 %r1077, %f781;
shfl.sync.bfly.b32 %r1078|%p233, %r1077, %r1035, %r1034, %r1036;
mov.b32 %f791, %r1078;
add.ftz.f32 %f792, %f781, %f791;
mov.b32 %r1079, %f792;
shfl.sync.bfly.b32 %r1080|%p234, %r1079, %r1039, %r1034, %r1036;
mov.b32 %f793, %r1080;
add.ftz.f32 %f794, %f792, %f793;
mov.b32 %r1081, %f794;
shfl.sync.bfly.b32 %r1082|%p235, %r1081, %r1042, %r1034, %r1036;
mov.b32 %f795, %r1082;
add.ftz.f32 %f796, %f794, %f795;
mov.b32 %r1083, %f796;
shfl.sync.bfly.b32 %r1084|%p236, %r1083, %r1045, %r1034, %r1036;
mov.b32 %f797, %r1084;
add.ftz.f32 %f798, %f796, %f797;
mov.b32 %r1085, %f798;
shfl.sync.bfly.b32 %r1086|%p237, %r1085, %r1048, %r1034, %r1036;
mov.b32 %f799, %r1086;
add.ftz.f32 %f67, %f798, %f799;
add.s32 %r1087, %r4, 1;
setp.ge.u32 %p238, %r1087, %r50;
setp.ne.s32 %p239, %r2, 1;
or.pred %p240, %p239, %p238;
@%p240 bra $L__BB252_66;

add.ftz.f32 %f800, %f5, %f66;
add.ftz.f32 %f801, %f4, %f67;
mul.ftz.f32 %f802, %f801, 0f3F000000;
mul.ftz.f32 %f803, %f801, 0f3F4C422A;
mul.ftz.f32 %f804, %f801, 0f3D372713;
fma.rn.ftz.f32 %f805, %f801, %f804, 0f3F800000;
mul.ftz.f32 %f806, %f803, %f805;
tanh.approx.f32 %f807, %f806;
add.ftz.f32 %f808, %f807, 0f3F800000;
mul.ftz.f32 %f809, %f802, %f808;
mul.ftz.f32 %f810, %f800, %f809;
st.global.f32 [%rd34+4], %f810;

$L__BB252_66:
ld.shared.f32 %f811, [%r32+256];
add.ftz.f32 %f812, %f811, %f28;
ld.shared.f32 %f813, [%r34+256];
add.ftz.f32 %f814, %f813, %f32;
ld.shared.f32 %f815, [%r32+768];
add.ftz.f32 %f816, %f815, %f812;
ld.shared.f32 %f817, [%r34+768];
add.ftz.f32 %f818, %f817, %f814;
ld.shared.f32 %f819, [%r32+1280];
add.ftz.f32 %f820, %f819, %f816;
ld.shared.f32 %f821, [%r34+1280];
add.ftz.f32 %f822, %f821, %f818;
mov.b32 %r1088, %f820;
mov.u32 %r1089, 31;
mov.u32 %r1090, 16;
mov.u32 %r1091, -1;
shfl.sync.bfly.b32 %r1092|%p241, %r1088, %r1090, %r1089, %r1091;
mov.b32 %f823, %r1092;
add.ftz.f32 %f824, %f820, %f823;
mov.b32 %r1093, %f824;
mov.u32 %r1094, 8;
shfl.sync.bfly.b32 %r1095|%p242, %r1093, %r1094, %r1089, %r1091;
mov.b32 %f825, %r1095;
add.ftz.f32 %f826, %f824, %f825;
mov.b32 %r1096, %f826;
mov.u32 %r1097, 4;
shfl.sync.bfly.b32 %r1098|%p243, %r1096, %r1097, %r1089, %r1091;
mov.b32 %f827, %r1098;
add.ftz.f32 %f828, %f826, %f827;
mov.b32 %r1099, %f828;
mov.u32 %r1100, 2;
shfl.sync.bfly.b32 %r1101|%p244, %r1099, %r1100, %r1089, %r1091;
mov.b32 %f829, %r1101;
add.ftz.f32 %f830, %f828, %f829;
mov.b32 %r1102, %f830;
mov.u32 %r1103, 1;
shfl.sync.bfly.b32 %r1104|%p245, %r1102, %r1103, %r1089, %r1091;
mov.b32 %f831, %r1104;
add.ftz.f32 %f68, %f830, %f831;
mov.b32 %r1105, %f822;
shfl.sync.bfly.b32 %r1106|%p246, %r1105, %r1090, %r1089, %r1091;
mov.b32 %f832, %r1106;
add.ftz.f32 %f833, %f822, %f832;
mov.b32 %r1107, %f833;
shfl.sync.bfly.b32 %r1108|%p247, %r1107, %r1094, %r1089, %r1091;
mov.b32 %f834, %r1108;
add.ftz.f32 %f835, %f833, %f834;
mov.b32 %r1109, %f835;
shfl.sync.bfly.b32 %r1110|%p248, %r1109, %r1097, %r1089, %r1091;
mov.b32 %f836, %r1110;
add.ftz.f32 %f837, %f835, %f836;
mov.b32 %r1111, %f837;
shfl.sync.bfly.b32 %r1112|%p249, %r1111, %r1100, %r1089, %r1091;
mov.b32 %f838, %r1112;
add.ftz.f32 %f839, %f837, %f838;
mov.b32 %r1113, %f839;
shfl.sync.bfly.b32 %r1114|%p250, %r1113, %r1103, %r1089, %r1091;
mov.b32 %f840, %r1114;
add.ftz.f32 %f69, %f839, %f840;
add.s32 %r1115, %r4, 2;
setp.ge.u32 %p251, %r1115, %r50;
setp.ne.s32 %p252, %r2, 2;
or.pred %p253, %p252, %p251;
@%p253 bra $L__BB252_68;

add.ftz.f32 %f841, %f5, %f68;
add.ftz.f32 %f842, %f4, %f69;
mul.ftz.f32 %f843, %f842, 0f3F000000;
mul.ftz.f32 %f844, %f842, 0f3F4C422A;
mul.ftz.f32 %f845, %f842, 0f3D372713;
fma.rn.ftz.f32 %f846, %f842, %f845, 0f3F800000;
mul.ftz.f32 %f847, %f844, %f846;
tanh.approx.f32 %f848, %f847;
add.ftz.f32 %f849, %f848, 0f3F800000;
mul.ftz.f32 %f850, %f843, %f849;
mul.ftz.f32 %f851, %f841, %f850;
st.global.f32 [%rd34+8], %f851;

$L__BB252_68:
ld.shared.f32 %f852, [%r32+384];
add.ftz.f32 %f853, %f852, %f27;
ld.shared.f32 %f854, [%r34+384];
add.ftz.f32 %f855, %f854, %f31;
ld.shared.f32 %f856, [%r32+896];
add.ftz.f32 %f857, %f856, %f853;
ld.shared.f32 %f858, [%r34+896];
add.ftz.f32 %f859, %f858, %f855;
ld.shared.f32 %f860, [%r32+1408];
add.ftz.f32 %f861, %f860, %f857;
ld.shared.f32 %f862, [%r34+1408];
add.ftz.f32 %f863, %f862, %f859;
mov.b32 %r1116, %f861;
shfl.sync.bfly.b32 %r1120|%p254, %r1116, %r1090, %r1089, %r1091;
mov.b32 %f864, %r1120;
add.ftz.f32 %f865, %f861, %f864;
mov.b32 %r1121, %f865;
shfl.sync.bfly.b32 %r1123|%p255, %r1121, %r1094, %r1089, %r1091;
mov.b32 %f866, %r1123;
add.ftz.f32 %f867, %f865, %f866;
mov.b32 %r1124, %f867;
shfl.sync.bfly.b32 %r1126|%p256, %r1124, %r1097, %r1089, %r1091;
mov.b32 %f868, %r1126;
add.ftz.f32 %f869, %f867, %f868;
mov.b32 %r1127, %f869;
shfl.sync.bfly.b32 %r1129|%p257, %r1127, %r1100, %r1089, %r1091;
mov.b32 %f870, %r1129;
add.ftz.f32 %f871, %f869, %f870;
mov.b32 %r1130, %f871;
shfl.sync.bfly.b32 %r1132|%p258, %r1130, %r1103, %r1089, %r1091;
mov.b32 %f872, %r1132;
add.ftz.f32 %f70, %f871, %f872;
mov.b32 %r1133, %f863;
shfl.sync.bfly.b32 %r1134|%p259, %r1133, %r1090, %r1089, %r1091;
mov.b32 %f873, %r1134;
add.ftz.f32 %f874, %f863, %f873;
mov.b32 %r1135, %f874;
shfl.sync.bfly.b32 %r1136|%p260, %r1135, %r1094, %r1089, %r1091;
mov.b32 %f875, %r1136;
add.ftz.f32 %f876, %f874, %f875;
mov.b32 %r1137, %f876;
shfl.sync.bfly.b32 %r1138|%p261, %r1137, %r1097, %r1089, %r1091;
mov.b32 %f877, %r1138;
add.ftz.f32 %f878, %f876, %f877;
mov.b32 %r1139, %f878;
shfl.sync.bfly.b32 %r1140|%p262, %r1139, %r1100, %r1089, %r1091;
mov.b32 %f879, %r1140;
add.ftz.f32 %f880, %f878, %f879;
mov.b32 %r1141, %f880;
shfl.sync.bfly.b32 %r1142|%p263, %r1141, %r1103, %r1089, %r1091;
mov.b32 %f881, %r1142;
add.ftz.f32 %f71, %f880, %f881;
add.s32 %r1143, %r4, 3;
setp.ge.u32 %p264, %r1143, %r50;
setp.ne.s32 %p265, %r2, 3;
or.pred %p266, %p265, %p264;
@%p266 bra $L__BB252_78;

add.ftz.f32 %f882, %f5, %f70;
add.ftz.f32 %f883, %f4, %f71;
mul.ftz.f32 %f884, %f883, 0f3F000000;
mul.ftz.f32 %f885, %f883, 0f3F4C422A;
mul.ftz.f32 %f886, %f883, 0f3D372713;
fma.rn.ftz.f32 %f887, %f883, %f886, 0f3F800000;
mul.ftz.f32 %f888, %f885, %f887;
tanh.approx.f32 %f889, %f888;
add.ftz.f32 %f890, %f889, 0f3F800000;
mul.ftz.f32 %f891, %f884, %f890;
mul.ftz.f32 %f892, %f882, %f891;
st.global.f32 [%rd34+12], %f892;
bra.uni $L__BB252_78;

$L__BB252_54:
ld.shared.f32 %f405, [%r32];
add.ftz.f32 %f406, %f405, %f30;
ld.shared.f32 %f407, [%r34];
add.ftz.f32 %f408, %f407, %f34;
ld.shared.f32 %f409, [%r32+512];
add.ftz.f32 %f410, %f409, %f406;
ld.shared.f32 %f411, [%r34+512];
add.ftz.f32 %f412, %f411, %f408;
ld.shared.f32 %f413, [%r32+1024];
add.ftz.f32 %f414, %f413, %f410;
ld.shared.f32 %f415, [%r34+1024];
add.ftz.f32 %f416, %f415, %f412;
mov.b32 %r811, %f414;
mov.u32 %r812, 31;
mov.u32 %r813, 16;
mov.u32 %r814, -1;
shfl.sync.bfly.b32 %r815|%p111, %r811, %r813, %r812, %r814;
mov.b32 %f417, %r815;
add.ftz.f32 %f418, %f414, %f417;
mov.b32 %r816, %f418;
mov.u32 %r817, 8;
shfl.sync.bfly.b32 %r818|%p112, %r816, %r817, %r812, %r814;
mov.b32 %f419, %r818;
add.ftz.f32 %f420, %f418, %f419;
mov.b32 %r819, %f420;
mov.u32 %r820, 4;
shfl.sync.bfly.b32 %r821|%p113, %r819, %r820, %r812, %r814;
mov.b32 %f421, %r821;
add.ftz.f32 %f422, %f420, %f421;
mov.b32 %r822, %f422;
mov.u32 %r823, 2;
shfl.sync.bfly.b32 %r824|%p114, %r822, %r823, %r812, %r814;
mov.b32 %f423, %r824;
add.ftz.f32 %f424, %f422, %f423;
mov.b32 %r825, %f424;
mov.u32 %r826, 1;
shfl.sync.bfly.b32 %r827|%p115, %r825, %r826, %r812, %r814;
mov.b32 %f425, %r827;
add.ftz.f32 %f48, %f424, %f425;
mov.b32 %r828, %f416;
shfl.sync.bfly.b32 %r829|%p116, %r828, %r813, %r812, %r814;
mov.b32 %f426, %r829;
add.ftz.f32 %f427, %f416, %f426;
mov.b32 %r830, %f427;
shfl.sync.bfly.b32 %r831|%p117, %r830, %r817, %r812, %r814;
mov.b32 %f428, %r831;
add.ftz.f32 %f429, %f427, %f428;
mov.b32 %r832, %f429;
shfl.sync.bfly.b32 %r833|%p118, %r832, %r820, %r812, %r814;
mov.b32 %f430, %r833;
add.ftz.f32 %f431, %f429, %f430;
mov.b32 %r834, %f431;
shfl.sync.bfly.b32 %r835|%p119, %r834, %r823, %r812, %r814;
mov.b32 %f432, %r835;
add.ftz.f32 %f433, %f431, %f432;
mov.b32 %r836, %f433;
shfl.sync.bfly.b32 %r837|%p120, %r836, %r826, %r812, %r814;
mov.b32 %f434, %r837;
add.ftz.f32 %f49, %f433, %f434;
setp.ge.u32 %p121, %r4, %r50;
setp.ne.s32 %p122, %r2, 0;
or.pred %p123, %p122, %p121;
@%p123 bra $L__BB252_56;

add.ftz.f32 %f435, %f5, %f48;
add.ftz.f32 %f436, %f4, %f49;
mov.f32 %f437, 0f40E00000;
min.ftz.f32 %f438, %f436, %f437;
min.ftz.f32 %f439, %f435, %f437;
mov.f32 %f440, 0fC0E00000;
max.ftz.f32 %f441, %f439, %f440;
mul.ftz.f32 %f442, %f438, 0f3FD9DB23;
mul.ftz.f32 %f443, %f442, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f444, %f443;
add.ftz.f32 %f445, %f444, 0f3F800000;
div.approx.ftz.f32 %f446, %f438, %f445;
add.ftz.f32 %f447, %f441, 0f3F800000;
mul.ftz.f32 %f448, %f447, %f446;
st.global.f32 [%rd34], %f448;

$L__BB252_56:
ld.shared.f32 %f449, [%r32+128];
add.ftz.f32 %f450, %f449, %f29;
ld.shared.f32 %f451, [%r34+128];
add.ftz.f32 %f452, %f451, %f33;
ld.shared.f32 %f453, [%r32+640];
add.ftz.f32 %f454, %f453, %f450;
ld.shared.f32 %f455, [%r34+640];
add.ftz.f32 %f456, %f455, %f452;
ld.shared.f32 %f457, [%r32+1152];
add.ftz.f32 %f458, %f457, %f454;
ld.shared.f32 %f459, [%r34+1152];
add.ftz.f32 %f460, %f459, %f456;
mov.b32 %r838, %f458;
shfl.sync.bfly.b32 %r842|%p124, %r838, %r813, %r812, %r814;
mov.b32 %f461, %r842;
add.ftz.f32 %f462, %f458, %f461;
mov.b32 %r843, %f462;
shfl.sync.bfly.b32 %r845|%p125, %r843, %r817, %r812, %r814;
mov.b32 %f463, %r845;
add.ftz.f32 %f464, %f462, %f463;
mov.b32 %r846, %f464;
shfl.sync.bfly.b32 %r848|%p126, %r846, %r820, %r812, %r814;
mov.b32 %f465, %r848;
add.ftz.f32 %f466, %f464, %f465;
mov.b32 %r849, %f466;
shfl.sync.bfly.b32 %r851|%p127, %r849, %r823, %r812, %r814;
mov.b32 %f467, %r851;
add.ftz.f32 %f468, %f466, %f467;
mov.b32 %r852, %f468;
shfl.sync.bfly.b32 %r854|%p128, %r852, %r826, %r812, %r814;
mov.b32 %f469, %r854;
add.ftz.f32 %f50, %f468, %f469;
mov.b32 %r855, %f460;
shfl.sync.bfly.b32 %r856|%p129, %r855, %r813, %r812, %r814;
mov.b32 %f470, %r856;
add.ftz.f32 %f471, %f460, %f470;
mov.b32 %r857, %f471;
shfl.sync.bfly.b32 %r858|%p130, %r857, %r817, %r812, %r814;
mov.b32 %f472, %r858;
add.ftz.f32 %f473, %f471, %f472;
mov.b32 %r859, %f473;
shfl.sync.bfly.b32 %r860|%p131, %r859, %r820, %r812, %r814;
mov.b32 %f474, %r860;
add.ftz.f32 %f475, %f473, %f474;
mov.b32 %r861, %f475;
shfl.sync.bfly.b32 %r862|%p132, %r861, %r823, %r812, %r814;
mov.b32 %f476, %r862;
add.ftz.f32 %f477, %f475, %f476;
mov.b32 %r863, %f477;
shfl.sync.bfly.b32 %r864|%p133, %r863, %r826, %r812, %r814;
mov.b32 %f478, %r864;
add.ftz.f32 %f51, %f477, %f478;
add.s32 %r865, %r4, 1;
setp.ge.u32 %p134, %r865, %r50;
setp.ne.s32 %p135, %r2, 1;
or.pred %p136, %p135, %p134;
@%p136 bra $L__BB252_58;

add.ftz.f32 %f479, %f5, %f50;
add.ftz.f32 %f480, %f4, %f51;
mov.f32 %f481, 0f40E00000;
min.ftz.f32 %f482, %f480, %f481;
min.ftz.f32 %f483, %f479, %f481;
mov.f32 %f484, 0fC0E00000;
max.ftz.f32 %f485, %f483, %f484;
mul.ftz.f32 %f486, %f482, 0f3FD9DB23;
mul.ftz.f32 %f487, %f486, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f488, %f487;
add.ftz.f32 %f489, %f488, 0f3F800000;
div.approx.ftz.f32 %f490, %f482, %f489;
add.ftz.f32 %f491, %f485, 0f3F800000;
mul.ftz.f32 %f492, %f491, %f490;
st.global.f32 [%rd34+4], %f492;

$L__BB252_58:
ld.shared.f32 %f493, [%r32+256];
add.ftz.f32 %f494, %f493, %f28;
ld.shared.f32 %f495, [%r34+256];
add.ftz.f32 %f496, %f495, %f32;
ld.shared.f32 %f497, [%r32+768];
add.ftz.f32 %f498, %f497, %f494;
ld.shared.f32 %f499, [%r34+768];
add.ftz.f32 %f500, %f499, %f496;
ld.shared.f32 %f501, [%r32+1280];
add.ftz.f32 %f502, %f501, %f498;
ld.shared.f32 %f503, [%r34+1280];
add.ftz.f32 %f504, %f503, %f500;
mov.b32 %r866, %f502;
mov.u32 %r867, 31;
mov.u32 %r868, 16;
mov.u32 %r869, -1;
shfl.sync.bfly.b32 %r870|%p137, %r866, %r868, %r867, %r869;
mov.b32 %f505, %r870;
add.ftz.f32 %f506, %f502, %f505;
mov.b32 %r871, %f506;
mov.u32 %r872, 8;
shfl.sync.bfly.b32 %r873|%p138, %r871, %r872, %r867, %r869;
mov.b32 %f507, %r873;
add.ftz.f32 %f508, %f506, %f507;
mov.b32 %r874, %f508;
mov.u32 %r875, 4;
shfl.sync.bfly.b32 %r876|%p139, %r874, %r875, %r867, %r869;
mov.b32 %f509, %r876;
add.ftz.f32 %f510, %f508, %f509;
mov.b32 %r877, %f510;
mov.u32 %r878, 2;
shfl.sync.bfly.b32 %r879|%p140, %r877, %r878, %r867, %r869;
mov.b32 %f511, %r879;
add.ftz.f32 %f512, %f510, %f511;
mov.b32 %r880, %f512;
mov.u32 %r881, 1;
shfl.sync.bfly.b32 %r882|%p141, %r880, %r881, %r867, %r869;
mov.b32 %f513, %r882;
add.ftz.f32 %f52, %f512, %f513;
mov.b32 %r883, %f504;
shfl.sync.bfly.b32 %r884|%p142, %r883, %r868, %r867, %r869;
mov.b32 %f514, %r884;
add.ftz.f32 %f515, %f504, %f514;
mov.b32 %r885, %f515;
shfl.sync.bfly.b32 %r886|%p143, %r885, %r872, %r867, %r869;
mov.b32 %f516, %r886;
add.ftz.f32 %f517, %f515, %f516;
mov.b32 %r887, %f517;
shfl.sync.bfly.b32 %r888|%p144, %r887, %r875, %r867, %r869;
mov.b32 %f518, %r888;
add.ftz.f32 %f519, %f517, %f518;
mov.b32 %r889, %f519;
shfl.sync.bfly.b32 %r890|%p145, %r889, %r878, %r867, %r869;
mov.b32 %f520, %r890;
add.ftz.f32 %f521, %f519, %f520;
mov.b32 %r891, %f521;
shfl.sync.bfly.b32 %r892|%p146, %r891, %r881, %r867, %r869;
mov.b32 %f522, %r892;
add.ftz.f32 %f53, %f521, %f522;
add.s32 %r893, %r4, 2;
setp.ge.u32 %p147, %r893, %r50;
setp.ne.s32 %p148, %r2, 2;
or.pred %p149, %p148, %p147;
@%p149 bra $L__BB252_60;

add.ftz.f32 %f523, %f5, %f52;
add.ftz.f32 %f524, %f4, %f53;
mov.f32 %f525, 0f40E00000;
min.ftz.f32 %f526, %f524, %f525;
min.ftz.f32 %f527, %f523, %f525;
mov.f32 %f528, 0fC0E00000;
max.ftz.f32 %f529, %f527, %f528;
mul.ftz.f32 %f530, %f526, 0f3FD9DB23;
mul.ftz.f32 %f531, %f530, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f532, %f531;
add.ftz.f32 %f533, %f532, 0f3F800000;
div.approx.ftz.f32 %f534, %f526, %f533;
add.ftz.f32 %f535, %f529, 0f3F800000;
mul.ftz.f32 %f536, %f535, %f534;
st.global.f32 [%rd34+8], %f536;

$L__BB252_60:
ld.shared.f32 %f537, [%r32+384];
add.ftz.f32 %f538, %f537, %f27;
ld.shared.f32 %f539, [%r34+384];
add.ftz.f32 %f540, %f539, %f31;
ld.shared.f32 %f541, [%r32+896];
add.ftz.f32 %f542, %f541, %f538;
ld.shared.f32 %f543, [%r34+896];
add.ftz.f32 %f544, %f543, %f540;
ld.shared.f32 %f545, [%r32+1408];
add.ftz.f32 %f546, %f545, %f542;
ld.shared.f32 %f547, [%r34+1408];
add.ftz.f32 %f548, %f547, %f544;
mov.b32 %r894, %f546;
shfl.sync.bfly.b32 %r898|%p150, %r894, %r868, %r867, %r869;
mov.b32 %f549, %r898;
add.ftz.f32 %f550, %f546, %f549;
mov.b32 %r899, %f550;
shfl.sync.bfly.b32 %r901|%p151, %r899, %r872, %r867, %r869;
mov.b32 %f551, %r901;
add.ftz.f32 %f552, %f550, %f551;
mov.b32 %r902, %f552;
shfl.sync.bfly.b32 %r904|%p152, %r902, %r875, %r867, %r869;
mov.b32 %f553, %r904;
add.ftz.f32 %f554, %f552, %f553;
mov.b32 %r905, %f554;
shfl.sync.bfly.b32 %r907|%p153, %r905, %r878, %r867, %r869;
mov.b32 %f555, %r907;
add.ftz.f32 %f556, %f554, %f555;
mov.b32 %r908, %f556;
shfl.sync.bfly.b32 %r910|%p154, %r908, %r881, %r867, %r869;
mov.b32 %f557, %r910;
add.ftz.f32 %f54, %f556, %f557;
mov.b32 %r911, %f548;
shfl.sync.bfly.b32 %r912|%p155, %r911, %r868, %r867, %r869;
mov.b32 %f558, %r912;
add.ftz.f32 %f559, %f548, %f558;
mov.b32 %r913, %f559;
shfl.sync.bfly.b32 %r914|%p156, %r913, %r872, %r867, %r869;
mov.b32 %f560, %r914;
add.ftz.f32 %f561, %f559, %f560;
mov.b32 %r915, %f561;
shfl.sync.bfly.b32 %r916|%p157, %r915, %r875, %r867, %r869;
mov.b32 %f562, %r916;
add.ftz.f32 %f563, %f561, %f562;
mov.b32 %r917, %f563;
shfl.sync.bfly.b32 %r918|%p158, %r917, %r878, %r867, %r869;
mov.b32 %f564, %r918;
add.ftz.f32 %f565, %f563, %f564;
mov.b32 %r919, %f565;
shfl.sync.bfly.b32 %r920|%p159, %r919, %r881, %r867, %r869;
mov.b32 %f566, %r920;
add.ftz.f32 %f55, %f565, %f566;
add.s32 %r921, %r4, 3;
setp.ge.u32 %p160, %r921, %r50;
setp.ne.s32 %p161, %r2, 3;
or.pred %p162, %p161, %p160;
@%p162 bra $L__BB252_78;

add.ftz.f32 %f567, %f5, %f54;
add.ftz.f32 %f568, %f4, %f55;
mov.f32 %f569, 0f40E00000;
min.ftz.f32 %f570, %f568, %f569;
min.ftz.f32 %f571, %f567, %f569;
mov.f32 %f572, 0fC0E00000;
max.ftz.f32 %f573, %f571, %f572;
mul.ftz.f32 %f574, %f570, 0f3FD9DB23;
mul.ftz.f32 %f575, %f574, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f576, %f575;
add.ftz.f32 %f577, %f576, 0f3F800000;
div.approx.ftz.f32 %f578, %f570, %f577;
add.ftz.f32 %f579, %f573, 0f3F800000;
mul.ftz.f32 %f580, %f579, %f578;
st.global.f32 [%rd34+12], %f580;
bra.uni $L__BB252_78;

}
.entry _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<38>;
.reg .b16 %rs<33>;
.reg .f32 %f<138>;
.reg .b32 %r<402>;
.reg .b64 %rd<79>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

ld.param.u64 %rd12, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd13, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd14, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
mov.u32 %r1, %tid.y;
shl.b32 %r38, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r38, %r2;
mov.u32 %r39, %ctaid.x;
shl.b32 %r4, %r39, 2;
mov.u32 %r399, %ctaid.y;
setp.eq.s64 %p1, %rd14, 0;
@%p1 bra $L__BB253_2;

cvta.to.global.u64 %rd16, %rd14;
mul.wide.u32 %rd17, %r399, 4;
add.s64 %rd18, %rd16, %rd17;
ld.global.nc.u32 %r398, [%rd18];
mul.hi.u32 %r40, %r399, %r21;
add.s32 %r41, %r40, %r399;
shr.u32 %r42, %r41, %r22;
mul.lo.s32 %r43, %r42, %r23;
sub.s32 %r399, %r399, %r43;
bra.uni $L__BB253_3;

$L__BB253_2:
mul.hi.u32 %r44, %r399, %r26;
add.s32 %r45, %r44, %r399;
shr.u32 %r398, %r45, %r27;

$L__BB253_3:
mov.u32 %r11, %ctaid.z;
shr.s32 %r46, %r3, 31;
shr.u32 %r47, %r46, 29;
add.s32 %r48, %r3, %r47;
shr.s32 %r401, %r48, 3;
shr.u32 %r49, %r20, 8;
setp.ge.s32 %p2, %r401, %r49;
mov.f32 %f134, 0f00000000;
mov.f32 %f135, %f134;
mov.f32 %f136, %f134;
mov.f32 %f137, %f134;
@%p2 bra $L__BB253_6;

and.b32 %r53, %r48, -8;
sub.s32 %r54, %r3, %r53;
shl.b32 %r55, %r54, 1;
cvt.s64.s32 %rd1, %r54;
shl.b32 %r400, %r401, 3;
mul.wide.s32 %rd19, %r54, 2;
mad.lo.s32 %r56, %r4, %r24, %r401;
mul.hi.u32 %r57, %r11, %r32;
add.s32 %r58, %r57, %r11;
shr.u32 %r59, %r58, %r33;
mad.lo.s32 %r60, %r59, %r35, %r56;
mad.lo.s32 %r61, %r398, %r29, %r60;
mul.wide.s32 %rd2, %r61, 50;
mul.wide.s32 %rd20, %r55, 2;
add.s64 %rd21, %rd2, %rd20;
add.s64 %rd3, %rd21, 4;
add.s64 %rd22, %rd2, %rd19;
add.s64 %rd4, %rd22, 34;
mul.wide.s32 %rd5, %r24, 50;
cvta.to.global.u64 %rd6, %rd13;
cvta.to.global.u64 %rd78, %rd12;
mul.lo.s32 %r62, %r399, %r30;
cvt.u64.u32 %rd23, %r62;
mul.lo.s32 %r63, %r11, %r36;
cvt.u64.u32 %rd24, %r63;
add.s64 %rd8, %rd23, %rd24;
mov.f32 %f134, 0f00000000;

$L__BB253_5:
cvt.u64.u32 %rd25, %r400;
add.s64 %rd26, %rd8, %rd25;
add.s64 %rd27, %rd26, %rd1;
mul.lo.s64 %rd28, %rd27, 36;
add.s64 %rd29, %rd78, %rd3;
add.s64 %rd30, %rd29, -2;
ld.global.nc.u16 %rs5, [%rd29+-2];
shr.u16 %rs6, %rs5, 8;
cvt.u32.u16 %r200, %rs6;
ld.global.nc.u16 %rs7, [%rd29];
shr.u16 %rs8, %rs7, 8;
cvt.u32.u16 %r201, %rs8;
add.s64 %rd31, %rd78, %rd4;
ld.global.nc.u16 %rs9, [%rd31];
cvt.u32.u16 %r202, %rs9;
and.b16 %rs10, %rs5, 255;
cvt.u32.u16 %r203, %rs10;
mul.wide.u16 %r204, %rs9, 256;
and.b32 %r205, %r204, 1792;
or.b32 %r206, %r205, %r203;
mul.wide.u32 %rd32, %r206, 4;
mov.u64 %rd33, iq1s_grid_gpu;
add.s64 %rd34, %rd33, %rd32;
ld.global.nc.u32 %r207, [%rd34];
and.b32 %r65, %r207, 252645135;
shr.u32 %r208, %r207, 4;
and.b32 %r69, %r208, 252645135;
add.s64 %rd35, %rd6, %rd28;
ld.global.nc.u32 %r66, [%rd35+4];
ld.global.nc.u32 %r70, [%rd35+8];
mov.u32 %r169, 0;

dp4a.s32.s32 %r64, %r65, %r66, %r169;


dp4a.s32.s32 %r68, %r69, %r70, %r64;

mul.wide.u16 %r209, %rs9, 32;
and.b32 %r210, %r209, 1792;
or.b32 %r211, %r210, %r200;
mul.wide.u32 %rd36, %r211, 4;
add.s64 %rd37, %rd33, %rd36;
ld.global.nc.u32 %r212, [%rd37];
and.b32 %r73, %r212, 252645135;
shr.u32 %r213, %r212, 4;
and.b32 %r77, %r213, 252645135;
ld.global.nc.u32 %r74, [%rd35+12];
ld.global.nc.u32 %r78, [%rd35+16];

dp4a.s32.s32 %r72, %r73, %r74, %r68;


dp4a.s32.s32 %r76, %r77, %r78, %r72;

and.b16 %rs11, %rs7, 255;
cvt.u32.u16 %r214, %rs11;
mul.wide.u16 %r215, %rs9, 4;
and.b32 %r216, %r215, 1792;
or.b32 %r217, %r216, %r214;
mul.wide.u32 %rd38, %r217, 4;
add.s64 %rd39, %rd33, %rd38;
ld.global.nc.u32 %r218, [%rd39];
and.b32 %r81, %r218, 252645135;
shr.u32 %r219, %r218, 4;
and.b32 %r85, %r219, 252645135;
ld.global.nc.u32 %r82, [%rd35+20];
ld.global.nc.u32 %r86, [%rd35+24];

dp4a.s32.s32 %r80, %r81, %r82, %r76;


dp4a.s32.s32 %r84, %r85, %r86, %r80;

shr.u32 %r220, %r202, 1;
and.b32 %r221, %r220, 1792;
or.b32 %r222, %r221, %r201;
mul.wide.u32 %rd40, %r222, 4;
add.s64 %rd41, %rd33, %rd40;
ld.global.nc.u32 %r223, [%rd41];
and.b32 %r89, %r223, 252645135;
shr.u32 %r224, %r223, 4;
and.b32 %r93, %r224, 252645135;
ld.global.nc.u32 %r90, [%rd35+28];
ld.global.nc.u32 %r94, [%rd35+32];

dp4a.s32.s32 %r88, %r89, %r90, %r84;


dp4a.s32.s32 %r92, %r93, %r94, %r88;

add.s64 %rd42, %rd78, %rd2;
ld.global.nc.u16 %rs1, [%rd42];

{ cvt.f32.f16 %f25, %rs1;}


shr.u32 %r225, %r202, 11;
and.b32 %r226, %r225, 14;
or.b32 %r227, %r226, 1;
cvt.rn.f32.s32 %f37, %r227;
mul.ftz.f32 %f38, %f25, %f37;
and.b32 %r228, %r202, 32768;
cvt.rn.f32.s32 %f39, %r228;
mul.ftz.f32 %f40, %f39, 0f37000000;
mov.f32 %f41, 0fBF600000;
sub.ftz.f32 %f42, %f41, %f40;
ld.global.nc.u32 %r96, [%rd35];

{.reg .f16 low,high;
mov.b32 {low,high},%r96;
cvt.f32.f16 %f26, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r96;
cvt.f32.f16 %f27, high;}


cvt.rn.f32.s32 %f43, %r92;
mul.ftz.f32 %f44, %f42, %f27;
fma.rn.ftz.f32 %f45, %f26, %f43, %f44;
fma.rn.ftz.f32 %f137, %f38, %f45, %f137;
add.s64 %rd43, %rd30, %rd5;
ld.global.nc.u16 %rs12, [%rd43];
shr.u16 %rs13, %rs12, 8;
cvt.u32.u16 %r229, %rs13;
ld.global.nc.u16 %rs14, [%rd43+2];
shr.u16 %rs15, %rs14, 8;
cvt.u32.u16 %r230, %rs15;
add.s64 %rd44, %rd31, %rd5;
ld.global.nc.u16 %rs16, [%rd44];
cvt.u32.u16 %r231, %rs16;
and.b16 %rs17, %rs12, 255;
cvt.u32.u16 %r232, %rs17;
mul.wide.u16 %r233, %rs16, 256;
and.b32 %r234, %r233, 1792;
or.b32 %r235, %r234, %r232;
mul.wide.u32 %rd45, %r235, 4;
add.s64 %rd46, %rd33, %rd45;
ld.global.nc.u32 %r236, [%rd46];
and.b32 %r99, %r236, 252645135;
shr.u32 %r237, %r236, 4;
and.b32 %r103, %r237, 252645135;

dp4a.s32.s32 %r98, %r99, %r66, %r169;


dp4a.s32.s32 %r102, %r103, %r70, %r98;

mul.wide.u16 %r238, %rs16, 32;
and.b32 %r239, %r238, 1792;
or.b32 %r240, %r239, %r229;
mul.wide.u32 %rd47, %r240, 4;
add.s64 %rd48, %rd33, %rd47;
ld.global.nc.u32 %r241, [%rd48];
and.b32 %r107, %r241, 252645135;
shr.u32 %r242, %r241, 4;
and.b32 %r111, %r242, 252645135;

dp4a.s32.s32 %r106, %r107, %r74, %r102;


dp4a.s32.s32 %r110, %r111, %r78, %r106;

and.b16 %rs18, %rs14, 255;
cvt.u32.u16 %r243, %rs18;
mul.wide.u16 %r244, %rs16, 4;
and.b32 %r245, %r244, 1792;
or.b32 %r246, %r245, %r243;
mul.wide.u32 %rd49, %r246, 4;
add.s64 %rd50, %rd33, %rd49;
ld.global.nc.u32 %r247, [%rd50];
and.b32 %r115, %r247, 252645135;
shr.u32 %r248, %r247, 4;
and.b32 %r119, %r248, 252645135;

dp4a.s32.s32 %r114, %r115, %r82, %r110;


dp4a.s32.s32 %r118, %r119, %r86, %r114;

shr.u32 %r249, %r231, 1;
and.b32 %r250, %r249, 1792;
or.b32 %r251, %r250, %r230;
mul.wide.u32 %rd51, %r251, 4;
add.s64 %rd52, %rd33, %rd51;
ld.global.nc.u32 %r252, [%rd52];
and.b32 %r123, %r252, 252645135;
shr.u32 %r253, %r252, 4;
and.b32 %r127, %r253, 252645135;

dp4a.s32.s32 %r122, %r123, %r90, %r118;


dp4a.s32.s32 %r126, %r127, %r94, %r122;

add.s64 %rd53, %rd42, %rd5;
ld.global.nc.u16 %rs2, [%rd53];

{ cvt.f32.f16 %f28, %rs2;}


shr.u32 %r254, %r231, 11;
and.b32 %r255, %r254, 14;
or.b32 %r256, %r255, 1;
cvt.rn.f32.s32 %f46, %r256;
mul.ftz.f32 %f47, %f28, %f46;
and.b32 %r257, %r231, 32768;
cvt.rn.f32.s32 %f48, %r257;
mul.ftz.f32 %f49, %f48, 0f37000000;
sub.ftz.f32 %f50, %f41, %f49;

{.reg .f16 low,high;
mov.b32 {low,high},%r96;
cvt.f32.f16 %f29, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r96;
cvt.f32.f16 %f30, high;}


cvt.rn.f32.s32 %f51, %r126;
mul.ftz.f32 %f52, %f50, %f30;
fma.rn.ftz.f32 %f53, %f29, %f51, %f52;
fma.rn.ftz.f32 %f136, %f47, %f53, %f136;
add.s64 %rd54, %rd43, %rd5;
ld.global.nc.u16 %rs19, [%rd54];
shr.u16 %rs20, %rs19, 8;
cvt.u32.u16 %r258, %rs20;
ld.global.nc.u16 %rs21, [%rd54+2];
shr.u16 %rs22, %rs21, 8;
cvt.u32.u16 %r259, %rs22;
add.s64 %rd55, %rd44, %rd5;
ld.global.nc.u16 %rs23, [%rd55];
cvt.u32.u16 %r260, %rs23;
and.b16 %rs24, %rs19, 255;
cvt.u32.u16 %r261, %rs24;
mul.wide.u16 %r262, %rs23, 256;
and.b32 %r263, %r262, 1792;
or.b32 %r264, %r263, %r261;
mul.wide.u32 %rd56, %r264, 4;
add.s64 %rd57, %rd33, %rd56;
ld.global.nc.u32 %r265, [%rd57];
and.b32 %r133, %r265, 252645135;
shr.u32 %r266, %r265, 4;
and.b32 %r137, %r266, 252645135;

dp4a.s32.s32 %r132, %r133, %r66, %r169;


dp4a.s32.s32 %r136, %r137, %r70, %r132;

mul.wide.u16 %r267, %rs23, 32;
and.b32 %r268, %r267, 1792;
or.b32 %r269, %r268, %r258;
mul.wide.u32 %rd58, %r269, 4;
add.s64 %rd59, %rd33, %rd58;
ld.global.nc.u32 %r270, [%rd59];
and.b32 %r141, %r270, 252645135;
shr.u32 %r271, %r270, 4;
and.b32 %r145, %r271, 252645135;

dp4a.s32.s32 %r140, %r141, %r74, %r136;


dp4a.s32.s32 %r144, %r145, %r78, %r140;

and.b16 %rs25, %rs21, 255;
cvt.u32.u16 %r272, %rs25;
mul.wide.u16 %r273, %rs23, 4;
and.b32 %r274, %r273, 1792;
or.b32 %r275, %r274, %r272;
mul.wide.u32 %rd60, %r275, 4;
add.s64 %rd61, %rd33, %rd60;
ld.global.nc.u32 %r276, [%rd61];
and.b32 %r149, %r276, 252645135;
shr.u32 %r277, %r276, 4;
and.b32 %r153, %r277, 252645135;

dp4a.s32.s32 %r148, %r149, %r82, %r144;


dp4a.s32.s32 %r152, %r153, %r86, %r148;

shr.u32 %r278, %r260, 1;
and.b32 %r279, %r278, 1792;
or.b32 %r280, %r279, %r259;
mul.wide.u32 %rd62, %r280, 4;
add.s64 %rd63, %rd33, %rd62;
ld.global.nc.u32 %r281, [%rd63];
and.b32 %r157, %r281, 252645135;
shr.u32 %r282, %r281, 4;
and.b32 %r161, %r282, 252645135;

dp4a.s32.s32 %r156, %r157, %r90, %r152;


dp4a.s32.s32 %r160, %r161, %r94, %r156;

add.s64 %rd64, %rd53, %rd5;
ld.global.nc.u16 %rs3, [%rd64];

{ cvt.f32.f16 %f31, %rs3;}


shr.u32 %r283, %r260, 11;
and.b32 %r284, %r283, 14;
or.b32 %r285, %r284, 1;
cvt.rn.f32.s32 %f54, %r285;
mul.ftz.f32 %f55, %f31, %f54;
and.b32 %r286, %r260, 32768;
cvt.rn.f32.s32 %f56, %r286;
mul.ftz.f32 %f57, %f56, 0f37000000;
sub.ftz.f32 %f58, %f41, %f57;

{.reg .f16 low,high;
mov.b32 {low,high},%r96;
cvt.f32.f16 %f32, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r96;
cvt.f32.f16 %f33, high;}


cvt.rn.f32.s32 %f59, %r160;
mul.ftz.f32 %f60, %f58, %f33;
fma.rn.ftz.f32 %f61, %f32, %f59, %f60;
fma.rn.ftz.f32 %f135, %f55, %f61, %f135;
add.s64 %rd65, %rd54, %rd5;
ld.global.nc.u16 %rs26, [%rd65];
shr.u16 %rs27, %rs26, 8;
cvt.u32.u16 %r287, %rs27;
ld.global.nc.u16 %rs28, [%rd65+2];
shr.u16 %rs29, %rs28, 8;
cvt.u32.u16 %r288, %rs29;
add.s64 %rd66, %rd55, %rd5;
ld.global.nc.u16 %rs30, [%rd66];
cvt.u32.u16 %r289, %rs30;
and.b16 %rs31, %rs26, 255;
cvt.u32.u16 %r290, %rs31;
mul.wide.u16 %r291, %rs30, 256;
and.b32 %r292, %r291, 1792;
or.b32 %r293, %r292, %r290;
mul.wide.u32 %rd67, %r293, 4;
add.s64 %rd68, %rd33, %rd67;
ld.global.nc.u32 %r294, [%rd68];
and.b32 %r167, %r294, 252645135;
shr.u32 %r295, %r294, 4;
and.b32 %r171, %r295, 252645135;

dp4a.s32.s32 %r166, %r167, %r66, %r169;


dp4a.s32.s32 %r170, %r171, %r70, %r166;

mul.wide.u16 %r296, %rs30, 32;
and.b32 %r297, %r296, 1792;
or.b32 %r298, %r297, %r287;
mul.wide.u32 %rd69, %r298, 4;
add.s64 %rd70, %rd33, %rd69;
ld.global.nc.u32 %r299, [%rd70];
and.b32 %r175, %r299, 252645135;
shr.u32 %r300, %r299, 4;
and.b32 %r179, %r300, 252645135;

dp4a.s32.s32 %r174, %r175, %r74, %r170;


dp4a.s32.s32 %r178, %r179, %r78, %r174;

and.b16 %rs32, %rs28, 255;
cvt.u32.u16 %r301, %rs32;
mul.wide.u16 %r302, %rs30, 4;
and.b32 %r303, %r302, 1792;
or.b32 %r304, %r303, %r301;
mul.wide.u32 %rd71, %r304, 4;
add.s64 %rd72, %rd33, %rd71;
ld.global.nc.u32 %r305, [%rd72];
and.b32 %r183, %r305, 252645135;
shr.u32 %r306, %r305, 4;
and.b32 %r187, %r306, 252645135;

dp4a.s32.s32 %r182, %r183, %r82, %r178;


dp4a.s32.s32 %r186, %r187, %r86, %r182;

shr.u32 %r307, %r289, 1;
and.b32 %r308, %r307, 1792;
or.b32 %r309, %r308, %r288;
mul.wide.u32 %rd73, %r309, 4;
add.s64 %rd74, %rd33, %rd73;
ld.global.nc.u32 %r310, [%rd74];
and.b32 %r191, %r310, 252645135;
shr.u32 %r311, %r310, 4;
and.b32 %r195, %r311, 252645135;

dp4a.s32.s32 %r190, %r191, %r90, %r186;


dp4a.s32.s32 %r194, %r195, %r94, %r190;

add.s64 %rd75, %rd64, %rd5;
ld.global.nc.u16 %rs4, [%rd75];

{ cvt.f32.f16 %f34, %rs4;}


shr.u32 %r312, %r289, 11;
and.b32 %r313, %r312, 14;
or.b32 %r314, %r313, 1;
cvt.rn.f32.s32 %f62, %r314;
mul.ftz.f32 %f63, %f34, %f62;
and.b32 %r315, %r289, 32768;
cvt.rn.f32.s32 %f64, %r315;
mul.ftz.f32 %f65, %f64, 0f37000000;
sub.ftz.f32 %f66, %f41, %f65;

{.reg .f16 low,high;
mov.b32 {low,high},%r96;
cvt.f32.f16 %f35, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r96;
cvt.f32.f16 %f36, high;}


cvt.rn.f32.s32 %f67, %r194;
mul.ftz.f32 %f68, %f66, %f36;
fma.rn.ftz.f32 %f69, %f35, %f67, %f68;
fma.rn.ftz.f32 %f134, %f63, %f69, %f134;
add.s32 %r400, %r400, 128;
add.s64 %rd78, %rd78, 800;
add.s32 %r401, %r401, 16;
setp.lt.s32 %p3, %r401, %r49;
@%p3 bra $L__BB253_5;

$L__BB253_6:
setp.eq.s32 %p4, %r1, 0;
@%p4 bra $L__BB253_8;

shl.b32 %r317, %r1, 9;
mov.u32 %r318, _ZZ13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r319, %r318, %r317;
shl.b32 %r320, %r2, 2;
add.s32 %r321, %r319, %r320;
st.shared.f32 [%r321+-512], %f137;
st.shared.f32 [%r321+-384], %f136;
st.shared.f32 [%r321+-256], %f135;
st.shared.f32 [%r321+-128], %f134;

$L__BB253_8:
setp.ne.s32 %p5, %r1, 0;
bar.sync 0;
@%p5 bra $L__BB253_17;

mov.u32 %r322, %ctaid.y;
mad.lo.s32 %r323, %r322, %r31, %r4;
mad.lo.s32 %r324, %r11, %r37, %r323;
shl.b32 %r325, %r2, 2;
mov.u32 %r326, _ZZ13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r18, %r326, %r325;
mov.u32 %r327, 2;
ld.shared.f32 %f70, [%r18];
add.ftz.f32 %f71, %f70, %f137;
ld.shared.f32 %f72, [%r18+512];
add.ftz.f32 %f73, %f72, %f71;
ld.shared.f32 %f74, [%r18+1024];
add.ftz.f32 %f75, %f74, %f73;
mov.b32 %r328, %f75;
mov.u32 %r329, 31;
mov.u32 %r330, 16;
mov.u32 %r331, -1;
shfl.sync.bfly.b32 %r332|%p6, %r328, %r330, %r329, %r331;
mov.b32 %f76, %r332;
add.ftz.f32 %f77, %f75, %f76;
mov.b32 %r333, %f77;
mov.u32 %r334, 8;
shfl.sync.bfly.b32 %r335|%p7, %r333, %r334, %r329, %r331;
mov.b32 %f78, %r335;
add.ftz.f32 %f79, %f77, %f78;
mov.b32 %r336, %f79;
mov.u32 %r337, 4;
shfl.sync.bfly.b32 %r338|%p8, %r336, %r337, %r329, %r331;
mov.b32 %f80, %r338;
add.ftz.f32 %f81, %f79, %f80;
mov.b32 %r339, %f81;
shfl.sync.bfly.b32 %r340|%p9, %r339, %r327, %r329, %r331;
mov.b32 %f82, %r340;
add.ftz.f32 %f83, %f81, %f82;
mov.b32 %r341, %f83;
mov.u32 %r342, 1;
shfl.sync.bfly.b32 %r343|%p10, %r341, %r342, %r329, %r331;
mov.b32 %f84, %r343;
add.ftz.f32 %f13, %f83, %f84;
setp.ge.u32 %p11, %r4, %r25;
setp.ne.s32 %p12, %r2, 0;
cvta.to.global.u64 %rd76, %rd15;
mul.wide.u32 %rd77, %r324, 4;
add.s64 %rd11, %rd76, %rd77;
or.pred %p13, %p12, %p11;
@%p13 bra $L__BB253_11;

st.global.f32 [%rd11], %f13;

$L__BB253_11:
ld.shared.f32 %f85, [%r18+128];
add.ftz.f32 %f86, %f85, %f136;
ld.shared.f32 %f87, [%r18+640];
add.ftz.f32 %f88, %f87, %f86;
ld.shared.f32 %f89, [%r18+1152];
add.ftz.f32 %f90, %f89, %f88;
mov.b32 %r344, %f90;
shfl.sync.bfly.b32 %r348|%p14, %r344, %r330, %r329, %r331;
mov.b32 %f91, %r348;
add.ftz.f32 %f92, %f90, %f91;
mov.b32 %r349, %f92;
shfl.sync.bfly.b32 %r351|%p15, %r349, %r334, %r329, %r331;
mov.b32 %f93, %r351;
add.ftz.f32 %f94, %f92, %f93;
mov.b32 %r352, %f94;
shfl.sync.bfly.b32 %r354|%p16, %r352, %r337, %r329, %r331;
mov.b32 %f95, %r354;
add.ftz.f32 %f96, %f94, %f95;
mov.b32 %r355, %f96;
shfl.sync.bfly.b32 %r357|%p17, %r355, %r327, %r329, %r331;
mov.b32 %f97, %r357;
add.ftz.f32 %f98, %f96, %f97;
mov.b32 %r358, %f98;
shfl.sync.bfly.b32 %r360|%p18, %r358, %r342, %r329, %r331;
mov.b32 %f99, %r360;
add.ftz.f32 %f14, %f98, %f99;
add.s32 %r361, %r4, 1;
setp.ge.u32 %p19, %r361, %r25;
setp.ne.s32 %p20, %r2, 1;
or.pred %p21, %p20, %p19;
@%p21 bra $L__BB253_13;

st.global.f32 [%rd11+4], %f14;

$L__BB253_13:
ld.shared.f32 %f100, [%r18+256];
add.ftz.f32 %f101, %f100, %f135;
ld.shared.f32 %f102, [%r18+768];
add.ftz.f32 %f103, %f102, %f101;
ld.shared.f32 %f104, [%r18+1280];
add.ftz.f32 %f105, %f104, %f103;
mov.b32 %r362, %f105;
mov.u32 %r363, 31;
mov.u32 %r364, 16;
mov.u32 %r365, -1;
shfl.sync.bfly.b32 %r366|%p22, %r362, %r364, %r363, %r365;
mov.b32 %f106, %r366;
add.ftz.f32 %f107, %f105, %f106;
mov.b32 %r367, %f107;
mov.u32 %r368, 8;
shfl.sync.bfly.b32 %r369|%p23, %r367, %r368, %r363, %r365;
mov.b32 %f108, %r369;
add.ftz.f32 %f109, %f107, %f108;
mov.b32 %r370, %f109;
mov.u32 %r371, 4;
shfl.sync.bfly.b32 %r372|%p24, %r370, %r371, %r363, %r365;
mov.b32 %f110, %r372;
add.ftz.f32 %f111, %f109, %f110;
mov.b32 %r373, %f111;
mov.u32 %r374, 2;
shfl.sync.bfly.b32 %r375|%p25, %r373, %r374, %r363, %r365;
mov.b32 %f112, %r375;
add.ftz.f32 %f113, %f111, %f112;
mov.b32 %r376, %f113;
mov.u32 %r377, 1;
shfl.sync.bfly.b32 %r378|%p26, %r376, %r377, %r363, %r365;
mov.b32 %f114, %r378;
add.ftz.f32 %f15, %f113, %f114;
add.s32 %r379, %r4, 2;
setp.ge.u32 %p27, %r379, %r25;
setp.ne.s32 %p28, %r2, 2;
or.pred %p29, %p28, %p27;
@%p29 bra $L__BB253_15;

st.global.f32 [%rd11+8], %f15;

$L__BB253_15:
ld.shared.f32 %f115, [%r18+384];
add.ftz.f32 %f116, %f115, %f134;
ld.shared.f32 %f117, [%r18+896];
add.ftz.f32 %f118, %f117, %f116;
ld.shared.f32 %f119, [%r18+1408];
add.ftz.f32 %f120, %f119, %f118;
mov.b32 %r380, %f120;
shfl.sync.bfly.b32 %r384|%p30, %r380, %r364, %r363, %r365;
mov.b32 %f121, %r384;
add.ftz.f32 %f122, %f120, %f121;
mov.b32 %r385, %f122;
shfl.sync.bfly.b32 %r387|%p31, %r385, %r368, %r363, %r365;
mov.b32 %f123, %r387;
add.ftz.f32 %f124, %f122, %f123;
mov.b32 %r388, %f124;
shfl.sync.bfly.b32 %r390|%p32, %r388, %r371, %r363, %r365;
mov.b32 %f125, %r390;
add.ftz.f32 %f126, %f124, %f125;
mov.b32 %r391, %f126;
shfl.sync.bfly.b32 %r393|%p33, %r391, %r374, %r363, %r365;
mov.b32 %f127, %r393;
add.ftz.f32 %f128, %f126, %f127;
mov.b32 %r394, %f128;
shfl.sync.bfly.b32 %r396|%p34, %r394, %r377, %r363, %r365;
mov.b32 %f129, %r396;
add.ftz.f32 %f16, %f128, %f129;
add.s32 %r397, %r4, 3;
setp.ge.u32 %p35, %r397, %r25;
setp.ne.s32 %p36, %r2, 3;
or.pred %p37, %p36, %p35;
@%p37 bra $L__BB253_17;

st.global.f32 [%rd11+12], %f16;

$L__BB253_17:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<34>;
.reg .b16 %rs<17>;
.reg .f32 %f<133>;
.reg .b32 %r<286>;
.reg .b64 %rd<87>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[384];

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate[384];

ld.param.u64 %rd8, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd9, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd10, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
ld.param.u64 %rd13, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+16];
ld.param.u64 %rd12, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+8];
ld.param.u64 %rd11, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3];
mov.u32 %r1, %tid.y;
mov.u32 %r2, %ctaid.y;
setp.eq.s64 %p2, %rd10, 0;
@%p2 bra $L__BB254_2;

cvta.to.global.u64 %rd17, %rd10;
mul.wide.u32 %rd18, %r2, 4;
add.s64 %rd19, %rd17, %rd18;
ld.global.nc.u32 %r282, [%rd19];
mul.hi.u32 %r45, %r2, %r29;
add.s32 %r46, %r45, %r2;
shr.u32 %r47, %r46, %r30;
mul.lo.s32 %r48, %r47, %r31;
sub.s32 %r283, %r2, %r48;
bra.uni $L__BB254_3;

$L__BB254_2:
mul.hi.u32 %r49, %r2, %r33;
add.s32 %r50, %r49, %r2;
shr.u32 %r282, %r50, %r34;
mov.u32 %r283, %r2;

$L__BB254_3:
mov.u32 %r51, %tid.x;
or.b32 %r52, %r51, %r1;
setp.ne.s32 %p3, %r52, 0;
mov.f32 %f126, 0f00000000;
mov.f32 %f122, %f126;
mov.f32 %f123, %f126;
@%p3 bra $L__BB254_8;

setp.eq.s64 %p4, %rd11, 0;
mov.f32 %f122, 0f00000000;
mov.f32 %f123, %f122;
@%p4 bra $L__BB254_6;

cvta.to.global.u64 %rd20, %rd11;
mov.u32 %r53, %ctaid.z;
mul.lo.s32 %r54, %r53, %r44;
cvt.u64.u32 %rd21, %r54;
selp.b32 %r55, %r2, %r282, %p2;
mul.lo.s32 %r56, %r55, %r38;
cvt.u64.u32 %rd22, %r56;
mov.u32 %r57, %ctaid.x;
cvt.s64.s32 %rd23, %r57;
cvt.u64.u32 %rd24, %r51;
add.s64 %rd25, %rd24, %rd23;
add.s64 %rd26, %rd25, %rd21;
add.s64 %rd27, %rd26, %rd22;
shl.b64 %rd28, %rd27, 2;
add.s64 %rd29, %rd20, %rd28;
ld.global.f32 %f123, [%rd29];

$L__BB254_6:
setp.eq.s64 %p6, %rd13, 0;
setp.eq.s64 %p7, %rd12, 0;
or.pred %p8, %p6, %p7;
@%p8 bra $L__BB254_8;

cvta.to.global.u64 %rd30, %rd13;
mov.u32 %r59, %ctaid.z;
mul.lo.s32 %r60, %r59, %r44;
cvt.u64.u32 %rd31, %r60;
selp.b32 %r61, %r2, %r282, %p2;
mul.lo.s32 %r62, %r61, %r38;
cvt.u64.u32 %rd32, %r62;
mov.u32 %r63, %ctaid.x;
cvt.s64.s32 %rd33, %r63;
cvt.u64.u32 %rd34, %r51;
add.s64 %rd35, %rd34, %rd33;
add.s64 %rd36, %rd35, %rd31;
add.s64 %rd37, %rd36, %rd32;
shl.b64 %rd38, %rd37, 2;
add.s64 %rd39, %rd30, %rd38;
ld.global.f32 %f122, [%rd39];

$L__BB254_8:
shl.b32 %r66, %r1, 5;
add.s32 %r67, %r66, %r51;
shr.s32 %r68, %r67, 31;
shr.u32 %r69, %r68, 29;
add.s32 %r70, %r67, %r69;
shr.s32 %r285, %r70, 3;
shr.u32 %r71, %r28, 8;
setp.ge.s32 %p10, %r285, %r71;
mov.f32 %f128, %f126;
@%p10 bra $L__BB254_13;

mov.u32 %r72, %ctaid.z;
and.b32 %r79, %r70, -8;
sub.s32 %r9, %r67, %r79;
shl.b32 %r284, %r285, 3;
mul.hi.u32 %r80, %r72, %r39;
add.s32 %r81, %r80, %r72;
shr.u32 %r82, %r81, %r40;
mad.lo.s32 %r83, %r82, %r42, %r285;
mad.lo.s32 %r84, %r282, %r36, %r83;
mov.u32 %r85, %ctaid.x;
mad.lo.s32 %r86, %r85, %r32, %r84;
mul.wide.s32 %rd86, %r86, 50;
mul.lo.s32 %r87, %r283, %r37;
cvt.u64.u32 %rd40, %r87;
mul.lo.s32 %r88, %r72, %r43;
cvt.u64.u32 %rd41, %r88;
add.s64 %rd5, %rd40, %rd41;
mov.f32 %f128, 0f00000000;
cvta.to.global.u64 %rd42, %rd9;
cvta.to.global.u64 %rd46, %rd8;
cvt.s64.s32 %rd49, %r9;
mul.wide.s32 %rd50, %r9, 2;
add.s64 %rd51, %rd46, %rd50;
mov.f32 %f126, %f128;

$L__BB254_10:
cvt.u64.u32 %rd43, %r284;
add.s64 %rd44, %rd5, %rd43;
shl.b32 %r123, %r9, 1;
mul.wide.s32 %rd45, %r123, 2;
add.s64 %rd47, %rd46, %rd45;
add.s64 %rd48, %rd47, %rd86;
ld.global.nc.u16 %rs2, [%rd48+2];
shr.u16 %rs3, %rs2, 8;
cvt.u32.u16 %r124, %rs3;
ld.global.nc.u16 %rs4, [%rd48+4];
shr.u16 %rs5, %rs4, 8;
cvt.u32.u16 %r125, %rs5;
add.s64 %rd52, %rd51, %rd86;
ld.global.nc.u16 %rs6, [%rd52+34];
cvt.u32.u16 %r126, %rs6;
add.s64 %rd53, %rd44, %rd49;
mul.lo.s64 %rd54, %rd53, 36;
and.b16 %rs7, %rs2, 255;
cvt.u32.u16 %r127, %rs7;
mul.wide.u16 %r128, %rs6, 256;
and.b32 %r129, %r128, 1792;
or.b32 %r130, %r129, %r127;
mul.wide.u32 %rd55, %r130, 4;
mov.u64 %rd56, iq1s_grid_gpu;
add.s64 %rd57, %rd56, %rd55;
ld.global.nc.u32 %r131, [%rd57];
and.b32 %r90, %r131, 252645135;
shr.u32 %r132, %r131, 4;
and.b32 %r94, %r132, 252645135;
add.s64 %rd58, %rd42, %rd54;
ld.global.nc.u32 %r13, [%rd58+4];
ld.global.nc.u32 %r14, [%rd58+8];
mov.u32 %r92, 0;

dp4a.s32.s32 %r89, %r90, %r13, %r92;


dp4a.s32.s32 %r93, %r94, %r14, %r89;

mul.wide.u16 %r133, %rs6, 32;
and.b32 %r134, %r133, 1792;
or.b32 %r135, %r134, %r124;
mul.wide.u32 %rd59, %r135, 4;
add.s64 %rd60, %rd56, %rd59;
ld.global.nc.u32 %r136, [%rd60];
and.b32 %r98, %r136, 252645135;
shr.u32 %r137, %r136, 4;
and.b32 %r102, %r137, 252645135;
ld.global.nc.u32 %r15, [%rd58+12];
ld.global.nc.u32 %r16, [%rd58+16];

dp4a.s32.s32 %r97, %r98, %r15, %r93;


dp4a.s32.s32 %r101, %r102, %r16, %r97;

and.b16 %rs8, %rs4, 255;
cvt.u32.u16 %r138, %rs8;
mul.wide.u16 %r139, %rs6, 4;
and.b32 %r140, %r139, 1792;
or.b32 %r141, %r140, %r138;
mul.wide.u32 %rd61, %r141, 4;
add.s64 %rd62, %rd56, %rd61;
ld.global.nc.u32 %r142, [%rd62];
and.b32 %r106, %r142, 252645135;
shr.u32 %r143, %r142, 4;
and.b32 %r110, %r143, 252645135;
ld.global.nc.u32 %r17, [%rd58+20];
ld.global.nc.u32 %r18, [%rd58+24];

dp4a.s32.s32 %r105, %r106, %r17, %r101;


dp4a.s32.s32 %r109, %r110, %r18, %r105;

shr.u32 %r144, %r126, 1;
and.b32 %r145, %r144, 1792;
or.b32 %r146, %r145, %r125;
mul.wide.u32 %rd63, %r146, 4;
add.s64 %rd64, %rd56, %rd63;
ld.global.nc.u32 %r147, [%rd64];
and.b32 %r114, %r147, 252645135;
shr.u32 %r148, %r147, 4;
and.b32 %r118, %r148, 252645135;
ld.global.nc.u32 %r19, [%rd58+28];
ld.global.nc.u32 %r20, [%rd58+32];

dp4a.s32.s32 %r113, %r114, %r19, %r109;


dp4a.s32.s32 %r117, %r118, %r20, %r113;

add.s64 %rd65, %rd46, %rd86;
ld.global.nc.u16 %rs1, [%rd65];

{ cvt.f32.f16 %f40, %rs1;}


shr.u32 %r149, %r126, 11;
and.b32 %r150, %r149, 14;
or.b32 %r151, %r150, 1;
cvt.rn.f32.s32 %f43, %r151;
mul.ftz.f32 %f44, %f40, %f43;
and.b32 %r152, %r126, 32768;
cvt.rn.f32.s32 %f45, %r152;
fma.rn.ftz.f32 %f46, %f45, 0fB7000000, 0fBF600000;
ld.global.nc.u32 %r21, [%rd58];

{.reg .f16 low,high;
mov.b32 {low,high},%r21;
cvt.f32.f16 %f41, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r21;
cvt.f32.f16 %f42, high;}


cvt.rn.f32.s32 %f47, %r117;
mul.ftz.f32 %f48, %f46, %f42;
fma.rn.ftz.f32 %f49, %f41, %f47, %f48;
fma.rn.ftz.f32 %f128, %f44, %f49, %f128;
setp.eq.s64 %p11, %rd12, 0;
@%p11 bra $L__BB254_12;

cvta.to.global.u64 %rd67, %rd12;
add.s64 %rd68, %rd67, %rd45;
add.s64 %rd69, %rd68, %rd86;
ld.global.u16 %rs10, [%rd69+2];
shr.u16 %rs11, %rs10, 8;
cvt.u32.u16 %r188, %rs11;
ld.global.u16 %rs12, [%rd69+4];
shr.u16 %rs13, %rs12, 8;
cvt.u32.u16 %r189, %rs13;
add.s64 %rd71, %rd67, %rd50;
add.s64 %rd72, %rd71, %rd86;
ld.global.u16 %rs14, [%rd72+34];
cvt.u32.u16 %r190, %rs14;
and.b16 %rs15, %rs10, 255;
cvt.u32.u16 %r191, %rs15;
mul.wide.u16 %r192, %rs14, 256;
and.b32 %r193, %r192, 1792;
or.b32 %r194, %r193, %r191;
mul.wide.u32 %rd73, %r194, 4;
add.s64 %rd75, %rd56, %rd73;
ld.global.nc.u32 %r195, [%rd75];
and.b32 %r154, %r195, 252645135;
shr.u32 %r196, %r195, 4;
and.b32 %r158, %r196, 252645135;

dp4a.s32.s32 %r153, %r154, %r13, %r92;


dp4a.s32.s32 %r157, %r158, %r14, %r153;

mul.wide.u16 %r197, %rs14, 32;
and.b32 %r198, %r197, 1792;
or.b32 %r199, %r198, %r188;
mul.wide.u32 %rd76, %r199, 4;
add.s64 %rd77, %rd56, %rd76;
ld.global.nc.u32 %r200, [%rd77];
and.b32 %r162, %r200, 252645135;
shr.u32 %r201, %r200, 4;
and.b32 %r166, %r201, 252645135;

dp4a.s32.s32 %r161, %r162, %r15, %r157;


dp4a.s32.s32 %r165, %r166, %r16, %r161;

and.b16 %rs16, %rs12, 255;
cvt.u32.u16 %r202, %rs16;
mul.wide.u16 %r203, %rs14, 4;
and.b32 %r204, %r203, 1792;
or.b32 %r205, %r204, %r202;
mul.wide.u32 %rd78, %r205, 4;
add.s64 %rd79, %rd56, %rd78;
ld.global.nc.u32 %r206, [%rd79];
and.b32 %r170, %r206, 252645135;
shr.u32 %r207, %r206, 4;
and.b32 %r174, %r207, 252645135;

dp4a.s32.s32 %r169, %r170, %r17, %r165;


dp4a.s32.s32 %r173, %r174, %r18, %r169;

shr.u32 %r208, %r190, 1;
and.b32 %r209, %r208, 1792;
or.b32 %r210, %r209, %r189;
mul.wide.u32 %rd80, %r210, 4;
add.s64 %rd81, %rd56, %rd80;
ld.global.nc.u32 %r211, [%rd81];
and.b32 %r178, %r211, 252645135;
shr.u32 %r212, %r211, 4;
and.b32 %r182, %r212, 252645135;

dp4a.s32.s32 %r177, %r178, %r19, %r173;


dp4a.s32.s32 %r181, %r182, %r20, %r177;

add.s64 %rd82, %rd67, %rd86;
ld.global.u16 %rs9, [%rd82];

{ cvt.f32.f16 %f50, %rs9;}


shr.u32 %r213, %r190, 11;
and.b32 %r214, %r213, 14;
or.b32 %r215, %r214, 1;
cvt.rn.f32.s32 %f53, %r215;
mul.ftz.f32 %f54, %f50, %f53;
and.b32 %r216, %r190, 32768;
cvt.rn.f32.s32 %f55, %r216;
fma.rn.ftz.f32 %f56, %f55, 0fB7000000, 0fBF600000;

{.reg .f16 low,high;
mov.b32 {low,high},%r21;
cvt.f32.f16 %f51, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r21;
cvt.f32.f16 %f52, high;}


cvt.rn.f32.s32 %f57, %r181;
mul.ftz.f32 %f58, %f56, %f52;
fma.rn.ftz.f32 %f59, %f51, %f57, %f58;
fma.rn.ftz.f32 %f126, %f54, %f59, %f126;

$L__BB254_12:
add.s32 %r284, %r284, 128;
add.s64 %rd86, %rd86, 800;
add.s32 %r285, %r285, 16;
setp.lt.s32 %p12, %r285, %r71;
@%p12 bra $L__BB254_10;

$L__BB254_13:
setp.eq.s32 %p13, %r1, 0;
@%p13 bra $L__BB254_16;

setp.eq.s64 %p14, %rd12, 0;
add.s32 %r24, %r1, -1;
shl.b32 %r221, %r1, 7;
mov.u32 %r222, _ZZ13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r223, %r222, %r221;
shl.b32 %r224, %r51, 2;
add.s32 %r225, %r223, %r224;
st.shared.f32 [%r225+-128], %f128;
@%p14 bra $L__BB254_16;

shl.b32 %r227, %r24, 7;
mov.u32 %r228, _ZZ13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r229, %r228, %r227;
add.s32 %r231, %r229, %r224;
st.shared.f32 [%r231], %f126;

$L__BB254_16:
setp.ne.s32 %p15, %r1, 0;
bar.sync 0;
@%p15 bra $L__BB254_35;

setp.ne.s64 %p16, %rd12, 0;
shl.b32 %r234, %r51, 2;
mov.u32 %r235, _ZZ13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r25, %r235, %r234;
ld.shared.f32 %f60, [%r25];
add.ftz.f32 %f13, %f128, %f60;
@%p16 bra $L__BB254_19;
bra.uni $L__BB254_18;

$L__BB254_19:
mov.u32 %r238, _ZZ13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r239, %r238, %r234;
ld.shared.f32 %f64, [%r239];
add.ftz.f32 %f65, %f126, %f64;
ld.shared.f32 %f66, [%r25+128];
add.ftz.f32 %f67, %f13, %f66;
ld.shared.f32 %f68, [%r239+128];
add.ftz.f32 %f69, %f65, %f68;
ld.shared.f32 %f70, [%r25+256];
add.ftz.f32 %f130, %f67, %f70;
ld.shared.f32 %f71, [%r239+256];
add.ftz.f32 %f126, %f69, %f71;
bra.uni $L__BB254_20;

$L__BB254_18:
ld.shared.f32 %f61, [%r25+128];
add.ftz.f32 %f62, %f13, %f61;
ld.shared.f32 %f63, [%r25+256];
add.ftz.f32 %f130, %f62, %f63;

$L__BB254_20:
mov.b32 %r240, %f130;
mov.u32 %r241, 31;
mov.u32 %r242, 16;
mov.u32 %r243, -1;
shfl.sync.bfly.b32 %r244|%p17, %r240, %r242, %r241, %r243;
mov.b32 %f72, %r244;
add.ftz.f32 %f73, %f130, %f72;
mov.b32 %r245, %f73;
mov.u32 %r246, 8;
shfl.sync.bfly.b32 %r247|%p18, %r245, %r246, %r241, %r243;
mov.b32 %f74, %r247;
add.ftz.f32 %f75, %f73, %f74;
mov.b32 %r248, %f75;
mov.u32 %r249, 4;
shfl.sync.bfly.b32 %r250|%p19, %r248, %r249, %r241, %r243;
mov.b32 %f76, %r250;
add.ftz.f32 %f77, %f75, %f76;
mov.b32 %r251, %f77;
mov.u32 %r252, 2;
shfl.sync.bfly.b32 %r253|%p20, %r251, %r252, %r241, %r243;
mov.b32 %f78, %r253;
add.ftz.f32 %f19, %f77, %f78;
mov.b32 %r254, %f19;
mov.u32 %r255, 1;
shfl.sync.bfly.b32 %r26|%p1, %r254, %r255, %r241, %r243;
setp.eq.s64 %p21, %rd12, 0;
@%p21 bra $L__BB254_22;

mov.b32 %r256, %f126;
shfl.sync.bfly.b32 %r260|%p22, %r256, %r242, %r241, %r243;
mov.b32 %f79, %r260;
add.ftz.f32 %f80, %f126, %f79;
mov.b32 %r261, %f80;
shfl.sync.bfly.b32 %r263|%p23, %r261, %r246, %r241, %r243;
mov.b32 %f81, %r263;
add.ftz.f32 %f82, %f80, %f81;
mov.b32 %r264, %f82;
shfl.sync.bfly.b32 %r266|%p24, %r264, %r249, %r241, %r243;
mov.b32 %f83, %r266;
add.ftz.f32 %f84, %f82, %f83;
mov.b32 %r267, %f84;
shfl.sync.bfly.b32 %r269|%p25, %r267, %r252, %r241, %r243;
mov.b32 %f85, %r269;
add.ftz.f32 %f86, %f84, %f85;
mov.b32 %r270, %f86;
shfl.sync.bfly.b32 %r272|%p26, %r270, %r255, %r241, %r243;
mov.b32 %f87, %r272;
add.ftz.f32 %f126, %f86, %f87;

$L__BB254_22:
setp.ne.s32 %p27, %r51, 0;
mov.b32 %f88, %r26;
add.ftz.f32 %f22, %f19, %f88;
@%p27 bra $L__BB254_35;

add.ftz.f32 %f132, %f123, %f22;
@%p21 bra $L__BB254_34;

add.ftz.f32 %f24, %f122, %f126;
setp.gt.s32 %p29, %r27, 2;
@%p29 bra $L__BB254_28;

setp.eq.s32 %p32, %r27, 1;
@%p32 bra $L__BB254_32;

setp.eq.s32 %p33, %r27, 2;
@%p33 bra $L__BB254_27;
bra.uni $L__BB254_33;

$L__BB254_27:
mul.ftz.f32 %f108, %f24, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f109, %f108;
add.ftz.f32 %f110, %f109, 0f3F800000;
div.approx.ftz.f32 %f111, %f24, %f110;
mul.ftz.f32 %f132, %f132, %f111;
bra.uni $L__BB254_34;

$L__BB254_28:
setp.eq.s32 %p30, %r27, 3;
@%p30 bra $L__BB254_31;

setp.ne.s32 %p31, %r27, 6;
@%p31 bra $L__BB254_33;

ld.param.f32 %f120, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
min.ftz.f32 %f89, %f24, %f120;
neg.ftz.f32 %f90, %f120;
min.ftz.f32 %f91, %f132, %f120;
max.ftz.f32 %f92, %f91, %f90;
mul.ftz.f32 %f93, %f89, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f94, %f93;
add.ftz.f32 %f95, %f94, 0f3F800000;
div.approx.ftz.f32 %f96, %f89, %f95;
mul.ftz.f32 %f132, %f92, %f96;
bra.uni $L__BB254_34;

$L__BB254_32:
mul.ftz.f32 %f112, %f24, 0f3F000000;
mul.ftz.f32 %f113, %f24, 0f3D372713;
fma.rn.ftz.f32 %f114, %f24, %f113, 0f3F800000;
mul.ftz.f32 %f115, %f24, 0f3F4C422A;
mul.ftz.f32 %f116, %f115, %f114;
tanh.approx.f32 %f117, %f116;
add.ftz.f32 %f118, %f117, 0f3F800000;
mul.ftz.f32 %f119, %f112, %f118;
mul.ftz.f32 %f132, %f132, %f119;
bra.uni $L__BB254_34;

$L__BB254_31:
mov.f32 %f97, 0f40E00000;
min.ftz.f32 %f98, %f24, %f97;
min.ftz.f32 %f99, %f132, %f97;
mov.f32 %f100, 0fC0E00000;
max.ftz.f32 %f101, %f99, %f100;
mul.ftz.f32 %f102, %f98, 0f3FD9DB23;
mul.ftz.f32 %f103, %f102, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f104, %f103;
add.ftz.f32 %f105, %f104, 0f3F800000;
div.approx.ftz.f32 %f106, %f98, %f105;
add.ftz.f32 %f107, %f101, 0f3F800000;
mul.ftz.f32 %f132, %f107, %f106;
bra.uni $L__BB254_34;

$L__BB254_33:
mul.ftz.f32 %f132, %f132, %f24;

$L__BB254_34:
ld.param.u32 %r281, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r280, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
cvta.to.global.u64 %rd83, %rd16;
mov.u32 %r275, %ctaid.z;
mov.u32 %r276, %ctaid.y;
mov.u32 %r277, %ctaid.x;
mad.lo.s32 %r278, %r276, %r281, %r277;
mad.lo.s32 %r279, %r275, %r280, %r278;
mul.wide.u32 %rd84, %r279, 4;
add.s64 %rd85, %rd83, %rd84;
st.global.f32 [%rd85], %f132;

$L__BB254_35:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<12>;
.reg .b16 %rs<9>;
.reg .f32 %f<34>;
.reg .b32 %r<157>;
.reg .b64 %rd<42>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[384];

ld.param.u64 %rd9, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd10, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd11, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd12, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r19, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
mov.u32 %r1, %tid.y;
shl.b32 %r36, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r36, %r2;
mov.u32 %r154, %ctaid.y;
setp.eq.s64 %p1, %rd11, 0;
@%p1 bra $L__BB255_2;

cvta.to.global.u64 %rd13, %rd11;
mul.wide.u32 %rd14, %r154, 4;
add.s64 %rd15, %rd13, %rd14;
ld.global.nc.u32 %r153, [%rd15];
mul.hi.u32 %r37, %r154, %r20;
add.s32 %r38, %r37, %r154;
shr.u32 %r39, %r38, %r21;
mul.lo.s32 %r40, %r39, %r22;
sub.s32 %r154, %r154, %r40;
bra.uni $L__BB255_3;

$L__BB255_2:
mul.hi.u32 %r41, %r154, %r24;
add.s32 %r42, %r41, %r154;
shr.u32 %r153, %r42, %r25;

$L__BB255_3:
mov.u32 %r10, %ctaid.z;
shr.s32 %r43, %r3, 31;
shr.u32 %r44, %r43, 29;
add.s32 %r45, %r3, %r44;
shr.s32 %r156, %r45, 3;
shr.u32 %r46, %r19, 8;
setp.ge.s32 %p2, %r156, %r46;
mov.f32 %f33, 0f00000000;
@%p2 bra $L__BB255_6;

and.b32 %r50, %r45, -8;
sub.s32 %r51, %r3, %r50;
shl.b32 %r52, %r51, 1;
cvt.s64.s32 %rd1, %r51;
mul.hi.u32 %r53, %r10, %r30;
add.s32 %r54, %r53, %r10;
shr.u32 %r55, %r54, %r31;
mad.lo.s32 %r56, %r55, %r33, %r156;
mad.lo.s32 %r57, %r153, %r27, %r56;
mov.u32 %r58, %ctaid.x;
mad.lo.s32 %r59, %r58, %r23, %r57;
mul.wide.s32 %rd16, %r59, 50;
cvta.to.global.u64 %rd17, %rd9;
add.s64 %rd41, %rd17, %rd16;
mul.wide.s32 %rd18, %r51, 2;
add.s64 %rd3, %rd18, 34;
shl.b32 %r155, %r156, 3;
mul.wide.s32 %rd19, %r52, 2;
or.b64 %rd4, %rd19, 2;
cvta.to.global.u64 %rd5, %rd10;
mul.lo.s32 %r60, %r154, %r28;
cvt.u64.u32 %rd20, %r60;
mul.lo.s32 %r61, %r10, %r34;
cvt.u64.u32 %rd21, %r61;
add.s64 %rd6, %rd20, %rd21;
mov.f32 %f33, 0f00000000;

$L__BB255_5:
cvt.u64.u32 %rd22, %r155;
add.s64 %rd23, %rd6, %rd22;
add.s64 %rd24, %rd41, %rd4;
ld.global.nc.u16 %rs2, [%rd24];
shr.u16 %rs3, %rs2, 8;
cvt.u32.u16 %r96, %rs3;
ld.global.nc.u16 %rs4, [%rd24+2];
shr.u16 %rs5, %rs4, 8;
cvt.u32.u16 %r97, %rs5;
add.s64 %rd25, %rd41, %rd3;
ld.global.nc.u16 %rs6, [%rd25];
cvt.u32.u16 %r98, %rs6;
add.s64 %rd26, %rd23, %rd1;
mul.lo.s64 %rd27, %rd26, 36;
and.b16 %rs7, %rs2, 255;
cvt.u32.u16 %r99, %rs7;
mul.wide.u16 %r100, %rs6, 256;
and.b32 %r101, %r100, 1792;
or.b32 %r102, %r101, %r99;
mul.wide.u32 %rd28, %r102, 4;
mov.u64 %rd29, iq1s_grid_gpu;
add.s64 %rd30, %rd29, %rd28;
ld.global.nc.u32 %r103, [%rd30];
and.b32 %r63, %r103, 252645135;
shr.u32 %r104, %r103, 4;
and.b32 %r67, %r104, 252645135;
add.s64 %rd31, %rd5, %rd27;
ld.global.nc.u32 %r64, [%rd31+4];
ld.global.nc.u32 %r68, [%rd31+8];
mov.u32 %r65, 0;

dp4a.s32.s32 %r62, %r63, %r64, %r65;


dp4a.s32.s32 %r66, %r67, %r68, %r62;

mul.wide.u16 %r105, %rs6, 32;
and.b32 %r106, %r105, 1792;
or.b32 %r107, %r106, %r96;
mul.wide.u32 %rd32, %r107, 4;
add.s64 %rd33, %rd29, %rd32;
ld.global.nc.u32 %r108, [%rd33];
and.b32 %r71, %r108, 252645135;
shr.u32 %r109, %r108, 4;
and.b32 %r75, %r109, 252645135;
ld.global.nc.u32 %r72, [%rd31+12];
ld.global.nc.u32 %r76, [%rd31+16];

dp4a.s32.s32 %r70, %r71, %r72, %r66;


dp4a.s32.s32 %r74, %r75, %r76, %r70;

and.b16 %rs8, %rs4, 255;
cvt.u32.u16 %r110, %rs8;
mul.wide.u16 %r111, %rs6, 4;
and.b32 %r112, %r111, 1792;
or.b32 %r113, %r112, %r110;
mul.wide.u32 %rd34, %r113, 4;
add.s64 %rd35, %rd29, %rd34;
ld.global.nc.u32 %r114, [%rd35];
and.b32 %r79, %r114, 252645135;
shr.u32 %r115, %r114, 4;
and.b32 %r83, %r115, 252645135;
ld.global.nc.u32 %r80, [%rd31+20];
ld.global.nc.u32 %r84, [%rd31+24];

dp4a.s32.s32 %r78, %r79, %r80, %r74;


dp4a.s32.s32 %r82, %r83, %r84, %r78;

shr.u32 %r116, %r98, 1;
and.b32 %r117, %r116, 1792;
or.b32 %r118, %r117, %r97;
mul.wide.u32 %rd36, %r118, 4;
add.s64 %rd37, %rd29, %rd36;
ld.global.nc.u32 %r119, [%rd37];
and.b32 %r87, %r119, 252645135;
shr.u32 %r120, %r119, 4;
and.b32 %r91, %r120, 252645135;
ld.global.nc.u32 %r88, [%rd31+28];
ld.global.nc.u32 %r92, [%rd31+32];

dp4a.s32.s32 %r86, %r87, %r88, %r82;


dp4a.s32.s32 %r90, %r91, %r92, %r86;

ld.global.nc.u16 %rs1, [%rd41];

{ cvt.f32.f16 %f7, %rs1;}


shr.u32 %r121, %r98, 11;
and.b32 %r122, %r121, 14;
or.b32 %r123, %r122, 1;
cvt.rn.f32.s32 %f10, %r123;
mul.ftz.f32 %f11, %f7, %f10;
and.b32 %r124, %r98, 32768;
cvt.rn.f32.s32 %f12, %r124;
fma.rn.ftz.f32 %f13, %f12, 0fB7000000, 0fBF600000;
ld.global.nc.u32 %r94, [%rd31];

{.reg .f16 low,high;
mov.b32 {low,high},%r94;
cvt.f32.f16 %f8, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r94;
cvt.f32.f16 %f9, high;}


cvt.rn.f32.s32 %f14, %r90;
mul.ftz.f32 %f15, %f13, %f9;
fma.rn.ftz.f32 %f16, %f8, %f14, %f15;
fma.rn.ftz.f32 %f33, %f11, %f16, %f33;
add.s64 %rd41, %rd41, 800;
add.s32 %r155, %r155, 128;
add.s32 %r156, %r156, 16;
setp.lt.s32 %p3, %r156, %r46;
@%p3 bra $L__BB255_5;

$L__BB255_6:
setp.eq.s32 %p4, %r1, 0;
@%p4 bra $L__BB255_8;

shl.b32 %r125, %r1, 7;
mov.u32 %r126, _ZZ13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r127, %r126, %r125;
shl.b32 %r128, %r2, 2;
add.s32 %r129, %r127, %r128;
st.shared.f32 [%r129+-128], %f33;

$L__BB255_8:
setp.ne.s32 %p5, %r1, 0;
bar.sync 0;
@%p5 bra $L__BB255_11;

mov.u32 %r130, %ctaid.y;
mov.u32 %r131, %ctaid.x;
mad.lo.s32 %r132, %r130, %r29, %r131;
mad.lo.s32 %r18, %r10, %r35, %r132;
shl.b32 %r133, %r2, 2;
mov.u32 %r134, _ZZ13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r135, %r134, %r133;
mov.u32 %r136, 2;
ld.shared.f32 %f17, [%r135];
add.ftz.f32 %f18, %f33, %f17;
ld.shared.f32 %f19, [%r135+128];
add.ftz.f32 %f20, %f18, %f19;
ld.shared.f32 %f21, [%r135+256];
add.ftz.f32 %f22, %f20, %f21;
mov.b32 %r137, %f22;
mov.u32 %r138, 31;
mov.u32 %r139, 16;
mov.u32 %r140, -1;
shfl.sync.bfly.b32 %r141|%p6, %r137, %r139, %r138, %r140;
mov.b32 %f23, %r141;
add.ftz.f32 %f24, %f22, %f23;
mov.b32 %r142, %f24;
mov.u32 %r143, 8;
shfl.sync.bfly.b32 %r144|%p7, %r142, %r143, %r138, %r140;
mov.b32 %f25, %r144;
add.ftz.f32 %f26, %f24, %f25;
mov.b32 %r145, %f26;
mov.u32 %r146, 4;
shfl.sync.bfly.b32 %r147|%p8, %r145, %r146, %r138, %r140;
mov.b32 %f27, %r147;
add.ftz.f32 %f28, %f26, %f27;
mov.b32 %r148, %f28;
shfl.sync.bfly.b32 %r149|%p9, %r148, %r136, %r138, %r140;
mov.b32 %f29, %r149;
add.ftz.f32 %f30, %f28, %f29;
mov.b32 %r150, %f30;
mov.u32 %r151, 1;
shfl.sync.bfly.b32 %r152|%p10, %r150, %r151, %r138, %r140;
mov.b32 %f31, %r152;
add.ftz.f32 %f4, %f30, %f31;
setp.ne.s32 %p11, %r2, 0;
@%p11 bra $L__BB255_11;

cvta.to.global.u64 %rd38, %rd12;
mul.wide.u32 %rd39, %r18, 4;
add.s64 %rd40, %rd38, %rd39;
st.global.f32 [%rd40], %f4;

$L__BB255_11:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type19ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type19ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type19ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type19ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type19ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type19ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type19ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type19ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type19ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<35>;
.reg .b16 %rs<19>;
.reg .f32 %f<130>;
.reg .b32 %r<343>;
.reg .b64 %rd<70>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type19ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type19ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type19ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type19ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r17, [_Z13mul_mat_vec_qIL9ggml_type19ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r18, [_Z13mul_mat_vec_qIL9ggml_type19ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r19, [_Z13mul_mat_vec_qIL9ggml_type19ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type19ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type19ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type19ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type19ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type19ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type19ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type19ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type19ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type19ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type19ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type19ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
cvta.to.global.u64 %rd1, %rd17;
mov.u32 %r1, %tid.y;
shl.b32 %r33, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r33, %r2;
mov.u32 %r34, %ctaid.x;
shl.b32 %r4, %r34, 1;
mov.u32 %r5, %ctaid.y;
shr.s32 %r35, %r3, 31;
shr.u32 %r36, %r35, 29;
add.s32 %r37, %r3, %r36;
shr.s32 %r342, %r37, 3;
shr.u32 %r38, %r17, 8;
setp.ge.s32 %p3, %r342, %r38;
mov.f32 %f126, 0f00000000;
mov.f32 %f127, %f126;
mov.f32 %f128, %f126;
mov.f32 %f129, %f126;
@%p3 bra $L__BB256_3;

mov.u32 %r39, %ctaid.z;
mul.hi.u32 %r40, %r39, %r27;
add.s32 %r41, %r40, %r39;
shr.u32 %r42, %r41, %r28;
mul.lo.s32 %r43, %r42, %r30;
mul.hi.u32 %r44, %r5, %r21;
add.s32 %r45, %r44, %r5;
shr.u32 %r46, %r45, %r22;
mul.lo.s32 %r47, %r46, %r24;
cvta.to.global.u64 %rd69, %rd15;
and.b32 %r51, %r37, -8;
sub.s32 %r52, %r3, %r51;
shl.b32 %r53, %r52, 1;
cvt.s64.s32 %rd3, %r52;
shl.b32 %r340, %r342, 3;
add.s32 %r341, %r19, %r340;
add.s32 %r54, %r342, %r43;
add.s32 %r55, %r54, %r47;
add.s32 %r56, %r4, 1;
mad.lo.s32 %r57, %r18, %r56, %r55;
mul.wide.s32 %rd4, %r57, 50;
mul.wide.s32 %rd18, %r52, 2;
add.s64 %rd19, %rd4, %rd18;
add.s64 %rd5, %rd19, 34;
mad.lo.s32 %r58, %r4, %r18, %r342;
add.s32 %r59, %r58, %r43;
add.s32 %r60, %r59, %r47;
mul.wide.s32 %rd6, %r60, 50;
mul.wide.s32 %rd20, %r53, 2;
add.s64 %rd21, %rd6, %rd20;
add.s64 %rd7, %rd21, 4;
add.s64 %rd22, %rd4, %rd20;
add.s64 %rd8, %rd22, 2;
add.s64 %rd23, %rd6, %rd18;
add.s64 %rd9, %rd23, 34;
cvta.to.global.u64 %rd10, %rd16;
mov.f32 %f126, 0f00000000;

$L__BB256_2:
mul.lo.s32 %r198, %r39, %r31;
cvt.u64.u32 %rd24, %r198;
mul.lo.s32 %r200, %r5, %r25;
cvt.u64.u32 %rd25, %r200;
add.s64 %rd26, %rd24, %rd25;
cvt.u64.u32 %rd27, %r340;
add.s64 %rd28, %rd26, %rd27;
add.s64 %rd29, %rd28, %rd3;
mul.lo.s64 %rd30, %rd29, 36;
add.s64 %rd31, %rd69, %rd7;
ld.global.nc.u16 %rs5, [%rd31+-2];
shr.u16 %rs6, %rs5, 8;
cvt.u32.u16 %r201, %rs6;
ld.global.nc.u16 %rs7, [%rd31];
shr.u16 %rs8, %rs7, 8;
cvt.u32.u16 %r202, %rs8;
add.s64 %rd32, %rd69, %rd9;
ld.global.nc.u16 %rs9, [%rd32];
cvt.u32.u16 %r203, %rs9;
and.b16 %rs10, %rs5, 255;
cvt.u32.u16 %r204, %rs10;
mul.wide.u16 %r205, %rs9, 256;
and.b32 %r206, %r205, 1792;
or.b32 %r207, %r206, %r204;
mul.wide.u32 %rd33, %r207, 4;
mov.u64 %rd34, iq1s_grid_gpu;
add.s64 %rd35, %rd34, %rd33;
ld.global.nc.u32 %r208, [%rd35];
and.b32 %r130, %r208, 252645135;
shr.u32 %r209, %r208, 4;
and.b32 %r134, %r209, 252645135;
add.s64 %rd36, %rd10, %rd30;
ld.global.nc.u32 %r63, [%rd36+4];
ld.global.nc.u32 %r67, [%rd36+8];
mov.u32 %r166, 0;

dp4a.s32.s32 %r61, %r130, %r63, %r166;


dp4a.s32.s32 %r65, %r134, %r67, %r61;

mul.wide.u16 %r210, %rs9, 32;
and.b32 %r211, %r210, 1792;
or.b32 %r212, %r211, %r201;
mul.wide.u32 %rd37, %r212, 4;
add.s64 %rd38, %rd34, %rd37;
ld.global.nc.u32 %r213, [%rd38];
and.b32 %r138, %r213, 252645135;
shr.u32 %r214, %r213, 4;
and.b32 %r142, %r214, 252645135;
ld.global.nc.u32 %r71, [%rd36+12];
ld.global.nc.u32 %r75, [%rd36+16];

dp4a.s32.s32 %r69, %r138, %r71, %r65;


dp4a.s32.s32 %r73, %r142, %r75, %r69;

and.b16 %rs11, %rs7, 255;
cvt.u32.u16 %r215, %rs11;
mul.wide.u16 %r216, %rs9, 4;
and.b32 %r217, %r216, 1792;
or.b32 %r218, %r217, %r215;
mul.wide.u32 %rd39, %r218, 4;
add.s64 %rd40, %rd34, %rd39;
ld.global.nc.u32 %r219, [%rd40];
and.b32 %r146, %r219, 252645135;
shr.u32 %r220, %r219, 4;
and.b32 %r150, %r220, 252645135;
ld.global.nc.u32 %r79, [%rd36+20];
ld.global.nc.u32 %r83, [%rd36+24];

dp4a.s32.s32 %r77, %r146, %r79, %r73;


dp4a.s32.s32 %r81, %r150, %r83, %r77;

shr.u32 %r221, %r203, 1;
and.b32 %r222, %r221, 1792;
or.b32 %r223, %r222, %r202;
mul.wide.u32 %rd41, %r223, 4;
add.s64 %rd42, %rd34, %rd41;
ld.global.nc.u32 %r224, [%rd42];
and.b32 %r154, %r224, 252645135;
shr.u32 %r225, %r224, 4;
and.b32 %r158, %r225, 252645135;
ld.global.nc.u32 %r87, [%rd36+28];
ld.global.nc.u32 %r91, [%rd36+32];

dp4a.s32.s32 %r85, %r154, %r87, %r81;


dp4a.s32.s32 %r89, %r158, %r91, %r85;

add.s64 %rd43, %rd69, %rd6;
ld.global.nc.u16 %rs1, [%rd43];

{ cvt.f32.f16 %f25, %rs1;}


shr.u32 %r226, %r203, 11;
and.b32 %r227, %r226, 14;
or.b32 %r228, %r227, 1;
cvt.rn.f32.s32 %f37, %r228;
mul.ftz.f32 %f38, %f25, %f37;
and.b32 %r229, %r203, 32768;
cvt.rn.f32.s32 %f39, %r229;
mul.ftz.f32 %f40, %f39, 0f37000000;
mov.f32 %f41, 0fBF600000;
sub.ftz.f32 %f42, %f41, %f40;
ld.global.nc.u32 %r93, [%rd36];

{.reg .f16 low,high;
mov.b32 {low,high},%r93;
cvt.f32.f16 %f26, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r93;
cvt.f32.f16 %f27, high;}


cvt.rn.f32.s32 %f43, %r89;
mul.ftz.f32 %f44, %f42, %f27;
fma.rn.ftz.f32 %f45, %f26, %f43, %f44;
fma.rn.ftz.f32 %f129, %f38, %f45, %f129;
add.s64 %rd44, %rd69, %rd8;
ld.global.nc.u16 %rs12, [%rd44];
shr.u16 %rs13, %rs12, 8;
cvt.u32.u16 %r230, %rs13;
ld.global.nc.u16 %rs14, [%rd44+2];
shr.u16 %rs15, %rs14, 8;
cvt.u32.u16 %r231, %rs15;
add.s64 %rd45, %rd69, %rd5;
ld.global.nc.u16 %rs16, [%rd45];
cvt.u32.u16 %r232, %rs16;
and.b16 %rs17, %rs12, 255;
cvt.u32.u16 %r233, %rs17;
mul.wide.u16 %r234, %rs16, 256;
and.b32 %r235, %r234, 1792;
or.b32 %r236, %r235, %r233;
mul.wide.u32 %rd46, %r236, 4;
add.s64 %rd47, %rd34, %rd46;
ld.global.nc.u32 %r237, [%rd47];
and.b32 %r164, %r237, 252645135;
shr.u32 %r238, %r237, 4;
and.b32 %r168, %r238, 252645135;

dp4a.s32.s32 %r95, %r164, %r63, %r166;


dp4a.s32.s32 %r99, %r168, %r67, %r95;

mul.wide.u16 %r239, %rs16, 32;
and.b32 %r240, %r239, 1792;
or.b32 %r241, %r240, %r230;
mul.wide.u32 %rd48, %r241, 4;
add.s64 %rd49, %rd34, %rd48;
ld.global.nc.u32 %r242, [%rd49];
and.b32 %r172, %r242, 252645135;
shr.u32 %r243, %r242, 4;
and.b32 %r176, %r243, 252645135;

dp4a.s32.s32 %r103, %r172, %r71, %r99;


dp4a.s32.s32 %r107, %r176, %r75, %r103;

and.b16 %rs18, %rs14, 255;
cvt.u32.u16 %r244, %rs18;
mul.wide.u16 %r245, %rs16, 4;
and.b32 %r246, %r245, 1792;
or.b32 %r247, %r246, %r244;
mul.wide.u32 %rd50, %r247, 4;
add.s64 %rd51, %rd34, %rd50;
ld.global.nc.u32 %r248, [%rd51];
and.b32 %r180, %r248, 252645135;
shr.u32 %r249, %r248, 4;
and.b32 %r184, %r249, 252645135;

dp4a.s32.s32 %r111, %r180, %r79, %r107;


dp4a.s32.s32 %r115, %r184, %r83, %r111;

shr.u32 %r250, %r232, 1;
and.b32 %r251, %r250, 1792;
or.b32 %r252, %r251, %r231;
mul.wide.u32 %rd52, %r252, 4;
add.s64 %rd53, %rd34, %rd52;
ld.global.nc.u32 %r253, [%rd53];
and.b32 %r188, %r253, 252645135;
shr.u32 %r254, %r253, 4;
and.b32 %r192, %r254, 252645135;

dp4a.s32.s32 %r119, %r188, %r87, %r115;


dp4a.s32.s32 %r123, %r192, %r91, %r119;

add.s64 %rd54, %rd69, %rd4;
ld.global.nc.u16 %rs2, [%rd54];

{ cvt.f32.f16 %f28, %rs2;}


shr.u32 %r255, %r232, 11;
and.b32 %r256, %r255, 14;
or.b32 %r257, %r256, 1;
cvt.rn.f32.s32 %f46, %r257;
mul.ftz.f32 %f47, %f28, %f46;
and.b32 %r258, %r232, 32768;
cvt.rn.f32.s32 %f48, %r258;
mul.ftz.f32 %f49, %f48, 0f37000000;
sub.ftz.f32 %f50, %f41, %f49;

{.reg .f16 low,high;
mov.b32 {low,high},%r93;
cvt.f32.f16 %f29, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r93;
cvt.f32.f16 %f30, high;}


cvt.rn.f32.s32 %f51, %r123;
mul.ftz.f32 %f52, %f50, %f30;
fma.rn.ftz.f32 %f53, %f29, %f51, %f52;
fma.rn.ftz.f32 %f128, %f47, %f53, %f128;
cvt.u64.u32 %rd55, %r341;
add.s64 %rd56, %rd26, %rd55;
add.s64 %rd57, %rd56, %rd3;
mul.lo.s64 %rd58, %rd57, 36;
add.s64 %rd59, %rd10, %rd58;
ld.global.nc.u32 %r131, [%rd59+4];
ld.global.nc.u32 %r135, [%rd59+8];

dp4a.s32.s32 %r129, %r130, %r131, %r166;


dp4a.s32.s32 %r133, %r134, %r135, %r129;

ld.global.nc.u32 %r139, [%rd59+12];
ld.global.nc.u32 %r143, [%rd59+16];

dp4a.s32.s32 %r137, %r138, %r139, %r133;


dp4a.s32.s32 %r141, %r142, %r143, %r137;

ld.global.nc.u32 %r147, [%rd59+20];
ld.global.nc.u32 %r151, [%rd59+24];

dp4a.s32.s32 %r145, %r146, %r147, %r141;


dp4a.s32.s32 %r149, %r150, %r151, %r145;

ld.global.nc.u32 %r155, [%rd59+28];
ld.global.nc.u32 %r159, [%rd59+32];

dp4a.s32.s32 %r153, %r154, %r155, %r149;


dp4a.s32.s32 %r157, %r158, %r159, %r153;


{ cvt.f32.f16 %f31, %rs1;}


mul.ftz.f32 %f54, %f31, %f37;
ld.global.nc.u32 %r161, [%rd59];

{.reg .f16 low,high;
mov.b32 {low,high},%r161;
cvt.f32.f16 %f32, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r161;
cvt.f32.f16 %f33, high;}


cvt.rn.f32.s32 %f55, %r157;
mul.ftz.f32 %f56, %f42, %f33;
fma.rn.ftz.f32 %f57, %f32, %f55, %f56;
fma.rn.ftz.f32 %f127, %f54, %f57, %f127;

dp4a.s32.s32 %r163, %r164, %r131, %r166;


dp4a.s32.s32 %r167, %r168, %r135, %r163;


dp4a.s32.s32 %r171, %r172, %r139, %r167;


dp4a.s32.s32 %r175, %r176, %r143, %r171;


dp4a.s32.s32 %r179, %r180, %r147, %r175;


dp4a.s32.s32 %r183, %r184, %r151, %r179;


dp4a.s32.s32 %r187, %r188, %r155, %r183;


dp4a.s32.s32 %r191, %r192, %r159, %r187;


{ cvt.f32.f16 %f34, %rs2;}


mul.ftz.f32 %f58, %f34, %f46;

{.reg .f16 low,high;
mov.b32 {low,high},%r161;
cvt.f32.f16 %f35, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r161;
cvt.f32.f16 %f36, high;}


cvt.rn.f32.s32 %f59, %r191;
mul.ftz.f32 %f60, %f50, %f36;
fma.rn.ftz.f32 %f61, %f35, %f59, %f60;
fma.rn.ftz.f32 %f126, %f58, %f61, %f126;
add.s32 %r341, %r341, 128;
add.s32 %r340, %r340, 128;
add.s64 %rd69, %rd69, 800;
add.s32 %r342, %r342, 16;
setp.lt.s32 %p4, %r342, %r38;
@%p4 bra $L__BB256_2;

$L__BB256_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB256_5;

shl.b32 %r260, %r1, 9;
mov.u32 %r261, _ZZ13mul_mat_vec_qIL9ggml_type19ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r262, %r261, %r260;
shl.b32 %r263, %r2, 2;
add.s32 %r264, %r262, %r263;
st.shared.f32 [%r264+-512], %f129;
st.shared.f32 [%r264+-384], %f128;
st.shared.f32 [%r264+-256], %f127;
st.shared.f32 [%r264+-128], %f126;

$L__BB256_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB256_14;

mov.u32 %r265, %ctaid.z;
mad.lo.s32 %r266, %r5, %r26, %r4;
mad.lo.s32 %r267, %r265, %r32, %r266;
cvt.u64.u32 %rd13, %r267;
shl.b32 %r268, %r2, 2;
mov.u32 %r269, _ZZ13mul_mat_vec_qIL9ggml_type19ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r15, %r269, %r268;
mov.u32 %r270, 2;
ld.shared.f32 %f62, [%r15];
add.ftz.f32 %f63, %f62, %f129;
ld.shared.f32 %f64, [%r15+512];
add.ftz.f32 %f65, %f64, %f63;
ld.shared.f32 %f66, [%r15+1024];
add.ftz.f32 %f67, %f66, %f65;
mov.b32 %r271, %f67;
mov.u32 %r272, 31;
mov.u32 %r273, 16;
mov.u32 %r274, -1;
shfl.sync.bfly.b32 %r275|%p7, %r271, %r273, %r272, %r274;
mov.b32 %f68, %r275;
add.ftz.f32 %f69, %f67, %f68;
mov.b32 %r276, %f69;
mov.u32 %r277, 8;
shfl.sync.bfly.b32 %r278|%p8, %r276, %r277, %r272, %r274;
mov.b32 %f70, %r278;
add.ftz.f32 %f71, %f69, %f70;
mov.b32 %r279, %f71;
mov.u32 %r280, 4;
shfl.sync.bfly.b32 %r281|%p9, %r279, %r280, %r272, %r274;
mov.b32 %f72, %r281;
add.ftz.f32 %f73, %f71, %f72;
mov.b32 %r282, %f73;
shfl.sync.bfly.b32 %r283|%p10, %r282, %r270, %r272, %r274;
mov.b32 %f74, %r283;
add.ftz.f32 %f75, %f73, %f74;
mov.b32 %r284, %f75;
mov.u32 %r285, 1;
shfl.sync.bfly.b32 %r286|%p11, %r284, %r285, %r272, %r274;
mov.b32 %f76, %r286;
add.ftz.f32 %f13, %f75, %f76;
setp.lt.u32 %p12, %r4, %r20;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
mul.wide.u32 %rd60, %r267, 4;
add.s64 %rd14, %rd1, %rd60;
not.pred %p14, %p1;
@%p14 bra $L__BB256_8;

st.global.f32 [%rd14], %f13;

$L__BB256_8:
ld.shared.f32 %f77, [%r15+128];
add.ftz.f32 %f78, %f77, %f128;
ld.shared.f32 %f79, [%r15+640];
add.ftz.f32 %f80, %f79, %f78;
ld.shared.f32 %f81, [%r15+1152];
add.ftz.f32 %f82, %f81, %f80;
mov.b32 %r287, %f82;
shfl.sync.bfly.b32 %r291|%p15, %r287, %r273, %r272, %r274;
mov.b32 %f83, %r291;
add.ftz.f32 %f84, %f82, %f83;
mov.b32 %r292, %f84;
shfl.sync.bfly.b32 %r294|%p16, %r292, %r277, %r272, %r274;
mov.b32 %f85, %r294;
add.ftz.f32 %f86, %f84, %f85;
mov.b32 %r295, %f86;
shfl.sync.bfly.b32 %r297|%p17, %r295, %r280, %r272, %r274;
mov.b32 %f87, %r297;
add.ftz.f32 %f88, %f86, %f87;
mov.b32 %r298, %f88;
shfl.sync.bfly.b32 %r300|%p18, %r298, %r270, %r272, %r274;
mov.b32 %f89, %r300;
add.ftz.f32 %f90, %f88, %f89;
mov.b32 %r301, %f90;
shfl.sync.bfly.b32 %r303|%p19, %r301, %r285, %r272, %r274;
mov.b32 %f91, %r303;
add.ftz.f32 %f14, %f90, %f91;
add.s32 %r304, %r4, 1;
setp.lt.u32 %p20, %r304, %r20;
setp.eq.s32 %p21, %r2, 1;
and.pred %p2, %p21, %p20;
not.pred %p22, %p2;
@%p22 bra $L__BB256_10;

st.global.f32 [%rd14+4], %f14;

$L__BB256_10:
ld.shared.f32 %f92, [%r15+256];
add.ftz.f32 %f93, %f92, %f127;
ld.shared.f32 %f94, [%r15+768];
add.ftz.f32 %f95, %f94, %f93;
ld.shared.f32 %f96, [%r15+1280];
add.ftz.f32 %f97, %f96, %f95;
mov.b32 %r305, %f97;
mov.u32 %r306, 31;
mov.u32 %r307, 16;
mov.u32 %r308, -1;
shfl.sync.bfly.b32 %r309|%p23, %r305, %r307, %r306, %r308;
mov.b32 %f98, %r309;
add.ftz.f32 %f99, %f97, %f98;
mov.b32 %r310, %f99;
mov.u32 %r311, 8;
shfl.sync.bfly.b32 %r312|%p24, %r310, %r311, %r306, %r308;
mov.b32 %f100, %r312;
add.ftz.f32 %f101, %f99, %f100;
mov.b32 %r313, %f101;
mov.u32 %r314, 4;
shfl.sync.bfly.b32 %r315|%p25, %r313, %r314, %r306, %r308;
mov.b32 %f102, %r315;
add.ftz.f32 %f103, %f101, %f102;
mov.b32 %r316, %f103;
mov.u32 %r317, 2;
shfl.sync.bfly.b32 %r318|%p26, %r316, %r317, %r306, %r308;
mov.b32 %f104, %r318;
add.ftz.f32 %f105, %f103, %f104;
mov.b32 %r319, %f105;
mov.u32 %r320, 1;
shfl.sync.bfly.b32 %r321|%p27, %r319, %r320, %r306, %r308;
mov.b32 %f106, %r321;
add.ftz.f32 %f15, %f105, %f106;
@%p14 bra $L__BB256_12;

cvt.u64.u32 %rd61, %r20;
add.s64 %rd62, %rd61, %rd13;
shl.b64 %rd63, %rd62, 2;
add.s64 %rd64, %rd1, %rd63;
st.global.f32 [%rd64], %f15;

$L__BB256_12:
ld.shared.f32 %f107, [%r15+384];
add.ftz.f32 %f108, %f107, %f126;
ld.shared.f32 %f109, [%r15+896];
add.ftz.f32 %f110, %f109, %f108;
ld.shared.f32 %f111, [%r15+1408];
add.ftz.f32 %f112, %f111, %f110;
mov.b32 %r322, %f112;
shfl.sync.bfly.b32 %r326|%p29, %r322, %r307, %r306, %r308;
mov.b32 %f113, %r326;
add.ftz.f32 %f114, %f112, %f113;
mov.b32 %r327, %f114;
shfl.sync.bfly.b32 %r329|%p30, %r327, %r311, %r306, %r308;
mov.b32 %f115, %r329;
add.ftz.f32 %f116, %f114, %f115;
mov.b32 %r330, %f116;
shfl.sync.bfly.b32 %r332|%p31, %r330, %r314, %r306, %r308;
mov.b32 %f117, %r332;
add.ftz.f32 %f118, %f116, %f117;
mov.b32 %r333, %f118;
shfl.sync.bfly.b32 %r335|%p32, %r333, %r317, %r306, %r308;
mov.b32 %f119, %r335;
add.ftz.f32 %f120, %f118, %f119;
mov.b32 %r336, %f120;
shfl.sync.bfly.b32 %r338|%p33, %r336, %r320, %r306, %r308;
mov.b32 %f121, %r338;
add.ftz.f32 %f16, %f120, %f121;
@%p22 bra $L__BB256_14;

add.s32 %r339, %r20, 1;
cvt.u64.u32 %rd65, %r339;
add.s64 %rd66, %rd65, %rd13;
shl.b64 %rd67, %rd66, 2;
add.s64 %rd68, %rd1, %rd67;
st.global.f32 [%rd68], %f16;

$L__BB256_14:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type19ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type19ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type19ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type19ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type19ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type19ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type19ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type19ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type19ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<47>;
.reg .b16 %rs<21>;
.reg .f32 %f<190>;
.reg .b32 %r<450>;
.reg .b64 %rd<79>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type19ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[2304];

ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type19ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type19ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type19ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type19ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type19ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type19ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type19ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type19ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type19ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type19ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type19ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type19ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type19ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type19ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type19ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type19ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type19ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
cvta.to.global.u64 %rd1, %rd18;
mov.u32 %r1, %tid.y;
shl.b32 %r36, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r36, %r2;
mov.u32 %r37, %ctaid.x;
shl.b32 %r4, %r37, 1;
mov.u32 %r5, %ctaid.y;
shr.s32 %r38, %r3, 31;
shr.u32 %r39, %r38, 29;
add.s32 %r40, %r3, %r39;
shr.s32 %r449, %r40, 3;
shr.u32 %r41, %r20, 8;
setp.ge.s32 %p3, %r449, %r41;
mov.f32 %f184, 0f00000000;
mov.f32 %f185, %f184;
mov.f32 %f186, %f184;
mov.f32 %f187, %f184;
mov.f32 %f188, %f184;
mov.f32 %f189, %f184;
@%p3 bra $L__BB257_3;

mov.u32 %r42, %ctaid.z;
mul.hi.u32 %r43, %r42, %r30;
add.s32 %r44, %r43, %r42;
shr.u32 %r45, %r44, %r31;
mul.lo.s32 %r46, %r45, %r33;
mul.hi.u32 %r47, %r5, %r24;
add.s32 %r48, %r47, %r5;
shr.u32 %r49, %r48, %r25;
mul.lo.s32 %r50, %r49, %r27;
cvta.to.global.u64 %rd78, %rd16;
and.b32 %r54, %r40, -8;
sub.s32 %r55, %r3, %r54;
shl.b32 %r56, %r55, 1;
cvt.s64.s32 %rd3, %r55;
shl.b32 %r57, %r22, 1;
shl.b32 %r446, %r449, 3;
add.s32 %r448, %r446, %r57;
add.s32 %r447, %r22, %r446;
add.s32 %r58, %r449, %r46;
add.s32 %r59, %r58, %r50;
add.s32 %r60, %r4, 1;
mad.lo.s32 %r61, %r21, %r60, %r59;
mul.wide.s32 %rd4, %r61, 50;
mul.wide.s32 %rd19, %r55, 2;
add.s64 %rd20, %rd4, %rd19;
add.s64 %rd5, %rd20, 34;
mad.lo.s32 %r62, %r4, %r21, %r449;
add.s32 %r63, %r62, %r46;
add.s32 %r64, %r63, %r50;
mul.wide.s32 %rd6, %r64, 50;
mul.wide.s32 %rd21, %r56, 2;
add.s64 %rd22, %rd6, %rd21;
add.s64 %rd7, %rd22, 4;
add.s64 %rd23, %rd4, %rd21;
add.s64 %rd8, %rd23, 2;
add.s64 %rd24, %rd6, %rd19;
add.s64 %rd9, %rd24, 34;
cvta.to.global.u64 %rd10, %rd17;
mov.f32 %f184, 0f00000000;

$L__BB257_2:
mul.lo.s32 %r270, %r42, %r34;
cvt.u64.u32 %rd25, %r270;
mul.lo.s32 %r272, %r5, %r28;
cvt.u64.u32 %rd26, %r272;
add.s64 %rd27, %rd25, %rd26;
cvt.u64.u32 %rd28, %r446;
add.s64 %rd29, %rd27, %rd28;
add.s64 %rd30, %rd29, %rd3;
mul.lo.s64 %rd31, %rd30, 36;
add.s64 %rd32, %rd78, %rd7;
ld.global.nc.u16 %rs7, [%rd32+-2];
shr.u16 %rs8, %rs7, 8;
cvt.u32.u16 %r273, %rs8;
ld.global.nc.u16 %rs9, [%rd32];
shr.u16 %rs10, %rs9, 8;
cvt.u32.u16 %r274, %rs10;
add.s64 %rd33, %rd78, %rd9;
ld.global.nc.u16 %rs11, [%rd33];
cvt.u32.u16 %r275, %rs11;
and.b16 %rs12, %rs7, 255;
cvt.u32.u16 %r276, %rs12;
mul.wide.u16 %r277, %rs11, 256;
and.b32 %r278, %r277, 1792;
or.b32 %r279, %r278, %r276;
mul.wide.u32 %rd34, %r279, 4;
mov.u64 %rd35, iq1s_grid_gpu;
add.s64 %rd36, %rd35, %rd34;
ld.global.nc.u32 %r280, [%rd36];
and.b32 %r202, %r280, 252645135;
shr.u32 %r281, %r280, 4;
and.b32 %r206, %r281, 252645135;
add.s64 %rd37, %rd10, %rd31;
ld.global.nc.u32 %r67, [%rd37+4];
ld.global.nc.u32 %r71, [%rd37+8];
mov.u32 %r238, 0;

dp4a.s32.s32 %r65, %r202, %r67, %r238;


dp4a.s32.s32 %r69, %r206, %r71, %r65;

mul.wide.u16 %r282, %rs11, 32;
and.b32 %r283, %r282, 1792;
or.b32 %r284, %r283, %r273;
mul.wide.u32 %rd38, %r284, 4;
add.s64 %rd39, %rd35, %rd38;
ld.global.nc.u32 %r285, [%rd39];
and.b32 %r210, %r285, 252645135;
shr.u32 %r286, %r285, 4;
and.b32 %r214, %r286, 252645135;
ld.global.nc.u32 %r75, [%rd37+12];
ld.global.nc.u32 %r79, [%rd37+16];

dp4a.s32.s32 %r73, %r210, %r75, %r69;


dp4a.s32.s32 %r77, %r214, %r79, %r73;

and.b16 %rs13, %rs9, 255;
cvt.u32.u16 %r287, %rs13;
mul.wide.u16 %r288, %rs11, 4;
and.b32 %r289, %r288, 1792;
or.b32 %r290, %r289, %r287;
mul.wide.u32 %rd40, %r290, 4;
add.s64 %rd41, %rd35, %rd40;
ld.global.nc.u32 %r291, [%rd41];
and.b32 %r218, %r291, 252645135;
shr.u32 %r292, %r291, 4;
and.b32 %r222, %r292, 252645135;
ld.global.nc.u32 %r83, [%rd37+20];
ld.global.nc.u32 %r87, [%rd37+24];

dp4a.s32.s32 %r81, %r218, %r83, %r77;


dp4a.s32.s32 %r85, %r222, %r87, %r81;

shr.u32 %r293, %r275, 1;
and.b32 %r294, %r293, 1792;
or.b32 %r295, %r294, %r274;
mul.wide.u32 %rd42, %r295, 4;
add.s64 %rd43, %rd35, %rd42;
ld.global.nc.u32 %r296, [%rd43];
and.b32 %r226, %r296, 252645135;
shr.u32 %r297, %r296, 4;
and.b32 %r230, %r297, 252645135;
ld.global.nc.u32 %r91, [%rd37+28];
ld.global.nc.u32 %r95, [%rd37+32];

dp4a.s32.s32 %r89, %r226, %r91, %r85;


dp4a.s32.s32 %r93, %r230, %r95, %r89;

add.s64 %rd44, %rd78, %rd6;
ld.global.nc.u16 %rs1, [%rd44];

{ cvt.f32.f16 %f37, %rs1;}


shr.u32 %r298, %r275, 11;
and.b32 %r299, %r298, 14;
or.b32 %r300, %r299, 1;
cvt.rn.f32.s32 %f55, %r300;
mul.ftz.f32 %f56, %f37, %f55;
and.b32 %r301, %r275, 32768;
cvt.rn.f32.s32 %f57, %r301;
mul.ftz.f32 %f58, %f57, 0f37000000;
mov.f32 %f59, 0fBF600000;
sub.ftz.f32 %f60, %f59, %f58;
ld.global.nc.u32 %r97, [%rd37];

{.reg .f16 low,high;
mov.b32 {low,high},%r97;
cvt.f32.f16 %f38, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r97;
cvt.f32.f16 %f39, high;}


cvt.rn.f32.s32 %f61, %r93;
mul.ftz.f32 %f62, %f60, %f39;
fma.rn.ftz.f32 %f63, %f38, %f61, %f62;
fma.rn.ftz.f32 %f189, %f56, %f63, %f189;
add.s64 %rd45, %rd78, %rd8;
ld.global.nc.u16 %rs14, [%rd45];
shr.u16 %rs15, %rs14, 8;
cvt.u32.u16 %r302, %rs15;
ld.global.nc.u16 %rs16, [%rd45+2];
shr.u16 %rs17, %rs16, 8;
cvt.u32.u16 %r303, %rs17;
add.s64 %rd46, %rd78, %rd5;
ld.global.nc.u16 %rs18, [%rd46];
cvt.u32.u16 %r304, %rs18;
and.b16 %rs19, %rs14, 255;
cvt.u32.u16 %r305, %rs19;
mul.wide.u16 %r306, %rs18, 256;
and.b32 %r307, %r306, 1792;
or.b32 %r308, %r307, %r305;
mul.wide.u32 %rd47, %r308, 4;
add.s64 %rd48, %rd35, %rd47;
ld.global.nc.u32 %r309, [%rd48];
and.b32 %r236, %r309, 252645135;
shr.u32 %r310, %r309, 4;
and.b32 %r240, %r310, 252645135;

dp4a.s32.s32 %r99, %r236, %r67, %r238;


dp4a.s32.s32 %r103, %r240, %r71, %r99;

mul.wide.u16 %r311, %rs18, 32;
and.b32 %r312, %r311, 1792;
or.b32 %r313, %r312, %r302;
mul.wide.u32 %rd49, %r313, 4;
add.s64 %rd50, %rd35, %rd49;
ld.global.nc.u32 %r314, [%rd50];
and.b32 %r244, %r314, 252645135;
shr.u32 %r315, %r314, 4;
and.b32 %r248, %r315, 252645135;

dp4a.s32.s32 %r107, %r244, %r75, %r103;


dp4a.s32.s32 %r111, %r248, %r79, %r107;

and.b16 %rs20, %rs16, 255;
cvt.u32.u16 %r316, %rs20;
mul.wide.u16 %r317, %rs18, 4;
and.b32 %r318, %r317, 1792;
or.b32 %r319, %r318, %r316;
mul.wide.u32 %rd51, %r319, 4;
add.s64 %rd52, %rd35, %rd51;
ld.global.nc.u32 %r320, [%rd52];
and.b32 %r252, %r320, 252645135;
shr.u32 %r321, %r320, 4;
and.b32 %r256, %r321, 252645135;

dp4a.s32.s32 %r115, %r252, %r83, %r111;


dp4a.s32.s32 %r119, %r256, %r87, %r115;

shr.u32 %r322, %r304, 1;
and.b32 %r323, %r322, 1792;
or.b32 %r324, %r323, %r303;
mul.wide.u32 %rd53, %r324, 4;
add.s64 %rd54, %rd35, %rd53;
ld.global.nc.u32 %r325, [%rd54];
and.b32 %r260, %r325, 252645135;
shr.u32 %r326, %r325, 4;
and.b32 %r264, %r326, 252645135;

dp4a.s32.s32 %r123, %r260, %r91, %r119;


dp4a.s32.s32 %r127, %r264, %r95, %r123;

add.s64 %rd55, %rd78, %rd4;
ld.global.nc.u16 %rs2, [%rd55];

{ cvt.f32.f16 %f40, %rs2;}


shr.u32 %r327, %r304, 11;
and.b32 %r328, %r327, 14;
or.b32 %r329, %r328, 1;
cvt.rn.f32.s32 %f64, %r329;
mul.ftz.f32 %f65, %f40, %f64;
and.b32 %r330, %r304, 32768;
cvt.rn.f32.s32 %f66, %r330;
mul.ftz.f32 %f67, %f66, 0f37000000;
sub.ftz.f32 %f68, %f59, %f67;

{.reg .f16 low,high;
mov.b32 {low,high},%r97;
cvt.f32.f16 %f41, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r97;
cvt.f32.f16 %f42, high;}


cvt.rn.f32.s32 %f69, %r127;
mul.ftz.f32 %f70, %f68, %f42;
fma.rn.ftz.f32 %f71, %f41, %f69, %f70;
fma.rn.ftz.f32 %f188, %f65, %f71, %f188;
cvt.u64.u32 %rd56, %r447;
add.s64 %rd57, %rd27, %rd56;
add.s64 %rd58, %rd57, %rd3;
mul.lo.s64 %rd59, %rd58, 36;
add.s64 %rd60, %rd10, %rd59;
ld.global.nc.u32 %r135, [%rd60+4];
ld.global.nc.u32 %r139, [%rd60+8];

dp4a.s32.s32 %r133, %r202, %r135, %r238;


dp4a.s32.s32 %r137, %r206, %r139, %r133;

ld.global.nc.u32 %r143, [%rd60+12];
ld.global.nc.u32 %r147, [%rd60+16];

dp4a.s32.s32 %r141, %r210, %r143, %r137;


dp4a.s32.s32 %r145, %r214, %r147, %r141;

ld.global.nc.u32 %r151, [%rd60+20];
ld.global.nc.u32 %r155, [%rd60+24];

dp4a.s32.s32 %r149, %r218, %r151, %r145;


dp4a.s32.s32 %r153, %r222, %r155, %r149;

ld.global.nc.u32 %r159, [%rd60+28];
ld.global.nc.u32 %r163, [%rd60+32];

dp4a.s32.s32 %r157, %r226, %r159, %r153;


dp4a.s32.s32 %r161, %r230, %r163, %r157;


{ cvt.f32.f16 %f43, %rs1;}


mul.ftz.f32 %f72, %f43, %f55;
ld.global.nc.u32 %r165, [%rd60];

{.reg .f16 low,high;
mov.b32 {low,high},%r165;
cvt.f32.f16 %f44, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r165;
cvt.f32.f16 %f45, high;}


cvt.rn.f32.s32 %f73, %r161;
mul.ftz.f32 %f74, %f60, %f45;
fma.rn.ftz.f32 %f75, %f44, %f73, %f74;
fma.rn.ftz.f32 %f187, %f72, %f75, %f187;

dp4a.s32.s32 %r167, %r236, %r135, %r238;


dp4a.s32.s32 %r171, %r240, %r139, %r167;


dp4a.s32.s32 %r175, %r244, %r143, %r171;


dp4a.s32.s32 %r179, %r248, %r147, %r175;


dp4a.s32.s32 %r183, %r252, %r151, %r179;


dp4a.s32.s32 %r187, %r256, %r155, %r183;


dp4a.s32.s32 %r191, %r260, %r159, %r187;


dp4a.s32.s32 %r195, %r264, %r163, %r191;


{ cvt.f32.f16 %f46, %rs2;}


mul.ftz.f32 %f76, %f46, %f64;

{.reg .f16 low,high;
mov.b32 {low,high},%r165;
cvt.f32.f16 %f47, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r165;
cvt.f32.f16 %f48, high;}


cvt.rn.f32.s32 %f77, %r195;
mul.ftz.f32 %f78, %f68, %f48;
fma.rn.ftz.f32 %f79, %f47, %f77, %f78;
fma.rn.ftz.f32 %f186, %f76, %f79, %f186;
cvt.u64.u32 %rd61, %r448;
add.s64 %rd62, %rd27, %rd61;
add.s64 %rd63, %rd62, %rd3;
mul.lo.s64 %rd64, %rd63, 36;
add.s64 %rd65, %rd10, %rd64;
ld.global.nc.u32 %r203, [%rd65+4];
ld.global.nc.u32 %r207, [%rd65+8];

dp4a.s32.s32 %r201, %r202, %r203, %r238;


dp4a.s32.s32 %r205, %r206, %r207, %r201;

ld.global.nc.u32 %r211, [%rd65+12];
ld.global.nc.u32 %r215, [%rd65+16];

dp4a.s32.s32 %r209, %r210, %r211, %r205;


dp4a.s32.s32 %r213, %r214, %r215, %r209;

ld.global.nc.u32 %r219, [%rd65+20];
ld.global.nc.u32 %r223, [%rd65+24];

dp4a.s32.s32 %r217, %r218, %r219, %r213;


dp4a.s32.s32 %r221, %r222, %r223, %r217;

ld.global.nc.u32 %r227, [%rd65+28];
ld.global.nc.u32 %r231, [%rd65+32];

dp4a.s32.s32 %r225, %r226, %r227, %r221;


dp4a.s32.s32 %r229, %r230, %r231, %r225;


{ cvt.f32.f16 %f49, %rs1;}


mul.ftz.f32 %f80, %f49, %f55;
ld.global.nc.u32 %r233, [%rd65];

{.reg .f16 low,high;
mov.b32 {low,high},%r233;
cvt.f32.f16 %f50, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r233;
cvt.f32.f16 %f51, high;}


cvt.rn.f32.s32 %f81, %r229;
mul.ftz.f32 %f82, %f60, %f51;
fma.rn.ftz.f32 %f83, %f50, %f81, %f82;
fma.rn.ftz.f32 %f185, %f80, %f83, %f185;

dp4a.s32.s32 %r235, %r236, %r203, %r238;


dp4a.s32.s32 %r239, %r240, %r207, %r235;


dp4a.s32.s32 %r243, %r244, %r211, %r239;


dp4a.s32.s32 %r247, %r248, %r215, %r243;


dp4a.s32.s32 %r251, %r252, %r219, %r247;


dp4a.s32.s32 %r255, %r256, %r223, %r251;


dp4a.s32.s32 %r259, %r260, %r227, %r255;


dp4a.s32.s32 %r263, %r264, %r231, %r259;


{ cvt.f32.f16 %f52, %rs2;}


mul.ftz.f32 %f84, %f52, %f64;

{.reg .f16 low,high;
mov.b32 {low,high},%r233;
cvt.f32.f16 %f53, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r233;
cvt.f32.f16 %f54, high;}


cvt.rn.f32.s32 %f85, %r263;
mul.ftz.f32 %f86, %f68, %f54;
fma.rn.ftz.f32 %f87, %f53, %f85, %f86;
fma.rn.ftz.f32 %f184, %f84, %f87, %f184;
add.s32 %r448, %r448, 128;
add.s32 %r447, %r447, 128;
add.s32 %r446, %r446, 128;
add.s64 %rd78, %rd78, 800;
add.s32 %r449, %r449, 16;
setp.lt.s32 %p4, %r449, %r41;
@%p4 bra $L__BB257_2;

$L__BB257_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB257_5;

mov.u32 %r332, _ZZ13mul_mat_vec_qIL9ggml_type19ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r333, %r1, 768, %r332;
shl.b32 %r334, %r2, 2;
add.s32 %r335, %r333, %r334;
st.shared.f32 [%r335+-768], %f189;
st.shared.f32 [%r335+-640], %f188;
st.shared.f32 [%r335+-512], %f187;
st.shared.f32 [%r335+-384], %f186;
st.shared.f32 [%r335+-256], %f185;
st.shared.f32 [%r335+-128], %f184;

$L__BB257_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB257_18;

mov.u32 %r336, %ctaid.z;
mad.lo.s32 %r337, %r5, %r29, %r4;
mad.lo.s32 %r338, %r336, %r35, %r337;
cvt.u64.u32 %rd13, %r338;
shl.b32 %r339, %r2, 2;
mov.u32 %r340, _ZZ13mul_mat_vec_qIL9ggml_type19ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r18, %r340, %r339;
mov.u32 %r341, 2;
ld.shared.f32 %f88, [%r18];
add.ftz.f32 %f89, %f88, %f189;
ld.shared.f32 %f90, [%r18+768];
add.ftz.f32 %f91, %f90, %f89;
ld.shared.f32 %f92, [%r18+1536];
add.ftz.f32 %f93, %f92, %f91;
mov.b32 %r342, %f93;
mov.u32 %r343, 31;
mov.u32 %r344, 16;
mov.u32 %r345, -1;
shfl.sync.bfly.b32 %r346|%p7, %r342, %r344, %r343, %r345;
mov.b32 %f94, %r346;
add.ftz.f32 %f95, %f93, %f94;
mov.b32 %r347, %f95;
mov.u32 %r348, 8;
shfl.sync.bfly.b32 %r349|%p8, %r347, %r348, %r343, %r345;
mov.b32 %f96, %r349;
add.ftz.f32 %f97, %f95, %f96;
mov.b32 %r350, %f97;
mov.u32 %r351, 4;
shfl.sync.bfly.b32 %r352|%p9, %r350, %r351, %r343, %r345;
mov.b32 %f98, %r352;
add.ftz.f32 %f99, %f97, %f98;
mov.b32 %r353, %f99;
shfl.sync.bfly.b32 %r354|%p10, %r353, %r341, %r343, %r345;
mov.b32 %f100, %r354;
add.ftz.f32 %f101, %f99, %f100;
mov.b32 %r355, %f101;
mov.u32 %r356, 1;
shfl.sync.bfly.b32 %r357|%p11, %r355, %r356, %r343, %r345;
mov.b32 %f102, %r357;
add.ftz.f32 %f19, %f101, %f102;
setp.lt.u32 %p12, %r4, %r23;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
mul.wide.u32 %rd66, %r338, 4;
add.s64 %rd14, %rd1, %rd66;
not.pred %p14, %p1;
@%p14 bra $L__BB257_8;

st.global.f32 [%rd14], %f19;

$L__BB257_8:
ld.shared.f32 %f103, [%r18+128];
add.ftz.f32 %f104, %f103, %f188;
ld.shared.f32 %f105, [%r18+896];
add.ftz.f32 %f106, %f105, %f104;
ld.shared.f32 %f107, [%r18+1664];
add.ftz.f32 %f108, %f107, %f106;
mov.b32 %r358, %f108;
shfl.sync.bfly.b32 %r362|%p15, %r358, %r344, %r343, %r345;
mov.b32 %f109, %r362;
add.ftz.f32 %f110, %f108, %f109;
mov.b32 %r363, %f110;
shfl.sync.bfly.b32 %r365|%p16, %r363, %r348, %r343, %r345;
mov.b32 %f111, %r365;
add.ftz.f32 %f112, %f110, %f111;
mov.b32 %r366, %f112;
shfl.sync.bfly.b32 %r368|%p17, %r366, %r351, %r343, %r345;
mov.b32 %f113, %r368;
add.ftz.f32 %f114, %f112, %f113;
mov.b32 %r369, %f114;
shfl.sync.bfly.b32 %r371|%p18, %r369, %r341, %r343, %r345;
mov.b32 %f115, %r371;
add.ftz.f32 %f116, %f114, %f115;
mov.b32 %r372, %f116;
shfl.sync.bfly.b32 %r374|%p19, %r372, %r356, %r343, %r345;
mov.b32 %f117, %r374;
add.ftz.f32 %f20, %f116, %f117;
add.s32 %r375, %r4, 1;
setp.lt.u32 %p20, %r375, %r23;
setp.eq.s32 %p21, %r2, 1;
and.pred %p2, %p21, %p20;
not.pred %p22, %p2;
@%p22 bra $L__BB257_10;

st.global.f32 [%rd14+4], %f20;

$L__BB257_10:
ld.shared.f32 %f118, [%r18+256];
add.ftz.f32 %f119, %f118, %f187;
ld.shared.f32 %f120, [%r18+1024];
add.ftz.f32 %f121, %f120, %f119;
ld.shared.f32 %f122, [%r18+1792];
add.ftz.f32 %f123, %f122, %f121;
mov.b32 %r376, %f123;
mov.u32 %r377, 31;
mov.u32 %r378, 16;
mov.u32 %r379, -1;
shfl.sync.bfly.b32 %r380|%p23, %r376, %r378, %r377, %r379;
mov.b32 %f124, %r380;
add.ftz.f32 %f125, %f123, %f124;
mov.b32 %r381, %f125;
mov.u32 %r382, 8;
shfl.sync.bfly.b32 %r383|%p24, %r381, %r382, %r377, %r379;
mov.b32 %f126, %r383;
add.ftz.f32 %f127, %f125, %f126;
mov.b32 %r384, %f127;
mov.u32 %r385, 4;
shfl.sync.bfly.b32 %r386|%p25, %r384, %r385, %r377, %r379;
mov.b32 %f128, %r386;
add.ftz.f32 %f129, %f127, %f128;
mov.b32 %r387, %f129;
mov.u32 %r388, 2;
shfl.sync.bfly.b32 %r389|%p26, %r387, %r388, %r377, %r379;
mov.b32 %f130, %r389;
add.ftz.f32 %f131, %f129, %f130;
mov.b32 %r390, %f131;
mov.u32 %r391, 1;
shfl.sync.bfly.b32 %r392|%p27, %r390, %r391, %r377, %r379;
mov.b32 %f132, %r392;
add.ftz.f32 %f21, %f131, %f132;
@%p14 bra $L__BB257_12;

cvt.u64.u32 %rd67, %r23;
add.s64 %rd68, %rd67, %rd13;
shl.b64 %rd69, %rd68, 2;
add.s64 %rd70, %rd1, %rd69;
st.global.f32 [%rd70], %f21;

$L__BB257_12:
ld.shared.f32 %f133, [%r18+384];
add.ftz.f32 %f134, %f133, %f186;
ld.shared.f32 %f135, [%r18+1152];
add.ftz.f32 %f136, %f135, %f134;
ld.shared.f32 %f137, [%r18+1920];
add.ftz.f32 %f138, %f137, %f136;
mov.b32 %r393, %f138;
shfl.sync.bfly.b32 %r397|%p29, %r393, %r378, %r377, %r379;
mov.b32 %f139, %r397;
add.ftz.f32 %f140, %f138, %f139;
mov.b32 %r398, %f140;
shfl.sync.bfly.b32 %r400|%p30, %r398, %r382, %r377, %r379;
mov.b32 %f141, %r400;
add.ftz.f32 %f142, %f140, %f141;
mov.b32 %r401, %f142;
shfl.sync.bfly.b32 %r403|%p31, %r401, %r385, %r377, %r379;
mov.b32 %f143, %r403;
add.ftz.f32 %f144, %f142, %f143;
mov.b32 %r404, %f144;
shfl.sync.bfly.b32 %r406|%p32, %r404, %r388, %r377, %r379;
mov.b32 %f145, %r406;
add.ftz.f32 %f146, %f144, %f145;
mov.b32 %r407, %f146;
shfl.sync.bfly.b32 %r409|%p33, %r407, %r391, %r377, %r379;
mov.b32 %f147, %r409;
add.ftz.f32 %f22, %f146, %f147;
@%p22 bra $L__BB257_14;

add.s32 %r410, %r23, 1;
cvt.u64.u32 %rd71, %r410;
add.s64 %rd72, %rd71, %rd13;
shl.b64 %rd73, %rd72, 2;
add.s64 %rd74, %rd1, %rd73;
st.global.f32 [%rd74], %f22;

$L__BB257_14:
shl.b32 %r411, %r23, 1;
mov.u32 %r412, 1;
ld.shared.f32 %f148, [%r18+512];
add.ftz.f32 %f149, %f148, %f185;
ld.shared.f32 %f150, [%r18+1280];
add.ftz.f32 %f151, %f150, %f149;
ld.shared.f32 %f152, [%r18+2048];
add.ftz.f32 %f153, %f152, %f151;
mov.b32 %r413, %f153;
mov.u32 %r414, 31;
mov.u32 %r415, 16;
mov.u32 %r416, -1;
shfl.sync.bfly.b32 %r417|%p35, %r413, %r415, %r414, %r416;
mov.b32 %f154, %r417;
add.ftz.f32 %f155, %f153, %f154;
mov.b32 %r418, %f155;
mov.u32 %r419, 8;
shfl.sync.bfly.b32 %r420|%p36, %r418, %r419, %r414, %r416;
mov.b32 %f156, %r420;
add.ftz.f32 %f157, %f155, %f156;
mov.b32 %r421, %f157;
mov.u32 %r422, 4;
shfl.sync.bfly.b32 %r423|%p37, %r421, %r422, %r414, %r416;
mov.b32 %f158, %r423;
add.ftz.f32 %f159, %f157, %f158;
mov.b32 %r424, %f159;
mov.u32 %r425, 2;
shfl.sync.bfly.b32 %r426|%p38, %r424, %r425, %r414, %r416;
mov.b32 %f160, %r426;
add.ftz.f32 %f161, %f159, %f160;
mov.b32 %r427, %f161;
shfl.sync.bfly.b32 %r428|%p39, %r427, %r412, %r414, %r416;
mov.b32 %f162, %r428;
add.ftz.f32 %f23, %f161, %f162;
cvt.u64.u32 %rd75, %r411;
add.s64 %rd76, %rd75, %rd13;
shl.b64 %rd77, %rd76, 2;
add.s64 %rd15, %rd1, %rd77;
@%p14 bra $L__BB257_16;

st.global.f32 [%rd15], %f23;

$L__BB257_16:
ld.shared.f32 %f163, [%r18+640];
add.ftz.f32 %f164, %f163, %f184;
ld.shared.f32 %f165, [%r18+1408];
add.ftz.f32 %f166, %f165, %f164;
ld.shared.f32 %f167, [%r18+2176];
add.ftz.f32 %f168, %f167, %f166;
mov.b32 %r429, %f168;
shfl.sync.bfly.b32 %r433|%p41, %r429, %r415, %r414, %r416;
mov.b32 %f169, %r433;
add.ftz.f32 %f170, %f168, %f169;
mov.b32 %r434, %f170;
shfl.sync.bfly.b32 %r436|%p42, %r434, %r419, %r414, %r416;
mov.b32 %f171, %r436;
add.ftz.f32 %f172, %f170, %f171;
mov.b32 %r437, %f172;
shfl.sync.bfly.b32 %r439|%p43, %r437, %r422, %r414, %r416;
mov.b32 %f173, %r439;
add.ftz.f32 %f174, %f172, %f173;
mov.b32 %r440, %f174;
shfl.sync.bfly.b32 %r442|%p44, %r440, %r425, %r414, %r416;
mov.b32 %f175, %r442;
add.ftz.f32 %f176, %f174, %f175;
mov.b32 %r443, %f176;
shfl.sync.bfly.b32 %r445|%p45, %r443, %r412, %r414, %r416;
mov.b32 %f177, %r445;
add.ftz.f32 %f24, %f176, %f177;
@%p22 bra $L__BB257_18;

st.global.f32 [%rd15+4], %f24;

$L__BB257_18:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type19ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type19ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type19ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type19ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type19ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type19ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type19ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type19ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type19ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<59>;
.reg .b16 %rs<23>;
.reg .f32 %f<250>;
.reg .b32 %r<562>;
.reg .b64 %rd<92>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type19ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[3072];

ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type19ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type19ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type19ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type19ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type19ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type19ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type19ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type19ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type19ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type19ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type19ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type19ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type19ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type19ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type19ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type19ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type19ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
cvta.to.global.u64 %rd1, %rd18;
mov.u32 %r1, %tid.y;
shl.b32 %r40, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r40, %r2;
mov.u32 %r41, %ctaid.x;
shl.b32 %r4, %r41, 1;
mov.u32 %r5, %ctaid.y;
shr.s32 %r42, %r3, 31;
shr.u32 %r43, %r42, 29;
add.s32 %r44, %r3, %r43;
shr.s32 %r561, %r44, 3;
shr.u32 %r45, %r24, 8;
setp.ge.s32 %p3, %r561, %r45;
mov.f32 %f242, 0f00000000;
mov.f32 %f243, %f242;
mov.f32 %f244, %f242;
mov.f32 %f245, %f242;
mov.f32 %f246, %f242;
mov.f32 %f247, %f242;
mov.f32 %f248, %f242;
mov.f32 %f249, %f242;
@%p3 bra $L__BB258_3;

mov.u32 %r46, %ctaid.z;
mul.hi.u32 %r47, %r46, %r34;
add.s32 %r48, %r47, %r46;
shr.u32 %r49, %r48, %r35;
mul.lo.s32 %r50, %r49, %r37;
mul.hi.u32 %r51, %r5, %r28;
add.s32 %r52, %r51, %r5;
shr.u32 %r53, %r52, %r29;
mul.lo.s32 %r54, %r53, %r31;
cvta.to.global.u64 %rd91, %rd16;
and.b32 %r58, %r44, -8;
sub.s32 %r59, %r3, %r58;
shl.b32 %r60, %r59, 1;
cvt.s64.s32 %rd3, %r59;
shl.b32 %r61, %r26, 1;
shl.b32 %r557, %r561, 3;
add.s32 %r560, %r557, %r61;
mad.lo.s32 %r559, %r26, 3, %r557;
add.s32 %r558, %r26, %r557;
add.s32 %r62, %r561, %r50;
add.s32 %r63, %r62, %r54;
add.s32 %r64, %r4, 1;
mad.lo.s32 %r65, %r25, %r64, %r63;
mul.wide.s32 %rd4, %r65, 50;
mul.wide.s32 %rd19, %r59, 2;
add.s64 %rd20, %rd4, %rd19;
add.s64 %rd5, %rd20, 34;
mad.lo.s32 %r66, %r4, %r25, %r561;
add.s32 %r67, %r66, %r50;
add.s32 %r68, %r67, %r54;
mul.wide.s32 %rd6, %r68, 50;
mul.wide.s32 %rd21, %r60, 2;
add.s64 %rd22, %rd6, %rd21;
add.s64 %rd7, %rd22, 4;
add.s64 %rd23, %rd4, %rd21;
add.s64 %rd8, %rd23, 2;
add.s64 %rd24, %rd6, %rd19;
add.s64 %rd9, %rd24, 34;
cvta.to.global.u64 %rd10, %rd17;
mov.f32 %f242, 0f00000000;

$L__BB258_2:
mul.lo.s32 %r342, %r46, %r38;
cvt.u64.u32 %rd25, %r342;
mul.lo.s32 %r344, %r5, %r32;
cvt.u64.u32 %rd26, %r344;
add.s64 %rd27, %rd25, %rd26;
cvt.u64.u32 %rd28, %r557;
add.s64 %rd29, %rd27, %rd28;
add.s64 %rd30, %rd29, %rd3;
mul.lo.s64 %rd31, %rd30, 36;
add.s64 %rd32, %rd91, %rd7;
ld.global.nc.u16 %rs9, [%rd32+-2];
shr.u16 %rs10, %rs9, 8;
cvt.u32.u16 %r345, %rs10;
ld.global.nc.u16 %rs11, [%rd32];
shr.u16 %rs12, %rs11, 8;
cvt.u32.u16 %r346, %rs12;
add.s64 %rd33, %rd91, %rd9;
ld.global.nc.u16 %rs13, [%rd33];
cvt.u32.u16 %r347, %rs13;
and.b16 %rs14, %rs9, 255;
cvt.u32.u16 %r348, %rs14;
mul.wide.u16 %r349, %rs13, 256;
and.b32 %r350, %r349, 1792;
or.b32 %r351, %r350, %r348;
mul.wide.u32 %rd34, %r351, 4;
mov.u64 %rd35, iq1s_grid_gpu;
add.s64 %rd36, %rd35, %rd34;
ld.global.nc.u32 %r352, [%rd36];
and.b32 %r274, %r352, 252645135;
shr.u32 %r353, %r352, 4;
and.b32 %r278, %r353, 252645135;
add.s64 %rd37, %rd10, %rd31;
ld.global.nc.u32 %r71, [%rd37+4];
ld.global.nc.u32 %r75, [%rd37+8];
mov.u32 %r310, 0;

dp4a.s32.s32 %r69, %r274, %r71, %r310;


dp4a.s32.s32 %r73, %r278, %r75, %r69;

mul.wide.u16 %r354, %rs13, 32;
and.b32 %r355, %r354, 1792;
or.b32 %r356, %r355, %r345;
mul.wide.u32 %rd38, %r356, 4;
add.s64 %rd39, %rd35, %rd38;
ld.global.nc.u32 %r357, [%rd39];
and.b32 %r282, %r357, 252645135;
shr.u32 %r358, %r357, 4;
and.b32 %r286, %r358, 252645135;
ld.global.nc.u32 %r79, [%rd37+12];
ld.global.nc.u32 %r83, [%rd37+16];

dp4a.s32.s32 %r77, %r282, %r79, %r73;


dp4a.s32.s32 %r81, %r286, %r83, %r77;

and.b16 %rs15, %rs11, 255;
cvt.u32.u16 %r359, %rs15;
mul.wide.u16 %r360, %rs13, 4;
and.b32 %r361, %r360, 1792;
or.b32 %r362, %r361, %r359;
mul.wide.u32 %rd40, %r362, 4;
add.s64 %rd41, %rd35, %rd40;
ld.global.nc.u32 %r363, [%rd41];
and.b32 %r290, %r363, 252645135;
shr.u32 %r364, %r363, 4;
and.b32 %r294, %r364, 252645135;
ld.global.nc.u32 %r87, [%rd37+20];
ld.global.nc.u32 %r91, [%rd37+24];

dp4a.s32.s32 %r85, %r290, %r87, %r81;


dp4a.s32.s32 %r89, %r294, %r91, %r85;

shr.u32 %r365, %r347, 1;
and.b32 %r366, %r365, 1792;
or.b32 %r367, %r366, %r346;
mul.wide.u32 %rd42, %r367, 4;
add.s64 %rd43, %rd35, %rd42;
ld.global.nc.u32 %r368, [%rd43];
and.b32 %r298, %r368, 252645135;
shr.u32 %r369, %r368, 4;
and.b32 %r302, %r369, 252645135;
ld.global.nc.u32 %r95, [%rd37+28];
ld.global.nc.u32 %r99, [%rd37+32];

dp4a.s32.s32 %r93, %r298, %r95, %r89;


dp4a.s32.s32 %r97, %r302, %r99, %r93;

add.s64 %rd44, %rd91, %rd6;
ld.global.nc.u16 %rs1, [%rd44];

{ cvt.f32.f16 %f49, %rs1;}


shr.u32 %r370, %r347, 11;
and.b32 %r371, %r370, 14;
or.b32 %r372, %r371, 1;
cvt.rn.f32.s32 %f73, %r372;
mul.ftz.f32 %f74, %f49, %f73;
and.b32 %r373, %r347, 32768;
cvt.rn.f32.s32 %f75, %r373;
mul.ftz.f32 %f76, %f75, 0f37000000;
mov.f32 %f77, 0fBF600000;
sub.ftz.f32 %f78, %f77, %f76;
ld.global.nc.u32 %r101, [%rd37];

{.reg .f16 low,high;
mov.b32 {low,high},%r101;
cvt.f32.f16 %f50, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r101;
cvt.f32.f16 %f51, high;}


cvt.rn.f32.s32 %f79, %r97;
mul.ftz.f32 %f80, %f78, %f51;
fma.rn.ftz.f32 %f81, %f50, %f79, %f80;
fma.rn.ftz.f32 %f249, %f74, %f81, %f249;
add.s64 %rd45, %rd91, %rd8;
ld.global.nc.u16 %rs16, [%rd45];
shr.u16 %rs17, %rs16, 8;
cvt.u32.u16 %r374, %rs17;
ld.global.nc.u16 %rs18, [%rd45+2];
shr.u16 %rs19, %rs18, 8;
cvt.u32.u16 %r375, %rs19;
add.s64 %rd46, %rd91, %rd5;
ld.global.nc.u16 %rs20, [%rd46];
cvt.u32.u16 %r376, %rs20;
and.b16 %rs21, %rs16, 255;
cvt.u32.u16 %r377, %rs21;
mul.wide.u16 %r378, %rs20, 256;
and.b32 %r379, %r378, 1792;
or.b32 %r380, %r379, %r377;
mul.wide.u32 %rd47, %r380, 4;
add.s64 %rd48, %rd35, %rd47;
ld.global.nc.u32 %r381, [%rd48];
and.b32 %r308, %r381, 252645135;
shr.u32 %r382, %r381, 4;
and.b32 %r312, %r382, 252645135;

dp4a.s32.s32 %r103, %r308, %r71, %r310;


dp4a.s32.s32 %r107, %r312, %r75, %r103;

mul.wide.u16 %r383, %rs20, 32;
and.b32 %r384, %r383, 1792;
or.b32 %r385, %r384, %r374;
mul.wide.u32 %rd49, %r385, 4;
add.s64 %rd50, %rd35, %rd49;
ld.global.nc.u32 %r386, [%rd50];
and.b32 %r316, %r386, 252645135;
shr.u32 %r387, %r386, 4;
and.b32 %r320, %r387, 252645135;

dp4a.s32.s32 %r111, %r316, %r79, %r107;


dp4a.s32.s32 %r115, %r320, %r83, %r111;

and.b16 %rs22, %rs18, 255;
cvt.u32.u16 %r388, %rs22;
mul.wide.u16 %r389, %rs20, 4;
and.b32 %r390, %r389, 1792;
or.b32 %r391, %r390, %r388;
mul.wide.u32 %rd51, %r391, 4;
add.s64 %rd52, %rd35, %rd51;
ld.global.nc.u32 %r392, [%rd52];
and.b32 %r324, %r392, 252645135;
shr.u32 %r393, %r392, 4;
and.b32 %r328, %r393, 252645135;

dp4a.s32.s32 %r119, %r324, %r87, %r115;


dp4a.s32.s32 %r123, %r328, %r91, %r119;

shr.u32 %r394, %r376, 1;
and.b32 %r395, %r394, 1792;
or.b32 %r396, %r395, %r375;
mul.wide.u32 %rd53, %r396, 4;
add.s64 %rd54, %rd35, %rd53;
ld.global.nc.u32 %r397, [%rd54];
and.b32 %r332, %r397, 252645135;
shr.u32 %r398, %r397, 4;
and.b32 %r336, %r398, 252645135;

dp4a.s32.s32 %r127, %r332, %r95, %r123;


dp4a.s32.s32 %r131, %r336, %r99, %r127;

add.s64 %rd55, %rd91, %rd4;
ld.global.nc.u16 %rs2, [%rd55];

{ cvt.f32.f16 %f52, %rs2;}


shr.u32 %r399, %r376, 11;
and.b32 %r400, %r399, 14;
or.b32 %r401, %r400, 1;
cvt.rn.f32.s32 %f82, %r401;
mul.ftz.f32 %f83, %f52, %f82;
and.b32 %r402, %r376, 32768;
cvt.rn.f32.s32 %f84, %r402;
mul.ftz.f32 %f85, %f84, 0f37000000;
sub.ftz.f32 %f86, %f77, %f85;

{.reg .f16 low,high;
mov.b32 {low,high},%r101;
cvt.f32.f16 %f53, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r101;
cvt.f32.f16 %f54, high;}


cvt.rn.f32.s32 %f87, %r131;
mul.ftz.f32 %f88, %f86, %f54;
fma.rn.ftz.f32 %f89, %f53, %f87, %f88;
fma.rn.ftz.f32 %f248, %f83, %f89, %f248;
cvt.u64.u32 %rd56, %r558;
add.s64 %rd57, %rd27, %rd56;
add.s64 %rd58, %rd57, %rd3;
mul.lo.s64 %rd59, %rd58, 36;
add.s64 %rd60, %rd10, %rd59;
ld.global.nc.u32 %r139, [%rd60+4];
ld.global.nc.u32 %r143, [%rd60+8];

dp4a.s32.s32 %r137, %r274, %r139, %r310;


dp4a.s32.s32 %r141, %r278, %r143, %r137;

ld.global.nc.u32 %r147, [%rd60+12];
ld.global.nc.u32 %r151, [%rd60+16];

dp4a.s32.s32 %r145, %r282, %r147, %r141;


dp4a.s32.s32 %r149, %r286, %r151, %r145;

ld.global.nc.u32 %r155, [%rd60+20];
ld.global.nc.u32 %r159, [%rd60+24];

dp4a.s32.s32 %r153, %r290, %r155, %r149;


dp4a.s32.s32 %r157, %r294, %r159, %r153;

ld.global.nc.u32 %r163, [%rd60+28];
ld.global.nc.u32 %r167, [%rd60+32];

dp4a.s32.s32 %r161, %r298, %r163, %r157;


dp4a.s32.s32 %r165, %r302, %r167, %r161;


{ cvt.f32.f16 %f55, %rs1;}


mul.ftz.f32 %f90, %f55, %f73;
ld.global.nc.u32 %r169, [%rd60];

{.reg .f16 low,high;
mov.b32 {low,high},%r169;
cvt.f32.f16 %f56, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r169;
cvt.f32.f16 %f57, high;}


cvt.rn.f32.s32 %f91, %r165;
mul.ftz.f32 %f92, %f78, %f57;
fma.rn.ftz.f32 %f93, %f56, %f91, %f92;
fma.rn.ftz.f32 %f247, %f90, %f93, %f247;

dp4a.s32.s32 %r171, %r308, %r139, %r310;


dp4a.s32.s32 %r175, %r312, %r143, %r171;


dp4a.s32.s32 %r179, %r316, %r147, %r175;


dp4a.s32.s32 %r183, %r320, %r151, %r179;


dp4a.s32.s32 %r187, %r324, %r155, %r183;


dp4a.s32.s32 %r191, %r328, %r159, %r187;


dp4a.s32.s32 %r195, %r332, %r163, %r191;


dp4a.s32.s32 %r199, %r336, %r167, %r195;


{ cvt.f32.f16 %f58, %rs2;}


mul.ftz.f32 %f94, %f58, %f82;

{.reg .f16 low,high;
mov.b32 {low,high},%r169;
cvt.f32.f16 %f59, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r169;
cvt.f32.f16 %f60, high;}


cvt.rn.f32.s32 %f95, %r199;
mul.ftz.f32 %f96, %f86, %f60;
fma.rn.ftz.f32 %f97, %f59, %f95, %f96;
fma.rn.ftz.f32 %f246, %f94, %f97, %f246;
cvt.u64.u32 %rd61, %r560;
add.s64 %rd62, %rd27, %rd61;
add.s64 %rd63, %rd62, %rd3;
mul.lo.s64 %rd64, %rd63, 36;
add.s64 %rd65, %rd10, %rd64;
ld.global.nc.u32 %r207, [%rd65+4];
ld.global.nc.u32 %r211, [%rd65+8];

dp4a.s32.s32 %r205, %r274, %r207, %r310;


dp4a.s32.s32 %r209, %r278, %r211, %r205;

ld.global.nc.u32 %r215, [%rd65+12];
ld.global.nc.u32 %r219, [%rd65+16];

dp4a.s32.s32 %r213, %r282, %r215, %r209;


dp4a.s32.s32 %r217, %r286, %r219, %r213;

ld.global.nc.u32 %r223, [%rd65+20];
ld.global.nc.u32 %r227, [%rd65+24];

dp4a.s32.s32 %r221, %r290, %r223, %r217;


dp4a.s32.s32 %r225, %r294, %r227, %r221;

ld.global.nc.u32 %r231, [%rd65+28];
ld.global.nc.u32 %r235, [%rd65+32];

dp4a.s32.s32 %r229, %r298, %r231, %r225;


dp4a.s32.s32 %r233, %r302, %r235, %r229;


{ cvt.f32.f16 %f61, %rs1;}


mul.ftz.f32 %f98, %f61, %f73;
ld.global.nc.u32 %r237, [%rd65];

{.reg .f16 low,high;
mov.b32 {low,high},%r237;
cvt.f32.f16 %f62, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r237;
cvt.f32.f16 %f63, high;}


cvt.rn.f32.s32 %f99, %r233;
mul.ftz.f32 %f100, %f78, %f63;
fma.rn.ftz.f32 %f101, %f62, %f99, %f100;
fma.rn.ftz.f32 %f245, %f98, %f101, %f245;

dp4a.s32.s32 %r239, %r308, %r207, %r310;


dp4a.s32.s32 %r243, %r312, %r211, %r239;


dp4a.s32.s32 %r247, %r316, %r215, %r243;


dp4a.s32.s32 %r251, %r320, %r219, %r247;


dp4a.s32.s32 %r255, %r324, %r223, %r251;


dp4a.s32.s32 %r259, %r328, %r227, %r255;


dp4a.s32.s32 %r263, %r332, %r231, %r259;


dp4a.s32.s32 %r267, %r336, %r235, %r263;


{ cvt.f32.f16 %f64, %rs2;}


mul.ftz.f32 %f102, %f64, %f82;

{.reg .f16 low,high;
mov.b32 {low,high},%r237;
cvt.f32.f16 %f65, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r237;
cvt.f32.f16 %f66, high;}


cvt.rn.f32.s32 %f103, %r267;
mul.ftz.f32 %f104, %f86, %f66;
fma.rn.ftz.f32 %f105, %f65, %f103, %f104;
fma.rn.ftz.f32 %f244, %f102, %f105, %f244;
cvt.u64.u32 %rd66, %r559;
add.s64 %rd67, %rd27, %rd66;
add.s64 %rd68, %rd67, %rd3;
mul.lo.s64 %rd69, %rd68, 36;
add.s64 %rd70, %rd10, %rd69;
ld.global.nc.u32 %r275, [%rd70+4];
ld.global.nc.u32 %r279, [%rd70+8];

dp4a.s32.s32 %r273, %r274, %r275, %r310;


dp4a.s32.s32 %r277, %r278, %r279, %r273;

ld.global.nc.u32 %r283, [%rd70+12];
ld.global.nc.u32 %r287, [%rd70+16];

dp4a.s32.s32 %r281, %r282, %r283, %r277;


dp4a.s32.s32 %r285, %r286, %r287, %r281;

ld.global.nc.u32 %r291, [%rd70+20];
ld.global.nc.u32 %r295, [%rd70+24];

dp4a.s32.s32 %r289, %r290, %r291, %r285;


dp4a.s32.s32 %r293, %r294, %r295, %r289;

ld.global.nc.u32 %r299, [%rd70+28];
ld.global.nc.u32 %r303, [%rd70+32];

dp4a.s32.s32 %r297, %r298, %r299, %r293;


dp4a.s32.s32 %r301, %r302, %r303, %r297;


{ cvt.f32.f16 %f67, %rs1;}


mul.ftz.f32 %f106, %f67, %f73;
ld.global.nc.u32 %r305, [%rd70];

{.reg .f16 low,high;
mov.b32 {low,high},%r305;
cvt.f32.f16 %f68, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r305;
cvt.f32.f16 %f69, high;}


cvt.rn.f32.s32 %f107, %r301;
mul.ftz.f32 %f108, %f78, %f69;
fma.rn.ftz.f32 %f109, %f68, %f107, %f108;
fma.rn.ftz.f32 %f243, %f106, %f109, %f243;

dp4a.s32.s32 %r307, %r308, %r275, %r310;


dp4a.s32.s32 %r311, %r312, %r279, %r307;


dp4a.s32.s32 %r315, %r316, %r283, %r311;


dp4a.s32.s32 %r319, %r320, %r287, %r315;


dp4a.s32.s32 %r323, %r324, %r291, %r319;


dp4a.s32.s32 %r327, %r328, %r295, %r323;


dp4a.s32.s32 %r331, %r332, %r299, %r327;


dp4a.s32.s32 %r335, %r336, %r303, %r331;


{ cvt.f32.f16 %f70, %rs2;}


mul.ftz.f32 %f110, %f70, %f82;

{.reg .f16 low,high;
mov.b32 {low,high},%r305;
cvt.f32.f16 %f71, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r305;
cvt.f32.f16 %f72, high;}


cvt.rn.f32.s32 %f111, %r335;
mul.ftz.f32 %f112, %f86, %f72;
fma.rn.ftz.f32 %f113, %f71, %f111, %f112;
fma.rn.ftz.f32 %f242, %f110, %f113, %f242;
add.s32 %r560, %r560, 128;
add.s32 %r559, %r559, 128;
add.s32 %r558, %r558, 128;
add.s32 %r557, %r557, 128;
add.s64 %rd91, %rd91, 800;
add.s32 %r561, %r561, 16;
setp.lt.s32 %p4, %r561, %r45;
@%p4 bra $L__BB258_2;

$L__BB258_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB258_5;

shl.b32 %r404, %r1, 10;
mov.u32 %r405, _ZZ13mul_mat_vec_qIL9ggml_type19ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r406, %r405, %r404;
shl.b32 %r407, %r2, 2;
add.s32 %r408, %r406, %r407;
st.shared.f32 [%r408+-1024], %f249;
st.shared.f32 [%r408+-896], %f248;
st.shared.f32 [%r408+-768], %f247;
st.shared.f32 [%r408+-640], %f246;
st.shared.f32 [%r408+-512], %f245;
st.shared.f32 [%r408+-384], %f244;
st.shared.f32 [%r408+-256], %f243;
st.shared.f32 [%r408+-128], %f242;

$L__BB258_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB258_22;

mov.u32 %r409, %ctaid.z;
mad.lo.s32 %r410, %r5, %r33, %r4;
mad.lo.s32 %r411, %r409, %r39, %r410;
cvt.u64.u32 %rd13, %r411;
shl.b32 %r412, %r2, 2;
mov.u32 %r413, _ZZ13mul_mat_vec_qIL9ggml_type19ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r21, %r413, %r412;
mov.u32 %r414, 2;
ld.shared.f32 %f114, [%r21];
add.ftz.f32 %f115, %f114, %f249;
ld.shared.f32 %f116, [%r21+1024];
add.ftz.f32 %f117, %f116, %f115;
ld.shared.f32 %f118, [%r21+2048];
add.ftz.f32 %f119, %f118, %f117;
mov.b32 %r415, %f119;
mov.u32 %r416, 31;
mov.u32 %r417, 16;
mov.u32 %r418, -1;
shfl.sync.bfly.b32 %r419|%p7, %r415, %r417, %r416, %r418;
mov.b32 %f120, %r419;
add.ftz.f32 %f121, %f119, %f120;
mov.b32 %r420, %f121;
mov.u32 %r421, 8;
shfl.sync.bfly.b32 %r422|%p8, %r420, %r421, %r416, %r418;
mov.b32 %f122, %r422;
add.ftz.f32 %f123, %f121, %f122;
mov.b32 %r423, %f123;
mov.u32 %r424, 4;
shfl.sync.bfly.b32 %r425|%p9, %r423, %r424, %r416, %r418;
mov.b32 %f124, %r425;
add.ftz.f32 %f125, %f123, %f124;
mov.b32 %r426, %f125;
shfl.sync.bfly.b32 %r427|%p10, %r426, %r414, %r416, %r418;
mov.b32 %f126, %r427;
add.ftz.f32 %f127, %f125, %f126;
mov.b32 %r428, %f127;
mov.u32 %r429, 1;
shfl.sync.bfly.b32 %r430|%p11, %r428, %r429, %r416, %r418;
mov.b32 %f128, %r430;
add.ftz.f32 %f25, %f127, %f128;
setp.lt.u32 %p12, %r4, %r27;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
mul.wide.u32 %rd71, %r411, 4;
add.s64 %rd14, %rd1, %rd71;
not.pred %p14, %p1;
@%p14 bra $L__BB258_8;

st.global.f32 [%rd14], %f25;

$L__BB258_8:
ld.shared.f32 %f129, [%r21+128];
add.ftz.f32 %f130, %f129, %f248;
ld.shared.f32 %f131, [%r21+1152];
add.ftz.f32 %f132, %f131, %f130;
ld.shared.f32 %f133, [%r21+2176];
add.ftz.f32 %f134, %f133, %f132;
mov.b32 %r431, %f134;
shfl.sync.bfly.b32 %r435|%p15, %r431, %r417, %r416, %r418;
mov.b32 %f135, %r435;
add.ftz.f32 %f136, %f134, %f135;
mov.b32 %r436, %f136;
shfl.sync.bfly.b32 %r438|%p16, %r436, %r421, %r416, %r418;
mov.b32 %f137, %r438;
add.ftz.f32 %f138, %f136, %f137;
mov.b32 %r439, %f138;
shfl.sync.bfly.b32 %r441|%p17, %r439, %r424, %r416, %r418;
mov.b32 %f139, %r441;
add.ftz.f32 %f140, %f138, %f139;
mov.b32 %r442, %f140;
shfl.sync.bfly.b32 %r444|%p18, %r442, %r414, %r416, %r418;
mov.b32 %f141, %r444;
add.ftz.f32 %f142, %f140, %f141;
mov.b32 %r445, %f142;
shfl.sync.bfly.b32 %r447|%p19, %r445, %r429, %r416, %r418;
mov.b32 %f143, %r447;
add.ftz.f32 %f26, %f142, %f143;
setp.eq.s32 %p20, %r2, 1;
or.b32 %r451, %r4, 1;
setp.lt.u32 %p21, %r451, %r27;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB258_10;

st.global.f32 [%rd14+4], %f26;

$L__BB258_10:
ld.shared.f32 %f144, [%r21+256];
add.ftz.f32 %f145, %f144, %f247;
ld.shared.f32 %f146, [%r21+1280];
add.ftz.f32 %f147, %f146, %f145;
ld.shared.f32 %f148, [%r21+2304];
add.ftz.f32 %f149, %f148, %f147;
mov.b32 %r452, %f149;
mov.u32 %r453, 31;
mov.u32 %r454, 16;
mov.u32 %r455, -1;
shfl.sync.bfly.b32 %r456|%p23, %r452, %r454, %r453, %r455;
mov.b32 %f150, %r456;
add.ftz.f32 %f151, %f149, %f150;
mov.b32 %r457, %f151;
mov.u32 %r458, 8;
shfl.sync.bfly.b32 %r459|%p24, %r457, %r458, %r453, %r455;
mov.b32 %f152, %r459;
add.ftz.f32 %f153, %f151, %f152;
mov.b32 %r460, %f153;
mov.u32 %r461, 4;
shfl.sync.bfly.b32 %r462|%p25, %r460, %r461, %r453, %r455;
mov.b32 %f154, %r462;
add.ftz.f32 %f155, %f153, %f154;
mov.b32 %r463, %f155;
mov.u32 %r464, 2;
shfl.sync.bfly.b32 %r465|%p26, %r463, %r464, %r453, %r455;
mov.b32 %f156, %r465;
add.ftz.f32 %f157, %f155, %f156;
mov.b32 %r466, %f157;
mov.u32 %r467, 1;
shfl.sync.bfly.b32 %r468|%p27, %r466, %r467, %r453, %r455;
mov.b32 %f158, %r468;
add.ftz.f32 %f27, %f157, %f158;
@%p14 bra $L__BB258_12;

cvt.u64.u32 %rd72, %r27;
add.s64 %rd73, %rd72, %rd13;
shl.b64 %rd74, %rd73, 2;
add.s64 %rd75, %rd1, %rd74;
st.global.f32 [%rd75], %f27;

$L__BB258_12:
ld.shared.f32 %f159, [%r21+384];
add.ftz.f32 %f160, %f159, %f246;
ld.shared.f32 %f161, [%r21+1408];
add.ftz.f32 %f162, %f161, %f160;
ld.shared.f32 %f163, [%r21+2432];
add.ftz.f32 %f164, %f163, %f162;
mov.b32 %r469, %f164;
shfl.sync.bfly.b32 %r473|%p29, %r469, %r454, %r453, %r455;
mov.b32 %f165, %r473;
add.ftz.f32 %f166, %f164, %f165;
mov.b32 %r474, %f166;
shfl.sync.bfly.b32 %r476|%p30, %r474, %r458, %r453, %r455;
mov.b32 %f167, %r476;
add.ftz.f32 %f168, %f166, %f167;
mov.b32 %r477, %f168;
shfl.sync.bfly.b32 %r479|%p31, %r477, %r461, %r453, %r455;
mov.b32 %f169, %r479;
add.ftz.f32 %f170, %f168, %f169;
mov.b32 %r480, %f170;
shfl.sync.bfly.b32 %r482|%p32, %r480, %r464, %r453, %r455;
mov.b32 %f171, %r482;
add.ftz.f32 %f172, %f170, %f171;
mov.b32 %r483, %f172;
shfl.sync.bfly.b32 %r485|%p33, %r483, %r467, %r453, %r455;
mov.b32 %f173, %r485;
add.ftz.f32 %f28, %f172, %f173;
@%p22 bra $L__BB258_14;

add.s32 %r486, %r27, 1;
cvt.u64.u32 %rd76, %r486;
add.s64 %rd77, %rd76, %rd13;
shl.b64 %rd78, %rd77, 2;
add.s64 %rd79, %rd1, %rd78;
st.global.f32 [%rd79], %f28;

$L__BB258_14:
shl.b32 %r487, %r27, 1;
mov.u32 %r488, 1;
ld.shared.f32 %f174, [%r21+512];
add.ftz.f32 %f175, %f174, %f245;
ld.shared.f32 %f176, [%r21+1536];
add.ftz.f32 %f177, %f176, %f175;
ld.shared.f32 %f178, [%r21+2560];
add.ftz.f32 %f179, %f178, %f177;
mov.b32 %r489, %f179;
mov.u32 %r490, 31;
mov.u32 %r491, 16;
mov.u32 %r492, -1;
shfl.sync.bfly.b32 %r493|%p35, %r489, %r491, %r490, %r492;
mov.b32 %f180, %r493;
add.ftz.f32 %f181, %f179, %f180;
mov.b32 %r494, %f181;
mov.u32 %r495, 8;
shfl.sync.bfly.b32 %r496|%p36, %r494, %r495, %r490, %r492;
mov.b32 %f182, %r496;
add.ftz.f32 %f183, %f181, %f182;
mov.b32 %r497, %f183;
mov.u32 %r498, 4;
shfl.sync.bfly.b32 %r499|%p37, %r497, %r498, %r490, %r492;
mov.b32 %f184, %r499;
add.ftz.f32 %f185, %f183, %f184;
mov.b32 %r500, %f185;
mov.u32 %r501, 2;
shfl.sync.bfly.b32 %r502|%p38, %r500, %r501, %r490, %r492;
mov.b32 %f186, %r502;
add.ftz.f32 %f187, %f185, %f186;
mov.b32 %r503, %f187;
shfl.sync.bfly.b32 %r504|%p39, %r503, %r488, %r490, %r492;
mov.b32 %f188, %r504;
add.ftz.f32 %f29, %f187, %f188;
cvt.u64.u32 %rd80, %r487;
add.s64 %rd81, %rd80, %rd13;
shl.b64 %rd82, %rd81, 2;
add.s64 %rd15, %rd1, %rd82;
@%p14 bra $L__BB258_16;

st.global.f32 [%rd15], %f29;

$L__BB258_16:
ld.shared.f32 %f189, [%r21+640];
add.ftz.f32 %f190, %f189, %f244;
ld.shared.f32 %f191, [%r21+1664];
add.ftz.f32 %f192, %f191, %f190;
ld.shared.f32 %f193, [%r21+2688];
add.ftz.f32 %f194, %f193, %f192;
mov.b32 %r505, %f194;
shfl.sync.bfly.b32 %r509|%p41, %r505, %r491, %r490, %r492;
mov.b32 %f195, %r509;
add.ftz.f32 %f196, %f194, %f195;
mov.b32 %r510, %f196;
shfl.sync.bfly.b32 %r512|%p42, %r510, %r495, %r490, %r492;
mov.b32 %f197, %r512;
add.ftz.f32 %f198, %f196, %f197;
mov.b32 %r513, %f198;
shfl.sync.bfly.b32 %r515|%p43, %r513, %r498, %r490, %r492;
mov.b32 %f199, %r515;
add.ftz.f32 %f200, %f198, %f199;
mov.b32 %r516, %f200;
shfl.sync.bfly.b32 %r518|%p44, %r516, %r501, %r490, %r492;
mov.b32 %f201, %r518;
add.ftz.f32 %f202, %f200, %f201;
mov.b32 %r519, %f202;
shfl.sync.bfly.b32 %r521|%p45, %r519, %r488, %r490, %r492;
mov.b32 %f203, %r521;
add.ftz.f32 %f30, %f202, %f203;
@%p22 bra $L__BB258_18;

st.global.f32 [%rd15+4], %f30;

$L__BB258_18:
mul.lo.s32 %r23, %r27, 3;
ld.shared.f32 %f204, [%r21+768];
add.ftz.f32 %f205, %f204, %f243;
ld.shared.f32 %f206, [%r21+1792];
add.ftz.f32 %f207, %f206, %f205;
ld.shared.f32 %f208, [%r21+2816];
add.ftz.f32 %f209, %f208, %f207;
mov.b32 %r522, %f209;
mov.u32 %r523, 31;
mov.u32 %r524, 16;
mov.u32 %r525, -1;
shfl.sync.bfly.b32 %r526|%p47, %r522, %r524, %r523, %r525;
mov.b32 %f210, %r526;
add.ftz.f32 %f211, %f209, %f210;
mov.b32 %r527, %f211;
mov.u32 %r528, 8;
shfl.sync.bfly.b32 %r529|%p48, %r527, %r528, %r523, %r525;
mov.b32 %f212, %r529;
add.ftz.f32 %f213, %f211, %f212;
mov.b32 %r530, %f213;
mov.u32 %r531, 4;
shfl.sync.bfly.b32 %r532|%p49, %r530, %r531, %r523, %r525;
mov.b32 %f214, %r532;
add.ftz.f32 %f215, %f213, %f214;
mov.b32 %r533, %f215;
mov.u32 %r534, 2;
shfl.sync.bfly.b32 %r535|%p50, %r533, %r534, %r523, %r525;
mov.b32 %f216, %r535;
add.ftz.f32 %f217, %f215, %f216;
mov.b32 %r536, %f217;
mov.u32 %r537, 1;
shfl.sync.bfly.b32 %r538|%p51, %r536, %r537, %r523, %r525;
mov.b32 %f218, %r538;
add.ftz.f32 %f31, %f217, %f218;
@%p14 bra $L__BB258_20;

cvt.u64.u32 %rd83, %r23;
add.s64 %rd84, %rd83, %rd13;
shl.b64 %rd85, %rd84, 2;
add.s64 %rd86, %rd1, %rd85;
st.global.f32 [%rd86], %f31;

$L__BB258_20:
ld.shared.f32 %f219, [%r21+896];
add.ftz.f32 %f220, %f219, %f242;
ld.shared.f32 %f221, [%r21+1920];
add.ftz.f32 %f222, %f221, %f220;
ld.shared.f32 %f223, [%r21+2944];
add.ftz.f32 %f224, %f223, %f222;
mov.b32 %r539, %f224;
shfl.sync.bfly.b32 %r543|%p53, %r539, %r524, %r523, %r525;
mov.b32 %f225, %r543;
add.ftz.f32 %f226, %f224, %f225;
mov.b32 %r544, %f226;
shfl.sync.bfly.b32 %r546|%p54, %r544, %r528, %r523, %r525;
mov.b32 %f227, %r546;
add.ftz.f32 %f228, %f226, %f227;
mov.b32 %r547, %f228;
shfl.sync.bfly.b32 %r549|%p55, %r547, %r531, %r523, %r525;
mov.b32 %f229, %r549;
add.ftz.f32 %f230, %f228, %f229;
mov.b32 %r550, %f230;
shfl.sync.bfly.b32 %r552|%p56, %r550, %r534, %r523, %r525;
mov.b32 %f231, %r552;
add.ftz.f32 %f232, %f230, %f231;
mov.b32 %r553, %f232;
shfl.sync.bfly.b32 %r555|%p57, %r553, %r537, %r523, %r525;
mov.b32 %f233, %r555;
add.ftz.f32 %f32, %f232, %f233;
@%p22 bra $L__BB258_22;

add.s32 %r556, %r23, 1;
cvt.u64.u32 %rd87, %r556;
add.s64 %rd88, %rd87, %rd13;
shl.b64 %rd89, %rd88, 2;
add.s64 %rd90, %rd1, %rd89;
st.global.f32 [%rd90], %f32;

$L__BB258_22:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type19ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type19ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type19ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type19ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type19ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type19ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type19ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type19ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type19ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<71>;
.reg .b16 %rs<25>;
.reg .f32 %f<270>;
.reg .b32 %r<666>;
.reg .b64 %rd<107>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type19ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1280];

ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type19ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type19ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type19ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type19ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type19ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type19ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type19ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type19ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type19ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type19ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type19ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type19ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type19ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type19ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type19ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type19ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type19ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r42, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r42, %r2;
mov.u32 %r43, %ctaid.x;
shl.b32 %r4, %r43, 1;
shr.s32 %r44, %r3, 31;
shr.u32 %r45, %r44, 29;
add.s32 %r46, %r3, %r45;
shr.s32 %r665, %r46, 3;
shr.u32 %r6, %r41, 8;
setp.ge.s32 %p3, %r665, %r6;
mov.f32 %f260, 0f00000000;
mov.f32 %f261, %f260;
mov.f32 %f262, %f260;
mov.f32 %f263, %f260;
mov.f32 %f264, %f260;
mov.f32 %f265, %f260;
mov.f32 %f266, %f260;
mov.f32 %f267, %f260;
mov.f32 %f268, %f260;
mov.f32 %f269, %f260;
@%p3 bra $L__BB259_3;

mov.u32 %r47, %ctaid.y;
mov.u32 %r48, %ctaid.z;
mul.lo.s32 %r49, %r47, %r33;
cvt.u64.u32 %rd20, %r49;
mul.lo.s32 %r50, %r48, %r39;
cvt.u64.u32 %rd21, %r50;
add.s64 %rd1, %rd21, %rd20;
mul.hi.u32 %r51, %r48, %r35;
add.s32 %r52, %r51, %r48;
shr.u32 %r53, %r52, %r36;
mul.lo.s32 %r54, %r53, %r38;
mul.hi.u32 %r55, %r47, %r29;
add.s32 %r56, %r55, %r47;
shr.u32 %r57, %r56, %r30;
mul.lo.s32 %r58, %r57, %r32;
cvta.to.global.u64 %rd106, %rd17;
and.b32 %r62, %r46, -8;
sub.s32 %r63, %r3, %r62;
shl.b32 %r64, %r63, 1;
cvt.s64.s32 %rd3, %r63;
shl.b32 %r65, %r27, 1;
shl.b32 %r660, %r665, 3;
add.s32 %r664, %r660, %r65;
mad.lo.s32 %r663, %r27, 3, %r660;
shl.b32 %r66, %r27, 2;
add.s32 %r662, %r660, %r66;
add.s32 %r661, %r27, %r660;
add.s32 %r67, %r665, %r54;
add.s32 %r68, %r67, %r58;
add.s32 %r69, %r4, 1;
mad.lo.s32 %r70, %r26, %r69, %r68;
mul.wide.s32 %rd4, %r70, 50;
mul.wide.s32 %rd22, %r63, 2;
add.s64 %rd23, %rd4, %rd22;
add.s64 %rd5, %rd23, 34;
mad.lo.s32 %r71, %r4, %r26, %r665;
add.s32 %r72, %r71, %r54;
add.s32 %r73, %r72, %r58;
mul.wide.s32 %rd6, %r73, 50;
mul.wide.s32 %rd24, %r64, 2;
add.s64 %rd25, %rd6, %rd24;
add.s64 %rd7, %rd25, 4;
add.s64 %rd26, %rd4, %rd24;
add.s64 %rd8, %rd26, 2;
add.s64 %rd27, %rd6, %rd22;
add.s64 %rd9, %rd27, 34;
cvta.to.global.u64 %rd10, %rd18;
mov.f32 %f260, 0f00000000;

$L__BB259_2:
cvt.u64.u32 %rd28, %r660;
add.s64 %rd29, %rd1, %rd28;
add.s64 %rd30, %rd29, %rd3;
mul.lo.s64 %rd31, %rd30, 36;
add.s64 %rd32, %rd106, %rd7;
ld.global.nc.u16 %rs11, [%rd32+-2];
shr.u16 %rs12, %rs11, 8;
cvt.u32.u16 %r414, %rs12;
ld.global.nc.u16 %rs13, [%rd32];
shr.u16 %rs14, %rs13, 8;
cvt.u32.u16 %r415, %rs14;
add.s64 %rd33, %rd106, %rd9;
ld.global.nc.u16 %rs15, [%rd33];
cvt.u32.u16 %r416, %rs15;
and.b16 %rs16, %rs11, 255;
cvt.u32.u16 %r417, %rs16;
mul.wide.u16 %r418, %rs15, 256;
and.b32 %r419, %r418, 1792;
or.b32 %r420, %r419, %r417;
mul.wide.u32 %rd34, %r420, 4;
mov.u64 %rd35, iq1s_grid_gpu;
add.s64 %rd36, %rd35, %rd34;
ld.global.nc.u32 %r421, [%rd36];
and.b32 %r347, %r421, 252645135;
shr.u32 %r422, %r421, 4;
and.b32 %r351, %r422, 252645135;
add.s64 %rd37, %rd10, %rd31;
ld.global.nc.u32 %r76, [%rd37+4];
ld.global.nc.u32 %r80, [%rd37+8];
mov.u32 %r383, 0;

dp4a.s32.s32 %r74, %r347, %r76, %r383;


dp4a.s32.s32 %r78, %r351, %r80, %r74;

mul.wide.u16 %r423, %rs15, 32;
and.b32 %r424, %r423, 1792;
or.b32 %r425, %r424, %r414;
mul.wide.u32 %rd38, %r425, 4;
add.s64 %rd39, %rd35, %rd38;
ld.global.nc.u32 %r426, [%rd39];
and.b32 %r355, %r426, 252645135;
shr.u32 %r427, %r426, 4;
and.b32 %r359, %r427, 252645135;
ld.global.nc.u32 %r84, [%rd37+12];
ld.global.nc.u32 %r88, [%rd37+16];

dp4a.s32.s32 %r82, %r355, %r84, %r78;


dp4a.s32.s32 %r86, %r359, %r88, %r82;

and.b16 %rs17, %rs13, 255;
cvt.u32.u16 %r428, %rs17;
mul.wide.u16 %r429, %rs15, 4;
and.b32 %r430, %r429, 1792;
or.b32 %r431, %r430, %r428;
mul.wide.u32 %rd40, %r431, 4;
add.s64 %rd41, %rd35, %rd40;
ld.global.nc.u32 %r432, [%rd41];
and.b32 %r363, %r432, 252645135;
shr.u32 %r433, %r432, 4;
and.b32 %r367, %r433, 252645135;
ld.global.nc.u32 %r92, [%rd37+20];
ld.global.nc.u32 %r96, [%rd37+24];

dp4a.s32.s32 %r90, %r363, %r92, %r86;


dp4a.s32.s32 %r94, %r367, %r96, %r90;

shr.u32 %r434, %r416, 1;
and.b32 %r435, %r434, 1792;
or.b32 %r436, %r435, %r415;
mul.wide.u32 %rd42, %r436, 4;
add.s64 %rd43, %rd35, %rd42;
ld.global.nc.u32 %r437, [%rd43];
and.b32 %r371, %r437, 252645135;
shr.u32 %r438, %r437, 4;
and.b32 %r375, %r438, 252645135;
ld.global.nc.u32 %r100, [%rd37+28];
ld.global.nc.u32 %r104, [%rd37+32];

dp4a.s32.s32 %r98, %r371, %r100, %r94;


dp4a.s32.s32 %r102, %r375, %r104, %r98;

add.s64 %rd44, %rd106, %rd6;
ld.global.nc.u16 %rs1, [%rd44];

{ cvt.f32.f16 %f61, %rs1;}


shr.u32 %r439, %r416, 11;
and.b32 %r440, %r439, 14;
or.b32 %r441, %r440, 1;
cvt.rn.f32.s32 %f91, %r441;
mul.ftz.f32 %f92, %f61, %f91;
and.b32 %r442, %r416, 32768;
cvt.rn.f32.s32 %f93, %r442;
mul.ftz.f32 %f94, %f93, 0f37000000;
mov.f32 %f95, 0fBF600000;
sub.ftz.f32 %f96, %f95, %f94;
ld.global.nc.u32 %r106, [%rd37];

{.reg .f16 low,high;
mov.b32 {low,high},%r106;
cvt.f32.f16 %f62, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r106;
cvt.f32.f16 %f63, high;}


cvt.rn.f32.s32 %f97, %r102;
mul.ftz.f32 %f98, %f96, %f63;
fma.rn.ftz.f32 %f99, %f62, %f97, %f98;
fma.rn.ftz.f32 %f269, %f92, %f99, %f269;
add.s64 %rd45, %rd106, %rd8;
ld.global.nc.u16 %rs18, [%rd45];
shr.u16 %rs19, %rs18, 8;
cvt.u32.u16 %r443, %rs19;
ld.global.nc.u16 %rs20, [%rd45+2];
shr.u16 %rs21, %rs20, 8;
cvt.u32.u16 %r444, %rs21;
add.s64 %rd46, %rd106, %rd5;
ld.global.nc.u16 %rs22, [%rd46];
cvt.u32.u16 %r445, %rs22;
and.b16 %rs23, %rs18, 255;
cvt.u32.u16 %r446, %rs23;
mul.wide.u16 %r447, %rs22, 256;
and.b32 %r448, %r447, 1792;
or.b32 %r449, %r448, %r446;
mul.wide.u32 %rd47, %r449, 4;
add.s64 %rd48, %rd35, %rd47;
ld.global.nc.u32 %r450, [%rd48];
and.b32 %r381, %r450, 252645135;
shr.u32 %r451, %r450, 4;
and.b32 %r385, %r451, 252645135;

dp4a.s32.s32 %r108, %r381, %r76, %r383;


dp4a.s32.s32 %r112, %r385, %r80, %r108;

mul.wide.u16 %r452, %rs22, 32;
and.b32 %r453, %r452, 1792;
or.b32 %r454, %r453, %r443;
mul.wide.u32 %rd49, %r454, 4;
add.s64 %rd50, %rd35, %rd49;
ld.global.nc.u32 %r455, [%rd50];
and.b32 %r389, %r455, 252645135;
shr.u32 %r456, %r455, 4;
and.b32 %r393, %r456, 252645135;

dp4a.s32.s32 %r116, %r389, %r84, %r112;


dp4a.s32.s32 %r120, %r393, %r88, %r116;

and.b16 %rs24, %rs20, 255;
cvt.u32.u16 %r457, %rs24;
mul.wide.u16 %r458, %rs22, 4;
and.b32 %r459, %r458, 1792;
or.b32 %r460, %r459, %r457;
mul.wide.u32 %rd51, %r460, 4;
add.s64 %rd52, %rd35, %rd51;
ld.global.nc.u32 %r461, [%rd52];
and.b32 %r397, %r461, 252645135;
shr.u32 %r462, %r461, 4;
and.b32 %r401, %r462, 252645135;

dp4a.s32.s32 %r124, %r397, %r92, %r120;


dp4a.s32.s32 %r128, %r401, %r96, %r124;

shr.u32 %r463, %r445, 1;
and.b32 %r464, %r463, 1792;
or.b32 %r465, %r464, %r444;
mul.wide.u32 %rd53, %r465, 4;
add.s64 %rd54, %rd35, %rd53;
ld.global.nc.u32 %r466, [%rd54];
and.b32 %r405, %r466, 252645135;
shr.u32 %r467, %r466, 4;
and.b32 %r409, %r467, 252645135;

dp4a.s32.s32 %r132, %r405, %r100, %r128;


dp4a.s32.s32 %r136, %r409, %r104, %r132;

add.s64 %rd55, %rd106, %rd4;
ld.global.nc.u16 %rs2, [%rd55];

{ cvt.f32.f16 %f64, %rs2;}


shr.u32 %r468, %r445, 11;
and.b32 %r469, %r468, 14;
or.b32 %r470, %r469, 1;
cvt.rn.f32.s32 %f100, %r470;
mul.ftz.f32 %f101, %f64, %f100;
and.b32 %r471, %r445, 32768;
cvt.rn.f32.s32 %f102, %r471;
mul.ftz.f32 %f103, %f102, 0f37000000;
sub.ftz.f32 %f104, %f95, %f103;

{.reg .f16 low,high;
mov.b32 {low,high},%r106;
cvt.f32.f16 %f65, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r106;
cvt.f32.f16 %f66, high;}


cvt.rn.f32.s32 %f105, %r136;
mul.ftz.f32 %f106, %f104, %f66;
fma.rn.ftz.f32 %f107, %f65, %f105, %f106;
fma.rn.ftz.f32 %f268, %f101, %f107, %f268;
cvt.u64.u32 %rd56, %r661;
add.s64 %rd57, %rd1, %rd56;
add.s64 %rd58, %rd57, %rd3;
mul.lo.s64 %rd59, %rd58, 36;
add.s64 %rd60, %rd10, %rd59;
ld.global.nc.u32 %r144, [%rd60+4];
ld.global.nc.u32 %r148, [%rd60+8];

dp4a.s32.s32 %r142, %r347, %r144, %r383;


dp4a.s32.s32 %r146, %r351, %r148, %r142;

ld.global.nc.u32 %r152, [%rd60+12];
ld.global.nc.u32 %r156, [%rd60+16];

dp4a.s32.s32 %r150, %r355, %r152, %r146;


dp4a.s32.s32 %r154, %r359, %r156, %r150;

ld.global.nc.u32 %r160, [%rd60+20];
ld.global.nc.u32 %r164, [%rd60+24];

dp4a.s32.s32 %r158, %r363, %r160, %r154;


dp4a.s32.s32 %r162, %r367, %r164, %r158;

ld.global.nc.u32 %r168, [%rd60+28];
ld.global.nc.u32 %r172, [%rd60+32];

dp4a.s32.s32 %r166, %r371, %r168, %r162;


dp4a.s32.s32 %r170, %r375, %r172, %r166;


{ cvt.f32.f16 %f67, %rs1;}


mul.ftz.f32 %f108, %f67, %f91;
ld.global.nc.u32 %r174, [%rd60];

{.reg .f16 low,high;
mov.b32 {low,high},%r174;
cvt.f32.f16 %f68, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r174;
cvt.f32.f16 %f69, high;}


cvt.rn.f32.s32 %f109, %r170;
mul.ftz.f32 %f110, %f96, %f69;
fma.rn.ftz.f32 %f111, %f68, %f109, %f110;
fma.rn.ftz.f32 %f267, %f108, %f111, %f267;

dp4a.s32.s32 %r176, %r381, %r144, %r383;


dp4a.s32.s32 %r180, %r385, %r148, %r176;


dp4a.s32.s32 %r184, %r389, %r152, %r180;


dp4a.s32.s32 %r188, %r393, %r156, %r184;


dp4a.s32.s32 %r192, %r397, %r160, %r188;


dp4a.s32.s32 %r196, %r401, %r164, %r192;


dp4a.s32.s32 %r200, %r405, %r168, %r196;


dp4a.s32.s32 %r204, %r409, %r172, %r200;


{ cvt.f32.f16 %f70, %rs2;}


mul.ftz.f32 %f112, %f70, %f100;

{.reg .f16 low,high;
mov.b32 {low,high},%r174;
cvt.f32.f16 %f71, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r174;
cvt.f32.f16 %f72, high;}


cvt.rn.f32.s32 %f113, %r204;
mul.ftz.f32 %f114, %f104, %f72;
fma.rn.ftz.f32 %f115, %f71, %f113, %f114;
fma.rn.ftz.f32 %f266, %f112, %f115, %f266;
cvt.u64.u32 %rd61, %r664;
add.s64 %rd62, %rd1, %rd61;
add.s64 %rd63, %rd62, %rd3;
mul.lo.s64 %rd64, %rd63, 36;
add.s64 %rd65, %rd10, %rd64;
ld.global.nc.u32 %r212, [%rd65+4];
ld.global.nc.u32 %r216, [%rd65+8];

dp4a.s32.s32 %r210, %r347, %r212, %r383;


dp4a.s32.s32 %r214, %r351, %r216, %r210;

ld.global.nc.u32 %r220, [%rd65+12];
ld.global.nc.u32 %r224, [%rd65+16];

dp4a.s32.s32 %r218, %r355, %r220, %r214;


dp4a.s32.s32 %r222, %r359, %r224, %r218;

ld.global.nc.u32 %r228, [%rd65+20];
ld.global.nc.u32 %r232, [%rd65+24];

dp4a.s32.s32 %r226, %r363, %r228, %r222;


dp4a.s32.s32 %r230, %r367, %r232, %r226;

ld.global.nc.u32 %r236, [%rd65+28];
ld.global.nc.u32 %r240, [%rd65+32];

dp4a.s32.s32 %r234, %r371, %r236, %r230;


dp4a.s32.s32 %r238, %r375, %r240, %r234;


{ cvt.f32.f16 %f73, %rs1;}


mul.ftz.f32 %f116, %f73, %f91;
ld.global.nc.u32 %r242, [%rd65];

{.reg .f16 low,high;
mov.b32 {low,high},%r242;
cvt.f32.f16 %f74, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r242;
cvt.f32.f16 %f75, high;}


cvt.rn.f32.s32 %f117, %r238;
mul.ftz.f32 %f118, %f96, %f75;
fma.rn.ftz.f32 %f119, %f74, %f117, %f118;
fma.rn.ftz.f32 %f265, %f116, %f119, %f265;

dp4a.s32.s32 %r244, %r381, %r212, %r383;


dp4a.s32.s32 %r248, %r385, %r216, %r244;


dp4a.s32.s32 %r252, %r389, %r220, %r248;


dp4a.s32.s32 %r256, %r393, %r224, %r252;


dp4a.s32.s32 %r260, %r397, %r228, %r256;


dp4a.s32.s32 %r264, %r401, %r232, %r260;


dp4a.s32.s32 %r268, %r405, %r236, %r264;


dp4a.s32.s32 %r272, %r409, %r240, %r268;


{ cvt.f32.f16 %f76, %rs2;}


mul.ftz.f32 %f120, %f76, %f100;

{.reg .f16 low,high;
mov.b32 {low,high},%r242;
cvt.f32.f16 %f77, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r242;
cvt.f32.f16 %f78, high;}


cvt.rn.f32.s32 %f121, %r272;
mul.ftz.f32 %f122, %f104, %f78;
fma.rn.ftz.f32 %f123, %f77, %f121, %f122;
fma.rn.ftz.f32 %f264, %f120, %f123, %f264;
cvt.u64.u32 %rd66, %r663;
add.s64 %rd67, %rd1, %rd66;
add.s64 %rd68, %rd67, %rd3;
mul.lo.s64 %rd69, %rd68, 36;
add.s64 %rd70, %rd10, %rd69;
ld.global.nc.u32 %r280, [%rd70+4];
ld.global.nc.u32 %r284, [%rd70+8];

dp4a.s32.s32 %r278, %r347, %r280, %r383;


dp4a.s32.s32 %r282, %r351, %r284, %r278;

ld.global.nc.u32 %r288, [%rd70+12];
ld.global.nc.u32 %r292, [%rd70+16];

dp4a.s32.s32 %r286, %r355, %r288, %r282;


dp4a.s32.s32 %r290, %r359, %r292, %r286;

ld.global.nc.u32 %r296, [%rd70+20];
ld.global.nc.u32 %r300, [%rd70+24];

dp4a.s32.s32 %r294, %r363, %r296, %r290;


dp4a.s32.s32 %r298, %r367, %r300, %r294;

ld.global.nc.u32 %r304, [%rd70+28];
ld.global.nc.u32 %r308, [%rd70+32];

dp4a.s32.s32 %r302, %r371, %r304, %r298;


dp4a.s32.s32 %r306, %r375, %r308, %r302;


{ cvt.f32.f16 %f79, %rs1;}


mul.ftz.f32 %f124, %f79, %f91;
ld.global.nc.u32 %r310, [%rd70];

{.reg .f16 low,high;
mov.b32 {low,high},%r310;
cvt.f32.f16 %f80, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r310;
cvt.f32.f16 %f81, high;}


cvt.rn.f32.s32 %f125, %r306;
mul.ftz.f32 %f126, %f96, %f81;
fma.rn.ftz.f32 %f127, %f80, %f125, %f126;
fma.rn.ftz.f32 %f263, %f124, %f127, %f263;

dp4a.s32.s32 %r312, %r381, %r280, %r383;


dp4a.s32.s32 %r316, %r385, %r284, %r312;


dp4a.s32.s32 %r320, %r389, %r288, %r316;


dp4a.s32.s32 %r324, %r393, %r292, %r320;


dp4a.s32.s32 %r328, %r397, %r296, %r324;


dp4a.s32.s32 %r332, %r401, %r300, %r328;


dp4a.s32.s32 %r336, %r405, %r304, %r332;


dp4a.s32.s32 %r340, %r409, %r308, %r336;


{ cvt.f32.f16 %f82, %rs2;}


mul.ftz.f32 %f128, %f82, %f100;

{.reg .f16 low,high;
mov.b32 {low,high},%r310;
cvt.f32.f16 %f83, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r310;
cvt.f32.f16 %f84, high;}


cvt.rn.f32.s32 %f129, %r340;
mul.ftz.f32 %f130, %f104, %f84;
fma.rn.ftz.f32 %f131, %f83, %f129, %f130;
fma.rn.ftz.f32 %f262, %f128, %f131, %f262;
cvt.u64.u32 %rd71, %r662;
add.s64 %rd72, %rd1, %rd71;
add.s64 %rd73, %rd72, %rd3;
mul.lo.s64 %rd74, %rd73, 36;
add.s64 %rd75, %rd10, %rd74;
ld.global.nc.u32 %r348, [%rd75+4];
ld.global.nc.u32 %r352, [%rd75+8];

dp4a.s32.s32 %r346, %r347, %r348, %r383;


dp4a.s32.s32 %r350, %r351, %r352, %r346;

ld.global.nc.u32 %r356, [%rd75+12];
ld.global.nc.u32 %r360, [%rd75+16];

dp4a.s32.s32 %r354, %r355, %r356, %r350;


dp4a.s32.s32 %r358, %r359, %r360, %r354;

ld.global.nc.u32 %r364, [%rd75+20];
ld.global.nc.u32 %r368, [%rd75+24];

dp4a.s32.s32 %r362, %r363, %r364, %r358;


dp4a.s32.s32 %r366, %r367, %r368, %r362;

ld.global.nc.u32 %r372, [%rd75+28];
ld.global.nc.u32 %r376, [%rd75+32];

dp4a.s32.s32 %r370, %r371, %r372, %r366;


dp4a.s32.s32 %r374, %r375, %r376, %r370;


{ cvt.f32.f16 %f85, %rs1;}


mul.ftz.f32 %f132, %f85, %f91;
ld.global.nc.u32 %r378, [%rd75];

{.reg .f16 low,high;
mov.b32 {low,high},%r378;
cvt.f32.f16 %f86, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r378;
cvt.f32.f16 %f87, high;}


cvt.rn.f32.s32 %f133, %r374;
mul.ftz.f32 %f134, %f96, %f87;
fma.rn.ftz.f32 %f135, %f86, %f133, %f134;
fma.rn.ftz.f32 %f261, %f132, %f135, %f261;

dp4a.s32.s32 %r380, %r381, %r348, %r383;


dp4a.s32.s32 %r384, %r385, %r352, %r380;


dp4a.s32.s32 %r388, %r389, %r356, %r384;


dp4a.s32.s32 %r392, %r393, %r360, %r388;


dp4a.s32.s32 %r396, %r397, %r364, %r392;


dp4a.s32.s32 %r400, %r401, %r368, %r396;


dp4a.s32.s32 %r404, %r405, %r372, %r400;


dp4a.s32.s32 %r408, %r409, %r376, %r404;


{ cvt.f32.f16 %f88, %rs2;}


mul.ftz.f32 %f136, %f88, %f100;

{.reg .f16 low,high;
mov.b32 {low,high},%r378;
cvt.f32.f16 %f89, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r378;
cvt.f32.f16 %f90, high;}


cvt.rn.f32.s32 %f137, %r408;
mul.ftz.f32 %f138, %f104, %f90;
fma.rn.ftz.f32 %f139, %f89, %f137, %f138;
fma.rn.ftz.f32 %f260, %f136, %f139, %f260;
add.s32 %r664, %r664, 64;
add.s32 %r663, %r663, 64;
add.s32 %r662, %r662, 64;
add.s32 %r661, %r661, 64;
add.s32 %r660, %r660, 64;
add.s64 %rd106, %rd106, 400;
add.s32 %r665, %r665, 8;
setp.lt.s32 %p4, %r665, %r6;
@%p4 bra $L__BB259_2;

$L__BB259_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB259_5;

mov.u32 %r472, _ZZ13mul_mat_vec_qIL9ggml_type19ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r473, %r1, 1280, %r472;
shl.b32 %r474, %r2, 2;
add.s32 %r475, %r473, %r474;
st.shared.f32 [%r475+-1280], %f269;
st.shared.f32 [%r475+-1152], %f268;
st.shared.f32 [%r475+-1024], %f267;
st.shared.f32 [%r475+-896], %f266;
st.shared.f32 [%r475+-768], %f265;
st.shared.f32 [%r475+-640], %f264;
st.shared.f32 [%r475+-512], %f263;
st.shared.f32 [%r475+-384], %f262;
st.shared.f32 [%r475+-256], %f261;
st.shared.f32 [%r475+-128], %f260;

$L__BB259_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB259_26;

mov.u32 %r476, %ctaid.z;
mov.u32 %r477, %ctaid.y;
mad.lo.s32 %r478, %r477, %r34, %r4;
mad.lo.s32 %r479, %r476, %r40, %r478;
cvt.u64.u32 %rd13, %r479;
shl.b32 %r480, %r2, 2;
mov.u32 %r481, _ZZ13mul_mat_vec_qIL9ggml_type19ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r24, %r481, %r480;
mov.u32 %r482, 2;
ld.shared.f32 %f140, [%r24];
add.ftz.f32 %f141, %f140, %f269;
mov.b32 %r483, %f141;
mov.u32 %r484, 31;
mov.u32 %r485, 16;
mov.u32 %r486, -1;
shfl.sync.bfly.b32 %r487|%p7, %r483, %r485, %r484, %r486;
mov.b32 %f142, %r487;
add.ftz.f32 %f143, %f141, %f142;
mov.b32 %r488, %f143;
mov.u32 %r489, 8;
shfl.sync.bfly.b32 %r490|%p8, %r488, %r489, %r484, %r486;
mov.b32 %f144, %r490;
add.ftz.f32 %f145, %f143, %f144;
mov.b32 %r491, %f145;
mov.u32 %r492, 4;
shfl.sync.bfly.b32 %r493|%p9, %r491, %r492, %r484, %r486;
mov.b32 %f146, %r493;
add.ftz.f32 %f147, %f145, %f146;
mov.b32 %r494, %f147;
shfl.sync.bfly.b32 %r495|%p10, %r494, %r482, %r484, %r486;
mov.b32 %f148, %r495;
add.ftz.f32 %f149, %f147, %f148;
mov.b32 %r496, %f149;
mov.u32 %r497, 1;
shfl.sync.bfly.b32 %r498|%p11, %r496, %r497, %r484, %r486;
mov.b32 %f150, %r498;
add.ftz.f32 %f31, %f149, %f150;
setp.lt.u32 %p12, %r4, %r28;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
cvta.to.global.u64 %rd76, %rd19;
mul.wide.u32 %rd77, %r479, 4;
add.s64 %rd14, %rd76, %rd77;
not.pred %p14, %p1;
@%p14 bra $L__BB259_8;

st.global.f32 [%rd14], %f31;

$L__BB259_8:
ld.shared.f32 %f151, [%r24+128];
add.ftz.f32 %f152, %f151, %f268;
mov.b32 %r499, %f152;
shfl.sync.bfly.b32 %r503|%p15, %r499, %r485, %r484, %r486;
mov.b32 %f153, %r503;
add.ftz.f32 %f154, %f152, %f153;
mov.b32 %r504, %f154;
shfl.sync.bfly.b32 %r506|%p16, %r504, %r489, %r484, %r486;
mov.b32 %f155, %r506;
add.ftz.f32 %f156, %f154, %f155;
mov.b32 %r507, %f156;
shfl.sync.bfly.b32 %r509|%p17, %r507, %r492, %r484, %r486;
mov.b32 %f157, %r509;
add.ftz.f32 %f158, %f156, %f157;
mov.b32 %r510, %f158;
shfl.sync.bfly.b32 %r512|%p18, %r510, %r482, %r484, %r486;
mov.b32 %f159, %r512;
add.ftz.f32 %f160, %f158, %f159;
mov.b32 %r513, %f160;
shfl.sync.bfly.b32 %r515|%p19, %r513, %r497, %r484, %r486;
mov.b32 %f161, %r515;
add.ftz.f32 %f32, %f160, %f161;
setp.eq.s32 %p20, %r2, 1;
or.b32 %r519, %r4, 1;
setp.lt.u32 %p21, %r519, %r28;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB259_10;

st.global.f32 [%rd14+4], %f32;

$L__BB259_10:
ld.shared.f32 %f162, [%r24+256];
add.ftz.f32 %f163, %f162, %f267;
mov.b32 %r520, %f163;
mov.u32 %r521, 31;
mov.u32 %r522, 16;
mov.u32 %r523, -1;
shfl.sync.bfly.b32 %r524|%p23, %r520, %r522, %r521, %r523;
mov.b32 %f164, %r524;
add.ftz.f32 %f165, %f163, %f164;
mov.b32 %r525, %f165;
mov.u32 %r526, 8;
shfl.sync.bfly.b32 %r527|%p24, %r525, %r526, %r521, %r523;
mov.b32 %f166, %r527;
add.ftz.f32 %f167, %f165, %f166;
mov.b32 %r528, %f167;
mov.u32 %r529, 4;
shfl.sync.bfly.b32 %r530|%p25, %r528, %r529, %r521, %r523;
mov.b32 %f168, %r530;
add.ftz.f32 %f169, %f167, %f168;
mov.b32 %r531, %f169;
mov.u32 %r532, 2;
shfl.sync.bfly.b32 %r533|%p26, %r531, %r532, %r521, %r523;
mov.b32 %f170, %r533;
add.ftz.f32 %f171, %f169, %f170;
mov.b32 %r534, %f171;
mov.u32 %r535, 1;
shfl.sync.bfly.b32 %r536|%p27, %r534, %r535, %r521, %r523;
mov.b32 %f172, %r536;
add.ftz.f32 %f33, %f171, %f172;
@%p14 bra $L__BB259_12;

cvt.u64.u32 %rd78, %r28;
add.s64 %rd79, %rd78, %rd13;
shl.b64 %rd81, %rd79, 2;
add.s64 %rd82, %rd76, %rd81;
st.global.f32 [%rd82], %f33;

$L__BB259_12:
ld.shared.f32 %f173, [%r24+384];
add.ftz.f32 %f174, %f173, %f266;
mov.b32 %r537, %f174;
shfl.sync.bfly.b32 %r541|%p29, %r537, %r522, %r521, %r523;
mov.b32 %f175, %r541;
add.ftz.f32 %f176, %f174, %f175;
mov.b32 %r542, %f176;
shfl.sync.bfly.b32 %r544|%p30, %r542, %r526, %r521, %r523;
mov.b32 %f177, %r544;
add.ftz.f32 %f178, %f176, %f177;
mov.b32 %r545, %f178;
shfl.sync.bfly.b32 %r547|%p31, %r545, %r529, %r521, %r523;
mov.b32 %f179, %r547;
add.ftz.f32 %f180, %f178, %f179;
mov.b32 %r548, %f180;
shfl.sync.bfly.b32 %r550|%p32, %r548, %r532, %r521, %r523;
mov.b32 %f181, %r550;
add.ftz.f32 %f182, %f180, %f181;
mov.b32 %r551, %f182;
shfl.sync.bfly.b32 %r553|%p33, %r551, %r535, %r521, %r523;
mov.b32 %f183, %r553;
add.ftz.f32 %f34, %f182, %f183;
@%p22 bra $L__BB259_14;

add.s32 %r554, %r28, 1;
cvt.u64.u32 %rd83, %r554;
add.s64 %rd84, %rd83, %rd13;
shl.b64 %rd86, %rd84, 2;
add.s64 %rd87, %rd76, %rd86;
st.global.f32 [%rd87], %f34;

$L__BB259_14:
shl.b32 %r555, %r28, 1;
mov.u32 %r556, 1;
ld.shared.f32 %f184, [%r24+512];
add.ftz.f32 %f185, %f184, %f265;
mov.b32 %r557, %f185;
mov.u32 %r558, 31;
mov.u32 %r559, 16;
mov.u32 %r560, -1;
shfl.sync.bfly.b32 %r561|%p35, %r557, %r559, %r558, %r560;
mov.b32 %f186, %r561;
add.ftz.f32 %f187, %f185, %f186;
mov.b32 %r562, %f187;
mov.u32 %r563, 8;
shfl.sync.bfly.b32 %r564|%p36, %r562, %r563, %r558, %r560;
mov.b32 %f188, %r564;
add.ftz.f32 %f189, %f187, %f188;
mov.b32 %r565, %f189;
mov.u32 %r566, 4;
shfl.sync.bfly.b32 %r567|%p37, %r565, %r566, %r558, %r560;
mov.b32 %f190, %r567;
add.ftz.f32 %f191, %f189, %f190;
mov.b32 %r568, %f191;
mov.u32 %r569, 2;
shfl.sync.bfly.b32 %r570|%p38, %r568, %r569, %r558, %r560;
mov.b32 %f192, %r570;
add.ftz.f32 %f193, %f191, %f192;
mov.b32 %r571, %f193;
shfl.sync.bfly.b32 %r572|%p39, %r571, %r556, %r558, %r560;
mov.b32 %f194, %r572;
add.ftz.f32 %f35, %f193, %f194;
cvt.u64.u32 %rd88, %r555;
add.s64 %rd89, %rd88, %rd13;
shl.b64 %rd91, %rd89, 2;
add.s64 %rd15, %rd76, %rd91;
@%p14 bra $L__BB259_16;

st.global.f32 [%rd15], %f35;

$L__BB259_16:
ld.shared.f32 %f195, [%r24+640];
add.ftz.f32 %f196, %f195, %f264;
mov.b32 %r573, %f196;
shfl.sync.bfly.b32 %r577|%p41, %r573, %r559, %r558, %r560;
mov.b32 %f197, %r577;
add.ftz.f32 %f198, %f196, %f197;
mov.b32 %r578, %f198;
shfl.sync.bfly.b32 %r580|%p42, %r578, %r563, %r558, %r560;
mov.b32 %f199, %r580;
add.ftz.f32 %f200, %f198, %f199;
mov.b32 %r581, %f200;
shfl.sync.bfly.b32 %r583|%p43, %r581, %r566, %r558, %r560;
mov.b32 %f201, %r583;
add.ftz.f32 %f202, %f200, %f201;
mov.b32 %r584, %f202;
shfl.sync.bfly.b32 %r586|%p44, %r584, %r569, %r558, %r560;
mov.b32 %f203, %r586;
add.ftz.f32 %f204, %f202, %f203;
mov.b32 %r587, %f204;
shfl.sync.bfly.b32 %r589|%p45, %r587, %r556, %r558, %r560;
mov.b32 %f205, %r589;
add.ftz.f32 %f36, %f204, %f205;
@%p22 bra $L__BB259_18;

st.global.f32 [%rd15+4], %f36;

$L__BB259_18:
mul.lo.s32 %r25, %r28, 3;
ld.shared.f32 %f206, [%r24+768];
add.ftz.f32 %f207, %f206, %f263;
mov.b32 %r590, %f207;
mov.u32 %r591, 31;
mov.u32 %r592, 16;
mov.u32 %r593, -1;
shfl.sync.bfly.b32 %r594|%p47, %r590, %r592, %r591, %r593;
mov.b32 %f208, %r594;
add.ftz.f32 %f209, %f207, %f208;
mov.b32 %r595, %f209;
mov.u32 %r596, 8;
shfl.sync.bfly.b32 %r597|%p48, %r595, %r596, %r591, %r593;
mov.b32 %f210, %r597;
add.ftz.f32 %f211, %f209, %f210;
mov.b32 %r598, %f211;
mov.u32 %r599, 4;
shfl.sync.bfly.b32 %r600|%p49, %r598, %r599, %r591, %r593;
mov.b32 %f212, %r600;
add.ftz.f32 %f213, %f211, %f212;
mov.b32 %r601, %f213;
mov.u32 %r602, 2;
shfl.sync.bfly.b32 %r603|%p50, %r601, %r602, %r591, %r593;
mov.b32 %f214, %r603;
add.ftz.f32 %f215, %f213, %f214;
mov.b32 %r604, %f215;
mov.u32 %r605, 1;
shfl.sync.bfly.b32 %r606|%p51, %r604, %r605, %r591, %r593;
mov.b32 %f216, %r606;
add.ftz.f32 %f37, %f215, %f216;
@%p14 bra $L__BB259_20;

cvt.u64.u32 %rd92, %r25;
add.s64 %rd93, %rd92, %rd13;
shl.b64 %rd95, %rd93, 2;
add.s64 %rd96, %rd76, %rd95;
st.global.f32 [%rd96], %f37;

$L__BB259_20:
ld.shared.f32 %f217, [%r24+896];
add.ftz.f32 %f218, %f217, %f262;
mov.b32 %r607, %f218;
shfl.sync.bfly.b32 %r611|%p53, %r607, %r592, %r591, %r593;
mov.b32 %f219, %r611;
add.ftz.f32 %f220, %f218, %f219;
mov.b32 %r612, %f220;
shfl.sync.bfly.b32 %r614|%p54, %r612, %r596, %r591, %r593;
mov.b32 %f221, %r614;
add.ftz.f32 %f222, %f220, %f221;
mov.b32 %r615, %f222;
shfl.sync.bfly.b32 %r617|%p55, %r615, %r599, %r591, %r593;
mov.b32 %f223, %r617;
add.ftz.f32 %f224, %f222, %f223;
mov.b32 %r618, %f224;
shfl.sync.bfly.b32 %r620|%p56, %r618, %r602, %r591, %r593;
mov.b32 %f225, %r620;
add.ftz.f32 %f226, %f224, %f225;
mov.b32 %r621, %f226;
shfl.sync.bfly.b32 %r623|%p57, %r621, %r605, %r591, %r593;
mov.b32 %f227, %r623;
add.ftz.f32 %f38, %f226, %f227;
@%p22 bra $L__BB259_22;

add.s32 %r624, %r25, 1;
cvt.u64.u32 %rd97, %r624;
add.s64 %rd98, %rd97, %rd13;
shl.b64 %rd100, %rd98, 2;
add.s64 %rd101, %rd76, %rd100;
st.global.f32 [%rd101], %f38;

$L__BB259_22:
shl.b32 %r625, %r28, 2;
mov.u32 %r626, 2;
ld.shared.f32 %f228, [%r24+1024];
add.ftz.f32 %f229, %f228, %f261;
mov.b32 %r627, %f229;
mov.u32 %r628, 31;
mov.u32 %r629, 16;
mov.u32 %r630, -1;
shfl.sync.bfly.b32 %r631|%p59, %r627, %r629, %r628, %r630;
mov.b32 %f230, %r631;
add.ftz.f32 %f231, %f229, %f230;
mov.b32 %r632, %f231;
mov.u32 %r633, 8;
shfl.sync.bfly.b32 %r634|%p60, %r632, %r633, %r628, %r630;
mov.b32 %f232, %r634;
add.ftz.f32 %f233, %f231, %f232;
mov.b32 %r635, %f233;
mov.u32 %r636, 4;
shfl.sync.bfly.b32 %r637|%p61, %r635, %r636, %r628, %r630;
mov.b32 %f234, %r637;
add.ftz.f32 %f235, %f233, %f234;
mov.b32 %r638, %f235;
shfl.sync.bfly.b32 %r639|%p62, %r638, %r626, %r628, %r630;
mov.b32 %f236, %r639;
add.ftz.f32 %f237, %f235, %f236;
mov.b32 %r640, %f237;
mov.u32 %r641, 1;
shfl.sync.bfly.b32 %r642|%p63, %r640, %r641, %r628, %r630;
mov.b32 %f238, %r642;
add.ftz.f32 %f39, %f237, %f238;
cvt.u64.u32 %rd102, %r625;
add.s64 %rd103, %rd102, %rd13;
shl.b64 %rd105, %rd103, 2;
add.s64 %rd16, %rd76, %rd105;
@%p14 bra $L__BB259_24;

st.global.f32 [%rd16], %f39;

$L__BB259_24:
ld.shared.f32 %f239, [%r24+1152];
add.ftz.f32 %f240, %f239, %f260;
mov.b32 %r643, %f240;
shfl.sync.bfly.b32 %r647|%p65, %r643, %r629, %r628, %r630;
mov.b32 %f241, %r647;
add.ftz.f32 %f242, %f240, %f241;
mov.b32 %r648, %f242;
shfl.sync.bfly.b32 %r650|%p66, %r648, %r633, %r628, %r630;
mov.b32 %f243, %r650;
add.ftz.f32 %f244, %f242, %f243;
mov.b32 %r651, %f244;
shfl.sync.bfly.b32 %r653|%p67, %r651, %r636, %r628, %r630;
mov.b32 %f245, %r653;
add.ftz.f32 %f246, %f244, %f245;
mov.b32 %r654, %f246;
shfl.sync.bfly.b32 %r656|%p68, %r654, %r626, %r628, %r630;
mov.b32 %f247, %r656;
add.ftz.f32 %f248, %f246, %f247;
mov.b32 %r657, %f248;
shfl.sync.bfly.b32 %r659|%p69, %r657, %r641, %r628, %r630;
mov.b32 %f249, %r659;
add.ftz.f32 %f40, %f248, %f249;
@%p22 bra $L__BB259_26;

st.global.f32 [%rd16+4], %f40;

$L__BB259_26:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type19ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type19ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type19ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type19ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type19ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type19ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type19ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type19ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type19ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<83>;
.reg .b16 %rs<27>;
.reg .f32 %f<322>;
.reg .b32 %r<834>;
.reg .b64 %rd<125>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type19ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type19ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type19ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type19ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type19ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type19ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type19ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type19ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type19ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type19ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type19ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type19ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type19ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type19ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type19ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type19ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type19ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type19ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r43, %r1, 5;
mov.u32 %r44, %tid.x;
add.s32 %r2, %r43, %r44;
mov.u32 %r45, %ctaid.x;
shl.b32 %r3, %r45, 1;
shr.s32 %r46, %r2, 31;
shr.u32 %r47, %r46, 29;
add.s32 %r48, %r2, %r47;
shr.s32 %r833, %r48, 3;
shr.u32 %r49, %r27, 8;
setp.ge.s32 %p3, %r833, %r49;
mov.f32 %f310, 0f00000000;
mov.f32 %f311, %f310;
mov.f32 %f312, %f310;
mov.f32 %f313, %f310;
mov.f32 %f314, %f310;
mov.f32 %f315, %f310;
mov.f32 %f316, %f310;
mov.f32 %f317, %f310;
mov.f32 %f318, %f310;
mov.f32 %f319, %f310;
mov.f32 %f320, %f310;
mov.f32 %f321, %f310;
@%p3 bra $L__BB260_3;

mov.u32 %r50, %ctaid.y;
mov.u32 %r51, %ctaid.z;
mul.hi.u32 %r52, %r51, %r37;
add.s32 %r53, %r52, %r51;
shr.u32 %r54, %r53, %r38;
mul.lo.s32 %r55, %r54, %r40;
mul.hi.u32 %r56, %r50, %r31;
add.s32 %r57, %r56, %r50;
shr.u32 %r58, %r57, %r32;
mul.lo.s32 %r59, %r58, %r34;
cvta.to.global.u64 %rd124, %rd15;
and.b32 %r63, %r48, -8;
sub.s32 %r64, %r2, %r63;
shl.b32 %r65, %r64, 1;
cvt.s64.s32 %rd2, %r64;
shl.b32 %r66, %r29, 1;
shl.b32 %r827, %r833, 3;
add.s32 %r832, %r827, %r66;
mad.lo.s32 %r831, %r29, 3, %r827;
shl.b32 %r67, %r29, 2;
add.s32 %r830, %r827, %r67;
mad.lo.s32 %r829, %r29, 5, %r827;
add.s32 %r828, %r29, %r827;
add.s32 %r68, %r833, %r55;
add.s32 %r69, %r68, %r59;
add.s32 %r70, %r3, 1;
mad.lo.s32 %r71, %r28, %r70, %r69;
mul.wide.s32 %rd3, %r71, 50;
mul.wide.s32 %rd18, %r64, 2;
add.s64 %rd19, %rd3, %rd18;
add.s64 %rd4, %rd19, 34;
mad.lo.s32 %r72, %r3, %r28, %r833;
add.s32 %r73, %r72, %r55;
add.s32 %r74, %r73, %r59;
mul.wide.s32 %rd5, %r74, 50;
mul.wide.s32 %rd20, %r65, 2;
add.s64 %rd21, %rd5, %rd20;
add.s64 %rd6, %rd21, 4;
add.s64 %rd22, %rd3, %rd20;
add.s64 %rd7, %rd22, 2;
add.s64 %rd23, %rd5, %rd18;
add.s64 %rd8, %rd23, 34;
cvta.to.global.u64 %rd9, %rd16;
mov.f32 %f310, 0f00000000;

$L__BB260_2:
mul.lo.s32 %r484, %r51, %r41;
cvt.u64.u32 %rd24, %r484;
mul.lo.s32 %r486, %r50, %r35;
cvt.u64.u32 %rd25, %r486;
add.s64 %rd26, %rd24, %rd25;
cvt.u64.u32 %rd27, %r827;
add.s64 %rd28, %rd26, %rd27;
add.s64 %rd29, %rd28, %rd2;
mul.lo.s64 %rd30, %rd29, 36;
add.s64 %rd31, %rd124, %rd6;
ld.global.nc.u16 %rs13, [%rd31+-2];
shr.u16 %rs14, %rs13, 8;
cvt.u32.u16 %r487, %rs14;
ld.global.nc.u16 %rs15, [%rd31];
shr.u16 %rs16, %rs15, 8;
cvt.u32.u16 %r488, %rs16;
add.s64 %rd32, %rd124, %rd8;
ld.global.nc.u16 %rs17, [%rd32];
cvt.u32.u16 %r489, %rs17;
and.b16 %rs18, %rs13, 255;
cvt.u32.u16 %r490, %rs18;
mul.wide.u16 %r491, %rs17, 256;
and.b32 %r492, %r491, 1792;
or.b32 %r493, %r492, %r490;
mul.wide.u32 %rd33, %r493, 4;
mov.u64 %rd34, iq1s_grid_gpu;
add.s64 %rd35, %rd34, %rd33;
ld.global.nc.u32 %r494, [%rd35];
and.b32 %r416, %r494, 252645135;
shr.u32 %r495, %r494, 4;
and.b32 %r420, %r495, 252645135;
add.s64 %rd36, %rd9, %rd30;
ld.global.nc.u32 %r77, [%rd36+4];
ld.global.nc.u32 %r81, [%rd36+8];
mov.u32 %r452, 0;

dp4a.s32.s32 %r75, %r416, %r77, %r452;


dp4a.s32.s32 %r79, %r420, %r81, %r75;

mul.wide.u16 %r496, %rs17, 32;
and.b32 %r497, %r496, 1792;
or.b32 %r498, %r497, %r487;
mul.wide.u32 %rd37, %r498, 4;
add.s64 %rd38, %rd34, %rd37;
ld.global.nc.u32 %r499, [%rd38];
and.b32 %r424, %r499, 252645135;
shr.u32 %r500, %r499, 4;
and.b32 %r428, %r500, 252645135;
ld.global.nc.u32 %r85, [%rd36+12];
ld.global.nc.u32 %r89, [%rd36+16];

dp4a.s32.s32 %r83, %r424, %r85, %r79;


dp4a.s32.s32 %r87, %r428, %r89, %r83;

and.b16 %rs19, %rs15, 255;
cvt.u32.u16 %r501, %rs19;
mul.wide.u16 %r502, %rs17, 4;
and.b32 %r503, %r502, 1792;
or.b32 %r504, %r503, %r501;
mul.wide.u32 %rd39, %r504, 4;
add.s64 %rd40, %rd34, %rd39;
ld.global.nc.u32 %r505, [%rd40];
and.b32 %r432, %r505, 252645135;
shr.u32 %r506, %r505, 4;
and.b32 %r436, %r506, 252645135;
ld.global.nc.u32 %r93, [%rd36+20];
ld.global.nc.u32 %r97, [%rd36+24];

dp4a.s32.s32 %r91, %r432, %r93, %r87;


dp4a.s32.s32 %r95, %r436, %r97, %r91;

shr.u32 %r507, %r489, 1;
and.b32 %r508, %r507, 1792;
or.b32 %r509, %r508, %r488;
mul.wide.u32 %rd41, %r509, 4;
add.s64 %rd42, %rd34, %rd41;
ld.global.nc.u32 %r510, [%rd42];
and.b32 %r440, %r510, 252645135;
shr.u32 %r511, %r510, 4;
and.b32 %r444, %r511, 252645135;
ld.global.nc.u32 %r101, [%rd36+28];
ld.global.nc.u32 %r105, [%rd36+32];

dp4a.s32.s32 %r99, %r440, %r101, %r95;


dp4a.s32.s32 %r103, %r444, %r105, %r99;

add.s64 %rd43, %rd124, %rd5;
ld.global.nc.u16 %rs1, [%rd43];

{ cvt.f32.f16 %f73, %rs1;}


shr.u32 %r512, %r489, 11;
and.b32 %r513, %r512, 14;
or.b32 %r514, %r513, 1;
cvt.rn.f32.s32 %f109, %r514;
mul.ftz.f32 %f110, %f73, %f109;
and.b32 %r515, %r489, 32768;
cvt.rn.f32.s32 %f111, %r515;
mul.ftz.f32 %f112, %f111, 0f37000000;
mov.f32 %f113, 0fBF600000;
sub.ftz.f32 %f114, %f113, %f112;
ld.global.nc.u32 %r107, [%rd36];

{.reg .f16 low,high;
mov.b32 {low,high},%r107;
cvt.f32.f16 %f74, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r107;
cvt.f32.f16 %f75, high;}


cvt.rn.f32.s32 %f115, %r103;
mul.ftz.f32 %f116, %f114, %f75;
fma.rn.ftz.f32 %f117, %f74, %f115, %f116;
fma.rn.ftz.f32 %f321, %f110, %f117, %f321;
add.s64 %rd44, %rd124, %rd7;
ld.global.nc.u16 %rs20, [%rd44];
shr.u16 %rs21, %rs20, 8;
cvt.u32.u16 %r516, %rs21;
ld.global.nc.u16 %rs22, [%rd44+2];
shr.u16 %rs23, %rs22, 8;
cvt.u32.u16 %r517, %rs23;
add.s64 %rd45, %rd124, %rd4;
ld.global.nc.u16 %rs24, [%rd45];
cvt.u32.u16 %r518, %rs24;
and.b16 %rs25, %rs20, 255;
cvt.u32.u16 %r519, %rs25;
mul.wide.u16 %r520, %rs24, 256;
and.b32 %r521, %r520, 1792;
or.b32 %r522, %r521, %r519;
mul.wide.u32 %rd46, %r522, 4;
add.s64 %rd47, %rd34, %rd46;
ld.global.nc.u32 %r523, [%rd47];
and.b32 %r450, %r523, 252645135;
shr.u32 %r524, %r523, 4;
and.b32 %r454, %r524, 252645135;

dp4a.s32.s32 %r109, %r450, %r77, %r452;


dp4a.s32.s32 %r113, %r454, %r81, %r109;

mul.wide.u16 %r525, %rs24, 32;
and.b32 %r526, %r525, 1792;
or.b32 %r527, %r526, %r516;
mul.wide.u32 %rd48, %r527, 4;
add.s64 %rd49, %rd34, %rd48;
ld.global.nc.u32 %r528, [%rd49];
and.b32 %r458, %r528, 252645135;
shr.u32 %r529, %r528, 4;
and.b32 %r462, %r529, 252645135;

dp4a.s32.s32 %r117, %r458, %r85, %r113;


dp4a.s32.s32 %r121, %r462, %r89, %r117;

and.b16 %rs26, %rs22, 255;
cvt.u32.u16 %r530, %rs26;
mul.wide.u16 %r531, %rs24, 4;
and.b32 %r532, %r531, 1792;
or.b32 %r533, %r532, %r530;
mul.wide.u32 %rd50, %r533, 4;
add.s64 %rd51, %rd34, %rd50;
ld.global.nc.u32 %r534, [%rd51];
and.b32 %r466, %r534, 252645135;
shr.u32 %r535, %r534, 4;
and.b32 %r470, %r535, 252645135;

dp4a.s32.s32 %r125, %r466, %r93, %r121;


dp4a.s32.s32 %r129, %r470, %r97, %r125;

shr.u32 %r536, %r518, 1;
and.b32 %r537, %r536, 1792;
or.b32 %r538, %r537, %r517;
mul.wide.u32 %rd52, %r538, 4;
add.s64 %rd53, %rd34, %rd52;
ld.global.nc.u32 %r539, [%rd53];
and.b32 %r474, %r539, 252645135;
shr.u32 %r540, %r539, 4;
and.b32 %r478, %r540, 252645135;

dp4a.s32.s32 %r133, %r474, %r101, %r129;


dp4a.s32.s32 %r137, %r478, %r105, %r133;

add.s64 %rd54, %rd124, %rd3;
ld.global.nc.u16 %rs2, [%rd54];

{ cvt.f32.f16 %f76, %rs2;}


shr.u32 %r541, %r518, 11;
and.b32 %r542, %r541, 14;
or.b32 %r543, %r542, 1;
cvt.rn.f32.s32 %f118, %r543;
mul.ftz.f32 %f119, %f76, %f118;
and.b32 %r544, %r518, 32768;
cvt.rn.f32.s32 %f120, %r544;
mul.ftz.f32 %f121, %f120, 0f37000000;
sub.ftz.f32 %f122, %f113, %f121;

{.reg .f16 low,high;
mov.b32 {low,high},%r107;
cvt.f32.f16 %f77, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r107;
cvt.f32.f16 %f78, high;}


cvt.rn.f32.s32 %f123, %r137;
mul.ftz.f32 %f124, %f122, %f78;
fma.rn.ftz.f32 %f125, %f77, %f123, %f124;
fma.rn.ftz.f32 %f320, %f119, %f125, %f320;
cvt.u64.u32 %rd55, %r828;
add.s64 %rd56, %rd26, %rd55;
add.s64 %rd57, %rd56, %rd2;
mul.lo.s64 %rd58, %rd57, 36;
add.s64 %rd59, %rd9, %rd58;
ld.global.nc.u32 %r145, [%rd59+4];
ld.global.nc.u32 %r149, [%rd59+8];

dp4a.s32.s32 %r143, %r416, %r145, %r452;


dp4a.s32.s32 %r147, %r420, %r149, %r143;

ld.global.nc.u32 %r153, [%rd59+12];
ld.global.nc.u32 %r157, [%rd59+16];

dp4a.s32.s32 %r151, %r424, %r153, %r147;


dp4a.s32.s32 %r155, %r428, %r157, %r151;

ld.global.nc.u32 %r161, [%rd59+20];
ld.global.nc.u32 %r165, [%rd59+24];

dp4a.s32.s32 %r159, %r432, %r161, %r155;


dp4a.s32.s32 %r163, %r436, %r165, %r159;

ld.global.nc.u32 %r169, [%rd59+28];
ld.global.nc.u32 %r173, [%rd59+32];

dp4a.s32.s32 %r167, %r440, %r169, %r163;


dp4a.s32.s32 %r171, %r444, %r173, %r167;


{ cvt.f32.f16 %f79, %rs1;}


mul.ftz.f32 %f126, %f79, %f109;
ld.global.nc.u32 %r175, [%rd59];

{.reg .f16 low,high;
mov.b32 {low,high},%r175;
cvt.f32.f16 %f80, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r175;
cvt.f32.f16 %f81, high;}


cvt.rn.f32.s32 %f127, %r171;
mul.ftz.f32 %f128, %f114, %f81;
fma.rn.ftz.f32 %f129, %f80, %f127, %f128;
fma.rn.ftz.f32 %f319, %f126, %f129, %f319;

dp4a.s32.s32 %r177, %r450, %r145, %r452;


dp4a.s32.s32 %r181, %r454, %r149, %r177;


dp4a.s32.s32 %r185, %r458, %r153, %r181;


dp4a.s32.s32 %r189, %r462, %r157, %r185;


dp4a.s32.s32 %r193, %r466, %r161, %r189;


dp4a.s32.s32 %r197, %r470, %r165, %r193;


dp4a.s32.s32 %r201, %r474, %r169, %r197;


dp4a.s32.s32 %r205, %r478, %r173, %r201;


{ cvt.f32.f16 %f82, %rs2;}


mul.ftz.f32 %f130, %f82, %f118;

{.reg .f16 low,high;
mov.b32 {low,high},%r175;
cvt.f32.f16 %f83, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r175;
cvt.f32.f16 %f84, high;}


cvt.rn.f32.s32 %f131, %r205;
mul.ftz.f32 %f132, %f122, %f84;
fma.rn.ftz.f32 %f133, %f83, %f131, %f132;
fma.rn.ftz.f32 %f318, %f130, %f133, %f318;
cvt.u64.u32 %rd60, %r832;
add.s64 %rd61, %rd26, %rd60;
add.s64 %rd62, %rd61, %rd2;
mul.lo.s64 %rd63, %rd62, 36;
add.s64 %rd64, %rd9, %rd63;
ld.global.nc.u32 %r213, [%rd64+4];
ld.global.nc.u32 %r217, [%rd64+8];

dp4a.s32.s32 %r211, %r416, %r213, %r452;


dp4a.s32.s32 %r215, %r420, %r217, %r211;

ld.global.nc.u32 %r221, [%rd64+12];
ld.global.nc.u32 %r225, [%rd64+16];

dp4a.s32.s32 %r219, %r424, %r221, %r215;


dp4a.s32.s32 %r223, %r428, %r225, %r219;

ld.global.nc.u32 %r229, [%rd64+20];
ld.global.nc.u32 %r233, [%rd64+24];

dp4a.s32.s32 %r227, %r432, %r229, %r223;


dp4a.s32.s32 %r231, %r436, %r233, %r227;

ld.global.nc.u32 %r237, [%rd64+28];
ld.global.nc.u32 %r241, [%rd64+32];

dp4a.s32.s32 %r235, %r440, %r237, %r231;


dp4a.s32.s32 %r239, %r444, %r241, %r235;


{ cvt.f32.f16 %f85, %rs1;}


mul.ftz.f32 %f134, %f85, %f109;
ld.global.nc.u32 %r243, [%rd64];

{.reg .f16 low,high;
mov.b32 {low,high},%r243;
cvt.f32.f16 %f86, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r243;
cvt.f32.f16 %f87, high;}


cvt.rn.f32.s32 %f135, %r239;
mul.ftz.f32 %f136, %f114, %f87;
fma.rn.ftz.f32 %f137, %f86, %f135, %f136;
fma.rn.ftz.f32 %f317, %f134, %f137, %f317;

dp4a.s32.s32 %r245, %r450, %r213, %r452;


dp4a.s32.s32 %r249, %r454, %r217, %r245;


dp4a.s32.s32 %r253, %r458, %r221, %r249;


dp4a.s32.s32 %r257, %r462, %r225, %r253;


dp4a.s32.s32 %r261, %r466, %r229, %r257;


dp4a.s32.s32 %r265, %r470, %r233, %r261;


dp4a.s32.s32 %r269, %r474, %r237, %r265;


dp4a.s32.s32 %r273, %r478, %r241, %r269;


{ cvt.f32.f16 %f88, %rs2;}


mul.ftz.f32 %f138, %f88, %f118;

{.reg .f16 low,high;
mov.b32 {low,high},%r243;
cvt.f32.f16 %f89, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r243;
cvt.f32.f16 %f90, high;}


cvt.rn.f32.s32 %f139, %r273;
mul.ftz.f32 %f140, %f122, %f90;
fma.rn.ftz.f32 %f141, %f89, %f139, %f140;
fma.rn.ftz.f32 %f316, %f138, %f141, %f316;
cvt.u64.u32 %rd65, %r831;
add.s64 %rd66, %rd26, %rd65;
add.s64 %rd67, %rd66, %rd2;
mul.lo.s64 %rd68, %rd67, 36;
add.s64 %rd69, %rd9, %rd68;
ld.global.nc.u32 %r281, [%rd69+4];
ld.global.nc.u32 %r285, [%rd69+8];

dp4a.s32.s32 %r279, %r416, %r281, %r452;


dp4a.s32.s32 %r283, %r420, %r285, %r279;

ld.global.nc.u32 %r289, [%rd69+12];
ld.global.nc.u32 %r293, [%rd69+16];

dp4a.s32.s32 %r287, %r424, %r289, %r283;


dp4a.s32.s32 %r291, %r428, %r293, %r287;

ld.global.nc.u32 %r297, [%rd69+20];
ld.global.nc.u32 %r301, [%rd69+24];

dp4a.s32.s32 %r295, %r432, %r297, %r291;


dp4a.s32.s32 %r299, %r436, %r301, %r295;

ld.global.nc.u32 %r305, [%rd69+28];
ld.global.nc.u32 %r309, [%rd69+32];

dp4a.s32.s32 %r303, %r440, %r305, %r299;


dp4a.s32.s32 %r307, %r444, %r309, %r303;


{ cvt.f32.f16 %f91, %rs1;}


mul.ftz.f32 %f142, %f91, %f109;
ld.global.nc.u32 %r311, [%rd69];

{.reg .f16 low,high;
mov.b32 {low,high},%r311;
cvt.f32.f16 %f92, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r311;
cvt.f32.f16 %f93, high;}


cvt.rn.f32.s32 %f143, %r307;
mul.ftz.f32 %f144, %f114, %f93;
fma.rn.ftz.f32 %f145, %f92, %f143, %f144;
fma.rn.ftz.f32 %f315, %f142, %f145, %f315;

dp4a.s32.s32 %r313, %r450, %r281, %r452;


dp4a.s32.s32 %r317, %r454, %r285, %r313;


dp4a.s32.s32 %r321, %r458, %r289, %r317;


dp4a.s32.s32 %r325, %r462, %r293, %r321;


dp4a.s32.s32 %r329, %r466, %r297, %r325;


dp4a.s32.s32 %r333, %r470, %r301, %r329;


dp4a.s32.s32 %r337, %r474, %r305, %r333;


dp4a.s32.s32 %r341, %r478, %r309, %r337;


{ cvt.f32.f16 %f94, %rs2;}


mul.ftz.f32 %f146, %f94, %f118;

{.reg .f16 low,high;
mov.b32 {low,high},%r311;
cvt.f32.f16 %f95, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r311;
cvt.f32.f16 %f96, high;}


cvt.rn.f32.s32 %f147, %r341;
mul.ftz.f32 %f148, %f122, %f96;
fma.rn.ftz.f32 %f149, %f95, %f147, %f148;
fma.rn.ftz.f32 %f314, %f146, %f149, %f314;
cvt.u64.u32 %rd70, %r830;
add.s64 %rd71, %rd26, %rd70;
add.s64 %rd72, %rd71, %rd2;
mul.lo.s64 %rd73, %rd72, 36;
add.s64 %rd74, %rd9, %rd73;
ld.global.nc.u32 %r349, [%rd74+4];
ld.global.nc.u32 %r353, [%rd74+8];

dp4a.s32.s32 %r347, %r416, %r349, %r452;


dp4a.s32.s32 %r351, %r420, %r353, %r347;

ld.global.nc.u32 %r357, [%rd74+12];
ld.global.nc.u32 %r361, [%rd74+16];

dp4a.s32.s32 %r355, %r424, %r357, %r351;


dp4a.s32.s32 %r359, %r428, %r361, %r355;

ld.global.nc.u32 %r365, [%rd74+20];
ld.global.nc.u32 %r369, [%rd74+24];

dp4a.s32.s32 %r363, %r432, %r365, %r359;


dp4a.s32.s32 %r367, %r436, %r369, %r363;

ld.global.nc.u32 %r373, [%rd74+28];
ld.global.nc.u32 %r377, [%rd74+32];

dp4a.s32.s32 %r371, %r440, %r373, %r367;


dp4a.s32.s32 %r375, %r444, %r377, %r371;


{ cvt.f32.f16 %f97, %rs1;}


mul.ftz.f32 %f150, %f97, %f109;
ld.global.nc.u32 %r379, [%rd74];

{.reg .f16 low,high;
mov.b32 {low,high},%r379;
cvt.f32.f16 %f98, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r379;
cvt.f32.f16 %f99, high;}


cvt.rn.f32.s32 %f151, %r375;
mul.ftz.f32 %f152, %f114, %f99;
fma.rn.ftz.f32 %f153, %f98, %f151, %f152;
fma.rn.ftz.f32 %f313, %f150, %f153, %f313;

dp4a.s32.s32 %r381, %r450, %r349, %r452;


dp4a.s32.s32 %r385, %r454, %r353, %r381;


dp4a.s32.s32 %r389, %r458, %r357, %r385;


dp4a.s32.s32 %r393, %r462, %r361, %r389;


dp4a.s32.s32 %r397, %r466, %r365, %r393;


dp4a.s32.s32 %r401, %r470, %r369, %r397;


dp4a.s32.s32 %r405, %r474, %r373, %r401;


dp4a.s32.s32 %r409, %r478, %r377, %r405;


{ cvt.f32.f16 %f100, %rs2;}


mul.ftz.f32 %f154, %f100, %f118;

{.reg .f16 low,high;
mov.b32 {low,high},%r379;
cvt.f32.f16 %f101, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r379;
cvt.f32.f16 %f102, high;}


cvt.rn.f32.s32 %f155, %r409;
mul.ftz.f32 %f156, %f122, %f102;
fma.rn.ftz.f32 %f157, %f101, %f155, %f156;
fma.rn.ftz.f32 %f312, %f154, %f157, %f312;
cvt.u64.u32 %rd75, %r829;
add.s64 %rd76, %rd26, %rd75;
add.s64 %rd77, %rd76, %rd2;
mul.lo.s64 %rd78, %rd77, 36;
add.s64 %rd79, %rd9, %rd78;
ld.global.nc.u32 %r417, [%rd79+4];
ld.global.nc.u32 %r421, [%rd79+8];

dp4a.s32.s32 %r415, %r416, %r417, %r452;


dp4a.s32.s32 %r419, %r420, %r421, %r415;

ld.global.nc.u32 %r425, [%rd79+12];
ld.global.nc.u32 %r429, [%rd79+16];

dp4a.s32.s32 %r423, %r424, %r425, %r419;


dp4a.s32.s32 %r427, %r428, %r429, %r423;

ld.global.nc.u32 %r433, [%rd79+20];
ld.global.nc.u32 %r437, [%rd79+24];

dp4a.s32.s32 %r431, %r432, %r433, %r427;


dp4a.s32.s32 %r435, %r436, %r437, %r431;

ld.global.nc.u32 %r441, [%rd79+28];
ld.global.nc.u32 %r445, [%rd79+32];

dp4a.s32.s32 %r439, %r440, %r441, %r435;


dp4a.s32.s32 %r443, %r444, %r445, %r439;


{ cvt.f32.f16 %f103, %rs1;}


mul.ftz.f32 %f158, %f103, %f109;
ld.global.nc.u32 %r447, [%rd79];

{.reg .f16 low,high;
mov.b32 {low,high},%r447;
cvt.f32.f16 %f104, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r447;
cvt.f32.f16 %f105, high;}


cvt.rn.f32.s32 %f159, %r443;
mul.ftz.f32 %f160, %f114, %f105;
fma.rn.ftz.f32 %f161, %f104, %f159, %f160;
fma.rn.ftz.f32 %f311, %f158, %f161, %f311;

dp4a.s32.s32 %r449, %r450, %r417, %r452;


dp4a.s32.s32 %r453, %r454, %r421, %r449;


dp4a.s32.s32 %r457, %r458, %r425, %r453;


dp4a.s32.s32 %r461, %r462, %r429, %r457;


dp4a.s32.s32 %r465, %r466, %r433, %r461;


dp4a.s32.s32 %r469, %r470, %r437, %r465;


dp4a.s32.s32 %r473, %r474, %r441, %r469;


dp4a.s32.s32 %r477, %r478, %r445, %r473;


{ cvt.f32.f16 %f106, %rs2;}


mul.ftz.f32 %f162, %f106, %f118;

{.reg .f16 low,high;
mov.b32 {low,high},%r447;
cvt.f32.f16 %f107, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r447;
cvt.f32.f16 %f108, high;}


cvt.rn.f32.s32 %f163, %r477;
mul.ftz.f32 %f164, %f122, %f108;
fma.rn.ftz.f32 %f165, %f107, %f163, %f164;
fma.rn.ftz.f32 %f310, %f162, %f165, %f310;
add.s32 %r832, %r832, 64;
add.s32 %r831, %r831, 64;
add.s32 %r830, %r830, 64;
add.s32 %r829, %r829, 64;
add.s32 %r828, %r828, 64;
add.s32 %r827, %r827, 64;
add.s64 %rd124, %rd124, 400;
add.s32 %r833, %r833, 8;
setp.lt.s32 %p4, %r833, %r49;
@%p4 bra $L__BB260_2;

$L__BB260_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB260_5;

mov.u32 %r547, _ZZ13mul_mat_vec_qIL9ggml_type19ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r548, %r1, 1536, %r547;
shl.b32 %r549, %r44, 2;
add.s32 %r550, %r548, %r549;
st.shared.f32 [%r550+-1536], %f321;
st.shared.f32 [%r550+-1408], %f320;
st.shared.f32 [%r550+-1280], %f319;
st.shared.f32 [%r550+-1152], %f318;
st.shared.f32 [%r550+-1024], %f317;
st.shared.f32 [%r550+-896], %f316;
st.shared.f32 [%r550+-768], %f315;
st.shared.f32 [%r550+-640], %f314;
st.shared.f32 [%r550+-512], %f313;
st.shared.f32 [%r550+-384], %f312;
st.shared.f32 [%r550+-256], %f311;
st.shared.f32 [%r550+-128], %f310;

$L__BB260_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB260_30;

mov.u32 %r551, %ctaid.z;
mov.u32 %r552, %ctaid.y;
mad.lo.s32 %r553, %r552, %r36, %r3;
mad.lo.s32 %r554, %r551, %r42, %r553;
cvt.u64.u32 %rd12, %r554;
shl.b32 %r556, %r44, 2;
mov.u32 %r557, _ZZ13mul_mat_vec_qIL9ggml_type19ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r558, %r557, %r556;
mov.u32 %r559, 2;
ld.shared.f32 %f166, [%r558];
add.ftz.f32 %f167, %f166, %f321;
mov.b32 %r560, %f167;
mov.u32 %r561, 31;
mov.u32 %r562, 16;
mov.u32 %r563, -1;
shfl.sync.bfly.b32 %r564|%p7, %r560, %r562, %r561, %r563;
mov.b32 %f168, %r564;
add.ftz.f32 %f169, %f167, %f168;
mov.b32 %r565, %f169;
mov.u32 %r566, 8;
shfl.sync.bfly.b32 %r567|%p8, %r565, %r566, %r561, %r563;
mov.b32 %f170, %r567;
add.ftz.f32 %f171, %f169, %f170;
mov.b32 %r568, %f171;
mov.u32 %r569, 4;
shfl.sync.bfly.b32 %r570|%p9, %r568, %r569, %r561, %r563;
mov.b32 %f172, %r570;
add.ftz.f32 %f173, %f171, %f172;
mov.b32 %r571, %f173;
shfl.sync.bfly.b32 %r572|%p10, %r571, %r559, %r561, %r563;
mov.b32 %f174, %r572;
add.ftz.f32 %f175, %f173, %f174;
mov.b32 %r573, %f175;
mov.u32 %r574, 1;
shfl.sync.bfly.b32 %r575|%p11, %r573, %r574, %r561, %r563;
mov.b32 %f176, %r575;
add.ftz.f32 %f37, %f175, %f176;
setp.eq.s32 %p12, %r44, 0;
setp.lt.u32 %p13, %r3, %r30;
and.pred %p1, %p12, %p13;
not.pred %p14, %p1;
@%p14 bra $L__BB260_8;

cvta.to.global.u64 %rd80, %rd17;
mul.wide.u32 %rd81, %r554, 4;
add.s64 %rd82, %rd80, %rd81;
st.global.f32 [%rd82], %f37;

$L__BB260_8:
ld.shared.f32 %f177, [%r558+128];
add.ftz.f32 %f178, %f177, %f320;
mov.b32 %r587, %f178;
shfl.sync.bfly.b32 %r591|%p15, %r587, %r562, %r561, %r563;
mov.b32 %f179, %r591;
add.ftz.f32 %f180, %f178, %f179;
mov.b32 %r592, %f180;
shfl.sync.bfly.b32 %r594|%p16, %r592, %r566, %r561, %r563;
mov.b32 %f181, %r594;
add.ftz.f32 %f182, %f180, %f181;
mov.b32 %r595, %f182;
shfl.sync.bfly.b32 %r597|%p17, %r595, %r569, %r561, %r563;
mov.b32 %f183, %r597;
add.ftz.f32 %f184, %f182, %f183;
mov.b32 %r598, %f184;
shfl.sync.bfly.b32 %r599|%p18, %r598, %r559, %r561, %r563;
mov.b32 %f185, %r599;
add.ftz.f32 %f186, %f184, %f185;
mov.b32 %r600, %f186;
shfl.sync.bfly.b32 %r602|%p19, %r600, %r574, %r561, %r563;
mov.b32 %f187, %r602;
add.ftz.f32 %f38, %f186, %f187;
setp.eq.s32 %p20, %r44, 1;
or.b32 %r605, %r3, 1;
setp.lt.u32 %p21, %r605, %r30;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB260_10;

cvta.to.global.u64 %rd83, %rd17;
mul.wide.u32 %rd84, %r554, 4;
add.s64 %rd85, %rd83, %rd84;
st.global.f32 [%rd85+4], %f38;

$L__BB260_10:
mov.u32 %r616, 2;
ld.shared.f32 %f188, [%r558+256];
add.ftz.f32 %f189, %f188, %f319;
mov.b32 %r617, %f189;
mov.u32 %r618, 31;
mov.u32 %r619, 16;
mov.u32 %r620, -1;
shfl.sync.bfly.b32 %r621|%p23, %r617, %r619, %r618, %r620;
mov.b32 %f190, %r621;
add.ftz.f32 %f191, %f189, %f190;
mov.b32 %r622, %f191;
mov.u32 %r623, 8;
shfl.sync.bfly.b32 %r624|%p24, %r622, %r623, %r618, %r620;
mov.b32 %f192, %r624;
add.ftz.f32 %f193, %f191, %f192;
mov.b32 %r625, %f193;
mov.u32 %r626, 4;
shfl.sync.bfly.b32 %r627|%p25, %r625, %r626, %r618, %r620;
mov.b32 %f194, %r627;
add.ftz.f32 %f195, %f193, %f194;
mov.b32 %r628, %f195;
shfl.sync.bfly.b32 %r629|%p26, %r628, %r616, %r618, %r620;
mov.b32 %f196, %r629;
add.ftz.f32 %f197, %f195, %f196;
mov.b32 %r630, %f197;
mov.u32 %r631, 1;
shfl.sync.bfly.b32 %r632|%p27, %r630, %r631, %r618, %r620;
mov.b32 %f198, %r632;
add.ftz.f32 %f39, %f197, %f198;
@%p14 bra $L__BB260_12;

cvt.u64.u32 %rd86, %r30;
add.s64 %rd87, %rd86, %rd12;
cvta.to.global.u64 %rd88, %rd17;
shl.b64 %rd89, %rd87, 2;
add.s64 %rd90, %rd88, %rd89;
st.global.f32 [%rd90], %f39;

$L__BB260_12:
ld.shared.f32 %f199, [%r558+384];
add.ftz.f32 %f200, %f199, %f318;
mov.b32 %r638, %f200;
shfl.sync.bfly.b32 %r642|%p29, %r638, %r619, %r618, %r620;
mov.b32 %f201, %r642;
add.ftz.f32 %f202, %f200, %f201;
mov.b32 %r643, %f202;
shfl.sync.bfly.b32 %r645|%p30, %r643, %r623, %r618, %r620;
mov.b32 %f203, %r645;
add.ftz.f32 %f204, %f202, %f203;
mov.b32 %r646, %f204;
shfl.sync.bfly.b32 %r648|%p31, %r646, %r626, %r618, %r620;
mov.b32 %f205, %r648;
add.ftz.f32 %f206, %f204, %f205;
mov.b32 %r649, %f206;
shfl.sync.bfly.b32 %r650|%p32, %r649, %r616, %r618, %r620;
mov.b32 %f207, %r650;
add.ftz.f32 %f208, %f206, %f207;
mov.b32 %r651, %f208;
shfl.sync.bfly.b32 %r653|%p33, %r651, %r631, %r618, %r620;
mov.b32 %f209, %r653;
add.ftz.f32 %f40, %f208, %f209;
@%p22 bra $L__BB260_14;

add.s32 %r654, %r30, 1;
cvt.u64.u32 %rd91, %r654;
add.s64 %rd92, %rd91, %rd12;
cvta.to.global.u64 %rd93, %rd17;
shl.b64 %rd94, %rd92, 2;
add.s64 %rd95, %rd93, %rd94;
st.global.f32 [%rd95], %f40;

$L__BB260_14:
shl.b32 %r655, %r30, 1;
mov.u32 %r656, 1;
mov.u32 %r661, 2;
ld.shared.f32 %f210, [%r558+512];
add.ftz.f32 %f211, %f210, %f317;
mov.b32 %r662, %f211;
mov.u32 %r663, 31;
mov.u32 %r664, 16;
mov.u32 %r665, -1;
shfl.sync.bfly.b32 %r666|%p35, %r662, %r664, %r663, %r665;
mov.b32 %f212, %r666;
add.ftz.f32 %f213, %f211, %f212;
mov.b32 %r667, %f213;
mov.u32 %r668, 8;
shfl.sync.bfly.b32 %r669|%p36, %r667, %r668, %r663, %r665;
mov.b32 %f214, %r669;
add.ftz.f32 %f215, %f213, %f214;
mov.b32 %r670, %f215;
mov.u32 %r671, 4;
shfl.sync.bfly.b32 %r672|%p37, %r670, %r671, %r663, %r665;
mov.b32 %f216, %r672;
add.ftz.f32 %f217, %f215, %f216;
mov.b32 %r673, %f217;
shfl.sync.bfly.b32 %r674|%p38, %r673, %r661, %r663, %r665;
mov.b32 %f218, %r674;
add.ftz.f32 %f219, %f217, %f218;
mov.b32 %r675, %f219;
shfl.sync.bfly.b32 %r676|%p39, %r675, %r656, %r663, %r665;
mov.b32 %f220, %r676;
add.ftz.f32 %f41, %f219, %f220;
cvt.u64.u32 %rd96, %r655;
add.s64 %rd97, %rd96, %rd12;
cvta.to.global.u64 %rd98, %rd17;
shl.b64 %rd99, %rd97, 2;
add.s64 %rd13, %rd98, %rd99;
@%p14 bra $L__BB260_16;

st.global.f32 [%rd13], %f41;

$L__BB260_16:
ld.shared.f32 %f221, [%r558+640];
add.ftz.f32 %f222, %f221, %f316;
mov.b32 %r682, %f222;
shfl.sync.bfly.b32 %r686|%p41, %r682, %r664, %r663, %r665;
mov.b32 %f223, %r686;
add.ftz.f32 %f224, %f222, %f223;
mov.b32 %r687, %f224;
shfl.sync.bfly.b32 %r689|%p42, %r687, %r668, %r663, %r665;
mov.b32 %f225, %r689;
add.ftz.f32 %f226, %f224, %f225;
mov.b32 %r690, %f226;
shfl.sync.bfly.b32 %r692|%p43, %r690, %r671, %r663, %r665;
mov.b32 %f227, %r692;
add.ftz.f32 %f228, %f226, %f227;
mov.b32 %r693, %f228;
shfl.sync.bfly.b32 %r694|%p44, %r693, %r661, %r663, %r665;
mov.b32 %f229, %r694;
add.ftz.f32 %f230, %f228, %f229;
mov.b32 %r695, %f230;
shfl.sync.bfly.b32 %r697|%p45, %r695, %r656, %r663, %r665;
mov.b32 %f231, %r697;
add.ftz.f32 %f42, %f230, %f231;
@%p22 bra $L__BB260_18;

st.global.f32 [%rd13+4], %f42;

$L__BB260_18:
mul.lo.s32 %r25, %r30, 3;
mov.u32 %r702, 2;
ld.shared.f32 %f232, [%r558+768];
add.ftz.f32 %f233, %f232, %f315;
mov.b32 %r703, %f233;
mov.u32 %r704, 31;
mov.u32 %r705, 16;
mov.u32 %r706, -1;
shfl.sync.bfly.b32 %r707|%p47, %r703, %r705, %r704, %r706;
mov.b32 %f234, %r707;
add.ftz.f32 %f235, %f233, %f234;
mov.b32 %r708, %f235;
mov.u32 %r709, 8;
shfl.sync.bfly.b32 %r710|%p48, %r708, %r709, %r704, %r706;
mov.b32 %f236, %r710;
add.ftz.f32 %f237, %f235, %f236;
mov.b32 %r711, %f237;
mov.u32 %r712, 4;
shfl.sync.bfly.b32 %r713|%p49, %r711, %r712, %r704, %r706;
mov.b32 %f238, %r713;
add.ftz.f32 %f239, %f237, %f238;
mov.b32 %r714, %f239;
shfl.sync.bfly.b32 %r715|%p50, %r714, %r702, %r704, %r706;
mov.b32 %f240, %r715;
add.ftz.f32 %f241, %f239, %f240;
mov.b32 %r716, %f241;
mov.u32 %r717, 1;
shfl.sync.bfly.b32 %r718|%p51, %r716, %r717, %r704, %r706;
mov.b32 %f242, %r718;
add.ftz.f32 %f43, %f241, %f242;
@%p14 bra $L__BB260_20;

cvt.u64.u32 %rd100, %r25;
add.s64 %rd101, %rd100, %rd12;
shl.b64 %rd103, %rd101, 2;
add.s64 %rd104, %rd98, %rd103;
st.global.f32 [%rd104], %f43;

$L__BB260_20:
ld.shared.f32 %f243, [%r558+896];
add.ftz.f32 %f244, %f243, %f314;
mov.b32 %r724, %f244;
shfl.sync.bfly.b32 %r728|%p53, %r724, %r705, %r704, %r706;
mov.b32 %f245, %r728;
add.ftz.f32 %f246, %f244, %f245;
mov.b32 %r729, %f246;
shfl.sync.bfly.b32 %r731|%p54, %r729, %r709, %r704, %r706;
mov.b32 %f247, %r731;
add.ftz.f32 %f248, %f246, %f247;
mov.b32 %r732, %f248;
shfl.sync.bfly.b32 %r734|%p55, %r732, %r712, %r704, %r706;
mov.b32 %f249, %r734;
add.ftz.f32 %f250, %f248, %f249;
mov.b32 %r735, %f250;
shfl.sync.bfly.b32 %r736|%p56, %r735, %r702, %r704, %r706;
mov.b32 %f251, %r736;
add.ftz.f32 %f252, %f250, %f251;
mov.b32 %r737, %f252;
shfl.sync.bfly.b32 %r739|%p57, %r737, %r717, %r704, %r706;
mov.b32 %f253, %r739;
add.ftz.f32 %f44, %f252, %f253;
@%p22 bra $L__BB260_22;

add.s32 %r740, %r25, 1;
cvt.u64.u32 %rd105, %r740;
add.s64 %rd106, %rd105, %rd12;
shl.b64 %rd108, %rd106, 2;
add.s64 %rd109, %rd98, %rd108;
st.global.f32 [%rd109], %f44;

$L__BB260_22:
shl.b32 %r741, %r30, 2;
mov.u32 %r742, 2;
ld.shared.f32 %f254, [%r558+1024];
add.ftz.f32 %f255, %f254, %f313;
mov.b32 %r747, %f255;
mov.u32 %r748, 31;
mov.u32 %r749, 16;
mov.u32 %r750, -1;
shfl.sync.bfly.b32 %r751|%p59, %r747, %r749, %r748, %r750;
mov.b32 %f256, %r751;
add.ftz.f32 %f257, %f255, %f256;
mov.b32 %r752, %f257;
mov.u32 %r753, 8;
shfl.sync.bfly.b32 %r754|%p60, %r752, %r753, %r748, %r750;
mov.b32 %f258, %r754;
add.ftz.f32 %f259, %f257, %f258;
mov.b32 %r755, %f259;
mov.u32 %r756, 4;
shfl.sync.bfly.b32 %r757|%p61, %r755, %r756, %r748, %r750;
mov.b32 %f260, %r757;
add.ftz.f32 %f261, %f259, %f260;
mov.b32 %r758, %f261;
shfl.sync.bfly.b32 %r759|%p62, %r758, %r742, %r748, %r750;
mov.b32 %f262, %r759;
add.ftz.f32 %f263, %f261, %f262;
mov.b32 %r760, %f263;
mov.u32 %r761, 1;
shfl.sync.bfly.b32 %r762|%p63, %r760, %r761, %r748, %r750;
mov.b32 %f264, %r762;
add.ftz.f32 %f45, %f263, %f264;
cvt.u64.u32 %rd110, %r741;
add.s64 %rd111, %rd110, %rd12;
shl.b64 %rd113, %rd111, 2;
add.s64 %rd14, %rd98, %rd113;
@%p14 bra $L__BB260_24;

st.global.f32 [%rd14], %f45;

$L__BB260_24:
ld.shared.f32 %f265, [%r558+1152];
add.ftz.f32 %f266, %f265, %f312;
mov.b32 %r768, %f266;
shfl.sync.bfly.b32 %r772|%p65, %r768, %r749, %r748, %r750;
mov.b32 %f267, %r772;
add.ftz.f32 %f268, %f266, %f267;
mov.b32 %r773, %f268;
shfl.sync.bfly.b32 %r775|%p66, %r773, %r753, %r748, %r750;
mov.b32 %f269, %r775;
add.ftz.f32 %f270, %f268, %f269;
mov.b32 %r776, %f270;
shfl.sync.bfly.b32 %r778|%p67, %r776, %r756, %r748, %r750;
mov.b32 %f271, %r778;
add.ftz.f32 %f272, %f270, %f271;
mov.b32 %r779, %f272;
shfl.sync.bfly.b32 %r780|%p68, %r779, %r742, %r748, %r750;
mov.b32 %f273, %r780;
add.ftz.f32 %f274, %f272, %f273;
mov.b32 %r781, %f274;
shfl.sync.bfly.b32 %r783|%p69, %r781, %r761, %r748, %r750;
mov.b32 %f275, %r783;
add.ftz.f32 %f46, %f274, %f275;
@%p22 bra $L__BB260_26;

st.global.f32 [%rd14+4], %f46;

$L__BB260_26:
mul.lo.s32 %r26, %r30, 5;
mov.u32 %r788, 2;
ld.shared.f32 %f276, [%r558+1280];
add.ftz.f32 %f277, %f276, %f311;
mov.b32 %r789, %f277;
mov.u32 %r790, 31;
mov.u32 %r791, 16;
mov.u32 %r792, -1;
shfl.sync.bfly.b32 %r793|%p71, %r789, %r791, %r790, %r792;
mov.b32 %f278, %r793;
add.ftz.f32 %f279, %f277, %f278;
mov.b32 %r794, %f279;
mov.u32 %r795, 8;
shfl.sync.bfly.b32 %r796|%p72, %r794, %r795, %r790, %r792;
mov.b32 %f280, %r796;
add.ftz.f32 %f281, %f279, %f280;
mov.b32 %r797, %f281;
mov.u32 %r798, 4;
shfl.sync.bfly.b32 %r799|%p73, %r797, %r798, %r790, %r792;
mov.b32 %f282, %r799;
add.ftz.f32 %f283, %f281, %f282;
mov.b32 %r800, %f283;
shfl.sync.bfly.b32 %r801|%p74, %r800, %r788, %r790, %r792;
mov.b32 %f284, %r801;
add.ftz.f32 %f285, %f283, %f284;
mov.b32 %r802, %f285;
mov.u32 %r803, 1;
shfl.sync.bfly.b32 %r804|%p75, %r802, %r803, %r790, %r792;
mov.b32 %f286, %r804;
add.ftz.f32 %f47, %f285, %f286;
@%p14 bra $L__BB260_28;

cvt.u64.u32 %rd114, %r26;
add.s64 %rd115, %rd114, %rd12;
shl.b64 %rd117, %rd115, 2;
add.s64 %rd118, %rd98, %rd117;
st.global.f32 [%rd118], %f47;

$L__BB260_28:
ld.shared.f32 %f287, [%r558+1408];
add.ftz.f32 %f288, %f287, %f310;
mov.b32 %r810, %f288;
shfl.sync.bfly.b32 %r814|%p77, %r810, %r791, %r790, %r792;
mov.b32 %f289, %r814;
add.ftz.f32 %f290, %f288, %f289;
mov.b32 %r815, %f290;
shfl.sync.bfly.b32 %r817|%p78, %r815, %r795, %r790, %r792;
mov.b32 %f291, %r817;
add.ftz.f32 %f292, %f290, %f291;
mov.b32 %r818, %f292;
shfl.sync.bfly.b32 %r820|%p79, %r818, %r798, %r790, %r792;
mov.b32 %f293, %r820;
add.ftz.f32 %f294, %f292, %f293;
mov.b32 %r821, %f294;
shfl.sync.bfly.b32 %r822|%p80, %r821, %r788, %r790, %r792;
mov.b32 %f295, %r822;
add.ftz.f32 %f296, %f294, %f295;
mov.b32 %r823, %f296;
shfl.sync.bfly.b32 %r825|%p81, %r823, %r803, %r790, %r792;
mov.b32 %f297, %r825;
add.ftz.f32 %f48, %f296, %f297;
@%p22 bra $L__BB260_30;

add.s32 %r826, %r26, 1;
cvt.u64.u32 %rd119, %r826;
add.s64 %rd120, %rd119, %rd12;
shl.b64 %rd122, %rd120, 2;
add.s64 %rd123, %rd98, %rd122;
st.global.f32 [%rd123], %f48;

$L__BB260_30:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type19ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type19ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type19ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type19ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type19ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type19ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type19ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type19ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type19ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<106>;
.reg .b16 %rs<29>;
.reg .f32 %f<374>;
.reg .b32 %r<1009>;
.reg .b64 %rd<143>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type19ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1792];

ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type19ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type19ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type19ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type19ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type19ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type19ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type19ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type19ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type19ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type19ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type19ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type19ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type19ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type19ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type19ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type19ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type19ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r45, %r1, 5;
mov.u32 %r46, %tid.x;
add.s32 %r2, %r45, %r46;
shr.s32 %r47, %r2, 31;
shr.u32 %r48, %r47, 29;
add.s32 %r49, %r2, %r48;
shr.s32 %r1008, %r49, 3;
shr.u32 %r50, %r29, 8;
setp.ge.s32 %p2, %r1008, %r50;
mov.f32 %f360, 0f00000000;
mov.f32 %f361, %f360;
mov.f32 %f362, %f360;
mov.f32 %f363, %f360;
mov.f32 %f364, %f360;
mov.f32 %f365, %f360;
mov.f32 %f366, %f360;
mov.f32 %f367, %f360;
mov.f32 %f368, %f360;
mov.f32 %f369, %f360;
mov.f32 %f370, %f360;
mov.f32 %f371, %f360;
mov.f32 %f372, %f360;
mov.f32 %f373, %f360;
@%p2 bra $L__BB261_3;

mov.u32 %r51, %ctaid.x;
shl.b32 %r52, %r51, 1;
mov.u32 %r53, %ctaid.y;
mov.u32 %r54, %ctaid.z;
mul.hi.u32 %r55, %r54, %r39;
add.s32 %r56, %r55, %r54;
shr.u32 %r57, %r56, %r40;
mul.lo.s32 %r58, %r57, %r42;
mul.hi.u32 %r59, %r53, %r33;
add.s32 %r60, %r59, %r53;
shr.u32 %r61, %r60, %r34;
mul.lo.s32 %r62, %r61, %r36;
cvta.to.global.u64 %rd142, %rd15;
and.b32 %r66, %r49, -8;
sub.s32 %r67, %r2, %r66;
shl.b32 %r68, %r67, 1;
cvt.s64.s32 %rd2, %r67;
shl.b32 %r69, %r31, 1;
shl.b32 %r1001, %r1008, 3;
add.s32 %r1007, %r1001, %r69;
mad.lo.s32 %r1006, %r31, 3, %r1001;
shl.b32 %r70, %r31, 2;
add.s32 %r1005, %r1001, %r70;
mad.lo.s32 %r1004, %r31, 5, %r1001;
mad.lo.s32 %r1003, %r31, 6, %r1001;
add.s32 %r1002, %r31, %r1001;
add.s32 %r71, %r1008, %r58;
add.s32 %r72, %r71, %r62;
or.b32 %r73, %r52, 1;
mad.lo.s32 %r74, %r30, %r73, %r72;
mul.wide.s32 %rd3, %r74, 50;
mul.wide.s32 %rd18, %r67, 2;
add.s64 %rd19, %rd3, %rd18;
add.s64 %rd4, %rd19, 34;
mad.lo.s32 %r75, %r52, %r30, %r1008;
add.s32 %r76, %r75, %r58;
add.s32 %r77, %r76, %r62;
mul.wide.s32 %rd5, %r77, 50;
mul.wide.s32 %rd20, %r68, 2;
add.s64 %rd21, %rd5, %rd20;
add.s64 %rd6, %rd21, 4;
add.s64 %rd22, %rd3, %rd20;
add.s64 %rd7, %rd22, 2;
add.s64 %rd23, %rd5, %rd18;
add.s64 %rd8, %rd23, 34;
cvta.to.global.u64 %rd9, %rd16;
mov.f32 %f360, 0f00000000;

$L__BB261_2:
mul.lo.s32 %r555, %r54, %r43;
cvt.u64.u32 %rd24, %r555;
mul.lo.s32 %r557, %r53, %r37;
cvt.u64.u32 %rd25, %r557;
add.s64 %rd26, %rd24, %rd25;
cvt.u64.u32 %rd27, %r1001;
add.s64 %rd28, %rd26, %rd27;
add.s64 %rd29, %rd28, %rd2;
mul.lo.s64 %rd30, %rd29, 36;
add.s64 %rd31, %rd142, %rd6;
ld.global.nc.u16 %rs15, [%rd31+-2];
shr.u16 %rs16, %rs15, 8;
cvt.u32.u16 %r558, %rs16;
ld.global.nc.u16 %rs17, [%rd31];
shr.u16 %rs18, %rs17, 8;
cvt.u32.u16 %r559, %rs18;
add.s64 %rd32, %rd142, %rd8;
ld.global.nc.u16 %rs19, [%rd32];
cvt.u32.u16 %r560, %rs19;
and.b16 %rs20, %rs15, 255;
cvt.u32.u16 %r561, %rs20;
mul.wide.u16 %r562, %rs19, 256;
and.b32 %r563, %r562, 1792;
or.b32 %r564, %r563, %r561;
mul.wide.u32 %rd33, %r564, 4;
mov.u64 %rd34, iq1s_grid_gpu;
add.s64 %rd35, %rd34, %rd33;
ld.global.nc.u32 %r565, [%rd35];
and.b32 %r487, %r565, 252645135;
shr.u32 %r566, %r565, 4;
and.b32 %r491, %r566, 252645135;
add.s64 %rd36, %rd9, %rd30;
ld.global.nc.u32 %r80, [%rd36+4];
ld.global.nc.u32 %r84, [%rd36+8];
mov.u32 %r523, 0;

dp4a.s32.s32 %r78, %r487, %r80, %r523;


dp4a.s32.s32 %r82, %r491, %r84, %r78;

mul.wide.u16 %r567, %rs19, 32;
and.b32 %r568, %r567, 1792;
or.b32 %r569, %r568, %r558;
mul.wide.u32 %rd37, %r569, 4;
add.s64 %rd38, %rd34, %rd37;
ld.global.nc.u32 %r570, [%rd38];
and.b32 %r495, %r570, 252645135;
shr.u32 %r571, %r570, 4;
and.b32 %r499, %r571, 252645135;
ld.global.nc.u32 %r88, [%rd36+12];
ld.global.nc.u32 %r92, [%rd36+16];

dp4a.s32.s32 %r86, %r495, %r88, %r82;


dp4a.s32.s32 %r90, %r499, %r92, %r86;

and.b16 %rs21, %rs17, 255;
cvt.u32.u16 %r572, %rs21;
mul.wide.u16 %r573, %rs19, 4;
and.b32 %r574, %r573, 1792;
or.b32 %r575, %r574, %r572;
mul.wide.u32 %rd39, %r575, 4;
add.s64 %rd40, %rd34, %rd39;
ld.global.nc.u32 %r576, [%rd40];
and.b32 %r503, %r576, 252645135;
shr.u32 %r577, %r576, 4;
and.b32 %r507, %r577, 252645135;
ld.global.nc.u32 %r96, [%rd36+20];
ld.global.nc.u32 %r100, [%rd36+24];

dp4a.s32.s32 %r94, %r503, %r96, %r90;


dp4a.s32.s32 %r98, %r507, %r100, %r94;

shr.u32 %r578, %r560, 1;
and.b32 %r579, %r578, 1792;
or.b32 %r580, %r579, %r559;
mul.wide.u32 %rd41, %r580, 4;
add.s64 %rd42, %rd34, %rd41;
ld.global.nc.u32 %r581, [%rd42];
and.b32 %r511, %r581, 252645135;
shr.u32 %r582, %r581, 4;
and.b32 %r515, %r582, 252645135;
ld.global.nc.u32 %r104, [%rd36+28];
ld.global.nc.u32 %r108, [%rd36+32];

dp4a.s32.s32 %r102, %r511, %r104, %r98;


dp4a.s32.s32 %r106, %r515, %r108, %r102;

add.s64 %rd43, %rd142, %rd5;
ld.global.nc.u16 %rs1, [%rd43];

{ cvt.f32.f16 %f85, %rs1;}


shr.u32 %r583, %r560, 11;
and.b32 %r584, %r583, 14;
or.b32 %r585, %r584, 1;
cvt.rn.f32.s32 %f127, %r585;
mul.ftz.f32 %f128, %f85, %f127;
and.b32 %r586, %r560, 32768;
cvt.rn.f32.s32 %f129, %r586;
mul.ftz.f32 %f130, %f129, 0f37000000;
mov.f32 %f131, 0fBF600000;
sub.ftz.f32 %f132, %f131, %f130;
ld.global.nc.u32 %r110, [%rd36];

{.reg .f16 low,high;
mov.b32 {low,high},%r110;
cvt.f32.f16 %f86, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r110;
cvt.f32.f16 %f87, high;}


cvt.rn.f32.s32 %f133, %r106;
mul.ftz.f32 %f134, %f132, %f87;
fma.rn.ftz.f32 %f135, %f86, %f133, %f134;
fma.rn.ftz.f32 %f373, %f128, %f135, %f373;
add.s64 %rd44, %rd142, %rd7;
ld.global.nc.u16 %rs22, [%rd44];
shr.u16 %rs23, %rs22, 8;
cvt.u32.u16 %r587, %rs23;
ld.global.nc.u16 %rs24, [%rd44+2];
shr.u16 %rs25, %rs24, 8;
cvt.u32.u16 %r588, %rs25;
add.s64 %rd45, %rd142, %rd4;
ld.global.nc.u16 %rs26, [%rd45];
cvt.u32.u16 %r589, %rs26;
and.b16 %rs27, %rs22, 255;
cvt.u32.u16 %r590, %rs27;
mul.wide.u16 %r591, %rs26, 256;
and.b32 %r592, %r591, 1792;
or.b32 %r593, %r592, %r590;
mul.wide.u32 %rd46, %r593, 4;
add.s64 %rd47, %rd34, %rd46;
ld.global.nc.u32 %r594, [%rd47];
and.b32 %r521, %r594, 252645135;
shr.u32 %r595, %r594, 4;
and.b32 %r525, %r595, 252645135;

dp4a.s32.s32 %r112, %r521, %r80, %r523;


dp4a.s32.s32 %r116, %r525, %r84, %r112;

mul.wide.u16 %r596, %rs26, 32;
and.b32 %r597, %r596, 1792;
or.b32 %r598, %r597, %r587;
mul.wide.u32 %rd48, %r598, 4;
add.s64 %rd49, %rd34, %rd48;
ld.global.nc.u32 %r599, [%rd49];
and.b32 %r529, %r599, 252645135;
shr.u32 %r600, %r599, 4;
and.b32 %r533, %r600, 252645135;

dp4a.s32.s32 %r120, %r529, %r88, %r116;


dp4a.s32.s32 %r124, %r533, %r92, %r120;

and.b16 %rs28, %rs24, 255;
cvt.u32.u16 %r601, %rs28;
mul.wide.u16 %r602, %rs26, 4;
and.b32 %r603, %r602, 1792;
or.b32 %r604, %r603, %r601;
mul.wide.u32 %rd50, %r604, 4;
add.s64 %rd51, %rd34, %rd50;
ld.global.nc.u32 %r605, [%rd51];
and.b32 %r537, %r605, 252645135;
shr.u32 %r606, %r605, 4;
and.b32 %r541, %r606, 252645135;

dp4a.s32.s32 %r128, %r537, %r96, %r124;


dp4a.s32.s32 %r132, %r541, %r100, %r128;

shr.u32 %r607, %r589, 1;
and.b32 %r608, %r607, 1792;
or.b32 %r609, %r608, %r588;
mul.wide.u32 %rd52, %r609, 4;
add.s64 %rd53, %rd34, %rd52;
ld.global.nc.u32 %r610, [%rd53];
and.b32 %r545, %r610, 252645135;
shr.u32 %r611, %r610, 4;
and.b32 %r549, %r611, 252645135;

dp4a.s32.s32 %r136, %r545, %r104, %r132;


dp4a.s32.s32 %r140, %r549, %r108, %r136;

add.s64 %rd54, %rd142, %rd3;
ld.global.nc.u16 %rs2, [%rd54];

{ cvt.f32.f16 %f88, %rs2;}


shr.u32 %r612, %r589, 11;
and.b32 %r613, %r612, 14;
or.b32 %r614, %r613, 1;
cvt.rn.f32.s32 %f136, %r614;
mul.ftz.f32 %f137, %f88, %f136;
and.b32 %r615, %r589, 32768;
cvt.rn.f32.s32 %f138, %r615;
mul.ftz.f32 %f139, %f138, 0f37000000;
sub.ftz.f32 %f140, %f131, %f139;

{.reg .f16 low,high;
mov.b32 {low,high},%r110;
cvt.f32.f16 %f89, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r110;
cvt.f32.f16 %f90, high;}


cvt.rn.f32.s32 %f141, %r140;
mul.ftz.f32 %f142, %f140, %f90;
fma.rn.ftz.f32 %f143, %f89, %f141, %f142;
fma.rn.ftz.f32 %f372, %f137, %f143, %f372;
cvt.u64.u32 %rd55, %r1002;
add.s64 %rd56, %rd26, %rd55;
add.s64 %rd57, %rd56, %rd2;
mul.lo.s64 %rd58, %rd57, 36;
add.s64 %rd59, %rd9, %rd58;
ld.global.nc.u32 %r148, [%rd59+4];
ld.global.nc.u32 %r152, [%rd59+8];

dp4a.s32.s32 %r146, %r487, %r148, %r523;


dp4a.s32.s32 %r150, %r491, %r152, %r146;

ld.global.nc.u32 %r156, [%rd59+12];
ld.global.nc.u32 %r160, [%rd59+16];

dp4a.s32.s32 %r154, %r495, %r156, %r150;


dp4a.s32.s32 %r158, %r499, %r160, %r154;

ld.global.nc.u32 %r164, [%rd59+20];
ld.global.nc.u32 %r168, [%rd59+24];

dp4a.s32.s32 %r162, %r503, %r164, %r158;


dp4a.s32.s32 %r166, %r507, %r168, %r162;

ld.global.nc.u32 %r172, [%rd59+28];
ld.global.nc.u32 %r176, [%rd59+32];

dp4a.s32.s32 %r170, %r511, %r172, %r166;


dp4a.s32.s32 %r174, %r515, %r176, %r170;


{ cvt.f32.f16 %f91, %rs1;}


mul.ftz.f32 %f144, %f91, %f127;
ld.global.nc.u32 %r178, [%rd59];

{.reg .f16 low,high;
mov.b32 {low,high},%r178;
cvt.f32.f16 %f92, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r178;
cvt.f32.f16 %f93, high;}


cvt.rn.f32.s32 %f145, %r174;
mul.ftz.f32 %f146, %f132, %f93;
fma.rn.ftz.f32 %f147, %f92, %f145, %f146;
fma.rn.ftz.f32 %f371, %f144, %f147, %f371;

dp4a.s32.s32 %r180, %r521, %r148, %r523;


dp4a.s32.s32 %r184, %r525, %r152, %r180;


dp4a.s32.s32 %r188, %r529, %r156, %r184;


dp4a.s32.s32 %r192, %r533, %r160, %r188;


dp4a.s32.s32 %r196, %r537, %r164, %r192;


dp4a.s32.s32 %r200, %r541, %r168, %r196;


dp4a.s32.s32 %r204, %r545, %r172, %r200;


dp4a.s32.s32 %r208, %r549, %r176, %r204;


{ cvt.f32.f16 %f94, %rs2;}


mul.ftz.f32 %f148, %f94, %f136;

{.reg .f16 low,high;
mov.b32 {low,high},%r178;
cvt.f32.f16 %f95, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r178;
cvt.f32.f16 %f96, high;}


cvt.rn.f32.s32 %f149, %r208;
mul.ftz.f32 %f150, %f140, %f96;
fma.rn.ftz.f32 %f151, %f95, %f149, %f150;
fma.rn.ftz.f32 %f370, %f148, %f151, %f370;
cvt.u64.u32 %rd60, %r1007;
add.s64 %rd61, %rd26, %rd60;
add.s64 %rd62, %rd61, %rd2;
mul.lo.s64 %rd63, %rd62, 36;
add.s64 %rd64, %rd9, %rd63;
ld.global.nc.u32 %r216, [%rd64+4];
ld.global.nc.u32 %r220, [%rd64+8];

dp4a.s32.s32 %r214, %r487, %r216, %r523;


dp4a.s32.s32 %r218, %r491, %r220, %r214;

ld.global.nc.u32 %r224, [%rd64+12];
ld.global.nc.u32 %r228, [%rd64+16];

dp4a.s32.s32 %r222, %r495, %r224, %r218;


dp4a.s32.s32 %r226, %r499, %r228, %r222;

ld.global.nc.u32 %r232, [%rd64+20];
ld.global.nc.u32 %r236, [%rd64+24];

dp4a.s32.s32 %r230, %r503, %r232, %r226;


dp4a.s32.s32 %r234, %r507, %r236, %r230;

ld.global.nc.u32 %r240, [%rd64+28];
ld.global.nc.u32 %r244, [%rd64+32];

dp4a.s32.s32 %r238, %r511, %r240, %r234;


dp4a.s32.s32 %r242, %r515, %r244, %r238;


{ cvt.f32.f16 %f97, %rs1;}


mul.ftz.f32 %f152, %f97, %f127;
ld.global.nc.u32 %r246, [%rd64];

{.reg .f16 low,high;
mov.b32 {low,high},%r246;
cvt.f32.f16 %f98, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r246;
cvt.f32.f16 %f99, high;}


cvt.rn.f32.s32 %f153, %r242;
mul.ftz.f32 %f154, %f132, %f99;
fma.rn.ftz.f32 %f155, %f98, %f153, %f154;
fma.rn.ftz.f32 %f369, %f152, %f155, %f369;

dp4a.s32.s32 %r248, %r521, %r216, %r523;


dp4a.s32.s32 %r252, %r525, %r220, %r248;


dp4a.s32.s32 %r256, %r529, %r224, %r252;


dp4a.s32.s32 %r260, %r533, %r228, %r256;


dp4a.s32.s32 %r264, %r537, %r232, %r260;


dp4a.s32.s32 %r268, %r541, %r236, %r264;


dp4a.s32.s32 %r272, %r545, %r240, %r268;


dp4a.s32.s32 %r276, %r549, %r244, %r272;


{ cvt.f32.f16 %f100, %rs2;}


mul.ftz.f32 %f156, %f100, %f136;

{.reg .f16 low,high;
mov.b32 {low,high},%r246;
cvt.f32.f16 %f101, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r246;
cvt.f32.f16 %f102, high;}


cvt.rn.f32.s32 %f157, %r276;
mul.ftz.f32 %f158, %f140, %f102;
fma.rn.ftz.f32 %f159, %f101, %f157, %f158;
fma.rn.ftz.f32 %f368, %f156, %f159, %f368;
cvt.u64.u32 %rd65, %r1006;
add.s64 %rd66, %rd26, %rd65;
add.s64 %rd67, %rd66, %rd2;
mul.lo.s64 %rd68, %rd67, 36;
add.s64 %rd69, %rd9, %rd68;
ld.global.nc.u32 %r284, [%rd69+4];
ld.global.nc.u32 %r288, [%rd69+8];

dp4a.s32.s32 %r282, %r487, %r284, %r523;


dp4a.s32.s32 %r286, %r491, %r288, %r282;

ld.global.nc.u32 %r292, [%rd69+12];
ld.global.nc.u32 %r296, [%rd69+16];

dp4a.s32.s32 %r290, %r495, %r292, %r286;


dp4a.s32.s32 %r294, %r499, %r296, %r290;

ld.global.nc.u32 %r300, [%rd69+20];
ld.global.nc.u32 %r304, [%rd69+24];

dp4a.s32.s32 %r298, %r503, %r300, %r294;


dp4a.s32.s32 %r302, %r507, %r304, %r298;

ld.global.nc.u32 %r308, [%rd69+28];
ld.global.nc.u32 %r312, [%rd69+32];

dp4a.s32.s32 %r306, %r511, %r308, %r302;


dp4a.s32.s32 %r310, %r515, %r312, %r306;


{ cvt.f32.f16 %f103, %rs1;}


mul.ftz.f32 %f160, %f103, %f127;
ld.global.nc.u32 %r314, [%rd69];

{.reg .f16 low,high;
mov.b32 {low,high},%r314;
cvt.f32.f16 %f104, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r314;
cvt.f32.f16 %f105, high;}


cvt.rn.f32.s32 %f161, %r310;
mul.ftz.f32 %f162, %f132, %f105;
fma.rn.ftz.f32 %f163, %f104, %f161, %f162;
fma.rn.ftz.f32 %f367, %f160, %f163, %f367;

dp4a.s32.s32 %r316, %r521, %r284, %r523;


dp4a.s32.s32 %r320, %r525, %r288, %r316;


dp4a.s32.s32 %r324, %r529, %r292, %r320;


dp4a.s32.s32 %r328, %r533, %r296, %r324;


dp4a.s32.s32 %r332, %r537, %r300, %r328;


dp4a.s32.s32 %r336, %r541, %r304, %r332;


dp4a.s32.s32 %r340, %r545, %r308, %r336;


dp4a.s32.s32 %r344, %r549, %r312, %r340;


{ cvt.f32.f16 %f106, %rs2;}


mul.ftz.f32 %f164, %f106, %f136;

{.reg .f16 low,high;
mov.b32 {low,high},%r314;
cvt.f32.f16 %f107, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r314;
cvt.f32.f16 %f108, high;}


cvt.rn.f32.s32 %f165, %r344;
mul.ftz.f32 %f166, %f140, %f108;
fma.rn.ftz.f32 %f167, %f107, %f165, %f166;
fma.rn.ftz.f32 %f366, %f164, %f167, %f366;
cvt.u64.u32 %rd70, %r1005;
add.s64 %rd71, %rd26, %rd70;
add.s64 %rd72, %rd71, %rd2;
mul.lo.s64 %rd73, %rd72, 36;
add.s64 %rd74, %rd9, %rd73;
ld.global.nc.u32 %r352, [%rd74+4];
ld.global.nc.u32 %r356, [%rd74+8];

dp4a.s32.s32 %r350, %r487, %r352, %r523;


dp4a.s32.s32 %r354, %r491, %r356, %r350;

ld.global.nc.u32 %r360, [%rd74+12];
ld.global.nc.u32 %r364, [%rd74+16];

dp4a.s32.s32 %r358, %r495, %r360, %r354;


dp4a.s32.s32 %r362, %r499, %r364, %r358;

ld.global.nc.u32 %r368, [%rd74+20];
ld.global.nc.u32 %r372, [%rd74+24];

dp4a.s32.s32 %r366, %r503, %r368, %r362;


dp4a.s32.s32 %r370, %r507, %r372, %r366;

ld.global.nc.u32 %r376, [%rd74+28];
ld.global.nc.u32 %r380, [%rd74+32];

dp4a.s32.s32 %r374, %r511, %r376, %r370;


dp4a.s32.s32 %r378, %r515, %r380, %r374;


{ cvt.f32.f16 %f109, %rs1;}


mul.ftz.f32 %f168, %f109, %f127;
ld.global.nc.u32 %r382, [%rd74];

{.reg .f16 low,high;
mov.b32 {low,high},%r382;
cvt.f32.f16 %f110, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r382;
cvt.f32.f16 %f111, high;}


cvt.rn.f32.s32 %f169, %r378;
mul.ftz.f32 %f170, %f132, %f111;
fma.rn.ftz.f32 %f171, %f110, %f169, %f170;
fma.rn.ftz.f32 %f365, %f168, %f171, %f365;

dp4a.s32.s32 %r384, %r521, %r352, %r523;


dp4a.s32.s32 %r388, %r525, %r356, %r384;


dp4a.s32.s32 %r392, %r529, %r360, %r388;


dp4a.s32.s32 %r396, %r533, %r364, %r392;


dp4a.s32.s32 %r400, %r537, %r368, %r396;


dp4a.s32.s32 %r404, %r541, %r372, %r400;


dp4a.s32.s32 %r408, %r545, %r376, %r404;


dp4a.s32.s32 %r412, %r549, %r380, %r408;


{ cvt.f32.f16 %f112, %rs2;}


mul.ftz.f32 %f172, %f112, %f136;

{.reg .f16 low,high;
mov.b32 {low,high},%r382;
cvt.f32.f16 %f113, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r382;
cvt.f32.f16 %f114, high;}


cvt.rn.f32.s32 %f173, %r412;
mul.ftz.f32 %f174, %f140, %f114;
fma.rn.ftz.f32 %f175, %f113, %f173, %f174;
fma.rn.ftz.f32 %f364, %f172, %f175, %f364;
cvt.u64.u32 %rd75, %r1004;
add.s64 %rd76, %rd26, %rd75;
add.s64 %rd77, %rd76, %rd2;
mul.lo.s64 %rd78, %rd77, 36;
add.s64 %rd79, %rd9, %rd78;
ld.global.nc.u32 %r420, [%rd79+4];
ld.global.nc.u32 %r424, [%rd79+8];

dp4a.s32.s32 %r418, %r487, %r420, %r523;


dp4a.s32.s32 %r422, %r491, %r424, %r418;

ld.global.nc.u32 %r428, [%rd79+12];
ld.global.nc.u32 %r432, [%rd79+16];

dp4a.s32.s32 %r426, %r495, %r428, %r422;


dp4a.s32.s32 %r430, %r499, %r432, %r426;

ld.global.nc.u32 %r436, [%rd79+20];
ld.global.nc.u32 %r440, [%rd79+24];

dp4a.s32.s32 %r434, %r503, %r436, %r430;


dp4a.s32.s32 %r438, %r507, %r440, %r434;

ld.global.nc.u32 %r444, [%rd79+28];
ld.global.nc.u32 %r448, [%rd79+32];

dp4a.s32.s32 %r442, %r511, %r444, %r438;


dp4a.s32.s32 %r446, %r515, %r448, %r442;


{ cvt.f32.f16 %f115, %rs1;}


mul.ftz.f32 %f176, %f115, %f127;
ld.global.nc.u32 %r450, [%rd79];

{.reg .f16 low,high;
mov.b32 {low,high},%r450;
cvt.f32.f16 %f116, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r450;
cvt.f32.f16 %f117, high;}


cvt.rn.f32.s32 %f177, %r446;
mul.ftz.f32 %f178, %f132, %f117;
fma.rn.ftz.f32 %f179, %f116, %f177, %f178;
fma.rn.ftz.f32 %f363, %f176, %f179, %f363;

dp4a.s32.s32 %r452, %r521, %r420, %r523;


dp4a.s32.s32 %r456, %r525, %r424, %r452;


dp4a.s32.s32 %r460, %r529, %r428, %r456;


dp4a.s32.s32 %r464, %r533, %r432, %r460;


dp4a.s32.s32 %r468, %r537, %r436, %r464;


dp4a.s32.s32 %r472, %r541, %r440, %r468;


dp4a.s32.s32 %r476, %r545, %r444, %r472;


dp4a.s32.s32 %r480, %r549, %r448, %r476;


{ cvt.f32.f16 %f118, %rs2;}


mul.ftz.f32 %f180, %f118, %f136;

{.reg .f16 low,high;
mov.b32 {low,high},%r450;
cvt.f32.f16 %f119, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r450;
cvt.f32.f16 %f120, high;}


cvt.rn.f32.s32 %f181, %r480;
mul.ftz.f32 %f182, %f140, %f120;
fma.rn.ftz.f32 %f183, %f119, %f181, %f182;
fma.rn.ftz.f32 %f362, %f180, %f183, %f362;
cvt.u64.u32 %rd80, %r1003;
add.s64 %rd81, %rd26, %rd80;
add.s64 %rd82, %rd81, %rd2;
mul.lo.s64 %rd83, %rd82, 36;
add.s64 %rd84, %rd9, %rd83;
ld.global.nc.u32 %r488, [%rd84+4];
ld.global.nc.u32 %r492, [%rd84+8];

dp4a.s32.s32 %r486, %r487, %r488, %r523;


dp4a.s32.s32 %r490, %r491, %r492, %r486;

ld.global.nc.u32 %r496, [%rd84+12];
ld.global.nc.u32 %r500, [%rd84+16];

dp4a.s32.s32 %r494, %r495, %r496, %r490;


dp4a.s32.s32 %r498, %r499, %r500, %r494;

ld.global.nc.u32 %r504, [%rd84+20];
ld.global.nc.u32 %r508, [%rd84+24];

dp4a.s32.s32 %r502, %r503, %r504, %r498;


dp4a.s32.s32 %r506, %r507, %r508, %r502;

ld.global.nc.u32 %r512, [%rd84+28];
ld.global.nc.u32 %r516, [%rd84+32];

dp4a.s32.s32 %r510, %r511, %r512, %r506;


dp4a.s32.s32 %r514, %r515, %r516, %r510;


{ cvt.f32.f16 %f121, %rs1;}


mul.ftz.f32 %f184, %f121, %f127;
ld.global.nc.u32 %r518, [%rd84];

{.reg .f16 low,high;
mov.b32 {low,high},%r518;
cvt.f32.f16 %f122, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r518;
cvt.f32.f16 %f123, high;}


cvt.rn.f32.s32 %f185, %r514;
mul.ftz.f32 %f186, %f132, %f123;
fma.rn.ftz.f32 %f187, %f122, %f185, %f186;
fma.rn.ftz.f32 %f361, %f184, %f187, %f361;

dp4a.s32.s32 %r520, %r521, %r488, %r523;


dp4a.s32.s32 %r524, %r525, %r492, %r520;


dp4a.s32.s32 %r528, %r529, %r496, %r524;


dp4a.s32.s32 %r532, %r533, %r500, %r528;


dp4a.s32.s32 %r536, %r537, %r504, %r532;


dp4a.s32.s32 %r540, %r541, %r508, %r536;


dp4a.s32.s32 %r544, %r545, %r512, %r540;


dp4a.s32.s32 %r548, %r549, %r516, %r544;


{ cvt.f32.f16 %f124, %rs2;}


mul.ftz.f32 %f188, %f124, %f136;

{.reg .f16 low,high;
mov.b32 {low,high},%r518;
cvt.f32.f16 %f125, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r518;
cvt.f32.f16 %f126, high;}


cvt.rn.f32.s32 %f189, %r548;
mul.ftz.f32 %f190, %f140, %f126;
fma.rn.ftz.f32 %f191, %f125, %f189, %f190;
fma.rn.ftz.f32 %f360, %f188, %f191, %f360;
add.s32 %r1007, %r1007, 64;
add.s32 %r1006, %r1006, 64;
add.s32 %r1005, %r1005, 64;
add.s32 %r1004, %r1004, 64;
add.s32 %r1003, %r1003, 64;
add.s32 %r1002, %r1002, 64;
add.s32 %r1001, %r1001, 64;
add.s64 %rd142, %rd142, 400;
add.s32 %r1008, %r1008, 8;
setp.lt.s32 %p3, %r1008, %r50;
@%p3 bra $L__BB261_2;

$L__BB261_3:
setp.eq.s32 %p4, %r1, 0;
@%p4 bra $L__BB261_5;

mov.u32 %r619, _ZZ13mul_mat_vec_qIL9ggml_type19ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r620, %r1, 1792, %r619;
shl.b32 %r621, %r46, 2;
add.s32 %r622, %r620, %r621;
st.shared.f32 [%r622+-1792], %f373;
st.shared.f32 [%r622+-1664], %f372;
st.shared.f32 [%r622+-1536], %f371;
st.shared.f32 [%r622+-1408], %f370;
st.shared.f32 [%r622+-1280], %f369;
st.shared.f32 [%r622+-1152], %f368;
st.shared.f32 [%r622+-1024], %f367;
st.shared.f32 [%r622+-896], %f366;
st.shared.f32 [%r622+-768], %f365;
st.shared.f32 [%r622+-640], %f364;
st.shared.f32 [%r622+-512], %f363;
st.shared.f32 [%r622+-384], %f362;
st.shared.f32 [%r622+-256], %f361;
st.shared.f32 [%r622+-128], %f360;

$L__BB261_5:
setp.ne.s32 %p5, %r1, 0;
bar.sync 0;
@%p5 bra $L__BB261_34;

mov.u32 %r624, %ctaid.x;
shl.b32 %r625, %r624, 1;
mov.u32 %r626, 1;
shl.b32 %r628, %r46, 2;
mov.u32 %r629, _ZZ13mul_mat_vec_qIL9ggml_type19ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r630, %r629, %r628;
mov.u32 %r631, 2;
ld.shared.f32 %f192, [%r630];
add.ftz.f32 %f193, %f192, %f373;
mov.b32 %r632, %f193;
mov.u32 %r633, 31;
mov.u32 %r634, 16;
mov.u32 %r635, -1;
shfl.sync.bfly.b32 %r636|%p6, %r632, %r634, %r633, %r635;
mov.b32 %f194, %r636;
add.ftz.f32 %f195, %f193, %f194;
mov.b32 %r637, %f195;
mov.u32 %r638, 8;
shfl.sync.bfly.b32 %r639|%p7, %r637, %r638, %r633, %r635;
mov.b32 %f196, %r639;
add.ftz.f32 %f197, %f195, %f196;
mov.b32 %r640, %f197;
mov.u32 %r641, 4;
shfl.sync.bfly.b32 %r642|%p8, %r640, %r641, %r633, %r635;
mov.b32 %f198, %r642;
add.ftz.f32 %f199, %f197, %f198;
mov.b32 %r643, %f199;
shfl.sync.bfly.b32 %r644|%p9, %r643, %r631, %r633, %r635;
mov.b32 %f200, %r644;
add.ftz.f32 %f201, %f199, %f200;
mov.b32 %r645, %f201;
shfl.sync.bfly.b32 %r646|%p10, %r645, %r626, %r633, %r635;
mov.b32 %f202, %r646;
add.ftz.f32 %f43, %f201, %f202;
setp.ne.s32 %p11, %r46, 0;
setp.ge.u32 %p12, %r625, %r32;
or.pred %p13, %p11, %p12;
@%p13 bra $L__BB261_8;

cvta.to.global.u64 %rd85, %rd17;
mov.u32 %r647, %ctaid.y;
mad.lo.s32 %r650, %r647, %r38, %r625;
mov.u32 %r651, %ctaid.z;
mad.lo.s32 %r652, %r651, %r44, %r650;
mul.wide.u32 %rd86, %r652, 4;
add.s64 %rd87, %rd85, %rd86;
st.global.f32 [%rd87], %f43;

$L__BB261_8:
ld.shared.f32 %f203, [%r630+128];
add.ftz.f32 %f204, %f203, %f372;
mov.b32 %r658, %f204;
shfl.sync.bfly.b32 %r662|%p14, %r658, %r634, %r633, %r635;
mov.b32 %f205, %r662;
add.ftz.f32 %f206, %f204, %f205;
mov.b32 %r663, %f206;
shfl.sync.bfly.b32 %r665|%p15, %r663, %r638, %r633, %r635;
mov.b32 %f207, %r665;
add.ftz.f32 %f208, %f206, %f207;
mov.b32 %r666, %f208;
shfl.sync.bfly.b32 %r668|%p16, %r666, %r641, %r633, %r635;
mov.b32 %f209, %r668;
add.ftz.f32 %f210, %f208, %f209;
mov.b32 %r669, %f210;
shfl.sync.bfly.b32 %r670|%p17, %r669, %r631, %r633, %r635;
mov.b32 %f211, %r670;
add.ftz.f32 %f212, %f210, %f211;
mov.b32 %r671, %f212;
shfl.sync.bfly.b32 %r673|%p18, %r671, %r626, %r633, %r635;
mov.b32 %f213, %r673;
add.ftz.f32 %f44, %f212, %f213;
setp.eq.s32 %p19, %r46, 1;
or.b32 %r676, %r625, 1;
setp.lt.u32 %p20, %r676, %r32;
and.pred %p1, %p19, %p20;
not.pred %p21, %p1;
@%p21 bra $L__BB261_10;

cvta.to.global.u64 %rd88, %rd17;
mov.u32 %r677, %ctaid.y;
mad.lo.s32 %r680, %r677, %r38, %r625;
mov.u32 %r681, %ctaid.z;
mad.lo.s32 %r682, %r681, %r44, %r680;
mul.wide.u32 %rd89, %r682, 4;
add.s64 %rd90, %rd88, %rd89;
st.global.f32 [%rd90+4], %f44;

$L__BB261_10:
mov.u32 %r687, 2;
ld.shared.f32 %f214, [%r630+256];
add.ftz.f32 %f215, %f214, %f371;
mov.b32 %r688, %f215;
mov.u32 %r689, 31;
mov.u32 %r690, 16;
mov.u32 %r691, -1;
shfl.sync.bfly.b32 %r692|%p22, %r688, %r690, %r689, %r691;
mov.b32 %f216, %r692;
add.ftz.f32 %f217, %f215, %f216;
mov.b32 %r693, %f217;
mov.u32 %r694, 8;
shfl.sync.bfly.b32 %r695|%p23, %r693, %r694, %r689, %r691;
mov.b32 %f218, %r695;
add.ftz.f32 %f219, %f217, %f218;
mov.b32 %r696, %f219;
mov.u32 %r697, 4;
shfl.sync.bfly.b32 %r698|%p24, %r696, %r697, %r689, %r691;
mov.b32 %f220, %r698;
add.ftz.f32 %f221, %f219, %f220;
mov.b32 %r699, %f221;
shfl.sync.bfly.b32 %r700|%p25, %r699, %r687, %r689, %r691;
mov.b32 %f222, %r700;
add.ftz.f32 %f223, %f221, %f222;
mov.b32 %r701, %f223;
mov.u32 %r702, 1;
shfl.sync.bfly.b32 %r703|%p26, %r701, %r702, %r689, %r691;
mov.b32 %f224, %r703;
add.ftz.f32 %f45, %f223, %f224;
@%p13 bra $L__BB261_12;

cvt.u64.u32 %rd91, %r32;
mov.u32 %r706, %ctaid.y;
mad.lo.s32 %r709, %r706, %r38, %r625;
mov.u32 %r710, %ctaid.z;
mad.lo.s32 %r711, %r710, %r44, %r709;
cvt.u64.u32 %rd92, %r711;
add.s64 %rd93, %rd91, %rd92;
cvta.to.global.u64 %rd94, %rd17;
shl.b64 %rd95, %rd93, 2;
add.s64 %rd96, %rd94, %rd95;
st.global.f32 [%rd96], %f45;

$L__BB261_12:
ld.shared.f32 %f225, [%r630+384];
add.ftz.f32 %f226, %f225, %f370;
mov.b32 %r717, %f226;
shfl.sync.bfly.b32 %r721|%p30, %r717, %r690, %r689, %r691;
mov.b32 %f227, %r721;
add.ftz.f32 %f228, %f226, %f227;
mov.b32 %r722, %f228;
shfl.sync.bfly.b32 %r724|%p31, %r722, %r694, %r689, %r691;
mov.b32 %f229, %r724;
add.ftz.f32 %f230, %f228, %f229;
mov.b32 %r725, %f230;
shfl.sync.bfly.b32 %r727|%p32, %r725, %r697, %r689, %r691;
mov.b32 %f231, %r727;
add.ftz.f32 %f232, %f230, %f231;
mov.b32 %r728, %f232;
shfl.sync.bfly.b32 %r729|%p33, %r728, %r687, %r689, %r691;
mov.b32 %f233, %r729;
add.ftz.f32 %f234, %f232, %f233;
mov.b32 %r730, %f234;
shfl.sync.bfly.b32 %r732|%p34, %r730, %r702, %r689, %r691;
mov.b32 %f235, %r732;
add.ftz.f32 %f46, %f234, %f235;
@%p21 bra $L__BB261_14;

add.s32 %r733, %r32, 1;
cvt.u64.u32 %rd97, %r733;
mov.u32 %r734, %ctaid.y;
mad.lo.s32 %r737, %r734, %r38, %r625;
mov.u32 %r738, %ctaid.z;
mad.lo.s32 %r739, %r738, %r44, %r737;
cvt.u64.u32 %rd98, %r739;
add.s64 %rd99, %rd97, %rd98;
cvta.to.global.u64 %rd100, %rd17;
shl.b64 %rd101, %rd99, 2;
add.s64 %rd102, %rd100, %rd101;
st.global.f32 [%rd102], %f46;

$L__BB261_14:
shl.b32 %r740, %r32, 1;
mov.u32 %r741, 1;
mov.u32 %r746, 2;
ld.shared.f32 %f236, [%r630+512];
add.ftz.f32 %f237, %f236, %f369;
mov.b32 %r747, %f237;
mov.u32 %r748, 31;
mov.u32 %r749, 16;
mov.u32 %r750, -1;
shfl.sync.bfly.b32 %r751|%p36, %r747, %r749, %r748, %r750;
mov.b32 %f238, %r751;
add.ftz.f32 %f239, %f237, %f238;
mov.b32 %r752, %f239;
mov.u32 %r753, 8;
shfl.sync.bfly.b32 %r754|%p37, %r752, %r753, %r748, %r750;
mov.b32 %f240, %r754;
add.ftz.f32 %f241, %f239, %f240;
mov.b32 %r755, %f241;
mov.u32 %r756, 4;
shfl.sync.bfly.b32 %r757|%p38, %r755, %r756, %r748, %r750;
mov.b32 %f242, %r757;
add.ftz.f32 %f243, %f241, %f242;
mov.b32 %r758, %f243;
shfl.sync.bfly.b32 %r759|%p39, %r758, %r746, %r748, %r750;
mov.b32 %f244, %r759;
add.ftz.f32 %f245, %f243, %f244;
mov.b32 %r760, %f245;
shfl.sync.bfly.b32 %r761|%p40, %r760, %r741, %r748, %r750;
mov.b32 %f246, %r761;
add.ftz.f32 %f47, %f245, %f246;
cvt.u64.u32 %rd103, %r740;
mov.u32 %r762, %ctaid.y;
mad.lo.s32 %r765, %r762, %r38, %r625;
mov.u32 %r766, %ctaid.z;
mad.lo.s32 %r767, %r766, %r44, %r765;
cvt.u64.u32 %rd104, %r767;
add.s64 %rd105, %rd103, %rd104;
cvta.to.global.u64 %rd106, %rd17;
shl.b64 %rd107, %rd105, 2;
add.s64 %rd12, %rd106, %rd107;
@%p13 bra $L__BB261_16;

st.global.f32 [%rd12], %f47;

$L__BB261_16:
ld.shared.f32 %f247, [%r630+640];
add.ftz.f32 %f248, %f247, %f368;
mov.b32 %r773, %f248;
shfl.sync.bfly.b32 %r777|%p44, %r773, %r749, %r748, %r750;
mov.b32 %f249, %r777;
add.ftz.f32 %f250, %f248, %f249;
mov.b32 %r778, %f250;
shfl.sync.bfly.b32 %r780|%p45, %r778, %r753, %r748, %r750;
mov.b32 %f251, %r780;
add.ftz.f32 %f252, %f250, %f251;
mov.b32 %r781, %f252;
shfl.sync.bfly.b32 %r783|%p46, %r781, %r756, %r748, %r750;
mov.b32 %f253, %r783;
add.ftz.f32 %f254, %f252, %f253;
mov.b32 %r784, %f254;
shfl.sync.bfly.b32 %r785|%p47, %r784, %r746, %r748, %r750;
mov.b32 %f255, %r785;
add.ftz.f32 %f256, %f254, %f255;
mov.b32 %r786, %f256;
shfl.sync.bfly.b32 %r788|%p48, %r786, %r741, %r748, %r750;
mov.b32 %f257, %r788;
add.ftz.f32 %f48, %f256, %f257;
@%p21 bra $L__BB261_18;

st.global.f32 [%rd12+4], %f48;

$L__BB261_18:
mul.lo.s32 %r27, %r32, 3;
mov.u32 %r793, 2;
ld.shared.f32 %f258, [%r630+768];
add.ftz.f32 %f259, %f258, %f367;
mov.b32 %r794, %f259;
mov.u32 %r795, 31;
mov.u32 %r796, 16;
mov.u32 %r797, -1;
shfl.sync.bfly.b32 %r798|%p50, %r794, %r796, %r795, %r797;
mov.b32 %f260, %r798;
add.ftz.f32 %f261, %f259, %f260;
mov.b32 %r799, %f261;
mov.u32 %r800, 8;
shfl.sync.bfly.b32 %r801|%p51, %r799, %r800, %r795, %r797;
mov.b32 %f262, %r801;
add.ftz.f32 %f263, %f261, %f262;
mov.b32 %r802, %f263;
mov.u32 %r803, 4;
shfl.sync.bfly.b32 %r804|%p52, %r802, %r803, %r795, %r797;
mov.b32 %f264, %r804;
add.ftz.f32 %f265, %f263, %f264;
mov.b32 %r805, %f265;
shfl.sync.bfly.b32 %r806|%p53, %r805, %r793, %r795, %r797;
mov.b32 %f266, %r806;
add.ftz.f32 %f267, %f265, %f266;
mov.b32 %r807, %f267;
mov.u32 %r808, 1;
shfl.sync.bfly.b32 %r809|%p54, %r807, %r808, %r795, %r797;
mov.b32 %f268, %r809;
add.ftz.f32 %f49, %f267, %f268;
@%p13 bra $L__BB261_20;

cvt.u64.u32 %rd108, %r27;
add.s64 %rd110, %rd108, %rd104;
shl.b64 %rd112, %rd110, 2;
add.s64 %rd113, %rd106, %rd112;
st.global.f32 [%rd113], %f49;

$L__BB261_20:
ld.shared.f32 %f269, [%r630+896];
add.ftz.f32 %f270, %f269, %f366;
mov.b32 %r823, %f270;
shfl.sync.bfly.b32 %r827|%p58, %r823, %r796, %r795, %r797;
mov.b32 %f271, %r827;
add.ftz.f32 %f272, %f270, %f271;
mov.b32 %r828, %f272;
shfl.sync.bfly.b32 %r830|%p59, %r828, %r800, %r795, %r797;
mov.b32 %f273, %r830;
add.ftz.f32 %f274, %f272, %f273;
mov.b32 %r831, %f274;
shfl.sync.bfly.b32 %r833|%p60, %r831, %r803, %r795, %r797;
mov.b32 %f275, %r833;
add.ftz.f32 %f276, %f274, %f275;
mov.b32 %r834, %f276;
shfl.sync.bfly.b32 %r835|%p61, %r834, %r793, %r795, %r797;
mov.b32 %f277, %r835;
add.ftz.f32 %f278, %f276, %f277;
mov.b32 %r836, %f278;
shfl.sync.bfly.b32 %r838|%p62, %r836, %r808, %r795, %r797;
mov.b32 %f279, %r838;
add.ftz.f32 %f50, %f278, %f279;
@%p21 bra $L__BB261_22;

add.s32 %r839, %r27, 1;
cvt.u64.u32 %rd114, %r839;
add.s64 %rd116, %rd114, %rd104;
shl.b64 %rd118, %rd116, 2;
add.s64 %rd119, %rd106, %rd118;
st.global.f32 [%rd119], %f50;

$L__BB261_22:
shl.b32 %r846, %r32, 2;
mov.u32 %r847, 2;
ld.shared.f32 %f280, [%r630+1024];
add.ftz.f32 %f281, %f280, %f365;
mov.b32 %r852, %f281;
mov.u32 %r853, 31;
mov.u32 %r854, 16;
mov.u32 %r855, -1;
shfl.sync.bfly.b32 %r856|%p64, %r852, %r854, %r853, %r855;
mov.b32 %f282, %r856;
add.ftz.f32 %f283, %f281, %f282;
mov.b32 %r857, %f283;
mov.u32 %r858, 8;
shfl.sync.bfly.b32 %r859|%p65, %r857, %r858, %r853, %r855;
mov.b32 %f284, %r859;
add.ftz.f32 %f285, %f283, %f284;
mov.b32 %r860, %f285;
mov.u32 %r861, 4;
shfl.sync.bfly.b32 %r862|%p66, %r860, %r861, %r853, %r855;
mov.b32 %f286, %r862;
add.ftz.f32 %f287, %f285, %f286;
mov.b32 %r863, %f287;
shfl.sync.bfly.b32 %r864|%p67, %r863, %r847, %r853, %r855;
mov.b32 %f288, %r864;
add.ftz.f32 %f289, %f287, %f288;
mov.b32 %r865, %f289;
mov.u32 %r866, 1;
shfl.sync.bfly.b32 %r867|%p68, %r865, %r866, %r853, %r855;
mov.b32 %f290, %r867;
add.ftz.f32 %f51, %f289, %f290;
cvt.u64.u32 %rd120, %r846;
add.s64 %rd122, %rd120, %rd104;
shl.b64 %rd124, %rd122, 2;
add.s64 %rd13, %rd106, %rd124;
@%p13 bra $L__BB261_24;

st.global.f32 [%rd13], %f51;

$L__BB261_24:
ld.shared.f32 %f291, [%r630+1152];
add.ftz.f32 %f292, %f291, %f364;
mov.b32 %r879, %f292;
shfl.sync.bfly.b32 %r883|%p72, %r879, %r854, %r853, %r855;
mov.b32 %f293, %r883;
add.ftz.f32 %f294, %f292, %f293;
mov.b32 %r884, %f294;
shfl.sync.bfly.b32 %r886|%p73, %r884, %r858, %r853, %r855;
mov.b32 %f295, %r886;
add.ftz.f32 %f296, %f294, %f295;
mov.b32 %r887, %f296;
shfl.sync.bfly.b32 %r889|%p74, %r887, %r861, %r853, %r855;
mov.b32 %f297, %r889;
add.ftz.f32 %f298, %f296, %f297;
mov.b32 %r890, %f298;
shfl.sync.bfly.b32 %r891|%p75, %r890, %r847, %r853, %r855;
mov.b32 %f299, %r891;
add.ftz.f32 %f300, %f298, %f299;
mov.b32 %r892, %f300;
shfl.sync.bfly.b32 %r894|%p76, %r892, %r866, %r853, %r855;
mov.b32 %f301, %r894;
add.ftz.f32 %f52, %f300, %f301;
@%p21 bra $L__BB261_26;

st.global.f32 [%rd13+4], %f52;

$L__BB261_26:
mul.lo.s32 %r28, %r32, 5;
mov.u32 %r899, 2;
ld.shared.f32 %f302, [%r630+1280];
add.ftz.f32 %f303, %f302, %f363;
mov.b32 %r900, %f303;
mov.u32 %r901, 31;
mov.u32 %r902, 16;
mov.u32 %r903, -1;
shfl.sync.bfly.b32 %r904|%p78, %r900, %r902, %r901, %r903;
mov.b32 %f304, %r904;
add.ftz.f32 %f305, %f303, %f304;
mov.b32 %r905, %f305;
mov.u32 %r906, 8;
shfl.sync.bfly.b32 %r907|%p79, %r905, %r906, %r901, %r903;
mov.b32 %f306, %r907;
add.ftz.f32 %f307, %f305, %f306;
mov.b32 %r908, %f307;
mov.u32 %r909, 4;
shfl.sync.bfly.b32 %r910|%p80, %r908, %r909, %r901, %r903;
mov.b32 %f308, %r910;
add.ftz.f32 %f309, %f307, %f308;
mov.b32 %r911, %f309;
shfl.sync.bfly.b32 %r912|%p81, %r911, %r899, %r901, %r903;
mov.b32 %f310, %r912;
add.ftz.f32 %f311, %f309, %f310;
mov.b32 %r913, %f311;
mov.u32 %r914, 1;
shfl.sync.bfly.b32 %r915|%p82, %r913, %r914, %r901, %r903;
mov.b32 %f312, %r915;
add.ftz.f32 %f53, %f311, %f312;
@%p13 bra $L__BB261_28;

cvt.u64.u32 %rd125, %r28;
add.s64 %rd127, %rd125, %rd104;
shl.b64 %rd129, %rd127, 2;
add.s64 %rd130, %rd106, %rd129;
st.global.f32 [%rd130], %f53;

$L__BB261_28:
ld.shared.f32 %f313, [%r630+1408];
add.ftz.f32 %f314, %f313, %f362;
mov.b32 %r929, %f314;
shfl.sync.bfly.b32 %r933|%p86, %r929, %r902, %r901, %r903;
mov.b32 %f315, %r933;
add.ftz.f32 %f316, %f314, %f315;
mov.b32 %r934, %f316;
shfl.sync.bfly.b32 %r936|%p87, %r934, %r906, %r901, %r903;
mov.b32 %f317, %r936;
add.ftz.f32 %f318, %f316, %f317;
mov.b32 %r937, %f318;
shfl.sync.bfly.b32 %r939|%p88, %r937, %r909, %r901, %r903;
mov.b32 %f319, %r939;
add.ftz.f32 %f320, %f318, %f319;
mov.b32 %r940, %f320;
shfl.sync.bfly.b32 %r941|%p89, %r940, %r899, %r901, %r903;
mov.b32 %f321, %r941;
add.ftz.f32 %f322, %f320, %f321;
mov.b32 %r942, %f322;
shfl.sync.bfly.b32 %r944|%p90, %r942, %r914, %r901, %r903;
mov.b32 %f323, %r944;
add.ftz.f32 %f54, %f322, %f323;
@%p21 bra $L__BB261_30;

add.s32 %r945, %r28, 1;
cvt.u64.u32 %rd131, %r945;
add.s64 %rd133, %rd131, %rd104;
shl.b64 %rd135, %rd133, 2;
add.s64 %rd136, %rd106, %rd135;
st.global.f32 [%rd136], %f54;

$L__BB261_30:
mul.lo.s32 %r952, %r32, 6;
mov.u32 %r957, 2;
ld.shared.f32 %f324, [%r630+1536];
add.ftz.f32 %f325, %f324, %f361;
mov.b32 %r958, %f325;
mov.u32 %r959, 31;
mov.u32 %r960, 16;
mov.u32 %r961, -1;
shfl.sync.bfly.b32 %r962|%p92, %r958, %r960, %r959, %r961;
mov.b32 %f326, %r962;
add.ftz.f32 %f327, %f325, %f326;
mov.b32 %r963, %f327;
mov.u32 %r964, 8;
shfl.sync.bfly.b32 %r965|%p93, %r963, %r964, %r959, %r961;
mov.b32 %f328, %r965;
add.ftz.f32 %f329, %f327, %f328;
mov.b32 %r966, %f329;
mov.u32 %r967, 4;
shfl.sync.bfly.b32 %r968|%p94, %r966, %r967, %r959, %r961;
mov.b32 %f330, %r968;
add.ftz.f32 %f331, %f329, %f330;
mov.b32 %r969, %f331;
shfl.sync.bfly.b32 %r970|%p95, %r969, %r957, %r959, %r961;
mov.b32 %f332, %r970;
add.ftz.f32 %f333, %f331, %f332;
mov.b32 %r971, %f333;
mov.u32 %r972, 1;
shfl.sync.bfly.b32 %r973|%p96, %r971, %r972, %r959, %r961;
mov.b32 %f334, %r973;
add.ftz.f32 %f55, %f333, %f334;
cvt.u64.u32 %rd137, %r952;
add.s64 %rd139, %rd137, %rd104;
shl.b64 %rd141, %rd139, 2;
add.s64 %rd14, %rd106, %rd141;
@%p13 bra $L__BB261_32;

st.global.f32 [%rd14], %f55;

$L__BB261_32:
ld.shared.f32 %f335, [%r630+1664];
add.ftz.f32 %f336, %f335, %f360;
mov.b32 %r985, %f336;
shfl.sync.bfly.b32 %r989|%p100, %r985, %r960, %r959, %r961;
mov.b32 %f337, %r989;
add.ftz.f32 %f338, %f336, %f337;
mov.b32 %r990, %f338;
shfl.sync.bfly.b32 %r992|%p101, %r990, %r964, %r959, %r961;
mov.b32 %f339, %r992;
add.ftz.f32 %f340, %f338, %f339;
mov.b32 %r993, %f340;
shfl.sync.bfly.b32 %r995|%p102, %r993, %r967, %r959, %r961;
mov.b32 %f341, %r995;
add.ftz.f32 %f342, %f340, %f341;
mov.b32 %r996, %f342;
shfl.sync.bfly.b32 %r997|%p103, %r996, %r957, %r959, %r961;
mov.b32 %f343, %r997;
add.ftz.f32 %f344, %f342, %f343;
mov.b32 %r998, %f344;
shfl.sync.bfly.b32 %r1000|%p104, %r998, %r972, %r959, %r961;
mov.b32 %f345, %r1000;
add.ftz.f32 %f56, %f344, %f345;
@%p21 bra $L__BB261_34;

st.global.f32 [%rd14+4], %f56;

$L__BB261_34:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type19ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type19ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type19ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type19ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type19ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type19ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type19ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type19ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type19ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type19ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<120>;
.reg .b16 %rs<31>;
.reg .f32 %f<426>;
.reg .b32 %r<1140>;
.reg .b64 %rd<160>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type19ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[2048];

ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type19ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type19ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type19ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type19ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type19ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type19ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type19ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type19ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type19ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type19ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r46, [_Z13mul_mat_vec_qIL9ggml_type19ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r47, [_Z13mul_mat_vec_qIL9ggml_type19ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r48, [_Z13mul_mat_vec_qIL9ggml_type19ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type19ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type19ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type19ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type19ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r49, %r1, 5;
mov.u32 %r50, %tid.x;
add.s32 %r2, %r49, %r50;
shr.s32 %r51, %r2, 31;
shr.u32 %r52, %r51, 29;
add.s32 %r53, %r2, %r52;
shr.s32 %r1139, %r53, 3;
shr.u32 %r54, %r33, 8;
setp.ge.s32 %p2, %r1139, %r54;
mov.f32 %f410, 0f00000000;
mov.f32 %f411, %f410;
mov.f32 %f412, %f410;
mov.f32 %f413, %f410;
mov.f32 %f414, %f410;
mov.f32 %f415, %f410;
mov.f32 %f416, %f410;
mov.f32 %f417, %f410;
mov.f32 %f418, %f410;
mov.f32 %f419, %f410;
mov.f32 %f420, %f410;
mov.f32 %f421, %f410;
mov.f32 %f422, %f410;
mov.f32 %f423, %f410;
mov.f32 %f424, %f410;
mov.f32 %f425, %f410;
@%p2 bra $L__BB262_3;

mov.u32 %r55, %ctaid.x;
shl.b32 %r56, %r55, 1;
mov.u32 %r57, %ctaid.y;
mov.u32 %r58, %ctaid.z;
mul.hi.u32 %r59, %r58, %r43;
add.s32 %r60, %r59, %r58;
shr.u32 %r61, %r60, %r44;
mul.lo.s32 %r62, %r61, %r46;
mul.hi.u32 %r63, %r57, %r37;
add.s32 %r64, %r63, %r57;
shr.u32 %r65, %r64, %r38;
mul.lo.s32 %r66, %r65, %r40;
cvta.to.global.u64 %rd159, %rd15;
and.b32 %r70, %r53, -8;
sub.s32 %r71, %r2, %r70;
shl.b32 %r72, %r71, 1;
cvt.s64.s32 %rd2, %r71;
shl.b32 %r73, %r35, 1;
shl.b32 %r1131, %r1139, 3;
add.s32 %r1138, %r1131, %r73;
mad.lo.s32 %r1137, %r35, 3, %r1131;
shl.b32 %r74, %r35, 2;
add.s32 %r1136, %r1131, %r74;
mad.lo.s32 %r1135, %r35, 5, %r1131;
mad.lo.s32 %r1134, %r35, 6, %r1131;
mad.lo.s32 %r1133, %r35, 7, %r1131;
add.s32 %r1132, %r35, %r1131;
add.s32 %r75, %r1139, %r62;
add.s32 %r76, %r75, %r66;
or.b32 %r77, %r56, 1;
mad.lo.s32 %r78, %r34, %r77, %r76;
mul.wide.s32 %rd3, %r78, 50;
mul.wide.s32 %rd18, %r71, 2;
add.s64 %rd19, %rd3, %rd18;
add.s64 %rd4, %rd19, 34;
mad.lo.s32 %r79, %r56, %r34, %r1139;
add.s32 %r80, %r79, %r62;
add.s32 %r81, %r80, %r66;
mul.wide.s32 %rd5, %r81, 50;
mul.wide.s32 %rd20, %r72, 2;
add.s64 %rd21, %rd5, %rd20;
add.s64 %rd6, %rd21, 4;
add.s64 %rd22, %rd3, %rd20;
add.s64 %rd7, %rd22, 2;
add.s64 %rd23, %rd5, %rd18;
add.s64 %rd8, %rd23, 34;
cvta.to.global.u64 %rd9, %rd16;
mov.f32 %f410, 0f00000000;

$L__BB262_2:
mul.lo.s32 %r627, %r58, %r47;
cvt.u64.u32 %rd24, %r627;
mul.lo.s32 %r629, %r57, %r41;
cvt.u64.u32 %rd25, %r629;
add.s64 %rd26, %rd24, %rd25;
cvt.u64.u32 %rd27, %r1131;
add.s64 %rd28, %rd26, %rd27;
add.s64 %rd29, %rd28, %rd2;
mul.lo.s64 %rd30, %rd29, 36;
add.s64 %rd31, %rd159, %rd6;
ld.global.nc.u16 %rs17, [%rd31+-2];
shr.u16 %rs18, %rs17, 8;
cvt.u32.u16 %r630, %rs18;
ld.global.nc.u16 %rs19, [%rd31];
shr.u16 %rs20, %rs19, 8;
cvt.u32.u16 %r631, %rs20;
add.s64 %rd32, %rd159, %rd8;
ld.global.nc.u16 %rs21, [%rd32];
cvt.u32.u16 %r632, %rs21;
and.b16 %rs22, %rs17, 255;
cvt.u32.u16 %r633, %rs22;
mul.wide.u16 %r634, %rs21, 256;
and.b32 %r635, %r634, 1792;
or.b32 %r636, %r635, %r633;
mul.wide.u32 %rd33, %r636, 4;
mov.u64 %rd34, iq1s_grid_gpu;
add.s64 %rd35, %rd34, %rd33;
ld.global.nc.u32 %r637, [%rd35];
and.b32 %r559, %r637, 252645135;
shr.u32 %r638, %r637, 4;
and.b32 %r563, %r638, 252645135;
add.s64 %rd36, %rd9, %rd30;
ld.global.nc.u32 %r84, [%rd36+4];
ld.global.nc.u32 %r88, [%rd36+8];
mov.u32 %r595, 0;

dp4a.s32.s32 %r82, %r559, %r84, %r595;


dp4a.s32.s32 %r86, %r563, %r88, %r82;

mul.wide.u16 %r639, %rs21, 32;
and.b32 %r640, %r639, 1792;
or.b32 %r641, %r640, %r630;
mul.wide.u32 %rd37, %r641, 4;
add.s64 %rd38, %rd34, %rd37;
ld.global.nc.u32 %r642, [%rd38];
and.b32 %r567, %r642, 252645135;
shr.u32 %r643, %r642, 4;
and.b32 %r571, %r643, 252645135;
ld.global.nc.u32 %r92, [%rd36+12];
ld.global.nc.u32 %r96, [%rd36+16];

dp4a.s32.s32 %r90, %r567, %r92, %r86;


dp4a.s32.s32 %r94, %r571, %r96, %r90;

and.b16 %rs23, %rs19, 255;
cvt.u32.u16 %r644, %rs23;
mul.wide.u16 %r645, %rs21, 4;
and.b32 %r646, %r645, 1792;
or.b32 %r647, %r646, %r644;
mul.wide.u32 %rd39, %r647, 4;
add.s64 %rd40, %rd34, %rd39;
ld.global.nc.u32 %r648, [%rd40];
and.b32 %r575, %r648, 252645135;
shr.u32 %r649, %r648, 4;
and.b32 %r579, %r649, 252645135;
ld.global.nc.u32 %r100, [%rd36+20];
ld.global.nc.u32 %r104, [%rd36+24];

dp4a.s32.s32 %r98, %r575, %r100, %r94;


dp4a.s32.s32 %r102, %r579, %r104, %r98;

shr.u32 %r650, %r632, 1;
and.b32 %r651, %r650, 1792;
or.b32 %r652, %r651, %r631;
mul.wide.u32 %rd41, %r652, 4;
add.s64 %rd42, %rd34, %rd41;
ld.global.nc.u32 %r653, [%rd42];
and.b32 %r583, %r653, 252645135;
shr.u32 %r654, %r653, 4;
and.b32 %r587, %r654, 252645135;
ld.global.nc.u32 %r108, [%rd36+28];
ld.global.nc.u32 %r112, [%rd36+32];

dp4a.s32.s32 %r106, %r583, %r108, %r102;


dp4a.s32.s32 %r110, %r587, %r112, %r106;

add.s64 %rd43, %rd159, %rd5;
ld.global.nc.u16 %rs1, [%rd43];

{ cvt.f32.f16 %f97, %rs1;}


shr.u32 %r655, %r632, 11;
and.b32 %r656, %r655, 14;
or.b32 %r657, %r656, 1;
cvt.rn.f32.s32 %f145, %r657;
mul.ftz.f32 %f146, %f97, %f145;
and.b32 %r658, %r632, 32768;
cvt.rn.f32.s32 %f147, %r658;
mul.ftz.f32 %f148, %f147, 0f37000000;
mov.f32 %f149, 0fBF600000;
sub.ftz.f32 %f150, %f149, %f148;
ld.global.nc.u32 %r114, [%rd36];

{.reg .f16 low,high;
mov.b32 {low,high},%r114;
cvt.f32.f16 %f98, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r114;
cvt.f32.f16 %f99, high;}


cvt.rn.f32.s32 %f151, %r110;
mul.ftz.f32 %f152, %f150, %f99;
fma.rn.ftz.f32 %f153, %f98, %f151, %f152;
fma.rn.ftz.f32 %f425, %f146, %f153, %f425;
add.s64 %rd44, %rd159, %rd7;
ld.global.nc.u16 %rs24, [%rd44];
shr.u16 %rs25, %rs24, 8;
cvt.u32.u16 %r659, %rs25;
ld.global.nc.u16 %rs26, [%rd44+2];
shr.u16 %rs27, %rs26, 8;
cvt.u32.u16 %r660, %rs27;
add.s64 %rd45, %rd159, %rd4;
ld.global.nc.u16 %rs28, [%rd45];
cvt.u32.u16 %r661, %rs28;
and.b16 %rs29, %rs24, 255;
cvt.u32.u16 %r662, %rs29;
mul.wide.u16 %r663, %rs28, 256;
and.b32 %r664, %r663, 1792;
or.b32 %r665, %r664, %r662;
mul.wide.u32 %rd46, %r665, 4;
add.s64 %rd47, %rd34, %rd46;
ld.global.nc.u32 %r666, [%rd47];
and.b32 %r593, %r666, 252645135;
shr.u32 %r667, %r666, 4;
and.b32 %r597, %r667, 252645135;

dp4a.s32.s32 %r116, %r593, %r84, %r595;


dp4a.s32.s32 %r120, %r597, %r88, %r116;

mul.wide.u16 %r668, %rs28, 32;
and.b32 %r669, %r668, 1792;
or.b32 %r670, %r669, %r659;
mul.wide.u32 %rd48, %r670, 4;
add.s64 %rd49, %rd34, %rd48;
ld.global.nc.u32 %r671, [%rd49];
and.b32 %r601, %r671, 252645135;
shr.u32 %r672, %r671, 4;
and.b32 %r605, %r672, 252645135;

dp4a.s32.s32 %r124, %r601, %r92, %r120;


dp4a.s32.s32 %r128, %r605, %r96, %r124;

and.b16 %rs30, %rs26, 255;
cvt.u32.u16 %r673, %rs30;
mul.wide.u16 %r674, %rs28, 4;
and.b32 %r675, %r674, 1792;
or.b32 %r676, %r675, %r673;
mul.wide.u32 %rd50, %r676, 4;
add.s64 %rd51, %rd34, %rd50;
ld.global.nc.u32 %r677, [%rd51];
and.b32 %r609, %r677, 252645135;
shr.u32 %r678, %r677, 4;
and.b32 %r613, %r678, 252645135;

dp4a.s32.s32 %r132, %r609, %r100, %r128;


dp4a.s32.s32 %r136, %r613, %r104, %r132;

shr.u32 %r679, %r661, 1;
and.b32 %r680, %r679, 1792;
or.b32 %r681, %r680, %r660;
mul.wide.u32 %rd52, %r681, 4;
add.s64 %rd53, %rd34, %rd52;
ld.global.nc.u32 %r682, [%rd53];
and.b32 %r617, %r682, 252645135;
shr.u32 %r683, %r682, 4;
and.b32 %r621, %r683, 252645135;

dp4a.s32.s32 %r140, %r617, %r108, %r136;


dp4a.s32.s32 %r144, %r621, %r112, %r140;

add.s64 %rd54, %rd159, %rd3;
ld.global.nc.u16 %rs2, [%rd54];

{ cvt.f32.f16 %f100, %rs2;}


shr.u32 %r684, %r661, 11;
and.b32 %r685, %r684, 14;
or.b32 %r686, %r685, 1;
cvt.rn.f32.s32 %f154, %r686;
mul.ftz.f32 %f155, %f100, %f154;
and.b32 %r687, %r661, 32768;
cvt.rn.f32.s32 %f156, %r687;
mul.ftz.f32 %f157, %f156, 0f37000000;
sub.ftz.f32 %f158, %f149, %f157;

{.reg .f16 low,high;
mov.b32 {low,high},%r114;
cvt.f32.f16 %f101, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r114;
cvt.f32.f16 %f102, high;}


cvt.rn.f32.s32 %f159, %r144;
mul.ftz.f32 %f160, %f158, %f102;
fma.rn.ftz.f32 %f161, %f101, %f159, %f160;
fma.rn.ftz.f32 %f424, %f155, %f161, %f424;
cvt.u64.u32 %rd55, %r1132;
add.s64 %rd56, %rd26, %rd55;
add.s64 %rd57, %rd56, %rd2;
mul.lo.s64 %rd58, %rd57, 36;
add.s64 %rd59, %rd9, %rd58;
ld.global.nc.u32 %r152, [%rd59+4];
ld.global.nc.u32 %r156, [%rd59+8];

dp4a.s32.s32 %r150, %r559, %r152, %r595;


dp4a.s32.s32 %r154, %r563, %r156, %r150;

ld.global.nc.u32 %r160, [%rd59+12];
ld.global.nc.u32 %r164, [%rd59+16];

dp4a.s32.s32 %r158, %r567, %r160, %r154;


dp4a.s32.s32 %r162, %r571, %r164, %r158;

ld.global.nc.u32 %r168, [%rd59+20];
ld.global.nc.u32 %r172, [%rd59+24];

dp4a.s32.s32 %r166, %r575, %r168, %r162;


dp4a.s32.s32 %r170, %r579, %r172, %r166;

ld.global.nc.u32 %r176, [%rd59+28];
ld.global.nc.u32 %r180, [%rd59+32];

dp4a.s32.s32 %r174, %r583, %r176, %r170;


dp4a.s32.s32 %r178, %r587, %r180, %r174;


{ cvt.f32.f16 %f103, %rs1;}


mul.ftz.f32 %f162, %f103, %f145;
ld.global.nc.u32 %r182, [%rd59];

{.reg .f16 low,high;
mov.b32 {low,high},%r182;
cvt.f32.f16 %f104, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r182;
cvt.f32.f16 %f105, high;}


cvt.rn.f32.s32 %f163, %r178;
mul.ftz.f32 %f164, %f150, %f105;
fma.rn.ftz.f32 %f165, %f104, %f163, %f164;
fma.rn.ftz.f32 %f423, %f162, %f165, %f423;

dp4a.s32.s32 %r184, %r593, %r152, %r595;


dp4a.s32.s32 %r188, %r597, %r156, %r184;


dp4a.s32.s32 %r192, %r601, %r160, %r188;


dp4a.s32.s32 %r196, %r605, %r164, %r192;


dp4a.s32.s32 %r200, %r609, %r168, %r196;


dp4a.s32.s32 %r204, %r613, %r172, %r200;


dp4a.s32.s32 %r208, %r617, %r176, %r204;


dp4a.s32.s32 %r212, %r621, %r180, %r208;


{ cvt.f32.f16 %f106, %rs2;}


mul.ftz.f32 %f166, %f106, %f154;

{.reg .f16 low,high;
mov.b32 {low,high},%r182;
cvt.f32.f16 %f107, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r182;
cvt.f32.f16 %f108, high;}


cvt.rn.f32.s32 %f167, %r212;
mul.ftz.f32 %f168, %f158, %f108;
fma.rn.ftz.f32 %f169, %f107, %f167, %f168;
fma.rn.ftz.f32 %f422, %f166, %f169, %f422;
cvt.u64.u32 %rd60, %r1138;
add.s64 %rd61, %rd26, %rd60;
add.s64 %rd62, %rd61, %rd2;
mul.lo.s64 %rd63, %rd62, 36;
add.s64 %rd64, %rd9, %rd63;
ld.global.nc.u32 %r220, [%rd64+4];
ld.global.nc.u32 %r224, [%rd64+8];

dp4a.s32.s32 %r218, %r559, %r220, %r595;


dp4a.s32.s32 %r222, %r563, %r224, %r218;

ld.global.nc.u32 %r228, [%rd64+12];
ld.global.nc.u32 %r232, [%rd64+16];

dp4a.s32.s32 %r226, %r567, %r228, %r222;


dp4a.s32.s32 %r230, %r571, %r232, %r226;

ld.global.nc.u32 %r236, [%rd64+20];
ld.global.nc.u32 %r240, [%rd64+24];

dp4a.s32.s32 %r234, %r575, %r236, %r230;


dp4a.s32.s32 %r238, %r579, %r240, %r234;

ld.global.nc.u32 %r244, [%rd64+28];
ld.global.nc.u32 %r248, [%rd64+32];

dp4a.s32.s32 %r242, %r583, %r244, %r238;


dp4a.s32.s32 %r246, %r587, %r248, %r242;


{ cvt.f32.f16 %f109, %rs1;}


mul.ftz.f32 %f170, %f109, %f145;
ld.global.nc.u32 %r250, [%rd64];

{.reg .f16 low,high;
mov.b32 {low,high},%r250;
cvt.f32.f16 %f110, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r250;
cvt.f32.f16 %f111, high;}


cvt.rn.f32.s32 %f171, %r246;
mul.ftz.f32 %f172, %f150, %f111;
fma.rn.ftz.f32 %f173, %f110, %f171, %f172;
fma.rn.ftz.f32 %f421, %f170, %f173, %f421;

dp4a.s32.s32 %r252, %r593, %r220, %r595;


dp4a.s32.s32 %r256, %r597, %r224, %r252;


dp4a.s32.s32 %r260, %r601, %r228, %r256;


dp4a.s32.s32 %r264, %r605, %r232, %r260;


dp4a.s32.s32 %r268, %r609, %r236, %r264;


dp4a.s32.s32 %r272, %r613, %r240, %r268;


dp4a.s32.s32 %r276, %r617, %r244, %r272;


dp4a.s32.s32 %r280, %r621, %r248, %r276;


{ cvt.f32.f16 %f112, %rs2;}


mul.ftz.f32 %f174, %f112, %f154;

{.reg .f16 low,high;
mov.b32 {low,high},%r250;
cvt.f32.f16 %f113, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r250;
cvt.f32.f16 %f114, high;}


cvt.rn.f32.s32 %f175, %r280;
mul.ftz.f32 %f176, %f158, %f114;
fma.rn.ftz.f32 %f177, %f113, %f175, %f176;
fma.rn.ftz.f32 %f420, %f174, %f177, %f420;
cvt.u64.u32 %rd65, %r1137;
add.s64 %rd66, %rd26, %rd65;
add.s64 %rd67, %rd66, %rd2;
mul.lo.s64 %rd68, %rd67, 36;
add.s64 %rd69, %rd9, %rd68;
ld.global.nc.u32 %r288, [%rd69+4];
ld.global.nc.u32 %r292, [%rd69+8];

dp4a.s32.s32 %r286, %r559, %r288, %r595;


dp4a.s32.s32 %r290, %r563, %r292, %r286;

ld.global.nc.u32 %r296, [%rd69+12];
ld.global.nc.u32 %r300, [%rd69+16];

dp4a.s32.s32 %r294, %r567, %r296, %r290;


dp4a.s32.s32 %r298, %r571, %r300, %r294;

ld.global.nc.u32 %r304, [%rd69+20];
ld.global.nc.u32 %r308, [%rd69+24];

dp4a.s32.s32 %r302, %r575, %r304, %r298;


dp4a.s32.s32 %r306, %r579, %r308, %r302;

ld.global.nc.u32 %r312, [%rd69+28];
ld.global.nc.u32 %r316, [%rd69+32];

dp4a.s32.s32 %r310, %r583, %r312, %r306;


dp4a.s32.s32 %r314, %r587, %r316, %r310;


{ cvt.f32.f16 %f115, %rs1;}


mul.ftz.f32 %f178, %f115, %f145;
ld.global.nc.u32 %r318, [%rd69];

{.reg .f16 low,high;
mov.b32 {low,high},%r318;
cvt.f32.f16 %f116, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r318;
cvt.f32.f16 %f117, high;}


cvt.rn.f32.s32 %f179, %r314;
mul.ftz.f32 %f180, %f150, %f117;
fma.rn.ftz.f32 %f181, %f116, %f179, %f180;
fma.rn.ftz.f32 %f419, %f178, %f181, %f419;

dp4a.s32.s32 %r320, %r593, %r288, %r595;


dp4a.s32.s32 %r324, %r597, %r292, %r320;


dp4a.s32.s32 %r328, %r601, %r296, %r324;


dp4a.s32.s32 %r332, %r605, %r300, %r328;


dp4a.s32.s32 %r336, %r609, %r304, %r332;


dp4a.s32.s32 %r340, %r613, %r308, %r336;


dp4a.s32.s32 %r344, %r617, %r312, %r340;


dp4a.s32.s32 %r348, %r621, %r316, %r344;


{ cvt.f32.f16 %f118, %rs2;}


mul.ftz.f32 %f182, %f118, %f154;

{.reg .f16 low,high;
mov.b32 {low,high},%r318;
cvt.f32.f16 %f119, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r318;
cvt.f32.f16 %f120, high;}


cvt.rn.f32.s32 %f183, %r348;
mul.ftz.f32 %f184, %f158, %f120;
fma.rn.ftz.f32 %f185, %f119, %f183, %f184;
fma.rn.ftz.f32 %f418, %f182, %f185, %f418;
cvt.u64.u32 %rd70, %r1136;
add.s64 %rd71, %rd26, %rd70;
add.s64 %rd72, %rd71, %rd2;
mul.lo.s64 %rd73, %rd72, 36;
add.s64 %rd74, %rd9, %rd73;
ld.global.nc.u32 %r356, [%rd74+4];
ld.global.nc.u32 %r360, [%rd74+8];

dp4a.s32.s32 %r354, %r559, %r356, %r595;


dp4a.s32.s32 %r358, %r563, %r360, %r354;

ld.global.nc.u32 %r364, [%rd74+12];
ld.global.nc.u32 %r368, [%rd74+16];

dp4a.s32.s32 %r362, %r567, %r364, %r358;


dp4a.s32.s32 %r366, %r571, %r368, %r362;

ld.global.nc.u32 %r372, [%rd74+20];
ld.global.nc.u32 %r376, [%rd74+24];

dp4a.s32.s32 %r370, %r575, %r372, %r366;


dp4a.s32.s32 %r374, %r579, %r376, %r370;

ld.global.nc.u32 %r380, [%rd74+28];
ld.global.nc.u32 %r384, [%rd74+32];

dp4a.s32.s32 %r378, %r583, %r380, %r374;


dp4a.s32.s32 %r382, %r587, %r384, %r378;


{ cvt.f32.f16 %f121, %rs1;}


mul.ftz.f32 %f186, %f121, %f145;
ld.global.nc.u32 %r386, [%rd74];

{.reg .f16 low,high;
mov.b32 {low,high},%r386;
cvt.f32.f16 %f122, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r386;
cvt.f32.f16 %f123, high;}


cvt.rn.f32.s32 %f187, %r382;
mul.ftz.f32 %f188, %f150, %f123;
fma.rn.ftz.f32 %f189, %f122, %f187, %f188;
fma.rn.ftz.f32 %f417, %f186, %f189, %f417;

dp4a.s32.s32 %r388, %r593, %r356, %r595;


dp4a.s32.s32 %r392, %r597, %r360, %r388;


dp4a.s32.s32 %r396, %r601, %r364, %r392;


dp4a.s32.s32 %r400, %r605, %r368, %r396;


dp4a.s32.s32 %r404, %r609, %r372, %r400;


dp4a.s32.s32 %r408, %r613, %r376, %r404;


dp4a.s32.s32 %r412, %r617, %r380, %r408;


dp4a.s32.s32 %r416, %r621, %r384, %r412;


{ cvt.f32.f16 %f124, %rs2;}


mul.ftz.f32 %f190, %f124, %f154;

{.reg .f16 low,high;
mov.b32 {low,high},%r386;
cvt.f32.f16 %f125, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r386;
cvt.f32.f16 %f126, high;}


cvt.rn.f32.s32 %f191, %r416;
mul.ftz.f32 %f192, %f158, %f126;
fma.rn.ftz.f32 %f193, %f125, %f191, %f192;
fma.rn.ftz.f32 %f416, %f190, %f193, %f416;
cvt.u64.u32 %rd75, %r1135;
add.s64 %rd76, %rd26, %rd75;
add.s64 %rd77, %rd76, %rd2;
mul.lo.s64 %rd78, %rd77, 36;
add.s64 %rd79, %rd9, %rd78;
ld.global.nc.u32 %r424, [%rd79+4];
ld.global.nc.u32 %r428, [%rd79+8];

dp4a.s32.s32 %r422, %r559, %r424, %r595;


dp4a.s32.s32 %r426, %r563, %r428, %r422;

ld.global.nc.u32 %r432, [%rd79+12];
ld.global.nc.u32 %r436, [%rd79+16];

dp4a.s32.s32 %r430, %r567, %r432, %r426;


dp4a.s32.s32 %r434, %r571, %r436, %r430;

ld.global.nc.u32 %r440, [%rd79+20];
ld.global.nc.u32 %r444, [%rd79+24];

dp4a.s32.s32 %r438, %r575, %r440, %r434;


dp4a.s32.s32 %r442, %r579, %r444, %r438;

ld.global.nc.u32 %r448, [%rd79+28];
ld.global.nc.u32 %r452, [%rd79+32];

dp4a.s32.s32 %r446, %r583, %r448, %r442;


dp4a.s32.s32 %r450, %r587, %r452, %r446;


{ cvt.f32.f16 %f127, %rs1;}


mul.ftz.f32 %f194, %f127, %f145;
ld.global.nc.u32 %r454, [%rd79];

{.reg .f16 low,high;
mov.b32 {low,high},%r454;
cvt.f32.f16 %f128, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r454;
cvt.f32.f16 %f129, high;}


cvt.rn.f32.s32 %f195, %r450;
mul.ftz.f32 %f196, %f150, %f129;
fma.rn.ftz.f32 %f197, %f128, %f195, %f196;
fma.rn.ftz.f32 %f415, %f194, %f197, %f415;

dp4a.s32.s32 %r456, %r593, %r424, %r595;


dp4a.s32.s32 %r460, %r597, %r428, %r456;


dp4a.s32.s32 %r464, %r601, %r432, %r460;


dp4a.s32.s32 %r468, %r605, %r436, %r464;


dp4a.s32.s32 %r472, %r609, %r440, %r468;


dp4a.s32.s32 %r476, %r613, %r444, %r472;


dp4a.s32.s32 %r480, %r617, %r448, %r476;


dp4a.s32.s32 %r484, %r621, %r452, %r480;


{ cvt.f32.f16 %f130, %rs2;}


mul.ftz.f32 %f198, %f130, %f154;

{.reg .f16 low,high;
mov.b32 {low,high},%r454;
cvt.f32.f16 %f131, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r454;
cvt.f32.f16 %f132, high;}


cvt.rn.f32.s32 %f199, %r484;
mul.ftz.f32 %f200, %f158, %f132;
fma.rn.ftz.f32 %f201, %f131, %f199, %f200;
fma.rn.ftz.f32 %f414, %f198, %f201, %f414;
cvt.u64.u32 %rd80, %r1134;
add.s64 %rd81, %rd26, %rd80;
add.s64 %rd82, %rd81, %rd2;
mul.lo.s64 %rd83, %rd82, 36;
add.s64 %rd84, %rd9, %rd83;
ld.global.nc.u32 %r492, [%rd84+4];
ld.global.nc.u32 %r496, [%rd84+8];

dp4a.s32.s32 %r490, %r559, %r492, %r595;


dp4a.s32.s32 %r494, %r563, %r496, %r490;

ld.global.nc.u32 %r500, [%rd84+12];
ld.global.nc.u32 %r504, [%rd84+16];

dp4a.s32.s32 %r498, %r567, %r500, %r494;


dp4a.s32.s32 %r502, %r571, %r504, %r498;

ld.global.nc.u32 %r508, [%rd84+20];
ld.global.nc.u32 %r512, [%rd84+24];

dp4a.s32.s32 %r506, %r575, %r508, %r502;


dp4a.s32.s32 %r510, %r579, %r512, %r506;

ld.global.nc.u32 %r516, [%rd84+28];
ld.global.nc.u32 %r520, [%rd84+32];

dp4a.s32.s32 %r514, %r583, %r516, %r510;


dp4a.s32.s32 %r518, %r587, %r520, %r514;


{ cvt.f32.f16 %f133, %rs1;}


mul.ftz.f32 %f202, %f133, %f145;
ld.global.nc.u32 %r522, [%rd84];

{.reg .f16 low,high;
mov.b32 {low,high},%r522;
cvt.f32.f16 %f134, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r522;
cvt.f32.f16 %f135, high;}


cvt.rn.f32.s32 %f203, %r518;
mul.ftz.f32 %f204, %f150, %f135;
fma.rn.ftz.f32 %f205, %f134, %f203, %f204;
fma.rn.ftz.f32 %f413, %f202, %f205, %f413;

dp4a.s32.s32 %r524, %r593, %r492, %r595;


dp4a.s32.s32 %r528, %r597, %r496, %r524;


dp4a.s32.s32 %r532, %r601, %r500, %r528;


dp4a.s32.s32 %r536, %r605, %r504, %r532;


dp4a.s32.s32 %r540, %r609, %r508, %r536;


dp4a.s32.s32 %r544, %r613, %r512, %r540;


dp4a.s32.s32 %r548, %r617, %r516, %r544;


dp4a.s32.s32 %r552, %r621, %r520, %r548;


{ cvt.f32.f16 %f136, %rs2;}


mul.ftz.f32 %f206, %f136, %f154;

{.reg .f16 low,high;
mov.b32 {low,high},%r522;
cvt.f32.f16 %f137, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r522;
cvt.f32.f16 %f138, high;}


cvt.rn.f32.s32 %f207, %r552;
mul.ftz.f32 %f208, %f158, %f138;
fma.rn.ftz.f32 %f209, %f137, %f207, %f208;
fma.rn.ftz.f32 %f412, %f206, %f209, %f412;
cvt.u64.u32 %rd85, %r1133;
add.s64 %rd86, %rd26, %rd85;
add.s64 %rd87, %rd86, %rd2;
mul.lo.s64 %rd88, %rd87, 36;
add.s64 %rd89, %rd9, %rd88;
ld.global.nc.u32 %r560, [%rd89+4];
ld.global.nc.u32 %r564, [%rd89+8];

dp4a.s32.s32 %r558, %r559, %r560, %r595;


dp4a.s32.s32 %r562, %r563, %r564, %r558;

ld.global.nc.u32 %r568, [%rd89+12];
ld.global.nc.u32 %r572, [%rd89+16];

dp4a.s32.s32 %r566, %r567, %r568, %r562;


dp4a.s32.s32 %r570, %r571, %r572, %r566;

ld.global.nc.u32 %r576, [%rd89+20];
ld.global.nc.u32 %r580, [%rd89+24];

dp4a.s32.s32 %r574, %r575, %r576, %r570;


dp4a.s32.s32 %r578, %r579, %r580, %r574;

ld.global.nc.u32 %r584, [%rd89+28];
ld.global.nc.u32 %r588, [%rd89+32];

dp4a.s32.s32 %r582, %r583, %r584, %r578;


dp4a.s32.s32 %r586, %r587, %r588, %r582;


{ cvt.f32.f16 %f139, %rs1;}


mul.ftz.f32 %f210, %f139, %f145;
ld.global.nc.u32 %r590, [%rd89];

{.reg .f16 low,high;
mov.b32 {low,high},%r590;
cvt.f32.f16 %f140, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r590;
cvt.f32.f16 %f141, high;}


cvt.rn.f32.s32 %f211, %r586;
mul.ftz.f32 %f212, %f150, %f141;
fma.rn.ftz.f32 %f213, %f140, %f211, %f212;
fma.rn.ftz.f32 %f411, %f210, %f213, %f411;

dp4a.s32.s32 %r592, %r593, %r560, %r595;


dp4a.s32.s32 %r596, %r597, %r564, %r592;


dp4a.s32.s32 %r600, %r601, %r568, %r596;


dp4a.s32.s32 %r604, %r605, %r572, %r600;


dp4a.s32.s32 %r608, %r609, %r576, %r604;


dp4a.s32.s32 %r612, %r613, %r580, %r608;


dp4a.s32.s32 %r616, %r617, %r584, %r612;


dp4a.s32.s32 %r620, %r621, %r588, %r616;


{ cvt.f32.f16 %f142, %rs2;}


mul.ftz.f32 %f214, %f142, %f154;

{.reg .f16 low,high;
mov.b32 {low,high},%r590;
cvt.f32.f16 %f143, low;}



{.reg .f16 low,high;
mov.b32 {low,high},%r590;
cvt.f32.f16 %f144, high;}


cvt.rn.f32.s32 %f215, %r620;
mul.ftz.f32 %f216, %f158, %f144;
fma.rn.ftz.f32 %f217, %f143, %f215, %f216;
fma.rn.ftz.f32 %f410, %f214, %f217, %f410;
add.s32 %r1138, %r1138, 64;
add.s32 %r1137, %r1137, 64;
add.s32 %r1136, %r1136, 64;
add.s32 %r1135, %r1135, 64;
add.s32 %r1134, %r1134, 64;
add.s32 %r1133, %r1133, 64;
add.s32 %r1132, %r1132, 64;
add.s32 %r1131, %r1131, 64;
add.s64 %rd159, %rd159, 400;
add.s32 %r1139, %r1139, 8;
setp.lt.s32 %p3, %r1139, %r54;
@%p3 bra $L__BB262_2;

$L__BB262_3:
setp.eq.s32 %p4, %r1, 0;
@%p4 bra $L__BB262_5;

shl.b32 %r691, %r1, 11;
mov.u32 %r692, _ZZ13mul_mat_vec_qIL9ggml_type19ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r693, %r692, %r691;
shl.b32 %r694, %r50, 2;
add.s32 %r695, %r693, %r694;
st.shared.f32 [%r695+-2048], %f425;
st.shared.f32 [%r695+-1920], %f424;
st.shared.f32 [%r695+-1792], %f423;
st.shared.f32 [%r695+-1664], %f422;
st.shared.f32 [%r695+-1536], %f421;
st.shared.f32 [%r695+-1408], %f420;
st.shared.f32 [%r695+-1280], %f419;
st.shared.f32 [%r695+-1152], %f418;
st.shared.f32 [%r695+-1024], %f417;
st.shared.f32 [%r695+-896], %f416;
st.shared.f32 [%r695+-768], %f415;
st.shared.f32 [%r695+-640], %f414;
st.shared.f32 [%r695+-512], %f413;
st.shared.f32 [%r695+-384], %f412;
st.shared.f32 [%r695+-256], %f411;
st.shared.f32 [%r695+-128], %f410;

$L__BB262_5:
setp.ne.s32 %p5, %r1, 0;
bar.sync 0;
@%p5 bra $L__BB262_38;

mov.u32 %r697, %ctaid.x;
shl.b32 %r698, %r697, 1;
mov.u32 %r699, 1;
shl.b32 %r701, %r50, 2;
mov.u32 %r702, _ZZ13mul_mat_vec_qIL9ggml_type19ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r703, %r702, %r701;
mov.u32 %r704, 2;
ld.shared.f32 %f218, [%r703];
add.ftz.f32 %f219, %f218, %f425;
mov.b32 %r705, %f219;
mov.u32 %r706, 31;
mov.u32 %r707, 16;
mov.u32 %r708, -1;
shfl.sync.bfly.b32 %r709|%p6, %r705, %r707, %r706, %r708;
mov.b32 %f220, %r709;
add.ftz.f32 %f221, %f219, %f220;
mov.b32 %r710, %f221;
mov.u32 %r711, 8;
shfl.sync.bfly.b32 %r712|%p7, %r710, %r711, %r706, %r708;
mov.b32 %f222, %r712;
add.ftz.f32 %f223, %f221, %f222;
mov.b32 %r713, %f223;
mov.u32 %r714, 4;
shfl.sync.bfly.b32 %r715|%p8, %r713, %r714, %r706, %r708;
mov.b32 %f224, %r715;
add.ftz.f32 %f225, %f223, %f224;
mov.b32 %r716, %f225;
shfl.sync.bfly.b32 %r717|%p9, %r716, %r704, %r706, %r708;
mov.b32 %f226, %r717;
add.ftz.f32 %f227, %f225, %f226;
mov.b32 %r718, %f227;
shfl.sync.bfly.b32 %r719|%p10, %r718, %r699, %r706, %r708;
mov.b32 %f228, %r719;
add.ftz.f32 %f49, %f227, %f228;
setp.ne.s32 %p11, %r50, 0;
setp.ge.u32 %p12, %r698, %r36;
or.pred %p13, %p11, %p12;
@%p13 bra $L__BB262_8;

cvta.to.global.u64 %rd90, %rd17;
mov.u32 %r720, %ctaid.y;
mad.lo.s32 %r723, %r720, %r42, %r698;
mov.u32 %r724, %ctaid.z;
mad.lo.s32 %r725, %r724, %r48, %r723;
mul.wide.u32 %rd91, %r725, 4;
add.s64 %rd92, %rd90, %rd91;
st.global.f32 [%rd92], %f49;

$L__BB262_8:
ld.shared.f32 %f229, [%r703+128];
add.ftz.f32 %f230, %f229, %f424;
mov.b32 %r731, %f230;
shfl.sync.bfly.b32 %r735|%p14, %r731, %r707, %r706, %r708;
mov.b32 %f231, %r735;
add.ftz.f32 %f232, %f230, %f231;
mov.b32 %r736, %f232;
shfl.sync.bfly.b32 %r738|%p15, %r736, %r711, %r706, %r708;
mov.b32 %f233, %r738;
add.ftz.f32 %f234, %f232, %f233;
mov.b32 %r739, %f234;
shfl.sync.bfly.b32 %r741|%p16, %r739, %r714, %r706, %r708;
mov.b32 %f235, %r741;
add.ftz.f32 %f236, %f234, %f235;
mov.b32 %r742, %f236;
shfl.sync.bfly.b32 %r743|%p17, %r742, %r704, %r706, %r708;
mov.b32 %f237, %r743;
add.ftz.f32 %f238, %f236, %f237;
mov.b32 %r744, %f238;
shfl.sync.bfly.b32 %r746|%p18, %r744, %r699, %r706, %r708;
mov.b32 %f239, %r746;
add.ftz.f32 %f50, %f238, %f239;
setp.eq.s32 %p19, %r50, 1;
or.b32 %r749, %r698, 1;
setp.lt.u32 %p20, %r749, %r36;
and.pred %p1, %p19, %p20;
not.pred %p21, %p1;
@%p21 bra $L__BB262_10;

cvta.to.global.u64 %rd93, %rd17;
mov.u32 %r750, %ctaid.y;
mad.lo.s32 %r753, %r750, %r42, %r698;
mov.u32 %r754, %ctaid.z;
mad.lo.s32 %r755, %r754, %r48, %r753;
mul.wide.u32 %rd94, %r755, 4;
add.s64 %rd95, %rd93, %rd94;
st.global.f32 [%rd95+4], %f50;

$L__BB262_10:
mov.u32 %r760, 2;
ld.shared.f32 %f240, [%r703+256];
add.ftz.f32 %f241, %f240, %f423;
mov.b32 %r761, %f241;
mov.u32 %r762, 31;
mov.u32 %r763, 16;
mov.u32 %r764, -1;
shfl.sync.bfly.b32 %r765|%p22, %r761, %r763, %r762, %r764;
mov.b32 %f242, %r765;
add.ftz.f32 %f243, %f241, %f242;
mov.b32 %r766, %f243;
mov.u32 %r767, 8;
shfl.sync.bfly.b32 %r768|%p23, %r766, %r767, %r762, %r764;
mov.b32 %f244, %r768;
add.ftz.f32 %f245, %f243, %f244;
mov.b32 %r769, %f245;
mov.u32 %r770, 4;
shfl.sync.bfly.b32 %r771|%p24, %r769, %r770, %r762, %r764;
mov.b32 %f246, %r771;
add.ftz.f32 %f247, %f245, %f246;
mov.b32 %r772, %f247;
shfl.sync.bfly.b32 %r773|%p25, %r772, %r760, %r762, %r764;
mov.b32 %f248, %r773;
add.ftz.f32 %f249, %f247, %f248;
mov.b32 %r774, %f249;
mov.u32 %r775, 1;
shfl.sync.bfly.b32 %r776|%p26, %r774, %r775, %r762, %r764;
mov.b32 %f250, %r776;
add.ftz.f32 %f51, %f249, %f250;
@%p13 bra $L__BB262_12;

cvt.u64.u32 %rd96, %r36;
mov.u32 %r779, %ctaid.y;
mad.lo.s32 %r782, %r779, %r42, %r698;
mov.u32 %r783, %ctaid.z;
mad.lo.s32 %r784, %r783, %r48, %r782;
cvt.u64.u32 %rd97, %r784;
add.s64 %rd98, %rd96, %rd97;
cvta.to.global.u64 %rd99, %rd17;
shl.b64 %rd100, %rd98, 2;
add.s64 %rd101, %rd99, %rd100;
st.global.f32 [%rd101], %f51;

$L__BB262_12:
ld.shared.f32 %f251, [%r703+384];
add.ftz.f32 %f252, %f251, %f422;
mov.b32 %r790, %f252;
shfl.sync.bfly.b32 %r794|%p30, %r790, %r763, %r762, %r764;
mov.b32 %f253, %r794;
add.ftz.f32 %f254, %f252, %f253;
mov.b32 %r795, %f254;
shfl.sync.bfly.b32 %r797|%p31, %r795, %r767, %r762, %r764;
mov.b32 %f255, %r797;
add.ftz.f32 %f256, %f254, %f255;
mov.b32 %r798, %f256;
shfl.sync.bfly.b32 %r800|%p32, %r798, %r770, %r762, %r764;
mov.b32 %f257, %r800;
add.ftz.f32 %f258, %f256, %f257;
mov.b32 %r801, %f258;
shfl.sync.bfly.b32 %r802|%p33, %r801, %r760, %r762, %r764;
mov.b32 %f259, %r802;
add.ftz.f32 %f260, %f258, %f259;
mov.b32 %r803, %f260;
shfl.sync.bfly.b32 %r805|%p34, %r803, %r775, %r762, %r764;
mov.b32 %f261, %r805;
add.ftz.f32 %f52, %f260, %f261;
@%p21 bra $L__BB262_14;

add.s32 %r806, %r36, 1;
cvt.u64.u32 %rd102, %r806;
mov.u32 %r807, %ctaid.y;
mad.lo.s32 %r810, %r807, %r42, %r698;
mov.u32 %r811, %ctaid.z;
mad.lo.s32 %r812, %r811, %r48, %r810;
cvt.u64.u32 %rd103, %r812;
add.s64 %rd104, %rd102, %rd103;
cvta.to.global.u64 %rd105, %rd17;
shl.b64 %rd106, %rd104, 2;
add.s64 %rd107, %rd105, %rd106;
st.global.f32 [%rd107], %f52;

$L__BB262_14:
shl.b32 %r813, %r36, 1;
mov.u32 %r814, 1;
mov.u32 %r819, 2;
ld.shared.f32 %f262, [%r703+512];
add.ftz.f32 %f263, %f262, %f421;
mov.b32 %r820, %f263;
mov.u32 %r821, 31;
mov.u32 %r822, 16;
mov.u32 %r823, -1;
shfl.sync.bfly.b32 %r824|%p36, %r820, %r822, %r821, %r823;
mov.b32 %f264, %r824;
add.ftz.f32 %f265, %f263, %f264;
mov.b32 %r825, %f265;
mov.u32 %r826, 8;
shfl.sync.bfly.b32 %r827|%p37, %r825, %r826, %r821, %r823;
mov.b32 %f266, %r827;
add.ftz.f32 %f267, %f265, %f266;
mov.b32 %r828, %f267;
mov.u32 %r829, 4;
shfl.sync.bfly.b32 %r830|%p38, %r828, %r829, %r821, %r823;
mov.b32 %f268, %r830;
add.ftz.f32 %f269, %f267, %f268;
mov.b32 %r831, %f269;
shfl.sync.bfly.b32 %r832|%p39, %r831, %r819, %r821, %r823;
mov.b32 %f270, %r832;
add.ftz.f32 %f271, %f269, %f270;
mov.b32 %r833, %f271;
shfl.sync.bfly.b32 %r834|%p40, %r833, %r814, %r821, %r823;
mov.b32 %f272, %r834;
add.ftz.f32 %f53, %f271, %f272;
cvt.u64.u32 %rd108, %r813;
mov.u32 %r835, %ctaid.y;
mad.lo.s32 %r838, %r835, %r42, %r698;
mov.u32 %r839, %ctaid.z;
mad.lo.s32 %r840, %r839, %r48, %r838;
cvt.u64.u32 %rd109, %r840;
add.s64 %rd110, %rd108, %rd109;
cvta.to.global.u64 %rd111, %rd17;
shl.b64 %rd112, %rd110, 2;
add.s64 %rd12, %rd111, %rd112;
@%p13 bra $L__BB262_16;

st.global.f32 [%rd12], %f53;

$L__BB262_16:
ld.shared.f32 %f273, [%r703+640];
add.ftz.f32 %f274, %f273, %f420;
mov.b32 %r846, %f274;
shfl.sync.bfly.b32 %r850|%p44, %r846, %r822, %r821, %r823;
mov.b32 %f275, %r850;
add.ftz.f32 %f276, %f274, %f275;
mov.b32 %r851, %f276;
shfl.sync.bfly.b32 %r853|%p45, %r851, %r826, %r821, %r823;
mov.b32 %f277, %r853;
add.ftz.f32 %f278, %f276, %f277;
mov.b32 %r854, %f278;
shfl.sync.bfly.b32 %r856|%p46, %r854, %r829, %r821, %r823;
mov.b32 %f279, %r856;
add.ftz.f32 %f280, %f278, %f279;
mov.b32 %r857, %f280;
shfl.sync.bfly.b32 %r858|%p47, %r857, %r819, %r821, %r823;
mov.b32 %f281, %r858;
add.ftz.f32 %f282, %f280, %f281;
mov.b32 %r859, %f282;
shfl.sync.bfly.b32 %r861|%p48, %r859, %r814, %r821, %r823;
mov.b32 %f283, %r861;
add.ftz.f32 %f54, %f282, %f283;
@%p21 bra $L__BB262_18;

st.global.f32 [%rd12+4], %f54;

$L__BB262_18:
mul.lo.s32 %r30, %r36, 3;
mov.u32 %r866, 2;
ld.shared.f32 %f284, [%r703+768];
add.ftz.f32 %f285, %f284, %f419;
mov.b32 %r867, %f285;
mov.u32 %r868, 31;
mov.u32 %r869, 16;
mov.u32 %r870, -1;
shfl.sync.bfly.b32 %r871|%p50, %r867, %r869, %r868, %r870;
mov.b32 %f286, %r871;
add.ftz.f32 %f287, %f285, %f286;
mov.b32 %r872, %f287;
mov.u32 %r873, 8;
shfl.sync.bfly.b32 %r874|%p51, %r872, %r873, %r868, %r870;
mov.b32 %f288, %r874;
add.ftz.f32 %f289, %f287, %f288;
mov.b32 %r875, %f289;
mov.u32 %r876, 4;
shfl.sync.bfly.b32 %r877|%p52, %r875, %r876, %r868, %r870;
mov.b32 %f290, %r877;
add.ftz.f32 %f291, %f289, %f290;
mov.b32 %r878, %f291;
shfl.sync.bfly.b32 %r879|%p53, %r878, %r866, %r868, %r870;
mov.b32 %f292, %r879;
add.ftz.f32 %f293, %f291, %f292;
mov.b32 %r880, %f293;
mov.u32 %r881, 1;
shfl.sync.bfly.b32 %r882|%p54, %r880, %r881, %r868, %r870;
mov.b32 %f294, %r882;
add.ftz.f32 %f55, %f293, %f294;
@%p13 bra $L__BB262_20;

cvt.u64.u32 %rd113, %r30;
add.s64 %rd115, %rd113, %rd109;
shl.b64 %rd117, %rd115, 2;
add.s64 %rd118, %rd111, %rd117;
st.global.f32 [%rd118], %f55;

$L__BB262_20:
ld.shared.f32 %f295, [%r703+896];
add.ftz.f32 %f296, %f295, %f418;
mov.b32 %r896, %f296;
shfl.sync.bfly.b32 %r900|%p58, %r896, %r869, %r868, %r870;
mov.b32 %f297, %r900;
add.ftz.f32 %f298, %f296, %f297;
mov.b32 %r901, %f298;
shfl.sync.bfly.b32 %r903|%p59, %r901, %r873, %r868, %r870;
mov.b32 %f299, %r903;
add.ftz.f32 %f300, %f298, %f299;
mov.b32 %r904, %f300;
shfl.sync.bfly.b32 %r906|%p60, %r904, %r876, %r868, %r870;
mov.b32 %f301, %r906;
add.ftz.f32 %f302, %f300, %f301;
mov.b32 %r907, %f302;
shfl.sync.bfly.b32 %r908|%p61, %r907, %r866, %r868, %r870;
mov.b32 %f303, %r908;
add.ftz.f32 %f304, %f302, %f303;
mov.b32 %r909, %f304;
shfl.sync.bfly.b32 %r911|%p62, %r909, %r881, %r868, %r870;
mov.b32 %f305, %r911;
add.ftz.f32 %f56, %f304, %f305;
@%p21 bra $L__BB262_22;

add.s32 %r912, %r30, 1;
cvt.u64.u32 %rd119, %r912;
add.s64 %rd121, %rd119, %rd109;
shl.b64 %rd123, %rd121, 2;
add.s64 %rd124, %rd111, %rd123;
st.global.f32 [%rd124], %f56;

$L__BB262_22:
shl.b32 %r919, %r36, 2;
mov.u32 %r920, 2;
ld.shared.f32 %f306, [%r703+1024];
add.ftz.f32 %f307, %f306, %f417;
mov.b32 %r925, %f307;
mov.u32 %r926, 31;
mov.u32 %r927, 16;
mov.u32 %r928, -1;
shfl.sync.bfly.b32 %r929|%p64, %r925, %r927, %r926, %r928;
mov.b32 %f308, %r929;
add.ftz.f32 %f309, %f307, %f308;
mov.b32 %r930, %f309;
mov.u32 %r931, 8;
shfl.sync.bfly.b32 %r932|%p65, %r930, %r931, %r926, %r928;
mov.b32 %f310, %r932;
add.ftz.f32 %f311, %f309, %f310;
mov.b32 %r933, %f311;
mov.u32 %r934, 4;
shfl.sync.bfly.b32 %r935|%p66, %r933, %r934, %r926, %r928;
mov.b32 %f312, %r935;
add.ftz.f32 %f313, %f311, %f312;
mov.b32 %r936, %f313;
shfl.sync.bfly.b32 %r937|%p67, %r936, %r920, %r926, %r928;
mov.b32 %f314, %r937;
add.ftz.f32 %f315, %f313, %f314;
mov.b32 %r938, %f315;
mov.u32 %r939, 1;
shfl.sync.bfly.b32 %r940|%p68, %r938, %r939, %r926, %r928;
mov.b32 %f316, %r940;
add.ftz.f32 %f57, %f315, %f316;
cvt.u64.u32 %rd125, %r919;
add.s64 %rd127, %rd125, %rd109;
shl.b64 %rd129, %rd127, 2;
add.s64 %rd13, %rd111, %rd129;
@%p13 bra $L__BB262_24;

st.global.f32 [%rd13], %f57;

$L__BB262_24:
ld.shared.f32 %f317, [%r703+1152];
add.ftz.f32 %f318, %f317, %f416;
mov.b32 %r952, %f318;
shfl.sync.bfly.b32 %r956|%p72, %r952, %r927, %r926, %r928;
mov.b32 %f319, %r956;
add.ftz.f32 %f320, %f318, %f319;
mov.b32 %r957, %f320;
shfl.sync.bfly.b32 %r959|%p73, %r957, %r931, %r926, %r928;
mov.b32 %f321, %r959;
add.ftz.f32 %f322, %f320, %f321;
mov.b32 %r960, %f322;
shfl.sync.bfly.b32 %r962|%p74, %r960, %r934, %r926, %r928;
mov.b32 %f323, %r962;
add.ftz.f32 %f324, %f322, %f323;
mov.b32 %r963, %f324;
shfl.sync.bfly.b32 %r964|%p75, %r963, %r920, %r926, %r928;
mov.b32 %f325, %r964;
add.ftz.f32 %f326, %f324, %f325;
mov.b32 %r965, %f326;
shfl.sync.bfly.b32 %r967|%p76, %r965, %r939, %r926, %r928;
mov.b32 %f327, %r967;
add.ftz.f32 %f58, %f326, %f327;
@%p21 bra $L__BB262_26;

st.global.f32 [%rd13+4], %f58;

$L__BB262_26:
mul.lo.s32 %r31, %r36, 5;
mov.u32 %r972, 2;
ld.shared.f32 %f328, [%r703+1280];
add.ftz.f32 %f329, %f328, %f415;
mov.b32 %r973, %f329;
mov.u32 %r974, 31;
mov.u32 %r975, 16;
mov.u32 %r976, -1;
shfl.sync.bfly.b32 %r977|%p78, %r973, %r975, %r974, %r976;
mov.b32 %f330, %r977;
add.ftz.f32 %f331, %f329, %f330;
mov.b32 %r978, %f331;
mov.u32 %r979, 8;
shfl.sync.bfly.b32 %r980|%p79, %r978, %r979, %r974, %r976;
mov.b32 %f332, %r980;
add.ftz.f32 %f333, %f331, %f332;
mov.b32 %r981, %f333;
mov.u32 %r982, 4;
shfl.sync.bfly.b32 %r983|%p80, %r981, %r982, %r974, %r976;
mov.b32 %f334, %r983;
add.ftz.f32 %f335, %f333, %f334;
mov.b32 %r984, %f335;
shfl.sync.bfly.b32 %r985|%p81, %r984, %r972, %r974, %r976;
mov.b32 %f336, %r985;
add.ftz.f32 %f337, %f335, %f336;
mov.b32 %r986, %f337;
mov.u32 %r987, 1;
shfl.sync.bfly.b32 %r988|%p82, %r986, %r987, %r974, %r976;
mov.b32 %f338, %r988;
add.ftz.f32 %f59, %f337, %f338;
@%p13 bra $L__BB262_28;

cvt.u64.u32 %rd130, %r31;
add.s64 %rd132, %rd130, %rd109;
shl.b64 %rd134, %rd132, 2;
add.s64 %rd135, %rd111, %rd134;
st.global.f32 [%rd135], %f59;

$L__BB262_28:
ld.shared.f32 %f339, [%r703+1408];
add.ftz.f32 %f340, %f339, %f414;
mov.b32 %r1002, %f340;
shfl.sync.bfly.b32 %r1006|%p86, %r1002, %r975, %r974, %r976;
mov.b32 %f341, %r1006;
add.ftz.f32 %f342, %f340, %f341;
mov.b32 %r1007, %f342;
shfl.sync.bfly.b32 %r1009|%p87, %r1007, %r979, %r974, %r976;
mov.b32 %f343, %r1009;
add.ftz.f32 %f344, %f342, %f343;
mov.b32 %r1010, %f344;
shfl.sync.bfly.b32 %r1012|%p88, %r1010, %r982, %r974, %r976;
mov.b32 %f345, %r1012;
add.ftz.f32 %f346, %f344, %f345;
mov.b32 %r1013, %f346;
shfl.sync.bfly.b32 %r1014|%p89, %r1013, %r972, %r974, %r976;
mov.b32 %f347, %r1014;
add.ftz.f32 %f348, %f346, %f347;
mov.b32 %r1015, %f348;
shfl.sync.bfly.b32 %r1017|%p90, %r1015, %r987, %r974, %r976;
mov.b32 %f349, %r1017;
add.ftz.f32 %f60, %f348, %f349;
@%p21 bra $L__BB262_30;

add.s32 %r1018, %r31, 1;
cvt.u64.u32 %rd136, %r1018;
add.s64 %rd138, %rd136, %rd109;
shl.b64 %rd140, %rd138, 2;
add.s64 %rd141, %rd111, %rd140;
st.global.f32 [%rd141], %f60;

$L__BB262_30:
mul.lo.s32 %r1025, %r36, 6;
mov.u32 %r1030, 2;
ld.shared.f32 %f350, [%r703+1536];
add.ftz.f32 %f351, %f350, %f413;
mov.b32 %r1031, %f351;
mov.u32 %r1032, 31;
mov.u32 %r1033, 16;
mov.u32 %r1034, -1;
shfl.sync.bfly.b32 %r1035|%p92, %r1031, %r1033, %r1032, %r1034;
mov.b32 %f352, %r1035;
add.ftz.f32 %f353, %f351, %f352;
mov.b32 %r1036, %f353;
mov.u32 %r1037, 8;
shfl.sync.bfly.b32 %r1038|%p93, %r1036, %r1037, %r1032, %r1034;
mov.b32 %f354, %r1038;
add.ftz.f32 %f355, %f353, %f354;
mov.b32 %r1039, %f355;
mov.u32 %r1040, 4;
shfl.sync.bfly.b32 %r1041|%p94, %r1039, %r1040, %r1032, %r1034;
mov.b32 %f356, %r1041;
add.ftz.f32 %f357, %f355, %f356;
mov.b32 %r1042, %f357;
shfl.sync.bfly.b32 %r1043|%p95, %r1042, %r1030, %r1032, %r1034;
mov.b32 %f358, %r1043;
add.ftz.f32 %f359, %f357, %f358;
mov.b32 %r1044, %f359;
mov.u32 %r1045, 1;
shfl.sync.bfly.b32 %r1046|%p96, %r1044, %r1045, %r1032, %r1034;
mov.b32 %f360, %r1046;
add.ftz.f32 %f61, %f359, %f360;
cvt.u64.u32 %rd142, %r1025;
add.s64 %rd144, %rd142, %rd109;
shl.b64 %rd146, %rd144, 2;
add.s64 %rd14, %rd111, %rd146;
@%p13 bra $L__BB262_32;

st.global.f32 [%rd14], %f61;

$L__BB262_32:
ld.shared.f32 %f361, [%r703+1664];
add.ftz.f32 %f362, %f361, %f412;
mov.b32 %r1058, %f362;
shfl.sync.bfly.b32 %r1062|%p100, %r1058, %r1033, %r1032, %r1034;
mov.b32 %f363, %r1062;
add.ftz.f32 %f364, %f362, %f363;
mov.b32 %r1063, %f364;
shfl.sync.bfly.b32 %r1065|%p101, %r1063, %r1037, %r1032, %r1034;
mov.b32 %f365, %r1065;
add.ftz.f32 %f366, %f364, %f365;
mov.b32 %r1066, %f366;
shfl.sync.bfly.b32 %r1068|%p102, %r1066, %r1040, %r1032, %r1034;
mov.b32 %f367, %r1068;
add.ftz.f32 %f368, %f366, %f367;
mov.b32 %r1069, %f368;
shfl.sync.bfly.b32 %r1070|%p103, %r1069, %r1030, %r1032, %r1034;
mov.b32 %f369, %r1070;
add.ftz.f32 %f370, %f368, %f369;
mov.b32 %r1071, %f370;
shfl.sync.bfly.b32 %r1073|%p104, %r1071, %r1045, %r1032, %r1034;
mov.b32 %f371, %r1073;
add.ftz.f32 %f62, %f370, %f371;
@%p21 bra $L__BB262_34;

st.global.f32 [%rd14+4], %f62;

$L__BB262_34:
mul.lo.s32 %r32, %r36, 7;
mov.u32 %r1078, 2;
ld.shared.f32 %f372, [%r703+1792];
add.ftz.f32 %f373, %f372, %f411;
mov.b32 %r1079, %f373;
mov.u32 %r1080, 31;
mov.u32 %r1081, 16;
mov.u32 %r1082, -1;
shfl.sync.bfly.b32 %r1083|%p106, %r1079, %r1081, %r1080, %r1082;
mov.b32 %f374, %r1083;
add.ftz.f32 %f375, %f373, %f374;
mov.b32 %r1084, %f375;
mov.u32 %r1085, 8;
shfl.sync.bfly.b32 %r1086|%p107, %r1084, %r1085, %r1080, %r1082;
mov.b32 %f376, %r1086;
add.ftz.f32 %f377, %f375, %f376;
mov.b32 %r1087, %f377;
mov.u32 %r1088, 4;
shfl.sync.bfly.b32 %r1089|%p108, %r1087, %r1088, %r1080, %r1082;
mov.b32 %f378, %r1089;
add.ftz.f32 %f379, %f377, %f378;
mov.b32 %r1090, %f379;
shfl.sync.bfly.b32 %r1091|%p109, %r1090, %r1078, %r1080, %r1082;
mov.b32 %f380, %r1091;
add.ftz.f32 %f381, %f379, %f380;
mov.b32 %r1092, %f381;
mov.u32 %r1093, 1;
shfl.sync.bfly.b32 %r1094|%p110, %r1092, %r1093, %r1080, %r1082;
mov.b32 %f382, %r1094;
add.ftz.f32 %f63, %f381, %f382;
@%p13 bra $L__BB262_36;

cvt.u64.u32 %rd147, %r32;
add.s64 %rd149, %rd147, %rd109;
shl.b64 %rd151, %rd149, 2;
add.s64 %rd152, %rd111, %rd151;
st.global.f32 [%rd152], %f63;

$L__BB262_36:
ld.shared.f32 %f383, [%r703+1920];
add.ftz.f32 %f384, %f383, %f410;
mov.b32 %r1108, %f384;
shfl.sync.bfly.b32 %r1112|%p114, %r1108, %r1081, %r1080, %r1082;
mov.b32 %f385, %r1112;
add.ftz.f32 %f386, %f384, %f385;
mov.b32 %r1113, %f386;
shfl.sync.bfly.b32 %r1115|%p115, %r1113, %r1085, %r1080, %r1082;
mov.b32 %f387, %r1115;
add.ftz.f32 %f388, %f386, %f387;
mov.b32 %r1116, %f388;
shfl.sync.bfly.b32 %r1118|%p116, %r1116, %r1088, %r1080, %r1082;
mov.b32 %f389, %r1118;
add.ftz.f32 %f390, %f388, %f389;
mov.b32 %r1119, %f390;
shfl.sync.bfly.b32 %r1120|%p117, %r1119, %r1078, %r1080, %r1082;
mov.b32 %f391, %r1120;
add.ftz.f32 %f392, %f390, %f391;
mov.b32 %r1121, %f392;
shfl.sync.bfly.b32 %r1123|%p118, %r1121, %r1093, %r1080, %r1082;
mov.b32 %f393, %r1123;
add.ftz.f32 %f64, %f392, %f393;
@%p21 bra $L__BB262_38;

add.s32 %r1124, %r32, 1;
cvt.u64.u32 %rd153, %r1124;
add.s64 %rd155, %rd153, %rd109;
shl.b64 %rd157, %rd155, 2;
add.s64 %rd158, %rd111, %rd157;
st.global.f32 [%rd158], %f64;

$L__BB262_38:
ret;

}
.entry _Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj(
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2,
.param .align 8 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3[48],
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5,
.param .align 4 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6[12],
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15
)
.maxntid 256, 1, 1
.minnctapersm 1
{
.local .align 8 .b8 __local_depot263[16];
.reg .b64 %SP;
.reg .b64 %SPL;
.reg .pred %p<48>;
.reg .b16 %rs<73>;
.reg .f32 %f<281>;
.reg .b32 %r<572>;
.reg .b64 %rd<146>;


mov.u64 %SPL, __local_depot263;
ld.param.u64 %rd26, [_Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0];
ld.param.u64 %rd27, [_Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1];
ld.param.u64 %rd28, [_Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2];
ld.param.u32 %r35, [_Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5];
ld.param.u32 %r28, [_Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8];
ld.param.u32 %r29, [_Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9];
ld.param.u32 %r31, [_Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11];
ld.param.u32 %r32, [_Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12];
ld.param.u32 %r36, [_Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14];
ld.param.u32 %r34, [_Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15];
ld.param.u32 %r26, [_Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+8];
ld.param.u32 %r25, [_Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+4];
ld.param.u32 %r24, [_Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6];
ld.param.u64 %rd30, [_Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+8];
add.u64 %rd1, %SPL, 0;
add.u64 %rd2, %SPL, 8;
mov.u32 %r37, %ctaid.x;
shl.b32 %r1, %r37, 1;
shr.u32 %r2, %r35, 8;
mov.u32 %r3, %tid.y;
setp.ge.u32 %p1, %r3, %r36;
@%p1 bra $L__BB263_29;

mov.u32 %r4, %ctaid.y;
cvta.to.global.u64 %rd37, %rd28;
mad.lo.s32 %r38, %r3, %r34, %r4;
mul.wide.u32 %rd38, %r38, 4;
add.s64 %rd39, %rd37, %rd38;
ld.global.nc.u32 %r5, [%rd39];
mov.f32 %f274, 0f00000000;
st.local.v2.f32 [%rd1], {%f274, %f274};
st.local.v2.f32 [%rd2], {%f274, %f274};
mov.u32 %r6, %tid.x;
shr.u32 %r571, %r6, 3;
setp.ge.u32 %p2, %r571, %r2;
mov.f32 %f275, %f274;
mov.f32 %f276, %f274;
mov.f32 %f277, %f274;
@%p2 bra $L__BB263_9;

mul.lo.s32 %r39, %r5, %r31;
and.b32 %r40, %r6, 7;
shl.b32 %r41, %r6, 1;
and.b32 %r42, %r41, 14;
cvt.u64.u32 %rd40, %r42;
and.b32 %r43, %r6, 1;
mul.lo.s32 %r8, %r43, 6;
shl.b32 %r44, %r6, 2;
cvt.u64.u32 %rd41, %r44;
and.b64 %rd42, %rd41, 28;
mad.lo.s32 %r45, %r1, %r28, %r571;
add.s32 %r46, %r45, %r39;
mul.wide.s32 %rd6, %r46, 56;
add.s64 %rd7, %rd42, %rd6;
and.b32 %r47, %r6, 6;
cvt.u64.u32 %rd43, %r47;
or.b64 %rd44, %rd43, %rd6;
add.s64 %rd8, %rd44, 48;
add.s64 %rd45, %rd40, %rd6;
add.s64 %rd9, %rd45, 33;
mul.hi.u32 %r48, %r4, %r24;
add.s32 %r49, %r48, %r4;
shr.u32 %r50, %r49, %r25;
mul.lo.s32 %r51, %r50, %r26;
sub.s32 %r52, %r4, %r51;
mul.lo.s32 %r53, %r52, %r32;
mul.wide.u32 %rd46, %r53, 36;
mul.lo.s32 %r54, %r3, %r29;
mul.wide.u32 %rd47, %r54, 36;
add.s64 %rd48, %rd47, %rd46;
mul.wide.s32 %rd49, %r40, 36;
add.s64 %rd50, %rd48, %rd49;
shl.b32 %r55, %r571, 3;
mul.wide.s32 %rd51, %r55, 36;
add.s64 %rd52, %rd50, %rd51;
cvta.to.global.u64 %rd53, %rd27;
add.s64 %rd54, %rd53, %rd52;
add.s64 %rd143, %rd54, 16;
add.s32 %r56, %r1, 1;
add.s32 %r57, %r571, %r39;
mad.lo.s32 %r58, %r28, %r56, %r57;
mul.wide.s32 %rd11, %r58, 56;
add.s64 %rd12, %rd42, %rd11;
add.s64 %rd13, %rd40, %rd11;
or.b64 %rd55, %rd43, %rd11;
add.s64 %rd14, %rd55, 48;
cvta.to.global.u64 %rd145, %rd26;
cvta.to.global.u64 %rd144, %rd30;
mov.f32 %f274, 0f00000000;

$L__BB263_3:
add.s64 %rd56, %rd145, %rd7;
ld.global.nc.u32 %r124, [%rd56];
shr.u32 %r125, %r124, 8;
shr.u32 %r126, %r124, 16;
shr.u32 %r127, %r124, 24;
add.s64 %rd57, %rd145, %rd9;
ld.global.nc.u8 %rs2, [%rd57+-1];
and.b16 %rs3, %rs2, 240;
cvt.u32.u16 %r128, %rs2;
and.b32 %r129, %r124, 255;
and.b32 %r130, %r128, 7;
bfi.b32 %r131, %r130, %r129, 8, 3;
mul.wide.u32 %rd58, %r131, 4;
mov.u64 %rd59, iq1s_grid_gpu;
add.s64 %rd60, %rd59, %rd58;
ld.global.nc.u32 %r132, [%rd60];
and.b32 %r60, %r132, 252645135;
shr.u32 %r133, %r132, 4;
and.b32 %r64, %r133, 252645135;
ld.global.nc.u32 %r10, [%rd143+-12];
ld.global.nc.u32 %r11, [%rd143+-8];
mov.u32 %r118, 0;

dp4a.s32.s32 %r59, %r60, %r10, %r118;


dp4a.s32.s32 %r63, %r64, %r11, %r59;

and.b32 %r134, %r128, 8;
cvt.rn.f32.s32 %f44, %r134;
mul.ftz.f32 %f45, %f44, 0f3D000000;
mov.f32 %f46, 0fBF600000;
sub.ftz.f32 %f47, %f46, %f45;
mov.u32 %r121, 16843009;

dp4a.s32.s32 %r67, %r10, %r121, %r118;


dp4a.s32.s32 %r71, %r11, %r121, %r67;

cvt.rn.f32.s32 %f48, %r71;
fma.rn.ftz.f32 %f49, %f47, %f48, 0f00000000;
shr.u16 %rs4, %rs3, 4;
cvt.u32.u16 %r135, %rs4;
and.b32 %r136, %r125, 255;
and.b32 %r137, %r135, 7;
bfi.b32 %r138, %r137, %r136, 8, 3;
mul.wide.u32 %rd61, %r138, 4;
add.s64 %rd62, %rd59, %rd61;
ld.global.nc.u32 %r139, [%rd62];
and.b32 %r76, %r139, 252645135;
shr.u32 %r140, %r139, 4;
and.b32 %r80, %r140, 252645135;
ld.global.nc.u32 %r12, [%rd143+-4];
ld.global.nc.u32 %r13, [%rd143];

dp4a.s32.s32 %r75, %r76, %r12, %r63;


dp4a.s32.s32 %r79, %r80, %r13, %r75;

and.b32 %r141, %r135, 8;
cvt.rn.f32.s32 %f50, %r141;
mul.ftz.f32 %f51, %f50, 0f3D000000;
sub.ftz.f32 %f52, %f46, %f51;

dp4a.s32.s32 %r83, %r12, %r121, %r118;


dp4a.s32.s32 %r87, %r13, %r121, %r83;

cvt.rn.f32.s32 %f53, %r87;
fma.rn.ftz.f32 %f54, %f52, %f53, %f49;
ld.global.nc.u8 %rs5, [%rd57];
and.b16 %rs6, %rs5, 240;
cvt.u32.u16 %r142, %rs5;
and.b32 %r143, %r126, 255;
and.b32 %r144, %r142, 7;
bfi.b32 %r145, %r144, %r143, 8, 3;
mul.wide.u32 %rd63, %r145, 4;
add.s64 %rd64, %rd59, %rd63;
ld.global.nc.u32 %r146, [%rd64];
and.b32 %r92, %r146, 252645135;
shr.u32 %r147, %r146, 4;
and.b32 %r96, %r147, 252645135;
ld.global.nc.u32 %r14, [%rd143+4];
ld.global.nc.u32 %r15, [%rd143+8];

dp4a.s32.s32 %r91, %r92, %r14, %r118;


dp4a.s32.s32 %r95, %r96, %r15, %r91;

and.b32 %r148, %r142, 8;
cvt.rn.f32.s32 %f55, %r148;
mul.ftz.f32 %f56, %f55, 0f3D000000;
sub.ftz.f32 %f57, %f46, %f56;

dp4a.s32.s32 %r99, %r14, %r121, %r118;


dp4a.s32.s32 %r103, %r15, %r121, %r99;

cvt.rn.f32.s32 %f58, %r103;
fma.rn.ftz.f32 %f59, %f57, %f58, 0f00000000;
shr.u16 %rs7, %rs6, 4;
cvt.u32.u16 %r149, %rs7;
mul.wide.u16 %r150, %rs7, 256;
and.b32 %r151, %r150, 1792;
or.b32 %r152, %r151, %r127;
mul.wide.u32 %rd65, %r152, 4;
add.s64 %rd66, %rd59, %rd65;
ld.global.nc.u32 %r153, [%rd66];
and.b32 %r108, %r153, 252645135;
shr.u32 %r154, %r153, 4;
and.b32 %r112, %r154, 252645135;
ld.global.nc.u32 %r16, [%rd143+12];
ld.global.nc.u32 %r17, [%rd143+16];

dp4a.s32.s32 %r107, %r108, %r16, %r95;


dp4a.s32.s32 %r111, %r112, %r17, %r107;

and.b32 %r155, %r149, 8;
cvt.rn.f32.s32 %f60, %r155;
mul.ftz.f32 %f61, %f60, 0f3D000000;
sub.ftz.f32 %f62, %f46, %f61;

dp4a.s32.s32 %r115, %r16, %r121, %r118;


dp4a.s32.s32 %r119, %r17, %r121, %r115;

cvt.rn.f32.s32 %f63, %r119;
fma.rn.ftz.f32 %f64, %f62, %f63, %f59;
add.s64 %rd67, %rd145, %rd6;
ld.global.nc.u16 %rs8, [%rd67+48];
shr.u16 %rs9, %rs8, 12;
ld.global.nc.u16 %rs10, [%rd67+50];
shr.u16 %rs11, %rs10, 8;
and.b16 %rs12, %rs11, 240;
or.b16 %rs13, %rs12, %rs9;
ld.global.nc.u16 %rs14, [%rd67+52];
shr.u16 %rs15, %rs14, 4;
and.b16 %rs16, %rs15, 3840;
or.b16 %rs17, %rs13, %rs16;
ld.global.nc.u16 %rs18, [%rd67+54];
and.b16 %rs19, %rs18, -4096;
or.b16 %rs1, %rs17, %rs19;

{ cvt.f32.f16 %f42, %rs1;}


ld.global.nc.u32 %r18, [%rd143+-16];

{.reg .f16 low,high;
mov.b32 {low,high},%r18;
cvt.f32.f16 %f43, low;}


mul.ftz.f32 %f65, %f42, %f43;
add.s64 %rd68, %rd145, %rd8;
ld.global.nc.u16 %rs20, [%rd68];
cvt.u32.u16 %r156, %rs20;
shr.u32 %r157, %r156, %r8;
shl.b32 %r158, %r157, 1;
and.b32 %r159, %r158, 14;
or.b32 %r160, %r159, 1;
shr.u32 %r161, %r157, 2;
and.b32 %r162, %r161, 14;
or.b32 %r163, %r162, 1;
cvt.rn.f32.s32 %f66, %r79;
add.ftz.f32 %f67, %f54, %f66;
cvt.rn.f32.s32 %f68, %r160;
cvt.rn.f32.s32 %f69, %r111;
add.ftz.f32 %f70, %f64, %f69;
cvt.rn.f32.s32 %f71, %r163;
mul.ftz.f32 %f72, %f70, %f71;
fma.rn.ftz.f32 %f73, %f67, %f68, %f72;
fma.rn.ftz.f32 %f277, %f65, %f73, %f277;
setp.eq.s64 %p3, %rd30, 0;
@%p3 bra $L__BB263_5;

mov.u64 %rd127, iq1s_grid_gpu;
add.s64 %rd69, %rd144, %rd7;
ld.global.u32 %r229, [%rd69];
shr.u32 %r230, %r229, 8;
shr.u32 %r231, %r229, 16;
shr.u32 %r232, %r229, 24;
add.s64 %rd70, %rd144, %rd9;
ld.global.u8 %rs22, [%rd70+-1];
cvt.u32.u16 %r233, %rs22;
and.b32 %r234, %r229, 255;
and.b32 %r235, %r233, 7;
bfi.b32 %r236, %r235, %r234, 8, 3;
mul.wide.u32 %rd71, %r236, 4;
add.s64 %rd73, %rd127, %rd71;
ld.global.nc.u32 %r237, [%rd73];
and.b32 %r165, %r237, 252645135;
shr.u32 %r238, %r237, 4;
and.b32 %r169, %r238, 252645135;

dp4a.s32.s32 %r164, %r165, %r10, %r118;


dp4a.s32.s32 %r168, %r169, %r11, %r164;

and.b32 %r239, %r233, 8;
cvt.rn.f32.s32 %f76, %r239;
mul.ftz.f32 %f77, %f76, 0f3D000000;
sub.ftz.f32 %f79, %f46, %f77;

dp4a.s32.s32 %r172, %r10, %r121, %r118;


dp4a.s32.s32 %r176, %r11, %r121, %r172;

cvt.rn.f32.s32 %f80, %r176;
fma.rn.ftz.f32 %f81, %f79, %f80, 0f00000000;
shr.u16 %rs23, %rs22, 4;
cvt.u32.u16 %r240, %rs23;
and.b32 %r241, %r230, 255;
and.b32 %r242, %r240, 7;
bfi.b32 %r243, %r242, %r241, 8, 3;
mul.wide.u32 %rd74, %r243, 4;
add.s64 %rd75, %rd127, %rd74;
ld.global.nc.u32 %r244, [%rd75];
and.b32 %r181, %r244, 252645135;
shr.u32 %r245, %r244, 4;
and.b32 %r185, %r245, 252645135;

dp4a.s32.s32 %r180, %r181, %r12, %r168;


dp4a.s32.s32 %r184, %r185, %r13, %r180;

and.b32 %r246, %r240, 8;
cvt.rn.f32.s32 %f82, %r246;
mul.ftz.f32 %f83, %f82, 0f3D000000;
sub.ftz.f32 %f84, %f46, %f83;

dp4a.s32.s32 %r188, %r12, %r121, %r118;


dp4a.s32.s32 %r192, %r13, %r121, %r188;

cvt.rn.f32.s32 %f85, %r192;
fma.rn.ftz.f32 %f86, %f84, %f85, %f81;
ld.global.u8 %rs24, [%rd70];
cvt.u32.u16 %r247, %rs24;
and.b32 %r248, %r231, 255;
and.b32 %r249, %r247, 7;
bfi.b32 %r250, %r249, %r248, 8, 3;
mul.wide.u32 %rd76, %r250, 4;
add.s64 %rd77, %rd127, %rd76;
ld.global.nc.u32 %r251, [%rd77];
and.b32 %r197, %r251, 252645135;
shr.u32 %r252, %r251, 4;
and.b32 %r201, %r252, 252645135;

dp4a.s32.s32 %r196, %r197, %r14, %r118;


dp4a.s32.s32 %r200, %r201, %r15, %r196;

and.b32 %r253, %r247, 8;
cvt.rn.f32.s32 %f87, %r253;
mul.ftz.f32 %f88, %f87, 0f3D000000;
sub.ftz.f32 %f89, %f46, %f88;

dp4a.s32.s32 %r204, %r14, %r121, %r118;


dp4a.s32.s32 %r208, %r15, %r121, %r204;

cvt.rn.f32.s32 %f90, %r208;
fma.rn.ftz.f32 %f91, %f89, %f90, 0f00000000;
shr.u16 %rs25, %rs24, 4;
cvt.u32.u16 %r254, %rs25;
mul.wide.u16 %r255, %rs25, 256;
and.b32 %r256, %r255, 1792;
or.b32 %r257, %r256, %r232;
mul.wide.u32 %rd78, %r257, 4;
add.s64 %rd79, %rd127, %rd78;
ld.global.nc.u32 %r258, [%rd79];
and.b32 %r213, %r258, 252645135;
shr.u32 %r259, %r258, 4;
and.b32 %r217, %r259, 252645135;

dp4a.s32.s32 %r212, %r213, %r16, %r200;


dp4a.s32.s32 %r216, %r217, %r17, %r212;

and.b32 %r260, %r254, 8;
cvt.rn.f32.s32 %f92, %r260;
mul.ftz.f32 %f93, %f92, 0f3D000000;
sub.ftz.f32 %f94, %f46, %f93;

dp4a.s32.s32 %r220, %r16, %r121, %r118;


dp4a.s32.s32 %r224, %r17, %r121, %r220;

cvt.rn.f32.s32 %f95, %r224;
fma.rn.ftz.f32 %f96, %f94, %f95, %f91;
add.s64 %rd80, %rd144, %rd6;
ld.global.u16 %rs26, [%rd80+48];
shr.u16 %rs27, %rs26, 12;
ld.global.u8 %rs28, [%rd80+51];
and.b16 %rs29, %rs28, 240;
or.b16 %rs30, %rs29, %rs27;
ld.global.u16 %rs31, [%rd80+52];
shr.u16 %rs32, %rs31, 4;
and.b16 %rs33, %rs32, 3840;
or.b16 %rs34, %rs30, %rs33;
ld.global.u16 %rs35, [%rd80+54];
and.b16 %rs36, %rs35, -4096;
or.b16 %rs21, %rs34, %rs36;

{ cvt.f32.f16 %f74, %rs21;}



{.reg .f16 low,high;
mov.b32 {low,high},%r18;
cvt.f32.f16 %f75, low;}


mul.ftz.f32 %f97, %f74, %f75;
add.s64 %rd81, %rd144, %rd8;
ld.global.u16 %r261, [%rd81];
shr.u32 %r262, %r261, %r8;
shl.b32 %r263, %r262, 1;
and.b32 %r264, %r263, 14;
or.b32 %r265, %r264, 1;
shr.u32 %r266, %r262, 2;
and.b32 %r267, %r266, 14;
or.b32 %r268, %r267, 1;
cvt.rn.f32.s32 %f98, %r184;
add.ftz.f32 %f99, %f86, %f98;
cvt.rn.f32.s32 %f100, %r265;
cvt.rn.f32.s32 %f101, %r216;
add.ftz.f32 %f102, %f96, %f101;
cvt.rn.f32.s32 %f103, %r268;
mul.ftz.f32 %f104, %f102, %f103;
fma.rn.ftz.f32 %f105, %f99, %f100, %f104;
fma.rn.ftz.f32 %f275, %f97, %f105, %f275;

$L__BB263_5:
mov.u64 %rd121, iq1s_grid_gpu;
add.s64 %rd82, %rd145, %rd12;
ld.global.nc.u32 %r334, [%rd82];
shr.u32 %r335, %r334, 8;
shr.u32 %r336, %r334, 16;
shr.u32 %r337, %r334, 24;
add.s64 %rd83, %rd145, %rd13;
ld.global.nc.u8 %rs38, [%rd83+32];
and.b16 %rs39, %rs38, 240;
cvt.u32.u16 %r338, %rs38;
and.b32 %r339, %r334, 255;
and.b32 %r340, %r338, 7;
bfi.b32 %r341, %r340, %r339, 8, 3;
mul.wide.u32 %rd84, %r341, 4;
add.s64 %rd86, %rd121, %rd84;
ld.global.nc.u32 %r342, [%rd86];
and.b32 %r270, %r342, 252645135;
shr.u32 %r343, %r342, 4;
and.b32 %r274, %r343, 252645135;

dp4a.s32.s32 %r269, %r270, %r10, %r118;


dp4a.s32.s32 %r273, %r274, %r11, %r269;

and.b32 %r344, %r338, 8;
cvt.rn.f32.s32 %f108, %r344;
mul.ftz.f32 %f109, %f108, 0f3D000000;
sub.ftz.f32 %f111, %f46, %f109;

dp4a.s32.s32 %r277, %r10, %r121, %r118;


dp4a.s32.s32 %r281, %r11, %r121, %r277;

cvt.rn.f32.s32 %f112, %r281;
fma.rn.ftz.f32 %f113, %f111, %f112, 0f00000000;
shr.u16 %rs40, %rs39, 4;
cvt.u32.u16 %r345, %rs40;
and.b32 %r346, %r335, 255;
and.b32 %r347, %r345, 7;
bfi.b32 %r348, %r347, %r346, 8, 3;
mul.wide.u32 %rd87, %r348, 4;
add.s64 %rd88, %rd121, %rd87;
ld.global.nc.u32 %r349, [%rd88];
and.b32 %r286, %r349, 252645135;
shr.u32 %r350, %r349, 4;
and.b32 %r290, %r350, 252645135;

dp4a.s32.s32 %r285, %r286, %r12, %r273;


dp4a.s32.s32 %r289, %r290, %r13, %r285;

and.b32 %r351, %r345, 8;
cvt.rn.f32.s32 %f114, %r351;
mul.ftz.f32 %f115, %f114, 0f3D000000;
sub.ftz.f32 %f116, %f46, %f115;

dp4a.s32.s32 %r293, %r12, %r121, %r118;


dp4a.s32.s32 %r297, %r13, %r121, %r293;

cvt.rn.f32.s32 %f117, %r297;
fma.rn.ftz.f32 %f118, %f116, %f117, %f113;
ld.global.nc.u8 %rs41, [%rd83+33];
and.b16 %rs42, %rs41, 240;
cvt.u32.u16 %r352, %rs41;
and.b32 %r353, %r336, 255;
and.b32 %r354, %r352, 7;
bfi.b32 %r355, %r354, %r353, 8, 3;
mul.wide.u32 %rd89, %r355, 4;
add.s64 %rd90, %rd121, %rd89;
ld.global.nc.u32 %r356, [%rd90];
and.b32 %r302, %r356, 252645135;
shr.u32 %r357, %r356, 4;
and.b32 %r306, %r357, 252645135;

dp4a.s32.s32 %r301, %r302, %r14, %r118;


dp4a.s32.s32 %r305, %r306, %r15, %r301;

and.b32 %r358, %r352, 8;
cvt.rn.f32.s32 %f119, %r358;
mul.ftz.f32 %f120, %f119, 0f3D000000;
sub.ftz.f32 %f121, %f46, %f120;

dp4a.s32.s32 %r309, %r14, %r121, %r118;


dp4a.s32.s32 %r313, %r15, %r121, %r309;

cvt.rn.f32.s32 %f122, %r313;
fma.rn.ftz.f32 %f123, %f121, %f122, 0f00000000;
shr.u16 %rs43, %rs42, 4;
cvt.u32.u16 %r359, %rs43;
mul.wide.u16 %r360, %rs43, 256;
and.b32 %r361, %r360, 1792;
or.b32 %r362, %r361, %r337;
mul.wide.u32 %rd91, %r362, 4;
add.s64 %rd92, %rd121, %rd91;
ld.global.nc.u32 %r363, [%rd92];
and.b32 %r318, %r363, 252645135;
shr.u32 %r364, %r363, 4;
and.b32 %r322, %r364, 252645135;

dp4a.s32.s32 %r317, %r318, %r16, %r305;


dp4a.s32.s32 %r321, %r322, %r17, %r317;

and.b32 %r365, %r359, 8;
cvt.rn.f32.s32 %f124, %r365;
mul.ftz.f32 %f125, %f124, 0f3D000000;
sub.ftz.f32 %f126, %f46, %f125;

dp4a.s32.s32 %r325, %r16, %r121, %r118;


dp4a.s32.s32 %r329, %r17, %r121, %r325;

cvt.rn.f32.s32 %f127, %r329;
fma.rn.ftz.f32 %f128, %f126, %f127, %f123;
add.s64 %rd93, %rd145, %rd11;
ld.global.nc.u16 %rs44, [%rd93+48];
shr.u16 %rs45, %rs44, 12;
ld.global.nc.u16 %rs46, [%rd93+50];
shr.u16 %rs47, %rs46, 8;
and.b16 %rs48, %rs47, 240;
or.b16 %rs49, %rs48, %rs45;
ld.global.nc.u16 %rs50, [%rd93+52];
shr.u16 %rs51, %rs50, 4;
and.b16 %rs52, %rs51, 3840;
or.b16 %rs53, %rs49, %rs52;
ld.global.nc.u16 %rs54, [%rd93+54];
and.b16 %rs55, %rs54, -4096;
or.b16 %rs37, %rs53, %rs55;

{ cvt.f32.f16 %f106, %rs37;}



{.reg .f16 low,high;
mov.b32 {low,high},%r18;
cvt.f32.f16 %f107, low;}


mul.ftz.f32 %f129, %f106, %f107;
add.s64 %rd94, %rd145, %rd14;
ld.global.nc.u16 %rs56, [%rd94];
cvt.u32.u16 %r366, %rs56;
shr.u32 %r367, %r366, %r8;
shl.b32 %r368, %r367, 1;
and.b32 %r369, %r368, 14;
or.b32 %r370, %r369, 1;
shr.u32 %r371, %r367, 2;
and.b32 %r372, %r371, 14;
or.b32 %r373, %r372, 1;
cvt.rn.f32.s32 %f130, %r289;
add.ftz.f32 %f131, %f118, %f130;
cvt.rn.f32.s32 %f132, %r370;
cvt.rn.f32.s32 %f133, %r321;
add.ftz.f32 %f134, %f128, %f133;
cvt.rn.f32.s32 %f135, %r373;
mul.ftz.f32 %f136, %f134, %f135;
fma.rn.ftz.f32 %f137, %f131, %f132, %f136;
fma.rn.ftz.f32 %f276, %f129, %f137, %f276;
@%p3 bra $L__BB263_7;

mov.u64 %rd128, iq1s_grid_gpu;
add.s64 %rd95, %rd144, %rd12;
ld.global.u32 %r439, [%rd95];
shr.u32 %r440, %r439, 8;
shr.u32 %r441, %r439, 16;
shr.u32 %r442, %r439, 24;
add.s64 %rd96, %rd144, %rd13;
ld.global.u8 %rs58, [%rd96+32];
cvt.u32.u16 %r443, %rs58;
and.b32 %r444, %r439, 255;
and.b32 %r445, %r443, 7;
bfi.b32 %r446, %r445, %r444, 8, 3;
mul.wide.u32 %rd97, %r446, 4;
add.s64 %rd99, %rd128, %rd97;
ld.global.nc.u32 %r447, [%rd99];
and.b32 %r375, %r447, 252645135;
shr.u32 %r448, %r447, 4;
and.b32 %r379, %r448, 252645135;
mov.u32 %r433, 0;

dp4a.s32.s32 %r374, %r375, %r10, %r433;


dp4a.s32.s32 %r378, %r379, %r11, %r374;

and.b32 %r449, %r443, 8;
cvt.rn.f32.s32 %f140, %r449;
mul.ftz.f32 %f141, %f140, 0f3D000000;
mov.f32 %f142, 0fBF600000;
sub.ftz.f32 %f143, %f142, %f141;
mov.u32 %r436, 16843009;

dp4a.s32.s32 %r382, %r10, %r436, %r433;


dp4a.s32.s32 %r386, %r11, %r436, %r382;

cvt.rn.f32.s32 %f144, %r386;
fma.rn.ftz.f32 %f145, %f143, %f144, 0f00000000;
shr.u16 %rs59, %rs58, 4;
cvt.u32.u16 %r450, %rs59;
and.b32 %r451, %r440, 255;
and.b32 %r452, %r450, 7;
bfi.b32 %r453, %r452, %r451, 8, 3;
mul.wide.u32 %rd100, %r453, 4;
add.s64 %rd101, %rd128, %rd100;
ld.global.nc.u32 %r454, [%rd101];
and.b32 %r391, %r454, 252645135;
shr.u32 %r455, %r454, 4;
and.b32 %r395, %r455, 252645135;

dp4a.s32.s32 %r390, %r391, %r12, %r378;


dp4a.s32.s32 %r394, %r395, %r13, %r390;

and.b32 %r456, %r450, 8;
cvt.rn.f32.s32 %f146, %r456;
mul.ftz.f32 %f147, %f146, 0f3D000000;
sub.ftz.f32 %f148, %f142, %f147;

dp4a.s32.s32 %r398, %r12, %r436, %r433;


dp4a.s32.s32 %r402, %r13, %r436, %r398;

cvt.rn.f32.s32 %f149, %r402;
fma.rn.ftz.f32 %f150, %f148, %f149, %f145;
ld.global.u8 %rs60, [%rd96+33];
cvt.u32.u16 %r457, %rs60;
and.b32 %r458, %r441, 255;
and.b32 %r459, %r457, 7;
bfi.b32 %r460, %r459, %r458, 8, 3;
mul.wide.u32 %rd102, %r460, 4;
add.s64 %rd103, %rd128, %rd102;
ld.global.nc.u32 %r461, [%rd103];
and.b32 %r407, %r461, 252645135;
shr.u32 %r462, %r461, 4;
and.b32 %r411, %r462, 252645135;

dp4a.s32.s32 %r406, %r407, %r14, %r433;


dp4a.s32.s32 %r410, %r411, %r15, %r406;

and.b32 %r463, %r457, 8;
cvt.rn.f32.s32 %f151, %r463;
mul.ftz.f32 %f152, %f151, 0f3D000000;
sub.ftz.f32 %f153, %f142, %f152;

dp4a.s32.s32 %r414, %r14, %r436, %r433;


dp4a.s32.s32 %r418, %r15, %r436, %r414;

cvt.rn.f32.s32 %f154, %r418;
fma.rn.ftz.f32 %f155, %f153, %f154, 0f00000000;
shr.u16 %rs61, %rs60, 4;
cvt.u32.u16 %r464, %rs61;
mul.wide.u16 %r465, %rs61, 256;
and.b32 %r466, %r465, 1792;
or.b32 %r467, %r466, %r442;
mul.wide.u32 %rd104, %r467, 4;
add.s64 %rd105, %rd128, %rd104;
ld.global.nc.u32 %r468, [%rd105];
and.b32 %r423, %r468, 252645135;
shr.u32 %r469, %r468, 4;
and.b32 %r427, %r469, 252645135;

dp4a.s32.s32 %r422, %r423, %r16, %r410;


dp4a.s32.s32 %r426, %r427, %r17, %r422;

and.b32 %r470, %r464, 8;
cvt.rn.f32.s32 %f156, %r470;
mul.ftz.f32 %f157, %f156, 0f3D000000;
sub.ftz.f32 %f158, %f142, %f157;

dp4a.s32.s32 %r430, %r16, %r436, %r433;


dp4a.s32.s32 %r434, %r17, %r436, %r430;

cvt.rn.f32.s32 %f159, %r434;
fma.rn.ftz.f32 %f160, %f158, %f159, %f155;
add.s64 %rd106, %rd144, %rd11;
ld.global.u16 %rs62, [%rd106+48];
shr.u16 %rs63, %rs62, 12;
ld.global.u8 %rs64, [%rd106+51];
and.b16 %rs65, %rs64, 240;
or.b16 %rs66, %rs65, %rs63;
ld.global.u16 %rs67, [%rd106+52];
shr.u16 %rs68, %rs67, 4;
and.b16 %rs69, %rs68, 3840;
or.b16 %rs70, %rs66, %rs69;
ld.global.u16 %rs71, [%rd106+54];
and.b16 %rs72, %rs71, -4096;
or.b16 %rs57, %rs70, %rs72;

{ cvt.f32.f16 %f138, %rs57;}



{.reg .f16 low,high;
mov.b32 {low,high},%r18;
cvt.f32.f16 %f139, low;}


mul.ftz.f32 %f161, %f138, %f139;
add.s64 %rd107, %rd144, %rd14;
ld.global.u16 %r471, [%rd107];
shr.u32 %r472, %r471, %r8;
shl.b32 %r473, %r472, 1;
and.b32 %r474, %r473, 14;
or.b32 %r475, %r474, 1;
shr.u32 %r476, %r472, 2;
and.b32 %r477, %r476, 14;
or.b32 %r478, %r477, 1;
cvt.rn.f32.s32 %f162, %r394;
add.ftz.f32 %f163, %f150, %f162;
cvt.rn.f32.s32 %f164, %r475;
cvt.rn.f32.s32 %f165, %r426;
add.ftz.f32 %f166, %f160, %f165;
cvt.rn.f32.s32 %f167, %r478;
mul.ftz.f32 %f168, %f166, %f167;
fma.rn.ftz.f32 %f169, %f163, %f164, %f168;
fma.rn.ftz.f32 %f274, %f161, %f169, %f274;

$L__BB263_7:
add.s64 %rd145, %rd145, 224;
add.s64 %rd144, %rd144, 224;
add.s64 %rd143, %rd143, 1152;
add.s32 %r571, %r571, 4;
setp.lt.u32 %p5, %r571, %r2;
@%p5 bra $L__BB263_3;

add.u64 %rd135, %SPL, 0;
add.u64 %rd129, %SPL, 8;
st.local.v2.f32 [%rd135], {%f277, %f276};
st.local.v2.f32 [%rd129], {%f275, %f274};

$L__BB263_9:
mov.b32 %r20, %f277;
setp.ne.s64 %p6, %rd30, 0;
@%p6 bra $L__BB263_11;
bra.uni $L__BB263_10;

$L__BB263_11:
add.u64 %rd141, %SPL, 0;
add.u64 %rd133, %SPL, 8;
mov.u32 %r505, 31;
mov.u32 %r506, 16;
mov.u32 %r507, -1;
shfl.sync.bfly.b32 %r508|%p17, %r20, %r506, %r505, %r507;
mov.b32 %f190, %r508;
add.ftz.f32 %f191, %f277, %f190;
mov.b32 %r509, %f191;
mov.u32 %r510, 8;
shfl.sync.bfly.b32 %r511|%p18, %r509, %r510, %r505, %r507;
mov.b32 %f192, %r511;
add.ftz.f32 %f193, %f191, %f192;
mov.b32 %r512, %f193;
mov.u32 %r513, 4;
shfl.sync.bfly.b32 %r514|%p19, %r512, %r513, %r505, %r507;
mov.b32 %f194, %r514;
add.ftz.f32 %f195, %f193, %f194;
mov.b32 %r515, %f195;
mov.u32 %r516, 2;
shfl.sync.bfly.b32 %r517|%p20, %r515, %r516, %r505, %r507;
mov.b32 %f196, %r517;
add.ftz.f32 %f197, %f195, %f196;
mov.b32 %r518, %f197;
mov.u32 %r519, 1;
shfl.sync.bfly.b32 %r520|%p21, %r518, %r519, %r505, %r507;
mov.b32 %f198, %r520;
add.ftz.f32 %f199, %f197, %f198;
st.local.f32 [%rd141], %f199;
mov.b32 %r521, %f275;
shfl.sync.bfly.b32 %r522|%p22, %r521, %r506, %r505, %r507;
mov.b32 %f200, %r522;
add.ftz.f32 %f201, %f275, %f200;
mov.b32 %r523, %f201;
shfl.sync.bfly.b32 %r524|%p23, %r523, %r510, %r505, %r507;
mov.b32 %f202, %r524;
add.ftz.f32 %f203, %f201, %f202;
mov.b32 %r525, %f203;
shfl.sync.bfly.b32 %r526|%p24, %r525, %r513, %r505, %r507;
mov.b32 %f204, %r526;
add.ftz.f32 %f205, %f203, %f204;
mov.b32 %r527, %f205;
shfl.sync.bfly.b32 %r528|%p25, %r527, %r516, %r505, %r507;
mov.b32 %f206, %r528;
add.ftz.f32 %f207, %f205, %f206;
mov.b32 %r529, %f207;
shfl.sync.bfly.b32 %r530|%p26, %r529, %r519, %r505, %r507;
mov.b32 %f208, %r530;
add.ftz.f32 %f209, %f207, %f208;
st.local.f32 [%rd133], %f209;
mov.b32 %r531, %f276;
shfl.sync.bfly.b32 %r532|%p27, %r531, %r506, %r505, %r507;
mov.b32 %f210, %r532;
add.ftz.f32 %f211, %f276, %f210;
mov.b32 %r533, %f211;
shfl.sync.bfly.b32 %r534|%p28, %r533, %r510, %r505, %r507;
mov.b32 %f212, %r534;
add.ftz.f32 %f213, %f211, %f212;
mov.b32 %r535, %f213;
shfl.sync.bfly.b32 %r536|%p29, %r535, %r513, %r505, %r507;
mov.b32 %f214, %r536;
add.ftz.f32 %f215, %f213, %f214;
mov.b32 %r537, %f215;
shfl.sync.bfly.b32 %r538|%p30, %r537, %r516, %r505, %r507;
mov.b32 %f216, %r538;
add.ftz.f32 %f217, %f215, %f216;
mov.b32 %r539, %f217;
shfl.sync.bfly.b32 %r540|%p31, %r539, %r519, %r505, %r507;
mov.b32 %f218, %r540;
add.ftz.f32 %f219, %f217, %f218;
st.local.f32 [%rd141+4], %f219;
mov.b32 %r541, %f274;
shfl.sync.bfly.b32 %r542|%p32, %r541, %r506, %r505, %r507;
mov.b32 %f220, %r542;
add.ftz.f32 %f221, %f274, %f220;
mov.b32 %r543, %f221;
shfl.sync.bfly.b32 %r544|%p33, %r543, %r510, %r505, %r507;
mov.b32 %f222, %r544;
add.ftz.f32 %f223, %f221, %f222;
mov.b32 %r545, %f223;
shfl.sync.bfly.b32 %r546|%p34, %r545, %r513, %r505, %r507;
mov.b32 %f224, %r546;
add.ftz.f32 %f225, %f223, %f224;
mov.b32 %r547, %f225;
shfl.sync.bfly.b32 %r548|%p35, %r547, %r516, %r505, %r507;
mov.b32 %f226, %r548;
add.ftz.f32 %f227, %f225, %f226;
mov.b32 %r549, %f227;
shfl.sync.bfly.b32 %r550|%p36, %r549, %r519, %r505, %r507;
mov.b32 %f228, %r550;
add.ftz.f32 %f229, %f227, %f228;
st.local.f32 [%rd133+4], %f229;
bra.uni $L__BB263_12;

$L__BB263_10:
add.u64 %rd137, %SPL, 0;
mov.u32 %r479, 31;
mov.u32 %r480, 16;
mov.u32 %r481, -1;
shfl.sync.bfly.b32 %r482|%p7, %r20, %r480, %r479, %r481;
mov.b32 %f170, %r482;
add.ftz.f32 %f171, %f277, %f170;
mov.b32 %r483, %f171;
mov.u32 %r484, 8;
shfl.sync.bfly.b32 %r485|%p8, %r483, %r484, %r479, %r481;
mov.b32 %f172, %r485;
add.ftz.f32 %f173, %f171, %f172;
mov.b32 %r486, %f173;
mov.u32 %r487, 4;
shfl.sync.bfly.b32 %r488|%p9, %r486, %r487, %r479, %r481;
mov.b32 %f174, %r488;
add.ftz.f32 %f175, %f173, %f174;
mov.b32 %r489, %f175;
mov.u32 %r490, 2;
shfl.sync.bfly.b32 %r491|%p10, %r489, %r490, %r479, %r481;
mov.b32 %f176, %r491;
add.ftz.f32 %f177, %f175, %f176;
mov.b32 %r492, %f177;
mov.u32 %r493, 1;
shfl.sync.bfly.b32 %r494|%p11, %r492, %r493, %r479, %r481;
mov.b32 %f178, %r494;
add.ftz.f32 %f179, %f177, %f178;
st.local.f32 [%rd137], %f179;
mov.b32 %r495, %f276;
shfl.sync.bfly.b32 %r496|%p12, %r495, %r480, %r479, %r481;
mov.b32 %f180, %r496;
add.ftz.f32 %f181, %f276, %f180;
mov.b32 %r497, %f181;
shfl.sync.bfly.b32 %r498|%p13, %r497, %r484, %r479, %r481;
mov.b32 %f182, %r498;
add.ftz.f32 %f183, %f181, %f182;
mov.b32 %r499, %f183;
shfl.sync.bfly.b32 %r500|%p14, %r499, %r487, %r479, %r481;
mov.b32 %f184, %r500;
add.ftz.f32 %f185, %f183, %f184;
mov.b32 %r501, %f185;
shfl.sync.bfly.b32 %r502|%p15, %r501, %r490, %r479, %r481;
mov.b32 %f186, %r502;
add.ftz.f32 %f187, %f185, %f186;
mov.b32 %r503, %f187;
shfl.sync.bfly.b32 %r504|%p16, %r503, %r493, %r479, %r481;
mov.b32 %f188, %r504;
add.ftz.f32 %f189, %f187, %f188;
st.local.f32 [%rd137+4], %f189;

$L__BB263_12:
ld.param.u32 %r559, [_Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7];
mov.u32 %r558, %ctaid.x;
shl.b32 %r557, %r558, 1;
mov.u32 %r556, %tid.x;
setp.gt.u32 %p37, %r556, 1;
add.s32 %r21, %r557, %r556;
setp.ge.u32 %p38, %r21, %r559;
or.pred %p39, %p37, %p38;
@%p39 bra $L__BB263_29;

add.u64 %rd139, %SPL, 0;
ld.param.u64 %rd122, [_Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3];
ld.param.u32 %r561, [_Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13];
mov.u32 %r560, %tid.x;
cvt.u64.u32 %rd24, %r560;
mul.wide.u32 %rd108, %r560, 4;
add.s64 %rd109, %rd139, %rd108;
ld.local.f32 %f280, [%rd109];
mad.lo.s32 %r22, %r5, %r561, %r21;
setp.eq.s64 %p40, %rd122, 0;
@%p40 bra $L__BB263_15;

ld.param.u64 %rd123, [_Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3];
cvta.to.global.u64 %rd110, %rd123;
mul.wide.u32 %rd111, %r22, 4;
add.s64 %rd112, %rd110, %rd111;
ld.global.f32 %f230, [%rd112];
add.ftz.f32 %f280, %f280, %f230;

$L__BB263_15:
setp.eq.s64 %p41, %rd30, 0;
@%p41 bra $L__BB263_28;

add.u64 %rd131, %SPL, 8;
ld.param.u64 %rd125, [_Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+16];
shl.b64 %rd113, %rd24, 2;
add.s64 %rd114, %rd131, %rd113;
ld.local.f32 %f24, [%rd114];
setp.eq.s64 %p42, %rd125, 0;
@%p42 bra $L__BB263_18;

ld.param.u64 %rd126, [_Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+16];
cvta.to.global.u64 %rd115, %rd126;
mul.wide.u32 %rd116, %r22, 4;
add.s64 %rd117, %rd115, %rd116;
ld.global.f32 %f231, [%rd117];
add.ftz.f32 %f24, %f24, %f231;

$L__BB263_18:
ld.param.u32 %r566, [_Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.gt.s32 %p43, %r566, 2;
@%p43 bra $L__BB263_22;

ld.param.u32 %r569, [_Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.eq.s32 %p46, %r569, 1;
@%p46 bra $L__BB263_26;

ld.param.u32 %r570, [_Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.eq.s32 %p47, %r570, 2;
@%p47 bra $L__BB263_21;
bra.uni $L__BB263_27;

$L__BB263_21:
mul.ftz.f32 %f251, %f24, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f252, %f251;
add.ftz.f32 %f253, %f252, 0f3F800000;
div.approx.ftz.f32 %f254, %f24, %f253;
mul.ftz.f32 %f280, %f280, %f254;
bra.uni $L__BB263_28;

$L__BB263_22:
ld.param.u32 %r567, [_Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.eq.s32 %p44, %r567, 3;
@%p44 bra $L__BB263_25;

ld.param.u32 %r568, [_Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.ne.s32 %p45, %r568, 6;
@%p45 bra $L__BB263_27;

ld.param.f32 %f263, [_Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+44];
min.ftz.f32 %f232, %f24, %f263;
neg.ftz.f32 %f233, %f263;
min.ftz.f32 %f234, %f280, %f263;
max.ftz.f32 %f235, %f234, %f233;
mul.ftz.f32 %f236, %f232, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f237, %f236;
add.ftz.f32 %f238, %f237, 0f3F800000;
div.approx.ftz.f32 %f239, %f232, %f238;
mul.ftz.f32 %f280, %f235, %f239;
bra.uni $L__BB263_28;

$L__BB263_27:
mul.ftz.f32 %f280, %f280, %f24;
bra.uni $L__BB263_28;

$L__BB263_26:
mul.ftz.f32 %f255, %f24, 0f3F000000;
mul.ftz.f32 %f256, %f24, 0f3D372713;
fma.rn.ftz.f32 %f257, %f24, %f256, 0f3F800000;
mul.ftz.f32 %f258, %f24, 0f3F4C422A;
mul.ftz.f32 %f259, %f258, %f257;
tanh.approx.f32 %f260, %f259;
add.ftz.f32 %f261, %f260, 0f3F800000;
mul.ftz.f32 %f262, %f255, %f261;
mul.ftz.f32 %f280, %f280, %f262;
bra.uni $L__BB263_28;

$L__BB263_25:
mov.f32 %f240, 0f40E00000;
min.ftz.f32 %f241, %f24, %f240;
min.ftz.f32 %f242, %f280, %f240;
mov.f32 %f243, 0fC0E00000;
max.ftz.f32 %f244, %f242, %f243;
mul.ftz.f32 %f245, %f241, 0f3FD9DB23;
mul.ftz.f32 %f246, %f245, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f247, %f246;
add.ftz.f32 %f248, %f247, 0f3F800000;
div.approx.ftz.f32 %f249, %f241, %f248;
add.ftz.f32 %f250, %f244, 0f3F800000;
mul.ftz.f32 %f280, %f250, %f249;

$L__BB263_28:
ld.param.u64 %rd124, [_Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4];
ld.param.u32 %r565, [_Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10];
ld.param.u32 %r564, [_Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13];
mov.u32 %r563, %ctaid.y;
mov.u32 %r562, %tid.y;
mad.lo.s32 %r554, %r562, %r565, %r21;
mad.lo.s32 %r555, %r563, %r564, %r554;
cvta.to.global.u64 %rd118, %rd124;
mul.wide.u32 %rd119, %r555, 4;
add.s64 %rd120, %rd118, %rd119;
st.global.f32 [%rd120], %f280;

$L__BB263_29:
ret;

}
.entry _Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj(
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2,
.param .align 8 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3[48],
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5,
.param .align 4 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6[12],
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15
)
.maxntid 256, 1, 1
.minnctapersm 1
{
.local .align 8 .b8 __local_depot264[8];
.reg .b64 %SP;
.reg .b64 %SPL;
.reg .pred %p<17>;
.reg .b16 %rs<41>;
.reg .f32 %f<99>;
.reg .b32 %r<289>;
.reg .b64 %rd<74>;


mov.u64 %SPL, __local_depot264;
ld.param.u64 %rd17, [_Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0];
ld.param.u64 %rd18, [_Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1];
ld.param.u64 %rd19, [_Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2];
ld.param.u64 %rd20, [_Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4];
ld.param.u32 %r9, [_Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5];
ld.param.u32 %r13, [_Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7];
ld.param.u32 %r14, [_Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8];
ld.param.u32 %r15, [_Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9];
ld.param.u32 %r16, [_Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10];
ld.param.u32 %r17, [_Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11];
ld.param.u32 %r18, [_Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12];
ld.param.u32 %r19, [_Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13];
ld.param.u32 %r21, [_Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14];
ld.param.u32 %r20, [_Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15];
ld.param.u32 %r12, [_Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+8];
ld.param.u32 %r11, [_Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+4];
ld.param.u32 %r10, [_Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6];
add.u64 %rd1, %SPL, 0;
mov.u32 %r22, %ctaid.x;
shl.b32 %r1, %r22, 1;
mov.u32 %r2, %tid.y;
setp.ge.u32 %p1, %r2, %r21;
@%p1 bra $L__BB264_7;

mov.f32 %f97, 0f00000000;
st.local.v2.f32 [%rd1], {%f97, %f97};
mov.u32 %r3, %tid.x;
shr.u32 %r288, %r3, 3;
shr.u32 %r23, %r9, 8;
setp.ge.u32 %p2, %r288, %r23;
mov.f32 %f98, %f97;
@%p2 bra $L__BB264_5;

mov.u32 %r24, %ctaid.y;
mad.lo.s32 %r25, %r2, %r20, %r24;
cvta.to.global.u64 %rd22, %rd19;
mul.wide.u32 %rd23, %r25, 4;
add.s64 %rd24, %rd22, %rd23;
ld.global.nc.u32 %r26, [%rd24];
mul.lo.s32 %r27, %r26, %r17;
and.b32 %r28, %r3, 7;
shl.b32 %r29, %r3, 1;
and.b32 %r30, %r29, 14;
and.b32 %r31, %r3, 1;
mul.lo.s32 %r5, %r31, 6;
cvt.u64.u32 %rd25, %r30;
shl.b32 %r32, %r3, 2;
cvt.u64.u32 %rd26, %r32;
and.b64 %rd27, %rd26, 28;
mad.lo.s32 %r33, %r1, %r14, %r288;
add.s32 %r34, %r33, %r27;
mul.wide.s32 %rd3, %r34, 56;
add.s64 %rd4, %rd27, %rd3;
and.b32 %r35, %r3, 6;
cvt.u64.u32 %rd28, %r35;
or.b64 %rd29, %rd28, %rd3;
add.s64 %rd5, %rd29, 48;
add.s64 %rd30, %rd25, %rd3;
add.s64 %rd6, %rd30, 33;
add.s32 %r36, %r1, 1;
add.s32 %r37, %r288, %r27;
mad.lo.s32 %r38, %r14, %r36, %r37;
mul.wide.s32 %rd7, %r38, 56;
add.s64 %rd8, %rd27, %rd7;
mul.hi.u32 %r39, %r24, %r10;
add.s32 %r40, %r39, %r24;
shr.u32 %r41, %r40, %r11;
mul.lo.s32 %r42, %r41, %r12;
sub.s32 %r43, %r24, %r42;
mul.lo.s32 %r44, %r43, %r18;
mul.wide.u32 %rd31, %r44, 36;
mul.lo.s32 %r45, %r2, %r15;
mul.wide.u32 %rd32, %r45, 36;
add.s64 %rd33, %rd32, %rd31;
mul.wide.s32 %rd34, %r28, 36;
add.s64 %rd35, %rd33, %rd34;
shl.b32 %r46, %r288, 3;
mul.wide.s32 %rd36, %r46, 36;
add.s64 %rd37, %rd35, %rd36;
cvta.to.global.u64 %rd38, %rd18;
add.s64 %rd39, %rd38, %rd37;
add.s64 %rd72, %rd39, 16;
add.s64 %rd40, %rd25, %rd7;
add.s64 %rd10, %rd40, 33;
or.b64 %rd41, %rd28, %rd7;
add.s64 %rd11, %rd41, 48;
cvta.to.global.u64 %rd73, %rd17;
mov.f32 %f97, 0f00000000;

$L__BB264_3:
add.s64 %rd42, %rd73, %rd4;
ld.global.nc.u32 %r177, [%rd42];
shr.u32 %r178, %r177, 8;
shr.u32 %r179, %r177, 16;
shr.u32 %r180, %r177, 24;
add.s64 %rd43, %rd73, %rd6;
ld.global.nc.u8 %rs3, [%rd43+-1];
and.b16 %rs4, %rs3, 240;
cvt.u32.u16 %r181, %rs3;
and.b32 %r182, %r177, 255;
and.b32 %r183, %r181, 7;
bfi.b32 %r184, %r183, %r182, 8, 3;
mul.wide.u32 %rd44, %r184, 4;
mov.u64 %rd45, iq1s_grid_gpu;
add.s64 %rd46, %rd45, %rd44;
ld.global.nc.u32 %r185, [%rd46];
and.b32 %r48, %r185, 252645135;
shr.u32 %r186, %r185, 4;
and.b32 %r52, %r186, 252645135;
ld.global.nc.u32 %r49, [%rd72+-12];
ld.global.nc.u32 %r53, [%rd72+-8];
mov.u32 %r171, 0;

dp4a.s32.s32 %r47, %r48, %r49, %r171;


dp4a.s32.s32 %r51, %r52, %r53, %r47;

and.b32 %r187, %r181, 8;
cvt.rn.f32.s32 %f15, %r187;
mul.ftz.f32 %f16, %f15, 0f3D000000;
mov.f32 %f17, 0fBF600000;
sub.ftz.f32 %f18, %f17, %f16;
mov.u32 %r174, 16843009;

dp4a.s32.s32 %r55, %r49, %r174, %r171;


dp4a.s32.s32 %r59, %r53, %r174, %r55;

cvt.rn.f32.s32 %f19, %r59;
fma.rn.ftz.f32 %f20, %f18, %f19, 0f00000000;
shr.u16 %rs5, %rs4, 4;
cvt.u32.u16 %r188, %rs5;
and.b32 %r189, %r178, 255;
and.b32 %r190, %r188, 7;
bfi.b32 %r191, %r190, %r189, 8, 3;
mul.wide.u32 %rd47, %r191, 4;
add.s64 %rd48, %rd45, %rd47;
ld.global.nc.u32 %r192, [%rd48];
and.b32 %r64, %r192, 252645135;
shr.u32 %r193, %r192, 4;
and.b32 %r68, %r193, 252645135;
ld.global.nc.u32 %r65, [%rd72+-4];
ld.global.nc.u32 %r69, [%rd72];

dp4a.s32.s32 %r63, %r64, %r65, %r51;


dp4a.s32.s32 %r67, %r68, %r69, %r63;

and.b32 %r194, %r188, 8;
cvt.rn.f32.s32 %f21, %r194;
mul.ftz.f32 %f22, %f21, 0f3D000000;
sub.ftz.f32 %f23, %f17, %f22;

dp4a.s32.s32 %r71, %r65, %r174, %r171;


dp4a.s32.s32 %r75, %r69, %r174, %r71;

cvt.rn.f32.s32 %f24, %r75;
fma.rn.ftz.f32 %f25, %f23, %f24, %f20;
ld.global.nc.u8 %rs6, [%rd43];
and.b16 %rs7, %rs6, 240;
cvt.u32.u16 %r195, %rs6;
and.b32 %r196, %r179, 255;
and.b32 %r197, %r195, 7;
bfi.b32 %r198, %r197, %r196, 8, 3;
mul.wide.u32 %rd49, %r198, 4;
add.s64 %rd50, %rd45, %rd49;
ld.global.nc.u32 %r199, [%rd50];
and.b32 %r80, %r199, 252645135;
shr.u32 %r200, %r199, 4;
and.b32 %r84, %r200, 252645135;
ld.global.nc.u32 %r81, [%rd72+4];
ld.global.nc.u32 %r85, [%rd72+8];

dp4a.s32.s32 %r79, %r80, %r81, %r171;


dp4a.s32.s32 %r83, %r84, %r85, %r79;

and.b32 %r201, %r195, 8;
cvt.rn.f32.s32 %f26, %r201;
mul.ftz.f32 %f27, %f26, 0f3D000000;
sub.ftz.f32 %f28, %f17, %f27;

dp4a.s32.s32 %r87, %r81, %r174, %r171;


dp4a.s32.s32 %r91, %r85, %r174, %r87;

cvt.rn.f32.s32 %f29, %r91;
fma.rn.ftz.f32 %f30, %f28, %f29, 0f00000000;
shr.u16 %rs8, %rs7, 4;
cvt.u32.u16 %r202, %rs8;
mul.wide.u16 %r203, %rs8, 256;
and.b32 %r204, %r203, 1792;
or.b32 %r205, %r204, %r180;
mul.wide.u32 %rd51, %r205, 4;
add.s64 %rd52, %rd45, %rd51;
ld.global.nc.u32 %r206, [%rd52];
and.b32 %r96, %r206, 252645135;
shr.u32 %r207, %r206, 4;
and.b32 %r100, %r207, 252645135;
ld.global.nc.u32 %r97, [%rd72+12];
ld.global.nc.u32 %r101, [%rd72+16];

dp4a.s32.s32 %r95, %r96, %r97, %r83;


dp4a.s32.s32 %r99, %r100, %r101, %r95;

and.b32 %r208, %r202, 8;
cvt.rn.f32.s32 %f31, %r208;
mul.ftz.f32 %f32, %f31, 0f3D000000;
sub.ftz.f32 %f33, %f17, %f32;

dp4a.s32.s32 %r103, %r97, %r174, %r171;


dp4a.s32.s32 %r107, %r101, %r174, %r103;

cvt.rn.f32.s32 %f34, %r107;
fma.rn.ftz.f32 %f35, %f33, %f34, %f30;
add.s64 %rd53, %rd73, %rd3;
ld.global.nc.u16 %rs9, [%rd53+48];
shr.u16 %rs10, %rs9, 12;
ld.global.nc.u16 %rs11, [%rd53+50];
shr.u16 %rs12, %rs11, 8;
and.b16 %rs13, %rs12, 240;
or.b16 %rs14, %rs13, %rs10;
ld.global.nc.u16 %rs15, [%rd53+52];
shr.u16 %rs16, %rs15, 4;
and.b16 %rs17, %rs16, 3840;
or.b16 %rs18, %rs14, %rs17;
ld.global.nc.u16 %rs19, [%rd53+54];
and.b16 %rs20, %rs19, -4096;
or.b16 %rs1, %rs18, %rs20;

{ cvt.f32.f16 %f11, %rs1;}


ld.global.nc.u32 %r111, [%rd72+-16];

{.reg .f16 low,high;
mov.b32 {low,high},%r111;
cvt.f32.f16 %f12, low;}


mul.ftz.f32 %f36, %f11, %f12;
add.s64 %rd54, %rd73, %rd5;
ld.global.nc.u16 %rs21, [%rd54];
cvt.u32.u16 %r209, %rs21;
shr.u32 %r210, %r209, %r5;
shl.b32 %r211, %r210, 1;
and.b32 %r212, %r211, 14;
or.b32 %r213, %r212, 1;
shr.u32 %r214, %r210, 2;
and.b32 %r215, %r214, 14;
or.b32 %r216, %r215, 1;
cvt.rn.f32.s32 %f37, %r67;
add.ftz.f32 %f38, %f25, %f37;
cvt.rn.f32.s32 %f39, %r213;
cvt.rn.f32.s32 %f40, %r99;
add.ftz.f32 %f41, %f35, %f40;
cvt.rn.f32.s32 %f42, %r216;
mul.ftz.f32 %f43, %f41, %f42;
fma.rn.ftz.f32 %f44, %f38, %f39, %f43;
fma.rn.ftz.f32 %f98, %f36, %f44, %f98;
add.s64 %rd55, %rd73, %rd8;
ld.global.nc.u32 %r217, [%rd55];
shr.u32 %r218, %r217, 8;
shr.u32 %r219, %r217, 16;
shr.u32 %r220, %r217, 24;
add.s64 %rd56, %rd73, %rd10;
ld.global.nc.u8 %rs22, [%rd56+-1];
and.b16 %rs23, %rs22, 240;
cvt.u32.u16 %r221, %rs22;
and.b32 %r222, %r217, 255;
and.b32 %r223, %r221, 7;
bfi.b32 %r224, %r223, %r222, 8, 3;
mul.wide.u32 %rd57, %r224, 4;
add.s64 %rd58, %rd45, %rd57;
ld.global.nc.u32 %r225, [%rd58];
and.b32 %r113, %r225, 252645135;
shr.u32 %r226, %r225, 4;
and.b32 %r117, %r226, 252645135;

dp4a.s32.s32 %r112, %r113, %r49, %r171;


dp4a.s32.s32 %r116, %r117, %r53, %r112;

and.b32 %r227, %r221, 8;
cvt.rn.f32.s32 %f45, %r227;
mul.ftz.f32 %f46, %f45, 0f3D000000;
sub.ftz.f32 %f47, %f17, %f46;

dp4a.s32.s32 %r120, %r49, %r174, %r171;


dp4a.s32.s32 %r124, %r53, %r174, %r120;

cvt.rn.f32.s32 %f48, %r124;
fma.rn.ftz.f32 %f49, %f47, %f48, 0f00000000;
shr.u16 %rs24, %rs23, 4;
cvt.u32.u16 %r228, %rs24;
and.b32 %r229, %r218, 255;
and.b32 %r230, %r228, 7;
bfi.b32 %r231, %r230, %r229, 8, 3;
mul.wide.u32 %rd59, %r231, 4;
add.s64 %rd60, %rd45, %rd59;
ld.global.nc.u32 %r232, [%rd60];
and.b32 %r129, %r232, 252645135;
shr.u32 %r233, %r232, 4;
and.b32 %r133, %r233, 252645135;

dp4a.s32.s32 %r128, %r129, %r65, %r116;


dp4a.s32.s32 %r132, %r133, %r69, %r128;

and.b32 %r234, %r228, 8;
cvt.rn.f32.s32 %f50, %r234;
mul.ftz.f32 %f51, %f50, 0f3D000000;
sub.ftz.f32 %f52, %f17, %f51;

dp4a.s32.s32 %r136, %r65, %r174, %r171;


dp4a.s32.s32 %r140, %r69, %r174, %r136;

cvt.rn.f32.s32 %f53, %r140;
fma.rn.ftz.f32 %f54, %f52, %f53, %f49;
ld.global.nc.u8 %rs25, [%rd56];
and.b16 %rs26, %rs25, 240;
cvt.u32.u16 %r235, %rs25;
and.b32 %r236, %r219, 255;
and.b32 %r237, %r235, 7;
bfi.b32 %r238, %r237, %r236, 8, 3;
mul.wide.u32 %rd61, %r238, 4;
add.s64 %rd62, %rd45, %rd61;
ld.global.nc.u32 %r239, [%rd62];
and.b32 %r145, %r239, 252645135;
shr.u32 %r240, %r239, 4;
and.b32 %r149, %r240, 252645135;

dp4a.s32.s32 %r144, %r145, %r81, %r171;


dp4a.s32.s32 %r148, %r149, %r85, %r144;

and.b32 %r241, %r235, 8;
cvt.rn.f32.s32 %f55, %r241;
mul.ftz.f32 %f56, %f55, 0f3D000000;
sub.ftz.f32 %f57, %f17, %f56;

dp4a.s32.s32 %r152, %r81, %r174, %r171;


dp4a.s32.s32 %r156, %r85, %r174, %r152;

cvt.rn.f32.s32 %f58, %r156;
fma.rn.ftz.f32 %f59, %f57, %f58, 0f00000000;
shr.u16 %rs27, %rs26, 4;
cvt.u32.u16 %r242, %rs27;
mul.wide.u16 %r243, %rs27, 256;
and.b32 %r244, %r243, 1792;
or.b32 %r245, %r244, %r220;
mul.wide.u32 %rd63, %r245, 4;
add.s64 %rd64, %rd45, %rd63;
ld.global.nc.u32 %r246, [%rd64];
and.b32 %r161, %r246, 252645135;
shr.u32 %r247, %r246, 4;
and.b32 %r165, %r247, 252645135;

dp4a.s32.s32 %r160, %r161, %r97, %r148;


dp4a.s32.s32 %r164, %r165, %r101, %r160;

and.b32 %r248, %r242, 8;
cvt.rn.f32.s32 %f60, %r248;
mul.ftz.f32 %f61, %f60, 0f3D000000;
sub.ftz.f32 %f62, %f17, %f61;

dp4a.s32.s32 %r168, %r97, %r174, %r171;


dp4a.s32.s32 %r172, %r101, %r174, %r168;

cvt.rn.f32.s32 %f63, %r172;
fma.rn.ftz.f32 %f64, %f62, %f63, %f59;
add.s64 %rd65, %rd73, %rd7;
ld.global.nc.u16 %rs28, [%rd65+48];
shr.u16 %rs29, %rs28, 12;
ld.global.nc.u16 %rs30, [%rd65+50];
shr.u16 %rs31, %rs30, 8;
and.b16 %rs32, %rs31, 240;
or.b16 %rs33, %rs32, %rs29;
ld.global.nc.u16 %rs34, [%rd65+52];
shr.u16 %rs35, %rs34, 4;
and.b16 %rs36, %rs35, 3840;
or.b16 %rs37, %rs33, %rs36;
ld.global.nc.u16 %rs38, [%rd65+54];
and.b16 %rs39, %rs38, -4096;
or.b16 %rs2, %rs37, %rs39;

{ cvt.f32.f16 %f13, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r111;
cvt.f32.f16 %f14, low;}


mul.ftz.f32 %f65, %f13, %f14;
add.s64 %rd66, %rd73, %rd11;
ld.global.nc.u16 %rs40, [%rd66];
cvt.u32.u16 %r249, %rs40;
shr.u32 %r250, %r249, %r5;
shl.b32 %r251, %r250, 1;
and.b32 %r252, %r251, 14;
or.b32 %r253, %r252, 1;
shr.u32 %r254, %r250, 2;
and.b32 %r255, %r254, 14;
or.b32 %r256, %r255, 1;
cvt.rn.f32.s32 %f66, %r132;
add.ftz.f32 %f67, %f54, %f66;
cvt.rn.f32.s32 %f68, %r253;
cvt.rn.f32.s32 %f69, %r164;
add.ftz.f32 %f70, %f64, %f69;
cvt.rn.f32.s32 %f71, %r256;
mul.ftz.f32 %f72, %f70, %f71;
fma.rn.ftz.f32 %f73, %f67, %f68, %f72;
fma.rn.ftz.f32 %f97, %f65, %f73, %f97;
add.s64 %rd73, %rd73, 224;
add.s64 %rd72, %rd72, 1152;
add.s32 %r288, %r288, 4;
setp.lt.u32 %p3, %r288, %r23;
@%p3 bra $L__BB264_3;

st.local.v2.f32 [%rd1], {%f98, %f97};

$L__BB264_5:
mov.b32 %r258, %f98;
mov.u32 %r259, 31;
mov.u32 %r260, 16;
mov.u32 %r261, -1;
shfl.sync.bfly.b32 %r262|%p4, %r258, %r260, %r259, %r261;
mov.b32 %f74, %r262;
add.ftz.f32 %f75, %f98, %f74;
mov.b32 %r263, %f75;
mov.u32 %r264, 8;
shfl.sync.bfly.b32 %r265|%p5, %r263, %r264, %r259, %r261;
mov.b32 %f76, %r265;
add.ftz.f32 %f77, %f75, %f76;
mov.b32 %r266, %f77;
mov.u32 %r267, 4;
shfl.sync.bfly.b32 %r268|%p6, %r266, %r267, %r259, %r261;
mov.b32 %f78, %r268;
add.ftz.f32 %f79, %f77, %f78;
mov.b32 %r269, %f79;
mov.u32 %r270, 2;
shfl.sync.bfly.b32 %r271|%p7, %r269, %r270, %r259, %r261;
mov.b32 %f80, %r271;
add.ftz.f32 %f81, %f79, %f80;
mov.b32 %r272, %f81;
mov.u32 %r273, 1;
shfl.sync.bfly.b32 %r274|%p8, %r272, %r273, %r259, %r261;
mov.b32 %f82, %r274;
add.ftz.f32 %f83, %f81, %f82;
st.local.f32 [%rd1], %f83;
mov.b32 %r275, %f97;
shfl.sync.bfly.b32 %r276|%p9, %r275, %r260, %r259, %r261;
mov.b32 %f84, %r276;
add.ftz.f32 %f85, %f97, %f84;
mov.b32 %r277, %f85;
shfl.sync.bfly.b32 %r278|%p10, %r277, %r264, %r259, %r261;
mov.b32 %f86, %r278;
add.ftz.f32 %f87, %f85, %f86;
mov.b32 %r279, %f87;
shfl.sync.bfly.b32 %r280|%p11, %r279, %r267, %r259, %r261;
mov.b32 %f88, %r280;
add.ftz.f32 %f89, %f87, %f88;
mov.b32 %r281, %f89;
shfl.sync.bfly.b32 %r282|%p12, %r281, %r270, %r259, %r261;
mov.b32 %f90, %r282;
add.ftz.f32 %f91, %f89, %f90;
mov.b32 %r283, %f91;
shfl.sync.bfly.b32 %r284|%p13, %r283, %r273, %r259, %r261;
mov.b32 %f92, %r284;
add.ftz.f32 %f93, %f91, %f92;
st.local.f32 [%rd1+4], %f93;
add.s32 %r8, %r1, %r3;
setp.ge.u32 %p14, %r8, %r13;
setp.gt.u32 %p15, %r3, 1;
or.pred %p16, %p15, %p14;
@%p16 bra $L__BB264_7;

mul.wide.u32 %rd67, %r3, 4;
add.s64 %rd68, %rd1, %rd67;
ld.local.f32 %f94, [%rd68];
mov.u32 %r285, %ctaid.y;
mad.lo.s32 %r286, %r2, %r16, %r8;
mad.lo.s32 %r287, %r285, %r19, %r286;
cvta.to.global.u64 %rd69, %rd20;
mul.wide.u32 %rd70, %r287, 4;
add.s64 %rd71, %rd69, %rd70;
st.global.f32 [%rd71], %f94;

$L__BB264_7:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<320>;
.reg .b16 %rs<145>;
.reg .f32 %f<1229>;
.reg .b32 %r<1709>;
.reg .b64 %rd<210>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate[1536];

ld.param.u64 %rd38, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd39, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd40, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u32 %r47, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r51, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r52, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r56, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r57, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r58, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r62, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r63, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r64, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r60, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r59, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r54, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r53, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r50, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r49, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r48, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
ld.param.u64 %rd43, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+16];
ld.param.u64 %rd42, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+8];
ld.param.u64 %rd41, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3];
mov.u32 %r1, %tid.y;
shl.b32 %r65, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r65, %r2;
mov.u32 %r66, %ctaid.x;
shl.b32 %r4, %r66, 2;
mov.u32 %r5, %ctaid.y;
setp.eq.s64 %p1, %rd40, 0;
@%p1 bra $L__BB265_2;

cvta.to.global.u64 %rd47, %rd40;
mul.wide.u32 %rd48, %r5, 4;
add.s64 %rd49, %rd47, %rd48;
ld.global.nc.u32 %r1705, [%rd49];
mul.hi.u32 %r67, %r5, %r48;
add.s32 %r68, %r67, %r5;
shr.u32 %r69, %r68, %r49;
mul.lo.s32 %r70, %r69, %r50;
sub.s32 %r1706, %r5, %r70;
bra.uni $L__BB265_3;

$L__BB265_2:
mul.hi.u32 %r71, %r5, %r53;
add.s32 %r72, %r71, %r5;
shr.u32 %r1705, %r72, %r54;
mov.u32 %r1706, %r5;

$L__BB265_3:
mov.u32 %r11, %ctaid.z;
selp.b32 %r12, %r5, %r1705, %p1;
setp.ne.s32 %p3, %r1, 0;
setp.gt.u32 %p4, %r2, 3;
or.pred %p5, %p4, %p3;
add.s32 %r73, %r4, %r2;
setp.ge.u32 %p6, %r73, %r52;
mov.f32 %f27, 0f00000000;
or.pred %p7, %p5, %p6;
mov.f32 %f4, %f27;
mov.f32 %f5, %f27;
@%p7 bra $L__BB265_8;

setp.eq.s64 %p8, %rd41, 0;
mov.f32 %f4, 0f00000000;
mov.f32 %f5, %f4;
@%p8 bra $L__BB265_6;

cvta.to.global.u64 %rd50, %rd41;
mul.lo.s32 %r74, %r11, %r64;
cvt.u64.u32 %rd51, %r74;
mul.lo.s32 %r75, %r12, %r58;
cvt.u64.u32 %rd52, %r75;
cvt.u64.u32 %rd53, %r2;
cvt.s64.s32 %rd54, %r4;
add.s64 %rd55, %rd54, %rd53;
add.s64 %rd56, %rd55, %rd51;
add.s64 %rd57, %rd56, %rd52;
shl.b64 %rd58, %rd57, 2;
add.s64 %rd59, %rd50, %rd58;
ld.global.f32 %f5, [%rd59];

$L__BB265_6:
setp.eq.s64 %p9, %rd43, 0;
setp.eq.s64 %p10, %rd42, 0;
or.pred %p11, %p9, %p10;
@%p11 bra $L__BB265_8;

cvta.to.global.u64 %rd60, %rd43;
mul.lo.s32 %r76, %r11, %r64;
cvt.u64.u32 %rd61, %r76;
mul.lo.s32 %r77, %r12, %r58;
cvt.u64.u32 %rd62, %r77;
cvt.u64.u32 %rd63, %r2;
cvt.s64.s32 %rd64, %r4;
add.s64 %rd65, %rd64, %rd63;
add.s64 %rd66, %rd65, %rd61;
add.s64 %rd67, %rd66, %rd62;
shl.b64 %rd68, %rd67, 2;
add.s64 %rd69, %rd60, %rd68;
ld.global.f32 %f4, [%rd69];

$L__BB265_8:
mul.lo.s32 %r78, %r11, %r63;
cvt.u64.u32 %rd70, %r78;
mul.lo.s32 %r79, %r1706, %r57;
cvt.u64.u32 %rd71, %r79;
add.s64 %rd4, %rd71, %rd70;
mul.hi.u32 %r80, %r11, %r59;
add.s32 %r81, %r80, %r11;
shr.u32 %r13, %r81, %r60;
shr.s32 %r82, %r3, 31;
shr.u32 %r83, %r82, 29;
add.s32 %r84, %r3, %r83;
shr.s32 %r1708, %r84, 3;
shr.u32 %r85, %r47, 8;
setp.ge.s32 %p12, %r1708, %r85;
mov.f32 %f28, %f27;
mov.f32 %f29, %f27;
mov.f32 %f30, %f27;
mov.f32 %f31, %f27;
mov.f32 %f32, %f27;
mov.f32 %f33, %f27;
mov.f32 %f34, %f27;
@%p12 bra $L__BB265_19;

mul.lo.s32 %r86, %r13, %r62;
mul.lo.s32 %r87, %r1705, %r56;
and.b32 %r91, %r84, -8;
sub.s32 %r92, %r3, %r91;
cvt.s64.s32 %rd5, %r92;
shl.b32 %r93, %r92, 1;
cvt.s64.s32 %rd72, %r93;
shr.u32 %r94, %r92, 31;
add.s32 %r95, %r92, %r94;
shr.s32 %r96, %r95, 1;
and.b32 %r97, %r95, -2;
sub.s32 %r98, %r92, %r97;
mul.lo.s32 %r15, %r98, 6;
shl.b32 %r1707, %r1708, 3;
add.s32 %r99, %r1708, %r86;
add.s32 %r17, %r99, %r87;
add.s32 %r100, %r4, 1;
mad.lo.s32 %r101, %r51, %r100, %r17;
mul.wide.s32 %rd73, %r101, 56;
mad.lo.s32 %r102, %r4, %r51, %r1708;
add.s32 %r103, %r102, %r86;
add.s32 %r104, %r103, %r87;
mul.wide.s32 %rd6, %r104, 56;
mul.wide.s32 %rd74, %r92, 4;
add.s64 %rd7, %rd6, %rd74;
mul.wide.s32 %rd75, %r96, 2;
add.s64 %rd76, %rd6, %rd75;
add.s64 %rd8, %rd76, 48;
add.s64 %rd77, %rd73, %rd75;
add.s64 %rd9, %rd77, 48;
add.s64 %rd78, %rd6, %rd72;
add.s64 %rd10, %rd78, 33;
add.s64 %rd79, %rd73, %rd72;
add.s64 %rd11, %rd79, 32;
add.s64 %rd12, %rd73, %rd74;
add.s32 %r105, %r4, 2;
mad.lo.s32 %r106, %r51, %r105, %r17;
mul.wide.s32 %rd80, %r106, 56;
add.s32 %r107, %r4, 3;
mad.lo.s32 %r108, %r51, %r107, %r17;
mul.wide.s32 %rd81, %r108, 56;
add.s64 %rd13, %rd81, %rd74;
add.s64 %rd82, %rd81, %rd72;
add.s64 %rd14, %rd82, 33;
add.s64 %rd83, %rd80, %rd75;
add.s64 %rd15, %rd83, 48;
add.s64 %rd84, %rd81, %rd75;
add.s64 %rd16, %rd84, 48;
add.s64 %rd85, %rd80, %rd72;
add.s64 %rd17, %rd85, 32;
add.s64 %rd18, %rd80, %rd74;
cvta.to.global.u64 %rd209, %rd42;
cvta.to.global.u64 %rd208, %rd38;
mov.f32 %f27, 0f00000000;
cvta.to.global.u64 %rd86, %rd39;
mov.f32 %f28, %f27;
mov.f32 %f29, %f27;
mov.f32 %f30, %f27;
mov.f32 %f31, %f27;
mov.f32 %f32, %f27;
mov.f32 %f33, %f27;
mov.f32 %f34, %f27;

$L__BB265_10:
cvt.u64.u32 %rd87, %r1707;
add.s64 %rd88, %rd4, %rd87;
add.s64 %rd23, %rd208, %rd7;
ld.global.nc.u32 %r174, [%rd23];
shr.u32 %r175, %r174, 8;
shr.u32 %r176, %r174, 16;
shr.u32 %r177, %r174, 24;
add.s64 %rd89, %rd88, %rd5;
mul.lo.s64 %rd90, %rd89, 36;
add.s64 %rd91, %rd208, %rd10;
ld.global.nc.u8 %rs2, [%rd91+-1];
and.b16 %rs3, %rs2, 240;
cvt.u32.u16 %r178, %rs2;
and.b32 %r179, %r174, 255;
and.b32 %r180, %r178, 7;
bfi.b32 %r181, %r180, %r179, 8, 3;
mul.wide.u32 %rd92, %r181, 4;
mov.u64 %rd93, iq1s_grid_gpu;
add.s64 %rd94, %rd93, %rd92;
ld.global.nc.u32 %r182, [%rd94];
and.b32 %r110, %r182, 252645135;
shr.u32 %r183, %r182, 4;
and.b32 %r114, %r183, 252645135;
add.s64 %rd95, %rd86, %rd90;
ld.global.nc.u32 %r20, [%rd95+4];
ld.global.nc.u32 %r21, [%rd95+8];
mov.u32 %r168, 0;

dp4a.s32.s32 %r109, %r110, %r20, %r168;


dp4a.s32.s32 %r113, %r114, %r21, %r109;

and.b32 %r184, %r178, 8;
cvt.rn.f32.s32 %f103, %r184;
mul.ftz.f32 %f104, %f103, 0f3D000000;
mov.f32 %f105, 0fBF600000;
sub.ftz.f32 %f106, %f105, %f104;
mov.u32 %r171, 16843009;

dp4a.s32.s32 %r117, %r20, %r171, %r168;


dp4a.s32.s32 %r121, %r21, %r171, %r117;

cvt.rn.f32.s32 %f107, %r121;
fma.rn.ftz.f32 %f108, %f106, %f107, 0f00000000;
shr.u16 %rs4, %rs3, 4;
cvt.u32.u16 %r185, %rs4;
and.b32 %r186, %r175, 255;
and.b32 %r187, %r185, 7;
bfi.b32 %r188, %r187, %r186, 8, 3;
mul.wide.u32 %rd96, %r188, 4;
add.s64 %rd97, %rd93, %rd96;
ld.global.nc.u32 %r189, [%rd97];
and.b32 %r126, %r189, 252645135;
shr.u32 %r190, %r189, 4;
and.b32 %r130, %r190, 252645135;
ld.global.nc.u32 %r22, [%rd95+12];
ld.global.nc.u32 %r23, [%rd95+16];

dp4a.s32.s32 %r125, %r126, %r22, %r113;


dp4a.s32.s32 %r129, %r130, %r23, %r125;

and.b32 %r191, %r185, 8;
cvt.rn.f32.s32 %f109, %r191;
mul.ftz.f32 %f110, %f109, 0f3D000000;
sub.ftz.f32 %f111, %f105, %f110;

dp4a.s32.s32 %r133, %r22, %r171, %r168;


dp4a.s32.s32 %r137, %r23, %r171, %r133;

cvt.rn.f32.s32 %f112, %r137;
fma.rn.ftz.f32 %f113, %f111, %f112, %f108;
ld.global.nc.u8 %rs5, [%rd91];
and.b16 %rs6, %rs5, 240;
cvt.u32.u16 %r192, %rs5;
and.b32 %r193, %r176, 255;
and.b32 %r194, %r192, 7;
bfi.b32 %r195, %r194, %r193, 8, 3;
mul.wide.u32 %rd98, %r195, 4;
add.s64 %rd99, %rd93, %rd98;
ld.global.nc.u32 %r196, [%rd99];
and.b32 %r142, %r196, 252645135;
shr.u32 %r197, %r196, 4;
and.b32 %r146, %r197, 252645135;
ld.global.nc.u32 %r24, [%rd95+20];
ld.global.nc.u32 %r25, [%rd95+24];

dp4a.s32.s32 %r141, %r142, %r24, %r168;


dp4a.s32.s32 %r145, %r146, %r25, %r141;

and.b32 %r198, %r192, 8;
cvt.rn.f32.s32 %f114, %r198;
mul.ftz.f32 %f115, %f114, 0f3D000000;
sub.ftz.f32 %f116, %f105, %f115;

dp4a.s32.s32 %r149, %r24, %r171, %r168;


dp4a.s32.s32 %r153, %r25, %r171, %r149;

cvt.rn.f32.s32 %f117, %r153;
fma.rn.ftz.f32 %f118, %f116, %f117, 0f00000000;
shr.u16 %rs7, %rs6, 4;
cvt.u32.u16 %r199, %rs7;
mul.wide.u16 %r200, %rs7, 256;
and.b32 %r201, %r200, 1792;
or.b32 %r202, %r201, %r177;
mul.wide.u32 %rd100, %r202, 4;
add.s64 %rd101, %rd93, %rd100;
ld.global.nc.u32 %r203, [%rd101];
and.b32 %r158, %r203, 252645135;
shr.u32 %r204, %r203, 4;
and.b32 %r162, %r204, 252645135;
ld.global.nc.u32 %r26, [%rd95+28];
ld.global.nc.u32 %r27, [%rd95+32];

dp4a.s32.s32 %r157, %r158, %r26, %r145;


dp4a.s32.s32 %r161, %r162, %r27, %r157;

and.b32 %r205, %r199, 8;
cvt.rn.f32.s32 %f119, %r205;
mul.ftz.f32 %f120, %f119, 0f3D000000;
sub.ftz.f32 %f121, %f105, %f120;

dp4a.s32.s32 %r165, %r26, %r171, %r168;


dp4a.s32.s32 %r169, %r27, %r171, %r165;

cvt.rn.f32.s32 %f122, %r169;
fma.rn.ftz.f32 %f123, %f121, %f122, %f118;
add.s64 %rd102, %rd208, %rd6;
ld.global.nc.u16 %rs8, [%rd102+48];
shr.u16 %rs9, %rs8, 12;
ld.global.nc.u16 %rs10, [%rd102+50];
shr.u16 %rs11, %rs10, 8;
and.b16 %rs12, %rs11, 240;
or.b16 %rs13, %rs12, %rs9;
ld.global.nc.u16 %rs14, [%rd102+52];
shr.u16 %rs15, %rs14, 4;
and.b16 %rs16, %rs15, 3840;
or.b16 %rs17, %rs13, %rs16;
ld.global.nc.u16 %rs18, [%rd102+54];
and.b16 %rs19, %rs18, -4096;
or.b16 %rs1, %rs17, %rs19;

{ cvt.f32.f16 %f101, %rs1;}


ld.global.nc.u32 %r28, [%rd95];

{.reg .f16 low,high;
mov.b32 {low,high},%r28;
cvt.f32.f16 %f102, low;}


mul.ftz.f32 %f124, %f101, %f102;
add.s64 %rd26, %rd208, %rd8;
ld.global.nc.u16 %rs20, [%rd26];
cvt.u32.u16 %r206, %rs20;
shr.u32 %r207, %r206, %r15;
shl.b32 %r208, %r207, 1;
and.b32 %r209, %r208, 14;
or.b32 %r210, %r209, 1;
shr.u32 %r211, %r207, 2;
and.b32 %r212, %r211, 14;
or.b32 %r213, %r212, 1;
cvt.rn.f32.s32 %f125, %r129;
add.ftz.f32 %f126, %f113, %f125;
cvt.rn.f32.s32 %f127, %r210;
cvt.rn.f32.s32 %f128, %r161;
add.ftz.f32 %f129, %f123, %f128;
cvt.rn.f32.s32 %f130, %r213;
mul.ftz.f32 %f131, %f129, %f130;
fma.rn.ftz.f32 %f132, %f126, %f127, %f131;
fma.rn.ftz.f32 %f30, %f124, %f132, %f30;
setp.eq.s64 %p13, %rd42, 0;
@%p13 bra $L__BB265_12;

mov.u64 %rd199, iq1s_grid_gpu;
add.s64 %rd103, %rd209, %rd7;
ld.global.u32 %r279, [%rd103];
shr.u32 %r280, %r279, 8;
shr.u32 %r281, %r279, 16;
shr.u32 %r282, %r279, 24;
add.s64 %rd104, %rd209, %rd10;
ld.global.u8 %rs22, [%rd104+-1];
cvt.u32.u16 %r283, %rs22;
and.b32 %r284, %r279, 255;
and.b32 %r285, %r283, 7;
bfi.b32 %r286, %r285, %r284, 8, 3;
mul.wide.u32 %rd105, %r286, 4;
add.s64 %rd107, %rd199, %rd105;
ld.global.nc.u32 %r287, [%rd107];
and.b32 %r215, %r287, 252645135;
shr.u32 %r288, %r287, 4;
and.b32 %r219, %r288, 252645135;

dp4a.s32.s32 %r214, %r215, %r20, %r168;


dp4a.s32.s32 %r218, %r219, %r21, %r214;

and.b32 %r289, %r283, 8;
cvt.rn.f32.s32 %f135, %r289;
mul.ftz.f32 %f136, %f135, 0f3D000000;
sub.ftz.f32 %f138, %f105, %f136;

dp4a.s32.s32 %r222, %r20, %r171, %r168;


dp4a.s32.s32 %r226, %r21, %r171, %r222;

cvt.rn.f32.s32 %f139, %r226;
fma.rn.ftz.f32 %f140, %f138, %f139, 0f00000000;
shr.u16 %rs23, %rs22, 4;
cvt.u32.u16 %r290, %rs23;
and.b32 %r291, %r280, 255;
and.b32 %r292, %r290, 7;
bfi.b32 %r293, %r292, %r291, 8, 3;
mul.wide.u32 %rd108, %r293, 4;
add.s64 %rd109, %rd199, %rd108;
ld.global.nc.u32 %r294, [%rd109];
and.b32 %r231, %r294, 252645135;
shr.u32 %r295, %r294, 4;
and.b32 %r235, %r295, 252645135;

dp4a.s32.s32 %r230, %r231, %r22, %r218;


dp4a.s32.s32 %r234, %r235, %r23, %r230;

and.b32 %r296, %r290, 8;
cvt.rn.f32.s32 %f141, %r296;
mul.ftz.f32 %f142, %f141, 0f3D000000;
sub.ftz.f32 %f143, %f105, %f142;

dp4a.s32.s32 %r238, %r22, %r171, %r168;


dp4a.s32.s32 %r242, %r23, %r171, %r238;

cvt.rn.f32.s32 %f144, %r242;
fma.rn.ftz.f32 %f145, %f143, %f144, %f140;
ld.global.u8 %rs24, [%rd104];
cvt.u32.u16 %r297, %rs24;
and.b32 %r298, %r281, 255;
and.b32 %r299, %r297, 7;
bfi.b32 %r300, %r299, %r298, 8, 3;
mul.wide.u32 %rd110, %r300, 4;
add.s64 %rd111, %rd199, %rd110;
ld.global.nc.u32 %r301, [%rd111];
and.b32 %r247, %r301, 252645135;
shr.u32 %r302, %r301, 4;
and.b32 %r251, %r302, 252645135;

dp4a.s32.s32 %r246, %r247, %r24, %r168;


dp4a.s32.s32 %r250, %r251, %r25, %r246;

and.b32 %r303, %r297, 8;
cvt.rn.f32.s32 %f146, %r303;
mul.ftz.f32 %f147, %f146, 0f3D000000;
sub.ftz.f32 %f148, %f105, %f147;

dp4a.s32.s32 %r254, %r24, %r171, %r168;


dp4a.s32.s32 %r258, %r25, %r171, %r254;

cvt.rn.f32.s32 %f149, %r258;
fma.rn.ftz.f32 %f150, %f148, %f149, 0f00000000;
shr.u16 %rs25, %rs24, 4;
cvt.u32.u16 %r304, %rs25;
mul.wide.u16 %r305, %rs25, 256;
and.b32 %r306, %r305, 1792;
or.b32 %r307, %r306, %r282;
mul.wide.u32 %rd112, %r307, 4;
add.s64 %rd113, %rd199, %rd112;
ld.global.nc.u32 %r308, [%rd113];
and.b32 %r263, %r308, 252645135;
shr.u32 %r309, %r308, 4;
and.b32 %r267, %r309, 252645135;

dp4a.s32.s32 %r262, %r263, %r26, %r250;


dp4a.s32.s32 %r266, %r267, %r27, %r262;

and.b32 %r310, %r304, 8;
cvt.rn.f32.s32 %f151, %r310;
mul.ftz.f32 %f152, %f151, 0f3D000000;
sub.ftz.f32 %f153, %f105, %f152;

dp4a.s32.s32 %r270, %r26, %r171, %r168;


dp4a.s32.s32 %r274, %r27, %r171, %r270;

cvt.rn.f32.s32 %f154, %r274;
fma.rn.ftz.f32 %f155, %f153, %f154, %f150;
add.s64 %rd114, %rd209, %rd6;
ld.global.u16 %rs26, [%rd114+48];
shr.u16 %rs27, %rs26, 12;
ld.global.u8 %rs28, [%rd114+51];
and.b16 %rs29, %rs28, 240;
or.b16 %rs30, %rs29, %rs27;
ld.global.u16 %rs31, [%rd114+52];
shr.u16 %rs32, %rs31, 4;
and.b16 %rs33, %rs32, 3840;
or.b16 %rs34, %rs30, %rs33;
ld.global.u16 %rs35, [%rd114+54];
and.b16 %rs36, %rs35, -4096;
or.b16 %rs21, %rs34, %rs36;

{ cvt.f32.f16 %f133, %rs21;}



{.reg .f16 low,high;
mov.b32 {low,high},%r28;
cvt.f32.f16 %f134, low;}


mul.ftz.f32 %f156, %f133, %f134;
add.s64 %rd115, %rd209, %rd8;
ld.global.u16 %r311, [%rd115];
shr.u32 %r312, %r311, %r15;
shl.b32 %r313, %r312, 1;
and.b32 %r314, %r313, 14;
or.b32 %r315, %r314, 1;
shr.u32 %r316, %r312, 2;
and.b32 %r317, %r316, 14;
or.b32 %r318, %r317, 1;
cvt.rn.f32.s32 %f157, %r234;
add.ftz.f32 %f158, %f145, %f157;
cvt.rn.f32.s32 %f159, %r315;
cvt.rn.f32.s32 %f160, %r266;
add.ftz.f32 %f161, %f155, %f160;
cvt.rn.f32.s32 %f162, %r318;
mul.ftz.f32 %f163, %f161, %f162;
fma.rn.ftz.f32 %f164, %f158, %f159, %f163;
fma.rn.ftz.f32 %f34, %f156, %f164, %f34;

$L__BB265_12:
add.s64 %rd205, %rd208, %rd6;
add.s64 %rd204, %rd205, 48;
add.s64 %rd203, %rd208, %rd10;
add.s64 %rd202, %rd203, -1;
mov.u64 %rd196, iq1s_grid_gpu;
mul.wide.s32 %rd116, %r51, 56;
add.s64 %rd27, %rd23, %rd116;
ld.global.nc.u32 %r384, [%rd27];
shr.u32 %r385, %r384, 8;
shr.u32 %r386, %r384, 16;
shr.u32 %r387, %r384, 24;
add.s64 %rd28, %rd202, %rd116;
ld.global.nc.u8 %rs38, [%rd28];
and.b16 %rs39, %rs38, 240;
cvt.u32.u16 %r388, %rs38;
and.b32 %r389, %r384, 255;
and.b32 %r390, %r388, 7;
bfi.b32 %r391, %r390, %r389, 8, 3;
mul.wide.u32 %rd117, %r391, 4;
add.s64 %rd119, %rd196, %rd117;
ld.global.nc.u32 %r392, [%rd119];
and.b32 %r320, %r392, 252645135;
shr.u32 %r393, %r392, 4;
and.b32 %r324, %r393, 252645135;

dp4a.s32.s32 %r319, %r320, %r20, %r168;


dp4a.s32.s32 %r323, %r324, %r21, %r319;

and.b32 %r394, %r388, 8;
cvt.rn.f32.s32 %f167, %r394;
mul.ftz.f32 %f168, %f167, 0f3D000000;
sub.ftz.f32 %f170, %f105, %f168;

dp4a.s32.s32 %r327, %r20, %r171, %r168;


dp4a.s32.s32 %r331, %r21, %r171, %r327;

cvt.rn.f32.s32 %f171, %r331;
fma.rn.ftz.f32 %f172, %f170, %f171, 0f00000000;
shr.u16 %rs40, %rs39, 4;
cvt.u32.u16 %r395, %rs40;
and.b32 %r396, %r385, 255;
and.b32 %r397, %r395, 7;
bfi.b32 %r398, %r397, %r396, 8, 3;
mul.wide.u32 %rd120, %r398, 4;
add.s64 %rd121, %rd196, %rd120;
ld.global.nc.u32 %r399, [%rd121];
and.b32 %r336, %r399, 252645135;
shr.u32 %r400, %r399, 4;
and.b32 %r340, %r400, 252645135;

dp4a.s32.s32 %r335, %r336, %r22, %r323;


dp4a.s32.s32 %r339, %r340, %r23, %r335;

and.b32 %r401, %r395, 8;
cvt.rn.f32.s32 %f173, %r401;
mul.ftz.f32 %f174, %f173, 0f3D000000;
sub.ftz.f32 %f175, %f105, %f174;

dp4a.s32.s32 %r343, %r22, %r171, %r168;


dp4a.s32.s32 %r347, %r23, %r171, %r343;

cvt.rn.f32.s32 %f176, %r347;
fma.rn.ftz.f32 %f177, %f175, %f176, %f172;
ld.global.nc.u8 %rs41, [%rd28+1];
and.b16 %rs42, %rs41, 240;
cvt.u32.u16 %r402, %rs41;
and.b32 %r403, %r386, 255;
and.b32 %r404, %r402, 7;
bfi.b32 %r405, %r404, %r403, 8, 3;
mul.wide.u32 %rd122, %r405, 4;
add.s64 %rd123, %rd196, %rd122;
ld.global.nc.u32 %r406, [%rd123];
and.b32 %r352, %r406, 252645135;
shr.u32 %r407, %r406, 4;
and.b32 %r356, %r407, 252645135;

dp4a.s32.s32 %r351, %r352, %r24, %r168;


dp4a.s32.s32 %r355, %r356, %r25, %r351;

and.b32 %r408, %r402, 8;
cvt.rn.f32.s32 %f178, %r408;
mul.ftz.f32 %f179, %f178, 0f3D000000;
sub.ftz.f32 %f180, %f105, %f179;

dp4a.s32.s32 %r359, %r24, %r171, %r168;


dp4a.s32.s32 %r363, %r25, %r171, %r359;

cvt.rn.f32.s32 %f181, %r363;
fma.rn.ftz.f32 %f182, %f180, %f181, 0f00000000;
shr.u16 %rs43, %rs42, 4;
cvt.u32.u16 %r409, %rs43;
mul.wide.u16 %r410, %rs43, 256;
and.b32 %r411, %r410, 1792;
or.b32 %r412, %r411, %r387;
mul.wide.u32 %rd124, %r412, 4;
add.s64 %rd125, %rd196, %rd124;
ld.global.nc.u32 %r413, [%rd125];
and.b32 %r368, %r413, 252645135;
shr.u32 %r414, %r413, 4;
and.b32 %r372, %r414, 252645135;

dp4a.s32.s32 %r367, %r368, %r26, %r355;


dp4a.s32.s32 %r371, %r372, %r27, %r367;

and.b32 %r415, %r409, 8;
cvt.rn.f32.s32 %f183, %r415;
mul.ftz.f32 %f184, %f183, 0f3D000000;
sub.ftz.f32 %f185, %f105, %f184;

dp4a.s32.s32 %r375, %r26, %r171, %r168;


dp4a.s32.s32 %r379, %r27, %r171, %r375;

cvt.rn.f32.s32 %f186, %r379;
fma.rn.ftz.f32 %f187, %f185, %f186, %f182;
add.s64 %rd29, %rd204, %rd116;
ld.global.nc.u16 %rs44, [%rd29];
shr.u16 %rs45, %rs44, 12;
ld.global.nc.u16 %rs46, [%rd29+2];
shr.u16 %rs47, %rs46, 8;
and.b16 %rs48, %rs47, 240;
or.b16 %rs49, %rs48, %rs45;
ld.global.nc.u16 %rs50, [%rd29+4];
shr.u16 %rs51, %rs50, 4;
and.b16 %rs52, %rs51, 3840;
or.b16 %rs53, %rs49, %rs52;
ld.global.nc.u16 %rs54, [%rd29+6];
and.b16 %rs55, %rs54, -4096;
or.b16 %rs37, %rs53, %rs55;

{ cvt.f32.f16 %f165, %rs37;}



{.reg .f16 low,high;
mov.b32 {low,high},%r28;
cvt.f32.f16 %f166, low;}


mul.ftz.f32 %f188, %f165, %f166;
add.s64 %rd30, %rd26, %rd116;
ld.global.nc.u16 %rs56, [%rd30];
cvt.u32.u16 %r416, %rs56;
shr.u32 %r417, %r416, %r15;
shl.b32 %r418, %r417, 1;
and.b32 %r419, %r418, 14;
or.b32 %r420, %r419, 1;
shr.u32 %r421, %r417, 2;
and.b32 %r422, %r421, 14;
or.b32 %r423, %r422, 1;
cvt.rn.f32.s32 %f189, %r339;
add.ftz.f32 %f190, %f177, %f189;
cvt.rn.f32.s32 %f191, %r420;
cvt.rn.f32.s32 %f192, %r371;
add.ftz.f32 %f193, %f187, %f192;
cvt.rn.f32.s32 %f194, %r423;
mul.ftz.f32 %f195, %f193, %f194;
fma.rn.ftz.f32 %f196, %f190, %f191, %f195;
fma.rn.ftz.f32 %f29, %f188, %f196, %f29;
@%p13 bra $L__BB265_14;

mov.u64 %rd198, iq1s_grid_gpu;
add.s64 %rd126, %rd209, %rd12;
ld.global.u32 %r489, [%rd126];
shr.u32 %r490, %r489, 8;
shr.u32 %r491, %r489, 16;
shr.u32 %r492, %r489, 24;
add.s64 %rd127, %rd209, %rd11;
ld.global.u8 %rs58, [%rd127];
cvt.u32.u16 %r493, %rs58;
and.b32 %r494, %r489, 255;
and.b32 %r495, %r493, 7;
bfi.b32 %r496, %r495, %r494, 8, 3;
mul.wide.u32 %rd128, %r496, 4;
add.s64 %rd130, %rd198, %rd128;
ld.global.nc.u32 %r497, [%rd130];
and.b32 %r425, %r497, 252645135;
shr.u32 %r498, %r497, 4;
and.b32 %r429, %r498, 252645135;
mov.u32 %r483, 0;

dp4a.s32.s32 %r424, %r425, %r20, %r483;


dp4a.s32.s32 %r428, %r429, %r21, %r424;

and.b32 %r499, %r493, 8;
cvt.rn.f32.s32 %f199, %r499;
mul.ftz.f32 %f200, %f199, 0f3D000000;
mov.f32 %f201, 0fBF600000;
sub.ftz.f32 %f202, %f201, %f200;
mov.u32 %r486, 16843009;

dp4a.s32.s32 %r432, %r20, %r486, %r483;


dp4a.s32.s32 %r436, %r21, %r486, %r432;

cvt.rn.f32.s32 %f203, %r436;
fma.rn.ftz.f32 %f204, %f202, %f203, 0f00000000;
shr.u16 %rs59, %rs58, 4;
cvt.u32.u16 %r500, %rs59;
and.b32 %r501, %r490, 255;
and.b32 %r502, %r500, 7;
bfi.b32 %r503, %r502, %r501, 8, 3;
mul.wide.u32 %rd131, %r503, 4;
add.s64 %rd132, %rd198, %rd131;
ld.global.nc.u32 %r504, [%rd132];
and.b32 %r441, %r504, 252645135;
shr.u32 %r505, %r504, 4;
and.b32 %r445, %r505, 252645135;

dp4a.s32.s32 %r440, %r441, %r22, %r428;


dp4a.s32.s32 %r444, %r445, %r23, %r440;

and.b32 %r506, %r500, 8;
cvt.rn.f32.s32 %f205, %r506;
mul.ftz.f32 %f206, %f205, 0f3D000000;
sub.ftz.f32 %f207, %f201, %f206;

dp4a.s32.s32 %r448, %r22, %r486, %r483;


dp4a.s32.s32 %r452, %r23, %r486, %r448;

cvt.rn.f32.s32 %f208, %r452;
fma.rn.ftz.f32 %f209, %f207, %f208, %f204;
ld.global.u8 %rs60, [%rd127+1];
cvt.u32.u16 %r507, %rs60;
and.b32 %r508, %r491, 255;
and.b32 %r509, %r507, 7;
bfi.b32 %r510, %r509, %r508, 8, 3;
mul.wide.u32 %rd133, %r510, 4;
add.s64 %rd134, %rd198, %rd133;
ld.global.nc.u32 %r511, [%rd134];
and.b32 %r457, %r511, 252645135;
shr.u32 %r512, %r511, 4;
and.b32 %r461, %r512, 252645135;

dp4a.s32.s32 %r456, %r457, %r24, %r483;


dp4a.s32.s32 %r460, %r461, %r25, %r456;

and.b32 %r513, %r507, 8;
cvt.rn.f32.s32 %f210, %r513;
mul.ftz.f32 %f211, %f210, 0f3D000000;
sub.ftz.f32 %f212, %f201, %f211;

dp4a.s32.s32 %r464, %r24, %r486, %r483;


dp4a.s32.s32 %r468, %r25, %r486, %r464;

cvt.rn.f32.s32 %f213, %r468;
fma.rn.ftz.f32 %f214, %f212, %f213, 0f00000000;
shr.u16 %rs61, %rs60, 4;
cvt.u32.u16 %r514, %rs61;
mul.wide.u16 %r515, %rs61, 256;
and.b32 %r516, %r515, 1792;
or.b32 %r517, %r516, %r492;
mul.wide.u32 %rd135, %r517, 4;
add.s64 %rd136, %rd198, %rd135;
ld.global.nc.u32 %r518, [%rd136];
and.b32 %r473, %r518, 252645135;
shr.u32 %r519, %r518, 4;
and.b32 %r477, %r519, 252645135;

dp4a.s32.s32 %r472, %r473, %r26, %r460;


dp4a.s32.s32 %r476, %r477, %r27, %r472;

and.b32 %r520, %r514, 8;
cvt.rn.f32.s32 %f215, %r520;
mul.ftz.f32 %f216, %f215, 0f3D000000;
sub.ftz.f32 %f217, %f201, %f216;

dp4a.s32.s32 %r480, %r26, %r486, %r483;


dp4a.s32.s32 %r484, %r27, %r486, %r480;

cvt.rn.f32.s32 %f218, %r484;
fma.rn.ftz.f32 %f219, %f217, %f218, %f214;
or.b32 %r523, %r4, 1;
mad.lo.s32 %r524, %r51, %r523, %r17;
mul.wide.s32 %rd137, %r524, 56;
add.s64 %rd138, %rd209, %rd137;
ld.global.u16 %rs62, [%rd138+48];
shr.u16 %rs63, %rs62, 12;
ld.global.u8 %rs64, [%rd138+51];
and.b16 %rs65, %rs64, 240;
or.b16 %rs66, %rs65, %rs63;
ld.global.u16 %rs67, [%rd138+52];
shr.u16 %rs68, %rs67, 4;
and.b16 %rs69, %rs68, 3840;
or.b16 %rs70, %rs66, %rs69;
ld.global.u16 %rs71, [%rd138+54];
and.b16 %rs72, %rs71, -4096;
or.b16 %rs57, %rs70, %rs72;

{ cvt.f32.f16 %f197, %rs57;}



{.reg .f16 low,high;
mov.b32 {low,high},%r28;
cvt.f32.f16 %f198, low;}


mul.ftz.f32 %f220, %f197, %f198;
add.s64 %rd139, %rd209, %rd9;
ld.global.u16 %r525, [%rd139];
shr.u32 %r526, %r525, %r15;
shl.b32 %r527, %r526, 1;
and.b32 %r528, %r527, 14;
or.b32 %r529, %r528, 1;
shr.u32 %r530, %r526, 2;
and.b32 %r531, %r530, 14;
or.b32 %r532, %r531, 1;
cvt.rn.f32.s32 %f221, %r444;
add.ftz.f32 %f222, %f209, %f221;
cvt.rn.f32.s32 %f223, %r529;
cvt.rn.f32.s32 %f224, %r476;
add.ftz.f32 %f225, %f219, %f224;
cvt.rn.f32.s32 %f226, %r532;
mul.ftz.f32 %f227, %f225, %f226;
fma.rn.ftz.f32 %f228, %f222, %f223, %f227;
fma.rn.ftz.f32 %f33, %f220, %f228, %f33;

$L__BB265_14:
mul.wide.s32 %rd206, %r51, 56;
mov.u64 %rd197, iq1s_grid_gpu;
add.s64 %rd31, %rd27, %rd206;
ld.global.nc.u32 %r598, [%rd31];
shr.u32 %r599, %r598, 8;
shr.u32 %r600, %r598, 16;
shr.u32 %r601, %r598, 24;
add.s64 %rd32, %rd28, %rd206;
ld.global.nc.u8 %rs74, [%rd32];
and.b16 %rs75, %rs74, 240;
cvt.u32.u16 %r602, %rs74;
and.b32 %r603, %r598, 255;
and.b32 %r604, %r602, 7;
bfi.b32 %r605, %r604, %r603, 8, 3;
mul.wide.u32 %rd141, %r605, 4;
add.s64 %rd143, %rd197, %rd141;
ld.global.nc.u32 %r606, [%rd143];
and.b32 %r534, %r606, 252645135;
shr.u32 %r607, %r606, 4;
and.b32 %r538, %r607, 252645135;
mov.u32 %r592, 0;

dp4a.s32.s32 %r533, %r534, %r20, %r592;


dp4a.s32.s32 %r537, %r538, %r21, %r533;

and.b32 %r608, %r602, 8;
cvt.rn.f32.s32 %f231, %r608;
mul.ftz.f32 %f232, %f231, 0f3D000000;
mov.f32 %f233, 0fBF600000;
sub.ftz.f32 %f234, %f233, %f232;
mov.u32 %r595, 16843009;

dp4a.s32.s32 %r541, %r20, %r595, %r592;


dp4a.s32.s32 %r545, %r21, %r595, %r541;

cvt.rn.f32.s32 %f235, %r545;
fma.rn.ftz.f32 %f236, %f234, %f235, 0f00000000;
shr.u16 %rs76, %rs75, 4;
cvt.u32.u16 %r609, %rs76;
and.b32 %r610, %r599, 255;
and.b32 %r611, %r609, 7;
bfi.b32 %r612, %r611, %r610, 8, 3;
mul.wide.u32 %rd144, %r612, 4;
add.s64 %rd145, %rd197, %rd144;
ld.global.nc.u32 %r613, [%rd145];
and.b32 %r550, %r613, 252645135;
shr.u32 %r614, %r613, 4;
and.b32 %r554, %r614, 252645135;

dp4a.s32.s32 %r549, %r550, %r22, %r537;


dp4a.s32.s32 %r553, %r554, %r23, %r549;

and.b32 %r615, %r609, 8;
cvt.rn.f32.s32 %f237, %r615;
mul.ftz.f32 %f238, %f237, 0f3D000000;
sub.ftz.f32 %f239, %f233, %f238;

dp4a.s32.s32 %r557, %r22, %r595, %r592;


dp4a.s32.s32 %r561, %r23, %r595, %r557;

cvt.rn.f32.s32 %f240, %r561;
fma.rn.ftz.f32 %f241, %f239, %f240, %f236;
ld.global.nc.u8 %rs77, [%rd32+1];
and.b16 %rs78, %rs77, 240;
cvt.u32.u16 %r616, %rs77;
and.b32 %r617, %r600, 255;
and.b32 %r618, %r616, 7;
bfi.b32 %r619, %r618, %r617, 8, 3;
mul.wide.u32 %rd146, %r619, 4;
add.s64 %rd147, %rd197, %rd146;
ld.global.nc.u32 %r620, [%rd147];
and.b32 %r566, %r620, 252645135;
shr.u32 %r621, %r620, 4;
and.b32 %r570, %r621, 252645135;

dp4a.s32.s32 %r565, %r566, %r24, %r592;


dp4a.s32.s32 %r569, %r570, %r25, %r565;

and.b32 %r622, %r616, 8;
cvt.rn.f32.s32 %f242, %r622;
mul.ftz.f32 %f243, %f242, 0f3D000000;
sub.ftz.f32 %f244, %f233, %f243;

dp4a.s32.s32 %r573, %r24, %r595, %r592;


dp4a.s32.s32 %r577, %r25, %r595, %r573;

cvt.rn.f32.s32 %f245, %r577;
fma.rn.ftz.f32 %f246, %f244, %f245, 0f00000000;
shr.u16 %rs79, %rs78, 4;
cvt.u32.u16 %r623, %rs79;
mul.wide.u16 %r624, %rs79, 256;
and.b32 %r625, %r624, 1792;
or.b32 %r626, %r625, %r601;
mul.wide.u32 %rd148, %r626, 4;
add.s64 %rd149, %rd197, %rd148;
ld.global.nc.u32 %r627, [%rd149];
and.b32 %r582, %r627, 252645135;
shr.u32 %r628, %r627, 4;
and.b32 %r586, %r628, 252645135;

dp4a.s32.s32 %r581, %r582, %r26, %r569;


dp4a.s32.s32 %r585, %r586, %r27, %r581;

and.b32 %r629, %r623, 8;
cvt.rn.f32.s32 %f247, %r629;
mul.ftz.f32 %f248, %f247, 0f3D000000;
sub.ftz.f32 %f249, %f233, %f248;

dp4a.s32.s32 %r589, %r26, %r595, %r592;


dp4a.s32.s32 %r593, %r27, %r595, %r589;

cvt.rn.f32.s32 %f250, %r593;
fma.rn.ftz.f32 %f251, %f249, %f250, %f246;
add.s64 %rd33, %rd29, %rd206;
ld.global.nc.u16 %rs80, [%rd33];
shr.u16 %rs81, %rs80, 12;
ld.global.nc.u16 %rs82, [%rd33+2];
shr.u16 %rs83, %rs82, 8;
and.b16 %rs84, %rs83, 240;
or.b16 %rs85, %rs84, %rs81;
ld.global.nc.u16 %rs86, [%rd33+4];
shr.u16 %rs87, %rs86, 4;
and.b16 %rs88, %rs87, 3840;
or.b16 %rs89, %rs85, %rs88;
ld.global.nc.u16 %rs90, [%rd33+6];
and.b16 %rs91, %rs90, -4096;
or.b16 %rs73, %rs89, %rs91;

{ cvt.f32.f16 %f229, %rs73;}



{.reg .f16 low,high;
mov.b32 {low,high},%r28;
cvt.f32.f16 %f230, low;}


mul.ftz.f32 %f252, %f229, %f230;
add.s64 %rd34, %rd30, %rd206;
ld.global.nc.u16 %rs92, [%rd34];
cvt.u32.u16 %r630, %rs92;
shr.u32 %r631, %r630, %r15;
shl.b32 %r632, %r631, 1;
and.b32 %r633, %r632, 14;
or.b32 %r634, %r633, 1;
shr.u32 %r635, %r631, 2;
and.b32 %r636, %r635, 14;
or.b32 %r637, %r636, 1;
cvt.rn.f32.s32 %f253, %r553;
add.ftz.f32 %f254, %f241, %f253;
cvt.rn.f32.s32 %f255, %r634;
cvt.rn.f32.s32 %f256, %r585;
add.ftz.f32 %f257, %f251, %f256;
cvt.rn.f32.s32 %f258, %r637;
mul.ftz.f32 %f259, %f257, %f258;
fma.rn.ftz.f32 %f260, %f254, %f255, %f259;
fma.rn.ftz.f32 %f28, %f252, %f260, %f28;
@%p13 bra $L__BB265_16;

mov.u64 %rd201, iq1s_grid_gpu;
add.s64 %rd150, %rd209, %rd18;
ld.global.u32 %r703, [%rd150];
shr.u32 %r704, %r703, 8;
shr.u32 %r705, %r703, 16;
shr.u32 %r706, %r703, 24;
add.s64 %rd151, %rd209, %rd17;
ld.global.u8 %rs94, [%rd151];
cvt.u32.u16 %r707, %rs94;
and.b32 %r708, %r703, 255;
and.b32 %r709, %r707, 7;
bfi.b32 %r710, %r709, %r708, 8, 3;
mul.wide.u32 %rd152, %r710, 4;
add.s64 %rd154, %rd201, %rd152;
ld.global.nc.u32 %r711, [%rd154];
and.b32 %r639, %r711, 252645135;
shr.u32 %r712, %r711, 4;
and.b32 %r643, %r712, 252645135;

dp4a.s32.s32 %r638, %r639, %r20, %r592;


dp4a.s32.s32 %r642, %r643, %r21, %r638;

and.b32 %r713, %r707, 8;
cvt.rn.f32.s32 %f263, %r713;
mul.ftz.f32 %f264, %f263, 0f3D000000;
sub.ftz.f32 %f266, %f233, %f264;

dp4a.s32.s32 %r646, %r20, %r595, %r592;


dp4a.s32.s32 %r650, %r21, %r595, %r646;

cvt.rn.f32.s32 %f267, %r650;
fma.rn.ftz.f32 %f268, %f266, %f267, 0f00000000;
shr.u16 %rs95, %rs94, 4;
cvt.u32.u16 %r714, %rs95;
and.b32 %r715, %r704, 255;
and.b32 %r716, %r714, 7;
bfi.b32 %r717, %r716, %r715, 8, 3;
mul.wide.u32 %rd155, %r717, 4;
add.s64 %rd156, %rd201, %rd155;
ld.global.nc.u32 %r718, [%rd156];
and.b32 %r655, %r718, 252645135;
shr.u32 %r719, %r718, 4;
and.b32 %r659, %r719, 252645135;

dp4a.s32.s32 %r654, %r655, %r22, %r642;


dp4a.s32.s32 %r658, %r659, %r23, %r654;

and.b32 %r720, %r714, 8;
cvt.rn.f32.s32 %f269, %r720;
mul.ftz.f32 %f270, %f269, 0f3D000000;
sub.ftz.f32 %f271, %f233, %f270;

dp4a.s32.s32 %r662, %r22, %r595, %r592;


dp4a.s32.s32 %r666, %r23, %r595, %r662;

cvt.rn.f32.s32 %f272, %r666;
fma.rn.ftz.f32 %f273, %f271, %f272, %f268;
ld.global.u8 %rs96, [%rd151+1];
cvt.u32.u16 %r721, %rs96;
and.b32 %r722, %r705, 255;
and.b32 %r723, %r721, 7;
bfi.b32 %r724, %r723, %r722, 8, 3;
mul.wide.u32 %rd157, %r724, 4;
add.s64 %rd158, %rd201, %rd157;
ld.global.nc.u32 %r725, [%rd158];
and.b32 %r671, %r725, 252645135;
shr.u32 %r726, %r725, 4;
and.b32 %r675, %r726, 252645135;

dp4a.s32.s32 %r670, %r671, %r24, %r592;


dp4a.s32.s32 %r674, %r675, %r25, %r670;

and.b32 %r727, %r721, 8;
cvt.rn.f32.s32 %f274, %r727;
mul.ftz.f32 %f275, %f274, 0f3D000000;
sub.ftz.f32 %f276, %f233, %f275;

dp4a.s32.s32 %r678, %r24, %r595, %r592;


dp4a.s32.s32 %r682, %r25, %r595, %r678;

cvt.rn.f32.s32 %f277, %r682;
fma.rn.ftz.f32 %f278, %f276, %f277, 0f00000000;
shr.u16 %rs97, %rs96, 4;
cvt.u32.u16 %r728, %rs97;
mul.wide.u16 %r729, %rs97, 256;
and.b32 %r730, %r729, 1792;
or.b32 %r731, %r730, %r706;
mul.wide.u32 %rd159, %r731, 4;
add.s64 %rd160, %rd201, %rd159;
ld.global.nc.u32 %r732, [%rd160];
and.b32 %r687, %r732, 252645135;
shr.u32 %r733, %r732, 4;
and.b32 %r691, %r733, 252645135;

dp4a.s32.s32 %r686, %r687, %r26, %r674;


dp4a.s32.s32 %r690, %r691, %r27, %r686;

and.b32 %r734, %r728, 8;
cvt.rn.f32.s32 %f279, %r734;
mul.ftz.f32 %f280, %f279, 0f3D000000;
sub.ftz.f32 %f281, %f233, %f280;

dp4a.s32.s32 %r694, %r26, %r595, %r592;


dp4a.s32.s32 %r698, %r27, %r595, %r694;

cvt.rn.f32.s32 %f282, %r698;
fma.rn.ftz.f32 %f283, %f281, %f282, %f278;
or.b32 %r737, %r4, 2;
mad.lo.s32 %r738, %r51, %r737, %r17;
mul.wide.s32 %rd161, %r738, 56;
add.s64 %rd162, %rd209, %rd161;
ld.global.u16 %rs98, [%rd162+48];
shr.u16 %rs99, %rs98, 12;
ld.global.u8 %rs100, [%rd162+51];
and.b16 %rs101, %rs100, 240;
or.b16 %rs102, %rs101, %rs99;
ld.global.u16 %rs103, [%rd162+52];
shr.u16 %rs104, %rs103, 4;
and.b16 %rs105, %rs104, 3840;
or.b16 %rs106, %rs102, %rs105;
ld.global.u16 %rs107, [%rd162+54];
and.b16 %rs108, %rs107, -4096;
or.b16 %rs93, %rs106, %rs108;

{ cvt.f32.f16 %f261, %rs93;}



{.reg .f16 low,high;
mov.b32 {low,high},%r28;
cvt.f32.f16 %f262, low;}


mul.ftz.f32 %f284, %f261, %f262;
add.s64 %rd163, %rd209, %rd15;
ld.global.u16 %r739, [%rd163];
shr.u32 %r740, %r739, %r15;
shl.b32 %r741, %r740, 1;
and.b32 %r742, %r741, 14;
or.b32 %r743, %r742, 1;
shr.u32 %r744, %r740, 2;
and.b32 %r745, %r744, 14;
or.b32 %r746, %r745, 1;
cvt.rn.f32.s32 %f285, %r658;
add.ftz.f32 %f286, %f273, %f285;
cvt.rn.f32.s32 %f287, %r743;
cvt.rn.f32.s32 %f288, %r690;
add.ftz.f32 %f289, %f283, %f288;
cvt.rn.f32.s32 %f290, %r746;
mul.ftz.f32 %f291, %f289, %f290;
fma.rn.ftz.f32 %f292, %f286, %f287, %f291;
fma.rn.ftz.f32 %f32, %f284, %f292, %f32;

$L__BB265_16:
mul.wide.s32 %rd207, %r51, 56;
mov.u64 %rd194, iq1s_grid_gpu;
add.s64 %rd165, %rd31, %rd207;
ld.global.nc.u32 %r812, [%rd165];
shr.u32 %r813, %r812, 8;
shr.u32 %r814, %r812, 16;
shr.u32 %r815, %r812, 24;
add.s64 %rd166, %rd32, %rd207;
ld.global.nc.u8 %rs110, [%rd166];
and.b16 %rs111, %rs110, 240;
cvt.u32.u16 %r816, %rs110;
and.b32 %r817, %r812, 255;
and.b32 %r818, %r816, 7;
bfi.b32 %r819, %r818, %r817, 8, 3;
mul.wide.u32 %rd167, %r819, 4;
add.s64 %rd169, %rd194, %rd167;
ld.global.nc.u32 %r820, [%rd169];
and.b32 %r748, %r820, 252645135;
shr.u32 %r821, %r820, 4;
and.b32 %r752, %r821, 252645135;

dp4a.s32.s32 %r747, %r748, %r20, %r592;


dp4a.s32.s32 %r751, %r752, %r21, %r747;

and.b32 %r822, %r816, 8;
cvt.rn.f32.s32 %f295, %r822;
mul.ftz.f32 %f296, %f295, 0f3D000000;
sub.ftz.f32 %f298, %f233, %f296;

dp4a.s32.s32 %r755, %r20, %r595, %r592;


dp4a.s32.s32 %r759, %r21, %r595, %r755;

cvt.rn.f32.s32 %f299, %r759;
fma.rn.ftz.f32 %f300, %f298, %f299, 0f00000000;
shr.u16 %rs112, %rs111, 4;
cvt.u32.u16 %r823, %rs112;
and.b32 %r824, %r813, 255;
and.b32 %r825, %r823, 7;
bfi.b32 %r826, %r825, %r824, 8, 3;
mul.wide.u32 %rd170, %r826, 4;
add.s64 %rd171, %rd194, %rd170;
ld.global.nc.u32 %r827, [%rd171];
and.b32 %r764, %r827, 252645135;
shr.u32 %r828, %r827, 4;
and.b32 %r768, %r828, 252645135;

dp4a.s32.s32 %r763, %r764, %r22, %r751;


dp4a.s32.s32 %r767, %r768, %r23, %r763;

and.b32 %r829, %r823, 8;
cvt.rn.f32.s32 %f301, %r829;
mul.ftz.f32 %f302, %f301, 0f3D000000;
sub.ftz.f32 %f303, %f233, %f302;

dp4a.s32.s32 %r771, %r22, %r595, %r592;


dp4a.s32.s32 %r775, %r23, %r595, %r771;

cvt.rn.f32.s32 %f304, %r775;
fma.rn.ftz.f32 %f305, %f303, %f304, %f300;
ld.global.nc.u8 %rs113, [%rd166+1];
and.b16 %rs114, %rs113, 240;
cvt.u32.u16 %r830, %rs113;
and.b32 %r831, %r814, 255;
and.b32 %r832, %r830, 7;
bfi.b32 %r833, %r832, %r831, 8, 3;
mul.wide.u32 %rd172, %r833, 4;
add.s64 %rd173, %rd194, %rd172;
ld.global.nc.u32 %r834, [%rd173];
and.b32 %r780, %r834, 252645135;
shr.u32 %r835, %r834, 4;
and.b32 %r784, %r835, 252645135;

dp4a.s32.s32 %r779, %r780, %r24, %r592;


dp4a.s32.s32 %r783, %r784, %r25, %r779;

and.b32 %r836, %r830, 8;
cvt.rn.f32.s32 %f306, %r836;
mul.ftz.f32 %f307, %f306, 0f3D000000;
sub.ftz.f32 %f308, %f233, %f307;

dp4a.s32.s32 %r787, %r24, %r595, %r592;


dp4a.s32.s32 %r791, %r25, %r595, %r787;

cvt.rn.f32.s32 %f309, %r791;
fma.rn.ftz.f32 %f310, %f308, %f309, 0f00000000;
shr.u16 %rs115, %rs114, 4;
cvt.u32.u16 %r837, %rs115;
mul.wide.u16 %r838, %rs115, 256;
and.b32 %r839, %r838, 1792;
or.b32 %r840, %r839, %r815;
mul.wide.u32 %rd174, %r840, 4;
add.s64 %rd175, %rd194, %rd174;
ld.global.nc.u32 %r841, [%rd175];
and.b32 %r796, %r841, 252645135;
shr.u32 %r842, %r841, 4;
and.b32 %r800, %r842, 252645135;

dp4a.s32.s32 %r795, %r796, %r26, %r783;


dp4a.s32.s32 %r799, %r800, %r27, %r795;

and.b32 %r843, %r837, 8;
cvt.rn.f32.s32 %f311, %r843;
mul.ftz.f32 %f312, %f311, 0f3D000000;
sub.ftz.f32 %f313, %f233, %f312;

dp4a.s32.s32 %r803, %r26, %r595, %r592;


dp4a.s32.s32 %r807, %r27, %r595, %r803;

cvt.rn.f32.s32 %f314, %r807;
fma.rn.ftz.f32 %f315, %f313, %f314, %f310;
add.s64 %rd176, %rd33, %rd207;
ld.global.nc.u16 %rs116, [%rd176];
shr.u16 %rs117, %rs116, 12;
ld.global.nc.u16 %rs118, [%rd176+2];
shr.u16 %rs119, %rs118, 8;
and.b16 %rs120, %rs119, 240;
or.b16 %rs121, %rs120, %rs117;
ld.global.nc.u16 %rs122, [%rd176+4];
shr.u16 %rs123, %rs122, 4;
and.b16 %rs124, %rs123, 3840;
or.b16 %rs125, %rs121, %rs124;
ld.global.nc.u16 %rs126, [%rd176+6];
and.b16 %rs127, %rs126, -4096;
or.b16 %rs109, %rs125, %rs127;

{ cvt.f32.f16 %f293, %rs109;}



{.reg .f16 low,high;
mov.b32 {low,high},%r28;
cvt.f32.f16 %f294, low;}


mul.ftz.f32 %f316, %f293, %f294;
add.s64 %rd177, %rd34, %rd207;
ld.global.nc.u16 %rs128, [%rd177];
cvt.u32.u16 %r844, %rs128;
shr.u32 %r845, %r844, %r15;
shl.b32 %r846, %r845, 1;
and.b32 %r847, %r846, 14;
or.b32 %r848, %r847, 1;
shr.u32 %r849, %r845, 2;
and.b32 %r850, %r849, 14;
or.b32 %r851, %r850, 1;
cvt.rn.f32.s32 %f317, %r767;
add.ftz.f32 %f318, %f305, %f317;
cvt.rn.f32.s32 %f319, %r848;
cvt.rn.f32.s32 %f320, %r799;
add.ftz.f32 %f321, %f315, %f320;
cvt.rn.f32.s32 %f322, %r851;
mul.ftz.f32 %f323, %f321, %f322;
fma.rn.ftz.f32 %f324, %f318, %f319, %f323;
fma.rn.ftz.f32 %f27, %f316, %f324, %f27;
@%p13 bra $L__BB265_18;

mov.u64 %rd200, iq1s_grid_gpu;
add.s64 %rd178, %rd209, %rd13;
ld.global.u32 %r917, [%rd178];
shr.u32 %r918, %r917, 8;
shr.u32 %r919, %r917, 16;
shr.u32 %r920, %r917, 24;
add.s64 %rd179, %rd209, %rd14;
ld.global.u8 %rs130, [%rd179+-1];
cvt.u32.u16 %r921, %rs130;
and.b32 %r922, %r917, 255;
and.b32 %r923, %r921, 7;
bfi.b32 %r924, %r923, %r922, 8, 3;
mul.wide.u32 %rd180, %r924, 4;
add.s64 %rd182, %rd200, %rd180;
ld.global.nc.u32 %r925, [%rd182];
and.b32 %r853, %r925, 252645135;
shr.u32 %r926, %r925, 4;
and.b32 %r857, %r926, 252645135;
mov.u32 %r911, 0;

dp4a.s32.s32 %r852, %r853, %r20, %r911;


dp4a.s32.s32 %r856, %r857, %r21, %r852;

and.b32 %r927, %r921, 8;
cvt.rn.f32.s32 %f327, %r927;
mul.ftz.f32 %f328, %f327, 0f3D000000;
mov.f32 %f329, 0fBF600000;
sub.ftz.f32 %f330, %f329, %f328;
mov.u32 %r914, 16843009;

dp4a.s32.s32 %r860, %r20, %r914, %r911;


dp4a.s32.s32 %r864, %r21, %r914, %r860;

cvt.rn.f32.s32 %f331, %r864;
fma.rn.ftz.f32 %f332, %f330, %f331, 0f00000000;
shr.u16 %rs131, %rs130, 4;
cvt.u32.u16 %r928, %rs131;
and.b32 %r929, %r918, 255;
and.b32 %r930, %r928, 7;
bfi.b32 %r931, %r930, %r929, 8, 3;
mul.wide.u32 %rd183, %r931, 4;
add.s64 %rd184, %rd200, %rd183;
ld.global.nc.u32 %r932, [%rd184];
and.b32 %r869, %r932, 252645135;
shr.u32 %r933, %r932, 4;
and.b32 %r873, %r933, 252645135;

dp4a.s32.s32 %r868, %r869, %r22, %r856;


dp4a.s32.s32 %r872, %r873, %r23, %r868;

and.b32 %r934, %r928, 8;
cvt.rn.f32.s32 %f333, %r934;
mul.ftz.f32 %f334, %f333, 0f3D000000;
sub.ftz.f32 %f335, %f329, %f334;

dp4a.s32.s32 %r876, %r22, %r914, %r911;


dp4a.s32.s32 %r880, %r23, %r914, %r876;

cvt.rn.f32.s32 %f336, %r880;
fma.rn.ftz.f32 %f337, %f335, %f336, %f332;
ld.global.u8 %rs132, [%rd179];
cvt.u32.u16 %r935, %rs132;
and.b32 %r936, %r919, 255;
and.b32 %r937, %r935, 7;
bfi.b32 %r938, %r937, %r936, 8, 3;
mul.wide.u32 %rd185, %r938, 4;
add.s64 %rd186, %rd200, %rd185;
ld.global.nc.u32 %r939, [%rd186];
and.b32 %r885, %r939, 252645135;
shr.u32 %r940, %r939, 4;
and.b32 %r889, %r940, 252645135;

dp4a.s32.s32 %r884, %r885, %r24, %r911;


dp4a.s32.s32 %r888, %r889, %r25, %r884;

and.b32 %r941, %r935, 8;
cvt.rn.f32.s32 %f338, %r941;
mul.ftz.f32 %f339, %f338, 0f3D000000;
sub.ftz.f32 %f340, %f329, %f339;

dp4a.s32.s32 %r892, %r24, %r914, %r911;


dp4a.s32.s32 %r896, %r25, %r914, %r892;

cvt.rn.f32.s32 %f341, %r896;
fma.rn.ftz.f32 %f342, %f340, %f341, 0f00000000;
shr.u16 %rs133, %rs132, 4;
cvt.u32.u16 %r942, %rs133;
mul.wide.u16 %r943, %rs133, 256;
and.b32 %r944, %r943, 1792;
or.b32 %r945, %r944, %r920;
mul.wide.u32 %rd187, %r945, 4;
add.s64 %rd188, %rd200, %rd187;
ld.global.nc.u32 %r946, [%rd188];
and.b32 %r901, %r946, 252645135;
shr.u32 %r947, %r946, 4;
and.b32 %r905, %r947, 252645135;

dp4a.s32.s32 %r900, %r901, %r26, %r888;


dp4a.s32.s32 %r904, %r905, %r27, %r900;

and.b32 %r948, %r942, 8;
cvt.rn.f32.s32 %f343, %r948;
mul.ftz.f32 %f344, %f343, 0f3D000000;
sub.ftz.f32 %f345, %f329, %f344;

dp4a.s32.s32 %r908, %r26, %r914, %r911;


dp4a.s32.s32 %r912, %r27, %r914, %r908;

cvt.rn.f32.s32 %f346, %r912;
fma.rn.ftz.f32 %f347, %f345, %f346, %f342;
or.b32 %r951, %r4, 3;
mad.lo.s32 %r952, %r51, %r951, %r17;
mul.wide.s32 %rd189, %r952, 56;
add.s64 %rd190, %rd209, %rd189;
ld.global.u16 %rs134, [%rd190+48];
shr.u16 %rs135, %rs134, 12;
ld.global.u8 %rs136, [%rd190+51];
and.b16 %rs137, %rs136, 240;
or.b16 %rs138, %rs137, %rs135;
ld.global.u16 %rs139, [%rd190+52];
shr.u16 %rs140, %rs139, 4;
and.b16 %rs141, %rs140, 3840;
or.b16 %rs142, %rs138, %rs141;
ld.global.u16 %rs143, [%rd190+54];
and.b16 %rs144, %rs143, -4096;
or.b16 %rs129, %rs142, %rs144;

{ cvt.f32.f16 %f325, %rs129;}



{.reg .f16 low,high;
mov.b32 {low,high},%r28;
cvt.f32.f16 %f326, low;}


mul.ftz.f32 %f348, %f325, %f326;
add.s64 %rd191, %rd209, %rd16;
ld.global.u16 %r953, [%rd191];
shr.u32 %r954, %r953, %r15;
shl.b32 %r955, %r954, 1;
and.b32 %r956, %r955, 14;
or.b32 %r957, %r956, 1;
shr.u32 %r958, %r954, 2;
and.b32 %r959, %r958, 14;
or.b32 %r960, %r959, 1;
cvt.rn.f32.s32 %f349, %r872;
add.ftz.f32 %f350, %f337, %f349;
cvt.rn.f32.s32 %f351, %r957;
cvt.rn.f32.s32 %f352, %r904;
add.ftz.f32 %f353, %f347, %f352;
cvt.rn.f32.s32 %f354, %r960;
mul.ftz.f32 %f355, %f353, %f354;
fma.rn.ftz.f32 %f356, %f350, %f351, %f355;
fma.rn.ftz.f32 %f31, %f348, %f356, %f31;

$L__BB265_18:
add.s32 %r1707, %r1707, 128;
add.s64 %rd209, %rd209, 896;
add.s64 %rd208, %rd208, 896;
add.s32 %r1708, %r1708, 16;
setp.lt.s32 %p17, %r1708, %r85;
@%p17 bra $L__BB265_10;

$L__BB265_19:
mov.u32 %r1694, %tid.y;
setp.eq.s32 %p18, %r1694, 0;
@%p18 bra $L__BB265_23;

mov.u32 %r1695, %tid.y;
setp.ne.s64 %p19, %rd42, 0;
add.s32 %r31, %r1695, -1;
shl.b32 %r965, %r1695, 9;
mov.u32 %r966, _ZZ13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r967, %r966, %r965;
shl.b32 %r968, %r2, 2;
add.s32 %r969, %r967, %r968;
st.shared.f32 [%r969+-512], %f30;
@%p19 bra $L__BB265_22;
bra.uni $L__BB265_21;

$L__BB265_22:
shl.b32 %r971, %r31, 9;
mov.u32 %r972, _ZZ13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r973, %r972, %r971;
add.s32 %r975, %r973, %r968;
st.shared.f32 [%r975], %f34;
add.s32 %r1691, %r969, -512;
st.shared.f32 [%r1691+128], %f29;
st.shared.f32 [%r975+128], %f33;
add.s32 %r1692, %r969, -512;
st.shared.f32 [%r1692+256], %f28;
st.shared.f32 [%r975+256], %f32;
add.s32 %r1693, %r969, -512;
st.shared.f32 [%r1693+384], %f27;
st.shared.f32 [%r975+384], %f31;
bra.uni $L__BB265_23;

$L__BB265_21:
add.s32 %r1688, %r969, -512;
st.shared.f32 [%r1688+128], %f29;
add.s32 %r1689, %r969, -512;
st.shared.f32 [%r1689+256], %f28;
add.s32 %r1690, %r969, -512;
st.shared.f32 [%r1690+384], %f27;

$L__BB265_23:
mov.u32 %r1696, %tid.y;
setp.ne.s32 %p319, %r1696, 0;
bar.sync 0;
@%p319 bra $L__BB265_78;
bra.uni $L__BB265_24;

$L__BB265_78:
ret;

$L__BB265_24:
ld.param.u64 %rd195, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.f32 %f1201, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
ld.param.u32 %r1699, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r1698, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
mov.u32 %r1697, %ctaid.z;
setp.ne.s64 %p21, %rd42, 0;
mov.u32 %r978, %ctaid.y;
mad.lo.s32 %r981, %r978, %r1699, %r4;
mad.lo.s32 %r982, %r1697, %r1698, %r981;
neg.ftz.f32 %f35, %f1201;
shl.b32 %r984, %r2, 2;
mov.u32 %r985, _ZZ13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r33, %r985, %r984;
cvta.to.global.u64 %rd192, %rd195;
mul.wide.u32 %rd193, %r982, 4;
add.s64 %rd37, %rd192, %rd193;
@%p21 bra $L__BB265_33;
bra.uni $L__BB265_25;

$L__BB265_33:
ld.param.u32 %r1700, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
mov.u32 %r1072, _ZZ13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r35, %r1072, %r984;
setp.gt.s32 %p54, %r1700, 2;
@%p54 bra $L__BB265_44;

ld.param.u32 %r1703, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p57, %r1703, 1;
@%p57 bra $L__BB265_62;

ld.param.u32 %r1704, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p58, %r1704, 2;
@%p58 bra $L__BB265_36;
bra.uni $L__BB265_70;

$L__BB265_36:
ld.shared.f32 %f757, [%r33];
add.ftz.f32 %f758, %f757, %f30;
ld.shared.f32 %f759, [%r35];
add.ftz.f32 %f760, %f759, %f34;
ld.shared.f32 %f761, [%r33+512];
add.ftz.f32 %f762, %f761, %f758;
ld.shared.f32 %f763, [%r35+512];
add.ftz.f32 %f764, %f763, %f760;
ld.shared.f32 %f765, [%r33+1024];
add.ftz.f32 %f766, %f765, %f762;
ld.shared.f32 %f767, [%r35+1024];
add.ftz.f32 %f768, %f767, %f764;
mov.b32 %r1319, %f766;
mov.u32 %r1320, 31;
mov.u32 %r1321, 16;
mov.u32 %r1322, -1;
shfl.sync.bfly.b32 %r1323|%p163, %r1319, %r1321, %r1320, %r1322;
mov.b32 %f769, %r1323;
add.ftz.f32 %f770, %f766, %f769;
mov.b32 %r1324, %f770;
mov.u32 %r1325, 8;
shfl.sync.bfly.b32 %r1326|%p164, %r1324, %r1325, %r1320, %r1322;
mov.b32 %f771, %r1326;
add.ftz.f32 %f772, %f770, %f771;
mov.b32 %r1327, %f772;
mov.u32 %r1328, 4;
shfl.sync.bfly.b32 %r1329|%p165, %r1327, %r1328, %r1320, %r1322;
mov.b32 %f773, %r1329;
add.ftz.f32 %f774, %f772, %f773;
mov.b32 %r1330, %f774;
mov.u32 %r1331, 2;
shfl.sync.bfly.b32 %r1332|%p166, %r1330, %r1331, %r1320, %r1322;
mov.b32 %f775, %r1332;
add.ftz.f32 %f776, %f774, %f775;
mov.b32 %r1333, %f776;
mov.u32 %r1334, 1;
shfl.sync.bfly.b32 %r1335|%p167, %r1333, %r1334, %r1320, %r1322;
mov.b32 %f777, %r1335;
add.ftz.f32 %f56, %f776, %f777;
mov.b32 %r1336, %f768;
shfl.sync.bfly.b32 %r1337|%p168, %r1336, %r1321, %r1320, %r1322;
mov.b32 %f778, %r1337;
add.ftz.f32 %f779, %f768, %f778;
mov.b32 %r1338, %f779;
shfl.sync.bfly.b32 %r1339|%p169, %r1338, %r1325, %r1320, %r1322;
mov.b32 %f780, %r1339;
add.ftz.f32 %f781, %f779, %f780;
mov.b32 %r1340, %f781;
shfl.sync.bfly.b32 %r1341|%p170, %r1340, %r1328, %r1320, %r1322;
mov.b32 %f782, %r1341;
add.ftz.f32 %f783, %f781, %f782;
mov.b32 %r1342, %f783;
shfl.sync.bfly.b32 %r1343|%p171, %r1342, %r1331, %r1320, %r1322;
mov.b32 %f784, %r1343;
add.ftz.f32 %f785, %f783, %f784;
mov.b32 %r1344, %f785;
shfl.sync.bfly.b32 %r1345|%p172, %r1344, %r1334, %r1320, %r1322;
mov.b32 %f786, %r1345;
add.ftz.f32 %f57, %f785, %f786;
setp.ne.s32 %p173, %r2, 0;
setp.ge.u32 %p174, %r4, %r52;
or.pred %p175, %p173, %p174;
@%p175 bra $L__BB265_38;

add.ftz.f32 %f787, %f5, %f56;
add.ftz.f32 %f788, %f4, %f57;
mul.ftz.f32 %f789, %f788, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f790, %f789;
add.ftz.f32 %f791, %f790, 0f3F800000;
div.approx.ftz.f32 %f792, %f788, %f791;
mul.ftz.f32 %f793, %f787, %f792;
st.global.f32 [%rd37], %f793;

$L__BB265_38:
ld.shared.f32 %f794, [%r33+128];
add.ftz.f32 %f795, %f794, %f29;
ld.shared.f32 %f796, [%r35+128];
add.ftz.f32 %f797, %f796, %f33;
ld.shared.f32 %f798, [%r33+640];
add.ftz.f32 %f799, %f798, %f795;
ld.shared.f32 %f800, [%r35+640];
add.ftz.f32 %f801, %f800, %f797;
ld.shared.f32 %f802, [%r33+1152];
add.ftz.f32 %f803, %f802, %f799;
ld.shared.f32 %f804, [%r35+1152];
add.ftz.f32 %f805, %f804, %f801;
mov.b32 %r1349, %f803;
shfl.sync.bfly.b32 %r1353|%p176, %r1349, %r1321, %r1320, %r1322;
mov.b32 %f806, %r1353;
add.ftz.f32 %f807, %f803, %f806;
mov.b32 %r1354, %f807;
shfl.sync.bfly.b32 %r1356|%p177, %r1354, %r1325, %r1320, %r1322;
mov.b32 %f808, %r1356;
add.ftz.f32 %f809, %f807, %f808;
mov.b32 %r1357, %f809;
shfl.sync.bfly.b32 %r1359|%p178, %r1357, %r1328, %r1320, %r1322;
mov.b32 %f810, %r1359;
add.ftz.f32 %f811, %f809, %f810;
mov.b32 %r1360, %f811;
shfl.sync.bfly.b32 %r1362|%p179, %r1360, %r1331, %r1320, %r1322;
mov.b32 %f812, %r1362;
add.ftz.f32 %f813, %f811, %f812;
mov.b32 %r1363, %f813;
shfl.sync.bfly.b32 %r1365|%p180, %r1363, %r1334, %r1320, %r1322;
mov.b32 %f814, %r1365;
add.ftz.f32 %f58, %f813, %f814;
mov.b32 %r1366, %f805;
shfl.sync.bfly.b32 %r1367|%p181, %r1366, %r1321, %r1320, %r1322;
mov.b32 %f815, %r1367;
add.ftz.f32 %f816, %f805, %f815;
mov.b32 %r1368, %f816;
shfl.sync.bfly.b32 %r1369|%p182, %r1368, %r1325, %r1320, %r1322;
mov.b32 %f817, %r1369;
add.ftz.f32 %f818, %f816, %f817;
mov.b32 %r1370, %f818;
shfl.sync.bfly.b32 %r1371|%p183, %r1370, %r1328, %r1320, %r1322;
mov.b32 %f819, %r1371;
add.ftz.f32 %f820, %f818, %f819;
mov.b32 %r1372, %f820;
shfl.sync.bfly.b32 %r1373|%p184, %r1372, %r1331, %r1320, %r1322;
mov.b32 %f821, %r1373;
add.ftz.f32 %f822, %f820, %f821;
mov.b32 %r1374, %f822;
shfl.sync.bfly.b32 %r1375|%p185, %r1374, %r1334, %r1320, %r1322;
mov.b32 %f823, %r1375;
add.ftz.f32 %f59, %f822, %f823;
setp.ne.s32 %p186, %r2, 1;
or.b32 %r1379, %r4, 1;
setp.ge.u32 %p187, %r1379, %r52;
or.pred %p188, %p186, %p187;
@%p188 bra $L__BB265_40;

add.ftz.f32 %f824, %f5, %f58;
add.ftz.f32 %f825, %f4, %f59;
mul.ftz.f32 %f826, %f825, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f827, %f826;
add.ftz.f32 %f828, %f827, 0f3F800000;
div.approx.ftz.f32 %f829, %f825, %f828;
mul.ftz.f32 %f830, %f824, %f829;
st.global.f32 [%rd37+4], %f830;

$L__BB265_40:
ld.shared.f32 %f831, [%r33+256];
add.ftz.f32 %f832, %f831, %f28;
ld.shared.f32 %f833, [%r35+256];
add.ftz.f32 %f834, %f833, %f32;
ld.shared.f32 %f835, [%r33+768];
add.ftz.f32 %f836, %f835, %f832;
ld.shared.f32 %f837, [%r35+768];
add.ftz.f32 %f838, %f837, %f834;
ld.shared.f32 %f839, [%r33+1280];
add.ftz.f32 %f840, %f839, %f836;
ld.shared.f32 %f841, [%r35+1280];
add.ftz.f32 %f842, %f841, %f838;
mov.b32 %r1380, %f840;
mov.u32 %r1381, 31;
mov.u32 %r1382, 16;
mov.u32 %r1383, -1;
shfl.sync.bfly.b32 %r1384|%p189, %r1380, %r1382, %r1381, %r1383;
mov.b32 %f843, %r1384;
add.ftz.f32 %f844, %f840, %f843;
mov.b32 %r1385, %f844;
mov.u32 %r1386, 8;
shfl.sync.bfly.b32 %r1387|%p190, %r1385, %r1386, %r1381, %r1383;
mov.b32 %f845, %r1387;
add.ftz.f32 %f846, %f844, %f845;
mov.b32 %r1388, %f846;
mov.u32 %r1389, 4;
shfl.sync.bfly.b32 %r1390|%p191, %r1388, %r1389, %r1381, %r1383;
mov.b32 %f847, %r1390;
add.ftz.f32 %f848, %f846, %f847;
mov.b32 %r1391, %f848;
mov.u32 %r1392, 2;
shfl.sync.bfly.b32 %r1393|%p192, %r1391, %r1392, %r1381, %r1383;
mov.b32 %f849, %r1393;
add.ftz.f32 %f850, %f848, %f849;
mov.b32 %r1394, %f850;
mov.u32 %r1395, 1;
shfl.sync.bfly.b32 %r1396|%p193, %r1394, %r1395, %r1381, %r1383;
mov.b32 %f851, %r1396;
add.ftz.f32 %f60, %f850, %f851;
mov.b32 %r1397, %f842;
shfl.sync.bfly.b32 %r1398|%p194, %r1397, %r1382, %r1381, %r1383;
mov.b32 %f852, %r1398;
add.ftz.f32 %f853, %f842, %f852;
mov.b32 %r1399, %f853;
shfl.sync.bfly.b32 %r1400|%p195, %r1399, %r1386, %r1381, %r1383;
mov.b32 %f854, %r1400;
add.ftz.f32 %f855, %f853, %f854;
mov.b32 %r1401, %f855;
shfl.sync.bfly.b32 %r1402|%p196, %r1401, %r1389, %r1381, %r1383;
mov.b32 %f856, %r1402;
add.ftz.f32 %f857, %f855, %f856;
mov.b32 %r1403, %f857;
shfl.sync.bfly.b32 %r1404|%p197, %r1403, %r1392, %r1381, %r1383;
mov.b32 %f858, %r1404;
add.ftz.f32 %f859, %f857, %f858;
mov.b32 %r1405, %f859;
shfl.sync.bfly.b32 %r1406|%p198, %r1405, %r1395, %r1381, %r1383;
mov.b32 %f860, %r1406;
add.ftz.f32 %f61, %f859, %f860;
setp.ne.s32 %p199, %r2, 2;
or.b32 %r1410, %r4, 2;
setp.ge.u32 %p200, %r1410, %r52;
or.pred %p201, %p199, %p200;
@%p201 bra $L__BB265_42;

add.ftz.f32 %f861, %f5, %f60;
add.ftz.f32 %f862, %f4, %f61;
mul.ftz.f32 %f863, %f862, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f864, %f863;
add.ftz.f32 %f865, %f864, 0f3F800000;
div.approx.ftz.f32 %f866, %f862, %f865;
mul.ftz.f32 %f867, %f861, %f866;
st.global.f32 [%rd37+8], %f867;

$L__BB265_42:
ld.shared.f32 %f868, [%r33+384];
add.ftz.f32 %f869, %f868, %f27;
ld.shared.f32 %f870, [%r35+384];
add.ftz.f32 %f871, %f870, %f31;
ld.shared.f32 %f872, [%r33+896];
add.ftz.f32 %f873, %f872, %f869;
ld.shared.f32 %f874, [%r35+896];
add.ftz.f32 %f875, %f874, %f871;
ld.shared.f32 %f876, [%r33+1408];
add.ftz.f32 %f877, %f876, %f873;
ld.shared.f32 %f878, [%r35+1408];
add.ftz.f32 %f879, %f878, %f875;
mov.b32 %r1411, %f877;
shfl.sync.bfly.b32 %r1415|%p202, %r1411, %r1382, %r1381, %r1383;
mov.b32 %f880, %r1415;
add.ftz.f32 %f881, %f877, %f880;
mov.b32 %r1416, %f881;
shfl.sync.bfly.b32 %r1418|%p203, %r1416, %r1386, %r1381, %r1383;
mov.b32 %f882, %r1418;
add.ftz.f32 %f883, %f881, %f882;
mov.b32 %r1419, %f883;
shfl.sync.bfly.b32 %r1421|%p204, %r1419, %r1389, %r1381, %r1383;
mov.b32 %f884, %r1421;
add.ftz.f32 %f885, %f883, %f884;
mov.b32 %r1422, %f885;
shfl.sync.bfly.b32 %r1424|%p205, %r1422, %r1392, %r1381, %r1383;
mov.b32 %f886, %r1424;
add.ftz.f32 %f887, %f885, %f886;
mov.b32 %r1425, %f887;
shfl.sync.bfly.b32 %r1427|%p206, %r1425, %r1395, %r1381, %r1383;
mov.b32 %f888, %r1427;
add.ftz.f32 %f62, %f887, %f888;
mov.b32 %r1428, %f879;
shfl.sync.bfly.b32 %r1429|%p207, %r1428, %r1382, %r1381, %r1383;
mov.b32 %f889, %r1429;
add.ftz.f32 %f890, %f879, %f889;
mov.b32 %r1430, %f890;
shfl.sync.bfly.b32 %r1431|%p208, %r1430, %r1386, %r1381, %r1383;
mov.b32 %f891, %r1431;
add.ftz.f32 %f892, %f890, %f891;
mov.b32 %r1432, %f892;
shfl.sync.bfly.b32 %r1433|%p209, %r1432, %r1389, %r1381, %r1383;
mov.b32 %f893, %r1433;
add.ftz.f32 %f894, %f892, %f893;
mov.b32 %r1434, %f894;
shfl.sync.bfly.b32 %r1435|%p210, %r1434, %r1392, %r1381, %r1383;
mov.b32 %f895, %r1435;
add.ftz.f32 %f896, %f894, %f895;
mov.b32 %r1436, %f896;
shfl.sync.bfly.b32 %r1437|%p211, %r1436, %r1395, %r1381, %r1383;
mov.b32 %f897, %r1437;
add.ftz.f32 %f63, %f896, %f897;
setp.ne.s32 %p212, %r2, 3;
or.b32 %r1441, %r4, 3;
setp.ge.u32 %p213, %r1441, %r52;
or.pred %p214, %p212, %p213;
@%p214 bra $L__BB265_78;

add.ftz.f32 %f898, %f5, %f62;
add.ftz.f32 %f899, %f4, %f63;
mul.ftz.f32 %f900, %f899, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f901, %f900;
add.ftz.f32 %f902, %f901, 0f3F800000;
div.approx.ftz.f32 %f903, %f899, %f902;
mul.ftz.f32 %f904, %f898, %f903;
st.global.f32 [%rd37+12], %f904;
bra.uni $L__BB265_78;

$L__BB265_25:
ld.shared.f32 %f357, [%r33];
add.ftz.f32 %f358, %f357, %f30;
ld.shared.f32 %f359, [%r33+512];
add.ftz.f32 %f360, %f359, %f358;
ld.shared.f32 %f361, [%r33+1024];
add.ftz.f32 %f362, %f361, %f360;
mov.b32 %r986, %f362;
mov.u32 %r987, 31;
mov.u32 %r988, 16;
mov.u32 %r989, -1;
shfl.sync.bfly.b32 %r990|%p22, %r986, %r988, %r987, %r989;
mov.b32 %f363, %r990;
add.ftz.f32 %f364, %f362, %f363;
mov.b32 %r991, %f364;
mov.u32 %r992, 8;
shfl.sync.bfly.b32 %r993|%p23, %r991, %r992, %r987, %r989;
mov.b32 %f365, %r993;
add.ftz.f32 %f366, %f364, %f365;
mov.b32 %r994, %f366;
mov.u32 %r995, 4;
shfl.sync.bfly.b32 %r996|%p24, %r994, %r995, %r987, %r989;
mov.b32 %f367, %r996;
add.ftz.f32 %f368, %f366, %f367;
mov.b32 %r997, %f368;
mov.u32 %r998, 2;
shfl.sync.bfly.b32 %r999|%p25, %r997, %r998, %r987, %r989;
mov.b32 %f369, %r999;
add.ftz.f32 %f370, %f368, %f369;
mov.b32 %r1000, %f370;
mov.u32 %r1001, 1;
shfl.sync.bfly.b32 %r1002|%p26, %r1000, %r1001, %r987, %r989;
mov.b32 %f371, %r1002;
add.ftz.f32 %f36, %f370, %f371;
setp.ne.s32 %p27, %r2, 0;
setp.ge.u32 %p28, %r4, %r52;
or.pred %p29, %p27, %p28;
@%p29 bra $L__BB265_27;

add.ftz.f32 %f372, %f5, %f36;
st.global.f32 [%rd37], %f372;

$L__BB265_27:
ld.shared.f32 %f373, [%r33+128];
add.ftz.f32 %f374, %f373, %f29;
ld.shared.f32 %f375, [%r33+640];
add.ftz.f32 %f376, %f375, %f374;
ld.shared.f32 %f377, [%r33+1152];
add.ftz.f32 %f378, %f377, %f376;
mov.b32 %r1006, %f378;
shfl.sync.bfly.b32 %r1010|%p30, %r1006, %r988, %r987, %r989;
mov.b32 %f379, %r1010;
add.ftz.f32 %f380, %f378, %f379;
mov.b32 %r1011, %f380;
shfl.sync.bfly.b32 %r1013|%p31, %r1011, %r992, %r987, %r989;
mov.b32 %f381, %r1013;
add.ftz.f32 %f382, %f380, %f381;
mov.b32 %r1014, %f382;
shfl.sync.bfly.b32 %r1016|%p32, %r1014, %r995, %r987, %r989;
mov.b32 %f383, %r1016;
add.ftz.f32 %f384, %f382, %f383;
mov.b32 %r1017, %f384;
shfl.sync.bfly.b32 %r1019|%p33, %r1017, %r998, %r987, %r989;
mov.b32 %f385, %r1019;
add.ftz.f32 %f386, %f384, %f385;
mov.b32 %r1020, %f386;
shfl.sync.bfly.b32 %r1022|%p34, %r1020, %r1001, %r987, %r989;
mov.b32 %f387, %r1022;
add.ftz.f32 %f37, %f386, %f387;
setp.ne.s32 %p35, %r2, 1;
or.b32 %r1026, %r4, 1;
setp.ge.u32 %p36, %r1026, %r52;
or.pred %p37, %p35, %p36;
@%p37 bra $L__BB265_29;

add.ftz.f32 %f388, %f5, %f37;
st.global.f32 [%rd37+4], %f388;

$L__BB265_29:
ld.shared.f32 %f389, [%r33+256];
add.ftz.f32 %f390, %f389, %f28;
ld.shared.f32 %f391, [%r33+768];
add.ftz.f32 %f392, %f391, %f390;
ld.shared.f32 %f393, [%r33+1280];
add.ftz.f32 %f394, %f393, %f392;
mov.b32 %r1027, %f394;
mov.u32 %r1028, 31;
mov.u32 %r1029, 16;
mov.u32 %r1030, -1;
shfl.sync.bfly.b32 %r1031|%p38, %r1027, %r1029, %r1028, %r1030;
mov.b32 %f395, %r1031;
add.ftz.f32 %f396, %f394, %f395;
mov.b32 %r1032, %f396;
mov.u32 %r1033, 8;
shfl.sync.bfly.b32 %r1034|%p39, %r1032, %r1033, %r1028, %r1030;
mov.b32 %f397, %r1034;
add.ftz.f32 %f398, %f396, %f397;
mov.b32 %r1035, %f398;
mov.u32 %r1036, 4;
shfl.sync.bfly.b32 %r1037|%p40, %r1035, %r1036, %r1028, %r1030;
mov.b32 %f399, %r1037;
add.ftz.f32 %f400, %f398, %f399;
mov.b32 %r1038, %f400;
mov.u32 %r1039, 2;
shfl.sync.bfly.b32 %r1040|%p41, %r1038, %r1039, %r1028, %r1030;
mov.b32 %f401, %r1040;
add.ftz.f32 %f402, %f400, %f401;
mov.b32 %r1041, %f402;
mov.u32 %r1042, 1;
shfl.sync.bfly.b32 %r1043|%p42, %r1041, %r1042, %r1028, %r1030;
mov.b32 %f403, %r1043;
add.ftz.f32 %f38, %f402, %f403;
setp.ne.s32 %p43, %r2, 2;
or.b32 %r1047, %r4, 2;
setp.ge.u32 %p44, %r1047, %r52;
or.pred %p45, %p43, %p44;
@%p45 bra $L__BB265_31;

add.ftz.f32 %f404, %f5, %f38;
st.global.f32 [%rd37+8], %f404;

$L__BB265_31:
ld.shared.f32 %f405, [%r33+384];
add.ftz.f32 %f406, %f405, %f27;
ld.shared.f32 %f407, [%r33+896];
add.ftz.f32 %f408, %f407, %f406;
ld.shared.f32 %f409, [%r33+1408];
add.ftz.f32 %f410, %f409, %f408;
mov.b32 %r1048, %f410;
shfl.sync.bfly.b32 %r1052|%p46, %r1048, %r1029, %r1028, %r1030;
mov.b32 %f411, %r1052;
add.ftz.f32 %f412, %f410, %f411;
mov.b32 %r1053, %f412;
shfl.sync.bfly.b32 %r1055|%p47, %r1053, %r1033, %r1028, %r1030;
mov.b32 %f413, %r1055;
add.ftz.f32 %f414, %f412, %f413;
mov.b32 %r1056, %f414;
shfl.sync.bfly.b32 %r1058|%p48, %r1056, %r1036, %r1028, %r1030;
mov.b32 %f415, %r1058;
add.ftz.f32 %f416, %f414, %f415;
mov.b32 %r1059, %f416;
shfl.sync.bfly.b32 %r1061|%p49, %r1059, %r1039, %r1028, %r1030;
mov.b32 %f417, %r1061;
add.ftz.f32 %f418, %f416, %f417;
mov.b32 %r1062, %f418;
shfl.sync.bfly.b32 %r1064|%p50, %r1062, %r1042, %r1028, %r1030;
mov.b32 %f419, %r1064;
add.ftz.f32 %f39, %f418, %f419;
setp.ne.s32 %p51, %r2, 3;
or.b32 %r1068, %r4, 3;
setp.ge.u32 %p52, %r1068, %r52;
or.pred %p53, %p51, %p52;
@%p53 bra $L__BB265_78;

add.ftz.f32 %f420, %f5, %f39;
st.global.f32 [%rd37+12], %f420;
bra.uni $L__BB265_78;

$L__BB265_44:
ld.param.u32 %r1701, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p55, %r1701, 3;
@%p55 bra $L__BB265_54;

ld.param.u32 %r1702, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.ne.s32 %p56, %r1702, 6;
@%p56 bra $L__BB265_70;

ld.shared.f32 %f421, [%r33];
add.ftz.f32 %f422, %f421, %f30;
ld.shared.f32 %f423, [%r35];
add.ftz.f32 %f424, %f423, %f34;
ld.shared.f32 %f425, [%r33+512];
add.ftz.f32 %f426, %f425, %f422;
ld.shared.f32 %f427, [%r35+512];
add.ftz.f32 %f428, %f427, %f424;
ld.shared.f32 %f429, [%r33+1024];
add.ftz.f32 %f430, %f429, %f426;
ld.shared.f32 %f431, [%r35+1024];
add.ftz.f32 %f432, %f431, %f428;
mov.b32 %r1073, %f430;
mov.u32 %r1074, 31;
mov.u32 %r1075, 16;
mov.u32 %r1076, -1;
shfl.sync.bfly.b32 %r1077|%p59, %r1073, %r1075, %r1074, %r1076;
mov.b32 %f433, %r1077;
add.ftz.f32 %f434, %f430, %f433;
mov.b32 %r1078, %f434;
mov.u32 %r1079, 8;
shfl.sync.bfly.b32 %r1080|%p60, %r1078, %r1079, %r1074, %r1076;
mov.b32 %f435, %r1080;
add.ftz.f32 %f436, %f434, %f435;
mov.b32 %r1081, %f436;
mov.u32 %r1082, 4;
shfl.sync.bfly.b32 %r1083|%p61, %r1081, %r1082, %r1074, %r1076;
mov.b32 %f437, %r1083;
add.ftz.f32 %f438, %f436, %f437;
mov.b32 %r1084, %f438;
mov.u32 %r1085, 2;
shfl.sync.bfly.b32 %r1086|%p62, %r1084, %r1085, %r1074, %r1076;
mov.b32 %f439, %r1086;
add.ftz.f32 %f440, %f438, %f439;
mov.b32 %r1087, %f440;
mov.u32 %r1088, 1;
shfl.sync.bfly.b32 %r1089|%p63, %r1087, %r1088, %r1074, %r1076;
mov.b32 %f441, %r1089;
add.ftz.f32 %f40, %f440, %f441;
mov.b32 %r1090, %f432;
shfl.sync.bfly.b32 %r1091|%p64, %r1090, %r1075, %r1074, %r1076;
mov.b32 %f442, %r1091;
add.ftz.f32 %f443, %f432, %f442;
mov.b32 %r1092, %f443;
shfl.sync.bfly.b32 %r1093|%p65, %r1092, %r1079, %r1074, %r1076;
mov.b32 %f444, %r1093;
add.ftz.f32 %f445, %f443, %f444;
mov.b32 %r1094, %f445;
shfl.sync.bfly.b32 %r1095|%p66, %r1094, %r1082, %r1074, %r1076;
mov.b32 %f446, %r1095;
add.ftz.f32 %f447, %f445, %f446;
mov.b32 %r1096, %f447;
shfl.sync.bfly.b32 %r1097|%p67, %r1096, %r1085, %r1074, %r1076;
mov.b32 %f448, %r1097;
add.ftz.f32 %f449, %f447, %f448;
mov.b32 %r1098, %f449;
shfl.sync.bfly.b32 %r1099|%p68, %r1098, %r1088, %r1074, %r1076;
mov.b32 %f450, %r1099;
add.ftz.f32 %f41, %f449, %f450;
setp.ne.s32 %p69, %r2, 0;
setp.ge.u32 %p70, %r4, %r52;
or.pred %p71, %p69, %p70;
@%p71 bra $L__BB265_48;

ld.param.f32 %f1205, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f451, %f5, %f40;
add.ftz.f32 %f452, %f4, %f41;
min.ftz.f32 %f453, %f452, %f1205;
min.ftz.f32 %f454, %f451, %f1205;
max.ftz.f32 %f455, %f454, %f35;
mul.ftz.f32 %f456, %f453, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f457, %f456;
add.ftz.f32 %f458, %f457, 0f3F800000;
div.approx.ftz.f32 %f459, %f453, %f458;
mul.ftz.f32 %f460, %f455, %f459;
st.global.f32 [%rd37], %f460;

$L__BB265_48:
ld.shared.f32 %f461, [%r33+128];
add.ftz.f32 %f462, %f461, %f29;
ld.shared.f32 %f463, [%r35+128];
add.ftz.f32 %f464, %f463, %f33;
ld.shared.f32 %f465, [%r33+640];
add.ftz.f32 %f466, %f465, %f462;
ld.shared.f32 %f467, [%r35+640];
add.ftz.f32 %f468, %f467, %f464;
ld.shared.f32 %f469, [%r33+1152];
add.ftz.f32 %f470, %f469, %f466;
ld.shared.f32 %f471, [%r35+1152];
add.ftz.f32 %f472, %f471, %f468;
mov.b32 %r1103, %f470;
shfl.sync.bfly.b32 %r1107|%p72, %r1103, %r1075, %r1074, %r1076;
mov.b32 %f473, %r1107;
add.ftz.f32 %f474, %f470, %f473;
mov.b32 %r1108, %f474;
shfl.sync.bfly.b32 %r1110|%p73, %r1108, %r1079, %r1074, %r1076;
mov.b32 %f475, %r1110;
add.ftz.f32 %f476, %f474, %f475;
mov.b32 %r1111, %f476;
shfl.sync.bfly.b32 %r1113|%p74, %r1111, %r1082, %r1074, %r1076;
mov.b32 %f477, %r1113;
add.ftz.f32 %f478, %f476, %f477;
mov.b32 %r1114, %f478;
shfl.sync.bfly.b32 %r1116|%p75, %r1114, %r1085, %r1074, %r1076;
mov.b32 %f479, %r1116;
add.ftz.f32 %f480, %f478, %f479;
mov.b32 %r1117, %f480;
shfl.sync.bfly.b32 %r1119|%p76, %r1117, %r1088, %r1074, %r1076;
mov.b32 %f481, %r1119;
add.ftz.f32 %f42, %f480, %f481;
mov.b32 %r1120, %f472;
shfl.sync.bfly.b32 %r1121|%p77, %r1120, %r1075, %r1074, %r1076;
mov.b32 %f482, %r1121;
add.ftz.f32 %f483, %f472, %f482;
mov.b32 %r1122, %f483;
shfl.sync.bfly.b32 %r1123|%p78, %r1122, %r1079, %r1074, %r1076;
mov.b32 %f484, %r1123;
add.ftz.f32 %f485, %f483, %f484;
mov.b32 %r1124, %f485;
shfl.sync.bfly.b32 %r1125|%p79, %r1124, %r1082, %r1074, %r1076;
mov.b32 %f486, %r1125;
add.ftz.f32 %f487, %f485, %f486;
mov.b32 %r1126, %f487;
shfl.sync.bfly.b32 %r1127|%p80, %r1126, %r1085, %r1074, %r1076;
mov.b32 %f488, %r1127;
add.ftz.f32 %f489, %f487, %f488;
mov.b32 %r1128, %f489;
shfl.sync.bfly.b32 %r1129|%p81, %r1128, %r1088, %r1074, %r1076;
mov.b32 %f490, %r1129;
add.ftz.f32 %f43, %f489, %f490;
setp.ne.s32 %p82, %r2, 1;
or.b32 %r1133, %r4, 1;
setp.ge.u32 %p83, %r1133, %r52;
or.pred %p84, %p82, %p83;
@%p84 bra $L__BB265_50;

ld.param.f32 %f1204, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f491, %f5, %f42;
add.ftz.f32 %f492, %f4, %f43;
min.ftz.f32 %f493, %f492, %f1204;
min.ftz.f32 %f494, %f491, %f1204;
max.ftz.f32 %f495, %f494, %f35;
mul.ftz.f32 %f496, %f493, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f497, %f496;
add.ftz.f32 %f498, %f497, 0f3F800000;
div.approx.ftz.f32 %f499, %f493, %f498;
mul.ftz.f32 %f500, %f495, %f499;
st.global.f32 [%rd37+4], %f500;

$L__BB265_50:
ld.shared.f32 %f501, [%r33+256];
add.ftz.f32 %f502, %f501, %f28;
ld.shared.f32 %f503, [%r35+256];
add.ftz.f32 %f504, %f503, %f32;
ld.shared.f32 %f505, [%r33+768];
add.ftz.f32 %f506, %f505, %f502;
ld.shared.f32 %f507, [%r35+768];
add.ftz.f32 %f508, %f507, %f504;
ld.shared.f32 %f509, [%r33+1280];
add.ftz.f32 %f510, %f509, %f506;
ld.shared.f32 %f511, [%r35+1280];
add.ftz.f32 %f512, %f511, %f508;
mov.b32 %r1134, %f510;
mov.u32 %r1135, 31;
mov.u32 %r1136, 16;
mov.u32 %r1137, -1;
shfl.sync.bfly.b32 %r1138|%p85, %r1134, %r1136, %r1135, %r1137;
mov.b32 %f513, %r1138;
add.ftz.f32 %f514, %f510, %f513;
mov.b32 %r1139, %f514;
mov.u32 %r1140, 8;
shfl.sync.bfly.b32 %r1141|%p86, %r1139, %r1140, %r1135, %r1137;
mov.b32 %f515, %r1141;
add.ftz.f32 %f516, %f514, %f515;
mov.b32 %r1142, %f516;
mov.u32 %r1143, 4;
shfl.sync.bfly.b32 %r1144|%p87, %r1142, %r1143, %r1135, %r1137;
mov.b32 %f517, %r1144;
add.ftz.f32 %f518, %f516, %f517;
mov.b32 %r1145, %f518;
mov.u32 %r1146, 2;
shfl.sync.bfly.b32 %r1147|%p88, %r1145, %r1146, %r1135, %r1137;
mov.b32 %f519, %r1147;
add.ftz.f32 %f520, %f518, %f519;
mov.b32 %r1148, %f520;
mov.u32 %r1149, 1;
shfl.sync.bfly.b32 %r1150|%p89, %r1148, %r1149, %r1135, %r1137;
mov.b32 %f521, %r1150;
add.ftz.f32 %f44, %f520, %f521;
mov.b32 %r1151, %f512;
shfl.sync.bfly.b32 %r1152|%p90, %r1151, %r1136, %r1135, %r1137;
mov.b32 %f522, %r1152;
add.ftz.f32 %f523, %f512, %f522;
mov.b32 %r1153, %f523;
shfl.sync.bfly.b32 %r1154|%p91, %r1153, %r1140, %r1135, %r1137;
mov.b32 %f524, %r1154;
add.ftz.f32 %f525, %f523, %f524;
mov.b32 %r1155, %f525;
shfl.sync.bfly.b32 %r1156|%p92, %r1155, %r1143, %r1135, %r1137;
mov.b32 %f526, %r1156;
add.ftz.f32 %f527, %f525, %f526;
mov.b32 %r1157, %f527;
shfl.sync.bfly.b32 %r1158|%p93, %r1157, %r1146, %r1135, %r1137;
mov.b32 %f528, %r1158;
add.ftz.f32 %f529, %f527, %f528;
mov.b32 %r1159, %f529;
shfl.sync.bfly.b32 %r1160|%p94, %r1159, %r1149, %r1135, %r1137;
mov.b32 %f530, %r1160;
add.ftz.f32 %f45, %f529, %f530;
setp.ne.s32 %p95, %r2, 2;
or.b32 %r1164, %r4, 2;
setp.ge.u32 %p96, %r1164, %r52;
or.pred %p97, %p95, %p96;
@%p97 bra $L__BB265_52;

ld.param.f32 %f1203, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f531, %f5, %f44;
add.ftz.f32 %f532, %f4, %f45;
min.ftz.f32 %f533, %f532, %f1203;
min.ftz.f32 %f534, %f531, %f1203;
max.ftz.f32 %f535, %f534, %f35;
mul.ftz.f32 %f536, %f533, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f537, %f536;
add.ftz.f32 %f538, %f537, 0f3F800000;
div.approx.ftz.f32 %f539, %f533, %f538;
mul.ftz.f32 %f540, %f535, %f539;
st.global.f32 [%rd37+8], %f540;

$L__BB265_52:
ld.shared.f32 %f541, [%r33+384];
add.ftz.f32 %f542, %f541, %f27;
ld.shared.f32 %f543, [%r35+384];
add.ftz.f32 %f544, %f543, %f31;
ld.shared.f32 %f545, [%r33+896];
add.ftz.f32 %f546, %f545, %f542;
ld.shared.f32 %f547, [%r35+896];
add.ftz.f32 %f548, %f547, %f544;
ld.shared.f32 %f549, [%r33+1408];
add.ftz.f32 %f550, %f549, %f546;
ld.shared.f32 %f551, [%r35+1408];
add.ftz.f32 %f552, %f551, %f548;
mov.b32 %r1165, %f550;
shfl.sync.bfly.b32 %r1169|%p98, %r1165, %r1136, %r1135, %r1137;
mov.b32 %f553, %r1169;
add.ftz.f32 %f554, %f550, %f553;
mov.b32 %r1170, %f554;
shfl.sync.bfly.b32 %r1172|%p99, %r1170, %r1140, %r1135, %r1137;
mov.b32 %f555, %r1172;
add.ftz.f32 %f556, %f554, %f555;
mov.b32 %r1173, %f556;
shfl.sync.bfly.b32 %r1175|%p100, %r1173, %r1143, %r1135, %r1137;
mov.b32 %f557, %r1175;
add.ftz.f32 %f558, %f556, %f557;
mov.b32 %r1176, %f558;
shfl.sync.bfly.b32 %r1178|%p101, %r1176, %r1146, %r1135, %r1137;
mov.b32 %f559, %r1178;
add.ftz.f32 %f560, %f558, %f559;
mov.b32 %r1179, %f560;
shfl.sync.bfly.b32 %r1181|%p102, %r1179, %r1149, %r1135, %r1137;
mov.b32 %f561, %r1181;
add.ftz.f32 %f46, %f560, %f561;
mov.b32 %r1182, %f552;
shfl.sync.bfly.b32 %r1183|%p103, %r1182, %r1136, %r1135, %r1137;
mov.b32 %f562, %r1183;
add.ftz.f32 %f563, %f552, %f562;
mov.b32 %r1184, %f563;
shfl.sync.bfly.b32 %r1185|%p104, %r1184, %r1140, %r1135, %r1137;
mov.b32 %f564, %r1185;
add.ftz.f32 %f565, %f563, %f564;
mov.b32 %r1186, %f565;
shfl.sync.bfly.b32 %r1187|%p105, %r1186, %r1143, %r1135, %r1137;
mov.b32 %f566, %r1187;
add.ftz.f32 %f567, %f565, %f566;
mov.b32 %r1188, %f567;
shfl.sync.bfly.b32 %r1189|%p106, %r1188, %r1146, %r1135, %r1137;
mov.b32 %f568, %r1189;
add.ftz.f32 %f569, %f567, %f568;
mov.b32 %r1190, %f569;
shfl.sync.bfly.b32 %r1191|%p107, %r1190, %r1149, %r1135, %r1137;
mov.b32 %f570, %r1191;
add.ftz.f32 %f47, %f569, %f570;
setp.ne.s32 %p108, %r2, 3;
or.b32 %r1195, %r4, 3;
setp.ge.u32 %p109, %r1195, %r52;
or.pred %p110, %p108, %p109;
@%p110 bra $L__BB265_78;

ld.param.f32 %f1202, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f571, %f5, %f46;
add.ftz.f32 %f572, %f4, %f47;
min.ftz.f32 %f573, %f572, %f1202;
min.ftz.f32 %f574, %f571, %f1202;
max.ftz.f32 %f575, %f574, %f35;
mul.ftz.f32 %f576, %f573, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f577, %f576;
add.ftz.f32 %f578, %f577, 0f3F800000;
div.approx.ftz.f32 %f579, %f573, %f578;
mul.ftz.f32 %f580, %f575, %f579;
st.global.f32 [%rd37+12], %f580;
bra.uni $L__BB265_78;

$L__BB265_70:
ld.shared.f32 %f1069, [%r33];
add.ftz.f32 %f1070, %f1069, %f30;
ld.shared.f32 %f1071, [%r35];
add.ftz.f32 %f1072, %f1071, %f34;
ld.shared.f32 %f1073, [%r33+512];
add.ftz.f32 %f1074, %f1073, %f1070;
ld.shared.f32 %f1075, [%r35+512];
add.ftz.f32 %f1076, %f1075, %f1072;
ld.shared.f32 %f1077, [%r33+1024];
add.ftz.f32 %f1078, %f1077, %f1074;
ld.shared.f32 %f1079, [%r35+1024];
add.ftz.f32 %f1080, %f1079, %f1076;
mov.b32 %r1565, %f1078;
mov.u32 %r1566, 31;
mov.u32 %r1567, 16;
mov.u32 %r1568, -1;
shfl.sync.bfly.b32 %r1569|%p267, %r1565, %r1567, %r1566, %r1568;
mov.b32 %f1081, %r1569;
add.ftz.f32 %f1082, %f1078, %f1081;
mov.b32 %r1570, %f1082;
mov.u32 %r1571, 8;
shfl.sync.bfly.b32 %r1572|%p268, %r1570, %r1571, %r1566, %r1568;
mov.b32 %f1083, %r1572;
add.ftz.f32 %f1084, %f1082, %f1083;
mov.b32 %r1573, %f1084;
mov.u32 %r1574, 4;
shfl.sync.bfly.b32 %r1575|%p269, %r1573, %r1574, %r1566, %r1568;
mov.b32 %f1085, %r1575;
add.ftz.f32 %f1086, %f1084, %f1085;
mov.b32 %r1576, %f1086;
mov.u32 %r1577, 2;
shfl.sync.bfly.b32 %r1578|%p270, %r1576, %r1577, %r1566, %r1568;
mov.b32 %f1087, %r1578;
add.ftz.f32 %f1088, %f1086, %f1087;
mov.b32 %r1579, %f1088;
mov.u32 %r1580, 1;
shfl.sync.bfly.b32 %r1581|%p271, %r1579, %r1580, %r1566, %r1568;
mov.b32 %f1089, %r1581;
add.ftz.f32 %f72, %f1088, %f1089;
mov.b32 %r1582, %f1080;
shfl.sync.bfly.b32 %r1583|%p272, %r1582, %r1567, %r1566, %r1568;
mov.b32 %f1090, %r1583;
add.ftz.f32 %f1091, %f1080, %f1090;
mov.b32 %r1584, %f1091;
shfl.sync.bfly.b32 %r1585|%p273, %r1584, %r1571, %r1566, %r1568;
mov.b32 %f1092, %r1585;
add.ftz.f32 %f1093, %f1091, %f1092;
mov.b32 %r1586, %f1093;
shfl.sync.bfly.b32 %r1587|%p274, %r1586, %r1574, %r1566, %r1568;
mov.b32 %f1094, %r1587;
add.ftz.f32 %f1095, %f1093, %f1094;
mov.b32 %r1588, %f1095;
shfl.sync.bfly.b32 %r1589|%p275, %r1588, %r1577, %r1566, %r1568;
mov.b32 %f1096, %r1589;
add.ftz.f32 %f1097, %f1095, %f1096;
mov.b32 %r1590, %f1097;
shfl.sync.bfly.b32 %r1591|%p276, %r1590, %r1580, %r1566, %r1568;
mov.b32 %f1098, %r1591;
add.ftz.f32 %f73, %f1097, %f1098;
setp.ne.s32 %p277, %r2, 0;
setp.ge.u32 %p278, %r4, %r52;
or.pred %p279, %p277, %p278;
@%p279 bra $L__BB265_72;

add.ftz.f32 %f1099, %f5, %f72;
add.ftz.f32 %f1100, %f4, %f73;
mul.ftz.f32 %f1101, %f1099, %f1100;
st.global.f32 [%rd37], %f1101;

$L__BB265_72:
ld.shared.f32 %f1102, [%r33+128];
add.ftz.f32 %f1103, %f1102, %f29;
ld.shared.f32 %f1104, [%r35+128];
add.ftz.f32 %f1105, %f1104, %f33;
ld.shared.f32 %f1106, [%r33+640];
add.ftz.f32 %f1107, %f1106, %f1103;
ld.shared.f32 %f1108, [%r35+640];
add.ftz.f32 %f1109, %f1108, %f1105;
ld.shared.f32 %f1110, [%r33+1152];
add.ftz.f32 %f1111, %f1110, %f1107;
ld.shared.f32 %f1112, [%r35+1152];
add.ftz.f32 %f1113, %f1112, %f1109;
mov.b32 %r1595, %f1111;
shfl.sync.bfly.b32 %r1599|%p280, %r1595, %r1567, %r1566, %r1568;
mov.b32 %f1114, %r1599;
add.ftz.f32 %f1115, %f1111, %f1114;
mov.b32 %r1600, %f1115;
shfl.sync.bfly.b32 %r1602|%p281, %r1600, %r1571, %r1566, %r1568;
mov.b32 %f1116, %r1602;
add.ftz.f32 %f1117, %f1115, %f1116;
mov.b32 %r1603, %f1117;
shfl.sync.bfly.b32 %r1605|%p282, %r1603, %r1574, %r1566, %r1568;
mov.b32 %f1118, %r1605;
add.ftz.f32 %f1119, %f1117, %f1118;
mov.b32 %r1606, %f1119;
shfl.sync.bfly.b32 %r1608|%p283, %r1606, %r1577, %r1566, %r1568;
mov.b32 %f1120, %r1608;
add.ftz.f32 %f1121, %f1119, %f1120;
mov.b32 %r1609, %f1121;
shfl.sync.bfly.b32 %r1611|%p284, %r1609, %r1580, %r1566, %r1568;
mov.b32 %f1122, %r1611;
add.ftz.f32 %f74, %f1121, %f1122;
mov.b32 %r1612, %f1113;
shfl.sync.bfly.b32 %r1613|%p285, %r1612, %r1567, %r1566, %r1568;
mov.b32 %f1123, %r1613;
add.ftz.f32 %f1124, %f1113, %f1123;
mov.b32 %r1614, %f1124;
shfl.sync.bfly.b32 %r1615|%p286, %r1614, %r1571, %r1566, %r1568;
mov.b32 %f1125, %r1615;
add.ftz.f32 %f1126, %f1124, %f1125;
mov.b32 %r1616, %f1126;
shfl.sync.bfly.b32 %r1617|%p287, %r1616, %r1574, %r1566, %r1568;
mov.b32 %f1127, %r1617;
add.ftz.f32 %f1128, %f1126, %f1127;
mov.b32 %r1618, %f1128;
shfl.sync.bfly.b32 %r1619|%p288, %r1618, %r1577, %r1566, %r1568;
mov.b32 %f1129, %r1619;
add.ftz.f32 %f1130, %f1128, %f1129;
mov.b32 %r1620, %f1130;
shfl.sync.bfly.b32 %r1621|%p289, %r1620, %r1580, %r1566, %r1568;
mov.b32 %f1131, %r1621;
add.ftz.f32 %f75, %f1130, %f1131;
setp.ne.s32 %p290, %r2, 1;
or.b32 %r1625, %r4, 1;
setp.ge.u32 %p291, %r1625, %r52;
or.pred %p292, %p290, %p291;
@%p292 bra $L__BB265_74;

add.ftz.f32 %f1132, %f5, %f74;
add.ftz.f32 %f1133, %f4, %f75;
mul.ftz.f32 %f1134, %f1132, %f1133;
st.global.f32 [%rd37+4], %f1134;

$L__BB265_74:
ld.shared.f32 %f1135, [%r33+256];
add.ftz.f32 %f1136, %f1135, %f28;
ld.shared.f32 %f1137, [%r35+256];
add.ftz.f32 %f1138, %f1137, %f32;
ld.shared.f32 %f1139, [%r33+768];
add.ftz.f32 %f1140, %f1139, %f1136;
ld.shared.f32 %f1141, [%r35+768];
add.ftz.f32 %f1142, %f1141, %f1138;
ld.shared.f32 %f1143, [%r33+1280];
add.ftz.f32 %f1144, %f1143, %f1140;
ld.shared.f32 %f1145, [%r35+1280];
add.ftz.f32 %f1146, %f1145, %f1142;
mov.b32 %r1626, %f1144;
mov.u32 %r1627, 31;
mov.u32 %r1628, 16;
mov.u32 %r1629, -1;
shfl.sync.bfly.b32 %r1630|%p293, %r1626, %r1628, %r1627, %r1629;
mov.b32 %f1147, %r1630;
add.ftz.f32 %f1148, %f1144, %f1147;
mov.b32 %r1631, %f1148;
mov.u32 %r1632, 8;
shfl.sync.bfly.b32 %r1633|%p294, %r1631, %r1632, %r1627, %r1629;
mov.b32 %f1149, %r1633;
add.ftz.f32 %f1150, %f1148, %f1149;
mov.b32 %r1634, %f1150;
mov.u32 %r1635, 4;
shfl.sync.bfly.b32 %r1636|%p295, %r1634, %r1635, %r1627, %r1629;
mov.b32 %f1151, %r1636;
add.ftz.f32 %f1152, %f1150, %f1151;
mov.b32 %r1637, %f1152;
mov.u32 %r1638, 2;
shfl.sync.bfly.b32 %r1639|%p296, %r1637, %r1638, %r1627, %r1629;
mov.b32 %f1153, %r1639;
add.ftz.f32 %f1154, %f1152, %f1153;
mov.b32 %r1640, %f1154;
mov.u32 %r1641, 1;
shfl.sync.bfly.b32 %r1642|%p297, %r1640, %r1641, %r1627, %r1629;
mov.b32 %f1155, %r1642;
add.ftz.f32 %f76, %f1154, %f1155;
mov.b32 %r1643, %f1146;
shfl.sync.bfly.b32 %r1644|%p298, %r1643, %r1628, %r1627, %r1629;
mov.b32 %f1156, %r1644;
add.ftz.f32 %f1157, %f1146, %f1156;
mov.b32 %r1645, %f1157;
shfl.sync.bfly.b32 %r1646|%p299, %r1645, %r1632, %r1627, %r1629;
mov.b32 %f1158, %r1646;
add.ftz.f32 %f1159, %f1157, %f1158;
mov.b32 %r1647, %f1159;
shfl.sync.bfly.b32 %r1648|%p300, %r1647, %r1635, %r1627, %r1629;
mov.b32 %f1160, %r1648;
add.ftz.f32 %f1161, %f1159, %f1160;
mov.b32 %r1649, %f1161;
shfl.sync.bfly.b32 %r1650|%p301, %r1649, %r1638, %r1627, %r1629;
mov.b32 %f1162, %r1650;
add.ftz.f32 %f1163, %f1161, %f1162;
mov.b32 %r1651, %f1163;
shfl.sync.bfly.b32 %r1652|%p302, %r1651, %r1641, %r1627, %r1629;
mov.b32 %f1164, %r1652;
add.ftz.f32 %f77, %f1163, %f1164;
setp.ne.s32 %p303, %r2, 2;
or.b32 %r1656, %r4, 2;
setp.ge.u32 %p304, %r1656, %r52;
or.pred %p305, %p303, %p304;
@%p305 bra $L__BB265_76;

add.ftz.f32 %f1165, %f5, %f76;
add.ftz.f32 %f1166, %f4, %f77;
mul.ftz.f32 %f1167, %f1165, %f1166;
st.global.f32 [%rd37+8], %f1167;

$L__BB265_76:
ld.shared.f32 %f1168, [%r33+384];
add.ftz.f32 %f1169, %f1168, %f27;
ld.shared.f32 %f1170, [%r35+384];
add.ftz.f32 %f1171, %f1170, %f31;
ld.shared.f32 %f1172, [%r33+896];
add.ftz.f32 %f1173, %f1172, %f1169;
ld.shared.f32 %f1174, [%r35+896];
add.ftz.f32 %f1175, %f1174, %f1171;
ld.shared.f32 %f1176, [%r33+1408];
add.ftz.f32 %f1177, %f1176, %f1173;
ld.shared.f32 %f1178, [%r35+1408];
add.ftz.f32 %f1179, %f1178, %f1175;
mov.b32 %r1657, %f1177;
shfl.sync.bfly.b32 %r1661|%p306, %r1657, %r1628, %r1627, %r1629;
mov.b32 %f1180, %r1661;
add.ftz.f32 %f1181, %f1177, %f1180;
mov.b32 %r1662, %f1181;
shfl.sync.bfly.b32 %r1664|%p307, %r1662, %r1632, %r1627, %r1629;
mov.b32 %f1182, %r1664;
add.ftz.f32 %f1183, %f1181, %f1182;
mov.b32 %r1665, %f1183;
shfl.sync.bfly.b32 %r1667|%p308, %r1665, %r1635, %r1627, %r1629;
mov.b32 %f1184, %r1667;
add.ftz.f32 %f1185, %f1183, %f1184;
mov.b32 %r1668, %f1185;
shfl.sync.bfly.b32 %r1670|%p309, %r1668, %r1638, %r1627, %r1629;
mov.b32 %f1186, %r1670;
add.ftz.f32 %f1187, %f1185, %f1186;
mov.b32 %r1671, %f1187;
shfl.sync.bfly.b32 %r1673|%p310, %r1671, %r1641, %r1627, %r1629;
mov.b32 %f1188, %r1673;
add.ftz.f32 %f78, %f1187, %f1188;
mov.b32 %r1674, %f1179;
shfl.sync.bfly.b32 %r1675|%p311, %r1674, %r1628, %r1627, %r1629;
mov.b32 %f1189, %r1675;
add.ftz.f32 %f1190, %f1179, %f1189;
mov.b32 %r1676, %f1190;
shfl.sync.bfly.b32 %r1677|%p312, %r1676, %r1632, %r1627, %r1629;
mov.b32 %f1191, %r1677;
add.ftz.f32 %f1192, %f1190, %f1191;
mov.b32 %r1678, %f1192;
shfl.sync.bfly.b32 %r1679|%p313, %r1678, %r1635, %r1627, %r1629;
mov.b32 %f1193, %r1679;
add.ftz.f32 %f1194, %f1192, %f1193;
mov.b32 %r1680, %f1194;
shfl.sync.bfly.b32 %r1681|%p314, %r1680, %r1638, %r1627, %r1629;
mov.b32 %f1195, %r1681;
add.ftz.f32 %f1196, %f1194, %f1195;
mov.b32 %r1682, %f1196;
shfl.sync.bfly.b32 %r1683|%p315, %r1682, %r1641, %r1627, %r1629;
mov.b32 %f1197, %r1683;
add.ftz.f32 %f79, %f1196, %f1197;
setp.ne.s32 %p316, %r2, 3;
or.b32 %r1687, %r4, 3;
setp.ge.u32 %p317, %r1687, %r52;
or.pred %p318, %p316, %p317;
@%p318 bra $L__BB265_78;

add.ftz.f32 %f1198, %f5, %f78;
add.ftz.f32 %f1199, %f4, %f79;
mul.ftz.f32 %f1200, %f1198, %f1199;
st.global.f32 [%rd37+12], %f1200;
bra.uni $L__BB265_78;

$L__BB265_62:
ld.shared.f32 %f905, [%r33];
add.ftz.f32 %f906, %f905, %f30;
ld.shared.f32 %f907, [%r35];
add.ftz.f32 %f908, %f907, %f34;
ld.shared.f32 %f909, [%r33+512];
add.ftz.f32 %f910, %f909, %f906;
ld.shared.f32 %f911, [%r35+512];
add.ftz.f32 %f912, %f911, %f908;
ld.shared.f32 %f913, [%r33+1024];
add.ftz.f32 %f914, %f913, %f910;
ld.shared.f32 %f915, [%r35+1024];
add.ftz.f32 %f916, %f915, %f912;
mov.b32 %r1442, %f914;
mov.u32 %r1443, 31;
mov.u32 %r1444, 16;
mov.u32 %r1445, -1;
shfl.sync.bfly.b32 %r1446|%p215, %r1442, %r1444, %r1443, %r1445;
mov.b32 %f917, %r1446;
add.ftz.f32 %f918, %f914, %f917;
mov.b32 %r1447, %f918;
mov.u32 %r1448, 8;
shfl.sync.bfly.b32 %r1449|%p216, %r1447, %r1448, %r1443, %r1445;
mov.b32 %f919, %r1449;
add.ftz.f32 %f920, %f918, %f919;
mov.b32 %r1450, %f920;
mov.u32 %r1451, 4;
shfl.sync.bfly.b32 %r1452|%p217, %r1450, %r1451, %r1443, %r1445;
mov.b32 %f921, %r1452;
add.ftz.f32 %f922, %f920, %f921;
mov.b32 %r1453, %f922;
mov.u32 %r1454, 2;
shfl.sync.bfly.b32 %r1455|%p218, %r1453, %r1454, %r1443, %r1445;
mov.b32 %f923, %r1455;
add.ftz.f32 %f924, %f922, %f923;
mov.b32 %r1456, %f924;
mov.u32 %r1457, 1;
shfl.sync.bfly.b32 %r1458|%p219, %r1456, %r1457, %r1443, %r1445;
mov.b32 %f925, %r1458;
add.ftz.f32 %f64, %f924, %f925;
mov.b32 %r1459, %f916;
shfl.sync.bfly.b32 %r1460|%p220, %r1459, %r1444, %r1443, %r1445;
mov.b32 %f926, %r1460;
add.ftz.f32 %f927, %f916, %f926;
mov.b32 %r1461, %f927;
shfl.sync.bfly.b32 %r1462|%p221, %r1461, %r1448, %r1443, %r1445;
mov.b32 %f928, %r1462;
add.ftz.f32 %f929, %f927, %f928;
mov.b32 %r1463, %f929;
shfl.sync.bfly.b32 %r1464|%p222, %r1463, %r1451, %r1443, %r1445;
mov.b32 %f930, %r1464;
add.ftz.f32 %f931, %f929, %f930;
mov.b32 %r1465, %f931;
shfl.sync.bfly.b32 %r1466|%p223, %r1465, %r1454, %r1443, %r1445;
mov.b32 %f932, %r1466;
add.ftz.f32 %f933, %f931, %f932;
mov.b32 %r1467, %f933;
shfl.sync.bfly.b32 %r1468|%p224, %r1467, %r1457, %r1443, %r1445;
mov.b32 %f934, %r1468;
add.ftz.f32 %f65, %f933, %f934;
setp.ne.s32 %p225, %r2, 0;
setp.ge.u32 %p226, %r4, %r52;
or.pred %p227, %p225, %p226;
@%p227 bra $L__BB265_64;

add.ftz.f32 %f935, %f5, %f64;
add.ftz.f32 %f936, %f4, %f65;
mul.ftz.f32 %f937, %f936, 0f3F000000;
mul.ftz.f32 %f938, %f936, 0f3F4C422A;
mul.ftz.f32 %f939, %f936, 0f3D372713;
fma.rn.ftz.f32 %f940, %f936, %f939, 0f3F800000;
mul.ftz.f32 %f941, %f938, %f940;
tanh.approx.f32 %f942, %f941;
add.ftz.f32 %f943, %f942, 0f3F800000;
mul.ftz.f32 %f944, %f937, %f943;
mul.ftz.f32 %f945, %f935, %f944;
st.global.f32 [%rd37], %f945;

$L__BB265_64:
ld.shared.f32 %f946, [%r33+128];
add.ftz.f32 %f947, %f946, %f29;
ld.shared.f32 %f948, [%r35+128];
add.ftz.f32 %f949, %f948, %f33;
ld.shared.f32 %f950, [%r33+640];
add.ftz.f32 %f951, %f950, %f947;
ld.shared.f32 %f952, [%r35+640];
add.ftz.f32 %f953, %f952, %f949;
ld.shared.f32 %f954, [%r33+1152];
add.ftz.f32 %f955, %f954, %f951;
ld.shared.f32 %f956, [%r35+1152];
add.ftz.f32 %f957, %f956, %f953;
mov.b32 %r1472, %f955;
shfl.sync.bfly.b32 %r1476|%p228, %r1472, %r1444, %r1443, %r1445;
mov.b32 %f958, %r1476;
add.ftz.f32 %f959, %f955, %f958;
mov.b32 %r1477, %f959;
shfl.sync.bfly.b32 %r1479|%p229, %r1477, %r1448, %r1443, %r1445;
mov.b32 %f960, %r1479;
add.ftz.f32 %f961, %f959, %f960;
mov.b32 %r1480, %f961;
shfl.sync.bfly.b32 %r1482|%p230, %r1480, %r1451, %r1443, %r1445;
mov.b32 %f962, %r1482;
add.ftz.f32 %f963, %f961, %f962;
mov.b32 %r1483, %f963;
shfl.sync.bfly.b32 %r1485|%p231, %r1483, %r1454, %r1443, %r1445;
mov.b32 %f964, %r1485;
add.ftz.f32 %f965, %f963, %f964;
mov.b32 %r1486, %f965;
shfl.sync.bfly.b32 %r1488|%p232, %r1486, %r1457, %r1443, %r1445;
mov.b32 %f966, %r1488;
add.ftz.f32 %f66, %f965, %f966;
mov.b32 %r1489, %f957;
shfl.sync.bfly.b32 %r1490|%p233, %r1489, %r1444, %r1443, %r1445;
mov.b32 %f967, %r1490;
add.ftz.f32 %f968, %f957, %f967;
mov.b32 %r1491, %f968;
shfl.sync.bfly.b32 %r1492|%p234, %r1491, %r1448, %r1443, %r1445;
mov.b32 %f969, %r1492;
add.ftz.f32 %f970, %f968, %f969;
mov.b32 %r1493, %f970;
shfl.sync.bfly.b32 %r1494|%p235, %r1493, %r1451, %r1443, %r1445;
mov.b32 %f971, %r1494;
add.ftz.f32 %f972, %f970, %f971;
mov.b32 %r1495, %f972;
shfl.sync.bfly.b32 %r1496|%p236, %r1495, %r1454, %r1443, %r1445;
mov.b32 %f973, %r1496;
add.ftz.f32 %f974, %f972, %f973;
mov.b32 %r1497, %f974;
shfl.sync.bfly.b32 %r1498|%p237, %r1497, %r1457, %r1443, %r1445;
mov.b32 %f975, %r1498;
add.ftz.f32 %f67, %f974, %f975;
setp.ne.s32 %p238, %r2, 1;
or.b32 %r1502, %r4, 1;
setp.ge.u32 %p239, %r1502, %r52;
or.pred %p240, %p238, %p239;
@%p240 bra $L__BB265_66;

add.ftz.f32 %f976, %f5, %f66;
add.ftz.f32 %f977, %f4, %f67;
mul.ftz.f32 %f978, %f977, 0f3F000000;
mul.ftz.f32 %f979, %f977, 0f3F4C422A;
mul.ftz.f32 %f980, %f977, 0f3D372713;
fma.rn.ftz.f32 %f981, %f977, %f980, 0f3F800000;
mul.ftz.f32 %f982, %f979, %f981;
tanh.approx.f32 %f983, %f982;
add.ftz.f32 %f984, %f983, 0f3F800000;
mul.ftz.f32 %f985, %f978, %f984;
mul.ftz.f32 %f986, %f976, %f985;
st.global.f32 [%rd37+4], %f986;

$L__BB265_66:
ld.shared.f32 %f987, [%r33+256];
add.ftz.f32 %f988, %f987, %f28;
ld.shared.f32 %f989, [%r35+256];
add.ftz.f32 %f990, %f989, %f32;
ld.shared.f32 %f991, [%r33+768];
add.ftz.f32 %f992, %f991, %f988;
ld.shared.f32 %f993, [%r35+768];
add.ftz.f32 %f994, %f993, %f990;
ld.shared.f32 %f995, [%r33+1280];
add.ftz.f32 %f996, %f995, %f992;
ld.shared.f32 %f997, [%r35+1280];
add.ftz.f32 %f998, %f997, %f994;
mov.b32 %r1503, %f996;
mov.u32 %r1504, 31;
mov.u32 %r1505, 16;
mov.u32 %r1506, -1;
shfl.sync.bfly.b32 %r1507|%p241, %r1503, %r1505, %r1504, %r1506;
mov.b32 %f999, %r1507;
add.ftz.f32 %f1000, %f996, %f999;
mov.b32 %r1508, %f1000;
mov.u32 %r1509, 8;
shfl.sync.bfly.b32 %r1510|%p242, %r1508, %r1509, %r1504, %r1506;
mov.b32 %f1001, %r1510;
add.ftz.f32 %f1002, %f1000, %f1001;
mov.b32 %r1511, %f1002;
mov.u32 %r1512, 4;
shfl.sync.bfly.b32 %r1513|%p243, %r1511, %r1512, %r1504, %r1506;
mov.b32 %f1003, %r1513;
add.ftz.f32 %f1004, %f1002, %f1003;
mov.b32 %r1514, %f1004;
mov.u32 %r1515, 2;
shfl.sync.bfly.b32 %r1516|%p244, %r1514, %r1515, %r1504, %r1506;
mov.b32 %f1005, %r1516;
add.ftz.f32 %f1006, %f1004, %f1005;
mov.b32 %r1517, %f1006;
mov.u32 %r1518, 1;
shfl.sync.bfly.b32 %r1519|%p245, %r1517, %r1518, %r1504, %r1506;
mov.b32 %f1007, %r1519;
add.ftz.f32 %f68, %f1006, %f1007;
mov.b32 %r1520, %f998;
shfl.sync.bfly.b32 %r1521|%p246, %r1520, %r1505, %r1504, %r1506;
mov.b32 %f1008, %r1521;
add.ftz.f32 %f1009, %f998, %f1008;
mov.b32 %r1522, %f1009;
shfl.sync.bfly.b32 %r1523|%p247, %r1522, %r1509, %r1504, %r1506;
mov.b32 %f1010, %r1523;
add.ftz.f32 %f1011, %f1009, %f1010;
mov.b32 %r1524, %f1011;
shfl.sync.bfly.b32 %r1525|%p248, %r1524, %r1512, %r1504, %r1506;
mov.b32 %f1012, %r1525;
add.ftz.f32 %f1013, %f1011, %f1012;
mov.b32 %r1526, %f1013;
shfl.sync.bfly.b32 %r1527|%p249, %r1526, %r1515, %r1504, %r1506;
mov.b32 %f1014, %r1527;
add.ftz.f32 %f1015, %f1013, %f1014;
mov.b32 %r1528, %f1015;
shfl.sync.bfly.b32 %r1529|%p250, %r1528, %r1518, %r1504, %r1506;
mov.b32 %f1016, %r1529;
add.ftz.f32 %f69, %f1015, %f1016;
setp.ne.s32 %p251, %r2, 2;
or.b32 %r1533, %r4, 2;
setp.ge.u32 %p252, %r1533, %r52;
or.pred %p253, %p251, %p252;
@%p253 bra $L__BB265_68;

add.ftz.f32 %f1017, %f5, %f68;
add.ftz.f32 %f1018, %f4, %f69;
mul.ftz.f32 %f1019, %f1018, 0f3F000000;
mul.ftz.f32 %f1020, %f1018, 0f3F4C422A;
mul.ftz.f32 %f1021, %f1018, 0f3D372713;
fma.rn.ftz.f32 %f1022, %f1018, %f1021, 0f3F800000;
mul.ftz.f32 %f1023, %f1020, %f1022;
tanh.approx.f32 %f1024, %f1023;
add.ftz.f32 %f1025, %f1024, 0f3F800000;
mul.ftz.f32 %f1026, %f1019, %f1025;
mul.ftz.f32 %f1027, %f1017, %f1026;
st.global.f32 [%rd37+8], %f1027;

$L__BB265_68:
ld.shared.f32 %f1028, [%r33+384];
add.ftz.f32 %f1029, %f1028, %f27;
ld.shared.f32 %f1030, [%r35+384];
add.ftz.f32 %f1031, %f1030, %f31;
ld.shared.f32 %f1032, [%r33+896];
add.ftz.f32 %f1033, %f1032, %f1029;
ld.shared.f32 %f1034, [%r35+896];
add.ftz.f32 %f1035, %f1034, %f1031;
ld.shared.f32 %f1036, [%r33+1408];
add.ftz.f32 %f1037, %f1036, %f1033;
ld.shared.f32 %f1038, [%r35+1408];
add.ftz.f32 %f1039, %f1038, %f1035;
mov.b32 %r1534, %f1037;
shfl.sync.bfly.b32 %r1538|%p254, %r1534, %r1505, %r1504, %r1506;
mov.b32 %f1040, %r1538;
add.ftz.f32 %f1041, %f1037, %f1040;
mov.b32 %r1539, %f1041;
shfl.sync.bfly.b32 %r1541|%p255, %r1539, %r1509, %r1504, %r1506;
mov.b32 %f1042, %r1541;
add.ftz.f32 %f1043, %f1041, %f1042;
mov.b32 %r1542, %f1043;
shfl.sync.bfly.b32 %r1544|%p256, %r1542, %r1512, %r1504, %r1506;
mov.b32 %f1044, %r1544;
add.ftz.f32 %f1045, %f1043, %f1044;
mov.b32 %r1545, %f1045;
shfl.sync.bfly.b32 %r1547|%p257, %r1545, %r1515, %r1504, %r1506;
mov.b32 %f1046, %r1547;
add.ftz.f32 %f1047, %f1045, %f1046;
mov.b32 %r1548, %f1047;
shfl.sync.bfly.b32 %r1550|%p258, %r1548, %r1518, %r1504, %r1506;
mov.b32 %f1048, %r1550;
add.ftz.f32 %f70, %f1047, %f1048;
mov.b32 %r1551, %f1039;
shfl.sync.bfly.b32 %r1552|%p259, %r1551, %r1505, %r1504, %r1506;
mov.b32 %f1049, %r1552;
add.ftz.f32 %f1050, %f1039, %f1049;
mov.b32 %r1553, %f1050;
shfl.sync.bfly.b32 %r1554|%p260, %r1553, %r1509, %r1504, %r1506;
mov.b32 %f1051, %r1554;
add.ftz.f32 %f1052, %f1050, %f1051;
mov.b32 %r1555, %f1052;
shfl.sync.bfly.b32 %r1556|%p261, %r1555, %r1512, %r1504, %r1506;
mov.b32 %f1053, %r1556;
add.ftz.f32 %f1054, %f1052, %f1053;
mov.b32 %r1557, %f1054;
shfl.sync.bfly.b32 %r1558|%p262, %r1557, %r1515, %r1504, %r1506;
mov.b32 %f1055, %r1558;
add.ftz.f32 %f1056, %f1054, %f1055;
mov.b32 %r1559, %f1056;
shfl.sync.bfly.b32 %r1560|%p263, %r1559, %r1518, %r1504, %r1506;
mov.b32 %f1057, %r1560;
add.ftz.f32 %f71, %f1056, %f1057;
setp.ne.s32 %p264, %r2, 3;
or.b32 %r1564, %r4, 3;
setp.ge.u32 %p265, %r1564, %r52;
or.pred %p266, %p264, %p265;
@%p266 bra $L__BB265_78;

add.ftz.f32 %f1058, %f5, %f70;
add.ftz.f32 %f1059, %f4, %f71;
mul.ftz.f32 %f1060, %f1059, 0f3F000000;
mul.ftz.f32 %f1061, %f1059, 0f3F4C422A;
mul.ftz.f32 %f1062, %f1059, 0f3D372713;
fma.rn.ftz.f32 %f1063, %f1059, %f1062, 0f3F800000;
mul.ftz.f32 %f1064, %f1061, %f1063;
tanh.approx.f32 %f1065, %f1064;
add.ftz.f32 %f1066, %f1065, 0f3F800000;
mul.ftz.f32 %f1067, %f1060, %f1066;
mul.ftz.f32 %f1068, %f1058, %f1067;
st.global.f32 [%rd37+12], %f1068;
bra.uni $L__BB265_78;

$L__BB265_54:
ld.shared.f32 %f581, [%r33];
add.ftz.f32 %f582, %f581, %f30;
ld.shared.f32 %f583, [%r35];
add.ftz.f32 %f584, %f583, %f34;
ld.shared.f32 %f585, [%r33+512];
add.ftz.f32 %f586, %f585, %f582;
ld.shared.f32 %f587, [%r35+512];
add.ftz.f32 %f588, %f587, %f584;
ld.shared.f32 %f589, [%r33+1024];
add.ftz.f32 %f590, %f589, %f586;
ld.shared.f32 %f591, [%r35+1024];
add.ftz.f32 %f592, %f591, %f588;
mov.b32 %r1196, %f590;
mov.u32 %r1197, 31;
mov.u32 %r1198, 16;
mov.u32 %r1199, -1;
shfl.sync.bfly.b32 %r1200|%p111, %r1196, %r1198, %r1197, %r1199;
mov.b32 %f593, %r1200;
add.ftz.f32 %f594, %f590, %f593;
mov.b32 %r1201, %f594;
mov.u32 %r1202, 8;
shfl.sync.bfly.b32 %r1203|%p112, %r1201, %r1202, %r1197, %r1199;
mov.b32 %f595, %r1203;
add.ftz.f32 %f596, %f594, %f595;
mov.b32 %r1204, %f596;
mov.u32 %r1205, 4;
shfl.sync.bfly.b32 %r1206|%p113, %r1204, %r1205, %r1197, %r1199;
mov.b32 %f597, %r1206;
add.ftz.f32 %f598, %f596, %f597;
mov.b32 %r1207, %f598;
mov.u32 %r1208, 2;
shfl.sync.bfly.b32 %r1209|%p114, %r1207, %r1208, %r1197, %r1199;
mov.b32 %f599, %r1209;
add.ftz.f32 %f600, %f598, %f599;
mov.b32 %r1210, %f600;
mov.u32 %r1211, 1;
shfl.sync.bfly.b32 %r1212|%p115, %r1210, %r1211, %r1197, %r1199;
mov.b32 %f601, %r1212;
add.ftz.f32 %f48, %f600, %f601;
mov.b32 %r1213, %f592;
shfl.sync.bfly.b32 %r1214|%p116, %r1213, %r1198, %r1197, %r1199;
mov.b32 %f602, %r1214;
add.ftz.f32 %f603, %f592, %f602;
mov.b32 %r1215, %f603;
shfl.sync.bfly.b32 %r1216|%p117, %r1215, %r1202, %r1197, %r1199;
mov.b32 %f604, %r1216;
add.ftz.f32 %f605, %f603, %f604;
mov.b32 %r1217, %f605;
shfl.sync.bfly.b32 %r1218|%p118, %r1217, %r1205, %r1197, %r1199;
mov.b32 %f606, %r1218;
add.ftz.f32 %f607, %f605, %f606;
mov.b32 %r1219, %f607;
shfl.sync.bfly.b32 %r1220|%p119, %r1219, %r1208, %r1197, %r1199;
mov.b32 %f608, %r1220;
add.ftz.f32 %f609, %f607, %f608;
mov.b32 %r1221, %f609;
shfl.sync.bfly.b32 %r1222|%p120, %r1221, %r1211, %r1197, %r1199;
mov.b32 %f610, %r1222;
add.ftz.f32 %f49, %f609, %f610;
setp.ne.s32 %p121, %r2, 0;
setp.ge.u32 %p122, %r4, %r52;
or.pred %p123, %p121, %p122;
@%p123 bra $L__BB265_56;

add.ftz.f32 %f611, %f5, %f48;
add.ftz.f32 %f612, %f4, %f49;
mov.f32 %f613, 0f40E00000;
min.ftz.f32 %f614, %f612, %f613;
min.ftz.f32 %f615, %f611, %f613;
mov.f32 %f616, 0fC0E00000;
max.ftz.f32 %f617, %f615, %f616;
mul.ftz.f32 %f618, %f614, 0f3FD9DB23;
mul.ftz.f32 %f619, %f618, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f620, %f619;
add.ftz.f32 %f621, %f620, 0f3F800000;
div.approx.ftz.f32 %f622, %f614, %f621;
add.ftz.f32 %f623, %f617, 0f3F800000;
mul.ftz.f32 %f624, %f623, %f622;
st.global.f32 [%rd37], %f624;

$L__BB265_56:
ld.shared.f32 %f625, [%r33+128];
add.ftz.f32 %f626, %f625, %f29;
ld.shared.f32 %f627, [%r35+128];
add.ftz.f32 %f628, %f627, %f33;
ld.shared.f32 %f629, [%r33+640];
add.ftz.f32 %f630, %f629, %f626;
ld.shared.f32 %f631, [%r35+640];
add.ftz.f32 %f632, %f631, %f628;
ld.shared.f32 %f633, [%r33+1152];
add.ftz.f32 %f634, %f633, %f630;
ld.shared.f32 %f635, [%r35+1152];
add.ftz.f32 %f636, %f635, %f632;
mov.b32 %r1226, %f634;
shfl.sync.bfly.b32 %r1230|%p124, %r1226, %r1198, %r1197, %r1199;
mov.b32 %f637, %r1230;
add.ftz.f32 %f638, %f634, %f637;
mov.b32 %r1231, %f638;
shfl.sync.bfly.b32 %r1233|%p125, %r1231, %r1202, %r1197, %r1199;
mov.b32 %f639, %r1233;
add.ftz.f32 %f640, %f638, %f639;
mov.b32 %r1234, %f640;
shfl.sync.bfly.b32 %r1236|%p126, %r1234, %r1205, %r1197, %r1199;
mov.b32 %f641, %r1236;
add.ftz.f32 %f642, %f640, %f641;
mov.b32 %r1237, %f642;
shfl.sync.bfly.b32 %r1239|%p127, %r1237, %r1208, %r1197, %r1199;
mov.b32 %f643, %r1239;
add.ftz.f32 %f644, %f642, %f643;
mov.b32 %r1240, %f644;
shfl.sync.bfly.b32 %r1242|%p128, %r1240, %r1211, %r1197, %r1199;
mov.b32 %f645, %r1242;
add.ftz.f32 %f50, %f644, %f645;
mov.b32 %r1243, %f636;
shfl.sync.bfly.b32 %r1244|%p129, %r1243, %r1198, %r1197, %r1199;
mov.b32 %f646, %r1244;
add.ftz.f32 %f647, %f636, %f646;
mov.b32 %r1245, %f647;
shfl.sync.bfly.b32 %r1246|%p130, %r1245, %r1202, %r1197, %r1199;
mov.b32 %f648, %r1246;
add.ftz.f32 %f649, %f647, %f648;
mov.b32 %r1247, %f649;
shfl.sync.bfly.b32 %r1248|%p131, %r1247, %r1205, %r1197, %r1199;
mov.b32 %f650, %r1248;
add.ftz.f32 %f651, %f649, %f650;
mov.b32 %r1249, %f651;
shfl.sync.bfly.b32 %r1250|%p132, %r1249, %r1208, %r1197, %r1199;
mov.b32 %f652, %r1250;
add.ftz.f32 %f653, %f651, %f652;
mov.b32 %r1251, %f653;
shfl.sync.bfly.b32 %r1252|%p133, %r1251, %r1211, %r1197, %r1199;
mov.b32 %f654, %r1252;
add.ftz.f32 %f51, %f653, %f654;
setp.ne.s32 %p134, %r2, 1;
or.b32 %r1256, %r4, 1;
setp.ge.u32 %p135, %r1256, %r52;
or.pred %p136, %p134, %p135;
@%p136 bra $L__BB265_58;

add.ftz.f32 %f655, %f5, %f50;
add.ftz.f32 %f656, %f4, %f51;
mov.f32 %f657, 0f40E00000;
min.ftz.f32 %f658, %f656, %f657;
min.ftz.f32 %f659, %f655, %f657;
mov.f32 %f660, 0fC0E00000;
max.ftz.f32 %f661, %f659, %f660;
mul.ftz.f32 %f662, %f658, 0f3FD9DB23;
mul.ftz.f32 %f663, %f662, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f664, %f663;
add.ftz.f32 %f665, %f664, 0f3F800000;
div.approx.ftz.f32 %f666, %f658, %f665;
add.ftz.f32 %f667, %f661, 0f3F800000;
mul.ftz.f32 %f668, %f667, %f666;
st.global.f32 [%rd37+4], %f668;

$L__BB265_58:
ld.shared.f32 %f669, [%r33+256];
add.ftz.f32 %f670, %f669, %f28;
ld.shared.f32 %f671, [%r35+256];
add.ftz.f32 %f672, %f671, %f32;
ld.shared.f32 %f673, [%r33+768];
add.ftz.f32 %f674, %f673, %f670;
ld.shared.f32 %f675, [%r35+768];
add.ftz.f32 %f676, %f675, %f672;
ld.shared.f32 %f677, [%r33+1280];
add.ftz.f32 %f678, %f677, %f674;
ld.shared.f32 %f679, [%r35+1280];
add.ftz.f32 %f680, %f679, %f676;
mov.b32 %r1257, %f678;
mov.u32 %r1258, 31;
mov.u32 %r1259, 16;
mov.u32 %r1260, -1;
shfl.sync.bfly.b32 %r1261|%p137, %r1257, %r1259, %r1258, %r1260;
mov.b32 %f681, %r1261;
add.ftz.f32 %f682, %f678, %f681;
mov.b32 %r1262, %f682;
mov.u32 %r1263, 8;
shfl.sync.bfly.b32 %r1264|%p138, %r1262, %r1263, %r1258, %r1260;
mov.b32 %f683, %r1264;
add.ftz.f32 %f684, %f682, %f683;
mov.b32 %r1265, %f684;
mov.u32 %r1266, 4;
shfl.sync.bfly.b32 %r1267|%p139, %r1265, %r1266, %r1258, %r1260;
mov.b32 %f685, %r1267;
add.ftz.f32 %f686, %f684, %f685;
mov.b32 %r1268, %f686;
mov.u32 %r1269, 2;
shfl.sync.bfly.b32 %r1270|%p140, %r1268, %r1269, %r1258, %r1260;
mov.b32 %f687, %r1270;
add.ftz.f32 %f688, %f686, %f687;
mov.b32 %r1271, %f688;
mov.u32 %r1272, 1;
shfl.sync.bfly.b32 %r1273|%p141, %r1271, %r1272, %r1258, %r1260;
mov.b32 %f689, %r1273;
add.ftz.f32 %f52, %f688, %f689;
mov.b32 %r1274, %f680;
shfl.sync.bfly.b32 %r1275|%p142, %r1274, %r1259, %r1258, %r1260;
mov.b32 %f690, %r1275;
add.ftz.f32 %f691, %f680, %f690;
mov.b32 %r1276, %f691;
shfl.sync.bfly.b32 %r1277|%p143, %r1276, %r1263, %r1258, %r1260;
mov.b32 %f692, %r1277;
add.ftz.f32 %f693, %f691, %f692;
mov.b32 %r1278, %f693;
shfl.sync.bfly.b32 %r1279|%p144, %r1278, %r1266, %r1258, %r1260;
mov.b32 %f694, %r1279;
add.ftz.f32 %f695, %f693, %f694;
mov.b32 %r1280, %f695;
shfl.sync.bfly.b32 %r1281|%p145, %r1280, %r1269, %r1258, %r1260;
mov.b32 %f696, %r1281;
add.ftz.f32 %f697, %f695, %f696;
mov.b32 %r1282, %f697;
shfl.sync.bfly.b32 %r1283|%p146, %r1282, %r1272, %r1258, %r1260;
mov.b32 %f698, %r1283;
add.ftz.f32 %f53, %f697, %f698;
setp.ne.s32 %p147, %r2, 2;
or.b32 %r1287, %r4, 2;
setp.ge.u32 %p148, %r1287, %r52;
or.pred %p149, %p147, %p148;
@%p149 bra $L__BB265_60;

add.ftz.f32 %f699, %f5, %f52;
add.ftz.f32 %f700, %f4, %f53;
mov.f32 %f701, 0f40E00000;
min.ftz.f32 %f702, %f700, %f701;
min.ftz.f32 %f703, %f699, %f701;
mov.f32 %f704, 0fC0E00000;
max.ftz.f32 %f705, %f703, %f704;
mul.ftz.f32 %f706, %f702, 0f3FD9DB23;
mul.ftz.f32 %f707, %f706, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f708, %f707;
add.ftz.f32 %f709, %f708, 0f3F800000;
div.approx.ftz.f32 %f710, %f702, %f709;
add.ftz.f32 %f711, %f705, 0f3F800000;
mul.ftz.f32 %f712, %f711, %f710;
st.global.f32 [%rd37+8], %f712;

$L__BB265_60:
ld.shared.f32 %f713, [%r33+384];
add.ftz.f32 %f714, %f713, %f27;
ld.shared.f32 %f715, [%r35+384];
add.ftz.f32 %f716, %f715, %f31;
ld.shared.f32 %f717, [%r33+896];
add.ftz.f32 %f718, %f717, %f714;
ld.shared.f32 %f719, [%r35+896];
add.ftz.f32 %f720, %f719, %f716;
ld.shared.f32 %f721, [%r33+1408];
add.ftz.f32 %f722, %f721, %f718;
ld.shared.f32 %f723, [%r35+1408];
add.ftz.f32 %f724, %f723, %f720;
mov.b32 %r1288, %f722;
shfl.sync.bfly.b32 %r1292|%p150, %r1288, %r1259, %r1258, %r1260;
mov.b32 %f725, %r1292;
add.ftz.f32 %f726, %f722, %f725;
mov.b32 %r1293, %f726;
shfl.sync.bfly.b32 %r1295|%p151, %r1293, %r1263, %r1258, %r1260;
mov.b32 %f727, %r1295;
add.ftz.f32 %f728, %f726, %f727;
mov.b32 %r1296, %f728;
shfl.sync.bfly.b32 %r1298|%p152, %r1296, %r1266, %r1258, %r1260;
mov.b32 %f729, %r1298;
add.ftz.f32 %f730, %f728, %f729;
mov.b32 %r1299, %f730;
shfl.sync.bfly.b32 %r1301|%p153, %r1299, %r1269, %r1258, %r1260;
mov.b32 %f731, %r1301;
add.ftz.f32 %f732, %f730, %f731;
mov.b32 %r1302, %f732;
shfl.sync.bfly.b32 %r1304|%p154, %r1302, %r1272, %r1258, %r1260;
mov.b32 %f733, %r1304;
add.ftz.f32 %f54, %f732, %f733;
mov.b32 %r1305, %f724;
shfl.sync.bfly.b32 %r1306|%p155, %r1305, %r1259, %r1258, %r1260;
mov.b32 %f734, %r1306;
add.ftz.f32 %f735, %f724, %f734;
mov.b32 %r1307, %f735;
shfl.sync.bfly.b32 %r1308|%p156, %r1307, %r1263, %r1258, %r1260;
mov.b32 %f736, %r1308;
add.ftz.f32 %f737, %f735, %f736;
mov.b32 %r1309, %f737;
shfl.sync.bfly.b32 %r1310|%p157, %r1309, %r1266, %r1258, %r1260;
mov.b32 %f738, %r1310;
add.ftz.f32 %f739, %f737, %f738;
mov.b32 %r1311, %f739;
shfl.sync.bfly.b32 %r1312|%p158, %r1311, %r1269, %r1258, %r1260;
mov.b32 %f740, %r1312;
add.ftz.f32 %f741, %f739, %f740;
mov.b32 %r1313, %f741;
shfl.sync.bfly.b32 %r1314|%p159, %r1313, %r1272, %r1258, %r1260;
mov.b32 %f742, %r1314;
add.ftz.f32 %f55, %f741, %f742;
setp.ne.s32 %p160, %r2, 3;
or.b32 %r1318, %r4, 3;
setp.ge.u32 %p161, %r1318, %r52;
or.pred %p162, %p160, %p161;
@%p162 bra $L__BB265_78;

add.ftz.f32 %f743, %f5, %f54;
add.ftz.f32 %f744, %f4, %f55;
mov.f32 %f745, 0f40E00000;
min.ftz.f32 %f746, %f744, %f745;
min.ftz.f32 %f747, %f743, %f745;
mov.f32 %f748, 0fC0E00000;
max.ftz.f32 %f749, %f747, %f748;
mul.ftz.f32 %f750, %f746, 0f3FD9DB23;
mul.ftz.f32 %f751, %f750, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f752, %f751;
add.ftz.f32 %f753, %f752, 0f3F800000;
div.approx.ftz.f32 %f754, %f746, %f753;
add.ftz.f32 %f755, %f749, 0f3F800000;
mul.ftz.f32 %f756, %f755, %f754;
st.global.f32 [%rd37+12], %f756;
bra.uni $L__BB265_78;

}
.entry _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<38>;
.reg .b16 %rs<81>;
.reg .f32 %f<218>;
.reg .b32 %r<576>;
.reg .b64 %rd<86>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

ld.param.u64 %rd13, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd14, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
mov.u32 %r1, %tid.y;
shl.b32 %r39, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r39, %r2;
mov.u32 %r40, %ctaid.x;
shl.b32 %r4, %r40, 2;
shr.u32 %r5, %r38, 8;
mov.u32 %r573, %ctaid.y;
setp.eq.s64 %p1, %rd15, 0;
@%p1 bra $L__BB266_2;

cvta.to.global.u64 %rd17, %rd15;
mul.wide.u32 %rd18, %r573, 4;
add.s64 %rd19, %rd17, %rd18;
ld.global.nc.u32 %r572, [%rd19];
mul.hi.u32 %r41, %r573, %r21;
add.s32 %r42, %r41, %r573;
shr.u32 %r43, %r42, %r22;
mul.lo.s32 %r44, %r43, %r23;
sub.s32 %r573, %r573, %r44;
bra.uni $L__BB266_3;

$L__BB266_2:
mul.hi.u32 %r45, %r573, %r26;
add.s32 %r46, %r45, %r573;
shr.u32 %r572, %r46, %r27;

$L__BB266_3:
shr.s32 %r47, %r3, 31;
shr.u32 %r48, %r47, 29;
add.s32 %r49, %r3, %r48;
shr.s32 %r575, %r49, 3;
setp.ge.s32 %p2, %r575, %r5;
mov.f32 %f214, 0f00000000;
mov.f32 %f215, %f214;
mov.f32 %f216, %f214;
mov.f32 %f217, %f214;
@%p2 bra $L__BB266_6;

mov.u32 %r50, %ctaid.z;
and.b32 %r54, %r49, -8;
sub.s32 %r55, %r3, %r54;
shl.b32 %r56, %r55, 1;
cvt.s64.s32 %rd20, %r56;
cvt.s64.s32 %rd1, %r55;
shr.u32 %r57, %r55, 31;
add.s32 %r58, %r55, %r57;
shr.s32 %r59, %r58, 1;
and.b32 %r60, %r58, -2;
sub.s32 %r61, %r55, %r60;
mul.lo.s32 %r13, %r61, 6;
shl.b32 %r574, %r575, 3;
mad.lo.s32 %r62, %r4, %r24, %r575;
mul.hi.u32 %r63, %r50, %r32;
add.s32 %r64, %r63, %r50;
shr.u32 %r65, %r64, %r33;
mad.lo.s32 %r66, %r65, %r35, %r62;
mad.lo.s32 %r67, %r572, %r29, %r66;
mul.wide.s32 %rd2, %r67, 56;
mul.wide.s32 %rd21, %r55, 4;
add.s64 %rd3, %rd2, %rd21;
mul.wide.s32 %rd22, %r59, 2;
add.s64 %rd23, %rd2, %rd22;
add.s64 %rd4, %rd23, 48;
add.s64 %rd24, %rd2, %rd20;
add.s64 %rd5, %rd24, 33;
mul.wide.s32 %rd6, %r24, 56;
cvta.to.global.u64 %rd7, %rd14;
cvta.to.global.u64 %rd85, %rd13;
mul.lo.s32 %r68, %r573, %r30;
cvt.u64.u32 %rd25, %r68;
mul.lo.s32 %r69, %r50, %r36;
cvt.u64.u32 %rd26, %r69;
add.s64 %rd9, %rd25, %rd26;
mov.f32 %f214, 0f00000000;

$L__BB266_5:
cvt.u64.u32 %rd27, %r574;
add.s64 %rd28, %rd9, %rd27;
add.s64 %rd29, %rd85, %rd3;
ld.global.nc.u32 %r330, [%rd29];
shr.u32 %r331, %r330, 8;
shr.u32 %r332, %r330, 16;
shr.u32 %r333, %r330, 24;
add.s64 %rd30, %rd28, %rd1;
mul.lo.s64 %rd31, %rd30, 36;
add.s64 %rd32, %rd85, %rd5;
add.s64 %rd33, %rd32, -1;
ld.global.nc.u8 %rs5, [%rd32+-1];
and.b16 %rs6, %rs5, 240;
cvt.u32.u16 %r334, %rs5;
and.b32 %r335, %r330, 255;
and.b32 %r336, %r334, 7;
bfi.b32 %r337, %r336, %r335, 8, 3;
mul.wide.u32 %rd34, %r337, 4;
mov.u64 %rd35, iq1s_grid_gpu;
add.s64 %rd36, %rd35, %rd34;
ld.global.nc.u32 %r338, [%rd36];
and.b32 %r71, %r338, 252645135;
shr.u32 %r339, %r338, 4;
and.b32 %r75, %r339, 252645135;
add.s64 %rd37, %rd7, %rd31;
ld.global.nc.u32 %r72, [%rd37+4];
ld.global.nc.u32 %r76, [%rd37+8];
mov.u32 %r324, 0;

dp4a.s32.s32 %r70, %r71, %r72, %r324;


dp4a.s32.s32 %r74, %r75, %r76, %r70;

and.b32 %r340, %r334, 8;
cvt.rn.f32.s32 %f33, %r340;
mul.ftz.f32 %f34, %f33, 0f3D000000;
mov.f32 %f35, 0fBF600000;
sub.ftz.f32 %f36, %f35, %f34;
mov.u32 %r327, 16843009;

dp4a.s32.s32 %r78, %r72, %r327, %r324;


dp4a.s32.s32 %r82, %r76, %r327, %r78;

cvt.rn.f32.s32 %f37, %r82;
fma.rn.ftz.f32 %f38, %f36, %f37, 0f00000000;
shr.u16 %rs7, %rs6, 4;
cvt.u32.u16 %r341, %rs7;
and.b32 %r342, %r331, 255;
and.b32 %r343, %r341, 7;
bfi.b32 %r344, %r343, %r342, 8, 3;
mul.wide.u32 %rd38, %r344, 4;
add.s64 %rd39, %rd35, %rd38;
ld.global.nc.u32 %r345, [%rd39];
and.b32 %r87, %r345, 252645135;
shr.u32 %r346, %r345, 4;
and.b32 %r91, %r346, 252645135;
ld.global.nc.u32 %r88, [%rd37+12];
ld.global.nc.u32 %r92, [%rd37+16];

dp4a.s32.s32 %r86, %r87, %r88, %r74;


dp4a.s32.s32 %r90, %r91, %r92, %r86;

and.b32 %r347, %r341, 8;
cvt.rn.f32.s32 %f39, %r347;
mul.ftz.f32 %f40, %f39, 0f3D000000;
sub.ftz.f32 %f41, %f35, %f40;

dp4a.s32.s32 %r94, %r88, %r327, %r324;


dp4a.s32.s32 %r98, %r92, %r327, %r94;

cvt.rn.f32.s32 %f42, %r98;
fma.rn.ftz.f32 %f43, %f41, %f42, %f38;
ld.global.nc.u8 %rs8, [%rd32];
and.b16 %rs9, %rs8, 240;
cvt.u32.u16 %r348, %rs8;
and.b32 %r349, %r332, 255;
and.b32 %r350, %r348, 7;
bfi.b32 %r351, %r350, %r349, 8, 3;
mul.wide.u32 %rd40, %r351, 4;
add.s64 %rd41, %rd35, %rd40;
ld.global.nc.u32 %r352, [%rd41];
and.b32 %r103, %r352, 252645135;
shr.u32 %r353, %r352, 4;
and.b32 %r107, %r353, 252645135;
ld.global.nc.u32 %r104, [%rd37+20];
ld.global.nc.u32 %r108, [%rd37+24];

dp4a.s32.s32 %r102, %r103, %r104, %r324;


dp4a.s32.s32 %r106, %r107, %r108, %r102;

and.b32 %r354, %r348, 8;
cvt.rn.f32.s32 %f44, %r354;
mul.ftz.f32 %f45, %f44, 0f3D000000;
sub.ftz.f32 %f46, %f35, %f45;

dp4a.s32.s32 %r110, %r104, %r327, %r324;


dp4a.s32.s32 %r114, %r108, %r327, %r110;

cvt.rn.f32.s32 %f47, %r114;
fma.rn.ftz.f32 %f48, %f46, %f47, 0f00000000;
shr.u16 %rs10, %rs9, 4;
cvt.u32.u16 %r355, %rs10;
mul.wide.u16 %r356, %rs10, 256;
and.b32 %r357, %r356, 1792;
or.b32 %r358, %r357, %r333;
mul.wide.u32 %rd42, %r358, 4;
add.s64 %rd43, %rd35, %rd42;
ld.global.nc.u32 %r359, [%rd43];
and.b32 %r119, %r359, 252645135;
shr.u32 %r360, %r359, 4;
and.b32 %r123, %r360, 252645135;
ld.global.nc.u32 %r120, [%rd37+28];
ld.global.nc.u32 %r124, [%rd37+32];

dp4a.s32.s32 %r118, %r119, %r120, %r106;


dp4a.s32.s32 %r122, %r123, %r124, %r118;

and.b32 %r361, %r355, 8;
cvt.rn.f32.s32 %f49, %r361;
mul.ftz.f32 %f50, %f49, 0f3D000000;
sub.ftz.f32 %f51, %f35, %f50;

dp4a.s32.s32 %r126, %r120, %r327, %r324;


dp4a.s32.s32 %r130, %r124, %r327, %r126;

cvt.rn.f32.s32 %f52, %r130;
fma.rn.ftz.f32 %f53, %f51, %f52, %f48;
add.s64 %rd44, %rd85, %rd2;
add.s64 %rd45, %rd44, 48;
ld.global.nc.u16 %rs11, [%rd44+48];
shr.u16 %rs12, %rs11, 12;
ld.global.nc.u16 %rs13, [%rd44+50];
shr.u16 %rs14, %rs13, 8;
and.b16 %rs15, %rs14, 240;
or.b16 %rs16, %rs15, %rs12;
ld.global.nc.u16 %rs17, [%rd44+52];
shr.u16 %rs18, %rs17, 4;
and.b16 %rs19, %rs18, 3840;
or.b16 %rs20, %rs16, %rs19;
ld.global.nc.u16 %rs21, [%rd44+54];
and.b16 %rs22, %rs21, -4096;
or.b16 %rs1, %rs20, %rs22;

{ cvt.f32.f16 %f25, %rs1;}


ld.global.nc.u32 %r134, [%rd37];

{.reg .f16 low,high;
mov.b32 {low,high},%r134;
cvt.f32.f16 %f26, low;}


mul.ftz.f32 %f54, %f25, %f26;
add.s64 %rd46, %rd85, %rd4;
ld.global.nc.u16 %rs23, [%rd46];
cvt.u32.u16 %r362, %rs23;
shr.u32 %r363, %r362, %r13;
shl.b32 %r364, %r363, 1;
and.b32 %r365, %r364, 14;
or.b32 %r366, %r365, 1;
shr.u32 %r367, %r363, 2;
and.b32 %r368, %r367, 14;
or.b32 %r369, %r368, 1;
cvt.rn.f32.s32 %f55, %r90;
add.ftz.f32 %f56, %f43, %f55;
cvt.rn.f32.s32 %f57, %r366;
cvt.rn.f32.s32 %f58, %r122;
add.ftz.f32 %f59, %f53, %f58;
cvt.rn.f32.s32 %f60, %r369;
mul.ftz.f32 %f61, %f59, %f60;
fma.rn.ftz.f32 %f62, %f56, %f57, %f61;
fma.rn.ftz.f32 %f217, %f54, %f62, %f217;
add.s64 %rd47, %rd29, %rd6;
ld.global.nc.u32 %r370, [%rd47];
shr.u32 %r371, %r370, 8;
shr.u32 %r372, %r370, 16;
shr.u32 %r373, %r370, 24;
add.s64 %rd48, %rd33, %rd6;
ld.global.nc.u8 %rs24, [%rd48];
and.b16 %rs25, %rs24, 240;
cvt.u32.u16 %r374, %rs24;
and.b32 %r375, %r370, 255;
and.b32 %r376, %r374, 7;
bfi.b32 %r377, %r376, %r375, 8, 3;
mul.wide.u32 %rd49, %r377, 4;
add.s64 %rd50, %rd35, %rd49;
ld.global.nc.u32 %r378, [%rd50];
and.b32 %r136, %r378, 252645135;
shr.u32 %r379, %r378, 4;
and.b32 %r140, %r379, 252645135;

dp4a.s32.s32 %r135, %r136, %r72, %r324;


dp4a.s32.s32 %r139, %r140, %r76, %r135;

and.b32 %r380, %r374, 8;
cvt.rn.f32.s32 %f63, %r380;
mul.ftz.f32 %f64, %f63, 0f3D000000;
sub.ftz.f32 %f65, %f35, %f64;

dp4a.s32.s32 %r143, %r72, %r327, %r324;


dp4a.s32.s32 %r147, %r76, %r327, %r143;

cvt.rn.f32.s32 %f66, %r147;
fma.rn.ftz.f32 %f67, %f65, %f66, 0f00000000;
shr.u16 %rs26, %rs25, 4;
cvt.u32.u16 %r381, %rs26;
and.b32 %r382, %r371, 255;
and.b32 %r383, %r381, 7;
bfi.b32 %r384, %r383, %r382, 8, 3;
mul.wide.u32 %rd51, %r384, 4;
add.s64 %rd52, %rd35, %rd51;
ld.global.nc.u32 %r385, [%rd52];
and.b32 %r152, %r385, 252645135;
shr.u32 %r386, %r385, 4;
and.b32 %r156, %r386, 252645135;

dp4a.s32.s32 %r151, %r152, %r88, %r139;


dp4a.s32.s32 %r155, %r156, %r92, %r151;

and.b32 %r387, %r381, 8;
cvt.rn.f32.s32 %f68, %r387;
mul.ftz.f32 %f69, %f68, 0f3D000000;
sub.ftz.f32 %f70, %f35, %f69;

dp4a.s32.s32 %r159, %r88, %r327, %r324;


dp4a.s32.s32 %r163, %r92, %r327, %r159;

cvt.rn.f32.s32 %f71, %r163;
fma.rn.ftz.f32 %f72, %f70, %f71, %f67;
ld.global.nc.u8 %rs27, [%rd48+1];
and.b16 %rs28, %rs27, 240;
cvt.u32.u16 %r388, %rs27;
and.b32 %r389, %r372, 255;
and.b32 %r390, %r388, 7;
bfi.b32 %r391, %r390, %r389, 8, 3;
mul.wide.u32 %rd53, %r391, 4;
add.s64 %rd54, %rd35, %rd53;
ld.global.nc.u32 %r392, [%rd54];
and.b32 %r168, %r392, 252645135;
shr.u32 %r393, %r392, 4;
and.b32 %r172, %r393, 252645135;

dp4a.s32.s32 %r167, %r168, %r104, %r324;


dp4a.s32.s32 %r171, %r172, %r108, %r167;

and.b32 %r394, %r388, 8;
cvt.rn.f32.s32 %f73, %r394;
mul.ftz.f32 %f74, %f73, 0f3D000000;
sub.ftz.f32 %f75, %f35, %f74;

dp4a.s32.s32 %r175, %r104, %r327, %r324;


dp4a.s32.s32 %r179, %r108, %r327, %r175;

cvt.rn.f32.s32 %f76, %r179;
fma.rn.ftz.f32 %f77, %f75, %f76, 0f00000000;
shr.u16 %rs29, %rs28, 4;
cvt.u32.u16 %r395, %rs29;
mul.wide.u16 %r396, %rs29, 256;
and.b32 %r397, %r396, 1792;
or.b32 %r398, %r397, %r373;
mul.wide.u32 %rd55, %r398, 4;
add.s64 %rd56, %rd35, %rd55;
ld.global.nc.u32 %r399, [%rd56];
and.b32 %r184, %r399, 252645135;
shr.u32 %r400, %r399, 4;
and.b32 %r188, %r400, 252645135;

dp4a.s32.s32 %r183, %r184, %r120, %r171;


dp4a.s32.s32 %r187, %r188, %r124, %r183;

and.b32 %r401, %r395, 8;
cvt.rn.f32.s32 %f78, %r401;
mul.ftz.f32 %f79, %f78, 0f3D000000;
sub.ftz.f32 %f80, %f35, %f79;

dp4a.s32.s32 %r191, %r120, %r327, %r324;


dp4a.s32.s32 %r195, %r124, %r327, %r191;

cvt.rn.f32.s32 %f81, %r195;
fma.rn.ftz.f32 %f82, %f80, %f81, %f77;
add.s64 %rd57, %rd45, %rd6;
ld.global.nc.u16 %rs30, [%rd57];
shr.u16 %rs31, %rs30, 12;
ld.global.nc.u16 %rs32, [%rd57+2];
shr.u16 %rs33, %rs32, 8;
and.b16 %rs34, %rs33, 240;
or.b16 %rs35, %rs34, %rs31;
ld.global.nc.u16 %rs36, [%rd57+4];
shr.u16 %rs37, %rs36, 4;
and.b16 %rs38, %rs37, 3840;
or.b16 %rs39, %rs35, %rs38;
ld.global.nc.u16 %rs40, [%rd57+6];
and.b16 %rs41, %rs40, -4096;
or.b16 %rs2, %rs39, %rs41;

{ cvt.f32.f16 %f27, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r134;
cvt.f32.f16 %f28, low;}


mul.ftz.f32 %f83, %f27, %f28;
add.s64 %rd58, %rd46, %rd6;
ld.global.nc.u16 %rs42, [%rd58];
cvt.u32.u16 %r402, %rs42;
shr.u32 %r403, %r402, %r13;
shl.b32 %r404, %r403, 1;
and.b32 %r405, %r404, 14;
or.b32 %r406, %r405, 1;
shr.u32 %r407, %r403, 2;
and.b32 %r408, %r407, 14;
or.b32 %r409, %r408, 1;
cvt.rn.f32.s32 %f84, %r155;
add.ftz.f32 %f85, %f72, %f84;
cvt.rn.f32.s32 %f86, %r406;
cvt.rn.f32.s32 %f87, %r187;
add.ftz.f32 %f88, %f82, %f87;
cvt.rn.f32.s32 %f89, %r409;
mul.ftz.f32 %f90, %f88, %f89;
fma.rn.ftz.f32 %f91, %f85, %f86, %f90;
fma.rn.ftz.f32 %f216, %f83, %f91, %f216;
add.s64 %rd59, %rd47, %rd6;
ld.global.nc.u32 %r410, [%rd59];
shr.u32 %r411, %r410, 8;
shr.u32 %r412, %r410, 16;
shr.u32 %r413, %r410, 24;
add.s64 %rd60, %rd48, %rd6;
ld.global.nc.u8 %rs43, [%rd60];
and.b16 %rs44, %rs43, 240;
cvt.u32.u16 %r414, %rs43;
and.b32 %r415, %r410, 255;
and.b32 %r416, %r414, 7;
bfi.b32 %r417, %r416, %r415, 8, 3;
mul.wide.u32 %rd61, %r417, 4;
add.s64 %rd62, %rd35, %rd61;
ld.global.nc.u32 %r418, [%rd62];
and.b32 %r201, %r418, 252645135;
shr.u32 %r419, %r418, 4;
and.b32 %r205, %r419, 252645135;

dp4a.s32.s32 %r200, %r201, %r72, %r324;


dp4a.s32.s32 %r204, %r205, %r76, %r200;

and.b32 %r420, %r414, 8;
cvt.rn.f32.s32 %f92, %r420;
mul.ftz.f32 %f93, %f92, 0f3D000000;
sub.ftz.f32 %f94, %f35, %f93;

dp4a.s32.s32 %r208, %r72, %r327, %r324;


dp4a.s32.s32 %r212, %r76, %r327, %r208;

cvt.rn.f32.s32 %f95, %r212;
fma.rn.ftz.f32 %f96, %f94, %f95, 0f00000000;
shr.u16 %rs45, %rs44, 4;
cvt.u32.u16 %r421, %rs45;
and.b32 %r422, %r411, 255;
and.b32 %r423, %r421, 7;
bfi.b32 %r424, %r423, %r422, 8, 3;
mul.wide.u32 %rd63, %r424, 4;
add.s64 %rd64, %rd35, %rd63;
ld.global.nc.u32 %r425, [%rd64];
and.b32 %r217, %r425, 252645135;
shr.u32 %r426, %r425, 4;
and.b32 %r221, %r426, 252645135;

dp4a.s32.s32 %r216, %r217, %r88, %r204;


dp4a.s32.s32 %r220, %r221, %r92, %r216;

and.b32 %r427, %r421, 8;
cvt.rn.f32.s32 %f97, %r427;
mul.ftz.f32 %f98, %f97, 0f3D000000;
sub.ftz.f32 %f99, %f35, %f98;

dp4a.s32.s32 %r224, %r88, %r327, %r324;


dp4a.s32.s32 %r228, %r92, %r327, %r224;

cvt.rn.f32.s32 %f100, %r228;
fma.rn.ftz.f32 %f101, %f99, %f100, %f96;
ld.global.nc.u8 %rs46, [%rd60+1];
and.b16 %rs47, %rs46, 240;
cvt.u32.u16 %r428, %rs46;
and.b32 %r429, %r412, 255;
and.b32 %r430, %r428, 7;
bfi.b32 %r431, %r430, %r429, 8, 3;
mul.wide.u32 %rd65, %r431, 4;
add.s64 %rd66, %rd35, %rd65;
ld.global.nc.u32 %r432, [%rd66];
and.b32 %r233, %r432, 252645135;
shr.u32 %r433, %r432, 4;
and.b32 %r237, %r433, 252645135;

dp4a.s32.s32 %r232, %r233, %r104, %r324;


dp4a.s32.s32 %r236, %r237, %r108, %r232;

and.b32 %r434, %r428, 8;
cvt.rn.f32.s32 %f102, %r434;
mul.ftz.f32 %f103, %f102, 0f3D000000;
sub.ftz.f32 %f104, %f35, %f103;

dp4a.s32.s32 %r240, %r104, %r327, %r324;


dp4a.s32.s32 %r244, %r108, %r327, %r240;

cvt.rn.f32.s32 %f105, %r244;
fma.rn.ftz.f32 %f106, %f104, %f105, 0f00000000;
shr.u16 %rs48, %rs47, 4;
cvt.u32.u16 %r435, %rs48;
mul.wide.u16 %r436, %rs48, 256;
and.b32 %r437, %r436, 1792;
or.b32 %r438, %r437, %r413;
mul.wide.u32 %rd67, %r438, 4;
add.s64 %rd68, %rd35, %rd67;
ld.global.nc.u32 %r439, [%rd68];
and.b32 %r249, %r439, 252645135;
shr.u32 %r440, %r439, 4;
and.b32 %r253, %r440, 252645135;

dp4a.s32.s32 %r248, %r249, %r120, %r236;


dp4a.s32.s32 %r252, %r253, %r124, %r248;

and.b32 %r441, %r435, 8;
cvt.rn.f32.s32 %f107, %r441;
mul.ftz.f32 %f108, %f107, 0f3D000000;
sub.ftz.f32 %f109, %f35, %f108;

dp4a.s32.s32 %r256, %r120, %r327, %r324;


dp4a.s32.s32 %r260, %r124, %r327, %r256;

cvt.rn.f32.s32 %f110, %r260;
fma.rn.ftz.f32 %f111, %f109, %f110, %f106;
add.s64 %rd69, %rd57, %rd6;
ld.global.nc.u16 %rs49, [%rd69];
shr.u16 %rs50, %rs49, 12;
ld.global.nc.u16 %rs51, [%rd69+2];
shr.u16 %rs52, %rs51, 8;
and.b16 %rs53, %rs52, 240;
or.b16 %rs54, %rs53, %rs50;
ld.global.nc.u16 %rs55, [%rd69+4];
shr.u16 %rs56, %rs55, 4;
and.b16 %rs57, %rs56, 3840;
or.b16 %rs58, %rs54, %rs57;
ld.global.nc.u16 %rs59, [%rd69+6];
and.b16 %rs60, %rs59, -4096;
or.b16 %rs3, %rs58, %rs60;

{ cvt.f32.f16 %f29, %rs3;}



{.reg .f16 low,high;
mov.b32 {low,high},%r134;
cvt.f32.f16 %f30, low;}


mul.ftz.f32 %f112, %f29, %f30;
add.s64 %rd70, %rd58, %rd6;
ld.global.nc.u16 %rs61, [%rd70];
cvt.u32.u16 %r442, %rs61;
shr.u32 %r443, %r442, %r13;
shl.b32 %r444, %r443, 1;
and.b32 %r445, %r444, 14;
or.b32 %r446, %r445, 1;
shr.u32 %r447, %r443, 2;
and.b32 %r448, %r447, 14;
or.b32 %r449, %r448, 1;
cvt.rn.f32.s32 %f113, %r220;
add.ftz.f32 %f114, %f101, %f113;
cvt.rn.f32.s32 %f115, %r446;
cvt.rn.f32.s32 %f116, %r252;
add.ftz.f32 %f117, %f111, %f116;
cvt.rn.f32.s32 %f118, %r449;
mul.ftz.f32 %f119, %f117, %f118;
fma.rn.ftz.f32 %f120, %f114, %f115, %f119;
fma.rn.ftz.f32 %f215, %f112, %f120, %f215;
add.s64 %rd71, %rd59, %rd6;
ld.global.nc.u32 %r450, [%rd71];
shr.u32 %r451, %r450, 8;
shr.u32 %r452, %r450, 16;
shr.u32 %r453, %r450, 24;
add.s64 %rd72, %rd60, %rd6;
ld.global.nc.u8 %rs62, [%rd72];
and.b16 %rs63, %rs62, 240;
cvt.u32.u16 %r454, %rs62;
and.b32 %r455, %r450, 255;
and.b32 %r456, %r454, 7;
bfi.b32 %r457, %r456, %r455, 8, 3;
mul.wide.u32 %rd73, %r457, 4;
add.s64 %rd74, %rd35, %rd73;
ld.global.nc.u32 %r458, [%rd74];
and.b32 %r266, %r458, 252645135;
shr.u32 %r459, %r458, 4;
and.b32 %r270, %r459, 252645135;

dp4a.s32.s32 %r265, %r266, %r72, %r324;


dp4a.s32.s32 %r269, %r270, %r76, %r265;

and.b32 %r460, %r454, 8;
cvt.rn.f32.s32 %f121, %r460;
mul.ftz.f32 %f122, %f121, 0f3D000000;
sub.ftz.f32 %f123, %f35, %f122;

dp4a.s32.s32 %r273, %r72, %r327, %r324;


dp4a.s32.s32 %r277, %r76, %r327, %r273;

cvt.rn.f32.s32 %f124, %r277;
fma.rn.ftz.f32 %f125, %f123, %f124, 0f00000000;
shr.u16 %rs64, %rs63, 4;
cvt.u32.u16 %r461, %rs64;
and.b32 %r462, %r451, 255;
and.b32 %r463, %r461, 7;
bfi.b32 %r464, %r463, %r462, 8, 3;
mul.wide.u32 %rd75, %r464, 4;
add.s64 %rd76, %rd35, %rd75;
ld.global.nc.u32 %r465, [%rd76];
and.b32 %r282, %r465, 252645135;
shr.u32 %r466, %r465, 4;
and.b32 %r286, %r466, 252645135;

dp4a.s32.s32 %r281, %r282, %r88, %r269;


dp4a.s32.s32 %r285, %r286, %r92, %r281;

and.b32 %r467, %r461, 8;
cvt.rn.f32.s32 %f126, %r467;
mul.ftz.f32 %f127, %f126, 0f3D000000;
sub.ftz.f32 %f128, %f35, %f127;

dp4a.s32.s32 %r289, %r88, %r327, %r324;


dp4a.s32.s32 %r293, %r92, %r327, %r289;

cvt.rn.f32.s32 %f129, %r293;
fma.rn.ftz.f32 %f130, %f128, %f129, %f125;
ld.global.nc.u8 %rs65, [%rd72+1];
and.b16 %rs66, %rs65, 240;
cvt.u32.u16 %r468, %rs65;
and.b32 %r469, %r452, 255;
and.b32 %r470, %r468, 7;
bfi.b32 %r471, %r470, %r469, 8, 3;
mul.wide.u32 %rd77, %r471, 4;
add.s64 %rd78, %rd35, %rd77;
ld.global.nc.u32 %r472, [%rd78];
and.b32 %r298, %r472, 252645135;
shr.u32 %r473, %r472, 4;
and.b32 %r302, %r473, 252645135;

dp4a.s32.s32 %r297, %r298, %r104, %r324;


dp4a.s32.s32 %r301, %r302, %r108, %r297;

and.b32 %r474, %r468, 8;
cvt.rn.f32.s32 %f131, %r474;
mul.ftz.f32 %f132, %f131, 0f3D000000;
sub.ftz.f32 %f133, %f35, %f132;

dp4a.s32.s32 %r305, %r104, %r327, %r324;


dp4a.s32.s32 %r309, %r108, %r327, %r305;

cvt.rn.f32.s32 %f134, %r309;
fma.rn.ftz.f32 %f135, %f133, %f134, 0f00000000;
shr.u16 %rs67, %rs66, 4;
cvt.u32.u16 %r475, %rs67;
mul.wide.u16 %r476, %rs67, 256;
and.b32 %r477, %r476, 1792;
or.b32 %r478, %r477, %r453;
mul.wide.u32 %rd79, %r478, 4;
add.s64 %rd80, %rd35, %rd79;
ld.global.nc.u32 %r479, [%rd80];
and.b32 %r314, %r479, 252645135;
shr.u32 %r480, %r479, 4;
and.b32 %r318, %r480, 252645135;

dp4a.s32.s32 %r313, %r314, %r120, %r301;


dp4a.s32.s32 %r317, %r318, %r124, %r313;

and.b32 %r481, %r475, 8;
cvt.rn.f32.s32 %f136, %r481;
mul.ftz.f32 %f137, %f136, 0f3D000000;
sub.ftz.f32 %f138, %f35, %f137;

dp4a.s32.s32 %r321, %r120, %r327, %r324;


dp4a.s32.s32 %r325, %r124, %r327, %r321;

cvt.rn.f32.s32 %f139, %r325;
fma.rn.ftz.f32 %f140, %f138, %f139, %f135;
add.s64 %rd81, %rd69, %rd6;
ld.global.nc.u16 %rs68, [%rd81];
shr.u16 %rs69, %rs68, 12;
ld.global.nc.u16 %rs70, [%rd81+2];
shr.u16 %rs71, %rs70, 8;
and.b16 %rs72, %rs71, 240;
or.b16 %rs73, %rs72, %rs69;
ld.global.nc.u16 %rs74, [%rd81+4];
shr.u16 %rs75, %rs74, 4;
and.b16 %rs76, %rs75, 3840;
or.b16 %rs77, %rs73, %rs76;
ld.global.nc.u16 %rs78, [%rd81+6];
and.b16 %rs79, %rs78, -4096;
or.b16 %rs4, %rs77, %rs79;

{ cvt.f32.f16 %f31, %rs4;}



{.reg .f16 low,high;
mov.b32 {low,high},%r134;
cvt.f32.f16 %f32, low;}


mul.ftz.f32 %f141, %f31, %f32;
add.s64 %rd82, %rd70, %rd6;
ld.global.nc.u16 %rs80, [%rd82];
cvt.u32.u16 %r482, %rs80;
shr.u32 %r483, %r482, %r13;
shl.b32 %r484, %r483, 1;
and.b32 %r485, %r484, 14;
or.b32 %r486, %r485, 1;
shr.u32 %r487, %r483, 2;
and.b32 %r488, %r487, 14;
or.b32 %r489, %r488, 1;
cvt.rn.f32.s32 %f142, %r285;
add.ftz.f32 %f143, %f130, %f142;
cvt.rn.f32.s32 %f144, %r486;
cvt.rn.f32.s32 %f145, %r317;
add.ftz.f32 %f146, %f140, %f145;
cvt.rn.f32.s32 %f147, %r489;
mul.ftz.f32 %f148, %f146, %f147;
fma.rn.ftz.f32 %f149, %f143, %f144, %f148;
fma.rn.ftz.f32 %f214, %f141, %f149, %f214;
add.s32 %r574, %r574, 128;
add.s64 %rd85, %rd85, 896;
add.s32 %r575, %r575, 16;
setp.lt.s32 %p3, %r575, %r5;
@%p3 bra $L__BB266_5;

$L__BB266_6:
setp.eq.s32 %p4, %r1, 0;
@%p4 bra $L__BB266_8;

shl.b32 %r490, %r1, 9;
mov.u32 %r491, _ZZ13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r492, %r491, %r490;
shl.b32 %r493, %r2, 2;
add.s32 %r494, %r492, %r493;
st.shared.f32 [%r494+-512], %f217;
st.shared.f32 [%r494+-384], %f216;
st.shared.f32 [%r494+-256], %f215;
st.shared.f32 [%r494+-128], %f214;

$L__BB266_8:
setp.ne.s32 %p5, %r1, 0;
bar.sync 0;
@%p5 bra $L__BB266_17;

mov.u32 %r495, %ctaid.z;
mov.u32 %r496, %ctaid.y;
mad.lo.s32 %r497, %r496, %r31, %r4;
mad.lo.s32 %r498, %r495, %r37, %r497;
shl.b32 %r499, %r2, 2;
mov.u32 %r500, _ZZ13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r19, %r500, %r499;
mov.u32 %r501, 2;
ld.shared.f32 %f150, [%r19];
add.ftz.f32 %f151, %f150, %f217;
ld.shared.f32 %f152, [%r19+512];
add.ftz.f32 %f153, %f152, %f151;
ld.shared.f32 %f154, [%r19+1024];
add.ftz.f32 %f155, %f154, %f153;
mov.b32 %r502, %f155;
mov.u32 %r503, 31;
mov.u32 %r504, 16;
mov.u32 %r505, -1;
shfl.sync.bfly.b32 %r506|%p6, %r502, %r504, %r503, %r505;
mov.b32 %f156, %r506;
add.ftz.f32 %f157, %f155, %f156;
mov.b32 %r507, %f157;
mov.u32 %r508, 8;
shfl.sync.bfly.b32 %r509|%p7, %r507, %r508, %r503, %r505;
mov.b32 %f158, %r509;
add.ftz.f32 %f159, %f157, %f158;
mov.b32 %r510, %f159;
mov.u32 %r511, 4;
shfl.sync.bfly.b32 %r512|%p8, %r510, %r511, %r503, %r505;
mov.b32 %f160, %r512;
add.ftz.f32 %f161, %f159, %f160;
mov.b32 %r513, %f161;
shfl.sync.bfly.b32 %r514|%p9, %r513, %r501, %r503, %r505;
mov.b32 %f162, %r514;
add.ftz.f32 %f163, %f161, %f162;
mov.b32 %r515, %f163;
mov.u32 %r516, 1;
shfl.sync.bfly.b32 %r517|%p10, %r515, %r516, %r503, %r505;
mov.b32 %f164, %r517;
add.ftz.f32 %f13, %f163, %f164;
setp.ge.u32 %p11, %r4, %r25;
setp.ne.s32 %p12, %r2, 0;
cvta.to.global.u64 %rd83, %rd16;
mul.wide.u32 %rd84, %r498, 4;
add.s64 %rd12, %rd83, %rd84;
or.pred %p13, %p12, %p11;
@%p13 bra $L__BB266_11;

st.global.f32 [%rd12], %f13;

$L__BB266_11:
ld.shared.f32 %f165, [%r19+128];
add.ftz.f32 %f166, %f165, %f216;
ld.shared.f32 %f167, [%r19+640];
add.ftz.f32 %f168, %f167, %f166;
ld.shared.f32 %f169, [%r19+1152];
add.ftz.f32 %f170, %f169, %f168;
mov.b32 %r518, %f170;
shfl.sync.bfly.b32 %r522|%p14, %r518, %r504, %r503, %r505;
mov.b32 %f171, %r522;
add.ftz.f32 %f172, %f170, %f171;
mov.b32 %r523, %f172;
shfl.sync.bfly.b32 %r525|%p15, %r523, %r508, %r503, %r505;
mov.b32 %f173, %r525;
add.ftz.f32 %f174, %f172, %f173;
mov.b32 %r526, %f174;
shfl.sync.bfly.b32 %r528|%p16, %r526, %r511, %r503, %r505;
mov.b32 %f175, %r528;
add.ftz.f32 %f176, %f174, %f175;
mov.b32 %r529, %f176;
shfl.sync.bfly.b32 %r531|%p17, %r529, %r501, %r503, %r505;
mov.b32 %f177, %r531;
add.ftz.f32 %f178, %f176, %f177;
mov.b32 %r532, %f178;
shfl.sync.bfly.b32 %r534|%p18, %r532, %r516, %r503, %r505;
mov.b32 %f179, %r534;
add.ftz.f32 %f14, %f178, %f179;
add.s32 %r535, %r4, 1;
setp.ge.u32 %p19, %r535, %r25;
setp.ne.s32 %p20, %r2, 1;
or.pred %p21, %p20, %p19;
@%p21 bra $L__BB266_13;

st.global.f32 [%rd12+4], %f14;

$L__BB266_13:
ld.shared.f32 %f180, [%r19+256];
add.ftz.f32 %f181, %f180, %f215;
ld.shared.f32 %f182, [%r19+768];
add.ftz.f32 %f183, %f182, %f181;
ld.shared.f32 %f184, [%r19+1280];
add.ftz.f32 %f185, %f184, %f183;
mov.b32 %r536, %f185;
mov.u32 %r537, 31;
mov.u32 %r538, 16;
mov.u32 %r539, -1;
shfl.sync.bfly.b32 %r540|%p22, %r536, %r538, %r537, %r539;
mov.b32 %f186, %r540;
add.ftz.f32 %f187, %f185, %f186;
mov.b32 %r541, %f187;
mov.u32 %r542, 8;
shfl.sync.bfly.b32 %r543|%p23, %r541, %r542, %r537, %r539;
mov.b32 %f188, %r543;
add.ftz.f32 %f189, %f187, %f188;
mov.b32 %r544, %f189;
mov.u32 %r545, 4;
shfl.sync.bfly.b32 %r546|%p24, %r544, %r545, %r537, %r539;
mov.b32 %f190, %r546;
add.ftz.f32 %f191, %f189, %f190;
mov.b32 %r547, %f191;
mov.u32 %r548, 2;
shfl.sync.bfly.b32 %r549|%p25, %r547, %r548, %r537, %r539;
mov.b32 %f192, %r549;
add.ftz.f32 %f193, %f191, %f192;
mov.b32 %r550, %f193;
mov.u32 %r551, 1;
shfl.sync.bfly.b32 %r552|%p26, %r550, %r551, %r537, %r539;
mov.b32 %f194, %r552;
add.ftz.f32 %f15, %f193, %f194;
add.s32 %r553, %r4, 2;
setp.ge.u32 %p27, %r553, %r25;
setp.ne.s32 %p28, %r2, 2;
or.pred %p29, %p28, %p27;
@%p29 bra $L__BB266_15;

st.global.f32 [%rd12+8], %f15;

$L__BB266_15:
ld.shared.f32 %f195, [%r19+384];
add.ftz.f32 %f196, %f195, %f214;
ld.shared.f32 %f197, [%r19+896];
add.ftz.f32 %f198, %f197, %f196;
ld.shared.f32 %f199, [%r19+1408];
add.ftz.f32 %f200, %f199, %f198;
mov.b32 %r554, %f200;
shfl.sync.bfly.b32 %r558|%p30, %r554, %r538, %r537, %r539;
mov.b32 %f201, %r558;
add.ftz.f32 %f202, %f200, %f201;
mov.b32 %r559, %f202;
shfl.sync.bfly.b32 %r561|%p31, %r559, %r542, %r537, %r539;
mov.b32 %f203, %r561;
add.ftz.f32 %f204, %f202, %f203;
mov.b32 %r562, %f204;
shfl.sync.bfly.b32 %r564|%p32, %r562, %r545, %r537, %r539;
mov.b32 %f205, %r564;
add.ftz.f32 %f206, %f204, %f205;
mov.b32 %r565, %f206;
shfl.sync.bfly.b32 %r567|%p33, %r565, %r548, %r537, %r539;
mov.b32 %f207, %r567;
add.ftz.f32 %f208, %f206, %f207;
mov.b32 %r568, %f208;
shfl.sync.bfly.b32 %r570|%p34, %r568, %r551, %r537, %r539;
mov.b32 %f209, %r570;
add.ftz.f32 %f16, %f208, %f209;
add.s32 %r571, %r4, 3;
setp.ge.u32 %p35, %r571, %r25;
setp.ne.s32 %p36, %r2, 3;
or.pred %p37, %p36, %p35;
@%p37 bra $L__BB266_17;

st.global.f32 [%rd12+12], %f16;

$L__BB266_17:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<33>;
.reg .b16 %rs<37>;
.reg .f32 %f<177>;
.reg .b32 %r<376>;
.reg .b64 %rd<94>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[384];

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate[384];

ld.param.u64 %rd8, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd9, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd10, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r48, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r49, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r50, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r46, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r45, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
ld.param.u64 %rd13, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+16];
ld.param.u64 %rd12, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+8];
ld.param.u64 %rd11, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3];
mov.u32 %r1, %tid.y;
shl.b32 %r51, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r51, %r2;
mov.u32 %r4, %ctaid.x;
mov.u32 %r5, %ctaid.y;
setp.eq.s64 %p2, %rd10, 0;
@%p2 bra $L__BB267_2;

cvta.to.global.u64 %rd17, %rd10;
mul.wide.u32 %rd18, %r5, 4;
add.s64 %rd19, %rd17, %rd18;
ld.global.nc.u32 %r372, [%rd19];
mul.hi.u32 %r52, %r5, %r35;
add.s32 %r53, %r52, %r5;
shr.u32 %r54, %r53, %r36;
mul.lo.s32 %r55, %r54, %r37;
sub.s32 %r373, %r5, %r55;
bra.uni $L__BB267_3;

$L__BB267_2:
mul.hi.u32 %r56, %r5, %r39;
add.s32 %r57, %r56, %r5;
shr.u32 %r372, %r57, %r40;
mov.u32 %r373, %r5;

$L__BB267_3:
selp.b32 %r11, %r5, %r372, %p2;
or.b32 %r58, %r2, %r1;
setp.ne.s32 %p4, %r58, 0;
mov.f32 %f170, 0f00000000;
mov.f32 %f166, %f170;
mov.f32 %f167, %f170;
@%p4 bra $L__BB267_8;

setp.eq.s64 %p5, %rd11, 0;
mov.f32 %f166, 0f00000000;
mov.f32 %f167, %f166;
@%p5 bra $L__BB267_6;

cvta.to.global.u64 %rd20, %rd11;
mov.u32 %r59, %ctaid.z;
mul.lo.s32 %r60, %r59, %r50;
cvt.u64.u32 %rd21, %r60;
mul.lo.s32 %r61, %r11, %r44;
cvt.u64.u32 %rd22, %r61;
cvt.u64.u32 %rd23, %r2;
cvt.s64.s32 %rd24, %r4;
add.s64 %rd25, %rd23, %rd24;
add.s64 %rd26, %rd25, %rd21;
add.s64 %rd27, %rd26, %rd22;
shl.b64 %rd28, %rd27, 2;
add.s64 %rd29, %rd20, %rd28;
ld.global.f32 %f167, [%rd29];

$L__BB267_6:
setp.eq.s64 %p6, %rd13, 0;
setp.eq.s64 %p7, %rd12, 0;
or.pred %p8, %p6, %p7;
@%p8 bra $L__BB267_8;

cvta.to.global.u64 %rd30, %rd13;
mov.u32 %r63, %ctaid.z;
mul.lo.s32 %r64, %r63, %r50;
cvt.u64.u32 %rd31, %r64;
mul.lo.s32 %r65, %r11, %r44;
cvt.u64.u32 %rd32, %r65;
cvt.u64.u32 %rd33, %r2;
cvt.s64.s32 %rd34, %r4;
add.s64 %rd35, %rd33, %rd34;
add.s64 %rd36, %rd35, %rd31;
add.s64 %rd37, %rd36, %rd32;
shl.b64 %rd38, %rd37, 2;
add.s64 %rd39, %rd30, %rd38;
ld.global.f32 %f166, [%rd39];

$L__BB267_8:
shr.u32 %r67, %r34, 8;
shr.s32 %r68, %r3, 31;
shr.u32 %r69, %r68, 29;
add.s32 %r70, %r3, %r69;
shr.s32 %r375, %r70, 3;
setp.ge.s32 %p9, %r375, %r67;
mov.f32 %f172, %f170;
@%p9 bra $L__BB267_13;

mov.u32 %r71, %ctaid.z;
and.b32 %r75, %r70, -8;
sub.s32 %r13, %r3, %r75;
shr.u32 %r76, %r13, 31;
add.s32 %r77, %r13, %r76;
shr.s32 %r14, %r77, 1;
and.b32 %r78, %r77, -2;
sub.s32 %r79, %r13, %r78;
mul.lo.s32 %r15, %r79, 6;
mul.hi.u32 %r80, %r71, %r45;
add.s32 %r81, %r80, %r71;
shr.u32 %r82, %r81, %r46;
mad.lo.s32 %r83, %r82, %r48, %r375;
mad.lo.s32 %r84, %r372, %r42, %r83;
mad.lo.s32 %r85, %r4, %r38, %r84;
mul.wide.s32 %rd93, %r85, 56;
shl.b32 %r374, %r375, 3;
mul.lo.s32 %r86, %r373, %r43;
cvt.u64.u32 %rd40, %r86;
mul.lo.s32 %r87, %r71, %r49;
cvt.u64.u32 %rd41, %r87;
add.s64 %rd5, %rd40, %rd41;
mov.f32 %f172, 0f00000000;
cvta.to.global.u64 %rd42, %rd9;
cvt.s64.s32 %rd45, %r13;
mul.wide.s32 %rd46, %r13, 4;
cvta.to.global.u64 %rd47, %rd8;
add.s64 %rd48, %rd47, %rd46;
mov.f32 %f170, %f172;

$L__BB267_10:
cvt.u64.u32 %rd43, %r374;
add.s64 %rd44, %rd5, %rd43;
add.s64 %rd49, %rd48, %rd93;
ld.global.nc.u32 %r153, [%rd49];
shr.u32 %r154, %r153, 8;
shr.u32 %r155, %r153, 16;
shr.u32 %r156, %r153, 24;
add.s64 %rd50, %rd44, %rd45;
mul.lo.s64 %rd51, %rd50, 36;
shl.b32 %r157, %r13, 1;
cvt.s64.s32 %rd52, %r157;
add.s64 %rd53, %rd47, %rd52;
add.s64 %rd54, %rd53, %rd93;
ld.global.nc.u8 %rs2, [%rd54+32];
and.b16 %rs3, %rs2, 240;
cvt.u32.u16 %r158, %rs2;
and.b32 %r159, %r153, 255;
and.b32 %r160, %r158, 7;
bfi.b32 %r161, %r160, %r159, 8, 3;
mul.wide.u32 %rd55, %r161, 4;
mov.u64 %rd56, iq1s_grid_gpu;
add.s64 %rd57, %rd56, %rd55;
ld.global.nc.u32 %r162, [%rd57];
and.b32 %r89, %r162, 252645135;
shr.u32 %r163, %r162, 4;
and.b32 %r93, %r163, 252645135;
add.s64 %rd58, %rd42, %rd51;
ld.global.nc.u32 %r19, [%rd58+4];
ld.global.nc.u32 %r20, [%rd58+8];
mov.u32 %r147, 0;

dp4a.s32.s32 %r88, %r89, %r19, %r147;


dp4a.s32.s32 %r92, %r93, %r20, %r88;

and.b32 %r164, %r158, 8;
cvt.rn.f32.s32 %f42, %r164;
mul.ftz.f32 %f43, %f42, 0f3D000000;
mov.f32 %f44, 0fBF600000;
sub.ftz.f32 %f45, %f44, %f43;
mov.u32 %r150, 16843009;

dp4a.s32.s32 %r96, %r19, %r150, %r147;


dp4a.s32.s32 %r100, %r20, %r150, %r96;

cvt.rn.f32.s32 %f46, %r100;
fma.rn.ftz.f32 %f47, %f45, %f46, 0f00000000;
shr.u16 %rs4, %rs3, 4;
cvt.u32.u16 %r165, %rs4;
and.b32 %r166, %r154, 255;
and.b32 %r167, %r165, 7;
bfi.b32 %r168, %r167, %r166, 8, 3;
mul.wide.u32 %rd59, %r168, 4;
add.s64 %rd60, %rd56, %rd59;
ld.global.nc.u32 %r169, [%rd60];
and.b32 %r105, %r169, 252645135;
shr.u32 %r170, %r169, 4;
and.b32 %r109, %r170, 252645135;
ld.global.nc.u32 %r21, [%rd58+12];
ld.global.nc.u32 %r22, [%rd58+16];

dp4a.s32.s32 %r104, %r105, %r21, %r92;


dp4a.s32.s32 %r108, %r109, %r22, %r104;

and.b32 %r171, %r165, 8;
cvt.rn.f32.s32 %f48, %r171;
mul.ftz.f32 %f49, %f48, 0f3D000000;
sub.ftz.f32 %f50, %f44, %f49;

dp4a.s32.s32 %r112, %r21, %r150, %r147;


dp4a.s32.s32 %r116, %r22, %r150, %r112;

cvt.rn.f32.s32 %f51, %r116;
fma.rn.ftz.f32 %f52, %f50, %f51, %f47;
ld.global.nc.u8 %rs5, [%rd54+33];
and.b16 %rs6, %rs5, 240;
cvt.u32.u16 %r172, %rs5;
and.b32 %r173, %r155, 255;
and.b32 %r174, %r172, 7;
bfi.b32 %r175, %r174, %r173, 8, 3;
mul.wide.u32 %rd61, %r175, 4;
add.s64 %rd62, %rd56, %rd61;
ld.global.nc.u32 %r176, [%rd62];
and.b32 %r121, %r176, 252645135;
shr.u32 %r177, %r176, 4;
and.b32 %r125, %r177, 252645135;
ld.global.nc.u32 %r23, [%rd58+20];
ld.global.nc.u32 %r24, [%rd58+24];

dp4a.s32.s32 %r120, %r121, %r23, %r147;


dp4a.s32.s32 %r124, %r125, %r24, %r120;

and.b32 %r178, %r172, 8;
cvt.rn.f32.s32 %f53, %r178;
mul.ftz.f32 %f54, %f53, 0f3D000000;
sub.ftz.f32 %f55, %f44, %f54;

dp4a.s32.s32 %r128, %r23, %r150, %r147;


dp4a.s32.s32 %r132, %r24, %r150, %r128;

cvt.rn.f32.s32 %f56, %r132;
fma.rn.ftz.f32 %f57, %f55, %f56, 0f00000000;
shr.u16 %rs7, %rs6, 4;
cvt.u32.u16 %r179, %rs7;
mul.wide.u16 %r180, %rs7, 256;
and.b32 %r181, %r180, 1792;
or.b32 %r182, %r181, %r156;
mul.wide.u32 %rd63, %r182, 4;
add.s64 %rd64, %rd56, %rd63;
ld.global.nc.u32 %r183, [%rd64];
and.b32 %r137, %r183, 252645135;
shr.u32 %r184, %r183, 4;
and.b32 %r141, %r184, 252645135;
ld.global.nc.u32 %r25, [%rd58+28];
ld.global.nc.u32 %r26, [%rd58+32];

dp4a.s32.s32 %r136, %r137, %r25, %r124;


dp4a.s32.s32 %r140, %r141, %r26, %r136;

and.b32 %r185, %r179, 8;
cvt.rn.f32.s32 %f58, %r185;
mul.ftz.f32 %f59, %f58, 0f3D000000;
sub.ftz.f32 %f60, %f44, %f59;

dp4a.s32.s32 %r144, %r25, %r150, %r147;


dp4a.s32.s32 %r148, %r26, %r150, %r144;

cvt.rn.f32.s32 %f61, %r148;
fma.rn.ftz.f32 %f62, %f60, %f61, %f57;
add.s64 %rd65, %rd47, %rd93;
ld.global.nc.u16 %rs8, [%rd65+48];
shr.u16 %rs9, %rs8, 12;
ld.global.nc.u16 %rs10, [%rd65+50];
shr.u16 %rs11, %rs10, 8;
and.b16 %rs12, %rs11, 240;
or.b16 %rs13, %rs12, %rs9;
ld.global.nc.u16 %rs14, [%rd65+52];
shr.u16 %rs15, %rs14, 4;
and.b16 %rs16, %rs15, 3840;
or.b16 %rs17, %rs13, %rs16;
ld.global.nc.u16 %rs18, [%rd65+54];
and.b16 %rs19, %rs18, -4096;
or.b16 %rs1, %rs17, %rs19;

{ cvt.f32.f16 %f40, %rs1;}


ld.global.nc.u32 %r27, [%rd58];

{.reg .f16 low,high;
mov.b32 {low,high},%r27;
cvt.f32.f16 %f41, low;}


mul.ftz.f32 %f63, %f40, %f41;
mul.wide.s32 %rd66, %r14, 2;
add.s64 %rd67, %rd47, %rd66;
add.s64 %rd68, %rd67, %rd93;
ld.global.nc.u16 %rs20, [%rd68+48];
cvt.u32.u16 %r186, %rs20;
shr.u32 %r187, %r186, %r15;
shl.b32 %r188, %r187, 1;
and.b32 %r189, %r188, 14;
or.b32 %r190, %r189, 1;
shr.u32 %r191, %r187, 2;
and.b32 %r192, %r191, 14;
or.b32 %r193, %r192, 1;
cvt.rn.f32.s32 %f64, %r108;
add.ftz.f32 %f65, %f52, %f64;
cvt.rn.f32.s32 %f66, %r190;
cvt.rn.f32.s32 %f67, %r140;
add.ftz.f32 %f68, %f62, %f67;
cvt.rn.f32.s32 %f69, %r193;
mul.ftz.f32 %f70, %f68, %f69;
fma.rn.ftz.f32 %f71, %f65, %f66, %f70;
fma.rn.ftz.f32 %f172, %f63, %f71, %f172;
setp.eq.s64 %p10, %rd12, 0;
@%p10 bra $L__BB267_12;

cvta.to.global.u64 %rd70, %rd12;
add.s64 %rd71, %rd70, %rd46;
add.s64 %rd72, %rd71, %rd93;
ld.global.u32 %r259, [%rd72];
shr.u32 %r260, %r259, 8;
shr.u32 %r261, %r259, 16;
shr.u32 %r262, %r259, 24;
add.s64 %rd74, %rd70, %rd52;
add.s64 %rd75, %rd74, %rd93;
ld.global.u8 %rs22, [%rd75+32];
cvt.u32.u16 %r264, %rs22;
and.b32 %r265, %r259, 255;
and.b32 %r266, %r264, 7;
bfi.b32 %r267, %r266, %r265, 8, 3;
mul.wide.u32 %rd76, %r267, 4;
add.s64 %rd78, %rd56, %rd76;
ld.global.nc.u32 %r268, [%rd78];
and.b32 %r195, %r268, 252645135;
shr.u32 %r269, %r268, 4;
and.b32 %r199, %r269, 252645135;

dp4a.s32.s32 %r194, %r195, %r19, %r147;


dp4a.s32.s32 %r198, %r199, %r20, %r194;

and.b32 %r270, %r264, 8;
cvt.rn.f32.s32 %f74, %r270;
mul.ftz.f32 %f75, %f74, 0f3D000000;
sub.ftz.f32 %f77, %f44, %f75;

dp4a.s32.s32 %r202, %r19, %r150, %r147;


dp4a.s32.s32 %r206, %r20, %r150, %r202;

cvt.rn.f32.s32 %f78, %r206;
fma.rn.ftz.f32 %f79, %f77, %f78, 0f00000000;
shr.u16 %rs23, %rs22, 4;
cvt.u32.u16 %r271, %rs23;
and.b32 %r272, %r260, 255;
and.b32 %r273, %r271, 7;
bfi.b32 %r274, %r273, %r272, 8, 3;
mul.wide.u32 %rd79, %r274, 4;
add.s64 %rd80, %rd56, %rd79;
ld.global.nc.u32 %r275, [%rd80];
and.b32 %r211, %r275, 252645135;
shr.u32 %r276, %r275, 4;
and.b32 %r215, %r276, 252645135;

dp4a.s32.s32 %r210, %r211, %r21, %r198;


dp4a.s32.s32 %r214, %r215, %r22, %r210;

and.b32 %r277, %r271, 8;
cvt.rn.f32.s32 %f80, %r277;
mul.ftz.f32 %f81, %f80, 0f3D000000;
sub.ftz.f32 %f82, %f44, %f81;

dp4a.s32.s32 %r218, %r21, %r150, %r147;


dp4a.s32.s32 %r222, %r22, %r150, %r218;

cvt.rn.f32.s32 %f83, %r222;
fma.rn.ftz.f32 %f84, %f82, %f83, %f79;
ld.global.u8 %rs24, [%rd75+33];
cvt.u32.u16 %r278, %rs24;
and.b32 %r279, %r261, 255;
and.b32 %r280, %r278, 7;
bfi.b32 %r281, %r280, %r279, 8, 3;
mul.wide.u32 %rd81, %r281, 4;
add.s64 %rd82, %rd56, %rd81;
ld.global.nc.u32 %r282, [%rd82];
and.b32 %r227, %r282, 252645135;
shr.u32 %r283, %r282, 4;
and.b32 %r231, %r283, 252645135;

dp4a.s32.s32 %r226, %r227, %r23, %r147;


dp4a.s32.s32 %r230, %r231, %r24, %r226;

and.b32 %r284, %r278, 8;
cvt.rn.f32.s32 %f85, %r284;
mul.ftz.f32 %f86, %f85, 0f3D000000;
sub.ftz.f32 %f87, %f44, %f86;

dp4a.s32.s32 %r234, %r23, %r150, %r147;


dp4a.s32.s32 %r238, %r24, %r150, %r234;

cvt.rn.f32.s32 %f88, %r238;
fma.rn.ftz.f32 %f89, %f87, %f88, 0f00000000;
shr.u16 %rs25, %rs24, 4;
cvt.u32.u16 %r285, %rs25;
mul.wide.u16 %r286, %rs25, 256;
and.b32 %r287, %r286, 1792;
or.b32 %r288, %r287, %r262;
mul.wide.u32 %rd83, %r288, 4;
add.s64 %rd84, %rd56, %rd83;
ld.global.nc.u32 %r289, [%rd84];
and.b32 %r243, %r289, 252645135;
shr.u32 %r290, %r289, 4;
and.b32 %r247, %r290, 252645135;

dp4a.s32.s32 %r242, %r243, %r25, %r230;


dp4a.s32.s32 %r246, %r247, %r26, %r242;

and.b32 %r291, %r285, 8;
cvt.rn.f32.s32 %f90, %r291;
mul.ftz.f32 %f91, %f90, 0f3D000000;
sub.ftz.f32 %f92, %f44, %f91;

dp4a.s32.s32 %r250, %r25, %r150, %r147;


dp4a.s32.s32 %r254, %r26, %r150, %r250;

cvt.rn.f32.s32 %f93, %r254;
fma.rn.ftz.f32 %f94, %f92, %f93, %f89;
add.s64 %rd85, %rd70, %rd93;
ld.global.u16 %rs26, [%rd85+48];
shr.u16 %rs27, %rs26, 12;
ld.global.u8 %rs28, [%rd85+51];
and.b16 %rs29, %rs28, 240;
or.b16 %rs30, %rs29, %rs27;
ld.global.u16 %rs31, [%rd85+52];
shr.u16 %rs32, %rs31, 4;
and.b16 %rs33, %rs32, 3840;
or.b16 %rs34, %rs30, %rs33;
ld.global.u16 %rs35, [%rd85+54];
and.b16 %rs36, %rs35, -4096;
or.b16 %rs21, %rs34, %rs36;

{ cvt.f32.f16 %f72, %rs21;}



{.reg .f16 low,high;
mov.b32 {low,high},%r27;
cvt.f32.f16 %f73, low;}


mul.ftz.f32 %f95, %f72, %f73;
add.s64 %rd87, %rd70, %rd66;
add.s64 %rd88, %rd87, %rd93;
ld.global.u16 %r292, [%rd88+48];
shr.u32 %r293, %r292, %r15;
shl.b32 %r294, %r293, 1;
and.b32 %r295, %r294, 14;
or.b32 %r296, %r295, 1;
shr.u32 %r297, %r293, 2;
and.b32 %r298, %r297, 14;
or.b32 %r299, %r298, 1;
cvt.rn.f32.s32 %f96, %r214;
add.ftz.f32 %f97, %f84, %f96;
cvt.rn.f32.s32 %f98, %r296;
cvt.rn.f32.s32 %f99, %r246;
add.ftz.f32 %f100, %f94, %f99;
cvt.rn.f32.s32 %f101, %r299;
mul.ftz.f32 %f102, %f100, %f101;
fma.rn.ftz.f32 %f103, %f97, %f98, %f102;
fma.rn.ftz.f32 %f170, %f95, %f103, %f170;

$L__BB267_12:
add.s64 %rd93, %rd93, 896;
add.s32 %r374, %r374, 128;
add.s32 %r375, %r375, 16;
setp.lt.s32 %p11, %r375, %r67;
@%p11 bra $L__BB267_10;

$L__BB267_13:
mov.u32 %r363, %tid.y;
setp.eq.s32 %p12, %r363, 0;
@%p12 bra $L__BB267_16;

mov.u32 %r365, %tid.y;
mov.u32 %r364, %tid.x;
setp.eq.s64 %p13, %rd12, 0;
add.s32 %r30, %r365, -1;
shl.b32 %r304, %r365, 7;
mov.u32 %r305, _ZZ13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r306, %r305, %r304;
shl.b32 %r307, %r364, 2;
add.s32 %r308, %r306, %r307;
st.shared.f32 [%r308+-128], %f172;
@%p13 bra $L__BB267_16;

shl.b32 %r310, %r30, 7;
mov.u32 %r311, _ZZ13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r312, %r311, %r310;
add.s32 %r314, %r312, %r307;
st.shared.f32 [%r314], %f170;

$L__BB267_16:
mov.u32 %r366, %tid.y;
setp.ne.s32 %p14, %r366, 0;
bar.sync 0;
@%p14 bra $L__BB267_35;

mov.u32 %r367, %tid.x;
setp.ne.s64 %p15, %rd12, 0;
shl.b32 %r317, %r367, 2;
mov.u32 %r318, _ZZ13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r31, %r318, %r317;
ld.shared.f32 %f104, [%r31];
add.ftz.f32 %f13, %f172, %f104;
@%p15 bra $L__BB267_19;
bra.uni $L__BB267_18;

$L__BB267_19:
mov.u32 %r321, _ZZ13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r322, %r321, %r317;
ld.shared.f32 %f108, [%r322];
add.ftz.f32 %f109, %f170, %f108;
ld.shared.f32 %f110, [%r31+128];
add.ftz.f32 %f111, %f13, %f110;
ld.shared.f32 %f112, [%r322+128];
add.ftz.f32 %f113, %f109, %f112;
ld.shared.f32 %f114, [%r31+256];
add.ftz.f32 %f174, %f111, %f114;
ld.shared.f32 %f115, [%r322+256];
add.ftz.f32 %f170, %f113, %f115;
bra.uni $L__BB267_20;

$L__BB267_18:
ld.shared.f32 %f105, [%r31+128];
add.ftz.f32 %f106, %f13, %f105;
ld.shared.f32 %f107, [%r31+256];
add.ftz.f32 %f174, %f106, %f107;

$L__BB267_20:
mov.b32 %r323, %f174;
mov.u32 %r324, 31;
mov.u32 %r325, 16;
mov.u32 %r326, -1;
shfl.sync.bfly.b32 %r327|%p16, %r323, %r325, %r324, %r326;
mov.b32 %f116, %r327;
add.ftz.f32 %f117, %f174, %f116;
mov.b32 %r328, %f117;
mov.u32 %r329, 8;
shfl.sync.bfly.b32 %r330|%p17, %r328, %r329, %r324, %r326;
mov.b32 %f118, %r330;
add.ftz.f32 %f119, %f117, %f118;
mov.b32 %r331, %f119;
mov.u32 %r332, 4;
shfl.sync.bfly.b32 %r333|%p18, %r331, %r332, %r324, %r326;
mov.b32 %f120, %r333;
add.ftz.f32 %f121, %f119, %f120;
mov.b32 %r334, %f121;
mov.u32 %r335, 2;
shfl.sync.bfly.b32 %r336|%p19, %r334, %r335, %r324, %r326;
mov.b32 %f122, %r336;
add.ftz.f32 %f19, %f121, %f122;
mov.b32 %r337, %f19;
mov.u32 %r338, 1;
shfl.sync.bfly.b32 %r32|%p1, %r337, %r338, %r324, %r326;
setp.eq.s64 %p20, %rd12, 0;
@%p20 bra $L__BB267_22;

mov.b32 %r339, %f170;
shfl.sync.bfly.b32 %r343|%p21, %r339, %r325, %r324, %r326;
mov.b32 %f123, %r343;
add.ftz.f32 %f124, %f170, %f123;
mov.b32 %r344, %f124;
shfl.sync.bfly.b32 %r346|%p22, %r344, %r329, %r324, %r326;
mov.b32 %f125, %r346;
add.ftz.f32 %f126, %f124, %f125;
mov.b32 %r347, %f126;
shfl.sync.bfly.b32 %r349|%p23, %r347, %r332, %r324, %r326;
mov.b32 %f127, %r349;
add.ftz.f32 %f128, %f126, %f127;
mov.b32 %r350, %f128;
shfl.sync.bfly.b32 %r352|%p24, %r350, %r335, %r324, %r326;
mov.b32 %f129, %r352;
add.ftz.f32 %f130, %f128, %f129;
mov.b32 %r353, %f130;
shfl.sync.bfly.b32 %r355|%p25, %r353, %r338, %r324, %r326;
mov.b32 %f131, %r355;
add.ftz.f32 %f170, %f130, %f131;

$L__BB267_22:
mov.u32 %r368, %tid.x;
setp.ne.s32 %p26, %r368, 0;
mov.b32 %f132, %r32;
add.ftz.f32 %f22, %f19, %f132;
@%p26 bra $L__BB267_35;

add.ftz.f32 %f176, %f167, %f22;
@%p20 bra $L__BB267_34;

add.ftz.f32 %f24, %f166, %f170;
setp.gt.s32 %p28, %r33, 2;
@%p28 bra $L__BB267_28;

setp.eq.s32 %p31, %r33, 1;
@%p31 bra $L__BB267_32;

setp.eq.s32 %p32, %r33, 2;
@%p32 bra $L__BB267_27;
bra.uni $L__BB267_33;

$L__BB267_27:
mul.ftz.f32 %f152, %f24, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f153, %f152;
add.ftz.f32 %f154, %f153, 0f3F800000;
div.approx.ftz.f32 %f155, %f24, %f154;
mul.ftz.f32 %f176, %f176, %f155;
bra.uni $L__BB267_34;

$L__BB267_28:
setp.eq.s32 %p29, %r33, 3;
@%p29 bra $L__BB267_31;

setp.ne.s32 %p30, %r33, 6;
@%p30 bra $L__BB267_33;

ld.param.f32 %f164, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
min.ftz.f32 %f133, %f24, %f164;
neg.ftz.f32 %f134, %f164;
min.ftz.f32 %f135, %f176, %f164;
max.ftz.f32 %f136, %f135, %f134;
mul.ftz.f32 %f137, %f133, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f138, %f137;
add.ftz.f32 %f139, %f138, 0f3F800000;
div.approx.ftz.f32 %f140, %f133, %f139;
mul.ftz.f32 %f176, %f136, %f140;
bra.uni $L__BB267_34;

$L__BB267_32:
mul.ftz.f32 %f156, %f24, 0f3F000000;
mul.ftz.f32 %f157, %f24, 0f3D372713;
fma.rn.ftz.f32 %f158, %f24, %f157, 0f3F800000;
mul.ftz.f32 %f159, %f24, 0f3F4C422A;
mul.ftz.f32 %f160, %f159, %f158;
tanh.approx.f32 %f161, %f160;
add.ftz.f32 %f162, %f161, 0f3F800000;
mul.ftz.f32 %f163, %f156, %f162;
mul.ftz.f32 %f176, %f176, %f163;
bra.uni $L__BB267_34;

$L__BB267_31:
mov.f32 %f141, 0f40E00000;
min.ftz.f32 %f142, %f24, %f141;
min.ftz.f32 %f143, %f176, %f141;
mov.f32 %f144, 0fC0E00000;
max.ftz.f32 %f145, %f143, %f144;
mul.ftz.f32 %f146, %f142, 0f3FD9DB23;
mul.ftz.f32 %f147, %f146, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f148, %f147;
add.ftz.f32 %f149, %f148, 0f3F800000;
div.approx.ftz.f32 %f150, %f142, %f149;
add.ftz.f32 %f151, %f145, 0f3F800000;
mul.ftz.f32 %f176, %f151, %f150;
bra.uni $L__BB267_34;

$L__BB267_33:
mul.ftz.f32 %f176, %f176, %f24;

$L__BB267_34:
ld.param.u64 %rd92, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
mov.u32 %r371, %ctaid.x;
ld.param.u32 %r370, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r369, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
cvta.to.global.u64 %rd89, %rd92;
mov.u32 %r358, %ctaid.z;
mov.u32 %r359, %ctaid.y;
mad.lo.s32 %r361, %r359, %r370, %r371;
mad.lo.s32 %r362, %r358, %r369, %r361;
mul.wide.u32 %rd90, %r362, 4;
add.s64 %rd91, %rd89, %rd90;
st.global.f32 [%rd91], %f176;

$L__BB267_35:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<12>;
.reg .b16 %rs<21>;
.reg .f32 %f<56>;
.reg .b32 %r<205>;
.reg .b64 %rd<48>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[384];

ld.param.u64 %rd12, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd13, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd14, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
mov.u32 %r1, %tid.y;
shl.b32 %r37, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r37, %r2;
mov.u32 %r202, %ctaid.y;
setp.eq.s64 %p1, %rd14, 0;
@%p1 bra $L__BB268_2;

cvta.to.global.u64 %rd16, %rd14;
mul.wide.u32 %rd17, %r202, 4;
add.s64 %rd18, %rd16, %rd17;
ld.global.nc.u32 %r201, [%rd18];
mul.hi.u32 %r38, %r202, %r21;
add.s32 %r39, %r38, %r202;
shr.u32 %r40, %r39, %r22;
mul.lo.s32 %r41, %r40, %r23;
sub.s32 %r202, %r202, %r41;
bra.uni $L__BB268_3;

$L__BB268_2:
mul.hi.u32 %r42, %r202, %r25;
add.s32 %r43, %r42, %r202;
shr.u32 %r201, %r43, %r26;

$L__BB268_3:
mov.u32 %r10, %ctaid.z;
shr.s32 %r44, %r3, 31;
shr.u32 %r45, %r44, 29;
add.s32 %r46, %r3, %r45;
shr.s32 %r204, %r46, 3;
shr.u32 %r47, %r20, 8;
setp.ge.s32 %p2, %r204, %r47;
mov.f32 %f55, 0f00000000;
@%p2 bra $L__BB268_6;

and.b32 %r51, %r46, -8;
sub.s32 %r52, %r3, %r51;
cvt.s64.s32 %rd1, %r52;
shl.b32 %r53, %r52, 1;
shr.u32 %r54, %r52, 31;
add.s32 %r55, %r52, %r54;
shr.s32 %r56, %r55, 1;
and.b32 %r57, %r55, -2;
sub.s32 %r58, %r52, %r57;
mul.lo.s32 %r12, %r58, 6;
cvt.s64.s32 %rd19, %r53;
shl.b32 %r203, %r204, 3;
mul.hi.u32 %r59, %r10, %r31;
add.s32 %r60, %r59, %r10;
shr.u32 %r61, %r60, %r32;
mad.lo.s32 %r62, %r61, %r34, %r204;
mad.lo.s32 %r63, %r201, %r28, %r62;
mov.u32 %r64, %ctaid.x;
mad.lo.s32 %r65, %r64, %r24, %r63;
mul.wide.s32 %rd20, %r65, 56;
add.s64 %rd21, %rd20, %rd19;
cvta.to.global.u64 %rd22, %rd12;
add.s64 %rd23, %rd22, %rd21;
add.s64 %rd47, %rd23, 33;
add.s64 %rd46, %rd22, %rd20;
mul.wide.s32 %rd4, %r52, 4;
mul.wide.s32 %rd24, %r56, 2;
add.s64 %rd5, %rd24, 48;
cvta.to.global.u64 %rd6, %rd13;
mul.lo.s32 %r66, %r202, %r29;
cvt.u64.u32 %rd25, %r66;
mul.lo.s32 %r67, %r10, %r35;
cvt.u64.u32 %rd26, %r67;
add.s64 %rd7, %rd25, %rd26;
mov.f32 %f55, 0f00000000;

$L__BB268_5:
cvt.u64.u32 %rd27, %r203;
add.s64 %rd28, %rd7, %rd27;
add.s64 %rd29, %rd46, %rd4;
ld.global.nc.u32 %r133, [%rd29];
shr.u32 %r134, %r133, 8;
shr.u32 %r135, %r133, 16;
shr.u32 %r136, %r133, 24;
add.s64 %rd30, %rd28, %rd1;
mul.lo.s64 %rd31, %rd30, 36;
ld.global.nc.u8 %rs2, [%rd47+-1];
and.b16 %rs3, %rs2, 240;
cvt.u32.u16 %r137, %rs2;
and.b32 %r138, %r133, 255;
and.b32 %r139, %r137, 7;
bfi.b32 %r140, %r139, %r138, 8, 3;
mul.wide.u32 %rd32, %r140, 4;
mov.u64 %rd33, iq1s_grid_gpu;
add.s64 %rd34, %rd33, %rd32;
ld.global.nc.u32 %r141, [%rd34];
and.b32 %r69, %r141, 252645135;
shr.u32 %r142, %r141, 4;
and.b32 %r73, %r142, 252645135;
add.s64 %rd35, %rd6, %rd31;
ld.global.nc.u32 %r70, [%rd35+4];
ld.global.nc.u32 %r74, [%rd35+8];
mov.u32 %r127, 0;

dp4a.s32.s32 %r68, %r69, %r70, %r127;


dp4a.s32.s32 %r72, %r73, %r74, %r68;

and.b32 %r143, %r137, 8;
cvt.rn.f32.s32 %f9, %r143;
mul.ftz.f32 %f10, %f9, 0f3D000000;
mov.f32 %f11, 0fBF600000;
sub.ftz.f32 %f12, %f11, %f10;
mov.u32 %r130, 16843009;

dp4a.s32.s32 %r76, %r70, %r130, %r127;


dp4a.s32.s32 %r80, %r74, %r130, %r76;

cvt.rn.f32.s32 %f13, %r80;
fma.rn.ftz.f32 %f14, %f12, %f13, 0f00000000;
shr.u16 %rs4, %rs3, 4;
cvt.u32.u16 %r144, %rs4;
and.b32 %r145, %r134, 255;
and.b32 %r146, %r144, 7;
bfi.b32 %r147, %r146, %r145, 8, 3;
mul.wide.u32 %rd36, %r147, 4;
add.s64 %rd37, %rd33, %rd36;
ld.global.nc.u32 %r148, [%rd37];
and.b32 %r85, %r148, 252645135;
shr.u32 %r149, %r148, 4;
and.b32 %r89, %r149, 252645135;
ld.global.nc.u32 %r86, [%rd35+12];
ld.global.nc.u32 %r90, [%rd35+16];

dp4a.s32.s32 %r84, %r85, %r86, %r72;


dp4a.s32.s32 %r88, %r89, %r90, %r84;

and.b32 %r150, %r144, 8;
cvt.rn.f32.s32 %f15, %r150;
mul.ftz.f32 %f16, %f15, 0f3D000000;
sub.ftz.f32 %f17, %f11, %f16;

dp4a.s32.s32 %r92, %r86, %r130, %r127;


dp4a.s32.s32 %r96, %r90, %r130, %r92;

cvt.rn.f32.s32 %f18, %r96;
fma.rn.ftz.f32 %f19, %f17, %f18, %f14;
ld.global.nc.u8 %rs5, [%rd47];
and.b16 %rs6, %rs5, 240;
cvt.u32.u16 %r151, %rs5;
and.b32 %r152, %r135, 255;
and.b32 %r153, %r151, 7;
bfi.b32 %r154, %r153, %r152, 8, 3;
mul.wide.u32 %rd38, %r154, 4;
add.s64 %rd39, %rd33, %rd38;
ld.global.nc.u32 %r155, [%rd39];
and.b32 %r101, %r155, 252645135;
shr.u32 %r156, %r155, 4;
and.b32 %r105, %r156, 252645135;
ld.global.nc.u32 %r102, [%rd35+20];
ld.global.nc.u32 %r106, [%rd35+24];

dp4a.s32.s32 %r100, %r101, %r102, %r127;


dp4a.s32.s32 %r104, %r105, %r106, %r100;

and.b32 %r157, %r151, 8;
cvt.rn.f32.s32 %f20, %r157;
mul.ftz.f32 %f21, %f20, 0f3D000000;
sub.ftz.f32 %f22, %f11, %f21;

dp4a.s32.s32 %r108, %r102, %r130, %r127;


dp4a.s32.s32 %r112, %r106, %r130, %r108;

cvt.rn.f32.s32 %f23, %r112;
fma.rn.ftz.f32 %f24, %f22, %f23, 0f00000000;
shr.u16 %rs7, %rs6, 4;
cvt.u32.u16 %r158, %rs7;
mul.wide.u16 %r159, %rs7, 256;
and.b32 %r160, %r159, 1792;
or.b32 %r161, %r160, %r136;
mul.wide.u32 %rd40, %r161, 4;
add.s64 %rd41, %rd33, %rd40;
ld.global.nc.u32 %r162, [%rd41];
and.b32 %r117, %r162, 252645135;
shr.u32 %r163, %r162, 4;
and.b32 %r121, %r163, 252645135;
ld.global.nc.u32 %r118, [%rd35+28];
ld.global.nc.u32 %r122, [%rd35+32];

dp4a.s32.s32 %r116, %r117, %r118, %r104;


dp4a.s32.s32 %r120, %r121, %r122, %r116;

and.b32 %r164, %r158, 8;
cvt.rn.f32.s32 %f25, %r164;
mul.ftz.f32 %f26, %f25, 0f3D000000;
sub.ftz.f32 %f27, %f11, %f26;

dp4a.s32.s32 %r124, %r118, %r130, %r127;


dp4a.s32.s32 %r128, %r122, %r130, %r124;

cvt.rn.f32.s32 %f28, %r128;
fma.rn.ftz.f32 %f29, %f27, %f28, %f24;
ld.global.nc.u16 %rs8, [%rd46+48];
shr.u16 %rs9, %rs8, 12;
ld.global.nc.u16 %rs10, [%rd46+50];
shr.u16 %rs11, %rs10, 8;
and.b16 %rs12, %rs11, 240;
or.b16 %rs13, %rs12, %rs9;
ld.global.nc.u16 %rs14, [%rd46+52];
shr.u16 %rs15, %rs14, 4;
and.b16 %rs16, %rs15, 3840;
or.b16 %rs17, %rs13, %rs16;
ld.global.nc.u16 %rs18, [%rd46+54];
and.b16 %rs19, %rs18, -4096;
or.b16 %rs1, %rs17, %rs19;

{ cvt.f32.f16 %f7, %rs1;}


ld.global.nc.u32 %r132, [%rd35];

{.reg .f16 low,high;
mov.b32 {low,high},%r132;
cvt.f32.f16 %f8, low;}


mul.ftz.f32 %f30, %f7, %f8;
add.s64 %rd42, %rd46, %rd5;
ld.global.nc.u16 %rs20, [%rd42];
cvt.u32.u16 %r165, %rs20;
shr.u32 %r166, %r165, %r12;
shl.b32 %r167, %r166, 1;
and.b32 %r168, %r167, 14;
or.b32 %r169, %r168, 1;
shr.u32 %r170, %r166, 2;
and.b32 %r171, %r170, 14;
or.b32 %r172, %r171, 1;
cvt.rn.f32.s32 %f31, %r88;
add.ftz.f32 %f32, %f19, %f31;
cvt.rn.f32.s32 %f33, %r169;
cvt.rn.f32.s32 %f34, %r120;
add.ftz.f32 %f35, %f29, %f34;
cvt.rn.f32.s32 %f36, %r172;
mul.ftz.f32 %f37, %f35, %f36;
fma.rn.ftz.f32 %f38, %f32, %f33, %f37;
fma.rn.ftz.f32 %f55, %f30, %f38, %f55;
add.s32 %r203, %r203, 128;
add.s64 %rd47, %rd47, 896;
add.s64 %rd46, %rd46, 896;
add.s32 %r204, %r204, 16;
setp.lt.s32 %p3, %r204, %r47;
@%p3 bra $L__BB268_5;

$L__BB268_6:
setp.eq.s32 %p4, %r1, 0;
@%p4 bra $L__BB268_8;

shl.b32 %r173, %r1, 7;
mov.u32 %r174, _ZZ13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r175, %r174, %r173;
shl.b32 %r176, %r2, 2;
add.s32 %r177, %r175, %r176;
st.shared.f32 [%r177+-128], %f55;

$L__BB268_8:
setp.ne.s32 %p5, %r1, 0;
bar.sync 0;
@%p5 bra $L__BB268_11;

mov.u32 %r178, %ctaid.y;
mov.u32 %r179, %ctaid.x;
mad.lo.s32 %r180, %r178, %r30, %r179;
mad.lo.s32 %r19, %r10, %r36, %r180;
shl.b32 %r181, %r2, 2;
mov.u32 %r182, _ZZ13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r183, %r182, %r181;
mov.u32 %r184, 2;
ld.shared.f32 %f39, [%r183];
add.ftz.f32 %f40, %f55, %f39;
ld.shared.f32 %f41, [%r183+128];
add.ftz.f32 %f42, %f40, %f41;
ld.shared.f32 %f43, [%r183+256];
add.ftz.f32 %f44, %f42, %f43;
mov.b32 %r185, %f44;
mov.u32 %r186, 31;
mov.u32 %r187, 16;
mov.u32 %r188, -1;
shfl.sync.bfly.b32 %r189|%p6, %r185, %r187, %r186, %r188;
mov.b32 %f45, %r189;
add.ftz.f32 %f46, %f44, %f45;
mov.b32 %r190, %f46;
mov.u32 %r191, 8;
shfl.sync.bfly.b32 %r192|%p7, %r190, %r191, %r186, %r188;
mov.b32 %f47, %r192;
add.ftz.f32 %f48, %f46, %f47;
mov.b32 %r193, %f48;
mov.u32 %r194, 4;
shfl.sync.bfly.b32 %r195|%p8, %r193, %r194, %r186, %r188;
mov.b32 %f49, %r195;
add.ftz.f32 %f50, %f48, %f49;
mov.b32 %r196, %f50;
shfl.sync.bfly.b32 %r197|%p9, %r196, %r184, %r186, %r188;
mov.b32 %f51, %r197;
add.ftz.f32 %f52, %f50, %f51;
mov.b32 %r198, %f52;
mov.u32 %r199, 1;
shfl.sync.bfly.b32 %r200|%p10, %r198, %r199, %r186, %r188;
mov.b32 %f53, %r200;
add.ftz.f32 %f4, %f52, %f53;
setp.ne.s32 %p11, %r2, 0;
@%p11 bra $L__BB268_11;

cvta.to.global.u64 %rd43, %rd15;
mul.wide.u32 %rd44, %r19, 4;
add.s64 %rd45, %rd43, %rd44;
st.global.f32 [%rd45], %f4;

$L__BB268_11:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type29ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type29ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type29ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type29ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type29ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type29ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type29ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type29ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type29ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<35>;
.reg .b16 %rs<43>;
.reg .f32 %f<190>;
.reg .b32 %r<492>;
.reg .b64 %rd<75>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type29ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type29ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type29ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd20, [_Z13mul_mat_vec_qIL9ggml_type29ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r19, [_Z13mul_mat_vec_qIL9ggml_type29ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type29ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type29ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type29ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type29ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type29ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type29ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type29ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type29ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type29ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type29ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type29ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type29ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type29ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
cvta.to.global.u64 %rd1, %rd20;
mov.u32 %r1, %tid.y;
shl.b32 %r35, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r35, %r2;
mov.u32 %r36, %ctaid.x;
shl.b32 %r4, %r36, 1;
mov.u32 %r5, %ctaid.y;
mov.u32 %r6, %ctaid.z;
shr.s32 %r37, %r3, 31;
shr.u32 %r38, %r37, 29;
add.s32 %r39, %r3, %r38;
shr.s32 %r491, %r39, 3;
shr.u32 %r40, %r19, 8;
setp.ge.s32 %p3, %r491, %r40;
mov.f32 %f186, 0f00000000;
mov.f32 %f187, %f186;
mov.f32 %f188, %f186;
mov.f32 %f189, %f186;
@%p3 bra $L__BB269_3;

mul.lo.s32 %r41, %r5, %r27;
cvt.u64.u32 %rd21, %r41;
mul.lo.s32 %r42, %r6, %r33;
cvt.u64.u32 %rd22, %r42;
add.s64 %rd2, %rd22, %rd21;
mul.hi.u32 %r43, %r6, %r29;
add.s32 %r44, %r43, %r6;
shr.u32 %r45, %r44, %r30;
mul.lo.s32 %r46, %r45, %r32;
mul.hi.u32 %r47, %r5, %r23;
add.s32 %r48, %r47, %r5;
shr.u32 %r49, %r48, %r24;
mul.lo.s32 %r50, %r49, %r26;
cvta.to.global.u64 %rd74, %rd18;
and.b32 %r54, %r39, -8;
sub.s32 %r55, %r3, %r54;
cvt.s64.s32 %rd4, %r55;
shl.b32 %r56, %r55, 1;
shr.u32 %r57, %r55, 31;
add.s32 %r58, %r55, %r57;
shr.s32 %r59, %r58, 1;
and.b32 %r60, %r58, -2;
sub.s32 %r61, %r55, %r60;
mul.lo.s32 %r8, %r61, 6;
add.s32 %r62, %r4, 1;
cvt.s64.s32 %rd23, %r56;
shl.b32 %r489, %r491, 3;
add.s32 %r490, %r21, %r489;
add.s32 %r63, %r491, %r46;
add.s32 %r64, %r63, %r50;
mad.lo.s32 %r65, %r62, %r20, %r64;
mul.wide.s32 %rd5, %r65, 56;
mad.lo.s32 %r66, %r4, %r20, %r491;
add.s32 %r67, %r66, %r46;
add.s32 %r68, %r67, %r50;
mul.wide.s32 %rd6, %r68, 56;
mul.wide.s32 %rd24, %r55, 4;
add.s64 %rd7, %rd6, %rd24;
mul.wide.s32 %rd25, %r59, 2;
add.s64 %rd26, %rd6, %rd25;
add.s64 %rd8, %rd26, 48;
add.s64 %rd27, %rd5, %rd25;
add.s64 %rd9, %rd27, 48;
add.s64 %rd28, %rd6, %rd23;
add.s64 %rd10, %rd28, 33;
add.s64 %rd29, %rd5, %rd23;
add.s64 %rd11, %rd29, 32;
add.s64 %rd12, %rd5, %rd24;
cvta.to.global.u64 %rd13, %rd19;
mov.f32 %f186, 0f00000000;

$L__BB269_2:
cvt.u64.u32 %rd30, %r489;
add.s64 %rd31, %rd2, %rd30;
add.s64 %rd32, %rd31, %rd4;
mul.lo.s64 %rd33, %rd32, 36;
add.s64 %rd34, %rd74, %rd7;
ld.global.nc.u32 %r329, [%rd34];
shr.u32 %r330, %r329, 8;
shr.u32 %r331, %r329, 16;
shr.u32 %r332, %r329, 24;
add.s64 %rd35, %rd74, %rd10;
ld.global.nc.u8 %rs5, [%rd35+-1];
and.b16 %rs6, %rs5, 240;
cvt.u32.u16 %r333, %rs5;
and.b32 %r334, %r329, 255;
and.b32 %r335, %r333, 7;
bfi.b32 %r336, %r335, %r334, 8, 3;
mul.wide.u32 %rd36, %r336, 4;
mov.u64 %rd37, iq1s_grid_gpu;
add.s64 %rd38, %rd37, %rd36;
ld.global.nc.u32 %r337, [%rd38];
and.b32 %r200, %r337, 252645135;
shr.u32 %r338, %r337, 4;
and.b32 %r204, %r338, 252645135;
add.s64 %rd39, %rd13, %rd33;
ld.global.nc.u32 %r71, [%rd39+4];
ld.global.nc.u32 %r75, [%rd39+8];
mov.u32 %r323, 0;

dp4a.s32.s32 %r69, %r200, %r71, %r323;


dp4a.s32.s32 %r73, %r204, %r75, %r69;

and.b32 %r339, %r333, 8;
cvt.rn.f32.s32 %f33, %r339;
mul.ftz.f32 %f34, %f33, 0f3D000000;
mov.f32 %f35, 0fBF600000;
sub.ftz.f32 %f36, %f35, %f34;
mov.u32 %r326, 16843009;

dp4a.s32.s32 %r77, %r71, %r326, %r323;


dp4a.s32.s32 %r81, %r75, %r326, %r77;

cvt.rn.f32.s32 %f37, %r81;
fma.rn.ftz.f32 %f38, %f36, %f37, 0f00000000;
shr.u16 %rs7, %rs6, 4;
cvt.u32.u16 %r340, %rs7;
and.b32 %r341, %r330, 255;
and.b32 %r342, %r340, 7;
bfi.b32 %r343, %r342, %r341, 8, 3;
mul.wide.u32 %rd40, %r343, 4;
add.s64 %rd41, %rd37, %rd40;
ld.global.nc.u32 %r344, [%rd41];
and.b32 %r216, %r344, 252645135;
shr.u32 %r345, %r344, 4;
and.b32 %r220, %r345, 252645135;
ld.global.nc.u32 %r87, [%rd39+12];
ld.global.nc.u32 %r91, [%rd39+16];

dp4a.s32.s32 %r85, %r216, %r87, %r73;


dp4a.s32.s32 %r89, %r220, %r91, %r85;

and.b32 %r346, %r340, 8;
cvt.rn.f32.s32 %f39, %r346;
mul.ftz.f32 %f40, %f39, 0f3D000000;
sub.ftz.f32 %f41, %f35, %f40;

dp4a.s32.s32 %r93, %r87, %r326, %r323;


dp4a.s32.s32 %r97, %r91, %r326, %r93;

cvt.rn.f32.s32 %f42, %r97;
fma.rn.ftz.f32 %f43, %f41, %f42, %f38;
ld.global.nc.u8 %rs8, [%rd35];
and.b16 %rs9, %rs8, 240;
cvt.u32.u16 %r347, %rs8;
and.b32 %r348, %r331, 255;
and.b32 %r349, %r347, 7;
bfi.b32 %r350, %r349, %r348, 8, 3;
mul.wide.u32 %rd42, %r350, 4;
add.s64 %rd43, %rd37, %rd42;
ld.global.nc.u32 %r351, [%rd43];
and.b32 %r232, %r351, 252645135;
shr.u32 %r352, %r351, 4;
and.b32 %r236, %r352, 252645135;
ld.global.nc.u32 %r103, [%rd39+20];
ld.global.nc.u32 %r107, [%rd39+24];

dp4a.s32.s32 %r101, %r232, %r103, %r323;


dp4a.s32.s32 %r105, %r236, %r107, %r101;

and.b32 %r353, %r347, 8;
cvt.rn.f32.s32 %f44, %r353;
mul.ftz.f32 %f45, %f44, 0f3D000000;
sub.ftz.f32 %f46, %f35, %f45;

dp4a.s32.s32 %r109, %r103, %r326, %r323;


dp4a.s32.s32 %r113, %r107, %r326, %r109;

cvt.rn.f32.s32 %f47, %r113;
fma.rn.ftz.f32 %f48, %f46, %f47, 0f00000000;
shr.u16 %rs10, %rs9, 4;
cvt.u32.u16 %r354, %rs10;
mul.wide.u16 %r355, %rs10, 256;
and.b32 %r356, %r355, 1792;
or.b32 %r357, %r356, %r332;
mul.wide.u32 %rd44, %r357, 4;
add.s64 %rd45, %rd37, %rd44;
ld.global.nc.u32 %r358, [%rd45];
and.b32 %r248, %r358, 252645135;
shr.u32 %r359, %r358, 4;
and.b32 %r252, %r359, 252645135;
ld.global.nc.u32 %r119, [%rd39+28];
ld.global.nc.u32 %r123, [%rd39+32];

dp4a.s32.s32 %r117, %r248, %r119, %r105;


dp4a.s32.s32 %r121, %r252, %r123, %r117;

and.b32 %r360, %r354, 8;
cvt.rn.f32.s32 %f49, %r360;
mul.ftz.f32 %f50, %f49, 0f3D000000;
sub.ftz.f32 %f51, %f35, %f50;

dp4a.s32.s32 %r125, %r119, %r326, %r323;


dp4a.s32.s32 %r129, %r123, %r326, %r125;

cvt.rn.f32.s32 %f52, %r129;
fma.rn.ftz.f32 %f53, %f51, %f52, %f48;
add.s64 %rd46, %rd74, %rd6;
ld.global.nc.u16 %rs11, [%rd46+48];
shr.u16 %rs12, %rs11, 12;
ld.global.nc.u16 %rs13, [%rd46+50];
shr.u16 %rs14, %rs13, 8;
and.b16 %rs15, %rs14, 240;
or.b16 %rs16, %rs15, %rs12;
ld.global.nc.u16 %rs17, [%rd46+52];
shr.u16 %rs18, %rs17, 4;
and.b16 %rs19, %rs18, 3840;
or.b16 %rs20, %rs16, %rs19;
ld.global.nc.u16 %rs21, [%rd46+54];
and.b16 %rs22, %rs21, -4096;
or.b16 %rs3, %rs20, %rs22;

{ cvt.f32.f16 %f25, %rs3;}


ld.global.nc.u32 %r133, [%rd39];

{.reg .f16 low,high;
mov.b32 {low,high},%r133;
cvt.f32.f16 %f26, low;}


mul.ftz.f32 %f54, %f25, %f26;
add.s64 %rd47, %rd74, %rd8;
ld.global.nc.u16 %rs23, [%rd47];
cvt.u32.u16 %r361, %rs23;
shr.u32 %r362, %r361, %r8;
shl.b32 %r363, %r362, 1;
and.b32 %r364, %r363, 14;
or.b32 %r365, %r364, 1;
shr.u32 %r366, %r362, 2;
and.b32 %r367, %r366, 14;
or.b32 %r368, %r367, 1;
cvt.rn.f32.s32 %f55, %r89;
add.ftz.f32 %f56, %f43, %f55;
cvt.rn.f32.s32 %f57, %r365;
cvt.rn.f32.s32 %f58, %r121;
add.ftz.f32 %f59, %f53, %f58;
cvt.rn.f32.s32 %f60, %r368;
mul.ftz.f32 %f61, %f59, %f60;
fma.rn.ftz.f32 %f62, %f56, %f57, %f61;
fma.rn.ftz.f32 %f189, %f54, %f62, %f189;
add.s64 %rd48, %rd74, %rd12;
ld.global.nc.u32 %r369, [%rd48];
shr.u32 %r370, %r369, 8;
shr.u32 %r371, %r369, 16;
shr.u32 %r372, %r369, 24;
add.s64 %rd49, %rd74, %rd11;
ld.global.nc.u8 %rs24, [%rd49];
and.b16 %rs25, %rs24, 240;
cvt.u32.u16 %r373, %rs24;
and.b32 %r374, %r369, 255;
and.b32 %r375, %r373, 7;
bfi.b32 %r376, %r375, %r374, 8, 3;
mul.wide.u32 %rd50, %r376, 4;
add.s64 %rd51, %rd37, %rd50;
ld.global.nc.u32 %r377, [%rd51];
and.b32 %r265, %r377, 252645135;
shr.u32 %r378, %r377, 4;
and.b32 %r269, %r378, 252645135;

dp4a.s32.s32 %r134, %r265, %r71, %r323;


dp4a.s32.s32 %r138, %r269, %r75, %r134;

and.b32 %r379, %r373, 8;
cvt.rn.f32.s32 %f63, %r379;
mul.ftz.f32 %f64, %f63, 0f3D000000;
sub.ftz.f32 %f65, %f35, %f64;

dp4a.s32.s32 %r142, %r71, %r326, %r323;


dp4a.s32.s32 %r146, %r75, %r326, %r142;

cvt.rn.f32.s32 %f66, %r146;
fma.rn.ftz.f32 %f67, %f65, %f66, 0f00000000;
shr.u16 %rs26, %rs25, 4;
cvt.u32.u16 %r380, %rs26;
and.b32 %r381, %r370, 255;
and.b32 %r382, %r380, 7;
bfi.b32 %r383, %r382, %r381, 8, 3;
mul.wide.u32 %rd52, %r383, 4;
add.s64 %rd53, %rd37, %rd52;
ld.global.nc.u32 %r384, [%rd53];
and.b32 %r281, %r384, 252645135;
shr.u32 %r385, %r384, 4;
and.b32 %r285, %r385, 252645135;

dp4a.s32.s32 %r150, %r281, %r87, %r138;


dp4a.s32.s32 %r154, %r285, %r91, %r150;

and.b32 %r386, %r380, 8;
cvt.rn.f32.s32 %f68, %r386;
mul.ftz.f32 %f69, %f68, 0f3D000000;
sub.ftz.f32 %f70, %f35, %f69;

dp4a.s32.s32 %r158, %r87, %r326, %r323;


dp4a.s32.s32 %r162, %r91, %r326, %r158;

cvt.rn.f32.s32 %f71, %r162;
fma.rn.ftz.f32 %f72, %f70, %f71, %f67;
ld.global.nc.u8 %rs27, [%rd49+1];
and.b16 %rs28, %rs27, 240;
cvt.u32.u16 %r387, %rs27;
and.b32 %r388, %r371, 255;
and.b32 %r389, %r387, 7;
bfi.b32 %r390, %r389, %r388, 8, 3;
mul.wide.u32 %rd54, %r390, 4;
add.s64 %rd55, %rd37, %rd54;
ld.global.nc.u32 %r391, [%rd55];
and.b32 %r297, %r391, 252645135;
shr.u32 %r392, %r391, 4;
and.b32 %r301, %r392, 252645135;

dp4a.s32.s32 %r166, %r297, %r103, %r323;


dp4a.s32.s32 %r170, %r301, %r107, %r166;

and.b32 %r393, %r387, 8;
cvt.rn.f32.s32 %f73, %r393;
mul.ftz.f32 %f74, %f73, 0f3D000000;
sub.ftz.f32 %f75, %f35, %f74;

dp4a.s32.s32 %r174, %r103, %r326, %r323;


dp4a.s32.s32 %r178, %r107, %r326, %r174;

cvt.rn.f32.s32 %f76, %r178;
fma.rn.ftz.f32 %f77, %f75, %f76, 0f00000000;
shr.u16 %rs29, %rs28, 4;
cvt.u32.u16 %r394, %rs29;
mul.wide.u16 %r395, %rs29, 256;
and.b32 %r396, %r395, 1792;
or.b32 %r397, %r396, %r372;
mul.wide.u32 %rd56, %r397, 4;
add.s64 %rd57, %rd37, %rd56;
ld.global.nc.u32 %r398, [%rd57];
and.b32 %r313, %r398, 252645135;
shr.u32 %r399, %r398, 4;
and.b32 %r317, %r399, 252645135;

dp4a.s32.s32 %r182, %r313, %r119, %r170;


dp4a.s32.s32 %r186, %r317, %r123, %r182;

and.b32 %r400, %r394, 8;
cvt.rn.f32.s32 %f78, %r400;
mul.ftz.f32 %f79, %f78, 0f3D000000;
sub.ftz.f32 %f80, %f35, %f79;

dp4a.s32.s32 %r190, %r119, %r326, %r323;


dp4a.s32.s32 %r194, %r123, %r326, %r190;

cvt.rn.f32.s32 %f81, %r194;
fma.rn.ftz.f32 %f82, %f80, %f81, %f77;
add.s64 %rd58, %rd74, %rd5;
ld.global.nc.u16 %rs30, [%rd58+48];
shr.u16 %rs31, %rs30, 12;
ld.global.nc.u16 %rs32, [%rd58+50];
shr.u16 %rs33, %rs32, 8;
and.b16 %rs34, %rs33, 240;
or.b16 %rs35, %rs34, %rs31;
ld.global.nc.u16 %rs36, [%rd58+52];
shr.u16 %rs37, %rs36, 4;
and.b16 %rs38, %rs37, 3840;
or.b16 %rs39, %rs35, %rs38;
ld.global.nc.u16 %rs40, [%rd58+54];
and.b16 %rs41, %rs40, -4096;
or.b16 %rs4, %rs39, %rs41;

{ cvt.f32.f16 %f27, %rs4;}



{.reg .f16 low,high;
mov.b32 {low,high},%r133;
cvt.f32.f16 %f28, low;}


mul.ftz.f32 %f83, %f27, %f28;
add.s64 %rd59, %rd74, %rd9;
ld.global.nc.u16 %rs42, [%rd59];
cvt.u32.u16 %r401, %rs42;
shr.u32 %r402, %r401, %r8;
shl.b32 %r403, %r402, 1;
and.b32 %r404, %r403, 14;
or.b32 %r405, %r404, 1;
shr.u32 %r406, %r402, 2;
and.b32 %r407, %r406, 14;
or.b32 %r408, %r407, 1;
cvt.rn.f32.s32 %f84, %r154;
add.ftz.f32 %f85, %f72, %f84;
cvt.rn.f32.s32 %f86, %r405;
cvt.rn.f32.s32 %f87, %r186;
add.ftz.f32 %f88, %f82, %f87;
cvt.rn.f32.s32 %f89, %r408;
mul.ftz.f32 %f90, %f88, %f89;
fma.rn.ftz.f32 %f91, %f85, %f86, %f90;
fma.rn.ftz.f32 %f188, %f83, %f91, %f188;
cvt.u64.u32 %rd60, %r490;
add.s64 %rd61, %rd2, %rd60;
add.s64 %rd62, %rd61, %rd4;
mul.lo.s64 %rd63, %rd62, 36;
add.s64 %rd64, %rd13, %rd63;
ld.global.nc.u32 %r201, [%rd64+4];
ld.global.nc.u32 %r205, [%rd64+8];

dp4a.s32.s32 %r199, %r200, %r201, %r323;


dp4a.s32.s32 %r203, %r204, %r205, %r199;


dp4a.s32.s32 %r207, %r201, %r326, %r323;


dp4a.s32.s32 %r211, %r205, %r326, %r207;

cvt.rn.f32.s32 %f92, %r211;
fma.rn.ftz.f32 %f93, %f36, %f92, 0f00000000;
ld.global.nc.u32 %r217, [%rd64+12];
ld.global.nc.u32 %r221, [%rd64+16];

dp4a.s32.s32 %r215, %r216, %r217, %r203;


dp4a.s32.s32 %r219, %r220, %r221, %r215;


dp4a.s32.s32 %r223, %r217, %r326, %r323;


dp4a.s32.s32 %r227, %r221, %r326, %r223;

cvt.rn.f32.s32 %f94, %r227;
fma.rn.ftz.f32 %f95, %f41, %f94, %f93;
ld.global.nc.u32 %r233, [%rd64+20];
ld.global.nc.u32 %r237, [%rd64+24];

dp4a.s32.s32 %r231, %r232, %r233, %r323;


dp4a.s32.s32 %r235, %r236, %r237, %r231;


dp4a.s32.s32 %r239, %r233, %r326, %r323;


dp4a.s32.s32 %r243, %r237, %r326, %r239;

cvt.rn.f32.s32 %f96, %r243;
fma.rn.ftz.f32 %f97, %f46, %f96, 0f00000000;
ld.global.nc.u32 %r249, [%rd64+28];
ld.global.nc.u32 %r253, [%rd64+32];

dp4a.s32.s32 %r247, %r248, %r249, %r235;


dp4a.s32.s32 %r251, %r252, %r253, %r247;


dp4a.s32.s32 %r255, %r249, %r326, %r323;


dp4a.s32.s32 %r259, %r253, %r326, %r255;

cvt.rn.f32.s32 %f98, %r259;
fma.rn.ftz.f32 %f99, %f51, %f98, %f97;

{ cvt.f32.f16 %f29, %rs3;}


ld.global.nc.u32 %r263, [%rd64];

{.reg .f16 low,high;
mov.b32 {low,high},%r263;
cvt.f32.f16 %f30, low;}


mul.ftz.f32 %f100, %f29, %f30;
cvt.rn.f32.s32 %f101, %r219;
add.ftz.f32 %f102, %f95, %f101;
cvt.rn.f32.s32 %f103, %r251;
add.ftz.f32 %f104, %f99, %f103;
mul.ftz.f32 %f105, %f104, %f60;
fma.rn.ftz.f32 %f106, %f102, %f57, %f105;
fma.rn.ftz.f32 %f187, %f100, %f106, %f187;

dp4a.s32.s32 %r264, %r265, %r201, %r323;


dp4a.s32.s32 %r268, %r269, %r205, %r264;


dp4a.s32.s32 %r272, %r201, %r326, %r323;


dp4a.s32.s32 %r276, %r205, %r326, %r272;

cvt.rn.f32.s32 %f107, %r276;
fma.rn.ftz.f32 %f108, %f65, %f107, 0f00000000;

dp4a.s32.s32 %r280, %r281, %r217, %r268;


dp4a.s32.s32 %r284, %r285, %r221, %r280;


dp4a.s32.s32 %r288, %r217, %r326, %r323;


dp4a.s32.s32 %r292, %r221, %r326, %r288;

cvt.rn.f32.s32 %f109, %r292;
fma.rn.ftz.f32 %f110, %f70, %f109, %f108;

dp4a.s32.s32 %r296, %r297, %r233, %r323;


dp4a.s32.s32 %r300, %r301, %r237, %r296;


dp4a.s32.s32 %r304, %r233, %r326, %r323;


dp4a.s32.s32 %r308, %r237, %r326, %r304;

cvt.rn.f32.s32 %f111, %r308;
fma.rn.ftz.f32 %f112, %f75, %f111, 0f00000000;

dp4a.s32.s32 %r312, %r313, %r249, %r300;


dp4a.s32.s32 %r316, %r317, %r253, %r312;


dp4a.s32.s32 %r320, %r249, %r326, %r323;


dp4a.s32.s32 %r324, %r253, %r326, %r320;

cvt.rn.f32.s32 %f113, %r324;
fma.rn.ftz.f32 %f114, %f80, %f113, %f112;

{ cvt.f32.f16 %f31, %rs4;}



{.reg .f16 low,high;
mov.b32 {low,high},%r263;
cvt.f32.f16 %f32, low;}


mul.ftz.f32 %f115, %f31, %f32;
cvt.rn.f32.s32 %f116, %r284;
add.ftz.f32 %f117, %f110, %f116;
cvt.rn.f32.s32 %f118, %r316;
add.ftz.f32 %f119, %f114, %f118;
mul.ftz.f32 %f120, %f119, %f89;
fma.rn.ftz.f32 %f121, %f117, %f86, %f120;
fma.rn.ftz.f32 %f186, %f115, %f121, %f186;
add.s32 %r490, %r490, 128;
add.s32 %r489, %r489, 128;
add.s64 %rd74, %rd74, 896;
add.s32 %r491, %r491, 16;
setp.lt.s32 %p4, %r491, %r40;
@%p4 bra $L__BB269_2;

$L__BB269_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB269_5;

shl.b32 %r410, %r1, 9;
mov.u32 %r411, _ZZ13mul_mat_vec_qIL9ggml_type29ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r412, %r411, %r410;
shl.b32 %r413, %r2, 2;
add.s32 %r414, %r412, %r413;
st.shared.f32 [%r414+-512], %f189;
st.shared.f32 [%r414+-384], %f188;
st.shared.f32 [%r414+-256], %f187;
st.shared.f32 [%r414+-128], %f186;

$L__BB269_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB269_14;

mad.lo.s32 %r415, %r5, %r28, %r4;
mad.lo.s32 %r416, %r6, %r34, %r415;
cvt.u64.u32 %rd16, %r416;
shl.b32 %r417, %r2, 2;
mov.u32 %r418, _ZZ13mul_mat_vec_qIL9ggml_type29ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r17, %r418, %r417;
mov.u32 %r419, 2;
ld.shared.f32 %f122, [%r17];
add.ftz.f32 %f123, %f122, %f189;
ld.shared.f32 %f124, [%r17+512];
add.ftz.f32 %f125, %f124, %f123;
ld.shared.f32 %f126, [%r17+1024];
add.ftz.f32 %f127, %f126, %f125;
mov.b32 %r420, %f127;
mov.u32 %r421, 31;
mov.u32 %r422, 16;
mov.u32 %r423, -1;
shfl.sync.bfly.b32 %r424|%p7, %r420, %r422, %r421, %r423;
mov.b32 %f128, %r424;
add.ftz.f32 %f129, %f127, %f128;
mov.b32 %r425, %f129;
mov.u32 %r426, 8;
shfl.sync.bfly.b32 %r427|%p8, %r425, %r426, %r421, %r423;
mov.b32 %f130, %r427;
add.ftz.f32 %f131, %f129, %f130;
mov.b32 %r428, %f131;
mov.u32 %r429, 4;
shfl.sync.bfly.b32 %r430|%p9, %r428, %r429, %r421, %r423;
mov.b32 %f132, %r430;
add.ftz.f32 %f133, %f131, %f132;
mov.b32 %r431, %f133;
shfl.sync.bfly.b32 %r432|%p10, %r431, %r419, %r421, %r423;
mov.b32 %f134, %r432;
add.ftz.f32 %f135, %f133, %f134;
mov.b32 %r433, %f135;
mov.u32 %r434, 1;
shfl.sync.bfly.b32 %r435|%p11, %r433, %r434, %r421, %r423;
mov.b32 %f136, %r435;
add.ftz.f32 %f13, %f135, %f136;
setp.lt.u32 %p12, %r4, %r22;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
mul.wide.u32 %rd65, %r416, 4;
add.s64 %rd17, %rd1, %rd65;
not.pred %p14, %p1;
@%p14 bra $L__BB269_8;

st.global.f32 [%rd17], %f13;

$L__BB269_8:
ld.shared.f32 %f137, [%r17+128];
add.ftz.f32 %f138, %f137, %f188;
ld.shared.f32 %f139, [%r17+640];
add.ftz.f32 %f140, %f139, %f138;
ld.shared.f32 %f141, [%r17+1152];
add.ftz.f32 %f142, %f141, %f140;
mov.b32 %r436, %f142;
shfl.sync.bfly.b32 %r440|%p15, %r436, %r422, %r421, %r423;
mov.b32 %f143, %r440;
add.ftz.f32 %f144, %f142, %f143;
mov.b32 %r441, %f144;
shfl.sync.bfly.b32 %r443|%p16, %r441, %r426, %r421, %r423;
mov.b32 %f145, %r443;
add.ftz.f32 %f146, %f144, %f145;
mov.b32 %r444, %f146;
shfl.sync.bfly.b32 %r446|%p17, %r444, %r429, %r421, %r423;
mov.b32 %f147, %r446;
add.ftz.f32 %f148, %f146, %f147;
mov.b32 %r447, %f148;
shfl.sync.bfly.b32 %r449|%p18, %r447, %r419, %r421, %r423;
mov.b32 %f149, %r449;
add.ftz.f32 %f150, %f148, %f149;
mov.b32 %r450, %f150;
shfl.sync.bfly.b32 %r452|%p19, %r450, %r434, %r421, %r423;
mov.b32 %f151, %r452;
add.ftz.f32 %f14, %f150, %f151;
add.s32 %r453, %r4, 1;
setp.lt.u32 %p20, %r453, %r22;
setp.eq.s32 %p21, %r2, 1;
and.pred %p2, %p21, %p20;
not.pred %p22, %p2;
@%p22 bra $L__BB269_10;

st.global.f32 [%rd17+4], %f14;

$L__BB269_10:
ld.shared.f32 %f152, [%r17+256];
add.ftz.f32 %f153, %f152, %f187;
ld.shared.f32 %f154, [%r17+768];
add.ftz.f32 %f155, %f154, %f153;
ld.shared.f32 %f156, [%r17+1280];
add.ftz.f32 %f157, %f156, %f155;
mov.b32 %r454, %f157;
mov.u32 %r455, 31;
mov.u32 %r456, 16;
mov.u32 %r457, -1;
shfl.sync.bfly.b32 %r458|%p23, %r454, %r456, %r455, %r457;
mov.b32 %f158, %r458;
add.ftz.f32 %f159, %f157, %f158;
mov.b32 %r459, %f159;
mov.u32 %r460, 8;
shfl.sync.bfly.b32 %r461|%p24, %r459, %r460, %r455, %r457;
mov.b32 %f160, %r461;
add.ftz.f32 %f161, %f159, %f160;
mov.b32 %r462, %f161;
mov.u32 %r463, 4;
shfl.sync.bfly.b32 %r464|%p25, %r462, %r463, %r455, %r457;
mov.b32 %f162, %r464;
add.ftz.f32 %f163, %f161, %f162;
mov.b32 %r465, %f163;
mov.u32 %r466, 2;
shfl.sync.bfly.b32 %r467|%p26, %r465, %r466, %r455, %r457;
mov.b32 %f164, %r467;
add.ftz.f32 %f165, %f163, %f164;
mov.b32 %r468, %f165;
mov.u32 %r469, 1;
shfl.sync.bfly.b32 %r470|%p27, %r468, %r469, %r455, %r457;
mov.b32 %f166, %r470;
add.ftz.f32 %f15, %f165, %f166;
@%p14 bra $L__BB269_12;

cvt.u64.u32 %rd66, %r22;
add.s64 %rd67, %rd66, %rd16;
shl.b64 %rd68, %rd67, 2;
add.s64 %rd69, %rd1, %rd68;
st.global.f32 [%rd69], %f15;

$L__BB269_12:
ld.shared.f32 %f167, [%r17+384];
add.ftz.f32 %f168, %f167, %f186;
ld.shared.f32 %f169, [%r17+896];
add.ftz.f32 %f170, %f169, %f168;
ld.shared.f32 %f171, [%r17+1408];
add.ftz.f32 %f172, %f171, %f170;
mov.b32 %r471, %f172;
shfl.sync.bfly.b32 %r475|%p29, %r471, %r456, %r455, %r457;
mov.b32 %f173, %r475;
add.ftz.f32 %f174, %f172, %f173;
mov.b32 %r476, %f174;
shfl.sync.bfly.b32 %r478|%p30, %r476, %r460, %r455, %r457;
mov.b32 %f175, %r478;
add.ftz.f32 %f176, %f174, %f175;
mov.b32 %r479, %f176;
shfl.sync.bfly.b32 %r481|%p31, %r479, %r463, %r455, %r457;
mov.b32 %f177, %r481;
add.ftz.f32 %f178, %f176, %f177;
mov.b32 %r482, %f178;
shfl.sync.bfly.b32 %r484|%p32, %r482, %r466, %r455, %r457;
mov.b32 %f179, %r484;
add.ftz.f32 %f180, %f178, %f179;
mov.b32 %r485, %f180;
shfl.sync.bfly.b32 %r487|%p33, %r485, %r469, %r455, %r457;
mov.b32 %f181, %r487;
add.ftz.f32 %f16, %f180, %f181;
@%p22 bra $L__BB269_14;

add.s32 %r488, %r22, 1;
cvt.u64.u32 %rd70, %r488;
add.s64 %rd71, %rd70, %rd16;
shl.b64 %rd72, %rd71, 2;
add.s64 %rd73, %rd1, %rd72;
st.global.f32 [%rd73], %f16;

$L__BB269_14:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type29ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type29ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type29ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type29ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type29ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type29ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type29ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type29ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type29ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<47>;
.reg .b16 %rs<45>;
.reg .f32 %f<270>;
.reg .b32 %r<661>;
.reg .b64 %rd<84>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type29ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[2304];

ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type29ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd20, [_Z13mul_mat_vec_qIL9ggml_type29ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd21, [_Z13mul_mat_vec_qIL9ggml_type29ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type29ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type29ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type29ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type29ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type29ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type29ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type29ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type29ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type29ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type29ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type29ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type29ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type29ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type29ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
cvta.to.global.u64 %rd1, %rd21;
mov.u32 %r1, %tid.y;
shl.b32 %r38, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r38, %r2;
mov.u32 %r39, %ctaid.x;
shl.b32 %r4, %r39, 1;
mov.u32 %r5, %ctaid.y;
shr.s32 %r40, %r3, 31;
shr.u32 %r41, %r40, 29;
add.s32 %r42, %r3, %r41;
shr.s32 %r660, %r42, 3;
shr.u32 %r7, %r37, 8;
setp.ge.s32 %p3, %r660, %r7;
mov.f32 %f264, 0f00000000;
mov.f32 %f265, %f264;
mov.f32 %f266, %f264;
mov.f32 %f267, %f264;
mov.f32 %f268, %f264;
mov.f32 %f269, %f264;
@%p3 bra $L__BB270_3;

mov.u32 %r43, %ctaid.z;
mul.lo.s32 %r44, %r5, %r29;
cvt.u64.u32 %rd22, %r44;
mul.lo.s32 %r45, %r43, %r35;
cvt.u64.u32 %rd23, %r45;
add.s64 %rd2, %rd23, %rd22;
mul.hi.u32 %r46, %r43, %r31;
add.s32 %r47, %r46, %r43;
shr.u32 %r48, %r47, %r32;
mul.lo.s32 %r49, %r48, %r34;
mul.hi.u32 %r50, %r5, %r25;
add.s32 %r51, %r50, %r5;
shr.u32 %r52, %r51, %r26;
mul.lo.s32 %r53, %r52, %r28;
cvta.to.global.u64 %rd83, %rd19;
and.b32 %r57, %r42, -8;
sub.s32 %r58, %r3, %r57;
cvt.s64.s32 %rd4, %r58;
shl.b32 %r59, %r58, 1;
shr.u32 %r60, %r58, 31;
add.s32 %r61, %r58, %r60;
shr.s32 %r62, %r61, 1;
and.b32 %r63, %r61, -2;
sub.s32 %r64, %r58, %r63;
mul.lo.s32 %r8, %r64, 6;
add.s32 %r65, %r4, 1;
cvt.s64.s32 %rd24, %r59;
shl.b32 %r66, %r23, 1;
shl.b32 %r657, %r660, 3;
add.s32 %r659, %r657, %r66;
add.s32 %r658, %r23, %r657;
add.s32 %r67, %r660, %r49;
add.s32 %r68, %r67, %r53;
mad.lo.s32 %r69, %r65, %r22, %r68;
mul.wide.s32 %rd5, %r69, 56;
mad.lo.s32 %r70, %r4, %r22, %r660;
add.s32 %r71, %r70, %r49;
add.s32 %r72, %r71, %r53;
mul.wide.s32 %rd6, %r72, 56;
mul.wide.s32 %rd25, %r58, 4;
add.s64 %rd7, %rd6, %rd25;
mul.wide.s32 %rd26, %r62, 2;
add.s64 %rd27, %rd6, %rd26;
add.s64 %rd8, %rd27, 48;
add.s64 %rd28, %rd5, %rd26;
add.s64 %rd9, %rd28, 48;
add.s64 %rd29, %rd6, %rd24;
add.s64 %rd10, %rd29, 33;
add.s64 %rd30, %rd5, %rd24;
add.s64 %rd11, %rd30, 32;
add.s64 %rd12, %rd5, %rd25;
cvta.to.global.u64 %rd13, %rd20;
mov.f32 %f264, 0f00000000;

$L__BB270_2:
cvt.u64.u32 %rd31, %r657;
add.s64 %rd32, %rd2, %rd31;
add.s64 %rd33, %rd32, %rd4;
mul.lo.s64 %rd34, %rd33, 36;
add.s64 %rd35, %rd83, %rd7;
ld.global.nc.u32 %r463, [%rd35];
shr.u32 %r464, %r463, 8;
shr.u32 %r465, %r463, 16;
shr.u32 %r466, %r463, 24;
add.s64 %rd36, %rd83, %rd10;
ld.global.nc.u8 %rs7, [%rd36+-1];
and.b16 %rs8, %rs7, 240;
cvt.u32.u16 %r467, %rs7;
and.b32 %r468, %r463, 255;
and.b32 %r469, %r467, 7;
bfi.b32 %r470, %r469, %r468, 8, 3;
mul.wide.u32 %rd37, %r470, 4;
mov.u64 %rd38, iq1s_grid_gpu;
add.s64 %rd39, %rd38, %rd37;
ld.global.nc.u32 %r471, [%rd39];
and.b32 %r334, %r471, 252645135;
shr.u32 %r472, %r471, 4;
and.b32 %r338, %r472, 252645135;
add.s64 %rd40, %rd13, %rd34;
ld.global.nc.u32 %r75, [%rd40+4];
ld.global.nc.u32 %r79, [%rd40+8];
mov.u32 %r457, 0;

dp4a.s32.s32 %r73, %r334, %r75, %r457;


dp4a.s32.s32 %r77, %r338, %r79, %r73;

and.b32 %r473, %r467, 8;
cvt.rn.f32.s32 %f49, %r473;
mul.ftz.f32 %f50, %f49, 0f3D000000;
mov.f32 %f51, 0fBF600000;
sub.ftz.f32 %f52, %f51, %f50;
mov.u32 %r460, 16843009;

dp4a.s32.s32 %r81, %r75, %r460, %r457;


dp4a.s32.s32 %r85, %r79, %r460, %r81;

cvt.rn.f32.s32 %f53, %r85;
fma.rn.ftz.f32 %f54, %f52, %f53, 0f00000000;
shr.u16 %rs9, %rs8, 4;
cvt.u32.u16 %r474, %rs9;
and.b32 %r475, %r464, 255;
and.b32 %r476, %r474, 7;
bfi.b32 %r477, %r476, %r475, 8, 3;
mul.wide.u32 %rd41, %r477, 4;
add.s64 %rd42, %rd38, %rd41;
ld.global.nc.u32 %r478, [%rd42];
and.b32 %r350, %r478, 252645135;
shr.u32 %r479, %r478, 4;
and.b32 %r354, %r479, 252645135;
ld.global.nc.u32 %r91, [%rd40+12];
ld.global.nc.u32 %r95, [%rd40+16];

dp4a.s32.s32 %r89, %r350, %r91, %r77;


dp4a.s32.s32 %r93, %r354, %r95, %r89;

and.b32 %r480, %r474, 8;
cvt.rn.f32.s32 %f55, %r480;
mul.ftz.f32 %f56, %f55, 0f3D000000;
sub.ftz.f32 %f57, %f51, %f56;

dp4a.s32.s32 %r97, %r91, %r460, %r457;


dp4a.s32.s32 %r101, %r95, %r460, %r97;

cvt.rn.f32.s32 %f58, %r101;
fma.rn.ftz.f32 %f59, %f57, %f58, %f54;
ld.global.nc.u8 %rs10, [%rd36];
and.b16 %rs11, %rs10, 240;
cvt.u32.u16 %r481, %rs10;
and.b32 %r482, %r465, 255;
and.b32 %r483, %r481, 7;
bfi.b32 %r484, %r483, %r482, 8, 3;
mul.wide.u32 %rd43, %r484, 4;
add.s64 %rd44, %rd38, %rd43;
ld.global.nc.u32 %r485, [%rd44];
and.b32 %r366, %r485, 252645135;
shr.u32 %r486, %r485, 4;
and.b32 %r370, %r486, 252645135;
ld.global.nc.u32 %r107, [%rd40+20];
ld.global.nc.u32 %r111, [%rd40+24];

dp4a.s32.s32 %r105, %r366, %r107, %r457;


dp4a.s32.s32 %r109, %r370, %r111, %r105;

and.b32 %r487, %r481, 8;
cvt.rn.f32.s32 %f60, %r487;
mul.ftz.f32 %f61, %f60, 0f3D000000;
sub.ftz.f32 %f62, %f51, %f61;

dp4a.s32.s32 %r113, %r107, %r460, %r457;


dp4a.s32.s32 %r117, %r111, %r460, %r113;

cvt.rn.f32.s32 %f63, %r117;
fma.rn.ftz.f32 %f64, %f62, %f63, 0f00000000;
shr.u16 %rs12, %rs11, 4;
cvt.u32.u16 %r488, %rs12;
mul.wide.u16 %r489, %rs12, 256;
and.b32 %r490, %r489, 1792;
or.b32 %r491, %r490, %r466;
mul.wide.u32 %rd45, %r491, 4;
add.s64 %rd46, %rd38, %rd45;
ld.global.nc.u32 %r492, [%rd46];
and.b32 %r382, %r492, 252645135;
shr.u32 %r493, %r492, 4;
and.b32 %r386, %r493, 252645135;
ld.global.nc.u32 %r123, [%rd40+28];
ld.global.nc.u32 %r127, [%rd40+32];

dp4a.s32.s32 %r121, %r382, %r123, %r109;


dp4a.s32.s32 %r125, %r386, %r127, %r121;

and.b32 %r494, %r488, 8;
cvt.rn.f32.s32 %f65, %r494;
mul.ftz.f32 %f66, %f65, 0f3D000000;
sub.ftz.f32 %f67, %f51, %f66;

dp4a.s32.s32 %r129, %r123, %r460, %r457;


dp4a.s32.s32 %r133, %r127, %r460, %r129;

cvt.rn.f32.s32 %f68, %r133;
fma.rn.ftz.f32 %f69, %f67, %f68, %f64;
add.s64 %rd47, %rd83, %rd6;
ld.global.nc.u16 %rs13, [%rd47+48];
shr.u16 %rs14, %rs13, 12;
ld.global.nc.u16 %rs15, [%rd47+50];
shr.u16 %rs16, %rs15, 8;
and.b16 %rs17, %rs16, 240;
or.b16 %rs18, %rs17, %rs14;
ld.global.nc.u16 %rs19, [%rd47+52];
shr.u16 %rs20, %rs19, 4;
and.b16 %rs21, %rs20, 3840;
or.b16 %rs22, %rs18, %rs21;
ld.global.nc.u16 %rs23, [%rd47+54];
and.b16 %rs24, %rs23, -4096;
or.b16 %rs5, %rs22, %rs24;

{ cvt.f32.f16 %f37, %rs5;}


ld.global.nc.u32 %r137, [%rd40];

{.reg .f16 low,high;
mov.b32 {low,high},%r137;
cvt.f32.f16 %f38, low;}


mul.ftz.f32 %f70, %f37, %f38;
add.s64 %rd48, %rd83, %rd8;
ld.global.nc.u16 %rs25, [%rd48];
cvt.u32.u16 %r495, %rs25;
shr.u32 %r496, %r495, %r8;
shl.b32 %r497, %r496, 1;
and.b32 %r498, %r497, 14;
or.b32 %r499, %r498, 1;
shr.u32 %r500, %r496, 2;
and.b32 %r501, %r500, 14;
or.b32 %r502, %r501, 1;
cvt.rn.f32.s32 %f71, %r93;
add.ftz.f32 %f72, %f59, %f71;
cvt.rn.f32.s32 %f73, %r499;
cvt.rn.f32.s32 %f74, %r125;
add.ftz.f32 %f75, %f69, %f74;
cvt.rn.f32.s32 %f76, %r502;
mul.ftz.f32 %f77, %f75, %f76;
fma.rn.ftz.f32 %f78, %f72, %f73, %f77;
fma.rn.ftz.f32 %f269, %f70, %f78, %f269;
add.s64 %rd49, %rd83, %rd12;
ld.global.nc.u32 %r503, [%rd49];
shr.u32 %r504, %r503, 8;
shr.u32 %r505, %r503, 16;
shr.u32 %r506, %r503, 24;
add.s64 %rd50, %rd83, %rd11;
ld.global.nc.u8 %rs26, [%rd50];
and.b16 %rs27, %rs26, 240;
cvt.u32.u16 %r507, %rs26;
and.b32 %r508, %r503, 255;
and.b32 %r509, %r507, 7;
bfi.b32 %r510, %r509, %r508, 8, 3;
mul.wide.u32 %rd51, %r510, 4;
add.s64 %rd52, %rd38, %rd51;
ld.global.nc.u32 %r511, [%rd52];
and.b32 %r399, %r511, 252645135;
shr.u32 %r512, %r511, 4;
and.b32 %r403, %r512, 252645135;

dp4a.s32.s32 %r138, %r399, %r75, %r457;


dp4a.s32.s32 %r142, %r403, %r79, %r138;

and.b32 %r513, %r507, 8;
cvt.rn.f32.s32 %f79, %r513;
mul.ftz.f32 %f80, %f79, 0f3D000000;
sub.ftz.f32 %f81, %f51, %f80;

dp4a.s32.s32 %r146, %r75, %r460, %r457;


dp4a.s32.s32 %r150, %r79, %r460, %r146;

cvt.rn.f32.s32 %f82, %r150;
fma.rn.ftz.f32 %f83, %f81, %f82, 0f00000000;
shr.u16 %rs28, %rs27, 4;
cvt.u32.u16 %r514, %rs28;
and.b32 %r515, %r504, 255;
and.b32 %r516, %r514, 7;
bfi.b32 %r517, %r516, %r515, 8, 3;
mul.wide.u32 %rd53, %r517, 4;
add.s64 %rd54, %rd38, %rd53;
ld.global.nc.u32 %r518, [%rd54];
and.b32 %r415, %r518, 252645135;
shr.u32 %r519, %r518, 4;
and.b32 %r419, %r519, 252645135;

dp4a.s32.s32 %r154, %r415, %r91, %r142;


dp4a.s32.s32 %r158, %r419, %r95, %r154;

and.b32 %r520, %r514, 8;
cvt.rn.f32.s32 %f84, %r520;
mul.ftz.f32 %f85, %f84, 0f3D000000;
sub.ftz.f32 %f86, %f51, %f85;

dp4a.s32.s32 %r162, %r91, %r460, %r457;


dp4a.s32.s32 %r166, %r95, %r460, %r162;

cvt.rn.f32.s32 %f87, %r166;
fma.rn.ftz.f32 %f88, %f86, %f87, %f83;
ld.global.nc.u8 %rs29, [%rd50+1];
and.b16 %rs30, %rs29, 240;
cvt.u32.u16 %r521, %rs29;
and.b32 %r522, %r505, 255;
and.b32 %r523, %r521, 7;
bfi.b32 %r524, %r523, %r522, 8, 3;
mul.wide.u32 %rd55, %r524, 4;
add.s64 %rd56, %rd38, %rd55;
ld.global.nc.u32 %r525, [%rd56];
and.b32 %r431, %r525, 252645135;
shr.u32 %r526, %r525, 4;
and.b32 %r435, %r526, 252645135;

dp4a.s32.s32 %r170, %r431, %r107, %r457;


dp4a.s32.s32 %r174, %r435, %r111, %r170;

and.b32 %r527, %r521, 8;
cvt.rn.f32.s32 %f89, %r527;
mul.ftz.f32 %f90, %f89, 0f3D000000;
sub.ftz.f32 %f91, %f51, %f90;

dp4a.s32.s32 %r178, %r107, %r460, %r457;


dp4a.s32.s32 %r182, %r111, %r460, %r178;

cvt.rn.f32.s32 %f92, %r182;
fma.rn.ftz.f32 %f93, %f91, %f92, 0f00000000;
shr.u16 %rs31, %rs30, 4;
cvt.u32.u16 %r528, %rs31;
mul.wide.u16 %r529, %rs31, 256;
and.b32 %r530, %r529, 1792;
or.b32 %r531, %r530, %r506;
mul.wide.u32 %rd57, %r531, 4;
add.s64 %rd58, %rd38, %rd57;
ld.global.nc.u32 %r532, [%rd58];
and.b32 %r447, %r532, 252645135;
shr.u32 %r533, %r532, 4;
and.b32 %r451, %r533, 252645135;

dp4a.s32.s32 %r186, %r447, %r123, %r174;


dp4a.s32.s32 %r190, %r451, %r127, %r186;

and.b32 %r534, %r528, 8;
cvt.rn.f32.s32 %f94, %r534;
mul.ftz.f32 %f95, %f94, 0f3D000000;
sub.ftz.f32 %f96, %f51, %f95;

dp4a.s32.s32 %r194, %r123, %r460, %r457;


dp4a.s32.s32 %r198, %r127, %r460, %r194;

cvt.rn.f32.s32 %f97, %r198;
fma.rn.ftz.f32 %f98, %f96, %f97, %f93;
add.s64 %rd59, %rd83, %rd5;
ld.global.nc.u16 %rs32, [%rd59+48];
shr.u16 %rs33, %rs32, 12;
ld.global.nc.u16 %rs34, [%rd59+50];
shr.u16 %rs35, %rs34, 8;
and.b16 %rs36, %rs35, 240;
or.b16 %rs37, %rs36, %rs33;
ld.global.nc.u16 %rs38, [%rd59+52];
shr.u16 %rs39, %rs38, 4;
and.b16 %rs40, %rs39, 3840;
or.b16 %rs41, %rs37, %rs40;
ld.global.nc.u16 %rs42, [%rd59+54];
and.b16 %rs43, %rs42, -4096;
or.b16 %rs6, %rs41, %rs43;

{ cvt.f32.f16 %f39, %rs6;}



{.reg .f16 low,high;
mov.b32 {low,high},%r137;
cvt.f32.f16 %f40, low;}


mul.ftz.f32 %f99, %f39, %f40;
add.s64 %rd60, %rd83, %rd9;
ld.global.nc.u16 %rs44, [%rd60];
cvt.u32.u16 %r535, %rs44;
shr.u32 %r536, %r535, %r8;
shl.b32 %r537, %r536, 1;
and.b32 %r538, %r537, 14;
or.b32 %r539, %r538, 1;
shr.u32 %r540, %r536, 2;
and.b32 %r541, %r540, 14;
or.b32 %r542, %r541, 1;
cvt.rn.f32.s32 %f100, %r158;
add.ftz.f32 %f101, %f88, %f100;
cvt.rn.f32.s32 %f102, %r539;
cvt.rn.f32.s32 %f103, %r190;
add.ftz.f32 %f104, %f98, %f103;
cvt.rn.f32.s32 %f105, %r542;
mul.ftz.f32 %f106, %f104, %f105;
fma.rn.ftz.f32 %f107, %f101, %f102, %f106;
fma.rn.ftz.f32 %f268, %f99, %f107, %f268;
cvt.u64.u32 %rd61, %r658;
add.s64 %rd62, %rd2, %rd61;
add.s64 %rd63, %rd62, %rd4;
mul.lo.s64 %rd64, %rd63, 36;
add.s64 %rd65, %rd13, %rd64;
ld.global.nc.u32 %r205, [%rd65+4];
ld.global.nc.u32 %r209, [%rd65+8];

dp4a.s32.s32 %r203, %r334, %r205, %r457;


dp4a.s32.s32 %r207, %r338, %r209, %r203;


dp4a.s32.s32 %r211, %r205, %r460, %r457;


dp4a.s32.s32 %r215, %r209, %r460, %r211;

cvt.rn.f32.s32 %f108, %r215;
fma.rn.ftz.f32 %f109, %f52, %f108, 0f00000000;
ld.global.nc.u32 %r221, [%rd65+12];
ld.global.nc.u32 %r225, [%rd65+16];

dp4a.s32.s32 %r219, %r350, %r221, %r207;


dp4a.s32.s32 %r223, %r354, %r225, %r219;


dp4a.s32.s32 %r227, %r221, %r460, %r457;


dp4a.s32.s32 %r231, %r225, %r460, %r227;

cvt.rn.f32.s32 %f110, %r231;
fma.rn.ftz.f32 %f111, %f57, %f110, %f109;
ld.global.nc.u32 %r237, [%rd65+20];
ld.global.nc.u32 %r241, [%rd65+24];

dp4a.s32.s32 %r235, %r366, %r237, %r457;


dp4a.s32.s32 %r239, %r370, %r241, %r235;


dp4a.s32.s32 %r243, %r237, %r460, %r457;


dp4a.s32.s32 %r247, %r241, %r460, %r243;

cvt.rn.f32.s32 %f112, %r247;
fma.rn.ftz.f32 %f113, %f62, %f112, 0f00000000;
ld.global.nc.u32 %r253, [%rd65+28];
ld.global.nc.u32 %r257, [%rd65+32];

dp4a.s32.s32 %r251, %r382, %r253, %r239;


dp4a.s32.s32 %r255, %r386, %r257, %r251;


dp4a.s32.s32 %r259, %r253, %r460, %r457;


dp4a.s32.s32 %r263, %r257, %r460, %r259;

cvt.rn.f32.s32 %f114, %r263;
fma.rn.ftz.f32 %f115, %f67, %f114, %f113;

{ cvt.f32.f16 %f41, %rs5;}


ld.global.nc.u32 %r267, [%rd65];

{.reg .f16 low,high;
mov.b32 {low,high},%r267;
cvt.f32.f16 %f42, low;}


mul.ftz.f32 %f116, %f41, %f42;
cvt.rn.f32.s32 %f117, %r223;
add.ftz.f32 %f118, %f111, %f117;
cvt.rn.f32.s32 %f119, %r255;
add.ftz.f32 %f120, %f115, %f119;
mul.ftz.f32 %f121, %f120, %f76;
fma.rn.ftz.f32 %f122, %f118, %f73, %f121;
fma.rn.ftz.f32 %f267, %f116, %f122, %f267;

dp4a.s32.s32 %r268, %r399, %r205, %r457;


dp4a.s32.s32 %r272, %r403, %r209, %r268;


dp4a.s32.s32 %r276, %r205, %r460, %r457;


dp4a.s32.s32 %r280, %r209, %r460, %r276;

cvt.rn.f32.s32 %f123, %r280;
fma.rn.ftz.f32 %f124, %f81, %f123, 0f00000000;

dp4a.s32.s32 %r284, %r415, %r221, %r272;


dp4a.s32.s32 %r288, %r419, %r225, %r284;


dp4a.s32.s32 %r292, %r221, %r460, %r457;


dp4a.s32.s32 %r296, %r225, %r460, %r292;

cvt.rn.f32.s32 %f125, %r296;
fma.rn.ftz.f32 %f126, %f86, %f125, %f124;

dp4a.s32.s32 %r300, %r431, %r237, %r457;


dp4a.s32.s32 %r304, %r435, %r241, %r300;


dp4a.s32.s32 %r308, %r237, %r460, %r457;


dp4a.s32.s32 %r312, %r241, %r460, %r308;

cvt.rn.f32.s32 %f127, %r312;
fma.rn.ftz.f32 %f128, %f91, %f127, 0f00000000;

dp4a.s32.s32 %r316, %r447, %r253, %r304;


dp4a.s32.s32 %r320, %r451, %r257, %r316;


dp4a.s32.s32 %r324, %r253, %r460, %r457;


dp4a.s32.s32 %r328, %r257, %r460, %r324;

cvt.rn.f32.s32 %f129, %r328;
fma.rn.ftz.f32 %f130, %f96, %f129, %f128;

{ cvt.f32.f16 %f43, %rs6;}



{.reg .f16 low,high;
mov.b32 {low,high},%r267;
cvt.f32.f16 %f44, low;}


mul.ftz.f32 %f131, %f43, %f44;
cvt.rn.f32.s32 %f132, %r288;
add.ftz.f32 %f133, %f126, %f132;
cvt.rn.f32.s32 %f134, %r320;
add.ftz.f32 %f135, %f130, %f134;
mul.ftz.f32 %f136, %f135, %f105;
fma.rn.ftz.f32 %f137, %f133, %f102, %f136;
fma.rn.ftz.f32 %f266, %f131, %f137, %f266;
cvt.u64.u32 %rd66, %r659;
add.s64 %rd67, %rd2, %rd66;
add.s64 %rd68, %rd67, %rd4;
mul.lo.s64 %rd69, %rd68, 36;
add.s64 %rd70, %rd13, %rd69;
ld.global.nc.u32 %r335, [%rd70+4];
ld.global.nc.u32 %r339, [%rd70+8];

dp4a.s32.s32 %r333, %r334, %r335, %r457;


dp4a.s32.s32 %r337, %r338, %r339, %r333;


dp4a.s32.s32 %r341, %r335, %r460, %r457;


dp4a.s32.s32 %r345, %r339, %r460, %r341;

cvt.rn.f32.s32 %f138, %r345;
fma.rn.ftz.f32 %f139, %f52, %f138, 0f00000000;
ld.global.nc.u32 %r351, [%rd70+12];
ld.global.nc.u32 %r355, [%rd70+16];

dp4a.s32.s32 %r349, %r350, %r351, %r337;


dp4a.s32.s32 %r353, %r354, %r355, %r349;


dp4a.s32.s32 %r357, %r351, %r460, %r457;


dp4a.s32.s32 %r361, %r355, %r460, %r357;

cvt.rn.f32.s32 %f140, %r361;
fma.rn.ftz.f32 %f141, %f57, %f140, %f139;
ld.global.nc.u32 %r367, [%rd70+20];
ld.global.nc.u32 %r371, [%rd70+24];

dp4a.s32.s32 %r365, %r366, %r367, %r457;


dp4a.s32.s32 %r369, %r370, %r371, %r365;


dp4a.s32.s32 %r373, %r367, %r460, %r457;


dp4a.s32.s32 %r377, %r371, %r460, %r373;

cvt.rn.f32.s32 %f142, %r377;
fma.rn.ftz.f32 %f143, %f62, %f142, 0f00000000;
ld.global.nc.u32 %r383, [%rd70+28];
ld.global.nc.u32 %r387, [%rd70+32];

dp4a.s32.s32 %r381, %r382, %r383, %r369;


dp4a.s32.s32 %r385, %r386, %r387, %r381;


dp4a.s32.s32 %r389, %r383, %r460, %r457;


dp4a.s32.s32 %r393, %r387, %r460, %r389;

cvt.rn.f32.s32 %f144, %r393;
fma.rn.ftz.f32 %f145, %f67, %f144, %f143;

{ cvt.f32.f16 %f45, %rs5;}


ld.global.nc.u32 %r397, [%rd70];

{.reg .f16 low,high;
mov.b32 {low,high},%r397;
cvt.f32.f16 %f46, low;}


mul.ftz.f32 %f146, %f45, %f46;
cvt.rn.f32.s32 %f147, %r353;
add.ftz.f32 %f148, %f141, %f147;
cvt.rn.f32.s32 %f149, %r385;
add.ftz.f32 %f150, %f145, %f149;
mul.ftz.f32 %f151, %f150, %f76;
fma.rn.ftz.f32 %f152, %f148, %f73, %f151;
fma.rn.ftz.f32 %f265, %f146, %f152, %f265;

dp4a.s32.s32 %r398, %r399, %r335, %r457;


dp4a.s32.s32 %r402, %r403, %r339, %r398;


dp4a.s32.s32 %r406, %r335, %r460, %r457;


dp4a.s32.s32 %r410, %r339, %r460, %r406;

cvt.rn.f32.s32 %f153, %r410;
fma.rn.ftz.f32 %f154, %f81, %f153, 0f00000000;

dp4a.s32.s32 %r414, %r415, %r351, %r402;


dp4a.s32.s32 %r418, %r419, %r355, %r414;


dp4a.s32.s32 %r422, %r351, %r460, %r457;


dp4a.s32.s32 %r426, %r355, %r460, %r422;

cvt.rn.f32.s32 %f155, %r426;
fma.rn.ftz.f32 %f156, %f86, %f155, %f154;

dp4a.s32.s32 %r430, %r431, %r367, %r457;


dp4a.s32.s32 %r434, %r435, %r371, %r430;


dp4a.s32.s32 %r438, %r367, %r460, %r457;


dp4a.s32.s32 %r442, %r371, %r460, %r438;

cvt.rn.f32.s32 %f157, %r442;
fma.rn.ftz.f32 %f158, %f91, %f157, 0f00000000;

dp4a.s32.s32 %r446, %r447, %r383, %r434;


dp4a.s32.s32 %r450, %r451, %r387, %r446;


dp4a.s32.s32 %r454, %r383, %r460, %r457;


dp4a.s32.s32 %r458, %r387, %r460, %r454;

cvt.rn.f32.s32 %f159, %r458;
fma.rn.ftz.f32 %f160, %f96, %f159, %f158;

{ cvt.f32.f16 %f47, %rs6;}



{.reg .f16 low,high;
mov.b32 {low,high},%r397;
cvt.f32.f16 %f48, low;}


mul.ftz.f32 %f161, %f47, %f48;
cvt.rn.f32.s32 %f162, %r418;
add.ftz.f32 %f163, %f156, %f162;
cvt.rn.f32.s32 %f164, %r450;
add.ftz.f32 %f165, %f160, %f164;
mul.ftz.f32 %f166, %f165, %f105;
fma.rn.ftz.f32 %f167, %f163, %f102, %f166;
fma.rn.ftz.f32 %f264, %f161, %f167, %f264;
add.s32 %r659, %r659, 128;
add.s32 %r658, %r658, 128;
add.s32 %r657, %r657, 128;
add.s64 %rd83, %rd83, 896;
add.s32 %r660, %r660, 16;
setp.lt.s32 %p4, %r660, %r7;
@%p4 bra $L__BB270_2;

$L__BB270_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB270_5;

mov.u32 %r543, _ZZ13mul_mat_vec_qIL9ggml_type29ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r544, %r1, 768, %r543;
shl.b32 %r545, %r2, 2;
add.s32 %r546, %r544, %r545;
st.shared.f32 [%r546+-768], %f269;
st.shared.f32 [%r546+-640], %f268;
st.shared.f32 [%r546+-512], %f267;
st.shared.f32 [%r546+-384], %f266;
st.shared.f32 [%r546+-256], %f265;
st.shared.f32 [%r546+-128], %f264;

$L__BB270_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB270_18;

mov.u32 %r547, %ctaid.z;
mad.lo.s32 %r548, %r5, %r30, %r4;
mad.lo.s32 %r549, %r547, %r36, %r548;
cvt.u64.u32 %rd16, %r549;
shl.b32 %r550, %r2, 2;
mov.u32 %r551, _ZZ13mul_mat_vec_qIL9ggml_type29ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r20, %r551, %r550;
mov.u32 %r552, 2;
ld.shared.f32 %f168, [%r20];
add.ftz.f32 %f169, %f168, %f269;
ld.shared.f32 %f170, [%r20+768];
add.ftz.f32 %f171, %f170, %f169;
ld.shared.f32 %f172, [%r20+1536];
add.ftz.f32 %f173, %f172, %f171;
mov.b32 %r553, %f173;
mov.u32 %r554, 31;
mov.u32 %r555, 16;
mov.u32 %r556, -1;
shfl.sync.bfly.b32 %r557|%p7, %r553, %r555, %r554, %r556;
mov.b32 %f174, %r557;
add.ftz.f32 %f175, %f173, %f174;
mov.b32 %r558, %f175;
mov.u32 %r559, 8;
shfl.sync.bfly.b32 %r560|%p8, %r558, %r559, %r554, %r556;
mov.b32 %f176, %r560;
add.ftz.f32 %f177, %f175, %f176;
mov.b32 %r561, %f177;
mov.u32 %r562, 4;
shfl.sync.bfly.b32 %r563|%p9, %r561, %r562, %r554, %r556;
mov.b32 %f178, %r563;
add.ftz.f32 %f179, %f177, %f178;
mov.b32 %r564, %f179;
shfl.sync.bfly.b32 %r565|%p10, %r564, %r552, %r554, %r556;
mov.b32 %f180, %r565;
add.ftz.f32 %f181, %f179, %f180;
mov.b32 %r566, %f181;
mov.u32 %r567, 1;
shfl.sync.bfly.b32 %r568|%p11, %r566, %r567, %r554, %r556;
mov.b32 %f182, %r568;
add.ftz.f32 %f19, %f181, %f182;
setp.lt.u32 %p12, %r4, %r24;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
mul.wide.u32 %rd71, %r549, 4;
add.s64 %rd17, %rd1, %rd71;
not.pred %p14, %p1;
@%p14 bra $L__BB270_8;

st.global.f32 [%rd17], %f19;

$L__BB270_8:
ld.shared.f32 %f183, [%r20+128];
add.ftz.f32 %f184, %f183, %f268;
ld.shared.f32 %f185, [%r20+896];
add.ftz.f32 %f186, %f185, %f184;
ld.shared.f32 %f187, [%r20+1664];
add.ftz.f32 %f188, %f187, %f186;
mov.b32 %r569, %f188;
shfl.sync.bfly.b32 %r573|%p15, %r569, %r555, %r554, %r556;
mov.b32 %f189, %r573;
add.ftz.f32 %f190, %f188, %f189;
mov.b32 %r574, %f190;
shfl.sync.bfly.b32 %r576|%p16, %r574, %r559, %r554, %r556;
mov.b32 %f191, %r576;
add.ftz.f32 %f192, %f190, %f191;
mov.b32 %r577, %f192;
shfl.sync.bfly.b32 %r579|%p17, %r577, %r562, %r554, %r556;
mov.b32 %f193, %r579;
add.ftz.f32 %f194, %f192, %f193;
mov.b32 %r580, %f194;
shfl.sync.bfly.b32 %r582|%p18, %r580, %r552, %r554, %r556;
mov.b32 %f195, %r582;
add.ftz.f32 %f196, %f194, %f195;
mov.b32 %r583, %f196;
shfl.sync.bfly.b32 %r585|%p19, %r583, %r567, %r554, %r556;
mov.b32 %f197, %r585;
add.ftz.f32 %f20, %f196, %f197;
add.s32 %r586, %r4, 1;
setp.lt.u32 %p20, %r586, %r24;
setp.eq.s32 %p21, %r2, 1;
and.pred %p2, %p21, %p20;
not.pred %p22, %p2;
@%p22 bra $L__BB270_10;

st.global.f32 [%rd17+4], %f20;

$L__BB270_10:
ld.shared.f32 %f198, [%r20+256];
add.ftz.f32 %f199, %f198, %f267;
ld.shared.f32 %f200, [%r20+1024];
add.ftz.f32 %f201, %f200, %f199;
ld.shared.f32 %f202, [%r20+1792];
add.ftz.f32 %f203, %f202, %f201;
mov.b32 %r587, %f203;
mov.u32 %r588, 31;
mov.u32 %r589, 16;
mov.u32 %r590, -1;
shfl.sync.bfly.b32 %r591|%p23, %r587, %r589, %r588, %r590;
mov.b32 %f204, %r591;
add.ftz.f32 %f205, %f203, %f204;
mov.b32 %r592, %f205;
mov.u32 %r593, 8;
shfl.sync.bfly.b32 %r594|%p24, %r592, %r593, %r588, %r590;
mov.b32 %f206, %r594;
add.ftz.f32 %f207, %f205, %f206;
mov.b32 %r595, %f207;
mov.u32 %r596, 4;
shfl.sync.bfly.b32 %r597|%p25, %r595, %r596, %r588, %r590;
mov.b32 %f208, %r597;
add.ftz.f32 %f209, %f207, %f208;
mov.b32 %r598, %f209;
mov.u32 %r599, 2;
shfl.sync.bfly.b32 %r600|%p26, %r598, %r599, %r588, %r590;
mov.b32 %f210, %r600;
add.ftz.f32 %f211, %f209, %f210;
mov.b32 %r601, %f211;
mov.u32 %r602, 1;
shfl.sync.bfly.b32 %r603|%p27, %r601, %r602, %r588, %r590;
mov.b32 %f212, %r603;
add.ftz.f32 %f21, %f211, %f212;
@%p14 bra $L__BB270_12;

cvt.u64.u32 %rd72, %r24;
add.s64 %rd73, %rd72, %rd16;
shl.b64 %rd74, %rd73, 2;
add.s64 %rd75, %rd1, %rd74;
st.global.f32 [%rd75], %f21;

$L__BB270_12:
ld.shared.f32 %f213, [%r20+384];
add.ftz.f32 %f214, %f213, %f266;
ld.shared.f32 %f215, [%r20+1152];
add.ftz.f32 %f216, %f215, %f214;
ld.shared.f32 %f217, [%r20+1920];
add.ftz.f32 %f218, %f217, %f216;
mov.b32 %r604, %f218;
shfl.sync.bfly.b32 %r608|%p29, %r604, %r589, %r588, %r590;
mov.b32 %f219, %r608;
add.ftz.f32 %f220, %f218, %f219;
mov.b32 %r609, %f220;
shfl.sync.bfly.b32 %r611|%p30, %r609, %r593, %r588, %r590;
mov.b32 %f221, %r611;
add.ftz.f32 %f222, %f220, %f221;
mov.b32 %r612, %f222;
shfl.sync.bfly.b32 %r614|%p31, %r612, %r596, %r588, %r590;
mov.b32 %f223, %r614;
add.ftz.f32 %f224, %f222, %f223;
mov.b32 %r615, %f224;
shfl.sync.bfly.b32 %r617|%p32, %r615, %r599, %r588, %r590;
mov.b32 %f225, %r617;
add.ftz.f32 %f226, %f224, %f225;
mov.b32 %r618, %f226;
shfl.sync.bfly.b32 %r620|%p33, %r618, %r602, %r588, %r590;
mov.b32 %f227, %r620;
add.ftz.f32 %f22, %f226, %f227;
@%p22 bra $L__BB270_14;

add.s32 %r621, %r24, 1;
cvt.u64.u32 %rd76, %r621;
add.s64 %rd77, %rd76, %rd16;
shl.b64 %rd78, %rd77, 2;
add.s64 %rd79, %rd1, %rd78;
st.global.f32 [%rd79], %f22;

$L__BB270_14:
shl.b32 %r622, %r24, 1;
mov.u32 %r623, 1;
ld.shared.f32 %f228, [%r20+512];
add.ftz.f32 %f229, %f228, %f265;
ld.shared.f32 %f230, [%r20+1280];
add.ftz.f32 %f231, %f230, %f229;
ld.shared.f32 %f232, [%r20+2048];
add.ftz.f32 %f233, %f232, %f231;
mov.b32 %r624, %f233;
mov.u32 %r625, 31;
mov.u32 %r626, 16;
mov.u32 %r627, -1;
shfl.sync.bfly.b32 %r628|%p35, %r624, %r626, %r625, %r627;
mov.b32 %f234, %r628;
add.ftz.f32 %f235, %f233, %f234;
mov.b32 %r629, %f235;
mov.u32 %r630, 8;
shfl.sync.bfly.b32 %r631|%p36, %r629, %r630, %r625, %r627;
mov.b32 %f236, %r631;
add.ftz.f32 %f237, %f235, %f236;
mov.b32 %r632, %f237;
mov.u32 %r633, 4;
shfl.sync.bfly.b32 %r634|%p37, %r632, %r633, %r625, %r627;
mov.b32 %f238, %r634;
add.ftz.f32 %f239, %f237, %f238;
mov.b32 %r635, %f239;
mov.u32 %r636, 2;
shfl.sync.bfly.b32 %r637|%p38, %r635, %r636, %r625, %r627;
mov.b32 %f240, %r637;
add.ftz.f32 %f241, %f239, %f240;
mov.b32 %r638, %f241;
shfl.sync.bfly.b32 %r639|%p39, %r638, %r623, %r625, %r627;
mov.b32 %f242, %r639;
add.ftz.f32 %f23, %f241, %f242;
cvt.u64.u32 %rd80, %r622;
add.s64 %rd81, %rd80, %rd16;
shl.b64 %rd82, %rd81, 2;
add.s64 %rd18, %rd1, %rd82;
@%p14 bra $L__BB270_16;

st.global.f32 [%rd18], %f23;

$L__BB270_16:
ld.shared.f32 %f243, [%r20+640];
add.ftz.f32 %f244, %f243, %f264;
ld.shared.f32 %f245, [%r20+1408];
add.ftz.f32 %f246, %f245, %f244;
ld.shared.f32 %f247, [%r20+2176];
add.ftz.f32 %f248, %f247, %f246;
mov.b32 %r640, %f248;
shfl.sync.bfly.b32 %r644|%p41, %r640, %r626, %r625, %r627;
mov.b32 %f249, %r644;
add.ftz.f32 %f250, %f248, %f249;
mov.b32 %r645, %f250;
shfl.sync.bfly.b32 %r647|%p42, %r645, %r630, %r625, %r627;
mov.b32 %f251, %r647;
add.ftz.f32 %f252, %f250, %f251;
mov.b32 %r648, %f252;
shfl.sync.bfly.b32 %r650|%p43, %r648, %r633, %r625, %r627;
mov.b32 %f253, %r650;
add.ftz.f32 %f254, %f252, %f253;
mov.b32 %r651, %f254;
shfl.sync.bfly.b32 %r653|%p44, %r651, %r636, %r625, %r627;
mov.b32 %f255, %r653;
add.ftz.f32 %f256, %f254, %f255;
mov.b32 %r654, %f256;
shfl.sync.bfly.b32 %r656|%p45, %r654, %r623, %r625, %r627;
mov.b32 %f257, %r656;
add.ftz.f32 %f24, %f256, %f257;
@%p22 bra $L__BB270_18;

st.global.f32 [%rd18+4], %f24;

$L__BB270_18:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type29ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type29ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type29ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type29ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type29ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type29ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type29ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type29ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type29ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<59>;
.reg .b16 %rs<47>;
.reg .f32 %f<350>;
.reg .b32 %r<832>;
.reg .b64 %rd<97>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type29ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[3072];

ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type29ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd20, [_Z13mul_mat_vec_qIL9ggml_type29ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd21, [_Z13mul_mat_vec_qIL9ggml_type29ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type29ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type29ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type29ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type29ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type29ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type29ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type29ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type29ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type29ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type29ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type29ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type29ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type29ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type29ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
cvta.to.global.u64 %rd1, %rd21;
mov.u32 %r1, %tid.y;
shl.b32 %r42, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r42, %r2;
mov.u32 %r43, %ctaid.x;
shl.b32 %r4, %r43, 1;
mov.u32 %r5, %ctaid.y;
shr.s32 %r44, %r3, 31;
shr.u32 %r45, %r44, 29;
add.s32 %r46, %r3, %r45;
shr.s32 %r831, %r46, 3;
shr.u32 %r7, %r41, 8;
setp.ge.s32 %p3, %r831, %r7;
mov.f32 %f342, 0f00000000;
mov.f32 %f343, %f342;
mov.f32 %f344, %f342;
mov.f32 %f345, %f342;
mov.f32 %f346, %f342;
mov.f32 %f347, %f342;
mov.f32 %f348, %f342;
mov.f32 %f349, %f342;
@%p3 bra $L__BB271_3;

mov.u32 %r47, %ctaid.z;
mul.lo.s32 %r48, %r5, %r33;
cvt.u64.u32 %rd22, %r48;
mul.lo.s32 %r49, %r47, %r39;
cvt.u64.u32 %rd23, %r49;
add.s64 %rd2, %rd23, %rd22;
mul.hi.u32 %r50, %r47, %r35;
add.s32 %r51, %r50, %r47;
shr.u32 %r52, %r51, %r36;
mul.lo.s32 %r53, %r52, %r38;
mul.hi.u32 %r54, %r5, %r29;
add.s32 %r55, %r54, %r5;
shr.u32 %r56, %r55, %r30;
mul.lo.s32 %r57, %r56, %r32;
cvta.to.global.u64 %rd96, %rd19;
add.s32 %r58, %r4, 1;
and.b32 %r62, %r46, -8;
sub.s32 %r63, %r3, %r62;
shl.b32 %r64, %r63, 1;
cvt.s64.s32 %rd24, %r64;
cvt.s64.s32 %rd4, %r63;
shr.u32 %r65, %r63, 31;
add.s32 %r66, %r63, %r65;
shr.s32 %r67, %r66, 1;
and.b32 %r68, %r66, -2;
sub.s32 %r69, %r63, %r68;
mul.lo.s32 %r8, %r69, 6;
shl.b32 %r70, %r27, 1;
shl.b32 %r827, %r831, 3;
add.s32 %r830, %r827, %r70;
mad.lo.s32 %r829, %r27, 3, %r827;
add.s32 %r828, %r27, %r827;
add.s32 %r71, %r831, %r53;
add.s32 %r72, %r71, %r57;
mad.lo.s32 %r73, %r58, %r26, %r72;
mul.wide.s32 %rd5, %r73, 56;
mad.lo.s32 %r74, %r4, %r26, %r831;
add.s32 %r75, %r74, %r53;
add.s32 %r76, %r75, %r57;
mul.wide.s32 %rd6, %r76, 56;
mul.wide.s32 %rd25, %r63, 4;
add.s64 %rd7, %rd6, %rd25;
mul.wide.s32 %rd26, %r67, 2;
add.s64 %rd27, %rd6, %rd26;
add.s64 %rd8, %rd27, 48;
add.s64 %rd28, %rd5, %rd26;
add.s64 %rd9, %rd28, 48;
add.s64 %rd29, %rd6, %rd24;
add.s64 %rd10, %rd29, 33;
add.s64 %rd30, %rd5, %rd24;
add.s64 %rd11, %rd30, 32;
add.s64 %rd12, %rd5, %rd25;
cvta.to.global.u64 %rd13, %rd20;
mov.f32 %f342, 0f00000000;

$L__BB271_2:
cvt.u64.u32 %rd31, %r827;
add.s64 %rd32, %rd2, %rd31;
add.s64 %rd33, %rd32, %rd4;
mul.lo.s64 %rd34, %rd33, 36;
add.s64 %rd35, %rd96, %rd7;
ld.global.nc.u32 %r597, [%rd35];
shr.u32 %r598, %r597, 8;
shr.u32 %r599, %r597, 16;
shr.u32 %r600, %r597, 24;
add.s64 %rd36, %rd96, %rd10;
ld.global.nc.u8 %rs9, [%rd36+-1];
and.b16 %rs10, %rs9, 240;
cvt.u32.u16 %r601, %rs9;
and.b32 %r602, %r597, 255;
and.b32 %r603, %r601, 7;
bfi.b32 %r604, %r603, %r602, 8, 3;
mul.wide.u32 %rd37, %r604, 4;
mov.u64 %rd38, iq1s_grid_gpu;
add.s64 %rd39, %rd38, %rd37;
ld.global.nc.u32 %r605, [%rd39];
and.b32 %r468, %r605, 252645135;
shr.u32 %r606, %r605, 4;
and.b32 %r472, %r606, 252645135;
add.s64 %rd40, %rd13, %rd34;
ld.global.nc.u32 %r79, [%rd40+4];
ld.global.nc.u32 %r83, [%rd40+8];
mov.u32 %r591, 0;

dp4a.s32.s32 %r77, %r468, %r79, %r591;


dp4a.s32.s32 %r81, %r472, %r83, %r77;

and.b32 %r607, %r601, 8;
cvt.rn.f32.s32 %f65, %r607;
mul.ftz.f32 %f66, %f65, 0f3D000000;
mov.f32 %f67, 0fBF600000;
sub.ftz.f32 %f68, %f67, %f66;
mov.u32 %r594, 16843009;

dp4a.s32.s32 %r85, %r79, %r594, %r591;


dp4a.s32.s32 %r89, %r83, %r594, %r85;

cvt.rn.f32.s32 %f69, %r89;
fma.rn.ftz.f32 %f70, %f68, %f69, 0f00000000;
shr.u16 %rs11, %rs10, 4;
cvt.u32.u16 %r608, %rs11;
and.b32 %r609, %r598, 255;
and.b32 %r610, %r608, 7;
bfi.b32 %r611, %r610, %r609, 8, 3;
mul.wide.u32 %rd41, %r611, 4;
add.s64 %rd42, %rd38, %rd41;
ld.global.nc.u32 %r612, [%rd42];
and.b32 %r484, %r612, 252645135;
shr.u32 %r613, %r612, 4;
and.b32 %r488, %r613, 252645135;
ld.global.nc.u32 %r95, [%rd40+12];
ld.global.nc.u32 %r99, [%rd40+16];

dp4a.s32.s32 %r93, %r484, %r95, %r81;


dp4a.s32.s32 %r97, %r488, %r99, %r93;

and.b32 %r614, %r608, 8;
cvt.rn.f32.s32 %f71, %r614;
mul.ftz.f32 %f72, %f71, 0f3D000000;
sub.ftz.f32 %f73, %f67, %f72;

dp4a.s32.s32 %r101, %r95, %r594, %r591;


dp4a.s32.s32 %r105, %r99, %r594, %r101;

cvt.rn.f32.s32 %f74, %r105;
fma.rn.ftz.f32 %f75, %f73, %f74, %f70;
ld.global.nc.u8 %rs12, [%rd36];
and.b16 %rs13, %rs12, 240;
cvt.u32.u16 %r615, %rs12;
and.b32 %r616, %r599, 255;
and.b32 %r617, %r615, 7;
bfi.b32 %r618, %r617, %r616, 8, 3;
mul.wide.u32 %rd43, %r618, 4;
add.s64 %rd44, %rd38, %rd43;
ld.global.nc.u32 %r619, [%rd44];
and.b32 %r500, %r619, 252645135;
shr.u32 %r620, %r619, 4;
and.b32 %r504, %r620, 252645135;
ld.global.nc.u32 %r111, [%rd40+20];
ld.global.nc.u32 %r115, [%rd40+24];

dp4a.s32.s32 %r109, %r500, %r111, %r591;


dp4a.s32.s32 %r113, %r504, %r115, %r109;

and.b32 %r621, %r615, 8;
cvt.rn.f32.s32 %f76, %r621;
mul.ftz.f32 %f77, %f76, 0f3D000000;
sub.ftz.f32 %f78, %f67, %f77;

dp4a.s32.s32 %r117, %r111, %r594, %r591;


dp4a.s32.s32 %r121, %r115, %r594, %r117;

cvt.rn.f32.s32 %f79, %r121;
fma.rn.ftz.f32 %f80, %f78, %f79, 0f00000000;
shr.u16 %rs14, %rs13, 4;
cvt.u32.u16 %r622, %rs14;
mul.wide.u16 %r623, %rs14, 256;
and.b32 %r624, %r623, 1792;
or.b32 %r625, %r624, %r600;
mul.wide.u32 %rd45, %r625, 4;
add.s64 %rd46, %rd38, %rd45;
ld.global.nc.u32 %r626, [%rd46];
and.b32 %r516, %r626, 252645135;
shr.u32 %r627, %r626, 4;
and.b32 %r520, %r627, 252645135;
ld.global.nc.u32 %r127, [%rd40+28];
ld.global.nc.u32 %r131, [%rd40+32];

dp4a.s32.s32 %r125, %r516, %r127, %r113;


dp4a.s32.s32 %r129, %r520, %r131, %r125;

and.b32 %r628, %r622, 8;
cvt.rn.f32.s32 %f81, %r628;
mul.ftz.f32 %f82, %f81, 0f3D000000;
sub.ftz.f32 %f83, %f67, %f82;

dp4a.s32.s32 %r133, %r127, %r594, %r591;


dp4a.s32.s32 %r137, %r131, %r594, %r133;

cvt.rn.f32.s32 %f84, %r137;
fma.rn.ftz.f32 %f85, %f83, %f84, %f80;
add.s64 %rd47, %rd96, %rd6;
ld.global.nc.u16 %rs15, [%rd47+48];
shr.u16 %rs16, %rs15, 12;
ld.global.nc.u16 %rs17, [%rd47+50];
shr.u16 %rs18, %rs17, 8;
and.b16 %rs19, %rs18, 240;
or.b16 %rs20, %rs19, %rs16;
ld.global.nc.u16 %rs21, [%rd47+52];
shr.u16 %rs22, %rs21, 4;
and.b16 %rs23, %rs22, 3840;
or.b16 %rs24, %rs20, %rs23;
ld.global.nc.u16 %rs25, [%rd47+54];
and.b16 %rs26, %rs25, -4096;
or.b16 %rs7, %rs24, %rs26;

{ cvt.f32.f16 %f49, %rs7;}


ld.global.nc.u32 %r141, [%rd40];

{.reg .f16 low,high;
mov.b32 {low,high},%r141;
cvt.f32.f16 %f50, low;}


mul.ftz.f32 %f86, %f49, %f50;
add.s64 %rd48, %rd96, %rd8;
ld.global.nc.u16 %rs27, [%rd48];
cvt.u32.u16 %r629, %rs27;
shr.u32 %r630, %r629, %r8;
shl.b32 %r631, %r630, 1;
and.b32 %r632, %r631, 14;
or.b32 %r633, %r632, 1;
shr.u32 %r634, %r630, 2;
and.b32 %r635, %r634, 14;
or.b32 %r636, %r635, 1;
cvt.rn.f32.s32 %f87, %r97;
add.ftz.f32 %f88, %f75, %f87;
cvt.rn.f32.s32 %f89, %r633;
cvt.rn.f32.s32 %f90, %r129;
add.ftz.f32 %f91, %f85, %f90;
cvt.rn.f32.s32 %f92, %r636;
mul.ftz.f32 %f93, %f91, %f92;
fma.rn.ftz.f32 %f94, %f88, %f89, %f93;
fma.rn.ftz.f32 %f349, %f86, %f94, %f349;
add.s64 %rd49, %rd96, %rd12;
ld.global.nc.u32 %r637, [%rd49];
shr.u32 %r638, %r637, 8;
shr.u32 %r639, %r637, 16;
shr.u32 %r640, %r637, 24;
add.s64 %rd50, %rd96, %rd11;
ld.global.nc.u8 %rs28, [%rd50];
and.b16 %rs29, %rs28, 240;
cvt.u32.u16 %r641, %rs28;
and.b32 %r642, %r637, 255;
and.b32 %r643, %r641, 7;
bfi.b32 %r644, %r643, %r642, 8, 3;
mul.wide.u32 %rd51, %r644, 4;
add.s64 %rd52, %rd38, %rd51;
ld.global.nc.u32 %r645, [%rd52];
and.b32 %r533, %r645, 252645135;
shr.u32 %r646, %r645, 4;
and.b32 %r537, %r646, 252645135;

dp4a.s32.s32 %r142, %r533, %r79, %r591;


dp4a.s32.s32 %r146, %r537, %r83, %r142;

and.b32 %r647, %r641, 8;
cvt.rn.f32.s32 %f95, %r647;
mul.ftz.f32 %f96, %f95, 0f3D000000;
sub.ftz.f32 %f97, %f67, %f96;

dp4a.s32.s32 %r150, %r79, %r594, %r591;


dp4a.s32.s32 %r154, %r83, %r594, %r150;

cvt.rn.f32.s32 %f98, %r154;
fma.rn.ftz.f32 %f99, %f97, %f98, 0f00000000;
shr.u16 %rs30, %rs29, 4;
cvt.u32.u16 %r648, %rs30;
and.b32 %r649, %r638, 255;
and.b32 %r650, %r648, 7;
bfi.b32 %r651, %r650, %r649, 8, 3;
mul.wide.u32 %rd53, %r651, 4;
add.s64 %rd54, %rd38, %rd53;
ld.global.nc.u32 %r652, [%rd54];
and.b32 %r549, %r652, 252645135;
shr.u32 %r653, %r652, 4;
and.b32 %r553, %r653, 252645135;

dp4a.s32.s32 %r158, %r549, %r95, %r146;


dp4a.s32.s32 %r162, %r553, %r99, %r158;

and.b32 %r654, %r648, 8;
cvt.rn.f32.s32 %f100, %r654;
mul.ftz.f32 %f101, %f100, 0f3D000000;
sub.ftz.f32 %f102, %f67, %f101;

dp4a.s32.s32 %r166, %r95, %r594, %r591;


dp4a.s32.s32 %r170, %r99, %r594, %r166;

cvt.rn.f32.s32 %f103, %r170;
fma.rn.ftz.f32 %f104, %f102, %f103, %f99;
ld.global.nc.u8 %rs31, [%rd50+1];
and.b16 %rs32, %rs31, 240;
cvt.u32.u16 %r655, %rs31;
and.b32 %r656, %r639, 255;
and.b32 %r657, %r655, 7;
bfi.b32 %r658, %r657, %r656, 8, 3;
mul.wide.u32 %rd55, %r658, 4;
add.s64 %rd56, %rd38, %rd55;
ld.global.nc.u32 %r659, [%rd56];
and.b32 %r565, %r659, 252645135;
shr.u32 %r660, %r659, 4;
and.b32 %r569, %r660, 252645135;

dp4a.s32.s32 %r174, %r565, %r111, %r591;


dp4a.s32.s32 %r178, %r569, %r115, %r174;

and.b32 %r661, %r655, 8;
cvt.rn.f32.s32 %f105, %r661;
mul.ftz.f32 %f106, %f105, 0f3D000000;
sub.ftz.f32 %f107, %f67, %f106;

dp4a.s32.s32 %r182, %r111, %r594, %r591;


dp4a.s32.s32 %r186, %r115, %r594, %r182;

cvt.rn.f32.s32 %f108, %r186;
fma.rn.ftz.f32 %f109, %f107, %f108, 0f00000000;
shr.u16 %rs33, %rs32, 4;
cvt.u32.u16 %r662, %rs33;
mul.wide.u16 %r663, %rs33, 256;
and.b32 %r664, %r663, 1792;
or.b32 %r665, %r664, %r640;
mul.wide.u32 %rd57, %r665, 4;
add.s64 %rd58, %rd38, %rd57;
ld.global.nc.u32 %r666, [%rd58];
and.b32 %r581, %r666, 252645135;
shr.u32 %r667, %r666, 4;
and.b32 %r585, %r667, 252645135;

dp4a.s32.s32 %r190, %r581, %r127, %r178;


dp4a.s32.s32 %r194, %r585, %r131, %r190;

and.b32 %r668, %r662, 8;
cvt.rn.f32.s32 %f110, %r668;
mul.ftz.f32 %f111, %f110, 0f3D000000;
sub.ftz.f32 %f112, %f67, %f111;

dp4a.s32.s32 %r198, %r127, %r594, %r591;


dp4a.s32.s32 %r202, %r131, %r594, %r198;

cvt.rn.f32.s32 %f113, %r202;
fma.rn.ftz.f32 %f114, %f112, %f113, %f109;
add.s64 %rd59, %rd96, %rd5;
ld.global.nc.u16 %rs34, [%rd59+48];
shr.u16 %rs35, %rs34, 12;
ld.global.nc.u16 %rs36, [%rd59+50];
shr.u16 %rs37, %rs36, 8;
and.b16 %rs38, %rs37, 240;
or.b16 %rs39, %rs38, %rs35;
ld.global.nc.u16 %rs40, [%rd59+52];
shr.u16 %rs41, %rs40, 4;
and.b16 %rs42, %rs41, 3840;
or.b16 %rs43, %rs39, %rs42;
ld.global.nc.u16 %rs44, [%rd59+54];
and.b16 %rs45, %rs44, -4096;
or.b16 %rs8, %rs43, %rs45;

{ cvt.f32.f16 %f51, %rs8;}



{.reg .f16 low,high;
mov.b32 {low,high},%r141;
cvt.f32.f16 %f52, low;}


mul.ftz.f32 %f115, %f51, %f52;
add.s64 %rd60, %rd96, %rd9;
ld.global.nc.u16 %rs46, [%rd60];
cvt.u32.u16 %r669, %rs46;
shr.u32 %r670, %r669, %r8;
shl.b32 %r671, %r670, 1;
and.b32 %r672, %r671, 14;
or.b32 %r673, %r672, 1;
shr.u32 %r674, %r670, 2;
and.b32 %r675, %r674, 14;
or.b32 %r676, %r675, 1;
cvt.rn.f32.s32 %f116, %r162;
add.ftz.f32 %f117, %f104, %f116;
cvt.rn.f32.s32 %f118, %r673;
cvt.rn.f32.s32 %f119, %r194;
add.ftz.f32 %f120, %f114, %f119;
cvt.rn.f32.s32 %f121, %r676;
mul.ftz.f32 %f122, %f120, %f121;
fma.rn.ftz.f32 %f123, %f117, %f118, %f122;
fma.rn.ftz.f32 %f348, %f115, %f123, %f348;
cvt.u64.u32 %rd61, %r828;
add.s64 %rd62, %rd2, %rd61;
add.s64 %rd63, %rd62, %rd4;
mul.lo.s64 %rd64, %rd63, 36;
add.s64 %rd65, %rd13, %rd64;
ld.global.nc.u32 %r209, [%rd65+4];
ld.global.nc.u32 %r213, [%rd65+8];

dp4a.s32.s32 %r207, %r468, %r209, %r591;


dp4a.s32.s32 %r211, %r472, %r213, %r207;


dp4a.s32.s32 %r215, %r209, %r594, %r591;


dp4a.s32.s32 %r219, %r213, %r594, %r215;

cvt.rn.f32.s32 %f124, %r219;
fma.rn.ftz.f32 %f125, %f68, %f124, 0f00000000;
ld.global.nc.u32 %r225, [%rd65+12];
ld.global.nc.u32 %r229, [%rd65+16];

dp4a.s32.s32 %r223, %r484, %r225, %r211;


dp4a.s32.s32 %r227, %r488, %r229, %r223;


dp4a.s32.s32 %r231, %r225, %r594, %r591;


dp4a.s32.s32 %r235, %r229, %r594, %r231;

cvt.rn.f32.s32 %f126, %r235;
fma.rn.ftz.f32 %f127, %f73, %f126, %f125;
ld.global.nc.u32 %r241, [%rd65+20];
ld.global.nc.u32 %r245, [%rd65+24];

dp4a.s32.s32 %r239, %r500, %r241, %r591;


dp4a.s32.s32 %r243, %r504, %r245, %r239;


dp4a.s32.s32 %r247, %r241, %r594, %r591;


dp4a.s32.s32 %r251, %r245, %r594, %r247;

cvt.rn.f32.s32 %f128, %r251;
fma.rn.ftz.f32 %f129, %f78, %f128, 0f00000000;
ld.global.nc.u32 %r257, [%rd65+28];
ld.global.nc.u32 %r261, [%rd65+32];

dp4a.s32.s32 %r255, %r516, %r257, %r243;


dp4a.s32.s32 %r259, %r520, %r261, %r255;


dp4a.s32.s32 %r263, %r257, %r594, %r591;


dp4a.s32.s32 %r267, %r261, %r594, %r263;

cvt.rn.f32.s32 %f130, %r267;
fma.rn.ftz.f32 %f131, %f83, %f130, %f129;

{ cvt.f32.f16 %f53, %rs7;}


ld.global.nc.u32 %r271, [%rd65];

{.reg .f16 low,high;
mov.b32 {low,high},%r271;
cvt.f32.f16 %f54, low;}


mul.ftz.f32 %f132, %f53, %f54;
cvt.rn.f32.s32 %f133, %r227;
add.ftz.f32 %f134, %f127, %f133;
cvt.rn.f32.s32 %f135, %r259;
add.ftz.f32 %f136, %f131, %f135;
mul.ftz.f32 %f137, %f136, %f92;
fma.rn.ftz.f32 %f138, %f134, %f89, %f137;
fma.rn.ftz.f32 %f347, %f132, %f138, %f347;

dp4a.s32.s32 %r272, %r533, %r209, %r591;


dp4a.s32.s32 %r276, %r537, %r213, %r272;


dp4a.s32.s32 %r280, %r209, %r594, %r591;


dp4a.s32.s32 %r284, %r213, %r594, %r280;

cvt.rn.f32.s32 %f139, %r284;
fma.rn.ftz.f32 %f140, %f97, %f139, 0f00000000;

dp4a.s32.s32 %r288, %r549, %r225, %r276;


dp4a.s32.s32 %r292, %r553, %r229, %r288;


dp4a.s32.s32 %r296, %r225, %r594, %r591;


dp4a.s32.s32 %r300, %r229, %r594, %r296;

cvt.rn.f32.s32 %f141, %r300;
fma.rn.ftz.f32 %f142, %f102, %f141, %f140;

dp4a.s32.s32 %r304, %r565, %r241, %r591;


dp4a.s32.s32 %r308, %r569, %r245, %r304;


dp4a.s32.s32 %r312, %r241, %r594, %r591;


dp4a.s32.s32 %r316, %r245, %r594, %r312;

cvt.rn.f32.s32 %f143, %r316;
fma.rn.ftz.f32 %f144, %f107, %f143, 0f00000000;

dp4a.s32.s32 %r320, %r581, %r257, %r308;


dp4a.s32.s32 %r324, %r585, %r261, %r320;


dp4a.s32.s32 %r328, %r257, %r594, %r591;


dp4a.s32.s32 %r332, %r261, %r594, %r328;

cvt.rn.f32.s32 %f145, %r332;
fma.rn.ftz.f32 %f146, %f112, %f145, %f144;

{ cvt.f32.f16 %f55, %rs8;}



{.reg .f16 low,high;
mov.b32 {low,high},%r271;
cvt.f32.f16 %f56, low;}


mul.ftz.f32 %f147, %f55, %f56;
cvt.rn.f32.s32 %f148, %r292;
add.ftz.f32 %f149, %f142, %f148;
cvt.rn.f32.s32 %f150, %r324;
add.ftz.f32 %f151, %f146, %f150;
mul.ftz.f32 %f152, %f151, %f121;
fma.rn.ftz.f32 %f153, %f149, %f118, %f152;
fma.rn.ftz.f32 %f346, %f147, %f153, %f346;
cvt.u64.u32 %rd66, %r830;
add.s64 %rd67, %rd2, %rd66;
add.s64 %rd68, %rd67, %rd4;
mul.lo.s64 %rd69, %rd68, 36;
add.s64 %rd70, %rd13, %rd69;
ld.global.nc.u32 %r339, [%rd70+4];
ld.global.nc.u32 %r343, [%rd70+8];

dp4a.s32.s32 %r337, %r468, %r339, %r591;


dp4a.s32.s32 %r341, %r472, %r343, %r337;


dp4a.s32.s32 %r345, %r339, %r594, %r591;


dp4a.s32.s32 %r349, %r343, %r594, %r345;

cvt.rn.f32.s32 %f154, %r349;
fma.rn.ftz.f32 %f155, %f68, %f154, 0f00000000;
ld.global.nc.u32 %r355, [%rd70+12];
ld.global.nc.u32 %r359, [%rd70+16];

dp4a.s32.s32 %r353, %r484, %r355, %r341;


dp4a.s32.s32 %r357, %r488, %r359, %r353;


dp4a.s32.s32 %r361, %r355, %r594, %r591;


dp4a.s32.s32 %r365, %r359, %r594, %r361;

cvt.rn.f32.s32 %f156, %r365;
fma.rn.ftz.f32 %f157, %f73, %f156, %f155;
ld.global.nc.u32 %r371, [%rd70+20];
ld.global.nc.u32 %r375, [%rd70+24];

dp4a.s32.s32 %r369, %r500, %r371, %r591;


dp4a.s32.s32 %r373, %r504, %r375, %r369;


dp4a.s32.s32 %r377, %r371, %r594, %r591;


dp4a.s32.s32 %r381, %r375, %r594, %r377;

cvt.rn.f32.s32 %f158, %r381;
fma.rn.ftz.f32 %f159, %f78, %f158, 0f00000000;
ld.global.nc.u32 %r387, [%rd70+28];
ld.global.nc.u32 %r391, [%rd70+32];

dp4a.s32.s32 %r385, %r516, %r387, %r373;


dp4a.s32.s32 %r389, %r520, %r391, %r385;


dp4a.s32.s32 %r393, %r387, %r594, %r591;


dp4a.s32.s32 %r397, %r391, %r594, %r393;

cvt.rn.f32.s32 %f160, %r397;
fma.rn.ftz.f32 %f161, %f83, %f160, %f159;

{ cvt.f32.f16 %f57, %rs7;}


ld.global.nc.u32 %r401, [%rd70];

{.reg .f16 low,high;
mov.b32 {low,high},%r401;
cvt.f32.f16 %f58, low;}


mul.ftz.f32 %f162, %f57, %f58;
cvt.rn.f32.s32 %f163, %r357;
add.ftz.f32 %f164, %f157, %f163;
cvt.rn.f32.s32 %f165, %r389;
add.ftz.f32 %f166, %f161, %f165;
mul.ftz.f32 %f167, %f166, %f92;
fma.rn.ftz.f32 %f168, %f164, %f89, %f167;
fma.rn.ftz.f32 %f345, %f162, %f168, %f345;

dp4a.s32.s32 %r402, %r533, %r339, %r591;


dp4a.s32.s32 %r406, %r537, %r343, %r402;


dp4a.s32.s32 %r410, %r339, %r594, %r591;


dp4a.s32.s32 %r414, %r343, %r594, %r410;

cvt.rn.f32.s32 %f169, %r414;
fma.rn.ftz.f32 %f170, %f97, %f169, 0f00000000;

dp4a.s32.s32 %r418, %r549, %r355, %r406;


dp4a.s32.s32 %r422, %r553, %r359, %r418;


dp4a.s32.s32 %r426, %r355, %r594, %r591;


dp4a.s32.s32 %r430, %r359, %r594, %r426;

cvt.rn.f32.s32 %f171, %r430;
fma.rn.ftz.f32 %f172, %f102, %f171, %f170;

dp4a.s32.s32 %r434, %r565, %r371, %r591;


dp4a.s32.s32 %r438, %r569, %r375, %r434;


dp4a.s32.s32 %r442, %r371, %r594, %r591;


dp4a.s32.s32 %r446, %r375, %r594, %r442;

cvt.rn.f32.s32 %f173, %r446;
fma.rn.ftz.f32 %f174, %f107, %f173, 0f00000000;

dp4a.s32.s32 %r450, %r581, %r387, %r438;


dp4a.s32.s32 %r454, %r585, %r391, %r450;


dp4a.s32.s32 %r458, %r387, %r594, %r591;


dp4a.s32.s32 %r462, %r391, %r594, %r458;

cvt.rn.f32.s32 %f175, %r462;
fma.rn.ftz.f32 %f176, %f112, %f175, %f174;

{ cvt.f32.f16 %f59, %rs8;}



{.reg .f16 low,high;
mov.b32 {low,high},%r401;
cvt.f32.f16 %f60, low;}


mul.ftz.f32 %f177, %f59, %f60;
cvt.rn.f32.s32 %f178, %r422;
add.ftz.f32 %f179, %f172, %f178;
cvt.rn.f32.s32 %f180, %r454;
add.ftz.f32 %f181, %f176, %f180;
mul.ftz.f32 %f182, %f181, %f121;
fma.rn.ftz.f32 %f183, %f179, %f118, %f182;
fma.rn.ftz.f32 %f344, %f177, %f183, %f344;
cvt.u64.u32 %rd71, %r829;
add.s64 %rd72, %rd2, %rd71;
add.s64 %rd73, %rd72, %rd4;
mul.lo.s64 %rd74, %rd73, 36;
add.s64 %rd75, %rd13, %rd74;
ld.global.nc.u32 %r469, [%rd75+4];
ld.global.nc.u32 %r473, [%rd75+8];

dp4a.s32.s32 %r467, %r468, %r469, %r591;


dp4a.s32.s32 %r471, %r472, %r473, %r467;


dp4a.s32.s32 %r475, %r469, %r594, %r591;


dp4a.s32.s32 %r479, %r473, %r594, %r475;

cvt.rn.f32.s32 %f184, %r479;
fma.rn.ftz.f32 %f185, %f68, %f184, 0f00000000;
ld.global.nc.u32 %r485, [%rd75+12];
ld.global.nc.u32 %r489, [%rd75+16];

dp4a.s32.s32 %r483, %r484, %r485, %r471;


dp4a.s32.s32 %r487, %r488, %r489, %r483;


dp4a.s32.s32 %r491, %r485, %r594, %r591;


dp4a.s32.s32 %r495, %r489, %r594, %r491;

cvt.rn.f32.s32 %f186, %r495;
fma.rn.ftz.f32 %f187, %f73, %f186, %f185;
ld.global.nc.u32 %r501, [%rd75+20];
ld.global.nc.u32 %r505, [%rd75+24];

dp4a.s32.s32 %r499, %r500, %r501, %r591;


dp4a.s32.s32 %r503, %r504, %r505, %r499;


dp4a.s32.s32 %r507, %r501, %r594, %r591;


dp4a.s32.s32 %r511, %r505, %r594, %r507;

cvt.rn.f32.s32 %f188, %r511;
fma.rn.ftz.f32 %f189, %f78, %f188, 0f00000000;
ld.global.nc.u32 %r517, [%rd75+28];
ld.global.nc.u32 %r521, [%rd75+32];

dp4a.s32.s32 %r515, %r516, %r517, %r503;


dp4a.s32.s32 %r519, %r520, %r521, %r515;


dp4a.s32.s32 %r523, %r517, %r594, %r591;


dp4a.s32.s32 %r527, %r521, %r594, %r523;

cvt.rn.f32.s32 %f190, %r527;
fma.rn.ftz.f32 %f191, %f83, %f190, %f189;

{ cvt.f32.f16 %f61, %rs7;}


ld.global.nc.u32 %r531, [%rd75];

{.reg .f16 low,high;
mov.b32 {low,high},%r531;
cvt.f32.f16 %f62, low;}


mul.ftz.f32 %f192, %f61, %f62;
cvt.rn.f32.s32 %f193, %r487;
add.ftz.f32 %f194, %f187, %f193;
cvt.rn.f32.s32 %f195, %r519;
add.ftz.f32 %f196, %f191, %f195;
mul.ftz.f32 %f197, %f196, %f92;
fma.rn.ftz.f32 %f198, %f194, %f89, %f197;
fma.rn.ftz.f32 %f343, %f192, %f198, %f343;

dp4a.s32.s32 %r532, %r533, %r469, %r591;


dp4a.s32.s32 %r536, %r537, %r473, %r532;


dp4a.s32.s32 %r540, %r469, %r594, %r591;


dp4a.s32.s32 %r544, %r473, %r594, %r540;

cvt.rn.f32.s32 %f199, %r544;
fma.rn.ftz.f32 %f200, %f97, %f199, 0f00000000;

dp4a.s32.s32 %r548, %r549, %r485, %r536;


dp4a.s32.s32 %r552, %r553, %r489, %r548;


dp4a.s32.s32 %r556, %r485, %r594, %r591;


dp4a.s32.s32 %r560, %r489, %r594, %r556;

cvt.rn.f32.s32 %f201, %r560;
fma.rn.ftz.f32 %f202, %f102, %f201, %f200;

dp4a.s32.s32 %r564, %r565, %r501, %r591;


dp4a.s32.s32 %r568, %r569, %r505, %r564;


dp4a.s32.s32 %r572, %r501, %r594, %r591;


dp4a.s32.s32 %r576, %r505, %r594, %r572;

cvt.rn.f32.s32 %f203, %r576;
fma.rn.ftz.f32 %f204, %f107, %f203, 0f00000000;

dp4a.s32.s32 %r580, %r581, %r517, %r568;


dp4a.s32.s32 %r584, %r585, %r521, %r580;


dp4a.s32.s32 %r588, %r517, %r594, %r591;


dp4a.s32.s32 %r592, %r521, %r594, %r588;

cvt.rn.f32.s32 %f205, %r592;
fma.rn.ftz.f32 %f206, %f112, %f205, %f204;

{ cvt.f32.f16 %f63, %rs8;}



{.reg .f16 low,high;
mov.b32 {low,high},%r531;
cvt.f32.f16 %f64, low;}


mul.ftz.f32 %f207, %f63, %f64;
cvt.rn.f32.s32 %f208, %r552;
add.ftz.f32 %f209, %f202, %f208;
cvt.rn.f32.s32 %f210, %r584;
add.ftz.f32 %f211, %f206, %f210;
mul.ftz.f32 %f212, %f211, %f121;
fma.rn.ftz.f32 %f213, %f209, %f118, %f212;
fma.rn.ftz.f32 %f342, %f207, %f213, %f342;
add.s32 %r830, %r830, 128;
add.s32 %r829, %r829, 128;
add.s32 %r828, %r828, 128;
add.s32 %r827, %r827, 128;
add.s64 %rd96, %rd96, 896;
add.s32 %r831, %r831, 16;
setp.lt.s32 %p4, %r831, %r7;
@%p4 bra $L__BB271_2;

$L__BB271_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB271_5;

shl.b32 %r677, %r1, 10;
mov.u32 %r678, _ZZ13mul_mat_vec_qIL9ggml_type29ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r679, %r678, %r677;
shl.b32 %r680, %r2, 2;
add.s32 %r681, %r679, %r680;
st.shared.f32 [%r681+-1024], %f349;
st.shared.f32 [%r681+-896], %f348;
st.shared.f32 [%r681+-768], %f347;
st.shared.f32 [%r681+-640], %f346;
st.shared.f32 [%r681+-512], %f345;
st.shared.f32 [%r681+-384], %f344;
st.shared.f32 [%r681+-256], %f343;
st.shared.f32 [%r681+-128], %f342;

$L__BB271_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB271_22;

mov.u32 %r682, %ctaid.z;
mad.lo.s32 %r683, %r5, %r34, %r4;
mad.lo.s32 %r684, %r682, %r40, %r683;
cvt.u64.u32 %rd16, %r684;
shl.b32 %r685, %r2, 2;
mov.u32 %r686, _ZZ13mul_mat_vec_qIL9ggml_type29ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r23, %r686, %r685;
mov.u32 %r687, 2;
ld.shared.f32 %f214, [%r23];
add.ftz.f32 %f215, %f214, %f349;
ld.shared.f32 %f216, [%r23+1024];
add.ftz.f32 %f217, %f216, %f215;
ld.shared.f32 %f218, [%r23+2048];
add.ftz.f32 %f219, %f218, %f217;
mov.b32 %r688, %f219;
mov.u32 %r689, 31;
mov.u32 %r690, 16;
mov.u32 %r691, -1;
shfl.sync.bfly.b32 %r692|%p7, %r688, %r690, %r689, %r691;
mov.b32 %f220, %r692;
add.ftz.f32 %f221, %f219, %f220;
mov.b32 %r693, %f221;
mov.u32 %r694, 8;
shfl.sync.bfly.b32 %r695|%p8, %r693, %r694, %r689, %r691;
mov.b32 %f222, %r695;
add.ftz.f32 %f223, %f221, %f222;
mov.b32 %r696, %f223;
mov.u32 %r697, 4;
shfl.sync.bfly.b32 %r698|%p9, %r696, %r697, %r689, %r691;
mov.b32 %f224, %r698;
add.ftz.f32 %f225, %f223, %f224;
mov.b32 %r699, %f225;
shfl.sync.bfly.b32 %r700|%p10, %r699, %r687, %r689, %r691;
mov.b32 %f226, %r700;
add.ftz.f32 %f227, %f225, %f226;
mov.b32 %r701, %f227;
mov.u32 %r702, 1;
shfl.sync.bfly.b32 %r703|%p11, %r701, %r702, %r689, %r691;
mov.b32 %f228, %r703;
add.ftz.f32 %f25, %f227, %f228;
setp.lt.u32 %p12, %r4, %r28;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
mul.wide.u32 %rd76, %r684, 4;
add.s64 %rd17, %rd1, %rd76;
not.pred %p14, %p1;
@%p14 bra $L__BB271_8;

st.global.f32 [%rd17], %f25;

$L__BB271_8:
ld.shared.f32 %f229, [%r23+128];
add.ftz.f32 %f230, %f229, %f348;
ld.shared.f32 %f231, [%r23+1152];
add.ftz.f32 %f232, %f231, %f230;
ld.shared.f32 %f233, [%r23+2176];
add.ftz.f32 %f234, %f233, %f232;
mov.b32 %r704, %f234;
shfl.sync.bfly.b32 %r708|%p15, %r704, %r690, %r689, %r691;
mov.b32 %f235, %r708;
add.ftz.f32 %f236, %f234, %f235;
mov.b32 %r709, %f236;
shfl.sync.bfly.b32 %r711|%p16, %r709, %r694, %r689, %r691;
mov.b32 %f237, %r711;
add.ftz.f32 %f238, %f236, %f237;
mov.b32 %r712, %f238;
shfl.sync.bfly.b32 %r714|%p17, %r712, %r697, %r689, %r691;
mov.b32 %f239, %r714;
add.ftz.f32 %f240, %f238, %f239;
mov.b32 %r715, %f240;
shfl.sync.bfly.b32 %r717|%p18, %r715, %r687, %r689, %r691;
mov.b32 %f241, %r717;
add.ftz.f32 %f242, %f240, %f241;
mov.b32 %r718, %f242;
shfl.sync.bfly.b32 %r720|%p19, %r718, %r702, %r689, %r691;
mov.b32 %f243, %r720;
add.ftz.f32 %f26, %f242, %f243;
add.s32 %r721, %r4, 1;
setp.lt.u32 %p20, %r721, %r28;
setp.eq.s32 %p21, %r2, 1;
and.pred %p2, %p21, %p20;
not.pred %p22, %p2;
@%p22 bra $L__BB271_10;

st.global.f32 [%rd17+4], %f26;

$L__BB271_10:
ld.shared.f32 %f244, [%r23+256];
add.ftz.f32 %f245, %f244, %f347;
ld.shared.f32 %f246, [%r23+1280];
add.ftz.f32 %f247, %f246, %f245;
ld.shared.f32 %f248, [%r23+2304];
add.ftz.f32 %f249, %f248, %f247;
mov.b32 %r722, %f249;
mov.u32 %r723, 31;
mov.u32 %r724, 16;
mov.u32 %r725, -1;
shfl.sync.bfly.b32 %r726|%p23, %r722, %r724, %r723, %r725;
mov.b32 %f250, %r726;
add.ftz.f32 %f251, %f249, %f250;
mov.b32 %r727, %f251;
mov.u32 %r728, 8;
shfl.sync.bfly.b32 %r729|%p24, %r727, %r728, %r723, %r725;
mov.b32 %f252, %r729;
add.ftz.f32 %f253, %f251, %f252;
mov.b32 %r730, %f253;
mov.u32 %r731, 4;
shfl.sync.bfly.b32 %r732|%p25, %r730, %r731, %r723, %r725;
mov.b32 %f254, %r732;
add.ftz.f32 %f255, %f253, %f254;
mov.b32 %r733, %f255;
mov.u32 %r734, 2;
shfl.sync.bfly.b32 %r735|%p26, %r733, %r734, %r723, %r725;
mov.b32 %f256, %r735;
add.ftz.f32 %f257, %f255, %f256;
mov.b32 %r736, %f257;
mov.u32 %r737, 1;
shfl.sync.bfly.b32 %r738|%p27, %r736, %r737, %r723, %r725;
mov.b32 %f258, %r738;
add.ftz.f32 %f27, %f257, %f258;
@%p14 bra $L__BB271_12;

cvt.u64.u32 %rd77, %r28;
add.s64 %rd78, %rd77, %rd16;
shl.b64 %rd79, %rd78, 2;
add.s64 %rd80, %rd1, %rd79;
st.global.f32 [%rd80], %f27;

$L__BB271_12:
ld.shared.f32 %f259, [%r23+384];
add.ftz.f32 %f260, %f259, %f346;
ld.shared.f32 %f261, [%r23+1408];
add.ftz.f32 %f262, %f261, %f260;
ld.shared.f32 %f263, [%r23+2432];
add.ftz.f32 %f264, %f263, %f262;
mov.b32 %r739, %f264;
shfl.sync.bfly.b32 %r743|%p29, %r739, %r724, %r723, %r725;
mov.b32 %f265, %r743;
add.ftz.f32 %f266, %f264, %f265;
mov.b32 %r744, %f266;
shfl.sync.bfly.b32 %r746|%p30, %r744, %r728, %r723, %r725;
mov.b32 %f267, %r746;
add.ftz.f32 %f268, %f266, %f267;
mov.b32 %r747, %f268;
shfl.sync.bfly.b32 %r749|%p31, %r747, %r731, %r723, %r725;
mov.b32 %f269, %r749;
add.ftz.f32 %f270, %f268, %f269;
mov.b32 %r750, %f270;
shfl.sync.bfly.b32 %r752|%p32, %r750, %r734, %r723, %r725;
mov.b32 %f271, %r752;
add.ftz.f32 %f272, %f270, %f271;
mov.b32 %r753, %f272;
shfl.sync.bfly.b32 %r755|%p33, %r753, %r737, %r723, %r725;
mov.b32 %f273, %r755;
add.ftz.f32 %f28, %f272, %f273;
@%p22 bra $L__BB271_14;

add.s32 %r756, %r28, 1;
cvt.u64.u32 %rd81, %r756;
add.s64 %rd82, %rd81, %rd16;
shl.b64 %rd83, %rd82, 2;
add.s64 %rd84, %rd1, %rd83;
st.global.f32 [%rd84], %f28;

$L__BB271_14:
shl.b32 %r757, %r28, 1;
mov.u32 %r758, 1;
ld.shared.f32 %f274, [%r23+512];
add.ftz.f32 %f275, %f274, %f345;
ld.shared.f32 %f276, [%r23+1536];
add.ftz.f32 %f277, %f276, %f275;
ld.shared.f32 %f278, [%r23+2560];
add.ftz.f32 %f279, %f278, %f277;
mov.b32 %r759, %f279;
mov.u32 %r760, 31;
mov.u32 %r761, 16;
mov.u32 %r762, -1;
shfl.sync.bfly.b32 %r763|%p35, %r759, %r761, %r760, %r762;
mov.b32 %f280, %r763;
add.ftz.f32 %f281, %f279, %f280;
mov.b32 %r764, %f281;
mov.u32 %r765, 8;
shfl.sync.bfly.b32 %r766|%p36, %r764, %r765, %r760, %r762;
mov.b32 %f282, %r766;
add.ftz.f32 %f283, %f281, %f282;
mov.b32 %r767, %f283;
mov.u32 %r768, 4;
shfl.sync.bfly.b32 %r769|%p37, %r767, %r768, %r760, %r762;
mov.b32 %f284, %r769;
add.ftz.f32 %f285, %f283, %f284;
mov.b32 %r770, %f285;
mov.u32 %r771, 2;
shfl.sync.bfly.b32 %r772|%p38, %r770, %r771, %r760, %r762;
mov.b32 %f286, %r772;
add.ftz.f32 %f287, %f285, %f286;
mov.b32 %r773, %f287;
shfl.sync.bfly.b32 %r774|%p39, %r773, %r758, %r760, %r762;
mov.b32 %f288, %r774;
add.ftz.f32 %f29, %f287, %f288;
cvt.u64.u32 %rd85, %r757;
add.s64 %rd86, %rd85, %rd16;
shl.b64 %rd87, %rd86, 2;
add.s64 %rd18, %rd1, %rd87;
@%p14 bra $L__BB271_16;

st.global.f32 [%rd18], %f29;

$L__BB271_16:
ld.shared.f32 %f289, [%r23+640];
add.ftz.f32 %f290, %f289, %f344;
ld.shared.f32 %f291, [%r23+1664];
add.ftz.f32 %f292, %f291, %f290;
ld.shared.f32 %f293, [%r23+2688];
add.ftz.f32 %f294, %f293, %f292;
mov.b32 %r775, %f294;
shfl.sync.bfly.b32 %r779|%p41, %r775, %r761, %r760, %r762;
mov.b32 %f295, %r779;
add.ftz.f32 %f296, %f294, %f295;
mov.b32 %r780, %f296;
shfl.sync.bfly.b32 %r782|%p42, %r780, %r765, %r760, %r762;
mov.b32 %f297, %r782;
add.ftz.f32 %f298, %f296, %f297;
mov.b32 %r783, %f298;
shfl.sync.bfly.b32 %r785|%p43, %r783, %r768, %r760, %r762;
mov.b32 %f299, %r785;
add.ftz.f32 %f300, %f298, %f299;
mov.b32 %r786, %f300;
shfl.sync.bfly.b32 %r788|%p44, %r786, %r771, %r760, %r762;
mov.b32 %f301, %r788;
add.ftz.f32 %f302, %f300, %f301;
mov.b32 %r789, %f302;
shfl.sync.bfly.b32 %r791|%p45, %r789, %r758, %r760, %r762;
mov.b32 %f303, %r791;
add.ftz.f32 %f30, %f302, %f303;
@%p22 bra $L__BB271_18;

st.global.f32 [%rd18+4], %f30;

$L__BB271_18:
mul.lo.s32 %r25, %r28, 3;
ld.shared.f32 %f304, [%r23+768];
add.ftz.f32 %f305, %f304, %f343;
ld.shared.f32 %f306, [%r23+1792];
add.ftz.f32 %f307, %f306, %f305;
ld.shared.f32 %f308, [%r23+2816];
add.ftz.f32 %f309, %f308, %f307;
mov.b32 %r792, %f309;
mov.u32 %r793, 31;
mov.u32 %r794, 16;
mov.u32 %r795, -1;
shfl.sync.bfly.b32 %r796|%p47, %r792, %r794, %r793, %r795;
mov.b32 %f310, %r796;
add.ftz.f32 %f311, %f309, %f310;
mov.b32 %r797, %f311;
mov.u32 %r798, 8;
shfl.sync.bfly.b32 %r799|%p48, %r797, %r798, %r793, %r795;
mov.b32 %f312, %r799;
add.ftz.f32 %f313, %f311, %f312;
mov.b32 %r800, %f313;
mov.u32 %r801, 4;
shfl.sync.bfly.b32 %r802|%p49, %r800, %r801, %r793, %r795;
mov.b32 %f314, %r802;
add.ftz.f32 %f315, %f313, %f314;
mov.b32 %r803, %f315;
mov.u32 %r804, 2;
shfl.sync.bfly.b32 %r805|%p50, %r803, %r804, %r793, %r795;
mov.b32 %f316, %r805;
add.ftz.f32 %f317, %f315, %f316;
mov.b32 %r806, %f317;
mov.u32 %r807, 1;
shfl.sync.bfly.b32 %r808|%p51, %r806, %r807, %r793, %r795;
mov.b32 %f318, %r808;
add.ftz.f32 %f31, %f317, %f318;
@%p14 bra $L__BB271_20;

cvt.u64.u32 %rd88, %r25;
add.s64 %rd89, %rd88, %rd16;
shl.b64 %rd90, %rd89, 2;
add.s64 %rd91, %rd1, %rd90;
st.global.f32 [%rd91], %f31;

$L__BB271_20:
ld.shared.f32 %f319, [%r23+896];
add.ftz.f32 %f320, %f319, %f342;
ld.shared.f32 %f321, [%r23+1920];
add.ftz.f32 %f322, %f321, %f320;
ld.shared.f32 %f323, [%r23+2944];
add.ftz.f32 %f324, %f323, %f322;
mov.b32 %r809, %f324;
shfl.sync.bfly.b32 %r813|%p53, %r809, %r794, %r793, %r795;
mov.b32 %f325, %r813;
add.ftz.f32 %f326, %f324, %f325;
mov.b32 %r814, %f326;
shfl.sync.bfly.b32 %r816|%p54, %r814, %r798, %r793, %r795;
mov.b32 %f327, %r816;
add.ftz.f32 %f328, %f326, %f327;
mov.b32 %r817, %f328;
shfl.sync.bfly.b32 %r819|%p55, %r817, %r801, %r793, %r795;
mov.b32 %f329, %r819;
add.ftz.f32 %f330, %f328, %f329;
mov.b32 %r820, %f330;
shfl.sync.bfly.b32 %r822|%p56, %r820, %r804, %r793, %r795;
mov.b32 %f331, %r822;
add.ftz.f32 %f332, %f330, %f331;
mov.b32 %r823, %f332;
shfl.sync.bfly.b32 %r825|%p57, %r823, %r807, %r793, %r795;
mov.b32 %f333, %r825;
add.ftz.f32 %f32, %f332, %f333;
@%p22 bra $L__BB271_22;

add.s32 %r826, %r25, 1;
cvt.u64.u32 %rd92, %r826;
add.s64 %rd93, %rd92, %rd16;
shl.b64 %rd94, %rd93, 2;
add.s64 %rd95, %rd1, %rd94;
st.global.f32 [%rd95], %f32;

$L__BB271_22:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type29ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type29ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type29ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type29ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type29ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type29ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type29ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type29ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type29ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<71>;
.reg .b16 %rs<49>;
.reg .f32 %f<390>;
.reg .b32 %r<1000>;
.reg .b64 %rd<106>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type29ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1280];

ld.param.u64 %rd20, [_Z13mul_mat_vec_qIL9ggml_type29ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd21, [_Z13mul_mat_vec_qIL9ggml_type29ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd22, [_Z13mul_mat_vec_qIL9ggml_type29ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type29ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type29ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type29ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type29ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type29ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type29ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type29ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type29ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type29ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type29ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type29ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type29ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type29ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type29ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
cvta.to.global.u64 %rd1, %rd22;
mov.u32 %r1, %tid.y;
shl.b32 %r44, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r44, %r2;
mov.u32 %r45, %ctaid.x;
shl.b32 %r4, %r45, 1;
mov.u32 %r5, %ctaid.y;
shr.s32 %r46, %r3, 31;
shr.u32 %r47, %r46, 29;
add.s32 %r48, %r3, %r47;
shr.s32 %r999, %r48, 3;
shr.u32 %r7, %r43, 8;
setp.ge.s32 %p3, %r999, %r7;
mov.f32 %f380, 0f00000000;
mov.f32 %f381, %f380;
mov.f32 %f382, %f380;
mov.f32 %f383, %f380;
mov.f32 %f384, %f380;
mov.f32 %f385, %f380;
mov.f32 %f386, %f380;
mov.f32 %f387, %f380;
mov.f32 %f388, %f380;
mov.f32 %f389, %f380;
@%p3 bra $L__BB272_3;

mov.u32 %r49, %ctaid.z;
mul.lo.s32 %r50, %r5, %r35;
cvt.u64.u32 %rd23, %r50;
mul.lo.s32 %r51, %r49, %r41;
cvt.u64.u32 %rd24, %r51;
add.s64 %rd2, %rd24, %rd23;
mul.hi.u32 %r52, %r49, %r37;
add.s32 %r53, %r52, %r49;
shr.u32 %r54, %r53, %r38;
mul.lo.s32 %r55, %r54, %r40;
mul.hi.u32 %r56, %r5, %r31;
add.s32 %r57, %r56, %r5;
shr.u32 %r58, %r57, %r32;
mul.lo.s32 %r59, %r58, %r34;
cvta.to.global.u64 %rd105, %rd20;
add.s32 %r60, %r4, 1;
and.b32 %r64, %r48, -8;
sub.s32 %r65, %r3, %r64;
shl.b32 %r66, %r65, 1;
cvt.s64.s32 %rd25, %r66;
cvt.s64.s32 %rd4, %r65;
shr.u32 %r67, %r65, 31;
add.s32 %r68, %r65, %r67;
shr.s32 %r69, %r68, 1;
and.b32 %r70, %r68, -2;
sub.s32 %r71, %r65, %r70;
mul.lo.s32 %r8, %r71, 6;
shl.b32 %r72, %r29, 1;
shl.b32 %r994, %r999, 3;
add.s32 %r998, %r994, %r72;
mad.lo.s32 %r997, %r29, 3, %r994;
shl.b32 %r73, %r29, 2;
add.s32 %r996, %r994, %r73;
add.s32 %r995, %r29, %r994;
add.s32 %r74, %r999, %r55;
add.s32 %r75, %r74, %r59;
mad.lo.s32 %r76, %r60, %r28, %r75;
mul.wide.s32 %rd5, %r76, 56;
mad.lo.s32 %r77, %r4, %r28, %r999;
add.s32 %r78, %r77, %r55;
add.s32 %r79, %r78, %r59;
mul.wide.s32 %rd6, %r79, 56;
mul.wide.s32 %rd26, %r65, 4;
add.s64 %rd7, %rd6, %rd26;
mul.wide.s32 %rd27, %r69, 2;
add.s64 %rd28, %rd6, %rd27;
add.s64 %rd8, %rd28, 48;
add.s64 %rd29, %rd5, %rd27;
add.s64 %rd9, %rd29, 48;
add.s64 %rd30, %rd6, %rd25;
add.s64 %rd10, %rd30, 33;
add.s64 %rd31, %rd5, %rd25;
add.s64 %rd11, %rd31, 32;
add.s64 %rd12, %rd5, %rd26;
cvta.to.global.u64 %rd13, %rd21;
mov.f32 %f380, 0f00000000;

$L__BB272_2:
cvt.u64.u32 %rd32, %r994;
add.s64 %rd33, %rd2, %rd32;
add.s64 %rd34, %rd33, %rd4;
mul.lo.s64 %rd35, %rd34, 36;
add.s64 %rd36, %rd105, %rd7;
ld.global.nc.u32 %r730, [%rd36];
shr.u32 %r731, %r730, 8;
shr.u32 %r732, %r730, 16;
shr.u32 %r733, %r730, 24;
add.s64 %rd37, %rd105, %rd10;
ld.global.nc.u8 %rs11, [%rd37+-1];
and.b16 %rs12, %rs11, 240;
cvt.u32.u16 %r734, %rs11;
and.b32 %r735, %r730, 255;
and.b32 %r736, %r734, 7;
bfi.b32 %r737, %r736, %r735, 8, 3;
mul.wide.u32 %rd38, %r737, 4;
mov.u64 %rd39, iq1s_grid_gpu;
add.s64 %rd40, %rd39, %rd38;
ld.global.nc.u32 %r738, [%rd40];
and.b32 %r601, %r738, 252645135;
shr.u32 %r739, %r738, 4;
and.b32 %r605, %r739, 252645135;
add.s64 %rd41, %rd13, %rd35;
ld.global.nc.u32 %r82, [%rd41+4];
ld.global.nc.u32 %r86, [%rd41+8];
mov.u32 %r724, 0;

dp4a.s32.s32 %r80, %r601, %r82, %r724;


dp4a.s32.s32 %r84, %r605, %r86, %r80;

and.b32 %r740, %r734, 8;
cvt.rn.f32.s32 %f81, %r740;
mul.ftz.f32 %f82, %f81, 0f3D000000;
mov.f32 %f83, 0fBF600000;
sub.ftz.f32 %f84, %f83, %f82;
mov.u32 %r727, 16843009;

dp4a.s32.s32 %r88, %r82, %r727, %r724;


dp4a.s32.s32 %r92, %r86, %r727, %r88;

cvt.rn.f32.s32 %f85, %r92;
fma.rn.ftz.f32 %f86, %f84, %f85, 0f00000000;
shr.u16 %rs13, %rs12, 4;
cvt.u32.u16 %r741, %rs13;
and.b32 %r742, %r731, 255;
and.b32 %r743, %r741, 7;
bfi.b32 %r744, %r743, %r742, 8, 3;
mul.wide.u32 %rd42, %r744, 4;
add.s64 %rd43, %rd39, %rd42;
ld.global.nc.u32 %r745, [%rd43];
and.b32 %r617, %r745, 252645135;
shr.u32 %r746, %r745, 4;
and.b32 %r621, %r746, 252645135;
ld.global.nc.u32 %r98, [%rd41+12];
ld.global.nc.u32 %r102, [%rd41+16];

dp4a.s32.s32 %r96, %r617, %r98, %r84;


dp4a.s32.s32 %r100, %r621, %r102, %r96;

and.b32 %r747, %r741, 8;
cvt.rn.f32.s32 %f87, %r747;
mul.ftz.f32 %f88, %f87, 0f3D000000;
sub.ftz.f32 %f89, %f83, %f88;

dp4a.s32.s32 %r104, %r98, %r727, %r724;


dp4a.s32.s32 %r108, %r102, %r727, %r104;

cvt.rn.f32.s32 %f90, %r108;
fma.rn.ftz.f32 %f91, %f89, %f90, %f86;
ld.global.nc.u8 %rs14, [%rd37];
and.b16 %rs15, %rs14, 240;
cvt.u32.u16 %r748, %rs14;
and.b32 %r749, %r732, 255;
and.b32 %r750, %r748, 7;
bfi.b32 %r751, %r750, %r749, 8, 3;
mul.wide.u32 %rd44, %r751, 4;
add.s64 %rd45, %rd39, %rd44;
ld.global.nc.u32 %r752, [%rd45];
and.b32 %r633, %r752, 252645135;
shr.u32 %r753, %r752, 4;
and.b32 %r637, %r753, 252645135;
ld.global.nc.u32 %r114, [%rd41+20];
ld.global.nc.u32 %r118, [%rd41+24];

dp4a.s32.s32 %r112, %r633, %r114, %r724;


dp4a.s32.s32 %r116, %r637, %r118, %r112;

and.b32 %r754, %r748, 8;
cvt.rn.f32.s32 %f92, %r754;
mul.ftz.f32 %f93, %f92, 0f3D000000;
sub.ftz.f32 %f94, %f83, %f93;

dp4a.s32.s32 %r120, %r114, %r727, %r724;


dp4a.s32.s32 %r124, %r118, %r727, %r120;

cvt.rn.f32.s32 %f95, %r124;
fma.rn.ftz.f32 %f96, %f94, %f95, 0f00000000;
shr.u16 %rs16, %rs15, 4;
cvt.u32.u16 %r755, %rs16;
mul.wide.u16 %r756, %rs16, 256;
and.b32 %r757, %r756, 1792;
or.b32 %r758, %r757, %r733;
mul.wide.u32 %rd46, %r758, 4;
add.s64 %rd47, %rd39, %rd46;
ld.global.nc.u32 %r759, [%rd47];
and.b32 %r649, %r759, 252645135;
shr.u32 %r760, %r759, 4;
and.b32 %r653, %r760, 252645135;
ld.global.nc.u32 %r130, [%rd41+28];
ld.global.nc.u32 %r134, [%rd41+32];

dp4a.s32.s32 %r128, %r649, %r130, %r116;


dp4a.s32.s32 %r132, %r653, %r134, %r128;

and.b32 %r761, %r755, 8;
cvt.rn.f32.s32 %f97, %r761;
mul.ftz.f32 %f98, %f97, 0f3D000000;
sub.ftz.f32 %f99, %f83, %f98;

dp4a.s32.s32 %r136, %r130, %r727, %r724;


dp4a.s32.s32 %r140, %r134, %r727, %r136;

cvt.rn.f32.s32 %f100, %r140;
fma.rn.ftz.f32 %f101, %f99, %f100, %f96;
add.s64 %rd48, %rd105, %rd6;
ld.global.nc.u16 %rs17, [%rd48+48];
shr.u16 %rs18, %rs17, 12;
ld.global.nc.u16 %rs19, [%rd48+50];
shr.u16 %rs20, %rs19, 8;
and.b16 %rs21, %rs20, 240;
or.b16 %rs22, %rs21, %rs18;
ld.global.nc.u16 %rs23, [%rd48+52];
shr.u16 %rs24, %rs23, 4;
and.b16 %rs25, %rs24, 3840;
or.b16 %rs26, %rs22, %rs25;
ld.global.nc.u16 %rs27, [%rd48+54];
and.b16 %rs28, %rs27, -4096;
or.b16 %rs9, %rs26, %rs28;

{ cvt.f32.f16 %f61, %rs9;}


ld.global.nc.u32 %r144, [%rd41];

{.reg .f16 low,high;
mov.b32 {low,high},%r144;
cvt.f32.f16 %f62, low;}


mul.ftz.f32 %f102, %f61, %f62;
add.s64 %rd49, %rd105, %rd8;
ld.global.nc.u16 %rs29, [%rd49];
cvt.u32.u16 %r762, %rs29;
shr.u32 %r763, %r762, %r8;
shl.b32 %r764, %r763, 1;
and.b32 %r765, %r764, 14;
or.b32 %r766, %r765, 1;
shr.u32 %r767, %r763, 2;
and.b32 %r768, %r767, 14;
or.b32 %r769, %r768, 1;
cvt.rn.f32.s32 %f103, %r100;
add.ftz.f32 %f104, %f91, %f103;
cvt.rn.f32.s32 %f105, %r766;
cvt.rn.f32.s32 %f106, %r132;
add.ftz.f32 %f107, %f101, %f106;
cvt.rn.f32.s32 %f108, %r769;
mul.ftz.f32 %f109, %f107, %f108;
fma.rn.ftz.f32 %f110, %f104, %f105, %f109;
fma.rn.ftz.f32 %f389, %f102, %f110, %f389;
add.s64 %rd50, %rd105, %rd12;
ld.global.nc.u32 %r770, [%rd50];
shr.u32 %r771, %r770, 8;
shr.u32 %r772, %r770, 16;
shr.u32 %r773, %r770, 24;
add.s64 %rd51, %rd105, %rd11;
ld.global.nc.u8 %rs30, [%rd51];
and.b16 %rs31, %rs30, 240;
cvt.u32.u16 %r774, %rs30;
and.b32 %r775, %r770, 255;
and.b32 %r776, %r774, 7;
bfi.b32 %r777, %r776, %r775, 8, 3;
mul.wide.u32 %rd52, %r777, 4;
add.s64 %rd53, %rd39, %rd52;
ld.global.nc.u32 %r778, [%rd53];
and.b32 %r666, %r778, 252645135;
shr.u32 %r779, %r778, 4;
and.b32 %r670, %r779, 252645135;

dp4a.s32.s32 %r145, %r666, %r82, %r724;


dp4a.s32.s32 %r149, %r670, %r86, %r145;

and.b32 %r780, %r774, 8;
cvt.rn.f32.s32 %f111, %r780;
mul.ftz.f32 %f112, %f111, 0f3D000000;
sub.ftz.f32 %f113, %f83, %f112;

dp4a.s32.s32 %r153, %r82, %r727, %r724;


dp4a.s32.s32 %r157, %r86, %r727, %r153;

cvt.rn.f32.s32 %f114, %r157;
fma.rn.ftz.f32 %f115, %f113, %f114, 0f00000000;
shr.u16 %rs32, %rs31, 4;
cvt.u32.u16 %r781, %rs32;
and.b32 %r782, %r771, 255;
and.b32 %r783, %r781, 7;
bfi.b32 %r784, %r783, %r782, 8, 3;
mul.wide.u32 %rd54, %r784, 4;
add.s64 %rd55, %rd39, %rd54;
ld.global.nc.u32 %r785, [%rd55];
and.b32 %r682, %r785, 252645135;
shr.u32 %r786, %r785, 4;
and.b32 %r686, %r786, 252645135;

dp4a.s32.s32 %r161, %r682, %r98, %r149;


dp4a.s32.s32 %r165, %r686, %r102, %r161;

and.b32 %r787, %r781, 8;
cvt.rn.f32.s32 %f116, %r787;
mul.ftz.f32 %f117, %f116, 0f3D000000;
sub.ftz.f32 %f118, %f83, %f117;

dp4a.s32.s32 %r169, %r98, %r727, %r724;


dp4a.s32.s32 %r173, %r102, %r727, %r169;

cvt.rn.f32.s32 %f119, %r173;
fma.rn.ftz.f32 %f120, %f118, %f119, %f115;
ld.global.nc.u8 %rs33, [%rd51+1];
and.b16 %rs34, %rs33, 240;
cvt.u32.u16 %r788, %rs33;
and.b32 %r789, %r772, 255;
and.b32 %r790, %r788, 7;
bfi.b32 %r791, %r790, %r789, 8, 3;
mul.wide.u32 %rd56, %r791, 4;
add.s64 %rd57, %rd39, %rd56;
ld.global.nc.u32 %r792, [%rd57];
and.b32 %r698, %r792, 252645135;
shr.u32 %r793, %r792, 4;
and.b32 %r702, %r793, 252645135;

dp4a.s32.s32 %r177, %r698, %r114, %r724;


dp4a.s32.s32 %r181, %r702, %r118, %r177;

and.b32 %r794, %r788, 8;
cvt.rn.f32.s32 %f121, %r794;
mul.ftz.f32 %f122, %f121, 0f3D000000;
sub.ftz.f32 %f123, %f83, %f122;

dp4a.s32.s32 %r185, %r114, %r727, %r724;


dp4a.s32.s32 %r189, %r118, %r727, %r185;

cvt.rn.f32.s32 %f124, %r189;
fma.rn.ftz.f32 %f125, %f123, %f124, 0f00000000;
shr.u16 %rs35, %rs34, 4;
cvt.u32.u16 %r795, %rs35;
mul.wide.u16 %r796, %rs35, 256;
and.b32 %r797, %r796, 1792;
or.b32 %r798, %r797, %r773;
mul.wide.u32 %rd58, %r798, 4;
add.s64 %rd59, %rd39, %rd58;
ld.global.nc.u32 %r799, [%rd59];
and.b32 %r714, %r799, 252645135;
shr.u32 %r800, %r799, 4;
and.b32 %r718, %r800, 252645135;

dp4a.s32.s32 %r193, %r714, %r130, %r181;


dp4a.s32.s32 %r197, %r718, %r134, %r193;

and.b32 %r801, %r795, 8;
cvt.rn.f32.s32 %f126, %r801;
mul.ftz.f32 %f127, %f126, 0f3D000000;
sub.ftz.f32 %f128, %f83, %f127;

dp4a.s32.s32 %r201, %r130, %r727, %r724;


dp4a.s32.s32 %r205, %r134, %r727, %r201;

cvt.rn.f32.s32 %f129, %r205;
fma.rn.ftz.f32 %f130, %f128, %f129, %f125;
add.s64 %rd60, %rd105, %rd5;
ld.global.nc.u16 %rs36, [%rd60+48];
shr.u16 %rs37, %rs36, 12;
ld.global.nc.u16 %rs38, [%rd60+50];
shr.u16 %rs39, %rs38, 8;
and.b16 %rs40, %rs39, 240;
or.b16 %rs41, %rs40, %rs37;
ld.global.nc.u16 %rs42, [%rd60+52];
shr.u16 %rs43, %rs42, 4;
and.b16 %rs44, %rs43, 3840;
or.b16 %rs45, %rs41, %rs44;
ld.global.nc.u16 %rs46, [%rd60+54];
and.b16 %rs47, %rs46, -4096;
or.b16 %rs10, %rs45, %rs47;

{ cvt.f32.f16 %f63, %rs10;}



{.reg .f16 low,high;
mov.b32 {low,high},%r144;
cvt.f32.f16 %f64, low;}


mul.ftz.f32 %f131, %f63, %f64;
add.s64 %rd61, %rd105, %rd9;
ld.global.nc.u16 %rs48, [%rd61];
cvt.u32.u16 %r802, %rs48;
shr.u32 %r803, %r802, %r8;
shl.b32 %r804, %r803, 1;
and.b32 %r805, %r804, 14;
or.b32 %r806, %r805, 1;
shr.u32 %r807, %r803, 2;
and.b32 %r808, %r807, 14;
or.b32 %r809, %r808, 1;
cvt.rn.f32.s32 %f132, %r165;
add.ftz.f32 %f133, %f120, %f132;
cvt.rn.f32.s32 %f134, %r806;
cvt.rn.f32.s32 %f135, %r197;
add.ftz.f32 %f136, %f130, %f135;
cvt.rn.f32.s32 %f137, %r809;
mul.ftz.f32 %f138, %f136, %f137;
fma.rn.ftz.f32 %f139, %f133, %f134, %f138;
fma.rn.ftz.f32 %f388, %f131, %f139, %f388;
cvt.u64.u32 %rd62, %r995;
add.s64 %rd63, %rd2, %rd62;
add.s64 %rd64, %rd63, %rd4;
mul.lo.s64 %rd65, %rd64, 36;
add.s64 %rd66, %rd13, %rd65;
ld.global.nc.u32 %r212, [%rd66+4];
ld.global.nc.u32 %r216, [%rd66+8];

dp4a.s32.s32 %r210, %r601, %r212, %r724;


dp4a.s32.s32 %r214, %r605, %r216, %r210;


dp4a.s32.s32 %r218, %r212, %r727, %r724;


dp4a.s32.s32 %r222, %r216, %r727, %r218;

cvt.rn.f32.s32 %f140, %r222;
fma.rn.ftz.f32 %f141, %f84, %f140, 0f00000000;
ld.global.nc.u32 %r228, [%rd66+12];
ld.global.nc.u32 %r232, [%rd66+16];

dp4a.s32.s32 %r226, %r617, %r228, %r214;


dp4a.s32.s32 %r230, %r621, %r232, %r226;


dp4a.s32.s32 %r234, %r228, %r727, %r724;


dp4a.s32.s32 %r238, %r232, %r727, %r234;

cvt.rn.f32.s32 %f142, %r238;
fma.rn.ftz.f32 %f143, %f89, %f142, %f141;
ld.global.nc.u32 %r244, [%rd66+20];
ld.global.nc.u32 %r248, [%rd66+24];

dp4a.s32.s32 %r242, %r633, %r244, %r724;


dp4a.s32.s32 %r246, %r637, %r248, %r242;


dp4a.s32.s32 %r250, %r244, %r727, %r724;


dp4a.s32.s32 %r254, %r248, %r727, %r250;

cvt.rn.f32.s32 %f144, %r254;
fma.rn.ftz.f32 %f145, %f94, %f144, 0f00000000;
ld.global.nc.u32 %r260, [%rd66+28];
ld.global.nc.u32 %r264, [%rd66+32];

dp4a.s32.s32 %r258, %r649, %r260, %r246;


dp4a.s32.s32 %r262, %r653, %r264, %r258;


dp4a.s32.s32 %r266, %r260, %r727, %r724;


dp4a.s32.s32 %r270, %r264, %r727, %r266;

cvt.rn.f32.s32 %f146, %r270;
fma.rn.ftz.f32 %f147, %f99, %f146, %f145;

{ cvt.f32.f16 %f65, %rs9;}


ld.global.nc.u32 %r274, [%rd66];

{.reg .f16 low,high;
mov.b32 {low,high},%r274;
cvt.f32.f16 %f66, low;}


mul.ftz.f32 %f148, %f65, %f66;
cvt.rn.f32.s32 %f149, %r230;
add.ftz.f32 %f150, %f143, %f149;
cvt.rn.f32.s32 %f151, %r262;
add.ftz.f32 %f152, %f147, %f151;
mul.ftz.f32 %f153, %f152, %f108;
fma.rn.ftz.f32 %f154, %f150, %f105, %f153;
fma.rn.ftz.f32 %f387, %f148, %f154, %f387;

dp4a.s32.s32 %r275, %r666, %r212, %r724;


dp4a.s32.s32 %r279, %r670, %r216, %r275;


dp4a.s32.s32 %r283, %r212, %r727, %r724;


dp4a.s32.s32 %r287, %r216, %r727, %r283;

cvt.rn.f32.s32 %f155, %r287;
fma.rn.ftz.f32 %f156, %f113, %f155, 0f00000000;

dp4a.s32.s32 %r291, %r682, %r228, %r279;


dp4a.s32.s32 %r295, %r686, %r232, %r291;


dp4a.s32.s32 %r299, %r228, %r727, %r724;


dp4a.s32.s32 %r303, %r232, %r727, %r299;

cvt.rn.f32.s32 %f157, %r303;
fma.rn.ftz.f32 %f158, %f118, %f157, %f156;

dp4a.s32.s32 %r307, %r698, %r244, %r724;


dp4a.s32.s32 %r311, %r702, %r248, %r307;


dp4a.s32.s32 %r315, %r244, %r727, %r724;


dp4a.s32.s32 %r319, %r248, %r727, %r315;

cvt.rn.f32.s32 %f159, %r319;
fma.rn.ftz.f32 %f160, %f123, %f159, 0f00000000;

dp4a.s32.s32 %r323, %r714, %r260, %r311;


dp4a.s32.s32 %r327, %r718, %r264, %r323;


dp4a.s32.s32 %r331, %r260, %r727, %r724;


dp4a.s32.s32 %r335, %r264, %r727, %r331;

cvt.rn.f32.s32 %f161, %r335;
fma.rn.ftz.f32 %f162, %f128, %f161, %f160;

{ cvt.f32.f16 %f67, %rs10;}



{.reg .f16 low,high;
mov.b32 {low,high},%r274;
cvt.f32.f16 %f68, low;}


mul.ftz.f32 %f163, %f67, %f68;
cvt.rn.f32.s32 %f164, %r295;
add.ftz.f32 %f165, %f158, %f164;
cvt.rn.f32.s32 %f166, %r327;
add.ftz.f32 %f167, %f162, %f166;
mul.ftz.f32 %f168, %f167, %f137;
fma.rn.ftz.f32 %f169, %f165, %f134, %f168;
fma.rn.ftz.f32 %f386, %f163, %f169, %f386;
cvt.u64.u32 %rd67, %r998;
add.s64 %rd68, %rd2, %rd67;
add.s64 %rd69, %rd68, %rd4;
mul.lo.s64 %rd70, %rd69, 36;
add.s64 %rd71, %rd13, %rd70;
ld.global.nc.u32 %r342, [%rd71+4];
ld.global.nc.u32 %r346, [%rd71+8];

dp4a.s32.s32 %r340, %r601, %r342, %r724;


dp4a.s32.s32 %r344, %r605, %r346, %r340;


dp4a.s32.s32 %r348, %r342, %r727, %r724;


dp4a.s32.s32 %r352, %r346, %r727, %r348;

cvt.rn.f32.s32 %f170, %r352;
fma.rn.ftz.f32 %f171, %f84, %f170, 0f00000000;
ld.global.nc.u32 %r358, [%rd71+12];
ld.global.nc.u32 %r362, [%rd71+16];

dp4a.s32.s32 %r356, %r617, %r358, %r344;


dp4a.s32.s32 %r360, %r621, %r362, %r356;


dp4a.s32.s32 %r364, %r358, %r727, %r724;


dp4a.s32.s32 %r368, %r362, %r727, %r364;

cvt.rn.f32.s32 %f172, %r368;
fma.rn.ftz.f32 %f173, %f89, %f172, %f171;
ld.global.nc.u32 %r374, [%rd71+20];
ld.global.nc.u32 %r378, [%rd71+24];

dp4a.s32.s32 %r372, %r633, %r374, %r724;


dp4a.s32.s32 %r376, %r637, %r378, %r372;


dp4a.s32.s32 %r380, %r374, %r727, %r724;


dp4a.s32.s32 %r384, %r378, %r727, %r380;

cvt.rn.f32.s32 %f174, %r384;
fma.rn.ftz.f32 %f175, %f94, %f174, 0f00000000;
ld.global.nc.u32 %r390, [%rd71+28];
ld.global.nc.u32 %r394, [%rd71+32];

dp4a.s32.s32 %r388, %r649, %r390, %r376;


dp4a.s32.s32 %r392, %r653, %r394, %r388;


dp4a.s32.s32 %r396, %r390, %r727, %r724;


dp4a.s32.s32 %r400, %r394, %r727, %r396;

cvt.rn.f32.s32 %f176, %r400;
fma.rn.ftz.f32 %f177, %f99, %f176, %f175;

{ cvt.f32.f16 %f69, %rs9;}


ld.global.nc.u32 %r404, [%rd71];

{.reg .f16 low,high;
mov.b32 {low,high},%r404;
cvt.f32.f16 %f70, low;}


mul.ftz.f32 %f178, %f69, %f70;
cvt.rn.f32.s32 %f179, %r360;
add.ftz.f32 %f180, %f173, %f179;
cvt.rn.f32.s32 %f181, %r392;
add.ftz.f32 %f182, %f177, %f181;
mul.ftz.f32 %f183, %f182, %f108;
fma.rn.ftz.f32 %f184, %f180, %f105, %f183;
fma.rn.ftz.f32 %f385, %f178, %f184, %f385;

dp4a.s32.s32 %r405, %r666, %r342, %r724;


dp4a.s32.s32 %r409, %r670, %r346, %r405;


dp4a.s32.s32 %r413, %r342, %r727, %r724;


dp4a.s32.s32 %r417, %r346, %r727, %r413;

cvt.rn.f32.s32 %f185, %r417;
fma.rn.ftz.f32 %f186, %f113, %f185, 0f00000000;

dp4a.s32.s32 %r421, %r682, %r358, %r409;


dp4a.s32.s32 %r425, %r686, %r362, %r421;


dp4a.s32.s32 %r429, %r358, %r727, %r724;


dp4a.s32.s32 %r433, %r362, %r727, %r429;

cvt.rn.f32.s32 %f187, %r433;
fma.rn.ftz.f32 %f188, %f118, %f187, %f186;

dp4a.s32.s32 %r437, %r698, %r374, %r724;


dp4a.s32.s32 %r441, %r702, %r378, %r437;


dp4a.s32.s32 %r445, %r374, %r727, %r724;


dp4a.s32.s32 %r449, %r378, %r727, %r445;

cvt.rn.f32.s32 %f189, %r449;
fma.rn.ftz.f32 %f190, %f123, %f189, 0f00000000;

dp4a.s32.s32 %r453, %r714, %r390, %r441;


dp4a.s32.s32 %r457, %r718, %r394, %r453;


dp4a.s32.s32 %r461, %r390, %r727, %r724;


dp4a.s32.s32 %r465, %r394, %r727, %r461;

cvt.rn.f32.s32 %f191, %r465;
fma.rn.ftz.f32 %f192, %f128, %f191, %f190;

{ cvt.f32.f16 %f71, %rs10;}



{.reg .f16 low,high;
mov.b32 {low,high},%r404;
cvt.f32.f16 %f72, low;}


mul.ftz.f32 %f193, %f71, %f72;
cvt.rn.f32.s32 %f194, %r425;
add.ftz.f32 %f195, %f188, %f194;
cvt.rn.f32.s32 %f196, %r457;
add.ftz.f32 %f197, %f192, %f196;
mul.ftz.f32 %f198, %f197, %f137;
fma.rn.ftz.f32 %f199, %f195, %f134, %f198;
fma.rn.ftz.f32 %f384, %f193, %f199, %f384;
cvt.u64.u32 %rd72, %r997;
add.s64 %rd73, %rd2, %rd72;
add.s64 %rd74, %rd73, %rd4;
mul.lo.s64 %rd75, %rd74, 36;
add.s64 %rd76, %rd13, %rd75;
ld.global.nc.u32 %r472, [%rd76+4];
ld.global.nc.u32 %r476, [%rd76+8];

dp4a.s32.s32 %r470, %r601, %r472, %r724;


dp4a.s32.s32 %r474, %r605, %r476, %r470;


dp4a.s32.s32 %r478, %r472, %r727, %r724;


dp4a.s32.s32 %r482, %r476, %r727, %r478;

cvt.rn.f32.s32 %f200, %r482;
fma.rn.ftz.f32 %f201, %f84, %f200, 0f00000000;
ld.global.nc.u32 %r488, [%rd76+12];
ld.global.nc.u32 %r492, [%rd76+16];

dp4a.s32.s32 %r486, %r617, %r488, %r474;


dp4a.s32.s32 %r490, %r621, %r492, %r486;


dp4a.s32.s32 %r494, %r488, %r727, %r724;


dp4a.s32.s32 %r498, %r492, %r727, %r494;

cvt.rn.f32.s32 %f202, %r498;
fma.rn.ftz.f32 %f203, %f89, %f202, %f201;
ld.global.nc.u32 %r504, [%rd76+20];
ld.global.nc.u32 %r508, [%rd76+24];

dp4a.s32.s32 %r502, %r633, %r504, %r724;


dp4a.s32.s32 %r506, %r637, %r508, %r502;


dp4a.s32.s32 %r510, %r504, %r727, %r724;


dp4a.s32.s32 %r514, %r508, %r727, %r510;

cvt.rn.f32.s32 %f204, %r514;
fma.rn.ftz.f32 %f205, %f94, %f204, 0f00000000;
ld.global.nc.u32 %r520, [%rd76+28];
ld.global.nc.u32 %r524, [%rd76+32];

dp4a.s32.s32 %r518, %r649, %r520, %r506;


dp4a.s32.s32 %r522, %r653, %r524, %r518;


dp4a.s32.s32 %r526, %r520, %r727, %r724;


dp4a.s32.s32 %r530, %r524, %r727, %r526;

cvt.rn.f32.s32 %f206, %r530;
fma.rn.ftz.f32 %f207, %f99, %f206, %f205;

{ cvt.f32.f16 %f73, %rs9;}


ld.global.nc.u32 %r534, [%rd76];

{.reg .f16 low,high;
mov.b32 {low,high},%r534;
cvt.f32.f16 %f74, low;}


mul.ftz.f32 %f208, %f73, %f74;
cvt.rn.f32.s32 %f209, %r490;
add.ftz.f32 %f210, %f203, %f209;
cvt.rn.f32.s32 %f211, %r522;
add.ftz.f32 %f212, %f207, %f211;
mul.ftz.f32 %f213, %f212, %f108;
fma.rn.ftz.f32 %f214, %f210, %f105, %f213;
fma.rn.ftz.f32 %f383, %f208, %f214, %f383;

dp4a.s32.s32 %r535, %r666, %r472, %r724;


dp4a.s32.s32 %r539, %r670, %r476, %r535;


dp4a.s32.s32 %r543, %r472, %r727, %r724;


dp4a.s32.s32 %r547, %r476, %r727, %r543;

cvt.rn.f32.s32 %f215, %r547;
fma.rn.ftz.f32 %f216, %f113, %f215, 0f00000000;

dp4a.s32.s32 %r551, %r682, %r488, %r539;


dp4a.s32.s32 %r555, %r686, %r492, %r551;


dp4a.s32.s32 %r559, %r488, %r727, %r724;


dp4a.s32.s32 %r563, %r492, %r727, %r559;

cvt.rn.f32.s32 %f217, %r563;
fma.rn.ftz.f32 %f218, %f118, %f217, %f216;

dp4a.s32.s32 %r567, %r698, %r504, %r724;


dp4a.s32.s32 %r571, %r702, %r508, %r567;


dp4a.s32.s32 %r575, %r504, %r727, %r724;


dp4a.s32.s32 %r579, %r508, %r727, %r575;

cvt.rn.f32.s32 %f219, %r579;
fma.rn.ftz.f32 %f220, %f123, %f219, 0f00000000;

dp4a.s32.s32 %r583, %r714, %r520, %r571;


dp4a.s32.s32 %r587, %r718, %r524, %r583;


dp4a.s32.s32 %r591, %r520, %r727, %r724;


dp4a.s32.s32 %r595, %r524, %r727, %r591;

cvt.rn.f32.s32 %f221, %r595;
fma.rn.ftz.f32 %f222, %f128, %f221, %f220;

{ cvt.f32.f16 %f75, %rs10;}



{.reg .f16 low,high;
mov.b32 {low,high},%r534;
cvt.f32.f16 %f76, low;}


mul.ftz.f32 %f223, %f75, %f76;
cvt.rn.f32.s32 %f224, %r555;
add.ftz.f32 %f225, %f218, %f224;
cvt.rn.f32.s32 %f226, %r587;
add.ftz.f32 %f227, %f222, %f226;
mul.ftz.f32 %f228, %f227, %f137;
fma.rn.ftz.f32 %f229, %f225, %f134, %f228;
fma.rn.ftz.f32 %f382, %f223, %f229, %f382;
cvt.u64.u32 %rd77, %r996;
add.s64 %rd78, %rd2, %rd77;
add.s64 %rd79, %rd78, %rd4;
mul.lo.s64 %rd80, %rd79, 36;
add.s64 %rd81, %rd13, %rd80;
ld.global.nc.u32 %r602, [%rd81+4];
ld.global.nc.u32 %r606, [%rd81+8];

dp4a.s32.s32 %r600, %r601, %r602, %r724;


dp4a.s32.s32 %r604, %r605, %r606, %r600;


dp4a.s32.s32 %r608, %r602, %r727, %r724;


dp4a.s32.s32 %r612, %r606, %r727, %r608;

cvt.rn.f32.s32 %f230, %r612;
fma.rn.ftz.f32 %f231, %f84, %f230, 0f00000000;
ld.global.nc.u32 %r618, [%rd81+12];
ld.global.nc.u32 %r622, [%rd81+16];

dp4a.s32.s32 %r616, %r617, %r618, %r604;


dp4a.s32.s32 %r620, %r621, %r622, %r616;


dp4a.s32.s32 %r624, %r618, %r727, %r724;


dp4a.s32.s32 %r628, %r622, %r727, %r624;

cvt.rn.f32.s32 %f232, %r628;
fma.rn.ftz.f32 %f233, %f89, %f232, %f231;
ld.global.nc.u32 %r634, [%rd81+20];
ld.global.nc.u32 %r638, [%rd81+24];

dp4a.s32.s32 %r632, %r633, %r634, %r724;


dp4a.s32.s32 %r636, %r637, %r638, %r632;


dp4a.s32.s32 %r640, %r634, %r727, %r724;


dp4a.s32.s32 %r644, %r638, %r727, %r640;

cvt.rn.f32.s32 %f234, %r644;
fma.rn.ftz.f32 %f235, %f94, %f234, 0f00000000;
ld.global.nc.u32 %r650, [%rd81+28];
ld.global.nc.u32 %r654, [%rd81+32];

dp4a.s32.s32 %r648, %r649, %r650, %r636;


dp4a.s32.s32 %r652, %r653, %r654, %r648;


dp4a.s32.s32 %r656, %r650, %r727, %r724;


dp4a.s32.s32 %r660, %r654, %r727, %r656;

cvt.rn.f32.s32 %f236, %r660;
fma.rn.ftz.f32 %f237, %f99, %f236, %f235;

{ cvt.f32.f16 %f77, %rs9;}


ld.global.nc.u32 %r664, [%rd81];

{.reg .f16 low,high;
mov.b32 {low,high},%r664;
cvt.f32.f16 %f78, low;}


mul.ftz.f32 %f238, %f77, %f78;
cvt.rn.f32.s32 %f239, %r620;
add.ftz.f32 %f240, %f233, %f239;
cvt.rn.f32.s32 %f241, %r652;
add.ftz.f32 %f242, %f237, %f241;
mul.ftz.f32 %f243, %f242, %f108;
fma.rn.ftz.f32 %f244, %f240, %f105, %f243;
fma.rn.ftz.f32 %f381, %f238, %f244, %f381;

dp4a.s32.s32 %r665, %r666, %r602, %r724;


dp4a.s32.s32 %r669, %r670, %r606, %r665;


dp4a.s32.s32 %r673, %r602, %r727, %r724;


dp4a.s32.s32 %r677, %r606, %r727, %r673;

cvt.rn.f32.s32 %f245, %r677;
fma.rn.ftz.f32 %f246, %f113, %f245, 0f00000000;

dp4a.s32.s32 %r681, %r682, %r618, %r669;


dp4a.s32.s32 %r685, %r686, %r622, %r681;


dp4a.s32.s32 %r689, %r618, %r727, %r724;


dp4a.s32.s32 %r693, %r622, %r727, %r689;

cvt.rn.f32.s32 %f247, %r693;
fma.rn.ftz.f32 %f248, %f118, %f247, %f246;

dp4a.s32.s32 %r697, %r698, %r634, %r724;


dp4a.s32.s32 %r701, %r702, %r638, %r697;


dp4a.s32.s32 %r705, %r634, %r727, %r724;


dp4a.s32.s32 %r709, %r638, %r727, %r705;

cvt.rn.f32.s32 %f249, %r709;
fma.rn.ftz.f32 %f250, %f123, %f249, 0f00000000;

dp4a.s32.s32 %r713, %r714, %r650, %r701;


dp4a.s32.s32 %r717, %r718, %r654, %r713;


dp4a.s32.s32 %r721, %r650, %r727, %r724;


dp4a.s32.s32 %r725, %r654, %r727, %r721;

cvt.rn.f32.s32 %f251, %r725;
fma.rn.ftz.f32 %f252, %f128, %f251, %f250;

{ cvt.f32.f16 %f79, %rs10;}



{.reg .f16 low,high;
mov.b32 {low,high},%r664;
cvt.f32.f16 %f80, low;}


mul.ftz.f32 %f253, %f79, %f80;
cvt.rn.f32.s32 %f254, %r685;
add.ftz.f32 %f255, %f248, %f254;
cvt.rn.f32.s32 %f256, %r717;
add.ftz.f32 %f257, %f252, %f256;
mul.ftz.f32 %f258, %f257, %f137;
fma.rn.ftz.f32 %f259, %f255, %f134, %f258;
fma.rn.ftz.f32 %f380, %f253, %f259, %f380;
add.s32 %r998, %r998, 64;
add.s32 %r997, %r997, 64;
add.s32 %r996, %r996, 64;
add.s32 %r995, %r995, 64;
add.s32 %r994, %r994, 64;
add.s64 %rd105, %rd105, 448;
add.s32 %r999, %r999, 8;
setp.lt.s32 %p4, %r999, %r7;
@%p4 bra $L__BB272_2;

$L__BB272_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB272_5;

mov.u32 %r810, _ZZ13mul_mat_vec_qIL9ggml_type29ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r811, %r1, 1280, %r810;
shl.b32 %r812, %r2, 2;
add.s32 %r813, %r811, %r812;
st.shared.f32 [%r813+-1280], %f389;
st.shared.f32 [%r813+-1152], %f388;
st.shared.f32 [%r813+-1024], %f387;
st.shared.f32 [%r813+-896], %f386;
st.shared.f32 [%r813+-768], %f385;
st.shared.f32 [%r813+-640], %f384;
st.shared.f32 [%r813+-512], %f383;
st.shared.f32 [%r813+-384], %f382;
st.shared.f32 [%r813+-256], %f381;
st.shared.f32 [%r813+-128], %f380;

$L__BB272_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB272_26;

mov.u32 %r814, %ctaid.z;
mad.lo.s32 %r815, %r5, %r36, %r4;
mad.lo.s32 %r816, %r814, %r42, %r815;
cvt.u64.u32 %rd16, %r816;
shl.b32 %r817, %r2, 2;
mov.u32 %r818, _ZZ13mul_mat_vec_qIL9ggml_type29ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r26, %r818, %r817;
mov.u32 %r819, 2;
ld.shared.f32 %f260, [%r26];
add.ftz.f32 %f261, %f260, %f389;
mov.b32 %r820, %f261;
mov.u32 %r821, 31;
mov.u32 %r822, 16;
mov.u32 %r823, -1;
shfl.sync.bfly.b32 %r824|%p7, %r820, %r822, %r821, %r823;
mov.b32 %f262, %r824;
add.ftz.f32 %f263, %f261, %f262;
mov.b32 %r825, %f263;
mov.u32 %r826, 8;
shfl.sync.bfly.b32 %r827|%p8, %r825, %r826, %r821, %r823;
mov.b32 %f264, %r827;
add.ftz.f32 %f265, %f263, %f264;
mov.b32 %r828, %f265;
mov.u32 %r829, 4;
shfl.sync.bfly.b32 %r830|%p9, %r828, %r829, %r821, %r823;
mov.b32 %f266, %r830;
add.ftz.f32 %f267, %f265, %f266;
mov.b32 %r831, %f267;
shfl.sync.bfly.b32 %r832|%p10, %r831, %r819, %r821, %r823;
mov.b32 %f268, %r832;
add.ftz.f32 %f269, %f267, %f268;
mov.b32 %r833, %f269;
mov.u32 %r834, 1;
shfl.sync.bfly.b32 %r835|%p11, %r833, %r834, %r821, %r823;
mov.b32 %f270, %r835;
add.ftz.f32 %f31, %f269, %f270;
setp.lt.u32 %p12, %r4, %r30;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
mul.wide.u32 %rd82, %r816, 4;
add.s64 %rd17, %rd1, %rd82;
not.pred %p14, %p1;
@%p14 bra $L__BB272_8;

st.global.f32 [%rd17], %f31;

$L__BB272_8:
ld.shared.f32 %f271, [%r26+128];
add.ftz.f32 %f272, %f271, %f388;
mov.b32 %r836, %f272;
shfl.sync.bfly.b32 %r840|%p15, %r836, %r822, %r821, %r823;
mov.b32 %f273, %r840;
add.ftz.f32 %f274, %f272, %f273;
mov.b32 %r841, %f274;
shfl.sync.bfly.b32 %r843|%p16, %r841, %r826, %r821, %r823;
mov.b32 %f275, %r843;
add.ftz.f32 %f276, %f274, %f275;
mov.b32 %r844, %f276;
shfl.sync.bfly.b32 %r846|%p17, %r844, %r829, %r821, %r823;
mov.b32 %f277, %r846;
add.ftz.f32 %f278, %f276, %f277;
mov.b32 %r847, %f278;
shfl.sync.bfly.b32 %r849|%p18, %r847, %r819, %r821, %r823;
mov.b32 %f279, %r849;
add.ftz.f32 %f280, %f278, %f279;
mov.b32 %r850, %f280;
shfl.sync.bfly.b32 %r852|%p19, %r850, %r834, %r821, %r823;
mov.b32 %f281, %r852;
add.ftz.f32 %f32, %f280, %f281;
add.s32 %r853, %r4, 1;
setp.lt.u32 %p20, %r853, %r30;
setp.eq.s32 %p21, %r2, 1;
and.pred %p2, %p21, %p20;
not.pred %p22, %p2;
@%p22 bra $L__BB272_10;

st.global.f32 [%rd17+4], %f32;

$L__BB272_10:
ld.shared.f32 %f282, [%r26+256];
add.ftz.f32 %f283, %f282, %f387;
mov.b32 %r854, %f283;
mov.u32 %r855, 31;
mov.u32 %r856, 16;
mov.u32 %r857, -1;
shfl.sync.bfly.b32 %r858|%p23, %r854, %r856, %r855, %r857;
mov.b32 %f284, %r858;
add.ftz.f32 %f285, %f283, %f284;
mov.b32 %r859, %f285;
mov.u32 %r860, 8;
shfl.sync.bfly.b32 %r861|%p24, %r859, %r860, %r855, %r857;
mov.b32 %f286, %r861;
add.ftz.f32 %f287, %f285, %f286;
mov.b32 %r862, %f287;
mov.u32 %r863, 4;
shfl.sync.bfly.b32 %r864|%p25, %r862, %r863, %r855, %r857;
mov.b32 %f288, %r864;
add.ftz.f32 %f289, %f287, %f288;
mov.b32 %r865, %f289;
mov.u32 %r866, 2;
shfl.sync.bfly.b32 %r867|%p26, %r865, %r866, %r855, %r857;
mov.b32 %f290, %r867;
add.ftz.f32 %f291, %f289, %f290;
mov.b32 %r868, %f291;
mov.u32 %r869, 1;
shfl.sync.bfly.b32 %r870|%p27, %r868, %r869, %r855, %r857;
mov.b32 %f292, %r870;
add.ftz.f32 %f33, %f291, %f292;
@%p14 bra $L__BB272_12;

cvt.u64.u32 %rd83, %r30;
add.s64 %rd84, %rd83, %rd16;
shl.b64 %rd85, %rd84, 2;
add.s64 %rd86, %rd1, %rd85;
st.global.f32 [%rd86], %f33;

$L__BB272_12:
ld.shared.f32 %f293, [%r26+384];
add.ftz.f32 %f294, %f293, %f386;
mov.b32 %r871, %f294;
shfl.sync.bfly.b32 %r875|%p29, %r871, %r856, %r855, %r857;
mov.b32 %f295, %r875;
add.ftz.f32 %f296, %f294, %f295;
mov.b32 %r876, %f296;
shfl.sync.bfly.b32 %r878|%p30, %r876, %r860, %r855, %r857;
mov.b32 %f297, %r878;
add.ftz.f32 %f298, %f296, %f297;
mov.b32 %r879, %f298;
shfl.sync.bfly.b32 %r881|%p31, %r879, %r863, %r855, %r857;
mov.b32 %f299, %r881;
add.ftz.f32 %f300, %f298, %f299;
mov.b32 %r882, %f300;
shfl.sync.bfly.b32 %r884|%p32, %r882, %r866, %r855, %r857;
mov.b32 %f301, %r884;
add.ftz.f32 %f302, %f300, %f301;
mov.b32 %r885, %f302;
shfl.sync.bfly.b32 %r887|%p33, %r885, %r869, %r855, %r857;
mov.b32 %f303, %r887;
add.ftz.f32 %f34, %f302, %f303;
@%p22 bra $L__BB272_14;

add.s32 %r888, %r30, 1;
cvt.u64.u32 %rd87, %r888;
add.s64 %rd88, %rd87, %rd16;
shl.b64 %rd89, %rd88, 2;
add.s64 %rd90, %rd1, %rd89;
st.global.f32 [%rd90], %f34;

$L__BB272_14:
shl.b32 %r889, %r30, 1;
mov.u32 %r890, 1;
ld.shared.f32 %f304, [%r26+512];
add.ftz.f32 %f305, %f304, %f385;
mov.b32 %r891, %f305;
mov.u32 %r892, 31;
mov.u32 %r893, 16;
mov.u32 %r894, -1;
shfl.sync.bfly.b32 %r895|%p35, %r891, %r893, %r892, %r894;
mov.b32 %f306, %r895;
add.ftz.f32 %f307, %f305, %f306;
mov.b32 %r896, %f307;
mov.u32 %r897, 8;
shfl.sync.bfly.b32 %r898|%p36, %r896, %r897, %r892, %r894;
mov.b32 %f308, %r898;
add.ftz.f32 %f309, %f307, %f308;
mov.b32 %r899, %f309;
mov.u32 %r900, 4;
shfl.sync.bfly.b32 %r901|%p37, %r899, %r900, %r892, %r894;
mov.b32 %f310, %r901;
add.ftz.f32 %f311, %f309, %f310;
mov.b32 %r902, %f311;
mov.u32 %r903, 2;
shfl.sync.bfly.b32 %r904|%p38, %r902, %r903, %r892, %r894;
mov.b32 %f312, %r904;
add.ftz.f32 %f313, %f311, %f312;
mov.b32 %r905, %f313;
shfl.sync.bfly.b32 %r906|%p39, %r905, %r890, %r892, %r894;
mov.b32 %f314, %r906;
add.ftz.f32 %f35, %f313, %f314;
cvt.u64.u32 %rd91, %r889;
add.s64 %rd92, %rd91, %rd16;
shl.b64 %rd93, %rd92, 2;
add.s64 %rd18, %rd1, %rd93;
@%p14 bra $L__BB272_16;

st.global.f32 [%rd18], %f35;

$L__BB272_16:
ld.shared.f32 %f315, [%r26+640];
add.ftz.f32 %f316, %f315, %f384;
mov.b32 %r907, %f316;
shfl.sync.bfly.b32 %r911|%p41, %r907, %r893, %r892, %r894;
mov.b32 %f317, %r911;
add.ftz.f32 %f318, %f316, %f317;
mov.b32 %r912, %f318;
shfl.sync.bfly.b32 %r914|%p42, %r912, %r897, %r892, %r894;
mov.b32 %f319, %r914;
add.ftz.f32 %f320, %f318, %f319;
mov.b32 %r915, %f320;
shfl.sync.bfly.b32 %r917|%p43, %r915, %r900, %r892, %r894;
mov.b32 %f321, %r917;
add.ftz.f32 %f322, %f320, %f321;
mov.b32 %r918, %f322;
shfl.sync.bfly.b32 %r920|%p44, %r918, %r903, %r892, %r894;
mov.b32 %f323, %r920;
add.ftz.f32 %f324, %f322, %f323;
mov.b32 %r921, %f324;
shfl.sync.bfly.b32 %r923|%p45, %r921, %r890, %r892, %r894;
mov.b32 %f325, %r923;
add.ftz.f32 %f36, %f324, %f325;
@%p22 bra $L__BB272_18;

st.global.f32 [%rd18+4], %f36;

$L__BB272_18:
mul.lo.s32 %r27, %r30, 3;
ld.shared.f32 %f326, [%r26+768];
add.ftz.f32 %f327, %f326, %f383;
mov.b32 %r924, %f327;
mov.u32 %r925, 31;
mov.u32 %r926, 16;
mov.u32 %r927, -1;
shfl.sync.bfly.b32 %r928|%p47, %r924, %r926, %r925, %r927;
mov.b32 %f328, %r928;
add.ftz.f32 %f329, %f327, %f328;
mov.b32 %r929, %f329;
mov.u32 %r930, 8;
shfl.sync.bfly.b32 %r931|%p48, %r929, %r930, %r925, %r927;
mov.b32 %f330, %r931;
add.ftz.f32 %f331, %f329, %f330;
mov.b32 %r932, %f331;
mov.u32 %r933, 4;
shfl.sync.bfly.b32 %r934|%p49, %r932, %r933, %r925, %r927;
mov.b32 %f332, %r934;
add.ftz.f32 %f333, %f331, %f332;
mov.b32 %r935, %f333;
mov.u32 %r936, 2;
shfl.sync.bfly.b32 %r937|%p50, %r935, %r936, %r925, %r927;
mov.b32 %f334, %r937;
add.ftz.f32 %f335, %f333, %f334;
mov.b32 %r938, %f335;
mov.u32 %r939, 1;
shfl.sync.bfly.b32 %r940|%p51, %r938, %r939, %r925, %r927;
mov.b32 %f336, %r940;
add.ftz.f32 %f37, %f335, %f336;
@%p14 bra $L__BB272_20;

cvt.u64.u32 %rd94, %r27;
add.s64 %rd95, %rd94, %rd16;
shl.b64 %rd96, %rd95, 2;
add.s64 %rd97, %rd1, %rd96;
st.global.f32 [%rd97], %f37;

$L__BB272_20:
ld.shared.f32 %f337, [%r26+896];
add.ftz.f32 %f338, %f337, %f382;
mov.b32 %r941, %f338;
shfl.sync.bfly.b32 %r945|%p53, %r941, %r926, %r925, %r927;
mov.b32 %f339, %r945;
add.ftz.f32 %f340, %f338, %f339;
mov.b32 %r946, %f340;
shfl.sync.bfly.b32 %r948|%p54, %r946, %r930, %r925, %r927;
mov.b32 %f341, %r948;
add.ftz.f32 %f342, %f340, %f341;
mov.b32 %r949, %f342;
shfl.sync.bfly.b32 %r951|%p55, %r949, %r933, %r925, %r927;
mov.b32 %f343, %r951;
add.ftz.f32 %f344, %f342, %f343;
mov.b32 %r952, %f344;
shfl.sync.bfly.b32 %r954|%p56, %r952, %r936, %r925, %r927;
mov.b32 %f345, %r954;
add.ftz.f32 %f346, %f344, %f345;
mov.b32 %r955, %f346;
shfl.sync.bfly.b32 %r957|%p57, %r955, %r939, %r925, %r927;
mov.b32 %f347, %r957;
add.ftz.f32 %f38, %f346, %f347;
@%p22 bra $L__BB272_22;

add.s32 %r958, %r27, 1;
cvt.u64.u32 %rd98, %r958;
add.s64 %rd99, %rd98, %rd16;
shl.b64 %rd100, %rd99, 2;
add.s64 %rd101, %rd1, %rd100;
st.global.f32 [%rd101], %f38;

$L__BB272_22:
shl.b32 %r959, %r30, 2;
mov.u32 %r960, 2;
ld.shared.f32 %f348, [%r26+1024];
add.ftz.f32 %f349, %f348, %f381;
mov.b32 %r961, %f349;
mov.u32 %r962, 31;
mov.u32 %r963, 16;
mov.u32 %r964, -1;
shfl.sync.bfly.b32 %r965|%p59, %r961, %r963, %r962, %r964;
mov.b32 %f350, %r965;
add.ftz.f32 %f351, %f349, %f350;
mov.b32 %r966, %f351;
mov.u32 %r967, 8;
shfl.sync.bfly.b32 %r968|%p60, %r966, %r967, %r962, %r964;
mov.b32 %f352, %r968;
add.ftz.f32 %f353, %f351, %f352;
mov.b32 %r969, %f353;
mov.u32 %r970, 4;
shfl.sync.bfly.b32 %r971|%p61, %r969, %r970, %r962, %r964;
mov.b32 %f354, %r971;
add.ftz.f32 %f355, %f353, %f354;
mov.b32 %r972, %f355;
shfl.sync.bfly.b32 %r973|%p62, %r972, %r960, %r962, %r964;
mov.b32 %f356, %r973;
add.ftz.f32 %f357, %f355, %f356;
mov.b32 %r974, %f357;
mov.u32 %r975, 1;
shfl.sync.bfly.b32 %r976|%p63, %r974, %r975, %r962, %r964;
mov.b32 %f358, %r976;
add.ftz.f32 %f39, %f357, %f358;
cvt.u64.u32 %rd102, %r959;
add.s64 %rd103, %rd102, %rd16;
shl.b64 %rd104, %rd103, 2;
add.s64 %rd19, %rd1, %rd104;
@%p14 bra $L__BB272_24;

st.global.f32 [%rd19], %f39;

$L__BB272_24:
ld.shared.f32 %f359, [%r26+1152];
add.ftz.f32 %f360, %f359, %f380;
mov.b32 %r977, %f360;
shfl.sync.bfly.b32 %r981|%p65, %r977, %r963, %r962, %r964;
mov.b32 %f361, %r981;
add.ftz.f32 %f362, %f360, %f361;
mov.b32 %r982, %f362;
shfl.sync.bfly.b32 %r984|%p66, %r982, %r967, %r962, %r964;
mov.b32 %f363, %r984;
add.ftz.f32 %f364, %f362, %f363;
mov.b32 %r985, %f364;
shfl.sync.bfly.b32 %r987|%p67, %r985, %r970, %r962, %r964;
mov.b32 %f365, %r987;
add.ftz.f32 %f366, %f364, %f365;
mov.b32 %r988, %f366;
shfl.sync.bfly.b32 %r990|%p68, %r988, %r960, %r962, %r964;
mov.b32 %f367, %r990;
add.ftz.f32 %f368, %f366, %f367;
mov.b32 %r991, %f368;
shfl.sync.bfly.b32 %r993|%p69, %r991, %r975, %r962, %r964;
mov.b32 %f369, %r993;
add.ftz.f32 %f40, %f368, %f369;
@%p22 bra $L__BB272_26;

st.global.f32 [%rd19+4], %f40;

$L__BB272_26:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type29ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type29ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type29ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type29ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type29ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type29ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type29ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type29ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type29ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<83>;
.reg .b16 %rs<51>;
.reg .f32 %f<462>;
.reg .b32 %r<1172>;
.reg .b64 %rd<119>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type29ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

ld.param.u64 %rd20, [_Z13mul_mat_vec_qIL9ggml_type29ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd21, [_Z13mul_mat_vec_qIL9ggml_type29ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd22, [_Z13mul_mat_vec_qIL9ggml_type29ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r47, [_Z13mul_mat_vec_qIL9ggml_type29ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type29ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type29ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type29ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type29ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type29ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type29ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type29ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r45, [_Z13mul_mat_vec_qIL9ggml_type29ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r46, [_Z13mul_mat_vec_qIL9ggml_type29ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type29ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type29ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type29ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type29ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
cvta.to.global.u64 %rd1, %rd22;
mov.u32 %r1, %tid.y;
shl.b32 %r48, %r1, 5;
mov.u32 %r49, %tid.x;
add.s32 %r2, %r48, %r49;
mov.u32 %r50, %ctaid.x;
shl.b32 %r3, %r50, 1;
mov.u32 %r4, %ctaid.y;
mov.u32 %r5, %ctaid.z;
shr.s32 %r51, %r2, 31;
shr.u32 %r52, %r51, 29;
add.s32 %r53, %r2, %r52;
shr.s32 %r1171, %r53, 3;
shr.u32 %r7, %r47, 8;
setp.ge.s32 %p3, %r1171, %r7;
mov.f32 %f450, 0f00000000;
mov.f32 %f451, %f450;
mov.f32 %f452, %f450;
mov.f32 %f453, %f450;
mov.f32 %f454, %f450;
mov.f32 %f455, %f450;
mov.f32 %f456, %f450;
mov.f32 %f457, %f450;
mov.f32 %f458, %f450;
mov.f32 %f459, %f450;
mov.f32 %f460, %f450;
mov.f32 %f461, %f450;
@%p3 bra $L__BB273_3;

mul.lo.s32 %r54, %r4, %r39;
cvt.u64.u32 %rd23, %r54;
mul.lo.s32 %r55, %r5, %r45;
cvt.u64.u32 %rd24, %r55;
add.s64 %rd2, %rd24, %rd23;
mul.hi.u32 %r56, %r5, %r41;
add.s32 %r57, %r56, %r5;
shr.u32 %r58, %r57, %r42;
mul.lo.s32 %r59, %r58, %r44;
mul.hi.u32 %r60, %r4, %r35;
add.s32 %r61, %r60, %r4;
shr.u32 %r62, %r61, %r36;
mul.lo.s32 %r63, %r62, %r38;
cvta.to.global.u64 %rd118, %rd20;
add.s32 %r64, %r3, 1;
and.b32 %r68, %r53, -8;
sub.s32 %r69, %r2, %r68;
shl.b32 %r70, %r69, 1;
cvt.s64.s32 %rd25, %r70;
cvt.s64.s32 %rd4, %r69;
shr.u32 %r71, %r69, 31;
add.s32 %r72, %r69, %r71;
shr.s32 %r73, %r72, 1;
and.b32 %r74, %r72, -2;
sub.s32 %r75, %r69, %r74;
mul.lo.s32 %r8, %r75, 6;
shl.b32 %r76, %r33, 1;
shl.b32 %r1165, %r1171, 3;
add.s32 %r1170, %r1165, %r76;
mad.lo.s32 %r1169, %r33, 3, %r1165;
shl.b32 %r77, %r33, 2;
add.s32 %r1168, %r1165, %r77;
mad.lo.s32 %r1167, %r33, 5, %r1165;
add.s32 %r1166, %r33, %r1165;
add.s32 %r78, %r1171, %r59;
add.s32 %r79, %r78, %r63;
mad.lo.s32 %r80, %r64, %r32, %r79;
mul.wide.s32 %rd5, %r80, 56;
mad.lo.s32 %r81, %r3, %r32, %r1171;
add.s32 %r82, %r81, %r59;
add.s32 %r83, %r82, %r63;
mul.wide.s32 %rd6, %r83, 56;
mul.wide.s32 %rd26, %r69, 4;
add.s64 %rd7, %rd6, %rd26;
mul.wide.s32 %rd27, %r73, 2;
add.s64 %rd28, %rd6, %rd27;
add.s64 %rd8, %rd28, 48;
add.s64 %rd29, %rd5, %rd27;
add.s64 %rd9, %rd29, 48;
add.s64 %rd30, %rd6, %rd25;
add.s64 %rd10, %rd30, 33;
add.s64 %rd31, %rd5, %rd25;
add.s64 %rd11, %rd31, 32;
add.s64 %rd12, %rd5, %rd26;
cvta.to.global.u64 %rd13, %rd21;
mov.f32 %f450, 0f00000000;

$L__BB273_2:
cvt.u64.u32 %rd32, %r1165;
add.s64 %rd33, %rd2, %rd32;
add.s64 %rd34, %rd33, %rd4;
mul.lo.s64 %rd35, %rd34, 36;
add.s64 %rd36, %rd118, %rd7;
ld.global.nc.u32 %r864, [%rd36];
shr.u32 %r865, %r864, 8;
shr.u32 %r866, %r864, 16;
shr.u32 %r867, %r864, 24;
add.s64 %rd37, %rd118, %rd10;
ld.global.nc.u8 %rs13, [%rd37+-1];
and.b16 %rs14, %rs13, 240;
cvt.u32.u16 %r868, %rs13;
and.b32 %r869, %r864, 255;
and.b32 %r870, %r868, 7;
bfi.b32 %r871, %r870, %r869, 8, 3;
mul.wide.u32 %rd38, %r871, 4;
mov.u64 %rd39, iq1s_grid_gpu;
add.s64 %rd40, %rd39, %rd38;
ld.global.nc.u32 %r872, [%rd40];
and.b32 %r735, %r872, 252645135;
shr.u32 %r873, %r872, 4;
and.b32 %r739, %r873, 252645135;
add.s64 %rd41, %rd13, %rd35;
ld.global.nc.u32 %r86, [%rd41+4];
ld.global.nc.u32 %r90, [%rd41+8];
mov.u32 %r858, 0;

dp4a.s32.s32 %r84, %r735, %r86, %r858;


dp4a.s32.s32 %r88, %r739, %r90, %r84;

and.b32 %r874, %r868, 8;
cvt.rn.f32.s32 %f97, %r874;
mul.ftz.f32 %f98, %f97, 0f3D000000;
mov.f32 %f99, 0fBF600000;
sub.ftz.f32 %f100, %f99, %f98;
mov.u32 %r861, 16843009;

dp4a.s32.s32 %r92, %r86, %r861, %r858;


dp4a.s32.s32 %r96, %r90, %r861, %r92;

cvt.rn.f32.s32 %f101, %r96;
fma.rn.ftz.f32 %f102, %f100, %f101, 0f00000000;
shr.u16 %rs15, %rs14, 4;
cvt.u32.u16 %r875, %rs15;
and.b32 %r876, %r865, 255;
and.b32 %r877, %r875, 7;
bfi.b32 %r878, %r877, %r876, 8, 3;
mul.wide.u32 %rd42, %r878, 4;
add.s64 %rd43, %rd39, %rd42;
ld.global.nc.u32 %r879, [%rd43];
and.b32 %r751, %r879, 252645135;
shr.u32 %r880, %r879, 4;
and.b32 %r755, %r880, 252645135;
ld.global.nc.u32 %r102, [%rd41+12];
ld.global.nc.u32 %r106, [%rd41+16];

dp4a.s32.s32 %r100, %r751, %r102, %r88;


dp4a.s32.s32 %r104, %r755, %r106, %r100;

and.b32 %r881, %r875, 8;
cvt.rn.f32.s32 %f103, %r881;
mul.ftz.f32 %f104, %f103, 0f3D000000;
sub.ftz.f32 %f105, %f99, %f104;

dp4a.s32.s32 %r108, %r102, %r861, %r858;


dp4a.s32.s32 %r112, %r106, %r861, %r108;

cvt.rn.f32.s32 %f106, %r112;
fma.rn.ftz.f32 %f107, %f105, %f106, %f102;
ld.global.nc.u8 %rs16, [%rd37];
and.b16 %rs17, %rs16, 240;
cvt.u32.u16 %r882, %rs16;
and.b32 %r883, %r866, 255;
and.b32 %r884, %r882, 7;
bfi.b32 %r885, %r884, %r883, 8, 3;
mul.wide.u32 %rd44, %r885, 4;
add.s64 %rd45, %rd39, %rd44;
ld.global.nc.u32 %r886, [%rd45];
and.b32 %r767, %r886, 252645135;
shr.u32 %r887, %r886, 4;
and.b32 %r771, %r887, 252645135;
ld.global.nc.u32 %r118, [%rd41+20];
ld.global.nc.u32 %r122, [%rd41+24];

dp4a.s32.s32 %r116, %r767, %r118, %r858;


dp4a.s32.s32 %r120, %r771, %r122, %r116;

and.b32 %r888, %r882, 8;
cvt.rn.f32.s32 %f108, %r888;
mul.ftz.f32 %f109, %f108, 0f3D000000;
sub.ftz.f32 %f110, %f99, %f109;

dp4a.s32.s32 %r124, %r118, %r861, %r858;


dp4a.s32.s32 %r128, %r122, %r861, %r124;

cvt.rn.f32.s32 %f111, %r128;
fma.rn.ftz.f32 %f112, %f110, %f111, 0f00000000;
shr.u16 %rs18, %rs17, 4;
cvt.u32.u16 %r889, %rs18;
mul.wide.u16 %r890, %rs18, 256;
and.b32 %r891, %r890, 1792;
or.b32 %r892, %r891, %r867;
mul.wide.u32 %rd46, %r892, 4;
add.s64 %rd47, %rd39, %rd46;
ld.global.nc.u32 %r893, [%rd47];
and.b32 %r783, %r893, 252645135;
shr.u32 %r894, %r893, 4;
and.b32 %r787, %r894, 252645135;
ld.global.nc.u32 %r134, [%rd41+28];
ld.global.nc.u32 %r138, [%rd41+32];

dp4a.s32.s32 %r132, %r783, %r134, %r120;


dp4a.s32.s32 %r136, %r787, %r138, %r132;

and.b32 %r895, %r889, 8;
cvt.rn.f32.s32 %f113, %r895;
mul.ftz.f32 %f114, %f113, 0f3D000000;
sub.ftz.f32 %f115, %f99, %f114;

dp4a.s32.s32 %r140, %r134, %r861, %r858;


dp4a.s32.s32 %r144, %r138, %r861, %r140;

cvt.rn.f32.s32 %f116, %r144;
fma.rn.ftz.f32 %f117, %f115, %f116, %f112;
add.s64 %rd48, %rd118, %rd6;
ld.global.nc.u16 %rs19, [%rd48+48];
shr.u16 %rs20, %rs19, 12;
ld.global.nc.u16 %rs21, [%rd48+50];
shr.u16 %rs22, %rs21, 8;
and.b16 %rs23, %rs22, 240;
or.b16 %rs24, %rs23, %rs20;
ld.global.nc.u16 %rs25, [%rd48+52];
shr.u16 %rs26, %rs25, 4;
and.b16 %rs27, %rs26, 3840;
or.b16 %rs28, %rs24, %rs27;
ld.global.nc.u16 %rs29, [%rd48+54];
and.b16 %rs30, %rs29, -4096;
or.b16 %rs11, %rs28, %rs30;

{ cvt.f32.f16 %f73, %rs11;}


ld.global.nc.u32 %r148, [%rd41];

{.reg .f16 low,high;
mov.b32 {low,high},%r148;
cvt.f32.f16 %f74, low;}


mul.ftz.f32 %f118, %f73, %f74;
add.s64 %rd49, %rd118, %rd8;
ld.global.nc.u16 %rs31, [%rd49];
cvt.u32.u16 %r896, %rs31;
shr.u32 %r897, %r896, %r8;
shl.b32 %r898, %r897, 1;
and.b32 %r899, %r898, 14;
or.b32 %r900, %r899, 1;
shr.u32 %r901, %r897, 2;
and.b32 %r902, %r901, 14;
or.b32 %r903, %r902, 1;
cvt.rn.f32.s32 %f119, %r104;
add.ftz.f32 %f120, %f107, %f119;
cvt.rn.f32.s32 %f121, %r900;
cvt.rn.f32.s32 %f122, %r136;
add.ftz.f32 %f123, %f117, %f122;
cvt.rn.f32.s32 %f124, %r903;
mul.ftz.f32 %f125, %f123, %f124;
fma.rn.ftz.f32 %f126, %f120, %f121, %f125;
fma.rn.ftz.f32 %f461, %f118, %f126, %f461;
add.s64 %rd50, %rd118, %rd12;
ld.global.nc.u32 %r904, [%rd50];
shr.u32 %r905, %r904, 8;
shr.u32 %r906, %r904, 16;
shr.u32 %r907, %r904, 24;
add.s64 %rd51, %rd118, %rd11;
ld.global.nc.u8 %rs32, [%rd51];
and.b16 %rs33, %rs32, 240;
cvt.u32.u16 %r908, %rs32;
and.b32 %r909, %r904, 255;
and.b32 %r910, %r908, 7;
bfi.b32 %r911, %r910, %r909, 8, 3;
mul.wide.u32 %rd52, %r911, 4;
add.s64 %rd53, %rd39, %rd52;
ld.global.nc.u32 %r912, [%rd53];
and.b32 %r800, %r912, 252645135;
shr.u32 %r913, %r912, 4;
and.b32 %r804, %r913, 252645135;

dp4a.s32.s32 %r149, %r800, %r86, %r858;


dp4a.s32.s32 %r153, %r804, %r90, %r149;

and.b32 %r914, %r908, 8;
cvt.rn.f32.s32 %f127, %r914;
mul.ftz.f32 %f128, %f127, 0f3D000000;
sub.ftz.f32 %f129, %f99, %f128;

dp4a.s32.s32 %r157, %r86, %r861, %r858;


dp4a.s32.s32 %r161, %r90, %r861, %r157;

cvt.rn.f32.s32 %f130, %r161;
fma.rn.ftz.f32 %f131, %f129, %f130, 0f00000000;
shr.u16 %rs34, %rs33, 4;
cvt.u32.u16 %r915, %rs34;
and.b32 %r916, %r905, 255;
and.b32 %r917, %r915, 7;
bfi.b32 %r918, %r917, %r916, 8, 3;
mul.wide.u32 %rd54, %r918, 4;
add.s64 %rd55, %rd39, %rd54;
ld.global.nc.u32 %r919, [%rd55];
and.b32 %r816, %r919, 252645135;
shr.u32 %r920, %r919, 4;
and.b32 %r820, %r920, 252645135;

dp4a.s32.s32 %r165, %r816, %r102, %r153;


dp4a.s32.s32 %r169, %r820, %r106, %r165;

and.b32 %r921, %r915, 8;
cvt.rn.f32.s32 %f132, %r921;
mul.ftz.f32 %f133, %f132, 0f3D000000;
sub.ftz.f32 %f134, %f99, %f133;

dp4a.s32.s32 %r173, %r102, %r861, %r858;


dp4a.s32.s32 %r177, %r106, %r861, %r173;

cvt.rn.f32.s32 %f135, %r177;
fma.rn.ftz.f32 %f136, %f134, %f135, %f131;
ld.global.nc.u8 %rs35, [%rd51+1];
and.b16 %rs36, %rs35, 240;
cvt.u32.u16 %r922, %rs35;
and.b32 %r923, %r906, 255;
and.b32 %r924, %r922, 7;
bfi.b32 %r925, %r924, %r923, 8, 3;
mul.wide.u32 %rd56, %r925, 4;
add.s64 %rd57, %rd39, %rd56;
ld.global.nc.u32 %r926, [%rd57];
and.b32 %r832, %r926, 252645135;
shr.u32 %r927, %r926, 4;
and.b32 %r836, %r927, 252645135;

dp4a.s32.s32 %r181, %r832, %r118, %r858;


dp4a.s32.s32 %r185, %r836, %r122, %r181;

and.b32 %r928, %r922, 8;
cvt.rn.f32.s32 %f137, %r928;
mul.ftz.f32 %f138, %f137, 0f3D000000;
sub.ftz.f32 %f139, %f99, %f138;

dp4a.s32.s32 %r189, %r118, %r861, %r858;


dp4a.s32.s32 %r193, %r122, %r861, %r189;

cvt.rn.f32.s32 %f140, %r193;
fma.rn.ftz.f32 %f141, %f139, %f140, 0f00000000;
shr.u16 %rs37, %rs36, 4;
cvt.u32.u16 %r929, %rs37;
mul.wide.u16 %r930, %rs37, 256;
and.b32 %r931, %r930, 1792;
or.b32 %r932, %r931, %r907;
mul.wide.u32 %rd58, %r932, 4;
add.s64 %rd59, %rd39, %rd58;
ld.global.nc.u32 %r933, [%rd59];
and.b32 %r848, %r933, 252645135;
shr.u32 %r934, %r933, 4;
and.b32 %r852, %r934, 252645135;

dp4a.s32.s32 %r197, %r848, %r134, %r185;


dp4a.s32.s32 %r201, %r852, %r138, %r197;

and.b32 %r935, %r929, 8;
cvt.rn.f32.s32 %f142, %r935;
mul.ftz.f32 %f143, %f142, 0f3D000000;
sub.ftz.f32 %f144, %f99, %f143;

dp4a.s32.s32 %r205, %r134, %r861, %r858;


dp4a.s32.s32 %r209, %r138, %r861, %r205;

cvt.rn.f32.s32 %f145, %r209;
fma.rn.ftz.f32 %f146, %f144, %f145, %f141;
add.s64 %rd60, %rd118, %rd5;
ld.global.nc.u16 %rs38, [%rd60+48];
shr.u16 %rs39, %rs38, 12;
ld.global.nc.u16 %rs40, [%rd60+50];
shr.u16 %rs41, %rs40, 8;
and.b16 %rs42, %rs41, 240;
or.b16 %rs43, %rs42, %rs39;
ld.global.nc.u16 %rs44, [%rd60+52];
shr.u16 %rs45, %rs44, 4;
and.b16 %rs46, %rs45, 3840;
or.b16 %rs47, %rs43, %rs46;
ld.global.nc.u16 %rs48, [%rd60+54];
and.b16 %rs49, %rs48, -4096;
or.b16 %rs12, %rs47, %rs49;

{ cvt.f32.f16 %f75, %rs12;}



{.reg .f16 low,high;
mov.b32 {low,high},%r148;
cvt.f32.f16 %f76, low;}


mul.ftz.f32 %f147, %f75, %f76;
add.s64 %rd61, %rd118, %rd9;
ld.global.nc.u16 %rs50, [%rd61];
cvt.u32.u16 %r936, %rs50;
shr.u32 %r937, %r936, %r8;
shl.b32 %r938, %r937, 1;
and.b32 %r939, %r938, 14;
or.b32 %r940, %r939, 1;
shr.u32 %r941, %r937, 2;
and.b32 %r942, %r941, 14;
or.b32 %r943, %r942, 1;
cvt.rn.f32.s32 %f148, %r169;
add.ftz.f32 %f149, %f136, %f148;
cvt.rn.f32.s32 %f150, %r940;
cvt.rn.f32.s32 %f151, %r201;
add.ftz.f32 %f152, %f146, %f151;
cvt.rn.f32.s32 %f153, %r943;
mul.ftz.f32 %f154, %f152, %f153;
fma.rn.ftz.f32 %f155, %f149, %f150, %f154;
fma.rn.ftz.f32 %f460, %f147, %f155, %f460;
cvt.u64.u32 %rd62, %r1166;
add.s64 %rd63, %rd2, %rd62;
add.s64 %rd64, %rd63, %rd4;
mul.lo.s64 %rd65, %rd64, 36;
add.s64 %rd66, %rd13, %rd65;
ld.global.nc.u32 %r216, [%rd66+4];
ld.global.nc.u32 %r220, [%rd66+8];

dp4a.s32.s32 %r214, %r735, %r216, %r858;


dp4a.s32.s32 %r218, %r739, %r220, %r214;


dp4a.s32.s32 %r222, %r216, %r861, %r858;


dp4a.s32.s32 %r226, %r220, %r861, %r222;

cvt.rn.f32.s32 %f156, %r226;
fma.rn.ftz.f32 %f157, %f100, %f156, 0f00000000;
ld.global.nc.u32 %r232, [%rd66+12];
ld.global.nc.u32 %r236, [%rd66+16];

dp4a.s32.s32 %r230, %r751, %r232, %r218;


dp4a.s32.s32 %r234, %r755, %r236, %r230;


dp4a.s32.s32 %r238, %r232, %r861, %r858;


dp4a.s32.s32 %r242, %r236, %r861, %r238;

cvt.rn.f32.s32 %f158, %r242;
fma.rn.ftz.f32 %f159, %f105, %f158, %f157;
ld.global.nc.u32 %r248, [%rd66+20];
ld.global.nc.u32 %r252, [%rd66+24];

dp4a.s32.s32 %r246, %r767, %r248, %r858;


dp4a.s32.s32 %r250, %r771, %r252, %r246;


dp4a.s32.s32 %r254, %r248, %r861, %r858;


dp4a.s32.s32 %r258, %r252, %r861, %r254;

cvt.rn.f32.s32 %f160, %r258;
fma.rn.ftz.f32 %f161, %f110, %f160, 0f00000000;
ld.global.nc.u32 %r264, [%rd66+28];
ld.global.nc.u32 %r268, [%rd66+32];

dp4a.s32.s32 %r262, %r783, %r264, %r250;


dp4a.s32.s32 %r266, %r787, %r268, %r262;


dp4a.s32.s32 %r270, %r264, %r861, %r858;


dp4a.s32.s32 %r274, %r268, %r861, %r270;

cvt.rn.f32.s32 %f162, %r274;
fma.rn.ftz.f32 %f163, %f115, %f162, %f161;

{ cvt.f32.f16 %f77, %rs11;}


ld.global.nc.u32 %r278, [%rd66];

{.reg .f16 low,high;
mov.b32 {low,high},%r278;
cvt.f32.f16 %f78, low;}


mul.ftz.f32 %f164, %f77, %f78;
cvt.rn.f32.s32 %f165, %r234;
add.ftz.f32 %f166, %f159, %f165;
cvt.rn.f32.s32 %f167, %r266;
add.ftz.f32 %f168, %f163, %f167;
mul.ftz.f32 %f169, %f168, %f124;
fma.rn.ftz.f32 %f170, %f166, %f121, %f169;
fma.rn.ftz.f32 %f459, %f164, %f170, %f459;

dp4a.s32.s32 %r279, %r800, %r216, %r858;


dp4a.s32.s32 %r283, %r804, %r220, %r279;


dp4a.s32.s32 %r287, %r216, %r861, %r858;


dp4a.s32.s32 %r291, %r220, %r861, %r287;

cvt.rn.f32.s32 %f171, %r291;
fma.rn.ftz.f32 %f172, %f129, %f171, 0f00000000;

dp4a.s32.s32 %r295, %r816, %r232, %r283;


dp4a.s32.s32 %r299, %r820, %r236, %r295;


dp4a.s32.s32 %r303, %r232, %r861, %r858;


dp4a.s32.s32 %r307, %r236, %r861, %r303;

cvt.rn.f32.s32 %f173, %r307;
fma.rn.ftz.f32 %f174, %f134, %f173, %f172;

dp4a.s32.s32 %r311, %r832, %r248, %r858;


dp4a.s32.s32 %r315, %r836, %r252, %r311;


dp4a.s32.s32 %r319, %r248, %r861, %r858;


dp4a.s32.s32 %r323, %r252, %r861, %r319;

cvt.rn.f32.s32 %f175, %r323;
fma.rn.ftz.f32 %f176, %f139, %f175, 0f00000000;

dp4a.s32.s32 %r327, %r848, %r264, %r315;


dp4a.s32.s32 %r331, %r852, %r268, %r327;


dp4a.s32.s32 %r335, %r264, %r861, %r858;


dp4a.s32.s32 %r339, %r268, %r861, %r335;

cvt.rn.f32.s32 %f177, %r339;
fma.rn.ftz.f32 %f178, %f144, %f177, %f176;

{ cvt.f32.f16 %f79, %rs12;}



{.reg .f16 low,high;
mov.b32 {low,high},%r278;
cvt.f32.f16 %f80, low;}


mul.ftz.f32 %f179, %f79, %f80;
cvt.rn.f32.s32 %f180, %r299;
add.ftz.f32 %f181, %f174, %f180;
cvt.rn.f32.s32 %f182, %r331;
add.ftz.f32 %f183, %f178, %f182;
mul.ftz.f32 %f184, %f183, %f153;
fma.rn.ftz.f32 %f185, %f181, %f150, %f184;
fma.rn.ftz.f32 %f458, %f179, %f185, %f458;
cvt.u64.u32 %rd67, %r1170;
add.s64 %rd68, %rd2, %rd67;
add.s64 %rd69, %rd68, %rd4;
mul.lo.s64 %rd70, %rd69, 36;
add.s64 %rd71, %rd13, %rd70;
ld.global.nc.u32 %r346, [%rd71+4];
ld.global.nc.u32 %r350, [%rd71+8];

dp4a.s32.s32 %r344, %r735, %r346, %r858;


dp4a.s32.s32 %r348, %r739, %r350, %r344;


dp4a.s32.s32 %r352, %r346, %r861, %r858;


dp4a.s32.s32 %r356, %r350, %r861, %r352;

cvt.rn.f32.s32 %f186, %r356;
fma.rn.ftz.f32 %f187, %f100, %f186, 0f00000000;
ld.global.nc.u32 %r362, [%rd71+12];
ld.global.nc.u32 %r366, [%rd71+16];

dp4a.s32.s32 %r360, %r751, %r362, %r348;


dp4a.s32.s32 %r364, %r755, %r366, %r360;


dp4a.s32.s32 %r368, %r362, %r861, %r858;


dp4a.s32.s32 %r372, %r366, %r861, %r368;

cvt.rn.f32.s32 %f188, %r372;
fma.rn.ftz.f32 %f189, %f105, %f188, %f187;
ld.global.nc.u32 %r378, [%rd71+20];
ld.global.nc.u32 %r382, [%rd71+24];

dp4a.s32.s32 %r376, %r767, %r378, %r858;


dp4a.s32.s32 %r380, %r771, %r382, %r376;


dp4a.s32.s32 %r384, %r378, %r861, %r858;


dp4a.s32.s32 %r388, %r382, %r861, %r384;

cvt.rn.f32.s32 %f190, %r388;
fma.rn.ftz.f32 %f191, %f110, %f190, 0f00000000;
ld.global.nc.u32 %r394, [%rd71+28];
ld.global.nc.u32 %r398, [%rd71+32];

dp4a.s32.s32 %r392, %r783, %r394, %r380;


dp4a.s32.s32 %r396, %r787, %r398, %r392;


dp4a.s32.s32 %r400, %r394, %r861, %r858;


dp4a.s32.s32 %r404, %r398, %r861, %r400;

cvt.rn.f32.s32 %f192, %r404;
fma.rn.ftz.f32 %f193, %f115, %f192, %f191;

{ cvt.f32.f16 %f81, %rs11;}


ld.global.nc.u32 %r408, [%rd71];

{.reg .f16 low,high;
mov.b32 {low,high},%r408;
cvt.f32.f16 %f82, low;}


mul.ftz.f32 %f194, %f81, %f82;
cvt.rn.f32.s32 %f195, %r364;
add.ftz.f32 %f196, %f189, %f195;
cvt.rn.f32.s32 %f197, %r396;
add.ftz.f32 %f198, %f193, %f197;
mul.ftz.f32 %f199, %f198, %f124;
fma.rn.ftz.f32 %f200, %f196, %f121, %f199;
fma.rn.ftz.f32 %f457, %f194, %f200, %f457;

dp4a.s32.s32 %r409, %r800, %r346, %r858;


dp4a.s32.s32 %r413, %r804, %r350, %r409;


dp4a.s32.s32 %r417, %r346, %r861, %r858;


dp4a.s32.s32 %r421, %r350, %r861, %r417;

cvt.rn.f32.s32 %f201, %r421;
fma.rn.ftz.f32 %f202, %f129, %f201, 0f00000000;

dp4a.s32.s32 %r425, %r816, %r362, %r413;


dp4a.s32.s32 %r429, %r820, %r366, %r425;


dp4a.s32.s32 %r433, %r362, %r861, %r858;


dp4a.s32.s32 %r437, %r366, %r861, %r433;

cvt.rn.f32.s32 %f203, %r437;
fma.rn.ftz.f32 %f204, %f134, %f203, %f202;

dp4a.s32.s32 %r441, %r832, %r378, %r858;


dp4a.s32.s32 %r445, %r836, %r382, %r441;


dp4a.s32.s32 %r449, %r378, %r861, %r858;


dp4a.s32.s32 %r453, %r382, %r861, %r449;

cvt.rn.f32.s32 %f205, %r453;
fma.rn.ftz.f32 %f206, %f139, %f205, 0f00000000;

dp4a.s32.s32 %r457, %r848, %r394, %r445;


dp4a.s32.s32 %r461, %r852, %r398, %r457;


dp4a.s32.s32 %r465, %r394, %r861, %r858;


dp4a.s32.s32 %r469, %r398, %r861, %r465;

cvt.rn.f32.s32 %f207, %r469;
fma.rn.ftz.f32 %f208, %f144, %f207, %f206;

{ cvt.f32.f16 %f83, %rs12;}



{.reg .f16 low,high;
mov.b32 {low,high},%r408;
cvt.f32.f16 %f84, low;}


mul.ftz.f32 %f209, %f83, %f84;
cvt.rn.f32.s32 %f210, %r429;
add.ftz.f32 %f211, %f204, %f210;
cvt.rn.f32.s32 %f212, %r461;
add.ftz.f32 %f213, %f208, %f212;
mul.ftz.f32 %f214, %f213, %f153;
fma.rn.ftz.f32 %f215, %f211, %f150, %f214;
fma.rn.ftz.f32 %f456, %f209, %f215, %f456;
cvt.u64.u32 %rd72, %r1169;
add.s64 %rd73, %rd2, %rd72;
add.s64 %rd74, %rd73, %rd4;
mul.lo.s64 %rd75, %rd74, 36;
add.s64 %rd76, %rd13, %rd75;
ld.global.nc.u32 %r476, [%rd76+4];
ld.global.nc.u32 %r480, [%rd76+8];

dp4a.s32.s32 %r474, %r735, %r476, %r858;


dp4a.s32.s32 %r478, %r739, %r480, %r474;


dp4a.s32.s32 %r482, %r476, %r861, %r858;


dp4a.s32.s32 %r486, %r480, %r861, %r482;

cvt.rn.f32.s32 %f216, %r486;
fma.rn.ftz.f32 %f217, %f100, %f216, 0f00000000;
ld.global.nc.u32 %r492, [%rd76+12];
ld.global.nc.u32 %r496, [%rd76+16];

dp4a.s32.s32 %r490, %r751, %r492, %r478;


dp4a.s32.s32 %r494, %r755, %r496, %r490;


dp4a.s32.s32 %r498, %r492, %r861, %r858;


dp4a.s32.s32 %r502, %r496, %r861, %r498;

cvt.rn.f32.s32 %f218, %r502;
fma.rn.ftz.f32 %f219, %f105, %f218, %f217;
ld.global.nc.u32 %r508, [%rd76+20];
ld.global.nc.u32 %r512, [%rd76+24];

dp4a.s32.s32 %r506, %r767, %r508, %r858;


dp4a.s32.s32 %r510, %r771, %r512, %r506;


dp4a.s32.s32 %r514, %r508, %r861, %r858;


dp4a.s32.s32 %r518, %r512, %r861, %r514;

cvt.rn.f32.s32 %f220, %r518;
fma.rn.ftz.f32 %f221, %f110, %f220, 0f00000000;
ld.global.nc.u32 %r524, [%rd76+28];
ld.global.nc.u32 %r528, [%rd76+32];

dp4a.s32.s32 %r522, %r783, %r524, %r510;


dp4a.s32.s32 %r526, %r787, %r528, %r522;


dp4a.s32.s32 %r530, %r524, %r861, %r858;


dp4a.s32.s32 %r534, %r528, %r861, %r530;

cvt.rn.f32.s32 %f222, %r534;
fma.rn.ftz.f32 %f223, %f115, %f222, %f221;

{ cvt.f32.f16 %f85, %rs11;}


ld.global.nc.u32 %r538, [%rd76];

{.reg .f16 low,high;
mov.b32 {low,high},%r538;
cvt.f32.f16 %f86, low;}


mul.ftz.f32 %f224, %f85, %f86;
cvt.rn.f32.s32 %f225, %r494;
add.ftz.f32 %f226, %f219, %f225;
cvt.rn.f32.s32 %f227, %r526;
add.ftz.f32 %f228, %f223, %f227;
mul.ftz.f32 %f229, %f228, %f124;
fma.rn.ftz.f32 %f230, %f226, %f121, %f229;
fma.rn.ftz.f32 %f455, %f224, %f230, %f455;

dp4a.s32.s32 %r539, %r800, %r476, %r858;


dp4a.s32.s32 %r543, %r804, %r480, %r539;


dp4a.s32.s32 %r547, %r476, %r861, %r858;


dp4a.s32.s32 %r551, %r480, %r861, %r547;

cvt.rn.f32.s32 %f231, %r551;
fma.rn.ftz.f32 %f232, %f129, %f231, 0f00000000;

dp4a.s32.s32 %r555, %r816, %r492, %r543;


dp4a.s32.s32 %r559, %r820, %r496, %r555;


dp4a.s32.s32 %r563, %r492, %r861, %r858;


dp4a.s32.s32 %r567, %r496, %r861, %r563;

cvt.rn.f32.s32 %f233, %r567;
fma.rn.ftz.f32 %f234, %f134, %f233, %f232;

dp4a.s32.s32 %r571, %r832, %r508, %r858;


dp4a.s32.s32 %r575, %r836, %r512, %r571;


dp4a.s32.s32 %r579, %r508, %r861, %r858;


dp4a.s32.s32 %r583, %r512, %r861, %r579;

cvt.rn.f32.s32 %f235, %r583;
fma.rn.ftz.f32 %f236, %f139, %f235, 0f00000000;

dp4a.s32.s32 %r587, %r848, %r524, %r575;


dp4a.s32.s32 %r591, %r852, %r528, %r587;


dp4a.s32.s32 %r595, %r524, %r861, %r858;


dp4a.s32.s32 %r599, %r528, %r861, %r595;

cvt.rn.f32.s32 %f237, %r599;
fma.rn.ftz.f32 %f238, %f144, %f237, %f236;

{ cvt.f32.f16 %f87, %rs12;}



{.reg .f16 low,high;
mov.b32 {low,high},%r538;
cvt.f32.f16 %f88, low;}


mul.ftz.f32 %f239, %f87, %f88;
cvt.rn.f32.s32 %f240, %r559;
add.ftz.f32 %f241, %f234, %f240;
cvt.rn.f32.s32 %f242, %r591;
add.ftz.f32 %f243, %f238, %f242;
mul.ftz.f32 %f244, %f243, %f153;
fma.rn.ftz.f32 %f245, %f241, %f150, %f244;
fma.rn.ftz.f32 %f454, %f239, %f245, %f454;
cvt.u64.u32 %rd77, %r1168;
add.s64 %rd78, %rd2, %rd77;
add.s64 %rd79, %rd78, %rd4;
mul.lo.s64 %rd80, %rd79, 36;
add.s64 %rd81, %rd13, %rd80;
ld.global.nc.u32 %r606, [%rd81+4];
ld.global.nc.u32 %r610, [%rd81+8];

dp4a.s32.s32 %r604, %r735, %r606, %r858;


dp4a.s32.s32 %r608, %r739, %r610, %r604;


dp4a.s32.s32 %r612, %r606, %r861, %r858;


dp4a.s32.s32 %r616, %r610, %r861, %r612;

cvt.rn.f32.s32 %f246, %r616;
fma.rn.ftz.f32 %f247, %f100, %f246, 0f00000000;
ld.global.nc.u32 %r622, [%rd81+12];
ld.global.nc.u32 %r626, [%rd81+16];

dp4a.s32.s32 %r620, %r751, %r622, %r608;


dp4a.s32.s32 %r624, %r755, %r626, %r620;


dp4a.s32.s32 %r628, %r622, %r861, %r858;


dp4a.s32.s32 %r632, %r626, %r861, %r628;

cvt.rn.f32.s32 %f248, %r632;
fma.rn.ftz.f32 %f249, %f105, %f248, %f247;
ld.global.nc.u32 %r638, [%rd81+20];
ld.global.nc.u32 %r642, [%rd81+24];

dp4a.s32.s32 %r636, %r767, %r638, %r858;


dp4a.s32.s32 %r640, %r771, %r642, %r636;


dp4a.s32.s32 %r644, %r638, %r861, %r858;


dp4a.s32.s32 %r648, %r642, %r861, %r644;

cvt.rn.f32.s32 %f250, %r648;
fma.rn.ftz.f32 %f251, %f110, %f250, 0f00000000;
ld.global.nc.u32 %r654, [%rd81+28];
ld.global.nc.u32 %r658, [%rd81+32];

dp4a.s32.s32 %r652, %r783, %r654, %r640;


dp4a.s32.s32 %r656, %r787, %r658, %r652;


dp4a.s32.s32 %r660, %r654, %r861, %r858;


dp4a.s32.s32 %r664, %r658, %r861, %r660;

cvt.rn.f32.s32 %f252, %r664;
fma.rn.ftz.f32 %f253, %f115, %f252, %f251;

{ cvt.f32.f16 %f89, %rs11;}


ld.global.nc.u32 %r668, [%rd81];

{.reg .f16 low,high;
mov.b32 {low,high},%r668;
cvt.f32.f16 %f90, low;}


mul.ftz.f32 %f254, %f89, %f90;
cvt.rn.f32.s32 %f255, %r624;
add.ftz.f32 %f256, %f249, %f255;
cvt.rn.f32.s32 %f257, %r656;
add.ftz.f32 %f258, %f253, %f257;
mul.ftz.f32 %f259, %f258, %f124;
fma.rn.ftz.f32 %f260, %f256, %f121, %f259;
fma.rn.ftz.f32 %f453, %f254, %f260, %f453;

dp4a.s32.s32 %r669, %r800, %r606, %r858;


dp4a.s32.s32 %r673, %r804, %r610, %r669;


dp4a.s32.s32 %r677, %r606, %r861, %r858;


dp4a.s32.s32 %r681, %r610, %r861, %r677;

cvt.rn.f32.s32 %f261, %r681;
fma.rn.ftz.f32 %f262, %f129, %f261, 0f00000000;

dp4a.s32.s32 %r685, %r816, %r622, %r673;


dp4a.s32.s32 %r689, %r820, %r626, %r685;


dp4a.s32.s32 %r693, %r622, %r861, %r858;


dp4a.s32.s32 %r697, %r626, %r861, %r693;

cvt.rn.f32.s32 %f263, %r697;
fma.rn.ftz.f32 %f264, %f134, %f263, %f262;

dp4a.s32.s32 %r701, %r832, %r638, %r858;


dp4a.s32.s32 %r705, %r836, %r642, %r701;


dp4a.s32.s32 %r709, %r638, %r861, %r858;


dp4a.s32.s32 %r713, %r642, %r861, %r709;

cvt.rn.f32.s32 %f265, %r713;
fma.rn.ftz.f32 %f266, %f139, %f265, 0f00000000;

dp4a.s32.s32 %r717, %r848, %r654, %r705;


dp4a.s32.s32 %r721, %r852, %r658, %r717;


dp4a.s32.s32 %r725, %r654, %r861, %r858;


dp4a.s32.s32 %r729, %r658, %r861, %r725;

cvt.rn.f32.s32 %f267, %r729;
fma.rn.ftz.f32 %f268, %f144, %f267, %f266;

{ cvt.f32.f16 %f91, %rs12;}



{.reg .f16 low,high;
mov.b32 {low,high},%r668;
cvt.f32.f16 %f92, low;}


mul.ftz.f32 %f269, %f91, %f92;
cvt.rn.f32.s32 %f270, %r689;
add.ftz.f32 %f271, %f264, %f270;
cvt.rn.f32.s32 %f272, %r721;
add.ftz.f32 %f273, %f268, %f272;
mul.ftz.f32 %f274, %f273, %f153;
fma.rn.ftz.f32 %f275, %f271, %f150, %f274;
fma.rn.ftz.f32 %f452, %f269, %f275, %f452;
cvt.u64.u32 %rd82, %r1167;
add.s64 %rd83, %rd2, %rd82;
add.s64 %rd84, %rd83, %rd4;
mul.lo.s64 %rd85, %rd84, 36;
add.s64 %rd86, %rd13, %rd85;
ld.global.nc.u32 %r736, [%rd86+4];
ld.global.nc.u32 %r740, [%rd86+8];

dp4a.s32.s32 %r734, %r735, %r736, %r858;


dp4a.s32.s32 %r738, %r739, %r740, %r734;


dp4a.s32.s32 %r742, %r736, %r861, %r858;


dp4a.s32.s32 %r746, %r740, %r861, %r742;

cvt.rn.f32.s32 %f276, %r746;
fma.rn.ftz.f32 %f277, %f100, %f276, 0f00000000;
ld.global.nc.u32 %r752, [%rd86+12];
ld.global.nc.u32 %r756, [%rd86+16];

dp4a.s32.s32 %r750, %r751, %r752, %r738;


dp4a.s32.s32 %r754, %r755, %r756, %r750;


dp4a.s32.s32 %r758, %r752, %r861, %r858;


dp4a.s32.s32 %r762, %r756, %r861, %r758;

cvt.rn.f32.s32 %f278, %r762;
fma.rn.ftz.f32 %f279, %f105, %f278, %f277;
ld.global.nc.u32 %r768, [%rd86+20];
ld.global.nc.u32 %r772, [%rd86+24];

dp4a.s32.s32 %r766, %r767, %r768, %r858;


dp4a.s32.s32 %r770, %r771, %r772, %r766;


dp4a.s32.s32 %r774, %r768, %r861, %r858;


dp4a.s32.s32 %r778, %r772, %r861, %r774;

cvt.rn.f32.s32 %f280, %r778;
fma.rn.ftz.f32 %f281, %f110, %f280, 0f00000000;
ld.global.nc.u32 %r784, [%rd86+28];
ld.global.nc.u32 %r788, [%rd86+32];

dp4a.s32.s32 %r782, %r783, %r784, %r770;


dp4a.s32.s32 %r786, %r787, %r788, %r782;


dp4a.s32.s32 %r790, %r784, %r861, %r858;


dp4a.s32.s32 %r794, %r788, %r861, %r790;

cvt.rn.f32.s32 %f282, %r794;
fma.rn.ftz.f32 %f283, %f115, %f282, %f281;

{ cvt.f32.f16 %f93, %rs11;}


ld.global.nc.u32 %r798, [%rd86];

{.reg .f16 low,high;
mov.b32 {low,high},%r798;
cvt.f32.f16 %f94, low;}


mul.ftz.f32 %f284, %f93, %f94;
cvt.rn.f32.s32 %f285, %r754;
add.ftz.f32 %f286, %f279, %f285;
cvt.rn.f32.s32 %f287, %r786;
add.ftz.f32 %f288, %f283, %f287;
mul.ftz.f32 %f289, %f288, %f124;
fma.rn.ftz.f32 %f290, %f286, %f121, %f289;
fma.rn.ftz.f32 %f451, %f284, %f290, %f451;

dp4a.s32.s32 %r799, %r800, %r736, %r858;


dp4a.s32.s32 %r803, %r804, %r740, %r799;


dp4a.s32.s32 %r807, %r736, %r861, %r858;


dp4a.s32.s32 %r811, %r740, %r861, %r807;

cvt.rn.f32.s32 %f291, %r811;
fma.rn.ftz.f32 %f292, %f129, %f291, 0f00000000;

dp4a.s32.s32 %r815, %r816, %r752, %r803;


dp4a.s32.s32 %r819, %r820, %r756, %r815;


dp4a.s32.s32 %r823, %r752, %r861, %r858;


dp4a.s32.s32 %r827, %r756, %r861, %r823;

cvt.rn.f32.s32 %f293, %r827;
fma.rn.ftz.f32 %f294, %f134, %f293, %f292;

dp4a.s32.s32 %r831, %r832, %r768, %r858;


dp4a.s32.s32 %r835, %r836, %r772, %r831;


dp4a.s32.s32 %r839, %r768, %r861, %r858;


dp4a.s32.s32 %r843, %r772, %r861, %r839;

cvt.rn.f32.s32 %f295, %r843;
fma.rn.ftz.f32 %f296, %f139, %f295, 0f00000000;

dp4a.s32.s32 %r847, %r848, %r784, %r835;


dp4a.s32.s32 %r851, %r852, %r788, %r847;


dp4a.s32.s32 %r855, %r784, %r861, %r858;


dp4a.s32.s32 %r859, %r788, %r861, %r855;

cvt.rn.f32.s32 %f297, %r859;
fma.rn.ftz.f32 %f298, %f144, %f297, %f296;

{ cvt.f32.f16 %f95, %rs12;}



{.reg .f16 low,high;
mov.b32 {low,high},%r798;
cvt.f32.f16 %f96, low;}


mul.ftz.f32 %f299, %f95, %f96;
cvt.rn.f32.s32 %f300, %r819;
add.ftz.f32 %f301, %f294, %f300;
cvt.rn.f32.s32 %f302, %r851;
add.ftz.f32 %f303, %f298, %f302;
mul.ftz.f32 %f304, %f303, %f153;
fma.rn.ftz.f32 %f305, %f301, %f150, %f304;
fma.rn.ftz.f32 %f450, %f299, %f305, %f450;
add.s32 %r1170, %r1170, 64;
add.s32 %r1169, %r1169, 64;
add.s32 %r1168, %r1168, 64;
add.s32 %r1167, %r1167, 64;
add.s32 %r1166, %r1166, 64;
add.s32 %r1165, %r1165, 64;
add.s64 %rd118, %rd118, 448;
add.s32 %r1171, %r1171, 8;
setp.lt.s32 %p4, %r1171, %r7;
@%p4 bra $L__BB273_2;

$L__BB273_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB273_5;

mov.u32 %r945, _ZZ13mul_mat_vec_qIL9ggml_type29ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r946, %r1, 1536, %r945;
shl.b32 %r947, %r49, 2;
add.s32 %r948, %r946, %r947;
st.shared.f32 [%r948+-1536], %f461;
st.shared.f32 [%r948+-1408], %f460;
st.shared.f32 [%r948+-1280], %f459;
st.shared.f32 [%r948+-1152], %f458;
st.shared.f32 [%r948+-1024], %f457;
st.shared.f32 [%r948+-896], %f456;
st.shared.f32 [%r948+-768], %f455;
st.shared.f32 [%r948+-640], %f454;
st.shared.f32 [%r948+-512], %f453;
st.shared.f32 [%r948+-384], %f452;
st.shared.f32 [%r948+-256], %f451;
st.shared.f32 [%r948+-128], %f450;

$L__BB273_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB273_30;

mad.lo.s32 %r949, %r4, %r40, %r3;
mad.lo.s32 %r950, %r5, %r46, %r949;
cvt.u64.u32 %rd16, %r950;
shl.b32 %r952, %r49, 2;
mov.u32 %r953, _ZZ13mul_mat_vec_qIL9ggml_type29ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r29, %r953, %r952;
mov.u32 %r954, 2;
ld.shared.f32 %f306, [%r29];
add.ftz.f32 %f307, %f306, %f461;
mov.b32 %r955, %f307;
mov.u32 %r956, 31;
mov.u32 %r957, 16;
mov.u32 %r958, -1;
shfl.sync.bfly.b32 %r959|%p7, %r955, %r957, %r956, %r958;
mov.b32 %f308, %r959;
add.ftz.f32 %f309, %f307, %f308;
mov.b32 %r960, %f309;
mov.u32 %r961, 8;
shfl.sync.bfly.b32 %r962|%p8, %r960, %r961, %r956, %r958;
mov.b32 %f310, %r962;
add.ftz.f32 %f311, %f309, %f310;
mov.b32 %r963, %f311;
mov.u32 %r964, 4;
shfl.sync.bfly.b32 %r965|%p9, %r963, %r964, %r956, %r958;
mov.b32 %f312, %r965;
add.ftz.f32 %f313, %f311, %f312;
mov.b32 %r966, %f313;
shfl.sync.bfly.b32 %r967|%p10, %r966, %r954, %r956, %r958;
mov.b32 %f314, %r967;
add.ftz.f32 %f315, %f313, %f314;
mov.b32 %r968, %f315;
mov.u32 %r969, 1;
shfl.sync.bfly.b32 %r970|%p11, %r968, %r969, %r956, %r958;
mov.b32 %f316, %r970;
add.ftz.f32 %f37, %f315, %f316;
setp.eq.s32 %p12, %r49, 0;
setp.lt.u32 %p13, %r3, %r34;
and.pred %p1, %p12, %p13;
mul.wide.u32 %rd87, %r950, 4;
add.s64 %rd17, %rd1, %rd87;
not.pred %p14, %p1;
@%p14 bra $L__BB273_8;

st.global.f32 [%rd17], %f37;

$L__BB273_8:
ld.shared.f32 %f317, [%r29+128];
add.ftz.f32 %f318, %f317, %f460;
mov.b32 %r971, %f318;
shfl.sync.bfly.b32 %r975|%p15, %r971, %r957, %r956, %r958;
mov.b32 %f319, %r975;
add.ftz.f32 %f320, %f318, %f319;
mov.b32 %r976, %f320;
shfl.sync.bfly.b32 %r978|%p16, %r976, %r961, %r956, %r958;
mov.b32 %f321, %r978;
add.ftz.f32 %f322, %f320, %f321;
mov.b32 %r979, %f322;
shfl.sync.bfly.b32 %r981|%p17, %r979, %r964, %r956, %r958;
mov.b32 %f323, %r981;
add.ftz.f32 %f324, %f322, %f323;
mov.b32 %r982, %f324;
shfl.sync.bfly.b32 %r984|%p18, %r982, %r954, %r956, %r958;
mov.b32 %f325, %r984;
add.ftz.f32 %f326, %f324, %f325;
mov.b32 %r985, %f326;
shfl.sync.bfly.b32 %r987|%p19, %r985, %r969, %r956, %r958;
mov.b32 %f327, %r987;
add.ftz.f32 %f38, %f326, %f327;
setp.eq.s32 %p20, %r49, 1;
add.s32 %r989, %r3, 1;
setp.lt.u32 %p21, %r989, %r34;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB273_10;

st.global.f32 [%rd17+4], %f38;

$L__BB273_10:
ld.shared.f32 %f328, [%r29+256];
add.ftz.f32 %f329, %f328, %f459;
mov.b32 %r990, %f329;
mov.u32 %r991, 31;
mov.u32 %r992, 16;
mov.u32 %r993, -1;
shfl.sync.bfly.b32 %r994|%p23, %r990, %r992, %r991, %r993;
mov.b32 %f330, %r994;
add.ftz.f32 %f331, %f329, %f330;
mov.b32 %r995, %f331;
mov.u32 %r996, 8;
shfl.sync.bfly.b32 %r997|%p24, %r995, %r996, %r991, %r993;
mov.b32 %f332, %r997;
add.ftz.f32 %f333, %f331, %f332;
mov.b32 %r998, %f333;
mov.u32 %r999, 4;
shfl.sync.bfly.b32 %r1000|%p25, %r998, %r999, %r991, %r993;
mov.b32 %f334, %r1000;
add.ftz.f32 %f335, %f333, %f334;
mov.b32 %r1001, %f335;
mov.u32 %r1002, 2;
shfl.sync.bfly.b32 %r1003|%p26, %r1001, %r1002, %r991, %r993;
mov.b32 %f336, %r1003;
add.ftz.f32 %f337, %f335, %f336;
mov.b32 %r1004, %f337;
mov.u32 %r1005, 1;
shfl.sync.bfly.b32 %r1006|%p27, %r1004, %r1005, %r991, %r993;
mov.b32 %f338, %r1006;
add.ftz.f32 %f39, %f337, %f338;
@%p14 bra $L__BB273_12;

cvt.u64.u32 %rd88, %r34;
add.s64 %rd89, %rd88, %rd16;
shl.b64 %rd90, %rd89, 2;
add.s64 %rd91, %rd1, %rd90;
st.global.f32 [%rd91], %f39;

$L__BB273_12:
ld.shared.f32 %f339, [%r29+384];
add.ftz.f32 %f340, %f339, %f458;
mov.b32 %r1007, %f340;
shfl.sync.bfly.b32 %r1011|%p29, %r1007, %r992, %r991, %r993;
mov.b32 %f341, %r1011;
add.ftz.f32 %f342, %f340, %f341;
mov.b32 %r1012, %f342;
shfl.sync.bfly.b32 %r1014|%p30, %r1012, %r996, %r991, %r993;
mov.b32 %f343, %r1014;
add.ftz.f32 %f344, %f342, %f343;
mov.b32 %r1015, %f344;
shfl.sync.bfly.b32 %r1017|%p31, %r1015, %r999, %r991, %r993;
mov.b32 %f345, %r1017;
add.ftz.f32 %f346, %f344, %f345;
mov.b32 %r1018, %f346;
shfl.sync.bfly.b32 %r1020|%p32, %r1018, %r1002, %r991, %r993;
mov.b32 %f347, %r1020;
add.ftz.f32 %f348, %f346, %f347;
mov.b32 %r1021, %f348;
shfl.sync.bfly.b32 %r1023|%p33, %r1021, %r1005, %r991, %r993;
mov.b32 %f349, %r1023;
add.ftz.f32 %f40, %f348, %f349;
@%p22 bra $L__BB273_14;

add.s32 %r1024, %r34, 1;
cvt.u64.u32 %rd92, %r1024;
add.s64 %rd93, %rd92, %rd16;
shl.b64 %rd94, %rd93, 2;
add.s64 %rd95, %rd1, %rd94;
st.global.f32 [%rd95], %f40;

$L__BB273_14:
shl.b32 %r1025, %r34, 1;
mov.u32 %r1026, 1;
ld.shared.f32 %f350, [%r29+512];
add.ftz.f32 %f351, %f350, %f457;
mov.b32 %r1027, %f351;
mov.u32 %r1028, 31;
mov.u32 %r1029, 16;
mov.u32 %r1030, -1;
shfl.sync.bfly.b32 %r1031|%p35, %r1027, %r1029, %r1028, %r1030;
mov.b32 %f352, %r1031;
add.ftz.f32 %f353, %f351, %f352;
mov.b32 %r1032, %f353;
mov.u32 %r1033, 8;
shfl.sync.bfly.b32 %r1034|%p36, %r1032, %r1033, %r1028, %r1030;
mov.b32 %f354, %r1034;
add.ftz.f32 %f355, %f353, %f354;
mov.b32 %r1035, %f355;
mov.u32 %r1036, 4;
shfl.sync.bfly.b32 %r1037|%p37, %r1035, %r1036, %r1028, %r1030;
mov.b32 %f356, %r1037;
add.ftz.f32 %f357, %f355, %f356;
mov.b32 %r1038, %f357;
mov.u32 %r1039, 2;
shfl.sync.bfly.b32 %r1040|%p38, %r1038, %r1039, %r1028, %r1030;
mov.b32 %f358, %r1040;
add.ftz.f32 %f359, %f357, %f358;
mov.b32 %r1041, %f359;
shfl.sync.bfly.b32 %r1042|%p39, %r1041, %r1026, %r1028, %r1030;
mov.b32 %f360, %r1042;
add.ftz.f32 %f41, %f359, %f360;
cvt.u64.u32 %rd96, %r1025;
add.s64 %rd97, %rd96, %rd16;
shl.b64 %rd98, %rd97, 2;
add.s64 %rd18, %rd1, %rd98;
@%p14 bra $L__BB273_16;

st.global.f32 [%rd18], %f41;

$L__BB273_16:
ld.shared.f32 %f361, [%r29+640];
add.ftz.f32 %f362, %f361, %f456;
mov.b32 %r1043, %f362;
shfl.sync.bfly.b32 %r1047|%p41, %r1043, %r1029, %r1028, %r1030;
mov.b32 %f363, %r1047;
add.ftz.f32 %f364, %f362, %f363;
mov.b32 %r1048, %f364;
shfl.sync.bfly.b32 %r1050|%p42, %r1048, %r1033, %r1028, %r1030;
mov.b32 %f365, %r1050;
add.ftz.f32 %f366, %f364, %f365;
mov.b32 %r1051, %f366;
shfl.sync.bfly.b32 %r1053|%p43, %r1051, %r1036, %r1028, %r1030;
mov.b32 %f367, %r1053;
add.ftz.f32 %f368, %f366, %f367;
mov.b32 %r1054, %f368;
shfl.sync.bfly.b32 %r1056|%p44, %r1054, %r1039, %r1028, %r1030;
mov.b32 %f369, %r1056;
add.ftz.f32 %f370, %f368, %f369;
mov.b32 %r1057, %f370;
shfl.sync.bfly.b32 %r1059|%p45, %r1057, %r1026, %r1028, %r1030;
mov.b32 %f371, %r1059;
add.ftz.f32 %f42, %f370, %f371;
@%p22 bra $L__BB273_18;

st.global.f32 [%rd18+4], %f42;

$L__BB273_18:
mul.lo.s32 %r30, %r34, 3;
ld.shared.f32 %f372, [%r29+768];
add.ftz.f32 %f373, %f372, %f455;
mov.b32 %r1060, %f373;
mov.u32 %r1061, 31;
mov.u32 %r1062, 16;
mov.u32 %r1063, -1;
shfl.sync.bfly.b32 %r1064|%p47, %r1060, %r1062, %r1061, %r1063;
mov.b32 %f374, %r1064;
add.ftz.f32 %f375, %f373, %f374;
mov.b32 %r1065, %f375;
mov.u32 %r1066, 8;
shfl.sync.bfly.b32 %r1067|%p48, %r1065, %r1066, %r1061, %r1063;
mov.b32 %f376, %r1067;
add.ftz.f32 %f377, %f375, %f376;
mov.b32 %r1068, %f377;
mov.u32 %r1069, 4;
shfl.sync.bfly.b32 %r1070|%p49, %r1068, %r1069, %r1061, %r1063;
mov.b32 %f378, %r1070;
add.ftz.f32 %f379, %f377, %f378;
mov.b32 %r1071, %f379;
mov.u32 %r1072, 2;
shfl.sync.bfly.b32 %r1073|%p50, %r1071, %r1072, %r1061, %r1063;
mov.b32 %f380, %r1073;
add.ftz.f32 %f381, %f379, %f380;
mov.b32 %r1074, %f381;
mov.u32 %r1075, 1;
shfl.sync.bfly.b32 %r1076|%p51, %r1074, %r1075, %r1061, %r1063;
mov.b32 %f382, %r1076;
add.ftz.f32 %f43, %f381, %f382;
@%p14 bra $L__BB273_20;

cvt.u64.u32 %rd99, %r30;
add.s64 %rd100, %rd99, %rd16;
shl.b64 %rd101, %rd100, 2;
add.s64 %rd102, %rd1, %rd101;
st.global.f32 [%rd102], %f43;

$L__BB273_20:
ld.shared.f32 %f383, [%r29+896];
add.ftz.f32 %f384, %f383, %f454;
mov.b32 %r1077, %f384;
shfl.sync.bfly.b32 %r1081|%p53, %r1077, %r1062, %r1061, %r1063;
mov.b32 %f385, %r1081;
add.ftz.f32 %f386, %f384, %f385;
mov.b32 %r1082, %f386;
shfl.sync.bfly.b32 %r1084|%p54, %r1082, %r1066, %r1061, %r1063;
mov.b32 %f387, %r1084;
add.ftz.f32 %f388, %f386, %f387;
mov.b32 %r1085, %f388;
shfl.sync.bfly.b32 %r1087|%p55, %r1085, %r1069, %r1061, %r1063;
mov.b32 %f389, %r1087;
add.ftz.f32 %f390, %f388, %f389;
mov.b32 %r1088, %f390;
shfl.sync.bfly.b32 %r1090|%p56, %r1088, %r1072, %r1061, %r1063;
mov.b32 %f391, %r1090;
add.ftz.f32 %f392, %f390, %f391;
mov.b32 %r1091, %f392;
shfl.sync.bfly.b32 %r1093|%p57, %r1091, %r1075, %r1061, %r1063;
mov.b32 %f393, %r1093;
add.ftz.f32 %f44, %f392, %f393;
@%p22 bra $L__BB273_22;

add.s32 %r1094, %r30, 1;
cvt.u64.u32 %rd103, %r1094;
add.s64 %rd104, %rd103, %rd16;
shl.b64 %rd105, %rd104, 2;
add.s64 %rd106, %rd1, %rd105;
st.global.f32 [%rd106], %f44;

$L__BB273_22:
shl.b32 %r1095, %r34, 2;
mov.u32 %r1096, 2;
ld.shared.f32 %f394, [%r29+1024];
add.ftz.f32 %f395, %f394, %f453;
mov.b32 %r1097, %f395;
mov.u32 %r1098, 31;
mov.u32 %r1099, 16;
mov.u32 %r1100, -1;
shfl.sync.bfly.b32 %r1101|%p59, %r1097, %r1099, %r1098, %r1100;
mov.b32 %f396, %r1101;
add.ftz.f32 %f397, %f395, %f396;
mov.b32 %r1102, %f397;
mov.u32 %r1103, 8;
shfl.sync.bfly.b32 %r1104|%p60, %r1102, %r1103, %r1098, %r1100;
mov.b32 %f398, %r1104;
add.ftz.f32 %f399, %f397, %f398;
mov.b32 %r1105, %f399;
mov.u32 %r1106, 4;
shfl.sync.bfly.b32 %r1107|%p61, %r1105, %r1106, %r1098, %r1100;
mov.b32 %f400, %r1107;
add.ftz.f32 %f401, %f399, %f400;
mov.b32 %r1108, %f401;
shfl.sync.bfly.b32 %r1109|%p62, %r1108, %r1096, %r1098, %r1100;
mov.b32 %f402, %r1109;
add.ftz.f32 %f403, %f401, %f402;
mov.b32 %r1110, %f403;
mov.u32 %r1111, 1;
shfl.sync.bfly.b32 %r1112|%p63, %r1110, %r1111, %r1098, %r1100;
mov.b32 %f404, %r1112;
add.ftz.f32 %f45, %f403, %f404;
cvt.u64.u32 %rd107, %r1095;
add.s64 %rd108, %rd107, %rd16;
shl.b64 %rd109, %rd108, 2;
add.s64 %rd19, %rd1, %rd109;
@%p14 bra $L__BB273_24;

st.global.f32 [%rd19], %f45;

$L__BB273_24:
ld.shared.f32 %f405, [%r29+1152];
add.ftz.f32 %f406, %f405, %f452;
mov.b32 %r1113, %f406;
shfl.sync.bfly.b32 %r1117|%p65, %r1113, %r1099, %r1098, %r1100;
mov.b32 %f407, %r1117;
add.ftz.f32 %f408, %f406, %f407;
mov.b32 %r1118, %f408;
shfl.sync.bfly.b32 %r1120|%p66, %r1118, %r1103, %r1098, %r1100;
mov.b32 %f409, %r1120;
add.ftz.f32 %f410, %f408, %f409;
mov.b32 %r1121, %f410;
shfl.sync.bfly.b32 %r1123|%p67, %r1121, %r1106, %r1098, %r1100;
mov.b32 %f411, %r1123;
add.ftz.f32 %f412, %f410, %f411;
mov.b32 %r1124, %f412;
shfl.sync.bfly.b32 %r1126|%p68, %r1124, %r1096, %r1098, %r1100;
mov.b32 %f413, %r1126;
add.ftz.f32 %f414, %f412, %f413;
mov.b32 %r1127, %f414;
shfl.sync.bfly.b32 %r1129|%p69, %r1127, %r1111, %r1098, %r1100;
mov.b32 %f415, %r1129;
add.ftz.f32 %f46, %f414, %f415;
@%p22 bra $L__BB273_26;

st.global.f32 [%rd19+4], %f46;

$L__BB273_26:
mul.lo.s32 %r31, %r34, 5;
ld.shared.f32 %f416, [%r29+1280];
add.ftz.f32 %f417, %f416, %f451;
mov.b32 %r1130, %f417;
mov.u32 %r1131, 31;
mov.u32 %r1132, 16;
mov.u32 %r1133, -1;
shfl.sync.bfly.b32 %r1134|%p71, %r1130, %r1132, %r1131, %r1133;
mov.b32 %f418, %r1134;
add.ftz.f32 %f419, %f417, %f418;
mov.b32 %r1135, %f419;
mov.u32 %r1136, 8;
shfl.sync.bfly.b32 %r1137|%p72, %r1135, %r1136, %r1131, %r1133;
mov.b32 %f420, %r1137;
add.ftz.f32 %f421, %f419, %f420;
mov.b32 %r1138, %f421;
mov.u32 %r1139, 4;
shfl.sync.bfly.b32 %r1140|%p73, %r1138, %r1139, %r1131, %r1133;
mov.b32 %f422, %r1140;
add.ftz.f32 %f423, %f421, %f422;
mov.b32 %r1141, %f423;
mov.u32 %r1142, 2;
shfl.sync.bfly.b32 %r1143|%p74, %r1141, %r1142, %r1131, %r1133;
mov.b32 %f424, %r1143;
add.ftz.f32 %f425, %f423, %f424;
mov.b32 %r1144, %f425;
mov.u32 %r1145, 1;
shfl.sync.bfly.b32 %r1146|%p75, %r1144, %r1145, %r1131, %r1133;
mov.b32 %f426, %r1146;
add.ftz.f32 %f47, %f425, %f426;
@%p14 bra $L__BB273_28;

cvt.u64.u32 %rd110, %r31;
add.s64 %rd111, %rd110, %rd16;
shl.b64 %rd112, %rd111, 2;
add.s64 %rd113, %rd1, %rd112;
st.global.f32 [%rd113], %f47;

$L__BB273_28:
ld.shared.f32 %f427, [%r29+1408];
add.ftz.f32 %f428, %f427, %f450;
mov.b32 %r1147, %f428;
shfl.sync.bfly.b32 %r1151|%p77, %r1147, %r1132, %r1131, %r1133;
mov.b32 %f429, %r1151;
add.ftz.f32 %f430, %f428, %f429;
mov.b32 %r1152, %f430;
shfl.sync.bfly.b32 %r1154|%p78, %r1152, %r1136, %r1131, %r1133;
mov.b32 %f431, %r1154;
add.ftz.f32 %f432, %f430, %f431;
mov.b32 %r1155, %f432;
shfl.sync.bfly.b32 %r1157|%p79, %r1155, %r1139, %r1131, %r1133;
mov.b32 %f433, %r1157;
add.ftz.f32 %f434, %f432, %f433;
mov.b32 %r1158, %f434;
shfl.sync.bfly.b32 %r1160|%p80, %r1158, %r1142, %r1131, %r1133;
mov.b32 %f435, %r1160;
add.ftz.f32 %f436, %f434, %f435;
mov.b32 %r1161, %f436;
shfl.sync.bfly.b32 %r1163|%p81, %r1161, %r1145, %r1131, %r1133;
mov.b32 %f437, %r1163;
add.ftz.f32 %f48, %f436, %f437;
@%p22 bra $L__BB273_30;

add.s32 %r1164, %r31, 1;
cvt.u64.u32 %rd114, %r1164;
add.s64 %rd115, %rd114, %rd16;
shl.b64 %rd116, %rd115, 2;
add.s64 %rd117, %rd1, %rd116;
st.global.f32 [%rd117], %f48;

$L__BB273_30:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type29ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type29ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type29ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type29ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type29ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type29ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type29ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type29ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type29ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<95>;
.reg .b16 %rs<53>;
.reg .f32 %f<534>;
.reg .b32 %r<1343>;
.reg .b64 %rd<137>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type29ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1792];

ld.param.u64 %rd20, [_Z13mul_mat_vec_qIL9ggml_type29ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd21, [_Z13mul_mat_vec_qIL9ggml_type29ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd22, [_Z13mul_mat_vec_qIL9ggml_type29ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r49, [_Z13mul_mat_vec_qIL9ggml_type29ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type29ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type29ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type29ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type29ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type29ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type29ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r46, [_Z13mul_mat_vec_qIL9ggml_type29ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r47, [_Z13mul_mat_vec_qIL9ggml_type29ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r48, [_Z13mul_mat_vec_qIL9ggml_type29ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type29ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type29ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type29ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type29ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r50, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r50, %r2;
mov.u32 %r51, %ctaid.x;
shl.b32 %r4, %r51, 1;
shr.s32 %r52, %r3, 31;
shr.u32 %r53, %r52, 29;
add.s32 %r54, %r3, %r53;
shr.s32 %r1342, %r54, 3;
shr.u32 %r6, %r49, 8;
setp.ge.s32 %p3, %r1342, %r6;
mov.f32 %f520, 0f00000000;
mov.f32 %f521, %f520;
mov.f32 %f522, %f520;
mov.f32 %f523, %f520;
mov.f32 %f524, %f520;
mov.f32 %f525, %f520;
mov.f32 %f526, %f520;
mov.f32 %f527, %f520;
mov.f32 %f528, %f520;
mov.f32 %f529, %f520;
mov.f32 %f530, %f520;
mov.f32 %f531, %f520;
mov.f32 %f532, %f520;
mov.f32 %f533, %f520;
@%p3 bra $L__BB274_3;

mov.u32 %r55, %ctaid.y;
mov.u32 %r56, %ctaid.z;
mul.lo.s32 %r57, %r55, %r41;
cvt.u64.u32 %rd23, %r57;
mul.lo.s32 %r58, %r56, %r47;
cvt.u64.u32 %rd24, %r58;
add.s64 %rd1, %rd24, %rd23;
mul.hi.u32 %r59, %r56, %r43;
add.s32 %r60, %r59, %r56;
shr.u32 %r61, %r60, %r44;
mul.lo.s32 %r62, %r61, %r46;
mul.hi.u32 %r63, %r55, %r37;
add.s32 %r64, %r63, %r55;
shr.u32 %r65, %r64, %r38;
mul.lo.s32 %r66, %r65, %r40;
cvta.to.global.u64 %rd136, %rd20;
add.s32 %r67, %r4, 1;
and.b32 %r71, %r54, -8;
sub.s32 %r72, %r3, %r71;
shl.b32 %r73, %r72, 1;
cvt.s64.s32 %rd25, %r73;
cvt.s64.s32 %rd3, %r72;
shr.u32 %r74, %r72, 31;
add.s32 %r75, %r72, %r74;
shr.s32 %r76, %r75, 1;
and.b32 %r77, %r75, -2;
sub.s32 %r78, %r72, %r77;
mul.lo.s32 %r7, %r78, 6;
shl.b32 %r79, %r35, 1;
shl.b32 %r1335, %r1342, 3;
add.s32 %r1341, %r1335, %r79;
mad.lo.s32 %r1340, %r35, 3, %r1335;
shl.b32 %r80, %r35, 2;
add.s32 %r1339, %r1335, %r80;
mad.lo.s32 %r1338, %r35, 5, %r1335;
mad.lo.s32 %r1337, %r35, 6, %r1335;
add.s32 %r1336, %r35, %r1335;
add.s32 %r81, %r1342, %r62;
add.s32 %r82, %r81, %r66;
mad.lo.s32 %r83, %r67, %r34, %r82;
mul.wide.s32 %rd4, %r83, 56;
mad.lo.s32 %r84, %r4, %r34, %r1342;
add.s32 %r85, %r84, %r62;
add.s32 %r86, %r85, %r66;
mul.wide.s32 %rd5, %r86, 56;
mul.wide.s32 %rd26, %r72, 4;
add.s64 %rd6, %rd5, %rd26;
mul.wide.s32 %rd27, %r76, 2;
add.s64 %rd28, %rd5, %rd27;
add.s64 %rd7, %rd28, 48;
add.s64 %rd29, %rd4, %rd27;
add.s64 %rd8, %rd29, 48;
add.s64 %rd30, %rd5, %rd25;
add.s64 %rd9, %rd30, 33;
add.s64 %rd31, %rd4, %rd25;
add.s64 %rd10, %rd31, 32;
add.s64 %rd11, %rd4, %rd26;
cvta.to.global.u64 %rd12, %rd21;
mov.f32 %f520, 0f00000000;

$L__BB274_2:
cvt.u64.u32 %rd32, %r1335;
add.s64 %rd33, %rd1, %rd32;
add.s64 %rd34, %rd33, %rd3;
mul.lo.s64 %rd35, %rd34, 36;
add.s64 %rd36, %rd136, %rd6;
ld.global.nc.u32 %r997, [%rd36];
shr.u32 %r998, %r997, 8;
shr.u32 %r999, %r997, 16;
shr.u32 %r1000, %r997, 24;
add.s64 %rd37, %rd136, %rd9;
ld.global.nc.u8 %rs15, [%rd37+-1];
and.b16 %rs16, %rs15, 240;
cvt.u32.u16 %r1001, %rs15;
and.b32 %r1002, %r997, 255;
and.b32 %r1003, %r1001, 7;
bfi.b32 %r1004, %r1003, %r1002, 8, 3;
mul.wide.u32 %rd38, %r1004, 4;
mov.u64 %rd39, iq1s_grid_gpu;
add.s64 %rd40, %rd39, %rd38;
ld.global.nc.u32 %r1005, [%rd40];
and.b32 %r868, %r1005, 252645135;
shr.u32 %r1006, %r1005, 4;
and.b32 %r872, %r1006, 252645135;
add.s64 %rd41, %rd12, %rd35;
ld.global.nc.u32 %r89, [%rd41+4];
ld.global.nc.u32 %r93, [%rd41+8];
mov.u32 %r991, 0;

dp4a.s32.s32 %r87, %r868, %r89, %r991;


dp4a.s32.s32 %r91, %r872, %r93, %r87;

and.b32 %r1007, %r1001, 8;
cvt.rn.f32.s32 %f113, %r1007;
mul.ftz.f32 %f114, %f113, 0f3D000000;
mov.f32 %f115, 0fBF600000;
sub.ftz.f32 %f116, %f115, %f114;
mov.u32 %r994, 16843009;

dp4a.s32.s32 %r95, %r89, %r994, %r991;


dp4a.s32.s32 %r99, %r93, %r994, %r95;

cvt.rn.f32.s32 %f117, %r99;
fma.rn.ftz.f32 %f118, %f116, %f117, 0f00000000;
shr.u16 %rs17, %rs16, 4;
cvt.u32.u16 %r1008, %rs17;
and.b32 %r1009, %r998, 255;
and.b32 %r1010, %r1008, 7;
bfi.b32 %r1011, %r1010, %r1009, 8, 3;
mul.wide.u32 %rd42, %r1011, 4;
add.s64 %rd43, %rd39, %rd42;
ld.global.nc.u32 %r1012, [%rd43];
and.b32 %r884, %r1012, 252645135;
shr.u32 %r1013, %r1012, 4;
and.b32 %r888, %r1013, 252645135;
ld.global.nc.u32 %r105, [%rd41+12];
ld.global.nc.u32 %r109, [%rd41+16];

dp4a.s32.s32 %r103, %r884, %r105, %r91;


dp4a.s32.s32 %r107, %r888, %r109, %r103;

and.b32 %r1014, %r1008, 8;
cvt.rn.f32.s32 %f119, %r1014;
mul.ftz.f32 %f120, %f119, 0f3D000000;
sub.ftz.f32 %f121, %f115, %f120;

dp4a.s32.s32 %r111, %r105, %r994, %r991;


dp4a.s32.s32 %r115, %r109, %r994, %r111;

cvt.rn.f32.s32 %f122, %r115;
fma.rn.ftz.f32 %f123, %f121, %f122, %f118;
ld.global.nc.u8 %rs18, [%rd37];
and.b16 %rs19, %rs18, 240;
cvt.u32.u16 %r1015, %rs18;
and.b32 %r1016, %r999, 255;
and.b32 %r1017, %r1015, 7;
bfi.b32 %r1018, %r1017, %r1016, 8, 3;
mul.wide.u32 %rd44, %r1018, 4;
add.s64 %rd45, %rd39, %rd44;
ld.global.nc.u32 %r1019, [%rd45];
and.b32 %r900, %r1019, 252645135;
shr.u32 %r1020, %r1019, 4;
and.b32 %r904, %r1020, 252645135;
ld.global.nc.u32 %r121, [%rd41+20];
ld.global.nc.u32 %r125, [%rd41+24];

dp4a.s32.s32 %r119, %r900, %r121, %r991;


dp4a.s32.s32 %r123, %r904, %r125, %r119;

and.b32 %r1021, %r1015, 8;
cvt.rn.f32.s32 %f124, %r1021;
mul.ftz.f32 %f125, %f124, 0f3D000000;
sub.ftz.f32 %f126, %f115, %f125;

dp4a.s32.s32 %r127, %r121, %r994, %r991;


dp4a.s32.s32 %r131, %r125, %r994, %r127;

cvt.rn.f32.s32 %f127, %r131;
fma.rn.ftz.f32 %f128, %f126, %f127, 0f00000000;
shr.u16 %rs20, %rs19, 4;
cvt.u32.u16 %r1022, %rs20;
mul.wide.u16 %r1023, %rs20, 256;
and.b32 %r1024, %r1023, 1792;
or.b32 %r1025, %r1024, %r1000;
mul.wide.u32 %rd46, %r1025, 4;
add.s64 %rd47, %rd39, %rd46;
ld.global.nc.u32 %r1026, [%rd47];
and.b32 %r916, %r1026, 252645135;
shr.u32 %r1027, %r1026, 4;
and.b32 %r920, %r1027, 252645135;
ld.global.nc.u32 %r137, [%rd41+28];
ld.global.nc.u32 %r141, [%rd41+32];

dp4a.s32.s32 %r135, %r916, %r137, %r123;


dp4a.s32.s32 %r139, %r920, %r141, %r135;

and.b32 %r1028, %r1022, 8;
cvt.rn.f32.s32 %f129, %r1028;
mul.ftz.f32 %f130, %f129, 0f3D000000;
sub.ftz.f32 %f131, %f115, %f130;

dp4a.s32.s32 %r143, %r137, %r994, %r991;


dp4a.s32.s32 %r147, %r141, %r994, %r143;

cvt.rn.f32.s32 %f132, %r147;
fma.rn.ftz.f32 %f133, %f131, %f132, %f128;
add.s64 %rd48, %rd136, %rd5;
ld.global.nc.u16 %rs21, [%rd48+48];
shr.u16 %rs22, %rs21, 12;
ld.global.nc.u16 %rs23, [%rd48+50];
shr.u16 %rs24, %rs23, 8;
and.b16 %rs25, %rs24, 240;
or.b16 %rs26, %rs25, %rs22;
ld.global.nc.u16 %rs27, [%rd48+52];
shr.u16 %rs28, %rs27, 4;
and.b16 %rs29, %rs28, 3840;
or.b16 %rs30, %rs26, %rs29;
ld.global.nc.u16 %rs31, [%rd48+54];
and.b16 %rs32, %rs31, -4096;
or.b16 %rs13, %rs30, %rs32;

{ cvt.f32.f16 %f85, %rs13;}


ld.global.nc.u32 %r151, [%rd41];

{.reg .f16 low,high;
mov.b32 {low,high},%r151;
cvt.f32.f16 %f86, low;}


mul.ftz.f32 %f134, %f85, %f86;
add.s64 %rd49, %rd136, %rd7;
ld.global.nc.u16 %rs33, [%rd49];
cvt.u32.u16 %r1029, %rs33;
shr.u32 %r1030, %r1029, %r7;
shl.b32 %r1031, %r1030, 1;
and.b32 %r1032, %r1031, 14;
or.b32 %r1033, %r1032, 1;
shr.u32 %r1034, %r1030, 2;
and.b32 %r1035, %r1034, 14;
or.b32 %r1036, %r1035, 1;
cvt.rn.f32.s32 %f135, %r107;
add.ftz.f32 %f136, %f123, %f135;
cvt.rn.f32.s32 %f137, %r1033;
cvt.rn.f32.s32 %f138, %r139;
add.ftz.f32 %f139, %f133, %f138;
cvt.rn.f32.s32 %f140, %r1036;
mul.ftz.f32 %f141, %f139, %f140;
fma.rn.ftz.f32 %f142, %f136, %f137, %f141;
fma.rn.ftz.f32 %f533, %f134, %f142, %f533;
add.s64 %rd50, %rd136, %rd11;
ld.global.nc.u32 %r1037, [%rd50];
shr.u32 %r1038, %r1037, 8;
shr.u32 %r1039, %r1037, 16;
shr.u32 %r1040, %r1037, 24;
add.s64 %rd51, %rd136, %rd10;
ld.global.nc.u8 %rs34, [%rd51];
and.b16 %rs35, %rs34, 240;
cvt.u32.u16 %r1041, %rs34;
and.b32 %r1042, %r1037, 255;
and.b32 %r1043, %r1041, 7;
bfi.b32 %r1044, %r1043, %r1042, 8, 3;
mul.wide.u32 %rd52, %r1044, 4;
add.s64 %rd53, %rd39, %rd52;
ld.global.nc.u32 %r1045, [%rd53];
and.b32 %r933, %r1045, 252645135;
shr.u32 %r1046, %r1045, 4;
and.b32 %r937, %r1046, 252645135;

dp4a.s32.s32 %r152, %r933, %r89, %r991;


dp4a.s32.s32 %r156, %r937, %r93, %r152;

and.b32 %r1047, %r1041, 8;
cvt.rn.f32.s32 %f143, %r1047;
mul.ftz.f32 %f144, %f143, 0f3D000000;
sub.ftz.f32 %f145, %f115, %f144;

dp4a.s32.s32 %r160, %r89, %r994, %r991;


dp4a.s32.s32 %r164, %r93, %r994, %r160;

cvt.rn.f32.s32 %f146, %r164;
fma.rn.ftz.f32 %f147, %f145, %f146, 0f00000000;
shr.u16 %rs36, %rs35, 4;
cvt.u32.u16 %r1048, %rs36;
and.b32 %r1049, %r1038, 255;
and.b32 %r1050, %r1048, 7;
bfi.b32 %r1051, %r1050, %r1049, 8, 3;
mul.wide.u32 %rd54, %r1051, 4;
add.s64 %rd55, %rd39, %rd54;
ld.global.nc.u32 %r1052, [%rd55];
and.b32 %r949, %r1052, 252645135;
shr.u32 %r1053, %r1052, 4;
and.b32 %r953, %r1053, 252645135;

dp4a.s32.s32 %r168, %r949, %r105, %r156;


dp4a.s32.s32 %r172, %r953, %r109, %r168;

and.b32 %r1054, %r1048, 8;
cvt.rn.f32.s32 %f148, %r1054;
mul.ftz.f32 %f149, %f148, 0f3D000000;
sub.ftz.f32 %f150, %f115, %f149;

dp4a.s32.s32 %r176, %r105, %r994, %r991;


dp4a.s32.s32 %r180, %r109, %r994, %r176;

cvt.rn.f32.s32 %f151, %r180;
fma.rn.ftz.f32 %f152, %f150, %f151, %f147;
ld.global.nc.u8 %rs37, [%rd51+1];
and.b16 %rs38, %rs37, 240;
cvt.u32.u16 %r1055, %rs37;
and.b32 %r1056, %r1039, 255;
and.b32 %r1057, %r1055, 7;
bfi.b32 %r1058, %r1057, %r1056, 8, 3;
mul.wide.u32 %rd56, %r1058, 4;
add.s64 %rd57, %rd39, %rd56;
ld.global.nc.u32 %r1059, [%rd57];
and.b32 %r965, %r1059, 252645135;
shr.u32 %r1060, %r1059, 4;
and.b32 %r969, %r1060, 252645135;

dp4a.s32.s32 %r184, %r965, %r121, %r991;


dp4a.s32.s32 %r188, %r969, %r125, %r184;

and.b32 %r1061, %r1055, 8;
cvt.rn.f32.s32 %f153, %r1061;
mul.ftz.f32 %f154, %f153, 0f3D000000;
sub.ftz.f32 %f155, %f115, %f154;

dp4a.s32.s32 %r192, %r121, %r994, %r991;


dp4a.s32.s32 %r196, %r125, %r994, %r192;

cvt.rn.f32.s32 %f156, %r196;
fma.rn.ftz.f32 %f157, %f155, %f156, 0f00000000;
shr.u16 %rs39, %rs38, 4;
cvt.u32.u16 %r1062, %rs39;
mul.wide.u16 %r1063, %rs39, 256;
and.b32 %r1064, %r1063, 1792;
or.b32 %r1065, %r1064, %r1040;
mul.wide.u32 %rd58, %r1065, 4;
add.s64 %rd59, %rd39, %rd58;
ld.global.nc.u32 %r1066, [%rd59];
and.b32 %r981, %r1066, 252645135;
shr.u32 %r1067, %r1066, 4;
and.b32 %r985, %r1067, 252645135;

dp4a.s32.s32 %r200, %r981, %r137, %r188;


dp4a.s32.s32 %r204, %r985, %r141, %r200;

and.b32 %r1068, %r1062, 8;
cvt.rn.f32.s32 %f158, %r1068;
mul.ftz.f32 %f159, %f158, 0f3D000000;
sub.ftz.f32 %f160, %f115, %f159;

dp4a.s32.s32 %r208, %r137, %r994, %r991;


dp4a.s32.s32 %r212, %r141, %r994, %r208;

cvt.rn.f32.s32 %f161, %r212;
fma.rn.ftz.f32 %f162, %f160, %f161, %f157;
add.s64 %rd60, %rd136, %rd4;
ld.global.nc.u16 %rs40, [%rd60+48];
shr.u16 %rs41, %rs40, 12;
ld.global.nc.u16 %rs42, [%rd60+50];
shr.u16 %rs43, %rs42, 8;
and.b16 %rs44, %rs43, 240;
or.b16 %rs45, %rs44, %rs41;
ld.global.nc.u16 %rs46, [%rd60+52];
shr.u16 %rs47, %rs46, 4;
and.b16 %rs48, %rs47, 3840;
or.b16 %rs49, %rs45, %rs48;
ld.global.nc.u16 %rs50, [%rd60+54];
and.b16 %rs51, %rs50, -4096;
or.b16 %rs14, %rs49, %rs51;

{ cvt.f32.f16 %f87, %rs14;}



{.reg .f16 low,high;
mov.b32 {low,high},%r151;
cvt.f32.f16 %f88, low;}


mul.ftz.f32 %f163, %f87, %f88;
add.s64 %rd61, %rd136, %rd8;
ld.global.nc.u16 %rs52, [%rd61];
cvt.u32.u16 %r1069, %rs52;
shr.u32 %r1070, %r1069, %r7;
shl.b32 %r1071, %r1070, 1;
and.b32 %r1072, %r1071, 14;
or.b32 %r1073, %r1072, 1;
shr.u32 %r1074, %r1070, 2;
and.b32 %r1075, %r1074, 14;
or.b32 %r1076, %r1075, 1;
cvt.rn.f32.s32 %f164, %r172;
add.ftz.f32 %f165, %f152, %f164;
cvt.rn.f32.s32 %f166, %r1073;
cvt.rn.f32.s32 %f167, %r204;
add.ftz.f32 %f168, %f162, %f167;
cvt.rn.f32.s32 %f169, %r1076;
mul.ftz.f32 %f170, %f168, %f169;
fma.rn.ftz.f32 %f171, %f165, %f166, %f170;
fma.rn.ftz.f32 %f532, %f163, %f171, %f532;
cvt.u64.u32 %rd62, %r1336;
add.s64 %rd63, %rd1, %rd62;
add.s64 %rd64, %rd63, %rd3;
mul.lo.s64 %rd65, %rd64, 36;
add.s64 %rd66, %rd12, %rd65;
ld.global.nc.u32 %r219, [%rd66+4];
ld.global.nc.u32 %r223, [%rd66+8];

dp4a.s32.s32 %r217, %r868, %r219, %r991;


dp4a.s32.s32 %r221, %r872, %r223, %r217;


dp4a.s32.s32 %r225, %r219, %r994, %r991;


dp4a.s32.s32 %r229, %r223, %r994, %r225;

cvt.rn.f32.s32 %f172, %r229;
fma.rn.ftz.f32 %f173, %f116, %f172, 0f00000000;
ld.global.nc.u32 %r235, [%rd66+12];
ld.global.nc.u32 %r239, [%rd66+16];

dp4a.s32.s32 %r233, %r884, %r235, %r221;


dp4a.s32.s32 %r237, %r888, %r239, %r233;


dp4a.s32.s32 %r241, %r235, %r994, %r991;


dp4a.s32.s32 %r245, %r239, %r994, %r241;

cvt.rn.f32.s32 %f174, %r245;
fma.rn.ftz.f32 %f175, %f121, %f174, %f173;
ld.global.nc.u32 %r251, [%rd66+20];
ld.global.nc.u32 %r255, [%rd66+24];

dp4a.s32.s32 %r249, %r900, %r251, %r991;


dp4a.s32.s32 %r253, %r904, %r255, %r249;


dp4a.s32.s32 %r257, %r251, %r994, %r991;


dp4a.s32.s32 %r261, %r255, %r994, %r257;

cvt.rn.f32.s32 %f176, %r261;
fma.rn.ftz.f32 %f177, %f126, %f176, 0f00000000;
ld.global.nc.u32 %r267, [%rd66+28];
ld.global.nc.u32 %r271, [%rd66+32];

dp4a.s32.s32 %r265, %r916, %r267, %r253;


dp4a.s32.s32 %r269, %r920, %r271, %r265;


dp4a.s32.s32 %r273, %r267, %r994, %r991;


dp4a.s32.s32 %r277, %r271, %r994, %r273;

cvt.rn.f32.s32 %f178, %r277;
fma.rn.ftz.f32 %f179, %f131, %f178, %f177;

{ cvt.f32.f16 %f89, %rs13;}


ld.global.nc.u32 %r281, [%rd66];

{.reg .f16 low,high;
mov.b32 {low,high},%r281;
cvt.f32.f16 %f90, low;}


mul.ftz.f32 %f180, %f89, %f90;
cvt.rn.f32.s32 %f181, %r237;
add.ftz.f32 %f182, %f175, %f181;
cvt.rn.f32.s32 %f183, %r269;
add.ftz.f32 %f184, %f179, %f183;
mul.ftz.f32 %f185, %f184, %f140;
fma.rn.ftz.f32 %f186, %f182, %f137, %f185;
fma.rn.ftz.f32 %f531, %f180, %f186, %f531;

dp4a.s32.s32 %r282, %r933, %r219, %r991;


dp4a.s32.s32 %r286, %r937, %r223, %r282;


dp4a.s32.s32 %r290, %r219, %r994, %r991;


dp4a.s32.s32 %r294, %r223, %r994, %r290;

cvt.rn.f32.s32 %f187, %r294;
fma.rn.ftz.f32 %f188, %f145, %f187, 0f00000000;

dp4a.s32.s32 %r298, %r949, %r235, %r286;


dp4a.s32.s32 %r302, %r953, %r239, %r298;


dp4a.s32.s32 %r306, %r235, %r994, %r991;


dp4a.s32.s32 %r310, %r239, %r994, %r306;

cvt.rn.f32.s32 %f189, %r310;
fma.rn.ftz.f32 %f190, %f150, %f189, %f188;

dp4a.s32.s32 %r314, %r965, %r251, %r991;


dp4a.s32.s32 %r318, %r969, %r255, %r314;


dp4a.s32.s32 %r322, %r251, %r994, %r991;


dp4a.s32.s32 %r326, %r255, %r994, %r322;

cvt.rn.f32.s32 %f191, %r326;
fma.rn.ftz.f32 %f192, %f155, %f191, 0f00000000;

dp4a.s32.s32 %r330, %r981, %r267, %r318;


dp4a.s32.s32 %r334, %r985, %r271, %r330;


dp4a.s32.s32 %r338, %r267, %r994, %r991;


dp4a.s32.s32 %r342, %r271, %r994, %r338;

cvt.rn.f32.s32 %f193, %r342;
fma.rn.ftz.f32 %f194, %f160, %f193, %f192;

{ cvt.f32.f16 %f91, %rs14;}



{.reg .f16 low,high;
mov.b32 {low,high},%r281;
cvt.f32.f16 %f92, low;}


mul.ftz.f32 %f195, %f91, %f92;
cvt.rn.f32.s32 %f196, %r302;
add.ftz.f32 %f197, %f190, %f196;
cvt.rn.f32.s32 %f198, %r334;
add.ftz.f32 %f199, %f194, %f198;
mul.ftz.f32 %f200, %f199, %f169;
fma.rn.ftz.f32 %f201, %f197, %f166, %f200;
fma.rn.ftz.f32 %f530, %f195, %f201, %f530;
cvt.u64.u32 %rd67, %r1341;
add.s64 %rd68, %rd1, %rd67;
add.s64 %rd69, %rd68, %rd3;
mul.lo.s64 %rd70, %rd69, 36;
add.s64 %rd71, %rd12, %rd70;
ld.global.nc.u32 %r349, [%rd71+4];
ld.global.nc.u32 %r353, [%rd71+8];

dp4a.s32.s32 %r347, %r868, %r349, %r991;


dp4a.s32.s32 %r351, %r872, %r353, %r347;


dp4a.s32.s32 %r355, %r349, %r994, %r991;


dp4a.s32.s32 %r359, %r353, %r994, %r355;

cvt.rn.f32.s32 %f202, %r359;
fma.rn.ftz.f32 %f203, %f116, %f202, 0f00000000;
ld.global.nc.u32 %r365, [%rd71+12];
ld.global.nc.u32 %r369, [%rd71+16];

dp4a.s32.s32 %r363, %r884, %r365, %r351;


dp4a.s32.s32 %r367, %r888, %r369, %r363;


dp4a.s32.s32 %r371, %r365, %r994, %r991;


dp4a.s32.s32 %r375, %r369, %r994, %r371;

cvt.rn.f32.s32 %f204, %r375;
fma.rn.ftz.f32 %f205, %f121, %f204, %f203;
ld.global.nc.u32 %r381, [%rd71+20];
ld.global.nc.u32 %r385, [%rd71+24];

dp4a.s32.s32 %r379, %r900, %r381, %r991;


dp4a.s32.s32 %r383, %r904, %r385, %r379;


dp4a.s32.s32 %r387, %r381, %r994, %r991;


dp4a.s32.s32 %r391, %r385, %r994, %r387;

cvt.rn.f32.s32 %f206, %r391;
fma.rn.ftz.f32 %f207, %f126, %f206, 0f00000000;
ld.global.nc.u32 %r397, [%rd71+28];
ld.global.nc.u32 %r401, [%rd71+32];

dp4a.s32.s32 %r395, %r916, %r397, %r383;


dp4a.s32.s32 %r399, %r920, %r401, %r395;


dp4a.s32.s32 %r403, %r397, %r994, %r991;


dp4a.s32.s32 %r407, %r401, %r994, %r403;

cvt.rn.f32.s32 %f208, %r407;
fma.rn.ftz.f32 %f209, %f131, %f208, %f207;

{ cvt.f32.f16 %f93, %rs13;}


ld.global.nc.u32 %r411, [%rd71];

{.reg .f16 low,high;
mov.b32 {low,high},%r411;
cvt.f32.f16 %f94, low;}


mul.ftz.f32 %f210, %f93, %f94;
cvt.rn.f32.s32 %f211, %r367;
add.ftz.f32 %f212, %f205, %f211;
cvt.rn.f32.s32 %f213, %r399;
add.ftz.f32 %f214, %f209, %f213;
mul.ftz.f32 %f215, %f214, %f140;
fma.rn.ftz.f32 %f216, %f212, %f137, %f215;
fma.rn.ftz.f32 %f529, %f210, %f216, %f529;

dp4a.s32.s32 %r412, %r933, %r349, %r991;


dp4a.s32.s32 %r416, %r937, %r353, %r412;


dp4a.s32.s32 %r420, %r349, %r994, %r991;


dp4a.s32.s32 %r424, %r353, %r994, %r420;

cvt.rn.f32.s32 %f217, %r424;
fma.rn.ftz.f32 %f218, %f145, %f217, 0f00000000;

dp4a.s32.s32 %r428, %r949, %r365, %r416;


dp4a.s32.s32 %r432, %r953, %r369, %r428;


dp4a.s32.s32 %r436, %r365, %r994, %r991;


dp4a.s32.s32 %r440, %r369, %r994, %r436;

cvt.rn.f32.s32 %f219, %r440;
fma.rn.ftz.f32 %f220, %f150, %f219, %f218;

dp4a.s32.s32 %r444, %r965, %r381, %r991;


dp4a.s32.s32 %r448, %r969, %r385, %r444;


dp4a.s32.s32 %r452, %r381, %r994, %r991;


dp4a.s32.s32 %r456, %r385, %r994, %r452;

cvt.rn.f32.s32 %f221, %r456;
fma.rn.ftz.f32 %f222, %f155, %f221, 0f00000000;

dp4a.s32.s32 %r460, %r981, %r397, %r448;


dp4a.s32.s32 %r464, %r985, %r401, %r460;


dp4a.s32.s32 %r468, %r397, %r994, %r991;


dp4a.s32.s32 %r472, %r401, %r994, %r468;

cvt.rn.f32.s32 %f223, %r472;
fma.rn.ftz.f32 %f224, %f160, %f223, %f222;

{ cvt.f32.f16 %f95, %rs14;}



{.reg .f16 low,high;
mov.b32 {low,high},%r411;
cvt.f32.f16 %f96, low;}


mul.ftz.f32 %f225, %f95, %f96;
cvt.rn.f32.s32 %f226, %r432;
add.ftz.f32 %f227, %f220, %f226;
cvt.rn.f32.s32 %f228, %r464;
add.ftz.f32 %f229, %f224, %f228;
mul.ftz.f32 %f230, %f229, %f169;
fma.rn.ftz.f32 %f231, %f227, %f166, %f230;
fma.rn.ftz.f32 %f528, %f225, %f231, %f528;
cvt.u64.u32 %rd72, %r1340;
add.s64 %rd73, %rd1, %rd72;
add.s64 %rd74, %rd73, %rd3;
mul.lo.s64 %rd75, %rd74, 36;
add.s64 %rd76, %rd12, %rd75;
ld.global.nc.u32 %r479, [%rd76+4];
ld.global.nc.u32 %r483, [%rd76+8];

dp4a.s32.s32 %r477, %r868, %r479, %r991;


dp4a.s32.s32 %r481, %r872, %r483, %r477;


dp4a.s32.s32 %r485, %r479, %r994, %r991;


dp4a.s32.s32 %r489, %r483, %r994, %r485;

cvt.rn.f32.s32 %f232, %r489;
fma.rn.ftz.f32 %f233, %f116, %f232, 0f00000000;
ld.global.nc.u32 %r495, [%rd76+12];
ld.global.nc.u32 %r499, [%rd76+16];

dp4a.s32.s32 %r493, %r884, %r495, %r481;


dp4a.s32.s32 %r497, %r888, %r499, %r493;


dp4a.s32.s32 %r501, %r495, %r994, %r991;


dp4a.s32.s32 %r505, %r499, %r994, %r501;

cvt.rn.f32.s32 %f234, %r505;
fma.rn.ftz.f32 %f235, %f121, %f234, %f233;
ld.global.nc.u32 %r511, [%rd76+20];
ld.global.nc.u32 %r515, [%rd76+24];

dp4a.s32.s32 %r509, %r900, %r511, %r991;


dp4a.s32.s32 %r513, %r904, %r515, %r509;


dp4a.s32.s32 %r517, %r511, %r994, %r991;


dp4a.s32.s32 %r521, %r515, %r994, %r517;

cvt.rn.f32.s32 %f236, %r521;
fma.rn.ftz.f32 %f237, %f126, %f236, 0f00000000;
ld.global.nc.u32 %r527, [%rd76+28];
ld.global.nc.u32 %r531, [%rd76+32];

dp4a.s32.s32 %r525, %r916, %r527, %r513;


dp4a.s32.s32 %r529, %r920, %r531, %r525;


dp4a.s32.s32 %r533, %r527, %r994, %r991;


dp4a.s32.s32 %r537, %r531, %r994, %r533;

cvt.rn.f32.s32 %f238, %r537;
fma.rn.ftz.f32 %f239, %f131, %f238, %f237;

{ cvt.f32.f16 %f97, %rs13;}


ld.global.nc.u32 %r541, [%rd76];

{.reg .f16 low,high;
mov.b32 {low,high},%r541;
cvt.f32.f16 %f98, low;}


mul.ftz.f32 %f240, %f97, %f98;
cvt.rn.f32.s32 %f241, %r497;
add.ftz.f32 %f242, %f235, %f241;
cvt.rn.f32.s32 %f243, %r529;
add.ftz.f32 %f244, %f239, %f243;
mul.ftz.f32 %f245, %f244, %f140;
fma.rn.ftz.f32 %f246, %f242, %f137, %f245;
fma.rn.ftz.f32 %f527, %f240, %f246, %f527;

dp4a.s32.s32 %r542, %r933, %r479, %r991;


dp4a.s32.s32 %r546, %r937, %r483, %r542;


dp4a.s32.s32 %r550, %r479, %r994, %r991;


dp4a.s32.s32 %r554, %r483, %r994, %r550;

cvt.rn.f32.s32 %f247, %r554;
fma.rn.ftz.f32 %f248, %f145, %f247, 0f00000000;

dp4a.s32.s32 %r558, %r949, %r495, %r546;


dp4a.s32.s32 %r562, %r953, %r499, %r558;


dp4a.s32.s32 %r566, %r495, %r994, %r991;


dp4a.s32.s32 %r570, %r499, %r994, %r566;

cvt.rn.f32.s32 %f249, %r570;
fma.rn.ftz.f32 %f250, %f150, %f249, %f248;

dp4a.s32.s32 %r574, %r965, %r511, %r991;


dp4a.s32.s32 %r578, %r969, %r515, %r574;


dp4a.s32.s32 %r582, %r511, %r994, %r991;


dp4a.s32.s32 %r586, %r515, %r994, %r582;

cvt.rn.f32.s32 %f251, %r586;
fma.rn.ftz.f32 %f252, %f155, %f251, 0f00000000;

dp4a.s32.s32 %r590, %r981, %r527, %r578;


dp4a.s32.s32 %r594, %r985, %r531, %r590;


dp4a.s32.s32 %r598, %r527, %r994, %r991;


dp4a.s32.s32 %r602, %r531, %r994, %r598;

cvt.rn.f32.s32 %f253, %r602;
fma.rn.ftz.f32 %f254, %f160, %f253, %f252;

{ cvt.f32.f16 %f99, %rs14;}



{.reg .f16 low,high;
mov.b32 {low,high},%r541;
cvt.f32.f16 %f100, low;}


mul.ftz.f32 %f255, %f99, %f100;
cvt.rn.f32.s32 %f256, %r562;
add.ftz.f32 %f257, %f250, %f256;
cvt.rn.f32.s32 %f258, %r594;
add.ftz.f32 %f259, %f254, %f258;
mul.ftz.f32 %f260, %f259, %f169;
fma.rn.ftz.f32 %f261, %f257, %f166, %f260;
fma.rn.ftz.f32 %f526, %f255, %f261, %f526;
cvt.u64.u32 %rd77, %r1339;
add.s64 %rd78, %rd1, %rd77;
add.s64 %rd79, %rd78, %rd3;
mul.lo.s64 %rd80, %rd79, 36;
add.s64 %rd81, %rd12, %rd80;
ld.global.nc.u32 %r609, [%rd81+4];
ld.global.nc.u32 %r613, [%rd81+8];

dp4a.s32.s32 %r607, %r868, %r609, %r991;


dp4a.s32.s32 %r611, %r872, %r613, %r607;


dp4a.s32.s32 %r615, %r609, %r994, %r991;


dp4a.s32.s32 %r619, %r613, %r994, %r615;

cvt.rn.f32.s32 %f262, %r619;
fma.rn.ftz.f32 %f263, %f116, %f262, 0f00000000;
ld.global.nc.u32 %r625, [%rd81+12];
ld.global.nc.u32 %r629, [%rd81+16];

dp4a.s32.s32 %r623, %r884, %r625, %r611;


dp4a.s32.s32 %r627, %r888, %r629, %r623;


dp4a.s32.s32 %r631, %r625, %r994, %r991;


dp4a.s32.s32 %r635, %r629, %r994, %r631;

cvt.rn.f32.s32 %f264, %r635;
fma.rn.ftz.f32 %f265, %f121, %f264, %f263;
ld.global.nc.u32 %r641, [%rd81+20];
ld.global.nc.u32 %r645, [%rd81+24];

dp4a.s32.s32 %r639, %r900, %r641, %r991;


dp4a.s32.s32 %r643, %r904, %r645, %r639;


dp4a.s32.s32 %r647, %r641, %r994, %r991;


dp4a.s32.s32 %r651, %r645, %r994, %r647;

cvt.rn.f32.s32 %f266, %r651;
fma.rn.ftz.f32 %f267, %f126, %f266, 0f00000000;
ld.global.nc.u32 %r657, [%rd81+28];
ld.global.nc.u32 %r661, [%rd81+32];

dp4a.s32.s32 %r655, %r916, %r657, %r643;


dp4a.s32.s32 %r659, %r920, %r661, %r655;


dp4a.s32.s32 %r663, %r657, %r994, %r991;


dp4a.s32.s32 %r667, %r661, %r994, %r663;

cvt.rn.f32.s32 %f268, %r667;
fma.rn.ftz.f32 %f269, %f131, %f268, %f267;

{ cvt.f32.f16 %f101, %rs13;}


ld.global.nc.u32 %r671, [%rd81];

{.reg .f16 low,high;
mov.b32 {low,high},%r671;
cvt.f32.f16 %f102, low;}


mul.ftz.f32 %f270, %f101, %f102;
cvt.rn.f32.s32 %f271, %r627;
add.ftz.f32 %f272, %f265, %f271;
cvt.rn.f32.s32 %f273, %r659;
add.ftz.f32 %f274, %f269, %f273;
mul.ftz.f32 %f275, %f274, %f140;
fma.rn.ftz.f32 %f276, %f272, %f137, %f275;
fma.rn.ftz.f32 %f525, %f270, %f276, %f525;

dp4a.s32.s32 %r672, %r933, %r609, %r991;


dp4a.s32.s32 %r676, %r937, %r613, %r672;


dp4a.s32.s32 %r680, %r609, %r994, %r991;


dp4a.s32.s32 %r684, %r613, %r994, %r680;

cvt.rn.f32.s32 %f277, %r684;
fma.rn.ftz.f32 %f278, %f145, %f277, 0f00000000;

dp4a.s32.s32 %r688, %r949, %r625, %r676;


dp4a.s32.s32 %r692, %r953, %r629, %r688;


dp4a.s32.s32 %r696, %r625, %r994, %r991;


dp4a.s32.s32 %r700, %r629, %r994, %r696;

cvt.rn.f32.s32 %f279, %r700;
fma.rn.ftz.f32 %f280, %f150, %f279, %f278;

dp4a.s32.s32 %r704, %r965, %r641, %r991;


dp4a.s32.s32 %r708, %r969, %r645, %r704;


dp4a.s32.s32 %r712, %r641, %r994, %r991;


dp4a.s32.s32 %r716, %r645, %r994, %r712;

cvt.rn.f32.s32 %f281, %r716;
fma.rn.ftz.f32 %f282, %f155, %f281, 0f00000000;

dp4a.s32.s32 %r720, %r981, %r657, %r708;


dp4a.s32.s32 %r724, %r985, %r661, %r720;


dp4a.s32.s32 %r728, %r657, %r994, %r991;


dp4a.s32.s32 %r732, %r661, %r994, %r728;

cvt.rn.f32.s32 %f283, %r732;
fma.rn.ftz.f32 %f284, %f160, %f283, %f282;

{ cvt.f32.f16 %f103, %rs14;}



{.reg .f16 low,high;
mov.b32 {low,high},%r671;
cvt.f32.f16 %f104, low;}


mul.ftz.f32 %f285, %f103, %f104;
cvt.rn.f32.s32 %f286, %r692;
add.ftz.f32 %f287, %f280, %f286;
cvt.rn.f32.s32 %f288, %r724;
add.ftz.f32 %f289, %f284, %f288;
mul.ftz.f32 %f290, %f289, %f169;
fma.rn.ftz.f32 %f291, %f287, %f166, %f290;
fma.rn.ftz.f32 %f524, %f285, %f291, %f524;
cvt.u64.u32 %rd82, %r1338;
add.s64 %rd83, %rd1, %rd82;
add.s64 %rd84, %rd83, %rd3;
mul.lo.s64 %rd85, %rd84, 36;
add.s64 %rd86, %rd12, %rd85;
ld.global.nc.u32 %r739, [%rd86+4];
ld.global.nc.u32 %r743, [%rd86+8];

dp4a.s32.s32 %r737, %r868, %r739, %r991;


dp4a.s32.s32 %r741, %r872, %r743, %r737;


dp4a.s32.s32 %r745, %r739, %r994, %r991;


dp4a.s32.s32 %r749, %r743, %r994, %r745;

cvt.rn.f32.s32 %f292, %r749;
fma.rn.ftz.f32 %f293, %f116, %f292, 0f00000000;
ld.global.nc.u32 %r755, [%rd86+12];
ld.global.nc.u32 %r759, [%rd86+16];

dp4a.s32.s32 %r753, %r884, %r755, %r741;


dp4a.s32.s32 %r757, %r888, %r759, %r753;


dp4a.s32.s32 %r761, %r755, %r994, %r991;


dp4a.s32.s32 %r765, %r759, %r994, %r761;

cvt.rn.f32.s32 %f294, %r765;
fma.rn.ftz.f32 %f295, %f121, %f294, %f293;
ld.global.nc.u32 %r771, [%rd86+20];
ld.global.nc.u32 %r775, [%rd86+24];

dp4a.s32.s32 %r769, %r900, %r771, %r991;


dp4a.s32.s32 %r773, %r904, %r775, %r769;


dp4a.s32.s32 %r777, %r771, %r994, %r991;


dp4a.s32.s32 %r781, %r775, %r994, %r777;

cvt.rn.f32.s32 %f296, %r781;
fma.rn.ftz.f32 %f297, %f126, %f296, 0f00000000;
ld.global.nc.u32 %r787, [%rd86+28];
ld.global.nc.u32 %r791, [%rd86+32];

dp4a.s32.s32 %r785, %r916, %r787, %r773;


dp4a.s32.s32 %r789, %r920, %r791, %r785;


dp4a.s32.s32 %r793, %r787, %r994, %r991;


dp4a.s32.s32 %r797, %r791, %r994, %r793;

cvt.rn.f32.s32 %f298, %r797;
fma.rn.ftz.f32 %f299, %f131, %f298, %f297;

{ cvt.f32.f16 %f105, %rs13;}


ld.global.nc.u32 %r801, [%rd86];

{.reg .f16 low,high;
mov.b32 {low,high},%r801;
cvt.f32.f16 %f106, low;}


mul.ftz.f32 %f300, %f105, %f106;
cvt.rn.f32.s32 %f301, %r757;
add.ftz.f32 %f302, %f295, %f301;
cvt.rn.f32.s32 %f303, %r789;
add.ftz.f32 %f304, %f299, %f303;
mul.ftz.f32 %f305, %f304, %f140;
fma.rn.ftz.f32 %f306, %f302, %f137, %f305;
fma.rn.ftz.f32 %f523, %f300, %f306, %f523;

dp4a.s32.s32 %r802, %r933, %r739, %r991;


dp4a.s32.s32 %r806, %r937, %r743, %r802;


dp4a.s32.s32 %r810, %r739, %r994, %r991;


dp4a.s32.s32 %r814, %r743, %r994, %r810;

cvt.rn.f32.s32 %f307, %r814;
fma.rn.ftz.f32 %f308, %f145, %f307, 0f00000000;

dp4a.s32.s32 %r818, %r949, %r755, %r806;


dp4a.s32.s32 %r822, %r953, %r759, %r818;


dp4a.s32.s32 %r826, %r755, %r994, %r991;


dp4a.s32.s32 %r830, %r759, %r994, %r826;

cvt.rn.f32.s32 %f309, %r830;
fma.rn.ftz.f32 %f310, %f150, %f309, %f308;

dp4a.s32.s32 %r834, %r965, %r771, %r991;


dp4a.s32.s32 %r838, %r969, %r775, %r834;


dp4a.s32.s32 %r842, %r771, %r994, %r991;


dp4a.s32.s32 %r846, %r775, %r994, %r842;

cvt.rn.f32.s32 %f311, %r846;
fma.rn.ftz.f32 %f312, %f155, %f311, 0f00000000;

dp4a.s32.s32 %r850, %r981, %r787, %r838;


dp4a.s32.s32 %r854, %r985, %r791, %r850;


dp4a.s32.s32 %r858, %r787, %r994, %r991;


dp4a.s32.s32 %r862, %r791, %r994, %r858;

cvt.rn.f32.s32 %f313, %r862;
fma.rn.ftz.f32 %f314, %f160, %f313, %f312;

{ cvt.f32.f16 %f107, %rs14;}



{.reg .f16 low,high;
mov.b32 {low,high},%r801;
cvt.f32.f16 %f108, low;}


mul.ftz.f32 %f315, %f107, %f108;
cvt.rn.f32.s32 %f316, %r822;
add.ftz.f32 %f317, %f310, %f316;
cvt.rn.f32.s32 %f318, %r854;
add.ftz.f32 %f319, %f314, %f318;
mul.ftz.f32 %f320, %f319, %f169;
fma.rn.ftz.f32 %f321, %f317, %f166, %f320;
fma.rn.ftz.f32 %f522, %f315, %f321, %f522;
cvt.u64.u32 %rd87, %r1337;
add.s64 %rd88, %rd1, %rd87;
add.s64 %rd89, %rd88, %rd3;
mul.lo.s64 %rd90, %rd89, 36;
add.s64 %rd91, %rd12, %rd90;
ld.global.nc.u32 %r869, [%rd91+4];
ld.global.nc.u32 %r873, [%rd91+8];

dp4a.s32.s32 %r867, %r868, %r869, %r991;


dp4a.s32.s32 %r871, %r872, %r873, %r867;


dp4a.s32.s32 %r875, %r869, %r994, %r991;


dp4a.s32.s32 %r879, %r873, %r994, %r875;

cvt.rn.f32.s32 %f322, %r879;
fma.rn.ftz.f32 %f323, %f116, %f322, 0f00000000;
ld.global.nc.u32 %r885, [%rd91+12];
ld.global.nc.u32 %r889, [%rd91+16];

dp4a.s32.s32 %r883, %r884, %r885, %r871;


dp4a.s32.s32 %r887, %r888, %r889, %r883;


dp4a.s32.s32 %r891, %r885, %r994, %r991;


dp4a.s32.s32 %r895, %r889, %r994, %r891;

cvt.rn.f32.s32 %f324, %r895;
fma.rn.ftz.f32 %f325, %f121, %f324, %f323;
ld.global.nc.u32 %r901, [%rd91+20];
ld.global.nc.u32 %r905, [%rd91+24];

dp4a.s32.s32 %r899, %r900, %r901, %r991;


dp4a.s32.s32 %r903, %r904, %r905, %r899;


dp4a.s32.s32 %r907, %r901, %r994, %r991;


dp4a.s32.s32 %r911, %r905, %r994, %r907;

cvt.rn.f32.s32 %f326, %r911;
fma.rn.ftz.f32 %f327, %f126, %f326, 0f00000000;
ld.global.nc.u32 %r917, [%rd91+28];
ld.global.nc.u32 %r921, [%rd91+32];

dp4a.s32.s32 %r915, %r916, %r917, %r903;


dp4a.s32.s32 %r919, %r920, %r921, %r915;


dp4a.s32.s32 %r923, %r917, %r994, %r991;


dp4a.s32.s32 %r927, %r921, %r994, %r923;

cvt.rn.f32.s32 %f328, %r927;
fma.rn.ftz.f32 %f329, %f131, %f328, %f327;

{ cvt.f32.f16 %f109, %rs13;}


ld.global.nc.u32 %r931, [%rd91];

{.reg .f16 low,high;
mov.b32 {low,high},%r931;
cvt.f32.f16 %f110, low;}


mul.ftz.f32 %f330, %f109, %f110;
cvt.rn.f32.s32 %f331, %r887;
add.ftz.f32 %f332, %f325, %f331;
cvt.rn.f32.s32 %f333, %r919;
add.ftz.f32 %f334, %f329, %f333;
mul.ftz.f32 %f335, %f334, %f140;
fma.rn.ftz.f32 %f336, %f332, %f137, %f335;
fma.rn.ftz.f32 %f521, %f330, %f336, %f521;

dp4a.s32.s32 %r932, %r933, %r869, %r991;


dp4a.s32.s32 %r936, %r937, %r873, %r932;


dp4a.s32.s32 %r940, %r869, %r994, %r991;


dp4a.s32.s32 %r944, %r873, %r994, %r940;

cvt.rn.f32.s32 %f337, %r944;
fma.rn.ftz.f32 %f338, %f145, %f337, 0f00000000;

dp4a.s32.s32 %r948, %r949, %r885, %r936;


dp4a.s32.s32 %r952, %r953, %r889, %r948;


dp4a.s32.s32 %r956, %r885, %r994, %r991;


dp4a.s32.s32 %r960, %r889, %r994, %r956;

cvt.rn.f32.s32 %f339, %r960;
fma.rn.ftz.f32 %f340, %f150, %f339, %f338;

dp4a.s32.s32 %r964, %r965, %r901, %r991;


dp4a.s32.s32 %r968, %r969, %r905, %r964;


dp4a.s32.s32 %r972, %r901, %r994, %r991;


dp4a.s32.s32 %r976, %r905, %r994, %r972;

cvt.rn.f32.s32 %f341, %r976;
fma.rn.ftz.f32 %f342, %f155, %f341, 0f00000000;

dp4a.s32.s32 %r980, %r981, %r917, %r968;


dp4a.s32.s32 %r984, %r985, %r921, %r980;


dp4a.s32.s32 %r988, %r917, %r994, %r991;


dp4a.s32.s32 %r992, %r921, %r994, %r988;

cvt.rn.f32.s32 %f343, %r992;
fma.rn.ftz.f32 %f344, %f160, %f343, %f342;

{ cvt.f32.f16 %f111, %rs14;}



{.reg .f16 low,high;
mov.b32 {low,high},%r931;
cvt.f32.f16 %f112, low;}


mul.ftz.f32 %f345, %f111, %f112;
cvt.rn.f32.s32 %f346, %r952;
add.ftz.f32 %f347, %f340, %f346;
cvt.rn.f32.s32 %f348, %r984;
add.ftz.f32 %f349, %f344, %f348;
mul.ftz.f32 %f350, %f349, %f169;
fma.rn.ftz.f32 %f351, %f347, %f166, %f350;
fma.rn.ftz.f32 %f520, %f345, %f351, %f520;
add.s32 %r1341, %r1341, 64;
add.s32 %r1340, %r1340, 64;
add.s32 %r1339, %r1339, 64;
add.s32 %r1338, %r1338, 64;
add.s32 %r1337, %r1337, 64;
add.s32 %r1336, %r1336, 64;
add.s32 %r1335, %r1335, 64;
add.s64 %rd136, %rd136, 448;
add.s32 %r1342, %r1342, 8;
setp.lt.s32 %p4, %r1342, %r6;
@%p4 bra $L__BB274_2;

$L__BB274_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB274_5;

mov.u32 %r1077, _ZZ13mul_mat_vec_qIL9ggml_type29ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r1078, %r1, 1792, %r1077;
shl.b32 %r1079, %r2, 2;
add.s32 %r1080, %r1078, %r1079;
st.shared.f32 [%r1080+-1792], %f533;
st.shared.f32 [%r1080+-1664], %f532;
st.shared.f32 [%r1080+-1536], %f531;
st.shared.f32 [%r1080+-1408], %f530;
st.shared.f32 [%r1080+-1280], %f529;
st.shared.f32 [%r1080+-1152], %f528;
st.shared.f32 [%r1080+-1024], %f527;
st.shared.f32 [%r1080+-896], %f526;
st.shared.f32 [%r1080+-768], %f525;
st.shared.f32 [%r1080+-640], %f524;
st.shared.f32 [%r1080+-512], %f523;
st.shared.f32 [%r1080+-384], %f522;
st.shared.f32 [%r1080+-256], %f521;
st.shared.f32 [%r1080+-128], %f520;

$L__BB274_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB274_34;

mov.u32 %r1081, %ctaid.z;
mov.u32 %r1082, %ctaid.y;
mad.lo.s32 %r1083, %r1082, %r42, %r4;
mad.lo.s32 %r1084, %r1081, %r48, %r1083;
cvt.u64.u32 %rd15, %r1084;
shl.b32 %r1085, %r2, 2;
mov.u32 %r1086, _ZZ13mul_mat_vec_qIL9ggml_type29ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r31, %r1086, %r1085;
mov.u32 %r1087, 2;
ld.shared.f32 %f352, [%r31];
add.ftz.f32 %f353, %f352, %f533;
mov.b32 %r1088, %f353;
mov.u32 %r1089, 31;
mov.u32 %r1090, 16;
mov.u32 %r1091, -1;
shfl.sync.bfly.b32 %r1092|%p7, %r1088, %r1090, %r1089, %r1091;
mov.b32 %f354, %r1092;
add.ftz.f32 %f355, %f353, %f354;
mov.b32 %r1093, %f355;
mov.u32 %r1094, 8;
shfl.sync.bfly.b32 %r1095|%p8, %r1093, %r1094, %r1089, %r1091;
mov.b32 %f356, %r1095;
add.ftz.f32 %f357, %f355, %f356;
mov.b32 %r1096, %f357;
mov.u32 %r1097, 4;
shfl.sync.bfly.b32 %r1098|%p9, %r1096, %r1097, %r1089, %r1091;
mov.b32 %f358, %r1098;
add.ftz.f32 %f359, %f357, %f358;
mov.b32 %r1099, %f359;
shfl.sync.bfly.b32 %r1100|%p10, %r1099, %r1087, %r1089, %r1091;
mov.b32 %f360, %r1100;
add.ftz.f32 %f361, %f359, %f360;
mov.b32 %r1101, %f361;
mov.u32 %r1102, 1;
shfl.sync.bfly.b32 %r1103|%p11, %r1101, %r1102, %r1089, %r1091;
mov.b32 %f362, %r1103;
add.ftz.f32 %f43, %f361, %f362;
setp.lt.u32 %p12, %r4, %r36;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
cvta.to.global.u64 %rd92, %rd22;
mul.wide.u32 %rd93, %r1084, 4;
add.s64 %rd16, %rd92, %rd93;
not.pred %p14, %p1;
@%p14 bra $L__BB274_8;

st.global.f32 [%rd16], %f43;

$L__BB274_8:
ld.shared.f32 %f363, [%r31+128];
add.ftz.f32 %f364, %f363, %f532;
mov.b32 %r1104, %f364;
shfl.sync.bfly.b32 %r1108|%p15, %r1104, %r1090, %r1089, %r1091;
mov.b32 %f365, %r1108;
add.ftz.f32 %f366, %f364, %f365;
mov.b32 %r1109, %f366;
shfl.sync.bfly.b32 %r1111|%p16, %r1109, %r1094, %r1089, %r1091;
mov.b32 %f367, %r1111;
add.ftz.f32 %f368, %f366, %f367;
mov.b32 %r1112, %f368;
shfl.sync.bfly.b32 %r1114|%p17, %r1112, %r1097, %r1089, %r1091;
mov.b32 %f369, %r1114;
add.ftz.f32 %f370, %f368, %f369;
mov.b32 %r1115, %f370;
shfl.sync.bfly.b32 %r1117|%p18, %r1115, %r1087, %r1089, %r1091;
mov.b32 %f371, %r1117;
add.ftz.f32 %f372, %f370, %f371;
mov.b32 %r1118, %f372;
shfl.sync.bfly.b32 %r1120|%p19, %r1118, %r1102, %r1089, %r1091;
mov.b32 %f373, %r1120;
add.ftz.f32 %f44, %f372, %f373;
setp.eq.s32 %p20, %r2, 1;
or.b32 %r1124, %r4, 1;
setp.lt.u32 %p21, %r1124, %r36;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB274_10;

st.global.f32 [%rd16+4], %f44;

$L__BB274_10:
ld.shared.f32 %f374, [%r31+256];
add.ftz.f32 %f375, %f374, %f531;
mov.b32 %r1125, %f375;
mov.u32 %r1126, 31;
mov.u32 %r1127, 16;
mov.u32 %r1128, -1;
shfl.sync.bfly.b32 %r1129|%p23, %r1125, %r1127, %r1126, %r1128;
mov.b32 %f376, %r1129;
add.ftz.f32 %f377, %f375, %f376;
mov.b32 %r1130, %f377;
mov.u32 %r1131, 8;
shfl.sync.bfly.b32 %r1132|%p24, %r1130, %r1131, %r1126, %r1128;
mov.b32 %f378, %r1132;
add.ftz.f32 %f379, %f377, %f378;
mov.b32 %r1133, %f379;
mov.u32 %r1134, 4;
shfl.sync.bfly.b32 %r1135|%p25, %r1133, %r1134, %r1126, %r1128;
mov.b32 %f380, %r1135;
add.ftz.f32 %f381, %f379, %f380;
mov.b32 %r1136, %f381;
mov.u32 %r1137, 2;
shfl.sync.bfly.b32 %r1138|%p26, %r1136, %r1137, %r1126, %r1128;
mov.b32 %f382, %r1138;
add.ftz.f32 %f383, %f381, %f382;
mov.b32 %r1139, %f383;
mov.u32 %r1140, 1;
shfl.sync.bfly.b32 %r1141|%p27, %r1139, %r1140, %r1126, %r1128;
mov.b32 %f384, %r1141;
add.ftz.f32 %f45, %f383, %f384;
@%p14 bra $L__BB274_12;

cvt.u64.u32 %rd94, %r36;
add.s64 %rd95, %rd94, %rd15;
shl.b64 %rd97, %rd95, 2;
add.s64 %rd98, %rd92, %rd97;
st.global.f32 [%rd98], %f45;

$L__BB274_12:
ld.shared.f32 %f385, [%r31+384];
add.ftz.f32 %f386, %f385, %f530;
mov.b32 %r1142, %f386;
shfl.sync.bfly.b32 %r1146|%p29, %r1142, %r1127, %r1126, %r1128;
mov.b32 %f387, %r1146;
add.ftz.f32 %f388, %f386, %f387;
mov.b32 %r1147, %f388;
shfl.sync.bfly.b32 %r1149|%p30, %r1147, %r1131, %r1126, %r1128;
mov.b32 %f389, %r1149;
add.ftz.f32 %f390, %f388, %f389;
mov.b32 %r1150, %f390;
shfl.sync.bfly.b32 %r1152|%p31, %r1150, %r1134, %r1126, %r1128;
mov.b32 %f391, %r1152;
add.ftz.f32 %f392, %f390, %f391;
mov.b32 %r1153, %f392;
shfl.sync.bfly.b32 %r1155|%p32, %r1153, %r1137, %r1126, %r1128;
mov.b32 %f393, %r1155;
add.ftz.f32 %f394, %f392, %f393;
mov.b32 %r1156, %f394;
shfl.sync.bfly.b32 %r1158|%p33, %r1156, %r1140, %r1126, %r1128;
mov.b32 %f395, %r1158;
add.ftz.f32 %f46, %f394, %f395;
@%p22 bra $L__BB274_14;

add.s32 %r1159, %r36, 1;
cvt.u64.u32 %rd99, %r1159;
add.s64 %rd100, %rd99, %rd15;
shl.b64 %rd102, %rd100, 2;
add.s64 %rd103, %rd92, %rd102;
st.global.f32 [%rd103], %f46;

$L__BB274_14:
shl.b32 %r1160, %r36, 1;
mov.u32 %r1161, 1;
ld.shared.f32 %f396, [%r31+512];
add.ftz.f32 %f397, %f396, %f529;
mov.b32 %r1162, %f397;
mov.u32 %r1163, 31;
mov.u32 %r1164, 16;
mov.u32 %r1165, -1;
shfl.sync.bfly.b32 %r1166|%p35, %r1162, %r1164, %r1163, %r1165;
mov.b32 %f398, %r1166;
add.ftz.f32 %f399, %f397, %f398;
mov.b32 %r1167, %f399;
mov.u32 %r1168, 8;
shfl.sync.bfly.b32 %r1169|%p36, %r1167, %r1168, %r1163, %r1165;
mov.b32 %f400, %r1169;
add.ftz.f32 %f401, %f399, %f400;
mov.b32 %r1170, %f401;
mov.u32 %r1171, 4;
shfl.sync.bfly.b32 %r1172|%p37, %r1170, %r1171, %r1163, %r1165;
mov.b32 %f402, %r1172;
add.ftz.f32 %f403, %f401, %f402;
mov.b32 %r1173, %f403;
mov.u32 %r1174, 2;
shfl.sync.bfly.b32 %r1175|%p38, %r1173, %r1174, %r1163, %r1165;
mov.b32 %f404, %r1175;
add.ftz.f32 %f405, %f403, %f404;
mov.b32 %r1176, %f405;
shfl.sync.bfly.b32 %r1177|%p39, %r1176, %r1161, %r1163, %r1165;
mov.b32 %f406, %r1177;
add.ftz.f32 %f47, %f405, %f406;
cvt.u64.u32 %rd104, %r1160;
add.s64 %rd105, %rd104, %rd15;
shl.b64 %rd107, %rd105, 2;
add.s64 %rd17, %rd92, %rd107;
@%p14 bra $L__BB274_16;

st.global.f32 [%rd17], %f47;

$L__BB274_16:
ld.shared.f32 %f407, [%r31+640];
add.ftz.f32 %f408, %f407, %f528;
mov.b32 %r1178, %f408;
shfl.sync.bfly.b32 %r1182|%p41, %r1178, %r1164, %r1163, %r1165;
mov.b32 %f409, %r1182;
add.ftz.f32 %f410, %f408, %f409;
mov.b32 %r1183, %f410;
shfl.sync.bfly.b32 %r1185|%p42, %r1183, %r1168, %r1163, %r1165;
mov.b32 %f411, %r1185;
add.ftz.f32 %f412, %f410, %f411;
mov.b32 %r1186, %f412;
shfl.sync.bfly.b32 %r1188|%p43, %r1186, %r1171, %r1163, %r1165;
mov.b32 %f413, %r1188;
add.ftz.f32 %f414, %f412, %f413;
mov.b32 %r1189, %f414;
shfl.sync.bfly.b32 %r1191|%p44, %r1189, %r1174, %r1163, %r1165;
mov.b32 %f415, %r1191;
add.ftz.f32 %f416, %f414, %f415;
mov.b32 %r1192, %f416;
shfl.sync.bfly.b32 %r1194|%p45, %r1192, %r1161, %r1163, %r1165;
mov.b32 %f417, %r1194;
add.ftz.f32 %f48, %f416, %f417;
@%p22 bra $L__BB274_18;

st.global.f32 [%rd17+4], %f48;

$L__BB274_18:
mul.lo.s32 %r32, %r36, 3;
ld.shared.f32 %f418, [%r31+768];
add.ftz.f32 %f419, %f418, %f527;
mov.b32 %r1195, %f419;
mov.u32 %r1196, 31;
mov.u32 %r1197, 16;
mov.u32 %r1198, -1;
shfl.sync.bfly.b32 %r1199|%p47, %r1195, %r1197, %r1196, %r1198;
mov.b32 %f420, %r1199;
add.ftz.f32 %f421, %f419, %f420;
mov.b32 %r1200, %f421;
mov.u32 %r1201, 8;
shfl.sync.bfly.b32 %r1202|%p48, %r1200, %r1201, %r1196, %r1198;
mov.b32 %f422, %r1202;
add.ftz.f32 %f423, %f421, %f422;
mov.b32 %r1203, %f423;
mov.u32 %r1204, 4;
shfl.sync.bfly.b32 %r1205|%p49, %r1203, %r1204, %r1196, %r1198;
mov.b32 %f424, %r1205;
add.ftz.f32 %f425, %f423, %f424;
mov.b32 %r1206, %f425;
mov.u32 %r1207, 2;
shfl.sync.bfly.b32 %r1208|%p50, %r1206, %r1207, %r1196, %r1198;
mov.b32 %f426, %r1208;
add.ftz.f32 %f427, %f425, %f426;
mov.b32 %r1209, %f427;
mov.u32 %r1210, 1;
shfl.sync.bfly.b32 %r1211|%p51, %r1209, %r1210, %r1196, %r1198;
mov.b32 %f428, %r1211;
add.ftz.f32 %f49, %f427, %f428;
@%p14 bra $L__BB274_20;

cvt.u64.u32 %rd108, %r32;
add.s64 %rd109, %rd108, %rd15;
shl.b64 %rd111, %rd109, 2;
add.s64 %rd112, %rd92, %rd111;
st.global.f32 [%rd112], %f49;

$L__BB274_20:
ld.shared.f32 %f429, [%r31+896];
add.ftz.f32 %f430, %f429, %f526;
mov.b32 %r1212, %f430;
shfl.sync.bfly.b32 %r1216|%p53, %r1212, %r1197, %r1196, %r1198;
mov.b32 %f431, %r1216;
add.ftz.f32 %f432, %f430, %f431;
mov.b32 %r1217, %f432;
shfl.sync.bfly.b32 %r1219|%p54, %r1217, %r1201, %r1196, %r1198;
mov.b32 %f433, %r1219;
add.ftz.f32 %f434, %f432, %f433;
mov.b32 %r1220, %f434;
shfl.sync.bfly.b32 %r1222|%p55, %r1220, %r1204, %r1196, %r1198;
mov.b32 %f435, %r1222;
add.ftz.f32 %f436, %f434, %f435;
mov.b32 %r1223, %f436;
shfl.sync.bfly.b32 %r1225|%p56, %r1223, %r1207, %r1196, %r1198;
mov.b32 %f437, %r1225;
add.ftz.f32 %f438, %f436, %f437;
mov.b32 %r1226, %f438;
shfl.sync.bfly.b32 %r1228|%p57, %r1226, %r1210, %r1196, %r1198;
mov.b32 %f439, %r1228;
add.ftz.f32 %f50, %f438, %f439;
@%p22 bra $L__BB274_22;

add.s32 %r1229, %r32, 1;
cvt.u64.u32 %rd113, %r1229;
add.s64 %rd114, %rd113, %rd15;
shl.b64 %rd116, %rd114, 2;
add.s64 %rd117, %rd92, %rd116;
st.global.f32 [%rd117], %f50;

$L__BB274_22:
shl.b32 %r1230, %r36, 2;
mov.u32 %r1231, 2;
ld.shared.f32 %f440, [%r31+1024];
add.ftz.f32 %f441, %f440, %f525;
mov.b32 %r1232, %f441;
mov.u32 %r1233, 31;
mov.u32 %r1234, 16;
mov.u32 %r1235, -1;
shfl.sync.bfly.b32 %r1236|%p59, %r1232, %r1234, %r1233, %r1235;
mov.b32 %f442, %r1236;
add.ftz.f32 %f443, %f441, %f442;
mov.b32 %r1237, %f443;
mov.u32 %r1238, 8;
shfl.sync.bfly.b32 %r1239|%p60, %r1237, %r1238, %r1233, %r1235;
mov.b32 %f444, %r1239;
add.ftz.f32 %f445, %f443, %f444;
mov.b32 %r1240, %f445;
mov.u32 %r1241, 4;
shfl.sync.bfly.b32 %r1242|%p61, %r1240, %r1241, %r1233, %r1235;
mov.b32 %f446, %r1242;
add.ftz.f32 %f447, %f445, %f446;
mov.b32 %r1243, %f447;
shfl.sync.bfly.b32 %r1244|%p62, %r1243, %r1231, %r1233, %r1235;
mov.b32 %f448, %r1244;
add.ftz.f32 %f449, %f447, %f448;
mov.b32 %r1245, %f449;
mov.u32 %r1246, 1;
shfl.sync.bfly.b32 %r1247|%p63, %r1245, %r1246, %r1233, %r1235;
mov.b32 %f450, %r1247;
add.ftz.f32 %f51, %f449, %f450;
cvt.u64.u32 %rd118, %r1230;
add.s64 %rd119, %rd118, %rd15;
shl.b64 %rd121, %rd119, 2;
add.s64 %rd18, %rd92, %rd121;
@%p14 bra $L__BB274_24;

st.global.f32 [%rd18], %f51;

$L__BB274_24:
ld.shared.f32 %f451, [%r31+1152];
add.ftz.f32 %f452, %f451, %f524;
mov.b32 %r1248, %f452;
shfl.sync.bfly.b32 %r1252|%p65, %r1248, %r1234, %r1233, %r1235;
mov.b32 %f453, %r1252;
add.ftz.f32 %f454, %f452, %f453;
mov.b32 %r1253, %f454;
shfl.sync.bfly.b32 %r1255|%p66, %r1253, %r1238, %r1233, %r1235;
mov.b32 %f455, %r1255;
add.ftz.f32 %f456, %f454, %f455;
mov.b32 %r1256, %f456;
shfl.sync.bfly.b32 %r1258|%p67, %r1256, %r1241, %r1233, %r1235;
mov.b32 %f457, %r1258;
add.ftz.f32 %f458, %f456, %f457;
mov.b32 %r1259, %f458;
shfl.sync.bfly.b32 %r1261|%p68, %r1259, %r1231, %r1233, %r1235;
mov.b32 %f459, %r1261;
add.ftz.f32 %f460, %f458, %f459;
mov.b32 %r1262, %f460;
shfl.sync.bfly.b32 %r1264|%p69, %r1262, %r1246, %r1233, %r1235;
mov.b32 %f461, %r1264;
add.ftz.f32 %f52, %f460, %f461;
@%p22 bra $L__BB274_26;

st.global.f32 [%rd18+4], %f52;

$L__BB274_26:
mul.lo.s32 %r33, %r36, 5;
ld.shared.f32 %f462, [%r31+1280];
add.ftz.f32 %f463, %f462, %f523;
mov.b32 %r1265, %f463;
mov.u32 %r1266, 31;
mov.u32 %r1267, 16;
mov.u32 %r1268, -1;
shfl.sync.bfly.b32 %r1269|%p71, %r1265, %r1267, %r1266, %r1268;
mov.b32 %f464, %r1269;
add.ftz.f32 %f465, %f463, %f464;
mov.b32 %r1270, %f465;
mov.u32 %r1271, 8;
shfl.sync.bfly.b32 %r1272|%p72, %r1270, %r1271, %r1266, %r1268;
mov.b32 %f466, %r1272;
add.ftz.f32 %f467, %f465, %f466;
mov.b32 %r1273, %f467;
mov.u32 %r1274, 4;
shfl.sync.bfly.b32 %r1275|%p73, %r1273, %r1274, %r1266, %r1268;
mov.b32 %f468, %r1275;
add.ftz.f32 %f469, %f467, %f468;
mov.b32 %r1276, %f469;
mov.u32 %r1277, 2;
shfl.sync.bfly.b32 %r1278|%p74, %r1276, %r1277, %r1266, %r1268;
mov.b32 %f470, %r1278;
add.ftz.f32 %f471, %f469, %f470;
mov.b32 %r1279, %f471;
mov.u32 %r1280, 1;
shfl.sync.bfly.b32 %r1281|%p75, %r1279, %r1280, %r1266, %r1268;
mov.b32 %f472, %r1281;
add.ftz.f32 %f53, %f471, %f472;
@%p14 bra $L__BB274_28;

cvt.u64.u32 %rd122, %r33;
add.s64 %rd123, %rd122, %rd15;
shl.b64 %rd125, %rd123, 2;
add.s64 %rd126, %rd92, %rd125;
st.global.f32 [%rd126], %f53;

$L__BB274_28:
ld.shared.f32 %f473, [%r31+1408];
add.ftz.f32 %f474, %f473, %f522;
mov.b32 %r1282, %f474;
shfl.sync.bfly.b32 %r1286|%p77, %r1282, %r1267, %r1266, %r1268;
mov.b32 %f475, %r1286;
add.ftz.f32 %f476, %f474, %f475;
mov.b32 %r1287, %f476;
shfl.sync.bfly.b32 %r1289|%p78, %r1287, %r1271, %r1266, %r1268;
mov.b32 %f477, %r1289;
add.ftz.f32 %f478, %f476, %f477;
mov.b32 %r1290, %f478;
shfl.sync.bfly.b32 %r1292|%p79, %r1290, %r1274, %r1266, %r1268;
mov.b32 %f479, %r1292;
add.ftz.f32 %f480, %f478, %f479;
mov.b32 %r1293, %f480;
shfl.sync.bfly.b32 %r1295|%p80, %r1293, %r1277, %r1266, %r1268;
mov.b32 %f481, %r1295;
add.ftz.f32 %f482, %f480, %f481;
mov.b32 %r1296, %f482;
shfl.sync.bfly.b32 %r1298|%p81, %r1296, %r1280, %r1266, %r1268;
mov.b32 %f483, %r1298;
add.ftz.f32 %f54, %f482, %f483;
@%p22 bra $L__BB274_30;

add.s32 %r1299, %r33, 1;
cvt.u64.u32 %rd127, %r1299;
add.s64 %rd128, %rd127, %rd15;
shl.b64 %rd130, %rd128, 2;
add.s64 %rd131, %rd92, %rd130;
st.global.f32 [%rd131], %f54;

$L__BB274_30:
mul.lo.s32 %r1300, %r36, 6;
ld.shared.f32 %f484, [%r31+1536];
add.ftz.f32 %f485, %f484, %f521;
mov.b32 %r1301, %f485;
mov.u32 %r1302, 31;
mov.u32 %r1303, 16;
mov.u32 %r1304, -1;
shfl.sync.bfly.b32 %r1305|%p83, %r1301, %r1303, %r1302, %r1304;
mov.b32 %f486, %r1305;
add.ftz.f32 %f487, %f485, %f486;
mov.b32 %r1306, %f487;
mov.u32 %r1307, 8;
shfl.sync.bfly.b32 %r1308|%p84, %r1306, %r1307, %r1302, %r1304;
mov.b32 %f488, %r1308;
add.ftz.f32 %f489, %f487, %f488;
mov.b32 %r1309, %f489;
mov.u32 %r1310, 4;
shfl.sync.bfly.b32 %r1311|%p85, %r1309, %r1310, %r1302, %r1304;
mov.b32 %f490, %r1311;
add.ftz.f32 %f491, %f489, %f490;
mov.b32 %r1312, %f491;
mov.u32 %r1313, 2;
shfl.sync.bfly.b32 %r1314|%p86, %r1312, %r1313, %r1302, %r1304;
mov.b32 %f492, %r1314;
add.ftz.f32 %f493, %f491, %f492;
mov.b32 %r1315, %f493;
mov.u32 %r1316, 1;
shfl.sync.bfly.b32 %r1317|%p87, %r1315, %r1316, %r1302, %r1304;
mov.b32 %f494, %r1317;
add.ftz.f32 %f55, %f493, %f494;
cvt.u64.u32 %rd132, %r1300;
add.s64 %rd133, %rd132, %rd15;
shl.b64 %rd135, %rd133, 2;
add.s64 %rd19, %rd92, %rd135;
@%p14 bra $L__BB274_32;

st.global.f32 [%rd19], %f55;

$L__BB274_32:
ld.shared.f32 %f495, [%r31+1664];
add.ftz.f32 %f496, %f495, %f520;
mov.b32 %r1318, %f496;
shfl.sync.bfly.b32 %r1322|%p89, %r1318, %r1303, %r1302, %r1304;
mov.b32 %f497, %r1322;
add.ftz.f32 %f498, %f496, %f497;
mov.b32 %r1323, %f498;
shfl.sync.bfly.b32 %r1325|%p90, %r1323, %r1307, %r1302, %r1304;
mov.b32 %f499, %r1325;
add.ftz.f32 %f500, %f498, %f499;
mov.b32 %r1326, %f500;
shfl.sync.bfly.b32 %r1328|%p91, %r1326, %r1310, %r1302, %r1304;
mov.b32 %f501, %r1328;
add.ftz.f32 %f502, %f500, %f501;
mov.b32 %r1329, %f502;
shfl.sync.bfly.b32 %r1331|%p92, %r1329, %r1313, %r1302, %r1304;
mov.b32 %f503, %r1331;
add.ftz.f32 %f504, %f502, %f503;
mov.b32 %r1332, %f504;
shfl.sync.bfly.b32 %r1334|%p93, %r1332, %r1316, %r1302, %r1304;
mov.b32 %f505, %r1334;
add.ftz.f32 %f56, %f504, %f505;
@%p22 bra $L__BB274_34;

st.global.f32 [%rd19+4], %f56;

$L__BB274_34:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type29ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type29ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type29ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type29ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type29ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type29ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type29ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type29ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type29ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type29ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<107>;
.reg .b16 %rs<55>;
.reg .f32 %f<606>;
.reg .b32 %r<1589>;
.reg .b64 %rd<155>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type29ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[2048];

ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type29ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd20, [_Z13mul_mat_vec_qIL9ggml_type29ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd21, [_Z13mul_mat_vec_qIL9ggml_type29ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r50, [_Z13mul_mat_vec_qIL9ggml_type29ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type29ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type29ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type29ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type29ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type29ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type29ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r47, [_Z13mul_mat_vec_qIL9ggml_type29ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r48, [_Z13mul_mat_vec_qIL9ggml_type29ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r49, [_Z13mul_mat_vec_qIL9ggml_type29ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r45, [_Z13mul_mat_vec_qIL9ggml_type29ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type29ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type29ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type29ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r51, %r1, 5;
mov.u32 %r52, %tid.x;
add.s32 %r2, %r51, %r52;
shr.s32 %r53, %r2, 31;
shr.u32 %r54, %r53, 29;
add.s32 %r55, %r2, %r54;
shr.s32 %r1588, %r55, 3;
shr.u32 %r4, %r50, 8;
setp.ge.s32 %p3, %r1588, %r4;
mov.f32 %f590, 0f00000000;
mov.f32 %f591, %f590;
mov.f32 %f592, %f590;
mov.f32 %f593, %f590;
mov.f32 %f594, %f590;
mov.f32 %f595, %f590;
mov.f32 %f596, %f590;
mov.f32 %f597, %f590;
mov.f32 %f598, %f590;
mov.f32 %f599, %f590;
mov.f32 %f600, %f590;
mov.f32 %f601, %f590;
mov.f32 %f602, %f590;
mov.f32 %f603, %f590;
mov.f32 %f604, %f590;
mov.f32 %f605, %f590;
@%p3 bra $L__BB275_3;

mov.u32 %r56, %ctaid.x;
shl.b32 %r57, %r56, 1;
mov.u32 %r58, %ctaid.y;
mov.u32 %r59, %ctaid.z;
mul.lo.s32 %r60, %r58, %r42;
cvt.u64.u32 %rd22, %r60;
mul.lo.s32 %r61, %r59, %r48;
cvt.u64.u32 %rd23, %r61;
add.s64 %rd1, %rd23, %rd22;
mul.hi.u32 %r62, %r59, %r44;
add.s32 %r63, %r62, %r59;
shr.u32 %r64, %r63, %r45;
mul.lo.s32 %r65, %r64, %r47;
mul.hi.u32 %r66, %r58, %r38;
add.s32 %r67, %r66, %r58;
shr.u32 %r68, %r67, %r39;
mul.lo.s32 %r69, %r68, %r41;
cvta.to.global.u64 %rd154, %rd19;
or.b32 %r70, %r57, 1;
and.b32 %r74, %r55, -8;
sub.s32 %r75, %r2, %r74;
shl.b32 %r76, %r75, 1;
cvt.s64.s32 %rd24, %r76;
cvt.s64.s32 %rd3, %r75;
shr.u32 %r77, %r75, 31;
add.s32 %r78, %r75, %r77;
shr.s32 %r79, %r78, 1;
and.b32 %r80, %r78, -2;
sub.s32 %r81, %r75, %r80;
mul.lo.s32 %r5, %r81, 6;
shl.b32 %r82, %r36, 1;
shl.b32 %r1580, %r1588, 3;
add.s32 %r1587, %r1580, %r82;
mad.lo.s32 %r1586, %r36, 3, %r1580;
shl.b32 %r83, %r36, 2;
add.s32 %r1585, %r1580, %r83;
mad.lo.s32 %r1584, %r36, 5, %r1580;
mad.lo.s32 %r1583, %r36, 6, %r1580;
mad.lo.s32 %r1582, %r36, 7, %r1580;
add.s32 %r1581, %r36, %r1580;
add.s32 %r84, %r1588, %r65;
add.s32 %r85, %r84, %r69;
mad.lo.s32 %r86, %r70, %r35, %r85;
mul.wide.s32 %rd4, %r86, 56;
mad.lo.s32 %r87, %r57, %r35, %r1588;
add.s32 %r88, %r87, %r65;
add.s32 %r89, %r88, %r69;
mul.wide.s32 %rd5, %r89, 56;
mul.wide.s32 %rd25, %r75, 4;
add.s64 %rd6, %rd5, %rd25;
mul.wide.s32 %rd26, %r79, 2;
add.s64 %rd27, %rd5, %rd26;
add.s64 %rd7, %rd27, 48;
add.s64 %rd28, %rd4, %rd26;
add.s64 %rd8, %rd28, 48;
add.s64 %rd29, %rd5, %rd24;
add.s64 %rd9, %rd29, 33;
add.s64 %rd30, %rd4, %rd24;
add.s64 %rd10, %rd30, 32;
add.s64 %rd11, %rd4, %rd25;
cvta.to.global.u64 %rd12, %rd20;
mov.f32 %f590, 0f00000000;

$L__BB275_2:
cvt.u64.u32 %rd31, %r1580;
add.s64 %rd32, %rd1, %rd31;
add.s64 %rd33, %rd32, %rd3;
mul.lo.s64 %rd34, %rd33, 36;
add.s64 %rd35, %rd154, %rd6;
ld.global.nc.u32 %r1130, [%rd35];
shr.u32 %r1131, %r1130, 8;
shr.u32 %r1132, %r1130, 16;
shr.u32 %r1133, %r1130, 24;
add.s64 %rd36, %rd154, %rd9;
ld.global.nc.u8 %rs17, [%rd36+-1];
and.b16 %rs18, %rs17, 240;
cvt.u32.u16 %r1134, %rs17;
and.b32 %r1135, %r1130, 255;
and.b32 %r1136, %r1134, 7;
bfi.b32 %r1137, %r1136, %r1135, 8, 3;
mul.wide.u32 %rd37, %r1137, 4;
mov.u64 %rd38, iq1s_grid_gpu;
add.s64 %rd39, %rd38, %rd37;
ld.global.nc.u32 %r1138, [%rd39];
and.b32 %r1001, %r1138, 252645135;
shr.u32 %r1139, %r1138, 4;
and.b32 %r1005, %r1139, 252645135;
add.s64 %rd40, %rd12, %rd34;
ld.global.nc.u32 %r92, [%rd40+4];
ld.global.nc.u32 %r96, [%rd40+8];
mov.u32 %r1124, 0;

dp4a.s32.s32 %r90, %r1001, %r92, %r1124;


dp4a.s32.s32 %r94, %r1005, %r96, %r90;

and.b32 %r1140, %r1134, 8;
cvt.rn.f32.s32 %f129, %r1140;
mul.ftz.f32 %f130, %f129, 0f3D000000;
mov.f32 %f131, 0fBF600000;
sub.ftz.f32 %f132, %f131, %f130;
mov.u32 %r1127, 16843009;

dp4a.s32.s32 %r98, %r92, %r1127, %r1124;


dp4a.s32.s32 %r102, %r96, %r1127, %r98;

cvt.rn.f32.s32 %f133, %r102;
fma.rn.ftz.f32 %f134, %f132, %f133, 0f00000000;
shr.u16 %rs19, %rs18, 4;
cvt.u32.u16 %r1141, %rs19;
and.b32 %r1142, %r1131, 255;
and.b32 %r1143, %r1141, 7;
bfi.b32 %r1144, %r1143, %r1142, 8, 3;
mul.wide.u32 %rd41, %r1144, 4;
add.s64 %rd42, %rd38, %rd41;
ld.global.nc.u32 %r1145, [%rd42];
and.b32 %r1017, %r1145, 252645135;
shr.u32 %r1146, %r1145, 4;
and.b32 %r1021, %r1146, 252645135;
ld.global.nc.u32 %r108, [%rd40+12];
ld.global.nc.u32 %r112, [%rd40+16];

dp4a.s32.s32 %r106, %r1017, %r108, %r94;


dp4a.s32.s32 %r110, %r1021, %r112, %r106;

and.b32 %r1147, %r1141, 8;
cvt.rn.f32.s32 %f135, %r1147;
mul.ftz.f32 %f136, %f135, 0f3D000000;
sub.ftz.f32 %f137, %f131, %f136;

dp4a.s32.s32 %r114, %r108, %r1127, %r1124;


dp4a.s32.s32 %r118, %r112, %r1127, %r114;

cvt.rn.f32.s32 %f138, %r118;
fma.rn.ftz.f32 %f139, %f137, %f138, %f134;
ld.global.nc.u8 %rs20, [%rd36];
and.b16 %rs21, %rs20, 240;
cvt.u32.u16 %r1148, %rs20;
and.b32 %r1149, %r1132, 255;
and.b32 %r1150, %r1148, 7;
bfi.b32 %r1151, %r1150, %r1149, 8, 3;
mul.wide.u32 %rd43, %r1151, 4;
add.s64 %rd44, %rd38, %rd43;
ld.global.nc.u32 %r1152, [%rd44];
and.b32 %r1033, %r1152, 252645135;
shr.u32 %r1153, %r1152, 4;
and.b32 %r1037, %r1153, 252645135;
ld.global.nc.u32 %r124, [%rd40+20];
ld.global.nc.u32 %r128, [%rd40+24];

dp4a.s32.s32 %r122, %r1033, %r124, %r1124;


dp4a.s32.s32 %r126, %r1037, %r128, %r122;

and.b32 %r1154, %r1148, 8;
cvt.rn.f32.s32 %f140, %r1154;
mul.ftz.f32 %f141, %f140, 0f3D000000;
sub.ftz.f32 %f142, %f131, %f141;

dp4a.s32.s32 %r130, %r124, %r1127, %r1124;


dp4a.s32.s32 %r134, %r128, %r1127, %r130;

cvt.rn.f32.s32 %f143, %r134;
fma.rn.ftz.f32 %f144, %f142, %f143, 0f00000000;
shr.u16 %rs22, %rs21, 4;
cvt.u32.u16 %r1155, %rs22;
mul.wide.u16 %r1156, %rs22, 256;
and.b32 %r1157, %r1156, 1792;
or.b32 %r1158, %r1157, %r1133;
mul.wide.u32 %rd45, %r1158, 4;
add.s64 %rd46, %rd38, %rd45;
ld.global.nc.u32 %r1159, [%rd46];
and.b32 %r1049, %r1159, 252645135;
shr.u32 %r1160, %r1159, 4;
and.b32 %r1053, %r1160, 252645135;
ld.global.nc.u32 %r140, [%rd40+28];
ld.global.nc.u32 %r144, [%rd40+32];

dp4a.s32.s32 %r138, %r1049, %r140, %r126;


dp4a.s32.s32 %r142, %r1053, %r144, %r138;

and.b32 %r1161, %r1155, 8;
cvt.rn.f32.s32 %f145, %r1161;
mul.ftz.f32 %f146, %f145, 0f3D000000;
sub.ftz.f32 %f147, %f131, %f146;

dp4a.s32.s32 %r146, %r140, %r1127, %r1124;


dp4a.s32.s32 %r150, %r144, %r1127, %r146;

cvt.rn.f32.s32 %f148, %r150;
fma.rn.ftz.f32 %f149, %f147, %f148, %f144;
add.s64 %rd47, %rd154, %rd5;
ld.global.nc.u16 %rs23, [%rd47+48];
shr.u16 %rs24, %rs23, 12;
ld.global.nc.u16 %rs25, [%rd47+50];
shr.u16 %rs26, %rs25, 8;
and.b16 %rs27, %rs26, 240;
or.b16 %rs28, %rs27, %rs24;
ld.global.nc.u16 %rs29, [%rd47+52];
shr.u16 %rs30, %rs29, 4;
and.b16 %rs31, %rs30, 3840;
or.b16 %rs32, %rs28, %rs31;
ld.global.nc.u16 %rs33, [%rd47+54];
and.b16 %rs34, %rs33, -4096;
or.b16 %rs15, %rs32, %rs34;

{ cvt.f32.f16 %f97, %rs15;}


ld.global.nc.u32 %r154, [%rd40];

{.reg .f16 low,high;
mov.b32 {low,high},%r154;
cvt.f32.f16 %f98, low;}


mul.ftz.f32 %f150, %f97, %f98;
add.s64 %rd48, %rd154, %rd7;
ld.global.nc.u16 %rs35, [%rd48];
cvt.u32.u16 %r1162, %rs35;
shr.u32 %r1163, %r1162, %r5;
shl.b32 %r1164, %r1163, 1;
and.b32 %r1165, %r1164, 14;
or.b32 %r1166, %r1165, 1;
shr.u32 %r1167, %r1163, 2;
and.b32 %r1168, %r1167, 14;
or.b32 %r1169, %r1168, 1;
cvt.rn.f32.s32 %f151, %r110;
add.ftz.f32 %f152, %f139, %f151;
cvt.rn.f32.s32 %f153, %r1166;
cvt.rn.f32.s32 %f154, %r142;
add.ftz.f32 %f155, %f149, %f154;
cvt.rn.f32.s32 %f156, %r1169;
mul.ftz.f32 %f157, %f155, %f156;
fma.rn.ftz.f32 %f158, %f152, %f153, %f157;
fma.rn.ftz.f32 %f605, %f150, %f158, %f605;
add.s64 %rd49, %rd154, %rd11;
ld.global.nc.u32 %r1170, [%rd49];
shr.u32 %r1171, %r1170, 8;
shr.u32 %r1172, %r1170, 16;
shr.u32 %r1173, %r1170, 24;
add.s64 %rd50, %rd154, %rd10;
ld.global.nc.u8 %rs36, [%rd50];
and.b16 %rs37, %rs36, 240;
cvt.u32.u16 %r1174, %rs36;
and.b32 %r1175, %r1170, 255;
and.b32 %r1176, %r1174, 7;
bfi.b32 %r1177, %r1176, %r1175, 8, 3;
mul.wide.u32 %rd51, %r1177, 4;
add.s64 %rd52, %rd38, %rd51;
ld.global.nc.u32 %r1178, [%rd52];
and.b32 %r1066, %r1178, 252645135;
shr.u32 %r1179, %r1178, 4;
and.b32 %r1070, %r1179, 252645135;

dp4a.s32.s32 %r155, %r1066, %r92, %r1124;


dp4a.s32.s32 %r159, %r1070, %r96, %r155;

and.b32 %r1180, %r1174, 8;
cvt.rn.f32.s32 %f159, %r1180;
mul.ftz.f32 %f160, %f159, 0f3D000000;
sub.ftz.f32 %f161, %f131, %f160;

dp4a.s32.s32 %r163, %r92, %r1127, %r1124;


dp4a.s32.s32 %r167, %r96, %r1127, %r163;

cvt.rn.f32.s32 %f162, %r167;
fma.rn.ftz.f32 %f163, %f161, %f162, 0f00000000;
shr.u16 %rs38, %rs37, 4;
cvt.u32.u16 %r1181, %rs38;
and.b32 %r1182, %r1171, 255;
and.b32 %r1183, %r1181, 7;
bfi.b32 %r1184, %r1183, %r1182, 8, 3;
mul.wide.u32 %rd53, %r1184, 4;
add.s64 %rd54, %rd38, %rd53;
ld.global.nc.u32 %r1185, [%rd54];
and.b32 %r1082, %r1185, 252645135;
shr.u32 %r1186, %r1185, 4;
and.b32 %r1086, %r1186, 252645135;

dp4a.s32.s32 %r171, %r1082, %r108, %r159;


dp4a.s32.s32 %r175, %r1086, %r112, %r171;

and.b32 %r1187, %r1181, 8;
cvt.rn.f32.s32 %f164, %r1187;
mul.ftz.f32 %f165, %f164, 0f3D000000;
sub.ftz.f32 %f166, %f131, %f165;

dp4a.s32.s32 %r179, %r108, %r1127, %r1124;


dp4a.s32.s32 %r183, %r112, %r1127, %r179;

cvt.rn.f32.s32 %f167, %r183;
fma.rn.ftz.f32 %f168, %f166, %f167, %f163;
ld.global.nc.u8 %rs39, [%rd50+1];
and.b16 %rs40, %rs39, 240;
cvt.u32.u16 %r1188, %rs39;
and.b32 %r1189, %r1172, 255;
and.b32 %r1190, %r1188, 7;
bfi.b32 %r1191, %r1190, %r1189, 8, 3;
mul.wide.u32 %rd55, %r1191, 4;
add.s64 %rd56, %rd38, %rd55;
ld.global.nc.u32 %r1192, [%rd56];
and.b32 %r1098, %r1192, 252645135;
shr.u32 %r1193, %r1192, 4;
and.b32 %r1102, %r1193, 252645135;

dp4a.s32.s32 %r187, %r1098, %r124, %r1124;


dp4a.s32.s32 %r191, %r1102, %r128, %r187;

and.b32 %r1194, %r1188, 8;
cvt.rn.f32.s32 %f169, %r1194;
mul.ftz.f32 %f170, %f169, 0f3D000000;
sub.ftz.f32 %f171, %f131, %f170;

dp4a.s32.s32 %r195, %r124, %r1127, %r1124;


dp4a.s32.s32 %r199, %r128, %r1127, %r195;

cvt.rn.f32.s32 %f172, %r199;
fma.rn.ftz.f32 %f173, %f171, %f172, 0f00000000;
shr.u16 %rs41, %rs40, 4;
cvt.u32.u16 %r1195, %rs41;
mul.wide.u16 %r1196, %rs41, 256;
and.b32 %r1197, %r1196, 1792;
or.b32 %r1198, %r1197, %r1173;
mul.wide.u32 %rd57, %r1198, 4;
add.s64 %rd58, %rd38, %rd57;
ld.global.nc.u32 %r1199, [%rd58];
and.b32 %r1114, %r1199, 252645135;
shr.u32 %r1200, %r1199, 4;
and.b32 %r1118, %r1200, 252645135;

dp4a.s32.s32 %r203, %r1114, %r140, %r191;


dp4a.s32.s32 %r207, %r1118, %r144, %r203;

and.b32 %r1201, %r1195, 8;
cvt.rn.f32.s32 %f174, %r1201;
mul.ftz.f32 %f175, %f174, 0f3D000000;
sub.ftz.f32 %f176, %f131, %f175;

dp4a.s32.s32 %r211, %r140, %r1127, %r1124;


dp4a.s32.s32 %r215, %r144, %r1127, %r211;

cvt.rn.f32.s32 %f177, %r215;
fma.rn.ftz.f32 %f178, %f176, %f177, %f173;
add.s64 %rd59, %rd154, %rd4;
ld.global.nc.u16 %rs42, [%rd59+48];
shr.u16 %rs43, %rs42, 12;
ld.global.nc.u16 %rs44, [%rd59+50];
shr.u16 %rs45, %rs44, 8;
and.b16 %rs46, %rs45, 240;
or.b16 %rs47, %rs46, %rs43;
ld.global.nc.u16 %rs48, [%rd59+52];
shr.u16 %rs49, %rs48, 4;
and.b16 %rs50, %rs49, 3840;
or.b16 %rs51, %rs47, %rs50;
ld.global.nc.u16 %rs52, [%rd59+54];
and.b16 %rs53, %rs52, -4096;
or.b16 %rs16, %rs51, %rs53;

{ cvt.f32.f16 %f99, %rs16;}



{.reg .f16 low,high;
mov.b32 {low,high},%r154;
cvt.f32.f16 %f100, low;}


mul.ftz.f32 %f179, %f99, %f100;
add.s64 %rd60, %rd154, %rd8;
ld.global.nc.u16 %rs54, [%rd60];
cvt.u32.u16 %r1202, %rs54;
shr.u32 %r1203, %r1202, %r5;
shl.b32 %r1204, %r1203, 1;
and.b32 %r1205, %r1204, 14;
or.b32 %r1206, %r1205, 1;
shr.u32 %r1207, %r1203, 2;
and.b32 %r1208, %r1207, 14;
or.b32 %r1209, %r1208, 1;
cvt.rn.f32.s32 %f180, %r175;
add.ftz.f32 %f181, %f168, %f180;
cvt.rn.f32.s32 %f182, %r1206;
cvt.rn.f32.s32 %f183, %r207;
add.ftz.f32 %f184, %f178, %f183;
cvt.rn.f32.s32 %f185, %r1209;
mul.ftz.f32 %f186, %f184, %f185;
fma.rn.ftz.f32 %f187, %f181, %f182, %f186;
fma.rn.ftz.f32 %f604, %f179, %f187, %f604;
cvt.u64.u32 %rd61, %r1581;
add.s64 %rd62, %rd1, %rd61;
add.s64 %rd63, %rd62, %rd3;
mul.lo.s64 %rd64, %rd63, 36;
add.s64 %rd65, %rd12, %rd64;
ld.global.nc.u32 %r222, [%rd65+4];
ld.global.nc.u32 %r226, [%rd65+8];

dp4a.s32.s32 %r220, %r1001, %r222, %r1124;


dp4a.s32.s32 %r224, %r1005, %r226, %r220;


dp4a.s32.s32 %r228, %r222, %r1127, %r1124;


dp4a.s32.s32 %r232, %r226, %r1127, %r228;

cvt.rn.f32.s32 %f188, %r232;
fma.rn.ftz.f32 %f189, %f132, %f188, 0f00000000;
ld.global.nc.u32 %r238, [%rd65+12];
ld.global.nc.u32 %r242, [%rd65+16];

dp4a.s32.s32 %r236, %r1017, %r238, %r224;


dp4a.s32.s32 %r240, %r1021, %r242, %r236;


dp4a.s32.s32 %r244, %r238, %r1127, %r1124;


dp4a.s32.s32 %r248, %r242, %r1127, %r244;

cvt.rn.f32.s32 %f190, %r248;
fma.rn.ftz.f32 %f191, %f137, %f190, %f189;
ld.global.nc.u32 %r254, [%rd65+20];
ld.global.nc.u32 %r258, [%rd65+24];

dp4a.s32.s32 %r252, %r1033, %r254, %r1124;


dp4a.s32.s32 %r256, %r1037, %r258, %r252;


dp4a.s32.s32 %r260, %r254, %r1127, %r1124;


dp4a.s32.s32 %r264, %r258, %r1127, %r260;

cvt.rn.f32.s32 %f192, %r264;
fma.rn.ftz.f32 %f193, %f142, %f192, 0f00000000;
ld.global.nc.u32 %r270, [%rd65+28];
ld.global.nc.u32 %r274, [%rd65+32];

dp4a.s32.s32 %r268, %r1049, %r270, %r256;


dp4a.s32.s32 %r272, %r1053, %r274, %r268;


dp4a.s32.s32 %r276, %r270, %r1127, %r1124;


dp4a.s32.s32 %r280, %r274, %r1127, %r276;

cvt.rn.f32.s32 %f194, %r280;
fma.rn.ftz.f32 %f195, %f147, %f194, %f193;

{ cvt.f32.f16 %f101, %rs15;}


ld.global.nc.u32 %r284, [%rd65];

{.reg .f16 low,high;
mov.b32 {low,high},%r284;
cvt.f32.f16 %f102, low;}


mul.ftz.f32 %f196, %f101, %f102;
cvt.rn.f32.s32 %f197, %r240;
add.ftz.f32 %f198, %f191, %f197;
cvt.rn.f32.s32 %f199, %r272;
add.ftz.f32 %f200, %f195, %f199;
mul.ftz.f32 %f201, %f200, %f156;
fma.rn.ftz.f32 %f202, %f198, %f153, %f201;
fma.rn.ftz.f32 %f603, %f196, %f202, %f603;

dp4a.s32.s32 %r285, %r1066, %r222, %r1124;


dp4a.s32.s32 %r289, %r1070, %r226, %r285;


dp4a.s32.s32 %r293, %r222, %r1127, %r1124;


dp4a.s32.s32 %r297, %r226, %r1127, %r293;

cvt.rn.f32.s32 %f203, %r297;
fma.rn.ftz.f32 %f204, %f161, %f203, 0f00000000;

dp4a.s32.s32 %r301, %r1082, %r238, %r289;


dp4a.s32.s32 %r305, %r1086, %r242, %r301;


dp4a.s32.s32 %r309, %r238, %r1127, %r1124;


dp4a.s32.s32 %r313, %r242, %r1127, %r309;

cvt.rn.f32.s32 %f205, %r313;
fma.rn.ftz.f32 %f206, %f166, %f205, %f204;

dp4a.s32.s32 %r317, %r1098, %r254, %r1124;


dp4a.s32.s32 %r321, %r1102, %r258, %r317;


dp4a.s32.s32 %r325, %r254, %r1127, %r1124;


dp4a.s32.s32 %r329, %r258, %r1127, %r325;

cvt.rn.f32.s32 %f207, %r329;
fma.rn.ftz.f32 %f208, %f171, %f207, 0f00000000;

dp4a.s32.s32 %r333, %r1114, %r270, %r321;


dp4a.s32.s32 %r337, %r1118, %r274, %r333;


dp4a.s32.s32 %r341, %r270, %r1127, %r1124;


dp4a.s32.s32 %r345, %r274, %r1127, %r341;

cvt.rn.f32.s32 %f209, %r345;
fma.rn.ftz.f32 %f210, %f176, %f209, %f208;

{ cvt.f32.f16 %f103, %rs16;}



{.reg .f16 low,high;
mov.b32 {low,high},%r284;
cvt.f32.f16 %f104, low;}


mul.ftz.f32 %f211, %f103, %f104;
cvt.rn.f32.s32 %f212, %r305;
add.ftz.f32 %f213, %f206, %f212;
cvt.rn.f32.s32 %f214, %r337;
add.ftz.f32 %f215, %f210, %f214;
mul.ftz.f32 %f216, %f215, %f185;
fma.rn.ftz.f32 %f217, %f213, %f182, %f216;
fma.rn.ftz.f32 %f602, %f211, %f217, %f602;
cvt.u64.u32 %rd66, %r1587;
add.s64 %rd67, %rd1, %rd66;
add.s64 %rd68, %rd67, %rd3;
mul.lo.s64 %rd69, %rd68, 36;
add.s64 %rd70, %rd12, %rd69;
ld.global.nc.u32 %r352, [%rd70+4];
ld.global.nc.u32 %r356, [%rd70+8];

dp4a.s32.s32 %r350, %r1001, %r352, %r1124;


dp4a.s32.s32 %r354, %r1005, %r356, %r350;


dp4a.s32.s32 %r358, %r352, %r1127, %r1124;


dp4a.s32.s32 %r362, %r356, %r1127, %r358;

cvt.rn.f32.s32 %f218, %r362;
fma.rn.ftz.f32 %f219, %f132, %f218, 0f00000000;
ld.global.nc.u32 %r368, [%rd70+12];
ld.global.nc.u32 %r372, [%rd70+16];

dp4a.s32.s32 %r366, %r1017, %r368, %r354;


dp4a.s32.s32 %r370, %r1021, %r372, %r366;


dp4a.s32.s32 %r374, %r368, %r1127, %r1124;


dp4a.s32.s32 %r378, %r372, %r1127, %r374;

cvt.rn.f32.s32 %f220, %r378;
fma.rn.ftz.f32 %f221, %f137, %f220, %f219;
ld.global.nc.u32 %r384, [%rd70+20];
ld.global.nc.u32 %r388, [%rd70+24];

dp4a.s32.s32 %r382, %r1033, %r384, %r1124;


dp4a.s32.s32 %r386, %r1037, %r388, %r382;


dp4a.s32.s32 %r390, %r384, %r1127, %r1124;


dp4a.s32.s32 %r394, %r388, %r1127, %r390;

cvt.rn.f32.s32 %f222, %r394;
fma.rn.ftz.f32 %f223, %f142, %f222, 0f00000000;
ld.global.nc.u32 %r400, [%rd70+28];
ld.global.nc.u32 %r404, [%rd70+32];

dp4a.s32.s32 %r398, %r1049, %r400, %r386;


dp4a.s32.s32 %r402, %r1053, %r404, %r398;


dp4a.s32.s32 %r406, %r400, %r1127, %r1124;


dp4a.s32.s32 %r410, %r404, %r1127, %r406;

cvt.rn.f32.s32 %f224, %r410;
fma.rn.ftz.f32 %f225, %f147, %f224, %f223;

{ cvt.f32.f16 %f105, %rs15;}


ld.global.nc.u32 %r414, [%rd70];

{.reg .f16 low,high;
mov.b32 {low,high},%r414;
cvt.f32.f16 %f106, low;}


mul.ftz.f32 %f226, %f105, %f106;
cvt.rn.f32.s32 %f227, %r370;
add.ftz.f32 %f228, %f221, %f227;
cvt.rn.f32.s32 %f229, %r402;
add.ftz.f32 %f230, %f225, %f229;
mul.ftz.f32 %f231, %f230, %f156;
fma.rn.ftz.f32 %f232, %f228, %f153, %f231;
fma.rn.ftz.f32 %f601, %f226, %f232, %f601;

dp4a.s32.s32 %r415, %r1066, %r352, %r1124;


dp4a.s32.s32 %r419, %r1070, %r356, %r415;


dp4a.s32.s32 %r423, %r352, %r1127, %r1124;


dp4a.s32.s32 %r427, %r356, %r1127, %r423;

cvt.rn.f32.s32 %f233, %r427;
fma.rn.ftz.f32 %f234, %f161, %f233, 0f00000000;

dp4a.s32.s32 %r431, %r1082, %r368, %r419;


dp4a.s32.s32 %r435, %r1086, %r372, %r431;


dp4a.s32.s32 %r439, %r368, %r1127, %r1124;


dp4a.s32.s32 %r443, %r372, %r1127, %r439;

cvt.rn.f32.s32 %f235, %r443;
fma.rn.ftz.f32 %f236, %f166, %f235, %f234;

dp4a.s32.s32 %r447, %r1098, %r384, %r1124;


dp4a.s32.s32 %r451, %r1102, %r388, %r447;


dp4a.s32.s32 %r455, %r384, %r1127, %r1124;


dp4a.s32.s32 %r459, %r388, %r1127, %r455;

cvt.rn.f32.s32 %f237, %r459;
fma.rn.ftz.f32 %f238, %f171, %f237, 0f00000000;

dp4a.s32.s32 %r463, %r1114, %r400, %r451;


dp4a.s32.s32 %r467, %r1118, %r404, %r463;


dp4a.s32.s32 %r471, %r400, %r1127, %r1124;


dp4a.s32.s32 %r475, %r404, %r1127, %r471;

cvt.rn.f32.s32 %f239, %r475;
fma.rn.ftz.f32 %f240, %f176, %f239, %f238;

{ cvt.f32.f16 %f107, %rs16;}



{.reg .f16 low,high;
mov.b32 {low,high},%r414;
cvt.f32.f16 %f108, low;}


mul.ftz.f32 %f241, %f107, %f108;
cvt.rn.f32.s32 %f242, %r435;
add.ftz.f32 %f243, %f236, %f242;
cvt.rn.f32.s32 %f244, %r467;
add.ftz.f32 %f245, %f240, %f244;
mul.ftz.f32 %f246, %f245, %f185;
fma.rn.ftz.f32 %f247, %f243, %f182, %f246;
fma.rn.ftz.f32 %f600, %f241, %f247, %f600;
cvt.u64.u32 %rd71, %r1586;
add.s64 %rd72, %rd1, %rd71;
add.s64 %rd73, %rd72, %rd3;
mul.lo.s64 %rd74, %rd73, 36;
add.s64 %rd75, %rd12, %rd74;
ld.global.nc.u32 %r482, [%rd75+4];
ld.global.nc.u32 %r486, [%rd75+8];

dp4a.s32.s32 %r480, %r1001, %r482, %r1124;


dp4a.s32.s32 %r484, %r1005, %r486, %r480;


dp4a.s32.s32 %r488, %r482, %r1127, %r1124;


dp4a.s32.s32 %r492, %r486, %r1127, %r488;

cvt.rn.f32.s32 %f248, %r492;
fma.rn.ftz.f32 %f249, %f132, %f248, 0f00000000;
ld.global.nc.u32 %r498, [%rd75+12];
ld.global.nc.u32 %r502, [%rd75+16];

dp4a.s32.s32 %r496, %r1017, %r498, %r484;


dp4a.s32.s32 %r500, %r1021, %r502, %r496;


dp4a.s32.s32 %r504, %r498, %r1127, %r1124;


dp4a.s32.s32 %r508, %r502, %r1127, %r504;

cvt.rn.f32.s32 %f250, %r508;
fma.rn.ftz.f32 %f251, %f137, %f250, %f249;
ld.global.nc.u32 %r514, [%rd75+20];
ld.global.nc.u32 %r518, [%rd75+24];

dp4a.s32.s32 %r512, %r1033, %r514, %r1124;


dp4a.s32.s32 %r516, %r1037, %r518, %r512;


dp4a.s32.s32 %r520, %r514, %r1127, %r1124;


dp4a.s32.s32 %r524, %r518, %r1127, %r520;

cvt.rn.f32.s32 %f252, %r524;
fma.rn.ftz.f32 %f253, %f142, %f252, 0f00000000;
ld.global.nc.u32 %r530, [%rd75+28];
ld.global.nc.u32 %r534, [%rd75+32];

dp4a.s32.s32 %r528, %r1049, %r530, %r516;


dp4a.s32.s32 %r532, %r1053, %r534, %r528;


dp4a.s32.s32 %r536, %r530, %r1127, %r1124;


dp4a.s32.s32 %r540, %r534, %r1127, %r536;

cvt.rn.f32.s32 %f254, %r540;
fma.rn.ftz.f32 %f255, %f147, %f254, %f253;

{ cvt.f32.f16 %f109, %rs15;}


ld.global.nc.u32 %r544, [%rd75];

{.reg .f16 low,high;
mov.b32 {low,high},%r544;
cvt.f32.f16 %f110, low;}


mul.ftz.f32 %f256, %f109, %f110;
cvt.rn.f32.s32 %f257, %r500;
add.ftz.f32 %f258, %f251, %f257;
cvt.rn.f32.s32 %f259, %r532;
add.ftz.f32 %f260, %f255, %f259;
mul.ftz.f32 %f261, %f260, %f156;
fma.rn.ftz.f32 %f262, %f258, %f153, %f261;
fma.rn.ftz.f32 %f599, %f256, %f262, %f599;

dp4a.s32.s32 %r545, %r1066, %r482, %r1124;


dp4a.s32.s32 %r549, %r1070, %r486, %r545;


dp4a.s32.s32 %r553, %r482, %r1127, %r1124;


dp4a.s32.s32 %r557, %r486, %r1127, %r553;

cvt.rn.f32.s32 %f263, %r557;
fma.rn.ftz.f32 %f264, %f161, %f263, 0f00000000;

dp4a.s32.s32 %r561, %r1082, %r498, %r549;


dp4a.s32.s32 %r565, %r1086, %r502, %r561;


dp4a.s32.s32 %r569, %r498, %r1127, %r1124;


dp4a.s32.s32 %r573, %r502, %r1127, %r569;

cvt.rn.f32.s32 %f265, %r573;
fma.rn.ftz.f32 %f266, %f166, %f265, %f264;

dp4a.s32.s32 %r577, %r1098, %r514, %r1124;


dp4a.s32.s32 %r581, %r1102, %r518, %r577;


dp4a.s32.s32 %r585, %r514, %r1127, %r1124;


dp4a.s32.s32 %r589, %r518, %r1127, %r585;

cvt.rn.f32.s32 %f267, %r589;
fma.rn.ftz.f32 %f268, %f171, %f267, 0f00000000;

dp4a.s32.s32 %r593, %r1114, %r530, %r581;


dp4a.s32.s32 %r597, %r1118, %r534, %r593;


dp4a.s32.s32 %r601, %r530, %r1127, %r1124;


dp4a.s32.s32 %r605, %r534, %r1127, %r601;

cvt.rn.f32.s32 %f269, %r605;
fma.rn.ftz.f32 %f270, %f176, %f269, %f268;

{ cvt.f32.f16 %f111, %rs16;}



{.reg .f16 low,high;
mov.b32 {low,high},%r544;
cvt.f32.f16 %f112, low;}


mul.ftz.f32 %f271, %f111, %f112;
cvt.rn.f32.s32 %f272, %r565;
add.ftz.f32 %f273, %f266, %f272;
cvt.rn.f32.s32 %f274, %r597;
add.ftz.f32 %f275, %f270, %f274;
mul.ftz.f32 %f276, %f275, %f185;
fma.rn.ftz.f32 %f277, %f273, %f182, %f276;
fma.rn.ftz.f32 %f598, %f271, %f277, %f598;
cvt.u64.u32 %rd76, %r1585;
add.s64 %rd77, %rd1, %rd76;
add.s64 %rd78, %rd77, %rd3;
mul.lo.s64 %rd79, %rd78, 36;
add.s64 %rd80, %rd12, %rd79;
ld.global.nc.u32 %r612, [%rd80+4];
ld.global.nc.u32 %r616, [%rd80+8];

dp4a.s32.s32 %r610, %r1001, %r612, %r1124;


dp4a.s32.s32 %r614, %r1005, %r616, %r610;


dp4a.s32.s32 %r618, %r612, %r1127, %r1124;


dp4a.s32.s32 %r622, %r616, %r1127, %r618;

cvt.rn.f32.s32 %f278, %r622;
fma.rn.ftz.f32 %f279, %f132, %f278, 0f00000000;
ld.global.nc.u32 %r628, [%rd80+12];
ld.global.nc.u32 %r632, [%rd80+16];

dp4a.s32.s32 %r626, %r1017, %r628, %r614;


dp4a.s32.s32 %r630, %r1021, %r632, %r626;


dp4a.s32.s32 %r634, %r628, %r1127, %r1124;


dp4a.s32.s32 %r638, %r632, %r1127, %r634;

cvt.rn.f32.s32 %f280, %r638;
fma.rn.ftz.f32 %f281, %f137, %f280, %f279;
ld.global.nc.u32 %r644, [%rd80+20];
ld.global.nc.u32 %r648, [%rd80+24];

dp4a.s32.s32 %r642, %r1033, %r644, %r1124;


dp4a.s32.s32 %r646, %r1037, %r648, %r642;


dp4a.s32.s32 %r650, %r644, %r1127, %r1124;


dp4a.s32.s32 %r654, %r648, %r1127, %r650;

cvt.rn.f32.s32 %f282, %r654;
fma.rn.ftz.f32 %f283, %f142, %f282, 0f00000000;
ld.global.nc.u32 %r660, [%rd80+28];
ld.global.nc.u32 %r664, [%rd80+32];

dp4a.s32.s32 %r658, %r1049, %r660, %r646;


dp4a.s32.s32 %r662, %r1053, %r664, %r658;


dp4a.s32.s32 %r666, %r660, %r1127, %r1124;


dp4a.s32.s32 %r670, %r664, %r1127, %r666;

cvt.rn.f32.s32 %f284, %r670;
fma.rn.ftz.f32 %f285, %f147, %f284, %f283;

{ cvt.f32.f16 %f113, %rs15;}


ld.global.nc.u32 %r674, [%rd80];

{.reg .f16 low,high;
mov.b32 {low,high},%r674;
cvt.f32.f16 %f114, low;}


mul.ftz.f32 %f286, %f113, %f114;
cvt.rn.f32.s32 %f287, %r630;
add.ftz.f32 %f288, %f281, %f287;
cvt.rn.f32.s32 %f289, %r662;
add.ftz.f32 %f290, %f285, %f289;
mul.ftz.f32 %f291, %f290, %f156;
fma.rn.ftz.f32 %f292, %f288, %f153, %f291;
fma.rn.ftz.f32 %f597, %f286, %f292, %f597;

dp4a.s32.s32 %r675, %r1066, %r612, %r1124;


dp4a.s32.s32 %r679, %r1070, %r616, %r675;


dp4a.s32.s32 %r683, %r612, %r1127, %r1124;


dp4a.s32.s32 %r687, %r616, %r1127, %r683;

cvt.rn.f32.s32 %f293, %r687;
fma.rn.ftz.f32 %f294, %f161, %f293, 0f00000000;

dp4a.s32.s32 %r691, %r1082, %r628, %r679;


dp4a.s32.s32 %r695, %r1086, %r632, %r691;


dp4a.s32.s32 %r699, %r628, %r1127, %r1124;


dp4a.s32.s32 %r703, %r632, %r1127, %r699;

cvt.rn.f32.s32 %f295, %r703;
fma.rn.ftz.f32 %f296, %f166, %f295, %f294;

dp4a.s32.s32 %r707, %r1098, %r644, %r1124;


dp4a.s32.s32 %r711, %r1102, %r648, %r707;


dp4a.s32.s32 %r715, %r644, %r1127, %r1124;


dp4a.s32.s32 %r719, %r648, %r1127, %r715;

cvt.rn.f32.s32 %f297, %r719;
fma.rn.ftz.f32 %f298, %f171, %f297, 0f00000000;

dp4a.s32.s32 %r723, %r1114, %r660, %r711;


dp4a.s32.s32 %r727, %r1118, %r664, %r723;


dp4a.s32.s32 %r731, %r660, %r1127, %r1124;


dp4a.s32.s32 %r735, %r664, %r1127, %r731;

cvt.rn.f32.s32 %f299, %r735;
fma.rn.ftz.f32 %f300, %f176, %f299, %f298;

{ cvt.f32.f16 %f115, %rs16;}



{.reg .f16 low,high;
mov.b32 {low,high},%r674;
cvt.f32.f16 %f116, low;}


mul.ftz.f32 %f301, %f115, %f116;
cvt.rn.f32.s32 %f302, %r695;
add.ftz.f32 %f303, %f296, %f302;
cvt.rn.f32.s32 %f304, %r727;
add.ftz.f32 %f305, %f300, %f304;
mul.ftz.f32 %f306, %f305, %f185;
fma.rn.ftz.f32 %f307, %f303, %f182, %f306;
fma.rn.ftz.f32 %f596, %f301, %f307, %f596;
cvt.u64.u32 %rd81, %r1584;
add.s64 %rd82, %rd1, %rd81;
add.s64 %rd83, %rd82, %rd3;
mul.lo.s64 %rd84, %rd83, 36;
add.s64 %rd85, %rd12, %rd84;
ld.global.nc.u32 %r742, [%rd85+4];
ld.global.nc.u32 %r746, [%rd85+8];

dp4a.s32.s32 %r740, %r1001, %r742, %r1124;


dp4a.s32.s32 %r744, %r1005, %r746, %r740;


dp4a.s32.s32 %r748, %r742, %r1127, %r1124;


dp4a.s32.s32 %r752, %r746, %r1127, %r748;

cvt.rn.f32.s32 %f308, %r752;
fma.rn.ftz.f32 %f309, %f132, %f308, 0f00000000;
ld.global.nc.u32 %r758, [%rd85+12];
ld.global.nc.u32 %r762, [%rd85+16];

dp4a.s32.s32 %r756, %r1017, %r758, %r744;


dp4a.s32.s32 %r760, %r1021, %r762, %r756;


dp4a.s32.s32 %r764, %r758, %r1127, %r1124;


dp4a.s32.s32 %r768, %r762, %r1127, %r764;

cvt.rn.f32.s32 %f310, %r768;
fma.rn.ftz.f32 %f311, %f137, %f310, %f309;
ld.global.nc.u32 %r774, [%rd85+20];
ld.global.nc.u32 %r778, [%rd85+24];

dp4a.s32.s32 %r772, %r1033, %r774, %r1124;


dp4a.s32.s32 %r776, %r1037, %r778, %r772;


dp4a.s32.s32 %r780, %r774, %r1127, %r1124;


dp4a.s32.s32 %r784, %r778, %r1127, %r780;

cvt.rn.f32.s32 %f312, %r784;
fma.rn.ftz.f32 %f313, %f142, %f312, 0f00000000;
ld.global.nc.u32 %r790, [%rd85+28];
ld.global.nc.u32 %r794, [%rd85+32];

dp4a.s32.s32 %r788, %r1049, %r790, %r776;


dp4a.s32.s32 %r792, %r1053, %r794, %r788;


dp4a.s32.s32 %r796, %r790, %r1127, %r1124;


dp4a.s32.s32 %r800, %r794, %r1127, %r796;

cvt.rn.f32.s32 %f314, %r800;
fma.rn.ftz.f32 %f315, %f147, %f314, %f313;

{ cvt.f32.f16 %f117, %rs15;}


ld.global.nc.u32 %r804, [%rd85];

{.reg .f16 low,high;
mov.b32 {low,high},%r804;
cvt.f32.f16 %f118, low;}


mul.ftz.f32 %f316, %f117, %f118;
cvt.rn.f32.s32 %f317, %r760;
add.ftz.f32 %f318, %f311, %f317;
cvt.rn.f32.s32 %f319, %r792;
add.ftz.f32 %f320, %f315, %f319;
mul.ftz.f32 %f321, %f320, %f156;
fma.rn.ftz.f32 %f322, %f318, %f153, %f321;
fma.rn.ftz.f32 %f595, %f316, %f322, %f595;

dp4a.s32.s32 %r805, %r1066, %r742, %r1124;


dp4a.s32.s32 %r809, %r1070, %r746, %r805;


dp4a.s32.s32 %r813, %r742, %r1127, %r1124;


dp4a.s32.s32 %r817, %r746, %r1127, %r813;

cvt.rn.f32.s32 %f323, %r817;
fma.rn.ftz.f32 %f324, %f161, %f323, 0f00000000;

dp4a.s32.s32 %r821, %r1082, %r758, %r809;


dp4a.s32.s32 %r825, %r1086, %r762, %r821;


dp4a.s32.s32 %r829, %r758, %r1127, %r1124;


dp4a.s32.s32 %r833, %r762, %r1127, %r829;

cvt.rn.f32.s32 %f325, %r833;
fma.rn.ftz.f32 %f326, %f166, %f325, %f324;

dp4a.s32.s32 %r837, %r1098, %r774, %r1124;


dp4a.s32.s32 %r841, %r1102, %r778, %r837;


dp4a.s32.s32 %r845, %r774, %r1127, %r1124;


dp4a.s32.s32 %r849, %r778, %r1127, %r845;

cvt.rn.f32.s32 %f327, %r849;
fma.rn.ftz.f32 %f328, %f171, %f327, 0f00000000;

dp4a.s32.s32 %r853, %r1114, %r790, %r841;


dp4a.s32.s32 %r857, %r1118, %r794, %r853;


dp4a.s32.s32 %r861, %r790, %r1127, %r1124;


dp4a.s32.s32 %r865, %r794, %r1127, %r861;

cvt.rn.f32.s32 %f329, %r865;
fma.rn.ftz.f32 %f330, %f176, %f329, %f328;

{ cvt.f32.f16 %f119, %rs16;}



{.reg .f16 low,high;
mov.b32 {low,high},%r804;
cvt.f32.f16 %f120, low;}


mul.ftz.f32 %f331, %f119, %f120;
cvt.rn.f32.s32 %f332, %r825;
add.ftz.f32 %f333, %f326, %f332;
cvt.rn.f32.s32 %f334, %r857;
add.ftz.f32 %f335, %f330, %f334;
mul.ftz.f32 %f336, %f335, %f185;
fma.rn.ftz.f32 %f337, %f333, %f182, %f336;
fma.rn.ftz.f32 %f594, %f331, %f337, %f594;
cvt.u64.u32 %rd86, %r1583;
add.s64 %rd87, %rd1, %rd86;
add.s64 %rd88, %rd87, %rd3;
mul.lo.s64 %rd89, %rd88, 36;
add.s64 %rd90, %rd12, %rd89;
ld.global.nc.u32 %r872, [%rd90+4];
ld.global.nc.u32 %r876, [%rd90+8];

dp4a.s32.s32 %r870, %r1001, %r872, %r1124;


dp4a.s32.s32 %r874, %r1005, %r876, %r870;


dp4a.s32.s32 %r878, %r872, %r1127, %r1124;


dp4a.s32.s32 %r882, %r876, %r1127, %r878;

cvt.rn.f32.s32 %f338, %r882;
fma.rn.ftz.f32 %f339, %f132, %f338, 0f00000000;
ld.global.nc.u32 %r888, [%rd90+12];
ld.global.nc.u32 %r892, [%rd90+16];

dp4a.s32.s32 %r886, %r1017, %r888, %r874;


dp4a.s32.s32 %r890, %r1021, %r892, %r886;


dp4a.s32.s32 %r894, %r888, %r1127, %r1124;


dp4a.s32.s32 %r898, %r892, %r1127, %r894;

cvt.rn.f32.s32 %f340, %r898;
fma.rn.ftz.f32 %f341, %f137, %f340, %f339;
ld.global.nc.u32 %r904, [%rd90+20];
ld.global.nc.u32 %r908, [%rd90+24];

dp4a.s32.s32 %r902, %r1033, %r904, %r1124;


dp4a.s32.s32 %r906, %r1037, %r908, %r902;


dp4a.s32.s32 %r910, %r904, %r1127, %r1124;


dp4a.s32.s32 %r914, %r908, %r1127, %r910;

cvt.rn.f32.s32 %f342, %r914;
fma.rn.ftz.f32 %f343, %f142, %f342, 0f00000000;
ld.global.nc.u32 %r920, [%rd90+28];
ld.global.nc.u32 %r924, [%rd90+32];

dp4a.s32.s32 %r918, %r1049, %r920, %r906;


dp4a.s32.s32 %r922, %r1053, %r924, %r918;


dp4a.s32.s32 %r926, %r920, %r1127, %r1124;


dp4a.s32.s32 %r930, %r924, %r1127, %r926;

cvt.rn.f32.s32 %f344, %r930;
fma.rn.ftz.f32 %f345, %f147, %f344, %f343;

{ cvt.f32.f16 %f121, %rs15;}


ld.global.nc.u32 %r934, [%rd90];

{.reg .f16 low,high;
mov.b32 {low,high},%r934;
cvt.f32.f16 %f122, low;}


mul.ftz.f32 %f346, %f121, %f122;
cvt.rn.f32.s32 %f347, %r890;
add.ftz.f32 %f348, %f341, %f347;
cvt.rn.f32.s32 %f349, %r922;
add.ftz.f32 %f350, %f345, %f349;
mul.ftz.f32 %f351, %f350, %f156;
fma.rn.ftz.f32 %f352, %f348, %f153, %f351;
fma.rn.ftz.f32 %f593, %f346, %f352, %f593;

dp4a.s32.s32 %r935, %r1066, %r872, %r1124;


dp4a.s32.s32 %r939, %r1070, %r876, %r935;


dp4a.s32.s32 %r943, %r872, %r1127, %r1124;


dp4a.s32.s32 %r947, %r876, %r1127, %r943;

cvt.rn.f32.s32 %f353, %r947;
fma.rn.ftz.f32 %f354, %f161, %f353, 0f00000000;

dp4a.s32.s32 %r951, %r1082, %r888, %r939;


dp4a.s32.s32 %r955, %r1086, %r892, %r951;


dp4a.s32.s32 %r959, %r888, %r1127, %r1124;


dp4a.s32.s32 %r963, %r892, %r1127, %r959;

cvt.rn.f32.s32 %f355, %r963;
fma.rn.ftz.f32 %f356, %f166, %f355, %f354;

dp4a.s32.s32 %r967, %r1098, %r904, %r1124;


dp4a.s32.s32 %r971, %r1102, %r908, %r967;


dp4a.s32.s32 %r975, %r904, %r1127, %r1124;


dp4a.s32.s32 %r979, %r908, %r1127, %r975;

cvt.rn.f32.s32 %f357, %r979;
fma.rn.ftz.f32 %f358, %f171, %f357, 0f00000000;

dp4a.s32.s32 %r983, %r1114, %r920, %r971;


dp4a.s32.s32 %r987, %r1118, %r924, %r983;


dp4a.s32.s32 %r991, %r920, %r1127, %r1124;


dp4a.s32.s32 %r995, %r924, %r1127, %r991;

cvt.rn.f32.s32 %f359, %r995;
fma.rn.ftz.f32 %f360, %f176, %f359, %f358;

{ cvt.f32.f16 %f123, %rs16;}



{.reg .f16 low,high;
mov.b32 {low,high},%r934;
cvt.f32.f16 %f124, low;}


mul.ftz.f32 %f361, %f123, %f124;
cvt.rn.f32.s32 %f362, %r955;
add.ftz.f32 %f363, %f356, %f362;
cvt.rn.f32.s32 %f364, %r987;
add.ftz.f32 %f365, %f360, %f364;
mul.ftz.f32 %f366, %f365, %f185;
fma.rn.ftz.f32 %f367, %f363, %f182, %f366;
fma.rn.ftz.f32 %f592, %f361, %f367, %f592;
cvt.u64.u32 %rd91, %r1582;
add.s64 %rd92, %rd1, %rd91;
add.s64 %rd93, %rd92, %rd3;
mul.lo.s64 %rd94, %rd93, 36;
add.s64 %rd95, %rd12, %rd94;
ld.global.nc.u32 %r1002, [%rd95+4];
ld.global.nc.u32 %r1006, [%rd95+8];

dp4a.s32.s32 %r1000, %r1001, %r1002, %r1124;


dp4a.s32.s32 %r1004, %r1005, %r1006, %r1000;


dp4a.s32.s32 %r1008, %r1002, %r1127, %r1124;


dp4a.s32.s32 %r1012, %r1006, %r1127, %r1008;

cvt.rn.f32.s32 %f368, %r1012;
fma.rn.ftz.f32 %f369, %f132, %f368, 0f00000000;
ld.global.nc.u32 %r1018, [%rd95+12];
ld.global.nc.u32 %r1022, [%rd95+16];

dp4a.s32.s32 %r1016, %r1017, %r1018, %r1004;


dp4a.s32.s32 %r1020, %r1021, %r1022, %r1016;


dp4a.s32.s32 %r1024, %r1018, %r1127, %r1124;


dp4a.s32.s32 %r1028, %r1022, %r1127, %r1024;

cvt.rn.f32.s32 %f370, %r1028;
fma.rn.ftz.f32 %f371, %f137, %f370, %f369;
ld.global.nc.u32 %r1034, [%rd95+20];
ld.global.nc.u32 %r1038, [%rd95+24];

dp4a.s32.s32 %r1032, %r1033, %r1034, %r1124;


dp4a.s32.s32 %r1036, %r1037, %r1038, %r1032;


dp4a.s32.s32 %r1040, %r1034, %r1127, %r1124;


dp4a.s32.s32 %r1044, %r1038, %r1127, %r1040;

cvt.rn.f32.s32 %f372, %r1044;
fma.rn.ftz.f32 %f373, %f142, %f372, 0f00000000;
ld.global.nc.u32 %r1050, [%rd95+28];
ld.global.nc.u32 %r1054, [%rd95+32];

dp4a.s32.s32 %r1048, %r1049, %r1050, %r1036;


dp4a.s32.s32 %r1052, %r1053, %r1054, %r1048;


dp4a.s32.s32 %r1056, %r1050, %r1127, %r1124;


dp4a.s32.s32 %r1060, %r1054, %r1127, %r1056;

cvt.rn.f32.s32 %f374, %r1060;
fma.rn.ftz.f32 %f375, %f147, %f374, %f373;

{ cvt.f32.f16 %f125, %rs15;}


ld.global.nc.u32 %r1064, [%rd95];

{.reg .f16 low,high;
mov.b32 {low,high},%r1064;
cvt.f32.f16 %f126, low;}


mul.ftz.f32 %f376, %f125, %f126;
cvt.rn.f32.s32 %f377, %r1020;
add.ftz.f32 %f378, %f371, %f377;
cvt.rn.f32.s32 %f379, %r1052;
add.ftz.f32 %f380, %f375, %f379;
mul.ftz.f32 %f381, %f380, %f156;
fma.rn.ftz.f32 %f382, %f378, %f153, %f381;
fma.rn.ftz.f32 %f591, %f376, %f382, %f591;

dp4a.s32.s32 %r1065, %r1066, %r1002, %r1124;


dp4a.s32.s32 %r1069, %r1070, %r1006, %r1065;


dp4a.s32.s32 %r1073, %r1002, %r1127, %r1124;


dp4a.s32.s32 %r1077, %r1006, %r1127, %r1073;

cvt.rn.f32.s32 %f383, %r1077;
fma.rn.ftz.f32 %f384, %f161, %f383, 0f00000000;

dp4a.s32.s32 %r1081, %r1082, %r1018, %r1069;


dp4a.s32.s32 %r1085, %r1086, %r1022, %r1081;


dp4a.s32.s32 %r1089, %r1018, %r1127, %r1124;


dp4a.s32.s32 %r1093, %r1022, %r1127, %r1089;

cvt.rn.f32.s32 %f385, %r1093;
fma.rn.ftz.f32 %f386, %f166, %f385, %f384;

dp4a.s32.s32 %r1097, %r1098, %r1034, %r1124;


dp4a.s32.s32 %r1101, %r1102, %r1038, %r1097;


dp4a.s32.s32 %r1105, %r1034, %r1127, %r1124;


dp4a.s32.s32 %r1109, %r1038, %r1127, %r1105;

cvt.rn.f32.s32 %f387, %r1109;
fma.rn.ftz.f32 %f388, %f171, %f387, 0f00000000;

dp4a.s32.s32 %r1113, %r1114, %r1050, %r1101;


dp4a.s32.s32 %r1117, %r1118, %r1054, %r1113;


dp4a.s32.s32 %r1121, %r1050, %r1127, %r1124;


dp4a.s32.s32 %r1125, %r1054, %r1127, %r1121;

cvt.rn.f32.s32 %f389, %r1125;
fma.rn.ftz.f32 %f390, %f176, %f389, %f388;

{ cvt.f32.f16 %f127, %rs16;}



{.reg .f16 low,high;
mov.b32 {low,high},%r1064;
cvt.f32.f16 %f128, low;}


mul.ftz.f32 %f391, %f127, %f128;
cvt.rn.f32.s32 %f392, %r1085;
add.ftz.f32 %f393, %f386, %f392;
cvt.rn.f32.s32 %f394, %r1117;
add.ftz.f32 %f395, %f390, %f394;
mul.ftz.f32 %f396, %f395, %f185;
fma.rn.ftz.f32 %f397, %f393, %f182, %f396;
fma.rn.ftz.f32 %f590, %f391, %f397, %f590;
add.s32 %r1587, %r1587, 64;
add.s32 %r1586, %r1586, 64;
add.s32 %r1585, %r1585, 64;
add.s32 %r1584, %r1584, 64;
add.s32 %r1583, %r1583, 64;
add.s32 %r1582, %r1582, 64;
add.s32 %r1581, %r1581, 64;
add.s32 %r1580, %r1580, 64;
add.s64 %rd154, %rd154, 448;
add.s32 %r1588, %r1588, 8;
setp.lt.s32 %p4, %r1588, %r4;
@%p4 bra $L__BB275_2;

$L__BB275_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB275_5;

shl.b32 %r1211, %r1, 11;
mov.u32 %r1212, _ZZ13mul_mat_vec_qIL9ggml_type29ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r1213, %r1212, %r1211;
shl.b32 %r1214, %r52, 2;
add.s32 %r1215, %r1213, %r1214;
st.shared.f32 [%r1215+-2048], %f605;
st.shared.f32 [%r1215+-1920], %f604;
st.shared.f32 [%r1215+-1792], %f603;
st.shared.f32 [%r1215+-1664], %f602;
st.shared.f32 [%r1215+-1536], %f601;
st.shared.f32 [%r1215+-1408], %f600;
st.shared.f32 [%r1215+-1280], %f599;
st.shared.f32 [%r1215+-1152], %f598;
st.shared.f32 [%r1215+-1024], %f597;
st.shared.f32 [%r1215+-896], %f596;
st.shared.f32 [%r1215+-768], %f595;
st.shared.f32 [%r1215+-640], %f594;
st.shared.f32 [%r1215+-512], %f593;
st.shared.f32 [%r1215+-384], %f592;
st.shared.f32 [%r1215+-256], %f591;
st.shared.f32 [%r1215+-128], %f590;

$L__BB275_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB275_38;

mov.u32 %r1216, %ctaid.z;
mov.u32 %r1217, %ctaid.y;
mov.u32 %r1218, %ctaid.x;
shl.b32 %r1219, %r1218, 1;
mov.u32 %r1220, 1;
mad.lo.s32 %r1221, %r1217, %r43, %r1219;
mad.lo.s32 %r1222, %r1216, %r49, %r1221;
cvt.u64.u32 %rd15, %r1222;
shl.b32 %r1224, %r52, 2;
mov.u32 %r1225, _ZZ13mul_mat_vec_qIL9ggml_type29ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r1226, %r1225, %r1224;
mov.u32 %r1227, 2;
ld.shared.f32 %f398, [%r1226];
add.ftz.f32 %f399, %f398, %f605;
mov.b32 %r1228, %f399;
mov.u32 %r1229, 31;
mov.u32 %r1230, 16;
mov.u32 %r1231, -1;
shfl.sync.bfly.b32 %r1232|%p7, %r1228, %r1230, %r1229, %r1231;
mov.b32 %f400, %r1232;
add.ftz.f32 %f401, %f399, %f400;
mov.b32 %r1233, %f401;
mov.u32 %r1234, 8;
shfl.sync.bfly.b32 %r1235|%p8, %r1233, %r1234, %r1229, %r1231;
mov.b32 %f402, %r1235;
add.ftz.f32 %f403, %f401, %f402;
mov.b32 %r1236, %f403;
mov.u32 %r1237, 4;
shfl.sync.bfly.b32 %r1238|%p9, %r1236, %r1237, %r1229, %r1231;
mov.b32 %f404, %r1238;
add.ftz.f32 %f405, %f403, %f404;
mov.b32 %r1239, %f405;
shfl.sync.bfly.b32 %r1240|%p10, %r1239, %r1227, %r1229, %r1231;
mov.b32 %f406, %r1240;
add.ftz.f32 %f407, %f405, %f406;
mov.b32 %r1241, %f407;
shfl.sync.bfly.b32 %r1242|%p11, %r1241, %r1220, %r1229, %r1231;
mov.b32 %f408, %r1242;
add.ftz.f32 %f49, %f407, %f408;
setp.eq.s32 %p12, %r52, 0;
setp.lt.u32 %p13, %r1219, %r37;
and.pred %p1, %p12, %p13;
not.pred %p14, %p1;
@%p14 bra $L__BB275_8;

cvta.to.global.u64 %rd96, %rd21;
mul.wide.u32 %rd97, %r1222, 4;
add.s64 %rd98, %rd96, %rd97;
st.global.f32 [%rd98], %f49;

$L__BB275_8:
ld.shared.f32 %f409, [%r1226+128];
add.ftz.f32 %f410, %f409, %f604;
mov.b32 %r1254, %f410;
shfl.sync.bfly.b32 %r1258|%p15, %r1254, %r1230, %r1229, %r1231;
mov.b32 %f411, %r1258;
add.ftz.f32 %f412, %f410, %f411;
mov.b32 %r1259, %f412;
shfl.sync.bfly.b32 %r1261|%p16, %r1259, %r1234, %r1229, %r1231;
mov.b32 %f413, %r1261;
add.ftz.f32 %f414, %f412, %f413;
mov.b32 %r1262, %f414;
shfl.sync.bfly.b32 %r1264|%p17, %r1262, %r1237, %r1229, %r1231;
mov.b32 %f415, %r1264;
add.ftz.f32 %f416, %f414, %f415;
mov.b32 %r1265, %f416;
shfl.sync.bfly.b32 %r1266|%p18, %r1265, %r1227, %r1229, %r1231;
mov.b32 %f417, %r1266;
add.ftz.f32 %f418, %f416, %f417;
mov.b32 %r1267, %f418;
shfl.sync.bfly.b32 %r1269|%p19, %r1267, %r1220, %r1229, %r1231;
mov.b32 %f419, %r1269;
add.ftz.f32 %f50, %f418, %f419;
setp.eq.s32 %p20, %r52, 1;
or.b32 %r1272, %r1219, 1;
setp.lt.u32 %p21, %r1272, %r37;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB275_10;

cvta.to.global.u64 %rd99, %rd21;
mul.wide.u32 %rd100, %r1222, 4;
add.s64 %rd101, %rd99, %rd100;
st.global.f32 [%rd101+4], %f50;

$L__BB275_10:
mov.u32 %r1283, 2;
ld.shared.f32 %f420, [%r1226+256];
add.ftz.f32 %f421, %f420, %f603;
mov.b32 %r1284, %f421;
mov.u32 %r1285, 31;
mov.u32 %r1286, 16;
mov.u32 %r1287, -1;
shfl.sync.bfly.b32 %r1288|%p23, %r1284, %r1286, %r1285, %r1287;
mov.b32 %f422, %r1288;
add.ftz.f32 %f423, %f421, %f422;
mov.b32 %r1289, %f423;
mov.u32 %r1290, 8;
shfl.sync.bfly.b32 %r1291|%p24, %r1289, %r1290, %r1285, %r1287;
mov.b32 %f424, %r1291;
add.ftz.f32 %f425, %f423, %f424;
mov.b32 %r1292, %f425;
mov.u32 %r1293, 4;
shfl.sync.bfly.b32 %r1294|%p25, %r1292, %r1293, %r1285, %r1287;
mov.b32 %f426, %r1294;
add.ftz.f32 %f427, %f425, %f426;
mov.b32 %r1295, %f427;
shfl.sync.bfly.b32 %r1296|%p26, %r1295, %r1283, %r1285, %r1287;
mov.b32 %f428, %r1296;
add.ftz.f32 %f429, %f427, %f428;
mov.b32 %r1297, %f429;
mov.u32 %r1298, 1;
shfl.sync.bfly.b32 %r1299|%p27, %r1297, %r1298, %r1285, %r1287;
mov.b32 %f430, %r1299;
add.ftz.f32 %f51, %f429, %f430;
@%p14 bra $L__BB275_12;

cvt.u64.u32 %rd102, %r37;
add.s64 %rd103, %rd102, %rd15;
cvta.to.global.u64 %rd104, %rd21;
shl.b64 %rd105, %rd103, 2;
add.s64 %rd106, %rd104, %rd105;
st.global.f32 [%rd106], %f51;

$L__BB275_12:
ld.shared.f32 %f431, [%r1226+384];
add.ftz.f32 %f432, %f431, %f602;
mov.b32 %r1305, %f432;
shfl.sync.bfly.b32 %r1309|%p29, %r1305, %r1286, %r1285, %r1287;
mov.b32 %f433, %r1309;
add.ftz.f32 %f434, %f432, %f433;
mov.b32 %r1310, %f434;
shfl.sync.bfly.b32 %r1312|%p30, %r1310, %r1290, %r1285, %r1287;
mov.b32 %f435, %r1312;
add.ftz.f32 %f436, %f434, %f435;
mov.b32 %r1313, %f436;
shfl.sync.bfly.b32 %r1315|%p31, %r1313, %r1293, %r1285, %r1287;
mov.b32 %f437, %r1315;
add.ftz.f32 %f438, %f436, %f437;
mov.b32 %r1316, %f438;
shfl.sync.bfly.b32 %r1317|%p32, %r1316, %r1283, %r1285, %r1287;
mov.b32 %f439, %r1317;
add.ftz.f32 %f440, %f438, %f439;
mov.b32 %r1318, %f440;
shfl.sync.bfly.b32 %r1320|%p33, %r1318, %r1298, %r1285, %r1287;
mov.b32 %f441, %r1320;
add.ftz.f32 %f52, %f440, %f441;
@%p22 bra $L__BB275_14;

add.s32 %r1321, %r37, 1;
cvt.u64.u32 %rd107, %r1321;
add.s64 %rd108, %rd107, %rd15;
cvta.to.global.u64 %rd109, %rd21;
shl.b64 %rd110, %rd108, 2;
add.s64 %rd111, %rd109, %rd110;
st.global.f32 [%rd111], %f52;

$L__BB275_14:
shl.b32 %r1322, %r37, 1;
mov.u32 %r1323, 1;
mov.u32 %r1328, 2;
ld.shared.f32 %f442, [%r1226+512];
add.ftz.f32 %f443, %f442, %f601;
mov.b32 %r1329, %f443;
mov.u32 %r1330, 31;
mov.u32 %r1331, 16;
mov.u32 %r1332, -1;
shfl.sync.bfly.b32 %r1333|%p35, %r1329, %r1331, %r1330, %r1332;
mov.b32 %f444, %r1333;
add.ftz.f32 %f445, %f443, %f444;
mov.b32 %r1334, %f445;
mov.u32 %r1335, 8;
shfl.sync.bfly.b32 %r1336|%p36, %r1334, %r1335, %r1330, %r1332;
mov.b32 %f446, %r1336;
add.ftz.f32 %f447, %f445, %f446;
mov.b32 %r1337, %f447;
mov.u32 %r1338, 4;
shfl.sync.bfly.b32 %r1339|%p37, %r1337, %r1338, %r1330, %r1332;
mov.b32 %f448, %r1339;
add.ftz.f32 %f449, %f447, %f448;
mov.b32 %r1340, %f449;
shfl.sync.bfly.b32 %r1341|%p38, %r1340, %r1328, %r1330, %r1332;
mov.b32 %f450, %r1341;
add.ftz.f32 %f451, %f449, %f450;
mov.b32 %r1342, %f451;
shfl.sync.bfly.b32 %r1343|%p39, %r1342, %r1323, %r1330, %r1332;
mov.b32 %f452, %r1343;
add.ftz.f32 %f53, %f451, %f452;
cvt.u64.u32 %rd112, %r1322;
add.s64 %rd113, %rd112, %rd15;
cvta.to.global.u64 %rd114, %rd21;
shl.b64 %rd115, %rd113, 2;
add.s64 %rd16, %rd114, %rd115;
@%p14 bra $L__BB275_16;

st.global.f32 [%rd16], %f53;

$L__BB275_16:
ld.shared.f32 %f453, [%r1226+640];
add.ftz.f32 %f454, %f453, %f600;
mov.b32 %r1349, %f454;
shfl.sync.bfly.b32 %r1353|%p41, %r1349, %r1331, %r1330, %r1332;
mov.b32 %f455, %r1353;
add.ftz.f32 %f456, %f454, %f455;
mov.b32 %r1354, %f456;
shfl.sync.bfly.b32 %r1356|%p42, %r1354, %r1335, %r1330, %r1332;
mov.b32 %f457, %r1356;
add.ftz.f32 %f458, %f456, %f457;
mov.b32 %r1357, %f458;
shfl.sync.bfly.b32 %r1359|%p43, %r1357, %r1338, %r1330, %r1332;
mov.b32 %f459, %r1359;
add.ftz.f32 %f460, %f458, %f459;
mov.b32 %r1360, %f460;
shfl.sync.bfly.b32 %r1361|%p44, %r1360, %r1328, %r1330, %r1332;
mov.b32 %f461, %r1361;
add.ftz.f32 %f462, %f460, %f461;
mov.b32 %r1362, %f462;
shfl.sync.bfly.b32 %r1364|%p45, %r1362, %r1323, %r1330, %r1332;
mov.b32 %f463, %r1364;
add.ftz.f32 %f54, %f462, %f463;
@%p22 bra $L__BB275_18;

st.global.f32 [%rd16+4], %f54;

$L__BB275_18:
mul.lo.s32 %r32, %r37, 3;
mov.u32 %r1369, 2;
ld.shared.f32 %f464, [%r1226+768];
add.ftz.f32 %f465, %f464, %f599;
mov.b32 %r1370, %f465;
mov.u32 %r1371, 31;
mov.u32 %r1372, 16;
mov.u32 %r1373, -1;
shfl.sync.bfly.b32 %r1374|%p47, %r1370, %r1372, %r1371, %r1373;
mov.b32 %f466, %r1374;
add.ftz.f32 %f467, %f465, %f466;
mov.b32 %r1375, %f467;
mov.u32 %r1376, 8;
shfl.sync.bfly.b32 %r1377|%p48, %r1375, %r1376, %r1371, %r1373;
mov.b32 %f468, %r1377;
add.ftz.f32 %f469, %f467, %f468;
mov.b32 %r1378, %f469;
mov.u32 %r1379, 4;
shfl.sync.bfly.b32 %r1380|%p49, %r1378, %r1379, %r1371, %r1373;
mov.b32 %f470, %r1380;
add.ftz.f32 %f471, %f469, %f470;
mov.b32 %r1381, %f471;
shfl.sync.bfly.b32 %r1382|%p50, %r1381, %r1369, %r1371, %r1373;
mov.b32 %f472, %r1382;
add.ftz.f32 %f473, %f471, %f472;
mov.b32 %r1383, %f473;
mov.u32 %r1384, 1;
shfl.sync.bfly.b32 %r1385|%p51, %r1383, %r1384, %r1371, %r1373;
mov.b32 %f474, %r1385;
add.ftz.f32 %f55, %f473, %f474;
@%p14 bra $L__BB275_20;

cvt.u64.u32 %rd116, %r32;
add.s64 %rd117, %rd116, %rd15;
shl.b64 %rd119, %rd117, 2;
add.s64 %rd120, %rd114, %rd119;
st.global.f32 [%rd120], %f55;

$L__BB275_20:
ld.shared.f32 %f475, [%r1226+896];
add.ftz.f32 %f476, %f475, %f598;
mov.b32 %r1391, %f476;
shfl.sync.bfly.b32 %r1395|%p53, %r1391, %r1372, %r1371, %r1373;
mov.b32 %f477, %r1395;
add.ftz.f32 %f478, %f476, %f477;
mov.b32 %r1396, %f478;
shfl.sync.bfly.b32 %r1398|%p54, %r1396, %r1376, %r1371, %r1373;
mov.b32 %f479, %r1398;
add.ftz.f32 %f480, %f478, %f479;
mov.b32 %r1399, %f480;
shfl.sync.bfly.b32 %r1401|%p55, %r1399, %r1379, %r1371, %r1373;
mov.b32 %f481, %r1401;
add.ftz.f32 %f482, %f480, %f481;
mov.b32 %r1402, %f482;
shfl.sync.bfly.b32 %r1403|%p56, %r1402, %r1369, %r1371, %r1373;
mov.b32 %f483, %r1403;
add.ftz.f32 %f484, %f482, %f483;
mov.b32 %r1404, %f484;
shfl.sync.bfly.b32 %r1406|%p57, %r1404, %r1384, %r1371, %r1373;
mov.b32 %f485, %r1406;
add.ftz.f32 %f56, %f484, %f485;
@%p22 bra $L__BB275_22;

add.s32 %r1407, %r32, 1;
cvt.u64.u32 %rd121, %r1407;
add.s64 %rd122, %rd121, %rd15;
shl.b64 %rd124, %rd122, 2;
add.s64 %rd125, %rd114, %rd124;
st.global.f32 [%rd125], %f56;

$L__BB275_22:
shl.b32 %r1408, %r37, 2;
mov.u32 %r1409, 2;
ld.shared.f32 %f486, [%r1226+1024];
add.ftz.f32 %f487, %f486, %f597;
mov.b32 %r1414, %f487;
mov.u32 %r1415, 31;
mov.u32 %r1416, 16;
mov.u32 %r1417, -1;
shfl.sync.bfly.b32 %r1418|%p59, %r1414, %r1416, %r1415, %r1417;
mov.b32 %f488, %r1418;
add.ftz.f32 %f489, %f487, %f488;
mov.b32 %r1419, %f489;
mov.u32 %r1420, 8;
shfl.sync.bfly.b32 %r1421|%p60, %r1419, %r1420, %r1415, %r1417;
mov.b32 %f490, %r1421;
add.ftz.f32 %f491, %f489, %f490;
mov.b32 %r1422, %f491;
mov.u32 %r1423, 4;
shfl.sync.bfly.b32 %r1424|%p61, %r1422, %r1423, %r1415, %r1417;
mov.b32 %f492, %r1424;
add.ftz.f32 %f493, %f491, %f492;
mov.b32 %r1425, %f493;
shfl.sync.bfly.b32 %r1426|%p62, %r1425, %r1409, %r1415, %r1417;
mov.b32 %f494, %r1426;
add.ftz.f32 %f495, %f493, %f494;
mov.b32 %r1427, %f495;
mov.u32 %r1428, 1;
shfl.sync.bfly.b32 %r1429|%p63, %r1427, %r1428, %r1415, %r1417;
mov.b32 %f496, %r1429;
add.ftz.f32 %f57, %f495, %f496;
cvt.u64.u32 %rd126, %r1408;
add.s64 %rd127, %rd126, %rd15;
shl.b64 %rd129, %rd127, 2;
add.s64 %rd17, %rd114, %rd129;
@%p14 bra $L__BB275_24;

st.global.f32 [%rd17], %f57;

$L__BB275_24:
ld.shared.f32 %f497, [%r1226+1152];
add.ftz.f32 %f498, %f497, %f596;
mov.b32 %r1435, %f498;
shfl.sync.bfly.b32 %r1439|%p65, %r1435, %r1416, %r1415, %r1417;
mov.b32 %f499, %r1439;
add.ftz.f32 %f500, %f498, %f499;
mov.b32 %r1440, %f500;
shfl.sync.bfly.b32 %r1442|%p66, %r1440, %r1420, %r1415, %r1417;
mov.b32 %f501, %r1442;
add.ftz.f32 %f502, %f500, %f501;
mov.b32 %r1443, %f502;
shfl.sync.bfly.b32 %r1445|%p67, %r1443, %r1423, %r1415, %r1417;
mov.b32 %f503, %r1445;
add.ftz.f32 %f504, %f502, %f503;
mov.b32 %r1446, %f504;
shfl.sync.bfly.b32 %r1447|%p68, %r1446, %r1409, %r1415, %r1417;
mov.b32 %f505, %r1447;
add.ftz.f32 %f506, %f504, %f505;
mov.b32 %r1448, %f506;
shfl.sync.bfly.b32 %r1450|%p69, %r1448, %r1428, %r1415, %r1417;
mov.b32 %f507, %r1450;
add.ftz.f32 %f58, %f506, %f507;
@%p22 bra $L__BB275_26;

st.global.f32 [%rd17+4], %f58;

$L__BB275_26:
mul.lo.s32 %r33, %r37, 5;
mov.u32 %r1455, 2;
ld.shared.f32 %f508, [%r1226+1280];
add.ftz.f32 %f509, %f508, %f595;
mov.b32 %r1456, %f509;
mov.u32 %r1457, 31;
mov.u32 %r1458, 16;
mov.u32 %r1459, -1;
shfl.sync.bfly.b32 %r1460|%p71, %r1456, %r1458, %r1457, %r1459;
mov.b32 %f510, %r1460;
add.ftz.f32 %f511, %f509, %f510;
mov.b32 %r1461, %f511;
mov.u32 %r1462, 8;
shfl.sync.bfly.b32 %r1463|%p72, %r1461, %r1462, %r1457, %r1459;
mov.b32 %f512, %r1463;
add.ftz.f32 %f513, %f511, %f512;
mov.b32 %r1464, %f513;
mov.u32 %r1465, 4;
shfl.sync.bfly.b32 %r1466|%p73, %r1464, %r1465, %r1457, %r1459;
mov.b32 %f514, %r1466;
add.ftz.f32 %f515, %f513, %f514;
mov.b32 %r1467, %f515;
shfl.sync.bfly.b32 %r1468|%p74, %r1467, %r1455, %r1457, %r1459;
mov.b32 %f516, %r1468;
add.ftz.f32 %f517, %f515, %f516;
mov.b32 %r1469, %f517;
mov.u32 %r1470, 1;
shfl.sync.bfly.b32 %r1471|%p75, %r1469, %r1470, %r1457, %r1459;
mov.b32 %f518, %r1471;
add.ftz.f32 %f59, %f517, %f518;
@%p14 bra $L__BB275_28;

cvt.u64.u32 %rd130, %r33;
add.s64 %rd131, %rd130, %rd15;
shl.b64 %rd133, %rd131, 2;
add.s64 %rd134, %rd114, %rd133;
st.global.f32 [%rd134], %f59;

$L__BB275_28:
ld.shared.f32 %f519, [%r1226+1408];
add.ftz.f32 %f520, %f519, %f594;
mov.b32 %r1477, %f520;
shfl.sync.bfly.b32 %r1481|%p77, %r1477, %r1458, %r1457, %r1459;
mov.b32 %f521, %r1481;
add.ftz.f32 %f522, %f520, %f521;
mov.b32 %r1482, %f522;
shfl.sync.bfly.b32 %r1484|%p78, %r1482, %r1462, %r1457, %r1459;
mov.b32 %f523, %r1484;
add.ftz.f32 %f524, %f522, %f523;
mov.b32 %r1485, %f524;
shfl.sync.bfly.b32 %r1487|%p79, %r1485, %r1465, %r1457, %r1459;
mov.b32 %f525, %r1487;
add.ftz.f32 %f526, %f524, %f525;
mov.b32 %r1488, %f526;
shfl.sync.bfly.b32 %r1489|%p80, %r1488, %r1455, %r1457, %r1459;
mov.b32 %f527, %r1489;
add.ftz.f32 %f528, %f526, %f527;
mov.b32 %r1490, %f528;
shfl.sync.bfly.b32 %r1492|%p81, %r1490, %r1470, %r1457, %r1459;
mov.b32 %f529, %r1492;
add.ftz.f32 %f60, %f528, %f529;
@%p22 bra $L__BB275_30;

add.s32 %r1493, %r33, 1;
cvt.u64.u32 %rd135, %r1493;
add.s64 %rd136, %rd135, %rd15;
shl.b64 %rd138, %rd136, 2;
add.s64 %rd139, %rd114, %rd138;
st.global.f32 [%rd139], %f60;

$L__BB275_30:
mul.lo.s32 %r1494, %r37, 6;
mov.u32 %r1499, 2;
ld.shared.f32 %f530, [%r1226+1536];
add.ftz.f32 %f531, %f530, %f593;
mov.b32 %r1500, %f531;
mov.u32 %r1501, 31;
mov.u32 %r1502, 16;
mov.u32 %r1503, -1;
shfl.sync.bfly.b32 %r1504|%p83, %r1500, %r1502, %r1501, %r1503;
mov.b32 %f532, %r1504;
add.ftz.f32 %f533, %f531, %f532;
mov.b32 %r1505, %f533;
mov.u32 %r1506, 8;
shfl.sync.bfly.b32 %r1507|%p84, %r1505, %r1506, %r1501, %r1503;
mov.b32 %f534, %r1507;
add.ftz.f32 %f535, %f533, %f534;
mov.b32 %r1508, %f535;
mov.u32 %r1509, 4;
shfl.sync.bfly.b32 %r1510|%p85, %r1508, %r1509, %r1501, %r1503;
mov.b32 %f536, %r1510;
add.ftz.f32 %f537, %f535, %f536;
mov.b32 %r1511, %f537;
shfl.sync.bfly.b32 %r1512|%p86, %r1511, %r1499, %r1501, %r1503;
mov.b32 %f538, %r1512;
add.ftz.f32 %f539, %f537, %f538;
mov.b32 %r1513, %f539;
mov.u32 %r1514, 1;
shfl.sync.bfly.b32 %r1515|%p87, %r1513, %r1514, %r1501, %r1503;
mov.b32 %f540, %r1515;
add.ftz.f32 %f61, %f539, %f540;
cvt.u64.u32 %rd140, %r1494;
add.s64 %rd141, %rd140, %rd15;
shl.b64 %rd143, %rd141, 2;
add.s64 %rd18, %rd114, %rd143;
@%p14 bra $L__BB275_32;

st.global.f32 [%rd18], %f61;

$L__BB275_32:
ld.shared.f32 %f541, [%r1226+1664];
add.ftz.f32 %f542, %f541, %f592;
mov.b32 %r1521, %f542;
shfl.sync.bfly.b32 %r1525|%p89, %r1521, %r1502, %r1501, %r1503;
mov.b32 %f543, %r1525;
add.ftz.f32 %f544, %f542, %f543;
mov.b32 %r1526, %f544;
shfl.sync.bfly.b32 %r1528|%p90, %r1526, %r1506, %r1501, %r1503;
mov.b32 %f545, %r1528;
add.ftz.f32 %f546, %f544, %f545;
mov.b32 %r1529, %f546;
shfl.sync.bfly.b32 %r1531|%p91, %r1529, %r1509, %r1501, %r1503;
mov.b32 %f547, %r1531;
add.ftz.f32 %f548, %f546, %f547;
mov.b32 %r1532, %f548;
shfl.sync.bfly.b32 %r1533|%p92, %r1532, %r1499, %r1501, %r1503;
mov.b32 %f549, %r1533;
add.ftz.f32 %f550, %f548, %f549;
mov.b32 %r1534, %f550;
shfl.sync.bfly.b32 %r1536|%p93, %r1534, %r1514, %r1501, %r1503;
mov.b32 %f551, %r1536;
add.ftz.f32 %f62, %f550, %f551;
@%p22 bra $L__BB275_34;

st.global.f32 [%rd18+4], %f62;

$L__BB275_34:
mul.lo.s32 %r34, %r37, 7;
mov.u32 %r1541, 2;
ld.shared.f32 %f552, [%r1226+1792];
add.ftz.f32 %f553, %f552, %f591;
mov.b32 %r1542, %f553;
mov.u32 %r1543, 31;
mov.u32 %r1544, 16;
mov.u32 %r1545, -1;
shfl.sync.bfly.b32 %r1546|%p95, %r1542, %r1544, %r1543, %r1545;
mov.b32 %f554, %r1546;
add.ftz.f32 %f555, %f553, %f554;
mov.b32 %r1547, %f555;
mov.u32 %r1548, 8;
shfl.sync.bfly.b32 %r1549|%p96, %r1547, %r1548, %r1543, %r1545;
mov.b32 %f556, %r1549;
add.ftz.f32 %f557, %f555, %f556;
mov.b32 %r1550, %f557;
mov.u32 %r1551, 4;
shfl.sync.bfly.b32 %r1552|%p97, %r1550, %r1551, %r1543, %r1545;
mov.b32 %f558, %r1552;
add.ftz.f32 %f559, %f557, %f558;
mov.b32 %r1553, %f559;
shfl.sync.bfly.b32 %r1554|%p98, %r1553, %r1541, %r1543, %r1545;
mov.b32 %f560, %r1554;
add.ftz.f32 %f561, %f559, %f560;
mov.b32 %r1555, %f561;
mov.u32 %r1556, 1;
shfl.sync.bfly.b32 %r1557|%p99, %r1555, %r1556, %r1543, %r1545;
mov.b32 %f562, %r1557;
add.ftz.f32 %f63, %f561, %f562;
@%p14 bra $L__BB275_36;

cvt.u64.u32 %rd144, %r34;
add.s64 %rd145, %rd144, %rd15;
shl.b64 %rd147, %rd145, 2;
add.s64 %rd148, %rd114, %rd147;
st.global.f32 [%rd148], %f63;

$L__BB275_36:
ld.shared.f32 %f563, [%r1226+1920];
add.ftz.f32 %f564, %f563, %f590;
mov.b32 %r1563, %f564;
shfl.sync.bfly.b32 %r1567|%p101, %r1563, %r1544, %r1543, %r1545;
mov.b32 %f565, %r1567;
add.ftz.f32 %f566, %f564, %f565;
mov.b32 %r1568, %f566;
shfl.sync.bfly.b32 %r1570|%p102, %r1568, %r1548, %r1543, %r1545;
mov.b32 %f567, %r1570;
add.ftz.f32 %f568, %f566, %f567;
mov.b32 %r1571, %f568;
shfl.sync.bfly.b32 %r1573|%p103, %r1571, %r1551, %r1543, %r1545;
mov.b32 %f569, %r1573;
add.ftz.f32 %f570, %f568, %f569;
mov.b32 %r1574, %f570;
shfl.sync.bfly.b32 %r1575|%p104, %r1574, %r1541, %r1543, %r1545;
mov.b32 %f571, %r1575;
add.ftz.f32 %f572, %f570, %f571;
mov.b32 %r1576, %f572;
shfl.sync.bfly.b32 %r1578|%p105, %r1576, %r1556, %r1543, %r1545;
mov.b32 %f573, %r1578;
add.ftz.f32 %f64, %f572, %f573;
@%p22 bra $L__BB275_38;

add.s32 %r1579, %r34, 1;
cvt.u64.u32 %rd149, %r1579;
add.s64 %rd150, %rd149, %rd15;
shl.b64 %rd152, %rd150, 2;
add.s64 %rd153, %rd114, %rd152;
st.global.f32 [%rd153], %f64;

$L__BB275_38:
ret;

}
.entry _Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj(
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2,
.param .align 8 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3[48],
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5,
.param .align 4 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6[12],
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15
)
.maxntid 256, 1, 1
.minnctapersm 1
{
.local .align 8 .b8 __local_depot276[16];
.reg .b64 %SP;
.reg .b64 %SPL;
.reg .pred %p<48>;
.reg .b16 %rs<29>;
.reg .f32 %f<169>;
.reg .b32 %r<378>;
.reg .b64 %rd<108>;


mov.u64 %SPL, __local_depot276;
ld.param.u64 %rd19, [_Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0];
ld.param.u64 %rd20, [_Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1];
ld.param.u64 %rd21, [_Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2];
ld.param.u32 %r17, [_Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5];
ld.param.u32 %r22, [_Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8];
ld.param.u32 %r23, [_Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9];
ld.param.u32 %r25, [_Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11];
ld.param.u32 %r26, [_Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12];
ld.param.u32 %r29, [_Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14];
ld.param.u32 %r28, [_Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15];
ld.param.u32 %r20, [_Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+8];
ld.param.u32 %r19, [_Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+4];
ld.param.u32 %r18, [_Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6];
ld.param.u64 %rd23, [_Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+8];
add.u64 %rd1, %SPL, 0;
add.u64 %rd2, %SPL, 8;
mov.u32 %r1, %tid.y;
setp.ge.u32 %p1, %r1, %r29;
@%p1 bra $L__BB276_29;

mov.u32 %r2, %ctaid.y;
cvta.to.global.u64 %rd30, %rd21;
mad.lo.s32 %r30, %r1, %r28, %r2;
mul.wide.u32 %rd31, %r30, 4;
add.s64 %rd32, %rd30, %rd31;
ld.global.nc.u32 %r3, [%rd32];
mov.f32 %f162, 0f00000000;
st.local.v2.f32 [%rd1], {%f162, %f162};
st.local.v2.f32 [%rd2], {%f162, %f162};
mov.u32 %r4, %tid.x;
shr.u32 %r5, %r4, 1;
shr.u32 %r31, %r17, 5;
setp.ge.u32 %p2, %r5, %r31;
mov.f32 %f163, %f162;
mov.f32 %f164, %f162;
mov.f32 %f165, %f162;
@%p2 bra $L__BB276_9;

mul.hi.u32 %r32, %r2, %r18;
add.s32 %r33, %r32, %r2;
shr.u32 %r34, %r33, %r19;
mul.lo.s32 %r35, %r34, %r20;
sub.s32 %r36, %r2, %r35;
mul.lo.s32 %r37, %r36, %r26;
cvt.u64.u32 %rd33, %r37;
cvt.u64.u32 %rd34, %r4;
shr.u64 %rd35, %rd34, 1;
mul.wide.u32 %rd107, %r37, 36;
and.b32 %r38, %r4, 1;
mul.wide.u32 %rd36, %r38, 8;
mul.lo.s32 %r39, %r1, %r23;
cvt.u64.u32 %rd37, %r39;
add.s64 %rd38, %rd35, %rd37;
add.s64 %rd39, %rd38, %rd33;
mul.lo.s64 %rd40, %rd39, 36;
add.s64 %rd41, %rd40, %rd36;
cvta.to.global.u64 %rd42, %rd20;
add.s64 %rd43, %rd42, %rd41;
add.s64 %rd105, %rd43, 20;
cvta.to.global.u64 %rd106, %rd19;
cvta.to.global.u64 %rd104, %rd23;
mov.f32 %f162, 0f00000000;
mov.u32 %r377, %r5;

$L__BB276_3:
mul.lo.s64 %rd46, %rd35, 36;
mul.wide.u32 %rd47, %r39, 36;
add.s64 %rd48, %rd46, %rd47;
add.s64 %rd50, %rd42, %rd48;
add.s64 %rd51, %rd50, %rd107;
mov.u32 %r62, %ctaid.x;
shl.b32 %r63, %r62, 1;
mad.lo.s32 %r64, %r63, %r22, %r5;
mad.lo.s32 %r65, %r3, %r25, %r64;
mul.wide.s32 %rd53, %r65, 18;
add.s64 %rd54, %rd36, %rd53;
add.s64 %rd55, %rd106, %rd54;
ld.global.nc.u16 %rs2, [%rd55+2];
cvt.u32.u16 %r66, %rs2;
ld.global.nc.u16 %rs3, [%rd55+4];
mov.b32 %r67, {%rs2, %rs3};
and.b32 %r68, %r67, -2004318072;
shr.u32 %r69, %r68, 1;
or.b32 %r70, %r69, 839922192;
and.b32 %r71, %r66, 30583;
mov.u32 %r72, -152379953;
mov.u32 %r73, -1079142271;
prmt.b32 %r74, %r73, %r72, %r71;
mov.u32 %r75, 1901675829;
mov.u32 %r76, 639175937;
prmt.b32 %r77, %r76, %r75, %r71;
and.b32 %r78, %r70, 30292;
prmt.b32 %r79, %r74, %r77, %r78;
and.b16 %rs4, %rs3, 30583;
cvt.u32.u16 %r80, %rs4;
prmt.b32 %r81, %r73, %r72, %r80;
prmt.b32 %r82, %r76, %r75, %r80;
shr.u32 %r83, %r70, 16;
prmt.b32 %r84, %r81, %r82, %r83;
mov.u32 %r85, 25632;
prmt.b32 %r41, %r79, %r84, %r85;
mov.u32 %r86, 30001;
prmt.b32 %r45, %r79, %r84, %r86;
ld.global.nc.u32 %r7, [%rd105+-16];
mov.u32 %r43, 0;

dp4a.s32.s32 %r40, %r41, %r7, %r43;

ld.global.nc.u32 %r8, [%rd105];

dp4a.s32.s32 %r44, %r45, %r8, %r40;

ld.global.nc.u16 %rs5, [%rd55+6];
cvt.u32.u16 %r87, %rs5;
ld.global.nc.u16 %rs6, [%rd55+8];
mov.b32 %r88, {%rs5, %rs6};
and.b32 %r89, %r88, -2004318072;
shr.u32 %r90, %r89, 1;
or.b32 %r91, %r90, 839922192;
and.b32 %r92, %r87, 30583;
prmt.b32 %r93, %r73, %r72, %r92;
prmt.b32 %r94, %r76, %r75, %r92;
and.b32 %r95, %r91, 30292;
prmt.b32 %r96, %r93, %r94, %r95;
and.b16 %rs7, %rs6, 30583;
cvt.u32.u16 %r97, %rs7;
prmt.b32 %r98, %r73, %r72, %r97;
prmt.b32 %r99, %r76, %r75, %r97;
shr.u32 %r100, %r91, 16;
prmt.b32 %r101, %r98, %r99, %r100;
prmt.b32 %r49, %r96, %r101, %r85;
prmt.b32 %r53, %r96, %r101, %r86;
ld.global.nc.u32 %r9, [%rd105+-12];

dp4a.s32.s32 %r48, %r49, %r9, %r44;

ld.global.nc.u32 %r10, [%rd105+4];

dp4a.s32.s32 %r52, %r53, %r10, %r48;

add.s64 %rd56, %rd106, %rd53;
ld.global.nc.u16 %rs1, [%rd56];

{ cvt.f32.f16 %f42, %rs1;}


ld.global.nc.u32 %r11, [%rd51];

{.reg .f16 low,high;
mov.b32 {low,high},%r11;
cvt.f32.f16 %f43, low;}


mul.ftz.f32 %f44, %f42, %f43;
cvt.rn.f32.s32 %f45, %r52;
fma.rn.ftz.f32 %f165, %f44, %f45, %f165;
setp.eq.s64 %p3, %rd23, 0;
@%p3 bra $L__BB276_5;

shr.u32 %r121, %r4, 1;
mad.lo.s32 %r124, %r63, %r22, %r121;
mad.lo.s32 %r125, %r3, %r25, %r124;
mul.wide.s32 %rd58, %r125, 18;
add.s64 %rd59, %rd36, %rd58;
add.s64 %rd60, %rd104, %rd59;
ld.global.u16 %rs9, [%rd60+2];
cvt.u32.u16 %r126, %rs9;
ld.global.u16 %rs10, [%rd60+4];
mov.b32 %r127, {%rs9, %rs10};
and.b32 %r128, %r127, -2004318072;
shr.u32 %r129, %r128, 1;
or.b32 %r130, %r129, 839922192;
and.b32 %r131, %r126, 30583;
prmt.b32 %r134, %r73, %r72, %r131;
prmt.b32 %r137, %r76, %r75, %r131;
and.b32 %r138, %r130, 30292;
prmt.b32 %r139, %r134, %r137, %r138;
and.b16 %rs11, %rs10, 30583;
cvt.u32.u16 %r140, %rs11;
prmt.b32 %r141, %r73, %r72, %r140;
prmt.b32 %r142, %r76, %r75, %r140;
shr.u32 %r143, %r130, 16;
prmt.b32 %r144, %r141, %r142, %r143;
prmt.b32 %r103, %r139, %r144, %r85;
prmt.b32 %r107, %r139, %r144, %r86;

dp4a.s32.s32 %r102, %r103, %r7, %r43;


dp4a.s32.s32 %r106, %r107, %r8, %r102;

ld.global.u16 %rs12, [%rd60+6];
cvt.u32.u16 %r147, %rs12;
ld.global.u16 %rs13, [%rd60+8];
mov.b32 %r148, {%rs12, %rs13};
and.b32 %r149, %r148, -2004318072;
shr.u32 %r150, %r149, 1;
or.b32 %r151, %r150, 839922192;
and.b32 %r152, %r147, 30583;
prmt.b32 %r153, %r73, %r72, %r152;
prmt.b32 %r154, %r76, %r75, %r152;
and.b32 %r155, %r151, 30292;
prmt.b32 %r156, %r153, %r154, %r155;
and.b16 %rs14, %rs13, 30583;
cvt.u32.u16 %r157, %rs14;
prmt.b32 %r158, %r73, %r72, %r157;
prmt.b32 %r159, %r76, %r75, %r157;
shr.u32 %r160, %r151, 16;
prmt.b32 %r161, %r158, %r159, %r160;
prmt.b32 %r111, %r156, %r161, %r85;
prmt.b32 %r115, %r156, %r161, %r86;

dp4a.s32.s32 %r110, %r111, %r9, %r106;


dp4a.s32.s32 %r114, %r115, %r10, %r110;

add.s64 %rd61, %rd104, %rd58;
ld.global.u16 %rs8, [%rd61];

{ cvt.f32.f16 %f46, %rs8;}



{.reg .f16 low,high;
mov.b32 {low,high},%r11;
cvt.f32.f16 %f47, low;}


mul.ftz.f32 %f48, %f46, %f47;
cvt.rn.f32.s32 %f49, %r114;
fma.rn.ftz.f32 %f163, %f48, %f49, %f163;

$L__BB276_5:
shr.u32 %r181, %r4, 1;
mad.lo.s32 %r182, %r3, %r25, %r181;
or.b32 %r185, %r63, 1;
mad.lo.s32 %r186, %r22, %r185, %r182;
mul.wide.s32 %rd63, %r186, 18;
add.s64 %rd64, %rd36, %rd63;
add.s64 %rd65, %rd106, %rd64;
ld.global.nc.u16 %rs16, [%rd65+2];
cvt.u32.u16 %r187, %rs16;
ld.global.nc.u16 %rs17, [%rd65+4];
mov.b32 %r188, {%rs16, %rs17};
and.b32 %r189, %r188, -2004318072;
shr.u32 %r190, %r189, 1;
or.b32 %r191, %r190, 839922192;
and.b32 %r192, %r187, 30583;
prmt.b32 %r195, %r73, %r72, %r192;
prmt.b32 %r198, %r76, %r75, %r192;
and.b32 %r199, %r191, 30292;
prmt.b32 %r200, %r195, %r198, %r199;
and.b16 %rs18, %rs17, 30583;
cvt.u32.u16 %r201, %rs18;
prmt.b32 %r202, %r73, %r72, %r201;
prmt.b32 %r203, %r76, %r75, %r201;
shr.u32 %r204, %r191, 16;
prmt.b32 %r205, %r202, %r203, %r204;
prmt.b32 %r163, %r200, %r205, %r85;
prmt.b32 %r167, %r200, %r205, %r86;

dp4a.s32.s32 %r162, %r163, %r7, %r43;


dp4a.s32.s32 %r166, %r167, %r8, %r162;

ld.global.nc.u16 %rs19, [%rd65+6];
cvt.u32.u16 %r208, %rs19;
ld.global.nc.u16 %rs20, [%rd65+8];
mov.b32 %r209, {%rs19, %rs20};
and.b32 %r210, %r209, -2004318072;
shr.u32 %r211, %r210, 1;
or.b32 %r212, %r211, 839922192;
and.b32 %r213, %r208, 30583;
prmt.b32 %r214, %r73, %r72, %r213;
prmt.b32 %r215, %r76, %r75, %r213;
and.b32 %r216, %r212, 30292;
prmt.b32 %r217, %r214, %r215, %r216;
and.b16 %rs21, %rs20, 30583;
cvt.u32.u16 %r218, %rs21;
prmt.b32 %r219, %r73, %r72, %r218;
prmt.b32 %r220, %r76, %r75, %r218;
shr.u32 %r221, %r212, 16;
prmt.b32 %r222, %r219, %r220, %r221;
prmt.b32 %r171, %r217, %r222, %r85;
prmt.b32 %r175, %r217, %r222, %r86;

dp4a.s32.s32 %r170, %r171, %r9, %r166;


dp4a.s32.s32 %r174, %r175, %r10, %r170;

add.s64 %rd66, %rd106, %rd63;
ld.global.nc.u16 %rs15, [%rd66];

{ cvt.f32.f16 %f50, %rs15;}



{.reg .f16 low,high;
mov.b32 {low,high},%r11;
cvt.f32.f16 %f51, low;}


mul.ftz.f32 %f52, %f50, %f51;
cvt.rn.f32.s32 %f53, %r174;
fma.rn.ftz.f32 %f164, %f52, %f53, %f164;
@%p3 bra $L__BB276_7;

add.s64 %rd70, %rd104, %rd64;
ld.global.u16 %rs23, [%rd70+2];
cvt.u32.u16 %r248, %rs23;
ld.global.u16 %rs24, [%rd70+4];
mov.b32 %r249, {%rs23, %rs24};
and.b32 %r250, %r249, -2004318072;
shr.u32 %r251, %r250, 1;
or.b32 %r252, %r251, 839922192;
and.b32 %r253, %r248, 30583;
mov.u32 %r254, -152379953;
mov.u32 %r255, -1079142271;
prmt.b32 %r256, %r255, %r254, %r253;
mov.u32 %r257, 1901675829;
mov.u32 %r258, 639175937;
prmt.b32 %r259, %r258, %r257, %r253;
and.b32 %r260, %r252, 30292;
prmt.b32 %r261, %r256, %r259, %r260;
and.b16 %rs25, %rs24, 30583;
cvt.u32.u16 %r262, %rs25;
prmt.b32 %r263, %r255, %r254, %r262;
prmt.b32 %r264, %r258, %r257, %r262;
shr.u32 %r265, %r252, 16;
prmt.b32 %r266, %r263, %r264, %r265;
mov.u32 %r267, 25632;
prmt.b32 %r224, %r261, %r266, %r267;
mov.u32 %r268, 30001;
prmt.b32 %r228, %r261, %r266, %r268;
mov.u32 %r226, 0;

dp4a.s32.s32 %r223, %r224, %r7, %r226;


dp4a.s32.s32 %r227, %r228, %r8, %r223;

ld.global.u16 %rs26, [%rd70+6];
cvt.u32.u16 %r269, %rs26;
ld.global.u16 %rs27, [%rd70+8];
mov.b32 %r270, {%rs26, %rs27};
and.b32 %r271, %r270, -2004318072;
shr.u32 %r272, %r271, 1;
or.b32 %r273, %r272, 839922192;
and.b32 %r274, %r269, 30583;
prmt.b32 %r275, %r255, %r254, %r274;
prmt.b32 %r276, %r258, %r257, %r274;
and.b32 %r277, %r273, 30292;
prmt.b32 %r278, %r275, %r276, %r277;
and.b16 %rs28, %rs27, 30583;
cvt.u32.u16 %r279, %rs28;
prmt.b32 %r280, %r255, %r254, %r279;
prmt.b32 %r281, %r258, %r257, %r279;
shr.u32 %r282, %r273, 16;
prmt.b32 %r283, %r280, %r281, %r282;
prmt.b32 %r232, %r278, %r283, %r267;
prmt.b32 %r236, %r278, %r283, %r268;

dp4a.s32.s32 %r231, %r232, %r9, %r227;


dp4a.s32.s32 %r235, %r236, %r10, %r231;

add.s64 %rd71, %rd104, %rd63;
ld.global.u16 %rs22, [%rd71];

{ cvt.f32.f16 %f54, %rs22;}



{.reg .f16 low,high;
mov.b32 {low,high},%r11;
cvt.f32.f16 %f55, low;}


mul.ftz.f32 %f56, %f54, %f55;
cvt.rn.f32.s32 %f57, %r235;
fma.rn.ftz.f32 %f162, %f56, %f57, %f162;

$L__BB276_7:
add.s64 %rd107, %rd107, 576;
add.s64 %rd106, %rd106, 288;
add.s64 %rd105, %rd105, 576;
add.s64 %rd104, %rd104, 288;
add.s32 %r377, %r377, 16;
setp.lt.u32 %p5, %r377, %r31;
@%p5 bra $L__BB276_3;

add.u64 %rd96, %SPL, 0;
add.u64 %rd90, %SPL, 8;
st.local.v2.f32 [%rd96], {%f165, %f164};
st.local.v2.f32 [%rd90], {%f163, %f162};

$L__BB276_9:
mov.b32 %r13, %f165;
setp.ne.s64 %p6, %rd23, 0;
@%p6 bra $L__BB276_11;
bra.uni $L__BB276_10;

$L__BB276_11:
add.u64 %rd102, %SPL, 0;
add.u64 %rd94, %SPL, 8;
mov.u32 %r311, 31;
mov.u32 %r312, 16;
mov.u32 %r313, -1;
shfl.sync.bfly.b32 %r314|%p17, %r13, %r312, %r311, %r313;
mov.b32 %f78, %r314;
add.ftz.f32 %f79, %f165, %f78;
mov.b32 %r315, %f79;
mov.u32 %r316, 8;
shfl.sync.bfly.b32 %r317|%p18, %r315, %r316, %r311, %r313;
mov.b32 %f80, %r317;
add.ftz.f32 %f81, %f79, %f80;
mov.b32 %r318, %f81;
mov.u32 %r319, 4;
shfl.sync.bfly.b32 %r320|%p19, %r318, %r319, %r311, %r313;
mov.b32 %f82, %r320;
add.ftz.f32 %f83, %f81, %f82;
mov.b32 %r321, %f83;
mov.u32 %r322, 2;
shfl.sync.bfly.b32 %r323|%p20, %r321, %r322, %r311, %r313;
mov.b32 %f84, %r323;
add.ftz.f32 %f85, %f83, %f84;
mov.b32 %r324, %f85;
mov.u32 %r325, 1;
shfl.sync.bfly.b32 %r326|%p21, %r324, %r325, %r311, %r313;
mov.b32 %f86, %r326;
add.ftz.f32 %f87, %f85, %f86;
st.local.f32 [%rd102], %f87;
mov.b32 %r327, %f163;
shfl.sync.bfly.b32 %r328|%p22, %r327, %r312, %r311, %r313;
mov.b32 %f88, %r328;
add.ftz.f32 %f89, %f163, %f88;
mov.b32 %r329, %f89;
shfl.sync.bfly.b32 %r330|%p23, %r329, %r316, %r311, %r313;
mov.b32 %f90, %r330;
add.ftz.f32 %f91, %f89, %f90;
mov.b32 %r331, %f91;
shfl.sync.bfly.b32 %r332|%p24, %r331, %r319, %r311, %r313;
mov.b32 %f92, %r332;
add.ftz.f32 %f93, %f91, %f92;
mov.b32 %r333, %f93;
shfl.sync.bfly.b32 %r334|%p25, %r333, %r322, %r311, %r313;
mov.b32 %f94, %r334;
add.ftz.f32 %f95, %f93, %f94;
mov.b32 %r335, %f95;
shfl.sync.bfly.b32 %r336|%p26, %r335, %r325, %r311, %r313;
mov.b32 %f96, %r336;
add.ftz.f32 %f97, %f95, %f96;
st.local.f32 [%rd94], %f97;
mov.b32 %r337, %f164;
shfl.sync.bfly.b32 %r338|%p27, %r337, %r312, %r311, %r313;
mov.b32 %f98, %r338;
add.ftz.f32 %f99, %f164, %f98;
mov.b32 %r339, %f99;
shfl.sync.bfly.b32 %r340|%p28, %r339, %r316, %r311, %r313;
mov.b32 %f100, %r340;
add.ftz.f32 %f101, %f99, %f100;
mov.b32 %r341, %f101;
shfl.sync.bfly.b32 %r342|%p29, %r341, %r319, %r311, %r313;
mov.b32 %f102, %r342;
add.ftz.f32 %f103, %f101, %f102;
mov.b32 %r343, %f103;
shfl.sync.bfly.b32 %r344|%p30, %r343, %r322, %r311, %r313;
mov.b32 %f104, %r344;
add.ftz.f32 %f105, %f103, %f104;
mov.b32 %r345, %f105;
shfl.sync.bfly.b32 %r346|%p31, %r345, %r325, %r311, %r313;
mov.b32 %f106, %r346;
add.ftz.f32 %f107, %f105, %f106;
st.local.f32 [%rd102+4], %f107;
mov.b32 %r347, %f162;
shfl.sync.bfly.b32 %r348|%p32, %r347, %r312, %r311, %r313;
mov.b32 %f108, %r348;
add.ftz.f32 %f109, %f162, %f108;
mov.b32 %r349, %f109;
shfl.sync.bfly.b32 %r350|%p33, %r349, %r316, %r311, %r313;
mov.b32 %f110, %r350;
add.ftz.f32 %f111, %f109, %f110;
mov.b32 %r351, %f111;
shfl.sync.bfly.b32 %r352|%p34, %r351, %r319, %r311, %r313;
mov.b32 %f112, %r352;
add.ftz.f32 %f113, %f111, %f112;
mov.b32 %r353, %f113;
shfl.sync.bfly.b32 %r354|%p35, %r353, %r322, %r311, %r313;
mov.b32 %f114, %r354;
add.ftz.f32 %f115, %f113, %f114;
mov.b32 %r355, %f115;
shfl.sync.bfly.b32 %r356|%p36, %r355, %r325, %r311, %r313;
mov.b32 %f116, %r356;
add.ftz.f32 %f117, %f115, %f116;
st.local.f32 [%rd94+4], %f117;
bra.uni $L__BB276_12;

$L__BB276_10:
add.u64 %rd98, %SPL, 0;
mov.u32 %r285, 31;
mov.u32 %r286, 16;
mov.u32 %r287, -1;
shfl.sync.bfly.b32 %r288|%p7, %r13, %r286, %r285, %r287;
mov.b32 %f58, %r288;
add.ftz.f32 %f59, %f165, %f58;
mov.b32 %r289, %f59;
mov.u32 %r290, 8;
shfl.sync.bfly.b32 %r291|%p8, %r289, %r290, %r285, %r287;
mov.b32 %f60, %r291;
add.ftz.f32 %f61, %f59, %f60;
mov.b32 %r292, %f61;
mov.u32 %r293, 4;
shfl.sync.bfly.b32 %r294|%p9, %r292, %r293, %r285, %r287;
mov.b32 %f62, %r294;
add.ftz.f32 %f63, %f61, %f62;
mov.b32 %r295, %f63;
mov.u32 %r296, 2;
shfl.sync.bfly.b32 %r297|%p10, %r295, %r296, %r285, %r287;
mov.b32 %f64, %r297;
add.ftz.f32 %f65, %f63, %f64;
mov.b32 %r298, %f65;
mov.u32 %r299, 1;
shfl.sync.bfly.b32 %r300|%p11, %r298, %r299, %r285, %r287;
mov.b32 %f66, %r300;
add.ftz.f32 %f67, %f65, %f66;
st.local.f32 [%rd98], %f67;
mov.b32 %r301, %f164;
shfl.sync.bfly.b32 %r302|%p12, %r301, %r286, %r285, %r287;
mov.b32 %f68, %r302;
add.ftz.f32 %f69, %f164, %f68;
mov.b32 %r303, %f69;
shfl.sync.bfly.b32 %r304|%p13, %r303, %r290, %r285, %r287;
mov.b32 %f70, %r304;
add.ftz.f32 %f71, %f69, %f70;
mov.b32 %r305, %f71;
shfl.sync.bfly.b32 %r306|%p14, %r305, %r293, %r285, %r287;
mov.b32 %f72, %r306;
add.ftz.f32 %f73, %f71, %f72;
mov.b32 %r307, %f73;
shfl.sync.bfly.b32 %r308|%p15, %r307, %r296, %r285, %r287;
mov.b32 %f74, %r308;
add.ftz.f32 %f75, %f73, %f74;
mov.b32 %r309, %f75;
shfl.sync.bfly.b32 %r310|%p16, %r309, %r299, %r285, %r287;
mov.b32 %f76, %r310;
add.ftz.f32 %f77, %f75, %f76;
st.local.f32 [%rd98+4], %f77;

$L__BB276_12:
ld.param.u32 %r366, [_Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7];
setp.gt.u32 %p37, %r4, 1;
mov.u32 %r358, %ctaid.x;
shl.b32 %r359, %r358, 1;
add.s32 %r14, %r359, %r4;
setp.ge.u32 %p38, %r14, %r366;
or.pred %p39, %p37, %p38;
@%p39 bra $L__BB276_29;

add.u64 %rd100, %SPL, 0;
ld.param.u64 %rd85, [_Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3];
ld.param.u32 %r367, [_Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13];
cvt.u64.u32 %rd17, %r4;
mul.wide.u32 %rd72, %r4, 4;
add.s64 %rd73, %rd100, %rd72;
ld.local.f32 %f168, [%rd73];
mad.lo.s32 %r15, %r3, %r367, %r14;
setp.eq.s64 %p40, %rd85, 0;
@%p40 bra $L__BB276_15;

ld.param.u64 %rd86, [_Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3];
cvta.to.global.u64 %rd74, %rd86;
mul.wide.u32 %rd75, %r15, 4;
add.s64 %rd76, %rd74, %rd75;
ld.global.f32 %f118, [%rd76];
add.ftz.f32 %f168, %f168, %f118;

$L__BB276_15:
setp.eq.s64 %p41, %rd23, 0;
@%p41 bra $L__BB276_28;

add.u64 %rd92, %SPL, 8;
ld.param.u64 %rd88, [_Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+16];
shl.b64 %rd77, %rd17, 2;
add.s64 %rd78, %rd92, %rd77;
ld.local.f32 %f24, [%rd78];
setp.eq.s64 %p42, %rd88, 0;
@%p42 bra $L__BB276_18;

ld.param.u64 %rd89, [_Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+16];
cvta.to.global.u64 %rd79, %rd89;
mul.wide.u32 %rd80, %r15, 4;
add.s64 %rd81, %rd79, %rd80;
ld.global.f32 %f119, [%rd81];
add.ftz.f32 %f24, %f24, %f119;

$L__BB276_18:
ld.param.u32 %r372, [_Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.gt.s32 %p43, %r372, 2;
@%p43 bra $L__BB276_22;

ld.param.u32 %r375, [_Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.eq.s32 %p46, %r375, 1;
@%p46 bra $L__BB276_26;

ld.param.u32 %r376, [_Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.eq.s32 %p47, %r376, 2;
@%p47 bra $L__BB276_21;
bra.uni $L__BB276_27;

$L__BB276_21:
mul.ftz.f32 %f139, %f24, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f140, %f139;
add.ftz.f32 %f141, %f140, 0f3F800000;
div.approx.ftz.f32 %f142, %f24, %f141;
mul.ftz.f32 %f168, %f168, %f142;
bra.uni $L__BB276_28;

$L__BB276_22:
ld.param.u32 %r373, [_Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.eq.s32 %p44, %r373, 3;
@%p44 bra $L__BB276_25;

ld.param.u32 %r374, [_Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.ne.s32 %p45, %r374, 6;
@%p45 bra $L__BB276_27;

ld.param.f32 %f151, [_Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+44];
min.ftz.f32 %f120, %f24, %f151;
neg.ftz.f32 %f121, %f151;
min.ftz.f32 %f122, %f168, %f151;
max.ftz.f32 %f123, %f122, %f121;
mul.ftz.f32 %f124, %f120, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f125, %f124;
add.ftz.f32 %f126, %f125, 0f3F800000;
div.approx.ftz.f32 %f127, %f120, %f126;
mul.ftz.f32 %f168, %f123, %f127;
bra.uni $L__BB276_28;

$L__BB276_27:
mul.ftz.f32 %f168, %f168, %f24;
bra.uni $L__BB276_28;

$L__BB276_26:
mul.ftz.f32 %f143, %f24, 0f3F000000;
mul.ftz.f32 %f144, %f24, 0f3D372713;
fma.rn.ftz.f32 %f145, %f24, %f144, 0f3F800000;
mul.ftz.f32 %f146, %f24, 0f3F4C422A;
mul.ftz.f32 %f147, %f146, %f145;
tanh.approx.f32 %f148, %f147;
add.ftz.f32 %f149, %f148, 0f3F800000;
mul.ftz.f32 %f150, %f143, %f149;
mul.ftz.f32 %f168, %f168, %f150;
bra.uni $L__BB276_28;

$L__BB276_25:
mov.f32 %f128, 0f40E00000;
min.ftz.f32 %f129, %f24, %f128;
min.ftz.f32 %f130, %f168, %f128;
mov.f32 %f131, 0fC0E00000;
max.ftz.f32 %f132, %f130, %f131;
mul.ftz.f32 %f133, %f129, 0f3FD9DB23;
mul.ftz.f32 %f134, %f133, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f135, %f134;
add.ftz.f32 %f136, %f135, 0f3F800000;
div.approx.ftz.f32 %f137, %f129, %f136;
add.ftz.f32 %f138, %f132, 0f3F800000;
mul.ftz.f32 %f168, %f138, %f137;

$L__BB276_28:
ld.param.u64 %rd87, [_Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4];
ld.param.u32 %r371, [_Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10];
ld.param.u32 %r370, [_Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13];
mov.u32 %r369, %ctaid.y;
mov.u32 %r368, %tid.y;
mad.lo.s32 %r364, %r368, %r371, %r14;
mad.lo.s32 %r365, %r369, %r370, %r364;
cvta.to.global.u64 %rd82, %rd87;
mul.wide.u32 %rd83, %r365, 4;
add.s64 %rd84, %rd82, %rd83;
st.global.f32 [%rd84], %f168;

$L__BB276_29:
ret;

}
.entry _Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj(
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2,
.param .align 8 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3[48],
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5,
.param .align 4 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6[12],
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15
)
.maxntid 256, 1, 1
.minnctapersm 1
{
.local .align 8 .b8 __local_depot277[8];
.reg .b64 %SP;
.reg .b64 %SPL;
.reg .pred %p<17>;
.reg .b16 %rs<15>;
.reg .f32 %f<44>;
.reg .b32 %r<172>;
.reg .b64 %rd<54>;


mov.u64 %SPL, __local_depot277;
ld.param.u64 %rd17, [_Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0];
ld.param.u64 %rd18, [_Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1];
ld.param.u64 %rd19, [_Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2];
ld.param.u64 %rd20, [_Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4];
ld.param.u32 %r20, [_Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5];
ld.param.u32 %r12, [_Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7];
ld.param.u32 %r13, [_Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8];
ld.param.u32 %r14, [_Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9];
ld.param.u32 %r15, [_Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10];
ld.param.u32 %r16, [_Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11];
ld.param.u32 %r17, [_Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12];
ld.param.u32 %r18, [_Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13];
ld.param.u32 %r21, [_Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14];
ld.param.u32 %r19, [_Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15];
ld.param.u32 %r11, [_Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+8];
ld.param.u32 %r10, [_Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+4];
ld.param.u32 %r9, [_Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6];
add.u64 %rd1, %SPL, 0;
mov.u32 %r22, %ctaid.x;
shl.b32 %r1, %r22, 1;
shr.u32 %r2, %r20, 5;
mov.u32 %r3, %tid.y;
setp.ge.u32 %p1, %r3, %r21;
@%p1 bra $L__BB277_7;

mov.f32 %f42, 0f00000000;
st.local.v2.f32 [%rd1], {%f42, %f42};
mov.u32 %r4, %tid.x;
shr.u32 %r171, %r4, 1;
setp.ge.u32 %p2, %r171, %r2;
mov.f32 %f43, %f42;
@%p2 bra $L__BB277_5;

mov.u32 %r23, %ctaid.y;
cvta.to.global.u64 %rd22, %rd18;
mul.hi.u32 %r24, %r23, %r9;
add.s32 %r25, %r24, %r23;
shr.u32 %r26, %r25, %r10;
mul.lo.s32 %r27, %r26, %r11;
sub.s32 %r28, %r23, %r27;
mul.lo.s32 %r29, %r28, %r17;
cvt.u64.u32 %rd23, %r29;
mul.lo.s32 %r30, %r3, %r14;
cvt.u64.u32 %rd24, %r30;
mad.lo.s32 %r31, %r3, %r19, %r23;
cvta.to.global.u64 %rd25, %rd19;
mul.wide.u32 %rd26, %r31, 4;
add.s64 %rd27, %rd25, %rd26;
ld.global.nc.u32 %r32, [%rd27];
mul.lo.s32 %r33, %r32, %r16;
cvt.u64.u32 %rd28, %r4;
shr.u64 %rd29, %rd28, 1;
mul.lo.s64 %rd30, %rd29, 36;
mul.wide.u32 %rd31, %r30, 36;
add.s64 %rd32, %rd30, %rd31;
add.s64 %rd3, %rd22, %rd32;
mul.wide.u32 %rd53, %r29, 36;
mad.lo.s32 %r34, %r1, %r13, %r171;
add.s32 %r35, %r34, %r33;
mul.wide.s32 %rd5, %r35, 18;
and.b32 %r36, %r4, 1;
mul.wide.u32 %rd33, %r36, 8;
add.s64 %rd34, %rd33, %rd5;
add.s64 %rd6, %rd34, 8;
add.s64 %rd35, %rd29, %rd24;
add.s64 %rd36, %rd35, %rd23;
mul.lo.s64 %rd37, %rd36, 36;
add.s64 %rd38, %rd37, %rd33;
add.s64 %rd39, %rd22, %rd38;
add.s64 %rd51, %rd39, 20;
add.s32 %r37, %r1, 1;
add.s32 %r38, %r171, %r33;
mad.lo.s32 %r39, %r13, %r37, %r38;
mul.wide.s32 %rd8, %r39, 18;
add.s64 %rd40, %rd33, %rd8;
add.s64 %rd9, %rd40, 8;
cvta.to.global.u64 %rd52, %rd17;
mov.f32 %f42, 0f00000000;

$L__BB277_3:
add.s64 %rd41, %rd3, %rd53;
add.s64 %rd42, %rd52, %rd6;
ld.global.nc.u16 %rs3, [%rd42+-6];
cvt.u32.u16 %r74, %rs3;
ld.global.nc.u16 %rs4, [%rd42+-4];
mov.b32 %r75, {%rs3, %rs4};
and.b32 %r76, %r75, -2004318072;
shr.u32 %r77, %r76, 1;
or.b32 %r78, %r77, 839922192;
and.b32 %r79, %r74, 30583;
mov.u32 %r80, -152379953;
mov.u32 %r81, -1079142271;
prmt.b32 %r82, %r81, %r80, %r79;
mov.u32 %r83, 1901675829;
mov.u32 %r84, 639175937;
prmt.b32 %r85, %r84, %r83, %r79;
and.b32 %r86, %r78, 30292;
prmt.b32 %r87, %r82, %r85, %r86;
and.b16 %rs5, %rs4, 30583;
cvt.u32.u16 %r88, %rs5;
prmt.b32 %r89, %r81, %r80, %r88;
prmt.b32 %r90, %r84, %r83, %r88;
shr.u32 %r91, %r78, 16;
prmt.b32 %r92, %r89, %r90, %r91;
mov.u32 %r93, 25632;
prmt.b32 %r41, %r87, %r92, %r93;
mov.u32 %r94, 30001;
prmt.b32 %r45, %r87, %r92, %r94;
ld.global.nc.u32 %r42, [%rd51+-16];
mov.u32 %r60, 0;

dp4a.s32.s32 %r40, %r41, %r42, %r60;

ld.global.nc.u32 %r46, [%rd51];

dp4a.s32.s32 %r44, %r45, %r46, %r40;

ld.global.nc.u16 %rs6, [%rd42+-2];
cvt.u32.u16 %r95, %rs6;
ld.global.nc.u16 %rs7, [%rd42];
mov.b32 %r96, {%rs6, %rs7};
and.b32 %r97, %r96, -2004318072;
shr.u32 %r98, %r97, 1;
or.b32 %r99, %r98, 839922192;
and.b32 %r100, %r95, 30583;
prmt.b32 %r101, %r81, %r80, %r100;
prmt.b32 %r102, %r84, %r83, %r100;
and.b32 %r103, %r99, 30292;
prmt.b32 %r104, %r101, %r102, %r103;
and.b16 %rs8, %rs7, 30583;
cvt.u32.u16 %r105, %rs8;
prmt.b32 %r106, %r81, %r80, %r105;
prmt.b32 %r107, %r84, %r83, %r105;
shr.u32 %r108, %r99, 16;
prmt.b32 %r109, %r106, %r107, %r108;
prmt.b32 %r49, %r104, %r109, %r93;
prmt.b32 %r53, %r104, %r109, %r94;
ld.global.nc.u32 %r50, [%rd51+-12];

dp4a.s32.s32 %r48, %r49, %r50, %r44;

ld.global.nc.u32 %r54, [%rd51+4];

dp4a.s32.s32 %r52, %r53, %r54, %r48;

add.s64 %rd43, %rd52, %rd5;
ld.global.nc.u16 %rs1, [%rd43];

{ cvt.f32.f16 %f11, %rs1;}


ld.global.nc.u32 %r56, [%rd41];

{.reg .f16 low,high;
mov.b32 {low,high},%r56;
cvt.f32.f16 %f12, low;}


mul.ftz.f32 %f15, %f11, %f12;
cvt.rn.f32.s32 %f16, %r52;
fma.rn.ftz.f32 %f43, %f15, %f16, %f43;
add.s64 %rd44, %rd52, %rd9;
ld.global.nc.u16 %rs9, [%rd44+-6];
cvt.u32.u16 %r110, %rs9;
ld.global.nc.u16 %rs10, [%rd44+-4];
mov.b32 %r111, {%rs9, %rs10};
and.b32 %r112, %r111, -2004318072;
shr.u32 %r113, %r112, 1;
or.b32 %r114, %r113, 839922192;
and.b32 %r115, %r110, 30583;
prmt.b32 %r116, %r81, %r80, %r115;
prmt.b32 %r117, %r84, %r83, %r115;
and.b32 %r118, %r114, 30292;
prmt.b32 %r119, %r116, %r117, %r118;
and.b16 %rs11, %rs10, 30583;
cvt.u32.u16 %r120, %rs11;
prmt.b32 %r121, %r81, %r80, %r120;
prmt.b32 %r122, %r84, %r83, %r120;
shr.u32 %r123, %r114, 16;
prmt.b32 %r124, %r121, %r122, %r123;
prmt.b32 %r58, %r119, %r124, %r93;
prmt.b32 %r62, %r119, %r124, %r94;

dp4a.s32.s32 %r57, %r58, %r42, %r60;


dp4a.s32.s32 %r61, %r62, %r46, %r57;

ld.global.nc.u16 %rs12, [%rd44+-2];
cvt.u32.u16 %r125, %rs12;
ld.global.nc.u16 %rs13, [%rd44];
mov.b32 %r126, {%rs12, %rs13};
and.b32 %r127, %r126, -2004318072;
shr.u32 %r128, %r127, 1;
or.b32 %r129, %r128, 839922192;
and.b32 %r130, %r125, 30583;
prmt.b32 %r131, %r81, %r80, %r130;
prmt.b32 %r132, %r84, %r83, %r130;
and.b32 %r133, %r129, 30292;
prmt.b32 %r134, %r131, %r132, %r133;
and.b16 %rs14, %rs13, 30583;
cvt.u32.u16 %r135, %rs14;
prmt.b32 %r136, %r81, %r80, %r135;
prmt.b32 %r137, %r84, %r83, %r135;
shr.u32 %r138, %r129, 16;
prmt.b32 %r139, %r136, %r137, %r138;
prmt.b32 %r66, %r134, %r139, %r93;
prmt.b32 %r70, %r134, %r139, %r94;

dp4a.s32.s32 %r65, %r66, %r50, %r61;


dp4a.s32.s32 %r69, %r70, %r54, %r65;

add.s64 %rd45, %rd52, %rd8;
ld.global.nc.u16 %rs2, [%rd45];

{ cvt.f32.f16 %f13, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r56;
cvt.f32.f16 %f14, low;}


mul.ftz.f32 %f17, %f13, %f14;
cvt.rn.f32.s32 %f18, %r69;
fma.rn.ftz.f32 %f42, %f17, %f18, %f42;
add.s64 %rd53, %rd53, 576;
add.s64 %rd52, %rd52, 288;
add.s64 %rd51, %rd51, 576;
add.s32 %r171, %r171, 16;
setp.lt.u32 %p3, %r171, %r2;
@%p3 bra $L__BB277_3;

st.local.v2.f32 [%rd1], {%f43, %f42};

$L__BB277_5:
mov.b32 %r140, %f43;
mov.u32 %r141, 31;
mov.u32 %r142, 16;
mov.u32 %r143, -1;
shfl.sync.bfly.b32 %r144|%p4, %r140, %r142, %r141, %r143;
mov.b32 %f19, %r144;
add.ftz.f32 %f20, %f43, %f19;
mov.b32 %r145, %f20;
mov.u32 %r146, 8;
shfl.sync.bfly.b32 %r147|%p5, %r145, %r146, %r141, %r143;
mov.b32 %f21, %r147;
add.ftz.f32 %f22, %f20, %f21;
mov.b32 %r148, %f22;
mov.u32 %r149, 4;
shfl.sync.bfly.b32 %r150|%p6, %r148, %r149, %r141, %r143;
mov.b32 %f23, %r150;
add.ftz.f32 %f24, %f22, %f23;
mov.b32 %r151, %f24;
mov.u32 %r152, 2;
shfl.sync.bfly.b32 %r153|%p7, %r151, %r152, %r141, %r143;
mov.b32 %f25, %r153;
add.ftz.f32 %f26, %f24, %f25;
mov.b32 %r154, %f26;
mov.u32 %r155, 1;
shfl.sync.bfly.b32 %r156|%p8, %r154, %r155, %r141, %r143;
mov.b32 %f27, %r156;
add.ftz.f32 %f28, %f26, %f27;
st.local.f32 [%rd1], %f28;
mov.b32 %r157, %f42;
shfl.sync.bfly.b32 %r158|%p9, %r157, %r142, %r141, %r143;
mov.b32 %f29, %r158;
add.ftz.f32 %f30, %f42, %f29;
mov.b32 %r159, %f30;
shfl.sync.bfly.b32 %r160|%p10, %r159, %r146, %r141, %r143;
mov.b32 %f31, %r160;
add.ftz.f32 %f32, %f30, %f31;
mov.b32 %r161, %f32;
shfl.sync.bfly.b32 %r162|%p11, %r161, %r149, %r141, %r143;
mov.b32 %f33, %r162;
add.ftz.f32 %f34, %f32, %f33;
mov.b32 %r163, %f34;
shfl.sync.bfly.b32 %r164|%p12, %r163, %r152, %r141, %r143;
mov.b32 %f35, %r164;
add.ftz.f32 %f36, %f34, %f35;
mov.b32 %r165, %f36;
shfl.sync.bfly.b32 %r166|%p13, %r165, %r155, %r141, %r143;
mov.b32 %f37, %r166;
add.ftz.f32 %f38, %f36, %f37;
st.local.f32 [%rd1+4], %f38;
add.s32 %r8, %r1, %r4;
setp.ge.u32 %p14, %r8, %r12;
setp.gt.u32 %p15, %r4, 1;
or.pred %p16, %p15, %p14;
@%p16 bra $L__BB277_7;

mul.wide.u32 %rd46, %r4, 4;
add.s64 %rd47, %rd1, %rd46;
ld.local.f32 %f39, [%rd47];
mov.u32 %r167, %ctaid.y;
mad.lo.s32 %r169, %r3, %r15, %r8;
mad.lo.s32 %r170, %r167, %r18, %r169;
cvta.to.global.u64 %rd48, %rd20;
mul.wide.u32 %rd49, %r170, 4;
add.s64 %rd50, %rd48, %rd49;
st.global.f32 [%rd50], %f39;

$L__BB277_7:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<320>;
.reg .b16 %rs<57>;
.reg .f32 %f<1005>;
.reg .b32 %r<1266>;
.reg .b64 %rd<91>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate[1536];

ld.param.u64 %rd28, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd29, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd30, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r47, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r48, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r49, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r53, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r54, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r55, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r51, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r50, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r45, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
ld.param.u64 %rd33, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+16];
ld.param.u64 %rd32, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+8];
ld.param.u64 %rd31, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3];
mov.u32 %r1, %tid.y;
shl.b32 %r56, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r56, %r2;
mov.u32 %r57, %ctaid.x;
shl.b32 %r4, %r57, 2;
mov.u32 %r5, %ctaid.y;
setp.eq.s64 %p1, %rd30, 0;
@%p1 bra $L__BB278_2;

cvta.to.global.u64 %rd37, %rd30;
mul.wide.u32 %rd38, %r5, 4;
add.s64 %rd39, %rd37, %rd38;
ld.global.nc.u32 %r1263, [%rd39];
mul.hi.u32 %r58, %r5, %r39;
add.s32 %r59, %r58, %r5;
shr.u32 %r60, %r59, %r40;
mul.lo.s32 %r61, %r60, %r41;
sub.s32 %r1264, %r5, %r61;
bra.uni $L__BB278_3;

$L__BB278_2:
mul.hi.u32 %r62, %r5, %r44;
add.s32 %r63, %r62, %r5;
shr.u32 %r1263, %r63, %r45;
mov.u32 %r1264, %r5;

$L__BB278_3:
mov.u32 %r11, %ctaid.z;
selp.b32 %r12, %r5, %r1263, %p1;
setp.ne.s32 %p3, %r1, 0;
setp.gt.u32 %p4, %r2, 3;
or.pred %p5, %p4, %p3;
add.s32 %r64, %r4, %r2;
setp.ge.u32 %p6, %r64, %r43;
mov.f32 %f27, 0f00000000;
or.pred %p7, %p5, %p6;
mov.f32 %f4, %f27;
mov.f32 %f5, %f27;
@%p7 bra $L__BB278_8;

setp.eq.s64 %p8, %rd31, 0;
mov.f32 %f4, 0f00000000;
mov.f32 %f5, %f4;
@%p8 bra $L__BB278_6;

cvta.to.global.u64 %rd40, %rd31;
mul.lo.s32 %r65, %r11, %r55;
cvt.u64.u32 %rd41, %r65;
mul.lo.s32 %r66, %r12, %r49;
cvt.u64.u32 %rd42, %r66;
cvt.u64.u32 %rd43, %r2;
cvt.s64.s32 %rd44, %r4;
add.s64 %rd45, %rd44, %rd43;
add.s64 %rd46, %rd45, %rd41;
add.s64 %rd47, %rd46, %rd42;
shl.b64 %rd48, %rd47, 2;
add.s64 %rd49, %rd40, %rd48;
ld.global.f32 %f5, [%rd49];

$L__BB278_6:
setp.eq.s64 %p9, %rd33, 0;
setp.eq.s64 %p10, %rd32, 0;
or.pred %p11, %p9, %p10;
@%p11 bra $L__BB278_8;

cvta.to.global.u64 %rd50, %rd33;
mul.lo.s32 %r67, %r11, %r55;
cvt.u64.u32 %rd51, %r67;
mul.lo.s32 %r68, %r12, %r49;
cvt.u64.u32 %rd52, %r68;
cvt.u64.u32 %rd53, %r2;
cvt.s64.s32 %rd54, %r4;
add.s64 %rd55, %rd54, %rd53;
add.s64 %rd56, %rd55, %rd51;
add.s64 %rd57, %rd56, %rd52;
shl.b64 %rd58, %rd57, 2;
add.s64 %rd59, %rd50, %rd58;
ld.global.f32 %f4, [%rd59];

$L__BB278_8:
mul.lo.s32 %r69, %r11, %r54;
cvt.u64.u32 %rd60, %r69;
mul.lo.s32 %r70, %r1264, %r48;
cvt.u64.u32 %rd61, %r70;
add.s64 %rd4, %rd61, %rd60;
mul.hi.u32 %r71, %r11, %r50;
add.s32 %r72, %r71, %r11;
shr.u32 %r13, %r72, %r51;
shr.u32 %r73, %r3, 31;
add.s32 %r74, %r3, %r73;
shr.s32 %r1265, %r74, 1;
shr.u32 %r75, %r38, 5;
setp.ge.s32 %p12, %r1265, %r75;
mov.f32 %f28, %f27;
mov.f32 %f29, %f27;
mov.f32 %f30, %f27;
mov.f32 %f31, %f27;
mov.f32 %f32, %f27;
mov.f32 %f33, %f27;
mov.f32 %f34, %f27;
@%p12 bra $L__BB278_19;

mul.lo.s32 %r76, %r13, %r53;
mul.lo.s32 %r77, %r1263, %r47;
and.b32 %r80, %r74, -2;
sub.s32 %r81, %r3, %r80;
shl.b32 %r82, %r81, 1;
cvt.s64.s32 %rd5, %r82;
shl.b32 %r83, %r81, 2;
add.s32 %r84, %r1265, %r76;
add.s32 %r85, %r84, %r77;
add.s32 %r86, %r4, 1;
mad.lo.s32 %r87, %r42, %r86, %r85;
mul.wide.s32 %rd6, %r87, 18;
mad.lo.s32 %r88, %r4, %r42, %r1265;
add.s32 %r89, %r88, %r76;
add.s32 %r90, %r89, %r77;
mul.wide.s32 %rd7, %r90, 18;
mul.wide.s32 %rd62, %r83, 2;
add.s64 %rd63, %rd7, %rd62;
add.s64 %rd8, %rd63, 8;
add.s64 %rd64, %rd6, %rd62;
add.s64 %rd9, %rd64, 4;
add.s32 %r91, %r4, 3;
mad.lo.s32 %r92, %r42, %r91, %r85;
mul.wide.s32 %rd10, %r92, 18;
add.s64 %rd65, %rd10, %rd62;
add.s64 %rd11, %rd65, 8;
add.s32 %r93, %r4, 2;
mad.lo.s32 %r94, %r42, %r93, %r85;
mul.wide.s32 %rd12, %r94, 18;
add.s64 %rd13, %rd12, %rd62;
mul.wide.s32 %rd14, %r42, 18;
cvta.to.global.u64 %rd90, %rd32;
cvta.to.global.u64 %rd89, %rd28;
mov.f32 %f27, 0f00000000;
cvta.to.global.u64 %rd66, %rd29;
shl.b64 %rd71, %rd5, 2;
mov.f32 %f28, %f27;
mov.f32 %f29, %f27;
mov.f32 %f30, %f27;
mov.f32 %f31, %f27;
mov.f32 %f32, %f27;
mov.f32 %f33, %f27;
mov.f32 %f34, %f27;

$L__BB278_10:
cvt.u64.u32 %rd67, %r1265;
add.s64 %rd68, %rd4, %rd67;
mul.lo.s64 %rd69, %rd68, 36;
add.s64 %rd70, %rd66, %rd69;
add.s64 %rd72, %rd70, %rd71;
add.s64 %rd73, %rd89, %rd8;
ld.global.nc.u16 %rs2, [%rd73+-6];
cvt.u32.u16 %r112, %rs2;
ld.global.nc.u16 %rs3, [%rd73+-4];
mov.b32 %r113, {%rs2, %rs3};
and.b32 %r114, %r113, -2004318072;
shr.u32 %r115, %r114, 1;
or.b32 %r116, %r115, 839922192;
and.b32 %r117, %r112, 30583;
mov.u32 %r118, -152379953;
mov.u32 %r119, -1079142271;
prmt.b32 %r120, %r119, %r118, %r117;
mov.u32 %r121, 1901675829;
mov.u32 %r122, 639175937;
prmt.b32 %r123, %r122, %r121, %r117;
and.b32 %r124, %r116, 30292;
prmt.b32 %r125, %r120, %r123, %r124;
and.b16 %rs4, %rs3, 30583;
cvt.u32.u16 %r126, %rs4;
prmt.b32 %r127, %r119, %r118, %r126;
prmt.b32 %r128, %r122, %r121, %r126;
shr.u32 %r129, %r116, 16;
prmt.b32 %r130, %r127, %r128, %r129;
mov.u32 %r131, 25632;
prmt.b32 %r96, %r125, %r130, %r131;
mov.u32 %r132, 30001;
prmt.b32 %r100, %r125, %r130, %r132;
ld.global.nc.u32 %r16, [%rd72+4];
mov.u32 %r98, 0;

dp4a.s32.s32 %r95, %r96, %r16, %r98;

ld.global.nc.u32 %r17, [%rd72+20];

dp4a.s32.s32 %r99, %r100, %r17, %r95;

ld.global.nc.u16 %rs5, [%rd73+-2];
cvt.u32.u16 %r133, %rs5;
ld.global.nc.u16 %rs6, [%rd73];
mov.b32 %r134, {%rs5, %rs6};
and.b32 %r135, %r134, -2004318072;
shr.u32 %r136, %r135, 1;
or.b32 %r137, %r136, 839922192;
and.b32 %r138, %r133, 30583;
prmt.b32 %r139, %r119, %r118, %r138;
prmt.b32 %r140, %r122, %r121, %r138;
and.b32 %r141, %r137, 30292;
prmt.b32 %r142, %r139, %r140, %r141;
and.b16 %rs7, %rs6, 30583;
cvt.u32.u16 %r143, %rs7;
prmt.b32 %r144, %r119, %r118, %r143;
prmt.b32 %r145, %r122, %r121, %r143;
shr.u32 %r146, %r137, 16;
prmt.b32 %r147, %r144, %r145, %r146;
prmt.b32 %r104, %r142, %r147, %r131;
prmt.b32 %r108, %r142, %r147, %r132;
ld.global.nc.u32 %r18, [%rd72+8];

dp4a.s32.s32 %r103, %r104, %r18, %r99;

ld.global.nc.u32 %r19, [%rd72+24];

dp4a.s32.s32 %r107, %r108, %r19, %r103;

add.s64 %rd20, %rd89, %rd7;
ld.global.nc.u16 %rs1, [%rd20];

{ cvt.f32.f16 %f101, %rs1;}


ld.global.nc.u32 %r20, [%rd70];

{.reg .f16 low,high;
mov.b32 {low,high},%r20;
cvt.f32.f16 %f102, low;}


mul.ftz.f32 %f103, %f101, %f102;
cvt.rn.f32.s32 %f104, %r107;
fma.rn.ftz.f32 %f30, %f103, %f104, %f30;
setp.eq.s64 %p13, %rd32, 0;
@%p13 bra $L__BB278_12;

add.s64 %rd74, %rd90, %rd8;
ld.global.u16 %rs9, [%rd74+-6];
cvt.u32.u16 %r165, %rs9;
ld.global.u16 %rs10, [%rd74+-4];
mov.b32 %r166, {%rs9, %rs10};
and.b32 %r167, %r166, -2004318072;
shr.u32 %r168, %r167, 1;
or.b32 %r169, %r168, 839922192;
and.b32 %r170, %r165, 30583;
prmt.b32 %r173, %r119, %r118, %r170;
prmt.b32 %r176, %r122, %r121, %r170;
and.b32 %r177, %r169, 30292;
prmt.b32 %r178, %r173, %r176, %r177;
and.b16 %rs11, %rs10, 30583;
cvt.u32.u16 %r179, %rs11;
prmt.b32 %r180, %r119, %r118, %r179;
prmt.b32 %r181, %r122, %r121, %r179;
shr.u32 %r182, %r169, 16;
prmt.b32 %r183, %r180, %r181, %r182;
prmt.b32 %r149, %r178, %r183, %r131;
prmt.b32 %r153, %r178, %r183, %r132;

dp4a.s32.s32 %r148, %r149, %r16, %r98;


dp4a.s32.s32 %r152, %r153, %r17, %r148;

ld.global.u16 %rs12, [%rd74+-2];
cvt.u32.u16 %r186, %rs12;
ld.global.u16 %rs13, [%rd74];
mov.b32 %r187, {%rs12, %rs13};
and.b32 %r188, %r187, -2004318072;
shr.u32 %r189, %r188, 1;
or.b32 %r190, %r189, 839922192;
and.b32 %r191, %r186, 30583;
prmt.b32 %r192, %r119, %r118, %r191;
prmt.b32 %r193, %r122, %r121, %r191;
and.b32 %r194, %r190, 30292;
prmt.b32 %r195, %r192, %r193, %r194;
and.b16 %rs14, %rs13, 30583;
cvt.u32.u16 %r196, %rs14;
prmt.b32 %r197, %r119, %r118, %r196;
prmt.b32 %r198, %r122, %r121, %r196;
shr.u32 %r199, %r190, 16;
prmt.b32 %r200, %r197, %r198, %r199;
prmt.b32 %r157, %r195, %r200, %r131;
prmt.b32 %r161, %r195, %r200, %r132;

dp4a.s32.s32 %r156, %r157, %r18, %r152;


dp4a.s32.s32 %r160, %r161, %r19, %r156;

add.s64 %rd75, %rd90, %rd7;
ld.global.u16 %rs8, [%rd75];

{ cvt.f32.f16 %f105, %rs8;}



{.reg .f16 low,high;
mov.b32 {low,high},%r20;
cvt.f32.f16 %f106, low;}


mul.ftz.f32 %f107, %f105, %f106;
cvt.rn.f32.s32 %f108, %r160;
fma.rn.ftz.f32 %f34, %f107, %f108, %f34;

$L__BB278_12:
add.s64 %rd87, %rd89, %rd8;
add.s64 %rd86, %rd87, -6;
add.s64 %rd21, %rd86, %rd14;
ld.global.nc.u16 %rs16, [%rd21];
cvt.u32.u16 %r218, %rs16;
ld.global.nc.u16 %rs17, [%rd21+2];
mov.b32 %r219, {%rs16, %rs17};
and.b32 %r220, %r219, -2004318072;
shr.u32 %r221, %r220, 1;
or.b32 %r222, %r221, 839922192;
and.b32 %r223, %r218, 30583;
prmt.b32 %r226, %r119, %r118, %r223;
prmt.b32 %r229, %r122, %r121, %r223;
and.b32 %r230, %r222, 30292;
prmt.b32 %r231, %r226, %r229, %r230;
and.b16 %rs18, %rs17, 30583;
cvt.u32.u16 %r232, %rs18;
prmt.b32 %r233, %r119, %r118, %r232;
prmt.b32 %r234, %r122, %r121, %r232;
shr.u32 %r235, %r222, 16;
prmt.b32 %r236, %r233, %r234, %r235;
prmt.b32 %r202, %r231, %r236, %r131;
prmt.b32 %r206, %r231, %r236, %r132;

dp4a.s32.s32 %r201, %r202, %r16, %r98;


dp4a.s32.s32 %r205, %r206, %r17, %r201;

ld.global.nc.u16 %rs19, [%rd21+4];
cvt.u32.u16 %r239, %rs19;
ld.global.nc.u16 %rs20, [%rd21+6];
mov.b32 %r240, {%rs19, %rs20};
and.b32 %r241, %r240, -2004318072;
shr.u32 %r242, %r241, 1;
or.b32 %r243, %r242, 839922192;
and.b32 %r244, %r239, 30583;
prmt.b32 %r245, %r119, %r118, %r244;
prmt.b32 %r246, %r122, %r121, %r244;
and.b32 %r247, %r243, 30292;
prmt.b32 %r248, %r245, %r246, %r247;
and.b16 %rs21, %rs20, 30583;
cvt.u32.u16 %r249, %rs21;
prmt.b32 %r250, %r119, %r118, %r249;
prmt.b32 %r251, %r122, %r121, %r249;
shr.u32 %r252, %r243, 16;
prmt.b32 %r253, %r250, %r251, %r252;
prmt.b32 %r210, %r248, %r253, %r131;
prmt.b32 %r214, %r248, %r253, %r132;

dp4a.s32.s32 %r209, %r210, %r18, %r205;


dp4a.s32.s32 %r213, %r214, %r19, %r209;

add.s64 %rd22, %rd20, %rd14;
ld.global.nc.u16 %rs15, [%rd22];

{ cvt.f32.f16 %f109, %rs15;}



{.reg .f16 low,high;
mov.b32 {low,high},%r20;
cvt.f32.f16 %f110, low;}


mul.ftz.f32 %f111, %f109, %f110;
cvt.rn.f32.s32 %f112, %r213;
fma.rn.ftz.f32 %f29, %f111, %f112, %f29;
@%p13 bra $L__BB278_14;

add.s64 %rd76, %rd90, %rd9;
ld.global.u16 %rs23, [%rd76+-2];
cvt.u32.u16 %r271, %rs23;
ld.global.u16 %rs24, [%rd76];
mov.b32 %r272, {%rs23, %rs24};
and.b32 %r273, %r272, -2004318072;
shr.u32 %r274, %r273, 1;
or.b32 %r275, %r274, 839922192;
and.b32 %r276, %r271, 30583;
mov.u32 %r277, -152379953;
mov.u32 %r278, -1079142271;
prmt.b32 %r279, %r278, %r277, %r276;
mov.u32 %r280, 1901675829;
mov.u32 %r281, 639175937;
prmt.b32 %r282, %r281, %r280, %r276;
and.b32 %r283, %r275, 30292;
prmt.b32 %r284, %r279, %r282, %r283;
and.b16 %rs25, %rs24, 30583;
cvt.u32.u16 %r285, %rs25;
prmt.b32 %r286, %r278, %r277, %r285;
prmt.b32 %r287, %r281, %r280, %r285;
shr.u32 %r288, %r275, 16;
prmt.b32 %r289, %r286, %r287, %r288;
mov.u32 %r290, 25632;
prmt.b32 %r255, %r284, %r289, %r290;
mov.u32 %r291, 30001;
prmt.b32 %r259, %r284, %r289, %r291;
mov.u32 %r257, 0;

dp4a.s32.s32 %r254, %r255, %r16, %r257;


dp4a.s32.s32 %r258, %r259, %r17, %r254;

ld.global.u16 %rs26, [%rd76+2];
cvt.u32.u16 %r292, %rs26;
ld.global.u16 %rs27, [%rd76+4];
mov.b32 %r293, {%rs26, %rs27};
and.b32 %r294, %r293, -2004318072;
shr.u32 %r295, %r294, 1;
or.b32 %r296, %r295, 839922192;
and.b32 %r297, %r292, 30583;
prmt.b32 %r298, %r278, %r277, %r297;
prmt.b32 %r299, %r281, %r280, %r297;
and.b32 %r300, %r296, 30292;
prmt.b32 %r301, %r298, %r299, %r300;
and.b16 %rs28, %rs27, 30583;
cvt.u32.u16 %r302, %rs28;
prmt.b32 %r303, %r278, %r277, %r302;
prmt.b32 %r304, %r281, %r280, %r302;
shr.u32 %r305, %r296, 16;
prmt.b32 %r306, %r303, %r304, %r305;
prmt.b32 %r263, %r301, %r306, %r290;
prmt.b32 %r267, %r301, %r306, %r291;

dp4a.s32.s32 %r262, %r263, %r18, %r258;


dp4a.s32.s32 %r266, %r267, %r19, %r262;

add.s64 %rd77, %rd90, %rd6;
ld.global.u16 %rs22, [%rd77];

{ cvt.f32.f16 %f113, %rs22;}



{.reg .f16 low,high;
mov.b32 {low,high},%r20;
cvt.f32.f16 %f114, low;}


mul.ftz.f32 %f115, %f113, %f114;
cvt.rn.f32.s32 %f116, %r266;
fma.rn.ftz.f32 %f33, %f115, %f116, %f33;

$L__BB278_14:
add.s64 %rd23, %rd21, %rd14;
ld.global.nc.u16 %rs30, [%rd23];
cvt.u32.u16 %r324, %rs30;
ld.global.nc.u16 %rs31, [%rd23+2];
mov.b32 %r325, {%rs30, %rs31};
and.b32 %r326, %r325, -2004318072;
shr.u32 %r327, %r326, 1;
or.b32 %r328, %r327, 839922192;
and.b32 %r329, %r324, 30583;
mov.u32 %r330, -152379953;
mov.u32 %r331, -1079142271;
prmt.b32 %r332, %r331, %r330, %r329;
mov.u32 %r333, 1901675829;
mov.u32 %r334, 639175937;
prmt.b32 %r335, %r334, %r333, %r329;
and.b32 %r336, %r328, 30292;
prmt.b32 %r337, %r332, %r335, %r336;
and.b16 %rs32, %rs31, 30583;
cvt.u32.u16 %r338, %rs32;
prmt.b32 %r339, %r331, %r330, %r338;
prmt.b32 %r340, %r334, %r333, %r338;
shr.u32 %r341, %r328, 16;
prmt.b32 %r342, %r339, %r340, %r341;
mov.u32 %r343, 25632;
prmt.b32 %r308, %r337, %r342, %r343;
mov.u32 %r344, 30001;
prmt.b32 %r312, %r337, %r342, %r344;
mov.u32 %r310, 0;

dp4a.s32.s32 %r307, %r308, %r16, %r310;


dp4a.s32.s32 %r311, %r312, %r17, %r307;

ld.global.nc.u16 %rs33, [%rd23+4];
cvt.u32.u16 %r345, %rs33;
ld.global.nc.u16 %rs34, [%rd23+6];
mov.b32 %r346, {%rs33, %rs34};
and.b32 %r347, %r346, -2004318072;
shr.u32 %r348, %r347, 1;
or.b32 %r349, %r348, 839922192;
and.b32 %r350, %r345, 30583;
prmt.b32 %r351, %r331, %r330, %r350;
prmt.b32 %r352, %r334, %r333, %r350;
and.b32 %r353, %r349, 30292;
prmt.b32 %r354, %r351, %r352, %r353;
and.b16 %rs35, %rs34, 30583;
cvt.u32.u16 %r355, %rs35;
prmt.b32 %r356, %r331, %r330, %r355;
prmt.b32 %r357, %r334, %r333, %r355;
shr.u32 %r358, %r349, 16;
prmt.b32 %r359, %r356, %r357, %r358;
prmt.b32 %r316, %r354, %r359, %r343;
prmt.b32 %r320, %r354, %r359, %r344;

dp4a.s32.s32 %r315, %r316, %r18, %r311;


dp4a.s32.s32 %r319, %r320, %r19, %r315;

add.s64 %rd24, %rd22, %rd14;
ld.global.nc.u16 %rs29, [%rd24];

{ cvt.f32.f16 %f117, %rs29;}



{.reg .f16 low,high;
mov.b32 {low,high},%r20;
cvt.f32.f16 %f118, low;}


mul.ftz.f32 %f119, %f117, %f118;
cvt.rn.f32.s32 %f120, %r319;
fma.rn.ftz.f32 %f28, %f119, %f120, %f28;
@%p13 bra $L__BB278_16;

add.s64 %rd78, %rd90, %rd13;
ld.global.u16 %rs37, [%rd78+2];
cvt.u32.u16 %r377, %rs37;
ld.global.u16 %rs38, [%rd78+4];
mov.b32 %r378, {%rs37, %rs38};
and.b32 %r379, %r378, -2004318072;
shr.u32 %r380, %r379, 1;
or.b32 %r381, %r380, 839922192;
and.b32 %r382, %r377, 30583;
prmt.b32 %r385, %r331, %r330, %r382;
prmt.b32 %r388, %r334, %r333, %r382;
and.b32 %r389, %r381, 30292;
prmt.b32 %r390, %r385, %r388, %r389;
and.b16 %rs39, %rs38, 30583;
cvt.u32.u16 %r391, %rs39;
prmt.b32 %r392, %r331, %r330, %r391;
prmt.b32 %r393, %r334, %r333, %r391;
shr.u32 %r394, %r381, 16;
prmt.b32 %r395, %r392, %r393, %r394;
prmt.b32 %r361, %r390, %r395, %r343;
prmt.b32 %r365, %r390, %r395, %r344;

dp4a.s32.s32 %r360, %r361, %r16, %r310;


dp4a.s32.s32 %r364, %r365, %r17, %r360;

ld.global.u16 %rs40, [%rd78+6];
cvt.u32.u16 %r398, %rs40;
ld.global.u16 %rs41, [%rd78+8];
mov.b32 %r399, {%rs40, %rs41};
and.b32 %r400, %r399, -2004318072;
shr.u32 %r401, %r400, 1;
or.b32 %r402, %r401, 839922192;
and.b32 %r403, %r398, 30583;
prmt.b32 %r404, %r331, %r330, %r403;
prmt.b32 %r405, %r334, %r333, %r403;
and.b32 %r406, %r402, 30292;
prmt.b32 %r407, %r404, %r405, %r406;
and.b16 %rs42, %rs41, 30583;
cvt.u32.u16 %r408, %rs42;
prmt.b32 %r409, %r331, %r330, %r408;
prmt.b32 %r410, %r334, %r333, %r408;
shr.u32 %r411, %r402, 16;
prmt.b32 %r412, %r409, %r410, %r411;
prmt.b32 %r369, %r407, %r412, %r343;
prmt.b32 %r373, %r407, %r412, %r344;

dp4a.s32.s32 %r368, %r369, %r18, %r364;


dp4a.s32.s32 %r372, %r373, %r19, %r368;

add.s64 %rd79, %rd90, %rd12;
ld.global.u16 %rs36, [%rd79];

{ cvt.f32.f16 %f121, %rs36;}



{.reg .f16 low,high;
mov.b32 {low,high},%r20;
cvt.f32.f16 %f122, low;}


mul.ftz.f32 %f123, %f121, %f122;
cvt.rn.f32.s32 %f124, %r372;
fma.rn.ftz.f32 %f32, %f123, %f124, %f32;

$L__BB278_16:
add.s64 %rd80, %rd23, %rd14;
ld.global.nc.u16 %rs44, [%rd80];
cvt.u32.u16 %r430, %rs44;
ld.global.nc.u16 %rs45, [%rd80+2];
mov.b32 %r431, {%rs44, %rs45};
and.b32 %r432, %r431, -2004318072;
shr.u32 %r433, %r432, 1;
or.b32 %r434, %r433, 839922192;
and.b32 %r435, %r430, 30583;
prmt.b32 %r438, %r331, %r330, %r435;
prmt.b32 %r441, %r334, %r333, %r435;
and.b32 %r442, %r434, 30292;
prmt.b32 %r443, %r438, %r441, %r442;
and.b16 %rs46, %rs45, 30583;
cvt.u32.u16 %r444, %rs46;
prmt.b32 %r445, %r331, %r330, %r444;
prmt.b32 %r446, %r334, %r333, %r444;
shr.u32 %r447, %r434, 16;
prmt.b32 %r448, %r445, %r446, %r447;
prmt.b32 %r414, %r443, %r448, %r343;
prmt.b32 %r418, %r443, %r448, %r344;

dp4a.s32.s32 %r413, %r414, %r16, %r310;


dp4a.s32.s32 %r417, %r418, %r17, %r413;

ld.global.nc.u16 %rs47, [%rd80+4];
cvt.u32.u16 %r451, %rs47;
ld.global.nc.u16 %rs48, [%rd80+6];
mov.b32 %r452, {%rs47, %rs48};
and.b32 %r453, %r452, -2004318072;
shr.u32 %r454, %r453, 1;
or.b32 %r455, %r454, 839922192;
and.b32 %r456, %r451, 30583;
prmt.b32 %r457, %r331, %r330, %r456;
prmt.b32 %r458, %r334, %r333, %r456;
and.b32 %r459, %r455, 30292;
prmt.b32 %r460, %r457, %r458, %r459;
and.b16 %rs49, %rs48, 30583;
cvt.u32.u16 %r461, %rs49;
prmt.b32 %r462, %r331, %r330, %r461;
prmt.b32 %r463, %r334, %r333, %r461;
shr.u32 %r464, %r455, 16;
prmt.b32 %r465, %r462, %r463, %r464;
prmt.b32 %r422, %r460, %r465, %r343;
prmt.b32 %r426, %r460, %r465, %r344;

dp4a.s32.s32 %r421, %r422, %r18, %r417;


dp4a.s32.s32 %r425, %r426, %r19, %r421;

add.s64 %rd81, %rd24, %rd14;
ld.global.nc.u16 %rs43, [%rd81];

{ cvt.f32.f16 %f125, %rs43;}



{.reg .f16 low,high;
mov.b32 {low,high},%r20;
cvt.f32.f16 %f126, low;}


mul.ftz.f32 %f127, %f125, %f126;
cvt.rn.f32.s32 %f128, %r425;
fma.rn.ftz.f32 %f27, %f127, %f128, %f27;
@%p13 bra $L__BB278_18;

add.s64 %rd82, %rd90, %rd11;
ld.global.u16 %rs51, [%rd82+-6];
cvt.u32.u16 %r483, %rs51;
ld.global.u16 %rs52, [%rd82+-4];
mov.b32 %r484, {%rs51, %rs52};
and.b32 %r485, %r484, -2004318072;
shr.u32 %r486, %r485, 1;
or.b32 %r487, %r486, 839922192;
and.b32 %r488, %r483, 30583;
mov.u32 %r489, -152379953;
mov.u32 %r490, -1079142271;
prmt.b32 %r491, %r490, %r489, %r488;
mov.u32 %r492, 1901675829;
mov.u32 %r493, 639175937;
prmt.b32 %r494, %r493, %r492, %r488;
and.b32 %r495, %r487, 30292;
prmt.b32 %r496, %r491, %r494, %r495;
and.b16 %rs53, %rs52, 30583;
cvt.u32.u16 %r497, %rs53;
prmt.b32 %r498, %r490, %r489, %r497;
prmt.b32 %r499, %r493, %r492, %r497;
shr.u32 %r500, %r487, 16;
prmt.b32 %r501, %r498, %r499, %r500;
mov.u32 %r502, 25632;
prmt.b32 %r467, %r496, %r501, %r502;
mov.u32 %r503, 30001;
prmt.b32 %r471, %r496, %r501, %r503;
mov.u32 %r469, 0;

dp4a.s32.s32 %r466, %r467, %r16, %r469;


dp4a.s32.s32 %r470, %r471, %r17, %r466;

ld.global.u16 %rs54, [%rd82+-2];
cvt.u32.u16 %r504, %rs54;
ld.global.u16 %rs55, [%rd82];
mov.b32 %r505, {%rs54, %rs55};
and.b32 %r506, %r505, -2004318072;
shr.u32 %r507, %r506, 1;
or.b32 %r508, %r507, 839922192;
and.b32 %r509, %r504, 30583;
prmt.b32 %r510, %r490, %r489, %r509;
prmt.b32 %r511, %r493, %r492, %r509;
and.b32 %r512, %r508, 30292;
prmt.b32 %r513, %r510, %r511, %r512;
and.b16 %rs56, %rs55, 30583;
cvt.u32.u16 %r514, %rs56;
prmt.b32 %r515, %r490, %r489, %r514;
prmt.b32 %r516, %r493, %r492, %r514;
shr.u32 %r517, %r508, 16;
prmt.b32 %r518, %r515, %r516, %r517;
prmt.b32 %r475, %r513, %r518, %r502;
prmt.b32 %r479, %r513, %r518, %r503;

dp4a.s32.s32 %r474, %r475, %r18, %r470;


dp4a.s32.s32 %r478, %r479, %r19, %r474;

add.s64 %rd83, %rd90, %rd10;
ld.global.u16 %rs50, [%rd83];

{ cvt.f32.f16 %f129, %rs50;}



{.reg .f16 low,high;
mov.b32 {low,high},%r20;
cvt.f32.f16 %f130, low;}


mul.ftz.f32 %f131, %f129, %f130;
cvt.rn.f32.s32 %f132, %r478;
fma.rn.ftz.f32 %f31, %f131, %f132, %f31;

$L__BB278_18:
add.s64 %rd90, %rd90, 1152;
add.s64 %rd89, %rd89, 1152;
add.s32 %r1265, %r1265, 64;
setp.lt.s32 %p17, %r1265, %r75;
@%p17 bra $L__BB278_10;

$L__BB278_19:
mov.u32 %r1252, %tid.y;
setp.eq.s32 %p18, %r1252, 0;
@%p18 bra $L__BB278_23;

mov.u32 %r1253, %tid.y;
setp.ne.s64 %p19, %rd32, 0;
add.s32 %r22, %r1253, -1;
shl.b32 %r523, %r1253, 9;
mov.u32 %r524, _ZZ13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r525, %r524, %r523;
shl.b32 %r526, %r2, 2;
add.s32 %r527, %r525, %r526;
st.shared.f32 [%r527+-512], %f30;
@%p19 bra $L__BB278_22;
bra.uni $L__BB278_21;

$L__BB278_22:
shl.b32 %r529, %r22, 9;
mov.u32 %r530, _ZZ13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r531, %r530, %r529;
add.s32 %r533, %r531, %r526;
st.shared.f32 [%r533], %f34;
add.s32 %r1249, %r527, -512;
st.shared.f32 [%r1249+128], %f29;
st.shared.f32 [%r533+128], %f33;
add.s32 %r1250, %r527, -512;
st.shared.f32 [%r1250+256], %f28;
st.shared.f32 [%r533+256], %f32;
add.s32 %r1251, %r527, -512;
st.shared.f32 [%r1251+384], %f27;
st.shared.f32 [%r533+384], %f31;
bra.uni $L__BB278_23;

$L__BB278_21:
add.s32 %r1246, %r527, -512;
st.shared.f32 [%r1246+128], %f29;
add.s32 %r1247, %r527, -512;
st.shared.f32 [%r1247+256], %f28;
add.s32 %r1248, %r527, -512;
st.shared.f32 [%r1248+384], %f27;

$L__BB278_23:
mov.u32 %r1254, %tid.y;
setp.ne.s32 %p319, %r1254, 0;
bar.sync 0;
@%p319 bra $L__BB278_78;
bra.uni $L__BB278_24;

$L__BB278_78:
ret;

$L__BB278_24:
ld.param.f32 %f977, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
ld.param.u64 %rd88, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r1257, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r1256, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
mov.u32 %r1255, %ctaid.z;
setp.ne.s64 %p21, %rd32, 0;
mov.u32 %r536, %ctaid.y;
mad.lo.s32 %r539, %r536, %r1257, %r4;
mad.lo.s32 %r540, %r1255, %r1256, %r539;
neg.ftz.f32 %f35, %f977;
shl.b32 %r542, %r2, 2;
mov.u32 %r543, _ZZ13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r24, %r543, %r542;
cvta.to.global.u64 %rd84, %rd88;
mul.wide.u32 %rd85, %r540, 4;
add.s64 %rd27, %rd84, %rd85;
@%p21 bra $L__BB278_33;
bra.uni $L__BB278_25;

$L__BB278_33:
ld.param.u32 %r1258, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
mov.u32 %r630, _ZZ13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r26, %r630, %r542;
setp.gt.s32 %p54, %r1258, 2;
@%p54 bra $L__BB278_44;

ld.param.u32 %r1261, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p57, %r1261, 1;
@%p57 bra $L__BB278_62;

ld.param.u32 %r1262, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p58, %r1262, 2;
@%p58 bra $L__BB278_36;
bra.uni $L__BB278_70;

$L__BB278_36:
ld.shared.f32 %f533, [%r24];
add.ftz.f32 %f534, %f533, %f30;
ld.shared.f32 %f535, [%r26];
add.ftz.f32 %f536, %f535, %f34;
ld.shared.f32 %f537, [%r24+512];
add.ftz.f32 %f538, %f537, %f534;
ld.shared.f32 %f539, [%r26+512];
add.ftz.f32 %f540, %f539, %f536;
ld.shared.f32 %f541, [%r24+1024];
add.ftz.f32 %f542, %f541, %f538;
ld.shared.f32 %f543, [%r26+1024];
add.ftz.f32 %f544, %f543, %f540;
mov.b32 %r877, %f542;
mov.u32 %r878, 31;
mov.u32 %r879, 16;
mov.u32 %r880, -1;
shfl.sync.bfly.b32 %r881|%p163, %r877, %r879, %r878, %r880;
mov.b32 %f545, %r881;
add.ftz.f32 %f546, %f542, %f545;
mov.b32 %r882, %f546;
mov.u32 %r883, 8;
shfl.sync.bfly.b32 %r884|%p164, %r882, %r883, %r878, %r880;
mov.b32 %f547, %r884;
add.ftz.f32 %f548, %f546, %f547;
mov.b32 %r885, %f548;
mov.u32 %r886, 4;
shfl.sync.bfly.b32 %r887|%p165, %r885, %r886, %r878, %r880;
mov.b32 %f549, %r887;
add.ftz.f32 %f550, %f548, %f549;
mov.b32 %r888, %f550;
mov.u32 %r889, 2;
shfl.sync.bfly.b32 %r890|%p166, %r888, %r889, %r878, %r880;
mov.b32 %f551, %r890;
add.ftz.f32 %f552, %f550, %f551;
mov.b32 %r891, %f552;
mov.u32 %r892, 1;
shfl.sync.bfly.b32 %r893|%p167, %r891, %r892, %r878, %r880;
mov.b32 %f553, %r893;
add.ftz.f32 %f56, %f552, %f553;
mov.b32 %r894, %f544;
shfl.sync.bfly.b32 %r895|%p168, %r894, %r879, %r878, %r880;
mov.b32 %f554, %r895;
add.ftz.f32 %f555, %f544, %f554;
mov.b32 %r896, %f555;
shfl.sync.bfly.b32 %r897|%p169, %r896, %r883, %r878, %r880;
mov.b32 %f556, %r897;
add.ftz.f32 %f557, %f555, %f556;
mov.b32 %r898, %f557;
shfl.sync.bfly.b32 %r899|%p170, %r898, %r886, %r878, %r880;
mov.b32 %f558, %r899;
add.ftz.f32 %f559, %f557, %f558;
mov.b32 %r900, %f559;
shfl.sync.bfly.b32 %r901|%p171, %r900, %r889, %r878, %r880;
mov.b32 %f560, %r901;
add.ftz.f32 %f561, %f559, %f560;
mov.b32 %r902, %f561;
shfl.sync.bfly.b32 %r903|%p172, %r902, %r892, %r878, %r880;
mov.b32 %f562, %r903;
add.ftz.f32 %f57, %f561, %f562;
setp.ne.s32 %p173, %r2, 0;
setp.ge.u32 %p174, %r4, %r43;
or.pred %p175, %p173, %p174;
@%p175 bra $L__BB278_38;

add.ftz.f32 %f563, %f5, %f56;
add.ftz.f32 %f564, %f4, %f57;
mul.ftz.f32 %f565, %f564, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f566, %f565;
add.ftz.f32 %f567, %f566, 0f3F800000;
div.approx.ftz.f32 %f568, %f564, %f567;
mul.ftz.f32 %f569, %f563, %f568;
st.global.f32 [%rd27], %f569;

$L__BB278_38:
ld.shared.f32 %f570, [%r24+128];
add.ftz.f32 %f571, %f570, %f29;
ld.shared.f32 %f572, [%r26+128];
add.ftz.f32 %f573, %f572, %f33;
ld.shared.f32 %f574, [%r24+640];
add.ftz.f32 %f575, %f574, %f571;
ld.shared.f32 %f576, [%r26+640];
add.ftz.f32 %f577, %f576, %f573;
ld.shared.f32 %f578, [%r24+1152];
add.ftz.f32 %f579, %f578, %f575;
ld.shared.f32 %f580, [%r26+1152];
add.ftz.f32 %f581, %f580, %f577;
mov.b32 %r907, %f579;
shfl.sync.bfly.b32 %r911|%p176, %r907, %r879, %r878, %r880;
mov.b32 %f582, %r911;
add.ftz.f32 %f583, %f579, %f582;
mov.b32 %r912, %f583;
shfl.sync.bfly.b32 %r914|%p177, %r912, %r883, %r878, %r880;
mov.b32 %f584, %r914;
add.ftz.f32 %f585, %f583, %f584;
mov.b32 %r915, %f585;
shfl.sync.bfly.b32 %r917|%p178, %r915, %r886, %r878, %r880;
mov.b32 %f586, %r917;
add.ftz.f32 %f587, %f585, %f586;
mov.b32 %r918, %f587;
shfl.sync.bfly.b32 %r920|%p179, %r918, %r889, %r878, %r880;
mov.b32 %f588, %r920;
add.ftz.f32 %f589, %f587, %f588;
mov.b32 %r921, %f589;
shfl.sync.bfly.b32 %r923|%p180, %r921, %r892, %r878, %r880;
mov.b32 %f590, %r923;
add.ftz.f32 %f58, %f589, %f590;
mov.b32 %r924, %f581;
shfl.sync.bfly.b32 %r925|%p181, %r924, %r879, %r878, %r880;
mov.b32 %f591, %r925;
add.ftz.f32 %f592, %f581, %f591;
mov.b32 %r926, %f592;
shfl.sync.bfly.b32 %r927|%p182, %r926, %r883, %r878, %r880;
mov.b32 %f593, %r927;
add.ftz.f32 %f594, %f592, %f593;
mov.b32 %r928, %f594;
shfl.sync.bfly.b32 %r929|%p183, %r928, %r886, %r878, %r880;
mov.b32 %f595, %r929;
add.ftz.f32 %f596, %f594, %f595;
mov.b32 %r930, %f596;
shfl.sync.bfly.b32 %r931|%p184, %r930, %r889, %r878, %r880;
mov.b32 %f597, %r931;
add.ftz.f32 %f598, %f596, %f597;
mov.b32 %r932, %f598;
shfl.sync.bfly.b32 %r933|%p185, %r932, %r892, %r878, %r880;
mov.b32 %f599, %r933;
add.ftz.f32 %f59, %f598, %f599;
setp.ne.s32 %p186, %r2, 1;
or.b32 %r937, %r4, 1;
setp.ge.u32 %p187, %r937, %r43;
or.pred %p188, %p186, %p187;
@%p188 bra $L__BB278_40;

add.ftz.f32 %f600, %f5, %f58;
add.ftz.f32 %f601, %f4, %f59;
mul.ftz.f32 %f602, %f601, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f603, %f602;
add.ftz.f32 %f604, %f603, 0f3F800000;
div.approx.ftz.f32 %f605, %f601, %f604;
mul.ftz.f32 %f606, %f600, %f605;
st.global.f32 [%rd27+4], %f606;

$L__BB278_40:
ld.shared.f32 %f607, [%r24+256];
add.ftz.f32 %f608, %f607, %f28;
ld.shared.f32 %f609, [%r26+256];
add.ftz.f32 %f610, %f609, %f32;
ld.shared.f32 %f611, [%r24+768];
add.ftz.f32 %f612, %f611, %f608;
ld.shared.f32 %f613, [%r26+768];
add.ftz.f32 %f614, %f613, %f610;
ld.shared.f32 %f615, [%r24+1280];
add.ftz.f32 %f616, %f615, %f612;
ld.shared.f32 %f617, [%r26+1280];
add.ftz.f32 %f618, %f617, %f614;
mov.b32 %r938, %f616;
mov.u32 %r939, 31;
mov.u32 %r940, 16;
mov.u32 %r941, -1;
shfl.sync.bfly.b32 %r942|%p189, %r938, %r940, %r939, %r941;
mov.b32 %f619, %r942;
add.ftz.f32 %f620, %f616, %f619;
mov.b32 %r943, %f620;
mov.u32 %r944, 8;
shfl.sync.bfly.b32 %r945|%p190, %r943, %r944, %r939, %r941;
mov.b32 %f621, %r945;
add.ftz.f32 %f622, %f620, %f621;
mov.b32 %r946, %f622;
mov.u32 %r947, 4;
shfl.sync.bfly.b32 %r948|%p191, %r946, %r947, %r939, %r941;
mov.b32 %f623, %r948;
add.ftz.f32 %f624, %f622, %f623;
mov.b32 %r949, %f624;
mov.u32 %r950, 2;
shfl.sync.bfly.b32 %r951|%p192, %r949, %r950, %r939, %r941;
mov.b32 %f625, %r951;
add.ftz.f32 %f626, %f624, %f625;
mov.b32 %r952, %f626;
mov.u32 %r953, 1;
shfl.sync.bfly.b32 %r954|%p193, %r952, %r953, %r939, %r941;
mov.b32 %f627, %r954;
add.ftz.f32 %f60, %f626, %f627;
mov.b32 %r955, %f618;
shfl.sync.bfly.b32 %r956|%p194, %r955, %r940, %r939, %r941;
mov.b32 %f628, %r956;
add.ftz.f32 %f629, %f618, %f628;
mov.b32 %r957, %f629;
shfl.sync.bfly.b32 %r958|%p195, %r957, %r944, %r939, %r941;
mov.b32 %f630, %r958;
add.ftz.f32 %f631, %f629, %f630;
mov.b32 %r959, %f631;
shfl.sync.bfly.b32 %r960|%p196, %r959, %r947, %r939, %r941;
mov.b32 %f632, %r960;
add.ftz.f32 %f633, %f631, %f632;
mov.b32 %r961, %f633;
shfl.sync.bfly.b32 %r962|%p197, %r961, %r950, %r939, %r941;
mov.b32 %f634, %r962;
add.ftz.f32 %f635, %f633, %f634;
mov.b32 %r963, %f635;
shfl.sync.bfly.b32 %r964|%p198, %r963, %r953, %r939, %r941;
mov.b32 %f636, %r964;
add.ftz.f32 %f61, %f635, %f636;
setp.ne.s32 %p199, %r2, 2;
or.b32 %r968, %r4, 2;
setp.ge.u32 %p200, %r968, %r43;
or.pred %p201, %p199, %p200;
@%p201 bra $L__BB278_42;

add.ftz.f32 %f637, %f5, %f60;
add.ftz.f32 %f638, %f4, %f61;
mul.ftz.f32 %f639, %f638, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f640, %f639;
add.ftz.f32 %f641, %f640, 0f3F800000;
div.approx.ftz.f32 %f642, %f638, %f641;
mul.ftz.f32 %f643, %f637, %f642;
st.global.f32 [%rd27+8], %f643;

$L__BB278_42:
ld.shared.f32 %f644, [%r24+384];
add.ftz.f32 %f645, %f644, %f27;
ld.shared.f32 %f646, [%r26+384];
add.ftz.f32 %f647, %f646, %f31;
ld.shared.f32 %f648, [%r24+896];
add.ftz.f32 %f649, %f648, %f645;
ld.shared.f32 %f650, [%r26+896];
add.ftz.f32 %f651, %f650, %f647;
ld.shared.f32 %f652, [%r24+1408];
add.ftz.f32 %f653, %f652, %f649;
ld.shared.f32 %f654, [%r26+1408];
add.ftz.f32 %f655, %f654, %f651;
mov.b32 %r969, %f653;
shfl.sync.bfly.b32 %r973|%p202, %r969, %r940, %r939, %r941;
mov.b32 %f656, %r973;
add.ftz.f32 %f657, %f653, %f656;
mov.b32 %r974, %f657;
shfl.sync.bfly.b32 %r976|%p203, %r974, %r944, %r939, %r941;
mov.b32 %f658, %r976;
add.ftz.f32 %f659, %f657, %f658;
mov.b32 %r977, %f659;
shfl.sync.bfly.b32 %r979|%p204, %r977, %r947, %r939, %r941;
mov.b32 %f660, %r979;
add.ftz.f32 %f661, %f659, %f660;
mov.b32 %r980, %f661;
shfl.sync.bfly.b32 %r982|%p205, %r980, %r950, %r939, %r941;
mov.b32 %f662, %r982;
add.ftz.f32 %f663, %f661, %f662;
mov.b32 %r983, %f663;
shfl.sync.bfly.b32 %r985|%p206, %r983, %r953, %r939, %r941;
mov.b32 %f664, %r985;
add.ftz.f32 %f62, %f663, %f664;
mov.b32 %r986, %f655;
shfl.sync.bfly.b32 %r987|%p207, %r986, %r940, %r939, %r941;
mov.b32 %f665, %r987;
add.ftz.f32 %f666, %f655, %f665;
mov.b32 %r988, %f666;
shfl.sync.bfly.b32 %r989|%p208, %r988, %r944, %r939, %r941;
mov.b32 %f667, %r989;
add.ftz.f32 %f668, %f666, %f667;
mov.b32 %r990, %f668;
shfl.sync.bfly.b32 %r991|%p209, %r990, %r947, %r939, %r941;
mov.b32 %f669, %r991;
add.ftz.f32 %f670, %f668, %f669;
mov.b32 %r992, %f670;
shfl.sync.bfly.b32 %r993|%p210, %r992, %r950, %r939, %r941;
mov.b32 %f671, %r993;
add.ftz.f32 %f672, %f670, %f671;
mov.b32 %r994, %f672;
shfl.sync.bfly.b32 %r995|%p211, %r994, %r953, %r939, %r941;
mov.b32 %f673, %r995;
add.ftz.f32 %f63, %f672, %f673;
setp.ne.s32 %p212, %r2, 3;
or.b32 %r999, %r4, 3;
setp.ge.u32 %p213, %r999, %r43;
or.pred %p214, %p212, %p213;
@%p214 bra $L__BB278_78;

add.ftz.f32 %f674, %f5, %f62;
add.ftz.f32 %f675, %f4, %f63;
mul.ftz.f32 %f676, %f675, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f677, %f676;
add.ftz.f32 %f678, %f677, 0f3F800000;
div.approx.ftz.f32 %f679, %f675, %f678;
mul.ftz.f32 %f680, %f674, %f679;
st.global.f32 [%rd27+12], %f680;
bra.uni $L__BB278_78;

$L__BB278_25:
ld.shared.f32 %f133, [%r24];
add.ftz.f32 %f134, %f133, %f30;
ld.shared.f32 %f135, [%r24+512];
add.ftz.f32 %f136, %f135, %f134;
ld.shared.f32 %f137, [%r24+1024];
add.ftz.f32 %f138, %f137, %f136;
mov.b32 %r544, %f138;
mov.u32 %r545, 31;
mov.u32 %r546, 16;
mov.u32 %r547, -1;
shfl.sync.bfly.b32 %r548|%p22, %r544, %r546, %r545, %r547;
mov.b32 %f139, %r548;
add.ftz.f32 %f140, %f138, %f139;
mov.b32 %r549, %f140;
mov.u32 %r550, 8;
shfl.sync.bfly.b32 %r551|%p23, %r549, %r550, %r545, %r547;
mov.b32 %f141, %r551;
add.ftz.f32 %f142, %f140, %f141;
mov.b32 %r552, %f142;
mov.u32 %r553, 4;
shfl.sync.bfly.b32 %r554|%p24, %r552, %r553, %r545, %r547;
mov.b32 %f143, %r554;
add.ftz.f32 %f144, %f142, %f143;
mov.b32 %r555, %f144;
mov.u32 %r556, 2;
shfl.sync.bfly.b32 %r557|%p25, %r555, %r556, %r545, %r547;
mov.b32 %f145, %r557;
add.ftz.f32 %f146, %f144, %f145;
mov.b32 %r558, %f146;
mov.u32 %r559, 1;
shfl.sync.bfly.b32 %r560|%p26, %r558, %r559, %r545, %r547;
mov.b32 %f147, %r560;
add.ftz.f32 %f36, %f146, %f147;
setp.ne.s32 %p27, %r2, 0;
setp.ge.u32 %p28, %r4, %r43;
or.pred %p29, %p27, %p28;
@%p29 bra $L__BB278_27;

add.ftz.f32 %f148, %f5, %f36;
st.global.f32 [%rd27], %f148;

$L__BB278_27:
ld.shared.f32 %f149, [%r24+128];
add.ftz.f32 %f150, %f149, %f29;
ld.shared.f32 %f151, [%r24+640];
add.ftz.f32 %f152, %f151, %f150;
ld.shared.f32 %f153, [%r24+1152];
add.ftz.f32 %f154, %f153, %f152;
mov.b32 %r564, %f154;
shfl.sync.bfly.b32 %r568|%p30, %r564, %r546, %r545, %r547;
mov.b32 %f155, %r568;
add.ftz.f32 %f156, %f154, %f155;
mov.b32 %r569, %f156;
shfl.sync.bfly.b32 %r571|%p31, %r569, %r550, %r545, %r547;
mov.b32 %f157, %r571;
add.ftz.f32 %f158, %f156, %f157;
mov.b32 %r572, %f158;
shfl.sync.bfly.b32 %r574|%p32, %r572, %r553, %r545, %r547;
mov.b32 %f159, %r574;
add.ftz.f32 %f160, %f158, %f159;
mov.b32 %r575, %f160;
shfl.sync.bfly.b32 %r577|%p33, %r575, %r556, %r545, %r547;
mov.b32 %f161, %r577;
add.ftz.f32 %f162, %f160, %f161;
mov.b32 %r578, %f162;
shfl.sync.bfly.b32 %r580|%p34, %r578, %r559, %r545, %r547;
mov.b32 %f163, %r580;
add.ftz.f32 %f37, %f162, %f163;
setp.ne.s32 %p35, %r2, 1;
or.b32 %r584, %r4, 1;
setp.ge.u32 %p36, %r584, %r43;
or.pred %p37, %p35, %p36;
@%p37 bra $L__BB278_29;

add.ftz.f32 %f164, %f5, %f37;
st.global.f32 [%rd27+4], %f164;

$L__BB278_29:
ld.shared.f32 %f165, [%r24+256];
add.ftz.f32 %f166, %f165, %f28;
ld.shared.f32 %f167, [%r24+768];
add.ftz.f32 %f168, %f167, %f166;
ld.shared.f32 %f169, [%r24+1280];
add.ftz.f32 %f170, %f169, %f168;
mov.b32 %r585, %f170;
mov.u32 %r586, 31;
mov.u32 %r587, 16;
mov.u32 %r588, -1;
shfl.sync.bfly.b32 %r589|%p38, %r585, %r587, %r586, %r588;
mov.b32 %f171, %r589;
add.ftz.f32 %f172, %f170, %f171;
mov.b32 %r590, %f172;
mov.u32 %r591, 8;
shfl.sync.bfly.b32 %r592|%p39, %r590, %r591, %r586, %r588;
mov.b32 %f173, %r592;
add.ftz.f32 %f174, %f172, %f173;
mov.b32 %r593, %f174;
mov.u32 %r594, 4;
shfl.sync.bfly.b32 %r595|%p40, %r593, %r594, %r586, %r588;
mov.b32 %f175, %r595;
add.ftz.f32 %f176, %f174, %f175;
mov.b32 %r596, %f176;
mov.u32 %r597, 2;
shfl.sync.bfly.b32 %r598|%p41, %r596, %r597, %r586, %r588;
mov.b32 %f177, %r598;
add.ftz.f32 %f178, %f176, %f177;
mov.b32 %r599, %f178;
mov.u32 %r600, 1;
shfl.sync.bfly.b32 %r601|%p42, %r599, %r600, %r586, %r588;
mov.b32 %f179, %r601;
add.ftz.f32 %f38, %f178, %f179;
setp.ne.s32 %p43, %r2, 2;
or.b32 %r605, %r4, 2;
setp.ge.u32 %p44, %r605, %r43;
or.pred %p45, %p43, %p44;
@%p45 bra $L__BB278_31;

add.ftz.f32 %f180, %f5, %f38;
st.global.f32 [%rd27+8], %f180;

$L__BB278_31:
ld.shared.f32 %f181, [%r24+384];
add.ftz.f32 %f182, %f181, %f27;
ld.shared.f32 %f183, [%r24+896];
add.ftz.f32 %f184, %f183, %f182;
ld.shared.f32 %f185, [%r24+1408];
add.ftz.f32 %f186, %f185, %f184;
mov.b32 %r606, %f186;
shfl.sync.bfly.b32 %r610|%p46, %r606, %r587, %r586, %r588;
mov.b32 %f187, %r610;
add.ftz.f32 %f188, %f186, %f187;
mov.b32 %r611, %f188;
shfl.sync.bfly.b32 %r613|%p47, %r611, %r591, %r586, %r588;
mov.b32 %f189, %r613;
add.ftz.f32 %f190, %f188, %f189;
mov.b32 %r614, %f190;
shfl.sync.bfly.b32 %r616|%p48, %r614, %r594, %r586, %r588;
mov.b32 %f191, %r616;
add.ftz.f32 %f192, %f190, %f191;
mov.b32 %r617, %f192;
shfl.sync.bfly.b32 %r619|%p49, %r617, %r597, %r586, %r588;
mov.b32 %f193, %r619;
add.ftz.f32 %f194, %f192, %f193;
mov.b32 %r620, %f194;
shfl.sync.bfly.b32 %r622|%p50, %r620, %r600, %r586, %r588;
mov.b32 %f195, %r622;
add.ftz.f32 %f39, %f194, %f195;
setp.ne.s32 %p51, %r2, 3;
or.b32 %r626, %r4, 3;
setp.ge.u32 %p52, %r626, %r43;
or.pred %p53, %p51, %p52;
@%p53 bra $L__BB278_78;

add.ftz.f32 %f196, %f5, %f39;
st.global.f32 [%rd27+12], %f196;
bra.uni $L__BB278_78;

$L__BB278_44:
ld.param.u32 %r1259, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p55, %r1259, 3;
@%p55 bra $L__BB278_54;

ld.param.u32 %r1260, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.ne.s32 %p56, %r1260, 6;
@%p56 bra $L__BB278_70;

ld.shared.f32 %f197, [%r24];
add.ftz.f32 %f198, %f197, %f30;
ld.shared.f32 %f199, [%r26];
add.ftz.f32 %f200, %f199, %f34;
ld.shared.f32 %f201, [%r24+512];
add.ftz.f32 %f202, %f201, %f198;
ld.shared.f32 %f203, [%r26+512];
add.ftz.f32 %f204, %f203, %f200;
ld.shared.f32 %f205, [%r24+1024];
add.ftz.f32 %f206, %f205, %f202;
ld.shared.f32 %f207, [%r26+1024];
add.ftz.f32 %f208, %f207, %f204;
mov.b32 %r631, %f206;
mov.u32 %r632, 31;
mov.u32 %r633, 16;
mov.u32 %r634, -1;
shfl.sync.bfly.b32 %r635|%p59, %r631, %r633, %r632, %r634;
mov.b32 %f209, %r635;
add.ftz.f32 %f210, %f206, %f209;
mov.b32 %r636, %f210;
mov.u32 %r637, 8;
shfl.sync.bfly.b32 %r638|%p60, %r636, %r637, %r632, %r634;
mov.b32 %f211, %r638;
add.ftz.f32 %f212, %f210, %f211;
mov.b32 %r639, %f212;
mov.u32 %r640, 4;
shfl.sync.bfly.b32 %r641|%p61, %r639, %r640, %r632, %r634;
mov.b32 %f213, %r641;
add.ftz.f32 %f214, %f212, %f213;
mov.b32 %r642, %f214;
mov.u32 %r643, 2;
shfl.sync.bfly.b32 %r644|%p62, %r642, %r643, %r632, %r634;
mov.b32 %f215, %r644;
add.ftz.f32 %f216, %f214, %f215;
mov.b32 %r645, %f216;
mov.u32 %r646, 1;
shfl.sync.bfly.b32 %r647|%p63, %r645, %r646, %r632, %r634;
mov.b32 %f217, %r647;
add.ftz.f32 %f40, %f216, %f217;
mov.b32 %r648, %f208;
shfl.sync.bfly.b32 %r649|%p64, %r648, %r633, %r632, %r634;
mov.b32 %f218, %r649;
add.ftz.f32 %f219, %f208, %f218;
mov.b32 %r650, %f219;
shfl.sync.bfly.b32 %r651|%p65, %r650, %r637, %r632, %r634;
mov.b32 %f220, %r651;
add.ftz.f32 %f221, %f219, %f220;
mov.b32 %r652, %f221;
shfl.sync.bfly.b32 %r653|%p66, %r652, %r640, %r632, %r634;
mov.b32 %f222, %r653;
add.ftz.f32 %f223, %f221, %f222;
mov.b32 %r654, %f223;
shfl.sync.bfly.b32 %r655|%p67, %r654, %r643, %r632, %r634;
mov.b32 %f224, %r655;
add.ftz.f32 %f225, %f223, %f224;
mov.b32 %r656, %f225;
shfl.sync.bfly.b32 %r657|%p68, %r656, %r646, %r632, %r634;
mov.b32 %f226, %r657;
add.ftz.f32 %f41, %f225, %f226;
setp.ne.s32 %p69, %r2, 0;
setp.ge.u32 %p70, %r4, %r43;
or.pred %p71, %p69, %p70;
@%p71 bra $L__BB278_48;

ld.param.f32 %f981, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f227, %f5, %f40;
add.ftz.f32 %f228, %f4, %f41;
min.ftz.f32 %f229, %f228, %f981;
min.ftz.f32 %f230, %f227, %f981;
max.ftz.f32 %f231, %f230, %f35;
mul.ftz.f32 %f232, %f229, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f233, %f232;
add.ftz.f32 %f234, %f233, 0f3F800000;
div.approx.ftz.f32 %f235, %f229, %f234;
mul.ftz.f32 %f236, %f231, %f235;
st.global.f32 [%rd27], %f236;

$L__BB278_48:
ld.shared.f32 %f237, [%r24+128];
add.ftz.f32 %f238, %f237, %f29;
ld.shared.f32 %f239, [%r26+128];
add.ftz.f32 %f240, %f239, %f33;
ld.shared.f32 %f241, [%r24+640];
add.ftz.f32 %f242, %f241, %f238;
ld.shared.f32 %f243, [%r26+640];
add.ftz.f32 %f244, %f243, %f240;
ld.shared.f32 %f245, [%r24+1152];
add.ftz.f32 %f246, %f245, %f242;
ld.shared.f32 %f247, [%r26+1152];
add.ftz.f32 %f248, %f247, %f244;
mov.b32 %r661, %f246;
shfl.sync.bfly.b32 %r665|%p72, %r661, %r633, %r632, %r634;
mov.b32 %f249, %r665;
add.ftz.f32 %f250, %f246, %f249;
mov.b32 %r666, %f250;
shfl.sync.bfly.b32 %r668|%p73, %r666, %r637, %r632, %r634;
mov.b32 %f251, %r668;
add.ftz.f32 %f252, %f250, %f251;
mov.b32 %r669, %f252;
shfl.sync.bfly.b32 %r671|%p74, %r669, %r640, %r632, %r634;
mov.b32 %f253, %r671;
add.ftz.f32 %f254, %f252, %f253;
mov.b32 %r672, %f254;
shfl.sync.bfly.b32 %r674|%p75, %r672, %r643, %r632, %r634;
mov.b32 %f255, %r674;
add.ftz.f32 %f256, %f254, %f255;
mov.b32 %r675, %f256;
shfl.sync.bfly.b32 %r677|%p76, %r675, %r646, %r632, %r634;
mov.b32 %f257, %r677;
add.ftz.f32 %f42, %f256, %f257;
mov.b32 %r678, %f248;
shfl.sync.bfly.b32 %r679|%p77, %r678, %r633, %r632, %r634;
mov.b32 %f258, %r679;
add.ftz.f32 %f259, %f248, %f258;
mov.b32 %r680, %f259;
shfl.sync.bfly.b32 %r681|%p78, %r680, %r637, %r632, %r634;
mov.b32 %f260, %r681;
add.ftz.f32 %f261, %f259, %f260;
mov.b32 %r682, %f261;
shfl.sync.bfly.b32 %r683|%p79, %r682, %r640, %r632, %r634;
mov.b32 %f262, %r683;
add.ftz.f32 %f263, %f261, %f262;
mov.b32 %r684, %f263;
shfl.sync.bfly.b32 %r685|%p80, %r684, %r643, %r632, %r634;
mov.b32 %f264, %r685;
add.ftz.f32 %f265, %f263, %f264;
mov.b32 %r686, %f265;
shfl.sync.bfly.b32 %r687|%p81, %r686, %r646, %r632, %r634;
mov.b32 %f266, %r687;
add.ftz.f32 %f43, %f265, %f266;
setp.ne.s32 %p82, %r2, 1;
or.b32 %r691, %r4, 1;
setp.ge.u32 %p83, %r691, %r43;
or.pred %p84, %p82, %p83;
@%p84 bra $L__BB278_50;

ld.param.f32 %f980, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f267, %f5, %f42;
add.ftz.f32 %f268, %f4, %f43;
min.ftz.f32 %f269, %f268, %f980;
min.ftz.f32 %f270, %f267, %f980;
max.ftz.f32 %f271, %f270, %f35;
mul.ftz.f32 %f272, %f269, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f273, %f272;
add.ftz.f32 %f274, %f273, 0f3F800000;
div.approx.ftz.f32 %f275, %f269, %f274;
mul.ftz.f32 %f276, %f271, %f275;
st.global.f32 [%rd27+4], %f276;

$L__BB278_50:
ld.shared.f32 %f277, [%r24+256];
add.ftz.f32 %f278, %f277, %f28;
ld.shared.f32 %f279, [%r26+256];
add.ftz.f32 %f280, %f279, %f32;
ld.shared.f32 %f281, [%r24+768];
add.ftz.f32 %f282, %f281, %f278;
ld.shared.f32 %f283, [%r26+768];
add.ftz.f32 %f284, %f283, %f280;
ld.shared.f32 %f285, [%r24+1280];
add.ftz.f32 %f286, %f285, %f282;
ld.shared.f32 %f287, [%r26+1280];
add.ftz.f32 %f288, %f287, %f284;
mov.b32 %r692, %f286;
mov.u32 %r693, 31;
mov.u32 %r694, 16;
mov.u32 %r695, -1;
shfl.sync.bfly.b32 %r696|%p85, %r692, %r694, %r693, %r695;
mov.b32 %f289, %r696;
add.ftz.f32 %f290, %f286, %f289;
mov.b32 %r697, %f290;
mov.u32 %r698, 8;
shfl.sync.bfly.b32 %r699|%p86, %r697, %r698, %r693, %r695;
mov.b32 %f291, %r699;
add.ftz.f32 %f292, %f290, %f291;
mov.b32 %r700, %f292;
mov.u32 %r701, 4;
shfl.sync.bfly.b32 %r702|%p87, %r700, %r701, %r693, %r695;
mov.b32 %f293, %r702;
add.ftz.f32 %f294, %f292, %f293;
mov.b32 %r703, %f294;
mov.u32 %r704, 2;
shfl.sync.bfly.b32 %r705|%p88, %r703, %r704, %r693, %r695;
mov.b32 %f295, %r705;
add.ftz.f32 %f296, %f294, %f295;
mov.b32 %r706, %f296;
mov.u32 %r707, 1;
shfl.sync.bfly.b32 %r708|%p89, %r706, %r707, %r693, %r695;
mov.b32 %f297, %r708;
add.ftz.f32 %f44, %f296, %f297;
mov.b32 %r709, %f288;
shfl.sync.bfly.b32 %r710|%p90, %r709, %r694, %r693, %r695;
mov.b32 %f298, %r710;
add.ftz.f32 %f299, %f288, %f298;
mov.b32 %r711, %f299;
shfl.sync.bfly.b32 %r712|%p91, %r711, %r698, %r693, %r695;
mov.b32 %f300, %r712;
add.ftz.f32 %f301, %f299, %f300;
mov.b32 %r713, %f301;
shfl.sync.bfly.b32 %r714|%p92, %r713, %r701, %r693, %r695;
mov.b32 %f302, %r714;
add.ftz.f32 %f303, %f301, %f302;
mov.b32 %r715, %f303;
shfl.sync.bfly.b32 %r716|%p93, %r715, %r704, %r693, %r695;
mov.b32 %f304, %r716;
add.ftz.f32 %f305, %f303, %f304;
mov.b32 %r717, %f305;
shfl.sync.bfly.b32 %r718|%p94, %r717, %r707, %r693, %r695;
mov.b32 %f306, %r718;
add.ftz.f32 %f45, %f305, %f306;
setp.ne.s32 %p95, %r2, 2;
or.b32 %r722, %r4, 2;
setp.ge.u32 %p96, %r722, %r43;
or.pred %p97, %p95, %p96;
@%p97 bra $L__BB278_52;

ld.param.f32 %f979, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f307, %f5, %f44;
add.ftz.f32 %f308, %f4, %f45;
min.ftz.f32 %f309, %f308, %f979;
min.ftz.f32 %f310, %f307, %f979;
max.ftz.f32 %f311, %f310, %f35;
mul.ftz.f32 %f312, %f309, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f313, %f312;
add.ftz.f32 %f314, %f313, 0f3F800000;
div.approx.ftz.f32 %f315, %f309, %f314;
mul.ftz.f32 %f316, %f311, %f315;
st.global.f32 [%rd27+8], %f316;

$L__BB278_52:
ld.shared.f32 %f317, [%r24+384];
add.ftz.f32 %f318, %f317, %f27;
ld.shared.f32 %f319, [%r26+384];
add.ftz.f32 %f320, %f319, %f31;
ld.shared.f32 %f321, [%r24+896];
add.ftz.f32 %f322, %f321, %f318;
ld.shared.f32 %f323, [%r26+896];
add.ftz.f32 %f324, %f323, %f320;
ld.shared.f32 %f325, [%r24+1408];
add.ftz.f32 %f326, %f325, %f322;
ld.shared.f32 %f327, [%r26+1408];
add.ftz.f32 %f328, %f327, %f324;
mov.b32 %r723, %f326;
shfl.sync.bfly.b32 %r727|%p98, %r723, %r694, %r693, %r695;
mov.b32 %f329, %r727;
add.ftz.f32 %f330, %f326, %f329;
mov.b32 %r728, %f330;
shfl.sync.bfly.b32 %r730|%p99, %r728, %r698, %r693, %r695;
mov.b32 %f331, %r730;
add.ftz.f32 %f332, %f330, %f331;
mov.b32 %r731, %f332;
shfl.sync.bfly.b32 %r733|%p100, %r731, %r701, %r693, %r695;
mov.b32 %f333, %r733;
add.ftz.f32 %f334, %f332, %f333;
mov.b32 %r734, %f334;
shfl.sync.bfly.b32 %r736|%p101, %r734, %r704, %r693, %r695;
mov.b32 %f335, %r736;
add.ftz.f32 %f336, %f334, %f335;
mov.b32 %r737, %f336;
shfl.sync.bfly.b32 %r739|%p102, %r737, %r707, %r693, %r695;
mov.b32 %f337, %r739;
add.ftz.f32 %f46, %f336, %f337;
mov.b32 %r740, %f328;
shfl.sync.bfly.b32 %r741|%p103, %r740, %r694, %r693, %r695;
mov.b32 %f338, %r741;
add.ftz.f32 %f339, %f328, %f338;
mov.b32 %r742, %f339;
shfl.sync.bfly.b32 %r743|%p104, %r742, %r698, %r693, %r695;
mov.b32 %f340, %r743;
add.ftz.f32 %f341, %f339, %f340;
mov.b32 %r744, %f341;
shfl.sync.bfly.b32 %r745|%p105, %r744, %r701, %r693, %r695;
mov.b32 %f342, %r745;
add.ftz.f32 %f343, %f341, %f342;
mov.b32 %r746, %f343;
shfl.sync.bfly.b32 %r747|%p106, %r746, %r704, %r693, %r695;
mov.b32 %f344, %r747;
add.ftz.f32 %f345, %f343, %f344;
mov.b32 %r748, %f345;
shfl.sync.bfly.b32 %r749|%p107, %r748, %r707, %r693, %r695;
mov.b32 %f346, %r749;
add.ftz.f32 %f47, %f345, %f346;
setp.ne.s32 %p108, %r2, 3;
or.b32 %r753, %r4, 3;
setp.ge.u32 %p109, %r753, %r43;
or.pred %p110, %p108, %p109;
@%p110 bra $L__BB278_78;

ld.param.f32 %f978, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f347, %f5, %f46;
add.ftz.f32 %f348, %f4, %f47;
min.ftz.f32 %f349, %f348, %f978;
min.ftz.f32 %f350, %f347, %f978;
max.ftz.f32 %f351, %f350, %f35;
mul.ftz.f32 %f352, %f349, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f353, %f352;
add.ftz.f32 %f354, %f353, 0f3F800000;
div.approx.ftz.f32 %f355, %f349, %f354;
mul.ftz.f32 %f356, %f351, %f355;
st.global.f32 [%rd27+12], %f356;
bra.uni $L__BB278_78;

$L__BB278_70:
ld.shared.f32 %f845, [%r24];
add.ftz.f32 %f846, %f845, %f30;
ld.shared.f32 %f847, [%r26];
add.ftz.f32 %f848, %f847, %f34;
ld.shared.f32 %f849, [%r24+512];
add.ftz.f32 %f850, %f849, %f846;
ld.shared.f32 %f851, [%r26+512];
add.ftz.f32 %f852, %f851, %f848;
ld.shared.f32 %f853, [%r24+1024];
add.ftz.f32 %f854, %f853, %f850;
ld.shared.f32 %f855, [%r26+1024];
add.ftz.f32 %f856, %f855, %f852;
mov.b32 %r1123, %f854;
mov.u32 %r1124, 31;
mov.u32 %r1125, 16;
mov.u32 %r1126, -1;
shfl.sync.bfly.b32 %r1127|%p267, %r1123, %r1125, %r1124, %r1126;
mov.b32 %f857, %r1127;
add.ftz.f32 %f858, %f854, %f857;
mov.b32 %r1128, %f858;
mov.u32 %r1129, 8;
shfl.sync.bfly.b32 %r1130|%p268, %r1128, %r1129, %r1124, %r1126;
mov.b32 %f859, %r1130;
add.ftz.f32 %f860, %f858, %f859;
mov.b32 %r1131, %f860;
mov.u32 %r1132, 4;
shfl.sync.bfly.b32 %r1133|%p269, %r1131, %r1132, %r1124, %r1126;
mov.b32 %f861, %r1133;
add.ftz.f32 %f862, %f860, %f861;
mov.b32 %r1134, %f862;
mov.u32 %r1135, 2;
shfl.sync.bfly.b32 %r1136|%p270, %r1134, %r1135, %r1124, %r1126;
mov.b32 %f863, %r1136;
add.ftz.f32 %f864, %f862, %f863;
mov.b32 %r1137, %f864;
mov.u32 %r1138, 1;
shfl.sync.bfly.b32 %r1139|%p271, %r1137, %r1138, %r1124, %r1126;
mov.b32 %f865, %r1139;
add.ftz.f32 %f72, %f864, %f865;
mov.b32 %r1140, %f856;
shfl.sync.bfly.b32 %r1141|%p272, %r1140, %r1125, %r1124, %r1126;
mov.b32 %f866, %r1141;
add.ftz.f32 %f867, %f856, %f866;
mov.b32 %r1142, %f867;
shfl.sync.bfly.b32 %r1143|%p273, %r1142, %r1129, %r1124, %r1126;
mov.b32 %f868, %r1143;
add.ftz.f32 %f869, %f867, %f868;
mov.b32 %r1144, %f869;
shfl.sync.bfly.b32 %r1145|%p274, %r1144, %r1132, %r1124, %r1126;
mov.b32 %f870, %r1145;
add.ftz.f32 %f871, %f869, %f870;
mov.b32 %r1146, %f871;
shfl.sync.bfly.b32 %r1147|%p275, %r1146, %r1135, %r1124, %r1126;
mov.b32 %f872, %r1147;
add.ftz.f32 %f873, %f871, %f872;
mov.b32 %r1148, %f873;
shfl.sync.bfly.b32 %r1149|%p276, %r1148, %r1138, %r1124, %r1126;
mov.b32 %f874, %r1149;
add.ftz.f32 %f73, %f873, %f874;
setp.ne.s32 %p277, %r2, 0;
setp.ge.u32 %p278, %r4, %r43;
or.pred %p279, %p277, %p278;
@%p279 bra $L__BB278_72;

add.ftz.f32 %f875, %f5, %f72;
add.ftz.f32 %f876, %f4, %f73;
mul.ftz.f32 %f877, %f875, %f876;
st.global.f32 [%rd27], %f877;

$L__BB278_72:
ld.shared.f32 %f878, [%r24+128];
add.ftz.f32 %f879, %f878, %f29;
ld.shared.f32 %f880, [%r26+128];
add.ftz.f32 %f881, %f880, %f33;
ld.shared.f32 %f882, [%r24+640];
add.ftz.f32 %f883, %f882, %f879;
ld.shared.f32 %f884, [%r26+640];
add.ftz.f32 %f885, %f884, %f881;
ld.shared.f32 %f886, [%r24+1152];
add.ftz.f32 %f887, %f886, %f883;
ld.shared.f32 %f888, [%r26+1152];
add.ftz.f32 %f889, %f888, %f885;
mov.b32 %r1153, %f887;
shfl.sync.bfly.b32 %r1157|%p280, %r1153, %r1125, %r1124, %r1126;
mov.b32 %f890, %r1157;
add.ftz.f32 %f891, %f887, %f890;
mov.b32 %r1158, %f891;
shfl.sync.bfly.b32 %r1160|%p281, %r1158, %r1129, %r1124, %r1126;
mov.b32 %f892, %r1160;
add.ftz.f32 %f893, %f891, %f892;
mov.b32 %r1161, %f893;
shfl.sync.bfly.b32 %r1163|%p282, %r1161, %r1132, %r1124, %r1126;
mov.b32 %f894, %r1163;
add.ftz.f32 %f895, %f893, %f894;
mov.b32 %r1164, %f895;
shfl.sync.bfly.b32 %r1166|%p283, %r1164, %r1135, %r1124, %r1126;
mov.b32 %f896, %r1166;
add.ftz.f32 %f897, %f895, %f896;
mov.b32 %r1167, %f897;
shfl.sync.bfly.b32 %r1169|%p284, %r1167, %r1138, %r1124, %r1126;
mov.b32 %f898, %r1169;
add.ftz.f32 %f74, %f897, %f898;
mov.b32 %r1170, %f889;
shfl.sync.bfly.b32 %r1171|%p285, %r1170, %r1125, %r1124, %r1126;
mov.b32 %f899, %r1171;
add.ftz.f32 %f900, %f889, %f899;
mov.b32 %r1172, %f900;
shfl.sync.bfly.b32 %r1173|%p286, %r1172, %r1129, %r1124, %r1126;
mov.b32 %f901, %r1173;
add.ftz.f32 %f902, %f900, %f901;
mov.b32 %r1174, %f902;
shfl.sync.bfly.b32 %r1175|%p287, %r1174, %r1132, %r1124, %r1126;
mov.b32 %f903, %r1175;
add.ftz.f32 %f904, %f902, %f903;
mov.b32 %r1176, %f904;
shfl.sync.bfly.b32 %r1177|%p288, %r1176, %r1135, %r1124, %r1126;
mov.b32 %f905, %r1177;
add.ftz.f32 %f906, %f904, %f905;
mov.b32 %r1178, %f906;
shfl.sync.bfly.b32 %r1179|%p289, %r1178, %r1138, %r1124, %r1126;
mov.b32 %f907, %r1179;
add.ftz.f32 %f75, %f906, %f907;
setp.ne.s32 %p290, %r2, 1;
or.b32 %r1183, %r4, 1;
setp.ge.u32 %p291, %r1183, %r43;
or.pred %p292, %p290, %p291;
@%p292 bra $L__BB278_74;

add.ftz.f32 %f908, %f5, %f74;
add.ftz.f32 %f909, %f4, %f75;
mul.ftz.f32 %f910, %f908, %f909;
st.global.f32 [%rd27+4], %f910;

$L__BB278_74:
ld.shared.f32 %f911, [%r24+256];
add.ftz.f32 %f912, %f911, %f28;
ld.shared.f32 %f913, [%r26+256];
add.ftz.f32 %f914, %f913, %f32;
ld.shared.f32 %f915, [%r24+768];
add.ftz.f32 %f916, %f915, %f912;
ld.shared.f32 %f917, [%r26+768];
add.ftz.f32 %f918, %f917, %f914;
ld.shared.f32 %f919, [%r24+1280];
add.ftz.f32 %f920, %f919, %f916;
ld.shared.f32 %f921, [%r26+1280];
add.ftz.f32 %f922, %f921, %f918;
mov.b32 %r1184, %f920;
mov.u32 %r1185, 31;
mov.u32 %r1186, 16;
mov.u32 %r1187, -1;
shfl.sync.bfly.b32 %r1188|%p293, %r1184, %r1186, %r1185, %r1187;
mov.b32 %f923, %r1188;
add.ftz.f32 %f924, %f920, %f923;
mov.b32 %r1189, %f924;
mov.u32 %r1190, 8;
shfl.sync.bfly.b32 %r1191|%p294, %r1189, %r1190, %r1185, %r1187;
mov.b32 %f925, %r1191;
add.ftz.f32 %f926, %f924, %f925;
mov.b32 %r1192, %f926;
mov.u32 %r1193, 4;
shfl.sync.bfly.b32 %r1194|%p295, %r1192, %r1193, %r1185, %r1187;
mov.b32 %f927, %r1194;
add.ftz.f32 %f928, %f926, %f927;
mov.b32 %r1195, %f928;
mov.u32 %r1196, 2;
shfl.sync.bfly.b32 %r1197|%p296, %r1195, %r1196, %r1185, %r1187;
mov.b32 %f929, %r1197;
add.ftz.f32 %f930, %f928, %f929;
mov.b32 %r1198, %f930;
mov.u32 %r1199, 1;
shfl.sync.bfly.b32 %r1200|%p297, %r1198, %r1199, %r1185, %r1187;
mov.b32 %f931, %r1200;
add.ftz.f32 %f76, %f930, %f931;
mov.b32 %r1201, %f922;
shfl.sync.bfly.b32 %r1202|%p298, %r1201, %r1186, %r1185, %r1187;
mov.b32 %f932, %r1202;
add.ftz.f32 %f933, %f922, %f932;
mov.b32 %r1203, %f933;
shfl.sync.bfly.b32 %r1204|%p299, %r1203, %r1190, %r1185, %r1187;
mov.b32 %f934, %r1204;
add.ftz.f32 %f935, %f933, %f934;
mov.b32 %r1205, %f935;
shfl.sync.bfly.b32 %r1206|%p300, %r1205, %r1193, %r1185, %r1187;
mov.b32 %f936, %r1206;
add.ftz.f32 %f937, %f935, %f936;
mov.b32 %r1207, %f937;
shfl.sync.bfly.b32 %r1208|%p301, %r1207, %r1196, %r1185, %r1187;
mov.b32 %f938, %r1208;
add.ftz.f32 %f939, %f937, %f938;
mov.b32 %r1209, %f939;
shfl.sync.bfly.b32 %r1210|%p302, %r1209, %r1199, %r1185, %r1187;
mov.b32 %f940, %r1210;
add.ftz.f32 %f77, %f939, %f940;
setp.ne.s32 %p303, %r2, 2;
or.b32 %r1214, %r4, 2;
setp.ge.u32 %p304, %r1214, %r43;
or.pred %p305, %p303, %p304;
@%p305 bra $L__BB278_76;

add.ftz.f32 %f941, %f5, %f76;
add.ftz.f32 %f942, %f4, %f77;
mul.ftz.f32 %f943, %f941, %f942;
st.global.f32 [%rd27+8], %f943;

$L__BB278_76:
ld.shared.f32 %f944, [%r24+384];
add.ftz.f32 %f945, %f944, %f27;
ld.shared.f32 %f946, [%r26+384];
add.ftz.f32 %f947, %f946, %f31;
ld.shared.f32 %f948, [%r24+896];
add.ftz.f32 %f949, %f948, %f945;
ld.shared.f32 %f950, [%r26+896];
add.ftz.f32 %f951, %f950, %f947;
ld.shared.f32 %f952, [%r24+1408];
add.ftz.f32 %f953, %f952, %f949;
ld.shared.f32 %f954, [%r26+1408];
add.ftz.f32 %f955, %f954, %f951;
mov.b32 %r1215, %f953;
shfl.sync.bfly.b32 %r1219|%p306, %r1215, %r1186, %r1185, %r1187;
mov.b32 %f956, %r1219;
add.ftz.f32 %f957, %f953, %f956;
mov.b32 %r1220, %f957;
shfl.sync.bfly.b32 %r1222|%p307, %r1220, %r1190, %r1185, %r1187;
mov.b32 %f958, %r1222;
add.ftz.f32 %f959, %f957, %f958;
mov.b32 %r1223, %f959;
shfl.sync.bfly.b32 %r1225|%p308, %r1223, %r1193, %r1185, %r1187;
mov.b32 %f960, %r1225;
add.ftz.f32 %f961, %f959, %f960;
mov.b32 %r1226, %f961;
shfl.sync.bfly.b32 %r1228|%p309, %r1226, %r1196, %r1185, %r1187;
mov.b32 %f962, %r1228;
add.ftz.f32 %f963, %f961, %f962;
mov.b32 %r1229, %f963;
shfl.sync.bfly.b32 %r1231|%p310, %r1229, %r1199, %r1185, %r1187;
mov.b32 %f964, %r1231;
add.ftz.f32 %f78, %f963, %f964;
mov.b32 %r1232, %f955;
shfl.sync.bfly.b32 %r1233|%p311, %r1232, %r1186, %r1185, %r1187;
mov.b32 %f965, %r1233;
add.ftz.f32 %f966, %f955, %f965;
mov.b32 %r1234, %f966;
shfl.sync.bfly.b32 %r1235|%p312, %r1234, %r1190, %r1185, %r1187;
mov.b32 %f967, %r1235;
add.ftz.f32 %f968, %f966, %f967;
mov.b32 %r1236, %f968;
shfl.sync.bfly.b32 %r1237|%p313, %r1236, %r1193, %r1185, %r1187;
mov.b32 %f969, %r1237;
add.ftz.f32 %f970, %f968, %f969;
mov.b32 %r1238, %f970;
shfl.sync.bfly.b32 %r1239|%p314, %r1238, %r1196, %r1185, %r1187;
mov.b32 %f971, %r1239;
add.ftz.f32 %f972, %f970, %f971;
mov.b32 %r1240, %f972;
shfl.sync.bfly.b32 %r1241|%p315, %r1240, %r1199, %r1185, %r1187;
mov.b32 %f973, %r1241;
add.ftz.f32 %f79, %f972, %f973;
setp.ne.s32 %p316, %r2, 3;
or.b32 %r1245, %r4, 3;
setp.ge.u32 %p317, %r1245, %r43;
or.pred %p318, %p316, %p317;
@%p318 bra $L__BB278_78;

add.ftz.f32 %f974, %f5, %f78;
add.ftz.f32 %f975, %f4, %f79;
mul.ftz.f32 %f976, %f974, %f975;
st.global.f32 [%rd27+12], %f976;
bra.uni $L__BB278_78;

$L__BB278_62:
ld.shared.f32 %f681, [%r24];
add.ftz.f32 %f682, %f681, %f30;
ld.shared.f32 %f683, [%r26];
add.ftz.f32 %f684, %f683, %f34;
ld.shared.f32 %f685, [%r24+512];
add.ftz.f32 %f686, %f685, %f682;
ld.shared.f32 %f687, [%r26+512];
add.ftz.f32 %f688, %f687, %f684;
ld.shared.f32 %f689, [%r24+1024];
add.ftz.f32 %f690, %f689, %f686;
ld.shared.f32 %f691, [%r26+1024];
add.ftz.f32 %f692, %f691, %f688;
mov.b32 %r1000, %f690;
mov.u32 %r1001, 31;
mov.u32 %r1002, 16;
mov.u32 %r1003, -1;
shfl.sync.bfly.b32 %r1004|%p215, %r1000, %r1002, %r1001, %r1003;
mov.b32 %f693, %r1004;
add.ftz.f32 %f694, %f690, %f693;
mov.b32 %r1005, %f694;
mov.u32 %r1006, 8;
shfl.sync.bfly.b32 %r1007|%p216, %r1005, %r1006, %r1001, %r1003;
mov.b32 %f695, %r1007;
add.ftz.f32 %f696, %f694, %f695;
mov.b32 %r1008, %f696;
mov.u32 %r1009, 4;
shfl.sync.bfly.b32 %r1010|%p217, %r1008, %r1009, %r1001, %r1003;
mov.b32 %f697, %r1010;
add.ftz.f32 %f698, %f696, %f697;
mov.b32 %r1011, %f698;
mov.u32 %r1012, 2;
shfl.sync.bfly.b32 %r1013|%p218, %r1011, %r1012, %r1001, %r1003;
mov.b32 %f699, %r1013;
add.ftz.f32 %f700, %f698, %f699;
mov.b32 %r1014, %f700;
mov.u32 %r1015, 1;
shfl.sync.bfly.b32 %r1016|%p219, %r1014, %r1015, %r1001, %r1003;
mov.b32 %f701, %r1016;
add.ftz.f32 %f64, %f700, %f701;
mov.b32 %r1017, %f692;
shfl.sync.bfly.b32 %r1018|%p220, %r1017, %r1002, %r1001, %r1003;
mov.b32 %f702, %r1018;
add.ftz.f32 %f703, %f692, %f702;
mov.b32 %r1019, %f703;
shfl.sync.bfly.b32 %r1020|%p221, %r1019, %r1006, %r1001, %r1003;
mov.b32 %f704, %r1020;
add.ftz.f32 %f705, %f703, %f704;
mov.b32 %r1021, %f705;
shfl.sync.bfly.b32 %r1022|%p222, %r1021, %r1009, %r1001, %r1003;
mov.b32 %f706, %r1022;
add.ftz.f32 %f707, %f705, %f706;
mov.b32 %r1023, %f707;
shfl.sync.bfly.b32 %r1024|%p223, %r1023, %r1012, %r1001, %r1003;
mov.b32 %f708, %r1024;
add.ftz.f32 %f709, %f707, %f708;
mov.b32 %r1025, %f709;
shfl.sync.bfly.b32 %r1026|%p224, %r1025, %r1015, %r1001, %r1003;
mov.b32 %f710, %r1026;
add.ftz.f32 %f65, %f709, %f710;
setp.ne.s32 %p225, %r2, 0;
setp.ge.u32 %p226, %r4, %r43;
or.pred %p227, %p225, %p226;
@%p227 bra $L__BB278_64;

add.ftz.f32 %f711, %f5, %f64;
add.ftz.f32 %f712, %f4, %f65;
mul.ftz.f32 %f713, %f712, 0f3F000000;
mul.ftz.f32 %f714, %f712, 0f3F4C422A;
mul.ftz.f32 %f715, %f712, 0f3D372713;
fma.rn.ftz.f32 %f716, %f712, %f715, 0f3F800000;
mul.ftz.f32 %f717, %f714, %f716;
tanh.approx.f32 %f718, %f717;
add.ftz.f32 %f719, %f718, 0f3F800000;
mul.ftz.f32 %f720, %f713, %f719;
mul.ftz.f32 %f721, %f711, %f720;
st.global.f32 [%rd27], %f721;

$L__BB278_64:
ld.shared.f32 %f722, [%r24+128];
add.ftz.f32 %f723, %f722, %f29;
ld.shared.f32 %f724, [%r26+128];
add.ftz.f32 %f725, %f724, %f33;
ld.shared.f32 %f726, [%r24+640];
add.ftz.f32 %f727, %f726, %f723;
ld.shared.f32 %f728, [%r26+640];
add.ftz.f32 %f729, %f728, %f725;
ld.shared.f32 %f730, [%r24+1152];
add.ftz.f32 %f731, %f730, %f727;
ld.shared.f32 %f732, [%r26+1152];
add.ftz.f32 %f733, %f732, %f729;
mov.b32 %r1030, %f731;
shfl.sync.bfly.b32 %r1034|%p228, %r1030, %r1002, %r1001, %r1003;
mov.b32 %f734, %r1034;
add.ftz.f32 %f735, %f731, %f734;
mov.b32 %r1035, %f735;
shfl.sync.bfly.b32 %r1037|%p229, %r1035, %r1006, %r1001, %r1003;
mov.b32 %f736, %r1037;
add.ftz.f32 %f737, %f735, %f736;
mov.b32 %r1038, %f737;
shfl.sync.bfly.b32 %r1040|%p230, %r1038, %r1009, %r1001, %r1003;
mov.b32 %f738, %r1040;
add.ftz.f32 %f739, %f737, %f738;
mov.b32 %r1041, %f739;
shfl.sync.bfly.b32 %r1043|%p231, %r1041, %r1012, %r1001, %r1003;
mov.b32 %f740, %r1043;
add.ftz.f32 %f741, %f739, %f740;
mov.b32 %r1044, %f741;
shfl.sync.bfly.b32 %r1046|%p232, %r1044, %r1015, %r1001, %r1003;
mov.b32 %f742, %r1046;
add.ftz.f32 %f66, %f741, %f742;
mov.b32 %r1047, %f733;
shfl.sync.bfly.b32 %r1048|%p233, %r1047, %r1002, %r1001, %r1003;
mov.b32 %f743, %r1048;
add.ftz.f32 %f744, %f733, %f743;
mov.b32 %r1049, %f744;
shfl.sync.bfly.b32 %r1050|%p234, %r1049, %r1006, %r1001, %r1003;
mov.b32 %f745, %r1050;
add.ftz.f32 %f746, %f744, %f745;
mov.b32 %r1051, %f746;
shfl.sync.bfly.b32 %r1052|%p235, %r1051, %r1009, %r1001, %r1003;
mov.b32 %f747, %r1052;
add.ftz.f32 %f748, %f746, %f747;
mov.b32 %r1053, %f748;
shfl.sync.bfly.b32 %r1054|%p236, %r1053, %r1012, %r1001, %r1003;
mov.b32 %f749, %r1054;
add.ftz.f32 %f750, %f748, %f749;
mov.b32 %r1055, %f750;
shfl.sync.bfly.b32 %r1056|%p237, %r1055, %r1015, %r1001, %r1003;
mov.b32 %f751, %r1056;
add.ftz.f32 %f67, %f750, %f751;
setp.ne.s32 %p238, %r2, 1;
or.b32 %r1060, %r4, 1;
setp.ge.u32 %p239, %r1060, %r43;
or.pred %p240, %p238, %p239;
@%p240 bra $L__BB278_66;

add.ftz.f32 %f752, %f5, %f66;
add.ftz.f32 %f753, %f4, %f67;
mul.ftz.f32 %f754, %f753, 0f3F000000;
mul.ftz.f32 %f755, %f753, 0f3F4C422A;
mul.ftz.f32 %f756, %f753, 0f3D372713;
fma.rn.ftz.f32 %f757, %f753, %f756, 0f3F800000;
mul.ftz.f32 %f758, %f755, %f757;
tanh.approx.f32 %f759, %f758;
add.ftz.f32 %f760, %f759, 0f3F800000;
mul.ftz.f32 %f761, %f754, %f760;
mul.ftz.f32 %f762, %f752, %f761;
st.global.f32 [%rd27+4], %f762;

$L__BB278_66:
ld.shared.f32 %f763, [%r24+256];
add.ftz.f32 %f764, %f763, %f28;
ld.shared.f32 %f765, [%r26+256];
add.ftz.f32 %f766, %f765, %f32;
ld.shared.f32 %f767, [%r24+768];
add.ftz.f32 %f768, %f767, %f764;
ld.shared.f32 %f769, [%r26+768];
add.ftz.f32 %f770, %f769, %f766;
ld.shared.f32 %f771, [%r24+1280];
add.ftz.f32 %f772, %f771, %f768;
ld.shared.f32 %f773, [%r26+1280];
add.ftz.f32 %f774, %f773, %f770;
mov.b32 %r1061, %f772;
mov.u32 %r1062, 31;
mov.u32 %r1063, 16;
mov.u32 %r1064, -1;
shfl.sync.bfly.b32 %r1065|%p241, %r1061, %r1063, %r1062, %r1064;
mov.b32 %f775, %r1065;
add.ftz.f32 %f776, %f772, %f775;
mov.b32 %r1066, %f776;
mov.u32 %r1067, 8;
shfl.sync.bfly.b32 %r1068|%p242, %r1066, %r1067, %r1062, %r1064;
mov.b32 %f777, %r1068;
add.ftz.f32 %f778, %f776, %f777;
mov.b32 %r1069, %f778;
mov.u32 %r1070, 4;
shfl.sync.bfly.b32 %r1071|%p243, %r1069, %r1070, %r1062, %r1064;
mov.b32 %f779, %r1071;
add.ftz.f32 %f780, %f778, %f779;
mov.b32 %r1072, %f780;
mov.u32 %r1073, 2;
shfl.sync.bfly.b32 %r1074|%p244, %r1072, %r1073, %r1062, %r1064;
mov.b32 %f781, %r1074;
add.ftz.f32 %f782, %f780, %f781;
mov.b32 %r1075, %f782;
mov.u32 %r1076, 1;
shfl.sync.bfly.b32 %r1077|%p245, %r1075, %r1076, %r1062, %r1064;
mov.b32 %f783, %r1077;
add.ftz.f32 %f68, %f782, %f783;
mov.b32 %r1078, %f774;
shfl.sync.bfly.b32 %r1079|%p246, %r1078, %r1063, %r1062, %r1064;
mov.b32 %f784, %r1079;
add.ftz.f32 %f785, %f774, %f784;
mov.b32 %r1080, %f785;
shfl.sync.bfly.b32 %r1081|%p247, %r1080, %r1067, %r1062, %r1064;
mov.b32 %f786, %r1081;
add.ftz.f32 %f787, %f785, %f786;
mov.b32 %r1082, %f787;
shfl.sync.bfly.b32 %r1083|%p248, %r1082, %r1070, %r1062, %r1064;
mov.b32 %f788, %r1083;
add.ftz.f32 %f789, %f787, %f788;
mov.b32 %r1084, %f789;
shfl.sync.bfly.b32 %r1085|%p249, %r1084, %r1073, %r1062, %r1064;
mov.b32 %f790, %r1085;
add.ftz.f32 %f791, %f789, %f790;
mov.b32 %r1086, %f791;
shfl.sync.bfly.b32 %r1087|%p250, %r1086, %r1076, %r1062, %r1064;
mov.b32 %f792, %r1087;
add.ftz.f32 %f69, %f791, %f792;
setp.ne.s32 %p251, %r2, 2;
or.b32 %r1091, %r4, 2;
setp.ge.u32 %p252, %r1091, %r43;
or.pred %p253, %p251, %p252;
@%p253 bra $L__BB278_68;

add.ftz.f32 %f793, %f5, %f68;
add.ftz.f32 %f794, %f4, %f69;
mul.ftz.f32 %f795, %f794, 0f3F000000;
mul.ftz.f32 %f796, %f794, 0f3F4C422A;
mul.ftz.f32 %f797, %f794, 0f3D372713;
fma.rn.ftz.f32 %f798, %f794, %f797, 0f3F800000;
mul.ftz.f32 %f799, %f796, %f798;
tanh.approx.f32 %f800, %f799;
add.ftz.f32 %f801, %f800, 0f3F800000;
mul.ftz.f32 %f802, %f795, %f801;
mul.ftz.f32 %f803, %f793, %f802;
st.global.f32 [%rd27+8], %f803;

$L__BB278_68:
ld.shared.f32 %f804, [%r24+384];
add.ftz.f32 %f805, %f804, %f27;
ld.shared.f32 %f806, [%r26+384];
add.ftz.f32 %f807, %f806, %f31;
ld.shared.f32 %f808, [%r24+896];
add.ftz.f32 %f809, %f808, %f805;
ld.shared.f32 %f810, [%r26+896];
add.ftz.f32 %f811, %f810, %f807;
ld.shared.f32 %f812, [%r24+1408];
add.ftz.f32 %f813, %f812, %f809;
ld.shared.f32 %f814, [%r26+1408];
add.ftz.f32 %f815, %f814, %f811;
mov.b32 %r1092, %f813;
shfl.sync.bfly.b32 %r1096|%p254, %r1092, %r1063, %r1062, %r1064;
mov.b32 %f816, %r1096;
add.ftz.f32 %f817, %f813, %f816;
mov.b32 %r1097, %f817;
shfl.sync.bfly.b32 %r1099|%p255, %r1097, %r1067, %r1062, %r1064;
mov.b32 %f818, %r1099;
add.ftz.f32 %f819, %f817, %f818;
mov.b32 %r1100, %f819;
shfl.sync.bfly.b32 %r1102|%p256, %r1100, %r1070, %r1062, %r1064;
mov.b32 %f820, %r1102;
add.ftz.f32 %f821, %f819, %f820;
mov.b32 %r1103, %f821;
shfl.sync.bfly.b32 %r1105|%p257, %r1103, %r1073, %r1062, %r1064;
mov.b32 %f822, %r1105;
add.ftz.f32 %f823, %f821, %f822;
mov.b32 %r1106, %f823;
shfl.sync.bfly.b32 %r1108|%p258, %r1106, %r1076, %r1062, %r1064;
mov.b32 %f824, %r1108;
add.ftz.f32 %f70, %f823, %f824;
mov.b32 %r1109, %f815;
shfl.sync.bfly.b32 %r1110|%p259, %r1109, %r1063, %r1062, %r1064;
mov.b32 %f825, %r1110;
add.ftz.f32 %f826, %f815, %f825;
mov.b32 %r1111, %f826;
shfl.sync.bfly.b32 %r1112|%p260, %r1111, %r1067, %r1062, %r1064;
mov.b32 %f827, %r1112;
add.ftz.f32 %f828, %f826, %f827;
mov.b32 %r1113, %f828;
shfl.sync.bfly.b32 %r1114|%p261, %r1113, %r1070, %r1062, %r1064;
mov.b32 %f829, %r1114;
add.ftz.f32 %f830, %f828, %f829;
mov.b32 %r1115, %f830;
shfl.sync.bfly.b32 %r1116|%p262, %r1115, %r1073, %r1062, %r1064;
mov.b32 %f831, %r1116;
add.ftz.f32 %f832, %f830, %f831;
mov.b32 %r1117, %f832;
shfl.sync.bfly.b32 %r1118|%p263, %r1117, %r1076, %r1062, %r1064;
mov.b32 %f833, %r1118;
add.ftz.f32 %f71, %f832, %f833;
setp.ne.s32 %p264, %r2, 3;
or.b32 %r1122, %r4, 3;
setp.ge.u32 %p265, %r1122, %r43;
or.pred %p266, %p264, %p265;
@%p266 bra $L__BB278_78;

add.ftz.f32 %f834, %f5, %f70;
add.ftz.f32 %f835, %f4, %f71;
mul.ftz.f32 %f836, %f835, 0f3F000000;
mul.ftz.f32 %f837, %f835, 0f3F4C422A;
mul.ftz.f32 %f838, %f835, 0f3D372713;
fma.rn.ftz.f32 %f839, %f835, %f838, 0f3F800000;
mul.ftz.f32 %f840, %f837, %f839;
tanh.approx.f32 %f841, %f840;
add.ftz.f32 %f842, %f841, 0f3F800000;
mul.ftz.f32 %f843, %f836, %f842;
mul.ftz.f32 %f844, %f834, %f843;
st.global.f32 [%rd27+12], %f844;
bra.uni $L__BB278_78;

$L__BB278_54:
ld.shared.f32 %f357, [%r24];
add.ftz.f32 %f358, %f357, %f30;
ld.shared.f32 %f359, [%r26];
add.ftz.f32 %f360, %f359, %f34;
ld.shared.f32 %f361, [%r24+512];
add.ftz.f32 %f362, %f361, %f358;
ld.shared.f32 %f363, [%r26+512];
add.ftz.f32 %f364, %f363, %f360;
ld.shared.f32 %f365, [%r24+1024];
add.ftz.f32 %f366, %f365, %f362;
ld.shared.f32 %f367, [%r26+1024];
add.ftz.f32 %f368, %f367, %f364;
mov.b32 %r754, %f366;
mov.u32 %r755, 31;
mov.u32 %r756, 16;
mov.u32 %r757, -1;
shfl.sync.bfly.b32 %r758|%p111, %r754, %r756, %r755, %r757;
mov.b32 %f369, %r758;
add.ftz.f32 %f370, %f366, %f369;
mov.b32 %r759, %f370;
mov.u32 %r760, 8;
shfl.sync.bfly.b32 %r761|%p112, %r759, %r760, %r755, %r757;
mov.b32 %f371, %r761;
add.ftz.f32 %f372, %f370, %f371;
mov.b32 %r762, %f372;
mov.u32 %r763, 4;
shfl.sync.bfly.b32 %r764|%p113, %r762, %r763, %r755, %r757;
mov.b32 %f373, %r764;
add.ftz.f32 %f374, %f372, %f373;
mov.b32 %r765, %f374;
mov.u32 %r766, 2;
shfl.sync.bfly.b32 %r767|%p114, %r765, %r766, %r755, %r757;
mov.b32 %f375, %r767;
add.ftz.f32 %f376, %f374, %f375;
mov.b32 %r768, %f376;
mov.u32 %r769, 1;
shfl.sync.bfly.b32 %r770|%p115, %r768, %r769, %r755, %r757;
mov.b32 %f377, %r770;
add.ftz.f32 %f48, %f376, %f377;
mov.b32 %r771, %f368;
shfl.sync.bfly.b32 %r772|%p116, %r771, %r756, %r755, %r757;
mov.b32 %f378, %r772;
add.ftz.f32 %f379, %f368, %f378;
mov.b32 %r773, %f379;
shfl.sync.bfly.b32 %r774|%p117, %r773, %r760, %r755, %r757;
mov.b32 %f380, %r774;
add.ftz.f32 %f381, %f379, %f380;
mov.b32 %r775, %f381;
shfl.sync.bfly.b32 %r776|%p118, %r775, %r763, %r755, %r757;
mov.b32 %f382, %r776;
add.ftz.f32 %f383, %f381, %f382;
mov.b32 %r777, %f383;
shfl.sync.bfly.b32 %r778|%p119, %r777, %r766, %r755, %r757;
mov.b32 %f384, %r778;
add.ftz.f32 %f385, %f383, %f384;
mov.b32 %r779, %f385;
shfl.sync.bfly.b32 %r780|%p120, %r779, %r769, %r755, %r757;
mov.b32 %f386, %r780;
add.ftz.f32 %f49, %f385, %f386;
setp.ne.s32 %p121, %r2, 0;
setp.ge.u32 %p122, %r4, %r43;
or.pred %p123, %p121, %p122;
@%p123 bra $L__BB278_56;

add.ftz.f32 %f387, %f5, %f48;
add.ftz.f32 %f388, %f4, %f49;
mov.f32 %f389, 0f40E00000;
min.ftz.f32 %f390, %f388, %f389;
min.ftz.f32 %f391, %f387, %f389;
mov.f32 %f392, 0fC0E00000;
max.ftz.f32 %f393, %f391, %f392;
mul.ftz.f32 %f394, %f390, 0f3FD9DB23;
mul.ftz.f32 %f395, %f394, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f396, %f395;
add.ftz.f32 %f397, %f396, 0f3F800000;
div.approx.ftz.f32 %f398, %f390, %f397;
add.ftz.f32 %f399, %f393, 0f3F800000;
mul.ftz.f32 %f400, %f399, %f398;
st.global.f32 [%rd27], %f400;

$L__BB278_56:
ld.shared.f32 %f401, [%r24+128];
add.ftz.f32 %f402, %f401, %f29;
ld.shared.f32 %f403, [%r26+128];
add.ftz.f32 %f404, %f403, %f33;
ld.shared.f32 %f405, [%r24+640];
add.ftz.f32 %f406, %f405, %f402;
ld.shared.f32 %f407, [%r26+640];
add.ftz.f32 %f408, %f407, %f404;
ld.shared.f32 %f409, [%r24+1152];
add.ftz.f32 %f410, %f409, %f406;
ld.shared.f32 %f411, [%r26+1152];
add.ftz.f32 %f412, %f411, %f408;
mov.b32 %r784, %f410;
shfl.sync.bfly.b32 %r788|%p124, %r784, %r756, %r755, %r757;
mov.b32 %f413, %r788;
add.ftz.f32 %f414, %f410, %f413;
mov.b32 %r789, %f414;
shfl.sync.bfly.b32 %r791|%p125, %r789, %r760, %r755, %r757;
mov.b32 %f415, %r791;
add.ftz.f32 %f416, %f414, %f415;
mov.b32 %r792, %f416;
shfl.sync.bfly.b32 %r794|%p126, %r792, %r763, %r755, %r757;
mov.b32 %f417, %r794;
add.ftz.f32 %f418, %f416, %f417;
mov.b32 %r795, %f418;
shfl.sync.bfly.b32 %r797|%p127, %r795, %r766, %r755, %r757;
mov.b32 %f419, %r797;
add.ftz.f32 %f420, %f418, %f419;
mov.b32 %r798, %f420;
shfl.sync.bfly.b32 %r800|%p128, %r798, %r769, %r755, %r757;
mov.b32 %f421, %r800;
add.ftz.f32 %f50, %f420, %f421;
mov.b32 %r801, %f412;
shfl.sync.bfly.b32 %r802|%p129, %r801, %r756, %r755, %r757;
mov.b32 %f422, %r802;
add.ftz.f32 %f423, %f412, %f422;
mov.b32 %r803, %f423;
shfl.sync.bfly.b32 %r804|%p130, %r803, %r760, %r755, %r757;
mov.b32 %f424, %r804;
add.ftz.f32 %f425, %f423, %f424;
mov.b32 %r805, %f425;
shfl.sync.bfly.b32 %r806|%p131, %r805, %r763, %r755, %r757;
mov.b32 %f426, %r806;
add.ftz.f32 %f427, %f425, %f426;
mov.b32 %r807, %f427;
shfl.sync.bfly.b32 %r808|%p132, %r807, %r766, %r755, %r757;
mov.b32 %f428, %r808;
add.ftz.f32 %f429, %f427, %f428;
mov.b32 %r809, %f429;
shfl.sync.bfly.b32 %r810|%p133, %r809, %r769, %r755, %r757;
mov.b32 %f430, %r810;
add.ftz.f32 %f51, %f429, %f430;
setp.ne.s32 %p134, %r2, 1;
or.b32 %r814, %r4, 1;
setp.ge.u32 %p135, %r814, %r43;
or.pred %p136, %p134, %p135;
@%p136 bra $L__BB278_58;

add.ftz.f32 %f431, %f5, %f50;
add.ftz.f32 %f432, %f4, %f51;
mov.f32 %f433, 0f40E00000;
min.ftz.f32 %f434, %f432, %f433;
min.ftz.f32 %f435, %f431, %f433;
mov.f32 %f436, 0fC0E00000;
max.ftz.f32 %f437, %f435, %f436;
mul.ftz.f32 %f438, %f434, 0f3FD9DB23;
mul.ftz.f32 %f439, %f438, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f440, %f439;
add.ftz.f32 %f441, %f440, 0f3F800000;
div.approx.ftz.f32 %f442, %f434, %f441;
add.ftz.f32 %f443, %f437, 0f3F800000;
mul.ftz.f32 %f444, %f443, %f442;
st.global.f32 [%rd27+4], %f444;

$L__BB278_58:
ld.shared.f32 %f445, [%r24+256];
add.ftz.f32 %f446, %f445, %f28;
ld.shared.f32 %f447, [%r26+256];
add.ftz.f32 %f448, %f447, %f32;
ld.shared.f32 %f449, [%r24+768];
add.ftz.f32 %f450, %f449, %f446;
ld.shared.f32 %f451, [%r26+768];
add.ftz.f32 %f452, %f451, %f448;
ld.shared.f32 %f453, [%r24+1280];
add.ftz.f32 %f454, %f453, %f450;
ld.shared.f32 %f455, [%r26+1280];
add.ftz.f32 %f456, %f455, %f452;
mov.b32 %r815, %f454;
mov.u32 %r816, 31;
mov.u32 %r817, 16;
mov.u32 %r818, -1;
shfl.sync.bfly.b32 %r819|%p137, %r815, %r817, %r816, %r818;
mov.b32 %f457, %r819;
add.ftz.f32 %f458, %f454, %f457;
mov.b32 %r820, %f458;
mov.u32 %r821, 8;
shfl.sync.bfly.b32 %r822|%p138, %r820, %r821, %r816, %r818;
mov.b32 %f459, %r822;
add.ftz.f32 %f460, %f458, %f459;
mov.b32 %r823, %f460;
mov.u32 %r824, 4;
shfl.sync.bfly.b32 %r825|%p139, %r823, %r824, %r816, %r818;
mov.b32 %f461, %r825;
add.ftz.f32 %f462, %f460, %f461;
mov.b32 %r826, %f462;
mov.u32 %r827, 2;
shfl.sync.bfly.b32 %r828|%p140, %r826, %r827, %r816, %r818;
mov.b32 %f463, %r828;
add.ftz.f32 %f464, %f462, %f463;
mov.b32 %r829, %f464;
mov.u32 %r830, 1;
shfl.sync.bfly.b32 %r831|%p141, %r829, %r830, %r816, %r818;
mov.b32 %f465, %r831;
add.ftz.f32 %f52, %f464, %f465;
mov.b32 %r832, %f456;
shfl.sync.bfly.b32 %r833|%p142, %r832, %r817, %r816, %r818;
mov.b32 %f466, %r833;
add.ftz.f32 %f467, %f456, %f466;
mov.b32 %r834, %f467;
shfl.sync.bfly.b32 %r835|%p143, %r834, %r821, %r816, %r818;
mov.b32 %f468, %r835;
add.ftz.f32 %f469, %f467, %f468;
mov.b32 %r836, %f469;
shfl.sync.bfly.b32 %r837|%p144, %r836, %r824, %r816, %r818;
mov.b32 %f470, %r837;
add.ftz.f32 %f471, %f469, %f470;
mov.b32 %r838, %f471;
shfl.sync.bfly.b32 %r839|%p145, %r838, %r827, %r816, %r818;
mov.b32 %f472, %r839;
add.ftz.f32 %f473, %f471, %f472;
mov.b32 %r840, %f473;
shfl.sync.bfly.b32 %r841|%p146, %r840, %r830, %r816, %r818;
mov.b32 %f474, %r841;
add.ftz.f32 %f53, %f473, %f474;
setp.ne.s32 %p147, %r2, 2;
or.b32 %r845, %r4, 2;
setp.ge.u32 %p148, %r845, %r43;
or.pred %p149, %p147, %p148;
@%p149 bra $L__BB278_60;

add.ftz.f32 %f475, %f5, %f52;
add.ftz.f32 %f476, %f4, %f53;
mov.f32 %f477, 0f40E00000;
min.ftz.f32 %f478, %f476, %f477;
min.ftz.f32 %f479, %f475, %f477;
mov.f32 %f480, 0fC0E00000;
max.ftz.f32 %f481, %f479, %f480;
mul.ftz.f32 %f482, %f478, 0f3FD9DB23;
mul.ftz.f32 %f483, %f482, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f484, %f483;
add.ftz.f32 %f485, %f484, 0f3F800000;
div.approx.ftz.f32 %f486, %f478, %f485;
add.ftz.f32 %f487, %f481, 0f3F800000;
mul.ftz.f32 %f488, %f487, %f486;
st.global.f32 [%rd27+8], %f488;

$L__BB278_60:
ld.shared.f32 %f489, [%r24+384];
add.ftz.f32 %f490, %f489, %f27;
ld.shared.f32 %f491, [%r26+384];
add.ftz.f32 %f492, %f491, %f31;
ld.shared.f32 %f493, [%r24+896];
add.ftz.f32 %f494, %f493, %f490;
ld.shared.f32 %f495, [%r26+896];
add.ftz.f32 %f496, %f495, %f492;
ld.shared.f32 %f497, [%r24+1408];
add.ftz.f32 %f498, %f497, %f494;
ld.shared.f32 %f499, [%r26+1408];
add.ftz.f32 %f500, %f499, %f496;
mov.b32 %r846, %f498;
shfl.sync.bfly.b32 %r850|%p150, %r846, %r817, %r816, %r818;
mov.b32 %f501, %r850;
add.ftz.f32 %f502, %f498, %f501;
mov.b32 %r851, %f502;
shfl.sync.bfly.b32 %r853|%p151, %r851, %r821, %r816, %r818;
mov.b32 %f503, %r853;
add.ftz.f32 %f504, %f502, %f503;
mov.b32 %r854, %f504;
shfl.sync.bfly.b32 %r856|%p152, %r854, %r824, %r816, %r818;
mov.b32 %f505, %r856;
add.ftz.f32 %f506, %f504, %f505;
mov.b32 %r857, %f506;
shfl.sync.bfly.b32 %r859|%p153, %r857, %r827, %r816, %r818;
mov.b32 %f507, %r859;
add.ftz.f32 %f508, %f506, %f507;
mov.b32 %r860, %f508;
shfl.sync.bfly.b32 %r862|%p154, %r860, %r830, %r816, %r818;
mov.b32 %f509, %r862;
add.ftz.f32 %f54, %f508, %f509;
mov.b32 %r863, %f500;
shfl.sync.bfly.b32 %r864|%p155, %r863, %r817, %r816, %r818;
mov.b32 %f510, %r864;
add.ftz.f32 %f511, %f500, %f510;
mov.b32 %r865, %f511;
shfl.sync.bfly.b32 %r866|%p156, %r865, %r821, %r816, %r818;
mov.b32 %f512, %r866;
add.ftz.f32 %f513, %f511, %f512;
mov.b32 %r867, %f513;
shfl.sync.bfly.b32 %r868|%p157, %r867, %r824, %r816, %r818;
mov.b32 %f514, %r868;
add.ftz.f32 %f515, %f513, %f514;
mov.b32 %r869, %f515;
shfl.sync.bfly.b32 %r870|%p158, %r869, %r827, %r816, %r818;
mov.b32 %f516, %r870;
add.ftz.f32 %f517, %f515, %f516;
mov.b32 %r871, %f517;
shfl.sync.bfly.b32 %r872|%p159, %r871, %r830, %r816, %r818;
mov.b32 %f518, %r872;
add.ftz.f32 %f55, %f517, %f518;
setp.ne.s32 %p160, %r2, 3;
or.b32 %r876, %r4, 3;
setp.ge.u32 %p161, %r876, %r43;
or.pred %p162, %p160, %p161;
@%p162 bra $L__BB278_78;

add.ftz.f32 %f519, %f5, %f54;
add.ftz.f32 %f520, %f4, %f55;
mov.f32 %f521, 0f40E00000;
min.ftz.f32 %f522, %f520, %f521;
min.ftz.f32 %f523, %f519, %f521;
mov.f32 %f524, 0fC0E00000;
max.ftz.f32 %f525, %f523, %f524;
mul.ftz.f32 %f526, %f522, 0f3FD9DB23;
mul.ftz.f32 %f527, %f526, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f528, %f527;
add.ftz.f32 %f529, %f528, 0f3F800000;
div.approx.ftz.f32 %f530, %f522, %f529;
add.ftz.f32 %f531, %f525, 0f3F800000;
mul.ftz.f32 %f532, %f531, %f530;
st.global.f32 [%rd27+12], %f532;
bra.uni $L__BB278_78;

}
.entry _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<38>;
.reg .b16 %rs<29>;
.reg .f32 %f<109>;
.reg .b32 %r<339>;
.reg .b64 %rd<40>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

ld.param.u64 %rd11, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd12, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd13, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd14, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r19, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r18, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r17, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
mov.u32 %r1, %tid.y;
shl.b32 %r35, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r35, %r2;
mov.u32 %r36, %ctaid.x;
shl.b32 %r4, %r36, 2;
shr.u32 %r5, %r34, 5;
mov.u32 %r337, %ctaid.y;
setp.eq.s64 %p1, %rd13, 0;
@%p1 bra $L__BB279_2;

cvta.to.global.u64 %rd15, %rd13;
mul.wide.u32 %rd16, %r337, 4;
add.s64 %rd17, %rd15, %rd16;
ld.global.nc.u32 %r336, [%rd17];
mul.hi.u32 %r37, %r337, %r17;
add.s32 %r38, %r37, %r337;
shr.u32 %r39, %r38, %r18;
mul.lo.s32 %r40, %r39, %r19;
sub.s32 %r337, %r337, %r40;
bra.uni $L__BB279_3;

$L__BB279_2:
mul.hi.u32 %r41, %r337, %r22;
add.s32 %r42, %r41, %r337;
shr.u32 %r336, %r42, %r23;

$L__BB279_3:
shr.u32 %r43, %r3, 31;
add.s32 %r44, %r3, %r43;
shr.s32 %r338, %r44, 1;
setp.ge.s32 %p2, %r338, %r5;
mov.f32 %f105, 0f00000000;
mov.f32 %f106, %f105;
mov.f32 %f107, %f105;
mov.f32 %f108, %f105;
@%p2 bra $L__BB279_6;

mov.u32 %r45, %ctaid.z;
and.b32 %r48, %r44, -2;
sub.s32 %r49, %r3, %r48;
shl.b32 %r50, %r49, 1;
cvt.s64.s32 %rd1, %r50;
shl.b32 %r51, %r49, 2;
mad.lo.s32 %r52, %r4, %r20, %r338;
mul.hi.u32 %r53, %r45, %r28;
add.s32 %r54, %r53, %r45;
shr.u32 %r55, %r54, %r29;
mad.lo.s32 %r56, %r55, %r31, %r52;
mad.lo.s32 %r57, %r336, %r25, %r56;
mul.wide.s32 %rd2, %r57, 18;
mul.wide.s32 %rd18, %r51, 2;
add.s64 %rd19, %rd2, %rd18;
add.s64 %rd3, %rd19, 8;
mul.wide.s32 %rd4, %r20, 18;
cvta.to.global.u64 %rd39, %rd11;
cvta.to.global.u64 %rd6, %rd12;
mul.lo.s32 %r58, %r337, %r26;
cvt.u64.u32 %rd20, %r58;
mul.lo.s32 %r59, %r45, %r32;
cvt.u64.u32 %rd21, %r59;
add.s64 %rd7, %rd20, %rd21;
mov.f32 %f105, 0f00000000;
shl.b64 %rd26, %rd1, 2;

$L__BB279_5:
cvt.u64.u32 %rd22, %r338;
add.s64 %rd23, %rd7, %rd22;
mul.lo.s64 %rd24, %rd23, 36;
add.s64 %rd25, %rd6, %rd24;
add.s64 %rd27, %rd25, %rd26;
add.s64 %rd28, %rd39, %rd3;
add.s64 %rd29, %rd28, -6;
ld.global.nc.u16 %rs5, [%rd28+-6];
cvt.u32.u16 %r128, %rs5;
ld.global.nc.u16 %rs6, [%rd28+-4];
mov.b32 %r129, {%rs5, %rs6};
and.b32 %r130, %r129, -2004318072;
shr.u32 %r131, %r130, 1;
or.b32 %r132, %r131, 839922192;
and.b32 %r133, %r128, 30583;
mov.u32 %r134, -152379953;
mov.u32 %r135, -1079142271;
prmt.b32 %r136, %r135, %r134, %r133;
mov.u32 %r137, 1901675829;
mov.u32 %r138, 639175937;
prmt.b32 %r139, %r138, %r137, %r133;
and.b32 %r140, %r132, 30292;
prmt.b32 %r141, %r136, %r139, %r140;
and.b16 %rs7, %rs6, 30583;
cvt.u32.u16 %r142, %rs7;
prmt.b32 %r143, %r135, %r134, %r142;
prmt.b32 %r144, %r138, %r137, %r142;
shr.u32 %r145, %r132, 16;
prmt.b32 %r146, %r143, %r144, %r145;
mov.u32 %r147, 25632;
prmt.b32 %r61, %r141, %r146, %r147;
mov.u32 %r148, 30001;
prmt.b32 %r65, %r141, %r146, %r148;
ld.global.nc.u32 %r62, [%rd27+4];
mov.u32 %r114, 0;

dp4a.s32.s32 %r60, %r61, %r62, %r114;

ld.global.nc.u32 %r66, [%rd27+20];

dp4a.s32.s32 %r64, %r65, %r66, %r60;

ld.global.nc.u16 %rs8, [%rd28+-2];
cvt.u32.u16 %r149, %rs8;
ld.global.nc.u16 %rs9, [%rd28];
mov.b32 %r150, {%rs8, %rs9};
and.b32 %r151, %r150, -2004318072;
shr.u32 %r152, %r151, 1;
or.b32 %r153, %r152, 839922192;
and.b32 %r154, %r149, 30583;
prmt.b32 %r155, %r135, %r134, %r154;
prmt.b32 %r156, %r138, %r137, %r154;
and.b32 %r157, %r153, 30292;
prmt.b32 %r158, %r155, %r156, %r157;
and.b16 %rs10, %rs9, 30583;
cvt.u32.u16 %r159, %rs10;
prmt.b32 %r160, %r135, %r134, %r159;
prmt.b32 %r161, %r138, %r137, %r159;
shr.u32 %r162, %r153, 16;
prmt.b32 %r163, %r160, %r161, %r162;
prmt.b32 %r69, %r158, %r163, %r147;
prmt.b32 %r73, %r158, %r163, %r148;
ld.global.nc.u32 %r70, [%rd27+8];

dp4a.s32.s32 %r68, %r69, %r70, %r64;

ld.global.nc.u32 %r74, [%rd27+24];

dp4a.s32.s32 %r72, %r73, %r74, %r68;

add.s64 %rd30, %rd39, %rd2;
ld.global.nc.u16 %rs1, [%rd30];

{ cvt.f32.f16 %f25, %rs1;}


ld.global.nc.u32 %r76, [%rd25];

{.reg .f16 low,high;
mov.b32 {low,high},%r76;
cvt.f32.f16 %f26, low;}


mul.ftz.f32 %f33, %f25, %f26;
cvt.rn.f32.s32 %f34, %r72;
fma.rn.ftz.f32 %f108, %f33, %f34, %f108;
add.s64 %rd31, %rd29, %rd4;
ld.global.nc.u16 %rs11, [%rd31];
cvt.u32.u16 %r164, %rs11;
ld.global.nc.u16 %rs12, [%rd31+2];
mov.b32 %r165, {%rs11, %rs12};
and.b32 %r166, %r165, -2004318072;
shr.u32 %r167, %r166, 1;
or.b32 %r168, %r167, 839922192;
and.b32 %r169, %r164, 30583;
prmt.b32 %r170, %r135, %r134, %r169;
prmt.b32 %r171, %r138, %r137, %r169;
and.b32 %r172, %r168, 30292;
prmt.b32 %r173, %r170, %r171, %r172;
and.b16 %rs13, %rs12, 30583;
cvt.u32.u16 %r174, %rs13;
prmt.b32 %r175, %r135, %r134, %r174;
prmt.b32 %r176, %r138, %r137, %r174;
shr.u32 %r177, %r168, 16;
prmt.b32 %r178, %r175, %r176, %r177;
prmt.b32 %r78, %r173, %r178, %r147;
prmt.b32 %r82, %r173, %r178, %r148;

dp4a.s32.s32 %r77, %r78, %r62, %r114;


dp4a.s32.s32 %r81, %r82, %r66, %r77;

ld.global.nc.u16 %rs14, [%rd31+4];
cvt.u32.u16 %r179, %rs14;
ld.global.nc.u16 %rs15, [%rd31+6];
mov.b32 %r180, {%rs14, %rs15};
and.b32 %r181, %r180, -2004318072;
shr.u32 %r182, %r181, 1;
or.b32 %r183, %r182, 839922192;
and.b32 %r184, %r179, 30583;
prmt.b32 %r185, %r135, %r134, %r184;
prmt.b32 %r186, %r138, %r137, %r184;
and.b32 %r187, %r183, 30292;
prmt.b32 %r188, %r185, %r186, %r187;
and.b16 %rs16, %rs15, 30583;
cvt.u32.u16 %r189, %rs16;
prmt.b32 %r190, %r135, %r134, %r189;
prmt.b32 %r191, %r138, %r137, %r189;
shr.u32 %r192, %r183, 16;
prmt.b32 %r193, %r190, %r191, %r192;
prmt.b32 %r86, %r188, %r193, %r147;
prmt.b32 %r90, %r188, %r193, %r148;

dp4a.s32.s32 %r85, %r86, %r70, %r81;


dp4a.s32.s32 %r89, %r90, %r74, %r85;

add.s64 %rd32, %rd30, %rd4;
ld.global.nc.u16 %rs2, [%rd32];

{ cvt.f32.f16 %f27, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r76;
cvt.f32.f16 %f28, low;}


mul.ftz.f32 %f35, %f27, %f28;
cvt.rn.f32.s32 %f36, %r89;
fma.rn.ftz.f32 %f107, %f35, %f36, %f107;
add.s64 %rd33, %rd31, %rd4;
ld.global.nc.u16 %rs17, [%rd33];
cvt.u32.u16 %r194, %rs17;
ld.global.nc.u16 %rs18, [%rd33+2];
mov.b32 %r195, {%rs17, %rs18};
and.b32 %r196, %r195, -2004318072;
shr.u32 %r197, %r196, 1;
or.b32 %r198, %r197, 839922192;
and.b32 %r199, %r194, 30583;
prmt.b32 %r200, %r135, %r134, %r199;
prmt.b32 %r201, %r138, %r137, %r199;
and.b32 %r202, %r198, 30292;
prmt.b32 %r203, %r200, %r201, %r202;
and.b16 %rs19, %rs18, 30583;
cvt.u32.u16 %r204, %rs19;
prmt.b32 %r205, %r135, %r134, %r204;
prmt.b32 %r206, %r138, %r137, %r204;
shr.u32 %r207, %r198, 16;
prmt.b32 %r208, %r205, %r206, %r207;
prmt.b32 %r95, %r203, %r208, %r147;
prmt.b32 %r99, %r203, %r208, %r148;

dp4a.s32.s32 %r94, %r95, %r62, %r114;


dp4a.s32.s32 %r98, %r99, %r66, %r94;

ld.global.nc.u16 %rs20, [%rd33+4];
cvt.u32.u16 %r209, %rs20;
ld.global.nc.u16 %rs21, [%rd33+6];
mov.b32 %r210, {%rs20, %rs21};
and.b32 %r211, %r210, -2004318072;
shr.u32 %r212, %r211, 1;
or.b32 %r213, %r212, 839922192;
and.b32 %r214, %r209, 30583;
prmt.b32 %r215, %r135, %r134, %r214;
prmt.b32 %r216, %r138, %r137, %r214;
and.b32 %r217, %r213, 30292;
prmt.b32 %r218, %r215, %r216, %r217;
and.b16 %rs22, %rs21, 30583;
cvt.u32.u16 %r219, %rs22;
prmt.b32 %r220, %r135, %r134, %r219;
prmt.b32 %r221, %r138, %r137, %r219;
shr.u32 %r222, %r213, 16;
prmt.b32 %r223, %r220, %r221, %r222;
prmt.b32 %r103, %r218, %r223, %r147;
prmt.b32 %r107, %r218, %r223, %r148;

dp4a.s32.s32 %r102, %r103, %r70, %r98;


dp4a.s32.s32 %r106, %r107, %r74, %r102;

add.s64 %rd34, %rd32, %rd4;
ld.global.nc.u16 %rs3, [%rd34];

{ cvt.f32.f16 %f29, %rs3;}



{.reg .f16 low,high;
mov.b32 {low,high},%r76;
cvt.f32.f16 %f30, low;}


mul.ftz.f32 %f37, %f29, %f30;
cvt.rn.f32.s32 %f38, %r106;
fma.rn.ftz.f32 %f106, %f37, %f38, %f106;
add.s64 %rd35, %rd33, %rd4;
ld.global.nc.u16 %rs23, [%rd35];
cvt.u32.u16 %r224, %rs23;
ld.global.nc.u16 %rs24, [%rd35+2];
mov.b32 %r225, {%rs23, %rs24};
and.b32 %r226, %r225, -2004318072;
shr.u32 %r227, %r226, 1;
or.b32 %r228, %r227, 839922192;
and.b32 %r229, %r224, 30583;
prmt.b32 %r230, %r135, %r134, %r229;
prmt.b32 %r231, %r138, %r137, %r229;
and.b32 %r232, %r228, 30292;
prmt.b32 %r233, %r230, %r231, %r232;
and.b16 %rs25, %rs24, 30583;
cvt.u32.u16 %r234, %rs25;
prmt.b32 %r235, %r135, %r134, %r234;
prmt.b32 %r236, %r138, %r137, %r234;
shr.u32 %r237, %r228, 16;
prmt.b32 %r238, %r235, %r236, %r237;
prmt.b32 %r112, %r233, %r238, %r147;
prmt.b32 %r116, %r233, %r238, %r148;

dp4a.s32.s32 %r111, %r112, %r62, %r114;


dp4a.s32.s32 %r115, %r116, %r66, %r111;

ld.global.nc.u16 %rs26, [%rd35+4];
cvt.u32.u16 %r239, %rs26;
ld.global.nc.u16 %rs27, [%rd35+6];
mov.b32 %r240, {%rs26, %rs27};
and.b32 %r241, %r240, -2004318072;
shr.u32 %r242, %r241, 1;
or.b32 %r243, %r242, 839922192;
and.b32 %r244, %r239, 30583;
prmt.b32 %r245, %r135, %r134, %r244;
prmt.b32 %r246, %r138, %r137, %r244;
and.b32 %r247, %r243, 30292;
prmt.b32 %r248, %r245, %r246, %r247;
and.b16 %rs28, %rs27, 30583;
cvt.u32.u16 %r249, %rs28;
prmt.b32 %r250, %r135, %r134, %r249;
prmt.b32 %r251, %r138, %r137, %r249;
shr.u32 %r252, %r243, 16;
prmt.b32 %r253, %r250, %r251, %r252;
prmt.b32 %r120, %r248, %r253, %r147;
prmt.b32 %r124, %r248, %r253, %r148;

dp4a.s32.s32 %r119, %r120, %r70, %r115;


dp4a.s32.s32 %r123, %r124, %r74, %r119;

add.s64 %rd36, %rd34, %rd4;
ld.global.nc.u16 %rs4, [%rd36];

{ cvt.f32.f16 %f31, %rs4;}



{.reg .f16 low,high;
mov.b32 {low,high},%r76;
cvt.f32.f16 %f32, low;}


mul.ftz.f32 %f39, %f31, %f32;
cvt.rn.f32.s32 %f40, %r123;
fma.rn.ftz.f32 %f105, %f39, %f40, %f105;
add.s64 %rd39, %rd39, 1152;
add.s32 %r338, %r338, 64;
setp.lt.s32 %p3, %r338, %r5;
@%p3 bra $L__BB279_5;

$L__BB279_6:
setp.eq.s32 %p4, %r1, 0;
@%p4 bra $L__BB279_8;

shl.b32 %r254, %r1, 9;
mov.u32 %r255, _ZZ13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r256, %r255, %r254;
shl.b32 %r257, %r2, 2;
add.s32 %r258, %r256, %r257;
st.shared.f32 [%r258+-512], %f108;
st.shared.f32 [%r258+-384], %f107;
st.shared.f32 [%r258+-256], %f106;
st.shared.f32 [%r258+-128], %f105;

$L__BB279_8:
setp.ne.s32 %p5, %r1, 0;
bar.sync 0;
@%p5 bra $L__BB279_17;

mov.u32 %r259, %ctaid.z;
mov.u32 %r260, %ctaid.y;
mad.lo.s32 %r261, %r260, %r27, %r4;
mad.lo.s32 %r262, %r259, %r33, %r261;
shl.b32 %r263, %r2, 2;
mov.u32 %r264, _ZZ13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r15, %r264, %r263;
mov.u32 %r265, 2;
ld.shared.f32 %f41, [%r15];
add.ftz.f32 %f42, %f41, %f108;
ld.shared.f32 %f43, [%r15+512];
add.ftz.f32 %f44, %f43, %f42;
ld.shared.f32 %f45, [%r15+1024];
add.ftz.f32 %f46, %f45, %f44;
mov.b32 %r266, %f46;
mov.u32 %r267, 31;
mov.u32 %r268, 16;
mov.u32 %r269, -1;
shfl.sync.bfly.b32 %r270|%p6, %r266, %r268, %r267, %r269;
mov.b32 %f47, %r270;
add.ftz.f32 %f48, %f46, %f47;
mov.b32 %r271, %f48;
mov.u32 %r272, 8;
shfl.sync.bfly.b32 %r273|%p7, %r271, %r272, %r267, %r269;
mov.b32 %f49, %r273;
add.ftz.f32 %f50, %f48, %f49;
mov.b32 %r274, %f50;
mov.u32 %r275, 4;
shfl.sync.bfly.b32 %r276|%p8, %r274, %r275, %r267, %r269;
mov.b32 %f51, %r276;
add.ftz.f32 %f52, %f50, %f51;
mov.b32 %r277, %f52;
shfl.sync.bfly.b32 %r278|%p9, %r277, %r265, %r267, %r269;
mov.b32 %f53, %r278;
add.ftz.f32 %f54, %f52, %f53;
mov.b32 %r279, %f54;
mov.u32 %r280, 1;
shfl.sync.bfly.b32 %r281|%p10, %r279, %r280, %r267, %r269;
mov.b32 %f55, %r281;
add.ftz.f32 %f13, %f54, %f55;
setp.ge.u32 %p11, %r4, %r21;
setp.ne.s32 %p12, %r2, 0;
cvta.to.global.u64 %rd37, %rd14;
mul.wide.u32 %rd38, %r262, 4;
add.s64 %rd10, %rd37, %rd38;
or.pred %p13, %p12, %p11;
@%p13 bra $L__BB279_11;

st.global.f32 [%rd10], %f13;

$L__BB279_11:
ld.shared.f32 %f56, [%r15+128];
add.ftz.f32 %f57, %f56, %f107;
ld.shared.f32 %f58, [%r15+640];
add.ftz.f32 %f59, %f58, %f57;
ld.shared.f32 %f60, [%r15+1152];
add.ftz.f32 %f61, %f60, %f59;
mov.b32 %r282, %f61;
shfl.sync.bfly.b32 %r286|%p14, %r282, %r268, %r267, %r269;
mov.b32 %f62, %r286;
add.ftz.f32 %f63, %f61, %f62;
mov.b32 %r287, %f63;
shfl.sync.bfly.b32 %r289|%p15, %r287, %r272, %r267, %r269;
mov.b32 %f64, %r289;
add.ftz.f32 %f65, %f63, %f64;
mov.b32 %r290, %f65;
shfl.sync.bfly.b32 %r292|%p16, %r290, %r275, %r267, %r269;
mov.b32 %f66, %r292;
add.ftz.f32 %f67, %f65, %f66;
mov.b32 %r293, %f67;
shfl.sync.bfly.b32 %r295|%p17, %r293, %r265, %r267, %r269;
mov.b32 %f68, %r295;
add.ftz.f32 %f69, %f67, %f68;
mov.b32 %r296, %f69;
shfl.sync.bfly.b32 %r298|%p18, %r296, %r280, %r267, %r269;
mov.b32 %f70, %r298;
add.ftz.f32 %f14, %f69, %f70;
add.s32 %r299, %r4, 1;
setp.ge.u32 %p19, %r299, %r21;
setp.ne.s32 %p20, %r2, 1;
or.pred %p21, %p20, %p19;
@%p21 bra $L__BB279_13;

st.global.f32 [%rd10+4], %f14;

$L__BB279_13:
ld.shared.f32 %f71, [%r15+256];
add.ftz.f32 %f72, %f71, %f106;
ld.shared.f32 %f73, [%r15+768];
add.ftz.f32 %f74, %f73, %f72;
ld.shared.f32 %f75, [%r15+1280];
add.ftz.f32 %f76, %f75, %f74;
mov.b32 %r300, %f76;
mov.u32 %r301, 31;
mov.u32 %r302, 16;
mov.u32 %r303, -1;
shfl.sync.bfly.b32 %r304|%p22, %r300, %r302, %r301, %r303;
mov.b32 %f77, %r304;
add.ftz.f32 %f78, %f76, %f77;
mov.b32 %r305, %f78;
mov.u32 %r306, 8;
shfl.sync.bfly.b32 %r307|%p23, %r305, %r306, %r301, %r303;
mov.b32 %f79, %r307;
add.ftz.f32 %f80, %f78, %f79;
mov.b32 %r308, %f80;
mov.u32 %r309, 4;
shfl.sync.bfly.b32 %r310|%p24, %r308, %r309, %r301, %r303;
mov.b32 %f81, %r310;
add.ftz.f32 %f82, %f80, %f81;
mov.b32 %r311, %f82;
mov.u32 %r312, 2;
shfl.sync.bfly.b32 %r313|%p25, %r311, %r312, %r301, %r303;
mov.b32 %f83, %r313;
add.ftz.f32 %f84, %f82, %f83;
mov.b32 %r314, %f84;
mov.u32 %r315, 1;
shfl.sync.bfly.b32 %r316|%p26, %r314, %r315, %r301, %r303;
mov.b32 %f85, %r316;
add.ftz.f32 %f15, %f84, %f85;
add.s32 %r317, %r4, 2;
setp.ge.u32 %p27, %r317, %r21;
setp.ne.s32 %p28, %r2, 2;
or.pred %p29, %p28, %p27;
@%p29 bra $L__BB279_15;

st.global.f32 [%rd10+8], %f15;

$L__BB279_15:
ld.shared.f32 %f86, [%r15+384];
add.ftz.f32 %f87, %f86, %f105;
ld.shared.f32 %f88, [%r15+896];
add.ftz.f32 %f89, %f88, %f87;
ld.shared.f32 %f90, [%r15+1408];
add.ftz.f32 %f91, %f90, %f89;
mov.b32 %r318, %f91;
shfl.sync.bfly.b32 %r322|%p30, %r318, %r302, %r301, %r303;
mov.b32 %f92, %r322;
add.ftz.f32 %f93, %f91, %f92;
mov.b32 %r323, %f93;
shfl.sync.bfly.b32 %r325|%p31, %r323, %r306, %r301, %r303;
mov.b32 %f94, %r325;
add.ftz.f32 %f95, %f93, %f94;
mov.b32 %r326, %f95;
shfl.sync.bfly.b32 %r328|%p32, %r326, %r309, %r301, %r303;
mov.b32 %f96, %r328;
add.ftz.f32 %f97, %f95, %f96;
mov.b32 %r329, %f97;
shfl.sync.bfly.b32 %r331|%p33, %r329, %r312, %r301, %r303;
mov.b32 %f98, %r331;
add.ftz.f32 %f99, %f97, %f98;
mov.b32 %r332, %f99;
shfl.sync.bfly.b32 %r334|%p34, %r332, %r315, %r301, %r303;
mov.b32 %f100, %r334;
add.ftz.f32 %f16, %f99, %f100;
add.s32 %r335, %r4, 3;
setp.ge.u32 %p35, %r335, %r21;
setp.ne.s32 %p36, %r2, 3;
or.pred %p37, %p36, %p35;
@%p37 bra $L__BB279_17;

st.global.f32 [%rd10+12], %f16;

$L__BB279_17:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<34>;
.reg .b16 %rs<15>;
.reg .f32 %f<120>;
.reg .b32 %r<262>;
.reg .b64 %rd<70>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[384];

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate[384];

ld.param.u64 %rd14, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd22, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
ld.param.f32 %f31, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+16];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+8];
ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3];
setp.eq.s64 %p2, %rd16, 0;
@%p2 bra $L__BB280_2;

cvta.to.global.u64 %rd23, %rd16;
mov.u32 %r39, %ctaid.y;
mul.wide.u32 %rd24, %r39, 4;
add.s64 %rd25, %rd23, %rd24;
ld.global.nc.u32 %r259, [%rd25];
mul.hi.u32 %r40, %r39, %r23;
add.s32 %r41, %r40, %r39;
shr.u32 %r42, %r41, %r24;
mul.lo.s32 %r43, %r42, %r25;
sub.s32 %r260, %r39, %r43;
bra.uni $L__BB280_3;

$L__BB280_2:
mov.u32 %r260, %ctaid.y;
mul.hi.u32 %r44, %r260, %r27;
add.s32 %r45, %r44, %r260;
shr.u32 %r259, %r45, %r28;

$L__BB280_3:
mov.u32 %r46, %tid.y;
mov.u32 %r47, %tid.x;
or.b32 %r48, %r47, %r46;
setp.ne.s32 %p3, %r48, 0;
mov.f32 %f113, 0f00000000;
mov.f32 %f109, %f113;
mov.f32 %f110, %f113;
@%p3 bra $L__BB280_8;

setp.eq.s64 %p4, %rd17, 0;
mov.f32 %f109, 0f00000000;
mov.f32 %f110, %f109;
@%p4 bra $L__BB280_6;

cvta.to.global.u64 %rd26, %rd17;
mov.u32 %r49, %ctaid.z;
mul.lo.s32 %r50, %r49, %r38;
cvt.u64.u32 %rd27, %r50;
mov.u32 %r51, %ctaid.y;
selp.b32 %r52, %r51, %r259, %p2;
mul.lo.s32 %r53, %r52, %r32;
cvt.u64.u32 %rd28, %r53;
mov.u32 %r54, %ctaid.x;
cvt.s64.s32 %rd29, %r54;
cvt.u64.u32 %rd30, %r47;
add.s64 %rd31, %rd30, %rd29;
add.s64 %rd32, %rd31, %rd27;
add.s64 %rd33, %rd32, %rd28;
shl.b64 %rd34, %rd33, 2;
add.s64 %rd35, %rd26, %rd34;
ld.global.f32 %f110, [%rd35];

$L__BB280_6:
setp.eq.s64 %p6, %rd19, 0;
setp.eq.s64 %p7, %rd18, 0;
or.pred %p8, %p6, %p7;
@%p8 bra $L__BB280_8;

cvta.to.global.u64 %rd36, %rd19;
mov.u32 %r56, %ctaid.z;
mul.lo.s32 %r57, %r56, %r38;
cvt.u64.u32 %rd37, %r57;
mov.u32 %r58, %ctaid.y;
selp.b32 %r59, %r58, %r259, %p2;
mul.lo.s32 %r60, %r59, %r32;
cvt.u64.u32 %rd38, %r60;
mov.u32 %r61, %ctaid.x;
cvt.s64.s32 %rd39, %r61;
cvt.u64.u32 %rd40, %r47;
add.s64 %rd41, %rd40, %rd39;
add.s64 %rd42, %rd41, %rd37;
add.s64 %rd43, %rd42, %rd38;
shl.b64 %rd44, %rd43, 2;
add.s64 %rd45, %rd36, %rd44;
ld.global.f32 %f109, [%rd45];

$L__BB280_8:
shl.b32 %r64, %r46, 5;
add.s32 %r66, %r64, %r47;
shr.u32 %r67, %r66, 31;
add.s32 %r68, %r66, %r67;
shr.s32 %r261, %r68, 1;
shr.u32 %r69, %r22, 5;
setp.ge.s32 %p10, %r261, %r69;
mov.f32 %f115, %f113;
@%p10 bra $L__BB280_13;

mov.u32 %r70, %ctaid.z;
and.b32 %r77, %r68, -2;
sub.s32 %r10, %r66, %r77;
shl.b32 %r78, %r10, 2;
mul.hi.u32 %r79, %r70, %r33;
add.s32 %r80, %r79, %r70;
shr.u32 %r81, %r80, %r34;
mad.lo.s32 %r82, %r81, %r36, %r261;
mad.lo.s32 %r83, %r259, %r30, %r82;
mov.u32 %r84, %ctaid.x;
mad.lo.s32 %r85, %r84, %r26, %r83;
mul.wide.s32 %rd46, %r85, 18;
cvta.to.global.u64 %rd47, %rd18;
add.s64 %rd69, %rd47, %rd46;
cvta.to.global.u64 %rd48, %rd14;
add.s64 %rd49, %rd48, %rd46;
add.s64 %rd68, %rd49, 4;
mul.wide.s32 %rd50, %r78, 2;
add.s64 %rd51, %rd46, %rd50;
add.s64 %rd52, %rd48, %rd51;
add.s64 %rd67, %rd52, 8;
mul.lo.s32 %r86, %r260, %r31;
cvt.u64.u32 %rd53, %r86;
mul.lo.s32 %r87, %r70, %r37;
cvt.u64.u32 %rd54, %r87;
add.s64 %rd7, %rd53, %rd54;
mov.f32 %f115, 0f00000000;
cvta.to.global.u64 %rd55, %rd15;
mov.f32 %f113, %f115;

$L__BB280_10:
cvt.u64.u32 %rd56, %r261;
add.s64 %rd57, %rd7, %rd56;
mul.lo.s64 %rd58, %rd57, 36;
add.s64 %rd59, %rd55, %rd58;
shl.b32 %r105, %r10, 1;
mul.wide.s32 %rd60, %r105, 4;
add.s64 %rd61, %rd59, %rd60;
ld.global.nc.u16 %rs2, [%rd67+-6];
cvt.u32.u16 %r106, %rs2;
ld.global.nc.u16 %rs3, [%rd67+-4];
mov.b32 %r107, {%rs2, %rs3};
and.b32 %r108, %r107, -2004318072;
shr.u32 %r109, %r108, 1;
or.b32 %r110, %r109, 839922192;
and.b32 %r111, %r106, 30583;
mov.u32 %r112, -152379953;
mov.u32 %r113, -1079142271;
prmt.b32 %r114, %r113, %r112, %r111;
mov.u32 %r115, 1901675829;
mov.u32 %r116, 639175937;
prmt.b32 %r117, %r116, %r115, %r111;
and.b32 %r118, %r110, 30292;
prmt.b32 %r119, %r114, %r117, %r118;
and.b16 %rs4, %rs3, 30583;
cvt.u32.u16 %r120, %rs4;
prmt.b32 %r121, %r113, %r112, %r120;
prmt.b32 %r122, %r116, %r115, %r120;
shr.u32 %r123, %r110, 16;
prmt.b32 %r124, %r121, %r122, %r123;
mov.u32 %r125, 25632;
prmt.b32 %r89, %r119, %r124, %r125;
mov.u32 %r126, 30001;
prmt.b32 %r93, %r119, %r124, %r126;
ld.global.nc.u32 %r12, [%rd61+4];
mov.u32 %r91, 0;

dp4a.s32.s32 %r88, %r89, %r12, %r91;

ld.global.nc.u32 %r13, [%rd61+20];

dp4a.s32.s32 %r92, %r93, %r13, %r88;

ld.global.nc.u16 %rs5, [%rd67+-2];
cvt.u32.u16 %r127, %rs5;
ld.global.nc.u16 %rs6, [%rd67];
mov.b32 %r128, {%rs5, %rs6};
and.b32 %r129, %r128, -2004318072;
shr.u32 %r130, %r129, 1;
or.b32 %r131, %r130, 839922192;
and.b32 %r132, %r127, 30583;
prmt.b32 %r133, %r113, %r112, %r132;
prmt.b32 %r134, %r116, %r115, %r132;
and.b32 %r135, %r131, 30292;
prmt.b32 %r136, %r133, %r134, %r135;
and.b16 %rs7, %rs6, 30583;
cvt.u32.u16 %r137, %rs7;
prmt.b32 %r138, %r113, %r112, %r137;
prmt.b32 %r139, %r116, %r115, %r137;
shr.u32 %r140, %r131, 16;
prmt.b32 %r141, %r138, %r139, %r140;
prmt.b32 %r97, %r136, %r141, %r125;
prmt.b32 %r101, %r136, %r141, %r126;
ld.global.nc.u32 %r14, [%rd61+8];

dp4a.s32.s32 %r96, %r97, %r14, %r92;

ld.global.nc.u32 %r15, [%rd61+24];

dp4a.s32.s32 %r100, %r101, %r15, %r96;

ld.global.nc.u16 %rs1, [%rd68+-4];

{ cvt.f32.f16 %f40, %rs1;}


ld.global.nc.u32 %r16, [%rd59];

{.reg .f16 low,high;
mov.b32 {low,high},%r16;
cvt.f32.f16 %f41, low;}


mul.ftz.f32 %f42, %f40, %f41;
cvt.rn.f32.s32 %f43, %r100;
fma.rn.ftz.f32 %f115, %f42, %f43, %f115;
setp.eq.s64 %p11, %rd18, 0;
@%p11 bra $L__BB280_12;

add.s64 %rd63, %rd69, %rd50;
ld.global.u16 %rs9, [%rd63+2];
cvt.u32.u16 %r160, %rs9;
ld.global.u16 %rs10, [%rd63+4];
mov.b32 %r161, {%rs9, %rs10};
and.b32 %r162, %r161, -2004318072;
shr.u32 %r163, %r162, 1;
or.b32 %r164, %r163, 839922192;
and.b32 %r165, %r160, 30583;
prmt.b32 %r168, %r113, %r112, %r165;
prmt.b32 %r171, %r116, %r115, %r165;
and.b32 %r172, %r164, 30292;
prmt.b32 %r173, %r168, %r171, %r172;
and.b16 %rs11, %rs10, 30583;
cvt.u32.u16 %r174, %rs11;
prmt.b32 %r175, %r113, %r112, %r174;
prmt.b32 %r176, %r116, %r115, %r174;
shr.u32 %r177, %r164, 16;
prmt.b32 %r178, %r175, %r176, %r177;
prmt.b32 %r143, %r173, %r178, %r125;
prmt.b32 %r147, %r173, %r178, %r126;

dp4a.s32.s32 %r142, %r143, %r12, %r91;


dp4a.s32.s32 %r146, %r147, %r13, %r142;

ld.global.u16 %rs12, [%rd63+6];
cvt.u32.u16 %r181, %rs12;
ld.global.u16 %rs13, [%rd63+8];
mov.b32 %r182, {%rs12, %rs13};
and.b32 %r183, %r182, -2004318072;
shr.u32 %r184, %r183, 1;
or.b32 %r185, %r184, 839922192;
and.b32 %r186, %r181, 30583;
prmt.b32 %r187, %r113, %r112, %r186;
prmt.b32 %r188, %r116, %r115, %r186;
and.b32 %r189, %r185, 30292;
prmt.b32 %r190, %r187, %r188, %r189;
and.b16 %rs14, %rs13, 30583;
cvt.u32.u16 %r191, %rs14;
prmt.b32 %r192, %r113, %r112, %r191;
prmt.b32 %r193, %r116, %r115, %r191;
shr.u32 %r194, %r185, 16;
prmt.b32 %r195, %r192, %r193, %r194;
prmt.b32 %r151, %r190, %r195, %r125;
prmt.b32 %r155, %r190, %r195, %r126;

dp4a.s32.s32 %r150, %r151, %r14, %r146;


dp4a.s32.s32 %r154, %r155, %r15, %r150;

ld.global.u16 %rs8, [%rd69];

{ cvt.f32.f16 %f44, %rs8;}



{.reg .f16 low,high;
mov.b32 {low,high},%r16;
cvt.f32.f16 %f45, low;}


mul.ftz.f32 %f46, %f44, %f45;
cvt.rn.f32.s32 %f47, %r154;
fma.rn.ftz.f32 %f113, %f46, %f47, %f113;

$L__BB280_12:
add.s64 %rd69, %rd69, 1152;
add.s64 %rd68, %rd68, 1152;
add.s64 %rd67, %rd67, 1152;
add.s32 %r261, %r261, 64;
setp.lt.s32 %p12, %r261, %r69;
@%p12 bra $L__BB280_10;

$L__BB280_13:
setp.eq.s32 %p13, %r46, 0;
@%p13 bra $L__BB280_16;

setp.eq.s64 %p14, %rd18, 0;
add.s32 %r18, %r46, -1;
shl.b32 %r200, %r46, 7;
mov.u32 %r201, _ZZ13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r202, %r201, %r200;
shl.b32 %r203, %r47, 2;
add.s32 %r204, %r202, %r203;
st.shared.f32 [%r204+-128], %f115;
@%p14 bra $L__BB280_16;

shl.b32 %r206, %r18, 7;
mov.u32 %r207, _ZZ13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r208, %r207, %r206;
add.s32 %r210, %r208, %r203;
st.shared.f32 [%r210], %f113;

$L__BB280_16:
setp.ne.s32 %p15, %r46, 0;
bar.sync 0;
@%p15 bra $L__BB280_35;

setp.ne.s64 %p16, %rd18, 0;
shl.b32 %r213, %r47, 2;
mov.u32 %r214, _ZZ13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r19, %r214, %r213;
ld.shared.f32 %f48, [%r19];
add.ftz.f32 %f13, %f115, %f48;
@%p16 bra $L__BB280_19;
bra.uni $L__BB280_18;

$L__BB280_19:
mov.u32 %r217, _ZZ13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r218, %r217, %r213;
ld.shared.f32 %f52, [%r218];
add.ftz.f32 %f53, %f113, %f52;
ld.shared.f32 %f54, [%r19+128];
add.ftz.f32 %f55, %f13, %f54;
ld.shared.f32 %f56, [%r218+128];
add.ftz.f32 %f57, %f53, %f56;
ld.shared.f32 %f58, [%r19+256];
add.ftz.f32 %f117, %f55, %f58;
ld.shared.f32 %f59, [%r218+256];
add.ftz.f32 %f113, %f57, %f59;
bra.uni $L__BB280_20;

$L__BB280_18:
ld.shared.f32 %f49, [%r19+128];
add.ftz.f32 %f50, %f13, %f49;
ld.shared.f32 %f51, [%r19+256];
add.ftz.f32 %f117, %f50, %f51;

$L__BB280_20:
mov.b32 %r219, %f117;
mov.u32 %r220, 31;
mov.u32 %r221, 16;
mov.u32 %r222, -1;
shfl.sync.bfly.b32 %r223|%p17, %r219, %r221, %r220, %r222;
mov.b32 %f60, %r223;
add.ftz.f32 %f61, %f117, %f60;
mov.b32 %r224, %f61;
mov.u32 %r225, 8;
shfl.sync.bfly.b32 %r226|%p18, %r224, %r225, %r220, %r222;
mov.b32 %f62, %r226;
add.ftz.f32 %f63, %f61, %f62;
mov.b32 %r227, %f63;
mov.u32 %r228, 4;
shfl.sync.bfly.b32 %r229|%p19, %r227, %r228, %r220, %r222;
mov.b32 %f64, %r229;
add.ftz.f32 %f65, %f63, %f64;
mov.b32 %r230, %f65;
mov.u32 %r231, 2;
shfl.sync.bfly.b32 %r232|%p20, %r230, %r231, %r220, %r222;
mov.b32 %f66, %r232;
add.ftz.f32 %f19, %f65, %f66;
mov.b32 %r233, %f19;
mov.u32 %r234, 1;
shfl.sync.bfly.b32 %r20|%p1, %r233, %r234, %r220, %r222;
setp.eq.s64 %p21, %rd18, 0;
@%p21 bra $L__BB280_22;

mov.b32 %r235, %f113;
shfl.sync.bfly.b32 %r239|%p22, %r235, %r221, %r220, %r222;
mov.b32 %f67, %r239;
add.ftz.f32 %f68, %f113, %f67;
mov.b32 %r240, %f68;
shfl.sync.bfly.b32 %r242|%p23, %r240, %r225, %r220, %r222;
mov.b32 %f69, %r242;
add.ftz.f32 %f70, %f68, %f69;
mov.b32 %r243, %f70;
shfl.sync.bfly.b32 %r245|%p24, %r243, %r228, %r220, %r222;
mov.b32 %f71, %r245;
add.ftz.f32 %f72, %f70, %f71;
mov.b32 %r246, %f72;
shfl.sync.bfly.b32 %r248|%p25, %r246, %r231, %r220, %r222;
mov.b32 %f73, %r248;
add.ftz.f32 %f74, %f72, %f73;
mov.b32 %r249, %f74;
shfl.sync.bfly.b32 %r251|%p26, %r249, %r234, %r220, %r222;
mov.b32 %f75, %r251;
add.ftz.f32 %f113, %f74, %f75;

$L__BB280_22:
setp.ne.s32 %p27, %r47, 0;
mov.b32 %f76, %r20;
add.ftz.f32 %f22, %f19, %f76;
@%p27 bra $L__BB280_35;

add.ftz.f32 %f119, %f110, %f22;
@%p21 bra $L__BB280_34;

add.ftz.f32 %f24, %f109, %f113;
setp.gt.s32 %p29, %r21, 2;
@%p29 bra $L__BB280_28;

setp.eq.s32 %p32, %r21, 1;
@%p32 bra $L__BB280_32;

setp.eq.s32 %p33, %r21, 2;
@%p33 bra $L__BB280_27;
bra.uni $L__BB280_33;

$L__BB280_27:
mul.ftz.f32 %f96, %f24, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f97, %f96;
add.ftz.f32 %f98, %f97, 0f3F800000;
div.approx.ftz.f32 %f99, %f24, %f98;
mul.ftz.f32 %f119, %f119, %f99;
bra.uni $L__BB280_34;

$L__BB280_28:
setp.eq.s32 %p30, %r21, 3;
@%p30 bra $L__BB280_31;

setp.ne.s32 %p31, %r21, 6;
@%p31 bra $L__BB280_33;

min.ftz.f32 %f77, %f24, %f31;
neg.ftz.f32 %f78, %f31;
min.ftz.f32 %f79, %f119, %f31;
max.ftz.f32 %f80, %f79, %f78;
mul.ftz.f32 %f81, %f77, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f82, %f81;
add.ftz.f32 %f83, %f82, 0f3F800000;
div.approx.ftz.f32 %f84, %f77, %f83;
mul.ftz.f32 %f119, %f80, %f84;
bra.uni $L__BB280_34;

$L__BB280_32:
mul.ftz.f32 %f100, %f24, 0f3F000000;
mul.ftz.f32 %f101, %f24, 0f3D372713;
fma.rn.ftz.f32 %f102, %f24, %f101, 0f3F800000;
mul.ftz.f32 %f103, %f24, 0f3F4C422A;
mul.ftz.f32 %f104, %f103, %f102;
tanh.approx.f32 %f105, %f104;
add.ftz.f32 %f106, %f105, 0f3F800000;
mul.ftz.f32 %f107, %f100, %f106;
mul.ftz.f32 %f119, %f119, %f107;
bra.uni $L__BB280_34;

$L__BB280_31:
mov.f32 %f85, 0f40E00000;
min.ftz.f32 %f86, %f24, %f85;
min.ftz.f32 %f87, %f119, %f85;
mov.f32 %f88, 0fC0E00000;
max.ftz.f32 %f89, %f87, %f88;
mul.ftz.f32 %f90, %f86, 0f3FD9DB23;
mul.ftz.f32 %f91, %f90, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f92, %f91;
add.ftz.f32 %f93, %f92, 0f3F800000;
div.approx.ftz.f32 %f94, %f86, %f93;
add.ftz.f32 %f95, %f89, 0f3F800000;
mul.ftz.f32 %f119, %f95, %f94;
bra.uni $L__BB280_34;

$L__BB280_33:
mul.ftz.f32 %f119, %f119, %f24;

$L__BB280_34:
cvta.to.global.u64 %rd64, %rd22;
mov.u32 %r254, %ctaid.z;
mov.u32 %r255, %ctaid.y;
mov.u32 %r256, %ctaid.x;
mad.lo.s32 %r257, %r255, %r32, %r256;
mad.lo.s32 %r258, %r254, %r38, %r257;
mul.wide.u32 %rd65, %r258, 4;
add.s64 %rd66, %rd64, %rd65;
st.global.f32 [%rd66], %f119;

$L__BB280_35:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<12>;
.reg .b16 %rs<8>;
.reg .f32 %f<28>;
.reg .b32 %r<143>;
.reg .b64 %rd<36>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[384];

ld.param.u64 %rd10, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd11, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd12, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd13, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r15, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r19, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r18, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r17, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r16, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
mov.u32 %r1, %tid.y;
shl.b32 %r32, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r32, %r2;
mov.u32 %r141, %ctaid.y;
setp.eq.s64 %p2, %rd12, 0;
@%p2 bra $L__BB281_2;

cvta.to.global.u64 %rd14, %rd12;
mul.wide.u32 %rd15, %r141, 4;
add.s64 %rd16, %rd14, %rd15;
ld.global.nc.u32 %r140, [%rd16];
mul.hi.u32 %r33, %r141, %r16;
add.s32 %r34, %r33, %r141;
shr.u32 %r35, %r34, %r17;
mul.lo.s32 %r36, %r35, %r18;
sub.s32 %r141, %r141, %r36;
bra.uni $L__BB281_3;

$L__BB281_2:
mul.hi.u32 %r37, %r141, %r20;
add.s32 %r38, %r37, %r141;
shr.u32 %r140, %r38, %r21;

$L__BB281_3:
shr.u32 %r39, %r3, 31;
add.s32 %r40, %r3, %r39;
shr.s32 %r142, %r40, 1;
shr.u32 %r41, %r15, 5;
setp.ge.s32 %p3, %r142, %r41;
mov.f32 %f27, 0f00000000;
@%p3 bra $L__BB281_6;

mov.u32 %r42, %ctaid.z;
and.b32 %r45, %r40, -2;
sub.s32 %r46, %r3, %r45;
shl.b32 %r47, %r46, 1;
cvt.s64.s32 %rd1, %r47;
shl.b32 %r48, %r46, 2;
mul.hi.u32 %r49, %r42, %r26;
add.s32 %r50, %r49, %r42;
shr.u32 %r51, %r50, %r27;
mad.lo.s32 %r52, %r51, %r29, %r142;
mad.lo.s32 %r53, %r140, %r23, %r52;
mov.u32 %r54, %ctaid.x;
mad.lo.s32 %r55, %r54, %r19, %r53;
mul.wide.s32 %rd17, %r55, 18;
cvta.to.global.u64 %rd18, %rd10;
add.s64 %rd19, %rd18, %rd17;
add.s64 %rd35, %rd19, 8;
mul.wide.s32 %rd20, %r48, 2;
add.s64 %rd21, %rd17, %rd20;
add.s64 %rd22, %rd18, %rd21;
add.s64 %rd34, %rd22, 4;
cvta.to.global.u64 %rd4, %rd11;
mul.lo.s32 %r56, %r141, %r24;
cvt.u64.u32 %rd23, %r56;
mul.lo.s32 %r57, %r42, %r30;
cvt.u64.u32 %rd24, %r57;
add.s64 %rd5, %rd23, %rd24;
mov.f32 %f27, 0f00000000;
shl.b64 %rd29, %rd1, 2;

$L__BB281_5:
cvt.u64.u32 %rd25, %r142;
add.s64 %rd26, %rd5, %rd25;
mul.lo.s64 %rd27, %rd26, 36;
add.s64 %rd28, %rd4, %rd27;
add.s64 %rd30, %rd28, %rd29;
ld.global.nc.u16 %rs2, [%rd34+-2];
cvt.u32.u16 %r75, %rs2;
ld.global.nc.u16 %rs3, [%rd34];
mov.b32 %r76, {%rs2, %rs3};
and.b32 %r77, %r76, -2004318072;
shr.u32 %r78, %r77, 1;
or.b32 %r79, %r78, 839922192;
and.b32 %r80, %r75, 30583;
mov.u32 %r81, -152379953;
mov.u32 %r82, -1079142271;
prmt.b32 %r83, %r82, %r81, %r80;
mov.u32 %r84, 1901675829;
mov.u32 %r85, 639175937;
prmt.b32 %r86, %r85, %r84, %r80;
and.b32 %r87, %r79, 30292;
prmt.b32 %r88, %r83, %r86, %r87;
and.b16 %rs4, %rs3, 30583;
cvt.u32.u16 %r89, %rs4;
prmt.b32 %r90, %r82, %r81, %r89;
prmt.b32 %r91, %r85, %r84, %r89;
shr.u32 %r92, %r79, 16;
prmt.b32 %r93, %r90, %r91, %r92;
mov.u32 %r94, 25632;
prmt.b32 %r59, %r88, %r93, %r94;
mov.u32 %r95, 30001;
prmt.b32 %r63, %r88, %r93, %r95;
ld.global.nc.u32 %r60, [%rd30+4];
mov.u32 %r61, 0;

dp4a.s32.s32 %r58, %r59, %r60, %r61;

ld.global.nc.u32 %r64, [%rd30+20];

dp4a.s32.s32 %r62, %r63, %r64, %r58;

ld.global.nc.u16 %rs5, [%rd34+2];
cvt.u32.u16 %r96, %rs5;
ld.global.nc.u16 %rs6, [%rd34+4];
mov.b32 %r97, {%rs5, %rs6};
and.b32 %r98, %r97, -2004318072;
shr.u32 %r99, %r98, 1;
or.b32 %r100, %r99, 839922192;
and.b32 %r101, %r96, 30583;
prmt.b32 %r102, %r82, %r81, %r101;
prmt.b32 %r103, %r85, %r84, %r101;
and.b32 %r104, %r100, 30292;
prmt.b32 %r105, %r102, %r103, %r104;
and.b16 %rs7, %rs6, 30583;
cvt.u32.u16 %r106, %rs7;
prmt.b32 %r107, %r82, %r81, %r106;
prmt.b32 %r108, %r85, %r84, %r106;
shr.u32 %r109, %r100, 16;
prmt.b32 %r110, %r107, %r108, %r109;
prmt.b32 %r67, %r105, %r110, %r94;
prmt.b32 %r71, %r105, %r110, %r95;
ld.global.nc.u32 %r68, [%rd30+8];

dp4a.s32.s32 %r66, %r67, %r68, %r62;

ld.global.nc.u32 %r72, [%rd30+24];

dp4a.s32.s32 %r70, %r71, %r72, %r66;

ld.global.nc.u16 %rs1, [%rd35+-8];

{ cvt.f32.f16 %f7, %rs1;}


ld.global.nc.u32 %r74, [%rd28];

{.reg .f16 low,high;
mov.b32 {low,high},%r74;
cvt.f32.f16 %f8, low;}


mul.ftz.f32 %f9, %f7, %f8;
cvt.rn.f32.s32 %f10, %r70;
fma.rn.ftz.f32 %f27, %f9, %f10, %f27;
add.s64 %rd35, %rd35, 1152;
add.s64 %rd34, %rd34, 1152;
add.s32 %r142, %r142, 64;
setp.lt.s32 %p4, %r142, %r41;
@%p4 bra $L__BB281_5;

$L__BB281_6:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB281_8;

shl.b32 %r111, %r1, 7;
mov.u32 %r112, _ZZ13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r113, %r112, %r111;
shl.b32 %r114, %r2, 2;
add.s32 %r115, %r113, %r114;
st.shared.f32 [%r115+-128], %f27;

$L__BB281_8:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB281_11;

shl.b32 %r116, %r2, 2;
mov.u32 %r117, _ZZ13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r118, %r117, %r116;
mov.u32 %r119, 2;
ld.shared.f32 %f11, [%r118];
add.ftz.f32 %f12, %f27, %f11;
ld.shared.f32 %f13, [%r118+128];
add.ftz.f32 %f14, %f12, %f13;
ld.shared.f32 %f15, [%r118+256];
add.ftz.f32 %f16, %f14, %f15;
mov.b32 %r120, %f16;
mov.u32 %r121, 31;
mov.u32 %r122, 16;
mov.u32 %r123, -1;
shfl.sync.bfly.b32 %r124|%p7, %r120, %r122, %r121, %r123;
mov.b32 %f17, %r124;
add.ftz.f32 %f18, %f16, %f17;
mov.b32 %r125, %f18;
mov.u32 %r126, 8;
shfl.sync.bfly.b32 %r127|%p8, %r125, %r126, %r121, %r123;
mov.b32 %f19, %r127;
add.ftz.f32 %f20, %f18, %f19;
mov.b32 %r128, %f20;
mov.u32 %r129, 4;
shfl.sync.bfly.b32 %r130|%p9, %r128, %r129, %r121, %r123;
mov.b32 %f21, %r130;
add.ftz.f32 %f22, %f20, %f21;
mov.b32 %r131, %f22;
shfl.sync.bfly.b32 %r132|%p10, %r131, %r119, %r121, %r123;
mov.b32 %f23, %r132;
add.ftz.f32 %f4, %f22, %f23;
mov.b32 %r133, %f4;
mov.u32 %r134, 1;
shfl.sync.bfly.b32 %r14|%p1, %r133, %r134, %r121, %r123;
setp.ne.s32 %p11, %r2, 0;
@%p11 bra $L__BB281_11;

mov.u32 %r135, %ctaid.x;
mov.u32 %r136, %ctaid.y;
mad.lo.s32 %r137, %r136, %r25, %r135;
mov.u32 %r138, %ctaid.z;
mad.lo.s32 %r139, %r138, %r31, %r137;
cvta.to.global.u64 %rd31, %rd13;
mul.wide.u32 %rd32, %r139, 4;
add.s64 %rd33, %rd31, %rd32;
mov.b32 %f24, %r14;
add.ftz.f32 %f25, %f4, %f24;
st.global.f32 [%rd33], %f25;

$L__BB281_11:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<34>;
.reg .b16 %rs<15>;
.reg .f32 %f<120>;
.reg .b32 %r<262>;
.reg .b64 %rd<70>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[384];

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate[384];

ld.param.u64 %rd14, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd22, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
ld.param.f32 %f31, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+16];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+8];
ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3];
setp.eq.s64 %p2, %rd16, 0;
@%p2 bra $L__BB282_2;

cvta.to.global.u64 %rd23, %rd16;
mov.u32 %r39, %ctaid.y;
mul.wide.u32 %rd24, %r39, 4;
add.s64 %rd25, %rd23, %rd24;
ld.global.nc.u32 %r259, [%rd25];
mul.hi.u32 %r40, %r39, %r23;
add.s32 %r41, %r40, %r39;
shr.u32 %r42, %r41, %r24;
mul.lo.s32 %r43, %r42, %r25;
sub.s32 %r260, %r39, %r43;
bra.uni $L__BB282_3;

$L__BB282_2:
mov.u32 %r260, %ctaid.y;
mul.hi.u32 %r44, %r260, %r27;
add.s32 %r45, %r44, %r260;
shr.u32 %r259, %r45, %r28;

$L__BB282_3:
mov.u32 %r46, %tid.y;
mov.u32 %r47, %tid.x;
or.b32 %r48, %r47, %r46;
setp.ne.s32 %p3, %r48, 0;
mov.f32 %f113, 0f00000000;
mov.f32 %f109, %f113;
mov.f32 %f110, %f113;
@%p3 bra $L__BB282_8;

setp.eq.s64 %p4, %rd17, 0;
mov.f32 %f109, 0f00000000;
mov.f32 %f110, %f109;
@%p4 bra $L__BB282_6;

cvta.to.global.u64 %rd26, %rd17;
mov.u32 %r49, %ctaid.z;
mul.lo.s32 %r50, %r49, %r38;
cvt.u64.u32 %rd27, %r50;
mov.u32 %r51, %ctaid.y;
selp.b32 %r52, %r51, %r259, %p2;
mul.lo.s32 %r53, %r52, %r32;
cvt.u64.u32 %rd28, %r53;
mov.u32 %r54, %ctaid.x;
cvt.s64.s32 %rd29, %r54;
cvt.u64.u32 %rd30, %r47;
add.s64 %rd31, %rd30, %rd29;
add.s64 %rd32, %rd31, %rd27;
add.s64 %rd33, %rd32, %rd28;
shl.b64 %rd34, %rd33, 2;
add.s64 %rd35, %rd26, %rd34;
ld.global.f32 %f110, [%rd35];

$L__BB282_6:
setp.eq.s64 %p6, %rd19, 0;
setp.eq.s64 %p7, %rd18, 0;
or.pred %p8, %p6, %p7;
@%p8 bra $L__BB282_8;

cvta.to.global.u64 %rd36, %rd19;
mov.u32 %r56, %ctaid.z;
mul.lo.s32 %r57, %r56, %r38;
cvt.u64.u32 %rd37, %r57;
mov.u32 %r58, %ctaid.y;
selp.b32 %r59, %r58, %r259, %p2;
mul.lo.s32 %r60, %r59, %r32;
cvt.u64.u32 %rd38, %r60;
mov.u32 %r61, %ctaid.x;
cvt.s64.s32 %rd39, %r61;
cvt.u64.u32 %rd40, %r47;
add.s64 %rd41, %rd40, %rd39;
add.s64 %rd42, %rd41, %rd37;
add.s64 %rd43, %rd42, %rd38;
shl.b64 %rd44, %rd43, 2;
add.s64 %rd45, %rd36, %rd44;
ld.global.f32 %f109, [%rd45];

$L__BB282_8:
shl.b32 %r64, %r46, 5;
add.s32 %r66, %r64, %r47;
shr.u32 %r67, %r66, 31;
add.s32 %r68, %r66, %r67;
shr.s32 %r261, %r68, 1;
shr.u32 %r69, %r22, 5;
setp.ge.s32 %p10, %r261, %r69;
mov.f32 %f115, %f113;
@%p10 bra $L__BB282_13;

mov.u32 %r70, %ctaid.z;
and.b32 %r77, %r68, -2;
sub.s32 %r10, %r66, %r77;
shl.b32 %r78, %r10, 2;
mul.hi.u32 %r79, %r70, %r33;
add.s32 %r80, %r79, %r70;
shr.u32 %r81, %r80, %r34;
mad.lo.s32 %r82, %r81, %r36, %r261;
mad.lo.s32 %r83, %r259, %r30, %r82;
mov.u32 %r84, %ctaid.x;
mad.lo.s32 %r85, %r84, %r26, %r83;
mul.wide.s32 %rd46, %r85, 18;
cvta.to.global.u64 %rd47, %rd18;
add.s64 %rd69, %rd47, %rd46;
cvta.to.global.u64 %rd48, %rd14;
add.s64 %rd49, %rd48, %rd46;
add.s64 %rd68, %rd49, 4;
mul.wide.s32 %rd50, %r78, 2;
add.s64 %rd51, %rd46, %rd50;
add.s64 %rd52, %rd48, %rd51;
add.s64 %rd67, %rd52, 8;
mul.lo.s32 %r86, %r260, %r31;
cvt.u64.u32 %rd53, %r86;
mul.lo.s32 %r87, %r70, %r37;
cvt.u64.u32 %rd54, %r87;
add.s64 %rd7, %rd53, %rd54;
mov.f32 %f115, 0f00000000;
cvta.to.global.u64 %rd55, %rd15;
mov.f32 %f113, %f115;

$L__BB282_10:
cvt.u64.u32 %rd56, %r261;
add.s64 %rd57, %rd7, %rd56;
mul.lo.s64 %rd58, %rd57, 36;
add.s64 %rd59, %rd55, %rd58;
shl.b32 %r105, %r10, 1;
mul.wide.s32 %rd60, %r105, 4;
add.s64 %rd61, %rd59, %rd60;
ld.global.nc.u16 %rs2, [%rd67+-6];
cvt.u32.u16 %r106, %rs2;
ld.global.nc.u16 %rs3, [%rd67+-4];
mov.b32 %r107, {%rs2, %rs3};
and.b32 %r108, %r107, -2004318072;
shr.u32 %r109, %r108, 1;
or.b32 %r110, %r109, 839922192;
and.b32 %r111, %r106, 30583;
mov.u32 %r112, -152379953;
mov.u32 %r113, -1079142271;
prmt.b32 %r114, %r113, %r112, %r111;
mov.u32 %r115, 1901675829;
mov.u32 %r116, 639175937;
prmt.b32 %r117, %r116, %r115, %r111;
and.b32 %r118, %r110, 30292;
prmt.b32 %r119, %r114, %r117, %r118;
and.b16 %rs4, %rs3, 30583;
cvt.u32.u16 %r120, %rs4;
prmt.b32 %r121, %r113, %r112, %r120;
prmt.b32 %r122, %r116, %r115, %r120;
shr.u32 %r123, %r110, 16;
prmt.b32 %r124, %r121, %r122, %r123;
mov.u32 %r125, 25632;
prmt.b32 %r89, %r119, %r124, %r125;
mov.u32 %r126, 30001;
prmt.b32 %r93, %r119, %r124, %r126;
ld.global.nc.u32 %r12, [%rd61+4];
mov.u32 %r91, 0;

dp4a.s32.s32 %r88, %r89, %r12, %r91;

ld.global.nc.u32 %r13, [%rd61+20];

dp4a.s32.s32 %r92, %r93, %r13, %r88;

ld.global.nc.u16 %rs5, [%rd67+-2];
cvt.u32.u16 %r127, %rs5;
ld.global.nc.u16 %rs6, [%rd67];
mov.b32 %r128, {%rs5, %rs6};
and.b32 %r129, %r128, -2004318072;
shr.u32 %r130, %r129, 1;
or.b32 %r131, %r130, 839922192;
and.b32 %r132, %r127, 30583;
prmt.b32 %r133, %r113, %r112, %r132;
prmt.b32 %r134, %r116, %r115, %r132;
and.b32 %r135, %r131, 30292;
prmt.b32 %r136, %r133, %r134, %r135;
and.b16 %rs7, %rs6, 30583;
cvt.u32.u16 %r137, %rs7;
prmt.b32 %r138, %r113, %r112, %r137;
prmt.b32 %r139, %r116, %r115, %r137;
shr.u32 %r140, %r131, 16;
prmt.b32 %r141, %r138, %r139, %r140;
prmt.b32 %r97, %r136, %r141, %r125;
prmt.b32 %r101, %r136, %r141, %r126;
ld.global.nc.u32 %r14, [%rd61+8];

dp4a.s32.s32 %r96, %r97, %r14, %r92;

ld.global.nc.u32 %r15, [%rd61+24];

dp4a.s32.s32 %r100, %r101, %r15, %r96;

ld.global.nc.u16 %rs1, [%rd68+-4];

{ cvt.f32.f16 %f40, %rs1;}


ld.global.nc.u32 %r16, [%rd59];

{.reg .f16 low,high;
mov.b32 {low,high},%r16;
cvt.f32.f16 %f41, low;}


mul.ftz.f32 %f42, %f40, %f41;
cvt.rn.f32.s32 %f43, %r100;
fma.rn.ftz.f32 %f115, %f42, %f43, %f115;
setp.eq.s64 %p11, %rd18, 0;
@%p11 bra $L__BB282_12;

add.s64 %rd63, %rd69, %rd50;
ld.global.u16 %rs9, [%rd63+2];
cvt.u32.u16 %r160, %rs9;
ld.global.u16 %rs10, [%rd63+4];
mov.b32 %r161, {%rs9, %rs10};
and.b32 %r162, %r161, -2004318072;
shr.u32 %r163, %r162, 1;
or.b32 %r164, %r163, 839922192;
and.b32 %r165, %r160, 30583;
prmt.b32 %r168, %r113, %r112, %r165;
prmt.b32 %r171, %r116, %r115, %r165;
and.b32 %r172, %r164, 30292;
prmt.b32 %r173, %r168, %r171, %r172;
and.b16 %rs11, %rs10, 30583;
cvt.u32.u16 %r174, %rs11;
prmt.b32 %r175, %r113, %r112, %r174;
prmt.b32 %r176, %r116, %r115, %r174;
shr.u32 %r177, %r164, 16;
prmt.b32 %r178, %r175, %r176, %r177;
prmt.b32 %r143, %r173, %r178, %r125;
prmt.b32 %r147, %r173, %r178, %r126;

dp4a.s32.s32 %r142, %r143, %r12, %r91;


dp4a.s32.s32 %r146, %r147, %r13, %r142;

ld.global.u16 %rs12, [%rd63+6];
cvt.u32.u16 %r181, %rs12;
ld.global.u16 %rs13, [%rd63+8];
mov.b32 %r182, {%rs12, %rs13};
and.b32 %r183, %r182, -2004318072;
shr.u32 %r184, %r183, 1;
or.b32 %r185, %r184, 839922192;
and.b32 %r186, %r181, 30583;
prmt.b32 %r187, %r113, %r112, %r186;
prmt.b32 %r188, %r116, %r115, %r186;
and.b32 %r189, %r185, 30292;
prmt.b32 %r190, %r187, %r188, %r189;
and.b16 %rs14, %rs13, 30583;
cvt.u32.u16 %r191, %rs14;
prmt.b32 %r192, %r113, %r112, %r191;
prmt.b32 %r193, %r116, %r115, %r191;
shr.u32 %r194, %r185, 16;
prmt.b32 %r195, %r192, %r193, %r194;
prmt.b32 %r151, %r190, %r195, %r125;
prmt.b32 %r155, %r190, %r195, %r126;

dp4a.s32.s32 %r150, %r151, %r14, %r146;


dp4a.s32.s32 %r154, %r155, %r15, %r150;

ld.global.u16 %rs8, [%rd69];

{ cvt.f32.f16 %f44, %rs8;}



{.reg .f16 low,high;
mov.b32 {low,high},%r16;
cvt.f32.f16 %f45, low;}


mul.ftz.f32 %f46, %f44, %f45;
cvt.rn.f32.s32 %f47, %r154;
fma.rn.ftz.f32 %f113, %f46, %f47, %f113;

$L__BB282_12:
add.s64 %rd69, %rd69, 1152;
add.s64 %rd68, %rd68, 1152;
add.s64 %rd67, %rd67, 1152;
add.s32 %r261, %r261, 64;
setp.lt.s32 %p12, %r261, %r69;
@%p12 bra $L__BB282_10;

$L__BB282_13:
setp.eq.s32 %p13, %r46, 0;
@%p13 bra $L__BB282_16;

setp.eq.s64 %p14, %rd18, 0;
add.s32 %r18, %r46, -1;
shl.b32 %r200, %r46, 7;
mov.u32 %r201, _ZZ13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r202, %r201, %r200;
shl.b32 %r203, %r47, 2;
add.s32 %r204, %r202, %r203;
st.shared.f32 [%r204+-128], %f115;
@%p14 bra $L__BB282_16;

shl.b32 %r206, %r18, 7;
mov.u32 %r207, _ZZ13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r208, %r207, %r206;
add.s32 %r210, %r208, %r203;
st.shared.f32 [%r210], %f113;

$L__BB282_16:
setp.ne.s32 %p15, %r46, 0;
bar.sync 0;
@%p15 bra $L__BB282_35;

setp.ne.s64 %p16, %rd18, 0;
shl.b32 %r213, %r47, 2;
mov.u32 %r214, _ZZ13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r19, %r214, %r213;
ld.shared.f32 %f48, [%r19];
add.ftz.f32 %f13, %f115, %f48;
@%p16 bra $L__BB282_19;
bra.uni $L__BB282_18;

$L__BB282_19:
mov.u32 %r217, _ZZ13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r218, %r217, %r213;
ld.shared.f32 %f52, [%r218];
add.ftz.f32 %f53, %f113, %f52;
ld.shared.f32 %f54, [%r19+128];
add.ftz.f32 %f55, %f13, %f54;
ld.shared.f32 %f56, [%r218+128];
add.ftz.f32 %f57, %f53, %f56;
ld.shared.f32 %f58, [%r19+256];
add.ftz.f32 %f117, %f55, %f58;
ld.shared.f32 %f59, [%r218+256];
add.ftz.f32 %f113, %f57, %f59;
bra.uni $L__BB282_20;

$L__BB282_18:
ld.shared.f32 %f49, [%r19+128];
add.ftz.f32 %f50, %f13, %f49;
ld.shared.f32 %f51, [%r19+256];
add.ftz.f32 %f117, %f50, %f51;

$L__BB282_20:
mov.b32 %r219, %f117;
mov.u32 %r220, 31;
mov.u32 %r221, 16;
mov.u32 %r222, -1;
shfl.sync.bfly.b32 %r223|%p17, %r219, %r221, %r220, %r222;
mov.b32 %f60, %r223;
add.ftz.f32 %f61, %f117, %f60;
mov.b32 %r224, %f61;
mov.u32 %r225, 8;
shfl.sync.bfly.b32 %r226|%p18, %r224, %r225, %r220, %r222;
mov.b32 %f62, %r226;
add.ftz.f32 %f63, %f61, %f62;
mov.b32 %r227, %f63;
mov.u32 %r228, 4;
shfl.sync.bfly.b32 %r229|%p19, %r227, %r228, %r220, %r222;
mov.b32 %f64, %r229;
add.ftz.f32 %f65, %f63, %f64;
mov.b32 %r230, %f65;
mov.u32 %r231, 2;
shfl.sync.bfly.b32 %r232|%p20, %r230, %r231, %r220, %r222;
mov.b32 %f66, %r232;
add.ftz.f32 %f19, %f65, %f66;
mov.b32 %r233, %f19;
mov.u32 %r234, 1;
shfl.sync.bfly.b32 %r20|%p1, %r233, %r234, %r220, %r222;
setp.eq.s64 %p21, %rd18, 0;
@%p21 bra $L__BB282_22;

mov.b32 %r235, %f113;
shfl.sync.bfly.b32 %r239|%p22, %r235, %r221, %r220, %r222;
mov.b32 %f67, %r239;
add.ftz.f32 %f68, %f113, %f67;
mov.b32 %r240, %f68;
shfl.sync.bfly.b32 %r242|%p23, %r240, %r225, %r220, %r222;
mov.b32 %f69, %r242;
add.ftz.f32 %f70, %f68, %f69;
mov.b32 %r243, %f70;
shfl.sync.bfly.b32 %r245|%p24, %r243, %r228, %r220, %r222;
mov.b32 %f71, %r245;
add.ftz.f32 %f72, %f70, %f71;
mov.b32 %r246, %f72;
shfl.sync.bfly.b32 %r248|%p25, %r246, %r231, %r220, %r222;
mov.b32 %f73, %r248;
add.ftz.f32 %f74, %f72, %f73;
mov.b32 %r249, %f74;
shfl.sync.bfly.b32 %r251|%p26, %r249, %r234, %r220, %r222;
mov.b32 %f75, %r251;
add.ftz.f32 %f113, %f74, %f75;

$L__BB282_22:
setp.ne.s32 %p27, %r47, 0;
mov.b32 %f76, %r20;
add.ftz.f32 %f22, %f19, %f76;
@%p27 bra $L__BB282_35;

add.ftz.f32 %f119, %f110, %f22;
@%p21 bra $L__BB282_34;

add.ftz.f32 %f24, %f109, %f113;
setp.gt.s32 %p29, %r21, 2;
@%p29 bra $L__BB282_28;

setp.eq.s32 %p32, %r21, 1;
@%p32 bra $L__BB282_32;

setp.eq.s32 %p33, %r21, 2;
@%p33 bra $L__BB282_27;
bra.uni $L__BB282_33;

$L__BB282_27:
mul.ftz.f32 %f96, %f24, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f97, %f96;
add.ftz.f32 %f98, %f97, 0f3F800000;
div.approx.ftz.f32 %f99, %f24, %f98;
mul.ftz.f32 %f119, %f119, %f99;
bra.uni $L__BB282_34;

$L__BB282_28:
setp.eq.s32 %p30, %r21, 3;
@%p30 bra $L__BB282_31;

setp.ne.s32 %p31, %r21, 6;
@%p31 bra $L__BB282_33;

min.ftz.f32 %f77, %f24, %f31;
neg.ftz.f32 %f78, %f31;
min.ftz.f32 %f79, %f119, %f31;
max.ftz.f32 %f80, %f79, %f78;
mul.ftz.f32 %f81, %f77, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f82, %f81;
add.ftz.f32 %f83, %f82, 0f3F800000;
div.approx.ftz.f32 %f84, %f77, %f83;
mul.ftz.f32 %f119, %f80, %f84;
bra.uni $L__BB282_34;

$L__BB282_32:
mul.ftz.f32 %f100, %f24, 0f3F000000;
mul.ftz.f32 %f101, %f24, 0f3D372713;
fma.rn.ftz.f32 %f102, %f24, %f101, 0f3F800000;
mul.ftz.f32 %f103, %f24, 0f3F4C422A;
mul.ftz.f32 %f104, %f103, %f102;
tanh.approx.f32 %f105, %f104;
add.ftz.f32 %f106, %f105, 0f3F800000;
mul.ftz.f32 %f107, %f100, %f106;
mul.ftz.f32 %f119, %f119, %f107;
bra.uni $L__BB282_34;

$L__BB282_31:
mov.f32 %f85, 0f40E00000;
min.ftz.f32 %f86, %f24, %f85;
min.ftz.f32 %f87, %f119, %f85;
mov.f32 %f88, 0fC0E00000;
max.ftz.f32 %f89, %f87, %f88;
mul.ftz.f32 %f90, %f86, 0f3FD9DB23;
mul.ftz.f32 %f91, %f90, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f92, %f91;
add.ftz.f32 %f93, %f92, 0f3F800000;
div.approx.ftz.f32 %f94, %f86, %f93;
add.ftz.f32 %f95, %f89, 0f3F800000;
mul.ftz.f32 %f119, %f95, %f94;
bra.uni $L__BB282_34;

$L__BB282_33:
mul.ftz.f32 %f119, %f119, %f24;

$L__BB282_34:
cvta.to.global.u64 %rd64, %rd22;
mov.u32 %r254, %ctaid.z;
mov.u32 %r255, %ctaid.y;
mov.u32 %r256, %ctaid.x;
mad.lo.s32 %r257, %r255, %r32, %r256;
mad.lo.s32 %r258, %r254, %r38, %r257;
mul.wide.u32 %rd65, %r258, 4;
add.s64 %rd66, %rd64, %rd65;
st.global.f32 [%rd66], %f119;

$L__BB282_35:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<12>;
.reg .b16 %rs<8>;
.reg .f32 %f<28>;
.reg .b32 %r<143>;
.reg .b64 %rd<36>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[384];

ld.param.u64 %rd10, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd11, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd12, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd13, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r15, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r19, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r18, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r17, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r16, [_Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
mov.u32 %r1, %tid.y;
shl.b32 %r32, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r32, %r2;
mov.u32 %r141, %ctaid.y;
setp.eq.s64 %p2, %rd12, 0;
@%p2 bra $L__BB283_2;

cvta.to.global.u64 %rd14, %rd12;
mul.wide.u32 %rd15, %r141, 4;
add.s64 %rd16, %rd14, %rd15;
ld.global.nc.u32 %r140, [%rd16];
mul.hi.u32 %r33, %r141, %r16;
add.s32 %r34, %r33, %r141;
shr.u32 %r35, %r34, %r17;
mul.lo.s32 %r36, %r35, %r18;
sub.s32 %r141, %r141, %r36;
bra.uni $L__BB283_3;

$L__BB283_2:
mul.hi.u32 %r37, %r141, %r20;
add.s32 %r38, %r37, %r141;
shr.u32 %r140, %r38, %r21;

$L__BB283_3:
shr.u32 %r39, %r3, 31;
add.s32 %r40, %r3, %r39;
shr.s32 %r142, %r40, 1;
shr.u32 %r41, %r15, 5;
setp.ge.s32 %p3, %r142, %r41;
mov.f32 %f27, 0f00000000;
@%p3 bra $L__BB283_6;

mov.u32 %r42, %ctaid.z;
and.b32 %r45, %r40, -2;
sub.s32 %r46, %r3, %r45;
shl.b32 %r47, %r46, 1;
cvt.s64.s32 %rd1, %r47;
shl.b32 %r48, %r46, 2;
mul.hi.u32 %r49, %r42, %r26;
add.s32 %r50, %r49, %r42;
shr.u32 %r51, %r50, %r27;
mad.lo.s32 %r52, %r51, %r29, %r142;
mad.lo.s32 %r53, %r140, %r23, %r52;
mov.u32 %r54, %ctaid.x;
mad.lo.s32 %r55, %r54, %r19, %r53;
mul.wide.s32 %rd17, %r55, 18;
cvta.to.global.u64 %rd18, %rd10;
add.s64 %rd19, %rd18, %rd17;
add.s64 %rd35, %rd19, 8;
mul.wide.s32 %rd20, %r48, 2;
add.s64 %rd21, %rd17, %rd20;
add.s64 %rd22, %rd18, %rd21;
add.s64 %rd34, %rd22, 4;
cvta.to.global.u64 %rd4, %rd11;
mul.lo.s32 %r56, %r141, %r24;
cvt.u64.u32 %rd23, %r56;
mul.lo.s32 %r57, %r42, %r30;
cvt.u64.u32 %rd24, %r57;
add.s64 %rd5, %rd23, %rd24;
mov.f32 %f27, 0f00000000;
shl.b64 %rd29, %rd1, 2;

$L__BB283_5:
cvt.u64.u32 %rd25, %r142;
add.s64 %rd26, %rd5, %rd25;
mul.lo.s64 %rd27, %rd26, 36;
add.s64 %rd28, %rd4, %rd27;
add.s64 %rd30, %rd28, %rd29;
ld.global.nc.u16 %rs2, [%rd34+-2];
cvt.u32.u16 %r75, %rs2;
ld.global.nc.u16 %rs3, [%rd34];
mov.b32 %r76, {%rs2, %rs3};
and.b32 %r77, %r76, -2004318072;
shr.u32 %r78, %r77, 1;
or.b32 %r79, %r78, 839922192;
and.b32 %r80, %r75, 30583;
mov.u32 %r81, -152379953;
mov.u32 %r82, -1079142271;
prmt.b32 %r83, %r82, %r81, %r80;
mov.u32 %r84, 1901675829;
mov.u32 %r85, 639175937;
prmt.b32 %r86, %r85, %r84, %r80;
and.b32 %r87, %r79, 30292;
prmt.b32 %r88, %r83, %r86, %r87;
and.b16 %rs4, %rs3, 30583;
cvt.u32.u16 %r89, %rs4;
prmt.b32 %r90, %r82, %r81, %r89;
prmt.b32 %r91, %r85, %r84, %r89;
shr.u32 %r92, %r79, 16;
prmt.b32 %r93, %r90, %r91, %r92;
mov.u32 %r94, 25632;
prmt.b32 %r59, %r88, %r93, %r94;
mov.u32 %r95, 30001;
prmt.b32 %r63, %r88, %r93, %r95;
ld.global.nc.u32 %r60, [%rd30+4];
mov.u32 %r61, 0;

dp4a.s32.s32 %r58, %r59, %r60, %r61;

ld.global.nc.u32 %r64, [%rd30+20];

dp4a.s32.s32 %r62, %r63, %r64, %r58;

ld.global.nc.u16 %rs5, [%rd34+2];
cvt.u32.u16 %r96, %rs5;
ld.global.nc.u16 %rs6, [%rd34+4];
mov.b32 %r97, {%rs5, %rs6};
and.b32 %r98, %r97, -2004318072;
shr.u32 %r99, %r98, 1;
or.b32 %r100, %r99, 839922192;
and.b32 %r101, %r96, 30583;
prmt.b32 %r102, %r82, %r81, %r101;
prmt.b32 %r103, %r85, %r84, %r101;
and.b32 %r104, %r100, 30292;
prmt.b32 %r105, %r102, %r103, %r104;
and.b16 %rs7, %rs6, 30583;
cvt.u32.u16 %r106, %rs7;
prmt.b32 %r107, %r82, %r81, %r106;
prmt.b32 %r108, %r85, %r84, %r106;
shr.u32 %r109, %r100, 16;
prmt.b32 %r110, %r107, %r108, %r109;
prmt.b32 %r67, %r105, %r110, %r94;
prmt.b32 %r71, %r105, %r110, %r95;
ld.global.nc.u32 %r68, [%rd30+8];

dp4a.s32.s32 %r66, %r67, %r68, %r62;

ld.global.nc.u32 %r72, [%rd30+24];

dp4a.s32.s32 %r70, %r71, %r72, %r66;

ld.global.nc.u16 %rs1, [%rd35+-8];

{ cvt.f32.f16 %f7, %rs1;}


ld.global.nc.u32 %r74, [%rd28];

{.reg .f16 low,high;
mov.b32 {low,high},%r74;
cvt.f32.f16 %f8, low;}


mul.ftz.f32 %f9, %f7, %f8;
cvt.rn.f32.s32 %f10, %r70;
fma.rn.ftz.f32 %f27, %f9, %f10, %f27;
add.s64 %rd35, %rd35, 1152;
add.s64 %rd34, %rd34, 1152;
add.s32 %r142, %r142, 64;
setp.lt.s32 %p4, %r142, %r41;
@%p4 bra $L__BB283_5;

$L__BB283_6:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB283_8;

shl.b32 %r111, %r1, 7;
mov.u32 %r112, _ZZ13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r113, %r112, %r111;
shl.b32 %r114, %r2, 2;
add.s32 %r115, %r113, %r114;
st.shared.f32 [%r115+-128], %f27;

$L__BB283_8:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB283_11;

shl.b32 %r116, %r2, 2;
mov.u32 %r117, _ZZ13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r118, %r117, %r116;
mov.u32 %r119, 2;
ld.shared.f32 %f11, [%r118];
add.ftz.f32 %f12, %f27, %f11;
ld.shared.f32 %f13, [%r118+128];
add.ftz.f32 %f14, %f12, %f13;
ld.shared.f32 %f15, [%r118+256];
add.ftz.f32 %f16, %f14, %f15;
mov.b32 %r120, %f16;
mov.u32 %r121, 31;
mov.u32 %r122, 16;
mov.u32 %r123, -1;
shfl.sync.bfly.b32 %r124|%p7, %r120, %r122, %r121, %r123;
mov.b32 %f17, %r124;
add.ftz.f32 %f18, %f16, %f17;
mov.b32 %r125, %f18;
mov.u32 %r126, 8;
shfl.sync.bfly.b32 %r127|%p8, %r125, %r126, %r121, %r123;
mov.b32 %f19, %r127;
add.ftz.f32 %f20, %f18, %f19;
mov.b32 %r128, %f20;
mov.u32 %r129, 4;
shfl.sync.bfly.b32 %r130|%p9, %r128, %r129, %r121, %r123;
mov.b32 %f21, %r130;
add.ftz.f32 %f22, %f20, %f21;
mov.b32 %r131, %f22;
shfl.sync.bfly.b32 %r132|%p10, %r131, %r119, %r121, %r123;
mov.b32 %f23, %r132;
add.ftz.f32 %f4, %f22, %f23;
mov.b32 %r133, %f4;
mov.u32 %r134, 1;
shfl.sync.bfly.b32 %r14|%p1, %r133, %r134, %r121, %r123;
setp.ne.s32 %p11, %r2, 0;
@%p11 bra $L__BB283_11;

mov.u32 %r135, %ctaid.x;
mov.u32 %r136, %ctaid.y;
mad.lo.s32 %r137, %r136, %r25, %r135;
mov.u32 %r138, %ctaid.z;
mad.lo.s32 %r139, %r138, %r31, %r137;
cvta.to.global.u64 %rd31, %rd13;
mul.wide.u32 %rd32, %r139, 4;
add.s64 %rd33, %rd31, %rd32;
mov.b32 %f24, %r14;
add.ftz.f32 %f25, %f4, %f24;
st.global.f32 [%rd33], %f25;

$L__BB283_11:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type20ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type20ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type20ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type20ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type20ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type20ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type20ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type20ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type20ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<35>;
.reg .b16 %rs<17>;
.reg .f32 %f<109>;
.reg .b32 %r<280>;
.reg .b64 %rd<53>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type20ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type20ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type20ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type20ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r13, [_Z13mul_mat_vec_qIL9ggml_type20ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r14, [_Z13mul_mat_vec_qIL9ggml_type20ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r15, [_Z13mul_mat_vec_qIL9ggml_type20ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r16, [_Z13mul_mat_vec_qIL9ggml_type20ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type20ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type20ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type20ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type20ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type20ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type20ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type20ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type20ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r18, [_Z13mul_mat_vec_qIL9ggml_type20ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r17, [_Z13mul_mat_vec_qIL9ggml_type20ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
cvta.to.global.u64 %rd1, %rd19;
mov.u32 %r1, %tid.y;
shl.b32 %r29, %r1, 5;
mov.u32 %r30, %tid.x;
add.s32 %r2, %r29, %r30;
mov.u32 %r31, %ctaid.x;
shl.b32 %r3, %r31, 1;
shr.u32 %r32, %r2, 31;
add.s32 %r33, %r2, %r32;
shr.s32 %r279, %r33, 1;
shr.u32 %r34, %r13, 5;
setp.ge.s32 %p3, %r279, %r34;
mov.f32 %f105, 0f00000000;
mov.f32 %f106, %f105;
mov.f32 %f107, %f105;
mov.f32 %f108, %f105;
@%p3 bra $L__BB284_3;

mov.u32 %r35, %ctaid.y;
mov.u32 %r36, %ctaid.z;
mul.hi.u32 %r37, %r36, %r23;
add.s32 %r38, %r37, %r36;
shr.u32 %r39, %r38, %r24;
mul.lo.s32 %r40, %r39, %r26;
mul.hi.u32 %r41, %r35, %r17;
add.s32 %r42, %r41, %r35;
shr.u32 %r43, %r42, %r18;
mul.lo.s32 %r44, %r43, %r20;
cvta.to.global.u64 %rd20, %rd17;
and.b32 %r47, %r33, -2;
sub.s32 %r48, %r2, %r47;
shl.b32 %r49, %r48, 1;
cvt.s64.s32 %rd2, %r49;
shl.b32 %r50, %r48, 2;
add.s32 %r278, %r279, %r15;
add.s32 %r51, %r279, %r40;
add.s32 %r52, %r51, %r44;
add.s32 %r53, %r3, 1;
mad.lo.s32 %r54, %r14, %r53, %r52;
mul.wide.s32 %rd21, %r54, 18;
add.s64 %rd52, %rd20, %rd21;
mad.lo.s32 %r55, %r3, %r14, %r279;
add.s32 %r56, %r55, %r40;
add.s32 %r57, %r56, %r44;
mul.wide.s32 %rd22, %r57, 18;
add.s64 %rd23, %rd20, %rd22;
add.s64 %rd51, %rd23, 4;
mul.wide.s32 %rd24, %r50, 2;
add.s64 %rd25, %rd22, %rd24;
add.s64 %rd26, %rd20, %rd25;
add.s64 %rd50, %rd26, 8;
or.b64 %rd6, %rd24, 4;
mul.lo.s32 %r58, %r35, %r21;
cvt.u64.u32 %rd27, %r58;
mul.lo.s32 %r59, %r36, %r27;
cvt.u64.u32 %rd28, %r59;
add.s64 %rd7, %rd28, %rd27;
cvta.to.global.u64 %rd8, %rd18;
mov.f32 %f105, 0f00000000;
shl.b64 %rd33, %rd2, 2;

$L__BB284_2:
cvt.u64.u32 %rd29, %r279;
add.s64 %rd30, %rd7, %rd29;
mul.lo.s64 %rd31, %rd30, 36;
add.s64 %rd32, %rd8, %rd31;
add.s64 %rd34, %rd32, %rd33;
ld.global.nc.u16 %rs5, [%rd50+-6];
cvt.u32.u16 %r128, %rs5;
ld.global.nc.u16 %rs6, [%rd50+-4];
mov.b32 %r129, {%rs5, %rs6};
and.b32 %r130, %r129, -2004318072;
shr.u32 %r131, %r130, 1;
or.b32 %r132, %r131, 839922192;
and.b32 %r133, %r128, 30583;
mov.u32 %r134, -152379953;
mov.u32 %r135, -1079142271;
prmt.b32 %r136, %r135, %r134, %r133;
mov.u32 %r137, 1901675829;
mov.u32 %r138, 639175937;
prmt.b32 %r139, %r138, %r137, %r133;
and.b32 %r140, %r132, 30292;
prmt.b32 %r141, %r136, %r139, %r140;
and.b16 %rs7, %rs6, 30583;
cvt.u32.u16 %r142, %rs7;
prmt.b32 %r143, %r135, %r134, %r142;
prmt.b32 %r144, %r138, %r137, %r142;
shr.u32 %r145, %r132, 16;
prmt.b32 %r146, %r143, %r144, %r145;
mov.u32 %r147, 25632;
prmt.b32 %r95, %r141, %r146, %r147;
mov.u32 %r148, 30001;
prmt.b32 %r99, %r141, %r146, %r148;
ld.global.nc.u32 %r62, [%rd34+4];
mov.u32 %r114, 0;

dp4a.s32.s32 %r60, %r95, %r62, %r114;

ld.global.nc.u32 %r66, [%rd34+20];

dp4a.s32.s32 %r64, %r99, %r66, %r60;

ld.global.nc.u16 %rs8, [%rd50+-2];
cvt.u32.u16 %r149, %rs8;
ld.global.nc.u16 %rs9, [%rd50];
mov.b32 %r150, {%rs8, %rs9};
and.b32 %r151, %r150, -2004318072;
shr.u32 %r152, %r151, 1;
or.b32 %r153, %r152, 839922192;
and.b32 %r154, %r149, 30583;
prmt.b32 %r155, %r135, %r134, %r154;
prmt.b32 %r156, %r138, %r137, %r154;
and.b32 %r157, %r153, 30292;
prmt.b32 %r158, %r155, %r156, %r157;
and.b16 %rs10, %rs9, 30583;
cvt.u32.u16 %r159, %rs10;
prmt.b32 %r160, %r135, %r134, %r159;
prmt.b32 %r161, %r138, %r137, %r159;
shr.u32 %r162, %r153, 16;
prmt.b32 %r163, %r160, %r161, %r162;
prmt.b32 %r103, %r158, %r163, %r147;
prmt.b32 %r107, %r158, %r163, %r148;
ld.global.nc.u32 %r70, [%rd34+8];

dp4a.s32.s32 %r68, %r103, %r70, %r64;

ld.global.nc.u32 %r74, [%rd34+24];

dp4a.s32.s32 %r72, %r107, %r74, %r68;

ld.global.nc.u16 %rs1, [%rd51+-4];

{ cvt.f32.f16 %f25, %rs1;}


ld.global.nc.u32 %r76, [%rd32];

{.reg .f16 low,high;
mov.b32 {low,high},%r76;
cvt.f32.f16 %f26, low;}


mul.ftz.f32 %f33, %f25, %f26;
cvt.rn.f32.s32 %f34, %r72;
fma.rn.ftz.f32 %f108, %f33, %f34, %f108;
add.s64 %rd35, %rd52, %rd6;
ld.global.nc.u16 %rs11, [%rd35+-2];
cvt.u32.u16 %r164, %rs11;
ld.global.nc.u16 %rs12, [%rd35];
mov.b32 %r165, {%rs11, %rs12};
and.b32 %r166, %r165, -2004318072;
shr.u32 %r167, %r166, 1;
or.b32 %r168, %r167, 839922192;
and.b32 %r169, %r164, 30583;
prmt.b32 %r170, %r135, %r134, %r169;
prmt.b32 %r171, %r138, %r137, %r169;
and.b32 %r172, %r168, 30292;
prmt.b32 %r173, %r170, %r171, %r172;
and.b16 %rs13, %rs12, 30583;
cvt.u32.u16 %r174, %rs13;
prmt.b32 %r175, %r135, %r134, %r174;
prmt.b32 %r176, %r138, %r137, %r174;
shr.u32 %r177, %r168, 16;
prmt.b32 %r178, %r175, %r176, %r177;
prmt.b32 %r112, %r173, %r178, %r147;
prmt.b32 %r116, %r173, %r178, %r148;

dp4a.s32.s32 %r77, %r112, %r62, %r114;


dp4a.s32.s32 %r81, %r116, %r66, %r77;

ld.global.nc.u16 %rs14, [%rd35+2];
cvt.u32.u16 %r179, %rs14;
ld.global.nc.u16 %rs15, [%rd35+4];
mov.b32 %r180, {%rs14, %rs15};
and.b32 %r181, %r180, -2004318072;
shr.u32 %r182, %r181, 1;
or.b32 %r183, %r182, 839922192;
and.b32 %r184, %r179, 30583;
prmt.b32 %r185, %r135, %r134, %r184;
prmt.b32 %r186, %r138, %r137, %r184;
and.b32 %r187, %r183, 30292;
prmt.b32 %r188, %r185, %r186, %r187;
and.b16 %rs16, %rs15, 30583;
cvt.u32.u16 %r189, %rs16;
prmt.b32 %r190, %r135, %r134, %r189;
prmt.b32 %r191, %r138, %r137, %r189;
shr.u32 %r192, %r183, 16;
prmt.b32 %r193, %r190, %r191, %r192;
prmt.b32 %r120, %r188, %r193, %r147;
prmt.b32 %r124, %r188, %r193, %r148;

dp4a.s32.s32 %r85, %r120, %r70, %r81;


dp4a.s32.s32 %r89, %r124, %r74, %r85;

ld.global.nc.u16 %rs2, [%rd52];

{ cvt.f32.f16 %f27, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r76;
cvt.f32.f16 %f28, low;}


mul.ftz.f32 %f35, %f27, %f28;
cvt.rn.f32.s32 %f36, %r89;
fma.rn.ftz.f32 %f107, %f35, %f36, %f107;
cvt.u64.u32 %rd36, %r278;
add.s64 %rd37, %rd7, %rd36;
mul.lo.s64 %rd38, %rd37, 36;
add.s64 %rd39, %rd8, %rd38;
add.s64 %rd40, %rd39, %rd33;
ld.global.nc.u32 %r96, [%rd40+4];

dp4a.s32.s32 %r94, %r95, %r96, %r114;

ld.global.nc.u32 %r100, [%rd40+20];

dp4a.s32.s32 %r98, %r99, %r100, %r94;

ld.global.nc.u32 %r104, [%rd40+8];

dp4a.s32.s32 %r102, %r103, %r104, %r98;

ld.global.nc.u32 %r108, [%rd40+24];

dp4a.s32.s32 %r106, %r107, %r108, %r102;


{ cvt.f32.f16 %f29, %rs1;}


ld.global.nc.u32 %r110, [%rd39];

{.reg .f16 low,high;
mov.b32 {low,high},%r110;
cvt.f32.f16 %f30, low;}


mul.ftz.f32 %f37, %f29, %f30;
cvt.rn.f32.s32 %f38, %r106;
fma.rn.ftz.f32 %f106, %f37, %f38, %f106;

dp4a.s32.s32 %r111, %r112, %r96, %r114;


dp4a.s32.s32 %r115, %r116, %r100, %r111;


dp4a.s32.s32 %r119, %r120, %r104, %r115;


dp4a.s32.s32 %r123, %r124, %r108, %r119;


{ cvt.f32.f16 %f31, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r110;
cvt.f32.f16 %f32, low;}


mul.ftz.f32 %f39, %f31, %f32;
cvt.rn.f32.s32 %f40, %r123;
fma.rn.ftz.f32 %f105, %f39, %f40, %f105;
add.s32 %r278, %r278, 64;
add.s64 %rd52, %rd52, 1152;
add.s64 %rd51, %rd51, 1152;
add.s64 %rd50, %rd50, 1152;
add.s32 %r279, %r279, 64;
setp.lt.s32 %p4, %r279, %r34;
@%p4 bra $L__BB284_2;

$L__BB284_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB284_5;

shl.b32 %r195, %r1, 9;
mov.u32 %r196, _ZZ13mul_mat_vec_qIL9ggml_type20ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r197, %r196, %r195;
shl.b32 %r198, %r30, 2;
add.s32 %r199, %r197, %r198;
st.shared.f32 [%r199+-512], %f108;
st.shared.f32 [%r199+-384], %f107;
st.shared.f32 [%r199+-256], %f106;
st.shared.f32 [%r199+-128], %f105;

$L__BB284_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB284_14;

mov.u32 %r200, %ctaid.z;
mov.u32 %r201, %ctaid.y;
mad.lo.s32 %r202, %r201, %r22, %r3;
mad.lo.s32 %r203, %r200, %r28, %r202;
cvt.u64.u32 %rd15, %r203;
shl.b32 %r205, %r30, 2;
mov.u32 %r206, _ZZ13mul_mat_vec_qIL9ggml_type20ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r11, %r206, %r205;
mov.u32 %r207, 2;
ld.shared.f32 %f41, [%r11];
add.ftz.f32 %f42, %f41, %f108;
ld.shared.f32 %f43, [%r11+512];
add.ftz.f32 %f44, %f43, %f42;
ld.shared.f32 %f45, [%r11+1024];
add.ftz.f32 %f46, %f45, %f44;
mov.b32 %r208, %f46;
mov.u32 %r209, 31;
mov.u32 %r210, 16;
mov.u32 %r211, -1;
shfl.sync.bfly.b32 %r212|%p7, %r208, %r210, %r209, %r211;
mov.b32 %f47, %r212;
add.ftz.f32 %f48, %f46, %f47;
mov.b32 %r213, %f48;
mov.u32 %r214, 8;
shfl.sync.bfly.b32 %r215|%p8, %r213, %r214, %r209, %r211;
mov.b32 %f49, %r215;
add.ftz.f32 %f50, %f48, %f49;
mov.b32 %r216, %f50;
mov.u32 %r217, 4;
shfl.sync.bfly.b32 %r218|%p9, %r216, %r217, %r209, %r211;
mov.b32 %f51, %r218;
add.ftz.f32 %f52, %f50, %f51;
mov.b32 %r219, %f52;
shfl.sync.bfly.b32 %r220|%p10, %r219, %r207, %r209, %r211;
mov.b32 %f53, %r220;
add.ftz.f32 %f54, %f52, %f53;
mov.b32 %r221, %f54;
mov.u32 %r222, 1;
shfl.sync.bfly.b32 %r223|%p11, %r221, %r222, %r209, %r211;
mov.b32 %f55, %r223;
add.ftz.f32 %f13, %f54, %f55;
setp.eq.s32 %p12, %r30, 0;
setp.lt.u32 %p13, %r3, %r16;
and.pred %p1, %p12, %p13;
mul.wide.u32 %rd41, %r203, 4;
add.s64 %rd16, %rd1, %rd41;
not.pred %p14, %p1;
@%p14 bra $L__BB284_8;

st.global.f32 [%rd16], %f13;

$L__BB284_8:
ld.shared.f32 %f56, [%r11+128];
add.ftz.f32 %f57, %f56, %f107;
ld.shared.f32 %f58, [%r11+640];
add.ftz.f32 %f59, %f58, %f57;
ld.shared.f32 %f60, [%r11+1152];
add.ftz.f32 %f61, %f60, %f59;
mov.b32 %r224, %f61;
shfl.sync.bfly.b32 %r228|%p15, %r224, %r210, %r209, %r211;
mov.b32 %f62, %r228;
add.ftz.f32 %f63, %f61, %f62;
mov.b32 %r229, %f63;
shfl.sync.bfly.b32 %r231|%p16, %r229, %r214, %r209, %r211;
mov.b32 %f64, %r231;
add.ftz.f32 %f65, %f63, %f64;
mov.b32 %r232, %f65;
shfl.sync.bfly.b32 %r234|%p17, %r232, %r217, %r209, %r211;
mov.b32 %f66, %r234;
add.ftz.f32 %f67, %f65, %f66;
mov.b32 %r235, %f67;
shfl.sync.bfly.b32 %r237|%p18, %r235, %r207, %r209, %r211;
mov.b32 %f68, %r237;
add.ftz.f32 %f69, %f67, %f68;
mov.b32 %r238, %f69;
shfl.sync.bfly.b32 %r240|%p19, %r238, %r222, %r209, %r211;
mov.b32 %f70, %r240;
add.ftz.f32 %f14, %f69, %f70;
setp.eq.s32 %p20, %r30, 1;
add.s32 %r242, %r3, 1;
setp.lt.u32 %p21, %r242, %r16;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB284_10;

st.global.f32 [%rd16+4], %f14;

$L__BB284_10:
ld.shared.f32 %f71, [%r11+256];
add.ftz.f32 %f72, %f71, %f106;
ld.shared.f32 %f73, [%r11+768];
add.ftz.f32 %f74, %f73, %f72;
ld.shared.f32 %f75, [%r11+1280];
add.ftz.f32 %f76, %f75, %f74;
mov.b32 %r243, %f76;
mov.u32 %r244, 31;
mov.u32 %r245, 16;
mov.u32 %r246, -1;
shfl.sync.bfly.b32 %r247|%p23, %r243, %r245, %r244, %r246;
mov.b32 %f77, %r247;
add.ftz.f32 %f78, %f76, %f77;
mov.b32 %r248, %f78;
mov.u32 %r249, 8;
shfl.sync.bfly.b32 %r250|%p24, %r248, %r249, %r244, %r246;
mov.b32 %f79, %r250;
add.ftz.f32 %f80, %f78, %f79;
mov.b32 %r251, %f80;
mov.u32 %r252, 4;
shfl.sync.bfly.b32 %r253|%p25, %r251, %r252, %r244, %r246;
mov.b32 %f81, %r253;
add.ftz.f32 %f82, %f80, %f81;
mov.b32 %r254, %f82;
mov.u32 %r255, 2;
shfl.sync.bfly.b32 %r256|%p26, %r254, %r255, %r244, %r246;
mov.b32 %f83, %r256;
add.ftz.f32 %f84, %f82, %f83;
mov.b32 %r257, %f84;
mov.u32 %r258, 1;
shfl.sync.bfly.b32 %r259|%p27, %r257, %r258, %r244, %r246;
mov.b32 %f85, %r259;
add.ftz.f32 %f15, %f84, %f85;
@%p14 bra $L__BB284_12;

cvt.u64.u32 %rd42, %r16;
add.s64 %rd43, %rd42, %rd15;
shl.b64 %rd44, %rd43, 2;
add.s64 %rd45, %rd1, %rd44;
st.global.f32 [%rd45], %f15;

$L__BB284_12:
ld.shared.f32 %f86, [%r11+384];
add.ftz.f32 %f87, %f86, %f105;
ld.shared.f32 %f88, [%r11+896];
add.ftz.f32 %f89, %f88, %f87;
ld.shared.f32 %f90, [%r11+1408];
add.ftz.f32 %f91, %f90, %f89;
mov.b32 %r260, %f91;
shfl.sync.bfly.b32 %r264|%p29, %r260, %r245, %r244, %r246;
mov.b32 %f92, %r264;
add.ftz.f32 %f93, %f91, %f92;
mov.b32 %r265, %f93;
shfl.sync.bfly.b32 %r267|%p30, %r265, %r249, %r244, %r246;
mov.b32 %f94, %r267;
add.ftz.f32 %f95, %f93, %f94;
mov.b32 %r268, %f95;
shfl.sync.bfly.b32 %r270|%p31, %r268, %r252, %r244, %r246;
mov.b32 %f96, %r270;
add.ftz.f32 %f97, %f95, %f96;
mov.b32 %r271, %f97;
shfl.sync.bfly.b32 %r273|%p32, %r271, %r255, %r244, %r246;
mov.b32 %f98, %r273;
add.ftz.f32 %f99, %f97, %f98;
mov.b32 %r274, %f99;
shfl.sync.bfly.b32 %r276|%p33, %r274, %r258, %r244, %r246;
mov.b32 %f100, %r276;
add.ftz.f32 %f16, %f99, %f100;
@%p22 bra $L__BB284_14;

add.s32 %r277, %r16, 1;
cvt.u64.u32 %rd46, %r277;
add.s64 %rd47, %rd46, %rd15;
shl.b64 %rd48, %rd47, 2;
add.s64 %rd49, %rd1, %rd48;
st.global.f32 [%rd49], %f16;

$L__BB284_14:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type20ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type20ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type20ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type20ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type20ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type20ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type20ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type20ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type20ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<47>;
.reg .b16 %rs<19>;
.reg .f32 %f<163>;
.reg .b32 %r<354>;
.reg .b64 %rd<65>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type20ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[2304];

ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type20ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type20ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type20ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type20ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r17, [_Z13mul_mat_vec_qIL9ggml_type20ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r18, [_Z13mul_mat_vec_qIL9ggml_type20ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r19, [_Z13mul_mat_vec_qIL9ggml_type20ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type20ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type20ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type20ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type20ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type20ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type20ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type20ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type20ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type20ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type20ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r33, %r1, 5;
mov.u32 %r34, %tid.x;
add.s32 %r2, %r33, %r34;
mov.u32 %r3, %ctaid.y;
mov.u32 %r4, %ctaid.z;
shr.u32 %r35, %r2, 31;
add.s32 %r36, %r2, %r35;
shr.s32 %r353, %r36, 1;
shr.u32 %r6, %r32, 5;
setp.ge.s32 %p3, %r353, %r6;
mov.f32 %f157, 0f00000000;
mov.f32 %f158, %f157;
mov.f32 %f159, %f157;
mov.f32 %f160, %f157;
mov.f32 %f161, %f157;
mov.f32 %f162, %f157;
@%p3 bra $L__BB285_3;

mul.lo.s32 %r37, %r3, %r24;
cvt.u64.u32 %rd20, %r37;
mul.lo.s32 %r38, %r4, %r30;
cvt.u64.u32 %rd21, %r38;
add.s64 %rd1, %rd21, %rd20;
mul.hi.u32 %r39, %r4, %r26;
add.s32 %r40, %r39, %r4;
shr.u32 %r41, %r40, %r27;
mul.lo.s32 %r42, %r41, %r29;
mul.hi.u32 %r43, %r3, %r20;
add.s32 %r44, %r43, %r3;
shr.u32 %r45, %r44, %r21;
mul.lo.s32 %r46, %r45, %r23;
cvta.to.global.u64 %rd22, %rd17;
and.b32 %r49, %r36, -2;
sub.s32 %r50, %r2, %r49;
shl.b32 %r51, %r50, 1;
cvt.s64.s32 %rd2, %r51;
mov.u32 %r52, %ctaid.x;
shl.b32 %r53, %r52, 1;
shl.b32 %r54, %r50, 2;
shl.b32 %r55, %r18, 1;
add.s32 %r352, %r353, %r55;
add.s32 %r351, %r353, %r18;
add.s32 %r56, %r353, %r42;
add.s32 %r57, %r56, %r46;
or.b32 %r58, %r53, 1;
mad.lo.s32 %r59, %r17, %r58, %r57;
mul.wide.s32 %rd23, %r59, 18;
add.s64 %rd64, %rd22, %rd23;
mad.lo.s32 %r60, %r53, %r17, %r353;
add.s32 %r61, %r60, %r42;
add.s32 %r62, %r61, %r46;
mul.wide.s32 %rd24, %r62, 18;
add.s64 %rd25, %rd22, %rd24;
add.s64 %rd63, %rd25, 4;
mul.wide.s32 %rd26, %r54, 2;
add.s64 %rd27, %rd24, %rd26;
add.s64 %rd28, %rd22, %rd27;
add.s64 %rd62, %rd28, 8;
or.b64 %rd6, %rd26, 4;
cvta.to.global.u64 %rd7, %rd18;
mov.f32 %f157, 0f00000000;
shl.b64 %rd33, %rd2, 2;

$L__BB285_2:
cvt.u64.u32 %rd29, %r353;
add.s64 %rd30, %rd1, %rd29;
mul.lo.s64 %rd31, %rd30, 36;
add.s64 %rd32, %rd7, %rd31;
add.s64 %rd34, %rd32, %rd33;
ld.global.nc.u16 %rs7, [%rd62+-6];
cvt.u32.u16 %r165, %rs7;
ld.global.nc.u16 %rs8, [%rd62+-4];
mov.b32 %r166, {%rs7, %rs8};
and.b32 %r167, %r166, -2004318072;
shr.u32 %r168, %r167, 1;
or.b32 %r169, %r168, 839922192;
and.b32 %r170, %r165, 30583;
mov.u32 %r171, -152379953;
mov.u32 %r172, -1079142271;
prmt.b32 %r173, %r172, %r171, %r170;
mov.u32 %r174, 1901675829;
mov.u32 %r175, 639175937;
prmt.b32 %r176, %r175, %r174, %r170;
and.b32 %r177, %r169, 30292;
prmt.b32 %r178, %r173, %r176, %r177;
and.b16 %rs9, %rs8, 30583;
cvt.u32.u16 %r179, %rs9;
prmt.b32 %r180, %r172, %r171, %r179;
prmt.b32 %r181, %r175, %r174, %r179;
shr.u32 %r182, %r169, 16;
prmt.b32 %r183, %r180, %r181, %r182;
mov.u32 %r184, 25632;
prmt.b32 %r132, %r178, %r183, %r184;
mov.u32 %r185, 30001;
prmt.b32 %r136, %r178, %r183, %r185;
ld.global.nc.u32 %r65, [%rd34+4];
mov.u32 %r151, 0;

dp4a.s32.s32 %r63, %r132, %r65, %r151;

ld.global.nc.u32 %r69, [%rd34+20];

dp4a.s32.s32 %r67, %r136, %r69, %r63;

ld.global.nc.u16 %rs10, [%rd62+-2];
cvt.u32.u16 %r186, %rs10;
ld.global.nc.u16 %rs11, [%rd62];
mov.b32 %r187, {%rs10, %rs11};
and.b32 %r188, %r187, -2004318072;
shr.u32 %r189, %r188, 1;
or.b32 %r190, %r189, 839922192;
and.b32 %r191, %r186, 30583;
prmt.b32 %r192, %r172, %r171, %r191;
prmt.b32 %r193, %r175, %r174, %r191;
and.b32 %r194, %r190, 30292;
prmt.b32 %r195, %r192, %r193, %r194;
and.b16 %rs12, %rs11, 30583;
cvt.u32.u16 %r196, %rs12;
prmt.b32 %r197, %r172, %r171, %r196;
prmt.b32 %r198, %r175, %r174, %r196;
shr.u32 %r199, %r190, 16;
prmt.b32 %r200, %r197, %r198, %r199;
prmt.b32 %r140, %r195, %r200, %r184;
prmt.b32 %r144, %r195, %r200, %r185;
ld.global.nc.u32 %r73, [%rd34+8];

dp4a.s32.s32 %r71, %r140, %r73, %r67;

ld.global.nc.u32 %r77, [%rd34+24];

dp4a.s32.s32 %r75, %r144, %r77, %r71;

ld.global.nc.u16 %rs1, [%rd63+-4];

{ cvt.f32.f16 %f37, %rs1;}


ld.global.nc.u32 %r79, [%rd32];

{.reg .f16 low,high;
mov.b32 {low,high},%r79;
cvt.f32.f16 %f38, low;}


mul.ftz.f32 %f49, %f37, %f38;
cvt.rn.f32.s32 %f50, %r75;
fma.rn.ftz.f32 %f162, %f49, %f50, %f162;
add.s64 %rd35, %rd64, %rd6;
ld.global.nc.u16 %rs13, [%rd35+-2];
cvt.u32.u16 %r201, %rs13;
ld.global.nc.u16 %rs14, [%rd35];
mov.b32 %r202, {%rs13, %rs14};
and.b32 %r203, %r202, -2004318072;
shr.u32 %r204, %r203, 1;
or.b32 %r205, %r204, 839922192;
and.b32 %r206, %r201, 30583;
prmt.b32 %r207, %r172, %r171, %r206;
prmt.b32 %r208, %r175, %r174, %r206;
and.b32 %r209, %r205, 30292;
prmt.b32 %r210, %r207, %r208, %r209;
and.b16 %rs15, %rs14, 30583;
cvt.u32.u16 %r211, %rs15;
prmt.b32 %r212, %r172, %r171, %r211;
prmt.b32 %r213, %r175, %r174, %r211;
shr.u32 %r214, %r205, 16;
prmt.b32 %r215, %r212, %r213, %r214;
prmt.b32 %r149, %r210, %r215, %r184;
prmt.b32 %r153, %r210, %r215, %r185;

dp4a.s32.s32 %r80, %r149, %r65, %r151;


dp4a.s32.s32 %r84, %r153, %r69, %r80;

ld.global.nc.u16 %rs16, [%rd35+2];
cvt.u32.u16 %r216, %rs16;
ld.global.nc.u16 %rs17, [%rd35+4];
mov.b32 %r217, {%rs16, %rs17};
and.b32 %r218, %r217, -2004318072;
shr.u32 %r219, %r218, 1;
or.b32 %r220, %r219, 839922192;
and.b32 %r221, %r216, 30583;
prmt.b32 %r222, %r172, %r171, %r221;
prmt.b32 %r223, %r175, %r174, %r221;
and.b32 %r224, %r220, 30292;
prmt.b32 %r225, %r222, %r223, %r224;
and.b16 %rs18, %rs17, 30583;
cvt.u32.u16 %r226, %rs18;
prmt.b32 %r227, %r172, %r171, %r226;
prmt.b32 %r228, %r175, %r174, %r226;
shr.u32 %r229, %r220, 16;
prmt.b32 %r230, %r227, %r228, %r229;
prmt.b32 %r157, %r225, %r230, %r184;
prmt.b32 %r161, %r225, %r230, %r185;

dp4a.s32.s32 %r88, %r157, %r73, %r84;


dp4a.s32.s32 %r92, %r161, %r77, %r88;

ld.global.nc.u16 %rs2, [%rd64];

{ cvt.f32.f16 %f39, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r79;
cvt.f32.f16 %f40, low;}


mul.ftz.f32 %f51, %f39, %f40;
cvt.rn.f32.s32 %f52, %r92;
fma.rn.ftz.f32 %f161, %f51, %f52, %f161;
cvt.u64.u32 %rd36, %r351;
add.s64 %rd37, %rd1, %rd36;
mul.lo.s64 %rd38, %rd37, 36;
add.s64 %rd39, %rd7, %rd38;
add.s64 %rd40, %rd39, %rd33;
ld.global.nc.u32 %r99, [%rd40+4];

dp4a.s32.s32 %r97, %r132, %r99, %r151;

ld.global.nc.u32 %r103, [%rd40+20];

dp4a.s32.s32 %r101, %r136, %r103, %r97;

ld.global.nc.u32 %r107, [%rd40+8];

dp4a.s32.s32 %r105, %r140, %r107, %r101;

ld.global.nc.u32 %r111, [%rd40+24];

dp4a.s32.s32 %r109, %r144, %r111, %r105;


{ cvt.f32.f16 %f41, %rs1;}


ld.global.nc.u32 %r113, [%rd39];

{.reg .f16 low,high;
mov.b32 {low,high},%r113;
cvt.f32.f16 %f42, low;}


mul.ftz.f32 %f53, %f41, %f42;
cvt.rn.f32.s32 %f54, %r109;
fma.rn.ftz.f32 %f160, %f53, %f54, %f160;

dp4a.s32.s32 %r114, %r149, %r99, %r151;


dp4a.s32.s32 %r118, %r153, %r103, %r114;


dp4a.s32.s32 %r122, %r157, %r107, %r118;


dp4a.s32.s32 %r126, %r161, %r111, %r122;


{ cvt.f32.f16 %f43, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r113;
cvt.f32.f16 %f44, low;}


mul.ftz.f32 %f55, %f43, %f44;
cvt.rn.f32.s32 %f56, %r126;
fma.rn.ftz.f32 %f159, %f55, %f56, %f159;
cvt.u64.u32 %rd41, %r352;
add.s64 %rd42, %rd1, %rd41;
mul.lo.s64 %rd43, %rd42, 36;
add.s64 %rd44, %rd7, %rd43;
add.s64 %rd45, %rd44, %rd33;
ld.global.nc.u32 %r133, [%rd45+4];

dp4a.s32.s32 %r131, %r132, %r133, %r151;

ld.global.nc.u32 %r137, [%rd45+20];

dp4a.s32.s32 %r135, %r136, %r137, %r131;

ld.global.nc.u32 %r141, [%rd45+8];

dp4a.s32.s32 %r139, %r140, %r141, %r135;

ld.global.nc.u32 %r145, [%rd45+24];

dp4a.s32.s32 %r143, %r144, %r145, %r139;


{ cvt.f32.f16 %f45, %rs1;}


ld.global.nc.u32 %r147, [%rd44];

{.reg .f16 low,high;
mov.b32 {low,high},%r147;
cvt.f32.f16 %f46, low;}


mul.ftz.f32 %f57, %f45, %f46;
cvt.rn.f32.s32 %f58, %r143;
fma.rn.ftz.f32 %f158, %f57, %f58, %f158;

dp4a.s32.s32 %r148, %r149, %r133, %r151;


dp4a.s32.s32 %r152, %r153, %r137, %r148;


dp4a.s32.s32 %r156, %r157, %r141, %r152;


dp4a.s32.s32 %r160, %r161, %r145, %r156;


{ cvt.f32.f16 %f47, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r147;
cvt.f32.f16 %f48, low;}


mul.ftz.f32 %f59, %f47, %f48;
cvt.rn.f32.s32 %f60, %r160;
fma.rn.ftz.f32 %f157, %f59, %f60, %f157;
add.s32 %r352, %r352, 64;
add.s32 %r351, %r351, 64;
add.s64 %rd64, %rd64, 1152;
add.s64 %rd63, %rd63, 1152;
add.s64 %rd62, %rd62, 1152;
add.s32 %r353, %r353, 64;
setp.lt.s32 %p4, %r353, %r6;
@%p4 bra $L__BB285_2;

$L__BB285_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB285_5;

mov.u32 %r232, _ZZ13mul_mat_vec_qIL9ggml_type20ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r233, %r1, 768, %r232;
shl.b32 %r234, %r34, 2;
add.s32 %r235, %r233, %r234;
st.shared.f32 [%r235+-768], %f162;
st.shared.f32 [%r235+-640], %f161;
st.shared.f32 [%r235+-512], %f160;
st.shared.f32 [%r235+-384], %f159;
st.shared.f32 [%r235+-256], %f158;
st.shared.f32 [%r235+-128], %f157;

$L__BB285_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB285_18;

mov.u32 %r236, %ctaid.x;
shl.b32 %r237, %r236, 1;
mov.u32 %r238, 1;
mad.lo.s32 %r239, %r3, %r25, %r237;
mad.lo.s32 %r240, %r4, %r31, %r239;
cvt.u64.u32 %rd14, %r240;
shl.b32 %r242, %r34, 2;
mov.u32 %r243, _ZZ13mul_mat_vec_qIL9ggml_type20ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r15, %r243, %r242;
mov.u32 %r244, 2;
ld.shared.f32 %f61, [%r15];
add.ftz.f32 %f62, %f61, %f162;
ld.shared.f32 %f63, [%r15+768];
add.ftz.f32 %f64, %f63, %f62;
ld.shared.f32 %f65, [%r15+1536];
add.ftz.f32 %f66, %f65, %f64;
mov.b32 %r245, %f66;
mov.u32 %r246, 31;
mov.u32 %r247, 16;
mov.u32 %r248, -1;
shfl.sync.bfly.b32 %r249|%p7, %r245, %r247, %r246, %r248;
mov.b32 %f67, %r249;
add.ftz.f32 %f68, %f66, %f67;
mov.b32 %r250, %f68;
mov.u32 %r251, 8;
shfl.sync.bfly.b32 %r252|%p8, %r250, %r251, %r246, %r248;
mov.b32 %f69, %r252;
add.ftz.f32 %f70, %f68, %f69;
mov.b32 %r253, %f70;
mov.u32 %r254, 4;
shfl.sync.bfly.b32 %r255|%p9, %r253, %r254, %r246, %r248;
mov.b32 %f71, %r255;
add.ftz.f32 %f72, %f70, %f71;
mov.b32 %r256, %f72;
shfl.sync.bfly.b32 %r257|%p10, %r256, %r244, %r246, %r248;
mov.b32 %f73, %r257;
add.ftz.f32 %f74, %f72, %f73;
mov.b32 %r258, %f74;
shfl.sync.bfly.b32 %r259|%p11, %r258, %r238, %r246, %r248;
mov.b32 %f75, %r259;
add.ftz.f32 %f19, %f74, %f75;
setp.eq.s32 %p12, %r34, 0;
setp.lt.u32 %p13, %r237, %r19;
and.pred %p1, %p12, %p13;
cvta.to.global.u64 %rd46, %rd19;
mul.wide.u32 %rd47, %r240, 4;
add.s64 %rd15, %rd46, %rd47;
not.pred %p14, %p1;
@%p14 bra $L__BB285_8;

st.global.f32 [%rd15], %f19;

$L__BB285_8:
ld.shared.f32 %f76, [%r15+128];
add.ftz.f32 %f77, %f76, %f161;
ld.shared.f32 %f78, [%r15+896];
add.ftz.f32 %f79, %f78, %f77;
ld.shared.f32 %f80, [%r15+1664];
add.ftz.f32 %f81, %f80, %f79;
mov.b32 %r260, %f81;
shfl.sync.bfly.b32 %r264|%p15, %r260, %r247, %r246, %r248;
mov.b32 %f82, %r264;
add.ftz.f32 %f83, %f81, %f82;
mov.b32 %r265, %f83;
shfl.sync.bfly.b32 %r267|%p16, %r265, %r251, %r246, %r248;
mov.b32 %f84, %r267;
add.ftz.f32 %f85, %f83, %f84;
mov.b32 %r268, %f85;
shfl.sync.bfly.b32 %r270|%p17, %r268, %r254, %r246, %r248;
mov.b32 %f86, %r270;
add.ftz.f32 %f87, %f85, %f86;
mov.b32 %r271, %f87;
shfl.sync.bfly.b32 %r273|%p18, %r271, %r244, %r246, %r248;
mov.b32 %f88, %r273;
add.ftz.f32 %f89, %f87, %f88;
mov.b32 %r274, %f89;
shfl.sync.bfly.b32 %r276|%p19, %r274, %r238, %r246, %r248;
mov.b32 %f90, %r276;
add.ftz.f32 %f20, %f89, %f90;
setp.eq.s32 %p20, %r34, 1;
or.b32 %r280, %r237, 1;
setp.lt.u32 %p21, %r280, %r19;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB285_10;

st.global.f32 [%rd15+4], %f20;

$L__BB285_10:
ld.shared.f32 %f91, [%r15+256];
add.ftz.f32 %f92, %f91, %f160;
ld.shared.f32 %f93, [%r15+1024];
add.ftz.f32 %f94, %f93, %f92;
ld.shared.f32 %f95, [%r15+1792];
add.ftz.f32 %f96, %f95, %f94;
mov.b32 %r281, %f96;
mov.u32 %r282, 31;
mov.u32 %r283, 16;
mov.u32 %r284, -1;
shfl.sync.bfly.b32 %r285|%p23, %r281, %r283, %r282, %r284;
mov.b32 %f97, %r285;
add.ftz.f32 %f98, %f96, %f97;
mov.b32 %r286, %f98;
mov.u32 %r287, 8;
shfl.sync.bfly.b32 %r288|%p24, %r286, %r287, %r282, %r284;
mov.b32 %f99, %r288;
add.ftz.f32 %f100, %f98, %f99;
mov.b32 %r289, %f100;
mov.u32 %r290, 4;
shfl.sync.bfly.b32 %r291|%p25, %r289, %r290, %r282, %r284;
mov.b32 %f101, %r291;
add.ftz.f32 %f102, %f100, %f101;
mov.b32 %r292, %f102;
mov.u32 %r293, 2;
shfl.sync.bfly.b32 %r294|%p26, %r292, %r293, %r282, %r284;
mov.b32 %f103, %r294;
add.ftz.f32 %f104, %f102, %f103;
mov.b32 %r295, %f104;
mov.u32 %r296, 1;
shfl.sync.bfly.b32 %r297|%p27, %r295, %r296, %r282, %r284;
mov.b32 %f105, %r297;
add.ftz.f32 %f21, %f104, %f105;
@%p14 bra $L__BB285_12;

cvt.u64.u32 %rd48, %r19;
add.s64 %rd49, %rd48, %rd14;
shl.b64 %rd51, %rd49, 2;
add.s64 %rd52, %rd46, %rd51;
st.global.f32 [%rd52], %f21;

$L__BB285_12:
ld.shared.f32 %f106, [%r15+384];
add.ftz.f32 %f107, %f106, %f159;
ld.shared.f32 %f108, [%r15+1152];
add.ftz.f32 %f109, %f108, %f107;
ld.shared.f32 %f110, [%r15+1920];
add.ftz.f32 %f111, %f110, %f109;
mov.b32 %r298, %f111;
shfl.sync.bfly.b32 %r302|%p29, %r298, %r283, %r282, %r284;
mov.b32 %f112, %r302;
add.ftz.f32 %f113, %f111, %f112;
mov.b32 %r303, %f113;
shfl.sync.bfly.b32 %r305|%p30, %r303, %r287, %r282, %r284;
mov.b32 %f114, %r305;
add.ftz.f32 %f115, %f113, %f114;
mov.b32 %r306, %f115;
shfl.sync.bfly.b32 %r308|%p31, %r306, %r290, %r282, %r284;
mov.b32 %f116, %r308;
add.ftz.f32 %f117, %f115, %f116;
mov.b32 %r309, %f117;
shfl.sync.bfly.b32 %r311|%p32, %r309, %r293, %r282, %r284;
mov.b32 %f118, %r311;
add.ftz.f32 %f119, %f117, %f118;
mov.b32 %r312, %f119;
shfl.sync.bfly.b32 %r314|%p33, %r312, %r296, %r282, %r284;
mov.b32 %f120, %r314;
add.ftz.f32 %f22, %f119, %f120;
@%p22 bra $L__BB285_14;

add.s32 %r315, %r19, 1;
cvt.u64.u32 %rd53, %r315;
add.s64 %rd54, %rd53, %rd14;
shl.b64 %rd56, %rd54, 2;
add.s64 %rd57, %rd46, %rd56;
st.global.f32 [%rd57], %f22;

$L__BB285_14:
shl.b32 %r316, %r19, 1;
mov.u32 %r317, 1;
ld.shared.f32 %f121, [%r15+512];
add.ftz.f32 %f122, %f121, %f158;
ld.shared.f32 %f123, [%r15+1280];
add.ftz.f32 %f124, %f123, %f122;
ld.shared.f32 %f125, [%r15+2048];
add.ftz.f32 %f126, %f125, %f124;
mov.b32 %r318, %f126;
mov.u32 %r319, 31;
mov.u32 %r320, 16;
mov.u32 %r321, -1;
shfl.sync.bfly.b32 %r322|%p35, %r318, %r320, %r319, %r321;
mov.b32 %f127, %r322;
add.ftz.f32 %f128, %f126, %f127;
mov.b32 %r323, %f128;
mov.u32 %r324, 8;
shfl.sync.bfly.b32 %r325|%p36, %r323, %r324, %r319, %r321;
mov.b32 %f129, %r325;
add.ftz.f32 %f130, %f128, %f129;
mov.b32 %r326, %f130;
mov.u32 %r327, 4;
shfl.sync.bfly.b32 %r328|%p37, %r326, %r327, %r319, %r321;
mov.b32 %f131, %r328;
add.ftz.f32 %f132, %f130, %f131;
mov.b32 %r329, %f132;
mov.u32 %r330, 2;
shfl.sync.bfly.b32 %r331|%p38, %r329, %r330, %r319, %r321;
mov.b32 %f133, %r331;
add.ftz.f32 %f134, %f132, %f133;
mov.b32 %r332, %f134;
shfl.sync.bfly.b32 %r333|%p39, %r332, %r317, %r319, %r321;
mov.b32 %f135, %r333;
add.ftz.f32 %f23, %f134, %f135;
cvt.u64.u32 %rd58, %r316;
add.s64 %rd59, %rd58, %rd14;
shl.b64 %rd61, %rd59, 2;
add.s64 %rd16, %rd46, %rd61;
@%p14 bra $L__BB285_16;

st.global.f32 [%rd16], %f23;

$L__BB285_16:
ld.shared.f32 %f136, [%r15+640];
add.ftz.f32 %f137, %f136, %f157;
ld.shared.f32 %f138, [%r15+1408];
add.ftz.f32 %f139, %f138, %f137;
ld.shared.f32 %f140, [%r15+2176];
add.ftz.f32 %f141, %f140, %f139;
mov.b32 %r334, %f141;
shfl.sync.bfly.b32 %r338|%p41, %r334, %r320, %r319, %r321;
mov.b32 %f142, %r338;
add.ftz.f32 %f143, %f141, %f142;
mov.b32 %r339, %f143;
shfl.sync.bfly.b32 %r341|%p42, %r339, %r324, %r319, %r321;
mov.b32 %f144, %r341;
add.ftz.f32 %f145, %f143, %f144;
mov.b32 %r342, %f145;
shfl.sync.bfly.b32 %r344|%p43, %r342, %r327, %r319, %r321;
mov.b32 %f146, %r344;
add.ftz.f32 %f147, %f145, %f146;
mov.b32 %r345, %f147;
shfl.sync.bfly.b32 %r347|%p44, %r345, %r330, %r319, %r321;
mov.b32 %f148, %r347;
add.ftz.f32 %f149, %f147, %f148;
mov.b32 %r348, %f149;
shfl.sync.bfly.b32 %r350|%p45, %r348, %r317, %r319, %r321;
mov.b32 %f150, %r350;
add.ftz.f32 %f24, %f149, %f150;
@%p22 bra $L__BB285_18;

st.global.f32 [%rd16+4], %f24;

$L__BB285_18:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type20ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type20ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type20ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type20ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type20ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type20ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type20ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type20ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type20ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<64>;
.reg .b16 %rs<21>;
.reg .f32 %f<217>;
.reg .b32 %r<451>;
.reg .b64 %rd<83>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type20ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[3072];

ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type20ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type20ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type20ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r18, [_Z13mul_mat_vec_qIL9ggml_type20ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r19, [_Z13mul_mat_vec_qIL9ggml_type20ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type20ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type20ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type20ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type20ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type20ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type20ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type20ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type20ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type20ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type20ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type20ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type20ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r34, %r1, 5;
mov.u32 %r35, %tid.x;
mov.u32 %r36, %ctaid.x;
shl.b32 %r2, %r36, 1;
add.s32 %r37, %r34, %r35;
shr.u32 %r38, %r37, 31;
add.s32 %r39, %r37, %r38;
shr.s32 %r450, %r39, 1;
shr.u32 %r40, %r18, 5;
setp.ge.s32 %p2, %r450, %r40;
mov.f32 %f209, 0f00000000;
mov.f32 %f210, %f209;
mov.f32 %f211, %f209;
mov.f32 %f212, %f209;
mov.f32 %f213, %f209;
mov.f32 %f214, %f209;
mov.f32 %f215, %f209;
mov.f32 %f216, %f209;
@%p2 bra $L__BB286_3;

mov.u32 %r41, %ctaid.y;
mov.u32 %r42, %ctaid.z;
mul.hi.u32 %r43, %r42, %r28;
add.s32 %r44, %r43, %r42;
shr.u32 %r45, %r44, %r29;
mul.lo.s32 %r46, %r45, %r31;
mul.hi.u32 %r47, %r41, %r22;
add.s32 %r48, %r47, %r41;
shr.u32 %r49, %r48, %r23;
mul.lo.s32 %r50, %r49, %r25;
cvta.to.global.u64 %rd20, %rd17;
and.b32 %r57, %r39, -2;
sub.s32 %r58, %r37, %r57;
shl.b32 %r59, %r58, 1;
cvt.s64.s32 %rd1, %r59;
shl.b32 %r60, %r58, 2;
shl.b32 %r61, %r20, 1;
add.s32 %r449, %r450, %r61;
mad.lo.s32 %r448, %r20, 3, %r450;
add.s32 %r447, %r450, %r20;
add.s32 %r62, %r450, %r46;
add.s32 %r63, %r62, %r50;
add.s32 %r64, %r2, 1;
mad.lo.s32 %r65, %r19, %r64, %r63;
mul.wide.s32 %rd21, %r65, 18;
add.s64 %rd82, %rd20, %rd21;
mad.lo.s32 %r66, %r2, %r19, %r450;
add.s32 %r67, %r66, %r46;
add.s32 %r68, %r67, %r50;
mul.wide.s32 %rd22, %r68, 18;
add.s64 %rd23, %rd20, %rd22;
add.s64 %rd81, %rd23, 4;
mul.wide.s32 %rd24, %r60, 2;
add.s64 %rd25, %rd22, %rd24;
add.s64 %rd26, %rd20, %rd25;
add.s64 %rd80, %rd26, 8;
or.b64 %rd5, %rd24, 4;
mul.lo.s32 %r69, %r41, %r26;
cvt.u64.u32 %rd27, %r69;
mul.lo.s32 %r70, %r42, %r32;
cvt.u64.u32 %rd28, %r70;
add.s64 %rd6, %rd28, %rd27;
cvta.to.global.u64 %rd7, %rd18;
mov.f32 %f209, 0f00000000;
shl.b64 %rd33, %rd1, 2;

$L__BB286_2:
cvt.u64.u32 %rd29, %r450;
add.s64 %rd30, %rd6, %rd29;
mul.lo.s64 %rd31, %rd30, 36;
add.s64 %rd32, %rd7, %rd31;
add.s64 %rd34, %rd32, %rd33;
ld.global.nc.u16 %rs9, [%rd80+-6];
cvt.u32.u16 %r207, %rs9;
ld.global.nc.u16 %rs10, [%rd80+-4];
mov.b32 %r208, {%rs9, %rs10};
and.b32 %r209, %r208, -2004318072;
shr.u32 %r210, %r209, 1;
or.b32 %r211, %r210, 839922192;
and.b32 %r212, %r207, 30583;
mov.u32 %r213, -152379953;
mov.u32 %r214, -1079142271;
prmt.b32 %r215, %r214, %r213, %r212;
mov.u32 %r216, 1901675829;
mov.u32 %r217, 639175937;
prmt.b32 %r218, %r217, %r216, %r212;
and.b32 %r219, %r211, 30292;
prmt.b32 %r220, %r215, %r218, %r219;
and.b16 %rs11, %rs10, 30583;
cvt.u32.u16 %r221, %rs11;
prmt.b32 %r222, %r214, %r213, %r221;
prmt.b32 %r223, %r217, %r216, %r221;
shr.u32 %r224, %r211, 16;
prmt.b32 %r225, %r222, %r223, %r224;
mov.u32 %r226, 25632;
prmt.b32 %r174, %r220, %r225, %r226;
mov.u32 %r227, 30001;
prmt.b32 %r178, %r220, %r225, %r227;
ld.global.nc.u32 %r73, [%rd34+4];
mov.u32 %r193, 0;

dp4a.s32.s32 %r71, %r174, %r73, %r193;

ld.global.nc.u32 %r77, [%rd34+20];

dp4a.s32.s32 %r75, %r178, %r77, %r71;

ld.global.nc.u16 %rs12, [%rd80+-2];
cvt.u32.u16 %r228, %rs12;
ld.global.nc.u16 %rs13, [%rd80];
mov.b32 %r229, {%rs12, %rs13};
and.b32 %r230, %r229, -2004318072;
shr.u32 %r231, %r230, 1;
or.b32 %r232, %r231, 839922192;
and.b32 %r233, %r228, 30583;
prmt.b32 %r234, %r214, %r213, %r233;
prmt.b32 %r235, %r217, %r216, %r233;
and.b32 %r236, %r232, 30292;
prmt.b32 %r237, %r234, %r235, %r236;
and.b16 %rs14, %rs13, 30583;
cvt.u32.u16 %r238, %rs14;
prmt.b32 %r239, %r214, %r213, %r238;
prmt.b32 %r240, %r217, %r216, %r238;
shr.u32 %r241, %r232, 16;
prmt.b32 %r242, %r239, %r240, %r241;
prmt.b32 %r182, %r237, %r242, %r226;
prmt.b32 %r186, %r237, %r242, %r227;
ld.global.nc.u32 %r81, [%rd34+8];

dp4a.s32.s32 %r79, %r182, %r81, %r75;

ld.global.nc.u32 %r85, [%rd34+24];

dp4a.s32.s32 %r83, %r186, %r85, %r79;

ld.global.nc.u16 %rs1, [%rd81+-4];

{ cvt.f32.f16 %f49, %rs1;}


ld.global.nc.u32 %r87, [%rd32];

{.reg .f16 low,high;
mov.b32 {low,high},%r87;
cvt.f32.f16 %f50, low;}


mul.ftz.f32 %f65, %f49, %f50;
cvt.rn.f32.s32 %f66, %r83;
fma.rn.ftz.f32 %f216, %f65, %f66, %f216;
add.s64 %rd35, %rd82, %rd5;
ld.global.nc.u16 %rs15, [%rd35+-2];
cvt.u32.u16 %r243, %rs15;
ld.global.nc.u16 %rs16, [%rd35];
mov.b32 %r244, {%rs15, %rs16};
and.b32 %r245, %r244, -2004318072;
shr.u32 %r246, %r245, 1;
or.b32 %r247, %r246, 839922192;
and.b32 %r248, %r243, 30583;
prmt.b32 %r249, %r214, %r213, %r248;
prmt.b32 %r250, %r217, %r216, %r248;
and.b32 %r251, %r247, 30292;
prmt.b32 %r252, %r249, %r250, %r251;
and.b16 %rs17, %rs16, 30583;
cvt.u32.u16 %r253, %rs17;
prmt.b32 %r254, %r214, %r213, %r253;
prmt.b32 %r255, %r217, %r216, %r253;
shr.u32 %r256, %r247, 16;
prmt.b32 %r257, %r254, %r255, %r256;
prmt.b32 %r191, %r252, %r257, %r226;
prmt.b32 %r195, %r252, %r257, %r227;

dp4a.s32.s32 %r88, %r191, %r73, %r193;


dp4a.s32.s32 %r92, %r195, %r77, %r88;

ld.global.nc.u16 %rs18, [%rd35+2];
cvt.u32.u16 %r258, %rs18;
ld.global.nc.u16 %rs19, [%rd35+4];
mov.b32 %r259, {%rs18, %rs19};
and.b32 %r260, %r259, -2004318072;
shr.u32 %r261, %r260, 1;
or.b32 %r262, %r261, 839922192;
and.b32 %r263, %r258, 30583;
prmt.b32 %r264, %r214, %r213, %r263;
prmt.b32 %r265, %r217, %r216, %r263;
and.b32 %r266, %r262, 30292;
prmt.b32 %r267, %r264, %r265, %r266;
and.b16 %rs20, %rs19, 30583;
cvt.u32.u16 %r268, %rs20;
prmt.b32 %r269, %r214, %r213, %r268;
prmt.b32 %r270, %r217, %r216, %r268;
shr.u32 %r271, %r262, 16;
prmt.b32 %r272, %r269, %r270, %r271;
prmt.b32 %r199, %r267, %r272, %r226;
prmt.b32 %r203, %r267, %r272, %r227;

dp4a.s32.s32 %r96, %r199, %r81, %r92;


dp4a.s32.s32 %r100, %r203, %r85, %r96;

ld.global.nc.u16 %rs2, [%rd82];

{ cvt.f32.f16 %f51, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r87;
cvt.f32.f16 %f52, low;}


mul.ftz.f32 %f67, %f51, %f52;
cvt.rn.f32.s32 %f68, %r100;
fma.rn.ftz.f32 %f215, %f67, %f68, %f215;
cvt.u64.u32 %rd36, %r447;
add.s64 %rd37, %rd6, %rd36;
mul.lo.s64 %rd38, %rd37, 36;
add.s64 %rd39, %rd7, %rd38;
add.s64 %rd40, %rd39, %rd33;
ld.global.nc.u32 %r107, [%rd40+4];

dp4a.s32.s32 %r105, %r174, %r107, %r193;

ld.global.nc.u32 %r111, [%rd40+20];

dp4a.s32.s32 %r109, %r178, %r111, %r105;

ld.global.nc.u32 %r115, [%rd40+8];

dp4a.s32.s32 %r113, %r182, %r115, %r109;

ld.global.nc.u32 %r119, [%rd40+24];

dp4a.s32.s32 %r117, %r186, %r119, %r113;


{ cvt.f32.f16 %f53, %rs1;}


ld.global.nc.u32 %r121, [%rd39];

{.reg .f16 low,high;
mov.b32 {low,high},%r121;
cvt.f32.f16 %f54, low;}


mul.ftz.f32 %f69, %f53, %f54;
cvt.rn.f32.s32 %f70, %r117;
fma.rn.ftz.f32 %f214, %f69, %f70, %f214;

dp4a.s32.s32 %r122, %r191, %r107, %r193;


dp4a.s32.s32 %r126, %r195, %r111, %r122;


dp4a.s32.s32 %r130, %r199, %r115, %r126;


dp4a.s32.s32 %r134, %r203, %r119, %r130;


{ cvt.f32.f16 %f55, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r121;
cvt.f32.f16 %f56, low;}


mul.ftz.f32 %f71, %f55, %f56;
cvt.rn.f32.s32 %f72, %r134;
fma.rn.ftz.f32 %f213, %f71, %f72, %f213;
cvt.u64.u32 %rd41, %r449;
add.s64 %rd42, %rd6, %rd41;
mul.lo.s64 %rd43, %rd42, 36;
add.s64 %rd44, %rd7, %rd43;
add.s64 %rd45, %rd44, %rd33;
ld.global.nc.u32 %r141, [%rd45+4];

dp4a.s32.s32 %r139, %r174, %r141, %r193;

ld.global.nc.u32 %r145, [%rd45+20];

dp4a.s32.s32 %r143, %r178, %r145, %r139;

ld.global.nc.u32 %r149, [%rd45+8];

dp4a.s32.s32 %r147, %r182, %r149, %r143;

ld.global.nc.u32 %r153, [%rd45+24];

dp4a.s32.s32 %r151, %r186, %r153, %r147;


{ cvt.f32.f16 %f57, %rs1;}


ld.global.nc.u32 %r155, [%rd44];

{.reg .f16 low,high;
mov.b32 {low,high},%r155;
cvt.f32.f16 %f58, low;}


mul.ftz.f32 %f73, %f57, %f58;
cvt.rn.f32.s32 %f74, %r151;
fma.rn.ftz.f32 %f212, %f73, %f74, %f212;

dp4a.s32.s32 %r156, %r191, %r141, %r193;


dp4a.s32.s32 %r160, %r195, %r145, %r156;


dp4a.s32.s32 %r164, %r199, %r149, %r160;


dp4a.s32.s32 %r168, %r203, %r153, %r164;


{ cvt.f32.f16 %f59, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r155;
cvt.f32.f16 %f60, low;}


mul.ftz.f32 %f75, %f59, %f60;
cvt.rn.f32.s32 %f76, %r168;
fma.rn.ftz.f32 %f211, %f75, %f76, %f211;
cvt.u64.u32 %rd46, %r448;
add.s64 %rd47, %rd6, %rd46;
mul.lo.s64 %rd48, %rd47, 36;
add.s64 %rd49, %rd7, %rd48;
add.s64 %rd50, %rd49, %rd33;
ld.global.nc.u32 %r175, [%rd50+4];

dp4a.s32.s32 %r173, %r174, %r175, %r193;

ld.global.nc.u32 %r179, [%rd50+20];

dp4a.s32.s32 %r177, %r178, %r179, %r173;

ld.global.nc.u32 %r183, [%rd50+8];

dp4a.s32.s32 %r181, %r182, %r183, %r177;

ld.global.nc.u32 %r187, [%rd50+24];

dp4a.s32.s32 %r185, %r186, %r187, %r181;


{ cvt.f32.f16 %f61, %rs1;}


ld.global.nc.u32 %r189, [%rd49];

{.reg .f16 low,high;
mov.b32 {low,high},%r189;
cvt.f32.f16 %f62, low;}


mul.ftz.f32 %f77, %f61, %f62;
cvt.rn.f32.s32 %f78, %r185;
fma.rn.ftz.f32 %f210, %f77, %f78, %f210;

dp4a.s32.s32 %r190, %r191, %r175, %r193;


dp4a.s32.s32 %r194, %r195, %r179, %r190;


dp4a.s32.s32 %r198, %r199, %r183, %r194;


dp4a.s32.s32 %r202, %r203, %r187, %r198;


{ cvt.f32.f16 %f63, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r189;
cvt.f32.f16 %f64, low;}


mul.ftz.f32 %f79, %f63, %f64;
cvt.rn.f32.s32 %f80, %r202;
fma.rn.ftz.f32 %f209, %f79, %f80, %f209;
add.s32 %r449, %r449, 64;
add.s32 %r448, %r448, 64;
add.s32 %r447, %r447, 64;
add.s64 %rd82, %rd82, 1152;
add.s64 %rd81, %rd81, 1152;
add.s64 %rd80, %rd80, 1152;
add.s32 %r450, %r450, 64;
setp.lt.s32 %p3, %r450, %r40;
@%p3 bra $L__BB286_2;

$L__BB286_3:
setp.eq.s32 %p4, %r1, 0;
@%p4 bra $L__BB286_5;

shl.b32 %r274, %r1, 10;
mov.u32 %r275, _ZZ13mul_mat_vec_qIL9ggml_type20ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r276, %r275, %r274;
shl.b32 %r277, %r35, 2;
add.s32 %r278, %r276, %r277;
st.shared.f32 [%r278+-1024], %f216;
st.shared.f32 [%r278+-896], %f215;
st.shared.f32 [%r278+-768], %f214;
st.shared.f32 [%r278+-640], %f213;
st.shared.f32 [%r278+-512], %f212;
st.shared.f32 [%r278+-384], %f211;
st.shared.f32 [%r278+-256], %f210;
st.shared.f32 [%r278+-128], %f209;

$L__BB286_5:
setp.ne.s32 %p5, %r1, 0;
bar.sync 0;
@%p5 bra $L__BB286_22;

mov.u32 %r279, %ctaid.z;
mov.u32 %r280, %ctaid.y;
mad.lo.s32 %r281, %r280, %r27, %r2;
mad.lo.s32 %r282, %r279, %r33, %r281;
cvt.u64.u32 %rd14, %r282;
shl.b32 %r284, %r35, 2;
mov.u32 %r285, _ZZ13mul_mat_vec_qIL9ggml_type20ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r286, %r285, %r284;
mov.u32 %r287, 2;
ld.shared.f32 %f81, [%r286];
add.ftz.f32 %f82, %f81, %f216;
ld.shared.f32 %f83, [%r286+1024];
add.ftz.f32 %f84, %f83, %f82;
ld.shared.f32 %f85, [%r286+2048];
add.ftz.f32 %f86, %f85, %f84;
mov.b32 %r288, %f86;
mov.u32 %r289, 31;
mov.u32 %r290, 16;
mov.u32 %r291, -1;
shfl.sync.bfly.b32 %r292|%p6, %r288, %r290, %r289, %r291;
mov.b32 %f87, %r292;
add.ftz.f32 %f88, %f86, %f87;
mov.b32 %r293, %f88;
mov.u32 %r294, 8;
shfl.sync.bfly.b32 %r295|%p7, %r293, %r294, %r289, %r291;
mov.b32 %f89, %r295;
add.ftz.f32 %f90, %f88, %f89;
mov.b32 %r296, %f90;
mov.u32 %r297, 4;
shfl.sync.bfly.b32 %r298|%p8, %r296, %r297, %r289, %r291;
mov.b32 %f91, %r298;
add.ftz.f32 %f92, %f90, %f91;
mov.b32 %r299, %f92;
shfl.sync.bfly.b32 %r300|%p9, %r299, %r287, %r289, %r291;
mov.b32 %f93, %r300;
add.ftz.f32 %f94, %f92, %f93;
mov.b32 %r301, %f94;
mov.u32 %r302, 1;
shfl.sync.bfly.b32 %r303|%p10, %r301, %r302, %r289, %r291;
mov.b32 %f95, %r303;
add.ftz.f32 %f25, %f94, %f95;
setp.ne.s32 %p11, %r35, 0;
setp.ge.u32 %p12, %r2, %r21;
cvta.to.global.u64 %rd51, %rd19;
mul.wide.u32 %rd52, %r282, 4;
add.s64 %rd15, %rd51, %rd52;
or.pred %p13, %p11, %p12;
@%p13 bra $L__BB286_8;

st.global.f32 [%rd15], %f25;

$L__BB286_8:
ld.shared.f32 %f96, [%r286+128];
add.ftz.f32 %f97, %f96, %f215;
ld.shared.f32 %f98, [%r286+1152];
add.ftz.f32 %f99, %f98, %f97;
ld.shared.f32 %f100, [%r286+2176];
add.ftz.f32 %f101, %f100, %f99;
mov.b32 %r308, %f101;
shfl.sync.bfly.b32 %r312|%p14, %r308, %r290, %r289, %r291;
mov.b32 %f102, %r312;
add.ftz.f32 %f103, %f101, %f102;
mov.b32 %r313, %f103;
shfl.sync.bfly.b32 %r315|%p15, %r313, %r294, %r289, %r291;
mov.b32 %f104, %r315;
add.ftz.f32 %f105, %f103, %f104;
mov.b32 %r316, %f105;
shfl.sync.bfly.b32 %r318|%p16, %r316, %r297, %r289, %r291;
mov.b32 %f106, %r318;
add.ftz.f32 %f107, %f105, %f106;
mov.b32 %r319, %f107;
shfl.sync.bfly.b32 %r320|%p17, %r319, %r287, %r289, %r291;
mov.b32 %f108, %r320;
add.ftz.f32 %f109, %f107, %f108;
mov.b32 %r321, %f109;
shfl.sync.bfly.b32 %r323|%p18, %r321, %r302, %r289, %r291;
mov.b32 %f110, %r323;
add.ftz.f32 %f26, %f109, %f110;
setp.eq.s32 %p19, %r35, 1;
or.b32 %r326, %r2, 1;
setp.lt.u32 %p20, %r326, %r21;
and.pred %p1, %p19, %p20;
not.pred %p21, %p1;
@%p21 bra $L__BB286_10;

st.global.f32 [%rd15+4], %f26;

$L__BB286_10:
ld.shared.f32 %f111, [%r286+256];
add.ftz.f32 %f112, %f111, %f214;
ld.shared.f32 %f113, [%r286+1280];
add.ftz.f32 %f114, %f113, %f112;
ld.shared.f32 %f115, [%r286+2304];
add.ftz.f32 %f116, %f115, %f114;
mov.b32 %r333, %f116;
mov.u32 %r334, 31;
mov.u32 %r335, 16;
mov.u32 %r336, -1;
shfl.sync.bfly.b32 %r337|%p22, %r333, %r335, %r334, %r336;
mov.b32 %f117, %r337;
add.ftz.f32 %f118, %f116, %f117;
mov.b32 %r338, %f118;
mov.u32 %r339, 8;
shfl.sync.bfly.b32 %r340|%p23, %r338, %r339, %r334, %r336;
mov.b32 %f119, %r340;
add.ftz.f32 %f120, %f118, %f119;
mov.b32 %r341, %f120;
mov.u32 %r342, 4;
shfl.sync.bfly.b32 %r343|%p24, %r341, %r342, %r334, %r336;
mov.b32 %f121, %r343;
add.ftz.f32 %f122, %f120, %f121;
mov.b32 %r344, %f122;
mov.u32 %r345, 2;
shfl.sync.bfly.b32 %r346|%p25, %r344, %r345, %r334, %r336;
mov.b32 %f123, %r346;
add.ftz.f32 %f124, %f122, %f123;
mov.b32 %r347, %f124;
mov.u32 %r348, 1;
shfl.sync.bfly.b32 %r349|%p26, %r347, %r348, %r334, %r336;
mov.b32 %f125, %r349;
add.ftz.f32 %f27, %f124, %f125;
@%p13 bra $L__BB286_12;

cvt.u64.u32 %rd56, %r21;
add.s64 %rd57, %rd56, %rd14;
shl.b64 %rd59, %rd57, 2;
add.s64 %rd60, %rd51, %rd59;
st.global.f32 [%rd60], %f27;

$L__BB286_12:
ld.shared.f32 %f126, [%r286+384];
add.ftz.f32 %f127, %f126, %f213;
ld.shared.f32 %f128, [%r286+1408];
add.ftz.f32 %f129, %f128, %f127;
ld.shared.f32 %f130, [%r286+2432];
add.ftz.f32 %f131, %f130, %f129;
mov.b32 %r353, %f131;
shfl.sync.bfly.b32 %r357|%p30, %r353, %r335, %r334, %r336;
mov.b32 %f132, %r357;
add.ftz.f32 %f133, %f131, %f132;
mov.b32 %r358, %f133;
shfl.sync.bfly.b32 %r360|%p31, %r358, %r339, %r334, %r336;
mov.b32 %f134, %r360;
add.ftz.f32 %f135, %f133, %f134;
mov.b32 %r361, %f135;
shfl.sync.bfly.b32 %r363|%p32, %r361, %r342, %r334, %r336;
mov.b32 %f136, %r363;
add.ftz.f32 %f137, %f135, %f136;
mov.b32 %r364, %f137;
shfl.sync.bfly.b32 %r366|%p33, %r364, %r345, %r334, %r336;
mov.b32 %f138, %r366;
add.ftz.f32 %f139, %f137, %f138;
mov.b32 %r367, %f139;
shfl.sync.bfly.b32 %r369|%p34, %r367, %r348, %r334, %r336;
mov.b32 %f140, %r369;
add.ftz.f32 %f28, %f139, %f140;
@%p21 bra $L__BB286_14;

add.s32 %r370, %r21, 1;
cvt.u64.u32 %rd61, %r370;
add.s64 %rd62, %rd61, %rd14;
shl.b64 %rd64, %rd62, 2;
add.s64 %rd65, %rd51, %rd64;
st.global.f32 [%rd65], %f28;

$L__BB286_14:
shl.b32 %r371, %r21, 1;
mov.u32 %r372, 1;
ld.shared.f32 %f141, [%r286+512];
add.ftz.f32 %f142, %f141, %f212;
ld.shared.f32 %f143, [%r286+1536];
add.ftz.f32 %f144, %f143, %f142;
ld.shared.f32 %f145, [%r286+2560];
add.ftz.f32 %f146, %f145, %f144;
mov.b32 %r373, %f146;
mov.u32 %r374, 31;
mov.u32 %r375, 16;
mov.u32 %r376, -1;
shfl.sync.bfly.b32 %r377|%p36, %r373, %r375, %r374, %r376;
mov.b32 %f147, %r377;
add.ftz.f32 %f148, %f146, %f147;
mov.b32 %r378, %f148;
mov.u32 %r379, 8;
shfl.sync.bfly.b32 %r380|%p37, %r378, %r379, %r374, %r376;
mov.b32 %f149, %r380;
add.ftz.f32 %f150, %f148, %f149;
mov.b32 %r381, %f150;
mov.u32 %r382, 4;
shfl.sync.bfly.b32 %r383|%p38, %r381, %r382, %r374, %r376;
mov.b32 %f151, %r383;
add.ftz.f32 %f152, %f150, %f151;
mov.b32 %r384, %f152;
mov.u32 %r385, 2;
shfl.sync.bfly.b32 %r386|%p39, %r384, %r385, %r374, %r376;
mov.b32 %f153, %r386;
add.ftz.f32 %f154, %f152, %f153;
mov.b32 %r387, %f154;
shfl.sync.bfly.b32 %r388|%p40, %r387, %r372, %r374, %r376;
mov.b32 %f155, %r388;
add.ftz.f32 %f29, %f154, %f155;
cvt.u64.u32 %rd66, %r371;
add.s64 %rd67, %rd66, %rd14;
shl.b64 %rd69, %rd67, 2;
add.s64 %rd16, %rd51, %rd69;
@%p13 bra $L__BB286_16;

st.global.f32 [%rd16], %f29;

$L__BB286_16:
ld.shared.f32 %f156, [%r286+640];
add.ftz.f32 %f157, %f156, %f211;
ld.shared.f32 %f158, [%r286+1664];
add.ftz.f32 %f159, %f158, %f157;
ld.shared.f32 %f160, [%r286+2688];
add.ftz.f32 %f161, %f160, %f159;
mov.b32 %r392, %f161;
shfl.sync.bfly.b32 %r396|%p44, %r392, %r375, %r374, %r376;
mov.b32 %f162, %r396;
add.ftz.f32 %f163, %f161, %f162;
mov.b32 %r397, %f163;
shfl.sync.bfly.b32 %r399|%p45, %r397, %r379, %r374, %r376;
mov.b32 %f164, %r399;
add.ftz.f32 %f165, %f163, %f164;
mov.b32 %r400, %f165;
shfl.sync.bfly.b32 %r402|%p46, %r400, %r382, %r374, %r376;
mov.b32 %f166, %r402;
add.ftz.f32 %f167, %f165, %f166;
mov.b32 %r403, %f167;
shfl.sync.bfly.b32 %r405|%p47, %r403, %r385, %r374, %r376;
mov.b32 %f168, %r405;
add.ftz.f32 %f169, %f167, %f168;
mov.b32 %r406, %f169;
shfl.sync.bfly.b32 %r408|%p48, %r406, %r372, %r374, %r376;
mov.b32 %f170, %r408;
add.ftz.f32 %f30, %f169, %f170;
@%p21 bra $L__BB286_18;

st.global.f32 [%rd16+4], %f30;

$L__BB286_18:
mul.lo.s32 %r17, %r21, 3;
ld.shared.f32 %f171, [%r286+768];
add.ftz.f32 %f172, %f171, %f210;
ld.shared.f32 %f173, [%r286+1792];
add.ftz.f32 %f174, %f173, %f172;
ld.shared.f32 %f175, [%r286+2816];
add.ftz.f32 %f176, %f175, %f174;
mov.b32 %r409, %f176;
mov.u32 %r410, 31;
mov.u32 %r411, 16;
mov.u32 %r412, -1;
shfl.sync.bfly.b32 %r413|%p50, %r409, %r411, %r410, %r412;
mov.b32 %f177, %r413;
add.ftz.f32 %f178, %f176, %f177;
mov.b32 %r414, %f178;
mov.u32 %r415, 8;
shfl.sync.bfly.b32 %r416|%p51, %r414, %r415, %r410, %r412;
mov.b32 %f179, %r416;
add.ftz.f32 %f180, %f178, %f179;
mov.b32 %r417, %f180;
mov.u32 %r418, 4;
shfl.sync.bfly.b32 %r419|%p52, %r417, %r418, %r410, %r412;
mov.b32 %f181, %r419;
add.ftz.f32 %f182, %f180, %f181;
mov.b32 %r420, %f182;
mov.u32 %r421, 2;
shfl.sync.bfly.b32 %r422|%p53, %r420, %r421, %r410, %r412;
mov.b32 %f183, %r422;
add.ftz.f32 %f184, %f182, %f183;
mov.b32 %r423, %f184;
mov.u32 %r424, 1;
shfl.sync.bfly.b32 %r425|%p54, %r423, %r424, %r410, %r412;
mov.b32 %f185, %r425;
add.ftz.f32 %f31, %f184, %f185;
@%p13 bra $L__BB286_20;

cvt.u64.u32 %rd70, %r17;
add.s64 %rd71, %rd70, %rd14;
shl.b64 %rd73, %rd71, 2;
add.s64 %rd74, %rd51, %rd73;
st.global.f32 [%rd74], %f31;

$L__BB286_20:
ld.shared.f32 %f186, [%r286+896];
add.ftz.f32 %f187, %f186, %f209;
ld.shared.f32 %f188, [%r286+1920];
add.ftz.f32 %f189, %f188, %f187;
ld.shared.f32 %f190, [%r286+2944];
add.ftz.f32 %f191, %f190, %f189;
mov.b32 %r429, %f191;
shfl.sync.bfly.b32 %r433|%p58, %r429, %r411, %r410, %r412;
mov.b32 %f192, %r433;
add.ftz.f32 %f193, %f191, %f192;
mov.b32 %r434, %f193;
shfl.sync.bfly.b32 %r436|%p59, %r434, %r415, %r410, %r412;
mov.b32 %f194, %r436;
add.ftz.f32 %f195, %f193, %f194;
mov.b32 %r437, %f195;
shfl.sync.bfly.b32 %r439|%p60, %r437, %r418, %r410, %r412;
mov.b32 %f196, %r439;
add.ftz.f32 %f197, %f195, %f196;
mov.b32 %r440, %f197;
shfl.sync.bfly.b32 %r442|%p61, %r440, %r421, %r410, %r412;
mov.b32 %f198, %r442;
add.ftz.f32 %f199, %f197, %f198;
mov.b32 %r443, %f199;
shfl.sync.bfly.b32 %r445|%p62, %r443, %r424, %r410, %r412;
mov.b32 %f200, %r445;
add.ftz.f32 %f32, %f199, %f200;
@%p21 bra $L__BB286_22;

add.s32 %r446, %r17, 1;
cvt.u64.u32 %rd75, %r446;
add.s64 %rd76, %rd75, %rd14;
shl.b64 %rd78, %rd76, 2;
add.s64 %rd79, %rd51, %rd78;
st.global.f32 [%rd79], %f32;

$L__BB286_22:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type20ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type20ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type20ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type20ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type20ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type20ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type20ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type20ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type20ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<78>;
.reg .b16 %rs<23>;
.reg .f32 %f<231>;
.reg .b32 %r<589>;
.reg .b64 %rd<98>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type20ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1280];

ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type20ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type20ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type20ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type20ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type20ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type20ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type20ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type20ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type20ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type20ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type20ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type20ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type20ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type20ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type20ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type20ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type20ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r36, %r1, 5;
mov.u32 %r37, %tid.x;
add.s32 %r2, %r36, %r37;
shr.u32 %r38, %r2, 31;
add.s32 %r39, %r2, %r38;
shr.s32 %r588, %r39, 1;
shr.u32 %r40, %r20, 5;
setp.ge.s32 %p2, %r588, %r40;
mov.f32 %f221, 0f00000000;
mov.f32 %f222, %f221;
mov.f32 %f223, %f221;
mov.f32 %f224, %f221;
mov.f32 %f225, %f221;
mov.f32 %f226, %f221;
mov.f32 %f227, %f221;
mov.f32 %f228, %f221;
mov.f32 %f229, %f221;
mov.f32 %f230, %f221;
@%p2 bra $L__BB287_3;

mov.u32 %r41, %ctaid.y;
mov.u32 %r42, %ctaid.z;
mul.hi.u32 %r43, %r42, %r30;
add.s32 %r44, %r43, %r42;
shr.u32 %r45, %r44, %r31;
mul.lo.s32 %r46, %r45, %r33;
mul.hi.u32 %r47, %r41, %r24;
add.s32 %r48, %r47, %r41;
shr.u32 %r49, %r48, %r25;
mul.lo.s32 %r50, %r49, %r27;
cvta.to.global.u64 %rd19, %rd16;
and.b32 %r53, %r39, -2;
sub.s32 %r54, %r2, %r53;
shl.b32 %r55, %r54, 1;
cvt.s64.s32 %rd1, %r55;
mov.u32 %r56, %ctaid.x;
shl.b32 %r57, %r56, 1;
shl.b32 %r58, %r54, 2;
shl.b32 %r59, %r22, 1;
add.s32 %r587, %r588, %r59;
mad.lo.s32 %r586, %r22, 3, %r588;
shl.b32 %r60, %r22, 2;
add.s32 %r585, %r588, %r60;
add.s32 %r584, %r588, %r22;
add.s32 %r61, %r588, %r46;
add.s32 %r62, %r61, %r50;
or.b32 %r63, %r57, 1;
mad.lo.s32 %r64, %r21, %r63, %r62;
mul.wide.s32 %rd20, %r64, 18;
add.s64 %rd97, %rd19, %rd20;
mad.lo.s32 %r65, %r57, %r21, %r588;
add.s32 %r66, %r65, %r46;
add.s32 %r67, %r66, %r50;
mul.wide.s32 %rd21, %r67, 18;
add.s64 %rd22, %rd19, %rd21;
add.s64 %rd96, %rd22, 4;
mul.wide.s32 %rd23, %r58, 2;
add.s64 %rd24, %rd21, %rd23;
add.s64 %rd25, %rd19, %rd24;
add.s64 %rd95, %rd25, 8;
or.b64 %rd5, %rd23, 4;
mul.lo.s32 %r68, %r41, %r28;
cvt.u64.u32 %rd26, %r68;
mul.lo.s32 %r69, %r42, %r34;
cvt.u64.u32 %rd27, %r69;
add.s64 %rd6, %rd27, %rd26;
cvta.to.global.u64 %rd7, %rd17;
mov.f32 %f221, 0f00000000;
shl.b64 %rd32, %rd1, 2;

$L__BB287_2:
cvt.u64.u32 %rd28, %r588;
add.s64 %rd29, %rd6, %rd28;
mul.lo.s64 %rd30, %rd29, 36;
add.s64 %rd31, %rd7, %rd30;
add.s64 %rd33, %rd31, %rd32;
ld.global.nc.u16 %rs11, [%rd95+-6];
cvt.u32.u16 %r240, %rs11;
ld.global.nc.u16 %rs12, [%rd95+-4];
mov.b32 %r241, {%rs11, %rs12};
and.b32 %r242, %r241, -2004318072;
shr.u32 %r243, %r242, 1;
or.b32 %r244, %r243, 839922192;
and.b32 %r245, %r240, 30583;
mov.u32 %r246, -152379953;
mov.u32 %r247, -1079142271;
prmt.b32 %r248, %r247, %r246, %r245;
mov.u32 %r249, 1901675829;
mov.u32 %r250, 639175937;
prmt.b32 %r251, %r250, %r249, %r245;
and.b32 %r252, %r244, 30292;
prmt.b32 %r253, %r248, %r251, %r252;
and.b16 %rs13, %rs12, 30583;
cvt.u32.u16 %r254, %rs13;
prmt.b32 %r255, %r247, %r246, %r254;
prmt.b32 %r256, %r250, %r249, %r254;
shr.u32 %r257, %r244, 16;
prmt.b32 %r258, %r255, %r256, %r257;
mov.u32 %r259, 25632;
prmt.b32 %r207, %r253, %r258, %r259;
mov.u32 %r260, 30001;
prmt.b32 %r211, %r253, %r258, %r260;
ld.global.nc.u32 %r72, [%rd33+4];
mov.u32 %r226, 0;

dp4a.s32.s32 %r70, %r207, %r72, %r226;

ld.global.nc.u32 %r76, [%rd33+20];

dp4a.s32.s32 %r74, %r211, %r76, %r70;

ld.global.nc.u16 %rs14, [%rd95+-2];
cvt.u32.u16 %r261, %rs14;
ld.global.nc.u16 %rs15, [%rd95];
mov.b32 %r262, {%rs14, %rs15};
and.b32 %r263, %r262, -2004318072;
shr.u32 %r264, %r263, 1;
or.b32 %r265, %r264, 839922192;
and.b32 %r266, %r261, 30583;
prmt.b32 %r267, %r247, %r246, %r266;
prmt.b32 %r268, %r250, %r249, %r266;
and.b32 %r269, %r265, 30292;
prmt.b32 %r270, %r267, %r268, %r269;
and.b16 %rs16, %rs15, 30583;
cvt.u32.u16 %r271, %rs16;
prmt.b32 %r272, %r247, %r246, %r271;
prmt.b32 %r273, %r250, %r249, %r271;
shr.u32 %r274, %r265, 16;
prmt.b32 %r275, %r272, %r273, %r274;
prmt.b32 %r215, %r270, %r275, %r259;
prmt.b32 %r219, %r270, %r275, %r260;
ld.global.nc.u32 %r80, [%rd33+8];

dp4a.s32.s32 %r78, %r215, %r80, %r74;

ld.global.nc.u32 %r84, [%rd33+24];

dp4a.s32.s32 %r82, %r219, %r84, %r78;

ld.global.nc.u16 %rs1, [%rd96+-4];

{ cvt.f32.f16 %f61, %rs1;}


ld.global.nc.u32 %r86, [%rd31];

{.reg .f16 low,high;
mov.b32 {low,high},%r86;
cvt.f32.f16 %f62, low;}


mul.ftz.f32 %f81, %f61, %f62;
cvt.rn.f32.s32 %f82, %r82;
fma.rn.ftz.f32 %f230, %f81, %f82, %f230;
add.s64 %rd34, %rd97, %rd5;
ld.global.nc.u16 %rs17, [%rd34+-2];
cvt.u32.u16 %r276, %rs17;
ld.global.nc.u16 %rs18, [%rd34];
mov.b32 %r277, {%rs17, %rs18};
and.b32 %r278, %r277, -2004318072;
shr.u32 %r279, %r278, 1;
or.b32 %r280, %r279, 839922192;
and.b32 %r281, %r276, 30583;
prmt.b32 %r282, %r247, %r246, %r281;
prmt.b32 %r283, %r250, %r249, %r281;
and.b32 %r284, %r280, 30292;
prmt.b32 %r285, %r282, %r283, %r284;
and.b16 %rs19, %rs18, 30583;
cvt.u32.u16 %r286, %rs19;
prmt.b32 %r287, %r247, %r246, %r286;
prmt.b32 %r288, %r250, %r249, %r286;
shr.u32 %r289, %r280, 16;
prmt.b32 %r290, %r287, %r288, %r289;
prmt.b32 %r224, %r285, %r290, %r259;
prmt.b32 %r228, %r285, %r290, %r260;

dp4a.s32.s32 %r87, %r224, %r72, %r226;


dp4a.s32.s32 %r91, %r228, %r76, %r87;

ld.global.nc.u16 %rs20, [%rd34+2];
cvt.u32.u16 %r291, %rs20;
ld.global.nc.u16 %rs21, [%rd34+4];
mov.b32 %r292, {%rs20, %rs21};
and.b32 %r293, %r292, -2004318072;
shr.u32 %r294, %r293, 1;
or.b32 %r295, %r294, 839922192;
and.b32 %r296, %r291, 30583;
prmt.b32 %r297, %r247, %r246, %r296;
prmt.b32 %r298, %r250, %r249, %r296;
and.b32 %r299, %r295, 30292;
prmt.b32 %r300, %r297, %r298, %r299;
and.b16 %rs22, %rs21, 30583;
cvt.u32.u16 %r301, %rs22;
prmt.b32 %r302, %r247, %r246, %r301;
prmt.b32 %r303, %r250, %r249, %r301;
shr.u32 %r304, %r295, 16;
prmt.b32 %r305, %r302, %r303, %r304;
prmt.b32 %r232, %r300, %r305, %r259;
prmt.b32 %r236, %r300, %r305, %r260;

dp4a.s32.s32 %r95, %r232, %r80, %r91;


dp4a.s32.s32 %r99, %r236, %r84, %r95;

ld.global.nc.u16 %rs2, [%rd97];

{ cvt.f32.f16 %f63, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r86;
cvt.f32.f16 %f64, low;}


mul.ftz.f32 %f83, %f63, %f64;
cvt.rn.f32.s32 %f84, %r99;
fma.rn.ftz.f32 %f229, %f83, %f84, %f229;
cvt.u64.u32 %rd35, %r584;
add.s64 %rd36, %rd6, %rd35;
mul.lo.s64 %rd37, %rd36, 36;
add.s64 %rd38, %rd7, %rd37;
add.s64 %rd39, %rd38, %rd32;
ld.global.nc.u32 %r106, [%rd39+4];

dp4a.s32.s32 %r104, %r207, %r106, %r226;

ld.global.nc.u32 %r110, [%rd39+20];

dp4a.s32.s32 %r108, %r211, %r110, %r104;

ld.global.nc.u32 %r114, [%rd39+8];

dp4a.s32.s32 %r112, %r215, %r114, %r108;

ld.global.nc.u32 %r118, [%rd39+24];

dp4a.s32.s32 %r116, %r219, %r118, %r112;


{ cvt.f32.f16 %f65, %rs1;}


ld.global.nc.u32 %r120, [%rd38];

{.reg .f16 low,high;
mov.b32 {low,high},%r120;
cvt.f32.f16 %f66, low;}


mul.ftz.f32 %f85, %f65, %f66;
cvt.rn.f32.s32 %f86, %r116;
fma.rn.ftz.f32 %f228, %f85, %f86, %f228;

dp4a.s32.s32 %r121, %r224, %r106, %r226;


dp4a.s32.s32 %r125, %r228, %r110, %r121;


dp4a.s32.s32 %r129, %r232, %r114, %r125;


dp4a.s32.s32 %r133, %r236, %r118, %r129;


{ cvt.f32.f16 %f67, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r120;
cvt.f32.f16 %f68, low;}


mul.ftz.f32 %f87, %f67, %f68;
cvt.rn.f32.s32 %f88, %r133;
fma.rn.ftz.f32 %f227, %f87, %f88, %f227;
cvt.u64.u32 %rd40, %r587;
add.s64 %rd41, %rd6, %rd40;
mul.lo.s64 %rd42, %rd41, 36;
add.s64 %rd43, %rd7, %rd42;
add.s64 %rd44, %rd43, %rd32;
ld.global.nc.u32 %r140, [%rd44+4];

dp4a.s32.s32 %r138, %r207, %r140, %r226;

ld.global.nc.u32 %r144, [%rd44+20];

dp4a.s32.s32 %r142, %r211, %r144, %r138;

ld.global.nc.u32 %r148, [%rd44+8];

dp4a.s32.s32 %r146, %r215, %r148, %r142;

ld.global.nc.u32 %r152, [%rd44+24];

dp4a.s32.s32 %r150, %r219, %r152, %r146;


{ cvt.f32.f16 %f69, %rs1;}


ld.global.nc.u32 %r154, [%rd43];

{.reg .f16 low,high;
mov.b32 {low,high},%r154;
cvt.f32.f16 %f70, low;}


mul.ftz.f32 %f89, %f69, %f70;
cvt.rn.f32.s32 %f90, %r150;
fma.rn.ftz.f32 %f226, %f89, %f90, %f226;

dp4a.s32.s32 %r155, %r224, %r140, %r226;


dp4a.s32.s32 %r159, %r228, %r144, %r155;


dp4a.s32.s32 %r163, %r232, %r148, %r159;


dp4a.s32.s32 %r167, %r236, %r152, %r163;


{ cvt.f32.f16 %f71, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r154;
cvt.f32.f16 %f72, low;}


mul.ftz.f32 %f91, %f71, %f72;
cvt.rn.f32.s32 %f92, %r167;
fma.rn.ftz.f32 %f225, %f91, %f92, %f225;
cvt.u64.u32 %rd45, %r586;
add.s64 %rd46, %rd6, %rd45;
mul.lo.s64 %rd47, %rd46, 36;
add.s64 %rd48, %rd7, %rd47;
add.s64 %rd49, %rd48, %rd32;
ld.global.nc.u32 %r174, [%rd49+4];

dp4a.s32.s32 %r172, %r207, %r174, %r226;

ld.global.nc.u32 %r178, [%rd49+20];

dp4a.s32.s32 %r176, %r211, %r178, %r172;

ld.global.nc.u32 %r182, [%rd49+8];

dp4a.s32.s32 %r180, %r215, %r182, %r176;

ld.global.nc.u32 %r186, [%rd49+24];

dp4a.s32.s32 %r184, %r219, %r186, %r180;


{ cvt.f32.f16 %f73, %rs1;}


ld.global.nc.u32 %r188, [%rd48];

{.reg .f16 low,high;
mov.b32 {low,high},%r188;
cvt.f32.f16 %f74, low;}


mul.ftz.f32 %f93, %f73, %f74;
cvt.rn.f32.s32 %f94, %r184;
fma.rn.ftz.f32 %f224, %f93, %f94, %f224;

dp4a.s32.s32 %r189, %r224, %r174, %r226;


dp4a.s32.s32 %r193, %r228, %r178, %r189;


dp4a.s32.s32 %r197, %r232, %r182, %r193;


dp4a.s32.s32 %r201, %r236, %r186, %r197;


{ cvt.f32.f16 %f75, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r188;
cvt.f32.f16 %f76, low;}


mul.ftz.f32 %f95, %f75, %f76;
cvt.rn.f32.s32 %f96, %r201;
fma.rn.ftz.f32 %f223, %f95, %f96, %f223;
cvt.u64.u32 %rd50, %r585;
add.s64 %rd51, %rd6, %rd50;
mul.lo.s64 %rd52, %rd51, 36;
add.s64 %rd53, %rd7, %rd52;
add.s64 %rd54, %rd53, %rd32;
ld.global.nc.u32 %r208, [%rd54+4];

dp4a.s32.s32 %r206, %r207, %r208, %r226;

ld.global.nc.u32 %r212, [%rd54+20];

dp4a.s32.s32 %r210, %r211, %r212, %r206;

ld.global.nc.u32 %r216, [%rd54+8];

dp4a.s32.s32 %r214, %r215, %r216, %r210;

ld.global.nc.u32 %r220, [%rd54+24];

dp4a.s32.s32 %r218, %r219, %r220, %r214;


{ cvt.f32.f16 %f77, %rs1;}


ld.global.nc.u32 %r222, [%rd53];

{.reg .f16 low,high;
mov.b32 {low,high},%r222;
cvt.f32.f16 %f78, low;}


mul.ftz.f32 %f97, %f77, %f78;
cvt.rn.f32.s32 %f98, %r218;
fma.rn.ftz.f32 %f222, %f97, %f98, %f222;

dp4a.s32.s32 %r223, %r224, %r208, %r226;


dp4a.s32.s32 %r227, %r228, %r212, %r223;


dp4a.s32.s32 %r231, %r232, %r216, %r227;


dp4a.s32.s32 %r235, %r236, %r220, %r231;


{ cvt.f32.f16 %f79, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r222;
cvt.f32.f16 %f80, low;}


mul.ftz.f32 %f99, %f79, %f80;
cvt.rn.f32.s32 %f100, %r235;
fma.rn.ftz.f32 %f221, %f99, %f100, %f221;
add.s32 %r587, %r587, 32;
add.s32 %r586, %r586, 32;
add.s32 %r585, %r585, 32;
add.s32 %r584, %r584, 32;
add.s64 %rd97, %rd97, 576;
add.s64 %rd96, %rd96, 576;
add.s64 %rd95, %rd95, 576;
add.s32 %r588, %r588, 32;
setp.lt.s32 %p3, %r588, %r40;
@%p3 bra $L__BB287_2;

$L__BB287_3:
setp.eq.s32 %p4, %r1, 0;
@%p4 bra $L__BB287_5;

mov.u32 %r308, _ZZ13mul_mat_vec_qIL9ggml_type20ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r309, %r1, 1280, %r308;
shl.b32 %r310, %r37, 2;
add.s32 %r311, %r309, %r310;
st.shared.f32 [%r311+-1280], %f230;
st.shared.f32 [%r311+-1152], %f229;
st.shared.f32 [%r311+-1024], %f228;
st.shared.f32 [%r311+-896], %f227;
st.shared.f32 [%r311+-768], %f226;
st.shared.f32 [%r311+-640], %f225;
st.shared.f32 [%r311+-512], %f224;
st.shared.f32 [%r311+-384], %f223;
st.shared.f32 [%r311+-256], %f222;
st.shared.f32 [%r311+-128], %f221;

$L__BB287_5:
setp.ne.s32 %p5, %r1, 0;
bar.sync 0;
@%p5 bra $L__BB287_26;

mov.u32 %r313, %ctaid.x;
shl.b32 %r314, %r313, 1;
mov.u32 %r315, 1;
shl.b32 %r317, %r37, 2;
mov.u32 %r318, _ZZ13mul_mat_vec_qIL9ggml_type20ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r319, %r318, %r317;
mov.u32 %r320, 2;
ld.shared.f32 %f101, [%r319];
add.ftz.f32 %f102, %f101, %f230;
mov.b32 %r321, %f102;
mov.u32 %r322, 31;
mov.u32 %r323, 16;
mov.u32 %r324, -1;
shfl.sync.bfly.b32 %r325|%p6, %r321, %r323, %r322, %r324;
mov.b32 %f103, %r325;
add.ftz.f32 %f104, %f102, %f103;
mov.b32 %r326, %f104;
mov.u32 %r327, 8;
shfl.sync.bfly.b32 %r328|%p7, %r326, %r327, %r322, %r324;
mov.b32 %f105, %r328;
add.ftz.f32 %f106, %f104, %f105;
mov.b32 %r329, %f106;
mov.u32 %r330, 4;
shfl.sync.bfly.b32 %r331|%p8, %r329, %r330, %r322, %r324;
mov.b32 %f107, %r331;
add.ftz.f32 %f108, %f106, %f107;
mov.b32 %r332, %f108;
shfl.sync.bfly.b32 %r333|%p9, %r332, %r320, %r322, %r324;
mov.b32 %f109, %r333;
add.ftz.f32 %f110, %f108, %f109;
mov.b32 %r334, %f110;
shfl.sync.bfly.b32 %r335|%p10, %r334, %r315, %r322, %r324;
mov.b32 %f111, %r335;
add.ftz.f32 %f31, %f110, %f111;
setp.ne.s32 %p11, %r37, 0;
setp.ge.u32 %p12, %r314, %r23;
or.pred %p13, %p11, %p12;
@%p13 bra $L__BB287_8;

cvta.to.global.u64 %rd55, %rd18;
mov.u32 %r336, %ctaid.y;
mad.lo.s32 %r339, %r336, %r29, %r314;
mov.u32 %r340, %ctaid.z;
mad.lo.s32 %r341, %r340, %r35, %r339;
mul.wide.u32 %rd56, %r341, 4;
add.s64 %rd57, %rd55, %rd56;
st.global.f32 [%rd57], %f31;

$L__BB287_8:
ld.shared.f32 %f112, [%r319+128];
add.ftz.f32 %f113, %f112, %f229;
mov.b32 %r347, %f113;
shfl.sync.bfly.b32 %r351|%p14, %r347, %r323, %r322, %r324;
mov.b32 %f114, %r351;
add.ftz.f32 %f115, %f113, %f114;
mov.b32 %r352, %f115;
shfl.sync.bfly.b32 %r354|%p15, %r352, %r327, %r322, %r324;
mov.b32 %f116, %r354;
add.ftz.f32 %f117, %f115, %f116;
mov.b32 %r355, %f117;
shfl.sync.bfly.b32 %r357|%p16, %r355, %r330, %r322, %r324;
mov.b32 %f118, %r357;
add.ftz.f32 %f119, %f117, %f118;
mov.b32 %r358, %f119;
shfl.sync.bfly.b32 %r359|%p17, %r358, %r320, %r322, %r324;
mov.b32 %f120, %r359;
add.ftz.f32 %f121, %f119, %f120;
mov.b32 %r360, %f121;
shfl.sync.bfly.b32 %r362|%p18, %r360, %r315, %r322, %r324;
mov.b32 %f122, %r362;
add.ftz.f32 %f32, %f121, %f122;
setp.eq.s32 %p19, %r37, 1;
or.b32 %r365, %r314, 1;
setp.lt.u32 %p20, %r365, %r23;
and.pred %p1, %p19, %p20;
not.pred %p21, %p1;
@%p21 bra $L__BB287_10;

cvta.to.global.u64 %rd58, %rd18;
mov.u32 %r366, %ctaid.y;
mad.lo.s32 %r369, %r366, %r29, %r314;
mov.u32 %r370, %ctaid.z;
mad.lo.s32 %r371, %r370, %r35, %r369;
mul.wide.u32 %rd59, %r371, 4;
add.s64 %rd60, %rd58, %rd59;
st.global.f32 [%rd60+4], %f32;

$L__BB287_10:
mov.u32 %r376, 2;
ld.shared.f32 %f123, [%r319+256];
add.ftz.f32 %f124, %f123, %f228;
mov.b32 %r377, %f124;
mov.u32 %r378, 31;
mov.u32 %r379, 16;
mov.u32 %r380, -1;
shfl.sync.bfly.b32 %r381|%p22, %r377, %r379, %r378, %r380;
mov.b32 %f125, %r381;
add.ftz.f32 %f126, %f124, %f125;
mov.b32 %r382, %f126;
mov.u32 %r383, 8;
shfl.sync.bfly.b32 %r384|%p23, %r382, %r383, %r378, %r380;
mov.b32 %f127, %r384;
add.ftz.f32 %f128, %f126, %f127;
mov.b32 %r385, %f128;
mov.u32 %r386, 4;
shfl.sync.bfly.b32 %r387|%p24, %r385, %r386, %r378, %r380;
mov.b32 %f129, %r387;
add.ftz.f32 %f130, %f128, %f129;
mov.b32 %r388, %f130;
shfl.sync.bfly.b32 %r389|%p25, %r388, %r376, %r378, %r380;
mov.b32 %f131, %r389;
add.ftz.f32 %f132, %f130, %f131;
mov.b32 %r390, %f132;
mov.u32 %r391, 1;
shfl.sync.bfly.b32 %r392|%p26, %r390, %r391, %r378, %r380;
mov.b32 %f133, %r392;
add.ftz.f32 %f33, %f132, %f133;
@%p13 bra $L__BB287_12;

cvt.u64.u32 %rd61, %r23;
mov.u32 %r395, %ctaid.y;
mad.lo.s32 %r398, %r395, %r29, %r314;
mov.u32 %r399, %ctaid.z;
mad.lo.s32 %r400, %r399, %r35, %r398;
cvt.u64.u32 %rd62, %r400;
add.s64 %rd63, %rd61, %rd62;
cvta.to.global.u64 %rd64, %rd18;
shl.b64 %rd65, %rd63, 2;
add.s64 %rd66, %rd64, %rd65;
st.global.f32 [%rd66], %f33;

$L__BB287_12:
ld.shared.f32 %f134, [%r319+384];
add.ftz.f32 %f135, %f134, %f227;
mov.b32 %r406, %f135;
shfl.sync.bfly.b32 %r410|%p30, %r406, %r379, %r378, %r380;
mov.b32 %f136, %r410;
add.ftz.f32 %f137, %f135, %f136;
mov.b32 %r411, %f137;
shfl.sync.bfly.b32 %r413|%p31, %r411, %r383, %r378, %r380;
mov.b32 %f138, %r413;
add.ftz.f32 %f139, %f137, %f138;
mov.b32 %r414, %f139;
shfl.sync.bfly.b32 %r416|%p32, %r414, %r386, %r378, %r380;
mov.b32 %f140, %r416;
add.ftz.f32 %f141, %f139, %f140;
mov.b32 %r417, %f141;
shfl.sync.bfly.b32 %r418|%p33, %r417, %r376, %r378, %r380;
mov.b32 %f142, %r418;
add.ftz.f32 %f143, %f141, %f142;
mov.b32 %r419, %f143;
shfl.sync.bfly.b32 %r421|%p34, %r419, %r391, %r378, %r380;
mov.b32 %f144, %r421;
add.ftz.f32 %f34, %f143, %f144;
@%p21 bra $L__BB287_14;

add.s32 %r422, %r23, 1;
cvt.u64.u32 %rd67, %r422;
mov.u32 %r423, %ctaid.y;
mad.lo.s32 %r426, %r423, %r29, %r314;
mov.u32 %r427, %ctaid.z;
mad.lo.s32 %r428, %r427, %r35, %r426;
cvt.u64.u32 %rd68, %r428;
add.s64 %rd69, %rd67, %rd68;
cvta.to.global.u64 %rd70, %rd18;
shl.b64 %rd71, %rd69, 2;
add.s64 %rd72, %rd70, %rd71;
st.global.f32 [%rd72], %f34;

$L__BB287_14:
shl.b32 %r429, %r23, 1;
mov.u32 %r430, 1;
mov.u32 %r435, 2;
ld.shared.f32 %f145, [%r319+512];
add.ftz.f32 %f146, %f145, %f226;
mov.b32 %r436, %f146;
mov.u32 %r437, 31;
mov.u32 %r438, 16;
mov.u32 %r439, -1;
shfl.sync.bfly.b32 %r440|%p36, %r436, %r438, %r437, %r439;
mov.b32 %f147, %r440;
add.ftz.f32 %f148, %f146, %f147;
mov.b32 %r441, %f148;
mov.u32 %r442, 8;
shfl.sync.bfly.b32 %r443|%p37, %r441, %r442, %r437, %r439;
mov.b32 %f149, %r443;
add.ftz.f32 %f150, %f148, %f149;
mov.b32 %r444, %f150;
mov.u32 %r445, 4;
shfl.sync.bfly.b32 %r446|%p38, %r444, %r445, %r437, %r439;
mov.b32 %f151, %r446;
add.ftz.f32 %f152, %f150, %f151;
mov.b32 %r447, %f152;
shfl.sync.bfly.b32 %r448|%p39, %r447, %r435, %r437, %r439;
mov.b32 %f153, %r448;
add.ftz.f32 %f154, %f152, %f153;
mov.b32 %r449, %f154;
shfl.sync.bfly.b32 %r450|%p40, %r449, %r430, %r437, %r439;
mov.b32 %f155, %r450;
add.ftz.f32 %f35, %f154, %f155;
cvt.u64.u32 %rd73, %r429;
mov.u32 %r451, %ctaid.y;
mad.lo.s32 %r454, %r451, %r29, %r314;
mov.u32 %r455, %ctaid.z;
mad.lo.s32 %r456, %r455, %r35, %r454;
cvt.u64.u32 %rd74, %r456;
add.s64 %rd75, %rd73, %rd74;
cvta.to.global.u64 %rd76, %rd18;
shl.b64 %rd77, %rd75, 2;
add.s64 %rd14, %rd76, %rd77;
@%p13 bra $L__BB287_16;

st.global.f32 [%rd14], %f35;

$L__BB287_16:
ld.shared.f32 %f156, [%r319+640];
add.ftz.f32 %f157, %f156, %f225;
mov.b32 %r462, %f157;
shfl.sync.bfly.b32 %r466|%p44, %r462, %r438, %r437, %r439;
mov.b32 %f158, %r466;
add.ftz.f32 %f159, %f157, %f158;
mov.b32 %r467, %f159;
shfl.sync.bfly.b32 %r469|%p45, %r467, %r442, %r437, %r439;
mov.b32 %f160, %r469;
add.ftz.f32 %f161, %f159, %f160;
mov.b32 %r470, %f161;
shfl.sync.bfly.b32 %r472|%p46, %r470, %r445, %r437, %r439;
mov.b32 %f162, %r472;
add.ftz.f32 %f163, %f161, %f162;
mov.b32 %r473, %f163;
shfl.sync.bfly.b32 %r474|%p47, %r473, %r435, %r437, %r439;
mov.b32 %f164, %r474;
add.ftz.f32 %f165, %f163, %f164;
mov.b32 %r475, %f165;
shfl.sync.bfly.b32 %r477|%p48, %r475, %r430, %r437, %r439;
mov.b32 %f166, %r477;
add.ftz.f32 %f36, %f165, %f166;
@%p21 bra $L__BB287_18;

st.global.f32 [%rd14+4], %f36;

$L__BB287_18:
mul.lo.s32 %r19, %r23, 3;
mov.u32 %r482, 2;
ld.shared.f32 %f167, [%r319+768];
add.ftz.f32 %f168, %f167, %f224;
mov.b32 %r483, %f168;
mov.u32 %r484, 31;
mov.u32 %r485, 16;
mov.u32 %r486, -1;
shfl.sync.bfly.b32 %r487|%p50, %r483, %r485, %r484, %r486;
mov.b32 %f169, %r487;
add.ftz.f32 %f170, %f168, %f169;
mov.b32 %r488, %f170;
mov.u32 %r489, 8;
shfl.sync.bfly.b32 %r490|%p51, %r488, %r489, %r484, %r486;
mov.b32 %f171, %r490;
add.ftz.f32 %f172, %f170, %f171;
mov.b32 %r491, %f172;
mov.u32 %r492, 4;
shfl.sync.bfly.b32 %r493|%p52, %r491, %r492, %r484, %r486;
mov.b32 %f173, %r493;
add.ftz.f32 %f174, %f172, %f173;
mov.b32 %r494, %f174;
shfl.sync.bfly.b32 %r495|%p53, %r494, %r482, %r484, %r486;
mov.b32 %f175, %r495;
add.ftz.f32 %f176, %f174, %f175;
mov.b32 %r496, %f176;
mov.u32 %r497, 1;
shfl.sync.bfly.b32 %r498|%p54, %r496, %r497, %r484, %r486;
mov.b32 %f177, %r498;
add.ftz.f32 %f37, %f176, %f177;
@%p13 bra $L__BB287_20;

cvt.u64.u32 %rd78, %r19;
add.s64 %rd80, %rd78, %rd74;
shl.b64 %rd82, %rd80, 2;
add.s64 %rd83, %rd76, %rd82;
st.global.f32 [%rd83], %f37;

$L__BB287_20:
ld.shared.f32 %f178, [%r319+896];
add.ftz.f32 %f179, %f178, %f223;
mov.b32 %r512, %f179;
shfl.sync.bfly.b32 %r516|%p58, %r512, %r485, %r484, %r486;
mov.b32 %f180, %r516;
add.ftz.f32 %f181, %f179, %f180;
mov.b32 %r517, %f181;
shfl.sync.bfly.b32 %r519|%p59, %r517, %r489, %r484, %r486;
mov.b32 %f182, %r519;
add.ftz.f32 %f183, %f181, %f182;
mov.b32 %r520, %f183;
shfl.sync.bfly.b32 %r522|%p60, %r520, %r492, %r484, %r486;
mov.b32 %f184, %r522;
add.ftz.f32 %f185, %f183, %f184;
mov.b32 %r523, %f185;
shfl.sync.bfly.b32 %r524|%p61, %r523, %r482, %r484, %r486;
mov.b32 %f186, %r524;
add.ftz.f32 %f187, %f185, %f186;
mov.b32 %r525, %f187;
shfl.sync.bfly.b32 %r527|%p62, %r525, %r497, %r484, %r486;
mov.b32 %f188, %r527;
add.ftz.f32 %f38, %f187, %f188;
@%p21 bra $L__BB287_22;

add.s32 %r528, %r19, 1;
cvt.u64.u32 %rd84, %r528;
add.s64 %rd86, %rd84, %rd74;
shl.b64 %rd88, %rd86, 2;
add.s64 %rd89, %rd76, %rd88;
st.global.f32 [%rd89], %f38;

$L__BB287_22:
shl.b32 %r535, %r23, 2;
mov.u32 %r536, 2;
ld.shared.f32 %f189, [%r319+1024];
add.ftz.f32 %f190, %f189, %f222;
mov.b32 %r541, %f190;
mov.u32 %r542, 31;
mov.u32 %r543, 16;
mov.u32 %r544, -1;
shfl.sync.bfly.b32 %r545|%p64, %r541, %r543, %r542, %r544;
mov.b32 %f191, %r545;
add.ftz.f32 %f192, %f190, %f191;
mov.b32 %r546, %f192;
mov.u32 %r547, 8;
shfl.sync.bfly.b32 %r548|%p65, %r546, %r547, %r542, %r544;
mov.b32 %f193, %r548;
add.ftz.f32 %f194, %f192, %f193;
mov.b32 %r549, %f194;
mov.u32 %r550, 4;
shfl.sync.bfly.b32 %r551|%p66, %r549, %r550, %r542, %r544;
mov.b32 %f195, %r551;
add.ftz.f32 %f196, %f194, %f195;
mov.b32 %r552, %f196;
shfl.sync.bfly.b32 %r553|%p67, %r552, %r536, %r542, %r544;
mov.b32 %f197, %r553;
add.ftz.f32 %f198, %f196, %f197;
mov.b32 %r554, %f198;
mov.u32 %r555, 1;
shfl.sync.bfly.b32 %r556|%p68, %r554, %r555, %r542, %r544;
mov.b32 %f199, %r556;
add.ftz.f32 %f39, %f198, %f199;
cvt.u64.u32 %rd90, %r535;
add.s64 %rd92, %rd90, %rd74;
shl.b64 %rd94, %rd92, 2;
add.s64 %rd15, %rd76, %rd94;
@%p13 bra $L__BB287_24;

st.global.f32 [%rd15], %f39;

$L__BB287_24:
ld.shared.f32 %f200, [%r319+1152];
add.ftz.f32 %f201, %f200, %f221;
mov.b32 %r568, %f201;
shfl.sync.bfly.b32 %r572|%p72, %r568, %r543, %r542, %r544;
mov.b32 %f202, %r572;
add.ftz.f32 %f203, %f201, %f202;
mov.b32 %r573, %f203;
shfl.sync.bfly.b32 %r575|%p73, %r573, %r547, %r542, %r544;
mov.b32 %f204, %r575;
add.ftz.f32 %f205, %f203, %f204;
mov.b32 %r576, %f205;
shfl.sync.bfly.b32 %r578|%p74, %r576, %r550, %r542, %r544;
mov.b32 %f206, %r578;
add.ftz.f32 %f207, %f205, %f206;
mov.b32 %r579, %f207;
shfl.sync.bfly.b32 %r580|%p75, %r579, %r536, %r542, %r544;
mov.b32 %f208, %r580;
add.ftz.f32 %f209, %f207, %f208;
mov.b32 %r581, %f209;
shfl.sync.bfly.b32 %r583|%p76, %r581, %r555, %r542, %r544;
mov.b32 %f210, %r583;
add.ftz.f32 %f40, %f209, %f210;
@%p21 bra $L__BB287_26;

st.global.f32 [%rd15+4], %f40;

$L__BB287_26:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type20ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type20ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type20ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type20ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type20ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type20ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type20ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type20ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type20ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<83>;
.reg .b16 %rs<25>;
.reg .f32 %f<277>;
.reg .b32 %r<585>;
.reg .b64 %rd<108>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type20ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type20ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type20ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type20ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type20ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type20ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type20ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type20ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type20ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type20ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type20ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type20ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type20ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type20ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type20ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type20ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type20ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type20ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r43, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r43, %r2;
mov.u32 %r44, %ctaid.x;
shl.b32 %r4, %r44, 1;
shr.u32 %r45, %r3, 31;
add.s32 %r46, %r3, %r45;
shr.s32 %r584, %r46, 1;
shr.u32 %r6, %r42, 5;
setp.ge.s32 %p3, %r584, %r6;
mov.f32 %f265, 0f00000000;
mov.f32 %f266, %f265;
mov.f32 %f267, %f265;
mov.f32 %f268, %f265;
mov.f32 %f269, %f265;
mov.f32 %f270, %f265;
mov.f32 %f271, %f265;
mov.f32 %f272, %f265;
mov.f32 %f273, %f265;
mov.f32 %f274, %f265;
mov.f32 %f275, %f265;
mov.f32 %f276, %f265;
@%p3 bra $L__BB288_3;

mov.u32 %r47, %ctaid.y;
mov.u32 %r48, %ctaid.z;
mul.lo.s32 %r49, %r47, %r34;
cvt.u64.u32 %rd20, %r49;
mul.lo.s32 %r50, %r48, %r40;
cvt.u64.u32 %rd21, %r50;
add.s64 %rd1, %rd21, %rd20;
mul.hi.u32 %r51, %r48, %r36;
add.s32 %r52, %r51, %r48;
shr.u32 %r53, %r52, %r37;
mul.lo.s32 %r54, %r53, %r39;
mul.hi.u32 %r55, %r47, %r30;
add.s32 %r56, %r55, %r47;
shr.u32 %r57, %r56, %r31;
mul.lo.s32 %r58, %r57, %r33;
cvta.to.global.u64 %rd22, %rd17;
and.b32 %r61, %r46, -2;
sub.s32 %r62, %r3, %r61;
shl.b32 %r63, %r62, 1;
cvt.s64.s32 %rd2, %r63;
shl.b32 %r64, %r62, 2;
shl.b32 %r65, %r28, 1;
add.s32 %r583, %r584, %r65;
mad.lo.s32 %r582, %r28, 3, %r584;
shl.b32 %r66, %r28, 2;
add.s32 %r581, %r584, %r66;
mad.lo.s32 %r580, %r28, 5, %r584;
add.s32 %r579, %r584, %r28;
add.s32 %r67, %r584, %r54;
add.s32 %r68, %r67, %r58;
add.s32 %r69, %r4, 1;
mad.lo.s32 %r70, %r27, %r69, %r68;
mul.wide.s32 %rd23, %r70, 18;
add.s64 %rd107, %rd22, %rd23;
mad.lo.s32 %r71, %r4, %r27, %r584;
add.s32 %r72, %r71, %r54;
add.s32 %r73, %r72, %r58;
mul.wide.s32 %rd24, %r73, 18;
add.s64 %rd25, %rd22, %rd24;
add.s64 %rd106, %rd25, 4;
mul.wide.s32 %rd26, %r64, 2;
add.s64 %rd27, %rd24, %rd26;
add.s64 %rd28, %rd22, %rd27;
add.s64 %rd105, %rd28, 8;
or.b64 %rd6, %rd26, 4;
cvta.to.global.u64 %rd7, %rd18;
mov.f32 %f265, 0f00000000;
shl.b64 %rd33, %rd2, 2;

$L__BB288_2:
cvt.u64.u32 %rd29, %r584;
add.s64 %rd30, %rd1, %rd29;
mul.lo.s64 %rd31, %rd30, 36;
add.s64 %rd32, %rd7, %rd31;
add.s64 %rd34, %rd32, %rd33;
ld.global.nc.u16 %rs13, [%rd105+-6];
cvt.u32.u16 %r278, %rs13;
ld.global.nc.u16 %rs14, [%rd105+-4];
mov.b32 %r279, {%rs13, %rs14};
and.b32 %r280, %r279, -2004318072;
shr.u32 %r281, %r280, 1;
or.b32 %r282, %r281, 839922192;
and.b32 %r283, %r278, 30583;
mov.u32 %r284, -152379953;
mov.u32 %r285, -1079142271;
prmt.b32 %r286, %r285, %r284, %r283;
mov.u32 %r287, 1901675829;
mov.u32 %r288, 639175937;
prmt.b32 %r289, %r288, %r287, %r283;
and.b32 %r290, %r282, 30292;
prmt.b32 %r291, %r286, %r289, %r290;
and.b16 %rs15, %rs14, 30583;
cvt.u32.u16 %r292, %rs15;
prmt.b32 %r293, %r285, %r284, %r292;
prmt.b32 %r294, %r288, %r287, %r292;
shr.u32 %r295, %r282, 16;
prmt.b32 %r296, %r293, %r294, %r295;
mov.u32 %r297, 25632;
prmt.b32 %r245, %r291, %r296, %r297;
mov.u32 %r298, 30001;
prmt.b32 %r249, %r291, %r296, %r298;
ld.global.nc.u32 %r76, [%rd34+4];
mov.u32 %r264, 0;

dp4a.s32.s32 %r74, %r245, %r76, %r264;

ld.global.nc.u32 %r80, [%rd34+20];

dp4a.s32.s32 %r78, %r249, %r80, %r74;

ld.global.nc.u16 %rs16, [%rd105+-2];
cvt.u32.u16 %r299, %rs16;
ld.global.nc.u16 %rs17, [%rd105];
mov.b32 %r300, {%rs16, %rs17};
and.b32 %r301, %r300, -2004318072;
shr.u32 %r302, %r301, 1;
or.b32 %r303, %r302, 839922192;
and.b32 %r304, %r299, 30583;
prmt.b32 %r305, %r285, %r284, %r304;
prmt.b32 %r306, %r288, %r287, %r304;
and.b32 %r307, %r303, 30292;
prmt.b32 %r308, %r305, %r306, %r307;
and.b16 %rs18, %rs17, 30583;
cvt.u32.u16 %r309, %rs18;
prmt.b32 %r310, %r285, %r284, %r309;
prmt.b32 %r311, %r288, %r287, %r309;
shr.u32 %r312, %r303, 16;
prmt.b32 %r313, %r310, %r311, %r312;
prmt.b32 %r253, %r308, %r313, %r297;
prmt.b32 %r257, %r308, %r313, %r298;
ld.global.nc.u32 %r84, [%rd34+8];

dp4a.s32.s32 %r82, %r253, %r84, %r78;

ld.global.nc.u32 %r88, [%rd34+24];

dp4a.s32.s32 %r86, %r257, %r88, %r82;

ld.global.nc.u16 %rs1, [%rd106+-4];

{ cvt.f32.f16 %f73, %rs1;}


ld.global.nc.u32 %r90, [%rd32];

{.reg .f16 low,high;
mov.b32 {low,high},%r90;
cvt.f32.f16 %f74, low;}


mul.ftz.f32 %f97, %f73, %f74;
cvt.rn.f32.s32 %f98, %r86;
fma.rn.ftz.f32 %f276, %f97, %f98, %f276;
add.s64 %rd35, %rd107, %rd6;
ld.global.nc.u16 %rs19, [%rd35+-2];
cvt.u32.u16 %r314, %rs19;
ld.global.nc.u16 %rs20, [%rd35];
mov.b32 %r315, {%rs19, %rs20};
and.b32 %r316, %r315, -2004318072;
shr.u32 %r317, %r316, 1;
or.b32 %r318, %r317, 839922192;
and.b32 %r319, %r314, 30583;
prmt.b32 %r320, %r285, %r284, %r319;
prmt.b32 %r321, %r288, %r287, %r319;
and.b32 %r322, %r318, 30292;
prmt.b32 %r323, %r320, %r321, %r322;
and.b16 %rs21, %rs20, 30583;
cvt.u32.u16 %r324, %rs21;
prmt.b32 %r325, %r285, %r284, %r324;
prmt.b32 %r326, %r288, %r287, %r324;
shr.u32 %r327, %r318, 16;
prmt.b32 %r328, %r325, %r326, %r327;
prmt.b32 %r262, %r323, %r328, %r297;
prmt.b32 %r266, %r323, %r328, %r298;

dp4a.s32.s32 %r91, %r262, %r76, %r264;


dp4a.s32.s32 %r95, %r266, %r80, %r91;

ld.global.nc.u16 %rs22, [%rd35+2];
cvt.u32.u16 %r329, %rs22;
ld.global.nc.u16 %rs23, [%rd35+4];
mov.b32 %r330, {%rs22, %rs23};
and.b32 %r331, %r330, -2004318072;
shr.u32 %r332, %r331, 1;
or.b32 %r333, %r332, 839922192;
and.b32 %r334, %r329, 30583;
prmt.b32 %r335, %r285, %r284, %r334;
prmt.b32 %r336, %r288, %r287, %r334;
and.b32 %r337, %r333, 30292;
prmt.b32 %r338, %r335, %r336, %r337;
and.b16 %rs24, %rs23, 30583;
cvt.u32.u16 %r339, %rs24;
prmt.b32 %r340, %r285, %r284, %r339;
prmt.b32 %r341, %r288, %r287, %r339;
shr.u32 %r342, %r333, 16;
prmt.b32 %r343, %r340, %r341, %r342;
prmt.b32 %r270, %r338, %r343, %r297;
prmt.b32 %r274, %r338, %r343, %r298;

dp4a.s32.s32 %r99, %r270, %r84, %r95;


dp4a.s32.s32 %r103, %r274, %r88, %r99;

ld.global.nc.u16 %rs2, [%rd107];

{ cvt.f32.f16 %f75, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r90;
cvt.f32.f16 %f76, low;}


mul.ftz.f32 %f99, %f75, %f76;
cvt.rn.f32.s32 %f100, %r103;
fma.rn.ftz.f32 %f275, %f99, %f100, %f275;
cvt.u64.u32 %rd36, %r579;
add.s64 %rd37, %rd1, %rd36;
mul.lo.s64 %rd38, %rd37, 36;
add.s64 %rd39, %rd7, %rd38;
add.s64 %rd40, %rd39, %rd33;
ld.global.nc.u32 %r110, [%rd40+4];

dp4a.s32.s32 %r108, %r245, %r110, %r264;

ld.global.nc.u32 %r114, [%rd40+20];

dp4a.s32.s32 %r112, %r249, %r114, %r108;

ld.global.nc.u32 %r118, [%rd40+8];

dp4a.s32.s32 %r116, %r253, %r118, %r112;

ld.global.nc.u32 %r122, [%rd40+24];

dp4a.s32.s32 %r120, %r257, %r122, %r116;


{ cvt.f32.f16 %f77, %rs1;}


ld.global.nc.u32 %r124, [%rd39];

{.reg .f16 low,high;
mov.b32 {low,high},%r124;
cvt.f32.f16 %f78, low;}


mul.ftz.f32 %f101, %f77, %f78;
cvt.rn.f32.s32 %f102, %r120;
fma.rn.ftz.f32 %f274, %f101, %f102, %f274;

dp4a.s32.s32 %r125, %r262, %r110, %r264;


dp4a.s32.s32 %r129, %r266, %r114, %r125;


dp4a.s32.s32 %r133, %r270, %r118, %r129;


dp4a.s32.s32 %r137, %r274, %r122, %r133;


{ cvt.f32.f16 %f79, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r124;
cvt.f32.f16 %f80, low;}


mul.ftz.f32 %f103, %f79, %f80;
cvt.rn.f32.s32 %f104, %r137;
fma.rn.ftz.f32 %f273, %f103, %f104, %f273;
cvt.u64.u32 %rd41, %r583;
add.s64 %rd42, %rd1, %rd41;
mul.lo.s64 %rd43, %rd42, 36;
add.s64 %rd44, %rd7, %rd43;
add.s64 %rd45, %rd44, %rd33;
ld.global.nc.u32 %r144, [%rd45+4];

dp4a.s32.s32 %r142, %r245, %r144, %r264;

ld.global.nc.u32 %r148, [%rd45+20];

dp4a.s32.s32 %r146, %r249, %r148, %r142;

ld.global.nc.u32 %r152, [%rd45+8];

dp4a.s32.s32 %r150, %r253, %r152, %r146;

ld.global.nc.u32 %r156, [%rd45+24];

dp4a.s32.s32 %r154, %r257, %r156, %r150;


{ cvt.f32.f16 %f81, %rs1;}


ld.global.nc.u32 %r158, [%rd44];

{.reg .f16 low,high;
mov.b32 {low,high},%r158;
cvt.f32.f16 %f82, low;}


mul.ftz.f32 %f105, %f81, %f82;
cvt.rn.f32.s32 %f106, %r154;
fma.rn.ftz.f32 %f272, %f105, %f106, %f272;

dp4a.s32.s32 %r159, %r262, %r144, %r264;


dp4a.s32.s32 %r163, %r266, %r148, %r159;


dp4a.s32.s32 %r167, %r270, %r152, %r163;


dp4a.s32.s32 %r171, %r274, %r156, %r167;


{ cvt.f32.f16 %f83, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r158;
cvt.f32.f16 %f84, low;}


mul.ftz.f32 %f107, %f83, %f84;
cvt.rn.f32.s32 %f108, %r171;
fma.rn.ftz.f32 %f271, %f107, %f108, %f271;
cvt.u64.u32 %rd46, %r582;
add.s64 %rd47, %rd1, %rd46;
mul.lo.s64 %rd48, %rd47, 36;
add.s64 %rd49, %rd7, %rd48;
add.s64 %rd50, %rd49, %rd33;
ld.global.nc.u32 %r178, [%rd50+4];

dp4a.s32.s32 %r176, %r245, %r178, %r264;

ld.global.nc.u32 %r182, [%rd50+20];

dp4a.s32.s32 %r180, %r249, %r182, %r176;

ld.global.nc.u32 %r186, [%rd50+8];

dp4a.s32.s32 %r184, %r253, %r186, %r180;

ld.global.nc.u32 %r190, [%rd50+24];

dp4a.s32.s32 %r188, %r257, %r190, %r184;


{ cvt.f32.f16 %f85, %rs1;}


ld.global.nc.u32 %r192, [%rd49];

{.reg .f16 low,high;
mov.b32 {low,high},%r192;
cvt.f32.f16 %f86, low;}


mul.ftz.f32 %f109, %f85, %f86;
cvt.rn.f32.s32 %f110, %r188;
fma.rn.ftz.f32 %f270, %f109, %f110, %f270;

dp4a.s32.s32 %r193, %r262, %r178, %r264;


dp4a.s32.s32 %r197, %r266, %r182, %r193;


dp4a.s32.s32 %r201, %r270, %r186, %r197;


dp4a.s32.s32 %r205, %r274, %r190, %r201;


{ cvt.f32.f16 %f87, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r192;
cvt.f32.f16 %f88, low;}


mul.ftz.f32 %f111, %f87, %f88;
cvt.rn.f32.s32 %f112, %r205;
fma.rn.ftz.f32 %f269, %f111, %f112, %f269;
cvt.u64.u32 %rd51, %r581;
add.s64 %rd52, %rd1, %rd51;
mul.lo.s64 %rd53, %rd52, 36;
add.s64 %rd54, %rd7, %rd53;
add.s64 %rd55, %rd54, %rd33;
ld.global.nc.u32 %r212, [%rd55+4];

dp4a.s32.s32 %r210, %r245, %r212, %r264;

ld.global.nc.u32 %r216, [%rd55+20];

dp4a.s32.s32 %r214, %r249, %r216, %r210;

ld.global.nc.u32 %r220, [%rd55+8];

dp4a.s32.s32 %r218, %r253, %r220, %r214;

ld.global.nc.u32 %r224, [%rd55+24];

dp4a.s32.s32 %r222, %r257, %r224, %r218;


{ cvt.f32.f16 %f89, %rs1;}


ld.global.nc.u32 %r226, [%rd54];

{.reg .f16 low,high;
mov.b32 {low,high},%r226;
cvt.f32.f16 %f90, low;}


mul.ftz.f32 %f113, %f89, %f90;
cvt.rn.f32.s32 %f114, %r222;
fma.rn.ftz.f32 %f268, %f113, %f114, %f268;

dp4a.s32.s32 %r227, %r262, %r212, %r264;


dp4a.s32.s32 %r231, %r266, %r216, %r227;


dp4a.s32.s32 %r235, %r270, %r220, %r231;


dp4a.s32.s32 %r239, %r274, %r224, %r235;


{ cvt.f32.f16 %f91, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r226;
cvt.f32.f16 %f92, low;}


mul.ftz.f32 %f115, %f91, %f92;
cvt.rn.f32.s32 %f116, %r239;
fma.rn.ftz.f32 %f267, %f115, %f116, %f267;
cvt.u64.u32 %rd56, %r580;
add.s64 %rd57, %rd1, %rd56;
mul.lo.s64 %rd58, %rd57, 36;
add.s64 %rd59, %rd7, %rd58;
add.s64 %rd60, %rd59, %rd33;
ld.global.nc.u32 %r246, [%rd60+4];

dp4a.s32.s32 %r244, %r245, %r246, %r264;

ld.global.nc.u32 %r250, [%rd60+20];

dp4a.s32.s32 %r248, %r249, %r250, %r244;

ld.global.nc.u32 %r254, [%rd60+8];

dp4a.s32.s32 %r252, %r253, %r254, %r248;

ld.global.nc.u32 %r258, [%rd60+24];

dp4a.s32.s32 %r256, %r257, %r258, %r252;


{ cvt.f32.f16 %f93, %rs1;}


ld.global.nc.u32 %r260, [%rd59];

{.reg .f16 low,high;
mov.b32 {low,high},%r260;
cvt.f32.f16 %f94, low;}


mul.ftz.f32 %f117, %f93, %f94;
cvt.rn.f32.s32 %f118, %r256;
fma.rn.ftz.f32 %f266, %f117, %f118, %f266;

dp4a.s32.s32 %r261, %r262, %r246, %r264;


dp4a.s32.s32 %r265, %r266, %r250, %r261;


dp4a.s32.s32 %r269, %r270, %r254, %r265;


dp4a.s32.s32 %r273, %r274, %r258, %r269;


{ cvt.f32.f16 %f95, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r260;
cvt.f32.f16 %f96, low;}


mul.ftz.f32 %f119, %f95, %f96;
cvt.rn.f32.s32 %f120, %r273;
fma.rn.ftz.f32 %f265, %f119, %f120, %f265;
add.s32 %r583, %r583, 32;
add.s32 %r582, %r582, 32;
add.s32 %r581, %r581, 32;
add.s32 %r580, %r580, 32;
add.s32 %r579, %r579, 32;
add.s64 %rd107, %rd107, 576;
add.s64 %rd106, %rd106, 576;
add.s64 %rd105, %rd105, 576;
add.s32 %r584, %r584, 32;
setp.lt.s32 %p4, %r584, %r6;
@%p4 bra $L__BB288_2;

$L__BB288_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB288_5;

mov.u32 %r344, _ZZ13mul_mat_vec_qIL9ggml_type20ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r345, %r1, 1536, %r344;
shl.b32 %r346, %r2, 2;
add.s32 %r347, %r345, %r346;
st.shared.f32 [%r347+-1536], %f276;
st.shared.f32 [%r347+-1408], %f275;
st.shared.f32 [%r347+-1280], %f274;
st.shared.f32 [%r347+-1152], %f273;
st.shared.f32 [%r347+-1024], %f272;
st.shared.f32 [%r347+-896], %f271;
st.shared.f32 [%r347+-768], %f270;
st.shared.f32 [%r347+-640], %f269;
st.shared.f32 [%r347+-512], %f268;
st.shared.f32 [%r347+-384], %f267;
st.shared.f32 [%r347+-256], %f266;
st.shared.f32 [%r347+-128], %f265;

$L__BB288_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB288_30;

mov.u32 %r348, %ctaid.z;
mov.u32 %r349, %ctaid.y;
mad.lo.s32 %r350, %r349, %r35, %r4;
mad.lo.s32 %r351, %r348, %r41, %r350;
cvt.u64.u32 %rd14, %r351;
shl.b32 %r352, %r2, 2;
mov.u32 %r353, _ZZ13mul_mat_vec_qIL9ggml_type20ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r24, %r353, %r352;
mov.u32 %r354, 2;
ld.shared.f32 %f121, [%r24];
add.ftz.f32 %f122, %f121, %f276;
mov.b32 %r355, %f122;
mov.u32 %r356, 31;
mov.u32 %r357, 16;
mov.u32 %r358, -1;
shfl.sync.bfly.b32 %r359|%p7, %r355, %r357, %r356, %r358;
mov.b32 %f123, %r359;
add.ftz.f32 %f124, %f122, %f123;
mov.b32 %r360, %f124;
mov.u32 %r361, 8;
shfl.sync.bfly.b32 %r362|%p8, %r360, %r361, %r356, %r358;
mov.b32 %f125, %r362;
add.ftz.f32 %f126, %f124, %f125;
mov.b32 %r363, %f126;
mov.u32 %r364, 4;
shfl.sync.bfly.b32 %r365|%p9, %r363, %r364, %r356, %r358;
mov.b32 %f127, %r365;
add.ftz.f32 %f128, %f126, %f127;
mov.b32 %r366, %f128;
shfl.sync.bfly.b32 %r367|%p10, %r366, %r354, %r356, %r358;
mov.b32 %f129, %r367;
add.ftz.f32 %f130, %f128, %f129;
mov.b32 %r368, %f130;
mov.u32 %r369, 1;
shfl.sync.bfly.b32 %r370|%p11, %r368, %r369, %r356, %r358;
mov.b32 %f131, %r370;
add.ftz.f32 %f37, %f130, %f131;
setp.lt.u32 %p12, %r4, %r29;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
not.pred %p14, %p1;
@%p14 bra $L__BB288_8;

cvta.to.global.u64 %rd61, %rd19;
mul.wide.u32 %rd62, %r351, 4;
add.s64 %rd63, %rd61, %rd62;
st.global.f32 [%rd63], %f37;

$L__BB288_8:
ld.shared.f32 %f132, [%r24+128];
add.ftz.f32 %f133, %f132, %f275;
mov.b32 %r377, %f133;
shfl.sync.bfly.b32 %r381|%p15, %r377, %r357, %r356, %r358;
mov.b32 %f134, %r381;
add.ftz.f32 %f135, %f133, %f134;
mov.b32 %r382, %f135;
shfl.sync.bfly.b32 %r384|%p16, %r382, %r361, %r356, %r358;
mov.b32 %f136, %r384;
add.ftz.f32 %f137, %f135, %f136;
mov.b32 %r385, %f137;
shfl.sync.bfly.b32 %r387|%p17, %r385, %r364, %r356, %r358;
mov.b32 %f138, %r387;
add.ftz.f32 %f139, %f137, %f138;
mov.b32 %r388, %f139;
shfl.sync.bfly.b32 %r390|%p18, %r388, %r354, %r356, %r358;
mov.b32 %f140, %r390;
add.ftz.f32 %f141, %f139, %f140;
mov.b32 %r391, %f141;
shfl.sync.bfly.b32 %r393|%p19, %r391, %r369, %r356, %r358;
mov.b32 %f142, %r393;
add.ftz.f32 %f38, %f141, %f142;
setp.eq.s32 %p20, %r2, 1;
or.b32 %r397, %r4, 1;
setp.lt.u32 %p21, %r397, %r29;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB288_10;

cvta.to.global.u64 %rd64, %rd19;
mul.wide.u32 %rd65, %r351, 4;
add.s64 %rd66, %rd64, %rd65;
st.global.f32 [%rd66+4], %f38;

$L__BB288_10:
ld.shared.f32 %f143, [%r24+256];
add.ftz.f32 %f144, %f143, %f274;
mov.b32 %r404, %f144;
mov.u32 %r405, 31;
mov.u32 %r406, 16;
mov.u32 %r407, -1;
shfl.sync.bfly.b32 %r408|%p23, %r404, %r406, %r405, %r407;
mov.b32 %f145, %r408;
add.ftz.f32 %f146, %f144, %f145;
mov.b32 %r409, %f146;
mov.u32 %r410, 8;
shfl.sync.bfly.b32 %r411|%p24, %r409, %r410, %r405, %r407;
mov.b32 %f147, %r411;
add.ftz.f32 %f148, %f146, %f147;
mov.b32 %r412, %f148;
mov.u32 %r413, 4;
shfl.sync.bfly.b32 %r414|%p25, %r412, %r413, %r405, %r407;
mov.b32 %f149, %r414;
add.ftz.f32 %f150, %f148, %f149;
mov.b32 %r415, %f150;
mov.u32 %r416, 2;
shfl.sync.bfly.b32 %r417|%p26, %r415, %r416, %r405, %r407;
mov.b32 %f151, %r417;
add.ftz.f32 %f152, %f150, %f151;
mov.b32 %r418, %f152;
mov.u32 %r419, 1;
shfl.sync.bfly.b32 %r420|%p27, %r418, %r419, %r405, %r407;
mov.b32 %f153, %r420;
add.ftz.f32 %f39, %f152, %f153;
@%p14 bra $L__BB288_12;

cvt.u64.u32 %rd67, %r29;
add.s64 %rd68, %rd67, %rd14;
cvta.to.global.u64 %rd69, %rd19;
shl.b64 %rd70, %rd68, 2;
add.s64 %rd71, %rd69, %rd70;
st.global.f32 [%rd71], %f39;

$L__BB288_12:
ld.shared.f32 %f154, [%r24+384];
add.ftz.f32 %f155, %f154, %f273;
mov.b32 %r421, %f155;
shfl.sync.bfly.b32 %r425|%p29, %r421, %r406, %r405, %r407;
mov.b32 %f156, %r425;
add.ftz.f32 %f157, %f155, %f156;
mov.b32 %r426, %f157;
shfl.sync.bfly.b32 %r428|%p30, %r426, %r410, %r405, %r407;
mov.b32 %f158, %r428;
add.ftz.f32 %f159, %f157, %f158;
mov.b32 %r429, %f159;
shfl.sync.bfly.b32 %r431|%p31, %r429, %r413, %r405, %r407;
mov.b32 %f160, %r431;
add.ftz.f32 %f161, %f159, %f160;
mov.b32 %r432, %f161;
shfl.sync.bfly.b32 %r434|%p32, %r432, %r416, %r405, %r407;
mov.b32 %f162, %r434;
add.ftz.f32 %f163, %f161, %f162;
mov.b32 %r435, %f163;
shfl.sync.bfly.b32 %r437|%p33, %r435, %r419, %r405, %r407;
mov.b32 %f164, %r437;
add.ftz.f32 %f40, %f163, %f164;
@%p22 bra $L__BB288_14;

add.s32 %r438, %r29, 1;
cvt.u64.u32 %rd72, %r438;
add.s64 %rd73, %rd72, %rd14;
cvta.to.global.u64 %rd74, %rd19;
shl.b64 %rd75, %rd73, 2;
add.s64 %rd76, %rd74, %rd75;
st.global.f32 [%rd76], %f40;

$L__BB288_14:
shl.b32 %r439, %r29, 1;
mov.u32 %r440, 1;
ld.shared.f32 %f165, [%r24+512];
add.ftz.f32 %f166, %f165, %f272;
mov.b32 %r441, %f166;
mov.u32 %r442, 31;
mov.u32 %r443, 16;
mov.u32 %r444, -1;
shfl.sync.bfly.b32 %r445|%p35, %r441, %r443, %r442, %r444;
mov.b32 %f167, %r445;
add.ftz.f32 %f168, %f166, %f167;
mov.b32 %r446, %f168;
mov.u32 %r447, 8;
shfl.sync.bfly.b32 %r448|%p36, %r446, %r447, %r442, %r444;
mov.b32 %f169, %r448;
add.ftz.f32 %f170, %f168, %f169;
mov.b32 %r449, %f170;
mov.u32 %r450, 4;
shfl.sync.bfly.b32 %r451|%p37, %r449, %r450, %r442, %r444;
mov.b32 %f171, %r451;
add.ftz.f32 %f172, %f170, %f171;
mov.b32 %r452, %f172;
mov.u32 %r453, 2;
shfl.sync.bfly.b32 %r454|%p38, %r452, %r453, %r442, %r444;
mov.b32 %f173, %r454;
add.ftz.f32 %f174, %f172, %f173;
mov.b32 %r455, %f174;
shfl.sync.bfly.b32 %r456|%p39, %r455, %r440, %r442, %r444;
mov.b32 %f175, %r456;
add.ftz.f32 %f41, %f174, %f175;
cvt.u64.u32 %rd77, %r439;
add.s64 %rd78, %rd77, %rd14;
cvta.to.global.u64 %rd79, %rd19;
shl.b64 %rd80, %rd78, 2;
add.s64 %rd15, %rd79, %rd80;
@%p14 bra $L__BB288_16;

st.global.f32 [%rd15], %f41;

$L__BB288_16:
ld.shared.f32 %f176, [%r24+640];
add.ftz.f32 %f177, %f176, %f271;
mov.b32 %r457, %f177;
shfl.sync.bfly.b32 %r461|%p41, %r457, %r443, %r442, %r444;
mov.b32 %f178, %r461;
add.ftz.f32 %f179, %f177, %f178;
mov.b32 %r462, %f179;
shfl.sync.bfly.b32 %r464|%p42, %r462, %r447, %r442, %r444;
mov.b32 %f180, %r464;
add.ftz.f32 %f181, %f179, %f180;
mov.b32 %r465, %f181;
shfl.sync.bfly.b32 %r467|%p43, %r465, %r450, %r442, %r444;
mov.b32 %f182, %r467;
add.ftz.f32 %f183, %f181, %f182;
mov.b32 %r468, %f183;
shfl.sync.bfly.b32 %r470|%p44, %r468, %r453, %r442, %r444;
mov.b32 %f184, %r470;
add.ftz.f32 %f185, %f183, %f184;
mov.b32 %r471, %f185;
shfl.sync.bfly.b32 %r473|%p45, %r471, %r440, %r442, %r444;
mov.b32 %f186, %r473;
add.ftz.f32 %f42, %f185, %f186;
@%p22 bra $L__BB288_18;

st.global.f32 [%rd15+4], %f42;

$L__BB288_18:
mul.lo.s32 %r25, %r29, 3;
ld.shared.f32 %f187, [%r24+768];
add.ftz.f32 %f188, %f187, %f270;
mov.b32 %r474, %f188;
mov.u32 %r475, 31;
mov.u32 %r476, 16;
mov.u32 %r477, -1;
shfl.sync.bfly.b32 %r478|%p47, %r474, %r476, %r475, %r477;
mov.b32 %f189, %r478;
add.ftz.f32 %f190, %f188, %f189;
mov.b32 %r479, %f190;
mov.u32 %r480, 8;
shfl.sync.bfly.b32 %r481|%p48, %r479, %r480, %r475, %r477;
mov.b32 %f191, %r481;
add.ftz.f32 %f192, %f190, %f191;
mov.b32 %r482, %f192;
mov.u32 %r483, 4;
shfl.sync.bfly.b32 %r484|%p49, %r482, %r483, %r475, %r477;
mov.b32 %f193, %r484;
add.ftz.f32 %f194, %f192, %f193;
mov.b32 %r485, %f194;
mov.u32 %r486, 2;
shfl.sync.bfly.b32 %r487|%p50, %r485, %r486, %r475, %r477;
mov.b32 %f195, %r487;
add.ftz.f32 %f196, %f194, %f195;
mov.b32 %r488, %f196;
mov.u32 %r489, 1;
shfl.sync.bfly.b32 %r490|%p51, %r488, %r489, %r475, %r477;
mov.b32 %f197, %r490;
add.ftz.f32 %f43, %f196, %f197;
@%p14 bra $L__BB288_20;

cvt.u64.u32 %rd81, %r25;
add.s64 %rd82, %rd81, %rd14;
shl.b64 %rd84, %rd82, 2;
add.s64 %rd85, %rd79, %rd84;
st.global.f32 [%rd85], %f43;

$L__BB288_20:
ld.shared.f32 %f198, [%r24+896];
add.ftz.f32 %f199, %f198, %f269;
mov.b32 %r491, %f199;
shfl.sync.bfly.b32 %r495|%p53, %r491, %r476, %r475, %r477;
mov.b32 %f200, %r495;
add.ftz.f32 %f201, %f199, %f200;
mov.b32 %r496, %f201;
shfl.sync.bfly.b32 %r498|%p54, %r496, %r480, %r475, %r477;
mov.b32 %f202, %r498;
add.ftz.f32 %f203, %f201, %f202;
mov.b32 %r499, %f203;
shfl.sync.bfly.b32 %r501|%p55, %r499, %r483, %r475, %r477;
mov.b32 %f204, %r501;
add.ftz.f32 %f205, %f203, %f204;
mov.b32 %r502, %f205;
shfl.sync.bfly.b32 %r504|%p56, %r502, %r486, %r475, %r477;
mov.b32 %f206, %r504;
add.ftz.f32 %f207, %f205, %f206;
mov.b32 %r505, %f207;
shfl.sync.bfly.b32 %r507|%p57, %r505, %r489, %r475, %r477;
mov.b32 %f208, %r507;
add.ftz.f32 %f44, %f207, %f208;
@%p22 bra $L__BB288_22;

add.s32 %r508, %r25, 1;
cvt.u64.u32 %rd86, %r508;
add.s64 %rd87, %rd86, %rd14;
shl.b64 %rd89, %rd87, 2;
add.s64 %rd90, %rd79, %rd89;
st.global.f32 [%rd90], %f44;

$L__BB288_22:
shl.b32 %r509, %r29, 2;
mov.u32 %r510, 2;
ld.shared.f32 %f209, [%r24+1024];
add.ftz.f32 %f210, %f209, %f268;
mov.b32 %r511, %f210;
mov.u32 %r512, 31;
mov.u32 %r513, 16;
mov.u32 %r514, -1;
shfl.sync.bfly.b32 %r515|%p59, %r511, %r513, %r512, %r514;
mov.b32 %f211, %r515;
add.ftz.f32 %f212, %f210, %f211;
mov.b32 %r516, %f212;
mov.u32 %r517, 8;
shfl.sync.bfly.b32 %r518|%p60, %r516, %r517, %r512, %r514;
mov.b32 %f213, %r518;
add.ftz.f32 %f214, %f212, %f213;
mov.b32 %r519, %f214;
mov.u32 %r520, 4;
shfl.sync.bfly.b32 %r521|%p61, %r519, %r520, %r512, %r514;
mov.b32 %f215, %r521;
add.ftz.f32 %f216, %f214, %f215;
mov.b32 %r522, %f216;
shfl.sync.bfly.b32 %r523|%p62, %r522, %r510, %r512, %r514;
mov.b32 %f217, %r523;
add.ftz.f32 %f218, %f216, %f217;
mov.b32 %r524, %f218;
mov.u32 %r525, 1;
shfl.sync.bfly.b32 %r526|%p63, %r524, %r525, %r512, %r514;
mov.b32 %f219, %r526;
add.ftz.f32 %f45, %f218, %f219;
cvt.u64.u32 %rd91, %r509;
add.s64 %rd92, %rd91, %rd14;
shl.b64 %rd94, %rd92, 2;
add.s64 %rd16, %rd79, %rd94;
@%p14 bra $L__BB288_24;

st.global.f32 [%rd16], %f45;

$L__BB288_24:
ld.shared.f32 %f220, [%r24+1152];
add.ftz.f32 %f221, %f220, %f267;
mov.b32 %r527, %f221;
shfl.sync.bfly.b32 %r531|%p65, %r527, %r513, %r512, %r514;
mov.b32 %f222, %r531;
add.ftz.f32 %f223, %f221, %f222;
mov.b32 %r532, %f223;
shfl.sync.bfly.b32 %r534|%p66, %r532, %r517, %r512, %r514;
mov.b32 %f224, %r534;
add.ftz.f32 %f225, %f223, %f224;
mov.b32 %r535, %f225;
shfl.sync.bfly.b32 %r537|%p67, %r535, %r520, %r512, %r514;
mov.b32 %f226, %r537;
add.ftz.f32 %f227, %f225, %f226;
mov.b32 %r538, %f227;
shfl.sync.bfly.b32 %r540|%p68, %r538, %r510, %r512, %r514;
mov.b32 %f228, %r540;
add.ftz.f32 %f229, %f227, %f228;
mov.b32 %r541, %f229;
shfl.sync.bfly.b32 %r543|%p69, %r541, %r525, %r512, %r514;
mov.b32 %f230, %r543;
add.ftz.f32 %f46, %f229, %f230;
@%p22 bra $L__BB288_26;

st.global.f32 [%rd16+4], %f46;

$L__BB288_26:
mul.lo.s32 %r26, %r29, 5;
ld.shared.f32 %f231, [%r24+1280];
add.ftz.f32 %f232, %f231, %f266;
mov.b32 %r544, %f232;
mov.u32 %r545, 31;
mov.u32 %r546, 16;
mov.u32 %r547, -1;
shfl.sync.bfly.b32 %r548|%p71, %r544, %r546, %r545, %r547;
mov.b32 %f233, %r548;
add.ftz.f32 %f234, %f232, %f233;
mov.b32 %r549, %f234;
mov.u32 %r550, 8;
shfl.sync.bfly.b32 %r551|%p72, %r549, %r550, %r545, %r547;
mov.b32 %f235, %r551;
add.ftz.f32 %f236, %f234, %f235;
mov.b32 %r552, %f236;
mov.u32 %r553, 4;
shfl.sync.bfly.b32 %r554|%p73, %r552, %r553, %r545, %r547;
mov.b32 %f237, %r554;
add.ftz.f32 %f238, %f236, %f237;
mov.b32 %r555, %f238;
mov.u32 %r556, 2;
shfl.sync.bfly.b32 %r557|%p74, %r555, %r556, %r545, %r547;
mov.b32 %f239, %r557;
add.ftz.f32 %f240, %f238, %f239;
mov.b32 %r558, %f240;
mov.u32 %r559, 1;
shfl.sync.bfly.b32 %r560|%p75, %r558, %r559, %r545, %r547;
mov.b32 %f241, %r560;
add.ftz.f32 %f47, %f240, %f241;
@%p14 bra $L__BB288_28;

cvt.u64.u32 %rd95, %r26;
add.s64 %rd96, %rd95, %rd14;
shl.b64 %rd98, %rd96, 2;
add.s64 %rd99, %rd79, %rd98;
st.global.f32 [%rd99], %f47;

$L__BB288_28:
ld.shared.f32 %f242, [%r24+1408];
add.ftz.f32 %f243, %f242, %f265;
mov.b32 %r561, %f243;
shfl.sync.bfly.b32 %r565|%p77, %r561, %r546, %r545, %r547;
mov.b32 %f244, %r565;
add.ftz.f32 %f245, %f243, %f244;
mov.b32 %r566, %f245;
shfl.sync.bfly.b32 %r568|%p78, %r566, %r550, %r545, %r547;
mov.b32 %f246, %r568;
add.ftz.f32 %f247, %f245, %f246;
mov.b32 %r569, %f247;
shfl.sync.bfly.b32 %r571|%p79, %r569, %r553, %r545, %r547;
mov.b32 %f248, %r571;
add.ftz.f32 %f249, %f247, %f248;
mov.b32 %r572, %f249;
shfl.sync.bfly.b32 %r574|%p80, %r572, %r556, %r545, %r547;
mov.b32 %f250, %r574;
add.ftz.f32 %f251, %f249, %f250;
mov.b32 %r575, %f251;
shfl.sync.bfly.b32 %r577|%p81, %r575, %r559, %r545, %r547;
mov.b32 %f252, %r577;
add.ftz.f32 %f48, %f251, %f252;
@%p22 bra $L__BB288_30;

add.s32 %r578, %r26, 1;
cvt.u64.u32 %rd100, %r578;
add.s64 %rd101, %rd100, %rd14;
shl.b64 %rd103, %rd101, 2;
add.s64 %rd104, %rd79, %rd103;
st.global.f32 [%rd104], %f48;

$L__BB288_30:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type20ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type20ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type20ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type20ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type20ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type20ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type20ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type20ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type20ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<95>;
.reg .b16 %rs<27>;
.reg .f32 %f<323>;
.reg .b32 %r<646>;
.reg .b64 %rd<115>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type20ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1792];

ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type20ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd20, [_Z13mul_mat_vec_qIL9ggml_type20ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd21, [_Z13mul_mat_vec_qIL9ggml_type20ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r45, [_Z13mul_mat_vec_qIL9ggml_type20ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type20ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type20ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type20ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type20ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type20ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type20ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type20ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type20ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type20ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type20ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type20ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type20ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type20ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r46, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r46, %r2;
mov.u32 %r47, %ctaid.x;
shl.b32 %r4, %r47, 1;
shr.u32 %r48, %r3, 31;
add.s32 %r49, %r3, %r48;
shr.s32 %r645, %r49, 1;
shr.u32 %r6, %r45, 5;
setp.ge.s32 %p3, %r645, %r6;
mov.f32 %f309, 0f00000000;
mov.f32 %f310, %f309;
mov.f32 %f311, %f309;
mov.f32 %f312, %f309;
mov.f32 %f313, %f309;
mov.f32 %f314, %f309;
mov.f32 %f315, %f309;
mov.f32 %f316, %f309;
mov.f32 %f317, %f309;
mov.f32 %f318, %f309;
mov.f32 %f319, %f309;
mov.f32 %f320, %f309;
mov.f32 %f321, %f309;
mov.f32 %f322, %f309;
@%p3 bra $L__BB289_3;

mov.u32 %r50, %ctaid.y;
mov.u32 %r51, %ctaid.z;
mul.lo.s32 %r52, %r50, %r37;
cvt.u64.u32 %rd22, %r52;
mul.lo.s32 %r53, %r51, %r43;
cvt.u64.u32 %rd23, %r53;
add.s64 %rd1, %rd23, %rd22;
mul.hi.u32 %r54, %r51, %r39;
add.s32 %r55, %r54, %r51;
shr.u32 %r56, %r55, %r40;
mul.lo.s32 %r57, %r56, %r42;
mul.hi.u32 %r58, %r50, %r33;
add.s32 %r59, %r58, %r50;
shr.u32 %r60, %r59, %r34;
mul.lo.s32 %r61, %r60, %r36;
cvta.to.global.u64 %rd24, %rd19;
and.b32 %r64, %r49, -2;
sub.s32 %r65, %r3, %r64;
shl.b32 %r66, %r65, 1;
cvt.s64.s32 %rd2, %r66;
shl.b32 %r67, %r65, 2;
shl.b32 %r68, %r31, 1;
add.s32 %r644, %r645, %r68;
mad.lo.s32 %r643, %r31, 3, %r645;
shl.b32 %r69, %r31, 2;
add.s32 %r642, %r645, %r69;
mad.lo.s32 %r641, %r31, 5, %r645;
mad.lo.s32 %r640, %r31, 6, %r645;
add.s32 %r639, %r645, %r31;
add.s32 %r70, %r645, %r57;
add.s32 %r71, %r70, %r61;
add.s32 %r72, %r4, 1;
mad.lo.s32 %r73, %r30, %r72, %r71;
mul.wide.s32 %rd25, %r73, 18;
add.s64 %rd114, %rd24, %rd25;
mad.lo.s32 %r74, %r4, %r30, %r645;
add.s32 %r75, %r74, %r57;
add.s32 %r76, %r75, %r61;
mul.wide.s32 %rd26, %r76, 18;
add.s64 %rd27, %rd24, %rd26;
add.s64 %rd113, %rd27, 4;
mul.wide.s32 %rd28, %r67, 2;
add.s64 %rd29, %rd26, %rd28;
add.s64 %rd30, %rd24, %rd29;
add.s64 %rd112, %rd30, 8;
or.b64 %rd6, %rd28, 4;
cvta.to.global.u64 %rd7, %rd20;
mov.f32 %f309, 0f00000000;
shl.b64 %rd35, %rd2, 2;

$L__BB289_2:
cvt.u64.u32 %rd31, %r645;
add.s64 %rd32, %rd1, %rd31;
mul.lo.s64 %rd33, %rd32, 36;
add.s64 %rd34, %rd7, %rd33;
add.s64 %rd36, %rd34, %rd35;
ld.global.nc.u16 %rs15, [%rd112+-6];
cvt.u32.u16 %r315, %rs15;
ld.global.nc.u16 %rs16, [%rd112+-4];
mov.b32 %r316, {%rs15, %rs16};
and.b32 %r317, %r316, -2004318072;
shr.u32 %r318, %r317, 1;
or.b32 %r319, %r318, 839922192;
and.b32 %r320, %r315, 30583;
mov.u32 %r321, -152379953;
mov.u32 %r322, -1079142271;
prmt.b32 %r323, %r322, %r321, %r320;
mov.u32 %r324, 1901675829;
mov.u32 %r325, 639175937;
prmt.b32 %r326, %r325, %r324, %r320;
and.b32 %r327, %r319, 30292;
prmt.b32 %r328, %r323, %r326, %r327;
and.b16 %rs17, %rs16, 30583;
cvt.u32.u16 %r329, %rs17;
prmt.b32 %r330, %r322, %r321, %r329;
prmt.b32 %r331, %r325, %r324, %r329;
shr.u32 %r332, %r319, 16;
prmt.b32 %r333, %r330, %r331, %r332;
mov.u32 %r334, 25632;
prmt.b32 %r282, %r328, %r333, %r334;
mov.u32 %r335, 30001;
prmt.b32 %r286, %r328, %r333, %r335;
ld.global.nc.u32 %r79, [%rd36+4];
mov.u32 %r301, 0;

dp4a.s32.s32 %r77, %r282, %r79, %r301;

ld.global.nc.u32 %r83, [%rd36+20];

dp4a.s32.s32 %r81, %r286, %r83, %r77;

ld.global.nc.u16 %rs18, [%rd112+-2];
cvt.u32.u16 %r336, %rs18;
ld.global.nc.u16 %rs19, [%rd112];
mov.b32 %r337, {%rs18, %rs19};
and.b32 %r338, %r337, -2004318072;
shr.u32 %r339, %r338, 1;
or.b32 %r340, %r339, 839922192;
and.b32 %r341, %r336, 30583;
prmt.b32 %r342, %r322, %r321, %r341;
prmt.b32 %r343, %r325, %r324, %r341;
and.b32 %r344, %r340, 30292;
prmt.b32 %r345, %r342, %r343, %r344;
and.b16 %rs20, %rs19, 30583;
cvt.u32.u16 %r346, %rs20;
prmt.b32 %r347, %r322, %r321, %r346;
prmt.b32 %r348, %r325, %r324, %r346;
shr.u32 %r349, %r340, 16;
prmt.b32 %r350, %r347, %r348, %r349;
prmt.b32 %r290, %r345, %r350, %r334;
prmt.b32 %r294, %r345, %r350, %r335;
ld.global.nc.u32 %r87, [%rd36+8];

dp4a.s32.s32 %r85, %r290, %r87, %r81;

ld.global.nc.u32 %r91, [%rd36+24];

dp4a.s32.s32 %r89, %r294, %r91, %r85;

ld.global.nc.u16 %rs1, [%rd113+-4];

{ cvt.f32.f16 %f85, %rs1;}


ld.global.nc.u32 %r93, [%rd34];

{.reg .f16 low,high;
mov.b32 {low,high},%r93;
cvt.f32.f16 %f86, low;}


mul.ftz.f32 %f113, %f85, %f86;
cvt.rn.f32.s32 %f114, %r89;
fma.rn.ftz.f32 %f322, %f113, %f114, %f322;
add.s64 %rd37, %rd114, %rd6;
ld.global.nc.u16 %rs21, [%rd37+-2];
cvt.u32.u16 %r351, %rs21;
ld.global.nc.u16 %rs22, [%rd37];
mov.b32 %r352, {%rs21, %rs22};
and.b32 %r353, %r352, -2004318072;
shr.u32 %r354, %r353, 1;
or.b32 %r355, %r354, 839922192;
and.b32 %r356, %r351, 30583;
prmt.b32 %r357, %r322, %r321, %r356;
prmt.b32 %r358, %r325, %r324, %r356;
and.b32 %r359, %r355, 30292;
prmt.b32 %r360, %r357, %r358, %r359;
and.b16 %rs23, %rs22, 30583;
cvt.u32.u16 %r361, %rs23;
prmt.b32 %r362, %r322, %r321, %r361;
prmt.b32 %r363, %r325, %r324, %r361;
shr.u32 %r364, %r355, 16;
prmt.b32 %r365, %r362, %r363, %r364;
prmt.b32 %r299, %r360, %r365, %r334;
prmt.b32 %r303, %r360, %r365, %r335;

dp4a.s32.s32 %r94, %r299, %r79, %r301;


dp4a.s32.s32 %r98, %r303, %r83, %r94;

ld.global.nc.u16 %rs24, [%rd37+2];
cvt.u32.u16 %r366, %rs24;
ld.global.nc.u16 %rs25, [%rd37+4];
mov.b32 %r367, {%rs24, %rs25};
and.b32 %r368, %r367, -2004318072;
shr.u32 %r369, %r368, 1;
or.b32 %r370, %r369, 839922192;
and.b32 %r371, %r366, 30583;
prmt.b32 %r372, %r322, %r321, %r371;
prmt.b32 %r373, %r325, %r324, %r371;
and.b32 %r374, %r370, 30292;
prmt.b32 %r375, %r372, %r373, %r374;
and.b16 %rs26, %rs25, 30583;
cvt.u32.u16 %r376, %rs26;
prmt.b32 %r377, %r322, %r321, %r376;
prmt.b32 %r378, %r325, %r324, %r376;
shr.u32 %r379, %r370, 16;
prmt.b32 %r380, %r377, %r378, %r379;
prmt.b32 %r307, %r375, %r380, %r334;
prmt.b32 %r311, %r375, %r380, %r335;

dp4a.s32.s32 %r102, %r307, %r87, %r98;


dp4a.s32.s32 %r106, %r311, %r91, %r102;

ld.global.nc.u16 %rs2, [%rd114];

{ cvt.f32.f16 %f87, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r93;
cvt.f32.f16 %f88, low;}


mul.ftz.f32 %f115, %f87, %f88;
cvt.rn.f32.s32 %f116, %r106;
fma.rn.ftz.f32 %f321, %f115, %f116, %f321;
cvt.u64.u32 %rd38, %r639;
add.s64 %rd39, %rd1, %rd38;
mul.lo.s64 %rd40, %rd39, 36;
add.s64 %rd41, %rd7, %rd40;
add.s64 %rd42, %rd41, %rd35;
ld.global.nc.u32 %r113, [%rd42+4];

dp4a.s32.s32 %r111, %r282, %r113, %r301;

ld.global.nc.u32 %r117, [%rd42+20];

dp4a.s32.s32 %r115, %r286, %r117, %r111;

ld.global.nc.u32 %r121, [%rd42+8];

dp4a.s32.s32 %r119, %r290, %r121, %r115;

ld.global.nc.u32 %r125, [%rd42+24];

dp4a.s32.s32 %r123, %r294, %r125, %r119;


{ cvt.f32.f16 %f89, %rs1;}


ld.global.nc.u32 %r127, [%rd41];

{.reg .f16 low,high;
mov.b32 {low,high},%r127;
cvt.f32.f16 %f90, low;}


mul.ftz.f32 %f117, %f89, %f90;
cvt.rn.f32.s32 %f118, %r123;
fma.rn.ftz.f32 %f320, %f117, %f118, %f320;

dp4a.s32.s32 %r128, %r299, %r113, %r301;


dp4a.s32.s32 %r132, %r303, %r117, %r128;


dp4a.s32.s32 %r136, %r307, %r121, %r132;


dp4a.s32.s32 %r140, %r311, %r125, %r136;


{ cvt.f32.f16 %f91, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r127;
cvt.f32.f16 %f92, low;}


mul.ftz.f32 %f119, %f91, %f92;
cvt.rn.f32.s32 %f120, %r140;
fma.rn.ftz.f32 %f319, %f119, %f120, %f319;
cvt.u64.u32 %rd43, %r644;
add.s64 %rd44, %rd1, %rd43;
mul.lo.s64 %rd45, %rd44, 36;
add.s64 %rd46, %rd7, %rd45;
add.s64 %rd47, %rd46, %rd35;
ld.global.nc.u32 %r147, [%rd47+4];

dp4a.s32.s32 %r145, %r282, %r147, %r301;

ld.global.nc.u32 %r151, [%rd47+20];

dp4a.s32.s32 %r149, %r286, %r151, %r145;

ld.global.nc.u32 %r155, [%rd47+8];

dp4a.s32.s32 %r153, %r290, %r155, %r149;

ld.global.nc.u32 %r159, [%rd47+24];

dp4a.s32.s32 %r157, %r294, %r159, %r153;


{ cvt.f32.f16 %f93, %rs1;}


ld.global.nc.u32 %r161, [%rd46];

{.reg .f16 low,high;
mov.b32 {low,high},%r161;
cvt.f32.f16 %f94, low;}


mul.ftz.f32 %f121, %f93, %f94;
cvt.rn.f32.s32 %f122, %r157;
fma.rn.ftz.f32 %f318, %f121, %f122, %f318;

dp4a.s32.s32 %r162, %r299, %r147, %r301;


dp4a.s32.s32 %r166, %r303, %r151, %r162;


dp4a.s32.s32 %r170, %r307, %r155, %r166;


dp4a.s32.s32 %r174, %r311, %r159, %r170;


{ cvt.f32.f16 %f95, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r161;
cvt.f32.f16 %f96, low;}


mul.ftz.f32 %f123, %f95, %f96;
cvt.rn.f32.s32 %f124, %r174;
fma.rn.ftz.f32 %f317, %f123, %f124, %f317;
cvt.u64.u32 %rd48, %r643;
add.s64 %rd49, %rd1, %rd48;
mul.lo.s64 %rd50, %rd49, 36;
add.s64 %rd51, %rd7, %rd50;
add.s64 %rd52, %rd51, %rd35;
ld.global.nc.u32 %r181, [%rd52+4];

dp4a.s32.s32 %r179, %r282, %r181, %r301;

ld.global.nc.u32 %r185, [%rd52+20];

dp4a.s32.s32 %r183, %r286, %r185, %r179;

ld.global.nc.u32 %r189, [%rd52+8];

dp4a.s32.s32 %r187, %r290, %r189, %r183;

ld.global.nc.u32 %r193, [%rd52+24];

dp4a.s32.s32 %r191, %r294, %r193, %r187;


{ cvt.f32.f16 %f97, %rs1;}


ld.global.nc.u32 %r195, [%rd51];

{.reg .f16 low,high;
mov.b32 {low,high},%r195;
cvt.f32.f16 %f98, low;}


mul.ftz.f32 %f125, %f97, %f98;
cvt.rn.f32.s32 %f126, %r191;
fma.rn.ftz.f32 %f316, %f125, %f126, %f316;

dp4a.s32.s32 %r196, %r299, %r181, %r301;


dp4a.s32.s32 %r200, %r303, %r185, %r196;


dp4a.s32.s32 %r204, %r307, %r189, %r200;


dp4a.s32.s32 %r208, %r311, %r193, %r204;


{ cvt.f32.f16 %f99, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r195;
cvt.f32.f16 %f100, low;}


mul.ftz.f32 %f127, %f99, %f100;
cvt.rn.f32.s32 %f128, %r208;
fma.rn.ftz.f32 %f315, %f127, %f128, %f315;
cvt.u64.u32 %rd53, %r642;
add.s64 %rd54, %rd1, %rd53;
mul.lo.s64 %rd55, %rd54, 36;
add.s64 %rd56, %rd7, %rd55;
add.s64 %rd57, %rd56, %rd35;
ld.global.nc.u32 %r215, [%rd57+4];

dp4a.s32.s32 %r213, %r282, %r215, %r301;

ld.global.nc.u32 %r219, [%rd57+20];

dp4a.s32.s32 %r217, %r286, %r219, %r213;

ld.global.nc.u32 %r223, [%rd57+8];

dp4a.s32.s32 %r221, %r290, %r223, %r217;

ld.global.nc.u32 %r227, [%rd57+24];

dp4a.s32.s32 %r225, %r294, %r227, %r221;


{ cvt.f32.f16 %f101, %rs1;}


ld.global.nc.u32 %r229, [%rd56];

{.reg .f16 low,high;
mov.b32 {low,high},%r229;
cvt.f32.f16 %f102, low;}


mul.ftz.f32 %f129, %f101, %f102;
cvt.rn.f32.s32 %f130, %r225;
fma.rn.ftz.f32 %f314, %f129, %f130, %f314;

dp4a.s32.s32 %r230, %r299, %r215, %r301;


dp4a.s32.s32 %r234, %r303, %r219, %r230;


dp4a.s32.s32 %r238, %r307, %r223, %r234;


dp4a.s32.s32 %r242, %r311, %r227, %r238;


{ cvt.f32.f16 %f103, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r229;
cvt.f32.f16 %f104, low;}


mul.ftz.f32 %f131, %f103, %f104;
cvt.rn.f32.s32 %f132, %r242;
fma.rn.ftz.f32 %f313, %f131, %f132, %f313;
cvt.u64.u32 %rd58, %r641;
add.s64 %rd59, %rd1, %rd58;
mul.lo.s64 %rd60, %rd59, 36;
add.s64 %rd61, %rd7, %rd60;
add.s64 %rd62, %rd61, %rd35;
ld.global.nc.u32 %r249, [%rd62+4];

dp4a.s32.s32 %r247, %r282, %r249, %r301;

ld.global.nc.u32 %r253, [%rd62+20];

dp4a.s32.s32 %r251, %r286, %r253, %r247;

ld.global.nc.u32 %r257, [%rd62+8];

dp4a.s32.s32 %r255, %r290, %r257, %r251;

ld.global.nc.u32 %r261, [%rd62+24];

dp4a.s32.s32 %r259, %r294, %r261, %r255;


{ cvt.f32.f16 %f105, %rs1;}


ld.global.nc.u32 %r263, [%rd61];

{.reg .f16 low,high;
mov.b32 {low,high},%r263;
cvt.f32.f16 %f106, low;}


mul.ftz.f32 %f133, %f105, %f106;
cvt.rn.f32.s32 %f134, %r259;
fma.rn.ftz.f32 %f312, %f133, %f134, %f312;

dp4a.s32.s32 %r264, %r299, %r249, %r301;


dp4a.s32.s32 %r268, %r303, %r253, %r264;


dp4a.s32.s32 %r272, %r307, %r257, %r268;


dp4a.s32.s32 %r276, %r311, %r261, %r272;


{ cvt.f32.f16 %f107, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r263;
cvt.f32.f16 %f108, low;}


mul.ftz.f32 %f135, %f107, %f108;
cvt.rn.f32.s32 %f136, %r276;
fma.rn.ftz.f32 %f311, %f135, %f136, %f311;
cvt.u64.u32 %rd63, %r640;
add.s64 %rd64, %rd1, %rd63;
mul.lo.s64 %rd65, %rd64, 36;
add.s64 %rd66, %rd7, %rd65;
add.s64 %rd67, %rd66, %rd35;
ld.global.nc.u32 %r283, [%rd67+4];

dp4a.s32.s32 %r281, %r282, %r283, %r301;

ld.global.nc.u32 %r287, [%rd67+20];

dp4a.s32.s32 %r285, %r286, %r287, %r281;

ld.global.nc.u32 %r291, [%rd67+8];

dp4a.s32.s32 %r289, %r290, %r291, %r285;

ld.global.nc.u32 %r295, [%rd67+24];

dp4a.s32.s32 %r293, %r294, %r295, %r289;


{ cvt.f32.f16 %f109, %rs1;}


ld.global.nc.u32 %r297, [%rd66];

{.reg .f16 low,high;
mov.b32 {low,high},%r297;
cvt.f32.f16 %f110, low;}


mul.ftz.f32 %f137, %f109, %f110;
cvt.rn.f32.s32 %f138, %r293;
fma.rn.ftz.f32 %f310, %f137, %f138, %f310;

dp4a.s32.s32 %r298, %r299, %r283, %r301;


dp4a.s32.s32 %r302, %r303, %r287, %r298;


dp4a.s32.s32 %r306, %r307, %r291, %r302;


dp4a.s32.s32 %r310, %r311, %r295, %r306;


{ cvt.f32.f16 %f111, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r297;
cvt.f32.f16 %f112, low;}


mul.ftz.f32 %f139, %f111, %f112;
cvt.rn.f32.s32 %f140, %r310;
fma.rn.ftz.f32 %f309, %f139, %f140, %f309;
add.s32 %r644, %r644, 32;
add.s32 %r643, %r643, 32;
add.s32 %r642, %r642, 32;
add.s32 %r641, %r641, 32;
add.s32 %r640, %r640, 32;
add.s32 %r639, %r639, 32;
add.s64 %rd114, %rd114, 576;
add.s64 %rd113, %rd113, 576;
add.s64 %rd112, %rd112, 576;
add.s32 %r645, %r645, 32;
setp.lt.s32 %p4, %r645, %r6;
@%p4 bra $L__BB289_2;

$L__BB289_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB289_5;

mov.u32 %r381, _ZZ13mul_mat_vec_qIL9ggml_type20ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r382, %r1, 1792, %r381;
shl.b32 %r383, %r2, 2;
add.s32 %r384, %r382, %r383;
st.shared.f32 [%r384+-1792], %f322;
st.shared.f32 [%r384+-1664], %f321;
st.shared.f32 [%r384+-1536], %f320;
st.shared.f32 [%r384+-1408], %f319;
st.shared.f32 [%r384+-1280], %f318;
st.shared.f32 [%r384+-1152], %f317;
st.shared.f32 [%r384+-1024], %f316;
st.shared.f32 [%r384+-896], %f315;
st.shared.f32 [%r384+-768], %f314;
st.shared.f32 [%r384+-640], %f313;
st.shared.f32 [%r384+-512], %f312;
st.shared.f32 [%r384+-384], %f311;
st.shared.f32 [%r384+-256], %f310;
st.shared.f32 [%r384+-128], %f309;

$L__BB289_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB289_34;

mov.u32 %r385, %ctaid.z;
mov.u32 %r386, %ctaid.y;
mad.lo.s32 %r387, %r386, %r38, %r4;
mad.lo.s32 %r388, %r385, %r44, %r387;
cvt.u64.u32 %rd14, %r388;
shl.b32 %r389, %r2, 2;
mov.u32 %r390, _ZZ13mul_mat_vec_qIL9ggml_type20ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r27, %r390, %r389;
mov.u32 %r391, 2;
ld.shared.f32 %f141, [%r27];
add.ftz.f32 %f142, %f141, %f322;
mov.b32 %r392, %f142;
mov.u32 %r393, 31;
mov.u32 %r394, 16;
mov.u32 %r395, -1;
shfl.sync.bfly.b32 %r396|%p7, %r392, %r394, %r393, %r395;
mov.b32 %f143, %r396;
add.ftz.f32 %f144, %f142, %f143;
mov.b32 %r397, %f144;
mov.u32 %r398, 8;
shfl.sync.bfly.b32 %r399|%p8, %r397, %r398, %r393, %r395;
mov.b32 %f145, %r399;
add.ftz.f32 %f146, %f144, %f145;
mov.b32 %r400, %f146;
mov.u32 %r401, 4;
shfl.sync.bfly.b32 %r402|%p9, %r400, %r401, %r393, %r395;
mov.b32 %f147, %r402;
add.ftz.f32 %f148, %f146, %f147;
mov.b32 %r403, %f148;
shfl.sync.bfly.b32 %r404|%p10, %r403, %r391, %r393, %r395;
mov.b32 %f149, %r404;
add.ftz.f32 %f150, %f148, %f149;
mov.b32 %r405, %f150;
mov.u32 %r406, 1;
shfl.sync.bfly.b32 %r407|%p11, %r405, %r406, %r393, %r395;
mov.b32 %f151, %r407;
add.ftz.f32 %f43, %f150, %f151;
setp.lt.u32 %p12, %r4, %r32;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
cvta.to.global.u64 %rd68, %rd21;
mul.wide.u32 %rd69, %r388, 4;
add.s64 %rd15, %rd68, %rd69;
not.pred %p14, %p1;
@%p14 bra $L__BB289_8;

st.global.f32 [%rd15], %f43;

$L__BB289_8:
ld.shared.f32 %f152, [%r27+128];
add.ftz.f32 %f153, %f152, %f321;
mov.b32 %r408, %f153;
shfl.sync.bfly.b32 %r412|%p15, %r408, %r394, %r393, %r395;
mov.b32 %f154, %r412;
add.ftz.f32 %f155, %f153, %f154;
mov.b32 %r413, %f155;
shfl.sync.bfly.b32 %r415|%p16, %r413, %r398, %r393, %r395;
mov.b32 %f156, %r415;
add.ftz.f32 %f157, %f155, %f156;
mov.b32 %r416, %f157;
shfl.sync.bfly.b32 %r418|%p17, %r416, %r401, %r393, %r395;
mov.b32 %f158, %r418;
add.ftz.f32 %f159, %f157, %f158;
mov.b32 %r419, %f159;
shfl.sync.bfly.b32 %r421|%p18, %r419, %r391, %r393, %r395;
mov.b32 %f160, %r421;
add.ftz.f32 %f161, %f159, %f160;
mov.b32 %r422, %f161;
shfl.sync.bfly.b32 %r424|%p19, %r422, %r406, %r393, %r395;
mov.b32 %f162, %r424;
add.ftz.f32 %f44, %f161, %f162;
setp.eq.s32 %p20, %r2, 1;
or.b32 %r428, %r4, 1;
setp.lt.u32 %p21, %r428, %r32;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB289_10;

st.global.f32 [%rd15+4], %f44;

$L__BB289_10:
ld.shared.f32 %f163, [%r27+256];
add.ftz.f32 %f164, %f163, %f320;
mov.b32 %r429, %f164;
mov.u32 %r430, 31;
mov.u32 %r431, 16;
mov.u32 %r432, -1;
shfl.sync.bfly.b32 %r433|%p23, %r429, %r431, %r430, %r432;
mov.b32 %f165, %r433;
add.ftz.f32 %f166, %f164, %f165;
mov.b32 %r434, %f166;
mov.u32 %r435, 8;
shfl.sync.bfly.b32 %r436|%p24, %r434, %r435, %r430, %r432;
mov.b32 %f167, %r436;
add.ftz.f32 %f168, %f166, %f167;
mov.b32 %r437, %f168;
mov.u32 %r438, 4;
shfl.sync.bfly.b32 %r439|%p25, %r437, %r438, %r430, %r432;
mov.b32 %f169, %r439;
add.ftz.f32 %f170, %f168, %f169;
mov.b32 %r440, %f170;
mov.u32 %r441, 2;
shfl.sync.bfly.b32 %r442|%p26, %r440, %r441, %r430, %r432;
mov.b32 %f171, %r442;
add.ftz.f32 %f172, %f170, %f171;
mov.b32 %r443, %f172;
mov.u32 %r444, 1;
shfl.sync.bfly.b32 %r445|%p27, %r443, %r444, %r430, %r432;
mov.b32 %f173, %r445;
add.ftz.f32 %f45, %f172, %f173;
@%p14 bra $L__BB289_12;

cvt.u64.u32 %rd70, %r32;
add.s64 %rd71, %rd70, %rd14;
shl.b64 %rd73, %rd71, 2;
add.s64 %rd74, %rd68, %rd73;
st.global.f32 [%rd74], %f45;

$L__BB289_12:
ld.shared.f32 %f174, [%r27+384];
add.ftz.f32 %f175, %f174, %f319;
mov.b32 %r446, %f175;
shfl.sync.bfly.b32 %r450|%p29, %r446, %r431, %r430, %r432;
mov.b32 %f176, %r450;
add.ftz.f32 %f177, %f175, %f176;
mov.b32 %r451, %f177;
shfl.sync.bfly.b32 %r453|%p30, %r451, %r435, %r430, %r432;
mov.b32 %f178, %r453;
add.ftz.f32 %f179, %f177, %f178;
mov.b32 %r454, %f179;
shfl.sync.bfly.b32 %r456|%p31, %r454, %r438, %r430, %r432;
mov.b32 %f180, %r456;
add.ftz.f32 %f181, %f179, %f180;
mov.b32 %r457, %f181;
shfl.sync.bfly.b32 %r459|%p32, %r457, %r441, %r430, %r432;
mov.b32 %f182, %r459;
add.ftz.f32 %f183, %f181, %f182;
mov.b32 %r460, %f183;
shfl.sync.bfly.b32 %r462|%p33, %r460, %r444, %r430, %r432;
mov.b32 %f184, %r462;
add.ftz.f32 %f46, %f183, %f184;
@%p22 bra $L__BB289_14;

add.s32 %r463, %r32, 1;
cvt.u64.u32 %rd75, %r463;
add.s64 %rd76, %rd75, %rd14;
shl.b64 %rd78, %rd76, 2;
add.s64 %rd79, %rd68, %rd78;
st.global.f32 [%rd79], %f46;

$L__BB289_14:
shl.b32 %r464, %r32, 1;
mov.u32 %r465, 1;
ld.shared.f32 %f185, [%r27+512];
add.ftz.f32 %f186, %f185, %f318;
mov.b32 %r466, %f186;
mov.u32 %r467, 31;
mov.u32 %r468, 16;
mov.u32 %r469, -1;
shfl.sync.bfly.b32 %r470|%p35, %r466, %r468, %r467, %r469;
mov.b32 %f187, %r470;
add.ftz.f32 %f188, %f186, %f187;
mov.b32 %r471, %f188;
mov.u32 %r472, 8;
shfl.sync.bfly.b32 %r473|%p36, %r471, %r472, %r467, %r469;
mov.b32 %f189, %r473;
add.ftz.f32 %f190, %f188, %f189;
mov.b32 %r474, %f190;
mov.u32 %r475, 4;
shfl.sync.bfly.b32 %r476|%p37, %r474, %r475, %r467, %r469;
mov.b32 %f191, %r476;
add.ftz.f32 %f192, %f190, %f191;
mov.b32 %r477, %f192;
mov.u32 %r478, 2;
shfl.sync.bfly.b32 %r479|%p38, %r477, %r478, %r467, %r469;
mov.b32 %f193, %r479;
add.ftz.f32 %f194, %f192, %f193;
mov.b32 %r480, %f194;
shfl.sync.bfly.b32 %r481|%p39, %r480, %r465, %r467, %r469;
mov.b32 %f195, %r481;
add.ftz.f32 %f47, %f194, %f195;
cvt.u64.u32 %rd80, %r464;
add.s64 %rd81, %rd80, %rd14;
shl.b64 %rd83, %rd81, 2;
add.s64 %rd16, %rd68, %rd83;
@%p14 bra $L__BB289_16;

st.global.f32 [%rd16], %f47;

$L__BB289_16:
ld.shared.f32 %f196, [%r27+640];
add.ftz.f32 %f197, %f196, %f317;
mov.b32 %r482, %f197;
shfl.sync.bfly.b32 %r486|%p41, %r482, %r468, %r467, %r469;
mov.b32 %f198, %r486;
add.ftz.f32 %f199, %f197, %f198;
mov.b32 %r487, %f199;
shfl.sync.bfly.b32 %r489|%p42, %r487, %r472, %r467, %r469;
mov.b32 %f200, %r489;
add.ftz.f32 %f201, %f199, %f200;
mov.b32 %r490, %f201;
shfl.sync.bfly.b32 %r492|%p43, %r490, %r475, %r467, %r469;
mov.b32 %f202, %r492;
add.ftz.f32 %f203, %f201, %f202;
mov.b32 %r493, %f203;
shfl.sync.bfly.b32 %r495|%p44, %r493, %r478, %r467, %r469;
mov.b32 %f204, %r495;
add.ftz.f32 %f205, %f203, %f204;
mov.b32 %r496, %f205;
shfl.sync.bfly.b32 %r498|%p45, %r496, %r465, %r467, %r469;
mov.b32 %f206, %r498;
add.ftz.f32 %f48, %f205, %f206;
@%p22 bra $L__BB289_18;

st.global.f32 [%rd16+4], %f48;

$L__BB289_18:
mul.lo.s32 %r28, %r32, 3;
ld.shared.f32 %f207, [%r27+768];
add.ftz.f32 %f208, %f207, %f316;
mov.b32 %r499, %f208;
mov.u32 %r500, 31;
mov.u32 %r501, 16;
mov.u32 %r502, -1;
shfl.sync.bfly.b32 %r503|%p47, %r499, %r501, %r500, %r502;
mov.b32 %f209, %r503;
add.ftz.f32 %f210, %f208, %f209;
mov.b32 %r504, %f210;
mov.u32 %r505, 8;
shfl.sync.bfly.b32 %r506|%p48, %r504, %r505, %r500, %r502;
mov.b32 %f211, %r506;
add.ftz.f32 %f212, %f210, %f211;
mov.b32 %r507, %f212;
mov.u32 %r508, 4;
shfl.sync.bfly.b32 %r509|%p49, %r507, %r508, %r500, %r502;
mov.b32 %f213, %r509;
add.ftz.f32 %f214, %f212, %f213;
mov.b32 %r510, %f214;
mov.u32 %r511, 2;
shfl.sync.bfly.b32 %r512|%p50, %r510, %r511, %r500, %r502;
mov.b32 %f215, %r512;
add.ftz.f32 %f216, %f214, %f215;
mov.b32 %r513, %f216;
mov.u32 %r514, 1;
shfl.sync.bfly.b32 %r515|%p51, %r513, %r514, %r500, %r502;
mov.b32 %f217, %r515;
add.ftz.f32 %f49, %f216, %f217;
@%p14 bra $L__BB289_20;

cvt.u64.u32 %rd84, %r28;
add.s64 %rd85, %rd84, %rd14;
shl.b64 %rd87, %rd85, 2;
add.s64 %rd88, %rd68, %rd87;
st.global.f32 [%rd88], %f49;

$L__BB289_20:
ld.shared.f32 %f218, [%r27+896];
add.ftz.f32 %f219, %f218, %f315;
mov.b32 %r516, %f219;
shfl.sync.bfly.b32 %r520|%p53, %r516, %r501, %r500, %r502;
mov.b32 %f220, %r520;
add.ftz.f32 %f221, %f219, %f220;
mov.b32 %r521, %f221;
shfl.sync.bfly.b32 %r523|%p54, %r521, %r505, %r500, %r502;
mov.b32 %f222, %r523;
add.ftz.f32 %f223, %f221, %f222;
mov.b32 %r524, %f223;
shfl.sync.bfly.b32 %r526|%p55, %r524, %r508, %r500, %r502;
mov.b32 %f224, %r526;
add.ftz.f32 %f225, %f223, %f224;
mov.b32 %r527, %f225;
shfl.sync.bfly.b32 %r529|%p56, %r527, %r511, %r500, %r502;
mov.b32 %f226, %r529;
add.ftz.f32 %f227, %f225, %f226;
mov.b32 %r530, %f227;
shfl.sync.bfly.b32 %r532|%p57, %r530, %r514, %r500, %r502;
mov.b32 %f228, %r532;
add.ftz.f32 %f50, %f227, %f228;
@%p22 bra $L__BB289_22;

add.s32 %r533, %r28, 1;
cvt.u64.u32 %rd89, %r533;
add.s64 %rd90, %rd89, %rd14;
shl.b64 %rd92, %rd90, 2;
add.s64 %rd93, %rd68, %rd92;
st.global.f32 [%rd93], %f50;

$L__BB289_22:
shl.b32 %r534, %r32, 2;
mov.u32 %r535, 2;
ld.shared.f32 %f229, [%r27+1024];
add.ftz.f32 %f230, %f229, %f314;
mov.b32 %r536, %f230;
mov.u32 %r537, 31;
mov.u32 %r538, 16;
mov.u32 %r539, -1;
shfl.sync.bfly.b32 %r540|%p59, %r536, %r538, %r537, %r539;
mov.b32 %f231, %r540;
add.ftz.f32 %f232, %f230, %f231;
mov.b32 %r541, %f232;
mov.u32 %r542, 8;
shfl.sync.bfly.b32 %r543|%p60, %r541, %r542, %r537, %r539;
mov.b32 %f233, %r543;
add.ftz.f32 %f234, %f232, %f233;
mov.b32 %r544, %f234;
mov.u32 %r545, 4;
shfl.sync.bfly.b32 %r546|%p61, %r544, %r545, %r537, %r539;
mov.b32 %f235, %r546;
add.ftz.f32 %f236, %f234, %f235;
mov.b32 %r547, %f236;
shfl.sync.bfly.b32 %r548|%p62, %r547, %r535, %r537, %r539;
mov.b32 %f237, %r548;
add.ftz.f32 %f238, %f236, %f237;
mov.b32 %r549, %f238;
mov.u32 %r550, 1;
shfl.sync.bfly.b32 %r551|%p63, %r549, %r550, %r537, %r539;
mov.b32 %f239, %r551;
add.ftz.f32 %f51, %f238, %f239;
cvt.u64.u32 %rd94, %r534;
add.s64 %rd95, %rd94, %rd14;
shl.b64 %rd97, %rd95, 2;
add.s64 %rd17, %rd68, %rd97;
@%p14 bra $L__BB289_24;

st.global.f32 [%rd17], %f51;

$L__BB289_24:
ld.shared.f32 %f240, [%r27+1152];
add.ftz.f32 %f241, %f240, %f313;
mov.b32 %r552, %f241;
shfl.sync.bfly.b32 %r556|%p65, %r552, %r538, %r537, %r539;
mov.b32 %f242, %r556;
add.ftz.f32 %f243, %f241, %f242;
mov.b32 %r557, %f243;
shfl.sync.bfly.b32 %r559|%p66, %r557, %r542, %r537, %r539;
mov.b32 %f244, %r559;
add.ftz.f32 %f245, %f243, %f244;
mov.b32 %r560, %f245;
shfl.sync.bfly.b32 %r562|%p67, %r560, %r545, %r537, %r539;
mov.b32 %f246, %r562;
add.ftz.f32 %f247, %f245, %f246;
mov.b32 %r563, %f247;
shfl.sync.bfly.b32 %r565|%p68, %r563, %r535, %r537, %r539;
mov.b32 %f248, %r565;
add.ftz.f32 %f249, %f247, %f248;
mov.b32 %r566, %f249;
shfl.sync.bfly.b32 %r568|%p69, %r566, %r550, %r537, %r539;
mov.b32 %f250, %r568;
add.ftz.f32 %f52, %f249, %f250;
@%p22 bra $L__BB289_26;

st.global.f32 [%rd17+4], %f52;

$L__BB289_26:
mul.lo.s32 %r29, %r32, 5;
ld.shared.f32 %f251, [%r27+1280];
add.ftz.f32 %f252, %f251, %f312;
mov.b32 %r569, %f252;
mov.u32 %r570, 31;
mov.u32 %r571, 16;
mov.u32 %r572, -1;
shfl.sync.bfly.b32 %r573|%p71, %r569, %r571, %r570, %r572;
mov.b32 %f253, %r573;
add.ftz.f32 %f254, %f252, %f253;
mov.b32 %r574, %f254;
mov.u32 %r575, 8;
shfl.sync.bfly.b32 %r576|%p72, %r574, %r575, %r570, %r572;
mov.b32 %f255, %r576;
add.ftz.f32 %f256, %f254, %f255;
mov.b32 %r577, %f256;
mov.u32 %r578, 4;
shfl.sync.bfly.b32 %r579|%p73, %r577, %r578, %r570, %r572;
mov.b32 %f257, %r579;
add.ftz.f32 %f258, %f256, %f257;
mov.b32 %r580, %f258;
mov.u32 %r581, 2;
shfl.sync.bfly.b32 %r582|%p74, %r580, %r581, %r570, %r572;
mov.b32 %f259, %r582;
add.ftz.f32 %f260, %f258, %f259;
mov.b32 %r583, %f260;
mov.u32 %r584, 1;
shfl.sync.bfly.b32 %r585|%p75, %r583, %r584, %r570, %r572;
mov.b32 %f261, %r585;
add.ftz.f32 %f53, %f260, %f261;
@%p14 bra $L__BB289_28;

cvt.u64.u32 %rd98, %r29;
add.s64 %rd99, %rd98, %rd14;
shl.b64 %rd101, %rd99, 2;
add.s64 %rd102, %rd68, %rd101;
st.global.f32 [%rd102], %f53;

$L__BB289_28:
ld.shared.f32 %f262, [%r27+1408];
add.ftz.f32 %f263, %f262, %f311;
mov.b32 %r586, %f263;
shfl.sync.bfly.b32 %r590|%p77, %r586, %r571, %r570, %r572;
mov.b32 %f264, %r590;
add.ftz.f32 %f265, %f263, %f264;
mov.b32 %r591, %f265;
shfl.sync.bfly.b32 %r593|%p78, %r591, %r575, %r570, %r572;
mov.b32 %f266, %r593;
add.ftz.f32 %f267, %f265, %f266;
mov.b32 %r594, %f267;
shfl.sync.bfly.b32 %r596|%p79, %r594, %r578, %r570, %r572;
mov.b32 %f268, %r596;
add.ftz.f32 %f269, %f267, %f268;
mov.b32 %r597, %f269;
shfl.sync.bfly.b32 %r599|%p80, %r597, %r581, %r570, %r572;
mov.b32 %f270, %r599;
add.ftz.f32 %f271, %f269, %f270;
mov.b32 %r600, %f271;
shfl.sync.bfly.b32 %r602|%p81, %r600, %r584, %r570, %r572;
mov.b32 %f272, %r602;
add.ftz.f32 %f54, %f271, %f272;
@%p22 bra $L__BB289_30;

add.s32 %r603, %r29, 1;
cvt.u64.u32 %rd103, %r603;
add.s64 %rd104, %rd103, %rd14;
shl.b64 %rd106, %rd104, 2;
add.s64 %rd107, %rd68, %rd106;
st.global.f32 [%rd107], %f54;

$L__BB289_30:
mul.lo.s32 %r604, %r32, 6;
ld.shared.f32 %f273, [%r27+1536];
add.ftz.f32 %f274, %f273, %f310;
mov.b32 %r605, %f274;
mov.u32 %r606, 31;
mov.u32 %r607, 16;
mov.u32 %r608, -1;
shfl.sync.bfly.b32 %r609|%p83, %r605, %r607, %r606, %r608;
mov.b32 %f275, %r609;
add.ftz.f32 %f276, %f274, %f275;
mov.b32 %r610, %f276;
mov.u32 %r611, 8;
shfl.sync.bfly.b32 %r612|%p84, %r610, %r611, %r606, %r608;
mov.b32 %f277, %r612;
add.ftz.f32 %f278, %f276, %f277;
mov.b32 %r613, %f278;
mov.u32 %r614, 4;
shfl.sync.bfly.b32 %r615|%p85, %r613, %r614, %r606, %r608;
mov.b32 %f279, %r615;
add.ftz.f32 %f280, %f278, %f279;
mov.b32 %r616, %f280;
mov.u32 %r617, 2;
shfl.sync.bfly.b32 %r618|%p86, %r616, %r617, %r606, %r608;
mov.b32 %f281, %r618;
add.ftz.f32 %f282, %f280, %f281;
mov.b32 %r619, %f282;
mov.u32 %r620, 1;
shfl.sync.bfly.b32 %r621|%p87, %r619, %r620, %r606, %r608;
mov.b32 %f283, %r621;
add.ftz.f32 %f55, %f282, %f283;
cvt.u64.u32 %rd108, %r604;
add.s64 %rd109, %rd108, %rd14;
shl.b64 %rd111, %rd109, 2;
add.s64 %rd18, %rd68, %rd111;
@%p14 bra $L__BB289_32;

st.global.f32 [%rd18], %f55;

$L__BB289_32:
ld.shared.f32 %f284, [%r27+1664];
add.ftz.f32 %f285, %f284, %f309;
mov.b32 %r622, %f285;
shfl.sync.bfly.b32 %r626|%p89, %r622, %r607, %r606, %r608;
mov.b32 %f286, %r626;
add.ftz.f32 %f287, %f285, %f286;
mov.b32 %r627, %f287;
shfl.sync.bfly.b32 %r629|%p90, %r627, %r611, %r606, %r608;
mov.b32 %f288, %r629;
add.ftz.f32 %f289, %f287, %f288;
mov.b32 %r630, %f289;
shfl.sync.bfly.b32 %r632|%p91, %r630, %r614, %r606, %r608;
mov.b32 %f290, %r632;
add.ftz.f32 %f291, %f289, %f290;
mov.b32 %r633, %f291;
shfl.sync.bfly.b32 %r635|%p92, %r633, %r617, %r606, %r608;
mov.b32 %f292, %r635;
add.ftz.f32 %f293, %f291, %f292;
mov.b32 %r636, %f293;
shfl.sync.bfly.b32 %r638|%p93, %r636, %r620, %r606, %r608;
mov.b32 %f294, %r638;
add.ftz.f32 %f56, %f293, %f294;
@%p22 bra $L__BB289_34;

st.global.f32 [%rd18+4], %f56;

$L__BB289_34:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type20ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type20ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type20ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type20ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type20ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type20ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type20ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type20ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type20ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type20ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<107>;
.reg .b16 %rs<29>;
.reg .f32 %f<369>;
.reg .b32 %r<721>;
.reg .b64 %rd<130>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type20ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[2048];

ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type20ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd20, [_Z13mul_mat_vec_qIL9ggml_type20ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd21, [_Z13mul_mat_vec_qIL9ggml_type20ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r49, [_Z13mul_mat_vec_qIL9ggml_type20ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type20ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type20ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type20ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type20ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type20ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type20ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r46, [_Z13mul_mat_vec_qIL9ggml_type20ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r47, [_Z13mul_mat_vec_qIL9ggml_type20ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r48, [_Z13mul_mat_vec_qIL9ggml_type20ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type20ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type20ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type20ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type20ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r50, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r50, %r2;
mov.u32 %r51, %ctaid.x;
shl.b32 %r4, %r51, 1;
shr.u32 %r52, %r3, 31;
add.s32 %r53, %r3, %r52;
shr.s32 %r720, %r53, 1;
shr.u32 %r6, %r49, 5;
setp.ge.s32 %p3, %r720, %r6;
mov.f32 %f353, 0f00000000;
mov.f32 %f354, %f353;
mov.f32 %f355, %f353;
mov.f32 %f356, %f353;
mov.f32 %f357, %f353;
mov.f32 %f358, %f353;
mov.f32 %f359, %f353;
mov.f32 %f360, %f353;
mov.f32 %f361, %f353;
mov.f32 %f362, %f353;
mov.f32 %f363, %f353;
mov.f32 %f364, %f353;
mov.f32 %f365, %f353;
mov.f32 %f366, %f353;
mov.f32 %f367, %f353;
mov.f32 %f368, %f353;
@%p3 bra $L__BB290_3;

mov.u32 %r54, %ctaid.y;
mov.u32 %r55, %ctaid.z;
mul.lo.s32 %r56, %r54, %r41;
cvt.u64.u32 %rd22, %r56;
mul.lo.s32 %r57, %r55, %r47;
cvt.u64.u32 %rd23, %r57;
add.s64 %rd1, %rd23, %rd22;
mul.hi.u32 %r58, %r55, %r43;
add.s32 %r59, %r58, %r55;
shr.u32 %r60, %r59, %r44;
mul.lo.s32 %r61, %r60, %r46;
mul.hi.u32 %r62, %r54, %r37;
add.s32 %r63, %r62, %r54;
shr.u32 %r64, %r63, %r38;
mul.lo.s32 %r65, %r64, %r40;
cvta.to.global.u64 %rd24, %rd19;
and.b32 %r68, %r53, -2;
sub.s32 %r69, %r3, %r68;
shl.b32 %r70, %r69, 1;
cvt.s64.s32 %rd2, %r70;
shl.b32 %r71, %r69, 2;
shl.b32 %r72, %r35, 1;
add.s32 %r719, %r720, %r72;
mad.lo.s32 %r718, %r35, 3, %r720;
shl.b32 %r73, %r35, 2;
add.s32 %r717, %r720, %r73;
mad.lo.s32 %r716, %r35, 5, %r720;
mad.lo.s32 %r715, %r35, 6, %r720;
mad.lo.s32 %r714, %r35, 7, %r720;
add.s32 %r713, %r720, %r35;
add.s32 %r74, %r720, %r61;
add.s32 %r75, %r74, %r65;
add.s32 %r76, %r4, 1;
mad.lo.s32 %r77, %r34, %r76, %r75;
mul.wide.s32 %rd25, %r77, 18;
add.s64 %rd129, %rd24, %rd25;
mad.lo.s32 %r78, %r4, %r34, %r720;
add.s32 %r79, %r78, %r61;
add.s32 %r80, %r79, %r65;
mul.wide.s32 %rd26, %r80, 18;
add.s64 %rd27, %rd24, %rd26;
add.s64 %rd128, %rd27, 4;
mul.wide.s32 %rd28, %r71, 2;
add.s64 %rd29, %rd26, %rd28;
add.s64 %rd30, %rd24, %rd29;
add.s64 %rd127, %rd30, 8;
or.b64 %rd6, %rd28, 4;
cvta.to.global.u64 %rd7, %rd20;
mov.f32 %f353, 0f00000000;
shl.b64 %rd35, %rd2, 2;

$L__BB290_2:
cvt.u64.u32 %rd31, %r720;
add.s64 %rd32, %rd1, %rd31;
mul.lo.s64 %rd33, %rd32, 36;
add.s64 %rd34, %rd7, %rd33;
add.s64 %rd36, %rd34, %rd35;
ld.global.nc.u16 %rs17, [%rd127+-6];
cvt.u32.u16 %r353, %rs17;
ld.global.nc.u16 %rs18, [%rd127+-4];
mov.b32 %r354, {%rs17, %rs18};
and.b32 %r355, %r354, -2004318072;
shr.u32 %r356, %r355, 1;
or.b32 %r357, %r356, 839922192;
and.b32 %r358, %r353, 30583;
mov.u32 %r359, -152379953;
mov.u32 %r360, -1079142271;
prmt.b32 %r361, %r360, %r359, %r358;
mov.u32 %r362, 1901675829;
mov.u32 %r363, 639175937;
prmt.b32 %r364, %r363, %r362, %r358;
and.b32 %r365, %r357, 30292;
prmt.b32 %r366, %r361, %r364, %r365;
and.b16 %rs19, %rs18, 30583;
cvt.u32.u16 %r367, %rs19;
prmt.b32 %r368, %r360, %r359, %r367;
prmt.b32 %r369, %r363, %r362, %r367;
shr.u32 %r370, %r357, 16;
prmt.b32 %r371, %r368, %r369, %r370;
mov.u32 %r372, 25632;
prmt.b32 %r320, %r366, %r371, %r372;
mov.u32 %r373, 30001;
prmt.b32 %r324, %r366, %r371, %r373;
ld.global.nc.u32 %r83, [%rd36+4];
mov.u32 %r339, 0;

dp4a.s32.s32 %r81, %r320, %r83, %r339;

ld.global.nc.u32 %r87, [%rd36+20];

dp4a.s32.s32 %r85, %r324, %r87, %r81;

ld.global.nc.u16 %rs20, [%rd127+-2];
cvt.u32.u16 %r374, %rs20;
ld.global.nc.u16 %rs21, [%rd127];
mov.b32 %r375, {%rs20, %rs21};
and.b32 %r376, %r375, -2004318072;
shr.u32 %r377, %r376, 1;
or.b32 %r378, %r377, 839922192;
and.b32 %r379, %r374, 30583;
prmt.b32 %r380, %r360, %r359, %r379;
prmt.b32 %r381, %r363, %r362, %r379;
and.b32 %r382, %r378, 30292;
prmt.b32 %r383, %r380, %r381, %r382;
and.b16 %rs22, %rs21, 30583;
cvt.u32.u16 %r384, %rs22;
prmt.b32 %r385, %r360, %r359, %r384;
prmt.b32 %r386, %r363, %r362, %r384;
shr.u32 %r387, %r378, 16;
prmt.b32 %r388, %r385, %r386, %r387;
prmt.b32 %r328, %r383, %r388, %r372;
prmt.b32 %r332, %r383, %r388, %r373;
ld.global.nc.u32 %r91, [%rd36+8];

dp4a.s32.s32 %r89, %r328, %r91, %r85;

ld.global.nc.u32 %r95, [%rd36+24];

dp4a.s32.s32 %r93, %r332, %r95, %r89;

ld.global.nc.u16 %rs1, [%rd128+-4];

{ cvt.f32.f16 %f97, %rs1;}


ld.global.nc.u32 %r97, [%rd34];

{.reg .f16 low,high;
mov.b32 {low,high},%r97;
cvt.f32.f16 %f98, low;}


mul.ftz.f32 %f129, %f97, %f98;
cvt.rn.f32.s32 %f130, %r93;
fma.rn.ftz.f32 %f368, %f129, %f130, %f368;
add.s64 %rd37, %rd129, %rd6;
ld.global.nc.u16 %rs23, [%rd37+-2];
cvt.u32.u16 %r389, %rs23;
ld.global.nc.u16 %rs24, [%rd37];
mov.b32 %r390, {%rs23, %rs24};
and.b32 %r391, %r390, -2004318072;
shr.u32 %r392, %r391, 1;
or.b32 %r393, %r392, 839922192;
and.b32 %r394, %r389, 30583;
prmt.b32 %r395, %r360, %r359, %r394;
prmt.b32 %r396, %r363, %r362, %r394;
and.b32 %r397, %r393, 30292;
prmt.b32 %r398, %r395, %r396, %r397;
and.b16 %rs25, %rs24, 30583;
cvt.u32.u16 %r399, %rs25;
prmt.b32 %r400, %r360, %r359, %r399;
prmt.b32 %r401, %r363, %r362, %r399;
shr.u32 %r402, %r393, 16;
prmt.b32 %r403, %r400, %r401, %r402;
prmt.b32 %r337, %r398, %r403, %r372;
prmt.b32 %r341, %r398, %r403, %r373;

dp4a.s32.s32 %r98, %r337, %r83, %r339;


dp4a.s32.s32 %r102, %r341, %r87, %r98;

ld.global.nc.u16 %rs26, [%rd37+2];
cvt.u32.u16 %r404, %rs26;
ld.global.nc.u16 %rs27, [%rd37+4];
mov.b32 %r405, {%rs26, %rs27};
and.b32 %r406, %r405, -2004318072;
shr.u32 %r407, %r406, 1;
or.b32 %r408, %r407, 839922192;
and.b32 %r409, %r404, 30583;
prmt.b32 %r410, %r360, %r359, %r409;
prmt.b32 %r411, %r363, %r362, %r409;
and.b32 %r412, %r408, 30292;
prmt.b32 %r413, %r410, %r411, %r412;
and.b16 %rs28, %rs27, 30583;
cvt.u32.u16 %r414, %rs28;
prmt.b32 %r415, %r360, %r359, %r414;
prmt.b32 %r416, %r363, %r362, %r414;
shr.u32 %r417, %r408, 16;
prmt.b32 %r418, %r415, %r416, %r417;
prmt.b32 %r345, %r413, %r418, %r372;
prmt.b32 %r349, %r413, %r418, %r373;

dp4a.s32.s32 %r106, %r345, %r91, %r102;


dp4a.s32.s32 %r110, %r349, %r95, %r106;

ld.global.nc.u16 %rs2, [%rd129];

{ cvt.f32.f16 %f99, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r97;
cvt.f32.f16 %f100, low;}


mul.ftz.f32 %f131, %f99, %f100;
cvt.rn.f32.s32 %f132, %r110;
fma.rn.ftz.f32 %f367, %f131, %f132, %f367;
cvt.u64.u32 %rd38, %r713;
add.s64 %rd39, %rd1, %rd38;
mul.lo.s64 %rd40, %rd39, 36;
add.s64 %rd41, %rd7, %rd40;
add.s64 %rd42, %rd41, %rd35;
ld.global.nc.u32 %r117, [%rd42+4];

dp4a.s32.s32 %r115, %r320, %r117, %r339;

ld.global.nc.u32 %r121, [%rd42+20];

dp4a.s32.s32 %r119, %r324, %r121, %r115;

ld.global.nc.u32 %r125, [%rd42+8];

dp4a.s32.s32 %r123, %r328, %r125, %r119;

ld.global.nc.u32 %r129, [%rd42+24];

dp4a.s32.s32 %r127, %r332, %r129, %r123;


{ cvt.f32.f16 %f101, %rs1;}


ld.global.nc.u32 %r131, [%rd41];

{.reg .f16 low,high;
mov.b32 {low,high},%r131;
cvt.f32.f16 %f102, low;}


mul.ftz.f32 %f133, %f101, %f102;
cvt.rn.f32.s32 %f134, %r127;
fma.rn.ftz.f32 %f366, %f133, %f134, %f366;

dp4a.s32.s32 %r132, %r337, %r117, %r339;


dp4a.s32.s32 %r136, %r341, %r121, %r132;


dp4a.s32.s32 %r140, %r345, %r125, %r136;


dp4a.s32.s32 %r144, %r349, %r129, %r140;


{ cvt.f32.f16 %f103, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r131;
cvt.f32.f16 %f104, low;}


mul.ftz.f32 %f135, %f103, %f104;
cvt.rn.f32.s32 %f136, %r144;
fma.rn.ftz.f32 %f365, %f135, %f136, %f365;
cvt.u64.u32 %rd43, %r719;
add.s64 %rd44, %rd1, %rd43;
mul.lo.s64 %rd45, %rd44, 36;
add.s64 %rd46, %rd7, %rd45;
add.s64 %rd47, %rd46, %rd35;
ld.global.nc.u32 %r151, [%rd47+4];

dp4a.s32.s32 %r149, %r320, %r151, %r339;

ld.global.nc.u32 %r155, [%rd47+20];

dp4a.s32.s32 %r153, %r324, %r155, %r149;

ld.global.nc.u32 %r159, [%rd47+8];

dp4a.s32.s32 %r157, %r328, %r159, %r153;

ld.global.nc.u32 %r163, [%rd47+24];

dp4a.s32.s32 %r161, %r332, %r163, %r157;


{ cvt.f32.f16 %f105, %rs1;}


ld.global.nc.u32 %r165, [%rd46];

{.reg .f16 low,high;
mov.b32 {low,high},%r165;
cvt.f32.f16 %f106, low;}


mul.ftz.f32 %f137, %f105, %f106;
cvt.rn.f32.s32 %f138, %r161;
fma.rn.ftz.f32 %f364, %f137, %f138, %f364;

dp4a.s32.s32 %r166, %r337, %r151, %r339;


dp4a.s32.s32 %r170, %r341, %r155, %r166;


dp4a.s32.s32 %r174, %r345, %r159, %r170;


dp4a.s32.s32 %r178, %r349, %r163, %r174;


{ cvt.f32.f16 %f107, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r165;
cvt.f32.f16 %f108, low;}


mul.ftz.f32 %f139, %f107, %f108;
cvt.rn.f32.s32 %f140, %r178;
fma.rn.ftz.f32 %f363, %f139, %f140, %f363;
cvt.u64.u32 %rd48, %r718;
add.s64 %rd49, %rd1, %rd48;
mul.lo.s64 %rd50, %rd49, 36;
add.s64 %rd51, %rd7, %rd50;
add.s64 %rd52, %rd51, %rd35;
ld.global.nc.u32 %r185, [%rd52+4];

dp4a.s32.s32 %r183, %r320, %r185, %r339;

ld.global.nc.u32 %r189, [%rd52+20];

dp4a.s32.s32 %r187, %r324, %r189, %r183;

ld.global.nc.u32 %r193, [%rd52+8];

dp4a.s32.s32 %r191, %r328, %r193, %r187;

ld.global.nc.u32 %r197, [%rd52+24];

dp4a.s32.s32 %r195, %r332, %r197, %r191;


{ cvt.f32.f16 %f109, %rs1;}


ld.global.nc.u32 %r199, [%rd51];

{.reg .f16 low,high;
mov.b32 {low,high},%r199;
cvt.f32.f16 %f110, low;}


mul.ftz.f32 %f141, %f109, %f110;
cvt.rn.f32.s32 %f142, %r195;
fma.rn.ftz.f32 %f362, %f141, %f142, %f362;

dp4a.s32.s32 %r200, %r337, %r185, %r339;


dp4a.s32.s32 %r204, %r341, %r189, %r200;


dp4a.s32.s32 %r208, %r345, %r193, %r204;


dp4a.s32.s32 %r212, %r349, %r197, %r208;


{ cvt.f32.f16 %f111, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r199;
cvt.f32.f16 %f112, low;}


mul.ftz.f32 %f143, %f111, %f112;
cvt.rn.f32.s32 %f144, %r212;
fma.rn.ftz.f32 %f361, %f143, %f144, %f361;
cvt.u64.u32 %rd53, %r717;
add.s64 %rd54, %rd1, %rd53;
mul.lo.s64 %rd55, %rd54, 36;
add.s64 %rd56, %rd7, %rd55;
add.s64 %rd57, %rd56, %rd35;
ld.global.nc.u32 %r219, [%rd57+4];

dp4a.s32.s32 %r217, %r320, %r219, %r339;

ld.global.nc.u32 %r223, [%rd57+20];

dp4a.s32.s32 %r221, %r324, %r223, %r217;

ld.global.nc.u32 %r227, [%rd57+8];

dp4a.s32.s32 %r225, %r328, %r227, %r221;

ld.global.nc.u32 %r231, [%rd57+24];

dp4a.s32.s32 %r229, %r332, %r231, %r225;


{ cvt.f32.f16 %f113, %rs1;}


ld.global.nc.u32 %r233, [%rd56];

{.reg .f16 low,high;
mov.b32 {low,high},%r233;
cvt.f32.f16 %f114, low;}


mul.ftz.f32 %f145, %f113, %f114;
cvt.rn.f32.s32 %f146, %r229;
fma.rn.ftz.f32 %f360, %f145, %f146, %f360;

dp4a.s32.s32 %r234, %r337, %r219, %r339;


dp4a.s32.s32 %r238, %r341, %r223, %r234;


dp4a.s32.s32 %r242, %r345, %r227, %r238;


dp4a.s32.s32 %r246, %r349, %r231, %r242;


{ cvt.f32.f16 %f115, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r233;
cvt.f32.f16 %f116, low;}


mul.ftz.f32 %f147, %f115, %f116;
cvt.rn.f32.s32 %f148, %r246;
fma.rn.ftz.f32 %f359, %f147, %f148, %f359;
cvt.u64.u32 %rd58, %r716;
add.s64 %rd59, %rd1, %rd58;
mul.lo.s64 %rd60, %rd59, 36;
add.s64 %rd61, %rd7, %rd60;
add.s64 %rd62, %rd61, %rd35;
ld.global.nc.u32 %r253, [%rd62+4];

dp4a.s32.s32 %r251, %r320, %r253, %r339;

ld.global.nc.u32 %r257, [%rd62+20];

dp4a.s32.s32 %r255, %r324, %r257, %r251;

ld.global.nc.u32 %r261, [%rd62+8];

dp4a.s32.s32 %r259, %r328, %r261, %r255;

ld.global.nc.u32 %r265, [%rd62+24];

dp4a.s32.s32 %r263, %r332, %r265, %r259;


{ cvt.f32.f16 %f117, %rs1;}


ld.global.nc.u32 %r267, [%rd61];

{.reg .f16 low,high;
mov.b32 {low,high},%r267;
cvt.f32.f16 %f118, low;}


mul.ftz.f32 %f149, %f117, %f118;
cvt.rn.f32.s32 %f150, %r263;
fma.rn.ftz.f32 %f358, %f149, %f150, %f358;

dp4a.s32.s32 %r268, %r337, %r253, %r339;


dp4a.s32.s32 %r272, %r341, %r257, %r268;


dp4a.s32.s32 %r276, %r345, %r261, %r272;


dp4a.s32.s32 %r280, %r349, %r265, %r276;


{ cvt.f32.f16 %f119, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r267;
cvt.f32.f16 %f120, low;}


mul.ftz.f32 %f151, %f119, %f120;
cvt.rn.f32.s32 %f152, %r280;
fma.rn.ftz.f32 %f357, %f151, %f152, %f357;
cvt.u64.u32 %rd63, %r715;
add.s64 %rd64, %rd1, %rd63;
mul.lo.s64 %rd65, %rd64, 36;
add.s64 %rd66, %rd7, %rd65;
add.s64 %rd67, %rd66, %rd35;
ld.global.nc.u32 %r287, [%rd67+4];

dp4a.s32.s32 %r285, %r320, %r287, %r339;

ld.global.nc.u32 %r291, [%rd67+20];

dp4a.s32.s32 %r289, %r324, %r291, %r285;

ld.global.nc.u32 %r295, [%rd67+8];

dp4a.s32.s32 %r293, %r328, %r295, %r289;

ld.global.nc.u32 %r299, [%rd67+24];

dp4a.s32.s32 %r297, %r332, %r299, %r293;


{ cvt.f32.f16 %f121, %rs1;}


ld.global.nc.u32 %r301, [%rd66];

{.reg .f16 low,high;
mov.b32 {low,high},%r301;
cvt.f32.f16 %f122, low;}


mul.ftz.f32 %f153, %f121, %f122;
cvt.rn.f32.s32 %f154, %r297;
fma.rn.ftz.f32 %f356, %f153, %f154, %f356;

dp4a.s32.s32 %r302, %r337, %r287, %r339;


dp4a.s32.s32 %r306, %r341, %r291, %r302;


dp4a.s32.s32 %r310, %r345, %r295, %r306;


dp4a.s32.s32 %r314, %r349, %r299, %r310;


{ cvt.f32.f16 %f123, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r301;
cvt.f32.f16 %f124, low;}


mul.ftz.f32 %f155, %f123, %f124;
cvt.rn.f32.s32 %f156, %r314;
fma.rn.ftz.f32 %f355, %f155, %f156, %f355;
cvt.u64.u32 %rd68, %r714;
add.s64 %rd69, %rd1, %rd68;
mul.lo.s64 %rd70, %rd69, 36;
add.s64 %rd71, %rd7, %rd70;
add.s64 %rd72, %rd71, %rd35;
ld.global.nc.u32 %r321, [%rd72+4];

dp4a.s32.s32 %r319, %r320, %r321, %r339;

ld.global.nc.u32 %r325, [%rd72+20];

dp4a.s32.s32 %r323, %r324, %r325, %r319;

ld.global.nc.u32 %r329, [%rd72+8];

dp4a.s32.s32 %r327, %r328, %r329, %r323;

ld.global.nc.u32 %r333, [%rd72+24];

dp4a.s32.s32 %r331, %r332, %r333, %r327;


{ cvt.f32.f16 %f125, %rs1;}


ld.global.nc.u32 %r335, [%rd71];

{.reg .f16 low,high;
mov.b32 {low,high},%r335;
cvt.f32.f16 %f126, low;}


mul.ftz.f32 %f157, %f125, %f126;
cvt.rn.f32.s32 %f158, %r331;
fma.rn.ftz.f32 %f354, %f157, %f158, %f354;

dp4a.s32.s32 %r336, %r337, %r321, %r339;


dp4a.s32.s32 %r340, %r341, %r325, %r336;


dp4a.s32.s32 %r344, %r345, %r329, %r340;


dp4a.s32.s32 %r348, %r349, %r333, %r344;


{ cvt.f32.f16 %f127, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r335;
cvt.f32.f16 %f128, low;}


mul.ftz.f32 %f159, %f127, %f128;
cvt.rn.f32.s32 %f160, %r348;
fma.rn.ftz.f32 %f353, %f159, %f160, %f353;
add.s32 %r719, %r719, 32;
add.s32 %r718, %r718, 32;
add.s32 %r717, %r717, 32;
add.s32 %r716, %r716, 32;
add.s32 %r715, %r715, 32;
add.s32 %r714, %r714, 32;
add.s32 %r713, %r713, 32;
add.s64 %rd129, %rd129, 576;
add.s64 %rd128, %rd128, 576;
add.s64 %rd127, %rd127, 576;
add.s32 %r720, %r720, 32;
setp.lt.s32 %p4, %r720, %r6;
@%p4 bra $L__BB290_2;

$L__BB290_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB290_5;

shl.b32 %r419, %r1, 11;
mov.u32 %r420, _ZZ13mul_mat_vec_qIL9ggml_type20ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r421, %r420, %r419;
shl.b32 %r422, %r2, 2;
add.s32 %r423, %r421, %r422;
st.shared.f32 [%r423+-2048], %f368;
st.shared.f32 [%r423+-1920], %f367;
st.shared.f32 [%r423+-1792], %f366;
st.shared.f32 [%r423+-1664], %f365;
st.shared.f32 [%r423+-1536], %f364;
st.shared.f32 [%r423+-1408], %f363;
st.shared.f32 [%r423+-1280], %f362;
st.shared.f32 [%r423+-1152], %f361;
st.shared.f32 [%r423+-1024], %f360;
st.shared.f32 [%r423+-896], %f359;
st.shared.f32 [%r423+-768], %f358;
st.shared.f32 [%r423+-640], %f357;
st.shared.f32 [%r423+-512], %f356;
st.shared.f32 [%r423+-384], %f355;
st.shared.f32 [%r423+-256], %f354;
st.shared.f32 [%r423+-128], %f353;

$L__BB290_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB290_38;

mov.u32 %r424, %ctaid.z;
mov.u32 %r425, %ctaid.y;
mad.lo.s32 %r426, %r425, %r42, %r4;
mad.lo.s32 %r427, %r424, %r48, %r426;
cvt.u64.u32 %rd14, %r427;
shl.b32 %r428, %r2, 2;
mov.u32 %r429, _ZZ13mul_mat_vec_qIL9ggml_type20ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r30, %r429, %r428;
mov.u32 %r430, 2;
ld.shared.f32 %f161, [%r30];
add.ftz.f32 %f162, %f161, %f368;
mov.b32 %r431, %f162;
mov.u32 %r432, 31;
mov.u32 %r433, 16;
mov.u32 %r434, -1;
shfl.sync.bfly.b32 %r435|%p7, %r431, %r433, %r432, %r434;
mov.b32 %f163, %r435;
add.ftz.f32 %f164, %f162, %f163;
mov.b32 %r436, %f164;
mov.u32 %r437, 8;
shfl.sync.bfly.b32 %r438|%p8, %r436, %r437, %r432, %r434;
mov.b32 %f165, %r438;
add.ftz.f32 %f166, %f164, %f165;
mov.b32 %r439, %f166;
mov.u32 %r440, 4;
shfl.sync.bfly.b32 %r441|%p9, %r439, %r440, %r432, %r434;
mov.b32 %f167, %r441;
add.ftz.f32 %f168, %f166, %f167;
mov.b32 %r442, %f168;
shfl.sync.bfly.b32 %r443|%p10, %r442, %r430, %r432, %r434;
mov.b32 %f169, %r443;
add.ftz.f32 %f170, %f168, %f169;
mov.b32 %r444, %f170;
mov.u32 %r445, 1;
shfl.sync.bfly.b32 %r446|%p11, %r444, %r445, %r432, %r434;
mov.b32 %f171, %r446;
add.ftz.f32 %f49, %f170, %f171;
setp.lt.u32 %p12, %r4, %r36;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
cvta.to.global.u64 %rd73, %rd21;
mul.wide.u32 %rd74, %r427, 4;
add.s64 %rd15, %rd73, %rd74;
not.pred %p14, %p1;
@%p14 bra $L__BB290_8;

st.global.f32 [%rd15], %f49;

$L__BB290_8:
ld.shared.f32 %f172, [%r30+128];
add.ftz.f32 %f173, %f172, %f367;
mov.b32 %r447, %f173;
shfl.sync.bfly.b32 %r451|%p15, %r447, %r433, %r432, %r434;
mov.b32 %f174, %r451;
add.ftz.f32 %f175, %f173, %f174;
mov.b32 %r452, %f175;
shfl.sync.bfly.b32 %r454|%p16, %r452, %r437, %r432, %r434;
mov.b32 %f176, %r454;
add.ftz.f32 %f177, %f175, %f176;
mov.b32 %r455, %f177;
shfl.sync.bfly.b32 %r457|%p17, %r455, %r440, %r432, %r434;
mov.b32 %f178, %r457;
add.ftz.f32 %f179, %f177, %f178;
mov.b32 %r458, %f179;
shfl.sync.bfly.b32 %r460|%p18, %r458, %r430, %r432, %r434;
mov.b32 %f180, %r460;
add.ftz.f32 %f181, %f179, %f180;
mov.b32 %r461, %f181;
shfl.sync.bfly.b32 %r463|%p19, %r461, %r445, %r432, %r434;
mov.b32 %f182, %r463;
add.ftz.f32 %f50, %f181, %f182;
setp.eq.s32 %p20, %r2, 1;
or.b32 %r467, %r4, 1;
setp.lt.u32 %p21, %r467, %r36;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB290_10;

st.global.f32 [%rd15+4], %f50;

$L__BB290_10:
ld.shared.f32 %f183, [%r30+256];
add.ftz.f32 %f184, %f183, %f366;
mov.b32 %r468, %f184;
mov.u32 %r469, 31;
mov.u32 %r470, 16;
mov.u32 %r471, -1;
shfl.sync.bfly.b32 %r472|%p23, %r468, %r470, %r469, %r471;
mov.b32 %f185, %r472;
add.ftz.f32 %f186, %f184, %f185;
mov.b32 %r473, %f186;
mov.u32 %r474, 8;
shfl.sync.bfly.b32 %r475|%p24, %r473, %r474, %r469, %r471;
mov.b32 %f187, %r475;
add.ftz.f32 %f188, %f186, %f187;
mov.b32 %r476, %f188;
mov.u32 %r477, 4;
shfl.sync.bfly.b32 %r478|%p25, %r476, %r477, %r469, %r471;
mov.b32 %f189, %r478;
add.ftz.f32 %f190, %f188, %f189;
mov.b32 %r479, %f190;
mov.u32 %r480, 2;
shfl.sync.bfly.b32 %r481|%p26, %r479, %r480, %r469, %r471;
mov.b32 %f191, %r481;
add.ftz.f32 %f192, %f190, %f191;
mov.b32 %r482, %f192;
mov.u32 %r483, 1;
shfl.sync.bfly.b32 %r484|%p27, %r482, %r483, %r469, %r471;
mov.b32 %f193, %r484;
add.ftz.f32 %f51, %f192, %f193;
@%p14 bra $L__BB290_12;

cvt.u64.u32 %rd75, %r36;
add.s64 %rd76, %rd75, %rd14;
shl.b64 %rd78, %rd76, 2;
add.s64 %rd79, %rd73, %rd78;
st.global.f32 [%rd79], %f51;

$L__BB290_12:
ld.shared.f32 %f194, [%r30+384];
add.ftz.f32 %f195, %f194, %f365;
mov.b32 %r485, %f195;
shfl.sync.bfly.b32 %r489|%p29, %r485, %r470, %r469, %r471;
mov.b32 %f196, %r489;
add.ftz.f32 %f197, %f195, %f196;
mov.b32 %r490, %f197;
shfl.sync.bfly.b32 %r492|%p30, %r490, %r474, %r469, %r471;
mov.b32 %f198, %r492;
add.ftz.f32 %f199, %f197, %f198;
mov.b32 %r493, %f199;
shfl.sync.bfly.b32 %r495|%p31, %r493, %r477, %r469, %r471;
mov.b32 %f200, %r495;
add.ftz.f32 %f201, %f199, %f200;
mov.b32 %r496, %f201;
shfl.sync.bfly.b32 %r498|%p32, %r496, %r480, %r469, %r471;
mov.b32 %f202, %r498;
add.ftz.f32 %f203, %f201, %f202;
mov.b32 %r499, %f203;
shfl.sync.bfly.b32 %r501|%p33, %r499, %r483, %r469, %r471;
mov.b32 %f204, %r501;
add.ftz.f32 %f52, %f203, %f204;
@%p22 bra $L__BB290_14;

add.s32 %r502, %r36, 1;
cvt.u64.u32 %rd80, %r502;
add.s64 %rd81, %rd80, %rd14;
shl.b64 %rd83, %rd81, 2;
add.s64 %rd84, %rd73, %rd83;
st.global.f32 [%rd84], %f52;

$L__BB290_14:
shl.b32 %r503, %r36, 1;
mov.u32 %r504, 1;
ld.shared.f32 %f205, [%r30+512];
add.ftz.f32 %f206, %f205, %f364;
mov.b32 %r505, %f206;
mov.u32 %r506, 31;
mov.u32 %r507, 16;
mov.u32 %r508, -1;
shfl.sync.bfly.b32 %r509|%p35, %r505, %r507, %r506, %r508;
mov.b32 %f207, %r509;
add.ftz.f32 %f208, %f206, %f207;
mov.b32 %r510, %f208;
mov.u32 %r511, 8;
shfl.sync.bfly.b32 %r512|%p36, %r510, %r511, %r506, %r508;
mov.b32 %f209, %r512;
add.ftz.f32 %f210, %f208, %f209;
mov.b32 %r513, %f210;
mov.u32 %r514, 4;
shfl.sync.bfly.b32 %r515|%p37, %r513, %r514, %r506, %r508;
mov.b32 %f211, %r515;
add.ftz.f32 %f212, %f210, %f211;
mov.b32 %r516, %f212;
mov.u32 %r517, 2;
shfl.sync.bfly.b32 %r518|%p38, %r516, %r517, %r506, %r508;
mov.b32 %f213, %r518;
add.ftz.f32 %f214, %f212, %f213;
mov.b32 %r519, %f214;
shfl.sync.bfly.b32 %r520|%p39, %r519, %r504, %r506, %r508;
mov.b32 %f215, %r520;
add.ftz.f32 %f53, %f214, %f215;
cvt.u64.u32 %rd85, %r503;
add.s64 %rd86, %rd85, %rd14;
shl.b64 %rd88, %rd86, 2;
add.s64 %rd16, %rd73, %rd88;
@%p14 bra $L__BB290_16;

st.global.f32 [%rd16], %f53;

$L__BB290_16:
ld.shared.f32 %f216, [%r30+640];
add.ftz.f32 %f217, %f216, %f363;
mov.b32 %r521, %f217;
shfl.sync.bfly.b32 %r525|%p41, %r521, %r507, %r506, %r508;
mov.b32 %f218, %r525;
add.ftz.f32 %f219, %f217, %f218;
mov.b32 %r526, %f219;
shfl.sync.bfly.b32 %r528|%p42, %r526, %r511, %r506, %r508;
mov.b32 %f220, %r528;
add.ftz.f32 %f221, %f219, %f220;
mov.b32 %r529, %f221;
shfl.sync.bfly.b32 %r531|%p43, %r529, %r514, %r506, %r508;
mov.b32 %f222, %r531;
add.ftz.f32 %f223, %f221, %f222;
mov.b32 %r532, %f223;
shfl.sync.bfly.b32 %r534|%p44, %r532, %r517, %r506, %r508;
mov.b32 %f224, %r534;
add.ftz.f32 %f225, %f223, %f224;
mov.b32 %r535, %f225;
shfl.sync.bfly.b32 %r537|%p45, %r535, %r504, %r506, %r508;
mov.b32 %f226, %r537;
add.ftz.f32 %f54, %f225, %f226;
@%p22 bra $L__BB290_18;

st.global.f32 [%rd16+4], %f54;

$L__BB290_18:
mul.lo.s32 %r31, %r36, 3;
ld.shared.f32 %f227, [%r30+768];
add.ftz.f32 %f228, %f227, %f362;
mov.b32 %r538, %f228;
mov.u32 %r539, 31;
mov.u32 %r540, 16;
mov.u32 %r541, -1;
shfl.sync.bfly.b32 %r542|%p47, %r538, %r540, %r539, %r541;
mov.b32 %f229, %r542;
add.ftz.f32 %f230, %f228, %f229;
mov.b32 %r543, %f230;
mov.u32 %r544, 8;
shfl.sync.bfly.b32 %r545|%p48, %r543, %r544, %r539, %r541;
mov.b32 %f231, %r545;
add.ftz.f32 %f232, %f230, %f231;
mov.b32 %r546, %f232;
mov.u32 %r547, 4;
shfl.sync.bfly.b32 %r548|%p49, %r546, %r547, %r539, %r541;
mov.b32 %f233, %r548;
add.ftz.f32 %f234, %f232, %f233;
mov.b32 %r549, %f234;
mov.u32 %r550, 2;
shfl.sync.bfly.b32 %r551|%p50, %r549, %r550, %r539, %r541;
mov.b32 %f235, %r551;
add.ftz.f32 %f236, %f234, %f235;
mov.b32 %r552, %f236;
mov.u32 %r553, 1;
shfl.sync.bfly.b32 %r554|%p51, %r552, %r553, %r539, %r541;
mov.b32 %f237, %r554;
add.ftz.f32 %f55, %f236, %f237;
@%p14 bra $L__BB290_20;

cvt.u64.u32 %rd89, %r31;
add.s64 %rd90, %rd89, %rd14;
shl.b64 %rd92, %rd90, 2;
add.s64 %rd93, %rd73, %rd92;
st.global.f32 [%rd93], %f55;

$L__BB290_20:
ld.shared.f32 %f238, [%r30+896];
add.ftz.f32 %f239, %f238, %f361;
mov.b32 %r555, %f239;
shfl.sync.bfly.b32 %r559|%p53, %r555, %r540, %r539, %r541;
mov.b32 %f240, %r559;
add.ftz.f32 %f241, %f239, %f240;
mov.b32 %r560, %f241;
shfl.sync.bfly.b32 %r562|%p54, %r560, %r544, %r539, %r541;
mov.b32 %f242, %r562;
add.ftz.f32 %f243, %f241, %f242;
mov.b32 %r563, %f243;
shfl.sync.bfly.b32 %r565|%p55, %r563, %r547, %r539, %r541;
mov.b32 %f244, %r565;
add.ftz.f32 %f245, %f243, %f244;
mov.b32 %r566, %f245;
shfl.sync.bfly.b32 %r568|%p56, %r566, %r550, %r539, %r541;
mov.b32 %f246, %r568;
add.ftz.f32 %f247, %f245, %f246;
mov.b32 %r569, %f247;
shfl.sync.bfly.b32 %r571|%p57, %r569, %r553, %r539, %r541;
mov.b32 %f248, %r571;
add.ftz.f32 %f56, %f247, %f248;
@%p22 bra $L__BB290_22;

add.s32 %r572, %r31, 1;
cvt.u64.u32 %rd94, %r572;
add.s64 %rd95, %rd94, %rd14;
shl.b64 %rd97, %rd95, 2;
add.s64 %rd98, %rd73, %rd97;
st.global.f32 [%rd98], %f56;

$L__BB290_22:
shl.b32 %r573, %r36, 2;
mov.u32 %r574, 2;
ld.shared.f32 %f249, [%r30+1024];
add.ftz.f32 %f250, %f249, %f360;
mov.b32 %r575, %f250;
mov.u32 %r576, 31;
mov.u32 %r577, 16;
mov.u32 %r578, -1;
shfl.sync.bfly.b32 %r579|%p59, %r575, %r577, %r576, %r578;
mov.b32 %f251, %r579;
add.ftz.f32 %f252, %f250, %f251;
mov.b32 %r580, %f252;
mov.u32 %r581, 8;
shfl.sync.bfly.b32 %r582|%p60, %r580, %r581, %r576, %r578;
mov.b32 %f253, %r582;
add.ftz.f32 %f254, %f252, %f253;
mov.b32 %r583, %f254;
mov.u32 %r584, 4;
shfl.sync.bfly.b32 %r585|%p61, %r583, %r584, %r576, %r578;
mov.b32 %f255, %r585;
add.ftz.f32 %f256, %f254, %f255;
mov.b32 %r586, %f256;
shfl.sync.bfly.b32 %r587|%p62, %r586, %r574, %r576, %r578;
mov.b32 %f257, %r587;
add.ftz.f32 %f258, %f256, %f257;
mov.b32 %r588, %f258;
mov.u32 %r589, 1;
shfl.sync.bfly.b32 %r590|%p63, %r588, %r589, %r576, %r578;
mov.b32 %f259, %r590;
add.ftz.f32 %f57, %f258, %f259;
cvt.u64.u32 %rd99, %r573;
add.s64 %rd100, %rd99, %rd14;
shl.b64 %rd102, %rd100, 2;
add.s64 %rd17, %rd73, %rd102;
@%p14 bra $L__BB290_24;

st.global.f32 [%rd17], %f57;

$L__BB290_24:
ld.shared.f32 %f260, [%r30+1152];
add.ftz.f32 %f261, %f260, %f359;
mov.b32 %r591, %f261;
shfl.sync.bfly.b32 %r595|%p65, %r591, %r577, %r576, %r578;
mov.b32 %f262, %r595;
add.ftz.f32 %f263, %f261, %f262;
mov.b32 %r596, %f263;
shfl.sync.bfly.b32 %r598|%p66, %r596, %r581, %r576, %r578;
mov.b32 %f264, %r598;
add.ftz.f32 %f265, %f263, %f264;
mov.b32 %r599, %f265;
shfl.sync.bfly.b32 %r601|%p67, %r599, %r584, %r576, %r578;
mov.b32 %f266, %r601;
add.ftz.f32 %f267, %f265, %f266;
mov.b32 %r602, %f267;
shfl.sync.bfly.b32 %r604|%p68, %r602, %r574, %r576, %r578;
mov.b32 %f268, %r604;
add.ftz.f32 %f269, %f267, %f268;
mov.b32 %r605, %f269;
shfl.sync.bfly.b32 %r607|%p69, %r605, %r589, %r576, %r578;
mov.b32 %f270, %r607;
add.ftz.f32 %f58, %f269, %f270;
@%p22 bra $L__BB290_26;

st.global.f32 [%rd17+4], %f58;

$L__BB290_26:
mul.lo.s32 %r32, %r36, 5;
ld.shared.f32 %f271, [%r30+1280];
add.ftz.f32 %f272, %f271, %f358;
mov.b32 %r608, %f272;
mov.u32 %r609, 31;
mov.u32 %r610, 16;
mov.u32 %r611, -1;
shfl.sync.bfly.b32 %r612|%p71, %r608, %r610, %r609, %r611;
mov.b32 %f273, %r612;
add.ftz.f32 %f274, %f272, %f273;
mov.b32 %r613, %f274;
mov.u32 %r614, 8;
shfl.sync.bfly.b32 %r615|%p72, %r613, %r614, %r609, %r611;
mov.b32 %f275, %r615;
add.ftz.f32 %f276, %f274, %f275;
mov.b32 %r616, %f276;
mov.u32 %r617, 4;
shfl.sync.bfly.b32 %r618|%p73, %r616, %r617, %r609, %r611;
mov.b32 %f277, %r618;
add.ftz.f32 %f278, %f276, %f277;
mov.b32 %r619, %f278;
mov.u32 %r620, 2;
shfl.sync.bfly.b32 %r621|%p74, %r619, %r620, %r609, %r611;
mov.b32 %f279, %r621;
add.ftz.f32 %f280, %f278, %f279;
mov.b32 %r622, %f280;
mov.u32 %r623, 1;
shfl.sync.bfly.b32 %r624|%p75, %r622, %r623, %r609, %r611;
mov.b32 %f281, %r624;
add.ftz.f32 %f59, %f280, %f281;
@%p14 bra $L__BB290_28;

cvt.u64.u32 %rd103, %r32;
add.s64 %rd104, %rd103, %rd14;
shl.b64 %rd106, %rd104, 2;
add.s64 %rd107, %rd73, %rd106;
st.global.f32 [%rd107], %f59;

$L__BB290_28:
ld.shared.f32 %f282, [%r30+1408];
add.ftz.f32 %f283, %f282, %f357;
mov.b32 %r625, %f283;
shfl.sync.bfly.b32 %r629|%p77, %r625, %r610, %r609, %r611;
mov.b32 %f284, %r629;
add.ftz.f32 %f285, %f283, %f284;
mov.b32 %r630, %f285;
shfl.sync.bfly.b32 %r632|%p78, %r630, %r614, %r609, %r611;
mov.b32 %f286, %r632;
add.ftz.f32 %f287, %f285, %f286;
mov.b32 %r633, %f287;
shfl.sync.bfly.b32 %r635|%p79, %r633, %r617, %r609, %r611;
mov.b32 %f288, %r635;
add.ftz.f32 %f289, %f287, %f288;
mov.b32 %r636, %f289;
shfl.sync.bfly.b32 %r638|%p80, %r636, %r620, %r609, %r611;
mov.b32 %f290, %r638;
add.ftz.f32 %f291, %f289, %f290;
mov.b32 %r639, %f291;
shfl.sync.bfly.b32 %r641|%p81, %r639, %r623, %r609, %r611;
mov.b32 %f292, %r641;
add.ftz.f32 %f60, %f291, %f292;
@%p22 bra $L__BB290_30;

add.s32 %r642, %r32, 1;
cvt.u64.u32 %rd108, %r642;
add.s64 %rd109, %rd108, %rd14;
shl.b64 %rd111, %rd109, 2;
add.s64 %rd112, %rd73, %rd111;
st.global.f32 [%rd112], %f60;

$L__BB290_30:
mul.lo.s32 %r643, %r36, 6;
ld.shared.f32 %f293, [%r30+1536];
add.ftz.f32 %f294, %f293, %f356;
mov.b32 %r644, %f294;
mov.u32 %r645, 31;
mov.u32 %r646, 16;
mov.u32 %r647, -1;
shfl.sync.bfly.b32 %r648|%p83, %r644, %r646, %r645, %r647;
mov.b32 %f295, %r648;
add.ftz.f32 %f296, %f294, %f295;
mov.b32 %r649, %f296;
mov.u32 %r650, 8;
shfl.sync.bfly.b32 %r651|%p84, %r649, %r650, %r645, %r647;
mov.b32 %f297, %r651;
add.ftz.f32 %f298, %f296, %f297;
mov.b32 %r652, %f298;
mov.u32 %r653, 4;
shfl.sync.bfly.b32 %r654|%p85, %r652, %r653, %r645, %r647;
mov.b32 %f299, %r654;
add.ftz.f32 %f300, %f298, %f299;
mov.b32 %r655, %f300;
mov.u32 %r656, 2;
shfl.sync.bfly.b32 %r657|%p86, %r655, %r656, %r645, %r647;
mov.b32 %f301, %r657;
add.ftz.f32 %f302, %f300, %f301;
mov.b32 %r658, %f302;
mov.u32 %r659, 1;
shfl.sync.bfly.b32 %r660|%p87, %r658, %r659, %r645, %r647;
mov.b32 %f303, %r660;
add.ftz.f32 %f61, %f302, %f303;
cvt.u64.u32 %rd113, %r643;
add.s64 %rd114, %rd113, %rd14;
shl.b64 %rd116, %rd114, 2;
add.s64 %rd18, %rd73, %rd116;
@%p14 bra $L__BB290_32;

st.global.f32 [%rd18], %f61;

$L__BB290_32:
ld.shared.f32 %f304, [%r30+1664];
add.ftz.f32 %f305, %f304, %f355;
mov.b32 %r661, %f305;
shfl.sync.bfly.b32 %r665|%p89, %r661, %r646, %r645, %r647;
mov.b32 %f306, %r665;
add.ftz.f32 %f307, %f305, %f306;
mov.b32 %r666, %f307;
shfl.sync.bfly.b32 %r668|%p90, %r666, %r650, %r645, %r647;
mov.b32 %f308, %r668;
add.ftz.f32 %f309, %f307, %f308;
mov.b32 %r669, %f309;
shfl.sync.bfly.b32 %r671|%p91, %r669, %r653, %r645, %r647;
mov.b32 %f310, %r671;
add.ftz.f32 %f311, %f309, %f310;
mov.b32 %r672, %f311;
shfl.sync.bfly.b32 %r674|%p92, %r672, %r656, %r645, %r647;
mov.b32 %f312, %r674;
add.ftz.f32 %f313, %f311, %f312;
mov.b32 %r675, %f313;
shfl.sync.bfly.b32 %r677|%p93, %r675, %r659, %r645, %r647;
mov.b32 %f314, %r677;
add.ftz.f32 %f62, %f313, %f314;
@%p22 bra $L__BB290_34;

st.global.f32 [%rd18+4], %f62;

$L__BB290_34:
mul.lo.s32 %r33, %r36, 7;
ld.shared.f32 %f315, [%r30+1792];
add.ftz.f32 %f316, %f315, %f354;
mov.b32 %r678, %f316;
mov.u32 %r679, 31;
mov.u32 %r680, 16;
mov.u32 %r681, -1;
shfl.sync.bfly.b32 %r682|%p95, %r678, %r680, %r679, %r681;
mov.b32 %f317, %r682;
add.ftz.f32 %f318, %f316, %f317;
mov.b32 %r683, %f318;
mov.u32 %r684, 8;
shfl.sync.bfly.b32 %r685|%p96, %r683, %r684, %r679, %r681;
mov.b32 %f319, %r685;
add.ftz.f32 %f320, %f318, %f319;
mov.b32 %r686, %f320;
mov.u32 %r687, 4;
shfl.sync.bfly.b32 %r688|%p97, %r686, %r687, %r679, %r681;
mov.b32 %f321, %r688;
add.ftz.f32 %f322, %f320, %f321;
mov.b32 %r689, %f322;
mov.u32 %r690, 2;
shfl.sync.bfly.b32 %r691|%p98, %r689, %r690, %r679, %r681;
mov.b32 %f323, %r691;
add.ftz.f32 %f324, %f322, %f323;
mov.b32 %r692, %f324;
mov.u32 %r693, 1;
shfl.sync.bfly.b32 %r694|%p99, %r692, %r693, %r679, %r681;
mov.b32 %f325, %r694;
add.ftz.f32 %f63, %f324, %f325;
@%p14 bra $L__BB290_36;

cvt.u64.u32 %rd117, %r33;
add.s64 %rd118, %rd117, %rd14;
shl.b64 %rd120, %rd118, 2;
add.s64 %rd121, %rd73, %rd120;
st.global.f32 [%rd121], %f63;

$L__BB290_36:
ld.shared.f32 %f326, [%r30+1920];
add.ftz.f32 %f327, %f326, %f353;
mov.b32 %r695, %f327;
shfl.sync.bfly.b32 %r699|%p101, %r695, %r680, %r679, %r681;
mov.b32 %f328, %r699;
add.ftz.f32 %f329, %f327, %f328;
mov.b32 %r700, %f329;
shfl.sync.bfly.b32 %r702|%p102, %r700, %r684, %r679, %r681;
mov.b32 %f330, %r702;
add.ftz.f32 %f331, %f329, %f330;
mov.b32 %r703, %f331;
shfl.sync.bfly.b32 %r705|%p103, %r703, %r687, %r679, %r681;
mov.b32 %f332, %r705;
add.ftz.f32 %f333, %f331, %f332;
mov.b32 %r706, %f333;
shfl.sync.bfly.b32 %r708|%p104, %r706, %r690, %r679, %r681;
mov.b32 %f334, %r708;
add.ftz.f32 %f335, %f333, %f334;
mov.b32 %r709, %f335;
shfl.sync.bfly.b32 %r711|%p105, %r709, %r693, %r679, %r681;
mov.b32 %f336, %r711;
add.ftz.f32 %f64, %f335, %f336;
@%p22 bra $L__BB290_38;

add.s32 %r712, %r33, 1;
cvt.u64.u32 %rd122, %r712;
add.s64 %rd123, %rd122, %rd14;
shl.b64 %rd125, %rd123, 2;
add.s64 %rd126, %rd73, %rd125;
st.global.f32 [%rd126], %f64;

$L__BB290_38:
ret;

}
.entry _Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj(
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2,
.param .align 8 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3[48],
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5,
.param .align 4 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6[12],
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15
)
.maxntid 256, 1, 1
.minnctapersm 1
{
.local .align 8 .b8 __local_depot291[16];
.reg .b64 %SP;
.reg .b64 %SPL;
.reg .pred %p<48>;
.reg .b16 %rs<9>;
.reg .f32 %f<169>;
.reg .b32 %r<588>;
.reg .b64 %rd<100>;


mov.u64 %SPL, __local_depot291;
ld.param.u64 %rd24, [_Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0];
ld.param.u64 %rd25, [_Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1];
ld.param.u64 %rd26, [_Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2];
ld.param.u32 %r35, [_Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5];
ld.param.u32 %r28, [_Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8];
ld.param.u32 %r29, [_Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9];
ld.param.u32 %r31, [_Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11];
ld.param.u32 %r32, [_Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12];
ld.param.u32 %r36, [_Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14];
ld.param.u32 %r34, [_Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15];
ld.param.u32 %r26, [_Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+8];
ld.param.u32 %r25, [_Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+4];
ld.param.u32 %r24, [_Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6];
ld.param.u64 %rd28, [_Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+8];
add.u64 %rd1, %SPL, 0;
add.u64 %rd2, %SPL, 8;
shr.u32 %r1, %r35, 8;
mov.u32 %r2, %tid.y;
setp.ge.u32 %p1, %r2, %r36;
@%p1 bra $L__BB291_29;

mov.u32 %r3, %ctaid.y;
cvta.to.global.u64 %rd35, %rd26;
mad.lo.s32 %r37, %r2, %r34, %r3;
mul.wide.u32 %rd36, %r37, 4;
add.s64 %rd37, %rd35, %rd36;
ld.global.nc.u32 %r4, [%rd37];
mov.f32 %f162, 0f00000000;
st.local.v2.f32 [%rd1], {%f162, %f162};
st.local.v2.f32 [%rd2], {%f162, %f162};
mov.u32 %r5, %tid.x;
shr.u32 %r587, %r5, 3;
setp.ge.u32 %p2, %r587, %r1;
mov.f32 %f163, %f162;
mov.f32 %f164, %f162;
mov.f32 %f165, %f162;
@%p2 bra $L__BB291_9;

mov.u32 %r38, %ctaid.x;
shl.b32 %r39, %r38, 1;
mul.lo.s32 %r40, %r4, %r31;
and.b32 %r41, %r5, 7;
shl.b32 %r42, %r5, 2;
shr.u32 %r43, %r41, 1;
cvt.u64.u32 %rd38, %r43;
and.b32 %r7, %r42, 4;
shl.b32 %r44, %r5, 1;
and.b32 %r8, %r44, 14;
mad.lo.s32 %r45, %r39, %r28, %r587;
add.s32 %r46, %r45, %r40;
mul.wide.s32 %rd6, %r46, 136;
or.b64 %rd39, %rd38, %rd6;
or.b64 %rd7, %rd39, 4;
mul.hi.u32 %r47, %r3, %r24;
add.s32 %r48, %r47, %r3;
shr.u32 %r49, %r48, %r25;
mul.lo.s32 %r50, %r49, %r26;
sub.s32 %r51, %r3, %r50;
mul.lo.s32 %r52, %r51, %r32;
mul.wide.u32 %rd40, %r52, 36;
mul.lo.s32 %r53, %r2, %r29;
mul.wide.u32 %rd41, %r53, 36;
add.s64 %rd42, %rd41, %rd40;
mul.wide.u32 %rd43, %r41, 36;
add.s64 %rd44, %rd42, %rd43;
shl.b32 %r54, %r587, 3;
mul.wide.s32 %rd45, %r54, 36;
add.s64 %rd46, %rd44, %rd45;
cvta.to.global.u64 %rd47, %rd25;
add.s64 %rd48, %rd47, %rd46;
add.s64 %rd98, %rd48, 16;
shl.b32 %r55, %r5, 4;
cvt.u64.u32 %rd49, %r55;
and.b64 %rd50, %rd49, 112;
add.s64 %rd51, %rd50, %rd6;
add.s64 %rd9, %rd51, 20;
or.b32 %r56, %r39, 1;
add.s32 %r57, %r587, %r40;
mad.lo.s32 %r58, %r28, %r56, %r57;
mul.wide.s32 %rd10, %r58, 136;
add.s64 %rd11, %rd50, %rd10;
or.b64 %rd52, %rd38, %rd10;
or.b64 %rd12, %rd52, 4;
cvta.to.global.u64 %rd99, %rd24;
cvta.to.global.u64 %rd97, %rd28;
mov.f32 %f162, 0f00000000;

$L__BB291_3:
add.s64 %rd53, %rd99, %rd9;
ld.global.nc.u32 %r92, [%rd53+-12];
and.b32 %r93, %r92, -2004318072;
shr.u32 %r94, %r93, 1;
or.b32 %r95, %r94, 839922192;
and.b32 %r96, %r92, 30583;
mov.u32 %r97, -152379953;
mov.u32 %r98, -1079142271;
prmt.b32 %r99, %r98, %r97, %r96;
mov.u32 %r100, 1901675829;
mov.u32 %r101, 639175937;
prmt.b32 %r102, %r101, %r100, %r96;
and.b32 %r103, %r95, 30292;
prmt.b32 %r104, %r99, %r102, %r103;
shr.u32 %r105, %r92, 16;
and.b32 %r106, %r105, 30583;
prmt.b32 %r107, %r98, %r97, %r106;
prmt.b32 %r108, %r101, %r100, %r106;
shr.u32 %r109, %r95, 16;
prmt.b32 %r110, %r107, %r108, %r109;
mov.u32 %r111, 25632;
prmt.b32 %r60, %r104, %r110, %r111;
mov.u32 %r112, 30001;
prmt.b32 %r64, %r104, %r110, %r112;
ld.global.nc.u32 %r10, [%rd98+-12];
ld.global.nc.u32 %r11, [%rd98+4];
mov.u32 %r62, 0;

dp4a.s32.s32 %r59, %r60, %r10, %r62;


dp4a.s32.s32 %r63, %r64, %r11, %r59;

ld.global.nc.u32 %r113, [%rd53+-8];
and.b32 %r114, %r113, -2004318072;
shr.u32 %r115, %r114, 1;
or.b32 %r116, %r115, 839922192;
and.b32 %r117, %r113, 30583;
prmt.b32 %r118, %r98, %r97, %r117;
prmt.b32 %r119, %r101, %r100, %r117;
and.b32 %r120, %r116, 30292;
prmt.b32 %r121, %r118, %r119, %r120;
shr.u32 %r122, %r113, 16;
and.b32 %r123, %r122, 30583;
prmt.b32 %r124, %r98, %r97, %r123;
prmt.b32 %r125, %r101, %r100, %r123;
shr.u32 %r126, %r116, 16;
prmt.b32 %r127, %r124, %r125, %r126;
prmt.b32 %r68, %r121, %r127, %r111;
prmt.b32 %r72, %r121, %r127, %r112;
ld.global.nc.u32 %r12, [%rd98+-8];
ld.global.nc.u32 %r13, [%rd98+8];

dp4a.s32.s32 %r67, %r68, %r12, %r63;


dp4a.s32.s32 %r71, %r72, %r13, %r67;

ld.global.nc.u32 %r128, [%rd53+-4];
and.b32 %r129, %r128, -2004318072;
shr.u32 %r130, %r129, 1;
or.b32 %r131, %r130, 839922192;
and.b32 %r132, %r128, 30583;
prmt.b32 %r133, %r98, %r97, %r132;
prmt.b32 %r134, %r101, %r100, %r132;
and.b32 %r135, %r131, 30292;
prmt.b32 %r136, %r133, %r134, %r135;
shr.u32 %r137, %r128, 16;
and.b32 %r138, %r137, 30583;
prmt.b32 %r139, %r98, %r97, %r138;
prmt.b32 %r140, %r101, %r100, %r138;
shr.u32 %r141, %r131, 16;
prmt.b32 %r142, %r139, %r140, %r141;
prmt.b32 %r76, %r136, %r142, %r111;
prmt.b32 %r80, %r136, %r142, %r112;
ld.global.nc.u32 %r14, [%rd98+-4];
ld.global.nc.u32 %r15, [%rd98+12];

dp4a.s32.s32 %r75, %r76, %r14, %r71;


dp4a.s32.s32 %r79, %r80, %r15, %r75;

ld.global.nc.u32 %r143, [%rd53];
and.b32 %r144, %r143, -2004318072;
shr.u32 %r145, %r144, 1;
or.b32 %r146, %r145, 839922192;
and.b32 %r147, %r143, 30583;
prmt.b32 %r148, %r98, %r97, %r147;
prmt.b32 %r149, %r101, %r100, %r147;
and.b32 %r150, %r146, 30292;
prmt.b32 %r151, %r148, %r149, %r150;
shr.u32 %r152, %r143, 16;
and.b32 %r153, %r152, 30583;
prmt.b32 %r154, %r98, %r97, %r153;
prmt.b32 %r155, %r101, %r100, %r153;
shr.u32 %r156, %r146, 16;
prmt.b32 %r157, %r154, %r155, %r156;
prmt.b32 %r84, %r151, %r157, %r111;
prmt.b32 %r88, %r151, %r157, %r112;
ld.global.nc.u32 %r16, [%rd98];
ld.global.nc.u32 %r17, [%rd98+16];

dp4a.s32.s32 %r83, %r84, %r16, %r79;


dp4a.s32.s32 %r87, %r88, %r17, %r83;

add.s64 %rd54, %rd99, %rd7;
ld.global.nc.u8 %rs2, [%rd54];
cvt.u32.u16 %r158, %rs2;
and.b32 %r159, %r158, 255;
shr.u32 %r160, %r159, %r7;
and.b32 %r161, %r160, 15;
add.s64 %rd55, %rd99, %rd6;
ld.global.nc.u16 %rs3, [%rd55+2];
cvt.u32.u16 %r162, %rs3;
shr.u32 %r163, %r162, %r8;
and.b32 %r164, %r163, 3;
bfi.b32 %r165, %r164, %r161, 4, 2;
add.s32 %r166, %r165, -32;
mul.lo.s32 %r167, %r166, %r87;
ld.global.nc.u16 %rs1, [%rd55];

{ cvt.f32.f16 %f42, %rs1;}


ld.global.nc.u32 %r18, [%rd98+-16];

{.reg .f16 low,high;
mov.b32 {low,high},%r18;
cvt.f32.f16 %f43, low;}


mul.ftz.f32 %f44, %f42, %f43;
cvt.rn.f32.s32 %f45, %r167;
fma.rn.ftz.f32 %f165, %f44, %f45, %f165;
setp.eq.s64 %p3, %rd28, 0;
@%p3 bra $L__BB291_5;

add.s64 %rd56, %rd97, %rd9;
ld.global.u32 %r201, [%rd56+-12];
and.b32 %r202, %r201, -2004318072;
shr.u32 %r203, %r202, 1;
or.b32 %r204, %r203, 839922192;
and.b32 %r205, %r201, 30583;
prmt.b32 %r208, %r98, %r97, %r205;
prmt.b32 %r211, %r101, %r100, %r205;
and.b32 %r212, %r204, 30292;
prmt.b32 %r213, %r208, %r211, %r212;
shr.u32 %r214, %r201, 16;
and.b32 %r215, %r214, 30583;
prmt.b32 %r216, %r98, %r97, %r215;
prmt.b32 %r217, %r101, %r100, %r215;
shr.u32 %r218, %r204, 16;
prmt.b32 %r219, %r216, %r217, %r218;
prmt.b32 %r169, %r213, %r219, %r111;
prmt.b32 %r173, %r213, %r219, %r112;

dp4a.s32.s32 %r168, %r169, %r10, %r62;


dp4a.s32.s32 %r172, %r173, %r11, %r168;

ld.global.u32 %r222, [%rd56+-8];
and.b32 %r223, %r222, -2004318072;
shr.u32 %r224, %r223, 1;
or.b32 %r225, %r224, 839922192;
and.b32 %r226, %r222, 30583;
prmt.b32 %r227, %r98, %r97, %r226;
prmt.b32 %r228, %r101, %r100, %r226;
and.b32 %r229, %r225, 30292;
prmt.b32 %r230, %r227, %r228, %r229;
shr.u32 %r231, %r222, 16;
and.b32 %r232, %r231, 30583;
prmt.b32 %r233, %r98, %r97, %r232;
prmt.b32 %r234, %r101, %r100, %r232;
shr.u32 %r235, %r225, 16;
prmt.b32 %r236, %r233, %r234, %r235;
prmt.b32 %r177, %r230, %r236, %r111;
prmt.b32 %r181, %r230, %r236, %r112;

dp4a.s32.s32 %r176, %r177, %r12, %r172;


dp4a.s32.s32 %r180, %r181, %r13, %r176;

ld.global.u32 %r237, [%rd56+-4];
and.b32 %r238, %r237, -2004318072;
shr.u32 %r239, %r238, 1;
or.b32 %r240, %r239, 839922192;
and.b32 %r241, %r237, 30583;
prmt.b32 %r242, %r98, %r97, %r241;
prmt.b32 %r243, %r101, %r100, %r241;
and.b32 %r244, %r240, 30292;
prmt.b32 %r245, %r242, %r243, %r244;
shr.u32 %r246, %r237, 16;
and.b32 %r247, %r246, 30583;
prmt.b32 %r248, %r98, %r97, %r247;
prmt.b32 %r249, %r101, %r100, %r247;
shr.u32 %r250, %r240, 16;
prmt.b32 %r251, %r248, %r249, %r250;
prmt.b32 %r185, %r245, %r251, %r111;
prmt.b32 %r189, %r245, %r251, %r112;

dp4a.s32.s32 %r184, %r185, %r14, %r180;


dp4a.s32.s32 %r188, %r189, %r15, %r184;

ld.global.u32 %r252, [%rd56];
and.b32 %r253, %r252, -2004318072;
shr.u32 %r254, %r253, 1;
or.b32 %r255, %r254, 839922192;
and.b32 %r256, %r252, 30583;
prmt.b32 %r257, %r98, %r97, %r256;
prmt.b32 %r258, %r101, %r100, %r256;
and.b32 %r259, %r255, 30292;
prmt.b32 %r260, %r257, %r258, %r259;
shr.u32 %r261, %r252, 16;
and.b32 %r262, %r261, 30583;
prmt.b32 %r263, %r98, %r97, %r262;
prmt.b32 %r264, %r101, %r100, %r262;
shr.u32 %r265, %r255, 16;
prmt.b32 %r266, %r263, %r264, %r265;
prmt.b32 %r193, %r260, %r266, %r111;
prmt.b32 %r197, %r260, %r266, %r112;

dp4a.s32.s32 %r192, %r193, %r16, %r188;


dp4a.s32.s32 %r196, %r197, %r17, %r192;

add.s64 %rd57, %rd97, %rd7;
ld.global.u8 %r267, [%rd57];
shr.u32 %r268, %r267, %r7;
and.b32 %r269, %r268, 15;
add.s64 %rd58, %rd97, %rd6;
ld.global.u16 %r270, [%rd58+2];
shr.u32 %r271, %r270, %r8;
and.b32 %r272, %r271, 3;
bfi.b32 %r273, %r272, %r269, 4, 2;
add.s32 %r274, %r273, -32;
mul.lo.s32 %r275, %r274, %r196;
ld.global.u16 %rs4, [%rd58];

{ cvt.f32.f16 %f46, %rs4;}



{.reg .f16 low,high;
mov.b32 {low,high},%r18;
cvt.f32.f16 %f47, low;}


mul.ftz.f32 %f48, %f46, %f47;
cvt.rn.f32.s32 %f49, %r275;
fma.rn.ftz.f32 %f163, %f48, %f49, %f163;

$L__BB291_5:
add.s64 %rd59, %rd99, %rd11;
ld.global.nc.u32 %r309, [%rd59+8];
and.b32 %r310, %r309, -2004318072;
shr.u32 %r311, %r310, 1;
or.b32 %r312, %r311, 839922192;
and.b32 %r313, %r309, 30583;
prmt.b32 %r316, %r98, %r97, %r313;
prmt.b32 %r319, %r101, %r100, %r313;
and.b32 %r320, %r312, 30292;
prmt.b32 %r321, %r316, %r319, %r320;
shr.u32 %r322, %r309, 16;
and.b32 %r323, %r322, 30583;
prmt.b32 %r324, %r98, %r97, %r323;
prmt.b32 %r325, %r101, %r100, %r323;
shr.u32 %r326, %r312, 16;
prmt.b32 %r327, %r324, %r325, %r326;
prmt.b32 %r277, %r321, %r327, %r111;
prmt.b32 %r281, %r321, %r327, %r112;

dp4a.s32.s32 %r276, %r277, %r10, %r62;


dp4a.s32.s32 %r280, %r281, %r11, %r276;

ld.global.nc.u32 %r330, [%rd59+12];
and.b32 %r331, %r330, -2004318072;
shr.u32 %r332, %r331, 1;
or.b32 %r333, %r332, 839922192;
and.b32 %r334, %r330, 30583;
prmt.b32 %r335, %r98, %r97, %r334;
prmt.b32 %r336, %r101, %r100, %r334;
and.b32 %r337, %r333, 30292;
prmt.b32 %r338, %r335, %r336, %r337;
shr.u32 %r339, %r330, 16;
and.b32 %r340, %r339, 30583;
prmt.b32 %r341, %r98, %r97, %r340;
prmt.b32 %r342, %r101, %r100, %r340;
shr.u32 %r343, %r333, 16;
prmt.b32 %r344, %r341, %r342, %r343;
prmt.b32 %r285, %r338, %r344, %r111;
prmt.b32 %r289, %r338, %r344, %r112;

dp4a.s32.s32 %r284, %r285, %r12, %r280;


dp4a.s32.s32 %r288, %r289, %r13, %r284;

ld.global.nc.u32 %r345, [%rd59+16];
and.b32 %r346, %r345, -2004318072;
shr.u32 %r347, %r346, 1;
or.b32 %r348, %r347, 839922192;
and.b32 %r349, %r345, 30583;
prmt.b32 %r350, %r98, %r97, %r349;
prmt.b32 %r351, %r101, %r100, %r349;
and.b32 %r352, %r348, 30292;
prmt.b32 %r353, %r350, %r351, %r352;
shr.u32 %r354, %r345, 16;
and.b32 %r355, %r354, 30583;
prmt.b32 %r356, %r98, %r97, %r355;
prmt.b32 %r357, %r101, %r100, %r355;
shr.u32 %r358, %r348, 16;
prmt.b32 %r359, %r356, %r357, %r358;
prmt.b32 %r293, %r353, %r359, %r111;
prmt.b32 %r297, %r353, %r359, %r112;

dp4a.s32.s32 %r292, %r293, %r14, %r288;


dp4a.s32.s32 %r296, %r297, %r15, %r292;

ld.global.nc.u32 %r360, [%rd59+20];
and.b32 %r361, %r360, -2004318072;
shr.u32 %r362, %r361, 1;
or.b32 %r363, %r362, 839922192;
and.b32 %r364, %r360, 30583;
prmt.b32 %r365, %r98, %r97, %r364;
prmt.b32 %r366, %r101, %r100, %r364;
and.b32 %r367, %r363, 30292;
prmt.b32 %r368, %r365, %r366, %r367;
shr.u32 %r369, %r360, 16;
and.b32 %r370, %r369, 30583;
prmt.b32 %r371, %r98, %r97, %r370;
prmt.b32 %r372, %r101, %r100, %r370;
shr.u32 %r373, %r363, 16;
prmt.b32 %r374, %r371, %r372, %r373;
prmt.b32 %r301, %r368, %r374, %r111;
prmt.b32 %r305, %r368, %r374, %r112;

dp4a.s32.s32 %r300, %r301, %r16, %r296;


dp4a.s32.s32 %r304, %r305, %r17, %r300;

add.s64 %rd60, %rd99, %rd12;
ld.global.nc.u8 %rs6, [%rd60];
cvt.u32.u16 %r375, %rs6;
and.b32 %r376, %r375, 255;
shr.u32 %r377, %r376, %r7;
and.b32 %r378, %r377, 15;
add.s64 %rd61, %rd99, %rd10;
ld.global.nc.u16 %rs7, [%rd61+2];
cvt.u32.u16 %r379, %rs7;
shr.u32 %r380, %r379, %r8;
and.b32 %r381, %r380, 3;
bfi.b32 %r382, %r381, %r378, 4, 2;
add.s32 %r383, %r382, -32;
mul.lo.s32 %r384, %r383, %r304;
ld.global.nc.u16 %rs5, [%rd61];

{ cvt.f32.f16 %f50, %rs5;}



{.reg .f16 low,high;
mov.b32 {low,high},%r18;
cvt.f32.f16 %f51, low;}


mul.ftz.f32 %f52, %f50, %f51;
cvt.rn.f32.s32 %f53, %r384;
fma.rn.ftz.f32 %f164, %f52, %f53, %f164;
@%p3 bra $L__BB291_7;

add.s64 %rd62, %rd97, %rd11;
ld.global.u32 %r418, [%rd62+8];
and.b32 %r419, %r418, -2004318072;
shr.u32 %r420, %r419, 1;
or.b32 %r421, %r420, 839922192;
and.b32 %r422, %r418, 30583;
mov.u32 %r423, -152379953;
mov.u32 %r424, -1079142271;
prmt.b32 %r425, %r424, %r423, %r422;
mov.u32 %r426, 1901675829;
mov.u32 %r427, 639175937;
prmt.b32 %r428, %r427, %r426, %r422;
and.b32 %r429, %r421, 30292;
prmt.b32 %r430, %r425, %r428, %r429;
shr.u32 %r431, %r418, 16;
and.b32 %r432, %r431, 30583;
prmt.b32 %r433, %r424, %r423, %r432;
prmt.b32 %r434, %r427, %r426, %r432;
shr.u32 %r435, %r421, 16;
prmt.b32 %r436, %r433, %r434, %r435;
mov.u32 %r437, 25632;
prmt.b32 %r386, %r430, %r436, %r437;
mov.u32 %r438, 30001;
prmt.b32 %r390, %r430, %r436, %r438;
mov.u32 %r388, 0;

dp4a.s32.s32 %r385, %r386, %r10, %r388;


dp4a.s32.s32 %r389, %r390, %r11, %r385;

ld.global.u32 %r439, [%rd62+12];
and.b32 %r440, %r439, -2004318072;
shr.u32 %r441, %r440, 1;
or.b32 %r442, %r441, 839922192;
and.b32 %r443, %r439, 30583;
prmt.b32 %r444, %r424, %r423, %r443;
prmt.b32 %r445, %r427, %r426, %r443;
and.b32 %r446, %r442, 30292;
prmt.b32 %r447, %r444, %r445, %r446;
shr.u32 %r448, %r439, 16;
and.b32 %r449, %r448, 30583;
prmt.b32 %r450, %r424, %r423, %r449;
prmt.b32 %r451, %r427, %r426, %r449;
shr.u32 %r452, %r442, 16;
prmt.b32 %r453, %r450, %r451, %r452;
prmt.b32 %r394, %r447, %r453, %r437;
prmt.b32 %r398, %r447, %r453, %r438;

dp4a.s32.s32 %r393, %r394, %r12, %r389;


dp4a.s32.s32 %r397, %r398, %r13, %r393;

ld.global.u32 %r454, [%rd62+16];
and.b32 %r455, %r454, -2004318072;
shr.u32 %r456, %r455, 1;
or.b32 %r457, %r456, 839922192;
and.b32 %r458, %r454, 30583;
prmt.b32 %r459, %r424, %r423, %r458;
prmt.b32 %r460, %r427, %r426, %r458;
and.b32 %r461, %r457, 30292;
prmt.b32 %r462, %r459, %r460, %r461;
shr.u32 %r463, %r454, 16;
and.b32 %r464, %r463, 30583;
prmt.b32 %r465, %r424, %r423, %r464;
prmt.b32 %r466, %r427, %r426, %r464;
shr.u32 %r467, %r457, 16;
prmt.b32 %r468, %r465, %r466, %r467;
prmt.b32 %r402, %r462, %r468, %r437;
prmt.b32 %r406, %r462, %r468, %r438;

dp4a.s32.s32 %r401, %r402, %r14, %r397;


dp4a.s32.s32 %r405, %r406, %r15, %r401;

ld.global.u32 %r469, [%rd62+20];
and.b32 %r470, %r469, -2004318072;
shr.u32 %r471, %r470, 1;
or.b32 %r472, %r471, 839922192;
and.b32 %r473, %r469, 30583;
prmt.b32 %r474, %r424, %r423, %r473;
prmt.b32 %r475, %r427, %r426, %r473;
and.b32 %r476, %r472, 30292;
prmt.b32 %r477, %r474, %r475, %r476;
shr.u32 %r478, %r469, 16;
and.b32 %r479, %r478, 30583;
prmt.b32 %r480, %r424, %r423, %r479;
prmt.b32 %r481, %r427, %r426, %r479;
shr.u32 %r482, %r472, 16;
prmt.b32 %r483, %r480, %r481, %r482;
prmt.b32 %r410, %r477, %r483, %r437;
prmt.b32 %r414, %r477, %r483, %r438;

dp4a.s32.s32 %r409, %r410, %r16, %r405;


dp4a.s32.s32 %r413, %r414, %r17, %r409;

add.s64 %rd63, %rd97, %rd12;
ld.global.u8 %r484, [%rd63];
shr.u32 %r485, %r484, %r7;
and.b32 %r486, %r485, 15;
add.s64 %rd64, %rd97, %rd10;
ld.global.u16 %r487, [%rd64+2];
shr.u32 %r488, %r487, %r8;
and.b32 %r489, %r488, 3;
bfi.b32 %r490, %r489, %r486, 4, 2;
add.s32 %r491, %r490, -32;
mul.lo.s32 %r492, %r491, %r413;
ld.global.u16 %rs8, [%rd64];

{ cvt.f32.f16 %f54, %rs8;}



{.reg .f16 low,high;
mov.b32 {low,high},%r18;
cvt.f32.f16 %f55, low;}


mul.ftz.f32 %f56, %f54, %f55;
cvt.rn.f32.s32 %f57, %r492;
fma.rn.ftz.f32 %f162, %f56, %f57, %f162;

$L__BB291_7:
add.s64 %rd99, %rd99, 544;
add.s64 %rd98, %rd98, 1152;
add.s64 %rd97, %rd97, 544;
add.s32 %r587, %r587, 4;
setp.lt.u32 %p5, %r587, %r1;
@%p5 bra $L__BB291_3;

add.u64 %rd89, %SPL, 0;
add.u64 %rd83, %SPL, 8;
st.local.v2.f32 [%rd89], {%f165, %f164};
st.local.v2.f32 [%rd83], {%f163, %f162};

$L__BB291_9:
mov.b32 %r20, %f165;
setp.ne.s64 %p6, %rd28, 0;
@%p6 bra $L__BB291_11;
bra.uni $L__BB291_10;

$L__BB291_11:
add.u64 %rd95, %SPL, 0;
add.u64 %rd87, %SPL, 8;
mov.u32 %r519, 31;
mov.u32 %r520, 16;
mov.u32 %r521, -1;
shfl.sync.bfly.b32 %r522|%p17, %r20, %r520, %r519, %r521;
mov.b32 %f78, %r522;
add.ftz.f32 %f79, %f165, %f78;
mov.b32 %r523, %f79;
mov.u32 %r524, 8;
shfl.sync.bfly.b32 %r525|%p18, %r523, %r524, %r519, %r521;
mov.b32 %f80, %r525;
add.ftz.f32 %f81, %f79, %f80;
mov.b32 %r526, %f81;
mov.u32 %r527, 4;
shfl.sync.bfly.b32 %r528|%p19, %r526, %r527, %r519, %r521;
mov.b32 %f82, %r528;
add.ftz.f32 %f83, %f81, %f82;
mov.b32 %r529, %f83;
mov.u32 %r530, 2;
shfl.sync.bfly.b32 %r531|%p20, %r529, %r530, %r519, %r521;
mov.b32 %f84, %r531;
add.ftz.f32 %f85, %f83, %f84;
mov.b32 %r532, %f85;
mov.u32 %r533, 1;
shfl.sync.bfly.b32 %r534|%p21, %r532, %r533, %r519, %r521;
mov.b32 %f86, %r534;
add.ftz.f32 %f87, %f85, %f86;
st.local.f32 [%rd95], %f87;
mov.b32 %r535, %f163;
shfl.sync.bfly.b32 %r536|%p22, %r535, %r520, %r519, %r521;
mov.b32 %f88, %r536;
add.ftz.f32 %f89, %f163, %f88;
mov.b32 %r537, %f89;
shfl.sync.bfly.b32 %r538|%p23, %r537, %r524, %r519, %r521;
mov.b32 %f90, %r538;
add.ftz.f32 %f91, %f89, %f90;
mov.b32 %r539, %f91;
shfl.sync.bfly.b32 %r540|%p24, %r539, %r527, %r519, %r521;
mov.b32 %f92, %r540;
add.ftz.f32 %f93, %f91, %f92;
mov.b32 %r541, %f93;
shfl.sync.bfly.b32 %r542|%p25, %r541, %r530, %r519, %r521;
mov.b32 %f94, %r542;
add.ftz.f32 %f95, %f93, %f94;
mov.b32 %r543, %f95;
shfl.sync.bfly.b32 %r544|%p26, %r543, %r533, %r519, %r521;
mov.b32 %f96, %r544;
add.ftz.f32 %f97, %f95, %f96;
st.local.f32 [%rd87], %f97;
mov.b32 %r545, %f164;
shfl.sync.bfly.b32 %r546|%p27, %r545, %r520, %r519, %r521;
mov.b32 %f98, %r546;
add.ftz.f32 %f99, %f164, %f98;
mov.b32 %r547, %f99;
shfl.sync.bfly.b32 %r548|%p28, %r547, %r524, %r519, %r521;
mov.b32 %f100, %r548;
add.ftz.f32 %f101, %f99, %f100;
mov.b32 %r549, %f101;
shfl.sync.bfly.b32 %r550|%p29, %r549, %r527, %r519, %r521;
mov.b32 %f102, %r550;
add.ftz.f32 %f103, %f101, %f102;
mov.b32 %r551, %f103;
shfl.sync.bfly.b32 %r552|%p30, %r551, %r530, %r519, %r521;
mov.b32 %f104, %r552;
add.ftz.f32 %f105, %f103, %f104;
mov.b32 %r553, %f105;
shfl.sync.bfly.b32 %r554|%p31, %r553, %r533, %r519, %r521;
mov.b32 %f106, %r554;
add.ftz.f32 %f107, %f105, %f106;
st.local.f32 [%rd95+4], %f107;
mov.b32 %r555, %f162;
shfl.sync.bfly.b32 %r556|%p32, %r555, %r520, %r519, %r521;
mov.b32 %f108, %r556;
add.ftz.f32 %f109, %f162, %f108;
mov.b32 %r557, %f109;
shfl.sync.bfly.b32 %r558|%p33, %r557, %r524, %r519, %r521;
mov.b32 %f110, %r558;
add.ftz.f32 %f111, %f109, %f110;
mov.b32 %r559, %f111;
shfl.sync.bfly.b32 %r560|%p34, %r559, %r527, %r519, %r521;
mov.b32 %f112, %r560;
add.ftz.f32 %f113, %f111, %f112;
mov.b32 %r561, %f113;
shfl.sync.bfly.b32 %r562|%p35, %r561, %r530, %r519, %r521;
mov.b32 %f114, %r562;
add.ftz.f32 %f115, %f113, %f114;
mov.b32 %r563, %f115;
shfl.sync.bfly.b32 %r564|%p36, %r563, %r533, %r519, %r521;
mov.b32 %f116, %r564;
add.ftz.f32 %f117, %f115, %f116;
st.local.f32 [%rd87+4], %f117;
bra.uni $L__BB291_12;

$L__BB291_10:
add.u64 %rd91, %SPL, 0;
mov.u32 %r493, 31;
mov.u32 %r494, 16;
mov.u32 %r495, -1;
shfl.sync.bfly.b32 %r496|%p7, %r20, %r494, %r493, %r495;
mov.b32 %f58, %r496;
add.ftz.f32 %f59, %f165, %f58;
mov.b32 %r497, %f59;
mov.u32 %r498, 8;
shfl.sync.bfly.b32 %r499|%p8, %r497, %r498, %r493, %r495;
mov.b32 %f60, %r499;
add.ftz.f32 %f61, %f59, %f60;
mov.b32 %r500, %f61;
mov.u32 %r501, 4;
shfl.sync.bfly.b32 %r502|%p9, %r500, %r501, %r493, %r495;
mov.b32 %f62, %r502;
add.ftz.f32 %f63, %f61, %f62;
mov.b32 %r503, %f63;
mov.u32 %r504, 2;
shfl.sync.bfly.b32 %r505|%p10, %r503, %r504, %r493, %r495;
mov.b32 %f64, %r505;
add.ftz.f32 %f65, %f63, %f64;
mov.b32 %r506, %f65;
mov.u32 %r507, 1;
shfl.sync.bfly.b32 %r508|%p11, %r506, %r507, %r493, %r495;
mov.b32 %f66, %r508;
add.ftz.f32 %f67, %f65, %f66;
st.local.f32 [%rd91], %f67;
mov.b32 %r509, %f164;
shfl.sync.bfly.b32 %r510|%p12, %r509, %r494, %r493, %r495;
mov.b32 %f68, %r510;
add.ftz.f32 %f69, %f164, %f68;
mov.b32 %r511, %f69;
shfl.sync.bfly.b32 %r512|%p13, %r511, %r498, %r493, %r495;
mov.b32 %f70, %r512;
add.ftz.f32 %f71, %f69, %f70;
mov.b32 %r513, %f71;
shfl.sync.bfly.b32 %r514|%p14, %r513, %r501, %r493, %r495;
mov.b32 %f72, %r514;
add.ftz.f32 %f73, %f71, %f72;
mov.b32 %r515, %f73;
shfl.sync.bfly.b32 %r516|%p15, %r515, %r504, %r493, %r495;
mov.b32 %f74, %r516;
add.ftz.f32 %f75, %f73, %f74;
mov.b32 %r517, %f75;
shfl.sync.bfly.b32 %r518|%p16, %r517, %r507, %r493, %r495;
mov.b32 %f76, %r518;
add.ftz.f32 %f77, %f75, %f76;
st.local.f32 [%rd91+4], %f77;

$L__BB291_12:
ld.param.u32 %r575, [_Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7];
mov.u32 %r574, %tid.x;
setp.gt.u32 %p37, %r574, 1;
mov.u32 %r566, %ctaid.x;
shl.b32 %r567, %r566, 1;
add.s32 %r21, %r567, %r574;
setp.ge.u32 %p38, %r21, %r575;
or.pred %p39, %p37, %p38;
@%p39 bra $L__BB291_29;

add.u64 %rd93, %SPL, 0;
ld.param.u64 %rd78, [_Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3];
ld.param.u32 %r577, [_Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13];
mov.u32 %r576, %tid.x;
cvt.u64.u32 %rd22, %r576;
mul.wide.u32 %rd65, %r576, 4;
add.s64 %rd66, %rd93, %rd65;
ld.local.f32 %f168, [%rd66];
mad.lo.s32 %r22, %r4, %r577, %r21;
setp.eq.s64 %p40, %rd78, 0;
@%p40 bra $L__BB291_15;

ld.param.u64 %rd79, [_Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3];
cvta.to.global.u64 %rd67, %rd79;
mul.wide.u32 %rd68, %r22, 4;
add.s64 %rd69, %rd67, %rd68;
ld.global.f32 %f118, [%rd69];
add.ftz.f32 %f168, %f168, %f118;

$L__BB291_15:
setp.eq.s64 %p41, %rd28, 0;
@%p41 bra $L__BB291_28;

add.u64 %rd85, %SPL, 8;
ld.param.u64 %rd81, [_Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+16];
shl.b64 %rd70, %rd22, 2;
add.s64 %rd71, %rd85, %rd70;
ld.local.f32 %f24, [%rd71];
setp.eq.s64 %p42, %rd81, 0;
@%p42 bra $L__BB291_18;

ld.param.u64 %rd82, [_Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+16];
cvta.to.global.u64 %rd72, %rd82;
mul.wide.u32 %rd73, %r22, 4;
add.s64 %rd74, %rd72, %rd73;
ld.global.f32 %f119, [%rd74];
add.ftz.f32 %f24, %f24, %f119;

$L__BB291_18:
ld.param.u32 %r582, [_Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.gt.s32 %p43, %r582, 2;
@%p43 bra $L__BB291_22;

ld.param.u32 %r585, [_Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.eq.s32 %p46, %r585, 1;
@%p46 bra $L__BB291_26;

ld.param.u32 %r586, [_Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.eq.s32 %p47, %r586, 2;
@%p47 bra $L__BB291_21;
bra.uni $L__BB291_27;

$L__BB291_21:
mul.ftz.f32 %f139, %f24, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f140, %f139;
add.ftz.f32 %f141, %f140, 0f3F800000;
div.approx.ftz.f32 %f142, %f24, %f141;
mul.ftz.f32 %f168, %f168, %f142;
bra.uni $L__BB291_28;

$L__BB291_22:
ld.param.u32 %r583, [_Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.eq.s32 %p44, %r583, 3;
@%p44 bra $L__BB291_25;

ld.param.u32 %r584, [_Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.ne.s32 %p45, %r584, 6;
@%p45 bra $L__BB291_27;

ld.param.f32 %f151, [_Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+44];
min.ftz.f32 %f120, %f24, %f151;
neg.ftz.f32 %f121, %f151;
min.ftz.f32 %f122, %f168, %f151;
max.ftz.f32 %f123, %f122, %f121;
mul.ftz.f32 %f124, %f120, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f125, %f124;
add.ftz.f32 %f126, %f125, 0f3F800000;
div.approx.ftz.f32 %f127, %f120, %f126;
mul.ftz.f32 %f168, %f123, %f127;
bra.uni $L__BB291_28;

$L__BB291_27:
mul.ftz.f32 %f168, %f168, %f24;
bra.uni $L__BB291_28;

$L__BB291_26:
mul.ftz.f32 %f143, %f24, 0f3F000000;
mul.ftz.f32 %f144, %f24, 0f3D372713;
fma.rn.ftz.f32 %f145, %f24, %f144, 0f3F800000;
mul.ftz.f32 %f146, %f24, 0f3F4C422A;
mul.ftz.f32 %f147, %f146, %f145;
tanh.approx.f32 %f148, %f147;
add.ftz.f32 %f149, %f148, 0f3F800000;
mul.ftz.f32 %f150, %f143, %f149;
mul.ftz.f32 %f168, %f168, %f150;
bra.uni $L__BB291_28;

$L__BB291_25:
mov.f32 %f128, 0f40E00000;
min.ftz.f32 %f129, %f24, %f128;
min.ftz.f32 %f130, %f168, %f128;
mov.f32 %f131, 0fC0E00000;
max.ftz.f32 %f132, %f130, %f131;
mul.ftz.f32 %f133, %f129, 0f3FD9DB23;
mul.ftz.f32 %f134, %f133, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f135, %f134;
add.ftz.f32 %f136, %f135, 0f3F800000;
div.approx.ftz.f32 %f137, %f129, %f136;
add.ftz.f32 %f138, %f132, 0f3F800000;
mul.ftz.f32 %f168, %f138, %f137;

$L__BB291_28:
ld.param.u64 %rd80, [_Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4];
ld.param.u32 %r581, [_Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10];
ld.param.u32 %r580, [_Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13];
mov.u32 %r579, %ctaid.y;
mov.u32 %r578, %tid.y;
mad.lo.s32 %r572, %r578, %r581, %r21;
mad.lo.s32 %r573, %r579, %r580, %r572;
cvta.to.global.u64 %rd75, %rd80;
mul.wide.u32 %rd76, %r573, 4;
add.s64 %rd77, %rd75, %rd76;
st.global.f32 [%rd77], %f168;

$L__BB291_29:
ret;

}
.entry _Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj(
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2,
.param .align 8 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3[48],
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5,
.param .align 4 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6[12],
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15
)
.maxntid 256, 1, 1
.minnctapersm 1
{
.local .align 8 .b8 __local_depot292[8];
.reg .b64 %SP;
.reg .b64 %SPL;
.reg .pred %p<17>;
.reg .b16 %rs<7>;
.reg .f32 %f<44>;
.reg .b32 %r<291>;
.reg .b64 %rd<52>;


mov.u64 %SPL, __local_depot292;
ld.param.u64 %rd15, [_Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0];
ld.param.u64 %rd16, [_Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1];
ld.param.u64 %rd17, [_Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2];
ld.param.u64 %rd18, [_Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4];
ld.param.u32 %r22, [_Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5];
ld.param.u32 %r14, [_Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7];
ld.param.u32 %r15, [_Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8];
ld.param.u32 %r16, [_Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9];
ld.param.u32 %r17, [_Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10];
ld.param.u32 %r18, [_Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11];
ld.param.u32 %r19, [_Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12];
ld.param.u32 %r20, [_Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13];
ld.param.u32 %r23, [_Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14];
ld.param.u32 %r21, [_Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15];
ld.param.u32 %r13, [_Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+8];
ld.param.u32 %r12, [_Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+4];
ld.param.u32 %r11, [_Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6];
add.u64 %rd1, %SPL, 0;
mov.u32 %r24, %ctaid.x;
shl.b32 %r1, %r24, 1;
shr.u32 %r2, %r22, 8;
mov.u32 %r3, %tid.y;
setp.ge.u32 %p1, %r3, %r23;
@%p1 bra $L__BB292_7;

mov.f32 %f42, 0f00000000;
st.local.v2.f32 [%rd1], {%f42, %f42};
mov.u32 %r4, %tid.x;
shr.u32 %r290, %r4, 3;
setp.ge.u32 %p2, %r290, %r2;
mov.f32 %f43, %f42;
@%p2 bra $L__BB292_5;

mov.u32 %r25, %ctaid.y;
mad.lo.s32 %r26, %r3, %r21, %r25;
cvta.to.global.u64 %rd20, %rd17;
mul.wide.u32 %rd21, %r26, 4;
add.s64 %rd22, %rd20, %rd21;
ld.global.nc.u32 %r27, [%rd22];
mul.lo.s32 %r28, %r27, %r18;
and.b32 %r29, %r4, 7;
shl.b32 %r30, %r4, 2;
shr.u32 %r31, %r29, 1;
cvt.u64.u32 %rd23, %r31;
and.b32 %r6, %r30, 4;
shl.b32 %r32, %r4, 1;
and.b32 %r7, %r32, 14;
mad.lo.s32 %r33, %r1, %r15, %r290;
add.s32 %r34, %r33, %r28;
mul.wide.s32 %rd3, %r34, 136;
or.b64 %rd24, %rd23, %rd3;
or.b64 %rd4, %rd24, 4;
mul.hi.u32 %r35, %r25, %r11;
add.s32 %r36, %r35, %r25;
shr.u32 %r37, %r36, %r12;
mul.lo.s32 %r38, %r37, %r13;
sub.s32 %r39, %r25, %r38;
mul.lo.s32 %r40, %r39, %r19;
mul.wide.u32 %rd25, %r40, 36;
mul.lo.s32 %r41, %r3, %r16;
mul.wide.u32 %rd26, %r41, 36;
add.s64 %rd27, %rd26, %rd25;
mul.wide.u32 %rd28, %r29, 36;
add.s64 %rd29, %rd27, %rd28;
shl.b32 %r42, %r290, 3;
mul.wide.s32 %rd30, %r42, 36;
add.s64 %rd31, %rd29, %rd30;
cvta.to.global.u64 %rd32, %rd16;
add.s64 %rd33, %rd32, %rd31;
add.s64 %rd50, %rd33, 16;
shl.b32 %r43, %r4, 4;
cvt.u64.u32 %rd34, %r43;
and.b64 %rd35, %rd34, 112;
add.s64 %rd36, %rd35, %rd3;
add.s64 %rd6, %rd36, 20;
add.s32 %r44, %r1, 1;
add.s32 %r45, %r290, %r28;
mad.lo.s32 %r46, %r15, %r44, %r45;
mul.wide.s32 %rd7, %r46, 136;
add.s64 %rd37, %rd35, %rd7;
add.s64 %rd8, %rd37, 20;
or.b64 %rd38, %rd23, %rd7;
or.b64 %rd9, %rd38, 4;
cvta.to.global.u64 %rd51, %rd15;
mov.f32 %f42, 0f00000000;

$L__BB292_3:
add.s64 %rd39, %rd51, %rd6;
ld.global.nc.u32 %r113, [%rd39+-12];
and.b32 %r114, %r113, -2004318072;
shr.u32 %r115, %r114, 1;
or.b32 %r116, %r115, 839922192;
and.b32 %r117, %r113, 30583;
mov.u32 %r118, -152379953;
mov.u32 %r119, -1079142271;
prmt.b32 %r120, %r119, %r118, %r117;
mov.u32 %r121, 1901675829;
mov.u32 %r122, 639175937;
prmt.b32 %r123, %r122, %r121, %r117;
and.b32 %r124, %r116, 30292;
prmt.b32 %r125, %r120, %r123, %r124;
shr.u32 %r126, %r113, 16;
and.b32 %r127, %r126, 30583;
prmt.b32 %r128, %r119, %r118, %r127;
prmt.b32 %r129, %r122, %r121, %r127;
shr.u32 %r130, %r116, 16;
prmt.b32 %r131, %r128, %r129, %r130;
mov.u32 %r132, 25632;
prmt.b32 %r48, %r125, %r131, %r132;
mov.u32 %r133, 30001;
prmt.b32 %r52, %r125, %r131, %r133;
ld.global.nc.u32 %r49, [%rd50+-12];
ld.global.nc.u32 %r53, [%rd50+4];
mov.u32 %r83, 0;

dp4a.s32.s32 %r47, %r48, %r49, %r83;


dp4a.s32.s32 %r51, %r52, %r53, %r47;

ld.global.nc.u32 %r134, [%rd39+-8];
and.b32 %r135, %r134, -2004318072;
shr.u32 %r136, %r135, 1;
or.b32 %r137, %r136, 839922192;
and.b32 %r138, %r134, 30583;
prmt.b32 %r139, %r119, %r118, %r138;
prmt.b32 %r140, %r122, %r121, %r138;
and.b32 %r141, %r137, 30292;
prmt.b32 %r142, %r139, %r140, %r141;
shr.u32 %r143, %r134, 16;
and.b32 %r144, %r143, 30583;
prmt.b32 %r145, %r119, %r118, %r144;
prmt.b32 %r146, %r122, %r121, %r144;
shr.u32 %r147, %r137, 16;
prmt.b32 %r148, %r145, %r146, %r147;
prmt.b32 %r56, %r142, %r148, %r132;
prmt.b32 %r60, %r142, %r148, %r133;
ld.global.nc.u32 %r57, [%rd50+-8];
ld.global.nc.u32 %r61, [%rd50+8];

dp4a.s32.s32 %r55, %r56, %r57, %r51;


dp4a.s32.s32 %r59, %r60, %r61, %r55;

ld.global.nc.u32 %r149, [%rd39+-4];
and.b32 %r150, %r149, -2004318072;
shr.u32 %r151, %r150, 1;
or.b32 %r152, %r151, 839922192;
and.b32 %r153, %r149, 30583;
prmt.b32 %r154, %r119, %r118, %r153;
prmt.b32 %r155, %r122, %r121, %r153;
and.b32 %r156, %r152, 30292;
prmt.b32 %r157, %r154, %r155, %r156;
shr.u32 %r158, %r149, 16;
and.b32 %r159, %r158, 30583;
prmt.b32 %r160, %r119, %r118, %r159;
prmt.b32 %r161, %r122, %r121, %r159;
shr.u32 %r162, %r152, 16;
prmt.b32 %r163, %r160, %r161, %r162;
prmt.b32 %r64, %r157, %r163, %r132;
prmt.b32 %r68, %r157, %r163, %r133;
ld.global.nc.u32 %r65, [%rd50+-4];
ld.global.nc.u32 %r69, [%rd50+12];

dp4a.s32.s32 %r63, %r64, %r65, %r59;


dp4a.s32.s32 %r67, %r68, %r69, %r63;

ld.global.nc.u32 %r164, [%rd39];
and.b32 %r165, %r164, -2004318072;
shr.u32 %r166, %r165, 1;
or.b32 %r167, %r166, 839922192;
and.b32 %r168, %r164, 30583;
prmt.b32 %r169, %r119, %r118, %r168;
prmt.b32 %r170, %r122, %r121, %r168;
and.b32 %r171, %r167, 30292;
prmt.b32 %r172, %r169, %r170, %r171;
shr.u32 %r173, %r164, 16;
and.b32 %r174, %r173, 30583;
prmt.b32 %r175, %r119, %r118, %r174;
prmt.b32 %r176, %r122, %r121, %r174;
shr.u32 %r177, %r167, 16;
prmt.b32 %r178, %r175, %r176, %r177;
prmt.b32 %r72, %r172, %r178, %r132;
prmt.b32 %r76, %r172, %r178, %r133;
ld.global.nc.u32 %r73, [%rd50];
ld.global.nc.u32 %r77, [%rd50+16];

dp4a.s32.s32 %r71, %r72, %r73, %r67;


dp4a.s32.s32 %r75, %r76, %r77, %r71;

add.s64 %rd40, %rd51, %rd4;
ld.global.nc.u8 %rs3, [%rd40];
cvt.u32.u16 %r179, %rs3;
and.b32 %r180, %r179, 255;
shr.u32 %r181, %r180, %r6;
and.b32 %r182, %r181, 15;
add.s64 %rd41, %rd51, %rd3;
ld.global.nc.u16 %rs4, [%rd41+2];
cvt.u32.u16 %r183, %rs4;
shr.u32 %r184, %r183, %r7;
and.b32 %r185, %r184, 3;
bfi.b32 %r186, %r185, %r182, 4, 2;
add.s32 %r187, %r186, -32;
mul.lo.s32 %r188, %r187, %r75;
ld.global.nc.u16 %rs1, [%rd41];

{ cvt.f32.f16 %f11, %rs1;}


ld.global.nc.u32 %r79, [%rd50+-16];

{.reg .f16 low,high;
mov.b32 {low,high},%r79;
cvt.f32.f16 %f12, low;}


mul.ftz.f32 %f15, %f11, %f12;
cvt.rn.f32.s32 %f16, %r188;
fma.rn.ftz.f32 %f43, %f15, %f16, %f43;
add.s64 %rd42, %rd51, %rd8;
ld.global.nc.u32 %r189, [%rd42+-12];
and.b32 %r190, %r189, -2004318072;
shr.u32 %r191, %r190, 1;
or.b32 %r192, %r191, 839922192;
and.b32 %r193, %r189, 30583;
prmt.b32 %r194, %r119, %r118, %r193;
prmt.b32 %r195, %r122, %r121, %r193;
and.b32 %r196, %r192, 30292;
prmt.b32 %r197, %r194, %r195, %r196;
shr.u32 %r198, %r189, 16;
and.b32 %r199, %r198, 30583;
prmt.b32 %r200, %r119, %r118, %r199;
prmt.b32 %r201, %r122, %r121, %r199;
shr.u32 %r202, %r192, 16;
prmt.b32 %r203, %r200, %r201, %r202;
prmt.b32 %r81, %r197, %r203, %r132;
prmt.b32 %r85, %r197, %r203, %r133;

dp4a.s32.s32 %r80, %r81, %r49, %r83;


dp4a.s32.s32 %r84, %r85, %r53, %r80;

ld.global.nc.u32 %r204, [%rd42+-8];
and.b32 %r205, %r204, -2004318072;
shr.u32 %r206, %r205, 1;
or.b32 %r207, %r206, 839922192;
and.b32 %r208, %r204, 30583;
prmt.b32 %r209, %r119, %r118, %r208;
prmt.b32 %r210, %r122, %r121, %r208;
and.b32 %r211, %r207, 30292;
prmt.b32 %r212, %r209, %r210, %r211;
shr.u32 %r213, %r204, 16;
and.b32 %r214, %r213, 30583;
prmt.b32 %r215, %r119, %r118, %r214;
prmt.b32 %r216, %r122, %r121, %r214;
shr.u32 %r217, %r207, 16;
prmt.b32 %r218, %r215, %r216, %r217;
prmt.b32 %r89, %r212, %r218, %r132;
prmt.b32 %r93, %r212, %r218, %r133;

dp4a.s32.s32 %r88, %r89, %r57, %r84;


dp4a.s32.s32 %r92, %r93, %r61, %r88;

ld.global.nc.u32 %r219, [%rd42+-4];
and.b32 %r220, %r219, -2004318072;
shr.u32 %r221, %r220, 1;
or.b32 %r222, %r221, 839922192;
and.b32 %r223, %r219, 30583;
prmt.b32 %r224, %r119, %r118, %r223;
prmt.b32 %r225, %r122, %r121, %r223;
and.b32 %r226, %r222, 30292;
prmt.b32 %r227, %r224, %r225, %r226;
shr.u32 %r228, %r219, 16;
and.b32 %r229, %r228, 30583;
prmt.b32 %r230, %r119, %r118, %r229;
prmt.b32 %r231, %r122, %r121, %r229;
shr.u32 %r232, %r222, 16;
prmt.b32 %r233, %r230, %r231, %r232;
prmt.b32 %r97, %r227, %r233, %r132;
prmt.b32 %r101, %r227, %r233, %r133;

dp4a.s32.s32 %r96, %r97, %r65, %r92;


dp4a.s32.s32 %r100, %r101, %r69, %r96;

ld.global.nc.u32 %r234, [%rd42];
and.b32 %r235, %r234, -2004318072;
shr.u32 %r236, %r235, 1;
or.b32 %r237, %r236, 839922192;
and.b32 %r238, %r234, 30583;
prmt.b32 %r239, %r119, %r118, %r238;
prmt.b32 %r240, %r122, %r121, %r238;
and.b32 %r241, %r237, 30292;
prmt.b32 %r242, %r239, %r240, %r241;
shr.u32 %r243, %r234, 16;
and.b32 %r244, %r243, 30583;
prmt.b32 %r245, %r119, %r118, %r244;
prmt.b32 %r246, %r122, %r121, %r244;
shr.u32 %r247, %r237, 16;
prmt.b32 %r248, %r245, %r246, %r247;
prmt.b32 %r105, %r242, %r248, %r132;
prmt.b32 %r109, %r242, %r248, %r133;

dp4a.s32.s32 %r104, %r105, %r73, %r100;


dp4a.s32.s32 %r108, %r109, %r77, %r104;

add.s64 %rd43, %rd51, %rd9;
ld.global.nc.u8 %rs5, [%rd43];
cvt.u32.u16 %r249, %rs5;
and.b32 %r250, %r249, 255;
shr.u32 %r251, %r250, %r6;
and.b32 %r252, %r251, 15;
add.s64 %rd44, %rd51, %rd7;
ld.global.nc.u16 %rs6, [%rd44+2];
cvt.u32.u16 %r253, %rs6;
shr.u32 %r254, %r253, %r7;
and.b32 %r255, %r254, 3;
bfi.b32 %r256, %r255, %r252, 4, 2;
add.s32 %r257, %r256, -32;
mul.lo.s32 %r258, %r257, %r108;
ld.global.nc.u16 %rs2, [%rd44];

{ cvt.f32.f16 %f13, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r79;
cvt.f32.f16 %f14, low;}


mul.ftz.f32 %f17, %f13, %f14;
cvt.rn.f32.s32 %f18, %r258;
fma.rn.ftz.f32 %f42, %f17, %f18, %f42;
add.s64 %rd51, %rd51, 544;
add.s64 %rd50, %rd50, 1152;
add.s32 %r290, %r290, 4;
setp.lt.u32 %p3, %r290, %r2;
@%p3 bra $L__BB292_3;

st.local.v2.f32 [%rd1], {%f43, %f42};

$L__BB292_5:
mov.b32 %r259, %f43;
mov.u32 %r260, 31;
mov.u32 %r261, 16;
mov.u32 %r262, -1;
shfl.sync.bfly.b32 %r263|%p4, %r259, %r261, %r260, %r262;
mov.b32 %f19, %r263;
add.ftz.f32 %f20, %f43, %f19;
mov.b32 %r264, %f20;
mov.u32 %r265, 8;
shfl.sync.bfly.b32 %r266|%p5, %r264, %r265, %r260, %r262;
mov.b32 %f21, %r266;
add.ftz.f32 %f22, %f20, %f21;
mov.b32 %r267, %f22;
mov.u32 %r268, 4;
shfl.sync.bfly.b32 %r269|%p6, %r267, %r268, %r260, %r262;
mov.b32 %f23, %r269;
add.ftz.f32 %f24, %f22, %f23;
mov.b32 %r270, %f24;
mov.u32 %r271, 2;
shfl.sync.bfly.b32 %r272|%p7, %r270, %r271, %r260, %r262;
mov.b32 %f25, %r272;
add.ftz.f32 %f26, %f24, %f25;
mov.b32 %r273, %f26;
mov.u32 %r274, 1;
shfl.sync.bfly.b32 %r275|%p8, %r273, %r274, %r260, %r262;
mov.b32 %f27, %r275;
add.ftz.f32 %f28, %f26, %f27;
st.local.f32 [%rd1], %f28;
mov.b32 %r276, %f42;
shfl.sync.bfly.b32 %r277|%p9, %r276, %r261, %r260, %r262;
mov.b32 %f29, %r277;
add.ftz.f32 %f30, %f42, %f29;
mov.b32 %r278, %f30;
shfl.sync.bfly.b32 %r279|%p10, %r278, %r265, %r260, %r262;
mov.b32 %f31, %r279;
add.ftz.f32 %f32, %f30, %f31;
mov.b32 %r280, %f32;
shfl.sync.bfly.b32 %r281|%p11, %r280, %r268, %r260, %r262;
mov.b32 %f33, %r281;
add.ftz.f32 %f34, %f32, %f33;
mov.b32 %r282, %f34;
shfl.sync.bfly.b32 %r283|%p12, %r282, %r271, %r260, %r262;
mov.b32 %f35, %r283;
add.ftz.f32 %f36, %f34, %f35;
mov.b32 %r284, %f36;
shfl.sync.bfly.b32 %r285|%p13, %r284, %r274, %r260, %r262;
mov.b32 %f37, %r285;
add.ftz.f32 %f38, %f36, %f37;
st.local.f32 [%rd1+4], %f38;
add.s32 %r10, %r1, %r4;
setp.ge.u32 %p14, %r10, %r14;
setp.gt.u32 %p15, %r4, 1;
or.pred %p16, %p15, %p14;
@%p16 bra $L__BB292_7;

mul.wide.u32 %rd45, %r4, 4;
add.s64 %rd46, %rd1, %rd45;
ld.local.f32 %f39, [%rd46];
mov.u32 %r286, %ctaid.y;
mad.lo.s32 %r288, %r3, %r17, %r10;
mad.lo.s32 %r289, %r286, %r20, %r288;
cvta.to.global.u64 %rd47, %rd18;
mul.wide.u32 %rd48, %r289, 4;
add.s64 %rd49, %rd47, %rd48;
st.global.f32 [%rd49], %f39;

$L__BB292_7:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<320>;
.reg .b16 %rs<17>;
.reg .f32 %f<1005>;
.reg .b32 %r<1751>;
.reg .b64 %rd<121>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate[1536];

ld.param.u64 %rd29, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd30, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd31, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u32 %r49, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r53, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r54, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r58, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r59, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r60, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r64, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r65, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r66, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r62, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r61, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r56, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r55, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r52, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r51, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r50, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
ld.param.u64 %rd34, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+16];
ld.param.u64 %rd33, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+8];
ld.param.u64 %rd32, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3];
mov.u32 %r1, %tid.y;
shl.b32 %r67, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r67, %r2;
mov.u32 %r68, %ctaid.x;
shl.b32 %r4, %r68, 2;
mov.u32 %r5, %ctaid.y;
setp.eq.s64 %p1, %rd31, 0;
@%p1 bra $L__BB293_2;

cvta.to.global.u64 %rd38, %rd31;
mul.wide.u32 %rd39, %r5, 4;
add.s64 %rd40, %rd38, %rd39;
ld.global.nc.u32 %r1747, [%rd40];
mul.hi.u32 %r69, %r5, %r50;
add.s32 %r70, %r69, %r5;
shr.u32 %r71, %r70, %r51;
mul.lo.s32 %r72, %r71, %r52;
sub.s32 %r1748, %r5, %r72;
bra.uni $L__BB293_3;

$L__BB293_2:
mul.hi.u32 %r73, %r5, %r55;
add.s32 %r74, %r73, %r5;
shr.u32 %r1747, %r74, %r56;
mov.u32 %r1748, %r5;

$L__BB293_3:
mov.u32 %r11, %ctaid.z;
selp.b32 %r12, %r5, %r1747, %p1;
setp.ne.s32 %p3, %r1, 0;
setp.gt.u32 %p4, %r2, 3;
or.pred %p5, %p4, %p3;
add.s32 %r75, %r4, %r2;
setp.ge.u32 %p6, %r75, %r54;
mov.f32 %f27, 0f00000000;
or.pred %p7, %p5, %p6;
mov.f32 %f4, %f27;
mov.f32 %f5, %f27;
@%p7 bra $L__BB293_8;

setp.eq.s64 %p8, %rd32, 0;
mov.f32 %f4, 0f00000000;
mov.f32 %f5, %f4;
@%p8 bra $L__BB293_6;

cvta.to.global.u64 %rd41, %rd32;
mul.lo.s32 %r76, %r11, %r66;
cvt.u64.u32 %rd42, %r76;
mul.lo.s32 %r77, %r12, %r60;
cvt.u64.u32 %rd43, %r77;
cvt.u64.u32 %rd44, %r2;
cvt.s64.s32 %rd45, %r4;
add.s64 %rd46, %rd45, %rd44;
add.s64 %rd47, %rd46, %rd42;
add.s64 %rd48, %rd47, %rd43;
shl.b64 %rd49, %rd48, 2;
add.s64 %rd50, %rd41, %rd49;
ld.global.f32 %f5, [%rd50];

$L__BB293_6:
setp.eq.s64 %p9, %rd34, 0;
setp.eq.s64 %p10, %rd33, 0;
or.pred %p11, %p9, %p10;
@%p11 bra $L__BB293_8;

cvta.to.global.u64 %rd51, %rd34;
mul.lo.s32 %r78, %r11, %r66;
cvt.u64.u32 %rd52, %r78;
mul.lo.s32 %r79, %r12, %r60;
cvt.u64.u32 %rd53, %r79;
cvt.u64.u32 %rd54, %r2;
cvt.s64.s32 %rd55, %r4;
add.s64 %rd56, %rd55, %rd54;
add.s64 %rd57, %rd56, %rd52;
add.s64 %rd58, %rd57, %rd53;
shl.b64 %rd59, %rd58, 2;
add.s64 %rd60, %rd51, %rd59;
ld.global.f32 %f4, [%rd60];

$L__BB293_8:
mul.lo.s32 %r80, %r11, %r65;
cvt.u64.u32 %rd61, %r80;
mul.lo.s32 %r81, %r1748, %r59;
cvt.u64.u32 %rd62, %r81;
add.s64 %rd4, %rd62, %rd61;
mul.hi.u32 %r82, %r11, %r61;
add.s32 %r83, %r82, %r11;
shr.u32 %r13, %r83, %r62;
shr.s32 %r84, %r3, 31;
shr.u32 %r85, %r84, 29;
add.s32 %r86, %r3, %r85;
shr.s32 %r1750, %r86, 3;
shr.u32 %r87, %r49, 8;
setp.ge.s32 %p12, %r1750, %r87;
mov.f32 %f28, %f27;
mov.f32 %f29, %f27;
mov.f32 %f30, %f27;
mov.f32 %f31, %f27;
mov.f32 %f32, %f27;
mov.f32 %f33, %f27;
mov.f32 %f34, %f27;
@%p12 bra $L__BB293_19;

mul.lo.s32 %r88, %r13, %r64;
mul.lo.s32 %r89, %r1747, %r58;
and.b32 %r93, %r86, -8;
sub.s32 %r15, %r3, %r93;
shl.b32 %r94, %r15, 2;
shr.u32 %r95, %r15, 31;
add.s32 %r96, %r15, %r95;
shr.s32 %r97, %r96, 1;
cvt.s64.s32 %rd63, %r97;
and.b32 %r16, %r94, 4;
shl.b32 %r1749, %r1750, 3;
add.s32 %r98, %r1750, %r88;
add.s32 %r18, %r98, %r89;
add.s32 %r99, %r4, 1;
mad.lo.s32 %r100, %r53, %r99, %r18;
mul.wide.s32 %rd64, %r100, 136;
mad.lo.s32 %r101, %r4, %r53, %r1750;
add.s32 %r102, %r101, %r88;
add.s32 %r19, %r102, %r89;
mul.wide.s32 %rd65, %r19, 136;
add.s64 %rd66, %rd65, %rd63;
add.s64 %rd5, %rd66, 4;
add.s64 %rd67, %rd64, %rd63;
add.s64 %rd6, %rd67, 4;
mul.wide.s32 %rd68, %r94, 4;
add.s64 %rd69, %rd65, %rd68;
add.s64 %rd7, %rd69, 20;
add.s64 %rd70, %rd64, %rd68;
add.s64 %rd8, %rd70, 12;
add.s32 %r103, %r4, 3;
mad.lo.s32 %r104, %r53, %r103, %r18;
mul.wide.s32 %rd71, %r104, 136;
add.s64 %rd72, %rd71, %rd68;
add.s64 %rd9, %rd72, 20;
add.s64 %rd73, %rd71, %rd63;
add.s64 %rd10, %rd73, 4;
add.s32 %r105, %r4, 2;
mad.lo.s32 %r106, %r53, %r105, %r18;
mul.wide.s32 %rd74, %r106, 136;
add.s64 %rd75, %rd74, %rd63;
add.s64 %rd11, %rd75, 4;
mul.wide.s32 %rd76, %r53, 136;
or.b64 %rd12, %rd76, 2;
cvta.to.global.u64 %rd120, %rd33;
cvta.to.global.u64 %rd119, %rd29;
mov.f32 %f27, 0f00000000;
cvta.to.global.u64 %rd77, %rd30;
mov.f32 %f28, %f27;
mov.f32 %f29, %f27;
mov.f32 %f30, %f27;
mov.f32 %f31, %f27;
mov.f32 %f32, %f27;
mov.f32 %f33, %f27;
mov.f32 %f34, %f27;

$L__BB293_10:
cvt.u64.u32 %rd78, %r1749;
add.s64 %rd79, %rd4, %rd78;
cvt.s64.s32 %rd80, %r15;
add.s64 %rd81, %rd79, %rd80;
mul.lo.s64 %rd82, %rd81, 36;
add.s64 %rd83, %rd119, %rd7;
ld.global.nc.u32 %r140, [%rd83+-12];
and.b32 %r141, %r140, -2004318072;
shr.u32 %r142, %r141, 1;
or.b32 %r143, %r142, 839922192;
and.b32 %r144, %r140, 30583;
mov.u32 %r145, -152379953;
mov.u32 %r146, -1079142271;
prmt.b32 %r147, %r146, %r145, %r144;
mov.u32 %r148, 1901675829;
mov.u32 %r149, 639175937;
prmt.b32 %r150, %r149, %r148, %r144;
and.b32 %r151, %r143, 30292;
prmt.b32 %r152, %r147, %r150, %r151;
shr.u32 %r153, %r140, 16;
and.b32 %r154, %r153, 30583;
prmt.b32 %r155, %r146, %r145, %r154;
prmt.b32 %r156, %r149, %r148, %r154;
shr.u32 %r157, %r143, 16;
prmt.b32 %r158, %r155, %r156, %r157;
mov.u32 %r159, 25632;
prmt.b32 %r108, %r152, %r158, %r159;
mov.u32 %r160, 30001;
prmt.b32 %r112, %r152, %r158, %r160;
add.s64 %rd84, %rd77, %rd82;
ld.global.nc.u32 %r22, [%rd84+4];
ld.global.nc.u32 %r23, [%rd84+20];
mov.u32 %r110, 0;

dp4a.s32.s32 %r107, %r108, %r22, %r110;


dp4a.s32.s32 %r111, %r112, %r23, %r107;

ld.global.nc.u32 %r161, [%rd83+-8];
and.b32 %r162, %r161, -2004318072;
shr.u32 %r163, %r162, 1;
or.b32 %r164, %r163, 839922192;
and.b32 %r165, %r161, 30583;
prmt.b32 %r166, %r146, %r145, %r165;
prmt.b32 %r167, %r149, %r148, %r165;
and.b32 %r168, %r164, 30292;
prmt.b32 %r169, %r166, %r167, %r168;
shr.u32 %r170, %r161, 16;
and.b32 %r171, %r170, 30583;
prmt.b32 %r172, %r146, %r145, %r171;
prmt.b32 %r173, %r149, %r148, %r171;
shr.u32 %r174, %r164, 16;
prmt.b32 %r175, %r172, %r173, %r174;
prmt.b32 %r116, %r169, %r175, %r159;
prmt.b32 %r120, %r169, %r175, %r160;
ld.global.nc.u32 %r24, [%rd84+8];
ld.global.nc.u32 %r25, [%rd84+24];

dp4a.s32.s32 %r115, %r116, %r24, %r111;


dp4a.s32.s32 %r119, %r120, %r25, %r115;

ld.global.nc.u32 %r176, [%rd83+-4];
and.b32 %r177, %r176, -2004318072;
shr.u32 %r178, %r177, 1;
or.b32 %r179, %r178, 839922192;
and.b32 %r180, %r176, 30583;
prmt.b32 %r181, %r146, %r145, %r180;
prmt.b32 %r182, %r149, %r148, %r180;
and.b32 %r183, %r179, 30292;
prmt.b32 %r184, %r181, %r182, %r183;
shr.u32 %r185, %r176, 16;
and.b32 %r186, %r185, 30583;
prmt.b32 %r187, %r146, %r145, %r186;
prmt.b32 %r188, %r149, %r148, %r186;
shr.u32 %r189, %r179, 16;
prmt.b32 %r190, %r187, %r188, %r189;
prmt.b32 %r124, %r184, %r190, %r159;
prmt.b32 %r128, %r184, %r190, %r160;
ld.global.nc.u32 %r26, [%rd84+12];
ld.global.nc.u32 %r27, [%rd84+28];

dp4a.s32.s32 %r123, %r124, %r26, %r119;


dp4a.s32.s32 %r127, %r128, %r27, %r123;

ld.global.nc.u32 %r191, [%rd83];
and.b32 %r192, %r191, -2004318072;
shr.u32 %r193, %r192, 1;
or.b32 %r194, %r193, 839922192;
and.b32 %r195, %r191, 30583;
prmt.b32 %r196, %r146, %r145, %r195;
prmt.b32 %r197, %r149, %r148, %r195;
and.b32 %r198, %r194, 30292;
prmt.b32 %r199, %r196, %r197, %r198;
shr.u32 %r200, %r191, 16;
and.b32 %r201, %r200, 30583;
prmt.b32 %r202, %r146, %r145, %r201;
prmt.b32 %r203, %r149, %r148, %r201;
shr.u32 %r204, %r194, 16;
prmt.b32 %r205, %r202, %r203, %r204;
prmt.b32 %r132, %r199, %r205, %r159;
prmt.b32 %r136, %r199, %r205, %r160;
ld.global.nc.u32 %r28, [%rd84+16];
ld.global.nc.u32 %r29, [%rd84+32];

dp4a.s32.s32 %r131, %r132, %r28, %r127;


dp4a.s32.s32 %r135, %r136, %r29, %r131;

add.s64 %rd18, %rd119, %rd5;
ld.global.nc.u8 %rs2, [%rd18];
cvt.u32.u16 %r206, %rs2;
and.b32 %r207, %r206, 255;
shr.u32 %r208, %r207, %r16;
and.b32 %r209, %r208, 15;
add.s64 %rd19, %rd119, %rd65;
ld.global.nc.u16 %rs3, [%rd19+2];
cvt.u32.u16 %r210, %rs3;
shl.b32 %r211, %r15, 1;
shr.u32 %r212, %r210, %r211;
and.b32 %r213, %r212, 3;
bfi.b32 %r214, %r213, %r209, 4, 2;
add.s32 %r215, %r214, -32;
mul.lo.s32 %r216, %r215, %r135;
ld.global.nc.u16 %rs1, [%rd19];

{ cvt.f32.f16 %f101, %rs1;}


ld.global.nc.u32 %r30, [%rd84];

{.reg .f16 low,high;
mov.b32 {low,high},%r30;
cvt.f32.f16 %f102, low;}


mul.ftz.f32 %f103, %f101, %f102;
cvt.rn.f32.s32 %f104, %r216;
fma.rn.ftz.f32 %f30, %f103, %f104, %f30;
setp.eq.s64 %p13, %rd33, 0;
@%p13 bra $L__BB293_12;

shl.b32 %r1746, %r15, 1;
add.s64 %rd86, %rd120, %rd7;
ld.global.u32 %r250, [%rd86+-12];
and.b32 %r251, %r250, -2004318072;
shr.u32 %r252, %r251, 1;
or.b32 %r253, %r252, 839922192;
and.b32 %r254, %r250, 30583;
prmt.b32 %r257, %r146, %r145, %r254;
prmt.b32 %r260, %r149, %r148, %r254;
and.b32 %r261, %r253, 30292;
prmt.b32 %r262, %r257, %r260, %r261;
shr.u32 %r263, %r250, 16;
and.b32 %r264, %r263, 30583;
prmt.b32 %r265, %r146, %r145, %r264;
prmt.b32 %r266, %r149, %r148, %r264;
shr.u32 %r267, %r253, 16;
prmt.b32 %r268, %r265, %r266, %r267;
prmt.b32 %r218, %r262, %r268, %r159;
prmt.b32 %r222, %r262, %r268, %r160;

dp4a.s32.s32 %r217, %r218, %r22, %r110;


dp4a.s32.s32 %r221, %r222, %r23, %r217;

ld.global.u32 %r271, [%rd86+-8];
and.b32 %r272, %r271, -2004318072;
shr.u32 %r273, %r272, 1;
or.b32 %r274, %r273, 839922192;
and.b32 %r275, %r271, 30583;
prmt.b32 %r276, %r146, %r145, %r275;
prmt.b32 %r277, %r149, %r148, %r275;
and.b32 %r278, %r274, 30292;
prmt.b32 %r279, %r276, %r277, %r278;
shr.u32 %r280, %r271, 16;
and.b32 %r281, %r280, 30583;
prmt.b32 %r282, %r146, %r145, %r281;
prmt.b32 %r283, %r149, %r148, %r281;
shr.u32 %r284, %r274, 16;
prmt.b32 %r285, %r282, %r283, %r284;
prmt.b32 %r226, %r279, %r285, %r159;
prmt.b32 %r230, %r279, %r285, %r160;

dp4a.s32.s32 %r225, %r226, %r24, %r221;


dp4a.s32.s32 %r229, %r230, %r25, %r225;

ld.global.u32 %r286, [%rd86+-4];
and.b32 %r287, %r286, -2004318072;
shr.u32 %r288, %r287, 1;
or.b32 %r289, %r288, 839922192;
and.b32 %r290, %r286, 30583;
prmt.b32 %r291, %r146, %r145, %r290;
prmt.b32 %r292, %r149, %r148, %r290;
and.b32 %r293, %r289, 30292;
prmt.b32 %r294, %r291, %r292, %r293;
shr.u32 %r295, %r286, 16;
and.b32 %r296, %r295, 30583;
prmt.b32 %r297, %r146, %r145, %r296;
prmt.b32 %r298, %r149, %r148, %r296;
shr.u32 %r299, %r289, 16;
prmt.b32 %r300, %r297, %r298, %r299;
prmt.b32 %r234, %r294, %r300, %r159;
prmt.b32 %r238, %r294, %r300, %r160;

dp4a.s32.s32 %r233, %r234, %r26, %r229;


dp4a.s32.s32 %r237, %r238, %r27, %r233;

ld.global.u32 %r301, [%rd86];
and.b32 %r302, %r301, -2004318072;
shr.u32 %r303, %r302, 1;
or.b32 %r304, %r303, 839922192;
and.b32 %r305, %r301, 30583;
prmt.b32 %r306, %r146, %r145, %r305;
prmt.b32 %r307, %r149, %r148, %r305;
and.b32 %r308, %r304, 30292;
prmt.b32 %r309, %r306, %r307, %r308;
shr.u32 %r310, %r301, 16;
and.b32 %r311, %r310, 30583;
prmt.b32 %r312, %r146, %r145, %r311;
prmt.b32 %r313, %r149, %r148, %r311;
shr.u32 %r314, %r304, 16;
prmt.b32 %r315, %r312, %r313, %r314;
prmt.b32 %r242, %r309, %r315, %r159;
prmt.b32 %r246, %r309, %r315, %r160;

dp4a.s32.s32 %r241, %r242, %r28, %r237;


dp4a.s32.s32 %r245, %r246, %r29, %r241;

add.s64 %rd87, %rd120, %rd5;
ld.global.u8 %r316, [%rd87];
shr.u32 %r317, %r316, %r16;
and.b32 %r318, %r317, 15;
add.s64 %rd89, %rd120, %rd65;
ld.global.u16 %r319, [%rd89+2];
shr.u32 %r321, %r319, %r1746;
and.b32 %r322, %r321, 3;
bfi.b32 %r323, %r322, %r318, 4, 2;
add.s32 %r324, %r323, -32;
mul.lo.s32 %r325, %r324, %r245;
ld.global.u16 %rs4, [%rd89];

{ cvt.f32.f16 %f105, %rs4;}



{.reg .f16 low,high;
mov.b32 {low,high},%r30;
cvt.f32.f16 %f106, low;}


mul.ftz.f32 %f107, %f105, %f106;
cvt.rn.f32.s32 %f108, %r325;
fma.rn.ftz.f32 %f34, %f107, %f108, %f34;

$L__BB293_12:
add.s64 %rd118, %rd119, %rd7;
add.s64 %rd117, %rd118, -12;
shl.b32 %r1740, %r15, 1;
add.s64 %rd20, %rd117, %rd76;
ld.global.nc.u32 %r359, [%rd20];
and.b32 %r360, %r359, -2004318072;
shr.u32 %r361, %r360, 1;
or.b32 %r362, %r361, 839922192;
and.b32 %r363, %r359, 30583;
prmt.b32 %r366, %r146, %r145, %r363;
prmt.b32 %r369, %r149, %r148, %r363;
and.b32 %r370, %r362, 30292;
prmt.b32 %r371, %r366, %r369, %r370;
shr.u32 %r372, %r359, 16;
and.b32 %r373, %r372, 30583;
prmt.b32 %r374, %r146, %r145, %r373;
prmt.b32 %r375, %r149, %r148, %r373;
shr.u32 %r376, %r362, 16;
prmt.b32 %r377, %r374, %r375, %r376;
prmt.b32 %r327, %r371, %r377, %r159;
prmt.b32 %r331, %r371, %r377, %r160;

dp4a.s32.s32 %r326, %r327, %r22, %r110;


dp4a.s32.s32 %r330, %r331, %r23, %r326;

ld.global.nc.u32 %r380, [%rd20+4];
and.b32 %r381, %r380, -2004318072;
shr.u32 %r382, %r381, 1;
or.b32 %r383, %r382, 839922192;
and.b32 %r384, %r380, 30583;
prmt.b32 %r385, %r146, %r145, %r384;
prmt.b32 %r386, %r149, %r148, %r384;
and.b32 %r387, %r383, 30292;
prmt.b32 %r388, %r385, %r386, %r387;
shr.u32 %r389, %r380, 16;
and.b32 %r390, %r389, 30583;
prmt.b32 %r391, %r146, %r145, %r390;
prmt.b32 %r392, %r149, %r148, %r390;
shr.u32 %r393, %r383, 16;
prmt.b32 %r394, %r391, %r392, %r393;
prmt.b32 %r335, %r388, %r394, %r159;
prmt.b32 %r339, %r388, %r394, %r160;

dp4a.s32.s32 %r334, %r335, %r24, %r330;


dp4a.s32.s32 %r338, %r339, %r25, %r334;

ld.global.nc.u32 %r395, [%rd20+8];
and.b32 %r396, %r395, -2004318072;
shr.u32 %r397, %r396, 1;
or.b32 %r398, %r397, 839922192;
and.b32 %r399, %r395, 30583;
prmt.b32 %r400, %r146, %r145, %r399;
prmt.b32 %r401, %r149, %r148, %r399;
and.b32 %r402, %r398, 30292;
prmt.b32 %r403, %r400, %r401, %r402;
shr.u32 %r404, %r395, 16;
and.b32 %r405, %r404, 30583;
prmt.b32 %r406, %r146, %r145, %r405;
prmt.b32 %r407, %r149, %r148, %r405;
shr.u32 %r408, %r398, 16;
prmt.b32 %r409, %r406, %r407, %r408;
prmt.b32 %r343, %r403, %r409, %r159;
prmt.b32 %r347, %r403, %r409, %r160;

dp4a.s32.s32 %r342, %r343, %r26, %r338;


dp4a.s32.s32 %r346, %r347, %r27, %r342;

ld.global.nc.u32 %r410, [%rd20+12];
and.b32 %r411, %r410, -2004318072;
shr.u32 %r412, %r411, 1;
or.b32 %r413, %r412, 839922192;
and.b32 %r414, %r410, 30583;
prmt.b32 %r415, %r146, %r145, %r414;
prmt.b32 %r416, %r149, %r148, %r414;
and.b32 %r417, %r413, 30292;
prmt.b32 %r418, %r415, %r416, %r417;
shr.u32 %r419, %r410, 16;
and.b32 %r420, %r419, 30583;
prmt.b32 %r421, %r146, %r145, %r420;
prmt.b32 %r422, %r149, %r148, %r420;
shr.u32 %r423, %r413, 16;
prmt.b32 %r424, %r421, %r422, %r423;
prmt.b32 %r351, %r418, %r424, %r159;
prmt.b32 %r355, %r418, %r424, %r160;

dp4a.s32.s32 %r350, %r351, %r28, %r346;


dp4a.s32.s32 %r354, %r355, %r29, %r350;

add.s64 %rd21, %rd18, %rd76;
ld.global.nc.u8 %rs6, [%rd21];
cvt.u32.u16 %r425, %rs6;
and.b32 %r426, %r425, 255;
shr.u32 %r427, %r426, %r16;
and.b32 %r428, %r427, 15;
add.s64 %rd22, %rd19, %rd76;
ld.global.nc.u16 %rs7, [%rd22+2];
cvt.u32.u16 %r429, %rs7;
shr.u32 %r431, %r429, %r1740;
and.b32 %r432, %r431, 3;
bfi.b32 %r433, %r432, %r428, 4, 2;
add.s32 %r434, %r433, -32;
mul.lo.s32 %r435, %r434, %r354;
ld.global.nc.u16 %rs5, [%rd22];

{ cvt.f32.f16 %f109, %rs5;}



{.reg .f16 low,high;
mov.b32 {low,high},%r30;
cvt.f32.f16 %f110, low;}


mul.ftz.f32 %f111, %f109, %f110;
cvt.rn.f32.s32 %f112, %r435;
fma.rn.ftz.f32 %f29, %f111, %f112, %f29;
@%p13 bra $L__BB293_14;

shl.b32 %r1745, %r15, 1;
add.s64 %rd91, %rd120, %rd8;
ld.global.u32 %r469, [%rd91+-4];
and.b32 %r470, %r469, -2004318072;
shr.u32 %r471, %r470, 1;
or.b32 %r472, %r471, 839922192;
and.b32 %r473, %r469, 30583;
mov.u32 %r474, -152379953;
mov.u32 %r475, -1079142271;
prmt.b32 %r476, %r475, %r474, %r473;
mov.u32 %r477, 1901675829;
mov.u32 %r478, 639175937;
prmt.b32 %r479, %r478, %r477, %r473;
and.b32 %r480, %r472, 30292;
prmt.b32 %r481, %r476, %r479, %r480;
shr.u32 %r482, %r469, 16;
and.b32 %r483, %r482, 30583;
prmt.b32 %r484, %r475, %r474, %r483;
prmt.b32 %r485, %r478, %r477, %r483;
shr.u32 %r486, %r472, 16;
prmt.b32 %r487, %r484, %r485, %r486;
mov.u32 %r488, 25632;
prmt.b32 %r437, %r481, %r487, %r488;
mov.u32 %r489, 30001;
prmt.b32 %r441, %r481, %r487, %r489;
mov.u32 %r439, 0;

dp4a.s32.s32 %r436, %r437, %r22, %r439;


dp4a.s32.s32 %r440, %r441, %r23, %r436;

ld.global.u32 %r490, [%rd91];
and.b32 %r491, %r490, -2004318072;
shr.u32 %r492, %r491, 1;
or.b32 %r493, %r492, 839922192;
and.b32 %r494, %r490, 30583;
prmt.b32 %r495, %r475, %r474, %r494;
prmt.b32 %r496, %r478, %r477, %r494;
and.b32 %r497, %r493, 30292;
prmt.b32 %r498, %r495, %r496, %r497;
shr.u32 %r499, %r490, 16;
and.b32 %r500, %r499, 30583;
prmt.b32 %r501, %r475, %r474, %r500;
prmt.b32 %r502, %r478, %r477, %r500;
shr.u32 %r503, %r493, 16;
prmt.b32 %r504, %r501, %r502, %r503;
prmt.b32 %r445, %r498, %r504, %r488;
prmt.b32 %r449, %r498, %r504, %r489;

dp4a.s32.s32 %r444, %r445, %r24, %r440;


dp4a.s32.s32 %r448, %r449, %r25, %r444;

ld.global.u32 %r505, [%rd91+4];
and.b32 %r506, %r505, -2004318072;
shr.u32 %r507, %r506, 1;
or.b32 %r508, %r507, 839922192;
and.b32 %r509, %r505, 30583;
prmt.b32 %r510, %r475, %r474, %r509;
prmt.b32 %r511, %r478, %r477, %r509;
and.b32 %r512, %r508, 30292;
prmt.b32 %r513, %r510, %r511, %r512;
shr.u32 %r514, %r505, 16;
and.b32 %r515, %r514, 30583;
prmt.b32 %r516, %r475, %r474, %r515;
prmt.b32 %r517, %r478, %r477, %r515;
shr.u32 %r518, %r508, 16;
prmt.b32 %r519, %r516, %r517, %r518;
prmt.b32 %r453, %r513, %r519, %r488;
prmt.b32 %r457, %r513, %r519, %r489;

dp4a.s32.s32 %r452, %r453, %r26, %r448;


dp4a.s32.s32 %r456, %r457, %r27, %r452;

ld.global.u32 %r520, [%rd91+8];
and.b32 %r521, %r520, -2004318072;
shr.u32 %r522, %r521, 1;
or.b32 %r523, %r522, 839922192;
and.b32 %r524, %r520, 30583;
prmt.b32 %r525, %r475, %r474, %r524;
prmt.b32 %r526, %r478, %r477, %r524;
and.b32 %r527, %r523, 30292;
prmt.b32 %r528, %r525, %r526, %r527;
shr.u32 %r529, %r520, 16;
and.b32 %r530, %r529, 30583;
prmt.b32 %r531, %r475, %r474, %r530;
prmt.b32 %r532, %r478, %r477, %r530;
shr.u32 %r533, %r523, 16;
prmt.b32 %r534, %r531, %r532, %r533;
prmt.b32 %r461, %r528, %r534, %r488;
prmt.b32 %r465, %r528, %r534, %r489;

dp4a.s32.s32 %r460, %r461, %r28, %r456;


dp4a.s32.s32 %r464, %r465, %r29, %r460;

add.s64 %rd92, %rd120, %rd6;
ld.global.u8 %r535, [%rd92];
shr.u32 %r536, %r535, %r16;
and.b32 %r537, %r536, 15;
or.b32 %r540, %r4, 1;
mad.lo.s32 %r541, %r53, %r540, %r18;
mul.wide.s32 %rd93, %r541, 136;
add.s64 %rd94, %rd120, %rd93;
ld.global.u16 %r542, [%rd94+2];
shr.u32 %r544, %r542, %r1745;
and.b32 %r545, %r544, 3;
bfi.b32 %r546, %r545, %r537, 4, 2;
add.s32 %r547, %r546, -32;
mul.lo.s32 %r548, %r547, %r464;
ld.global.u16 %rs8, [%rd94];

{ cvt.f32.f16 %f113, %rs8;}



{.reg .f16 low,high;
mov.b32 {low,high},%r30;
cvt.f32.f16 %f114, low;}


mul.ftz.f32 %f115, %f113, %f114;
cvt.rn.f32.s32 %f116, %r548;
fma.rn.ftz.f32 %f33, %f115, %f116, %f33;

$L__BB293_14:
shl.b32 %r1741, %r15, 1;
add.s64 %rd23, %rd20, %rd76;
ld.global.nc.u32 %r582, [%rd23];
and.b32 %r583, %r582, -2004318072;
shr.u32 %r584, %r583, 1;
or.b32 %r585, %r584, 839922192;
and.b32 %r586, %r582, 30583;
mov.u32 %r587, -152379953;
mov.u32 %r588, -1079142271;
prmt.b32 %r589, %r588, %r587, %r586;
mov.u32 %r590, 1901675829;
mov.u32 %r591, 639175937;
prmt.b32 %r592, %r591, %r590, %r586;
and.b32 %r593, %r585, 30292;
prmt.b32 %r594, %r589, %r592, %r593;
shr.u32 %r595, %r582, 16;
and.b32 %r596, %r595, 30583;
prmt.b32 %r597, %r588, %r587, %r596;
prmt.b32 %r598, %r591, %r590, %r596;
shr.u32 %r599, %r585, 16;
prmt.b32 %r600, %r597, %r598, %r599;
mov.u32 %r601, 25632;
prmt.b32 %r550, %r594, %r600, %r601;
mov.u32 %r602, 30001;
prmt.b32 %r554, %r594, %r600, %r602;
mov.u32 %r552, 0;

dp4a.s32.s32 %r549, %r550, %r22, %r552;


dp4a.s32.s32 %r553, %r554, %r23, %r549;

ld.global.nc.u32 %r603, [%rd23+4];
and.b32 %r604, %r603, -2004318072;
shr.u32 %r605, %r604, 1;
or.b32 %r606, %r605, 839922192;
and.b32 %r607, %r603, 30583;
prmt.b32 %r608, %r588, %r587, %r607;
prmt.b32 %r609, %r591, %r590, %r607;
and.b32 %r610, %r606, 30292;
prmt.b32 %r611, %r608, %r609, %r610;
shr.u32 %r612, %r603, 16;
and.b32 %r613, %r612, 30583;
prmt.b32 %r614, %r588, %r587, %r613;
prmt.b32 %r615, %r591, %r590, %r613;
shr.u32 %r616, %r606, 16;
prmt.b32 %r617, %r614, %r615, %r616;
prmt.b32 %r558, %r611, %r617, %r601;
prmt.b32 %r562, %r611, %r617, %r602;

dp4a.s32.s32 %r557, %r558, %r24, %r553;


dp4a.s32.s32 %r561, %r562, %r25, %r557;

ld.global.nc.u32 %r618, [%rd23+8];
and.b32 %r619, %r618, -2004318072;
shr.u32 %r620, %r619, 1;
or.b32 %r621, %r620, 839922192;
and.b32 %r622, %r618, 30583;
prmt.b32 %r623, %r588, %r587, %r622;
prmt.b32 %r624, %r591, %r590, %r622;
and.b32 %r625, %r621, 30292;
prmt.b32 %r626, %r623, %r624, %r625;
shr.u32 %r627, %r618, 16;
and.b32 %r628, %r627, 30583;
prmt.b32 %r629, %r588, %r587, %r628;
prmt.b32 %r630, %r591, %r590, %r628;
shr.u32 %r631, %r621, 16;
prmt.b32 %r632, %r629, %r630, %r631;
prmt.b32 %r566, %r626, %r632, %r601;
prmt.b32 %r570, %r626, %r632, %r602;

dp4a.s32.s32 %r565, %r566, %r26, %r561;


dp4a.s32.s32 %r569, %r570, %r27, %r565;

ld.global.nc.u32 %r633, [%rd23+12];
and.b32 %r634, %r633, -2004318072;
shr.u32 %r635, %r634, 1;
or.b32 %r636, %r635, 839922192;
and.b32 %r637, %r633, 30583;
prmt.b32 %r638, %r588, %r587, %r637;
prmt.b32 %r639, %r591, %r590, %r637;
and.b32 %r640, %r636, 30292;
prmt.b32 %r641, %r638, %r639, %r640;
shr.u32 %r642, %r633, 16;
and.b32 %r643, %r642, 30583;
prmt.b32 %r644, %r588, %r587, %r643;
prmt.b32 %r645, %r591, %r590, %r643;
shr.u32 %r646, %r636, 16;
prmt.b32 %r647, %r644, %r645, %r646;
prmt.b32 %r574, %r641, %r647, %r601;
prmt.b32 %r578, %r641, %r647, %r602;

dp4a.s32.s32 %r573, %r574, %r28, %r569;


dp4a.s32.s32 %r577, %r578, %r29, %r573;

add.s64 %rd24, %rd21, %rd76;
ld.global.nc.u8 %rs10, [%rd24];
cvt.u32.u16 %r648, %rs10;
and.b32 %r649, %r648, 255;
shr.u32 %r650, %r649, %r16;
and.b32 %r651, %r650, 15;
add.s64 %rd96, %rd22, %rd12;
ld.global.nc.u16 %rs11, [%rd96];
cvt.u32.u16 %r652, %rs11;
shr.u32 %r654, %r652, %r1741;
and.b32 %r655, %r654, 3;
bfi.b32 %r656, %r655, %r651, 4, 2;
add.s32 %r657, %r656, -32;
mul.lo.s32 %r658, %r657, %r577;
ld.global.nc.u16 %rs9, [%rd96+-2];

{ cvt.f32.f16 %f117, %rs9;}



{.reg .f16 low,high;
mov.b32 {low,high},%r30;
cvt.f32.f16 %f118, low;}


mul.ftz.f32 %f119, %f117, %f118;
cvt.rn.f32.s32 %f120, %r658;
fma.rn.ftz.f32 %f28, %f119, %f120, %f28;
@%p13 bra $L__BB293_16;

shl.b32 %r1744, %r15, 1;
or.b32 %r694, %r4, 2;
mad.lo.s32 %r695, %r53, %r694, %r18;
mul.wide.s32 %rd97, %r695, 136;
add.s64 %rd99, %rd97, %rd68;
add.s64 %rd100, %rd120, %rd99;
ld.global.u32 %r697, [%rd100+8];
and.b32 %r698, %r697, -2004318072;
shr.u32 %r699, %r698, 1;
or.b32 %r700, %r699, 839922192;
and.b32 %r701, %r697, 30583;
prmt.b32 %r704, %r588, %r587, %r701;
prmt.b32 %r707, %r591, %r590, %r701;
and.b32 %r708, %r700, 30292;
prmt.b32 %r709, %r704, %r707, %r708;
shr.u32 %r710, %r697, 16;
and.b32 %r711, %r710, 30583;
prmt.b32 %r712, %r588, %r587, %r711;
prmt.b32 %r713, %r591, %r590, %r711;
shr.u32 %r714, %r700, 16;
prmt.b32 %r715, %r712, %r713, %r714;
prmt.b32 %r660, %r709, %r715, %r601;
prmt.b32 %r664, %r709, %r715, %r602;

dp4a.s32.s32 %r659, %r660, %r22, %r552;


dp4a.s32.s32 %r663, %r664, %r23, %r659;

ld.global.u32 %r718, [%rd100+12];
and.b32 %r719, %r718, -2004318072;
shr.u32 %r720, %r719, 1;
or.b32 %r721, %r720, 839922192;
and.b32 %r722, %r718, 30583;
prmt.b32 %r723, %r588, %r587, %r722;
prmt.b32 %r724, %r591, %r590, %r722;
and.b32 %r725, %r721, 30292;
prmt.b32 %r726, %r723, %r724, %r725;
shr.u32 %r727, %r718, 16;
and.b32 %r728, %r727, 30583;
prmt.b32 %r729, %r588, %r587, %r728;
prmt.b32 %r730, %r591, %r590, %r728;
shr.u32 %r731, %r721, 16;
prmt.b32 %r732, %r729, %r730, %r731;
prmt.b32 %r668, %r726, %r732, %r601;
prmt.b32 %r672, %r726, %r732, %r602;

dp4a.s32.s32 %r667, %r668, %r24, %r663;


dp4a.s32.s32 %r671, %r672, %r25, %r667;

ld.global.u32 %r733, [%rd100+16];
and.b32 %r734, %r733, -2004318072;
shr.u32 %r735, %r734, 1;
or.b32 %r736, %r735, 839922192;
and.b32 %r737, %r733, 30583;
prmt.b32 %r738, %r588, %r587, %r737;
prmt.b32 %r739, %r591, %r590, %r737;
and.b32 %r740, %r736, 30292;
prmt.b32 %r741, %r738, %r739, %r740;
shr.u32 %r742, %r733, 16;
and.b32 %r743, %r742, 30583;
prmt.b32 %r744, %r588, %r587, %r743;
prmt.b32 %r745, %r591, %r590, %r743;
shr.u32 %r746, %r736, 16;
prmt.b32 %r747, %r744, %r745, %r746;
prmt.b32 %r676, %r741, %r747, %r601;
prmt.b32 %r680, %r741, %r747, %r602;

dp4a.s32.s32 %r675, %r676, %r26, %r671;


dp4a.s32.s32 %r679, %r680, %r27, %r675;

ld.global.u32 %r748, [%rd100+20];
and.b32 %r749, %r748, -2004318072;
shr.u32 %r750, %r749, 1;
or.b32 %r751, %r750, 839922192;
and.b32 %r752, %r748, 30583;
prmt.b32 %r753, %r588, %r587, %r752;
prmt.b32 %r754, %r591, %r590, %r752;
and.b32 %r755, %r751, 30292;
prmt.b32 %r756, %r753, %r754, %r755;
shr.u32 %r757, %r748, 16;
and.b32 %r758, %r757, 30583;
prmt.b32 %r759, %r588, %r587, %r758;
prmt.b32 %r760, %r591, %r590, %r758;
shr.u32 %r761, %r751, 16;
prmt.b32 %r762, %r759, %r760, %r761;
prmt.b32 %r684, %r756, %r762, %r601;
prmt.b32 %r688, %r756, %r762, %r602;

dp4a.s32.s32 %r683, %r684, %r28, %r679;


dp4a.s32.s32 %r687, %r688, %r29, %r683;

add.s64 %rd101, %rd120, %rd11;
ld.global.u8 %r763, [%rd101];
shr.u32 %r764, %r763, %r16;
and.b32 %r765, %r764, 15;
add.s64 %rd102, %rd120, %rd97;
ld.global.u16 %r766, [%rd102+2];
shr.u32 %r768, %r766, %r1744;
and.b32 %r769, %r768, 3;
bfi.b32 %r770, %r769, %r765, 4, 2;
add.s32 %r771, %r770, -32;
mul.lo.s32 %r772, %r771, %r687;
ld.global.u16 %rs12, [%rd102];

{ cvt.f32.f16 %f121, %rs12;}



{.reg .f16 low,high;
mov.b32 {low,high},%r30;
cvt.f32.f16 %f122, low;}


mul.ftz.f32 %f123, %f121, %f122;
cvt.rn.f32.s32 %f124, %r772;
fma.rn.ftz.f32 %f32, %f123, %f124, %f32;

$L__BB293_16:
shl.b32 %r1742, %r15, 1;
add.s64 %rd115, %rd119, %rd65;
add.s64 %rd114, %rd115, %rd76;
add.s64 %rd113, %rd114, %rd76;
add.s64 %rd104, %rd23, %rd76;
ld.global.nc.u32 %r806, [%rd104];
and.b32 %r807, %r806, -2004318072;
shr.u32 %r808, %r807, 1;
or.b32 %r809, %r808, 839922192;
and.b32 %r810, %r806, 30583;
prmt.b32 %r813, %r588, %r587, %r810;
prmt.b32 %r816, %r591, %r590, %r810;
and.b32 %r817, %r809, 30292;
prmt.b32 %r818, %r813, %r816, %r817;
shr.u32 %r819, %r806, 16;
and.b32 %r820, %r819, 30583;
prmt.b32 %r821, %r588, %r587, %r820;
prmt.b32 %r822, %r591, %r590, %r820;
shr.u32 %r823, %r809, 16;
prmt.b32 %r824, %r821, %r822, %r823;
prmt.b32 %r774, %r818, %r824, %r601;
prmt.b32 %r778, %r818, %r824, %r602;

dp4a.s32.s32 %r773, %r774, %r22, %r552;


dp4a.s32.s32 %r777, %r778, %r23, %r773;

ld.global.nc.u32 %r827, [%rd104+4];
and.b32 %r828, %r827, -2004318072;
shr.u32 %r829, %r828, 1;
or.b32 %r830, %r829, 839922192;
and.b32 %r831, %r827, 30583;
prmt.b32 %r832, %r588, %r587, %r831;
prmt.b32 %r833, %r591, %r590, %r831;
and.b32 %r834, %r830, 30292;
prmt.b32 %r835, %r832, %r833, %r834;
shr.u32 %r836, %r827, 16;
and.b32 %r837, %r836, 30583;
prmt.b32 %r838, %r588, %r587, %r837;
prmt.b32 %r839, %r591, %r590, %r837;
shr.u32 %r840, %r830, 16;
prmt.b32 %r841, %r838, %r839, %r840;
prmt.b32 %r782, %r835, %r841, %r601;
prmt.b32 %r786, %r835, %r841, %r602;

dp4a.s32.s32 %r781, %r782, %r24, %r777;


dp4a.s32.s32 %r785, %r786, %r25, %r781;

ld.global.nc.u32 %r842, [%rd104+8];
and.b32 %r843, %r842, -2004318072;
shr.u32 %r844, %r843, 1;
or.b32 %r845, %r844, 839922192;
and.b32 %r846, %r842, 30583;
prmt.b32 %r847, %r588, %r587, %r846;
prmt.b32 %r848, %r591, %r590, %r846;
and.b32 %r849, %r845, 30292;
prmt.b32 %r850, %r847, %r848, %r849;
shr.u32 %r851, %r842, 16;
and.b32 %r852, %r851, 30583;
prmt.b32 %r853, %r588, %r587, %r852;
prmt.b32 %r854, %r591, %r590, %r852;
shr.u32 %r855, %r845, 16;
prmt.b32 %r856, %r853, %r854, %r855;
prmt.b32 %r790, %r850, %r856, %r601;
prmt.b32 %r794, %r850, %r856, %r602;

dp4a.s32.s32 %r789, %r790, %r26, %r785;


dp4a.s32.s32 %r793, %r794, %r27, %r789;

ld.global.nc.u32 %r857, [%rd104+12];
and.b32 %r858, %r857, -2004318072;
shr.u32 %r859, %r858, 1;
or.b32 %r860, %r859, 839922192;
and.b32 %r861, %r857, 30583;
prmt.b32 %r862, %r588, %r587, %r861;
prmt.b32 %r863, %r591, %r590, %r861;
and.b32 %r864, %r860, 30292;
prmt.b32 %r865, %r862, %r863, %r864;
shr.u32 %r866, %r857, 16;
and.b32 %r867, %r866, 30583;
prmt.b32 %r868, %r588, %r587, %r867;
prmt.b32 %r869, %r591, %r590, %r867;
shr.u32 %r870, %r860, 16;
prmt.b32 %r871, %r868, %r869, %r870;
prmt.b32 %r798, %r865, %r871, %r601;
prmt.b32 %r802, %r865, %r871, %r602;

dp4a.s32.s32 %r797, %r798, %r28, %r793;


dp4a.s32.s32 %r801, %r802, %r29, %r797;

add.s64 %rd105, %rd24, %rd76;
ld.global.nc.u8 %rs14, [%rd105];
cvt.u32.u16 %r872, %rs14;
and.b32 %r873, %r872, 255;
shr.u32 %r874, %r873, %r16;
and.b32 %r875, %r874, 15;
add.s64 %rd106, %rd113, %rd12;
ld.global.nc.u16 %rs15, [%rd106];
cvt.u32.u16 %r876, %rs15;
shr.u32 %r878, %r876, %r1742;
and.b32 %r879, %r878, 3;
bfi.b32 %r880, %r879, %r875, 4, 2;
add.s32 %r881, %r880, -32;
mul.lo.s32 %r882, %r881, %r801;
ld.global.nc.u16 %rs13, [%rd106+-2];

{ cvt.f32.f16 %f125, %rs13;}



{.reg .f16 low,high;
mov.b32 {low,high},%r30;
cvt.f32.f16 %f126, low;}


mul.ftz.f32 %f127, %f125, %f126;
cvt.rn.f32.s32 %f128, %r882;
fma.rn.ftz.f32 %f27, %f127, %f128, %f27;
@%p13 bra $L__BB293_18;

shl.b32 %r1743, %r15, 1;
add.s64 %rd107, %rd120, %rd9;
ld.global.u32 %r916, [%rd107+-12];
and.b32 %r917, %r916, -2004318072;
shr.u32 %r918, %r917, 1;
or.b32 %r919, %r918, 839922192;
and.b32 %r920, %r916, 30583;
mov.u32 %r921, -152379953;
mov.u32 %r922, -1079142271;
prmt.b32 %r923, %r922, %r921, %r920;
mov.u32 %r924, 1901675829;
mov.u32 %r925, 639175937;
prmt.b32 %r926, %r925, %r924, %r920;
and.b32 %r927, %r919, 30292;
prmt.b32 %r928, %r923, %r926, %r927;
shr.u32 %r929, %r916, 16;
and.b32 %r930, %r929, 30583;
prmt.b32 %r931, %r922, %r921, %r930;
prmt.b32 %r932, %r925, %r924, %r930;
shr.u32 %r933, %r919, 16;
prmt.b32 %r934, %r931, %r932, %r933;
mov.u32 %r935, 25632;
prmt.b32 %r884, %r928, %r934, %r935;
mov.u32 %r936, 30001;
prmt.b32 %r888, %r928, %r934, %r936;
mov.u32 %r886, 0;

dp4a.s32.s32 %r883, %r884, %r22, %r886;


dp4a.s32.s32 %r887, %r888, %r23, %r883;

ld.global.u32 %r937, [%rd107+-8];
and.b32 %r938, %r937, -2004318072;
shr.u32 %r939, %r938, 1;
or.b32 %r940, %r939, 839922192;
and.b32 %r941, %r937, 30583;
prmt.b32 %r942, %r922, %r921, %r941;
prmt.b32 %r943, %r925, %r924, %r941;
and.b32 %r944, %r940, 30292;
prmt.b32 %r945, %r942, %r943, %r944;
shr.u32 %r946, %r937, 16;
and.b32 %r947, %r946, 30583;
prmt.b32 %r948, %r922, %r921, %r947;
prmt.b32 %r949, %r925, %r924, %r947;
shr.u32 %r950, %r940, 16;
prmt.b32 %r951, %r948, %r949, %r950;
prmt.b32 %r892, %r945, %r951, %r935;
prmt.b32 %r896, %r945, %r951, %r936;

dp4a.s32.s32 %r891, %r892, %r24, %r887;


dp4a.s32.s32 %r895, %r896, %r25, %r891;

ld.global.u32 %r952, [%rd107+-4];
and.b32 %r953, %r952, -2004318072;
shr.u32 %r954, %r953, 1;
or.b32 %r955, %r954, 839922192;
and.b32 %r956, %r952, 30583;
prmt.b32 %r957, %r922, %r921, %r956;
prmt.b32 %r958, %r925, %r924, %r956;
and.b32 %r959, %r955, 30292;
prmt.b32 %r960, %r957, %r958, %r959;
shr.u32 %r961, %r952, 16;
and.b32 %r962, %r961, 30583;
prmt.b32 %r963, %r922, %r921, %r962;
prmt.b32 %r964, %r925, %r924, %r962;
shr.u32 %r965, %r955, 16;
prmt.b32 %r966, %r963, %r964, %r965;
prmt.b32 %r900, %r960, %r966, %r935;
prmt.b32 %r904, %r960, %r966, %r936;

dp4a.s32.s32 %r899, %r900, %r26, %r895;


dp4a.s32.s32 %r903, %r904, %r27, %r899;

ld.global.u32 %r967, [%rd107];
and.b32 %r968, %r967, -2004318072;
shr.u32 %r969, %r968, 1;
or.b32 %r970, %r969, 839922192;
and.b32 %r971, %r967, 30583;
prmt.b32 %r972, %r922, %r921, %r971;
prmt.b32 %r973, %r925, %r924, %r971;
and.b32 %r974, %r970, 30292;
prmt.b32 %r975, %r972, %r973, %r974;
shr.u32 %r976, %r967, 16;
and.b32 %r977, %r976, 30583;
prmt.b32 %r978, %r922, %r921, %r977;
prmt.b32 %r979, %r925, %r924, %r977;
shr.u32 %r980, %r970, 16;
prmt.b32 %r981, %r978, %r979, %r980;
prmt.b32 %r908, %r975, %r981, %r935;
prmt.b32 %r912, %r975, %r981, %r936;

dp4a.s32.s32 %r907, %r908, %r28, %r903;


dp4a.s32.s32 %r911, %r912, %r29, %r907;

add.s64 %rd108, %rd120, %rd10;
ld.global.u8 %r982, [%rd108];
shr.u32 %r983, %r982, %r16;
and.b32 %r984, %r983, 15;
or.b32 %r987, %r4, 3;
mad.lo.s32 %r988, %r53, %r987, %r18;
mul.wide.s32 %rd109, %r988, 136;
add.s64 %rd110, %rd120, %rd109;
ld.global.u16 %r989, [%rd110+2];
shr.u32 %r991, %r989, %r1743;
and.b32 %r992, %r991, 3;
bfi.b32 %r993, %r992, %r984, 4, 2;
add.s32 %r994, %r993, -32;
mul.lo.s32 %r995, %r994, %r911;
ld.global.u16 %rs16, [%rd110];

{ cvt.f32.f16 %f129, %rs16;}



{.reg .f16 low,high;
mov.b32 {low,high},%r30;
cvt.f32.f16 %f130, low;}


mul.ftz.f32 %f131, %f129, %f130;
cvt.rn.f32.s32 %f132, %r995;
fma.rn.ftz.f32 %f31, %f131, %f132, %f31;

$L__BB293_18:
add.s32 %r1749, %r1749, 128;
add.s64 %rd120, %rd120, 2176;
add.s64 %rd119, %rd119, 2176;
add.s32 %r1750, %r1750, 16;
setp.lt.s32 %p17, %r1750, %r87;
@%p17 bra $L__BB293_10;

$L__BB293_19:
mov.u32 %r1729, %tid.y;
setp.eq.s32 %p18, %r1729, 0;
@%p18 bra $L__BB293_23;

mov.u32 %r1730, %tid.y;
setp.ne.s64 %p19, %rd33, 0;
add.s32 %r33, %r1730, -1;
shl.b32 %r1000, %r1730, 9;
mov.u32 %r1001, _ZZ13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r1002, %r1001, %r1000;
shl.b32 %r1003, %r2, 2;
add.s32 %r1004, %r1002, %r1003;
st.shared.f32 [%r1004+-512], %f30;
@%p19 bra $L__BB293_22;
bra.uni $L__BB293_21;

$L__BB293_22:
shl.b32 %r1006, %r33, 9;
mov.u32 %r1007, _ZZ13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r1008, %r1007, %r1006;
add.s32 %r1010, %r1008, %r1003;
st.shared.f32 [%r1010], %f34;
add.s32 %r1726, %r1004, -512;
st.shared.f32 [%r1726+128], %f29;
st.shared.f32 [%r1010+128], %f33;
add.s32 %r1727, %r1004, -512;
st.shared.f32 [%r1727+256], %f28;
st.shared.f32 [%r1010+256], %f32;
add.s32 %r1728, %r1004, -512;
st.shared.f32 [%r1728+384], %f27;
st.shared.f32 [%r1010+384], %f31;
bra.uni $L__BB293_23;

$L__BB293_21:
add.s32 %r1723, %r1004, -512;
st.shared.f32 [%r1723+128], %f29;
add.s32 %r1724, %r1004, -512;
st.shared.f32 [%r1724+256], %f28;
add.s32 %r1725, %r1004, -512;
st.shared.f32 [%r1725+384], %f27;

$L__BB293_23:
mov.u32 %r1731, %tid.y;
setp.ne.s32 %p319, %r1731, 0;
bar.sync 0;
@%p319 bra $L__BB293_78;
bra.uni $L__BB293_24;

$L__BB293_78:
ret;

$L__BB293_24:
ld.param.u64 %rd116, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.f32 %f977, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
ld.param.u32 %r1734, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r1733, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
mov.u32 %r1732, %ctaid.z;
setp.ne.s64 %p21, %rd33, 0;
mov.u32 %r1013, %ctaid.y;
mad.lo.s32 %r1016, %r1013, %r1734, %r4;
mad.lo.s32 %r1017, %r1732, %r1733, %r1016;
neg.ftz.f32 %f35, %f977;
shl.b32 %r1019, %r2, 2;
mov.u32 %r1020, _ZZ13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r35, %r1020, %r1019;
cvta.to.global.u64 %rd111, %rd116;
mul.wide.u32 %rd112, %r1017, 4;
add.s64 %rd28, %rd111, %rd112;
@%p21 bra $L__BB293_33;
bra.uni $L__BB293_25;

$L__BB293_33:
ld.param.u32 %r1735, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
mov.u32 %r1107, _ZZ13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r37, %r1107, %r1019;
setp.gt.s32 %p54, %r1735, 2;
@%p54 bra $L__BB293_44;

ld.param.u32 %r1738, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p57, %r1738, 1;
@%p57 bra $L__BB293_62;

ld.param.u32 %r1739, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p58, %r1739, 2;
@%p58 bra $L__BB293_36;
bra.uni $L__BB293_70;

$L__BB293_36:
ld.shared.f32 %f533, [%r35];
add.ftz.f32 %f534, %f533, %f30;
ld.shared.f32 %f535, [%r37];
add.ftz.f32 %f536, %f535, %f34;
ld.shared.f32 %f537, [%r35+512];
add.ftz.f32 %f538, %f537, %f534;
ld.shared.f32 %f539, [%r37+512];
add.ftz.f32 %f540, %f539, %f536;
ld.shared.f32 %f541, [%r35+1024];
add.ftz.f32 %f542, %f541, %f538;
ld.shared.f32 %f543, [%r37+1024];
add.ftz.f32 %f544, %f543, %f540;
mov.b32 %r1354, %f542;
mov.u32 %r1355, 31;
mov.u32 %r1356, 16;
mov.u32 %r1357, -1;
shfl.sync.bfly.b32 %r1358|%p163, %r1354, %r1356, %r1355, %r1357;
mov.b32 %f545, %r1358;
add.ftz.f32 %f546, %f542, %f545;
mov.b32 %r1359, %f546;
mov.u32 %r1360, 8;
shfl.sync.bfly.b32 %r1361|%p164, %r1359, %r1360, %r1355, %r1357;
mov.b32 %f547, %r1361;
add.ftz.f32 %f548, %f546, %f547;
mov.b32 %r1362, %f548;
mov.u32 %r1363, 4;
shfl.sync.bfly.b32 %r1364|%p165, %r1362, %r1363, %r1355, %r1357;
mov.b32 %f549, %r1364;
add.ftz.f32 %f550, %f548, %f549;
mov.b32 %r1365, %f550;
mov.u32 %r1366, 2;
shfl.sync.bfly.b32 %r1367|%p166, %r1365, %r1366, %r1355, %r1357;
mov.b32 %f551, %r1367;
add.ftz.f32 %f552, %f550, %f551;
mov.b32 %r1368, %f552;
mov.u32 %r1369, 1;
shfl.sync.bfly.b32 %r1370|%p167, %r1368, %r1369, %r1355, %r1357;
mov.b32 %f553, %r1370;
add.ftz.f32 %f56, %f552, %f553;
mov.b32 %r1371, %f544;
shfl.sync.bfly.b32 %r1372|%p168, %r1371, %r1356, %r1355, %r1357;
mov.b32 %f554, %r1372;
add.ftz.f32 %f555, %f544, %f554;
mov.b32 %r1373, %f555;
shfl.sync.bfly.b32 %r1374|%p169, %r1373, %r1360, %r1355, %r1357;
mov.b32 %f556, %r1374;
add.ftz.f32 %f557, %f555, %f556;
mov.b32 %r1375, %f557;
shfl.sync.bfly.b32 %r1376|%p170, %r1375, %r1363, %r1355, %r1357;
mov.b32 %f558, %r1376;
add.ftz.f32 %f559, %f557, %f558;
mov.b32 %r1377, %f559;
shfl.sync.bfly.b32 %r1378|%p171, %r1377, %r1366, %r1355, %r1357;
mov.b32 %f560, %r1378;
add.ftz.f32 %f561, %f559, %f560;
mov.b32 %r1379, %f561;
shfl.sync.bfly.b32 %r1380|%p172, %r1379, %r1369, %r1355, %r1357;
mov.b32 %f562, %r1380;
add.ftz.f32 %f57, %f561, %f562;
setp.ne.s32 %p173, %r2, 0;
setp.ge.u32 %p174, %r4, %r54;
or.pred %p175, %p173, %p174;
@%p175 bra $L__BB293_38;

add.ftz.f32 %f563, %f5, %f56;
add.ftz.f32 %f564, %f4, %f57;
mul.ftz.f32 %f565, %f564, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f566, %f565;
add.ftz.f32 %f567, %f566, 0f3F800000;
div.approx.ftz.f32 %f568, %f564, %f567;
mul.ftz.f32 %f569, %f563, %f568;
st.global.f32 [%rd28], %f569;

$L__BB293_38:
ld.shared.f32 %f570, [%r35+128];
add.ftz.f32 %f571, %f570, %f29;
ld.shared.f32 %f572, [%r37+128];
add.ftz.f32 %f573, %f572, %f33;
ld.shared.f32 %f574, [%r35+640];
add.ftz.f32 %f575, %f574, %f571;
ld.shared.f32 %f576, [%r37+640];
add.ftz.f32 %f577, %f576, %f573;
ld.shared.f32 %f578, [%r35+1152];
add.ftz.f32 %f579, %f578, %f575;
ld.shared.f32 %f580, [%r37+1152];
add.ftz.f32 %f581, %f580, %f577;
mov.b32 %r1384, %f579;
shfl.sync.bfly.b32 %r1388|%p176, %r1384, %r1356, %r1355, %r1357;
mov.b32 %f582, %r1388;
add.ftz.f32 %f583, %f579, %f582;
mov.b32 %r1389, %f583;
shfl.sync.bfly.b32 %r1391|%p177, %r1389, %r1360, %r1355, %r1357;
mov.b32 %f584, %r1391;
add.ftz.f32 %f585, %f583, %f584;
mov.b32 %r1392, %f585;
shfl.sync.bfly.b32 %r1394|%p178, %r1392, %r1363, %r1355, %r1357;
mov.b32 %f586, %r1394;
add.ftz.f32 %f587, %f585, %f586;
mov.b32 %r1395, %f587;
shfl.sync.bfly.b32 %r1397|%p179, %r1395, %r1366, %r1355, %r1357;
mov.b32 %f588, %r1397;
add.ftz.f32 %f589, %f587, %f588;
mov.b32 %r1398, %f589;
shfl.sync.bfly.b32 %r1400|%p180, %r1398, %r1369, %r1355, %r1357;
mov.b32 %f590, %r1400;
add.ftz.f32 %f58, %f589, %f590;
mov.b32 %r1401, %f581;
shfl.sync.bfly.b32 %r1402|%p181, %r1401, %r1356, %r1355, %r1357;
mov.b32 %f591, %r1402;
add.ftz.f32 %f592, %f581, %f591;
mov.b32 %r1403, %f592;
shfl.sync.bfly.b32 %r1404|%p182, %r1403, %r1360, %r1355, %r1357;
mov.b32 %f593, %r1404;
add.ftz.f32 %f594, %f592, %f593;
mov.b32 %r1405, %f594;
shfl.sync.bfly.b32 %r1406|%p183, %r1405, %r1363, %r1355, %r1357;
mov.b32 %f595, %r1406;
add.ftz.f32 %f596, %f594, %f595;
mov.b32 %r1407, %f596;
shfl.sync.bfly.b32 %r1408|%p184, %r1407, %r1366, %r1355, %r1357;
mov.b32 %f597, %r1408;
add.ftz.f32 %f598, %f596, %f597;
mov.b32 %r1409, %f598;
shfl.sync.bfly.b32 %r1410|%p185, %r1409, %r1369, %r1355, %r1357;
mov.b32 %f599, %r1410;
add.ftz.f32 %f59, %f598, %f599;
setp.ne.s32 %p186, %r2, 1;
or.b32 %r1414, %r4, 1;
setp.ge.u32 %p187, %r1414, %r54;
or.pred %p188, %p186, %p187;
@%p188 bra $L__BB293_40;

add.ftz.f32 %f600, %f5, %f58;
add.ftz.f32 %f601, %f4, %f59;
mul.ftz.f32 %f602, %f601, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f603, %f602;
add.ftz.f32 %f604, %f603, 0f3F800000;
div.approx.ftz.f32 %f605, %f601, %f604;
mul.ftz.f32 %f606, %f600, %f605;
st.global.f32 [%rd28+4], %f606;

$L__BB293_40:
ld.shared.f32 %f607, [%r35+256];
add.ftz.f32 %f608, %f607, %f28;
ld.shared.f32 %f609, [%r37+256];
add.ftz.f32 %f610, %f609, %f32;
ld.shared.f32 %f611, [%r35+768];
add.ftz.f32 %f612, %f611, %f608;
ld.shared.f32 %f613, [%r37+768];
add.ftz.f32 %f614, %f613, %f610;
ld.shared.f32 %f615, [%r35+1280];
add.ftz.f32 %f616, %f615, %f612;
ld.shared.f32 %f617, [%r37+1280];
add.ftz.f32 %f618, %f617, %f614;
mov.b32 %r1415, %f616;
mov.u32 %r1416, 31;
mov.u32 %r1417, 16;
mov.u32 %r1418, -1;
shfl.sync.bfly.b32 %r1419|%p189, %r1415, %r1417, %r1416, %r1418;
mov.b32 %f619, %r1419;
add.ftz.f32 %f620, %f616, %f619;
mov.b32 %r1420, %f620;
mov.u32 %r1421, 8;
shfl.sync.bfly.b32 %r1422|%p190, %r1420, %r1421, %r1416, %r1418;
mov.b32 %f621, %r1422;
add.ftz.f32 %f622, %f620, %f621;
mov.b32 %r1423, %f622;
mov.u32 %r1424, 4;
shfl.sync.bfly.b32 %r1425|%p191, %r1423, %r1424, %r1416, %r1418;
mov.b32 %f623, %r1425;
add.ftz.f32 %f624, %f622, %f623;
mov.b32 %r1426, %f624;
mov.u32 %r1427, 2;
shfl.sync.bfly.b32 %r1428|%p192, %r1426, %r1427, %r1416, %r1418;
mov.b32 %f625, %r1428;
add.ftz.f32 %f626, %f624, %f625;
mov.b32 %r1429, %f626;
mov.u32 %r1430, 1;
shfl.sync.bfly.b32 %r1431|%p193, %r1429, %r1430, %r1416, %r1418;
mov.b32 %f627, %r1431;
add.ftz.f32 %f60, %f626, %f627;
mov.b32 %r1432, %f618;
shfl.sync.bfly.b32 %r1433|%p194, %r1432, %r1417, %r1416, %r1418;
mov.b32 %f628, %r1433;
add.ftz.f32 %f629, %f618, %f628;
mov.b32 %r1434, %f629;
shfl.sync.bfly.b32 %r1435|%p195, %r1434, %r1421, %r1416, %r1418;
mov.b32 %f630, %r1435;
add.ftz.f32 %f631, %f629, %f630;
mov.b32 %r1436, %f631;
shfl.sync.bfly.b32 %r1437|%p196, %r1436, %r1424, %r1416, %r1418;
mov.b32 %f632, %r1437;
add.ftz.f32 %f633, %f631, %f632;
mov.b32 %r1438, %f633;
shfl.sync.bfly.b32 %r1439|%p197, %r1438, %r1427, %r1416, %r1418;
mov.b32 %f634, %r1439;
add.ftz.f32 %f635, %f633, %f634;
mov.b32 %r1440, %f635;
shfl.sync.bfly.b32 %r1441|%p198, %r1440, %r1430, %r1416, %r1418;
mov.b32 %f636, %r1441;
add.ftz.f32 %f61, %f635, %f636;
setp.ne.s32 %p199, %r2, 2;
or.b32 %r1445, %r4, 2;
setp.ge.u32 %p200, %r1445, %r54;
or.pred %p201, %p199, %p200;
@%p201 bra $L__BB293_42;

add.ftz.f32 %f637, %f5, %f60;
add.ftz.f32 %f638, %f4, %f61;
mul.ftz.f32 %f639, %f638, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f640, %f639;
add.ftz.f32 %f641, %f640, 0f3F800000;
div.approx.ftz.f32 %f642, %f638, %f641;
mul.ftz.f32 %f643, %f637, %f642;
st.global.f32 [%rd28+8], %f643;

$L__BB293_42:
ld.shared.f32 %f644, [%r35+384];
add.ftz.f32 %f645, %f644, %f27;
ld.shared.f32 %f646, [%r37+384];
add.ftz.f32 %f647, %f646, %f31;
ld.shared.f32 %f648, [%r35+896];
add.ftz.f32 %f649, %f648, %f645;
ld.shared.f32 %f650, [%r37+896];
add.ftz.f32 %f651, %f650, %f647;
ld.shared.f32 %f652, [%r35+1408];
add.ftz.f32 %f653, %f652, %f649;
ld.shared.f32 %f654, [%r37+1408];
add.ftz.f32 %f655, %f654, %f651;
mov.b32 %r1446, %f653;
shfl.sync.bfly.b32 %r1450|%p202, %r1446, %r1417, %r1416, %r1418;
mov.b32 %f656, %r1450;
add.ftz.f32 %f657, %f653, %f656;
mov.b32 %r1451, %f657;
shfl.sync.bfly.b32 %r1453|%p203, %r1451, %r1421, %r1416, %r1418;
mov.b32 %f658, %r1453;
add.ftz.f32 %f659, %f657, %f658;
mov.b32 %r1454, %f659;
shfl.sync.bfly.b32 %r1456|%p204, %r1454, %r1424, %r1416, %r1418;
mov.b32 %f660, %r1456;
add.ftz.f32 %f661, %f659, %f660;
mov.b32 %r1457, %f661;
shfl.sync.bfly.b32 %r1459|%p205, %r1457, %r1427, %r1416, %r1418;
mov.b32 %f662, %r1459;
add.ftz.f32 %f663, %f661, %f662;
mov.b32 %r1460, %f663;
shfl.sync.bfly.b32 %r1462|%p206, %r1460, %r1430, %r1416, %r1418;
mov.b32 %f664, %r1462;
add.ftz.f32 %f62, %f663, %f664;
mov.b32 %r1463, %f655;
shfl.sync.bfly.b32 %r1464|%p207, %r1463, %r1417, %r1416, %r1418;
mov.b32 %f665, %r1464;
add.ftz.f32 %f666, %f655, %f665;
mov.b32 %r1465, %f666;
shfl.sync.bfly.b32 %r1466|%p208, %r1465, %r1421, %r1416, %r1418;
mov.b32 %f667, %r1466;
add.ftz.f32 %f668, %f666, %f667;
mov.b32 %r1467, %f668;
shfl.sync.bfly.b32 %r1468|%p209, %r1467, %r1424, %r1416, %r1418;
mov.b32 %f669, %r1468;
add.ftz.f32 %f670, %f668, %f669;
mov.b32 %r1469, %f670;
shfl.sync.bfly.b32 %r1470|%p210, %r1469, %r1427, %r1416, %r1418;
mov.b32 %f671, %r1470;
add.ftz.f32 %f672, %f670, %f671;
mov.b32 %r1471, %f672;
shfl.sync.bfly.b32 %r1472|%p211, %r1471, %r1430, %r1416, %r1418;
mov.b32 %f673, %r1472;
add.ftz.f32 %f63, %f672, %f673;
setp.ne.s32 %p212, %r2, 3;
or.b32 %r1476, %r4, 3;
setp.ge.u32 %p213, %r1476, %r54;
or.pred %p214, %p212, %p213;
@%p214 bra $L__BB293_78;

add.ftz.f32 %f674, %f5, %f62;
add.ftz.f32 %f675, %f4, %f63;
mul.ftz.f32 %f676, %f675, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f677, %f676;
add.ftz.f32 %f678, %f677, 0f3F800000;
div.approx.ftz.f32 %f679, %f675, %f678;
mul.ftz.f32 %f680, %f674, %f679;
st.global.f32 [%rd28+12], %f680;
bra.uni $L__BB293_78;

$L__BB293_25:
ld.shared.f32 %f133, [%r35];
add.ftz.f32 %f134, %f133, %f30;
ld.shared.f32 %f135, [%r35+512];
add.ftz.f32 %f136, %f135, %f134;
ld.shared.f32 %f137, [%r35+1024];
add.ftz.f32 %f138, %f137, %f136;
mov.b32 %r1021, %f138;
mov.u32 %r1022, 31;
mov.u32 %r1023, 16;
mov.u32 %r1024, -1;
shfl.sync.bfly.b32 %r1025|%p22, %r1021, %r1023, %r1022, %r1024;
mov.b32 %f139, %r1025;
add.ftz.f32 %f140, %f138, %f139;
mov.b32 %r1026, %f140;
mov.u32 %r1027, 8;
shfl.sync.bfly.b32 %r1028|%p23, %r1026, %r1027, %r1022, %r1024;
mov.b32 %f141, %r1028;
add.ftz.f32 %f142, %f140, %f141;
mov.b32 %r1029, %f142;
mov.u32 %r1030, 4;
shfl.sync.bfly.b32 %r1031|%p24, %r1029, %r1030, %r1022, %r1024;
mov.b32 %f143, %r1031;
add.ftz.f32 %f144, %f142, %f143;
mov.b32 %r1032, %f144;
mov.u32 %r1033, 2;
shfl.sync.bfly.b32 %r1034|%p25, %r1032, %r1033, %r1022, %r1024;
mov.b32 %f145, %r1034;
add.ftz.f32 %f146, %f144, %f145;
mov.b32 %r1035, %f146;
mov.u32 %r1036, 1;
shfl.sync.bfly.b32 %r1037|%p26, %r1035, %r1036, %r1022, %r1024;
mov.b32 %f147, %r1037;
add.ftz.f32 %f36, %f146, %f147;
setp.ne.s32 %p27, %r2, 0;
setp.ge.u32 %p28, %r4, %r54;
or.pred %p29, %p27, %p28;
@%p29 bra $L__BB293_27;

add.ftz.f32 %f148, %f5, %f36;
st.global.f32 [%rd28], %f148;

$L__BB293_27:
ld.shared.f32 %f149, [%r35+128];
add.ftz.f32 %f150, %f149, %f29;
ld.shared.f32 %f151, [%r35+640];
add.ftz.f32 %f152, %f151, %f150;
ld.shared.f32 %f153, [%r35+1152];
add.ftz.f32 %f154, %f153, %f152;
mov.b32 %r1041, %f154;
shfl.sync.bfly.b32 %r1045|%p30, %r1041, %r1023, %r1022, %r1024;
mov.b32 %f155, %r1045;
add.ftz.f32 %f156, %f154, %f155;
mov.b32 %r1046, %f156;
shfl.sync.bfly.b32 %r1048|%p31, %r1046, %r1027, %r1022, %r1024;
mov.b32 %f157, %r1048;
add.ftz.f32 %f158, %f156, %f157;
mov.b32 %r1049, %f158;
shfl.sync.bfly.b32 %r1051|%p32, %r1049, %r1030, %r1022, %r1024;
mov.b32 %f159, %r1051;
add.ftz.f32 %f160, %f158, %f159;
mov.b32 %r1052, %f160;
shfl.sync.bfly.b32 %r1054|%p33, %r1052, %r1033, %r1022, %r1024;
mov.b32 %f161, %r1054;
add.ftz.f32 %f162, %f160, %f161;
mov.b32 %r1055, %f162;
shfl.sync.bfly.b32 %r1057|%p34, %r1055, %r1036, %r1022, %r1024;
mov.b32 %f163, %r1057;
add.ftz.f32 %f37, %f162, %f163;
setp.ne.s32 %p35, %r2, 1;
or.b32 %r1061, %r4, 1;
setp.ge.u32 %p36, %r1061, %r54;
or.pred %p37, %p35, %p36;
@%p37 bra $L__BB293_29;

add.ftz.f32 %f164, %f5, %f37;
st.global.f32 [%rd28+4], %f164;

$L__BB293_29:
ld.shared.f32 %f165, [%r35+256];
add.ftz.f32 %f166, %f165, %f28;
ld.shared.f32 %f167, [%r35+768];
add.ftz.f32 %f168, %f167, %f166;
ld.shared.f32 %f169, [%r35+1280];
add.ftz.f32 %f170, %f169, %f168;
mov.b32 %r1062, %f170;
mov.u32 %r1063, 31;
mov.u32 %r1064, 16;
mov.u32 %r1065, -1;
shfl.sync.bfly.b32 %r1066|%p38, %r1062, %r1064, %r1063, %r1065;
mov.b32 %f171, %r1066;
add.ftz.f32 %f172, %f170, %f171;
mov.b32 %r1067, %f172;
mov.u32 %r1068, 8;
shfl.sync.bfly.b32 %r1069|%p39, %r1067, %r1068, %r1063, %r1065;
mov.b32 %f173, %r1069;
add.ftz.f32 %f174, %f172, %f173;
mov.b32 %r1070, %f174;
mov.u32 %r1071, 4;
shfl.sync.bfly.b32 %r1072|%p40, %r1070, %r1071, %r1063, %r1065;
mov.b32 %f175, %r1072;
add.ftz.f32 %f176, %f174, %f175;
mov.b32 %r1073, %f176;
mov.u32 %r1074, 2;
shfl.sync.bfly.b32 %r1075|%p41, %r1073, %r1074, %r1063, %r1065;
mov.b32 %f177, %r1075;
add.ftz.f32 %f178, %f176, %f177;
mov.b32 %r1076, %f178;
mov.u32 %r1077, 1;
shfl.sync.bfly.b32 %r1078|%p42, %r1076, %r1077, %r1063, %r1065;
mov.b32 %f179, %r1078;
add.ftz.f32 %f38, %f178, %f179;
setp.ne.s32 %p43, %r2, 2;
or.b32 %r1082, %r4, 2;
setp.ge.u32 %p44, %r1082, %r54;
or.pred %p45, %p43, %p44;
@%p45 bra $L__BB293_31;

add.ftz.f32 %f180, %f5, %f38;
st.global.f32 [%rd28+8], %f180;

$L__BB293_31:
ld.shared.f32 %f181, [%r35+384];
add.ftz.f32 %f182, %f181, %f27;
ld.shared.f32 %f183, [%r35+896];
add.ftz.f32 %f184, %f183, %f182;
ld.shared.f32 %f185, [%r35+1408];
add.ftz.f32 %f186, %f185, %f184;
mov.b32 %r1083, %f186;
shfl.sync.bfly.b32 %r1087|%p46, %r1083, %r1064, %r1063, %r1065;
mov.b32 %f187, %r1087;
add.ftz.f32 %f188, %f186, %f187;
mov.b32 %r1088, %f188;
shfl.sync.bfly.b32 %r1090|%p47, %r1088, %r1068, %r1063, %r1065;
mov.b32 %f189, %r1090;
add.ftz.f32 %f190, %f188, %f189;
mov.b32 %r1091, %f190;
shfl.sync.bfly.b32 %r1093|%p48, %r1091, %r1071, %r1063, %r1065;
mov.b32 %f191, %r1093;
add.ftz.f32 %f192, %f190, %f191;
mov.b32 %r1094, %f192;
shfl.sync.bfly.b32 %r1096|%p49, %r1094, %r1074, %r1063, %r1065;
mov.b32 %f193, %r1096;
add.ftz.f32 %f194, %f192, %f193;
mov.b32 %r1097, %f194;
shfl.sync.bfly.b32 %r1099|%p50, %r1097, %r1077, %r1063, %r1065;
mov.b32 %f195, %r1099;
add.ftz.f32 %f39, %f194, %f195;
setp.ne.s32 %p51, %r2, 3;
or.b32 %r1103, %r4, 3;
setp.ge.u32 %p52, %r1103, %r54;
or.pred %p53, %p51, %p52;
@%p53 bra $L__BB293_78;

add.ftz.f32 %f196, %f5, %f39;
st.global.f32 [%rd28+12], %f196;
bra.uni $L__BB293_78;

$L__BB293_44:
ld.param.u32 %r1736, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p55, %r1736, 3;
@%p55 bra $L__BB293_54;

ld.param.u32 %r1737, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.ne.s32 %p56, %r1737, 6;
@%p56 bra $L__BB293_70;

ld.shared.f32 %f197, [%r35];
add.ftz.f32 %f198, %f197, %f30;
ld.shared.f32 %f199, [%r37];
add.ftz.f32 %f200, %f199, %f34;
ld.shared.f32 %f201, [%r35+512];
add.ftz.f32 %f202, %f201, %f198;
ld.shared.f32 %f203, [%r37+512];
add.ftz.f32 %f204, %f203, %f200;
ld.shared.f32 %f205, [%r35+1024];
add.ftz.f32 %f206, %f205, %f202;
ld.shared.f32 %f207, [%r37+1024];
add.ftz.f32 %f208, %f207, %f204;
mov.b32 %r1108, %f206;
mov.u32 %r1109, 31;
mov.u32 %r1110, 16;
mov.u32 %r1111, -1;
shfl.sync.bfly.b32 %r1112|%p59, %r1108, %r1110, %r1109, %r1111;
mov.b32 %f209, %r1112;
add.ftz.f32 %f210, %f206, %f209;
mov.b32 %r1113, %f210;
mov.u32 %r1114, 8;
shfl.sync.bfly.b32 %r1115|%p60, %r1113, %r1114, %r1109, %r1111;
mov.b32 %f211, %r1115;
add.ftz.f32 %f212, %f210, %f211;
mov.b32 %r1116, %f212;
mov.u32 %r1117, 4;
shfl.sync.bfly.b32 %r1118|%p61, %r1116, %r1117, %r1109, %r1111;
mov.b32 %f213, %r1118;
add.ftz.f32 %f214, %f212, %f213;
mov.b32 %r1119, %f214;
mov.u32 %r1120, 2;
shfl.sync.bfly.b32 %r1121|%p62, %r1119, %r1120, %r1109, %r1111;
mov.b32 %f215, %r1121;
add.ftz.f32 %f216, %f214, %f215;
mov.b32 %r1122, %f216;
mov.u32 %r1123, 1;
shfl.sync.bfly.b32 %r1124|%p63, %r1122, %r1123, %r1109, %r1111;
mov.b32 %f217, %r1124;
add.ftz.f32 %f40, %f216, %f217;
mov.b32 %r1125, %f208;
shfl.sync.bfly.b32 %r1126|%p64, %r1125, %r1110, %r1109, %r1111;
mov.b32 %f218, %r1126;
add.ftz.f32 %f219, %f208, %f218;
mov.b32 %r1127, %f219;
shfl.sync.bfly.b32 %r1128|%p65, %r1127, %r1114, %r1109, %r1111;
mov.b32 %f220, %r1128;
add.ftz.f32 %f221, %f219, %f220;
mov.b32 %r1129, %f221;
shfl.sync.bfly.b32 %r1130|%p66, %r1129, %r1117, %r1109, %r1111;
mov.b32 %f222, %r1130;
add.ftz.f32 %f223, %f221, %f222;
mov.b32 %r1131, %f223;
shfl.sync.bfly.b32 %r1132|%p67, %r1131, %r1120, %r1109, %r1111;
mov.b32 %f224, %r1132;
add.ftz.f32 %f225, %f223, %f224;
mov.b32 %r1133, %f225;
shfl.sync.bfly.b32 %r1134|%p68, %r1133, %r1123, %r1109, %r1111;
mov.b32 %f226, %r1134;
add.ftz.f32 %f41, %f225, %f226;
setp.ne.s32 %p69, %r2, 0;
setp.ge.u32 %p70, %r4, %r54;
or.pred %p71, %p69, %p70;
@%p71 bra $L__BB293_48;

ld.param.f32 %f981, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f227, %f5, %f40;
add.ftz.f32 %f228, %f4, %f41;
min.ftz.f32 %f229, %f228, %f981;
min.ftz.f32 %f230, %f227, %f981;
max.ftz.f32 %f231, %f230, %f35;
mul.ftz.f32 %f232, %f229, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f233, %f232;
add.ftz.f32 %f234, %f233, 0f3F800000;
div.approx.ftz.f32 %f235, %f229, %f234;
mul.ftz.f32 %f236, %f231, %f235;
st.global.f32 [%rd28], %f236;

$L__BB293_48:
ld.shared.f32 %f237, [%r35+128];
add.ftz.f32 %f238, %f237, %f29;
ld.shared.f32 %f239, [%r37+128];
add.ftz.f32 %f240, %f239, %f33;
ld.shared.f32 %f241, [%r35+640];
add.ftz.f32 %f242, %f241, %f238;
ld.shared.f32 %f243, [%r37+640];
add.ftz.f32 %f244, %f243, %f240;
ld.shared.f32 %f245, [%r35+1152];
add.ftz.f32 %f246, %f245, %f242;
ld.shared.f32 %f247, [%r37+1152];
add.ftz.f32 %f248, %f247, %f244;
mov.b32 %r1138, %f246;
shfl.sync.bfly.b32 %r1142|%p72, %r1138, %r1110, %r1109, %r1111;
mov.b32 %f249, %r1142;
add.ftz.f32 %f250, %f246, %f249;
mov.b32 %r1143, %f250;
shfl.sync.bfly.b32 %r1145|%p73, %r1143, %r1114, %r1109, %r1111;
mov.b32 %f251, %r1145;
add.ftz.f32 %f252, %f250, %f251;
mov.b32 %r1146, %f252;
shfl.sync.bfly.b32 %r1148|%p74, %r1146, %r1117, %r1109, %r1111;
mov.b32 %f253, %r1148;
add.ftz.f32 %f254, %f252, %f253;
mov.b32 %r1149, %f254;
shfl.sync.bfly.b32 %r1151|%p75, %r1149, %r1120, %r1109, %r1111;
mov.b32 %f255, %r1151;
add.ftz.f32 %f256, %f254, %f255;
mov.b32 %r1152, %f256;
shfl.sync.bfly.b32 %r1154|%p76, %r1152, %r1123, %r1109, %r1111;
mov.b32 %f257, %r1154;
add.ftz.f32 %f42, %f256, %f257;
mov.b32 %r1155, %f248;
shfl.sync.bfly.b32 %r1156|%p77, %r1155, %r1110, %r1109, %r1111;
mov.b32 %f258, %r1156;
add.ftz.f32 %f259, %f248, %f258;
mov.b32 %r1157, %f259;
shfl.sync.bfly.b32 %r1158|%p78, %r1157, %r1114, %r1109, %r1111;
mov.b32 %f260, %r1158;
add.ftz.f32 %f261, %f259, %f260;
mov.b32 %r1159, %f261;
shfl.sync.bfly.b32 %r1160|%p79, %r1159, %r1117, %r1109, %r1111;
mov.b32 %f262, %r1160;
add.ftz.f32 %f263, %f261, %f262;
mov.b32 %r1161, %f263;
shfl.sync.bfly.b32 %r1162|%p80, %r1161, %r1120, %r1109, %r1111;
mov.b32 %f264, %r1162;
add.ftz.f32 %f265, %f263, %f264;
mov.b32 %r1163, %f265;
shfl.sync.bfly.b32 %r1164|%p81, %r1163, %r1123, %r1109, %r1111;
mov.b32 %f266, %r1164;
add.ftz.f32 %f43, %f265, %f266;
setp.ne.s32 %p82, %r2, 1;
or.b32 %r1168, %r4, 1;
setp.ge.u32 %p83, %r1168, %r54;
or.pred %p84, %p82, %p83;
@%p84 bra $L__BB293_50;

ld.param.f32 %f980, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f267, %f5, %f42;
add.ftz.f32 %f268, %f4, %f43;
min.ftz.f32 %f269, %f268, %f980;
min.ftz.f32 %f270, %f267, %f980;
max.ftz.f32 %f271, %f270, %f35;
mul.ftz.f32 %f272, %f269, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f273, %f272;
add.ftz.f32 %f274, %f273, 0f3F800000;
div.approx.ftz.f32 %f275, %f269, %f274;
mul.ftz.f32 %f276, %f271, %f275;
st.global.f32 [%rd28+4], %f276;

$L__BB293_50:
ld.shared.f32 %f277, [%r35+256];
add.ftz.f32 %f278, %f277, %f28;
ld.shared.f32 %f279, [%r37+256];
add.ftz.f32 %f280, %f279, %f32;
ld.shared.f32 %f281, [%r35+768];
add.ftz.f32 %f282, %f281, %f278;
ld.shared.f32 %f283, [%r37+768];
add.ftz.f32 %f284, %f283, %f280;
ld.shared.f32 %f285, [%r35+1280];
add.ftz.f32 %f286, %f285, %f282;
ld.shared.f32 %f287, [%r37+1280];
add.ftz.f32 %f288, %f287, %f284;
mov.b32 %r1169, %f286;
mov.u32 %r1170, 31;
mov.u32 %r1171, 16;
mov.u32 %r1172, -1;
shfl.sync.bfly.b32 %r1173|%p85, %r1169, %r1171, %r1170, %r1172;
mov.b32 %f289, %r1173;
add.ftz.f32 %f290, %f286, %f289;
mov.b32 %r1174, %f290;
mov.u32 %r1175, 8;
shfl.sync.bfly.b32 %r1176|%p86, %r1174, %r1175, %r1170, %r1172;
mov.b32 %f291, %r1176;
add.ftz.f32 %f292, %f290, %f291;
mov.b32 %r1177, %f292;
mov.u32 %r1178, 4;
shfl.sync.bfly.b32 %r1179|%p87, %r1177, %r1178, %r1170, %r1172;
mov.b32 %f293, %r1179;
add.ftz.f32 %f294, %f292, %f293;
mov.b32 %r1180, %f294;
mov.u32 %r1181, 2;
shfl.sync.bfly.b32 %r1182|%p88, %r1180, %r1181, %r1170, %r1172;
mov.b32 %f295, %r1182;
add.ftz.f32 %f296, %f294, %f295;
mov.b32 %r1183, %f296;
mov.u32 %r1184, 1;
shfl.sync.bfly.b32 %r1185|%p89, %r1183, %r1184, %r1170, %r1172;
mov.b32 %f297, %r1185;
add.ftz.f32 %f44, %f296, %f297;
mov.b32 %r1186, %f288;
shfl.sync.bfly.b32 %r1187|%p90, %r1186, %r1171, %r1170, %r1172;
mov.b32 %f298, %r1187;
add.ftz.f32 %f299, %f288, %f298;
mov.b32 %r1188, %f299;
shfl.sync.bfly.b32 %r1189|%p91, %r1188, %r1175, %r1170, %r1172;
mov.b32 %f300, %r1189;
add.ftz.f32 %f301, %f299, %f300;
mov.b32 %r1190, %f301;
shfl.sync.bfly.b32 %r1191|%p92, %r1190, %r1178, %r1170, %r1172;
mov.b32 %f302, %r1191;
add.ftz.f32 %f303, %f301, %f302;
mov.b32 %r1192, %f303;
shfl.sync.bfly.b32 %r1193|%p93, %r1192, %r1181, %r1170, %r1172;
mov.b32 %f304, %r1193;
add.ftz.f32 %f305, %f303, %f304;
mov.b32 %r1194, %f305;
shfl.sync.bfly.b32 %r1195|%p94, %r1194, %r1184, %r1170, %r1172;
mov.b32 %f306, %r1195;
add.ftz.f32 %f45, %f305, %f306;
setp.ne.s32 %p95, %r2, 2;
or.b32 %r1199, %r4, 2;
setp.ge.u32 %p96, %r1199, %r54;
or.pred %p97, %p95, %p96;
@%p97 bra $L__BB293_52;

ld.param.f32 %f979, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f307, %f5, %f44;
add.ftz.f32 %f308, %f4, %f45;
min.ftz.f32 %f309, %f308, %f979;
min.ftz.f32 %f310, %f307, %f979;
max.ftz.f32 %f311, %f310, %f35;
mul.ftz.f32 %f312, %f309, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f313, %f312;
add.ftz.f32 %f314, %f313, 0f3F800000;
div.approx.ftz.f32 %f315, %f309, %f314;
mul.ftz.f32 %f316, %f311, %f315;
st.global.f32 [%rd28+8], %f316;

$L__BB293_52:
ld.shared.f32 %f317, [%r35+384];
add.ftz.f32 %f318, %f317, %f27;
ld.shared.f32 %f319, [%r37+384];
add.ftz.f32 %f320, %f319, %f31;
ld.shared.f32 %f321, [%r35+896];
add.ftz.f32 %f322, %f321, %f318;
ld.shared.f32 %f323, [%r37+896];
add.ftz.f32 %f324, %f323, %f320;
ld.shared.f32 %f325, [%r35+1408];
add.ftz.f32 %f326, %f325, %f322;
ld.shared.f32 %f327, [%r37+1408];
add.ftz.f32 %f328, %f327, %f324;
mov.b32 %r1200, %f326;
shfl.sync.bfly.b32 %r1204|%p98, %r1200, %r1171, %r1170, %r1172;
mov.b32 %f329, %r1204;
add.ftz.f32 %f330, %f326, %f329;
mov.b32 %r1205, %f330;
shfl.sync.bfly.b32 %r1207|%p99, %r1205, %r1175, %r1170, %r1172;
mov.b32 %f331, %r1207;
add.ftz.f32 %f332, %f330, %f331;
mov.b32 %r1208, %f332;
shfl.sync.bfly.b32 %r1210|%p100, %r1208, %r1178, %r1170, %r1172;
mov.b32 %f333, %r1210;
add.ftz.f32 %f334, %f332, %f333;
mov.b32 %r1211, %f334;
shfl.sync.bfly.b32 %r1213|%p101, %r1211, %r1181, %r1170, %r1172;
mov.b32 %f335, %r1213;
add.ftz.f32 %f336, %f334, %f335;
mov.b32 %r1214, %f336;
shfl.sync.bfly.b32 %r1216|%p102, %r1214, %r1184, %r1170, %r1172;
mov.b32 %f337, %r1216;
add.ftz.f32 %f46, %f336, %f337;
mov.b32 %r1217, %f328;
shfl.sync.bfly.b32 %r1218|%p103, %r1217, %r1171, %r1170, %r1172;
mov.b32 %f338, %r1218;
add.ftz.f32 %f339, %f328, %f338;
mov.b32 %r1219, %f339;
shfl.sync.bfly.b32 %r1220|%p104, %r1219, %r1175, %r1170, %r1172;
mov.b32 %f340, %r1220;
add.ftz.f32 %f341, %f339, %f340;
mov.b32 %r1221, %f341;
shfl.sync.bfly.b32 %r1222|%p105, %r1221, %r1178, %r1170, %r1172;
mov.b32 %f342, %r1222;
add.ftz.f32 %f343, %f341, %f342;
mov.b32 %r1223, %f343;
shfl.sync.bfly.b32 %r1224|%p106, %r1223, %r1181, %r1170, %r1172;
mov.b32 %f344, %r1224;
add.ftz.f32 %f345, %f343, %f344;
mov.b32 %r1225, %f345;
shfl.sync.bfly.b32 %r1226|%p107, %r1225, %r1184, %r1170, %r1172;
mov.b32 %f346, %r1226;
add.ftz.f32 %f47, %f345, %f346;
setp.ne.s32 %p108, %r2, 3;
or.b32 %r1230, %r4, 3;
setp.ge.u32 %p109, %r1230, %r54;
or.pred %p110, %p108, %p109;
@%p110 bra $L__BB293_78;

ld.param.f32 %f978, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f347, %f5, %f46;
add.ftz.f32 %f348, %f4, %f47;
min.ftz.f32 %f349, %f348, %f978;
min.ftz.f32 %f350, %f347, %f978;
max.ftz.f32 %f351, %f350, %f35;
mul.ftz.f32 %f352, %f349, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f353, %f352;
add.ftz.f32 %f354, %f353, 0f3F800000;
div.approx.ftz.f32 %f355, %f349, %f354;
mul.ftz.f32 %f356, %f351, %f355;
st.global.f32 [%rd28+12], %f356;
bra.uni $L__BB293_78;

$L__BB293_70:
ld.shared.f32 %f845, [%r35];
add.ftz.f32 %f846, %f845, %f30;
ld.shared.f32 %f847, [%r37];
add.ftz.f32 %f848, %f847, %f34;
ld.shared.f32 %f849, [%r35+512];
add.ftz.f32 %f850, %f849, %f846;
ld.shared.f32 %f851, [%r37+512];
add.ftz.f32 %f852, %f851, %f848;
ld.shared.f32 %f853, [%r35+1024];
add.ftz.f32 %f854, %f853, %f850;
ld.shared.f32 %f855, [%r37+1024];
add.ftz.f32 %f856, %f855, %f852;
mov.b32 %r1600, %f854;
mov.u32 %r1601, 31;
mov.u32 %r1602, 16;
mov.u32 %r1603, -1;
shfl.sync.bfly.b32 %r1604|%p267, %r1600, %r1602, %r1601, %r1603;
mov.b32 %f857, %r1604;
add.ftz.f32 %f858, %f854, %f857;
mov.b32 %r1605, %f858;
mov.u32 %r1606, 8;
shfl.sync.bfly.b32 %r1607|%p268, %r1605, %r1606, %r1601, %r1603;
mov.b32 %f859, %r1607;
add.ftz.f32 %f860, %f858, %f859;
mov.b32 %r1608, %f860;
mov.u32 %r1609, 4;
shfl.sync.bfly.b32 %r1610|%p269, %r1608, %r1609, %r1601, %r1603;
mov.b32 %f861, %r1610;
add.ftz.f32 %f862, %f860, %f861;
mov.b32 %r1611, %f862;
mov.u32 %r1612, 2;
shfl.sync.bfly.b32 %r1613|%p270, %r1611, %r1612, %r1601, %r1603;
mov.b32 %f863, %r1613;
add.ftz.f32 %f864, %f862, %f863;
mov.b32 %r1614, %f864;
mov.u32 %r1615, 1;
shfl.sync.bfly.b32 %r1616|%p271, %r1614, %r1615, %r1601, %r1603;
mov.b32 %f865, %r1616;
add.ftz.f32 %f72, %f864, %f865;
mov.b32 %r1617, %f856;
shfl.sync.bfly.b32 %r1618|%p272, %r1617, %r1602, %r1601, %r1603;
mov.b32 %f866, %r1618;
add.ftz.f32 %f867, %f856, %f866;
mov.b32 %r1619, %f867;
shfl.sync.bfly.b32 %r1620|%p273, %r1619, %r1606, %r1601, %r1603;
mov.b32 %f868, %r1620;
add.ftz.f32 %f869, %f867, %f868;
mov.b32 %r1621, %f869;
shfl.sync.bfly.b32 %r1622|%p274, %r1621, %r1609, %r1601, %r1603;
mov.b32 %f870, %r1622;
add.ftz.f32 %f871, %f869, %f870;
mov.b32 %r1623, %f871;
shfl.sync.bfly.b32 %r1624|%p275, %r1623, %r1612, %r1601, %r1603;
mov.b32 %f872, %r1624;
add.ftz.f32 %f873, %f871, %f872;
mov.b32 %r1625, %f873;
shfl.sync.bfly.b32 %r1626|%p276, %r1625, %r1615, %r1601, %r1603;
mov.b32 %f874, %r1626;
add.ftz.f32 %f73, %f873, %f874;
setp.ne.s32 %p277, %r2, 0;
setp.ge.u32 %p278, %r4, %r54;
or.pred %p279, %p277, %p278;
@%p279 bra $L__BB293_72;

add.ftz.f32 %f875, %f5, %f72;
add.ftz.f32 %f876, %f4, %f73;
mul.ftz.f32 %f877, %f875, %f876;
st.global.f32 [%rd28], %f877;

$L__BB293_72:
ld.shared.f32 %f878, [%r35+128];
add.ftz.f32 %f879, %f878, %f29;
ld.shared.f32 %f880, [%r37+128];
add.ftz.f32 %f881, %f880, %f33;
ld.shared.f32 %f882, [%r35+640];
add.ftz.f32 %f883, %f882, %f879;
ld.shared.f32 %f884, [%r37+640];
add.ftz.f32 %f885, %f884, %f881;
ld.shared.f32 %f886, [%r35+1152];
add.ftz.f32 %f887, %f886, %f883;
ld.shared.f32 %f888, [%r37+1152];
add.ftz.f32 %f889, %f888, %f885;
mov.b32 %r1630, %f887;
shfl.sync.bfly.b32 %r1634|%p280, %r1630, %r1602, %r1601, %r1603;
mov.b32 %f890, %r1634;
add.ftz.f32 %f891, %f887, %f890;
mov.b32 %r1635, %f891;
shfl.sync.bfly.b32 %r1637|%p281, %r1635, %r1606, %r1601, %r1603;
mov.b32 %f892, %r1637;
add.ftz.f32 %f893, %f891, %f892;
mov.b32 %r1638, %f893;
shfl.sync.bfly.b32 %r1640|%p282, %r1638, %r1609, %r1601, %r1603;
mov.b32 %f894, %r1640;
add.ftz.f32 %f895, %f893, %f894;
mov.b32 %r1641, %f895;
shfl.sync.bfly.b32 %r1643|%p283, %r1641, %r1612, %r1601, %r1603;
mov.b32 %f896, %r1643;
add.ftz.f32 %f897, %f895, %f896;
mov.b32 %r1644, %f897;
shfl.sync.bfly.b32 %r1646|%p284, %r1644, %r1615, %r1601, %r1603;
mov.b32 %f898, %r1646;
add.ftz.f32 %f74, %f897, %f898;
mov.b32 %r1647, %f889;
shfl.sync.bfly.b32 %r1648|%p285, %r1647, %r1602, %r1601, %r1603;
mov.b32 %f899, %r1648;
add.ftz.f32 %f900, %f889, %f899;
mov.b32 %r1649, %f900;
shfl.sync.bfly.b32 %r1650|%p286, %r1649, %r1606, %r1601, %r1603;
mov.b32 %f901, %r1650;
add.ftz.f32 %f902, %f900, %f901;
mov.b32 %r1651, %f902;
shfl.sync.bfly.b32 %r1652|%p287, %r1651, %r1609, %r1601, %r1603;
mov.b32 %f903, %r1652;
add.ftz.f32 %f904, %f902, %f903;
mov.b32 %r1653, %f904;
shfl.sync.bfly.b32 %r1654|%p288, %r1653, %r1612, %r1601, %r1603;
mov.b32 %f905, %r1654;
add.ftz.f32 %f906, %f904, %f905;
mov.b32 %r1655, %f906;
shfl.sync.bfly.b32 %r1656|%p289, %r1655, %r1615, %r1601, %r1603;
mov.b32 %f907, %r1656;
add.ftz.f32 %f75, %f906, %f907;
setp.ne.s32 %p290, %r2, 1;
or.b32 %r1660, %r4, 1;
setp.ge.u32 %p291, %r1660, %r54;
or.pred %p292, %p290, %p291;
@%p292 bra $L__BB293_74;

add.ftz.f32 %f908, %f5, %f74;
add.ftz.f32 %f909, %f4, %f75;
mul.ftz.f32 %f910, %f908, %f909;
st.global.f32 [%rd28+4], %f910;

$L__BB293_74:
ld.shared.f32 %f911, [%r35+256];
add.ftz.f32 %f912, %f911, %f28;
ld.shared.f32 %f913, [%r37+256];
add.ftz.f32 %f914, %f913, %f32;
ld.shared.f32 %f915, [%r35+768];
add.ftz.f32 %f916, %f915, %f912;
ld.shared.f32 %f917, [%r37+768];
add.ftz.f32 %f918, %f917, %f914;
ld.shared.f32 %f919, [%r35+1280];
add.ftz.f32 %f920, %f919, %f916;
ld.shared.f32 %f921, [%r37+1280];
add.ftz.f32 %f922, %f921, %f918;
mov.b32 %r1661, %f920;
mov.u32 %r1662, 31;
mov.u32 %r1663, 16;
mov.u32 %r1664, -1;
shfl.sync.bfly.b32 %r1665|%p293, %r1661, %r1663, %r1662, %r1664;
mov.b32 %f923, %r1665;
add.ftz.f32 %f924, %f920, %f923;
mov.b32 %r1666, %f924;
mov.u32 %r1667, 8;
shfl.sync.bfly.b32 %r1668|%p294, %r1666, %r1667, %r1662, %r1664;
mov.b32 %f925, %r1668;
add.ftz.f32 %f926, %f924, %f925;
mov.b32 %r1669, %f926;
mov.u32 %r1670, 4;
shfl.sync.bfly.b32 %r1671|%p295, %r1669, %r1670, %r1662, %r1664;
mov.b32 %f927, %r1671;
add.ftz.f32 %f928, %f926, %f927;
mov.b32 %r1672, %f928;
mov.u32 %r1673, 2;
shfl.sync.bfly.b32 %r1674|%p296, %r1672, %r1673, %r1662, %r1664;
mov.b32 %f929, %r1674;
add.ftz.f32 %f930, %f928, %f929;
mov.b32 %r1675, %f930;
mov.u32 %r1676, 1;
shfl.sync.bfly.b32 %r1677|%p297, %r1675, %r1676, %r1662, %r1664;
mov.b32 %f931, %r1677;
add.ftz.f32 %f76, %f930, %f931;
mov.b32 %r1678, %f922;
shfl.sync.bfly.b32 %r1679|%p298, %r1678, %r1663, %r1662, %r1664;
mov.b32 %f932, %r1679;
add.ftz.f32 %f933, %f922, %f932;
mov.b32 %r1680, %f933;
shfl.sync.bfly.b32 %r1681|%p299, %r1680, %r1667, %r1662, %r1664;
mov.b32 %f934, %r1681;
add.ftz.f32 %f935, %f933, %f934;
mov.b32 %r1682, %f935;
shfl.sync.bfly.b32 %r1683|%p300, %r1682, %r1670, %r1662, %r1664;
mov.b32 %f936, %r1683;
add.ftz.f32 %f937, %f935, %f936;
mov.b32 %r1684, %f937;
shfl.sync.bfly.b32 %r1685|%p301, %r1684, %r1673, %r1662, %r1664;
mov.b32 %f938, %r1685;
add.ftz.f32 %f939, %f937, %f938;
mov.b32 %r1686, %f939;
shfl.sync.bfly.b32 %r1687|%p302, %r1686, %r1676, %r1662, %r1664;
mov.b32 %f940, %r1687;
add.ftz.f32 %f77, %f939, %f940;
setp.ne.s32 %p303, %r2, 2;
or.b32 %r1691, %r4, 2;
setp.ge.u32 %p304, %r1691, %r54;
or.pred %p305, %p303, %p304;
@%p305 bra $L__BB293_76;

add.ftz.f32 %f941, %f5, %f76;
add.ftz.f32 %f942, %f4, %f77;
mul.ftz.f32 %f943, %f941, %f942;
st.global.f32 [%rd28+8], %f943;

$L__BB293_76:
ld.shared.f32 %f944, [%r35+384];
add.ftz.f32 %f945, %f944, %f27;
ld.shared.f32 %f946, [%r37+384];
add.ftz.f32 %f947, %f946, %f31;
ld.shared.f32 %f948, [%r35+896];
add.ftz.f32 %f949, %f948, %f945;
ld.shared.f32 %f950, [%r37+896];
add.ftz.f32 %f951, %f950, %f947;
ld.shared.f32 %f952, [%r35+1408];
add.ftz.f32 %f953, %f952, %f949;
ld.shared.f32 %f954, [%r37+1408];
add.ftz.f32 %f955, %f954, %f951;
mov.b32 %r1692, %f953;
shfl.sync.bfly.b32 %r1696|%p306, %r1692, %r1663, %r1662, %r1664;
mov.b32 %f956, %r1696;
add.ftz.f32 %f957, %f953, %f956;
mov.b32 %r1697, %f957;
shfl.sync.bfly.b32 %r1699|%p307, %r1697, %r1667, %r1662, %r1664;
mov.b32 %f958, %r1699;
add.ftz.f32 %f959, %f957, %f958;
mov.b32 %r1700, %f959;
shfl.sync.bfly.b32 %r1702|%p308, %r1700, %r1670, %r1662, %r1664;
mov.b32 %f960, %r1702;
add.ftz.f32 %f961, %f959, %f960;
mov.b32 %r1703, %f961;
shfl.sync.bfly.b32 %r1705|%p309, %r1703, %r1673, %r1662, %r1664;
mov.b32 %f962, %r1705;
add.ftz.f32 %f963, %f961, %f962;
mov.b32 %r1706, %f963;
shfl.sync.bfly.b32 %r1708|%p310, %r1706, %r1676, %r1662, %r1664;
mov.b32 %f964, %r1708;
add.ftz.f32 %f78, %f963, %f964;
mov.b32 %r1709, %f955;
shfl.sync.bfly.b32 %r1710|%p311, %r1709, %r1663, %r1662, %r1664;
mov.b32 %f965, %r1710;
add.ftz.f32 %f966, %f955, %f965;
mov.b32 %r1711, %f966;
shfl.sync.bfly.b32 %r1712|%p312, %r1711, %r1667, %r1662, %r1664;
mov.b32 %f967, %r1712;
add.ftz.f32 %f968, %f966, %f967;
mov.b32 %r1713, %f968;
shfl.sync.bfly.b32 %r1714|%p313, %r1713, %r1670, %r1662, %r1664;
mov.b32 %f969, %r1714;
add.ftz.f32 %f970, %f968, %f969;
mov.b32 %r1715, %f970;
shfl.sync.bfly.b32 %r1716|%p314, %r1715, %r1673, %r1662, %r1664;
mov.b32 %f971, %r1716;
add.ftz.f32 %f972, %f970, %f971;
mov.b32 %r1717, %f972;
shfl.sync.bfly.b32 %r1718|%p315, %r1717, %r1676, %r1662, %r1664;
mov.b32 %f973, %r1718;
add.ftz.f32 %f79, %f972, %f973;
setp.ne.s32 %p316, %r2, 3;
or.b32 %r1722, %r4, 3;
setp.ge.u32 %p317, %r1722, %r54;
or.pred %p318, %p316, %p317;
@%p318 bra $L__BB293_78;

add.ftz.f32 %f974, %f5, %f78;
add.ftz.f32 %f975, %f4, %f79;
mul.ftz.f32 %f976, %f974, %f975;
st.global.f32 [%rd28+12], %f976;
bra.uni $L__BB293_78;

$L__BB293_62:
ld.shared.f32 %f681, [%r35];
add.ftz.f32 %f682, %f681, %f30;
ld.shared.f32 %f683, [%r37];
add.ftz.f32 %f684, %f683, %f34;
ld.shared.f32 %f685, [%r35+512];
add.ftz.f32 %f686, %f685, %f682;
ld.shared.f32 %f687, [%r37+512];
add.ftz.f32 %f688, %f687, %f684;
ld.shared.f32 %f689, [%r35+1024];
add.ftz.f32 %f690, %f689, %f686;
ld.shared.f32 %f691, [%r37+1024];
add.ftz.f32 %f692, %f691, %f688;
mov.b32 %r1477, %f690;
mov.u32 %r1478, 31;
mov.u32 %r1479, 16;
mov.u32 %r1480, -1;
shfl.sync.bfly.b32 %r1481|%p215, %r1477, %r1479, %r1478, %r1480;
mov.b32 %f693, %r1481;
add.ftz.f32 %f694, %f690, %f693;
mov.b32 %r1482, %f694;
mov.u32 %r1483, 8;
shfl.sync.bfly.b32 %r1484|%p216, %r1482, %r1483, %r1478, %r1480;
mov.b32 %f695, %r1484;
add.ftz.f32 %f696, %f694, %f695;
mov.b32 %r1485, %f696;
mov.u32 %r1486, 4;
shfl.sync.bfly.b32 %r1487|%p217, %r1485, %r1486, %r1478, %r1480;
mov.b32 %f697, %r1487;
add.ftz.f32 %f698, %f696, %f697;
mov.b32 %r1488, %f698;
mov.u32 %r1489, 2;
shfl.sync.bfly.b32 %r1490|%p218, %r1488, %r1489, %r1478, %r1480;
mov.b32 %f699, %r1490;
add.ftz.f32 %f700, %f698, %f699;
mov.b32 %r1491, %f700;
mov.u32 %r1492, 1;
shfl.sync.bfly.b32 %r1493|%p219, %r1491, %r1492, %r1478, %r1480;
mov.b32 %f701, %r1493;
add.ftz.f32 %f64, %f700, %f701;
mov.b32 %r1494, %f692;
shfl.sync.bfly.b32 %r1495|%p220, %r1494, %r1479, %r1478, %r1480;
mov.b32 %f702, %r1495;
add.ftz.f32 %f703, %f692, %f702;
mov.b32 %r1496, %f703;
shfl.sync.bfly.b32 %r1497|%p221, %r1496, %r1483, %r1478, %r1480;
mov.b32 %f704, %r1497;
add.ftz.f32 %f705, %f703, %f704;
mov.b32 %r1498, %f705;
shfl.sync.bfly.b32 %r1499|%p222, %r1498, %r1486, %r1478, %r1480;
mov.b32 %f706, %r1499;
add.ftz.f32 %f707, %f705, %f706;
mov.b32 %r1500, %f707;
shfl.sync.bfly.b32 %r1501|%p223, %r1500, %r1489, %r1478, %r1480;
mov.b32 %f708, %r1501;
add.ftz.f32 %f709, %f707, %f708;
mov.b32 %r1502, %f709;
shfl.sync.bfly.b32 %r1503|%p224, %r1502, %r1492, %r1478, %r1480;
mov.b32 %f710, %r1503;
add.ftz.f32 %f65, %f709, %f710;
setp.ne.s32 %p225, %r2, 0;
setp.ge.u32 %p226, %r4, %r54;
or.pred %p227, %p225, %p226;
@%p227 bra $L__BB293_64;

add.ftz.f32 %f711, %f5, %f64;
add.ftz.f32 %f712, %f4, %f65;
mul.ftz.f32 %f713, %f712, 0f3F000000;
mul.ftz.f32 %f714, %f712, 0f3F4C422A;
mul.ftz.f32 %f715, %f712, 0f3D372713;
fma.rn.ftz.f32 %f716, %f712, %f715, 0f3F800000;
mul.ftz.f32 %f717, %f714, %f716;
tanh.approx.f32 %f718, %f717;
add.ftz.f32 %f719, %f718, 0f3F800000;
mul.ftz.f32 %f720, %f713, %f719;
mul.ftz.f32 %f721, %f711, %f720;
st.global.f32 [%rd28], %f721;

$L__BB293_64:
ld.shared.f32 %f722, [%r35+128];
add.ftz.f32 %f723, %f722, %f29;
ld.shared.f32 %f724, [%r37+128];
add.ftz.f32 %f725, %f724, %f33;
ld.shared.f32 %f726, [%r35+640];
add.ftz.f32 %f727, %f726, %f723;
ld.shared.f32 %f728, [%r37+640];
add.ftz.f32 %f729, %f728, %f725;
ld.shared.f32 %f730, [%r35+1152];
add.ftz.f32 %f731, %f730, %f727;
ld.shared.f32 %f732, [%r37+1152];
add.ftz.f32 %f733, %f732, %f729;
mov.b32 %r1507, %f731;
shfl.sync.bfly.b32 %r1511|%p228, %r1507, %r1479, %r1478, %r1480;
mov.b32 %f734, %r1511;
add.ftz.f32 %f735, %f731, %f734;
mov.b32 %r1512, %f735;
shfl.sync.bfly.b32 %r1514|%p229, %r1512, %r1483, %r1478, %r1480;
mov.b32 %f736, %r1514;
add.ftz.f32 %f737, %f735, %f736;
mov.b32 %r1515, %f737;
shfl.sync.bfly.b32 %r1517|%p230, %r1515, %r1486, %r1478, %r1480;
mov.b32 %f738, %r1517;
add.ftz.f32 %f739, %f737, %f738;
mov.b32 %r1518, %f739;
shfl.sync.bfly.b32 %r1520|%p231, %r1518, %r1489, %r1478, %r1480;
mov.b32 %f740, %r1520;
add.ftz.f32 %f741, %f739, %f740;
mov.b32 %r1521, %f741;
shfl.sync.bfly.b32 %r1523|%p232, %r1521, %r1492, %r1478, %r1480;
mov.b32 %f742, %r1523;
add.ftz.f32 %f66, %f741, %f742;
mov.b32 %r1524, %f733;
shfl.sync.bfly.b32 %r1525|%p233, %r1524, %r1479, %r1478, %r1480;
mov.b32 %f743, %r1525;
add.ftz.f32 %f744, %f733, %f743;
mov.b32 %r1526, %f744;
shfl.sync.bfly.b32 %r1527|%p234, %r1526, %r1483, %r1478, %r1480;
mov.b32 %f745, %r1527;
add.ftz.f32 %f746, %f744, %f745;
mov.b32 %r1528, %f746;
shfl.sync.bfly.b32 %r1529|%p235, %r1528, %r1486, %r1478, %r1480;
mov.b32 %f747, %r1529;
add.ftz.f32 %f748, %f746, %f747;
mov.b32 %r1530, %f748;
shfl.sync.bfly.b32 %r1531|%p236, %r1530, %r1489, %r1478, %r1480;
mov.b32 %f749, %r1531;
add.ftz.f32 %f750, %f748, %f749;
mov.b32 %r1532, %f750;
shfl.sync.bfly.b32 %r1533|%p237, %r1532, %r1492, %r1478, %r1480;
mov.b32 %f751, %r1533;
add.ftz.f32 %f67, %f750, %f751;
setp.ne.s32 %p238, %r2, 1;
or.b32 %r1537, %r4, 1;
setp.ge.u32 %p239, %r1537, %r54;
or.pred %p240, %p238, %p239;
@%p240 bra $L__BB293_66;

add.ftz.f32 %f752, %f5, %f66;
add.ftz.f32 %f753, %f4, %f67;
mul.ftz.f32 %f754, %f753, 0f3F000000;
mul.ftz.f32 %f755, %f753, 0f3F4C422A;
mul.ftz.f32 %f756, %f753, 0f3D372713;
fma.rn.ftz.f32 %f757, %f753, %f756, 0f3F800000;
mul.ftz.f32 %f758, %f755, %f757;
tanh.approx.f32 %f759, %f758;
add.ftz.f32 %f760, %f759, 0f3F800000;
mul.ftz.f32 %f761, %f754, %f760;
mul.ftz.f32 %f762, %f752, %f761;
st.global.f32 [%rd28+4], %f762;

$L__BB293_66:
ld.shared.f32 %f763, [%r35+256];
add.ftz.f32 %f764, %f763, %f28;
ld.shared.f32 %f765, [%r37+256];
add.ftz.f32 %f766, %f765, %f32;
ld.shared.f32 %f767, [%r35+768];
add.ftz.f32 %f768, %f767, %f764;
ld.shared.f32 %f769, [%r37+768];
add.ftz.f32 %f770, %f769, %f766;
ld.shared.f32 %f771, [%r35+1280];
add.ftz.f32 %f772, %f771, %f768;
ld.shared.f32 %f773, [%r37+1280];
add.ftz.f32 %f774, %f773, %f770;
mov.b32 %r1538, %f772;
mov.u32 %r1539, 31;
mov.u32 %r1540, 16;
mov.u32 %r1541, -1;
shfl.sync.bfly.b32 %r1542|%p241, %r1538, %r1540, %r1539, %r1541;
mov.b32 %f775, %r1542;
add.ftz.f32 %f776, %f772, %f775;
mov.b32 %r1543, %f776;
mov.u32 %r1544, 8;
shfl.sync.bfly.b32 %r1545|%p242, %r1543, %r1544, %r1539, %r1541;
mov.b32 %f777, %r1545;
add.ftz.f32 %f778, %f776, %f777;
mov.b32 %r1546, %f778;
mov.u32 %r1547, 4;
shfl.sync.bfly.b32 %r1548|%p243, %r1546, %r1547, %r1539, %r1541;
mov.b32 %f779, %r1548;
add.ftz.f32 %f780, %f778, %f779;
mov.b32 %r1549, %f780;
mov.u32 %r1550, 2;
shfl.sync.bfly.b32 %r1551|%p244, %r1549, %r1550, %r1539, %r1541;
mov.b32 %f781, %r1551;
add.ftz.f32 %f782, %f780, %f781;
mov.b32 %r1552, %f782;
mov.u32 %r1553, 1;
shfl.sync.bfly.b32 %r1554|%p245, %r1552, %r1553, %r1539, %r1541;
mov.b32 %f783, %r1554;
add.ftz.f32 %f68, %f782, %f783;
mov.b32 %r1555, %f774;
shfl.sync.bfly.b32 %r1556|%p246, %r1555, %r1540, %r1539, %r1541;
mov.b32 %f784, %r1556;
add.ftz.f32 %f785, %f774, %f784;
mov.b32 %r1557, %f785;
shfl.sync.bfly.b32 %r1558|%p247, %r1557, %r1544, %r1539, %r1541;
mov.b32 %f786, %r1558;
add.ftz.f32 %f787, %f785, %f786;
mov.b32 %r1559, %f787;
shfl.sync.bfly.b32 %r1560|%p248, %r1559, %r1547, %r1539, %r1541;
mov.b32 %f788, %r1560;
add.ftz.f32 %f789, %f787, %f788;
mov.b32 %r1561, %f789;
shfl.sync.bfly.b32 %r1562|%p249, %r1561, %r1550, %r1539, %r1541;
mov.b32 %f790, %r1562;
add.ftz.f32 %f791, %f789, %f790;
mov.b32 %r1563, %f791;
shfl.sync.bfly.b32 %r1564|%p250, %r1563, %r1553, %r1539, %r1541;
mov.b32 %f792, %r1564;
add.ftz.f32 %f69, %f791, %f792;
setp.ne.s32 %p251, %r2, 2;
or.b32 %r1568, %r4, 2;
setp.ge.u32 %p252, %r1568, %r54;
or.pred %p253, %p251, %p252;
@%p253 bra $L__BB293_68;

add.ftz.f32 %f793, %f5, %f68;
add.ftz.f32 %f794, %f4, %f69;
mul.ftz.f32 %f795, %f794, 0f3F000000;
mul.ftz.f32 %f796, %f794, 0f3F4C422A;
mul.ftz.f32 %f797, %f794, 0f3D372713;
fma.rn.ftz.f32 %f798, %f794, %f797, 0f3F800000;
mul.ftz.f32 %f799, %f796, %f798;
tanh.approx.f32 %f800, %f799;
add.ftz.f32 %f801, %f800, 0f3F800000;
mul.ftz.f32 %f802, %f795, %f801;
mul.ftz.f32 %f803, %f793, %f802;
st.global.f32 [%rd28+8], %f803;

$L__BB293_68:
ld.shared.f32 %f804, [%r35+384];
add.ftz.f32 %f805, %f804, %f27;
ld.shared.f32 %f806, [%r37+384];
add.ftz.f32 %f807, %f806, %f31;
ld.shared.f32 %f808, [%r35+896];
add.ftz.f32 %f809, %f808, %f805;
ld.shared.f32 %f810, [%r37+896];
add.ftz.f32 %f811, %f810, %f807;
ld.shared.f32 %f812, [%r35+1408];
add.ftz.f32 %f813, %f812, %f809;
ld.shared.f32 %f814, [%r37+1408];
add.ftz.f32 %f815, %f814, %f811;
mov.b32 %r1569, %f813;
shfl.sync.bfly.b32 %r1573|%p254, %r1569, %r1540, %r1539, %r1541;
mov.b32 %f816, %r1573;
add.ftz.f32 %f817, %f813, %f816;
mov.b32 %r1574, %f817;
shfl.sync.bfly.b32 %r1576|%p255, %r1574, %r1544, %r1539, %r1541;
mov.b32 %f818, %r1576;
add.ftz.f32 %f819, %f817, %f818;
mov.b32 %r1577, %f819;
shfl.sync.bfly.b32 %r1579|%p256, %r1577, %r1547, %r1539, %r1541;
mov.b32 %f820, %r1579;
add.ftz.f32 %f821, %f819, %f820;
mov.b32 %r1580, %f821;
shfl.sync.bfly.b32 %r1582|%p257, %r1580, %r1550, %r1539, %r1541;
mov.b32 %f822, %r1582;
add.ftz.f32 %f823, %f821, %f822;
mov.b32 %r1583, %f823;
shfl.sync.bfly.b32 %r1585|%p258, %r1583, %r1553, %r1539, %r1541;
mov.b32 %f824, %r1585;
add.ftz.f32 %f70, %f823, %f824;
mov.b32 %r1586, %f815;
shfl.sync.bfly.b32 %r1587|%p259, %r1586, %r1540, %r1539, %r1541;
mov.b32 %f825, %r1587;
add.ftz.f32 %f826, %f815, %f825;
mov.b32 %r1588, %f826;
shfl.sync.bfly.b32 %r1589|%p260, %r1588, %r1544, %r1539, %r1541;
mov.b32 %f827, %r1589;
add.ftz.f32 %f828, %f826, %f827;
mov.b32 %r1590, %f828;
shfl.sync.bfly.b32 %r1591|%p261, %r1590, %r1547, %r1539, %r1541;
mov.b32 %f829, %r1591;
add.ftz.f32 %f830, %f828, %f829;
mov.b32 %r1592, %f830;
shfl.sync.bfly.b32 %r1593|%p262, %r1592, %r1550, %r1539, %r1541;
mov.b32 %f831, %r1593;
add.ftz.f32 %f832, %f830, %f831;
mov.b32 %r1594, %f832;
shfl.sync.bfly.b32 %r1595|%p263, %r1594, %r1553, %r1539, %r1541;
mov.b32 %f833, %r1595;
add.ftz.f32 %f71, %f832, %f833;
setp.ne.s32 %p264, %r2, 3;
or.b32 %r1599, %r4, 3;
setp.ge.u32 %p265, %r1599, %r54;
or.pred %p266, %p264, %p265;
@%p266 bra $L__BB293_78;

add.ftz.f32 %f834, %f5, %f70;
add.ftz.f32 %f835, %f4, %f71;
mul.ftz.f32 %f836, %f835, 0f3F000000;
mul.ftz.f32 %f837, %f835, 0f3F4C422A;
mul.ftz.f32 %f838, %f835, 0f3D372713;
fma.rn.ftz.f32 %f839, %f835, %f838, 0f3F800000;
mul.ftz.f32 %f840, %f837, %f839;
tanh.approx.f32 %f841, %f840;
add.ftz.f32 %f842, %f841, 0f3F800000;
mul.ftz.f32 %f843, %f836, %f842;
mul.ftz.f32 %f844, %f834, %f843;
st.global.f32 [%rd28+12], %f844;
bra.uni $L__BB293_78;

$L__BB293_54:
ld.shared.f32 %f357, [%r35];
add.ftz.f32 %f358, %f357, %f30;
ld.shared.f32 %f359, [%r37];
add.ftz.f32 %f360, %f359, %f34;
ld.shared.f32 %f361, [%r35+512];
add.ftz.f32 %f362, %f361, %f358;
ld.shared.f32 %f363, [%r37+512];
add.ftz.f32 %f364, %f363, %f360;
ld.shared.f32 %f365, [%r35+1024];
add.ftz.f32 %f366, %f365, %f362;
ld.shared.f32 %f367, [%r37+1024];
add.ftz.f32 %f368, %f367, %f364;
mov.b32 %r1231, %f366;
mov.u32 %r1232, 31;
mov.u32 %r1233, 16;
mov.u32 %r1234, -1;
shfl.sync.bfly.b32 %r1235|%p111, %r1231, %r1233, %r1232, %r1234;
mov.b32 %f369, %r1235;
add.ftz.f32 %f370, %f366, %f369;
mov.b32 %r1236, %f370;
mov.u32 %r1237, 8;
shfl.sync.bfly.b32 %r1238|%p112, %r1236, %r1237, %r1232, %r1234;
mov.b32 %f371, %r1238;
add.ftz.f32 %f372, %f370, %f371;
mov.b32 %r1239, %f372;
mov.u32 %r1240, 4;
shfl.sync.bfly.b32 %r1241|%p113, %r1239, %r1240, %r1232, %r1234;
mov.b32 %f373, %r1241;
add.ftz.f32 %f374, %f372, %f373;
mov.b32 %r1242, %f374;
mov.u32 %r1243, 2;
shfl.sync.bfly.b32 %r1244|%p114, %r1242, %r1243, %r1232, %r1234;
mov.b32 %f375, %r1244;
add.ftz.f32 %f376, %f374, %f375;
mov.b32 %r1245, %f376;
mov.u32 %r1246, 1;
shfl.sync.bfly.b32 %r1247|%p115, %r1245, %r1246, %r1232, %r1234;
mov.b32 %f377, %r1247;
add.ftz.f32 %f48, %f376, %f377;
mov.b32 %r1248, %f368;
shfl.sync.bfly.b32 %r1249|%p116, %r1248, %r1233, %r1232, %r1234;
mov.b32 %f378, %r1249;
add.ftz.f32 %f379, %f368, %f378;
mov.b32 %r1250, %f379;
shfl.sync.bfly.b32 %r1251|%p117, %r1250, %r1237, %r1232, %r1234;
mov.b32 %f380, %r1251;
add.ftz.f32 %f381, %f379, %f380;
mov.b32 %r1252, %f381;
shfl.sync.bfly.b32 %r1253|%p118, %r1252, %r1240, %r1232, %r1234;
mov.b32 %f382, %r1253;
add.ftz.f32 %f383, %f381, %f382;
mov.b32 %r1254, %f383;
shfl.sync.bfly.b32 %r1255|%p119, %r1254, %r1243, %r1232, %r1234;
mov.b32 %f384, %r1255;
add.ftz.f32 %f385, %f383, %f384;
mov.b32 %r1256, %f385;
shfl.sync.bfly.b32 %r1257|%p120, %r1256, %r1246, %r1232, %r1234;
mov.b32 %f386, %r1257;
add.ftz.f32 %f49, %f385, %f386;
setp.ne.s32 %p121, %r2, 0;
setp.ge.u32 %p122, %r4, %r54;
or.pred %p123, %p121, %p122;
@%p123 bra $L__BB293_56;

add.ftz.f32 %f387, %f5, %f48;
add.ftz.f32 %f388, %f4, %f49;
mov.f32 %f389, 0f40E00000;
min.ftz.f32 %f390, %f388, %f389;
min.ftz.f32 %f391, %f387, %f389;
mov.f32 %f392, 0fC0E00000;
max.ftz.f32 %f393, %f391, %f392;
mul.ftz.f32 %f394, %f390, 0f3FD9DB23;
mul.ftz.f32 %f395, %f394, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f396, %f395;
add.ftz.f32 %f397, %f396, 0f3F800000;
div.approx.ftz.f32 %f398, %f390, %f397;
add.ftz.f32 %f399, %f393, 0f3F800000;
mul.ftz.f32 %f400, %f399, %f398;
st.global.f32 [%rd28], %f400;

$L__BB293_56:
ld.shared.f32 %f401, [%r35+128];
add.ftz.f32 %f402, %f401, %f29;
ld.shared.f32 %f403, [%r37+128];
add.ftz.f32 %f404, %f403, %f33;
ld.shared.f32 %f405, [%r35+640];
add.ftz.f32 %f406, %f405, %f402;
ld.shared.f32 %f407, [%r37+640];
add.ftz.f32 %f408, %f407, %f404;
ld.shared.f32 %f409, [%r35+1152];
add.ftz.f32 %f410, %f409, %f406;
ld.shared.f32 %f411, [%r37+1152];
add.ftz.f32 %f412, %f411, %f408;
mov.b32 %r1261, %f410;
shfl.sync.bfly.b32 %r1265|%p124, %r1261, %r1233, %r1232, %r1234;
mov.b32 %f413, %r1265;
add.ftz.f32 %f414, %f410, %f413;
mov.b32 %r1266, %f414;
shfl.sync.bfly.b32 %r1268|%p125, %r1266, %r1237, %r1232, %r1234;
mov.b32 %f415, %r1268;
add.ftz.f32 %f416, %f414, %f415;
mov.b32 %r1269, %f416;
shfl.sync.bfly.b32 %r1271|%p126, %r1269, %r1240, %r1232, %r1234;
mov.b32 %f417, %r1271;
add.ftz.f32 %f418, %f416, %f417;
mov.b32 %r1272, %f418;
shfl.sync.bfly.b32 %r1274|%p127, %r1272, %r1243, %r1232, %r1234;
mov.b32 %f419, %r1274;
add.ftz.f32 %f420, %f418, %f419;
mov.b32 %r1275, %f420;
shfl.sync.bfly.b32 %r1277|%p128, %r1275, %r1246, %r1232, %r1234;
mov.b32 %f421, %r1277;
add.ftz.f32 %f50, %f420, %f421;
mov.b32 %r1278, %f412;
shfl.sync.bfly.b32 %r1279|%p129, %r1278, %r1233, %r1232, %r1234;
mov.b32 %f422, %r1279;
add.ftz.f32 %f423, %f412, %f422;
mov.b32 %r1280, %f423;
shfl.sync.bfly.b32 %r1281|%p130, %r1280, %r1237, %r1232, %r1234;
mov.b32 %f424, %r1281;
add.ftz.f32 %f425, %f423, %f424;
mov.b32 %r1282, %f425;
shfl.sync.bfly.b32 %r1283|%p131, %r1282, %r1240, %r1232, %r1234;
mov.b32 %f426, %r1283;
add.ftz.f32 %f427, %f425, %f426;
mov.b32 %r1284, %f427;
shfl.sync.bfly.b32 %r1285|%p132, %r1284, %r1243, %r1232, %r1234;
mov.b32 %f428, %r1285;
add.ftz.f32 %f429, %f427, %f428;
mov.b32 %r1286, %f429;
shfl.sync.bfly.b32 %r1287|%p133, %r1286, %r1246, %r1232, %r1234;
mov.b32 %f430, %r1287;
add.ftz.f32 %f51, %f429, %f430;
setp.ne.s32 %p134, %r2, 1;
or.b32 %r1291, %r4, 1;
setp.ge.u32 %p135, %r1291, %r54;
or.pred %p136, %p134, %p135;
@%p136 bra $L__BB293_58;

add.ftz.f32 %f431, %f5, %f50;
add.ftz.f32 %f432, %f4, %f51;
mov.f32 %f433, 0f40E00000;
min.ftz.f32 %f434, %f432, %f433;
min.ftz.f32 %f435, %f431, %f433;
mov.f32 %f436, 0fC0E00000;
max.ftz.f32 %f437, %f435, %f436;
mul.ftz.f32 %f438, %f434, 0f3FD9DB23;
mul.ftz.f32 %f439, %f438, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f440, %f439;
add.ftz.f32 %f441, %f440, 0f3F800000;
div.approx.ftz.f32 %f442, %f434, %f441;
add.ftz.f32 %f443, %f437, 0f3F800000;
mul.ftz.f32 %f444, %f443, %f442;
st.global.f32 [%rd28+4], %f444;

$L__BB293_58:
ld.shared.f32 %f445, [%r35+256];
add.ftz.f32 %f446, %f445, %f28;
ld.shared.f32 %f447, [%r37+256];
add.ftz.f32 %f448, %f447, %f32;
ld.shared.f32 %f449, [%r35+768];
add.ftz.f32 %f450, %f449, %f446;
ld.shared.f32 %f451, [%r37+768];
add.ftz.f32 %f452, %f451, %f448;
ld.shared.f32 %f453, [%r35+1280];
add.ftz.f32 %f454, %f453, %f450;
ld.shared.f32 %f455, [%r37+1280];
add.ftz.f32 %f456, %f455, %f452;
mov.b32 %r1292, %f454;
mov.u32 %r1293, 31;
mov.u32 %r1294, 16;
mov.u32 %r1295, -1;
shfl.sync.bfly.b32 %r1296|%p137, %r1292, %r1294, %r1293, %r1295;
mov.b32 %f457, %r1296;
add.ftz.f32 %f458, %f454, %f457;
mov.b32 %r1297, %f458;
mov.u32 %r1298, 8;
shfl.sync.bfly.b32 %r1299|%p138, %r1297, %r1298, %r1293, %r1295;
mov.b32 %f459, %r1299;
add.ftz.f32 %f460, %f458, %f459;
mov.b32 %r1300, %f460;
mov.u32 %r1301, 4;
shfl.sync.bfly.b32 %r1302|%p139, %r1300, %r1301, %r1293, %r1295;
mov.b32 %f461, %r1302;
add.ftz.f32 %f462, %f460, %f461;
mov.b32 %r1303, %f462;
mov.u32 %r1304, 2;
shfl.sync.bfly.b32 %r1305|%p140, %r1303, %r1304, %r1293, %r1295;
mov.b32 %f463, %r1305;
add.ftz.f32 %f464, %f462, %f463;
mov.b32 %r1306, %f464;
mov.u32 %r1307, 1;
shfl.sync.bfly.b32 %r1308|%p141, %r1306, %r1307, %r1293, %r1295;
mov.b32 %f465, %r1308;
add.ftz.f32 %f52, %f464, %f465;
mov.b32 %r1309, %f456;
shfl.sync.bfly.b32 %r1310|%p142, %r1309, %r1294, %r1293, %r1295;
mov.b32 %f466, %r1310;
add.ftz.f32 %f467, %f456, %f466;
mov.b32 %r1311, %f467;
shfl.sync.bfly.b32 %r1312|%p143, %r1311, %r1298, %r1293, %r1295;
mov.b32 %f468, %r1312;
add.ftz.f32 %f469, %f467, %f468;
mov.b32 %r1313, %f469;
shfl.sync.bfly.b32 %r1314|%p144, %r1313, %r1301, %r1293, %r1295;
mov.b32 %f470, %r1314;
add.ftz.f32 %f471, %f469, %f470;
mov.b32 %r1315, %f471;
shfl.sync.bfly.b32 %r1316|%p145, %r1315, %r1304, %r1293, %r1295;
mov.b32 %f472, %r1316;
add.ftz.f32 %f473, %f471, %f472;
mov.b32 %r1317, %f473;
shfl.sync.bfly.b32 %r1318|%p146, %r1317, %r1307, %r1293, %r1295;
mov.b32 %f474, %r1318;
add.ftz.f32 %f53, %f473, %f474;
setp.ne.s32 %p147, %r2, 2;
or.b32 %r1322, %r4, 2;
setp.ge.u32 %p148, %r1322, %r54;
or.pred %p149, %p147, %p148;
@%p149 bra $L__BB293_60;

add.ftz.f32 %f475, %f5, %f52;
add.ftz.f32 %f476, %f4, %f53;
mov.f32 %f477, 0f40E00000;
min.ftz.f32 %f478, %f476, %f477;
min.ftz.f32 %f479, %f475, %f477;
mov.f32 %f480, 0fC0E00000;
max.ftz.f32 %f481, %f479, %f480;
mul.ftz.f32 %f482, %f478, 0f3FD9DB23;
mul.ftz.f32 %f483, %f482, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f484, %f483;
add.ftz.f32 %f485, %f484, 0f3F800000;
div.approx.ftz.f32 %f486, %f478, %f485;
add.ftz.f32 %f487, %f481, 0f3F800000;
mul.ftz.f32 %f488, %f487, %f486;
st.global.f32 [%rd28+8], %f488;

$L__BB293_60:
ld.shared.f32 %f489, [%r35+384];
add.ftz.f32 %f490, %f489, %f27;
ld.shared.f32 %f491, [%r37+384];
add.ftz.f32 %f492, %f491, %f31;
ld.shared.f32 %f493, [%r35+896];
add.ftz.f32 %f494, %f493, %f490;
ld.shared.f32 %f495, [%r37+896];
add.ftz.f32 %f496, %f495, %f492;
ld.shared.f32 %f497, [%r35+1408];
add.ftz.f32 %f498, %f497, %f494;
ld.shared.f32 %f499, [%r37+1408];
add.ftz.f32 %f500, %f499, %f496;
mov.b32 %r1323, %f498;
shfl.sync.bfly.b32 %r1327|%p150, %r1323, %r1294, %r1293, %r1295;
mov.b32 %f501, %r1327;
add.ftz.f32 %f502, %f498, %f501;
mov.b32 %r1328, %f502;
shfl.sync.bfly.b32 %r1330|%p151, %r1328, %r1298, %r1293, %r1295;
mov.b32 %f503, %r1330;
add.ftz.f32 %f504, %f502, %f503;
mov.b32 %r1331, %f504;
shfl.sync.bfly.b32 %r1333|%p152, %r1331, %r1301, %r1293, %r1295;
mov.b32 %f505, %r1333;
add.ftz.f32 %f506, %f504, %f505;
mov.b32 %r1334, %f506;
shfl.sync.bfly.b32 %r1336|%p153, %r1334, %r1304, %r1293, %r1295;
mov.b32 %f507, %r1336;
add.ftz.f32 %f508, %f506, %f507;
mov.b32 %r1337, %f508;
shfl.sync.bfly.b32 %r1339|%p154, %r1337, %r1307, %r1293, %r1295;
mov.b32 %f509, %r1339;
add.ftz.f32 %f54, %f508, %f509;
mov.b32 %r1340, %f500;
shfl.sync.bfly.b32 %r1341|%p155, %r1340, %r1294, %r1293, %r1295;
mov.b32 %f510, %r1341;
add.ftz.f32 %f511, %f500, %f510;
mov.b32 %r1342, %f511;
shfl.sync.bfly.b32 %r1343|%p156, %r1342, %r1298, %r1293, %r1295;
mov.b32 %f512, %r1343;
add.ftz.f32 %f513, %f511, %f512;
mov.b32 %r1344, %f513;
shfl.sync.bfly.b32 %r1345|%p157, %r1344, %r1301, %r1293, %r1295;
mov.b32 %f514, %r1345;
add.ftz.f32 %f515, %f513, %f514;
mov.b32 %r1346, %f515;
shfl.sync.bfly.b32 %r1347|%p158, %r1346, %r1304, %r1293, %r1295;
mov.b32 %f516, %r1347;
add.ftz.f32 %f517, %f515, %f516;
mov.b32 %r1348, %f517;
shfl.sync.bfly.b32 %r1349|%p159, %r1348, %r1307, %r1293, %r1295;
mov.b32 %f518, %r1349;
add.ftz.f32 %f55, %f517, %f518;
setp.ne.s32 %p160, %r2, 3;
or.b32 %r1353, %r4, 3;
setp.ge.u32 %p161, %r1353, %r54;
or.pred %p162, %p160, %p161;
@%p162 bra $L__BB293_78;

add.ftz.f32 %f519, %f5, %f54;
add.ftz.f32 %f520, %f4, %f55;
mov.f32 %f521, 0f40E00000;
min.ftz.f32 %f522, %f520, %f521;
min.ftz.f32 %f523, %f519, %f521;
mov.f32 %f524, 0fC0E00000;
max.ftz.f32 %f525, %f523, %f524;
mul.ftz.f32 %f526, %f522, 0f3FD9DB23;
mul.ftz.f32 %f527, %f526, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f528, %f527;
add.ftz.f32 %f529, %f528, 0f3F800000;
div.approx.ftz.f32 %f530, %f522, %f529;
add.ftz.f32 %f531, %f525, 0f3F800000;
mul.ftz.f32 %f532, %f531, %f530;
st.global.f32 [%rd28+12], %f532;
bra.uni $L__BB293_78;

}
.entry _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<38>;
.reg .b16 %rs<13>;
.reg .f32 %f<109>;
.reg .b32 %r<572>;
.reg .b64 %rd<51>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

ld.param.u64 %rd13, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd14, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
mov.u32 %r1, %tid.y;
shl.b32 %r41, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r41, %r2;
mov.u32 %r42, %ctaid.x;
shl.b32 %r4, %r42, 2;
shr.u32 %r5, %r40, 8;
mov.u32 %r569, %ctaid.y;
setp.eq.s64 %p1, %rd15, 0;
@%p1 bra $L__BB294_2;

cvta.to.global.u64 %rd17, %rd15;
mul.wide.u32 %rd18, %r569, 4;
add.s64 %rd19, %rd17, %rd18;
ld.global.nc.u32 %r568, [%rd19];
mul.hi.u32 %r43, %r569, %r23;
add.s32 %r44, %r43, %r569;
shr.u32 %r45, %r44, %r24;
mul.lo.s32 %r46, %r45, %r25;
sub.s32 %r569, %r569, %r46;
bra.uni $L__BB294_3;

$L__BB294_2:
mul.hi.u32 %r47, %r569, %r28;
add.s32 %r48, %r47, %r569;
shr.u32 %r568, %r48, %r29;

$L__BB294_3:
mov.u32 %r12, %ctaid.z;
shr.s32 %r49, %r3, 31;
shr.u32 %r50, %r49, 29;
add.s32 %r51, %r3, %r50;
shr.s32 %r571, %r51, 3;
setp.ge.s32 %p2, %r571, %r5;
mov.f32 %f105, 0f00000000;
mov.f32 %f106, %f105;
mov.f32 %f107, %f105;
mov.f32 %f108, %f105;
@%p2 bra $L__BB294_6;

and.b32 %r55, %r51, -8;
sub.s32 %r56, %r3, %r55;
shl.b32 %r57, %r56, 2;
shr.u32 %r58, %r56, 31;
add.s32 %r59, %r56, %r58;
shr.s32 %r60, %r59, 1;
cvt.s64.s32 %rd20, %r60;
cvt.s64.s32 %rd1, %r56;
and.b32 %r14, %r57, 4;
shl.b32 %r15, %r56, 1;
shl.b32 %r570, %r571, 3;
mad.lo.s32 %r61, %r4, %r26, %r571;
mul.hi.u32 %r62, %r12, %r34;
add.s32 %r63, %r62, %r12;
shr.u32 %r64, %r63, %r35;
mad.lo.s32 %r65, %r64, %r37, %r61;
mad.lo.s32 %r66, %r568, %r31, %r65;
mul.wide.s32 %rd2, %r66, 136;
add.s64 %rd21, %rd2, %rd20;
add.s64 %rd3, %rd21, 4;
mul.wide.s32 %rd22, %r57, 4;
add.s64 %rd23, %rd2, %rd22;
add.s64 %rd4, %rd23, 20;
mul.wide.s32 %rd5, %r26, 136;
or.b64 %rd6, %rd5, 2;
cvta.to.global.u64 %rd7, %rd14;
cvta.to.global.u64 %rd50, %rd13;
mul.lo.s32 %r67, %r569, %r32;
cvt.u64.u32 %rd24, %r67;
mul.lo.s32 %r68, %r12, %r38;
cvt.u64.u32 %rd25, %r68;
add.s64 %rd9, %rd24, %rd25;
mov.f32 %f105, 0f00000000;

$L__BB294_5:
cvt.u64.u32 %rd26, %r570;
add.s64 %rd27, %rd9, %rd26;
add.s64 %rd28, %rd27, %rd1;
mul.lo.s64 %rd29, %rd28, 36;
add.s64 %rd30, %rd50, %rd4;
add.s64 %rd31, %rd30, -12;
ld.global.nc.u32 %r201, [%rd30+-12];
and.b32 %r202, %r201, -2004318072;
shr.u32 %r203, %r202, 1;
or.b32 %r204, %r203, 839922192;
and.b32 %r205, %r201, 30583;
mov.u32 %r206, -152379953;
mov.u32 %r207, -1079142271;
prmt.b32 %r208, %r207, %r206, %r205;
mov.u32 %r209, 1901675829;
mov.u32 %r210, 639175937;
prmt.b32 %r211, %r210, %r209, %r205;
and.b32 %r212, %r204, 30292;
prmt.b32 %r213, %r208, %r211, %r212;
shr.u32 %r214, %r201, 16;
and.b32 %r215, %r214, 30583;
prmt.b32 %r216, %r207, %r206, %r215;
prmt.b32 %r217, %r210, %r209, %r215;
shr.u32 %r218, %r204, 16;
prmt.b32 %r219, %r216, %r217, %r218;
mov.u32 %r220, 25632;
prmt.b32 %r70, %r213, %r219, %r220;
mov.u32 %r221, 30001;
prmt.b32 %r74, %r213, %r219, %r221;
add.s64 %rd32, %rd7, %rd29;
ld.global.nc.u32 %r71, [%rd32+4];
ld.global.nc.u32 %r75, [%rd32+20];
mov.u32 %r171, 0;

dp4a.s32.s32 %r69, %r70, %r71, %r171;


dp4a.s32.s32 %r73, %r74, %r75, %r69;

ld.global.nc.u32 %r222, [%rd30+-8];
and.b32 %r223, %r222, -2004318072;
shr.u32 %r224, %r223, 1;
or.b32 %r225, %r224, 839922192;
and.b32 %r226, %r222, 30583;
prmt.b32 %r227, %r207, %r206, %r226;
prmt.b32 %r228, %r210, %r209, %r226;
and.b32 %r229, %r225, 30292;
prmt.b32 %r230, %r227, %r228, %r229;
shr.u32 %r231, %r222, 16;
and.b32 %r232, %r231, 30583;
prmt.b32 %r233, %r207, %r206, %r232;
prmt.b32 %r234, %r210, %r209, %r232;
shr.u32 %r235, %r225, 16;
prmt.b32 %r236, %r233, %r234, %r235;
prmt.b32 %r78, %r230, %r236, %r220;
prmt.b32 %r82, %r230, %r236, %r221;
ld.global.nc.u32 %r79, [%rd32+8];
ld.global.nc.u32 %r83, [%rd32+24];

dp4a.s32.s32 %r77, %r78, %r79, %r73;


dp4a.s32.s32 %r81, %r82, %r83, %r77;

ld.global.nc.u32 %r237, [%rd30+-4];
and.b32 %r238, %r237, -2004318072;
shr.u32 %r239, %r238, 1;
or.b32 %r240, %r239, 839922192;
and.b32 %r241, %r237, 30583;
prmt.b32 %r242, %r207, %r206, %r241;
prmt.b32 %r243, %r210, %r209, %r241;
and.b32 %r244, %r240, 30292;
prmt.b32 %r245, %r242, %r243, %r244;
shr.u32 %r246, %r237, 16;
and.b32 %r247, %r246, 30583;
prmt.b32 %r248, %r207, %r206, %r247;
prmt.b32 %r249, %r210, %r209, %r247;
shr.u32 %r250, %r240, 16;
prmt.b32 %r251, %r248, %r249, %r250;
prmt.b32 %r86, %r245, %r251, %r220;
prmt.b32 %r90, %r245, %r251, %r221;
ld.global.nc.u32 %r87, [%rd32+12];
ld.global.nc.u32 %r91, [%rd32+28];

dp4a.s32.s32 %r85, %r86, %r87, %r81;


dp4a.s32.s32 %r89, %r90, %r91, %r85;

ld.global.nc.u32 %r252, [%rd30];
and.b32 %r253, %r252, -2004318072;
shr.u32 %r254, %r253, 1;
or.b32 %r255, %r254, 839922192;
and.b32 %r256, %r252, 30583;
prmt.b32 %r257, %r207, %r206, %r256;
prmt.b32 %r258, %r210, %r209, %r256;
and.b32 %r259, %r255, 30292;
prmt.b32 %r260, %r257, %r258, %r259;
shr.u32 %r261, %r252, 16;
and.b32 %r262, %r261, 30583;
prmt.b32 %r263, %r207, %r206, %r262;
prmt.b32 %r264, %r210, %r209, %r262;
shr.u32 %r265, %r255, 16;
prmt.b32 %r266, %r263, %r264, %r265;
prmt.b32 %r94, %r260, %r266, %r220;
prmt.b32 %r98, %r260, %r266, %r221;
ld.global.nc.u32 %r95, [%rd32+16];
ld.global.nc.u32 %r99, [%rd32+32];

dp4a.s32.s32 %r93, %r94, %r95, %r89;


dp4a.s32.s32 %r97, %r98, %r99, %r93;

add.s64 %rd33, %rd50, %rd3;
ld.global.nc.u8 %rs5, [%rd33];
cvt.u32.u16 %r267, %rs5;
and.b32 %r268, %r267, 255;
shr.u32 %r269, %r268, %r14;
and.b32 %r270, %r269, 15;
add.s64 %rd34, %rd50, %rd2;
add.s64 %rd35, %rd34, 2;
ld.global.nc.u16 %rs6, [%rd34+2];
cvt.u32.u16 %r271, %rs6;
shr.u32 %r272, %r271, %r15;
and.b32 %r273, %r272, 3;
bfi.b32 %r274, %r273, %r270, 4, 2;
add.s32 %r275, %r274, -32;
mul.lo.s32 %r276, %r275, %r97;
ld.global.nc.u16 %rs1, [%rd34];

{ cvt.f32.f16 %f25, %rs1;}


ld.global.nc.u32 %r101, [%rd32];

{.reg .f16 low,high;
mov.b32 {low,high},%r101;
cvt.f32.f16 %f26, low;}


mul.ftz.f32 %f33, %f25, %f26;
cvt.rn.f32.s32 %f34, %r276;
fma.rn.ftz.f32 %f108, %f33, %f34, %f108;
add.s64 %rd36, %rd31, %rd5;
ld.global.nc.u32 %r277, [%rd36];
and.b32 %r278, %r277, -2004318072;
shr.u32 %r279, %r278, 1;
or.b32 %r280, %r279, 839922192;
and.b32 %r281, %r277, 30583;
prmt.b32 %r282, %r207, %r206, %r281;
prmt.b32 %r283, %r210, %r209, %r281;
and.b32 %r284, %r280, 30292;
prmt.b32 %r285, %r282, %r283, %r284;
shr.u32 %r286, %r277, 16;
and.b32 %r287, %r286, 30583;
prmt.b32 %r288, %r207, %r206, %r287;
prmt.b32 %r289, %r210, %r209, %r287;
shr.u32 %r290, %r280, 16;
prmt.b32 %r291, %r288, %r289, %r290;
prmt.b32 %r103, %r285, %r291, %r220;
prmt.b32 %r107, %r285, %r291, %r221;

dp4a.s32.s32 %r102, %r103, %r71, %r171;


dp4a.s32.s32 %r106, %r107, %r75, %r102;

ld.global.nc.u32 %r292, [%rd36+4];
and.b32 %r293, %r292, -2004318072;
shr.u32 %r294, %r293, 1;
or.b32 %r295, %r294, 839922192;
and.b32 %r296, %r292, 30583;
prmt.b32 %r297, %r207, %r206, %r296;
prmt.b32 %r298, %r210, %r209, %r296;
and.b32 %r299, %r295, 30292;
prmt.b32 %r300, %r297, %r298, %r299;
shr.u32 %r301, %r292, 16;
and.b32 %r302, %r301, 30583;
prmt.b32 %r303, %r207, %r206, %r302;
prmt.b32 %r304, %r210, %r209, %r302;
shr.u32 %r305, %r295, 16;
prmt.b32 %r306, %r303, %r304, %r305;
prmt.b32 %r111, %r300, %r306, %r220;
prmt.b32 %r115, %r300, %r306, %r221;

dp4a.s32.s32 %r110, %r111, %r79, %r106;


dp4a.s32.s32 %r114, %r115, %r83, %r110;

ld.global.nc.u32 %r307, [%rd36+8];
and.b32 %r308, %r307, -2004318072;
shr.u32 %r309, %r308, 1;
or.b32 %r310, %r309, 839922192;
and.b32 %r311, %r307, 30583;
prmt.b32 %r312, %r207, %r206, %r311;
prmt.b32 %r313, %r210, %r209, %r311;
and.b32 %r314, %r310, 30292;
prmt.b32 %r315, %r312, %r313, %r314;
shr.u32 %r316, %r307, 16;
and.b32 %r317, %r316, 30583;
prmt.b32 %r318, %r207, %r206, %r317;
prmt.b32 %r319, %r210, %r209, %r317;
shr.u32 %r320, %r310, 16;
prmt.b32 %r321, %r318, %r319, %r320;
prmt.b32 %r119, %r315, %r321, %r220;
prmt.b32 %r123, %r315, %r321, %r221;

dp4a.s32.s32 %r118, %r119, %r87, %r114;


dp4a.s32.s32 %r122, %r123, %r91, %r118;

ld.global.nc.u32 %r322, [%rd36+12];
and.b32 %r323, %r322, -2004318072;
shr.u32 %r324, %r323, 1;
or.b32 %r325, %r324, 839922192;
and.b32 %r326, %r322, 30583;
prmt.b32 %r327, %r207, %r206, %r326;
prmt.b32 %r328, %r210, %r209, %r326;
and.b32 %r329, %r325, 30292;
prmt.b32 %r330, %r327, %r328, %r329;
shr.u32 %r331, %r322, 16;
and.b32 %r332, %r331, 30583;
prmt.b32 %r333, %r207, %r206, %r332;
prmt.b32 %r334, %r210, %r209, %r332;
shr.u32 %r335, %r325, 16;
prmt.b32 %r336, %r333, %r334, %r335;
prmt.b32 %r127, %r330, %r336, %r220;
prmt.b32 %r131, %r330, %r336, %r221;

dp4a.s32.s32 %r126, %r127, %r95, %r122;


dp4a.s32.s32 %r130, %r131, %r99, %r126;

add.s64 %rd37, %rd33, %rd5;
ld.global.nc.u8 %rs7, [%rd37];
cvt.u32.u16 %r337, %rs7;
and.b32 %r338, %r337, 255;
shr.u32 %r339, %r338, %r14;
and.b32 %r340, %r339, 15;
add.s64 %rd38, %rd35, %rd5;
add.s64 %rd39, %rd38, -2;
add.s64 %rd40, %rd34, %rd5;
ld.global.nc.u16 %rs8, [%rd40+2];
cvt.u32.u16 %r341, %rs8;
shr.u32 %r342, %r341, %r15;
and.b32 %r343, %r342, 3;
bfi.b32 %r344, %r343, %r340, 4, 2;
add.s32 %r345, %r344, -32;
mul.lo.s32 %r346, %r345, %r130;
ld.global.nc.u16 %rs2, [%rd40];

{ cvt.f32.f16 %f27, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r101;
cvt.f32.f16 %f28, low;}


mul.ftz.f32 %f35, %f27, %f28;
cvt.rn.f32.s32 %f36, %r346;
fma.rn.ftz.f32 %f107, %f35, %f36, %f107;
add.s64 %rd41, %rd36, %rd5;
ld.global.nc.u32 %r347, [%rd41];
and.b32 %r348, %r347, -2004318072;
shr.u32 %r349, %r348, 1;
or.b32 %r350, %r349, 839922192;
and.b32 %r351, %r347, 30583;
prmt.b32 %r352, %r207, %r206, %r351;
prmt.b32 %r353, %r210, %r209, %r351;
and.b32 %r354, %r350, 30292;
prmt.b32 %r355, %r352, %r353, %r354;
shr.u32 %r356, %r347, 16;
and.b32 %r357, %r356, 30583;
prmt.b32 %r358, %r207, %r206, %r357;
prmt.b32 %r359, %r210, %r209, %r357;
shr.u32 %r360, %r350, 16;
prmt.b32 %r361, %r358, %r359, %r360;
prmt.b32 %r136, %r355, %r361, %r220;
prmt.b32 %r140, %r355, %r361, %r221;

dp4a.s32.s32 %r135, %r136, %r71, %r171;


dp4a.s32.s32 %r139, %r140, %r75, %r135;

ld.global.nc.u32 %r362, [%rd41+4];
and.b32 %r363, %r362, -2004318072;
shr.u32 %r364, %r363, 1;
or.b32 %r365, %r364, 839922192;
and.b32 %r366, %r362, 30583;
prmt.b32 %r367, %r207, %r206, %r366;
prmt.b32 %r368, %r210, %r209, %r366;
and.b32 %r369, %r365, 30292;
prmt.b32 %r370, %r367, %r368, %r369;
shr.u32 %r371, %r362, 16;
and.b32 %r372, %r371, 30583;
prmt.b32 %r373, %r207, %r206, %r372;
prmt.b32 %r374, %r210, %r209, %r372;
shr.u32 %r375, %r365, 16;
prmt.b32 %r376, %r373, %r374, %r375;
prmt.b32 %r144, %r370, %r376, %r220;
prmt.b32 %r148, %r370, %r376, %r221;

dp4a.s32.s32 %r143, %r144, %r79, %r139;


dp4a.s32.s32 %r147, %r148, %r83, %r143;

ld.global.nc.u32 %r377, [%rd41+8];
and.b32 %r378, %r377, -2004318072;
shr.u32 %r379, %r378, 1;
or.b32 %r380, %r379, 839922192;
and.b32 %r381, %r377, 30583;
prmt.b32 %r382, %r207, %r206, %r381;
prmt.b32 %r383, %r210, %r209, %r381;
and.b32 %r384, %r380, 30292;
prmt.b32 %r385, %r382, %r383, %r384;
shr.u32 %r386, %r377, 16;
and.b32 %r387, %r386, 30583;
prmt.b32 %r388, %r207, %r206, %r387;
prmt.b32 %r389, %r210, %r209, %r387;
shr.u32 %r390, %r380, 16;
prmt.b32 %r391, %r388, %r389, %r390;
prmt.b32 %r152, %r385, %r391, %r220;
prmt.b32 %r156, %r385, %r391, %r221;

dp4a.s32.s32 %r151, %r152, %r87, %r147;


dp4a.s32.s32 %r155, %r156, %r91, %r151;

ld.global.nc.u32 %r392, [%rd41+12];
and.b32 %r393, %r392, -2004318072;
shr.u32 %r394, %r393, 1;
or.b32 %r395, %r394, 839922192;
and.b32 %r396, %r392, 30583;
prmt.b32 %r397, %r207, %r206, %r396;
prmt.b32 %r398, %r210, %r209, %r396;
and.b32 %r399, %r395, 30292;
prmt.b32 %r400, %r397, %r398, %r399;
shr.u32 %r401, %r392, 16;
and.b32 %r402, %r401, 30583;
prmt.b32 %r403, %r207, %r206, %r402;
prmt.b32 %r404, %r210, %r209, %r402;
shr.u32 %r405, %r395, 16;
prmt.b32 %r406, %r403, %r404, %r405;
prmt.b32 %r160, %r400, %r406, %r220;
prmt.b32 %r164, %r400, %r406, %r221;

dp4a.s32.s32 %r159, %r160, %r95, %r155;


dp4a.s32.s32 %r163, %r164, %r99, %r159;

add.s64 %rd42, %rd37, %rd5;
ld.global.nc.u8 %rs9, [%rd42];
cvt.u32.u16 %r407, %rs9;
and.b32 %r408, %r407, 255;
shr.u32 %r409, %r408, %r14;
and.b32 %r410, %r409, 15;
add.s64 %rd43, %rd39, %rd5;
add.s64 %rd44, %rd39, %rd6;
ld.global.nc.u16 %rs10, [%rd44];
cvt.u32.u16 %r411, %rs10;
shr.u32 %r412, %r411, %r15;
and.b32 %r413, %r412, 3;
bfi.b32 %r414, %r413, %r410, 4, 2;
add.s32 %r415, %r414, -32;
mul.lo.s32 %r416, %r415, %r163;
ld.global.nc.u16 %rs3, [%rd44+-2];

{ cvt.f32.f16 %f29, %rs3;}



{.reg .f16 low,high;
mov.b32 {low,high},%r101;
cvt.f32.f16 %f30, low;}


mul.ftz.f32 %f37, %f29, %f30;
cvt.rn.f32.s32 %f38, %r416;
fma.rn.ftz.f32 %f106, %f37, %f38, %f106;
add.s64 %rd45, %rd41, %rd5;
ld.global.nc.u32 %r417, [%rd45];
and.b32 %r418, %r417, -2004318072;
shr.u32 %r419, %r418, 1;
or.b32 %r420, %r419, 839922192;
and.b32 %r421, %r417, 30583;
prmt.b32 %r422, %r207, %r206, %r421;
prmt.b32 %r423, %r210, %r209, %r421;
and.b32 %r424, %r420, 30292;
prmt.b32 %r425, %r422, %r423, %r424;
shr.u32 %r426, %r417, 16;
and.b32 %r427, %r426, 30583;
prmt.b32 %r428, %r207, %r206, %r427;
prmt.b32 %r429, %r210, %r209, %r427;
shr.u32 %r430, %r420, 16;
prmt.b32 %r431, %r428, %r429, %r430;
prmt.b32 %r169, %r425, %r431, %r220;
prmt.b32 %r173, %r425, %r431, %r221;

dp4a.s32.s32 %r168, %r169, %r71, %r171;


dp4a.s32.s32 %r172, %r173, %r75, %r168;

ld.global.nc.u32 %r432, [%rd45+4];
and.b32 %r433, %r432, -2004318072;
shr.u32 %r434, %r433, 1;
or.b32 %r435, %r434, 839922192;
and.b32 %r436, %r432, 30583;
prmt.b32 %r437, %r207, %r206, %r436;
prmt.b32 %r438, %r210, %r209, %r436;
and.b32 %r439, %r435, 30292;
prmt.b32 %r440, %r437, %r438, %r439;
shr.u32 %r441, %r432, 16;
and.b32 %r442, %r441, 30583;
prmt.b32 %r443, %r207, %r206, %r442;
prmt.b32 %r444, %r210, %r209, %r442;
shr.u32 %r445, %r435, 16;
prmt.b32 %r446, %r443, %r444, %r445;
prmt.b32 %r177, %r440, %r446, %r220;
prmt.b32 %r181, %r440, %r446, %r221;

dp4a.s32.s32 %r176, %r177, %r79, %r172;


dp4a.s32.s32 %r180, %r181, %r83, %r176;

ld.global.nc.u32 %r447, [%rd45+8];
and.b32 %r448, %r447, -2004318072;
shr.u32 %r449, %r448, 1;
or.b32 %r450, %r449, 839922192;
and.b32 %r451, %r447, 30583;
prmt.b32 %r452, %r207, %r206, %r451;
prmt.b32 %r453, %r210, %r209, %r451;
and.b32 %r454, %r450, 30292;
prmt.b32 %r455, %r452, %r453, %r454;
shr.u32 %r456, %r447, 16;
and.b32 %r457, %r456, 30583;
prmt.b32 %r458, %r207, %r206, %r457;
prmt.b32 %r459, %r210, %r209, %r457;
shr.u32 %r460, %r450, 16;
prmt.b32 %r461, %r458, %r459, %r460;
prmt.b32 %r185, %r455, %r461, %r220;
prmt.b32 %r189, %r455, %r461, %r221;

dp4a.s32.s32 %r184, %r185, %r87, %r180;


dp4a.s32.s32 %r188, %r189, %r91, %r184;

ld.global.nc.u32 %r462, [%rd45+12];
and.b32 %r463, %r462, -2004318072;
shr.u32 %r464, %r463, 1;
or.b32 %r465, %r464, 839922192;
and.b32 %r466, %r462, 30583;
prmt.b32 %r467, %r207, %r206, %r466;
prmt.b32 %r468, %r210, %r209, %r466;
and.b32 %r469, %r465, 30292;
prmt.b32 %r470, %r467, %r468, %r469;
shr.u32 %r471, %r462, 16;
and.b32 %r472, %r471, 30583;
prmt.b32 %r473, %r207, %r206, %r472;
prmt.b32 %r474, %r210, %r209, %r472;
shr.u32 %r475, %r465, 16;
prmt.b32 %r476, %r473, %r474, %r475;
prmt.b32 %r193, %r470, %r476, %r220;
prmt.b32 %r197, %r470, %r476, %r221;

dp4a.s32.s32 %r192, %r193, %r95, %r188;


dp4a.s32.s32 %r196, %r197, %r99, %r192;

add.s64 %rd46, %rd42, %rd5;
ld.global.nc.u8 %rs11, [%rd46];
cvt.u32.u16 %r477, %rs11;
and.b32 %r478, %r477, 255;
shr.u32 %r479, %r478, %r14;
and.b32 %r480, %r479, 15;
add.s64 %rd47, %rd43, %rd6;
ld.global.nc.u16 %rs12, [%rd47];
cvt.u32.u16 %r481, %rs12;
shr.u32 %r482, %r481, %r15;
and.b32 %r483, %r482, 3;
bfi.b32 %r484, %r483, %r480, 4, 2;
add.s32 %r485, %r484, -32;
mul.lo.s32 %r486, %r485, %r196;
ld.global.nc.u16 %rs4, [%rd47+-2];

{ cvt.f32.f16 %f31, %rs4;}



{.reg .f16 low,high;
mov.b32 {low,high},%r101;
cvt.f32.f16 %f32, low;}


mul.ftz.f32 %f39, %f31, %f32;
cvt.rn.f32.s32 %f40, %r486;
fma.rn.ftz.f32 %f105, %f39, %f40, %f105;
add.s32 %r570, %r570, 128;
add.s64 %rd50, %rd50, 2176;
add.s32 %r571, %r571, 16;
setp.lt.s32 %p3, %r571, %r5;
@%p3 bra $L__BB294_5;

$L__BB294_6:
setp.eq.s32 %p4, %r1, 0;
@%p4 bra $L__BB294_8;

shl.b32 %r487, %r1, 9;
mov.u32 %r488, _ZZ13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r489, %r488, %r487;
shl.b32 %r490, %r2, 2;
add.s32 %r491, %r489, %r490;
st.shared.f32 [%r491+-512], %f108;
st.shared.f32 [%r491+-384], %f107;
st.shared.f32 [%r491+-256], %f106;
st.shared.f32 [%r491+-128], %f105;

$L__BB294_8:
setp.ne.s32 %p5, %r1, 0;
bar.sync 0;
@%p5 bra $L__BB294_17;

mov.u32 %r492, %ctaid.y;
mad.lo.s32 %r493, %r492, %r33, %r4;
mad.lo.s32 %r494, %r12, %r39, %r493;
shl.b32 %r495, %r2, 2;
mov.u32 %r496, _ZZ13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r21, %r496, %r495;
mov.u32 %r497, 2;
ld.shared.f32 %f41, [%r21];
add.ftz.f32 %f42, %f41, %f108;
ld.shared.f32 %f43, [%r21+512];
add.ftz.f32 %f44, %f43, %f42;
ld.shared.f32 %f45, [%r21+1024];
add.ftz.f32 %f46, %f45, %f44;
mov.b32 %r498, %f46;
mov.u32 %r499, 31;
mov.u32 %r500, 16;
mov.u32 %r501, -1;
shfl.sync.bfly.b32 %r502|%p6, %r498, %r500, %r499, %r501;
mov.b32 %f47, %r502;
add.ftz.f32 %f48, %f46, %f47;
mov.b32 %r503, %f48;
mov.u32 %r504, 8;
shfl.sync.bfly.b32 %r505|%p7, %r503, %r504, %r499, %r501;
mov.b32 %f49, %r505;
add.ftz.f32 %f50, %f48, %f49;
mov.b32 %r506, %f50;
mov.u32 %r507, 4;
shfl.sync.bfly.b32 %r508|%p8, %r506, %r507, %r499, %r501;
mov.b32 %f51, %r508;
add.ftz.f32 %f52, %f50, %f51;
mov.b32 %r509, %f52;
shfl.sync.bfly.b32 %r510|%p9, %r509, %r497, %r499, %r501;
mov.b32 %f53, %r510;
add.ftz.f32 %f54, %f52, %f53;
mov.b32 %r511, %f54;
mov.u32 %r512, 1;
shfl.sync.bfly.b32 %r513|%p10, %r511, %r512, %r499, %r501;
mov.b32 %f55, %r513;
add.ftz.f32 %f13, %f54, %f55;
setp.ge.u32 %p11, %r4, %r27;
setp.ne.s32 %p12, %r2, 0;
cvta.to.global.u64 %rd48, %rd16;
mul.wide.u32 %rd49, %r494, 4;
add.s64 %rd12, %rd48, %rd49;
or.pred %p13, %p12, %p11;
@%p13 bra $L__BB294_11;

st.global.f32 [%rd12], %f13;

$L__BB294_11:
ld.shared.f32 %f56, [%r21+128];
add.ftz.f32 %f57, %f56, %f107;
ld.shared.f32 %f58, [%r21+640];
add.ftz.f32 %f59, %f58, %f57;
ld.shared.f32 %f60, [%r21+1152];
add.ftz.f32 %f61, %f60, %f59;
mov.b32 %r514, %f61;
shfl.sync.bfly.b32 %r518|%p14, %r514, %r500, %r499, %r501;
mov.b32 %f62, %r518;
add.ftz.f32 %f63, %f61, %f62;
mov.b32 %r519, %f63;
shfl.sync.bfly.b32 %r521|%p15, %r519, %r504, %r499, %r501;
mov.b32 %f64, %r521;
add.ftz.f32 %f65, %f63, %f64;
mov.b32 %r522, %f65;
shfl.sync.bfly.b32 %r524|%p16, %r522, %r507, %r499, %r501;
mov.b32 %f66, %r524;
add.ftz.f32 %f67, %f65, %f66;
mov.b32 %r525, %f67;
shfl.sync.bfly.b32 %r527|%p17, %r525, %r497, %r499, %r501;
mov.b32 %f68, %r527;
add.ftz.f32 %f69, %f67, %f68;
mov.b32 %r528, %f69;
shfl.sync.bfly.b32 %r530|%p18, %r528, %r512, %r499, %r501;
mov.b32 %f70, %r530;
add.ftz.f32 %f14, %f69, %f70;
add.s32 %r531, %r4, 1;
setp.ge.u32 %p19, %r531, %r27;
setp.ne.s32 %p20, %r2, 1;
or.pred %p21, %p20, %p19;
@%p21 bra $L__BB294_13;

st.global.f32 [%rd12+4], %f14;

$L__BB294_13:
ld.shared.f32 %f71, [%r21+256];
add.ftz.f32 %f72, %f71, %f106;
ld.shared.f32 %f73, [%r21+768];
add.ftz.f32 %f74, %f73, %f72;
ld.shared.f32 %f75, [%r21+1280];
add.ftz.f32 %f76, %f75, %f74;
mov.b32 %r532, %f76;
mov.u32 %r533, 31;
mov.u32 %r534, 16;
mov.u32 %r535, -1;
shfl.sync.bfly.b32 %r536|%p22, %r532, %r534, %r533, %r535;
mov.b32 %f77, %r536;
add.ftz.f32 %f78, %f76, %f77;
mov.b32 %r537, %f78;
mov.u32 %r538, 8;
shfl.sync.bfly.b32 %r539|%p23, %r537, %r538, %r533, %r535;
mov.b32 %f79, %r539;
add.ftz.f32 %f80, %f78, %f79;
mov.b32 %r540, %f80;
mov.u32 %r541, 4;
shfl.sync.bfly.b32 %r542|%p24, %r540, %r541, %r533, %r535;
mov.b32 %f81, %r542;
add.ftz.f32 %f82, %f80, %f81;
mov.b32 %r543, %f82;
mov.u32 %r544, 2;
shfl.sync.bfly.b32 %r545|%p25, %r543, %r544, %r533, %r535;
mov.b32 %f83, %r545;
add.ftz.f32 %f84, %f82, %f83;
mov.b32 %r546, %f84;
mov.u32 %r547, 1;
shfl.sync.bfly.b32 %r548|%p26, %r546, %r547, %r533, %r535;
mov.b32 %f85, %r548;
add.ftz.f32 %f15, %f84, %f85;
add.s32 %r549, %r4, 2;
setp.ge.u32 %p27, %r549, %r27;
setp.ne.s32 %p28, %r2, 2;
or.pred %p29, %p28, %p27;
@%p29 bra $L__BB294_15;

st.global.f32 [%rd12+8], %f15;

$L__BB294_15:
ld.shared.f32 %f86, [%r21+384];
add.ftz.f32 %f87, %f86, %f105;
ld.shared.f32 %f88, [%r21+896];
add.ftz.f32 %f89, %f88, %f87;
ld.shared.f32 %f90, [%r21+1408];
add.ftz.f32 %f91, %f90, %f89;
mov.b32 %r550, %f91;
shfl.sync.bfly.b32 %r554|%p30, %r550, %r534, %r533, %r535;
mov.b32 %f92, %r554;
add.ftz.f32 %f93, %f91, %f92;
mov.b32 %r555, %f93;
shfl.sync.bfly.b32 %r557|%p31, %r555, %r538, %r533, %r535;
mov.b32 %f94, %r557;
add.ftz.f32 %f95, %f93, %f94;
mov.b32 %r558, %f95;
shfl.sync.bfly.b32 %r560|%p32, %r558, %r541, %r533, %r535;
mov.b32 %f96, %r560;
add.ftz.f32 %f97, %f95, %f96;
mov.b32 %r561, %f97;
shfl.sync.bfly.b32 %r563|%p33, %r561, %r544, %r533, %r535;
mov.b32 %f98, %r563;
add.ftz.f32 %f99, %f97, %f98;
mov.b32 %r564, %f99;
shfl.sync.bfly.b32 %r566|%p34, %r564, %r547, %r533, %r535;
mov.b32 %f100, %r566;
add.ftz.f32 %f16, %f99, %f100;
add.s32 %r567, %r4, 3;
setp.ge.u32 %p35, %r567, %r27;
setp.ne.s32 %p36, %r2, 3;
or.pred %p37, %p36, %p35;
@%p37 bra $L__BB294_17;

st.global.f32 [%rd12+12], %f16;

$L__BB294_17:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<33>;
.reg .b16 %rs<5>;
.reg .f32 %f<121>;
.reg .b32 %r<386>;
.reg .b64 %rd<70>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[384];

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate[384];

ld.param.u64 %rd8, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd9, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd10, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r48, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r49, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r50, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r46, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r45, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
ld.param.u64 %rd13, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+16];
ld.param.u64 %rd12, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+8];
ld.param.u64 %rd11, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3];
mov.u32 %r1, %tid.y;
shl.b32 %r51, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r51, %r2;
mov.u32 %r4, %ctaid.x;
mov.u32 %r5, %ctaid.y;
setp.eq.s64 %p2, %rd10, 0;
@%p2 bra $L__BB295_2;

cvta.to.global.u64 %rd17, %rd10;
mul.wide.u32 %rd18, %r5, 4;
add.s64 %rd19, %rd17, %rd18;
ld.global.nc.u32 %r382, [%rd19];
mul.hi.u32 %r52, %r5, %r35;
add.s32 %r53, %r52, %r5;
shr.u32 %r54, %r53, %r36;
mul.lo.s32 %r55, %r54, %r37;
sub.s32 %r383, %r5, %r55;
bra.uni $L__BB295_3;

$L__BB295_2:
mul.hi.u32 %r56, %r5, %r39;
add.s32 %r57, %r56, %r5;
shr.u32 %r382, %r57, %r40;
mov.u32 %r383, %r5;

$L__BB295_3:
mov.u32 %r11, %ctaid.z;
selp.b32 %r12, %r5, %r382, %p2;
or.b32 %r58, %r2, %r1;
setp.ne.s32 %p4, %r58, 0;
mov.f32 %f114, 0f00000000;
mov.f32 %f110, %f114;
mov.f32 %f111, %f114;
@%p4 bra $L__BB295_8;

setp.eq.s64 %p5, %rd11, 0;
mov.f32 %f110, 0f00000000;
mov.f32 %f111, %f110;
@%p5 bra $L__BB295_6;

cvta.to.global.u64 %rd20, %rd11;
mul.lo.s32 %r59, %r11, %r50;
cvt.u64.u32 %rd21, %r59;
mul.lo.s32 %r60, %r12, %r44;
cvt.u64.u32 %rd22, %r60;
cvt.u64.u32 %rd23, %r2;
cvt.s64.s32 %rd24, %r4;
add.s64 %rd25, %rd23, %rd24;
add.s64 %rd26, %rd25, %rd21;
add.s64 %rd27, %rd26, %rd22;
shl.b64 %rd28, %rd27, 2;
add.s64 %rd29, %rd20, %rd28;
ld.global.f32 %f111, [%rd29];

$L__BB295_6:
setp.eq.s64 %p6, %rd13, 0;
setp.eq.s64 %p7, %rd12, 0;
or.pred %p8, %p6, %p7;
@%p8 bra $L__BB295_8;

cvta.to.global.u64 %rd30, %rd13;
mul.lo.s32 %r61, %r11, %r50;
cvt.u64.u32 %rd31, %r61;
mul.lo.s32 %r62, %r12, %r44;
cvt.u64.u32 %rd32, %r62;
cvt.u64.u32 %rd33, %r2;
cvt.s64.s32 %rd34, %r4;
add.s64 %rd35, %rd33, %rd34;
add.s64 %rd36, %rd35, %rd31;
add.s64 %rd37, %rd36, %rd32;
shl.b64 %rd38, %rd37, 2;
add.s64 %rd39, %rd30, %rd38;
ld.global.f32 %f110, [%rd39];

$L__BB295_8:
shr.s32 %r63, %r3, 31;
shr.u32 %r64, %r63, 29;
add.s32 %r65, %r3, %r64;
shr.s32 %r385, %r65, 3;
shr.u32 %r66, %r34, 8;
setp.ge.s32 %p9, %r385, %r66;
mov.f32 %f116, %f114;
@%p9 bra $L__BB295_13;

and.b32 %r70, %r65, -8;
sub.s32 %r14, %r3, %r70;
shr.u32 %r71, %r14, 31;
add.s32 %r72, %r14, %r71;
shr.s32 %r15, %r72, 1;
mul.hi.u32 %r73, %r11, %r45;
add.s32 %r74, %r73, %r11;
shr.u32 %r75, %r74, %r46;
mad.lo.s32 %r76, %r75, %r48, %r385;
mad.lo.s32 %r77, %r382, %r42, %r76;
mad.lo.s32 %r78, %r4, %r38, %r77;
mul.wide.s32 %rd69, %r78, 136;
shl.b32 %r384, %r385, 3;
mul.lo.s32 %r79, %r383, %r43;
cvt.u64.u32 %rd40, %r79;
mul.lo.s32 %r80, %r11, %r49;
cvt.u64.u32 %rd41, %r80;
add.s64 %rd5, %rd40, %rd41;
mov.f32 %f116, 0f00000000;
cvta.to.global.u64 %rd42, %rd9;
cvt.s64.s32 %rd45, %r14;
cvta.to.global.u64 %rd49, %rd8;
cvt.s64.s32 %rd53, %r15;
add.s64 %rd54, %rd49, %rd53;
mov.f32 %f114, %f116;

$L__BB295_10:
cvt.u64.u32 %rd43, %r384;
add.s64 %rd44, %rd5, %rd43;
add.s64 %rd46, %rd44, %rd45;
mul.lo.s64 %rd47, %rd46, 36;
shl.b32 %r114, %r14, 2;
mul.wide.s32 %rd48, %r114, 4;
add.s64 %rd50, %rd49, %rd48;
add.s64 %rd51, %rd50, %rd69;
ld.global.nc.u32 %r115, [%rd51+8];
and.b32 %r116, %r115, -2004318072;
shr.u32 %r117, %r116, 1;
or.b32 %r118, %r117, 839922192;
and.b32 %r119, %r115, 30583;
mov.u32 %r120, -152379953;
mov.u32 %r121, -1079142271;
prmt.b32 %r122, %r121, %r120, %r119;
mov.u32 %r123, 1901675829;
mov.u32 %r124, 639175937;
prmt.b32 %r125, %r124, %r123, %r119;
and.b32 %r126, %r118, 30292;
prmt.b32 %r127, %r122, %r125, %r126;
shr.u32 %r128, %r115, 16;
and.b32 %r129, %r128, 30583;
prmt.b32 %r130, %r121, %r120, %r129;
prmt.b32 %r131, %r124, %r123, %r129;
shr.u32 %r132, %r118, 16;
prmt.b32 %r133, %r130, %r131, %r132;
mov.u32 %r134, 25632;
prmt.b32 %r82, %r127, %r133, %r134;
mov.u32 %r135, 30001;
prmt.b32 %r86, %r127, %r133, %r135;
add.s64 %rd52, %rd42, %rd47;
ld.global.nc.u32 %r19, [%rd52+4];
ld.global.nc.u32 %r20, [%rd52+20];
mov.u32 %r84, 0;

dp4a.s32.s32 %r81, %r82, %r19, %r84;


dp4a.s32.s32 %r85, %r86, %r20, %r81;

ld.global.nc.u32 %r136, [%rd51+12];
and.b32 %r137, %r136, -2004318072;
shr.u32 %r138, %r137, 1;
or.b32 %r139, %r138, 839922192;
and.b32 %r140, %r136, 30583;
prmt.b32 %r141, %r121, %r120, %r140;
prmt.b32 %r142, %r124, %r123, %r140;
and.b32 %r143, %r139, 30292;
prmt.b32 %r144, %r141, %r142, %r143;
shr.u32 %r145, %r136, 16;
and.b32 %r146, %r145, 30583;
prmt.b32 %r147, %r121, %r120, %r146;
prmt.b32 %r148, %r124, %r123, %r146;
shr.u32 %r149, %r139, 16;
prmt.b32 %r150, %r147, %r148, %r149;
prmt.b32 %r90, %r144, %r150, %r134;
prmt.b32 %r94, %r144, %r150, %r135;
ld.global.nc.u32 %r21, [%rd52+8];
ld.global.nc.u32 %r22, [%rd52+24];

dp4a.s32.s32 %r89, %r90, %r21, %r85;


dp4a.s32.s32 %r93, %r94, %r22, %r89;

ld.global.nc.u32 %r151, [%rd51+16];
and.b32 %r152, %r151, -2004318072;
shr.u32 %r153, %r152, 1;
or.b32 %r154, %r153, 839922192;
and.b32 %r155, %r151, 30583;
prmt.b32 %r156, %r121, %r120, %r155;
prmt.b32 %r157, %r124, %r123, %r155;
and.b32 %r158, %r154, 30292;
prmt.b32 %r159, %r156, %r157, %r158;
shr.u32 %r160, %r151, 16;
and.b32 %r161, %r160, 30583;
prmt.b32 %r162, %r121, %r120, %r161;
prmt.b32 %r163, %r124, %r123, %r161;
shr.u32 %r164, %r154, 16;
prmt.b32 %r165, %r162, %r163, %r164;
prmt.b32 %r98, %r159, %r165, %r134;
prmt.b32 %r102, %r159, %r165, %r135;
ld.global.nc.u32 %r23, [%rd52+12];
ld.global.nc.u32 %r24, [%rd52+28];

dp4a.s32.s32 %r97, %r98, %r23, %r93;


dp4a.s32.s32 %r101, %r102, %r24, %r97;

ld.global.nc.u32 %r166, [%rd51+20];
and.b32 %r167, %r166, -2004318072;
shr.u32 %r168, %r167, 1;
or.b32 %r169, %r168, 839922192;
and.b32 %r170, %r166, 30583;
prmt.b32 %r171, %r121, %r120, %r170;
prmt.b32 %r172, %r124, %r123, %r170;
and.b32 %r173, %r169, 30292;
prmt.b32 %r174, %r171, %r172, %r173;
shr.u32 %r175, %r166, 16;
and.b32 %r176, %r175, 30583;
prmt.b32 %r177, %r121, %r120, %r176;
prmt.b32 %r178, %r124, %r123, %r176;
shr.u32 %r179, %r169, 16;
prmt.b32 %r180, %r177, %r178, %r179;
prmt.b32 %r106, %r174, %r180, %r134;
prmt.b32 %r110, %r174, %r180, %r135;
ld.global.nc.u32 %r25, [%rd52+16];
ld.global.nc.u32 %r26, [%rd52+32];

dp4a.s32.s32 %r105, %r106, %r25, %r101;


dp4a.s32.s32 %r109, %r110, %r26, %r105;

add.s64 %rd55, %rd54, %rd69;
ld.global.nc.u8 %rs2, [%rd55+4];
cvt.u32.u16 %r181, %rs2;
and.b32 %r182, %r181, 255;
and.b32 %r183, %r114, 4;
shr.u32 %r184, %r182, %r183;
and.b32 %r185, %r184, 15;
add.s64 %rd56, %rd49, %rd69;
ld.global.nc.u16 %rs3, [%rd56+2];
cvt.u32.u16 %r186, %rs3;
shl.b32 %r187, %r14, 1;
shr.u32 %r188, %r186, %r187;
and.b32 %r189, %r188, 3;
bfi.b32 %r190, %r189, %r185, 4, 2;
add.s32 %r191, %r190, -32;
mul.lo.s32 %r192, %r191, %r109;
ld.global.nc.u16 %rs1, [%rd56];

{ cvt.f32.f16 %f40, %rs1;}


ld.global.nc.u32 %r27, [%rd52];

{.reg .f16 low,high;
mov.b32 {low,high},%r27;
cvt.f32.f16 %f41, low;}


mul.ftz.f32 %f42, %f40, %f41;
cvt.rn.f32.s32 %f43, %r192;
fma.rn.ftz.f32 %f116, %f42, %f43, %f116;
setp.eq.s64 %p10, %rd12, 0;
@%p10 bra $L__BB295_12;

cvta.to.global.u64 %rd58, %rd12;
add.s64 %rd59, %rd58, %rd48;
add.s64 %rd60, %rd59, %rd69;
ld.global.u32 %r227, [%rd60+8];
and.b32 %r228, %r227, -2004318072;
shr.u32 %r229, %r228, 1;
or.b32 %r230, %r229, 839922192;
and.b32 %r231, %r227, 30583;
prmt.b32 %r234, %r121, %r120, %r231;
prmt.b32 %r237, %r124, %r123, %r231;
and.b32 %r238, %r230, 30292;
prmt.b32 %r239, %r234, %r237, %r238;
shr.u32 %r240, %r227, 16;
and.b32 %r241, %r240, 30583;
prmt.b32 %r242, %r121, %r120, %r241;
prmt.b32 %r243, %r124, %r123, %r241;
shr.u32 %r244, %r230, 16;
prmt.b32 %r245, %r242, %r243, %r244;
prmt.b32 %r194, %r239, %r245, %r134;
prmt.b32 %r198, %r239, %r245, %r135;

dp4a.s32.s32 %r193, %r194, %r19, %r84;


dp4a.s32.s32 %r197, %r198, %r20, %r193;

ld.global.u32 %r248, [%rd60+12];
and.b32 %r249, %r248, -2004318072;
shr.u32 %r250, %r249, 1;
or.b32 %r251, %r250, 839922192;
and.b32 %r252, %r248, 30583;
prmt.b32 %r253, %r121, %r120, %r252;
prmt.b32 %r254, %r124, %r123, %r252;
and.b32 %r255, %r251, 30292;
prmt.b32 %r256, %r253, %r254, %r255;
shr.u32 %r257, %r248, 16;
and.b32 %r258, %r257, 30583;
prmt.b32 %r259, %r121, %r120, %r258;
prmt.b32 %r260, %r124, %r123, %r258;
shr.u32 %r261, %r251, 16;
prmt.b32 %r262, %r259, %r260, %r261;
prmt.b32 %r202, %r256, %r262, %r134;
prmt.b32 %r206, %r256, %r262, %r135;

dp4a.s32.s32 %r201, %r202, %r21, %r197;


dp4a.s32.s32 %r205, %r206, %r22, %r201;

ld.global.u32 %r263, [%rd60+16];
and.b32 %r264, %r263, -2004318072;
shr.u32 %r265, %r264, 1;
or.b32 %r266, %r265, 839922192;
and.b32 %r267, %r263, 30583;
prmt.b32 %r268, %r121, %r120, %r267;
prmt.b32 %r269, %r124, %r123, %r267;
and.b32 %r270, %r266, 30292;
prmt.b32 %r271, %r268, %r269, %r270;
shr.u32 %r272, %r263, 16;
and.b32 %r273, %r272, 30583;
prmt.b32 %r274, %r121, %r120, %r273;
prmt.b32 %r275, %r124, %r123, %r273;
shr.u32 %r276, %r266, 16;
prmt.b32 %r277, %r274, %r275, %r276;
prmt.b32 %r210, %r271, %r277, %r134;
prmt.b32 %r214, %r271, %r277, %r135;

dp4a.s32.s32 %r209, %r210, %r23, %r205;


dp4a.s32.s32 %r213, %r214, %r24, %r209;

ld.global.u32 %r278, [%rd60+20];
and.b32 %r279, %r278, -2004318072;
shr.u32 %r280, %r279, 1;
or.b32 %r281, %r280, 839922192;
and.b32 %r282, %r278, 30583;
prmt.b32 %r283, %r121, %r120, %r282;
prmt.b32 %r284, %r124, %r123, %r282;
and.b32 %r285, %r281, 30292;
prmt.b32 %r286, %r283, %r284, %r285;
shr.u32 %r287, %r278, 16;
and.b32 %r288, %r287, 30583;
prmt.b32 %r289, %r121, %r120, %r288;
prmt.b32 %r290, %r124, %r123, %r288;
shr.u32 %r291, %r281, 16;
prmt.b32 %r292, %r289, %r290, %r291;
prmt.b32 %r218, %r286, %r292, %r134;
prmt.b32 %r222, %r286, %r292, %r135;

dp4a.s32.s32 %r217, %r218, %r25, %r213;


dp4a.s32.s32 %r221, %r222, %r26, %r217;

add.s64 %rd62, %rd58, %rd53;
add.s64 %rd63, %rd62, %rd69;
ld.global.u8 %r293, [%rd63+4];
shr.u32 %r295, %r293, %r183;
and.b32 %r296, %r295, 15;
add.s64 %rd64, %rd58, %rd69;
ld.global.u16 %r297, [%rd64+2];
shr.u32 %r299, %r297, %r187;
and.b32 %r300, %r299, 3;
bfi.b32 %r301, %r300, %r296, 4, 2;
add.s32 %r302, %r301, -32;
mul.lo.s32 %r303, %r302, %r221;
ld.global.u16 %rs4, [%rd64];

{ cvt.f32.f16 %f44, %rs4;}



{.reg .f16 low,high;
mov.b32 {low,high},%r27;
cvt.f32.f16 %f45, low;}


mul.ftz.f32 %f46, %f44, %f45;
cvt.rn.f32.s32 %f47, %r303;
fma.rn.ftz.f32 %f114, %f46, %f47, %f114;

$L__BB295_12:
add.s64 %rd69, %rd69, 2176;
add.s32 %r384, %r384, 128;
add.s32 %r385, %r385, 16;
setp.lt.s32 %p11, %r385, %r66;
@%p11 bra $L__BB295_10;

$L__BB295_13:
mov.u32 %r367, %tid.y;
setp.eq.s32 %p12, %r367, 0;
@%p12 bra $L__BB295_16;

mov.u32 %r369, %tid.y;
mov.u32 %r368, %tid.x;
setp.eq.s64 %p13, %rd12, 0;
add.s32 %r30, %r369, -1;
shl.b32 %r308, %r369, 7;
mov.u32 %r309, _ZZ13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r310, %r309, %r308;
shl.b32 %r311, %r368, 2;
add.s32 %r312, %r310, %r311;
st.shared.f32 [%r312+-128], %f116;
@%p13 bra $L__BB295_16;

shl.b32 %r314, %r30, 7;
mov.u32 %r315, _ZZ13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r316, %r315, %r314;
add.s32 %r318, %r316, %r311;
st.shared.f32 [%r318], %f114;

$L__BB295_16:
mov.u32 %r370, %tid.y;
setp.ne.s32 %p14, %r370, 0;
bar.sync 0;
@%p14 bra $L__BB295_35;

mov.u32 %r371, %tid.x;
setp.ne.s64 %p15, %rd12, 0;
shl.b32 %r321, %r371, 2;
mov.u32 %r322, _ZZ13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r31, %r322, %r321;
ld.shared.f32 %f48, [%r31];
add.ftz.f32 %f13, %f116, %f48;
@%p15 bra $L__BB295_19;
bra.uni $L__BB295_18;

$L__BB295_19:
mov.u32 %r325, _ZZ13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r326, %r325, %r321;
ld.shared.f32 %f52, [%r326];
add.ftz.f32 %f53, %f114, %f52;
ld.shared.f32 %f54, [%r31+128];
add.ftz.f32 %f55, %f13, %f54;
ld.shared.f32 %f56, [%r326+128];
add.ftz.f32 %f57, %f53, %f56;
ld.shared.f32 %f58, [%r31+256];
add.ftz.f32 %f118, %f55, %f58;
ld.shared.f32 %f59, [%r326+256];
add.ftz.f32 %f114, %f57, %f59;
bra.uni $L__BB295_20;

$L__BB295_18:
ld.shared.f32 %f49, [%r31+128];
add.ftz.f32 %f50, %f13, %f49;
ld.shared.f32 %f51, [%r31+256];
add.ftz.f32 %f118, %f50, %f51;

$L__BB295_20:
mov.b32 %r327, %f118;
mov.u32 %r328, 31;
mov.u32 %r329, 16;
mov.u32 %r330, -1;
shfl.sync.bfly.b32 %r331|%p16, %r327, %r329, %r328, %r330;
mov.b32 %f60, %r331;
add.ftz.f32 %f61, %f118, %f60;
mov.b32 %r332, %f61;
mov.u32 %r333, 8;
shfl.sync.bfly.b32 %r334|%p17, %r332, %r333, %r328, %r330;
mov.b32 %f62, %r334;
add.ftz.f32 %f63, %f61, %f62;
mov.b32 %r335, %f63;
mov.u32 %r336, 4;
shfl.sync.bfly.b32 %r337|%p18, %r335, %r336, %r328, %r330;
mov.b32 %f64, %r337;
add.ftz.f32 %f65, %f63, %f64;
mov.b32 %r338, %f65;
mov.u32 %r339, 2;
shfl.sync.bfly.b32 %r340|%p19, %r338, %r339, %r328, %r330;
mov.b32 %f66, %r340;
add.ftz.f32 %f19, %f65, %f66;
mov.b32 %r341, %f19;
mov.u32 %r342, 1;
shfl.sync.bfly.b32 %r32|%p1, %r341, %r342, %r328, %r330;
setp.eq.s64 %p20, %rd12, 0;
@%p20 bra $L__BB295_22;

mov.b32 %r343, %f114;
shfl.sync.bfly.b32 %r347|%p21, %r343, %r329, %r328, %r330;
mov.b32 %f67, %r347;
add.ftz.f32 %f68, %f114, %f67;
mov.b32 %r348, %f68;
shfl.sync.bfly.b32 %r350|%p22, %r348, %r333, %r328, %r330;
mov.b32 %f69, %r350;
add.ftz.f32 %f70, %f68, %f69;
mov.b32 %r351, %f70;
shfl.sync.bfly.b32 %r353|%p23, %r351, %r336, %r328, %r330;
mov.b32 %f71, %r353;
add.ftz.f32 %f72, %f70, %f71;
mov.b32 %r354, %f72;
shfl.sync.bfly.b32 %r356|%p24, %r354, %r339, %r328, %r330;
mov.b32 %f73, %r356;
add.ftz.f32 %f74, %f72, %f73;
mov.b32 %r357, %f74;
shfl.sync.bfly.b32 %r359|%p25, %r357, %r342, %r328, %r330;
mov.b32 %f75, %r359;
add.ftz.f32 %f114, %f74, %f75;

$L__BB295_22:
mov.u32 %r372, %tid.x;
setp.ne.s32 %p26, %r372, 0;
mov.b32 %f76, %r32;
add.ftz.f32 %f22, %f19, %f76;
@%p26 bra $L__BB295_35;

add.ftz.f32 %f120, %f111, %f22;
@%p20 bra $L__BB295_34;

ld.param.u32 %r377, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
add.ftz.f32 %f24, %f110, %f114;
setp.gt.s32 %p28, %r377, 2;
@%p28 bra $L__BB295_28;

ld.param.u32 %r380, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p31, %r380, 1;
@%p31 bra $L__BB295_32;

ld.param.u32 %r381, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p32, %r381, 2;
@%p32 bra $L__BB295_27;
bra.uni $L__BB295_33;

$L__BB295_27:
mul.ftz.f32 %f96, %f24, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f97, %f96;
add.ftz.f32 %f98, %f97, 0f3F800000;
div.approx.ftz.f32 %f99, %f24, %f98;
mul.ftz.f32 %f120, %f120, %f99;
bra.uni $L__BB295_34;

$L__BB295_28:
ld.param.u32 %r378, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p29, %r378, 3;
@%p29 bra $L__BB295_31;

ld.param.u32 %r379, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.ne.s32 %p30, %r379, 6;
@%p30 bra $L__BB295_33;

ld.param.f32 %f108, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
min.ftz.f32 %f77, %f24, %f108;
neg.ftz.f32 %f78, %f108;
min.ftz.f32 %f79, %f120, %f108;
max.ftz.f32 %f80, %f79, %f78;
mul.ftz.f32 %f81, %f77, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f82, %f81;
add.ftz.f32 %f83, %f82, 0f3F800000;
div.approx.ftz.f32 %f84, %f77, %f83;
mul.ftz.f32 %f120, %f80, %f84;
bra.uni $L__BB295_34;

$L__BB295_32:
mul.ftz.f32 %f100, %f24, 0f3F000000;
mul.ftz.f32 %f101, %f24, 0f3D372713;
fma.rn.ftz.f32 %f102, %f24, %f101, 0f3F800000;
mul.ftz.f32 %f103, %f24, 0f3F4C422A;
mul.ftz.f32 %f104, %f103, %f102;
tanh.approx.f32 %f105, %f104;
add.ftz.f32 %f106, %f105, 0f3F800000;
mul.ftz.f32 %f107, %f100, %f106;
mul.ftz.f32 %f120, %f120, %f107;
bra.uni $L__BB295_34;

$L__BB295_31:
mov.f32 %f85, 0f40E00000;
min.ftz.f32 %f86, %f24, %f85;
min.ftz.f32 %f87, %f120, %f85;
mov.f32 %f88, 0fC0E00000;
max.ftz.f32 %f89, %f87, %f88;
mul.ftz.f32 %f90, %f86, 0f3FD9DB23;
mul.ftz.f32 %f91, %f90, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f92, %f91;
add.ftz.f32 %f93, %f92, 0f3F800000;
div.approx.ftz.f32 %f94, %f86, %f93;
add.ftz.f32 %f95, %f89, 0f3F800000;
mul.ftz.f32 %f120, %f95, %f94;
bra.uni $L__BB295_34;

$L__BB295_33:
mul.ftz.f32 %f120, %f120, %f24;

$L__BB295_34:
ld.param.u64 %rd68, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
mov.u32 %r376, %ctaid.x;
ld.param.u32 %r375, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r374, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
mov.u32 %r373, %ctaid.z;
cvta.to.global.u64 %rd65, %rd68;
mov.u32 %r363, %ctaid.y;
mad.lo.s32 %r365, %r363, %r375, %r376;
mad.lo.s32 %r366, %r373, %r374, %r365;
mul.wide.u32 %rd66, %r366, 4;
add.s64 %rd67, %rd65, %rd66;
st.global.f32 [%rd67], %f120;

$L__BB295_35:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<12>;
.reg .b16 %rs<4>;
.reg .f32 %f<28>;
.reg .b32 %r<208>;
.reg .b64 %rd<41>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[384];

ld.param.u64 %rd13, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd14, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
mov.u32 %r1, %tid.y;
shl.b32 %r38, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r38, %r2;
mov.u32 %r205, %ctaid.y;
setp.eq.s64 %p1, %rd15, 0;
@%p1 bra $L__BB296_2;

cvta.to.global.u64 %rd17, %rd15;
mul.wide.u32 %rd18, %r205, 4;
add.s64 %rd19, %rd17, %rd18;
ld.global.nc.u32 %r204, [%rd19];
mul.hi.u32 %r39, %r205, %r22;
add.s32 %r40, %r39, %r205;
shr.u32 %r41, %r40, %r23;
mul.lo.s32 %r42, %r41, %r24;
sub.s32 %r205, %r205, %r42;
bra.uni $L__BB296_3;

$L__BB296_2:
mul.hi.u32 %r43, %r205, %r26;
add.s32 %r44, %r43, %r205;
shr.u32 %r204, %r44, %r27;

$L__BB296_3:
mov.u32 %r10, %ctaid.z;
shr.s32 %r45, %r3, 31;
shr.u32 %r46, %r45, 29;
add.s32 %r47, %r3, %r46;
shr.s32 %r207, %r47, 3;
shr.u32 %r48, %r21, 8;
setp.ge.s32 %p2, %r207, %r48;
mov.f32 %f27, 0f00000000;
@%p2 bra $L__BB296_6;

and.b32 %r52, %r47, -8;
sub.s32 %r53, %r3, %r52;
shr.u32 %r54, %r53, 31;
add.s32 %r55, %r53, %r54;
shr.s32 %r56, %r55, 1;
cvt.s64.s32 %rd20, %r56;
shl.b32 %r57, %r53, 2;
cvt.s64.s32 %rd1, %r53;
and.b32 %r12, %r57, 4;
shl.b32 %r13, %r53, 1;
shl.b32 %r206, %r207, 3;
mul.hi.u32 %r58, %r10, %r32;
add.s32 %r59, %r58, %r10;
shr.u32 %r60, %r59, %r33;
mad.lo.s32 %r61, %r60, %r35, %r207;
mad.lo.s32 %r62, %r204, %r29, %r61;
mov.u32 %r63, %ctaid.x;
mad.lo.s32 %r64, %r63, %r25, %r62;
mul.wide.s32 %rd21, %r64, 136;
mul.wide.s32 %rd22, %r57, 4;
add.s64 %rd23, %rd21, %rd22;
cvta.to.global.u64 %rd24, %rd13;
add.s64 %rd25, %rd24, %rd23;
add.s64 %rd40, %rd25, 12;
add.s64 %rd26, %rd21, %rd20;
add.s64 %rd27, %rd24, %rd26;
add.s64 %rd39, %rd27, 4;
add.s64 %rd38, %rd24, %rd21;
cvta.to.global.u64 %rd5, %rd14;
mul.lo.s32 %r65, %r205, %r30;
cvt.u64.u32 %rd28, %r65;
mul.lo.s32 %r66, %r10, %r36;
cvt.u64.u32 %rd29, %r66;
add.s64 %rd6, %rd28, %rd29;
mov.f32 %f27, 0f00000000;

$L__BB296_5:
cvt.u64.u32 %rd30, %r206;
add.s64 %rd31, %rd6, %rd30;
add.s64 %rd32, %rd31, %rd1;
mul.lo.s64 %rd33, %rd32, 36;
ld.global.nc.u32 %r100, [%rd40+-4];
and.b32 %r101, %r100, -2004318072;
shr.u32 %r102, %r101, 1;
or.b32 %r103, %r102, 839922192;
and.b32 %r104, %r100, 30583;
mov.u32 %r105, -152379953;
mov.u32 %r106, -1079142271;
prmt.b32 %r107, %r106, %r105, %r104;
mov.u32 %r108, 1901675829;
mov.u32 %r109, 639175937;
prmt.b32 %r110, %r109, %r108, %r104;
and.b32 %r111, %r103, 30292;
prmt.b32 %r112, %r107, %r110, %r111;
shr.u32 %r113, %r100, 16;
and.b32 %r114, %r113, 30583;
prmt.b32 %r115, %r106, %r105, %r114;
prmt.b32 %r116, %r109, %r108, %r114;
shr.u32 %r117, %r103, 16;
prmt.b32 %r118, %r115, %r116, %r117;
mov.u32 %r119, 25632;
prmt.b32 %r68, %r112, %r118, %r119;
mov.u32 %r120, 30001;
prmt.b32 %r72, %r112, %r118, %r120;
add.s64 %rd34, %rd5, %rd33;
ld.global.nc.u32 %r69, [%rd34+4];
ld.global.nc.u32 %r73, [%rd34+20];
mov.u32 %r70, 0;

dp4a.s32.s32 %r67, %r68, %r69, %r70;


dp4a.s32.s32 %r71, %r72, %r73, %r67;

ld.global.nc.u32 %r121, [%rd40];
and.b32 %r122, %r121, -2004318072;
shr.u32 %r123, %r122, 1;
or.b32 %r124, %r123, 839922192;
and.b32 %r125, %r121, 30583;
prmt.b32 %r126, %r106, %r105, %r125;
prmt.b32 %r127, %r109, %r108, %r125;
and.b32 %r128, %r124, 30292;
prmt.b32 %r129, %r126, %r127, %r128;
shr.u32 %r130, %r121, 16;
and.b32 %r131, %r130, 30583;
prmt.b32 %r132, %r106, %r105, %r131;
prmt.b32 %r133, %r109, %r108, %r131;
shr.u32 %r134, %r124, 16;
prmt.b32 %r135, %r132, %r133, %r134;
prmt.b32 %r76, %r129, %r135, %r119;
prmt.b32 %r80, %r129, %r135, %r120;
ld.global.nc.u32 %r77, [%rd34+8];
ld.global.nc.u32 %r81, [%rd34+24];

dp4a.s32.s32 %r75, %r76, %r77, %r71;


dp4a.s32.s32 %r79, %r80, %r81, %r75;

ld.global.nc.u32 %r136, [%rd40+4];
and.b32 %r137, %r136, -2004318072;
shr.u32 %r138, %r137, 1;
or.b32 %r139, %r138, 839922192;
and.b32 %r140, %r136, 30583;
prmt.b32 %r141, %r106, %r105, %r140;
prmt.b32 %r142, %r109, %r108, %r140;
and.b32 %r143, %r139, 30292;
prmt.b32 %r144, %r141, %r142, %r143;
shr.u32 %r145, %r136, 16;
and.b32 %r146, %r145, 30583;
prmt.b32 %r147, %r106, %r105, %r146;
prmt.b32 %r148, %r109, %r108, %r146;
shr.u32 %r149, %r139, 16;
prmt.b32 %r150, %r147, %r148, %r149;
prmt.b32 %r84, %r144, %r150, %r119;
prmt.b32 %r88, %r144, %r150, %r120;
ld.global.nc.u32 %r85, [%rd34+12];
ld.global.nc.u32 %r89, [%rd34+28];

dp4a.s32.s32 %r83, %r84, %r85, %r79;


dp4a.s32.s32 %r87, %r88, %r89, %r83;

ld.global.nc.u32 %r151, [%rd40+8];
and.b32 %r152, %r151, -2004318072;
shr.u32 %r153, %r152, 1;
or.b32 %r154, %r153, 839922192;
and.b32 %r155, %r151, 30583;
prmt.b32 %r156, %r106, %r105, %r155;
prmt.b32 %r157, %r109, %r108, %r155;
and.b32 %r158, %r154, 30292;
prmt.b32 %r159, %r156, %r157, %r158;
shr.u32 %r160, %r151, 16;
and.b32 %r161, %r160, 30583;
prmt.b32 %r162, %r106, %r105, %r161;
prmt.b32 %r163, %r109, %r108, %r161;
shr.u32 %r164, %r154, 16;
prmt.b32 %r165, %r162, %r163, %r164;
prmt.b32 %r92, %r159, %r165, %r119;
prmt.b32 %r96, %r159, %r165, %r120;
ld.global.nc.u32 %r93, [%rd34+16];
ld.global.nc.u32 %r97, [%rd34+32];

dp4a.s32.s32 %r91, %r92, %r93, %r87;


dp4a.s32.s32 %r95, %r96, %r97, %r91;

ld.global.nc.u8 %rs2, [%rd39];
cvt.u32.u16 %r166, %rs2;
and.b32 %r167, %r166, 255;
shr.u32 %r168, %r167, %r12;
and.b32 %r169, %r168, 15;
ld.global.nc.u16 %rs3, [%rd38+2];
cvt.u32.u16 %r170, %rs3;
shr.u32 %r171, %r170, %r13;
and.b32 %r172, %r171, 3;
bfi.b32 %r173, %r172, %r169, 4, 2;
add.s32 %r174, %r173, -32;
mul.lo.s32 %r175, %r174, %r95;
ld.global.nc.u16 %rs1, [%rd38];

{ cvt.f32.f16 %f7, %rs1;}


ld.global.nc.u32 %r99, [%rd34];

{.reg .f16 low,high;
mov.b32 {low,high},%r99;
cvt.f32.f16 %f8, low;}


mul.ftz.f32 %f9, %f7, %f8;
cvt.rn.f32.s32 %f10, %r175;
fma.rn.ftz.f32 %f27, %f9, %f10, %f27;
add.s32 %r206, %r206, 128;
add.s64 %rd40, %rd40, 2176;
add.s64 %rd39, %rd39, 2176;
add.s64 %rd38, %rd38, 2176;
add.s32 %r207, %r207, 16;
setp.lt.s32 %p3, %r207, %r48;
@%p3 bra $L__BB296_5;

$L__BB296_6:
setp.eq.s32 %p4, %r1, 0;
@%p4 bra $L__BB296_8;

shl.b32 %r176, %r1, 7;
mov.u32 %r177, _ZZ13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r178, %r177, %r176;
shl.b32 %r179, %r2, 2;
add.s32 %r180, %r178, %r179;
st.shared.f32 [%r180+-128], %f27;

$L__BB296_8:
setp.ne.s32 %p5, %r1, 0;
bar.sync 0;
@%p5 bra $L__BB296_11;

mov.u32 %r181, %ctaid.y;
mov.u32 %r182, %ctaid.x;
mad.lo.s32 %r183, %r181, %r31, %r182;
mad.lo.s32 %r20, %r10, %r37, %r183;
shl.b32 %r184, %r2, 2;
mov.u32 %r185, _ZZ13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r186, %r185, %r184;
mov.u32 %r187, 2;
ld.shared.f32 %f11, [%r186];
add.ftz.f32 %f12, %f27, %f11;
ld.shared.f32 %f13, [%r186+128];
add.ftz.f32 %f14, %f12, %f13;
ld.shared.f32 %f15, [%r186+256];
add.ftz.f32 %f16, %f14, %f15;
mov.b32 %r188, %f16;
mov.u32 %r189, 31;
mov.u32 %r190, 16;
mov.u32 %r191, -1;
shfl.sync.bfly.b32 %r192|%p6, %r188, %r190, %r189, %r191;
mov.b32 %f17, %r192;
add.ftz.f32 %f18, %f16, %f17;
mov.b32 %r193, %f18;
mov.u32 %r194, 8;
shfl.sync.bfly.b32 %r195|%p7, %r193, %r194, %r189, %r191;
mov.b32 %f19, %r195;
add.ftz.f32 %f20, %f18, %f19;
mov.b32 %r196, %f20;
mov.u32 %r197, 4;
shfl.sync.bfly.b32 %r198|%p8, %r196, %r197, %r189, %r191;
mov.b32 %f21, %r198;
add.ftz.f32 %f22, %f20, %f21;
mov.b32 %r199, %f22;
shfl.sync.bfly.b32 %r200|%p9, %r199, %r187, %r189, %r191;
mov.b32 %f23, %r200;
add.ftz.f32 %f24, %f22, %f23;
mov.b32 %r201, %f24;
mov.u32 %r202, 1;
shfl.sync.bfly.b32 %r203|%p10, %r201, %r202, %r189, %r191;
mov.b32 %f25, %r203;
add.ftz.f32 %f4, %f24, %f25;
setp.ne.s32 %p11, %r2, 0;
@%p11 bra $L__BB296_11;

cvta.to.global.u64 %rd35, %rd16;
mul.wide.u32 %rd36, %r20, 4;
add.s64 %rd37, %rd35, %rd36;
st.global.f32 [%rd37], %f4;

$L__BB296_11:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type23ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type23ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type23ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type23ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type23ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type23ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type23ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type23ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type23ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<35>;
.reg .b16 %rs<9>;
.reg .f32 %f<109>;
.reg .b32 %r<432>;
.reg .b64 %rd<53>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type23ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type23ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type23ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type23ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r19, [_Z13mul_mat_vec_qIL9ggml_type23ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type23ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type23ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type23ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type23ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type23ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type23ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type23ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type23ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type23ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type23ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type23ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type23ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type23ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
cvta.to.global.u64 %rd1, %rd17;
mov.u32 %r1, %tid.y;
shl.b32 %r35, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r35, %r2;
mov.u32 %r36, %ctaid.x;
shl.b32 %r4, %r36, 1;
mov.u32 %r5, %ctaid.y;
mov.u32 %r6, %ctaid.z;
shr.s32 %r37, %r3, 31;
shr.u32 %r38, %r37, 29;
add.s32 %r39, %r3, %r38;
shr.s32 %r431, %r39, 3;
shr.u32 %r40, %r19, 8;
setp.ge.s32 %p3, %r431, %r40;
mov.f32 %f105, 0f00000000;
mov.f32 %f106, %f105;
mov.f32 %f107, %f105;
mov.f32 %f108, %f105;
@%p3 bra $L__BB297_3;

mul.lo.s32 %r41, %r5, %r27;
cvt.u64.u32 %rd18, %r41;
mul.lo.s32 %r42, %r6, %r33;
cvt.u64.u32 %rd19, %r42;
add.s64 %rd2, %rd19, %rd18;
mul.hi.u32 %r43, %r6, %r29;
add.s32 %r44, %r43, %r6;
shr.u32 %r45, %r44, %r30;
mul.lo.s32 %r46, %r45, %r32;
mul.hi.u32 %r47, %r5, %r23;
add.s32 %r48, %r47, %r5;
shr.u32 %r49, %r48, %r24;
mul.lo.s32 %r50, %r49, %r26;
cvta.to.global.u64 %rd52, %rd15;
and.b32 %r54, %r39, -8;
sub.s32 %r8, %r3, %r54;
shl.b32 %r55, %r8, 2;
shr.u32 %r56, %r8, 31;
add.s32 %r57, %r8, %r56;
shr.s32 %r58, %r57, 1;
cvt.s64.s32 %rd20, %r58;
shl.b32 %r429, %r431, 3;
add.s32 %r430, %r21, %r429;
add.s32 %r59, %r431, %r46;
add.s32 %r60, %r59, %r50;
add.s32 %r61, %r4, 1;
mad.lo.s32 %r62, %r20, %r61, %r60;
mul.wide.s32 %rd4, %r62, 136;
mad.lo.s32 %r63, %r4, %r20, %r431;
add.s32 %r64, %r63, %r46;
add.s32 %r65, %r64, %r50;
mul.wide.s32 %rd5, %r65, 136;
add.s64 %rd21, %rd5, %rd20;
add.s64 %rd6, %rd21, 4;
add.s64 %rd22, %rd4, %rd20;
add.s64 %rd7, %rd22, 4;
mul.wide.s32 %rd23, %r55, 4;
add.s64 %rd24, %rd5, %rd23;
add.s64 %rd8, %rd24, 20;
add.s64 %rd25, %rd4, %rd23;
add.s64 %rd9, %rd25, 12;
cvta.to.global.u64 %rd10, %rd16;
mov.f32 %f105, 0f00000000;

$L__BB297_2:
cvt.u64.u32 %rd26, %r429;
add.s64 %rd27, %rd2, %rd26;
cvt.s64.s32 %rd28, %r8;
add.s64 %rd29, %rd27, %rd28;
mul.lo.s64 %rd30, %rd29, 36;
add.s64 %rd31, %rd52, %rd8;
ld.global.nc.u32 %r198, [%rd31+-12];
and.b32 %r199, %r198, -2004318072;
shr.u32 %r200, %r199, 1;
or.b32 %r201, %r200, 839922192;
and.b32 %r202, %r198, 30583;
mov.u32 %r203, -152379953;
mov.u32 %r204, -1079142271;
prmt.b32 %r205, %r204, %r203, %r202;
mov.u32 %r206, 1901675829;
mov.u32 %r207, 639175937;
prmt.b32 %r208, %r207, %r206, %r202;
and.b32 %r209, %r201, 30292;
prmt.b32 %r210, %r205, %r208, %r209;
shr.u32 %r211, %r198, 16;
and.b32 %r212, %r211, 30583;
prmt.b32 %r213, %r204, %r203, %r212;
prmt.b32 %r214, %r207, %r206, %r212;
shr.u32 %r215, %r201, 16;
prmt.b32 %r216, %r213, %r214, %r215;
mov.u32 %r217, 25632;
prmt.b32 %r133, %r210, %r216, %r217;
mov.u32 %r218, 30001;
prmt.b32 %r137, %r210, %r216, %r218;
add.s64 %rd32, %rd10, %rd30;
ld.global.nc.u32 %r68, [%rd32+4];
ld.global.nc.u32 %r72, [%rd32+20];
mov.u32 %r168, 0;

dp4a.s32.s32 %r66, %r133, %r68, %r168;


dp4a.s32.s32 %r70, %r137, %r72, %r66;

ld.global.nc.u32 %r219, [%rd31+-8];
and.b32 %r220, %r219, -2004318072;
shr.u32 %r221, %r220, 1;
or.b32 %r222, %r221, 839922192;
and.b32 %r223, %r219, 30583;
prmt.b32 %r224, %r204, %r203, %r223;
prmt.b32 %r225, %r207, %r206, %r223;
and.b32 %r226, %r222, 30292;
prmt.b32 %r227, %r224, %r225, %r226;
shr.u32 %r228, %r219, 16;
and.b32 %r229, %r228, 30583;
prmt.b32 %r230, %r204, %r203, %r229;
prmt.b32 %r231, %r207, %r206, %r229;
shr.u32 %r232, %r222, 16;
prmt.b32 %r233, %r230, %r231, %r232;
prmt.b32 %r141, %r227, %r233, %r217;
prmt.b32 %r145, %r227, %r233, %r218;
ld.global.nc.u32 %r76, [%rd32+8];
ld.global.nc.u32 %r80, [%rd32+24];

dp4a.s32.s32 %r74, %r141, %r76, %r70;


dp4a.s32.s32 %r78, %r145, %r80, %r74;

ld.global.nc.u32 %r234, [%rd31+-4];
and.b32 %r235, %r234, -2004318072;
shr.u32 %r236, %r235, 1;
or.b32 %r237, %r236, 839922192;
and.b32 %r238, %r234, 30583;
prmt.b32 %r239, %r204, %r203, %r238;
prmt.b32 %r240, %r207, %r206, %r238;
and.b32 %r241, %r237, 30292;
prmt.b32 %r242, %r239, %r240, %r241;
shr.u32 %r243, %r234, 16;
and.b32 %r244, %r243, 30583;
prmt.b32 %r245, %r204, %r203, %r244;
prmt.b32 %r246, %r207, %r206, %r244;
shr.u32 %r247, %r237, 16;
prmt.b32 %r248, %r245, %r246, %r247;
prmt.b32 %r149, %r242, %r248, %r217;
prmt.b32 %r153, %r242, %r248, %r218;
ld.global.nc.u32 %r84, [%rd32+12];
ld.global.nc.u32 %r88, [%rd32+28];

dp4a.s32.s32 %r82, %r149, %r84, %r78;


dp4a.s32.s32 %r86, %r153, %r88, %r82;

ld.global.nc.u32 %r249, [%rd31];
and.b32 %r250, %r249, -2004318072;
shr.u32 %r251, %r250, 1;
or.b32 %r252, %r251, 839922192;
and.b32 %r253, %r249, 30583;
prmt.b32 %r254, %r204, %r203, %r253;
prmt.b32 %r255, %r207, %r206, %r253;
and.b32 %r256, %r252, 30292;
prmt.b32 %r257, %r254, %r255, %r256;
shr.u32 %r258, %r249, 16;
and.b32 %r259, %r258, 30583;
prmt.b32 %r260, %r204, %r203, %r259;
prmt.b32 %r261, %r207, %r206, %r259;
shr.u32 %r262, %r252, 16;
prmt.b32 %r263, %r260, %r261, %r262;
prmt.b32 %r157, %r257, %r263, %r217;
prmt.b32 %r161, %r257, %r263, %r218;
ld.global.nc.u32 %r92, [%rd32+16];
ld.global.nc.u32 %r96, [%rd32+32];

dp4a.s32.s32 %r90, %r157, %r92, %r86;


dp4a.s32.s32 %r94, %r161, %r96, %r90;

add.s64 %rd33, %rd52, %rd6;
ld.global.nc.u8 %rs5, [%rd33];
cvt.u32.u16 %r264, %rs5;
and.b32 %r265, %r264, 255;
and.b32 %r267, %r55, 4;
shr.u32 %r268, %r265, %r267;
and.b32 %r269, %r268, 15;
add.s64 %rd34, %rd52, %rd5;
ld.global.nc.u16 %rs6, [%rd34+2];
cvt.u32.u16 %r270, %rs6;
shl.b32 %r271, %r8, 1;
shr.u32 %r272, %r270, %r271;
and.b32 %r273, %r272, 3;
bfi.b32 %r274, %r273, %r269, 4, 2;
add.s32 %r275, %r274, -32;
mul.lo.s32 %r276, %r275, %r94;
ld.global.nc.u16 %rs1, [%rd34];

{ cvt.f32.f16 %f25, %rs1;}


ld.global.nc.u32 %r98, [%rd32];

{.reg .f16 low,high;
mov.b32 {low,high},%r98;
cvt.f32.f16 %f26, low;}


mul.ftz.f32 %f33, %f25, %f26;
cvt.rn.f32.s32 %f34, %r276;
fma.rn.ftz.f32 %f108, %f33, %f34, %f108;
add.s64 %rd35, %rd52, %rd9;
ld.global.nc.u32 %r277, [%rd35+-4];
and.b32 %r278, %r277, -2004318072;
shr.u32 %r279, %r278, 1;
or.b32 %r280, %r279, 839922192;
and.b32 %r281, %r277, 30583;
prmt.b32 %r282, %r204, %r203, %r281;
prmt.b32 %r283, %r207, %r206, %r281;
and.b32 %r284, %r280, 30292;
prmt.b32 %r285, %r282, %r283, %r284;
shr.u32 %r286, %r277, 16;
and.b32 %r287, %r286, 30583;
prmt.b32 %r288, %r204, %r203, %r287;
prmt.b32 %r289, %r207, %r206, %r287;
shr.u32 %r290, %r280, 16;
prmt.b32 %r291, %r288, %r289, %r290;
prmt.b32 %r166, %r285, %r291, %r217;
prmt.b32 %r170, %r285, %r291, %r218;

dp4a.s32.s32 %r99, %r166, %r68, %r168;


dp4a.s32.s32 %r103, %r170, %r72, %r99;

ld.global.nc.u32 %r292, [%rd35];
and.b32 %r293, %r292, -2004318072;
shr.u32 %r294, %r293, 1;
or.b32 %r295, %r294, 839922192;
and.b32 %r296, %r292, 30583;
prmt.b32 %r297, %r204, %r203, %r296;
prmt.b32 %r298, %r207, %r206, %r296;
and.b32 %r299, %r295, 30292;
prmt.b32 %r300, %r297, %r298, %r299;
shr.u32 %r301, %r292, 16;
and.b32 %r302, %r301, 30583;
prmt.b32 %r303, %r204, %r203, %r302;
prmt.b32 %r304, %r207, %r206, %r302;
shr.u32 %r305, %r295, 16;
prmt.b32 %r306, %r303, %r304, %r305;
prmt.b32 %r174, %r300, %r306, %r217;
prmt.b32 %r178, %r300, %r306, %r218;

dp4a.s32.s32 %r107, %r174, %r76, %r103;


dp4a.s32.s32 %r111, %r178, %r80, %r107;

ld.global.nc.u32 %r307, [%rd35+4];
and.b32 %r308, %r307, -2004318072;
shr.u32 %r309, %r308, 1;
or.b32 %r310, %r309, 839922192;
and.b32 %r311, %r307, 30583;
prmt.b32 %r312, %r204, %r203, %r311;
prmt.b32 %r313, %r207, %r206, %r311;
and.b32 %r314, %r310, 30292;
prmt.b32 %r315, %r312, %r313, %r314;
shr.u32 %r316, %r307, 16;
and.b32 %r317, %r316, 30583;
prmt.b32 %r318, %r204, %r203, %r317;
prmt.b32 %r319, %r207, %r206, %r317;
shr.u32 %r320, %r310, 16;
prmt.b32 %r321, %r318, %r319, %r320;
prmt.b32 %r182, %r315, %r321, %r217;
prmt.b32 %r186, %r315, %r321, %r218;

dp4a.s32.s32 %r115, %r182, %r84, %r111;


dp4a.s32.s32 %r119, %r186, %r88, %r115;

ld.global.nc.u32 %r322, [%rd35+8];
and.b32 %r323, %r322, -2004318072;
shr.u32 %r324, %r323, 1;
or.b32 %r325, %r324, 839922192;
and.b32 %r326, %r322, 30583;
prmt.b32 %r327, %r204, %r203, %r326;
prmt.b32 %r328, %r207, %r206, %r326;
and.b32 %r329, %r325, 30292;
prmt.b32 %r330, %r327, %r328, %r329;
shr.u32 %r331, %r322, 16;
and.b32 %r332, %r331, 30583;
prmt.b32 %r333, %r204, %r203, %r332;
prmt.b32 %r334, %r207, %r206, %r332;
shr.u32 %r335, %r325, 16;
prmt.b32 %r336, %r333, %r334, %r335;
prmt.b32 %r190, %r330, %r336, %r217;
prmt.b32 %r194, %r330, %r336, %r218;

dp4a.s32.s32 %r123, %r190, %r92, %r119;


dp4a.s32.s32 %r127, %r194, %r96, %r123;

add.s64 %rd36, %rd52, %rd7;
ld.global.nc.u8 %rs7, [%rd36];
cvt.u32.u16 %r337, %rs7;
and.b32 %r338, %r337, 255;
shr.u32 %r339, %r338, %r267;
and.b32 %r340, %r339, 15;
add.s64 %rd37, %rd52, %rd4;
ld.global.nc.u16 %rs8, [%rd37+2];
cvt.u32.u16 %r341, %rs8;
shr.u32 %r342, %r341, %r271;
and.b32 %r343, %r342, 3;
bfi.b32 %r344, %r343, %r340, 4, 2;
add.s32 %r345, %r344, -32;
mul.lo.s32 %r346, %r345, %r127;
ld.global.nc.u16 %rs2, [%rd37];

{ cvt.f32.f16 %f27, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r98;
cvt.f32.f16 %f28, low;}


mul.ftz.f32 %f35, %f27, %f28;
cvt.rn.f32.s32 %f36, %r346;
fma.rn.ftz.f32 %f107, %f35, %f36, %f107;
cvt.u64.u32 %rd38, %r430;
add.s64 %rd39, %rd2, %rd38;
add.s64 %rd40, %rd39, %rd28;
mul.lo.s64 %rd41, %rd40, 36;
add.s64 %rd42, %rd10, %rd41;
ld.global.nc.u32 %r134, [%rd42+4];
ld.global.nc.u32 %r138, [%rd42+20];

dp4a.s32.s32 %r132, %r133, %r134, %r168;


dp4a.s32.s32 %r136, %r137, %r138, %r132;

ld.global.nc.u32 %r142, [%rd42+8];
ld.global.nc.u32 %r146, [%rd42+24];

dp4a.s32.s32 %r140, %r141, %r142, %r136;


dp4a.s32.s32 %r144, %r145, %r146, %r140;

ld.global.nc.u32 %r150, [%rd42+12];
ld.global.nc.u32 %r154, [%rd42+28];

dp4a.s32.s32 %r148, %r149, %r150, %r144;


dp4a.s32.s32 %r152, %r153, %r154, %r148;

ld.global.nc.u32 %r158, [%rd42+16];
ld.global.nc.u32 %r162, [%rd42+32];

dp4a.s32.s32 %r156, %r157, %r158, %r152;


dp4a.s32.s32 %r160, %r161, %r162, %r156;

mul.lo.s32 %r347, %r275, %r160;

{ cvt.f32.f16 %f29, %rs1;}


ld.global.nc.u32 %r164, [%rd42];

{.reg .f16 low,high;
mov.b32 {low,high},%r164;
cvt.f32.f16 %f30, low;}


mul.ftz.f32 %f37, %f29, %f30;
cvt.rn.f32.s32 %f38, %r347;
fma.rn.ftz.f32 %f106, %f37, %f38, %f106;

dp4a.s32.s32 %r165, %r166, %r134, %r168;


dp4a.s32.s32 %r169, %r170, %r138, %r165;


dp4a.s32.s32 %r173, %r174, %r142, %r169;


dp4a.s32.s32 %r177, %r178, %r146, %r173;


dp4a.s32.s32 %r181, %r182, %r150, %r177;


dp4a.s32.s32 %r185, %r186, %r154, %r181;


dp4a.s32.s32 %r189, %r190, %r158, %r185;


dp4a.s32.s32 %r193, %r194, %r162, %r189;

mul.lo.s32 %r348, %r345, %r193;

{ cvt.f32.f16 %f31, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r164;
cvt.f32.f16 %f32, low;}


mul.ftz.f32 %f39, %f31, %f32;
cvt.rn.f32.s32 %f40, %r348;
fma.rn.ftz.f32 %f105, %f39, %f40, %f105;
add.s32 %r430, %r430, 128;
add.s32 %r429, %r429, 128;
add.s64 %rd52, %rd52, 2176;
add.s32 %r431, %r431, 16;
setp.lt.s32 %p4, %r431, %r40;
@%p4 bra $L__BB297_2;

$L__BB297_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB297_5;

shl.b32 %r350, %r1, 9;
mov.u32 %r351, _ZZ13mul_mat_vec_qIL9ggml_type23ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r352, %r351, %r350;
shl.b32 %r353, %r2, 2;
add.s32 %r354, %r352, %r353;
st.shared.f32 [%r354+-512], %f108;
st.shared.f32 [%r354+-384], %f107;
st.shared.f32 [%r354+-256], %f106;
st.shared.f32 [%r354+-128], %f105;

$L__BB297_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB297_14;

mad.lo.s32 %r355, %r5, %r28, %r4;
mad.lo.s32 %r356, %r6, %r34, %r355;
cvt.u64.u32 %rd13, %r356;
shl.b32 %r357, %r2, 2;
mov.u32 %r358, _ZZ13mul_mat_vec_qIL9ggml_type23ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r17, %r358, %r357;
mov.u32 %r359, 2;
ld.shared.f32 %f41, [%r17];
add.ftz.f32 %f42, %f41, %f108;
ld.shared.f32 %f43, [%r17+512];
add.ftz.f32 %f44, %f43, %f42;
ld.shared.f32 %f45, [%r17+1024];
add.ftz.f32 %f46, %f45, %f44;
mov.b32 %r360, %f46;
mov.u32 %r361, 31;
mov.u32 %r362, 16;
mov.u32 %r363, -1;
shfl.sync.bfly.b32 %r364|%p7, %r360, %r362, %r361, %r363;
mov.b32 %f47, %r364;
add.ftz.f32 %f48, %f46, %f47;
mov.b32 %r365, %f48;
mov.u32 %r366, 8;
shfl.sync.bfly.b32 %r367|%p8, %r365, %r366, %r361, %r363;
mov.b32 %f49, %r367;
add.ftz.f32 %f50, %f48, %f49;
mov.b32 %r368, %f50;
mov.u32 %r369, 4;
shfl.sync.bfly.b32 %r370|%p9, %r368, %r369, %r361, %r363;
mov.b32 %f51, %r370;
add.ftz.f32 %f52, %f50, %f51;
mov.b32 %r371, %f52;
shfl.sync.bfly.b32 %r372|%p10, %r371, %r359, %r361, %r363;
mov.b32 %f53, %r372;
add.ftz.f32 %f54, %f52, %f53;
mov.b32 %r373, %f54;
mov.u32 %r374, 1;
shfl.sync.bfly.b32 %r375|%p11, %r373, %r374, %r361, %r363;
mov.b32 %f55, %r375;
add.ftz.f32 %f13, %f54, %f55;
setp.lt.u32 %p12, %r4, %r22;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
mul.wide.u32 %rd43, %r356, 4;
add.s64 %rd14, %rd1, %rd43;
not.pred %p14, %p1;
@%p14 bra $L__BB297_8;

st.global.f32 [%rd14], %f13;

$L__BB297_8:
ld.shared.f32 %f56, [%r17+128];
add.ftz.f32 %f57, %f56, %f107;
ld.shared.f32 %f58, [%r17+640];
add.ftz.f32 %f59, %f58, %f57;
ld.shared.f32 %f60, [%r17+1152];
add.ftz.f32 %f61, %f60, %f59;
mov.b32 %r376, %f61;
shfl.sync.bfly.b32 %r380|%p15, %r376, %r362, %r361, %r363;
mov.b32 %f62, %r380;
add.ftz.f32 %f63, %f61, %f62;
mov.b32 %r381, %f63;
shfl.sync.bfly.b32 %r383|%p16, %r381, %r366, %r361, %r363;
mov.b32 %f64, %r383;
add.ftz.f32 %f65, %f63, %f64;
mov.b32 %r384, %f65;
shfl.sync.bfly.b32 %r386|%p17, %r384, %r369, %r361, %r363;
mov.b32 %f66, %r386;
add.ftz.f32 %f67, %f65, %f66;
mov.b32 %r387, %f67;
shfl.sync.bfly.b32 %r389|%p18, %r387, %r359, %r361, %r363;
mov.b32 %f68, %r389;
add.ftz.f32 %f69, %f67, %f68;
mov.b32 %r390, %f69;
shfl.sync.bfly.b32 %r392|%p19, %r390, %r374, %r361, %r363;
mov.b32 %f70, %r392;
add.ftz.f32 %f14, %f69, %f70;
add.s32 %r393, %r4, 1;
setp.lt.u32 %p20, %r393, %r22;
setp.eq.s32 %p21, %r2, 1;
and.pred %p2, %p21, %p20;
not.pred %p22, %p2;
@%p22 bra $L__BB297_10;

st.global.f32 [%rd14+4], %f14;

$L__BB297_10:
ld.shared.f32 %f71, [%r17+256];
add.ftz.f32 %f72, %f71, %f106;
ld.shared.f32 %f73, [%r17+768];
add.ftz.f32 %f74, %f73, %f72;
ld.shared.f32 %f75, [%r17+1280];
add.ftz.f32 %f76, %f75, %f74;
mov.b32 %r394, %f76;
mov.u32 %r395, 31;
mov.u32 %r396, 16;
mov.u32 %r397, -1;
shfl.sync.bfly.b32 %r398|%p23, %r394, %r396, %r395, %r397;
mov.b32 %f77, %r398;
add.ftz.f32 %f78, %f76, %f77;
mov.b32 %r399, %f78;
mov.u32 %r400, 8;
shfl.sync.bfly.b32 %r401|%p24, %r399, %r400, %r395, %r397;
mov.b32 %f79, %r401;
add.ftz.f32 %f80, %f78, %f79;
mov.b32 %r402, %f80;
mov.u32 %r403, 4;
shfl.sync.bfly.b32 %r404|%p25, %r402, %r403, %r395, %r397;
mov.b32 %f81, %r404;
add.ftz.f32 %f82, %f80, %f81;
mov.b32 %r405, %f82;
mov.u32 %r406, 2;
shfl.sync.bfly.b32 %r407|%p26, %r405, %r406, %r395, %r397;
mov.b32 %f83, %r407;
add.ftz.f32 %f84, %f82, %f83;
mov.b32 %r408, %f84;
mov.u32 %r409, 1;
shfl.sync.bfly.b32 %r410|%p27, %r408, %r409, %r395, %r397;
mov.b32 %f85, %r410;
add.ftz.f32 %f15, %f84, %f85;
@%p14 bra $L__BB297_12;

cvt.u64.u32 %rd44, %r22;
add.s64 %rd45, %rd44, %rd13;
shl.b64 %rd46, %rd45, 2;
add.s64 %rd47, %rd1, %rd46;
st.global.f32 [%rd47], %f15;

$L__BB297_12:
ld.shared.f32 %f86, [%r17+384];
add.ftz.f32 %f87, %f86, %f105;
ld.shared.f32 %f88, [%r17+896];
add.ftz.f32 %f89, %f88, %f87;
ld.shared.f32 %f90, [%r17+1408];
add.ftz.f32 %f91, %f90, %f89;
mov.b32 %r411, %f91;
shfl.sync.bfly.b32 %r415|%p29, %r411, %r396, %r395, %r397;
mov.b32 %f92, %r415;
add.ftz.f32 %f93, %f91, %f92;
mov.b32 %r416, %f93;
shfl.sync.bfly.b32 %r418|%p30, %r416, %r400, %r395, %r397;
mov.b32 %f94, %r418;
add.ftz.f32 %f95, %f93, %f94;
mov.b32 %r419, %f95;
shfl.sync.bfly.b32 %r421|%p31, %r419, %r403, %r395, %r397;
mov.b32 %f96, %r421;
add.ftz.f32 %f97, %f95, %f96;
mov.b32 %r422, %f97;
shfl.sync.bfly.b32 %r424|%p32, %r422, %r406, %r395, %r397;
mov.b32 %f98, %r424;
add.ftz.f32 %f99, %f97, %f98;
mov.b32 %r425, %f99;
shfl.sync.bfly.b32 %r427|%p33, %r425, %r409, %r395, %r397;
mov.b32 %f100, %r427;
add.ftz.f32 %f16, %f99, %f100;
@%p22 bra $L__BB297_14;

add.s32 %r428, %r22, 1;
cvt.u64.u32 %rd48, %r428;
add.s64 %rd49, %rd48, %rd13;
shl.b64 %rd50, %rd49, 2;
add.s64 %rd51, %rd1, %rd50;
st.global.f32 [%rd51], %f16;

$L__BB297_14:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type23ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type23ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type23ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type23ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type23ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type23ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type23ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type23ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type23ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<47>;
.reg .b16 %rs<11>;
.reg .f32 %f<163>;
.reg .b32 %r<539>;
.reg .b64 %rd<62>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type23ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[2304];

ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type23ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type23ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type23ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type23ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type23ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type23ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type23ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type23ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type23ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type23ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type23ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type23ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type23ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type23ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type23ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type23ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type23ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
cvta.to.global.u64 %rd1, %rd18;
mov.u32 %r1, %tid.y;
shl.b32 %r38, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r38, %r2;
mov.u32 %r39, %ctaid.x;
shl.b32 %r4, %r39, 1;
mov.u32 %r5, %ctaid.y;
mov.u32 %r6, %ctaid.z;
shr.s32 %r40, %r3, 31;
shr.u32 %r41, %r40, 29;
add.s32 %r42, %r3, %r41;
shr.s32 %r538, %r42, 3;
shr.u32 %r43, %r22, 8;
setp.ge.s32 %p3, %r538, %r43;
mov.f32 %f157, 0f00000000;
mov.f32 %f158, %f157;
mov.f32 %f159, %f157;
mov.f32 %f160, %f157;
mov.f32 %f161, %f157;
mov.f32 %f162, %f157;
@%p3 bra $L__BB298_3;

mul.lo.s32 %r44, %r5, %r30;
cvt.u64.u32 %rd19, %r44;
mul.lo.s32 %r45, %r6, %r36;
cvt.u64.u32 %rd20, %r45;
add.s64 %rd2, %rd20, %rd19;
mul.hi.u32 %r46, %r6, %r32;
add.s32 %r47, %r46, %r6;
shr.u32 %r48, %r47, %r33;
mul.lo.s32 %r49, %r48, %r35;
mul.hi.u32 %r50, %r5, %r26;
add.s32 %r51, %r50, %r5;
shr.u32 %r52, %r51, %r27;
mul.lo.s32 %r53, %r52, %r29;
cvta.to.global.u64 %rd61, %rd16;
and.b32 %r57, %r42, -8;
sub.s32 %r8, %r3, %r57;
shl.b32 %r58, %r8, 2;
shr.u32 %r59, %r8, 31;
add.s32 %r60, %r8, %r59;
shr.s32 %r61, %r60, 1;
cvt.s64.s32 %rd21, %r61;
shl.b32 %r62, %r24, 1;
shl.b32 %r535, %r538, 3;
add.s32 %r537, %r535, %r62;
add.s32 %r536, %r24, %r535;
add.s32 %r63, %r538, %r49;
add.s32 %r64, %r63, %r53;
add.s32 %r65, %r4, 1;
mad.lo.s32 %r66, %r23, %r65, %r64;
mul.wide.s32 %rd4, %r66, 136;
mad.lo.s32 %r67, %r4, %r23, %r538;
add.s32 %r68, %r67, %r49;
add.s32 %r69, %r68, %r53;
mul.wide.s32 %rd5, %r69, 136;
add.s64 %rd22, %rd5, %rd21;
add.s64 %rd6, %rd22, 4;
add.s64 %rd23, %rd4, %rd21;
add.s64 %rd7, %rd23, 4;
mul.wide.s32 %rd24, %r58, 4;
add.s64 %rd25, %rd5, %rd24;
add.s64 %rd8, %rd25, 20;
add.s64 %rd26, %rd4, %rd24;
add.s64 %rd9, %rd26, 12;
cvta.to.global.u64 %rd10, %rd17;
mov.f32 %f157, 0f00000000;

$L__BB298_2:
cvt.u64.u32 %rd27, %r535;
add.s64 %rd28, %rd2, %rd27;
cvt.s64.s32 %rd29, %r8;
add.s64 %rd30, %rd28, %rd29;
mul.lo.s64 %rd31, %rd30, 36;
add.s64 %rd32, %rd61, %rd8;
ld.global.nc.u32 %r268, [%rd32+-12];
and.b32 %r269, %r268, -2004318072;
shr.u32 %r270, %r269, 1;
or.b32 %r271, %r270, 839922192;
and.b32 %r272, %r268, 30583;
mov.u32 %r273, -152379953;
mov.u32 %r274, -1079142271;
prmt.b32 %r275, %r274, %r273, %r272;
mov.u32 %r276, 1901675829;
mov.u32 %r277, 639175937;
prmt.b32 %r278, %r277, %r276, %r272;
and.b32 %r279, %r271, 30292;
prmt.b32 %r280, %r275, %r278, %r279;
shr.u32 %r281, %r268, 16;
and.b32 %r282, %r281, 30583;
prmt.b32 %r283, %r274, %r273, %r282;
prmt.b32 %r284, %r277, %r276, %r282;
shr.u32 %r285, %r271, 16;
prmt.b32 %r286, %r283, %r284, %r285;
mov.u32 %r287, 25632;
prmt.b32 %r203, %r280, %r286, %r287;
mov.u32 %r288, 30001;
prmt.b32 %r207, %r280, %r286, %r288;
add.s64 %rd33, %rd10, %rd31;
ld.global.nc.u32 %r72, [%rd33+4];
ld.global.nc.u32 %r76, [%rd33+20];
mov.u32 %r238, 0;

dp4a.s32.s32 %r70, %r203, %r72, %r238;


dp4a.s32.s32 %r74, %r207, %r76, %r70;

ld.global.nc.u32 %r289, [%rd32+-8];
and.b32 %r290, %r289, -2004318072;
shr.u32 %r291, %r290, 1;
or.b32 %r292, %r291, 839922192;
and.b32 %r293, %r289, 30583;
prmt.b32 %r294, %r274, %r273, %r293;
prmt.b32 %r295, %r277, %r276, %r293;
and.b32 %r296, %r292, 30292;
prmt.b32 %r297, %r294, %r295, %r296;
shr.u32 %r298, %r289, 16;
and.b32 %r299, %r298, 30583;
prmt.b32 %r300, %r274, %r273, %r299;
prmt.b32 %r301, %r277, %r276, %r299;
shr.u32 %r302, %r292, 16;
prmt.b32 %r303, %r300, %r301, %r302;
prmt.b32 %r211, %r297, %r303, %r287;
prmt.b32 %r215, %r297, %r303, %r288;
ld.global.nc.u32 %r80, [%rd33+8];
ld.global.nc.u32 %r84, [%rd33+24];

dp4a.s32.s32 %r78, %r211, %r80, %r74;


dp4a.s32.s32 %r82, %r215, %r84, %r78;

ld.global.nc.u32 %r304, [%rd32+-4];
and.b32 %r305, %r304, -2004318072;
shr.u32 %r306, %r305, 1;
or.b32 %r307, %r306, 839922192;
and.b32 %r308, %r304, 30583;
prmt.b32 %r309, %r274, %r273, %r308;
prmt.b32 %r310, %r277, %r276, %r308;
and.b32 %r311, %r307, 30292;
prmt.b32 %r312, %r309, %r310, %r311;
shr.u32 %r313, %r304, 16;
and.b32 %r314, %r313, 30583;
prmt.b32 %r315, %r274, %r273, %r314;
prmt.b32 %r316, %r277, %r276, %r314;
shr.u32 %r317, %r307, 16;
prmt.b32 %r318, %r315, %r316, %r317;
prmt.b32 %r219, %r312, %r318, %r287;
prmt.b32 %r223, %r312, %r318, %r288;
ld.global.nc.u32 %r88, [%rd33+12];
ld.global.nc.u32 %r92, [%rd33+28];

dp4a.s32.s32 %r86, %r219, %r88, %r82;


dp4a.s32.s32 %r90, %r223, %r92, %r86;

ld.global.nc.u32 %r319, [%rd32];
and.b32 %r320, %r319, -2004318072;
shr.u32 %r321, %r320, 1;
or.b32 %r322, %r321, 839922192;
and.b32 %r323, %r319, 30583;
prmt.b32 %r324, %r274, %r273, %r323;
prmt.b32 %r325, %r277, %r276, %r323;
and.b32 %r326, %r322, 30292;
prmt.b32 %r327, %r324, %r325, %r326;
shr.u32 %r328, %r319, 16;
and.b32 %r329, %r328, 30583;
prmt.b32 %r330, %r274, %r273, %r329;
prmt.b32 %r331, %r277, %r276, %r329;
shr.u32 %r332, %r322, 16;
prmt.b32 %r333, %r330, %r331, %r332;
prmt.b32 %r227, %r327, %r333, %r287;
prmt.b32 %r231, %r327, %r333, %r288;
ld.global.nc.u32 %r96, [%rd33+16];
ld.global.nc.u32 %r100, [%rd33+32];

dp4a.s32.s32 %r94, %r227, %r96, %r90;


dp4a.s32.s32 %r98, %r231, %r100, %r94;

add.s64 %rd34, %rd61, %rd6;
ld.global.nc.u8 %rs7, [%rd34];
cvt.u32.u16 %r334, %rs7;
and.b32 %r335, %r334, 255;
and.b32 %r337, %r58, 4;
shr.u32 %r338, %r335, %r337;
and.b32 %r339, %r338, 15;
add.s64 %rd35, %rd61, %rd5;
ld.global.nc.u16 %rs8, [%rd35+2];
cvt.u32.u16 %r340, %rs8;
shl.b32 %r341, %r8, 1;
shr.u32 %r342, %r340, %r341;
and.b32 %r343, %r342, 3;
bfi.b32 %r344, %r343, %r339, 4, 2;
add.s32 %r345, %r344, -32;
mul.lo.s32 %r346, %r345, %r98;
ld.global.nc.u16 %rs1, [%rd35];

{ cvt.f32.f16 %f37, %rs1;}


ld.global.nc.u32 %r102, [%rd33];

{.reg .f16 low,high;
mov.b32 {low,high},%r102;
cvt.f32.f16 %f38, low;}


mul.ftz.f32 %f49, %f37, %f38;
cvt.rn.f32.s32 %f50, %r346;
fma.rn.ftz.f32 %f162, %f49, %f50, %f162;
add.s64 %rd36, %rd61, %rd9;
ld.global.nc.u32 %r347, [%rd36+-4];
and.b32 %r348, %r347, -2004318072;
shr.u32 %r349, %r348, 1;
or.b32 %r350, %r349, 839922192;
and.b32 %r351, %r347, 30583;
prmt.b32 %r352, %r274, %r273, %r351;
prmt.b32 %r353, %r277, %r276, %r351;
and.b32 %r354, %r350, 30292;
prmt.b32 %r355, %r352, %r353, %r354;
shr.u32 %r356, %r347, 16;
and.b32 %r357, %r356, 30583;
prmt.b32 %r358, %r274, %r273, %r357;
prmt.b32 %r359, %r277, %r276, %r357;
shr.u32 %r360, %r350, 16;
prmt.b32 %r361, %r358, %r359, %r360;
prmt.b32 %r236, %r355, %r361, %r287;
prmt.b32 %r240, %r355, %r361, %r288;

dp4a.s32.s32 %r103, %r236, %r72, %r238;


dp4a.s32.s32 %r107, %r240, %r76, %r103;

ld.global.nc.u32 %r362, [%rd36];
and.b32 %r363, %r362, -2004318072;
shr.u32 %r364, %r363, 1;
or.b32 %r365, %r364, 839922192;
and.b32 %r366, %r362, 30583;
prmt.b32 %r367, %r274, %r273, %r366;
prmt.b32 %r368, %r277, %r276, %r366;
and.b32 %r369, %r365, 30292;
prmt.b32 %r370, %r367, %r368, %r369;
shr.u32 %r371, %r362, 16;
and.b32 %r372, %r371, 30583;
prmt.b32 %r373, %r274, %r273, %r372;
prmt.b32 %r374, %r277, %r276, %r372;
shr.u32 %r375, %r365, 16;
prmt.b32 %r376, %r373, %r374, %r375;
prmt.b32 %r244, %r370, %r376, %r287;
prmt.b32 %r248, %r370, %r376, %r288;

dp4a.s32.s32 %r111, %r244, %r80, %r107;


dp4a.s32.s32 %r115, %r248, %r84, %r111;

ld.global.nc.u32 %r377, [%rd36+4];
and.b32 %r378, %r377, -2004318072;
shr.u32 %r379, %r378, 1;
or.b32 %r380, %r379, 839922192;
and.b32 %r381, %r377, 30583;
prmt.b32 %r382, %r274, %r273, %r381;
prmt.b32 %r383, %r277, %r276, %r381;
and.b32 %r384, %r380, 30292;
prmt.b32 %r385, %r382, %r383, %r384;
shr.u32 %r386, %r377, 16;
and.b32 %r387, %r386, 30583;
prmt.b32 %r388, %r274, %r273, %r387;
prmt.b32 %r389, %r277, %r276, %r387;
shr.u32 %r390, %r380, 16;
prmt.b32 %r391, %r388, %r389, %r390;
prmt.b32 %r252, %r385, %r391, %r287;
prmt.b32 %r256, %r385, %r391, %r288;

dp4a.s32.s32 %r119, %r252, %r88, %r115;


dp4a.s32.s32 %r123, %r256, %r92, %r119;

ld.global.nc.u32 %r392, [%rd36+8];
and.b32 %r393, %r392, -2004318072;
shr.u32 %r394, %r393, 1;
or.b32 %r395, %r394, 839922192;
and.b32 %r396, %r392, 30583;
prmt.b32 %r397, %r274, %r273, %r396;
prmt.b32 %r398, %r277, %r276, %r396;
and.b32 %r399, %r395, 30292;
prmt.b32 %r400, %r397, %r398, %r399;
shr.u32 %r401, %r392, 16;
and.b32 %r402, %r401, 30583;
prmt.b32 %r403, %r274, %r273, %r402;
prmt.b32 %r404, %r277, %r276, %r402;
shr.u32 %r405, %r395, 16;
prmt.b32 %r406, %r403, %r404, %r405;
prmt.b32 %r260, %r400, %r406, %r287;
prmt.b32 %r264, %r400, %r406, %r288;

dp4a.s32.s32 %r127, %r260, %r96, %r123;


dp4a.s32.s32 %r131, %r264, %r100, %r127;

add.s64 %rd37, %rd61, %rd7;
ld.global.nc.u8 %rs9, [%rd37];
cvt.u32.u16 %r407, %rs9;
and.b32 %r408, %r407, 255;
shr.u32 %r409, %r408, %r337;
and.b32 %r410, %r409, 15;
add.s64 %rd38, %rd61, %rd4;
ld.global.nc.u16 %rs10, [%rd38+2];
cvt.u32.u16 %r411, %rs10;
shr.u32 %r412, %r411, %r341;
and.b32 %r413, %r412, 3;
bfi.b32 %r414, %r413, %r410, 4, 2;
add.s32 %r415, %r414, -32;
mul.lo.s32 %r416, %r415, %r131;
ld.global.nc.u16 %rs2, [%rd38];

{ cvt.f32.f16 %f39, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r102;
cvt.f32.f16 %f40, low;}


mul.ftz.f32 %f51, %f39, %f40;
cvt.rn.f32.s32 %f52, %r416;
fma.rn.ftz.f32 %f161, %f51, %f52, %f161;
cvt.u64.u32 %rd39, %r536;
add.s64 %rd40, %rd2, %rd39;
add.s64 %rd41, %rd40, %rd29;
mul.lo.s64 %rd42, %rd41, 36;
add.s64 %rd43, %rd10, %rd42;
ld.global.nc.u32 %r138, [%rd43+4];
ld.global.nc.u32 %r142, [%rd43+20];

dp4a.s32.s32 %r136, %r203, %r138, %r238;


dp4a.s32.s32 %r140, %r207, %r142, %r136;

ld.global.nc.u32 %r146, [%rd43+8];
ld.global.nc.u32 %r150, [%rd43+24];

dp4a.s32.s32 %r144, %r211, %r146, %r140;


dp4a.s32.s32 %r148, %r215, %r150, %r144;

ld.global.nc.u32 %r154, [%rd43+12];
ld.global.nc.u32 %r158, [%rd43+28];

dp4a.s32.s32 %r152, %r219, %r154, %r148;


dp4a.s32.s32 %r156, %r223, %r158, %r152;

ld.global.nc.u32 %r162, [%rd43+16];
ld.global.nc.u32 %r166, [%rd43+32];

dp4a.s32.s32 %r160, %r227, %r162, %r156;


dp4a.s32.s32 %r164, %r231, %r166, %r160;

mul.lo.s32 %r417, %r345, %r164;

{ cvt.f32.f16 %f41, %rs1;}


ld.global.nc.u32 %r168, [%rd43];

{.reg .f16 low,high;
mov.b32 {low,high},%r168;
cvt.f32.f16 %f42, low;}


mul.ftz.f32 %f53, %f41, %f42;
cvt.rn.f32.s32 %f54, %r417;
fma.rn.ftz.f32 %f160, %f53, %f54, %f160;

dp4a.s32.s32 %r169, %r236, %r138, %r238;


dp4a.s32.s32 %r173, %r240, %r142, %r169;


dp4a.s32.s32 %r177, %r244, %r146, %r173;


dp4a.s32.s32 %r181, %r248, %r150, %r177;


dp4a.s32.s32 %r185, %r252, %r154, %r181;


dp4a.s32.s32 %r189, %r256, %r158, %r185;


dp4a.s32.s32 %r193, %r260, %r162, %r189;


dp4a.s32.s32 %r197, %r264, %r166, %r193;

mul.lo.s32 %r418, %r415, %r197;

{ cvt.f32.f16 %f43, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r168;
cvt.f32.f16 %f44, low;}


mul.ftz.f32 %f55, %f43, %f44;
cvt.rn.f32.s32 %f56, %r418;
fma.rn.ftz.f32 %f159, %f55, %f56, %f159;
cvt.u64.u32 %rd44, %r537;
add.s64 %rd45, %rd2, %rd44;
add.s64 %rd46, %rd45, %rd29;
mul.lo.s64 %rd47, %rd46, 36;
add.s64 %rd48, %rd10, %rd47;
ld.global.nc.u32 %r204, [%rd48+4];
ld.global.nc.u32 %r208, [%rd48+20];

dp4a.s32.s32 %r202, %r203, %r204, %r238;


dp4a.s32.s32 %r206, %r207, %r208, %r202;

ld.global.nc.u32 %r212, [%rd48+8];
ld.global.nc.u32 %r216, [%rd48+24];

dp4a.s32.s32 %r210, %r211, %r212, %r206;


dp4a.s32.s32 %r214, %r215, %r216, %r210;

ld.global.nc.u32 %r220, [%rd48+12];
ld.global.nc.u32 %r224, [%rd48+28];

dp4a.s32.s32 %r218, %r219, %r220, %r214;


dp4a.s32.s32 %r222, %r223, %r224, %r218;

ld.global.nc.u32 %r228, [%rd48+16];
ld.global.nc.u32 %r232, [%rd48+32];

dp4a.s32.s32 %r226, %r227, %r228, %r222;


dp4a.s32.s32 %r230, %r231, %r232, %r226;

mul.lo.s32 %r419, %r345, %r230;

{ cvt.f32.f16 %f45, %rs1;}


ld.global.nc.u32 %r234, [%rd48];

{.reg .f16 low,high;
mov.b32 {low,high},%r234;
cvt.f32.f16 %f46, low;}


mul.ftz.f32 %f57, %f45, %f46;
cvt.rn.f32.s32 %f58, %r419;
fma.rn.ftz.f32 %f158, %f57, %f58, %f158;

dp4a.s32.s32 %r235, %r236, %r204, %r238;


dp4a.s32.s32 %r239, %r240, %r208, %r235;


dp4a.s32.s32 %r243, %r244, %r212, %r239;


dp4a.s32.s32 %r247, %r248, %r216, %r243;


dp4a.s32.s32 %r251, %r252, %r220, %r247;


dp4a.s32.s32 %r255, %r256, %r224, %r251;


dp4a.s32.s32 %r259, %r260, %r228, %r255;


dp4a.s32.s32 %r263, %r264, %r232, %r259;

mul.lo.s32 %r420, %r415, %r263;

{ cvt.f32.f16 %f47, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r234;
cvt.f32.f16 %f48, low;}


mul.ftz.f32 %f59, %f47, %f48;
cvt.rn.f32.s32 %f60, %r420;
fma.rn.ftz.f32 %f157, %f59, %f60, %f157;
add.s32 %r537, %r537, 128;
add.s32 %r536, %r536, 128;
add.s32 %r535, %r535, 128;
add.s64 %rd61, %rd61, 2176;
add.s32 %r538, %r538, 16;
setp.lt.s32 %p4, %r538, %r43;
@%p4 bra $L__BB298_2;

$L__BB298_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB298_5;

mov.u32 %r422, _ZZ13mul_mat_vec_qIL9ggml_type23ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r423, %r1, 768, %r422;
shl.b32 %r424, %r2, 2;
add.s32 %r425, %r423, %r424;
st.shared.f32 [%r425+-768], %f162;
st.shared.f32 [%r425+-640], %f161;
st.shared.f32 [%r425+-512], %f160;
st.shared.f32 [%r425+-384], %f159;
st.shared.f32 [%r425+-256], %f158;
st.shared.f32 [%r425+-128], %f157;

$L__BB298_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB298_18;

mad.lo.s32 %r426, %r5, %r31, %r4;
mad.lo.s32 %r427, %r6, %r37, %r426;
cvt.u64.u32 %rd13, %r427;
shl.b32 %r428, %r2, 2;
mov.u32 %r429, _ZZ13mul_mat_vec_qIL9ggml_type23ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r20, %r429, %r428;
mov.u32 %r430, 2;
ld.shared.f32 %f61, [%r20];
add.ftz.f32 %f62, %f61, %f162;
ld.shared.f32 %f63, [%r20+768];
add.ftz.f32 %f64, %f63, %f62;
ld.shared.f32 %f65, [%r20+1536];
add.ftz.f32 %f66, %f65, %f64;
mov.b32 %r431, %f66;
mov.u32 %r432, 31;
mov.u32 %r433, 16;
mov.u32 %r434, -1;
shfl.sync.bfly.b32 %r435|%p7, %r431, %r433, %r432, %r434;
mov.b32 %f67, %r435;
add.ftz.f32 %f68, %f66, %f67;
mov.b32 %r436, %f68;
mov.u32 %r437, 8;
shfl.sync.bfly.b32 %r438|%p8, %r436, %r437, %r432, %r434;
mov.b32 %f69, %r438;
add.ftz.f32 %f70, %f68, %f69;
mov.b32 %r439, %f70;
mov.u32 %r440, 4;
shfl.sync.bfly.b32 %r441|%p9, %r439, %r440, %r432, %r434;
mov.b32 %f71, %r441;
add.ftz.f32 %f72, %f70, %f71;
mov.b32 %r442, %f72;
shfl.sync.bfly.b32 %r443|%p10, %r442, %r430, %r432, %r434;
mov.b32 %f73, %r443;
add.ftz.f32 %f74, %f72, %f73;
mov.b32 %r444, %f74;
mov.u32 %r445, 1;
shfl.sync.bfly.b32 %r446|%p11, %r444, %r445, %r432, %r434;
mov.b32 %f75, %r446;
add.ftz.f32 %f19, %f74, %f75;
setp.lt.u32 %p12, %r4, %r25;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
mul.wide.u32 %rd49, %r427, 4;
add.s64 %rd14, %rd1, %rd49;
not.pred %p14, %p1;
@%p14 bra $L__BB298_8;

st.global.f32 [%rd14], %f19;

$L__BB298_8:
ld.shared.f32 %f76, [%r20+128];
add.ftz.f32 %f77, %f76, %f161;
ld.shared.f32 %f78, [%r20+896];
add.ftz.f32 %f79, %f78, %f77;
ld.shared.f32 %f80, [%r20+1664];
add.ftz.f32 %f81, %f80, %f79;
mov.b32 %r447, %f81;
shfl.sync.bfly.b32 %r451|%p15, %r447, %r433, %r432, %r434;
mov.b32 %f82, %r451;
add.ftz.f32 %f83, %f81, %f82;
mov.b32 %r452, %f83;
shfl.sync.bfly.b32 %r454|%p16, %r452, %r437, %r432, %r434;
mov.b32 %f84, %r454;
add.ftz.f32 %f85, %f83, %f84;
mov.b32 %r455, %f85;
shfl.sync.bfly.b32 %r457|%p17, %r455, %r440, %r432, %r434;
mov.b32 %f86, %r457;
add.ftz.f32 %f87, %f85, %f86;
mov.b32 %r458, %f87;
shfl.sync.bfly.b32 %r460|%p18, %r458, %r430, %r432, %r434;
mov.b32 %f88, %r460;
add.ftz.f32 %f89, %f87, %f88;
mov.b32 %r461, %f89;
shfl.sync.bfly.b32 %r463|%p19, %r461, %r445, %r432, %r434;
mov.b32 %f90, %r463;
add.ftz.f32 %f20, %f89, %f90;
add.s32 %r464, %r4, 1;
setp.lt.u32 %p20, %r464, %r25;
setp.eq.s32 %p21, %r2, 1;
and.pred %p2, %p21, %p20;
not.pred %p22, %p2;
@%p22 bra $L__BB298_10;

st.global.f32 [%rd14+4], %f20;

$L__BB298_10:
ld.shared.f32 %f91, [%r20+256];
add.ftz.f32 %f92, %f91, %f160;
ld.shared.f32 %f93, [%r20+1024];
add.ftz.f32 %f94, %f93, %f92;
ld.shared.f32 %f95, [%r20+1792];
add.ftz.f32 %f96, %f95, %f94;
mov.b32 %r465, %f96;
mov.u32 %r466, 31;
mov.u32 %r467, 16;
mov.u32 %r468, -1;
shfl.sync.bfly.b32 %r469|%p23, %r465, %r467, %r466, %r468;
mov.b32 %f97, %r469;
add.ftz.f32 %f98, %f96, %f97;
mov.b32 %r470, %f98;
mov.u32 %r471, 8;
shfl.sync.bfly.b32 %r472|%p24, %r470, %r471, %r466, %r468;
mov.b32 %f99, %r472;
add.ftz.f32 %f100, %f98, %f99;
mov.b32 %r473, %f100;
mov.u32 %r474, 4;
shfl.sync.bfly.b32 %r475|%p25, %r473, %r474, %r466, %r468;
mov.b32 %f101, %r475;
add.ftz.f32 %f102, %f100, %f101;
mov.b32 %r476, %f102;
mov.u32 %r477, 2;
shfl.sync.bfly.b32 %r478|%p26, %r476, %r477, %r466, %r468;
mov.b32 %f103, %r478;
add.ftz.f32 %f104, %f102, %f103;
mov.b32 %r479, %f104;
mov.u32 %r480, 1;
shfl.sync.bfly.b32 %r481|%p27, %r479, %r480, %r466, %r468;
mov.b32 %f105, %r481;
add.ftz.f32 %f21, %f104, %f105;
@%p14 bra $L__BB298_12;

cvt.u64.u32 %rd50, %r25;
add.s64 %rd51, %rd50, %rd13;
shl.b64 %rd52, %rd51, 2;
add.s64 %rd53, %rd1, %rd52;
st.global.f32 [%rd53], %f21;

$L__BB298_12:
ld.shared.f32 %f106, [%r20+384];
add.ftz.f32 %f107, %f106, %f159;
ld.shared.f32 %f108, [%r20+1152];
add.ftz.f32 %f109, %f108, %f107;
ld.shared.f32 %f110, [%r20+1920];
add.ftz.f32 %f111, %f110, %f109;
mov.b32 %r482, %f111;
shfl.sync.bfly.b32 %r486|%p29, %r482, %r467, %r466, %r468;
mov.b32 %f112, %r486;
add.ftz.f32 %f113, %f111, %f112;
mov.b32 %r487, %f113;
shfl.sync.bfly.b32 %r489|%p30, %r487, %r471, %r466, %r468;
mov.b32 %f114, %r489;
add.ftz.f32 %f115, %f113, %f114;
mov.b32 %r490, %f115;
shfl.sync.bfly.b32 %r492|%p31, %r490, %r474, %r466, %r468;
mov.b32 %f116, %r492;
add.ftz.f32 %f117, %f115, %f116;
mov.b32 %r493, %f117;
shfl.sync.bfly.b32 %r495|%p32, %r493, %r477, %r466, %r468;
mov.b32 %f118, %r495;
add.ftz.f32 %f119, %f117, %f118;
mov.b32 %r496, %f119;
shfl.sync.bfly.b32 %r498|%p33, %r496, %r480, %r466, %r468;
mov.b32 %f120, %r498;
add.ftz.f32 %f22, %f119, %f120;
@%p22 bra $L__BB298_14;

add.s32 %r499, %r25, 1;
cvt.u64.u32 %rd54, %r499;
add.s64 %rd55, %rd54, %rd13;
shl.b64 %rd56, %rd55, 2;
add.s64 %rd57, %rd1, %rd56;
st.global.f32 [%rd57], %f22;

$L__BB298_14:
shl.b32 %r500, %r25, 1;
mov.u32 %r501, 1;
ld.shared.f32 %f121, [%r20+512];
add.ftz.f32 %f122, %f121, %f158;
ld.shared.f32 %f123, [%r20+1280];
add.ftz.f32 %f124, %f123, %f122;
ld.shared.f32 %f125, [%r20+2048];
add.ftz.f32 %f126, %f125, %f124;
mov.b32 %r502, %f126;
mov.u32 %r503, 31;
mov.u32 %r504, 16;
mov.u32 %r505, -1;
shfl.sync.bfly.b32 %r506|%p35, %r502, %r504, %r503, %r505;
mov.b32 %f127, %r506;
add.ftz.f32 %f128, %f126, %f127;
mov.b32 %r507, %f128;
mov.u32 %r508, 8;
shfl.sync.bfly.b32 %r509|%p36, %r507, %r508, %r503, %r505;
mov.b32 %f129, %r509;
add.ftz.f32 %f130, %f128, %f129;
mov.b32 %r510, %f130;
mov.u32 %r511, 4;
shfl.sync.bfly.b32 %r512|%p37, %r510, %r511, %r503, %r505;
mov.b32 %f131, %r512;
add.ftz.f32 %f132, %f130, %f131;
mov.b32 %r513, %f132;
mov.u32 %r514, 2;
shfl.sync.bfly.b32 %r515|%p38, %r513, %r514, %r503, %r505;
mov.b32 %f133, %r515;
add.ftz.f32 %f134, %f132, %f133;
mov.b32 %r516, %f134;
shfl.sync.bfly.b32 %r517|%p39, %r516, %r501, %r503, %r505;
mov.b32 %f135, %r517;
add.ftz.f32 %f23, %f134, %f135;
cvt.u64.u32 %rd58, %r500;
add.s64 %rd59, %rd58, %rd13;
shl.b64 %rd60, %rd59, 2;
add.s64 %rd15, %rd1, %rd60;
@%p14 bra $L__BB298_16;

st.global.f32 [%rd15], %f23;

$L__BB298_16:
ld.shared.f32 %f136, [%r20+640];
add.ftz.f32 %f137, %f136, %f157;
ld.shared.f32 %f138, [%r20+1408];
add.ftz.f32 %f139, %f138, %f137;
ld.shared.f32 %f140, [%r20+2176];
add.ftz.f32 %f141, %f140, %f139;
mov.b32 %r518, %f141;
shfl.sync.bfly.b32 %r522|%p41, %r518, %r504, %r503, %r505;
mov.b32 %f142, %r522;
add.ftz.f32 %f143, %f141, %f142;
mov.b32 %r523, %f143;
shfl.sync.bfly.b32 %r525|%p42, %r523, %r508, %r503, %r505;
mov.b32 %f144, %r525;
add.ftz.f32 %f145, %f143, %f144;
mov.b32 %r526, %f145;
shfl.sync.bfly.b32 %r528|%p43, %r526, %r511, %r503, %r505;
mov.b32 %f146, %r528;
add.ftz.f32 %f147, %f145, %f146;
mov.b32 %r529, %f147;
shfl.sync.bfly.b32 %r531|%p44, %r529, %r514, %r503, %r505;
mov.b32 %f148, %r531;
add.ftz.f32 %f149, %f147, %f148;
mov.b32 %r532, %f149;
shfl.sync.bfly.b32 %r534|%p45, %r532, %r501, %r503, %r505;
mov.b32 %f150, %r534;
add.ftz.f32 %f24, %f149, %f150;
@%p22 bra $L__BB298_18;

st.global.f32 [%rd15+4], %f24;

$L__BB298_18:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type23ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type23ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type23ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type23ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type23ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type23ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type23ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type23ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type23ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<59>;
.reg .b16 %rs<13>;
.reg .f32 %f<217>;
.reg .b32 %r<648>;
.reg .b64 %rd<75>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type23ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[3072];

ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type23ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type23ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type23ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type23ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type23ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type23ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type23ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type23ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type23ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type23ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type23ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type23ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type23ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type23ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type23ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type23ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type23ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
cvta.to.global.u64 %rd1, %rd18;
mov.u32 %r1, %tid.y;
shl.b32 %r42, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r42, %r2;
mov.u32 %r43, %ctaid.x;
shl.b32 %r4, %r43, 1;
mov.u32 %r5, %ctaid.y;
mov.u32 %r6, %ctaid.z;
shr.s32 %r44, %r3, 31;
shr.u32 %r45, %r44, 29;
add.s32 %r46, %r3, %r45;
shr.s32 %r647, %r46, 3;
shr.u32 %r47, %r26, 8;
setp.ge.s32 %p3, %r647, %r47;
mov.f32 %f209, 0f00000000;
mov.f32 %f210, %f209;
mov.f32 %f211, %f209;
mov.f32 %f212, %f209;
mov.f32 %f213, %f209;
mov.f32 %f214, %f209;
mov.f32 %f215, %f209;
mov.f32 %f216, %f209;
@%p3 bra $L__BB299_3;

mul.lo.s32 %r48, %r5, %r34;
cvt.u64.u32 %rd19, %r48;
mul.lo.s32 %r49, %r6, %r40;
cvt.u64.u32 %rd20, %r49;
add.s64 %rd2, %rd20, %rd19;
mul.hi.u32 %r50, %r6, %r36;
add.s32 %r51, %r50, %r6;
shr.u32 %r52, %r51, %r37;
mul.lo.s32 %r53, %r52, %r39;
mul.hi.u32 %r54, %r5, %r30;
add.s32 %r55, %r54, %r5;
shr.u32 %r56, %r55, %r31;
mul.lo.s32 %r57, %r56, %r33;
cvta.to.global.u64 %rd74, %rd16;
and.b32 %r61, %r46, -8;
sub.s32 %r8, %r3, %r61;
shl.b32 %r62, %r8, 2;
shr.u32 %r63, %r8, 31;
add.s32 %r64, %r8, %r63;
shr.s32 %r65, %r64, 1;
cvt.s64.s32 %rd21, %r65;
shl.b32 %r66, %r28, 1;
shl.b32 %r643, %r647, 3;
add.s32 %r646, %r643, %r66;
mad.lo.s32 %r645, %r28, 3, %r643;
add.s32 %r644, %r28, %r643;
add.s32 %r67, %r647, %r53;
add.s32 %r68, %r67, %r57;
add.s32 %r69, %r4, 1;
mad.lo.s32 %r70, %r27, %r69, %r68;
mul.wide.s32 %rd4, %r70, 136;
mad.lo.s32 %r71, %r4, %r27, %r647;
add.s32 %r72, %r71, %r53;
add.s32 %r73, %r72, %r57;
mul.wide.s32 %rd5, %r73, 136;
add.s64 %rd22, %rd5, %rd21;
add.s64 %rd6, %rd22, 4;
add.s64 %rd23, %rd4, %rd21;
add.s64 %rd7, %rd23, 4;
mul.wide.s32 %rd24, %r62, 4;
add.s64 %rd25, %rd5, %rd24;
add.s64 %rd8, %rd25, 20;
add.s64 %rd26, %rd4, %rd24;
add.s64 %rd9, %rd26, 12;
cvta.to.global.u64 %rd10, %rd17;
mov.f32 %f209, 0f00000000;

$L__BB299_2:
cvt.u64.u32 %rd27, %r643;
add.s64 %rd28, %rd2, %rd27;
cvt.s64.s32 %rd29, %r8;
add.s64 %rd30, %rd28, %rd29;
mul.lo.s64 %rd31, %rd30, 36;
add.s64 %rd32, %rd74, %rd8;
ld.global.nc.u32 %r338, [%rd32+-12];
and.b32 %r339, %r338, -2004318072;
shr.u32 %r340, %r339, 1;
or.b32 %r341, %r340, 839922192;
and.b32 %r342, %r338, 30583;
mov.u32 %r343, -152379953;
mov.u32 %r344, -1079142271;
prmt.b32 %r345, %r344, %r343, %r342;
mov.u32 %r346, 1901675829;
mov.u32 %r347, 639175937;
prmt.b32 %r348, %r347, %r346, %r342;
and.b32 %r349, %r341, 30292;
prmt.b32 %r350, %r345, %r348, %r349;
shr.u32 %r351, %r338, 16;
and.b32 %r352, %r351, 30583;
prmt.b32 %r353, %r344, %r343, %r352;
prmt.b32 %r354, %r347, %r346, %r352;
shr.u32 %r355, %r341, 16;
prmt.b32 %r356, %r353, %r354, %r355;
mov.u32 %r357, 25632;
prmt.b32 %r273, %r350, %r356, %r357;
mov.u32 %r358, 30001;
prmt.b32 %r277, %r350, %r356, %r358;
add.s64 %rd33, %rd10, %rd31;
ld.global.nc.u32 %r76, [%rd33+4];
ld.global.nc.u32 %r80, [%rd33+20];
mov.u32 %r308, 0;

dp4a.s32.s32 %r74, %r273, %r76, %r308;


dp4a.s32.s32 %r78, %r277, %r80, %r74;

ld.global.nc.u32 %r359, [%rd32+-8];
and.b32 %r360, %r359, -2004318072;
shr.u32 %r361, %r360, 1;
or.b32 %r362, %r361, 839922192;
and.b32 %r363, %r359, 30583;
prmt.b32 %r364, %r344, %r343, %r363;
prmt.b32 %r365, %r347, %r346, %r363;
and.b32 %r366, %r362, 30292;
prmt.b32 %r367, %r364, %r365, %r366;
shr.u32 %r368, %r359, 16;
and.b32 %r369, %r368, 30583;
prmt.b32 %r370, %r344, %r343, %r369;
prmt.b32 %r371, %r347, %r346, %r369;
shr.u32 %r372, %r362, 16;
prmt.b32 %r373, %r370, %r371, %r372;
prmt.b32 %r281, %r367, %r373, %r357;
prmt.b32 %r285, %r367, %r373, %r358;
ld.global.nc.u32 %r84, [%rd33+8];
ld.global.nc.u32 %r88, [%rd33+24];

dp4a.s32.s32 %r82, %r281, %r84, %r78;


dp4a.s32.s32 %r86, %r285, %r88, %r82;

ld.global.nc.u32 %r374, [%rd32+-4];
and.b32 %r375, %r374, -2004318072;
shr.u32 %r376, %r375, 1;
or.b32 %r377, %r376, 839922192;
and.b32 %r378, %r374, 30583;
prmt.b32 %r379, %r344, %r343, %r378;
prmt.b32 %r380, %r347, %r346, %r378;
and.b32 %r381, %r377, 30292;
prmt.b32 %r382, %r379, %r380, %r381;
shr.u32 %r383, %r374, 16;
and.b32 %r384, %r383, 30583;
prmt.b32 %r385, %r344, %r343, %r384;
prmt.b32 %r386, %r347, %r346, %r384;
shr.u32 %r387, %r377, 16;
prmt.b32 %r388, %r385, %r386, %r387;
prmt.b32 %r289, %r382, %r388, %r357;
prmt.b32 %r293, %r382, %r388, %r358;
ld.global.nc.u32 %r92, [%rd33+12];
ld.global.nc.u32 %r96, [%rd33+28];

dp4a.s32.s32 %r90, %r289, %r92, %r86;


dp4a.s32.s32 %r94, %r293, %r96, %r90;

ld.global.nc.u32 %r389, [%rd32];
and.b32 %r390, %r389, -2004318072;
shr.u32 %r391, %r390, 1;
or.b32 %r392, %r391, 839922192;
and.b32 %r393, %r389, 30583;
prmt.b32 %r394, %r344, %r343, %r393;
prmt.b32 %r395, %r347, %r346, %r393;
and.b32 %r396, %r392, 30292;
prmt.b32 %r397, %r394, %r395, %r396;
shr.u32 %r398, %r389, 16;
and.b32 %r399, %r398, 30583;
prmt.b32 %r400, %r344, %r343, %r399;
prmt.b32 %r401, %r347, %r346, %r399;
shr.u32 %r402, %r392, 16;
prmt.b32 %r403, %r400, %r401, %r402;
prmt.b32 %r297, %r397, %r403, %r357;
prmt.b32 %r301, %r397, %r403, %r358;
ld.global.nc.u32 %r100, [%rd33+16];
ld.global.nc.u32 %r104, [%rd33+32];

dp4a.s32.s32 %r98, %r297, %r100, %r94;


dp4a.s32.s32 %r102, %r301, %r104, %r98;

add.s64 %rd34, %rd74, %rd6;
ld.global.nc.u8 %rs9, [%rd34];
cvt.u32.u16 %r404, %rs9;
and.b32 %r405, %r404, 255;
and.b32 %r407, %r62, 4;
shr.u32 %r408, %r405, %r407;
and.b32 %r409, %r408, 15;
add.s64 %rd35, %rd74, %rd5;
ld.global.nc.u16 %rs10, [%rd35+2];
cvt.u32.u16 %r410, %rs10;
shl.b32 %r411, %r8, 1;
shr.u32 %r412, %r410, %r411;
and.b32 %r413, %r412, 3;
bfi.b32 %r414, %r413, %r409, 4, 2;
add.s32 %r415, %r414, -32;
mul.lo.s32 %r416, %r415, %r102;
ld.global.nc.u16 %rs1, [%rd35];

{ cvt.f32.f16 %f49, %rs1;}


ld.global.nc.u32 %r106, [%rd33];

{.reg .f16 low,high;
mov.b32 {low,high},%r106;
cvt.f32.f16 %f50, low;}


mul.ftz.f32 %f65, %f49, %f50;
cvt.rn.f32.s32 %f66, %r416;
fma.rn.ftz.f32 %f216, %f65, %f66, %f216;
add.s64 %rd36, %rd74, %rd9;
ld.global.nc.u32 %r417, [%rd36+-4];
and.b32 %r418, %r417, -2004318072;
shr.u32 %r419, %r418, 1;
or.b32 %r420, %r419, 839922192;
and.b32 %r421, %r417, 30583;
prmt.b32 %r422, %r344, %r343, %r421;
prmt.b32 %r423, %r347, %r346, %r421;
and.b32 %r424, %r420, 30292;
prmt.b32 %r425, %r422, %r423, %r424;
shr.u32 %r426, %r417, 16;
and.b32 %r427, %r426, 30583;
prmt.b32 %r428, %r344, %r343, %r427;
prmt.b32 %r429, %r347, %r346, %r427;
shr.u32 %r430, %r420, 16;
prmt.b32 %r431, %r428, %r429, %r430;
prmt.b32 %r306, %r425, %r431, %r357;
prmt.b32 %r310, %r425, %r431, %r358;

dp4a.s32.s32 %r107, %r306, %r76, %r308;


dp4a.s32.s32 %r111, %r310, %r80, %r107;

ld.global.nc.u32 %r432, [%rd36];
and.b32 %r433, %r432, -2004318072;
shr.u32 %r434, %r433, 1;
or.b32 %r435, %r434, 839922192;
and.b32 %r436, %r432, 30583;
prmt.b32 %r437, %r344, %r343, %r436;
prmt.b32 %r438, %r347, %r346, %r436;
and.b32 %r439, %r435, 30292;
prmt.b32 %r440, %r437, %r438, %r439;
shr.u32 %r441, %r432, 16;
and.b32 %r442, %r441, 30583;
prmt.b32 %r443, %r344, %r343, %r442;
prmt.b32 %r444, %r347, %r346, %r442;
shr.u32 %r445, %r435, 16;
prmt.b32 %r446, %r443, %r444, %r445;
prmt.b32 %r314, %r440, %r446, %r357;
prmt.b32 %r318, %r440, %r446, %r358;

dp4a.s32.s32 %r115, %r314, %r84, %r111;


dp4a.s32.s32 %r119, %r318, %r88, %r115;

ld.global.nc.u32 %r447, [%rd36+4];
and.b32 %r448, %r447, -2004318072;
shr.u32 %r449, %r448, 1;
or.b32 %r450, %r449, 839922192;
and.b32 %r451, %r447, 30583;
prmt.b32 %r452, %r344, %r343, %r451;
prmt.b32 %r453, %r347, %r346, %r451;
and.b32 %r454, %r450, 30292;
prmt.b32 %r455, %r452, %r453, %r454;
shr.u32 %r456, %r447, 16;
and.b32 %r457, %r456, 30583;
prmt.b32 %r458, %r344, %r343, %r457;
prmt.b32 %r459, %r347, %r346, %r457;
shr.u32 %r460, %r450, 16;
prmt.b32 %r461, %r458, %r459, %r460;
prmt.b32 %r322, %r455, %r461, %r357;
prmt.b32 %r326, %r455, %r461, %r358;

dp4a.s32.s32 %r123, %r322, %r92, %r119;


dp4a.s32.s32 %r127, %r326, %r96, %r123;

ld.global.nc.u32 %r462, [%rd36+8];
and.b32 %r463, %r462, -2004318072;
shr.u32 %r464, %r463, 1;
or.b32 %r465, %r464, 839922192;
and.b32 %r466, %r462, 30583;
prmt.b32 %r467, %r344, %r343, %r466;
prmt.b32 %r468, %r347, %r346, %r466;
and.b32 %r469, %r465, 30292;
prmt.b32 %r470, %r467, %r468, %r469;
shr.u32 %r471, %r462, 16;
and.b32 %r472, %r471, 30583;
prmt.b32 %r473, %r344, %r343, %r472;
prmt.b32 %r474, %r347, %r346, %r472;
shr.u32 %r475, %r465, 16;
prmt.b32 %r476, %r473, %r474, %r475;
prmt.b32 %r330, %r470, %r476, %r357;
prmt.b32 %r334, %r470, %r476, %r358;

dp4a.s32.s32 %r131, %r330, %r100, %r127;


dp4a.s32.s32 %r135, %r334, %r104, %r131;

add.s64 %rd37, %rd74, %rd7;
ld.global.nc.u8 %rs11, [%rd37];
cvt.u32.u16 %r477, %rs11;
and.b32 %r478, %r477, 255;
shr.u32 %r479, %r478, %r407;
and.b32 %r480, %r479, 15;
add.s64 %rd38, %rd74, %rd4;
ld.global.nc.u16 %rs12, [%rd38+2];
cvt.u32.u16 %r481, %rs12;
shr.u32 %r482, %r481, %r411;
and.b32 %r483, %r482, 3;
bfi.b32 %r484, %r483, %r480, 4, 2;
add.s32 %r485, %r484, -32;
mul.lo.s32 %r486, %r485, %r135;
ld.global.nc.u16 %rs2, [%rd38];

{ cvt.f32.f16 %f51, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r106;
cvt.f32.f16 %f52, low;}


mul.ftz.f32 %f67, %f51, %f52;
cvt.rn.f32.s32 %f68, %r486;
fma.rn.ftz.f32 %f215, %f67, %f68, %f215;
cvt.u64.u32 %rd39, %r644;
add.s64 %rd40, %rd2, %rd39;
add.s64 %rd41, %rd40, %rd29;
mul.lo.s64 %rd42, %rd41, 36;
add.s64 %rd43, %rd10, %rd42;
ld.global.nc.u32 %r142, [%rd43+4];
ld.global.nc.u32 %r146, [%rd43+20];

dp4a.s32.s32 %r140, %r273, %r142, %r308;


dp4a.s32.s32 %r144, %r277, %r146, %r140;

ld.global.nc.u32 %r150, [%rd43+8];
ld.global.nc.u32 %r154, [%rd43+24];

dp4a.s32.s32 %r148, %r281, %r150, %r144;


dp4a.s32.s32 %r152, %r285, %r154, %r148;

ld.global.nc.u32 %r158, [%rd43+12];
ld.global.nc.u32 %r162, [%rd43+28];

dp4a.s32.s32 %r156, %r289, %r158, %r152;


dp4a.s32.s32 %r160, %r293, %r162, %r156;

ld.global.nc.u32 %r166, [%rd43+16];
ld.global.nc.u32 %r170, [%rd43+32];

dp4a.s32.s32 %r164, %r297, %r166, %r160;


dp4a.s32.s32 %r168, %r301, %r170, %r164;

mul.lo.s32 %r487, %r415, %r168;

{ cvt.f32.f16 %f53, %rs1;}


ld.global.nc.u32 %r172, [%rd43];

{.reg .f16 low,high;
mov.b32 {low,high},%r172;
cvt.f32.f16 %f54, low;}


mul.ftz.f32 %f69, %f53, %f54;
cvt.rn.f32.s32 %f70, %r487;
fma.rn.ftz.f32 %f214, %f69, %f70, %f214;

dp4a.s32.s32 %r173, %r306, %r142, %r308;


dp4a.s32.s32 %r177, %r310, %r146, %r173;


dp4a.s32.s32 %r181, %r314, %r150, %r177;


dp4a.s32.s32 %r185, %r318, %r154, %r181;


dp4a.s32.s32 %r189, %r322, %r158, %r185;


dp4a.s32.s32 %r193, %r326, %r162, %r189;


dp4a.s32.s32 %r197, %r330, %r166, %r193;


dp4a.s32.s32 %r201, %r334, %r170, %r197;

mul.lo.s32 %r488, %r485, %r201;

{ cvt.f32.f16 %f55, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r172;
cvt.f32.f16 %f56, low;}


mul.ftz.f32 %f71, %f55, %f56;
cvt.rn.f32.s32 %f72, %r488;
fma.rn.ftz.f32 %f213, %f71, %f72, %f213;
cvt.u64.u32 %rd44, %r646;
add.s64 %rd45, %rd2, %rd44;
add.s64 %rd46, %rd45, %rd29;
mul.lo.s64 %rd47, %rd46, 36;
add.s64 %rd48, %rd10, %rd47;
ld.global.nc.u32 %r208, [%rd48+4];
ld.global.nc.u32 %r212, [%rd48+20];

dp4a.s32.s32 %r206, %r273, %r208, %r308;


dp4a.s32.s32 %r210, %r277, %r212, %r206;

ld.global.nc.u32 %r216, [%rd48+8];
ld.global.nc.u32 %r220, [%rd48+24];

dp4a.s32.s32 %r214, %r281, %r216, %r210;


dp4a.s32.s32 %r218, %r285, %r220, %r214;

ld.global.nc.u32 %r224, [%rd48+12];
ld.global.nc.u32 %r228, [%rd48+28];

dp4a.s32.s32 %r222, %r289, %r224, %r218;


dp4a.s32.s32 %r226, %r293, %r228, %r222;

ld.global.nc.u32 %r232, [%rd48+16];
ld.global.nc.u32 %r236, [%rd48+32];

dp4a.s32.s32 %r230, %r297, %r232, %r226;


dp4a.s32.s32 %r234, %r301, %r236, %r230;

mul.lo.s32 %r489, %r415, %r234;

{ cvt.f32.f16 %f57, %rs1;}


ld.global.nc.u32 %r238, [%rd48];

{.reg .f16 low,high;
mov.b32 {low,high},%r238;
cvt.f32.f16 %f58, low;}


mul.ftz.f32 %f73, %f57, %f58;
cvt.rn.f32.s32 %f74, %r489;
fma.rn.ftz.f32 %f212, %f73, %f74, %f212;

dp4a.s32.s32 %r239, %r306, %r208, %r308;


dp4a.s32.s32 %r243, %r310, %r212, %r239;


dp4a.s32.s32 %r247, %r314, %r216, %r243;


dp4a.s32.s32 %r251, %r318, %r220, %r247;


dp4a.s32.s32 %r255, %r322, %r224, %r251;


dp4a.s32.s32 %r259, %r326, %r228, %r255;


dp4a.s32.s32 %r263, %r330, %r232, %r259;


dp4a.s32.s32 %r267, %r334, %r236, %r263;

mul.lo.s32 %r490, %r485, %r267;

{ cvt.f32.f16 %f59, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r238;
cvt.f32.f16 %f60, low;}


mul.ftz.f32 %f75, %f59, %f60;
cvt.rn.f32.s32 %f76, %r490;
fma.rn.ftz.f32 %f211, %f75, %f76, %f211;
cvt.u64.u32 %rd49, %r645;
add.s64 %rd50, %rd2, %rd49;
add.s64 %rd51, %rd50, %rd29;
mul.lo.s64 %rd52, %rd51, 36;
add.s64 %rd53, %rd10, %rd52;
ld.global.nc.u32 %r274, [%rd53+4];
ld.global.nc.u32 %r278, [%rd53+20];

dp4a.s32.s32 %r272, %r273, %r274, %r308;


dp4a.s32.s32 %r276, %r277, %r278, %r272;

ld.global.nc.u32 %r282, [%rd53+8];
ld.global.nc.u32 %r286, [%rd53+24];

dp4a.s32.s32 %r280, %r281, %r282, %r276;


dp4a.s32.s32 %r284, %r285, %r286, %r280;

ld.global.nc.u32 %r290, [%rd53+12];
ld.global.nc.u32 %r294, [%rd53+28];

dp4a.s32.s32 %r288, %r289, %r290, %r284;


dp4a.s32.s32 %r292, %r293, %r294, %r288;

ld.global.nc.u32 %r298, [%rd53+16];
ld.global.nc.u32 %r302, [%rd53+32];

dp4a.s32.s32 %r296, %r297, %r298, %r292;


dp4a.s32.s32 %r300, %r301, %r302, %r296;

mul.lo.s32 %r491, %r415, %r300;

{ cvt.f32.f16 %f61, %rs1;}


ld.global.nc.u32 %r304, [%rd53];

{.reg .f16 low,high;
mov.b32 {low,high},%r304;
cvt.f32.f16 %f62, low;}


mul.ftz.f32 %f77, %f61, %f62;
cvt.rn.f32.s32 %f78, %r491;
fma.rn.ftz.f32 %f210, %f77, %f78, %f210;

dp4a.s32.s32 %r305, %r306, %r274, %r308;


dp4a.s32.s32 %r309, %r310, %r278, %r305;


dp4a.s32.s32 %r313, %r314, %r282, %r309;


dp4a.s32.s32 %r317, %r318, %r286, %r313;


dp4a.s32.s32 %r321, %r322, %r290, %r317;


dp4a.s32.s32 %r325, %r326, %r294, %r321;


dp4a.s32.s32 %r329, %r330, %r298, %r325;


dp4a.s32.s32 %r333, %r334, %r302, %r329;

mul.lo.s32 %r492, %r485, %r333;

{ cvt.f32.f16 %f63, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r304;
cvt.f32.f16 %f64, low;}


mul.ftz.f32 %f79, %f63, %f64;
cvt.rn.f32.s32 %f80, %r492;
fma.rn.ftz.f32 %f209, %f79, %f80, %f209;
add.s32 %r646, %r646, 128;
add.s32 %r645, %r645, 128;
add.s32 %r644, %r644, 128;
add.s32 %r643, %r643, 128;
add.s64 %rd74, %rd74, 2176;
add.s32 %r647, %r647, 16;
setp.lt.s32 %p4, %r647, %r47;
@%p4 bra $L__BB299_2;

$L__BB299_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB299_5;

shl.b32 %r494, %r1, 10;
mov.u32 %r495, _ZZ13mul_mat_vec_qIL9ggml_type23ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r496, %r495, %r494;
shl.b32 %r497, %r2, 2;
add.s32 %r498, %r496, %r497;
st.shared.f32 [%r498+-1024], %f216;
st.shared.f32 [%r498+-896], %f215;
st.shared.f32 [%r498+-768], %f214;
st.shared.f32 [%r498+-640], %f213;
st.shared.f32 [%r498+-512], %f212;
st.shared.f32 [%r498+-384], %f211;
st.shared.f32 [%r498+-256], %f210;
st.shared.f32 [%r498+-128], %f209;

$L__BB299_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB299_22;

mad.lo.s32 %r499, %r5, %r35, %r4;
mad.lo.s32 %r500, %r6, %r41, %r499;
cvt.u64.u32 %rd13, %r500;
shl.b32 %r501, %r2, 2;
mov.u32 %r502, _ZZ13mul_mat_vec_qIL9ggml_type23ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r23, %r502, %r501;
mov.u32 %r503, 2;
ld.shared.f32 %f81, [%r23];
add.ftz.f32 %f82, %f81, %f216;
ld.shared.f32 %f83, [%r23+1024];
add.ftz.f32 %f84, %f83, %f82;
ld.shared.f32 %f85, [%r23+2048];
add.ftz.f32 %f86, %f85, %f84;
mov.b32 %r504, %f86;
mov.u32 %r505, 31;
mov.u32 %r506, 16;
mov.u32 %r507, -1;
shfl.sync.bfly.b32 %r508|%p7, %r504, %r506, %r505, %r507;
mov.b32 %f87, %r508;
add.ftz.f32 %f88, %f86, %f87;
mov.b32 %r509, %f88;
mov.u32 %r510, 8;
shfl.sync.bfly.b32 %r511|%p8, %r509, %r510, %r505, %r507;
mov.b32 %f89, %r511;
add.ftz.f32 %f90, %f88, %f89;
mov.b32 %r512, %f90;
mov.u32 %r513, 4;
shfl.sync.bfly.b32 %r514|%p9, %r512, %r513, %r505, %r507;
mov.b32 %f91, %r514;
add.ftz.f32 %f92, %f90, %f91;
mov.b32 %r515, %f92;
shfl.sync.bfly.b32 %r516|%p10, %r515, %r503, %r505, %r507;
mov.b32 %f93, %r516;
add.ftz.f32 %f94, %f92, %f93;
mov.b32 %r517, %f94;
mov.u32 %r518, 1;
shfl.sync.bfly.b32 %r519|%p11, %r517, %r518, %r505, %r507;
mov.b32 %f95, %r519;
add.ftz.f32 %f25, %f94, %f95;
setp.lt.u32 %p12, %r4, %r29;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
mul.wide.u32 %rd54, %r500, 4;
add.s64 %rd14, %rd1, %rd54;
not.pred %p14, %p1;
@%p14 bra $L__BB299_8;

st.global.f32 [%rd14], %f25;

$L__BB299_8:
ld.shared.f32 %f96, [%r23+128];
add.ftz.f32 %f97, %f96, %f215;
ld.shared.f32 %f98, [%r23+1152];
add.ftz.f32 %f99, %f98, %f97;
ld.shared.f32 %f100, [%r23+2176];
add.ftz.f32 %f101, %f100, %f99;
mov.b32 %r520, %f101;
shfl.sync.bfly.b32 %r524|%p15, %r520, %r506, %r505, %r507;
mov.b32 %f102, %r524;
add.ftz.f32 %f103, %f101, %f102;
mov.b32 %r525, %f103;
shfl.sync.bfly.b32 %r527|%p16, %r525, %r510, %r505, %r507;
mov.b32 %f104, %r527;
add.ftz.f32 %f105, %f103, %f104;
mov.b32 %r528, %f105;
shfl.sync.bfly.b32 %r530|%p17, %r528, %r513, %r505, %r507;
mov.b32 %f106, %r530;
add.ftz.f32 %f107, %f105, %f106;
mov.b32 %r531, %f107;
shfl.sync.bfly.b32 %r533|%p18, %r531, %r503, %r505, %r507;
mov.b32 %f108, %r533;
add.ftz.f32 %f109, %f107, %f108;
mov.b32 %r534, %f109;
shfl.sync.bfly.b32 %r536|%p19, %r534, %r518, %r505, %r507;
mov.b32 %f110, %r536;
add.ftz.f32 %f26, %f109, %f110;
add.s32 %r537, %r4, 1;
setp.lt.u32 %p20, %r537, %r29;
setp.eq.s32 %p21, %r2, 1;
and.pred %p2, %p21, %p20;
not.pred %p22, %p2;
@%p22 bra $L__BB299_10;

st.global.f32 [%rd14+4], %f26;

$L__BB299_10:
ld.shared.f32 %f111, [%r23+256];
add.ftz.f32 %f112, %f111, %f214;
ld.shared.f32 %f113, [%r23+1280];
add.ftz.f32 %f114, %f113, %f112;
ld.shared.f32 %f115, [%r23+2304];
add.ftz.f32 %f116, %f115, %f114;
mov.b32 %r538, %f116;
mov.u32 %r539, 31;
mov.u32 %r540, 16;
mov.u32 %r541, -1;
shfl.sync.bfly.b32 %r542|%p23, %r538, %r540, %r539, %r541;
mov.b32 %f117, %r542;
add.ftz.f32 %f118, %f116, %f117;
mov.b32 %r543, %f118;
mov.u32 %r544, 8;
shfl.sync.bfly.b32 %r545|%p24, %r543, %r544, %r539, %r541;
mov.b32 %f119, %r545;
add.ftz.f32 %f120, %f118, %f119;
mov.b32 %r546, %f120;
mov.u32 %r547, 4;
shfl.sync.bfly.b32 %r548|%p25, %r546, %r547, %r539, %r541;
mov.b32 %f121, %r548;
add.ftz.f32 %f122, %f120, %f121;
mov.b32 %r549, %f122;
mov.u32 %r550, 2;
shfl.sync.bfly.b32 %r551|%p26, %r549, %r550, %r539, %r541;
mov.b32 %f123, %r551;
add.ftz.f32 %f124, %f122, %f123;
mov.b32 %r552, %f124;
mov.u32 %r553, 1;
shfl.sync.bfly.b32 %r554|%p27, %r552, %r553, %r539, %r541;
mov.b32 %f125, %r554;
add.ftz.f32 %f27, %f124, %f125;
@%p14 bra $L__BB299_12;

cvt.u64.u32 %rd55, %r29;
add.s64 %rd56, %rd55, %rd13;
shl.b64 %rd57, %rd56, 2;
add.s64 %rd58, %rd1, %rd57;
st.global.f32 [%rd58], %f27;

$L__BB299_12:
ld.shared.f32 %f126, [%r23+384];
add.ftz.f32 %f127, %f126, %f213;
ld.shared.f32 %f128, [%r23+1408];
add.ftz.f32 %f129, %f128, %f127;
ld.shared.f32 %f130, [%r23+2432];
add.ftz.f32 %f131, %f130, %f129;
mov.b32 %r555, %f131;
shfl.sync.bfly.b32 %r559|%p29, %r555, %r540, %r539, %r541;
mov.b32 %f132, %r559;
add.ftz.f32 %f133, %f131, %f132;
mov.b32 %r560, %f133;
shfl.sync.bfly.b32 %r562|%p30, %r560, %r544, %r539, %r541;
mov.b32 %f134, %r562;
add.ftz.f32 %f135, %f133, %f134;
mov.b32 %r563, %f135;
shfl.sync.bfly.b32 %r565|%p31, %r563, %r547, %r539, %r541;
mov.b32 %f136, %r565;
add.ftz.f32 %f137, %f135, %f136;
mov.b32 %r566, %f137;
shfl.sync.bfly.b32 %r568|%p32, %r566, %r550, %r539, %r541;
mov.b32 %f138, %r568;
add.ftz.f32 %f139, %f137, %f138;
mov.b32 %r569, %f139;
shfl.sync.bfly.b32 %r571|%p33, %r569, %r553, %r539, %r541;
mov.b32 %f140, %r571;
add.ftz.f32 %f28, %f139, %f140;
@%p22 bra $L__BB299_14;

add.s32 %r572, %r29, 1;
cvt.u64.u32 %rd59, %r572;
add.s64 %rd60, %rd59, %rd13;
shl.b64 %rd61, %rd60, 2;
add.s64 %rd62, %rd1, %rd61;
st.global.f32 [%rd62], %f28;

$L__BB299_14:
shl.b32 %r573, %r29, 1;
mov.u32 %r574, 1;
ld.shared.f32 %f141, [%r23+512];
add.ftz.f32 %f142, %f141, %f212;
ld.shared.f32 %f143, [%r23+1536];
add.ftz.f32 %f144, %f143, %f142;
ld.shared.f32 %f145, [%r23+2560];
add.ftz.f32 %f146, %f145, %f144;
mov.b32 %r575, %f146;
mov.u32 %r576, 31;
mov.u32 %r577, 16;
mov.u32 %r578, -1;
shfl.sync.bfly.b32 %r579|%p35, %r575, %r577, %r576, %r578;
mov.b32 %f147, %r579;
add.ftz.f32 %f148, %f146, %f147;
mov.b32 %r580, %f148;
mov.u32 %r581, 8;
shfl.sync.bfly.b32 %r582|%p36, %r580, %r581, %r576, %r578;
mov.b32 %f149, %r582;
add.ftz.f32 %f150, %f148, %f149;
mov.b32 %r583, %f150;
mov.u32 %r584, 4;
shfl.sync.bfly.b32 %r585|%p37, %r583, %r584, %r576, %r578;
mov.b32 %f151, %r585;
add.ftz.f32 %f152, %f150, %f151;
mov.b32 %r586, %f152;
mov.u32 %r587, 2;
shfl.sync.bfly.b32 %r588|%p38, %r586, %r587, %r576, %r578;
mov.b32 %f153, %r588;
add.ftz.f32 %f154, %f152, %f153;
mov.b32 %r589, %f154;
shfl.sync.bfly.b32 %r590|%p39, %r589, %r574, %r576, %r578;
mov.b32 %f155, %r590;
add.ftz.f32 %f29, %f154, %f155;
cvt.u64.u32 %rd63, %r573;
add.s64 %rd64, %rd63, %rd13;
shl.b64 %rd65, %rd64, 2;
add.s64 %rd15, %rd1, %rd65;
@%p14 bra $L__BB299_16;

st.global.f32 [%rd15], %f29;

$L__BB299_16:
ld.shared.f32 %f156, [%r23+640];
add.ftz.f32 %f157, %f156, %f211;
ld.shared.f32 %f158, [%r23+1664];
add.ftz.f32 %f159, %f158, %f157;
ld.shared.f32 %f160, [%r23+2688];
add.ftz.f32 %f161, %f160, %f159;
mov.b32 %r591, %f161;
shfl.sync.bfly.b32 %r595|%p41, %r591, %r577, %r576, %r578;
mov.b32 %f162, %r595;
add.ftz.f32 %f163, %f161, %f162;
mov.b32 %r596, %f163;
shfl.sync.bfly.b32 %r598|%p42, %r596, %r581, %r576, %r578;
mov.b32 %f164, %r598;
add.ftz.f32 %f165, %f163, %f164;
mov.b32 %r599, %f165;
shfl.sync.bfly.b32 %r601|%p43, %r599, %r584, %r576, %r578;
mov.b32 %f166, %r601;
add.ftz.f32 %f167, %f165, %f166;
mov.b32 %r602, %f167;
shfl.sync.bfly.b32 %r604|%p44, %r602, %r587, %r576, %r578;
mov.b32 %f168, %r604;
add.ftz.f32 %f169, %f167, %f168;
mov.b32 %r605, %f169;
shfl.sync.bfly.b32 %r607|%p45, %r605, %r574, %r576, %r578;
mov.b32 %f170, %r607;
add.ftz.f32 %f30, %f169, %f170;
@%p22 bra $L__BB299_18;

st.global.f32 [%rd15+4], %f30;

$L__BB299_18:
mul.lo.s32 %r25, %r29, 3;
ld.shared.f32 %f171, [%r23+768];
add.ftz.f32 %f172, %f171, %f210;
ld.shared.f32 %f173, [%r23+1792];
add.ftz.f32 %f174, %f173, %f172;
ld.shared.f32 %f175, [%r23+2816];
add.ftz.f32 %f176, %f175, %f174;
mov.b32 %r608, %f176;
mov.u32 %r609, 31;
mov.u32 %r610, 16;
mov.u32 %r611, -1;
shfl.sync.bfly.b32 %r612|%p47, %r608, %r610, %r609, %r611;
mov.b32 %f177, %r612;
add.ftz.f32 %f178, %f176, %f177;
mov.b32 %r613, %f178;
mov.u32 %r614, 8;
shfl.sync.bfly.b32 %r615|%p48, %r613, %r614, %r609, %r611;
mov.b32 %f179, %r615;
add.ftz.f32 %f180, %f178, %f179;
mov.b32 %r616, %f180;
mov.u32 %r617, 4;
shfl.sync.bfly.b32 %r618|%p49, %r616, %r617, %r609, %r611;
mov.b32 %f181, %r618;
add.ftz.f32 %f182, %f180, %f181;
mov.b32 %r619, %f182;
mov.u32 %r620, 2;
shfl.sync.bfly.b32 %r621|%p50, %r619, %r620, %r609, %r611;
mov.b32 %f183, %r621;
add.ftz.f32 %f184, %f182, %f183;
mov.b32 %r622, %f184;
mov.u32 %r623, 1;
shfl.sync.bfly.b32 %r624|%p51, %r622, %r623, %r609, %r611;
mov.b32 %f185, %r624;
add.ftz.f32 %f31, %f184, %f185;
@%p14 bra $L__BB299_20;

cvt.u64.u32 %rd66, %r25;
add.s64 %rd67, %rd66, %rd13;
shl.b64 %rd68, %rd67, 2;
add.s64 %rd69, %rd1, %rd68;
st.global.f32 [%rd69], %f31;

$L__BB299_20:
ld.shared.f32 %f186, [%r23+896];
add.ftz.f32 %f187, %f186, %f209;
ld.shared.f32 %f188, [%r23+1920];
add.ftz.f32 %f189, %f188, %f187;
ld.shared.f32 %f190, [%r23+2944];
add.ftz.f32 %f191, %f190, %f189;
mov.b32 %r625, %f191;
shfl.sync.bfly.b32 %r629|%p53, %r625, %r610, %r609, %r611;
mov.b32 %f192, %r629;
add.ftz.f32 %f193, %f191, %f192;
mov.b32 %r630, %f193;
shfl.sync.bfly.b32 %r632|%p54, %r630, %r614, %r609, %r611;
mov.b32 %f194, %r632;
add.ftz.f32 %f195, %f193, %f194;
mov.b32 %r633, %f195;
shfl.sync.bfly.b32 %r635|%p55, %r633, %r617, %r609, %r611;
mov.b32 %f196, %r635;
add.ftz.f32 %f197, %f195, %f196;
mov.b32 %r636, %f197;
shfl.sync.bfly.b32 %r638|%p56, %r636, %r620, %r609, %r611;
mov.b32 %f198, %r638;
add.ftz.f32 %f199, %f197, %f198;
mov.b32 %r639, %f199;
shfl.sync.bfly.b32 %r641|%p57, %r639, %r623, %r609, %r611;
mov.b32 %f200, %r641;
add.ftz.f32 %f32, %f199, %f200;
@%p22 bra $L__BB299_22;

add.s32 %r642, %r25, 1;
cvt.u64.u32 %rd70, %r642;
add.s64 %rd71, %rd70, %rd13;
shl.b64 %rd72, %rd71, 2;
add.s64 %rd73, %rd1, %rd72;
st.global.f32 [%rd73], %f32;

$L__BB299_22:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type23ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type23ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type23ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type23ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type23ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type23ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type23ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type23ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type23ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<71>;
.reg .b16 %rs<15>;
.reg .f32 %f<231>;
.reg .b32 %r<757>;
.reg .b64 %rd<84>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type23ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1280];

ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type23ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type23ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd19, [_Z13mul_mat_vec_qIL9ggml_type23ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type23ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type23ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type23ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type23ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type23ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type23ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type23ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type23ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type23ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type23ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type23ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type23ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type23ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type23ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
cvta.to.global.u64 %rd1, %rd19;
mov.u32 %r1, %tid.y;
shl.b32 %r44, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r44, %r2;
mov.u32 %r45, %ctaid.x;
shl.b32 %r4, %r45, 1;
mov.u32 %r5, %ctaid.y;
shr.s32 %r46, %r3, 31;
shr.u32 %r47, %r46, 29;
add.s32 %r48, %r3, %r47;
shr.s32 %r756, %r48, 3;
shr.u32 %r49, %r28, 8;
setp.ge.s32 %p3, %r756, %r49;
mov.f32 %f221, 0f00000000;
mov.f32 %f222, %f221;
mov.f32 %f223, %f221;
mov.f32 %f224, %f221;
mov.f32 %f225, %f221;
mov.f32 %f226, %f221;
mov.f32 %f227, %f221;
mov.f32 %f228, %f221;
mov.f32 %f229, %f221;
mov.f32 %f230, %f221;
@%p3 bra $L__BB300_3;

mov.u32 %r50, %ctaid.z;
mul.lo.s32 %r51, %r5, %r36;
cvt.u64.u32 %rd20, %r51;
mul.lo.s32 %r52, %r50, %r42;
cvt.u64.u32 %rd21, %r52;
add.s64 %rd2, %rd21, %rd20;
mul.hi.u32 %r53, %r50, %r38;
add.s32 %r54, %r53, %r50;
shr.u32 %r55, %r54, %r39;
mul.lo.s32 %r56, %r55, %r41;
mul.hi.u32 %r57, %r5, %r32;
add.s32 %r58, %r57, %r5;
shr.u32 %r59, %r58, %r33;
mul.lo.s32 %r60, %r59, %r35;
cvta.to.global.u64 %rd83, %rd17;
and.b32 %r64, %r48, -8;
sub.s32 %r7, %r3, %r64;
shl.b32 %r65, %r7, 2;
shr.u32 %r66, %r7, 31;
add.s32 %r67, %r7, %r66;
shr.s32 %r68, %r67, 1;
cvt.s64.s32 %rd22, %r68;
and.b32 %r8, %r65, 4;
shl.b32 %r69, %r30, 1;
shl.b32 %r751, %r756, 3;
add.s32 %r755, %r751, %r69;
mad.lo.s32 %r754, %r30, 3, %r751;
shl.b32 %r70, %r30, 2;
add.s32 %r753, %r751, %r70;
add.s32 %r752, %r30, %r751;
add.s32 %r71, %r756, %r56;
add.s32 %r72, %r71, %r60;
add.s32 %r73, %r4, 1;
mad.lo.s32 %r74, %r29, %r73, %r72;
mul.wide.s32 %rd4, %r74, 136;
mad.lo.s32 %r75, %r4, %r29, %r756;
add.s32 %r76, %r75, %r56;
add.s32 %r77, %r76, %r60;
mul.wide.s32 %rd5, %r77, 136;
add.s64 %rd23, %rd5, %rd22;
add.s64 %rd6, %rd23, 4;
add.s64 %rd24, %rd4, %rd22;
add.s64 %rd7, %rd24, 4;
mul.wide.s32 %rd25, %r65, 4;
add.s64 %rd26, %rd5, %rd25;
add.s64 %rd8, %rd26, 20;
add.s64 %rd27, %rd4, %rd25;
add.s64 %rd9, %rd27, 12;
cvta.to.global.u64 %rd10, %rd18;
mov.f32 %f221, 0f00000000;

$L__BB300_2:
cvt.u64.u32 %rd28, %r751;
add.s64 %rd29, %rd2, %rd28;
cvt.s64.s32 %rd30, %r7;
add.s64 %rd31, %rd29, %rd30;
mul.lo.s64 %rd32, %rd31, 36;
add.s64 %rd33, %rd83, %rd8;
ld.global.nc.u32 %r408, [%rd33+-12];
and.b32 %r409, %r408, -2004318072;
shr.u32 %r410, %r409, 1;
or.b32 %r411, %r410, 839922192;
and.b32 %r412, %r408, 30583;
mov.u32 %r413, -152379953;
mov.u32 %r414, -1079142271;
prmt.b32 %r415, %r414, %r413, %r412;
mov.u32 %r416, 1901675829;
mov.u32 %r417, 639175937;
prmt.b32 %r418, %r417, %r416, %r412;
and.b32 %r419, %r411, 30292;
prmt.b32 %r420, %r415, %r418, %r419;
shr.u32 %r421, %r408, 16;
and.b32 %r422, %r421, 30583;
prmt.b32 %r423, %r414, %r413, %r422;
prmt.b32 %r424, %r417, %r416, %r422;
shr.u32 %r425, %r411, 16;
prmt.b32 %r426, %r423, %r424, %r425;
mov.u32 %r427, 25632;
prmt.b32 %r343, %r420, %r426, %r427;
mov.u32 %r428, 30001;
prmt.b32 %r347, %r420, %r426, %r428;
add.s64 %rd34, %rd10, %rd32;
ld.global.nc.u32 %r80, [%rd34+4];
ld.global.nc.u32 %r84, [%rd34+20];
mov.u32 %r378, 0;

dp4a.s32.s32 %r78, %r343, %r80, %r378;


dp4a.s32.s32 %r82, %r347, %r84, %r78;

ld.global.nc.u32 %r429, [%rd33+-8];
and.b32 %r430, %r429, -2004318072;
shr.u32 %r431, %r430, 1;
or.b32 %r432, %r431, 839922192;
and.b32 %r433, %r429, 30583;
prmt.b32 %r434, %r414, %r413, %r433;
prmt.b32 %r435, %r417, %r416, %r433;
and.b32 %r436, %r432, 30292;
prmt.b32 %r437, %r434, %r435, %r436;
shr.u32 %r438, %r429, 16;
and.b32 %r439, %r438, 30583;
prmt.b32 %r440, %r414, %r413, %r439;
prmt.b32 %r441, %r417, %r416, %r439;
shr.u32 %r442, %r432, 16;
prmt.b32 %r443, %r440, %r441, %r442;
prmt.b32 %r351, %r437, %r443, %r427;
prmt.b32 %r355, %r437, %r443, %r428;
ld.global.nc.u32 %r88, [%rd34+8];
ld.global.nc.u32 %r92, [%rd34+24];

dp4a.s32.s32 %r86, %r351, %r88, %r82;


dp4a.s32.s32 %r90, %r355, %r92, %r86;

ld.global.nc.u32 %r444, [%rd33+-4];
and.b32 %r445, %r444, -2004318072;
shr.u32 %r446, %r445, 1;
or.b32 %r447, %r446, 839922192;
and.b32 %r448, %r444, 30583;
prmt.b32 %r449, %r414, %r413, %r448;
prmt.b32 %r450, %r417, %r416, %r448;
and.b32 %r451, %r447, 30292;
prmt.b32 %r452, %r449, %r450, %r451;
shr.u32 %r453, %r444, 16;
and.b32 %r454, %r453, 30583;
prmt.b32 %r455, %r414, %r413, %r454;
prmt.b32 %r456, %r417, %r416, %r454;
shr.u32 %r457, %r447, 16;
prmt.b32 %r458, %r455, %r456, %r457;
prmt.b32 %r359, %r452, %r458, %r427;
prmt.b32 %r363, %r452, %r458, %r428;
ld.global.nc.u32 %r96, [%rd34+12];
ld.global.nc.u32 %r100, [%rd34+28];

dp4a.s32.s32 %r94, %r359, %r96, %r90;


dp4a.s32.s32 %r98, %r363, %r100, %r94;

ld.global.nc.u32 %r459, [%rd33];
and.b32 %r460, %r459, -2004318072;
shr.u32 %r461, %r460, 1;
or.b32 %r462, %r461, 839922192;
and.b32 %r463, %r459, 30583;
prmt.b32 %r464, %r414, %r413, %r463;
prmt.b32 %r465, %r417, %r416, %r463;
and.b32 %r466, %r462, 30292;
prmt.b32 %r467, %r464, %r465, %r466;
shr.u32 %r468, %r459, 16;
and.b32 %r469, %r468, 30583;
prmt.b32 %r470, %r414, %r413, %r469;
prmt.b32 %r471, %r417, %r416, %r469;
shr.u32 %r472, %r462, 16;
prmt.b32 %r473, %r470, %r471, %r472;
prmt.b32 %r367, %r467, %r473, %r427;
prmt.b32 %r371, %r467, %r473, %r428;
ld.global.nc.u32 %r104, [%rd34+16];
ld.global.nc.u32 %r108, [%rd34+32];

dp4a.s32.s32 %r102, %r367, %r104, %r98;


dp4a.s32.s32 %r106, %r371, %r108, %r102;

add.s64 %rd35, %rd83, %rd6;
ld.global.nc.u8 %rs11, [%rd35];
cvt.u32.u16 %r474, %rs11;
and.b32 %r475, %r474, 255;
shr.u32 %r476, %r475, %r8;
and.b32 %r477, %r476, 15;
add.s64 %rd36, %rd83, %rd5;
ld.global.nc.u16 %rs12, [%rd36+2];
cvt.u32.u16 %r478, %rs12;
shl.b32 %r479, %r7, 1;
shr.u32 %r480, %r478, %r479;
and.b32 %r481, %r480, 3;
bfi.b32 %r482, %r481, %r477, 4, 2;
add.s32 %r483, %r482, -32;
mul.lo.s32 %r484, %r483, %r106;
ld.global.nc.u16 %rs1, [%rd36];

{ cvt.f32.f16 %f61, %rs1;}


ld.global.nc.u32 %r110, [%rd34];

{.reg .f16 low,high;
mov.b32 {low,high},%r110;
cvt.f32.f16 %f62, low;}


mul.ftz.f32 %f81, %f61, %f62;
cvt.rn.f32.s32 %f82, %r484;
fma.rn.ftz.f32 %f230, %f81, %f82, %f230;
add.s64 %rd37, %rd83, %rd9;
ld.global.nc.u32 %r485, [%rd37+-4];
and.b32 %r486, %r485, -2004318072;
shr.u32 %r487, %r486, 1;
or.b32 %r488, %r487, 839922192;
and.b32 %r489, %r485, 30583;
prmt.b32 %r490, %r414, %r413, %r489;
prmt.b32 %r491, %r417, %r416, %r489;
and.b32 %r492, %r488, 30292;
prmt.b32 %r493, %r490, %r491, %r492;
shr.u32 %r494, %r485, 16;
and.b32 %r495, %r494, 30583;
prmt.b32 %r496, %r414, %r413, %r495;
prmt.b32 %r497, %r417, %r416, %r495;
shr.u32 %r498, %r488, 16;
prmt.b32 %r499, %r496, %r497, %r498;
prmt.b32 %r376, %r493, %r499, %r427;
prmt.b32 %r380, %r493, %r499, %r428;

dp4a.s32.s32 %r111, %r376, %r80, %r378;


dp4a.s32.s32 %r115, %r380, %r84, %r111;

ld.global.nc.u32 %r500, [%rd37];
and.b32 %r501, %r500, -2004318072;
shr.u32 %r502, %r501, 1;
or.b32 %r503, %r502, 839922192;
and.b32 %r504, %r500, 30583;
prmt.b32 %r505, %r414, %r413, %r504;
prmt.b32 %r506, %r417, %r416, %r504;
and.b32 %r507, %r503, 30292;
prmt.b32 %r508, %r505, %r506, %r507;
shr.u32 %r509, %r500, 16;
and.b32 %r510, %r509, 30583;
prmt.b32 %r511, %r414, %r413, %r510;
prmt.b32 %r512, %r417, %r416, %r510;
shr.u32 %r513, %r503, 16;
prmt.b32 %r514, %r511, %r512, %r513;
prmt.b32 %r384, %r508, %r514, %r427;
prmt.b32 %r388, %r508, %r514, %r428;

dp4a.s32.s32 %r119, %r384, %r88, %r115;


dp4a.s32.s32 %r123, %r388, %r92, %r119;

ld.global.nc.u32 %r515, [%rd37+4];
and.b32 %r516, %r515, -2004318072;
shr.u32 %r517, %r516, 1;
or.b32 %r518, %r517, 839922192;
and.b32 %r519, %r515, 30583;
prmt.b32 %r520, %r414, %r413, %r519;
prmt.b32 %r521, %r417, %r416, %r519;
and.b32 %r522, %r518, 30292;
prmt.b32 %r523, %r520, %r521, %r522;
shr.u32 %r524, %r515, 16;
and.b32 %r525, %r524, 30583;
prmt.b32 %r526, %r414, %r413, %r525;
prmt.b32 %r527, %r417, %r416, %r525;
shr.u32 %r528, %r518, 16;
prmt.b32 %r529, %r526, %r527, %r528;
prmt.b32 %r392, %r523, %r529, %r427;
prmt.b32 %r396, %r523, %r529, %r428;

dp4a.s32.s32 %r127, %r392, %r96, %r123;


dp4a.s32.s32 %r131, %r396, %r100, %r127;

ld.global.nc.u32 %r530, [%rd37+8];
and.b32 %r531, %r530, -2004318072;
shr.u32 %r532, %r531, 1;
or.b32 %r533, %r532, 839922192;
and.b32 %r534, %r530, 30583;
prmt.b32 %r535, %r414, %r413, %r534;
prmt.b32 %r536, %r417, %r416, %r534;
and.b32 %r537, %r533, 30292;
prmt.b32 %r538, %r535, %r536, %r537;
shr.u32 %r539, %r530, 16;
and.b32 %r540, %r539, 30583;
prmt.b32 %r541, %r414, %r413, %r540;
prmt.b32 %r542, %r417, %r416, %r540;
shr.u32 %r543, %r533, 16;
prmt.b32 %r544, %r541, %r542, %r543;
prmt.b32 %r400, %r538, %r544, %r427;
prmt.b32 %r404, %r538, %r544, %r428;

dp4a.s32.s32 %r135, %r400, %r104, %r131;


dp4a.s32.s32 %r139, %r404, %r108, %r135;

add.s64 %rd38, %rd83, %rd7;
ld.global.nc.u8 %rs13, [%rd38];
cvt.u32.u16 %r545, %rs13;
and.b32 %r546, %r545, 255;
shr.u32 %r547, %r546, %r8;
and.b32 %r548, %r547, 15;
add.s64 %rd39, %rd83, %rd4;
ld.global.nc.u16 %rs14, [%rd39+2];
cvt.u32.u16 %r549, %rs14;
shr.u32 %r550, %r549, %r479;
and.b32 %r551, %r550, 3;
bfi.b32 %r552, %r551, %r548, 4, 2;
add.s32 %r553, %r552, -32;
mul.lo.s32 %r554, %r553, %r139;
ld.global.nc.u16 %rs2, [%rd39];

{ cvt.f32.f16 %f63, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r110;
cvt.f32.f16 %f64, low;}


mul.ftz.f32 %f83, %f63, %f64;
cvt.rn.f32.s32 %f84, %r554;
fma.rn.ftz.f32 %f229, %f83, %f84, %f229;
cvt.u64.u32 %rd40, %r752;
add.s64 %rd41, %rd2, %rd40;
add.s64 %rd42, %rd41, %rd30;
mul.lo.s64 %rd43, %rd42, 36;
add.s64 %rd44, %rd10, %rd43;
ld.global.nc.u32 %r146, [%rd44+4];
ld.global.nc.u32 %r150, [%rd44+20];

dp4a.s32.s32 %r144, %r343, %r146, %r378;


dp4a.s32.s32 %r148, %r347, %r150, %r144;

ld.global.nc.u32 %r154, [%rd44+8];
ld.global.nc.u32 %r158, [%rd44+24];

dp4a.s32.s32 %r152, %r351, %r154, %r148;


dp4a.s32.s32 %r156, %r355, %r158, %r152;

ld.global.nc.u32 %r162, [%rd44+12];
ld.global.nc.u32 %r166, [%rd44+28];

dp4a.s32.s32 %r160, %r359, %r162, %r156;


dp4a.s32.s32 %r164, %r363, %r166, %r160;

ld.global.nc.u32 %r170, [%rd44+16];
ld.global.nc.u32 %r174, [%rd44+32];

dp4a.s32.s32 %r168, %r367, %r170, %r164;


dp4a.s32.s32 %r172, %r371, %r174, %r168;

mul.lo.s32 %r555, %r483, %r172;

{ cvt.f32.f16 %f65, %rs1;}


ld.global.nc.u32 %r176, [%rd44];

{.reg .f16 low,high;
mov.b32 {low,high},%r176;
cvt.f32.f16 %f66, low;}


mul.ftz.f32 %f85, %f65, %f66;
cvt.rn.f32.s32 %f86, %r555;
fma.rn.ftz.f32 %f228, %f85, %f86, %f228;

dp4a.s32.s32 %r177, %r376, %r146, %r378;


dp4a.s32.s32 %r181, %r380, %r150, %r177;


dp4a.s32.s32 %r185, %r384, %r154, %r181;


dp4a.s32.s32 %r189, %r388, %r158, %r185;


dp4a.s32.s32 %r193, %r392, %r162, %r189;


dp4a.s32.s32 %r197, %r396, %r166, %r193;


dp4a.s32.s32 %r201, %r400, %r170, %r197;


dp4a.s32.s32 %r205, %r404, %r174, %r201;

mul.lo.s32 %r556, %r553, %r205;

{ cvt.f32.f16 %f67, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r176;
cvt.f32.f16 %f68, low;}


mul.ftz.f32 %f87, %f67, %f68;
cvt.rn.f32.s32 %f88, %r556;
fma.rn.ftz.f32 %f227, %f87, %f88, %f227;
cvt.u64.u32 %rd45, %r755;
add.s64 %rd46, %rd2, %rd45;
add.s64 %rd47, %rd46, %rd30;
mul.lo.s64 %rd48, %rd47, 36;
add.s64 %rd49, %rd10, %rd48;
ld.global.nc.u32 %r212, [%rd49+4];
ld.global.nc.u32 %r216, [%rd49+20];

dp4a.s32.s32 %r210, %r343, %r212, %r378;


dp4a.s32.s32 %r214, %r347, %r216, %r210;

ld.global.nc.u32 %r220, [%rd49+8];
ld.global.nc.u32 %r224, [%rd49+24];

dp4a.s32.s32 %r218, %r351, %r220, %r214;


dp4a.s32.s32 %r222, %r355, %r224, %r218;

ld.global.nc.u32 %r228, [%rd49+12];
ld.global.nc.u32 %r232, [%rd49+28];

dp4a.s32.s32 %r226, %r359, %r228, %r222;


dp4a.s32.s32 %r230, %r363, %r232, %r226;

ld.global.nc.u32 %r236, [%rd49+16];
ld.global.nc.u32 %r240, [%rd49+32];

dp4a.s32.s32 %r234, %r367, %r236, %r230;


dp4a.s32.s32 %r238, %r371, %r240, %r234;

mul.lo.s32 %r557, %r483, %r238;

{ cvt.f32.f16 %f69, %rs1;}


ld.global.nc.u32 %r242, [%rd49];

{.reg .f16 low,high;
mov.b32 {low,high},%r242;
cvt.f32.f16 %f70, low;}


mul.ftz.f32 %f89, %f69, %f70;
cvt.rn.f32.s32 %f90, %r557;
fma.rn.ftz.f32 %f226, %f89, %f90, %f226;

dp4a.s32.s32 %r243, %r376, %r212, %r378;


dp4a.s32.s32 %r247, %r380, %r216, %r243;


dp4a.s32.s32 %r251, %r384, %r220, %r247;


dp4a.s32.s32 %r255, %r388, %r224, %r251;


dp4a.s32.s32 %r259, %r392, %r228, %r255;


dp4a.s32.s32 %r263, %r396, %r232, %r259;


dp4a.s32.s32 %r267, %r400, %r236, %r263;


dp4a.s32.s32 %r271, %r404, %r240, %r267;

mul.lo.s32 %r558, %r553, %r271;

{ cvt.f32.f16 %f71, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r242;
cvt.f32.f16 %f72, low;}


mul.ftz.f32 %f91, %f71, %f72;
cvt.rn.f32.s32 %f92, %r558;
fma.rn.ftz.f32 %f225, %f91, %f92, %f225;
cvt.u64.u32 %rd50, %r754;
add.s64 %rd51, %rd2, %rd50;
add.s64 %rd52, %rd51, %rd30;
mul.lo.s64 %rd53, %rd52, 36;
add.s64 %rd54, %rd10, %rd53;
ld.global.nc.u32 %r278, [%rd54+4];
ld.global.nc.u32 %r282, [%rd54+20];

dp4a.s32.s32 %r276, %r343, %r278, %r378;


dp4a.s32.s32 %r280, %r347, %r282, %r276;

ld.global.nc.u32 %r286, [%rd54+8];
ld.global.nc.u32 %r290, [%rd54+24];

dp4a.s32.s32 %r284, %r351, %r286, %r280;


dp4a.s32.s32 %r288, %r355, %r290, %r284;

ld.global.nc.u32 %r294, [%rd54+12];
ld.global.nc.u32 %r298, [%rd54+28];

dp4a.s32.s32 %r292, %r359, %r294, %r288;


dp4a.s32.s32 %r296, %r363, %r298, %r292;

ld.global.nc.u32 %r302, [%rd54+16];
ld.global.nc.u32 %r306, [%rd54+32];

dp4a.s32.s32 %r300, %r367, %r302, %r296;


dp4a.s32.s32 %r304, %r371, %r306, %r300;

mul.lo.s32 %r559, %r483, %r304;

{ cvt.f32.f16 %f73, %rs1;}


ld.global.nc.u32 %r308, [%rd54];

{.reg .f16 low,high;
mov.b32 {low,high},%r308;
cvt.f32.f16 %f74, low;}


mul.ftz.f32 %f93, %f73, %f74;
cvt.rn.f32.s32 %f94, %r559;
fma.rn.ftz.f32 %f224, %f93, %f94, %f224;

dp4a.s32.s32 %r309, %r376, %r278, %r378;


dp4a.s32.s32 %r313, %r380, %r282, %r309;


dp4a.s32.s32 %r317, %r384, %r286, %r313;


dp4a.s32.s32 %r321, %r388, %r290, %r317;


dp4a.s32.s32 %r325, %r392, %r294, %r321;


dp4a.s32.s32 %r329, %r396, %r298, %r325;


dp4a.s32.s32 %r333, %r400, %r302, %r329;


dp4a.s32.s32 %r337, %r404, %r306, %r333;

mul.lo.s32 %r560, %r553, %r337;

{ cvt.f32.f16 %f75, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r308;
cvt.f32.f16 %f76, low;}


mul.ftz.f32 %f95, %f75, %f76;
cvt.rn.f32.s32 %f96, %r560;
fma.rn.ftz.f32 %f223, %f95, %f96, %f223;
cvt.u64.u32 %rd55, %r753;
add.s64 %rd56, %rd2, %rd55;
add.s64 %rd57, %rd56, %rd30;
mul.lo.s64 %rd58, %rd57, 36;
add.s64 %rd59, %rd10, %rd58;
ld.global.nc.u32 %r344, [%rd59+4];
ld.global.nc.u32 %r348, [%rd59+20];

dp4a.s32.s32 %r342, %r343, %r344, %r378;


dp4a.s32.s32 %r346, %r347, %r348, %r342;

ld.global.nc.u32 %r352, [%rd59+8];
ld.global.nc.u32 %r356, [%rd59+24];

dp4a.s32.s32 %r350, %r351, %r352, %r346;


dp4a.s32.s32 %r354, %r355, %r356, %r350;

ld.global.nc.u32 %r360, [%rd59+12];
ld.global.nc.u32 %r364, [%rd59+28];

dp4a.s32.s32 %r358, %r359, %r360, %r354;


dp4a.s32.s32 %r362, %r363, %r364, %r358;

ld.global.nc.u32 %r368, [%rd59+16];
ld.global.nc.u32 %r372, [%rd59+32];

dp4a.s32.s32 %r366, %r367, %r368, %r362;


dp4a.s32.s32 %r370, %r371, %r372, %r366;

mul.lo.s32 %r561, %r483, %r370;

{ cvt.f32.f16 %f77, %rs1;}


ld.global.nc.u32 %r374, [%rd59];

{.reg .f16 low,high;
mov.b32 {low,high},%r374;
cvt.f32.f16 %f78, low;}


mul.ftz.f32 %f97, %f77, %f78;
cvt.rn.f32.s32 %f98, %r561;
fma.rn.ftz.f32 %f222, %f97, %f98, %f222;

dp4a.s32.s32 %r375, %r376, %r344, %r378;


dp4a.s32.s32 %r379, %r380, %r348, %r375;


dp4a.s32.s32 %r383, %r384, %r352, %r379;


dp4a.s32.s32 %r387, %r388, %r356, %r383;


dp4a.s32.s32 %r391, %r392, %r360, %r387;


dp4a.s32.s32 %r395, %r396, %r364, %r391;


dp4a.s32.s32 %r399, %r400, %r368, %r395;


dp4a.s32.s32 %r403, %r404, %r372, %r399;

mul.lo.s32 %r562, %r553, %r403;

{ cvt.f32.f16 %f79, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r374;
cvt.f32.f16 %f80, low;}


mul.ftz.f32 %f99, %f79, %f80;
cvt.rn.f32.s32 %f100, %r562;
fma.rn.ftz.f32 %f221, %f99, %f100, %f221;
add.s32 %r755, %r755, 64;
add.s32 %r754, %r754, 64;
add.s32 %r753, %r753, 64;
add.s32 %r752, %r752, 64;
add.s32 %r751, %r751, 64;
add.s64 %rd83, %rd83, 1088;
add.s32 %r756, %r756, 8;
setp.lt.s32 %p4, %r756, %r49;
@%p4 bra $L__BB300_2;

$L__BB300_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB300_5;

mov.u32 %r564, _ZZ13mul_mat_vec_qIL9ggml_type23ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r565, %r1, 1280, %r564;
shl.b32 %r566, %r2, 2;
add.s32 %r567, %r565, %r566;
st.shared.f32 [%r567+-1280], %f230;
st.shared.f32 [%r567+-1152], %f229;
st.shared.f32 [%r567+-1024], %f228;
st.shared.f32 [%r567+-896], %f227;
st.shared.f32 [%r567+-768], %f226;
st.shared.f32 [%r567+-640], %f225;
st.shared.f32 [%r567+-512], %f224;
st.shared.f32 [%r567+-384], %f223;
st.shared.f32 [%r567+-256], %f222;
st.shared.f32 [%r567+-128], %f221;

$L__BB300_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB300_26;

mov.u32 %r568, %ctaid.z;
mad.lo.s32 %r569, %r5, %r37, %r4;
mad.lo.s32 %r570, %r568, %r43, %r569;
cvt.u64.u32 %rd13, %r570;
shl.b32 %r571, %r2, 2;
mov.u32 %r572, _ZZ13mul_mat_vec_qIL9ggml_type23ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r26, %r572, %r571;
mov.u32 %r573, 2;
ld.shared.f32 %f101, [%r26];
add.ftz.f32 %f102, %f101, %f230;
mov.b32 %r574, %f102;
mov.u32 %r575, 31;
mov.u32 %r576, 16;
mov.u32 %r577, -1;
shfl.sync.bfly.b32 %r578|%p7, %r574, %r576, %r575, %r577;
mov.b32 %f103, %r578;
add.ftz.f32 %f104, %f102, %f103;
mov.b32 %r579, %f104;
mov.u32 %r580, 8;
shfl.sync.bfly.b32 %r581|%p8, %r579, %r580, %r575, %r577;
mov.b32 %f105, %r581;
add.ftz.f32 %f106, %f104, %f105;
mov.b32 %r582, %f106;
mov.u32 %r583, 4;
shfl.sync.bfly.b32 %r584|%p9, %r582, %r583, %r575, %r577;
mov.b32 %f107, %r584;
add.ftz.f32 %f108, %f106, %f107;
mov.b32 %r585, %f108;
shfl.sync.bfly.b32 %r586|%p10, %r585, %r573, %r575, %r577;
mov.b32 %f109, %r586;
add.ftz.f32 %f110, %f108, %f109;
mov.b32 %r587, %f110;
mov.u32 %r588, 1;
shfl.sync.bfly.b32 %r589|%p11, %r587, %r588, %r575, %r577;
mov.b32 %f111, %r589;
add.ftz.f32 %f31, %f110, %f111;
setp.lt.u32 %p12, %r4, %r31;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
mul.wide.u32 %rd60, %r570, 4;
add.s64 %rd14, %rd1, %rd60;
not.pred %p14, %p1;
@%p14 bra $L__BB300_8;

st.global.f32 [%rd14], %f31;

$L__BB300_8:
ld.shared.f32 %f112, [%r26+128];
add.ftz.f32 %f113, %f112, %f229;
mov.b32 %r590, %f113;
shfl.sync.bfly.b32 %r594|%p15, %r590, %r576, %r575, %r577;
mov.b32 %f114, %r594;
add.ftz.f32 %f115, %f113, %f114;
mov.b32 %r595, %f115;
shfl.sync.bfly.b32 %r597|%p16, %r595, %r580, %r575, %r577;
mov.b32 %f116, %r597;
add.ftz.f32 %f117, %f115, %f116;
mov.b32 %r598, %f117;
shfl.sync.bfly.b32 %r600|%p17, %r598, %r583, %r575, %r577;
mov.b32 %f118, %r600;
add.ftz.f32 %f119, %f117, %f118;
mov.b32 %r601, %f119;
shfl.sync.bfly.b32 %r603|%p18, %r601, %r573, %r575, %r577;
mov.b32 %f120, %r603;
add.ftz.f32 %f121, %f119, %f120;
mov.b32 %r604, %f121;
shfl.sync.bfly.b32 %r606|%p19, %r604, %r588, %r575, %r577;
mov.b32 %f122, %r606;
add.ftz.f32 %f32, %f121, %f122;
setp.eq.s32 %p20, %r2, 1;
or.b32 %r610, %r4, 1;
setp.lt.u32 %p21, %r610, %r31;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB300_10;

st.global.f32 [%rd14+4], %f32;

$L__BB300_10:
ld.shared.f32 %f123, [%r26+256];
add.ftz.f32 %f124, %f123, %f228;
mov.b32 %r611, %f124;
mov.u32 %r612, 31;
mov.u32 %r613, 16;
mov.u32 %r614, -1;
shfl.sync.bfly.b32 %r615|%p23, %r611, %r613, %r612, %r614;
mov.b32 %f125, %r615;
add.ftz.f32 %f126, %f124, %f125;
mov.b32 %r616, %f126;
mov.u32 %r617, 8;
shfl.sync.bfly.b32 %r618|%p24, %r616, %r617, %r612, %r614;
mov.b32 %f127, %r618;
add.ftz.f32 %f128, %f126, %f127;
mov.b32 %r619, %f128;
mov.u32 %r620, 4;
shfl.sync.bfly.b32 %r621|%p25, %r619, %r620, %r612, %r614;
mov.b32 %f129, %r621;
add.ftz.f32 %f130, %f128, %f129;
mov.b32 %r622, %f130;
mov.u32 %r623, 2;
shfl.sync.bfly.b32 %r624|%p26, %r622, %r623, %r612, %r614;
mov.b32 %f131, %r624;
add.ftz.f32 %f132, %f130, %f131;
mov.b32 %r625, %f132;
mov.u32 %r626, 1;
shfl.sync.bfly.b32 %r627|%p27, %r625, %r626, %r612, %r614;
mov.b32 %f133, %r627;
add.ftz.f32 %f33, %f132, %f133;
@%p14 bra $L__BB300_12;

cvt.u64.u32 %rd61, %r31;
add.s64 %rd62, %rd61, %rd13;
shl.b64 %rd63, %rd62, 2;
add.s64 %rd64, %rd1, %rd63;
st.global.f32 [%rd64], %f33;

$L__BB300_12:
ld.shared.f32 %f134, [%r26+384];
add.ftz.f32 %f135, %f134, %f227;
mov.b32 %r628, %f135;
shfl.sync.bfly.b32 %r632|%p29, %r628, %r613, %r612, %r614;
mov.b32 %f136, %r632;
add.ftz.f32 %f137, %f135, %f136;
mov.b32 %r633, %f137;
shfl.sync.bfly.b32 %r635|%p30, %r633, %r617, %r612, %r614;
mov.b32 %f138, %r635;
add.ftz.f32 %f139, %f137, %f138;
mov.b32 %r636, %f139;
shfl.sync.bfly.b32 %r638|%p31, %r636, %r620, %r612, %r614;
mov.b32 %f140, %r638;
add.ftz.f32 %f141, %f139, %f140;
mov.b32 %r639, %f141;
shfl.sync.bfly.b32 %r641|%p32, %r639, %r623, %r612, %r614;
mov.b32 %f142, %r641;
add.ftz.f32 %f143, %f141, %f142;
mov.b32 %r642, %f143;
shfl.sync.bfly.b32 %r644|%p33, %r642, %r626, %r612, %r614;
mov.b32 %f144, %r644;
add.ftz.f32 %f34, %f143, %f144;
@%p22 bra $L__BB300_14;

add.s32 %r645, %r31, 1;
cvt.u64.u32 %rd65, %r645;
add.s64 %rd66, %rd65, %rd13;
shl.b64 %rd67, %rd66, 2;
add.s64 %rd68, %rd1, %rd67;
st.global.f32 [%rd68], %f34;

$L__BB300_14:
shl.b32 %r646, %r31, 1;
mov.u32 %r647, 1;
ld.shared.f32 %f145, [%r26+512];
add.ftz.f32 %f146, %f145, %f226;
mov.b32 %r648, %f146;
mov.u32 %r649, 31;
mov.u32 %r650, 16;
mov.u32 %r651, -1;
shfl.sync.bfly.b32 %r652|%p35, %r648, %r650, %r649, %r651;
mov.b32 %f147, %r652;
add.ftz.f32 %f148, %f146, %f147;
mov.b32 %r653, %f148;
mov.u32 %r654, 8;
shfl.sync.bfly.b32 %r655|%p36, %r653, %r654, %r649, %r651;
mov.b32 %f149, %r655;
add.ftz.f32 %f150, %f148, %f149;
mov.b32 %r656, %f150;
mov.u32 %r657, 4;
shfl.sync.bfly.b32 %r658|%p37, %r656, %r657, %r649, %r651;
mov.b32 %f151, %r658;
add.ftz.f32 %f152, %f150, %f151;
mov.b32 %r659, %f152;
mov.u32 %r660, 2;
shfl.sync.bfly.b32 %r661|%p38, %r659, %r660, %r649, %r651;
mov.b32 %f153, %r661;
add.ftz.f32 %f154, %f152, %f153;
mov.b32 %r662, %f154;
shfl.sync.bfly.b32 %r663|%p39, %r662, %r647, %r649, %r651;
mov.b32 %f155, %r663;
add.ftz.f32 %f35, %f154, %f155;
cvt.u64.u32 %rd69, %r646;
add.s64 %rd70, %rd69, %rd13;
shl.b64 %rd71, %rd70, 2;
add.s64 %rd15, %rd1, %rd71;
@%p14 bra $L__BB300_16;

st.global.f32 [%rd15], %f35;

$L__BB300_16:
ld.shared.f32 %f156, [%r26+640];
add.ftz.f32 %f157, %f156, %f225;
mov.b32 %r664, %f157;
shfl.sync.bfly.b32 %r668|%p41, %r664, %r650, %r649, %r651;
mov.b32 %f158, %r668;
add.ftz.f32 %f159, %f157, %f158;
mov.b32 %r669, %f159;
shfl.sync.bfly.b32 %r671|%p42, %r669, %r654, %r649, %r651;
mov.b32 %f160, %r671;
add.ftz.f32 %f161, %f159, %f160;
mov.b32 %r672, %f161;
shfl.sync.bfly.b32 %r674|%p43, %r672, %r657, %r649, %r651;
mov.b32 %f162, %r674;
add.ftz.f32 %f163, %f161, %f162;
mov.b32 %r675, %f163;
shfl.sync.bfly.b32 %r677|%p44, %r675, %r660, %r649, %r651;
mov.b32 %f164, %r677;
add.ftz.f32 %f165, %f163, %f164;
mov.b32 %r678, %f165;
shfl.sync.bfly.b32 %r680|%p45, %r678, %r647, %r649, %r651;
mov.b32 %f166, %r680;
add.ftz.f32 %f36, %f165, %f166;
@%p22 bra $L__BB300_18;

st.global.f32 [%rd15+4], %f36;

$L__BB300_18:
mul.lo.s32 %r27, %r31, 3;
ld.shared.f32 %f167, [%r26+768];
add.ftz.f32 %f168, %f167, %f224;
mov.b32 %r681, %f168;
mov.u32 %r682, 31;
mov.u32 %r683, 16;
mov.u32 %r684, -1;
shfl.sync.bfly.b32 %r685|%p47, %r681, %r683, %r682, %r684;
mov.b32 %f169, %r685;
add.ftz.f32 %f170, %f168, %f169;
mov.b32 %r686, %f170;
mov.u32 %r687, 8;
shfl.sync.bfly.b32 %r688|%p48, %r686, %r687, %r682, %r684;
mov.b32 %f171, %r688;
add.ftz.f32 %f172, %f170, %f171;
mov.b32 %r689, %f172;
mov.u32 %r690, 4;
shfl.sync.bfly.b32 %r691|%p49, %r689, %r690, %r682, %r684;
mov.b32 %f173, %r691;
add.ftz.f32 %f174, %f172, %f173;
mov.b32 %r692, %f174;
mov.u32 %r693, 2;
shfl.sync.bfly.b32 %r694|%p50, %r692, %r693, %r682, %r684;
mov.b32 %f175, %r694;
add.ftz.f32 %f176, %f174, %f175;
mov.b32 %r695, %f176;
mov.u32 %r696, 1;
shfl.sync.bfly.b32 %r697|%p51, %r695, %r696, %r682, %r684;
mov.b32 %f177, %r697;
add.ftz.f32 %f37, %f176, %f177;
@%p14 bra $L__BB300_20;

cvt.u64.u32 %rd72, %r27;
add.s64 %rd73, %rd72, %rd13;
shl.b64 %rd74, %rd73, 2;
add.s64 %rd75, %rd1, %rd74;
st.global.f32 [%rd75], %f37;

$L__BB300_20:
ld.shared.f32 %f178, [%r26+896];
add.ftz.f32 %f179, %f178, %f223;
mov.b32 %r698, %f179;
shfl.sync.bfly.b32 %r702|%p53, %r698, %r683, %r682, %r684;
mov.b32 %f180, %r702;
add.ftz.f32 %f181, %f179, %f180;
mov.b32 %r703, %f181;
shfl.sync.bfly.b32 %r705|%p54, %r703, %r687, %r682, %r684;
mov.b32 %f182, %r705;
add.ftz.f32 %f183, %f181, %f182;
mov.b32 %r706, %f183;
shfl.sync.bfly.b32 %r708|%p55, %r706, %r690, %r682, %r684;
mov.b32 %f184, %r708;
add.ftz.f32 %f185, %f183, %f184;
mov.b32 %r709, %f185;
shfl.sync.bfly.b32 %r711|%p56, %r709, %r693, %r682, %r684;
mov.b32 %f186, %r711;
add.ftz.f32 %f187, %f185, %f186;
mov.b32 %r712, %f187;
shfl.sync.bfly.b32 %r714|%p57, %r712, %r696, %r682, %r684;
mov.b32 %f188, %r714;
add.ftz.f32 %f38, %f187, %f188;
@%p22 bra $L__BB300_22;

add.s32 %r715, %r27, 1;
cvt.u64.u32 %rd76, %r715;
add.s64 %rd77, %rd76, %rd13;
shl.b64 %rd78, %rd77, 2;
add.s64 %rd79, %rd1, %rd78;
st.global.f32 [%rd79], %f38;

$L__BB300_22:
shl.b32 %r716, %r31, 2;
mov.u32 %r717, 2;
ld.shared.f32 %f189, [%r26+1024];
add.ftz.f32 %f190, %f189, %f222;
mov.b32 %r718, %f190;
mov.u32 %r719, 31;
mov.u32 %r720, 16;
mov.u32 %r721, -1;
shfl.sync.bfly.b32 %r722|%p59, %r718, %r720, %r719, %r721;
mov.b32 %f191, %r722;
add.ftz.f32 %f192, %f190, %f191;
mov.b32 %r723, %f192;
mov.u32 %r724, 8;
shfl.sync.bfly.b32 %r725|%p60, %r723, %r724, %r719, %r721;
mov.b32 %f193, %r725;
add.ftz.f32 %f194, %f192, %f193;
mov.b32 %r726, %f194;
mov.u32 %r727, 4;
shfl.sync.bfly.b32 %r728|%p61, %r726, %r727, %r719, %r721;
mov.b32 %f195, %r728;
add.ftz.f32 %f196, %f194, %f195;
mov.b32 %r729, %f196;
shfl.sync.bfly.b32 %r730|%p62, %r729, %r717, %r719, %r721;
mov.b32 %f197, %r730;
add.ftz.f32 %f198, %f196, %f197;
mov.b32 %r731, %f198;
mov.u32 %r732, 1;
shfl.sync.bfly.b32 %r733|%p63, %r731, %r732, %r719, %r721;
mov.b32 %f199, %r733;
add.ftz.f32 %f39, %f198, %f199;
cvt.u64.u32 %rd80, %r716;
add.s64 %rd81, %rd80, %rd13;
shl.b64 %rd82, %rd81, 2;
add.s64 %rd16, %rd1, %rd82;
@%p14 bra $L__BB300_24;

st.global.f32 [%rd16], %f39;

$L__BB300_24:
ld.shared.f32 %f200, [%r26+1152];
add.ftz.f32 %f201, %f200, %f221;
mov.b32 %r734, %f201;
shfl.sync.bfly.b32 %r738|%p65, %r734, %r720, %r719, %r721;
mov.b32 %f202, %r738;
add.ftz.f32 %f203, %f201, %f202;
mov.b32 %r739, %f203;
shfl.sync.bfly.b32 %r741|%p66, %r739, %r724, %r719, %r721;
mov.b32 %f204, %r741;
add.ftz.f32 %f205, %f203, %f204;
mov.b32 %r742, %f205;
shfl.sync.bfly.b32 %r744|%p67, %r742, %r727, %r719, %r721;
mov.b32 %f206, %r744;
add.ftz.f32 %f207, %f205, %f206;
mov.b32 %r745, %f207;
shfl.sync.bfly.b32 %r747|%p68, %r745, %r717, %r719, %r721;
mov.b32 %f208, %r747;
add.ftz.f32 %f209, %f207, %f208;
mov.b32 %r748, %f209;
shfl.sync.bfly.b32 %r750|%p69, %r748, %r732, %r719, %r721;
mov.b32 %f210, %r750;
add.ftz.f32 %f40, %f209, %f210;
@%p22 bra $L__BB300_26;

st.global.f32 [%rd16+4], %f40;

$L__BB300_26:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type23ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type23ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type23ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type23ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type23ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type23ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type23ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type23ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type23ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<83>;
.reg .b16 %rs<17>;
.reg .f32 %f<277>;
.reg .b32 %r<877>;
.reg .b64 %rd<108>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type23ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type23ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type23ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type23ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r47, [_Z13mul_mat_vec_qIL9ggml_type23ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type23ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type23ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type23ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type23ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type23ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type23ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type23ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r45, [_Z13mul_mat_vec_qIL9ggml_type23ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r46, [_Z13mul_mat_vec_qIL9ggml_type23ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type23ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type23ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type23ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type23ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r48, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r48, %r2;
mov.u32 %r49, %ctaid.x;
shl.b32 %r4, %r49, 1;
shr.s32 %r50, %r3, 31;
shr.u32 %r51, %r50, 29;
add.s32 %r52, %r3, %r51;
shr.s32 %r876, %r52, 3;
shr.u32 %r6, %r47, 8;
setp.ge.s32 %p3, %r876, %r6;
mov.f32 %f265, 0f00000000;
mov.f32 %f266, %f265;
mov.f32 %f267, %f265;
mov.f32 %f268, %f265;
mov.f32 %f269, %f265;
mov.f32 %f270, %f265;
mov.f32 %f271, %f265;
mov.f32 %f272, %f265;
mov.f32 %f273, %f265;
mov.f32 %f274, %f265;
mov.f32 %f275, %f265;
mov.f32 %f276, %f265;
@%p3 bra $L__BB301_3;

mov.u32 %r53, %ctaid.y;
mov.u32 %r54, %ctaid.z;
mul.lo.s32 %r55, %r53, %r39;
cvt.u64.u32 %rd19, %r55;
mul.lo.s32 %r56, %r54, %r45;
cvt.u64.u32 %rd20, %r56;
add.s64 %rd1, %rd20, %rd19;
mul.hi.u32 %r57, %r54, %r41;
add.s32 %r58, %r57, %r54;
shr.u32 %r59, %r58, %r42;
mul.lo.s32 %r60, %r59, %r44;
mul.hi.u32 %r61, %r53, %r35;
add.s32 %r62, %r61, %r53;
shr.u32 %r63, %r62, %r36;
mul.lo.s32 %r64, %r63, %r38;
cvta.to.global.u64 %rd107, %rd16;
and.b32 %r68, %r52, -8;
sub.s32 %r69, %r3, %r68;
shl.b32 %r70, %r69, 2;
shr.u32 %r71, %r69, 31;
add.s32 %r72, %r69, %r71;
shr.s32 %r73, %r72, 1;
cvt.s64.s32 %rd21, %r73;
cvt.s64.s32 %rd3, %r69;
and.b32 %r7, %r70, 4;
shl.b32 %r8, %r69, 1;
shl.b32 %r74, %r33, 1;
shl.b32 %r870, %r876, 3;
add.s32 %r875, %r870, %r74;
mad.lo.s32 %r874, %r33, 3, %r870;
shl.b32 %r75, %r33, 2;
add.s32 %r873, %r870, %r75;
mad.lo.s32 %r872, %r33, 5, %r870;
add.s32 %r871, %r33, %r870;
add.s32 %r76, %r876, %r60;
add.s32 %r77, %r76, %r64;
add.s32 %r78, %r4, 1;
mad.lo.s32 %r79, %r32, %r78, %r77;
mul.wide.s32 %rd4, %r79, 136;
mad.lo.s32 %r80, %r4, %r32, %r876;
add.s32 %r81, %r80, %r60;
add.s32 %r82, %r81, %r64;
mul.wide.s32 %rd5, %r82, 136;
add.s64 %rd22, %rd5, %rd21;
add.s64 %rd6, %rd22, 4;
add.s64 %rd23, %rd4, %rd21;
add.s64 %rd7, %rd23, 4;
mul.wide.s32 %rd24, %r70, 4;
add.s64 %rd25, %rd5, %rd24;
add.s64 %rd8, %rd25, 20;
add.s64 %rd26, %rd4, %rd24;
add.s64 %rd9, %rd26, 12;
cvta.to.global.u64 %rd10, %rd17;
mov.f32 %f265, 0f00000000;

$L__BB301_2:
cvt.u64.u32 %rd27, %r870;
add.s64 %rd28, %rd1, %rd27;
add.s64 %rd29, %rd28, %rd3;
mul.lo.s64 %rd30, %rd29, 36;
add.s64 %rd31, %rd107, %rd8;
ld.global.nc.u32 %r479, [%rd31+-12];
and.b32 %r480, %r479, -2004318072;
shr.u32 %r481, %r480, 1;
or.b32 %r482, %r481, 839922192;
and.b32 %r483, %r479, 30583;
mov.u32 %r484, -152379953;
mov.u32 %r485, -1079142271;
prmt.b32 %r486, %r485, %r484, %r483;
mov.u32 %r487, 1901675829;
mov.u32 %r488, 639175937;
prmt.b32 %r489, %r488, %r487, %r483;
and.b32 %r490, %r482, 30292;
prmt.b32 %r491, %r486, %r489, %r490;
shr.u32 %r492, %r479, 16;
and.b32 %r493, %r492, 30583;
prmt.b32 %r494, %r485, %r484, %r493;
prmt.b32 %r495, %r488, %r487, %r493;
shr.u32 %r496, %r482, 16;
prmt.b32 %r497, %r494, %r495, %r496;
mov.u32 %r498, 25632;
prmt.b32 %r414, %r491, %r497, %r498;
mov.u32 %r499, 30001;
prmt.b32 %r418, %r491, %r497, %r499;
add.s64 %rd32, %rd10, %rd30;
ld.global.nc.u32 %r85, [%rd32+4];
ld.global.nc.u32 %r89, [%rd32+20];
mov.u32 %r449, 0;

dp4a.s32.s32 %r83, %r414, %r85, %r449;


dp4a.s32.s32 %r87, %r418, %r89, %r83;

ld.global.nc.u32 %r500, [%rd31+-8];
and.b32 %r501, %r500, -2004318072;
shr.u32 %r502, %r501, 1;
or.b32 %r503, %r502, 839922192;
and.b32 %r504, %r500, 30583;
prmt.b32 %r505, %r485, %r484, %r504;
prmt.b32 %r506, %r488, %r487, %r504;
and.b32 %r507, %r503, 30292;
prmt.b32 %r508, %r505, %r506, %r507;
shr.u32 %r509, %r500, 16;
and.b32 %r510, %r509, 30583;
prmt.b32 %r511, %r485, %r484, %r510;
prmt.b32 %r512, %r488, %r487, %r510;
shr.u32 %r513, %r503, 16;
prmt.b32 %r514, %r511, %r512, %r513;
prmt.b32 %r422, %r508, %r514, %r498;
prmt.b32 %r426, %r508, %r514, %r499;
ld.global.nc.u32 %r93, [%rd32+8];
ld.global.nc.u32 %r97, [%rd32+24];

dp4a.s32.s32 %r91, %r422, %r93, %r87;


dp4a.s32.s32 %r95, %r426, %r97, %r91;

ld.global.nc.u32 %r515, [%rd31+-4];
and.b32 %r516, %r515, -2004318072;
shr.u32 %r517, %r516, 1;
or.b32 %r518, %r517, 839922192;
and.b32 %r519, %r515, 30583;
prmt.b32 %r520, %r485, %r484, %r519;
prmt.b32 %r521, %r488, %r487, %r519;
and.b32 %r522, %r518, 30292;
prmt.b32 %r523, %r520, %r521, %r522;
shr.u32 %r524, %r515, 16;
and.b32 %r525, %r524, 30583;
prmt.b32 %r526, %r485, %r484, %r525;
prmt.b32 %r527, %r488, %r487, %r525;
shr.u32 %r528, %r518, 16;
prmt.b32 %r529, %r526, %r527, %r528;
prmt.b32 %r430, %r523, %r529, %r498;
prmt.b32 %r434, %r523, %r529, %r499;
ld.global.nc.u32 %r101, [%rd32+12];
ld.global.nc.u32 %r105, [%rd32+28];

dp4a.s32.s32 %r99, %r430, %r101, %r95;


dp4a.s32.s32 %r103, %r434, %r105, %r99;

ld.global.nc.u32 %r530, [%rd31];
and.b32 %r531, %r530, -2004318072;
shr.u32 %r532, %r531, 1;
or.b32 %r533, %r532, 839922192;
and.b32 %r534, %r530, 30583;
prmt.b32 %r535, %r485, %r484, %r534;
prmt.b32 %r536, %r488, %r487, %r534;
and.b32 %r537, %r533, 30292;
prmt.b32 %r538, %r535, %r536, %r537;
shr.u32 %r539, %r530, 16;
and.b32 %r540, %r539, 30583;
prmt.b32 %r541, %r485, %r484, %r540;
prmt.b32 %r542, %r488, %r487, %r540;
shr.u32 %r543, %r533, 16;
prmt.b32 %r544, %r541, %r542, %r543;
prmt.b32 %r438, %r538, %r544, %r498;
prmt.b32 %r442, %r538, %r544, %r499;
ld.global.nc.u32 %r109, [%rd32+16];
ld.global.nc.u32 %r113, [%rd32+32];

dp4a.s32.s32 %r107, %r438, %r109, %r103;


dp4a.s32.s32 %r111, %r442, %r113, %r107;

add.s64 %rd33, %rd107, %rd6;
ld.global.nc.u8 %rs13, [%rd33];
cvt.u32.u16 %r545, %rs13;
and.b32 %r546, %r545, 255;
shr.u32 %r547, %r546, %r7;
and.b32 %r548, %r547, 15;
add.s64 %rd34, %rd107, %rd5;
ld.global.nc.u16 %rs14, [%rd34+2];
cvt.u32.u16 %r549, %rs14;
shr.u32 %r550, %r549, %r8;
and.b32 %r551, %r550, 3;
bfi.b32 %r552, %r551, %r548, 4, 2;
add.s32 %r553, %r552, -32;
mul.lo.s32 %r554, %r553, %r111;
ld.global.nc.u16 %rs1, [%rd34];

{ cvt.f32.f16 %f73, %rs1;}


ld.global.nc.u32 %r115, [%rd32];

{.reg .f16 low,high;
mov.b32 {low,high},%r115;
cvt.f32.f16 %f74, low;}


mul.ftz.f32 %f97, %f73, %f74;
cvt.rn.f32.s32 %f98, %r554;
fma.rn.ftz.f32 %f276, %f97, %f98, %f276;
add.s64 %rd35, %rd107, %rd9;
ld.global.nc.u32 %r555, [%rd35+-4];
and.b32 %r556, %r555, -2004318072;
shr.u32 %r557, %r556, 1;
or.b32 %r558, %r557, 839922192;
and.b32 %r559, %r555, 30583;
prmt.b32 %r560, %r485, %r484, %r559;
prmt.b32 %r561, %r488, %r487, %r559;
and.b32 %r562, %r558, 30292;
prmt.b32 %r563, %r560, %r561, %r562;
shr.u32 %r564, %r555, 16;
and.b32 %r565, %r564, 30583;
prmt.b32 %r566, %r485, %r484, %r565;
prmt.b32 %r567, %r488, %r487, %r565;
shr.u32 %r568, %r558, 16;
prmt.b32 %r569, %r566, %r567, %r568;
prmt.b32 %r447, %r563, %r569, %r498;
prmt.b32 %r451, %r563, %r569, %r499;

dp4a.s32.s32 %r116, %r447, %r85, %r449;


dp4a.s32.s32 %r120, %r451, %r89, %r116;

ld.global.nc.u32 %r570, [%rd35];
and.b32 %r571, %r570, -2004318072;
shr.u32 %r572, %r571, 1;
or.b32 %r573, %r572, 839922192;
and.b32 %r574, %r570, 30583;
prmt.b32 %r575, %r485, %r484, %r574;
prmt.b32 %r576, %r488, %r487, %r574;
and.b32 %r577, %r573, 30292;
prmt.b32 %r578, %r575, %r576, %r577;
shr.u32 %r579, %r570, 16;
and.b32 %r580, %r579, 30583;
prmt.b32 %r581, %r485, %r484, %r580;
prmt.b32 %r582, %r488, %r487, %r580;
shr.u32 %r583, %r573, 16;
prmt.b32 %r584, %r581, %r582, %r583;
prmt.b32 %r455, %r578, %r584, %r498;
prmt.b32 %r459, %r578, %r584, %r499;

dp4a.s32.s32 %r124, %r455, %r93, %r120;


dp4a.s32.s32 %r128, %r459, %r97, %r124;

ld.global.nc.u32 %r585, [%rd35+4];
and.b32 %r586, %r585, -2004318072;
shr.u32 %r587, %r586, 1;
or.b32 %r588, %r587, 839922192;
and.b32 %r589, %r585, 30583;
prmt.b32 %r590, %r485, %r484, %r589;
prmt.b32 %r591, %r488, %r487, %r589;
and.b32 %r592, %r588, 30292;
prmt.b32 %r593, %r590, %r591, %r592;
shr.u32 %r594, %r585, 16;
and.b32 %r595, %r594, 30583;
prmt.b32 %r596, %r485, %r484, %r595;
prmt.b32 %r597, %r488, %r487, %r595;
shr.u32 %r598, %r588, 16;
prmt.b32 %r599, %r596, %r597, %r598;
prmt.b32 %r463, %r593, %r599, %r498;
prmt.b32 %r467, %r593, %r599, %r499;

dp4a.s32.s32 %r132, %r463, %r101, %r128;


dp4a.s32.s32 %r136, %r467, %r105, %r132;

ld.global.nc.u32 %r600, [%rd35+8];
and.b32 %r601, %r600, -2004318072;
shr.u32 %r602, %r601, 1;
or.b32 %r603, %r602, 839922192;
and.b32 %r604, %r600, 30583;
prmt.b32 %r605, %r485, %r484, %r604;
prmt.b32 %r606, %r488, %r487, %r604;
and.b32 %r607, %r603, 30292;
prmt.b32 %r608, %r605, %r606, %r607;
shr.u32 %r609, %r600, 16;
and.b32 %r610, %r609, 30583;
prmt.b32 %r611, %r485, %r484, %r610;
prmt.b32 %r612, %r488, %r487, %r610;
shr.u32 %r613, %r603, 16;
prmt.b32 %r614, %r611, %r612, %r613;
prmt.b32 %r471, %r608, %r614, %r498;
prmt.b32 %r475, %r608, %r614, %r499;

dp4a.s32.s32 %r140, %r471, %r109, %r136;


dp4a.s32.s32 %r144, %r475, %r113, %r140;

add.s64 %rd36, %rd107, %rd7;
ld.global.nc.u8 %rs15, [%rd36];
cvt.u32.u16 %r615, %rs15;
and.b32 %r616, %r615, 255;
shr.u32 %r617, %r616, %r7;
and.b32 %r618, %r617, 15;
add.s64 %rd37, %rd107, %rd4;
ld.global.nc.u16 %rs16, [%rd37+2];
cvt.u32.u16 %r619, %rs16;
shr.u32 %r620, %r619, %r8;
and.b32 %r621, %r620, 3;
bfi.b32 %r622, %r621, %r618, 4, 2;
add.s32 %r623, %r622, -32;
mul.lo.s32 %r624, %r623, %r144;
ld.global.nc.u16 %rs2, [%rd37];

{ cvt.f32.f16 %f75, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r115;
cvt.f32.f16 %f76, low;}


mul.ftz.f32 %f99, %f75, %f76;
cvt.rn.f32.s32 %f100, %r624;
fma.rn.ftz.f32 %f275, %f99, %f100, %f275;
cvt.u64.u32 %rd38, %r871;
add.s64 %rd39, %rd1, %rd38;
add.s64 %rd40, %rd39, %rd3;
mul.lo.s64 %rd41, %rd40, 36;
add.s64 %rd42, %rd10, %rd41;
ld.global.nc.u32 %r151, [%rd42+4];
ld.global.nc.u32 %r155, [%rd42+20];

dp4a.s32.s32 %r149, %r414, %r151, %r449;


dp4a.s32.s32 %r153, %r418, %r155, %r149;

ld.global.nc.u32 %r159, [%rd42+8];
ld.global.nc.u32 %r163, [%rd42+24];

dp4a.s32.s32 %r157, %r422, %r159, %r153;


dp4a.s32.s32 %r161, %r426, %r163, %r157;

ld.global.nc.u32 %r167, [%rd42+12];
ld.global.nc.u32 %r171, [%rd42+28];

dp4a.s32.s32 %r165, %r430, %r167, %r161;


dp4a.s32.s32 %r169, %r434, %r171, %r165;

ld.global.nc.u32 %r175, [%rd42+16];
ld.global.nc.u32 %r179, [%rd42+32];

dp4a.s32.s32 %r173, %r438, %r175, %r169;


dp4a.s32.s32 %r177, %r442, %r179, %r173;

mul.lo.s32 %r625, %r553, %r177;

{ cvt.f32.f16 %f77, %rs1;}


ld.global.nc.u32 %r181, [%rd42];

{.reg .f16 low,high;
mov.b32 {low,high},%r181;
cvt.f32.f16 %f78, low;}


mul.ftz.f32 %f101, %f77, %f78;
cvt.rn.f32.s32 %f102, %r625;
fma.rn.ftz.f32 %f274, %f101, %f102, %f274;

dp4a.s32.s32 %r182, %r447, %r151, %r449;


dp4a.s32.s32 %r186, %r451, %r155, %r182;


dp4a.s32.s32 %r190, %r455, %r159, %r186;


dp4a.s32.s32 %r194, %r459, %r163, %r190;


dp4a.s32.s32 %r198, %r463, %r167, %r194;


dp4a.s32.s32 %r202, %r467, %r171, %r198;


dp4a.s32.s32 %r206, %r471, %r175, %r202;


dp4a.s32.s32 %r210, %r475, %r179, %r206;

mul.lo.s32 %r626, %r623, %r210;

{ cvt.f32.f16 %f79, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r181;
cvt.f32.f16 %f80, low;}


mul.ftz.f32 %f103, %f79, %f80;
cvt.rn.f32.s32 %f104, %r626;
fma.rn.ftz.f32 %f273, %f103, %f104, %f273;
cvt.u64.u32 %rd43, %r875;
add.s64 %rd44, %rd1, %rd43;
add.s64 %rd45, %rd44, %rd3;
mul.lo.s64 %rd46, %rd45, 36;
add.s64 %rd47, %rd10, %rd46;
ld.global.nc.u32 %r217, [%rd47+4];
ld.global.nc.u32 %r221, [%rd47+20];

dp4a.s32.s32 %r215, %r414, %r217, %r449;


dp4a.s32.s32 %r219, %r418, %r221, %r215;

ld.global.nc.u32 %r225, [%rd47+8];
ld.global.nc.u32 %r229, [%rd47+24];

dp4a.s32.s32 %r223, %r422, %r225, %r219;


dp4a.s32.s32 %r227, %r426, %r229, %r223;

ld.global.nc.u32 %r233, [%rd47+12];
ld.global.nc.u32 %r237, [%rd47+28];

dp4a.s32.s32 %r231, %r430, %r233, %r227;


dp4a.s32.s32 %r235, %r434, %r237, %r231;

ld.global.nc.u32 %r241, [%rd47+16];
ld.global.nc.u32 %r245, [%rd47+32];

dp4a.s32.s32 %r239, %r438, %r241, %r235;


dp4a.s32.s32 %r243, %r442, %r245, %r239;

mul.lo.s32 %r627, %r553, %r243;

{ cvt.f32.f16 %f81, %rs1;}


ld.global.nc.u32 %r247, [%rd47];

{.reg .f16 low,high;
mov.b32 {low,high},%r247;
cvt.f32.f16 %f82, low;}


mul.ftz.f32 %f105, %f81, %f82;
cvt.rn.f32.s32 %f106, %r627;
fma.rn.ftz.f32 %f272, %f105, %f106, %f272;

dp4a.s32.s32 %r248, %r447, %r217, %r449;


dp4a.s32.s32 %r252, %r451, %r221, %r248;


dp4a.s32.s32 %r256, %r455, %r225, %r252;


dp4a.s32.s32 %r260, %r459, %r229, %r256;


dp4a.s32.s32 %r264, %r463, %r233, %r260;


dp4a.s32.s32 %r268, %r467, %r237, %r264;


dp4a.s32.s32 %r272, %r471, %r241, %r268;


dp4a.s32.s32 %r276, %r475, %r245, %r272;

mul.lo.s32 %r628, %r623, %r276;

{ cvt.f32.f16 %f83, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r247;
cvt.f32.f16 %f84, low;}


mul.ftz.f32 %f107, %f83, %f84;
cvt.rn.f32.s32 %f108, %r628;
fma.rn.ftz.f32 %f271, %f107, %f108, %f271;
cvt.u64.u32 %rd48, %r874;
add.s64 %rd49, %rd1, %rd48;
add.s64 %rd50, %rd49, %rd3;
mul.lo.s64 %rd51, %rd50, 36;
add.s64 %rd52, %rd10, %rd51;
ld.global.nc.u32 %r283, [%rd52+4];
ld.global.nc.u32 %r287, [%rd52+20];

dp4a.s32.s32 %r281, %r414, %r283, %r449;


dp4a.s32.s32 %r285, %r418, %r287, %r281;

ld.global.nc.u32 %r291, [%rd52+8];
ld.global.nc.u32 %r295, [%rd52+24];

dp4a.s32.s32 %r289, %r422, %r291, %r285;


dp4a.s32.s32 %r293, %r426, %r295, %r289;

ld.global.nc.u32 %r299, [%rd52+12];
ld.global.nc.u32 %r303, [%rd52+28];

dp4a.s32.s32 %r297, %r430, %r299, %r293;


dp4a.s32.s32 %r301, %r434, %r303, %r297;

ld.global.nc.u32 %r307, [%rd52+16];
ld.global.nc.u32 %r311, [%rd52+32];

dp4a.s32.s32 %r305, %r438, %r307, %r301;


dp4a.s32.s32 %r309, %r442, %r311, %r305;

mul.lo.s32 %r629, %r553, %r309;

{ cvt.f32.f16 %f85, %rs1;}


ld.global.nc.u32 %r313, [%rd52];

{.reg .f16 low,high;
mov.b32 {low,high},%r313;
cvt.f32.f16 %f86, low;}


mul.ftz.f32 %f109, %f85, %f86;
cvt.rn.f32.s32 %f110, %r629;
fma.rn.ftz.f32 %f270, %f109, %f110, %f270;

dp4a.s32.s32 %r314, %r447, %r283, %r449;


dp4a.s32.s32 %r318, %r451, %r287, %r314;


dp4a.s32.s32 %r322, %r455, %r291, %r318;


dp4a.s32.s32 %r326, %r459, %r295, %r322;


dp4a.s32.s32 %r330, %r463, %r299, %r326;


dp4a.s32.s32 %r334, %r467, %r303, %r330;


dp4a.s32.s32 %r338, %r471, %r307, %r334;


dp4a.s32.s32 %r342, %r475, %r311, %r338;

mul.lo.s32 %r630, %r623, %r342;

{ cvt.f32.f16 %f87, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r313;
cvt.f32.f16 %f88, low;}


mul.ftz.f32 %f111, %f87, %f88;
cvt.rn.f32.s32 %f112, %r630;
fma.rn.ftz.f32 %f269, %f111, %f112, %f269;
cvt.u64.u32 %rd53, %r873;
add.s64 %rd54, %rd1, %rd53;
add.s64 %rd55, %rd54, %rd3;
mul.lo.s64 %rd56, %rd55, 36;
add.s64 %rd57, %rd10, %rd56;
ld.global.nc.u32 %r349, [%rd57+4];
ld.global.nc.u32 %r353, [%rd57+20];

dp4a.s32.s32 %r347, %r414, %r349, %r449;


dp4a.s32.s32 %r351, %r418, %r353, %r347;

ld.global.nc.u32 %r357, [%rd57+8];
ld.global.nc.u32 %r361, [%rd57+24];

dp4a.s32.s32 %r355, %r422, %r357, %r351;


dp4a.s32.s32 %r359, %r426, %r361, %r355;

ld.global.nc.u32 %r365, [%rd57+12];
ld.global.nc.u32 %r369, [%rd57+28];

dp4a.s32.s32 %r363, %r430, %r365, %r359;


dp4a.s32.s32 %r367, %r434, %r369, %r363;

ld.global.nc.u32 %r373, [%rd57+16];
ld.global.nc.u32 %r377, [%rd57+32];

dp4a.s32.s32 %r371, %r438, %r373, %r367;


dp4a.s32.s32 %r375, %r442, %r377, %r371;

mul.lo.s32 %r631, %r553, %r375;

{ cvt.f32.f16 %f89, %rs1;}


ld.global.nc.u32 %r379, [%rd57];

{.reg .f16 low,high;
mov.b32 {low,high},%r379;
cvt.f32.f16 %f90, low;}


mul.ftz.f32 %f113, %f89, %f90;
cvt.rn.f32.s32 %f114, %r631;
fma.rn.ftz.f32 %f268, %f113, %f114, %f268;

dp4a.s32.s32 %r380, %r447, %r349, %r449;


dp4a.s32.s32 %r384, %r451, %r353, %r380;


dp4a.s32.s32 %r388, %r455, %r357, %r384;


dp4a.s32.s32 %r392, %r459, %r361, %r388;


dp4a.s32.s32 %r396, %r463, %r365, %r392;


dp4a.s32.s32 %r400, %r467, %r369, %r396;


dp4a.s32.s32 %r404, %r471, %r373, %r400;


dp4a.s32.s32 %r408, %r475, %r377, %r404;

mul.lo.s32 %r632, %r623, %r408;

{ cvt.f32.f16 %f91, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r379;
cvt.f32.f16 %f92, low;}


mul.ftz.f32 %f115, %f91, %f92;
cvt.rn.f32.s32 %f116, %r632;
fma.rn.ftz.f32 %f267, %f115, %f116, %f267;
cvt.u64.u32 %rd58, %r872;
add.s64 %rd59, %rd1, %rd58;
add.s64 %rd60, %rd59, %rd3;
mul.lo.s64 %rd61, %rd60, 36;
add.s64 %rd62, %rd10, %rd61;
ld.global.nc.u32 %r415, [%rd62+4];
ld.global.nc.u32 %r419, [%rd62+20];

dp4a.s32.s32 %r413, %r414, %r415, %r449;


dp4a.s32.s32 %r417, %r418, %r419, %r413;

ld.global.nc.u32 %r423, [%rd62+8];
ld.global.nc.u32 %r427, [%rd62+24];

dp4a.s32.s32 %r421, %r422, %r423, %r417;


dp4a.s32.s32 %r425, %r426, %r427, %r421;

ld.global.nc.u32 %r431, [%rd62+12];
ld.global.nc.u32 %r435, [%rd62+28];

dp4a.s32.s32 %r429, %r430, %r431, %r425;


dp4a.s32.s32 %r433, %r434, %r435, %r429;

ld.global.nc.u32 %r439, [%rd62+16];
ld.global.nc.u32 %r443, [%rd62+32];

dp4a.s32.s32 %r437, %r438, %r439, %r433;


dp4a.s32.s32 %r441, %r442, %r443, %r437;

mul.lo.s32 %r633, %r553, %r441;

{ cvt.f32.f16 %f93, %rs1;}


ld.global.nc.u32 %r445, [%rd62];

{.reg .f16 low,high;
mov.b32 {low,high},%r445;
cvt.f32.f16 %f94, low;}


mul.ftz.f32 %f117, %f93, %f94;
cvt.rn.f32.s32 %f118, %r633;
fma.rn.ftz.f32 %f266, %f117, %f118, %f266;

dp4a.s32.s32 %r446, %r447, %r415, %r449;


dp4a.s32.s32 %r450, %r451, %r419, %r446;


dp4a.s32.s32 %r454, %r455, %r423, %r450;


dp4a.s32.s32 %r458, %r459, %r427, %r454;


dp4a.s32.s32 %r462, %r463, %r431, %r458;


dp4a.s32.s32 %r466, %r467, %r435, %r462;


dp4a.s32.s32 %r470, %r471, %r439, %r466;


dp4a.s32.s32 %r474, %r475, %r443, %r470;

mul.lo.s32 %r634, %r623, %r474;

{ cvt.f32.f16 %f95, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r445;
cvt.f32.f16 %f96, low;}


mul.ftz.f32 %f119, %f95, %f96;
cvt.rn.f32.s32 %f120, %r634;
fma.rn.ftz.f32 %f265, %f119, %f120, %f265;
add.s32 %r875, %r875, 64;
add.s32 %r874, %r874, 64;
add.s32 %r873, %r873, 64;
add.s32 %r872, %r872, 64;
add.s32 %r871, %r871, 64;
add.s32 %r870, %r870, 64;
add.s64 %rd107, %rd107, 1088;
add.s32 %r876, %r876, 8;
setp.lt.s32 %p4, %r876, %r6;
@%p4 bra $L__BB301_2;

$L__BB301_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB301_5;

mov.u32 %r635, _ZZ13mul_mat_vec_qIL9ggml_type23ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r636, %r1, 1536, %r635;
shl.b32 %r637, %r2, 2;
add.s32 %r638, %r636, %r637;
st.shared.f32 [%r638+-1536], %f276;
st.shared.f32 [%r638+-1408], %f275;
st.shared.f32 [%r638+-1280], %f274;
st.shared.f32 [%r638+-1152], %f273;
st.shared.f32 [%r638+-1024], %f272;
st.shared.f32 [%r638+-896], %f271;
st.shared.f32 [%r638+-768], %f270;
st.shared.f32 [%r638+-640], %f269;
st.shared.f32 [%r638+-512], %f268;
st.shared.f32 [%r638+-384], %f267;
st.shared.f32 [%r638+-256], %f266;
st.shared.f32 [%r638+-128], %f265;

$L__BB301_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB301_30;

mov.u32 %r639, %ctaid.z;
mov.u32 %r640, %ctaid.y;
mad.lo.s32 %r641, %r640, %r40, %r4;
mad.lo.s32 %r642, %r639, %r46, %r641;
cvt.u64.u32 %rd13, %r642;
shl.b32 %r643, %r2, 2;
mov.u32 %r644, _ZZ13mul_mat_vec_qIL9ggml_type23ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r29, %r644, %r643;
mov.u32 %r645, 2;
ld.shared.f32 %f121, [%r29];
add.ftz.f32 %f122, %f121, %f276;
mov.b32 %r646, %f122;
mov.u32 %r647, 31;
mov.u32 %r648, 16;
mov.u32 %r649, -1;
shfl.sync.bfly.b32 %r650|%p7, %r646, %r648, %r647, %r649;
mov.b32 %f123, %r650;
add.ftz.f32 %f124, %f122, %f123;
mov.b32 %r651, %f124;
mov.u32 %r652, 8;
shfl.sync.bfly.b32 %r653|%p8, %r651, %r652, %r647, %r649;
mov.b32 %f125, %r653;
add.ftz.f32 %f126, %f124, %f125;
mov.b32 %r654, %f126;
mov.u32 %r655, 4;
shfl.sync.bfly.b32 %r656|%p9, %r654, %r655, %r647, %r649;
mov.b32 %f127, %r656;
add.ftz.f32 %f128, %f126, %f127;
mov.b32 %r657, %f128;
shfl.sync.bfly.b32 %r658|%p10, %r657, %r645, %r647, %r649;
mov.b32 %f129, %r658;
add.ftz.f32 %f130, %f128, %f129;
mov.b32 %r659, %f130;
mov.u32 %r660, 1;
shfl.sync.bfly.b32 %r661|%p11, %r659, %r660, %r647, %r649;
mov.b32 %f131, %r661;
add.ftz.f32 %f37, %f130, %f131;
setp.lt.u32 %p12, %r4, %r34;
setp.eq.s32 %p13, %r2, 0;
and.pred %p1, %p13, %p12;
not.pred %p14, %p1;
@%p14 bra $L__BB301_8;

cvta.to.global.u64 %rd63, %rd18;
mul.wide.u32 %rd64, %r642, 4;
add.s64 %rd65, %rd63, %rd64;
st.global.f32 [%rd65], %f37;

$L__BB301_8:
ld.shared.f32 %f132, [%r29+128];
add.ftz.f32 %f133, %f132, %f275;
mov.b32 %r668, %f133;
shfl.sync.bfly.b32 %r672|%p15, %r668, %r648, %r647, %r649;
mov.b32 %f134, %r672;
add.ftz.f32 %f135, %f133, %f134;
mov.b32 %r673, %f135;
shfl.sync.bfly.b32 %r675|%p16, %r673, %r652, %r647, %r649;
mov.b32 %f136, %r675;
add.ftz.f32 %f137, %f135, %f136;
mov.b32 %r676, %f137;
shfl.sync.bfly.b32 %r678|%p17, %r676, %r655, %r647, %r649;
mov.b32 %f138, %r678;
add.ftz.f32 %f139, %f137, %f138;
mov.b32 %r679, %f139;
shfl.sync.bfly.b32 %r681|%p18, %r679, %r645, %r647, %r649;
mov.b32 %f140, %r681;
add.ftz.f32 %f141, %f139, %f140;
mov.b32 %r682, %f141;
shfl.sync.bfly.b32 %r684|%p19, %r682, %r660, %r647, %r649;
mov.b32 %f142, %r684;
add.ftz.f32 %f38, %f141, %f142;
setp.eq.s32 %p20, %r2, 1;
or.b32 %r688, %r4, 1;
setp.lt.u32 %p21, %r688, %r34;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB301_10;

cvta.to.global.u64 %rd66, %rd18;
mul.wide.u32 %rd67, %r642, 4;
add.s64 %rd68, %rd66, %rd67;
st.global.f32 [%rd68+4], %f38;

$L__BB301_10:
ld.shared.f32 %f143, [%r29+256];
add.ftz.f32 %f144, %f143, %f274;
mov.b32 %r695, %f144;
mov.u32 %r696, 31;
mov.u32 %r697, 16;
mov.u32 %r698, -1;
shfl.sync.bfly.b32 %r699|%p23, %r695, %r697, %r696, %r698;
mov.b32 %f145, %r699;
add.ftz.f32 %f146, %f144, %f145;
mov.b32 %r700, %f146;
mov.u32 %r701, 8;
shfl.sync.bfly.b32 %r702|%p24, %r700, %r701, %r696, %r698;
mov.b32 %f147, %r702;
add.ftz.f32 %f148, %f146, %f147;
mov.b32 %r703, %f148;
mov.u32 %r704, 4;
shfl.sync.bfly.b32 %r705|%p25, %r703, %r704, %r696, %r698;
mov.b32 %f149, %r705;
add.ftz.f32 %f150, %f148, %f149;
mov.b32 %r706, %f150;
mov.u32 %r707, 2;
shfl.sync.bfly.b32 %r708|%p26, %r706, %r707, %r696, %r698;
mov.b32 %f151, %r708;
add.ftz.f32 %f152, %f150, %f151;
mov.b32 %r709, %f152;
mov.u32 %r710, 1;
shfl.sync.bfly.b32 %r711|%p27, %r709, %r710, %r696, %r698;
mov.b32 %f153, %r711;
add.ftz.f32 %f39, %f152, %f153;
@%p14 bra $L__BB301_12;

cvt.u64.u32 %rd69, %r34;
add.s64 %rd70, %rd69, %rd13;
cvta.to.global.u64 %rd71, %rd18;
shl.b64 %rd72, %rd70, 2;
add.s64 %rd73, %rd71, %rd72;
st.global.f32 [%rd73], %f39;

$L__BB301_12:
ld.shared.f32 %f154, [%r29+384];
add.ftz.f32 %f155, %f154, %f273;
mov.b32 %r712, %f155;
shfl.sync.bfly.b32 %r716|%p29, %r712, %r697, %r696, %r698;
mov.b32 %f156, %r716;
add.ftz.f32 %f157, %f155, %f156;
mov.b32 %r717, %f157;
shfl.sync.bfly.b32 %r719|%p30, %r717, %r701, %r696, %r698;
mov.b32 %f158, %r719;
add.ftz.f32 %f159, %f157, %f158;
mov.b32 %r720, %f159;
shfl.sync.bfly.b32 %r722|%p31, %r720, %r704, %r696, %r698;
mov.b32 %f160, %r722;
add.ftz.f32 %f161, %f159, %f160;
mov.b32 %r723, %f161;
shfl.sync.bfly.b32 %r725|%p32, %r723, %r707, %r696, %r698;
mov.b32 %f162, %r725;
add.ftz.f32 %f163, %f161, %f162;
mov.b32 %r726, %f163;
shfl.sync.bfly.b32 %r728|%p33, %r726, %r710, %r696, %r698;
mov.b32 %f164, %r728;
add.ftz.f32 %f40, %f163, %f164;
@%p22 bra $L__BB301_14;

add.s32 %r729, %r34, 1;
cvt.u64.u32 %rd74, %r729;
add.s64 %rd75, %rd74, %rd13;
cvta.to.global.u64 %rd76, %rd18;
shl.b64 %rd77, %rd75, 2;
add.s64 %rd78, %rd76, %rd77;
st.global.f32 [%rd78], %f40;

$L__BB301_14:
shl.b32 %r730, %r34, 1;
mov.u32 %r731, 1;
ld.shared.f32 %f165, [%r29+512];
add.ftz.f32 %f166, %f165, %f272;
mov.b32 %r732, %f166;
mov.u32 %r733, 31;
mov.u32 %r734, 16;
mov.u32 %r735, -1;
shfl.sync.bfly.b32 %r736|%p35, %r732, %r734, %r733, %r735;
mov.b32 %f167, %r736;
add.ftz.f32 %f168, %f166, %f167;
mov.b32 %r737, %f168;
mov.u32 %r738, 8;
shfl.sync.bfly.b32 %r739|%p36, %r737, %r738, %r733, %r735;
mov.b32 %f169, %r739;
add.ftz.f32 %f170, %f168, %f169;
mov.b32 %r740, %f170;
mov.u32 %r741, 4;
shfl.sync.bfly.b32 %r742|%p37, %r740, %r741, %r733, %r735;
mov.b32 %f171, %r742;
add.ftz.f32 %f172, %f170, %f171;
mov.b32 %r743, %f172;
mov.u32 %r744, 2;
shfl.sync.bfly.b32 %r745|%p38, %r743, %r744, %r733, %r735;
mov.b32 %f173, %r745;
add.ftz.f32 %f174, %f172, %f173;
mov.b32 %r746, %f174;
shfl.sync.bfly.b32 %r747|%p39, %r746, %r731, %r733, %r735;
mov.b32 %f175, %r747;
add.ftz.f32 %f41, %f174, %f175;
cvt.u64.u32 %rd79, %r730;
add.s64 %rd80, %rd79, %rd13;
cvta.to.global.u64 %rd81, %rd18;
shl.b64 %rd82, %rd80, 2;
add.s64 %rd14, %rd81, %rd82;
@%p14 bra $L__BB301_16;

st.global.f32 [%rd14], %f41;

$L__BB301_16:
ld.shared.f32 %f176, [%r29+640];
add.ftz.f32 %f177, %f176, %f271;
mov.b32 %r748, %f177;
shfl.sync.bfly.b32 %r752|%p41, %r748, %r734, %r733, %r735;
mov.b32 %f178, %r752;
add.ftz.f32 %f179, %f177, %f178;
mov.b32 %r753, %f179;
shfl.sync.bfly.b32 %r755|%p42, %r753, %r738, %r733, %r735;
mov.b32 %f180, %r755;
add.ftz.f32 %f181, %f179, %f180;
mov.b32 %r756, %f181;
shfl.sync.bfly.b32 %r758|%p43, %r756, %r741, %r733, %r735;
mov.b32 %f182, %r758;
add.ftz.f32 %f183, %f181, %f182;
mov.b32 %r759, %f183;
shfl.sync.bfly.b32 %r761|%p44, %r759, %r744, %r733, %r735;
mov.b32 %f184, %r761;
add.ftz.f32 %f185, %f183, %f184;
mov.b32 %r762, %f185;
shfl.sync.bfly.b32 %r764|%p45, %r762, %r731, %r733, %r735;
mov.b32 %f186, %r764;
add.ftz.f32 %f42, %f185, %f186;
@%p22 bra $L__BB301_18;

st.global.f32 [%rd14+4], %f42;

$L__BB301_18:
mul.lo.s32 %r30, %r34, 3;
ld.shared.f32 %f187, [%r29+768];
add.ftz.f32 %f188, %f187, %f270;
mov.b32 %r765, %f188;
mov.u32 %r766, 31;
mov.u32 %r767, 16;
mov.u32 %r768, -1;
shfl.sync.bfly.b32 %r769|%p47, %r765, %r767, %r766, %r768;
mov.b32 %f189, %r769;
add.ftz.f32 %f190, %f188, %f189;
mov.b32 %r770, %f190;
mov.u32 %r771, 8;
shfl.sync.bfly.b32 %r772|%p48, %r770, %r771, %r766, %r768;
mov.b32 %f191, %r772;
add.ftz.f32 %f192, %f190, %f191;
mov.b32 %r773, %f192;
mov.u32 %r774, 4;
shfl.sync.bfly.b32 %r775|%p49, %r773, %r774, %r766, %r768;
mov.b32 %f193, %r775;
add.ftz.f32 %f194, %f192, %f193;
mov.b32 %r776, %f194;
mov.u32 %r777, 2;
shfl.sync.bfly.b32 %r778|%p50, %r776, %r777, %r766, %r768;
mov.b32 %f195, %r778;
add.ftz.f32 %f196, %f194, %f195;
mov.b32 %r779, %f196;
mov.u32 %r780, 1;
shfl.sync.bfly.b32 %r781|%p51, %r779, %r780, %r766, %r768;
mov.b32 %f197, %r781;
add.ftz.f32 %f43, %f196, %f197;
@%p14 bra $L__BB301_20;

cvt.u64.u32 %rd83, %r30;
add.s64 %rd84, %rd83, %rd13;
shl.b64 %rd86, %rd84, 2;
add.s64 %rd87, %rd81, %rd86;
st.global.f32 [%rd87], %f43;

$L__BB301_20:
ld.shared.f32 %f198, [%r29+896];
add.ftz.f32 %f199, %f198, %f269;
mov.b32 %r782, %f199;
shfl.sync.bfly.b32 %r786|%p53, %r782, %r767, %r766, %r768;
mov.b32 %f200, %r786;
add.ftz.f32 %f201, %f199, %f200;
mov.b32 %r787, %f201;
shfl.sync.bfly.b32 %r789|%p54, %r787, %r771, %r766, %r768;
mov.b32 %f202, %r789;
add.ftz.f32 %f203, %f201, %f202;
mov.b32 %r790, %f203;
shfl.sync.bfly.b32 %r792|%p55, %r790, %r774, %r766, %r768;
mov.b32 %f204, %r792;
add.ftz.f32 %f205, %f203, %f204;
mov.b32 %r793, %f205;
shfl.sync.bfly.b32 %r795|%p56, %r793, %r777, %r766, %r768;
mov.b32 %f206, %r795;
add.ftz.f32 %f207, %f205, %f206;
mov.b32 %r796, %f207;
shfl.sync.bfly.b32 %r798|%p57, %r796, %r780, %r766, %r768;
mov.b32 %f208, %r798;
add.ftz.f32 %f44, %f207, %f208;
@%p22 bra $L__BB301_22;

add.s32 %r799, %r30, 1;
cvt.u64.u32 %rd88, %r799;
add.s64 %rd89, %rd88, %rd13;
shl.b64 %rd91, %rd89, 2;
add.s64 %rd92, %rd81, %rd91;
st.global.f32 [%rd92], %f44;

$L__BB301_22:
shl.b32 %r800, %r34, 2;
mov.u32 %r801, 2;
ld.shared.f32 %f209, [%r29+1024];
add.ftz.f32 %f210, %f209, %f268;
mov.b32 %r802, %f210;
mov.u32 %r803, 31;
mov.u32 %r804, 16;
mov.u32 %r805, -1;
shfl.sync.bfly.b32 %r806|%p59, %r802, %r804, %r803, %r805;
mov.b32 %f211, %r806;
add.ftz.f32 %f212, %f210, %f211;
mov.b32 %r807, %f212;
mov.u32 %r808, 8;
shfl.sync.bfly.b32 %r809|%p60, %r807, %r808, %r803, %r805;
mov.b32 %f213, %r809;
add.ftz.f32 %f214, %f212, %f213;
mov.b32 %r810, %f214;
mov.u32 %r811, 4;
shfl.sync.bfly.b32 %r812|%p61, %r810, %r811, %r803, %r805;
mov.b32 %f215, %r812;
add.ftz.f32 %f216, %f214, %f215;
mov.b32 %r813, %f216;
shfl.sync.bfly.b32 %r814|%p62, %r813, %r801, %r803, %r805;
mov.b32 %f217, %r814;
add.ftz.f32 %f218, %f216, %f217;
mov.b32 %r815, %f218;
mov.u32 %r816, 1;
shfl.sync.bfly.b32 %r817|%p63, %r815, %r816, %r803, %r805;
mov.b32 %f219, %r817;
add.ftz.f32 %f45, %f218, %f219;
cvt.u64.u32 %rd93, %r800;
add.s64 %rd94, %rd93, %rd13;
shl.b64 %rd96, %rd94, 2;
add.s64 %rd15, %rd81, %rd96;
@%p14 bra $L__BB301_24;

st.global.f32 [%rd15], %f45;

$L__BB301_24:
ld.shared.f32 %f220, [%r29+1152];
add.ftz.f32 %f221, %f220, %f267;
mov.b32 %r818, %f221;
shfl.sync.bfly.b32 %r822|%p65, %r818, %r804, %r803, %r805;
mov.b32 %f222, %r822;
add.ftz.f32 %f223, %f221, %f222;
mov.b32 %r823, %f223;
shfl.sync.bfly.b32 %r825|%p66, %r823, %r808, %r803, %r805;
mov.b32 %f224, %r825;
add.ftz.f32 %f225, %f223, %f224;
mov.b32 %r826, %f225;
shfl.sync.bfly.b32 %r828|%p67, %r826, %r811, %r803, %r805;
mov.b32 %f226, %r828;
add.ftz.f32 %f227, %f225, %f226;
mov.b32 %r829, %f227;
shfl.sync.bfly.b32 %r831|%p68, %r829, %r801, %r803, %r805;
mov.b32 %f228, %r831;
add.ftz.f32 %f229, %f227, %f228;
mov.b32 %r832, %f229;
shfl.sync.bfly.b32 %r834|%p69, %r832, %r816, %r803, %r805;
mov.b32 %f230, %r834;
add.ftz.f32 %f46, %f229, %f230;
@%p22 bra $L__BB301_26;

st.global.f32 [%rd15+4], %f46;

$L__BB301_26:
mul.lo.s32 %r31, %r34, 5;
ld.shared.f32 %f231, [%r29+1280];
add.ftz.f32 %f232, %f231, %f266;
mov.b32 %r835, %f232;
mov.u32 %r836, 31;
mov.u32 %r837, 16;
mov.u32 %r838, -1;
shfl.sync.bfly.b32 %r839|%p71, %r835, %r837, %r836, %r838;
mov.b32 %f233, %r839;
add.ftz.f32 %f234, %f232, %f233;
mov.b32 %r840, %f234;
mov.u32 %r841, 8;
shfl.sync.bfly.b32 %r842|%p72, %r840, %r841, %r836, %r838;
mov.b32 %f235, %r842;
add.ftz.f32 %f236, %f234, %f235;
mov.b32 %r843, %f236;
mov.u32 %r844, 4;
shfl.sync.bfly.b32 %r845|%p73, %r843, %r844, %r836, %r838;
mov.b32 %f237, %r845;
add.ftz.f32 %f238, %f236, %f237;
mov.b32 %r846, %f238;
mov.u32 %r847, 2;
shfl.sync.bfly.b32 %r848|%p74, %r846, %r847, %r836, %r838;
mov.b32 %f239, %r848;
add.ftz.f32 %f240, %f238, %f239;
mov.b32 %r849, %f240;
mov.u32 %r850, 1;
shfl.sync.bfly.b32 %r851|%p75, %r849, %r850, %r836, %r838;
mov.b32 %f241, %r851;
add.ftz.f32 %f47, %f240, %f241;
@%p14 bra $L__BB301_28;

cvt.u64.u32 %rd97, %r31;
add.s64 %rd98, %rd97, %rd13;
shl.b64 %rd100, %rd98, 2;
add.s64 %rd101, %rd81, %rd100;
st.global.f32 [%rd101], %f47;

$L__BB301_28:
ld.shared.f32 %f242, [%r29+1408];
add.ftz.f32 %f243, %f242, %f265;
mov.b32 %r852, %f243;
shfl.sync.bfly.b32 %r856|%p77, %r852, %r837, %r836, %r838;
mov.b32 %f244, %r856;
add.ftz.f32 %f245, %f243, %f244;
mov.b32 %r857, %f245;
shfl.sync.bfly.b32 %r859|%p78, %r857, %r841, %r836, %r838;
mov.b32 %f246, %r859;
add.ftz.f32 %f247, %f245, %f246;
mov.b32 %r860, %f247;
shfl.sync.bfly.b32 %r862|%p79, %r860, %r844, %r836, %r838;
mov.b32 %f248, %r862;
add.ftz.f32 %f249, %f247, %f248;
mov.b32 %r863, %f249;
shfl.sync.bfly.b32 %r865|%p80, %r863, %r847, %r836, %r838;
mov.b32 %f250, %r865;
add.ftz.f32 %f251, %f249, %f250;
mov.b32 %r866, %f251;
shfl.sync.bfly.b32 %r868|%p81, %r866, %r850, %r836, %r838;
mov.b32 %f252, %r868;
add.ftz.f32 %f48, %f251, %f252;
@%p22 bra $L__BB301_30;

add.s32 %r869, %r31, 1;
cvt.u64.u32 %rd102, %r869;
add.s64 %rd103, %rd102, %rd13;
shl.b64 %rd105, %rd103, 2;
add.s64 %rd106, %rd81, %rd105;
st.global.f32 [%rd106], %f48;

$L__BB301_30:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type23ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type23ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type23ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type23ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type23ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type23ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type23ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type23ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type23ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<106>;
.reg .b16 %rs<19>;
.reg .f32 %f<323>;
.reg .b32 %r<1052>;
.reg .b64 %rd<118>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type23ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1792];

ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type23ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type23ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type23ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type23ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type23ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type23ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type23ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type23ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type23ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type23ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type23ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r45, [_Z13mul_mat_vec_qIL9ggml_type23ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r46, [_Z13mul_mat_vec_qIL9ggml_type23ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type23ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type23ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type23ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type23ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r47, %r1, 5;
mov.u32 %r48, %tid.x;
add.s32 %r2, %r47, %r48;
shr.s32 %r49, %r2, 31;
shr.u32 %r50, %r49, 29;
add.s32 %r51, %r2, %r50;
shr.s32 %r1051, %r51, 3;
shr.u32 %r52, %r31, 8;
setp.ge.s32 %p2, %r1051, %r52;
mov.f32 %f309, 0f00000000;
mov.f32 %f310, %f309;
mov.f32 %f311, %f309;
mov.f32 %f312, %f309;
mov.f32 %f313, %f309;
mov.f32 %f314, %f309;
mov.f32 %f315, %f309;
mov.f32 %f316, %f309;
mov.f32 %f317, %f309;
mov.f32 %f318, %f309;
mov.f32 %f319, %f309;
mov.f32 %f320, %f309;
mov.f32 %f321, %f309;
mov.f32 %f322, %f309;
@%p2 bra $L__BB302_3;

mov.u32 %r53, %ctaid.x;
shl.b32 %r54, %r53, 1;
mov.u32 %r55, %ctaid.y;
mov.u32 %r56, %ctaid.z;
mul.lo.s32 %r57, %r55, %r39;
cvt.u64.u32 %rd19, %r57;
mul.lo.s32 %r58, %r56, %r45;
cvt.u64.u32 %rd20, %r58;
add.s64 %rd1, %rd20, %rd19;
mul.hi.u32 %r59, %r56, %r41;
add.s32 %r60, %r59, %r56;
shr.u32 %r61, %r60, %r42;
mul.lo.s32 %r62, %r61, %r44;
mul.hi.u32 %r63, %r55, %r35;
add.s32 %r64, %r63, %r55;
shr.u32 %r65, %r64, %r36;
mul.lo.s32 %r66, %r65, %r38;
cvta.to.global.u64 %rd117, %rd16;
and.b32 %r70, %r51, -8;
sub.s32 %r4, %r2, %r70;
shl.b32 %r71, %r4, 2;
shr.u32 %r72, %r4, 31;
add.s32 %r73, %r4, %r72;
shr.s32 %r74, %r73, 1;
cvt.s64.s32 %rd21, %r74;
and.b32 %r5, %r71, 4;
shl.b32 %r75, %r33, 1;
shl.b32 %r1044, %r1051, 3;
add.s32 %r1050, %r1044, %r75;
mad.lo.s32 %r1049, %r33, 3, %r1044;
shl.b32 %r76, %r33, 2;
add.s32 %r1048, %r1044, %r76;
mad.lo.s32 %r1047, %r33, 5, %r1044;
mad.lo.s32 %r1046, %r33, 6, %r1044;
add.s32 %r1045, %r33, %r1044;
add.s32 %r77, %r1051, %r62;
add.s32 %r78, %r77, %r66;
or.b32 %r79, %r54, 1;
mad.lo.s32 %r80, %r32, %r79, %r78;
mul.wide.s32 %rd3, %r80, 136;
mad.lo.s32 %r81, %r54, %r32, %r1051;
add.s32 %r82, %r81, %r62;
add.s32 %r83, %r82, %r66;
mul.wide.s32 %rd4, %r83, 136;
add.s64 %rd22, %rd4, %rd21;
add.s64 %rd5, %rd22, 4;
add.s64 %rd23, %rd3, %rd21;
add.s64 %rd6, %rd23, 4;
mul.wide.s32 %rd24, %r71, 4;
add.s64 %rd25, %rd4, %rd24;
add.s64 %rd7, %rd25, 20;
add.s64 %rd26, %rd3, %rd24;
add.s64 %rd8, %rd26, 12;
cvta.to.global.u64 %rd9, %rd17;
mov.f32 %f309, 0f00000000;

$L__BB302_2:
cvt.u64.u32 %rd27, %r1044;
add.s64 %rd28, %rd1, %rd27;
cvt.s64.s32 %rd29, %r4;
add.s64 %rd30, %rd28, %rd29;
mul.lo.s64 %rd31, %rd30, 36;
add.s64 %rd32, %rd117, %rd7;
ld.global.nc.u32 %r546, [%rd32+-12];
and.b32 %r547, %r546, -2004318072;
shr.u32 %r548, %r547, 1;
or.b32 %r549, %r548, 839922192;
and.b32 %r550, %r546, 30583;
mov.u32 %r551, -152379953;
mov.u32 %r552, -1079142271;
prmt.b32 %r553, %r552, %r551, %r550;
mov.u32 %r554, 1901675829;
mov.u32 %r555, 639175937;
prmt.b32 %r556, %r555, %r554, %r550;
and.b32 %r557, %r549, 30292;
prmt.b32 %r558, %r553, %r556, %r557;
shr.u32 %r559, %r546, 16;
and.b32 %r560, %r559, 30583;
prmt.b32 %r561, %r552, %r551, %r560;
prmt.b32 %r562, %r555, %r554, %r560;
shr.u32 %r563, %r549, 16;
prmt.b32 %r564, %r561, %r562, %r563;
mov.u32 %r565, 25632;
prmt.b32 %r481, %r558, %r564, %r565;
mov.u32 %r566, 30001;
prmt.b32 %r485, %r558, %r564, %r566;
add.s64 %rd33, %rd9, %rd31;
ld.global.nc.u32 %r86, [%rd33+4];
ld.global.nc.u32 %r90, [%rd33+20];
mov.u32 %r516, 0;

dp4a.s32.s32 %r84, %r481, %r86, %r516;


dp4a.s32.s32 %r88, %r485, %r90, %r84;

ld.global.nc.u32 %r567, [%rd32+-8];
and.b32 %r568, %r567, -2004318072;
shr.u32 %r569, %r568, 1;
or.b32 %r570, %r569, 839922192;
and.b32 %r571, %r567, 30583;
prmt.b32 %r572, %r552, %r551, %r571;
prmt.b32 %r573, %r555, %r554, %r571;
and.b32 %r574, %r570, 30292;
prmt.b32 %r575, %r572, %r573, %r574;
shr.u32 %r576, %r567, 16;
and.b32 %r577, %r576, 30583;
prmt.b32 %r578, %r552, %r551, %r577;
prmt.b32 %r579, %r555, %r554, %r577;
shr.u32 %r580, %r570, 16;
prmt.b32 %r581, %r578, %r579, %r580;
prmt.b32 %r489, %r575, %r581, %r565;
prmt.b32 %r493, %r575, %r581, %r566;
ld.global.nc.u32 %r94, [%rd33+8];
ld.global.nc.u32 %r98, [%rd33+24];

dp4a.s32.s32 %r92, %r489, %r94, %r88;


dp4a.s32.s32 %r96, %r493, %r98, %r92;

ld.global.nc.u32 %r582, [%rd32+-4];
and.b32 %r583, %r582, -2004318072;
shr.u32 %r584, %r583, 1;
or.b32 %r585, %r584, 839922192;
and.b32 %r586, %r582, 30583;
prmt.b32 %r587, %r552, %r551, %r586;
prmt.b32 %r588, %r555, %r554, %r586;
and.b32 %r589, %r585, 30292;
prmt.b32 %r590, %r587, %r588, %r589;
shr.u32 %r591, %r582, 16;
and.b32 %r592, %r591, 30583;
prmt.b32 %r593, %r552, %r551, %r592;
prmt.b32 %r594, %r555, %r554, %r592;
shr.u32 %r595, %r585, 16;
prmt.b32 %r596, %r593, %r594, %r595;
prmt.b32 %r497, %r590, %r596, %r565;
prmt.b32 %r501, %r590, %r596, %r566;
ld.global.nc.u32 %r102, [%rd33+12];
ld.global.nc.u32 %r106, [%rd33+28];

dp4a.s32.s32 %r100, %r497, %r102, %r96;


dp4a.s32.s32 %r104, %r501, %r106, %r100;

ld.global.nc.u32 %r597, [%rd32];
and.b32 %r598, %r597, -2004318072;
shr.u32 %r599, %r598, 1;
or.b32 %r600, %r599, 839922192;
and.b32 %r601, %r597, 30583;
prmt.b32 %r602, %r552, %r551, %r601;
prmt.b32 %r603, %r555, %r554, %r601;
and.b32 %r604, %r600, 30292;
prmt.b32 %r605, %r602, %r603, %r604;
shr.u32 %r606, %r597, 16;
and.b32 %r607, %r606, 30583;
prmt.b32 %r608, %r552, %r551, %r607;
prmt.b32 %r609, %r555, %r554, %r607;
shr.u32 %r610, %r600, 16;
prmt.b32 %r611, %r608, %r609, %r610;
prmt.b32 %r505, %r605, %r611, %r565;
prmt.b32 %r509, %r605, %r611, %r566;
ld.global.nc.u32 %r110, [%rd33+16];
ld.global.nc.u32 %r114, [%rd33+32];

dp4a.s32.s32 %r108, %r505, %r110, %r104;


dp4a.s32.s32 %r112, %r509, %r114, %r108;

add.s64 %rd34, %rd117, %rd5;
ld.global.nc.u8 %rs15, [%rd34];
cvt.u32.u16 %r612, %rs15;
and.b32 %r613, %r612, 255;
shr.u32 %r614, %r613, %r5;
and.b32 %r615, %r614, 15;
add.s64 %rd35, %rd117, %rd4;
ld.global.nc.u16 %rs16, [%rd35+2];
cvt.u32.u16 %r616, %rs16;
shl.b32 %r617, %r4, 1;
shr.u32 %r618, %r616, %r617;
and.b32 %r619, %r618, 3;
bfi.b32 %r620, %r619, %r615, 4, 2;
add.s32 %r621, %r620, -32;
mul.lo.s32 %r622, %r621, %r112;
ld.global.nc.u16 %rs1, [%rd35];

{ cvt.f32.f16 %f85, %rs1;}


ld.global.nc.u32 %r116, [%rd33];

{.reg .f16 low,high;
mov.b32 {low,high},%r116;
cvt.f32.f16 %f86, low;}


mul.ftz.f32 %f113, %f85, %f86;
cvt.rn.f32.s32 %f114, %r622;
fma.rn.ftz.f32 %f322, %f113, %f114, %f322;
add.s64 %rd36, %rd117, %rd8;
ld.global.nc.u32 %r623, [%rd36+-4];
and.b32 %r624, %r623, -2004318072;
shr.u32 %r625, %r624, 1;
or.b32 %r626, %r625, 839922192;
and.b32 %r627, %r623, 30583;
prmt.b32 %r628, %r552, %r551, %r627;
prmt.b32 %r629, %r555, %r554, %r627;
and.b32 %r630, %r626, 30292;
prmt.b32 %r631, %r628, %r629, %r630;
shr.u32 %r632, %r623, 16;
and.b32 %r633, %r632, 30583;
prmt.b32 %r634, %r552, %r551, %r633;
prmt.b32 %r635, %r555, %r554, %r633;
shr.u32 %r636, %r626, 16;
prmt.b32 %r637, %r634, %r635, %r636;
prmt.b32 %r514, %r631, %r637, %r565;
prmt.b32 %r518, %r631, %r637, %r566;

dp4a.s32.s32 %r117, %r514, %r86, %r516;


dp4a.s32.s32 %r121, %r518, %r90, %r117;

ld.global.nc.u32 %r638, [%rd36];
and.b32 %r639, %r638, -2004318072;
shr.u32 %r640, %r639, 1;
or.b32 %r641, %r640, 839922192;
and.b32 %r642, %r638, 30583;
prmt.b32 %r643, %r552, %r551, %r642;
prmt.b32 %r644, %r555, %r554, %r642;
and.b32 %r645, %r641, 30292;
prmt.b32 %r646, %r643, %r644, %r645;
shr.u32 %r647, %r638, 16;
and.b32 %r648, %r647, 30583;
prmt.b32 %r649, %r552, %r551, %r648;
prmt.b32 %r650, %r555, %r554, %r648;
shr.u32 %r651, %r641, 16;
prmt.b32 %r652, %r649, %r650, %r651;
prmt.b32 %r522, %r646, %r652, %r565;
prmt.b32 %r526, %r646, %r652, %r566;

dp4a.s32.s32 %r125, %r522, %r94, %r121;


dp4a.s32.s32 %r129, %r526, %r98, %r125;

ld.global.nc.u32 %r653, [%rd36+4];
and.b32 %r654, %r653, -2004318072;
shr.u32 %r655, %r654, 1;
or.b32 %r656, %r655, 839922192;
and.b32 %r657, %r653, 30583;
prmt.b32 %r658, %r552, %r551, %r657;
prmt.b32 %r659, %r555, %r554, %r657;
and.b32 %r660, %r656, 30292;
prmt.b32 %r661, %r658, %r659, %r660;
shr.u32 %r662, %r653, 16;
and.b32 %r663, %r662, 30583;
prmt.b32 %r664, %r552, %r551, %r663;
prmt.b32 %r665, %r555, %r554, %r663;
shr.u32 %r666, %r656, 16;
prmt.b32 %r667, %r664, %r665, %r666;
prmt.b32 %r530, %r661, %r667, %r565;
prmt.b32 %r534, %r661, %r667, %r566;

dp4a.s32.s32 %r133, %r530, %r102, %r129;


dp4a.s32.s32 %r137, %r534, %r106, %r133;

ld.global.nc.u32 %r668, [%rd36+8];
and.b32 %r669, %r668, -2004318072;
shr.u32 %r670, %r669, 1;
or.b32 %r671, %r670, 839922192;
and.b32 %r672, %r668, 30583;
prmt.b32 %r673, %r552, %r551, %r672;
prmt.b32 %r674, %r555, %r554, %r672;
and.b32 %r675, %r671, 30292;
prmt.b32 %r676, %r673, %r674, %r675;
shr.u32 %r677, %r668, 16;
and.b32 %r678, %r677, 30583;
prmt.b32 %r679, %r552, %r551, %r678;
prmt.b32 %r680, %r555, %r554, %r678;
shr.u32 %r681, %r671, 16;
prmt.b32 %r682, %r679, %r680, %r681;
prmt.b32 %r538, %r676, %r682, %r565;
prmt.b32 %r542, %r676, %r682, %r566;

dp4a.s32.s32 %r141, %r538, %r110, %r137;


dp4a.s32.s32 %r145, %r542, %r114, %r141;

add.s64 %rd37, %rd117, %rd6;
ld.global.nc.u8 %rs17, [%rd37];
cvt.u32.u16 %r683, %rs17;
and.b32 %r684, %r683, 255;
shr.u32 %r685, %r684, %r5;
and.b32 %r686, %r685, 15;
add.s64 %rd38, %rd117, %rd3;
ld.global.nc.u16 %rs18, [%rd38+2];
cvt.u32.u16 %r687, %rs18;
shr.u32 %r688, %r687, %r617;
and.b32 %r689, %r688, 3;
bfi.b32 %r690, %r689, %r686, 4, 2;
add.s32 %r691, %r690, -32;
mul.lo.s32 %r692, %r691, %r145;
ld.global.nc.u16 %rs2, [%rd38];

{ cvt.f32.f16 %f87, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r116;
cvt.f32.f16 %f88, low;}


mul.ftz.f32 %f115, %f87, %f88;
cvt.rn.f32.s32 %f116, %r692;
fma.rn.ftz.f32 %f321, %f115, %f116, %f321;
cvt.u64.u32 %rd39, %r1045;
add.s64 %rd40, %rd1, %rd39;
add.s64 %rd41, %rd40, %rd29;
mul.lo.s64 %rd42, %rd41, 36;
add.s64 %rd43, %rd9, %rd42;
ld.global.nc.u32 %r152, [%rd43+4];
ld.global.nc.u32 %r156, [%rd43+20];

dp4a.s32.s32 %r150, %r481, %r152, %r516;


dp4a.s32.s32 %r154, %r485, %r156, %r150;

ld.global.nc.u32 %r160, [%rd43+8];
ld.global.nc.u32 %r164, [%rd43+24];

dp4a.s32.s32 %r158, %r489, %r160, %r154;


dp4a.s32.s32 %r162, %r493, %r164, %r158;

ld.global.nc.u32 %r168, [%rd43+12];
ld.global.nc.u32 %r172, [%rd43+28];

dp4a.s32.s32 %r166, %r497, %r168, %r162;


dp4a.s32.s32 %r170, %r501, %r172, %r166;

ld.global.nc.u32 %r176, [%rd43+16];
ld.global.nc.u32 %r180, [%rd43+32];

dp4a.s32.s32 %r174, %r505, %r176, %r170;


dp4a.s32.s32 %r178, %r509, %r180, %r174;

mul.lo.s32 %r693, %r621, %r178;

{ cvt.f32.f16 %f89, %rs1;}


ld.global.nc.u32 %r182, [%rd43];

{.reg .f16 low,high;
mov.b32 {low,high},%r182;
cvt.f32.f16 %f90, low;}


mul.ftz.f32 %f117, %f89, %f90;
cvt.rn.f32.s32 %f118, %r693;
fma.rn.ftz.f32 %f320, %f117, %f118, %f320;

dp4a.s32.s32 %r183, %r514, %r152, %r516;


dp4a.s32.s32 %r187, %r518, %r156, %r183;


dp4a.s32.s32 %r191, %r522, %r160, %r187;


dp4a.s32.s32 %r195, %r526, %r164, %r191;


dp4a.s32.s32 %r199, %r530, %r168, %r195;


dp4a.s32.s32 %r203, %r534, %r172, %r199;


dp4a.s32.s32 %r207, %r538, %r176, %r203;


dp4a.s32.s32 %r211, %r542, %r180, %r207;

mul.lo.s32 %r694, %r691, %r211;

{ cvt.f32.f16 %f91, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r182;
cvt.f32.f16 %f92, low;}


mul.ftz.f32 %f119, %f91, %f92;
cvt.rn.f32.s32 %f120, %r694;
fma.rn.ftz.f32 %f319, %f119, %f120, %f319;
cvt.u64.u32 %rd44, %r1050;
add.s64 %rd45, %rd1, %rd44;
add.s64 %rd46, %rd45, %rd29;
mul.lo.s64 %rd47, %rd46, 36;
add.s64 %rd48, %rd9, %rd47;
ld.global.nc.u32 %r218, [%rd48+4];
ld.global.nc.u32 %r222, [%rd48+20];

dp4a.s32.s32 %r216, %r481, %r218, %r516;


dp4a.s32.s32 %r220, %r485, %r222, %r216;

ld.global.nc.u32 %r226, [%rd48+8];
ld.global.nc.u32 %r230, [%rd48+24];

dp4a.s32.s32 %r224, %r489, %r226, %r220;


dp4a.s32.s32 %r228, %r493, %r230, %r224;

ld.global.nc.u32 %r234, [%rd48+12];
ld.global.nc.u32 %r238, [%rd48+28];

dp4a.s32.s32 %r232, %r497, %r234, %r228;


dp4a.s32.s32 %r236, %r501, %r238, %r232;

ld.global.nc.u32 %r242, [%rd48+16];
ld.global.nc.u32 %r246, [%rd48+32];

dp4a.s32.s32 %r240, %r505, %r242, %r236;


dp4a.s32.s32 %r244, %r509, %r246, %r240;

mul.lo.s32 %r695, %r621, %r244;

{ cvt.f32.f16 %f93, %rs1;}


ld.global.nc.u32 %r248, [%rd48];

{.reg .f16 low,high;
mov.b32 {low,high},%r248;
cvt.f32.f16 %f94, low;}


mul.ftz.f32 %f121, %f93, %f94;
cvt.rn.f32.s32 %f122, %r695;
fma.rn.ftz.f32 %f318, %f121, %f122, %f318;

dp4a.s32.s32 %r249, %r514, %r218, %r516;


dp4a.s32.s32 %r253, %r518, %r222, %r249;


dp4a.s32.s32 %r257, %r522, %r226, %r253;


dp4a.s32.s32 %r261, %r526, %r230, %r257;


dp4a.s32.s32 %r265, %r530, %r234, %r261;


dp4a.s32.s32 %r269, %r534, %r238, %r265;


dp4a.s32.s32 %r273, %r538, %r242, %r269;


dp4a.s32.s32 %r277, %r542, %r246, %r273;

mul.lo.s32 %r696, %r691, %r277;

{ cvt.f32.f16 %f95, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r248;
cvt.f32.f16 %f96, low;}


mul.ftz.f32 %f123, %f95, %f96;
cvt.rn.f32.s32 %f124, %r696;
fma.rn.ftz.f32 %f317, %f123, %f124, %f317;
cvt.u64.u32 %rd49, %r1049;
add.s64 %rd50, %rd1, %rd49;
add.s64 %rd51, %rd50, %rd29;
mul.lo.s64 %rd52, %rd51, 36;
add.s64 %rd53, %rd9, %rd52;
ld.global.nc.u32 %r284, [%rd53+4];
ld.global.nc.u32 %r288, [%rd53+20];

dp4a.s32.s32 %r282, %r481, %r284, %r516;


dp4a.s32.s32 %r286, %r485, %r288, %r282;

ld.global.nc.u32 %r292, [%rd53+8];
ld.global.nc.u32 %r296, [%rd53+24];

dp4a.s32.s32 %r290, %r489, %r292, %r286;


dp4a.s32.s32 %r294, %r493, %r296, %r290;

ld.global.nc.u32 %r300, [%rd53+12];
ld.global.nc.u32 %r304, [%rd53+28];

dp4a.s32.s32 %r298, %r497, %r300, %r294;


dp4a.s32.s32 %r302, %r501, %r304, %r298;

ld.global.nc.u32 %r308, [%rd53+16];
ld.global.nc.u32 %r312, [%rd53+32];

dp4a.s32.s32 %r306, %r505, %r308, %r302;


dp4a.s32.s32 %r310, %r509, %r312, %r306;

mul.lo.s32 %r697, %r621, %r310;

{ cvt.f32.f16 %f97, %rs1;}


ld.global.nc.u32 %r314, [%rd53];

{.reg .f16 low,high;
mov.b32 {low,high},%r314;
cvt.f32.f16 %f98, low;}


mul.ftz.f32 %f125, %f97, %f98;
cvt.rn.f32.s32 %f126, %r697;
fma.rn.ftz.f32 %f316, %f125, %f126, %f316;

dp4a.s32.s32 %r315, %r514, %r284, %r516;


dp4a.s32.s32 %r319, %r518, %r288, %r315;


dp4a.s32.s32 %r323, %r522, %r292, %r319;


dp4a.s32.s32 %r327, %r526, %r296, %r323;


dp4a.s32.s32 %r331, %r530, %r300, %r327;


dp4a.s32.s32 %r335, %r534, %r304, %r331;


dp4a.s32.s32 %r339, %r538, %r308, %r335;


dp4a.s32.s32 %r343, %r542, %r312, %r339;

mul.lo.s32 %r698, %r691, %r343;

{ cvt.f32.f16 %f99, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r314;
cvt.f32.f16 %f100, low;}


mul.ftz.f32 %f127, %f99, %f100;
cvt.rn.f32.s32 %f128, %r698;
fma.rn.ftz.f32 %f315, %f127, %f128, %f315;
cvt.u64.u32 %rd54, %r1048;
add.s64 %rd55, %rd1, %rd54;
add.s64 %rd56, %rd55, %rd29;
mul.lo.s64 %rd57, %rd56, 36;
add.s64 %rd58, %rd9, %rd57;
ld.global.nc.u32 %r350, [%rd58+4];
ld.global.nc.u32 %r354, [%rd58+20];

dp4a.s32.s32 %r348, %r481, %r350, %r516;


dp4a.s32.s32 %r352, %r485, %r354, %r348;

ld.global.nc.u32 %r358, [%rd58+8];
ld.global.nc.u32 %r362, [%rd58+24];

dp4a.s32.s32 %r356, %r489, %r358, %r352;


dp4a.s32.s32 %r360, %r493, %r362, %r356;

ld.global.nc.u32 %r366, [%rd58+12];
ld.global.nc.u32 %r370, [%rd58+28];

dp4a.s32.s32 %r364, %r497, %r366, %r360;


dp4a.s32.s32 %r368, %r501, %r370, %r364;

ld.global.nc.u32 %r374, [%rd58+16];
ld.global.nc.u32 %r378, [%rd58+32];

dp4a.s32.s32 %r372, %r505, %r374, %r368;


dp4a.s32.s32 %r376, %r509, %r378, %r372;

mul.lo.s32 %r699, %r621, %r376;

{ cvt.f32.f16 %f101, %rs1;}


ld.global.nc.u32 %r380, [%rd58];

{.reg .f16 low,high;
mov.b32 {low,high},%r380;
cvt.f32.f16 %f102, low;}


mul.ftz.f32 %f129, %f101, %f102;
cvt.rn.f32.s32 %f130, %r699;
fma.rn.ftz.f32 %f314, %f129, %f130, %f314;

dp4a.s32.s32 %r381, %r514, %r350, %r516;


dp4a.s32.s32 %r385, %r518, %r354, %r381;


dp4a.s32.s32 %r389, %r522, %r358, %r385;


dp4a.s32.s32 %r393, %r526, %r362, %r389;


dp4a.s32.s32 %r397, %r530, %r366, %r393;


dp4a.s32.s32 %r401, %r534, %r370, %r397;


dp4a.s32.s32 %r405, %r538, %r374, %r401;


dp4a.s32.s32 %r409, %r542, %r378, %r405;

mul.lo.s32 %r700, %r691, %r409;

{ cvt.f32.f16 %f103, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r380;
cvt.f32.f16 %f104, low;}


mul.ftz.f32 %f131, %f103, %f104;
cvt.rn.f32.s32 %f132, %r700;
fma.rn.ftz.f32 %f313, %f131, %f132, %f313;
cvt.u64.u32 %rd59, %r1047;
add.s64 %rd60, %rd1, %rd59;
add.s64 %rd61, %rd60, %rd29;
mul.lo.s64 %rd62, %rd61, 36;
add.s64 %rd63, %rd9, %rd62;
ld.global.nc.u32 %r416, [%rd63+4];
ld.global.nc.u32 %r420, [%rd63+20];

dp4a.s32.s32 %r414, %r481, %r416, %r516;


dp4a.s32.s32 %r418, %r485, %r420, %r414;

ld.global.nc.u32 %r424, [%rd63+8];
ld.global.nc.u32 %r428, [%rd63+24];

dp4a.s32.s32 %r422, %r489, %r424, %r418;


dp4a.s32.s32 %r426, %r493, %r428, %r422;

ld.global.nc.u32 %r432, [%rd63+12];
ld.global.nc.u32 %r436, [%rd63+28];

dp4a.s32.s32 %r430, %r497, %r432, %r426;


dp4a.s32.s32 %r434, %r501, %r436, %r430;

ld.global.nc.u32 %r440, [%rd63+16];
ld.global.nc.u32 %r444, [%rd63+32];

dp4a.s32.s32 %r438, %r505, %r440, %r434;


dp4a.s32.s32 %r442, %r509, %r444, %r438;

mul.lo.s32 %r701, %r621, %r442;

{ cvt.f32.f16 %f105, %rs1;}


ld.global.nc.u32 %r446, [%rd63];

{.reg .f16 low,high;
mov.b32 {low,high},%r446;
cvt.f32.f16 %f106, low;}


mul.ftz.f32 %f133, %f105, %f106;
cvt.rn.f32.s32 %f134, %r701;
fma.rn.ftz.f32 %f312, %f133, %f134, %f312;

dp4a.s32.s32 %r447, %r514, %r416, %r516;


dp4a.s32.s32 %r451, %r518, %r420, %r447;


dp4a.s32.s32 %r455, %r522, %r424, %r451;


dp4a.s32.s32 %r459, %r526, %r428, %r455;


dp4a.s32.s32 %r463, %r530, %r432, %r459;


dp4a.s32.s32 %r467, %r534, %r436, %r463;


dp4a.s32.s32 %r471, %r538, %r440, %r467;


dp4a.s32.s32 %r475, %r542, %r444, %r471;

mul.lo.s32 %r702, %r691, %r475;

{ cvt.f32.f16 %f107, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r446;
cvt.f32.f16 %f108, low;}


mul.ftz.f32 %f135, %f107, %f108;
cvt.rn.f32.s32 %f136, %r702;
fma.rn.ftz.f32 %f311, %f135, %f136, %f311;
cvt.u64.u32 %rd64, %r1046;
add.s64 %rd65, %rd1, %rd64;
add.s64 %rd66, %rd65, %rd29;
mul.lo.s64 %rd67, %rd66, 36;
add.s64 %rd68, %rd9, %rd67;
ld.global.nc.u32 %r482, [%rd68+4];
ld.global.nc.u32 %r486, [%rd68+20];

dp4a.s32.s32 %r480, %r481, %r482, %r516;


dp4a.s32.s32 %r484, %r485, %r486, %r480;

ld.global.nc.u32 %r490, [%rd68+8];
ld.global.nc.u32 %r494, [%rd68+24];

dp4a.s32.s32 %r488, %r489, %r490, %r484;


dp4a.s32.s32 %r492, %r493, %r494, %r488;

ld.global.nc.u32 %r498, [%rd68+12];
ld.global.nc.u32 %r502, [%rd68+28];

dp4a.s32.s32 %r496, %r497, %r498, %r492;


dp4a.s32.s32 %r500, %r501, %r502, %r496;

ld.global.nc.u32 %r506, [%rd68+16];
ld.global.nc.u32 %r510, [%rd68+32];

dp4a.s32.s32 %r504, %r505, %r506, %r500;


dp4a.s32.s32 %r508, %r509, %r510, %r504;

mul.lo.s32 %r703, %r621, %r508;

{ cvt.f32.f16 %f109, %rs1;}


ld.global.nc.u32 %r512, [%rd68];

{.reg .f16 low,high;
mov.b32 {low,high},%r512;
cvt.f32.f16 %f110, low;}


mul.ftz.f32 %f137, %f109, %f110;
cvt.rn.f32.s32 %f138, %r703;
fma.rn.ftz.f32 %f310, %f137, %f138, %f310;

dp4a.s32.s32 %r513, %r514, %r482, %r516;


dp4a.s32.s32 %r517, %r518, %r486, %r513;


dp4a.s32.s32 %r521, %r522, %r490, %r517;


dp4a.s32.s32 %r525, %r526, %r494, %r521;


dp4a.s32.s32 %r529, %r530, %r498, %r525;


dp4a.s32.s32 %r533, %r534, %r502, %r529;


dp4a.s32.s32 %r537, %r538, %r506, %r533;


dp4a.s32.s32 %r541, %r542, %r510, %r537;

mul.lo.s32 %r704, %r691, %r541;

{ cvt.f32.f16 %f111, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r512;
cvt.f32.f16 %f112, low;}


mul.ftz.f32 %f139, %f111, %f112;
cvt.rn.f32.s32 %f140, %r704;
fma.rn.ftz.f32 %f309, %f139, %f140, %f309;
add.s32 %r1050, %r1050, 64;
add.s32 %r1049, %r1049, 64;
add.s32 %r1048, %r1048, 64;
add.s32 %r1047, %r1047, 64;
add.s32 %r1046, %r1046, 64;
add.s32 %r1045, %r1045, 64;
add.s32 %r1044, %r1044, 64;
add.s64 %rd117, %rd117, 1088;
add.s32 %r1051, %r1051, 8;
setp.lt.s32 %p3, %r1051, %r52;
@%p3 bra $L__BB302_2;

$L__BB302_3:
setp.eq.s32 %p4, %r1, 0;
@%p4 bra $L__BB302_5;

mov.u32 %r707, _ZZ13mul_mat_vec_qIL9ggml_type23ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r708, %r1, 1792, %r707;
shl.b32 %r709, %r48, 2;
add.s32 %r710, %r708, %r709;
st.shared.f32 [%r710+-1792], %f322;
st.shared.f32 [%r710+-1664], %f321;
st.shared.f32 [%r710+-1536], %f320;
st.shared.f32 [%r710+-1408], %f319;
st.shared.f32 [%r710+-1280], %f318;
st.shared.f32 [%r710+-1152], %f317;
st.shared.f32 [%r710+-1024], %f316;
st.shared.f32 [%r710+-896], %f315;
st.shared.f32 [%r710+-768], %f314;
st.shared.f32 [%r710+-640], %f313;
st.shared.f32 [%r710+-512], %f312;
st.shared.f32 [%r710+-384], %f311;
st.shared.f32 [%r710+-256], %f310;
st.shared.f32 [%r710+-128], %f309;

$L__BB302_5:
setp.ne.s32 %p5, %r1, 0;
bar.sync 0;
@%p5 bra $L__BB302_34;

mov.u32 %r711, %ctaid.z;
mov.u32 %r712, %ctaid.y;
mov.u32 %r713, %ctaid.x;
shl.b32 %r714, %r713, 1;
mov.u32 %r715, 1;
mad.lo.s32 %r716, %r712, %r40, %r714;
mad.lo.s32 %r717, %r711, %r46, %r716;
cvt.u64.u32 %rd12, %r717;
shl.b32 %r719, %r48, 2;
mov.u32 %r720, _ZZ13mul_mat_vec_qIL9ggml_type23ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r721, %r720, %r719;
mov.u32 %r722, 2;
ld.shared.f32 %f141, [%r721];
add.ftz.f32 %f142, %f141, %f322;
mov.b32 %r723, %f142;
mov.u32 %r724, 31;
mov.u32 %r725, 16;
mov.u32 %r726, -1;
shfl.sync.bfly.b32 %r727|%p6, %r723, %r725, %r724, %r726;
mov.b32 %f143, %r727;
add.ftz.f32 %f144, %f142, %f143;
mov.b32 %r728, %f144;
mov.u32 %r729, 8;
shfl.sync.bfly.b32 %r730|%p7, %r728, %r729, %r724, %r726;
mov.b32 %f145, %r730;
add.ftz.f32 %f146, %f144, %f145;
mov.b32 %r731, %f146;
mov.u32 %r732, 4;
shfl.sync.bfly.b32 %r733|%p8, %r731, %r732, %r724, %r726;
mov.b32 %f147, %r733;
add.ftz.f32 %f148, %f146, %f147;
mov.b32 %r734, %f148;
shfl.sync.bfly.b32 %r735|%p9, %r734, %r722, %r724, %r726;
mov.b32 %f149, %r735;
add.ftz.f32 %f150, %f148, %f149;
mov.b32 %r736, %f150;
shfl.sync.bfly.b32 %r737|%p10, %r736, %r715, %r724, %r726;
mov.b32 %f151, %r737;
add.ftz.f32 %f43, %f150, %f151;
setp.ne.s32 %p11, %r48, 0;
setp.ge.u32 %p12, %r714, %r34;
or.pred %p13, %p11, %p12;
@%p13 bra $L__BB302_8;

cvta.to.global.u64 %rd69, %rd18;
mul.wide.u32 %rd70, %r717, 4;
add.s64 %rd71, %rd69, %rd70;
st.global.f32 [%rd71], %f43;

$L__BB302_8:
ld.shared.f32 %f152, [%r721+128];
add.ftz.f32 %f153, %f152, %f321;
mov.b32 %r749, %f153;
shfl.sync.bfly.b32 %r753|%p14, %r749, %r725, %r724, %r726;
mov.b32 %f154, %r753;
add.ftz.f32 %f155, %f153, %f154;
mov.b32 %r754, %f155;
shfl.sync.bfly.b32 %r756|%p15, %r754, %r729, %r724, %r726;
mov.b32 %f156, %r756;
add.ftz.f32 %f157, %f155, %f156;
mov.b32 %r757, %f157;
shfl.sync.bfly.b32 %r759|%p16, %r757, %r732, %r724, %r726;
mov.b32 %f158, %r759;
add.ftz.f32 %f159, %f157, %f158;
mov.b32 %r760, %f159;
shfl.sync.bfly.b32 %r761|%p17, %r760, %r722, %r724, %r726;
mov.b32 %f160, %r761;
add.ftz.f32 %f161, %f159, %f160;
mov.b32 %r762, %f161;
shfl.sync.bfly.b32 %r764|%p18, %r762, %r715, %r724, %r726;
mov.b32 %f162, %r764;
add.ftz.f32 %f44, %f161, %f162;
setp.eq.s32 %p19, %r48, 1;
or.b32 %r767, %r714, 1;
setp.lt.u32 %p20, %r767, %r34;
and.pred %p1, %p19, %p20;
not.pred %p21, %p1;
@%p21 bra $L__BB302_10;

cvta.to.global.u64 %rd72, %rd18;
mul.wide.u32 %rd73, %r717, 4;
add.s64 %rd74, %rd72, %rd73;
st.global.f32 [%rd74+4], %f44;

$L__BB302_10:
mov.u32 %r778, 2;
ld.shared.f32 %f163, [%r721+256];
add.ftz.f32 %f164, %f163, %f320;
mov.b32 %r779, %f164;
mov.u32 %r780, 31;
mov.u32 %r781, 16;
mov.u32 %r782, -1;
shfl.sync.bfly.b32 %r783|%p22, %r779, %r781, %r780, %r782;
mov.b32 %f165, %r783;
add.ftz.f32 %f166, %f164, %f165;
mov.b32 %r784, %f166;
mov.u32 %r785, 8;
shfl.sync.bfly.b32 %r786|%p23, %r784, %r785, %r780, %r782;
mov.b32 %f167, %r786;
add.ftz.f32 %f168, %f166, %f167;
mov.b32 %r787, %f168;
mov.u32 %r788, 4;
shfl.sync.bfly.b32 %r789|%p24, %r787, %r788, %r780, %r782;
mov.b32 %f169, %r789;
add.ftz.f32 %f170, %f168, %f169;
mov.b32 %r790, %f170;
shfl.sync.bfly.b32 %r791|%p25, %r790, %r778, %r780, %r782;
mov.b32 %f171, %r791;
add.ftz.f32 %f172, %f170, %f171;
mov.b32 %r792, %f172;
mov.u32 %r793, 1;
shfl.sync.bfly.b32 %r794|%p26, %r792, %r793, %r780, %r782;
mov.b32 %f173, %r794;
add.ftz.f32 %f45, %f172, %f173;
@%p13 bra $L__BB302_12;

cvt.u64.u32 %rd75, %r34;
add.s64 %rd76, %rd75, %rd12;
cvta.to.global.u64 %rd77, %rd18;
shl.b64 %rd78, %rd76, 2;
add.s64 %rd79, %rd77, %rd78;
st.global.f32 [%rd79], %f45;

$L__BB302_12:
ld.shared.f32 %f174, [%r721+384];
add.ftz.f32 %f175, %f174, %f319;
mov.b32 %r802, %f175;
shfl.sync.bfly.b32 %r806|%p30, %r802, %r781, %r780, %r782;
mov.b32 %f176, %r806;
add.ftz.f32 %f177, %f175, %f176;
mov.b32 %r807, %f177;
shfl.sync.bfly.b32 %r809|%p31, %r807, %r785, %r780, %r782;
mov.b32 %f178, %r809;
add.ftz.f32 %f179, %f177, %f178;
mov.b32 %r810, %f179;
shfl.sync.bfly.b32 %r812|%p32, %r810, %r788, %r780, %r782;
mov.b32 %f180, %r812;
add.ftz.f32 %f181, %f179, %f180;
mov.b32 %r813, %f181;
shfl.sync.bfly.b32 %r814|%p33, %r813, %r778, %r780, %r782;
mov.b32 %f182, %r814;
add.ftz.f32 %f183, %f181, %f182;
mov.b32 %r815, %f183;
shfl.sync.bfly.b32 %r817|%p34, %r815, %r793, %r780, %r782;
mov.b32 %f184, %r817;
add.ftz.f32 %f46, %f183, %f184;
@%p21 bra $L__BB302_14;

add.s32 %r818, %r34, 1;
cvt.u64.u32 %rd80, %r818;
add.s64 %rd81, %rd80, %rd12;
cvta.to.global.u64 %rd82, %rd18;
shl.b64 %rd83, %rd81, 2;
add.s64 %rd84, %rd82, %rd83;
st.global.f32 [%rd84], %f46;

$L__BB302_14:
shl.b32 %r819, %r34, 1;
mov.u32 %r820, 1;
mov.u32 %r825, 2;
ld.shared.f32 %f185, [%r721+512];
add.ftz.f32 %f186, %f185, %f318;
mov.b32 %r826, %f186;
mov.u32 %r827, 31;
mov.u32 %r828, 16;
mov.u32 %r829, -1;
shfl.sync.bfly.b32 %r830|%p36, %r826, %r828, %r827, %r829;
mov.b32 %f187, %r830;
add.ftz.f32 %f188, %f186, %f187;
mov.b32 %r831, %f188;
mov.u32 %r832, 8;
shfl.sync.bfly.b32 %r833|%p37, %r831, %r832, %r827, %r829;
mov.b32 %f189, %r833;
add.ftz.f32 %f190, %f188, %f189;
mov.b32 %r834, %f190;
mov.u32 %r835, 4;
shfl.sync.bfly.b32 %r836|%p38, %r834, %r835, %r827, %r829;
mov.b32 %f191, %r836;
add.ftz.f32 %f192, %f190, %f191;
mov.b32 %r837, %f192;
shfl.sync.bfly.b32 %r838|%p39, %r837, %r825, %r827, %r829;
mov.b32 %f193, %r838;
add.ftz.f32 %f194, %f192, %f193;
mov.b32 %r839, %f194;
shfl.sync.bfly.b32 %r840|%p40, %r839, %r820, %r827, %r829;
mov.b32 %f195, %r840;
add.ftz.f32 %f47, %f194, %f195;
cvt.u64.u32 %rd85, %r819;
add.s64 %rd86, %rd85, %rd12;
cvta.to.global.u64 %rd87, %rd18;
shl.b64 %rd88, %rd86, 2;
add.s64 %rd13, %rd87, %rd88;
@%p13 bra $L__BB302_16;

st.global.f32 [%rd13], %f47;

$L__BB302_16:
ld.shared.f32 %f196, [%r721+640];
add.ftz.f32 %f197, %f196, %f317;
mov.b32 %r848, %f197;
shfl.sync.bfly.b32 %r852|%p44, %r848, %r828, %r827, %r829;
mov.b32 %f198, %r852;
add.ftz.f32 %f199, %f197, %f198;
mov.b32 %r853, %f199;
shfl.sync.bfly.b32 %r855|%p45, %r853, %r832, %r827, %r829;
mov.b32 %f200, %r855;
add.ftz.f32 %f201, %f199, %f200;
mov.b32 %r856, %f201;
shfl.sync.bfly.b32 %r858|%p46, %r856, %r835, %r827, %r829;
mov.b32 %f202, %r858;
add.ftz.f32 %f203, %f201, %f202;
mov.b32 %r859, %f203;
shfl.sync.bfly.b32 %r860|%p47, %r859, %r825, %r827, %r829;
mov.b32 %f204, %r860;
add.ftz.f32 %f205, %f203, %f204;
mov.b32 %r861, %f205;
shfl.sync.bfly.b32 %r863|%p48, %r861, %r820, %r827, %r829;
mov.b32 %f206, %r863;
add.ftz.f32 %f48, %f205, %f206;
@%p21 bra $L__BB302_18;

st.global.f32 [%rd13+4], %f48;

$L__BB302_18:
mul.lo.s32 %r29, %r34, 3;
mov.u32 %r868, 2;
ld.shared.f32 %f207, [%r721+768];
add.ftz.f32 %f208, %f207, %f316;
mov.b32 %r869, %f208;
mov.u32 %r870, 31;
mov.u32 %r871, 16;
mov.u32 %r872, -1;
shfl.sync.bfly.b32 %r873|%p50, %r869, %r871, %r870, %r872;
mov.b32 %f209, %r873;
add.ftz.f32 %f210, %f208, %f209;
mov.b32 %r874, %f210;
mov.u32 %r875, 8;
shfl.sync.bfly.b32 %r876|%p51, %r874, %r875, %r870, %r872;
mov.b32 %f211, %r876;
add.ftz.f32 %f212, %f210, %f211;
mov.b32 %r877, %f212;
mov.u32 %r878, 4;
shfl.sync.bfly.b32 %r879|%p52, %r877, %r878, %r870, %r872;
mov.b32 %f213, %r879;
add.ftz.f32 %f214, %f212, %f213;
mov.b32 %r880, %f214;
shfl.sync.bfly.b32 %r881|%p53, %r880, %r868, %r870, %r872;
mov.b32 %f215, %r881;
add.ftz.f32 %f216, %f214, %f215;
mov.b32 %r882, %f216;
mov.u32 %r883, 1;
shfl.sync.bfly.b32 %r884|%p54, %r882, %r883, %r870, %r872;
mov.b32 %f217, %r884;
add.ftz.f32 %f49, %f216, %f217;
@%p13 bra $L__BB302_20;

cvt.u64.u32 %rd89, %r29;
add.s64 %rd90, %rd89, %rd12;
shl.b64 %rd92, %rd90, 2;
add.s64 %rd93, %rd87, %rd92;
st.global.f32 [%rd93], %f49;

$L__BB302_20:
ld.shared.f32 %f218, [%r721+896];
add.ftz.f32 %f219, %f218, %f315;
mov.b32 %r892, %f219;
shfl.sync.bfly.b32 %r896|%p58, %r892, %r871, %r870, %r872;
mov.b32 %f220, %r896;
add.ftz.f32 %f221, %f219, %f220;
mov.b32 %r897, %f221;
shfl.sync.bfly.b32 %r899|%p59, %r897, %r875, %r870, %r872;
mov.b32 %f222, %r899;
add.ftz.f32 %f223, %f221, %f222;
mov.b32 %r900, %f223;
shfl.sync.bfly.b32 %r902|%p60, %r900, %r878, %r870, %r872;
mov.b32 %f224, %r902;
add.ftz.f32 %f225, %f223, %f224;
mov.b32 %r903, %f225;
shfl.sync.bfly.b32 %r904|%p61, %r903, %r868, %r870, %r872;
mov.b32 %f226, %r904;
add.ftz.f32 %f227, %f225, %f226;
mov.b32 %r905, %f227;
shfl.sync.bfly.b32 %r907|%p62, %r905, %r883, %r870, %r872;
mov.b32 %f228, %r907;
add.ftz.f32 %f50, %f227, %f228;
@%p21 bra $L__BB302_22;

add.s32 %r908, %r29, 1;
cvt.u64.u32 %rd94, %r908;
add.s64 %rd95, %rd94, %rd12;
shl.b64 %rd97, %rd95, 2;
add.s64 %rd98, %rd87, %rd97;
st.global.f32 [%rd98], %f50;

$L__BB302_22:
shl.b32 %r909, %r34, 2;
mov.u32 %r910, 2;
ld.shared.f32 %f229, [%r721+1024];
add.ftz.f32 %f230, %f229, %f314;
mov.b32 %r915, %f230;
mov.u32 %r916, 31;
mov.u32 %r917, 16;
mov.u32 %r918, -1;
shfl.sync.bfly.b32 %r919|%p64, %r915, %r917, %r916, %r918;
mov.b32 %f231, %r919;
add.ftz.f32 %f232, %f230, %f231;
mov.b32 %r920, %f232;
mov.u32 %r921, 8;
shfl.sync.bfly.b32 %r922|%p65, %r920, %r921, %r916, %r918;
mov.b32 %f233, %r922;
add.ftz.f32 %f234, %f232, %f233;
mov.b32 %r923, %f234;
mov.u32 %r924, 4;
shfl.sync.bfly.b32 %r925|%p66, %r923, %r924, %r916, %r918;
mov.b32 %f235, %r925;
add.ftz.f32 %f236, %f234, %f235;
mov.b32 %r926, %f236;
shfl.sync.bfly.b32 %r927|%p67, %r926, %r910, %r916, %r918;
mov.b32 %f237, %r927;
add.ftz.f32 %f238, %f236, %f237;
mov.b32 %r928, %f238;
mov.u32 %r929, 1;
shfl.sync.bfly.b32 %r930|%p68, %r928, %r929, %r916, %r918;
mov.b32 %f239, %r930;
add.ftz.f32 %f51, %f238, %f239;
cvt.u64.u32 %rd99, %r909;
add.s64 %rd100, %rd99, %rd12;
shl.b64 %rd102, %rd100, 2;
add.s64 %rd14, %rd87, %rd102;
@%p13 bra $L__BB302_24;

st.global.f32 [%rd14], %f51;

$L__BB302_24:
ld.shared.f32 %f240, [%r721+1152];
add.ftz.f32 %f241, %f240, %f313;
mov.b32 %r938, %f241;
shfl.sync.bfly.b32 %r942|%p72, %r938, %r917, %r916, %r918;
mov.b32 %f242, %r942;
add.ftz.f32 %f243, %f241, %f242;
mov.b32 %r943, %f243;
shfl.sync.bfly.b32 %r945|%p73, %r943, %r921, %r916, %r918;
mov.b32 %f244, %r945;
add.ftz.f32 %f245, %f243, %f244;
mov.b32 %r946, %f245;
shfl.sync.bfly.b32 %r948|%p74, %r946, %r924, %r916, %r918;
mov.b32 %f246, %r948;
add.ftz.f32 %f247, %f245, %f246;
mov.b32 %r949, %f247;
shfl.sync.bfly.b32 %r950|%p75, %r949, %r910, %r916, %r918;
mov.b32 %f248, %r950;
add.ftz.f32 %f249, %f247, %f248;
mov.b32 %r951, %f249;
shfl.sync.bfly.b32 %r953|%p76, %r951, %r929, %r916, %r918;
mov.b32 %f250, %r953;
add.ftz.f32 %f52, %f249, %f250;
@%p21 bra $L__BB302_26;

st.global.f32 [%rd14+4], %f52;

$L__BB302_26:
mul.lo.s32 %r30, %r34, 5;
mov.u32 %r958, 2;
ld.shared.f32 %f251, [%r721+1280];
add.ftz.f32 %f252, %f251, %f312;
mov.b32 %r959, %f252;
mov.u32 %r960, 31;
mov.u32 %r961, 16;
mov.u32 %r962, -1;
shfl.sync.bfly.b32 %r963|%p78, %r959, %r961, %r960, %r962;
mov.b32 %f253, %r963;
add.ftz.f32 %f254, %f252, %f253;
mov.b32 %r964, %f254;
mov.u32 %r965, 8;
shfl.sync.bfly.b32 %r966|%p79, %r964, %r965, %r960, %r962;
mov.b32 %f255, %r966;
add.ftz.f32 %f256, %f254, %f255;
mov.b32 %r967, %f256;
mov.u32 %r968, 4;
shfl.sync.bfly.b32 %r969|%p80, %r967, %r968, %r960, %r962;
mov.b32 %f257, %r969;
add.ftz.f32 %f258, %f256, %f257;
mov.b32 %r970, %f258;
shfl.sync.bfly.b32 %r971|%p81, %r970, %r958, %r960, %r962;
mov.b32 %f259, %r971;
add.ftz.f32 %f260, %f258, %f259;
mov.b32 %r972, %f260;
mov.u32 %r973, 1;
shfl.sync.bfly.b32 %r974|%p82, %r972, %r973, %r960, %r962;
mov.b32 %f261, %r974;
add.ftz.f32 %f53, %f260, %f261;
@%p13 bra $L__BB302_28;

cvt.u64.u32 %rd103, %r30;
add.s64 %rd104, %rd103, %rd12;
shl.b64 %rd106, %rd104, 2;
add.s64 %rd107, %rd87, %rd106;
st.global.f32 [%rd107], %f53;

$L__BB302_28:
ld.shared.f32 %f262, [%r721+1408];
add.ftz.f32 %f263, %f262, %f311;
mov.b32 %r982, %f263;
shfl.sync.bfly.b32 %r986|%p86, %r982, %r961, %r960, %r962;
mov.b32 %f264, %r986;
add.ftz.f32 %f265, %f263, %f264;
mov.b32 %r987, %f265;
shfl.sync.bfly.b32 %r989|%p87, %r987, %r965, %r960, %r962;
mov.b32 %f266, %r989;
add.ftz.f32 %f267, %f265, %f266;
mov.b32 %r990, %f267;
shfl.sync.bfly.b32 %r992|%p88, %r990, %r968, %r960, %r962;
mov.b32 %f268, %r992;
add.ftz.f32 %f269, %f267, %f268;
mov.b32 %r993, %f269;
shfl.sync.bfly.b32 %r994|%p89, %r993, %r958, %r960, %r962;
mov.b32 %f270, %r994;
add.ftz.f32 %f271, %f269, %f270;
mov.b32 %r995, %f271;
shfl.sync.bfly.b32 %r997|%p90, %r995, %r973, %r960, %r962;
mov.b32 %f272, %r997;
add.ftz.f32 %f54, %f271, %f272;
@%p21 bra $L__BB302_30;

add.s32 %r998, %r30, 1;
cvt.u64.u32 %rd108, %r998;
add.s64 %rd109, %rd108, %rd12;
shl.b64 %rd111, %rd109, 2;
add.s64 %rd112, %rd87, %rd111;
st.global.f32 [%rd112], %f54;

$L__BB302_30:
mul.lo.s32 %r999, %r34, 6;
mov.u32 %r1004, 2;
ld.shared.f32 %f273, [%r721+1536];
add.ftz.f32 %f274, %f273, %f310;
mov.b32 %r1005, %f274;
mov.u32 %r1006, 31;
mov.u32 %r1007, 16;
mov.u32 %r1008, -1;
shfl.sync.bfly.b32 %r1009|%p92, %r1005, %r1007, %r1006, %r1008;
mov.b32 %f275, %r1009;
add.ftz.f32 %f276, %f274, %f275;
mov.b32 %r1010, %f276;
mov.u32 %r1011, 8;
shfl.sync.bfly.b32 %r1012|%p93, %r1010, %r1011, %r1006, %r1008;
mov.b32 %f277, %r1012;
add.ftz.f32 %f278, %f276, %f277;
mov.b32 %r1013, %f278;
mov.u32 %r1014, 4;
shfl.sync.bfly.b32 %r1015|%p94, %r1013, %r1014, %r1006, %r1008;
mov.b32 %f279, %r1015;
add.ftz.f32 %f280, %f278, %f279;
mov.b32 %r1016, %f280;
shfl.sync.bfly.b32 %r1017|%p95, %r1016, %r1004, %r1006, %r1008;
mov.b32 %f281, %r1017;
add.ftz.f32 %f282, %f280, %f281;
mov.b32 %r1018, %f282;
mov.u32 %r1019, 1;
shfl.sync.bfly.b32 %r1020|%p96, %r1018, %r1019, %r1006, %r1008;
mov.b32 %f283, %r1020;
add.ftz.f32 %f55, %f282, %f283;
cvt.u64.u32 %rd113, %r999;
add.s64 %rd114, %rd113, %rd12;
shl.b64 %rd116, %rd114, 2;
add.s64 %rd15, %rd87, %rd116;
@%p13 bra $L__BB302_32;

st.global.f32 [%rd15], %f55;

$L__BB302_32:
ld.shared.f32 %f284, [%r721+1664];
add.ftz.f32 %f285, %f284, %f309;
mov.b32 %r1028, %f285;
shfl.sync.bfly.b32 %r1032|%p100, %r1028, %r1007, %r1006, %r1008;
mov.b32 %f286, %r1032;
add.ftz.f32 %f287, %f285, %f286;
mov.b32 %r1033, %f287;
shfl.sync.bfly.b32 %r1035|%p101, %r1033, %r1011, %r1006, %r1008;
mov.b32 %f288, %r1035;
add.ftz.f32 %f289, %f287, %f288;
mov.b32 %r1036, %f289;
shfl.sync.bfly.b32 %r1038|%p102, %r1036, %r1014, %r1006, %r1008;
mov.b32 %f290, %r1038;
add.ftz.f32 %f291, %f289, %f290;
mov.b32 %r1039, %f291;
shfl.sync.bfly.b32 %r1040|%p103, %r1039, %r1004, %r1006, %r1008;
mov.b32 %f292, %r1040;
add.ftz.f32 %f293, %f291, %f292;
mov.b32 %r1041, %f293;
shfl.sync.bfly.b32 %r1043|%p104, %r1041, %r1019, %r1006, %r1008;
mov.b32 %f294, %r1043;
add.ftz.f32 %f56, %f293, %f294;
@%p21 bra $L__BB302_34;

st.global.f32 [%rd15+4], %f56;

$L__BB302_34:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type23ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type23ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type23ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type23ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type23ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type23ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type23ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type23ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type23ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type23ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<120>;
.reg .b16 %rs<21>;
.reg .f32 %f<369>;
.reg .b32 %r<1229>;
.reg .b64 %rd<143>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type23ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[2048];

ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type23ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type23ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type23ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type23ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type23ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type23ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type23ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type23ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type23ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type23ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r48, [_Z13mul_mat_vec_qIL9ggml_type23ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r49, [_Z13mul_mat_vec_qIL9ggml_type23ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r50, [_Z13mul_mat_vec_qIL9ggml_type23ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r46, [_Z13mul_mat_vec_qIL9ggml_type23ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r45, [_Z13mul_mat_vec_qIL9ggml_type23ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type23ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type23ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r51, %r1, 5;
mov.u32 %r52, %tid.x;
add.s32 %r2, %r51, %r52;
shr.s32 %r53, %r2, 31;
shr.u32 %r54, %r53, 29;
add.s32 %r55, %r2, %r54;
shr.s32 %r1228, %r55, 3;
shr.u32 %r56, %r35, 8;
setp.ge.s32 %p2, %r1228, %r56;
mov.f32 %f353, 0f00000000;
mov.f32 %f354, %f353;
mov.f32 %f355, %f353;
mov.f32 %f356, %f353;
mov.f32 %f357, %f353;
mov.f32 %f358, %f353;
mov.f32 %f359, %f353;
mov.f32 %f360, %f353;
mov.f32 %f361, %f353;
mov.f32 %f362, %f353;
mov.f32 %f363, %f353;
mov.f32 %f364, %f353;
mov.f32 %f365, %f353;
mov.f32 %f366, %f353;
mov.f32 %f367, %f353;
mov.f32 %f368, %f353;
@%p2 bra $L__BB303_3;

mov.u32 %r57, %ctaid.x;
shl.b32 %r58, %r57, 1;
mov.u32 %r59, %ctaid.y;
mov.u32 %r60, %ctaid.z;
mul.lo.s32 %r61, %r59, %r43;
cvt.u64.u32 %rd18, %r61;
mul.lo.s32 %r62, %r60, %r49;
cvt.u64.u32 %rd19, %r62;
add.s64 %rd1, %rd19, %rd18;
mul.hi.u32 %r63, %r60, %r45;
add.s32 %r64, %r63, %r60;
shr.u32 %r65, %r64, %r46;
mul.lo.s32 %r66, %r65, %r48;
mul.hi.u32 %r67, %r59, %r39;
add.s32 %r68, %r67, %r59;
shr.u32 %r69, %r68, %r40;
mul.lo.s32 %r70, %r69, %r42;
cvta.to.global.u64 %rd142, %rd15;
and.b32 %r74, %r55, -8;
sub.s32 %r4, %r2, %r74;
shl.b32 %r75, %r4, 2;
shr.u32 %r76, %r4, 31;
add.s32 %r77, %r4, %r76;
shr.s32 %r78, %r77, 1;
cvt.s64.s32 %rd20, %r78;
and.b32 %r5, %r75, 4;
shl.b32 %r79, %r37, 1;
shl.b32 %r1220, %r1228, 3;
add.s32 %r1227, %r1220, %r79;
mad.lo.s32 %r1226, %r37, 3, %r1220;
shl.b32 %r80, %r37, 2;
add.s32 %r1225, %r1220, %r80;
mad.lo.s32 %r1224, %r37, 5, %r1220;
mad.lo.s32 %r1223, %r37, 6, %r1220;
mad.lo.s32 %r1222, %r37, 7, %r1220;
add.s32 %r1221, %r37, %r1220;
add.s32 %r81, %r1228, %r66;
add.s32 %r82, %r81, %r70;
or.b32 %r83, %r58, 1;
mad.lo.s32 %r84, %r36, %r83, %r82;
mul.wide.s32 %rd3, %r84, 136;
mad.lo.s32 %r85, %r58, %r36, %r1228;
add.s32 %r86, %r85, %r66;
add.s32 %r87, %r86, %r70;
mul.wide.s32 %rd4, %r87, 136;
add.s64 %rd21, %rd4, %rd20;
add.s64 %rd5, %rd21, 4;
add.s64 %rd22, %rd3, %rd20;
add.s64 %rd6, %rd22, 4;
mul.wide.s32 %rd23, %r75, 4;
add.s64 %rd24, %rd4, %rd23;
add.s64 %rd7, %rd24, 20;
add.s64 %rd25, %rd3, %rd23;
add.s64 %rd8, %rd25, 12;
cvta.to.global.u64 %rd9, %rd16;
mov.f32 %f353, 0f00000000;

$L__BB303_2:
cvt.u64.u32 %rd26, %r1220;
add.s64 %rd27, %rd1, %rd26;
cvt.s64.s32 %rd28, %r4;
add.s64 %rd29, %rd27, %rd28;
mul.lo.s64 %rd30, %rd29, 36;
add.s64 %rd31, %rd142, %rd7;
ld.global.nc.u32 %r616, [%rd31+-12];
and.b32 %r617, %r616, -2004318072;
shr.u32 %r618, %r617, 1;
or.b32 %r619, %r618, 839922192;
and.b32 %r620, %r616, 30583;
mov.u32 %r621, -152379953;
mov.u32 %r622, -1079142271;
prmt.b32 %r623, %r622, %r621, %r620;
mov.u32 %r624, 1901675829;
mov.u32 %r625, 639175937;
prmt.b32 %r626, %r625, %r624, %r620;
and.b32 %r627, %r619, 30292;
prmt.b32 %r628, %r623, %r626, %r627;
shr.u32 %r629, %r616, 16;
and.b32 %r630, %r629, 30583;
prmt.b32 %r631, %r622, %r621, %r630;
prmt.b32 %r632, %r625, %r624, %r630;
shr.u32 %r633, %r619, 16;
prmt.b32 %r634, %r631, %r632, %r633;
mov.u32 %r635, 25632;
prmt.b32 %r551, %r628, %r634, %r635;
mov.u32 %r636, 30001;
prmt.b32 %r555, %r628, %r634, %r636;
add.s64 %rd32, %rd9, %rd30;
ld.global.nc.u32 %r90, [%rd32+4];
ld.global.nc.u32 %r94, [%rd32+20];
mov.u32 %r586, 0;

dp4a.s32.s32 %r88, %r551, %r90, %r586;


dp4a.s32.s32 %r92, %r555, %r94, %r88;

ld.global.nc.u32 %r637, [%rd31+-8];
and.b32 %r638, %r637, -2004318072;
shr.u32 %r639, %r638, 1;
or.b32 %r640, %r639, 839922192;
and.b32 %r641, %r637, 30583;
prmt.b32 %r642, %r622, %r621, %r641;
prmt.b32 %r643, %r625, %r624, %r641;
and.b32 %r644, %r640, 30292;
prmt.b32 %r645, %r642, %r643, %r644;
shr.u32 %r646, %r637, 16;
and.b32 %r647, %r646, 30583;
prmt.b32 %r648, %r622, %r621, %r647;
prmt.b32 %r649, %r625, %r624, %r647;
shr.u32 %r650, %r640, 16;
prmt.b32 %r651, %r648, %r649, %r650;
prmt.b32 %r559, %r645, %r651, %r635;
prmt.b32 %r563, %r645, %r651, %r636;
ld.global.nc.u32 %r98, [%rd32+8];
ld.global.nc.u32 %r102, [%rd32+24];

dp4a.s32.s32 %r96, %r559, %r98, %r92;


dp4a.s32.s32 %r100, %r563, %r102, %r96;

ld.global.nc.u32 %r652, [%rd31+-4];
and.b32 %r653, %r652, -2004318072;
shr.u32 %r654, %r653, 1;
or.b32 %r655, %r654, 839922192;
and.b32 %r656, %r652, 30583;
prmt.b32 %r657, %r622, %r621, %r656;
prmt.b32 %r658, %r625, %r624, %r656;
and.b32 %r659, %r655, 30292;
prmt.b32 %r660, %r657, %r658, %r659;
shr.u32 %r661, %r652, 16;
and.b32 %r662, %r661, 30583;
prmt.b32 %r663, %r622, %r621, %r662;
prmt.b32 %r664, %r625, %r624, %r662;
shr.u32 %r665, %r655, 16;
prmt.b32 %r666, %r663, %r664, %r665;
prmt.b32 %r567, %r660, %r666, %r635;
prmt.b32 %r571, %r660, %r666, %r636;
ld.global.nc.u32 %r106, [%rd32+12];
ld.global.nc.u32 %r110, [%rd32+28];

dp4a.s32.s32 %r104, %r567, %r106, %r100;


dp4a.s32.s32 %r108, %r571, %r110, %r104;

ld.global.nc.u32 %r667, [%rd31];
and.b32 %r668, %r667, -2004318072;
shr.u32 %r669, %r668, 1;
or.b32 %r670, %r669, 839922192;
and.b32 %r671, %r667, 30583;
prmt.b32 %r672, %r622, %r621, %r671;
prmt.b32 %r673, %r625, %r624, %r671;
and.b32 %r674, %r670, 30292;
prmt.b32 %r675, %r672, %r673, %r674;
shr.u32 %r676, %r667, 16;
and.b32 %r677, %r676, 30583;
prmt.b32 %r678, %r622, %r621, %r677;
prmt.b32 %r679, %r625, %r624, %r677;
shr.u32 %r680, %r670, 16;
prmt.b32 %r681, %r678, %r679, %r680;
prmt.b32 %r575, %r675, %r681, %r635;
prmt.b32 %r579, %r675, %r681, %r636;
ld.global.nc.u32 %r114, [%rd32+16];
ld.global.nc.u32 %r118, [%rd32+32];

dp4a.s32.s32 %r112, %r575, %r114, %r108;


dp4a.s32.s32 %r116, %r579, %r118, %r112;

add.s64 %rd33, %rd142, %rd5;
ld.global.nc.u8 %rs17, [%rd33];
cvt.u32.u16 %r682, %rs17;
and.b32 %r683, %r682, 255;
shr.u32 %r684, %r683, %r5;
and.b32 %r685, %r684, 15;
add.s64 %rd34, %rd142, %rd4;
ld.global.nc.u16 %rs18, [%rd34+2];
cvt.u32.u16 %r686, %rs18;
shl.b32 %r687, %r4, 1;
shr.u32 %r688, %r686, %r687;
and.b32 %r689, %r688, 3;
bfi.b32 %r690, %r689, %r685, 4, 2;
add.s32 %r691, %r690, -32;
mul.lo.s32 %r692, %r691, %r116;
ld.global.nc.u16 %rs1, [%rd34];

{ cvt.f32.f16 %f97, %rs1;}


ld.global.nc.u32 %r120, [%rd32];

{.reg .f16 low,high;
mov.b32 {low,high},%r120;
cvt.f32.f16 %f98, low;}


mul.ftz.f32 %f129, %f97, %f98;
cvt.rn.f32.s32 %f130, %r692;
fma.rn.ftz.f32 %f368, %f129, %f130, %f368;
add.s64 %rd35, %rd142, %rd8;
ld.global.nc.u32 %r693, [%rd35+-4];
and.b32 %r694, %r693, -2004318072;
shr.u32 %r695, %r694, 1;
or.b32 %r696, %r695, 839922192;
and.b32 %r697, %r693, 30583;
prmt.b32 %r698, %r622, %r621, %r697;
prmt.b32 %r699, %r625, %r624, %r697;
and.b32 %r700, %r696, 30292;
prmt.b32 %r701, %r698, %r699, %r700;
shr.u32 %r702, %r693, 16;
and.b32 %r703, %r702, 30583;
prmt.b32 %r704, %r622, %r621, %r703;
prmt.b32 %r705, %r625, %r624, %r703;
shr.u32 %r706, %r696, 16;
prmt.b32 %r707, %r704, %r705, %r706;
prmt.b32 %r584, %r701, %r707, %r635;
prmt.b32 %r588, %r701, %r707, %r636;

dp4a.s32.s32 %r121, %r584, %r90, %r586;


dp4a.s32.s32 %r125, %r588, %r94, %r121;

ld.global.nc.u32 %r708, [%rd35];
and.b32 %r709, %r708, -2004318072;
shr.u32 %r710, %r709, 1;
or.b32 %r711, %r710, 839922192;
and.b32 %r712, %r708, 30583;
prmt.b32 %r713, %r622, %r621, %r712;
prmt.b32 %r714, %r625, %r624, %r712;
and.b32 %r715, %r711, 30292;
prmt.b32 %r716, %r713, %r714, %r715;
shr.u32 %r717, %r708, 16;
and.b32 %r718, %r717, 30583;
prmt.b32 %r719, %r622, %r621, %r718;
prmt.b32 %r720, %r625, %r624, %r718;
shr.u32 %r721, %r711, 16;
prmt.b32 %r722, %r719, %r720, %r721;
prmt.b32 %r592, %r716, %r722, %r635;
prmt.b32 %r596, %r716, %r722, %r636;

dp4a.s32.s32 %r129, %r592, %r98, %r125;


dp4a.s32.s32 %r133, %r596, %r102, %r129;

ld.global.nc.u32 %r723, [%rd35+4];
and.b32 %r724, %r723, -2004318072;
shr.u32 %r725, %r724, 1;
or.b32 %r726, %r725, 839922192;
and.b32 %r727, %r723, 30583;
prmt.b32 %r728, %r622, %r621, %r727;
prmt.b32 %r729, %r625, %r624, %r727;
and.b32 %r730, %r726, 30292;
prmt.b32 %r731, %r728, %r729, %r730;
shr.u32 %r732, %r723, 16;
and.b32 %r733, %r732, 30583;
prmt.b32 %r734, %r622, %r621, %r733;
prmt.b32 %r735, %r625, %r624, %r733;
shr.u32 %r736, %r726, 16;
prmt.b32 %r737, %r734, %r735, %r736;
prmt.b32 %r600, %r731, %r737, %r635;
prmt.b32 %r604, %r731, %r737, %r636;

dp4a.s32.s32 %r137, %r600, %r106, %r133;


dp4a.s32.s32 %r141, %r604, %r110, %r137;

ld.global.nc.u32 %r738, [%rd35+8];
and.b32 %r739, %r738, -2004318072;
shr.u32 %r740, %r739, 1;
or.b32 %r741, %r740, 839922192;
and.b32 %r742, %r738, 30583;
prmt.b32 %r743, %r622, %r621, %r742;
prmt.b32 %r744, %r625, %r624, %r742;
and.b32 %r745, %r741, 30292;
prmt.b32 %r746, %r743, %r744, %r745;
shr.u32 %r747, %r738, 16;
and.b32 %r748, %r747, 30583;
prmt.b32 %r749, %r622, %r621, %r748;
prmt.b32 %r750, %r625, %r624, %r748;
shr.u32 %r751, %r741, 16;
prmt.b32 %r752, %r749, %r750, %r751;
prmt.b32 %r608, %r746, %r752, %r635;
prmt.b32 %r612, %r746, %r752, %r636;

dp4a.s32.s32 %r145, %r608, %r114, %r141;


dp4a.s32.s32 %r149, %r612, %r118, %r145;

add.s64 %rd36, %rd142, %rd6;
ld.global.nc.u8 %rs19, [%rd36];
cvt.u32.u16 %r753, %rs19;
and.b32 %r754, %r753, 255;
shr.u32 %r755, %r754, %r5;
and.b32 %r756, %r755, 15;
add.s64 %rd37, %rd142, %rd3;
ld.global.nc.u16 %rs20, [%rd37+2];
cvt.u32.u16 %r757, %rs20;
shr.u32 %r758, %r757, %r687;
and.b32 %r759, %r758, 3;
bfi.b32 %r760, %r759, %r756, 4, 2;
add.s32 %r761, %r760, -32;
mul.lo.s32 %r762, %r761, %r149;
ld.global.nc.u16 %rs2, [%rd37];

{ cvt.f32.f16 %f99, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r120;
cvt.f32.f16 %f100, low;}


mul.ftz.f32 %f131, %f99, %f100;
cvt.rn.f32.s32 %f132, %r762;
fma.rn.ftz.f32 %f367, %f131, %f132, %f367;
cvt.u64.u32 %rd38, %r1221;
add.s64 %rd39, %rd1, %rd38;
add.s64 %rd40, %rd39, %rd28;
mul.lo.s64 %rd41, %rd40, 36;
add.s64 %rd42, %rd9, %rd41;
ld.global.nc.u32 %r156, [%rd42+4];
ld.global.nc.u32 %r160, [%rd42+20];

dp4a.s32.s32 %r154, %r551, %r156, %r586;


dp4a.s32.s32 %r158, %r555, %r160, %r154;

ld.global.nc.u32 %r164, [%rd42+8];
ld.global.nc.u32 %r168, [%rd42+24];

dp4a.s32.s32 %r162, %r559, %r164, %r158;


dp4a.s32.s32 %r166, %r563, %r168, %r162;

ld.global.nc.u32 %r172, [%rd42+12];
ld.global.nc.u32 %r176, [%rd42+28];

dp4a.s32.s32 %r170, %r567, %r172, %r166;


dp4a.s32.s32 %r174, %r571, %r176, %r170;

ld.global.nc.u32 %r180, [%rd42+16];
ld.global.nc.u32 %r184, [%rd42+32];

dp4a.s32.s32 %r178, %r575, %r180, %r174;


dp4a.s32.s32 %r182, %r579, %r184, %r178;

mul.lo.s32 %r763, %r691, %r182;

{ cvt.f32.f16 %f101, %rs1;}


ld.global.nc.u32 %r186, [%rd42];

{.reg .f16 low,high;
mov.b32 {low,high},%r186;
cvt.f32.f16 %f102, low;}


mul.ftz.f32 %f133, %f101, %f102;
cvt.rn.f32.s32 %f134, %r763;
fma.rn.ftz.f32 %f366, %f133, %f134, %f366;

dp4a.s32.s32 %r187, %r584, %r156, %r586;


dp4a.s32.s32 %r191, %r588, %r160, %r187;


dp4a.s32.s32 %r195, %r592, %r164, %r191;


dp4a.s32.s32 %r199, %r596, %r168, %r195;


dp4a.s32.s32 %r203, %r600, %r172, %r199;


dp4a.s32.s32 %r207, %r604, %r176, %r203;


dp4a.s32.s32 %r211, %r608, %r180, %r207;


dp4a.s32.s32 %r215, %r612, %r184, %r211;

mul.lo.s32 %r764, %r761, %r215;

{ cvt.f32.f16 %f103, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r186;
cvt.f32.f16 %f104, low;}


mul.ftz.f32 %f135, %f103, %f104;
cvt.rn.f32.s32 %f136, %r764;
fma.rn.ftz.f32 %f365, %f135, %f136, %f365;
cvt.u64.u32 %rd43, %r1227;
add.s64 %rd44, %rd1, %rd43;
add.s64 %rd45, %rd44, %rd28;
mul.lo.s64 %rd46, %rd45, 36;
add.s64 %rd47, %rd9, %rd46;
ld.global.nc.u32 %r222, [%rd47+4];
ld.global.nc.u32 %r226, [%rd47+20];

dp4a.s32.s32 %r220, %r551, %r222, %r586;


dp4a.s32.s32 %r224, %r555, %r226, %r220;

ld.global.nc.u32 %r230, [%rd47+8];
ld.global.nc.u32 %r234, [%rd47+24];

dp4a.s32.s32 %r228, %r559, %r230, %r224;


dp4a.s32.s32 %r232, %r563, %r234, %r228;

ld.global.nc.u32 %r238, [%rd47+12];
ld.global.nc.u32 %r242, [%rd47+28];

dp4a.s32.s32 %r236, %r567, %r238, %r232;


dp4a.s32.s32 %r240, %r571, %r242, %r236;

ld.global.nc.u32 %r246, [%rd47+16];
ld.global.nc.u32 %r250, [%rd47+32];

dp4a.s32.s32 %r244, %r575, %r246, %r240;


dp4a.s32.s32 %r248, %r579, %r250, %r244;

mul.lo.s32 %r765, %r691, %r248;

{ cvt.f32.f16 %f105, %rs1;}


ld.global.nc.u32 %r252, [%rd47];

{.reg .f16 low,high;
mov.b32 {low,high},%r252;
cvt.f32.f16 %f106, low;}


mul.ftz.f32 %f137, %f105, %f106;
cvt.rn.f32.s32 %f138, %r765;
fma.rn.ftz.f32 %f364, %f137, %f138, %f364;

dp4a.s32.s32 %r253, %r584, %r222, %r586;


dp4a.s32.s32 %r257, %r588, %r226, %r253;


dp4a.s32.s32 %r261, %r592, %r230, %r257;


dp4a.s32.s32 %r265, %r596, %r234, %r261;


dp4a.s32.s32 %r269, %r600, %r238, %r265;


dp4a.s32.s32 %r273, %r604, %r242, %r269;


dp4a.s32.s32 %r277, %r608, %r246, %r273;


dp4a.s32.s32 %r281, %r612, %r250, %r277;

mul.lo.s32 %r766, %r761, %r281;

{ cvt.f32.f16 %f107, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r252;
cvt.f32.f16 %f108, low;}


mul.ftz.f32 %f139, %f107, %f108;
cvt.rn.f32.s32 %f140, %r766;
fma.rn.ftz.f32 %f363, %f139, %f140, %f363;
cvt.u64.u32 %rd48, %r1226;
add.s64 %rd49, %rd1, %rd48;
add.s64 %rd50, %rd49, %rd28;
mul.lo.s64 %rd51, %rd50, 36;
add.s64 %rd52, %rd9, %rd51;
ld.global.nc.u32 %r288, [%rd52+4];
ld.global.nc.u32 %r292, [%rd52+20];

dp4a.s32.s32 %r286, %r551, %r288, %r586;


dp4a.s32.s32 %r290, %r555, %r292, %r286;

ld.global.nc.u32 %r296, [%rd52+8];
ld.global.nc.u32 %r300, [%rd52+24];

dp4a.s32.s32 %r294, %r559, %r296, %r290;


dp4a.s32.s32 %r298, %r563, %r300, %r294;

ld.global.nc.u32 %r304, [%rd52+12];
ld.global.nc.u32 %r308, [%rd52+28];

dp4a.s32.s32 %r302, %r567, %r304, %r298;


dp4a.s32.s32 %r306, %r571, %r308, %r302;

ld.global.nc.u32 %r312, [%rd52+16];
ld.global.nc.u32 %r316, [%rd52+32];

dp4a.s32.s32 %r310, %r575, %r312, %r306;


dp4a.s32.s32 %r314, %r579, %r316, %r310;

mul.lo.s32 %r767, %r691, %r314;

{ cvt.f32.f16 %f109, %rs1;}


ld.global.nc.u32 %r318, [%rd52];

{.reg .f16 low,high;
mov.b32 {low,high},%r318;
cvt.f32.f16 %f110, low;}


mul.ftz.f32 %f141, %f109, %f110;
cvt.rn.f32.s32 %f142, %r767;
fma.rn.ftz.f32 %f362, %f141, %f142, %f362;

dp4a.s32.s32 %r319, %r584, %r288, %r586;


dp4a.s32.s32 %r323, %r588, %r292, %r319;


dp4a.s32.s32 %r327, %r592, %r296, %r323;


dp4a.s32.s32 %r331, %r596, %r300, %r327;


dp4a.s32.s32 %r335, %r600, %r304, %r331;


dp4a.s32.s32 %r339, %r604, %r308, %r335;


dp4a.s32.s32 %r343, %r608, %r312, %r339;


dp4a.s32.s32 %r347, %r612, %r316, %r343;

mul.lo.s32 %r768, %r761, %r347;

{ cvt.f32.f16 %f111, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r318;
cvt.f32.f16 %f112, low;}


mul.ftz.f32 %f143, %f111, %f112;
cvt.rn.f32.s32 %f144, %r768;
fma.rn.ftz.f32 %f361, %f143, %f144, %f361;
cvt.u64.u32 %rd53, %r1225;
add.s64 %rd54, %rd1, %rd53;
add.s64 %rd55, %rd54, %rd28;
mul.lo.s64 %rd56, %rd55, 36;
add.s64 %rd57, %rd9, %rd56;
ld.global.nc.u32 %r354, [%rd57+4];
ld.global.nc.u32 %r358, [%rd57+20];

dp4a.s32.s32 %r352, %r551, %r354, %r586;


dp4a.s32.s32 %r356, %r555, %r358, %r352;

ld.global.nc.u32 %r362, [%rd57+8];
ld.global.nc.u32 %r366, [%rd57+24];

dp4a.s32.s32 %r360, %r559, %r362, %r356;


dp4a.s32.s32 %r364, %r563, %r366, %r360;

ld.global.nc.u32 %r370, [%rd57+12];
ld.global.nc.u32 %r374, [%rd57+28];

dp4a.s32.s32 %r368, %r567, %r370, %r364;


dp4a.s32.s32 %r372, %r571, %r374, %r368;

ld.global.nc.u32 %r378, [%rd57+16];
ld.global.nc.u32 %r382, [%rd57+32];

dp4a.s32.s32 %r376, %r575, %r378, %r372;


dp4a.s32.s32 %r380, %r579, %r382, %r376;

mul.lo.s32 %r769, %r691, %r380;

{ cvt.f32.f16 %f113, %rs1;}


ld.global.nc.u32 %r384, [%rd57];

{.reg .f16 low,high;
mov.b32 {low,high},%r384;
cvt.f32.f16 %f114, low;}


mul.ftz.f32 %f145, %f113, %f114;
cvt.rn.f32.s32 %f146, %r769;
fma.rn.ftz.f32 %f360, %f145, %f146, %f360;

dp4a.s32.s32 %r385, %r584, %r354, %r586;


dp4a.s32.s32 %r389, %r588, %r358, %r385;


dp4a.s32.s32 %r393, %r592, %r362, %r389;


dp4a.s32.s32 %r397, %r596, %r366, %r393;


dp4a.s32.s32 %r401, %r600, %r370, %r397;


dp4a.s32.s32 %r405, %r604, %r374, %r401;


dp4a.s32.s32 %r409, %r608, %r378, %r405;


dp4a.s32.s32 %r413, %r612, %r382, %r409;

mul.lo.s32 %r770, %r761, %r413;

{ cvt.f32.f16 %f115, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r384;
cvt.f32.f16 %f116, low;}


mul.ftz.f32 %f147, %f115, %f116;
cvt.rn.f32.s32 %f148, %r770;
fma.rn.ftz.f32 %f359, %f147, %f148, %f359;
cvt.u64.u32 %rd58, %r1224;
add.s64 %rd59, %rd1, %rd58;
add.s64 %rd60, %rd59, %rd28;
mul.lo.s64 %rd61, %rd60, 36;
add.s64 %rd62, %rd9, %rd61;
ld.global.nc.u32 %r420, [%rd62+4];
ld.global.nc.u32 %r424, [%rd62+20];

dp4a.s32.s32 %r418, %r551, %r420, %r586;


dp4a.s32.s32 %r422, %r555, %r424, %r418;

ld.global.nc.u32 %r428, [%rd62+8];
ld.global.nc.u32 %r432, [%rd62+24];

dp4a.s32.s32 %r426, %r559, %r428, %r422;


dp4a.s32.s32 %r430, %r563, %r432, %r426;

ld.global.nc.u32 %r436, [%rd62+12];
ld.global.nc.u32 %r440, [%rd62+28];

dp4a.s32.s32 %r434, %r567, %r436, %r430;


dp4a.s32.s32 %r438, %r571, %r440, %r434;

ld.global.nc.u32 %r444, [%rd62+16];
ld.global.nc.u32 %r448, [%rd62+32];

dp4a.s32.s32 %r442, %r575, %r444, %r438;


dp4a.s32.s32 %r446, %r579, %r448, %r442;

mul.lo.s32 %r771, %r691, %r446;

{ cvt.f32.f16 %f117, %rs1;}


ld.global.nc.u32 %r450, [%rd62];

{.reg .f16 low,high;
mov.b32 {low,high},%r450;
cvt.f32.f16 %f118, low;}


mul.ftz.f32 %f149, %f117, %f118;
cvt.rn.f32.s32 %f150, %r771;
fma.rn.ftz.f32 %f358, %f149, %f150, %f358;

dp4a.s32.s32 %r451, %r584, %r420, %r586;


dp4a.s32.s32 %r455, %r588, %r424, %r451;


dp4a.s32.s32 %r459, %r592, %r428, %r455;


dp4a.s32.s32 %r463, %r596, %r432, %r459;


dp4a.s32.s32 %r467, %r600, %r436, %r463;


dp4a.s32.s32 %r471, %r604, %r440, %r467;


dp4a.s32.s32 %r475, %r608, %r444, %r471;


dp4a.s32.s32 %r479, %r612, %r448, %r475;

mul.lo.s32 %r772, %r761, %r479;

{ cvt.f32.f16 %f119, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r450;
cvt.f32.f16 %f120, low;}


mul.ftz.f32 %f151, %f119, %f120;
cvt.rn.f32.s32 %f152, %r772;
fma.rn.ftz.f32 %f357, %f151, %f152, %f357;
cvt.u64.u32 %rd63, %r1223;
add.s64 %rd64, %rd1, %rd63;
add.s64 %rd65, %rd64, %rd28;
mul.lo.s64 %rd66, %rd65, 36;
add.s64 %rd67, %rd9, %rd66;
ld.global.nc.u32 %r486, [%rd67+4];
ld.global.nc.u32 %r490, [%rd67+20];

dp4a.s32.s32 %r484, %r551, %r486, %r586;


dp4a.s32.s32 %r488, %r555, %r490, %r484;

ld.global.nc.u32 %r494, [%rd67+8];
ld.global.nc.u32 %r498, [%rd67+24];

dp4a.s32.s32 %r492, %r559, %r494, %r488;


dp4a.s32.s32 %r496, %r563, %r498, %r492;

ld.global.nc.u32 %r502, [%rd67+12];
ld.global.nc.u32 %r506, [%rd67+28];

dp4a.s32.s32 %r500, %r567, %r502, %r496;


dp4a.s32.s32 %r504, %r571, %r506, %r500;

ld.global.nc.u32 %r510, [%rd67+16];
ld.global.nc.u32 %r514, [%rd67+32];

dp4a.s32.s32 %r508, %r575, %r510, %r504;


dp4a.s32.s32 %r512, %r579, %r514, %r508;

mul.lo.s32 %r773, %r691, %r512;

{ cvt.f32.f16 %f121, %rs1;}


ld.global.nc.u32 %r516, [%rd67];

{.reg .f16 low,high;
mov.b32 {low,high},%r516;
cvt.f32.f16 %f122, low;}


mul.ftz.f32 %f153, %f121, %f122;
cvt.rn.f32.s32 %f154, %r773;
fma.rn.ftz.f32 %f356, %f153, %f154, %f356;

dp4a.s32.s32 %r517, %r584, %r486, %r586;


dp4a.s32.s32 %r521, %r588, %r490, %r517;


dp4a.s32.s32 %r525, %r592, %r494, %r521;


dp4a.s32.s32 %r529, %r596, %r498, %r525;


dp4a.s32.s32 %r533, %r600, %r502, %r529;


dp4a.s32.s32 %r537, %r604, %r506, %r533;


dp4a.s32.s32 %r541, %r608, %r510, %r537;


dp4a.s32.s32 %r545, %r612, %r514, %r541;

mul.lo.s32 %r774, %r761, %r545;

{ cvt.f32.f16 %f123, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r516;
cvt.f32.f16 %f124, low;}


mul.ftz.f32 %f155, %f123, %f124;
cvt.rn.f32.s32 %f156, %r774;
fma.rn.ftz.f32 %f355, %f155, %f156, %f355;
cvt.u64.u32 %rd68, %r1222;
add.s64 %rd69, %rd1, %rd68;
add.s64 %rd70, %rd69, %rd28;
mul.lo.s64 %rd71, %rd70, 36;
add.s64 %rd72, %rd9, %rd71;
ld.global.nc.u32 %r552, [%rd72+4];
ld.global.nc.u32 %r556, [%rd72+20];

dp4a.s32.s32 %r550, %r551, %r552, %r586;


dp4a.s32.s32 %r554, %r555, %r556, %r550;

ld.global.nc.u32 %r560, [%rd72+8];
ld.global.nc.u32 %r564, [%rd72+24];

dp4a.s32.s32 %r558, %r559, %r560, %r554;


dp4a.s32.s32 %r562, %r563, %r564, %r558;

ld.global.nc.u32 %r568, [%rd72+12];
ld.global.nc.u32 %r572, [%rd72+28];

dp4a.s32.s32 %r566, %r567, %r568, %r562;


dp4a.s32.s32 %r570, %r571, %r572, %r566;

ld.global.nc.u32 %r576, [%rd72+16];
ld.global.nc.u32 %r580, [%rd72+32];

dp4a.s32.s32 %r574, %r575, %r576, %r570;


dp4a.s32.s32 %r578, %r579, %r580, %r574;

mul.lo.s32 %r775, %r691, %r578;

{ cvt.f32.f16 %f125, %rs1;}


ld.global.nc.u32 %r582, [%rd72];

{.reg .f16 low,high;
mov.b32 {low,high},%r582;
cvt.f32.f16 %f126, low;}


mul.ftz.f32 %f157, %f125, %f126;
cvt.rn.f32.s32 %f158, %r775;
fma.rn.ftz.f32 %f354, %f157, %f158, %f354;

dp4a.s32.s32 %r583, %r584, %r552, %r586;


dp4a.s32.s32 %r587, %r588, %r556, %r583;


dp4a.s32.s32 %r591, %r592, %r560, %r587;


dp4a.s32.s32 %r595, %r596, %r564, %r591;


dp4a.s32.s32 %r599, %r600, %r568, %r595;


dp4a.s32.s32 %r603, %r604, %r572, %r599;


dp4a.s32.s32 %r607, %r608, %r576, %r603;


dp4a.s32.s32 %r611, %r612, %r580, %r607;

mul.lo.s32 %r776, %r761, %r611;

{ cvt.f32.f16 %f127, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r582;
cvt.f32.f16 %f128, low;}


mul.ftz.f32 %f159, %f127, %f128;
cvt.rn.f32.s32 %f160, %r776;
fma.rn.ftz.f32 %f353, %f159, %f160, %f353;
add.s32 %r1227, %r1227, 64;
add.s32 %r1226, %r1226, 64;
add.s32 %r1225, %r1225, 64;
add.s32 %r1224, %r1224, 64;
add.s32 %r1223, %r1223, 64;
add.s32 %r1222, %r1222, 64;
add.s32 %r1221, %r1221, 64;
add.s32 %r1220, %r1220, 64;
add.s64 %rd142, %rd142, 1088;
add.s32 %r1228, %r1228, 8;
setp.lt.s32 %p3, %r1228, %r56;
@%p3 bra $L__BB303_2;

$L__BB303_3:
setp.eq.s32 %p4, %r1, 0;
@%p4 bra $L__BB303_5;

shl.b32 %r780, %r1, 11;
mov.u32 %r781, _ZZ13mul_mat_vec_qIL9ggml_type23ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r782, %r781, %r780;
shl.b32 %r783, %r52, 2;
add.s32 %r784, %r782, %r783;
st.shared.f32 [%r784+-2048], %f368;
st.shared.f32 [%r784+-1920], %f367;
st.shared.f32 [%r784+-1792], %f366;
st.shared.f32 [%r784+-1664], %f365;
st.shared.f32 [%r784+-1536], %f364;
st.shared.f32 [%r784+-1408], %f363;
st.shared.f32 [%r784+-1280], %f362;
st.shared.f32 [%r784+-1152], %f361;
st.shared.f32 [%r784+-1024], %f360;
st.shared.f32 [%r784+-896], %f359;
st.shared.f32 [%r784+-768], %f358;
st.shared.f32 [%r784+-640], %f357;
st.shared.f32 [%r784+-512], %f356;
st.shared.f32 [%r784+-384], %f355;
st.shared.f32 [%r784+-256], %f354;
st.shared.f32 [%r784+-128], %f353;

$L__BB303_5:
setp.ne.s32 %p5, %r1, 0;
bar.sync 0;
@%p5 bra $L__BB303_38;

mov.u32 %r786, %ctaid.x;
shl.b32 %r787, %r786, 1;
mov.u32 %r788, 1;
shl.b32 %r790, %r52, 2;
mov.u32 %r791, _ZZ13mul_mat_vec_qIL9ggml_type23ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r792, %r791, %r790;
mov.u32 %r793, 2;
ld.shared.f32 %f161, [%r792];
add.ftz.f32 %f162, %f161, %f368;
mov.b32 %r794, %f162;
mov.u32 %r795, 31;
mov.u32 %r796, 16;
mov.u32 %r797, -1;
shfl.sync.bfly.b32 %r798|%p6, %r794, %r796, %r795, %r797;
mov.b32 %f163, %r798;
add.ftz.f32 %f164, %f162, %f163;
mov.b32 %r799, %f164;
mov.u32 %r800, 8;
shfl.sync.bfly.b32 %r801|%p7, %r799, %r800, %r795, %r797;
mov.b32 %f165, %r801;
add.ftz.f32 %f166, %f164, %f165;
mov.b32 %r802, %f166;
mov.u32 %r803, 4;
shfl.sync.bfly.b32 %r804|%p8, %r802, %r803, %r795, %r797;
mov.b32 %f167, %r804;
add.ftz.f32 %f168, %f166, %f167;
mov.b32 %r805, %f168;
shfl.sync.bfly.b32 %r806|%p9, %r805, %r793, %r795, %r797;
mov.b32 %f169, %r806;
add.ftz.f32 %f170, %f168, %f169;
mov.b32 %r807, %f170;
shfl.sync.bfly.b32 %r808|%p10, %r807, %r788, %r795, %r797;
mov.b32 %f171, %r808;
add.ftz.f32 %f49, %f170, %f171;
setp.ne.s32 %p11, %r52, 0;
setp.ge.u32 %p12, %r787, %r38;
or.pred %p13, %p11, %p12;
@%p13 bra $L__BB303_8;

cvta.to.global.u64 %rd73, %rd17;
mov.u32 %r809, %ctaid.y;
mad.lo.s32 %r812, %r809, %r44, %r787;
mov.u32 %r813, %ctaid.z;
mad.lo.s32 %r814, %r813, %r50, %r812;
mul.wide.u32 %rd74, %r814, 4;
add.s64 %rd75, %rd73, %rd74;
st.global.f32 [%rd75], %f49;

$L__BB303_8:
ld.shared.f32 %f172, [%r792+128];
add.ftz.f32 %f173, %f172, %f367;
mov.b32 %r820, %f173;
shfl.sync.bfly.b32 %r824|%p14, %r820, %r796, %r795, %r797;
mov.b32 %f174, %r824;
add.ftz.f32 %f175, %f173, %f174;
mov.b32 %r825, %f175;
shfl.sync.bfly.b32 %r827|%p15, %r825, %r800, %r795, %r797;
mov.b32 %f176, %r827;
add.ftz.f32 %f177, %f175, %f176;
mov.b32 %r828, %f177;
shfl.sync.bfly.b32 %r830|%p16, %r828, %r803, %r795, %r797;
mov.b32 %f178, %r830;
add.ftz.f32 %f179, %f177, %f178;
mov.b32 %r831, %f179;
shfl.sync.bfly.b32 %r832|%p17, %r831, %r793, %r795, %r797;
mov.b32 %f180, %r832;
add.ftz.f32 %f181, %f179, %f180;
mov.b32 %r833, %f181;
shfl.sync.bfly.b32 %r835|%p18, %r833, %r788, %r795, %r797;
mov.b32 %f182, %r835;
add.ftz.f32 %f50, %f181, %f182;
setp.eq.s32 %p19, %r52, 1;
or.b32 %r838, %r787, 1;
setp.lt.u32 %p20, %r838, %r38;
and.pred %p1, %p19, %p20;
not.pred %p21, %p1;
@%p21 bra $L__BB303_10;

cvta.to.global.u64 %rd76, %rd17;
mov.u32 %r839, %ctaid.y;
mad.lo.s32 %r842, %r839, %r44, %r787;
mov.u32 %r843, %ctaid.z;
mad.lo.s32 %r844, %r843, %r50, %r842;
mul.wide.u32 %rd77, %r844, 4;
add.s64 %rd78, %rd76, %rd77;
st.global.f32 [%rd78+4], %f50;

$L__BB303_10:
mov.u32 %r849, 2;
ld.shared.f32 %f183, [%r792+256];
add.ftz.f32 %f184, %f183, %f366;
mov.b32 %r850, %f184;
mov.u32 %r851, 31;
mov.u32 %r852, 16;
mov.u32 %r853, -1;
shfl.sync.bfly.b32 %r854|%p22, %r850, %r852, %r851, %r853;
mov.b32 %f185, %r854;
add.ftz.f32 %f186, %f184, %f185;
mov.b32 %r855, %f186;
mov.u32 %r856, 8;
shfl.sync.bfly.b32 %r857|%p23, %r855, %r856, %r851, %r853;
mov.b32 %f187, %r857;
add.ftz.f32 %f188, %f186, %f187;
mov.b32 %r858, %f188;
mov.u32 %r859, 4;
shfl.sync.bfly.b32 %r860|%p24, %r858, %r859, %r851, %r853;
mov.b32 %f189, %r860;
add.ftz.f32 %f190, %f188, %f189;
mov.b32 %r861, %f190;
shfl.sync.bfly.b32 %r862|%p25, %r861, %r849, %r851, %r853;
mov.b32 %f191, %r862;
add.ftz.f32 %f192, %f190, %f191;
mov.b32 %r863, %f192;
mov.u32 %r864, 1;
shfl.sync.bfly.b32 %r865|%p26, %r863, %r864, %r851, %r853;
mov.b32 %f193, %r865;
add.ftz.f32 %f51, %f192, %f193;
@%p13 bra $L__BB303_12;

cvt.u64.u32 %rd79, %r38;
mov.u32 %r868, %ctaid.y;
mad.lo.s32 %r871, %r868, %r44, %r787;
mov.u32 %r872, %ctaid.z;
mad.lo.s32 %r873, %r872, %r50, %r871;
cvt.u64.u32 %rd80, %r873;
add.s64 %rd81, %rd79, %rd80;
cvta.to.global.u64 %rd82, %rd17;
shl.b64 %rd83, %rd81, 2;
add.s64 %rd84, %rd82, %rd83;
st.global.f32 [%rd84], %f51;

$L__BB303_12:
ld.shared.f32 %f194, [%r792+384];
add.ftz.f32 %f195, %f194, %f365;
mov.b32 %r879, %f195;
shfl.sync.bfly.b32 %r883|%p30, %r879, %r852, %r851, %r853;
mov.b32 %f196, %r883;
add.ftz.f32 %f197, %f195, %f196;
mov.b32 %r884, %f197;
shfl.sync.bfly.b32 %r886|%p31, %r884, %r856, %r851, %r853;
mov.b32 %f198, %r886;
add.ftz.f32 %f199, %f197, %f198;
mov.b32 %r887, %f199;
shfl.sync.bfly.b32 %r889|%p32, %r887, %r859, %r851, %r853;
mov.b32 %f200, %r889;
add.ftz.f32 %f201, %f199, %f200;
mov.b32 %r890, %f201;
shfl.sync.bfly.b32 %r891|%p33, %r890, %r849, %r851, %r853;
mov.b32 %f202, %r891;
add.ftz.f32 %f203, %f201, %f202;
mov.b32 %r892, %f203;
shfl.sync.bfly.b32 %r894|%p34, %r892, %r864, %r851, %r853;
mov.b32 %f204, %r894;
add.ftz.f32 %f52, %f203, %f204;
@%p21 bra $L__BB303_14;

add.s32 %r895, %r38, 1;
cvt.u64.u32 %rd85, %r895;
mov.u32 %r896, %ctaid.y;
mad.lo.s32 %r899, %r896, %r44, %r787;
mov.u32 %r900, %ctaid.z;
mad.lo.s32 %r901, %r900, %r50, %r899;
cvt.u64.u32 %rd86, %r901;
add.s64 %rd87, %rd85, %rd86;
cvta.to.global.u64 %rd88, %rd17;
shl.b64 %rd89, %rd87, 2;
add.s64 %rd90, %rd88, %rd89;
st.global.f32 [%rd90], %f52;

$L__BB303_14:
shl.b32 %r902, %r38, 1;
mov.u32 %r903, 1;
mov.u32 %r908, 2;
ld.shared.f32 %f205, [%r792+512];
add.ftz.f32 %f206, %f205, %f364;
mov.b32 %r909, %f206;
mov.u32 %r910, 31;
mov.u32 %r911, 16;
mov.u32 %r912, -1;
shfl.sync.bfly.b32 %r913|%p36, %r909, %r911, %r910, %r912;
mov.b32 %f207, %r913;
add.ftz.f32 %f208, %f206, %f207;
mov.b32 %r914, %f208;
mov.u32 %r915, 8;
shfl.sync.bfly.b32 %r916|%p37, %r914, %r915, %r910, %r912;
mov.b32 %f209, %r916;
add.ftz.f32 %f210, %f208, %f209;
mov.b32 %r917, %f210;
mov.u32 %r918, 4;
shfl.sync.bfly.b32 %r919|%p38, %r917, %r918, %r910, %r912;
mov.b32 %f211, %r919;
add.ftz.f32 %f212, %f210, %f211;
mov.b32 %r920, %f212;
shfl.sync.bfly.b32 %r921|%p39, %r920, %r908, %r910, %r912;
mov.b32 %f213, %r921;
add.ftz.f32 %f214, %f212, %f213;
mov.b32 %r922, %f214;
shfl.sync.bfly.b32 %r923|%p40, %r922, %r903, %r910, %r912;
mov.b32 %f215, %r923;
add.ftz.f32 %f53, %f214, %f215;
cvt.u64.u32 %rd91, %r902;
mov.u32 %r924, %ctaid.y;
mad.lo.s32 %r927, %r924, %r44, %r787;
mov.u32 %r928, %ctaid.z;
mad.lo.s32 %r929, %r928, %r50, %r927;
cvt.u64.u32 %rd92, %r929;
add.s64 %rd93, %rd91, %rd92;
cvta.to.global.u64 %rd94, %rd17;
shl.b64 %rd95, %rd93, 2;
add.s64 %rd12, %rd94, %rd95;
@%p13 bra $L__BB303_16;

st.global.f32 [%rd12], %f53;

$L__BB303_16:
ld.shared.f32 %f216, [%r792+640];
add.ftz.f32 %f217, %f216, %f363;
mov.b32 %r935, %f217;
shfl.sync.bfly.b32 %r939|%p44, %r935, %r911, %r910, %r912;
mov.b32 %f218, %r939;
add.ftz.f32 %f219, %f217, %f218;
mov.b32 %r940, %f219;
shfl.sync.bfly.b32 %r942|%p45, %r940, %r915, %r910, %r912;
mov.b32 %f220, %r942;
add.ftz.f32 %f221, %f219, %f220;
mov.b32 %r943, %f221;
shfl.sync.bfly.b32 %r945|%p46, %r943, %r918, %r910, %r912;
mov.b32 %f222, %r945;
add.ftz.f32 %f223, %f221, %f222;
mov.b32 %r946, %f223;
shfl.sync.bfly.b32 %r947|%p47, %r946, %r908, %r910, %r912;
mov.b32 %f224, %r947;
add.ftz.f32 %f225, %f223, %f224;
mov.b32 %r948, %f225;
shfl.sync.bfly.b32 %r950|%p48, %r948, %r903, %r910, %r912;
mov.b32 %f226, %r950;
add.ftz.f32 %f54, %f225, %f226;
@%p21 bra $L__BB303_18;

st.global.f32 [%rd12+4], %f54;

$L__BB303_18:
mul.lo.s32 %r32, %r38, 3;
mov.u32 %r955, 2;
ld.shared.f32 %f227, [%r792+768];
add.ftz.f32 %f228, %f227, %f362;
mov.b32 %r956, %f228;
mov.u32 %r957, 31;
mov.u32 %r958, 16;
mov.u32 %r959, -1;
shfl.sync.bfly.b32 %r960|%p50, %r956, %r958, %r957, %r959;
mov.b32 %f229, %r960;
add.ftz.f32 %f230, %f228, %f229;
mov.b32 %r961, %f230;
mov.u32 %r962, 8;
shfl.sync.bfly.b32 %r963|%p51, %r961, %r962, %r957, %r959;
mov.b32 %f231, %r963;
add.ftz.f32 %f232, %f230, %f231;
mov.b32 %r964, %f232;
mov.u32 %r965, 4;
shfl.sync.bfly.b32 %r966|%p52, %r964, %r965, %r957, %r959;
mov.b32 %f233, %r966;
add.ftz.f32 %f234, %f232, %f233;
mov.b32 %r967, %f234;
shfl.sync.bfly.b32 %r968|%p53, %r967, %r955, %r957, %r959;
mov.b32 %f235, %r968;
add.ftz.f32 %f236, %f234, %f235;
mov.b32 %r969, %f236;
mov.u32 %r970, 1;
shfl.sync.bfly.b32 %r971|%p54, %r969, %r970, %r957, %r959;
mov.b32 %f237, %r971;
add.ftz.f32 %f55, %f236, %f237;
@%p13 bra $L__BB303_20;

cvt.u64.u32 %rd96, %r32;
add.s64 %rd98, %rd96, %rd92;
shl.b64 %rd100, %rd98, 2;
add.s64 %rd101, %rd94, %rd100;
st.global.f32 [%rd101], %f55;

$L__BB303_20:
ld.shared.f32 %f238, [%r792+896];
add.ftz.f32 %f239, %f238, %f361;
mov.b32 %r985, %f239;
shfl.sync.bfly.b32 %r989|%p58, %r985, %r958, %r957, %r959;
mov.b32 %f240, %r989;
add.ftz.f32 %f241, %f239, %f240;
mov.b32 %r990, %f241;
shfl.sync.bfly.b32 %r992|%p59, %r990, %r962, %r957, %r959;
mov.b32 %f242, %r992;
add.ftz.f32 %f243, %f241, %f242;
mov.b32 %r993, %f243;
shfl.sync.bfly.b32 %r995|%p60, %r993, %r965, %r957, %r959;
mov.b32 %f244, %r995;
add.ftz.f32 %f245, %f243, %f244;
mov.b32 %r996, %f245;
shfl.sync.bfly.b32 %r997|%p61, %r996, %r955, %r957, %r959;
mov.b32 %f246, %r997;
add.ftz.f32 %f247, %f245, %f246;
mov.b32 %r998, %f247;
shfl.sync.bfly.b32 %r1000|%p62, %r998, %r970, %r957, %r959;
mov.b32 %f248, %r1000;
add.ftz.f32 %f56, %f247, %f248;
@%p21 bra $L__BB303_22;

add.s32 %r1001, %r32, 1;
cvt.u64.u32 %rd102, %r1001;
add.s64 %rd104, %rd102, %rd92;
shl.b64 %rd106, %rd104, 2;
add.s64 %rd107, %rd94, %rd106;
st.global.f32 [%rd107], %f56;

$L__BB303_22:
shl.b32 %r1008, %r38, 2;
mov.u32 %r1009, 2;
ld.shared.f32 %f249, [%r792+1024];
add.ftz.f32 %f250, %f249, %f360;
mov.b32 %r1014, %f250;
mov.u32 %r1015, 31;
mov.u32 %r1016, 16;
mov.u32 %r1017, -1;
shfl.sync.bfly.b32 %r1018|%p64, %r1014, %r1016, %r1015, %r1017;
mov.b32 %f251, %r1018;
add.ftz.f32 %f252, %f250, %f251;
mov.b32 %r1019, %f252;
mov.u32 %r1020, 8;
shfl.sync.bfly.b32 %r1021|%p65, %r1019, %r1020, %r1015, %r1017;
mov.b32 %f253, %r1021;
add.ftz.f32 %f254, %f252, %f253;
mov.b32 %r1022, %f254;
mov.u32 %r1023, 4;
shfl.sync.bfly.b32 %r1024|%p66, %r1022, %r1023, %r1015, %r1017;
mov.b32 %f255, %r1024;
add.ftz.f32 %f256, %f254, %f255;
mov.b32 %r1025, %f256;
shfl.sync.bfly.b32 %r1026|%p67, %r1025, %r1009, %r1015, %r1017;
mov.b32 %f257, %r1026;
add.ftz.f32 %f258, %f256, %f257;
mov.b32 %r1027, %f258;
mov.u32 %r1028, 1;
shfl.sync.bfly.b32 %r1029|%p68, %r1027, %r1028, %r1015, %r1017;
mov.b32 %f259, %r1029;
add.ftz.f32 %f57, %f258, %f259;
cvt.u64.u32 %rd108, %r1008;
add.s64 %rd110, %rd108, %rd92;
shl.b64 %rd112, %rd110, 2;
add.s64 %rd13, %rd94, %rd112;
@%p13 bra $L__BB303_24;

st.global.f32 [%rd13], %f57;

$L__BB303_24:
ld.shared.f32 %f260, [%r792+1152];
add.ftz.f32 %f261, %f260, %f359;
mov.b32 %r1041, %f261;
shfl.sync.bfly.b32 %r1045|%p72, %r1041, %r1016, %r1015, %r1017;
mov.b32 %f262, %r1045;
add.ftz.f32 %f263, %f261, %f262;
mov.b32 %r1046, %f263;
shfl.sync.bfly.b32 %r1048|%p73, %r1046, %r1020, %r1015, %r1017;
mov.b32 %f264, %r1048;
add.ftz.f32 %f265, %f263, %f264;
mov.b32 %r1049, %f265;
shfl.sync.bfly.b32 %r1051|%p74, %r1049, %r1023, %r1015, %r1017;
mov.b32 %f266, %r1051;
add.ftz.f32 %f267, %f265, %f266;
mov.b32 %r1052, %f267;
shfl.sync.bfly.b32 %r1053|%p75, %r1052, %r1009, %r1015, %r1017;
mov.b32 %f268, %r1053;
add.ftz.f32 %f269, %f267, %f268;
mov.b32 %r1054, %f269;
shfl.sync.bfly.b32 %r1056|%p76, %r1054, %r1028, %r1015, %r1017;
mov.b32 %f270, %r1056;
add.ftz.f32 %f58, %f269, %f270;
@%p21 bra $L__BB303_26;

st.global.f32 [%rd13+4], %f58;

$L__BB303_26:
mul.lo.s32 %r33, %r38, 5;
mov.u32 %r1061, 2;
ld.shared.f32 %f271, [%r792+1280];
add.ftz.f32 %f272, %f271, %f358;
mov.b32 %r1062, %f272;
mov.u32 %r1063, 31;
mov.u32 %r1064, 16;
mov.u32 %r1065, -1;
shfl.sync.bfly.b32 %r1066|%p78, %r1062, %r1064, %r1063, %r1065;
mov.b32 %f273, %r1066;
add.ftz.f32 %f274, %f272, %f273;
mov.b32 %r1067, %f274;
mov.u32 %r1068, 8;
shfl.sync.bfly.b32 %r1069|%p79, %r1067, %r1068, %r1063, %r1065;
mov.b32 %f275, %r1069;
add.ftz.f32 %f276, %f274, %f275;
mov.b32 %r1070, %f276;
mov.u32 %r1071, 4;
shfl.sync.bfly.b32 %r1072|%p80, %r1070, %r1071, %r1063, %r1065;
mov.b32 %f277, %r1072;
add.ftz.f32 %f278, %f276, %f277;
mov.b32 %r1073, %f278;
shfl.sync.bfly.b32 %r1074|%p81, %r1073, %r1061, %r1063, %r1065;
mov.b32 %f279, %r1074;
add.ftz.f32 %f280, %f278, %f279;
mov.b32 %r1075, %f280;
mov.u32 %r1076, 1;
shfl.sync.bfly.b32 %r1077|%p82, %r1075, %r1076, %r1063, %r1065;
mov.b32 %f281, %r1077;
add.ftz.f32 %f59, %f280, %f281;
@%p13 bra $L__BB303_28;

cvt.u64.u32 %rd113, %r33;
add.s64 %rd115, %rd113, %rd92;
shl.b64 %rd117, %rd115, 2;
add.s64 %rd118, %rd94, %rd117;
st.global.f32 [%rd118], %f59;

$L__BB303_28:
ld.shared.f32 %f282, [%r792+1408];
add.ftz.f32 %f283, %f282, %f357;
mov.b32 %r1091, %f283;
shfl.sync.bfly.b32 %r1095|%p86, %r1091, %r1064, %r1063, %r1065;
mov.b32 %f284, %r1095;
add.ftz.f32 %f285, %f283, %f284;
mov.b32 %r1096, %f285;
shfl.sync.bfly.b32 %r1098|%p87, %r1096, %r1068, %r1063, %r1065;
mov.b32 %f286, %r1098;
add.ftz.f32 %f287, %f285, %f286;
mov.b32 %r1099, %f287;
shfl.sync.bfly.b32 %r1101|%p88, %r1099, %r1071, %r1063, %r1065;
mov.b32 %f288, %r1101;
add.ftz.f32 %f289, %f287, %f288;
mov.b32 %r1102, %f289;
shfl.sync.bfly.b32 %r1103|%p89, %r1102, %r1061, %r1063, %r1065;
mov.b32 %f290, %r1103;
add.ftz.f32 %f291, %f289, %f290;
mov.b32 %r1104, %f291;
shfl.sync.bfly.b32 %r1106|%p90, %r1104, %r1076, %r1063, %r1065;
mov.b32 %f292, %r1106;
add.ftz.f32 %f60, %f291, %f292;
@%p21 bra $L__BB303_30;

add.s32 %r1107, %r33, 1;
cvt.u64.u32 %rd119, %r1107;
add.s64 %rd121, %rd119, %rd92;
shl.b64 %rd123, %rd121, 2;
add.s64 %rd124, %rd94, %rd123;
st.global.f32 [%rd124], %f60;

$L__BB303_30:
mul.lo.s32 %r1114, %r38, 6;
mov.u32 %r1119, 2;
ld.shared.f32 %f293, [%r792+1536];
add.ftz.f32 %f294, %f293, %f356;
mov.b32 %r1120, %f294;
mov.u32 %r1121, 31;
mov.u32 %r1122, 16;
mov.u32 %r1123, -1;
shfl.sync.bfly.b32 %r1124|%p92, %r1120, %r1122, %r1121, %r1123;
mov.b32 %f295, %r1124;
add.ftz.f32 %f296, %f294, %f295;
mov.b32 %r1125, %f296;
mov.u32 %r1126, 8;
shfl.sync.bfly.b32 %r1127|%p93, %r1125, %r1126, %r1121, %r1123;
mov.b32 %f297, %r1127;
add.ftz.f32 %f298, %f296, %f297;
mov.b32 %r1128, %f298;
mov.u32 %r1129, 4;
shfl.sync.bfly.b32 %r1130|%p94, %r1128, %r1129, %r1121, %r1123;
mov.b32 %f299, %r1130;
add.ftz.f32 %f300, %f298, %f299;
mov.b32 %r1131, %f300;
shfl.sync.bfly.b32 %r1132|%p95, %r1131, %r1119, %r1121, %r1123;
mov.b32 %f301, %r1132;
add.ftz.f32 %f302, %f300, %f301;
mov.b32 %r1133, %f302;
mov.u32 %r1134, 1;
shfl.sync.bfly.b32 %r1135|%p96, %r1133, %r1134, %r1121, %r1123;
mov.b32 %f303, %r1135;
add.ftz.f32 %f61, %f302, %f303;
cvt.u64.u32 %rd125, %r1114;
add.s64 %rd127, %rd125, %rd92;
shl.b64 %rd129, %rd127, 2;
add.s64 %rd14, %rd94, %rd129;
@%p13 bra $L__BB303_32;

st.global.f32 [%rd14], %f61;

$L__BB303_32:
ld.shared.f32 %f304, [%r792+1664];
add.ftz.f32 %f305, %f304, %f355;
mov.b32 %r1147, %f305;
shfl.sync.bfly.b32 %r1151|%p100, %r1147, %r1122, %r1121, %r1123;
mov.b32 %f306, %r1151;
add.ftz.f32 %f307, %f305, %f306;
mov.b32 %r1152, %f307;
shfl.sync.bfly.b32 %r1154|%p101, %r1152, %r1126, %r1121, %r1123;
mov.b32 %f308, %r1154;
add.ftz.f32 %f309, %f307, %f308;
mov.b32 %r1155, %f309;
shfl.sync.bfly.b32 %r1157|%p102, %r1155, %r1129, %r1121, %r1123;
mov.b32 %f310, %r1157;
add.ftz.f32 %f311, %f309, %f310;
mov.b32 %r1158, %f311;
shfl.sync.bfly.b32 %r1159|%p103, %r1158, %r1119, %r1121, %r1123;
mov.b32 %f312, %r1159;
add.ftz.f32 %f313, %f311, %f312;
mov.b32 %r1160, %f313;
shfl.sync.bfly.b32 %r1162|%p104, %r1160, %r1134, %r1121, %r1123;
mov.b32 %f314, %r1162;
add.ftz.f32 %f62, %f313, %f314;
@%p21 bra $L__BB303_34;

st.global.f32 [%rd14+4], %f62;

$L__BB303_34:
mul.lo.s32 %r34, %r38, 7;
mov.u32 %r1167, 2;
ld.shared.f32 %f315, [%r792+1792];
add.ftz.f32 %f316, %f315, %f354;
mov.b32 %r1168, %f316;
mov.u32 %r1169, 31;
mov.u32 %r1170, 16;
mov.u32 %r1171, -1;
shfl.sync.bfly.b32 %r1172|%p106, %r1168, %r1170, %r1169, %r1171;
mov.b32 %f317, %r1172;
add.ftz.f32 %f318, %f316, %f317;
mov.b32 %r1173, %f318;
mov.u32 %r1174, 8;
shfl.sync.bfly.b32 %r1175|%p107, %r1173, %r1174, %r1169, %r1171;
mov.b32 %f319, %r1175;
add.ftz.f32 %f320, %f318, %f319;
mov.b32 %r1176, %f320;
mov.u32 %r1177, 4;
shfl.sync.bfly.b32 %r1178|%p108, %r1176, %r1177, %r1169, %r1171;
mov.b32 %f321, %r1178;
add.ftz.f32 %f322, %f320, %f321;
mov.b32 %r1179, %f322;
shfl.sync.bfly.b32 %r1180|%p109, %r1179, %r1167, %r1169, %r1171;
mov.b32 %f323, %r1180;
add.ftz.f32 %f324, %f322, %f323;
mov.b32 %r1181, %f324;
mov.u32 %r1182, 1;
shfl.sync.bfly.b32 %r1183|%p110, %r1181, %r1182, %r1169, %r1171;
mov.b32 %f325, %r1183;
add.ftz.f32 %f63, %f324, %f325;
@%p13 bra $L__BB303_36;

cvt.u64.u32 %rd130, %r34;
add.s64 %rd132, %rd130, %rd92;
shl.b64 %rd134, %rd132, 2;
add.s64 %rd135, %rd94, %rd134;
st.global.f32 [%rd135], %f63;

$L__BB303_36:
ld.shared.f32 %f326, [%r792+1920];
add.ftz.f32 %f327, %f326, %f353;
mov.b32 %r1197, %f327;
shfl.sync.bfly.b32 %r1201|%p114, %r1197, %r1170, %r1169, %r1171;
mov.b32 %f328, %r1201;
add.ftz.f32 %f329, %f327, %f328;
mov.b32 %r1202, %f329;
shfl.sync.bfly.b32 %r1204|%p115, %r1202, %r1174, %r1169, %r1171;
mov.b32 %f330, %r1204;
add.ftz.f32 %f331, %f329, %f330;
mov.b32 %r1205, %f331;
shfl.sync.bfly.b32 %r1207|%p116, %r1205, %r1177, %r1169, %r1171;
mov.b32 %f332, %r1207;
add.ftz.f32 %f333, %f331, %f332;
mov.b32 %r1208, %f333;
shfl.sync.bfly.b32 %r1209|%p117, %r1208, %r1167, %r1169, %r1171;
mov.b32 %f334, %r1209;
add.ftz.f32 %f335, %f333, %f334;
mov.b32 %r1210, %f335;
shfl.sync.bfly.b32 %r1212|%p118, %r1210, %r1182, %r1169, %r1171;
mov.b32 %f336, %r1212;
add.ftz.f32 %f64, %f335, %f336;
@%p21 bra $L__BB303_38;

add.s32 %r1213, %r34, 1;
cvt.u64.u32 %rd136, %r1213;
add.s64 %rd138, %rd136, %rd92;
shl.b64 %rd140, %rd138, 2;
add.s64 %rd141, %rd94, %rd140;
st.global.f32 [%rd141], %f64;

$L__BB303_38:
ret;

}
.entry _Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj(
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2,
.param .align 8 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3[48],
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5,
.param .align 4 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6[12],
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15
)
.maxntid 256, 1, 1
.minnctapersm 1
{
.local .align 8 .b8 __local_depot304[16];
.reg .b64 %SP;
.reg .b64 %SPL;
.reg .pred %p<48>;
.reg .b16 %rs<89>;
.reg .f32 %f<169>;
.reg .b32 %r<959>;
.reg .b64 %rd<185>;


mov.u64 %SPL, __local_depot304;
ld.param.u64 %rd24, [_Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0];
ld.param.u64 %rd25, [_Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1];
ld.param.u64 %rd26, [_Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2];
ld.param.u32 %r24, [_Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5];
ld.param.u32 %r29, [_Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8];
ld.param.u32 %r30, [_Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9];
ld.param.u32 %r32, [_Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11];
ld.param.u32 %r33, [_Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12];
ld.param.u32 %r36, [_Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14];
ld.param.u32 %r35, [_Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15];
ld.param.u32 %r27, [_Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+8];
ld.param.u32 %r26, [_Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+4];
ld.param.u32 %r25, [_Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6];
ld.param.u64 %rd28, [_Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+8];
add.u64 %rd1, %SPL, 0;
add.u64 %rd2, %SPL, 8;
mov.u32 %r1, %tid.y;
setp.ge.u32 %p1, %r1, %r36;
@%p1 bra $L__BB304_29;

mov.u32 %r2, %ctaid.y;
cvta.to.global.u64 %rd35, %rd26;
mad.lo.s32 %r37, %r1, %r35, %r2;
mul.wide.u32 %rd36, %r37, 4;
add.s64 %rd37, %rd35, %rd36;
ld.global.nc.u32 %r3, [%rd37];
mov.f32 %f162, 0f00000000;
st.local.v2.f32 [%rd1], {%f162, %f162};
st.local.v2.f32 [%rd2], {%f162, %f162};
mov.u32 %r4, %tid.x;
shr.u32 %r958, %r4, 3;
shr.u32 %r38, %r24, 8;
setp.ge.u32 %p2, %r958, %r38;
mov.f32 %f163, %f162;
mov.f32 %f164, %f162;
mov.f32 %f165, %f162;
@%p2 bra $L__BB304_9;

mov.u32 %r39, %ctaid.x;
shl.b32 %r40, %r39, 1;
mul.lo.s32 %r41, %r3, %r32;
and.b32 %r42, %r4, 7;
shl.b32 %r43, %r4, 1;
and.b32 %r44, %r43, 14;
shl.b32 %r45, %r4, 2;
and.b32 %r46, %r45, 28;
or.b32 %r47, %r46, 1;
sub.s32 %r48, %r47, %r44;
shl.b32 %r49, %r48, 1;
cvt.s64.s32 %rd38, %r49;
cvt.u64.u32 %rd39, %r42;
shr.u32 %r50, %r42, 1;
cvt.u64.u32 %rd40, %r50;
shl.b32 %r957, %r958, 3;
or.b32 %r51, %r40, 1;
add.s32 %r52, %r958, %r41;
mad.lo.s32 %r53, %r29, %r51, %r52;
mul.wide.s32 %rd4, %r53, 110;
add.s64 %rd41, %rd40, %rd4;
add.s64 %rd5, %rd41, 106;
shl.b32 %r54, %r4, 3;
cvt.u64.u32 %rd42, %r54;
and.b64 %rd43, %rd42, 56;
mad.lo.s32 %r55, %r40, %r29, %r958;
add.s32 %r56, %r55, %r41;
mul.wide.s32 %rd6, %r56, 110;
add.s64 %rd7, %rd43, %rd6;
add.s64 %rd44, %rd38, -2;
add.s64 %rd45, %rd44, %rd4;
add.s64 %rd8, %rd45, 74;
add.s64 %rd46, %rd4, %rd39;
add.s64 %rd9, %rd46, 66;
add.s64 %rd47, %rd6, %rd39;
add.s64 %rd10, %rd47, 66;
add.s64 %rd48, %rd44, %rd6;
add.s64 %rd11, %rd48, 76;
add.s64 %rd49, %rd43, %rd4;
add.s64 %rd12, %rd49, 4;
add.s64 %rd50, %rd40, %rd6;
add.s64 %rd13, %rd50, 106;
cvta.to.global.u64 %rd183, %rd24;
cvta.to.global.u64 %rd184, %rd28;
mul.lo.s32 %r57, %r1, %r30;
cvt.u64.u32 %rd51, %r57;
mul.hi.u32 %r58, %r2, %r25;
add.s32 %r59, %r58, %r2;
shr.u32 %r60, %r59, %r26;
mul.lo.s32 %r61, %r60, %r27;
sub.s32 %r62, %r2, %r61;
mul.lo.s32 %r63, %r62, %r33;
cvt.u64.u32 %rd52, %r63;
add.s64 %rd16, %rd52, %rd51;
mov.f32 %f162, 0f00000000;
cvta.to.global.u64 %rd53, %rd25;

$L__BB304_3:
cvt.s64.s32 %rd54, %r957;
add.s64 %rd55, %rd16, %rd54;
add.s64 %rd56, %rd183, %rd7;
ld.global.nc.u16 %rs2, [%rd56+4];
ld.global.nc.u16 %rs3, [%rd56+6];
ld.global.nc.u16 %rs4, [%rd56+8];
add.s64 %rd57, %rd183, %rd10;
ld.global.nc.u8 %rs5, [%rd57];
add.s64 %rd58, %rd183, %rd11;
ld.global.nc.u16 %rs6, [%rd58+-2];
ld.global.nc.u16 %rs7, [%rd58];
add.s64 %rd60, %rd55, %rd39;
mul.lo.s64 %rd61, %rd60, 36;
ld.global.nc.u16 %rs8, [%rd56+2];
and.b16 %rs9, %rs8, 255;
cvt.u32.u16 %r115, %rs9;
shl.b16 %rs10, %rs5, 8;
cvt.u32.u16 %r116, %rs10;
and.b32 %r117, %r116, 256;
or.b32 %r118, %r117, %r115;
mul.wide.u32 %rd62, %r118, 4;
mov.u64 %rd63, iq3s_grid;
add.s64 %rd64, %rd63, %rd62;
ld.global.nc.u32 %r119, [%rd64];
shr.u16 %rs11, %rs8, 8;
cvt.u32.u16 %r120, %rs11;
mul.wide.u16 %r121, %rs5, 128;
and.b32 %r122, %r121, 256;
or.b32 %r123, %r122, %r120;
mul.wide.u32 %rd65, %r123, 4;
add.s64 %rd66, %rd63, %rd65;
ld.global.nc.u32 %r124, [%rd66];
and.b16 %rs13, %rs6, 255;
mul.wide.u16 %r125, %rs13, 128;
cvt.u32.u16 %r126, %rs6;
shl.b32 %r127, %r126, 21;
or.b32 %r128, %r125, %r127;
and.b32 %r129, %r128, 25166208;
or.b32 %r130, %r129, -2139062144;
add.s32 %r131, %r130, -16843009;
or.b32 %r64, %r131, %r129;

prmt.b32 %r64, %r64, 0, 0xba98;

mul.wide.u16 %r132, %rs13, 8;
shl.b32 %r133, %r126, 17;
or.b32 %r134, %r132, %r133;
and.b32 %r135, %r134, 25166208;
or.b32 %r136, %r135, -2139062144;
add.s32 %r137, %r136, -16843009;
or.b32 %r66, %r137, %r135;

prmt.b32 %r66, %r66, 0, 0xba98;

xor.b32 %r138, %r64, %r119;
or.b32 %r139, %r138, -2139062144;
and.b32 %r140, %r64, 2139062143;
and.b32 %r141, %r119, -2139062144;
sub.s32 %r142, %r139, %r140;
xor.b32 %r143, %r141, %r142;
xor.b32 %r69, %r143, -2139062144;
xor.b32 %r144, %r66, %r124;
or.b32 %r145, %r144, -2139062144;
and.b32 %r146, %r66, 2139062143;
and.b32 %r147, %r124, -2139062144;
sub.s32 %r148, %r145, %r146;
xor.b32 %r149, %r147, %r148;
xor.b32 %r73, %r149, -2139062144;
add.s64 %rd67, %rd53, %rd61;
ld.global.nc.u32 %r9, [%rd67+4];
ld.global.nc.u32 %r10, [%rd67+8];
mov.u32 %r71, 0;

dp4a.s32.s32 %r68, %r69, %r9, %r71;


dp4a.s32.s32 %r72, %r73, %r10, %r68;

and.b16 %rs14, %rs2, 255;
cvt.u32.u16 %r150, %rs14;
mul.wide.u16 %r151, %rs5, 64;
and.b32 %r152, %r151, 256;
or.b32 %r153, %r152, %r150;
mul.wide.u32 %rd68, %r153, 4;
add.s64 %rd69, %rd63, %rd68;
ld.global.nc.u32 %r154, [%rd69];
shr.u16 %rs15, %rs2, 8;
cvt.u32.u16 %r155, %rs15;
mul.wide.u16 %r156, %rs5, 32;
and.b32 %r157, %r156, 256;
or.b32 %r158, %r157, %r155;
mul.wide.u32 %rd70, %r158, 4;
add.s64 %rd71, %rd63, %rd70;
ld.global.nc.u32 %r159, [%rd71];
shr.u16 %rs16, %rs6, 8;
mul.wide.u16 %r160, %rs16, 128;
cvt.u32.u16 %r161, %rs16;
shl.b32 %r162, %r161, 21;
or.b32 %r163, %r160, %r162;
and.b32 %r164, %r163, 25166208;
or.b32 %r165, %r164, -2139062144;
add.s32 %r166, %r165, -16843009;
or.b32 %r76, %r166, %r164;

prmt.b32 %r76, %r76, 0, 0xba98;

mul.wide.u16 %r167, %rs16, 8;
shl.b32 %r168, %r161, 17;
or.b32 %r169, %r167, %r168;
and.b32 %r170, %r169, 25166208;
or.b32 %r171, %r170, -2139062144;
add.s32 %r172, %r171, -16843009;
or.b32 %r78, %r172, %r170;

prmt.b32 %r78, %r78, 0, 0xba98;

xor.b32 %r173, %r76, %r154;
or.b32 %r174, %r173, -2139062144;
and.b32 %r175, %r76, 2139062143;
and.b32 %r176, %r154, -2139062144;
sub.s32 %r177, %r174, %r175;
xor.b32 %r178, %r176, %r177;
xor.b32 %r81, %r178, -2139062144;
xor.b32 %r179, %r78, %r159;
or.b32 %r180, %r179, -2139062144;
and.b32 %r181, %r78, 2139062143;
and.b32 %r182, %r159, -2139062144;
sub.s32 %r183, %r180, %r181;
xor.b32 %r184, %r182, %r183;
xor.b32 %r85, %r184, -2139062144;
ld.global.nc.u32 %r11, [%rd67+12];
ld.global.nc.u32 %r12, [%rd67+16];

dp4a.s32.s32 %r80, %r81, %r11, %r72;


dp4a.s32.s32 %r84, %r85, %r12, %r80;

and.b16 %rs17, %rs3, 255;
cvt.u32.u16 %r185, %rs17;
mul.wide.u16 %r186, %rs5, 16;
and.b32 %r187, %r186, 256;
or.b32 %r188, %r187, %r185;
mul.wide.u32 %rd72, %r188, 4;
add.s64 %rd73, %rd63, %rd72;
ld.global.nc.u32 %r189, [%rd73];
shr.u16 %rs18, %rs3, 8;
cvt.u32.u16 %r190, %rs18;
mul.wide.u16 %r191, %rs5, 8;
and.b32 %r192, %r191, 256;
or.b32 %r193, %r192, %r190;
mul.wide.u32 %rd74, %r193, 4;
add.s64 %rd75, %rd63, %rd74;
ld.global.nc.u32 %r194, [%rd75];
and.b16 %rs19, %rs7, 255;
mul.wide.u16 %r195, %rs19, 128;
cvt.u32.u16 %r196, %rs7;
shl.b32 %r197, %r196, 21;
or.b32 %r198, %r195, %r197;
and.b32 %r199, %r198, 25166208;
or.b32 %r200, %r199, -2139062144;
add.s32 %r201, %r200, -16843009;
or.b32 %r88, %r201, %r199;

prmt.b32 %r88, %r88, 0, 0xba98;

mul.wide.u16 %r202, %rs19, 8;
shl.b32 %r203, %r196, 17;
or.b32 %r204, %r202, %r203;
and.b32 %r205, %r204, 25166208;
or.b32 %r206, %r205, -2139062144;
add.s32 %r207, %r206, -16843009;
or.b32 %r90, %r207, %r205;

prmt.b32 %r90, %r90, 0, 0xba98;

xor.b32 %r208, %r88, %r189;
or.b32 %r209, %r208, -2139062144;
and.b32 %r210, %r88, 2139062143;
and.b32 %r211, %r189, -2139062144;
sub.s32 %r212, %r209, %r210;
xor.b32 %r213, %r211, %r212;
xor.b32 %r93, %r213, -2139062144;
xor.b32 %r214, %r90, %r194;
or.b32 %r215, %r214, -2139062144;
and.b32 %r216, %r90, 2139062143;
and.b32 %r217, %r194, -2139062144;
sub.s32 %r218, %r215, %r216;
xor.b32 %r219, %r217, %r218;
xor.b32 %r97, %r219, -2139062144;
ld.global.nc.u32 %r13, [%rd67+20];
ld.global.nc.u32 %r14, [%rd67+24];

dp4a.s32.s32 %r92, %r93, %r13, %r84;


dp4a.s32.s32 %r96, %r97, %r14, %r92;

and.b16 %rs20, %rs4, 255;
cvt.u32.u16 %r220, %rs20;
mul.wide.u16 %r221, %rs5, 4;
and.b32 %r222, %r221, 256;
or.b32 %r223, %r222, %r220;
mul.wide.u32 %rd76, %r223, 4;
add.s64 %rd77, %rd63, %rd76;
ld.global.nc.u32 %r224, [%rd77];
shr.u16 %rs21, %rs4, 8;
cvt.u32.u16 %r225, %rs21;
mul.wide.u16 %r226, %rs5, 2;
and.b32 %r227, %r226, 256;
or.b32 %r228, %r227, %r225;
mul.wide.u32 %rd78, %r228, 4;
add.s64 %rd79, %rd63, %rd78;
ld.global.nc.u32 %r229, [%rd79];
shr.u16 %rs22, %rs7, 8;
mul.wide.u16 %r230, %rs22, 128;
cvt.u32.u16 %r231, %rs22;
shl.b32 %r232, %r231, 21;
or.b32 %r233, %r230, %r232;
and.b32 %r234, %r233, 25166208;
or.b32 %r235, %r234, -2139062144;
add.s32 %r236, %r235, -16843009;
or.b32 %r100, %r236, %r234;

prmt.b32 %r100, %r100, 0, 0xba98;

mul.wide.u16 %r237, %rs22, 8;
shl.b32 %r238, %r231, 17;
or.b32 %r239, %r237, %r238;
and.b32 %r240, %r239, 25166208;
or.b32 %r241, %r240, -2139062144;
add.s32 %r242, %r241, -16843009;
or.b32 %r102, %r242, %r240;

prmt.b32 %r102, %r102, 0, 0xba98;

xor.b32 %r243, %r100, %r224;
or.b32 %r244, %r243, -2139062144;
and.b32 %r245, %r100, 2139062143;
and.b32 %r246, %r224, -2139062144;
sub.s32 %r247, %r244, %r245;
xor.b32 %r248, %r246, %r247;
xor.b32 %r105, %r248, -2139062144;
xor.b32 %r249, %r102, %r229;
or.b32 %r250, %r249, -2139062144;
and.b32 %r251, %r102, 2139062143;
and.b32 %r252, %r229, -2139062144;
sub.s32 %r253, %r250, %r251;
xor.b32 %r254, %r252, %r253;
xor.b32 %r109, %r254, -2139062144;
ld.global.nc.u32 %r15, [%rd67+28];
ld.global.nc.u32 %r16, [%rd67+32];

dp4a.s32.s32 %r104, %r105, %r15, %r96;


dp4a.s32.s32 %r108, %r109, %r16, %r104;

add.s64 %rd80, %rd183, %rd13;
ld.global.nc.u8 %rs23, [%rd80];
cvt.u32.u16 %r255, %rs23;
and.b32 %r256, %r255, 255;
and.b32 %r258, %r45, 4;
shr.u32 %r259, %r256, %r258;
shl.b32 %r260, %r259, 1;
and.b32 %r261, %r260, 30;
or.b32 %r262, %r261, 1;
mul.lo.s32 %r263, %r262, %r108;
add.s64 %rd81, %rd183, %rd6;
ld.global.nc.u16 %rs1, [%rd81];

{ cvt.f32.f16 %f42, %rs1;}


ld.global.nc.u32 %r17, [%rd67];

{.reg .f16 low,high;
mov.b32 {low,high},%r17;
cvt.f32.f16 %f43, low;}


mul.ftz.f32 %f44, %f42, %f43;
cvt.rn.f32.s32 %f45, %r263;
fma.rn.ftz.f32 %f165, %f44, %f45, %f165;
setp.eq.s64 %p3, %rd28, 0;
@%p3 bra $L__BB304_5;

and.b32 %r956, %r45, 4;
mov.u64 %rd167, iq3s_grid;
add.s64 %rd82, %rd184, %rd7;
ld.global.u16 %rs25, [%rd82+4];
ld.global.u16 %rs26, [%rd82+6];
ld.global.u16 %rs27, [%rd82+8];
add.s64 %rd83, %rd184, %rd10;
add.s64 %rd84, %rd184, %rd11;
ld.global.u16 %rs28, [%rd84+-2];
ld.global.u16 %rs29, [%rd84];
ld.global.u16 %rs30, [%rd82+2];
and.b16 %rs31, %rs30, 255;
cvt.u32.u16 %r313, %rs31;
ld.global.u8 %rs32, [%rd83];
shl.b16 %rs33, %rs32, 8;
cvt.u32.u16 %r314, %rs33;
and.b32 %r315, %r314, 256;
or.b32 %r316, %r315, %r313;
mul.wide.u32 %rd85, %r316, 4;
add.s64 %rd87, %rd167, %rd85;
ld.global.nc.u32 %r317, [%rd87];
shr.u16 %rs34, %rs30, 8;
cvt.u32.u16 %r318, %rs34;
mul.wide.u16 %r319, %rs32, 128;
and.b32 %r320, %r319, 256;
or.b32 %r321, %r320, %r318;
mul.wide.u32 %rd88, %r321, 4;
add.s64 %rd89, %rd167, %rd88;
ld.global.nc.u32 %r322, [%rd89];
and.b16 %rs35, %rs28, 255;
mul.wide.u16 %r323, %rs35, 128;
cvt.u32.u16 %r324, %rs28;
shl.b32 %r325, %r324, 21;
or.b32 %r326, %r323, %r325;
and.b32 %r327, %r326, 25166208;
or.b32 %r328, %r327, -2139062144;
add.s32 %r329, %r328, -16843009;
or.b32 %r264, %r329, %r327;

prmt.b32 %r264, %r264, 0, 0xba98;

mul.wide.u16 %r330, %rs35, 8;
shl.b32 %r331, %r324, 17;
or.b32 %r332, %r330, %r331;
and.b32 %r333, %r332, 25166208;
or.b32 %r334, %r333, -2139062144;
add.s32 %r335, %r334, -16843009;
or.b32 %r266, %r335, %r333;

prmt.b32 %r266, %r266, 0, 0xba98;

xor.b32 %r336, %r264, %r317;
or.b32 %r337, %r336, -2139062144;
and.b32 %r338, %r264, 2139062143;
and.b32 %r339, %r317, -2139062144;
sub.s32 %r340, %r337, %r338;
xor.b32 %r341, %r339, %r340;
xor.b32 %r269, %r341, -2139062144;
xor.b32 %r342, %r266, %r322;
or.b32 %r343, %r342, -2139062144;
and.b32 %r344, %r266, 2139062143;
and.b32 %r345, %r322, -2139062144;
sub.s32 %r346, %r343, %r344;
xor.b32 %r347, %r345, %r346;
xor.b32 %r273, %r347, -2139062144;

dp4a.s32.s32 %r268, %r269, %r9, %r71;


dp4a.s32.s32 %r272, %r273, %r10, %r268;

and.b16 %rs36, %rs25, 255;
cvt.u32.u16 %r348, %rs36;
mul.wide.u16 %r349, %rs32, 64;
and.b32 %r350, %r349, 256;
or.b32 %r351, %r350, %r348;
mul.wide.u32 %rd90, %r351, 4;
add.s64 %rd91, %rd167, %rd90;
ld.global.nc.u32 %r352, [%rd91];
shr.u16 %rs37, %rs25, 8;
cvt.u32.u16 %r353, %rs37;
mul.wide.u16 %r354, %rs32, 32;
and.b32 %r355, %r354, 256;
or.b32 %r356, %r355, %r353;
mul.wide.u32 %rd92, %r356, 4;
add.s64 %rd93, %rd167, %rd92;
ld.global.nc.u32 %r357, [%rd93];
shr.u16 %rs38, %rs28, 8;
mul.wide.u16 %r358, %rs38, 128;
cvt.u32.u16 %r359, %rs38;
shl.b32 %r360, %r359, 21;
or.b32 %r361, %r358, %r360;
and.b32 %r362, %r361, 25166208;
or.b32 %r363, %r362, -2139062144;
add.s32 %r364, %r363, -16843009;
or.b32 %r276, %r364, %r362;

prmt.b32 %r276, %r276, 0, 0xba98;

mul.wide.u16 %r365, %rs38, 8;
shl.b32 %r366, %r359, 17;
or.b32 %r367, %r365, %r366;
and.b32 %r368, %r367, 25166208;
or.b32 %r369, %r368, -2139062144;
add.s32 %r370, %r369, -16843009;
or.b32 %r278, %r370, %r368;

prmt.b32 %r278, %r278, 0, 0xba98;

xor.b32 %r371, %r276, %r352;
or.b32 %r372, %r371, -2139062144;
and.b32 %r373, %r276, 2139062143;
and.b32 %r374, %r352, -2139062144;
sub.s32 %r375, %r372, %r373;
xor.b32 %r376, %r374, %r375;
xor.b32 %r281, %r376, -2139062144;
xor.b32 %r377, %r278, %r357;
or.b32 %r378, %r377, -2139062144;
and.b32 %r379, %r278, 2139062143;
and.b32 %r380, %r357, -2139062144;
sub.s32 %r381, %r378, %r379;
xor.b32 %r382, %r380, %r381;
xor.b32 %r285, %r382, -2139062144;

dp4a.s32.s32 %r280, %r281, %r11, %r272;


dp4a.s32.s32 %r284, %r285, %r12, %r280;

and.b16 %rs39, %rs26, 255;
cvt.u32.u16 %r383, %rs39;
mul.wide.u16 %r384, %rs32, 16;
and.b32 %r385, %r384, 256;
or.b32 %r386, %r385, %r383;
mul.wide.u32 %rd94, %r386, 4;
add.s64 %rd95, %rd167, %rd94;
ld.global.nc.u32 %r387, [%rd95];
shr.u16 %rs40, %rs26, 8;
cvt.u32.u16 %r388, %rs40;
mul.wide.u16 %r389, %rs32, 8;
and.b32 %r390, %r389, 256;
or.b32 %r391, %r390, %r388;
mul.wide.u32 %rd96, %r391, 4;
add.s64 %rd97, %rd167, %rd96;
ld.global.nc.u32 %r392, [%rd97];
and.b16 %rs41, %rs29, 255;
mul.wide.u16 %r393, %rs41, 128;
cvt.u32.u16 %r394, %rs29;
shl.b32 %r395, %r394, 21;
or.b32 %r396, %r393, %r395;
and.b32 %r397, %r396, 25166208;
or.b32 %r398, %r397, -2139062144;
add.s32 %r399, %r398, -16843009;
or.b32 %r288, %r399, %r397;

prmt.b32 %r288, %r288, 0, 0xba98;

mul.wide.u16 %r400, %rs41, 8;
shl.b32 %r401, %r394, 17;
or.b32 %r402, %r400, %r401;
and.b32 %r403, %r402, 25166208;
or.b32 %r404, %r403, -2139062144;
add.s32 %r405, %r404, -16843009;
or.b32 %r290, %r405, %r403;

prmt.b32 %r290, %r290, 0, 0xba98;

xor.b32 %r406, %r288, %r387;
or.b32 %r407, %r406, -2139062144;
and.b32 %r408, %r288, 2139062143;
and.b32 %r409, %r387, -2139062144;
sub.s32 %r410, %r407, %r408;
xor.b32 %r411, %r409, %r410;
xor.b32 %r293, %r411, -2139062144;
xor.b32 %r412, %r290, %r392;
or.b32 %r413, %r412, -2139062144;
and.b32 %r414, %r290, 2139062143;
and.b32 %r415, %r392, -2139062144;
sub.s32 %r416, %r413, %r414;
xor.b32 %r417, %r415, %r416;
xor.b32 %r297, %r417, -2139062144;

dp4a.s32.s32 %r292, %r293, %r13, %r284;


dp4a.s32.s32 %r296, %r297, %r14, %r292;

and.b16 %rs42, %rs27, 255;
cvt.u32.u16 %r418, %rs42;
mul.wide.u16 %r419, %rs32, 4;
and.b32 %r420, %r419, 256;
or.b32 %r421, %r420, %r418;
mul.wide.u32 %rd98, %r421, 4;
add.s64 %rd99, %rd167, %rd98;
ld.global.nc.u32 %r422, [%rd99];
shr.u16 %rs43, %rs27, 8;
cvt.u32.u16 %r423, %rs43;
mul.wide.u16 %r424, %rs32, 2;
and.b32 %r425, %r424, 256;
or.b32 %r426, %r425, %r423;
mul.wide.u32 %rd100, %r426, 4;
add.s64 %rd101, %rd167, %rd100;
ld.global.nc.u32 %r427, [%rd101];
shr.u16 %rs44, %rs29, 8;
mul.wide.u16 %r428, %rs44, 128;
cvt.u32.u16 %r429, %rs44;
shl.b32 %r430, %r429, 21;
or.b32 %r431, %r428, %r430;
and.b32 %r432, %r431, 25166208;
or.b32 %r433, %r432, -2139062144;
add.s32 %r434, %r433, -16843009;
or.b32 %r300, %r434, %r432;

prmt.b32 %r300, %r300, 0, 0xba98;

mul.wide.u16 %r435, %rs44, 8;
shl.b32 %r436, %r429, 17;
or.b32 %r437, %r435, %r436;
and.b32 %r438, %r437, 25166208;
or.b32 %r439, %r438, -2139062144;
add.s32 %r440, %r439, -16843009;
or.b32 %r302, %r440, %r438;

prmt.b32 %r302, %r302, 0, 0xba98;

xor.b32 %r441, %r300, %r422;
or.b32 %r442, %r441, -2139062144;
and.b32 %r443, %r300, 2139062143;
and.b32 %r444, %r422, -2139062144;
sub.s32 %r445, %r442, %r443;
xor.b32 %r446, %r444, %r445;
xor.b32 %r305, %r446, -2139062144;
xor.b32 %r447, %r302, %r427;
or.b32 %r448, %r447, -2139062144;
and.b32 %r449, %r302, 2139062143;
and.b32 %r450, %r427, -2139062144;
sub.s32 %r451, %r448, %r449;
xor.b32 %r452, %r450, %r451;
xor.b32 %r309, %r452, -2139062144;

dp4a.s32.s32 %r304, %r305, %r15, %r296;


dp4a.s32.s32 %r308, %r309, %r16, %r304;

add.s64 %rd102, %rd184, %rd13;
ld.global.u8 %r453, [%rd102];
shr.u32 %r457, %r453, %r956;
shl.b32 %r458, %r457, 1;
and.b32 %r459, %r458, 30;
or.b32 %r460, %r459, 1;
mul.lo.s32 %r461, %r460, %r308;
add.s64 %rd103, %rd184, %rd6;
ld.global.u16 %rs24, [%rd103];

{ cvt.f32.f16 %f46, %rs24;}



{.reg .f16 low,high;
mov.b32 {low,high},%r17;
cvt.f32.f16 %f47, low;}


mul.ftz.f32 %f48, %f46, %f47;
cvt.rn.f32.s32 %f49, %r461;
fma.rn.ftz.f32 %f163, %f48, %f49, %f163;

$L__BB304_5:
and.b32 %r949, %r45, 4;
mov.u64 %rd161, iq3s_grid;
add.s64 %rd104, %rd183, %rd12;
ld.global.nc.u16 %rs46, [%rd104];
ld.global.nc.u16 %rs47, [%rd104+2];
ld.global.nc.u16 %rs48, [%rd104+4];
add.s64 %rd105, %rd183, %rd9;
ld.global.nc.u8 %rs49, [%rd105];
add.s64 %rd106, %rd183, %rd8;
ld.global.nc.u16 %rs50, [%rd106];
ld.global.nc.u16 %rs51, [%rd106+2];
ld.global.nc.u16 %rs52, [%rd104+-2];
and.b16 %rs53, %rs52, 255;
cvt.u32.u16 %r511, %rs53;
shl.b16 %rs54, %rs49, 8;
cvt.u32.u16 %r512, %rs54;
and.b32 %r513, %r512, 256;
or.b32 %r514, %r513, %r511;
mul.wide.u32 %rd107, %r514, 4;
add.s64 %rd109, %rd161, %rd107;
ld.global.nc.u32 %r515, [%rd109];
shr.u16 %rs55, %rs52, 8;
cvt.u32.u16 %r516, %rs55;
mul.wide.u16 %r517, %rs49, 128;
and.b32 %r518, %r517, 256;
or.b32 %r519, %r518, %r516;
mul.wide.u32 %rd110, %r519, 4;
add.s64 %rd111, %rd161, %rd110;
ld.global.nc.u32 %r520, [%rd111];
and.b16 %rs57, %rs50, 255;
mul.wide.u16 %r521, %rs57, 128;
cvt.u32.u16 %r522, %rs50;
shl.b32 %r523, %r522, 21;
or.b32 %r524, %r521, %r523;
and.b32 %r525, %r524, 25166208;
or.b32 %r526, %r525, -2139062144;
add.s32 %r527, %r526, -16843009;
or.b32 %r462, %r527, %r525;

prmt.b32 %r462, %r462, 0, 0xba98;

mul.wide.u16 %r528, %rs57, 8;
shl.b32 %r529, %r522, 17;
or.b32 %r530, %r528, %r529;
and.b32 %r531, %r530, 25166208;
or.b32 %r532, %r531, -2139062144;
add.s32 %r533, %r532, -16843009;
or.b32 %r464, %r533, %r531;

prmt.b32 %r464, %r464, 0, 0xba98;

xor.b32 %r534, %r462, %r515;
or.b32 %r535, %r534, -2139062144;
and.b32 %r536, %r462, 2139062143;
and.b32 %r537, %r515, -2139062144;
sub.s32 %r538, %r535, %r536;
xor.b32 %r539, %r537, %r538;
xor.b32 %r467, %r539, -2139062144;
xor.b32 %r540, %r464, %r520;
or.b32 %r541, %r540, -2139062144;
and.b32 %r542, %r464, 2139062143;
and.b32 %r543, %r520, -2139062144;
sub.s32 %r544, %r541, %r542;
xor.b32 %r545, %r543, %r544;
xor.b32 %r471, %r545, -2139062144;

dp4a.s32.s32 %r466, %r467, %r9, %r71;


dp4a.s32.s32 %r470, %r471, %r10, %r466;

and.b16 %rs58, %rs46, 255;
cvt.u32.u16 %r546, %rs58;
mul.wide.u16 %r547, %rs49, 64;
and.b32 %r548, %r547, 256;
or.b32 %r549, %r548, %r546;
mul.wide.u32 %rd112, %r549, 4;
add.s64 %rd113, %rd161, %rd112;
ld.global.nc.u32 %r550, [%rd113];
shr.u16 %rs59, %rs46, 8;
cvt.u32.u16 %r551, %rs59;
mul.wide.u16 %r552, %rs49, 32;
and.b32 %r553, %r552, 256;
or.b32 %r554, %r553, %r551;
mul.wide.u32 %rd114, %r554, 4;
add.s64 %rd115, %rd161, %rd114;
ld.global.nc.u32 %r555, [%rd115];
shr.u16 %rs60, %rs50, 8;
mul.wide.u16 %r556, %rs60, 128;
cvt.u32.u16 %r557, %rs60;
shl.b32 %r558, %r557, 21;
or.b32 %r559, %r556, %r558;
and.b32 %r560, %r559, 25166208;
or.b32 %r561, %r560, -2139062144;
add.s32 %r562, %r561, -16843009;
or.b32 %r474, %r562, %r560;

prmt.b32 %r474, %r474, 0, 0xba98;

mul.wide.u16 %r563, %rs60, 8;
shl.b32 %r564, %r557, 17;
or.b32 %r565, %r563, %r564;
and.b32 %r566, %r565, 25166208;
or.b32 %r567, %r566, -2139062144;
add.s32 %r568, %r567, -16843009;
or.b32 %r476, %r568, %r566;

prmt.b32 %r476, %r476, 0, 0xba98;

xor.b32 %r569, %r474, %r550;
or.b32 %r570, %r569, -2139062144;
and.b32 %r571, %r474, 2139062143;
and.b32 %r572, %r550, -2139062144;
sub.s32 %r573, %r570, %r571;
xor.b32 %r574, %r572, %r573;
xor.b32 %r479, %r574, -2139062144;
xor.b32 %r575, %r476, %r555;
or.b32 %r576, %r575, -2139062144;
and.b32 %r577, %r476, 2139062143;
and.b32 %r578, %r555, -2139062144;
sub.s32 %r579, %r576, %r577;
xor.b32 %r580, %r578, %r579;
xor.b32 %r483, %r580, -2139062144;

dp4a.s32.s32 %r478, %r479, %r11, %r470;


dp4a.s32.s32 %r482, %r483, %r12, %r478;

and.b16 %rs61, %rs47, 255;
cvt.u32.u16 %r581, %rs61;
mul.wide.u16 %r582, %rs49, 16;
and.b32 %r583, %r582, 256;
or.b32 %r584, %r583, %r581;
mul.wide.u32 %rd116, %r584, 4;
add.s64 %rd117, %rd161, %rd116;
ld.global.nc.u32 %r585, [%rd117];
shr.u16 %rs62, %rs47, 8;
cvt.u32.u16 %r586, %rs62;
mul.wide.u16 %r587, %rs49, 8;
and.b32 %r588, %r587, 256;
or.b32 %r589, %r588, %r586;
mul.wide.u32 %rd118, %r589, 4;
add.s64 %rd119, %rd161, %rd118;
ld.global.nc.u32 %r590, [%rd119];
and.b16 %rs63, %rs51, 255;
mul.wide.u16 %r591, %rs63, 128;
cvt.u32.u16 %r592, %rs51;
shl.b32 %r593, %r592, 21;
or.b32 %r594, %r591, %r593;
and.b32 %r595, %r594, 25166208;
or.b32 %r596, %r595, -2139062144;
add.s32 %r597, %r596, -16843009;
or.b32 %r486, %r597, %r595;

prmt.b32 %r486, %r486, 0, 0xba98;

mul.wide.u16 %r598, %rs63, 8;
shl.b32 %r599, %r592, 17;
or.b32 %r600, %r598, %r599;
and.b32 %r601, %r600, 25166208;
or.b32 %r602, %r601, -2139062144;
add.s32 %r603, %r602, -16843009;
or.b32 %r488, %r603, %r601;

prmt.b32 %r488, %r488, 0, 0xba98;

xor.b32 %r604, %r486, %r585;
or.b32 %r605, %r604, -2139062144;
and.b32 %r606, %r486, 2139062143;
and.b32 %r607, %r585, -2139062144;
sub.s32 %r608, %r605, %r606;
xor.b32 %r609, %r607, %r608;
xor.b32 %r491, %r609, -2139062144;
xor.b32 %r610, %r488, %r590;
or.b32 %r611, %r610, -2139062144;
and.b32 %r612, %r488, 2139062143;
and.b32 %r613, %r590, -2139062144;
sub.s32 %r614, %r611, %r612;
xor.b32 %r615, %r613, %r614;
xor.b32 %r495, %r615, -2139062144;

dp4a.s32.s32 %r490, %r491, %r13, %r482;


dp4a.s32.s32 %r494, %r495, %r14, %r490;

and.b16 %rs64, %rs48, 255;
cvt.u32.u16 %r616, %rs64;
mul.wide.u16 %r617, %rs49, 4;
and.b32 %r618, %r617, 256;
or.b32 %r619, %r618, %r616;
mul.wide.u32 %rd120, %r619, 4;
add.s64 %rd121, %rd161, %rd120;
ld.global.nc.u32 %r620, [%rd121];
shr.u16 %rs65, %rs48, 8;
cvt.u32.u16 %r621, %rs65;
mul.wide.u16 %r622, %rs49, 2;
and.b32 %r623, %r622, 256;
or.b32 %r624, %r623, %r621;
mul.wide.u32 %rd122, %r624, 4;
add.s64 %rd123, %rd161, %rd122;
ld.global.nc.u32 %r625, [%rd123];
shr.u16 %rs66, %rs51, 8;
mul.wide.u16 %r626, %rs66, 128;
cvt.u32.u16 %r627, %rs66;
shl.b32 %r628, %r627, 21;
or.b32 %r629, %r626, %r628;
and.b32 %r630, %r629, 25166208;
or.b32 %r631, %r630, -2139062144;
add.s32 %r632, %r631, -16843009;
or.b32 %r498, %r632, %r630;

prmt.b32 %r498, %r498, 0, 0xba98;

mul.wide.u16 %r633, %rs66, 8;
shl.b32 %r634, %r627, 17;
or.b32 %r635, %r633, %r634;
and.b32 %r636, %r635, 25166208;
or.b32 %r637, %r636, -2139062144;
add.s32 %r638, %r637, -16843009;
or.b32 %r500, %r638, %r636;

prmt.b32 %r500, %r500, 0, 0xba98;

xor.b32 %r639, %r498, %r620;
or.b32 %r640, %r639, -2139062144;
and.b32 %r641, %r498, 2139062143;
and.b32 %r642, %r620, -2139062144;
sub.s32 %r643, %r640, %r641;
xor.b32 %r644, %r642, %r643;
xor.b32 %r503, %r644, -2139062144;
xor.b32 %r645, %r500, %r625;
or.b32 %r646, %r645, -2139062144;
and.b32 %r647, %r500, 2139062143;
and.b32 %r648, %r625, -2139062144;
sub.s32 %r649, %r646, %r647;
xor.b32 %r650, %r648, %r649;
xor.b32 %r507, %r650, -2139062144;

dp4a.s32.s32 %r502, %r503, %r15, %r494;


dp4a.s32.s32 %r506, %r507, %r16, %r502;

add.s64 %rd124, %rd183, %rd5;
ld.global.nc.u8 %rs67, [%rd124];
cvt.u32.u16 %r651, %rs67;
and.b32 %r652, %r651, 255;
shr.u32 %r656, %r652, %r949;
shl.b32 %r657, %r656, 1;
and.b32 %r658, %r657, 30;
or.b32 %r659, %r658, 1;
mul.lo.s32 %r660, %r659, %r506;
add.s64 %rd125, %rd183, %rd4;
ld.global.nc.u16 %rs45, [%rd125];

{ cvt.f32.f16 %f50, %rs45;}



{.reg .f16 low,high;
mov.b32 {low,high},%r17;
cvt.f32.f16 %f51, low;}


mul.ftz.f32 %f52, %f50, %f51;
cvt.rn.f32.s32 %f53, %r660;
fma.rn.ftz.f32 %f164, %f52, %f53, %f164;
@%p3 bra $L__BB304_7;

and.b32 %r955, %r45, 4;
mov.u64 %rd168, iq3s_grid;
add.s64 %rd126, %rd184, %rd12;
ld.global.u16 %rs69, [%rd126];
ld.global.u16 %rs70, [%rd126+2];
ld.global.u16 %rs71, [%rd126+4];
add.s64 %rd127, %rd184, %rd9;
add.s64 %rd128, %rd184, %rd8;
ld.global.u16 %rs72, [%rd128];
ld.global.u16 %rs73, [%rd128+2];
ld.global.u16 %rs74, [%rd126+-2];
and.b16 %rs75, %rs74, 255;
cvt.u32.u16 %r710, %rs75;
ld.global.u8 %rs76, [%rd127];
shl.b16 %rs77, %rs76, 8;
cvt.u32.u16 %r711, %rs77;
and.b32 %r712, %r711, 256;
or.b32 %r713, %r712, %r710;
mul.wide.u32 %rd129, %r713, 4;
add.s64 %rd131, %rd168, %rd129;
ld.global.nc.u32 %r714, [%rd131];
shr.u16 %rs78, %rs74, 8;
cvt.u32.u16 %r715, %rs78;
mul.wide.u16 %r716, %rs76, 128;
and.b32 %r717, %r716, 256;
or.b32 %r718, %r717, %r715;
mul.wide.u32 %rd132, %r718, 4;
add.s64 %rd133, %rd168, %rd132;
ld.global.nc.u32 %r719, [%rd133];
and.b16 %rs79, %rs72, 255;
mul.wide.u16 %r720, %rs79, 128;
cvt.u32.u16 %r721, %rs72;
shl.b32 %r722, %r721, 21;
or.b32 %r723, %r720, %r722;
and.b32 %r724, %r723, 25166208;
or.b32 %r725, %r724, -2139062144;
add.s32 %r726, %r725, -16843009;
or.b32 %r661, %r726, %r724;

prmt.b32 %r661, %r661, 0, 0xba98;

mul.wide.u16 %r727, %rs79, 8;
shl.b32 %r728, %r721, 17;
or.b32 %r729, %r727, %r728;
and.b32 %r730, %r729, 25166208;
or.b32 %r731, %r730, -2139062144;
add.s32 %r732, %r731, -16843009;
or.b32 %r663, %r732, %r730;

prmt.b32 %r663, %r663, 0, 0xba98;

xor.b32 %r733, %r661, %r714;
or.b32 %r734, %r733, -2139062144;
and.b32 %r735, %r661, 2139062143;
and.b32 %r736, %r714, -2139062144;
sub.s32 %r737, %r734, %r735;
xor.b32 %r738, %r736, %r737;
xor.b32 %r666, %r738, -2139062144;
xor.b32 %r739, %r663, %r719;
or.b32 %r740, %r739, -2139062144;
and.b32 %r741, %r663, 2139062143;
and.b32 %r742, %r719, -2139062144;
sub.s32 %r743, %r740, %r741;
xor.b32 %r744, %r742, %r743;
xor.b32 %r670, %r744, -2139062144;
mov.u32 %r668, 0;

dp4a.s32.s32 %r665, %r666, %r9, %r668;


dp4a.s32.s32 %r669, %r670, %r10, %r665;

and.b16 %rs80, %rs69, 255;
cvt.u32.u16 %r745, %rs80;
mul.wide.u16 %r746, %rs76, 64;
and.b32 %r747, %r746, 256;
or.b32 %r748, %r747, %r745;
mul.wide.u32 %rd134, %r748, 4;
add.s64 %rd135, %rd168, %rd134;
ld.global.nc.u32 %r749, [%rd135];
shr.u16 %rs81, %rs69, 8;
cvt.u32.u16 %r750, %rs81;
mul.wide.u16 %r751, %rs76, 32;
and.b32 %r752, %r751, 256;
or.b32 %r753, %r752, %r750;
mul.wide.u32 %rd136, %r753, 4;
add.s64 %rd137, %rd168, %rd136;
ld.global.nc.u32 %r754, [%rd137];
shr.u16 %rs82, %rs72, 8;
mul.wide.u16 %r755, %rs82, 128;
cvt.u32.u16 %r756, %rs82;
shl.b32 %r757, %r756, 21;
or.b32 %r758, %r755, %r757;
and.b32 %r759, %r758, 25166208;
or.b32 %r760, %r759, -2139062144;
add.s32 %r761, %r760, -16843009;
or.b32 %r673, %r761, %r759;

prmt.b32 %r673, %r673, 0, 0xba98;

mul.wide.u16 %r762, %rs82, 8;
shl.b32 %r763, %r756, 17;
or.b32 %r764, %r762, %r763;
and.b32 %r765, %r764, 25166208;
or.b32 %r766, %r765, -2139062144;
add.s32 %r767, %r766, -16843009;
or.b32 %r675, %r767, %r765;

prmt.b32 %r675, %r675, 0, 0xba98;

xor.b32 %r768, %r673, %r749;
or.b32 %r769, %r768, -2139062144;
and.b32 %r770, %r673, 2139062143;
and.b32 %r771, %r749, -2139062144;
sub.s32 %r772, %r769, %r770;
xor.b32 %r773, %r771, %r772;
xor.b32 %r678, %r773, -2139062144;
xor.b32 %r774, %r675, %r754;
or.b32 %r775, %r774, -2139062144;
and.b32 %r776, %r675, 2139062143;
and.b32 %r777, %r754, -2139062144;
sub.s32 %r778, %r775, %r776;
xor.b32 %r779, %r777, %r778;
xor.b32 %r682, %r779, -2139062144;

dp4a.s32.s32 %r677, %r678, %r11, %r669;


dp4a.s32.s32 %r681, %r682, %r12, %r677;

and.b16 %rs83, %rs70, 255;
cvt.u32.u16 %r780, %rs83;
mul.wide.u16 %r781, %rs76, 16;
and.b32 %r782, %r781, 256;
or.b32 %r783, %r782, %r780;
mul.wide.u32 %rd138, %r783, 4;
add.s64 %rd139, %rd168, %rd138;
ld.global.nc.u32 %r784, [%rd139];
shr.u16 %rs84, %rs70, 8;
cvt.u32.u16 %r785, %rs84;
mul.wide.u16 %r786, %rs76, 8;
and.b32 %r787, %r786, 256;
or.b32 %r788, %r787, %r785;
mul.wide.u32 %rd140, %r788, 4;
add.s64 %rd141, %rd168, %rd140;
ld.global.nc.u32 %r789, [%rd141];
and.b16 %rs85, %rs73, 255;
mul.wide.u16 %r790, %rs85, 128;
cvt.u32.u16 %r791, %rs73;
shl.b32 %r792, %r791, 21;
or.b32 %r793, %r790, %r792;
and.b32 %r794, %r793, 25166208;
or.b32 %r795, %r794, -2139062144;
add.s32 %r796, %r795, -16843009;
or.b32 %r685, %r796, %r794;

prmt.b32 %r685, %r685, 0, 0xba98;

mul.wide.u16 %r797, %rs85, 8;
shl.b32 %r798, %r791, 17;
or.b32 %r799, %r797, %r798;
and.b32 %r800, %r799, 25166208;
or.b32 %r801, %r800, -2139062144;
add.s32 %r802, %r801, -16843009;
or.b32 %r687, %r802, %r800;

prmt.b32 %r687, %r687, 0, 0xba98;

xor.b32 %r803, %r685, %r784;
or.b32 %r804, %r803, -2139062144;
and.b32 %r805, %r685, 2139062143;
and.b32 %r806, %r784, -2139062144;
sub.s32 %r807, %r804, %r805;
xor.b32 %r808, %r806, %r807;
xor.b32 %r690, %r808, -2139062144;
xor.b32 %r809, %r687, %r789;
or.b32 %r810, %r809, -2139062144;
and.b32 %r811, %r687, 2139062143;
and.b32 %r812, %r789, -2139062144;
sub.s32 %r813, %r810, %r811;
xor.b32 %r814, %r812, %r813;
xor.b32 %r694, %r814, -2139062144;

dp4a.s32.s32 %r689, %r690, %r13, %r681;


dp4a.s32.s32 %r693, %r694, %r14, %r689;

and.b16 %rs86, %rs71, 255;
cvt.u32.u16 %r815, %rs86;
mul.wide.u16 %r816, %rs76, 4;
and.b32 %r817, %r816, 256;
or.b32 %r818, %r817, %r815;
mul.wide.u32 %rd142, %r818, 4;
add.s64 %rd143, %rd168, %rd142;
ld.global.nc.u32 %r819, [%rd143];
shr.u16 %rs87, %rs71, 8;
cvt.u32.u16 %r820, %rs87;
mul.wide.u16 %r821, %rs76, 2;
and.b32 %r822, %r821, 256;
or.b32 %r823, %r822, %r820;
mul.wide.u32 %rd144, %r823, 4;
add.s64 %rd145, %rd168, %rd144;
ld.global.nc.u32 %r824, [%rd145];
shr.u16 %rs88, %rs73, 8;
mul.wide.u16 %r825, %rs88, 128;
cvt.u32.u16 %r826, %rs88;
shl.b32 %r827, %r826, 21;
or.b32 %r828, %r825, %r827;
and.b32 %r829, %r828, 25166208;
or.b32 %r830, %r829, -2139062144;
add.s32 %r831, %r830, -16843009;
or.b32 %r697, %r831, %r829;

prmt.b32 %r697, %r697, 0, 0xba98;

mul.wide.u16 %r832, %rs88, 8;
shl.b32 %r833, %r826, 17;
or.b32 %r834, %r832, %r833;
and.b32 %r835, %r834, 25166208;
or.b32 %r836, %r835, -2139062144;
add.s32 %r837, %r836, -16843009;
or.b32 %r699, %r837, %r835;

prmt.b32 %r699, %r699, 0, 0xba98;

xor.b32 %r838, %r697, %r819;
or.b32 %r839, %r838, -2139062144;
and.b32 %r840, %r697, 2139062143;
and.b32 %r841, %r819, -2139062144;
sub.s32 %r842, %r839, %r840;
xor.b32 %r843, %r841, %r842;
xor.b32 %r702, %r843, -2139062144;
xor.b32 %r844, %r699, %r824;
or.b32 %r845, %r844, -2139062144;
and.b32 %r846, %r699, 2139062143;
and.b32 %r847, %r824, -2139062144;
sub.s32 %r848, %r845, %r846;
xor.b32 %r849, %r847, %r848;
xor.b32 %r706, %r849, -2139062144;

dp4a.s32.s32 %r701, %r702, %r15, %r693;


dp4a.s32.s32 %r705, %r706, %r16, %r701;

add.s64 %rd146, %rd184, %rd5;
ld.global.u8 %r850, [%rd146];
shr.u32 %r854, %r850, %r955;
shl.b32 %r855, %r854, 1;
and.b32 %r856, %r855, 30;
or.b32 %r857, %r856, 1;
mul.lo.s32 %r858, %r857, %r705;
add.s64 %rd147, %rd184, %rd4;
ld.global.u16 %rs68, [%rd147];

{ cvt.f32.f16 %f54, %rs68;}



{.reg .f16 low,high;
mov.b32 {low,high},%r17;
cvt.f32.f16 %f55, low;}


mul.ftz.f32 %f56, %f54, %f55;
cvt.rn.f32.s32 %f57, %r858;
fma.rn.ftz.f32 %f162, %f56, %f57, %f162;

$L__BB304_7:
add.s32 %r957, %r957, 32;
add.s64 %rd184, %rd184, 440;
add.s64 %rd183, %rd183, 440;
add.s32 %r958, %r958, 4;
setp.lt.u32 %p5, %r958, %r38;
@%p5 bra $L__BB304_3;

add.u64 %rd175, %SPL, 0;
add.u64 %rd169, %SPL, 8;
st.local.v2.f32 [%rd175], {%f165, %f164};
st.local.v2.f32 [%rd169], {%f163, %f162};

$L__BB304_9:
mov.b32 %r20, %f165;
setp.ne.s64 %p6, %rd28, 0;
@%p6 bra $L__BB304_11;
bra.uni $L__BB304_10;

$L__BB304_11:
add.u64 %rd181, %SPL, 0;
add.u64 %rd173, %SPL, 8;
mov.u32 %r886, 31;
mov.u32 %r887, 16;
mov.u32 %r888, -1;
shfl.sync.bfly.b32 %r889|%p17, %r20, %r887, %r886, %r888;
mov.b32 %f78, %r889;
add.ftz.f32 %f79, %f165, %f78;
mov.b32 %r890, %f79;
mov.u32 %r891, 8;
shfl.sync.bfly.b32 %r892|%p18, %r890, %r891, %r886, %r888;
mov.b32 %f80, %r892;
add.ftz.f32 %f81, %f79, %f80;
mov.b32 %r893, %f81;
mov.u32 %r894, 4;
shfl.sync.bfly.b32 %r895|%p19, %r893, %r894, %r886, %r888;
mov.b32 %f82, %r895;
add.ftz.f32 %f83, %f81, %f82;
mov.b32 %r896, %f83;
mov.u32 %r897, 2;
shfl.sync.bfly.b32 %r898|%p20, %r896, %r897, %r886, %r888;
mov.b32 %f84, %r898;
add.ftz.f32 %f85, %f83, %f84;
mov.b32 %r899, %f85;
mov.u32 %r900, 1;
shfl.sync.bfly.b32 %r901|%p21, %r899, %r900, %r886, %r888;
mov.b32 %f86, %r901;
add.ftz.f32 %f87, %f85, %f86;
st.local.f32 [%rd181], %f87;
mov.b32 %r902, %f163;
shfl.sync.bfly.b32 %r903|%p22, %r902, %r887, %r886, %r888;
mov.b32 %f88, %r903;
add.ftz.f32 %f89, %f163, %f88;
mov.b32 %r904, %f89;
shfl.sync.bfly.b32 %r905|%p23, %r904, %r891, %r886, %r888;
mov.b32 %f90, %r905;
add.ftz.f32 %f91, %f89, %f90;
mov.b32 %r906, %f91;
shfl.sync.bfly.b32 %r907|%p24, %r906, %r894, %r886, %r888;
mov.b32 %f92, %r907;
add.ftz.f32 %f93, %f91, %f92;
mov.b32 %r908, %f93;
shfl.sync.bfly.b32 %r909|%p25, %r908, %r897, %r886, %r888;
mov.b32 %f94, %r909;
add.ftz.f32 %f95, %f93, %f94;
mov.b32 %r910, %f95;
shfl.sync.bfly.b32 %r911|%p26, %r910, %r900, %r886, %r888;
mov.b32 %f96, %r911;
add.ftz.f32 %f97, %f95, %f96;
st.local.f32 [%rd173], %f97;
mov.b32 %r912, %f164;
shfl.sync.bfly.b32 %r913|%p27, %r912, %r887, %r886, %r888;
mov.b32 %f98, %r913;
add.ftz.f32 %f99, %f164, %f98;
mov.b32 %r914, %f99;
shfl.sync.bfly.b32 %r915|%p28, %r914, %r891, %r886, %r888;
mov.b32 %f100, %r915;
add.ftz.f32 %f101, %f99, %f100;
mov.b32 %r916, %f101;
shfl.sync.bfly.b32 %r917|%p29, %r916, %r894, %r886, %r888;
mov.b32 %f102, %r917;
add.ftz.f32 %f103, %f101, %f102;
mov.b32 %r918, %f103;
shfl.sync.bfly.b32 %r919|%p30, %r918, %r897, %r886, %r888;
mov.b32 %f104, %r919;
add.ftz.f32 %f105, %f103, %f104;
mov.b32 %r920, %f105;
shfl.sync.bfly.b32 %r921|%p31, %r920, %r900, %r886, %r888;
mov.b32 %f106, %r921;
add.ftz.f32 %f107, %f105, %f106;
st.local.f32 [%rd181+4], %f107;
mov.b32 %r922, %f162;
shfl.sync.bfly.b32 %r923|%p32, %r922, %r887, %r886, %r888;
mov.b32 %f108, %r923;
add.ftz.f32 %f109, %f162, %f108;
mov.b32 %r924, %f109;
shfl.sync.bfly.b32 %r925|%p33, %r924, %r891, %r886, %r888;
mov.b32 %f110, %r925;
add.ftz.f32 %f111, %f109, %f110;
mov.b32 %r926, %f111;
shfl.sync.bfly.b32 %r927|%p34, %r926, %r894, %r886, %r888;
mov.b32 %f112, %r927;
add.ftz.f32 %f113, %f111, %f112;
mov.b32 %r928, %f113;
shfl.sync.bfly.b32 %r929|%p35, %r928, %r897, %r886, %r888;
mov.b32 %f114, %r929;
add.ftz.f32 %f115, %f113, %f114;
mov.b32 %r930, %f115;
shfl.sync.bfly.b32 %r931|%p36, %r930, %r900, %r886, %r888;
mov.b32 %f116, %r931;
add.ftz.f32 %f117, %f115, %f116;
st.local.f32 [%rd173+4], %f117;
bra.uni $L__BB304_12;

$L__BB304_10:
add.u64 %rd177, %SPL, 0;
mov.u32 %r860, 31;
mov.u32 %r861, 16;
mov.u32 %r862, -1;
shfl.sync.bfly.b32 %r863|%p7, %r20, %r861, %r860, %r862;
mov.b32 %f58, %r863;
add.ftz.f32 %f59, %f165, %f58;
mov.b32 %r864, %f59;
mov.u32 %r865, 8;
shfl.sync.bfly.b32 %r866|%p8, %r864, %r865, %r860, %r862;
mov.b32 %f60, %r866;
add.ftz.f32 %f61, %f59, %f60;
mov.b32 %r867, %f61;
mov.u32 %r868, 4;
shfl.sync.bfly.b32 %r869|%p9, %r867, %r868, %r860, %r862;
mov.b32 %f62, %r869;
add.ftz.f32 %f63, %f61, %f62;
mov.b32 %r870, %f63;
mov.u32 %r871, 2;
shfl.sync.bfly.b32 %r872|%p10, %r870, %r871, %r860, %r862;
mov.b32 %f64, %r872;
add.ftz.f32 %f65, %f63, %f64;
mov.b32 %r873, %f65;
mov.u32 %r874, 1;
shfl.sync.bfly.b32 %r875|%p11, %r873, %r874, %r860, %r862;
mov.b32 %f66, %r875;
add.ftz.f32 %f67, %f65, %f66;
st.local.f32 [%rd177], %f67;
mov.b32 %r876, %f164;
shfl.sync.bfly.b32 %r877|%p12, %r876, %r861, %r860, %r862;
mov.b32 %f68, %r877;
add.ftz.f32 %f69, %f164, %f68;
mov.b32 %r878, %f69;
shfl.sync.bfly.b32 %r879|%p13, %r878, %r865, %r860, %r862;
mov.b32 %f70, %r879;
add.ftz.f32 %f71, %f69, %f70;
mov.b32 %r880, %f71;
shfl.sync.bfly.b32 %r881|%p14, %r880, %r868, %r860, %r862;
mov.b32 %f72, %r881;
add.ftz.f32 %f73, %f71, %f72;
mov.b32 %r882, %f73;
shfl.sync.bfly.b32 %r883|%p15, %r882, %r871, %r860, %r862;
mov.b32 %f74, %r883;
add.ftz.f32 %f75, %f73, %f74;
mov.b32 %r884, %f75;
shfl.sync.bfly.b32 %r885|%p16, %r884, %r874, %r860, %r862;
mov.b32 %f76, %r885;
add.ftz.f32 %f77, %f75, %f76;
st.local.f32 [%rd177+4], %f77;

$L__BB304_12:
ld.param.u32 %r942, [_Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7];
mov.u32 %r941, %tid.x;
setp.gt.u32 %p37, %r941, 1;
mov.u32 %r933, %ctaid.x;
shl.b32 %r934, %r933, 1;
add.s32 %r21, %r934, %r941;
setp.ge.u32 %p38, %r21, %r942;
or.pred %p39, %p37, %p38;
@%p39 bra $L__BB304_29;

add.u64 %rd179, %SPL, 0;
ld.param.u64 %rd162, [_Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3];
ld.param.u32 %r944, [_Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13];
mov.u32 %r943, %tid.x;
cvt.u64.u32 %rd22, %r943;
mul.wide.u32 %rd148, %r943, 4;
add.s64 %rd149, %rd179, %rd148;
ld.local.f32 %f168, [%rd149];
mad.lo.s32 %r22, %r3, %r944, %r21;
setp.eq.s64 %p40, %rd162, 0;
@%p40 bra $L__BB304_15;

ld.param.u64 %rd163, [_Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3];
cvta.to.global.u64 %rd150, %rd163;
mul.wide.u32 %rd151, %r22, 4;
add.s64 %rd152, %rd150, %rd151;
ld.global.f32 %f118, [%rd152];
add.ftz.f32 %f168, %f168, %f118;

$L__BB304_15:
setp.eq.s64 %p41, %rd28, 0;
@%p41 bra $L__BB304_28;

add.u64 %rd171, %SPL, 8;
ld.param.u64 %rd165, [_Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+16];
shl.b64 %rd153, %rd22, 2;
add.s64 %rd154, %rd171, %rd153;
ld.local.f32 %f24, [%rd154];
setp.eq.s64 %p42, %rd165, 0;
@%p42 bra $L__BB304_18;

ld.param.u64 %rd166, [_Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+16];
cvta.to.global.u64 %rd155, %rd166;
mul.wide.u32 %rd156, %r22, 4;
add.s64 %rd157, %rd155, %rd156;
ld.global.f32 %f119, [%rd157];
add.ftz.f32 %f24, %f24, %f119;

$L__BB304_18:
ld.param.u32 %r950, [_Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.gt.s32 %p43, %r950, 2;
@%p43 bra $L__BB304_22;

ld.param.u32 %r953, [_Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.eq.s32 %p46, %r953, 1;
@%p46 bra $L__BB304_26;

ld.param.u32 %r954, [_Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.eq.s32 %p47, %r954, 2;
@%p47 bra $L__BB304_21;
bra.uni $L__BB304_27;

$L__BB304_21:
mul.ftz.f32 %f139, %f24, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f140, %f139;
add.ftz.f32 %f141, %f140, 0f3F800000;
div.approx.ftz.f32 %f142, %f24, %f141;
mul.ftz.f32 %f168, %f168, %f142;
bra.uni $L__BB304_28;

$L__BB304_22:
ld.param.u32 %r951, [_Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.eq.s32 %p44, %r951, 3;
@%p44 bra $L__BB304_25;

ld.param.u32 %r952, [_Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+40];
setp.ne.s32 %p45, %r952, 6;
@%p45 bra $L__BB304_27;

ld.param.f32 %f151, [_Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3+44];
min.ftz.f32 %f120, %f24, %f151;
neg.ftz.f32 %f121, %f151;
min.ftz.f32 %f122, %f168, %f151;
max.ftz.f32 %f123, %f122, %f121;
mul.ftz.f32 %f124, %f120, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f125, %f124;
add.ftz.f32 %f126, %f125, 0f3F800000;
div.approx.ftz.f32 %f127, %f120, %f126;
mul.ftz.f32 %f168, %f123, %f127;
bra.uni $L__BB304_28;

$L__BB304_27:
mul.ftz.f32 %f168, %f168, %f24;
bra.uni $L__BB304_28;

$L__BB304_26:
mul.ftz.f32 %f143, %f24, 0f3F000000;
mul.ftz.f32 %f144, %f24, 0f3D372713;
fma.rn.ftz.f32 %f145, %f24, %f144, 0f3F800000;
mul.ftz.f32 %f146, %f24, 0f3F4C422A;
mul.ftz.f32 %f147, %f146, %f145;
tanh.approx.f32 %f148, %f147;
add.ftz.f32 %f149, %f148, 0f3F800000;
mul.ftz.f32 %f150, %f143, %f149;
mul.ftz.f32 %f168, %f168, %f150;
bra.uni $L__BB304_28;

$L__BB304_25:
mov.f32 %f128, 0f40E00000;
min.ftz.f32 %f129, %f24, %f128;
min.ftz.f32 %f130, %f168, %f128;
mov.f32 %f131, 0fC0E00000;
max.ftz.f32 %f132, %f130, %f131;
mul.ftz.f32 %f133, %f129, 0f3FD9DB23;
mul.ftz.f32 %f134, %f133, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f135, %f134;
add.ftz.f32 %f136, %f135, 0f3F800000;
div.approx.ftz.f32 %f137, %f129, %f136;
add.ftz.f32 %f138, %f132, 0f3F800000;
mul.ftz.f32 %f168, %f138, %f137;

$L__BB304_28:
ld.param.u64 %rd164, [_Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4];
ld.param.u32 %r948, [_Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10];
ld.param.u32 %r947, [_Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13];
mov.u32 %r946, %ctaid.y;
mov.u32 %r945, %tid.y;
mad.lo.s32 %r939, %r945, %r948, %r21;
mad.lo.s32 %r940, %r946, %r947, %r939;
cvta.to.global.u64 %rd158, %rd164;
mul.wide.u32 %rd159, %r940, 4;
add.s64 %rd160, %rd158, %rd159;
st.global.f32 [%rd160], %f168;

$L__BB304_29:
ret;

}
.entry _Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj(
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1,
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2,
.param .align 8 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_3[48],
.param .u64 _Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5,
.param .align 4 .b8 _Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6[12],
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14,
.param .u32 _Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15
)
.maxntid 256, 1, 1
.minnctapersm 1
{
.local .align 8 .b8 __local_depot305[8];
.reg .b64 %SP;
.reg .b64 %SPL;
.reg .pred %p<17>;
.reg .b16 %rs<47>;
.reg .f32 %f<44>;
.reg .b32 %r<483>;
.reg .b64 %rd<95>;


mov.u64 %SPL, __local_depot305;
ld.param.u64 %rd17, [_Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_0];
ld.param.u64 %rd18, [_Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_1];
ld.param.u64 %rd19, [_Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_2];
ld.param.u64 %rd20, [_Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_4];
ld.param.u32 %r11, [_Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_5];
ld.param.u32 %r15, [_Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_7];
ld.param.u32 %r16, [_Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_8];
ld.param.u32 %r17, [_Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_9];
ld.param.u32 %r18, [_Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_10];
ld.param.u32 %r19, [_Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_11];
ld.param.u32 %r20, [_Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_12];
ld.param.u32 %r21, [_Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_13];
ld.param.u32 %r23, [_Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_14];
ld.param.u32 %r22, [_Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_15];
ld.param.u32 %r14, [_Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+8];
ld.param.u32 %r13, [_Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6+4];
ld.param.u32 %r12, [_Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj_param_6];
add.u64 %rd1, %SPL, 0;
mov.u32 %r1, %tid.y;
setp.ge.u32 %p1, %r1, %r23;
@%p1 bra $L__BB305_7;

mov.f32 %f42, 0f00000000;
st.local.v2.f32 [%rd1], {%f42, %f42};
mov.u32 %r2, %tid.x;
shr.u32 %r482, %r2, 3;
shr.u32 %r24, %r11, 8;
setp.ge.u32 %p2, %r482, %r24;
mov.f32 %f43, %f42;
@%p2 bra $L__BB305_5;

mov.u32 %r25, %ctaid.x;
shl.b32 %r26, %r25, 1;
mov.u32 %r27, %ctaid.y;
mad.lo.s32 %r28, %r1, %r22, %r27;
cvta.to.global.u64 %rd22, %rd19;
mul.wide.u32 %rd23, %r28, 4;
add.s64 %rd24, %rd22, %rd23;
ld.global.nc.u32 %r29, [%rd24];
mul.lo.s32 %r30, %r29, %r19;
and.b32 %r31, %r2, 7;
shl.b32 %r32, %r2, 1;
and.b32 %r33, %r32, 14;
shr.u32 %r34, %r31, 1;
cvt.u64.u32 %rd25, %r34;
shl.b32 %r35, %r2, 2;
and.b32 %r36, %r35, 28;
or.b32 %r37, %r36, 1;
sub.s32 %r38, %r37, %r33;
shl.b32 %r39, %r38, 1;
cvt.s64.s32 %rd26, %r39;
cvt.u64.u32 %rd2, %r31;
and.b32 %r4, %r35, 4;
shl.b32 %r481, %r482, 3;
or.b32 %r40, %r26, 1;
add.s32 %r41, %r482, %r30;
mad.lo.s32 %r42, %r16, %r40, %r41;
mul.wide.s32 %rd3, %r42, 110;
add.s64 %rd27, %rd25, %rd3;
add.s64 %rd4, %rd27, 106;
shl.b32 %r43, %r2, 3;
cvt.u64.u32 %rd28, %r43;
and.b64 %rd29, %rd28, 56;
mad.lo.s32 %r44, %r26, %r16, %r482;
add.s32 %r45, %r44, %r30;
mul.wide.s32 %rd5, %r45, 110;
add.s64 %rd30, %rd29, %rd5;
add.s64 %rd6, %rd30, 8;
add.s64 %rd31, %rd26, -2;
add.s64 %rd32, %rd31, %rd3;
add.s64 %rd7, %rd32, 74;
add.s64 %rd33, %rd3, %rd2;
add.s64 %rd8, %rd33, 66;
add.s64 %rd34, %rd5, %rd2;
add.s64 %rd9, %rd34, 66;
add.s64 %rd35, %rd31, %rd5;
add.s64 %rd10, %rd35, 76;
add.s64 %rd36, %rd29, %rd3;
add.s64 %rd11, %rd36, 4;
add.s64 %rd37, %rd25, %rd5;
add.s64 %rd12, %rd37, 106;
cvta.to.global.u64 %rd94, %rd17;
mul.lo.s32 %r46, %r1, %r17;
cvt.u64.u32 %rd38, %r46;
mul.hi.u32 %r47, %r27, %r12;
add.s32 %r48, %r47, %r27;
shr.u32 %r49, %r48, %r13;
mul.lo.s32 %r50, %r49, %r14;
sub.s32 %r51, %r27, %r50;
mul.lo.s32 %r52, %r51, %r20;
cvt.u64.u32 %rd39, %r52;
add.s64 %rd14, %rd39, %rd38;
mov.f32 %f42, 0f00000000;
cvta.to.global.u64 %rd40, %rd18;

$L__BB305_3:
cvt.s64.s32 %rd41, %r481;
add.s64 %rd42, %rd14, %rd41;
add.s64 %rd43, %rd42, %rd2;
mul.lo.s64 %rd44, %rd43, 36;
add.s64 %rd45, %rd94, %rd6;
ld.global.nc.u16 %rs3, [%rd45+-4];
ld.global.nc.u16 %rs4, [%rd45+-2];
ld.global.nc.u16 %rs5, [%rd45];
add.s64 %rd46, %rd94, %rd9;
ld.global.nc.u8 %rs6, [%rd46];
add.s64 %rd47, %rd94, %rd10;
ld.global.nc.u16 %rs7, [%rd47+-2];
ld.global.nc.u16 %rs8, [%rd47];
ld.global.nc.u16 %rs9, [%rd45+-6];
and.b16 %rs10, %rs9, 255;
cvt.u32.u16 %r151, %rs10;
shl.b16 %rs11, %rs6, 8;
cvt.u32.u16 %r152, %rs11;
and.b32 %r153, %r152, 256;
or.b32 %r154, %r153, %r151;
mul.wide.u32 %rd48, %r154, 4;
mov.u64 %rd49, iq3s_grid;
add.s64 %rd50, %rd49, %rd48;
ld.global.nc.u32 %r155, [%rd50];
shr.u16 %rs12, %rs9, 8;
cvt.u32.u16 %r156, %rs12;
mul.wide.u16 %r157, %rs6, 128;
and.b32 %r158, %r157, 256;
or.b32 %r159, %r158, %r156;
mul.wide.u32 %rd51, %r159, 4;
add.s64 %rd52, %rd49, %rd51;
ld.global.nc.u32 %r160, [%rd52];
and.b16 %rs14, %rs7, 255;
mul.wide.u16 %r161, %rs14, 128;
cvt.u32.u16 %r162, %rs7;
shl.b32 %r163, %r162, 21;
or.b32 %r164, %r161, %r163;
and.b32 %r165, %r164, 25166208;
or.b32 %r166, %r165, -2139062144;
add.s32 %r167, %r166, -16843009;
or.b32 %r53, %r167, %r165;

prmt.b32 %r53, %r53, 0, 0xba98;

mul.wide.u16 %r168, %rs14, 8;
shl.b32 %r169, %r162, 17;
or.b32 %r170, %r168, %r169;
and.b32 %r171, %r170, 25166208;
or.b32 %r172, %r171, -2139062144;
add.s32 %r173, %r172, -16843009;
or.b32 %r55, %r173, %r171;

prmt.b32 %r55, %r55, 0, 0xba98;

xor.b32 %r174, %r53, %r155;
or.b32 %r175, %r174, -2139062144;
and.b32 %r176, %r53, 2139062143;
and.b32 %r177, %r155, -2139062144;
sub.s32 %r178, %r175, %r176;
xor.b32 %r179, %r177, %r178;
xor.b32 %r58, %r179, -2139062144;
xor.b32 %r180, %r55, %r160;
or.b32 %r181, %r180, -2139062144;
and.b32 %r182, %r55, 2139062143;
and.b32 %r183, %r160, -2139062144;
sub.s32 %r184, %r181, %r182;
xor.b32 %r185, %r183, %r184;
xor.b32 %r62, %r185, -2139062144;
add.s64 %rd53, %rd40, %rd44;
ld.global.nc.u32 %r59, [%rd53+4];
ld.global.nc.u32 %r63, [%rd53+8];
mov.u32 %r109, 0;

dp4a.s32.s32 %r57, %r58, %r59, %r109;


dp4a.s32.s32 %r61, %r62, %r63, %r57;

and.b16 %rs15, %rs3, 255;
cvt.u32.u16 %r186, %rs15;
mul.wide.u16 %r187, %rs6, 64;
and.b32 %r188, %r187, 256;
or.b32 %r189, %r188, %r186;
mul.wide.u32 %rd54, %r189, 4;
add.s64 %rd55, %rd49, %rd54;
ld.global.nc.u32 %r190, [%rd55];
shr.u16 %rs16, %rs3, 8;
cvt.u32.u16 %r191, %rs16;
mul.wide.u16 %r192, %rs6, 32;
and.b32 %r193, %r192, 256;
or.b32 %r194, %r193, %r191;
mul.wide.u32 %rd56, %r194, 4;
add.s64 %rd57, %rd49, %rd56;
ld.global.nc.u32 %r195, [%rd57];
shr.u16 %rs17, %rs7, 8;
mul.wide.u16 %r196, %rs17, 128;
cvt.u32.u16 %r197, %rs17;
shl.b32 %r198, %r197, 21;
or.b32 %r199, %r196, %r198;
and.b32 %r200, %r199, 25166208;
or.b32 %r201, %r200, -2139062144;
add.s32 %r202, %r201, -16843009;
or.b32 %r65, %r202, %r200;

prmt.b32 %r65, %r65, 0, 0xba98;

mul.wide.u16 %r203, %rs17, 8;
shl.b32 %r204, %r197, 17;
or.b32 %r205, %r203, %r204;
and.b32 %r206, %r205, 25166208;
or.b32 %r207, %r206, -2139062144;
add.s32 %r208, %r207, -16843009;
or.b32 %r67, %r208, %r206;

prmt.b32 %r67, %r67, 0, 0xba98;

xor.b32 %r209, %r65, %r190;
or.b32 %r210, %r209, -2139062144;
and.b32 %r211, %r65, 2139062143;
and.b32 %r212, %r190, -2139062144;
sub.s32 %r213, %r210, %r211;
xor.b32 %r214, %r212, %r213;
xor.b32 %r70, %r214, -2139062144;
xor.b32 %r215, %r67, %r195;
or.b32 %r216, %r215, -2139062144;
and.b32 %r217, %r67, 2139062143;
and.b32 %r218, %r195, -2139062144;
sub.s32 %r219, %r216, %r217;
xor.b32 %r220, %r218, %r219;
xor.b32 %r74, %r220, -2139062144;
ld.global.nc.u32 %r71, [%rd53+12];
ld.global.nc.u32 %r75, [%rd53+16];

dp4a.s32.s32 %r69, %r70, %r71, %r61;


dp4a.s32.s32 %r73, %r74, %r75, %r69;

and.b16 %rs18, %rs4, 255;
cvt.u32.u16 %r221, %rs18;
mul.wide.u16 %r222, %rs6, 16;
and.b32 %r223, %r222, 256;
or.b32 %r224, %r223, %r221;
mul.wide.u32 %rd58, %r224, 4;
add.s64 %rd59, %rd49, %rd58;
ld.global.nc.u32 %r225, [%rd59];
shr.u16 %rs19, %rs4, 8;
cvt.u32.u16 %r226, %rs19;
mul.wide.u16 %r227, %rs6, 8;
and.b32 %r228, %r227, 256;
or.b32 %r229, %r228, %r226;
mul.wide.u32 %rd60, %r229, 4;
add.s64 %rd61, %rd49, %rd60;
ld.global.nc.u32 %r230, [%rd61];
and.b16 %rs20, %rs8, 255;
mul.wide.u16 %r231, %rs20, 128;
cvt.u32.u16 %r232, %rs8;
shl.b32 %r233, %r232, 21;
or.b32 %r234, %r231, %r233;
and.b32 %r235, %r234, 25166208;
or.b32 %r236, %r235, -2139062144;
add.s32 %r237, %r236, -16843009;
or.b32 %r77, %r237, %r235;

prmt.b32 %r77, %r77, 0, 0xba98;

mul.wide.u16 %r238, %rs20, 8;
shl.b32 %r239, %r232, 17;
or.b32 %r240, %r238, %r239;
and.b32 %r241, %r240, 25166208;
or.b32 %r242, %r241, -2139062144;
add.s32 %r243, %r242, -16843009;
or.b32 %r79, %r243, %r241;

prmt.b32 %r79, %r79, 0, 0xba98;

xor.b32 %r244, %r77, %r225;
or.b32 %r245, %r244, -2139062144;
and.b32 %r246, %r77, 2139062143;
and.b32 %r247, %r225, -2139062144;
sub.s32 %r248, %r245, %r246;
xor.b32 %r249, %r247, %r248;
xor.b32 %r82, %r249, -2139062144;
xor.b32 %r250, %r79, %r230;
or.b32 %r251, %r250, -2139062144;
and.b32 %r252, %r79, 2139062143;
and.b32 %r253, %r230, -2139062144;
sub.s32 %r254, %r251, %r252;
xor.b32 %r255, %r253, %r254;
xor.b32 %r86, %r255, -2139062144;
ld.global.nc.u32 %r83, [%rd53+20];
ld.global.nc.u32 %r87, [%rd53+24];

dp4a.s32.s32 %r81, %r82, %r83, %r73;


dp4a.s32.s32 %r85, %r86, %r87, %r81;

and.b16 %rs21, %rs5, 255;
cvt.u32.u16 %r256, %rs21;
mul.wide.u16 %r257, %rs6, 4;
and.b32 %r258, %r257, 256;
or.b32 %r259, %r258, %r256;
mul.wide.u32 %rd62, %r259, 4;
add.s64 %rd63, %rd49, %rd62;
ld.global.nc.u32 %r260, [%rd63];
shr.u16 %rs22, %rs5, 8;
cvt.u32.u16 %r261, %rs22;
mul.wide.u16 %r262, %rs6, 2;
and.b32 %r263, %r262, 256;
or.b32 %r264, %r263, %r261;
mul.wide.u32 %rd64, %r264, 4;
add.s64 %rd65, %rd49, %rd64;
ld.global.nc.u32 %r265, [%rd65];
shr.u16 %rs23, %rs8, 8;
mul.wide.u16 %r266, %rs23, 128;
cvt.u32.u16 %r267, %rs23;
shl.b32 %r268, %r267, 21;
or.b32 %r269, %r266, %r268;
and.b32 %r270, %r269, 25166208;
or.b32 %r271, %r270, -2139062144;
add.s32 %r272, %r271, -16843009;
or.b32 %r89, %r272, %r270;

prmt.b32 %r89, %r89, 0, 0xba98;

mul.wide.u16 %r273, %rs23, 8;
shl.b32 %r274, %r267, 17;
or.b32 %r275, %r273, %r274;
and.b32 %r276, %r275, 25166208;
or.b32 %r277, %r276, -2139062144;
add.s32 %r278, %r277, -16843009;
or.b32 %r91, %r278, %r276;

prmt.b32 %r91, %r91, 0, 0xba98;

xor.b32 %r279, %r89, %r260;
or.b32 %r280, %r279, -2139062144;
and.b32 %r281, %r89, 2139062143;
and.b32 %r282, %r260, -2139062144;
sub.s32 %r283, %r280, %r281;
xor.b32 %r284, %r282, %r283;
xor.b32 %r94, %r284, -2139062144;
xor.b32 %r285, %r91, %r265;
or.b32 %r286, %r285, -2139062144;
and.b32 %r287, %r91, 2139062143;
and.b32 %r288, %r265, -2139062144;
sub.s32 %r289, %r286, %r287;
xor.b32 %r290, %r288, %r289;
xor.b32 %r98, %r290, -2139062144;
ld.global.nc.u32 %r95, [%rd53+28];
ld.global.nc.u32 %r99, [%rd53+32];

dp4a.s32.s32 %r93, %r94, %r95, %r85;


dp4a.s32.s32 %r97, %r98, %r99, %r93;

add.s64 %rd66, %rd94, %rd12;
ld.global.nc.u8 %rs24, [%rd66];
cvt.u32.u16 %r291, %rs24;
and.b32 %r292, %r291, 255;
shr.u32 %r293, %r292, %r4;
shl.b32 %r294, %r293, 1;
and.b32 %r295, %r294, 30;
or.b32 %r296, %r295, 1;
mul.lo.s32 %r297, %r296, %r97;
add.s64 %rd67, %rd94, %rd5;
ld.global.nc.u16 %rs1, [%rd67];

{ cvt.f32.f16 %f11, %rs1;}


ld.global.nc.u32 %r101, [%rd53];

{.reg .f16 low,high;
mov.b32 {low,high},%r101;
cvt.f32.f16 %f12, low;}


mul.ftz.f32 %f15, %f11, %f12;
cvt.rn.f32.s32 %f16, %r297;
fma.rn.ftz.f32 %f43, %f15, %f16, %f43;
add.s64 %rd68, %rd94, %rd11;
ld.global.nc.u16 %rs25, [%rd68];
ld.global.nc.u16 %rs26, [%rd68+2];
ld.global.nc.u16 %rs27, [%rd68+4];
add.s64 %rd69, %rd94, %rd8;
ld.global.nc.u8 %rs28, [%rd69];
add.s64 %rd70, %rd94, %rd7;
ld.global.nc.u16 %rs29, [%rd70];
ld.global.nc.u16 %rs30, [%rd70+2];
ld.global.nc.u16 %rs31, [%rd68+-2];
and.b16 %rs32, %rs31, 255;
cvt.u32.u16 %r298, %rs32;
shl.b16 %rs33, %rs28, 8;
cvt.u32.u16 %r299, %rs33;
and.b32 %r300, %r299, 256;
or.b32 %r301, %r300, %r298;
mul.wide.u32 %rd71, %r301, 4;
add.s64 %rd72, %rd49, %rd71;
ld.global.nc.u32 %r302, [%rd72];
shr.u16 %rs34, %rs31, 8;
cvt.u32.u16 %r303, %rs34;
mul.wide.u16 %r304, %rs28, 128;
and.b32 %r305, %r304, 256;
or.b32 %r306, %r305, %r303;
mul.wide.u32 %rd73, %r306, 4;
add.s64 %rd74, %rd49, %rd73;
ld.global.nc.u32 %r307, [%rd74];
and.b16 %rs36, %rs29, 255;
mul.wide.u16 %r308, %rs36, 128;
cvt.u32.u16 %r309, %rs29;
shl.b32 %r310, %r309, 21;
or.b32 %r311, %r308, %r310;
and.b32 %r312, %r311, 25166208;
or.b32 %r313, %r312, -2139062144;
add.s32 %r314, %r313, -16843009;
or.b32 %r102, %r314, %r312;

prmt.b32 %r102, %r102, 0, 0xba98;

mul.wide.u16 %r315, %rs36, 8;
shl.b32 %r316, %r309, 17;
or.b32 %r317, %r315, %r316;
and.b32 %r318, %r317, 25166208;
or.b32 %r319, %r318, -2139062144;
add.s32 %r320, %r319, -16843009;
or.b32 %r104, %r320, %r318;

prmt.b32 %r104, %r104, 0, 0xba98;

xor.b32 %r321, %r102, %r302;
or.b32 %r322, %r321, -2139062144;
and.b32 %r323, %r102, 2139062143;
and.b32 %r324, %r302, -2139062144;
sub.s32 %r325, %r322, %r323;
xor.b32 %r326, %r324, %r325;
xor.b32 %r107, %r326, -2139062144;
xor.b32 %r327, %r104, %r307;
or.b32 %r328, %r327, -2139062144;
and.b32 %r329, %r104, 2139062143;
and.b32 %r330, %r307, -2139062144;
sub.s32 %r331, %r328, %r329;
xor.b32 %r332, %r330, %r331;
xor.b32 %r111, %r332, -2139062144;

dp4a.s32.s32 %r106, %r107, %r59, %r109;


dp4a.s32.s32 %r110, %r111, %r63, %r106;

and.b16 %rs37, %rs25, 255;
cvt.u32.u16 %r333, %rs37;
mul.wide.u16 %r334, %rs28, 64;
and.b32 %r335, %r334, 256;
or.b32 %r336, %r335, %r333;
mul.wide.u32 %rd75, %r336, 4;
add.s64 %rd76, %rd49, %rd75;
ld.global.nc.u32 %r337, [%rd76];
shr.u16 %rs38, %rs25, 8;
cvt.u32.u16 %r338, %rs38;
mul.wide.u16 %r339, %rs28, 32;
and.b32 %r340, %r339, 256;
or.b32 %r341, %r340, %r338;
mul.wide.u32 %rd77, %r341, 4;
add.s64 %rd78, %rd49, %rd77;
ld.global.nc.u32 %r342, [%rd78];
shr.u16 %rs39, %rs29, 8;
mul.wide.u16 %r343, %rs39, 128;
cvt.u32.u16 %r344, %rs39;
shl.b32 %r345, %r344, 21;
or.b32 %r346, %r343, %r345;
and.b32 %r347, %r346, 25166208;
or.b32 %r348, %r347, -2139062144;
add.s32 %r349, %r348, -16843009;
or.b32 %r114, %r349, %r347;

prmt.b32 %r114, %r114, 0, 0xba98;

mul.wide.u16 %r350, %rs39, 8;
shl.b32 %r351, %r344, 17;
or.b32 %r352, %r350, %r351;
and.b32 %r353, %r352, 25166208;
or.b32 %r354, %r353, -2139062144;
add.s32 %r355, %r354, -16843009;
or.b32 %r116, %r355, %r353;

prmt.b32 %r116, %r116, 0, 0xba98;

xor.b32 %r356, %r114, %r337;
or.b32 %r357, %r356, -2139062144;
and.b32 %r358, %r114, 2139062143;
and.b32 %r359, %r337, -2139062144;
sub.s32 %r360, %r357, %r358;
xor.b32 %r361, %r359, %r360;
xor.b32 %r119, %r361, -2139062144;
xor.b32 %r362, %r116, %r342;
or.b32 %r363, %r362, -2139062144;
and.b32 %r364, %r116, 2139062143;
and.b32 %r365, %r342, -2139062144;
sub.s32 %r366, %r363, %r364;
xor.b32 %r367, %r365, %r366;
xor.b32 %r123, %r367, -2139062144;

dp4a.s32.s32 %r118, %r119, %r71, %r110;


dp4a.s32.s32 %r122, %r123, %r75, %r118;

and.b16 %rs40, %rs26, 255;
cvt.u32.u16 %r368, %rs40;
mul.wide.u16 %r369, %rs28, 16;
and.b32 %r370, %r369, 256;
or.b32 %r371, %r370, %r368;
mul.wide.u32 %rd79, %r371, 4;
add.s64 %rd80, %rd49, %rd79;
ld.global.nc.u32 %r372, [%rd80];
shr.u16 %rs41, %rs26, 8;
cvt.u32.u16 %r373, %rs41;
mul.wide.u16 %r374, %rs28, 8;
and.b32 %r375, %r374, 256;
or.b32 %r376, %r375, %r373;
mul.wide.u32 %rd81, %r376, 4;
add.s64 %rd82, %rd49, %rd81;
ld.global.nc.u32 %r377, [%rd82];
and.b16 %rs42, %rs30, 255;
mul.wide.u16 %r378, %rs42, 128;
cvt.u32.u16 %r379, %rs30;
shl.b32 %r380, %r379, 21;
or.b32 %r381, %r378, %r380;
and.b32 %r382, %r381, 25166208;
or.b32 %r383, %r382, -2139062144;
add.s32 %r384, %r383, -16843009;
or.b32 %r126, %r384, %r382;

prmt.b32 %r126, %r126, 0, 0xba98;

mul.wide.u16 %r385, %rs42, 8;
shl.b32 %r386, %r379, 17;
or.b32 %r387, %r385, %r386;
and.b32 %r388, %r387, 25166208;
or.b32 %r389, %r388, -2139062144;
add.s32 %r390, %r389, -16843009;
or.b32 %r128, %r390, %r388;

prmt.b32 %r128, %r128, 0, 0xba98;

xor.b32 %r391, %r126, %r372;
or.b32 %r392, %r391, -2139062144;
and.b32 %r393, %r126, 2139062143;
and.b32 %r394, %r372, -2139062144;
sub.s32 %r395, %r392, %r393;
xor.b32 %r396, %r394, %r395;
xor.b32 %r131, %r396, -2139062144;
xor.b32 %r397, %r128, %r377;
or.b32 %r398, %r397, -2139062144;
and.b32 %r399, %r128, 2139062143;
and.b32 %r400, %r377, -2139062144;
sub.s32 %r401, %r398, %r399;
xor.b32 %r402, %r400, %r401;
xor.b32 %r135, %r402, -2139062144;

dp4a.s32.s32 %r130, %r131, %r83, %r122;


dp4a.s32.s32 %r134, %r135, %r87, %r130;

and.b16 %rs43, %rs27, 255;
cvt.u32.u16 %r403, %rs43;
mul.wide.u16 %r404, %rs28, 4;
and.b32 %r405, %r404, 256;
or.b32 %r406, %r405, %r403;
mul.wide.u32 %rd83, %r406, 4;
add.s64 %rd84, %rd49, %rd83;
ld.global.nc.u32 %r407, [%rd84];
shr.u16 %rs44, %rs27, 8;
cvt.u32.u16 %r408, %rs44;
mul.wide.u16 %r409, %rs28, 2;
and.b32 %r410, %r409, 256;
or.b32 %r411, %r410, %r408;
mul.wide.u32 %rd85, %r411, 4;
add.s64 %rd86, %rd49, %rd85;
ld.global.nc.u32 %r412, [%rd86];
shr.u16 %rs45, %rs30, 8;
mul.wide.u16 %r413, %rs45, 128;
cvt.u32.u16 %r414, %rs45;
shl.b32 %r415, %r414, 21;
or.b32 %r416, %r413, %r415;
and.b32 %r417, %r416, 25166208;
or.b32 %r418, %r417, -2139062144;
add.s32 %r419, %r418, -16843009;
or.b32 %r138, %r419, %r417;

prmt.b32 %r138, %r138, 0, 0xba98;

mul.wide.u16 %r420, %rs45, 8;
shl.b32 %r421, %r414, 17;
or.b32 %r422, %r420, %r421;
and.b32 %r423, %r422, 25166208;
or.b32 %r424, %r423, -2139062144;
add.s32 %r425, %r424, -16843009;
or.b32 %r140, %r425, %r423;

prmt.b32 %r140, %r140, 0, 0xba98;

xor.b32 %r426, %r138, %r407;
or.b32 %r427, %r426, -2139062144;
and.b32 %r428, %r138, 2139062143;
and.b32 %r429, %r407, -2139062144;
sub.s32 %r430, %r427, %r428;
xor.b32 %r431, %r429, %r430;
xor.b32 %r143, %r431, -2139062144;
xor.b32 %r432, %r140, %r412;
or.b32 %r433, %r432, -2139062144;
and.b32 %r434, %r140, 2139062143;
and.b32 %r435, %r412, -2139062144;
sub.s32 %r436, %r433, %r434;
xor.b32 %r437, %r435, %r436;
xor.b32 %r147, %r437, -2139062144;

dp4a.s32.s32 %r142, %r143, %r95, %r134;


dp4a.s32.s32 %r146, %r147, %r99, %r142;

add.s64 %rd87, %rd94, %rd4;
ld.global.nc.u8 %rs46, [%rd87];
cvt.u32.u16 %r438, %rs46;
and.b32 %r439, %r438, 255;
shr.u32 %r440, %r439, %r4;
shl.b32 %r441, %r440, 1;
and.b32 %r442, %r441, 30;
or.b32 %r443, %r442, 1;
mul.lo.s32 %r444, %r443, %r146;
add.s64 %rd88, %rd94, %rd3;
ld.global.nc.u16 %rs2, [%rd88];

{ cvt.f32.f16 %f13, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r101;
cvt.f32.f16 %f14, low;}


mul.ftz.f32 %f17, %f13, %f14;
cvt.rn.f32.s32 %f18, %r444;
fma.rn.ftz.f32 %f42, %f17, %f18, %f42;
add.s32 %r481, %r481, 32;
add.s64 %rd94, %rd94, 440;
add.s32 %r482, %r482, 4;
setp.lt.u32 %p3, %r482, %r24;
@%p3 bra $L__BB305_3;

st.local.v2.f32 [%rd1], {%f43, %f42};

$L__BB305_5:
mov.b32 %r446, %f43;
mov.u32 %r447, 31;
mov.u32 %r448, 16;
mov.u32 %r449, -1;
shfl.sync.bfly.b32 %r450|%p4, %r446, %r448, %r447, %r449;
mov.b32 %f19, %r450;
add.ftz.f32 %f20, %f43, %f19;
mov.b32 %r451, %f20;
mov.u32 %r452, 8;
shfl.sync.bfly.b32 %r453|%p5, %r451, %r452, %r447, %r449;
mov.b32 %f21, %r453;
add.ftz.f32 %f22, %f20, %f21;
mov.b32 %r454, %f22;
mov.u32 %r455, 4;
shfl.sync.bfly.b32 %r456|%p6, %r454, %r455, %r447, %r449;
mov.b32 %f23, %r456;
add.ftz.f32 %f24, %f22, %f23;
mov.b32 %r457, %f24;
mov.u32 %r458, 2;
shfl.sync.bfly.b32 %r459|%p7, %r457, %r458, %r447, %r449;
mov.b32 %f25, %r459;
add.ftz.f32 %f26, %f24, %f25;
mov.b32 %r460, %f26;
mov.u32 %r461, 1;
shfl.sync.bfly.b32 %r462|%p8, %r460, %r461, %r447, %r449;
mov.b32 %f27, %r462;
add.ftz.f32 %f28, %f26, %f27;
st.local.f32 [%rd1], %f28;
mov.b32 %r463, %f42;
shfl.sync.bfly.b32 %r464|%p9, %r463, %r448, %r447, %r449;
mov.b32 %f29, %r464;
add.ftz.f32 %f30, %f42, %f29;
mov.b32 %r465, %f30;
shfl.sync.bfly.b32 %r466|%p10, %r465, %r452, %r447, %r449;
mov.b32 %f31, %r466;
add.ftz.f32 %f32, %f30, %f31;
mov.b32 %r467, %f32;
shfl.sync.bfly.b32 %r468|%p11, %r467, %r455, %r447, %r449;
mov.b32 %f33, %r468;
add.ftz.f32 %f34, %f32, %f33;
mov.b32 %r469, %f34;
shfl.sync.bfly.b32 %r470|%p12, %r469, %r458, %r447, %r449;
mov.b32 %f35, %r470;
add.ftz.f32 %f36, %f34, %f35;
mov.b32 %r471, %f36;
shfl.sync.bfly.b32 %r472|%p13, %r471, %r461, %r447, %r449;
mov.b32 %f37, %r472;
add.ftz.f32 %f38, %f36, %f37;
st.local.f32 [%rd1+4], %f38;
setp.gt.u32 %p14, %r2, 1;
mov.u32 %r474, %ctaid.x;
shl.b32 %r475, %r474, 1;
add.s32 %r10, %r475, %r2;
setp.ge.u32 %p15, %r10, %r15;
or.pred %p16, %p14, %p15;
@%p16 bra $L__BB305_7;

mul.wide.u32 %rd89, %r2, 4;
add.s64 %rd90, %rd1, %rd89;
ld.local.f32 %f39, [%rd90];
mov.u32 %r477, %ctaid.y;
mad.lo.s32 %r479, %r1, %r18, %r10;
mad.lo.s32 %r480, %r477, %r21, %r479;
cvta.to.global.u64 %rd91, %rd20;
mul.wide.u32 %rd92, %r480, 4;
add.s64 %rd93, %rd91, %rd92;
st.global.f32 [%rd93], %f39;

$L__BB305_7:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<320>;
.reg .b16 %rs<177>;
.reg .f32 %f<1005>;
.reg .b32 %r<2434>;
.reg .b64 %rd<311>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate[1536];

ld.param.u64 %rd48, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd49, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd50, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u32 %r49, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r53, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r54, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r58, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r59, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r60, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r64, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r65, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r66, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r62, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r61, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r56, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r55, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r52, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r51, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r50, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
ld.param.u64 %rd53, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+16];
ld.param.u64 %rd52, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+8];
ld.param.u64 %rd51, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3];
mov.u32 %r1, %tid.y;
shl.b32 %r67, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r67, %r2;
mov.u32 %r68, %ctaid.x;
shl.b32 %r4, %r68, 2;
mov.u32 %r5, %ctaid.y;
setp.eq.s64 %p1, %rd50, 0;
@%p1 bra $L__BB306_2;

cvta.to.global.u64 %rd57, %rd50;
mul.wide.u32 %rd58, %r5, 4;
add.s64 %rd59, %rd57, %rd58;
ld.global.nc.u32 %r2430, [%rd59];
mul.hi.u32 %r69, %r5, %r50;
add.s32 %r70, %r69, %r5;
shr.u32 %r71, %r70, %r51;
mul.lo.s32 %r72, %r71, %r52;
sub.s32 %r2431, %r5, %r72;
bra.uni $L__BB306_3;

$L__BB306_2:
mul.hi.u32 %r73, %r5, %r55;
add.s32 %r74, %r73, %r5;
shr.u32 %r2430, %r74, %r56;
mov.u32 %r2431, %r5;

$L__BB306_3:
mov.u32 %r11, %ctaid.z;
selp.b32 %r12, %r5, %r2430, %p1;
setp.ne.s32 %p3, %r1, 0;
setp.gt.u32 %p4, %r2, 3;
or.pred %p5, %p4, %p3;
add.s32 %r75, %r4, %r2;
setp.ge.u32 %p6, %r75, %r54;
mov.f32 %f27, 0f00000000;
or.pred %p7, %p5, %p6;
mov.f32 %f4, %f27;
mov.f32 %f5, %f27;
@%p7 bra $L__BB306_8;

setp.eq.s64 %p8, %rd51, 0;
mov.f32 %f4, 0f00000000;
mov.f32 %f5, %f4;
@%p8 bra $L__BB306_6;

cvta.to.global.u64 %rd60, %rd51;
mul.lo.s32 %r76, %r11, %r66;
cvt.u64.u32 %rd61, %r76;
mul.lo.s32 %r77, %r12, %r60;
cvt.u64.u32 %rd62, %r77;
cvt.u64.u32 %rd63, %r2;
cvt.s64.s32 %rd64, %r4;
add.s64 %rd65, %rd64, %rd63;
add.s64 %rd66, %rd65, %rd61;
add.s64 %rd67, %rd66, %rd62;
shl.b64 %rd68, %rd67, 2;
add.s64 %rd69, %rd60, %rd68;
ld.global.f32 %f5, [%rd69];

$L__BB306_6:
setp.eq.s64 %p9, %rd53, 0;
setp.eq.s64 %p10, %rd52, 0;
or.pred %p11, %p9, %p10;
@%p11 bra $L__BB306_8;

cvta.to.global.u64 %rd70, %rd53;
mul.lo.s32 %r78, %r11, %r66;
cvt.u64.u32 %rd71, %r78;
mul.lo.s32 %r79, %r12, %r60;
cvt.u64.u32 %rd72, %r79;
cvt.u64.u32 %rd73, %r2;
cvt.s64.s32 %rd74, %r4;
add.s64 %rd75, %rd74, %rd73;
add.s64 %rd76, %rd75, %rd71;
add.s64 %rd77, %rd76, %rd72;
shl.b64 %rd78, %rd77, 2;
add.s64 %rd79, %rd70, %rd78;
ld.global.f32 %f4, [%rd79];

$L__BB306_8:
mul.lo.s32 %r80, %r11, %r65;
cvt.u64.u32 %rd80, %r80;
mul.lo.s32 %r81, %r2431, %r59;
cvt.u64.u32 %rd81, %r81;
add.s64 %rd4, %rd81, %rd80;
mul.hi.u32 %r82, %r11, %r61;
add.s32 %r83, %r82, %r11;
shr.u32 %r13, %r83, %r62;
shr.s32 %r84, %r3, 31;
shr.u32 %r85, %r84, 29;
add.s32 %r86, %r3, %r85;
shr.s32 %r2433, %r86, 3;
shr.u32 %r87, %r49, 8;
setp.ge.s32 %p12, %r2433, %r87;
mov.f32 %f28, %f27;
mov.f32 %f29, %f27;
mov.f32 %f30, %f27;
mov.f32 %f31, %f27;
mov.f32 %f32, %f27;
mov.f32 %f33, %f27;
mov.f32 %f34, %f27;
@%p12 bra $L__BB306_19;

mul.lo.s32 %r88, %r13, %r64;
mul.lo.s32 %r89, %r2430, %r58;
and.b32 %r93, %r86, -8;
sub.s32 %r15, %r3, %r93;
shl.b32 %r94, %r15, 1;
shl.b32 %r95, %r15, 2;
cvt.s64.s32 %rd82, %r15;
shr.u32 %r96, %r15, 31;
add.s32 %r97, %r15, %r96;
shr.s32 %r98, %r97, 1;
cvt.s64.s32 %rd83, %r98;
and.b32 %r16, %r95, 4;
shl.b32 %r2432, %r2433, 3;
add.s32 %r99, %r2433, %r88;
add.s32 %r18, %r99, %r89;
add.s32 %r100, %r4, 1;
mad.lo.s32 %r101, %r53, %r100, %r18;
mul.wide.s32 %rd84, %r101, 110;
add.s64 %rd85, %rd84, %rd83;
add.s64 %rd5, %rd85, 106;
add.s64 %rd86, %rd84, 74;
not.b32 %r102, %r94;
mul.wide.s32 %rd87, %r102, 2;
sub.s64 %rd6, %rd86, %rd87;
mul.wide.s32 %rd88, %r94, 2;
add.s64 %rd89, %rd84, %rd88;
add.s64 %rd7, %rd89, 74;
add.s64 %rd90, %rd84, %rd82;
add.s64 %rd8, %rd90, 66;
mad.lo.s32 %r103, %r4, %r53, %r2433;
add.s32 %r104, %r103, %r88;
add.s32 %r19, %r104, %r89;
mul.wide.s32 %rd91, %r19, 110;
mul.wide.s32 %rd92, %r95, 2;
add.s64 %rd93, %rd91, %rd92;
add.s64 %rd9, %rd93, 4;
add.s64 %rd94, %rd84, %rd92;
add.s64 %rd10, %rd94, 2;
add.s64 %rd95, %rd91, %rd82;
add.s64 %rd11, %rd95, 66;
add.s64 %rd96, %rd91, %rd88;
add.s64 %rd12, %rd96, 74;
add.s32 %r105, %r4, 2;
mad.lo.s32 %r106, %r53, %r105, %r18;
mul.wide.s32 %rd97, %r106, 110;
add.s64 %rd98, %rd97, %rd83;
add.s64 %rd13, %rd98, 106;
add.s64 %rd99, %rd97, 74;
sub.s64 %rd14, %rd99, %rd87;
add.s64 %rd100, %rd97, %rd88;
add.s64 %rd15, %rd100, 74;
add.s64 %rd101, %rd97, %rd82;
add.s64 %rd16, %rd101, 66;
add.s64 %rd102, %rd91, 74;
sub.s64 %rd17, %rd102, %rd87;
add.s64 %rd103, %rd91, %rd83;
add.s64 %rd18, %rd103, 106;
add.s64 %rd104, %rd97, %rd92;
add.s64 %rd19, %rd104, 2;
add.s32 %r107, %r4, 3;
mad.lo.s32 %r108, %r53, %r107, %r18;
mul.wide.s32 %rd105, %r108, 110;
add.s64 %rd106, %rd105, %rd83;
add.s64 %rd20, %rd106, 106;
add.s64 %rd107, %rd105, 74;
sub.s64 %rd21, %rd107, %rd87;
add.s64 %rd108, %rd105, %rd88;
add.s64 %rd22, %rd108, 74;
cvta.to.global.u64 %rd310, %rd52;
cvta.to.global.u64 %rd309, %rd48;
mov.f32 %f27, 0f00000000;
cvta.to.global.u64 %rd109, %rd49;
mov.f32 %f28, %f27;
mov.f32 %f29, %f27;
mov.f32 %f30, %f27;
mov.f32 %f31, %f27;
mov.f32 %f32, %f27;
mov.f32 %f33, %f27;
mov.f32 %f34, %f27;

$L__BB306_10:
cvt.u64.u32 %rd110, %r2432;
add.s64 %rd111, %rd4, %rd110;
add.s64 %rd112, %rd309, %rd9;
ld.global.nc.u16 %rs2, [%rd112];
ld.global.nc.u16 %rs3, [%rd112+2];
ld.global.nc.u16 %rs4, [%rd112+4];
add.s64 %rd28, %rd309, %rd11;
ld.global.nc.u8 %rs5, [%rd28];
add.s64 %rd29, %rd309, %rd12;
add.s64 %rd30, %rd309, %rd17;
ld.global.nc.u16 %rs6, [%rd29];
ld.global.nc.u16 %rs7, [%rd30];
add.s64 %rd114, %rd111, %rd82;
mul.lo.s64 %rd115, %rd114, 36;
ld.global.nc.u16 %rs8, [%rd112+-2];
and.b16 %rs9, %rs8, 255;
cvt.u32.u16 %r158, %rs9;
shl.b16 %rs10, %rs5, 8;
cvt.u32.u16 %r159, %rs10;
and.b32 %r160, %r159, 256;
or.b32 %r161, %r160, %r158;
mul.wide.u32 %rd116, %r161, 4;
mov.u64 %rd117, iq3s_grid;
add.s64 %rd118, %rd117, %rd116;
ld.global.nc.u32 %r162, [%rd118];
shr.u16 %rs11, %rs8, 8;
cvt.u32.u16 %r163, %rs11;
mul.wide.u16 %r164, %rs5, 128;
and.b32 %r165, %r164, 256;
or.b32 %r166, %r165, %r163;
mul.wide.u32 %rd119, %r166, 4;
add.s64 %rd120, %rd117, %rd119;
ld.global.nc.u32 %r167, [%rd120];
and.b16 %rs13, %rs6, 255;
mul.wide.u16 %r168, %rs13, 128;
cvt.u32.u16 %r169, %rs6;
shl.b32 %r170, %r169, 21;
or.b32 %r171, %r168, %r170;
and.b32 %r172, %r171, 25166208;
or.b32 %r173, %r172, -2139062144;
add.s32 %r174, %r173, -16843009;
or.b32 %r109, %r174, %r172;

prmt.b32 %r109, %r109, 0, 0xba98;

mul.wide.u16 %r175, %rs13, 8;
shl.b32 %r176, %r169, 17;
or.b32 %r177, %r175, %r176;
and.b32 %r178, %r177, 25166208;
or.b32 %r179, %r178, -2139062144;
add.s32 %r180, %r179, -16843009;
or.b32 %r111, %r180, %r178;

prmt.b32 %r111, %r111, 0, 0xba98;

xor.b32 %r181, %r109, %r162;
or.b32 %r182, %r181, -2139062144;
and.b32 %r183, %r109, 2139062143;
and.b32 %r184, %r162, -2139062144;
sub.s32 %r185, %r182, %r183;
xor.b32 %r186, %r184, %r185;
xor.b32 %r114, %r186, -2139062144;
xor.b32 %r187, %r111, %r167;
or.b32 %r188, %r187, -2139062144;
and.b32 %r189, %r111, 2139062143;
and.b32 %r190, %r167, -2139062144;
sub.s32 %r191, %r188, %r189;
xor.b32 %r192, %r190, %r191;
xor.b32 %r118, %r192, -2139062144;
add.s64 %rd121, %rd109, %rd115;
ld.global.nc.u32 %r22, [%rd121+4];
ld.global.nc.u32 %r23, [%rd121+8];
mov.u32 %r116, 0;

dp4a.s32.s32 %r113, %r114, %r22, %r116;


dp4a.s32.s32 %r117, %r118, %r23, %r113;

and.b16 %rs14, %rs2, 255;
cvt.u32.u16 %r193, %rs14;
mul.wide.u16 %r194, %rs5, 64;
and.b32 %r195, %r194, 256;
or.b32 %r196, %r195, %r193;
mul.wide.u32 %rd122, %r196, 4;
add.s64 %rd123, %rd117, %rd122;
ld.global.nc.u32 %r197, [%rd123];
shr.u16 %rs15, %rs2, 8;
cvt.u32.u16 %r198, %rs15;
mul.wide.u16 %r199, %rs5, 32;
and.b32 %r200, %r199, 256;
or.b32 %r201, %r200, %r198;
mul.wide.u32 %rd124, %r201, 4;
add.s64 %rd125, %rd117, %rd124;
ld.global.nc.u32 %r202, [%rd125];
shr.u16 %rs16, %rs6, 8;
mul.wide.u16 %r203, %rs16, 128;
cvt.u32.u16 %r204, %rs16;
shl.b32 %r205, %r204, 21;
or.b32 %r206, %r203, %r205;
and.b32 %r207, %r206, 25166208;
or.b32 %r208, %r207, -2139062144;
add.s32 %r209, %r208, -16843009;
or.b32 %r121, %r209, %r207;

prmt.b32 %r121, %r121, 0, 0xba98;

mul.wide.u16 %r210, %rs16, 8;
shl.b32 %r211, %r204, 17;
or.b32 %r212, %r210, %r211;
and.b32 %r213, %r212, 25166208;
or.b32 %r214, %r213, -2139062144;
add.s32 %r215, %r214, -16843009;
or.b32 %r123, %r215, %r213;

prmt.b32 %r123, %r123, 0, 0xba98;

xor.b32 %r216, %r121, %r197;
or.b32 %r217, %r216, -2139062144;
and.b32 %r218, %r121, 2139062143;
and.b32 %r219, %r197, -2139062144;
sub.s32 %r220, %r217, %r218;
xor.b32 %r221, %r219, %r220;
xor.b32 %r126, %r221, -2139062144;
xor.b32 %r222, %r123, %r202;
or.b32 %r223, %r222, -2139062144;
and.b32 %r224, %r123, 2139062143;
and.b32 %r225, %r202, -2139062144;
sub.s32 %r226, %r223, %r224;
xor.b32 %r227, %r225, %r226;
xor.b32 %r130, %r227, -2139062144;
ld.global.nc.u32 %r24, [%rd121+12];
ld.global.nc.u32 %r25, [%rd121+16];

dp4a.s32.s32 %r125, %r126, %r24, %r117;


dp4a.s32.s32 %r129, %r130, %r25, %r125;

and.b16 %rs17, %rs3, 255;
cvt.u32.u16 %r228, %rs17;
mul.wide.u16 %r229, %rs5, 16;
and.b32 %r230, %r229, 256;
or.b32 %r231, %r230, %r228;
mul.wide.u32 %rd126, %r231, 4;
add.s64 %rd127, %rd117, %rd126;
ld.global.nc.u32 %r232, [%rd127];
shr.u16 %rs18, %rs3, 8;
cvt.u32.u16 %r233, %rs18;
mul.wide.u16 %r234, %rs5, 8;
and.b32 %r235, %r234, 256;
or.b32 %r236, %r235, %r233;
mul.wide.u32 %rd128, %r236, 4;
add.s64 %rd129, %rd117, %rd128;
ld.global.nc.u32 %r237, [%rd129];
and.b16 %rs19, %rs7, 255;
mul.wide.u16 %r238, %rs19, 128;
cvt.u32.u16 %r239, %rs7;
shl.b32 %r240, %r239, 21;
or.b32 %r241, %r238, %r240;
and.b32 %r242, %r241, 25166208;
or.b32 %r243, %r242, -2139062144;
add.s32 %r244, %r243, -16843009;
or.b32 %r133, %r244, %r242;

prmt.b32 %r133, %r133, 0, 0xba98;

mul.wide.u16 %r245, %rs19, 8;
shl.b32 %r246, %r239, 17;
or.b32 %r247, %r245, %r246;
and.b32 %r248, %r247, 25166208;
or.b32 %r249, %r248, -2139062144;
add.s32 %r250, %r249, -16843009;
or.b32 %r135, %r250, %r248;

prmt.b32 %r135, %r135, 0, 0xba98;

xor.b32 %r251, %r133, %r232;
or.b32 %r252, %r251, -2139062144;
and.b32 %r253, %r133, 2139062143;
and.b32 %r254, %r232, -2139062144;
sub.s32 %r255, %r252, %r253;
xor.b32 %r256, %r254, %r255;
xor.b32 %r138, %r256, -2139062144;
xor.b32 %r257, %r135, %r237;
or.b32 %r258, %r257, -2139062144;
and.b32 %r259, %r135, 2139062143;
and.b32 %r260, %r237, -2139062144;
sub.s32 %r261, %r258, %r259;
xor.b32 %r262, %r260, %r261;
xor.b32 %r142, %r262, -2139062144;
ld.global.nc.u32 %r26, [%rd121+20];
ld.global.nc.u32 %r27, [%rd121+24];

dp4a.s32.s32 %r137, %r138, %r26, %r129;


dp4a.s32.s32 %r141, %r142, %r27, %r137;

and.b16 %rs20, %rs4, 255;
cvt.u32.u16 %r263, %rs20;
mul.wide.u16 %r264, %rs5, 4;
and.b32 %r265, %r264, 256;
or.b32 %r266, %r265, %r263;
mul.wide.u32 %rd130, %r266, 4;
add.s64 %rd131, %rd117, %rd130;
ld.global.nc.u32 %r267, [%rd131];
shr.u16 %rs21, %rs4, 8;
cvt.u32.u16 %r268, %rs21;
mul.wide.u16 %r269, %rs5, 2;
and.b32 %r270, %r269, 256;
or.b32 %r271, %r270, %r268;
mul.wide.u32 %rd132, %r271, 4;
add.s64 %rd133, %rd117, %rd132;
ld.global.nc.u32 %r272, [%rd133];
shr.u16 %rs22, %rs7, 8;
mul.wide.u16 %r273, %rs22, 128;
cvt.u32.u16 %r274, %rs22;
shl.b32 %r275, %r274, 21;
or.b32 %r276, %r273, %r275;
and.b32 %r277, %r276, 25166208;
or.b32 %r278, %r277, -2139062144;
add.s32 %r279, %r278, -16843009;
or.b32 %r145, %r279, %r277;

prmt.b32 %r145, %r145, 0, 0xba98;

mul.wide.u16 %r280, %rs22, 8;
shl.b32 %r281, %r274, 17;
or.b32 %r282, %r280, %r281;
and.b32 %r283, %r282, 25166208;
or.b32 %r284, %r283, -2139062144;
add.s32 %r285, %r284, -16843009;
or.b32 %r147, %r285, %r283;

prmt.b32 %r147, %r147, 0, 0xba98;

xor.b32 %r286, %r145, %r267;
or.b32 %r287, %r286, -2139062144;
and.b32 %r288, %r145, 2139062143;
and.b32 %r289, %r267, -2139062144;
sub.s32 %r290, %r287, %r288;
xor.b32 %r291, %r289, %r290;
xor.b32 %r150, %r291, -2139062144;
xor.b32 %r292, %r147, %r272;
or.b32 %r293, %r292, -2139062144;
and.b32 %r294, %r147, 2139062143;
and.b32 %r295, %r272, -2139062144;
sub.s32 %r296, %r293, %r294;
xor.b32 %r297, %r295, %r296;
xor.b32 %r154, %r297, -2139062144;
ld.global.nc.u32 %r28, [%rd121+28];
ld.global.nc.u32 %r29, [%rd121+32];

dp4a.s32.s32 %r149, %r150, %r28, %r141;


dp4a.s32.s32 %r153, %r154, %r29, %r149;

add.s64 %rd31, %rd309, %rd18;
ld.global.nc.u8 %rs23, [%rd31];
cvt.u32.u16 %r298, %rs23;
and.b32 %r299, %r298, 255;
shr.u32 %r300, %r299, %r16;
shl.b32 %r301, %r300, 1;
and.b32 %r302, %r301, 30;
or.b32 %r303, %r302, 1;
mul.lo.s32 %r304, %r303, %r153;
add.s64 %rd32, %rd309, %rd91;
ld.global.nc.u16 %rs1, [%rd32];

{ cvt.f32.f16 %f101, %rs1;}


ld.global.nc.u32 %r30, [%rd121];

{.reg .f16 low,high;
mov.b32 {low,high},%r30;
cvt.f32.f16 %f102, low;}


mul.ftz.f32 %f103, %f101, %f102;
cvt.rn.f32.s32 %f104, %r304;
fma.rn.ftz.f32 %f30, %f103, %f104, %f30;
setp.eq.s64 %p13, %rd52, 0;
@%p13 bra $L__BB306_12;

mov.u64 %rd302, iq3s_grid;
add.s64 %rd135, %rd310, %rd9;
ld.global.u16 %rs25, [%rd135];
ld.global.u16 %rs26, [%rd135+2];
ld.global.u16 %rs27, [%rd135+4];
add.s64 %rd136, %rd310, %rd11;
add.s64 %rd137, %rd310, %rd12;
add.s64 %rd138, %rd310, %rd17;
ld.global.u16 %rs28, [%rd137];
ld.global.u16 %rs29, [%rd138];
ld.global.u16 %rs30, [%rd135+-2];
and.b16 %rs31, %rs30, 255;
cvt.u32.u16 %r354, %rs31;
ld.global.u8 %rs32, [%rd136];
shl.b16 %rs33, %rs32, 8;
cvt.u32.u16 %r355, %rs33;
and.b32 %r356, %r355, 256;
or.b32 %r357, %r356, %r354;
mul.wide.u32 %rd139, %r357, 4;
add.s64 %rd141, %rd302, %rd139;
ld.global.nc.u32 %r358, [%rd141];
shr.u16 %rs34, %rs30, 8;
cvt.u32.u16 %r359, %rs34;
mul.wide.u16 %r360, %rs32, 128;
and.b32 %r361, %r360, 256;
or.b32 %r362, %r361, %r359;
mul.wide.u32 %rd142, %r362, 4;
add.s64 %rd143, %rd302, %rd142;
ld.global.nc.u32 %r363, [%rd143];
and.b16 %rs35, %rs28, 255;
mul.wide.u16 %r364, %rs35, 128;
cvt.u32.u16 %r365, %rs28;
shl.b32 %r366, %r365, 21;
or.b32 %r367, %r364, %r366;
and.b32 %r368, %r367, 25166208;
or.b32 %r369, %r368, -2139062144;
add.s32 %r370, %r369, -16843009;
or.b32 %r305, %r370, %r368;

prmt.b32 %r305, %r305, 0, 0xba98;

mul.wide.u16 %r371, %rs35, 8;
shl.b32 %r372, %r365, 17;
or.b32 %r373, %r371, %r372;
and.b32 %r374, %r373, 25166208;
or.b32 %r375, %r374, -2139062144;
add.s32 %r376, %r375, -16843009;
or.b32 %r307, %r376, %r374;

prmt.b32 %r307, %r307, 0, 0xba98;

xor.b32 %r377, %r305, %r358;
or.b32 %r378, %r377, -2139062144;
and.b32 %r379, %r305, 2139062143;
and.b32 %r380, %r358, -2139062144;
sub.s32 %r381, %r378, %r379;
xor.b32 %r382, %r380, %r381;
xor.b32 %r310, %r382, -2139062144;
xor.b32 %r383, %r307, %r363;
or.b32 %r384, %r383, -2139062144;
and.b32 %r385, %r307, 2139062143;
and.b32 %r386, %r363, -2139062144;
sub.s32 %r387, %r384, %r385;
xor.b32 %r388, %r386, %r387;
xor.b32 %r314, %r388, -2139062144;

dp4a.s32.s32 %r309, %r310, %r22, %r116;


dp4a.s32.s32 %r313, %r314, %r23, %r309;

and.b16 %rs36, %rs25, 255;
cvt.u32.u16 %r389, %rs36;
mul.wide.u16 %r390, %rs32, 64;
and.b32 %r391, %r390, 256;
or.b32 %r392, %r391, %r389;
mul.wide.u32 %rd144, %r392, 4;
add.s64 %rd145, %rd302, %rd144;
ld.global.nc.u32 %r393, [%rd145];
shr.u16 %rs37, %rs25, 8;
cvt.u32.u16 %r394, %rs37;
mul.wide.u16 %r395, %rs32, 32;
and.b32 %r396, %r395, 256;
or.b32 %r397, %r396, %r394;
mul.wide.u32 %rd146, %r397, 4;
add.s64 %rd147, %rd302, %rd146;
ld.global.nc.u32 %r398, [%rd147];
shr.u16 %rs38, %rs28, 8;
mul.wide.u16 %r399, %rs38, 128;
cvt.u32.u16 %r400, %rs38;
shl.b32 %r401, %r400, 21;
or.b32 %r402, %r399, %r401;
and.b32 %r403, %r402, 25166208;
or.b32 %r404, %r403, -2139062144;
add.s32 %r405, %r404, -16843009;
or.b32 %r317, %r405, %r403;

prmt.b32 %r317, %r317, 0, 0xba98;

mul.wide.u16 %r406, %rs38, 8;
shl.b32 %r407, %r400, 17;
or.b32 %r408, %r406, %r407;
and.b32 %r409, %r408, 25166208;
or.b32 %r410, %r409, -2139062144;
add.s32 %r411, %r410, -16843009;
or.b32 %r319, %r411, %r409;

prmt.b32 %r319, %r319, 0, 0xba98;

xor.b32 %r412, %r317, %r393;
or.b32 %r413, %r412, -2139062144;
and.b32 %r414, %r317, 2139062143;
and.b32 %r415, %r393, -2139062144;
sub.s32 %r416, %r413, %r414;
xor.b32 %r417, %r415, %r416;
xor.b32 %r322, %r417, -2139062144;
xor.b32 %r418, %r319, %r398;
or.b32 %r419, %r418, -2139062144;
and.b32 %r420, %r319, 2139062143;
and.b32 %r421, %r398, -2139062144;
sub.s32 %r422, %r419, %r420;
xor.b32 %r423, %r421, %r422;
xor.b32 %r326, %r423, -2139062144;

dp4a.s32.s32 %r321, %r322, %r24, %r313;


dp4a.s32.s32 %r325, %r326, %r25, %r321;

and.b16 %rs39, %rs26, 255;
cvt.u32.u16 %r424, %rs39;
mul.wide.u16 %r425, %rs32, 16;
and.b32 %r426, %r425, 256;
or.b32 %r427, %r426, %r424;
mul.wide.u32 %rd148, %r427, 4;
add.s64 %rd149, %rd302, %rd148;
ld.global.nc.u32 %r428, [%rd149];
shr.u16 %rs40, %rs26, 8;
cvt.u32.u16 %r429, %rs40;
mul.wide.u16 %r430, %rs32, 8;
and.b32 %r431, %r430, 256;
or.b32 %r432, %r431, %r429;
mul.wide.u32 %rd150, %r432, 4;
add.s64 %rd151, %rd302, %rd150;
ld.global.nc.u32 %r433, [%rd151];
and.b16 %rs41, %rs29, 255;
mul.wide.u16 %r434, %rs41, 128;
cvt.u32.u16 %r435, %rs29;
shl.b32 %r436, %r435, 21;
or.b32 %r437, %r434, %r436;
and.b32 %r438, %r437, 25166208;
or.b32 %r439, %r438, -2139062144;
add.s32 %r440, %r439, -16843009;
or.b32 %r329, %r440, %r438;

prmt.b32 %r329, %r329, 0, 0xba98;

mul.wide.u16 %r441, %rs41, 8;
shl.b32 %r442, %r435, 17;
or.b32 %r443, %r441, %r442;
and.b32 %r444, %r443, 25166208;
or.b32 %r445, %r444, -2139062144;
add.s32 %r446, %r445, -16843009;
or.b32 %r331, %r446, %r444;

prmt.b32 %r331, %r331, 0, 0xba98;

xor.b32 %r447, %r329, %r428;
or.b32 %r448, %r447, -2139062144;
and.b32 %r449, %r329, 2139062143;
and.b32 %r450, %r428, -2139062144;
sub.s32 %r451, %r448, %r449;
xor.b32 %r452, %r450, %r451;
xor.b32 %r334, %r452, -2139062144;
xor.b32 %r453, %r331, %r433;
or.b32 %r454, %r453, -2139062144;
and.b32 %r455, %r331, 2139062143;
and.b32 %r456, %r433, -2139062144;
sub.s32 %r457, %r454, %r455;
xor.b32 %r458, %r456, %r457;
xor.b32 %r338, %r458, -2139062144;

dp4a.s32.s32 %r333, %r334, %r26, %r325;


dp4a.s32.s32 %r337, %r338, %r27, %r333;

and.b16 %rs42, %rs27, 255;
cvt.u32.u16 %r459, %rs42;
mul.wide.u16 %r460, %rs32, 4;
and.b32 %r461, %r460, 256;
or.b32 %r462, %r461, %r459;
mul.wide.u32 %rd152, %r462, 4;
add.s64 %rd153, %rd302, %rd152;
ld.global.nc.u32 %r463, [%rd153];
shr.u16 %rs43, %rs27, 8;
cvt.u32.u16 %r464, %rs43;
mul.wide.u16 %r465, %rs32, 2;
and.b32 %r466, %r465, 256;
or.b32 %r467, %r466, %r464;
mul.wide.u32 %rd154, %r467, 4;
add.s64 %rd155, %rd302, %rd154;
ld.global.nc.u32 %r468, [%rd155];
shr.u16 %rs44, %rs29, 8;
mul.wide.u16 %r469, %rs44, 128;
cvt.u32.u16 %r470, %rs44;
shl.b32 %r471, %r470, 21;
or.b32 %r472, %r469, %r471;
and.b32 %r473, %r472, 25166208;
or.b32 %r474, %r473, -2139062144;
add.s32 %r475, %r474, -16843009;
or.b32 %r341, %r475, %r473;

prmt.b32 %r341, %r341, 0, 0xba98;

mul.wide.u16 %r476, %rs44, 8;
shl.b32 %r477, %r470, 17;
or.b32 %r478, %r476, %r477;
and.b32 %r479, %r478, 25166208;
or.b32 %r480, %r479, -2139062144;
add.s32 %r481, %r480, -16843009;
or.b32 %r343, %r481, %r479;

prmt.b32 %r343, %r343, 0, 0xba98;

xor.b32 %r482, %r341, %r463;
or.b32 %r483, %r482, -2139062144;
and.b32 %r484, %r341, 2139062143;
and.b32 %r485, %r463, -2139062144;
sub.s32 %r486, %r483, %r484;
xor.b32 %r487, %r485, %r486;
xor.b32 %r346, %r487, -2139062144;
xor.b32 %r488, %r343, %r468;
or.b32 %r489, %r488, -2139062144;
and.b32 %r490, %r343, 2139062143;
and.b32 %r491, %r468, -2139062144;
sub.s32 %r492, %r489, %r490;
xor.b32 %r493, %r491, %r492;
xor.b32 %r350, %r493, -2139062144;

dp4a.s32.s32 %r345, %r346, %r28, %r337;


dp4a.s32.s32 %r349, %r350, %r29, %r345;

add.s64 %rd156, %rd310, %rd18;
ld.global.u8 %r494, [%rd156];
shr.u32 %r495, %r494, %r16;
shl.b32 %r496, %r495, 1;
and.b32 %r497, %r496, 30;
or.b32 %r498, %r497, 1;
mul.lo.s32 %r499, %r498, %r349;
add.s64 %rd158, %rd310, %rd91;
ld.global.u16 %rs24, [%rd158];

{ cvt.f32.f16 %f105, %rs24;}



{.reg .f16 low,high;
mov.b32 {low,high},%r30;
cvt.f32.f16 %f106, low;}


mul.ftz.f32 %f107, %f105, %f106;
cvt.rn.f32.s32 %f108, %r499;
fma.rn.ftz.f32 %f34, %f107, %f108, %f34;

$L__BB306_12:
add.s64 %rd306, %rd309, %rd9;
add.s64 %rd305, %rd306, -2;
mov.u64 %rd299, iq3s_grid;
mul.wide.s32 %rd159, %r53, 110;
add.s64 %rd33, %rd305, %rd159;
ld.global.nc.u16 %rs46, [%rd33+2];
ld.global.nc.u16 %rs47, [%rd33+4];
ld.global.nc.u16 %rs48, [%rd33+6];
add.s64 %rd34, %rd28, %rd159;
ld.global.nc.u8 %rs49, [%rd34];
add.s64 %rd35, %rd29, %rd159;
add.s64 %rd36, %rd30, %rd159;
ld.global.nc.u16 %rs50, [%rd35];
ld.global.nc.u16 %rs51, [%rd36];
ld.global.nc.u16 %rs52, [%rd33];
and.b16 %rs53, %rs52, 255;
cvt.u32.u16 %r549, %rs53;
shl.b16 %rs54, %rs49, 8;
cvt.u32.u16 %r550, %rs54;
and.b32 %r551, %r550, 256;
or.b32 %r552, %r551, %r549;
mul.wide.u32 %rd160, %r552, 4;
add.s64 %rd162, %rd299, %rd160;
ld.global.nc.u32 %r553, [%rd162];
shr.u16 %rs55, %rs52, 8;
cvt.u32.u16 %r554, %rs55;
mul.wide.u16 %r555, %rs49, 128;
and.b32 %r556, %r555, 256;
or.b32 %r557, %r556, %r554;
mul.wide.u32 %rd163, %r557, 4;
add.s64 %rd164, %rd299, %rd163;
ld.global.nc.u32 %r558, [%rd164];
and.b16 %rs57, %rs50, 255;
mul.wide.u16 %r559, %rs57, 128;
cvt.u32.u16 %r560, %rs50;
shl.b32 %r561, %r560, 21;
or.b32 %r562, %r559, %r561;
and.b32 %r563, %r562, 25166208;
or.b32 %r564, %r563, -2139062144;
add.s32 %r565, %r564, -16843009;
or.b32 %r500, %r565, %r563;

prmt.b32 %r500, %r500, 0, 0xba98;

mul.wide.u16 %r566, %rs57, 8;
shl.b32 %r567, %r560, 17;
or.b32 %r568, %r566, %r567;
and.b32 %r569, %r568, 25166208;
or.b32 %r570, %r569, -2139062144;
add.s32 %r571, %r570, -16843009;
or.b32 %r502, %r571, %r569;

prmt.b32 %r502, %r502, 0, 0xba98;

xor.b32 %r572, %r500, %r553;
or.b32 %r573, %r572, -2139062144;
and.b32 %r574, %r500, 2139062143;
and.b32 %r575, %r553, -2139062144;
sub.s32 %r576, %r573, %r574;
xor.b32 %r577, %r575, %r576;
xor.b32 %r505, %r577, -2139062144;
xor.b32 %r578, %r502, %r558;
or.b32 %r579, %r578, -2139062144;
and.b32 %r580, %r502, 2139062143;
and.b32 %r581, %r558, -2139062144;
sub.s32 %r582, %r579, %r580;
xor.b32 %r583, %r581, %r582;
xor.b32 %r509, %r583, -2139062144;

dp4a.s32.s32 %r504, %r505, %r22, %r116;


dp4a.s32.s32 %r508, %r509, %r23, %r504;

and.b16 %rs58, %rs46, 255;
cvt.u32.u16 %r584, %rs58;
mul.wide.u16 %r585, %rs49, 64;
and.b32 %r586, %r585, 256;
or.b32 %r587, %r586, %r584;
mul.wide.u32 %rd165, %r587, 4;
add.s64 %rd166, %rd299, %rd165;
ld.global.nc.u32 %r588, [%rd166];
shr.u16 %rs59, %rs46, 8;
cvt.u32.u16 %r589, %rs59;
mul.wide.u16 %r590, %rs49, 32;
and.b32 %r591, %r590, 256;
or.b32 %r592, %r591, %r589;
mul.wide.u32 %rd167, %r592, 4;
add.s64 %rd168, %rd299, %rd167;
ld.global.nc.u32 %r593, [%rd168];
shr.u16 %rs60, %rs50, 8;
mul.wide.u16 %r594, %rs60, 128;
cvt.u32.u16 %r595, %rs60;
shl.b32 %r596, %r595, 21;
or.b32 %r597, %r594, %r596;
and.b32 %r598, %r597, 25166208;
or.b32 %r599, %r598, -2139062144;
add.s32 %r600, %r599, -16843009;
or.b32 %r512, %r600, %r598;

prmt.b32 %r512, %r512, 0, 0xba98;

mul.wide.u16 %r601, %rs60, 8;
shl.b32 %r602, %r595, 17;
or.b32 %r603, %r601, %r602;
and.b32 %r604, %r603, 25166208;
or.b32 %r605, %r604, -2139062144;
add.s32 %r606, %r605, -16843009;
or.b32 %r514, %r606, %r604;

prmt.b32 %r514, %r514, 0, 0xba98;

xor.b32 %r607, %r512, %r588;
or.b32 %r608, %r607, -2139062144;
and.b32 %r609, %r512, 2139062143;
and.b32 %r610, %r588, -2139062144;
sub.s32 %r611, %r608, %r609;
xor.b32 %r612, %r610, %r611;
xor.b32 %r517, %r612, -2139062144;
xor.b32 %r613, %r514, %r593;
or.b32 %r614, %r613, -2139062144;
and.b32 %r615, %r514, 2139062143;
and.b32 %r616, %r593, -2139062144;
sub.s32 %r617, %r614, %r615;
xor.b32 %r618, %r616, %r617;
xor.b32 %r521, %r618, -2139062144;

dp4a.s32.s32 %r516, %r517, %r24, %r508;


dp4a.s32.s32 %r520, %r521, %r25, %r516;

and.b16 %rs61, %rs47, 255;
cvt.u32.u16 %r619, %rs61;
mul.wide.u16 %r620, %rs49, 16;
and.b32 %r621, %r620, 256;
or.b32 %r622, %r621, %r619;
mul.wide.u32 %rd169, %r622, 4;
add.s64 %rd170, %rd299, %rd169;
ld.global.nc.u32 %r623, [%rd170];
shr.u16 %rs62, %rs47, 8;
cvt.u32.u16 %r624, %rs62;
mul.wide.u16 %r625, %rs49, 8;
and.b32 %r626, %r625, 256;
or.b32 %r627, %r626, %r624;
mul.wide.u32 %rd171, %r627, 4;
add.s64 %rd172, %rd299, %rd171;
ld.global.nc.u32 %r628, [%rd172];
and.b16 %rs63, %rs51, 255;
mul.wide.u16 %r629, %rs63, 128;
cvt.u32.u16 %r630, %rs51;
shl.b32 %r631, %r630, 21;
or.b32 %r632, %r629, %r631;
and.b32 %r633, %r632, 25166208;
or.b32 %r634, %r633, -2139062144;
add.s32 %r635, %r634, -16843009;
or.b32 %r524, %r635, %r633;

prmt.b32 %r524, %r524, 0, 0xba98;

mul.wide.u16 %r636, %rs63, 8;
shl.b32 %r637, %r630, 17;
or.b32 %r638, %r636, %r637;
and.b32 %r639, %r638, 25166208;
or.b32 %r640, %r639, -2139062144;
add.s32 %r641, %r640, -16843009;
or.b32 %r526, %r641, %r639;

prmt.b32 %r526, %r526, 0, 0xba98;

xor.b32 %r642, %r524, %r623;
or.b32 %r643, %r642, -2139062144;
and.b32 %r644, %r524, 2139062143;
and.b32 %r645, %r623, -2139062144;
sub.s32 %r646, %r643, %r644;
xor.b32 %r647, %r645, %r646;
xor.b32 %r529, %r647, -2139062144;
xor.b32 %r648, %r526, %r628;
or.b32 %r649, %r648, -2139062144;
and.b32 %r650, %r526, 2139062143;
and.b32 %r651, %r628, -2139062144;
sub.s32 %r652, %r649, %r650;
xor.b32 %r653, %r651, %r652;
xor.b32 %r533, %r653, -2139062144;

dp4a.s32.s32 %r528, %r529, %r26, %r520;


dp4a.s32.s32 %r532, %r533, %r27, %r528;

and.b16 %rs64, %rs48, 255;
cvt.u32.u16 %r654, %rs64;
mul.wide.u16 %r655, %rs49, 4;
and.b32 %r656, %r655, 256;
or.b32 %r657, %r656, %r654;
mul.wide.u32 %rd173, %r657, 4;
add.s64 %rd174, %rd299, %rd173;
ld.global.nc.u32 %r658, [%rd174];
shr.u16 %rs65, %rs48, 8;
cvt.u32.u16 %r659, %rs65;
mul.wide.u16 %r660, %rs49, 2;
and.b32 %r661, %r660, 256;
or.b32 %r662, %r661, %r659;
mul.wide.u32 %rd175, %r662, 4;
add.s64 %rd176, %rd299, %rd175;
ld.global.nc.u32 %r663, [%rd176];
shr.u16 %rs66, %rs51, 8;
mul.wide.u16 %r664, %rs66, 128;
cvt.u32.u16 %r665, %rs66;
shl.b32 %r666, %r665, 21;
or.b32 %r667, %r664, %r666;
and.b32 %r668, %r667, 25166208;
or.b32 %r669, %r668, -2139062144;
add.s32 %r670, %r669, -16843009;
or.b32 %r536, %r670, %r668;

prmt.b32 %r536, %r536, 0, 0xba98;

mul.wide.u16 %r671, %rs66, 8;
shl.b32 %r672, %r665, 17;
or.b32 %r673, %r671, %r672;
and.b32 %r674, %r673, 25166208;
or.b32 %r675, %r674, -2139062144;
add.s32 %r676, %r675, -16843009;
or.b32 %r538, %r676, %r674;

prmt.b32 %r538, %r538, 0, 0xba98;

xor.b32 %r677, %r536, %r658;
or.b32 %r678, %r677, -2139062144;
and.b32 %r679, %r536, 2139062143;
and.b32 %r680, %r658, -2139062144;
sub.s32 %r681, %r678, %r679;
xor.b32 %r682, %r680, %r681;
xor.b32 %r541, %r682, -2139062144;
xor.b32 %r683, %r538, %r663;
or.b32 %r684, %r683, -2139062144;
and.b32 %r685, %r538, 2139062143;
and.b32 %r686, %r663, -2139062144;
sub.s32 %r687, %r684, %r685;
xor.b32 %r688, %r686, %r687;
xor.b32 %r545, %r688, -2139062144;

dp4a.s32.s32 %r540, %r541, %r28, %r532;


dp4a.s32.s32 %r544, %r545, %r29, %r540;

add.s64 %rd37, %rd31, %rd159;
ld.global.nc.u8 %rs67, [%rd37];
cvt.u32.u16 %r689, %rs67;
and.b32 %r690, %r689, 255;
shr.u32 %r691, %r690, %r16;
shl.b32 %r692, %r691, 1;
and.b32 %r693, %r692, 30;
or.b32 %r694, %r693, 1;
mul.lo.s32 %r695, %r694, %r544;
add.s64 %rd38, %rd32, %rd159;
ld.global.nc.u16 %rs45, [%rd38];

{ cvt.f32.f16 %f109, %rs45;}



{.reg .f16 low,high;
mov.b32 {low,high},%r30;
cvt.f32.f16 %f110, low;}


mul.ftz.f32 %f111, %f109, %f110;
cvt.rn.f32.s32 %f112, %r695;
fma.rn.ftz.f32 %f29, %f111, %f112, %f29;
@%p13 bra $L__BB306_14;

mov.u64 %rd301, iq3s_grid;
add.s64 %rd177, %rd310, %rd10;
ld.global.u16 %rs69, [%rd177+2];
ld.global.u16 %rs70, [%rd177+4];
ld.global.u16 %rs71, [%rd177+6];
add.s64 %rd178, %rd310, %rd8;
add.s64 %rd179, %rd310, %rd7;
add.s64 %rd180, %rd310, %rd6;
ld.global.u16 %rs72, [%rd179];
ld.global.u16 %rs73, [%rd180];
ld.global.u16 %rs74, [%rd177];
and.b16 %rs75, %rs74, 255;
cvt.u32.u16 %r745, %rs75;
ld.global.u8 %rs76, [%rd178];
shl.b16 %rs77, %rs76, 8;
cvt.u32.u16 %r746, %rs77;
and.b32 %r747, %r746, 256;
or.b32 %r748, %r747, %r745;
mul.wide.u32 %rd181, %r748, 4;
add.s64 %rd183, %rd301, %rd181;
ld.global.nc.u32 %r749, [%rd183];
shr.u16 %rs78, %rs74, 8;
cvt.u32.u16 %r750, %rs78;
mul.wide.u16 %r751, %rs76, 128;
and.b32 %r752, %r751, 256;
or.b32 %r753, %r752, %r750;
mul.wide.u32 %rd184, %r753, 4;
add.s64 %rd185, %rd301, %rd184;
ld.global.nc.u32 %r754, [%rd185];
and.b16 %rs79, %rs72, 255;
mul.wide.u16 %r755, %rs79, 128;
cvt.u32.u16 %r756, %rs72;
shl.b32 %r757, %r756, 21;
or.b32 %r758, %r755, %r757;
and.b32 %r759, %r758, 25166208;
or.b32 %r760, %r759, -2139062144;
add.s32 %r761, %r760, -16843009;
or.b32 %r696, %r761, %r759;

prmt.b32 %r696, %r696, 0, 0xba98;

mul.wide.u16 %r762, %rs79, 8;
shl.b32 %r763, %r756, 17;
or.b32 %r764, %r762, %r763;
and.b32 %r765, %r764, 25166208;
or.b32 %r766, %r765, -2139062144;
add.s32 %r767, %r766, -16843009;
or.b32 %r698, %r767, %r765;

prmt.b32 %r698, %r698, 0, 0xba98;

xor.b32 %r768, %r696, %r749;
or.b32 %r769, %r768, -2139062144;
and.b32 %r770, %r696, 2139062143;
and.b32 %r771, %r749, -2139062144;
sub.s32 %r772, %r769, %r770;
xor.b32 %r773, %r771, %r772;
xor.b32 %r701, %r773, -2139062144;
xor.b32 %r774, %r698, %r754;
or.b32 %r775, %r774, -2139062144;
and.b32 %r776, %r698, 2139062143;
and.b32 %r777, %r754, -2139062144;
sub.s32 %r778, %r775, %r776;
xor.b32 %r779, %r777, %r778;
xor.b32 %r705, %r779, -2139062144;
mov.u32 %r703, 0;

dp4a.s32.s32 %r700, %r701, %r22, %r703;


dp4a.s32.s32 %r704, %r705, %r23, %r700;

and.b16 %rs80, %rs69, 255;
cvt.u32.u16 %r780, %rs80;
mul.wide.u16 %r781, %rs76, 64;
and.b32 %r782, %r781, 256;
or.b32 %r783, %r782, %r780;
mul.wide.u32 %rd186, %r783, 4;
add.s64 %rd187, %rd301, %rd186;
ld.global.nc.u32 %r784, [%rd187];
shr.u16 %rs81, %rs69, 8;
cvt.u32.u16 %r785, %rs81;
mul.wide.u16 %r786, %rs76, 32;
and.b32 %r787, %r786, 256;
or.b32 %r788, %r787, %r785;
mul.wide.u32 %rd188, %r788, 4;
add.s64 %rd189, %rd301, %rd188;
ld.global.nc.u32 %r789, [%rd189];
shr.u16 %rs82, %rs72, 8;
mul.wide.u16 %r790, %rs82, 128;
cvt.u32.u16 %r791, %rs82;
shl.b32 %r792, %r791, 21;
or.b32 %r793, %r790, %r792;
and.b32 %r794, %r793, 25166208;
or.b32 %r795, %r794, -2139062144;
add.s32 %r796, %r795, -16843009;
or.b32 %r708, %r796, %r794;

prmt.b32 %r708, %r708, 0, 0xba98;

mul.wide.u16 %r797, %rs82, 8;
shl.b32 %r798, %r791, 17;
or.b32 %r799, %r797, %r798;
and.b32 %r800, %r799, 25166208;
or.b32 %r801, %r800, -2139062144;
add.s32 %r802, %r801, -16843009;
or.b32 %r710, %r802, %r800;

prmt.b32 %r710, %r710, 0, 0xba98;

xor.b32 %r803, %r708, %r784;
or.b32 %r804, %r803, -2139062144;
and.b32 %r805, %r708, 2139062143;
and.b32 %r806, %r784, -2139062144;
sub.s32 %r807, %r804, %r805;
xor.b32 %r808, %r806, %r807;
xor.b32 %r713, %r808, -2139062144;
xor.b32 %r809, %r710, %r789;
or.b32 %r810, %r809, -2139062144;
and.b32 %r811, %r710, 2139062143;
and.b32 %r812, %r789, -2139062144;
sub.s32 %r813, %r810, %r811;
xor.b32 %r814, %r812, %r813;
xor.b32 %r717, %r814, -2139062144;

dp4a.s32.s32 %r712, %r713, %r24, %r704;


dp4a.s32.s32 %r716, %r717, %r25, %r712;

and.b16 %rs83, %rs70, 255;
cvt.u32.u16 %r815, %rs83;
mul.wide.u16 %r816, %rs76, 16;
and.b32 %r817, %r816, 256;
or.b32 %r818, %r817, %r815;
mul.wide.u32 %rd190, %r818, 4;
add.s64 %rd191, %rd301, %rd190;
ld.global.nc.u32 %r819, [%rd191];
shr.u16 %rs84, %rs70, 8;
cvt.u32.u16 %r820, %rs84;
mul.wide.u16 %r821, %rs76, 8;
and.b32 %r822, %r821, 256;
or.b32 %r823, %r822, %r820;
mul.wide.u32 %rd192, %r823, 4;
add.s64 %rd193, %rd301, %rd192;
ld.global.nc.u32 %r824, [%rd193];
and.b16 %rs85, %rs73, 255;
mul.wide.u16 %r825, %rs85, 128;
cvt.u32.u16 %r826, %rs73;
shl.b32 %r827, %r826, 21;
or.b32 %r828, %r825, %r827;
and.b32 %r829, %r828, 25166208;
or.b32 %r830, %r829, -2139062144;
add.s32 %r831, %r830, -16843009;
or.b32 %r720, %r831, %r829;

prmt.b32 %r720, %r720, 0, 0xba98;

mul.wide.u16 %r832, %rs85, 8;
shl.b32 %r833, %r826, 17;
or.b32 %r834, %r832, %r833;
and.b32 %r835, %r834, 25166208;
or.b32 %r836, %r835, -2139062144;
add.s32 %r837, %r836, -16843009;
or.b32 %r722, %r837, %r835;

prmt.b32 %r722, %r722, 0, 0xba98;

xor.b32 %r838, %r720, %r819;
or.b32 %r839, %r838, -2139062144;
and.b32 %r840, %r720, 2139062143;
and.b32 %r841, %r819, -2139062144;
sub.s32 %r842, %r839, %r840;
xor.b32 %r843, %r841, %r842;
xor.b32 %r725, %r843, -2139062144;
xor.b32 %r844, %r722, %r824;
or.b32 %r845, %r844, -2139062144;
and.b32 %r846, %r722, 2139062143;
and.b32 %r847, %r824, -2139062144;
sub.s32 %r848, %r845, %r846;
xor.b32 %r849, %r847, %r848;
xor.b32 %r729, %r849, -2139062144;

dp4a.s32.s32 %r724, %r725, %r26, %r716;


dp4a.s32.s32 %r728, %r729, %r27, %r724;

and.b16 %rs86, %rs71, 255;
cvt.u32.u16 %r850, %rs86;
mul.wide.u16 %r851, %rs76, 4;
and.b32 %r852, %r851, 256;
or.b32 %r853, %r852, %r850;
mul.wide.u32 %rd194, %r853, 4;
add.s64 %rd195, %rd301, %rd194;
ld.global.nc.u32 %r854, [%rd195];
shr.u16 %rs87, %rs71, 8;
cvt.u32.u16 %r855, %rs87;
mul.wide.u16 %r856, %rs76, 2;
and.b32 %r857, %r856, 256;
or.b32 %r858, %r857, %r855;
mul.wide.u32 %rd196, %r858, 4;
add.s64 %rd197, %rd301, %rd196;
ld.global.nc.u32 %r859, [%rd197];
shr.u16 %rs88, %rs73, 8;
mul.wide.u16 %r860, %rs88, 128;
cvt.u32.u16 %r861, %rs88;
shl.b32 %r862, %r861, 21;
or.b32 %r863, %r860, %r862;
and.b32 %r864, %r863, 25166208;
or.b32 %r865, %r864, -2139062144;
add.s32 %r866, %r865, -16843009;
or.b32 %r732, %r866, %r864;

prmt.b32 %r732, %r732, 0, 0xba98;

mul.wide.u16 %r867, %rs88, 8;
shl.b32 %r868, %r861, 17;
or.b32 %r869, %r867, %r868;
and.b32 %r870, %r869, 25166208;
or.b32 %r871, %r870, -2139062144;
add.s32 %r872, %r871, -16843009;
or.b32 %r734, %r872, %r870;

prmt.b32 %r734, %r734, 0, 0xba98;

xor.b32 %r873, %r732, %r854;
or.b32 %r874, %r873, -2139062144;
and.b32 %r875, %r732, 2139062143;
and.b32 %r876, %r854, -2139062144;
sub.s32 %r877, %r874, %r875;
xor.b32 %r878, %r876, %r877;
xor.b32 %r737, %r878, -2139062144;
xor.b32 %r879, %r734, %r859;
or.b32 %r880, %r879, -2139062144;
and.b32 %r881, %r734, 2139062143;
and.b32 %r882, %r859, -2139062144;
sub.s32 %r883, %r880, %r881;
xor.b32 %r884, %r882, %r883;
xor.b32 %r741, %r884, -2139062144;

dp4a.s32.s32 %r736, %r737, %r28, %r728;


dp4a.s32.s32 %r740, %r741, %r29, %r736;

add.s64 %rd198, %rd310, %rd5;
ld.global.u8 %r885, [%rd198];
shr.u32 %r886, %r885, %r16;
shl.b32 %r887, %r886, 1;
and.b32 %r888, %r887, 30;
or.b32 %r889, %r888, 1;
mul.lo.s32 %r890, %r889, %r740;
or.b32 %r893, %r4, 1;
mad.lo.s32 %r894, %r53, %r893, %r18;
mul.wide.s32 %rd199, %r894, 110;
add.s64 %rd200, %rd310, %rd199;
ld.global.u16 %rs68, [%rd200];

{ cvt.f32.f16 %f113, %rs68;}



{.reg .f16 low,high;
mov.b32 {low,high},%r30;
cvt.f32.f16 %f114, low;}


mul.ftz.f32 %f115, %f113, %f114;
cvt.rn.f32.s32 %f116, %r890;
fma.rn.ftz.f32 %f33, %f115, %f116, %f33;

$L__BB306_14:
mul.wide.s32 %rd307, %r53, 110;
mov.u64 %rd300, iq3s_grid;
add.s64 %rd39, %rd33, %rd307;
ld.global.nc.u16 %rs90, [%rd39+2];
ld.global.nc.u16 %rs91, [%rd39+4];
ld.global.nc.u16 %rs92, [%rd39+6];
add.s64 %rd40, %rd34, %rd307;
ld.global.nc.u8 %rs93, [%rd40];
add.s64 %rd41, %rd35, %rd307;
add.s64 %rd42, %rd36, %rd307;
ld.global.nc.u16 %rs94, [%rd41];
ld.global.nc.u16 %rs95, [%rd42];
ld.global.nc.u16 %rs96, [%rd39];
and.b16 %rs97, %rs96, 255;
cvt.u32.u16 %r944, %rs97;
shl.b16 %rs98, %rs93, 8;
cvt.u32.u16 %r945, %rs98;
and.b32 %r946, %r945, 256;
or.b32 %r947, %r946, %r944;
mul.wide.u32 %rd202, %r947, 4;
add.s64 %rd204, %rd300, %rd202;
ld.global.nc.u32 %r948, [%rd204];
shr.u16 %rs99, %rs96, 8;
cvt.u32.u16 %r949, %rs99;
mul.wide.u16 %r950, %rs93, 128;
and.b32 %r951, %r950, 256;
or.b32 %r952, %r951, %r949;
mul.wide.u32 %rd205, %r952, 4;
add.s64 %rd206, %rd300, %rd205;
ld.global.nc.u32 %r953, [%rd206];
and.b16 %rs101, %rs94, 255;
mul.wide.u16 %r954, %rs101, 128;
cvt.u32.u16 %r955, %rs94;
shl.b32 %r956, %r955, 21;
or.b32 %r957, %r954, %r956;
and.b32 %r958, %r957, 25166208;
or.b32 %r959, %r958, -2139062144;
add.s32 %r960, %r959, -16843009;
or.b32 %r895, %r960, %r958;

prmt.b32 %r895, %r895, 0, 0xba98;

mul.wide.u16 %r961, %rs101, 8;
shl.b32 %r962, %r955, 17;
or.b32 %r963, %r961, %r962;
and.b32 %r964, %r963, 25166208;
or.b32 %r965, %r964, -2139062144;
add.s32 %r966, %r965, -16843009;
or.b32 %r897, %r966, %r964;

prmt.b32 %r897, %r897, 0, 0xba98;

xor.b32 %r967, %r895, %r948;
or.b32 %r968, %r967, -2139062144;
and.b32 %r969, %r895, 2139062143;
and.b32 %r970, %r948, -2139062144;
sub.s32 %r971, %r968, %r969;
xor.b32 %r972, %r970, %r971;
xor.b32 %r900, %r972, -2139062144;
xor.b32 %r973, %r897, %r953;
or.b32 %r974, %r973, -2139062144;
and.b32 %r975, %r897, 2139062143;
and.b32 %r976, %r953, -2139062144;
sub.s32 %r977, %r974, %r975;
xor.b32 %r978, %r976, %r977;
xor.b32 %r904, %r978, -2139062144;
mov.u32 %r902, 0;

dp4a.s32.s32 %r899, %r900, %r22, %r902;


dp4a.s32.s32 %r903, %r904, %r23, %r899;

and.b16 %rs102, %rs90, 255;
cvt.u32.u16 %r979, %rs102;
mul.wide.u16 %r980, %rs93, 64;
and.b32 %r981, %r980, 256;
or.b32 %r982, %r981, %r979;
mul.wide.u32 %rd207, %r982, 4;
add.s64 %rd208, %rd300, %rd207;
ld.global.nc.u32 %r983, [%rd208];
shr.u16 %rs103, %rs90, 8;
cvt.u32.u16 %r984, %rs103;
mul.wide.u16 %r985, %rs93, 32;
and.b32 %r986, %r985, 256;
or.b32 %r987, %r986, %r984;
mul.wide.u32 %rd209, %r987, 4;
add.s64 %rd210, %rd300, %rd209;
ld.global.nc.u32 %r988, [%rd210];
shr.u16 %rs104, %rs94, 8;
mul.wide.u16 %r989, %rs104, 128;
cvt.u32.u16 %r990, %rs104;
shl.b32 %r991, %r990, 21;
or.b32 %r992, %r989, %r991;
and.b32 %r993, %r992, 25166208;
or.b32 %r994, %r993, -2139062144;
add.s32 %r995, %r994, -16843009;
or.b32 %r907, %r995, %r993;

prmt.b32 %r907, %r907, 0, 0xba98;

mul.wide.u16 %r996, %rs104, 8;
shl.b32 %r997, %r990, 17;
or.b32 %r998, %r996, %r997;
and.b32 %r999, %r998, 25166208;
or.b32 %r1000, %r999, -2139062144;
add.s32 %r1001, %r1000, -16843009;
or.b32 %r909, %r1001, %r999;

prmt.b32 %r909, %r909, 0, 0xba98;

xor.b32 %r1002, %r907, %r983;
or.b32 %r1003, %r1002, -2139062144;
and.b32 %r1004, %r907, 2139062143;
and.b32 %r1005, %r983, -2139062144;
sub.s32 %r1006, %r1003, %r1004;
xor.b32 %r1007, %r1005, %r1006;
xor.b32 %r912, %r1007, -2139062144;
xor.b32 %r1008, %r909, %r988;
or.b32 %r1009, %r1008, -2139062144;
and.b32 %r1010, %r909, 2139062143;
and.b32 %r1011, %r988, -2139062144;
sub.s32 %r1012, %r1009, %r1010;
xor.b32 %r1013, %r1011, %r1012;
xor.b32 %r916, %r1013, -2139062144;

dp4a.s32.s32 %r911, %r912, %r24, %r903;


dp4a.s32.s32 %r915, %r916, %r25, %r911;

and.b16 %rs105, %rs91, 255;
cvt.u32.u16 %r1014, %rs105;
mul.wide.u16 %r1015, %rs93, 16;
and.b32 %r1016, %r1015, 256;
or.b32 %r1017, %r1016, %r1014;
mul.wide.u32 %rd211, %r1017, 4;
add.s64 %rd212, %rd300, %rd211;
ld.global.nc.u32 %r1018, [%rd212];
shr.u16 %rs106, %rs91, 8;
cvt.u32.u16 %r1019, %rs106;
mul.wide.u16 %r1020, %rs93, 8;
and.b32 %r1021, %r1020, 256;
or.b32 %r1022, %r1021, %r1019;
mul.wide.u32 %rd213, %r1022, 4;
add.s64 %rd214, %rd300, %rd213;
ld.global.nc.u32 %r1023, [%rd214];
and.b16 %rs107, %rs95, 255;
mul.wide.u16 %r1024, %rs107, 128;
cvt.u32.u16 %r1025, %rs95;
shl.b32 %r1026, %r1025, 21;
or.b32 %r1027, %r1024, %r1026;
and.b32 %r1028, %r1027, 25166208;
or.b32 %r1029, %r1028, -2139062144;
add.s32 %r1030, %r1029, -16843009;
or.b32 %r919, %r1030, %r1028;

prmt.b32 %r919, %r919, 0, 0xba98;

mul.wide.u16 %r1031, %rs107, 8;
shl.b32 %r1032, %r1025, 17;
or.b32 %r1033, %r1031, %r1032;
and.b32 %r1034, %r1033, 25166208;
or.b32 %r1035, %r1034, -2139062144;
add.s32 %r1036, %r1035, -16843009;
or.b32 %r921, %r1036, %r1034;

prmt.b32 %r921, %r921, 0, 0xba98;

xor.b32 %r1037, %r919, %r1018;
or.b32 %r1038, %r1037, -2139062144;
and.b32 %r1039, %r919, 2139062143;
and.b32 %r1040, %r1018, -2139062144;
sub.s32 %r1041, %r1038, %r1039;
xor.b32 %r1042, %r1040, %r1041;
xor.b32 %r924, %r1042, -2139062144;
xor.b32 %r1043, %r921, %r1023;
or.b32 %r1044, %r1043, -2139062144;
and.b32 %r1045, %r921, 2139062143;
and.b32 %r1046, %r1023, -2139062144;
sub.s32 %r1047, %r1044, %r1045;
xor.b32 %r1048, %r1046, %r1047;
xor.b32 %r928, %r1048, -2139062144;

dp4a.s32.s32 %r923, %r924, %r26, %r915;


dp4a.s32.s32 %r927, %r928, %r27, %r923;

and.b16 %rs108, %rs92, 255;
cvt.u32.u16 %r1049, %rs108;
mul.wide.u16 %r1050, %rs93, 4;
and.b32 %r1051, %r1050, 256;
or.b32 %r1052, %r1051, %r1049;
mul.wide.u32 %rd215, %r1052, 4;
add.s64 %rd216, %rd300, %rd215;
ld.global.nc.u32 %r1053, [%rd216];
shr.u16 %rs109, %rs92, 8;
cvt.u32.u16 %r1054, %rs109;
mul.wide.u16 %r1055, %rs93, 2;
and.b32 %r1056, %r1055, 256;
or.b32 %r1057, %r1056, %r1054;
mul.wide.u32 %rd217, %r1057, 4;
add.s64 %rd218, %rd300, %rd217;
ld.global.nc.u32 %r1058, [%rd218];
shr.u16 %rs110, %rs95, 8;
mul.wide.u16 %r1059, %rs110, 128;
cvt.u32.u16 %r1060, %rs110;
shl.b32 %r1061, %r1060, 21;
or.b32 %r1062, %r1059, %r1061;
and.b32 %r1063, %r1062, 25166208;
or.b32 %r1064, %r1063, -2139062144;
add.s32 %r1065, %r1064, -16843009;
or.b32 %r931, %r1065, %r1063;

prmt.b32 %r931, %r931, 0, 0xba98;

mul.wide.u16 %r1066, %rs110, 8;
shl.b32 %r1067, %r1060, 17;
or.b32 %r1068, %r1066, %r1067;
and.b32 %r1069, %r1068, 25166208;
or.b32 %r1070, %r1069, -2139062144;
add.s32 %r1071, %r1070, -16843009;
or.b32 %r933, %r1071, %r1069;

prmt.b32 %r933, %r933, 0, 0xba98;

xor.b32 %r1072, %r931, %r1053;
or.b32 %r1073, %r1072, -2139062144;
and.b32 %r1074, %r931, 2139062143;
and.b32 %r1075, %r1053, -2139062144;
sub.s32 %r1076, %r1073, %r1074;
xor.b32 %r1077, %r1075, %r1076;
xor.b32 %r936, %r1077, -2139062144;
xor.b32 %r1078, %r933, %r1058;
or.b32 %r1079, %r1078, -2139062144;
and.b32 %r1080, %r933, 2139062143;
and.b32 %r1081, %r1058, -2139062144;
sub.s32 %r1082, %r1079, %r1080;
xor.b32 %r1083, %r1081, %r1082;
xor.b32 %r940, %r1083, -2139062144;

dp4a.s32.s32 %r935, %r936, %r28, %r927;


dp4a.s32.s32 %r939, %r940, %r29, %r935;

add.s64 %rd43, %rd37, %rd307;
ld.global.nc.u8 %rs111, [%rd43];
cvt.u32.u16 %r1084, %rs111;
and.b32 %r1085, %r1084, 255;
shr.u32 %r1086, %r1085, %r16;
shl.b32 %r1087, %r1086, 1;
and.b32 %r1088, %r1087, 30;
or.b32 %r1089, %r1088, 1;
mul.lo.s32 %r1090, %r1089, %r939;
add.s64 %rd44, %rd38, %rd307;
ld.global.nc.u16 %rs89, [%rd44];

{ cvt.f32.f16 %f117, %rs89;}



{.reg .f16 low,high;
mov.b32 {low,high},%r30;
cvt.f32.f16 %f118, low;}


mul.ftz.f32 %f119, %f117, %f118;
cvt.rn.f32.s32 %f120, %r1090;
fma.rn.ftz.f32 %f28, %f119, %f120, %f28;
@%p13 bra $L__BB306_16;

mov.u64 %rd304, iq3s_grid;
add.s64 %rd219, %rd310, %rd19;
ld.global.u16 %rs113, [%rd219+2];
ld.global.u16 %rs114, [%rd219+4];
ld.global.u16 %rs115, [%rd219+6];
add.s64 %rd220, %rd310, %rd16;
add.s64 %rd221, %rd310, %rd15;
add.s64 %rd222, %rd310, %rd14;
ld.global.u16 %rs116, [%rd221];
ld.global.u16 %rs117, [%rd222];
ld.global.u16 %rs118, [%rd219];
and.b16 %rs119, %rs118, 255;
cvt.u32.u16 %r1140, %rs119;
ld.global.u8 %rs120, [%rd220];
shl.b16 %rs121, %rs120, 8;
cvt.u32.u16 %r1141, %rs121;
and.b32 %r1142, %r1141, 256;
or.b32 %r1143, %r1142, %r1140;
mul.wide.u32 %rd223, %r1143, 4;
add.s64 %rd225, %rd304, %rd223;
ld.global.nc.u32 %r1144, [%rd225];
shr.u16 %rs122, %rs118, 8;
cvt.u32.u16 %r1145, %rs122;
mul.wide.u16 %r1146, %rs120, 128;
and.b32 %r1147, %r1146, 256;
or.b32 %r1148, %r1147, %r1145;
mul.wide.u32 %rd226, %r1148, 4;
add.s64 %rd227, %rd304, %rd226;
ld.global.nc.u32 %r1149, [%rd227];
and.b16 %rs123, %rs116, 255;
mul.wide.u16 %r1150, %rs123, 128;
cvt.u32.u16 %r1151, %rs116;
shl.b32 %r1152, %r1151, 21;
or.b32 %r1153, %r1150, %r1152;
and.b32 %r1154, %r1153, 25166208;
or.b32 %r1155, %r1154, -2139062144;
add.s32 %r1156, %r1155, -16843009;
or.b32 %r1091, %r1156, %r1154;

prmt.b32 %r1091, %r1091, 0, 0xba98;

mul.wide.u16 %r1157, %rs123, 8;
shl.b32 %r1158, %r1151, 17;
or.b32 %r1159, %r1157, %r1158;
and.b32 %r1160, %r1159, 25166208;
or.b32 %r1161, %r1160, -2139062144;
add.s32 %r1162, %r1161, -16843009;
or.b32 %r1093, %r1162, %r1160;

prmt.b32 %r1093, %r1093, 0, 0xba98;

xor.b32 %r1163, %r1091, %r1144;
or.b32 %r1164, %r1163, -2139062144;
and.b32 %r1165, %r1091, 2139062143;
and.b32 %r1166, %r1144, -2139062144;
sub.s32 %r1167, %r1164, %r1165;
xor.b32 %r1168, %r1166, %r1167;
xor.b32 %r1096, %r1168, -2139062144;
xor.b32 %r1169, %r1093, %r1149;
or.b32 %r1170, %r1169, -2139062144;
and.b32 %r1171, %r1093, 2139062143;
and.b32 %r1172, %r1149, -2139062144;
sub.s32 %r1173, %r1170, %r1171;
xor.b32 %r1174, %r1172, %r1173;
xor.b32 %r1100, %r1174, -2139062144;

dp4a.s32.s32 %r1095, %r1096, %r22, %r902;


dp4a.s32.s32 %r1099, %r1100, %r23, %r1095;

and.b16 %rs124, %rs113, 255;
cvt.u32.u16 %r1175, %rs124;
mul.wide.u16 %r1176, %rs120, 64;
and.b32 %r1177, %r1176, 256;
or.b32 %r1178, %r1177, %r1175;
mul.wide.u32 %rd228, %r1178, 4;
add.s64 %rd229, %rd304, %rd228;
ld.global.nc.u32 %r1179, [%rd229];
shr.u16 %rs125, %rs113, 8;
cvt.u32.u16 %r1180, %rs125;
mul.wide.u16 %r1181, %rs120, 32;
and.b32 %r1182, %r1181, 256;
or.b32 %r1183, %r1182, %r1180;
mul.wide.u32 %rd230, %r1183, 4;
add.s64 %rd231, %rd304, %rd230;
ld.global.nc.u32 %r1184, [%rd231];
shr.u16 %rs126, %rs116, 8;
mul.wide.u16 %r1185, %rs126, 128;
cvt.u32.u16 %r1186, %rs126;
shl.b32 %r1187, %r1186, 21;
or.b32 %r1188, %r1185, %r1187;
and.b32 %r1189, %r1188, 25166208;
or.b32 %r1190, %r1189, -2139062144;
add.s32 %r1191, %r1190, -16843009;
or.b32 %r1103, %r1191, %r1189;

prmt.b32 %r1103, %r1103, 0, 0xba98;

mul.wide.u16 %r1192, %rs126, 8;
shl.b32 %r1193, %r1186, 17;
or.b32 %r1194, %r1192, %r1193;
and.b32 %r1195, %r1194, 25166208;
or.b32 %r1196, %r1195, -2139062144;
add.s32 %r1197, %r1196, -16843009;
or.b32 %r1105, %r1197, %r1195;

prmt.b32 %r1105, %r1105, 0, 0xba98;

xor.b32 %r1198, %r1103, %r1179;
or.b32 %r1199, %r1198, -2139062144;
and.b32 %r1200, %r1103, 2139062143;
and.b32 %r1201, %r1179, -2139062144;
sub.s32 %r1202, %r1199, %r1200;
xor.b32 %r1203, %r1201, %r1202;
xor.b32 %r1108, %r1203, -2139062144;
xor.b32 %r1204, %r1105, %r1184;
or.b32 %r1205, %r1204, -2139062144;
and.b32 %r1206, %r1105, 2139062143;
and.b32 %r1207, %r1184, -2139062144;
sub.s32 %r1208, %r1205, %r1206;
xor.b32 %r1209, %r1207, %r1208;
xor.b32 %r1112, %r1209, -2139062144;

dp4a.s32.s32 %r1107, %r1108, %r24, %r1099;


dp4a.s32.s32 %r1111, %r1112, %r25, %r1107;

and.b16 %rs127, %rs114, 255;
cvt.u32.u16 %r1210, %rs127;
mul.wide.u16 %r1211, %rs120, 16;
and.b32 %r1212, %r1211, 256;
or.b32 %r1213, %r1212, %r1210;
mul.wide.u32 %rd232, %r1213, 4;
add.s64 %rd233, %rd304, %rd232;
ld.global.nc.u32 %r1214, [%rd233];
shr.u16 %rs128, %rs114, 8;
cvt.u32.u16 %r1215, %rs128;
mul.wide.u16 %r1216, %rs120, 8;
and.b32 %r1217, %r1216, 256;
or.b32 %r1218, %r1217, %r1215;
mul.wide.u32 %rd234, %r1218, 4;
add.s64 %rd235, %rd304, %rd234;
ld.global.nc.u32 %r1219, [%rd235];
and.b16 %rs129, %rs117, 255;
mul.wide.u16 %r1220, %rs129, 128;
cvt.u32.u16 %r1221, %rs117;
shl.b32 %r1222, %r1221, 21;
or.b32 %r1223, %r1220, %r1222;
and.b32 %r1224, %r1223, 25166208;
or.b32 %r1225, %r1224, -2139062144;
add.s32 %r1226, %r1225, -16843009;
or.b32 %r1115, %r1226, %r1224;

prmt.b32 %r1115, %r1115, 0, 0xba98;

mul.wide.u16 %r1227, %rs129, 8;
shl.b32 %r1228, %r1221, 17;
or.b32 %r1229, %r1227, %r1228;
and.b32 %r1230, %r1229, 25166208;
or.b32 %r1231, %r1230, -2139062144;
add.s32 %r1232, %r1231, -16843009;
or.b32 %r1117, %r1232, %r1230;

prmt.b32 %r1117, %r1117, 0, 0xba98;

xor.b32 %r1233, %r1115, %r1214;
or.b32 %r1234, %r1233, -2139062144;
and.b32 %r1235, %r1115, 2139062143;
and.b32 %r1236, %r1214, -2139062144;
sub.s32 %r1237, %r1234, %r1235;
xor.b32 %r1238, %r1236, %r1237;
xor.b32 %r1120, %r1238, -2139062144;
xor.b32 %r1239, %r1117, %r1219;
or.b32 %r1240, %r1239, -2139062144;
and.b32 %r1241, %r1117, 2139062143;
and.b32 %r1242, %r1219, -2139062144;
sub.s32 %r1243, %r1240, %r1241;
xor.b32 %r1244, %r1242, %r1243;
xor.b32 %r1124, %r1244, -2139062144;

dp4a.s32.s32 %r1119, %r1120, %r26, %r1111;


dp4a.s32.s32 %r1123, %r1124, %r27, %r1119;

and.b16 %rs130, %rs115, 255;
cvt.u32.u16 %r1245, %rs130;
mul.wide.u16 %r1246, %rs120, 4;
and.b32 %r1247, %r1246, 256;
or.b32 %r1248, %r1247, %r1245;
mul.wide.u32 %rd236, %r1248, 4;
add.s64 %rd237, %rd304, %rd236;
ld.global.nc.u32 %r1249, [%rd237];
shr.u16 %rs131, %rs115, 8;
cvt.u32.u16 %r1250, %rs131;
mul.wide.u16 %r1251, %rs120, 2;
and.b32 %r1252, %r1251, 256;
or.b32 %r1253, %r1252, %r1250;
mul.wide.u32 %rd238, %r1253, 4;
add.s64 %rd239, %rd304, %rd238;
ld.global.nc.u32 %r1254, [%rd239];
shr.u16 %rs132, %rs117, 8;
mul.wide.u16 %r1255, %rs132, 128;
cvt.u32.u16 %r1256, %rs132;
shl.b32 %r1257, %r1256, 21;
or.b32 %r1258, %r1255, %r1257;
and.b32 %r1259, %r1258, 25166208;
or.b32 %r1260, %r1259, -2139062144;
add.s32 %r1261, %r1260, -16843009;
or.b32 %r1127, %r1261, %r1259;

prmt.b32 %r1127, %r1127, 0, 0xba98;

mul.wide.u16 %r1262, %rs132, 8;
shl.b32 %r1263, %r1256, 17;
or.b32 %r1264, %r1262, %r1263;
and.b32 %r1265, %r1264, 25166208;
or.b32 %r1266, %r1265, -2139062144;
add.s32 %r1267, %r1266, -16843009;
or.b32 %r1129, %r1267, %r1265;

prmt.b32 %r1129, %r1129, 0, 0xba98;

xor.b32 %r1268, %r1127, %r1249;
or.b32 %r1269, %r1268, -2139062144;
and.b32 %r1270, %r1127, 2139062143;
and.b32 %r1271, %r1249, -2139062144;
sub.s32 %r1272, %r1269, %r1270;
xor.b32 %r1273, %r1271, %r1272;
xor.b32 %r1132, %r1273, -2139062144;
xor.b32 %r1274, %r1129, %r1254;
or.b32 %r1275, %r1274, -2139062144;
and.b32 %r1276, %r1129, 2139062143;
and.b32 %r1277, %r1254, -2139062144;
sub.s32 %r1278, %r1275, %r1276;
xor.b32 %r1279, %r1277, %r1278;
xor.b32 %r1136, %r1279, -2139062144;

dp4a.s32.s32 %r1131, %r1132, %r28, %r1123;


dp4a.s32.s32 %r1135, %r1136, %r29, %r1131;

add.s64 %rd240, %rd310, %rd13;
ld.global.u8 %r1280, [%rd240];
shr.u32 %r1281, %r1280, %r16;
shl.b32 %r1282, %r1281, 1;
and.b32 %r1283, %r1282, 30;
or.b32 %r1284, %r1283, 1;
mul.lo.s32 %r1285, %r1284, %r1135;
or.b32 %r1288, %r4, 2;
mad.lo.s32 %r1289, %r53, %r1288, %r18;
mul.wide.s32 %rd241, %r1289, 110;
add.s64 %rd242, %rd310, %rd241;
ld.global.u16 %rs112, [%rd242];

{ cvt.f32.f16 %f121, %rs112;}



{.reg .f16 low,high;
mov.b32 {low,high},%r30;
cvt.f32.f16 %f122, low;}


mul.ftz.f32 %f123, %f121, %f122;
cvt.rn.f32.s32 %f124, %r1285;
fma.rn.ftz.f32 %f32, %f123, %f124, %f32;

$L__BB306_16:
mul.wide.s32 %rd308, %r53, 110;
mov.u64 %rd297, iq3s_grid;
add.s64 %rd244, %rd39, %rd308;
ld.global.nc.u16 %rs134, [%rd244+2];
ld.global.nc.u16 %rs135, [%rd244+4];
ld.global.nc.u16 %rs136, [%rd244+6];
add.s64 %rd245, %rd40, %rd308;
ld.global.nc.u8 %rs137, [%rd245];
add.s64 %rd246, %rd41, %rd308;
add.s64 %rd247, %rd42, %rd308;
ld.global.nc.u16 %rs138, [%rd246];
ld.global.nc.u16 %rs139, [%rd247];
ld.global.nc.u16 %rs140, [%rd244];
and.b16 %rs141, %rs140, 255;
cvt.u32.u16 %r1339, %rs141;
shl.b16 %rs142, %rs137, 8;
cvt.u32.u16 %r1340, %rs142;
and.b32 %r1341, %r1340, 256;
or.b32 %r1342, %r1341, %r1339;
mul.wide.u32 %rd248, %r1342, 4;
add.s64 %rd250, %rd297, %rd248;
ld.global.nc.u32 %r1343, [%rd250];
shr.u16 %rs143, %rs140, 8;
cvt.u32.u16 %r1344, %rs143;
mul.wide.u16 %r1345, %rs137, 128;
and.b32 %r1346, %r1345, 256;
or.b32 %r1347, %r1346, %r1344;
mul.wide.u32 %rd251, %r1347, 4;
add.s64 %rd252, %rd297, %rd251;
ld.global.nc.u32 %r1348, [%rd252];
and.b16 %rs145, %rs138, 255;
mul.wide.u16 %r1349, %rs145, 128;
cvt.u32.u16 %r1350, %rs138;
shl.b32 %r1351, %r1350, 21;
or.b32 %r1352, %r1349, %r1351;
and.b32 %r1353, %r1352, 25166208;
or.b32 %r1354, %r1353, -2139062144;
add.s32 %r1355, %r1354, -16843009;
or.b32 %r1290, %r1355, %r1353;

prmt.b32 %r1290, %r1290, 0, 0xba98;

mul.wide.u16 %r1356, %rs145, 8;
shl.b32 %r1357, %r1350, 17;
or.b32 %r1358, %r1356, %r1357;
and.b32 %r1359, %r1358, 25166208;
or.b32 %r1360, %r1359, -2139062144;
add.s32 %r1361, %r1360, -16843009;
or.b32 %r1292, %r1361, %r1359;

prmt.b32 %r1292, %r1292, 0, 0xba98;

xor.b32 %r1362, %r1290, %r1343;
or.b32 %r1363, %r1362, -2139062144;
and.b32 %r1364, %r1290, 2139062143;
and.b32 %r1365, %r1343, -2139062144;
sub.s32 %r1366, %r1363, %r1364;
xor.b32 %r1367, %r1365, %r1366;
xor.b32 %r1295, %r1367, -2139062144;
xor.b32 %r1368, %r1292, %r1348;
or.b32 %r1369, %r1368, -2139062144;
and.b32 %r1370, %r1292, 2139062143;
and.b32 %r1371, %r1348, -2139062144;
sub.s32 %r1372, %r1369, %r1370;
xor.b32 %r1373, %r1371, %r1372;
xor.b32 %r1299, %r1373, -2139062144;

dp4a.s32.s32 %r1294, %r1295, %r22, %r902;


dp4a.s32.s32 %r1298, %r1299, %r23, %r1294;

and.b16 %rs146, %rs134, 255;
cvt.u32.u16 %r1374, %rs146;
mul.wide.u16 %r1375, %rs137, 64;
and.b32 %r1376, %r1375, 256;
or.b32 %r1377, %r1376, %r1374;
mul.wide.u32 %rd253, %r1377, 4;
add.s64 %rd254, %rd297, %rd253;
ld.global.nc.u32 %r1378, [%rd254];
shr.u16 %rs147, %rs134, 8;
cvt.u32.u16 %r1379, %rs147;
mul.wide.u16 %r1380, %rs137, 32;
and.b32 %r1381, %r1380, 256;
or.b32 %r1382, %r1381, %r1379;
mul.wide.u32 %rd255, %r1382, 4;
add.s64 %rd256, %rd297, %rd255;
ld.global.nc.u32 %r1383, [%rd256];
shr.u16 %rs148, %rs138, 8;
mul.wide.u16 %r1384, %rs148, 128;
cvt.u32.u16 %r1385, %rs148;
shl.b32 %r1386, %r1385, 21;
or.b32 %r1387, %r1384, %r1386;
and.b32 %r1388, %r1387, 25166208;
or.b32 %r1389, %r1388, -2139062144;
add.s32 %r1390, %r1389, -16843009;
or.b32 %r1302, %r1390, %r1388;

prmt.b32 %r1302, %r1302, 0, 0xba98;

mul.wide.u16 %r1391, %rs148, 8;
shl.b32 %r1392, %r1385, 17;
or.b32 %r1393, %r1391, %r1392;
and.b32 %r1394, %r1393, 25166208;
or.b32 %r1395, %r1394, -2139062144;
add.s32 %r1396, %r1395, -16843009;
or.b32 %r1304, %r1396, %r1394;

prmt.b32 %r1304, %r1304, 0, 0xba98;

xor.b32 %r1397, %r1302, %r1378;
or.b32 %r1398, %r1397, -2139062144;
and.b32 %r1399, %r1302, 2139062143;
and.b32 %r1400, %r1378, -2139062144;
sub.s32 %r1401, %r1398, %r1399;
xor.b32 %r1402, %r1400, %r1401;
xor.b32 %r1307, %r1402, -2139062144;
xor.b32 %r1403, %r1304, %r1383;
or.b32 %r1404, %r1403, -2139062144;
and.b32 %r1405, %r1304, 2139062143;
and.b32 %r1406, %r1383, -2139062144;
sub.s32 %r1407, %r1404, %r1405;
xor.b32 %r1408, %r1406, %r1407;
xor.b32 %r1311, %r1408, -2139062144;

dp4a.s32.s32 %r1306, %r1307, %r24, %r1298;


dp4a.s32.s32 %r1310, %r1311, %r25, %r1306;

and.b16 %rs149, %rs135, 255;
cvt.u32.u16 %r1409, %rs149;
mul.wide.u16 %r1410, %rs137, 16;
and.b32 %r1411, %r1410, 256;
or.b32 %r1412, %r1411, %r1409;
mul.wide.u32 %rd257, %r1412, 4;
add.s64 %rd258, %rd297, %rd257;
ld.global.nc.u32 %r1413, [%rd258];
shr.u16 %rs150, %rs135, 8;
cvt.u32.u16 %r1414, %rs150;
mul.wide.u16 %r1415, %rs137, 8;
and.b32 %r1416, %r1415, 256;
or.b32 %r1417, %r1416, %r1414;
mul.wide.u32 %rd259, %r1417, 4;
add.s64 %rd260, %rd297, %rd259;
ld.global.nc.u32 %r1418, [%rd260];
and.b16 %rs151, %rs139, 255;
mul.wide.u16 %r1419, %rs151, 128;
cvt.u32.u16 %r1420, %rs139;
shl.b32 %r1421, %r1420, 21;
or.b32 %r1422, %r1419, %r1421;
and.b32 %r1423, %r1422, 25166208;
or.b32 %r1424, %r1423, -2139062144;
add.s32 %r1425, %r1424, -16843009;
or.b32 %r1314, %r1425, %r1423;

prmt.b32 %r1314, %r1314, 0, 0xba98;

mul.wide.u16 %r1426, %rs151, 8;
shl.b32 %r1427, %r1420, 17;
or.b32 %r1428, %r1426, %r1427;
and.b32 %r1429, %r1428, 25166208;
or.b32 %r1430, %r1429, -2139062144;
add.s32 %r1431, %r1430, -16843009;
or.b32 %r1316, %r1431, %r1429;

prmt.b32 %r1316, %r1316, 0, 0xba98;

xor.b32 %r1432, %r1314, %r1413;
or.b32 %r1433, %r1432, -2139062144;
and.b32 %r1434, %r1314, 2139062143;
and.b32 %r1435, %r1413, -2139062144;
sub.s32 %r1436, %r1433, %r1434;
xor.b32 %r1437, %r1435, %r1436;
xor.b32 %r1319, %r1437, -2139062144;
xor.b32 %r1438, %r1316, %r1418;
or.b32 %r1439, %r1438, -2139062144;
and.b32 %r1440, %r1316, 2139062143;
and.b32 %r1441, %r1418, -2139062144;
sub.s32 %r1442, %r1439, %r1440;
xor.b32 %r1443, %r1441, %r1442;
xor.b32 %r1323, %r1443, -2139062144;

dp4a.s32.s32 %r1318, %r1319, %r26, %r1310;


dp4a.s32.s32 %r1322, %r1323, %r27, %r1318;

and.b16 %rs152, %rs136, 255;
cvt.u32.u16 %r1444, %rs152;
mul.wide.u16 %r1445, %rs137, 4;
and.b32 %r1446, %r1445, 256;
or.b32 %r1447, %r1446, %r1444;
mul.wide.u32 %rd261, %r1447, 4;
add.s64 %rd262, %rd297, %rd261;
ld.global.nc.u32 %r1448, [%rd262];
shr.u16 %rs153, %rs136, 8;
cvt.u32.u16 %r1449, %rs153;
mul.wide.u16 %r1450, %rs137, 2;
and.b32 %r1451, %r1450, 256;
or.b32 %r1452, %r1451, %r1449;
mul.wide.u32 %rd263, %r1452, 4;
add.s64 %rd264, %rd297, %rd263;
ld.global.nc.u32 %r1453, [%rd264];
shr.u16 %rs154, %rs139, 8;
mul.wide.u16 %r1454, %rs154, 128;
cvt.u32.u16 %r1455, %rs154;
shl.b32 %r1456, %r1455, 21;
or.b32 %r1457, %r1454, %r1456;
and.b32 %r1458, %r1457, 25166208;
or.b32 %r1459, %r1458, -2139062144;
add.s32 %r1460, %r1459, -16843009;
or.b32 %r1326, %r1460, %r1458;

prmt.b32 %r1326, %r1326, 0, 0xba98;

mul.wide.u16 %r1461, %rs154, 8;
shl.b32 %r1462, %r1455, 17;
or.b32 %r1463, %r1461, %r1462;
and.b32 %r1464, %r1463, 25166208;
or.b32 %r1465, %r1464, -2139062144;
add.s32 %r1466, %r1465, -16843009;
or.b32 %r1328, %r1466, %r1464;

prmt.b32 %r1328, %r1328, 0, 0xba98;

xor.b32 %r1467, %r1326, %r1448;
or.b32 %r1468, %r1467, -2139062144;
and.b32 %r1469, %r1326, 2139062143;
and.b32 %r1470, %r1448, -2139062144;
sub.s32 %r1471, %r1468, %r1469;
xor.b32 %r1472, %r1470, %r1471;
xor.b32 %r1331, %r1472, -2139062144;
xor.b32 %r1473, %r1328, %r1453;
or.b32 %r1474, %r1473, -2139062144;
and.b32 %r1475, %r1328, 2139062143;
and.b32 %r1476, %r1453, -2139062144;
sub.s32 %r1477, %r1474, %r1475;
xor.b32 %r1478, %r1476, %r1477;
xor.b32 %r1335, %r1478, -2139062144;

dp4a.s32.s32 %r1330, %r1331, %r28, %r1322;


dp4a.s32.s32 %r1334, %r1335, %r29, %r1330;

add.s64 %rd265, %rd43, %rd308;
ld.global.nc.u8 %rs155, [%rd265];
cvt.u32.u16 %r1479, %rs155;
and.b32 %r1480, %r1479, 255;
shr.u32 %r1481, %r1480, %r16;
shl.b32 %r1482, %r1481, 1;
and.b32 %r1483, %r1482, 30;
or.b32 %r1484, %r1483, 1;
mul.lo.s32 %r1485, %r1484, %r1334;
add.s64 %rd266, %rd44, %rd308;
ld.global.nc.u16 %rs133, [%rd266];

{ cvt.f32.f16 %f125, %rs133;}



{.reg .f16 low,high;
mov.b32 {low,high},%r30;
cvt.f32.f16 %f126, low;}


mul.ftz.f32 %f127, %f125, %f126;
cvt.rn.f32.s32 %f128, %r1485;
fma.rn.ftz.f32 %f27, %f127, %f128, %f27;
@%p13 bra $L__BB306_18;

mov.u64 %rd303, iq3s_grid;
or.b32 %r1537, %r4, 3;
mad.lo.s32 %r1538, %r53, %r1537, %r18;
mul.wide.s32 %rd267, %r1538, 110;
add.s64 %rd269, %rd267, %rd92;
add.s64 %rd270, %rd310, %rd269;
ld.global.u16 %rs157, [%rd270+4];
ld.global.u16 %rs158, [%rd270+6];
ld.global.u16 %rs159, [%rd270+8];
add.s64 %rd272, %rd267, %rd82;
add.s64 %rd273, %rd310, %rd272;
add.s64 %rd274, %rd310, %rd22;
add.s64 %rd275, %rd310, %rd21;
ld.global.u16 %rs160, [%rd274];
ld.global.u16 %rs161, [%rd275];
ld.global.u16 %rs162, [%rd270+2];
and.b16 %rs163, %rs162, 255;
cvt.u32.u16 %r1540, %rs163;
ld.global.u8 %rs164, [%rd273+66];
shl.b16 %rs165, %rs164, 8;
cvt.u32.u16 %r1541, %rs165;
and.b32 %r1542, %r1541, 256;
or.b32 %r1543, %r1542, %r1540;
mul.wide.u32 %rd276, %r1543, 4;
add.s64 %rd278, %rd303, %rd276;
ld.global.nc.u32 %r1544, [%rd278];
shr.u16 %rs166, %rs162, 8;
cvt.u32.u16 %r1545, %rs166;
mul.wide.u16 %r1546, %rs164, 128;
and.b32 %r1547, %r1546, 256;
or.b32 %r1548, %r1547, %r1545;
mul.wide.u32 %rd279, %r1548, 4;
add.s64 %rd280, %rd303, %rd279;
ld.global.nc.u32 %r1549, [%rd280];
and.b16 %rs167, %rs160, 255;
mul.wide.u16 %r1550, %rs167, 128;
cvt.u32.u16 %r1551, %rs160;
shl.b32 %r1552, %r1551, 21;
or.b32 %r1553, %r1550, %r1552;
and.b32 %r1554, %r1553, 25166208;
or.b32 %r1555, %r1554, -2139062144;
add.s32 %r1556, %r1555, -16843009;
or.b32 %r1486, %r1556, %r1554;

prmt.b32 %r1486, %r1486, 0, 0xba98;

mul.wide.u16 %r1557, %rs167, 8;
shl.b32 %r1558, %r1551, 17;
or.b32 %r1559, %r1557, %r1558;
and.b32 %r1560, %r1559, 25166208;
or.b32 %r1561, %r1560, -2139062144;
add.s32 %r1562, %r1561, -16843009;
or.b32 %r1488, %r1562, %r1560;

prmt.b32 %r1488, %r1488, 0, 0xba98;

xor.b32 %r1563, %r1486, %r1544;
or.b32 %r1564, %r1563, -2139062144;
and.b32 %r1565, %r1486, 2139062143;
and.b32 %r1566, %r1544, -2139062144;
sub.s32 %r1567, %r1564, %r1565;
xor.b32 %r1568, %r1566, %r1567;
xor.b32 %r1491, %r1568, -2139062144;
xor.b32 %r1569, %r1488, %r1549;
or.b32 %r1570, %r1569, -2139062144;
and.b32 %r1571, %r1488, 2139062143;
and.b32 %r1572, %r1549, -2139062144;
sub.s32 %r1573, %r1570, %r1571;
xor.b32 %r1574, %r1572, %r1573;
xor.b32 %r1495, %r1574, -2139062144;
mov.u32 %r1493, 0;

dp4a.s32.s32 %r1490, %r1491, %r22, %r1493;


dp4a.s32.s32 %r1494, %r1495, %r23, %r1490;

and.b16 %rs168, %rs157, 255;
cvt.u32.u16 %r1575, %rs168;
mul.wide.u16 %r1576, %rs164, 64;
and.b32 %r1577, %r1576, 256;
or.b32 %r1578, %r1577, %r1575;
mul.wide.u32 %rd281, %r1578, 4;
add.s64 %rd282, %rd303, %rd281;
ld.global.nc.u32 %r1579, [%rd282];
shr.u16 %rs169, %rs157, 8;
cvt.u32.u16 %r1580, %rs169;
mul.wide.u16 %r1581, %rs164, 32;
and.b32 %r1582, %r1581, 256;
or.b32 %r1583, %r1582, %r1580;
mul.wide.u32 %rd283, %r1583, 4;
add.s64 %rd284, %rd303, %rd283;
ld.global.nc.u32 %r1584, [%rd284];
shr.u16 %rs170, %rs160, 8;
mul.wide.u16 %r1585, %rs170, 128;
cvt.u32.u16 %r1586, %rs170;
shl.b32 %r1587, %r1586, 21;
or.b32 %r1588, %r1585, %r1587;
and.b32 %r1589, %r1588, 25166208;
or.b32 %r1590, %r1589, -2139062144;
add.s32 %r1591, %r1590, -16843009;
or.b32 %r1498, %r1591, %r1589;

prmt.b32 %r1498, %r1498, 0, 0xba98;

mul.wide.u16 %r1592, %rs170, 8;
shl.b32 %r1593, %r1586, 17;
or.b32 %r1594, %r1592, %r1593;
and.b32 %r1595, %r1594, 25166208;
or.b32 %r1596, %r1595, -2139062144;
add.s32 %r1597, %r1596, -16843009;
or.b32 %r1500, %r1597, %r1595;

prmt.b32 %r1500, %r1500, 0, 0xba98;

xor.b32 %r1598, %r1498, %r1579;
or.b32 %r1599, %r1598, -2139062144;
and.b32 %r1600, %r1498, 2139062143;
and.b32 %r1601, %r1579, -2139062144;
sub.s32 %r1602, %r1599, %r1600;
xor.b32 %r1603, %r1601, %r1602;
xor.b32 %r1503, %r1603, -2139062144;
xor.b32 %r1604, %r1500, %r1584;
or.b32 %r1605, %r1604, -2139062144;
and.b32 %r1606, %r1500, 2139062143;
and.b32 %r1607, %r1584, -2139062144;
sub.s32 %r1608, %r1605, %r1606;
xor.b32 %r1609, %r1607, %r1608;
xor.b32 %r1507, %r1609, -2139062144;

dp4a.s32.s32 %r1502, %r1503, %r24, %r1494;


dp4a.s32.s32 %r1506, %r1507, %r25, %r1502;

and.b16 %rs171, %rs158, 255;
cvt.u32.u16 %r1610, %rs171;
mul.wide.u16 %r1611, %rs164, 16;
and.b32 %r1612, %r1611, 256;
or.b32 %r1613, %r1612, %r1610;
mul.wide.u32 %rd285, %r1613, 4;
add.s64 %rd286, %rd303, %rd285;
ld.global.nc.u32 %r1614, [%rd286];
shr.u16 %rs172, %rs158, 8;
cvt.u32.u16 %r1615, %rs172;
mul.wide.u16 %r1616, %rs164, 8;
and.b32 %r1617, %r1616, 256;
or.b32 %r1618, %r1617, %r1615;
mul.wide.u32 %rd287, %r1618, 4;
add.s64 %rd288, %rd303, %rd287;
ld.global.nc.u32 %r1619, [%rd288];
and.b16 %rs173, %rs161, 255;
mul.wide.u16 %r1620, %rs173, 128;
cvt.u32.u16 %r1621, %rs161;
shl.b32 %r1622, %r1621, 21;
or.b32 %r1623, %r1620, %r1622;
and.b32 %r1624, %r1623, 25166208;
or.b32 %r1625, %r1624, -2139062144;
add.s32 %r1626, %r1625, -16843009;
or.b32 %r1510, %r1626, %r1624;

prmt.b32 %r1510, %r1510, 0, 0xba98;

mul.wide.u16 %r1627, %rs173, 8;
shl.b32 %r1628, %r1621, 17;
or.b32 %r1629, %r1627, %r1628;
and.b32 %r1630, %r1629, 25166208;
or.b32 %r1631, %r1630, -2139062144;
add.s32 %r1632, %r1631, -16843009;
or.b32 %r1512, %r1632, %r1630;

prmt.b32 %r1512, %r1512, 0, 0xba98;

xor.b32 %r1633, %r1510, %r1614;
or.b32 %r1634, %r1633, -2139062144;
and.b32 %r1635, %r1510, 2139062143;
and.b32 %r1636, %r1614, -2139062144;
sub.s32 %r1637, %r1634, %r1635;
xor.b32 %r1638, %r1636, %r1637;
xor.b32 %r1515, %r1638, -2139062144;
xor.b32 %r1639, %r1512, %r1619;
or.b32 %r1640, %r1639, -2139062144;
and.b32 %r1641, %r1512, 2139062143;
and.b32 %r1642, %r1619, -2139062144;
sub.s32 %r1643, %r1640, %r1641;
xor.b32 %r1644, %r1642, %r1643;
xor.b32 %r1519, %r1644, -2139062144;

dp4a.s32.s32 %r1514, %r1515, %r26, %r1506;


dp4a.s32.s32 %r1518, %r1519, %r27, %r1514;

and.b16 %rs174, %rs159, 255;
cvt.u32.u16 %r1645, %rs174;
mul.wide.u16 %r1646, %rs164, 4;
and.b32 %r1647, %r1646, 256;
or.b32 %r1648, %r1647, %r1645;
mul.wide.u32 %rd289, %r1648, 4;
add.s64 %rd290, %rd303, %rd289;
ld.global.nc.u32 %r1649, [%rd290];
shr.u16 %rs175, %rs159, 8;
cvt.u32.u16 %r1650, %rs175;
mul.wide.u16 %r1651, %rs164, 2;
and.b32 %r1652, %r1651, 256;
or.b32 %r1653, %r1652, %r1650;
mul.wide.u32 %rd291, %r1653, 4;
add.s64 %rd292, %rd303, %rd291;
ld.global.nc.u32 %r1654, [%rd292];
shr.u16 %rs176, %rs161, 8;
mul.wide.u16 %r1655, %rs176, 128;
cvt.u32.u16 %r1656, %rs176;
shl.b32 %r1657, %r1656, 21;
or.b32 %r1658, %r1655, %r1657;
and.b32 %r1659, %r1658, 25166208;
or.b32 %r1660, %r1659, -2139062144;
add.s32 %r1661, %r1660, -16843009;
or.b32 %r1522, %r1661, %r1659;

prmt.b32 %r1522, %r1522, 0, 0xba98;

mul.wide.u16 %r1662, %rs176, 8;
shl.b32 %r1663, %r1656, 17;
or.b32 %r1664, %r1662, %r1663;
and.b32 %r1665, %r1664, 25166208;
or.b32 %r1666, %r1665, -2139062144;
add.s32 %r1667, %r1666, -16843009;
or.b32 %r1524, %r1667, %r1665;

prmt.b32 %r1524, %r1524, 0, 0xba98;

xor.b32 %r1668, %r1522, %r1649;
or.b32 %r1669, %r1668, -2139062144;
and.b32 %r1670, %r1522, 2139062143;
and.b32 %r1671, %r1649, -2139062144;
sub.s32 %r1672, %r1669, %r1670;
xor.b32 %r1673, %r1671, %r1672;
xor.b32 %r1527, %r1673, -2139062144;
xor.b32 %r1674, %r1524, %r1654;
or.b32 %r1675, %r1674, -2139062144;
and.b32 %r1676, %r1524, 2139062143;
and.b32 %r1677, %r1654, -2139062144;
sub.s32 %r1678, %r1675, %r1676;
xor.b32 %r1679, %r1677, %r1678;
xor.b32 %r1531, %r1679, -2139062144;

dp4a.s32.s32 %r1526, %r1527, %r28, %r1518;


dp4a.s32.s32 %r1530, %r1531, %r29, %r1526;

add.s64 %rd293, %rd310, %rd20;
ld.global.u8 %r1680, [%rd293];
shr.u32 %r1681, %r1680, %r16;
shl.b32 %r1682, %r1681, 1;
and.b32 %r1683, %r1682, 30;
or.b32 %r1684, %r1683, 1;
mul.lo.s32 %r1685, %r1684, %r1530;
add.s64 %rd294, %rd310, %rd267;
ld.global.u16 %rs156, [%rd294];

{ cvt.f32.f16 %f129, %rs156;}



{.reg .f16 low,high;
mov.b32 {low,high},%r30;
cvt.f32.f16 %f130, low;}


mul.ftz.f32 %f131, %f129, %f130;
cvt.rn.f32.s32 %f132, %r1685;
fma.rn.ftz.f32 %f31, %f131, %f132, %f31;

$L__BB306_18:
add.s32 %r2432, %r2432, 128;
add.s64 %rd310, %rd310, 1760;
add.s64 %rd309, %rd309, 1760;
add.s32 %r2433, %r2433, 16;
setp.lt.s32 %p17, %r2433, %r87;
@%p17 bra $L__BB306_10;

$L__BB306_19:
mov.u32 %r2419, %tid.y;
setp.eq.s32 %p18, %r2419, 0;
@%p18 bra $L__BB306_23;

mov.u32 %r2420, %tid.y;
setp.ne.s64 %p19, %rd52, 0;
add.s32 %r33, %r2420, -1;
shl.b32 %r1690, %r2420, 9;
mov.u32 %r1691, _ZZ13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r1692, %r1691, %r1690;
shl.b32 %r1693, %r2, 2;
add.s32 %r1694, %r1692, %r1693;
st.shared.f32 [%r1694+-512], %f30;
@%p19 bra $L__BB306_22;
bra.uni $L__BB306_21;

$L__BB306_22:
shl.b32 %r1696, %r33, 9;
mov.u32 %r1697, _ZZ13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r1698, %r1697, %r1696;
add.s32 %r1700, %r1698, %r1693;
st.shared.f32 [%r1700], %f34;
add.s32 %r2416, %r1694, -512;
st.shared.f32 [%r2416+128], %f29;
st.shared.f32 [%r1700+128], %f33;
add.s32 %r2417, %r1694, -512;
st.shared.f32 [%r2417+256], %f28;
st.shared.f32 [%r1700+256], %f32;
add.s32 %r2418, %r1694, -512;
st.shared.f32 [%r2418+384], %f27;
st.shared.f32 [%r1700+384], %f31;
bra.uni $L__BB306_23;

$L__BB306_21:
add.s32 %r2413, %r1694, -512;
st.shared.f32 [%r2413+128], %f29;
add.s32 %r2414, %r1694, -512;
st.shared.f32 [%r2414+256], %f28;
add.s32 %r2415, %r1694, -512;
st.shared.f32 [%r2415+384], %f27;

$L__BB306_23:
mov.u32 %r2421, %tid.y;
setp.ne.s32 %p319, %r2421, 0;
bar.sync 0;
@%p319 bra $L__BB306_78;
bra.uni $L__BB306_24;

$L__BB306_78:
ret;

$L__BB306_24:
ld.param.u64 %rd298, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.f32 %f977, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
ld.param.u32 %r2424, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r2423, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
mov.u32 %r2422, %ctaid.z;
setp.ne.s64 %p21, %rd52, 0;
mov.u32 %r1703, %ctaid.y;
mad.lo.s32 %r1706, %r1703, %r2424, %r4;
mad.lo.s32 %r1707, %r2422, %r2423, %r1706;
neg.ftz.f32 %f35, %f977;
shl.b32 %r1709, %r2, 2;
mov.u32 %r1710, _ZZ13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r35, %r1710, %r1709;
cvta.to.global.u64 %rd295, %rd298;
mul.wide.u32 %rd296, %r1707, 4;
add.s64 %rd47, %rd295, %rd296;
@%p21 bra $L__BB306_33;
bra.uni $L__BB306_25;

$L__BB306_33:
ld.param.u32 %r2425, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
mov.u32 %r1797, _ZZ13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r37, %r1797, %r1709;
setp.gt.s32 %p54, %r2425, 2;
@%p54 bra $L__BB306_44;

ld.param.u32 %r2428, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p57, %r2428, 1;
@%p57 bra $L__BB306_62;

ld.param.u32 %r2429, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p58, %r2429, 2;
@%p58 bra $L__BB306_36;
bra.uni $L__BB306_70;

$L__BB306_36:
ld.shared.f32 %f533, [%r35];
add.ftz.f32 %f534, %f533, %f30;
ld.shared.f32 %f535, [%r37];
add.ftz.f32 %f536, %f535, %f34;
ld.shared.f32 %f537, [%r35+512];
add.ftz.f32 %f538, %f537, %f534;
ld.shared.f32 %f539, [%r37+512];
add.ftz.f32 %f540, %f539, %f536;
ld.shared.f32 %f541, [%r35+1024];
add.ftz.f32 %f542, %f541, %f538;
ld.shared.f32 %f543, [%r37+1024];
add.ftz.f32 %f544, %f543, %f540;
mov.b32 %r2044, %f542;
mov.u32 %r2045, 31;
mov.u32 %r2046, 16;
mov.u32 %r2047, -1;
shfl.sync.bfly.b32 %r2048|%p163, %r2044, %r2046, %r2045, %r2047;
mov.b32 %f545, %r2048;
add.ftz.f32 %f546, %f542, %f545;
mov.b32 %r2049, %f546;
mov.u32 %r2050, 8;
shfl.sync.bfly.b32 %r2051|%p164, %r2049, %r2050, %r2045, %r2047;
mov.b32 %f547, %r2051;
add.ftz.f32 %f548, %f546, %f547;
mov.b32 %r2052, %f548;
mov.u32 %r2053, 4;
shfl.sync.bfly.b32 %r2054|%p165, %r2052, %r2053, %r2045, %r2047;
mov.b32 %f549, %r2054;
add.ftz.f32 %f550, %f548, %f549;
mov.b32 %r2055, %f550;
mov.u32 %r2056, 2;
shfl.sync.bfly.b32 %r2057|%p166, %r2055, %r2056, %r2045, %r2047;
mov.b32 %f551, %r2057;
add.ftz.f32 %f552, %f550, %f551;
mov.b32 %r2058, %f552;
mov.u32 %r2059, 1;
shfl.sync.bfly.b32 %r2060|%p167, %r2058, %r2059, %r2045, %r2047;
mov.b32 %f553, %r2060;
add.ftz.f32 %f56, %f552, %f553;
mov.b32 %r2061, %f544;
shfl.sync.bfly.b32 %r2062|%p168, %r2061, %r2046, %r2045, %r2047;
mov.b32 %f554, %r2062;
add.ftz.f32 %f555, %f544, %f554;
mov.b32 %r2063, %f555;
shfl.sync.bfly.b32 %r2064|%p169, %r2063, %r2050, %r2045, %r2047;
mov.b32 %f556, %r2064;
add.ftz.f32 %f557, %f555, %f556;
mov.b32 %r2065, %f557;
shfl.sync.bfly.b32 %r2066|%p170, %r2065, %r2053, %r2045, %r2047;
mov.b32 %f558, %r2066;
add.ftz.f32 %f559, %f557, %f558;
mov.b32 %r2067, %f559;
shfl.sync.bfly.b32 %r2068|%p171, %r2067, %r2056, %r2045, %r2047;
mov.b32 %f560, %r2068;
add.ftz.f32 %f561, %f559, %f560;
mov.b32 %r2069, %f561;
shfl.sync.bfly.b32 %r2070|%p172, %r2069, %r2059, %r2045, %r2047;
mov.b32 %f562, %r2070;
add.ftz.f32 %f57, %f561, %f562;
setp.ne.s32 %p173, %r2, 0;
setp.ge.u32 %p174, %r4, %r54;
or.pred %p175, %p173, %p174;
@%p175 bra $L__BB306_38;

add.ftz.f32 %f563, %f5, %f56;
add.ftz.f32 %f564, %f4, %f57;
mul.ftz.f32 %f565, %f564, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f566, %f565;
add.ftz.f32 %f567, %f566, 0f3F800000;
div.approx.ftz.f32 %f568, %f564, %f567;
mul.ftz.f32 %f569, %f563, %f568;
st.global.f32 [%rd47], %f569;

$L__BB306_38:
ld.shared.f32 %f570, [%r35+128];
add.ftz.f32 %f571, %f570, %f29;
ld.shared.f32 %f572, [%r37+128];
add.ftz.f32 %f573, %f572, %f33;
ld.shared.f32 %f574, [%r35+640];
add.ftz.f32 %f575, %f574, %f571;
ld.shared.f32 %f576, [%r37+640];
add.ftz.f32 %f577, %f576, %f573;
ld.shared.f32 %f578, [%r35+1152];
add.ftz.f32 %f579, %f578, %f575;
ld.shared.f32 %f580, [%r37+1152];
add.ftz.f32 %f581, %f580, %f577;
mov.b32 %r2074, %f579;
shfl.sync.bfly.b32 %r2078|%p176, %r2074, %r2046, %r2045, %r2047;
mov.b32 %f582, %r2078;
add.ftz.f32 %f583, %f579, %f582;
mov.b32 %r2079, %f583;
shfl.sync.bfly.b32 %r2081|%p177, %r2079, %r2050, %r2045, %r2047;
mov.b32 %f584, %r2081;
add.ftz.f32 %f585, %f583, %f584;
mov.b32 %r2082, %f585;
shfl.sync.bfly.b32 %r2084|%p178, %r2082, %r2053, %r2045, %r2047;
mov.b32 %f586, %r2084;
add.ftz.f32 %f587, %f585, %f586;
mov.b32 %r2085, %f587;
shfl.sync.bfly.b32 %r2087|%p179, %r2085, %r2056, %r2045, %r2047;
mov.b32 %f588, %r2087;
add.ftz.f32 %f589, %f587, %f588;
mov.b32 %r2088, %f589;
shfl.sync.bfly.b32 %r2090|%p180, %r2088, %r2059, %r2045, %r2047;
mov.b32 %f590, %r2090;
add.ftz.f32 %f58, %f589, %f590;
mov.b32 %r2091, %f581;
shfl.sync.bfly.b32 %r2092|%p181, %r2091, %r2046, %r2045, %r2047;
mov.b32 %f591, %r2092;
add.ftz.f32 %f592, %f581, %f591;
mov.b32 %r2093, %f592;
shfl.sync.bfly.b32 %r2094|%p182, %r2093, %r2050, %r2045, %r2047;
mov.b32 %f593, %r2094;
add.ftz.f32 %f594, %f592, %f593;
mov.b32 %r2095, %f594;
shfl.sync.bfly.b32 %r2096|%p183, %r2095, %r2053, %r2045, %r2047;
mov.b32 %f595, %r2096;
add.ftz.f32 %f596, %f594, %f595;
mov.b32 %r2097, %f596;
shfl.sync.bfly.b32 %r2098|%p184, %r2097, %r2056, %r2045, %r2047;
mov.b32 %f597, %r2098;
add.ftz.f32 %f598, %f596, %f597;
mov.b32 %r2099, %f598;
shfl.sync.bfly.b32 %r2100|%p185, %r2099, %r2059, %r2045, %r2047;
mov.b32 %f599, %r2100;
add.ftz.f32 %f59, %f598, %f599;
setp.ne.s32 %p186, %r2, 1;
or.b32 %r2104, %r4, 1;
setp.ge.u32 %p187, %r2104, %r54;
or.pred %p188, %p186, %p187;
@%p188 bra $L__BB306_40;

add.ftz.f32 %f600, %f5, %f58;
add.ftz.f32 %f601, %f4, %f59;
mul.ftz.f32 %f602, %f601, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f603, %f602;
add.ftz.f32 %f604, %f603, 0f3F800000;
div.approx.ftz.f32 %f605, %f601, %f604;
mul.ftz.f32 %f606, %f600, %f605;
st.global.f32 [%rd47+4], %f606;

$L__BB306_40:
ld.shared.f32 %f607, [%r35+256];
add.ftz.f32 %f608, %f607, %f28;
ld.shared.f32 %f609, [%r37+256];
add.ftz.f32 %f610, %f609, %f32;
ld.shared.f32 %f611, [%r35+768];
add.ftz.f32 %f612, %f611, %f608;
ld.shared.f32 %f613, [%r37+768];
add.ftz.f32 %f614, %f613, %f610;
ld.shared.f32 %f615, [%r35+1280];
add.ftz.f32 %f616, %f615, %f612;
ld.shared.f32 %f617, [%r37+1280];
add.ftz.f32 %f618, %f617, %f614;
mov.b32 %r2105, %f616;
mov.u32 %r2106, 31;
mov.u32 %r2107, 16;
mov.u32 %r2108, -1;
shfl.sync.bfly.b32 %r2109|%p189, %r2105, %r2107, %r2106, %r2108;
mov.b32 %f619, %r2109;
add.ftz.f32 %f620, %f616, %f619;
mov.b32 %r2110, %f620;
mov.u32 %r2111, 8;
shfl.sync.bfly.b32 %r2112|%p190, %r2110, %r2111, %r2106, %r2108;
mov.b32 %f621, %r2112;
add.ftz.f32 %f622, %f620, %f621;
mov.b32 %r2113, %f622;
mov.u32 %r2114, 4;
shfl.sync.bfly.b32 %r2115|%p191, %r2113, %r2114, %r2106, %r2108;
mov.b32 %f623, %r2115;
add.ftz.f32 %f624, %f622, %f623;
mov.b32 %r2116, %f624;
mov.u32 %r2117, 2;
shfl.sync.bfly.b32 %r2118|%p192, %r2116, %r2117, %r2106, %r2108;
mov.b32 %f625, %r2118;
add.ftz.f32 %f626, %f624, %f625;
mov.b32 %r2119, %f626;
mov.u32 %r2120, 1;
shfl.sync.bfly.b32 %r2121|%p193, %r2119, %r2120, %r2106, %r2108;
mov.b32 %f627, %r2121;
add.ftz.f32 %f60, %f626, %f627;
mov.b32 %r2122, %f618;
shfl.sync.bfly.b32 %r2123|%p194, %r2122, %r2107, %r2106, %r2108;
mov.b32 %f628, %r2123;
add.ftz.f32 %f629, %f618, %f628;
mov.b32 %r2124, %f629;
shfl.sync.bfly.b32 %r2125|%p195, %r2124, %r2111, %r2106, %r2108;
mov.b32 %f630, %r2125;
add.ftz.f32 %f631, %f629, %f630;
mov.b32 %r2126, %f631;
shfl.sync.bfly.b32 %r2127|%p196, %r2126, %r2114, %r2106, %r2108;
mov.b32 %f632, %r2127;
add.ftz.f32 %f633, %f631, %f632;
mov.b32 %r2128, %f633;
shfl.sync.bfly.b32 %r2129|%p197, %r2128, %r2117, %r2106, %r2108;
mov.b32 %f634, %r2129;
add.ftz.f32 %f635, %f633, %f634;
mov.b32 %r2130, %f635;
shfl.sync.bfly.b32 %r2131|%p198, %r2130, %r2120, %r2106, %r2108;
mov.b32 %f636, %r2131;
add.ftz.f32 %f61, %f635, %f636;
setp.ne.s32 %p199, %r2, 2;
or.b32 %r2135, %r4, 2;
setp.ge.u32 %p200, %r2135, %r54;
or.pred %p201, %p199, %p200;
@%p201 bra $L__BB306_42;

add.ftz.f32 %f637, %f5, %f60;
add.ftz.f32 %f638, %f4, %f61;
mul.ftz.f32 %f639, %f638, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f640, %f639;
add.ftz.f32 %f641, %f640, 0f3F800000;
div.approx.ftz.f32 %f642, %f638, %f641;
mul.ftz.f32 %f643, %f637, %f642;
st.global.f32 [%rd47+8], %f643;

$L__BB306_42:
ld.shared.f32 %f644, [%r35+384];
add.ftz.f32 %f645, %f644, %f27;
ld.shared.f32 %f646, [%r37+384];
add.ftz.f32 %f647, %f646, %f31;
ld.shared.f32 %f648, [%r35+896];
add.ftz.f32 %f649, %f648, %f645;
ld.shared.f32 %f650, [%r37+896];
add.ftz.f32 %f651, %f650, %f647;
ld.shared.f32 %f652, [%r35+1408];
add.ftz.f32 %f653, %f652, %f649;
ld.shared.f32 %f654, [%r37+1408];
add.ftz.f32 %f655, %f654, %f651;
mov.b32 %r2136, %f653;
shfl.sync.bfly.b32 %r2140|%p202, %r2136, %r2107, %r2106, %r2108;
mov.b32 %f656, %r2140;
add.ftz.f32 %f657, %f653, %f656;
mov.b32 %r2141, %f657;
shfl.sync.bfly.b32 %r2143|%p203, %r2141, %r2111, %r2106, %r2108;
mov.b32 %f658, %r2143;
add.ftz.f32 %f659, %f657, %f658;
mov.b32 %r2144, %f659;
shfl.sync.bfly.b32 %r2146|%p204, %r2144, %r2114, %r2106, %r2108;
mov.b32 %f660, %r2146;
add.ftz.f32 %f661, %f659, %f660;
mov.b32 %r2147, %f661;
shfl.sync.bfly.b32 %r2149|%p205, %r2147, %r2117, %r2106, %r2108;
mov.b32 %f662, %r2149;
add.ftz.f32 %f663, %f661, %f662;
mov.b32 %r2150, %f663;
shfl.sync.bfly.b32 %r2152|%p206, %r2150, %r2120, %r2106, %r2108;
mov.b32 %f664, %r2152;
add.ftz.f32 %f62, %f663, %f664;
mov.b32 %r2153, %f655;
shfl.sync.bfly.b32 %r2154|%p207, %r2153, %r2107, %r2106, %r2108;
mov.b32 %f665, %r2154;
add.ftz.f32 %f666, %f655, %f665;
mov.b32 %r2155, %f666;
shfl.sync.bfly.b32 %r2156|%p208, %r2155, %r2111, %r2106, %r2108;
mov.b32 %f667, %r2156;
add.ftz.f32 %f668, %f666, %f667;
mov.b32 %r2157, %f668;
shfl.sync.bfly.b32 %r2158|%p209, %r2157, %r2114, %r2106, %r2108;
mov.b32 %f669, %r2158;
add.ftz.f32 %f670, %f668, %f669;
mov.b32 %r2159, %f670;
shfl.sync.bfly.b32 %r2160|%p210, %r2159, %r2117, %r2106, %r2108;
mov.b32 %f671, %r2160;
add.ftz.f32 %f672, %f670, %f671;
mov.b32 %r2161, %f672;
shfl.sync.bfly.b32 %r2162|%p211, %r2161, %r2120, %r2106, %r2108;
mov.b32 %f673, %r2162;
add.ftz.f32 %f63, %f672, %f673;
setp.ne.s32 %p212, %r2, 3;
or.b32 %r2166, %r4, 3;
setp.ge.u32 %p213, %r2166, %r54;
or.pred %p214, %p212, %p213;
@%p214 bra $L__BB306_78;

add.ftz.f32 %f674, %f5, %f62;
add.ftz.f32 %f675, %f4, %f63;
mul.ftz.f32 %f676, %f675, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f677, %f676;
add.ftz.f32 %f678, %f677, 0f3F800000;
div.approx.ftz.f32 %f679, %f675, %f678;
mul.ftz.f32 %f680, %f674, %f679;
st.global.f32 [%rd47+12], %f680;
bra.uni $L__BB306_78;

$L__BB306_25:
ld.shared.f32 %f133, [%r35];
add.ftz.f32 %f134, %f133, %f30;
ld.shared.f32 %f135, [%r35+512];
add.ftz.f32 %f136, %f135, %f134;
ld.shared.f32 %f137, [%r35+1024];
add.ftz.f32 %f138, %f137, %f136;
mov.b32 %r1711, %f138;
mov.u32 %r1712, 31;
mov.u32 %r1713, 16;
mov.u32 %r1714, -1;
shfl.sync.bfly.b32 %r1715|%p22, %r1711, %r1713, %r1712, %r1714;
mov.b32 %f139, %r1715;
add.ftz.f32 %f140, %f138, %f139;
mov.b32 %r1716, %f140;
mov.u32 %r1717, 8;
shfl.sync.bfly.b32 %r1718|%p23, %r1716, %r1717, %r1712, %r1714;
mov.b32 %f141, %r1718;
add.ftz.f32 %f142, %f140, %f141;
mov.b32 %r1719, %f142;
mov.u32 %r1720, 4;
shfl.sync.bfly.b32 %r1721|%p24, %r1719, %r1720, %r1712, %r1714;
mov.b32 %f143, %r1721;
add.ftz.f32 %f144, %f142, %f143;
mov.b32 %r1722, %f144;
mov.u32 %r1723, 2;
shfl.sync.bfly.b32 %r1724|%p25, %r1722, %r1723, %r1712, %r1714;
mov.b32 %f145, %r1724;
add.ftz.f32 %f146, %f144, %f145;
mov.b32 %r1725, %f146;
mov.u32 %r1726, 1;
shfl.sync.bfly.b32 %r1727|%p26, %r1725, %r1726, %r1712, %r1714;
mov.b32 %f147, %r1727;
add.ftz.f32 %f36, %f146, %f147;
setp.ne.s32 %p27, %r2, 0;
setp.ge.u32 %p28, %r4, %r54;
or.pred %p29, %p27, %p28;
@%p29 bra $L__BB306_27;

add.ftz.f32 %f148, %f5, %f36;
st.global.f32 [%rd47], %f148;

$L__BB306_27:
ld.shared.f32 %f149, [%r35+128];
add.ftz.f32 %f150, %f149, %f29;
ld.shared.f32 %f151, [%r35+640];
add.ftz.f32 %f152, %f151, %f150;
ld.shared.f32 %f153, [%r35+1152];
add.ftz.f32 %f154, %f153, %f152;
mov.b32 %r1731, %f154;
shfl.sync.bfly.b32 %r1735|%p30, %r1731, %r1713, %r1712, %r1714;
mov.b32 %f155, %r1735;
add.ftz.f32 %f156, %f154, %f155;
mov.b32 %r1736, %f156;
shfl.sync.bfly.b32 %r1738|%p31, %r1736, %r1717, %r1712, %r1714;
mov.b32 %f157, %r1738;
add.ftz.f32 %f158, %f156, %f157;
mov.b32 %r1739, %f158;
shfl.sync.bfly.b32 %r1741|%p32, %r1739, %r1720, %r1712, %r1714;
mov.b32 %f159, %r1741;
add.ftz.f32 %f160, %f158, %f159;
mov.b32 %r1742, %f160;
shfl.sync.bfly.b32 %r1744|%p33, %r1742, %r1723, %r1712, %r1714;
mov.b32 %f161, %r1744;
add.ftz.f32 %f162, %f160, %f161;
mov.b32 %r1745, %f162;
shfl.sync.bfly.b32 %r1747|%p34, %r1745, %r1726, %r1712, %r1714;
mov.b32 %f163, %r1747;
add.ftz.f32 %f37, %f162, %f163;
setp.ne.s32 %p35, %r2, 1;
or.b32 %r1751, %r4, 1;
setp.ge.u32 %p36, %r1751, %r54;
or.pred %p37, %p35, %p36;
@%p37 bra $L__BB306_29;

add.ftz.f32 %f164, %f5, %f37;
st.global.f32 [%rd47+4], %f164;

$L__BB306_29:
ld.shared.f32 %f165, [%r35+256];
add.ftz.f32 %f166, %f165, %f28;
ld.shared.f32 %f167, [%r35+768];
add.ftz.f32 %f168, %f167, %f166;
ld.shared.f32 %f169, [%r35+1280];
add.ftz.f32 %f170, %f169, %f168;
mov.b32 %r1752, %f170;
mov.u32 %r1753, 31;
mov.u32 %r1754, 16;
mov.u32 %r1755, -1;
shfl.sync.bfly.b32 %r1756|%p38, %r1752, %r1754, %r1753, %r1755;
mov.b32 %f171, %r1756;
add.ftz.f32 %f172, %f170, %f171;
mov.b32 %r1757, %f172;
mov.u32 %r1758, 8;
shfl.sync.bfly.b32 %r1759|%p39, %r1757, %r1758, %r1753, %r1755;
mov.b32 %f173, %r1759;
add.ftz.f32 %f174, %f172, %f173;
mov.b32 %r1760, %f174;
mov.u32 %r1761, 4;
shfl.sync.bfly.b32 %r1762|%p40, %r1760, %r1761, %r1753, %r1755;
mov.b32 %f175, %r1762;
add.ftz.f32 %f176, %f174, %f175;
mov.b32 %r1763, %f176;
mov.u32 %r1764, 2;
shfl.sync.bfly.b32 %r1765|%p41, %r1763, %r1764, %r1753, %r1755;
mov.b32 %f177, %r1765;
add.ftz.f32 %f178, %f176, %f177;
mov.b32 %r1766, %f178;
mov.u32 %r1767, 1;
shfl.sync.bfly.b32 %r1768|%p42, %r1766, %r1767, %r1753, %r1755;
mov.b32 %f179, %r1768;
add.ftz.f32 %f38, %f178, %f179;
setp.ne.s32 %p43, %r2, 2;
or.b32 %r1772, %r4, 2;
setp.ge.u32 %p44, %r1772, %r54;
or.pred %p45, %p43, %p44;
@%p45 bra $L__BB306_31;

add.ftz.f32 %f180, %f5, %f38;
st.global.f32 [%rd47+8], %f180;

$L__BB306_31:
ld.shared.f32 %f181, [%r35+384];
add.ftz.f32 %f182, %f181, %f27;
ld.shared.f32 %f183, [%r35+896];
add.ftz.f32 %f184, %f183, %f182;
ld.shared.f32 %f185, [%r35+1408];
add.ftz.f32 %f186, %f185, %f184;
mov.b32 %r1773, %f186;
shfl.sync.bfly.b32 %r1777|%p46, %r1773, %r1754, %r1753, %r1755;
mov.b32 %f187, %r1777;
add.ftz.f32 %f188, %f186, %f187;
mov.b32 %r1778, %f188;
shfl.sync.bfly.b32 %r1780|%p47, %r1778, %r1758, %r1753, %r1755;
mov.b32 %f189, %r1780;
add.ftz.f32 %f190, %f188, %f189;
mov.b32 %r1781, %f190;
shfl.sync.bfly.b32 %r1783|%p48, %r1781, %r1761, %r1753, %r1755;
mov.b32 %f191, %r1783;
add.ftz.f32 %f192, %f190, %f191;
mov.b32 %r1784, %f192;
shfl.sync.bfly.b32 %r1786|%p49, %r1784, %r1764, %r1753, %r1755;
mov.b32 %f193, %r1786;
add.ftz.f32 %f194, %f192, %f193;
mov.b32 %r1787, %f194;
shfl.sync.bfly.b32 %r1789|%p50, %r1787, %r1767, %r1753, %r1755;
mov.b32 %f195, %r1789;
add.ftz.f32 %f39, %f194, %f195;
setp.ne.s32 %p51, %r2, 3;
or.b32 %r1793, %r4, 3;
setp.ge.u32 %p52, %r1793, %r54;
or.pred %p53, %p51, %p52;
@%p53 bra $L__BB306_78;

add.ftz.f32 %f196, %f5, %f39;
st.global.f32 [%rd47+12], %f196;
bra.uni $L__BB306_78;

$L__BB306_44:
ld.param.u32 %r2426, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p55, %r2426, 3;
@%p55 bra $L__BB306_54;

ld.param.u32 %r2427, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.ne.s32 %p56, %r2427, 6;
@%p56 bra $L__BB306_70;

ld.shared.f32 %f197, [%r35];
add.ftz.f32 %f198, %f197, %f30;
ld.shared.f32 %f199, [%r37];
add.ftz.f32 %f200, %f199, %f34;
ld.shared.f32 %f201, [%r35+512];
add.ftz.f32 %f202, %f201, %f198;
ld.shared.f32 %f203, [%r37+512];
add.ftz.f32 %f204, %f203, %f200;
ld.shared.f32 %f205, [%r35+1024];
add.ftz.f32 %f206, %f205, %f202;
ld.shared.f32 %f207, [%r37+1024];
add.ftz.f32 %f208, %f207, %f204;
mov.b32 %r1798, %f206;
mov.u32 %r1799, 31;
mov.u32 %r1800, 16;
mov.u32 %r1801, -1;
shfl.sync.bfly.b32 %r1802|%p59, %r1798, %r1800, %r1799, %r1801;
mov.b32 %f209, %r1802;
add.ftz.f32 %f210, %f206, %f209;
mov.b32 %r1803, %f210;
mov.u32 %r1804, 8;
shfl.sync.bfly.b32 %r1805|%p60, %r1803, %r1804, %r1799, %r1801;
mov.b32 %f211, %r1805;
add.ftz.f32 %f212, %f210, %f211;
mov.b32 %r1806, %f212;
mov.u32 %r1807, 4;
shfl.sync.bfly.b32 %r1808|%p61, %r1806, %r1807, %r1799, %r1801;
mov.b32 %f213, %r1808;
add.ftz.f32 %f214, %f212, %f213;
mov.b32 %r1809, %f214;
mov.u32 %r1810, 2;
shfl.sync.bfly.b32 %r1811|%p62, %r1809, %r1810, %r1799, %r1801;
mov.b32 %f215, %r1811;
add.ftz.f32 %f216, %f214, %f215;
mov.b32 %r1812, %f216;
mov.u32 %r1813, 1;
shfl.sync.bfly.b32 %r1814|%p63, %r1812, %r1813, %r1799, %r1801;
mov.b32 %f217, %r1814;
add.ftz.f32 %f40, %f216, %f217;
mov.b32 %r1815, %f208;
shfl.sync.bfly.b32 %r1816|%p64, %r1815, %r1800, %r1799, %r1801;
mov.b32 %f218, %r1816;
add.ftz.f32 %f219, %f208, %f218;
mov.b32 %r1817, %f219;
shfl.sync.bfly.b32 %r1818|%p65, %r1817, %r1804, %r1799, %r1801;
mov.b32 %f220, %r1818;
add.ftz.f32 %f221, %f219, %f220;
mov.b32 %r1819, %f221;
shfl.sync.bfly.b32 %r1820|%p66, %r1819, %r1807, %r1799, %r1801;
mov.b32 %f222, %r1820;
add.ftz.f32 %f223, %f221, %f222;
mov.b32 %r1821, %f223;
shfl.sync.bfly.b32 %r1822|%p67, %r1821, %r1810, %r1799, %r1801;
mov.b32 %f224, %r1822;
add.ftz.f32 %f225, %f223, %f224;
mov.b32 %r1823, %f225;
shfl.sync.bfly.b32 %r1824|%p68, %r1823, %r1813, %r1799, %r1801;
mov.b32 %f226, %r1824;
add.ftz.f32 %f41, %f225, %f226;
setp.ne.s32 %p69, %r2, 0;
setp.ge.u32 %p70, %r4, %r54;
or.pred %p71, %p69, %p70;
@%p71 bra $L__BB306_48;

ld.param.f32 %f981, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f227, %f5, %f40;
add.ftz.f32 %f228, %f4, %f41;
min.ftz.f32 %f229, %f228, %f981;
min.ftz.f32 %f230, %f227, %f981;
max.ftz.f32 %f231, %f230, %f35;
mul.ftz.f32 %f232, %f229, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f233, %f232;
add.ftz.f32 %f234, %f233, 0f3F800000;
div.approx.ftz.f32 %f235, %f229, %f234;
mul.ftz.f32 %f236, %f231, %f235;
st.global.f32 [%rd47], %f236;

$L__BB306_48:
ld.shared.f32 %f237, [%r35+128];
add.ftz.f32 %f238, %f237, %f29;
ld.shared.f32 %f239, [%r37+128];
add.ftz.f32 %f240, %f239, %f33;
ld.shared.f32 %f241, [%r35+640];
add.ftz.f32 %f242, %f241, %f238;
ld.shared.f32 %f243, [%r37+640];
add.ftz.f32 %f244, %f243, %f240;
ld.shared.f32 %f245, [%r35+1152];
add.ftz.f32 %f246, %f245, %f242;
ld.shared.f32 %f247, [%r37+1152];
add.ftz.f32 %f248, %f247, %f244;
mov.b32 %r1828, %f246;
shfl.sync.bfly.b32 %r1832|%p72, %r1828, %r1800, %r1799, %r1801;
mov.b32 %f249, %r1832;
add.ftz.f32 %f250, %f246, %f249;
mov.b32 %r1833, %f250;
shfl.sync.bfly.b32 %r1835|%p73, %r1833, %r1804, %r1799, %r1801;
mov.b32 %f251, %r1835;
add.ftz.f32 %f252, %f250, %f251;
mov.b32 %r1836, %f252;
shfl.sync.bfly.b32 %r1838|%p74, %r1836, %r1807, %r1799, %r1801;
mov.b32 %f253, %r1838;
add.ftz.f32 %f254, %f252, %f253;
mov.b32 %r1839, %f254;
shfl.sync.bfly.b32 %r1841|%p75, %r1839, %r1810, %r1799, %r1801;
mov.b32 %f255, %r1841;
add.ftz.f32 %f256, %f254, %f255;
mov.b32 %r1842, %f256;
shfl.sync.bfly.b32 %r1844|%p76, %r1842, %r1813, %r1799, %r1801;
mov.b32 %f257, %r1844;
add.ftz.f32 %f42, %f256, %f257;
mov.b32 %r1845, %f248;
shfl.sync.bfly.b32 %r1846|%p77, %r1845, %r1800, %r1799, %r1801;
mov.b32 %f258, %r1846;
add.ftz.f32 %f259, %f248, %f258;
mov.b32 %r1847, %f259;
shfl.sync.bfly.b32 %r1848|%p78, %r1847, %r1804, %r1799, %r1801;
mov.b32 %f260, %r1848;
add.ftz.f32 %f261, %f259, %f260;
mov.b32 %r1849, %f261;
shfl.sync.bfly.b32 %r1850|%p79, %r1849, %r1807, %r1799, %r1801;
mov.b32 %f262, %r1850;
add.ftz.f32 %f263, %f261, %f262;
mov.b32 %r1851, %f263;
shfl.sync.bfly.b32 %r1852|%p80, %r1851, %r1810, %r1799, %r1801;
mov.b32 %f264, %r1852;
add.ftz.f32 %f265, %f263, %f264;
mov.b32 %r1853, %f265;
shfl.sync.bfly.b32 %r1854|%p81, %r1853, %r1813, %r1799, %r1801;
mov.b32 %f266, %r1854;
add.ftz.f32 %f43, %f265, %f266;
setp.ne.s32 %p82, %r2, 1;
or.b32 %r1858, %r4, 1;
setp.ge.u32 %p83, %r1858, %r54;
or.pred %p84, %p82, %p83;
@%p84 bra $L__BB306_50;

ld.param.f32 %f980, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f267, %f5, %f42;
add.ftz.f32 %f268, %f4, %f43;
min.ftz.f32 %f269, %f268, %f980;
min.ftz.f32 %f270, %f267, %f980;
max.ftz.f32 %f271, %f270, %f35;
mul.ftz.f32 %f272, %f269, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f273, %f272;
add.ftz.f32 %f274, %f273, 0f3F800000;
div.approx.ftz.f32 %f275, %f269, %f274;
mul.ftz.f32 %f276, %f271, %f275;
st.global.f32 [%rd47+4], %f276;

$L__BB306_50:
ld.shared.f32 %f277, [%r35+256];
add.ftz.f32 %f278, %f277, %f28;
ld.shared.f32 %f279, [%r37+256];
add.ftz.f32 %f280, %f279, %f32;
ld.shared.f32 %f281, [%r35+768];
add.ftz.f32 %f282, %f281, %f278;
ld.shared.f32 %f283, [%r37+768];
add.ftz.f32 %f284, %f283, %f280;
ld.shared.f32 %f285, [%r35+1280];
add.ftz.f32 %f286, %f285, %f282;
ld.shared.f32 %f287, [%r37+1280];
add.ftz.f32 %f288, %f287, %f284;
mov.b32 %r1859, %f286;
mov.u32 %r1860, 31;
mov.u32 %r1861, 16;
mov.u32 %r1862, -1;
shfl.sync.bfly.b32 %r1863|%p85, %r1859, %r1861, %r1860, %r1862;
mov.b32 %f289, %r1863;
add.ftz.f32 %f290, %f286, %f289;
mov.b32 %r1864, %f290;
mov.u32 %r1865, 8;
shfl.sync.bfly.b32 %r1866|%p86, %r1864, %r1865, %r1860, %r1862;
mov.b32 %f291, %r1866;
add.ftz.f32 %f292, %f290, %f291;
mov.b32 %r1867, %f292;
mov.u32 %r1868, 4;
shfl.sync.bfly.b32 %r1869|%p87, %r1867, %r1868, %r1860, %r1862;
mov.b32 %f293, %r1869;
add.ftz.f32 %f294, %f292, %f293;
mov.b32 %r1870, %f294;
mov.u32 %r1871, 2;
shfl.sync.bfly.b32 %r1872|%p88, %r1870, %r1871, %r1860, %r1862;
mov.b32 %f295, %r1872;
add.ftz.f32 %f296, %f294, %f295;
mov.b32 %r1873, %f296;
mov.u32 %r1874, 1;
shfl.sync.bfly.b32 %r1875|%p89, %r1873, %r1874, %r1860, %r1862;
mov.b32 %f297, %r1875;
add.ftz.f32 %f44, %f296, %f297;
mov.b32 %r1876, %f288;
shfl.sync.bfly.b32 %r1877|%p90, %r1876, %r1861, %r1860, %r1862;
mov.b32 %f298, %r1877;
add.ftz.f32 %f299, %f288, %f298;
mov.b32 %r1878, %f299;
shfl.sync.bfly.b32 %r1879|%p91, %r1878, %r1865, %r1860, %r1862;
mov.b32 %f300, %r1879;
add.ftz.f32 %f301, %f299, %f300;
mov.b32 %r1880, %f301;
shfl.sync.bfly.b32 %r1881|%p92, %r1880, %r1868, %r1860, %r1862;
mov.b32 %f302, %r1881;
add.ftz.f32 %f303, %f301, %f302;
mov.b32 %r1882, %f303;
shfl.sync.bfly.b32 %r1883|%p93, %r1882, %r1871, %r1860, %r1862;
mov.b32 %f304, %r1883;
add.ftz.f32 %f305, %f303, %f304;
mov.b32 %r1884, %f305;
shfl.sync.bfly.b32 %r1885|%p94, %r1884, %r1874, %r1860, %r1862;
mov.b32 %f306, %r1885;
add.ftz.f32 %f45, %f305, %f306;
setp.ne.s32 %p95, %r2, 2;
or.b32 %r1889, %r4, 2;
setp.ge.u32 %p96, %r1889, %r54;
or.pred %p97, %p95, %p96;
@%p97 bra $L__BB306_52;

ld.param.f32 %f979, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f307, %f5, %f44;
add.ftz.f32 %f308, %f4, %f45;
min.ftz.f32 %f309, %f308, %f979;
min.ftz.f32 %f310, %f307, %f979;
max.ftz.f32 %f311, %f310, %f35;
mul.ftz.f32 %f312, %f309, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f313, %f312;
add.ftz.f32 %f314, %f313, 0f3F800000;
div.approx.ftz.f32 %f315, %f309, %f314;
mul.ftz.f32 %f316, %f311, %f315;
st.global.f32 [%rd47+8], %f316;

$L__BB306_52:
ld.shared.f32 %f317, [%r35+384];
add.ftz.f32 %f318, %f317, %f27;
ld.shared.f32 %f319, [%r37+384];
add.ftz.f32 %f320, %f319, %f31;
ld.shared.f32 %f321, [%r35+896];
add.ftz.f32 %f322, %f321, %f318;
ld.shared.f32 %f323, [%r37+896];
add.ftz.f32 %f324, %f323, %f320;
ld.shared.f32 %f325, [%r35+1408];
add.ftz.f32 %f326, %f325, %f322;
ld.shared.f32 %f327, [%r37+1408];
add.ftz.f32 %f328, %f327, %f324;
mov.b32 %r1890, %f326;
shfl.sync.bfly.b32 %r1894|%p98, %r1890, %r1861, %r1860, %r1862;
mov.b32 %f329, %r1894;
add.ftz.f32 %f330, %f326, %f329;
mov.b32 %r1895, %f330;
shfl.sync.bfly.b32 %r1897|%p99, %r1895, %r1865, %r1860, %r1862;
mov.b32 %f331, %r1897;
add.ftz.f32 %f332, %f330, %f331;
mov.b32 %r1898, %f332;
shfl.sync.bfly.b32 %r1900|%p100, %r1898, %r1868, %r1860, %r1862;
mov.b32 %f333, %r1900;
add.ftz.f32 %f334, %f332, %f333;
mov.b32 %r1901, %f334;
shfl.sync.bfly.b32 %r1903|%p101, %r1901, %r1871, %r1860, %r1862;
mov.b32 %f335, %r1903;
add.ftz.f32 %f336, %f334, %f335;
mov.b32 %r1904, %f336;
shfl.sync.bfly.b32 %r1906|%p102, %r1904, %r1874, %r1860, %r1862;
mov.b32 %f337, %r1906;
add.ftz.f32 %f46, %f336, %f337;
mov.b32 %r1907, %f328;
shfl.sync.bfly.b32 %r1908|%p103, %r1907, %r1861, %r1860, %r1862;
mov.b32 %f338, %r1908;
add.ftz.f32 %f339, %f328, %f338;
mov.b32 %r1909, %f339;
shfl.sync.bfly.b32 %r1910|%p104, %r1909, %r1865, %r1860, %r1862;
mov.b32 %f340, %r1910;
add.ftz.f32 %f341, %f339, %f340;
mov.b32 %r1911, %f341;
shfl.sync.bfly.b32 %r1912|%p105, %r1911, %r1868, %r1860, %r1862;
mov.b32 %f342, %r1912;
add.ftz.f32 %f343, %f341, %f342;
mov.b32 %r1913, %f343;
shfl.sync.bfly.b32 %r1914|%p106, %r1913, %r1871, %r1860, %r1862;
mov.b32 %f344, %r1914;
add.ftz.f32 %f345, %f343, %f344;
mov.b32 %r1915, %f345;
shfl.sync.bfly.b32 %r1916|%p107, %r1915, %r1874, %r1860, %r1862;
mov.b32 %f346, %r1916;
add.ftz.f32 %f47, %f345, %f346;
setp.ne.s32 %p108, %r2, 3;
or.b32 %r1920, %r4, 3;
setp.ge.u32 %p109, %r1920, %r54;
or.pred %p110, %p108, %p109;
@%p110 bra $L__BB306_78;

ld.param.f32 %f978, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
add.ftz.f32 %f347, %f5, %f46;
add.ftz.f32 %f348, %f4, %f47;
min.ftz.f32 %f349, %f348, %f978;
min.ftz.f32 %f350, %f347, %f978;
max.ftz.f32 %f351, %f350, %f35;
mul.ftz.f32 %f352, %f349, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f353, %f352;
add.ftz.f32 %f354, %f353, 0f3F800000;
div.approx.ftz.f32 %f355, %f349, %f354;
mul.ftz.f32 %f356, %f351, %f355;
st.global.f32 [%rd47+12], %f356;
bra.uni $L__BB306_78;

$L__BB306_70:
ld.shared.f32 %f845, [%r35];
add.ftz.f32 %f846, %f845, %f30;
ld.shared.f32 %f847, [%r37];
add.ftz.f32 %f848, %f847, %f34;
ld.shared.f32 %f849, [%r35+512];
add.ftz.f32 %f850, %f849, %f846;
ld.shared.f32 %f851, [%r37+512];
add.ftz.f32 %f852, %f851, %f848;
ld.shared.f32 %f853, [%r35+1024];
add.ftz.f32 %f854, %f853, %f850;
ld.shared.f32 %f855, [%r37+1024];
add.ftz.f32 %f856, %f855, %f852;
mov.b32 %r2290, %f854;
mov.u32 %r2291, 31;
mov.u32 %r2292, 16;
mov.u32 %r2293, -1;
shfl.sync.bfly.b32 %r2294|%p267, %r2290, %r2292, %r2291, %r2293;
mov.b32 %f857, %r2294;
add.ftz.f32 %f858, %f854, %f857;
mov.b32 %r2295, %f858;
mov.u32 %r2296, 8;
shfl.sync.bfly.b32 %r2297|%p268, %r2295, %r2296, %r2291, %r2293;
mov.b32 %f859, %r2297;
add.ftz.f32 %f860, %f858, %f859;
mov.b32 %r2298, %f860;
mov.u32 %r2299, 4;
shfl.sync.bfly.b32 %r2300|%p269, %r2298, %r2299, %r2291, %r2293;
mov.b32 %f861, %r2300;
add.ftz.f32 %f862, %f860, %f861;
mov.b32 %r2301, %f862;
mov.u32 %r2302, 2;
shfl.sync.bfly.b32 %r2303|%p270, %r2301, %r2302, %r2291, %r2293;
mov.b32 %f863, %r2303;
add.ftz.f32 %f864, %f862, %f863;
mov.b32 %r2304, %f864;
mov.u32 %r2305, 1;
shfl.sync.bfly.b32 %r2306|%p271, %r2304, %r2305, %r2291, %r2293;
mov.b32 %f865, %r2306;
add.ftz.f32 %f72, %f864, %f865;
mov.b32 %r2307, %f856;
shfl.sync.bfly.b32 %r2308|%p272, %r2307, %r2292, %r2291, %r2293;
mov.b32 %f866, %r2308;
add.ftz.f32 %f867, %f856, %f866;
mov.b32 %r2309, %f867;
shfl.sync.bfly.b32 %r2310|%p273, %r2309, %r2296, %r2291, %r2293;
mov.b32 %f868, %r2310;
add.ftz.f32 %f869, %f867, %f868;
mov.b32 %r2311, %f869;
shfl.sync.bfly.b32 %r2312|%p274, %r2311, %r2299, %r2291, %r2293;
mov.b32 %f870, %r2312;
add.ftz.f32 %f871, %f869, %f870;
mov.b32 %r2313, %f871;
shfl.sync.bfly.b32 %r2314|%p275, %r2313, %r2302, %r2291, %r2293;
mov.b32 %f872, %r2314;
add.ftz.f32 %f873, %f871, %f872;
mov.b32 %r2315, %f873;
shfl.sync.bfly.b32 %r2316|%p276, %r2315, %r2305, %r2291, %r2293;
mov.b32 %f874, %r2316;
add.ftz.f32 %f73, %f873, %f874;
setp.ne.s32 %p277, %r2, 0;
setp.ge.u32 %p278, %r4, %r54;
or.pred %p279, %p277, %p278;
@%p279 bra $L__BB306_72;

add.ftz.f32 %f875, %f5, %f72;
add.ftz.f32 %f876, %f4, %f73;
mul.ftz.f32 %f877, %f875, %f876;
st.global.f32 [%rd47], %f877;

$L__BB306_72:
ld.shared.f32 %f878, [%r35+128];
add.ftz.f32 %f879, %f878, %f29;
ld.shared.f32 %f880, [%r37+128];
add.ftz.f32 %f881, %f880, %f33;
ld.shared.f32 %f882, [%r35+640];
add.ftz.f32 %f883, %f882, %f879;
ld.shared.f32 %f884, [%r37+640];
add.ftz.f32 %f885, %f884, %f881;
ld.shared.f32 %f886, [%r35+1152];
add.ftz.f32 %f887, %f886, %f883;
ld.shared.f32 %f888, [%r37+1152];
add.ftz.f32 %f889, %f888, %f885;
mov.b32 %r2320, %f887;
shfl.sync.bfly.b32 %r2324|%p280, %r2320, %r2292, %r2291, %r2293;
mov.b32 %f890, %r2324;
add.ftz.f32 %f891, %f887, %f890;
mov.b32 %r2325, %f891;
shfl.sync.bfly.b32 %r2327|%p281, %r2325, %r2296, %r2291, %r2293;
mov.b32 %f892, %r2327;
add.ftz.f32 %f893, %f891, %f892;
mov.b32 %r2328, %f893;
shfl.sync.bfly.b32 %r2330|%p282, %r2328, %r2299, %r2291, %r2293;
mov.b32 %f894, %r2330;
add.ftz.f32 %f895, %f893, %f894;
mov.b32 %r2331, %f895;
shfl.sync.bfly.b32 %r2333|%p283, %r2331, %r2302, %r2291, %r2293;
mov.b32 %f896, %r2333;
add.ftz.f32 %f897, %f895, %f896;
mov.b32 %r2334, %f897;
shfl.sync.bfly.b32 %r2336|%p284, %r2334, %r2305, %r2291, %r2293;
mov.b32 %f898, %r2336;
add.ftz.f32 %f74, %f897, %f898;
mov.b32 %r2337, %f889;
shfl.sync.bfly.b32 %r2338|%p285, %r2337, %r2292, %r2291, %r2293;
mov.b32 %f899, %r2338;
add.ftz.f32 %f900, %f889, %f899;
mov.b32 %r2339, %f900;
shfl.sync.bfly.b32 %r2340|%p286, %r2339, %r2296, %r2291, %r2293;
mov.b32 %f901, %r2340;
add.ftz.f32 %f902, %f900, %f901;
mov.b32 %r2341, %f902;
shfl.sync.bfly.b32 %r2342|%p287, %r2341, %r2299, %r2291, %r2293;
mov.b32 %f903, %r2342;
add.ftz.f32 %f904, %f902, %f903;
mov.b32 %r2343, %f904;
shfl.sync.bfly.b32 %r2344|%p288, %r2343, %r2302, %r2291, %r2293;
mov.b32 %f905, %r2344;
add.ftz.f32 %f906, %f904, %f905;
mov.b32 %r2345, %f906;
shfl.sync.bfly.b32 %r2346|%p289, %r2345, %r2305, %r2291, %r2293;
mov.b32 %f907, %r2346;
add.ftz.f32 %f75, %f906, %f907;
setp.ne.s32 %p290, %r2, 1;
or.b32 %r2350, %r4, 1;
setp.ge.u32 %p291, %r2350, %r54;
or.pred %p292, %p290, %p291;
@%p292 bra $L__BB306_74;

add.ftz.f32 %f908, %f5, %f74;
add.ftz.f32 %f909, %f4, %f75;
mul.ftz.f32 %f910, %f908, %f909;
st.global.f32 [%rd47+4], %f910;

$L__BB306_74:
ld.shared.f32 %f911, [%r35+256];
add.ftz.f32 %f912, %f911, %f28;
ld.shared.f32 %f913, [%r37+256];
add.ftz.f32 %f914, %f913, %f32;
ld.shared.f32 %f915, [%r35+768];
add.ftz.f32 %f916, %f915, %f912;
ld.shared.f32 %f917, [%r37+768];
add.ftz.f32 %f918, %f917, %f914;
ld.shared.f32 %f919, [%r35+1280];
add.ftz.f32 %f920, %f919, %f916;
ld.shared.f32 %f921, [%r37+1280];
add.ftz.f32 %f922, %f921, %f918;
mov.b32 %r2351, %f920;
mov.u32 %r2352, 31;
mov.u32 %r2353, 16;
mov.u32 %r2354, -1;
shfl.sync.bfly.b32 %r2355|%p293, %r2351, %r2353, %r2352, %r2354;
mov.b32 %f923, %r2355;
add.ftz.f32 %f924, %f920, %f923;
mov.b32 %r2356, %f924;
mov.u32 %r2357, 8;
shfl.sync.bfly.b32 %r2358|%p294, %r2356, %r2357, %r2352, %r2354;
mov.b32 %f925, %r2358;
add.ftz.f32 %f926, %f924, %f925;
mov.b32 %r2359, %f926;
mov.u32 %r2360, 4;
shfl.sync.bfly.b32 %r2361|%p295, %r2359, %r2360, %r2352, %r2354;
mov.b32 %f927, %r2361;
add.ftz.f32 %f928, %f926, %f927;
mov.b32 %r2362, %f928;
mov.u32 %r2363, 2;
shfl.sync.bfly.b32 %r2364|%p296, %r2362, %r2363, %r2352, %r2354;
mov.b32 %f929, %r2364;
add.ftz.f32 %f930, %f928, %f929;
mov.b32 %r2365, %f930;
mov.u32 %r2366, 1;
shfl.sync.bfly.b32 %r2367|%p297, %r2365, %r2366, %r2352, %r2354;
mov.b32 %f931, %r2367;
add.ftz.f32 %f76, %f930, %f931;
mov.b32 %r2368, %f922;
shfl.sync.bfly.b32 %r2369|%p298, %r2368, %r2353, %r2352, %r2354;
mov.b32 %f932, %r2369;
add.ftz.f32 %f933, %f922, %f932;
mov.b32 %r2370, %f933;
shfl.sync.bfly.b32 %r2371|%p299, %r2370, %r2357, %r2352, %r2354;
mov.b32 %f934, %r2371;
add.ftz.f32 %f935, %f933, %f934;
mov.b32 %r2372, %f935;
shfl.sync.bfly.b32 %r2373|%p300, %r2372, %r2360, %r2352, %r2354;
mov.b32 %f936, %r2373;
add.ftz.f32 %f937, %f935, %f936;
mov.b32 %r2374, %f937;
shfl.sync.bfly.b32 %r2375|%p301, %r2374, %r2363, %r2352, %r2354;
mov.b32 %f938, %r2375;
add.ftz.f32 %f939, %f937, %f938;
mov.b32 %r2376, %f939;
shfl.sync.bfly.b32 %r2377|%p302, %r2376, %r2366, %r2352, %r2354;
mov.b32 %f940, %r2377;
add.ftz.f32 %f77, %f939, %f940;
setp.ne.s32 %p303, %r2, 2;
or.b32 %r2381, %r4, 2;
setp.ge.u32 %p304, %r2381, %r54;
or.pred %p305, %p303, %p304;
@%p305 bra $L__BB306_76;

add.ftz.f32 %f941, %f5, %f76;
add.ftz.f32 %f942, %f4, %f77;
mul.ftz.f32 %f943, %f941, %f942;
st.global.f32 [%rd47+8], %f943;

$L__BB306_76:
ld.shared.f32 %f944, [%r35+384];
add.ftz.f32 %f945, %f944, %f27;
ld.shared.f32 %f946, [%r37+384];
add.ftz.f32 %f947, %f946, %f31;
ld.shared.f32 %f948, [%r35+896];
add.ftz.f32 %f949, %f948, %f945;
ld.shared.f32 %f950, [%r37+896];
add.ftz.f32 %f951, %f950, %f947;
ld.shared.f32 %f952, [%r35+1408];
add.ftz.f32 %f953, %f952, %f949;
ld.shared.f32 %f954, [%r37+1408];
add.ftz.f32 %f955, %f954, %f951;
mov.b32 %r2382, %f953;
shfl.sync.bfly.b32 %r2386|%p306, %r2382, %r2353, %r2352, %r2354;
mov.b32 %f956, %r2386;
add.ftz.f32 %f957, %f953, %f956;
mov.b32 %r2387, %f957;
shfl.sync.bfly.b32 %r2389|%p307, %r2387, %r2357, %r2352, %r2354;
mov.b32 %f958, %r2389;
add.ftz.f32 %f959, %f957, %f958;
mov.b32 %r2390, %f959;
shfl.sync.bfly.b32 %r2392|%p308, %r2390, %r2360, %r2352, %r2354;
mov.b32 %f960, %r2392;
add.ftz.f32 %f961, %f959, %f960;
mov.b32 %r2393, %f961;
shfl.sync.bfly.b32 %r2395|%p309, %r2393, %r2363, %r2352, %r2354;
mov.b32 %f962, %r2395;
add.ftz.f32 %f963, %f961, %f962;
mov.b32 %r2396, %f963;
shfl.sync.bfly.b32 %r2398|%p310, %r2396, %r2366, %r2352, %r2354;
mov.b32 %f964, %r2398;
add.ftz.f32 %f78, %f963, %f964;
mov.b32 %r2399, %f955;
shfl.sync.bfly.b32 %r2400|%p311, %r2399, %r2353, %r2352, %r2354;
mov.b32 %f965, %r2400;
add.ftz.f32 %f966, %f955, %f965;
mov.b32 %r2401, %f966;
shfl.sync.bfly.b32 %r2402|%p312, %r2401, %r2357, %r2352, %r2354;
mov.b32 %f967, %r2402;
add.ftz.f32 %f968, %f966, %f967;
mov.b32 %r2403, %f968;
shfl.sync.bfly.b32 %r2404|%p313, %r2403, %r2360, %r2352, %r2354;
mov.b32 %f969, %r2404;
add.ftz.f32 %f970, %f968, %f969;
mov.b32 %r2405, %f970;
shfl.sync.bfly.b32 %r2406|%p314, %r2405, %r2363, %r2352, %r2354;
mov.b32 %f971, %r2406;
add.ftz.f32 %f972, %f970, %f971;
mov.b32 %r2407, %f972;
shfl.sync.bfly.b32 %r2408|%p315, %r2407, %r2366, %r2352, %r2354;
mov.b32 %f973, %r2408;
add.ftz.f32 %f79, %f972, %f973;
setp.ne.s32 %p316, %r2, 3;
or.b32 %r2412, %r4, 3;
setp.ge.u32 %p317, %r2412, %r54;
or.pred %p318, %p316, %p317;
@%p318 bra $L__BB306_78;

add.ftz.f32 %f974, %f5, %f78;
add.ftz.f32 %f975, %f4, %f79;
mul.ftz.f32 %f976, %f974, %f975;
st.global.f32 [%rd47+12], %f976;
bra.uni $L__BB306_78;

$L__BB306_62:
ld.shared.f32 %f681, [%r35];
add.ftz.f32 %f682, %f681, %f30;
ld.shared.f32 %f683, [%r37];
add.ftz.f32 %f684, %f683, %f34;
ld.shared.f32 %f685, [%r35+512];
add.ftz.f32 %f686, %f685, %f682;
ld.shared.f32 %f687, [%r37+512];
add.ftz.f32 %f688, %f687, %f684;
ld.shared.f32 %f689, [%r35+1024];
add.ftz.f32 %f690, %f689, %f686;
ld.shared.f32 %f691, [%r37+1024];
add.ftz.f32 %f692, %f691, %f688;
mov.b32 %r2167, %f690;
mov.u32 %r2168, 31;
mov.u32 %r2169, 16;
mov.u32 %r2170, -1;
shfl.sync.bfly.b32 %r2171|%p215, %r2167, %r2169, %r2168, %r2170;
mov.b32 %f693, %r2171;
add.ftz.f32 %f694, %f690, %f693;
mov.b32 %r2172, %f694;
mov.u32 %r2173, 8;
shfl.sync.bfly.b32 %r2174|%p216, %r2172, %r2173, %r2168, %r2170;
mov.b32 %f695, %r2174;
add.ftz.f32 %f696, %f694, %f695;
mov.b32 %r2175, %f696;
mov.u32 %r2176, 4;
shfl.sync.bfly.b32 %r2177|%p217, %r2175, %r2176, %r2168, %r2170;
mov.b32 %f697, %r2177;
add.ftz.f32 %f698, %f696, %f697;
mov.b32 %r2178, %f698;
mov.u32 %r2179, 2;
shfl.sync.bfly.b32 %r2180|%p218, %r2178, %r2179, %r2168, %r2170;
mov.b32 %f699, %r2180;
add.ftz.f32 %f700, %f698, %f699;
mov.b32 %r2181, %f700;
mov.u32 %r2182, 1;
shfl.sync.bfly.b32 %r2183|%p219, %r2181, %r2182, %r2168, %r2170;
mov.b32 %f701, %r2183;
add.ftz.f32 %f64, %f700, %f701;
mov.b32 %r2184, %f692;
shfl.sync.bfly.b32 %r2185|%p220, %r2184, %r2169, %r2168, %r2170;
mov.b32 %f702, %r2185;
add.ftz.f32 %f703, %f692, %f702;
mov.b32 %r2186, %f703;
shfl.sync.bfly.b32 %r2187|%p221, %r2186, %r2173, %r2168, %r2170;
mov.b32 %f704, %r2187;
add.ftz.f32 %f705, %f703, %f704;
mov.b32 %r2188, %f705;
shfl.sync.bfly.b32 %r2189|%p222, %r2188, %r2176, %r2168, %r2170;
mov.b32 %f706, %r2189;
add.ftz.f32 %f707, %f705, %f706;
mov.b32 %r2190, %f707;
shfl.sync.bfly.b32 %r2191|%p223, %r2190, %r2179, %r2168, %r2170;
mov.b32 %f708, %r2191;
add.ftz.f32 %f709, %f707, %f708;
mov.b32 %r2192, %f709;
shfl.sync.bfly.b32 %r2193|%p224, %r2192, %r2182, %r2168, %r2170;
mov.b32 %f710, %r2193;
add.ftz.f32 %f65, %f709, %f710;
setp.ne.s32 %p225, %r2, 0;
setp.ge.u32 %p226, %r4, %r54;
or.pred %p227, %p225, %p226;
@%p227 bra $L__BB306_64;

add.ftz.f32 %f711, %f5, %f64;
add.ftz.f32 %f712, %f4, %f65;
mul.ftz.f32 %f713, %f712, 0f3F000000;
mul.ftz.f32 %f714, %f712, 0f3F4C422A;
mul.ftz.f32 %f715, %f712, 0f3D372713;
fma.rn.ftz.f32 %f716, %f712, %f715, 0f3F800000;
mul.ftz.f32 %f717, %f714, %f716;
tanh.approx.f32 %f718, %f717;
add.ftz.f32 %f719, %f718, 0f3F800000;
mul.ftz.f32 %f720, %f713, %f719;
mul.ftz.f32 %f721, %f711, %f720;
st.global.f32 [%rd47], %f721;

$L__BB306_64:
ld.shared.f32 %f722, [%r35+128];
add.ftz.f32 %f723, %f722, %f29;
ld.shared.f32 %f724, [%r37+128];
add.ftz.f32 %f725, %f724, %f33;
ld.shared.f32 %f726, [%r35+640];
add.ftz.f32 %f727, %f726, %f723;
ld.shared.f32 %f728, [%r37+640];
add.ftz.f32 %f729, %f728, %f725;
ld.shared.f32 %f730, [%r35+1152];
add.ftz.f32 %f731, %f730, %f727;
ld.shared.f32 %f732, [%r37+1152];
add.ftz.f32 %f733, %f732, %f729;
mov.b32 %r2197, %f731;
shfl.sync.bfly.b32 %r2201|%p228, %r2197, %r2169, %r2168, %r2170;
mov.b32 %f734, %r2201;
add.ftz.f32 %f735, %f731, %f734;
mov.b32 %r2202, %f735;
shfl.sync.bfly.b32 %r2204|%p229, %r2202, %r2173, %r2168, %r2170;
mov.b32 %f736, %r2204;
add.ftz.f32 %f737, %f735, %f736;
mov.b32 %r2205, %f737;
shfl.sync.bfly.b32 %r2207|%p230, %r2205, %r2176, %r2168, %r2170;
mov.b32 %f738, %r2207;
add.ftz.f32 %f739, %f737, %f738;
mov.b32 %r2208, %f739;
shfl.sync.bfly.b32 %r2210|%p231, %r2208, %r2179, %r2168, %r2170;
mov.b32 %f740, %r2210;
add.ftz.f32 %f741, %f739, %f740;
mov.b32 %r2211, %f741;
shfl.sync.bfly.b32 %r2213|%p232, %r2211, %r2182, %r2168, %r2170;
mov.b32 %f742, %r2213;
add.ftz.f32 %f66, %f741, %f742;
mov.b32 %r2214, %f733;
shfl.sync.bfly.b32 %r2215|%p233, %r2214, %r2169, %r2168, %r2170;
mov.b32 %f743, %r2215;
add.ftz.f32 %f744, %f733, %f743;
mov.b32 %r2216, %f744;
shfl.sync.bfly.b32 %r2217|%p234, %r2216, %r2173, %r2168, %r2170;
mov.b32 %f745, %r2217;
add.ftz.f32 %f746, %f744, %f745;
mov.b32 %r2218, %f746;
shfl.sync.bfly.b32 %r2219|%p235, %r2218, %r2176, %r2168, %r2170;
mov.b32 %f747, %r2219;
add.ftz.f32 %f748, %f746, %f747;
mov.b32 %r2220, %f748;
shfl.sync.bfly.b32 %r2221|%p236, %r2220, %r2179, %r2168, %r2170;
mov.b32 %f749, %r2221;
add.ftz.f32 %f750, %f748, %f749;
mov.b32 %r2222, %f750;
shfl.sync.bfly.b32 %r2223|%p237, %r2222, %r2182, %r2168, %r2170;
mov.b32 %f751, %r2223;
add.ftz.f32 %f67, %f750, %f751;
setp.ne.s32 %p238, %r2, 1;
or.b32 %r2227, %r4, 1;
setp.ge.u32 %p239, %r2227, %r54;
or.pred %p240, %p238, %p239;
@%p240 bra $L__BB306_66;

add.ftz.f32 %f752, %f5, %f66;
add.ftz.f32 %f753, %f4, %f67;
mul.ftz.f32 %f754, %f753, 0f3F000000;
mul.ftz.f32 %f755, %f753, 0f3F4C422A;
mul.ftz.f32 %f756, %f753, 0f3D372713;
fma.rn.ftz.f32 %f757, %f753, %f756, 0f3F800000;
mul.ftz.f32 %f758, %f755, %f757;
tanh.approx.f32 %f759, %f758;
add.ftz.f32 %f760, %f759, 0f3F800000;
mul.ftz.f32 %f761, %f754, %f760;
mul.ftz.f32 %f762, %f752, %f761;
st.global.f32 [%rd47+4], %f762;

$L__BB306_66:
ld.shared.f32 %f763, [%r35+256];
add.ftz.f32 %f764, %f763, %f28;
ld.shared.f32 %f765, [%r37+256];
add.ftz.f32 %f766, %f765, %f32;
ld.shared.f32 %f767, [%r35+768];
add.ftz.f32 %f768, %f767, %f764;
ld.shared.f32 %f769, [%r37+768];
add.ftz.f32 %f770, %f769, %f766;
ld.shared.f32 %f771, [%r35+1280];
add.ftz.f32 %f772, %f771, %f768;
ld.shared.f32 %f773, [%r37+1280];
add.ftz.f32 %f774, %f773, %f770;
mov.b32 %r2228, %f772;
mov.u32 %r2229, 31;
mov.u32 %r2230, 16;
mov.u32 %r2231, -1;
shfl.sync.bfly.b32 %r2232|%p241, %r2228, %r2230, %r2229, %r2231;
mov.b32 %f775, %r2232;
add.ftz.f32 %f776, %f772, %f775;
mov.b32 %r2233, %f776;
mov.u32 %r2234, 8;
shfl.sync.bfly.b32 %r2235|%p242, %r2233, %r2234, %r2229, %r2231;
mov.b32 %f777, %r2235;
add.ftz.f32 %f778, %f776, %f777;
mov.b32 %r2236, %f778;
mov.u32 %r2237, 4;
shfl.sync.bfly.b32 %r2238|%p243, %r2236, %r2237, %r2229, %r2231;
mov.b32 %f779, %r2238;
add.ftz.f32 %f780, %f778, %f779;
mov.b32 %r2239, %f780;
mov.u32 %r2240, 2;
shfl.sync.bfly.b32 %r2241|%p244, %r2239, %r2240, %r2229, %r2231;
mov.b32 %f781, %r2241;
add.ftz.f32 %f782, %f780, %f781;
mov.b32 %r2242, %f782;
mov.u32 %r2243, 1;
shfl.sync.bfly.b32 %r2244|%p245, %r2242, %r2243, %r2229, %r2231;
mov.b32 %f783, %r2244;
add.ftz.f32 %f68, %f782, %f783;
mov.b32 %r2245, %f774;
shfl.sync.bfly.b32 %r2246|%p246, %r2245, %r2230, %r2229, %r2231;
mov.b32 %f784, %r2246;
add.ftz.f32 %f785, %f774, %f784;
mov.b32 %r2247, %f785;
shfl.sync.bfly.b32 %r2248|%p247, %r2247, %r2234, %r2229, %r2231;
mov.b32 %f786, %r2248;
add.ftz.f32 %f787, %f785, %f786;
mov.b32 %r2249, %f787;
shfl.sync.bfly.b32 %r2250|%p248, %r2249, %r2237, %r2229, %r2231;
mov.b32 %f788, %r2250;
add.ftz.f32 %f789, %f787, %f788;
mov.b32 %r2251, %f789;
shfl.sync.bfly.b32 %r2252|%p249, %r2251, %r2240, %r2229, %r2231;
mov.b32 %f790, %r2252;
add.ftz.f32 %f791, %f789, %f790;
mov.b32 %r2253, %f791;
shfl.sync.bfly.b32 %r2254|%p250, %r2253, %r2243, %r2229, %r2231;
mov.b32 %f792, %r2254;
add.ftz.f32 %f69, %f791, %f792;
setp.ne.s32 %p251, %r2, 2;
or.b32 %r2258, %r4, 2;
setp.ge.u32 %p252, %r2258, %r54;
or.pred %p253, %p251, %p252;
@%p253 bra $L__BB306_68;

add.ftz.f32 %f793, %f5, %f68;
add.ftz.f32 %f794, %f4, %f69;
mul.ftz.f32 %f795, %f794, 0f3F000000;
mul.ftz.f32 %f796, %f794, 0f3F4C422A;
mul.ftz.f32 %f797, %f794, 0f3D372713;
fma.rn.ftz.f32 %f798, %f794, %f797, 0f3F800000;
mul.ftz.f32 %f799, %f796, %f798;
tanh.approx.f32 %f800, %f799;
add.ftz.f32 %f801, %f800, 0f3F800000;
mul.ftz.f32 %f802, %f795, %f801;
mul.ftz.f32 %f803, %f793, %f802;
st.global.f32 [%rd47+8], %f803;

$L__BB306_68:
ld.shared.f32 %f804, [%r35+384];
add.ftz.f32 %f805, %f804, %f27;
ld.shared.f32 %f806, [%r37+384];
add.ftz.f32 %f807, %f806, %f31;
ld.shared.f32 %f808, [%r35+896];
add.ftz.f32 %f809, %f808, %f805;
ld.shared.f32 %f810, [%r37+896];
add.ftz.f32 %f811, %f810, %f807;
ld.shared.f32 %f812, [%r35+1408];
add.ftz.f32 %f813, %f812, %f809;
ld.shared.f32 %f814, [%r37+1408];
add.ftz.f32 %f815, %f814, %f811;
mov.b32 %r2259, %f813;
shfl.sync.bfly.b32 %r2263|%p254, %r2259, %r2230, %r2229, %r2231;
mov.b32 %f816, %r2263;
add.ftz.f32 %f817, %f813, %f816;
mov.b32 %r2264, %f817;
shfl.sync.bfly.b32 %r2266|%p255, %r2264, %r2234, %r2229, %r2231;
mov.b32 %f818, %r2266;
add.ftz.f32 %f819, %f817, %f818;
mov.b32 %r2267, %f819;
shfl.sync.bfly.b32 %r2269|%p256, %r2267, %r2237, %r2229, %r2231;
mov.b32 %f820, %r2269;
add.ftz.f32 %f821, %f819, %f820;
mov.b32 %r2270, %f821;
shfl.sync.bfly.b32 %r2272|%p257, %r2270, %r2240, %r2229, %r2231;
mov.b32 %f822, %r2272;
add.ftz.f32 %f823, %f821, %f822;
mov.b32 %r2273, %f823;
shfl.sync.bfly.b32 %r2275|%p258, %r2273, %r2243, %r2229, %r2231;
mov.b32 %f824, %r2275;
add.ftz.f32 %f70, %f823, %f824;
mov.b32 %r2276, %f815;
shfl.sync.bfly.b32 %r2277|%p259, %r2276, %r2230, %r2229, %r2231;
mov.b32 %f825, %r2277;
add.ftz.f32 %f826, %f815, %f825;
mov.b32 %r2278, %f826;
shfl.sync.bfly.b32 %r2279|%p260, %r2278, %r2234, %r2229, %r2231;
mov.b32 %f827, %r2279;
add.ftz.f32 %f828, %f826, %f827;
mov.b32 %r2280, %f828;
shfl.sync.bfly.b32 %r2281|%p261, %r2280, %r2237, %r2229, %r2231;
mov.b32 %f829, %r2281;
add.ftz.f32 %f830, %f828, %f829;
mov.b32 %r2282, %f830;
shfl.sync.bfly.b32 %r2283|%p262, %r2282, %r2240, %r2229, %r2231;
mov.b32 %f831, %r2283;
add.ftz.f32 %f832, %f830, %f831;
mov.b32 %r2284, %f832;
shfl.sync.bfly.b32 %r2285|%p263, %r2284, %r2243, %r2229, %r2231;
mov.b32 %f833, %r2285;
add.ftz.f32 %f71, %f832, %f833;
setp.ne.s32 %p264, %r2, 3;
or.b32 %r2289, %r4, 3;
setp.ge.u32 %p265, %r2289, %r54;
or.pred %p266, %p264, %p265;
@%p266 bra $L__BB306_78;

add.ftz.f32 %f834, %f5, %f70;
add.ftz.f32 %f835, %f4, %f71;
mul.ftz.f32 %f836, %f835, 0f3F000000;
mul.ftz.f32 %f837, %f835, 0f3F4C422A;
mul.ftz.f32 %f838, %f835, 0f3D372713;
fma.rn.ftz.f32 %f839, %f835, %f838, 0f3F800000;
mul.ftz.f32 %f840, %f837, %f839;
tanh.approx.f32 %f841, %f840;
add.ftz.f32 %f842, %f841, 0f3F800000;
mul.ftz.f32 %f843, %f836, %f842;
mul.ftz.f32 %f844, %f834, %f843;
st.global.f32 [%rd47+12], %f844;
bra.uni $L__BB306_78;

$L__BB306_54:
ld.shared.f32 %f357, [%r35];
add.ftz.f32 %f358, %f357, %f30;
ld.shared.f32 %f359, [%r37];
add.ftz.f32 %f360, %f359, %f34;
ld.shared.f32 %f361, [%r35+512];
add.ftz.f32 %f362, %f361, %f358;
ld.shared.f32 %f363, [%r37+512];
add.ftz.f32 %f364, %f363, %f360;
ld.shared.f32 %f365, [%r35+1024];
add.ftz.f32 %f366, %f365, %f362;
ld.shared.f32 %f367, [%r37+1024];
add.ftz.f32 %f368, %f367, %f364;
mov.b32 %r1921, %f366;
mov.u32 %r1922, 31;
mov.u32 %r1923, 16;
mov.u32 %r1924, -1;
shfl.sync.bfly.b32 %r1925|%p111, %r1921, %r1923, %r1922, %r1924;
mov.b32 %f369, %r1925;
add.ftz.f32 %f370, %f366, %f369;
mov.b32 %r1926, %f370;
mov.u32 %r1927, 8;
shfl.sync.bfly.b32 %r1928|%p112, %r1926, %r1927, %r1922, %r1924;
mov.b32 %f371, %r1928;
add.ftz.f32 %f372, %f370, %f371;
mov.b32 %r1929, %f372;
mov.u32 %r1930, 4;
shfl.sync.bfly.b32 %r1931|%p113, %r1929, %r1930, %r1922, %r1924;
mov.b32 %f373, %r1931;
add.ftz.f32 %f374, %f372, %f373;
mov.b32 %r1932, %f374;
mov.u32 %r1933, 2;
shfl.sync.bfly.b32 %r1934|%p114, %r1932, %r1933, %r1922, %r1924;
mov.b32 %f375, %r1934;
add.ftz.f32 %f376, %f374, %f375;
mov.b32 %r1935, %f376;
mov.u32 %r1936, 1;
shfl.sync.bfly.b32 %r1937|%p115, %r1935, %r1936, %r1922, %r1924;
mov.b32 %f377, %r1937;
add.ftz.f32 %f48, %f376, %f377;
mov.b32 %r1938, %f368;
shfl.sync.bfly.b32 %r1939|%p116, %r1938, %r1923, %r1922, %r1924;
mov.b32 %f378, %r1939;
add.ftz.f32 %f379, %f368, %f378;
mov.b32 %r1940, %f379;
shfl.sync.bfly.b32 %r1941|%p117, %r1940, %r1927, %r1922, %r1924;
mov.b32 %f380, %r1941;
add.ftz.f32 %f381, %f379, %f380;
mov.b32 %r1942, %f381;
shfl.sync.bfly.b32 %r1943|%p118, %r1942, %r1930, %r1922, %r1924;
mov.b32 %f382, %r1943;
add.ftz.f32 %f383, %f381, %f382;
mov.b32 %r1944, %f383;
shfl.sync.bfly.b32 %r1945|%p119, %r1944, %r1933, %r1922, %r1924;
mov.b32 %f384, %r1945;
add.ftz.f32 %f385, %f383, %f384;
mov.b32 %r1946, %f385;
shfl.sync.bfly.b32 %r1947|%p120, %r1946, %r1936, %r1922, %r1924;
mov.b32 %f386, %r1947;
add.ftz.f32 %f49, %f385, %f386;
setp.ne.s32 %p121, %r2, 0;
setp.ge.u32 %p122, %r4, %r54;
or.pred %p123, %p121, %p122;
@%p123 bra $L__BB306_56;

add.ftz.f32 %f387, %f5, %f48;
add.ftz.f32 %f388, %f4, %f49;
mov.f32 %f389, 0f40E00000;
min.ftz.f32 %f390, %f388, %f389;
min.ftz.f32 %f391, %f387, %f389;
mov.f32 %f392, 0fC0E00000;
max.ftz.f32 %f393, %f391, %f392;
mul.ftz.f32 %f394, %f390, 0f3FD9DB23;
mul.ftz.f32 %f395, %f394, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f396, %f395;
add.ftz.f32 %f397, %f396, 0f3F800000;
div.approx.ftz.f32 %f398, %f390, %f397;
add.ftz.f32 %f399, %f393, 0f3F800000;
mul.ftz.f32 %f400, %f399, %f398;
st.global.f32 [%rd47], %f400;

$L__BB306_56:
ld.shared.f32 %f401, [%r35+128];
add.ftz.f32 %f402, %f401, %f29;
ld.shared.f32 %f403, [%r37+128];
add.ftz.f32 %f404, %f403, %f33;
ld.shared.f32 %f405, [%r35+640];
add.ftz.f32 %f406, %f405, %f402;
ld.shared.f32 %f407, [%r37+640];
add.ftz.f32 %f408, %f407, %f404;
ld.shared.f32 %f409, [%r35+1152];
add.ftz.f32 %f410, %f409, %f406;
ld.shared.f32 %f411, [%r37+1152];
add.ftz.f32 %f412, %f411, %f408;
mov.b32 %r1951, %f410;
shfl.sync.bfly.b32 %r1955|%p124, %r1951, %r1923, %r1922, %r1924;
mov.b32 %f413, %r1955;
add.ftz.f32 %f414, %f410, %f413;
mov.b32 %r1956, %f414;
shfl.sync.bfly.b32 %r1958|%p125, %r1956, %r1927, %r1922, %r1924;
mov.b32 %f415, %r1958;
add.ftz.f32 %f416, %f414, %f415;
mov.b32 %r1959, %f416;
shfl.sync.bfly.b32 %r1961|%p126, %r1959, %r1930, %r1922, %r1924;
mov.b32 %f417, %r1961;
add.ftz.f32 %f418, %f416, %f417;
mov.b32 %r1962, %f418;
shfl.sync.bfly.b32 %r1964|%p127, %r1962, %r1933, %r1922, %r1924;
mov.b32 %f419, %r1964;
add.ftz.f32 %f420, %f418, %f419;
mov.b32 %r1965, %f420;
shfl.sync.bfly.b32 %r1967|%p128, %r1965, %r1936, %r1922, %r1924;
mov.b32 %f421, %r1967;
add.ftz.f32 %f50, %f420, %f421;
mov.b32 %r1968, %f412;
shfl.sync.bfly.b32 %r1969|%p129, %r1968, %r1923, %r1922, %r1924;
mov.b32 %f422, %r1969;
add.ftz.f32 %f423, %f412, %f422;
mov.b32 %r1970, %f423;
shfl.sync.bfly.b32 %r1971|%p130, %r1970, %r1927, %r1922, %r1924;
mov.b32 %f424, %r1971;
add.ftz.f32 %f425, %f423, %f424;
mov.b32 %r1972, %f425;
shfl.sync.bfly.b32 %r1973|%p131, %r1972, %r1930, %r1922, %r1924;
mov.b32 %f426, %r1973;
add.ftz.f32 %f427, %f425, %f426;
mov.b32 %r1974, %f427;
shfl.sync.bfly.b32 %r1975|%p132, %r1974, %r1933, %r1922, %r1924;
mov.b32 %f428, %r1975;
add.ftz.f32 %f429, %f427, %f428;
mov.b32 %r1976, %f429;
shfl.sync.bfly.b32 %r1977|%p133, %r1976, %r1936, %r1922, %r1924;
mov.b32 %f430, %r1977;
add.ftz.f32 %f51, %f429, %f430;
setp.ne.s32 %p134, %r2, 1;
or.b32 %r1981, %r4, 1;
setp.ge.u32 %p135, %r1981, %r54;
or.pred %p136, %p134, %p135;
@%p136 bra $L__BB306_58;

add.ftz.f32 %f431, %f5, %f50;
add.ftz.f32 %f432, %f4, %f51;
mov.f32 %f433, 0f40E00000;
min.ftz.f32 %f434, %f432, %f433;
min.ftz.f32 %f435, %f431, %f433;
mov.f32 %f436, 0fC0E00000;
max.ftz.f32 %f437, %f435, %f436;
mul.ftz.f32 %f438, %f434, 0f3FD9DB23;
mul.ftz.f32 %f439, %f438, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f440, %f439;
add.ftz.f32 %f441, %f440, 0f3F800000;
div.approx.ftz.f32 %f442, %f434, %f441;
add.ftz.f32 %f443, %f437, 0f3F800000;
mul.ftz.f32 %f444, %f443, %f442;
st.global.f32 [%rd47+4], %f444;

$L__BB306_58:
ld.shared.f32 %f445, [%r35+256];
add.ftz.f32 %f446, %f445, %f28;
ld.shared.f32 %f447, [%r37+256];
add.ftz.f32 %f448, %f447, %f32;
ld.shared.f32 %f449, [%r35+768];
add.ftz.f32 %f450, %f449, %f446;
ld.shared.f32 %f451, [%r37+768];
add.ftz.f32 %f452, %f451, %f448;
ld.shared.f32 %f453, [%r35+1280];
add.ftz.f32 %f454, %f453, %f450;
ld.shared.f32 %f455, [%r37+1280];
add.ftz.f32 %f456, %f455, %f452;
mov.b32 %r1982, %f454;
mov.u32 %r1983, 31;
mov.u32 %r1984, 16;
mov.u32 %r1985, -1;
shfl.sync.bfly.b32 %r1986|%p137, %r1982, %r1984, %r1983, %r1985;
mov.b32 %f457, %r1986;
add.ftz.f32 %f458, %f454, %f457;
mov.b32 %r1987, %f458;
mov.u32 %r1988, 8;
shfl.sync.bfly.b32 %r1989|%p138, %r1987, %r1988, %r1983, %r1985;
mov.b32 %f459, %r1989;
add.ftz.f32 %f460, %f458, %f459;
mov.b32 %r1990, %f460;
mov.u32 %r1991, 4;
shfl.sync.bfly.b32 %r1992|%p139, %r1990, %r1991, %r1983, %r1985;
mov.b32 %f461, %r1992;
add.ftz.f32 %f462, %f460, %f461;
mov.b32 %r1993, %f462;
mov.u32 %r1994, 2;
shfl.sync.bfly.b32 %r1995|%p140, %r1993, %r1994, %r1983, %r1985;
mov.b32 %f463, %r1995;
add.ftz.f32 %f464, %f462, %f463;
mov.b32 %r1996, %f464;
mov.u32 %r1997, 1;
shfl.sync.bfly.b32 %r1998|%p141, %r1996, %r1997, %r1983, %r1985;
mov.b32 %f465, %r1998;
add.ftz.f32 %f52, %f464, %f465;
mov.b32 %r1999, %f456;
shfl.sync.bfly.b32 %r2000|%p142, %r1999, %r1984, %r1983, %r1985;
mov.b32 %f466, %r2000;
add.ftz.f32 %f467, %f456, %f466;
mov.b32 %r2001, %f467;
shfl.sync.bfly.b32 %r2002|%p143, %r2001, %r1988, %r1983, %r1985;
mov.b32 %f468, %r2002;
add.ftz.f32 %f469, %f467, %f468;
mov.b32 %r2003, %f469;
shfl.sync.bfly.b32 %r2004|%p144, %r2003, %r1991, %r1983, %r1985;
mov.b32 %f470, %r2004;
add.ftz.f32 %f471, %f469, %f470;
mov.b32 %r2005, %f471;
shfl.sync.bfly.b32 %r2006|%p145, %r2005, %r1994, %r1983, %r1985;
mov.b32 %f472, %r2006;
add.ftz.f32 %f473, %f471, %f472;
mov.b32 %r2007, %f473;
shfl.sync.bfly.b32 %r2008|%p146, %r2007, %r1997, %r1983, %r1985;
mov.b32 %f474, %r2008;
add.ftz.f32 %f53, %f473, %f474;
setp.ne.s32 %p147, %r2, 2;
or.b32 %r2012, %r4, 2;
setp.ge.u32 %p148, %r2012, %r54;
or.pred %p149, %p147, %p148;
@%p149 bra $L__BB306_60;

add.ftz.f32 %f475, %f5, %f52;
add.ftz.f32 %f476, %f4, %f53;
mov.f32 %f477, 0f40E00000;
min.ftz.f32 %f478, %f476, %f477;
min.ftz.f32 %f479, %f475, %f477;
mov.f32 %f480, 0fC0E00000;
max.ftz.f32 %f481, %f479, %f480;
mul.ftz.f32 %f482, %f478, 0f3FD9DB23;
mul.ftz.f32 %f483, %f482, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f484, %f483;
add.ftz.f32 %f485, %f484, 0f3F800000;
div.approx.ftz.f32 %f486, %f478, %f485;
add.ftz.f32 %f487, %f481, 0f3F800000;
mul.ftz.f32 %f488, %f487, %f486;
st.global.f32 [%rd47+8], %f488;

$L__BB306_60:
ld.shared.f32 %f489, [%r35+384];
add.ftz.f32 %f490, %f489, %f27;
ld.shared.f32 %f491, [%r37+384];
add.ftz.f32 %f492, %f491, %f31;
ld.shared.f32 %f493, [%r35+896];
add.ftz.f32 %f494, %f493, %f490;
ld.shared.f32 %f495, [%r37+896];
add.ftz.f32 %f496, %f495, %f492;
ld.shared.f32 %f497, [%r35+1408];
add.ftz.f32 %f498, %f497, %f494;
ld.shared.f32 %f499, [%r37+1408];
add.ftz.f32 %f500, %f499, %f496;
mov.b32 %r2013, %f498;
shfl.sync.bfly.b32 %r2017|%p150, %r2013, %r1984, %r1983, %r1985;
mov.b32 %f501, %r2017;
add.ftz.f32 %f502, %f498, %f501;
mov.b32 %r2018, %f502;
shfl.sync.bfly.b32 %r2020|%p151, %r2018, %r1988, %r1983, %r1985;
mov.b32 %f503, %r2020;
add.ftz.f32 %f504, %f502, %f503;
mov.b32 %r2021, %f504;
shfl.sync.bfly.b32 %r2023|%p152, %r2021, %r1991, %r1983, %r1985;
mov.b32 %f505, %r2023;
add.ftz.f32 %f506, %f504, %f505;
mov.b32 %r2024, %f506;
shfl.sync.bfly.b32 %r2026|%p153, %r2024, %r1994, %r1983, %r1985;
mov.b32 %f507, %r2026;
add.ftz.f32 %f508, %f506, %f507;
mov.b32 %r2027, %f508;
shfl.sync.bfly.b32 %r2029|%p154, %r2027, %r1997, %r1983, %r1985;
mov.b32 %f509, %r2029;
add.ftz.f32 %f54, %f508, %f509;
mov.b32 %r2030, %f500;
shfl.sync.bfly.b32 %r2031|%p155, %r2030, %r1984, %r1983, %r1985;
mov.b32 %f510, %r2031;
add.ftz.f32 %f511, %f500, %f510;
mov.b32 %r2032, %f511;
shfl.sync.bfly.b32 %r2033|%p156, %r2032, %r1988, %r1983, %r1985;
mov.b32 %f512, %r2033;
add.ftz.f32 %f513, %f511, %f512;
mov.b32 %r2034, %f513;
shfl.sync.bfly.b32 %r2035|%p157, %r2034, %r1991, %r1983, %r1985;
mov.b32 %f514, %r2035;
add.ftz.f32 %f515, %f513, %f514;
mov.b32 %r2036, %f515;
shfl.sync.bfly.b32 %r2037|%p158, %r2036, %r1994, %r1983, %r1985;
mov.b32 %f516, %r2037;
add.ftz.f32 %f517, %f515, %f516;
mov.b32 %r2038, %f517;
shfl.sync.bfly.b32 %r2039|%p159, %r2038, %r1997, %r1983, %r1985;
mov.b32 %f518, %r2039;
add.ftz.f32 %f55, %f517, %f518;
setp.ne.s32 %p160, %r2, 3;
or.b32 %r2043, %r4, 3;
setp.ge.u32 %p161, %r2043, %r54;
or.pred %p162, %p160, %p161;
@%p162 bra $L__BB306_78;

add.ftz.f32 %f519, %f5, %f54;
add.ftz.f32 %f520, %f4, %f55;
mov.f32 %f521, 0f40E00000;
min.ftz.f32 %f522, %f520, %f521;
min.ftz.f32 %f523, %f519, %f521;
mov.f32 %f524, 0fC0E00000;
max.ftz.f32 %f525, %f523, %f524;
mul.ftz.f32 %f526, %f522, 0f3FD9DB23;
mul.ftz.f32 %f527, %f526, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f528, %f527;
add.ftz.f32 %f529, %f528, 0f3F800000;
div.approx.ftz.f32 %f530, %f522, %f529;
add.ftz.f32 %f531, %f525, 0f3F800000;
mul.ftz.f32 %f532, %f531, %f530;
st.global.f32 [%rd47+12], %f532;
bra.uni $L__BB306_78;

}
.entry _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<38>;
.reg .b16 %rs<93>;
.reg .f32 %f<109>;
.reg .b32 %r<939>;
.reg .b64 %rd<131>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd16, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd17, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd18, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
mov.u32 %r1, %tid.y;
shl.b32 %r40, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r40, %r2;
mov.u32 %r41, %ctaid.x;
shl.b32 %r4, %r41, 2;
shr.u32 %r5, %r39, 8;
mov.u32 %r936, %ctaid.y;
setp.eq.s64 %p1, %rd17, 0;
@%p1 bra $L__BB307_2;

cvta.to.global.u64 %rd19, %rd17;
mul.wide.u32 %rd20, %r936, 4;
add.s64 %rd21, %rd19, %rd20;
ld.global.nc.u32 %r935, [%rd21];
mul.hi.u32 %r42, %r936, %r22;
add.s32 %r43, %r42, %r936;
shr.u32 %r44, %r43, %r23;
mul.lo.s32 %r45, %r44, %r24;
sub.s32 %r936, %r936, %r45;
bra.uni $L__BB307_3;

$L__BB307_2:
mul.hi.u32 %r46, %r936, %r27;
add.s32 %r47, %r46, %r936;
shr.u32 %r935, %r47, %r28;

$L__BB307_3:
mov.u32 %r12, %ctaid.z;
shr.s32 %r48, %r3, 31;
shr.u32 %r49, %r48, 29;
add.s32 %r50, %r3, %r49;
shr.s32 %r938, %r50, 3;
setp.ge.s32 %p2, %r938, %r5;
mov.f32 %f105, 0f00000000;
mov.f32 %f106, %f105;
mov.f32 %f107, %f105;
mov.f32 %f108, %f105;
@%p2 bra $L__BB307_6;

and.b32 %r54, %r50, -8;
sub.s32 %r55, %r3, %r54;
shl.b32 %r56, %r55, 1;
shr.u32 %r57, %r55, 31;
add.s32 %r58, %r55, %r57;
shr.s32 %r59, %r58, 1;
cvt.s64.s32 %rd22, %r59;
shl.b32 %r60, %r55, 2;
cvt.s64.s32 %rd1, %r55;
and.b32 %r14, %r60, 4;
shl.b32 %r937, %r938, 3;
not.b32 %r61, %r56;
mul.wide.s32 %rd23, %r61, 2;
mul.wide.s32 %rd24, %r56, 2;
mad.lo.s32 %r62, %r4, %r25, %r938;
mul.hi.u32 %r63, %r12, %r33;
add.s32 %r64, %r63, %r12;
shr.u32 %r65, %r64, %r34;
mad.lo.s32 %r66, %r65, %r36, %r62;
mad.lo.s32 %r67, %r935, %r30, %r66;
mul.wide.s32 %rd2, %r67, 110;
mul.wide.s32 %rd25, %r60, 2;
add.s64 %rd26, %rd2, %rd25;
add.s64 %rd3, %rd26, 4;
add.s64 %rd27, %rd2, %rd1;
add.s64 %rd4, %rd27, 66;
add.s64 %rd28, %rd2, %rd24;
add.s64 %rd5, %rd28, 74;
add.s64 %rd29, %rd2, 74;
sub.s64 %rd6, %rd29, %rd23;
add.s64 %rd30, %rd2, %rd22;
add.s64 %rd7, %rd30, 106;
mul.wide.s32 %rd8, %r25, 110;
cvta.to.global.u64 %rd9, %rd16;
cvta.to.global.u64 %rd130, %rd15;
mul.lo.s32 %r68, %r936, %r31;
cvt.u64.u32 %rd31, %r68;
mul.lo.s32 %r69, %r12, %r37;
cvt.u64.u32 %rd32, %r69;
add.s64 %rd11, %rd31, %rd32;
mov.f32 %f105, 0f00000000;

$L__BB307_5:
cvt.u64.u32 %rd33, %r937;
add.s64 %rd34, %rd11, %rd33;
add.s64 %rd35, %rd130, %rd3;
add.s64 %rd36, %rd35, -2;
ld.global.nc.u16 %rs5, [%rd35];
ld.global.nc.u16 %rs6, [%rd35+2];
ld.global.nc.u16 %rs7, [%rd35+4];
add.s64 %rd37, %rd130, %rd4;
ld.global.nc.u8 %rs8, [%rd37];
add.s64 %rd38, %rd130, %rd5;
add.s64 %rd39, %rd130, %rd6;
ld.global.nc.u16 %rs9, [%rd38];
ld.global.nc.u16 %rs10, [%rd39];
add.s64 %rd40, %rd34, %rd1;
mul.lo.s64 %rd41, %rd40, 36;
ld.global.nc.u16 %rs11, [%rd35+-2];
and.b16 %rs12, %rs11, 255;
cvt.u32.u16 %r266, %rs12;
shl.b16 %rs13, %rs8, 8;
cvt.u32.u16 %r267, %rs13;
and.b32 %r268, %r267, 256;
or.b32 %r269, %r268, %r266;
mul.wide.u32 %rd42, %r269, 4;
mov.u64 %rd43, iq3s_grid;
add.s64 %rd44, %rd43, %rd42;
ld.global.nc.u32 %r270, [%rd44];
shr.u16 %rs14, %rs11, 8;
cvt.u32.u16 %r271, %rs14;
mul.wide.u16 %r272, %rs8, 128;
and.b32 %r273, %r272, 256;
or.b32 %r274, %r273, %r271;
mul.wide.u32 %rd45, %r274, 4;
add.s64 %rd46, %rd43, %rd45;
ld.global.nc.u32 %r275, [%rd46];
and.b16 %rs16, %rs9, 255;
mul.wide.u16 %r276, %rs16, 128;
cvt.u32.u16 %r277, %rs9;
shl.b32 %r278, %r277, 21;
or.b32 %r279, %r276, %r278;
and.b32 %r280, %r279, 25166208;
or.b32 %r281, %r280, -2139062144;
add.s32 %r282, %r281, -16843009;
or.b32 %r70, %r282, %r280;

prmt.b32 %r70, %r70, 0, 0xba98;

mul.wide.u16 %r283, %rs16, 8;
shl.b32 %r284, %r277, 17;
or.b32 %r285, %r283, %r284;
and.b32 %r286, %r285, 25166208;
or.b32 %r287, %r286, -2139062144;
add.s32 %r288, %r287, -16843009;
or.b32 %r72, %r288, %r286;

prmt.b32 %r72, %r72, 0, 0xba98;

xor.b32 %r289, %r70, %r270;
or.b32 %r290, %r289, -2139062144;
and.b32 %r291, %r70, 2139062143;
and.b32 %r292, %r270, -2139062144;
sub.s32 %r293, %r290, %r291;
xor.b32 %r294, %r292, %r293;
xor.b32 %r75, %r294, -2139062144;
xor.b32 %r295, %r72, %r275;
or.b32 %r296, %r295, -2139062144;
and.b32 %r297, %r72, 2139062143;
and.b32 %r298, %r275, -2139062144;
sub.s32 %r299, %r296, %r297;
xor.b32 %r300, %r298, %r299;
xor.b32 %r79, %r300, -2139062144;
add.s64 %rd47, %rd9, %rd41;
ld.global.nc.u32 %r76, [%rd47+4];
ld.global.nc.u32 %r80, [%rd47+8];
mov.u32 %r224, 0;

dp4a.s32.s32 %r74, %r75, %r76, %r224;


dp4a.s32.s32 %r78, %r79, %r80, %r74;

and.b16 %rs17, %rs5, 255;
cvt.u32.u16 %r301, %rs17;
mul.wide.u16 %r302, %rs8, 64;
and.b32 %r303, %r302, 256;
or.b32 %r304, %r303, %r301;
mul.wide.u32 %rd48, %r304, 4;
add.s64 %rd49, %rd43, %rd48;
ld.global.nc.u32 %r305, [%rd49];
shr.u16 %rs18, %rs5, 8;
cvt.u32.u16 %r306, %rs18;
mul.wide.u16 %r307, %rs8, 32;
and.b32 %r308, %r307, 256;
or.b32 %r309, %r308, %r306;
mul.wide.u32 %rd50, %r309, 4;
add.s64 %rd51, %rd43, %rd50;
ld.global.nc.u32 %r310, [%rd51];
shr.u16 %rs19, %rs9, 8;
mul.wide.u16 %r311, %rs19, 128;
cvt.u32.u16 %r312, %rs19;
shl.b32 %r313, %r312, 21;
or.b32 %r314, %r311, %r313;
and.b32 %r315, %r314, 25166208;
or.b32 %r316, %r315, -2139062144;
add.s32 %r317, %r316, -16843009;
or.b32 %r82, %r317, %r315;

prmt.b32 %r82, %r82, 0, 0xba98;

mul.wide.u16 %r318, %rs19, 8;
shl.b32 %r319, %r312, 17;
or.b32 %r320, %r318, %r319;
and.b32 %r321, %r320, 25166208;
or.b32 %r322, %r321, -2139062144;
add.s32 %r323, %r322, -16843009;
or.b32 %r84, %r323, %r321;

prmt.b32 %r84, %r84, 0, 0xba98;

xor.b32 %r324, %r82, %r305;
or.b32 %r325, %r324, -2139062144;
and.b32 %r326, %r82, 2139062143;
and.b32 %r327, %r305, -2139062144;
sub.s32 %r328, %r325, %r326;
xor.b32 %r329, %r327, %r328;
xor.b32 %r87, %r329, -2139062144;
xor.b32 %r330, %r84, %r310;
or.b32 %r331, %r330, -2139062144;
and.b32 %r332, %r84, 2139062143;
and.b32 %r333, %r310, -2139062144;
sub.s32 %r334, %r331, %r332;
xor.b32 %r335, %r333, %r334;
xor.b32 %r91, %r335, -2139062144;
ld.global.nc.u32 %r88, [%rd47+12];
ld.global.nc.u32 %r92, [%rd47+16];

dp4a.s32.s32 %r86, %r87, %r88, %r78;


dp4a.s32.s32 %r90, %r91, %r92, %r86;

and.b16 %rs20, %rs6, 255;
cvt.u32.u16 %r336, %rs20;
mul.wide.u16 %r337, %rs8, 16;
and.b32 %r338, %r337, 256;
or.b32 %r339, %r338, %r336;
mul.wide.u32 %rd52, %r339, 4;
add.s64 %rd53, %rd43, %rd52;
ld.global.nc.u32 %r340, [%rd53];
shr.u16 %rs21, %rs6, 8;
cvt.u32.u16 %r341, %rs21;
mul.wide.u16 %r342, %rs8, 8;
and.b32 %r343, %r342, 256;
or.b32 %r344, %r343, %r341;
mul.wide.u32 %rd54, %r344, 4;
add.s64 %rd55, %rd43, %rd54;
ld.global.nc.u32 %r345, [%rd55];
and.b16 %rs22, %rs10, 255;
mul.wide.u16 %r346, %rs22, 128;
cvt.u32.u16 %r347, %rs10;
shl.b32 %r348, %r347, 21;
or.b32 %r349, %r346, %r348;
and.b32 %r350, %r349, 25166208;
or.b32 %r351, %r350, -2139062144;
add.s32 %r352, %r351, -16843009;
or.b32 %r94, %r352, %r350;

prmt.b32 %r94, %r94, 0, 0xba98;

mul.wide.u16 %r353, %rs22, 8;
shl.b32 %r354, %r347, 17;
or.b32 %r355, %r353, %r354;
and.b32 %r356, %r355, 25166208;
or.b32 %r357, %r356, -2139062144;
add.s32 %r358, %r357, -16843009;
or.b32 %r96, %r358, %r356;

prmt.b32 %r96, %r96, 0, 0xba98;

xor.b32 %r359, %r94, %r340;
or.b32 %r360, %r359, -2139062144;
and.b32 %r361, %r94, 2139062143;
and.b32 %r362, %r340, -2139062144;
sub.s32 %r363, %r360, %r361;
xor.b32 %r364, %r362, %r363;
xor.b32 %r99, %r364, -2139062144;
xor.b32 %r365, %r96, %r345;
or.b32 %r366, %r365, -2139062144;
and.b32 %r367, %r96, 2139062143;
and.b32 %r368, %r345, -2139062144;
sub.s32 %r369, %r366, %r367;
xor.b32 %r370, %r368, %r369;
xor.b32 %r103, %r370, -2139062144;
ld.global.nc.u32 %r100, [%rd47+20];
ld.global.nc.u32 %r104, [%rd47+24];

dp4a.s32.s32 %r98, %r99, %r100, %r90;


dp4a.s32.s32 %r102, %r103, %r104, %r98;

and.b16 %rs23, %rs7, 255;
cvt.u32.u16 %r371, %rs23;
mul.wide.u16 %r372, %rs8, 4;
and.b32 %r373, %r372, 256;
or.b32 %r374, %r373, %r371;
mul.wide.u32 %rd56, %r374, 4;
add.s64 %rd57, %rd43, %rd56;
ld.global.nc.u32 %r375, [%rd57];
shr.u16 %rs24, %rs7, 8;
cvt.u32.u16 %r376, %rs24;
mul.wide.u16 %r377, %rs8, 2;
and.b32 %r378, %r377, 256;
or.b32 %r379, %r378, %r376;
mul.wide.u32 %rd58, %r379, 4;
add.s64 %rd59, %rd43, %rd58;
ld.global.nc.u32 %r380, [%rd59];
shr.u16 %rs25, %rs10, 8;
mul.wide.u16 %r381, %rs25, 128;
cvt.u32.u16 %r382, %rs25;
shl.b32 %r383, %r382, 21;
or.b32 %r384, %r381, %r383;
and.b32 %r385, %r384, 25166208;
or.b32 %r386, %r385, -2139062144;
add.s32 %r387, %r386, -16843009;
or.b32 %r106, %r387, %r385;

prmt.b32 %r106, %r106, 0, 0xba98;

mul.wide.u16 %r388, %rs25, 8;
shl.b32 %r389, %r382, 17;
or.b32 %r390, %r388, %r389;
and.b32 %r391, %r390, 25166208;
or.b32 %r392, %r391, -2139062144;
add.s32 %r393, %r392, -16843009;
or.b32 %r108, %r393, %r391;

prmt.b32 %r108, %r108, 0, 0xba98;

xor.b32 %r394, %r106, %r375;
or.b32 %r395, %r394, -2139062144;
and.b32 %r396, %r106, 2139062143;
and.b32 %r397, %r375, -2139062144;
sub.s32 %r398, %r395, %r396;
xor.b32 %r399, %r397, %r398;
xor.b32 %r111, %r399, -2139062144;
xor.b32 %r400, %r108, %r380;
or.b32 %r401, %r400, -2139062144;
and.b32 %r402, %r108, 2139062143;
and.b32 %r403, %r380, -2139062144;
sub.s32 %r404, %r401, %r402;
xor.b32 %r405, %r403, %r404;
xor.b32 %r115, %r405, -2139062144;
ld.global.nc.u32 %r112, [%rd47+28];
ld.global.nc.u32 %r116, [%rd47+32];

dp4a.s32.s32 %r110, %r111, %r112, %r102;


dp4a.s32.s32 %r114, %r115, %r116, %r110;

add.s64 %rd60, %rd130, %rd7;
ld.global.nc.u8 %rs26, [%rd60];
cvt.u32.u16 %r406, %rs26;
and.b32 %r407, %r406, 255;
shr.u32 %r408, %r407, %r14;
shl.b32 %r409, %r408, 1;
and.b32 %r410, %r409, 30;
or.b32 %r411, %r410, 1;
mul.lo.s32 %r412, %r411, %r114;
add.s64 %rd61, %rd130, %rd2;
ld.global.nc.u16 %rs1, [%rd61];

{ cvt.f32.f16 %f25, %rs1;}


ld.global.nc.u32 %r118, [%rd47];

{.reg .f16 low,high;
mov.b32 {low,high},%r118;
cvt.f32.f16 %f26, low;}


mul.ftz.f32 %f33, %f25, %f26;
cvt.rn.f32.s32 %f34, %r412;
fma.rn.ftz.f32 %f108, %f33, %f34, %f108;
add.s64 %rd62, %rd36, %rd8;
ld.global.nc.u16 %rs27, [%rd62+2];
ld.global.nc.u16 %rs28, [%rd62+4];
ld.global.nc.u16 %rs29, [%rd62+6];
add.s64 %rd63, %rd37, %rd8;
ld.global.nc.u8 %rs30, [%rd63];
add.s64 %rd64, %rd38, %rd8;
add.s64 %rd65, %rd39, %rd8;
ld.global.nc.u16 %rs31, [%rd64];
ld.global.nc.u16 %rs32, [%rd65];
ld.global.nc.u16 %rs33, [%rd62];
and.b16 %rs34, %rs33, 255;
cvt.u32.u16 %r413, %rs34;
shl.b16 %rs35, %rs30, 8;
cvt.u32.u16 %r414, %rs35;
and.b32 %r415, %r414, 256;
or.b32 %r416, %r415, %r413;
mul.wide.u32 %rd66, %r416, 4;
add.s64 %rd67, %rd43, %rd66;
ld.global.nc.u32 %r417, [%rd67];
shr.u16 %rs36, %rs33, 8;
cvt.u32.u16 %r418, %rs36;
mul.wide.u16 %r419, %rs30, 128;
and.b32 %r420, %r419, 256;
or.b32 %r421, %r420, %r418;
mul.wide.u32 %rd68, %r421, 4;
add.s64 %rd69, %rd43, %rd68;
ld.global.nc.u32 %r422, [%rd69];
and.b16 %rs38, %rs31, 255;
mul.wide.u16 %r423, %rs38, 128;
cvt.u32.u16 %r424, %rs31;
shl.b32 %r425, %r424, 21;
or.b32 %r426, %r423, %r425;
and.b32 %r427, %r426, 25166208;
or.b32 %r428, %r427, -2139062144;
add.s32 %r429, %r428, -16843009;
or.b32 %r119, %r429, %r427;

prmt.b32 %r119, %r119, 0, 0xba98;

mul.wide.u16 %r430, %rs38, 8;
shl.b32 %r431, %r424, 17;
or.b32 %r432, %r430, %r431;
and.b32 %r433, %r432, 25166208;
or.b32 %r434, %r433, -2139062144;
add.s32 %r435, %r434, -16843009;
or.b32 %r121, %r435, %r433;

prmt.b32 %r121, %r121, 0, 0xba98;

xor.b32 %r436, %r119, %r417;
or.b32 %r437, %r436, -2139062144;
and.b32 %r438, %r119, 2139062143;
and.b32 %r439, %r417, -2139062144;
sub.s32 %r440, %r437, %r438;
xor.b32 %r441, %r439, %r440;
xor.b32 %r124, %r441, -2139062144;
xor.b32 %r442, %r121, %r422;
or.b32 %r443, %r442, -2139062144;
and.b32 %r444, %r121, 2139062143;
and.b32 %r445, %r422, -2139062144;
sub.s32 %r446, %r443, %r444;
xor.b32 %r447, %r445, %r446;
xor.b32 %r128, %r447, -2139062144;

dp4a.s32.s32 %r123, %r124, %r76, %r224;


dp4a.s32.s32 %r127, %r128, %r80, %r123;

and.b16 %rs39, %rs27, 255;
cvt.u32.u16 %r448, %rs39;
mul.wide.u16 %r449, %rs30, 64;
and.b32 %r450, %r449, 256;
or.b32 %r451, %r450, %r448;
mul.wide.u32 %rd70, %r451, 4;
add.s64 %rd71, %rd43, %rd70;
ld.global.nc.u32 %r452, [%rd71];
shr.u16 %rs40, %rs27, 8;
cvt.u32.u16 %r453, %rs40;
mul.wide.u16 %r454, %rs30, 32;
and.b32 %r455, %r454, 256;
or.b32 %r456, %r455, %r453;
mul.wide.u32 %rd72, %r456, 4;
add.s64 %rd73, %rd43, %rd72;
ld.global.nc.u32 %r457, [%rd73];
shr.u16 %rs41, %rs31, 8;
mul.wide.u16 %r458, %rs41, 128;
cvt.u32.u16 %r459, %rs41;
shl.b32 %r460, %r459, 21;
or.b32 %r461, %r458, %r460;
and.b32 %r462, %r461, 25166208;
or.b32 %r463, %r462, -2139062144;
add.s32 %r464, %r463, -16843009;
or.b32 %r131, %r464, %r462;

prmt.b32 %r131, %r131, 0, 0xba98;

mul.wide.u16 %r465, %rs41, 8;
shl.b32 %r466, %r459, 17;
or.b32 %r467, %r465, %r466;
and.b32 %r468, %r467, 25166208;
or.b32 %r469, %r468, -2139062144;
add.s32 %r470, %r469, -16843009;
or.b32 %r133, %r470, %r468;

prmt.b32 %r133, %r133, 0, 0xba98;

xor.b32 %r471, %r131, %r452;
or.b32 %r472, %r471, -2139062144;
and.b32 %r473, %r131, 2139062143;
and.b32 %r474, %r452, -2139062144;
sub.s32 %r475, %r472, %r473;
xor.b32 %r476, %r474, %r475;
xor.b32 %r136, %r476, -2139062144;
xor.b32 %r477, %r133, %r457;
or.b32 %r478, %r477, -2139062144;
and.b32 %r479, %r133, 2139062143;
and.b32 %r480, %r457, -2139062144;
sub.s32 %r481, %r478, %r479;
xor.b32 %r482, %r480, %r481;
xor.b32 %r140, %r482, -2139062144;

dp4a.s32.s32 %r135, %r136, %r88, %r127;


dp4a.s32.s32 %r139, %r140, %r92, %r135;

and.b16 %rs42, %rs28, 255;
cvt.u32.u16 %r483, %rs42;
mul.wide.u16 %r484, %rs30, 16;
and.b32 %r485, %r484, 256;
or.b32 %r486, %r485, %r483;
mul.wide.u32 %rd74, %r486, 4;
add.s64 %rd75, %rd43, %rd74;
ld.global.nc.u32 %r487, [%rd75];
shr.u16 %rs43, %rs28, 8;
cvt.u32.u16 %r488, %rs43;
mul.wide.u16 %r489, %rs30, 8;
and.b32 %r490, %r489, 256;
or.b32 %r491, %r490, %r488;
mul.wide.u32 %rd76, %r491, 4;
add.s64 %rd77, %rd43, %rd76;
ld.global.nc.u32 %r492, [%rd77];
and.b16 %rs44, %rs32, 255;
mul.wide.u16 %r493, %rs44, 128;
cvt.u32.u16 %r494, %rs32;
shl.b32 %r495, %r494, 21;
or.b32 %r496, %r493, %r495;
and.b32 %r497, %r496, 25166208;
or.b32 %r498, %r497, -2139062144;
add.s32 %r499, %r498, -16843009;
or.b32 %r143, %r499, %r497;

prmt.b32 %r143, %r143, 0, 0xba98;

mul.wide.u16 %r500, %rs44, 8;
shl.b32 %r501, %r494, 17;
or.b32 %r502, %r500, %r501;
and.b32 %r503, %r502, 25166208;
or.b32 %r504, %r503, -2139062144;
add.s32 %r505, %r504, -16843009;
or.b32 %r145, %r505, %r503;

prmt.b32 %r145, %r145, 0, 0xba98;

xor.b32 %r506, %r143, %r487;
or.b32 %r507, %r506, -2139062144;
and.b32 %r508, %r143, 2139062143;
and.b32 %r509, %r487, -2139062144;
sub.s32 %r510, %r507, %r508;
xor.b32 %r511, %r509, %r510;
xor.b32 %r148, %r511, -2139062144;
xor.b32 %r512, %r145, %r492;
or.b32 %r513, %r512, -2139062144;
and.b32 %r514, %r145, 2139062143;
and.b32 %r515, %r492, -2139062144;
sub.s32 %r516, %r513, %r514;
xor.b32 %r517, %r515, %r516;
xor.b32 %r152, %r517, -2139062144;

dp4a.s32.s32 %r147, %r148, %r100, %r139;


dp4a.s32.s32 %r151, %r152, %r104, %r147;

and.b16 %rs45, %rs29, 255;
cvt.u32.u16 %r518, %rs45;
mul.wide.u16 %r519, %rs30, 4;
and.b32 %r520, %r519, 256;
or.b32 %r521, %r520, %r518;
mul.wide.u32 %rd78, %r521, 4;
add.s64 %rd79, %rd43, %rd78;
ld.global.nc.u32 %r522, [%rd79];
shr.u16 %rs46, %rs29, 8;
cvt.u32.u16 %r523, %rs46;
mul.wide.u16 %r524, %rs30, 2;
and.b32 %r525, %r524, 256;
or.b32 %r526, %r525, %r523;
mul.wide.u32 %rd80, %r526, 4;
add.s64 %rd81, %rd43, %rd80;
ld.global.nc.u32 %r527, [%rd81];
shr.u16 %rs47, %rs32, 8;
mul.wide.u16 %r528, %rs47, 128;
cvt.u32.u16 %r529, %rs47;
shl.b32 %r530, %r529, 21;
or.b32 %r531, %r528, %r530;
and.b32 %r532, %r531, 25166208;
or.b32 %r533, %r532, -2139062144;
add.s32 %r534, %r533, -16843009;
or.b32 %r155, %r534, %r532;

prmt.b32 %r155, %r155, 0, 0xba98;

mul.wide.u16 %r535, %rs47, 8;
shl.b32 %r536, %r529, 17;
or.b32 %r537, %r535, %r536;
and.b32 %r538, %r537, 25166208;
or.b32 %r539, %r538, -2139062144;
add.s32 %r540, %r539, -16843009;
or.b32 %r157, %r540, %r538;

prmt.b32 %r157, %r157, 0, 0xba98;

xor.b32 %r541, %r155, %r522;
or.b32 %r542, %r541, -2139062144;
and.b32 %r543, %r155, 2139062143;
and.b32 %r544, %r522, -2139062144;
sub.s32 %r545, %r542, %r543;
xor.b32 %r546, %r544, %r545;
xor.b32 %r160, %r546, -2139062144;
xor.b32 %r547, %r157, %r527;
or.b32 %r548, %r547, -2139062144;
and.b32 %r549, %r157, 2139062143;
and.b32 %r550, %r527, -2139062144;
sub.s32 %r551, %r548, %r549;
xor.b32 %r552, %r550, %r551;
xor.b32 %r164, %r552, -2139062144;

dp4a.s32.s32 %r159, %r160, %r112, %r151;


dp4a.s32.s32 %r163, %r164, %r116, %r159;

add.s64 %rd82, %rd60, %rd8;
ld.global.nc.u8 %rs48, [%rd82];
cvt.u32.u16 %r553, %rs48;
and.b32 %r554, %r553, 255;
shr.u32 %r555, %r554, %r14;
shl.b32 %r556, %r555, 1;
and.b32 %r557, %r556, 30;
or.b32 %r558, %r557, 1;
mul.lo.s32 %r559, %r558, %r163;
add.s64 %rd83, %rd61, %rd8;
ld.global.nc.u16 %rs2, [%rd83];

{ cvt.f32.f16 %f27, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r118;
cvt.f32.f16 %f28, low;}


mul.ftz.f32 %f35, %f27, %f28;
cvt.rn.f32.s32 %f36, %r559;
fma.rn.ftz.f32 %f107, %f35, %f36, %f107;
add.s64 %rd84, %rd62, %rd8;
ld.global.nc.u16 %rs49, [%rd84+2];
ld.global.nc.u16 %rs50, [%rd84+4];
ld.global.nc.u16 %rs51, [%rd84+6];
add.s64 %rd85, %rd63, %rd8;
ld.global.nc.u8 %rs52, [%rd85];
add.s64 %rd86, %rd64, %rd8;
add.s64 %rd87, %rd65, %rd8;
ld.global.nc.u16 %rs53, [%rd86];
ld.global.nc.u16 %rs54, [%rd87];
ld.global.nc.u16 %rs55, [%rd84];
and.b16 %rs56, %rs55, 255;
cvt.u32.u16 %r560, %rs56;
shl.b16 %rs57, %rs52, 8;
cvt.u32.u16 %r561, %rs57;
and.b32 %r562, %r561, 256;
or.b32 %r563, %r562, %r560;
mul.wide.u32 %rd88, %r563, 4;
add.s64 %rd89, %rd43, %rd88;
ld.global.nc.u32 %r564, [%rd89];
shr.u16 %rs58, %rs55, 8;
cvt.u32.u16 %r565, %rs58;
mul.wide.u16 %r566, %rs52, 128;
and.b32 %r567, %r566, 256;
or.b32 %r568, %r567, %r565;
mul.wide.u32 %rd90, %r568, 4;
add.s64 %rd91, %rd43, %rd90;
ld.global.nc.u32 %r569, [%rd91];
and.b16 %rs60, %rs53, 255;
mul.wide.u16 %r570, %rs60, 128;
cvt.u32.u16 %r571, %rs53;
shl.b32 %r572, %r571, 21;
or.b32 %r573, %r570, %r572;
and.b32 %r574, %r573, 25166208;
or.b32 %r575, %r574, -2139062144;
add.s32 %r576, %r575, -16843009;
or.b32 %r168, %r576, %r574;

prmt.b32 %r168, %r168, 0, 0xba98;

mul.wide.u16 %r577, %rs60, 8;
shl.b32 %r578, %r571, 17;
or.b32 %r579, %r577, %r578;
and.b32 %r580, %r579, 25166208;
or.b32 %r581, %r580, -2139062144;
add.s32 %r582, %r581, -16843009;
or.b32 %r170, %r582, %r580;

prmt.b32 %r170, %r170, 0, 0xba98;

xor.b32 %r583, %r168, %r564;
or.b32 %r584, %r583, -2139062144;
and.b32 %r585, %r168, 2139062143;
and.b32 %r586, %r564, -2139062144;
sub.s32 %r587, %r584, %r585;
xor.b32 %r588, %r586, %r587;
xor.b32 %r173, %r588, -2139062144;
xor.b32 %r589, %r170, %r569;
or.b32 %r590, %r589, -2139062144;
and.b32 %r591, %r170, 2139062143;
and.b32 %r592, %r569, -2139062144;
sub.s32 %r593, %r590, %r591;
xor.b32 %r594, %r592, %r593;
xor.b32 %r177, %r594, -2139062144;

dp4a.s32.s32 %r172, %r173, %r76, %r224;


dp4a.s32.s32 %r176, %r177, %r80, %r172;

and.b16 %rs61, %rs49, 255;
cvt.u32.u16 %r595, %rs61;
mul.wide.u16 %r596, %rs52, 64;
and.b32 %r597, %r596, 256;
or.b32 %r598, %r597, %r595;
mul.wide.u32 %rd92, %r598, 4;
add.s64 %rd93, %rd43, %rd92;
ld.global.nc.u32 %r599, [%rd93];
shr.u16 %rs62, %rs49, 8;
cvt.u32.u16 %r600, %rs62;
mul.wide.u16 %r601, %rs52, 32;
and.b32 %r602, %r601, 256;
or.b32 %r603, %r602, %r600;
mul.wide.u32 %rd94, %r603, 4;
add.s64 %rd95, %rd43, %rd94;
ld.global.nc.u32 %r604, [%rd95];
shr.u16 %rs63, %rs53, 8;
mul.wide.u16 %r605, %rs63, 128;
cvt.u32.u16 %r606, %rs63;
shl.b32 %r607, %r606, 21;
or.b32 %r608, %r605, %r607;
and.b32 %r609, %r608, 25166208;
or.b32 %r610, %r609, -2139062144;
add.s32 %r611, %r610, -16843009;
or.b32 %r180, %r611, %r609;

prmt.b32 %r180, %r180, 0, 0xba98;

mul.wide.u16 %r612, %rs63, 8;
shl.b32 %r613, %r606, 17;
or.b32 %r614, %r612, %r613;
and.b32 %r615, %r614, 25166208;
or.b32 %r616, %r615, -2139062144;
add.s32 %r617, %r616, -16843009;
or.b32 %r182, %r617, %r615;

prmt.b32 %r182, %r182, 0, 0xba98;

xor.b32 %r618, %r180, %r599;
or.b32 %r619, %r618, -2139062144;
and.b32 %r620, %r180, 2139062143;
and.b32 %r621, %r599, -2139062144;
sub.s32 %r622, %r619, %r620;
xor.b32 %r623, %r621, %r622;
xor.b32 %r185, %r623, -2139062144;
xor.b32 %r624, %r182, %r604;
or.b32 %r625, %r624, -2139062144;
and.b32 %r626, %r182, 2139062143;
and.b32 %r627, %r604, -2139062144;
sub.s32 %r628, %r625, %r626;
xor.b32 %r629, %r627, %r628;
xor.b32 %r189, %r629, -2139062144;

dp4a.s32.s32 %r184, %r185, %r88, %r176;


dp4a.s32.s32 %r188, %r189, %r92, %r184;

and.b16 %rs64, %rs50, 255;
cvt.u32.u16 %r630, %rs64;
mul.wide.u16 %r631, %rs52, 16;
and.b32 %r632, %r631, 256;
or.b32 %r633, %r632, %r630;
mul.wide.u32 %rd96, %r633, 4;
add.s64 %rd97, %rd43, %rd96;
ld.global.nc.u32 %r634, [%rd97];
shr.u16 %rs65, %rs50, 8;
cvt.u32.u16 %r635, %rs65;
mul.wide.u16 %r636, %rs52, 8;
and.b32 %r637, %r636, 256;
or.b32 %r638, %r637, %r635;
mul.wide.u32 %rd98, %r638, 4;
add.s64 %rd99, %rd43, %rd98;
ld.global.nc.u32 %r639, [%rd99];
and.b16 %rs66, %rs54, 255;
mul.wide.u16 %r640, %rs66, 128;
cvt.u32.u16 %r641, %rs54;
shl.b32 %r642, %r641, 21;
or.b32 %r643, %r640, %r642;
and.b32 %r644, %r643, 25166208;
or.b32 %r645, %r644, -2139062144;
add.s32 %r646, %r645, -16843009;
or.b32 %r192, %r646, %r644;

prmt.b32 %r192, %r192, 0, 0xba98;

mul.wide.u16 %r647, %rs66, 8;
shl.b32 %r648, %r641, 17;
or.b32 %r649, %r647, %r648;
and.b32 %r650, %r649, 25166208;
or.b32 %r651, %r650, -2139062144;
add.s32 %r652, %r651, -16843009;
or.b32 %r194, %r652, %r650;

prmt.b32 %r194, %r194, 0, 0xba98;

xor.b32 %r653, %r192, %r634;
or.b32 %r654, %r653, -2139062144;
and.b32 %r655, %r192, 2139062143;
and.b32 %r656, %r634, -2139062144;
sub.s32 %r657, %r654, %r655;
xor.b32 %r658, %r656, %r657;
xor.b32 %r197, %r658, -2139062144;
xor.b32 %r659, %r194, %r639;
or.b32 %r660, %r659, -2139062144;
and.b32 %r661, %r194, 2139062143;
and.b32 %r662, %r639, -2139062144;
sub.s32 %r663, %r660, %r661;
xor.b32 %r664, %r662, %r663;
xor.b32 %r201, %r664, -2139062144;

dp4a.s32.s32 %r196, %r197, %r100, %r188;


dp4a.s32.s32 %r200, %r201, %r104, %r196;

and.b16 %rs67, %rs51, 255;
cvt.u32.u16 %r665, %rs67;
mul.wide.u16 %r666, %rs52, 4;
and.b32 %r667, %r666, 256;
or.b32 %r668, %r667, %r665;
mul.wide.u32 %rd100, %r668, 4;
add.s64 %rd101, %rd43, %rd100;
ld.global.nc.u32 %r669, [%rd101];
shr.u16 %rs68, %rs51, 8;
cvt.u32.u16 %r670, %rs68;
mul.wide.u16 %r671, %rs52, 2;
and.b32 %r672, %r671, 256;
or.b32 %r673, %r672, %r670;
mul.wide.u32 %rd102, %r673, 4;
add.s64 %rd103, %rd43, %rd102;
ld.global.nc.u32 %r674, [%rd103];
shr.u16 %rs69, %rs54, 8;
mul.wide.u16 %r675, %rs69, 128;
cvt.u32.u16 %r676, %rs69;
shl.b32 %r677, %r676, 21;
or.b32 %r678, %r675, %r677;
and.b32 %r679, %r678, 25166208;
or.b32 %r680, %r679, -2139062144;
add.s32 %r681, %r680, -16843009;
or.b32 %r204, %r681, %r679;

prmt.b32 %r204, %r204, 0, 0xba98;

mul.wide.u16 %r682, %rs69, 8;
shl.b32 %r683, %r676, 17;
or.b32 %r684, %r682, %r683;
and.b32 %r685, %r684, 25166208;
or.b32 %r686, %r685, -2139062144;
add.s32 %r687, %r686, -16843009;
or.b32 %r206, %r687, %r685;

prmt.b32 %r206, %r206, 0, 0xba98;

xor.b32 %r688, %r204, %r669;
or.b32 %r689, %r688, -2139062144;
and.b32 %r690, %r204, 2139062143;
and.b32 %r691, %r669, -2139062144;
sub.s32 %r692, %r689, %r690;
xor.b32 %r693, %r691, %r692;
xor.b32 %r209, %r693, -2139062144;
xor.b32 %r694, %r206, %r674;
or.b32 %r695, %r694, -2139062144;
and.b32 %r696, %r206, 2139062143;
and.b32 %r697, %r674, -2139062144;
sub.s32 %r698, %r695, %r696;
xor.b32 %r699, %r697, %r698;
xor.b32 %r213, %r699, -2139062144;

dp4a.s32.s32 %r208, %r209, %r112, %r200;


dp4a.s32.s32 %r212, %r213, %r116, %r208;

add.s64 %rd104, %rd82, %rd8;
ld.global.nc.u8 %rs70, [%rd104];
cvt.u32.u16 %r700, %rs70;
and.b32 %r701, %r700, 255;
shr.u32 %r702, %r701, %r14;
shl.b32 %r703, %r702, 1;
and.b32 %r704, %r703, 30;
or.b32 %r705, %r704, 1;
mul.lo.s32 %r706, %r705, %r212;
add.s64 %rd105, %rd83, %rd8;
ld.global.nc.u16 %rs3, [%rd105];

{ cvt.f32.f16 %f29, %rs3;}



{.reg .f16 low,high;
mov.b32 {low,high},%r118;
cvt.f32.f16 %f30, low;}


mul.ftz.f32 %f37, %f29, %f30;
cvt.rn.f32.s32 %f38, %r706;
fma.rn.ftz.f32 %f106, %f37, %f38, %f106;
add.s64 %rd106, %rd84, %rd8;
ld.global.nc.u16 %rs71, [%rd106+2];
ld.global.nc.u16 %rs72, [%rd106+4];
ld.global.nc.u16 %rs73, [%rd106+6];
add.s64 %rd107, %rd85, %rd8;
ld.global.nc.u8 %rs74, [%rd107];
add.s64 %rd108, %rd86, %rd8;
add.s64 %rd109, %rd87, %rd8;
ld.global.nc.u16 %rs75, [%rd108];
ld.global.nc.u16 %rs76, [%rd109];
ld.global.nc.u16 %rs77, [%rd106];
and.b16 %rs78, %rs77, 255;
cvt.u32.u16 %r707, %rs78;
shl.b16 %rs79, %rs74, 8;
cvt.u32.u16 %r708, %rs79;
and.b32 %r709, %r708, 256;
or.b32 %r710, %r709, %r707;
mul.wide.u32 %rd110, %r710, 4;
add.s64 %rd111, %rd43, %rd110;
ld.global.nc.u32 %r711, [%rd111];
shr.u16 %rs80, %rs77, 8;
cvt.u32.u16 %r712, %rs80;
mul.wide.u16 %r713, %rs74, 128;
and.b32 %r714, %r713, 256;
or.b32 %r715, %r714, %r712;
mul.wide.u32 %rd112, %r715, 4;
add.s64 %rd113, %rd43, %rd112;
ld.global.nc.u32 %r716, [%rd113];
and.b16 %rs82, %rs75, 255;
mul.wide.u16 %r717, %rs82, 128;
cvt.u32.u16 %r718, %rs75;
shl.b32 %r719, %r718, 21;
or.b32 %r720, %r717, %r719;
and.b32 %r721, %r720, 25166208;
or.b32 %r722, %r721, -2139062144;
add.s32 %r723, %r722, -16843009;
or.b32 %r217, %r723, %r721;

prmt.b32 %r217, %r217, 0, 0xba98;

mul.wide.u16 %r724, %rs82, 8;
shl.b32 %r725, %r718, 17;
or.b32 %r726, %r724, %r725;
and.b32 %r727, %r726, 25166208;
or.b32 %r728, %r727, -2139062144;
add.s32 %r729, %r728, -16843009;
or.b32 %r219, %r729, %r727;

prmt.b32 %r219, %r219, 0, 0xba98;

xor.b32 %r730, %r217, %r711;
or.b32 %r731, %r730, -2139062144;
and.b32 %r732, %r217, 2139062143;
and.b32 %r733, %r711, -2139062144;
sub.s32 %r734, %r731, %r732;
xor.b32 %r735, %r733, %r734;
xor.b32 %r222, %r735, -2139062144;
xor.b32 %r736, %r219, %r716;
or.b32 %r737, %r736, -2139062144;
and.b32 %r738, %r219, 2139062143;
and.b32 %r739, %r716, -2139062144;
sub.s32 %r740, %r737, %r738;
xor.b32 %r741, %r739, %r740;
xor.b32 %r226, %r741, -2139062144;

dp4a.s32.s32 %r221, %r222, %r76, %r224;


dp4a.s32.s32 %r225, %r226, %r80, %r221;

and.b16 %rs83, %rs71, 255;
cvt.u32.u16 %r742, %rs83;
mul.wide.u16 %r743, %rs74, 64;
and.b32 %r744, %r743, 256;
or.b32 %r745, %r744, %r742;
mul.wide.u32 %rd114, %r745, 4;
add.s64 %rd115, %rd43, %rd114;
ld.global.nc.u32 %r746, [%rd115];
shr.u16 %rs84, %rs71, 8;
cvt.u32.u16 %r747, %rs84;
mul.wide.u16 %r748, %rs74, 32;
and.b32 %r749, %r748, 256;
or.b32 %r750, %r749, %r747;
mul.wide.u32 %rd116, %r750, 4;
add.s64 %rd117, %rd43, %rd116;
ld.global.nc.u32 %r751, [%rd117];
shr.u16 %rs85, %rs75, 8;
mul.wide.u16 %r752, %rs85, 128;
cvt.u32.u16 %r753, %rs85;
shl.b32 %r754, %r753, 21;
or.b32 %r755, %r752, %r754;
and.b32 %r756, %r755, 25166208;
or.b32 %r757, %r756, -2139062144;
add.s32 %r758, %r757, -16843009;
or.b32 %r229, %r758, %r756;

prmt.b32 %r229, %r229, 0, 0xba98;

mul.wide.u16 %r759, %rs85, 8;
shl.b32 %r760, %r753, 17;
or.b32 %r761, %r759, %r760;
and.b32 %r762, %r761, 25166208;
or.b32 %r763, %r762, -2139062144;
add.s32 %r764, %r763, -16843009;
or.b32 %r231, %r764, %r762;

prmt.b32 %r231, %r231, 0, 0xba98;

xor.b32 %r765, %r229, %r746;
or.b32 %r766, %r765, -2139062144;
and.b32 %r767, %r229, 2139062143;
and.b32 %r768, %r746, -2139062144;
sub.s32 %r769, %r766, %r767;
xor.b32 %r770, %r768, %r769;
xor.b32 %r234, %r770, -2139062144;
xor.b32 %r771, %r231, %r751;
or.b32 %r772, %r771, -2139062144;
and.b32 %r773, %r231, 2139062143;
and.b32 %r774, %r751, -2139062144;
sub.s32 %r775, %r772, %r773;
xor.b32 %r776, %r774, %r775;
xor.b32 %r238, %r776, -2139062144;

dp4a.s32.s32 %r233, %r234, %r88, %r225;


dp4a.s32.s32 %r237, %r238, %r92, %r233;

and.b16 %rs86, %rs72, 255;
cvt.u32.u16 %r777, %rs86;
mul.wide.u16 %r778, %rs74, 16;
and.b32 %r779, %r778, 256;
or.b32 %r780, %r779, %r777;
mul.wide.u32 %rd118, %r780, 4;
add.s64 %rd119, %rd43, %rd118;
ld.global.nc.u32 %r781, [%rd119];
shr.u16 %rs87, %rs72, 8;
cvt.u32.u16 %r782, %rs87;
mul.wide.u16 %r783, %rs74, 8;
and.b32 %r784, %r783, 256;
or.b32 %r785, %r784, %r782;
mul.wide.u32 %rd120, %r785, 4;
add.s64 %rd121, %rd43, %rd120;
ld.global.nc.u32 %r786, [%rd121];
and.b16 %rs88, %rs76, 255;
mul.wide.u16 %r787, %rs88, 128;
cvt.u32.u16 %r788, %rs76;
shl.b32 %r789, %r788, 21;
or.b32 %r790, %r787, %r789;
and.b32 %r791, %r790, 25166208;
or.b32 %r792, %r791, -2139062144;
add.s32 %r793, %r792, -16843009;
or.b32 %r241, %r793, %r791;

prmt.b32 %r241, %r241, 0, 0xba98;

mul.wide.u16 %r794, %rs88, 8;
shl.b32 %r795, %r788, 17;
or.b32 %r796, %r794, %r795;
and.b32 %r797, %r796, 25166208;
or.b32 %r798, %r797, -2139062144;
add.s32 %r799, %r798, -16843009;
or.b32 %r243, %r799, %r797;

prmt.b32 %r243, %r243, 0, 0xba98;

xor.b32 %r800, %r241, %r781;
or.b32 %r801, %r800, -2139062144;
and.b32 %r802, %r241, 2139062143;
and.b32 %r803, %r781, -2139062144;
sub.s32 %r804, %r801, %r802;
xor.b32 %r805, %r803, %r804;
xor.b32 %r246, %r805, -2139062144;
xor.b32 %r806, %r243, %r786;
or.b32 %r807, %r806, -2139062144;
and.b32 %r808, %r243, 2139062143;
and.b32 %r809, %r786, -2139062144;
sub.s32 %r810, %r807, %r808;
xor.b32 %r811, %r809, %r810;
xor.b32 %r250, %r811, -2139062144;

dp4a.s32.s32 %r245, %r246, %r100, %r237;


dp4a.s32.s32 %r249, %r250, %r104, %r245;

and.b16 %rs89, %rs73, 255;
cvt.u32.u16 %r812, %rs89;
mul.wide.u16 %r813, %rs74, 4;
and.b32 %r814, %r813, 256;
or.b32 %r815, %r814, %r812;
mul.wide.u32 %rd122, %r815, 4;
add.s64 %rd123, %rd43, %rd122;
ld.global.nc.u32 %r816, [%rd123];
shr.u16 %rs90, %rs73, 8;
cvt.u32.u16 %r817, %rs90;
mul.wide.u16 %r818, %rs74, 2;
and.b32 %r819, %r818, 256;
or.b32 %r820, %r819, %r817;
mul.wide.u32 %rd124, %r820, 4;
add.s64 %rd125, %rd43, %rd124;
ld.global.nc.u32 %r821, [%rd125];
shr.u16 %rs91, %rs76, 8;
mul.wide.u16 %r822, %rs91, 128;
cvt.u32.u16 %r823, %rs91;
shl.b32 %r824, %r823, 21;
or.b32 %r825, %r822, %r824;
and.b32 %r826, %r825, 25166208;
or.b32 %r827, %r826, -2139062144;
add.s32 %r828, %r827, -16843009;
or.b32 %r253, %r828, %r826;

prmt.b32 %r253, %r253, 0, 0xba98;

mul.wide.u16 %r829, %rs91, 8;
shl.b32 %r830, %r823, 17;
or.b32 %r831, %r829, %r830;
and.b32 %r832, %r831, 25166208;
or.b32 %r833, %r832, -2139062144;
add.s32 %r834, %r833, -16843009;
or.b32 %r255, %r834, %r832;

prmt.b32 %r255, %r255, 0, 0xba98;

xor.b32 %r835, %r253, %r816;
or.b32 %r836, %r835, -2139062144;
and.b32 %r837, %r253, 2139062143;
and.b32 %r838, %r816, -2139062144;
sub.s32 %r839, %r836, %r837;
xor.b32 %r840, %r838, %r839;
xor.b32 %r258, %r840, -2139062144;
xor.b32 %r841, %r255, %r821;
or.b32 %r842, %r841, -2139062144;
and.b32 %r843, %r255, 2139062143;
and.b32 %r844, %r821, -2139062144;
sub.s32 %r845, %r842, %r843;
xor.b32 %r846, %r844, %r845;
xor.b32 %r262, %r846, -2139062144;

dp4a.s32.s32 %r257, %r258, %r112, %r249;


dp4a.s32.s32 %r261, %r262, %r116, %r257;

add.s64 %rd126, %rd104, %rd8;
ld.global.nc.u8 %rs92, [%rd126];
cvt.u32.u16 %r847, %rs92;
and.b32 %r848, %r847, 255;
shr.u32 %r849, %r848, %r14;
shl.b32 %r850, %r849, 1;
and.b32 %r851, %r850, 30;
or.b32 %r852, %r851, 1;
mul.lo.s32 %r853, %r852, %r261;
add.s64 %rd127, %rd105, %rd8;
ld.global.nc.u16 %rs4, [%rd127];

{ cvt.f32.f16 %f31, %rs4;}



{.reg .f16 low,high;
mov.b32 {low,high},%r118;
cvt.f32.f16 %f32, low;}


mul.ftz.f32 %f39, %f31, %f32;
cvt.rn.f32.s32 %f40, %r853;
fma.rn.ftz.f32 %f105, %f39, %f40, %f105;
add.s32 %r937, %r937, 128;
add.s64 %rd130, %rd130, 1760;
add.s32 %r938, %r938, 16;
setp.lt.s32 %p3, %r938, %r5;
@%p3 bra $L__BB307_5;

$L__BB307_6:
setp.eq.s32 %p4, %r1, 0;
@%p4 bra $L__BB307_8;

shl.b32 %r854, %r1, 9;
mov.u32 %r855, _ZZ13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r856, %r855, %r854;
shl.b32 %r857, %r2, 2;
add.s32 %r858, %r856, %r857;
st.shared.f32 [%r858+-512], %f108;
st.shared.f32 [%r858+-384], %f107;
st.shared.f32 [%r858+-256], %f106;
st.shared.f32 [%r858+-128], %f105;

$L__BB307_8:
setp.ne.s32 %p5, %r1, 0;
bar.sync 0;
@%p5 bra $L__BB307_17;

mov.u32 %r859, %ctaid.y;
mad.lo.s32 %r860, %r859, %r32, %r4;
mad.lo.s32 %r861, %r12, %r38, %r860;
shl.b32 %r862, %r2, 2;
mov.u32 %r863, _ZZ13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r20, %r863, %r862;
mov.u32 %r864, 2;
ld.shared.f32 %f41, [%r20];
add.ftz.f32 %f42, %f41, %f108;
ld.shared.f32 %f43, [%r20+512];
add.ftz.f32 %f44, %f43, %f42;
ld.shared.f32 %f45, [%r20+1024];
add.ftz.f32 %f46, %f45, %f44;
mov.b32 %r865, %f46;
mov.u32 %r866, 31;
mov.u32 %r867, 16;
mov.u32 %r868, -1;
shfl.sync.bfly.b32 %r869|%p6, %r865, %r867, %r866, %r868;
mov.b32 %f47, %r869;
add.ftz.f32 %f48, %f46, %f47;
mov.b32 %r870, %f48;
mov.u32 %r871, 8;
shfl.sync.bfly.b32 %r872|%p7, %r870, %r871, %r866, %r868;
mov.b32 %f49, %r872;
add.ftz.f32 %f50, %f48, %f49;
mov.b32 %r873, %f50;
mov.u32 %r874, 4;
shfl.sync.bfly.b32 %r875|%p8, %r873, %r874, %r866, %r868;
mov.b32 %f51, %r875;
add.ftz.f32 %f52, %f50, %f51;
mov.b32 %r876, %f52;
shfl.sync.bfly.b32 %r877|%p9, %r876, %r864, %r866, %r868;
mov.b32 %f53, %r877;
add.ftz.f32 %f54, %f52, %f53;
mov.b32 %r878, %f54;
mov.u32 %r879, 1;
shfl.sync.bfly.b32 %r880|%p10, %r878, %r879, %r866, %r868;
mov.b32 %f55, %r880;
add.ftz.f32 %f13, %f54, %f55;
setp.ge.u32 %p11, %r4, %r26;
setp.ne.s32 %p12, %r2, 0;
cvta.to.global.u64 %rd128, %rd18;
mul.wide.u32 %rd129, %r861, 4;
add.s64 %rd14, %rd128, %rd129;
or.pred %p13, %p12, %p11;
@%p13 bra $L__BB307_11;

st.global.f32 [%rd14], %f13;

$L__BB307_11:
ld.shared.f32 %f56, [%r20+128];
add.ftz.f32 %f57, %f56, %f107;
ld.shared.f32 %f58, [%r20+640];
add.ftz.f32 %f59, %f58, %f57;
ld.shared.f32 %f60, [%r20+1152];
add.ftz.f32 %f61, %f60, %f59;
mov.b32 %r881, %f61;
shfl.sync.bfly.b32 %r885|%p14, %r881, %r867, %r866, %r868;
mov.b32 %f62, %r885;
add.ftz.f32 %f63, %f61, %f62;
mov.b32 %r886, %f63;
shfl.sync.bfly.b32 %r888|%p15, %r886, %r871, %r866, %r868;
mov.b32 %f64, %r888;
add.ftz.f32 %f65, %f63, %f64;
mov.b32 %r889, %f65;
shfl.sync.bfly.b32 %r891|%p16, %r889, %r874, %r866, %r868;
mov.b32 %f66, %r891;
add.ftz.f32 %f67, %f65, %f66;
mov.b32 %r892, %f67;
shfl.sync.bfly.b32 %r894|%p17, %r892, %r864, %r866, %r868;
mov.b32 %f68, %r894;
add.ftz.f32 %f69, %f67, %f68;
mov.b32 %r895, %f69;
shfl.sync.bfly.b32 %r897|%p18, %r895, %r879, %r866, %r868;
mov.b32 %f70, %r897;
add.ftz.f32 %f14, %f69, %f70;
add.s32 %r898, %r4, 1;
setp.ge.u32 %p19, %r898, %r26;
setp.ne.s32 %p20, %r2, 1;
or.pred %p21, %p20, %p19;
@%p21 bra $L__BB307_13;

st.global.f32 [%rd14+4], %f14;

$L__BB307_13:
ld.shared.f32 %f71, [%r20+256];
add.ftz.f32 %f72, %f71, %f106;
ld.shared.f32 %f73, [%r20+768];
add.ftz.f32 %f74, %f73, %f72;
ld.shared.f32 %f75, [%r20+1280];
add.ftz.f32 %f76, %f75, %f74;
mov.b32 %r899, %f76;
mov.u32 %r900, 31;
mov.u32 %r901, 16;
mov.u32 %r902, -1;
shfl.sync.bfly.b32 %r903|%p22, %r899, %r901, %r900, %r902;
mov.b32 %f77, %r903;
add.ftz.f32 %f78, %f76, %f77;
mov.b32 %r904, %f78;
mov.u32 %r905, 8;
shfl.sync.bfly.b32 %r906|%p23, %r904, %r905, %r900, %r902;
mov.b32 %f79, %r906;
add.ftz.f32 %f80, %f78, %f79;
mov.b32 %r907, %f80;
mov.u32 %r908, 4;
shfl.sync.bfly.b32 %r909|%p24, %r907, %r908, %r900, %r902;
mov.b32 %f81, %r909;
add.ftz.f32 %f82, %f80, %f81;
mov.b32 %r910, %f82;
mov.u32 %r911, 2;
shfl.sync.bfly.b32 %r912|%p25, %r910, %r911, %r900, %r902;
mov.b32 %f83, %r912;
add.ftz.f32 %f84, %f82, %f83;
mov.b32 %r913, %f84;
mov.u32 %r914, 1;
shfl.sync.bfly.b32 %r915|%p26, %r913, %r914, %r900, %r902;
mov.b32 %f85, %r915;
add.ftz.f32 %f15, %f84, %f85;
add.s32 %r916, %r4, 2;
setp.ge.u32 %p27, %r916, %r26;
setp.ne.s32 %p28, %r2, 2;
or.pred %p29, %p28, %p27;
@%p29 bra $L__BB307_15;

st.global.f32 [%rd14+8], %f15;

$L__BB307_15:
ld.shared.f32 %f86, [%r20+384];
add.ftz.f32 %f87, %f86, %f105;
ld.shared.f32 %f88, [%r20+896];
add.ftz.f32 %f89, %f88, %f87;
ld.shared.f32 %f90, [%r20+1408];
add.ftz.f32 %f91, %f90, %f89;
mov.b32 %r917, %f91;
shfl.sync.bfly.b32 %r921|%p30, %r917, %r901, %r900, %r902;
mov.b32 %f92, %r921;
add.ftz.f32 %f93, %f91, %f92;
mov.b32 %r922, %f93;
shfl.sync.bfly.b32 %r924|%p31, %r922, %r905, %r900, %r902;
mov.b32 %f94, %r924;
add.ftz.f32 %f95, %f93, %f94;
mov.b32 %r925, %f95;
shfl.sync.bfly.b32 %r927|%p32, %r925, %r908, %r900, %r902;
mov.b32 %f96, %r927;
add.ftz.f32 %f97, %f95, %f96;
mov.b32 %r928, %f97;
shfl.sync.bfly.b32 %r930|%p33, %r928, %r911, %r900, %r902;
mov.b32 %f98, %r930;
add.ftz.f32 %f99, %f97, %f98;
mov.b32 %r931, %f99;
shfl.sync.bfly.b32 %r933|%p34, %r931, %r914, %r900, %r902;
mov.b32 %f100, %r933;
add.ftz.f32 %f16, %f99, %f100;
add.s32 %r934, %r4, 3;
setp.ge.u32 %p35, %r934, %r26;
setp.ne.s32 %p36, %r2, 3;
or.pred %p37, %p36, %p35;
@%p37 bra $L__BB307_17;

st.global.f32 [%rd14+12], %f16;

$L__BB307_17:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<33>;
.reg .b16 %rs<45>;
.reg .f32 %f<121>;
.reg .b32 %r<563>;
.reg .b64 %rd<126>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[384];

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate[384];

ld.param.u64 %rd9, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd10, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd11, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r48, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r49, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r50, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r46, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r45, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
ld.param.u64 %rd14, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+16];
ld.param.u64 %rd13, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+8];
ld.param.u64 %rd12, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3];
mov.u32 %r1, %tid.y;
shl.b32 %r51, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r51, %r2;
mov.u32 %r4, %ctaid.x;
mov.u32 %r5, %ctaid.y;
setp.eq.s64 %p2, %rd11, 0;
@%p2 bra $L__BB308_2;

cvta.to.global.u64 %rd18, %rd11;
mul.wide.u32 %rd19, %r5, 4;
add.s64 %rd20, %rd18, %rd19;
ld.global.nc.u32 %r559, [%rd20];
mul.hi.u32 %r52, %r5, %r35;
add.s32 %r53, %r52, %r5;
shr.u32 %r54, %r53, %r36;
mul.lo.s32 %r55, %r54, %r37;
sub.s32 %r560, %r5, %r55;
bra.uni $L__BB308_3;

$L__BB308_2:
mul.hi.u32 %r56, %r5, %r39;
add.s32 %r57, %r56, %r5;
shr.u32 %r559, %r57, %r40;
mov.u32 %r560, %r5;

$L__BB308_3:
mov.u32 %r11, %ctaid.z;
selp.b32 %r12, %r5, %r559, %p2;
or.b32 %r58, %r2, %r1;
setp.ne.s32 %p4, %r58, 0;
mov.f32 %f114, 0f00000000;
mov.f32 %f110, %f114;
mov.f32 %f111, %f114;
@%p4 bra $L__BB308_8;

setp.eq.s64 %p5, %rd12, 0;
mov.f32 %f110, 0f00000000;
mov.f32 %f111, %f110;
@%p5 bra $L__BB308_6;

cvta.to.global.u64 %rd21, %rd12;
mul.lo.s32 %r59, %r11, %r50;
cvt.u64.u32 %rd22, %r59;
mul.lo.s32 %r60, %r12, %r44;
cvt.u64.u32 %rd23, %r60;
cvt.u64.u32 %rd24, %r2;
cvt.s64.s32 %rd25, %r4;
add.s64 %rd26, %rd24, %rd25;
add.s64 %rd27, %rd26, %rd22;
add.s64 %rd28, %rd27, %rd23;
shl.b64 %rd29, %rd28, 2;
add.s64 %rd30, %rd21, %rd29;
ld.global.f32 %f111, [%rd30];

$L__BB308_6:
setp.eq.s64 %p6, %rd14, 0;
setp.eq.s64 %p7, %rd13, 0;
or.pred %p8, %p6, %p7;
@%p8 bra $L__BB308_8;

cvta.to.global.u64 %rd31, %rd14;
mul.lo.s32 %r61, %r11, %r50;
cvt.u64.u32 %rd32, %r61;
mul.lo.s32 %r62, %r12, %r44;
cvt.u64.u32 %rd33, %r62;
cvt.u64.u32 %rd34, %r2;
cvt.s64.s32 %rd35, %r4;
add.s64 %rd36, %rd34, %rd35;
add.s64 %rd37, %rd36, %rd32;
add.s64 %rd38, %rd37, %rd33;
shl.b64 %rd39, %rd38, 2;
add.s64 %rd40, %rd31, %rd39;
ld.global.f32 %f110, [%rd40];

$L__BB308_8:
shr.s32 %r63, %r3, 31;
shr.u32 %r64, %r63, 29;
add.s32 %r65, %r3, %r64;
shr.s32 %r562, %r65, 3;
shr.u32 %r66, %r34, 8;
setp.ge.s32 %p9, %r562, %r66;
mov.f32 %f116, %f114;
@%p9 bra $L__BB308_13;

and.b32 %r70, %r65, -8;
sub.s32 %r14, %r3, %r70;
shl.b32 %r71, %r14, 1;
shr.u32 %r72, %r14, 31;
add.s32 %r73, %r14, %r72;
shr.s32 %r15, %r73, 1;
shl.b32 %r561, %r562, 3;
mul.hi.u32 %r74, %r11, %r45;
add.s32 %r75, %r74, %r11;
shr.u32 %r76, %r75, %r46;
mad.lo.s32 %r77, %r76, %r48, %r562;
mad.lo.s32 %r78, %r559, %r42, %r77;
mad.lo.s32 %r79, %r4, %r38, %r78;
mul.wide.s32 %rd125, %r79, 110;
not.b32 %r80, %r71;
mul.wide.s32 %rd41, %r80, 2;
mov.u64 %rd42, 74;
sub.s64 %rd43, %rd42, %rd41;
cvta.to.global.u64 %rd44, %rd9;
add.s64 %rd5, %rd44, %rd43;
mul.lo.s32 %r81, %r560, %r43;
cvt.u64.u32 %rd45, %r81;
mul.lo.s32 %r82, %r11, %r49;
cvt.u64.u32 %rd46, %r82;
add.s64 %rd6, %rd45, %rd46;
mov.f32 %f116, 0f00000000;
cvta.to.global.u64 %rd47, %rd10;
mov.f32 %f114, %f116;

$L__BB308_10:
cvt.u64.u32 %rd48, %r561;
add.s64 %rd49, %rd6, %rd48;
shl.b32 %r132, %r14, 2;
mul.wide.s32 %rd50, %r132, 2;
add.s64 %rd52, %rd44, %rd50;
add.s64 %rd53, %rd52, %rd125;
ld.global.nc.u16 %rs2, [%rd53+4];
ld.global.nc.u16 %rs3, [%rd53+6];
ld.global.nc.u16 %rs4, [%rd53+8];
cvt.s64.s32 %rd54, %r14;
add.s64 %rd55, %rd44, %rd54;
add.s64 %rd56, %rd55, %rd125;
ld.global.nc.u8 %rs5, [%rd56+66];
mul.wide.s32 %rd57, %r71, 2;
add.s64 %rd58, %rd44, %rd57;
add.s64 %rd59, %rd58, %rd125;
add.s64 %rd60, %rd5, %rd125;
ld.global.nc.u16 %rs6, [%rd59+74];
ld.global.nc.u16 %rs7, [%rd60];
add.s64 %rd61, %rd49, %rd54;
mul.lo.s64 %rd62, %rd61, 36;
ld.global.nc.u16 %rs8, [%rd53+2];
and.b16 %rs9, %rs8, 255;
cvt.u32.u16 %r134, %rs9;
shl.b16 %rs10, %rs5, 8;
cvt.u32.u16 %r135, %rs10;
and.b32 %r136, %r135, 256;
or.b32 %r137, %r136, %r134;
mul.wide.u32 %rd63, %r137, 4;
mov.u64 %rd64, iq3s_grid;
add.s64 %rd65, %rd64, %rd63;
ld.global.nc.u32 %r138, [%rd65];
shr.u16 %rs11, %rs8, 8;
cvt.u32.u16 %r139, %rs11;
mul.wide.u16 %r140, %rs5, 128;
and.b32 %r141, %r140, 256;
or.b32 %r142, %r141, %r139;
mul.wide.u32 %rd66, %r142, 4;
add.s64 %rd67, %rd64, %rd66;
ld.global.nc.u32 %r143, [%rd67];
and.b16 %rs13, %rs6, 255;
mul.wide.u16 %r144, %rs13, 128;
cvt.u32.u16 %r145, %rs6;
shl.b32 %r146, %r145, 21;
or.b32 %r147, %r144, %r146;
and.b32 %r148, %r147, 25166208;
or.b32 %r149, %r148, -2139062144;
add.s32 %r150, %r149, -16843009;
or.b32 %r83, %r150, %r148;

prmt.b32 %r83, %r83, 0, 0xba98;

mul.wide.u16 %r151, %rs13, 8;
shl.b32 %r152, %r145, 17;
or.b32 %r153, %r151, %r152;
and.b32 %r154, %r153, 25166208;
or.b32 %r155, %r154, -2139062144;
add.s32 %r156, %r155, -16843009;
or.b32 %r85, %r156, %r154;

prmt.b32 %r85, %r85, 0, 0xba98;

xor.b32 %r157, %r83, %r138;
or.b32 %r158, %r157, -2139062144;
and.b32 %r159, %r83, 2139062143;
and.b32 %r160, %r138, -2139062144;
sub.s32 %r161, %r158, %r159;
xor.b32 %r162, %r160, %r161;
xor.b32 %r88, %r162, -2139062144;
xor.b32 %r163, %r85, %r143;
or.b32 %r164, %r163, -2139062144;
and.b32 %r165, %r85, 2139062143;
and.b32 %r166, %r143, -2139062144;
sub.s32 %r167, %r164, %r165;
xor.b32 %r168, %r166, %r167;
xor.b32 %r92, %r168, -2139062144;
add.s64 %rd68, %rd47, %rd62;
ld.global.nc.u32 %r19, [%rd68+4];
ld.global.nc.u32 %r20, [%rd68+8];
mov.u32 %r90, 0;

dp4a.s32.s32 %r87, %r88, %r19, %r90;


dp4a.s32.s32 %r91, %r92, %r20, %r87;

and.b16 %rs14, %rs2, 255;
cvt.u32.u16 %r169, %rs14;
mul.wide.u16 %r170, %rs5, 64;
and.b32 %r171, %r170, 256;
or.b32 %r172, %r171, %r169;
mul.wide.u32 %rd69, %r172, 4;
add.s64 %rd70, %rd64, %rd69;
ld.global.nc.u32 %r173, [%rd70];
shr.u16 %rs15, %rs2, 8;
cvt.u32.u16 %r174, %rs15;
mul.wide.u16 %r175, %rs5, 32;
and.b32 %r176, %r175, 256;
or.b32 %r177, %r176, %r174;
mul.wide.u32 %rd71, %r177, 4;
add.s64 %rd72, %rd64, %rd71;
ld.global.nc.u32 %r178, [%rd72];
shr.u16 %rs16, %rs6, 8;
mul.wide.u16 %r179, %rs16, 128;
cvt.u32.u16 %r180, %rs16;
shl.b32 %r181, %r180, 21;
or.b32 %r182, %r179, %r181;
and.b32 %r183, %r182, 25166208;
or.b32 %r184, %r183, -2139062144;
add.s32 %r185, %r184, -16843009;
or.b32 %r95, %r185, %r183;

prmt.b32 %r95, %r95, 0, 0xba98;

mul.wide.u16 %r186, %rs16, 8;
shl.b32 %r187, %r180, 17;
or.b32 %r188, %r186, %r187;
and.b32 %r189, %r188, 25166208;
or.b32 %r190, %r189, -2139062144;
add.s32 %r191, %r190, -16843009;
or.b32 %r97, %r191, %r189;

prmt.b32 %r97, %r97, 0, 0xba98;

xor.b32 %r192, %r95, %r173;
or.b32 %r193, %r192, -2139062144;
and.b32 %r194, %r95, 2139062143;
and.b32 %r195, %r173, -2139062144;
sub.s32 %r196, %r193, %r194;
xor.b32 %r197, %r195, %r196;
xor.b32 %r100, %r197, -2139062144;
xor.b32 %r198, %r97, %r178;
or.b32 %r199, %r198, -2139062144;
and.b32 %r200, %r97, 2139062143;
and.b32 %r201, %r178, -2139062144;
sub.s32 %r202, %r199, %r200;
xor.b32 %r203, %r201, %r202;
xor.b32 %r104, %r203, -2139062144;
ld.global.nc.u32 %r21, [%rd68+12];
ld.global.nc.u32 %r22, [%rd68+16];

dp4a.s32.s32 %r99, %r100, %r21, %r91;


dp4a.s32.s32 %r103, %r104, %r22, %r99;

and.b16 %rs17, %rs3, 255;
cvt.u32.u16 %r204, %rs17;
mul.wide.u16 %r205, %rs5, 16;
and.b32 %r206, %r205, 256;
or.b32 %r207, %r206, %r204;
mul.wide.u32 %rd73, %r207, 4;
add.s64 %rd74, %rd64, %rd73;
ld.global.nc.u32 %r208, [%rd74];
shr.u16 %rs18, %rs3, 8;
cvt.u32.u16 %r209, %rs18;
mul.wide.u16 %r210, %rs5, 8;
and.b32 %r211, %r210, 256;
or.b32 %r212, %r211, %r209;
mul.wide.u32 %rd75, %r212, 4;
add.s64 %rd76, %rd64, %rd75;
ld.global.nc.u32 %r213, [%rd76];
and.b16 %rs19, %rs7, 255;
mul.wide.u16 %r214, %rs19, 128;
cvt.u32.u16 %r215, %rs7;
shl.b32 %r216, %r215, 21;
or.b32 %r217, %r214, %r216;
and.b32 %r218, %r217, 25166208;
or.b32 %r219, %r218, -2139062144;
add.s32 %r220, %r219, -16843009;
or.b32 %r107, %r220, %r218;

prmt.b32 %r107, %r107, 0, 0xba98;

mul.wide.u16 %r221, %rs19, 8;
shl.b32 %r222, %r215, 17;
or.b32 %r223, %r221, %r222;
and.b32 %r224, %r223, 25166208;
or.b32 %r225, %r224, -2139062144;
add.s32 %r226, %r225, -16843009;
or.b32 %r109, %r226, %r224;

prmt.b32 %r109, %r109, 0, 0xba98;

xor.b32 %r227, %r107, %r208;
or.b32 %r228, %r227, -2139062144;
and.b32 %r229, %r107, 2139062143;
and.b32 %r230, %r208, -2139062144;
sub.s32 %r231, %r228, %r229;
xor.b32 %r232, %r230, %r231;
xor.b32 %r112, %r232, -2139062144;
xor.b32 %r233, %r109, %r213;
or.b32 %r234, %r233, -2139062144;
and.b32 %r235, %r109, 2139062143;
and.b32 %r236, %r213, -2139062144;
sub.s32 %r237, %r234, %r235;
xor.b32 %r238, %r236, %r237;
xor.b32 %r116, %r238, -2139062144;
ld.global.nc.u32 %r23, [%rd68+20];
ld.global.nc.u32 %r24, [%rd68+24];

dp4a.s32.s32 %r111, %r112, %r23, %r103;


dp4a.s32.s32 %r115, %r116, %r24, %r111;

and.b16 %rs20, %rs4, 255;
cvt.u32.u16 %r239, %rs20;
mul.wide.u16 %r240, %rs5, 4;
and.b32 %r241, %r240, 256;
or.b32 %r242, %r241, %r239;
mul.wide.u32 %rd77, %r242, 4;
add.s64 %rd78, %rd64, %rd77;
ld.global.nc.u32 %r243, [%rd78];
shr.u16 %rs21, %rs4, 8;
cvt.u32.u16 %r244, %rs21;
mul.wide.u16 %r245, %rs5, 2;
and.b32 %r246, %r245, 256;
or.b32 %r247, %r246, %r244;
mul.wide.u32 %rd79, %r247, 4;
add.s64 %rd80, %rd64, %rd79;
ld.global.nc.u32 %r248, [%rd80];
shr.u16 %rs22, %rs7, 8;
mul.wide.u16 %r249, %rs22, 128;
cvt.u32.u16 %r250, %rs22;
shl.b32 %r251, %r250, 21;
or.b32 %r252, %r249, %r251;
and.b32 %r253, %r252, 25166208;
or.b32 %r254, %r253, -2139062144;
add.s32 %r255, %r254, -16843009;
or.b32 %r119, %r255, %r253;

prmt.b32 %r119, %r119, 0, 0xba98;

mul.wide.u16 %r256, %rs22, 8;
shl.b32 %r257, %r250, 17;
or.b32 %r258, %r256, %r257;
and.b32 %r259, %r258, 25166208;
or.b32 %r260, %r259, -2139062144;
add.s32 %r261, %r260, -16843009;
or.b32 %r121, %r261, %r259;

prmt.b32 %r121, %r121, 0, 0xba98;

xor.b32 %r262, %r119, %r243;
or.b32 %r263, %r262, -2139062144;
and.b32 %r264, %r119, 2139062143;
and.b32 %r265, %r243, -2139062144;
sub.s32 %r266, %r263, %r264;
xor.b32 %r267, %r265, %r266;
xor.b32 %r124, %r267, -2139062144;
xor.b32 %r268, %r121, %r248;
or.b32 %r269, %r268, -2139062144;
and.b32 %r270, %r121, 2139062143;
and.b32 %r271, %r248, -2139062144;
sub.s32 %r272, %r269, %r270;
xor.b32 %r273, %r271, %r272;
xor.b32 %r128, %r273, -2139062144;
ld.global.nc.u32 %r25, [%rd68+28];
ld.global.nc.u32 %r26, [%rd68+32];

dp4a.s32.s32 %r123, %r124, %r25, %r115;


dp4a.s32.s32 %r127, %r128, %r26, %r123;

cvt.s64.s32 %rd81, %r15;
add.s64 %rd82, %rd44, %rd81;
add.s64 %rd83, %rd82, %rd125;
ld.global.nc.u8 %rs23, [%rd83+106];
cvt.u32.u16 %r274, %rs23;
and.b32 %r275, %r274, 255;
and.b32 %r276, %r132, 4;
shr.u32 %r277, %r275, %r276;
shl.b32 %r278, %r277, 1;
and.b32 %r279, %r278, 30;
or.b32 %r280, %r279, 1;
mul.lo.s32 %r281, %r280, %r127;
add.s64 %rd84, %rd44, %rd125;
ld.global.nc.u16 %rs1, [%rd84];

{ cvt.f32.f16 %f40, %rs1;}


ld.global.nc.u32 %r27, [%rd68];

{.reg .f16 low,high;
mov.b32 {low,high},%r27;
cvt.f32.f16 %f41, low;}


mul.ftz.f32 %f42, %f40, %f41;
cvt.rn.f32.s32 %f43, %r281;
fma.rn.ftz.f32 %f116, %f42, %f43, %f116;
setp.eq.s64 %p10, %rd13, 0;
@%p10 bra $L__BB308_12;

cvta.to.global.u64 %rd86, %rd13;
add.s64 %rd87, %rd86, %rd50;
add.s64 %rd88, %rd87, %rd125;
ld.global.u16 %rs25, [%rd88+4];
ld.global.u16 %rs26, [%rd88+6];
ld.global.u16 %rs27, [%rd88+8];
add.s64 %rd90, %rd86, %rd54;
add.s64 %rd91, %rd90, %rd125;
add.s64 %rd93, %rd86, %rd57;
mov.u64 %rd94, 74;
add.s64 %rd95, %rd93, %rd125;
sub.s64 %rd97, %rd94, %rd41;
add.s64 %rd98, %rd86, %rd97;
add.s64 %rd99, %rd98, %rd125;
ld.global.u16 %rs28, [%rd95+74];
ld.global.u16 %rs29, [%rd99];
ld.global.u16 %rs30, [%rd88+2];
and.b16 %rs31, %rs30, 255;
cvt.u32.u16 %r334, %rs31;
ld.global.u8 %rs32, [%rd91+66];
shl.b16 %rs33, %rs32, 8;
cvt.u32.u16 %r335, %rs33;
and.b32 %r336, %r335, 256;
or.b32 %r337, %r336, %r334;
mul.wide.u32 %rd100, %r337, 4;
add.s64 %rd102, %rd64, %rd100;
ld.global.nc.u32 %r338, [%rd102];
shr.u16 %rs34, %rs30, 8;
cvt.u32.u16 %r339, %rs34;
mul.wide.u16 %r340, %rs32, 128;
and.b32 %r341, %r340, 256;
or.b32 %r342, %r341, %r339;
mul.wide.u32 %rd103, %r342, 4;
add.s64 %rd104, %rd64, %rd103;
ld.global.nc.u32 %r343, [%rd104];
and.b16 %rs35, %rs28, 255;
mul.wide.u16 %r344, %rs35, 128;
cvt.u32.u16 %r345, %rs28;
shl.b32 %r346, %r345, 21;
or.b32 %r347, %r344, %r346;
and.b32 %r348, %r347, 25166208;
or.b32 %r349, %r348, -2139062144;
add.s32 %r350, %r349, -16843009;
or.b32 %r282, %r350, %r348;

prmt.b32 %r282, %r282, 0, 0xba98;

mul.wide.u16 %r351, %rs35, 8;
shl.b32 %r352, %r345, 17;
or.b32 %r353, %r351, %r352;
and.b32 %r354, %r353, 25166208;
or.b32 %r355, %r354, -2139062144;
add.s32 %r356, %r355, -16843009;
or.b32 %r284, %r356, %r354;

prmt.b32 %r284, %r284, 0, 0xba98;

xor.b32 %r357, %r282, %r338;
or.b32 %r358, %r357, -2139062144;
and.b32 %r359, %r282, 2139062143;
and.b32 %r360, %r338, -2139062144;
sub.s32 %r361, %r358, %r359;
xor.b32 %r362, %r360, %r361;
xor.b32 %r287, %r362, -2139062144;
xor.b32 %r363, %r284, %r343;
or.b32 %r364, %r363, -2139062144;
and.b32 %r365, %r284, 2139062143;
and.b32 %r366, %r343, -2139062144;
sub.s32 %r367, %r364, %r365;
xor.b32 %r368, %r366, %r367;
xor.b32 %r291, %r368, -2139062144;

dp4a.s32.s32 %r286, %r287, %r19, %r90;


dp4a.s32.s32 %r290, %r291, %r20, %r286;

and.b16 %rs36, %rs25, 255;
cvt.u32.u16 %r369, %rs36;
mul.wide.u16 %r370, %rs32, 64;
and.b32 %r371, %r370, 256;
or.b32 %r372, %r371, %r369;
mul.wide.u32 %rd105, %r372, 4;
add.s64 %rd106, %rd64, %rd105;
ld.global.nc.u32 %r373, [%rd106];
shr.u16 %rs37, %rs25, 8;
cvt.u32.u16 %r374, %rs37;
mul.wide.u16 %r375, %rs32, 32;
and.b32 %r376, %r375, 256;
or.b32 %r377, %r376, %r374;
mul.wide.u32 %rd107, %r377, 4;
add.s64 %rd108, %rd64, %rd107;
ld.global.nc.u32 %r378, [%rd108];
shr.u16 %rs38, %rs28, 8;
mul.wide.u16 %r379, %rs38, 128;
cvt.u32.u16 %r380, %rs38;
shl.b32 %r381, %r380, 21;
or.b32 %r382, %r379, %r381;
and.b32 %r383, %r382, 25166208;
or.b32 %r384, %r383, -2139062144;
add.s32 %r385, %r384, -16843009;
or.b32 %r294, %r385, %r383;

prmt.b32 %r294, %r294, 0, 0xba98;

mul.wide.u16 %r386, %rs38, 8;
shl.b32 %r387, %r380, 17;
or.b32 %r388, %r386, %r387;
and.b32 %r389, %r388, 25166208;
or.b32 %r390, %r389, -2139062144;
add.s32 %r391, %r390, -16843009;
or.b32 %r296, %r391, %r389;

prmt.b32 %r296, %r296, 0, 0xba98;

xor.b32 %r392, %r294, %r373;
or.b32 %r393, %r392, -2139062144;
and.b32 %r394, %r294, 2139062143;
and.b32 %r395, %r373, -2139062144;
sub.s32 %r396, %r393, %r394;
xor.b32 %r397, %r395, %r396;
xor.b32 %r299, %r397, -2139062144;
xor.b32 %r398, %r296, %r378;
or.b32 %r399, %r398, -2139062144;
and.b32 %r400, %r296, 2139062143;
and.b32 %r401, %r378, -2139062144;
sub.s32 %r402, %r399, %r400;
xor.b32 %r403, %r401, %r402;
xor.b32 %r303, %r403, -2139062144;

dp4a.s32.s32 %r298, %r299, %r21, %r290;


dp4a.s32.s32 %r302, %r303, %r22, %r298;

and.b16 %rs39, %rs26, 255;
cvt.u32.u16 %r404, %rs39;
mul.wide.u16 %r405, %rs32, 16;
and.b32 %r406, %r405, 256;
or.b32 %r407, %r406, %r404;
mul.wide.u32 %rd109, %r407, 4;
add.s64 %rd110, %rd64, %rd109;
ld.global.nc.u32 %r408, [%rd110];
shr.u16 %rs40, %rs26, 8;
cvt.u32.u16 %r409, %rs40;
mul.wide.u16 %r410, %rs32, 8;
and.b32 %r411, %r410, 256;
or.b32 %r412, %r411, %r409;
mul.wide.u32 %rd111, %r412, 4;
add.s64 %rd112, %rd64, %rd111;
ld.global.nc.u32 %r413, [%rd112];
and.b16 %rs41, %rs29, 255;
mul.wide.u16 %r414, %rs41, 128;
cvt.u32.u16 %r415, %rs29;
shl.b32 %r416, %r415, 21;
or.b32 %r417, %r414, %r416;
and.b32 %r418, %r417, 25166208;
or.b32 %r419, %r418, -2139062144;
add.s32 %r420, %r419, -16843009;
or.b32 %r306, %r420, %r418;

prmt.b32 %r306, %r306, 0, 0xba98;

mul.wide.u16 %r421, %rs41, 8;
shl.b32 %r422, %r415, 17;
or.b32 %r423, %r421, %r422;
and.b32 %r424, %r423, 25166208;
or.b32 %r425, %r424, -2139062144;
add.s32 %r426, %r425, -16843009;
or.b32 %r308, %r426, %r424;

prmt.b32 %r308, %r308, 0, 0xba98;

xor.b32 %r427, %r306, %r408;
or.b32 %r428, %r427, -2139062144;
and.b32 %r429, %r306, 2139062143;
and.b32 %r430, %r408, -2139062144;
sub.s32 %r431, %r428, %r429;
xor.b32 %r432, %r430, %r431;
xor.b32 %r311, %r432, -2139062144;
xor.b32 %r433, %r308, %r413;
or.b32 %r434, %r433, -2139062144;
and.b32 %r435, %r308, 2139062143;
and.b32 %r436, %r413, -2139062144;
sub.s32 %r437, %r434, %r435;
xor.b32 %r438, %r436, %r437;
xor.b32 %r315, %r438, -2139062144;

dp4a.s32.s32 %r310, %r311, %r23, %r302;


dp4a.s32.s32 %r314, %r315, %r24, %r310;

and.b16 %rs42, %rs27, 255;
cvt.u32.u16 %r439, %rs42;
mul.wide.u16 %r440, %rs32, 4;
and.b32 %r441, %r440, 256;
or.b32 %r442, %r441, %r439;
mul.wide.u32 %rd113, %r442, 4;
add.s64 %rd114, %rd64, %rd113;
ld.global.nc.u32 %r443, [%rd114];
shr.u16 %rs43, %rs27, 8;
cvt.u32.u16 %r444, %rs43;
mul.wide.u16 %r445, %rs32, 2;
and.b32 %r446, %r445, 256;
or.b32 %r447, %r446, %r444;
mul.wide.u32 %rd115, %r447, 4;
add.s64 %rd116, %rd64, %rd115;
ld.global.nc.u32 %r448, [%rd116];
shr.u16 %rs44, %rs29, 8;
mul.wide.u16 %r449, %rs44, 128;
cvt.u32.u16 %r450, %rs44;
shl.b32 %r451, %r450, 21;
or.b32 %r452, %r449, %r451;
and.b32 %r453, %r452, 25166208;
or.b32 %r454, %r453, -2139062144;
add.s32 %r455, %r454, -16843009;
or.b32 %r318, %r455, %r453;

prmt.b32 %r318, %r318, 0, 0xba98;

mul.wide.u16 %r456, %rs44, 8;
shl.b32 %r457, %r450, 17;
or.b32 %r458, %r456, %r457;
and.b32 %r459, %r458, 25166208;
or.b32 %r460, %r459, -2139062144;
add.s32 %r461, %r460, -16843009;
or.b32 %r320, %r461, %r459;

prmt.b32 %r320, %r320, 0, 0xba98;

xor.b32 %r462, %r318, %r443;
or.b32 %r463, %r462, -2139062144;
and.b32 %r464, %r318, 2139062143;
and.b32 %r465, %r443, -2139062144;
sub.s32 %r466, %r463, %r464;
xor.b32 %r467, %r465, %r466;
xor.b32 %r323, %r467, -2139062144;
xor.b32 %r468, %r320, %r448;
or.b32 %r469, %r468, -2139062144;
and.b32 %r470, %r320, 2139062143;
and.b32 %r471, %r448, -2139062144;
sub.s32 %r472, %r469, %r470;
xor.b32 %r473, %r471, %r472;
xor.b32 %r327, %r473, -2139062144;

dp4a.s32.s32 %r322, %r323, %r25, %r314;


dp4a.s32.s32 %r326, %r327, %r26, %r322;

add.s64 %rd118, %rd86, %rd81;
add.s64 %rd119, %rd118, %rd125;
ld.global.u8 %r474, [%rd119+106];
shr.u32 %r476, %r474, %r276;
shl.b32 %r477, %r476, 1;
and.b32 %r478, %r477, 30;
or.b32 %r479, %r478, 1;
mul.lo.s32 %r480, %r479, %r326;
add.s64 %rd120, %rd86, %rd125;
ld.global.u16 %rs24, [%rd120];

{ cvt.f32.f16 %f44, %rs24;}



{.reg .f16 low,high;
mov.b32 {low,high},%r27;
cvt.f32.f16 %f45, low;}


mul.ftz.f32 %f46, %f44, %f45;
cvt.rn.f32.s32 %f47, %r480;
fma.rn.ftz.f32 %f114, %f46, %f47, %f114;

$L__BB308_12:
add.s32 %r561, %r561, 128;
add.s64 %rd125, %rd125, 1760;
add.s32 %r562, %r562, 16;
setp.lt.s32 %p11, %r562, %r66;
@%p11 bra $L__BB308_10;

$L__BB308_13:
mov.u32 %r544, %tid.y;
setp.eq.s32 %p12, %r544, 0;
@%p12 bra $L__BB308_16;

mov.u32 %r546, %tid.y;
mov.u32 %r545, %tid.x;
setp.eq.s64 %p13, %rd13, 0;
add.s32 %r30, %r546, -1;
shl.b32 %r485, %r546, 7;
mov.u32 %r486, _ZZ13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r487, %r486, %r485;
shl.b32 %r488, %r545, 2;
add.s32 %r489, %r487, %r488;
st.shared.f32 [%r489+-128], %f116;
@%p13 bra $L__BB308_16;

shl.b32 %r491, %r30, 7;
mov.u32 %r492, _ZZ13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r493, %r492, %r491;
add.s32 %r495, %r493, %r488;
st.shared.f32 [%r495], %f114;

$L__BB308_16:
mov.u32 %r547, %tid.y;
setp.ne.s32 %p14, %r547, 0;
bar.sync 0;
@%p14 bra $L__BB308_35;

mov.u32 %r548, %tid.x;
setp.ne.s64 %p15, %rd13, 0;
shl.b32 %r498, %r548, 2;
mov.u32 %r499, _ZZ13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r31, %r499, %r498;
ld.shared.f32 %f48, [%r31];
add.ftz.f32 %f13, %f116, %f48;
@%p15 bra $L__BB308_19;
bra.uni $L__BB308_18;

$L__BB308_19:
mov.u32 %r502, _ZZ13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE15tmp_shared_gate;
add.s32 %r503, %r502, %r498;
ld.shared.f32 %f52, [%r503];
add.ftz.f32 %f53, %f114, %f52;
ld.shared.f32 %f54, [%r31+128];
add.ftz.f32 %f55, %f13, %f54;
ld.shared.f32 %f56, [%r503+128];
add.ftz.f32 %f57, %f53, %f56;
ld.shared.f32 %f58, [%r31+256];
add.ftz.f32 %f118, %f55, %f58;
ld.shared.f32 %f59, [%r503+256];
add.ftz.f32 %f114, %f57, %f59;
bra.uni $L__BB308_20;

$L__BB308_18:
ld.shared.f32 %f49, [%r31+128];
add.ftz.f32 %f50, %f13, %f49;
ld.shared.f32 %f51, [%r31+256];
add.ftz.f32 %f118, %f50, %f51;

$L__BB308_20:
mov.b32 %r504, %f118;
mov.u32 %r505, 31;
mov.u32 %r506, 16;
mov.u32 %r507, -1;
shfl.sync.bfly.b32 %r508|%p16, %r504, %r506, %r505, %r507;
mov.b32 %f60, %r508;
add.ftz.f32 %f61, %f118, %f60;
mov.b32 %r509, %f61;
mov.u32 %r510, 8;
shfl.sync.bfly.b32 %r511|%p17, %r509, %r510, %r505, %r507;
mov.b32 %f62, %r511;
add.ftz.f32 %f63, %f61, %f62;
mov.b32 %r512, %f63;
mov.u32 %r513, 4;
shfl.sync.bfly.b32 %r514|%p18, %r512, %r513, %r505, %r507;
mov.b32 %f64, %r514;
add.ftz.f32 %f65, %f63, %f64;
mov.b32 %r515, %f65;
mov.u32 %r516, 2;
shfl.sync.bfly.b32 %r517|%p19, %r515, %r516, %r505, %r507;
mov.b32 %f66, %r517;
add.ftz.f32 %f19, %f65, %f66;
mov.b32 %r518, %f19;
mov.u32 %r519, 1;
shfl.sync.bfly.b32 %r32|%p1, %r518, %r519, %r505, %r507;
setp.eq.s64 %p20, %rd13, 0;
@%p20 bra $L__BB308_22;

mov.b32 %r520, %f114;
shfl.sync.bfly.b32 %r524|%p21, %r520, %r506, %r505, %r507;
mov.b32 %f67, %r524;
add.ftz.f32 %f68, %f114, %f67;
mov.b32 %r525, %f68;
shfl.sync.bfly.b32 %r527|%p22, %r525, %r510, %r505, %r507;
mov.b32 %f69, %r527;
add.ftz.f32 %f70, %f68, %f69;
mov.b32 %r528, %f70;
shfl.sync.bfly.b32 %r530|%p23, %r528, %r513, %r505, %r507;
mov.b32 %f71, %r530;
add.ftz.f32 %f72, %f70, %f71;
mov.b32 %r531, %f72;
shfl.sync.bfly.b32 %r533|%p24, %r531, %r516, %r505, %r507;
mov.b32 %f73, %r533;
add.ftz.f32 %f74, %f72, %f73;
mov.b32 %r534, %f74;
shfl.sync.bfly.b32 %r536|%p25, %r534, %r519, %r505, %r507;
mov.b32 %f75, %r536;
add.ftz.f32 %f114, %f74, %f75;

$L__BB308_22:
mov.u32 %r549, %tid.x;
setp.ne.s32 %p26, %r549, 0;
mov.b32 %f76, %r32;
add.ftz.f32 %f22, %f19, %f76;
@%p26 bra $L__BB308_35;

add.ftz.f32 %f120, %f111, %f22;
@%p20 bra $L__BB308_34;

ld.param.u32 %r554, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
add.ftz.f32 %f24, %f110, %f114;
setp.gt.s32 %p28, %r554, 2;
@%p28 bra $L__BB308_28;

ld.param.u32 %r557, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p31, %r557, 1;
@%p31 bra $L__BB308_32;

ld.param.u32 %r558, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p32, %r558, 2;
@%p32 bra $L__BB308_27;
bra.uni $L__BB308_33;

$L__BB308_27:
mul.ftz.f32 %f96, %f24, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f97, %f96;
add.ftz.f32 %f98, %f97, 0f3F800000;
div.approx.ftz.f32 %f99, %f24, %f98;
mul.ftz.f32 %f120, %f120, %f99;
bra.uni $L__BB308_34;

$L__BB308_28:
ld.param.u32 %r555, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.eq.s32 %p29, %r555, 3;
@%p29 bra $L__BB308_31;

ld.param.u32 %r556, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+40];
setp.ne.s32 %p30, %r556, 6;
@%p30 bra $L__BB308_33;

ld.param.f32 %f108, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3+44];
min.ftz.f32 %f77, %f24, %f108;
neg.ftz.f32 %f78, %f108;
min.ftz.f32 %f79, %f120, %f108;
max.ftz.f32 %f80, %f79, %f78;
mul.ftz.f32 %f81, %f77, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f82, %f81;
add.ftz.f32 %f83, %f82, 0f3F800000;
div.approx.ftz.f32 %f84, %f77, %f83;
mul.ftz.f32 %f120, %f80, %f84;
bra.uni $L__BB308_34;

$L__BB308_32:
mul.ftz.f32 %f100, %f24, 0f3F000000;
mul.ftz.f32 %f101, %f24, 0f3D372713;
fma.rn.ftz.f32 %f102, %f24, %f101, 0f3F800000;
mul.ftz.f32 %f103, %f24, 0f3F4C422A;
mul.ftz.f32 %f104, %f103, %f102;
tanh.approx.f32 %f105, %f104;
add.ftz.f32 %f106, %f105, 0f3F800000;
mul.ftz.f32 %f107, %f100, %f106;
mul.ftz.f32 %f120, %f120, %f107;
bra.uni $L__BB308_34;

$L__BB308_31:
mov.f32 %f85, 0f40E00000;
min.ftz.f32 %f86, %f24, %f85;
min.ftz.f32 %f87, %f120, %f85;
mov.f32 %f88, 0fC0E00000;
max.ftz.f32 %f89, %f87, %f88;
mul.ftz.f32 %f90, %f86, 0f3FD9DB23;
mul.ftz.f32 %f91, %f90, 0fBFB8AA3B;
ex2.approx.ftz.f32 %f92, %f91;
add.ftz.f32 %f93, %f92, 0f3F800000;
div.approx.ftz.f32 %f94, %f86, %f93;
add.ftz.f32 %f95, %f89, 0f3F800000;
mul.ftz.f32 %f120, %f95, %f94;
bra.uni $L__BB308_34;

$L__BB308_33:
mul.ftz.f32 %f120, %f120, %f24;

$L__BB308_34:
ld.param.u64 %rd124, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
mov.u32 %r553, %ctaid.x;
ld.param.u32 %r552, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r551, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
mov.u32 %r550, %ctaid.z;
cvta.to.global.u64 %rd121, %rd124;
mov.u32 %r540, %ctaid.y;
mad.lo.s32 %r542, %r540, %r552, %r553;
mad.lo.s32 %r543, %r550, %r551, %r542;
mul.wide.u32 %rd122, %r543, 4;
add.s64 %rd123, %rd121, %rd122;
st.global.f32 [%rd123], %f120;

$L__BB308_35:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<12>;
.reg .b16 %rs<24>;
.reg .f32 %f<28>;
.reg .b32 %r<295>;
.reg .b64 %rd<59>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[384];

ld.param.u64 %rd12, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd13, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd14, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2];
ld.param.u64 %rd15, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+8];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6+4];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6];
mov.u32 %r1, %tid.y;
shl.b32 %r37, %r1, 5;
mov.u32 %r2, %tid.x;
add.s32 %r3, %r37, %r2;
shr.u32 %r4, %r36, 8;
mov.u32 %r292, %ctaid.y;
setp.eq.s64 %p1, %rd14, 0;
@%p1 bra $L__BB309_2;

cvta.to.global.u64 %rd16, %rd14;
mul.wide.u32 %rd17, %r292, 4;
add.s64 %rd18, %rd16, %rd17;
ld.global.nc.u32 %r291, [%rd18];
mul.hi.u32 %r38, %r292, %r20;
add.s32 %r39, %r38, %r292;
shr.u32 %r40, %r39, %r21;
mul.lo.s32 %r41, %r40, %r22;
sub.s32 %r292, %r292, %r41;
bra.uni $L__BB309_3;

$L__BB309_2:
mul.hi.u32 %r42, %r292, %r24;
add.s32 %r43, %r42, %r292;
shr.u32 %r291, %r43, %r25;

$L__BB309_3:
mov.u32 %r11, %ctaid.z;
shr.s32 %r44, %r3, 31;
shr.u32 %r45, %r44, 29;
add.s32 %r46, %r3, %r45;
shr.s32 %r294, %r46, 3;
setp.ge.s32 %p2, %r294, %r4;
mov.f32 %f27, 0f00000000;
@%p2 bra $L__BB309_6;

and.b32 %r50, %r46, -8;
sub.s32 %r51, %r3, %r50;
shr.u32 %r52, %r51, 31;
add.s32 %r53, %r51, %r52;
shr.s32 %r54, %r53, 1;
cvt.s64.s32 %rd19, %r54;
shl.b32 %r55, %r51, 1;
shl.b32 %r56, %r51, 2;
cvt.s64.s32 %rd1, %r51;
and.b32 %r13, %r56, 4;
mul.hi.u32 %r57, %r11, %r30;
add.s32 %r58, %r57, %r11;
shr.u32 %r59, %r58, %r31;
mad.lo.s32 %r60, %r59, %r33, %r294;
mad.lo.s32 %r61, %r291, %r27, %r60;
mov.u32 %r62, %ctaid.x;
mad.lo.s32 %r63, %r62, %r23, %r61;
mul.wide.s32 %rd20, %r63, 110;
cvta.to.global.u64 %rd21, %rd12;
add.s64 %rd58, %rd21, %rd20;
add.s64 %rd3, %rd19, 106;
not.b32 %r64, %r55;
mul.wide.s32 %rd22, %r64, 2;
mov.u64 %rd23, 74;
sub.s64 %rd4, %rd23, %rd22;
mul.wide.s32 %rd24, %r55, 2;
add.s64 %rd5, %rd24, 74;
add.s64 %rd6, %rd1, 66;
shl.b32 %r293, %r294, 3;
mul.wide.s32 %rd25, %r56, 2;
or.b64 %rd7, %rd25, 2;
cvta.to.global.u64 %rd8, %rd13;
mul.lo.s32 %r65, %r292, %r28;
cvt.u64.u32 %rd26, %r65;
mul.lo.s32 %r66, %r11, %r34;
cvt.u64.u32 %rd27, %r66;
add.s64 %rd9, %rd26, %rd27;
mov.f32 %f27, 0f00000000;

$L__BB309_5:
cvt.u64.u32 %rd28, %r293;
add.s64 %rd29, %rd9, %rd28;
add.s64 %rd30, %rd58, %rd7;
ld.global.nc.u16 %rs2, [%rd30+2];
ld.global.nc.u16 %rs3, [%rd30+4];
ld.global.nc.u16 %rs4, [%rd30+6];
add.s64 %rd31, %rd58, %rd6;
ld.global.nc.u8 %rs5, [%rd31];
add.s64 %rd32, %rd58, %rd5;
add.s64 %rd33, %rd58, %rd4;
ld.global.nc.u16 %rs6, [%rd32];
ld.global.nc.u16 %rs7, [%rd33];
add.s64 %rd34, %rd29, %rd1;
mul.lo.s64 %rd35, %rd34, 36;
ld.global.nc.u16 %rs8, [%rd30];
and.b16 %rs9, %rs8, 255;
cvt.u32.u16 %r116, %rs9;
shl.b16 %rs10, %rs5, 8;
cvt.u32.u16 %r117, %rs10;
and.b32 %r118, %r117, 256;
or.b32 %r119, %r118, %r116;
mul.wide.u32 %rd36, %r119, 4;
mov.u64 %rd37, iq3s_grid;
add.s64 %rd38, %rd37, %rd36;
ld.global.nc.u32 %r120, [%rd38];
shr.u16 %rs11, %rs8, 8;
cvt.u32.u16 %r121, %rs11;
mul.wide.u16 %r122, %rs5, 128;
and.b32 %r123, %r122, 256;
or.b32 %r124, %r123, %r121;
mul.wide.u32 %rd39, %r124, 4;
add.s64 %rd40, %rd37, %rd39;
ld.global.nc.u32 %r125, [%rd40];
and.b16 %rs13, %rs6, 255;
mul.wide.u16 %r126, %rs13, 128;
cvt.u32.u16 %r127, %rs6;
shl.b32 %r128, %r127, 21;
or.b32 %r129, %r126, %r128;
and.b32 %r130, %r129, 25166208;
or.b32 %r131, %r130, -2139062144;
add.s32 %r132, %r131, -16843009;
or.b32 %r67, %r132, %r130;

prmt.b32 %r67, %r67, 0, 0xba98;

mul.wide.u16 %r133, %rs13, 8;
shl.b32 %r134, %r127, 17;
or.b32 %r135, %r133, %r134;
and.b32 %r136, %r135, 25166208;
or.b32 %r137, %r136, -2139062144;
add.s32 %r138, %r137, -16843009;
or.b32 %r69, %r138, %r136;

prmt.b32 %r69, %r69, 0, 0xba98;

xor.b32 %r139, %r67, %r120;
or.b32 %r140, %r139, -2139062144;
and.b32 %r141, %r67, 2139062143;
and.b32 %r142, %r120, -2139062144;
sub.s32 %r143, %r140, %r141;
xor.b32 %r144, %r142, %r143;
xor.b32 %r72, %r144, -2139062144;
xor.b32 %r145, %r69, %r125;
or.b32 %r146, %r145, -2139062144;
and.b32 %r147, %r69, 2139062143;
and.b32 %r148, %r125, -2139062144;
sub.s32 %r149, %r146, %r147;
xor.b32 %r150, %r148, %r149;
xor.b32 %r76, %r150, -2139062144;
add.s64 %rd41, %rd8, %rd35;
ld.global.nc.u32 %r73, [%rd41+4];
ld.global.nc.u32 %r77, [%rd41+8];
mov.u32 %r74, 0;

dp4a.s32.s32 %r71, %r72, %r73, %r74;


dp4a.s32.s32 %r75, %r76, %r77, %r71;

and.b16 %rs14, %rs2, 255;
cvt.u32.u16 %r151, %rs14;
mul.wide.u16 %r152, %rs5, 64;
and.b32 %r153, %r152, 256;
or.b32 %r154, %r153, %r151;
mul.wide.u32 %rd42, %r154, 4;
add.s64 %rd43, %rd37, %rd42;
ld.global.nc.u32 %r155, [%rd43];
shr.u16 %rs15, %rs2, 8;
cvt.u32.u16 %r156, %rs15;
mul.wide.u16 %r157, %rs5, 32;
and.b32 %r158, %r157, 256;
or.b32 %r159, %r158, %r156;
mul.wide.u32 %rd44, %r159, 4;
add.s64 %rd45, %rd37, %rd44;
ld.global.nc.u32 %r160, [%rd45];
shr.u16 %rs16, %rs6, 8;
mul.wide.u16 %r161, %rs16, 128;
cvt.u32.u16 %r162, %rs16;
shl.b32 %r163, %r162, 21;
or.b32 %r164, %r161, %r163;
and.b32 %r165, %r164, 25166208;
or.b32 %r166, %r165, -2139062144;
add.s32 %r167, %r166, -16843009;
or.b32 %r79, %r167, %r165;

prmt.b32 %r79, %r79, 0, 0xba98;

mul.wide.u16 %r168, %rs16, 8;
shl.b32 %r169, %r162, 17;
or.b32 %r170, %r168, %r169;
and.b32 %r171, %r170, 25166208;
or.b32 %r172, %r171, -2139062144;
add.s32 %r173, %r172, -16843009;
or.b32 %r81, %r173, %r171;

prmt.b32 %r81, %r81, 0, 0xba98;

xor.b32 %r174, %r79, %r155;
or.b32 %r175, %r174, -2139062144;
and.b32 %r176, %r79, 2139062143;
and.b32 %r177, %r155, -2139062144;
sub.s32 %r178, %r175, %r176;
xor.b32 %r179, %r177, %r178;
xor.b32 %r84, %r179, -2139062144;
xor.b32 %r180, %r81, %r160;
or.b32 %r181, %r180, -2139062144;
and.b32 %r182, %r81, 2139062143;
and.b32 %r183, %r160, -2139062144;
sub.s32 %r184, %r181, %r182;
xor.b32 %r185, %r183, %r184;
xor.b32 %r88, %r185, -2139062144;
ld.global.nc.u32 %r85, [%rd41+12];
ld.global.nc.u32 %r89, [%rd41+16];

dp4a.s32.s32 %r83, %r84, %r85, %r75;


dp4a.s32.s32 %r87, %r88, %r89, %r83;

and.b16 %rs17, %rs3, 255;
cvt.u32.u16 %r186, %rs17;
mul.wide.u16 %r187, %rs5, 16;
and.b32 %r188, %r187, 256;
or.b32 %r189, %r188, %r186;
mul.wide.u32 %rd46, %r189, 4;
add.s64 %rd47, %rd37, %rd46;
ld.global.nc.u32 %r190, [%rd47];
shr.u16 %rs18, %rs3, 8;
cvt.u32.u16 %r191, %rs18;
mul.wide.u16 %r192, %rs5, 8;
and.b32 %r193, %r192, 256;
or.b32 %r194, %r193, %r191;
mul.wide.u32 %rd48, %r194, 4;
add.s64 %rd49, %rd37, %rd48;
ld.global.nc.u32 %r195, [%rd49];
and.b16 %rs19, %rs7, 255;
mul.wide.u16 %r196, %rs19, 128;
cvt.u32.u16 %r197, %rs7;
shl.b32 %r198, %r197, 21;
or.b32 %r199, %r196, %r198;
and.b32 %r200, %r199, 25166208;
or.b32 %r201, %r200, -2139062144;
add.s32 %r202, %r201, -16843009;
or.b32 %r91, %r202, %r200;

prmt.b32 %r91, %r91, 0, 0xba98;

mul.wide.u16 %r203, %rs19, 8;
shl.b32 %r204, %r197, 17;
or.b32 %r205, %r203, %r204;
and.b32 %r206, %r205, 25166208;
or.b32 %r207, %r206, -2139062144;
add.s32 %r208, %r207, -16843009;
or.b32 %r93, %r208, %r206;

prmt.b32 %r93, %r93, 0, 0xba98;

xor.b32 %r209, %r91, %r190;
or.b32 %r210, %r209, -2139062144;
and.b32 %r211, %r91, 2139062143;
and.b32 %r212, %r190, -2139062144;
sub.s32 %r213, %r210, %r211;
xor.b32 %r214, %r212, %r213;
xor.b32 %r96, %r214, -2139062144;
xor.b32 %r215, %r93, %r195;
or.b32 %r216, %r215, -2139062144;
and.b32 %r217, %r93, 2139062143;
and.b32 %r218, %r195, -2139062144;
sub.s32 %r219, %r216, %r217;
xor.b32 %r220, %r218, %r219;
xor.b32 %r100, %r220, -2139062144;
ld.global.nc.u32 %r97, [%rd41+20];
ld.global.nc.u32 %r101, [%rd41+24];

dp4a.s32.s32 %r95, %r96, %r97, %r87;


dp4a.s32.s32 %r99, %r100, %r101, %r95;

and.b16 %rs20, %rs4, 255;
cvt.u32.u16 %r221, %rs20;
mul.wide.u16 %r222, %rs5, 4;
and.b32 %r223, %r222, 256;
or.b32 %r224, %r223, %r221;
mul.wide.u32 %rd50, %r224, 4;
add.s64 %rd51, %rd37, %rd50;
ld.global.nc.u32 %r225, [%rd51];
shr.u16 %rs21, %rs4, 8;
cvt.u32.u16 %r226, %rs21;
mul.wide.u16 %r227, %rs5, 2;
and.b32 %r228, %r227, 256;
or.b32 %r229, %r228, %r226;
mul.wide.u32 %rd52, %r229, 4;
add.s64 %rd53, %rd37, %rd52;
ld.global.nc.u32 %r230, [%rd53];
shr.u16 %rs22, %rs7, 8;
mul.wide.u16 %r231, %rs22, 128;
cvt.u32.u16 %r232, %rs22;
shl.b32 %r233, %r232, 21;
or.b32 %r234, %r231, %r233;
and.b32 %r235, %r234, 25166208;
or.b32 %r236, %r235, -2139062144;
add.s32 %r237, %r236, -16843009;
or.b32 %r103, %r237, %r235;

prmt.b32 %r103, %r103, 0, 0xba98;

mul.wide.u16 %r238, %rs22, 8;
shl.b32 %r239, %r232, 17;
or.b32 %r240, %r238, %r239;
and.b32 %r241, %r240, 25166208;
or.b32 %r242, %r241, -2139062144;
add.s32 %r243, %r242, -16843009;
or.b32 %r105, %r243, %r241;

prmt.b32 %r105, %r105, 0, 0xba98;

xor.b32 %r244, %r103, %r225;
or.b32 %r245, %r244, -2139062144;
and.b32 %r246, %r103, 2139062143;
and.b32 %r247, %r225, -2139062144;
sub.s32 %r248, %r245, %r246;
xor.b32 %r249, %r247, %r248;
xor.b32 %r108, %r249, -2139062144;
xor.b32 %r250, %r105, %r230;
or.b32 %r251, %r250, -2139062144;
and.b32 %r252, %r105, 2139062143;
and.b32 %r253, %r230, -2139062144;
sub.s32 %r254, %r251, %r252;
xor.b32 %r255, %r253, %r254;
xor.b32 %r112, %r255, -2139062144;
ld.global.nc.u32 %r109, [%rd41+28];
ld.global.nc.u32 %r113, [%rd41+32];

dp4a.s32.s32 %r107, %r108, %r109, %r99;


dp4a.s32.s32 %r111, %r112, %r113, %r107;

add.s64 %rd54, %rd58, %rd3;
ld.global.nc.u8 %rs23, [%rd54];
cvt.u32.u16 %r256, %rs23;
and.b32 %r257, %r256, 255;
shr.u32 %r258, %r257, %r13;
shl.b32 %r259, %r258, 1;
and.b32 %r260, %r259, 30;
or.b32 %r261, %r260, 1;
mul.lo.s32 %r262, %r261, %r111;
ld.global.nc.u16 %rs1, [%rd58];

{ cvt.f32.f16 %f7, %rs1;}


ld.global.nc.u32 %r115, [%rd41];

{.reg .f16 low,high;
mov.b32 {low,high},%r115;
cvt.f32.f16 %f8, low;}


mul.ftz.f32 %f9, %f7, %f8;
cvt.rn.f32.s32 %f10, %r262;
fma.rn.ftz.f32 %f27, %f9, %f10, %f27;
add.s64 %rd58, %rd58, 1760;
add.s32 %r293, %r293, 128;
add.s32 %r294, %r294, 16;
setp.lt.s32 %p3, %r294, %r4;
@%p3 bra $L__BB309_5;

$L__BB309_6:
setp.eq.s32 %p4, %r1, 0;
@%p4 bra $L__BB309_8;

shl.b32 %r263, %r1, 7;
mov.u32 %r264, _ZZ13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r265, %r264, %r263;
shl.b32 %r266, %r2, 2;
add.s32 %r267, %r265, %r266;
st.shared.f32 [%r267+-128], %f27;

$L__BB309_8:
setp.ne.s32 %p5, %r1, 0;
bar.sync 0;
@%p5 bra $L__BB309_11;

mov.u32 %r268, %ctaid.y;
mov.u32 %r269, %ctaid.x;
mad.lo.s32 %r270, %r268, %r29, %r269;
mad.lo.s32 %r19, %r11, %r35, %r270;
shl.b32 %r271, %r2, 2;
mov.u32 %r272, _ZZ13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r273, %r272, %r271;
mov.u32 %r274, 2;
ld.shared.f32 %f11, [%r273];
add.ftz.f32 %f12, %f27, %f11;
ld.shared.f32 %f13, [%r273+128];
add.ftz.f32 %f14, %f12, %f13;
ld.shared.f32 %f15, [%r273+256];
add.ftz.f32 %f16, %f14, %f15;
mov.b32 %r275, %f16;
mov.u32 %r276, 31;
mov.u32 %r277, 16;
mov.u32 %r278, -1;
shfl.sync.bfly.b32 %r279|%p6, %r275, %r277, %r276, %r278;
mov.b32 %f17, %r279;
add.ftz.f32 %f18, %f16, %f17;
mov.b32 %r280, %f18;
mov.u32 %r281, 8;
shfl.sync.bfly.b32 %r282|%p7, %r280, %r281, %r276, %r278;
mov.b32 %f19, %r282;
add.ftz.f32 %f20, %f18, %f19;
mov.b32 %r283, %f20;
mov.u32 %r284, 4;
shfl.sync.bfly.b32 %r285|%p8, %r283, %r284, %r276, %r278;
mov.b32 %f21, %r285;
add.ftz.f32 %f22, %f20, %f21;
mov.b32 %r286, %f22;
shfl.sync.bfly.b32 %r287|%p9, %r286, %r274, %r276, %r278;
mov.b32 %f23, %r287;
add.ftz.f32 %f24, %f22, %f23;
mov.b32 %r288, %f24;
mov.u32 %r289, 1;
shfl.sync.bfly.b32 %r290|%p10, %r288, %r289, %r276, %r278;
mov.b32 %f25, %r290;
add.ftz.f32 %f4, %f24, %f25;
setp.ne.s32 %p11, %r2, 0;
@%p11 bra $L__BB309_11;

cvta.to.global.u64 %rd55, %rd15;
mul.wide.u32 %rd56, %r19, 4;
add.s64 %rd57, %rd55, %rd56;
st.global.f32 [%rd57], %f4;

$L__BB309_11:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type21ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type21ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type21ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type21ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type21ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type21ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type21ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type21ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type21ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<35>;
.reg .b16 %rs<49>;
.reg .f32 %f<109>;
.reg .b32 %r<719>;
.reg .b64 %rd<108>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type21ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

ld.param.u64 %rd21, [_Z13mul_mat_vec_qIL9ggml_type21ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd22, [_Z13mul_mat_vec_qIL9ggml_type21ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd23, [_Z13mul_mat_vec_qIL9ggml_type21ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type21ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r15, [_Z13mul_mat_vec_qIL9ggml_type21ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r16, [_Z13mul_mat_vec_qIL9ggml_type21ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r17, [_Z13mul_mat_vec_qIL9ggml_type21ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type21ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type21ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type21ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type21ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type21ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type21ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type21ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type21ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r19, [_Z13mul_mat_vec_qIL9ggml_type21ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r18, [_Z13mul_mat_vec_qIL9ggml_type21ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r31, %tid.y;
shl.b32 %r32, %r31, 5;
mov.u32 %r33, %tid.x;
add.s32 %r1, %r32, %r33;
shr.s32 %r34, %r1, 31;
shr.u32 %r35, %r34, 29;
add.s32 %r36, %r1, %r35;
shr.s32 %r718, %r36, 3;
shr.u32 %r3, %r30, 8;
setp.ge.s32 %p3, %r718, %r3;
mov.f32 %f105, 0f00000000;
mov.f32 %f106, %f105;
mov.f32 %f107, %f105;
mov.f32 %f108, %f105;
@%p3 bra $L__BB310_3;

mov.u32 %r37, %ctaid.x;
shl.b32 %r38, %r37, 1;
mov.u32 %r39, %ctaid.y;
mov.u32 %r40, %ctaid.z;
mul.lo.s32 %r41, %r39, %r22;
cvt.u64.u32 %rd24, %r41;
mul.lo.s32 %r42, %r40, %r28;
cvt.u64.u32 %rd25, %r42;
add.s64 %rd1, %rd25, %rd24;
mul.hi.u32 %r43, %r40, %r24;
add.s32 %r44, %r43, %r40;
shr.u32 %r45, %r44, %r25;
mul.lo.s32 %r46, %r45, %r27;
mul.hi.u32 %r47, %r39, %r18;
add.s32 %r48, %r47, %r39;
shr.u32 %r49, %r48, %r19;
mul.lo.s32 %r50, %r49, %r21;
cvta.to.global.u64 %rd107, %rd21;
and.b32 %r54, %r36, -8;
sub.s32 %r55, %r1, %r54;
shl.b32 %r56, %r55, 1;
shr.u32 %r57, %r55, 31;
add.s32 %r58, %r55, %r57;
shr.s32 %r59, %r58, 1;
cvt.s64.s32 %rd26, %r59;
shl.b32 %r60, %r55, 2;
cvt.s64.s32 %rd3, %r55;
and.b32 %r4, %r60, 4;
shl.b32 %r716, %r718, 3;
add.s32 %r717, %r16, %r716;
add.s32 %r61, %r718, %r46;
add.s32 %r62, %r61, %r50;
or.b32 %r63, %r38, 1;
mad.lo.s32 %r64, %r15, %r63, %r62;
mul.wide.s32 %rd4, %r64, 110;
add.s64 %rd27, %rd4, %rd26;
add.s64 %rd5, %rd27, 106;
add.s64 %rd28, %rd4, 74;
not.b32 %r65, %r56;
mul.wide.s32 %rd29, %r65, 2;
sub.s64 %rd6, %rd28, %rd29;
mul.wide.s32 %rd30, %r56, 2;
add.s64 %rd31, %rd4, %rd30;
add.s64 %rd7, %rd31, 74;
add.s64 %rd32, %rd4, %rd3;
add.s64 %rd8, %rd32, 66;
mad.lo.s32 %r66, %r38, %r15, %r718;
add.s32 %r67, %r66, %r46;
add.s32 %r68, %r67, %r50;
mul.wide.s32 %rd9, %r68, 110;
mul.wide.s32 %rd33, %r60, 2;
add.s64 %rd34, %rd9, %rd33;
add.s64 %rd10, %rd34, 4;
add.s64 %rd35, %rd4, %rd33;
add.s64 %rd11, %rd35, 2;
add.s64 %rd36, %rd9, %rd26;
add.s64 %rd12, %rd36, 106;
add.s64 %rd37, %rd9, 74;
sub.s64 %rd13, %rd37, %rd29;
add.s64 %rd38, %rd9, %rd30;
add.s64 %rd14, %rd38, 74;
add.s64 %rd39, %rd9, %rd3;
add.s64 %rd15, %rd39, 66;
cvta.to.global.u64 %rd16, %rd22;
mov.f32 %f105, 0f00000000;

$L__BB310_2:
cvt.u64.u32 %rd40, %r716;
add.s64 %rd41, %rd1, %rd40;
add.s64 %rd42, %rd41, %rd3;
mul.lo.s64 %rd43, %rd42, 36;
add.s64 %rd44, %rd107, %rd10;
ld.global.nc.u16 %rs5, [%rd44];
ld.global.nc.u16 %rs6, [%rd44+2];
ld.global.nc.u16 %rs7, [%rd44+4];
add.s64 %rd45, %rd107, %rd15;
ld.global.nc.u8 %rs8, [%rd45];
add.s64 %rd46, %rd107, %rd14;
add.s64 %rd47, %rd107, %rd13;
ld.global.nc.u16 %rs9, [%rd46];
ld.global.nc.u16 %rs10, [%rd47];
ld.global.nc.u16 %rs11, [%rd44+-2];
and.b16 %rs12, %rs11, 255;
cvt.u32.u16 %r265, %rs12;
shl.b16 %rs13, %rs8, 8;
cvt.u32.u16 %r266, %rs13;
and.b32 %r267, %r266, 256;
or.b32 %r268, %r267, %r265;
mul.wide.u32 %rd48, %r268, 4;
mov.u64 %rd49, iq3s_grid;
add.s64 %rd50, %rd49, %rd48;
ld.global.nc.u32 %r269, [%rd50];
not.b32 %r270, %r269;
shr.u16 %rs14, %rs11, 8;
cvt.u32.u16 %r271, %rs14;
mul.wide.u16 %r272, %rs8, 128;
and.b32 %r273, %r272, 256;
or.b32 %r274, %r273, %r271;
mul.wide.u32 %rd51, %r274, 4;
add.s64 %rd52, %rd49, %rd51;
ld.global.nc.u32 %r275, [%rd52];
not.b32 %r276, %r275;
and.b16 %rs16, %rs9, 255;
mul.wide.u16 %r277, %rs16, 128;
cvt.u32.u16 %r278, %rs9;
shl.b32 %r279, %r278, 21;
or.b32 %r280, %r277, %r279;
and.b32 %r281, %r280, 25166208;
or.b32 %r282, %r281, -2139062144;
add.s32 %r283, %r282, -16843009;
or.b32 %r167, %r283, %r281;
mov.u32 %r69, %r167;

prmt.b32 %r69, %r69, 0, 0xba98;

mul.wide.u16 %r284, %rs16, 8;
shl.b32 %r285, %r278, 17;
or.b32 %r286, %r284, %r285;
and.b32 %r287, %r286, 25166208;
or.b32 %r288, %r287, -2139062144;
add.s32 %r289, %r288, -16843009;
or.b32 %r169, %r289, %r287;
mov.u32 %r71, %r169;

prmt.b32 %r71, %r71, 0, 0xba98;

xor.b32 %r290, %r69, %r269;
or.b32 %r291, %r290, -2139062144;
and.b32 %r292, %r69, 2139062143;
and.b32 %r293, %r270, -2139062144;
sub.s32 %r294, %r291, %r292;
xor.b32 %r74, %r293, %r294;
xor.b32 %r295, %r71, %r275;
or.b32 %r296, %r295, -2139062144;
and.b32 %r297, %r71, 2139062143;
and.b32 %r298, %r276, -2139062144;
sub.s32 %r299, %r296, %r297;
xor.b32 %r78, %r298, %r299;
add.s64 %rd53, %rd16, %rd43;
ld.global.nc.u32 %r75, [%rd53+4];
ld.global.nc.u32 %r79, [%rd53+8];
mov.u32 %r223, 0;

dp4a.s32.s32 %r73, %r74, %r75, %r223;


dp4a.s32.s32 %r77, %r78, %r79, %r73;

and.b16 %rs17, %rs5, 255;
cvt.u32.u16 %r300, %rs17;
mul.wide.u16 %r301, %rs8, 64;
and.b32 %r302, %r301, 256;
or.b32 %r303, %r302, %r300;
mul.wide.u32 %rd54, %r303, 4;
add.s64 %rd55, %rd49, %rd54;
ld.global.nc.u32 %r304, [%rd55];
not.b32 %r305, %r304;
shr.u16 %rs18, %rs5, 8;
cvt.u32.u16 %r306, %rs18;
mul.wide.u16 %r307, %rs8, 32;
and.b32 %r308, %r307, 256;
or.b32 %r309, %r308, %r306;
mul.wide.u32 %rd56, %r309, 4;
add.s64 %rd57, %rd49, %rd56;
ld.global.nc.u32 %r310, [%rd57];
not.b32 %r311, %r310;
shr.u16 %rs19, %rs9, 8;
mul.wide.u16 %r312, %rs19, 128;
cvt.u32.u16 %r313, %rs19;
shl.b32 %r314, %r313, 21;
or.b32 %r315, %r312, %r314;
and.b32 %r316, %r315, 25166208;
or.b32 %r317, %r316, -2139062144;
add.s32 %r318, %r317, -16843009;
or.b32 %r179, %r318, %r316;
mov.u32 %r81, %r179;

prmt.b32 %r81, %r81, 0, 0xba98;

mul.wide.u16 %r319, %rs19, 8;
shl.b32 %r320, %r313, 17;
or.b32 %r321, %r319, %r320;
and.b32 %r322, %r321, 25166208;
or.b32 %r323, %r322, -2139062144;
add.s32 %r324, %r323, -16843009;
or.b32 %r181, %r324, %r322;
mov.u32 %r83, %r181;

prmt.b32 %r83, %r83, 0, 0xba98;

xor.b32 %r325, %r81, %r304;
or.b32 %r326, %r325, -2139062144;
and.b32 %r327, %r81, 2139062143;
and.b32 %r328, %r305, -2139062144;
sub.s32 %r329, %r326, %r327;
xor.b32 %r86, %r328, %r329;
xor.b32 %r330, %r83, %r310;
or.b32 %r331, %r330, -2139062144;
and.b32 %r332, %r83, 2139062143;
and.b32 %r333, %r311, -2139062144;
sub.s32 %r334, %r331, %r332;
xor.b32 %r90, %r333, %r334;
ld.global.nc.u32 %r87, [%rd53+12];
ld.global.nc.u32 %r91, [%rd53+16];

dp4a.s32.s32 %r85, %r86, %r87, %r77;


dp4a.s32.s32 %r89, %r90, %r91, %r85;

and.b16 %rs20, %rs6, 255;
cvt.u32.u16 %r335, %rs20;
mul.wide.u16 %r336, %rs8, 16;
and.b32 %r337, %r336, 256;
or.b32 %r338, %r337, %r335;
mul.wide.u32 %rd58, %r338, 4;
add.s64 %rd59, %rd49, %rd58;
ld.global.nc.u32 %r339, [%rd59];
not.b32 %r340, %r339;
shr.u16 %rs21, %rs6, 8;
cvt.u32.u16 %r341, %rs21;
mul.wide.u16 %r342, %rs8, 8;
and.b32 %r343, %r342, 256;
or.b32 %r344, %r343, %r341;
mul.wide.u32 %rd60, %r344, 4;
add.s64 %rd61, %rd49, %rd60;
ld.global.nc.u32 %r345, [%rd61];
not.b32 %r346, %r345;
and.b16 %rs22, %rs10, 255;
mul.wide.u16 %r347, %rs22, 128;
cvt.u32.u16 %r348, %rs10;
shl.b32 %r349, %r348, 21;
or.b32 %r350, %r347, %r349;
and.b32 %r351, %r350, 25166208;
or.b32 %r352, %r351, -2139062144;
add.s32 %r353, %r352, -16843009;
or.b32 %r191, %r353, %r351;
mov.u32 %r93, %r191;

prmt.b32 %r93, %r93, 0, 0xba98;

mul.wide.u16 %r354, %rs22, 8;
shl.b32 %r355, %r348, 17;
or.b32 %r356, %r354, %r355;
and.b32 %r357, %r356, 25166208;
or.b32 %r358, %r357, -2139062144;
add.s32 %r359, %r358, -16843009;
or.b32 %r193, %r359, %r357;
mov.u32 %r95, %r193;

prmt.b32 %r95, %r95, 0, 0xba98;

xor.b32 %r360, %r93, %r339;
or.b32 %r361, %r360, -2139062144;
and.b32 %r362, %r93, 2139062143;
and.b32 %r363, %r340, -2139062144;
sub.s32 %r364, %r361, %r362;
xor.b32 %r98, %r363, %r364;
xor.b32 %r365, %r95, %r345;
or.b32 %r366, %r365, -2139062144;
and.b32 %r367, %r95, 2139062143;
and.b32 %r368, %r346, -2139062144;
sub.s32 %r369, %r366, %r367;
xor.b32 %r102, %r368, %r369;
ld.global.nc.u32 %r99, [%rd53+20];
ld.global.nc.u32 %r103, [%rd53+24];

dp4a.s32.s32 %r97, %r98, %r99, %r89;


dp4a.s32.s32 %r101, %r102, %r103, %r97;

and.b16 %rs23, %rs7, 255;
cvt.u32.u16 %r370, %rs23;
mul.wide.u16 %r371, %rs8, 4;
and.b32 %r372, %r371, 256;
or.b32 %r373, %r372, %r370;
mul.wide.u32 %rd62, %r373, 4;
add.s64 %rd63, %rd49, %rd62;
ld.global.nc.u32 %r374, [%rd63];
not.b32 %r375, %r374;
shr.u16 %rs24, %rs7, 8;
cvt.u32.u16 %r376, %rs24;
mul.wide.u16 %r377, %rs8, 2;
and.b32 %r378, %r377, 256;
or.b32 %r379, %r378, %r376;
mul.wide.u32 %rd64, %r379, 4;
add.s64 %rd65, %rd49, %rd64;
ld.global.nc.u32 %r380, [%rd65];
not.b32 %r381, %r380;
shr.u16 %rs25, %rs10, 8;
mul.wide.u16 %r382, %rs25, 128;
cvt.u32.u16 %r383, %rs25;
shl.b32 %r384, %r383, 21;
or.b32 %r385, %r382, %r384;
and.b32 %r386, %r385, 25166208;
or.b32 %r387, %r386, -2139062144;
add.s32 %r388, %r387, -16843009;
or.b32 %r203, %r388, %r386;
mov.u32 %r105, %r203;

prmt.b32 %r105, %r105, 0, 0xba98;

mul.wide.u16 %r389, %rs25, 8;
shl.b32 %r390, %r383, 17;
or.b32 %r391, %r389, %r390;
and.b32 %r392, %r391, 25166208;
or.b32 %r393, %r392, -2139062144;
add.s32 %r394, %r393, -16843009;
or.b32 %r205, %r394, %r392;
mov.u32 %r107, %r205;

prmt.b32 %r107, %r107, 0, 0xba98;

xor.b32 %r395, %r105, %r374;
or.b32 %r396, %r395, -2139062144;
and.b32 %r397, %r105, 2139062143;
and.b32 %r398, %r375, -2139062144;
sub.s32 %r399, %r396, %r397;
xor.b32 %r110, %r398, %r399;
xor.b32 %r400, %r107, %r380;
or.b32 %r401, %r400, -2139062144;
and.b32 %r402, %r107, 2139062143;
and.b32 %r403, %r381, -2139062144;
sub.s32 %r404, %r401, %r402;
xor.b32 %r114, %r403, %r404;
ld.global.nc.u32 %r111, [%rd53+28];
ld.global.nc.u32 %r115, [%rd53+32];

dp4a.s32.s32 %r109, %r110, %r111, %r101;


dp4a.s32.s32 %r113, %r114, %r115, %r109;

add.s64 %rd66, %rd107, %rd12;
ld.global.nc.u8 %rs26, [%rd66];
cvt.u32.u16 %r405, %rs26;
and.b32 %r406, %r405, 255;
shr.u32 %r407, %r406, %r4;
shl.b32 %r408, %r407, 1;
and.b32 %r409, %r408, 30;
or.b32 %r410, %r409, 1;
mul.lo.s32 %r411, %r410, %r113;
add.s64 %rd67, %rd107, %rd9;
ld.global.nc.u16 %rs1, [%rd67];

{ cvt.f32.f16 %f25, %rs1;}


ld.global.nc.u32 %r117, [%rd53];

{.reg .f16 low,high;
mov.b32 {low,high},%r117;
cvt.f32.f16 %f26, low;}


mul.ftz.f32 %f33, %f25, %f26;
cvt.rn.f32.s32 %f34, %r411;
fma.rn.ftz.f32 %f108, %f33, %f34, %f108;
add.s64 %rd68, %rd107, %rd11;
ld.global.nc.u16 %rs27, [%rd68+2];
ld.global.nc.u16 %rs28, [%rd68+4];
ld.global.nc.u16 %rs29, [%rd68+6];
add.s64 %rd69, %rd107, %rd8;
ld.global.nc.u8 %rs30, [%rd69];
add.s64 %rd70, %rd107, %rd7;
add.s64 %rd71, %rd107, %rd6;
ld.global.nc.u16 %rs31, [%rd70];
ld.global.nc.u16 %rs32, [%rd71];
ld.global.nc.u16 %rs33, [%rd68];
and.b16 %rs34, %rs33, 255;
cvt.u32.u16 %r412, %rs34;
shl.b16 %rs35, %rs30, 8;
cvt.u32.u16 %r413, %rs35;
and.b32 %r414, %r413, 256;
or.b32 %r415, %r414, %r412;
mul.wide.u32 %rd72, %r415, 4;
add.s64 %rd73, %rd49, %rd72;
ld.global.nc.u32 %r416, [%rd73];
not.b32 %r417, %r416;
shr.u16 %rs36, %rs33, 8;
cvt.u32.u16 %r418, %rs36;
mul.wide.u16 %r419, %rs30, 128;
and.b32 %r420, %r419, 256;
or.b32 %r421, %r420, %r418;
mul.wide.u32 %rd74, %r421, 4;
add.s64 %rd75, %rd49, %rd74;
ld.global.nc.u32 %r422, [%rd75];
not.b32 %r423, %r422;
and.b16 %rs38, %rs31, 255;
mul.wide.u16 %r424, %rs38, 128;
cvt.u32.u16 %r425, %rs31;
shl.b32 %r426, %r425, 21;
or.b32 %r427, %r424, %r426;
and.b32 %r428, %r427, 25166208;
or.b32 %r429, %r428, -2139062144;
add.s32 %r430, %r429, -16843009;
or.b32 %r216, %r430, %r428;
mov.u32 %r118, %r216;

prmt.b32 %r118, %r118, 0, 0xba98;

mul.wide.u16 %r431, %rs38, 8;
shl.b32 %r432, %r425, 17;
or.b32 %r433, %r431, %r432;
and.b32 %r434, %r433, 25166208;
or.b32 %r435, %r434, -2139062144;
add.s32 %r436, %r435, -16843009;
or.b32 %r218, %r436, %r434;
mov.u32 %r120, %r218;

prmt.b32 %r120, %r120, 0, 0xba98;

xor.b32 %r437, %r118, %r416;
or.b32 %r438, %r437, -2139062144;
and.b32 %r439, %r118, 2139062143;
and.b32 %r440, %r417, -2139062144;
sub.s32 %r441, %r438, %r439;
xor.b32 %r123, %r440, %r441;
xor.b32 %r442, %r120, %r422;
or.b32 %r443, %r442, -2139062144;
and.b32 %r444, %r120, 2139062143;
and.b32 %r445, %r423, -2139062144;
sub.s32 %r446, %r443, %r444;
xor.b32 %r127, %r445, %r446;

dp4a.s32.s32 %r122, %r123, %r75, %r223;


dp4a.s32.s32 %r126, %r127, %r79, %r122;

and.b16 %rs39, %rs27, 255;
cvt.u32.u16 %r447, %rs39;
mul.wide.u16 %r448, %rs30, 64;
and.b32 %r449, %r448, 256;
or.b32 %r450, %r449, %r447;
mul.wide.u32 %rd76, %r450, 4;
add.s64 %rd77, %rd49, %rd76;
ld.global.nc.u32 %r451, [%rd77];
not.b32 %r452, %r451;
shr.u16 %rs40, %rs27, 8;
cvt.u32.u16 %r453, %rs40;
mul.wide.u16 %r454, %rs30, 32;
and.b32 %r455, %r454, 256;
or.b32 %r456, %r455, %r453;
mul.wide.u32 %rd78, %r456, 4;
add.s64 %rd79, %rd49, %rd78;
ld.global.nc.u32 %r457, [%rd79];
not.b32 %r458, %r457;
shr.u16 %rs41, %rs31, 8;
mul.wide.u16 %r459, %rs41, 128;
cvt.u32.u16 %r460, %rs41;
shl.b32 %r461, %r460, 21;
or.b32 %r462, %r459, %r461;
and.b32 %r463, %r462, 25166208;
or.b32 %r464, %r463, -2139062144;
add.s32 %r465, %r464, -16843009;
or.b32 %r228, %r465, %r463;
mov.u32 %r130, %r228;

prmt.b32 %r130, %r130, 0, 0xba98;

mul.wide.u16 %r466, %rs41, 8;
shl.b32 %r467, %r460, 17;
or.b32 %r468, %r466, %r467;
and.b32 %r469, %r468, 25166208;
or.b32 %r470, %r469, -2139062144;
add.s32 %r471, %r470, -16843009;
or.b32 %r230, %r471, %r469;
mov.u32 %r132, %r230;

prmt.b32 %r132, %r132, 0, 0xba98;

xor.b32 %r472, %r130, %r451;
or.b32 %r473, %r472, -2139062144;
and.b32 %r474, %r130, 2139062143;
and.b32 %r475, %r452, -2139062144;
sub.s32 %r476, %r473, %r474;
xor.b32 %r135, %r475, %r476;
xor.b32 %r477, %r132, %r457;
or.b32 %r478, %r477, -2139062144;
and.b32 %r479, %r132, 2139062143;
and.b32 %r480, %r458, -2139062144;
sub.s32 %r481, %r478, %r479;
xor.b32 %r139, %r480, %r481;

dp4a.s32.s32 %r134, %r135, %r87, %r126;


dp4a.s32.s32 %r138, %r139, %r91, %r134;

and.b16 %rs42, %rs28, 255;
cvt.u32.u16 %r482, %rs42;
mul.wide.u16 %r483, %rs30, 16;
and.b32 %r484, %r483, 256;
or.b32 %r485, %r484, %r482;
mul.wide.u32 %rd80, %r485, 4;
add.s64 %rd81, %rd49, %rd80;
ld.global.nc.u32 %r486, [%rd81];
not.b32 %r487, %r486;
shr.u16 %rs43, %rs28, 8;
cvt.u32.u16 %r488, %rs43;
mul.wide.u16 %r489, %rs30, 8;
and.b32 %r490, %r489, 256;
or.b32 %r491, %r490, %r488;
mul.wide.u32 %rd82, %r491, 4;
add.s64 %rd83, %rd49, %rd82;
ld.global.nc.u32 %r492, [%rd83];
not.b32 %r493, %r492;
and.b16 %rs44, %rs32, 255;
mul.wide.u16 %r494, %rs44, 128;
cvt.u32.u16 %r495, %rs32;
shl.b32 %r496, %r495, 21;
or.b32 %r497, %r494, %r496;
and.b32 %r498, %r497, 25166208;
or.b32 %r499, %r498, -2139062144;
add.s32 %r500, %r499, -16843009;
or.b32 %r240, %r500, %r498;
mov.u32 %r142, %r240;

prmt.b32 %r142, %r142, 0, 0xba98;

mul.wide.u16 %r501, %rs44, 8;
shl.b32 %r502, %r495, 17;
or.b32 %r503, %r501, %r502;
and.b32 %r504, %r503, 25166208;
or.b32 %r505, %r504, -2139062144;
add.s32 %r506, %r505, -16843009;
or.b32 %r242, %r506, %r504;
mov.u32 %r144, %r242;

prmt.b32 %r144, %r144, 0, 0xba98;

xor.b32 %r507, %r142, %r486;
or.b32 %r508, %r507, -2139062144;
and.b32 %r509, %r142, 2139062143;
and.b32 %r510, %r487, -2139062144;
sub.s32 %r511, %r508, %r509;
xor.b32 %r147, %r510, %r511;
xor.b32 %r512, %r144, %r492;
or.b32 %r513, %r512, -2139062144;
and.b32 %r514, %r144, 2139062143;
and.b32 %r515, %r493, -2139062144;
sub.s32 %r516, %r513, %r514;
xor.b32 %r151, %r515, %r516;

dp4a.s32.s32 %r146, %r147, %r99, %r138;


dp4a.s32.s32 %r150, %r151, %r103, %r146;

and.b16 %rs45, %rs29, 255;
cvt.u32.u16 %r517, %rs45;
mul.wide.u16 %r518, %rs30, 4;
and.b32 %r519, %r518, 256;
or.b32 %r520, %r519, %r517;
mul.wide.u32 %rd84, %r520, 4;
add.s64 %rd85, %rd49, %rd84;
ld.global.nc.u32 %r521, [%rd85];
not.b32 %r522, %r521;
shr.u16 %rs46, %rs29, 8;
cvt.u32.u16 %r523, %rs46;
mul.wide.u16 %r524, %rs30, 2;
and.b32 %r525, %r524, 256;
or.b32 %r526, %r525, %r523;
mul.wide.u32 %rd86, %r526, 4;
add.s64 %rd87, %rd49, %rd86;
ld.global.nc.u32 %r527, [%rd87];
not.b32 %r528, %r527;
shr.u16 %rs47, %rs32, 8;
mul.wide.u16 %r529, %rs47, 128;
cvt.u32.u16 %r530, %rs47;
shl.b32 %r531, %r530, 21;
or.b32 %r532, %r529, %r531;
and.b32 %r533, %r532, 25166208;
or.b32 %r534, %r533, -2139062144;
add.s32 %r535, %r534, -16843009;
or.b32 %r252, %r535, %r533;
mov.u32 %r154, %r252;

prmt.b32 %r154, %r154, 0, 0xba98;

mul.wide.u16 %r536, %rs47, 8;
shl.b32 %r537, %r530, 17;
or.b32 %r538, %r536, %r537;
and.b32 %r539, %r538, 25166208;
or.b32 %r540, %r539, -2139062144;
add.s32 %r541, %r540, -16843009;
or.b32 %r254, %r541, %r539;
mov.u32 %r156, %r254;

prmt.b32 %r156, %r156, 0, 0xba98;

xor.b32 %r542, %r154, %r521;
or.b32 %r543, %r542, -2139062144;
and.b32 %r544, %r154, 2139062143;
and.b32 %r545, %r522, -2139062144;
sub.s32 %r546, %r543, %r544;
xor.b32 %r159, %r545, %r546;
xor.b32 %r547, %r156, %r527;
or.b32 %r548, %r547, -2139062144;
and.b32 %r549, %r156, 2139062143;
and.b32 %r550, %r528, -2139062144;
sub.s32 %r551, %r548, %r549;
xor.b32 %r163, %r550, %r551;

dp4a.s32.s32 %r158, %r159, %r111, %r150;


dp4a.s32.s32 %r162, %r163, %r115, %r158;

add.s64 %rd88, %rd107, %rd5;
ld.global.nc.u8 %rs48, [%rd88];
cvt.u32.u16 %r552, %rs48;
and.b32 %r553, %r552, 255;
shr.u32 %r554, %r553, %r4;
shl.b32 %r555, %r554, 1;
and.b32 %r556, %r555, 30;
or.b32 %r557, %r556, 1;
mul.lo.s32 %r558, %r557, %r162;
add.s64 %rd89, %rd107, %rd4;
ld.global.nc.u16 %rs2, [%rd89];

{ cvt.f32.f16 %f27, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r117;
cvt.f32.f16 %f28, low;}


mul.ftz.f32 %f35, %f27, %f28;
cvt.rn.f32.s32 %f36, %r558;
fma.rn.ftz.f32 %f107, %f35, %f36, %f107;
cvt.u64.u32 %rd90, %r717;
add.s64 %rd91, %rd1, %rd90;
add.s64 %rd92, %rd91, %rd3;
mul.lo.s64 %rd93, %rd92, 36;

prmt.b32 %r167, %r167, 0, 0xba98;


prmt.b32 %r169, %r169, 0, 0xba98;

xor.b32 %r559, %r167, %r269;
or.b32 %r560, %r559, -2139062144;
and.b32 %r561, %r167, 2139062143;
sub.s32 %r562, %r560, %r561;
xor.b32 %r172, %r293, %r562;
xor.b32 %r563, %r169, %r275;
or.b32 %r564, %r563, -2139062144;
and.b32 %r565, %r169, 2139062143;
sub.s32 %r566, %r564, %r565;
xor.b32 %r176, %r298, %r566;
add.s64 %rd94, %rd16, %rd93;
ld.global.nc.u32 %r173, [%rd94+4];
ld.global.nc.u32 %r177, [%rd94+8];

dp4a.s32.s32 %r171, %r172, %r173, %r223;


dp4a.s32.s32 %r175, %r176, %r177, %r171;


prmt.b32 %r179, %r179, 0, 0xba98;


prmt.b32 %r181, %r181, 0, 0xba98;

xor.b32 %r567, %r179, %r304;
or.b32 %r568, %r567, -2139062144;
and.b32 %r569, %r179, 2139062143;
sub.s32 %r570, %r568, %r569;
xor.b32 %r184, %r328, %r570;
xor.b32 %r571, %r181, %r310;
or.b32 %r572, %r571, -2139062144;
and.b32 %r573, %r181, 2139062143;
sub.s32 %r574, %r572, %r573;
xor.b32 %r188, %r333, %r574;
ld.global.nc.u32 %r185, [%rd94+12];
ld.global.nc.u32 %r189, [%rd94+16];

dp4a.s32.s32 %r183, %r184, %r185, %r175;


dp4a.s32.s32 %r187, %r188, %r189, %r183;


prmt.b32 %r191, %r191, 0, 0xba98;


prmt.b32 %r193, %r193, 0, 0xba98;

xor.b32 %r575, %r191, %r339;
or.b32 %r576, %r575, -2139062144;
and.b32 %r577, %r191, 2139062143;
sub.s32 %r578, %r576, %r577;
xor.b32 %r196, %r363, %r578;
xor.b32 %r579, %r193, %r345;
or.b32 %r580, %r579, -2139062144;
and.b32 %r581, %r193, 2139062143;
sub.s32 %r582, %r580, %r581;
xor.b32 %r200, %r368, %r582;
ld.global.nc.u32 %r197, [%rd94+20];
ld.global.nc.u32 %r201, [%rd94+24];

dp4a.s32.s32 %r195, %r196, %r197, %r187;


dp4a.s32.s32 %r199, %r200, %r201, %r195;


prmt.b32 %r203, %r203, 0, 0xba98;


prmt.b32 %r205, %r205, 0, 0xba98;

xor.b32 %r583, %r203, %r374;
or.b32 %r584, %r583, -2139062144;
and.b32 %r585, %r203, 2139062143;
sub.s32 %r586, %r584, %r585;
xor.b32 %r208, %r398, %r586;
xor.b32 %r587, %r205, %r380;
or.b32 %r588, %r587, -2139062144;
and.b32 %r589, %r205, 2139062143;
sub.s32 %r590, %r588, %r589;
xor.b32 %r212, %r403, %r590;
ld.global.nc.u32 %r209, [%rd94+28];
ld.global.nc.u32 %r213, [%rd94+32];

dp4a.s32.s32 %r207, %r208, %r209, %r199;


dp4a.s32.s32 %r211, %r212, %r213, %r207;

mul.lo.s32 %r591, %r410, %r211;

{ cvt.f32.f16 %f29, %rs1;}


ld.global.nc.u32 %r215, [%rd94];

{.reg .f16 low,high;
mov.b32 {low,high},%r215;
cvt.f32.f16 %f30, low;}


mul.ftz.f32 %f37, %f29, %f30;
cvt.rn.f32.s32 %f38, %r591;
fma.rn.ftz.f32 %f106, %f37, %f38, %f106;

prmt.b32 %r216, %r216, 0, 0xba98;


prmt.b32 %r218, %r218, 0, 0xba98;

xor.b32 %r592, %r216, %r416;
or.b32 %r593, %r592, -2139062144;
and.b32 %r594, %r216, 2139062143;
sub.s32 %r595, %r593, %r594;
xor.b32 %r221, %r440, %r595;
xor.b32 %r596, %r218, %r422;
or.b32 %r597, %r596, -2139062144;
and.b32 %r598, %r218, 2139062143;
sub.s32 %r599, %r597, %r598;
xor.b32 %r225, %r445, %r599;

dp4a.s32.s32 %r220, %r221, %r173, %r223;


dp4a.s32.s32 %r224, %r225, %r177, %r220;


prmt.b32 %r228, %r228, 0, 0xba98;


prmt.b32 %r230, %r230, 0, 0xba98;

xor.b32 %r600, %r228, %r451;
or.b32 %r601, %r600, -2139062144;
and.b32 %r602, %r228, 2139062143;
sub.s32 %r603, %r601, %r602;
xor.b32 %r233, %r475, %r603;
xor.b32 %r604, %r230, %r457;
or.b32 %r605, %r604, -2139062144;
and.b32 %r606, %r230, 2139062143;
sub.s32 %r607, %r605, %r606;
xor.b32 %r237, %r480, %r607;

dp4a.s32.s32 %r232, %r233, %r185, %r224;


dp4a.s32.s32 %r236, %r237, %r189, %r232;


prmt.b32 %r240, %r240, 0, 0xba98;


prmt.b32 %r242, %r242, 0, 0xba98;

xor.b32 %r608, %r240, %r486;
or.b32 %r609, %r608, -2139062144;
and.b32 %r610, %r240, 2139062143;
sub.s32 %r611, %r609, %r610;
xor.b32 %r245, %r510, %r611;
xor.b32 %r612, %r242, %r492;
or.b32 %r613, %r612, -2139062144;
and.b32 %r614, %r242, 2139062143;
sub.s32 %r615, %r613, %r614;
xor.b32 %r249, %r515, %r615;

dp4a.s32.s32 %r244, %r245, %r197, %r236;


dp4a.s32.s32 %r248, %r249, %r201, %r244;


prmt.b32 %r252, %r252, 0, 0xba98;


prmt.b32 %r254, %r254, 0, 0xba98;

xor.b32 %r616, %r252, %r521;
or.b32 %r617, %r616, -2139062144;
and.b32 %r618, %r252, 2139062143;
sub.s32 %r619, %r617, %r618;
xor.b32 %r257, %r545, %r619;
xor.b32 %r620, %r254, %r527;
or.b32 %r621, %r620, -2139062144;
and.b32 %r622, %r254, 2139062143;
sub.s32 %r623, %r621, %r622;
xor.b32 %r261, %r550, %r623;

dp4a.s32.s32 %r256, %r257, %r209, %r248;


dp4a.s32.s32 %r260, %r261, %r213, %r256;

mul.lo.s32 %r624, %r557, %r260;

{ cvt.f32.f16 %f31, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r215;
cvt.f32.f16 %f32, low;}


mul.ftz.f32 %f39, %f31, %f32;
cvt.rn.f32.s32 %f40, %r624;
fma.rn.ftz.f32 %f105, %f39, %f40, %f105;
add.s32 %r717, %r717, 128;
add.s32 %r716, %r716, 128;
add.s64 %rd107, %rd107, 1760;
add.s32 %r718, %r718, 16;
setp.lt.s32 %p4, %r718, %r3;
@%p4 bra $L__BB310_2;

$L__BB310_3:
setp.eq.s32 %p5, %r31, 0;
@%p5 bra $L__BB310_5;

shl.b32 %r628, %r31, 9;
mov.u32 %r629, _ZZ13mul_mat_vec_qIL9ggml_type21ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r630, %r629, %r628;
shl.b32 %r631, %r33, 2;
add.s32 %r632, %r630, %r631;
st.shared.f32 [%r632+-512], %f108;
st.shared.f32 [%r632+-384], %f107;
st.shared.f32 [%r632+-256], %f106;
st.shared.f32 [%r632+-128], %f105;

$L__BB310_5:
setp.ne.s32 %p6, %r31, 0;
bar.sync 0;
@%p6 bra $L__BB310_14;

mov.u32 %r634, %ctaid.z;
mov.u32 %r635, %ctaid.y;
mov.u32 %r636, %ctaid.x;
shl.b32 %r637, %r636, 1;
mov.u32 %r638, 1;
mad.lo.s32 %r639, %r635, %r23, %r637;
mad.lo.s32 %r640, %r634, %r29, %r639;
cvt.u64.u32 %rd19, %r640;
shl.b32 %r642, %r33, 2;
mov.u32 %r643, _ZZ13mul_mat_vec_qIL9ggml_type21ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r13, %r643, %r642;
mov.u32 %r644, 2;
ld.shared.f32 %f41, [%r13];
add.ftz.f32 %f42, %f41, %f108;
ld.shared.f32 %f43, [%r13+512];
add.ftz.f32 %f44, %f43, %f42;
ld.shared.f32 %f45, [%r13+1024];
add.ftz.f32 %f46, %f45, %f44;
mov.b32 %r645, %f46;
mov.u32 %r646, 31;
mov.u32 %r647, 16;
mov.u32 %r648, -1;
shfl.sync.bfly.b32 %r649|%p7, %r645, %r647, %r646, %r648;
mov.b32 %f47, %r649;
add.ftz.f32 %f48, %f46, %f47;
mov.b32 %r650, %f48;
mov.u32 %r651, 8;
shfl.sync.bfly.b32 %r652|%p8, %r650, %r651, %r646, %r648;
mov.b32 %f49, %r652;
add.ftz.f32 %f50, %f48, %f49;
mov.b32 %r653, %f50;
mov.u32 %r654, 4;
shfl.sync.bfly.b32 %r655|%p9, %r653, %r654, %r646, %r648;
mov.b32 %f51, %r655;
add.ftz.f32 %f52, %f50, %f51;
mov.b32 %r656, %f52;
shfl.sync.bfly.b32 %r657|%p10, %r656, %r644, %r646, %r648;
mov.b32 %f53, %r657;
add.ftz.f32 %f54, %f52, %f53;
mov.b32 %r658, %f54;
shfl.sync.bfly.b32 %r659|%p11, %r658, %r638, %r646, %r648;
mov.b32 %f55, %r659;
add.ftz.f32 %f13, %f54, %f55;
setp.eq.s32 %p12, %r33, 0;
setp.lt.u32 %p13, %r637, %r17;
and.pred %p1, %p12, %p13;
cvta.to.global.u64 %rd95, %rd23;
mul.wide.u32 %rd96, %r640, 4;
add.s64 %rd20, %rd95, %rd96;
not.pred %p14, %p1;
@%p14 bra $L__BB310_8;

st.global.f32 [%rd20], %f13;

$L__BB310_8:
ld.shared.f32 %f56, [%r13+128];
add.ftz.f32 %f57, %f56, %f107;
ld.shared.f32 %f58, [%r13+640];
add.ftz.f32 %f59, %f58, %f57;
ld.shared.f32 %f60, [%r13+1152];
add.ftz.f32 %f61, %f60, %f59;
mov.b32 %r660, %f61;
shfl.sync.bfly.b32 %r664|%p15, %r660, %r647, %r646, %r648;
mov.b32 %f62, %r664;
add.ftz.f32 %f63, %f61, %f62;
mov.b32 %r665, %f63;
shfl.sync.bfly.b32 %r667|%p16, %r665, %r651, %r646, %r648;
mov.b32 %f64, %r667;
add.ftz.f32 %f65, %f63, %f64;
mov.b32 %r668, %f65;
shfl.sync.bfly.b32 %r670|%p17, %r668, %r654, %r646, %r648;
mov.b32 %f66, %r670;
add.ftz.f32 %f67, %f65, %f66;
mov.b32 %r671, %f67;
shfl.sync.bfly.b32 %r673|%p18, %r671, %r644, %r646, %r648;
mov.b32 %f68, %r673;
add.ftz.f32 %f69, %f67, %f68;
mov.b32 %r674, %f69;
shfl.sync.bfly.b32 %r676|%p19, %r674, %r638, %r646, %r648;
mov.b32 %f70, %r676;
add.ftz.f32 %f14, %f69, %f70;
setp.eq.s32 %p20, %r33, 1;
or.b32 %r680, %r637, 1;
setp.lt.u32 %p21, %r680, %r17;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB310_10;

st.global.f32 [%rd20+4], %f14;

$L__BB310_10:
ld.shared.f32 %f71, [%r13+256];
add.ftz.f32 %f72, %f71, %f106;
ld.shared.f32 %f73, [%r13+768];
add.ftz.f32 %f74, %f73, %f72;
ld.shared.f32 %f75, [%r13+1280];
add.ftz.f32 %f76, %f75, %f74;
mov.b32 %r681, %f76;
mov.u32 %r682, 31;
mov.u32 %r683, 16;
mov.u32 %r684, -1;
shfl.sync.bfly.b32 %r685|%p23, %r681, %r683, %r682, %r684;
mov.b32 %f77, %r685;
add.ftz.f32 %f78, %f76, %f77;
mov.b32 %r686, %f78;
mov.u32 %r687, 8;
shfl.sync.bfly.b32 %r688|%p24, %r686, %r687, %r682, %r684;
mov.b32 %f79, %r688;
add.ftz.f32 %f80, %f78, %f79;
mov.b32 %r689, %f80;
mov.u32 %r690, 4;
shfl.sync.bfly.b32 %r691|%p25, %r689, %r690, %r682, %r684;
mov.b32 %f81, %r691;
add.ftz.f32 %f82, %f80, %f81;
mov.b32 %r692, %f82;
mov.u32 %r693, 2;
shfl.sync.bfly.b32 %r694|%p26, %r692, %r693, %r682, %r684;
mov.b32 %f83, %r694;
add.ftz.f32 %f84, %f82, %f83;
mov.b32 %r695, %f84;
mov.u32 %r696, 1;
shfl.sync.bfly.b32 %r697|%p27, %r695, %r696, %r682, %r684;
mov.b32 %f85, %r697;
add.ftz.f32 %f15, %f84, %f85;
@%p14 bra $L__BB310_12;

cvt.u64.u32 %rd97, %r17;
add.s64 %rd98, %rd97, %rd19;
shl.b64 %rd100, %rd98, 2;
add.s64 %rd101, %rd95, %rd100;
st.global.f32 [%rd101], %f15;

$L__BB310_12:
ld.shared.f32 %f86, [%r13+384];
add.ftz.f32 %f87, %f86, %f105;
ld.shared.f32 %f88, [%r13+896];
add.ftz.f32 %f89, %f88, %f87;
ld.shared.f32 %f90, [%r13+1408];
add.ftz.f32 %f91, %f90, %f89;
mov.b32 %r698, %f91;
shfl.sync.bfly.b32 %r702|%p29, %r698, %r683, %r682, %r684;
mov.b32 %f92, %r702;
add.ftz.f32 %f93, %f91, %f92;
mov.b32 %r703, %f93;
shfl.sync.bfly.b32 %r705|%p30, %r703, %r687, %r682, %r684;
mov.b32 %f94, %r705;
add.ftz.f32 %f95, %f93, %f94;
mov.b32 %r706, %f95;
shfl.sync.bfly.b32 %r708|%p31, %r706, %r690, %r682, %r684;
mov.b32 %f96, %r708;
add.ftz.f32 %f97, %f95, %f96;
mov.b32 %r709, %f97;
shfl.sync.bfly.b32 %r711|%p32, %r709, %r693, %r682, %r684;
mov.b32 %f98, %r711;
add.ftz.f32 %f99, %f97, %f98;
mov.b32 %r712, %f99;
shfl.sync.bfly.b32 %r714|%p33, %r712, %r696, %r682, %r684;
mov.b32 %f100, %r714;
add.ftz.f32 %f16, %f99, %f100;
@%p22 bra $L__BB310_14;

add.s32 %r715, %r17, 1;
cvt.u64.u32 %rd102, %r715;
add.s64 %rd103, %rd102, %rd19;
shl.b64 %rd105, %rd103, 2;
add.s64 %rd106, %rd95, %rd105;
st.global.f32 [%rd106], %f16;

$L__BB310_14:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type21ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type21ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type21ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type21ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type21ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type21ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type21ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type21ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type21ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<47>;
.reg .b16 %rs<51>;
.reg .f32 %f<163>;
.reg .b32 %r<922>;
.reg .b64 %rd<118>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type21ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[2304];

ld.param.u64 %rd22, [_Z13mul_mat_vec_qIL9ggml_type21ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd23, [_Z13mul_mat_vec_qIL9ggml_type21ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd24, [_Z13mul_mat_vec_qIL9ggml_type21ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type21ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r18, [_Z13mul_mat_vec_qIL9ggml_type21ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r19, [_Z13mul_mat_vec_qIL9ggml_type21ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r20, [_Z13mul_mat_vec_qIL9ggml_type21ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type21ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type21ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type21ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type21ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type21ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type21ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type21ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type21ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type21ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r21, [_Z13mul_mat_vec_qIL9ggml_type21ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r34, %tid.y;
shl.b32 %r35, %r34, 5;
mov.u32 %r36, %tid.x;
add.s32 %r1, %r35, %r36;
shr.s32 %r37, %r1, 31;
shr.u32 %r38, %r37, 29;
add.s32 %r39, %r1, %r38;
shr.s32 %r921, %r39, 3;
shr.u32 %r3, %r33, 8;
setp.ge.s32 %p3, %r921, %r3;
mov.f32 %f157, 0f00000000;
mov.f32 %f158, %f157;
mov.f32 %f159, %f157;
mov.f32 %f160, %f157;
mov.f32 %f161, %f157;
mov.f32 %f162, %f157;
@%p3 bra $L__BB311_3;

mov.u32 %r40, %ctaid.x;
shl.b32 %r41, %r40, 1;
mov.u32 %r42, %ctaid.y;
mov.u32 %r43, %ctaid.z;
mul.lo.s32 %r44, %r42, %r25;
cvt.u64.u32 %rd25, %r44;
mul.lo.s32 %r45, %r43, %r31;
cvt.u64.u32 %rd26, %r45;
add.s64 %rd1, %rd26, %rd25;
mul.hi.u32 %r46, %r43, %r27;
add.s32 %r47, %r46, %r43;
shr.u32 %r48, %r47, %r28;
mul.lo.s32 %r49, %r48, %r30;
mul.hi.u32 %r50, %r42, %r21;
add.s32 %r51, %r50, %r42;
shr.u32 %r52, %r51, %r22;
mul.lo.s32 %r53, %r52, %r24;
cvta.to.global.u64 %rd117, %rd22;
and.b32 %r57, %r39, -8;
sub.s32 %r58, %r1, %r57;
shl.b32 %r59, %r58, 1;
shr.u32 %r60, %r58, 31;
add.s32 %r61, %r58, %r60;
shr.s32 %r62, %r61, 1;
cvt.s64.s32 %rd27, %r62;
shl.b32 %r63, %r58, 2;
cvt.s64.s32 %rd3, %r58;
and.b32 %r4, %r63, 4;
shl.b32 %r64, %r19, 1;
shl.b32 %r918, %r921, 3;
add.s32 %r920, %r918, %r64;
add.s32 %r919, %r19, %r918;
add.s32 %r65, %r921, %r49;
add.s32 %r66, %r65, %r53;
or.b32 %r67, %r41, 1;
mad.lo.s32 %r68, %r18, %r67, %r66;
mul.wide.s32 %rd4, %r68, 110;
add.s64 %rd28, %rd4, %rd27;
add.s64 %rd5, %rd28, 106;
add.s64 %rd29, %rd4, 74;
not.b32 %r69, %r59;
mul.wide.s32 %rd30, %r69, 2;
sub.s64 %rd6, %rd29, %rd30;
mul.wide.s32 %rd31, %r59, 2;
add.s64 %rd32, %rd4, %rd31;
add.s64 %rd7, %rd32, 74;
add.s64 %rd33, %rd4, %rd3;
add.s64 %rd8, %rd33, 66;
mad.lo.s32 %r70, %r41, %r18, %r921;
add.s32 %r71, %r70, %r49;
add.s32 %r72, %r71, %r53;
mul.wide.s32 %rd9, %r72, 110;
mul.wide.s32 %rd34, %r63, 2;
add.s64 %rd35, %rd9, %rd34;
add.s64 %rd10, %rd35, 4;
add.s64 %rd36, %rd4, %rd34;
add.s64 %rd11, %rd36, 2;
add.s64 %rd37, %rd9, %rd27;
add.s64 %rd12, %rd37, 106;
add.s64 %rd38, %rd9, 74;
sub.s64 %rd13, %rd38, %rd30;
add.s64 %rd39, %rd9, %rd31;
add.s64 %rd14, %rd39, 74;
add.s64 %rd40, %rd9, %rd3;
add.s64 %rd15, %rd40, 66;
cvta.to.global.u64 %rd16, %rd23;
mov.f32 %f157, 0f00000000;

$L__BB311_2:
cvt.u64.u32 %rd41, %r918;
add.s64 %rd42, %rd1, %rd41;
add.s64 %rd43, %rd42, %rd3;
mul.lo.s64 %rd44, %rd43, 36;
add.s64 %rd45, %rd117, %rd10;
ld.global.nc.u16 %rs7, [%rd45];
ld.global.nc.u16 %rs8, [%rd45+2];
ld.global.nc.u16 %rs9, [%rd45+4];
add.s64 %rd46, %rd117, %rd15;
ld.global.nc.u8 %rs10, [%rd46];
add.s64 %rd47, %rd117, %rd14;
add.s64 %rd48, %rd117, %rd13;
ld.global.nc.u16 %rs11, [%rd47];
ld.global.nc.u16 %rs12, [%rd48];
ld.global.nc.u16 %rs13, [%rd45+-2];
and.b16 %rs14, %rs13, 255;
cvt.u32.u16 %r367, %rs14;
shl.b16 %rs15, %rs10, 8;
cvt.u32.u16 %r368, %rs15;
and.b32 %r369, %r368, 256;
or.b32 %r370, %r369, %r367;
mul.wide.u32 %rd49, %r370, 4;
mov.u64 %rd50, iq3s_grid;
add.s64 %rd51, %rd50, %rd49;
ld.global.nc.u32 %r371, [%rd51];
not.b32 %r372, %r371;
shr.u16 %rs16, %rs13, 8;
cvt.u32.u16 %r373, %rs16;
mul.wide.u16 %r374, %rs10, 128;
and.b32 %r375, %r374, 256;
or.b32 %r376, %r375, %r373;
mul.wide.u32 %rd52, %r376, 4;
add.s64 %rd53, %rd50, %rd52;
ld.global.nc.u32 %r377, [%rd53];
not.b32 %r378, %r377;
and.b16 %rs18, %rs11, 255;
mul.wide.u16 %r379, %rs18, 128;
cvt.u32.u16 %r380, %rs11;
shl.b32 %r381, %r380, 21;
or.b32 %r382, %r379, %r381;
and.b32 %r383, %r382, 25166208;
or.b32 %r384, %r383, -2139062144;
add.s32 %r385, %r384, -16843009;
or.b32 %r269, %r385, %r383;
mov.u32 %r73, %r269;

prmt.b32 %r73, %r73, 0, 0xba98;

mul.wide.u16 %r386, %rs18, 8;
shl.b32 %r387, %r380, 17;
or.b32 %r388, %r386, %r387;
and.b32 %r389, %r388, 25166208;
or.b32 %r390, %r389, -2139062144;
add.s32 %r391, %r390, -16843009;
or.b32 %r271, %r391, %r389;
mov.u32 %r75, %r271;

prmt.b32 %r75, %r75, 0, 0xba98;

xor.b32 %r392, %r73, %r371;
or.b32 %r393, %r392, -2139062144;
and.b32 %r394, %r73, 2139062143;
and.b32 %r395, %r372, -2139062144;
sub.s32 %r396, %r393, %r394;
xor.b32 %r78, %r395, %r396;
xor.b32 %r397, %r75, %r377;
or.b32 %r398, %r397, -2139062144;
and.b32 %r399, %r75, 2139062143;
and.b32 %r400, %r378, -2139062144;
sub.s32 %r401, %r398, %r399;
xor.b32 %r82, %r400, %r401;
add.s64 %rd54, %rd16, %rd44;
ld.global.nc.u32 %r79, [%rd54+4];
ld.global.nc.u32 %r83, [%rd54+8];
mov.u32 %r325, 0;

dp4a.s32.s32 %r77, %r78, %r79, %r325;


dp4a.s32.s32 %r81, %r82, %r83, %r77;

and.b16 %rs19, %rs7, 255;
cvt.u32.u16 %r402, %rs19;
mul.wide.u16 %r403, %rs10, 64;
and.b32 %r404, %r403, 256;
or.b32 %r405, %r404, %r402;
mul.wide.u32 %rd55, %r405, 4;
add.s64 %rd56, %rd50, %rd55;
ld.global.nc.u32 %r406, [%rd56];
not.b32 %r407, %r406;
shr.u16 %rs20, %rs7, 8;
cvt.u32.u16 %r408, %rs20;
mul.wide.u16 %r409, %rs10, 32;
and.b32 %r410, %r409, 256;
or.b32 %r411, %r410, %r408;
mul.wide.u32 %rd57, %r411, 4;
add.s64 %rd58, %rd50, %rd57;
ld.global.nc.u32 %r412, [%rd58];
not.b32 %r413, %r412;
shr.u16 %rs21, %rs11, 8;
mul.wide.u16 %r414, %rs21, 128;
cvt.u32.u16 %r415, %rs21;
shl.b32 %r416, %r415, 21;
or.b32 %r417, %r414, %r416;
and.b32 %r418, %r417, 25166208;
or.b32 %r419, %r418, -2139062144;
add.s32 %r420, %r419, -16843009;
or.b32 %r281, %r420, %r418;
mov.u32 %r85, %r281;

prmt.b32 %r85, %r85, 0, 0xba98;

mul.wide.u16 %r421, %rs21, 8;
shl.b32 %r422, %r415, 17;
or.b32 %r423, %r421, %r422;
and.b32 %r424, %r423, 25166208;
or.b32 %r425, %r424, -2139062144;
add.s32 %r426, %r425, -16843009;
or.b32 %r283, %r426, %r424;
mov.u32 %r87, %r283;

prmt.b32 %r87, %r87, 0, 0xba98;

xor.b32 %r427, %r85, %r406;
or.b32 %r428, %r427, -2139062144;
and.b32 %r429, %r85, 2139062143;
and.b32 %r430, %r407, -2139062144;
sub.s32 %r431, %r428, %r429;
xor.b32 %r90, %r430, %r431;
xor.b32 %r432, %r87, %r412;
or.b32 %r433, %r432, -2139062144;
and.b32 %r434, %r87, 2139062143;
and.b32 %r435, %r413, -2139062144;
sub.s32 %r436, %r433, %r434;
xor.b32 %r94, %r435, %r436;
ld.global.nc.u32 %r91, [%rd54+12];
ld.global.nc.u32 %r95, [%rd54+16];

dp4a.s32.s32 %r89, %r90, %r91, %r81;


dp4a.s32.s32 %r93, %r94, %r95, %r89;

and.b16 %rs22, %rs8, 255;
cvt.u32.u16 %r437, %rs22;
mul.wide.u16 %r438, %rs10, 16;
and.b32 %r439, %r438, 256;
or.b32 %r440, %r439, %r437;
mul.wide.u32 %rd59, %r440, 4;
add.s64 %rd60, %rd50, %rd59;
ld.global.nc.u32 %r441, [%rd60];
not.b32 %r442, %r441;
shr.u16 %rs23, %rs8, 8;
cvt.u32.u16 %r443, %rs23;
mul.wide.u16 %r444, %rs10, 8;
and.b32 %r445, %r444, 256;
or.b32 %r446, %r445, %r443;
mul.wide.u32 %rd61, %r446, 4;
add.s64 %rd62, %rd50, %rd61;
ld.global.nc.u32 %r447, [%rd62];
not.b32 %r448, %r447;
and.b16 %rs24, %rs12, 255;
mul.wide.u16 %r449, %rs24, 128;
cvt.u32.u16 %r450, %rs12;
shl.b32 %r451, %r450, 21;
or.b32 %r452, %r449, %r451;
and.b32 %r453, %r452, 25166208;
or.b32 %r454, %r453, -2139062144;
add.s32 %r455, %r454, -16843009;
or.b32 %r293, %r455, %r453;
mov.u32 %r97, %r293;

prmt.b32 %r97, %r97, 0, 0xba98;

mul.wide.u16 %r456, %rs24, 8;
shl.b32 %r457, %r450, 17;
or.b32 %r458, %r456, %r457;
and.b32 %r459, %r458, 25166208;
or.b32 %r460, %r459, -2139062144;
add.s32 %r461, %r460, -16843009;
or.b32 %r295, %r461, %r459;
mov.u32 %r99, %r295;

prmt.b32 %r99, %r99, 0, 0xba98;

xor.b32 %r462, %r97, %r441;
or.b32 %r463, %r462, -2139062144;
and.b32 %r464, %r97, 2139062143;
and.b32 %r465, %r442, -2139062144;
sub.s32 %r466, %r463, %r464;
xor.b32 %r102, %r465, %r466;
xor.b32 %r467, %r99, %r447;
or.b32 %r468, %r467, -2139062144;
and.b32 %r469, %r99, 2139062143;
and.b32 %r470, %r448, -2139062144;
sub.s32 %r471, %r468, %r469;
xor.b32 %r106, %r470, %r471;
ld.global.nc.u32 %r103, [%rd54+20];
ld.global.nc.u32 %r107, [%rd54+24];

dp4a.s32.s32 %r101, %r102, %r103, %r93;


dp4a.s32.s32 %r105, %r106, %r107, %r101;

and.b16 %rs25, %rs9, 255;
cvt.u32.u16 %r472, %rs25;
mul.wide.u16 %r473, %rs10, 4;
and.b32 %r474, %r473, 256;
or.b32 %r475, %r474, %r472;
mul.wide.u32 %rd63, %r475, 4;
add.s64 %rd64, %rd50, %rd63;
ld.global.nc.u32 %r476, [%rd64];
not.b32 %r477, %r476;
shr.u16 %rs26, %rs9, 8;
cvt.u32.u16 %r478, %rs26;
mul.wide.u16 %r479, %rs10, 2;
and.b32 %r480, %r479, 256;
or.b32 %r481, %r480, %r478;
mul.wide.u32 %rd65, %r481, 4;
add.s64 %rd66, %rd50, %rd65;
ld.global.nc.u32 %r482, [%rd66];
not.b32 %r483, %r482;
shr.u16 %rs27, %rs12, 8;
mul.wide.u16 %r484, %rs27, 128;
cvt.u32.u16 %r485, %rs27;
shl.b32 %r486, %r485, 21;
or.b32 %r487, %r484, %r486;
and.b32 %r488, %r487, 25166208;
or.b32 %r489, %r488, -2139062144;
add.s32 %r490, %r489, -16843009;
or.b32 %r305, %r490, %r488;
mov.u32 %r109, %r305;

prmt.b32 %r109, %r109, 0, 0xba98;

mul.wide.u16 %r491, %rs27, 8;
shl.b32 %r492, %r485, 17;
or.b32 %r493, %r491, %r492;
and.b32 %r494, %r493, 25166208;
or.b32 %r495, %r494, -2139062144;
add.s32 %r496, %r495, -16843009;
or.b32 %r307, %r496, %r494;
mov.u32 %r111, %r307;

prmt.b32 %r111, %r111, 0, 0xba98;

xor.b32 %r497, %r109, %r476;
or.b32 %r498, %r497, -2139062144;
and.b32 %r499, %r109, 2139062143;
and.b32 %r500, %r477, -2139062144;
sub.s32 %r501, %r498, %r499;
xor.b32 %r114, %r500, %r501;
xor.b32 %r502, %r111, %r482;
or.b32 %r503, %r502, -2139062144;
and.b32 %r504, %r111, 2139062143;
and.b32 %r505, %r483, -2139062144;
sub.s32 %r506, %r503, %r504;
xor.b32 %r118, %r505, %r506;
ld.global.nc.u32 %r115, [%rd54+28];
ld.global.nc.u32 %r119, [%rd54+32];

dp4a.s32.s32 %r113, %r114, %r115, %r105;


dp4a.s32.s32 %r117, %r118, %r119, %r113;

add.s64 %rd67, %rd117, %rd12;
ld.global.nc.u8 %rs28, [%rd67];
cvt.u32.u16 %r507, %rs28;
and.b32 %r508, %r507, 255;
shr.u32 %r509, %r508, %r4;
shl.b32 %r510, %r509, 1;
and.b32 %r511, %r510, 30;
or.b32 %r512, %r511, 1;
mul.lo.s32 %r513, %r512, %r117;
add.s64 %rd68, %rd117, %rd9;
ld.global.nc.u16 %rs1, [%rd68];

{ cvt.f32.f16 %f37, %rs1;}


ld.global.nc.u32 %r121, [%rd54];

{.reg .f16 low,high;
mov.b32 {low,high},%r121;
cvt.f32.f16 %f38, low;}


mul.ftz.f32 %f49, %f37, %f38;
cvt.rn.f32.s32 %f50, %r513;
fma.rn.ftz.f32 %f162, %f49, %f50, %f162;
add.s64 %rd69, %rd117, %rd11;
ld.global.nc.u16 %rs29, [%rd69+2];
ld.global.nc.u16 %rs30, [%rd69+4];
ld.global.nc.u16 %rs31, [%rd69+6];
add.s64 %rd70, %rd117, %rd8;
ld.global.nc.u8 %rs32, [%rd70];
add.s64 %rd71, %rd117, %rd7;
add.s64 %rd72, %rd117, %rd6;
ld.global.nc.u16 %rs33, [%rd71];
ld.global.nc.u16 %rs34, [%rd72];
ld.global.nc.u16 %rs35, [%rd69];
and.b16 %rs36, %rs35, 255;
cvt.u32.u16 %r514, %rs36;
shl.b16 %rs37, %rs32, 8;
cvt.u32.u16 %r515, %rs37;
and.b32 %r516, %r515, 256;
or.b32 %r517, %r516, %r514;
mul.wide.u32 %rd73, %r517, 4;
add.s64 %rd74, %rd50, %rd73;
ld.global.nc.u32 %r518, [%rd74];
not.b32 %r519, %r518;
shr.u16 %rs38, %rs35, 8;
cvt.u32.u16 %r520, %rs38;
mul.wide.u16 %r521, %rs32, 128;
and.b32 %r522, %r521, 256;
or.b32 %r523, %r522, %r520;
mul.wide.u32 %rd75, %r523, 4;
add.s64 %rd76, %rd50, %rd75;
ld.global.nc.u32 %r524, [%rd76];
not.b32 %r525, %r524;
and.b16 %rs40, %rs33, 255;
mul.wide.u16 %r526, %rs40, 128;
cvt.u32.u16 %r527, %rs33;
shl.b32 %r528, %r527, 21;
or.b32 %r529, %r526, %r528;
and.b32 %r530, %r529, 25166208;
or.b32 %r531, %r530, -2139062144;
add.s32 %r532, %r531, -16843009;
or.b32 %r318, %r532, %r530;
mov.u32 %r122, %r318;

prmt.b32 %r122, %r122, 0, 0xba98;

mul.wide.u16 %r533, %rs40, 8;
shl.b32 %r534, %r527, 17;
or.b32 %r535, %r533, %r534;
and.b32 %r536, %r535, 25166208;
or.b32 %r537, %r536, -2139062144;
add.s32 %r538, %r537, -16843009;
or.b32 %r320, %r538, %r536;
mov.u32 %r124, %r320;

prmt.b32 %r124, %r124, 0, 0xba98;

xor.b32 %r539, %r122, %r518;
or.b32 %r540, %r539, -2139062144;
and.b32 %r541, %r122, 2139062143;
and.b32 %r542, %r519, -2139062144;
sub.s32 %r543, %r540, %r541;
xor.b32 %r127, %r542, %r543;
xor.b32 %r544, %r124, %r524;
or.b32 %r545, %r544, -2139062144;
and.b32 %r546, %r124, 2139062143;
and.b32 %r547, %r525, -2139062144;
sub.s32 %r548, %r545, %r546;
xor.b32 %r131, %r547, %r548;

dp4a.s32.s32 %r126, %r127, %r79, %r325;


dp4a.s32.s32 %r130, %r131, %r83, %r126;

and.b16 %rs41, %rs29, 255;
cvt.u32.u16 %r549, %rs41;
mul.wide.u16 %r550, %rs32, 64;
and.b32 %r551, %r550, 256;
or.b32 %r552, %r551, %r549;
mul.wide.u32 %rd77, %r552, 4;
add.s64 %rd78, %rd50, %rd77;
ld.global.nc.u32 %r553, [%rd78];
not.b32 %r554, %r553;
shr.u16 %rs42, %rs29, 8;
cvt.u32.u16 %r555, %rs42;
mul.wide.u16 %r556, %rs32, 32;
and.b32 %r557, %r556, 256;
or.b32 %r558, %r557, %r555;
mul.wide.u32 %rd79, %r558, 4;
add.s64 %rd80, %rd50, %rd79;
ld.global.nc.u32 %r559, [%rd80];
not.b32 %r560, %r559;
shr.u16 %rs43, %rs33, 8;
mul.wide.u16 %r561, %rs43, 128;
cvt.u32.u16 %r562, %rs43;
shl.b32 %r563, %r562, 21;
or.b32 %r564, %r561, %r563;
and.b32 %r565, %r564, 25166208;
or.b32 %r566, %r565, -2139062144;
add.s32 %r567, %r566, -16843009;
or.b32 %r330, %r567, %r565;
mov.u32 %r134, %r330;

prmt.b32 %r134, %r134, 0, 0xba98;

mul.wide.u16 %r568, %rs43, 8;
shl.b32 %r569, %r562, 17;
or.b32 %r570, %r568, %r569;
and.b32 %r571, %r570, 25166208;
or.b32 %r572, %r571, -2139062144;
add.s32 %r573, %r572, -16843009;
or.b32 %r332, %r573, %r571;
mov.u32 %r136, %r332;

prmt.b32 %r136, %r136, 0, 0xba98;

xor.b32 %r574, %r134, %r553;
or.b32 %r575, %r574, -2139062144;
and.b32 %r576, %r134, 2139062143;
and.b32 %r577, %r554, -2139062144;
sub.s32 %r578, %r575, %r576;
xor.b32 %r139, %r577, %r578;
xor.b32 %r579, %r136, %r559;
or.b32 %r580, %r579, -2139062144;
and.b32 %r581, %r136, 2139062143;
and.b32 %r582, %r560, -2139062144;
sub.s32 %r583, %r580, %r581;
xor.b32 %r143, %r582, %r583;

dp4a.s32.s32 %r138, %r139, %r91, %r130;


dp4a.s32.s32 %r142, %r143, %r95, %r138;

and.b16 %rs44, %rs30, 255;
cvt.u32.u16 %r584, %rs44;
mul.wide.u16 %r585, %rs32, 16;
and.b32 %r586, %r585, 256;
or.b32 %r587, %r586, %r584;
mul.wide.u32 %rd81, %r587, 4;
add.s64 %rd82, %rd50, %rd81;
ld.global.nc.u32 %r588, [%rd82];
not.b32 %r589, %r588;
shr.u16 %rs45, %rs30, 8;
cvt.u32.u16 %r590, %rs45;
mul.wide.u16 %r591, %rs32, 8;
and.b32 %r592, %r591, 256;
or.b32 %r593, %r592, %r590;
mul.wide.u32 %rd83, %r593, 4;
add.s64 %rd84, %rd50, %rd83;
ld.global.nc.u32 %r594, [%rd84];
not.b32 %r595, %r594;
and.b16 %rs46, %rs34, 255;
mul.wide.u16 %r596, %rs46, 128;
cvt.u32.u16 %r597, %rs34;
shl.b32 %r598, %r597, 21;
or.b32 %r599, %r596, %r598;
and.b32 %r600, %r599, 25166208;
or.b32 %r601, %r600, -2139062144;
add.s32 %r602, %r601, -16843009;
or.b32 %r342, %r602, %r600;
mov.u32 %r146, %r342;

prmt.b32 %r146, %r146, 0, 0xba98;

mul.wide.u16 %r603, %rs46, 8;
shl.b32 %r604, %r597, 17;
or.b32 %r605, %r603, %r604;
and.b32 %r606, %r605, 25166208;
or.b32 %r607, %r606, -2139062144;
add.s32 %r608, %r607, -16843009;
or.b32 %r344, %r608, %r606;
mov.u32 %r148, %r344;

prmt.b32 %r148, %r148, 0, 0xba98;

xor.b32 %r609, %r146, %r588;
or.b32 %r610, %r609, -2139062144;
and.b32 %r611, %r146, 2139062143;
and.b32 %r612, %r589, -2139062144;
sub.s32 %r613, %r610, %r611;
xor.b32 %r151, %r612, %r613;
xor.b32 %r614, %r148, %r594;
or.b32 %r615, %r614, -2139062144;
and.b32 %r616, %r148, 2139062143;
and.b32 %r617, %r595, -2139062144;
sub.s32 %r618, %r615, %r616;
xor.b32 %r155, %r617, %r618;

dp4a.s32.s32 %r150, %r151, %r103, %r142;


dp4a.s32.s32 %r154, %r155, %r107, %r150;

and.b16 %rs47, %rs31, 255;
cvt.u32.u16 %r619, %rs47;
mul.wide.u16 %r620, %rs32, 4;
and.b32 %r621, %r620, 256;
or.b32 %r622, %r621, %r619;
mul.wide.u32 %rd85, %r622, 4;
add.s64 %rd86, %rd50, %rd85;
ld.global.nc.u32 %r623, [%rd86];
not.b32 %r624, %r623;
shr.u16 %rs48, %rs31, 8;
cvt.u32.u16 %r625, %rs48;
mul.wide.u16 %r626, %rs32, 2;
and.b32 %r627, %r626, 256;
or.b32 %r628, %r627, %r625;
mul.wide.u32 %rd87, %r628, 4;
add.s64 %rd88, %rd50, %rd87;
ld.global.nc.u32 %r629, [%rd88];
not.b32 %r630, %r629;
shr.u16 %rs49, %rs34, 8;
mul.wide.u16 %r631, %rs49, 128;
cvt.u32.u16 %r632, %rs49;
shl.b32 %r633, %r632, 21;
or.b32 %r634, %r631, %r633;
and.b32 %r635, %r634, 25166208;
or.b32 %r636, %r635, -2139062144;
add.s32 %r637, %r636, -16843009;
or.b32 %r354, %r637, %r635;
mov.u32 %r158, %r354;

prmt.b32 %r158, %r158, 0, 0xba98;

mul.wide.u16 %r638, %rs49, 8;
shl.b32 %r639, %r632, 17;
or.b32 %r640, %r638, %r639;
and.b32 %r641, %r640, 25166208;
or.b32 %r642, %r641, -2139062144;
add.s32 %r643, %r642, -16843009;
or.b32 %r356, %r643, %r641;
mov.u32 %r160, %r356;

prmt.b32 %r160, %r160, 0, 0xba98;

xor.b32 %r644, %r158, %r623;
or.b32 %r645, %r644, -2139062144;
and.b32 %r646, %r158, 2139062143;
and.b32 %r647, %r624, -2139062144;
sub.s32 %r648, %r645, %r646;
xor.b32 %r163, %r647, %r648;
xor.b32 %r649, %r160, %r629;
or.b32 %r650, %r649, -2139062144;
and.b32 %r651, %r160, 2139062143;
and.b32 %r652, %r630, -2139062144;
sub.s32 %r653, %r650, %r651;
xor.b32 %r167, %r652, %r653;

dp4a.s32.s32 %r162, %r163, %r115, %r154;


dp4a.s32.s32 %r166, %r167, %r119, %r162;

add.s64 %rd89, %rd117, %rd5;
ld.global.nc.u8 %rs50, [%rd89];
cvt.u32.u16 %r654, %rs50;
and.b32 %r655, %r654, 255;
shr.u32 %r656, %r655, %r4;
shl.b32 %r657, %r656, 1;
and.b32 %r658, %r657, 30;
or.b32 %r659, %r658, 1;
mul.lo.s32 %r660, %r659, %r166;
add.s64 %rd90, %rd117, %rd4;
ld.global.nc.u16 %rs2, [%rd90];

{ cvt.f32.f16 %f39, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r121;
cvt.f32.f16 %f40, low;}


mul.ftz.f32 %f51, %f39, %f40;
cvt.rn.f32.s32 %f52, %r660;
fma.rn.ftz.f32 %f161, %f51, %f52, %f161;
cvt.u64.u32 %rd91, %r919;
add.s64 %rd92, %rd1, %rd91;
add.s64 %rd93, %rd92, %rd3;
mul.lo.s64 %rd94, %rd93, 36;
mov.u32 %r171, %r269;

prmt.b32 %r171, %r171, 0, 0xba98;

mov.u32 %r173, %r271;

prmt.b32 %r173, %r173, 0, 0xba98;

xor.b32 %r661, %r171, %r371;
or.b32 %r662, %r661, -2139062144;
and.b32 %r663, %r171, 2139062143;
sub.s32 %r664, %r662, %r663;
xor.b32 %r176, %r395, %r664;
xor.b32 %r665, %r173, %r377;
or.b32 %r666, %r665, -2139062144;
and.b32 %r667, %r173, 2139062143;
sub.s32 %r668, %r666, %r667;
xor.b32 %r180, %r400, %r668;
add.s64 %rd95, %rd16, %rd94;
ld.global.nc.u32 %r177, [%rd95+4];
ld.global.nc.u32 %r181, [%rd95+8];

dp4a.s32.s32 %r175, %r176, %r177, %r325;


dp4a.s32.s32 %r179, %r180, %r181, %r175;

mov.u32 %r183, %r281;

prmt.b32 %r183, %r183, 0, 0xba98;

mov.u32 %r185, %r283;

prmt.b32 %r185, %r185, 0, 0xba98;

xor.b32 %r669, %r183, %r406;
or.b32 %r670, %r669, -2139062144;
and.b32 %r671, %r183, 2139062143;
sub.s32 %r672, %r670, %r671;
xor.b32 %r188, %r430, %r672;
xor.b32 %r673, %r185, %r412;
or.b32 %r674, %r673, -2139062144;
and.b32 %r675, %r185, 2139062143;
sub.s32 %r676, %r674, %r675;
xor.b32 %r192, %r435, %r676;
ld.global.nc.u32 %r189, [%rd95+12];
ld.global.nc.u32 %r193, [%rd95+16];

dp4a.s32.s32 %r187, %r188, %r189, %r179;


dp4a.s32.s32 %r191, %r192, %r193, %r187;

mov.u32 %r195, %r293;

prmt.b32 %r195, %r195, 0, 0xba98;

mov.u32 %r197, %r295;

prmt.b32 %r197, %r197, 0, 0xba98;

xor.b32 %r677, %r195, %r441;
or.b32 %r678, %r677, -2139062144;
and.b32 %r679, %r195, 2139062143;
sub.s32 %r680, %r678, %r679;
xor.b32 %r200, %r465, %r680;
xor.b32 %r681, %r197, %r447;
or.b32 %r682, %r681, -2139062144;
and.b32 %r683, %r197, 2139062143;
sub.s32 %r684, %r682, %r683;
xor.b32 %r204, %r470, %r684;
ld.global.nc.u32 %r201, [%rd95+20];
ld.global.nc.u32 %r205, [%rd95+24];

dp4a.s32.s32 %r199, %r200, %r201, %r191;


dp4a.s32.s32 %r203, %r204, %r205, %r199;

mov.u32 %r207, %r305;

prmt.b32 %r207, %r207, 0, 0xba98;

mov.u32 %r209, %r307;

prmt.b32 %r209, %r209, 0, 0xba98;

xor.b32 %r685, %r207, %r476;
or.b32 %r686, %r685, -2139062144;
and.b32 %r687, %r207, 2139062143;
sub.s32 %r688, %r686, %r687;
xor.b32 %r212, %r500, %r688;
xor.b32 %r689, %r209, %r482;
or.b32 %r690, %r689, -2139062144;
and.b32 %r691, %r209, 2139062143;
sub.s32 %r692, %r690, %r691;
xor.b32 %r216, %r505, %r692;
ld.global.nc.u32 %r213, [%rd95+28];
ld.global.nc.u32 %r217, [%rd95+32];

dp4a.s32.s32 %r211, %r212, %r213, %r203;


dp4a.s32.s32 %r215, %r216, %r217, %r211;

mul.lo.s32 %r693, %r512, %r215;

{ cvt.f32.f16 %f41, %rs1;}


ld.global.nc.u32 %r219, [%rd95];

{.reg .f16 low,high;
mov.b32 {low,high},%r219;
cvt.f32.f16 %f42, low;}


mul.ftz.f32 %f53, %f41, %f42;
cvt.rn.f32.s32 %f54, %r693;
fma.rn.ftz.f32 %f160, %f53, %f54, %f160;
mov.u32 %r220, %r318;

prmt.b32 %r220, %r220, 0, 0xba98;

mov.u32 %r222, %r320;

prmt.b32 %r222, %r222, 0, 0xba98;

xor.b32 %r694, %r220, %r518;
or.b32 %r695, %r694, -2139062144;
and.b32 %r696, %r220, 2139062143;
sub.s32 %r697, %r695, %r696;
xor.b32 %r225, %r542, %r697;
xor.b32 %r698, %r222, %r524;
or.b32 %r699, %r698, -2139062144;
and.b32 %r700, %r222, 2139062143;
sub.s32 %r701, %r699, %r700;
xor.b32 %r229, %r547, %r701;

dp4a.s32.s32 %r224, %r225, %r177, %r325;


dp4a.s32.s32 %r228, %r229, %r181, %r224;

mov.u32 %r232, %r330;

prmt.b32 %r232, %r232, 0, 0xba98;

mov.u32 %r234, %r332;

prmt.b32 %r234, %r234, 0, 0xba98;

xor.b32 %r702, %r232, %r553;
or.b32 %r703, %r702, -2139062144;
and.b32 %r704, %r232, 2139062143;
sub.s32 %r705, %r703, %r704;
xor.b32 %r237, %r577, %r705;
xor.b32 %r706, %r234, %r559;
or.b32 %r707, %r706, -2139062144;
and.b32 %r708, %r234, 2139062143;
sub.s32 %r709, %r707, %r708;
xor.b32 %r241, %r582, %r709;

dp4a.s32.s32 %r236, %r237, %r189, %r228;


dp4a.s32.s32 %r240, %r241, %r193, %r236;

mov.u32 %r244, %r342;

prmt.b32 %r244, %r244, 0, 0xba98;

mov.u32 %r246, %r344;

prmt.b32 %r246, %r246, 0, 0xba98;

xor.b32 %r710, %r244, %r588;
or.b32 %r711, %r710, -2139062144;
and.b32 %r712, %r244, 2139062143;
sub.s32 %r713, %r711, %r712;
xor.b32 %r249, %r612, %r713;
xor.b32 %r714, %r246, %r594;
or.b32 %r715, %r714, -2139062144;
and.b32 %r716, %r246, 2139062143;
sub.s32 %r717, %r715, %r716;
xor.b32 %r253, %r617, %r717;

dp4a.s32.s32 %r248, %r249, %r201, %r240;


dp4a.s32.s32 %r252, %r253, %r205, %r248;

mov.u32 %r256, %r354;

prmt.b32 %r256, %r256, 0, 0xba98;

mov.u32 %r258, %r356;

prmt.b32 %r258, %r258, 0, 0xba98;

xor.b32 %r718, %r256, %r623;
or.b32 %r719, %r718, -2139062144;
and.b32 %r720, %r256, 2139062143;
sub.s32 %r721, %r719, %r720;
xor.b32 %r261, %r647, %r721;
xor.b32 %r722, %r258, %r629;
or.b32 %r723, %r722, -2139062144;
and.b32 %r724, %r258, 2139062143;
sub.s32 %r725, %r723, %r724;
xor.b32 %r265, %r652, %r725;

dp4a.s32.s32 %r260, %r261, %r213, %r252;


dp4a.s32.s32 %r264, %r265, %r217, %r260;

mul.lo.s32 %r726, %r659, %r264;

{ cvt.f32.f16 %f43, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r219;
cvt.f32.f16 %f44, low;}


mul.ftz.f32 %f55, %f43, %f44;
cvt.rn.f32.s32 %f56, %r726;
fma.rn.ftz.f32 %f159, %f55, %f56, %f159;
cvt.u64.u32 %rd96, %r920;
add.s64 %rd97, %rd1, %rd96;
add.s64 %rd98, %rd97, %rd3;
mul.lo.s64 %rd99, %rd98, 36;

prmt.b32 %r269, %r269, 0, 0xba98;


prmt.b32 %r271, %r271, 0, 0xba98;

xor.b32 %r727, %r269, %r371;
or.b32 %r728, %r727, -2139062144;
and.b32 %r729, %r269, 2139062143;
sub.s32 %r730, %r728, %r729;
xor.b32 %r274, %r395, %r730;
xor.b32 %r731, %r271, %r377;
or.b32 %r732, %r731, -2139062144;
and.b32 %r733, %r271, 2139062143;
sub.s32 %r734, %r732, %r733;
xor.b32 %r278, %r400, %r734;
add.s64 %rd100, %rd16, %rd99;
ld.global.nc.u32 %r275, [%rd100+4];
ld.global.nc.u32 %r279, [%rd100+8];

dp4a.s32.s32 %r273, %r274, %r275, %r325;


dp4a.s32.s32 %r277, %r278, %r279, %r273;


prmt.b32 %r281, %r281, 0, 0xba98;


prmt.b32 %r283, %r283, 0, 0xba98;

xor.b32 %r735, %r281, %r406;
or.b32 %r736, %r735, -2139062144;
and.b32 %r737, %r281, 2139062143;
sub.s32 %r738, %r736, %r737;
xor.b32 %r286, %r430, %r738;
xor.b32 %r739, %r283, %r412;
or.b32 %r740, %r739, -2139062144;
and.b32 %r741, %r283, 2139062143;
sub.s32 %r742, %r740, %r741;
xor.b32 %r290, %r435, %r742;
ld.global.nc.u32 %r287, [%rd100+12];
ld.global.nc.u32 %r291, [%rd100+16];

dp4a.s32.s32 %r285, %r286, %r287, %r277;


dp4a.s32.s32 %r289, %r290, %r291, %r285;


prmt.b32 %r293, %r293, 0, 0xba98;


prmt.b32 %r295, %r295, 0, 0xba98;

xor.b32 %r743, %r293, %r441;
or.b32 %r744, %r743, -2139062144;
and.b32 %r745, %r293, 2139062143;
sub.s32 %r746, %r744, %r745;
xor.b32 %r298, %r465, %r746;
xor.b32 %r747, %r295, %r447;
or.b32 %r748, %r747, -2139062144;
and.b32 %r749, %r295, 2139062143;
sub.s32 %r750, %r748, %r749;
xor.b32 %r302, %r470, %r750;
ld.global.nc.u32 %r299, [%rd100+20];
ld.global.nc.u32 %r303, [%rd100+24];

dp4a.s32.s32 %r297, %r298, %r299, %r289;


dp4a.s32.s32 %r301, %r302, %r303, %r297;


prmt.b32 %r305, %r305, 0, 0xba98;


prmt.b32 %r307, %r307, 0, 0xba98;

xor.b32 %r751, %r305, %r476;
or.b32 %r752, %r751, -2139062144;
and.b32 %r753, %r305, 2139062143;
sub.s32 %r754, %r752, %r753;
xor.b32 %r310, %r500, %r754;
xor.b32 %r755, %r307, %r482;
or.b32 %r756, %r755, -2139062144;
and.b32 %r757, %r307, 2139062143;
sub.s32 %r758, %r756, %r757;
xor.b32 %r314, %r505, %r758;
ld.global.nc.u32 %r311, [%rd100+28];
ld.global.nc.u32 %r315, [%rd100+32];

dp4a.s32.s32 %r309, %r310, %r311, %r301;


dp4a.s32.s32 %r313, %r314, %r315, %r309;

mul.lo.s32 %r759, %r512, %r313;

{ cvt.f32.f16 %f45, %rs1;}


ld.global.nc.u32 %r317, [%rd100];

{.reg .f16 low,high;
mov.b32 {low,high},%r317;
cvt.f32.f16 %f46, low;}


mul.ftz.f32 %f57, %f45, %f46;
cvt.rn.f32.s32 %f58, %r759;
fma.rn.ftz.f32 %f158, %f57, %f58, %f158;

prmt.b32 %r318, %r318, 0, 0xba98;


prmt.b32 %r320, %r320, 0, 0xba98;

xor.b32 %r760, %r318, %r518;
or.b32 %r761, %r760, -2139062144;
and.b32 %r762, %r318, 2139062143;
sub.s32 %r763, %r761, %r762;
xor.b32 %r323, %r542, %r763;
xor.b32 %r764, %r320, %r524;
or.b32 %r765, %r764, -2139062144;
and.b32 %r766, %r320, 2139062143;
sub.s32 %r767, %r765, %r766;
xor.b32 %r327, %r547, %r767;

dp4a.s32.s32 %r322, %r323, %r275, %r325;


dp4a.s32.s32 %r326, %r327, %r279, %r322;


prmt.b32 %r330, %r330, 0, 0xba98;


prmt.b32 %r332, %r332, 0, 0xba98;

xor.b32 %r768, %r330, %r553;
or.b32 %r769, %r768, -2139062144;
and.b32 %r770, %r330, 2139062143;
sub.s32 %r771, %r769, %r770;
xor.b32 %r335, %r577, %r771;
xor.b32 %r772, %r332, %r559;
or.b32 %r773, %r772, -2139062144;
and.b32 %r774, %r332, 2139062143;
sub.s32 %r775, %r773, %r774;
xor.b32 %r339, %r582, %r775;

dp4a.s32.s32 %r334, %r335, %r287, %r326;


dp4a.s32.s32 %r338, %r339, %r291, %r334;


prmt.b32 %r342, %r342, 0, 0xba98;


prmt.b32 %r344, %r344, 0, 0xba98;

xor.b32 %r776, %r342, %r588;
or.b32 %r777, %r776, -2139062144;
and.b32 %r778, %r342, 2139062143;
sub.s32 %r779, %r777, %r778;
xor.b32 %r347, %r612, %r779;
xor.b32 %r780, %r344, %r594;
or.b32 %r781, %r780, -2139062144;
and.b32 %r782, %r344, 2139062143;
sub.s32 %r783, %r781, %r782;
xor.b32 %r351, %r617, %r783;

dp4a.s32.s32 %r346, %r347, %r299, %r338;


dp4a.s32.s32 %r350, %r351, %r303, %r346;


prmt.b32 %r354, %r354, 0, 0xba98;


prmt.b32 %r356, %r356, 0, 0xba98;

xor.b32 %r784, %r354, %r623;
or.b32 %r785, %r784, -2139062144;
and.b32 %r786, %r354, 2139062143;
sub.s32 %r787, %r785, %r786;
xor.b32 %r359, %r647, %r787;
xor.b32 %r788, %r356, %r629;
or.b32 %r789, %r788, -2139062144;
and.b32 %r790, %r356, 2139062143;
sub.s32 %r791, %r789, %r790;
xor.b32 %r363, %r652, %r791;

dp4a.s32.s32 %r358, %r359, %r311, %r350;


dp4a.s32.s32 %r362, %r363, %r315, %r358;

mul.lo.s32 %r792, %r659, %r362;

{ cvt.f32.f16 %f47, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r317;
cvt.f32.f16 %f48, low;}


mul.ftz.f32 %f59, %f47, %f48;
cvt.rn.f32.s32 %f60, %r792;
fma.rn.ftz.f32 %f157, %f59, %f60, %f157;
add.s32 %r920, %r920, 128;
add.s32 %r919, %r919, 128;
add.s32 %r918, %r918, 128;
add.s64 %rd117, %rd117, 1760;
add.s32 %r921, %r921, 16;
setp.lt.s32 %p4, %r921, %r3;
@%p4 bra $L__BB311_2;

$L__BB311_3:
setp.eq.s32 %p5, %r34, 0;
@%p5 bra $L__BB311_5;

mov.u32 %r796, _ZZ13mul_mat_vec_qIL9ggml_type21ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r797, %r34, 768, %r796;
shl.b32 %r798, %r36, 2;
add.s32 %r799, %r797, %r798;
st.shared.f32 [%r799+-768], %f162;
st.shared.f32 [%r799+-640], %f161;
st.shared.f32 [%r799+-512], %f160;
st.shared.f32 [%r799+-384], %f159;
st.shared.f32 [%r799+-256], %f158;
st.shared.f32 [%r799+-128], %f157;

$L__BB311_5:
setp.ne.s32 %p6, %r34, 0;
bar.sync 0;
@%p6 bra $L__BB311_18;

mov.u32 %r801, %ctaid.z;
mov.u32 %r802, %ctaid.y;
mov.u32 %r803, %ctaid.x;
shl.b32 %r804, %r803, 1;
mov.u32 %r805, 1;
mad.lo.s32 %r806, %r802, %r26, %r804;
mad.lo.s32 %r807, %r801, %r32, %r806;
cvt.u64.u32 %rd19, %r807;
shl.b32 %r809, %r36, 2;
mov.u32 %r810, _ZZ13mul_mat_vec_qIL9ggml_type21ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r16, %r810, %r809;
mov.u32 %r811, 2;
ld.shared.f32 %f61, [%r16];
add.ftz.f32 %f62, %f61, %f162;
ld.shared.f32 %f63, [%r16+768];
add.ftz.f32 %f64, %f63, %f62;
ld.shared.f32 %f65, [%r16+1536];
add.ftz.f32 %f66, %f65, %f64;
mov.b32 %r812, %f66;
mov.u32 %r813, 31;
mov.u32 %r814, 16;
mov.u32 %r815, -1;
shfl.sync.bfly.b32 %r816|%p7, %r812, %r814, %r813, %r815;
mov.b32 %f67, %r816;
add.ftz.f32 %f68, %f66, %f67;
mov.b32 %r817, %f68;
mov.u32 %r818, 8;
shfl.sync.bfly.b32 %r819|%p8, %r817, %r818, %r813, %r815;
mov.b32 %f69, %r819;
add.ftz.f32 %f70, %f68, %f69;
mov.b32 %r820, %f70;
mov.u32 %r821, 4;
shfl.sync.bfly.b32 %r822|%p9, %r820, %r821, %r813, %r815;
mov.b32 %f71, %r822;
add.ftz.f32 %f72, %f70, %f71;
mov.b32 %r823, %f72;
shfl.sync.bfly.b32 %r824|%p10, %r823, %r811, %r813, %r815;
mov.b32 %f73, %r824;
add.ftz.f32 %f74, %f72, %f73;
mov.b32 %r825, %f74;
shfl.sync.bfly.b32 %r826|%p11, %r825, %r805, %r813, %r815;
mov.b32 %f75, %r826;
add.ftz.f32 %f19, %f74, %f75;
setp.eq.s32 %p12, %r36, 0;
setp.lt.u32 %p13, %r804, %r20;
and.pred %p1, %p12, %p13;
cvta.to.global.u64 %rd101, %rd24;
mul.wide.u32 %rd102, %r807, 4;
add.s64 %rd20, %rd101, %rd102;
not.pred %p14, %p1;
@%p14 bra $L__BB311_8;

st.global.f32 [%rd20], %f19;

$L__BB311_8:
ld.shared.f32 %f76, [%r16+128];
add.ftz.f32 %f77, %f76, %f161;
ld.shared.f32 %f78, [%r16+896];
add.ftz.f32 %f79, %f78, %f77;
ld.shared.f32 %f80, [%r16+1664];
add.ftz.f32 %f81, %f80, %f79;
mov.b32 %r827, %f81;
shfl.sync.bfly.b32 %r831|%p15, %r827, %r814, %r813, %r815;
mov.b32 %f82, %r831;
add.ftz.f32 %f83, %f81, %f82;
mov.b32 %r832, %f83;
shfl.sync.bfly.b32 %r834|%p16, %r832, %r818, %r813, %r815;
mov.b32 %f84, %r834;
add.ftz.f32 %f85, %f83, %f84;
mov.b32 %r835, %f85;
shfl.sync.bfly.b32 %r837|%p17, %r835, %r821, %r813, %r815;
mov.b32 %f86, %r837;
add.ftz.f32 %f87, %f85, %f86;
mov.b32 %r838, %f87;
shfl.sync.bfly.b32 %r840|%p18, %r838, %r811, %r813, %r815;
mov.b32 %f88, %r840;
add.ftz.f32 %f89, %f87, %f88;
mov.b32 %r841, %f89;
shfl.sync.bfly.b32 %r843|%p19, %r841, %r805, %r813, %r815;
mov.b32 %f90, %r843;
add.ftz.f32 %f20, %f89, %f90;
setp.eq.s32 %p20, %r36, 1;
or.b32 %r847, %r804, 1;
setp.lt.u32 %p21, %r847, %r20;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB311_10;

st.global.f32 [%rd20+4], %f20;

$L__BB311_10:
ld.shared.f32 %f91, [%r16+256];
add.ftz.f32 %f92, %f91, %f160;
ld.shared.f32 %f93, [%r16+1024];
add.ftz.f32 %f94, %f93, %f92;
ld.shared.f32 %f95, [%r16+1792];
add.ftz.f32 %f96, %f95, %f94;
mov.b32 %r848, %f96;
mov.u32 %r849, 31;
mov.u32 %r850, 16;
mov.u32 %r851, -1;
shfl.sync.bfly.b32 %r852|%p23, %r848, %r850, %r849, %r851;
mov.b32 %f97, %r852;
add.ftz.f32 %f98, %f96, %f97;
mov.b32 %r853, %f98;
mov.u32 %r854, 8;
shfl.sync.bfly.b32 %r855|%p24, %r853, %r854, %r849, %r851;
mov.b32 %f99, %r855;
add.ftz.f32 %f100, %f98, %f99;
mov.b32 %r856, %f100;
mov.u32 %r857, 4;
shfl.sync.bfly.b32 %r858|%p25, %r856, %r857, %r849, %r851;
mov.b32 %f101, %r858;
add.ftz.f32 %f102, %f100, %f101;
mov.b32 %r859, %f102;
mov.u32 %r860, 2;
shfl.sync.bfly.b32 %r861|%p26, %r859, %r860, %r849, %r851;
mov.b32 %f103, %r861;
add.ftz.f32 %f104, %f102, %f103;
mov.b32 %r862, %f104;
mov.u32 %r863, 1;
shfl.sync.bfly.b32 %r864|%p27, %r862, %r863, %r849, %r851;
mov.b32 %f105, %r864;
add.ftz.f32 %f21, %f104, %f105;
@%p14 bra $L__BB311_12;

cvt.u64.u32 %rd103, %r20;
add.s64 %rd104, %rd103, %rd19;
shl.b64 %rd106, %rd104, 2;
add.s64 %rd107, %rd101, %rd106;
st.global.f32 [%rd107], %f21;

$L__BB311_12:
ld.shared.f32 %f106, [%r16+384];
add.ftz.f32 %f107, %f106, %f159;
ld.shared.f32 %f108, [%r16+1152];
add.ftz.f32 %f109, %f108, %f107;
ld.shared.f32 %f110, [%r16+1920];
add.ftz.f32 %f111, %f110, %f109;
mov.b32 %r865, %f111;
shfl.sync.bfly.b32 %r869|%p29, %r865, %r850, %r849, %r851;
mov.b32 %f112, %r869;
add.ftz.f32 %f113, %f111, %f112;
mov.b32 %r870, %f113;
shfl.sync.bfly.b32 %r872|%p30, %r870, %r854, %r849, %r851;
mov.b32 %f114, %r872;
add.ftz.f32 %f115, %f113, %f114;
mov.b32 %r873, %f115;
shfl.sync.bfly.b32 %r875|%p31, %r873, %r857, %r849, %r851;
mov.b32 %f116, %r875;
add.ftz.f32 %f117, %f115, %f116;
mov.b32 %r876, %f117;
shfl.sync.bfly.b32 %r878|%p32, %r876, %r860, %r849, %r851;
mov.b32 %f118, %r878;
add.ftz.f32 %f119, %f117, %f118;
mov.b32 %r879, %f119;
shfl.sync.bfly.b32 %r881|%p33, %r879, %r863, %r849, %r851;
mov.b32 %f120, %r881;
add.ftz.f32 %f22, %f119, %f120;
@%p22 bra $L__BB311_14;

add.s32 %r882, %r20, 1;
cvt.u64.u32 %rd108, %r882;
add.s64 %rd109, %rd108, %rd19;
shl.b64 %rd111, %rd109, 2;
add.s64 %rd112, %rd101, %rd111;
st.global.f32 [%rd112], %f22;

$L__BB311_14:
shl.b32 %r883, %r20, 1;
mov.u32 %r884, 1;
ld.shared.f32 %f121, [%r16+512];
add.ftz.f32 %f122, %f121, %f158;
ld.shared.f32 %f123, [%r16+1280];
add.ftz.f32 %f124, %f123, %f122;
ld.shared.f32 %f125, [%r16+2048];
add.ftz.f32 %f126, %f125, %f124;
mov.b32 %r885, %f126;
mov.u32 %r886, 31;
mov.u32 %r887, 16;
mov.u32 %r888, -1;
shfl.sync.bfly.b32 %r889|%p35, %r885, %r887, %r886, %r888;
mov.b32 %f127, %r889;
add.ftz.f32 %f128, %f126, %f127;
mov.b32 %r890, %f128;
mov.u32 %r891, 8;
shfl.sync.bfly.b32 %r892|%p36, %r890, %r891, %r886, %r888;
mov.b32 %f129, %r892;
add.ftz.f32 %f130, %f128, %f129;
mov.b32 %r893, %f130;
mov.u32 %r894, 4;
shfl.sync.bfly.b32 %r895|%p37, %r893, %r894, %r886, %r888;
mov.b32 %f131, %r895;
add.ftz.f32 %f132, %f130, %f131;
mov.b32 %r896, %f132;
mov.u32 %r897, 2;
shfl.sync.bfly.b32 %r898|%p38, %r896, %r897, %r886, %r888;
mov.b32 %f133, %r898;
add.ftz.f32 %f134, %f132, %f133;
mov.b32 %r899, %f134;
shfl.sync.bfly.b32 %r900|%p39, %r899, %r884, %r886, %r888;
mov.b32 %f135, %r900;
add.ftz.f32 %f23, %f134, %f135;
cvt.u64.u32 %rd113, %r883;
add.s64 %rd114, %rd113, %rd19;
shl.b64 %rd116, %rd114, 2;
add.s64 %rd21, %rd101, %rd116;
@%p14 bra $L__BB311_16;

st.global.f32 [%rd21], %f23;

$L__BB311_16:
ld.shared.f32 %f136, [%r16+640];
add.ftz.f32 %f137, %f136, %f157;
ld.shared.f32 %f138, [%r16+1408];
add.ftz.f32 %f139, %f138, %f137;
ld.shared.f32 %f140, [%r16+2176];
add.ftz.f32 %f141, %f140, %f139;
mov.b32 %r901, %f141;
shfl.sync.bfly.b32 %r905|%p41, %r901, %r887, %r886, %r888;
mov.b32 %f142, %r905;
add.ftz.f32 %f143, %f141, %f142;
mov.b32 %r906, %f143;
shfl.sync.bfly.b32 %r908|%p42, %r906, %r891, %r886, %r888;
mov.b32 %f144, %r908;
add.ftz.f32 %f145, %f143, %f144;
mov.b32 %r909, %f145;
shfl.sync.bfly.b32 %r911|%p43, %r909, %r894, %r886, %r888;
mov.b32 %f146, %r911;
add.ftz.f32 %f147, %f145, %f146;
mov.b32 %r912, %f147;
shfl.sync.bfly.b32 %r914|%p44, %r912, %r897, %r886, %r888;
mov.b32 %f148, %r914;
add.ftz.f32 %f149, %f147, %f148;
mov.b32 %r915, %f149;
shfl.sync.bfly.b32 %r917|%p45, %r915, %r884, %r886, %r888;
mov.b32 %f150, %r917;
add.ftz.f32 %f24, %f149, %f150;
@%p22 bra $L__BB311_18;

st.global.f32 [%rd21+4], %f24;

$L__BB311_18:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type21ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type21ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type21ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type21ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type21ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type21ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type21ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type21ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type21ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 128, 1, 1
.minnctapersm 1
{
.reg .pred %p<59>;
.reg .b16 %rs<53>;
.reg .f32 %f<217>;
.reg .b32 %r<1127>;
.reg .b64 %rd<133>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type21ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[3072];

ld.param.u64 %rd22, [_Z13mul_mat_vec_qIL9ggml_type21ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd23, [_Z13mul_mat_vec_qIL9ggml_type21ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd24, [_Z13mul_mat_vec_qIL9ggml_type21ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type21ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r22, [_Z13mul_mat_vec_qIL9ggml_type21ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r23, [_Z13mul_mat_vec_qIL9ggml_type21ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r24, [_Z13mul_mat_vec_qIL9ggml_type21ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type21ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type21ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type21ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type21ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type21ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type21ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type21ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type21ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type21ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r25, [_Z13mul_mat_vec_qIL9ggml_type21ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r38, %tid.y;
shl.b32 %r39, %r38, 5;
mov.u32 %r40, %tid.x;
add.s32 %r1, %r39, %r40;
shr.s32 %r41, %r1, 31;
shr.u32 %r42, %r41, 29;
add.s32 %r43, %r1, %r42;
shr.s32 %r1126, %r43, 3;
shr.u32 %r3, %r37, 8;
setp.ge.s32 %p3, %r1126, %r3;
mov.f32 %f209, 0f00000000;
mov.f32 %f210, %f209;
mov.f32 %f211, %f209;
mov.f32 %f212, %f209;
mov.f32 %f213, %f209;
mov.f32 %f214, %f209;
mov.f32 %f215, %f209;
mov.f32 %f216, %f209;
@%p3 bra $L__BB312_3;

mov.u32 %r44, %ctaid.x;
shl.b32 %r45, %r44, 1;
mov.u32 %r46, %ctaid.y;
mov.u32 %r47, %ctaid.z;
mul.lo.s32 %r48, %r46, %r29;
cvt.u64.u32 %rd25, %r48;
mul.lo.s32 %r49, %r47, %r35;
cvt.u64.u32 %rd26, %r49;
add.s64 %rd1, %rd26, %rd25;
mul.hi.u32 %r50, %r47, %r31;
add.s32 %r51, %r50, %r47;
shr.u32 %r52, %r51, %r32;
mul.lo.s32 %r53, %r52, %r34;
mul.hi.u32 %r54, %r46, %r25;
add.s32 %r55, %r54, %r46;
shr.u32 %r56, %r55, %r26;
mul.lo.s32 %r57, %r56, %r28;
cvta.to.global.u64 %rd132, %rd22;
and.b32 %r61, %r43, -8;
sub.s32 %r62, %r1, %r61;
shl.b32 %r63, %r62, 1;
shr.u32 %r64, %r62, 31;
add.s32 %r65, %r62, %r64;
shr.s32 %r66, %r65, 1;
cvt.s64.s32 %rd27, %r66;
shl.b32 %r67, %r62, 2;
cvt.s64.s32 %rd3, %r62;
and.b32 %r4, %r67, 4;
shl.b32 %r68, %r23, 1;
shl.b32 %r1122, %r1126, 3;
add.s32 %r1125, %r1122, %r68;
mad.lo.s32 %r1124, %r23, 3, %r1122;
add.s32 %r1123, %r23, %r1122;
add.s32 %r69, %r1126, %r53;
add.s32 %r70, %r69, %r57;
or.b32 %r71, %r45, 1;
mad.lo.s32 %r72, %r22, %r71, %r70;
mul.wide.s32 %rd4, %r72, 110;
add.s64 %rd28, %rd4, %rd27;
add.s64 %rd5, %rd28, 106;
add.s64 %rd29, %rd4, 74;
not.b32 %r73, %r63;
mul.wide.s32 %rd30, %r73, 2;
sub.s64 %rd6, %rd29, %rd30;
mul.wide.s32 %rd31, %r63, 2;
add.s64 %rd32, %rd4, %rd31;
add.s64 %rd7, %rd32, 74;
add.s64 %rd33, %rd4, %rd3;
add.s64 %rd8, %rd33, 66;
mad.lo.s32 %r74, %r45, %r22, %r1126;
add.s32 %r75, %r74, %r53;
add.s32 %r76, %r75, %r57;
mul.wide.s32 %rd9, %r76, 110;
mul.wide.s32 %rd34, %r67, 2;
add.s64 %rd35, %rd9, %rd34;
add.s64 %rd10, %rd35, 4;
add.s64 %rd36, %rd4, %rd34;
add.s64 %rd11, %rd36, 2;
add.s64 %rd37, %rd9, %rd27;
add.s64 %rd12, %rd37, 106;
add.s64 %rd38, %rd9, 74;
sub.s64 %rd13, %rd38, %rd30;
add.s64 %rd39, %rd9, %rd31;
add.s64 %rd14, %rd39, 74;
add.s64 %rd40, %rd9, %rd3;
add.s64 %rd15, %rd40, 66;
cvta.to.global.u64 %rd16, %rd23;
mov.f32 %f209, 0f00000000;

$L__BB312_2:
cvt.u64.u32 %rd41, %r1122;
add.s64 %rd42, %rd1, %rd41;
add.s64 %rd43, %rd42, %rd3;
mul.lo.s64 %rd44, %rd43, 36;
add.s64 %rd45, %rd132, %rd10;
ld.global.nc.u16 %rs9, [%rd45];
ld.global.nc.u16 %rs10, [%rd45+2];
ld.global.nc.u16 %rs11, [%rd45+4];
add.s64 %rd46, %rd132, %rd15;
ld.global.nc.u8 %rs12, [%rd46];
add.s64 %rd47, %rd132, %rd14;
add.s64 %rd48, %rd132, %rd13;
ld.global.nc.u16 %rs13, [%rd47];
ld.global.nc.u16 %rs14, [%rd48];
ld.global.nc.u16 %rs15, [%rd45+-2];
and.b16 %rs16, %rs15, 255;
cvt.u32.u16 %r469, %rs16;
shl.b16 %rs17, %rs12, 8;
cvt.u32.u16 %r470, %rs17;
and.b32 %r471, %r470, 256;
or.b32 %r472, %r471, %r469;
mul.wide.u32 %rd49, %r472, 4;
mov.u64 %rd50, iq3s_grid;
add.s64 %rd51, %rd50, %rd49;
ld.global.nc.u32 %r473, [%rd51];
not.b32 %r474, %r473;
shr.u16 %rs18, %rs15, 8;
cvt.u32.u16 %r475, %rs18;
mul.wide.u16 %r476, %rs12, 128;
and.b32 %r477, %r476, 256;
or.b32 %r478, %r477, %r475;
mul.wide.u32 %rd52, %r478, 4;
add.s64 %rd53, %rd50, %rd52;
ld.global.nc.u32 %r479, [%rd53];
not.b32 %r480, %r479;
and.b16 %rs20, %rs13, 255;
mul.wide.u16 %r481, %rs20, 128;
cvt.u32.u16 %r482, %rs13;
shl.b32 %r483, %r482, 21;
or.b32 %r484, %r481, %r483;
and.b32 %r485, %r484, 25166208;
or.b32 %r486, %r485, -2139062144;
add.s32 %r487, %r486, -16843009;
or.b32 %r371, %r487, %r485;
mov.u32 %r77, %r371;

prmt.b32 %r77, %r77, 0, 0xba98;

mul.wide.u16 %r488, %rs20, 8;
shl.b32 %r489, %r482, 17;
or.b32 %r490, %r488, %r489;
and.b32 %r491, %r490, 25166208;
or.b32 %r492, %r491, -2139062144;
add.s32 %r493, %r492, -16843009;
or.b32 %r373, %r493, %r491;
mov.u32 %r79, %r373;

prmt.b32 %r79, %r79, 0, 0xba98;

xor.b32 %r494, %r77, %r473;
or.b32 %r495, %r494, -2139062144;
and.b32 %r496, %r77, 2139062143;
and.b32 %r497, %r474, -2139062144;
sub.s32 %r498, %r495, %r496;
xor.b32 %r82, %r497, %r498;
xor.b32 %r499, %r79, %r479;
or.b32 %r500, %r499, -2139062144;
and.b32 %r501, %r79, 2139062143;
and.b32 %r502, %r480, -2139062144;
sub.s32 %r503, %r500, %r501;
xor.b32 %r86, %r502, %r503;
add.s64 %rd54, %rd16, %rd44;
ld.global.nc.u32 %r83, [%rd54+4];
ld.global.nc.u32 %r87, [%rd54+8];
mov.u32 %r427, 0;

dp4a.s32.s32 %r81, %r82, %r83, %r427;


dp4a.s32.s32 %r85, %r86, %r87, %r81;

and.b16 %rs21, %rs9, 255;
cvt.u32.u16 %r504, %rs21;
mul.wide.u16 %r505, %rs12, 64;
and.b32 %r506, %r505, 256;
or.b32 %r507, %r506, %r504;
mul.wide.u32 %rd55, %r507, 4;
add.s64 %rd56, %rd50, %rd55;
ld.global.nc.u32 %r508, [%rd56];
not.b32 %r509, %r508;
shr.u16 %rs22, %rs9, 8;
cvt.u32.u16 %r510, %rs22;
mul.wide.u16 %r511, %rs12, 32;
and.b32 %r512, %r511, 256;
or.b32 %r513, %r512, %r510;
mul.wide.u32 %rd57, %r513, 4;
add.s64 %rd58, %rd50, %rd57;
ld.global.nc.u32 %r514, [%rd58];
not.b32 %r515, %r514;
shr.u16 %rs23, %rs13, 8;
mul.wide.u16 %r516, %rs23, 128;
cvt.u32.u16 %r517, %rs23;
shl.b32 %r518, %r517, 21;
or.b32 %r519, %r516, %r518;
and.b32 %r520, %r519, 25166208;
or.b32 %r521, %r520, -2139062144;
add.s32 %r522, %r521, -16843009;
or.b32 %r383, %r522, %r520;
mov.u32 %r89, %r383;

prmt.b32 %r89, %r89, 0, 0xba98;

mul.wide.u16 %r523, %rs23, 8;
shl.b32 %r524, %r517, 17;
or.b32 %r525, %r523, %r524;
and.b32 %r526, %r525, 25166208;
or.b32 %r527, %r526, -2139062144;
add.s32 %r528, %r527, -16843009;
or.b32 %r385, %r528, %r526;
mov.u32 %r91, %r385;

prmt.b32 %r91, %r91, 0, 0xba98;

xor.b32 %r529, %r89, %r508;
or.b32 %r530, %r529, -2139062144;
and.b32 %r531, %r89, 2139062143;
and.b32 %r532, %r509, -2139062144;
sub.s32 %r533, %r530, %r531;
xor.b32 %r94, %r532, %r533;
xor.b32 %r534, %r91, %r514;
or.b32 %r535, %r534, -2139062144;
and.b32 %r536, %r91, 2139062143;
and.b32 %r537, %r515, -2139062144;
sub.s32 %r538, %r535, %r536;
xor.b32 %r98, %r537, %r538;
ld.global.nc.u32 %r95, [%rd54+12];
ld.global.nc.u32 %r99, [%rd54+16];

dp4a.s32.s32 %r93, %r94, %r95, %r85;


dp4a.s32.s32 %r97, %r98, %r99, %r93;

and.b16 %rs24, %rs10, 255;
cvt.u32.u16 %r539, %rs24;
mul.wide.u16 %r540, %rs12, 16;
and.b32 %r541, %r540, 256;
or.b32 %r542, %r541, %r539;
mul.wide.u32 %rd59, %r542, 4;
add.s64 %rd60, %rd50, %rd59;
ld.global.nc.u32 %r543, [%rd60];
not.b32 %r544, %r543;
shr.u16 %rs25, %rs10, 8;
cvt.u32.u16 %r545, %rs25;
mul.wide.u16 %r546, %rs12, 8;
and.b32 %r547, %r546, 256;
or.b32 %r548, %r547, %r545;
mul.wide.u32 %rd61, %r548, 4;
add.s64 %rd62, %rd50, %rd61;
ld.global.nc.u32 %r549, [%rd62];
not.b32 %r550, %r549;
and.b16 %rs26, %rs14, 255;
mul.wide.u16 %r551, %rs26, 128;
cvt.u32.u16 %r552, %rs14;
shl.b32 %r553, %r552, 21;
or.b32 %r554, %r551, %r553;
and.b32 %r555, %r554, 25166208;
or.b32 %r556, %r555, -2139062144;
add.s32 %r557, %r556, -16843009;
or.b32 %r395, %r557, %r555;
mov.u32 %r101, %r395;

prmt.b32 %r101, %r101, 0, 0xba98;

mul.wide.u16 %r558, %rs26, 8;
shl.b32 %r559, %r552, 17;
or.b32 %r560, %r558, %r559;
and.b32 %r561, %r560, 25166208;
or.b32 %r562, %r561, -2139062144;
add.s32 %r563, %r562, -16843009;
or.b32 %r397, %r563, %r561;
mov.u32 %r103, %r397;

prmt.b32 %r103, %r103, 0, 0xba98;

xor.b32 %r564, %r101, %r543;
or.b32 %r565, %r564, -2139062144;
and.b32 %r566, %r101, 2139062143;
and.b32 %r567, %r544, -2139062144;
sub.s32 %r568, %r565, %r566;
xor.b32 %r106, %r567, %r568;
xor.b32 %r569, %r103, %r549;
or.b32 %r570, %r569, -2139062144;
and.b32 %r571, %r103, 2139062143;
and.b32 %r572, %r550, -2139062144;
sub.s32 %r573, %r570, %r571;
xor.b32 %r110, %r572, %r573;
ld.global.nc.u32 %r107, [%rd54+20];
ld.global.nc.u32 %r111, [%rd54+24];

dp4a.s32.s32 %r105, %r106, %r107, %r97;


dp4a.s32.s32 %r109, %r110, %r111, %r105;

and.b16 %rs27, %rs11, 255;
cvt.u32.u16 %r574, %rs27;
mul.wide.u16 %r575, %rs12, 4;
and.b32 %r576, %r575, 256;
or.b32 %r577, %r576, %r574;
mul.wide.u32 %rd63, %r577, 4;
add.s64 %rd64, %rd50, %rd63;
ld.global.nc.u32 %r578, [%rd64];
not.b32 %r579, %r578;
shr.u16 %rs28, %rs11, 8;
cvt.u32.u16 %r580, %rs28;
mul.wide.u16 %r581, %rs12, 2;
and.b32 %r582, %r581, 256;
or.b32 %r583, %r582, %r580;
mul.wide.u32 %rd65, %r583, 4;
add.s64 %rd66, %rd50, %rd65;
ld.global.nc.u32 %r584, [%rd66];
not.b32 %r585, %r584;
shr.u16 %rs29, %rs14, 8;
mul.wide.u16 %r586, %rs29, 128;
cvt.u32.u16 %r587, %rs29;
shl.b32 %r588, %r587, 21;
or.b32 %r589, %r586, %r588;
and.b32 %r590, %r589, 25166208;
or.b32 %r591, %r590, -2139062144;
add.s32 %r592, %r591, -16843009;
or.b32 %r407, %r592, %r590;
mov.u32 %r113, %r407;

prmt.b32 %r113, %r113, 0, 0xba98;

mul.wide.u16 %r593, %rs29, 8;
shl.b32 %r594, %r587, 17;
or.b32 %r595, %r593, %r594;
and.b32 %r596, %r595, 25166208;
or.b32 %r597, %r596, -2139062144;
add.s32 %r598, %r597, -16843009;
or.b32 %r409, %r598, %r596;
mov.u32 %r115, %r409;

prmt.b32 %r115, %r115, 0, 0xba98;

xor.b32 %r599, %r113, %r578;
or.b32 %r600, %r599, -2139062144;
and.b32 %r601, %r113, 2139062143;
and.b32 %r602, %r579, -2139062144;
sub.s32 %r603, %r600, %r601;
xor.b32 %r118, %r602, %r603;
xor.b32 %r604, %r115, %r584;
or.b32 %r605, %r604, -2139062144;
and.b32 %r606, %r115, 2139062143;
and.b32 %r607, %r585, -2139062144;
sub.s32 %r608, %r605, %r606;
xor.b32 %r122, %r607, %r608;
ld.global.nc.u32 %r119, [%rd54+28];
ld.global.nc.u32 %r123, [%rd54+32];

dp4a.s32.s32 %r117, %r118, %r119, %r109;


dp4a.s32.s32 %r121, %r122, %r123, %r117;

add.s64 %rd67, %rd132, %rd12;
ld.global.nc.u8 %rs30, [%rd67];
cvt.u32.u16 %r609, %rs30;
and.b32 %r610, %r609, 255;
shr.u32 %r611, %r610, %r4;
shl.b32 %r612, %r611, 1;
and.b32 %r613, %r612, 30;
or.b32 %r614, %r613, 1;
mul.lo.s32 %r615, %r614, %r121;
add.s64 %rd68, %rd132, %rd9;
ld.global.nc.u16 %rs1, [%rd68];

{ cvt.f32.f16 %f49, %rs1;}


ld.global.nc.u32 %r125, [%rd54];

{.reg .f16 low,high;
mov.b32 {low,high},%r125;
cvt.f32.f16 %f50, low;}


mul.ftz.f32 %f65, %f49, %f50;
cvt.rn.f32.s32 %f66, %r615;
fma.rn.ftz.f32 %f216, %f65, %f66, %f216;
add.s64 %rd69, %rd132, %rd11;
ld.global.nc.u16 %rs31, [%rd69+2];
ld.global.nc.u16 %rs32, [%rd69+4];
ld.global.nc.u16 %rs33, [%rd69+6];
add.s64 %rd70, %rd132, %rd8;
ld.global.nc.u8 %rs34, [%rd70];
add.s64 %rd71, %rd132, %rd7;
add.s64 %rd72, %rd132, %rd6;
ld.global.nc.u16 %rs35, [%rd71];
ld.global.nc.u16 %rs36, [%rd72];
ld.global.nc.u16 %rs37, [%rd69];
and.b16 %rs38, %rs37, 255;
cvt.u32.u16 %r616, %rs38;
shl.b16 %rs39, %rs34, 8;
cvt.u32.u16 %r617, %rs39;
and.b32 %r618, %r617, 256;
or.b32 %r619, %r618, %r616;
mul.wide.u32 %rd73, %r619, 4;
add.s64 %rd74, %rd50, %rd73;
ld.global.nc.u32 %r620, [%rd74];
not.b32 %r621, %r620;
shr.u16 %rs40, %rs37, 8;
cvt.u32.u16 %r622, %rs40;
mul.wide.u16 %r623, %rs34, 128;
and.b32 %r624, %r623, 256;
or.b32 %r625, %r624, %r622;
mul.wide.u32 %rd75, %r625, 4;
add.s64 %rd76, %rd50, %rd75;
ld.global.nc.u32 %r626, [%rd76];
not.b32 %r627, %r626;
and.b16 %rs42, %rs35, 255;
mul.wide.u16 %r628, %rs42, 128;
cvt.u32.u16 %r629, %rs35;
shl.b32 %r630, %r629, 21;
or.b32 %r631, %r628, %r630;
and.b32 %r632, %r631, 25166208;
or.b32 %r633, %r632, -2139062144;
add.s32 %r634, %r633, -16843009;
or.b32 %r420, %r634, %r632;
mov.u32 %r126, %r420;

prmt.b32 %r126, %r126, 0, 0xba98;

mul.wide.u16 %r635, %rs42, 8;
shl.b32 %r636, %r629, 17;
or.b32 %r637, %r635, %r636;
and.b32 %r638, %r637, 25166208;
or.b32 %r639, %r638, -2139062144;
add.s32 %r640, %r639, -16843009;
or.b32 %r422, %r640, %r638;
mov.u32 %r128, %r422;

prmt.b32 %r128, %r128, 0, 0xba98;

xor.b32 %r641, %r126, %r620;
or.b32 %r642, %r641, -2139062144;
and.b32 %r643, %r126, 2139062143;
and.b32 %r644, %r621, -2139062144;
sub.s32 %r645, %r642, %r643;
xor.b32 %r131, %r644, %r645;
xor.b32 %r646, %r128, %r626;
or.b32 %r647, %r646, -2139062144;
and.b32 %r648, %r128, 2139062143;
and.b32 %r649, %r627, -2139062144;
sub.s32 %r650, %r647, %r648;
xor.b32 %r135, %r649, %r650;

dp4a.s32.s32 %r130, %r131, %r83, %r427;


dp4a.s32.s32 %r134, %r135, %r87, %r130;

and.b16 %rs43, %rs31, 255;
cvt.u32.u16 %r651, %rs43;
mul.wide.u16 %r652, %rs34, 64;
and.b32 %r653, %r652, 256;
or.b32 %r654, %r653, %r651;
mul.wide.u32 %rd77, %r654, 4;
add.s64 %rd78, %rd50, %rd77;
ld.global.nc.u32 %r655, [%rd78];
not.b32 %r656, %r655;
shr.u16 %rs44, %rs31, 8;
cvt.u32.u16 %r657, %rs44;
mul.wide.u16 %r658, %rs34, 32;
and.b32 %r659, %r658, 256;
or.b32 %r660, %r659, %r657;
mul.wide.u32 %rd79, %r660, 4;
add.s64 %rd80, %rd50, %rd79;
ld.global.nc.u32 %r661, [%rd80];
not.b32 %r662, %r661;
shr.u16 %rs45, %rs35, 8;
mul.wide.u16 %r663, %rs45, 128;
cvt.u32.u16 %r664, %rs45;
shl.b32 %r665, %r664, 21;
or.b32 %r666, %r663, %r665;
and.b32 %r667, %r666, 25166208;
or.b32 %r668, %r667, -2139062144;
add.s32 %r669, %r668, -16843009;
or.b32 %r432, %r669, %r667;
mov.u32 %r138, %r432;

prmt.b32 %r138, %r138, 0, 0xba98;

mul.wide.u16 %r670, %rs45, 8;
shl.b32 %r671, %r664, 17;
or.b32 %r672, %r670, %r671;
and.b32 %r673, %r672, 25166208;
or.b32 %r674, %r673, -2139062144;
add.s32 %r675, %r674, -16843009;
or.b32 %r434, %r675, %r673;
mov.u32 %r140, %r434;

prmt.b32 %r140, %r140, 0, 0xba98;

xor.b32 %r676, %r138, %r655;
or.b32 %r677, %r676, -2139062144;
and.b32 %r678, %r138, 2139062143;
and.b32 %r679, %r656, -2139062144;
sub.s32 %r680, %r677, %r678;
xor.b32 %r143, %r679, %r680;
xor.b32 %r681, %r140, %r661;
or.b32 %r682, %r681, -2139062144;
and.b32 %r683, %r140, 2139062143;
and.b32 %r684, %r662, -2139062144;
sub.s32 %r685, %r682, %r683;
xor.b32 %r147, %r684, %r685;

dp4a.s32.s32 %r142, %r143, %r95, %r134;


dp4a.s32.s32 %r146, %r147, %r99, %r142;

and.b16 %rs46, %rs32, 255;
cvt.u32.u16 %r686, %rs46;
mul.wide.u16 %r687, %rs34, 16;
and.b32 %r688, %r687, 256;
or.b32 %r689, %r688, %r686;
mul.wide.u32 %rd81, %r689, 4;
add.s64 %rd82, %rd50, %rd81;
ld.global.nc.u32 %r690, [%rd82];
not.b32 %r691, %r690;
shr.u16 %rs47, %rs32, 8;
cvt.u32.u16 %r692, %rs47;
mul.wide.u16 %r693, %rs34, 8;
and.b32 %r694, %r693, 256;
or.b32 %r695, %r694, %r692;
mul.wide.u32 %rd83, %r695, 4;
add.s64 %rd84, %rd50, %rd83;
ld.global.nc.u32 %r696, [%rd84];
not.b32 %r697, %r696;
and.b16 %rs48, %rs36, 255;
mul.wide.u16 %r698, %rs48, 128;
cvt.u32.u16 %r699, %rs36;
shl.b32 %r700, %r699, 21;
or.b32 %r701, %r698, %r700;
and.b32 %r702, %r701, 25166208;
or.b32 %r703, %r702, -2139062144;
add.s32 %r704, %r703, -16843009;
or.b32 %r444, %r704, %r702;
mov.u32 %r150, %r444;

prmt.b32 %r150, %r150, 0, 0xba98;

mul.wide.u16 %r705, %rs48, 8;
shl.b32 %r706, %r699, 17;
or.b32 %r707, %r705, %r706;
and.b32 %r708, %r707, 25166208;
or.b32 %r709, %r708, -2139062144;
add.s32 %r710, %r709, -16843009;
or.b32 %r446, %r710, %r708;
mov.u32 %r152, %r446;

prmt.b32 %r152, %r152, 0, 0xba98;

xor.b32 %r711, %r150, %r690;
or.b32 %r712, %r711, -2139062144;
and.b32 %r713, %r150, 2139062143;
and.b32 %r714, %r691, -2139062144;
sub.s32 %r715, %r712, %r713;
xor.b32 %r155, %r714, %r715;
xor.b32 %r716, %r152, %r696;
or.b32 %r717, %r716, -2139062144;
and.b32 %r718, %r152, 2139062143;
and.b32 %r719, %r697, -2139062144;
sub.s32 %r720, %r717, %r718;
xor.b32 %r159, %r719, %r720;

dp4a.s32.s32 %r154, %r155, %r107, %r146;


dp4a.s32.s32 %r158, %r159, %r111, %r154;

and.b16 %rs49, %rs33, 255;
cvt.u32.u16 %r721, %rs49;
mul.wide.u16 %r722, %rs34, 4;
and.b32 %r723, %r722, 256;
or.b32 %r724, %r723, %r721;
mul.wide.u32 %rd85, %r724, 4;
add.s64 %rd86, %rd50, %rd85;
ld.global.nc.u32 %r725, [%rd86];
not.b32 %r726, %r725;
shr.u16 %rs50, %rs33, 8;
cvt.u32.u16 %r727, %rs50;
mul.wide.u16 %r728, %rs34, 2;
and.b32 %r729, %r728, 256;
or.b32 %r730, %r729, %r727;
mul.wide.u32 %rd87, %r730, 4;
add.s64 %rd88, %rd50, %rd87;
ld.global.nc.u32 %r731, [%rd88];
not.b32 %r732, %r731;
shr.u16 %rs51, %rs36, 8;
mul.wide.u16 %r733, %rs51, 128;
cvt.u32.u16 %r734, %rs51;
shl.b32 %r735, %r734, 21;
or.b32 %r736, %r733, %r735;
and.b32 %r737, %r736, 25166208;
or.b32 %r738, %r737, -2139062144;
add.s32 %r739, %r738, -16843009;
or.b32 %r456, %r739, %r737;
mov.u32 %r162, %r456;

prmt.b32 %r162, %r162, 0, 0xba98;

mul.wide.u16 %r740, %rs51, 8;
shl.b32 %r741, %r734, 17;
or.b32 %r742, %r740, %r741;
and.b32 %r743, %r742, 25166208;
or.b32 %r744, %r743, -2139062144;
add.s32 %r745, %r744, -16843009;
or.b32 %r458, %r745, %r743;
mov.u32 %r164, %r458;

prmt.b32 %r164, %r164, 0, 0xba98;

xor.b32 %r746, %r162, %r725;
or.b32 %r747, %r746, -2139062144;
and.b32 %r748, %r162, 2139062143;
and.b32 %r749, %r726, -2139062144;
sub.s32 %r750, %r747, %r748;
xor.b32 %r167, %r749, %r750;
xor.b32 %r751, %r164, %r731;
or.b32 %r752, %r751, -2139062144;
and.b32 %r753, %r164, 2139062143;
and.b32 %r754, %r732, -2139062144;
sub.s32 %r755, %r752, %r753;
xor.b32 %r171, %r754, %r755;

dp4a.s32.s32 %r166, %r167, %r119, %r158;


dp4a.s32.s32 %r170, %r171, %r123, %r166;

add.s64 %rd89, %rd132, %rd5;
ld.global.nc.u8 %rs52, [%rd89];
cvt.u32.u16 %r756, %rs52;
and.b32 %r757, %r756, 255;
shr.u32 %r758, %r757, %r4;
shl.b32 %r759, %r758, 1;
and.b32 %r760, %r759, 30;
or.b32 %r761, %r760, 1;
mul.lo.s32 %r762, %r761, %r170;
add.s64 %rd90, %rd132, %rd4;
ld.global.nc.u16 %rs2, [%rd90];

{ cvt.f32.f16 %f51, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r125;
cvt.f32.f16 %f52, low;}


mul.ftz.f32 %f67, %f51, %f52;
cvt.rn.f32.s32 %f68, %r762;
fma.rn.ftz.f32 %f215, %f67, %f68, %f215;
cvt.u64.u32 %rd91, %r1123;
add.s64 %rd92, %rd1, %rd91;
add.s64 %rd93, %rd92, %rd3;
mul.lo.s64 %rd94, %rd93, 36;
mov.u32 %r175, %r371;

prmt.b32 %r175, %r175, 0, 0xba98;

mov.u32 %r177, %r373;

prmt.b32 %r177, %r177, 0, 0xba98;

xor.b32 %r763, %r175, %r473;
or.b32 %r764, %r763, -2139062144;
and.b32 %r765, %r175, 2139062143;
sub.s32 %r766, %r764, %r765;
xor.b32 %r180, %r497, %r766;
xor.b32 %r767, %r177, %r479;
or.b32 %r768, %r767, -2139062144;
and.b32 %r769, %r177, 2139062143;
sub.s32 %r770, %r768, %r769;
xor.b32 %r184, %r502, %r770;
add.s64 %rd95, %rd16, %rd94;
ld.global.nc.u32 %r181, [%rd95+4];
ld.global.nc.u32 %r185, [%rd95+8];

dp4a.s32.s32 %r179, %r180, %r181, %r427;


dp4a.s32.s32 %r183, %r184, %r185, %r179;

mov.u32 %r187, %r383;

prmt.b32 %r187, %r187, 0, 0xba98;

mov.u32 %r189, %r385;

prmt.b32 %r189, %r189, 0, 0xba98;

xor.b32 %r771, %r187, %r508;
or.b32 %r772, %r771, -2139062144;
and.b32 %r773, %r187, 2139062143;
sub.s32 %r774, %r772, %r773;
xor.b32 %r192, %r532, %r774;
xor.b32 %r775, %r189, %r514;
or.b32 %r776, %r775, -2139062144;
and.b32 %r777, %r189, 2139062143;
sub.s32 %r778, %r776, %r777;
xor.b32 %r196, %r537, %r778;
ld.global.nc.u32 %r193, [%rd95+12];
ld.global.nc.u32 %r197, [%rd95+16];

dp4a.s32.s32 %r191, %r192, %r193, %r183;


dp4a.s32.s32 %r195, %r196, %r197, %r191;

mov.u32 %r199, %r395;

prmt.b32 %r199, %r199, 0, 0xba98;

mov.u32 %r201, %r397;

prmt.b32 %r201, %r201, 0, 0xba98;

xor.b32 %r779, %r199, %r543;
or.b32 %r780, %r779, -2139062144;
and.b32 %r781, %r199, 2139062143;
sub.s32 %r782, %r780, %r781;
xor.b32 %r204, %r567, %r782;
xor.b32 %r783, %r201, %r549;
or.b32 %r784, %r783, -2139062144;
and.b32 %r785, %r201, 2139062143;
sub.s32 %r786, %r784, %r785;
xor.b32 %r208, %r572, %r786;
ld.global.nc.u32 %r205, [%rd95+20];
ld.global.nc.u32 %r209, [%rd95+24];

dp4a.s32.s32 %r203, %r204, %r205, %r195;


dp4a.s32.s32 %r207, %r208, %r209, %r203;

mov.u32 %r211, %r407;

prmt.b32 %r211, %r211, 0, 0xba98;

mov.u32 %r213, %r409;

prmt.b32 %r213, %r213, 0, 0xba98;

xor.b32 %r787, %r211, %r578;
or.b32 %r788, %r787, -2139062144;
and.b32 %r789, %r211, 2139062143;
sub.s32 %r790, %r788, %r789;
xor.b32 %r216, %r602, %r790;
xor.b32 %r791, %r213, %r584;
or.b32 %r792, %r791, -2139062144;
and.b32 %r793, %r213, 2139062143;
sub.s32 %r794, %r792, %r793;
xor.b32 %r220, %r607, %r794;
ld.global.nc.u32 %r217, [%rd95+28];
ld.global.nc.u32 %r221, [%rd95+32];

dp4a.s32.s32 %r215, %r216, %r217, %r207;


dp4a.s32.s32 %r219, %r220, %r221, %r215;

mul.lo.s32 %r795, %r614, %r219;

{ cvt.f32.f16 %f53, %rs1;}


ld.global.nc.u32 %r223, [%rd95];

{.reg .f16 low,high;
mov.b32 {low,high},%r223;
cvt.f32.f16 %f54, low;}


mul.ftz.f32 %f69, %f53, %f54;
cvt.rn.f32.s32 %f70, %r795;
fma.rn.ftz.f32 %f214, %f69, %f70, %f214;
mov.u32 %r224, %r420;

prmt.b32 %r224, %r224, 0, 0xba98;

mov.u32 %r226, %r422;

prmt.b32 %r226, %r226, 0, 0xba98;

xor.b32 %r796, %r224, %r620;
or.b32 %r797, %r796, -2139062144;
and.b32 %r798, %r224, 2139062143;
sub.s32 %r799, %r797, %r798;
xor.b32 %r229, %r644, %r799;
xor.b32 %r800, %r226, %r626;
or.b32 %r801, %r800, -2139062144;
and.b32 %r802, %r226, 2139062143;
sub.s32 %r803, %r801, %r802;
xor.b32 %r233, %r649, %r803;

dp4a.s32.s32 %r228, %r229, %r181, %r427;


dp4a.s32.s32 %r232, %r233, %r185, %r228;

mov.u32 %r236, %r432;

prmt.b32 %r236, %r236, 0, 0xba98;

mov.u32 %r238, %r434;

prmt.b32 %r238, %r238, 0, 0xba98;

xor.b32 %r804, %r236, %r655;
or.b32 %r805, %r804, -2139062144;
and.b32 %r806, %r236, 2139062143;
sub.s32 %r807, %r805, %r806;
xor.b32 %r241, %r679, %r807;
xor.b32 %r808, %r238, %r661;
or.b32 %r809, %r808, -2139062144;
and.b32 %r810, %r238, 2139062143;
sub.s32 %r811, %r809, %r810;
xor.b32 %r245, %r684, %r811;

dp4a.s32.s32 %r240, %r241, %r193, %r232;


dp4a.s32.s32 %r244, %r245, %r197, %r240;

mov.u32 %r248, %r444;

prmt.b32 %r248, %r248, 0, 0xba98;

mov.u32 %r250, %r446;

prmt.b32 %r250, %r250, 0, 0xba98;

xor.b32 %r812, %r248, %r690;
or.b32 %r813, %r812, -2139062144;
and.b32 %r814, %r248, 2139062143;
sub.s32 %r815, %r813, %r814;
xor.b32 %r253, %r714, %r815;
xor.b32 %r816, %r250, %r696;
or.b32 %r817, %r816, -2139062144;
and.b32 %r818, %r250, 2139062143;
sub.s32 %r819, %r817, %r818;
xor.b32 %r257, %r719, %r819;

dp4a.s32.s32 %r252, %r253, %r205, %r244;


dp4a.s32.s32 %r256, %r257, %r209, %r252;

mov.u32 %r260, %r456;

prmt.b32 %r260, %r260, 0, 0xba98;

mov.u32 %r262, %r458;

prmt.b32 %r262, %r262, 0, 0xba98;

xor.b32 %r820, %r260, %r725;
or.b32 %r821, %r820, -2139062144;
and.b32 %r822, %r260, 2139062143;
sub.s32 %r823, %r821, %r822;
xor.b32 %r265, %r749, %r823;
xor.b32 %r824, %r262, %r731;
or.b32 %r825, %r824, -2139062144;
and.b32 %r826, %r262, 2139062143;
sub.s32 %r827, %r825, %r826;
xor.b32 %r269, %r754, %r827;

dp4a.s32.s32 %r264, %r265, %r217, %r256;


dp4a.s32.s32 %r268, %r269, %r221, %r264;

mul.lo.s32 %r828, %r761, %r268;

{ cvt.f32.f16 %f55, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r223;
cvt.f32.f16 %f56, low;}


mul.ftz.f32 %f71, %f55, %f56;
cvt.rn.f32.s32 %f72, %r828;
fma.rn.ftz.f32 %f213, %f71, %f72, %f213;
cvt.u64.u32 %rd96, %r1125;
add.s64 %rd97, %rd1, %rd96;
add.s64 %rd98, %rd97, %rd3;
mul.lo.s64 %rd99, %rd98, 36;
mov.u32 %r273, %r371;

prmt.b32 %r273, %r273, 0, 0xba98;

mov.u32 %r275, %r373;

prmt.b32 %r275, %r275, 0, 0xba98;

xor.b32 %r829, %r273, %r473;
or.b32 %r830, %r829, -2139062144;
and.b32 %r831, %r273, 2139062143;
sub.s32 %r832, %r830, %r831;
xor.b32 %r278, %r497, %r832;
xor.b32 %r833, %r275, %r479;
or.b32 %r834, %r833, -2139062144;
and.b32 %r835, %r275, 2139062143;
sub.s32 %r836, %r834, %r835;
xor.b32 %r282, %r502, %r836;
add.s64 %rd100, %rd16, %rd99;
ld.global.nc.u32 %r279, [%rd100+4];
ld.global.nc.u32 %r283, [%rd100+8];

dp4a.s32.s32 %r277, %r278, %r279, %r427;


dp4a.s32.s32 %r281, %r282, %r283, %r277;

mov.u32 %r285, %r383;

prmt.b32 %r285, %r285, 0, 0xba98;

mov.u32 %r287, %r385;

prmt.b32 %r287, %r287, 0, 0xba98;

xor.b32 %r837, %r285, %r508;
or.b32 %r838, %r837, -2139062144;
and.b32 %r839, %r285, 2139062143;
sub.s32 %r840, %r838, %r839;
xor.b32 %r290, %r532, %r840;
xor.b32 %r841, %r287, %r514;
or.b32 %r842, %r841, -2139062144;
and.b32 %r843, %r287, 2139062143;
sub.s32 %r844, %r842, %r843;
xor.b32 %r294, %r537, %r844;
ld.global.nc.u32 %r291, [%rd100+12];
ld.global.nc.u32 %r295, [%rd100+16];

dp4a.s32.s32 %r289, %r290, %r291, %r281;


dp4a.s32.s32 %r293, %r294, %r295, %r289;

mov.u32 %r297, %r395;

prmt.b32 %r297, %r297, 0, 0xba98;

mov.u32 %r299, %r397;

prmt.b32 %r299, %r299, 0, 0xba98;

xor.b32 %r845, %r297, %r543;
or.b32 %r846, %r845, -2139062144;
and.b32 %r847, %r297, 2139062143;
sub.s32 %r848, %r846, %r847;
xor.b32 %r302, %r567, %r848;
xor.b32 %r849, %r299, %r549;
or.b32 %r850, %r849, -2139062144;
and.b32 %r851, %r299, 2139062143;
sub.s32 %r852, %r850, %r851;
xor.b32 %r306, %r572, %r852;
ld.global.nc.u32 %r303, [%rd100+20];
ld.global.nc.u32 %r307, [%rd100+24];

dp4a.s32.s32 %r301, %r302, %r303, %r293;


dp4a.s32.s32 %r305, %r306, %r307, %r301;

mov.u32 %r309, %r407;

prmt.b32 %r309, %r309, 0, 0xba98;

mov.u32 %r311, %r409;

prmt.b32 %r311, %r311, 0, 0xba98;

xor.b32 %r853, %r309, %r578;
or.b32 %r854, %r853, -2139062144;
and.b32 %r855, %r309, 2139062143;
sub.s32 %r856, %r854, %r855;
xor.b32 %r314, %r602, %r856;
xor.b32 %r857, %r311, %r584;
or.b32 %r858, %r857, -2139062144;
and.b32 %r859, %r311, 2139062143;
sub.s32 %r860, %r858, %r859;
xor.b32 %r318, %r607, %r860;
ld.global.nc.u32 %r315, [%rd100+28];
ld.global.nc.u32 %r319, [%rd100+32];

dp4a.s32.s32 %r313, %r314, %r315, %r305;


dp4a.s32.s32 %r317, %r318, %r319, %r313;

mul.lo.s32 %r861, %r614, %r317;

{ cvt.f32.f16 %f57, %rs1;}


ld.global.nc.u32 %r321, [%rd100];

{.reg .f16 low,high;
mov.b32 {low,high},%r321;
cvt.f32.f16 %f58, low;}


mul.ftz.f32 %f73, %f57, %f58;
cvt.rn.f32.s32 %f74, %r861;
fma.rn.ftz.f32 %f212, %f73, %f74, %f212;
mov.u32 %r322, %r420;

prmt.b32 %r322, %r322, 0, 0xba98;

mov.u32 %r324, %r422;

prmt.b32 %r324, %r324, 0, 0xba98;

xor.b32 %r862, %r322, %r620;
or.b32 %r863, %r862, -2139062144;
and.b32 %r864, %r322, 2139062143;
sub.s32 %r865, %r863, %r864;
xor.b32 %r327, %r644, %r865;
xor.b32 %r866, %r324, %r626;
or.b32 %r867, %r866, -2139062144;
and.b32 %r868, %r324, 2139062143;
sub.s32 %r869, %r867, %r868;
xor.b32 %r331, %r649, %r869;

dp4a.s32.s32 %r326, %r327, %r279, %r427;


dp4a.s32.s32 %r330, %r331, %r283, %r326;

mov.u32 %r334, %r432;

prmt.b32 %r334, %r334, 0, 0xba98;

mov.u32 %r336, %r434;

prmt.b32 %r336, %r336, 0, 0xba98;

xor.b32 %r870, %r334, %r655;
or.b32 %r871, %r870, -2139062144;
and.b32 %r872, %r334, 2139062143;
sub.s32 %r873, %r871, %r872;
xor.b32 %r339, %r679, %r873;
xor.b32 %r874, %r336, %r661;
or.b32 %r875, %r874, -2139062144;
and.b32 %r876, %r336, 2139062143;
sub.s32 %r877, %r875, %r876;
xor.b32 %r343, %r684, %r877;

dp4a.s32.s32 %r338, %r339, %r291, %r330;


dp4a.s32.s32 %r342, %r343, %r295, %r338;

mov.u32 %r346, %r444;

prmt.b32 %r346, %r346, 0, 0xba98;

mov.u32 %r348, %r446;

prmt.b32 %r348, %r348, 0, 0xba98;

xor.b32 %r878, %r346, %r690;
or.b32 %r879, %r878, -2139062144;
and.b32 %r880, %r346, 2139062143;
sub.s32 %r881, %r879, %r880;
xor.b32 %r351, %r714, %r881;
xor.b32 %r882, %r348, %r696;
or.b32 %r883, %r882, -2139062144;
and.b32 %r884, %r348, 2139062143;
sub.s32 %r885, %r883, %r884;
xor.b32 %r355, %r719, %r885;

dp4a.s32.s32 %r350, %r351, %r303, %r342;


dp4a.s32.s32 %r354, %r355, %r307, %r350;

mov.u32 %r358, %r456;

prmt.b32 %r358, %r358, 0, 0xba98;

mov.u32 %r360, %r458;

prmt.b32 %r360, %r360, 0, 0xba98;

xor.b32 %r886, %r358, %r725;
or.b32 %r887, %r886, -2139062144;
and.b32 %r888, %r358, 2139062143;
sub.s32 %r889, %r887, %r888;
xor.b32 %r363, %r749, %r889;
xor.b32 %r890, %r360, %r731;
or.b32 %r891, %r890, -2139062144;
and.b32 %r892, %r360, 2139062143;
sub.s32 %r893, %r891, %r892;
xor.b32 %r367, %r754, %r893;

dp4a.s32.s32 %r362, %r363, %r315, %r354;


dp4a.s32.s32 %r366, %r367, %r319, %r362;

mul.lo.s32 %r894, %r761, %r366;

{ cvt.f32.f16 %f59, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r321;
cvt.f32.f16 %f60, low;}


mul.ftz.f32 %f75, %f59, %f60;
cvt.rn.f32.s32 %f76, %r894;
fma.rn.ftz.f32 %f211, %f75, %f76, %f211;
cvt.u64.u32 %rd101, %r1124;
add.s64 %rd102, %rd1, %rd101;
add.s64 %rd103, %rd102, %rd3;
mul.lo.s64 %rd104, %rd103, 36;

prmt.b32 %r371, %r371, 0, 0xba98;


prmt.b32 %r373, %r373, 0, 0xba98;

xor.b32 %r895, %r371, %r473;
or.b32 %r896, %r895, -2139062144;
and.b32 %r897, %r371, 2139062143;
sub.s32 %r898, %r896, %r897;
xor.b32 %r376, %r497, %r898;
xor.b32 %r899, %r373, %r479;
or.b32 %r900, %r899, -2139062144;
and.b32 %r901, %r373, 2139062143;
sub.s32 %r902, %r900, %r901;
xor.b32 %r380, %r502, %r902;
add.s64 %rd105, %rd16, %rd104;
ld.global.nc.u32 %r377, [%rd105+4];
ld.global.nc.u32 %r381, [%rd105+8];

dp4a.s32.s32 %r375, %r376, %r377, %r427;


dp4a.s32.s32 %r379, %r380, %r381, %r375;


prmt.b32 %r383, %r383, 0, 0xba98;


prmt.b32 %r385, %r385, 0, 0xba98;

xor.b32 %r903, %r383, %r508;
or.b32 %r904, %r903, -2139062144;
and.b32 %r905, %r383, 2139062143;
sub.s32 %r906, %r904, %r905;
xor.b32 %r388, %r532, %r906;
xor.b32 %r907, %r385, %r514;
or.b32 %r908, %r907, -2139062144;
and.b32 %r909, %r385, 2139062143;
sub.s32 %r910, %r908, %r909;
xor.b32 %r392, %r537, %r910;
ld.global.nc.u32 %r389, [%rd105+12];
ld.global.nc.u32 %r393, [%rd105+16];

dp4a.s32.s32 %r387, %r388, %r389, %r379;


dp4a.s32.s32 %r391, %r392, %r393, %r387;


prmt.b32 %r395, %r395, 0, 0xba98;


prmt.b32 %r397, %r397, 0, 0xba98;

xor.b32 %r911, %r395, %r543;
or.b32 %r912, %r911, -2139062144;
and.b32 %r913, %r395, 2139062143;
sub.s32 %r914, %r912, %r913;
xor.b32 %r400, %r567, %r914;
xor.b32 %r915, %r397, %r549;
or.b32 %r916, %r915, -2139062144;
and.b32 %r917, %r397, 2139062143;
sub.s32 %r918, %r916, %r917;
xor.b32 %r404, %r572, %r918;
ld.global.nc.u32 %r401, [%rd105+20];
ld.global.nc.u32 %r405, [%rd105+24];

dp4a.s32.s32 %r399, %r400, %r401, %r391;


dp4a.s32.s32 %r403, %r404, %r405, %r399;


prmt.b32 %r407, %r407, 0, 0xba98;


prmt.b32 %r409, %r409, 0, 0xba98;

xor.b32 %r919, %r407, %r578;
or.b32 %r920, %r919, -2139062144;
and.b32 %r921, %r407, 2139062143;
sub.s32 %r922, %r920, %r921;
xor.b32 %r412, %r602, %r922;
xor.b32 %r923, %r409, %r584;
or.b32 %r924, %r923, -2139062144;
and.b32 %r925, %r409, 2139062143;
sub.s32 %r926, %r924, %r925;
xor.b32 %r416, %r607, %r926;
ld.global.nc.u32 %r413, [%rd105+28];
ld.global.nc.u32 %r417, [%rd105+32];

dp4a.s32.s32 %r411, %r412, %r413, %r403;


dp4a.s32.s32 %r415, %r416, %r417, %r411;

mul.lo.s32 %r927, %r614, %r415;

{ cvt.f32.f16 %f61, %rs1;}


ld.global.nc.u32 %r419, [%rd105];

{.reg .f16 low,high;
mov.b32 {low,high},%r419;
cvt.f32.f16 %f62, low;}


mul.ftz.f32 %f77, %f61, %f62;
cvt.rn.f32.s32 %f78, %r927;
fma.rn.ftz.f32 %f210, %f77, %f78, %f210;

prmt.b32 %r420, %r420, 0, 0xba98;


prmt.b32 %r422, %r422, 0, 0xba98;

xor.b32 %r928, %r420, %r620;
or.b32 %r929, %r928, -2139062144;
and.b32 %r930, %r420, 2139062143;
sub.s32 %r931, %r929, %r930;
xor.b32 %r425, %r644, %r931;
xor.b32 %r932, %r422, %r626;
or.b32 %r933, %r932, -2139062144;
and.b32 %r934, %r422, 2139062143;
sub.s32 %r935, %r933, %r934;
xor.b32 %r429, %r649, %r935;

dp4a.s32.s32 %r424, %r425, %r377, %r427;


dp4a.s32.s32 %r428, %r429, %r381, %r424;


prmt.b32 %r432, %r432, 0, 0xba98;


prmt.b32 %r434, %r434, 0, 0xba98;

xor.b32 %r936, %r432, %r655;
or.b32 %r937, %r936, -2139062144;
and.b32 %r938, %r432, 2139062143;
sub.s32 %r939, %r937, %r938;
xor.b32 %r437, %r679, %r939;
xor.b32 %r940, %r434, %r661;
or.b32 %r941, %r940, -2139062144;
and.b32 %r942, %r434, 2139062143;
sub.s32 %r943, %r941, %r942;
xor.b32 %r441, %r684, %r943;

dp4a.s32.s32 %r436, %r437, %r389, %r428;


dp4a.s32.s32 %r440, %r441, %r393, %r436;


prmt.b32 %r444, %r444, 0, 0xba98;


prmt.b32 %r446, %r446, 0, 0xba98;

xor.b32 %r944, %r444, %r690;
or.b32 %r945, %r944, -2139062144;
and.b32 %r946, %r444, 2139062143;
sub.s32 %r947, %r945, %r946;
xor.b32 %r449, %r714, %r947;
xor.b32 %r948, %r446, %r696;
or.b32 %r949, %r948, -2139062144;
and.b32 %r950, %r446, 2139062143;
sub.s32 %r951, %r949, %r950;
xor.b32 %r453, %r719, %r951;

dp4a.s32.s32 %r448, %r449, %r401, %r440;


dp4a.s32.s32 %r452, %r453, %r405, %r448;


prmt.b32 %r456, %r456, 0, 0xba98;


prmt.b32 %r458, %r458, 0, 0xba98;

xor.b32 %r952, %r456, %r725;
or.b32 %r953, %r952, -2139062144;
and.b32 %r954, %r456, 2139062143;
sub.s32 %r955, %r953, %r954;
xor.b32 %r461, %r749, %r955;
xor.b32 %r956, %r458, %r731;
or.b32 %r957, %r956, -2139062144;
and.b32 %r958, %r458, 2139062143;
sub.s32 %r959, %r957, %r958;
xor.b32 %r465, %r754, %r959;

dp4a.s32.s32 %r460, %r461, %r413, %r452;


dp4a.s32.s32 %r464, %r465, %r417, %r460;

mul.lo.s32 %r960, %r761, %r464;

{ cvt.f32.f16 %f63, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r419;
cvt.f32.f16 %f64, low;}


mul.ftz.f32 %f79, %f63, %f64;
cvt.rn.f32.s32 %f80, %r960;
fma.rn.ftz.f32 %f209, %f79, %f80, %f209;
add.s32 %r1125, %r1125, 128;
add.s32 %r1124, %r1124, 128;
add.s32 %r1123, %r1123, 128;
add.s32 %r1122, %r1122, 128;
add.s64 %rd132, %rd132, 1760;
add.s32 %r1126, %r1126, 16;
setp.lt.s32 %p4, %r1126, %r3;
@%p4 bra $L__BB312_2;

$L__BB312_3:
setp.eq.s32 %p5, %r38, 0;
@%p5 bra $L__BB312_5;

shl.b32 %r964, %r38, 10;
mov.u32 %r965, _ZZ13mul_mat_vec_qIL9ggml_type21ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r966, %r965, %r964;
shl.b32 %r967, %r40, 2;
add.s32 %r968, %r966, %r967;
st.shared.f32 [%r968+-1024], %f216;
st.shared.f32 [%r968+-896], %f215;
st.shared.f32 [%r968+-768], %f214;
st.shared.f32 [%r968+-640], %f213;
st.shared.f32 [%r968+-512], %f212;
st.shared.f32 [%r968+-384], %f211;
st.shared.f32 [%r968+-256], %f210;
st.shared.f32 [%r968+-128], %f209;

$L__BB312_5:
setp.ne.s32 %p6, %r38, 0;
bar.sync 0;
@%p6 bra $L__BB312_22;

mov.u32 %r970, %ctaid.z;
mov.u32 %r971, %ctaid.y;
mov.u32 %r972, %ctaid.x;
shl.b32 %r973, %r972, 1;
mov.u32 %r974, 1;
mad.lo.s32 %r975, %r971, %r30, %r973;
mad.lo.s32 %r976, %r970, %r36, %r975;
cvt.u64.u32 %rd19, %r976;
shl.b32 %r978, %r40, 2;
mov.u32 %r979, _ZZ13mul_mat_vec_qIL9ggml_type21ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r19, %r979, %r978;
mov.u32 %r980, 2;
ld.shared.f32 %f81, [%r19];
add.ftz.f32 %f82, %f81, %f216;
ld.shared.f32 %f83, [%r19+1024];
add.ftz.f32 %f84, %f83, %f82;
ld.shared.f32 %f85, [%r19+2048];
add.ftz.f32 %f86, %f85, %f84;
mov.b32 %r981, %f86;
mov.u32 %r982, 31;
mov.u32 %r983, 16;
mov.u32 %r984, -1;
shfl.sync.bfly.b32 %r985|%p7, %r981, %r983, %r982, %r984;
mov.b32 %f87, %r985;
add.ftz.f32 %f88, %f86, %f87;
mov.b32 %r986, %f88;
mov.u32 %r987, 8;
shfl.sync.bfly.b32 %r988|%p8, %r986, %r987, %r982, %r984;
mov.b32 %f89, %r988;
add.ftz.f32 %f90, %f88, %f89;
mov.b32 %r989, %f90;
mov.u32 %r990, 4;
shfl.sync.bfly.b32 %r991|%p9, %r989, %r990, %r982, %r984;
mov.b32 %f91, %r991;
add.ftz.f32 %f92, %f90, %f91;
mov.b32 %r992, %f92;
shfl.sync.bfly.b32 %r993|%p10, %r992, %r980, %r982, %r984;
mov.b32 %f93, %r993;
add.ftz.f32 %f94, %f92, %f93;
mov.b32 %r994, %f94;
shfl.sync.bfly.b32 %r995|%p11, %r994, %r974, %r982, %r984;
mov.b32 %f95, %r995;
add.ftz.f32 %f25, %f94, %f95;
setp.eq.s32 %p12, %r40, 0;
setp.lt.u32 %p13, %r973, %r24;
and.pred %p1, %p12, %p13;
cvta.to.global.u64 %rd106, %rd24;
mul.wide.u32 %rd107, %r976, 4;
add.s64 %rd20, %rd106, %rd107;
not.pred %p14, %p1;
@%p14 bra $L__BB312_8;

st.global.f32 [%rd20], %f25;

$L__BB312_8:
ld.shared.f32 %f96, [%r19+128];
add.ftz.f32 %f97, %f96, %f215;
ld.shared.f32 %f98, [%r19+1152];
add.ftz.f32 %f99, %f98, %f97;
ld.shared.f32 %f100, [%r19+2176];
add.ftz.f32 %f101, %f100, %f99;
mov.b32 %r996, %f101;
shfl.sync.bfly.b32 %r1000|%p15, %r996, %r983, %r982, %r984;
mov.b32 %f102, %r1000;
add.ftz.f32 %f103, %f101, %f102;
mov.b32 %r1001, %f103;
shfl.sync.bfly.b32 %r1003|%p16, %r1001, %r987, %r982, %r984;
mov.b32 %f104, %r1003;
add.ftz.f32 %f105, %f103, %f104;
mov.b32 %r1004, %f105;
shfl.sync.bfly.b32 %r1006|%p17, %r1004, %r990, %r982, %r984;
mov.b32 %f106, %r1006;
add.ftz.f32 %f107, %f105, %f106;
mov.b32 %r1007, %f107;
shfl.sync.bfly.b32 %r1009|%p18, %r1007, %r980, %r982, %r984;
mov.b32 %f108, %r1009;
add.ftz.f32 %f109, %f107, %f108;
mov.b32 %r1010, %f109;
shfl.sync.bfly.b32 %r1012|%p19, %r1010, %r974, %r982, %r984;
mov.b32 %f110, %r1012;
add.ftz.f32 %f26, %f109, %f110;
setp.eq.s32 %p20, %r40, 1;
or.b32 %r1016, %r973, 1;
setp.lt.u32 %p21, %r1016, %r24;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB312_10;

st.global.f32 [%rd20+4], %f26;

$L__BB312_10:
ld.shared.f32 %f111, [%r19+256];
add.ftz.f32 %f112, %f111, %f214;
ld.shared.f32 %f113, [%r19+1280];
add.ftz.f32 %f114, %f113, %f112;
ld.shared.f32 %f115, [%r19+2304];
add.ftz.f32 %f116, %f115, %f114;
mov.b32 %r1017, %f116;
mov.u32 %r1018, 31;
mov.u32 %r1019, 16;
mov.u32 %r1020, -1;
shfl.sync.bfly.b32 %r1021|%p23, %r1017, %r1019, %r1018, %r1020;
mov.b32 %f117, %r1021;
add.ftz.f32 %f118, %f116, %f117;
mov.b32 %r1022, %f118;
mov.u32 %r1023, 8;
shfl.sync.bfly.b32 %r1024|%p24, %r1022, %r1023, %r1018, %r1020;
mov.b32 %f119, %r1024;
add.ftz.f32 %f120, %f118, %f119;
mov.b32 %r1025, %f120;
mov.u32 %r1026, 4;
shfl.sync.bfly.b32 %r1027|%p25, %r1025, %r1026, %r1018, %r1020;
mov.b32 %f121, %r1027;
add.ftz.f32 %f122, %f120, %f121;
mov.b32 %r1028, %f122;
mov.u32 %r1029, 2;
shfl.sync.bfly.b32 %r1030|%p26, %r1028, %r1029, %r1018, %r1020;
mov.b32 %f123, %r1030;
add.ftz.f32 %f124, %f122, %f123;
mov.b32 %r1031, %f124;
mov.u32 %r1032, 1;
shfl.sync.bfly.b32 %r1033|%p27, %r1031, %r1032, %r1018, %r1020;
mov.b32 %f125, %r1033;
add.ftz.f32 %f27, %f124, %f125;
@%p14 bra $L__BB312_12;

cvt.u64.u32 %rd108, %r24;
add.s64 %rd109, %rd108, %rd19;
shl.b64 %rd111, %rd109, 2;
add.s64 %rd112, %rd106, %rd111;
st.global.f32 [%rd112], %f27;

$L__BB312_12:
ld.shared.f32 %f126, [%r19+384];
add.ftz.f32 %f127, %f126, %f213;
ld.shared.f32 %f128, [%r19+1408];
add.ftz.f32 %f129, %f128, %f127;
ld.shared.f32 %f130, [%r19+2432];
add.ftz.f32 %f131, %f130, %f129;
mov.b32 %r1034, %f131;
shfl.sync.bfly.b32 %r1038|%p29, %r1034, %r1019, %r1018, %r1020;
mov.b32 %f132, %r1038;
add.ftz.f32 %f133, %f131, %f132;
mov.b32 %r1039, %f133;
shfl.sync.bfly.b32 %r1041|%p30, %r1039, %r1023, %r1018, %r1020;
mov.b32 %f134, %r1041;
add.ftz.f32 %f135, %f133, %f134;
mov.b32 %r1042, %f135;
shfl.sync.bfly.b32 %r1044|%p31, %r1042, %r1026, %r1018, %r1020;
mov.b32 %f136, %r1044;
add.ftz.f32 %f137, %f135, %f136;
mov.b32 %r1045, %f137;
shfl.sync.bfly.b32 %r1047|%p32, %r1045, %r1029, %r1018, %r1020;
mov.b32 %f138, %r1047;
add.ftz.f32 %f139, %f137, %f138;
mov.b32 %r1048, %f139;
shfl.sync.bfly.b32 %r1050|%p33, %r1048, %r1032, %r1018, %r1020;
mov.b32 %f140, %r1050;
add.ftz.f32 %f28, %f139, %f140;
@%p22 bra $L__BB312_14;

add.s32 %r1051, %r24, 1;
cvt.u64.u32 %rd113, %r1051;
add.s64 %rd114, %rd113, %rd19;
shl.b64 %rd116, %rd114, 2;
add.s64 %rd117, %rd106, %rd116;
st.global.f32 [%rd117], %f28;

$L__BB312_14:
shl.b32 %r1052, %r24, 1;
mov.u32 %r1053, 1;
ld.shared.f32 %f141, [%r19+512];
add.ftz.f32 %f142, %f141, %f212;
ld.shared.f32 %f143, [%r19+1536];
add.ftz.f32 %f144, %f143, %f142;
ld.shared.f32 %f145, [%r19+2560];
add.ftz.f32 %f146, %f145, %f144;
mov.b32 %r1054, %f146;
mov.u32 %r1055, 31;
mov.u32 %r1056, 16;
mov.u32 %r1057, -1;
shfl.sync.bfly.b32 %r1058|%p35, %r1054, %r1056, %r1055, %r1057;
mov.b32 %f147, %r1058;
add.ftz.f32 %f148, %f146, %f147;
mov.b32 %r1059, %f148;
mov.u32 %r1060, 8;
shfl.sync.bfly.b32 %r1061|%p36, %r1059, %r1060, %r1055, %r1057;
mov.b32 %f149, %r1061;
add.ftz.f32 %f150, %f148, %f149;
mov.b32 %r1062, %f150;
mov.u32 %r1063, 4;
shfl.sync.bfly.b32 %r1064|%p37, %r1062, %r1063, %r1055, %r1057;
mov.b32 %f151, %r1064;
add.ftz.f32 %f152, %f150, %f151;
mov.b32 %r1065, %f152;
mov.u32 %r1066, 2;
shfl.sync.bfly.b32 %r1067|%p38, %r1065, %r1066, %r1055, %r1057;
mov.b32 %f153, %r1067;
add.ftz.f32 %f154, %f152, %f153;
mov.b32 %r1068, %f154;
shfl.sync.bfly.b32 %r1069|%p39, %r1068, %r1053, %r1055, %r1057;
mov.b32 %f155, %r1069;
add.ftz.f32 %f29, %f154, %f155;
cvt.u64.u32 %rd118, %r1052;
add.s64 %rd119, %rd118, %rd19;
shl.b64 %rd121, %rd119, 2;
add.s64 %rd21, %rd106, %rd121;
@%p14 bra $L__BB312_16;

st.global.f32 [%rd21], %f29;

$L__BB312_16:
ld.shared.f32 %f156, [%r19+640];
add.ftz.f32 %f157, %f156, %f211;
ld.shared.f32 %f158, [%r19+1664];
add.ftz.f32 %f159, %f158, %f157;
ld.shared.f32 %f160, [%r19+2688];
add.ftz.f32 %f161, %f160, %f159;
mov.b32 %r1070, %f161;
shfl.sync.bfly.b32 %r1074|%p41, %r1070, %r1056, %r1055, %r1057;
mov.b32 %f162, %r1074;
add.ftz.f32 %f163, %f161, %f162;
mov.b32 %r1075, %f163;
shfl.sync.bfly.b32 %r1077|%p42, %r1075, %r1060, %r1055, %r1057;
mov.b32 %f164, %r1077;
add.ftz.f32 %f165, %f163, %f164;
mov.b32 %r1078, %f165;
shfl.sync.bfly.b32 %r1080|%p43, %r1078, %r1063, %r1055, %r1057;
mov.b32 %f166, %r1080;
add.ftz.f32 %f167, %f165, %f166;
mov.b32 %r1081, %f167;
shfl.sync.bfly.b32 %r1083|%p44, %r1081, %r1066, %r1055, %r1057;
mov.b32 %f168, %r1083;
add.ftz.f32 %f169, %f167, %f168;
mov.b32 %r1084, %f169;
shfl.sync.bfly.b32 %r1086|%p45, %r1084, %r1053, %r1055, %r1057;
mov.b32 %f170, %r1086;
add.ftz.f32 %f30, %f169, %f170;
@%p22 bra $L__BB312_18;

st.global.f32 [%rd21+4], %f30;

$L__BB312_18:
mul.lo.s32 %r21, %r24, 3;
ld.shared.f32 %f171, [%r19+768];
add.ftz.f32 %f172, %f171, %f210;
ld.shared.f32 %f173, [%r19+1792];
add.ftz.f32 %f174, %f173, %f172;
ld.shared.f32 %f175, [%r19+2816];
add.ftz.f32 %f176, %f175, %f174;
mov.b32 %r1087, %f176;
mov.u32 %r1088, 31;
mov.u32 %r1089, 16;
mov.u32 %r1090, -1;
shfl.sync.bfly.b32 %r1091|%p47, %r1087, %r1089, %r1088, %r1090;
mov.b32 %f177, %r1091;
add.ftz.f32 %f178, %f176, %f177;
mov.b32 %r1092, %f178;
mov.u32 %r1093, 8;
shfl.sync.bfly.b32 %r1094|%p48, %r1092, %r1093, %r1088, %r1090;
mov.b32 %f179, %r1094;
add.ftz.f32 %f180, %f178, %f179;
mov.b32 %r1095, %f180;
mov.u32 %r1096, 4;
shfl.sync.bfly.b32 %r1097|%p49, %r1095, %r1096, %r1088, %r1090;
mov.b32 %f181, %r1097;
add.ftz.f32 %f182, %f180, %f181;
mov.b32 %r1098, %f182;
mov.u32 %r1099, 2;
shfl.sync.bfly.b32 %r1100|%p50, %r1098, %r1099, %r1088, %r1090;
mov.b32 %f183, %r1100;
add.ftz.f32 %f184, %f182, %f183;
mov.b32 %r1101, %f184;
mov.u32 %r1102, 1;
shfl.sync.bfly.b32 %r1103|%p51, %r1101, %r1102, %r1088, %r1090;
mov.b32 %f185, %r1103;
add.ftz.f32 %f31, %f184, %f185;
@%p14 bra $L__BB312_20;

cvt.u64.u32 %rd122, %r21;
add.s64 %rd123, %rd122, %rd19;
shl.b64 %rd125, %rd123, 2;
add.s64 %rd126, %rd106, %rd125;
st.global.f32 [%rd126], %f31;

$L__BB312_20:
ld.shared.f32 %f186, [%r19+896];
add.ftz.f32 %f187, %f186, %f209;
ld.shared.f32 %f188, [%r19+1920];
add.ftz.f32 %f189, %f188, %f187;
ld.shared.f32 %f190, [%r19+2944];
add.ftz.f32 %f191, %f190, %f189;
mov.b32 %r1104, %f191;
shfl.sync.bfly.b32 %r1108|%p53, %r1104, %r1089, %r1088, %r1090;
mov.b32 %f192, %r1108;
add.ftz.f32 %f193, %f191, %f192;
mov.b32 %r1109, %f193;
shfl.sync.bfly.b32 %r1111|%p54, %r1109, %r1093, %r1088, %r1090;
mov.b32 %f194, %r1111;
add.ftz.f32 %f195, %f193, %f194;
mov.b32 %r1112, %f195;
shfl.sync.bfly.b32 %r1114|%p55, %r1112, %r1096, %r1088, %r1090;
mov.b32 %f196, %r1114;
add.ftz.f32 %f197, %f195, %f196;
mov.b32 %r1115, %f197;
shfl.sync.bfly.b32 %r1117|%p56, %r1115, %r1099, %r1088, %r1090;
mov.b32 %f198, %r1117;
add.ftz.f32 %f199, %f197, %f198;
mov.b32 %r1118, %f199;
shfl.sync.bfly.b32 %r1120|%p57, %r1118, %r1102, %r1088, %r1090;
mov.b32 %f200, %r1120;
add.ftz.f32 %f32, %f199, %f200;
@%p22 bra $L__BB312_22;

add.s32 %r1121, %r21, 1;
cvt.u64.u32 %rd127, %r1121;
add.s64 %rd128, %rd127, %rd19;
shl.b64 %rd130, %rd128, 2;
add.s64 %rd131, %rd106, %rd130;
st.global.f32 [%rd131], %f32;

$L__BB312_22:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type21ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type21ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type21ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type21ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type21ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type21ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type21ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type21ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type21ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<71>;
.reg .b16 %rs<55>;
.reg .f32 %f<231>;
.reg .b32 %r<1322>;
.reg .b64 %rd<137>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type21ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1280];

ld.param.u64 %rd24, [_Z13mul_mat_vec_qIL9ggml_type21ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd25, [_Z13mul_mat_vec_qIL9ggml_type21ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd26, [_Z13mul_mat_vec_qIL9ggml_type21ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type21ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r26, [_Z13mul_mat_vec_qIL9ggml_type21ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r27, [_Z13mul_mat_vec_qIL9ggml_type21ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r28, [_Z13mul_mat_vec_qIL9ggml_type21ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type21ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type21ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type21ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type21ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type21ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type21ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type21ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type21ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type21ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r29, [_Z13mul_mat_vec_qIL9ggml_type21ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
cvta.to.global.u64 %rd1, %rd26;
mov.u32 %r1, %tid.y;
shl.b32 %r42, %r1, 5;
mov.u32 %r43, %tid.x;
add.s32 %r2, %r42, %r43;
mov.u32 %r44, %ctaid.x;
shl.b32 %r3, %r44, 1;
shr.s32 %r45, %r2, 31;
shr.u32 %r46, %r45, 29;
add.s32 %r47, %r2, %r46;
shr.s32 %r1321, %r47, 3;
shr.u32 %r5, %r41, 8;
setp.ge.s32 %p3, %r1321, %r5;
mov.f32 %f221, 0f00000000;
mov.f32 %f222, %f221;
mov.f32 %f223, %f221;
mov.f32 %f224, %f221;
mov.f32 %f225, %f221;
mov.f32 %f226, %f221;
mov.f32 %f227, %f221;
mov.f32 %f228, %f221;
mov.f32 %f229, %f221;
mov.f32 %f230, %f221;
@%p3 bra $L__BB313_3;

mov.u32 %r48, %ctaid.y;
mov.u32 %r49, %ctaid.z;
mul.lo.s32 %r50, %r48, %r33;
cvt.u64.u32 %rd27, %r50;
mul.lo.s32 %r51, %r49, %r39;
cvt.u64.u32 %rd28, %r51;
add.s64 %rd2, %rd28, %rd27;
mul.hi.u32 %r52, %r49, %r35;
add.s32 %r53, %r52, %r49;
shr.u32 %r54, %r53, %r36;
mul.lo.s32 %r55, %r54, %r38;
mul.hi.u32 %r56, %r48, %r29;
add.s32 %r57, %r56, %r48;
shr.u32 %r58, %r57, %r30;
mul.lo.s32 %r59, %r58, %r32;
cvta.to.global.u64 %rd136, %rd24;
and.b32 %r63, %r47, -8;
sub.s32 %r64, %r2, %r63;
shl.b32 %r65, %r64, 1;
shr.u32 %r66, %r64, 31;
add.s32 %r67, %r64, %r66;
shr.s32 %r68, %r67, 1;
cvt.s64.s32 %rd29, %r68;
shl.b32 %r69, %r64, 2;
cvt.s64.s32 %rd4, %r64;
and.b32 %r6, %r69, 4;
shl.b32 %r70, %r27, 1;
shl.b32 %r1316, %r1321, 3;
add.s32 %r1320, %r1316, %r70;
mad.lo.s32 %r1319, %r27, 3, %r1316;
shl.b32 %r71, %r27, 2;
add.s32 %r1318, %r1316, %r71;
add.s32 %r1317, %r27, %r1316;
add.s32 %r72, %r1321, %r55;
add.s32 %r73, %r72, %r59;
add.s32 %r74, %r3, 1;
mad.lo.s32 %r75, %r26, %r74, %r73;
mul.wide.s32 %rd5, %r75, 110;
add.s64 %rd30, %rd5, %rd29;
add.s64 %rd6, %rd30, 106;
add.s64 %rd31, %rd5, 74;
not.b32 %r76, %r65;
mul.wide.s32 %rd32, %r76, 2;
sub.s64 %rd7, %rd31, %rd32;
mul.wide.s32 %rd33, %r65, 2;
add.s64 %rd34, %rd5, %rd33;
add.s64 %rd8, %rd34, 74;
add.s64 %rd35, %rd5, %rd4;
add.s64 %rd9, %rd35, 66;
mad.lo.s32 %r77, %r3, %r26, %r1321;
add.s32 %r78, %r77, %r55;
add.s32 %r79, %r78, %r59;
mul.wide.s32 %rd10, %r79, 110;
mul.wide.s32 %rd36, %r69, 2;
add.s64 %rd37, %rd10, %rd36;
add.s64 %rd11, %rd37, 4;
add.s64 %rd38, %rd5, %rd36;
add.s64 %rd12, %rd38, 2;
add.s64 %rd39, %rd10, %rd29;
add.s64 %rd13, %rd39, 106;
add.s64 %rd40, %rd10, 74;
sub.s64 %rd14, %rd40, %rd32;
add.s64 %rd41, %rd10, %rd33;
add.s64 %rd15, %rd41, 74;
add.s64 %rd42, %rd10, %rd4;
add.s64 %rd16, %rd42, 66;
cvta.to.global.u64 %rd17, %rd25;
mov.f32 %f221, 0f00000000;

$L__BB313_2:
cvt.u64.u32 %rd43, %r1316;
add.s64 %rd44, %rd2, %rd43;
add.s64 %rd45, %rd44, %rd4;
mul.lo.s64 %rd46, %rd45, 36;
add.s64 %rd47, %rd136, %rd11;
ld.global.nc.u16 %rs11, [%rd47];
ld.global.nc.u16 %rs12, [%rd47+2];
ld.global.nc.u16 %rs13, [%rd47+4];
add.s64 %rd48, %rd136, %rd16;
ld.global.nc.u8 %rs14, [%rd48];
add.s64 %rd49, %rd136, %rd15;
add.s64 %rd50, %rd136, %rd14;
ld.global.nc.u16 %rs15, [%rd49];
ld.global.nc.u16 %rs16, [%rd50];
ld.global.nc.u16 %rs17, [%rd47+-2];
and.b16 %rs18, %rs17, 255;
cvt.u32.u16 %r570, %rs18;
shl.b16 %rs19, %rs14, 8;
cvt.u32.u16 %r571, %rs19;
and.b32 %r572, %r571, 256;
or.b32 %r573, %r572, %r570;
mul.wide.u32 %rd51, %r573, 4;
mov.u64 %rd52, iq3s_grid;
add.s64 %rd53, %rd52, %rd51;
ld.global.nc.u32 %r574, [%rd53];
not.b32 %r575, %r574;
shr.u16 %rs20, %rs17, 8;
cvt.u32.u16 %r576, %rs20;
mul.wide.u16 %r577, %rs14, 128;
and.b32 %r578, %r577, 256;
or.b32 %r579, %r578, %r576;
mul.wide.u32 %rd54, %r579, 4;
add.s64 %rd55, %rd52, %rd54;
ld.global.nc.u32 %r580, [%rd55];
not.b32 %r581, %r580;
and.b16 %rs22, %rs15, 255;
mul.wide.u16 %r582, %rs22, 128;
cvt.u32.u16 %r583, %rs15;
shl.b32 %r584, %r583, 21;
or.b32 %r585, %r582, %r584;
and.b32 %r586, %r585, 25166208;
or.b32 %r587, %r586, -2139062144;
add.s32 %r588, %r587, -16843009;
or.b32 %r472, %r588, %r586;
mov.u32 %r80, %r472;

prmt.b32 %r80, %r80, 0, 0xba98;

mul.wide.u16 %r589, %rs22, 8;
shl.b32 %r590, %r583, 17;
or.b32 %r591, %r589, %r590;
and.b32 %r592, %r591, 25166208;
or.b32 %r593, %r592, -2139062144;
add.s32 %r594, %r593, -16843009;
or.b32 %r474, %r594, %r592;
mov.u32 %r82, %r474;

prmt.b32 %r82, %r82, 0, 0xba98;

xor.b32 %r595, %r80, %r574;
or.b32 %r596, %r595, -2139062144;
and.b32 %r597, %r80, 2139062143;
and.b32 %r598, %r575, -2139062144;
sub.s32 %r599, %r596, %r597;
xor.b32 %r85, %r598, %r599;
xor.b32 %r600, %r82, %r580;
or.b32 %r601, %r600, -2139062144;
and.b32 %r602, %r82, 2139062143;
and.b32 %r603, %r581, -2139062144;
sub.s32 %r604, %r601, %r602;
xor.b32 %r89, %r603, %r604;
add.s64 %rd56, %rd17, %rd46;
ld.global.nc.u32 %r86, [%rd56+4];
ld.global.nc.u32 %r90, [%rd56+8];
mov.u32 %r528, 0;

dp4a.s32.s32 %r84, %r85, %r86, %r528;


dp4a.s32.s32 %r88, %r89, %r90, %r84;

and.b16 %rs23, %rs11, 255;
cvt.u32.u16 %r605, %rs23;
mul.wide.u16 %r606, %rs14, 64;
and.b32 %r607, %r606, 256;
or.b32 %r608, %r607, %r605;
mul.wide.u32 %rd57, %r608, 4;
add.s64 %rd58, %rd52, %rd57;
ld.global.nc.u32 %r609, [%rd58];
not.b32 %r610, %r609;
shr.u16 %rs24, %rs11, 8;
cvt.u32.u16 %r611, %rs24;
mul.wide.u16 %r612, %rs14, 32;
and.b32 %r613, %r612, 256;
or.b32 %r614, %r613, %r611;
mul.wide.u32 %rd59, %r614, 4;
add.s64 %rd60, %rd52, %rd59;
ld.global.nc.u32 %r615, [%rd60];
not.b32 %r616, %r615;
shr.u16 %rs25, %rs15, 8;
mul.wide.u16 %r617, %rs25, 128;
cvt.u32.u16 %r618, %rs25;
shl.b32 %r619, %r618, 21;
or.b32 %r620, %r617, %r619;
and.b32 %r621, %r620, 25166208;
or.b32 %r622, %r621, -2139062144;
add.s32 %r623, %r622, -16843009;
or.b32 %r484, %r623, %r621;
mov.u32 %r92, %r484;

prmt.b32 %r92, %r92, 0, 0xba98;

mul.wide.u16 %r624, %rs25, 8;
shl.b32 %r625, %r618, 17;
or.b32 %r626, %r624, %r625;
and.b32 %r627, %r626, 25166208;
or.b32 %r628, %r627, -2139062144;
add.s32 %r629, %r628, -16843009;
or.b32 %r486, %r629, %r627;
mov.u32 %r94, %r486;

prmt.b32 %r94, %r94, 0, 0xba98;

xor.b32 %r630, %r92, %r609;
or.b32 %r631, %r630, -2139062144;
and.b32 %r632, %r92, 2139062143;
and.b32 %r633, %r610, -2139062144;
sub.s32 %r634, %r631, %r632;
xor.b32 %r97, %r633, %r634;
xor.b32 %r635, %r94, %r615;
or.b32 %r636, %r635, -2139062144;
and.b32 %r637, %r94, 2139062143;
and.b32 %r638, %r616, -2139062144;
sub.s32 %r639, %r636, %r637;
xor.b32 %r101, %r638, %r639;
ld.global.nc.u32 %r98, [%rd56+12];
ld.global.nc.u32 %r102, [%rd56+16];

dp4a.s32.s32 %r96, %r97, %r98, %r88;


dp4a.s32.s32 %r100, %r101, %r102, %r96;

and.b16 %rs26, %rs12, 255;
cvt.u32.u16 %r640, %rs26;
mul.wide.u16 %r641, %rs14, 16;
and.b32 %r642, %r641, 256;
or.b32 %r643, %r642, %r640;
mul.wide.u32 %rd61, %r643, 4;
add.s64 %rd62, %rd52, %rd61;
ld.global.nc.u32 %r644, [%rd62];
not.b32 %r645, %r644;
shr.u16 %rs27, %rs12, 8;
cvt.u32.u16 %r646, %rs27;
mul.wide.u16 %r647, %rs14, 8;
and.b32 %r648, %r647, 256;
or.b32 %r649, %r648, %r646;
mul.wide.u32 %rd63, %r649, 4;
add.s64 %rd64, %rd52, %rd63;
ld.global.nc.u32 %r650, [%rd64];
not.b32 %r651, %r650;
and.b16 %rs28, %rs16, 255;
mul.wide.u16 %r652, %rs28, 128;
cvt.u32.u16 %r653, %rs16;
shl.b32 %r654, %r653, 21;
or.b32 %r655, %r652, %r654;
and.b32 %r656, %r655, 25166208;
or.b32 %r657, %r656, -2139062144;
add.s32 %r658, %r657, -16843009;
or.b32 %r496, %r658, %r656;
mov.u32 %r104, %r496;

prmt.b32 %r104, %r104, 0, 0xba98;

mul.wide.u16 %r659, %rs28, 8;
shl.b32 %r660, %r653, 17;
or.b32 %r661, %r659, %r660;
and.b32 %r662, %r661, 25166208;
or.b32 %r663, %r662, -2139062144;
add.s32 %r664, %r663, -16843009;
or.b32 %r498, %r664, %r662;
mov.u32 %r106, %r498;

prmt.b32 %r106, %r106, 0, 0xba98;

xor.b32 %r665, %r104, %r644;
or.b32 %r666, %r665, -2139062144;
and.b32 %r667, %r104, 2139062143;
and.b32 %r668, %r645, -2139062144;
sub.s32 %r669, %r666, %r667;
xor.b32 %r109, %r668, %r669;
xor.b32 %r670, %r106, %r650;
or.b32 %r671, %r670, -2139062144;
and.b32 %r672, %r106, 2139062143;
and.b32 %r673, %r651, -2139062144;
sub.s32 %r674, %r671, %r672;
xor.b32 %r113, %r673, %r674;
ld.global.nc.u32 %r110, [%rd56+20];
ld.global.nc.u32 %r114, [%rd56+24];

dp4a.s32.s32 %r108, %r109, %r110, %r100;


dp4a.s32.s32 %r112, %r113, %r114, %r108;

and.b16 %rs29, %rs13, 255;
cvt.u32.u16 %r675, %rs29;
mul.wide.u16 %r676, %rs14, 4;
and.b32 %r677, %r676, 256;
or.b32 %r678, %r677, %r675;
mul.wide.u32 %rd65, %r678, 4;
add.s64 %rd66, %rd52, %rd65;
ld.global.nc.u32 %r679, [%rd66];
not.b32 %r680, %r679;
shr.u16 %rs30, %rs13, 8;
cvt.u32.u16 %r681, %rs30;
mul.wide.u16 %r682, %rs14, 2;
and.b32 %r683, %r682, 256;
or.b32 %r684, %r683, %r681;
mul.wide.u32 %rd67, %r684, 4;
add.s64 %rd68, %rd52, %rd67;
ld.global.nc.u32 %r685, [%rd68];
not.b32 %r686, %r685;
shr.u16 %rs31, %rs16, 8;
mul.wide.u16 %r687, %rs31, 128;
cvt.u32.u16 %r688, %rs31;
shl.b32 %r689, %r688, 21;
or.b32 %r690, %r687, %r689;
and.b32 %r691, %r690, 25166208;
or.b32 %r692, %r691, -2139062144;
add.s32 %r693, %r692, -16843009;
or.b32 %r508, %r693, %r691;
mov.u32 %r116, %r508;

prmt.b32 %r116, %r116, 0, 0xba98;

mul.wide.u16 %r694, %rs31, 8;
shl.b32 %r695, %r688, 17;
or.b32 %r696, %r694, %r695;
and.b32 %r697, %r696, 25166208;
or.b32 %r698, %r697, -2139062144;
add.s32 %r699, %r698, -16843009;
or.b32 %r510, %r699, %r697;
mov.u32 %r118, %r510;

prmt.b32 %r118, %r118, 0, 0xba98;

xor.b32 %r700, %r116, %r679;
or.b32 %r701, %r700, -2139062144;
and.b32 %r702, %r116, 2139062143;
and.b32 %r703, %r680, -2139062144;
sub.s32 %r704, %r701, %r702;
xor.b32 %r121, %r703, %r704;
xor.b32 %r705, %r118, %r685;
or.b32 %r706, %r705, -2139062144;
and.b32 %r707, %r118, 2139062143;
and.b32 %r708, %r686, -2139062144;
sub.s32 %r709, %r706, %r707;
xor.b32 %r125, %r708, %r709;
ld.global.nc.u32 %r122, [%rd56+28];
ld.global.nc.u32 %r126, [%rd56+32];

dp4a.s32.s32 %r120, %r121, %r122, %r112;


dp4a.s32.s32 %r124, %r125, %r126, %r120;

add.s64 %rd69, %rd136, %rd13;
ld.global.nc.u8 %rs32, [%rd69];
cvt.u32.u16 %r710, %rs32;
and.b32 %r711, %r710, 255;
shr.u32 %r712, %r711, %r6;
shl.b32 %r713, %r712, 1;
and.b32 %r714, %r713, 30;
or.b32 %r715, %r714, 1;
mul.lo.s32 %r716, %r715, %r124;
add.s64 %rd70, %rd136, %rd10;
ld.global.nc.u16 %rs1, [%rd70];

{ cvt.f32.f16 %f61, %rs1;}


ld.global.nc.u32 %r128, [%rd56];

{.reg .f16 low,high;
mov.b32 {low,high},%r128;
cvt.f32.f16 %f62, low;}


mul.ftz.f32 %f81, %f61, %f62;
cvt.rn.f32.s32 %f82, %r716;
fma.rn.ftz.f32 %f230, %f81, %f82, %f230;
add.s64 %rd71, %rd136, %rd12;
ld.global.nc.u16 %rs33, [%rd71+2];
ld.global.nc.u16 %rs34, [%rd71+4];
ld.global.nc.u16 %rs35, [%rd71+6];
add.s64 %rd72, %rd136, %rd9;
ld.global.nc.u8 %rs36, [%rd72];
add.s64 %rd73, %rd136, %rd8;
add.s64 %rd74, %rd136, %rd7;
ld.global.nc.u16 %rs37, [%rd73];
ld.global.nc.u16 %rs38, [%rd74];
ld.global.nc.u16 %rs39, [%rd71];
and.b16 %rs40, %rs39, 255;
cvt.u32.u16 %r717, %rs40;
shl.b16 %rs41, %rs36, 8;
cvt.u32.u16 %r718, %rs41;
and.b32 %r719, %r718, 256;
or.b32 %r720, %r719, %r717;
mul.wide.u32 %rd75, %r720, 4;
add.s64 %rd76, %rd52, %rd75;
ld.global.nc.u32 %r721, [%rd76];
not.b32 %r722, %r721;
shr.u16 %rs42, %rs39, 8;
cvt.u32.u16 %r723, %rs42;
mul.wide.u16 %r724, %rs36, 128;
and.b32 %r725, %r724, 256;
or.b32 %r726, %r725, %r723;
mul.wide.u32 %rd77, %r726, 4;
add.s64 %rd78, %rd52, %rd77;
ld.global.nc.u32 %r727, [%rd78];
not.b32 %r728, %r727;
and.b16 %rs44, %rs37, 255;
mul.wide.u16 %r729, %rs44, 128;
cvt.u32.u16 %r730, %rs37;
shl.b32 %r731, %r730, 21;
or.b32 %r732, %r729, %r731;
and.b32 %r733, %r732, 25166208;
or.b32 %r734, %r733, -2139062144;
add.s32 %r735, %r734, -16843009;
or.b32 %r521, %r735, %r733;
mov.u32 %r129, %r521;

prmt.b32 %r129, %r129, 0, 0xba98;

mul.wide.u16 %r736, %rs44, 8;
shl.b32 %r737, %r730, 17;
or.b32 %r738, %r736, %r737;
and.b32 %r739, %r738, 25166208;
or.b32 %r740, %r739, -2139062144;
add.s32 %r741, %r740, -16843009;
or.b32 %r523, %r741, %r739;
mov.u32 %r131, %r523;

prmt.b32 %r131, %r131, 0, 0xba98;

xor.b32 %r742, %r129, %r721;
or.b32 %r743, %r742, -2139062144;
and.b32 %r744, %r129, 2139062143;
and.b32 %r745, %r722, -2139062144;
sub.s32 %r746, %r743, %r744;
xor.b32 %r134, %r745, %r746;
xor.b32 %r747, %r131, %r727;
or.b32 %r748, %r747, -2139062144;
and.b32 %r749, %r131, 2139062143;
and.b32 %r750, %r728, -2139062144;
sub.s32 %r751, %r748, %r749;
xor.b32 %r138, %r750, %r751;

dp4a.s32.s32 %r133, %r134, %r86, %r528;


dp4a.s32.s32 %r137, %r138, %r90, %r133;

and.b16 %rs45, %rs33, 255;
cvt.u32.u16 %r752, %rs45;
mul.wide.u16 %r753, %rs36, 64;
and.b32 %r754, %r753, 256;
or.b32 %r755, %r754, %r752;
mul.wide.u32 %rd79, %r755, 4;
add.s64 %rd80, %rd52, %rd79;
ld.global.nc.u32 %r756, [%rd80];
not.b32 %r757, %r756;
shr.u16 %rs46, %rs33, 8;
cvt.u32.u16 %r758, %rs46;
mul.wide.u16 %r759, %rs36, 32;
and.b32 %r760, %r759, 256;
or.b32 %r761, %r760, %r758;
mul.wide.u32 %rd81, %r761, 4;
add.s64 %rd82, %rd52, %rd81;
ld.global.nc.u32 %r762, [%rd82];
not.b32 %r763, %r762;
shr.u16 %rs47, %rs37, 8;
mul.wide.u16 %r764, %rs47, 128;
cvt.u32.u16 %r765, %rs47;
shl.b32 %r766, %r765, 21;
or.b32 %r767, %r764, %r766;
and.b32 %r768, %r767, 25166208;
or.b32 %r769, %r768, -2139062144;
add.s32 %r770, %r769, -16843009;
or.b32 %r533, %r770, %r768;
mov.u32 %r141, %r533;

prmt.b32 %r141, %r141, 0, 0xba98;

mul.wide.u16 %r771, %rs47, 8;
shl.b32 %r772, %r765, 17;
or.b32 %r773, %r771, %r772;
and.b32 %r774, %r773, 25166208;
or.b32 %r775, %r774, -2139062144;
add.s32 %r776, %r775, -16843009;
or.b32 %r535, %r776, %r774;
mov.u32 %r143, %r535;

prmt.b32 %r143, %r143, 0, 0xba98;

xor.b32 %r777, %r141, %r756;
or.b32 %r778, %r777, -2139062144;
and.b32 %r779, %r141, 2139062143;
and.b32 %r780, %r757, -2139062144;
sub.s32 %r781, %r778, %r779;
xor.b32 %r146, %r780, %r781;
xor.b32 %r782, %r143, %r762;
or.b32 %r783, %r782, -2139062144;
and.b32 %r784, %r143, 2139062143;
and.b32 %r785, %r763, -2139062144;
sub.s32 %r786, %r783, %r784;
xor.b32 %r150, %r785, %r786;

dp4a.s32.s32 %r145, %r146, %r98, %r137;


dp4a.s32.s32 %r149, %r150, %r102, %r145;

and.b16 %rs48, %rs34, 255;
cvt.u32.u16 %r787, %rs48;
mul.wide.u16 %r788, %rs36, 16;
and.b32 %r789, %r788, 256;
or.b32 %r790, %r789, %r787;
mul.wide.u32 %rd83, %r790, 4;
add.s64 %rd84, %rd52, %rd83;
ld.global.nc.u32 %r791, [%rd84];
not.b32 %r792, %r791;
shr.u16 %rs49, %rs34, 8;
cvt.u32.u16 %r793, %rs49;
mul.wide.u16 %r794, %rs36, 8;
and.b32 %r795, %r794, 256;
or.b32 %r796, %r795, %r793;
mul.wide.u32 %rd85, %r796, 4;
add.s64 %rd86, %rd52, %rd85;
ld.global.nc.u32 %r797, [%rd86];
not.b32 %r798, %r797;
and.b16 %rs50, %rs38, 255;
mul.wide.u16 %r799, %rs50, 128;
cvt.u32.u16 %r800, %rs38;
shl.b32 %r801, %r800, 21;
or.b32 %r802, %r799, %r801;
and.b32 %r803, %r802, 25166208;
or.b32 %r804, %r803, -2139062144;
add.s32 %r805, %r804, -16843009;
or.b32 %r545, %r805, %r803;
mov.u32 %r153, %r545;

prmt.b32 %r153, %r153, 0, 0xba98;

mul.wide.u16 %r806, %rs50, 8;
shl.b32 %r807, %r800, 17;
or.b32 %r808, %r806, %r807;
and.b32 %r809, %r808, 25166208;
or.b32 %r810, %r809, -2139062144;
add.s32 %r811, %r810, -16843009;
or.b32 %r547, %r811, %r809;
mov.u32 %r155, %r547;

prmt.b32 %r155, %r155, 0, 0xba98;

xor.b32 %r812, %r153, %r791;
or.b32 %r813, %r812, -2139062144;
and.b32 %r814, %r153, 2139062143;
and.b32 %r815, %r792, -2139062144;
sub.s32 %r816, %r813, %r814;
xor.b32 %r158, %r815, %r816;
xor.b32 %r817, %r155, %r797;
or.b32 %r818, %r817, -2139062144;
and.b32 %r819, %r155, 2139062143;
and.b32 %r820, %r798, -2139062144;
sub.s32 %r821, %r818, %r819;
xor.b32 %r162, %r820, %r821;

dp4a.s32.s32 %r157, %r158, %r110, %r149;


dp4a.s32.s32 %r161, %r162, %r114, %r157;

and.b16 %rs51, %rs35, 255;
cvt.u32.u16 %r822, %rs51;
mul.wide.u16 %r823, %rs36, 4;
and.b32 %r824, %r823, 256;
or.b32 %r825, %r824, %r822;
mul.wide.u32 %rd87, %r825, 4;
add.s64 %rd88, %rd52, %rd87;
ld.global.nc.u32 %r826, [%rd88];
not.b32 %r827, %r826;
shr.u16 %rs52, %rs35, 8;
cvt.u32.u16 %r828, %rs52;
mul.wide.u16 %r829, %rs36, 2;
and.b32 %r830, %r829, 256;
or.b32 %r831, %r830, %r828;
mul.wide.u32 %rd89, %r831, 4;
add.s64 %rd90, %rd52, %rd89;
ld.global.nc.u32 %r832, [%rd90];
not.b32 %r833, %r832;
shr.u16 %rs53, %rs38, 8;
mul.wide.u16 %r834, %rs53, 128;
cvt.u32.u16 %r835, %rs53;
shl.b32 %r836, %r835, 21;
or.b32 %r837, %r834, %r836;
and.b32 %r838, %r837, 25166208;
or.b32 %r839, %r838, -2139062144;
add.s32 %r840, %r839, -16843009;
or.b32 %r557, %r840, %r838;
mov.u32 %r165, %r557;

prmt.b32 %r165, %r165, 0, 0xba98;

mul.wide.u16 %r841, %rs53, 8;
shl.b32 %r842, %r835, 17;
or.b32 %r843, %r841, %r842;
and.b32 %r844, %r843, 25166208;
or.b32 %r845, %r844, -2139062144;
add.s32 %r846, %r845, -16843009;
or.b32 %r559, %r846, %r844;
mov.u32 %r167, %r559;

prmt.b32 %r167, %r167, 0, 0xba98;

xor.b32 %r847, %r165, %r826;
or.b32 %r848, %r847, -2139062144;
and.b32 %r849, %r165, 2139062143;
and.b32 %r850, %r827, -2139062144;
sub.s32 %r851, %r848, %r849;
xor.b32 %r170, %r850, %r851;
xor.b32 %r852, %r167, %r832;
or.b32 %r853, %r852, -2139062144;
and.b32 %r854, %r167, 2139062143;
and.b32 %r855, %r833, -2139062144;
sub.s32 %r856, %r853, %r854;
xor.b32 %r174, %r855, %r856;

dp4a.s32.s32 %r169, %r170, %r122, %r161;


dp4a.s32.s32 %r173, %r174, %r126, %r169;

add.s64 %rd91, %rd136, %rd6;
ld.global.nc.u8 %rs54, [%rd91];
cvt.u32.u16 %r857, %rs54;
and.b32 %r858, %r857, 255;
shr.u32 %r859, %r858, %r6;
shl.b32 %r860, %r859, 1;
and.b32 %r861, %r860, 30;
or.b32 %r862, %r861, 1;
mul.lo.s32 %r863, %r862, %r173;
add.s64 %rd92, %rd136, %rd5;
ld.global.nc.u16 %rs2, [%rd92];

{ cvt.f32.f16 %f63, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r128;
cvt.f32.f16 %f64, low;}


mul.ftz.f32 %f83, %f63, %f64;
cvt.rn.f32.s32 %f84, %r863;
fma.rn.ftz.f32 %f229, %f83, %f84, %f229;
cvt.u64.u32 %rd93, %r1317;
add.s64 %rd94, %rd2, %rd93;
add.s64 %rd95, %rd94, %rd4;
mul.lo.s64 %rd96, %rd95, 36;
mov.u32 %r178, %r472;

prmt.b32 %r178, %r178, 0, 0xba98;

mov.u32 %r180, %r474;

prmt.b32 %r180, %r180, 0, 0xba98;

xor.b32 %r864, %r178, %r574;
or.b32 %r865, %r864, -2139062144;
and.b32 %r866, %r178, 2139062143;
sub.s32 %r867, %r865, %r866;
xor.b32 %r183, %r598, %r867;
xor.b32 %r868, %r180, %r580;
or.b32 %r869, %r868, -2139062144;
and.b32 %r870, %r180, 2139062143;
sub.s32 %r871, %r869, %r870;
xor.b32 %r187, %r603, %r871;
add.s64 %rd97, %rd17, %rd96;
ld.global.nc.u32 %r184, [%rd97+4];
ld.global.nc.u32 %r188, [%rd97+8];

dp4a.s32.s32 %r182, %r183, %r184, %r528;


dp4a.s32.s32 %r186, %r187, %r188, %r182;

mov.u32 %r190, %r484;

prmt.b32 %r190, %r190, 0, 0xba98;

mov.u32 %r192, %r486;

prmt.b32 %r192, %r192, 0, 0xba98;

xor.b32 %r872, %r190, %r609;
or.b32 %r873, %r872, -2139062144;
and.b32 %r874, %r190, 2139062143;
sub.s32 %r875, %r873, %r874;
xor.b32 %r195, %r633, %r875;
xor.b32 %r876, %r192, %r615;
or.b32 %r877, %r876, -2139062144;
and.b32 %r878, %r192, 2139062143;
sub.s32 %r879, %r877, %r878;
xor.b32 %r199, %r638, %r879;
ld.global.nc.u32 %r196, [%rd97+12];
ld.global.nc.u32 %r200, [%rd97+16];

dp4a.s32.s32 %r194, %r195, %r196, %r186;


dp4a.s32.s32 %r198, %r199, %r200, %r194;

mov.u32 %r202, %r496;

prmt.b32 %r202, %r202, 0, 0xba98;

mov.u32 %r204, %r498;

prmt.b32 %r204, %r204, 0, 0xba98;

xor.b32 %r880, %r202, %r644;
or.b32 %r881, %r880, -2139062144;
and.b32 %r882, %r202, 2139062143;
sub.s32 %r883, %r881, %r882;
xor.b32 %r207, %r668, %r883;
xor.b32 %r884, %r204, %r650;
or.b32 %r885, %r884, -2139062144;
and.b32 %r886, %r204, 2139062143;
sub.s32 %r887, %r885, %r886;
xor.b32 %r211, %r673, %r887;
ld.global.nc.u32 %r208, [%rd97+20];
ld.global.nc.u32 %r212, [%rd97+24];

dp4a.s32.s32 %r206, %r207, %r208, %r198;


dp4a.s32.s32 %r210, %r211, %r212, %r206;

mov.u32 %r214, %r508;

prmt.b32 %r214, %r214, 0, 0xba98;

mov.u32 %r216, %r510;

prmt.b32 %r216, %r216, 0, 0xba98;

xor.b32 %r888, %r214, %r679;
or.b32 %r889, %r888, -2139062144;
and.b32 %r890, %r214, 2139062143;
sub.s32 %r891, %r889, %r890;
xor.b32 %r219, %r703, %r891;
xor.b32 %r892, %r216, %r685;
or.b32 %r893, %r892, -2139062144;
and.b32 %r894, %r216, 2139062143;
sub.s32 %r895, %r893, %r894;
xor.b32 %r223, %r708, %r895;
ld.global.nc.u32 %r220, [%rd97+28];
ld.global.nc.u32 %r224, [%rd97+32];

dp4a.s32.s32 %r218, %r219, %r220, %r210;


dp4a.s32.s32 %r222, %r223, %r224, %r218;

mul.lo.s32 %r896, %r715, %r222;

{ cvt.f32.f16 %f65, %rs1;}


ld.global.nc.u32 %r226, [%rd97];

{.reg .f16 low,high;
mov.b32 {low,high},%r226;
cvt.f32.f16 %f66, low;}


mul.ftz.f32 %f85, %f65, %f66;
cvt.rn.f32.s32 %f86, %r896;
fma.rn.ftz.f32 %f228, %f85, %f86, %f228;
mov.u32 %r227, %r521;

prmt.b32 %r227, %r227, 0, 0xba98;

mov.u32 %r229, %r523;

prmt.b32 %r229, %r229, 0, 0xba98;

xor.b32 %r897, %r227, %r721;
or.b32 %r898, %r897, -2139062144;
and.b32 %r899, %r227, 2139062143;
sub.s32 %r900, %r898, %r899;
xor.b32 %r232, %r745, %r900;
xor.b32 %r901, %r229, %r727;
or.b32 %r902, %r901, -2139062144;
and.b32 %r903, %r229, 2139062143;
sub.s32 %r904, %r902, %r903;
xor.b32 %r236, %r750, %r904;

dp4a.s32.s32 %r231, %r232, %r184, %r528;


dp4a.s32.s32 %r235, %r236, %r188, %r231;

mov.u32 %r239, %r533;

prmt.b32 %r239, %r239, 0, 0xba98;

mov.u32 %r241, %r535;

prmt.b32 %r241, %r241, 0, 0xba98;

xor.b32 %r905, %r239, %r756;
or.b32 %r906, %r905, -2139062144;
and.b32 %r907, %r239, 2139062143;
sub.s32 %r908, %r906, %r907;
xor.b32 %r244, %r780, %r908;
xor.b32 %r909, %r241, %r762;
or.b32 %r910, %r909, -2139062144;
and.b32 %r911, %r241, 2139062143;
sub.s32 %r912, %r910, %r911;
xor.b32 %r248, %r785, %r912;

dp4a.s32.s32 %r243, %r244, %r196, %r235;


dp4a.s32.s32 %r247, %r248, %r200, %r243;

mov.u32 %r251, %r545;

prmt.b32 %r251, %r251, 0, 0xba98;

mov.u32 %r253, %r547;

prmt.b32 %r253, %r253, 0, 0xba98;

xor.b32 %r913, %r251, %r791;
or.b32 %r914, %r913, -2139062144;
and.b32 %r915, %r251, 2139062143;
sub.s32 %r916, %r914, %r915;
xor.b32 %r256, %r815, %r916;
xor.b32 %r917, %r253, %r797;
or.b32 %r918, %r917, -2139062144;
and.b32 %r919, %r253, 2139062143;
sub.s32 %r920, %r918, %r919;
xor.b32 %r260, %r820, %r920;

dp4a.s32.s32 %r255, %r256, %r208, %r247;


dp4a.s32.s32 %r259, %r260, %r212, %r255;

mov.u32 %r263, %r557;

prmt.b32 %r263, %r263, 0, 0xba98;

mov.u32 %r265, %r559;

prmt.b32 %r265, %r265, 0, 0xba98;

xor.b32 %r921, %r263, %r826;
or.b32 %r922, %r921, -2139062144;
and.b32 %r923, %r263, 2139062143;
sub.s32 %r924, %r922, %r923;
xor.b32 %r268, %r850, %r924;
xor.b32 %r925, %r265, %r832;
or.b32 %r926, %r925, -2139062144;
and.b32 %r927, %r265, 2139062143;
sub.s32 %r928, %r926, %r927;
xor.b32 %r272, %r855, %r928;

dp4a.s32.s32 %r267, %r268, %r220, %r259;


dp4a.s32.s32 %r271, %r272, %r224, %r267;

mul.lo.s32 %r929, %r862, %r271;

{ cvt.f32.f16 %f67, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r226;
cvt.f32.f16 %f68, low;}


mul.ftz.f32 %f87, %f67, %f68;
cvt.rn.f32.s32 %f88, %r929;
fma.rn.ftz.f32 %f227, %f87, %f88, %f227;
cvt.u64.u32 %rd98, %r1320;
add.s64 %rd99, %rd2, %rd98;
add.s64 %rd100, %rd99, %rd4;
mul.lo.s64 %rd101, %rd100, 36;
mov.u32 %r276, %r472;

prmt.b32 %r276, %r276, 0, 0xba98;

mov.u32 %r278, %r474;

prmt.b32 %r278, %r278, 0, 0xba98;

xor.b32 %r930, %r276, %r574;
or.b32 %r931, %r930, -2139062144;
and.b32 %r932, %r276, 2139062143;
sub.s32 %r933, %r931, %r932;
xor.b32 %r281, %r598, %r933;
xor.b32 %r934, %r278, %r580;
or.b32 %r935, %r934, -2139062144;
and.b32 %r936, %r278, 2139062143;
sub.s32 %r937, %r935, %r936;
xor.b32 %r285, %r603, %r937;
add.s64 %rd102, %rd17, %rd101;
ld.global.nc.u32 %r282, [%rd102+4];
ld.global.nc.u32 %r286, [%rd102+8];

dp4a.s32.s32 %r280, %r281, %r282, %r528;


dp4a.s32.s32 %r284, %r285, %r286, %r280;

mov.u32 %r288, %r484;

prmt.b32 %r288, %r288, 0, 0xba98;

mov.u32 %r290, %r486;

prmt.b32 %r290, %r290, 0, 0xba98;

xor.b32 %r938, %r288, %r609;
or.b32 %r939, %r938, -2139062144;
and.b32 %r940, %r288, 2139062143;
sub.s32 %r941, %r939, %r940;
xor.b32 %r293, %r633, %r941;
xor.b32 %r942, %r290, %r615;
or.b32 %r943, %r942, -2139062144;
and.b32 %r944, %r290, 2139062143;
sub.s32 %r945, %r943, %r944;
xor.b32 %r297, %r638, %r945;
ld.global.nc.u32 %r294, [%rd102+12];
ld.global.nc.u32 %r298, [%rd102+16];

dp4a.s32.s32 %r292, %r293, %r294, %r284;


dp4a.s32.s32 %r296, %r297, %r298, %r292;

mov.u32 %r300, %r496;

prmt.b32 %r300, %r300, 0, 0xba98;

mov.u32 %r302, %r498;

prmt.b32 %r302, %r302, 0, 0xba98;

xor.b32 %r946, %r300, %r644;
or.b32 %r947, %r946, -2139062144;
and.b32 %r948, %r300, 2139062143;
sub.s32 %r949, %r947, %r948;
xor.b32 %r305, %r668, %r949;
xor.b32 %r950, %r302, %r650;
or.b32 %r951, %r950, -2139062144;
and.b32 %r952, %r302, 2139062143;
sub.s32 %r953, %r951, %r952;
xor.b32 %r309, %r673, %r953;
ld.global.nc.u32 %r306, [%rd102+20];
ld.global.nc.u32 %r310, [%rd102+24];

dp4a.s32.s32 %r304, %r305, %r306, %r296;


dp4a.s32.s32 %r308, %r309, %r310, %r304;

mov.u32 %r312, %r508;

prmt.b32 %r312, %r312, 0, 0xba98;

mov.u32 %r314, %r510;

prmt.b32 %r314, %r314, 0, 0xba98;

xor.b32 %r954, %r312, %r679;
or.b32 %r955, %r954, -2139062144;
and.b32 %r956, %r312, 2139062143;
sub.s32 %r957, %r955, %r956;
xor.b32 %r317, %r703, %r957;
xor.b32 %r958, %r314, %r685;
or.b32 %r959, %r958, -2139062144;
and.b32 %r960, %r314, 2139062143;
sub.s32 %r961, %r959, %r960;
xor.b32 %r321, %r708, %r961;
ld.global.nc.u32 %r318, [%rd102+28];
ld.global.nc.u32 %r322, [%rd102+32];

dp4a.s32.s32 %r316, %r317, %r318, %r308;


dp4a.s32.s32 %r320, %r321, %r322, %r316;

mul.lo.s32 %r962, %r715, %r320;

{ cvt.f32.f16 %f69, %rs1;}


ld.global.nc.u32 %r324, [%rd102];

{.reg .f16 low,high;
mov.b32 {low,high},%r324;
cvt.f32.f16 %f70, low;}


mul.ftz.f32 %f89, %f69, %f70;
cvt.rn.f32.s32 %f90, %r962;
fma.rn.ftz.f32 %f226, %f89, %f90, %f226;
mov.u32 %r325, %r521;

prmt.b32 %r325, %r325, 0, 0xba98;

mov.u32 %r327, %r523;

prmt.b32 %r327, %r327, 0, 0xba98;

xor.b32 %r963, %r325, %r721;
or.b32 %r964, %r963, -2139062144;
and.b32 %r965, %r325, 2139062143;
sub.s32 %r966, %r964, %r965;
xor.b32 %r330, %r745, %r966;
xor.b32 %r967, %r327, %r727;
or.b32 %r968, %r967, -2139062144;
and.b32 %r969, %r327, 2139062143;
sub.s32 %r970, %r968, %r969;
xor.b32 %r334, %r750, %r970;

dp4a.s32.s32 %r329, %r330, %r282, %r528;


dp4a.s32.s32 %r333, %r334, %r286, %r329;

mov.u32 %r337, %r533;

prmt.b32 %r337, %r337, 0, 0xba98;

mov.u32 %r339, %r535;

prmt.b32 %r339, %r339, 0, 0xba98;

xor.b32 %r971, %r337, %r756;
or.b32 %r972, %r971, -2139062144;
and.b32 %r973, %r337, 2139062143;
sub.s32 %r974, %r972, %r973;
xor.b32 %r342, %r780, %r974;
xor.b32 %r975, %r339, %r762;
or.b32 %r976, %r975, -2139062144;
and.b32 %r977, %r339, 2139062143;
sub.s32 %r978, %r976, %r977;
xor.b32 %r346, %r785, %r978;

dp4a.s32.s32 %r341, %r342, %r294, %r333;


dp4a.s32.s32 %r345, %r346, %r298, %r341;

mov.u32 %r349, %r545;

prmt.b32 %r349, %r349, 0, 0xba98;

mov.u32 %r351, %r547;

prmt.b32 %r351, %r351, 0, 0xba98;

xor.b32 %r979, %r349, %r791;
or.b32 %r980, %r979, -2139062144;
and.b32 %r981, %r349, 2139062143;
sub.s32 %r982, %r980, %r981;
xor.b32 %r354, %r815, %r982;
xor.b32 %r983, %r351, %r797;
or.b32 %r984, %r983, -2139062144;
and.b32 %r985, %r351, 2139062143;
sub.s32 %r986, %r984, %r985;
xor.b32 %r358, %r820, %r986;

dp4a.s32.s32 %r353, %r354, %r306, %r345;


dp4a.s32.s32 %r357, %r358, %r310, %r353;

mov.u32 %r361, %r557;

prmt.b32 %r361, %r361, 0, 0xba98;

mov.u32 %r363, %r559;

prmt.b32 %r363, %r363, 0, 0xba98;

xor.b32 %r987, %r361, %r826;
or.b32 %r988, %r987, -2139062144;
and.b32 %r989, %r361, 2139062143;
sub.s32 %r990, %r988, %r989;
xor.b32 %r366, %r850, %r990;
xor.b32 %r991, %r363, %r832;
or.b32 %r992, %r991, -2139062144;
and.b32 %r993, %r363, 2139062143;
sub.s32 %r994, %r992, %r993;
xor.b32 %r370, %r855, %r994;

dp4a.s32.s32 %r365, %r366, %r318, %r357;


dp4a.s32.s32 %r369, %r370, %r322, %r365;

mul.lo.s32 %r995, %r862, %r369;

{ cvt.f32.f16 %f71, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r324;
cvt.f32.f16 %f72, low;}


mul.ftz.f32 %f91, %f71, %f72;
cvt.rn.f32.s32 %f92, %r995;
fma.rn.ftz.f32 %f225, %f91, %f92, %f225;
cvt.u64.u32 %rd103, %r1319;
add.s64 %rd104, %rd2, %rd103;
add.s64 %rd105, %rd104, %rd4;
mul.lo.s64 %rd106, %rd105, 36;
mov.u32 %r374, %r472;

prmt.b32 %r374, %r374, 0, 0xba98;

mov.u32 %r376, %r474;

prmt.b32 %r376, %r376, 0, 0xba98;

xor.b32 %r996, %r374, %r574;
or.b32 %r997, %r996, -2139062144;
and.b32 %r998, %r374, 2139062143;
sub.s32 %r999, %r997, %r998;
xor.b32 %r379, %r598, %r999;
xor.b32 %r1000, %r376, %r580;
or.b32 %r1001, %r1000, -2139062144;
and.b32 %r1002, %r376, 2139062143;
sub.s32 %r1003, %r1001, %r1002;
xor.b32 %r383, %r603, %r1003;
add.s64 %rd107, %rd17, %rd106;
ld.global.nc.u32 %r380, [%rd107+4];
ld.global.nc.u32 %r384, [%rd107+8];

dp4a.s32.s32 %r378, %r379, %r380, %r528;


dp4a.s32.s32 %r382, %r383, %r384, %r378;

mov.u32 %r386, %r484;

prmt.b32 %r386, %r386, 0, 0xba98;

mov.u32 %r388, %r486;

prmt.b32 %r388, %r388, 0, 0xba98;

xor.b32 %r1004, %r386, %r609;
or.b32 %r1005, %r1004, -2139062144;
and.b32 %r1006, %r386, 2139062143;
sub.s32 %r1007, %r1005, %r1006;
xor.b32 %r391, %r633, %r1007;
xor.b32 %r1008, %r388, %r615;
or.b32 %r1009, %r1008, -2139062144;
and.b32 %r1010, %r388, 2139062143;
sub.s32 %r1011, %r1009, %r1010;
xor.b32 %r395, %r638, %r1011;
ld.global.nc.u32 %r392, [%rd107+12];
ld.global.nc.u32 %r396, [%rd107+16];

dp4a.s32.s32 %r390, %r391, %r392, %r382;


dp4a.s32.s32 %r394, %r395, %r396, %r390;

mov.u32 %r398, %r496;

prmt.b32 %r398, %r398, 0, 0xba98;

mov.u32 %r400, %r498;

prmt.b32 %r400, %r400, 0, 0xba98;

xor.b32 %r1012, %r398, %r644;
or.b32 %r1013, %r1012, -2139062144;
and.b32 %r1014, %r398, 2139062143;
sub.s32 %r1015, %r1013, %r1014;
xor.b32 %r403, %r668, %r1015;
xor.b32 %r1016, %r400, %r650;
or.b32 %r1017, %r1016, -2139062144;
and.b32 %r1018, %r400, 2139062143;
sub.s32 %r1019, %r1017, %r1018;
xor.b32 %r407, %r673, %r1019;
ld.global.nc.u32 %r404, [%rd107+20];
ld.global.nc.u32 %r408, [%rd107+24];

dp4a.s32.s32 %r402, %r403, %r404, %r394;


dp4a.s32.s32 %r406, %r407, %r408, %r402;

mov.u32 %r410, %r508;

prmt.b32 %r410, %r410, 0, 0xba98;

mov.u32 %r412, %r510;

prmt.b32 %r412, %r412, 0, 0xba98;

xor.b32 %r1020, %r410, %r679;
or.b32 %r1021, %r1020, -2139062144;
and.b32 %r1022, %r410, 2139062143;
sub.s32 %r1023, %r1021, %r1022;
xor.b32 %r415, %r703, %r1023;
xor.b32 %r1024, %r412, %r685;
or.b32 %r1025, %r1024, -2139062144;
and.b32 %r1026, %r412, 2139062143;
sub.s32 %r1027, %r1025, %r1026;
xor.b32 %r419, %r708, %r1027;
ld.global.nc.u32 %r416, [%rd107+28];
ld.global.nc.u32 %r420, [%rd107+32];

dp4a.s32.s32 %r414, %r415, %r416, %r406;


dp4a.s32.s32 %r418, %r419, %r420, %r414;

mul.lo.s32 %r1028, %r715, %r418;

{ cvt.f32.f16 %f73, %rs1;}


ld.global.nc.u32 %r422, [%rd107];

{.reg .f16 low,high;
mov.b32 {low,high},%r422;
cvt.f32.f16 %f74, low;}


mul.ftz.f32 %f93, %f73, %f74;
cvt.rn.f32.s32 %f94, %r1028;
fma.rn.ftz.f32 %f224, %f93, %f94, %f224;
mov.u32 %r423, %r521;

prmt.b32 %r423, %r423, 0, 0xba98;

mov.u32 %r425, %r523;

prmt.b32 %r425, %r425, 0, 0xba98;

xor.b32 %r1029, %r423, %r721;
or.b32 %r1030, %r1029, -2139062144;
and.b32 %r1031, %r423, 2139062143;
sub.s32 %r1032, %r1030, %r1031;
xor.b32 %r428, %r745, %r1032;
xor.b32 %r1033, %r425, %r727;
or.b32 %r1034, %r1033, -2139062144;
and.b32 %r1035, %r425, 2139062143;
sub.s32 %r1036, %r1034, %r1035;
xor.b32 %r432, %r750, %r1036;

dp4a.s32.s32 %r427, %r428, %r380, %r528;


dp4a.s32.s32 %r431, %r432, %r384, %r427;

mov.u32 %r435, %r533;

prmt.b32 %r435, %r435, 0, 0xba98;

mov.u32 %r437, %r535;

prmt.b32 %r437, %r437, 0, 0xba98;

xor.b32 %r1037, %r435, %r756;
or.b32 %r1038, %r1037, -2139062144;
and.b32 %r1039, %r435, 2139062143;
sub.s32 %r1040, %r1038, %r1039;
xor.b32 %r440, %r780, %r1040;
xor.b32 %r1041, %r437, %r762;
or.b32 %r1042, %r1041, -2139062144;
and.b32 %r1043, %r437, 2139062143;
sub.s32 %r1044, %r1042, %r1043;
xor.b32 %r444, %r785, %r1044;

dp4a.s32.s32 %r439, %r440, %r392, %r431;


dp4a.s32.s32 %r443, %r444, %r396, %r439;

mov.u32 %r447, %r545;

prmt.b32 %r447, %r447, 0, 0xba98;

mov.u32 %r449, %r547;

prmt.b32 %r449, %r449, 0, 0xba98;

xor.b32 %r1045, %r447, %r791;
or.b32 %r1046, %r1045, -2139062144;
and.b32 %r1047, %r447, 2139062143;
sub.s32 %r1048, %r1046, %r1047;
xor.b32 %r452, %r815, %r1048;
xor.b32 %r1049, %r449, %r797;
or.b32 %r1050, %r1049, -2139062144;
and.b32 %r1051, %r449, 2139062143;
sub.s32 %r1052, %r1050, %r1051;
xor.b32 %r456, %r820, %r1052;

dp4a.s32.s32 %r451, %r452, %r404, %r443;


dp4a.s32.s32 %r455, %r456, %r408, %r451;

mov.u32 %r459, %r557;

prmt.b32 %r459, %r459, 0, 0xba98;

mov.u32 %r461, %r559;

prmt.b32 %r461, %r461, 0, 0xba98;

xor.b32 %r1053, %r459, %r826;
or.b32 %r1054, %r1053, -2139062144;
and.b32 %r1055, %r459, 2139062143;
sub.s32 %r1056, %r1054, %r1055;
xor.b32 %r464, %r850, %r1056;
xor.b32 %r1057, %r461, %r832;
or.b32 %r1058, %r1057, -2139062144;
and.b32 %r1059, %r461, 2139062143;
sub.s32 %r1060, %r1058, %r1059;
xor.b32 %r468, %r855, %r1060;

dp4a.s32.s32 %r463, %r464, %r416, %r455;


dp4a.s32.s32 %r467, %r468, %r420, %r463;

mul.lo.s32 %r1061, %r862, %r467;

{ cvt.f32.f16 %f75, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r422;
cvt.f32.f16 %f76, low;}


mul.ftz.f32 %f95, %f75, %f76;
cvt.rn.f32.s32 %f96, %r1061;
fma.rn.ftz.f32 %f223, %f95, %f96, %f223;
cvt.u64.u32 %rd108, %r1318;
add.s64 %rd109, %rd2, %rd108;
add.s64 %rd110, %rd109, %rd4;
mul.lo.s64 %rd111, %rd110, 36;

prmt.b32 %r472, %r472, 0, 0xba98;


prmt.b32 %r474, %r474, 0, 0xba98;

xor.b32 %r1062, %r472, %r574;
or.b32 %r1063, %r1062, -2139062144;
and.b32 %r1064, %r472, 2139062143;
sub.s32 %r1065, %r1063, %r1064;
xor.b32 %r477, %r598, %r1065;
xor.b32 %r1066, %r474, %r580;
or.b32 %r1067, %r1066, -2139062144;
and.b32 %r1068, %r474, 2139062143;
sub.s32 %r1069, %r1067, %r1068;
xor.b32 %r481, %r603, %r1069;
add.s64 %rd112, %rd17, %rd111;
ld.global.nc.u32 %r478, [%rd112+4];
ld.global.nc.u32 %r482, [%rd112+8];

dp4a.s32.s32 %r476, %r477, %r478, %r528;


dp4a.s32.s32 %r480, %r481, %r482, %r476;


prmt.b32 %r484, %r484, 0, 0xba98;


prmt.b32 %r486, %r486, 0, 0xba98;

xor.b32 %r1070, %r484, %r609;
or.b32 %r1071, %r1070, -2139062144;
and.b32 %r1072, %r484, 2139062143;
sub.s32 %r1073, %r1071, %r1072;
xor.b32 %r489, %r633, %r1073;
xor.b32 %r1074, %r486, %r615;
or.b32 %r1075, %r1074, -2139062144;
and.b32 %r1076, %r486, 2139062143;
sub.s32 %r1077, %r1075, %r1076;
xor.b32 %r493, %r638, %r1077;
ld.global.nc.u32 %r490, [%rd112+12];
ld.global.nc.u32 %r494, [%rd112+16];

dp4a.s32.s32 %r488, %r489, %r490, %r480;


dp4a.s32.s32 %r492, %r493, %r494, %r488;


prmt.b32 %r496, %r496, 0, 0xba98;


prmt.b32 %r498, %r498, 0, 0xba98;

xor.b32 %r1078, %r496, %r644;
or.b32 %r1079, %r1078, -2139062144;
and.b32 %r1080, %r496, 2139062143;
sub.s32 %r1081, %r1079, %r1080;
xor.b32 %r501, %r668, %r1081;
xor.b32 %r1082, %r498, %r650;
or.b32 %r1083, %r1082, -2139062144;
and.b32 %r1084, %r498, 2139062143;
sub.s32 %r1085, %r1083, %r1084;
xor.b32 %r505, %r673, %r1085;
ld.global.nc.u32 %r502, [%rd112+20];
ld.global.nc.u32 %r506, [%rd112+24];

dp4a.s32.s32 %r500, %r501, %r502, %r492;


dp4a.s32.s32 %r504, %r505, %r506, %r500;


prmt.b32 %r508, %r508, 0, 0xba98;


prmt.b32 %r510, %r510, 0, 0xba98;

xor.b32 %r1086, %r508, %r679;
or.b32 %r1087, %r1086, -2139062144;
and.b32 %r1088, %r508, 2139062143;
sub.s32 %r1089, %r1087, %r1088;
xor.b32 %r513, %r703, %r1089;
xor.b32 %r1090, %r510, %r685;
or.b32 %r1091, %r1090, -2139062144;
and.b32 %r1092, %r510, 2139062143;
sub.s32 %r1093, %r1091, %r1092;
xor.b32 %r517, %r708, %r1093;
ld.global.nc.u32 %r514, [%rd112+28];
ld.global.nc.u32 %r518, [%rd112+32];

dp4a.s32.s32 %r512, %r513, %r514, %r504;


dp4a.s32.s32 %r516, %r517, %r518, %r512;

mul.lo.s32 %r1094, %r715, %r516;

{ cvt.f32.f16 %f77, %rs1;}


ld.global.nc.u32 %r520, [%rd112];

{.reg .f16 low,high;
mov.b32 {low,high},%r520;
cvt.f32.f16 %f78, low;}


mul.ftz.f32 %f97, %f77, %f78;
cvt.rn.f32.s32 %f98, %r1094;
fma.rn.ftz.f32 %f222, %f97, %f98, %f222;

prmt.b32 %r521, %r521, 0, 0xba98;


prmt.b32 %r523, %r523, 0, 0xba98;

xor.b32 %r1095, %r521, %r721;
or.b32 %r1096, %r1095, -2139062144;
and.b32 %r1097, %r521, 2139062143;
sub.s32 %r1098, %r1096, %r1097;
xor.b32 %r526, %r745, %r1098;
xor.b32 %r1099, %r523, %r727;
or.b32 %r1100, %r1099, -2139062144;
and.b32 %r1101, %r523, 2139062143;
sub.s32 %r1102, %r1100, %r1101;
xor.b32 %r530, %r750, %r1102;

dp4a.s32.s32 %r525, %r526, %r478, %r528;


dp4a.s32.s32 %r529, %r530, %r482, %r525;


prmt.b32 %r533, %r533, 0, 0xba98;


prmt.b32 %r535, %r535, 0, 0xba98;

xor.b32 %r1103, %r533, %r756;
or.b32 %r1104, %r1103, -2139062144;
and.b32 %r1105, %r533, 2139062143;
sub.s32 %r1106, %r1104, %r1105;
xor.b32 %r538, %r780, %r1106;
xor.b32 %r1107, %r535, %r762;
or.b32 %r1108, %r1107, -2139062144;
and.b32 %r1109, %r535, 2139062143;
sub.s32 %r1110, %r1108, %r1109;
xor.b32 %r542, %r785, %r1110;

dp4a.s32.s32 %r537, %r538, %r490, %r529;


dp4a.s32.s32 %r541, %r542, %r494, %r537;


prmt.b32 %r545, %r545, 0, 0xba98;


prmt.b32 %r547, %r547, 0, 0xba98;

xor.b32 %r1111, %r545, %r791;
or.b32 %r1112, %r1111, -2139062144;
and.b32 %r1113, %r545, 2139062143;
sub.s32 %r1114, %r1112, %r1113;
xor.b32 %r550, %r815, %r1114;
xor.b32 %r1115, %r547, %r797;
or.b32 %r1116, %r1115, -2139062144;
and.b32 %r1117, %r547, 2139062143;
sub.s32 %r1118, %r1116, %r1117;
xor.b32 %r554, %r820, %r1118;

dp4a.s32.s32 %r549, %r550, %r502, %r541;


dp4a.s32.s32 %r553, %r554, %r506, %r549;


prmt.b32 %r557, %r557, 0, 0xba98;


prmt.b32 %r559, %r559, 0, 0xba98;

xor.b32 %r1119, %r557, %r826;
or.b32 %r1120, %r1119, -2139062144;
and.b32 %r1121, %r557, 2139062143;
sub.s32 %r1122, %r1120, %r1121;
xor.b32 %r562, %r850, %r1122;
xor.b32 %r1123, %r559, %r832;
or.b32 %r1124, %r1123, -2139062144;
and.b32 %r1125, %r559, 2139062143;
sub.s32 %r1126, %r1124, %r1125;
xor.b32 %r566, %r855, %r1126;

dp4a.s32.s32 %r561, %r562, %r514, %r553;


dp4a.s32.s32 %r565, %r566, %r518, %r561;

mul.lo.s32 %r1127, %r862, %r565;

{ cvt.f32.f16 %f79, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r520;
cvt.f32.f16 %f80, low;}


mul.ftz.f32 %f99, %f79, %f80;
cvt.rn.f32.s32 %f100, %r1127;
fma.rn.ftz.f32 %f221, %f99, %f100, %f221;
add.s32 %r1320, %r1320, 64;
add.s32 %r1319, %r1319, 64;
add.s32 %r1318, %r1318, 64;
add.s32 %r1317, %r1317, 64;
add.s32 %r1316, %r1316, 64;
add.s64 %rd136, %rd136, 880;
add.s32 %r1321, %r1321, 8;
setp.lt.s32 %p4, %r1321, %r5;
@%p4 bra $L__BB313_2;

$L__BB313_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB313_5;

mov.u32 %r1129, _ZZ13mul_mat_vec_qIL9ggml_type21ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r1130, %r1, 1280, %r1129;
shl.b32 %r1131, %r43, 2;
add.s32 %r1132, %r1130, %r1131;
st.shared.f32 [%r1132+-1280], %f230;
st.shared.f32 [%r1132+-1152], %f229;
st.shared.f32 [%r1132+-1024], %f228;
st.shared.f32 [%r1132+-896], %f227;
st.shared.f32 [%r1132+-768], %f226;
st.shared.f32 [%r1132+-640], %f225;
st.shared.f32 [%r1132+-512], %f224;
st.shared.f32 [%r1132+-384], %f223;
st.shared.f32 [%r1132+-256], %f222;
st.shared.f32 [%r1132+-128], %f221;

$L__BB313_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB313_26;

mov.u32 %r1133, %ctaid.z;
mov.u32 %r1134, %ctaid.y;
mad.lo.s32 %r1135, %r1134, %r34, %r3;
mad.lo.s32 %r1136, %r1133, %r40, %r1135;
cvt.u64.u32 %rd20, %r1136;
shl.b32 %r1138, %r43, 2;
mov.u32 %r1139, _ZZ13mul_mat_vec_qIL9ggml_type21ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r24, %r1139, %r1138;
mov.u32 %r1140, 2;
ld.shared.f32 %f101, [%r24];
add.ftz.f32 %f102, %f101, %f230;
mov.b32 %r1141, %f102;
mov.u32 %r1142, 31;
mov.u32 %r1143, 16;
mov.u32 %r1144, -1;
shfl.sync.bfly.b32 %r1145|%p7, %r1141, %r1143, %r1142, %r1144;
mov.b32 %f103, %r1145;
add.ftz.f32 %f104, %f102, %f103;
mov.b32 %r1146, %f104;
mov.u32 %r1147, 8;
shfl.sync.bfly.b32 %r1148|%p8, %r1146, %r1147, %r1142, %r1144;
mov.b32 %f105, %r1148;
add.ftz.f32 %f106, %f104, %f105;
mov.b32 %r1149, %f106;
mov.u32 %r1150, 4;
shfl.sync.bfly.b32 %r1151|%p9, %r1149, %r1150, %r1142, %r1144;
mov.b32 %f107, %r1151;
add.ftz.f32 %f108, %f106, %f107;
mov.b32 %r1152, %f108;
shfl.sync.bfly.b32 %r1153|%p10, %r1152, %r1140, %r1142, %r1144;
mov.b32 %f109, %r1153;
add.ftz.f32 %f110, %f108, %f109;
mov.b32 %r1154, %f110;
mov.u32 %r1155, 1;
shfl.sync.bfly.b32 %r1156|%p11, %r1154, %r1155, %r1142, %r1144;
mov.b32 %f111, %r1156;
add.ftz.f32 %f31, %f110, %f111;
setp.eq.s32 %p12, %r43, 0;
setp.lt.u32 %p13, %r3, %r28;
and.pred %p1, %p12, %p13;
mul.wide.u32 %rd113, %r1136, 4;
add.s64 %rd21, %rd1, %rd113;
not.pred %p14, %p1;
@%p14 bra $L__BB313_8;

st.global.f32 [%rd21], %f31;

$L__BB313_8:
ld.shared.f32 %f112, [%r24+128];
add.ftz.f32 %f113, %f112, %f229;
mov.b32 %r1157, %f113;
shfl.sync.bfly.b32 %r1161|%p15, %r1157, %r1143, %r1142, %r1144;
mov.b32 %f114, %r1161;
add.ftz.f32 %f115, %f113, %f114;
mov.b32 %r1162, %f115;
shfl.sync.bfly.b32 %r1164|%p16, %r1162, %r1147, %r1142, %r1144;
mov.b32 %f116, %r1164;
add.ftz.f32 %f117, %f115, %f116;
mov.b32 %r1165, %f117;
shfl.sync.bfly.b32 %r1167|%p17, %r1165, %r1150, %r1142, %r1144;
mov.b32 %f118, %r1167;
add.ftz.f32 %f119, %f117, %f118;
mov.b32 %r1168, %f119;
shfl.sync.bfly.b32 %r1170|%p18, %r1168, %r1140, %r1142, %r1144;
mov.b32 %f120, %r1170;
add.ftz.f32 %f121, %f119, %f120;
mov.b32 %r1171, %f121;
shfl.sync.bfly.b32 %r1173|%p19, %r1171, %r1155, %r1142, %r1144;
mov.b32 %f122, %r1173;
add.ftz.f32 %f32, %f121, %f122;
setp.eq.s32 %p20, %r43, 1;
add.s32 %r1175, %r3, 1;
setp.lt.u32 %p21, %r1175, %r28;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB313_10;

st.global.f32 [%rd21+4], %f32;

$L__BB313_10:
ld.shared.f32 %f123, [%r24+256];
add.ftz.f32 %f124, %f123, %f228;
mov.b32 %r1176, %f124;
mov.u32 %r1177, 31;
mov.u32 %r1178, 16;
mov.u32 %r1179, -1;
shfl.sync.bfly.b32 %r1180|%p23, %r1176, %r1178, %r1177, %r1179;
mov.b32 %f125, %r1180;
add.ftz.f32 %f126, %f124, %f125;
mov.b32 %r1181, %f126;
mov.u32 %r1182, 8;
shfl.sync.bfly.b32 %r1183|%p24, %r1181, %r1182, %r1177, %r1179;
mov.b32 %f127, %r1183;
add.ftz.f32 %f128, %f126, %f127;
mov.b32 %r1184, %f128;
mov.u32 %r1185, 4;
shfl.sync.bfly.b32 %r1186|%p25, %r1184, %r1185, %r1177, %r1179;
mov.b32 %f129, %r1186;
add.ftz.f32 %f130, %f128, %f129;
mov.b32 %r1187, %f130;
mov.u32 %r1188, 2;
shfl.sync.bfly.b32 %r1189|%p26, %r1187, %r1188, %r1177, %r1179;
mov.b32 %f131, %r1189;
add.ftz.f32 %f132, %f130, %f131;
mov.b32 %r1190, %f132;
mov.u32 %r1191, 1;
shfl.sync.bfly.b32 %r1192|%p27, %r1190, %r1191, %r1177, %r1179;
mov.b32 %f133, %r1192;
add.ftz.f32 %f33, %f132, %f133;
@%p14 bra $L__BB313_12;

cvt.u64.u32 %rd114, %r28;
add.s64 %rd115, %rd114, %rd20;
shl.b64 %rd116, %rd115, 2;
add.s64 %rd117, %rd1, %rd116;
st.global.f32 [%rd117], %f33;

$L__BB313_12:
ld.shared.f32 %f134, [%r24+384];
add.ftz.f32 %f135, %f134, %f227;
mov.b32 %r1193, %f135;
shfl.sync.bfly.b32 %r1197|%p29, %r1193, %r1178, %r1177, %r1179;
mov.b32 %f136, %r1197;
add.ftz.f32 %f137, %f135, %f136;
mov.b32 %r1198, %f137;
shfl.sync.bfly.b32 %r1200|%p30, %r1198, %r1182, %r1177, %r1179;
mov.b32 %f138, %r1200;
add.ftz.f32 %f139, %f137, %f138;
mov.b32 %r1201, %f139;
shfl.sync.bfly.b32 %r1203|%p31, %r1201, %r1185, %r1177, %r1179;
mov.b32 %f140, %r1203;
add.ftz.f32 %f141, %f139, %f140;
mov.b32 %r1204, %f141;
shfl.sync.bfly.b32 %r1206|%p32, %r1204, %r1188, %r1177, %r1179;
mov.b32 %f142, %r1206;
add.ftz.f32 %f143, %f141, %f142;
mov.b32 %r1207, %f143;
shfl.sync.bfly.b32 %r1209|%p33, %r1207, %r1191, %r1177, %r1179;
mov.b32 %f144, %r1209;
add.ftz.f32 %f34, %f143, %f144;
@%p22 bra $L__BB313_14;

add.s32 %r1210, %r28, 1;
cvt.u64.u32 %rd118, %r1210;
add.s64 %rd119, %rd118, %rd20;
shl.b64 %rd120, %rd119, 2;
add.s64 %rd121, %rd1, %rd120;
st.global.f32 [%rd121], %f34;

$L__BB313_14:
shl.b32 %r1211, %r28, 1;
mov.u32 %r1212, 1;
ld.shared.f32 %f145, [%r24+512];
add.ftz.f32 %f146, %f145, %f226;
mov.b32 %r1213, %f146;
mov.u32 %r1214, 31;
mov.u32 %r1215, 16;
mov.u32 %r1216, -1;
shfl.sync.bfly.b32 %r1217|%p35, %r1213, %r1215, %r1214, %r1216;
mov.b32 %f147, %r1217;
add.ftz.f32 %f148, %f146, %f147;
mov.b32 %r1218, %f148;
mov.u32 %r1219, 8;
shfl.sync.bfly.b32 %r1220|%p36, %r1218, %r1219, %r1214, %r1216;
mov.b32 %f149, %r1220;
add.ftz.f32 %f150, %f148, %f149;
mov.b32 %r1221, %f150;
mov.u32 %r1222, 4;
shfl.sync.bfly.b32 %r1223|%p37, %r1221, %r1222, %r1214, %r1216;
mov.b32 %f151, %r1223;
add.ftz.f32 %f152, %f150, %f151;
mov.b32 %r1224, %f152;
mov.u32 %r1225, 2;
shfl.sync.bfly.b32 %r1226|%p38, %r1224, %r1225, %r1214, %r1216;
mov.b32 %f153, %r1226;
add.ftz.f32 %f154, %f152, %f153;
mov.b32 %r1227, %f154;
shfl.sync.bfly.b32 %r1228|%p39, %r1227, %r1212, %r1214, %r1216;
mov.b32 %f155, %r1228;
add.ftz.f32 %f35, %f154, %f155;
cvt.u64.u32 %rd122, %r1211;
add.s64 %rd123, %rd122, %rd20;
shl.b64 %rd124, %rd123, 2;
add.s64 %rd22, %rd1, %rd124;
@%p14 bra $L__BB313_16;

st.global.f32 [%rd22], %f35;

$L__BB313_16:
ld.shared.f32 %f156, [%r24+640];
add.ftz.f32 %f157, %f156, %f225;
mov.b32 %r1229, %f157;
shfl.sync.bfly.b32 %r1233|%p41, %r1229, %r1215, %r1214, %r1216;
mov.b32 %f158, %r1233;
add.ftz.f32 %f159, %f157, %f158;
mov.b32 %r1234, %f159;
shfl.sync.bfly.b32 %r1236|%p42, %r1234, %r1219, %r1214, %r1216;
mov.b32 %f160, %r1236;
add.ftz.f32 %f161, %f159, %f160;
mov.b32 %r1237, %f161;
shfl.sync.bfly.b32 %r1239|%p43, %r1237, %r1222, %r1214, %r1216;
mov.b32 %f162, %r1239;
add.ftz.f32 %f163, %f161, %f162;
mov.b32 %r1240, %f163;
shfl.sync.bfly.b32 %r1242|%p44, %r1240, %r1225, %r1214, %r1216;
mov.b32 %f164, %r1242;
add.ftz.f32 %f165, %f163, %f164;
mov.b32 %r1243, %f165;
shfl.sync.bfly.b32 %r1245|%p45, %r1243, %r1212, %r1214, %r1216;
mov.b32 %f166, %r1245;
add.ftz.f32 %f36, %f165, %f166;
@%p22 bra $L__BB313_18;

st.global.f32 [%rd22+4], %f36;

$L__BB313_18:
mul.lo.s32 %r25, %r28, 3;
ld.shared.f32 %f167, [%r24+768];
add.ftz.f32 %f168, %f167, %f224;
mov.b32 %r1246, %f168;
mov.u32 %r1247, 31;
mov.u32 %r1248, 16;
mov.u32 %r1249, -1;
shfl.sync.bfly.b32 %r1250|%p47, %r1246, %r1248, %r1247, %r1249;
mov.b32 %f169, %r1250;
add.ftz.f32 %f170, %f168, %f169;
mov.b32 %r1251, %f170;
mov.u32 %r1252, 8;
shfl.sync.bfly.b32 %r1253|%p48, %r1251, %r1252, %r1247, %r1249;
mov.b32 %f171, %r1253;
add.ftz.f32 %f172, %f170, %f171;
mov.b32 %r1254, %f172;
mov.u32 %r1255, 4;
shfl.sync.bfly.b32 %r1256|%p49, %r1254, %r1255, %r1247, %r1249;
mov.b32 %f173, %r1256;
add.ftz.f32 %f174, %f172, %f173;
mov.b32 %r1257, %f174;
mov.u32 %r1258, 2;
shfl.sync.bfly.b32 %r1259|%p50, %r1257, %r1258, %r1247, %r1249;
mov.b32 %f175, %r1259;
add.ftz.f32 %f176, %f174, %f175;
mov.b32 %r1260, %f176;
mov.u32 %r1261, 1;
shfl.sync.bfly.b32 %r1262|%p51, %r1260, %r1261, %r1247, %r1249;
mov.b32 %f177, %r1262;
add.ftz.f32 %f37, %f176, %f177;
@%p14 bra $L__BB313_20;

cvt.u64.u32 %rd125, %r25;
add.s64 %rd126, %rd125, %rd20;
shl.b64 %rd127, %rd126, 2;
add.s64 %rd128, %rd1, %rd127;
st.global.f32 [%rd128], %f37;

$L__BB313_20:
ld.shared.f32 %f178, [%r24+896];
add.ftz.f32 %f179, %f178, %f223;
mov.b32 %r1263, %f179;
shfl.sync.bfly.b32 %r1267|%p53, %r1263, %r1248, %r1247, %r1249;
mov.b32 %f180, %r1267;
add.ftz.f32 %f181, %f179, %f180;
mov.b32 %r1268, %f181;
shfl.sync.bfly.b32 %r1270|%p54, %r1268, %r1252, %r1247, %r1249;
mov.b32 %f182, %r1270;
add.ftz.f32 %f183, %f181, %f182;
mov.b32 %r1271, %f183;
shfl.sync.bfly.b32 %r1273|%p55, %r1271, %r1255, %r1247, %r1249;
mov.b32 %f184, %r1273;
add.ftz.f32 %f185, %f183, %f184;
mov.b32 %r1274, %f185;
shfl.sync.bfly.b32 %r1276|%p56, %r1274, %r1258, %r1247, %r1249;
mov.b32 %f186, %r1276;
add.ftz.f32 %f187, %f185, %f186;
mov.b32 %r1277, %f187;
shfl.sync.bfly.b32 %r1279|%p57, %r1277, %r1261, %r1247, %r1249;
mov.b32 %f188, %r1279;
add.ftz.f32 %f38, %f187, %f188;
@%p22 bra $L__BB313_22;

add.s32 %r1280, %r25, 1;
cvt.u64.u32 %rd129, %r1280;
add.s64 %rd130, %rd129, %rd20;
shl.b64 %rd131, %rd130, 2;
add.s64 %rd132, %rd1, %rd131;
st.global.f32 [%rd132], %f38;

$L__BB313_22:
shl.b32 %r1281, %r28, 2;
mov.u32 %r1282, 2;
ld.shared.f32 %f189, [%r24+1024];
add.ftz.f32 %f190, %f189, %f222;
mov.b32 %r1283, %f190;
mov.u32 %r1284, 31;
mov.u32 %r1285, 16;
mov.u32 %r1286, -1;
shfl.sync.bfly.b32 %r1287|%p59, %r1283, %r1285, %r1284, %r1286;
mov.b32 %f191, %r1287;
add.ftz.f32 %f192, %f190, %f191;
mov.b32 %r1288, %f192;
mov.u32 %r1289, 8;
shfl.sync.bfly.b32 %r1290|%p60, %r1288, %r1289, %r1284, %r1286;
mov.b32 %f193, %r1290;
add.ftz.f32 %f194, %f192, %f193;
mov.b32 %r1291, %f194;
mov.u32 %r1292, 4;
shfl.sync.bfly.b32 %r1293|%p61, %r1291, %r1292, %r1284, %r1286;
mov.b32 %f195, %r1293;
add.ftz.f32 %f196, %f194, %f195;
mov.b32 %r1294, %f196;
shfl.sync.bfly.b32 %r1295|%p62, %r1294, %r1282, %r1284, %r1286;
mov.b32 %f197, %r1295;
add.ftz.f32 %f198, %f196, %f197;
mov.b32 %r1296, %f198;
mov.u32 %r1297, 1;
shfl.sync.bfly.b32 %r1298|%p63, %r1296, %r1297, %r1284, %r1286;
mov.b32 %f199, %r1298;
add.ftz.f32 %f39, %f198, %f199;
cvt.u64.u32 %rd133, %r1281;
add.s64 %rd134, %rd133, %rd20;
shl.b64 %rd135, %rd134, 2;
add.s64 %rd23, %rd1, %rd135;
@%p14 bra $L__BB313_24;

st.global.f32 [%rd23], %f39;

$L__BB313_24:
ld.shared.f32 %f200, [%r24+1152];
add.ftz.f32 %f201, %f200, %f221;
mov.b32 %r1299, %f201;
shfl.sync.bfly.b32 %r1303|%p65, %r1299, %r1285, %r1284, %r1286;
mov.b32 %f202, %r1303;
add.ftz.f32 %f203, %f201, %f202;
mov.b32 %r1304, %f203;
shfl.sync.bfly.b32 %r1306|%p66, %r1304, %r1289, %r1284, %r1286;
mov.b32 %f204, %r1306;
add.ftz.f32 %f205, %f203, %f204;
mov.b32 %r1307, %f205;
shfl.sync.bfly.b32 %r1309|%p67, %r1307, %r1292, %r1284, %r1286;
mov.b32 %f206, %r1309;
add.ftz.f32 %f207, %f205, %f206;
mov.b32 %r1310, %f207;
shfl.sync.bfly.b32 %r1312|%p68, %r1310, %r1282, %r1284, %r1286;
mov.b32 %f208, %r1312;
add.ftz.f32 %f209, %f207, %f208;
mov.b32 %r1313, %f209;
shfl.sync.bfly.b32 %r1315|%p69, %r1313, %r1297, %r1284, %r1286;
mov.b32 %f210, %r1315;
add.ftz.f32 %f40, %f209, %f210;
@%p22 bra $L__BB313_26;

st.global.f32 [%rd23+4], %f40;

$L__BB313_26:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type21ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type21ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type21ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type21ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type21ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type21ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type21ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type21ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type21ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<83>;
.reg .b16 %rs<57>;
.reg .f32 %f<277>;
.reg .b32 %r<1526>;
.reg .b64 %rd<150>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type21ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1536];

ld.param.u64 %rd24, [_Z13mul_mat_vec_qIL9ggml_type21ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd25, [_Z13mul_mat_vec_qIL9ggml_type21ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd26, [_Z13mul_mat_vec_qIL9ggml_type21ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r45, [_Z13mul_mat_vec_qIL9ggml_type21ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r30, [_Z13mul_mat_vec_qIL9ggml_type21ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r31, [_Z13mul_mat_vec_qIL9ggml_type21ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r32, [_Z13mul_mat_vec_qIL9ggml_type21ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type21ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type21ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type21ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type21ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type21ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type21ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type21ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type21ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type21ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type21ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
cvta.to.global.u64 %rd1, %rd26;
mov.u32 %r1, %tid.y;
shl.b32 %r46, %r1, 5;
mov.u32 %r47, %tid.x;
add.s32 %r2, %r46, %r47;
mov.u32 %r48, %ctaid.x;
shl.b32 %r3, %r48, 1;
shr.s32 %r49, %r2, 31;
shr.u32 %r50, %r49, 29;
add.s32 %r51, %r2, %r50;
shr.s32 %r1525, %r51, 3;
shr.u32 %r5, %r45, 8;
setp.ge.s32 %p3, %r1525, %r5;
mov.f32 %f265, 0f00000000;
mov.f32 %f266, %f265;
mov.f32 %f267, %f265;
mov.f32 %f268, %f265;
mov.f32 %f269, %f265;
mov.f32 %f270, %f265;
mov.f32 %f271, %f265;
mov.f32 %f272, %f265;
mov.f32 %f273, %f265;
mov.f32 %f274, %f265;
mov.f32 %f275, %f265;
mov.f32 %f276, %f265;
@%p3 bra $L__BB314_3;

mov.u32 %r52, %ctaid.y;
mov.u32 %r53, %ctaid.z;
mul.lo.s32 %r54, %r52, %r37;
cvt.u64.u32 %rd27, %r54;
mul.lo.s32 %r55, %r53, %r43;
cvt.u64.u32 %rd28, %r55;
add.s64 %rd2, %rd28, %rd27;
mul.hi.u32 %r56, %r53, %r39;
add.s32 %r57, %r56, %r53;
shr.u32 %r58, %r57, %r40;
mul.lo.s32 %r59, %r58, %r42;
mul.hi.u32 %r60, %r52, %r33;
add.s32 %r61, %r60, %r52;
shr.u32 %r62, %r61, %r34;
mul.lo.s32 %r63, %r62, %r36;
cvta.to.global.u64 %rd149, %rd24;
and.b32 %r67, %r51, -8;
sub.s32 %r68, %r2, %r67;
shl.b32 %r69, %r68, 1;
shr.u32 %r70, %r68, 31;
add.s32 %r71, %r68, %r70;
shr.s32 %r72, %r71, 1;
cvt.s64.s32 %rd29, %r72;
shl.b32 %r73, %r68, 2;
cvt.s64.s32 %rd4, %r68;
and.b32 %r6, %r73, 4;
shl.b32 %r74, %r31, 1;
shl.b32 %r1519, %r1525, 3;
add.s32 %r1524, %r1519, %r74;
mad.lo.s32 %r1523, %r31, 3, %r1519;
shl.b32 %r75, %r31, 2;
add.s32 %r1522, %r1519, %r75;
mad.lo.s32 %r1521, %r31, 5, %r1519;
add.s32 %r1520, %r31, %r1519;
add.s32 %r76, %r1525, %r59;
add.s32 %r77, %r76, %r63;
add.s32 %r78, %r3, 1;
mad.lo.s32 %r79, %r30, %r78, %r77;
mul.wide.s32 %rd5, %r79, 110;
add.s64 %rd30, %rd5, %rd29;
add.s64 %rd6, %rd30, 106;
add.s64 %rd31, %rd5, 74;
not.b32 %r80, %r69;
mul.wide.s32 %rd32, %r80, 2;
sub.s64 %rd7, %rd31, %rd32;
mul.wide.s32 %rd33, %r69, 2;
add.s64 %rd34, %rd5, %rd33;
add.s64 %rd8, %rd34, 74;
add.s64 %rd35, %rd5, %rd4;
add.s64 %rd9, %rd35, 66;
mad.lo.s32 %r81, %r3, %r30, %r1525;
add.s32 %r82, %r81, %r59;
add.s32 %r83, %r82, %r63;
mul.wide.s32 %rd10, %r83, 110;
mul.wide.s32 %rd36, %r73, 2;
add.s64 %rd37, %rd10, %rd36;
add.s64 %rd11, %rd37, 4;
add.s64 %rd38, %rd5, %rd36;
add.s64 %rd12, %rd38, 2;
add.s64 %rd39, %rd10, %rd29;
add.s64 %rd13, %rd39, 106;
add.s64 %rd40, %rd10, 74;
sub.s64 %rd14, %rd40, %rd32;
add.s64 %rd41, %rd10, %rd33;
add.s64 %rd15, %rd41, 74;
add.s64 %rd42, %rd10, %rd4;
add.s64 %rd16, %rd42, 66;
cvta.to.global.u64 %rd17, %rd25;
mov.f32 %f265, 0f00000000;

$L__BB314_2:
cvt.u64.u32 %rd43, %r1519;
add.s64 %rd44, %rd2, %rd43;
add.s64 %rd45, %rd44, %rd4;
mul.lo.s64 %rd46, %rd45, 36;
add.s64 %rd47, %rd149, %rd11;
ld.global.nc.u16 %rs13, [%rd47];
ld.global.nc.u16 %rs14, [%rd47+2];
ld.global.nc.u16 %rs15, [%rd47+4];
add.s64 %rd48, %rd149, %rd16;
ld.global.nc.u8 %rs16, [%rd48];
add.s64 %rd49, %rd149, %rd15;
add.s64 %rd50, %rd149, %rd14;
ld.global.nc.u16 %rs17, [%rd49];
ld.global.nc.u16 %rs18, [%rd50];
ld.global.nc.u16 %rs19, [%rd47+-2];
and.b16 %rs20, %rs19, 255;
cvt.u32.u16 %r672, %rs20;
shl.b16 %rs21, %rs16, 8;
cvt.u32.u16 %r673, %rs21;
and.b32 %r674, %r673, 256;
or.b32 %r675, %r674, %r672;
mul.wide.u32 %rd51, %r675, 4;
mov.u64 %rd52, iq3s_grid;
add.s64 %rd53, %rd52, %rd51;
ld.global.nc.u32 %r676, [%rd53];
not.b32 %r677, %r676;
shr.u16 %rs22, %rs19, 8;
cvt.u32.u16 %r678, %rs22;
mul.wide.u16 %r679, %rs16, 128;
and.b32 %r680, %r679, 256;
or.b32 %r681, %r680, %r678;
mul.wide.u32 %rd54, %r681, 4;
add.s64 %rd55, %rd52, %rd54;
ld.global.nc.u32 %r682, [%rd55];
not.b32 %r683, %r682;
and.b16 %rs24, %rs17, 255;
mul.wide.u16 %r684, %rs24, 128;
cvt.u32.u16 %r685, %rs17;
shl.b32 %r686, %r685, 21;
or.b32 %r687, %r684, %r686;
and.b32 %r688, %r687, 25166208;
or.b32 %r689, %r688, -2139062144;
add.s32 %r690, %r689, -16843009;
or.b32 %r574, %r690, %r688;
mov.u32 %r84, %r574;

prmt.b32 %r84, %r84, 0, 0xba98;

mul.wide.u16 %r691, %rs24, 8;
shl.b32 %r692, %r685, 17;
or.b32 %r693, %r691, %r692;
and.b32 %r694, %r693, 25166208;
or.b32 %r695, %r694, -2139062144;
add.s32 %r696, %r695, -16843009;
or.b32 %r576, %r696, %r694;
mov.u32 %r86, %r576;

prmt.b32 %r86, %r86, 0, 0xba98;

xor.b32 %r697, %r84, %r676;
or.b32 %r698, %r697, -2139062144;
and.b32 %r699, %r84, 2139062143;
and.b32 %r700, %r677, -2139062144;
sub.s32 %r701, %r698, %r699;
xor.b32 %r89, %r700, %r701;
xor.b32 %r702, %r86, %r682;
or.b32 %r703, %r702, -2139062144;
and.b32 %r704, %r86, 2139062143;
and.b32 %r705, %r683, -2139062144;
sub.s32 %r706, %r703, %r704;
xor.b32 %r93, %r705, %r706;
add.s64 %rd56, %rd17, %rd46;
ld.global.nc.u32 %r90, [%rd56+4];
ld.global.nc.u32 %r94, [%rd56+8];
mov.u32 %r630, 0;

dp4a.s32.s32 %r88, %r89, %r90, %r630;


dp4a.s32.s32 %r92, %r93, %r94, %r88;

and.b16 %rs25, %rs13, 255;
cvt.u32.u16 %r707, %rs25;
mul.wide.u16 %r708, %rs16, 64;
and.b32 %r709, %r708, 256;
or.b32 %r710, %r709, %r707;
mul.wide.u32 %rd57, %r710, 4;
add.s64 %rd58, %rd52, %rd57;
ld.global.nc.u32 %r711, [%rd58];
not.b32 %r712, %r711;
shr.u16 %rs26, %rs13, 8;
cvt.u32.u16 %r713, %rs26;
mul.wide.u16 %r714, %rs16, 32;
and.b32 %r715, %r714, 256;
or.b32 %r716, %r715, %r713;
mul.wide.u32 %rd59, %r716, 4;
add.s64 %rd60, %rd52, %rd59;
ld.global.nc.u32 %r717, [%rd60];
not.b32 %r718, %r717;
shr.u16 %rs27, %rs17, 8;
mul.wide.u16 %r719, %rs27, 128;
cvt.u32.u16 %r720, %rs27;
shl.b32 %r721, %r720, 21;
or.b32 %r722, %r719, %r721;
and.b32 %r723, %r722, 25166208;
or.b32 %r724, %r723, -2139062144;
add.s32 %r725, %r724, -16843009;
or.b32 %r586, %r725, %r723;
mov.u32 %r96, %r586;

prmt.b32 %r96, %r96, 0, 0xba98;

mul.wide.u16 %r726, %rs27, 8;
shl.b32 %r727, %r720, 17;
or.b32 %r728, %r726, %r727;
and.b32 %r729, %r728, 25166208;
or.b32 %r730, %r729, -2139062144;
add.s32 %r731, %r730, -16843009;
or.b32 %r588, %r731, %r729;
mov.u32 %r98, %r588;

prmt.b32 %r98, %r98, 0, 0xba98;

xor.b32 %r732, %r96, %r711;
or.b32 %r733, %r732, -2139062144;
and.b32 %r734, %r96, 2139062143;
and.b32 %r735, %r712, -2139062144;
sub.s32 %r736, %r733, %r734;
xor.b32 %r101, %r735, %r736;
xor.b32 %r737, %r98, %r717;
or.b32 %r738, %r737, -2139062144;
and.b32 %r739, %r98, 2139062143;
and.b32 %r740, %r718, -2139062144;
sub.s32 %r741, %r738, %r739;
xor.b32 %r105, %r740, %r741;
ld.global.nc.u32 %r102, [%rd56+12];
ld.global.nc.u32 %r106, [%rd56+16];

dp4a.s32.s32 %r100, %r101, %r102, %r92;


dp4a.s32.s32 %r104, %r105, %r106, %r100;

and.b16 %rs28, %rs14, 255;
cvt.u32.u16 %r742, %rs28;
mul.wide.u16 %r743, %rs16, 16;
and.b32 %r744, %r743, 256;
or.b32 %r745, %r744, %r742;
mul.wide.u32 %rd61, %r745, 4;
add.s64 %rd62, %rd52, %rd61;
ld.global.nc.u32 %r746, [%rd62];
not.b32 %r747, %r746;
shr.u16 %rs29, %rs14, 8;
cvt.u32.u16 %r748, %rs29;
mul.wide.u16 %r749, %rs16, 8;
and.b32 %r750, %r749, 256;
or.b32 %r751, %r750, %r748;
mul.wide.u32 %rd63, %r751, 4;
add.s64 %rd64, %rd52, %rd63;
ld.global.nc.u32 %r752, [%rd64];
not.b32 %r753, %r752;
and.b16 %rs30, %rs18, 255;
mul.wide.u16 %r754, %rs30, 128;
cvt.u32.u16 %r755, %rs18;
shl.b32 %r756, %r755, 21;
or.b32 %r757, %r754, %r756;
and.b32 %r758, %r757, 25166208;
or.b32 %r759, %r758, -2139062144;
add.s32 %r760, %r759, -16843009;
or.b32 %r598, %r760, %r758;
mov.u32 %r108, %r598;

prmt.b32 %r108, %r108, 0, 0xba98;

mul.wide.u16 %r761, %rs30, 8;
shl.b32 %r762, %r755, 17;
or.b32 %r763, %r761, %r762;
and.b32 %r764, %r763, 25166208;
or.b32 %r765, %r764, -2139062144;
add.s32 %r766, %r765, -16843009;
or.b32 %r600, %r766, %r764;
mov.u32 %r110, %r600;

prmt.b32 %r110, %r110, 0, 0xba98;

xor.b32 %r767, %r108, %r746;
or.b32 %r768, %r767, -2139062144;
and.b32 %r769, %r108, 2139062143;
and.b32 %r770, %r747, -2139062144;
sub.s32 %r771, %r768, %r769;
xor.b32 %r113, %r770, %r771;
xor.b32 %r772, %r110, %r752;
or.b32 %r773, %r772, -2139062144;
and.b32 %r774, %r110, 2139062143;
and.b32 %r775, %r753, -2139062144;
sub.s32 %r776, %r773, %r774;
xor.b32 %r117, %r775, %r776;
ld.global.nc.u32 %r114, [%rd56+20];
ld.global.nc.u32 %r118, [%rd56+24];

dp4a.s32.s32 %r112, %r113, %r114, %r104;


dp4a.s32.s32 %r116, %r117, %r118, %r112;

and.b16 %rs31, %rs15, 255;
cvt.u32.u16 %r777, %rs31;
mul.wide.u16 %r778, %rs16, 4;
and.b32 %r779, %r778, 256;
or.b32 %r780, %r779, %r777;
mul.wide.u32 %rd65, %r780, 4;
add.s64 %rd66, %rd52, %rd65;
ld.global.nc.u32 %r781, [%rd66];
not.b32 %r782, %r781;
shr.u16 %rs32, %rs15, 8;
cvt.u32.u16 %r783, %rs32;
mul.wide.u16 %r784, %rs16, 2;
and.b32 %r785, %r784, 256;
or.b32 %r786, %r785, %r783;
mul.wide.u32 %rd67, %r786, 4;
add.s64 %rd68, %rd52, %rd67;
ld.global.nc.u32 %r787, [%rd68];
not.b32 %r788, %r787;
shr.u16 %rs33, %rs18, 8;
mul.wide.u16 %r789, %rs33, 128;
cvt.u32.u16 %r790, %rs33;
shl.b32 %r791, %r790, 21;
or.b32 %r792, %r789, %r791;
and.b32 %r793, %r792, 25166208;
or.b32 %r794, %r793, -2139062144;
add.s32 %r795, %r794, -16843009;
or.b32 %r610, %r795, %r793;
mov.u32 %r120, %r610;

prmt.b32 %r120, %r120, 0, 0xba98;

mul.wide.u16 %r796, %rs33, 8;
shl.b32 %r797, %r790, 17;
or.b32 %r798, %r796, %r797;
and.b32 %r799, %r798, 25166208;
or.b32 %r800, %r799, -2139062144;
add.s32 %r801, %r800, -16843009;
or.b32 %r612, %r801, %r799;
mov.u32 %r122, %r612;

prmt.b32 %r122, %r122, 0, 0xba98;

xor.b32 %r802, %r120, %r781;
or.b32 %r803, %r802, -2139062144;
and.b32 %r804, %r120, 2139062143;
and.b32 %r805, %r782, -2139062144;
sub.s32 %r806, %r803, %r804;
xor.b32 %r125, %r805, %r806;
xor.b32 %r807, %r122, %r787;
or.b32 %r808, %r807, -2139062144;
and.b32 %r809, %r122, 2139062143;
and.b32 %r810, %r788, -2139062144;
sub.s32 %r811, %r808, %r809;
xor.b32 %r129, %r810, %r811;
ld.global.nc.u32 %r126, [%rd56+28];
ld.global.nc.u32 %r130, [%rd56+32];

dp4a.s32.s32 %r124, %r125, %r126, %r116;


dp4a.s32.s32 %r128, %r129, %r130, %r124;

add.s64 %rd69, %rd149, %rd13;
ld.global.nc.u8 %rs34, [%rd69];
cvt.u32.u16 %r812, %rs34;
and.b32 %r813, %r812, 255;
shr.u32 %r814, %r813, %r6;
shl.b32 %r815, %r814, 1;
and.b32 %r816, %r815, 30;
or.b32 %r817, %r816, 1;
mul.lo.s32 %r818, %r817, %r128;
add.s64 %rd70, %rd149, %rd10;
ld.global.nc.u16 %rs1, [%rd70];

{ cvt.f32.f16 %f73, %rs1;}


ld.global.nc.u32 %r132, [%rd56];

{.reg .f16 low,high;
mov.b32 {low,high},%r132;
cvt.f32.f16 %f74, low;}


mul.ftz.f32 %f97, %f73, %f74;
cvt.rn.f32.s32 %f98, %r818;
fma.rn.ftz.f32 %f276, %f97, %f98, %f276;
add.s64 %rd71, %rd149, %rd12;
ld.global.nc.u16 %rs35, [%rd71+2];
ld.global.nc.u16 %rs36, [%rd71+4];
ld.global.nc.u16 %rs37, [%rd71+6];
add.s64 %rd72, %rd149, %rd9;
ld.global.nc.u8 %rs38, [%rd72];
add.s64 %rd73, %rd149, %rd8;
add.s64 %rd74, %rd149, %rd7;
ld.global.nc.u16 %rs39, [%rd73];
ld.global.nc.u16 %rs40, [%rd74];
ld.global.nc.u16 %rs41, [%rd71];
and.b16 %rs42, %rs41, 255;
cvt.u32.u16 %r819, %rs42;
shl.b16 %rs43, %rs38, 8;
cvt.u32.u16 %r820, %rs43;
and.b32 %r821, %r820, 256;
or.b32 %r822, %r821, %r819;
mul.wide.u32 %rd75, %r822, 4;
add.s64 %rd76, %rd52, %rd75;
ld.global.nc.u32 %r823, [%rd76];
not.b32 %r824, %r823;
shr.u16 %rs44, %rs41, 8;
cvt.u32.u16 %r825, %rs44;
mul.wide.u16 %r826, %rs38, 128;
and.b32 %r827, %r826, 256;
or.b32 %r828, %r827, %r825;
mul.wide.u32 %rd77, %r828, 4;
add.s64 %rd78, %rd52, %rd77;
ld.global.nc.u32 %r829, [%rd78];
not.b32 %r830, %r829;
and.b16 %rs46, %rs39, 255;
mul.wide.u16 %r831, %rs46, 128;
cvt.u32.u16 %r832, %rs39;
shl.b32 %r833, %r832, 21;
or.b32 %r834, %r831, %r833;
and.b32 %r835, %r834, 25166208;
or.b32 %r836, %r835, -2139062144;
add.s32 %r837, %r836, -16843009;
or.b32 %r623, %r837, %r835;
mov.u32 %r133, %r623;

prmt.b32 %r133, %r133, 0, 0xba98;

mul.wide.u16 %r838, %rs46, 8;
shl.b32 %r839, %r832, 17;
or.b32 %r840, %r838, %r839;
and.b32 %r841, %r840, 25166208;
or.b32 %r842, %r841, -2139062144;
add.s32 %r843, %r842, -16843009;
or.b32 %r625, %r843, %r841;
mov.u32 %r135, %r625;

prmt.b32 %r135, %r135, 0, 0xba98;

xor.b32 %r844, %r133, %r823;
or.b32 %r845, %r844, -2139062144;
and.b32 %r846, %r133, 2139062143;
and.b32 %r847, %r824, -2139062144;
sub.s32 %r848, %r845, %r846;
xor.b32 %r138, %r847, %r848;
xor.b32 %r849, %r135, %r829;
or.b32 %r850, %r849, -2139062144;
and.b32 %r851, %r135, 2139062143;
and.b32 %r852, %r830, -2139062144;
sub.s32 %r853, %r850, %r851;
xor.b32 %r142, %r852, %r853;

dp4a.s32.s32 %r137, %r138, %r90, %r630;


dp4a.s32.s32 %r141, %r142, %r94, %r137;

and.b16 %rs47, %rs35, 255;
cvt.u32.u16 %r854, %rs47;
mul.wide.u16 %r855, %rs38, 64;
and.b32 %r856, %r855, 256;
or.b32 %r857, %r856, %r854;
mul.wide.u32 %rd79, %r857, 4;
add.s64 %rd80, %rd52, %rd79;
ld.global.nc.u32 %r858, [%rd80];
not.b32 %r859, %r858;
shr.u16 %rs48, %rs35, 8;
cvt.u32.u16 %r860, %rs48;
mul.wide.u16 %r861, %rs38, 32;
and.b32 %r862, %r861, 256;
or.b32 %r863, %r862, %r860;
mul.wide.u32 %rd81, %r863, 4;
add.s64 %rd82, %rd52, %rd81;
ld.global.nc.u32 %r864, [%rd82];
not.b32 %r865, %r864;
shr.u16 %rs49, %rs39, 8;
mul.wide.u16 %r866, %rs49, 128;
cvt.u32.u16 %r867, %rs49;
shl.b32 %r868, %r867, 21;
or.b32 %r869, %r866, %r868;
and.b32 %r870, %r869, 25166208;
or.b32 %r871, %r870, -2139062144;
add.s32 %r872, %r871, -16843009;
or.b32 %r635, %r872, %r870;
mov.u32 %r145, %r635;

prmt.b32 %r145, %r145, 0, 0xba98;

mul.wide.u16 %r873, %rs49, 8;
shl.b32 %r874, %r867, 17;
or.b32 %r875, %r873, %r874;
and.b32 %r876, %r875, 25166208;
or.b32 %r877, %r876, -2139062144;
add.s32 %r878, %r877, -16843009;
or.b32 %r637, %r878, %r876;
mov.u32 %r147, %r637;

prmt.b32 %r147, %r147, 0, 0xba98;

xor.b32 %r879, %r145, %r858;
or.b32 %r880, %r879, -2139062144;
and.b32 %r881, %r145, 2139062143;
and.b32 %r882, %r859, -2139062144;
sub.s32 %r883, %r880, %r881;
xor.b32 %r150, %r882, %r883;
xor.b32 %r884, %r147, %r864;
or.b32 %r885, %r884, -2139062144;
and.b32 %r886, %r147, 2139062143;
and.b32 %r887, %r865, -2139062144;
sub.s32 %r888, %r885, %r886;
xor.b32 %r154, %r887, %r888;

dp4a.s32.s32 %r149, %r150, %r102, %r141;


dp4a.s32.s32 %r153, %r154, %r106, %r149;

and.b16 %rs50, %rs36, 255;
cvt.u32.u16 %r889, %rs50;
mul.wide.u16 %r890, %rs38, 16;
and.b32 %r891, %r890, 256;
or.b32 %r892, %r891, %r889;
mul.wide.u32 %rd83, %r892, 4;
add.s64 %rd84, %rd52, %rd83;
ld.global.nc.u32 %r893, [%rd84];
not.b32 %r894, %r893;
shr.u16 %rs51, %rs36, 8;
cvt.u32.u16 %r895, %rs51;
mul.wide.u16 %r896, %rs38, 8;
and.b32 %r897, %r896, 256;
or.b32 %r898, %r897, %r895;
mul.wide.u32 %rd85, %r898, 4;
add.s64 %rd86, %rd52, %rd85;
ld.global.nc.u32 %r899, [%rd86];
not.b32 %r900, %r899;
and.b16 %rs52, %rs40, 255;
mul.wide.u16 %r901, %rs52, 128;
cvt.u32.u16 %r902, %rs40;
shl.b32 %r903, %r902, 21;
or.b32 %r904, %r901, %r903;
and.b32 %r905, %r904, 25166208;
or.b32 %r906, %r905, -2139062144;
add.s32 %r907, %r906, -16843009;
or.b32 %r647, %r907, %r905;
mov.u32 %r157, %r647;

prmt.b32 %r157, %r157, 0, 0xba98;

mul.wide.u16 %r908, %rs52, 8;
shl.b32 %r909, %r902, 17;
or.b32 %r910, %r908, %r909;
and.b32 %r911, %r910, 25166208;
or.b32 %r912, %r911, -2139062144;
add.s32 %r913, %r912, -16843009;
or.b32 %r649, %r913, %r911;
mov.u32 %r159, %r649;

prmt.b32 %r159, %r159, 0, 0xba98;

xor.b32 %r914, %r157, %r893;
or.b32 %r915, %r914, -2139062144;
and.b32 %r916, %r157, 2139062143;
and.b32 %r917, %r894, -2139062144;
sub.s32 %r918, %r915, %r916;
xor.b32 %r162, %r917, %r918;
xor.b32 %r919, %r159, %r899;
or.b32 %r920, %r919, -2139062144;
and.b32 %r921, %r159, 2139062143;
and.b32 %r922, %r900, -2139062144;
sub.s32 %r923, %r920, %r921;
xor.b32 %r166, %r922, %r923;

dp4a.s32.s32 %r161, %r162, %r114, %r153;


dp4a.s32.s32 %r165, %r166, %r118, %r161;

and.b16 %rs53, %rs37, 255;
cvt.u32.u16 %r924, %rs53;
mul.wide.u16 %r925, %rs38, 4;
and.b32 %r926, %r925, 256;
or.b32 %r927, %r926, %r924;
mul.wide.u32 %rd87, %r927, 4;
add.s64 %rd88, %rd52, %rd87;
ld.global.nc.u32 %r928, [%rd88];
not.b32 %r929, %r928;
shr.u16 %rs54, %rs37, 8;
cvt.u32.u16 %r930, %rs54;
mul.wide.u16 %r931, %rs38, 2;
and.b32 %r932, %r931, 256;
or.b32 %r933, %r932, %r930;
mul.wide.u32 %rd89, %r933, 4;
add.s64 %rd90, %rd52, %rd89;
ld.global.nc.u32 %r934, [%rd90];
not.b32 %r935, %r934;
shr.u16 %rs55, %rs40, 8;
mul.wide.u16 %r936, %rs55, 128;
cvt.u32.u16 %r937, %rs55;
shl.b32 %r938, %r937, 21;
or.b32 %r939, %r936, %r938;
and.b32 %r940, %r939, 25166208;
or.b32 %r941, %r940, -2139062144;
add.s32 %r942, %r941, -16843009;
or.b32 %r659, %r942, %r940;
mov.u32 %r169, %r659;

prmt.b32 %r169, %r169, 0, 0xba98;

mul.wide.u16 %r943, %rs55, 8;
shl.b32 %r944, %r937, 17;
or.b32 %r945, %r943, %r944;
and.b32 %r946, %r945, 25166208;
or.b32 %r947, %r946, -2139062144;
add.s32 %r948, %r947, -16843009;
or.b32 %r661, %r948, %r946;
mov.u32 %r171, %r661;

prmt.b32 %r171, %r171, 0, 0xba98;

xor.b32 %r949, %r169, %r928;
or.b32 %r950, %r949, -2139062144;
and.b32 %r951, %r169, 2139062143;
and.b32 %r952, %r929, -2139062144;
sub.s32 %r953, %r950, %r951;
xor.b32 %r174, %r952, %r953;
xor.b32 %r954, %r171, %r934;
or.b32 %r955, %r954, -2139062144;
and.b32 %r956, %r171, 2139062143;
and.b32 %r957, %r935, -2139062144;
sub.s32 %r958, %r955, %r956;
xor.b32 %r178, %r957, %r958;

dp4a.s32.s32 %r173, %r174, %r126, %r165;


dp4a.s32.s32 %r177, %r178, %r130, %r173;

add.s64 %rd91, %rd149, %rd6;
ld.global.nc.u8 %rs56, [%rd91];
cvt.u32.u16 %r959, %rs56;
and.b32 %r960, %r959, 255;
shr.u32 %r961, %r960, %r6;
shl.b32 %r962, %r961, 1;
and.b32 %r963, %r962, 30;
or.b32 %r964, %r963, 1;
mul.lo.s32 %r965, %r964, %r177;
add.s64 %rd92, %rd149, %rd5;
ld.global.nc.u16 %rs2, [%rd92];

{ cvt.f32.f16 %f75, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r132;
cvt.f32.f16 %f76, low;}


mul.ftz.f32 %f99, %f75, %f76;
cvt.rn.f32.s32 %f100, %r965;
fma.rn.ftz.f32 %f275, %f99, %f100, %f275;
cvt.u64.u32 %rd93, %r1520;
add.s64 %rd94, %rd2, %rd93;
add.s64 %rd95, %rd94, %rd4;
mul.lo.s64 %rd96, %rd95, 36;
mov.u32 %r182, %r574;

prmt.b32 %r182, %r182, 0, 0xba98;

mov.u32 %r184, %r576;

prmt.b32 %r184, %r184, 0, 0xba98;

xor.b32 %r966, %r182, %r676;
or.b32 %r967, %r966, -2139062144;
and.b32 %r968, %r182, 2139062143;
sub.s32 %r969, %r967, %r968;
xor.b32 %r187, %r700, %r969;
xor.b32 %r970, %r184, %r682;
or.b32 %r971, %r970, -2139062144;
and.b32 %r972, %r184, 2139062143;
sub.s32 %r973, %r971, %r972;
xor.b32 %r191, %r705, %r973;
add.s64 %rd97, %rd17, %rd96;
ld.global.nc.u32 %r188, [%rd97+4];
ld.global.nc.u32 %r192, [%rd97+8];

dp4a.s32.s32 %r186, %r187, %r188, %r630;


dp4a.s32.s32 %r190, %r191, %r192, %r186;

mov.u32 %r194, %r586;

prmt.b32 %r194, %r194, 0, 0xba98;

mov.u32 %r196, %r588;

prmt.b32 %r196, %r196, 0, 0xba98;

xor.b32 %r974, %r194, %r711;
or.b32 %r975, %r974, -2139062144;
and.b32 %r976, %r194, 2139062143;
sub.s32 %r977, %r975, %r976;
xor.b32 %r199, %r735, %r977;
xor.b32 %r978, %r196, %r717;
or.b32 %r979, %r978, -2139062144;
and.b32 %r980, %r196, 2139062143;
sub.s32 %r981, %r979, %r980;
xor.b32 %r203, %r740, %r981;
ld.global.nc.u32 %r200, [%rd97+12];
ld.global.nc.u32 %r204, [%rd97+16];

dp4a.s32.s32 %r198, %r199, %r200, %r190;


dp4a.s32.s32 %r202, %r203, %r204, %r198;

mov.u32 %r206, %r598;

prmt.b32 %r206, %r206, 0, 0xba98;

mov.u32 %r208, %r600;

prmt.b32 %r208, %r208, 0, 0xba98;

xor.b32 %r982, %r206, %r746;
or.b32 %r983, %r982, -2139062144;
and.b32 %r984, %r206, 2139062143;
sub.s32 %r985, %r983, %r984;
xor.b32 %r211, %r770, %r985;
xor.b32 %r986, %r208, %r752;
or.b32 %r987, %r986, -2139062144;
and.b32 %r988, %r208, 2139062143;
sub.s32 %r989, %r987, %r988;
xor.b32 %r215, %r775, %r989;
ld.global.nc.u32 %r212, [%rd97+20];
ld.global.nc.u32 %r216, [%rd97+24];

dp4a.s32.s32 %r210, %r211, %r212, %r202;


dp4a.s32.s32 %r214, %r215, %r216, %r210;

mov.u32 %r218, %r610;

prmt.b32 %r218, %r218, 0, 0xba98;

mov.u32 %r220, %r612;

prmt.b32 %r220, %r220, 0, 0xba98;

xor.b32 %r990, %r218, %r781;
or.b32 %r991, %r990, -2139062144;
and.b32 %r992, %r218, 2139062143;
sub.s32 %r993, %r991, %r992;
xor.b32 %r223, %r805, %r993;
xor.b32 %r994, %r220, %r787;
or.b32 %r995, %r994, -2139062144;
and.b32 %r996, %r220, 2139062143;
sub.s32 %r997, %r995, %r996;
xor.b32 %r227, %r810, %r997;
ld.global.nc.u32 %r224, [%rd97+28];
ld.global.nc.u32 %r228, [%rd97+32];

dp4a.s32.s32 %r222, %r223, %r224, %r214;


dp4a.s32.s32 %r226, %r227, %r228, %r222;

mul.lo.s32 %r998, %r817, %r226;

{ cvt.f32.f16 %f77, %rs1;}


ld.global.nc.u32 %r230, [%rd97];

{.reg .f16 low,high;
mov.b32 {low,high},%r230;
cvt.f32.f16 %f78, low;}


mul.ftz.f32 %f101, %f77, %f78;
cvt.rn.f32.s32 %f102, %r998;
fma.rn.ftz.f32 %f274, %f101, %f102, %f274;
mov.u32 %r231, %r623;

prmt.b32 %r231, %r231, 0, 0xba98;

mov.u32 %r233, %r625;

prmt.b32 %r233, %r233, 0, 0xba98;

xor.b32 %r999, %r231, %r823;
or.b32 %r1000, %r999, -2139062144;
and.b32 %r1001, %r231, 2139062143;
sub.s32 %r1002, %r1000, %r1001;
xor.b32 %r236, %r847, %r1002;
xor.b32 %r1003, %r233, %r829;
or.b32 %r1004, %r1003, -2139062144;
and.b32 %r1005, %r233, 2139062143;
sub.s32 %r1006, %r1004, %r1005;
xor.b32 %r240, %r852, %r1006;

dp4a.s32.s32 %r235, %r236, %r188, %r630;


dp4a.s32.s32 %r239, %r240, %r192, %r235;

mov.u32 %r243, %r635;

prmt.b32 %r243, %r243, 0, 0xba98;

mov.u32 %r245, %r637;

prmt.b32 %r245, %r245, 0, 0xba98;

xor.b32 %r1007, %r243, %r858;
or.b32 %r1008, %r1007, -2139062144;
and.b32 %r1009, %r243, 2139062143;
sub.s32 %r1010, %r1008, %r1009;
xor.b32 %r248, %r882, %r1010;
xor.b32 %r1011, %r245, %r864;
or.b32 %r1012, %r1011, -2139062144;
and.b32 %r1013, %r245, 2139062143;
sub.s32 %r1014, %r1012, %r1013;
xor.b32 %r252, %r887, %r1014;

dp4a.s32.s32 %r247, %r248, %r200, %r239;


dp4a.s32.s32 %r251, %r252, %r204, %r247;

mov.u32 %r255, %r647;

prmt.b32 %r255, %r255, 0, 0xba98;

mov.u32 %r257, %r649;

prmt.b32 %r257, %r257, 0, 0xba98;

xor.b32 %r1015, %r255, %r893;
or.b32 %r1016, %r1015, -2139062144;
and.b32 %r1017, %r255, 2139062143;
sub.s32 %r1018, %r1016, %r1017;
xor.b32 %r260, %r917, %r1018;
xor.b32 %r1019, %r257, %r899;
or.b32 %r1020, %r1019, -2139062144;
and.b32 %r1021, %r257, 2139062143;
sub.s32 %r1022, %r1020, %r1021;
xor.b32 %r264, %r922, %r1022;

dp4a.s32.s32 %r259, %r260, %r212, %r251;


dp4a.s32.s32 %r263, %r264, %r216, %r259;

mov.u32 %r267, %r659;

prmt.b32 %r267, %r267, 0, 0xba98;

mov.u32 %r269, %r661;

prmt.b32 %r269, %r269, 0, 0xba98;

xor.b32 %r1023, %r267, %r928;
or.b32 %r1024, %r1023, -2139062144;
and.b32 %r1025, %r267, 2139062143;
sub.s32 %r1026, %r1024, %r1025;
xor.b32 %r272, %r952, %r1026;
xor.b32 %r1027, %r269, %r934;
or.b32 %r1028, %r1027, -2139062144;
and.b32 %r1029, %r269, 2139062143;
sub.s32 %r1030, %r1028, %r1029;
xor.b32 %r276, %r957, %r1030;

dp4a.s32.s32 %r271, %r272, %r224, %r263;


dp4a.s32.s32 %r275, %r276, %r228, %r271;

mul.lo.s32 %r1031, %r964, %r275;

{ cvt.f32.f16 %f79, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r230;
cvt.f32.f16 %f80, low;}


mul.ftz.f32 %f103, %f79, %f80;
cvt.rn.f32.s32 %f104, %r1031;
fma.rn.ftz.f32 %f273, %f103, %f104, %f273;
cvt.u64.u32 %rd98, %r1524;
add.s64 %rd99, %rd2, %rd98;
add.s64 %rd100, %rd99, %rd4;
mul.lo.s64 %rd101, %rd100, 36;
mov.u32 %r280, %r574;

prmt.b32 %r280, %r280, 0, 0xba98;

mov.u32 %r282, %r576;

prmt.b32 %r282, %r282, 0, 0xba98;

xor.b32 %r1032, %r280, %r676;
or.b32 %r1033, %r1032, -2139062144;
and.b32 %r1034, %r280, 2139062143;
sub.s32 %r1035, %r1033, %r1034;
xor.b32 %r285, %r700, %r1035;
xor.b32 %r1036, %r282, %r682;
or.b32 %r1037, %r1036, -2139062144;
and.b32 %r1038, %r282, 2139062143;
sub.s32 %r1039, %r1037, %r1038;
xor.b32 %r289, %r705, %r1039;
add.s64 %rd102, %rd17, %rd101;
ld.global.nc.u32 %r286, [%rd102+4];
ld.global.nc.u32 %r290, [%rd102+8];

dp4a.s32.s32 %r284, %r285, %r286, %r630;


dp4a.s32.s32 %r288, %r289, %r290, %r284;

mov.u32 %r292, %r586;

prmt.b32 %r292, %r292, 0, 0xba98;

mov.u32 %r294, %r588;

prmt.b32 %r294, %r294, 0, 0xba98;

xor.b32 %r1040, %r292, %r711;
or.b32 %r1041, %r1040, -2139062144;
and.b32 %r1042, %r292, 2139062143;
sub.s32 %r1043, %r1041, %r1042;
xor.b32 %r297, %r735, %r1043;
xor.b32 %r1044, %r294, %r717;
or.b32 %r1045, %r1044, -2139062144;
and.b32 %r1046, %r294, 2139062143;
sub.s32 %r1047, %r1045, %r1046;
xor.b32 %r301, %r740, %r1047;
ld.global.nc.u32 %r298, [%rd102+12];
ld.global.nc.u32 %r302, [%rd102+16];

dp4a.s32.s32 %r296, %r297, %r298, %r288;


dp4a.s32.s32 %r300, %r301, %r302, %r296;

mov.u32 %r304, %r598;

prmt.b32 %r304, %r304, 0, 0xba98;

mov.u32 %r306, %r600;

prmt.b32 %r306, %r306, 0, 0xba98;

xor.b32 %r1048, %r304, %r746;
or.b32 %r1049, %r1048, -2139062144;
and.b32 %r1050, %r304, 2139062143;
sub.s32 %r1051, %r1049, %r1050;
xor.b32 %r309, %r770, %r1051;
xor.b32 %r1052, %r306, %r752;
or.b32 %r1053, %r1052, -2139062144;
and.b32 %r1054, %r306, 2139062143;
sub.s32 %r1055, %r1053, %r1054;
xor.b32 %r313, %r775, %r1055;
ld.global.nc.u32 %r310, [%rd102+20];
ld.global.nc.u32 %r314, [%rd102+24];

dp4a.s32.s32 %r308, %r309, %r310, %r300;


dp4a.s32.s32 %r312, %r313, %r314, %r308;

mov.u32 %r316, %r610;

prmt.b32 %r316, %r316, 0, 0xba98;

mov.u32 %r318, %r612;

prmt.b32 %r318, %r318, 0, 0xba98;

xor.b32 %r1056, %r316, %r781;
or.b32 %r1057, %r1056, -2139062144;
and.b32 %r1058, %r316, 2139062143;
sub.s32 %r1059, %r1057, %r1058;
xor.b32 %r321, %r805, %r1059;
xor.b32 %r1060, %r318, %r787;
or.b32 %r1061, %r1060, -2139062144;
and.b32 %r1062, %r318, 2139062143;
sub.s32 %r1063, %r1061, %r1062;
xor.b32 %r325, %r810, %r1063;
ld.global.nc.u32 %r322, [%rd102+28];
ld.global.nc.u32 %r326, [%rd102+32];

dp4a.s32.s32 %r320, %r321, %r322, %r312;


dp4a.s32.s32 %r324, %r325, %r326, %r320;

mul.lo.s32 %r1064, %r817, %r324;

{ cvt.f32.f16 %f81, %rs1;}


ld.global.nc.u32 %r328, [%rd102];

{.reg .f16 low,high;
mov.b32 {low,high},%r328;
cvt.f32.f16 %f82, low;}


mul.ftz.f32 %f105, %f81, %f82;
cvt.rn.f32.s32 %f106, %r1064;
fma.rn.ftz.f32 %f272, %f105, %f106, %f272;
mov.u32 %r329, %r623;

prmt.b32 %r329, %r329, 0, 0xba98;

mov.u32 %r331, %r625;

prmt.b32 %r331, %r331, 0, 0xba98;

xor.b32 %r1065, %r329, %r823;
or.b32 %r1066, %r1065, -2139062144;
and.b32 %r1067, %r329, 2139062143;
sub.s32 %r1068, %r1066, %r1067;
xor.b32 %r334, %r847, %r1068;
xor.b32 %r1069, %r331, %r829;
or.b32 %r1070, %r1069, -2139062144;
and.b32 %r1071, %r331, 2139062143;
sub.s32 %r1072, %r1070, %r1071;
xor.b32 %r338, %r852, %r1072;

dp4a.s32.s32 %r333, %r334, %r286, %r630;


dp4a.s32.s32 %r337, %r338, %r290, %r333;

mov.u32 %r341, %r635;

prmt.b32 %r341, %r341, 0, 0xba98;

mov.u32 %r343, %r637;

prmt.b32 %r343, %r343, 0, 0xba98;

xor.b32 %r1073, %r341, %r858;
or.b32 %r1074, %r1073, -2139062144;
and.b32 %r1075, %r341, 2139062143;
sub.s32 %r1076, %r1074, %r1075;
xor.b32 %r346, %r882, %r1076;
xor.b32 %r1077, %r343, %r864;
or.b32 %r1078, %r1077, -2139062144;
and.b32 %r1079, %r343, 2139062143;
sub.s32 %r1080, %r1078, %r1079;
xor.b32 %r350, %r887, %r1080;

dp4a.s32.s32 %r345, %r346, %r298, %r337;


dp4a.s32.s32 %r349, %r350, %r302, %r345;

mov.u32 %r353, %r647;

prmt.b32 %r353, %r353, 0, 0xba98;

mov.u32 %r355, %r649;

prmt.b32 %r355, %r355, 0, 0xba98;

xor.b32 %r1081, %r353, %r893;
or.b32 %r1082, %r1081, -2139062144;
and.b32 %r1083, %r353, 2139062143;
sub.s32 %r1084, %r1082, %r1083;
xor.b32 %r358, %r917, %r1084;
xor.b32 %r1085, %r355, %r899;
or.b32 %r1086, %r1085, -2139062144;
and.b32 %r1087, %r355, 2139062143;
sub.s32 %r1088, %r1086, %r1087;
xor.b32 %r362, %r922, %r1088;

dp4a.s32.s32 %r357, %r358, %r310, %r349;


dp4a.s32.s32 %r361, %r362, %r314, %r357;

mov.u32 %r365, %r659;

prmt.b32 %r365, %r365, 0, 0xba98;

mov.u32 %r367, %r661;

prmt.b32 %r367, %r367, 0, 0xba98;

xor.b32 %r1089, %r365, %r928;
or.b32 %r1090, %r1089, -2139062144;
and.b32 %r1091, %r365, 2139062143;
sub.s32 %r1092, %r1090, %r1091;
xor.b32 %r370, %r952, %r1092;
xor.b32 %r1093, %r367, %r934;
or.b32 %r1094, %r1093, -2139062144;
and.b32 %r1095, %r367, 2139062143;
sub.s32 %r1096, %r1094, %r1095;
xor.b32 %r374, %r957, %r1096;

dp4a.s32.s32 %r369, %r370, %r322, %r361;


dp4a.s32.s32 %r373, %r374, %r326, %r369;

mul.lo.s32 %r1097, %r964, %r373;

{ cvt.f32.f16 %f83, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r328;
cvt.f32.f16 %f84, low;}


mul.ftz.f32 %f107, %f83, %f84;
cvt.rn.f32.s32 %f108, %r1097;
fma.rn.ftz.f32 %f271, %f107, %f108, %f271;
cvt.u64.u32 %rd103, %r1523;
add.s64 %rd104, %rd2, %rd103;
add.s64 %rd105, %rd104, %rd4;
mul.lo.s64 %rd106, %rd105, 36;
mov.u32 %r378, %r574;

prmt.b32 %r378, %r378, 0, 0xba98;

mov.u32 %r380, %r576;

prmt.b32 %r380, %r380, 0, 0xba98;

xor.b32 %r1098, %r378, %r676;
or.b32 %r1099, %r1098, -2139062144;
and.b32 %r1100, %r378, 2139062143;
sub.s32 %r1101, %r1099, %r1100;
xor.b32 %r383, %r700, %r1101;
xor.b32 %r1102, %r380, %r682;
or.b32 %r1103, %r1102, -2139062144;
and.b32 %r1104, %r380, 2139062143;
sub.s32 %r1105, %r1103, %r1104;
xor.b32 %r387, %r705, %r1105;
add.s64 %rd107, %rd17, %rd106;
ld.global.nc.u32 %r384, [%rd107+4];
ld.global.nc.u32 %r388, [%rd107+8];

dp4a.s32.s32 %r382, %r383, %r384, %r630;


dp4a.s32.s32 %r386, %r387, %r388, %r382;

mov.u32 %r390, %r586;

prmt.b32 %r390, %r390, 0, 0xba98;

mov.u32 %r392, %r588;

prmt.b32 %r392, %r392, 0, 0xba98;

xor.b32 %r1106, %r390, %r711;
or.b32 %r1107, %r1106, -2139062144;
and.b32 %r1108, %r390, 2139062143;
sub.s32 %r1109, %r1107, %r1108;
xor.b32 %r395, %r735, %r1109;
xor.b32 %r1110, %r392, %r717;
or.b32 %r1111, %r1110, -2139062144;
and.b32 %r1112, %r392, 2139062143;
sub.s32 %r1113, %r1111, %r1112;
xor.b32 %r399, %r740, %r1113;
ld.global.nc.u32 %r396, [%rd107+12];
ld.global.nc.u32 %r400, [%rd107+16];

dp4a.s32.s32 %r394, %r395, %r396, %r386;


dp4a.s32.s32 %r398, %r399, %r400, %r394;

mov.u32 %r402, %r598;

prmt.b32 %r402, %r402, 0, 0xba98;

mov.u32 %r404, %r600;

prmt.b32 %r404, %r404, 0, 0xba98;

xor.b32 %r1114, %r402, %r746;
or.b32 %r1115, %r1114, -2139062144;
and.b32 %r1116, %r402, 2139062143;
sub.s32 %r1117, %r1115, %r1116;
xor.b32 %r407, %r770, %r1117;
xor.b32 %r1118, %r404, %r752;
or.b32 %r1119, %r1118, -2139062144;
and.b32 %r1120, %r404, 2139062143;
sub.s32 %r1121, %r1119, %r1120;
xor.b32 %r411, %r775, %r1121;
ld.global.nc.u32 %r408, [%rd107+20];
ld.global.nc.u32 %r412, [%rd107+24];

dp4a.s32.s32 %r406, %r407, %r408, %r398;


dp4a.s32.s32 %r410, %r411, %r412, %r406;

mov.u32 %r414, %r610;

prmt.b32 %r414, %r414, 0, 0xba98;

mov.u32 %r416, %r612;

prmt.b32 %r416, %r416, 0, 0xba98;

xor.b32 %r1122, %r414, %r781;
or.b32 %r1123, %r1122, -2139062144;
and.b32 %r1124, %r414, 2139062143;
sub.s32 %r1125, %r1123, %r1124;
xor.b32 %r419, %r805, %r1125;
xor.b32 %r1126, %r416, %r787;
or.b32 %r1127, %r1126, -2139062144;
and.b32 %r1128, %r416, 2139062143;
sub.s32 %r1129, %r1127, %r1128;
xor.b32 %r423, %r810, %r1129;
ld.global.nc.u32 %r420, [%rd107+28];
ld.global.nc.u32 %r424, [%rd107+32];

dp4a.s32.s32 %r418, %r419, %r420, %r410;


dp4a.s32.s32 %r422, %r423, %r424, %r418;

mul.lo.s32 %r1130, %r817, %r422;

{ cvt.f32.f16 %f85, %rs1;}


ld.global.nc.u32 %r426, [%rd107];

{.reg .f16 low,high;
mov.b32 {low,high},%r426;
cvt.f32.f16 %f86, low;}


mul.ftz.f32 %f109, %f85, %f86;
cvt.rn.f32.s32 %f110, %r1130;
fma.rn.ftz.f32 %f270, %f109, %f110, %f270;
mov.u32 %r427, %r623;

prmt.b32 %r427, %r427, 0, 0xba98;

mov.u32 %r429, %r625;

prmt.b32 %r429, %r429, 0, 0xba98;

xor.b32 %r1131, %r427, %r823;
or.b32 %r1132, %r1131, -2139062144;
and.b32 %r1133, %r427, 2139062143;
sub.s32 %r1134, %r1132, %r1133;
xor.b32 %r432, %r847, %r1134;
xor.b32 %r1135, %r429, %r829;
or.b32 %r1136, %r1135, -2139062144;
and.b32 %r1137, %r429, 2139062143;
sub.s32 %r1138, %r1136, %r1137;
xor.b32 %r436, %r852, %r1138;

dp4a.s32.s32 %r431, %r432, %r384, %r630;


dp4a.s32.s32 %r435, %r436, %r388, %r431;

mov.u32 %r439, %r635;

prmt.b32 %r439, %r439, 0, 0xba98;

mov.u32 %r441, %r637;

prmt.b32 %r441, %r441, 0, 0xba98;

xor.b32 %r1139, %r439, %r858;
or.b32 %r1140, %r1139, -2139062144;
and.b32 %r1141, %r439, 2139062143;
sub.s32 %r1142, %r1140, %r1141;
xor.b32 %r444, %r882, %r1142;
xor.b32 %r1143, %r441, %r864;
or.b32 %r1144, %r1143, -2139062144;
and.b32 %r1145, %r441, 2139062143;
sub.s32 %r1146, %r1144, %r1145;
xor.b32 %r448, %r887, %r1146;

dp4a.s32.s32 %r443, %r444, %r396, %r435;


dp4a.s32.s32 %r447, %r448, %r400, %r443;

mov.u32 %r451, %r647;

prmt.b32 %r451, %r451, 0, 0xba98;

mov.u32 %r453, %r649;

prmt.b32 %r453, %r453, 0, 0xba98;

xor.b32 %r1147, %r451, %r893;
or.b32 %r1148, %r1147, -2139062144;
and.b32 %r1149, %r451, 2139062143;
sub.s32 %r1150, %r1148, %r1149;
xor.b32 %r456, %r917, %r1150;
xor.b32 %r1151, %r453, %r899;
or.b32 %r1152, %r1151, -2139062144;
and.b32 %r1153, %r453, 2139062143;
sub.s32 %r1154, %r1152, %r1153;
xor.b32 %r460, %r922, %r1154;

dp4a.s32.s32 %r455, %r456, %r408, %r447;


dp4a.s32.s32 %r459, %r460, %r412, %r455;

mov.u32 %r463, %r659;

prmt.b32 %r463, %r463, 0, 0xba98;

mov.u32 %r465, %r661;

prmt.b32 %r465, %r465, 0, 0xba98;

xor.b32 %r1155, %r463, %r928;
or.b32 %r1156, %r1155, -2139062144;
and.b32 %r1157, %r463, 2139062143;
sub.s32 %r1158, %r1156, %r1157;
xor.b32 %r468, %r952, %r1158;
xor.b32 %r1159, %r465, %r934;
or.b32 %r1160, %r1159, -2139062144;
and.b32 %r1161, %r465, 2139062143;
sub.s32 %r1162, %r1160, %r1161;
xor.b32 %r472, %r957, %r1162;

dp4a.s32.s32 %r467, %r468, %r420, %r459;


dp4a.s32.s32 %r471, %r472, %r424, %r467;

mul.lo.s32 %r1163, %r964, %r471;

{ cvt.f32.f16 %f87, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r426;
cvt.f32.f16 %f88, low;}


mul.ftz.f32 %f111, %f87, %f88;
cvt.rn.f32.s32 %f112, %r1163;
fma.rn.ftz.f32 %f269, %f111, %f112, %f269;
cvt.u64.u32 %rd108, %r1522;
add.s64 %rd109, %rd2, %rd108;
add.s64 %rd110, %rd109, %rd4;
mul.lo.s64 %rd111, %rd110, 36;
mov.u32 %r476, %r574;

prmt.b32 %r476, %r476, 0, 0xba98;

mov.u32 %r478, %r576;

prmt.b32 %r478, %r478, 0, 0xba98;

xor.b32 %r1164, %r476, %r676;
or.b32 %r1165, %r1164, -2139062144;
and.b32 %r1166, %r476, 2139062143;
sub.s32 %r1167, %r1165, %r1166;
xor.b32 %r481, %r700, %r1167;
xor.b32 %r1168, %r478, %r682;
or.b32 %r1169, %r1168, -2139062144;
and.b32 %r1170, %r478, 2139062143;
sub.s32 %r1171, %r1169, %r1170;
xor.b32 %r485, %r705, %r1171;
add.s64 %rd112, %rd17, %rd111;
ld.global.nc.u32 %r482, [%rd112+4];
ld.global.nc.u32 %r486, [%rd112+8];

dp4a.s32.s32 %r480, %r481, %r482, %r630;


dp4a.s32.s32 %r484, %r485, %r486, %r480;

mov.u32 %r488, %r586;

prmt.b32 %r488, %r488, 0, 0xba98;

mov.u32 %r490, %r588;

prmt.b32 %r490, %r490, 0, 0xba98;

xor.b32 %r1172, %r488, %r711;
or.b32 %r1173, %r1172, -2139062144;
and.b32 %r1174, %r488, 2139062143;
sub.s32 %r1175, %r1173, %r1174;
xor.b32 %r493, %r735, %r1175;
xor.b32 %r1176, %r490, %r717;
or.b32 %r1177, %r1176, -2139062144;
and.b32 %r1178, %r490, 2139062143;
sub.s32 %r1179, %r1177, %r1178;
xor.b32 %r497, %r740, %r1179;
ld.global.nc.u32 %r494, [%rd112+12];
ld.global.nc.u32 %r498, [%rd112+16];

dp4a.s32.s32 %r492, %r493, %r494, %r484;


dp4a.s32.s32 %r496, %r497, %r498, %r492;

mov.u32 %r500, %r598;

prmt.b32 %r500, %r500, 0, 0xba98;

mov.u32 %r502, %r600;

prmt.b32 %r502, %r502, 0, 0xba98;

xor.b32 %r1180, %r500, %r746;
or.b32 %r1181, %r1180, -2139062144;
and.b32 %r1182, %r500, 2139062143;
sub.s32 %r1183, %r1181, %r1182;
xor.b32 %r505, %r770, %r1183;
xor.b32 %r1184, %r502, %r752;
or.b32 %r1185, %r1184, -2139062144;
and.b32 %r1186, %r502, 2139062143;
sub.s32 %r1187, %r1185, %r1186;
xor.b32 %r509, %r775, %r1187;
ld.global.nc.u32 %r506, [%rd112+20];
ld.global.nc.u32 %r510, [%rd112+24];

dp4a.s32.s32 %r504, %r505, %r506, %r496;


dp4a.s32.s32 %r508, %r509, %r510, %r504;

mov.u32 %r512, %r610;

prmt.b32 %r512, %r512, 0, 0xba98;

mov.u32 %r514, %r612;

prmt.b32 %r514, %r514, 0, 0xba98;

xor.b32 %r1188, %r512, %r781;
or.b32 %r1189, %r1188, -2139062144;
and.b32 %r1190, %r512, 2139062143;
sub.s32 %r1191, %r1189, %r1190;
xor.b32 %r517, %r805, %r1191;
xor.b32 %r1192, %r514, %r787;
or.b32 %r1193, %r1192, -2139062144;
and.b32 %r1194, %r514, 2139062143;
sub.s32 %r1195, %r1193, %r1194;
xor.b32 %r521, %r810, %r1195;
ld.global.nc.u32 %r518, [%rd112+28];
ld.global.nc.u32 %r522, [%rd112+32];

dp4a.s32.s32 %r516, %r517, %r518, %r508;


dp4a.s32.s32 %r520, %r521, %r522, %r516;

mul.lo.s32 %r1196, %r817, %r520;

{ cvt.f32.f16 %f89, %rs1;}


ld.global.nc.u32 %r524, [%rd112];

{.reg .f16 low,high;
mov.b32 {low,high},%r524;
cvt.f32.f16 %f90, low;}


mul.ftz.f32 %f113, %f89, %f90;
cvt.rn.f32.s32 %f114, %r1196;
fma.rn.ftz.f32 %f268, %f113, %f114, %f268;
mov.u32 %r525, %r623;

prmt.b32 %r525, %r525, 0, 0xba98;

mov.u32 %r527, %r625;

prmt.b32 %r527, %r527, 0, 0xba98;

xor.b32 %r1197, %r525, %r823;
or.b32 %r1198, %r1197, -2139062144;
and.b32 %r1199, %r525, 2139062143;
sub.s32 %r1200, %r1198, %r1199;
xor.b32 %r530, %r847, %r1200;
xor.b32 %r1201, %r527, %r829;
or.b32 %r1202, %r1201, -2139062144;
and.b32 %r1203, %r527, 2139062143;
sub.s32 %r1204, %r1202, %r1203;
xor.b32 %r534, %r852, %r1204;

dp4a.s32.s32 %r529, %r530, %r482, %r630;


dp4a.s32.s32 %r533, %r534, %r486, %r529;

mov.u32 %r537, %r635;

prmt.b32 %r537, %r537, 0, 0xba98;

mov.u32 %r539, %r637;

prmt.b32 %r539, %r539, 0, 0xba98;

xor.b32 %r1205, %r537, %r858;
or.b32 %r1206, %r1205, -2139062144;
and.b32 %r1207, %r537, 2139062143;
sub.s32 %r1208, %r1206, %r1207;
xor.b32 %r542, %r882, %r1208;
xor.b32 %r1209, %r539, %r864;
or.b32 %r1210, %r1209, -2139062144;
and.b32 %r1211, %r539, 2139062143;
sub.s32 %r1212, %r1210, %r1211;
xor.b32 %r546, %r887, %r1212;

dp4a.s32.s32 %r541, %r542, %r494, %r533;


dp4a.s32.s32 %r545, %r546, %r498, %r541;

mov.u32 %r549, %r647;

prmt.b32 %r549, %r549, 0, 0xba98;

mov.u32 %r551, %r649;

prmt.b32 %r551, %r551, 0, 0xba98;

xor.b32 %r1213, %r549, %r893;
or.b32 %r1214, %r1213, -2139062144;
and.b32 %r1215, %r549, 2139062143;
sub.s32 %r1216, %r1214, %r1215;
xor.b32 %r554, %r917, %r1216;
xor.b32 %r1217, %r551, %r899;
or.b32 %r1218, %r1217, -2139062144;
and.b32 %r1219, %r551, 2139062143;
sub.s32 %r1220, %r1218, %r1219;
xor.b32 %r558, %r922, %r1220;

dp4a.s32.s32 %r553, %r554, %r506, %r545;


dp4a.s32.s32 %r557, %r558, %r510, %r553;

mov.u32 %r561, %r659;

prmt.b32 %r561, %r561, 0, 0xba98;

mov.u32 %r563, %r661;

prmt.b32 %r563, %r563, 0, 0xba98;

xor.b32 %r1221, %r561, %r928;
or.b32 %r1222, %r1221, -2139062144;
and.b32 %r1223, %r561, 2139062143;
sub.s32 %r1224, %r1222, %r1223;
xor.b32 %r566, %r952, %r1224;
xor.b32 %r1225, %r563, %r934;
or.b32 %r1226, %r1225, -2139062144;
and.b32 %r1227, %r563, 2139062143;
sub.s32 %r1228, %r1226, %r1227;
xor.b32 %r570, %r957, %r1228;

dp4a.s32.s32 %r565, %r566, %r518, %r557;


dp4a.s32.s32 %r569, %r570, %r522, %r565;

mul.lo.s32 %r1229, %r964, %r569;

{ cvt.f32.f16 %f91, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r524;
cvt.f32.f16 %f92, low;}


mul.ftz.f32 %f115, %f91, %f92;
cvt.rn.f32.s32 %f116, %r1229;
fma.rn.ftz.f32 %f267, %f115, %f116, %f267;
cvt.u64.u32 %rd113, %r1521;
add.s64 %rd114, %rd2, %rd113;
add.s64 %rd115, %rd114, %rd4;
mul.lo.s64 %rd116, %rd115, 36;

prmt.b32 %r574, %r574, 0, 0xba98;


prmt.b32 %r576, %r576, 0, 0xba98;

xor.b32 %r1230, %r574, %r676;
or.b32 %r1231, %r1230, -2139062144;
and.b32 %r1232, %r574, 2139062143;
sub.s32 %r1233, %r1231, %r1232;
xor.b32 %r579, %r700, %r1233;
xor.b32 %r1234, %r576, %r682;
or.b32 %r1235, %r1234, -2139062144;
and.b32 %r1236, %r576, 2139062143;
sub.s32 %r1237, %r1235, %r1236;
xor.b32 %r583, %r705, %r1237;
add.s64 %rd117, %rd17, %rd116;
ld.global.nc.u32 %r580, [%rd117+4];
ld.global.nc.u32 %r584, [%rd117+8];

dp4a.s32.s32 %r578, %r579, %r580, %r630;


dp4a.s32.s32 %r582, %r583, %r584, %r578;


prmt.b32 %r586, %r586, 0, 0xba98;


prmt.b32 %r588, %r588, 0, 0xba98;

xor.b32 %r1238, %r586, %r711;
or.b32 %r1239, %r1238, -2139062144;
and.b32 %r1240, %r586, 2139062143;
sub.s32 %r1241, %r1239, %r1240;
xor.b32 %r591, %r735, %r1241;
xor.b32 %r1242, %r588, %r717;
or.b32 %r1243, %r1242, -2139062144;
and.b32 %r1244, %r588, 2139062143;
sub.s32 %r1245, %r1243, %r1244;
xor.b32 %r595, %r740, %r1245;
ld.global.nc.u32 %r592, [%rd117+12];
ld.global.nc.u32 %r596, [%rd117+16];

dp4a.s32.s32 %r590, %r591, %r592, %r582;


dp4a.s32.s32 %r594, %r595, %r596, %r590;


prmt.b32 %r598, %r598, 0, 0xba98;


prmt.b32 %r600, %r600, 0, 0xba98;

xor.b32 %r1246, %r598, %r746;
or.b32 %r1247, %r1246, -2139062144;
and.b32 %r1248, %r598, 2139062143;
sub.s32 %r1249, %r1247, %r1248;
xor.b32 %r603, %r770, %r1249;
xor.b32 %r1250, %r600, %r752;
or.b32 %r1251, %r1250, -2139062144;
and.b32 %r1252, %r600, 2139062143;
sub.s32 %r1253, %r1251, %r1252;
xor.b32 %r607, %r775, %r1253;
ld.global.nc.u32 %r604, [%rd117+20];
ld.global.nc.u32 %r608, [%rd117+24];

dp4a.s32.s32 %r602, %r603, %r604, %r594;


dp4a.s32.s32 %r606, %r607, %r608, %r602;


prmt.b32 %r610, %r610, 0, 0xba98;


prmt.b32 %r612, %r612, 0, 0xba98;

xor.b32 %r1254, %r610, %r781;
or.b32 %r1255, %r1254, -2139062144;
and.b32 %r1256, %r610, 2139062143;
sub.s32 %r1257, %r1255, %r1256;
xor.b32 %r615, %r805, %r1257;
xor.b32 %r1258, %r612, %r787;
or.b32 %r1259, %r1258, -2139062144;
and.b32 %r1260, %r612, 2139062143;
sub.s32 %r1261, %r1259, %r1260;
xor.b32 %r619, %r810, %r1261;
ld.global.nc.u32 %r616, [%rd117+28];
ld.global.nc.u32 %r620, [%rd117+32];

dp4a.s32.s32 %r614, %r615, %r616, %r606;


dp4a.s32.s32 %r618, %r619, %r620, %r614;

mul.lo.s32 %r1262, %r817, %r618;

{ cvt.f32.f16 %f93, %rs1;}


ld.global.nc.u32 %r622, [%rd117];

{.reg .f16 low,high;
mov.b32 {low,high},%r622;
cvt.f32.f16 %f94, low;}


mul.ftz.f32 %f117, %f93, %f94;
cvt.rn.f32.s32 %f118, %r1262;
fma.rn.ftz.f32 %f266, %f117, %f118, %f266;

prmt.b32 %r623, %r623, 0, 0xba98;


prmt.b32 %r625, %r625, 0, 0xba98;

xor.b32 %r1263, %r623, %r823;
or.b32 %r1264, %r1263, -2139062144;
and.b32 %r1265, %r623, 2139062143;
sub.s32 %r1266, %r1264, %r1265;
xor.b32 %r628, %r847, %r1266;
xor.b32 %r1267, %r625, %r829;
or.b32 %r1268, %r1267, -2139062144;
and.b32 %r1269, %r625, 2139062143;
sub.s32 %r1270, %r1268, %r1269;
xor.b32 %r632, %r852, %r1270;

dp4a.s32.s32 %r627, %r628, %r580, %r630;


dp4a.s32.s32 %r631, %r632, %r584, %r627;


prmt.b32 %r635, %r635, 0, 0xba98;


prmt.b32 %r637, %r637, 0, 0xba98;

xor.b32 %r1271, %r635, %r858;
or.b32 %r1272, %r1271, -2139062144;
and.b32 %r1273, %r635, 2139062143;
sub.s32 %r1274, %r1272, %r1273;
xor.b32 %r640, %r882, %r1274;
xor.b32 %r1275, %r637, %r864;
or.b32 %r1276, %r1275, -2139062144;
and.b32 %r1277, %r637, 2139062143;
sub.s32 %r1278, %r1276, %r1277;
xor.b32 %r644, %r887, %r1278;

dp4a.s32.s32 %r639, %r640, %r592, %r631;


dp4a.s32.s32 %r643, %r644, %r596, %r639;


prmt.b32 %r647, %r647, 0, 0xba98;


prmt.b32 %r649, %r649, 0, 0xba98;

xor.b32 %r1279, %r647, %r893;
or.b32 %r1280, %r1279, -2139062144;
and.b32 %r1281, %r647, 2139062143;
sub.s32 %r1282, %r1280, %r1281;
xor.b32 %r652, %r917, %r1282;
xor.b32 %r1283, %r649, %r899;
or.b32 %r1284, %r1283, -2139062144;
and.b32 %r1285, %r649, 2139062143;
sub.s32 %r1286, %r1284, %r1285;
xor.b32 %r656, %r922, %r1286;

dp4a.s32.s32 %r651, %r652, %r604, %r643;


dp4a.s32.s32 %r655, %r656, %r608, %r651;


prmt.b32 %r659, %r659, 0, 0xba98;


prmt.b32 %r661, %r661, 0, 0xba98;

xor.b32 %r1287, %r659, %r928;
or.b32 %r1288, %r1287, -2139062144;
and.b32 %r1289, %r659, 2139062143;
sub.s32 %r1290, %r1288, %r1289;
xor.b32 %r664, %r952, %r1290;
xor.b32 %r1291, %r661, %r934;
or.b32 %r1292, %r1291, -2139062144;
and.b32 %r1293, %r661, 2139062143;
sub.s32 %r1294, %r1292, %r1293;
xor.b32 %r668, %r957, %r1294;

dp4a.s32.s32 %r663, %r664, %r616, %r655;


dp4a.s32.s32 %r667, %r668, %r620, %r663;

mul.lo.s32 %r1295, %r964, %r667;

{ cvt.f32.f16 %f95, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r622;
cvt.f32.f16 %f96, low;}


mul.ftz.f32 %f119, %f95, %f96;
cvt.rn.f32.s32 %f120, %r1295;
fma.rn.ftz.f32 %f265, %f119, %f120, %f265;
add.s32 %r1524, %r1524, 64;
add.s32 %r1523, %r1523, 64;
add.s32 %r1522, %r1522, 64;
add.s32 %r1521, %r1521, 64;
add.s32 %r1520, %r1520, 64;
add.s32 %r1519, %r1519, 64;
add.s64 %rd149, %rd149, 880;
add.s32 %r1525, %r1525, 8;
setp.lt.s32 %p4, %r1525, %r5;
@%p4 bra $L__BB314_2;

$L__BB314_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB314_5;

mov.u32 %r1297, _ZZ13mul_mat_vec_qIL9ggml_type21ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r1298, %r1, 1536, %r1297;
shl.b32 %r1299, %r47, 2;
add.s32 %r1300, %r1298, %r1299;
st.shared.f32 [%r1300+-1536], %f276;
st.shared.f32 [%r1300+-1408], %f275;
st.shared.f32 [%r1300+-1280], %f274;
st.shared.f32 [%r1300+-1152], %f273;
st.shared.f32 [%r1300+-1024], %f272;
st.shared.f32 [%r1300+-896], %f271;
st.shared.f32 [%r1300+-768], %f270;
st.shared.f32 [%r1300+-640], %f269;
st.shared.f32 [%r1300+-512], %f268;
st.shared.f32 [%r1300+-384], %f267;
st.shared.f32 [%r1300+-256], %f266;
st.shared.f32 [%r1300+-128], %f265;

$L__BB314_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB314_30;

mov.u32 %r1301, %ctaid.z;
mov.u32 %r1302, %ctaid.y;
mad.lo.s32 %r1303, %r1302, %r38, %r3;
mad.lo.s32 %r1304, %r1301, %r44, %r1303;
cvt.u64.u32 %rd20, %r1304;
shl.b32 %r1306, %r47, 2;
mov.u32 %r1307, _ZZ13mul_mat_vec_qIL9ggml_type21ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r27, %r1307, %r1306;
mov.u32 %r1308, 2;
ld.shared.f32 %f121, [%r27];
add.ftz.f32 %f122, %f121, %f276;
mov.b32 %r1309, %f122;
mov.u32 %r1310, 31;
mov.u32 %r1311, 16;
mov.u32 %r1312, -1;
shfl.sync.bfly.b32 %r1313|%p7, %r1309, %r1311, %r1310, %r1312;
mov.b32 %f123, %r1313;
add.ftz.f32 %f124, %f122, %f123;
mov.b32 %r1314, %f124;
mov.u32 %r1315, 8;
shfl.sync.bfly.b32 %r1316|%p8, %r1314, %r1315, %r1310, %r1312;
mov.b32 %f125, %r1316;
add.ftz.f32 %f126, %f124, %f125;
mov.b32 %r1317, %f126;
mov.u32 %r1318, 4;
shfl.sync.bfly.b32 %r1319|%p9, %r1317, %r1318, %r1310, %r1312;
mov.b32 %f127, %r1319;
add.ftz.f32 %f128, %f126, %f127;
mov.b32 %r1320, %f128;
shfl.sync.bfly.b32 %r1321|%p10, %r1320, %r1308, %r1310, %r1312;
mov.b32 %f129, %r1321;
add.ftz.f32 %f130, %f128, %f129;
mov.b32 %r1322, %f130;
mov.u32 %r1323, 1;
shfl.sync.bfly.b32 %r1324|%p11, %r1322, %r1323, %r1310, %r1312;
mov.b32 %f131, %r1324;
add.ftz.f32 %f37, %f130, %f131;
setp.eq.s32 %p12, %r47, 0;
setp.lt.u32 %p13, %r3, %r32;
and.pred %p1, %p12, %p13;
mul.wide.u32 %rd118, %r1304, 4;
add.s64 %rd21, %rd1, %rd118;
not.pred %p14, %p1;
@%p14 bra $L__BB314_8;

st.global.f32 [%rd21], %f37;

$L__BB314_8:
ld.shared.f32 %f132, [%r27+128];
add.ftz.f32 %f133, %f132, %f275;
mov.b32 %r1325, %f133;
shfl.sync.bfly.b32 %r1329|%p15, %r1325, %r1311, %r1310, %r1312;
mov.b32 %f134, %r1329;
add.ftz.f32 %f135, %f133, %f134;
mov.b32 %r1330, %f135;
shfl.sync.bfly.b32 %r1332|%p16, %r1330, %r1315, %r1310, %r1312;
mov.b32 %f136, %r1332;
add.ftz.f32 %f137, %f135, %f136;
mov.b32 %r1333, %f137;
shfl.sync.bfly.b32 %r1335|%p17, %r1333, %r1318, %r1310, %r1312;
mov.b32 %f138, %r1335;
add.ftz.f32 %f139, %f137, %f138;
mov.b32 %r1336, %f139;
shfl.sync.bfly.b32 %r1338|%p18, %r1336, %r1308, %r1310, %r1312;
mov.b32 %f140, %r1338;
add.ftz.f32 %f141, %f139, %f140;
mov.b32 %r1339, %f141;
shfl.sync.bfly.b32 %r1341|%p19, %r1339, %r1323, %r1310, %r1312;
mov.b32 %f142, %r1341;
add.ftz.f32 %f38, %f141, %f142;
setp.eq.s32 %p20, %r47, 1;
add.s32 %r1343, %r3, 1;
setp.lt.u32 %p21, %r1343, %r32;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB314_10;

st.global.f32 [%rd21+4], %f38;

$L__BB314_10:
ld.shared.f32 %f143, [%r27+256];
add.ftz.f32 %f144, %f143, %f274;
mov.b32 %r1344, %f144;
mov.u32 %r1345, 31;
mov.u32 %r1346, 16;
mov.u32 %r1347, -1;
shfl.sync.bfly.b32 %r1348|%p23, %r1344, %r1346, %r1345, %r1347;
mov.b32 %f145, %r1348;
add.ftz.f32 %f146, %f144, %f145;
mov.b32 %r1349, %f146;
mov.u32 %r1350, 8;
shfl.sync.bfly.b32 %r1351|%p24, %r1349, %r1350, %r1345, %r1347;
mov.b32 %f147, %r1351;
add.ftz.f32 %f148, %f146, %f147;
mov.b32 %r1352, %f148;
mov.u32 %r1353, 4;
shfl.sync.bfly.b32 %r1354|%p25, %r1352, %r1353, %r1345, %r1347;
mov.b32 %f149, %r1354;
add.ftz.f32 %f150, %f148, %f149;
mov.b32 %r1355, %f150;
mov.u32 %r1356, 2;
shfl.sync.bfly.b32 %r1357|%p26, %r1355, %r1356, %r1345, %r1347;
mov.b32 %f151, %r1357;
add.ftz.f32 %f152, %f150, %f151;
mov.b32 %r1358, %f152;
mov.u32 %r1359, 1;
shfl.sync.bfly.b32 %r1360|%p27, %r1358, %r1359, %r1345, %r1347;
mov.b32 %f153, %r1360;
add.ftz.f32 %f39, %f152, %f153;
@%p14 bra $L__BB314_12;

cvt.u64.u32 %rd119, %r32;
add.s64 %rd120, %rd119, %rd20;
shl.b64 %rd121, %rd120, 2;
add.s64 %rd122, %rd1, %rd121;
st.global.f32 [%rd122], %f39;

$L__BB314_12:
ld.shared.f32 %f154, [%r27+384];
add.ftz.f32 %f155, %f154, %f273;
mov.b32 %r1361, %f155;
shfl.sync.bfly.b32 %r1365|%p29, %r1361, %r1346, %r1345, %r1347;
mov.b32 %f156, %r1365;
add.ftz.f32 %f157, %f155, %f156;
mov.b32 %r1366, %f157;
shfl.sync.bfly.b32 %r1368|%p30, %r1366, %r1350, %r1345, %r1347;
mov.b32 %f158, %r1368;
add.ftz.f32 %f159, %f157, %f158;
mov.b32 %r1369, %f159;
shfl.sync.bfly.b32 %r1371|%p31, %r1369, %r1353, %r1345, %r1347;
mov.b32 %f160, %r1371;
add.ftz.f32 %f161, %f159, %f160;
mov.b32 %r1372, %f161;
shfl.sync.bfly.b32 %r1374|%p32, %r1372, %r1356, %r1345, %r1347;
mov.b32 %f162, %r1374;
add.ftz.f32 %f163, %f161, %f162;
mov.b32 %r1375, %f163;
shfl.sync.bfly.b32 %r1377|%p33, %r1375, %r1359, %r1345, %r1347;
mov.b32 %f164, %r1377;
add.ftz.f32 %f40, %f163, %f164;
@%p22 bra $L__BB314_14;

add.s32 %r1378, %r32, 1;
cvt.u64.u32 %rd123, %r1378;
add.s64 %rd124, %rd123, %rd20;
shl.b64 %rd125, %rd124, 2;
add.s64 %rd126, %rd1, %rd125;
st.global.f32 [%rd126], %f40;

$L__BB314_14:
shl.b32 %r1379, %r32, 1;
mov.u32 %r1380, 1;
ld.shared.f32 %f165, [%r27+512];
add.ftz.f32 %f166, %f165, %f272;
mov.b32 %r1381, %f166;
mov.u32 %r1382, 31;
mov.u32 %r1383, 16;
mov.u32 %r1384, -1;
shfl.sync.bfly.b32 %r1385|%p35, %r1381, %r1383, %r1382, %r1384;
mov.b32 %f167, %r1385;
add.ftz.f32 %f168, %f166, %f167;
mov.b32 %r1386, %f168;
mov.u32 %r1387, 8;
shfl.sync.bfly.b32 %r1388|%p36, %r1386, %r1387, %r1382, %r1384;
mov.b32 %f169, %r1388;
add.ftz.f32 %f170, %f168, %f169;
mov.b32 %r1389, %f170;
mov.u32 %r1390, 4;
shfl.sync.bfly.b32 %r1391|%p37, %r1389, %r1390, %r1382, %r1384;
mov.b32 %f171, %r1391;
add.ftz.f32 %f172, %f170, %f171;
mov.b32 %r1392, %f172;
mov.u32 %r1393, 2;
shfl.sync.bfly.b32 %r1394|%p38, %r1392, %r1393, %r1382, %r1384;
mov.b32 %f173, %r1394;
add.ftz.f32 %f174, %f172, %f173;
mov.b32 %r1395, %f174;
shfl.sync.bfly.b32 %r1396|%p39, %r1395, %r1380, %r1382, %r1384;
mov.b32 %f175, %r1396;
add.ftz.f32 %f41, %f174, %f175;
cvt.u64.u32 %rd127, %r1379;
add.s64 %rd128, %rd127, %rd20;
shl.b64 %rd129, %rd128, 2;
add.s64 %rd22, %rd1, %rd129;
@%p14 bra $L__BB314_16;

st.global.f32 [%rd22], %f41;

$L__BB314_16:
ld.shared.f32 %f176, [%r27+640];
add.ftz.f32 %f177, %f176, %f271;
mov.b32 %r1397, %f177;
shfl.sync.bfly.b32 %r1401|%p41, %r1397, %r1383, %r1382, %r1384;
mov.b32 %f178, %r1401;
add.ftz.f32 %f179, %f177, %f178;
mov.b32 %r1402, %f179;
shfl.sync.bfly.b32 %r1404|%p42, %r1402, %r1387, %r1382, %r1384;
mov.b32 %f180, %r1404;
add.ftz.f32 %f181, %f179, %f180;
mov.b32 %r1405, %f181;
shfl.sync.bfly.b32 %r1407|%p43, %r1405, %r1390, %r1382, %r1384;
mov.b32 %f182, %r1407;
add.ftz.f32 %f183, %f181, %f182;
mov.b32 %r1408, %f183;
shfl.sync.bfly.b32 %r1410|%p44, %r1408, %r1393, %r1382, %r1384;
mov.b32 %f184, %r1410;
add.ftz.f32 %f185, %f183, %f184;
mov.b32 %r1411, %f185;
shfl.sync.bfly.b32 %r1413|%p45, %r1411, %r1380, %r1382, %r1384;
mov.b32 %f186, %r1413;
add.ftz.f32 %f42, %f185, %f186;
@%p22 bra $L__BB314_18;

st.global.f32 [%rd22+4], %f42;

$L__BB314_18:
mul.lo.s32 %r28, %r32, 3;
ld.shared.f32 %f187, [%r27+768];
add.ftz.f32 %f188, %f187, %f270;
mov.b32 %r1414, %f188;
mov.u32 %r1415, 31;
mov.u32 %r1416, 16;
mov.u32 %r1417, -1;
shfl.sync.bfly.b32 %r1418|%p47, %r1414, %r1416, %r1415, %r1417;
mov.b32 %f189, %r1418;
add.ftz.f32 %f190, %f188, %f189;
mov.b32 %r1419, %f190;
mov.u32 %r1420, 8;
shfl.sync.bfly.b32 %r1421|%p48, %r1419, %r1420, %r1415, %r1417;
mov.b32 %f191, %r1421;
add.ftz.f32 %f192, %f190, %f191;
mov.b32 %r1422, %f192;
mov.u32 %r1423, 4;
shfl.sync.bfly.b32 %r1424|%p49, %r1422, %r1423, %r1415, %r1417;
mov.b32 %f193, %r1424;
add.ftz.f32 %f194, %f192, %f193;
mov.b32 %r1425, %f194;
mov.u32 %r1426, 2;
shfl.sync.bfly.b32 %r1427|%p50, %r1425, %r1426, %r1415, %r1417;
mov.b32 %f195, %r1427;
add.ftz.f32 %f196, %f194, %f195;
mov.b32 %r1428, %f196;
mov.u32 %r1429, 1;
shfl.sync.bfly.b32 %r1430|%p51, %r1428, %r1429, %r1415, %r1417;
mov.b32 %f197, %r1430;
add.ftz.f32 %f43, %f196, %f197;
@%p14 bra $L__BB314_20;

cvt.u64.u32 %rd130, %r28;
add.s64 %rd131, %rd130, %rd20;
shl.b64 %rd132, %rd131, 2;
add.s64 %rd133, %rd1, %rd132;
st.global.f32 [%rd133], %f43;

$L__BB314_20:
ld.shared.f32 %f198, [%r27+896];
add.ftz.f32 %f199, %f198, %f269;
mov.b32 %r1431, %f199;
shfl.sync.bfly.b32 %r1435|%p53, %r1431, %r1416, %r1415, %r1417;
mov.b32 %f200, %r1435;
add.ftz.f32 %f201, %f199, %f200;
mov.b32 %r1436, %f201;
shfl.sync.bfly.b32 %r1438|%p54, %r1436, %r1420, %r1415, %r1417;
mov.b32 %f202, %r1438;
add.ftz.f32 %f203, %f201, %f202;
mov.b32 %r1439, %f203;
shfl.sync.bfly.b32 %r1441|%p55, %r1439, %r1423, %r1415, %r1417;
mov.b32 %f204, %r1441;
add.ftz.f32 %f205, %f203, %f204;
mov.b32 %r1442, %f205;
shfl.sync.bfly.b32 %r1444|%p56, %r1442, %r1426, %r1415, %r1417;
mov.b32 %f206, %r1444;
add.ftz.f32 %f207, %f205, %f206;
mov.b32 %r1445, %f207;
shfl.sync.bfly.b32 %r1447|%p57, %r1445, %r1429, %r1415, %r1417;
mov.b32 %f208, %r1447;
add.ftz.f32 %f44, %f207, %f208;
@%p22 bra $L__BB314_22;

add.s32 %r1448, %r28, 1;
cvt.u64.u32 %rd134, %r1448;
add.s64 %rd135, %rd134, %rd20;
shl.b64 %rd136, %rd135, 2;
add.s64 %rd137, %rd1, %rd136;
st.global.f32 [%rd137], %f44;

$L__BB314_22:
shl.b32 %r1449, %r32, 2;
mov.u32 %r1450, 2;
ld.shared.f32 %f209, [%r27+1024];
add.ftz.f32 %f210, %f209, %f268;
mov.b32 %r1451, %f210;
mov.u32 %r1452, 31;
mov.u32 %r1453, 16;
mov.u32 %r1454, -1;
shfl.sync.bfly.b32 %r1455|%p59, %r1451, %r1453, %r1452, %r1454;
mov.b32 %f211, %r1455;
add.ftz.f32 %f212, %f210, %f211;
mov.b32 %r1456, %f212;
mov.u32 %r1457, 8;
shfl.sync.bfly.b32 %r1458|%p60, %r1456, %r1457, %r1452, %r1454;
mov.b32 %f213, %r1458;
add.ftz.f32 %f214, %f212, %f213;
mov.b32 %r1459, %f214;
mov.u32 %r1460, 4;
shfl.sync.bfly.b32 %r1461|%p61, %r1459, %r1460, %r1452, %r1454;
mov.b32 %f215, %r1461;
add.ftz.f32 %f216, %f214, %f215;
mov.b32 %r1462, %f216;
shfl.sync.bfly.b32 %r1463|%p62, %r1462, %r1450, %r1452, %r1454;
mov.b32 %f217, %r1463;
add.ftz.f32 %f218, %f216, %f217;
mov.b32 %r1464, %f218;
mov.u32 %r1465, 1;
shfl.sync.bfly.b32 %r1466|%p63, %r1464, %r1465, %r1452, %r1454;
mov.b32 %f219, %r1466;
add.ftz.f32 %f45, %f218, %f219;
cvt.u64.u32 %rd138, %r1449;
add.s64 %rd139, %rd138, %rd20;
shl.b64 %rd140, %rd139, 2;
add.s64 %rd23, %rd1, %rd140;
@%p14 bra $L__BB314_24;

st.global.f32 [%rd23], %f45;

$L__BB314_24:
ld.shared.f32 %f220, [%r27+1152];
add.ftz.f32 %f221, %f220, %f267;
mov.b32 %r1467, %f221;
shfl.sync.bfly.b32 %r1471|%p65, %r1467, %r1453, %r1452, %r1454;
mov.b32 %f222, %r1471;
add.ftz.f32 %f223, %f221, %f222;
mov.b32 %r1472, %f223;
shfl.sync.bfly.b32 %r1474|%p66, %r1472, %r1457, %r1452, %r1454;
mov.b32 %f224, %r1474;
add.ftz.f32 %f225, %f223, %f224;
mov.b32 %r1475, %f225;
shfl.sync.bfly.b32 %r1477|%p67, %r1475, %r1460, %r1452, %r1454;
mov.b32 %f226, %r1477;
add.ftz.f32 %f227, %f225, %f226;
mov.b32 %r1478, %f227;
shfl.sync.bfly.b32 %r1480|%p68, %r1478, %r1450, %r1452, %r1454;
mov.b32 %f228, %r1480;
add.ftz.f32 %f229, %f227, %f228;
mov.b32 %r1481, %f229;
shfl.sync.bfly.b32 %r1483|%p69, %r1481, %r1465, %r1452, %r1454;
mov.b32 %f230, %r1483;
add.ftz.f32 %f46, %f229, %f230;
@%p22 bra $L__BB314_26;

st.global.f32 [%rd23+4], %f46;

$L__BB314_26:
mul.lo.s32 %r29, %r32, 5;
ld.shared.f32 %f231, [%r27+1280];
add.ftz.f32 %f232, %f231, %f266;
mov.b32 %r1484, %f232;
mov.u32 %r1485, 31;
mov.u32 %r1486, 16;
mov.u32 %r1487, -1;
shfl.sync.bfly.b32 %r1488|%p71, %r1484, %r1486, %r1485, %r1487;
mov.b32 %f233, %r1488;
add.ftz.f32 %f234, %f232, %f233;
mov.b32 %r1489, %f234;
mov.u32 %r1490, 8;
shfl.sync.bfly.b32 %r1491|%p72, %r1489, %r1490, %r1485, %r1487;
mov.b32 %f235, %r1491;
add.ftz.f32 %f236, %f234, %f235;
mov.b32 %r1492, %f236;
mov.u32 %r1493, 4;
shfl.sync.bfly.b32 %r1494|%p73, %r1492, %r1493, %r1485, %r1487;
mov.b32 %f237, %r1494;
add.ftz.f32 %f238, %f236, %f237;
mov.b32 %r1495, %f238;
mov.u32 %r1496, 2;
shfl.sync.bfly.b32 %r1497|%p74, %r1495, %r1496, %r1485, %r1487;
mov.b32 %f239, %r1497;
add.ftz.f32 %f240, %f238, %f239;
mov.b32 %r1498, %f240;
mov.u32 %r1499, 1;
shfl.sync.bfly.b32 %r1500|%p75, %r1498, %r1499, %r1485, %r1487;
mov.b32 %f241, %r1500;
add.ftz.f32 %f47, %f240, %f241;
@%p14 bra $L__BB314_28;

cvt.u64.u32 %rd141, %r29;
add.s64 %rd142, %rd141, %rd20;
shl.b64 %rd143, %rd142, 2;
add.s64 %rd144, %rd1, %rd143;
st.global.f32 [%rd144], %f47;

$L__BB314_28:
ld.shared.f32 %f242, [%r27+1408];
add.ftz.f32 %f243, %f242, %f265;
mov.b32 %r1501, %f243;
shfl.sync.bfly.b32 %r1505|%p77, %r1501, %r1486, %r1485, %r1487;
mov.b32 %f244, %r1505;
add.ftz.f32 %f245, %f243, %f244;
mov.b32 %r1506, %f245;
shfl.sync.bfly.b32 %r1508|%p78, %r1506, %r1490, %r1485, %r1487;
mov.b32 %f246, %r1508;
add.ftz.f32 %f247, %f245, %f246;
mov.b32 %r1509, %f247;
shfl.sync.bfly.b32 %r1511|%p79, %r1509, %r1493, %r1485, %r1487;
mov.b32 %f248, %r1511;
add.ftz.f32 %f249, %f247, %f248;
mov.b32 %r1512, %f249;
shfl.sync.bfly.b32 %r1514|%p80, %r1512, %r1496, %r1485, %r1487;
mov.b32 %f250, %r1514;
add.ftz.f32 %f251, %f249, %f250;
mov.b32 %r1515, %f251;
shfl.sync.bfly.b32 %r1517|%p81, %r1515, %r1499, %r1485, %r1487;
mov.b32 %f252, %r1517;
add.ftz.f32 %f48, %f251, %f252;
@%p22 bra $L__BB314_30;

add.s32 %r1518, %r29, 1;
cvt.u64.u32 %rd145, %r1518;
add.s64 %rd146, %rd145, %rd20;
shl.b64 %rd147, %rd146, 2;
add.s64 %rd148, %rd1, %rd147;
st.global.f32 [%rd148], %f48;

$L__BB314_30:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type21ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type21ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type21ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type21ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type21ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type21ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type21ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type21ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type21ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<95>;
.reg .b16 %rs<59>;
.reg .f32 %f<323>;
.reg .b32 %r<1729>;
.reg .b64 %rd<159>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type21ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[1792];

ld.param.u64 %rd25, [_Z13mul_mat_vec_qIL9ggml_type21ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd26, [_Z13mul_mat_vec_qIL9ggml_type21ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd27, [_Z13mul_mat_vec_qIL9ggml_type21ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r48, [_Z13mul_mat_vec_qIL9ggml_type21ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r33, [_Z13mul_mat_vec_qIL9ggml_type21ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r34, [_Z13mul_mat_vec_qIL9ggml_type21ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r35, [_Z13mul_mat_vec_qIL9ggml_type21ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type21ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type21ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type21ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r45, [_Z13mul_mat_vec_qIL9ggml_type21ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r46, [_Z13mul_mat_vec_qIL9ggml_type21ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r47, [_Z13mul_mat_vec_qIL9ggml_type21ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r43, [_Z13mul_mat_vec_qIL9ggml_type21ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type21ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r37, [_Z13mul_mat_vec_qIL9ggml_type21ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r36, [_Z13mul_mat_vec_qIL9ggml_type21ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
cvta.to.global.u64 %rd1, %rd27;
mov.u32 %r1, %tid.y;
shl.b32 %r49, %r1, 5;
mov.u32 %r50, %tid.x;
add.s32 %r2, %r49, %r50;
mov.u32 %r51, %ctaid.x;
shl.b32 %r3, %r51, 1;
shr.s32 %r52, %r2, 31;
shr.u32 %r53, %r52, 29;
add.s32 %r54, %r2, %r53;
shr.s32 %r1728, %r54, 3;
shr.u32 %r5, %r48, 8;
setp.ge.s32 %p3, %r1728, %r5;
mov.f32 %f309, 0f00000000;
mov.f32 %f310, %f309;
mov.f32 %f311, %f309;
mov.f32 %f312, %f309;
mov.f32 %f313, %f309;
mov.f32 %f314, %f309;
mov.f32 %f315, %f309;
mov.f32 %f316, %f309;
mov.f32 %f317, %f309;
mov.f32 %f318, %f309;
mov.f32 %f319, %f309;
mov.f32 %f320, %f309;
mov.f32 %f321, %f309;
mov.f32 %f322, %f309;
@%p3 bra $L__BB315_3;

mov.u32 %r55, %ctaid.y;
mov.u32 %r56, %ctaid.z;
mul.lo.s32 %r57, %r55, %r40;
cvt.u64.u32 %rd28, %r57;
mul.lo.s32 %r58, %r56, %r46;
cvt.u64.u32 %rd29, %r58;
add.s64 %rd2, %rd29, %rd28;
mul.hi.u32 %r59, %r56, %r42;
add.s32 %r60, %r59, %r56;
shr.u32 %r61, %r60, %r43;
mul.lo.s32 %r62, %r61, %r45;
mul.hi.u32 %r63, %r55, %r36;
add.s32 %r64, %r63, %r55;
shr.u32 %r65, %r64, %r37;
mul.lo.s32 %r66, %r65, %r39;
cvta.to.global.u64 %rd158, %rd25;
and.b32 %r70, %r54, -8;
sub.s32 %r71, %r2, %r70;
shl.b32 %r72, %r71, 1;
shr.u32 %r73, %r71, 31;
add.s32 %r74, %r71, %r73;
shr.s32 %r75, %r74, 1;
cvt.s64.s32 %rd30, %r75;
shl.b32 %r76, %r71, 2;
cvt.s64.s32 %rd4, %r71;
and.b32 %r6, %r76, 4;
shl.b32 %r77, %r34, 1;
shl.b32 %r1721, %r1728, 3;
add.s32 %r1727, %r1721, %r77;
mad.lo.s32 %r1726, %r34, 3, %r1721;
shl.b32 %r78, %r34, 2;
add.s32 %r1725, %r1721, %r78;
mad.lo.s32 %r1724, %r34, 5, %r1721;
mad.lo.s32 %r1723, %r34, 6, %r1721;
add.s32 %r1722, %r34, %r1721;
add.s32 %r79, %r1728, %r62;
add.s32 %r80, %r79, %r66;
add.s32 %r81, %r3, 1;
mad.lo.s32 %r82, %r33, %r81, %r80;
mul.wide.s32 %rd5, %r82, 110;
add.s64 %rd31, %rd5, %rd30;
add.s64 %rd6, %rd31, 106;
add.s64 %rd32, %rd5, 74;
not.b32 %r83, %r72;
mul.wide.s32 %rd33, %r83, 2;
sub.s64 %rd7, %rd32, %rd33;
mul.wide.s32 %rd34, %r72, 2;
add.s64 %rd35, %rd5, %rd34;
add.s64 %rd8, %rd35, 74;
add.s64 %rd36, %rd5, %rd4;
add.s64 %rd9, %rd36, 66;
mad.lo.s32 %r84, %r3, %r33, %r1728;
add.s32 %r85, %r84, %r62;
add.s32 %r86, %r85, %r66;
mul.wide.s32 %rd10, %r86, 110;
mul.wide.s32 %rd37, %r76, 2;
add.s64 %rd38, %rd10, %rd37;
add.s64 %rd11, %rd38, 4;
add.s64 %rd39, %rd5, %rd37;
add.s64 %rd12, %rd39, 2;
add.s64 %rd40, %rd10, %rd30;
add.s64 %rd13, %rd40, 106;
add.s64 %rd41, %rd10, 74;
sub.s64 %rd14, %rd41, %rd33;
add.s64 %rd42, %rd10, %rd34;
add.s64 %rd15, %rd42, 74;
add.s64 %rd43, %rd10, %rd4;
add.s64 %rd16, %rd43, 66;
cvta.to.global.u64 %rd17, %rd26;
mov.f32 %f309, 0f00000000;

$L__BB315_2:
cvt.u64.u32 %rd44, %r1721;
add.s64 %rd45, %rd2, %rd44;
add.s64 %rd46, %rd45, %rd4;
mul.lo.s64 %rd47, %rd46, 36;
add.s64 %rd48, %rd158, %rd11;
ld.global.nc.u16 %rs15, [%rd48];
ld.global.nc.u16 %rs16, [%rd48+2];
ld.global.nc.u16 %rs17, [%rd48+4];
add.s64 %rd49, %rd158, %rd16;
ld.global.nc.u8 %rs18, [%rd49];
add.s64 %rd50, %rd158, %rd15;
add.s64 %rd51, %rd158, %rd14;
ld.global.nc.u16 %rs19, [%rd50];
ld.global.nc.u16 %rs20, [%rd51];
ld.global.nc.u16 %rs21, [%rd48+-2];
and.b16 %rs22, %rs21, 255;
cvt.u32.u16 %r773, %rs22;
shl.b16 %rs23, %rs18, 8;
cvt.u32.u16 %r774, %rs23;
and.b32 %r775, %r774, 256;
or.b32 %r776, %r775, %r773;
mul.wide.u32 %rd52, %r776, 4;
mov.u64 %rd53, iq3s_grid;
add.s64 %rd54, %rd53, %rd52;
ld.global.nc.u32 %r777, [%rd54];
not.b32 %r778, %r777;
shr.u16 %rs24, %rs21, 8;
cvt.u32.u16 %r779, %rs24;
mul.wide.u16 %r780, %rs18, 128;
and.b32 %r781, %r780, 256;
or.b32 %r782, %r781, %r779;
mul.wide.u32 %rd55, %r782, 4;
add.s64 %rd56, %rd53, %rd55;
ld.global.nc.u32 %r783, [%rd56];
not.b32 %r784, %r783;
and.b16 %rs26, %rs19, 255;
mul.wide.u16 %r785, %rs26, 128;
cvt.u32.u16 %r786, %rs19;
shl.b32 %r787, %r786, 21;
or.b32 %r788, %r785, %r787;
and.b32 %r789, %r788, 25166208;
or.b32 %r790, %r789, -2139062144;
add.s32 %r791, %r790, -16843009;
or.b32 %r675, %r791, %r789;
mov.u32 %r87, %r675;

prmt.b32 %r87, %r87, 0, 0xba98;

mul.wide.u16 %r792, %rs26, 8;
shl.b32 %r793, %r786, 17;
or.b32 %r794, %r792, %r793;
and.b32 %r795, %r794, 25166208;
or.b32 %r796, %r795, -2139062144;
add.s32 %r797, %r796, -16843009;
or.b32 %r677, %r797, %r795;
mov.u32 %r89, %r677;

prmt.b32 %r89, %r89, 0, 0xba98;

xor.b32 %r798, %r87, %r777;
or.b32 %r799, %r798, -2139062144;
and.b32 %r800, %r87, 2139062143;
and.b32 %r801, %r778, -2139062144;
sub.s32 %r802, %r799, %r800;
xor.b32 %r92, %r801, %r802;
xor.b32 %r803, %r89, %r783;
or.b32 %r804, %r803, -2139062144;
and.b32 %r805, %r89, 2139062143;
and.b32 %r806, %r784, -2139062144;
sub.s32 %r807, %r804, %r805;
xor.b32 %r96, %r806, %r807;
add.s64 %rd57, %rd17, %rd47;
ld.global.nc.u32 %r93, [%rd57+4];
ld.global.nc.u32 %r97, [%rd57+8];
mov.u32 %r731, 0;

dp4a.s32.s32 %r91, %r92, %r93, %r731;


dp4a.s32.s32 %r95, %r96, %r97, %r91;

and.b16 %rs27, %rs15, 255;
cvt.u32.u16 %r808, %rs27;
mul.wide.u16 %r809, %rs18, 64;
and.b32 %r810, %r809, 256;
or.b32 %r811, %r810, %r808;
mul.wide.u32 %rd58, %r811, 4;
add.s64 %rd59, %rd53, %rd58;
ld.global.nc.u32 %r812, [%rd59];
not.b32 %r813, %r812;
shr.u16 %rs28, %rs15, 8;
cvt.u32.u16 %r814, %rs28;
mul.wide.u16 %r815, %rs18, 32;
and.b32 %r816, %r815, 256;
or.b32 %r817, %r816, %r814;
mul.wide.u32 %rd60, %r817, 4;
add.s64 %rd61, %rd53, %rd60;
ld.global.nc.u32 %r818, [%rd61];
not.b32 %r819, %r818;
shr.u16 %rs29, %rs19, 8;
mul.wide.u16 %r820, %rs29, 128;
cvt.u32.u16 %r821, %rs29;
shl.b32 %r822, %r821, 21;
or.b32 %r823, %r820, %r822;
and.b32 %r824, %r823, 25166208;
or.b32 %r825, %r824, -2139062144;
add.s32 %r826, %r825, -16843009;
or.b32 %r687, %r826, %r824;
mov.u32 %r99, %r687;

prmt.b32 %r99, %r99, 0, 0xba98;

mul.wide.u16 %r827, %rs29, 8;
shl.b32 %r828, %r821, 17;
or.b32 %r829, %r827, %r828;
and.b32 %r830, %r829, 25166208;
or.b32 %r831, %r830, -2139062144;
add.s32 %r832, %r831, -16843009;
or.b32 %r689, %r832, %r830;
mov.u32 %r101, %r689;

prmt.b32 %r101, %r101, 0, 0xba98;

xor.b32 %r833, %r99, %r812;
or.b32 %r834, %r833, -2139062144;
and.b32 %r835, %r99, 2139062143;
and.b32 %r836, %r813, -2139062144;
sub.s32 %r837, %r834, %r835;
xor.b32 %r104, %r836, %r837;
xor.b32 %r838, %r101, %r818;
or.b32 %r839, %r838, -2139062144;
and.b32 %r840, %r101, 2139062143;
and.b32 %r841, %r819, -2139062144;
sub.s32 %r842, %r839, %r840;
xor.b32 %r108, %r841, %r842;
ld.global.nc.u32 %r105, [%rd57+12];
ld.global.nc.u32 %r109, [%rd57+16];

dp4a.s32.s32 %r103, %r104, %r105, %r95;


dp4a.s32.s32 %r107, %r108, %r109, %r103;

and.b16 %rs30, %rs16, 255;
cvt.u32.u16 %r843, %rs30;
mul.wide.u16 %r844, %rs18, 16;
and.b32 %r845, %r844, 256;
or.b32 %r846, %r845, %r843;
mul.wide.u32 %rd62, %r846, 4;
add.s64 %rd63, %rd53, %rd62;
ld.global.nc.u32 %r847, [%rd63];
not.b32 %r848, %r847;
shr.u16 %rs31, %rs16, 8;
cvt.u32.u16 %r849, %rs31;
mul.wide.u16 %r850, %rs18, 8;
and.b32 %r851, %r850, 256;
or.b32 %r852, %r851, %r849;
mul.wide.u32 %rd64, %r852, 4;
add.s64 %rd65, %rd53, %rd64;
ld.global.nc.u32 %r853, [%rd65];
not.b32 %r854, %r853;
and.b16 %rs32, %rs20, 255;
mul.wide.u16 %r855, %rs32, 128;
cvt.u32.u16 %r856, %rs20;
shl.b32 %r857, %r856, 21;
or.b32 %r858, %r855, %r857;
and.b32 %r859, %r858, 25166208;
or.b32 %r860, %r859, -2139062144;
add.s32 %r861, %r860, -16843009;
or.b32 %r699, %r861, %r859;
mov.u32 %r111, %r699;

prmt.b32 %r111, %r111, 0, 0xba98;

mul.wide.u16 %r862, %rs32, 8;
shl.b32 %r863, %r856, 17;
or.b32 %r864, %r862, %r863;
and.b32 %r865, %r864, 25166208;
or.b32 %r866, %r865, -2139062144;
add.s32 %r867, %r866, -16843009;
or.b32 %r701, %r867, %r865;
mov.u32 %r113, %r701;

prmt.b32 %r113, %r113, 0, 0xba98;

xor.b32 %r868, %r111, %r847;
or.b32 %r869, %r868, -2139062144;
and.b32 %r870, %r111, 2139062143;
and.b32 %r871, %r848, -2139062144;
sub.s32 %r872, %r869, %r870;
xor.b32 %r116, %r871, %r872;
xor.b32 %r873, %r113, %r853;
or.b32 %r874, %r873, -2139062144;
and.b32 %r875, %r113, 2139062143;
and.b32 %r876, %r854, -2139062144;
sub.s32 %r877, %r874, %r875;
xor.b32 %r120, %r876, %r877;
ld.global.nc.u32 %r117, [%rd57+20];
ld.global.nc.u32 %r121, [%rd57+24];

dp4a.s32.s32 %r115, %r116, %r117, %r107;


dp4a.s32.s32 %r119, %r120, %r121, %r115;

and.b16 %rs33, %rs17, 255;
cvt.u32.u16 %r878, %rs33;
mul.wide.u16 %r879, %rs18, 4;
and.b32 %r880, %r879, 256;
or.b32 %r881, %r880, %r878;
mul.wide.u32 %rd66, %r881, 4;
add.s64 %rd67, %rd53, %rd66;
ld.global.nc.u32 %r882, [%rd67];
not.b32 %r883, %r882;
shr.u16 %rs34, %rs17, 8;
cvt.u32.u16 %r884, %rs34;
mul.wide.u16 %r885, %rs18, 2;
and.b32 %r886, %r885, 256;
or.b32 %r887, %r886, %r884;
mul.wide.u32 %rd68, %r887, 4;
add.s64 %rd69, %rd53, %rd68;
ld.global.nc.u32 %r888, [%rd69];
not.b32 %r889, %r888;
shr.u16 %rs35, %rs20, 8;
mul.wide.u16 %r890, %rs35, 128;
cvt.u32.u16 %r891, %rs35;
shl.b32 %r892, %r891, 21;
or.b32 %r893, %r890, %r892;
and.b32 %r894, %r893, 25166208;
or.b32 %r895, %r894, -2139062144;
add.s32 %r896, %r895, -16843009;
or.b32 %r711, %r896, %r894;
mov.u32 %r123, %r711;

prmt.b32 %r123, %r123, 0, 0xba98;

mul.wide.u16 %r897, %rs35, 8;
shl.b32 %r898, %r891, 17;
or.b32 %r899, %r897, %r898;
and.b32 %r900, %r899, 25166208;
or.b32 %r901, %r900, -2139062144;
add.s32 %r902, %r901, -16843009;
or.b32 %r713, %r902, %r900;
mov.u32 %r125, %r713;

prmt.b32 %r125, %r125, 0, 0xba98;

xor.b32 %r903, %r123, %r882;
or.b32 %r904, %r903, -2139062144;
and.b32 %r905, %r123, 2139062143;
and.b32 %r906, %r883, -2139062144;
sub.s32 %r907, %r904, %r905;
xor.b32 %r128, %r906, %r907;
xor.b32 %r908, %r125, %r888;
or.b32 %r909, %r908, -2139062144;
and.b32 %r910, %r125, 2139062143;
and.b32 %r911, %r889, -2139062144;
sub.s32 %r912, %r909, %r910;
xor.b32 %r132, %r911, %r912;
ld.global.nc.u32 %r129, [%rd57+28];
ld.global.nc.u32 %r133, [%rd57+32];

dp4a.s32.s32 %r127, %r128, %r129, %r119;


dp4a.s32.s32 %r131, %r132, %r133, %r127;

add.s64 %rd70, %rd158, %rd13;
ld.global.nc.u8 %rs36, [%rd70];
cvt.u32.u16 %r913, %rs36;
and.b32 %r914, %r913, 255;
shr.u32 %r915, %r914, %r6;
shl.b32 %r916, %r915, 1;
and.b32 %r917, %r916, 30;
or.b32 %r918, %r917, 1;
mul.lo.s32 %r919, %r918, %r131;
add.s64 %rd71, %rd158, %rd10;
ld.global.nc.u16 %rs1, [%rd71];

{ cvt.f32.f16 %f85, %rs1;}


ld.global.nc.u32 %r135, [%rd57];

{.reg .f16 low,high;
mov.b32 {low,high},%r135;
cvt.f32.f16 %f86, low;}


mul.ftz.f32 %f113, %f85, %f86;
cvt.rn.f32.s32 %f114, %r919;
fma.rn.ftz.f32 %f322, %f113, %f114, %f322;
add.s64 %rd72, %rd158, %rd12;
ld.global.nc.u16 %rs37, [%rd72+2];
ld.global.nc.u16 %rs38, [%rd72+4];
ld.global.nc.u16 %rs39, [%rd72+6];
add.s64 %rd73, %rd158, %rd9;
ld.global.nc.u8 %rs40, [%rd73];
add.s64 %rd74, %rd158, %rd8;
add.s64 %rd75, %rd158, %rd7;
ld.global.nc.u16 %rs41, [%rd74];
ld.global.nc.u16 %rs42, [%rd75];
ld.global.nc.u16 %rs43, [%rd72];
and.b16 %rs44, %rs43, 255;
cvt.u32.u16 %r920, %rs44;
shl.b16 %rs45, %rs40, 8;
cvt.u32.u16 %r921, %rs45;
and.b32 %r922, %r921, 256;
or.b32 %r923, %r922, %r920;
mul.wide.u32 %rd76, %r923, 4;
add.s64 %rd77, %rd53, %rd76;
ld.global.nc.u32 %r924, [%rd77];
not.b32 %r925, %r924;
shr.u16 %rs46, %rs43, 8;
cvt.u32.u16 %r926, %rs46;
mul.wide.u16 %r927, %rs40, 128;
and.b32 %r928, %r927, 256;
or.b32 %r929, %r928, %r926;
mul.wide.u32 %rd78, %r929, 4;
add.s64 %rd79, %rd53, %rd78;
ld.global.nc.u32 %r930, [%rd79];
not.b32 %r931, %r930;
and.b16 %rs48, %rs41, 255;
mul.wide.u16 %r932, %rs48, 128;
cvt.u32.u16 %r933, %rs41;
shl.b32 %r934, %r933, 21;
or.b32 %r935, %r932, %r934;
and.b32 %r936, %r935, 25166208;
or.b32 %r937, %r936, -2139062144;
add.s32 %r938, %r937, -16843009;
or.b32 %r724, %r938, %r936;
mov.u32 %r136, %r724;

prmt.b32 %r136, %r136, 0, 0xba98;

mul.wide.u16 %r939, %rs48, 8;
shl.b32 %r940, %r933, 17;
or.b32 %r941, %r939, %r940;
and.b32 %r942, %r941, 25166208;
or.b32 %r943, %r942, -2139062144;
add.s32 %r944, %r943, -16843009;
or.b32 %r726, %r944, %r942;
mov.u32 %r138, %r726;

prmt.b32 %r138, %r138, 0, 0xba98;

xor.b32 %r945, %r136, %r924;
or.b32 %r946, %r945, -2139062144;
and.b32 %r947, %r136, 2139062143;
and.b32 %r948, %r925, -2139062144;
sub.s32 %r949, %r946, %r947;
xor.b32 %r141, %r948, %r949;
xor.b32 %r950, %r138, %r930;
or.b32 %r951, %r950, -2139062144;
and.b32 %r952, %r138, 2139062143;
and.b32 %r953, %r931, -2139062144;
sub.s32 %r954, %r951, %r952;
xor.b32 %r145, %r953, %r954;

dp4a.s32.s32 %r140, %r141, %r93, %r731;


dp4a.s32.s32 %r144, %r145, %r97, %r140;

and.b16 %rs49, %rs37, 255;
cvt.u32.u16 %r955, %rs49;
mul.wide.u16 %r956, %rs40, 64;
and.b32 %r957, %r956, 256;
or.b32 %r958, %r957, %r955;
mul.wide.u32 %rd80, %r958, 4;
add.s64 %rd81, %rd53, %rd80;
ld.global.nc.u32 %r959, [%rd81];
not.b32 %r960, %r959;
shr.u16 %rs50, %rs37, 8;
cvt.u32.u16 %r961, %rs50;
mul.wide.u16 %r962, %rs40, 32;
and.b32 %r963, %r962, 256;
or.b32 %r964, %r963, %r961;
mul.wide.u32 %rd82, %r964, 4;
add.s64 %rd83, %rd53, %rd82;
ld.global.nc.u32 %r965, [%rd83];
not.b32 %r966, %r965;
shr.u16 %rs51, %rs41, 8;
mul.wide.u16 %r967, %rs51, 128;
cvt.u32.u16 %r968, %rs51;
shl.b32 %r969, %r968, 21;
or.b32 %r970, %r967, %r969;
and.b32 %r971, %r970, 25166208;
or.b32 %r972, %r971, -2139062144;
add.s32 %r973, %r972, -16843009;
or.b32 %r736, %r973, %r971;
mov.u32 %r148, %r736;

prmt.b32 %r148, %r148, 0, 0xba98;

mul.wide.u16 %r974, %rs51, 8;
shl.b32 %r975, %r968, 17;
or.b32 %r976, %r974, %r975;
and.b32 %r977, %r976, 25166208;
or.b32 %r978, %r977, -2139062144;
add.s32 %r979, %r978, -16843009;
or.b32 %r738, %r979, %r977;
mov.u32 %r150, %r738;

prmt.b32 %r150, %r150, 0, 0xba98;

xor.b32 %r980, %r148, %r959;
or.b32 %r981, %r980, -2139062144;
and.b32 %r982, %r148, 2139062143;
and.b32 %r983, %r960, -2139062144;
sub.s32 %r984, %r981, %r982;
xor.b32 %r153, %r983, %r984;
xor.b32 %r985, %r150, %r965;
or.b32 %r986, %r985, -2139062144;
and.b32 %r987, %r150, 2139062143;
and.b32 %r988, %r966, -2139062144;
sub.s32 %r989, %r986, %r987;
xor.b32 %r157, %r988, %r989;

dp4a.s32.s32 %r152, %r153, %r105, %r144;


dp4a.s32.s32 %r156, %r157, %r109, %r152;

and.b16 %rs52, %rs38, 255;
cvt.u32.u16 %r990, %rs52;
mul.wide.u16 %r991, %rs40, 16;
and.b32 %r992, %r991, 256;
or.b32 %r993, %r992, %r990;
mul.wide.u32 %rd84, %r993, 4;
add.s64 %rd85, %rd53, %rd84;
ld.global.nc.u32 %r994, [%rd85];
not.b32 %r995, %r994;
shr.u16 %rs53, %rs38, 8;
cvt.u32.u16 %r996, %rs53;
mul.wide.u16 %r997, %rs40, 8;
and.b32 %r998, %r997, 256;
or.b32 %r999, %r998, %r996;
mul.wide.u32 %rd86, %r999, 4;
add.s64 %rd87, %rd53, %rd86;
ld.global.nc.u32 %r1000, [%rd87];
not.b32 %r1001, %r1000;
and.b16 %rs54, %rs42, 255;
mul.wide.u16 %r1002, %rs54, 128;
cvt.u32.u16 %r1003, %rs42;
shl.b32 %r1004, %r1003, 21;
or.b32 %r1005, %r1002, %r1004;
and.b32 %r1006, %r1005, 25166208;
or.b32 %r1007, %r1006, -2139062144;
add.s32 %r1008, %r1007, -16843009;
or.b32 %r748, %r1008, %r1006;
mov.u32 %r160, %r748;

prmt.b32 %r160, %r160, 0, 0xba98;

mul.wide.u16 %r1009, %rs54, 8;
shl.b32 %r1010, %r1003, 17;
or.b32 %r1011, %r1009, %r1010;
and.b32 %r1012, %r1011, 25166208;
or.b32 %r1013, %r1012, -2139062144;
add.s32 %r1014, %r1013, -16843009;
or.b32 %r750, %r1014, %r1012;
mov.u32 %r162, %r750;

prmt.b32 %r162, %r162, 0, 0xba98;

xor.b32 %r1015, %r160, %r994;
or.b32 %r1016, %r1015, -2139062144;
and.b32 %r1017, %r160, 2139062143;
and.b32 %r1018, %r995, -2139062144;
sub.s32 %r1019, %r1016, %r1017;
xor.b32 %r165, %r1018, %r1019;
xor.b32 %r1020, %r162, %r1000;
or.b32 %r1021, %r1020, -2139062144;
and.b32 %r1022, %r162, 2139062143;
and.b32 %r1023, %r1001, -2139062144;
sub.s32 %r1024, %r1021, %r1022;
xor.b32 %r169, %r1023, %r1024;

dp4a.s32.s32 %r164, %r165, %r117, %r156;


dp4a.s32.s32 %r168, %r169, %r121, %r164;

and.b16 %rs55, %rs39, 255;
cvt.u32.u16 %r1025, %rs55;
mul.wide.u16 %r1026, %rs40, 4;
and.b32 %r1027, %r1026, 256;
or.b32 %r1028, %r1027, %r1025;
mul.wide.u32 %rd88, %r1028, 4;
add.s64 %rd89, %rd53, %rd88;
ld.global.nc.u32 %r1029, [%rd89];
not.b32 %r1030, %r1029;
shr.u16 %rs56, %rs39, 8;
cvt.u32.u16 %r1031, %rs56;
mul.wide.u16 %r1032, %rs40, 2;
and.b32 %r1033, %r1032, 256;
or.b32 %r1034, %r1033, %r1031;
mul.wide.u32 %rd90, %r1034, 4;
add.s64 %rd91, %rd53, %rd90;
ld.global.nc.u32 %r1035, [%rd91];
not.b32 %r1036, %r1035;
shr.u16 %rs57, %rs42, 8;
mul.wide.u16 %r1037, %rs57, 128;
cvt.u32.u16 %r1038, %rs57;
shl.b32 %r1039, %r1038, 21;
or.b32 %r1040, %r1037, %r1039;
and.b32 %r1041, %r1040, 25166208;
or.b32 %r1042, %r1041, -2139062144;
add.s32 %r1043, %r1042, -16843009;
or.b32 %r760, %r1043, %r1041;
mov.u32 %r172, %r760;

prmt.b32 %r172, %r172, 0, 0xba98;

mul.wide.u16 %r1044, %rs57, 8;
shl.b32 %r1045, %r1038, 17;
or.b32 %r1046, %r1044, %r1045;
and.b32 %r1047, %r1046, 25166208;
or.b32 %r1048, %r1047, -2139062144;
add.s32 %r1049, %r1048, -16843009;
or.b32 %r762, %r1049, %r1047;
mov.u32 %r174, %r762;

prmt.b32 %r174, %r174, 0, 0xba98;

xor.b32 %r1050, %r172, %r1029;
or.b32 %r1051, %r1050, -2139062144;
and.b32 %r1052, %r172, 2139062143;
and.b32 %r1053, %r1030, -2139062144;
sub.s32 %r1054, %r1051, %r1052;
xor.b32 %r177, %r1053, %r1054;
xor.b32 %r1055, %r174, %r1035;
or.b32 %r1056, %r1055, -2139062144;
and.b32 %r1057, %r174, 2139062143;
and.b32 %r1058, %r1036, -2139062144;
sub.s32 %r1059, %r1056, %r1057;
xor.b32 %r181, %r1058, %r1059;

dp4a.s32.s32 %r176, %r177, %r129, %r168;


dp4a.s32.s32 %r180, %r181, %r133, %r176;

add.s64 %rd92, %rd158, %rd6;
ld.global.nc.u8 %rs58, [%rd92];
cvt.u32.u16 %r1060, %rs58;
and.b32 %r1061, %r1060, 255;
shr.u32 %r1062, %r1061, %r6;
shl.b32 %r1063, %r1062, 1;
and.b32 %r1064, %r1063, 30;
or.b32 %r1065, %r1064, 1;
mul.lo.s32 %r1066, %r1065, %r180;
add.s64 %rd93, %rd158, %rd5;
ld.global.nc.u16 %rs2, [%rd93];

{ cvt.f32.f16 %f87, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r135;
cvt.f32.f16 %f88, low;}


mul.ftz.f32 %f115, %f87, %f88;
cvt.rn.f32.s32 %f116, %r1066;
fma.rn.ftz.f32 %f321, %f115, %f116, %f321;
cvt.u64.u32 %rd94, %r1722;
add.s64 %rd95, %rd2, %rd94;
add.s64 %rd96, %rd95, %rd4;
mul.lo.s64 %rd97, %rd96, 36;
mov.u32 %r185, %r675;

prmt.b32 %r185, %r185, 0, 0xba98;

mov.u32 %r187, %r677;

prmt.b32 %r187, %r187, 0, 0xba98;

xor.b32 %r1067, %r185, %r777;
or.b32 %r1068, %r1067, -2139062144;
and.b32 %r1069, %r185, 2139062143;
sub.s32 %r1070, %r1068, %r1069;
xor.b32 %r190, %r801, %r1070;
xor.b32 %r1071, %r187, %r783;
or.b32 %r1072, %r1071, -2139062144;
and.b32 %r1073, %r187, 2139062143;
sub.s32 %r1074, %r1072, %r1073;
xor.b32 %r194, %r806, %r1074;
add.s64 %rd98, %rd17, %rd97;
ld.global.nc.u32 %r191, [%rd98+4];
ld.global.nc.u32 %r195, [%rd98+8];

dp4a.s32.s32 %r189, %r190, %r191, %r731;


dp4a.s32.s32 %r193, %r194, %r195, %r189;

mov.u32 %r197, %r687;

prmt.b32 %r197, %r197, 0, 0xba98;

mov.u32 %r199, %r689;

prmt.b32 %r199, %r199, 0, 0xba98;

xor.b32 %r1075, %r197, %r812;
or.b32 %r1076, %r1075, -2139062144;
and.b32 %r1077, %r197, 2139062143;
sub.s32 %r1078, %r1076, %r1077;
xor.b32 %r202, %r836, %r1078;
xor.b32 %r1079, %r199, %r818;
or.b32 %r1080, %r1079, -2139062144;
and.b32 %r1081, %r199, 2139062143;
sub.s32 %r1082, %r1080, %r1081;
xor.b32 %r206, %r841, %r1082;
ld.global.nc.u32 %r203, [%rd98+12];
ld.global.nc.u32 %r207, [%rd98+16];

dp4a.s32.s32 %r201, %r202, %r203, %r193;


dp4a.s32.s32 %r205, %r206, %r207, %r201;

mov.u32 %r209, %r699;

prmt.b32 %r209, %r209, 0, 0xba98;

mov.u32 %r211, %r701;

prmt.b32 %r211, %r211, 0, 0xba98;

xor.b32 %r1083, %r209, %r847;
or.b32 %r1084, %r1083, -2139062144;
and.b32 %r1085, %r209, 2139062143;
sub.s32 %r1086, %r1084, %r1085;
xor.b32 %r214, %r871, %r1086;
xor.b32 %r1087, %r211, %r853;
or.b32 %r1088, %r1087, -2139062144;
and.b32 %r1089, %r211, 2139062143;
sub.s32 %r1090, %r1088, %r1089;
xor.b32 %r218, %r876, %r1090;
ld.global.nc.u32 %r215, [%rd98+20];
ld.global.nc.u32 %r219, [%rd98+24];

dp4a.s32.s32 %r213, %r214, %r215, %r205;


dp4a.s32.s32 %r217, %r218, %r219, %r213;

mov.u32 %r221, %r711;

prmt.b32 %r221, %r221, 0, 0xba98;

mov.u32 %r223, %r713;

prmt.b32 %r223, %r223, 0, 0xba98;

xor.b32 %r1091, %r221, %r882;
or.b32 %r1092, %r1091, -2139062144;
and.b32 %r1093, %r221, 2139062143;
sub.s32 %r1094, %r1092, %r1093;
xor.b32 %r226, %r906, %r1094;
xor.b32 %r1095, %r223, %r888;
or.b32 %r1096, %r1095, -2139062144;
and.b32 %r1097, %r223, 2139062143;
sub.s32 %r1098, %r1096, %r1097;
xor.b32 %r230, %r911, %r1098;
ld.global.nc.u32 %r227, [%rd98+28];
ld.global.nc.u32 %r231, [%rd98+32];

dp4a.s32.s32 %r225, %r226, %r227, %r217;


dp4a.s32.s32 %r229, %r230, %r231, %r225;

mul.lo.s32 %r1099, %r918, %r229;

{ cvt.f32.f16 %f89, %rs1;}


ld.global.nc.u32 %r233, [%rd98];

{.reg .f16 low,high;
mov.b32 {low,high},%r233;
cvt.f32.f16 %f90, low;}


mul.ftz.f32 %f117, %f89, %f90;
cvt.rn.f32.s32 %f118, %r1099;
fma.rn.ftz.f32 %f320, %f117, %f118, %f320;
mov.u32 %r234, %r724;

prmt.b32 %r234, %r234, 0, 0xba98;

mov.u32 %r236, %r726;

prmt.b32 %r236, %r236, 0, 0xba98;

xor.b32 %r1100, %r234, %r924;
or.b32 %r1101, %r1100, -2139062144;
and.b32 %r1102, %r234, 2139062143;
sub.s32 %r1103, %r1101, %r1102;
xor.b32 %r239, %r948, %r1103;
xor.b32 %r1104, %r236, %r930;
or.b32 %r1105, %r1104, -2139062144;
and.b32 %r1106, %r236, 2139062143;
sub.s32 %r1107, %r1105, %r1106;
xor.b32 %r243, %r953, %r1107;

dp4a.s32.s32 %r238, %r239, %r191, %r731;


dp4a.s32.s32 %r242, %r243, %r195, %r238;

mov.u32 %r246, %r736;

prmt.b32 %r246, %r246, 0, 0xba98;

mov.u32 %r248, %r738;

prmt.b32 %r248, %r248, 0, 0xba98;

xor.b32 %r1108, %r246, %r959;
or.b32 %r1109, %r1108, -2139062144;
and.b32 %r1110, %r246, 2139062143;
sub.s32 %r1111, %r1109, %r1110;
xor.b32 %r251, %r983, %r1111;
xor.b32 %r1112, %r248, %r965;
or.b32 %r1113, %r1112, -2139062144;
and.b32 %r1114, %r248, 2139062143;
sub.s32 %r1115, %r1113, %r1114;
xor.b32 %r255, %r988, %r1115;

dp4a.s32.s32 %r250, %r251, %r203, %r242;


dp4a.s32.s32 %r254, %r255, %r207, %r250;

mov.u32 %r258, %r748;

prmt.b32 %r258, %r258, 0, 0xba98;

mov.u32 %r260, %r750;

prmt.b32 %r260, %r260, 0, 0xba98;

xor.b32 %r1116, %r258, %r994;
or.b32 %r1117, %r1116, -2139062144;
and.b32 %r1118, %r258, 2139062143;
sub.s32 %r1119, %r1117, %r1118;
xor.b32 %r263, %r1018, %r1119;
xor.b32 %r1120, %r260, %r1000;
or.b32 %r1121, %r1120, -2139062144;
and.b32 %r1122, %r260, 2139062143;
sub.s32 %r1123, %r1121, %r1122;
xor.b32 %r267, %r1023, %r1123;

dp4a.s32.s32 %r262, %r263, %r215, %r254;


dp4a.s32.s32 %r266, %r267, %r219, %r262;

mov.u32 %r270, %r760;

prmt.b32 %r270, %r270, 0, 0xba98;

mov.u32 %r272, %r762;

prmt.b32 %r272, %r272, 0, 0xba98;

xor.b32 %r1124, %r270, %r1029;
or.b32 %r1125, %r1124, -2139062144;
and.b32 %r1126, %r270, 2139062143;
sub.s32 %r1127, %r1125, %r1126;
xor.b32 %r275, %r1053, %r1127;
xor.b32 %r1128, %r272, %r1035;
or.b32 %r1129, %r1128, -2139062144;
and.b32 %r1130, %r272, 2139062143;
sub.s32 %r1131, %r1129, %r1130;
xor.b32 %r279, %r1058, %r1131;

dp4a.s32.s32 %r274, %r275, %r227, %r266;


dp4a.s32.s32 %r278, %r279, %r231, %r274;

mul.lo.s32 %r1132, %r1065, %r278;

{ cvt.f32.f16 %f91, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r233;
cvt.f32.f16 %f92, low;}


mul.ftz.f32 %f119, %f91, %f92;
cvt.rn.f32.s32 %f120, %r1132;
fma.rn.ftz.f32 %f319, %f119, %f120, %f319;
cvt.u64.u32 %rd99, %r1727;
add.s64 %rd100, %rd2, %rd99;
add.s64 %rd101, %rd100, %rd4;
mul.lo.s64 %rd102, %rd101, 36;
mov.u32 %r283, %r675;

prmt.b32 %r283, %r283, 0, 0xba98;

mov.u32 %r285, %r677;

prmt.b32 %r285, %r285, 0, 0xba98;

xor.b32 %r1133, %r283, %r777;
or.b32 %r1134, %r1133, -2139062144;
and.b32 %r1135, %r283, 2139062143;
sub.s32 %r1136, %r1134, %r1135;
xor.b32 %r288, %r801, %r1136;
xor.b32 %r1137, %r285, %r783;
or.b32 %r1138, %r1137, -2139062144;
and.b32 %r1139, %r285, 2139062143;
sub.s32 %r1140, %r1138, %r1139;
xor.b32 %r292, %r806, %r1140;
add.s64 %rd103, %rd17, %rd102;
ld.global.nc.u32 %r289, [%rd103+4];
ld.global.nc.u32 %r293, [%rd103+8];

dp4a.s32.s32 %r287, %r288, %r289, %r731;


dp4a.s32.s32 %r291, %r292, %r293, %r287;

mov.u32 %r295, %r687;

prmt.b32 %r295, %r295, 0, 0xba98;

mov.u32 %r297, %r689;

prmt.b32 %r297, %r297, 0, 0xba98;

xor.b32 %r1141, %r295, %r812;
or.b32 %r1142, %r1141, -2139062144;
and.b32 %r1143, %r295, 2139062143;
sub.s32 %r1144, %r1142, %r1143;
xor.b32 %r300, %r836, %r1144;
xor.b32 %r1145, %r297, %r818;
or.b32 %r1146, %r1145, -2139062144;
and.b32 %r1147, %r297, 2139062143;
sub.s32 %r1148, %r1146, %r1147;
xor.b32 %r304, %r841, %r1148;
ld.global.nc.u32 %r301, [%rd103+12];
ld.global.nc.u32 %r305, [%rd103+16];

dp4a.s32.s32 %r299, %r300, %r301, %r291;


dp4a.s32.s32 %r303, %r304, %r305, %r299;

mov.u32 %r307, %r699;

prmt.b32 %r307, %r307, 0, 0xba98;

mov.u32 %r309, %r701;

prmt.b32 %r309, %r309, 0, 0xba98;

xor.b32 %r1149, %r307, %r847;
or.b32 %r1150, %r1149, -2139062144;
and.b32 %r1151, %r307, 2139062143;
sub.s32 %r1152, %r1150, %r1151;
xor.b32 %r312, %r871, %r1152;
xor.b32 %r1153, %r309, %r853;
or.b32 %r1154, %r1153, -2139062144;
and.b32 %r1155, %r309, 2139062143;
sub.s32 %r1156, %r1154, %r1155;
xor.b32 %r316, %r876, %r1156;
ld.global.nc.u32 %r313, [%rd103+20];
ld.global.nc.u32 %r317, [%rd103+24];

dp4a.s32.s32 %r311, %r312, %r313, %r303;


dp4a.s32.s32 %r315, %r316, %r317, %r311;

mov.u32 %r319, %r711;

prmt.b32 %r319, %r319, 0, 0xba98;

mov.u32 %r321, %r713;

prmt.b32 %r321, %r321, 0, 0xba98;

xor.b32 %r1157, %r319, %r882;
or.b32 %r1158, %r1157, -2139062144;
and.b32 %r1159, %r319, 2139062143;
sub.s32 %r1160, %r1158, %r1159;
xor.b32 %r324, %r906, %r1160;
xor.b32 %r1161, %r321, %r888;
or.b32 %r1162, %r1161, -2139062144;
and.b32 %r1163, %r321, 2139062143;
sub.s32 %r1164, %r1162, %r1163;
xor.b32 %r328, %r911, %r1164;
ld.global.nc.u32 %r325, [%rd103+28];
ld.global.nc.u32 %r329, [%rd103+32];

dp4a.s32.s32 %r323, %r324, %r325, %r315;


dp4a.s32.s32 %r327, %r328, %r329, %r323;

mul.lo.s32 %r1165, %r918, %r327;

{ cvt.f32.f16 %f93, %rs1;}


ld.global.nc.u32 %r331, [%rd103];

{.reg .f16 low,high;
mov.b32 {low,high},%r331;
cvt.f32.f16 %f94, low;}


mul.ftz.f32 %f121, %f93, %f94;
cvt.rn.f32.s32 %f122, %r1165;
fma.rn.ftz.f32 %f318, %f121, %f122, %f318;
mov.u32 %r332, %r724;

prmt.b32 %r332, %r332, 0, 0xba98;

mov.u32 %r334, %r726;

prmt.b32 %r334, %r334, 0, 0xba98;

xor.b32 %r1166, %r332, %r924;
or.b32 %r1167, %r1166, -2139062144;
and.b32 %r1168, %r332, 2139062143;
sub.s32 %r1169, %r1167, %r1168;
xor.b32 %r337, %r948, %r1169;
xor.b32 %r1170, %r334, %r930;
or.b32 %r1171, %r1170, -2139062144;
and.b32 %r1172, %r334, 2139062143;
sub.s32 %r1173, %r1171, %r1172;
xor.b32 %r341, %r953, %r1173;

dp4a.s32.s32 %r336, %r337, %r289, %r731;


dp4a.s32.s32 %r340, %r341, %r293, %r336;

mov.u32 %r344, %r736;

prmt.b32 %r344, %r344, 0, 0xba98;

mov.u32 %r346, %r738;

prmt.b32 %r346, %r346, 0, 0xba98;

xor.b32 %r1174, %r344, %r959;
or.b32 %r1175, %r1174, -2139062144;
and.b32 %r1176, %r344, 2139062143;
sub.s32 %r1177, %r1175, %r1176;
xor.b32 %r349, %r983, %r1177;
xor.b32 %r1178, %r346, %r965;
or.b32 %r1179, %r1178, -2139062144;
and.b32 %r1180, %r346, 2139062143;
sub.s32 %r1181, %r1179, %r1180;
xor.b32 %r353, %r988, %r1181;

dp4a.s32.s32 %r348, %r349, %r301, %r340;


dp4a.s32.s32 %r352, %r353, %r305, %r348;

mov.u32 %r356, %r748;

prmt.b32 %r356, %r356, 0, 0xba98;

mov.u32 %r358, %r750;

prmt.b32 %r358, %r358, 0, 0xba98;

xor.b32 %r1182, %r356, %r994;
or.b32 %r1183, %r1182, -2139062144;
and.b32 %r1184, %r356, 2139062143;
sub.s32 %r1185, %r1183, %r1184;
xor.b32 %r361, %r1018, %r1185;
xor.b32 %r1186, %r358, %r1000;
or.b32 %r1187, %r1186, -2139062144;
and.b32 %r1188, %r358, 2139062143;
sub.s32 %r1189, %r1187, %r1188;
xor.b32 %r365, %r1023, %r1189;

dp4a.s32.s32 %r360, %r361, %r313, %r352;


dp4a.s32.s32 %r364, %r365, %r317, %r360;

mov.u32 %r368, %r760;

prmt.b32 %r368, %r368, 0, 0xba98;

mov.u32 %r370, %r762;

prmt.b32 %r370, %r370, 0, 0xba98;

xor.b32 %r1190, %r368, %r1029;
or.b32 %r1191, %r1190, -2139062144;
and.b32 %r1192, %r368, 2139062143;
sub.s32 %r1193, %r1191, %r1192;
xor.b32 %r373, %r1053, %r1193;
xor.b32 %r1194, %r370, %r1035;
or.b32 %r1195, %r1194, -2139062144;
and.b32 %r1196, %r370, 2139062143;
sub.s32 %r1197, %r1195, %r1196;
xor.b32 %r377, %r1058, %r1197;

dp4a.s32.s32 %r372, %r373, %r325, %r364;


dp4a.s32.s32 %r376, %r377, %r329, %r372;

mul.lo.s32 %r1198, %r1065, %r376;

{ cvt.f32.f16 %f95, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r331;
cvt.f32.f16 %f96, low;}


mul.ftz.f32 %f123, %f95, %f96;
cvt.rn.f32.s32 %f124, %r1198;
fma.rn.ftz.f32 %f317, %f123, %f124, %f317;
cvt.u64.u32 %rd104, %r1726;
add.s64 %rd105, %rd2, %rd104;
add.s64 %rd106, %rd105, %rd4;
mul.lo.s64 %rd107, %rd106, 36;
mov.u32 %r381, %r675;

prmt.b32 %r381, %r381, 0, 0xba98;

mov.u32 %r383, %r677;

prmt.b32 %r383, %r383, 0, 0xba98;

xor.b32 %r1199, %r381, %r777;
or.b32 %r1200, %r1199, -2139062144;
and.b32 %r1201, %r381, 2139062143;
sub.s32 %r1202, %r1200, %r1201;
xor.b32 %r386, %r801, %r1202;
xor.b32 %r1203, %r383, %r783;
or.b32 %r1204, %r1203, -2139062144;
and.b32 %r1205, %r383, 2139062143;
sub.s32 %r1206, %r1204, %r1205;
xor.b32 %r390, %r806, %r1206;
add.s64 %rd108, %rd17, %rd107;
ld.global.nc.u32 %r387, [%rd108+4];
ld.global.nc.u32 %r391, [%rd108+8];

dp4a.s32.s32 %r385, %r386, %r387, %r731;


dp4a.s32.s32 %r389, %r390, %r391, %r385;

mov.u32 %r393, %r687;

prmt.b32 %r393, %r393, 0, 0xba98;

mov.u32 %r395, %r689;

prmt.b32 %r395, %r395, 0, 0xba98;

xor.b32 %r1207, %r393, %r812;
or.b32 %r1208, %r1207, -2139062144;
and.b32 %r1209, %r393, 2139062143;
sub.s32 %r1210, %r1208, %r1209;
xor.b32 %r398, %r836, %r1210;
xor.b32 %r1211, %r395, %r818;
or.b32 %r1212, %r1211, -2139062144;
and.b32 %r1213, %r395, 2139062143;
sub.s32 %r1214, %r1212, %r1213;
xor.b32 %r402, %r841, %r1214;
ld.global.nc.u32 %r399, [%rd108+12];
ld.global.nc.u32 %r403, [%rd108+16];

dp4a.s32.s32 %r397, %r398, %r399, %r389;


dp4a.s32.s32 %r401, %r402, %r403, %r397;

mov.u32 %r405, %r699;

prmt.b32 %r405, %r405, 0, 0xba98;

mov.u32 %r407, %r701;

prmt.b32 %r407, %r407, 0, 0xba98;

xor.b32 %r1215, %r405, %r847;
or.b32 %r1216, %r1215, -2139062144;
and.b32 %r1217, %r405, 2139062143;
sub.s32 %r1218, %r1216, %r1217;
xor.b32 %r410, %r871, %r1218;
xor.b32 %r1219, %r407, %r853;
or.b32 %r1220, %r1219, -2139062144;
and.b32 %r1221, %r407, 2139062143;
sub.s32 %r1222, %r1220, %r1221;
xor.b32 %r414, %r876, %r1222;
ld.global.nc.u32 %r411, [%rd108+20];
ld.global.nc.u32 %r415, [%rd108+24];

dp4a.s32.s32 %r409, %r410, %r411, %r401;


dp4a.s32.s32 %r413, %r414, %r415, %r409;

mov.u32 %r417, %r711;

prmt.b32 %r417, %r417, 0, 0xba98;

mov.u32 %r419, %r713;

prmt.b32 %r419, %r419, 0, 0xba98;

xor.b32 %r1223, %r417, %r882;
or.b32 %r1224, %r1223, -2139062144;
and.b32 %r1225, %r417, 2139062143;
sub.s32 %r1226, %r1224, %r1225;
xor.b32 %r422, %r906, %r1226;
xor.b32 %r1227, %r419, %r888;
or.b32 %r1228, %r1227, -2139062144;
and.b32 %r1229, %r419, 2139062143;
sub.s32 %r1230, %r1228, %r1229;
xor.b32 %r426, %r911, %r1230;
ld.global.nc.u32 %r423, [%rd108+28];
ld.global.nc.u32 %r427, [%rd108+32];

dp4a.s32.s32 %r421, %r422, %r423, %r413;


dp4a.s32.s32 %r425, %r426, %r427, %r421;

mul.lo.s32 %r1231, %r918, %r425;

{ cvt.f32.f16 %f97, %rs1;}


ld.global.nc.u32 %r429, [%rd108];

{.reg .f16 low,high;
mov.b32 {low,high},%r429;
cvt.f32.f16 %f98, low;}


mul.ftz.f32 %f125, %f97, %f98;
cvt.rn.f32.s32 %f126, %r1231;
fma.rn.ftz.f32 %f316, %f125, %f126, %f316;
mov.u32 %r430, %r724;

prmt.b32 %r430, %r430, 0, 0xba98;

mov.u32 %r432, %r726;

prmt.b32 %r432, %r432, 0, 0xba98;

xor.b32 %r1232, %r430, %r924;
or.b32 %r1233, %r1232, -2139062144;
and.b32 %r1234, %r430, 2139062143;
sub.s32 %r1235, %r1233, %r1234;
xor.b32 %r435, %r948, %r1235;
xor.b32 %r1236, %r432, %r930;
or.b32 %r1237, %r1236, -2139062144;
and.b32 %r1238, %r432, 2139062143;
sub.s32 %r1239, %r1237, %r1238;
xor.b32 %r439, %r953, %r1239;

dp4a.s32.s32 %r434, %r435, %r387, %r731;


dp4a.s32.s32 %r438, %r439, %r391, %r434;

mov.u32 %r442, %r736;

prmt.b32 %r442, %r442, 0, 0xba98;

mov.u32 %r444, %r738;

prmt.b32 %r444, %r444, 0, 0xba98;

xor.b32 %r1240, %r442, %r959;
or.b32 %r1241, %r1240, -2139062144;
and.b32 %r1242, %r442, 2139062143;
sub.s32 %r1243, %r1241, %r1242;
xor.b32 %r447, %r983, %r1243;
xor.b32 %r1244, %r444, %r965;
or.b32 %r1245, %r1244, -2139062144;
and.b32 %r1246, %r444, 2139062143;
sub.s32 %r1247, %r1245, %r1246;
xor.b32 %r451, %r988, %r1247;

dp4a.s32.s32 %r446, %r447, %r399, %r438;


dp4a.s32.s32 %r450, %r451, %r403, %r446;

mov.u32 %r454, %r748;

prmt.b32 %r454, %r454, 0, 0xba98;

mov.u32 %r456, %r750;

prmt.b32 %r456, %r456, 0, 0xba98;

xor.b32 %r1248, %r454, %r994;
or.b32 %r1249, %r1248, -2139062144;
and.b32 %r1250, %r454, 2139062143;
sub.s32 %r1251, %r1249, %r1250;
xor.b32 %r459, %r1018, %r1251;
xor.b32 %r1252, %r456, %r1000;
or.b32 %r1253, %r1252, -2139062144;
and.b32 %r1254, %r456, 2139062143;
sub.s32 %r1255, %r1253, %r1254;
xor.b32 %r463, %r1023, %r1255;

dp4a.s32.s32 %r458, %r459, %r411, %r450;


dp4a.s32.s32 %r462, %r463, %r415, %r458;

mov.u32 %r466, %r760;

prmt.b32 %r466, %r466, 0, 0xba98;

mov.u32 %r468, %r762;

prmt.b32 %r468, %r468, 0, 0xba98;

xor.b32 %r1256, %r466, %r1029;
or.b32 %r1257, %r1256, -2139062144;
and.b32 %r1258, %r466, 2139062143;
sub.s32 %r1259, %r1257, %r1258;
xor.b32 %r471, %r1053, %r1259;
xor.b32 %r1260, %r468, %r1035;
or.b32 %r1261, %r1260, -2139062144;
and.b32 %r1262, %r468, 2139062143;
sub.s32 %r1263, %r1261, %r1262;
xor.b32 %r475, %r1058, %r1263;

dp4a.s32.s32 %r470, %r471, %r423, %r462;


dp4a.s32.s32 %r474, %r475, %r427, %r470;

mul.lo.s32 %r1264, %r1065, %r474;

{ cvt.f32.f16 %f99, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r429;
cvt.f32.f16 %f100, low;}


mul.ftz.f32 %f127, %f99, %f100;
cvt.rn.f32.s32 %f128, %r1264;
fma.rn.ftz.f32 %f315, %f127, %f128, %f315;
cvt.u64.u32 %rd109, %r1725;
add.s64 %rd110, %rd2, %rd109;
add.s64 %rd111, %rd110, %rd4;
mul.lo.s64 %rd112, %rd111, 36;
mov.u32 %r479, %r675;

prmt.b32 %r479, %r479, 0, 0xba98;

mov.u32 %r481, %r677;

prmt.b32 %r481, %r481, 0, 0xba98;

xor.b32 %r1265, %r479, %r777;
or.b32 %r1266, %r1265, -2139062144;
and.b32 %r1267, %r479, 2139062143;
sub.s32 %r1268, %r1266, %r1267;
xor.b32 %r484, %r801, %r1268;
xor.b32 %r1269, %r481, %r783;
or.b32 %r1270, %r1269, -2139062144;
and.b32 %r1271, %r481, 2139062143;
sub.s32 %r1272, %r1270, %r1271;
xor.b32 %r488, %r806, %r1272;
add.s64 %rd113, %rd17, %rd112;
ld.global.nc.u32 %r485, [%rd113+4];
ld.global.nc.u32 %r489, [%rd113+8];

dp4a.s32.s32 %r483, %r484, %r485, %r731;


dp4a.s32.s32 %r487, %r488, %r489, %r483;

mov.u32 %r491, %r687;

prmt.b32 %r491, %r491, 0, 0xba98;

mov.u32 %r493, %r689;

prmt.b32 %r493, %r493, 0, 0xba98;

xor.b32 %r1273, %r491, %r812;
or.b32 %r1274, %r1273, -2139062144;
and.b32 %r1275, %r491, 2139062143;
sub.s32 %r1276, %r1274, %r1275;
xor.b32 %r496, %r836, %r1276;
xor.b32 %r1277, %r493, %r818;
or.b32 %r1278, %r1277, -2139062144;
and.b32 %r1279, %r493, 2139062143;
sub.s32 %r1280, %r1278, %r1279;
xor.b32 %r500, %r841, %r1280;
ld.global.nc.u32 %r497, [%rd113+12];
ld.global.nc.u32 %r501, [%rd113+16];

dp4a.s32.s32 %r495, %r496, %r497, %r487;


dp4a.s32.s32 %r499, %r500, %r501, %r495;

mov.u32 %r503, %r699;

prmt.b32 %r503, %r503, 0, 0xba98;

mov.u32 %r505, %r701;

prmt.b32 %r505, %r505, 0, 0xba98;

xor.b32 %r1281, %r503, %r847;
or.b32 %r1282, %r1281, -2139062144;
and.b32 %r1283, %r503, 2139062143;
sub.s32 %r1284, %r1282, %r1283;
xor.b32 %r508, %r871, %r1284;
xor.b32 %r1285, %r505, %r853;
or.b32 %r1286, %r1285, -2139062144;
and.b32 %r1287, %r505, 2139062143;
sub.s32 %r1288, %r1286, %r1287;
xor.b32 %r512, %r876, %r1288;
ld.global.nc.u32 %r509, [%rd113+20];
ld.global.nc.u32 %r513, [%rd113+24];

dp4a.s32.s32 %r507, %r508, %r509, %r499;


dp4a.s32.s32 %r511, %r512, %r513, %r507;

mov.u32 %r515, %r711;

prmt.b32 %r515, %r515, 0, 0xba98;

mov.u32 %r517, %r713;

prmt.b32 %r517, %r517, 0, 0xba98;

xor.b32 %r1289, %r515, %r882;
or.b32 %r1290, %r1289, -2139062144;
and.b32 %r1291, %r515, 2139062143;
sub.s32 %r1292, %r1290, %r1291;
xor.b32 %r520, %r906, %r1292;
xor.b32 %r1293, %r517, %r888;
or.b32 %r1294, %r1293, -2139062144;
and.b32 %r1295, %r517, 2139062143;
sub.s32 %r1296, %r1294, %r1295;
xor.b32 %r524, %r911, %r1296;
ld.global.nc.u32 %r521, [%rd113+28];
ld.global.nc.u32 %r525, [%rd113+32];

dp4a.s32.s32 %r519, %r520, %r521, %r511;


dp4a.s32.s32 %r523, %r524, %r525, %r519;

mul.lo.s32 %r1297, %r918, %r523;

{ cvt.f32.f16 %f101, %rs1;}


ld.global.nc.u32 %r527, [%rd113];

{.reg .f16 low,high;
mov.b32 {low,high},%r527;
cvt.f32.f16 %f102, low;}


mul.ftz.f32 %f129, %f101, %f102;
cvt.rn.f32.s32 %f130, %r1297;
fma.rn.ftz.f32 %f314, %f129, %f130, %f314;
mov.u32 %r528, %r724;

prmt.b32 %r528, %r528, 0, 0xba98;

mov.u32 %r530, %r726;

prmt.b32 %r530, %r530, 0, 0xba98;

xor.b32 %r1298, %r528, %r924;
or.b32 %r1299, %r1298, -2139062144;
and.b32 %r1300, %r528, 2139062143;
sub.s32 %r1301, %r1299, %r1300;
xor.b32 %r533, %r948, %r1301;
xor.b32 %r1302, %r530, %r930;
or.b32 %r1303, %r1302, -2139062144;
and.b32 %r1304, %r530, 2139062143;
sub.s32 %r1305, %r1303, %r1304;
xor.b32 %r537, %r953, %r1305;

dp4a.s32.s32 %r532, %r533, %r485, %r731;


dp4a.s32.s32 %r536, %r537, %r489, %r532;

mov.u32 %r540, %r736;

prmt.b32 %r540, %r540, 0, 0xba98;

mov.u32 %r542, %r738;

prmt.b32 %r542, %r542, 0, 0xba98;

xor.b32 %r1306, %r540, %r959;
or.b32 %r1307, %r1306, -2139062144;
and.b32 %r1308, %r540, 2139062143;
sub.s32 %r1309, %r1307, %r1308;
xor.b32 %r545, %r983, %r1309;
xor.b32 %r1310, %r542, %r965;
or.b32 %r1311, %r1310, -2139062144;
and.b32 %r1312, %r542, 2139062143;
sub.s32 %r1313, %r1311, %r1312;
xor.b32 %r549, %r988, %r1313;

dp4a.s32.s32 %r544, %r545, %r497, %r536;


dp4a.s32.s32 %r548, %r549, %r501, %r544;

mov.u32 %r552, %r748;

prmt.b32 %r552, %r552, 0, 0xba98;

mov.u32 %r554, %r750;

prmt.b32 %r554, %r554, 0, 0xba98;

xor.b32 %r1314, %r552, %r994;
or.b32 %r1315, %r1314, -2139062144;
and.b32 %r1316, %r552, 2139062143;
sub.s32 %r1317, %r1315, %r1316;
xor.b32 %r557, %r1018, %r1317;
xor.b32 %r1318, %r554, %r1000;
or.b32 %r1319, %r1318, -2139062144;
and.b32 %r1320, %r554, 2139062143;
sub.s32 %r1321, %r1319, %r1320;
xor.b32 %r561, %r1023, %r1321;

dp4a.s32.s32 %r556, %r557, %r509, %r548;


dp4a.s32.s32 %r560, %r561, %r513, %r556;

mov.u32 %r564, %r760;

prmt.b32 %r564, %r564, 0, 0xba98;

mov.u32 %r566, %r762;

prmt.b32 %r566, %r566, 0, 0xba98;

xor.b32 %r1322, %r564, %r1029;
or.b32 %r1323, %r1322, -2139062144;
and.b32 %r1324, %r564, 2139062143;
sub.s32 %r1325, %r1323, %r1324;
xor.b32 %r569, %r1053, %r1325;
xor.b32 %r1326, %r566, %r1035;
or.b32 %r1327, %r1326, -2139062144;
and.b32 %r1328, %r566, 2139062143;
sub.s32 %r1329, %r1327, %r1328;
xor.b32 %r573, %r1058, %r1329;

dp4a.s32.s32 %r568, %r569, %r521, %r560;


dp4a.s32.s32 %r572, %r573, %r525, %r568;

mul.lo.s32 %r1330, %r1065, %r572;

{ cvt.f32.f16 %f103, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r527;
cvt.f32.f16 %f104, low;}


mul.ftz.f32 %f131, %f103, %f104;
cvt.rn.f32.s32 %f132, %r1330;
fma.rn.ftz.f32 %f313, %f131, %f132, %f313;
cvt.u64.u32 %rd114, %r1724;
add.s64 %rd115, %rd2, %rd114;
add.s64 %rd116, %rd115, %rd4;
mul.lo.s64 %rd117, %rd116, 36;
mov.u32 %r577, %r675;

prmt.b32 %r577, %r577, 0, 0xba98;

mov.u32 %r579, %r677;

prmt.b32 %r579, %r579, 0, 0xba98;

xor.b32 %r1331, %r577, %r777;
or.b32 %r1332, %r1331, -2139062144;
and.b32 %r1333, %r577, 2139062143;
sub.s32 %r1334, %r1332, %r1333;
xor.b32 %r582, %r801, %r1334;
xor.b32 %r1335, %r579, %r783;
or.b32 %r1336, %r1335, -2139062144;
and.b32 %r1337, %r579, 2139062143;
sub.s32 %r1338, %r1336, %r1337;
xor.b32 %r586, %r806, %r1338;
add.s64 %rd118, %rd17, %rd117;
ld.global.nc.u32 %r583, [%rd118+4];
ld.global.nc.u32 %r587, [%rd118+8];

dp4a.s32.s32 %r581, %r582, %r583, %r731;


dp4a.s32.s32 %r585, %r586, %r587, %r581;

mov.u32 %r589, %r687;

prmt.b32 %r589, %r589, 0, 0xba98;

mov.u32 %r591, %r689;

prmt.b32 %r591, %r591, 0, 0xba98;

xor.b32 %r1339, %r589, %r812;
or.b32 %r1340, %r1339, -2139062144;
and.b32 %r1341, %r589, 2139062143;
sub.s32 %r1342, %r1340, %r1341;
xor.b32 %r594, %r836, %r1342;
xor.b32 %r1343, %r591, %r818;
or.b32 %r1344, %r1343, -2139062144;
and.b32 %r1345, %r591, 2139062143;
sub.s32 %r1346, %r1344, %r1345;
xor.b32 %r598, %r841, %r1346;
ld.global.nc.u32 %r595, [%rd118+12];
ld.global.nc.u32 %r599, [%rd118+16];

dp4a.s32.s32 %r593, %r594, %r595, %r585;


dp4a.s32.s32 %r597, %r598, %r599, %r593;

mov.u32 %r601, %r699;

prmt.b32 %r601, %r601, 0, 0xba98;

mov.u32 %r603, %r701;

prmt.b32 %r603, %r603, 0, 0xba98;

xor.b32 %r1347, %r601, %r847;
or.b32 %r1348, %r1347, -2139062144;
and.b32 %r1349, %r601, 2139062143;
sub.s32 %r1350, %r1348, %r1349;
xor.b32 %r606, %r871, %r1350;
xor.b32 %r1351, %r603, %r853;
or.b32 %r1352, %r1351, -2139062144;
and.b32 %r1353, %r603, 2139062143;
sub.s32 %r1354, %r1352, %r1353;
xor.b32 %r610, %r876, %r1354;
ld.global.nc.u32 %r607, [%rd118+20];
ld.global.nc.u32 %r611, [%rd118+24];

dp4a.s32.s32 %r605, %r606, %r607, %r597;


dp4a.s32.s32 %r609, %r610, %r611, %r605;

mov.u32 %r613, %r711;

prmt.b32 %r613, %r613, 0, 0xba98;

mov.u32 %r615, %r713;

prmt.b32 %r615, %r615, 0, 0xba98;

xor.b32 %r1355, %r613, %r882;
or.b32 %r1356, %r1355, -2139062144;
and.b32 %r1357, %r613, 2139062143;
sub.s32 %r1358, %r1356, %r1357;
xor.b32 %r618, %r906, %r1358;
xor.b32 %r1359, %r615, %r888;
or.b32 %r1360, %r1359, -2139062144;
and.b32 %r1361, %r615, 2139062143;
sub.s32 %r1362, %r1360, %r1361;
xor.b32 %r622, %r911, %r1362;
ld.global.nc.u32 %r619, [%rd118+28];
ld.global.nc.u32 %r623, [%rd118+32];

dp4a.s32.s32 %r617, %r618, %r619, %r609;


dp4a.s32.s32 %r621, %r622, %r623, %r617;

mul.lo.s32 %r1363, %r918, %r621;

{ cvt.f32.f16 %f105, %rs1;}


ld.global.nc.u32 %r625, [%rd118];

{.reg .f16 low,high;
mov.b32 {low,high},%r625;
cvt.f32.f16 %f106, low;}


mul.ftz.f32 %f133, %f105, %f106;
cvt.rn.f32.s32 %f134, %r1363;
fma.rn.ftz.f32 %f312, %f133, %f134, %f312;
mov.u32 %r626, %r724;

prmt.b32 %r626, %r626, 0, 0xba98;

mov.u32 %r628, %r726;

prmt.b32 %r628, %r628, 0, 0xba98;

xor.b32 %r1364, %r626, %r924;
or.b32 %r1365, %r1364, -2139062144;
and.b32 %r1366, %r626, 2139062143;
sub.s32 %r1367, %r1365, %r1366;
xor.b32 %r631, %r948, %r1367;
xor.b32 %r1368, %r628, %r930;
or.b32 %r1369, %r1368, -2139062144;
and.b32 %r1370, %r628, 2139062143;
sub.s32 %r1371, %r1369, %r1370;
xor.b32 %r635, %r953, %r1371;

dp4a.s32.s32 %r630, %r631, %r583, %r731;


dp4a.s32.s32 %r634, %r635, %r587, %r630;

mov.u32 %r638, %r736;

prmt.b32 %r638, %r638, 0, 0xba98;

mov.u32 %r640, %r738;

prmt.b32 %r640, %r640, 0, 0xba98;

xor.b32 %r1372, %r638, %r959;
or.b32 %r1373, %r1372, -2139062144;
and.b32 %r1374, %r638, 2139062143;
sub.s32 %r1375, %r1373, %r1374;
xor.b32 %r643, %r983, %r1375;
xor.b32 %r1376, %r640, %r965;
or.b32 %r1377, %r1376, -2139062144;
and.b32 %r1378, %r640, 2139062143;
sub.s32 %r1379, %r1377, %r1378;
xor.b32 %r647, %r988, %r1379;

dp4a.s32.s32 %r642, %r643, %r595, %r634;


dp4a.s32.s32 %r646, %r647, %r599, %r642;

mov.u32 %r650, %r748;

prmt.b32 %r650, %r650, 0, 0xba98;

mov.u32 %r652, %r750;

prmt.b32 %r652, %r652, 0, 0xba98;

xor.b32 %r1380, %r650, %r994;
or.b32 %r1381, %r1380, -2139062144;
and.b32 %r1382, %r650, 2139062143;
sub.s32 %r1383, %r1381, %r1382;
xor.b32 %r655, %r1018, %r1383;
xor.b32 %r1384, %r652, %r1000;
or.b32 %r1385, %r1384, -2139062144;
and.b32 %r1386, %r652, 2139062143;
sub.s32 %r1387, %r1385, %r1386;
xor.b32 %r659, %r1023, %r1387;

dp4a.s32.s32 %r654, %r655, %r607, %r646;


dp4a.s32.s32 %r658, %r659, %r611, %r654;

mov.u32 %r662, %r760;

prmt.b32 %r662, %r662, 0, 0xba98;

mov.u32 %r664, %r762;

prmt.b32 %r664, %r664, 0, 0xba98;

xor.b32 %r1388, %r662, %r1029;
or.b32 %r1389, %r1388, -2139062144;
and.b32 %r1390, %r662, 2139062143;
sub.s32 %r1391, %r1389, %r1390;
xor.b32 %r667, %r1053, %r1391;
xor.b32 %r1392, %r664, %r1035;
or.b32 %r1393, %r1392, -2139062144;
and.b32 %r1394, %r664, 2139062143;
sub.s32 %r1395, %r1393, %r1394;
xor.b32 %r671, %r1058, %r1395;

dp4a.s32.s32 %r666, %r667, %r619, %r658;


dp4a.s32.s32 %r670, %r671, %r623, %r666;

mul.lo.s32 %r1396, %r1065, %r670;

{ cvt.f32.f16 %f107, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r625;
cvt.f32.f16 %f108, low;}


mul.ftz.f32 %f135, %f107, %f108;
cvt.rn.f32.s32 %f136, %r1396;
fma.rn.ftz.f32 %f311, %f135, %f136, %f311;
cvt.u64.u32 %rd119, %r1723;
add.s64 %rd120, %rd2, %rd119;
add.s64 %rd121, %rd120, %rd4;
mul.lo.s64 %rd122, %rd121, 36;

prmt.b32 %r675, %r675, 0, 0xba98;


prmt.b32 %r677, %r677, 0, 0xba98;

xor.b32 %r1397, %r675, %r777;
or.b32 %r1398, %r1397, -2139062144;
and.b32 %r1399, %r675, 2139062143;
sub.s32 %r1400, %r1398, %r1399;
xor.b32 %r680, %r801, %r1400;
xor.b32 %r1401, %r677, %r783;
or.b32 %r1402, %r1401, -2139062144;
and.b32 %r1403, %r677, 2139062143;
sub.s32 %r1404, %r1402, %r1403;
xor.b32 %r684, %r806, %r1404;
add.s64 %rd123, %rd17, %rd122;
ld.global.nc.u32 %r681, [%rd123+4];
ld.global.nc.u32 %r685, [%rd123+8];

dp4a.s32.s32 %r679, %r680, %r681, %r731;


dp4a.s32.s32 %r683, %r684, %r685, %r679;


prmt.b32 %r687, %r687, 0, 0xba98;


prmt.b32 %r689, %r689, 0, 0xba98;

xor.b32 %r1405, %r687, %r812;
or.b32 %r1406, %r1405, -2139062144;
and.b32 %r1407, %r687, 2139062143;
sub.s32 %r1408, %r1406, %r1407;
xor.b32 %r692, %r836, %r1408;
xor.b32 %r1409, %r689, %r818;
or.b32 %r1410, %r1409, -2139062144;
and.b32 %r1411, %r689, 2139062143;
sub.s32 %r1412, %r1410, %r1411;
xor.b32 %r696, %r841, %r1412;
ld.global.nc.u32 %r693, [%rd123+12];
ld.global.nc.u32 %r697, [%rd123+16];

dp4a.s32.s32 %r691, %r692, %r693, %r683;


dp4a.s32.s32 %r695, %r696, %r697, %r691;


prmt.b32 %r699, %r699, 0, 0xba98;


prmt.b32 %r701, %r701, 0, 0xba98;

xor.b32 %r1413, %r699, %r847;
or.b32 %r1414, %r1413, -2139062144;
and.b32 %r1415, %r699, 2139062143;
sub.s32 %r1416, %r1414, %r1415;
xor.b32 %r704, %r871, %r1416;
xor.b32 %r1417, %r701, %r853;
or.b32 %r1418, %r1417, -2139062144;
and.b32 %r1419, %r701, 2139062143;
sub.s32 %r1420, %r1418, %r1419;
xor.b32 %r708, %r876, %r1420;
ld.global.nc.u32 %r705, [%rd123+20];
ld.global.nc.u32 %r709, [%rd123+24];

dp4a.s32.s32 %r703, %r704, %r705, %r695;


dp4a.s32.s32 %r707, %r708, %r709, %r703;


prmt.b32 %r711, %r711, 0, 0xba98;


prmt.b32 %r713, %r713, 0, 0xba98;

xor.b32 %r1421, %r711, %r882;
or.b32 %r1422, %r1421, -2139062144;
and.b32 %r1423, %r711, 2139062143;
sub.s32 %r1424, %r1422, %r1423;
xor.b32 %r716, %r906, %r1424;
xor.b32 %r1425, %r713, %r888;
or.b32 %r1426, %r1425, -2139062144;
and.b32 %r1427, %r713, 2139062143;
sub.s32 %r1428, %r1426, %r1427;
xor.b32 %r720, %r911, %r1428;
ld.global.nc.u32 %r717, [%rd123+28];
ld.global.nc.u32 %r721, [%rd123+32];

dp4a.s32.s32 %r715, %r716, %r717, %r707;


dp4a.s32.s32 %r719, %r720, %r721, %r715;

mul.lo.s32 %r1429, %r918, %r719;

{ cvt.f32.f16 %f109, %rs1;}


ld.global.nc.u32 %r723, [%rd123];

{.reg .f16 low,high;
mov.b32 {low,high},%r723;
cvt.f32.f16 %f110, low;}


mul.ftz.f32 %f137, %f109, %f110;
cvt.rn.f32.s32 %f138, %r1429;
fma.rn.ftz.f32 %f310, %f137, %f138, %f310;

prmt.b32 %r724, %r724, 0, 0xba98;


prmt.b32 %r726, %r726, 0, 0xba98;

xor.b32 %r1430, %r724, %r924;
or.b32 %r1431, %r1430, -2139062144;
and.b32 %r1432, %r724, 2139062143;
sub.s32 %r1433, %r1431, %r1432;
xor.b32 %r729, %r948, %r1433;
xor.b32 %r1434, %r726, %r930;
or.b32 %r1435, %r1434, -2139062144;
and.b32 %r1436, %r726, 2139062143;
sub.s32 %r1437, %r1435, %r1436;
xor.b32 %r733, %r953, %r1437;

dp4a.s32.s32 %r728, %r729, %r681, %r731;


dp4a.s32.s32 %r732, %r733, %r685, %r728;


prmt.b32 %r736, %r736, 0, 0xba98;


prmt.b32 %r738, %r738, 0, 0xba98;

xor.b32 %r1438, %r736, %r959;
or.b32 %r1439, %r1438, -2139062144;
and.b32 %r1440, %r736, 2139062143;
sub.s32 %r1441, %r1439, %r1440;
xor.b32 %r741, %r983, %r1441;
xor.b32 %r1442, %r738, %r965;
or.b32 %r1443, %r1442, -2139062144;
and.b32 %r1444, %r738, 2139062143;
sub.s32 %r1445, %r1443, %r1444;
xor.b32 %r745, %r988, %r1445;

dp4a.s32.s32 %r740, %r741, %r693, %r732;


dp4a.s32.s32 %r744, %r745, %r697, %r740;


prmt.b32 %r748, %r748, 0, 0xba98;


prmt.b32 %r750, %r750, 0, 0xba98;

xor.b32 %r1446, %r748, %r994;
or.b32 %r1447, %r1446, -2139062144;
and.b32 %r1448, %r748, 2139062143;
sub.s32 %r1449, %r1447, %r1448;
xor.b32 %r753, %r1018, %r1449;
xor.b32 %r1450, %r750, %r1000;
or.b32 %r1451, %r1450, -2139062144;
and.b32 %r1452, %r750, 2139062143;
sub.s32 %r1453, %r1451, %r1452;
xor.b32 %r757, %r1023, %r1453;

dp4a.s32.s32 %r752, %r753, %r705, %r744;


dp4a.s32.s32 %r756, %r757, %r709, %r752;


prmt.b32 %r760, %r760, 0, 0xba98;


prmt.b32 %r762, %r762, 0, 0xba98;

xor.b32 %r1454, %r760, %r1029;
or.b32 %r1455, %r1454, -2139062144;
and.b32 %r1456, %r760, 2139062143;
sub.s32 %r1457, %r1455, %r1456;
xor.b32 %r765, %r1053, %r1457;
xor.b32 %r1458, %r762, %r1035;
or.b32 %r1459, %r1458, -2139062144;
and.b32 %r1460, %r762, 2139062143;
sub.s32 %r1461, %r1459, %r1460;
xor.b32 %r769, %r1058, %r1461;

dp4a.s32.s32 %r764, %r765, %r717, %r756;


dp4a.s32.s32 %r768, %r769, %r721, %r764;

mul.lo.s32 %r1462, %r1065, %r768;

{ cvt.f32.f16 %f111, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r723;
cvt.f32.f16 %f112, low;}


mul.ftz.f32 %f139, %f111, %f112;
cvt.rn.f32.s32 %f140, %r1462;
fma.rn.ftz.f32 %f309, %f139, %f140, %f309;
add.s32 %r1727, %r1727, 64;
add.s32 %r1726, %r1726, 64;
add.s32 %r1725, %r1725, 64;
add.s32 %r1724, %r1724, 64;
add.s32 %r1723, %r1723, 64;
add.s32 %r1722, %r1722, 64;
add.s32 %r1721, %r1721, 64;
add.s64 %rd158, %rd158, 880;
add.s32 %r1728, %r1728, 8;
setp.lt.s32 %p4, %r1728, %r5;
@%p4 bra $L__BB315_2;

$L__BB315_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB315_5;

mov.u32 %r1464, _ZZ13mul_mat_vec_qIL9ggml_type21ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
mad.lo.s32 %r1465, %r1, 1792, %r1464;
shl.b32 %r1466, %r50, 2;
add.s32 %r1467, %r1465, %r1466;
st.shared.f32 [%r1467+-1792], %f322;
st.shared.f32 [%r1467+-1664], %f321;
st.shared.f32 [%r1467+-1536], %f320;
st.shared.f32 [%r1467+-1408], %f319;
st.shared.f32 [%r1467+-1280], %f318;
st.shared.f32 [%r1467+-1152], %f317;
st.shared.f32 [%r1467+-1024], %f316;
st.shared.f32 [%r1467+-896], %f315;
st.shared.f32 [%r1467+-768], %f314;
st.shared.f32 [%r1467+-640], %f313;
st.shared.f32 [%r1467+-512], %f312;
st.shared.f32 [%r1467+-384], %f311;
st.shared.f32 [%r1467+-256], %f310;
st.shared.f32 [%r1467+-128], %f309;

$L__BB315_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB315_34;

mov.u32 %r1468, %ctaid.z;
mov.u32 %r1469, %ctaid.y;
mad.lo.s32 %r1470, %r1469, %r41, %r3;
mad.lo.s32 %r1471, %r1468, %r47, %r1470;
cvt.u64.u32 %rd20, %r1471;
shl.b32 %r1473, %r50, 2;
mov.u32 %r1474, _ZZ13mul_mat_vec_qIL9ggml_type21ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r30, %r1474, %r1473;
mov.u32 %r1475, 2;
ld.shared.f32 %f141, [%r30];
add.ftz.f32 %f142, %f141, %f322;
mov.b32 %r1476, %f142;
mov.u32 %r1477, 31;
mov.u32 %r1478, 16;
mov.u32 %r1479, -1;
shfl.sync.bfly.b32 %r1480|%p7, %r1476, %r1478, %r1477, %r1479;
mov.b32 %f143, %r1480;
add.ftz.f32 %f144, %f142, %f143;
mov.b32 %r1481, %f144;
mov.u32 %r1482, 8;
shfl.sync.bfly.b32 %r1483|%p8, %r1481, %r1482, %r1477, %r1479;
mov.b32 %f145, %r1483;
add.ftz.f32 %f146, %f144, %f145;
mov.b32 %r1484, %f146;
mov.u32 %r1485, 4;
shfl.sync.bfly.b32 %r1486|%p9, %r1484, %r1485, %r1477, %r1479;
mov.b32 %f147, %r1486;
add.ftz.f32 %f148, %f146, %f147;
mov.b32 %r1487, %f148;
shfl.sync.bfly.b32 %r1488|%p10, %r1487, %r1475, %r1477, %r1479;
mov.b32 %f149, %r1488;
add.ftz.f32 %f150, %f148, %f149;
mov.b32 %r1489, %f150;
mov.u32 %r1490, 1;
shfl.sync.bfly.b32 %r1491|%p11, %r1489, %r1490, %r1477, %r1479;
mov.b32 %f151, %r1491;
add.ftz.f32 %f43, %f150, %f151;
setp.eq.s32 %p12, %r50, 0;
setp.lt.u32 %p13, %r3, %r35;
and.pred %p1, %p12, %p13;
mul.wide.u32 %rd124, %r1471, 4;
add.s64 %rd21, %rd1, %rd124;
not.pred %p14, %p1;
@%p14 bra $L__BB315_8;

st.global.f32 [%rd21], %f43;

$L__BB315_8:
ld.shared.f32 %f152, [%r30+128];
add.ftz.f32 %f153, %f152, %f321;
mov.b32 %r1492, %f153;
shfl.sync.bfly.b32 %r1496|%p15, %r1492, %r1478, %r1477, %r1479;
mov.b32 %f154, %r1496;
add.ftz.f32 %f155, %f153, %f154;
mov.b32 %r1497, %f155;
shfl.sync.bfly.b32 %r1499|%p16, %r1497, %r1482, %r1477, %r1479;
mov.b32 %f156, %r1499;
add.ftz.f32 %f157, %f155, %f156;
mov.b32 %r1500, %f157;
shfl.sync.bfly.b32 %r1502|%p17, %r1500, %r1485, %r1477, %r1479;
mov.b32 %f158, %r1502;
add.ftz.f32 %f159, %f157, %f158;
mov.b32 %r1503, %f159;
shfl.sync.bfly.b32 %r1505|%p18, %r1503, %r1475, %r1477, %r1479;
mov.b32 %f160, %r1505;
add.ftz.f32 %f161, %f159, %f160;
mov.b32 %r1506, %f161;
shfl.sync.bfly.b32 %r1508|%p19, %r1506, %r1490, %r1477, %r1479;
mov.b32 %f162, %r1508;
add.ftz.f32 %f44, %f161, %f162;
setp.eq.s32 %p20, %r50, 1;
add.s32 %r1510, %r3, 1;
setp.lt.u32 %p21, %r1510, %r35;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB315_10;

st.global.f32 [%rd21+4], %f44;

$L__BB315_10:
ld.shared.f32 %f163, [%r30+256];
add.ftz.f32 %f164, %f163, %f320;
mov.b32 %r1511, %f164;
mov.u32 %r1512, 31;
mov.u32 %r1513, 16;
mov.u32 %r1514, -1;
shfl.sync.bfly.b32 %r1515|%p23, %r1511, %r1513, %r1512, %r1514;
mov.b32 %f165, %r1515;
add.ftz.f32 %f166, %f164, %f165;
mov.b32 %r1516, %f166;
mov.u32 %r1517, 8;
shfl.sync.bfly.b32 %r1518|%p24, %r1516, %r1517, %r1512, %r1514;
mov.b32 %f167, %r1518;
add.ftz.f32 %f168, %f166, %f167;
mov.b32 %r1519, %f168;
mov.u32 %r1520, 4;
shfl.sync.bfly.b32 %r1521|%p25, %r1519, %r1520, %r1512, %r1514;
mov.b32 %f169, %r1521;
add.ftz.f32 %f170, %f168, %f169;
mov.b32 %r1522, %f170;
mov.u32 %r1523, 2;
shfl.sync.bfly.b32 %r1524|%p26, %r1522, %r1523, %r1512, %r1514;
mov.b32 %f171, %r1524;
add.ftz.f32 %f172, %f170, %f171;
mov.b32 %r1525, %f172;
mov.u32 %r1526, 1;
shfl.sync.bfly.b32 %r1527|%p27, %r1525, %r1526, %r1512, %r1514;
mov.b32 %f173, %r1527;
add.ftz.f32 %f45, %f172, %f173;
@%p14 bra $L__BB315_12;

cvt.u64.u32 %rd125, %r35;
add.s64 %rd126, %rd125, %rd20;
shl.b64 %rd127, %rd126, 2;
add.s64 %rd128, %rd1, %rd127;
st.global.f32 [%rd128], %f45;

$L__BB315_12:
ld.shared.f32 %f174, [%r30+384];
add.ftz.f32 %f175, %f174, %f319;
mov.b32 %r1528, %f175;
shfl.sync.bfly.b32 %r1532|%p29, %r1528, %r1513, %r1512, %r1514;
mov.b32 %f176, %r1532;
add.ftz.f32 %f177, %f175, %f176;
mov.b32 %r1533, %f177;
shfl.sync.bfly.b32 %r1535|%p30, %r1533, %r1517, %r1512, %r1514;
mov.b32 %f178, %r1535;
add.ftz.f32 %f179, %f177, %f178;
mov.b32 %r1536, %f179;
shfl.sync.bfly.b32 %r1538|%p31, %r1536, %r1520, %r1512, %r1514;
mov.b32 %f180, %r1538;
add.ftz.f32 %f181, %f179, %f180;
mov.b32 %r1539, %f181;
shfl.sync.bfly.b32 %r1541|%p32, %r1539, %r1523, %r1512, %r1514;
mov.b32 %f182, %r1541;
add.ftz.f32 %f183, %f181, %f182;
mov.b32 %r1542, %f183;
shfl.sync.bfly.b32 %r1544|%p33, %r1542, %r1526, %r1512, %r1514;
mov.b32 %f184, %r1544;
add.ftz.f32 %f46, %f183, %f184;
@%p22 bra $L__BB315_14;

add.s32 %r1545, %r35, 1;
cvt.u64.u32 %rd129, %r1545;
add.s64 %rd130, %rd129, %rd20;
shl.b64 %rd131, %rd130, 2;
add.s64 %rd132, %rd1, %rd131;
st.global.f32 [%rd132], %f46;

$L__BB315_14:
shl.b32 %r1546, %r35, 1;
mov.u32 %r1547, 1;
ld.shared.f32 %f185, [%r30+512];
add.ftz.f32 %f186, %f185, %f318;
mov.b32 %r1548, %f186;
mov.u32 %r1549, 31;
mov.u32 %r1550, 16;
mov.u32 %r1551, -1;
shfl.sync.bfly.b32 %r1552|%p35, %r1548, %r1550, %r1549, %r1551;
mov.b32 %f187, %r1552;
add.ftz.f32 %f188, %f186, %f187;
mov.b32 %r1553, %f188;
mov.u32 %r1554, 8;
shfl.sync.bfly.b32 %r1555|%p36, %r1553, %r1554, %r1549, %r1551;
mov.b32 %f189, %r1555;
add.ftz.f32 %f190, %f188, %f189;
mov.b32 %r1556, %f190;
mov.u32 %r1557, 4;
shfl.sync.bfly.b32 %r1558|%p37, %r1556, %r1557, %r1549, %r1551;
mov.b32 %f191, %r1558;
add.ftz.f32 %f192, %f190, %f191;
mov.b32 %r1559, %f192;
mov.u32 %r1560, 2;
shfl.sync.bfly.b32 %r1561|%p38, %r1559, %r1560, %r1549, %r1551;
mov.b32 %f193, %r1561;
add.ftz.f32 %f194, %f192, %f193;
mov.b32 %r1562, %f194;
shfl.sync.bfly.b32 %r1563|%p39, %r1562, %r1547, %r1549, %r1551;
mov.b32 %f195, %r1563;
add.ftz.f32 %f47, %f194, %f195;
cvt.u64.u32 %rd133, %r1546;
add.s64 %rd134, %rd133, %rd20;
shl.b64 %rd135, %rd134, 2;
add.s64 %rd22, %rd1, %rd135;
@%p14 bra $L__BB315_16;

st.global.f32 [%rd22], %f47;

$L__BB315_16:
ld.shared.f32 %f196, [%r30+640];
add.ftz.f32 %f197, %f196, %f317;
mov.b32 %r1564, %f197;
shfl.sync.bfly.b32 %r1568|%p41, %r1564, %r1550, %r1549, %r1551;
mov.b32 %f198, %r1568;
add.ftz.f32 %f199, %f197, %f198;
mov.b32 %r1569, %f199;
shfl.sync.bfly.b32 %r1571|%p42, %r1569, %r1554, %r1549, %r1551;
mov.b32 %f200, %r1571;
add.ftz.f32 %f201, %f199, %f200;
mov.b32 %r1572, %f201;
shfl.sync.bfly.b32 %r1574|%p43, %r1572, %r1557, %r1549, %r1551;
mov.b32 %f202, %r1574;
add.ftz.f32 %f203, %f201, %f202;
mov.b32 %r1575, %f203;
shfl.sync.bfly.b32 %r1577|%p44, %r1575, %r1560, %r1549, %r1551;
mov.b32 %f204, %r1577;
add.ftz.f32 %f205, %f203, %f204;
mov.b32 %r1578, %f205;
shfl.sync.bfly.b32 %r1580|%p45, %r1578, %r1547, %r1549, %r1551;
mov.b32 %f206, %r1580;
add.ftz.f32 %f48, %f205, %f206;
@%p22 bra $L__BB315_18;

st.global.f32 [%rd22+4], %f48;

$L__BB315_18:
mul.lo.s32 %r31, %r35, 3;
ld.shared.f32 %f207, [%r30+768];
add.ftz.f32 %f208, %f207, %f316;
mov.b32 %r1581, %f208;
mov.u32 %r1582, 31;
mov.u32 %r1583, 16;
mov.u32 %r1584, -1;
shfl.sync.bfly.b32 %r1585|%p47, %r1581, %r1583, %r1582, %r1584;
mov.b32 %f209, %r1585;
add.ftz.f32 %f210, %f208, %f209;
mov.b32 %r1586, %f210;
mov.u32 %r1587, 8;
shfl.sync.bfly.b32 %r1588|%p48, %r1586, %r1587, %r1582, %r1584;
mov.b32 %f211, %r1588;
add.ftz.f32 %f212, %f210, %f211;
mov.b32 %r1589, %f212;
mov.u32 %r1590, 4;
shfl.sync.bfly.b32 %r1591|%p49, %r1589, %r1590, %r1582, %r1584;
mov.b32 %f213, %r1591;
add.ftz.f32 %f214, %f212, %f213;
mov.b32 %r1592, %f214;
mov.u32 %r1593, 2;
shfl.sync.bfly.b32 %r1594|%p50, %r1592, %r1593, %r1582, %r1584;
mov.b32 %f215, %r1594;
add.ftz.f32 %f216, %f214, %f215;
mov.b32 %r1595, %f216;
mov.u32 %r1596, 1;
shfl.sync.bfly.b32 %r1597|%p51, %r1595, %r1596, %r1582, %r1584;
mov.b32 %f217, %r1597;
add.ftz.f32 %f49, %f216, %f217;
@%p14 bra $L__BB315_20;

cvt.u64.u32 %rd136, %r31;
add.s64 %rd137, %rd136, %rd20;
shl.b64 %rd138, %rd137, 2;
add.s64 %rd139, %rd1, %rd138;
st.global.f32 [%rd139], %f49;

$L__BB315_20:
ld.shared.f32 %f218, [%r30+896];
add.ftz.f32 %f219, %f218, %f315;
mov.b32 %r1598, %f219;
shfl.sync.bfly.b32 %r1602|%p53, %r1598, %r1583, %r1582, %r1584;
mov.b32 %f220, %r1602;
add.ftz.f32 %f221, %f219, %f220;
mov.b32 %r1603, %f221;
shfl.sync.bfly.b32 %r1605|%p54, %r1603, %r1587, %r1582, %r1584;
mov.b32 %f222, %r1605;
add.ftz.f32 %f223, %f221, %f222;
mov.b32 %r1606, %f223;
shfl.sync.bfly.b32 %r1608|%p55, %r1606, %r1590, %r1582, %r1584;
mov.b32 %f224, %r1608;
add.ftz.f32 %f225, %f223, %f224;
mov.b32 %r1609, %f225;
shfl.sync.bfly.b32 %r1611|%p56, %r1609, %r1593, %r1582, %r1584;
mov.b32 %f226, %r1611;
add.ftz.f32 %f227, %f225, %f226;
mov.b32 %r1612, %f227;
shfl.sync.bfly.b32 %r1614|%p57, %r1612, %r1596, %r1582, %r1584;
mov.b32 %f228, %r1614;
add.ftz.f32 %f50, %f227, %f228;
@%p22 bra $L__BB315_22;

add.s32 %r1615, %r31, 1;
cvt.u64.u32 %rd140, %r1615;
add.s64 %rd141, %rd140, %rd20;
shl.b64 %rd142, %rd141, 2;
add.s64 %rd143, %rd1, %rd142;
st.global.f32 [%rd143], %f50;

$L__BB315_22:
shl.b32 %r1616, %r35, 2;
mov.u32 %r1617, 2;
ld.shared.f32 %f229, [%r30+1024];
add.ftz.f32 %f230, %f229, %f314;
mov.b32 %r1618, %f230;
mov.u32 %r1619, 31;
mov.u32 %r1620, 16;
mov.u32 %r1621, -1;
shfl.sync.bfly.b32 %r1622|%p59, %r1618, %r1620, %r1619, %r1621;
mov.b32 %f231, %r1622;
add.ftz.f32 %f232, %f230, %f231;
mov.b32 %r1623, %f232;
mov.u32 %r1624, 8;
shfl.sync.bfly.b32 %r1625|%p60, %r1623, %r1624, %r1619, %r1621;
mov.b32 %f233, %r1625;
add.ftz.f32 %f234, %f232, %f233;
mov.b32 %r1626, %f234;
mov.u32 %r1627, 4;
shfl.sync.bfly.b32 %r1628|%p61, %r1626, %r1627, %r1619, %r1621;
mov.b32 %f235, %r1628;
add.ftz.f32 %f236, %f234, %f235;
mov.b32 %r1629, %f236;
shfl.sync.bfly.b32 %r1630|%p62, %r1629, %r1617, %r1619, %r1621;
mov.b32 %f237, %r1630;
add.ftz.f32 %f238, %f236, %f237;
mov.b32 %r1631, %f238;
mov.u32 %r1632, 1;
shfl.sync.bfly.b32 %r1633|%p63, %r1631, %r1632, %r1619, %r1621;
mov.b32 %f239, %r1633;
add.ftz.f32 %f51, %f238, %f239;
cvt.u64.u32 %rd144, %r1616;
add.s64 %rd145, %rd144, %rd20;
shl.b64 %rd146, %rd145, 2;
add.s64 %rd23, %rd1, %rd146;
@%p14 bra $L__BB315_24;

st.global.f32 [%rd23], %f51;

$L__BB315_24:
ld.shared.f32 %f240, [%r30+1152];
add.ftz.f32 %f241, %f240, %f313;
mov.b32 %r1634, %f241;
shfl.sync.bfly.b32 %r1638|%p65, %r1634, %r1620, %r1619, %r1621;
mov.b32 %f242, %r1638;
add.ftz.f32 %f243, %f241, %f242;
mov.b32 %r1639, %f243;
shfl.sync.bfly.b32 %r1641|%p66, %r1639, %r1624, %r1619, %r1621;
mov.b32 %f244, %r1641;
add.ftz.f32 %f245, %f243, %f244;
mov.b32 %r1642, %f245;
shfl.sync.bfly.b32 %r1644|%p67, %r1642, %r1627, %r1619, %r1621;
mov.b32 %f246, %r1644;
add.ftz.f32 %f247, %f245, %f246;
mov.b32 %r1645, %f247;
shfl.sync.bfly.b32 %r1647|%p68, %r1645, %r1617, %r1619, %r1621;
mov.b32 %f248, %r1647;
add.ftz.f32 %f249, %f247, %f248;
mov.b32 %r1648, %f249;
shfl.sync.bfly.b32 %r1650|%p69, %r1648, %r1632, %r1619, %r1621;
mov.b32 %f250, %r1650;
add.ftz.f32 %f52, %f249, %f250;
@%p22 bra $L__BB315_26;

st.global.f32 [%rd23+4], %f52;

$L__BB315_26:
mul.lo.s32 %r32, %r35, 5;
ld.shared.f32 %f251, [%r30+1280];
add.ftz.f32 %f252, %f251, %f312;
mov.b32 %r1651, %f252;
mov.u32 %r1652, 31;
mov.u32 %r1653, 16;
mov.u32 %r1654, -1;
shfl.sync.bfly.b32 %r1655|%p71, %r1651, %r1653, %r1652, %r1654;
mov.b32 %f253, %r1655;
add.ftz.f32 %f254, %f252, %f253;
mov.b32 %r1656, %f254;
mov.u32 %r1657, 8;
shfl.sync.bfly.b32 %r1658|%p72, %r1656, %r1657, %r1652, %r1654;
mov.b32 %f255, %r1658;
add.ftz.f32 %f256, %f254, %f255;
mov.b32 %r1659, %f256;
mov.u32 %r1660, 4;
shfl.sync.bfly.b32 %r1661|%p73, %r1659, %r1660, %r1652, %r1654;
mov.b32 %f257, %r1661;
add.ftz.f32 %f258, %f256, %f257;
mov.b32 %r1662, %f258;
mov.u32 %r1663, 2;
shfl.sync.bfly.b32 %r1664|%p74, %r1662, %r1663, %r1652, %r1654;
mov.b32 %f259, %r1664;
add.ftz.f32 %f260, %f258, %f259;
mov.b32 %r1665, %f260;
mov.u32 %r1666, 1;
shfl.sync.bfly.b32 %r1667|%p75, %r1665, %r1666, %r1652, %r1654;
mov.b32 %f261, %r1667;
add.ftz.f32 %f53, %f260, %f261;
@%p14 bra $L__BB315_28;

cvt.u64.u32 %rd147, %r32;
add.s64 %rd148, %rd147, %rd20;
shl.b64 %rd149, %rd148, 2;
add.s64 %rd150, %rd1, %rd149;
st.global.f32 [%rd150], %f53;

$L__BB315_28:
ld.shared.f32 %f262, [%r30+1408];
add.ftz.f32 %f263, %f262, %f311;
mov.b32 %r1668, %f263;
shfl.sync.bfly.b32 %r1672|%p77, %r1668, %r1653, %r1652, %r1654;
mov.b32 %f264, %r1672;
add.ftz.f32 %f265, %f263, %f264;
mov.b32 %r1673, %f265;
shfl.sync.bfly.b32 %r1675|%p78, %r1673, %r1657, %r1652, %r1654;
mov.b32 %f266, %r1675;
add.ftz.f32 %f267, %f265, %f266;
mov.b32 %r1676, %f267;
shfl.sync.bfly.b32 %r1678|%p79, %r1676, %r1660, %r1652, %r1654;
mov.b32 %f268, %r1678;
add.ftz.f32 %f269, %f267, %f268;
mov.b32 %r1679, %f269;
shfl.sync.bfly.b32 %r1681|%p80, %r1679, %r1663, %r1652, %r1654;
mov.b32 %f270, %r1681;
add.ftz.f32 %f271, %f269, %f270;
mov.b32 %r1682, %f271;
shfl.sync.bfly.b32 %r1684|%p81, %r1682, %r1666, %r1652, %r1654;
mov.b32 %f272, %r1684;
add.ftz.f32 %f54, %f271, %f272;
@%p22 bra $L__BB315_30;

add.s32 %r1685, %r32, 1;
cvt.u64.u32 %rd151, %r1685;
add.s64 %rd152, %rd151, %rd20;
shl.b64 %rd153, %rd152, 2;
add.s64 %rd154, %rd1, %rd153;
st.global.f32 [%rd154], %f54;

$L__BB315_30:
mul.lo.s32 %r1686, %r35, 6;
ld.shared.f32 %f273, [%r30+1536];
add.ftz.f32 %f274, %f273, %f310;
mov.b32 %r1687, %f274;
mov.u32 %r1688, 31;
mov.u32 %r1689, 16;
mov.u32 %r1690, -1;
shfl.sync.bfly.b32 %r1691|%p83, %r1687, %r1689, %r1688, %r1690;
mov.b32 %f275, %r1691;
add.ftz.f32 %f276, %f274, %f275;
mov.b32 %r1692, %f276;
mov.u32 %r1693, 8;
shfl.sync.bfly.b32 %r1694|%p84, %r1692, %r1693, %r1688, %r1690;
mov.b32 %f277, %r1694;
add.ftz.f32 %f278, %f276, %f277;
mov.b32 %r1695, %f278;
mov.u32 %r1696, 4;
shfl.sync.bfly.b32 %r1697|%p85, %r1695, %r1696, %r1688, %r1690;
mov.b32 %f279, %r1697;
add.ftz.f32 %f280, %f278, %f279;
mov.b32 %r1698, %f280;
mov.u32 %r1699, 2;
shfl.sync.bfly.b32 %r1700|%p86, %r1698, %r1699, %r1688, %r1690;
mov.b32 %f281, %r1700;
add.ftz.f32 %f282, %f280, %f281;
mov.b32 %r1701, %f282;
mov.u32 %r1702, 1;
shfl.sync.bfly.b32 %r1703|%p87, %r1701, %r1702, %r1688, %r1690;
mov.b32 %f283, %r1703;
add.ftz.f32 %f55, %f282, %f283;
cvt.u64.u32 %rd155, %r1686;
add.s64 %rd156, %rd155, %rd20;
shl.b64 %rd157, %rd156, 2;
add.s64 %rd24, %rd1, %rd157;
@%p14 bra $L__BB315_32;

st.global.f32 [%rd24], %f55;

$L__BB315_32:
ld.shared.f32 %f284, [%r30+1664];
add.ftz.f32 %f285, %f284, %f309;
mov.b32 %r1704, %f285;
shfl.sync.bfly.b32 %r1708|%p89, %r1704, %r1689, %r1688, %r1690;
mov.b32 %f286, %r1708;
add.ftz.f32 %f287, %f285, %f286;
mov.b32 %r1709, %f287;
shfl.sync.bfly.b32 %r1711|%p90, %r1709, %r1693, %r1688, %r1690;
mov.b32 %f288, %r1711;
add.ftz.f32 %f289, %f287, %f288;
mov.b32 %r1712, %f289;
shfl.sync.bfly.b32 %r1714|%p91, %r1712, %r1696, %r1688, %r1690;
mov.b32 %f290, %r1714;
add.ftz.f32 %f291, %f289, %f290;
mov.b32 %r1715, %f291;
shfl.sync.bfly.b32 %r1717|%p92, %r1715, %r1699, %r1688, %r1690;
mov.b32 %f292, %r1717;
add.ftz.f32 %f293, %f291, %f292;
mov.b32 %r1718, %f293;
shfl.sync.bfly.b32 %r1720|%p93, %r1718, %r1702, %r1688, %r1690;
mov.b32 %f294, %r1720;
add.ftz.f32 %f56, %f293, %f294;
@%p22 bra $L__BB315_34;

st.global.f32 [%rd24+4], %f56;

$L__BB315_34:
ret;

}
.entry _Z13mul_mat_vec_qIL9ggml_type21ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj(
.param .u64 _Z13mul_mat_vec_qIL9ggml_type21ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type21ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1,
.param .u64 _Z13mul_mat_vec_qIL9ggml_type21ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_2,
.param .align 8 .b8 _Z13mul_mat_vec_qIL9ggml_type21ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_3[48],
.param .u64 _Z13mul_mat_vec_qIL9ggml_type21ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type21ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_6[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type21ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13,
.param .align 4 .b8 _Z13mul_mat_vec_qIL9ggml_type21ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14[12],
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17,
.param .u32 _Z13mul_mat_vec_qIL9ggml_type21ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_18
)
.maxntid 64, 1, 1
.minnctapersm 1
{
.reg .pred %p<107>;
.reg .b16 %rs<61>;
.reg .f32 %f<369>;
.reg .b32 %r<1936>;
.reg .b64 %rd<183>;

.shared .align 4 .b8 _ZZ13mul_mat_vec_qIL9ggml_type21ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared[2048];

ld.param.u64 %rd24, [_Z13mul_mat_vec_qIL9ggml_type21ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_0];
ld.param.u64 %rd25, [_Z13mul_mat_vec_qIL9ggml_type21ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_1];
ld.param.u64 %rd26, [_Z13mul_mat_vec_qIL9ggml_type21ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_4];
ld.param.u32 %r53, [_Z13mul_mat_vec_qIL9ggml_type21ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_5];
ld.param.u32 %r38, [_Z13mul_mat_vec_qIL9ggml_type21ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_7];
ld.param.u32 %r39, [_Z13mul_mat_vec_qIL9ggml_type21ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_8];
ld.param.u32 %r40, [_Z13mul_mat_vec_qIL9ggml_type21ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_9];
ld.param.u32 %r44, [_Z13mul_mat_vec_qIL9ggml_type21ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_11];
ld.param.u32 %r45, [_Z13mul_mat_vec_qIL9ggml_type21ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_12];
ld.param.u32 %r46, [_Z13mul_mat_vec_qIL9ggml_type21ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_13];
ld.param.u32 %r50, [_Z13mul_mat_vec_qIL9ggml_type21ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_15];
ld.param.u32 %r51, [_Z13mul_mat_vec_qIL9ggml_type21ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_16];
ld.param.u32 %r52, [_Z13mul_mat_vec_qIL9ggml_type21ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_17];
ld.param.u32 %r48, [_Z13mul_mat_vec_qIL9ggml_type21ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14+4];
ld.param.u32 %r47, [_Z13mul_mat_vec_qIL9ggml_type21ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_14];
ld.param.u32 %r42, [_Z13mul_mat_vec_qIL9ggml_type21ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10+4];
ld.param.u32 %r41, [_Z13mul_mat_vec_qIL9ggml_type21ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj_param_10];
mov.u32 %r1, %tid.y;
shl.b32 %r54, %r1, 5;
mov.u32 %r55, %tid.x;
add.s32 %r2, %r54, %r55;
mov.u32 %r3, %ctaid.y;
mov.u32 %r4, %ctaid.z;
shr.s32 %r56, %r2, 31;
shr.u32 %r57, %r56, 29;
add.s32 %r58, %r2, %r57;
shr.s32 %r1935, %r58, 3;
shr.u32 %r6, %r53, 8;
setp.ge.s32 %p3, %r1935, %r6;
mov.f32 %f353, 0f00000000;
mov.f32 %f354, %f353;
mov.f32 %f355, %f353;
mov.f32 %f356, %f353;
mov.f32 %f357, %f353;
mov.f32 %f358, %f353;
mov.f32 %f359, %f353;
mov.f32 %f360, %f353;
mov.f32 %f361, %f353;
mov.f32 %f362, %f353;
mov.f32 %f363, %f353;
mov.f32 %f364, %f353;
mov.f32 %f365, %f353;
mov.f32 %f366, %f353;
mov.f32 %f367, %f353;
mov.f32 %f368, %f353;
@%p3 bra $L__BB316_3;

mov.u32 %r59, %ctaid.x;
shl.b32 %r60, %r59, 1;
mul.lo.s32 %r61, %r3, %r45;
cvt.u64.u32 %rd27, %r61;
mul.lo.s32 %r62, %r4, %r51;
cvt.u64.u32 %rd28, %r62;
add.s64 %rd1, %rd28, %rd27;
mul.hi.u32 %r63, %r4, %r47;
add.s32 %r64, %r63, %r4;
shr.u32 %r65, %r64, %r48;
mul.lo.s32 %r66, %r65, %r50;
mul.hi.u32 %r67, %r3, %r41;
add.s32 %r68, %r67, %r3;
shr.u32 %r69, %r68, %r42;
mul.lo.s32 %r70, %r69, %r44;
cvta.to.global.u64 %rd182, %rd24;
and.b32 %r74, %r58, -8;
sub.s32 %r75, %r2, %r74;
shl.b32 %r76, %r75, 1;
shr.u32 %r77, %r75, 31;
add.s32 %r78, %r75, %r77;
shr.s32 %r79, %r78, 1;
cvt.s64.s32 %rd29, %r79;
shl.b32 %r80, %r75, 2;
cvt.s64.s32 %rd3, %r75;
and.b32 %r7, %r80, 4;
shl.b32 %r81, %r39, 1;
shl.b32 %r1927, %r1935, 3;
add.s32 %r1934, %r1927, %r81;
mad.lo.s32 %r1933, %r39, 3, %r1927;
shl.b32 %r82, %r39, 2;
add.s32 %r1932, %r1927, %r82;
mad.lo.s32 %r1931, %r39, 5, %r1927;
mad.lo.s32 %r1930, %r39, 6, %r1927;
mad.lo.s32 %r1929, %r39, 7, %r1927;
add.s32 %r1928, %r39, %r1927;
add.s32 %r83, %r1935, %r66;
add.s32 %r84, %r83, %r70;
or.b32 %r85, %r60, 1;
mad.lo.s32 %r86, %r38, %r85, %r84;
mul.wide.s32 %rd4, %r86, 110;
add.s64 %rd30, %rd4, %rd29;
add.s64 %rd5, %rd30, 106;
add.s64 %rd31, %rd4, 74;
not.b32 %r87, %r76;
mul.wide.s32 %rd32, %r87, 2;
sub.s64 %rd6, %rd31, %rd32;
mul.wide.s32 %rd33, %r76, 2;
add.s64 %rd34, %rd4, %rd33;
add.s64 %rd7, %rd34, 74;
add.s64 %rd35, %rd4, %rd3;
add.s64 %rd8, %rd35, 66;
mad.lo.s32 %r88, %r60, %r38, %r1935;
add.s32 %r89, %r88, %r66;
add.s32 %r90, %r89, %r70;
mul.wide.s32 %rd9, %r90, 110;
mul.wide.s32 %rd36, %r80, 2;
add.s64 %rd37, %rd9, %rd36;
add.s64 %rd10, %rd37, 4;
add.s64 %rd38, %rd4, %rd36;
add.s64 %rd11, %rd38, 2;
add.s64 %rd39, %rd9, %rd29;
add.s64 %rd12, %rd39, 106;
add.s64 %rd40, %rd9, 74;
sub.s64 %rd13, %rd40, %rd32;
add.s64 %rd41, %rd9, %rd33;
add.s64 %rd14, %rd41, 74;
add.s64 %rd42, %rd9, %rd3;
add.s64 %rd15, %rd42, 66;
cvta.to.global.u64 %rd16, %rd25;
mov.f32 %f353, 0f00000000;

$L__BB316_2:
cvt.u64.u32 %rd43, %r1927;
add.s64 %rd44, %rd1, %rd43;
add.s64 %rd45, %rd44, %rd3;
mul.lo.s64 %rd46, %rd45, 36;
add.s64 %rd47, %rd182, %rd10;
ld.global.nc.u16 %rs17, [%rd47];
ld.global.nc.u16 %rs18, [%rd47+2];
ld.global.nc.u16 %rs19, [%rd47+4];
add.s64 %rd48, %rd182, %rd15;
ld.global.nc.u8 %rs20, [%rd48];
add.s64 %rd49, %rd182, %rd14;
add.s64 %rd50, %rd182, %rd13;
ld.global.nc.u16 %rs21, [%rd49];
ld.global.nc.u16 %rs22, [%rd50];
ld.global.nc.u16 %rs23, [%rd47+-2];
and.b16 %rs24, %rs23, 255;
cvt.u32.u16 %r875, %rs24;
shl.b16 %rs25, %rs20, 8;
cvt.u32.u16 %r876, %rs25;
and.b32 %r877, %r876, 256;
or.b32 %r878, %r877, %r875;
mul.wide.u32 %rd51, %r878, 4;
mov.u64 %rd52, iq3s_grid;
add.s64 %rd53, %rd52, %rd51;
ld.global.nc.u32 %r879, [%rd53];
not.b32 %r880, %r879;
shr.u16 %rs26, %rs23, 8;
cvt.u32.u16 %r881, %rs26;
mul.wide.u16 %r882, %rs20, 128;
and.b32 %r883, %r882, 256;
or.b32 %r884, %r883, %r881;
mul.wide.u32 %rd54, %r884, 4;
add.s64 %rd55, %rd52, %rd54;
ld.global.nc.u32 %r885, [%rd55];
not.b32 %r886, %r885;
and.b16 %rs28, %rs21, 255;
mul.wide.u16 %r887, %rs28, 128;
cvt.u32.u16 %r888, %rs21;
shl.b32 %r889, %r888, 21;
or.b32 %r890, %r887, %r889;
and.b32 %r891, %r890, 25166208;
or.b32 %r892, %r891, -2139062144;
add.s32 %r893, %r892, -16843009;
or.b32 %r777, %r893, %r891;
mov.u32 %r91, %r777;

prmt.b32 %r91, %r91, 0, 0xba98;

mul.wide.u16 %r894, %rs28, 8;
shl.b32 %r895, %r888, 17;
or.b32 %r896, %r894, %r895;
and.b32 %r897, %r896, 25166208;
or.b32 %r898, %r897, -2139062144;
add.s32 %r899, %r898, -16843009;
or.b32 %r779, %r899, %r897;
mov.u32 %r93, %r779;

prmt.b32 %r93, %r93, 0, 0xba98;

xor.b32 %r900, %r91, %r879;
or.b32 %r901, %r900, -2139062144;
and.b32 %r902, %r91, 2139062143;
and.b32 %r903, %r880, -2139062144;
sub.s32 %r904, %r901, %r902;
xor.b32 %r96, %r903, %r904;
xor.b32 %r905, %r93, %r885;
or.b32 %r906, %r905, -2139062144;
and.b32 %r907, %r93, 2139062143;
and.b32 %r908, %r886, -2139062144;
sub.s32 %r909, %r906, %r907;
xor.b32 %r100, %r908, %r909;
add.s64 %rd56, %rd16, %rd46;
ld.global.nc.u32 %r97, [%rd56+4];
ld.global.nc.u32 %r101, [%rd56+8];
mov.u32 %r833, 0;

dp4a.s32.s32 %r95, %r96, %r97, %r833;


dp4a.s32.s32 %r99, %r100, %r101, %r95;

and.b16 %rs29, %rs17, 255;
cvt.u32.u16 %r910, %rs29;
mul.wide.u16 %r911, %rs20, 64;
and.b32 %r912, %r911, 256;
or.b32 %r913, %r912, %r910;
mul.wide.u32 %rd57, %r913, 4;
add.s64 %rd58, %rd52, %rd57;
ld.global.nc.u32 %r914, [%rd58];
not.b32 %r915, %r914;
shr.u16 %rs30, %rs17, 8;
cvt.u32.u16 %r916, %rs30;
mul.wide.u16 %r917, %rs20, 32;
and.b32 %r918, %r917, 256;
or.b32 %r919, %r918, %r916;
mul.wide.u32 %rd59, %r919, 4;
add.s64 %rd60, %rd52, %rd59;
ld.global.nc.u32 %r920, [%rd60];
not.b32 %r921, %r920;
shr.u16 %rs31, %rs21, 8;
mul.wide.u16 %r922, %rs31, 128;
cvt.u32.u16 %r923, %rs31;
shl.b32 %r924, %r923, 21;
or.b32 %r925, %r922, %r924;
and.b32 %r926, %r925, 25166208;
or.b32 %r927, %r926, -2139062144;
add.s32 %r928, %r927, -16843009;
or.b32 %r789, %r928, %r926;
mov.u32 %r103, %r789;

prmt.b32 %r103, %r103, 0, 0xba98;

mul.wide.u16 %r929, %rs31, 8;
shl.b32 %r930, %r923, 17;
or.b32 %r931, %r929, %r930;
and.b32 %r932, %r931, 25166208;
or.b32 %r933, %r932, -2139062144;
add.s32 %r934, %r933, -16843009;
or.b32 %r791, %r934, %r932;
mov.u32 %r105, %r791;

prmt.b32 %r105, %r105, 0, 0xba98;

xor.b32 %r935, %r103, %r914;
or.b32 %r936, %r935, -2139062144;
and.b32 %r937, %r103, 2139062143;
and.b32 %r938, %r915, -2139062144;
sub.s32 %r939, %r936, %r937;
xor.b32 %r108, %r938, %r939;
xor.b32 %r940, %r105, %r920;
or.b32 %r941, %r940, -2139062144;
and.b32 %r942, %r105, 2139062143;
and.b32 %r943, %r921, -2139062144;
sub.s32 %r944, %r941, %r942;
xor.b32 %r112, %r943, %r944;
ld.global.nc.u32 %r109, [%rd56+12];
ld.global.nc.u32 %r113, [%rd56+16];

dp4a.s32.s32 %r107, %r108, %r109, %r99;


dp4a.s32.s32 %r111, %r112, %r113, %r107;

and.b16 %rs32, %rs18, 255;
cvt.u32.u16 %r945, %rs32;
mul.wide.u16 %r946, %rs20, 16;
and.b32 %r947, %r946, 256;
or.b32 %r948, %r947, %r945;
mul.wide.u32 %rd61, %r948, 4;
add.s64 %rd62, %rd52, %rd61;
ld.global.nc.u32 %r949, [%rd62];
not.b32 %r950, %r949;
shr.u16 %rs33, %rs18, 8;
cvt.u32.u16 %r951, %rs33;
mul.wide.u16 %r952, %rs20, 8;
and.b32 %r953, %r952, 256;
or.b32 %r954, %r953, %r951;
mul.wide.u32 %rd63, %r954, 4;
add.s64 %rd64, %rd52, %rd63;
ld.global.nc.u32 %r955, [%rd64];
not.b32 %r956, %r955;
and.b16 %rs34, %rs22, 255;
mul.wide.u16 %r957, %rs34, 128;
cvt.u32.u16 %r958, %rs22;
shl.b32 %r959, %r958, 21;
or.b32 %r960, %r957, %r959;
and.b32 %r961, %r960, 25166208;
or.b32 %r962, %r961, -2139062144;
add.s32 %r963, %r962, -16843009;
or.b32 %r801, %r963, %r961;
mov.u32 %r115, %r801;

prmt.b32 %r115, %r115, 0, 0xba98;

mul.wide.u16 %r964, %rs34, 8;
shl.b32 %r965, %r958, 17;
or.b32 %r966, %r964, %r965;
and.b32 %r967, %r966, 25166208;
or.b32 %r968, %r967, -2139062144;
add.s32 %r969, %r968, -16843009;
or.b32 %r803, %r969, %r967;
mov.u32 %r117, %r803;

prmt.b32 %r117, %r117, 0, 0xba98;

xor.b32 %r970, %r115, %r949;
or.b32 %r971, %r970, -2139062144;
and.b32 %r972, %r115, 2139062143;
and.b32 %r973, %r950, -2139062144;
sub.s32 %r974, %r971, %r972;
xor.b32 %r120, %r973, %r974;
xor.b32 %r975, %r117, %r955;
or.b32 %r976, %r975, -2139062144;
and.b32 %r977, %r117, 2139062143;
and.b32 %r978, %r956, -2139062144;
sub.s32 %r979, %r976, %r977;
xor.b32 %r124, %r978, %r979;
ld.global.nc.u32 %r121, [%rd56+20];
ld.global.nc.u32 %r125, [%rd56+24];

dp4a.s32.s32 %r119, %r120, %r121, %r111;


dp4a.s32.s32 %r123, %r124, %r125, %r119;

and.b16 %rs35, %rs19, 255;
cvt.u32.u16 %r980, %rs35;
mul.wide.u16 %r981, %rs20, 4;
and.b32 %r982, %r981, 256;
or.b32 %r983, %r982, %r980;
mul.wide.u32 %rd65, %r983, 4;
add.s64 %rd66, %rd52, %rd65;
ld.global.nc.u32 %r984, [%rd66];
not.b32 %r985, %r984;
shr.u16 %rs36, %rs19, 8;
cvt.u32.u16 %r986, %rs36;
mul.wide.u16 %r987, %rs20, 2;
and.b32 %r988, %r987, 256;
or.b32 %r989, %r988, %r986;
mul.wide.u32 %rd67, %r989, 4;
add.s64 %rd68, %rd52, %rd67;
ld.global.nc.u32 %r990, [%rd68];
not.b32 %r991, %r990;
shr.u16 %rs37, %rs22, 8;
mul.wide.u16 %r992, %rs37, 128;
cvt.u32.u16 %r993, %rs37;
shl.b32 %r994, %r993, 21;
or.b32 %r995, %r992, %r994;
and.b32 %r996, %r995, 25166208;
or.b32 %r997, %r996, -2139062144;
add.s32 %r998, %r997, -16843009;
or.b32 %r813, %r998, %r996;
mov.u32 %r127, %r813;

prmt.b32 %r127, %r127, 0, 0xba98;

mul.wide.u16 %r999, %rs37, 8;
shl.b32 %r1000, %r993, 17;
or.b32 %r1001, %r999, %r1000;
and.b32 %r1002, %r1001, 25166208;
or.b32 %r1003, %r1002, -2139062144;
add.s32 %r1004, %r1003, -16843009;
or.b32 %r815, %r1004, %r1002;
mov.u32 %r129, %r815;

prmt.b32 %r129, %r129, 0, 0xba98;

xor.b32 %r1005, %r127, %r984;
or.b32 %r1006, %r1005, -2139062144;
and.b32 %r1007, %r127, 2139062143;
and.b32 %r1008, %r985, -2139062144;
sub.s32 %r1009, %r1006, %r1007;
xor.b32 %r132, %r1008, %r1009;
xor.b32 %r1010, %r129, %r990;
or.b32 %r1011, %r1010, -2139062144;
and.b32 %r1012, %r129, 2139062143;
and.b32 %r1013, %r991, -2139062144;
sub.s32 %r1014, %r1011, %r1012;
xor.b32 %r136, %r1013, %r1014;
ld.global.nc.u32 %r133, [%rd56+28];
ld.global.nc.u32 %r137, [%rd56+32];

dp4a.s32.s32 %r131, %r132, %r133, %r123;


dp4a.s32.s32 %r135, %r136, %r137, %r131;

add.s64 %rd69, %rd182, %rd12;
ld.global.nc.u8 %rs38, [%rd69];
cvt.u32.u16 %r1015, %rs38;
and.b32 %r1016, %r1015, 255;
shr.u32 %r1017, %r1016, %r7;
shl.b32 %r1018, %r1017, 1;
and.b32 %r1019, %r1018, 30;
or.b32 %r1020, %r1019, 1;
mul.lo.s32 %r1021, %r1020, %r135;
add.s64 %rd70, %rd182, %rd9;
ld.global.nc.u16 %rs1, [%rd70];

{ cvt.f32.f16 %f97, %rs1;}


ld.global.nc.u32 %r139, [%rd56];

{.reg .f16 low,high;
mov.b32 {low,high},%r139;
cvt.f32.f16 %f98, low;}


mul.ftz.f32 %f129, %f97, %f98;
cvt.rn.f32.s32 %f130, %r1021;
fma.rn.ftz.f32 %f368, %f129, %f130, %f368;
add.s64 %rd71, %rd182, %rd11;
ld.global.nc.u16 %rs39, [%rd71+2];
ld.global.nc.u16 %rs40, [%rd71+4];
ld.global.nc.u16 %rs41, [%rd71+6];
add.s64 %rd72, %rd182, %rd8;
ld.global.nc.u8 %rs42, [%rd72];
add.s64 %rd73, %rd182, %rd7;
add.s64 %rd74, %rd182, %rd6;
ld.global.nc.u16 %rs43, [%rd73];
ld.global.nc.u16 %rs44, [%rd74];
ld.global.nc.u16 %rs45, [%rd71];
and.b16 %rs46, %rs45, 255;
cvt.u32.u16 %r1022, %rs46;
shl.b16 %rs47, %rs42, 8;
cvt.u32.u16 %r1023, %rs47;
and.b32 %r1024, %r1023, 256;
or.b32 %r1025, %r1024, %r1022;
mul.wide.u32 %rd75, %r1025, 4;
add.s64 %rd76, %rd52, %rd75;
ld.global.nc.u32 %r1026, [%rd76];
not.b32 %r1027, %r1026;
shr.u16 %rs48, %rs45, 8;
cvt.u32.u16 %r1028, %rs48;
mul.wide.u16 %r1029, %rs42, 128;
and.b32 %r1030, %r1029, 256;
or.b32 %r1031, %r1030, %r1028;
mul.wide.u32 %rd77, %r1031, 4;
add.s64 %rd78, %rd52, %rd77;
ld.global.nc.u32 %r1032, [%rd78];
not.b32 %r1033, %r1032;
and.b16 %rs50, %rs43, 255;
mul.wide.u16 %r1034, %rs50, 128;
cvt.u32.u16 %r1035, %rs43;
shl.b32 %r1036, %r1035, 21;
or.b32 %r1037, %r1034, %r1036;
and.b32 %r1038, %r1037, 25166208;
or.b32 %r1039, %r1038, -2139062144;
add.s32 %r1040, %r1039, -16843009;
or.b32 %r826, %r1040, %r1038;
mov.u32 %r140, %r826;

prmt.b32 %r140, %r140, 0, 0xba98;

mul.wide.u16 %r1041, %rs50, 8;
shl.b32 %r1042, %r1035, 17;
or.b32 %r1043, %r1041, %r1042;
and.b32 %r1044, %r1043, 25166208;
or.b32 %r1045, %r1044, -2139062144;
add.s32 %r1046, %r1045, -16843009;
or.b32 %r828, %r1046, %r1044;
mov.u32 %r142, %r828;

prmt.b32 %r142, %r142, 0, 0xba98;

xor.b32 %r1047, %r140, %r1026;
or.b32 %r1048, %r1047, -2139062144;
and.b32 %r1049, %r140, 2139062143;
and.b32 %r1050, %r1027, -2139062144;
sub.s32 %r1051, %r1048, %r1049;
xor.b32 %r145, %r1050, %r1051;
xor.b32 %r1052, %r142, %r1032;
or.b32 %r1053, %r1052, -2139062144;
and.b32 %r1054, %r142, 2139062143;
and.b32 %r1055, %r1033, -2139062144;
sub.s32 %r1056, %r1053, %r1054;
xor.b32 %r149, %r1055, %r1056;

dp4a.s32.s32 %r144, %r145, %r97, %r833;


dp4a.s32.s32 %r148, %r149, %r101, %r144;

and.b16 %rs51, %rs39, 255;
cvt.u32.u16 %r1057, %rs51;
mul.wide.u16 %r1058, %rs42, 64;
and.b32 %r1059, %r1058, 256;
or.b32 %r1060, %r1059, %r1057;
mul.wide.u32 %rd79, %r1060, 4;
add.s64 %rd80, %rd52, %rd79;
ld.global.nc.u32 %r1061, [%rd80];
not.b32 %r1062, %r1061;
shr.u16 %rs52, %rs39, 8;
cvt.u32.u16 %r1063, %rs52;
mul.wide.u16 %r1064, %rs42, 32;
and.b32 %r1065, %r1064, 256;
or.b32 %r1066, %r1065, %r1063;
mul.wide.u32 %rd81, %r1066, 4;
add.s64 %rd82, %rd52, %rd81;
ld.global.nc.u32 %r1067, [%rd82];
not.b32 %r1068, %r1067;
shr.u16 %rs53, %rs43, 8;
mul.wide.u16 %r1069, %rs53, 128;
cvt.u32.u16 %r1070, %rs53;
shl.b32 %r1071, %r1070, 21;
or.b32 %r1072, %r1069, %r1071;
and.b32 %r1073, %r1072, 25166208;
or.b32 %r1074, %r1073, -2139062144;
add.s32 %r1075, %r1074, -16843009;
or.b32 %r838, %r1075, %r1073;
mov.u32 %r152, %r838;

prmt.b32 %r152, %r152, 0, 0xba98;

mul.wide.u16 %r1076, %rs53, 8;
shl.b32 %r1077, %r1070, 17;
or.b32 %r1078, %r1076, %r1077;
and.b32 %r1079, %r1078, 25166208;
or.b32 %r1080, %r1079, -2139062144;
add.s32 %r1081, %r1080, -16843009;
or.b32 %r840, %r1081, %r1079;
mov.u32 %r154, %r840;

prmt.b32 %r154, %r154, 0, 0xba98;

xor.b32 %r1082, %r152, %r1061;
or.b32 %r1083, %r1082, -2139062144;
and.b32 %r1084, %r152, 2139062143;
and.b32 %r1085, %r1062, -2139062144;
sub.s32 %r1086, %r1083, %r1084;
xor.b32 %r157, %r1085, %r1086;
xor.b32 %r1087, %r154, %r1067;
or.b32 %r1088, %r1087, -2139062144;
and.b32 %r1089, %r154, 2139062143;
and.b32 %r1090, %r1068, -2139062144;
sub.s32 %r1091, %r1088, %r1089;
xor.b32 %r161, %r1090, %r1091;

dp4a.s32.s32 %r156, %r157, %r109, %r148;


dp4a.s32.s32 %r160, %r161, %r113, %r156;

and.b16 %rs54, %rs40, 255;
cvt.u32.u16 %r1092, %rs54;
mul.wide.u16 %r1093, %rs42, 16;
and.b32 %r1094, %r1093, 256;
or.b32 %r1095, %r1094, %r1092;
mul.wide.u32 %rd83, %r1095, 4;
add.s64 %rd84, %rd52, %rd83;
ld.global.nc.u32 %r1096, [%rd84];
not.b32 %r1097, %r1096;
shr.u16 %rs55, %rs40, 8;
cvt.u32.u16 %r1098, %rs55;
mul.wide.u16 %r1099, %rs42, 8;
and.b32 %r1100, %r1099, 256;
or.b32 %r1101, %r1100, %r1098;
mul.wide.u32 %rd85, %r1101, 4;
add.s64 %rd86, %rd52, %rd85;
ld.global.nc.u32 %r1102, [%rd86];
not.b32 %r1103, %r1102;
and.b16 %rs56, %rs44, 255;
mul.wide.u16 %r1104, %rs56, 128;
cvt.u32.u16 %r1105, %rs44;
shl.b32 %r1106, %r1105, 21;
or.b32 %r1107, %r1104, %r1106;
and.b32 %r1108, %r1107, 25166208;
or.b32 %r1109, %r1108, -2139062144;
add.s32 %r1110, %r1109, -16843009;
or.b32 %r850, %r1110, %r1108;
mov.u32 %r164, %r850;

prmt.b32 %r164, %r164, 0, 0xba98;

mul.wide.u16 %r1111, %rs56, 8;
shl.b32 %r1112, %r1105, 17;
or.b32 %r1113, %r1111, %r1112;
and.b32 %r1114, %r1113, 25166208;
or.b32 %r1115, %r1114, -2139062144;
add.s32 %r1116, %r1115, -16843009;
or.b32 %r852, %r1116, %r1114;
mov.u32 %r166, %r852;

prmt.b32 %r166, %r166, 0, 0xba98;

xor.b32 %r1117, %r164, %r1096;
or.b32 %r1118, %r1117, -2139062144;
and.b32 %r1119, %r164, 2139062143;
and.b32 %r1120, %r1097, -2139062144;
sub.s32 %r1121, %r1118, %r1119;
xor.b32 %r169, %r1120, %r1121;
xor.b32 %r1122, %r166, %r1102;
or.b32 %r1123, %r1122, -2139062144;
and.b32 %r1124, %r166, 2139062143;
and.b32 %r1125, %r1103, -2139062144;
sub.s32 %r1126, %r1123, %r1124;
xor.b32 %r173, %r1125, %r1126;

dp4a.s32.s32 %r168, %r169, %r121, %r160;


dp4a.s32.s32 %r172, %r173, %r125, %r168;

and.b16 %rs57, %rs41, 255;
cvt.u32.u16 %r1127, %rs57;
mul.wide.u16 %r1128, %rs42, 4;
and.b32 %r1129, %r1128, 256;
or.b32 %r1130, %r1129, %r1127;
mul.wide.u32 %rd87, %r1130, 4;
add.s64 %rd88, %rd52, %rd87;
ld.global.nc.u32 %r1131, [%rd88];
not.b32 %r1132, %r1131;
shr.u16 %rs58, %rs41, 8;
cvt.u32.u16 %r1133, %rs58;
mul.wide.u16 %r1134, %rs42, 2;
and.b32 %r1135, %r1134, 256;
or.b32 %r1136, %r1135, %r1133;
mul.wide.u32 %rd89, %r1136, 4;
add.s64 %rd90, %rd52, %rd89;
ld.global.nc.u32 %r1137, [%rd90];
not.b32 %r1138, %r1137;
shr.u16 %rs59, %rs44, 8;
mul.wide.u16 %r1139, %rs59, 128;
cvt.u32.u16 %r1140, %rs59;
shl.b32 %r1141, %r1140, 21;
or.b32 %r1142, %r1139, %r1141;
and.b32 %r1143, %r1142, 25166208;
or.b32 %r1144, %r1143, -2139062144;
add.s32 %r1145, %r1144, -16843009;
or.b32 %r862, %r1145, %r1143;
mov.u32 %r176, %r862;

prmt.b32 %r176, %r176, 0, 0xba98;

mul.wide.u16 %r1146, %rs59, 8;
shl.b32 %r1147, %r1140, 17;
or.b32 %r1148, %r1146, %r1147;
and.b32 %r1149, %r1148, 25166208;
or.b32 %r1150, %r1149, -2139062144;
add.s32 %r1151, %r1150, -16843009;
or.b32 %r864, %r1151, %r1149;
mov.u32 %r178, %r864;

prmt.b32 %r178, %r178, 0, 0xba98;

xor.b32 %r1152, %r176, %r1131;
or.b32 %r1153, %r1152, -2139062144;
and.b32 %r1154, %r176, 2139062143;
and.b32 %r1155, %r1132, -2139062144;
sub.s32 %r1156, %r1153, %r1154;
xor.b32 %r181, %r1155, %r1156;
xor.b32 %r1157, %r178, %r1137;
or.b32 %r1158, %r1157, -2139062144;
and.b32 %r1159, %r178, 2139062143;
and.b32 %r1160, %r1138, -2139062144;
sub.s32 %r1161, %r1158, %r1159;
xor.b32 %r185, %r1160, %r1161;

dp4a.s32.s32 %r180, %r181, %r133, %r172;


dp4a.s32.s32 %r184, %r185, %r137, %r180;

add.s64 %rd91, %rd182, %rd5;
ld.global.nc.u8 %rs60, [%rd91];
cvt.u32.u16 %r1162, %rs60;
and.b32 %r1163, %r1162, 255;
shr.u32 %r1164, %r1163, %r7;
shl.b32 %r1165, %r1164, 1;
and.b32 %r1166, %r1165, 30;
or.b32 %r1167, %r1166, 1;
mul.lo.s32 %r1168, %r1167, %r184;
add.s64 %rd92, %rd182, %rd4;
ld.global.nc.u16 %rs2, [%rd92];

{ cvt.f32.f16 %f99, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r139;
cvt.f32.f16 %f100, low;}


mul.ftz.f32 %f131, %f99, %f100;
cvt.rn.f32.s32 %f132, %r1168;
fma.rn.ftz.f32 %f367, %f131, %f132, %f367;
cvt.u64.u32 %rd93, %r1928;
add.s64 %rd94, %rd1, %rd93;
add.s64 %rd95, %rd94, %rd3;
mul.lo.s64 %rd96, %rd95, 36;
mov.u32 %r189, %r777;

prmt.b32 %r189, %r189, 0, 0xba98;

mov.u32 %r191, %r779;

prmt.b32 %r191, %r191, 0, 0xba98;

xor.b32 %r1169, %r189, %r879;
or.b32 %r1170, %r1169, -2139062144;
and.b32 %r1171, %r189, 2139062143;
sub.s32 %r1172, %r1170, %r1171;
xor.b32 %r194, %r903, %r1172;
xor.b32 %r1173, %r191, %r885;
or.b32 %r1174, %r1173, -2139062144;
and.b32 %r1175, %r191, 2139062143;
sub.s32 %r1176, %r1174, %r1175;
xor.b32 %r198, %r908, %r1176;
add.s64 %rd97, %rd16, %rd96;
ld.global.nc.u32 %r195, [%rd97+4];
ld.global.nc.u32 %r199, [%rd97+8];

dp4a.s32.s32 %r193, %r194, %r195, %r833;


dp4a.s32.s32 %r197, %r198, %r199, %r193;

mov.u32 %r201, %r789;

prmt.b32 %r201, %r201, 0, 0xba98;

mov.u32 %r203, %r791;

prmt.b32 %r203, %r203, 0, 0xba98;

xor.b32 %r1177, %r201, %r914;
or.b32 %r1178, %r1177, -2139062144;
and.b32 %r1179, %r201, 2139062143;
sub.s32 %r1180, %r1178, %r1179;
xor.b32 %r206, %r938, %r1180;
xor.b32 %r1181, %r203, %r920;
or.b32 %r1182, %r1181, -2139062144;
and.b32 %r1183, %r203, 2139062143;
sub.s32 %r1184, %r1182, %r1183;
xor.b32 %r210, %r943, %r1184;
ld.global.nc.u32 %r207, [%rd97+12];
ld.global.nc.u32 %r211, [%rd97+16];

dp4a.s32.s32 %r205, %r206, %r207, %r197;


dp4a.s32.s32 %r209, %r210, %r211, %r205;

mov.u32 %r213, %r801;

prmt.b32 %r213, %r213, 0, 0xba98;

mov.u32 %r215, %r803;

prmt.b32 %r215, %r215, 0, 0xba98;

xor.b32 %r1185, %r213, %r949;
or.b32 %r1186, %r1185, -2139062144;
and.b32 %r1187, %r213, 2139062143;
sub.s32 %r1188, %r1186, %r1187;
xor.b32 %r218, %r973, %r1188;
xor.b32 %r1189, %r215, %r955;
or.b32 %r1190, %r1189, -2139062144;
and.b32 %r1191, %r215, 2139062143;
sub.s32 %r1192, %r1190, %r1191;
xor.b32 %r222, %r978, %r1192;
ld.global.nc.u32 %r219, [%rd97+20];
ld.global.nc.u32 %r223, [%rd97+24];

dp4a.s32.s32 %r217, %r218, %r219, %r209;


dp4a.s32.s32 %r221, %r222, %r223, %r217;

mov.u32 %r225, %r813;

prmt.b32 %r225, %r225, 0, 0xba98;

mov.u32 %r227, %r815;

prmt.b32 %r227, %r227, 0, 0xba98;

xor.b32 %r1193, %r225, %r984;
or.b32 %r1194, %r1193, -2139062144;
and.b32 %r1195, %r225, 2139062143;
sub.s32 %r1196, %r1194, %r1195;
xor.b32 %r230, %r1008, %r1196;
xor.b32 %r1197, %r227, %r990;
or.b32 %r1198, %r1197, -2139062144;
and.b32 %r1199, %r227, 2139062143;
sub.s32 %r1200, %r1198, %r1199;
xor.b32 %r234, %r1013, %r1200;
ld.global.nc.u32 %r231, [%rd97+28];
ld.global.nc.u32 %r235, [%rd97+32];

dp4a.s32.s32 %r229, %r230, %r231, %r221;


dp4a.s32.s32 %r233, %r234, %r235, %r229;

mul.lo.s32 %r1201, %r1020, %r233;

{ cvt.f32.f16 %f101, %rs1;}


ld.global.nc.u32 %r237, [%rd97];

{.reg .f16 low,high;
mov.b32 {low,high},%r237;
cvt.f32.f16 %f102, low;}


mul.ftz.f32 %f133, %f101, %f102;
cvt.rn.f32.s32 %f134, %r1201;
fma.rn.ftz.f32 %f366, %f133, %f134, %f366;
mov.u32 %r238, %r826;

prmt.b32 %r238, %r238, 0, 0xba98;

mov.u32 %r240, %r828;

prmt.b32 %r240, %r240, 0, 0xba98;

xor.b32 %r1202, %r238, %r1026;
or.b32 %r1203, %r1202, -2139062144;
and.b32 %r1204, %r238, 2139062143;
sub.s32 %r1205, %r1203, %r1204;
xor.b32 %r243, %r1050, %r1205;
xor.b32 %r1206, %r240, %r1032;
or.b32 %r1207, %r1206, -2139062144;
and.b32 %r1208, %r240, 2139062143;
sub.s32 %r1209, %r1207, %r1208;
xor.b32 %r247, %r1055, %r1209;

dp4a.s32.s32 %r242, %r243, %r195, %r833;


dp4a.s32.s32 %r246, %r247, %r199, %r242;

mov.u32 %r250, %r838;

prmt.b32 %r250, %r250, 0, 0xba98;

mov.u32 %r252, %r840;

prmt.b32 %r252, %r252, 0, 0xba98;

xor.b32 %r1210, %r250, %r1061;
or.b32 %r1211, %r1210, -2139062144;
and.b32 %r1212, %r250, 2139062143;
sub.s32 %r1213, %r1211, %r1212;
xor.b32 %r255, %r1085, %r1213;
xor.b32 %r1214, %r252, %r1067;
or.b32 %r1215, %r1214, -2139062144;
and.b32 %r1216, %r252, 2139062143;
sub.s32 %r1217, %r1215, %r1216;
xor.b32 %r259, %r1090, %r1217;

dp4a.s32.s32 %r254, %r255, %r207, %r246;


dp4a.s32.s32 %r258, %r259, %r211, %r254;

mov.u32 %r262, %r850;

prmt.b32 %r262, %r262, 0, 0xba98;

mov.u32 %r264, %r852;

prmt.b32 %r264, %r264, 0, 0xba98;

xor.b32 %r1218, %r262, %r1096;
or.b32 %r1219, %r1218, -2139062144;
and.b32 %r1220, %r262, 2139062143;
sub.s32 %r1221, %r1219, %r1220;
xor.b32 %r267, %r1120, %r1221;
xor.b32 %r1222, %r264, %r1102;
or.b32 %r1223, %r1222, -2139062144;
and.b32 %r1224, %r264, 2139062143;
sub.s32 %r1225, %r1223, %r1224;
xor.b32 %r271, %r1125, %r1225;

dp4a.s32.s32 %r266, %r267, %r219, %r258;


dp4a.s32.s32 %r270, %r271, %r223, %r266;

mov.u32 %r274, %r862;

prmt.b32 %r274, %r274, 0, 0xba98;

mov.u32 %r276, %r864;

prmt.b32 %r276, %r276, 0, 0xba98;

xor.b32 %r1226, %r274, %r1131;
or.b32 %r1227, %r1226, -2139062144;
and.b32 %r1228, %r274, 2139062143;
sub.s32 %r1229, %r1227, %r1228;
xor.b32 %r279, %r1155, %r1229;
xor.b32 %r1230, %r276, %r1137;
or.b32 %r1231, %r1230, -2139062144;
and.b32 %r1232, %r276, 2139062143;
sub.s32 %r1233, %r1231, %r1232;
xor.b32 %r283, %r1160, %r1233;

dp4a.s32.s32 %r278, %r279, %r231, %r270;


dp4a.s32.s32 %r282, %r283, %r235, %r278;

mul.lo.s32 %r1234, %r1167, %r282;

{ cvt.f32.f16 %f103, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r237;
cvt.f32.f16 %f104, low;}


mul.ftz.f32 %f135, %f103, %f104;
cvt.rn.f32.s32 %f136, %r1234;
fma.rn.ftz.f32 %f365, %f135, %f136, %f365;
cvt.u64.u32 %rd98, %r1934;
add.s64 %rd99, %rd1, %rd98;
add.s64 %rd100, %rd99, %rd3;
mul.lo.s64 %rd101, %rd100, 36;
mov.u32 %r287, %r777;

prmt.b32 %r287, %r287, 0, 0xba98;

mov.u32 %r289, %r779;

prmt.b32 %r289, %r289, 0, 0xba98;

xor.b32 %r1235, %r287, %r879;
or.b32 %r1236, %r1235, -2139062144;
and.b32 %r1237, %r287, 2139062143;
sub.s32 %r1238, %r1236, %r1237;
xor.b32 %r292, %r903, %r1238;
xor.b32 %r1239, %r289, %r885;
or.b32 %r1240, %r1239, -2139062144;
and.b32 %r1241, %r289, 2139062143;
sub.s32 %r1242, %r1240, %r1241;
xor.b32 %r296, %r908, %r1242;
add.s64 %rd102, %rd16, %rd101;
ld.global.nc.u32 %r293, [%rd102+4];
ld.global.nc.u32 %r297, [%rd102+8];

dp4a.s32.s32 %r291, %r292, %r293, %r833;


dp4a.s32.s32 %r295, %r296, %r297, %r291;

mov.u32 %r299, %r789;

prmt.b32 %r299, %r299, 0, 0xba98;

mov.u32 %r301, %r791;

prmt.b32 %r301, %r301, 0, 0xba98;

xor.b32 %r1243, %r299, %r914;
or.b32 %r1244, %r1243, -2139062144;
and.b32 %r1245, %r299, 2139062143;
sub.s32 %r1246, %r1244, %r1245;
xor.b32 %r304, %r938, %r1246;
xor.b32 %r1247, %r301, %r920;
or.b32 %r1248, %r1247, -2139062144;
and.b32 %r1249, %r301, 2139062143;
sub.s32 %r1250, %r1248, %r1249;
xor.b32 %r308, %r943, %r1250;
ld.global.nc.u32 %r305, [%rd102+12];
ld.global.nc.u32 %r309, [%rd102+16];

dp4a.s32.s32 %r303, %r304, %r305, %r295;


dp4a.s32.s32 %r307, %r308, %r309, %r303;

mov.u32 %r311, %r801;

prmt.b32 %r311, %r311, 0, 0xba98;

mov.u32 %r313, %r803;

prmt.b32 %r313, %r313, 0, 0xba98;

xor.b32 %r1251, %r311, %r949;
or.b32 %r1252, %r1251, -2139062144;
and.b32 %r1253, %r311, 2139062143;
sub.s32 %r1254, %r1252, %r1253;
xor.b32 %r316, %r973, %r1254;
xor.b32 %r1255, %r313, %r955;
or.b32 %r1256, %r1255, -2139062144;
and.b32 %r1257, %r313, 2139062143;
sub.s32 %r1258, %r1256, %r1257;
xor.b32 %r320, %r978, %r1258;
ld.global.nc.u32 %r317, [%rd102+20];
ld.global.nc.u32 %r321, [%rd102+24];

dp4a.s32.s32 %r315, %r316, %r317, %r307;


dp4a.s32.s32 %r319, %r320, %r321, %r315;

mov.u32 %r323, %r813;

prmt.b32 %r323, %r323, 0, 0xba98;

mov.u32 %r325, %r815;

prmt.b32 %r325, %r325, 0, 0xba98;

xor.b32 %r1259, %r323, %r984;
or.b32 %r1260, %r1259, -2139062144;
and.b32 %r1261, %r323, 2139062143;
sub.s32 %r1262, %r1260, %r1261;
xor.b32 %r328, %r1008, %r1262;
xor.b32 %r1263, %r325, %r990;
or.b32 %r1264, %r1263, -2139062144;
and.b32 %r1265, %r325, 2139062143;
sub.s32 %r1266, %r1264, %r1265;
xor.b32 %r332, %r1013, %r1266;
ld.global.nc.u32 %r329, [%rd102+28];
ld.global.nc.u32 %r333, [%rd102+32];

dp4a.s32.s32 %r327, %r328, %r329, %r319;


dp4a.s32.s32 %r331, %r332, %r333, %r327;

mul.lo.s32 %r1267, %r1020, %r331;

{ cvt.f32.f16 %f105, %rs1;}


ld.global.nc.u32 %r335, [%rd102];

{.reg .f16 low,high;
mov.b32 {low,high},%r335;
cvt.f32.f16 %f106, low;}


mul.ftz.f32 %f137, %f105, %f106;
cvt.rn.f32.s32 %f138, %r1267;
fma.rn.ftz.f32 %f364, %f137, %f138, %f364;
mov.u32 %r336, %r826;

prmt.b32 %r336, %r336, 0, 0xba98;

mov.u32 %r338, %r828;

prmt.b32 %r338, %r338, 0, 0xba98;

xor.b32 %r1268, %r336, %r1026;
or.b32 %r1269, %r1268, -2139062144;
and.b32 %r1270, %r336, 2139062143;
sub.s32 %r1271, %r1269, %r1270;
xor.b32 %r341, %r1050, %r1271;
xor.b32 %r1272, %r338, %r1032;
or.b32 %r1273, %r1272, -2139062144;
and.b32 %r1274, %r338, 2139062143;
sub.s32 %r1275, %r1273, %r1274;
xor.b32 %r345, %r1055, %r1275;

dp4a.s32.s32 %r340, %r341, %r293, %r833;


dp4a.s32.s32 %r344, %r345, %r297, %r340;

mov.u32 %r348, %r838;

prmt.b32 %r348, %r348, 0, 0xba98;

mov.u32 %r350, %r840;

prmt.b32 %r350, %r350, 0, 0xba98;

xor.b32 %r1276, %r348, %r1061;
or.b32 %r1277, %r1276, -2139062144;
and.b32 %r1278, %r348, 2139062143;
sub.s32 %r1279, %r1277, %r1278;
xor.b32 %r353, %r1085, %r1279;
xor.b32 %r1280, %r350, %r1067;
or.b32 %r1281, %r1280, -2139062144;
and.b32 %r1282, %r350, 2139062143;
sub.s32 %r1283, %r1281, %r1282;
xor.b32 %r357, %r1090, %r1283;

dp4a.s32.s32 %r352, %r353, %r305, %r344;


dp4a.s32.s32 %r356, %r357, %r309, %r352;

mov.u32 %r360, %r850;

prmt.b32 %r360, %r360, 0, 0xba98;

mov.u32 %r362, %r852;

prmt.b32 %r362, %r362, 0, 0xba98;

xor.b32 %r1284, %r360, %r1096;
or.b32 %r1285, %r1284, -2139062144;
and.b32 %r1286, %r360, 2139062143;
sub.s32 %r1287, %r1285, %r1286;
xor.b32 %r365, %r1120, %r1287;
xor.b32 %r1288, %r362, %r1102;
or.b32 %r1289, %r1288, -2139062144;
and.b32 %r1290, %r362, 2139062143;
sub.s32 %r1291, %r1289, %r1290;
xor.b32 %r369, %r1125, %r1291;

dp4a.s32.s32 %r364, %r365, %r317, %r356;


dp4a.s32.s32 %r368, %r369, %r321, %r364;

mov.u32 %r372, %r862;

prmt.b32 %r372, %r372, 0, 0xba98;

mov.u32 %r374, %r864;

prmt.b32 %r374, %r374, 0, 0xba98;

xor.b32 %r1292, %r372, %r1131;
or.b32 %r1293, %r1292, -2139062144;
and.b32 %r1294, %r372, 2139062143;
sub.s32 %r1295, %r1293, %r1294;
xor.b32 %r377, %r1155, %r1295;
xor.b32 %r1296, %r374, %r1137;
or.b32 %r1297, %r1296, -2139062144;
and.b32 %r1298, %r374, 2139062143;
sub.s32 %r1299, %r1297, %r1298;
xor.b32 %r381, %r1160, %r1299;

dp4a.s32.s32 %r376, %r377, %r329, %r368;


dp4a.s32.s32 %r380, %r381, %r333, %r376;

mul.lo.s32 %r1300, %r1167, %r380;

{ cvt.f32.f16 %f107, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r335;
cvt.f32.f16 %f108, low;}


mul.ftz.f32 %f139, %f107, %f108;
cvt.rn.f32.s32 %f140, %r1300;
fma.rn.ftz.f32 %f363, %f139, %f140, %f363;
cvt.u64.u32 %rd103, %r1933;
add.s64 %rd104, %rd1, %rd103;
add.s64 %rd105, %rd104, %rd3;
mul.lo.s64 %rd106, %rd105, 36;
mov.u32 %r385, %r777;

prmt.b32 %r385, %r385, 0, 0xba98;

mov.u32 %r387, %r779;

prmt.b32 %r387, %r387, 0, 0xba98;

xor.b32 %r1301, %r385, %r879;
or.b32 %r1302, %r1301, -2139062144;
and.b32 %r1303, %r385, 2139062143;
sub.s32 %r1304, %r1302, %r1303;
xor.b32 %r390, %r903, %r1304;
xor.b32 %r1305, %r387, %r885;
or.b32 %r1306, %r1305, -2139062144;
and.b32 %r1307, %r387, 2139062143;
sub.s32 %r1308, %r1306, %r1307;
xor.b32 %r394, %r908, %r1308;
add.s64 %rd107, %rd16, %rd106;
ld.global.nc.u32 %r391, [%rd107+4];
ld.global.nc.u32 %r395, [%rd107+8];

dp4a.s32.s32 %r389, %r390, %r391, %r833;


dp4a.s32.s32 %r393, %r394, %r395, %r389;

mov.u32 %r397, %r789;

prmt.b32 %r397, %r397, 0, 0xba98;

mov.u32 %r399, %r791;

prmt.b32 %r399, %r399, 0, 0xba98;

xor.b32 %r1309, %r397, %r914;
or.b32 %r1310, %r1309, -2139062144;
and.b32 %r1311, %r397, 2139062143;
sub.s32 %r1312, %r1310, %r1311;
xor.b32 %r402, %r938, %r1312;
xor.b32 %r1313, %r399, %r920;
or.b32 %r1314, %r1313, -2139062144;
and.b32 %r1315, %r399, 2139062143;
sub.s32 %r1316, %r1314, %r1315;
xor.b32 %r406, %r943, %r1316;
ld.global.nc.u32 %r403, [%rd107+12];
ld.global.nc.u32 %r407, [%rd107+16];

dp4a.s32.s32 %r401, %r402, %r403, %r393;


dp4a.s32.s32 %r405, %r406, %r407, %r401;

mov.u32 %r409, %r801;

prmt.b32 %r409, %r409, 0, 0xba98;

mov.u32 %r411, %r803;

prmt.b32 %r411, %r411, 0, 0xba98;

xor.b32 %r1317, %r409, %r949;
or.b32 %r1318, %r1317, -2139062144;
and.b32 %r1319, %r409, 2139062143;
sub.s32 %r1320, %r1318, %r1319;
xor.b32 %r414, %r973, %r1320;
xor.b32 %r1321, %r411, %r955;
or.b32 %r1322, %r1321, -2139062144;
and.b32 %r1323, %r411, 2139062143;
sub.s32 %r1324, %r1322, %r1323;
xor.b32 %r418, %r978, %r1324;
ld.global.nc.u32 %r415, [%rd107+20];
ld.global.nc.u32 %r419, [%rd107+24];

dp4a.s32.s32 %r413, %r414, %r415, %r405;


dp4a.s32.s32 %r417, %r418, %r419, %r413;

mov.u32 %r421, %r813;

prmt.b32 %r421, %r421, 0, 0xba98;

mov.u32 %r423, %r815;

prmt.b32 %r423, %r423, 0, 0xba98;

xor.b32 %r1325, %r421, %r984;
or.b32 %r1326, %r1325, -2139062144;
and.b32 %r1327, %r421, 2139062143;
sub.s32 %r1328, %r1326, %r1327;
xor.b32 %r426, %r1008, %r1328;
xor.b32 %r1329, %r423, %r990;
or.b32 %r1330, %r1329, -2139062144;
and.b32 %r1331, %r423, 2139062143;
sub.s32 %r1332, %r1330, %r1331;
xor.b32 %r430, %r1013, %r1332;
ld.global.nc.u32 %r427, [%rd107+28];
ld.global.nc.u32 %r431, [%rd107+32];

dp4a.s32.s32 %r425, %r426, %r427, %r417;


dp4a.s32.s32 %r429, %r430, %r431, %r425;

mul.lo.s32 %r1333, %r1020, %r429;

{ cvt.f32.f16 %f109, %rs1;}


ld.global.nc.u32 %r433, [%rd107];

{.reg .f16 low,high;
mov.b32 {low,high},%r433;
cvt.f32.f16 %f110, low;}


mul.ftz.f32 %f141, %f109, %f110;
cvt.rn.f32.s32 %f142, %r1333;
fma.rn.ftz.f32 %f362, %f141, %f142, %f362;
mov.u32 %r434, %r826;

prmt.b32 %r434, %r434, 0, 0xba98;

mov.u32 %r436, %r828;

prmt.b32 %r436, %r436, 0, 0xba98;

xor.b32 %r1334, %r434, %r1026;
or.b32 %r1335, %r1334, -2139062144;
and.b32 %r1336, %r434, 2139062143;
sub.s32 %r1337, %r1335, %r1336;
xor.b32 %r439, %r1050, %r1337;
xor.b32 %r1338, %r436, %r1032;
or.b32 %r1339, %r1338, -2139062144;
and.b32 %r1340, %r436, 2139062143;
sub.s32 %r1341, %r1339, %r1340;
xor.b32 %r443, %r1055, %r1341;

dp4a.s32.s32 %r438, %r439, %r391, %r833;


dp4a.s32.s32 %r442, %r443, %r395, %r438;

mov.u32 %r446, %r838;

prmt.b32 %r446, %r446, 0, 0xba98;

mov.u32 %r448, %r840;

prmt.b32 %r448, %r448, 0, 0xba98;

xor.b32 %r1342, %r446, %r1061;
or.b32 %r1343, %r1342, -2139062144;
and.b32 %r1344, %r446, 2139062143;
sub.s32 %r1345, %r1343, %r1344;
xor.b32 %r451, %r1085, %r1345;
xor.b32 %r1346, %r448, %r1067;
or.b32 %r1347, %r1346, -2139062144;
and.b32 %r1348, %r448, 2139062143;
sub.s32 %r1349, %r1347, %r1348;
xor.b32 %r455, %r1090, %r1349;

dp4a.s32.s32 %r450, %r451, %r403, %r442;


dp4a.s32.s32 %r454, %r455, %r407, %r450;

mov.u32 %r458, %r850;

prmt.b32 %r458, %r458, 0, 0xba98;

mov.u32 %r460, %r852;

prmt.b32 %r460, %r460, 0, 0xba98;

xor.b32 %r1350, %r458, %r1096;
or.b32 %r1351, %r1350, -2139062144;
and.b32 %r1352, %r458, 2139062143;
sub.s32 %r1353, %r1351, %r1352;
xor.b32 %r463, %r1120, %r1353;
xor.b32 %r1354, %r460, %r1102;
or.b32 %r1355, %r1354, -2139062144;
and.b32 %r1356, %r460, 2139062143;
sub.s32 %r1357, %r1355, %r1356;
xor.b32 %r467, %r1125, %r1357;

dp4a.s32.s32 %r462, %r463, %r415, %r454;


dp4a.s32.s32 %r466, %r467, %r419, %r462;

mov.u32 %r470, %r862;

prmt.b32 %r470, %r470, 0, 0xba98;

mov.u32 %r472, %r864;

prmt.b32 %r472, %r472, 0, 0xba98;

xor.b32 %r1358, %r470, %r1131;
or.b32 %r1359, %r1358, -2139062144;
and.b32 %r1360, %r470, 2139062143;
sub.s32 %r1361, %r1359, %r1360;
xor.b32 %r475, %r1155, %r1361;
xor.b32 %r1362, %r472, %r1137;
or.b32 %r1363, %r1362, -2139062144;
and.b32 %r1364, %r472, 2139062143;
sub.s32 %r1365, %r1363, %r1364;
xor.b32 %r479, %r1160, %r1365;

dp4a.s32.s32 %r474, %r475, %r427, %r466;


dp4a.s32.s32 %r478, %r479, %r431, %r474;

mul.lo.s32 %r1366, %r1167, %r478;

{ cvt.f32.f16 %f111, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r433;
cvt.f32.f16 %f112, low;}


mul.ftz.f32 %f143, %f111, %f112;
cvt.rn.f32.s32 %f144, %r1366;
fma.rn.ftz.f32 %f361, %f143, %f144, %f361;
cvt.u64.u32 %rd108, %r1932;
add.s64 %rd109, %rd1, %rd108;
add.s64 %rd110, %rd109, %rd3;
mul.lo.s64 %rd111, %rd110, 36;
mov.u32 %r483, %r777;

prmt.b32 %r483, %r483, 0, 0xba98;

mov.u32 %r485, %r779;

prmt.b32 %r485, %r485, 0, 0xba98;

xor.b32 %r1367, %r483, %r879;
or.b32 %r1368, %r1367, -2139062144;
and.b32 %r1369, %r483, 2139062143;
sub.s32 %r1370, %r1368, %r1369;
xor.b32 %r488, %r903, %r1370;
xor.b32 %r1371, %r485, %r885;
or.b32 %r1372, %r1371, -2139062144;
and.b32 %r1373, %r485, 2139062143;
sub.s32 %r1374, %r1372, %r1373;
xor.b32 %r492, %r908, %r1374;
add.s64 %rd112, %rd16, %rd111;
ld.global.nc.u32 %r489, [%rd112+4];
ld.global.nc.u32 %r493, [%rd112+8];

dp4a.s32.s32 %r487, %r488, %r489, %r833;


dp4a.s32.s32 %r491, %r492, %r493, %r487;

mov.u32 %r495, %r789;

prmt.b32 %r495, %r495, 0, 0xba98;

mov.u32 %r497, %r791;

prmt.b32 %r497, %r497, 0, 0xba98;

xor.b32 %r1375, %r495, %r914;
or.b32 %r1376, %r1375, -2139062144;
and.b32 %r1377, %r495, 2139062143;
sub.s32 %r1378, %r1376, %r1377;
xor.b32 %r500, %r938, %r1378;
xor.b32 %r1379, %r497, %r920;
or.b32 %r1380, %r1379, -2139062144;
and.b32 %r1381, %r497, 2139062143;
sub.s32 %r1382, %r1380, %r1381;
xor.b32 %r504, %r943, %r1382;
ld.global.nc.u32 %r501, [%rd112+12];
ld.global.nc.u32 %r505, [%rd112+16];

dp4a.s32.s32 %r499, %r500, %r501, %r491;


dp4a.s32.s32 %r503, %r504, %r505, %r499;

mov.u32 %r507, %r801;

prmt.b32 %r507, %r507, 0, 0xba98;

mov.u32 %r509, %r803;

prmt.b32 %r509, %r509, 0, 0xba98;

xor.b32 %r1383, %r507, %r949;
or.b32 %r1384, %r1383, -2139062144;
and.b32 %r1385, %r507, 2139062143;
sub.s32 %r1386, %r1384, %r1385;
xor.b32 %r512, %r973, %r1386;
xor.b32 %r1387, %r509, %r955;
or.b32 %r1388, %r1387, -2139062144;
and.b32 %r1389, %r509, 2139062143;
sub.s32 %r1390, %r1388, %r1389;
xor.b32 %r516, %r978, %r1390;
ld.global.nc.u32 %r513, [%rd112+20];
ld.global.nc.u32 %r517, [%rd112+24];

dp4a.s32.s32 %r511, %r512, %r513, %r503;


dp4a.s32.s32 %r515, %r516, %r517, %r511;

mov.u32 %r519, %r813;

prmt.b32 %r519, %r519, 0, 0xba98;

mov.u32 %r521, %r815;

prmt.b32 %r521, %r521, 0, 0xba98;

xor.b32 %r1391, %r519, %r984;
or.b32 %r1392, %r1391, -2139062144;
and.b32 %r1393, %r519, 2139062143;
sub.s32 %r1394, %r1392, %r1393;
xor.b32 %r524, %r1008, %r1394;
xor.b32 %r1395, %r521, %r990;
or.b32 %r1396, %r1395, -2139062144;
and.b32 %r1397, %r521, 2139062143;
sub.s32 %r1398, %r1396, %r1397;
xor.b32 %r528, %r1013, %r1398;
ld.global.nc.u32 %r525, [%rd112+28];
ld.global.nc.u32 %r529, [%rd112+32];

dp4a.s32.s32 %r523, %r524, %r525, %r515;


dp4a.s32.s32 %r527, %r528, %r529, %r523;

mul.lo.s32 %r1399, %r1020, %r527;

{ cvt.f32.f16 %f113, %rs1;}


ld.global.nc.u32 %r531, [%rd112];

{.reg .f16 low,high;
mov.b32 {low,high},%r531;
cvt.f32.f16 %f114, low;}


mul.ftz.f32 %f145, %f113, %f114;
cvt.rn.f32.s32 %f146, %r1399;
fma.rn.ftz.f32 %f360, %f145, %f146, %f360;
mov.u32 %r532, %r826;

prmt.b32 %r532, %r532, 0, 0xba98;

mov.u32 %r534, %r828;

prmt.b32 %r534, %r534, 0, 0xba98;

xor.b32 %r1400, %r532, %r1026;
or.b32 %r1401, %r1400, -2139062144;
and.b32 %r1402, %r532, 2139062143;
sub.s32 %r1403, %r1401, %r1402;
xor.b32 %r537, %r1050, %r1403;
xor.b32 %r1404, %r534, %r1032;
or.b32 %r1405, %r1404, -2139062144;
and.b32 %r1406, %r534, 2139062143;
sub.s32 %r1407, %r1405, %r1406;
xor.b32 %r541, %r1055, %r1407;

dp4a.s32.s32 %r536, %r537, %r489, %r833;


dp4a.s32.s32 %r540, %r541, %r493, %r536;

mov.u32 %r544, %r838;

prmt.b32 %r544, %r544, 0, 0xba98;

mov.u32 %r546, %r840;

prmt.b32 %r546, %r546, 0, 0xba98;

xor.b32 %r1408, %r544, %r1061;
or.b32 %r1409, %r1408, -2139062144;
and.b32 %r1410, %r544, 2139062143;
sub.s32 %r1411, %r1409, %r1410;
xor.b32 %r549, %r1085, %r1411;
xor.b32 %r1412, %r546, %r1067;
or.b32 %r1413, %r1412, -2139062144;
and.b32 %r1414, %r546, 2139062143;
sub.s32 %r1415, %r1413, %r1414;
xor.b32 %r553, %r1090, %r1415;

dp4a.s32.s32 %r548, %r549, %r501, %r540;


dp4a.s32.s32 %r552, %r553, %r505, %r548;

mov.u32 %r556, %r850;

prmt.b32 %r556, %r556, 0, 0xba98;

mov.u32 %r558, %r852;

prmt.b32 %r558, %r558, 0, 0xba98;

xor.b32 %r1416, %r556, %r1096;
or.b32 %r1417, %r1416, -2139062144;
and.b32 %r1418, %r556, 2139062143;
sub.s32 %r1419, %r1417, %r1418;
xor.b32 %r561, %r1120, %r1419;
xor.b32 %r1420, %r558, %r1102;
or.b32 %r1421, %r1420, -2139062144;
and.b32 %r1422, %r558, 2139062143;
sub.s32 %r1423, %r1421, %r1422;
xor.b32 %r565, %r1125, %r1423;

dp4a.s32.s32 %r560, %r561, %r513, %r552;


dp4a.s32.s32 %r564, %r565, %r517, %r560;

mov.u32 %r568, %r862;

prmt.b32 %r568, %r568, 0, 0xba98;

mov.u32 %r570, %r864;

prmt.b32 %r570, %r570, 0, 0xba98;

xor.b32 %r1424, %r568, %r1131;
or.b32 %r1425, %r1424, -2139062144;
and.b32 %r1426, %r568, 2139062143;
sub.s32 %r1427, %r1425, %r1426;
xor.b32 %r573, %r1155, %r1427;
xor.b32 %r1428, %r570, %r1137;
or.b32 %r1429, %r1428, -2139062144;
and.b32 %r1430, %r570, 2139062143;
sub.s32 %r1431, %r1429, %r1430;
xor.b32 %r577, %r1160, %r1431;

dp4a.s32.s32 %r572, %r573, %r525, %r564;


dp4a.s32.s32 %r576, %r577, %r529, %r572;

mul.lo.s32 %r1432, %r1167, %r576;

{ cvt.f32.f16 %f115, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r531;
cvt.f32.f16 %f116, low;}


mul.ftz.f32 %f147, %f115, %f116;
cvt.rn.f32.s32 %f148, %r1432;
fma.rn.ftz.f32 %f359, %f147, %f148, %f359;
cvt.u64.u32 %rd113, %r1931;
add.s64 %rd114, %rd1, %rd113;
add.s64 %rd115, %rd114, %rd3;
mul.lo.s64 %rd116, %rd115, 36;
mov.u32 %r581, %r777;

prmt.b32 %r581, %r581, 0, 0xba98;

mov.u32 %r583, %r779;

prmt.b32 %r583, %r583, 0, 0xba98;

xor.b32 %r1433, %r581, %r879;
or.b32 %r1434, %r1433, -2139062144;
and.b32 %r1435, %r581, 2139062143;
sub.s32 %r1436, %r1434, %r1435;
xor.b32 %r586, %r903, %r1436;
xor.b32 %r1437, %r583, %r885;
or.b32 %r1438, %r1437, -2139062144;
and.b32 %r1439, %r583, 2139062143;
sub.s32 %r1440, %r1438, %r1439;
xor.b32 %r590, %r908, %r1440;
add.s64 %rd117, %rd16, %rd116;
ld.global.nc.u32 %r587, [%rd117+4];
ld.global.nc.u32 %r591, [%rd117+8];

dp4a.s32.s32 %r585, %r586, %r587, %r833;


dp4a.s32.s32 %r589, %r590, %r591, %r585;

mov.u32 %r593, %r789;

prmt.b32 %r593, %r593, 0, 0xba98;

mov.u32 %r595, %r791;

prmt.b32 %r595, %r595, 0, 0xba98;

xor.b32 %r1441, %r593, %r914;
or.b32 %r1442, %r1441, -2139062144;
and.b32 %r1443, %r593, 2139062143;
sub.s32 %r1444, %r1442, %r1443;
xor.b32 %r598, %r938, %r1444;
xor.b32 %r1445, %r595, %r920;
or.b32 %r1446, %r1445, -2139062144;
and.b32 %r1447, %r595, 2139062143;
sub.s32 %r1448, %r1446, %r1447;
xor.b32 %r602, %r943, %r1448;
ld.global.nc.u32 %r599, [%rd117+12];
ld.global.nc.u32 %r603, [%rd117+16];

dp4a.s32.s32 %r597, %r598, %r599, %r589;


dp4a.s32.s32 %r601, %r602, %r603, %r597;

mov.u32 %r605, %r801;

prmt.b32 %r605, %r605, 0, 0xba98;

mov.u32 %r607, %r803;

prmt.b32 %r607, %r607, 0, 0xba98;

xor.b32 %r1449, %r605, %r949;
or.b32 %r1450, %r1449, -2139062144;
and.b32 %r1451, %r605, 2139062143;
sub.s32 %r1452, %r1450, %r1451;
xor.b32 %r610, %r973, %r1452;
xor.b32 %r1453, %r607, %r955;
or.b32 %r1454, %r1453, -2139062144;
and.b32 %r1455, %r607, 2139062143;
sub.s32 %r1456, %r1454, %r1455;
xor.b32 %r614, %r978, %r1456;
ld.global.nc.u32 %r611, [%rd117+20];
ld.global.nc.u32 %r615, [%rd117+24];

dp4a.s32.s32 %r609, %r610, %r611, %r601;


dp4a.s32.s32 %r613, %r614, %r615, %r609;

mov.u32 %r617, %r813;

prmt.b32 %r617, %r617, 0, 0xba98;

mov.u32 %r619, %r815;

prmt.b32 %r619, %r619, 0, 0xba98;

xor.b32 %r1457, %r617, %r984;
or.b32 %r1458, %r1457, -2139062144;
and.b32 %r1459, %r617, 2139062143;
sub.s32 %r1460, %r1458, %r1459;
xor.b32 %r622, %r1008, %r1460;
xor.b32 %r1461, %r619, %r990;
or.b32 %r1462, %r1461, -2139062144;
and.b32 %r1463, %r619, 2139062143;
sub.s32 %r1464, %r1462, %r1463;
xor.b32 %r626, %r1013, %r1464;
ld.global.nc.u32 %r623, [%rd117+28];
ld.global.nc.u32 %r627, [%rd117+32];

dp4a.s32.s32 %r621, %r622, %r623, %r613;


dp4a.s32.s32 %r625, %r626, %r627, %r621;

mul.lo.s32 %r1465, %r1020, %r625;

{ cvt.f32.f16 %f117, %rs1;}


ld.global.nc.u32 %r629, [%rd117];

{.reg .f16 low,high;
mov.b32 {low,high},%r629;
cvt.f32.f16 %f118, low;}


mul.ftz.f32 %f149, %f117, %f118;
cvt.rn.f32.s32 %f150, %r1465;
fma.rn.ftz.f32 %f358, %f149, %f150, %f358;
mov.u32 %r630, %r826;

prmt.b32 %r630, %r630, 0, 0xba98;

mov.u32 %r632, %r828;

prmt.b32 %r632, %r632, 0, 0xba98;

xor.b32 %r1466, %r630, %r1026;
or.b32 %r1467, %r1466, -2139062144;
and.b32 %r1468, %r630, 2139062143;
sub.s32 %r1469, %r1467, %r1468;
xor.b32 %r635, %r1050, %r1469;
xor.b32 %r1470, %r632, %r1032;
or.b32 %r1471, %r1470, -2139062144;
and.b32 %r1472, %r632, 2139062143;
sub.s32 %r1473, %r1471, %r1472;
xor.b32 %r639, %r1055, %r1473;

dp4a.s32.s32 %r634, %r635, %r587, %r833;


dp4a.s32.s32 %r638, %r639, %r591, %r634;

mov.u32 %r642, %r838;

prmt.b32 %r642, %r642, 0, 0xba98;

mov.u32 %r644, %r840;

prmt.b32 %r644, %r644, 0, 0xba98;

xor.b32 %r1474, %r642, %r1061;
or.b32 %r1475, %r1474, -2139062144;
and.b32 %r1476, %r642, 2139062143;
sub.s32 %r1477, %r1475, %r1476;
xor.b32 %r647, %r1085, %r1477;
xor.b32 %r1478, %r644, %r1067;
or.b32 %r1479, %r1478, -2139062144;
and.b32 %r1480, %r644, 2139062143;
sub.s32 %r1481, %r1479, %r1480;
xor.b32 %r651, %r1090, %r1481;

dp4a.s32.s32 %r646, %r647, %r599, %r638;


dp4a.s32.s32 %r650, %r651, %r603, %r646;

mov.u32 %r654, %r850;

prmt.b32 %r654, %r654, 0, 0xba98;

mov.u32 %r656, %r852;

prmt.b32 %r656, %r656, 0, 0xba98;

xor.b32 %r1482, %r654, %r1096;
or.b32 %r1483, %r1482, -2139062144;
and.b32 %r1484, %r654, 2139062143;
sub.s32 %r1485, %r1483, %r1484;
xor.b32 %r659, %r1120, %r1485;
xor.b32 %r1486, %r656, %r1102;
or.b32 %r1487, %r1486, -2139062144;
and.b32 %r1488, %r656, 2139062143;
sub.s32 %r1489, %r1487, %r1488;
xor.b32 %r663, %r1125, %r1489;

dp4a.s32.s32 %r658, %r659, %r611, %r650;


dp4a.s32.s32 %r662, %r663, %r615, %r658;

mov.u32 %r666, %r862;

prmt.b32 %r666, %r666, 0, 0xba98;

mov.u32 %r668, %r864;

prmt.b32 %r668, %r668, 0, 0xba98;

xor.b32 %r1490, %r666, %r1131;
or.b32 %r1491, %r1490, -2139062144;
and.b32 %r1492, %r666, 2139062143;
sub.s32 %r1493, %r1491, %r1492;
xor.b32 %r671, %r1155, %r1493;
xor.b32 %r1494, %r668, %r1137;
or.b32 %r1495, %r1494, -2139062144;
and.b32 %r1496, %r668, 2139062143;
sub.s32 %r1497, %r1495, %r1496;
xor.b32 %r675, %r1160, %r1497;

dp4a.s32.s32 %r670, %r671, %r623, %r662;


dp4a.s32.s32 %r674, %r675, %r627, %r670;

mul.lo.s32 %r1498, %r1167, %r674;

{ cvt.f32.f16 %f119, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r629;
cvt.f32.f16 %f120, low;}


mul.ftz.f32 %f151, %f119, %f120;
cvt.rn.f32.s32 %f152, %r1498;
fma.rn.ftz.f32 %f357, %f151, %f152, %f357;
cvt.u64.u32 %rd118, %r1930;
add.s64 %rd119, %rd1, %rd118;
add.s64 %rd120, %rd119, %rd3;
mul.lo.s64 %rd121, %rd120, 36;
mov.u32 %r679, %r777;

prmt.b32 %r679, %r679, 0, 0xba98;

mov.u32 %r681, %r779;

prmt.b32 %r681, %r681, 0, 0xba98;

xor.b32 %r1499, %r679, %r879;
or.b32 %r1500, %r1499, -2139062144;
and.b32 %r1501, %r679, 2139062143;
sub.s32 %r1502, %r1500, %r1501;
xor.b32 %r684, %r903, %r1502;
xor.b32 %r1503, %r681, %r885;
or.b32 %r1504, %r1503, -2139062144;
and.b32 %r1505, %r681, 2139062143;
sub.s32 %r1506, %r1504, %r1505;
xor.b32 %r688, %r908, %r1506;
add.s64 %rd122, %rd16, %rd121;
ld.global.nc.u32 %r685, [%rd122+4];
ld.global.nc.u32 %r689, [%rd122+8];

dp4a.s32.s32 %r683, %r684, %r685, %r833;


dp4a.s32.s32 %r687, %r688, %r689, %r683;

mov.u32 %r691, %r789;

prmt.b32 %r691, %r691, 0, 0xba98;

mov.u32 %r693, %r791;

prmt.b32 %r693, %r693, 0, 0xba98;

xor.b32 %r1507, %r691, %r914;
or.b32 %r1508, %r1507, -2139062144;
and.b32 %r1509, %r691, 2139062143;
sub.s32 %r1510, %r1508, %r1509;
xor.b32 %r696, %r938, %r1510;
xor.b32 %r1511, %r693, %r920;
or.b32 %r1512, %r1511, -2139062144;
and.b32 %r1513, %r693, 2139062143;
sub.s32 %r1514, %r1512, %r1513;
xor.b32 %r700, %r943, %r1514;
ld.global.nc.u32 %r697, [%rd122+12];
ld.global.nc.u32 %r701, [%rd122+16];

dp4a.s32.s32 %r695, %r696, %r697, %r687;


dp4a.s32.s32 %r699, %r700, %r701, %r695;

mov.u32 %r703, %r801;

prmt.b32 %r703, %r703, 0, 0xba98;

mov.u32 %r705, %r803;

prmt.b32 %r705, %r705, 0, 0xba98;

xor.b32 %r1515, %r703, %r949;
or.b32 %r1516, %r1515, -2139062144;
and.b32 %r1517, %r703, 2139062143;
sub.s32 %r1518, %r1516, %r1517;
xor.b32 %r708, %r973, %r1518;
xor.b32 %r1519, %r705, %r955;
or.b32 %r1520, %r1519, -2139062144;
and.b32 %r1521, %r705, 2139062143;
sub.s32 %r1522, %r1520, %r1521;
xor.b32 %r712, %r978, %r1522;
ld.global.nc.u32 %r709, [%rd122+20];
ld.global.nc.u32 %r713, [%rd122+24];

dp4a.s32.s32 %r707, %r708, %r709, %r699;


dp4a.s32.s32 %r711, %r712, %r713, %r707;

mov.u32 %r715, %r813;

prmt.b32 %r715, %r715, 0, 0xba98;

mov.u32 %r717, %r815;

prmt.b32 %r717, %r717, 0, 0xba98;

xor.b32 %r1523, %r715, %r984;
or.b32 %r1524, %r1523, -2139062144;
and.b32 %r1525, %r715, 2139062143;
sub.s32 %r1526, %r1524, %r1525;
xor.b32 %r720, %r1008, %r1526;
xor.b32 %r1527, %r717, %r990;
or.b32 %r1528, %r1527, -2139062144;
and.b32 %r1529, %r717, 2139062143;
sub.s32 %r1530, %r1528, %r1529;
xor.b32 %r724, %r1013, %r1530;
ld.global.nc.u32 %r721, [%rd122+28];
ld.global.nc.u32 %r725, [%rd122+32];

dp4a.s32.s32 %r719, %r720, %r721, %r711;


dp4a.s32.s32 %r723, %r724, %r725, %r719;

mul.lo.s32 %r1531, %r1020, %r723;

{ cvt.f32.f16 %f121, %rs1;}


ld.global.nc.u32 %r727, [%rd122];

{.reg .f16 low,high;
mov.b32 {low,high},%r727;
cvt.f32.f16 %f122, low;}


mul.ftz.f32 %f153, %f121, %f122;
cvt.rn.f32.s32 %f154, %r1531;
fma.rn.ftz.f32 %f356, %f153, %f154, %f356;
mov.u32 %r728, %r826;

prmt.b32 %r728, %r728, 0, 0xba98;

mov.u32 %r730, %r828;

prmt.b32 %r730, %r730, 0, 0xba98;

xor.b32 %r1532, %r728, %r1026;
or.b32 %r1533, %r1532, -2139062144;
and.b32 %r1534, %r728, 2139062143;
sub.s32 %r1535, %r1533, %r1534;
xor.b32 %r733, %r1050, %r1535;
xor.b32 %r1536, %r730, %r1032;
or.b32 %r1537, %r1536, -2139062144;
and.b32 %r1538, %r730, 2139062143;
sub.s32 %r1539, %r1537, %r1538;
xor.b32 %r737, %r1055, %r1539;

dp4a.s32.s32 %r732, %r733, %r685, %r833;


dp4a.s32.s32 %r736, %r737, %r689, %r732;

mov.u32 %r740, %r838;

prmt.b32 %r740, %r740, 0, 0xba98;

mov.u32 %r742, %r840;

prmt.b32 %r742, %r742, 0, 0xba98;

xor.b32 %r1540, %r740, %r1061;
or.b32 %r1541, %r1540, -2139062144;
and.b32 %r1542, %r740, 2139062143;
sub.s32 %r1543, %r1541, %r1542;
xor.b32 %r745, %r1085, %r1543;
xor.b32 %r1544, %r742, %r1067;
or.b32 %r1545, %r1544, -2139062144;
and.b32 %r1546, %r742, 2139062143;
sub.s32 %r1547, %r1545, %r1546;
xor.b32 %r749, %r1090, %r1547;

dp4a.s32.s32 %r744, %r745, %r697, %r736;


dp4a.s32.s32 %r748, %r749, %r701, %r744;

mov.u32 %r752, %r850;

prmt.b32 %r752, %r752, 0, 0xba98;

mov.u32 %r754, %r852;

prmt.b32 %r754, %r754, 0, 0xba98;

xor.b32 %r1548, %r752, %r1096;
or.b32 %r1549, %r1548, -2139062144;
and.b32 %r1550, %r752, 2139062143;
sub.s32 %r1551, %r1549, %r1550;
xor.b32 %r757, %r1120, %r1551;
xor.b32 %r1552, %r754, %r1102;
or.b32 %r1553, %r1552, -2139062144;
and.b32 %r1554, %r754, 2139062143;
sub.s32 %r1555, %r1553, %r1554;
xor.b32 %r761, %r1125, %r1555;

dp4a.s32.s32 %r756, %r757, %r709, %r748;


dp4a.s32.s32 %r760, %r761, %r713, %r756;

mov.u32 %r764, %r862;

prmt.b32 %r764, %r764, 0, 0xba98;

mov.u32 %r766, %r864;

prmt.b32 %r766, %r766, 0, 0xba98;

xor.b32 %r1556, %r764, %r1131;
or.b32 %r1557, %r1556, -2139062144;
and.b32 %r1558, %r764, 2139062143;
sub.s32 %r1559, %r1557, %r1558;
xor.b32 %r769, %r1155, %r1559;
xor.b32 %r1560, %r766, %r1137;
or.b32 %r1561, %r1560, -2139062144;
and.b32 %r1562, %r766, 2139062143;
sub.s32 %r1563, %r1561, %r1562;
xor.b32 %r773, %r1160, %r1563;

dp4a.s32.s32 %r768, %r769, %r721, %r760;


dp4a.s32.s32 %r772, %r773, %r725, %r768;

mul.lo.s32 %r1564, %r1167, %r772;

{ cvt.f32.f16 %f123, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r727;
cvt.f32.f16 %f124, low;}


mul.ftz.f32 %f155, %f123, %f124;
cvt.rn.f32.s32 %f156, %r1564;
fma.rn.ftz.f32 %f355, %f155, %f156, %f355;
cvt.u64.u32 %rd123, %r1929;
add.s64 %rd124, %rd1, %rd123;
add.s64 %rd125, %rd124, %rd3;
mul.lo.s64 %rd126, %rd125, 36;

prmt.b32 %r777, %r777, 0, 0xba98;


prmt.b32 %r779, %r779, 0, 0xba98;

xor.b32 %r1565, %r777, %r879;
or.b32 %r1566, %r1565, -2139062144;
and.b32 %r1567, %r777, 2139062143;
sub.s32 %r1568, %r1566, %r1567;
xor.b32 %r782, %r903, %r1568;
xor.b32 %r1569, %r779, %r885;
or.b32 %r1570, %r1569, -2139062144;
and.b32 %r1571, %r779, 2139062143;
sub.s32 %r1572, %r1570, %r1571;
xor.b32 %r786, %r908, %r1572;
add.s64 %rd127, %rd16, %rd126;
ld.global.nc.u32 %r783, [%rd127+4];
ld.global.nc.u32 %r787, [%rd127+8];

dp4a.s32.s32 %r781, %r782, %r783, %r833;


dp4a.s32.s32 %r785, %r786, %r787, %r781;


prmt.b32 %r789, %r789, 0, 0xba98;


prmt.b32 %r791, %r791, 0, 0xba98;

xor.b32 %r1573, %r789, %r914;
or.b32 %r1574, %r1573, -2139062144;
and.b32 %r1575, %r789, 2139062143;
sub.s32 %r1576, %r1574, %r1575;
xor.b32 %r794, %r938, %r1576;
xor.b32 %r1577, %r791, %r920;
or.b32 %r1578, %r1577, -2139062144;
and.b32 %r1579, %r791, 2139062143;
sub.s32 %r1580, %r1578, %r1579;
xor.b32 %r798, %r943, %r1580;
ld.global.nc.u32 %r795, [%rd127+12];
ld.global.nc.u32 %r799, [%rd127+16];

dp4a.s32.s32 %r793, %r794, %r795, %r785;


dp4a.s32.s32 %r797, %r798, %r799, %r793;


prmt.b32 %r801, %r801, 0, 0xba98;


prmt.b32 %r803, %r803, 0, 0xba98;

xor.b32 %r1581, %r801, %r949;
or.b32 %r1582, %r1581, -2139062144;
and.b32 %r1583, %r801, 2139062143;
sub.s32 %r1584, %r1582, %r1583;
xor.b32 %r806, %r973, %r1584;
xor.b32 %r1585, %r803, %r955;
or.b32 %r1586, %r1585, -2139062144;
and.b32 %r1587, %r803, 2139062143;
sub.s32 %r1588, %r1586, %r1587;
xor.b32 %r810, %r978, %r1588;
ld.global.nc.u32 %r807, [%rd127+20];
ld.global.nc.u32 %r811, [%rd127+24];

dp4a.s32.s32 %r805, %r806, %r807, %r797;


dp4a.s32.s32 %r809, %r810, %r811, %r805;


prmt.b32 %r813, %r813, 0, 0xba98;


prmt.b32 %r815, %r815, 0, 0xba98;

xor.b32 %r1589, %r813, %r984;
or.b32 %r1590, %r1589, -2139062144;
and.b32 %r1591, %r813, 2139062143;
sub.s32 %r1592, %r1590, %r1591;
xor.b32 %r818, %r1008, %r1592;
xor.b32 %r1593, %r815, %r990;
or.b32 %r1594, %r1593, -2139062144;
and.b32 %r1595, %r815, 2139062143;
sub.s32 %r1596, %r1594, %r1595;
xor.b32 %r822, %r1013, %r1596;
ld.global.nc.u32 %r819, [%rd127+28];
ld.global.nc.u32 %r823, [%rd127+32];

dp4a.s32.s32 %r817, %r818, %r819, %r809;


dp4a.s32.s32 %r821, %r822, %r823, %r817;

mul.lo.s32 %r1597, %r1020, %r821;

{ cvt.f32.f16 %f125, %rs1;}


ld.global.nc.u32 %r825, [%rd127];

{.reg .f16 low,high;
mov.b32 {low,high},%r825;
cvt.f32.f16 %f126, low;}


mul.ftz.f32 %f157, %f125, %f126;
cvt.rn.f32.s32 %f158, %r1597;
fma.rn.ftz.f32 %f354, %f157, %f158, %f354;

prmt.b32 %r826, %r826, 0, 0xba98;


prmt.b32 %r828, %r828, 0, 0xba98;

xor.b32 %r1598, %r826, %r1026;
or.b32 %r1599, %r1598, -2139062144;
and.b32 %r1600, %r826, 2139062143;
sub.s32 %r1601, %r1599, %r1600;
xor.b32 %r831, %r1050, %r1601;
xor.b32 %r1602, %r828, %r1032;
or.b32 %r1603, %r1602, -2139062144;
and.b32 %r1604, %r828, 2139062143;
sub.s32 %r1605, %r1603, %r1604;
xor.b32 %r835, %r1055, %r1605;

dp4a.s32.s32 %r830, %r831, %r783, %r833;


dp4a.s32.s32 %r834, %r835, %r787, %r830;


prmt.b32 %r838, %r838, 0, 0xba98;


prmt.b32 %r840, %r840, 0, 0xba98;

xor.b32 %r1606, %r838, %r1061;
or.b32 %r1607, %r1606, -2139062144;
and.b32 %r1608, %r838, 2139062143;
sub.s32 %r1609, %r1607, %r1608;
xor.b32 %r843, %r1085, %r1609;
xor.b32 %r1610, %r840, %r1067;
or.b32 %r1611, %r1610, -2139062144;
and.b32 %r1612, %r840, 2139062143;
sub.s32 %r1613, %r1611, %r1612;
xor.b32 %r847, %r1090, %r1613;

dp4a.s32.s32 %r842, %r843, %r795, %r834;


dp4a.s32.s32 %r846, %r847, %r799, %r842;


prmt.b32 %r850, %r850, 0, 0xba98;


prmt.b32 %r852, %r852, 0, 0xba98;

xor.b32 %r1614, %r850, %r1096;
or.b32 %r1615, %r1614, -2139062144;
and.b32 %r1616, %r850, 2139062143;
sub.s32 %r1617, %r1615, %r1616;
xor.b32 %r855, %r1120, %r1617;
xor.b32 %r1618, %r852, %r1102;
or.b32 %r1619, %r1618, -2139062144;
and.b32 %r1620, %r852, 2139062143;
sub.s32 %r1621, %r1619, %r1620;
xor.b32 %r859, %r1125, %r1621;

dp4a.s32.s32 %r854, %r855, %r807, %r846;


dp4a.s32.s32 %r858, %r859, %r811, %r854;


prmt.b32 %r862, %r862, 0, 0xba98;


prmt.b32 %r864, %r864, 0, 0xba98;

xor.b32 %r1622, %r862, %r1131;
or.b32 %r1623, %r1622, -2139062144;
and.b32 %r1624, %r862, 2139062143;
sub.s32 %r1625, %r1623, %r1624;
xor.b32 %r867, %r1155, %r1625;
xor.b32 %r1626, %r864, %r1137;
or.b32 %r1627, %r1626, -2139062144;
and.b32 %r1628, %r864, 2139062143;
sub.s32 %r1629, %r1627, %r1628;
xor.b32 %r871, %r1160, %r1629;

dp4a.s32.s32 %r866, %r867, %r819, %r858;


dp4a.s32.s32 %r870, %r871, %r823, %r866;

mul.lo.s32 %r1630, %r1167, %r870;

{ cvt.f32.f16 %f127, %rs2;}



{.reg .f16 low,high;
mov.b32 {low,high},%r825;
cvt.f32.f16 %f128, low;}


mul.ftz.f32 %f159, %f127, %f128;
cvt.rn.f32.s32 %f160, %r1630;
fma.rn.ftz.f32 %f353, %f159, %f160, %f353;
add.s32 %r1934, %r1934, 64;
add.s32 %r1933, %r1933, 64;
add.s32 %r1932, %r1932, 64;
add.s32 %r1931, %r1931, 64;
add.s32 %r1930, %r1930, 64;
add.s32 %r1929, %r1929, 64;
add.s32 %r1928, %r1928, 64;
add.s32 %r1927, %r1927, 64;
add.s64 %rd182, %rd182, 880;
add.s32 %r1935, %r1935, 8;
setp.lt.s32 %p4, %r1935, %r6;
@%p4 bra $L__BB316_2;

$L__BB316_3:
setp.eq.s32 %p5, %r1, 0;
@%p5 bra $L__BB316_5;

shl.b32 %r1632, %r1, 11;
mov.u32 %r1633, _ZZ13mul_mat_vec_qIL9ggml_type21ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r1634, %r1633, %r1632;
shl.b32 %r1635, %r55, 2;
add.s32 %r1636, %r1634, %r1635;
st.shared.f32 [%r1636+-2048], %f368;
st.shared.f32 [%r1636+-1920], %f367;
st.shared.f32 [%r1636+-1792], %f366;
st.shared.f32 [%r1636+-1664], %f365;
st.shared.f32 [%r1636+-1536], %f364;
st.shared.f32 [%r1636+-1408], %f363;
st.shared.f32 [%r1636+-1280], %f362;
st.shared.f32 [%r1636+-1152], %f361;
st.shared.f32 [%r1636+-1024], %f360;
st.shared.f32 [%r1636+-896], %f359;
st.shared.f32 [%r1636+-768], %f358;
st.shared.f32 [%r1636+-640], %f357;
st.shared.f32 [%r1636+-512], %f356;
st.shared.f32 [%r1636+-384], %f355;
st.shared.f32 [%r1636+-256], %f354;
st.shared.f32 [%r1636+-128], %f353;

$L__BB316_5:
setp.ne.s32 %p6, %r1, 0;
bar.sync 0;
@%p6 bra $L__BB316_38;

mov.u32 %r1637, %ctaid.x;
shl.b32 %r1638, %r1637, 1;
mov.u32 %r1639, 1;
mad.lo.s32 %r1640, %r3, %r46, %r1638;
mad.lo.s32 %r1641, %r4, %r52, %r1640;
cvt.u64.u32 %rd19, %r1641;
shl.b32 %r1643, %r55, 2;
mov.u32 %r1644, _ZZ13mul_mat_vec_qIL9ggml_type21ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjjE10tmp_shared;
add.s32 %r34, %r1644, %r1643;
mov.u32 %r1645, 2;
ld.shared.f32 %f161, [%r34];
add.ftz.f32 %f162, %f161, %f368;
mov.b32 %r1646, %f162;
mov.u32 %r1647, 31;
mov.u32 %r1648, 16;
mov.u32 %r1649, -1;
shfl.sync.bfly.b32 %r1650|%p7, %r1646, %r1648, %r1647, %r1649;
mov.b32 %f163, %r1650;
add.ftz.f32 %f164, %f162, %f163;
mov.b32 %r1651, %f164;
mov.u32 %r1652, 8;
shfl.sync.bfly.b32 %r1653|%p8, %r1651, %r1652, %r1647, %r1649;
mov.b32 %f165, %r1653;
add.ftz.f32 %f166, %f164, %f165;
mov.b32 %r1654, %f166;
mov.u32 %r1655, 4;
shfl.sync.bfly.b32 %r1656|%p9, %r1654, %r1655, %r1647, %r1649;
mov.b32 %f167, %r1656;
add.ftz.f32 %f168, %f166, %f167;
mov.b32 %r1657, %f168;
shfl.sync.bfly.b32 %r1658|%p10, %r1657, %r1645, %r1647, %r1649;
mov.b32 %f169, %r1658;
add.ftz.f32 %f170, %f168, %f169;
mov.b32 %r1659, %f170;
shfl.sync.bfly.b32 %r1660|%p11, %r1659, %r1639, %r1647, %r1649;
mov.b32 %f171, %r1660;
add.ftz.f32 %f49, %f170, %f171;
setp.eq.s32 %p12, %r55, 0;
setp.lt.u32 %p13, %r1638, %r40;
and.pred %p1, %p12, %p13;
cvta.to.global.u64 %rd128, %rd26;
mul.wide.u32 %rd129, %r1641, 4;
add.s64 %rd20, %rd128, %rd129;
not.pred %p14, %p1;
@%p14 bra $L__BB316_8;

st.global.f32 [%rd20], %f49;

$L__BB316_8:
ld.shared.f32 %f172, [%r34+128];
add.ftz.f32 %f173, %f172, %f367;
mov.b32 %r1661, %f173;
shfl.sync.bfly.b32 %r1665|%p15, %r1661, %r1648, %r1647, %r1649;
mov.b32 %f174, %r1665;
add.ftz.f32 %f175, %f173, %f174;
mov.b32 %r1666, %f175;
shfl.sync.bfly.b32 %r1668|%p16, %r1666, %r1652, %r1647, %r1649;
mov.b32 %f176, %r1668;
add.ftz.f32 %f177, %f175, %f176;
mov.b32 %r1669, %f177;
shfl.sync.bfly.b32 %r1671|%p17, %r1669, %r1655, %r1647, %r1649;
mov.b32 %f178, %r1671;
add.ftz.f32 %f179, %f177, %f178;
mov.b32 %r1672, %f179;
shfl.sync.bfly.b32 %r1674|%p18, %r1672, %r1645, %r1647, %r1649;
mov.b32 %f180, %r1674;
add.ftz.f32 %f181, %f179, %f180;
mov.b32 %r1675, %f181;
shfl.sync.bfly.b32 %r1677|%p19, %r1675, %r1639, %r1647, %r1649;
mov.b32 %f182, %r1677;
add.ftz.f32 %f50, %f181, %f182;
setp.eq.s32 %p20, %r55, 1;
or.b32 %r1681, %r1638, 1;
setp.lt.u32 %p21, %r1681, %r40;
and.pred %p2, %p20, %p21;
not.pred %p22, %p2;
@%p22 bra $L__BB316_10;

st.global.f32 [%rd20+4], %f50;

$L__BB316_10:
ld.shared.f32 %f183, [%r34+256];
add.ftz.f32 %f184, %f183, %f366;
mov.b32 %r1682, %f184;
mov.u32 %r1683, 31;
mov.u32 %r1684, 16;
mov.u32 %r1685, -1;
shfl.sync.bfly.b32 %r1686|%p23, %r1682, %r1684, %r1683, %r1685;
mov.b32 %f185, %r1686;
add.ftz.f32 %f186, %f184, %f185;
mov.b32 %r1687, %f186;
mov.u32 %r1688, 8;
shfl.sync.bfly.b32 %r1689|%p24, %r1687, %r1688, %r1683, %r1685;
mov.b32 %f187, %r1689;
add.ftz.f32 %f188, %f186, %f187;
mov.b32 %r1690, %f188;
mov.u32 %r1691, 4;
shfl.sync.bfly.b32 %r1692|%p25, %r1690, %r1691, %r1683, %r1685;
mov.b32 %f189, %r1692;
add.ftz.f32 %f190, %f188, %f189;
mov.b32 %r1693, %f190;
mov.u32 %r1694, 2;
shfl.sync.bfly.b32 %r1695|%p26, %r1693, %r1694, %r1683, %r1685;
mov.b32 %f191, %r1695;
add.ftz.f32 %f192, %f190, %f191;
mov.b32 %r1696, %f192;
mov.u32 %r1697, 1;
shfl.sync.bfly.b32 %r1698|%p27, %r1696, %r1697, %r1683, %r1685;
mov.b32 %f193, %r1698;
add.ftz.f32 %f51, %f192, %f193;
@%p14 bra $L__BB316_12;

cvt.u64.u32 %rd130, %r40;
add.s64 %rd131, %rd130, %rd19;
shl.b64 %rd133, %rd131, 2;
add.s64 %rd134, %rd128, %rd133;
st.global.f32 [%rd134], %f51;

$L__BB316_12:
ld.shared.f32 %f194, [%r34+384];
add.ftz.f32 %f195, %f194, %f365;
mov.b32 %r1699, %f195;
shfl.sync.bfly.b32 %r1703|%p29, %r1699, %r1684, %r1683, %r1685;
mov.b32 %f196, %r1703;
add.ftz.f32 %f197, %f195, %f196;
mov.b32 %r1704, %f197;
shfl.sync.bfly.b32 %r1706|%p30, %r1704, %r1688, %r1683, %r1685;
mov.b32 %f198, %r1706;
add.ftz.f32 %f199, %f197, %f198;
mov.b32 %r1707, %f199;
shfl.sync.bfly.b32 %r1709|%p31, %r1707, %r1691, %r1683, %r1685;
mov.b32 %f200, %r1709;
add.ftz.f32 %f201, %f199, %f200;
mov.b32 %r1710, %f201;
shfl.sync.bfly.b32 %r1712|%p32, %r1710, %r1694, %r1683, %r1685;
mov.b32 %f202, %r1712;
add.ftz.f32 %f203, %f201, %f202;
mov.b32 %r1713, %f203;
shfl.sync.bfly.b32 %r1715|%p33, %r1713, %r1697, %r1683, %r1685;
mov.b32 %f204, %r1715;
add.ftz.f32 %f52, %f203, %f204;
@%p22 bra $L__BB316_14;

add.s32 %r1716, %r40, 1;
cvt.u64.u32 %rd135, %r1716;
add.s64 %rd136, %rd135, %rd19;
shl.b64 %rd138, %rd136, 2;
add.s64 %rd139, %rd128, %rd138;
st.global.f32 [%rd139], %f52;

$L__BB316_14:
shl.b32 %r1717, %r40, 1;
mov.u32 %r1718, 1;
ld.shared.f32 %f205, [%r34+512];
add.ftz.f32 %f206, %f205, %f364;
mov.b32 %r1719, %f206;
mov.u32 %r1720, 31;
mov.u32 %r1721, 16;
mov.u32 %r1722, -1;
shfl.sync.bfly.b32 %r1723|%p35, %r1719, %r1721, %r1720, %r1722;
mov.b32 %f207, %r1723;
add.ftz.f32 %f208, %f206, %f207;
mov.b32 %r1724, %f208;
mov.u32 %r1725, 8;
shfl.sync.bfly.b32 %r1726|%p36, %r1724, %r1725, %r1720, %r1722;
mov.b32 %f209, %r1726;
add.ftz.f32 %f210, %f208, %f209;
mov.b32 %r1727, %f210;
mov.u32 %r1728, 4;
shfl.sync.bfly.b32 %r1729|%p37, %r1727, %r1728, %r1720, %r1722;
mov.b32 %f211, %r1729;
add.ftz.f32 %f212, %f210, %f211;
mov.b32 %r1730, %f212;
mov.u32 %r1731, 2;
shfl.sync.bfly.b32 %r1732|%p38, %r1730, %r1731, %r1720, %r1722;
mov.b32 %f213, %r1732;
add.ftz.f32 %f214, %f212, %f213;
mov.b32 %r1733, %f214;
shfl.sync.bfly.b32 %r1734|%p39, %r1733, %r1718, %r1720, %r1722;
mov.b32 %f215, %r1734;
add.ftz.f32 %f53, %f214, %f215;
cvt.u64.u32 %rd140, %r1717;
add.s64 %rd141, %rd140, %rd19;
shl.b64 %rd143, %rd141, 2;
add.s64 %rd21, %rd128, %rd143;
@%p14 bra $L__BB316_16;

st.global.f32 [%rd21], %f53;

$L__BB316_16:
ld.shared.f32 %f216, [%r34+640];
add.ftz.f32 %f217, %f216, %f363;
mov.b32 %r1735, %f217;
shfl.sync.bfly.b32 %r1739|%p41, %r1735, %r1721, %r1720, %r1722;
mov.b32 %f218, %r1739;
add.ftz.f32 %f219, %f217, %f218;
mov.b32 %r1740, %f219;
shfl.sync.bfly.b32 %r1742|%p42, %r1740, %r1725, %r1720, %r1722;
mov.b32 %f220, %r1742;
add.ftz.f32 %f221, %f219, %f220;
mov.b32 %r1743, %f221;
shfl.sync.bfly.b32 %r1745|%p43, %r1743, %r1728, %r1720, %r1722;
mov.b32 %f222, %r1745;
add.ftz.f32 %f223, %f221, %f222;
mov.b32 %r1746, %f223;
shfl.sync.bfly.b32 %r1748|%p44, %r1746, %r1731, %r1720, %r1722;
mov.b32 %f224, %r1748;
add.ftz.f32 %f225, %f223, %f224;
mov.b32 %r1749, %f225;
shfl.sync.bfly.b32 %r1751|%p45, %r1749, %r1718, %r1720, %r1722;
mov.b32 %f226, %r1751;
add.ftz.f32 %f54, %f225, %f226;
@%p22 bra $L__BB316_18;

st.global.f32 [%rd21+4], %f54;

$L__BB316_18:
mul.lo.s32 %r35, %r40, 3;
ld.shared.f32 %f227, [%r34+768];
add.ftz.f32 %f228, %f227, %f362;
mov.b32 %r1752, %f228;
mov.u32 %r1753, 31;
mov.u32 %r1754, 16;
mov.u32 %r1755, -1;
shfl.sync.bfly.b32 %r1756|%p47, %r1752, %r1754, %r1753, %r1755;
mov.b32 %f229, %r1756;
add.ftz.f32 %f230, %f228, %f229;
mov.b32 %r1757, %f230;
mov.u32 %r1758, 8;
shfl.sync.bfly.b32 %r1759|%p48, %r1757, %r1758, %r1753, %r1755;
mov.b32 %f231, %r1759;
add.ftz.f32 %f232, %f230, %f231;
mov.b32 %r1760, %f232;
mov.u32 %r1761, 4;
shfl.sync.bfly.b32 %r1762|%p49, %r1760, %r1761, %r1753, %r1755;
mov.b32 %f233, %r1762;
add.ftz.f32 %f234, %f232, %f233;
mov.b32 %r1763, %f234;
mov.u32 %r1764, 2;
shfl.sync.bfly.b32 %r1765|%p50, %r1763, %r1764, %r1753, %r1755;
mov.b32 %f235, %r1765;
add.ftz.f32 %f236, %f234, %f235;
mov.b32 %r1766, %f236;
mov.u32 %r1767, 1;
shfl.sync.bfly.b32 %r1768|%p51, %r1766, %r1767, %r1753, %r1755;
mov.b32 %f237, %r1768;
add.ftz.f32 %f55, %f236, %f237;
@%p14 bra $L__BB316_20;

cvt.u64.u32 %rd144, %r35;
add.s64 %rd145, %rd144, %rd19;
shl.b64 %rd147, %rd145, 2;
add.s64 %rd148, %rd128, %rd147;
st.global.f32 [%rd148], %f55;

$L__BB316_20:
ld.shared.f32 %f238, [%r34+896];
add.ftz.f32 %f239, %f238, %f361;
mov.b32 %r1769, %f239;
shfl.sync.bfly.b32 %r1773|%p53, %r1769, %r1754, %r1753, %r1755;
mov.b32 %f240, %r1773;
add.ftz.f32 %f241, %f239, %f240;
mov.b32 %r1774, %f241;
shfl.sync.bfly.b32 %r1776|%p54, %r1774, %r1758, %r1753, %r1755;
mov.b32 %f242, %r1776;
add.ftz.f32 %f243, %f241, %f242;
mov.b32 %r1777, %f243;
shfl.sync.bfly.b32 %r1779|%p55, %r1777, %r1761, %r1753, %r1755;
mov.b32 %f244, %r1779;
add.ftz.f32 %f245, %f243, %f244;
mov.b32 %r1780, %f245;
shfl.sync.bfly.b32 %r1782|%p56, %r1780, %r1764, %r1753, %r1755;
mov.b32 %f246, %r1782;
add.ftz.f32 %f247, %f245, %f246;
mov.b32 %r1783, %f247;
shfl.sync.bfly.b32 %r1785|%p57, %r1783, %r1767, %r1753, %r1755;
mov.b32 %f248, %r1785;
add.ftz.f32 %f56, %f247, %f248;
@%p22 bra $L__BB316_22;

add.s32 %r1786, %r35, 1;
cvt.u64.u32 %rd149, %r1786;
add.s64 %rd150, %rd149, %rd19;
shl.b64 %rd152, %rd150, 2;
add.s64 %rd153, %rd128, %rd152;
st.global.f32 [%rd153], %f56;

$L__BB316_22:
shl.b32 %r1787, %r40, 2;
mov.u32 %r1788, 2;
ld.shared.f32 %f249, [%r34+1024];
add.ftz.f32 %f250, %f249, %f360;
mov.b32 %r1789, %f250;
mov.u32 %r1790, 31;
mov.u32 %r1791, 16;
mov.u32 %r1792, -1;
shfl.sync.bfly.b32 %r1793|%p59, %r1789, %r1791, %r1790, %r1792;
mov.b32 %f251, %r1793;
add.ftz.f32 %f252, %f250, %f251;
mov.b32 %r1794, %f252;
mov.u32 %r1795, 8;
shfl.sync.bfly.b32 %r1796|%p60, %r1794, %r1795, %r1790, %r1792;
mov.b32 %f253, %r1796;
add.ftz.f32 %f254, %f252, %f253;
mov.b32 %r1797, %f254;
mov.u32 %r1798, 4;
shfl.sync.bfly.b32 %r1799|%p61, %r1797, %r1798, %r1790, %r1792;
mov.b32 %f255, %r1799;
add.ftz.f32 %f256, %f254, %f255;
mov.b32 %r1800, %f256;
shfl.sync.bfly.b32 %r1801|%p62, %r1800, %r1788, %r1790, %r1792;
mov.b32 %f257, %r1801;
add.ftz.f32 %f258, %f256, %f257;
mov.b32 %r1802, %f258;
mov.u32 %r1803, 1;
shfl.sync.bfly.b32 %r1804|%p63, %r1802, %r1803, %r1790, %r1792;
mov.b32 %f259, %r1804;
add.ftz.f32 %f57, %f258, %f259;
cvt.u64.u32 %rd154, %r1787;
add.s64 %rd155, %rd154, %rd19;
shl.b64 %rd157, %rd155, 2;
add.s64 %rd22, %rd128, %rd157;
@%p14 bra $L__BB316_24;

st.global.f32 [%rd22], %f57;

$L__BB316_24:
ld.shared.f32 %f260, [%r34+1152];
add.ftz.f32 %f261, %f260, %f359;
mov.b32 %r1805, %f261;
shfl.sync.bfly.b32 %r1809|%p65, %r1805, %r1791, %r1790, %r1792;
mov.b32 %f262, %r1809;
add.ftz.f32 %f263, %f261, %f262;
mov.b32 %r1810, %f263;
shfl.sync.bfly.b32 %r1812|%p66, %r1810, %r1795, %r1790, %r1792;
mov.b32 %f264, %r1812;
add.ftz.f32 %f265, %f263, %f264;
mov.b32 %r1813, %f265;
shfl.sync.bfly.b32 %r1815|%p67, %r1813, %r1798, %r1790, %r1792;
mov.b32 %f266, %r1815;
add.ftz.f32 %f267, %f265, %f266;
mov.b32 %r1816, %f267;
shfl.sync.bfly.b32 %r1818|%p68, %r1816, %r1788, %r1790, %r1792;
mov.b32 %f268, %r1818;
add.ftz.f32 %f269, %f267, %f268;
mov.b32 %r1819, %f269;
shfl.sync.bfly.b32 %r1821|%p69, %r1819, %r1803, %r1790, %r1792;
mov.b32 %f270, %r1821;
add.ftz.f32 %f58, %f269, %f270;
@%p22 bra $L__BB316_26;

st.global.f32 [%rd22+4], %f58;

$L__BB316_26:
mul.lo.s32 %r36, %r40, 5;
ld.shared.f32 %f271, [%r34+1280];
add.ftz.f32 %f272, %f271, %f358;
mov.b32 %r1822, %f272;
mov.u32 %r1823, 31;
mov.u32 %r1824, 16;
mov.u32 %r1825, -1;
shfl.sync.bfly.b32 %r1826|%p71, %r1822, %r1824, %r1823, %r1825;
mov.b32 %f273, %r1826;
add.ftz.f32 %f274, %f272, %f273;
mov.b32 %r1827, %f274;
mov.u32 %r1828, 8;
shfl.sync.bfly.b32 %r1829|%p72, %r1827, %r1828, %r1823, %r1825;
mov.b32 %f275, %r1829;
add.ftz.f32 %f276, %f274, %f275;
mov.b32 %r1830, %f276;
mov.u32 %r1831, 4;
shfl.sync.bfly.b32 %r1832|%p73, %r1830, %r1831, %r1823, %r1825;
mov.b32 %f277, %r1832;
add.ftz.f32 %f278, %f276, %f277;
mov.b32 %r1833, %f278;
mov.u32 %r1834, 2;
shfl.sync.bfly.b32 %r1835|%p74, %r1833, %r1834, %r1823, %r1825;
mov.b32 %f279, %r1835;
add.ftz.f32 %f280, %f278, %f279;
mov.b32 %r1836, %f280;
mov.u32 %r1837, 1;
shfl.sync.bfly.b32 %r1838|%p75, %r1836, %r1837, %r1823, %r1825;
mov.b32 %f281, %r1838;
add.ftz.f32 %f59, %f280, %f281;
@%p14 bra $L__BB316_28;

cvt.u64.u32 %rd158, %r36;
add.s64 %rd159, %rd158, %rd19;
shl.b64 %rd161, %rd159, 2;
add.s64 %rd162, %rd128, %rd161;
st.global.f32 [%rd162], %f59;

$L__BB316_28:
ld.shared.f32 %f282, [%r34+1408];
add.ftz.f32 %f283, %f282, %f357;
mov.b32 %r1839, %f283;
shfl.sync.bfly.b32 %r1843|%p77, %r1839, %r1824, %r1823, %r1825;
mov.b32 %f284, %r1843;
add.ftz.f32 %f285, %f283, %f284;
mov.b32 %r1844, %f285;
shfl.sync.bfly.b32 %r1846|%p78, %r1844, %r1828, %r1823, %r1825;
mov.b32 %f286, %r1846;
add.ftz.f32 %f287, %f285, %f286;
mov.b32 %r1847, %f287;
shfl.sync.bfly.b32 %r1849|%p79, %r1847, %r1831, %r1823, %r1825;
mov.b32 %f288, %r1849;
add.ftz.f32 %f289, %f287, %f288;
mov.b32 %r1850, %f289;
shfl.sync.bfly.b32 %r1852|%p80, %r1850, %r1834, %r1823, %r1825;
mov.b32 %f290, %r1852;
add.ftz.f32 %f291, %f289, %f290;
mov.b32 %r1853, %f291;
shfl.sync.bfly.b32 %r1855|%p81, %r1853, %r1837, %r1823, %r1825;
mov.b32 %f292, %r1855;
add.ftz.f32 %f60, %f291, %f292;
@%p22 bra $L__BB316_30;

add.s32 %r1856, %r36, 1;
cvt.u64.u32 %rd163, %r1856;
add.s64 %rd164, %rd163, %rd19;
shl.b64 %rd166, %rd164, 2;
add.s64 %rd167, %rd128, %rd166;
st.global.f32 [%rd167], %f60;

$L__BB316_30:
mul.lo.s32 %r1857, %r40, 6;
ld.shared.f32 %f293, [%r34+1536];
add.ftz.f32 %f294, %f293, %f356;
mov.b32 %r1858, %f294;
mov.u32 %r1859, 31;
mov.u32 %r1860, 16;
mov.u32 %r1861, -1;
shfl.sync.bfly.b32 %r1862|%p83, %r1858, %r1860, %r1859, %r1861;
mov.b32 %f295, %r1862;
add.ftz.f32 %f296, %f294, %f295;
mov.b32 %r1863, %f296;
mov.u32 %r1864, 8;
shfl.sync.bfly.b32 %r1865|%p84, %r1863, %r1864, %r1859, %r1861;
mov.b32 %f297, %r1865;
add.ftz.f32 %f298, %f296, %f297;
mov.b32 %r1866, %f298;
mov.u32 %r1867, 4;
shfl.sync.bfly.b32 %r1868|%p85, %r1866, %r1867, %r1859, %r1861;
mov.b32 %f299, %r1868;
add.ftz.f32 %f300, %f298, %f299;
mov.b32 %r1869, %f300;
mov.u32 %r1870, 2;
shfl.sync.bfly.b32 %r1871|%p86, %r1869, %r1870, %r1859, %r1861;
mov.b32 %f301, %r1871;
add.ftz.f32 %f302, %f300, %f301;
mov.b32 %r1872, %f302;
mov.u32 %r1873, 1;
shfl.sync.bfly.b32 %r1874|%p87, %r1872, %r1873, %r1859, %r1861;
mov.b32 %f303, %r1874;
add.ftz.f32 %f61, %f302, %f303;
cvt.u64.u32 %rd168, %r1857;
add.s64 %rd169, %rd168, %rd19;
shl.b64 %rd171, %rd169, 2;
add.s64 %rd23, %rd128, %rd171;
@%p14 bra $L__BB316_32;

st.global.f32 [%rd23], %f61;

$L__BB316_32:
ld.shared.f32 %f304, [%r34+1664];
add.ftz.f32 %f305, %f304, %f355;
mov.b32 %r1875, %f305;
shfl.sync.bfly.b32 %r1879|%p89, %r1875, %r1860, %r1859, %r1861;
mov.b32 %f306, %r1879;
add.ftz.f32 %f307, %f305, %f306;
mov.b32 %r1880, %f307;
shfl.sync.bfly.b32 %r1882|%p90, %r1880, %r1864, %r1859, %r1861;
mov.b32 %f308, %r1882;
add.ftz.f32 %f309, %f307, %f308;
mov.b32 %r1883, %f309;
shfl.sync.bfly.b32 %r1885|%p91, %r1883, %r1867, %r1859, %r1861;
mov.b32 %f310, %r1885;
add.ftz.f32 %f311, %f309, %f310;
mov.b32 %r1886, %f311;
shfl.sync.bfly.b32 %r1888|%p92, %r1886, %r1870, %r1859, %r1861;
mov.b32 %f312, %r1888;
add.ftz.f32 %f313, %f311, %f312;
mov.b32 %r1889, %f313;
shfl.sync.bfly.b32 %r1891|%p93, %r1889, %r1873, %r1859, %r1861;
mov.b32 %f314, %r1891;
add.ftz.f32 %f62, %f313, %f314;
@%p22 bra $L__BB316_34;

st.global.f32 [%rd23+4], %f62;

$L__BB316_34:
mul.lo.s32 %r37, %r40, 7;
ld.shared.f32 %f315, [%r34+1792];
add.ftz.f32 %f316, %f315, %f354;
mov.b32 %r1892, %f316;
mov.u32 %r1893, 31;
mov.u32 %r1894, 16;
mov.u32 %r1895, -1;
shfl.sync.bfly.b32 %r1896|%p95, %r1892, %r1894, %r1893, %r1895;
mov.b32 %f317, %r1896;
add.ftz.f32 %f318, %f316, %f317;
mov.b32 %r1897, %f318;
mov.u32 %r1898, 8;
shfl.sync.bfly.b32 %r1899|%p96, %r1897, %r1898, %r1893, %r1895;
mov.b32 %f319, %r1899;
add.ftz.f32 %f320, %f318, %f319;
mov.b32 %r1900, %f320;
mov.u32 %r1901, 4;
shfl.sync.bfly.b32 %r1902|%p97, %r1900, %r1901, %r1893, %r1895;
mov.b32 %f321, %r1902;
add.ftz.f32 %f322, %f320, %f321;
mov.b32 %r1903, %f322;
mov.u32 %r1904, 2;
shfl.sync.bfly.b32 %r1905|%p98, %r1903, %r1904, %r1893, %r1895;
mov.b32 %f323, %r1905;
add.ftz.f32 %f324, %f322, %f323;
mov.b32 %r1906, %f324;
mov.u32 %r1907, 1;
shfl.sync.bfly.b32 %r1908|%p99, %r1906, %r1907, %r1893, %r1895;
mov.b32 %f325, %r1908;
add.ftz.f32 %f63, %f324, %f325;
@%p14 bra $L__BB316_36;

cvt.u64.u32 %rd172, %r37;
add.s64 %rd173, %rd172, %rd19;
shl.b64 %rd175, %rd173, 2;
add.s64 %rd176, %rd128, %rd175;
st.global.f32 [%rd176], %f63;

$L__BB316_36:
ld.shared.f32 %f326, [%r34+1920];
add.ftz.f32 %f327, %f326, %f353;
mov.b32 %r1909, %f327;
shfl.sync.bfly.b32 %r1913|%p101, %r1909, %r1894, %r1893, %r1895;
mov.b32 %f328, %r1913;
add.ftz.f32 %f329, %f327, %f328;
mov.b32 %r1914, %f329;
shfl.sync.bfly.b32 %r1916|%p102, %r1914, %r1898, %r1893, %r1895;
mov.b32 %f330, %r1916;
add.ftz.f32 %f331, %f329, %f330;
mov.b32 %r1917, %f331;
shfl.sync.bfly.b32 %r1919|%p103, %r1917, %r1901, %r1893, %r1895;
mov.b32 %f332, %r1919;
add.ftz.f32 %f333, %f331, %f332;
mov.b32 %r1920, %f333;
shfl.sync.bfly.b32 %r1922|%p104, %r1920, %r1904, %r1893, %r1895;
mov.b32 %f334, %r1922;
add.ftz.f32 %f335, %f333, %f334;
mov.b32 %r1923, %f335;
shfl.sync.bfly.b32 %r1925|%p105, %r1923, %r1907, %r1893, %r1895;
mov.b32 %f336, %r1925;
add.ftz.f32 %f64, %f335, %f336;
@%p22 bra $L__BB316_38;

add.s32 %r1926, %r37, 1;
cvt.u64.u32 %rd177, %r1926;
add.s64 %rd178, %rd177, %rd19;
shl.b64 %rd180, %rd178, 2;
add.s64 %rd181, %rd128, %rd180;
st.global.f32 [%rd181], %f64;

$L__BB316_38:
ret;

}

   ELFA        ¾            h�B     ¨3A     Y @ 8  @ 7  .shstrtab .strtab .symtab .symtab_shndx .note.nv.tkinfo .note.nv.cuinfo .nv.info .text._Z13mul_mat_vec_qIL9ggml_type21ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type21ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type21ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.global.init .nv.constant4 .nv.constant2._Z13mul_mat_vec_qIL9ggml_type21ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type21ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type21ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type21ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type21ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type21ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type21ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type21ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type21ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type21ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type21ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type21ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type21ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type21ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type21ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type21ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type21ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type21ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type21ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type21ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type21ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type21ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type21ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type21ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type21ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type21ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type21ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type21ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type21ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type21ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type21ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type21ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type21ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type21ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type21ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type21ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type21ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type21ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type21ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z13mul_mat_vec_qIL9ggml_type23ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type23ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type23ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type23ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type23ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type23ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type23ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type23ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type23ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type23ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type23ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type23ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type23ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type23ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type23ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type23ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type23ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type23ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type23ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type23ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type23ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type23ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type23ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type23ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type23ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type23ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type23ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type23ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type23ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type23ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type23ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type23ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type23ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type23ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type23ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z13mul_mat_vec_qIL9ggml_type20ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type20ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type20ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type20ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type20ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type20ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type20ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type20ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type20ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type20ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type20ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type20ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type20ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type20ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type20ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type20ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type20ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type20ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type20ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type20ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type20ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type20ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type20ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type20ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type20ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type20ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type20ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type20ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type20ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type20ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type20ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type20ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type20ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type20ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type20ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z13mul_mat_vec_qIL9ggml_type29ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type29ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type29ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type29ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type29ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type29ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type29ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type29ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type29ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type29ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type29ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type29ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type29ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type29ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type29ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type29ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type29ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type29ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type29ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type29ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type29ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type29ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type29ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type29ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type29ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type29ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type29ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type29ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type29ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type29ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type29ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type29ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type29ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type29ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type29ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type29ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type29ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type29ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type29ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type29ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type29ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type29ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant2._Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant2._Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z13mul_mat_vec_qIL9ggml_type19ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type19ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type19ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type19ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type19ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type19ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type19ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type19ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type19ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type19ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type19ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type19ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type19ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type19ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type19ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type19ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type19ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type19ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type19ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type19ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type19ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type19ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type19ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type19ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type19ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type19ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type19ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type19ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type19ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type19ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type19ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type19ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type19ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type19ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type19ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z13mul_mat_vec_qIL9ggml_type18ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type18ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type18ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type18ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type18ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type18ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type18ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type18ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type18ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type18ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type18ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type18ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type18ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type18ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type18ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type18ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type18ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type18ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type18ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type18ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type18ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type18ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type18ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type18ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type18ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type18ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type18ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type18ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type18ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type18ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type18ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type18ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type18ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type18ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type18ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type18ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type18ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type18ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type18ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type18ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type18ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type18ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z13mul_mat_vec_qIL9ggml_type22ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type22ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type22ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type22ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type22ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type22ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type22ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type22ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type22ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type22ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type22ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type22ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type22ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type22ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type22ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type22ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type22ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type22ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type22ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type22ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type22ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type22ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type22ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type22ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type22ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type22ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type22ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type22ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type22ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type22ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type22ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type22ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type22ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type22ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type22ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z13mul_mat_vec_qIL9ggml_type17ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type17ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type17ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type17ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type17ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type17ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type17ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type17ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type17ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type17ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type17ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type17ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type17ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type17ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type17ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type17ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type17ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type17ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type17ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type17ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type17ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type17ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type17ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type17ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type17ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type17ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type17ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type17ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type17ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type17ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type17ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type17ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type17ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type17ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type17ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z13mul_mat_vec_qIL9ggml_type16ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type16ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type16ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type16ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type16ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type16ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type16ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type16ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type16ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type16ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type16ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type16ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type16ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type16ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type16ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type16ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type16ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type16ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type16ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type16ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type16ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type16ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type16ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type16ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type16ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type16ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type16ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type16ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type16ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type16ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type16ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type16ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type16ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type16ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type16ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z13mul_mat_vec_qIL9ggml_type14ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type14ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type14ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type14ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type14ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type14ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type14ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type14ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type14ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type14ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type14ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type14ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type14ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type14ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type14ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type14ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type14ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type14ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type14ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type14ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type14ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type14ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type14ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type14ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type14ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type14ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type14ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type14ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type14ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type14ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type14ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type14ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type14ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type14ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type14ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z13mul_mat_vec_qIL9ggml_type13ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type13ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type13ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type13ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type13ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type13ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type13ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type13ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type13ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type13ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type13ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type13ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type13ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type13ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type13ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type13ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type13ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type13ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type13ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type13ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type13ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type13ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type13ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type13ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type13ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type13ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type13ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type13ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type13ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type13ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type13ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type13ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type13ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type13ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type13ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type13ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type13ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type13ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type13ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type13ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type13ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type13ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant2._Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant2._Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z13mul_mat_vec_qIL9ggml_type12ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type12ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type12ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type12ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type12ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type12ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type12ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type12ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type12ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type12ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type12ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type12ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type12ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type12ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type12ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type12ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type12ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type12ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type12ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type12ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type12ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type12ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type12ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type12ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type12ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type12ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type12ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type12ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type12ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type12ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type12ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type12ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type12ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type12ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type12ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type12ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type12ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type12ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type12ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type12ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type12ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type12ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant2._Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant2._Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z13mul_mat_vec_qIL9ggml_type11ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type11ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type11ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type11ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type11ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type11ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type11ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type11ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type11ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type11ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type11ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type11ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type11ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type11ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type11ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type11ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type11ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type11ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type11ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type11ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type11ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type11ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type11ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type11ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type11ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type11ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type11ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type11ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type11ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type11ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type11ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type11ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type11ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type11ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type11ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z13mul_mat_vec_qIL9ggml_type10ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type10ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type10ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type10ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type10ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type10ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type10ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type10ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type10ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type10ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type10ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type10ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type10ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type10ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type10ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type10ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type10ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type10ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type10ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type10ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type10ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type10ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type10ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type10ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type10ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type10ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type10ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type10ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type10ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type10ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type10ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type10ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type10ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type10ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type10ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z13mul_mat_vec_qIL9ggml_type40ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type40ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type40ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type40ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type40ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type40ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type40ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type40ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type40ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type40ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type40ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type40ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type40ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type40ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type40ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type40ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type40ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type40ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type40ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type40ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type40ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type40ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type40ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type40ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type40ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type40ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type40ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type40ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type40ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type40ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type40ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type40ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type40ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type40ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type40ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z13mul_mat_vec_qIL9ggml_type39ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type39ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type39ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type39ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type39ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type39ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type39ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type39ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type39ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type39ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type39ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type39ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type39ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type39ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type39ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type39ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type39ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type39ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type39ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type39ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type39ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type39ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type39ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type39ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type39ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type39ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type39ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type39ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type39ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type39ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type39ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type39ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type39ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type39ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type39ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z13mul_mat_vec_qIL9ggml_type8ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type8ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type8ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type8ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type8ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type8ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type8ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type8ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type8ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type8ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type8ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type8ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type8ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type8ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type8ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type8ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type8ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type8ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type8ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type8ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type8ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type8ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type8ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type8ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type8ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type8ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type8ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type8ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type8ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type8ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type8ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type8ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type8ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type8ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type8ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant2._Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z13mul_mat_vec_qIL9ggml_type7ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type7ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type7ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type7ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type7ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type7ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type7ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type7ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type7ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type7ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type7ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type7ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type7ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type7ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type7ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type7ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type7ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type7ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type7ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type7ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type7ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type7ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type7ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type7ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type7ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type7ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type7ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type7ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type7ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type7ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type7ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type7ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type7ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type7ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type7ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z13mul_mat_vec_qIL9ggml_type6ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type6ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type6ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type6ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type6ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type6ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type6ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type6ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type6ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type6ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type6ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type6ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type6ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type6ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type6ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type6ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type6ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type6ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type6ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type6ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type6ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type6ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type6ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type6ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type6ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type6ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type6ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type6ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type6ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type6ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type6ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type6ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type6ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type6ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type6ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z13mul_mat_vec_qIL9ggml_type3ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type3ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type3ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type3ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type3ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type3ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type3ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type3ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type3ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type3ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type3ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type3ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type3ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type3ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type3ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type3ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type3ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type3ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type3ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type3ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type3ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type3ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type3ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type3ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type3ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type3ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type3ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type3ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type3ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type3ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type3ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type3ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type3ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type3ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type3ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant2._Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z13mul_mat_vec_qIL9ggml_type2ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type2ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type2ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type2ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type2ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type2ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type2ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type2ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type2ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type2ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type2ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type2ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type2ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type2ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type2ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type2ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type2ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type2ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type2ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type2ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type2ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type2ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type2ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type2ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type2ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type2ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type2ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type2ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type2ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type2ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type2ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type2ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type2ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type2ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type2ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant2._Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z13mul_mat_vec_qIL9ggml_type42ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type42ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type42ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type42ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type42ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type42ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type42ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type42ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type42ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type42ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type42ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type42ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type42ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type42ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type42ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type42ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type42ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type42ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type42ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type42ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type42ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type42ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type42ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type42ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type42ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type42ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type42ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type42ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type42ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type42ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type42ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type42ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type42ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type42ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type42ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z13mul_mat_vec_qIL9ggml_type41ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type41ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type41ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type41ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type41ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type41ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type41ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type41ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type41ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type41ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type41ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type41ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type41ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type41ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type41ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type41ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type41ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type41ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type41ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type41ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type41ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type41ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type41ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type41ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type41ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type41ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type41ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type41ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type41ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type41ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type41ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type41ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type41ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type41ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type41ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .debug_frame .rel.nv.constant4 .rel.debug_frame .rela.debug_frame .nv.callgraph .nv.prototype .nv.rel.action  .shstrtab .strtab .symtab .symtab_shndx .note.nv.tkinfo .note.nv.cuinfo .nv.info _Z13mul_mat_vec_qIL9ggml_type21ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type21ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type21ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type21ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.global.init kmask_iq2xs .nv.constant4 ksigns_iq2xs ksigns64 iq2xxs_grid iq2xs_grid iq2s_grid iq3xxs_grid iq3s_grid kvalues_iq4nl kvalues_fp4 iq1s_grid_gpu .nv.constant2._Z13mul_mat_vec_qIL9ggml_type21ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type21ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type21ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type21ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type21ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type21ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type21ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type21ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type21ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type21ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type21ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type21ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type21ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type21ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type21ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type21ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type21ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type21ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type21ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type21ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type21ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type21ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type21ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type21ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type21ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type21ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type21ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type21ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type21ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type21ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type21ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type21ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type21ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type21ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type21ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type21ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type21ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type21ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type21ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type21ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type21ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type21ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type21ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type21ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type21ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type21ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type21ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj _Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type21ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj _Z13mul_mat_vec_qIL9ggml_type23ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type23ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type23ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type23ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type23ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type23ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type23ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type23ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type23ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type23ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type23ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type23ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type23ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type23ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type23ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type23ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type23ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type23ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type23ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type23ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type23ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type23ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type23ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type23ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type23ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type23ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type23ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type23ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type23ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type23ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type23ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type23ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type23ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type23ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type23ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type23ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type23ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type23ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type23ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type23ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type23ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type23ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type23ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type23ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj _Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type23ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj _Z13mul_mat_vec_qIL9ggml_type20ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type20ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type20ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type20ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type20ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type20ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type20ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type20ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type20ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type20ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type20ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type20ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type20ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type20ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type20ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type20ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type20ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type20ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type20ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type20ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type20ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type20ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type20ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type20ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type20ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type20ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type20ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type20ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type20ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type20ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type20ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type20ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type20ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type20ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type20ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type20ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type20ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type20ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type20ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type20ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type20ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type20ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type20ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type20ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj _Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type20ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj _Z13mul_mat_vec_qIL9ggml_type29ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type29ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type29ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type29ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type29ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type29ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type29ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type29ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type29ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type29ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type29ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type29ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type29ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type29ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type29ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type29ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type29ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type29ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type29ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type29ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type29ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type29ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type29ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type29ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type29ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type29ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type29ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type29ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type29ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type29ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type29ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type29ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type29ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type29ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type29ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type29ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type29ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type29ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type29ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type29ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type29ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type29ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type29ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type29ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type29ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type29ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type29ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type29ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type29ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type29ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type29ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant2._Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj _Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant2._Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type29ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj _Z13mul_mat_vec_qIL9ggml_type19ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type19ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type19ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type19ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type19ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type19ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type19ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type19ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type19ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type19ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type19ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type19ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type19ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type19ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type19ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type19ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type19ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type19ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type19ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type19ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type19ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type19ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type19ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type19ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type19ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type19ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type19ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type19ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type19ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type19ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type19ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type19ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type19ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type19ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type19ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type19ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type19ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type19ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type19ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type19ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type19ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type19ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type19ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type19ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj _Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type19ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj _Z13mul_mat_vec_qIL9ggml_type18ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type18ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type18ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type18ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type18ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type18ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type18ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type18ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type18ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type18ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type18ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type18ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type18ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type18ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type18ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type18ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type18ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type18ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type18ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type18ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type18ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type18ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type18ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type18ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type18ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type18ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type18ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type18ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type18ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type18ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type18ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type18ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type18ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type18ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type18ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type18ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type18ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type18ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type18ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type18ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type18ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type18ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type18ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type18ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type18ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type18ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type18ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type18ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type18ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type18ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type18ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj _Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type18ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj _Z13mul_mat_vec_qIL9ggml_type22ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type22ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type22ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type22ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type22ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type22ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type22ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type22ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type22ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type22ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type22ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type22ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type22ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type22ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type22ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type22ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type22ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type22ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type22ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type22ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type22ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type22ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type22ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type22ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type22ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type22ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type22ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type22ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type22ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type22ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type22ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type22ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type22ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type22ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type22ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type22ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type22ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type22ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type22ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type22ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type22ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type22ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type22ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type22ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj _Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type22ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj _Z13mul_mat_vec_qIL9ggml_type17ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type17ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type17ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type17ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type17ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type17ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type17ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type17ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type17ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type17ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type17ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type17ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type17ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type17ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type17ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type17ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type17ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type17ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type17ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type17ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type17ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type17ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type17ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type17ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type17ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type17ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type17ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type17ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type17ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type17ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type17ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type17ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type17ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type17ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type17ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type17ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type17ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type17ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type17ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type17ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type17ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type17ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type17ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type17ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj _Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type17ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj _Z13mul_mat_vec_qIL9ggml_type16ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type16ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type16ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type16ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type16ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type16ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type16ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type16ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type16ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type16ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type16ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type16ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type16ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type16ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type16ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type16ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type16ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type16ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type16ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type16ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type16ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type16ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type16ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type16ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type16ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type16ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type16ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type16ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type16ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type16ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type16ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type16ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type16ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type16ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type16ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type16ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type16ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type16ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type16ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type16ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type16ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type16ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type16ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type16ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj _Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type16ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj _Z13mul_mat_vec_qIL9ggml_type14ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type14ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type14ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type14ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type14ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type14ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type14ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type14ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type14ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type14ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type14ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type14ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type14ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type14ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type14ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type14ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type14ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type14ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type14ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type14ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type14ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type14ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type14ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type14ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type14ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type14ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type14ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type14ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type14ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type14ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type14ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type14ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type14ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type14ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type14ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type14ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type14ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type14ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type14ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type14ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type14ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type14ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type14ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type14ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj _Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type14ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj _Z13mul_mat_vec_qIL9ggml_type13ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type13ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type13ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type13ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type13ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type13ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type13ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type13ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type13ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type13ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type13ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type13ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type13ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type13ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type13ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type13ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type13ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type13ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type13ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type13ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type13ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type13ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type13ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type13ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type13ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type13ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type13ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type13ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type13ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type13ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type13ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type13ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type13ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type13ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type13ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type13ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type13ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type13ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type13ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type13ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type13ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type13ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type13ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type13ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type13ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type13ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type13ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type13ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type13ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type13ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant2._Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj _Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant2._Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type13ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj _Z13mul_mat_vec_qIL9ggml_type12ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type12ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type12ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type12ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type12ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type12ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type12ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type12ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type12ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type12ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type12ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type12ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type12ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type12ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type12ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type12ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type12ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type12ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type12ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type12ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type12ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type12ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type12ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type12ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type12ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type12ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type12ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type12ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type12ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type12ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type12ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type12ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type12ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type12ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type12ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type12ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type12ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type12ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type12ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type12ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type12ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type12ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type12ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type12ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type12ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type12ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type12ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type12ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type12ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type12ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant2._Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj _Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant2._Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type12ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj _Z13mul_mat_vec_qIL9ggml_type11ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type11ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type11ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type11ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type11ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type11ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type11ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type11ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type11ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type11ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type11ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type11ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type11ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type11ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type11ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type11ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type11ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type11ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type11ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type11ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type11ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type11ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type11ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type11ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type11ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type11ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type11ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type11ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type11ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type11ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type11ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type11ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type11ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type11ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type11ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type11ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type11ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type11ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type11ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type11ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type11ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type11ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type11ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type11ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj _Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type11ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj _Z13mul_mat_vec_qIL9ggml_type10ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type10ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type10ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type10ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type10ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type10ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type10ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type10ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type10ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type10ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type10ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type10ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type10ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type10ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type10ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type10ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type10ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type10ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type10ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type10ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type10ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type10ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type10ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type10ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type10ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type10ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type10ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type10ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type10ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type10ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type10ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type10ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type10ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type10ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type10ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type10ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type10ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type10ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type10ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type10ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type10ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type10ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type10ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type10ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj _Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type10ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj _Z13mul_mat_vec_qIL9ggml_type40ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type40ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type40ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type40ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type40ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type40ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type40ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type40ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type40ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type40ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type40ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type40ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type40ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type40ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type40ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type40ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type40ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type40ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type40ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type40ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type40ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type40ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type40ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type40ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type40ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type40ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type40ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type40ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type40ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type40ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type40ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type40ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type40ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type40ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type40ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type40ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type40ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type40ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type40ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type40ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type40ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type40ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type40ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type40ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj _Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type40ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj _Z13mul_mat_vec_qIL9ggml_type39ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type39ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type39ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type39ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type39ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type39ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type39ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type39ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type39ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type39ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type39ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type39ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type39ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type39ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type39ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type39ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type39ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type39ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type39ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type39ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type39ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type39ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type39ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type39ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type39ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type39ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type39ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type39ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type39ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type39ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type39ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type39ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type39ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type39ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type39ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type39ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type39ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type39ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type39ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type39ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type39ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type39ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type39ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type39ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj _Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type39ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj _Z13mul_mat_vec_qIL9ggml_type8ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type8ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type8ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type8ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type8ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type8ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type8ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type8ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type8ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type8ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type8ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type8ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type8ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type8ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type8ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type8ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type8ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type8ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type8ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type8ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type8ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type8ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type8ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type8ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type8ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type8ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type8ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type8ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type8ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type8ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type8ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type8ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type8ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type8ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type8ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type8ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type8ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type8ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type8ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type8ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type8ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type8ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type8ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type8ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj _Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant2._Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type8ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj _Z13mul_mat_vec_qIL9ggml_type7ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type7ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type7ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type7ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type7ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type7ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type7ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type7ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type7ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type7ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type7ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type7ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type7ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type7ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type7ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type7ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type7ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type7ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type7ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type7ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type7ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type7ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type7ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type7ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type7ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type7ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type7ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type7ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type7ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type7ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type7ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type7ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type7ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type7ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type7ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type7ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type7ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type7ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type7ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type7ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type7ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type7ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type7ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type7ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj _Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type7ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj _Z13mul_mat_vec_qIL9ggml_type6ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type6ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type6ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type6ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type6ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type6ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type6ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type6ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type6ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type6ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type6ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type6ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type6ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type6ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type6ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type6ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type6ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type6ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type6ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type6ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type6ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type6ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type6ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type6ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type6ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type6ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type6ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type6ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type6ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type6ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type6ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type6ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type6ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type6ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type6ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type6ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type6ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type6ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type6ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type6ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type6ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type6ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type6ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type6ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj _Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type6ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj _Z13mul_mat_vec_qIL9ggml_type3ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type3ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type3ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type3ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type3ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type3ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type3ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type3ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type3ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type3ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type3ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type3ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type3ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type3ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type3ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type3ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type3ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type3ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type3ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type3ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type3ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type3ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type3ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type3ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type3ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type3ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type3ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type3ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type3ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type3ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type3ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type3ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type3ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type3ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type3ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type3ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type3ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type3ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type3ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type3ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type3ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type3ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type3ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type3ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj _Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant2._Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type3ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj _Z13mul_mat_vec_qIL9ggml_type2ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type2ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type2ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type2ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type2ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type2ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type2ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type2ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type2ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type2ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type2ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type2ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type2ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type2ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type2ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type2ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type2ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type2ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type2ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type2ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type2ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type2ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type2ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type2ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type2ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type2ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type2ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type2ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type2ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type2ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type2ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type2ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type2ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type2ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type2ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type2ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type2ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type2ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type2ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type2ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type2ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type2ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb0ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type2ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type2ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj _Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant2._Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type2ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj _Z13mul_mat_vec_qIL9ggml_type42ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type42ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type42ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type42ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type42ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type42ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type42ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type42ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type42ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type42ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type42ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type42ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type42ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type42ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type42ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type42ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type42ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type42ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type42ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type42ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type42ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type42ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type42ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type42ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type42ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type42ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type42ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type42ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type42ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type42ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type42ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type42ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type42ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type42ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type42ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type42ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type42ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type42ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type42ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type42ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type42ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type42ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type42ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type42ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj _Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type42ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj _Z13mul_mat_vec_qIL9ggml_type41ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type41ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type41ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type41ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type41ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type41ELi8ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type41ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type41ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type41ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type41ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type41ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type41ELi7ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type41ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type41ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type41ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type41ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type41ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type41ELi6ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type41ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type41ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type41ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type41ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type41ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type41ELi5ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type41ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type41ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type41ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type41ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type41ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type41ELi4ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type41ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type41ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type41ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type41ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type41ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type41ELi3ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type41ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type41ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type41ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type41ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type41ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type41ELi2ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant2._Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb0ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type41ELi1ELb0ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .text._Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.info._Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.shared._Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .rel.nv.constant0._Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj .nv.constant0._Z13mul_mat_vec_qIL9ggml_type41ELi1ELb1ELb1ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjS7_jjjS7_jjjj _Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb0EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj _Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .text._Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.info._Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.shared._Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .rel.nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .nv.constant0._Z17mul_mat_vec_q_moeIL9ggml_type41ELi2ELb1EEvPKvS2_PKi31ggml_cuda_mm_fusion_args_devicePfj5uint3jjjjjjjjj .debug_frame .rel.nv.constant4 .rel.debug_frame .rela.debug_frame .nv.callgraph .nv.prototype .nv.rel.action                               )                     9                     R   ê        €@      È    ê                Ã   (                D   '                T   ' h             `   I                n   '(h      €       {   '               „   '              �   '              ›   '               ¥   ' <             ±   ' @             »   'h             É   'h             Õ   ' H              ã   J                ï   K                s  ë        €;      é   ë                ä   )                e   L                q   M                õ  ì         7      k   ì                f	   *                ç	   N                ó
   O                w  í         2      í   í                è   +                i   P                u   Q                ù  î        €/      o   î                j   ,                ë   R                ÷   S                {  ï         *      ñ   ï                ì   -                m   T                y   U                ý  ð        €$      s   ð                n   .                ï   V                û   W                  ñ        €      õ   ñ                ð   /                q   X                }   Y                  ò        €%      w   ò                r   0                {   Z                ÿ  ó         5      u    ó                p!   1                y"   [                ý"  ô         ¢      s#   ô                n$   2                ï$   \                û%   ]                &  õ               ê&   õ                Â(   ^                ;)  ö         :      ¦)   ö                ~+   _                ÷+  ÷         :      m,   ÷                h-   3                q.   `                õ.  ø        €3      k/   ø                f0   4                o1   a                ó1  ù        €.      i2   ù                d3   5                m4   b                ñ4  ú         )      g5   ú                b6   6                k7   c                ï7  û        €&      e8   û                `9   7                i:   d                í:  ü         !      c;   ü                ^<   8                g=   e                ë=  ý               a>   ý                \?   9                e@   f                é@  þ               _A   þ                ZB   :                cC   g                çC  ÿ        €      ]D   ÿ                XE   ;                aF   h                åF            '      [G                    VH   <                _I   i                ãI           ƒ      YJ                   TK   =                ]L   j                áL                 LM                   $O   k                �O          €+      P                   àQ   l                YR          €,      ÏR                   ÊS   >                ÓT   m                WU          €(      ÍU                   ÈV   ?                ÑW   n                UX          €$      ËX                   ÆY   @                ÏZ   o                S[          €!      É[                   Ä\   A                Í]   p                Q^                 Ç^                   Â_   B                Ë`   q                Oa  	               Åa   	                Àb   C                Éc   r                Md  
        €      Ãd   
                ¾e   D                Çf   s                Kg                 Ág                   ¼h   E                Åi   t                Ij                 ¿j                   ºk   F                ;l   u                Gm   v                Ëm                 An                   <o   G                Ep   w                Ép                 ?q                   :r   H                »r   x                Çs   y                Kt                 Át                   ¼u   I                Åv   z                Iw          €g      ¿w                   ºx   J                Ãy   {                Gz          €      ²z                   Š|   |                }          €      n}                   F   }                ¿          €E      5€                   0�   K                ±�   ~                ½‚                   Aƒ          €>      ·ƒ                   ²„   L                3…   €                ?†   �                Ã†           8      9‡                   4ˆ   M                µˆ   ‚                Á‰   ƒ                EŠ           1      »Š                   ¶‹   N                7Œ   „                C�   …                Ç�           -      =Ž                   8�   O                ¹�   †                Å�   ‡                I‘          €%      ¿‘                   º’   P                ;“   ˆ                G”   ‰                Ë”          €      A•                   <–   Q                ½–   Š                É—   ‹                M˜          €      Ã˜                   ¾™   R                ?š   Œ                K›   �                Ï›                 Eœ                   @�   S                Á�   Ž                Íž   �                QŸ           &      ÇŸ                   Â    T                C¡   �                O¢   ‘                Ó¢          €�      I£                   D¤   U                Å¤   ’                Ñ¥   “                U¦                 À¦                   ¨   ”                ©   •                Š©           +      õ©                   P«   –                F¬   —                ¿¬            5      5­                    0®   V                9¯   ˜                ½¯  !         0      3°   !                .±   W                7²   ™                »²  "        €)      1³   "                ,´   X                5µ   š                ¹µ  #        €$      /¶   #                *·   Y                3¸   ›                ·¸  $        €!      -¹   $                (º   Z                1»   œ                µ»  %        €      +¼   %                &½   [                /¾   �                ³¾  &        €      )¿   &                $À   \                -Á   ž                ±Á  '        €      'Â   '                "Ã   ]                £Ã   Ÿ                ¯Ä                    3Å  (               ©Å   (                ¤Æ   ^                %Ç   ¡                1È   ¢                µÈ  )               +É   )                &Ê   _                /Ë   £                ³Ë  *        €l      )Ì   *                $Í   `                -Î   ¤                ±Î  +        €      Ï   +                ôÐ   ¥                mÑ  ,        €       ØÑ   ,                °Ó   ¦                )Ô  -         D      ŸÔ   -                šÕ   a                Ö   §                '×   ¨                «×  .        €>      !Ø   .                Ù   b                �Ù   ©                ©Ú   ª                -Û  /         8      £Û   /                žÜ   c                Ý   «                +Þ   ¬                ¯Þ  0         2      %ß   0                 à   d                ¡à   ­                ­á   ®                1â  1         /      §â   1                ¢ã   e                #ä   ¯                /å   °                ³å  2         )      )æ   2                $ç   f                ¥ç   ±                ±è   ²                5é  3         #      «é   3                ¦ê   g                'ë   ³                3ì   ´                ·ì  4        €      -í   4                (î   h                ©î   µ                µï   ¶                9ð  5         $      ¯ð   5                ªñ   i                ³ò   ·                7ó  6        €3      ­ó   6                ¨ô   j                ±õ   ¸                5ö  7         �      «ö   7                ¦÷   k                'ø   ¹                3ù   º                ·ù  8        €      "ú   8                úû   »                sü  9         8      Þü   9                ¶þ   ¼                /ÿ  :        €D      ¥ÿ   :                    l                ©  ½                - ;         =      £  ;                ž  m                §  ¾                + <        €7      ¡  <                œ  n                ¥  ¿                ) =        €1      Ÿ  =                š	  o                £
  À                ' >         .      �  >                ˜  p                ¡  Á                % ?        €'      ›  ?                –  q                Ÿ  Â                # @        €!      ™  @                ”  r                �  Ã                ! A        €      —  A                ’  s                  Ä                  Å                £ B         "        B                  t                •  Æ                ¡  Ç                % C        €/      ›  C                –  u                Ÿ  È                # D         •      ™  D                ”  v                �   É                !! E        €      Œ!  E                d#  Ê                Ý# F         4      H$  F                 &  Ë                ™& G         E      '  G                
(  w                )  Ì                —) H         ?      *  H                +  x                ,  Í                •, I         8      -  I                .  y                /  Î                “/ J        €1      	0  J                1  z                2  Ï                ‘2 K        €.      3  K                4  {                5  Ð                �5 L        €'      6  L                 7  |                	8  Ñ                �8 M        €!      9  M                þ9  }                ;  Ò                ‹; N        €      <  N                ü<  ~                >  Ó                ‰> O        €"      ÿ>  O                ú?                  A  Ô                ‡A P         0      ýA  P                øB  €                D  Õ                …D Q        €•      ûD  Q                öE  �                ÿF  Ö                ƒG R               îG  R                ÆI  ×                ?J S        €4      ªJ  S                ‚L  Ø                ûL T         ?      qM  T                lN  ‚                uO  Ù                ùO U        €;      oP  U                jQ  ƒ                sR  Ú                ÷R V        €6      mS  V                hT  „                qU  Û                õU W         1      kV  W                fW  …                oX  Ü                óX X         .      iY  X                dZ  †                m[  Ý                ñ[ Y         &      g\  Y                b]  ‡                k^  Þ                ï^ Z        €       e_  Z                ``  ˆ                ia  ß                ía [        €      cb  [                ^c  ‰                gd  à                ëd \         !      ae  \                \f  Š                Ýf  á                ég  â                mh ]        €.      ãh  ]                Þi  ‹                çj  ã                kk ^        €’      ák  ^                Ül  Œ                åm  ä                in _               Ôn  _                ¬p  å                %q `         3      �q  `                hs  æ                ás a        €/      Wt  a                Ru  �                [v  ç                ßv b         +      Uw  b                Px  Ž                Yy  è                Ýy c        €&      Sz  c                N{  �                W|  é                Û| d         "      Q}  d                L~  �                U  ê                Ù e        €      O€  e                J�  ‘                S‚  ë                ×‚ f        €      Mƒ  f                H„  ’                Q…  ì                Õ… g        €      K†  g                F‡  “                Oˆ  í                Óˆ h               I‰  h                DŠ  ”                ÅŠ  î                Ñ‹  ï                UŒ i        €      ËŒ  i                Æ�  •                GŽ  ð                S�  ñ                ×� j               M�  j                H‘  –                É‘  ò                Õ’  ó                Y“ k        €      Ï“  k                Ê”  —                K•  ô                W–  õ                Û– l        €      Q—  l                L˜  ˜                U™  ö                Ù™ m        €h      Oš  m                J›  ™                Sœ  ÷                ×œ n        €      B�  n                Ÿ  ø                “Ÿ o               þŸ  o                Ö¡  ù                O¢ p        €<      Å¢  p                À£  š                A¤  ú                M¥  û                Ñ¥ q        €6      G¦  q                B§  ›                Ã§  ü                Ï¨  ý                S© r        €0      É©  r                Äª  œ                E«  þ                Q¬  ÿ                Õ¬ s        €*      K­  s                F®  �                Ç®                   Ó¯                  W° t        €&      Í°  t                È±  ž                I²                  U³                  Ù³ u                O´  u                Jµ  Ÿ                Ëµ                  ×¶                  [· v        €      Ñ·  v                Ì¸                   M¹                  Yº                  Ýº w        €      S»  w                N¼  ¡                Ï¼                  Û½  	                _¾ x               Õ¾  x                Ð¿  ¢                QÀ  
                ]Á                  áÁ y        €      WÂ  y                RÃ  £                ÓÃ                  ßÄ                  cÅ z               ÙÅ  z                ÔÆ  ¤                UÇ                  aÈ                  åÈ {               [É  {                VÊ  ¥                ×Ê                  ãË                  gÌ |         q      ÝÌ  |                ØÍ  ¦                YÎ                  eÏ                  éÏ }               TÐ  }                ¯Ñ                  ¥Ò                  Ó ~        €"      ‰Ó  ~                äÔ                  ÚÕ                  SÖ         €:      ÉÖ                  Ä×  §                EØ                  QÙ                  ÕÙ €        €4      KÚ  €                FÛ  ¨                ÇÛ                  ÓÜ                  WÝ �        €.      ÍÝ  �                ÈÞ  ©                Iß                  Uà                  Ùà ‚        €(      Oá  ‚                Jâ  ª                Ëâ                  ×ã                  [ä ƒ        €$      Ñä  ƒ                Ìå  «                Mæ                   Yç  !                Ýç „        €      Sè  „                Né  ¬                Ïé  "                Ûê  #                _ë …        €      Õë  …                Ðì  ­                Qí  $                ]î  %                áî †        €      Wï  †                Rð  ®                Óð  &                ßñ  '                cò ‡               Ùò  ‡                Ôó  ¯                Uô  (                aõ  )                åõ ˆ        €      [ö  ˆ                V÷  °                ×÷  *                ãø  +                gù ‰               Ýù  ‰                Øú  ±                Yû  ,                eü  -                éü Š        €      _ý  Š                Zþ  ²                Ûþ  .                çÿ  /                k  ‹         i      á   ‹                Ü  ³                ]  0                i  1                í Œ               X  Œ                ³  2                ©  3                " �               �  �                è  4                Þ	  5                W
 Ž         G      Í
  Ž                È  ´                Ñ  6                U �        €A      Ë  �                Æ  µ                Ï  7                S �        €;      É  �                Ä  ¶                Í  8                Q ‘         6      Ç  ‘                Â  ·                Ë  9                O ’        €2      Å  ’                À  ¸                É  :                M “         ,      Ã  “                ¾  ¹                Ç  ;                K ”        €%      Á  ”                ¼  º                Å  <                I •               ¿  •                º   »                Ã!  =                G" –        €$      ½"  –                ¸#  ¼                Á$  >                E% —        €0      »%  —                ¶&  ½                7'  ?                C(  @                Ç( ˜         ”      =)  ˜                8*  ¾                ¹*  A                Å+  B                I, ™        €      ´,  ™                Œ.  C                / š        €8      p/  š                H1  D                Á1 ›         G      72  ›                23  ¿                ;4  E                ¿4 œ        €?      55  œ                06  À                97  F                ½7 �        €8      38  �                .9  Á                7:  G                »: ž        €0      1;  ž                ,<  Â                5=  H                ¹= Ÿ         +      />  Ÿ                *?  Ã                3@  I                ·@           #      -A                   (B  Ä                1C  J                µC ¡               +D  ¡                &E  Å                /F  K                ³F ¢        €      )G  ¢                $H  Æ                -I  L                ±I £        €      'J  £                "K  Ç                £K  M                ¯L  N                3M ¤               ©M  ¤                ¤N  È                ­O  O                1P ¥         q      §P  ¥                ¢Q  É                «R  P                /S ¦               šS  ¦                rU  Q                ëU §         #      VV  §                .X  R                §X ¨        €B      Y  ¨                Z  Ê                ![  S                ¥[ ©         ;      \  ©                ]  Ë                ^  T                £^ ª        €5      _  ª                `  Ì                a  U                ¡a «         /      b  «                c  Í                d  V                Ÿd ¬        €+      e  ¬                f  Î                g  W                �g ­        €$      h  ­                i  Ï                j  X                ›j ®        €      k  ®                l  Ð                m  Y                ™m ¯               n  ¯                
o  Ñ                p  Z                —p °        €       q  °                r  Ò                s  [                •s ±         )      t  ±                u  Ó                v  \                “v ²         ˆ      	w  ²                x  Ô                y  ]                ‘y ³        €      üy  ³                Ô{  ^                M| ´         /      ¸|  ´                �~  _                	 µ        €.        µ                z€  Õ                ƒ�  `                ‚ ¶         *      }‚  ¶                xƒ  Ö                �„  a                … ·        €&      {…  ·                v†  ×                ‡  b                ˆ ¸         $      yˆ  ¸                t‰  Ø                }Š  c                ‹ ¹        €       w‹  ¹                rŒ  Ù                {�  d                ÿ� º               uŽ  º                p�  Ú                y�  e                ý� »               s‘  »                n’  Û                w“  f                û“ ¼               q”  ¼                l•  Ü                u–  g                ù– ½               o—  ½                j˜  Ý                ë˜  h                ÷™  i                {š ¾               ñš  ¾                ì›  Þ                õœ  j                y� ¿         p      ï�  ¿                êž  ß                óŸ  k                w  À               â   À                º¢  l                3£ Á        €"      ž£  Á                v¥  m                ï¥ Â        €%      d¦  Â                ]§  à                d¨  n                ç¨ Ã         "      \©  Ã                Uª  á                \«  o                ß« Ä               T¬  Ä                M­  â                T®  p                ×® Å               L¯  Å                E°  ã                L±  q                Ï± Æ               D²  Æ                =³  ä                D´  r                Ç´ Ç        €      <µ  Ç                5¶  å                <·  s                ¿· È        €      4¸  È                -¹  æ                4º  t                ·º É        €      ,»  É                %¼  ç                ,½  u                ¯½ Ê               $¾  Ê                ¿  è                �¿  v                §À  w                *Á Ë        €      ŸÁ  Ë                ˜Â  é                ŸÃ  x                "Ä Ì               —Ä  Ì                �Å  ê                Æ  y                Ç  z                �Ç Í        €      È  Í                É  ë                Ê  {                •Ê Î        €U      
Ë  Î                Ì  ì                ƒÌ  |                �Í  }                Î Ï               zÎ  Ï                NÐ  ~                ÆÐ Ð               0Ñ  Ð                ˆÒ                  |Ó  €                ôÓ Ñ        €/      iÔ  Ñ                bÕ  í                iÖ  �                ìÖ Ò         +      a×  Ò                ZØ  î                aÙ  ‚                äÙ Ó        €&      YÚ  Ó                RÛ  ï                YÜ  ƒ                ÜÜ Ô        €!      QÝ  Ô                JÞ  ð                Qß  „                Ôß Õ               Ià  Õ                Bá  ñ                Iâ  …                Ìâ Ö        €      Aã  Ö                :ä  ò                Aå  †                Äå ×               9æ  ×                2ç  ó                9è  ‡                ¼è Ø               1é  Ø                *ê  ô                ªê  ˆ                ´ë  ‰                7ì Ù        €      ¬ì  Ù                ¥í  õ                %î  Š                /ï  ‹                ²ï Ú               'ð  Ú                 ñ  ö                 ñ  Œ                ªò  �                -ó Û        €      ¢ó  Û                ›ô  ÷                õ  Ž                %ö  �                ¨ö Ü        €      ÷  Ü                ø  ø                ù  �                 ù Ý         i      ú  Ý                û  ù                ü  ‘                ˜ü Þ               ý  Þ                Öþ  ’                Nÿ ß                ¸ÿ  ß                Œ  “                 à        €/      y  à                r  ú                y  ”                ü á        €*      q  á                j  û                q  •                ô â        €&      i  â                b	  ü                i
  –                ì
 ã         "      a  ã                Z  ý                a  —                ä ä        €      Y  ä                R  þ                Y  ˜                Ü å               Q  å                J  ÿ                Q  ™                Ô æ        €      I  æ                B                   I  š                Ì ç        €      A  ç                :                  º  ›                Ä  œ                G è        €      ¼  è                µ                  5  �                ?  ž                Â é        €      7  é                0                  °  Ÿ                º                    =! ê        €      ²!  ê                «"                  +#  ¡                5$  ¢                ¸$ ë               -%  ë                &&                  ¦&  £                °'  ¤                3( ì         k      ¨(  ì                ¡)                  ¨*  ¥                ++ í        €      •+  í                i-  ¦                á- î        €!      K.  î                0  §                —0 ï         +      1  ï                2                  3  ¨                �3 ð        €&      4  ð                ý4                  6  ©                ‡6 ñ         "      ü6  ñ                õ7  	                ü8  ª                9 ò               ô9  ò                í:  
                ô;  «                w< ó        €      ì<  ó                å=                  ì>  ¬                o? ô        €      ä?  ô                Ý@                  äA  ­                gB õ        €      ÜB  õ                ÕC                  ÜD  ®                _E ö        €      ÔE  ö                ÍF                  MG  ¯                WH  °                ÚH ÷        €      OI  ÷                HJ                  ÈJ  ±                ÒK  ²                UL ø        €      ÊL  ø                ÃM                  CN  ³                MO  ´                ÐO ù        €      EP  ù                >Q                  ¾Q  µ                ÈR  ¶                KS ú               ÀS  ú                ¹T                  ÀU  ·                CV û         X      ¸V  û                ±W                  ¸X  ¸                ;Y ü        €
      ¥Y  ü                y[  ¹                ñ[ ý        €      [\  ý                ³]  º                §^  »                _ þ        €*      ”_  þ                �`                  ”a  ¼                b ÿ         &      Œb  ÿ                …c                  Œd  ½                e           "      „e                   }f                  „g  ¾                h                |h                  ui                  |j  ¿                ÿj                tk                  ml                  tm  À                ÷m         €      ln                  eo                  lp  Á                ïp                dq                  ]r                  ds  Â                çs                \t                  Uu                  \v  Ã                ßv                Tw                  Mx                  Íx  Ä                ×y  Å                Zz                Ïz                  È{                  Ï|  Æ                R}                Ç}                  À~                  @  Ç                J€  È                Í€ 	               B�  	                ;‚                  Bƒ  É                Åƒ 
         Z      :„  
                3…                   :†  Ê                ½†                '‡                  ûˆ  Ë                s‰                Ý‰                  5‹  Ì                )Œ  Í                ¡Œ         €1      �                  Ž  !                �  Î                Ÿ�         €,      �                  ‘  "                ’  Ï                �’         €)      “                  ”  #                •  Ð                ›•          %      –                  —  $                ˜  Ñ                ™˜         €"      ™                  
š  %                ›  Ò                —›                œ                  �  &                ž  Ó                •ž                Ÿ                     '                ¡  Ô                “¡         €      	¢                  £  (                ¤  Õ                ‘¤         €      ¥                  ¦  )                ƒ¦  Ö                �§  ×                ¨                ‰¨                  „©  *                �ª  Ø                «          i      ‡«                  ‚¬  +                ‹­  Ù                ®         €      z®                  R°  Ú                Ë°                6±                  ³  Û                ‡³         €2      ý³                  ø´  ,                ¶  Ü                …¶          .      û¶                  ö·  -                ÿ¸  Ý                ƒ¹         €+      ù¹                  ôº  .                ý»  Þ                �¼         €&      ÷¼                  ò½  /                û¾  ß                ¿          $      õ¿                  ðÀ  0                ùÁ  à                }Â                óÂ                  îÃ  1                ÷Ä  á                {Å                 ñÅ                   ìÆ  2                õÇ  â                yÈ !               ïÈ  !                êÉ  3                óÊ  ã                wË "        €      íË  "                èÌ  4                iÍ  ä                uÎ  å                ùÎ #        €      oÏ  #                jÐ  5                sÑ  æ                ÷Ñ $        €o      mÒ  $                hÓ  6                qÔ  ç                õÔ %        €      `Õ  %                8×  è                ±× &        €       Ø  &                ôÙ  é                mÚ                   ¯Ú  E                ËÚ  F                ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4                       €@         l   �€€( ˆ         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       p               €;         d   �€€( L         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       à                7         d   �€€( $         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       P               2         d   �€€( è         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       À              €/         T   �€€( P         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       0               *         T   �€€( è	         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4                      €$         L   �€€( œ         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4                     €         „   �€€( ä         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       €              €%         d   �€€( À         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       ð               5         ˜   �€€( h         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       `               ¢         |   �€€( Ð'         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       Ð                           �€€(           ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       @               :            �€€( @         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       °               :         `   �€€( ô         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4                      €3         \   �€€( @         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       �              €.         \   �€€(          ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4                       )         `   �€€( °	         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       p              €&         d   �€€( 	         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       à               !         d   �€€(           ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       P                        \   �€€( `         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       À                        „   �€€(          ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       0	              €         d   �€€( ¸         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4        	               '         œ   �€€( ð         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       
               ƒ         x   �€€(           ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       €
                           �€€( ø         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       ð
              €+            �€€( ˜
         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       `              €,         `   �€€( x
         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       Ð              €(         `   �€€( x	         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       @              €$         \   �€€( |         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       °              €!         L   �€€( Ø         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4                                T   �€€( ø         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       �                        X   �€€( ´         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4                      €         P   �€€(           ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       p                        |   �€€(          ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       à                        h   �€€(          ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       P                        |   �€€(          ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       À                        h   �€€(          ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       0                        �   �€€( l         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4                      €g         x   �€€( 0         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4                     €            �€€( d         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       €              €            �€€( X         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       ð              €E         \   �€€( À         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       `              €>         `   �€€( ü         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       Ð               8         h   �€€( T         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       @               1         h   �€€( ¬         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       °               -         \   �€€( œ
         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4                      €%         \   �€€( ¼         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       �              €         X   �€€(           ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4                      €         „   �€€(          ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       p                        \   �€€( ì         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       à               &         ”   �€€( ´         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       P              €�         |   �€€( ¤         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       À                           �€€( 0         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       0               +            �€€( |
         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4                       5         h   �€€( ˜         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4                      0         \   �€€( `         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       €              €)         d   �€€( ¼	         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       ð              €$         `   �€€( x         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       `              €!         \   �€€( Ä         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       Ð              €         `   �€€( P         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       @              €         X   �€€(           ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       °              €         „   �€€( 0         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4                                X   �€€( |         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       �                        ˜   �€€( ø         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4                      €l         œ   �€€( <         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       p              €            �€€( ”         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       à              €             �€€( Ô         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       P               D         `   �€€( `         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       À              €>         \   �€€(          ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       0               8         h   �€€( h         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4                       2         `   �€€( è         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4                      /         `   �€€( 0         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       €               )         \   �€€( ¤	         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       ð               #         T   �€€( <         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       `               €         „   �€€( °         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       Ð                $         \   �€€( l         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       @!              €3         ˜   �€€(          ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       °!               �         ˜   �€€( h&         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4        "              €            �€€( ¨         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       �"               8            �€€( ¸         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4        #              €D         d   �€€( „         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       p#               =         `   �€€( ´         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       à#              €7         \   �€€( T         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       P$              €1         `   �€€( À         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       À$               .         \   �€€( ì
         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       0%              €'         \   �€€( H	         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4        %              €!         T   �€€( Ì         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       &              €         €   �€€( l         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       €&               "         `   �€€( ð         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       ð&              €/         ”   �€€(          ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       `'               •         |   �€€( ”$         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       Ð'              €            �€€( $         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       @(               4            �€€( À         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       °(               E         d   �€€( ”         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4        )               ?         `   �€€(           ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       �)               8         `   �€€( d         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4        *              €1         \   �€€( Ô         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       p*              €.         `   �€€( ø
         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       à*              €'         \   �€€( T	         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       P+              €!         T   �€€( Ø         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       À+              €         „   �€€( ”         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       0,              €"         \   �€€(           ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4        ,               0         ˜   �€€( 0         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       -              €•         ˜   �€€( Œ$         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       €-                           �€€( 4         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       ð-              €4            �€€( è         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       `.               ?         h   �€€(          ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       Ð.              €;         X   �€€( @         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       @/              €6         X   �€€(          ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       °/               1         T   �€€( ¨         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4        0               .         L   �€€( ð
         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       �0               &         T   �€€( ø         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4        1              €          T   �€€( ”         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       p1              €         €   �€€( \         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       à1               !         X   �€€( ¸         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       P2              €.         ˜   �€€( È
         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       À2              €’         |   �€€( Ü#         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       03                           �€€(          ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4        3               3            �€€( |         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       4              €/         p   �€€( <         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       €4               +         t   �€€( 
         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       ð4              €&         p   �€€( ð         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       `5               "         h   �€€( Ô         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       Ð5              €         d   �€€( P         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       @6              €         `   �€€(          ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       °6              €         X   �€€( Ø         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4        7                        |   �€€( T         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       �7              €         `   �€€( 8         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4        8                        |   �€€( T         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       p8              €         `   �€€( 8         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       à8              €         ”   �€€( ˆ         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       P9              €h         |   �€€( d         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       À9              €            �€€( ¤         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       0:                           �€€( €         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4        :              €<         d   �€€( €         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       ;              €6         d   �€€( ü         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       €;              €0         h   �€€( |         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       ð;              €*         d   �€€( ô	         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       `<              €&         \   �€€( 	         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       Ð<                         \   �€€( `         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       @=              €         T   �€€( Ì         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       °=              €         „   �€€( h         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4        >                        d   �€€( œ         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       �>              €         „   �€€( h         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4        ?                        d   �€€( œ         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       p?                        ”   �€€( p         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       à?               q         x   �€€( Œ         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       P@                           �€€( ô         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       À@              €"            �€€( T         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       0A              €:         h   �€€(           ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4        A              €4         d   �€€( x         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       B              €.         P   �€€(          ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       €B              €(         \   �€€( „	         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       ðB              €$         d   �€€( �         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       `C              €         `   �€€( Ô         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       ÐC              €         X   �€€( D         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       @D              €         €   �€€( ´         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       °D                        d   �€€(           ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4        E              €         €   �€€( ´         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       �E                        d   �€€(           ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4        F              €         ”   �€€( œ         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       pF               i         x   �€€( ˜         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       àF                           �€€( |         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       PG                           �€€( t         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       ÀG               G         d   �€€( 0         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       0H              €A         d   �€€( À         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4        H              €;         X   �€€( \         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       I               6         \   �€€( è         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       €I              €2         X   �€€(          ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       ðI               ,         T   �€€( l
         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       `J              €%         L   �€€( è         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       ÐJ                        „   �€€(          ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       @K              €$         d   �€€( �         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       °K              €0         Œ   �€€( P         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4        L               ”         ˜   �€€( <$         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       �L              €            �€€(          ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4        M              €8            �€€( è         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       pM               G         d   �€€(          ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       àM              €?         l   �€€( <         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       PN              €8         X   �€€( „         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       ÀN              €0         \   �€€( Œ         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       0O               +         \   �€€( 0
         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4        O               #         `   �€€(           ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       P                        T   �€€( (         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       €P              €         x   �€€(           ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       ðP              €         \   �€€( È         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       `Q                        ˜   �€€( ¸         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       ÐQ               q         x   �€€( „         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       @R                           �€€(          ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       °R               #            �€€( l         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4        S              €B         \   �€€(          ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       �S               ;         \   �€€( $         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4        T              €5         X   �€€( È         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       pT               /         d   �€€( 0         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       àT              €+         \   �€€( H
         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       PU              €$         T   �€€( ”         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       ÀU              €         L   �€€(           ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       0V                        |   �€€( <         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4        V              €          \   �€€( �         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       W               )         �   �€€( h	         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       €W               ˆ         œ   �€€( !         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       ðW              €            �€€( \         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       `X               /            �€€( |         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       ÐX              €.         h   �€€(           ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       @Y               *         h   �€€( è	         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       °Y              €&         \   �€€( 	         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4        Z               $         X   �€€( `         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       �Z              €          T   �€€( ”         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4        [                        X   �€€( <         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       p[                        P   �€€( 0         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       à[                        t   �€€( T         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       P\                        h   �€€( œ         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       À\                        �   �€€( €         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       0]               p         t   �€€( P         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4        ]                           �€€( ô         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       ^              €"            �€€( \         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       €^              €%         p   �€€( ¼         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       ð^               "         p   �€€( Ì         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       `_                        d   �€€( ð         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       Ð_                        d   �€€( à         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       @`                        T   �€€( €         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       °`              €         \   �€€( X         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4        a              €         T   �€€( D         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       �a              €         Œ   �€€( ”         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4        b                        d   �€€( °         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       pb              €         Œ   �€€( ”         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       àb                        d   �€€( °         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       Pc              €         œ   �€€(          ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       Àc              €U         |   �€€( ´         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       0d                           �€€( ,         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4        d                           �€€( ¸         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       e              €/         `   �€€( L         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       €e               +         d   �€€( 
         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       ðe              €&         `   �€€( 	         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       `f              €!         X   �€€( Ø         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       Ðf                        T   �€€( <         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       @g              €         \   �€€( Ø         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       °g                        P   �€€( ¨         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4        h                        €   �€€( ø         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       �h              €         T   �€€( L         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4        i                        €   �€€( ø         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       pi              €         T   �€€( L         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       ài              €         �   �€€( Ü         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       Pj               i         t   �€€( �         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       Àj                           �€€( ˆ         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       0k                            �€€( ¸         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4        k              €/         p   �€€( (         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       l              €*         d   �€€( 
         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       €l              €&         h   �€€( 	         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       ðl               "         `   �€€( ä         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       `m              €         d   �€€( <         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       Ðm                        `   �€€( ä         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       @n              €         L   �€€( è         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       °n              €         t   �€€( 8         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4        o              €         X   �€€( X         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       �o              €         t   �€€( 8         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4        p              €         X   �€€( X         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       pp                        �   �€€( 0         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       àp               k         t   �€€(          ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       Pq              €            �€€( Ì         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       Àq              €!            �€€(          ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       0r               +         p   �€€( 
         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4        r              €&         h   �€€( ð         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       s               "         d   �€€( Ü         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       €s                        \   �€€( °         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       ðs              €         T   �€€(          ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       `t              €         T   �€€( Ä         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       Ðt              €         L   �€€( �         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       @u              €         €   �€€( ¬         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       °u              €         h   �€€( ð         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4        v              €         €   �€€( ¬         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       �v              €         h   �€€( ð         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4        w                        ”   �€€( p         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       pw               X         x   �€€( T         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       àw              €
            �€€( \         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       Px              €            �€€( \         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       Àx              €*         p   �€€( ð	         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       0y               &         p   �€€( Ô         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4        y               "         h   �€€( Ø         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       z                        \   �€€( ´         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       €z                        `   �€€(           ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       ðz              €         \   �€€( Ð         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       `{                        T   �€€( ¬         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       Ð{                        €   �€€(          ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       @|                        x   �€€(          ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       °|                        €   �€€(          ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4        }                        x   �€€(          ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       �}                        ˜   �€€( ¤         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4        ~               Z         x   �€€( À         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       p~                           �€€( x         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       à~                           �€€( „         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       P              €1         d   �€€( ¸         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       À              €,         `   �€€( �
         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       0€              €)         T   �€€( Ì	         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4        €               %         P   �€€( À         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       �              €"         P   �€€(          ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       €�                        L   �€€( °         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       ð�                        P   �€€( ô         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       `‚              €         x   �€€(           ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       Ð‚              €         T   �€€( T         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       @ƒ                        �   �€€( è         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       °ƒ               i         x   �€€( �         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4        „              €            �€€(           ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       �„                           �€€( ,         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4        …              €2         h   �€€(          ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       p…               .         d   �€€( Ø
         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       à…              €+         T   �€€( L
         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       P†              €&         T   �€€( 	         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       À†               $         T   �€€( p         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       0‡                        P   �€€( ¬         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4        ‡                        T   �€€( 8         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       ˆ                        t   �€€(          ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       €ˆ              €         X   �€€( €         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       ðˆ              €         ˜   �€€( ”         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       `‰              €o         |   �€€( (         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       Ð‰              €            �€€( è         ÿÿÿÿ$       ÿÿÿÿÿÿÿÿ |ÿÿÿÿ�€€( ÿ�€(�€€(   ÿÿÿÿ4       @Š              €             �€€( Ø            Œ   Ð  NVIDIA Corp              6   f    ptxas Cuda compilation tools, release 13.4, V13.4.89 Build system must define TOOLS_VERSION_EXTENDED -arch sm_89         è  NVIDIA Corp  Y †   / 6  +    6      / 3  .    3      / /  @    /      / +  7    +      / &  *    &      / "  3    "      /   =          /   J          /   X          /   j          /   q          / 
  }    
      /   ˆ          /   .          /    ,           / ü  H    ü      / ø  8    ø      / ó  +    ó      / ï  8    ï      / ë  =    ë      / ç  H    ç      / ã  Y    ã      / ß  `    ß      / Û  o    Û      / ×  z    ×      / Ó  �    Ó      / Ï  /    Ï      / Ì  (    Ì      / È  H    È      / Ä  8    Ä      / ¿  '    ¿      / »  7    »      / ¶  '    ¶      / ²  7    ²      / ®  8    ®      / ª  @    ª      / ¦  H    ¦      / ¢  P    ¢      / ž  X    ž      / š  ^    š      / –  j    –      / ’  (    ’      / �  (    �      / ‹  8    ‹      / ‡  0    ‡      / ‚  "    ‚      / }  (    }      / x  "    x      / s  (    s      / o  ,    o      / k  8    k      / g  @    g      / c  G    c      / _  P    _      / [  U    [      / W  `    W      / T  7    T      / Q  /    Q      / M  P    M      / H  H    H      / C  (    C      / >  '    >      / 9  (    9      / 4  '    4      / 0  ?    0      / ,  P    ,      / (  P    (      / $  \    $      /    \           /   d          /   m          /   6          /   /          /   @          / 
  8    
      /   (          /    (           / û  (    û      / ö  (    ö      / ò  6    ò      / î  @    î      / ê  D    ê      / æ  H    æ      / â  R    â      / Þ  \    Þ      / Ú  c    Ú      / Ö  0    Ö      / Ó  P    Ó      / Î  @    Î      / Ê  8    Ê      / Å  0    Å      / Á  @    Á      / ¼  0    ¼      / ¸  @    ¸      / ´  0    ´      / °  8    °      / ¬  @    ¬      / ¨  H    ¨      / ¤  N    ¤      /    P           / œ  ^    œ      / ™  8    ™      / –  0    –      / ’  P    ’      / Ž  H    Ž      / ‰  (    ‰      / …  #    …      / �  8    �      / }  H    }      / y  H    y      / u  P    u      / q  ]    q      / m  d    m      / i  l    i      / f  @    f      / c  8    c      / _  e    _      / [  M    [      / W  7    W      / S  /    S      / O  O    O      / K  \    K      / G  n    G      / C  z    C      / ?  „    ?      / ;  “    ;      / 7  ž    7      / 4  8    4      / 1  0    1      / -  ]    -      / )  F    )      / $  0    $      /    (           /   P          /   _          /   i          /   v          /             /   Š          /   •          /   „          / þ  k    þ      / ù  Ë    ù      / ô  ~    ô      / ð  O    ð      / ì  @    ì      / è  ~    è      / ä  Š    ä      / à  •    à      / Ü  �    Ü      / Ø  §    Ø      / Ô  ¸    Ô      / Ð  À    Ð      / Ì  8    Ì      / È  0    È      / Ã  [    Ã      / ¾  @    ¾      / ¹  0    ¹      / ´  8    ´      / ¯  0    ¯      / ª  8    ª      / ¥  G    ¥      /    O           / ›  Y    ›      / –  `    –      / ‘  h    ‘      / Œ  q    Œ      / ‡      ‡      / ƒ  ;    ƒ      /   8          / z  n    z      / u  H    u      / p  8    p      / k  (    k      / f  8    f      / a  (    a      / \  H    \      / W  P    W      / R  ^    R      / M  h    M      / H  s    H      / C  v    C      / >  €    >      / ;  >    ;      / 8  8    8      / 4  `    4      / 0  H    0      / +  0    +      / &  (    &      / !  0    !      /   (          /   H          /   H          /   P          /   \          /   `          /   j          /    p           / ý  D    ý      / ú  @    ú      / ö  d    ö      / ò  V    ò      / í  3    í      / é  .    é      / å  P    å      / á  [    á      / Ý  h    Ý      / Ù  t    Ù      / Õ  €    Õ      / Ñ  Ž    Ñ      / Í  š    Í      / Ê  H    Ê      / Ç  G    Ç      / Ã  s    Ã      / ¿  `    ¿      / »  3    »      / ·  0    ·      / ³  W    ³      / ¯  `    ¯      / «  o    «      / §  {    §      / £  Š    £      / Ÿ  ’    Ÿ      / ›  ž    ›      / ˜  H    ˜      / •  G    •      / ‘  q    ‘      / �  ^    �      / ˆ  5    ˆ      / ƒ  /    ƒ      /   U          / {  `    {      / w  l    w      / s  y    s      / o  ‚    o      / k  �    k      / g  ›    g      / d  G    d      / a  G    a      / \  r    \      / X  [    X      / T  0    T      / O  7    O      / J  R    J      / E  `    E      / @  l    @      / ;  x    ;      / 6  …    6      / 1  Ž    1      / ,  �    ,      / )  >    )      / &  6    &      / "  ]    "      /   H          /   0          /   (          /   H          /   U          /   `          /   q          /    w           / ü   €    ü       / ø   �    ø       / ô   H    ô       / ð   G    ð       / ë   €    ë       / æ   `    æ       / á   ?    á       / Ü   0    Ü       / ×   _    ×       / Ò   m    Ò       / Í   x    Í       / È   €    È       / Ã   “    Ã       / ¾   ™    ¾       / ¹   £    ¹       / ¶   5    ¶       / ³   0    ³       / ¯   H    ¯       / «   8    «       / ¦   (    ¦       / ¢   &    ¢       / �   (    �       / ™   &    ™       / •   7    •       / ‘   @    ‘       / �   G    �       / ‰   N    ‰       / …   [    …       / �   _    �       / }   n    }       / z   @    z       / w   =    w       / s   d    s       / o   R    o       / k   0    k       / g   /    g       / c   S    c       / _   _    _       / [   k    [       / W   u    W       / S   ‚    S       / O   Š    O       / K   ™    K       / H   P    H       / E   N    E       / @   ‹    @       / <   k    <       / 8   7    8       / 3   7    3       / .   _    .       / )   o    )       / $   v    $       /    €           /    �           /    š           /    ¨                                            $        )        .        3        8        <        @        E        H        K        O        S        W        [        _        c        g        k        o        s        w        z        }        �        …        ‰        �        ‘        •        ™        �        ¢        ¦        «        ¯        ³        ¶        ¹        ¾        Ã        È        Í        Ò        ×        Ü        á        æ        ë        ð        ô        ø        ü                                                                 "       &       )       ,       1       6       ;       @       E       J       O       T       X       \       a       d       g       k       o       s       w       {              ƒ       ˆ       �       ‘       •       ˜       ›       Ÿ       £       §       «       ¯       ³       ·       »       ¿       Ã       Ç       Ê       Í       Ñ       Õ       Ù       Ý       á       å       é       í       ò       ö       ú       ý                                                                !       &       +       0       4       8       ;       >       C       H       M       R       W       \       a       f       k       p       u       z              ƒ       ‡       Œ       ‘       –       ›               ¥       ª       ¯       ´       ¹       ¾       Ã       È       Ì       Ð       Ô       Ø       Ü       à       ä       è       ì       ð       ô       ù       þ                                                                       $       )       -       1       4       7       ;       ?       C       G       K       O       S       W       [       _       c       f       i       m       q       u       y       }       �       …       ‰       Ž       ’       –       ™       œ               ¤       ¨       ¬       °       ´       ¸       ¼       Á       Å       Ê       Î       Ó       Ö       Ú       Þ       â       æ       ê       î       ò       ö       û                      
                                                  $       (       ,       0       4       9       >       C       H       M       Q       T       W       [       _       c       g       k       o       s       x       }       ‚       ‡       ‹       �       ’       –       š       ž       ¢       ¦       ª       ®       ²       ¶       »       ¿       Ä       È       Ì       Ï       Ó       ×       Û       ß       ã       ç       ë       ï       ó       ø       ü                             
                                          "       &       +       /       3       6      f    7 †   
    `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )@ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(@Ð/  ð/  `0  �0  À0   1  1  P1   1  Ð1   2  P2  Ð2  `3   3  à3  4  @4  p4  ð4  05  p5  �5  °5  Ð5   6  06  p6  °6  ð6  07  p7  °7  à7  8  @8  p8   8  À8  à8   9  @9  `9  €9  °9  Ð9  ð9   :   :  @:  `:  €:   :  À:  à:   ;   ;  P;  p;  �;  °;  Ð;  ð;   <   <  @<  P<  p<  �<   =  0=  P=  p=  �=  °=  à=   >   >  0>  @?   À.  `?  à?     @              f    7 †   
    `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ( -  À-  .  0.  P.  p.   .  à.  0/  `/  �/  À/  ð/  0   0  @0  `0  €0  °0  Ð0   1   1  P1  p1   1  Ð1  2  P2  €2  °2  à2  3  @3  p3   3  Ð3  4  @4  €4  °4  à4   5  P5  �5  À5  ð5   6  @6  p6   6  à6  7  @7  `7  €7   7  À7  à7   8   8  @8  `8  €8   8  À8  à8   9  `9  €9   9   ð+  °:  Ð:     @              f    7 †   
    `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )ð ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(ð @*  `*   *  À*  à*  +  @+  `+  �+  °+  Ð+  ð+  ,  `,  À,  ð,  -  0-  P-  €-  °-  à-   .   .  P.  �.  À.  à.   /   /  P/  €/  À/  ð/  00  `0   0  à0  1  01  P1  p1   1  Ð1  2  @2  p2   2  À2  à2  3  03  P3  p3   3  Ð3   4  @4  p4  �5   @)  °5  06     @              f    7 †   
 #   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )È ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(È  '  À'   (   (  @(  p(   (  À(  à(   )   )  `)  €)   )  À)  à)   *  *  0*  P*  p*  °*  à*   +  0+  `+  �+   +  Ð+   ,  0,  €,  À,   -  @-  €-  À-  à-  .  0.  `.  À.   /  @/  p/  �/  °/  À/  ð/  0   �&   1  @1     @              f    7 †   
 (   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )  ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(  @&  `&  €&   &  à&   '  0'  p'   '  °'  Ð'  ð'  (   (  @(  p(  �(  °(  Ð(  ð(  )   )  @)  Ð)   *   *  @*  `*  �*  Ð*   +  @+  �+  Ð+  ,  P,  �,  Ð,   -   .   °#   .   .     €              f    7 †   
 -   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )x ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(x �"   #  @#  p#  �#  °#  À#  ð#  $  0$  P$  p$  €$  °$  Ð$  ð$  %  0%  @%  p%  �%  À%   &  0&  p&  �&  P'  �'  Ð'   (   Ð   à(   )     €              f    7 †   
 2   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )P ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(P 0  P  p  °  Ð  ð      0   P   p   €   à    !   !  @!  °!  ð!   "  @"  #   à   #  °#     €              f    7 †   
 7   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  ) ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ( °  Ð  ð    @      P  °     €              f    7 †   
 ;   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )( ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ((   `  €     À  ð     0   P   p    À  �    $     €              f    7 †   
 ?   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )P ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(P 00  `0  €0   0  Ð0  ð0   1   1  P1  p1  €1   1  à1  ð1  2   2  `2  °2  ð2  03   À.  ð3  4     €              f    7 †   
 D   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )pÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(p€a   a  àa  0b  @b  Pb  pb   b  Àb  Ðb  ðb   c  @c  Pc  pc   c  àc   d  Pd  �d  @g  `g  Àg  Ðg   h  h  ph  €h  °h  Àh   i  Ài   j  0j  €j  �j   k  @k  �k   k  �l  °l  0m  @m  �m   m  àm  ðm  @n  `n  Po  Ào  ào   p   p  @p  `p  €p   p  Àp  Ðr  ðr  0s  @s  €s  �s   t  t  @t  Pt  pu  �u  v   v  €v  �v  ðv  w  pw  €w  �x  °x  0y  @y  €y  �y   z   z  `z  pz  €{  ð{  |  0|  P|  p|  �|  °|  Ð|  ð|  p  �  Ð  à   €  0€   €  °€  à€  ð€  ‚  0‚  °‚  À‚   ƒ  0ƒ  �ƒ  °ƒ  „   „  0…  P…  Ð…  à…   †  0†   †  À†   ‡  ‡   ˆ  �ˆ  °ˆ  Ðˆ  ðˆ  ‰  0‰  P‰  p‰  �‰  P�  €�   �  À�  à�   Ž  PŽ  �Ž  °Ž  àŽ  ðŽ   �   �  @�  `�  ��  À�   �  0�  `�  €�   �  À�  Ð�   ‘   ‘  @‘  `‘  €‘   ‘  °‘  Ð‘   ’   ’  @’  `’  €’   ’  °’  Ð’   •  À•   –  –  P–  `–  Ð–  à–  —   —  `˜  €˜   ™  ™  p™  €™  à™   š  `š  pš   ›  À›  @œ  Pœ  �œ   œ  �  0�  p�  €�  °ž   Ÿ  @Ÿ  `Ÿ  €Ÿ   Ÿ  ÀŸ  àŸ          4  _  Pe  €e  àp  �q  }  ð}  °‰  �Š  ”  `”  @   @¡     €              f    7 †   
 G   `„ „       €  ð       |  ð       x  ð       t  ð       p  ð      
 l  ð      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ _  )( ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ((      `  €  Ð  à       P  `   0   p  @                 f    7 †   
 J   `„ „       €  ð       |  ð       x  ð       t  ð       p  ð      
 l  ð      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ _  )x ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(x €/  �/  À/  Ð/   0  0  @0  P0  €0  �0  Ð0  à0  ð0   1  @1  `1  p1  €1  °1  Ð1  ð1   2  @2  `2  p2  €2  À2  à2  ð2   3   0    3  @9                 f    7 †   
 N   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )@ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(@ '   '  0'  p'  �'   '   (  @(  `(  à(   )  )  �)  °)  Ð)  0+  p+  �+  °+  Ð+   -  °-  À-   .   .  °.  Ð.  à.   /  0/  P/  p/  �/  À/  ð/  0  00  P0  p0   0  Ð0   1  01  `1  �1  À1  ð1   2  P2  €2  °2  à2   3  P3  �3  À3  ð3   4  `4  €4   4  À4   5  05  `5   5  à5   6  P6  €6  °6  Ð6  ð6  7  07  P7  `7  €7   7  À8   �&  à8  `9     @              f    7 †   
 R   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(p$  P%  p%  €%   %  Ð%  &  P&  €&  °&  à&   '  `'  �'  À'   (  0(  `(  �(  À(   )   )  P)  �)  À)  ð)   *  P*  €*  À*   +  @+  p+   +  Ð+   ,  0,  `,  �,  À,  ð,   -  P-  €-  °-  à-   .  0.  `.  €.   .  À.  à.   /   /  @/  `/  €/   /  À/  à/  0  00  P0  p0  �0  Ð0  ð0  1  P1   `#  `2  €2     @              f    7 †   
 V   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )ð ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(ð Ð!  ð!   "  @"  €"  °"  Ð"  ð"   #  0#  @#  `#  €#  Ð#   $  0$  €$   $  °$  Ð$  ð$   %  0%  `%   %  à%  &  @&  p&   &  Ð&   '  0'  `'  �'  À'  ð'  0(  `(   (  Ð(   )  0)  `)  �)  °)  à)   *   *  `*  €*   *  Ð*  +  P+  p+  �+  °+  Ð+  ð,   €   -  �-     @              f    7 †   
 Z   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )È ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(È    Ð        P  p  �  °  À  à      0   P   p   �   °   Ð   ð   !  @!  P!  €!  À!  0"  P"  p"  €"   "  À"  à"  #  `#   #  à#   $  `$   $  Ð$   %  0%  `%   %  Ð%   &  `&   &  À&  à&  '   '   °  0(  P(     @              f    7 †   
 ^   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )  ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(  ð    @  p     À  ð     P  p  �  °  Ð  ð         P   p   �   °   Ð   ð   !  @!  p!  �!  °!  Ð!  ð!   "  @"  p"  °"  à"  #  @#  p#  °#  ð#   %     @%  À%     €              f    7 †   
 b   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )x ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(x    �  °  à        0  `  €     À  à  ð     @  `  €     °  à     @  p  �     à     P  €      0             €              f    7 †   
 f   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )P ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(P 0  P  p     Ð  ð        P  p  €     ð       0  `  �  À  ð   p  p        €              f    7 †   
 j   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  ) ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ( 0  P  p  �  À      Ð  0     €              f    7 †   
 n   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )( ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ((   `  €     À  ð    0  P  p   À  �  €     €              f    7 †   
 r   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )P ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(P p"  �"  °"  Ð"   #   #  0#  P#  €#   #  °#  Ð#  @$  `$  €$  �$  à$  0%  p%   %   ð    &  @&     €              f    7 †   
 v   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )pÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(p°B  ÐB  ðB  @C  `C  pC  €C  ÀC  àC  ðC   D  0D  PD  €D   D  àD   E  PE  pE   E  0H  @H  �H   H  ÐH  àH  `I  pI  °I  ÀI  °J  ÐJ  @K  PK   K  ÀK  `L  pL   L  °L   M  ÀM  @N  PN   N  °N  O  0O  pO  €O  `P  ÐP  ðP  Q  0Q  PQ  pQ  �Q  °Q  ÐQ  àS  ðS  PT  pT  °T  ÀT   U  U  `U  pU  €V   V  W   W  �W   W   X  X  `X  €X   Y  ÀY  0Z  @Z  °Z  ÀZ   [  [  P[  p[  �\   ]   ]  @]  `]  €]   ]  À]  à]   ^  €`  �`  ð`  a  Pa  `a   a  °a   b  b   c  @c  °c  Àc  0d  @d   d  °d   e   e  @f  `f  àf  ðf  @g  Pg  �g   g  ðg  h  0i   i  Ài  ài   j   j  @j  `j  €j   j  @n  `n  €n   n  Àn  ðn  €o  °o  Ðo   p  0p  `p  €p   p  Ðp  ðp  q  0q  Pq  pq  €q   q  Ðq  ðq  r  0r  Pr  pr  €r   r  Ðr  ðr  s  0s  Ps  ps  €s   s  às   t  °v  Àv   w  @w  €w  �w  Ðw  àw  0x  @x  py  �y   z  z  €z  �z  ðz   {  P{  p{  °|  Ð|  P}  `}  Ð}  à}  ~   ~  `~  €~  À  0€  P€  p€  �€  °€  Ð€  ð€  �  0�  4 °@  `F  �F  ðQ   R   ^   _  Àj   k   u  pu  P�  P‚     €              f    7 †   
 y   `„ „       €  ð       |  ð       x  ð       t  ð       p  ð      
 l  ð      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ _  )( ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ((       p  €  °  À  ð     0  @   0   P                    f    7 †   
 |   `„ „       €  ð       |  ð       x  ð       t  ð       p  ð      
 l  ð      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ _  )x ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(x °   À   ð    !  0!  @!  p!  €!  °!  À!   "  "   "  0"  `"  €"   "  °"  ð"  #   #  0#  `#  €#   #  °#  à#   $   $  0$   0   Ð$   *                 f    7 †   
 €   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )@ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(@P  �     p    @  €  À     0  `  Ð    `  �  Ð      0   p       ð    !  @!  `!  €!   !  À!  à!   "  P"  €"  °"  à"   #  `#   #  Ð#   $  0$  P$  p$   $  À$  à$  %  0%  `%  €%  °%  Ð%  ð%  &  0&  P&  p&  �&  °&  à&   '   '  @'  `'  €'  �'  °'  Ð'  à'   (   (  `(  °(  à(   )  0)  @)  p)  �)  °)  À)  Ð*   P  ð*  p+     @              f    7 †   
 „   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(°  P  €  À  à  ð    @  p  Ð     0  `  �  Ð  à        @  p     Ð        @  p     à    P  �  À      0   `   �   Ð    !  @!  p!   !  Ð!   "  0"  `"  �"  À"  ð"  #  0#  P#  p#  �#  À#  à#  $  P$  €$   $  À$  à$   %   %  @%  `%  €%   %   &   &  @&   �  P'  p'     @              f    7 †   
 ˆ   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )ð ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(ð °  Ð  ð    P  €  °  à        @  `  �  Ð    0  P  p  €     À  ð     P  €  Ð     0  `  �  À  ð     @  p     Ð    @  p     Ð     0  p  �  À  ð     @   `   €       à    !  @!  p!  �!  °!  Ð"   `  ð"  p#     @              f    7 †   
 Œ   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )È ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(È °  Ð  à     @  P  �  À  ð  p     °     0  p  à  P  p  �  °  ð     0  @  P  `     À  à        @  p  �  °  Ð  ð     @  `  €     À  Ð  P     à     P  €   @  €          @              f    7 †   
 �   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )  ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(     0  P  �  °  Ð     0  `  €     À  à        0  `  €     À  à     0  �  À  à        @  p     ð  0  p  °  ð  0  p         p  À  @     €              f    7 †   
 ”   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )x ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(x    p  �  À  ð  0  `  �  °  Ð  ð       P  p  �  °  Ð  à    0  P  €  °  Ð     °  ð  0  p   0     @     €              f    7 †   
 ˜   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )P ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(P �  °  Ð     @  `  p     À  à  ð     `  �  °        @  p         @  Ð     €              f    7 †   
 œ   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  ) ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ( 0	  P	  p	  �	  °	   °  À	  P
     €              f    7 †   
 ¡   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )( ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(( €  Ð  ð    0  `  €     À  à   @       4 À         @  Ð  À     €              f    7 †   
 ¥   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  ) ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ( 0	  P	  p	  �	  °	   °  À	  P
     €              f    7 †   
 ª   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )( ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(( €  Ð  ð    0  `  €     À  à   @       4 À         @  Ð  À     €              f    7 †   
 ®   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )P ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(P 0  p  �  °  ð        0  `  €     °  à    @  p  à    @  P      à        €              f    7 †   
 ²   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )pÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(p '  0'  P'  �'   '  Ð'  à'  (  0(  P(  `(  €(  °(  à(   )  0)  �)  À)  Ð)   *   ,  À,  -   -  P-  `-  °-  À-  .   .  /  0/   /  °/  0  00  �0  °0   1  1   2   2  °2  À2   3  03  `3  p3  °3  Ð3  À4  05  P5  p5  �5  °5  Ð5  ð5  6  06  @8  `8  °8  À8   9  9  `9  p9  À9  Ð9  à:   ;  €;  �;  ð;  <  `<  €<  à<  ð<   >   >   >  °>  ?  0?  �?   ?  Ð?  à?  ð@  `A  €A   A  ÀA  àA   B   B  @B  `B  àD   E  PE  `E   E  °E   F  F  `F  pF  €G   G  H   H  �H   H   I   I  €I  �I   J  ÀJ  @K  PK   K  ÀK  0L  @L  pL  €L  �M   N   N  @N  `N  €N   N  ÀN  àN   O  ÀR  ðR  S  0S  PS  pS  ÀS  ðS   T  @T  pT   T  ÐT   U  0U  PU  pU  �U  °U  ÐU  àU   V  0V  PV  pV  �V  °V  ÐV  àV   W  0W  PW  pW  �W  °W  ÐW  àW   X  0X  `X  [  0[  €[  �[  Ð[  à[  0\  @\  �\   \  Ð]  ð]  `^  p^  à^  ð^  P_  p_  Ð_  à_  a  0a  °a  Àa  b  0b   b  °b  àb  ðb   d  �d  °d  Ðd  ðd  e  0e  Pe  pe  �e  4 %  À*  ð*  P6   7  €B  `C   O   P  €Y  ÐY  °e  °f     €              f    7 †   
 µ   `„ „       €  ð       |  ð       x  ð       t  ð       p  ð      
 l  ð      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ _  )( ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(( °  Ð     0  `  p     °  à  ð   0      Ð                 f    7 †   
 ¸   `„ „       €  ð       |  ð       x  ð       t  ð       p  ð      
 l  ð      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ _  )x ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(x À  Ð       @  P  €  �  À  Ð       0  @  p  �  °  À        0  @  p     °  À        0  @   0   Ð                    f    7 †   
 ½   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )@ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(@04  P4  �4  @5  €5  À5  Ð5  6  P6  €6   6  à6  `7   7  à7  8  @8  p8   8  à8  9  P9  �9  Ð9  :  P:  °:  ð:  0;  p;  °;  ð;   <  P<  €<   <  À<  à<   =   =  @=  p=  �=  °=  à=  >  @>  p>   >  À>  à>   ?   ?  @?  `?  €?  °?  Ð?  ð?  @  0@  P@  `@  €@   @  À@  à@   A  @A  `A  �A  ÐA   B  @B  `B  �B   B  ÀB  ÐB  àC    3   D  €D     @              f    7 †   
 Â   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(Ð/  P0   0  Ð0  ð0  1  @1  P1  �1  Ð1  2  P2  �2  À2  à2  ð2  3  03  `3  €3  °3  Ð3   4   4  P4  €4  °4  à4  5  @5  €5  °5  à5  6  @6  p6   6  à6  7  @7  p7  °7  ð7  08  `8  �8  À8  à8   9   9  @9  `9  €9   9  Ð9   :  @:  €:  °:  Ð:  ð:  ;  0;  P;  p;  �;  °;  <  0<  P<    .  `=  €=     @              f    7 †   
 Ç   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )ð ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(ð @+  `+   +  À+  à+  ,  P,  p,  °,  Ð,  ð,  -  0-  `-  �-  À-  ð-   .  @.  `.  �.  °.  Ð.  ð.  /  @/  `/  �/  À/  ð/   0  P0  €0  °0  ð0   1  P1  �1  À1  ð1   2  @2  p2   2  Ð2   3  03  `3  �3  °3  à3   4   4  @4  p4   4  à4  5  @5  `6   0*  €6   7     @              f    7 †   
 Ì   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )È ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(È À&  à&  '  0'  p'  �'  À'  Ð'  ð'  (   (  P(  p(  �(  °(  Ð(  ð(   )   )  @)  `)   )  Ð)   *  0*  €*  �*  °*  Ð*  ð*   +  @+  p+   +  Ð+   ,  0,  `,  �,  À,  ð,  0-  `-   -  à-   .  `.  �.  Ð.  /   À%  @0  `0     @              f    7 †   
 Ñ   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )  ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(  Ð#  ð#  $  P$  p$   $  Ð$   %  0%  P%  p%  �%  °%  Ð%  ð%   &  0&  P&  p&  �&  °&  Ð&  à&   '  0'  p'  À'   (  @(  p(  �(  À(   )  0)  `)  �)  À)  ð)   *  P+   0!  p+  ð+     €              f    7 †   
 Ö   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )x ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(x P  �  Ð    0  P  p  €  °  Ð  ð     0   @   p   �   °   Ð   ð    !  0!  �!  À!  ð!   "  "  �"  Ð"  #  P#   p  P$  p$     €              f    7 †   
 Û   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )P ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(P    À  à     @  `  p     À  à  ð  @  `  p  €  à    0  @  À   à  à  p     €              f    7 †   
 à   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  ) ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ( €     À  ð         0  �     €              f    7 †   
 å   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )( ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(( p  À  à        P  p  �  °  Ð      ð  0  4 Ð         P  à  à     €              f    7 †   
 ê   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )P ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(P `!   !  À!  à!   "  0"  P"  `"  �"  °"  Ð"  à"  #  @#  p#   #  $  @$  p$  €$   Ð  %  0%     €              f    7 †   
 ï   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )pÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(pð@  A  0A  €A   A  °A  ÀA   B   B  0B  @B  pB  �B  ÀB  àB   C  `C  �C  °C  àC  €F   F   G  G  @G  PG   G  °G  ðG   H  ðH  I  €I  �I  ðI  J   J  °J  àJ  ðJ  àK   L  €L  �L  àL  ðL  PM  pM  °M  ÀM   N  O  0O  PO  pO  �O  °O  ÐO  ðO  P   R  @R  �R   R  àR  ðR  @S  PS   S  °S  ÀT  àT  `U  pU  ÐU  ðU  @V  PV   V  ÀV  àW   X  pX  €X  ÀX  ÐX  @Y  PY  �Y  °Y  ÐZ  @[  `[  €[   [  À[  à[   \   \  @\  À^  à^  0_  @_  €_  �_  à_  ð_  @`  P`  `a  €a   b  b  pb  �b  àb  ðb  @c  `c  €d   d  e   e  `e  pe  Ðe  àe  0f  Pf  pg  àg   h   h  @h  `h  €h   h  Àh  àh  m  0m  Pm  pm  �m  °m   n  Pn  pn  �n   n  °n  Àn  àn  o  0o  Po  po  �o  °o  Ào  ào  p  0p  Pp  pp  �p  °p  Àp  àp  q  0q  Pq  pq  �q  °q  Àq  àq   r  @r  ðt  u  `u  pu  °u  Àu  v   v  pv  €v  °w  Ðw  Px  `x  Àx  àx  0y  @y  �y  °y  ðz  {  �{   {  |   |  P|  `|   |  À|   ~  p~  �~  °~  Ð~  ð~    0  P  p  4 ð>   D  ÐD  0P  àP  `\  @]   i  ài  `s  °s  �  �€     €              f    7 †   
 ó   `„ „       €  ð       |  ð       x  ð       t  ð       p  ð      
 l  ð      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ _  )( ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(( à     P  `  �     Ð  à        0   0                    f    7 †   
 ÷   `„ „       €  ð       |  ð       x  ð       t  ð       p  ð      
 l  ð      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ _  )x ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(x p   €   °   À   ð    !  0!  @!  p!  €!  À!  Ð!  à!  ð!  0"  P"  `"  p"   "  À"  à"  ð"  0#  P#  `#  p#   #  À#  à#  ð#   0   �$  0*                 f    7 †   
 û   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )@ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(@°!  Ð!  à!   "  @"  P"  °"  Ð"  à"  0#  `#   #   $  @$  P$  à%   &  @&  `&  €&  P(  `(  p(  À(  à(  )  @)  p)  �)   )  À)  à)  *  0*   *  À*  à*   +   +  @+  P+  �+  Ð+   ,  0,  `,   ,  Ð,   -  0-  `-  �-  À-  ð-  0.  `.  �.  À.  ð.   /  P/  p/  �/  À/  ð/  00  p0  °0  à0   1  `1  €1   1  À1  à1   2  2  02  P2  p3   @!  �3  4     @              f    7 †   
 ÿ   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(à       P  p  €  à      @   °   Ð   à   p!  �!  °!  #  �#  À#  à#   $  p%  °%  À%  Ð%  à%  ð%   &  0&  �&  °&  Ð&  ð&  '   '  `'  €'   '  À'  à'   (   (  P(  €(  À(  ð(  )  0)  P)  €)  °)  Ð)   *  `*   *  à*  +  @+  p+   +  Ð+  ,  @,  p,   ,  Ð,   -  0-  p-   -  Ð-   p  à.   /     @              f    7 †   
   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )ð ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(ð °  Ð        @  `  À  à        @  p  �  ð  @  `  €     °  Ð  ð      P   €   °   ð    !  P!  €!  °!  à!   "  0"  `"  �"  À"  ð"  0#  `#  €#   #  Ð#   $  0$  `$  �$  À$  ð$   %  `%  €%   %  Ð%  &  P&  p&  �&  °&  Ð&  ð'   €  (  �(     @              f    7 †   
   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )È ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(È à     `  €  °  Ð       0  P  `  �  °  Ð  ð    0  @  `  €     à    @  p     À  Ð     0  `  °  ð  0  p  °  ð     0   `   �   à   !  @!  p!  �!  °!  à!  "  @"   À  P#  p#     @              f    7 †   
   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )  ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(  p  �  °  ð    @  p     Ð  ð    0  P  p  �     Ð  ð    0  P  p  €  À    @  p  °  ð     @  p     à    @  p     Ð  ð   Ð     �      €              f    7 †   
   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )x ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(x °    0  `  €     °  ð    0  P  p  €  °  Ð  ð    0  @  €  Ð     @  P  €  À    P  €  À   Ð  °  Ð     €              f    7 †   
   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )P ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(P @  `  €  À  ð       P  p  �     Ð          P  °  ð    p   Ð  à  p     €              f    7 †   
   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  ) ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ( à	   
   
  @
  p
   P	  €
  à
     €              f    7 †   
   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )( ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ((    ð    0  P  €     À  à      `     `  4           €         €              f    7 †   
 !  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )P ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(P p     À  à     0  P  `  �  °  Ð  à    0  P  `     à     P      0  P     €              f    7 †   
 %  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )pÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(pÐ+  ð+  ,  `,  €,  �,   ,  à,   -  -   -  P-  €-  �-   -  À-   .  @.  p.  À.  p1  �1   2  2  @2  P2  �2   2  Ð2  à2  Ð3  ð3  `4  p4  Ð4  à4  @5  P5  °5  Ð5  À6  à6  `7  p7  Ð7  à7   8  08  p8  �8  €9  ð9  :  0:  P:  p:  �:  °:  Ð:  ð:  =  0=  �=   =  à=  ð=  @>  P>  €>  �>   ?  À?  @@  P@  °@  À@  A   A  €A   A  ÀB  àB  PC  `C  ÀC  ÐC   D  0D  pD  �D  °E   F  @F  `F  €F   F  ÀF  àF   G   G  °I  ÐI  0J  @J  €J  �J  àJ  ðJ   K  0K  @L  `L  àL  ðL  PM  `M  °M  ÀM   N  @N  `O  €O   P  P  pP  €P  ÀP  ÐP  Q  0Q  PR  ÀR  àR   S   S  @S  `S  €S   S  ÀS  `W  €W   W  ÀW  ðW   X  °X  ðX  0Y  `Y  €Y  �Y   Y  ÀY  ðY  Z  0Z  PZ  pZ  �Z   Z  ÀZ  ðZ  [  0[  P[  p[  �[   [  À[  ð[  \  0\  P\  p\  �\   \  À\   ]   ]  à_   `  ``  p`  °`  À`  a   a  Pa  `a  �b  °b  0c  @c   c  °c   d  d  pd  �d  Ðe  ðe  pf  €f  àf  ðf  0g  @g  €g   g  àh  Pi  pi  �i  °i  Ði  ði  j  0j  Pj  4 Ð)  €/  °/  ;  À;  @G   H  àS  ÀT  @^  �^  pj  pk     €              f    7 †   
 (  `„ „       €  ð       |  ð       x  ð       t  ð       p  ð      
 l  ð      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ _  )( ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(( €     ð     0  @  p  €  °  À   0   Ð  �                 f    7 †   
 +  `„ „       €  ð       |  ð       x  ð       t  ð       p  ð      
 l  ð      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ _  )x ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(x Ð  à       P  `  �     Ð  à     0  @  P  �  °  À  Ð        @  P  �  °  À  Ð    0  @  P   0   à  �                 f    7 †   
 0  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )@ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(@ð0  1   1  p1  �1   1  Ð1  ð1   2  P2  €2   2  3  03  p3  à4   5  @5  `5  €5  P7  `7  p7  Ð7   8  08  `8  €8  °8  À8  ð8  9  09  p9   9  À9  à9   :   :  @:  `:   :  Ð:  ;  @;  p;  °;  à;  <  @<  p<   <  Ð<   =  @=  p=   =  Ð=   >   >  P>  €>  °>  à>  ?  P?  �?  Ð?   @  @@  `@  €@   @  À@  à@  ð@  A  0A  PA  pB   @0  �B  C     @              f    7 †   
 5  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(à-   .  .  `.  €.  �.  À.  à.  ð.  @/  p/  �/   0   0  `0  Ð1  P2  €2   2  À2  04  `4  €4   4  °4  À4  Ð4   5  `5  €5   5  À5  à5  ð5  6  06  P6  �6  °6  Ð6  ð6  7  07  P7  �7  À7  ð7   8  P8  �8  À8  9  P9  €9  °9  Ð9  :  0:  `:  �:  Ð:   ;  @;  p;   ;  Ð;   <  @<  p<  �<   0-   =  À=     @              f    7 †   
 :  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )ð ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(ð P+  p+  °+  Ð+  ð+  0,  p,   ,  À,  à,   -  0-  `-  À-  à-   .  @.  P.  p.  �.  °.  À.  ð.   /  @/  €/  À/  ð/   0  `0  �0  À0   1  01  `1   1  Ð1  2  @2  €2  °2  à2  3   3  P3  p3   3  À3  à3   4  @4  `4  �4  À4  ð4   5  P5  p5  �5  °6    *  Ð6  P7     @              f    7 †   
 ?  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )È ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(È �'  Ð'  ð'  (  0(  `(  €(  °(  Ð(  ð(  )  0)  P)  p)  �)  °)  Ð)  à)   *   *  @*  p*  �*  À*   +  @+  P+  p+  �+  °+  à+   ,  `,   ,  à,   -  `-  �-  Ð-   .  0.  �.  Ð.  /  P/  €/   /  °/  à/   0   €&  1  01     @              f    7 †   
 D  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )  ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(   &  @&  `&  €&  À&  à&  '  @'  €'  �'  °'  Ð'  ð'  (   (  @(  p(  �(  °(  Ð(  ð(  )  @)   )  Ð)  ð)  *  0*  P*  €*  °*  ð*  @+  €+  À+   ,  @,  €,   ,  °-   €#  Ð-  P.     €              f    7 †   
 I  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )x ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(x à!  P"  p"  �"  À"  à"   #  0#  P#  `#  �#   #  À#  ð#  $   $  P$  `$  €$  °$  Ð$  à$  %   %  `%  À%  ð%  P&  `&  �&       ð'  (     €              f    7 †   
 N  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )P ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(P    @  `     À  à  ð     @  `  p     À  à  ð  @   €   °   à    !      À!  P"     €              f    7 †   
 S  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  ) ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ( à        @  p   P  €  à     €              f    7 †   
 W  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )( ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ((    ð    0  P  €     À  à      P     0#     €              f    7 †   
 [  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )P ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(P  .  Ð.  ð.  /  P/  `/  €/  �/  À/  à/   0  0  @0  `0  €0  �0  Ð0  1  P1  �1   0-  `2  €2     €              f    7 †   
 `  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )pÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(pP\  p\  °\   ]  ]   ]  @]  p]  �]   ]  À]  ð]  ^   ^  @^  p^  °^  ð^   _  `_   b   b  pb  €b  °b  Àb   c  0c  €c  �c  pd  �d  e   e  €e   e  àe  ðe  Pf  pf  `g  €g  ðg   h  @h  Ph  °h  Àh  i  0i   j  �j  °j  Ðj  ðj  k  0k  Pk  pk  �k  €m   m  àm  ðm  @n  Pn  °n  Àn  o   o  @p  `p  àp  ðp  Pq  pq  àq   r  @r  Pr  `s  €s   t  t  `t  pt  Ðt  ðt  0u  @u  Pv  Àv  àv   w   w  @w  `w  €w   w  Àw   z  @z  €z  �z  àz  ðz  P{  `{  °{  À{  à|   }  €}  �}  ð}  ~  €~   ~  à~  ð~   €   €   €  °€   �  �  p�  ��  Ð�  à�  ð‚  `ƒ  €ƒ   ƒ  Àƒ  àƒ   „   „  @„  `„  pˆ   ˆ  Ðˆ  ðˆ  ‰  0‰  €‰  °‰  à‰   Š   Š  0Š  @Š  `Š  �Š  °Š  ÐŠ  ðŠ  ‹  0‹  @‹  `‹  �‹  °‹  Ð‹  ð‹  Œ  0Œ  @Œ  `Œ  �Œ  °Œ  ÐŒ  ðŒ  �  0�  @�  `�   �  À�  P�  p�  °�  À�  ‘   ‘  €‘  �‘  à‘  ð‘  0“  P“  Ð“  à“  @”  `”  Ð”  ð”  0•  @•  p–  �–  —   —  p—  €—  à—   ˜  @˜  P˜  €™  ð™  š  0š  Pš  pš  �š  °š  Ðš  ðš  4 pZ   `  P`  °k  `l  àw  Àx  €„  `…  àŽ  0�  ›  œ     €              f    7 †   
 c  `„ „       €  ð       |  ð       x  ð       t  ð       p  ð      
 l  ð      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ _  )( ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(( °  À        p  €  °  À  ð      0     à                 f    7 †   
 f  `„ „       €  ð       |  ð       x  ð       t  ð       p  ð      
 l  ð      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ _  )x ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(x `-  p-   -  °-  à-  ð-   .  0.  `.  p.  °.  À.  Ð.  à.   /  @/  P/  `/   /  À/  Ð/  à/  0  00  P0  `0   0  À0  Ð0  à0   0   €1   7                 f    7 †   
 j  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )@ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(@P1  p1  €1  À1  à1  ð1  P2  �2  °2  03  P3  `3  à3   4   4  €5  À5  à5   6   6  à7  ð7   8  P8  �8  À8  ð8  9   9  P9  `9  �9  °9  Ð9  @:  `:  €:   :  À:  à:  ð:  0;  p;  °;  à;  <  P<  €<  °<  à<   =  0=  `=  �=  Ð=  >  @>  p>  °>  Ð>  ð>  ?  0?  `?  �?  Ð?  @  P@  �@  À@   A   A  @A  `A  €A   A  °A  ÐA  ðA  C   à0  0C  °C     @              f    7 †   
 n  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(0.  /  @/  `/  €/   /  à/   0  `0  �0  À0  1  @1  €1  °1  ð1   2  P2  €2  À2  à2  3  @3  p3   3  à3   4  P4  €4  À4   5  05  `5  �5  À5  ð5  06  `6  �6  À6  à6  7  @7  p7  �7  °7  Ð7   8  08  P8  p8  �8  À8  à8   9   9  @9  `9  €9   9  À9  ð9  :  0:  P:  p:  °:  Ð:  ð:  0;   @-  @<  `<     @              f    7 †   
 r  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )ð ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(ð Ð*  ð*  0+  P+  p+   +  Ð+  ð+  ,  0,  P,  p,  �,  -  @-  p-   -  Ð-  ð-  .  @.  p.  °.  à.  /  `/  �/  °/  à/  0  @0  P0  €0  À0  ð0   1  `1   1  Ð1   2  02  `2  �2   2  Ð2  ð2   3  @3  `3   3  À3  à3  4  P4  �4  °4  Ð4  ð4  5  06    )  P6  Ð6     @              f    7 †   
 v  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )È ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(È  '   '  P'  p'  °'  Ð'   (  (  0(  P(  `(  �(  °(  Ð(  ð(  )  0)  @)  `)  �)  Ð)  *  0*  p*  �*  °*  À*  à*   +   +  P+   +  à+   ,  `,   ,  à,  -  @-  p-  °-  ð-   .  p.  °.  à.   /   /  P/  `/   ð%  p0  �0     @              f    7 †   
 z  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )  ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(   %   %  @%  `%   %  À%  ð%   &  `&  p&  �&  °&  Ð&  ð&   '   '  P'  p'  �'  °'  Ð'  ð'   (  `(  �(  °(  Ð(  ð(   )  `)  �)  Ð)   *  `*   *  à*   +  `+  €+  �,   `"  °,  0-     €              f    7 †   
 ~  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )x ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(x p   Ð   ð   !  @!  `!  €!  °!  Ð!  à!  "   "  @"  p"  �"   "  Ð"  à"   #  0#  P#  `#   #  à#   $  @$  `$  À$   %  p%   �  €&   &     €              f    7 †   
 ‚  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )P ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(P P  p  �  à        0  `  €     °  à     P  p  À  à     0  à   à      �      €              f    7 †   
 ‡  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  ) ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(    À  à            0  À     €              f    7 †   
 Œ  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )( ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(( �  à        @  p  �  °  Ð  ð   @    P!  4 ð         p            €              f    7 †   
 �  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )P ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(P À*   +   +  @+  €+  �+  °+  À+  ð+  ,  0,  @,  p,   ,  Ð,   -  p-   -  Ð-  à-   0)  p.  �.     €              f    7 †   
 ”  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )pÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(p°T  ÐT  ðT  @U  `U  pU  €U  ÀU  àU  ðU   V  0V  PV  €V   V  àV   W  PW  pW   W  @Z  `Z  ÀZ  ÐZ   [  [  `[  p[  °[  À[  °\  Ð\  @]  P]  °]  Ð]  `^  p^   ^  °^   _  À_  @`  P`   `  °`  a  0a  pa  €a  `b  Ðb  ðb  c  0c  Pc  pc  �c  °c  Ðc  àe   f  Pf  `f   f  °f   g  g  `g  pg  €h   h   i  0i  �i  °i   j  j  `j  €j   k  Àk  0l  @l  €l  �l   m  m  Pm  pm  �n   o   o  @o  `o  €o   o  Ào  ào   p  €r   r  ðr   s  @s  Ps   s  °s   t  t   u  @u  Àu  Ðu  0v  Pv   v  °v   w   w  @x  `x  Ðx  àx   y  0y  �y   y  ðy  z  0{   {  À{  à{   |   |  @|  `|  €|   |  Ð€  ð€  �  0�  P�  p�  à�  ‚  0‚  P‚  `‚  p‚  €‚   ‚  Ð‚  ð‚  ƒ  0ƒ  Pƒ  pƒ  €ƒ   ƒ  Ðƒ  ðƒ  „  0„  P„  p„  €„   „  Ð„  ð„  …  0…  P…  p…  €…   …  à…   †  °ˆ  Ðˆ   ‰  0‰  p‰  €‰  Ð‰  à‰  0Š  @Š  p‹  �‹  Œ   Œ  €Œ   Œ  ðŒ   �  P�  p�  °Ž  ÐŽ  P�  `�  Ð�  à�  �   �  `�  €�  À‘  0’  P’  p’  �’  °’  Ð’  ð’  “  0“  4 °R  `X  �X  ðc   d   p   q  À|   }   ‡  p‡  P“  P”     €              f    7 †   
 —  `„ „       €  ð       |  ð       x  ð       t  ð       p  ð      
 l  ð      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ _  )( ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ((    À       P  `  �     Ð  à   0   ð  Ð                 f    7 †   
 š  `„ „       €  ð       |  ð       x  ð       t  ð       p  ð      
 l  ð      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ _  )x ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(x P)  `)  �)   )  Ð)  à)  *   *  P*  `*   *  °*  À*  Ð*  +  0+  @+  P+  €+   +  À+  Ð+  ,  0,  @,  P,  �,  °,  À,  Ð,   0   p-  @3                 f    7 †   
 ž  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )@ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(@�1  °1  À1   2   2  02  �2  Ð2  ð2  p3  �3   3   4  @4  `4  À5   6   6  @6  `6  08  @8  P8   8  À8  ð8   9  P9  p9  €9   9  À9  ð9  :  €:   :  À:  à:   ;   ;  0;  p;  °;  ð;   <  P<  �<  À<  ð<   =  @=  p=   =  Ð=  >  P>  €>  °>  ð>  ?  0?  P?  p?   ?  Ð?  @  P@  �@  Ð@   A  @A  `A  €A   A  ÀA  àA  ðA  B  0B  PC    1  pC  ðC     @              f    7 †   
 ¢  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(ð-  .   .  `.  €.  �.  ð.  0/  P/  Ð/  ð/   0  €0   0  À0   2   2  Ð2  ð2  3  €4  À4  Ð4  ð4   5  5   5  P5  Ð5  ð5  6  06  @6  P6  p6   6  À6  à6  7  07  `7  €7  À7  ð7  8  @8  p8  �8  °8  à8  9  `9   9  Ð9   :  0:  p:  �:  À:  ð:  @;  p;   ;  Ð;   <  @<  p<   <  À<  à<   €-  ð=  >     @              f    7 †   
 ¦  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )ð ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(ð  +  @+  `+  €+   +  À+  ð+   ,  P,  p,   ,  À,  Ð,  @-  €-   -  À-  Ð-   .  0.  `.  �.  À.  à.  /  P/  �/  Ð/   0  00  `0  �0  Ð0   1  01  `1  �1  Ð1  2  @2  p2   2  Ð2  ð2   3  @3  p3  �3  °3  ð3  4  04  `4  �4  À4  ð4   5  @5  `5  €6   ð)   6   7     @              f    7 †   
 ª  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )È ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(È P'  €'  °'  Ð'   (   (  @(  `(  p(  �(  °(  à(   )   )  @)  `)  €)  �)  °)  Ð)  ð)   *  @*  `*  À*  ð*   +   +  @+  `+  €+  À+   ,  @,  €,  À,   -  0-  p-   -  Ð-  0.  p.  °.  ð.   /  @/  P/  €/   /    &  °0  Ð0     @              f    7 †   
 ®  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )  ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(  @%  `%  €%   %  ð%  &  @&  p&  °&  À&  à&   '   '  @'  P'  p'   '  À'  à'   (   (  @(  p(  Ð(   )   )  @)  `)  €)  °)  Ð)  *  `*   *  à*   +  `+   +  Ð+  Ð,    "  ð,  p-     €              f    7 †   
 ²  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )x ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(x      !  @!  p!  �!  °!  À!   "   "  @"  `"  €"  �"  À"  à"   #   #  @#  P#  €#   #  À#  ð#  0$  `$  °$  `%   %  Ð%   &   À  °&  Ð&     €              f    7 †   
 ¶  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )P ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(P €     À    0  P  `  �  °  Ð  à    P  €     ð    0  `        0   À      €              f    7 †   
 º  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  ) ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ( P  p  �  °  Ð   Ð  à  p     €              f    7 †   
 ¾  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )( ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(( ð  @  `  €     Ð  ð    0  P      p  €!     €              f    7 †   
 Â  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )P ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(P P+  €+   +  À+   ,  ,  0,  @,  p,  �,  °,  À,  ð,  -  0-  @-  €-  À-   .  @.   à)  /  0/     €              f    7 †   
 Æ  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )pÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(p U   U  @U  pU   U  ÀU  ÐU  ðU   V  @V  PV  pV   V  ÐV  ðV   W  �W  °W  ÀW  X   Z  ÀZ  0[  @[  p[  €[  Ð[  à[  \   \   ]   ]  �]   ]  ð]   ^  p^  €^  à^   _  ð_  `  �`   `  à`  ð`  `a  €a  Àa  Ða  °b   c  @c  `c  €c   c  Àc  àc   d   d  0f  Pf   f  °f  ðf   g  `g  pg   g  °g  Ðh  ðh  `i  pi  Ði  ài  @j  Pj  °j  Ðj  ðk  l  �l   l  m   m  Pm  `m   m  Àm  àn  Po  po  �o  °o  Ðo  ðo  p  0p  Pp  Ðr  ðr  @s  Ps  �s   s   t  t  @t  Pt  pu  �u   v  v  pv  €v  àv  ðv  Pw  pw  �x  °x  0y  @y  °y  Ày  ðy   z  @z  `z  €{  ð{  |  0|  P|  p|  �|  °|  Ð|  ð|  �€  °€  Ð€  ð€  �  0�  €�  °�  à�  ‚  @‚  p‚   ‚  Ð‚  ƒ  @ƒ  `ƒ  €ƒ   ƒ  Àƒ  Ðƒ  ðƒ   „  @„  `„  €„   „  À„  Ð„  ð„   …  @…  `…  €…   …  À…  Ð…  ð…  0†  P†   ‰   ‰  p‰  €‰  À‰  Ð‰  0Š  @Š  pŠ  €Š  À‹  à‹  PŒ  `Œ  ÀŒ  ÐŒ  0�  @�   �  À�   �   �   �  °�   �  0�  `�  p�  °�  Ð�  ’  €’   ’  À’  à’   “   “  @“  `“  €“  4  S  °X  àX  @d  ðd  pp  Pq  }  ð}  p‡  À‡   “   ”     €              f    7 †   
 É  `„ „       €  ð       |  ð       x  ð       t  ð       p  ð      
 l  ð      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ _  )( ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(( à     P  `  �     Ð  à        0   0                   f    7 †   
 Ì  `„ „       €  ð       |  ð       x  ð       t  ð       p  ð      
 l  ð      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ _  )x ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(x ð)   *  0*  @*  p*  €*  °*  À*  ð*   +  @+  P+  `+  p+  °+  Ð+  à+  ð+   ,  @,  `,  p,  °,  Ð,  à,  ð,  0-  P-  `-  p-   0   .  à3                 f    7 †   
 Ð  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )@ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(@ .  `.  0/  P/  `/  €/  �/  À/  0  @0  €0  À0  @1   1  à1   2  `2  �2  Ð2  3  P3   3  Ð3   4  04  `4  �4  Ð4  5  @5  p5   5  Ð5   6   6  @6  `6  €6  °6  Ð6  ð6   7  @7  p7  �7  À7  ð7   8  P8  p8  �8  °8  Ð8  ð8  9  09  P9  €9   9  À9  à9   :   :  0:  P:  p:  €:   :  à:  0;  `;  p;  €;  °;  À;   <  @<  P<  p<  p=   ð,  �=  >     @              f    7 †   
 Ô  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(P*  p*  €*  À*  à*  ð*  P+  p+  €+  Ð+   ,  0,  �,  °,  -  €.   /  0/  P/  p/  à0  1   1  01  @1  P1  p1  �1   2   2  @2  `2  �2   2  À2  à2   3   3  @3  p3   3  Ð3  4  @4  `4  €4   4  À4  à4  5  P5  €5  °5  ð5   6  P6  €6  À6  ð6   7  `7  �7  Ð7   8  08  `8  �8  Ð8  9  @9   à)  P:  p:     @              f    7 †   
 Ø  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )ð ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(ð `'  €'  �'  Ð'  ð'   (  @(  p(  �(  ð(   )  @)  °)  Ð)  ð)  �+  ,  0,  P,  p,  à-   .   .  0.  @.  P.  `.  À.  à.   /   /  @/  `/  p/  �/  À/  à/   0   0  @0  `0  p0  �0  à0   1   1  @1  `1  €1   1  à1  �2  Ð2  3  P3  €3  °3  à3  4  ð4   ð&  5  �5     @              f    7 †   
 Ü  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )È ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(È P$  p$  €$  Ð$  ð$   %  0%  P%  `%  °%  ð%   &  �&  °&  À&  @(  °(  Ð(  ð(  )  P*  €*  �*   *  °*  À*  +  0+  P+  p+  €+   +  Ð+  ð+  ,  0,  P,  `,   ,  À,  à,   -   -  @-  Ð-  .  @.  p.   .  Ð.    #  à/   0     @              f    7 †   
 à  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )  ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(  °!  Ð!   "  @"  P"  p"  Ð"  à"  #  P#  p#  °#  $   $  �$  @&  À&   '   '  @'   )  )   )  p)  �)   )  °)  à)  *   *  0*  `*  �*  °*  À*  +  @+  p+   +  `,   €   €,   -     €              f    7 †   
 ä  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )x ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(x    À  P  €  °  Ð  à  ð      @   `   €       °   ð   !  0!  P!  p!  €!  °!  à!  "  @"  p"  °"  à"  0#  p#  °#   ð   %  @%     €              f    7 †   
 è  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )P ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(P p  �  °        @  P  €     À  Ð     @  p  �  à        P            °     €              f    7 †   
 ì  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  ) ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ( `  €     À  à   à  ð  €     €              f    7 †   
 ñ  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )( ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(( À    0  P  p     À  à         €  @  P   4           €          €              f    7 †   
 õ  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )P ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(P °)  à)   *   *  P*  p*  €*   *  Ð*  ð*   +   +  `+  p+  �+   +  à+  0,  p,  °,   @(  p-  �-     €              f    7 †   
 ù  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )pÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(pÐQ  ðQ  R  `R  €R  �R   R  àR   S  S   S  `S  €S  �S   S  ÐS  T  PT  €T  ÀT  `W  €W  °W  ÐW  X   X  pX  €X  ÐX  àX  ÐY  ðY  PZ  `Z  ÀZ  ÐZ  P[  p[  À[  Ð[  À\  à\  p]  €]  à]  ð]   ^  0^  p^  �^  €_  ð_  `  0`  P`  p`  �`  °`  Ð`  ð`  ðb  c  `c  pc  Àc  Ðc   d  0d  €d  �d   e  Àe  @f  Pf  °f  Ðf   g  @g   g  °g  Àh  àh  `i  pi  Ði  ài  @j  `j  �j   j  °k   l  @l  `l  €l   l  Àl  àl   m   m  �o  °o   p  p  `p  pp  Àp  Ðp   q  0q  @r  `r  àr  ðr  Ps  ps  Às  às  @t  Pt  `u  €u   v  v  pv  €v  àv   w  0w  @w  Px  Àx  àx   y   y  @y  `y  €y   y  Ày  €}  °}  Ð}  ð}  ~  0~  €~  °~  à~     0  `  �  À  ð  €  0€  P€  p€  �€   €  À€  ð€  �  0�  P�  p�  ��   �  À�  ð�  ‚  0‚  P‚  p‚  �‚   ‚  À‚   ƒ   ƒ  À…  à…  0†  @†  �†   †  ð†   ‡  P‡  `‡  �ˆ  °ˆ  0‰  @‰   ‰  À‰  Š  0Š  �Š   Š  Ð‹  ð‹  pŒ  €Œ  àŒ  ðŒ  P�  p�   �  °�  àŽ  P�  p�  ��  °�  Ð�  ð�  �  0�  P�  4 ÐO  €U  °U  a  Àa  @m   n  ày  Àz  @„  �„  p�  p‘     €              f    7 †   
 ü  `„ „       €  ð       |  ð       x  ð       t  ð       p  ð      
 l  ð      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ _  )( ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(( 0  P     °  à  ð     0  `  p   0   €  P                 f    7 †   
 ÿ  `„ „       €  ð       |  ð       x  ð       t  ð       p  ð      
 l  ð      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ _  )x ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(x @(  P(  €(  �(  À(  Ð(   )  )  @)  P)  �)   )  °)  À)  ð)  *  0*  @*  €*   *  °*  À*  ð*  +  0+  @+  p+  �+  °+  À+   0   `,  02                 f    7 †   
   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )@ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(@`  €     p  €         @   €   °   à   `!  °!  ð!  0"  p"   "  Ð"   #  0#  p#  °#  ð#  0$  p$  °$  %  P%  �%  Ð%  &  P&  €&  °&  Ð&  ð&   '  P'  €'   '  À'  ð'  (  0(  P(  p(  �(  °(  à(   )   )  @)  `)  €)   )  À)  ð)  *  0*  P*  p*  �*   *  À*  à*   +   +  @+  `+   +  À+   ,  0,  `,   ,  Ð,  ð,   -   -   .   `  @.  À.     @              f    7 †   
   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(@  À    P  p  �  °  Ð    `  �  À  ð     P  `  €     À  à     0   `   �   À   ð    !  P!  €!  À!   "  @"  €"  °"  à"  #  P#  €#  À#  ð#   $  P$  �$  À$  ð$   %  P%  €%   %  À%  à%   &   &  P&  p&   &  à&  '  0'  P'  p'  �'  °'  Ð'  ð'  (  0(  �(  °(  Ð(   0  à)   *     @              f    7 †   
   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )ð ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(ð       `  €     À  ð     0  p     Ð  ð     @  `  €     Ð  ð    0  P  €  °  à        P  €  °  ð     `  �  À  ð      @   `   €       Ð    !  0!  `!  �!  À!  ð!   "  `"  �"  °"  Ð"  #  P#  €#  °#  Ð#  ð$   °  %  �%     @              f    7 †   
   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )È ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(È `  �  °  Ð     @  p  �  °  À  Ð        @  `  €     °  Ð     0  `  �  À  ð    0  `  €     À  ð    0  `  �  À  ð     P  €  À  ð  0  `  �  Ð     @  €   `  à    !     @              f    7 †   
   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )  ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(  Ð        `  €     Ð     0  p  �  °  Ð  ð       @  p  �  °  Ð  ð       P  �  Ð  ð     P  €  °  ð     P  €     À     @      `  à     €              f    7 †   
   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )x ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(x �        P  p  �     Ð  ð    0  P  `     À  à          P  p     à     @  °  Ð    P  €   °  �  °     €              f    7 †   
   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )P ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(P    À  à    @  `  p     Ð  ð        p     À  P  p  €  �      0  @  Ð     €              f    7 †   
    `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  ) ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ( @
  `
  €
  °
  à
   À	  ð
  P     €              f    7 †   
 %  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )( ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(( Ð     @  `  €  °  Ð  ð    0   �  P  p  4 0         °  @  @     €              f    7 †   
 *  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  ) ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ( @
  `
  €
  °
  à
   À	  ð
  P     €              f    7 †   
 /  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )( ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(( Ð     @  `  €  °  Ð  ð    0   �  P  p  4 0         °  @  @     €              f    7 †   
 3  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )P ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(P °  ð    0  p  €     °  à        0  `  �  À  ð  `  �  À  Ð      `  €     €              f    7 †   
 7  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )pÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(pà'  (  0(  p(  €(  °(  À(  ð(  )  0)  @)  `)  �)  À)  à)  *  p*   *  °*  à*  €-   -  à-  ð-   .  0.   .  °.   /  /  ð/  0  €0  �0  ð0  1  p1  �1  à1  ð1  à2   3  �3   3   4  4  @4  P4  �4  °4   5  6  06  P6  p6  �6  °6  Ð6  ð6  7  9  09  `9  p9  °9  À9  0:  @:  �:   :  À;  à;  `<  p<  °<  À<  @=  `=  À=  Ð=  à>   ?  €?  �?  ð?   @  `@  €@  °@  À@  ÐA  @B  `B  €B   B  ÀB  àB   C   C  @C  °E  ÐE   F  F  PF  `F  ÐF  àF  0G  @G  `H  €H   I  I  PI  `I  àI   J  `J  pJ  €K   K   L  0L  �L   L   M   M  PM  `M  pN  àN   O   O  @O  `O  €O   O  ÀO  àO   S  ÐS  ðS  T  0T  PT  ÀT  ðT  U  @U  pU   U  ÀU  àU  V  0V  PV  pV  �V  °V  ÀV  àV  W  0W  PW  pW  �W  °W  ÀW  àW  X  0X  PX  pX  �X  °X  ÀX  àX   Y  @Y  à[   \  0\  @\  €\  �\   ]  ]  `]  p]  °^  Ð^  P_  `_   _  °_  0`  P`  °`  À`  ða  b  �b   b   c  c  pc  �c  Àc  Ðc   e  pe  �e  °e  Ðe  ðe  f  0f  Pf  pf  4 ð%   +  Ð+  07  à7  `C  @D   P  àP  `Z  °Z  �f  �g     €              f    7 †   
 :  `„ „       €  ð       |  ð       x  ð       t  ð       p  ð      
 l  ð      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ _  )( ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ((    À       P  `  �     Ð  à   0   ð  Ð                 f    7 †   
 =  `„ „       €  ð       |  ð       x  ð       t  ð       p  ð      
 l  ð      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ _  )x ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(x `  p     °  à  ð     0  `  p  °  À  Ð  à    0  P  `     À  Ð  à    0  P  `     À  Ð  à   0   p  @                 f    7 †   
 B  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )@ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(@°+  0,  P,  p,  �,  °,  ð,  P-  €-   .   .  P.  �.  Ð.   /  `/   /  à/   0  `0   0  à0   1  P1  €1  À1  ð1   2  P2  €2  °2  à2   3  P3  €3  °3  à3  4  04  P4  p4  °4  à4  5  @5  €5  °5  À5  à5   6   6  @6  `6  €6   6  À6  à6  7  07  P7  p7  �7  °7  À7  à7   8  8  08  P8  À8  ð8  9  09  P9  p9   9  À9  à9  ð9   ;   €*   ;   ;     @              f    7 †   
 G  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(À'  `(  �(  °(  Ð(  à(   )   )  p)  à)  *  @*  p*   *  Ð*   +   +  @+  `+  �+  À+  ð+  0,  `,  �,  À,  ð,   -  P-  �-  À-  ð-   .  P.  €.  °.  à.   /  P/  €/  °/  à/  0  @0  p0   0  Ð0   1  01  P1  p1  �1  °1  Ð1  ð1   2  P2  €2  °2  à2   3   3  @3  `3  €3   3  À3   4  @4  `4   °&  p5  �5     @              f    7 †   
 L  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )ð ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(ð  $   $  P$  €$   $  Ð$   %  %  @%  P%  p%  �%  À%   &  P&  p&  �&  À&  Ð&  ð&  '  @'  p'   '  Ð'   (   (  @(  `(  €(  °(  à(  )  P)  €)  °)  Ð)   *   *  P*  €*  °*  ð*   +  P+  €+  °+  à+  ,  @,  €,  °,  Ð,  ð,   -  `-  �-  À-  à-   /   À"   /   /     @              f    7 †   
 Q  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )È ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(È ð      @   p       À   à    !   !  @!  P!  €!   !  À!  à!   "   "  0"  P"  p"  �"  À"  à"  #  @#  p#   #  À#  à#   $   $  P$  €$  À$   %  @%  €%  °%  à%  &  @&  €&  °&  ð&  0'  p'  °'  à'  (  @(   Ð  P)  p)     @              f    7 †   
 V  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )  ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(     Ð  ð  0  P  p     Ð        @  `  €     À  Ð          @   `   €       Ð   0!  `!  €!   !  À!  à!  "  @"  €"  Ð"  #  P#  �#  Ð#  $  0$  @%     `%  à%     €              f    7 †   
 [  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )x ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(x Ð  P  €     À  à  ð     @  `  €     °  à        @  `  p     Ð  0  `  p  €        @  p  °   ð  à        €              f    7 †   
 `  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )P ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(P P  p  �  à          `  �     °        0  @  �  À  ð     à   à     �     €              f    7 †   
 e  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  ) ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ( À
  ð
    0  P   0
  `  À     €              f    7 †   
 j  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )( ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ((    ð    0  P  €     À  à      `       4 à         `  ð  à     €              f    7 †   
 o  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  ) ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ( À
  ð
    0  P   0
  `  À     €              f    7 †   
 t  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )( ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ((    ð    0  P  €     À  à      `       4 à         `  ð  à     €              f    7 †   
 y  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )P ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(P P  �  °  Ð       @  P  €     À  Ð     0  `  �     0  `  p   À           €              f    7 †   
 ~  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )pÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(p€0   0  À0  1  01  @1  P1  �1  °1  À1  Ð1   2   2  @2  P2  €2  À2   3  03  p3  6  06   6  °6  à6  ð6  07  @7  �7   7  €8   8   9  09  �9  °9  ð9   :  `:  €:  p;  �;  <   <  `<  €<   =  =  @=  P=  0>   >  À>  à>   ?   ?  @?  `?  €?   ?  �A  °A  ðA   B  @B  PB  ÐB  àB   C  0C  PD  pD  àD  ðD  `E  pE  àE   F  PF  `F  pG  �G  H   H  €H   H   I  I  @I  PI  `J  ÐJ  ðJ  K  0K  PK  pK  �K  °K  ÐK  0N  PN  �N   N  àN  ðN  pO  €O  ÀO  ÐO  ðP  Q  €Q  �Q   R  R  €R   R  ðR   S  T  0T  °T  ÀT   U  @U   U  °U  àU  ðU   W  pW  �W  °W  ÐW  ðW  X  0X  PX  pX  \  0\  P\  p\  �\  °\   ]  P]  €]  À]  ð]   ^  P^  €^   ^  °^  à^   _   _  @_  `_  €_  �_  °_  à_   `   `  @`  ``  €`  �`  °`  à`   a   a  @a  `a  €a  �a  °a  pd  �d  Ðd  àd   e  0e   e  °e   f  f  @g  `g  àg  ðg  Ph  `h  Ðh  ðh  @i  Pi  €j   j   k  0k  �k  °k  l   l  Pl  `l  �m   n   n  @n  `n  €n   n  Àn  àn   o  4 €.  04  `4  À?  p@  ðK  ÐL  �X  pY  ðb  @c   o   p     €              f    7 †   
 ‚  `„ „       €  ð       |  ð       x  ð       t  ð       p  ð      
 l  ð      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ _  )( ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(( à     P  `  �     Ð  à        0   0                   f    7 †   
 †  `„ „       €  ð       |  ð       x  ð       t  ð       p  ð      
 l  ð      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ _  )x ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(x    °  à  ð     0  `  p     °  ð          `  €  �     Ð  ð       `  €  �     à           0   À  �!                 f    7 †   
 ‹  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )@ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(@À)  à)   *   *  P*  À*   +  ,  P,  €,  °,  à,   -  0-   -  à-  .  @.  €.  À.  ð.  0/  p/  °/  à/  0  P0  �0  À0  ð0   1  P1  €1  °1  à1  2  02  P2  p2  �2  °2  à2   3   3  P3  €3  °3  Ð3  ð3  4  04  P4  p4  �4  °4  Ð4  ð4   5  @5  `5  €5   5  À5  Ð5  ð5  6  06  P6  p6  �6   7   7  @7  p7  €7  °7  Ð7  ð7   8  9   �(  09  °9     @              f    7 †   
 �  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(À%  `&  €&  À&  à&  ð&   '   '  p'  à'  (  @(  p(   (  Ð(  ð(  )  0)  `)  �)  À)  ð)  0*  `*  �*  À*  ð*   +  P+  �+  À+  ð+   ,  P,  €,  °,  à,   -  P-  €-  °-  à-  .  @.  p.   .  À.  ð.   /  @/  `/  €/   /  À/  à/  0  @0  p0   0  Ð0  ð0  1  01  P1  p1  �1  °1  2  02  P2    $  `3  €3     @              f    7 †   
 •  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )ð ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(ð À!  à!   "  0"  P"  €"  °"  Ð"   #   #  @#  P#  p#   $   $  @$  `$  €$  °$  à$  %  @%  `%  �%  À%   &  0&  `&  �&  À&  ð&   '  P'  €'  °'  à'  (  P(  €(  °(  Ð(   )  0)  `)  �)  À)  ð)   *  P*  �*  °*  Ð*   +  @+  €+   +  À+  à+   ,   -   �   @-  À-     @              f    7 †   
 š  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )È ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(È Ð  ð     @  `  °  à        @  P  p  �  °  Ð  ð         @   `   €   °   Ð   ð   !  @!  €!  À!  ð!   "  P"  �"  À"  ð"   #  P#  €#  °#  à#  $  @$  €$  °$  à$  %  P%  €%  °%  ð%  0&   à  p'  �'     @              f    7 †   
 Ÿ  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )  ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(  À  à     0  P  p     Ð        @  `  €     À  Ð        @  `  €     À     P  €     À  à     @   p   Ð   !  P!  �!  Ð!  "  0"  @#      `#  à#     €              f    7 †   
 ¤  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )x ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(x À     @  €     À  à  ð     @  `  €     °  à        @  `  p     ð    @  `  €  Ð    P  �   à  À  à     €              f    7 †   
 ©  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )P ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(P P  p  �  Ð  ð       `  €     °  à    @  `  �  °  à    Ð   à  ð  €     €              f    7 †   
 ®  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  ) ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ( ð    0  P  p   `  €  à     €              f    7 †   
 ³  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )( ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(( °        @  `  �  °  Ð  ð     p  0     4 ð         p     ð     €              f    7 †   
 ¸  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  ) ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ( ð    0  P  p   `  €  à     €              f    7 †   
 ½  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )( ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(( °        @  `  �  °  Ð  ð     p  0     4 ð         p     ð     €              f    7 †   
 Â  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )P ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(P    @  `  €  À  Ð  ð     0  P  p  €  °  à    @  °  à        p  °  Ð     €              f    7 †   
 Ç  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )pÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(p°(  Ð(  ð(  @)  `)  p)  €)  À)  à)  ð)   *  0*  P*  p*  �*  à*   +  P+  p+   +  0.  P.  °.  À.  ð.   /  `/  p/  °/  À/  °0  Ð0  P1  `1  À1  à1  02  @2  �2  °2   3  À3  @4  P4   4  À4  05  @5  p5  €5  `6  Ð6  ð6  7  07  P7  p7  �7  °7  Ð7  Ð9  ð9  0:  @:  €:  �:   ;  ;  `;  p;  €<   <   =  0=  �=   =  ð=   >  `>  €>   ?  À?  @@  P@  À@  Ð@   A  A  PA  pA  �B   C   C  @C  `C  €C   C  ÀC  àC   D  pF  �F  ÐF  àF   G  0G   G  °G   H  H   I  @I  °I  ÀI   J  0J  �J   J   K   K  @L  `L  àL  ðL  @M  PM  �M   M  ðM  N  0O   O  ÀO  àO   P   P  @P  `P  €P   P  @T  `T  €T   T  ÀT  àT  PU  �U  ÀU  ðU   V  PV  pV   V  ÐV  ðV  W  0W  PW  pW  €W   W  ÐW  ðW  X  0X  PX  pX  €X   X  ÐX  ðX  Y  0Y  PY  pY  €Y   Y  àY   Z   \  À\   ]  ]  P]  `]  Ð]  à]  0^  @^  p_  �_   `  `  p`  €`  à`  ð`  Pa  pa  °b  Ðb  Pc  `c  °c  Àc   d  d  `d  €d  Àe  0f  Pf  pf  �f  °f  Ðf  ðf  g  0g  4 °&  `,  �,  ð7   8   D   E  ÀP   Q   [  p[  Pg  Ph     €              f    7 †   
 Ë  `„ „       €  ð       |  ð       x  ð       t  ð       p  ð      
 l  ð      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ _  )( ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ((      P  p  À  Ð       @  P   0   `  0                 f    7 †   
 Ï  `„ „       €  ð       |  ð       x  ð       t  ð       p  ð      
 l  ð      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ _  )x ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(x 0  @  p  €  °  À  ð     0  @  €  �     °  ð       0  `  €     °  ð       0  p  �     °   0   @                   f    7 †   
 Ó  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )@ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(@06  `6  À6  Ð6  ð6  €7   7  P8  p8  À8   9  @9  p9  °9  @:  €:  À:  ð:  0;  p;   ;  à;   <  `<   <  Ð<  =  @=  p=   =  Ð=   >   >  P>  p>  �>  °>  Ð>  ð>  ?  0?  `?  €?   ?  Ð?   @  @@  p@   @  À@  à@   A   A  @A  `A  €A   A  ÐA  ðA  B  0B  PB  pB  €B   B  ÀB  ÐB  ðB  C  €C  °C  ÐC  ðC  D  0D  `D  €D   D  °D  ÀE    5  àE  `F     @              f    7 †   
 ×  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(°2   3  à3   4  4  04  P4  €4  à4  5  @5  p5   5  à5  ð5  6  06  P6  €6  °6  à6  7  @7  €7  °7  à7  8  08  `8   8  à8   9  P9  €9  °9  à9  0:  p:  °:  ð:  0;  P;  `;  €;   ;  À;  à;   <   <  @<  `<  €<   <  À<  ð<  0=  p=   =  Ð=  ð=  >  0>  P>  p>  �>  °>  Ð>  0?  P?  p?    1  €@   @     @              f    7 †   
 Û  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )ð ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(ð ð.  /  p/  �/  À/  ð/   0  @0  p0   0  °0  à0  1  p1  �1  °1  Ð1  ð1  2  02  P2  €2  °2  à2  3  @3  `3  €3   3  À3  ð3   4  `4  �4  Ð4  5  @5  €5  À5  à5   6   6  P6  €6  À6  ð6   7  P7  p7  �7  À7  à7   8   8  P8  €8  °8  ð8   9  @:   Ð-  `:  à:     @              f    7 †   
 ß  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )È ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(È `+   +  À+  à+   ,  @,  p,  °,  Ð,  ð,  -  0-  P-  p-  �-  °-  Ð-  à-   .   .  @.  p.  �.  À.   /  @/  P/  p/  �/  °/  Ð/  0  P0  �0  Ð0  1  P1  €1  °1  à1  2  p2  °2  ð2  03  `3  �3   3  Ð3  ð3   P*   5   5     @              f    7 †   
 ã  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )  ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(  P)  p)  �)  °)   *   *  P*  �*  À*  Ð*  ð*  +  0+  @+  `+  �+  °+  Ð+  ð+  ,  0,  @,  `,  ð,   -  @-  `-  €-  °-  ð-   .  `.  °.  ð.  0/  p/  °/  ð/   0   1   Ð&  @1  À1     €              f    7 †   
 ç  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )x ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(x  $   %  P%  €%   %  Ð%  ð%   &  P&  p&  �&  °&  À&  ð&  '  0'  P'  p'  €'  °'  ð'  (  @(  `(   (  °(   )  @)  p)  À)   à"  ð*  +     €              f    7 †   
 ë  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )P ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(P p   �   °   !  0!  @!  P!   !  À!  Ð!  à!  0"  `"  €"  �"  Ð"   #  @#  `#  0$     P$  à$     €              f    7 †   
 ï  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  ) ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ( P  p  �  °  Ð   °  à  @     €              f    7 †   
 ó  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )( ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(( P     À  à     0  P  p  �  °      Ð  à#     €              f    7 †   
 ø  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )P ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(P P+  €+   +  À+  à+  ,   ,  @,  p,  �,   ,  Ð,   -   -  0-  `-  �-  °-  Ð-   .    *  `/  €/     €              f    7 †   
 ý  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )pÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(pÀS  àS   T  PT  pT  €T  �T  ÐT  ðT   U  U  @U  `U  �U  °U  ðU  0V  `V  €V  °V  PY  pY  ÐY  àY  Z   Z  pZ  €Z  ÀZ  ÐZ  À[  à[  `\  €\  ð\   ]  `]  p]  °]  À]  °^  Ð^  P_  `_   _  °_   `  @`  €`  �`  pa  àa   b   b  @b  `b  €b   b  Àb  àb  ðd  e  `e  pe  °e  Àe  f   f  pf  €f  �g  °g   h  0h   h  °h   i  i  pi  �i  °j  Ðj  @k  Pk  °k  Àk  l   l  `l  €l   m  n  0n  Pn  pn  �n  °n  Ðn  ðn  o  �q  °q   r  r  Pr  `r  °r  Àr  s   s  0t  Pt  Àt  Ðt  @u  Pu   u  °u  v  0v  Pw  pw  ðw   x  Px  `x   x  °x   y   y  @z  °z  Ðz  ðz  {  0{  P{  p{  �{  °{  à   €   €  @€  `€  €€  ð€   �  @�  `�  p�  €�  ��  °�  à�   ‚   ‚  @‚  `‚  €‚  �‚  °‚  à‚   ƒ   ƒ  @ƒ  `ƒ  €ƒ  �ƒ  °ƒ  àƒ   „   „  @„  `„  €„  �„  °„  ð„  …  À‡  à‡  0ˆ  @ˆ  €ˆ  �ˆ  àˆ  ðˆ  @‰  P‰  €Š   Š  ‹   ‹  �‹   ‹  ð‹   Œ  `Œ  €Œ  À�  à�  `Ž  pŽ  ÐŽ  àŽ   �  0�  p�  ��  Ð�  @‘  `‘  €‘   ‘  À‘  à‘   ’   ’  @’  4 ÀQ  pW   W   c  °c  0o  p  Ð{  °|  0†  €†  `’  `“     €              f    7 †   
    `„ „       €  ð       |  ð       x  ð       t  ð       p  ð      
 l  ð      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ _  )( ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(( €  �  Ð  ð  @  P  €  �  À  Ð   0   à  °                 f    7 †   
   `„ „       €  ð       |  ð       x  ð       t  ð       p  ð      
 l  ð      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ _  )x ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(x  .  .  @.  P.  €.  �.  À.  Ð.   /  /  P/  `/  p/  €/  °/  Ð/  ð/   0  @0  `0  p0  €0  °0  Ð0  ð0   1  @1  `1  p1  €1   0   2  à7                 f    7 †   
   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )@ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(@ 6  @6  P6   7  `7   8   8  `8   8  Ð8   9  09  °9  ð9  0:  p:  °:  à:  ;  @;  p;  °;  à;   <   <  @<  p<  �<  °<  à<  =  @=  p=   =  Ð=  >  @>  p>   >  Ð>  ð>   ?  @?  `?  €?  °?  à?   @  0@  `@  �@  °@  Ð@  ð@  A  0A  PA  €A   A  ÀA  àA   B   B  0B  PB  pB  �B  °B  ÐB  ðB  `C  €C   C  ÐC  àC  D  0D  PD  `D  pE   ð4  �E  F     @              f    7 †   
   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ( 1   1  Ð1  2  02  @2  `2  €2  À2  3  @3  p3   3  Ð3  4  04  `4  €4  °4  à4  5  @5  €5  °5  à5  6  @6  p6   6  Ð6   7   7  P7  €7  °7  à7  8  P8  €8  °8  à8  9  @9  p9   9  Ð9  ð9   :  P:  p:  �:  °:  Ð:  ð:  ;  @;  p;   ;  Ð;   <   <  @<  `<  €<   <  À<  à<  @=  `=  €=   Ð/  �>  °>     @              f    7 †   
   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )ð ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(ð €+   +   ,   ,  P,  €,   ,  À,  à,   -  -  0-  P-   -  À-  à-   .   .  @.  €.  À.  ð.   /  P/  €/  À/  ð/   0  P0  €0  °0  à0   1  P1  €1  °1  à1   2  P2  €2  °2  à2  3  @3  p3   3  Ð3  ð3  4  P4  p4  �4  À4   5  @5  `5  €5   5  À5  à6   P*   7  €7     @              f    7 †   
   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )È ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(È à%   &   &  `&  €&  °&  Ð&   '   '  @'  €'  Ð'   (   (  @(  p(   (  °(  Ð(  ð(  )  P)  €)  °)  Ð)  ð)  *   *  P*  €*  °*  ð*   +  P+  €+  °+  à+  ,  @,  p,   ,  à,  -  P-  €-  °-  à-  .  @.  p.    %  �/  °/     @              f    7 †   
   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )  ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(   "  @"  `"  €"  °"  Ð"   #  0#  p#  €#   #  À#  à#   $  $  0$  `$  €$   $  À$  à$   %  %  P%  €%  À%  ð%   &  @&  p&   &  à&  P'  €'  °'  à'  (  @(  p(   )   €  À)  @*     €              f    7 †   
   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )x ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(x   `  €     Ð  ð    @  `  p     °  Ð        0  `  p  �  À  à  ð  0  `  �  °  Ð      `   À      ð!  "     €              f    7 †   
   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )P ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(P Ð  ð    P  p  �     à        0  `  €     °    P  €     P   `  p        €              f    7 †   
 #  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  ) ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ( P  p  �  °  Ð   Ð
  à  p     €              f    7 †   
 (  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )( ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ((    p  �  °  Ð        @  `  €   à        4 `         à  p  `     €              f    7 †   
 ,  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )P ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(P €     À  à    0  @  `  �  °  À  à     P  €  À  @  p     °      0  P     €              f    7 †   
 0  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )pÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(p`0  €0   0  ð0  1   1  01  p1  �1   1  °1  à1   2  02  P2  �2  Ð2   3   3  P3  ð5  6  p6  €6  °6  À6  7   7  `7  p7  `8  €8   9   9  �9   9   :  :  P:  `:  P;  p;  ð;   <  @<  P<  À<  à<   =  0=  >  €>   >  À>  à>   ?   ?  @?  `?  €?  �A  °A   B  B  PB  `B  °B  ÀB  C   C  0D  PD  ÀD  ÐD  0E  @E   E  °E  F  0F  PG  pG  ðG   H  pH  €H  °H  ÀH   I   I  @J  °J  ÐJ  ðJ  K  0K  PK  pK  �K  °K  0N  PN   N  °N  ðN   O  PO  `O  °O  ÀO  ÐP  ðP  `Q  pQ  àQ  ðQ  @R  PR  °R  ÐR  ðS  T  �T   T  ðT   U  @U  PU   U  ÀU  àV  PW  pW  �W  °W  ÐW  ðW  X  0X  PX  €\   \  À\  à\   ]   ]  �]  À]  à]   ^  ^   ^  0^  P^  €^   ^  À^  à^   _   _  0_  P_  €_   _  À_  à_   `   `  0`  P`  €`   `  À`  à`   a   a  0a  Pa  �a  °a  `d  €d  Ðd  àd   e  0e  €e  �e  àe  ðe   g  @g  °g  Àg  0h  @h  �h   h   i   i  `j  €j   k  k  `k  pk  °k  Àk  l  0l  pm  àm   n   n  @n  `n  €n   n  Àn  àn  4 `.  4  @4   ?  P@  ÐK  °L  pX  PY  Ðb   c   o   p     €              f    7 †   
 3  `„ „       €  ð       |  ð       x  ð       t  ð       p  ð      
 l  ð      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ _  )( ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(( 0  @  €     ð     0  @  p  €   0   �  `                 f    7 †   
 6  `„ „       €  ð       |  ð       x  ð       t  ð       p  ð      
 l  ð      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ _  )x ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(x      P  `  �     Ð  à       `  p  €  �  À  à       @  `  €  �  Ð  ð       P  p  €  �   0      ð!                 f    7 †   
 :  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )@ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(@`/  €/  �/  Ð/  ð/   0  `0  €0  �0  à0  1  P1  °1  Ð1   2  �3  Ð3  ð3  4  04   6  6   6  p6  �6  À6  ð6   7  @7  `7  €7  �7  À7  à7  P8  p8  �8  °8  Ð8  ð8   9  @9  €9  °9  à9  :  P:  €:  °:  à:  ;  P;  �;  À;   <  0<  `<  �<  Ð<  ð<  =  0=  `=  �=  À=   >  @>  €>  °>  ð>  ?  0?  P?  p?  �?   ?  À?  à?   @   A   ð.  @A  ÀA     @              f    7 †   
 >  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(`,  à,   -  @-  €-  À-  à-  .  P.  €.   .  À.  ð.   /  P/  �/  Ð/   0  @0  p0   0  Ð0  1  01  P1  €1  °1  à1  2  P2  �2  Ð2   3  03  `3  �3  Ð3   4  04  `4  �4  À4  ð4   5  P5  p5  �5  °5  Ð5  ð5  6  06  P6  p6   6  Ð6  ð6  7  07  P7  �7  °7  Ð7  ð7  8  08  `8   8  À8  à8   +  ð9  :     @              f    7 †   
 B  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )ð ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(ð ð(  )  0)  P)  p)  �)  À)  ð)   *  @*  p*  �*  °*  ð*  @+  €+   +  °+  Ð+  ð+   ,  0,  `,  �,  À,   -   -  @-  `-  €-  °-  ð-   .  `.  �.  À.  ð.   /  @/  `/  €/   /  Ð/   0  @0  p0   0  Ð0   1  01  p1  �1  °1  Ð1  2  P2  €2  °2  Ð2  ð3    '  4  �4     @              f    7 †   
 F  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )È ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(È ð$  %  @%  `%  €%  °%  Ð%   &  &  0&  P&  €&   &  À&  à&   '   '  0'  P'  p'  �'  À'  à'   (  0(  p(   (  Ð(  )  @)  p)  °)  à)  *  @*  p*   *  Ð*   +  0+  `+   +  Ð+  ,  P,  �,  À,  ð,   -  0-   Ð#  @.  `.     @              f    7 †   
 J  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )  ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(  Ð"   #  0#  `#  €#   #  Ð#   $  0$  P$  p$  �$  °$  Ð$  ð$   %  0%  P%  p%  �%  °%  Ð%  ð%   &  `&  €&  °&  à&  '  0'  P'  `'   '  Ð'   (  0(  p(   (  Ð(   *   ð   *   *     €              f    7 †   
 N  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )x ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(x €  à        P  p  �  À  à  ð     0  P  €     °  à  ð     @   `   p   °   ð   0!  P!  p!  Ð!  "  p"      �#  °#     €              f    7 †   
 R  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )P ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(P �  °  Ð     0  P  `  €  °  Ð  à     P  `  p  �  ð  0  `  À     0  À     €              f    7 †   
 V  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  ) ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ( Ð  ð    0  P   P  `  ð     €              f    7 †   
 Z  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )( ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(( P     À  à     0  P  p  �  °      Ð  À     €              f    7 †   
 ^  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )P ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(P $  @$  `$  €$  °$  Ð$  à$  %  @%  `%  p%   %   &  @&  `&  p&  À&  ð&   '  @'   �"  Ð'  ð'     €              f    7 †   
 b  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )pÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(pPG  pG  �G  àG   H  H   H  `H  €H  �H   H  ÐH  ðH   I  @I  €I  ÀI  ðI  J  @J  àL   M  `M  pM   M  °M   N  N  PN  `N  PO  pO  ðO  P  €P  �P  ðP   Q  @Q  PQ  @R  `R  àR  ðR  0S  @S  °S  ÐS  T   T   U  pU  �U  °U  ÐU  ðU  V  0V  PV  pV  €X   X  ðX   Y  @Y  PY   Y  °Y   Z  Z   [  @[  °[  À[  0\  @\  �\   \   ]   ]  @^  `^  Ð^  à^  @_  P_   _  °_  ð_  `  0a   a  Àa  àa   b   b  @b  `b  €b   b   e  @e  �e   e  àe  ðe  @f  Pf   f  °f  Àg  àg  Ph  `h  Ðh  àh  0i  @i   i  Ài  àj   k  €k  �k  àk  ðk  0l  @l  �l  °l  Ðm  @n  `n  €n   n  Àn  àn   o   o  @o  ps  �s  °s  Ðs  ðs  t  €t  °t  Ðt  ðt   u  u   u  @u  pu  �u  °u  Ðu  ðu  v   v  @v  pv  �v  °v  Ðv  ðv  w   w  @w  pw  �w  °w  Ðw  ðw  x   x  @x  €x   x  P{  p{  À{  Ð{  |   |  p|  €|  Ð|  à|  ~  0~   ~  °~     0  €  �  ð  €  P�  p�  ð�   ‚  `‚  p‚  °‚  À‚   ƒ   ƒ  `„  Ð„  ð„  …  0…  P…  p…  �…  °…  Ð…  4 PE   K  0K  �V  @W  Àb   c  `o  @p  Ày  z  ð…  ð†     €              f    7 †   
 e  `„ „       €  ð       |  ð       x  ð       t  ð       p  ð      
 l  ð      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ _  )( ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(( �  °       @  P  €  �  À  Ð   0   à  °                 f    7 †   
 h  `„ „       €  ð       |  ð       x  ð       t  ð       p  ð      
 l  ð      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ _  )x ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(x �#   #  Ð#  à#  $   $  P$  `$  �$   $  à$  ð$   %  %  @%  `%  €%  �%  À%  ð%   &  &  0&  `&  €&  �&  Ð&  ð&   '  '   0    '  0.                 f    7 †   
 l  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )@ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(@À  P  `  �  ð       Ð      0   p   !  P!  �!  À!   "  0"  `"  �"  À"   #  @#  p#  °#  à#   $  P$  �$  À$  ð$   %  P%  €%  °%  à%  &  @&  p&  �&  °&  Ð&   '   '  @'  p'   '  À'  Ð'  ð'  (  0(  P(  p(  �(  °(  Ð(  ð(   )  @)  `)  €)   )  À)  Ð)  ð)  *   *  @*  `*  Ð*   +   +  @+  `+  €+  °+  Ð+  ð+   ,  -   �  0-  °-     @              f    7 †   
 p  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(�    0  �  °  Ð  ð  0  `     Ð     0  `  �  °  Ð     0  P  p  �  À  ð      P   €   °   à    !  P!  �!  Ð!   "  0"  `"  °"  à"   #  P#  €#  °#  à#  $  @$  p$   $  Ð$  ð$  %  0%  P%  p%   %  À%  ð%  0&  `&  €&   &  À&  à&   '   '  @'  `'  €'  à'   (   (   p  0)  P)     @              f    7 †   
 t  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )ð ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(ð ð    @  p     Ð  ð    0  P  p     Ð  @  `  €     À  Ð  ð     P  €  °  ð  0  `  �  À  ð     0  `  �  À  ð      `       Ð    !  0!  `!  €!  °!  Ð!   "   "  @"  €"   "  À"  ð"  0#  p#  �#  °#  Ð#  ð#  %      0%  °%     @              f    7 †   
 x  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )È ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(È       0  p  �       P  `  Ð  ð     P  p  Ð  0     À  à     @  p  €  �     °  à        @  `  p  °  Ð  ð    0  @  �  °  Ð  ð             à    !  `!  �!  À!   �  Ð"  ð"     @              f    7 †   
 |  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )  ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(     @  p  �  °  Ð     0  `  €     À  à       @  €     À  à  ð    0  À  à        @  p  �  À  `  �  Ð    @  p  °  ð        0  °     €              f    7 †   
 €  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )x ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(x    €     Ð  ð       `  €     À  à  ð  P  �  °  Ð  ð     0  P  p     Ð  ð     à     `  �   P  @  `     €              f    7 †   
 „  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )P ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(P Ð  ð    @  p  �     Ð        0  P     Ð     �  °  À  Ð  `   `  €       €              f    7 †   
 ˆ  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  ) ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ( 
  0
  P
  p
  �
   �	   
  0     €              f    7 †   
 �  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )( ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(( �  à        @  p  �  °  Ð  ð   P       4 Ð         P  à  Ð     €              f    7 †   
 ‘  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )P ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(P €  À  à     @  P  p  €  °  Ð  ð     0  `  �  À  0  `  �      ð  0  P     €              f    7 †   
 •  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )pÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(p`/  �/  À/   0   0  00  P0  �0   0  À0  Ð0   1  01  `1  �1  ð1  2  02  @2  p2  5  05  �5   5  Ð5  à5  @6  P6  �6   6  €7   7   8  08  �8  °8   9  @9  €9  �9  p:  �:  ;   ;  p;  �;   <  <  @<  P<  0=   =  À=  à=   >   >  @>  `>  €>   >   @  À@  A   A  `A  pA  ÐA  àA   B  0B  PC  pC  àC  ðC  PD  `D  ÀD  ÐD  0E  PE  pF  �F  G   G  �G   G  ÐG  àG   H  @H  `I  ÐI  ðI  J  0J  PJ  pJ  �J  °J  ÐJ  PM  pM  ÀM  ÐM  N   N  pN  €N  ÐN  àN  ðO  P  €P  �P  ðP   Q  `Q  pQ  ÐQ  ðQ  S  0S  °S  ÀS  T   T  `T  pT  ÀT  àT   V  pV  �V  °V  ÐV  ðV  W  0W  PW  pW  @[  `[  €[   [  À[  à[  P\  €\  °\  à\  ]  0]  P]  p]   ]  À]  à]   ^   ^  @^  P^  p^   ^  À^  à^   _   _  @_  P_  p_   _  À_  à_   `   `  @`  P`  p`  °`  Ð`  €c   c  ðc   d  @d  Pd   d  °d   e  e  @f  `f  Ðf  àf  @g  Pg  °g  Àg   h  @h  €i   i   j  0j  €j  �j  Ðj  àj  0k  Pk  �l   m   m  @m  `m  €m   m  Àm  àm   n  4 €-  03  `3  À>  p?  ðJ  ÐK  �W  pX  ða  @b   n   o     €              f    7 †   
 ˜  `„ „       €  ð       |  ð       x  ð       t  ð       p  ð      
 l  ð      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ _  )( ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(( à  ð  0  P  �     Ð  à        0   0                   f    7 †   
 ›  `„ „       €  ð       |  ð       x  ð       t  ð       p  ð      
 l  ð      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ _  )x ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(x Ð  à       P  `  �     Ð  à     0  @  P  �  °  À  Ð        @  P  �  °  À  Ð    0  @  P   0   à  °!                 f    7 †   
 Ÿ  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )@ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(@°  ð    0     À  0  `  �  °  à     P  À    `     à    €  °  à    @  p     à    `  �  À  ð  0  p     Ð     0  `  �  À  ð    0  P  p  °  à    0  P  p  €     À  à      0   P   p   �   °   Ð   à    !   !  @!  €!   !  À!   "   "  0"  `"  €"  Ð"  ð"   #   #   $      @$  À$     @              f    7 †   
 £  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(      À  ð     `  �  °  ð    @  p     à    P  p     À  ð     P  �  À  ð     P  €  °  à    @  p     Ð     @  p     Ð  ð    @  p     Ð     0  `     Ð     0  `     °  à    0  P  p  �  °  Ð  ð    0  �  °  Ð      à    !     @              f    7 †   
 §  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )ð ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(ð p  �  Ð  ð    @  p  �  °  à       0  p  �  °  à    @  P  p  �  °  à    `  €     À  à       P  �  Ð    P     à     `     à    @  p  �  °  à     P  €  À  ð     @  `  €     À   P  à  `     @              f    7 †   
 «  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )È ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(È €  À  à        P  €  °  à        @  `  €     À  à  ð    0  `     À  ð     P  €     Ð     0  p     Ð     0  `  �  À  ð     p  °  ð  0  `  �     À  à   p           @              f    7 †   
 ¯  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )  ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(     0  P  €     À  ð     P  p  �  °  Ð  ð       @  p  �  Ð  ð       @  `     à     `  �  Ð     p     Ð     0  `  �  À   �  à  `     €              f    7 †   
 ³  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )x ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(x °  0  P  €     À  Ð        @  `  €  �  À  à  ð     @  P  €       P  p  °  à  P  �  °  à   Ð
  À  à     €              f    7 †   
 ·  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )P ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(P @
  `
  €
  °
  à
       @  p  �     À    @  `  ð       0  À   Ð  à  p     €              f    7 †   
 »  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  ) ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(    À  à           0  �     €              f    7 †   
 À  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )( ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(( à  0  P  p  �  À  à        @      `  `  4              0        €              f    7 †   
 Ä  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  ) ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(    À  à           0  �     €              f    7 †   
 É  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )( ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(( à  0  P  p  �  À  à        @      `  `  4              0        €              f    7 †   
 Í  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )P ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(P   @  `  €  °  Ð  à     0  P  `  €  °  à     0  p     À      �	  À  à     €              f    7 †   
 Ò  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )pÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(p  0  P     À  Ð  à     @  P  `  �  °  à     @  €  °  Ð     °  À       P  `  à  ð  0  @  0  P  À  Ð  0  P  à  ð     0      @   À   Ð    !  0!  �!  °!  ð!   "  à"  P#  p#  �#  °#  Ð#  ð#  $  0$  P$  p&  €&  à&   '  0'  @'  €'  �'  à'  ð'   )   )   )  °)  *   *  �*   *  à*   +   ,  @,  À,  Ð,  @-  P-  €-  �-  Ð-  ð-  /  €/   /  À/  à/   0   0  @0  `0  €0   3  3  p3  �3  Ð3  à3   4  04  €4  �4   5  À5  06  @6  °6  À6   7  07  €7   7  À8  à8  `9  p9  À9  Ð9  :   :  p:  �:  °;   <  @<  `<  €<   <  À<  à<   =   =  °@  à@   A   A  @A  `A  ÐA  B  @B  pB  €B  �B  °B  ÐB  ðB   C  0C  PC  pC  �C  °C  ÐC  àC   D  0D  PD  €D  °D  àD  E  0E  `E  �E  °E  ÐE  ðE  F  0F  @F  `F  0I  @I   I  ÀI   J  J  PJ  `J  °J  ÀJ  ðK  L  €L  �L   M  M  pM  €M  ÐM  ðM  0O  PO  ÐO  àO  0P  @P  €P  �P  àP   Q  @R  °R  ÐR  ðR  S  0S  PS  pS  �S  °S  4   À  ð  p$   %   0  €1  @=   >   G  ðG  ÐS  ÐT  4 €         0%  �  0>     €              f    7 †   
 Õ  `„ „       €  ð       |  ð       x  ð       t  ð       p  ð      
 l  ð      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ _  )( ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(( À  à  0  @  p  €  °  À  ð      0     ð                 f    7 †   
 Ù  `„ „       €  ð       |  ð       x  ð       t  ð       p  ð      
 l  ð      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ _  )x ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(x P	  `	  �	   	  Ð	  à	  
   
  P
  `
   
  °
  À
  Ð
        @  P  €     À  Ð        @  P  �  °  À  Ð   0   P     4           €                      f    7 †   
 Ý  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )@ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(@Ð  0  p  °  ð     `   °   ð    !  `!  �!  À!  0"  �"  Ð"  #  P#  �#  Ð#   $  @$  €$  À$  ð$   %  p%   %  Ð%   &  0&  `&  �&  À&  ð&   '  @'  `'  €'   '  À'  ð'  (  @(  p(   (  Ð(  à(   )   )  @)  `)  €)   )  À)  à)   *  0*  P*  p*  �*  °*  Ð*  à*   +   +  0+  P+  p+  à+  ,  0,  P,  p,  �,  À,  à,   -  -   .      @.  À.     @              f    7 †   
 á  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(@  à  0  P  p  �  À     P  €     À  ð     @  p  �  °  à     @   `   �   À   ð    !  P!  €!  °!  ð!  0"  p"   "  Ð"   #  0#  p#  °#  à#  $  @$  p$   $  Ð$   %  0%  `%  €%   %  À%  à%   &   &  P&  p&   &  à&  '  0'  P'  p'  �'  °'  Ð'  ð'  (  0(  �(  °(  Ð(      à)   *     @              f    7 †   
 å  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )ð ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(ð à     0  P  p     à        P  p  €     Ð    p     °  Ð  ð    0  `  �  À     0  `  �  À  ð     P  €  °  à     P   €   À   ð    !  P!  €!  °!  Ð!   "   "  @"  €"   "  À"  ð"  0#  p#  �#  °#  Ð#  ð#  %      0%  °%     @              f    7 †   
 é  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )È ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(È P  p  �  °  Ð  ð  0  `  €     À  à        @  `  €     °  Ð  ð     @  `  �  À  ð     @  `  €  °  Ð  ð     @  €  °  ð     P  �  À  ð  0  `  �  À     @      °   Ð      @              f    7 †   
 í  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )  ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(    @  €     Ð  ð     P  €     À  à        0  P  p     À  à     @  P  p     Ð     @  p     À     0  `  �  À  ð  0  p  °   €  Ð  P     €              f    7 †   
 ñ  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )x ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(x Ð  @  `  �  °  Ð  à    0  P  p  �     Ð  ð    0  P  `  °  ð  0  P  p  €  À  ð     P  �   ð  À  à     €              f    7 †   
 õ  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )P ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(P À  à     @  p  �     À  ð       @  �     °  Ð  0  p  �      P  `  ð     €              f    7 †   
 ú  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  ) ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ( ð  	  0	  P	  €	   `  �	  ð	     €              f    7 †   
 ÿ  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )( ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ((    P  p  �  °  à        @  `   À  €  �  4 @         À  P  @     €              f    7 †   
   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  ) ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ( ð  	  0	  P	  €	   `  �	  ð	     €              f    7 †   
 	  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )( ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ((    P  p  �  °  à        @  `   À  €  �  4 @         À  P  @     €              f    7 †   
   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )P ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(P à    0  P  €     °  Ð        0  P  €  °  Ð     p  °  ð     `     À     €              f    7 †   
   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )pÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(p€(   (  À(  )  0)  @)  P)  �)  °)  À)  Ð)   *   *  P*  p*  °*  ð*   +  @+  p+   .  @.  °.  À.  ð.   /  @/  P/  €/  �/  €0   0  1   1  €1  �1  ð1   2  `2  €2  p3  �3   4  4  p4  €4  À4  Ð4   5  @5  06   6  À6  à6   7   7  @7  `7  €7   7  °9  Ð9  0:  @:  €:  �:  à:  ð:  0;  @;  P<  p<  à<  ð<  0=  @=  À=  Ð=  0>  P>  p?  �?   @  @  p@  €@  Ð@  à@   A  @A  `B  ÐB  ðB  C  0C  PC  pC  �C  °C  ÐC  PF  pF  ÐF  àF   G  0G  €G  �G  ÐG  àG  ðH  I  €I  �I  ÐI  àI  `J  pJ  ÐJ  ðJ  L  0L  °L  ÀL  M   M  `M  pM  ÀM  àM   O  pO  �O  °O  ÐO  ðO  P  0P  PP  pP  T  0T  PT  pT  �T  °T   U  0U  `U  €U   U  °U  ÀU  àU  V  0V  PV  pV  �V  °V  ÀV  àV  W  0W  PW  pW  �W  °W  ÀW  àW   X  PX  €X  °X  àX  Y  0Y  `Y  °Y  ÐY  €\   \  ð\   ]  @]  P]  °]  À]   ^  ^  @_  `_  Ð_  à_  0`  @`  °`  À`   a  @a  €b   b   c  0c  �c   c  àc  ðc  0d  Pd  �e   f   f  @f  `f  €f   f  Àf  àf   g  4 €&  0,  `,  À7  p8  ðC  ÐD  �P  pQ  ðZ  @[   g   h     €              f    7 †   
   `„ „       €  ð       |  ð       x  ð       t  ð       p  ð      
 l  ð      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ _  )( ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(( @  `  °  À  ð     0  @  p  €   0   �  `                 f    7 †   
   `„ „       €  ð       |  ð       x  ð       t  ð       p  ð      
 l  ð      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ _  )x ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(x @  P  €  �  À  Ð       @  P  �     °  À  ð    0  @  €     °  À  ð     0  @  €     °  À   0   P                    f    7 †   
   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )@ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(@€    Ð      @   P   p       Ð   !  P!  Ð!   "  P"  �"  Ð"   #  @#  €#  °#  ð#   $  `$  �$  À$  ð$  0%  `%  €%  °%  à%   &  P&  €&  °&  Ð&  ð&  '  0'  P'  p'   '  À'  ð'   (  P(  €(  �(  °(  Ð(  ð(  )  0)  P)  p)  �)  °)  à)   *   *  @*  `*  €*  �*  °*  Ð*  à*   +   +  �+  À+  à+   ,   ,  @,  p,  �,  °,  À,  Ð-   P  ð-  p.     @              f    7 †   
   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(      Ð     0  `     Ð    0  P  p  �  À  ð     `  €     Ð  ð     P   €   °   à   !  @!  p!   !  Ð!  "  P"  €"  °"  à"  0#  `#  �#  À#  ð#  0$  `$  �$  À$  ð$  %  0%  P%  p%  �%  °%  à%  &  P&  €&  °&  Ð&  ð&  '  0'  P'  p'  �'  °'  Ð'  ð'  P(  p(  �(   à   )  À)     @              f    7 †   
 #  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )ð ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(ð   0  p  �  À  ð    0  @  p     Ð  ð  `  €     À  à  ð    @  `     Ð     @  p     Ð     0  P  €  °  à     @   €   °   à   !  @!  p!   !  Ð!  ð!   "  @"  `"   "  À"  à"  #  P#  �#  °#  Ð#  ð#  $  0%   À  P%  Ð%     @              f    7 †   
 '  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )È ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(È p  �  °  à     @  €  À  ð     @  `  €     À  à       0  P  p     Ð     0  `  €  �  À     0  p     Ð     0  `  �  À     0  p     Ð    @  €  °  à       p   !   !     @              f    7 †   
 +  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )  ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(  Ð     0  `  €     Ð     0  P  p  �  °  Ð  à        P  p  �  °  Ð  ð        P  p  �  °  à    P     à     `     Ð  ð  ð   à    �     €              f    7 †   
 /  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )x ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(x 0  �  À  ð    0  @  p  �  °  Ð  ð     0  P  p  �  °  À  ð    @  p     °  ð     à     P   0           €              f    7 †   
 3  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )P ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(P °  Ð  ð  0  `  p  €  À  à  ð     0  P  p  €  À     0  `      0  P  à     €              f    7 †   
 8  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  ) ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ( Ð	  ð	  
  0
  P
   0	  `
  À
     €              f    7 †   
 =  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )( ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ((    p  �  °  Ð        @  `  €   à     Ð  4 €            �  �     €              f    7 †   
 B  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  ) ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ( Ð	  ð	  
  0
  P
   0	  `
  À
     €              f    7 †   
 G  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )( ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ((    p  �  °  Ð        @  `  €   à     Ð  4 €            �  �     €              f    7 †   
 L  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )P ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(P 0  `  €     à  ð       P  p  �     Ð     0  `  Ð     @  `   °  ð       €              f    7 †   
 P  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )pÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(p�*  °*  Ð*   +  @+  P+  `+  �+  °+  Ð+  à+   ,  P,  €,  �,  à,  -  @-  P-  à-  00  @0  p0  �0   1  1  P1  `1  �1   1  �2  °2   3  03  €3   3  @4  P4  €4  �4  €5   5  6   6  p6  €6  Ð6  à6  07  P7  @8  °8  Ð8  ð8  9  09  P9  p9  �9  °9  °;  Ð;  0<  @<  �<   <  ð<   =  @=  P=  `>  €>  ?   ?  �?  °?   @   @  `@  p@  €A   A   B  0B  pB  €B  ÐB  àB  0C  PC  pD  àD   E   E  @E  `E  €E   E  ÀE  àE  PH  pH  ÐH  àH  0I  @I  �I   I  àI  ðI   K   K  °K  ÀK  0L  PL   L  ÀL   M  M   N  @N  ÀN  ÐN  O   O  pO  €O  ÐO  ðO  Q  €Q   Q  ÀQ  àQ   R   R  @R  `R  €R  PV  pV  �V  °V  ÐV  ðV  @W  €W   W  ÐW  àW  ðW  X  0X  PX  `X  €X  °X  ÐX  ðX  Y  0Y  PY  `Y  �Y  °Y  ÐY  ðY  Z  0Z  `Z  �Z  ÐZ   [  0[  P[  p[  �[   [  À[  €^   ^   _  _  `_  p_  À_  Ð_  `   `  Pa  pa   b  b  €b   b  ðb  c  Pc  `c  �d  °d  0e  @e  €e  �e  àe  ðe  @f  `f   g  h  0h  Ph  ph  �h  °h  Ðh  ðh  i  4 �(  @.  p.  Ð9  €:   F  àF   R  €S   ]  P]  0i  0j     €              f    7 †   
 S  `„ „       €  ð       |  ð       x  ð       t  ð       p  ð      
 l  ð      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ _  )( ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(( @  P  �  °       @  P  €  �   0      p                 f    7 †   
 V  `„ „       €  ð       |  ð       x  ð       t  ð       p  ð      
 l  ð      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ _  )x ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(x °  À  ð     0  @  p  €  °  À          0  p  �     °  ð       0  `  €     °  à        0   0   À  �                  f    7 †   
 Z  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )@ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(@0        p  �     `  �  Ð    @  €  À    `  �  Ð    P  €  °  ð  P   €   °   Ð   ð    !   !  @!  p!   !  Ð!  "  @"  p"   "  Ð"   #  0#  P#  €#   #  À#  à#  $  0$  `$  €$   $  À$  Ð$  ð$  %  0%  P%  p%   %  À%  à%   &   &  @&  P&  p&  �&  °&  Ð&  ð&  '  €'   '  À'  ð'   (  0(  P(  p(  €(  �)     °)  0*     @              f    7 †   
 ^  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(°  P  €     à     0  p  À  à     0  `  �  °  à        P  €  °  Ð     0  `  �  À  ð     `     à    @  p     à     P  €  °  à     @   p       Ð   ð   !  0!  P!  p!  �!  À!  à!  "  P"  €"   "  À"  à"   #   #  @#  `#  €#   #   $   $  @$   �  P%  p%     @              f    7 †   
 b  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )ð ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(ð @  `     À  à    @  `  �  °  Ð  ð    `  €  À  à          @  p     Ð     @  p     Ð     0  `     Ð    @  p  °  à    0  P  €  °  à    @  p     à     @  `  €     Ð    0  P  p       �   !     @              f    7 †   
 f  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )È ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(È À  à    0  P  €     Ð  ð    0  P  p  �  °  Ð  ð        @  p     À  ð     P  p  €     À  à  0  P  p  �  °  Ð  à     `     ð  0  p  °  ð  0  `     à   �     @     @              f    7 †   
 j  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )  ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(     Ð  ð  0  P  p     Ð        @  `  €     À  Ð        @  `  €     °  ð     @  `  €  °  à    `  �  À  ð     P  €  À  0      P  Ð     €              f    7 †   
 n  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )x ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(x    €     Ð  ð       P  p  �  °  Ð  à     @  `  €     °    P  p  �  À  à  0  P  Ð     0   `  P  p     €              f    7 †   
 r  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )P ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(P P  p  �  Ð  ð       P  p  �     à     0  @  °  ð    0  Ð   Ð	  ð  €     €              f    7 †   
 w  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  ) ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ( Ð  ð    0  P   @  `  À     €              f    7 †   
 |  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )( ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(( à  0  P  p  �  À  à        @    
  `  p  4              0        €              f    7 †   
 �  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  ) ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ( Ð  ð    0  P   @  `  À     €              f    7 †   
 †  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )( ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(( à  0  P  p  �  À  à        @    
  `  p  4              0        €              f    7 †   
 Š  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )P ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(P @  p  �  °  ð        0  `  €     °  à        0  p  °  ð      Ð
           €              f    7 †   
 Ž  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )pÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(p�  À  à     0  `  p     À  à  ð    @  p  �  À     P  `  �  0  P  �     Ð  à  P  `  °  À     À  0   @       À    !  @!  �!   !  �"  °"  @#  P#  °#  À#  ð#   $  @$  `$  P%  À%  à%   &   &  @&  `&  €&   &  À&  À(  à(  )   )  `)  p)  à)  ð)  @*  P*  p+  �+  ,   ,  `,  p,  ð,  -  p-  €-  �.  °.  0/  @/   /  °/  0  00  `0  p0  €1  ð1  2  02  P2  p2  �2  °2  Ð2  ð2  `5  €5  °5  À5   6  6  €6  �6  à6  ð6  8  08  °8  À8   9  9  �9  °9  :   :  0;  P;  Ð;  à;  @<  P<  °<  Ð<   =  =   >  �>  °>  Ð>  ð>  ?  0?  P?  p?  �?  0C  PC  pC  �C  °C  ÐC   D  PD  €D   D  ÀD  ÐD  àD   E  0E  PE  pE  �E  °E  ÐE  àE   F  0F  `F  �F  ÀF  àF   G  G  0G  `G  �G  ÀG  ðG   H  PH  pH  �H  ÐH  ðH  �K  °K  àK  ðK  0L  @L  °L  ÀL  M   M  `N  €N   O  O  PO  `O  àO   P  `P  pP   Q  ÀQ  @R  PR  °R  ÀR   S  @S  pS  €S  °T   U  @U  `U  €U   U  ÀU  àU   V   V  4    P  €  à&  �'  3  ð3  °?  �@  J  `J  @V  @W     €              f    7 †   
 ‘  `„ „       €  ð       |  ð       x  ð       t  ð       p  ð      
 l  ð      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ _  )( ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(( �  °       @  P  €  �  À  Ð   0   à  °	                 f    7 †   
 •  `„ „       €  ð       |  ð       x  ð       t  ð       p  ð      
 l  ð      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ _  )x ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(x °  À  ð     0  @  p  €  °  À          0  p  �     °  à        0  p  �     °  à        0   0   À  °  4 p         ð  €  p                 f    7 †   
 ™  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )@ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(@�  °     À  ð     P  �  Ð  `  Ð     0  `  �  À  ð     `  �  À  ð     P  €  À  ð      P   €   °   à    !  `!  �!  À!  ð!   "  @"  `"  �"  °"  Ð"  ð"  #  @#  `#  �#  °#  Ð#  ð#  $  0$  P$  p$  �$  °$  à$   %   %  @%  `%  �%  °%  Ð%  ð%   &   &  @&  €&  Ð&   '   '  P'  `'  �'  °'  Ð'  à'  ð(   p  )  �)     @              f    7 †   
 �  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(`  à  0  p  �  °  Ð  ð     €  °  à    @  p  €     À  à     0  P  €  °  à    @  p     à     `     Ð     0  p     à    @  p     Ð      0   `   �   À   à    !   !  @!  p!  �!  À!   "  0"  P"  p"  �"  °"  Ð"  ð"  #  0#  P#  °#  Ð#  ð#   @   %   %     @              f    7 †   
 ¡  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )ð ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(ð 0  P  p     À  ð     P  p  €  °  à    €     À  à       0  `  �  °  à     @  p     Ð  ð    @  p     Ð     0  p     Ð     0  `  �  À  ð     P  €  À  ð  0  `  €     Ð    0  P  p       �   !     @              f    7 †   
 ¥  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )È ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(È Ð    @  `  �  °  à  ð    0  @  p  �  °  Ð  ð       @  p     Ð  ð     P  €  °  Ð     0  `     Ð     0  `  �  À  ð     P     à     `  �  À  Ð            0  P     @              f    7 †   
 ©  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )  ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(  @  p     à        P  €  °  Ð  ð    0  `  �     À  ð    0  P  p  �     À  ð    0  P  €  °  à     P  €  À     @  p  p   `  �       €              f    7 †   
 ­  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )x ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(x À     @  p  �  °  À  ð    0  P  p  €  °  Ð  ð    0  @  p  �  °  à     P  €  À  0  p      Ð     À     €              f    7 †   
 ±  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )P ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(P P  p  �  À  ð       @  p  �     À          @  p  °  Ð      €
  €       €              f    7 †   
 µ  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  ) ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ( P  p  �  °  Ð   À  à  @     €              f    7 †   
 º  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )( ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(( P     À  à     0  P  p  �  °     Ð     4 °         0  À  À     €              f    7 †   
 ¾  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  ) ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ( P  p  �  °  Ð   À  à  @     €              f    7 †   
 Ã  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )( ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(( P     À  à     0  P  p  �  °     Ð     4 °         0  À  À     €              f    7 †   
 Ç  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )P ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(P    P  p  �  À  à  ð    @  `  p  �  Ð     0  p  ð     P  `   °  à        €              f    7 †   
 Ë  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )pÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(pP  p  �  à          `  €  �     Ð  ð     @  €  À  ð    @  Ð  à  0  @  p  €         P   `   P!  p!  à!  ð!  P"  p"   #  #  @#  P#  @$  `$  à$  ð$  @%  P%  °%  Ð%  &   &   '  p'  �'  °'  Ð'  ð'  (  0(  P(  p(  �*   *   +   +  P+  `+   +  °+   ,  ,   -  @-  À-  Ð-  0.  @.  °.  À.   /   /  @0  `0  à0  ð0  `1  p1   1  °1  ð1  2  03   3  À3  à3   4   4  @4  `4  €4   4   7  07  �7  °7  ð7   8  @8  P8   8  °8  À9  à9  P:  `:  Ð:  à:  @;  P;   ;  À;  à<   =  €=  �=  à=  ð=  0>  @>  �>  °>  Ð?  @@  `@  €@   @  À@  à@   A   A  @A  ÐD   E   E  @E  `E  €E  ðE  0F  `F  �F   F  °F  ÐF  ðF  G   G  PG  pG  �G  °G  ÐG  ðG   H   H  PH  pH   H  ÐH   I  0I  PI  €I  °I  ÐI  ðI  J  0J  PJ  `J  €J  PM  `M  ÀM  àM   N  0N  pN  €N  ÐN  àN  P  0P   P  °P   Q  0Q  �Q   Q  ðQ  R  PS  pS  ðS   T  PT  `T   T  °T   U   U  `V  ÐV  ðV  W  0W  PW  pW  �W  °W  ÐW  4 P     0  �(  @)  À4   5  `A  @B  ÀK  L  ðW  ðX     €              f    7 †   
 Î  `„ „       €  ð       |  ð       x  ð       t  ð       p  ð      
 l  ð      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ _  )( ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(( ð    `  p     °  à  ð   	  0	   0   @	   
                 f    7 †   
 Ò  `„ „       €  ð       |  ð       x  ð       t  ð       p  ð      
 l  ð      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ _  )x ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(x P  `  �     Ð  à       P  `     °  À  Ð        @  P  €     À  Ð    0  @  P  �  °  À  Ð   0   `  P  4          �                      f    7 †   
 Ö  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )@ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(@À   À!  ð!  "   "  @"  P"  €"  À"   #  �#  Ð#  $  @$  €$  °$  à$  %  @%  €%  Ð%   &  0&  P&  p&   &  Ð&   '  0'  `'  �'  À'  ð'   (  P(  €(  °(  à(  )  @)  `)  �)  °)  Ð)  *  @*  p*  �*  À*  à*   +   +  @+  `+  €+   +  À+  ð+  ,  0,  P,  p,  �,   ,  À,  à,  ð,  -  0-   -  Ð-  ð-  .  0.  P.  €.   .  À.  Ð.  à/   `   0  €0     @              f    7 †   
 Ú  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(   À  ð    P  p  �  À      0   `   �   À   ð    !  @!  `!  €!  °!  à!  "  @"  p"  °"  Ð"  ð"   #  P#  €#  °#  ð#   $  P$  €$  °$  à$  %  P%  €%  À%  ð%   &  P&  €&  À&  ð&   '  @'  p'   '  à'  (  @(  `(  €(   (  À(  à(   )   )  @)  `)  €)   )  À)  à)   *  `*  €*   *   ð  °+  Ð+     @              f    7 †   
 Þ  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )ð ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(ð `  €  �  Ð  ð     @  p  �  ð     @  °  Ð  ð  �    0  P  p  à    !   !  0!  @!  P!  `!  À!  à!   "   "  @"  `"  p"  �"  À"  à"   #   #  @#  `#  p#  �#  à#   $   $  @$  `$  €$   $  à$  �%  Ð%  &  P&  €&  °&  à&  '  ð'   ð  (  �(     @              f    7 †   
 â  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )È ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(È `  €  �  Ð  ð     @  p  �  ð     @  °  Ð  ð  �        @  `     Ð  à  ð       P  p  �  °  Ð  à  0   P   p   �   °   À    !   !  @!  `!  €!  �!  Ð!  @"  €"  À"  ð"  0#   ð  0$  P$     @              f    7 †   
 æ  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )  ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(  @  `     à  ð    @  P  p  °  à  @  °  Ð     à  `  �  À  à     °  À        0  @  p  �  °  À        @  `  À          P    !       !   !     €              f    7 †   
 ê  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )x ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(x `  €  À       0  `  p  �  Ð     `  Ð  ð        P  p  �  °       `  p  À  à     0  `  �   @  à        €              f    7 †   
 î  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )P ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(P à        `  €     °  ð    0  @  p  °  à     P  p     À  p   p  �        €              f    7 †   
 ò  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  ) ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ( €     À  à      ð    p     €              f    7 †   
 ÷  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )( ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ((    p  �  °  Ð        @  `  €   à     °  4 `         à  p  `     €              f    7 †   
 û  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )P ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(P    `  €     à  ð       P  p  �     Ð     0  `  Ð     0  @   �  Ð  ð     €              f    7 †   
 ÿ  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )pÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(pP(  °(  Ð(   )   )  P)  `)  �)   )  Ð)  à)   *  @*  €*  �*  à*   +  @+  P+  �+   .  @.   .  °.  à.  ð.  P/  `/   /  °/  �0  °0  01  @1   1  À1   2  2  p2  �2  €3   3   4  04  p4  €4  ð4  5  P5  `5  @6  °6  Ð6  ð6  7  07  P7  p7  �7  °7   9  °9  ð9   :  P:  `:  à:  ð:  0;  @;  `<  €<  ð<   =  P=  `=  Ð=  à=  @>  `>  €?   ?  @   @  €@  �@  Ð@  à@  0A  PA  pB  àB   C   C  @C  `C  €C   C  ÀC  àC  @F  PF  �F   F  ðF   G  €G  �G  ÐG  àG   I   I  �I   I  ðI   J  pJ  €J  àJ   K   L  @L  °L  ÀL   M  0M  pM  €M  ÐM  ðM  O  €O   O  ÀO  àO   P   P  @P  `P  €P  PT  pT   T  ÐT   U  0U   U  ÐU   V   V  @V  PV  pV  €V  °V  ÐV  ðV  W  0W  PW  `W  €W  °W  ÐW  ðW  X  0X  PX  `X  €X  °X  ÐX  ðX  Y  0Y  PY  `Y  €Y  ÀY  àY  p\  €\  À\  Ð\   ]  0]  °]  À]   ^  ^  P_  p_  à_  ð_  @`  P`  À`  Ð`  0a  Pa  �b  °b   c  0c  �c   c  àc  ðc  @d  `d   e  f  0f  Pf  pf  �f  °f  Ðf  ðf  g  4 �&  @,  p,  Ð7  €8   D  àD   P  €Q   [  P[  0g  0h     €              f    7 †   
   `„ „       €  ð       |  ð       x  ð       t  ð       p  ð      
 l  ð      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ _  )( ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ((    À       P  `  �     Ð  à   0   ð  À                 f    7 †   
   `„ „       €  ð       |  ð       x  ð       t  ð       p  ð      
 l  ð      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ _  )x ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(x      P  `  �     Ð  à       `  p  €  �  À  ð       @  `  €  �  À  à       @  `  €  �   0      ð                 f    7 †   
 	  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )@ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(@"  #  @#  `#  p#  �#  °#  à#   $  P$  �$  %  P%  �%  Ð%   &  0&  `&   &  Ð&   '  P'  €'   '  À'  ð'   (  P(  €(  °(  à(  )  @)  p)   )  Ð)   *  0*  `*  �*  °*  à*   +   +  `+  �+  À+  à+  ,  0,  P,  p,  �,  °,  Ð,  ð,  -  @-  `-  €-   -  À-  à-  ð-  .  0.  @.  `.  €.  ð.   /  @/  `/  €/   /  Ð/  ð/  0   0  01   °   P1  Ð1     @              f    7 †   
   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(P  Ð  ð  0   P   �   °   ð   0!  `!  �!  À!  ð!   "  @"  p"  �"  °"  Ð"   #  0#  `#  �#  Ð#   $  0$  `$  �$  À$  ð$   %  P%  €%  °%  à%  &  @&  €&  °&  ð&   '  P'  €'  °'  ð'   (  P(  p(   (  Ð(  )  @)  p)  �)  °)  Ð)  ð)  *  0*  P*  p*  �*  °*  Ð*  ð*  +  0+  �+  °+  Ð+      à,   -     @              f    7 †   
   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )ð ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(ð p  �     à       P  €     ð     P  °  Ð  ð  �   !  0!  P!  p!  à"   #   #  0#  @#  P#  `#  À#  à#   $   $  @$  `$  p$  �$  À$  à$   %   %  @%  `%  p%  �%  à%   &   &  @&  `&  €&   &  à&  �'  Ð'  (  P(  €(  °(  à(  )  ð)   ð  *  �*     @              f    7 †   
   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )È ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(È €     °  ð       €     °    @  p  À  ð  @  °     @  `  €  À  ð             0   `   €   À   à   ð    !  @!  `!  €!   !  À!  Ð!   "  @"  `"  €"   "  °"   #  `#   #  à#  $  @$     P%  p%     @              f    7 †   
   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )  ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(  À  à     `  p  �  À  Ð  ð  0  `  À  @  `  �  `  à    @  `     0  @  €     °  À  ð     0   @   €       À   à   @!  €!   !  Ð!  €"       "   #     €              f    7 †   
   `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )x ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(x `  €    @  p  �     °  à        @  `  p  °  Ð  ð    0  @  p  °  ð  0  `  �        P  �   °  à        €              f    7 †   
 !  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )P ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(P   0  P  �  °  Ð  à    0  P  `     à  ð     p  °  Ð  ð  �   �  °  @     €              f    7 †   
 %  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  ) ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ( P  p  �  °  Ð   À  à  @     €              f    7 †   
 *  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )( ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(( à  0  P  p  �  À  à        @      `  p  4              0        €              f    7 †   
 .  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )P ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(P à    0  P  €     °  à    0  @  p     À  Ð     `  €  À  ð   `     À     €              f    7 †   
 2  `          œ  ð       ˜  ð       ”  ð       �  ð       „  ð1       €  ð       |  ð       x  ð      
 l  ð1      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ L _  )pÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(pð.   /  @/  €/  �/  À/  Ð/   0   0  @0  P0  p0   0  Ð0  ð0   1  €1  °1  À1  ð1   4  À4  5   5  p5  €5  À5  Ð5   6  6   7   7   7  °7  8  08  p8  €8  à8   9  ð9  :  �:   :  ð:   ;  @;  P;   ;  À;  °<   =  @=  `=  €=   =  À=  à=   >   >  0@  P@  �@   @  à@  ð@  `A  pA   A  °A  ÐB  ðB  pC  €C  ÐC  àC  pD  �D  ÐD  àD  ðE  F  �F   F  àF  ðF  @G  PG   G  ÀG  àH  PI  pI  �I  °I  ÐI  ðI  J  0J  PJ  ÐL  ðL  0M  @M  €M  �M   N  N  @N  PN  pO  �O  P   P  pP  €P  Q  0Q  pQ  €Q  �R  °R  0S  @S  €S  �S  àS  ðS  @T  `T  €U  ðU  V  0V  PV  pV  �V  °V  ÐV  ðV  �Z  °Z  ÐZ  ðZ  [  0[  €[  °[  à[   \   \  0\  @\  `\  �\  °\  Ð\  ð\  ]  0]  @]  `]  �]  À]  ð]   ^  @^  `^  p^  �^  À^  ð^   _  P_  €_  °_  Ð_  ð_  0`  P`   c   c  `c  pc  °c  Àc  0d  @d  pd  €d  Àe  àe  `f  pf  Àf  Ðf  `g  €g  Àg  Ðg   i   i   i  °i  ði   j  Pj  `j  °j  Ðj  l  €l   l  Àl  àl   m   m  @m  `m  €m  4  -  °2  à2  @>  ð>  pJ  PK  W  ðW  pa  Àa   m   n     €              f    7 †   
 5  `„ „       €  ð       |  ð       x  ð       t  ð       p  ð      
 l  ð      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ _  )( ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(( À  à  0  @  p  €  °  À  ð      0     à                 f    7 †   
 8  `„ „       €  ð       |  ð       x  ð       t  ð       p  ð      
 l  ð      	 h  ð       d  ð       `  ð       T  ð1       P  ð       H  ð!         ðÁ         ð!         ð!           ð! ÿ _  )x ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ(x À  Ð       @  P  €  �  À  Ð       0  @  €     °  À  ð    0  @  €     °  À  ð    0  @   0   Ð                        ÿÿÿÿ    þÿÿÿ    ýÿÿÿ    üÿÿÿs          % 6P             H             @             8             0             (                                                  
             	                 „Š         6  Š         3  ¤‰         /  4‰         +  Äˆ         &  Tˆ         "  ä‡           t‡           ‡           ”†           $†           ´…         
  D…           Ô„           d„            ôƒ         ü  „ƒ         ø  ƒ         ó  ¤‚         ï  4‚         ë  Ä�         ç  T�         ã  ä€         ß  t€         Û  €         ×  ”         Ó  $         Ï  ´~         Ì  D~         È  Ô}         Ä  d}         ¿  ô|         »  „|         ¶  |         ²  ¤{         ®  4{         ª  Äz         ¦  Tz         ¢  äy         ž  ty         š  y         –  ”x         ’  $x         �  ´w         ‹  Dw         ‡  Ôv         ‚  dv         }  ôu         x  „u         s  u         o  ¤t         k  4t         g  Äs         c  Ts         _  är         [  tr         W  r         T  ”q         Q  $q         M  ´p         H  Dp         C  Ôo         >  do         9  ôn         4  „n         0  n         ,  ¤m         (  4m         $  Äl            Tl           äk           tk           k           ”j           $j         
  ´i           Di            Ôh         û  dh         ö  ôg         ò  „g         î  g         ê  ¤f         æ  4f         â  Äe         Þ  Te         Ú  äd         Ö  td         Ó  d         Î  ”c         Ê  $c         Å  ´b         Á  Db         ¼  Ôa         ¸  da         ´  ô`         °  „`         ¬  `         ¨  ¤_         ¤  4_            Ä^         œ  T^         ™  ä]         –  t]         ’  ]         Ž  ”\         ‰  $\         …  ´[         �  D[         }  ÔZ         y  dZ         u  ôY         q  „Y         m  Y         i  ¤X         f  4X         c  ÄW         _  TW         [  äV         W  tV         S  V         O  ”U         K  $U         G  ´T         C  DT         ?  ÔS         ;  dS         7  ôR         4  „R         1  R         -  ¤Q         )  4Q         $  ÄP            TP           äO           tO           O           ”N           $N           ´M           DM           ÔL         þ  dL         ù  ôK         ô  „K         ð  K         ì  ¤J         è  4J         ä  ÄI         à  TI         Ü  äH         Ø  tH         Ô  H         Ð  ”G         Ì  $G         È  ´F         Ã  DF         ¾  ÔE         ¹  dE         ´  ôD         ¯  „D         ª  D         ¥  ¤C            4C         ›  ÄB         –  TB         ‘  äA         Œ  tA         ‡  A         ƒ  ”@           $@         z  ´?         u  D?         p  Ô>         k  d>         f  ô=         a  „=         \  =         W  ¤<         R  4<         M  Ä;         H  T;         C  ä:         >  t:         ;  :         8  ”9         4  $9         0  ´8         +  D8         &  Ô7         !  d7           ô6           „6           6           ¤5           45           Ä4           T4            ä3         ý  t3         ú  3         ö  ”2         ò  $2         í  ´1         é  D1         å  Ô0         á  d0         Ý  ô/         Ù  „/         Õ  /         Ñ  ¤.         Í  4.         Ê  Ä-         Ç  T-         Ã  ä,         ¿  t,         »  ,         ·  ”+         ³  $+         ¯  ´*         «  D*         §  Ô)         £  d)         Ÿ  ô(         ›  „(         ˜  (         •  ¤'         ‘  4'         �  Ä&         ˆ  T&         ƒ  ä%           t%         {  %         w  ”$         s  $$         o  ´#         k  D#         g  Ô"         d  d"         a  ô!         \  „!         X  !         T  ¤          O  4          J  Ä         E  T         @  ä         ;  t         6           1  ”         ,  $         )  ´         &  D         "  Ô           d           ô           „                      ¤           4           Ä            T         ü   ä         ø   t         ô            ð   ”         ë   $         æ   ´         á   D         Ü   Ô         ×   d         Ò   ô         Í   „         È            Ã   ¤         ¾   4         ¹   Ä         ¶   T         ³   ä         ¯   t         «            ¦   ”         ¢   $         �   ´         ™   D         •   Ô         ‘   d         �   ô         ‰   „         …            �   ¤         }   4         z   Ä
         w   T
         s   ä	         o   t	         k   	         g   ”         c   $         _   ´         [   D         W   Ô         S   d         O   ô         K   „         H            E   ¤         @   4         <   Ä         8   T         3   ä         .   t         )            $   ”            $            ´             D                                                                                                     J                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                       J                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                       J                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                       J                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                       J                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                       J                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                       J                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                       J                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                       J                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                       @  Ð  À                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                  @  Ð  À                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                          0                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                       P  à  à                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                          "                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                       €                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                              þÿÿÿÿÿÿÿ                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                þÿÿÿÿÿÿÿ                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                þÿÿÿÿÿÿÿ                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                þÿÿÿÿÿÿÿ                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                þÿÿÿÿÿÿÿ                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                þÿÿÿÿÿÿÿ                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                þÿÿÿÿÿÿÿ                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                       þÿÿÿÿÿÿÿ                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                        "                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                       p                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                         €                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                               €                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                       °  @  @                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                  €                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                       °  @  @                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                          `  ð  à                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                  `  ð  à                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                          p     ð                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                  p     ð                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                à  p  `                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                  P  à  Ð                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                             0                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                        0                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                     0%  �  0>                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                      €                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                  À  P  @                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                         À  P  @                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                    �  �                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                            �  �                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                           0                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                               0                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                         ð  €  p                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                      0  À  À                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                  0  À  À                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                      �                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                           à  p  `                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                             0                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                 z  
      ä y'      "   " ¹z  l      â Ey  0-    € â ™x?    â y(      !   b ¹z  F   
   â x*      ÿ  â $r,ÿÿ   ÿ Ž â y      &   ¢ rE ÿ       â x.      ÿ  â $r0ÿÿ   ÿ Ž â y      '   â rc ÿ       â x       ÿ  â x"      ÿ  â x$      ÿ  â r& ÿ       Ä r 'ÿ   pRò är 'ÿ   pRð är3'(   ÿ(Ž È/ x ÿ   3  È r 3   ÿ� â $r ÿÿ   ÿ Ž Æ x)ÿ     È | )   pbô Ú G)   +    € ê y      %   "Î r ÿ       â $rÿÿ   ÿ Ž â x
øÿÿÿÿÀŽ â $z6 }  ÿŽ â xC)   ÿ   â 'z y   Ž â r33
  €ÿàÿ ä z=  X      â 'z	 s   Ž â rL33   ÿ� Ä x3   ÿ   â$x3   ÿ Ž â z	ÿ t  	  â $z w  ÿŽ â xLÿ   L  â $v9ÿ Y  ÿ Ž â rÿ   ÿ3Ž ä r66   ÿàõ ä xMÿ   L  Ä x   ÿ   ä zÿ z    ä rAÿÿ   ÿä â $z p  )Ž â xBÿ   3  â $z
 |  )Ž âxD   ÿÀŽ â $z |  Ž â x   ÿàÿ â $z
	 v  
Ž âr* ÿ       â $z	 v  Ž â x n       â $z p  
Ž â z@C q  ÿàÿ â %x
   ÿŽ È %xn   ÿŽ È %v  € Ž â x2
J   áÙ Æ %x   Ž â r4
  €ÿà÷ ä r5ÿ   ¥~ â $vÿ q  ÿ Ž â r7	  €ÿäÿ â %x
n   ÿŽ â rFL   ÿà÷ ärH3   ÿàù â$x;   CŽ ârJ3
   ÿàõ â $x<   CŽ âr1C   ÿŽ â$x8   CŽ ârNB   ÿä â$x>   CŽ âr:B   ÿä â $x?   CŽ Ä %xn   LŽ È %x   Ž È $xL   MŽ ä %x   
Ž È %x   
Ž È r=H   ÿàõ Ê $x9   N â r=   ÿàõ È �yMB   ‘ ¢  r9   ÿä Ê �yI   • è �y   • ( �y-   • h �yG   • " r=J   ÿà÷ Ä r=   ÿàõ ä r9:   ÿäÿ Æ $x   9 ä �yXB   ‘ h �yY   • h �yO   • h �y   • h �y   • b r=   ÿàõ È r9   ÿä Ê �yKJ   • b r`=2   ÿàõ È ra95   ÿä Ê �y``    • b xu $       ä xM   ÿ   ÈO x   ÿÀŽ ä/ xM   ÿ   â$tÿ   ÿ Ž â xÿ   IøŽ ä�xIÿ   I  ä xM   ÿ   ä x   ÿÀŽ â %v   Ž â x[I   øŽ Ä xÿ   øŽ â$xM    ÿ Ž â �yI    ™ ¢  xÿ     â %vZ[   Ž È %v   Ž âx   øŽ â �yZZ    ™ è �y]    ™ ¢ %vP   Ž Ì �yPP    ™ ¢ xM   ÿ   ä xÿ   -  äx   ÿÀŽ â $xM   ÿ Ž Ê x   øŽ ä xM   ÿ   È x   ÿÀŽ â %vT   Ž âxÿ   G  äxM   ÿ   äxÿ   GøŽ â �yTT    ™ " xÿ   -øŽ ä x   øŽ â %v   Ž â r=   ÿàõ Æ %v   Ž â�y-    ™ ¦
 %vR   Ž â �yV    ™ ¦  $x   9 ä �yRR    ™ ¨ �yWJ   • ¢ xX   ÿ   Ä xÿ   Y  ä x   ÿÀŽ â $xX@   ÿ Ž âxX   ÿ   È x_   øŽ ä x   ÿÀŽ äxÿ   O  ä xX   ÿ   È x   øŽ ä xÿ   OøŽ Æ %v   Ž â/ r=4   ÿàõ Æ %v^_   Ž âxÿ   YøŽ â �y\    ™ b r97   ÿä â %v   Ž ä�y^^    ™ ¤ %v   Ž ä �y[    ™ ¨  �yY    • ¨  �ya    ™ ¢  $xX   ÿ Ž âxbX   ÿ   Ä xMÿ     ä xX   ÿ   â$xXX   ÿ Ž â xGÿ     ä x   ÿÀŽ ä x   ÿÀŽ ä xMM   XøŽ Ä xG   bøŽ ä xÿ   øŽ ä xÿ   øŽ â %vM   Ž â rM36   CàÙ Æ %v   Ž ârOBA   ÿ¤~ â �yS    ™ ¤  %v   Ž â rG36   @àÙ â �yX    ™ ¤  %v   Ž â r=F   ÿà÷ â �y_    ™ ¢" r=   ÿàõ â %vM Z  u Ž Ä �yU    ™ ¤  $xOO$   ÿŽ â r9L   ÿäÿ â $x	   9 â rBA   ÿ¤~ â %vG Z  u Ž â rO   ÿàÿ â �yQj   ‘ ¤  $x$   ÿŽ Ä �yOj   ‘ ¦  r   ÿàÿ â �yM   ™ ¨  �ye   ™ ¨  �yG   ™ ¨  �yd   ™ ¨  �yb   ™ ¨  �yf   ™ ¨  �yg   ™ â  r=   ÿàõ Æ/ �yj   ™ â r
=   ÿà÷ Æ �yk   ™ â $x   9 Æ �yl   ™ â r9   ÿäÿ Æ �ym   ™ " r36   1àµ Æ �yn   ™ " r36   <àÙ Æ �yq   ™ ( �yh    • ( �yo    • ( �yi    ™ " rBA   ÿd~ Æ �ys   ™ " %v Z  u Ž â/ rBA   ÿ¤~ ä�yr    ™ " r36   ;àµ â $x$   ÿŽ ä �yp    • " rBA   ÿd~ â $x$   ÿŽ â r   ÿàÿ â �yt    • " %v Z  u Ž Æ �yv   ™ " %v Z  u Ž â/ r36   8àÙ ä�y|   ™ " $x$   ÿŽ â r   ÿàÿ ä �yz   ™ " rxBA   ÿ¤~ ä r36   >àµ â�y‚   ™ " r36   ?àÙ â $x   Ž â r        â %v Z  u Ž â ryBA   ÿ¤~ â �y€   ™ $ %v Z  u Ž Ä �yw   ™ $ %v Z  u Ž â ruBA   ÿd~ â �y~   ™ $ $x{x$   ÿŽ ä �yx   ™ " $x}u$   ÿŽ Æ �yu   ™ " $xyy$   ÿŽ ä $x   {Ž â �y�   ™ ( �y{   ™ b ry   ÿàÿ ä r}   ÿàÿ â �yy   ™ h �y}   ™ h �y…   ™ h �y    ™ h �y„    ™ h �yƒ    • h �y�    • h �y†   ™ h �yŽ   ™ h �y‡   ™ h �y   ™ h# �yŒ   ™ h �y“   ™ h �y‹   ™ h �y’   ™ h �yŠ   ™ h �y‘   ™ h �y‰   ™ h �y�   ™ h �yˆ   ™ h �y   ™ h �y    ™ b xœKÿ   ÿÀŽ Æ �y�    • d $x¡K @  œŽ ä �y    ™ h �yš    • b x£¡   ÿ   Æ �y   ™ h# �y™   ™ b xÿ   K  Æ �y˜   ™ h �y—   ™ h �y–   ™ h �y•   ™ h �y”   ™ h �y    ™ h �y›    • h �y   ™ b x££€€ÿÀŽ Æ �y    ™ b r   ÿpŽ â/ $x¡¡   ÿ Ž ä �yŸ   ™ b x¢   ÿ   ä x¤£€€€€ÿüŽ â �yž   ™ b x¢¢€€ÿÀŽ ä x   ÿ   â �y�   ™ b x¡¡€€ÿÀŽ Ä x¥¤ÿþþþÿàÿ â �yœ   ™ b x€€ÿÀŽ Æ �yK   ™ h �y    ™ h �y    • b x¤¡€€€€ÿüŽ ä r¥¥£   ÿüŽ Ä x¢€€€€ÿüŽ ä x€€€€ÿüŽ ä x£ÿþþþÿàÿ ä x¤¤ÿþþþÿàÿ ä x¥˜º  ÿ    ä r££¢   ÿüŽ Ä x¢ÿþþþÿàÿ ä r¤¤¡   ÿüŽ ä xÿÀŽ äx€€€€IÞŽ äO r¢   ÿüŽ ä x¢£˜º  ÿ    Ä x¡¤˜º  ÿ    â $x¤   
Ž â x¢ÿÀŽ äx¡ÿÀŽ äx¢¢€€€€]ÞŽ ä x£`ÿ   ÿÀŽ ä x¡¡€€€€ZÞŽ Ä� x˜º  ÿ    ä r¢¢  €ÿàÿ â $x` @  £Ž â r¡¡  €ÿàÿ ä xÿÀŽ äx€€€€PÞŽ ä xZ¡€€€€Z´Ž Ä x¡   ÿ   äx]¢€€€€]´Ž ä x£ÿ   `  ä x¢   ÿ   â $x   
Ž â x`¡€€ÿÀŽ ä r£££   ÿpŽ Ä xP€€€€P´Ž ä x`€€€€ÿüŽ ä x£   ÿ   äx¢€€ÿÀŽ ä x¡ÿþþþÿàÿ â $x££   ÿ Ž â x€€ÿÀŽ Ä x¢€€€€ÿüŽ ä r¡¡`   ÿüŽ ä x`€€€€ÿüŽ ä x¢¢ÿþþþÿàÿ ä x£€€ÿÀŽ ä x``ÿþþþÿàÿ Ä r¢¢   ÿüŽ ä x€€€€ÿüŽ ä r``   ÿüŽ ä xÿþþþÿàÿ ä x``˜º  ÿ    ä x£¡˜º  ÿ    Ä r   ÿüŽ ä x¥¢˜º  ÿ    ä x`ÿÀŽ ä xI¤€€€€I´Ž ä x``€€€€-ÞŽ ä x¡£ÿÀŽ Ä x¤£€€€€VÞŽ ä x˜º  ÿ    ä x¢¥ÿÀŽ äx£¥€€€€TÞŽ â$x``   
Ž â xÿÀŽ ä r££¢  €ÿàÿ Ä x€€€€RÞŽ äxÿ   W  ä x¢Wÿ   ÿÀŽ ä r  €ÿàÿ â $x @  Ž â r¢W¢   ÿpŽ ä xR€€€€R´Ž â $x€   ÿ Ž â x¢   ÿ   Ä x   ÿ   ä x€€ÿÀŽ ä x€€ÿÀŽ ä r¡¤¡  €ÿàÿ ä x-`€€€€-´Ž ä x¤€€€€ÿüŽ Ä x`€€€€ÿüŽ ä x€€ÿÀŽ ä xW¢   ÿ   ä x¤¤ÿþþþÿàÿ ä xV¡€€€€V´Ž ä x``ÿþþþÿàÿ Ä x¡€€€€ÿüŽ ä xWW€€ÿÀŽ ä r¤¤   ÿüŽ ä r`   ÿüŽ ä x¡¡ÿþþþÿàÿ ä x¢W€€€€ÿüŽ Ä x˜º  ÿ    ä r¡¡   ÿüŽ ä x¢¢ÿþþþÿàÿ ä x¤˜º  ÿ    ä x€€€€\ÞŽ äx`¡˜º  ÿ    Ä xÿÀŽ ä r¢¢W   ÿüŽ ä x¡€€€€^ÞŽ äx¤ÿÀŽ ä x`€€€€[ÞŽ ä r  €ÿàÿ Ä x¢¢˜º  ÿ    ä x``ÿÀŽ ä xYÿ   ÿÀŽ ä xW¢€€€€aÞŽ är``  €ÿàÿ â $xY @  Ž â x¢¢ÿÀŽ Ä xÿ   Y  ä r¢W¢  €ÿàÿ ä rY   ÿpŽ ä xW   ÿ   â$x   ÿ Ž â x\€€€€\´Ž â $x¡¡   ¤
Ž â xY   ÿ   Ä xY   ÿ   ä x€€ÿÀŽ ä xWW€€ÿÀŽ ä x€€ÿÀŽ ä x€€ÿÀŽ ä x^¡€€€€^´Ž Ä xa¢€€€€a´Ž ä x¡€€€€ÿüŽ ä xT£€€€€T´Ž ä xYW€€€€ÿüŽ ä x¢€€€€ÿüŽ ä x£€€€€ÿüŽ Ä x[`€€€€[´Ž ä x¡¡ÿþþþÿàÿ ä x`Yÿþþþÿàÿ ä x¢¢ÿþþþÿàÿ ä x¤£ÿþþþÿàÿ ä r¡¡   ÿüŽ Ä r``W   ÿüŽ ä r¢¢   ÿüŽ ä r¤¤   ÿüŽ ä x¡¡˜º  ÿ    ä x``˜º  ÿ    ä x¢˜º  ÿ    Ä x¤¤˜º  ÿ    ä x¡€€€€XÞŽ äx¢¡ÿÀŽ ä xW`€€€€_ÞŽ ä x€€€€UÞŽ äxYÿÀŽ Ä x``ÿÀŽ ä x¤€€€€SÞŽ äx¡¤ÿÀŽ ä r¢  €ÿàÿ â $x`W   `
Ž â rQÿD   Q  â$x   ¡
Ž â xX€€€€X´Ž â &rIM   ÿ   â rOÿD   O  Ä r¢Y  €ÿàÿ â $xQ   ÿ Ž â x_`€€€€_´Ž â &rZe      â xS€€€€S´Ž ä x`        ä xO   ÿ   â &rOIG   ÿ   ä&rGaG   ÿ   â x   `ìŽ â &r]d      Ä &r`Zb   O   ä&rb^b   G   ä &rPf      ä &rMaM   ÿ   ä &r`]g   `   ä&rb[g   b   Ä &rVj      ä &rM^e   M   ä &r`Pk   `   ä&rb\k   b   ä &rTl      ä &rM[d   M   Ä &r`Vm   `   ä&rm_m   b   â xU¢€€€€U´Ž â &r-n      ä &rM\f   M   ä &rbTq   `   ä&rdXq   m   â xW        â &r`Ri      Ä &rM_j   M   ä &r-s   b   ä&rdUs   d   â x   WìŽ â &rMXl   M   ä &rbRr      ä&rGSr   d   Ä $r``   ÿŽ ä 0rhÿh    A   ä 0rpÿp    A   ä &rnUn   M   ä $rbb   ÿŽ ä 0rtÿt    A   Ä $rMG   ÿŽ â Er` `       â 0roÿo    A   â  rhp     A  âErb b       â  rGht     A  âErM M       â  rtot     A  â #r&`   &   â &rIv   ÿ   â #r$Gb   $   â #r#tM   #   â &rGav   ÿ   Ä &rMIw   ÿ   ä &rZx      ä&rG^x   G   ä &r`Zu   M   ä &rwaw   ÿ   ä &r]|      Ä &rG[|   G   ä &r`]{   `   ä &rb^u   w   ä &rPz      ä&rz\z   G   ä &rGPy   `   Ä &rb[{   b   ä &rV}      ä&rGV~   G   ä &rb\y   b   ä &rT‚      ä&r`T�   G   Ä &rb_~   b   ä &r-€      ä&rG-…   `   ä &rbX�   b   ä &r`R      ä&rdR„   G   Ä &rbU…   b   ä $r``   ÿŽ ä0rƒÿƒ    A   ä $rdd   ÿŽ â Er` `       â &rGS„   b   ä 0r�ÿ�    A   â  rhƒ     A  â $rMG   ÿŽ â Erd d       Ä  rGh�     A  â #r"`   "   â &rI†   ÿ   ä #r Gd       â &rGIŽ   ÿ   ä&rZ‡      ä&rŽaŽ   ÿ   ä &r`Z   G   Ä &r]Œ      ä &r†a†   ÿ   ä &rŽ^   Ž   ä &r]“   `   ä &rP‹      ä &r†^‡   †   Ä &rP’      ä &rVŠ      ä&r†[Œ   †   ä &rV‘      ä &rT‰      ä &r†\‹   †   Ä &r`T�      ä &r-ˆ      ä &r†_Š   †   ä &rG-   `   ä &r`R      ä &r†X‰   †   Ä $r``   ÿŽ ä &r†Uˆ   †   ä 0r�ÿ�    A   â Er` `       â &rS   †   ä &r}_}   z   â  rh�     A  â $r   ÿŽ Ä &r}X‚   }   â #rc`   c   â &rI   ÿ   ä&ra   ÿ   ä &rII   ÿ   ä&raa   ÿ   ä &r€U€   }   Ä &rZ™      â Er        â &r^™      ä &rZZ    I   â r�o�     A  â &r^^    a   ä &rS   €   ä &r]˜      Ä &r“[“   Ž   ä&r[˜      ä &r]]Ÿ   Z   â #r0�   0   â &r[[Ÿ   ^   ä $r   ÿŽ ä &rP—      Ä &r’\’   “   ä&r\—      ä &r]Pž   ]   ä&r\\ž   [   â Er        â &rV–      â  rboƒ     A  â &r‘_‘   ’   Ä &r_–      ä&r]V�   ]   ä&r__�   \   ä &rT•      â #r!b   !   â &r�X�   ‘   ä&rX•      Ä &rTTœ   ]   ä&rXXœ   _   ä &rbR   G   â x))   ÿàÿ â &r�U   �   ä &r-”      ä&rT-K   T   Ä &rU”      ä&rXUK   X   ä $rbb   ÿŽ â | )   pbô â 0ršÿš    A   ä &rS   �   ä &rR      Ä &rTR   T   ä &riSi   n   ä&rS      ä&rS   X   â Erb b       â $r   ÿŽ â  rhš     A  â 0r›ÿ›    A   Ä $rii   ÿŽ ä $r   ÿŽ ä$rTT   ÿŽ ä $r   ÿŽ ä$r   ÿŽ ä 0rÿ    A   â #r/b   /   â Er        â  rh›     A  â  rppo     A  â Eri i       â  r�o�     A  âErM M       â  roš     A  â r›o›     A  âEr        â  rhh     A  âEr        â  roo     A  â ErT T       Ä Er        ä x==p  ÿà÷ â #r%pi   %   â #r �M       â #r.   .   â #rE   E   â #r,›   ,   â #r+hT   +   â #r*o   *   â x11@   ÿàÿ Ä x;;@   ÿàÿ ä x<<@   ÿàÿ ä x88@   ÿàÿ ä x>>@   ÿàÿ ä x??@   ÿàÿ ä x@@@   ÿàÿ Ä xCC@   ÿàÿ ä r9ÿ9   ÿäÿ â D)       À â Gy  0Øÿÿÿÿƒ ê Ay        € êÏ Ey  0    € â G™      € ê $x''   (Ž Ê ˆs '& øÿ H   è ˆs '%€øÿ H   è ˆs '$ ùÿ H   è ˆs '#€ùÿ H   è ˆs '" úÿ H   è ˆs '!€úÿ H   è ˆs '  ûÿ H   è ˆs ' €ûÿ H   è ˆs 'c üÿ H   è ˆs '0€üÿ H   è ˆs '/ ýÿ H   è ˆs '.€ýÿ H   è ˆs 'E þÿ H   è ˆs ',€þÿ H   è ˆs '+ ÿÿ H   è ˆs '*€ÿÿ H   ä Ay        € ê {          ì M	        € ê „y(     H   b z1  r      Æ „y( €   H   ¤ $x51   ÿŽ ä„y(    H   â $x31   ÿŽ Æ „y(    H   ( „y(    H   h y
      %   ( „y( €  H   ( „y( €  H   ( „y( €  H   è „y(    H   â !r&       Æ/ „y(    H   â !r%       ÆO ‰      b !r$       Æ� ‰      ¢ !r        Æ„y( €  H   â !r"       Æ„y (    H   è „y"(    H   â !r##       Æ ‰      " !r!!       â !r       ä/ ‰      b !r00       â !r       äO ‰	      ¢ !rcc       Æ „y( €  H   â !r//       Æ ‰      è ‰!      â !r       Æ „y( €  H   â !r..       Æ ‰#      " !r       â/ !rE E       â !r		       âO x
   ÿ   ä ‰       â !r+"+       â z  r  p`ð ä‰
      b $z x  Ž â x   ÿàÿ Ä r (ÿ   p"p â $z ~  Ž â ‰	 €    ¢ !r!!       Æ� r1   ÿ† â‰&c      æ rÿÿ   ÿäÿ â !r        â z j  ÿ† â $x1   ÿŽ â 
 r  ÿàó â !r##       â z k  � â ‰"      â 
 j  ÿ„ â $ÿÿ   ÿŽ  â z  r  p`ò â !r
       â/ 3   ÿàù â !r',       â 
 k   â !r       â ‰      " !r	       âO r	1   ÿŠ â!r        â ‰ €    b r    ÿàõ â !r&c&       â� $rÿÿ   ÿŽ â ‰ @    ¢ z	 j  ÿŠ Ä 5   ÿà÷ â ‰0      â z		 k  � ä rÿÿ   ÿä â ‰,/      " ÿÿ   ÿäÿ ä z  j  ÿ„ â ‰%.      b 

 j  ÿ† Ä $ÿÿ   ÿä â !r       â z  k   â „y( €  H   â 
 j  ÿˆ ä 
 k  � â ‰2E      " !r       â/ !r        âO 
 k  $ â ‰4'      b !r$"       â x (   p"ò â !r00       â� ‰#      ¢ !r,/,       Æ‰6+      â !r..%       Æ‰      & 1   ÿàÿ â ‰!      b 3   ÿàÿ ä 8   ÿà÷ â‰       b !r2E2       â (   ÿàù Ä ‰ @    " !r4'4       â/ 7ÿÿ   ÿä â !r##       âO ‰%$      b ( j  ÿˆ â!r6+6       â� !r**       â ‰'&      ¢ ÿÿ   ÿäÿ â !r       â$x1   ÿŽ â ‰)0      â ( k  7 â !r!!       â8 j  ÿ† â ‰+,      "    ÿàõ â !r9        â 8 k  � â ‰ €    b 5   ÿàÿ â !r        â $ÿÿ   ÿ â ‰-.      " 
 j  ÿ„ â !r$$%       Ä/ ‰/2      b 
 k   ä    ÿàõ â ‰# €    b !r&&'       äO (ÿÿ   ÿä â ‰14      ¢ !r00)       â�  j  ÿ„ Ä ‰36      â !r,,+       â k  ( ä ‰*      " !r       Æ‰ €    b !r..-       Æ ‰ ! €    " !r22/       Æ/ ‰" €    b !r#       â $t#ÿ   ÿ Ž â !r441       âO ‰'$ €    ¢ !r663       Æ� ‰)& €    â !r7*       Æ‰+0 €    " %v j  # Ž â !r       ä‰-, €    b !r !        Æ ‰ @    " !r""       Æ/ ‰/. €    h ‰ @    b !r'$'       ÆO ‰12 €    ¢ !r)&)       Æ� ‰34 €    â !r+0+       Æ‰56 €    " !r-,-       Æ‰87      b !r       Æ ‰       " !r/./       Æ/ ‰#  @    b !r       Æ ‰! @    b !r121       ÆO ‰%" @    ¢ !r343       Æ� ‰$' @    â !r565       â!r878       ä‰&) @    " !r;        Æ ‰(+ @    " !r# #       Æ/ ‰       â !r!!       Æ ‰*- @    b !r%"%       ÆO ‰ / @    ¢ !r$'$       Æ� ‰      è ‰"1 @    b !r&)&       Æ‰,3 @    " !r(+(       Æ ‰78 €    ( ‰.5 @    " !r*-*       Æ/ ‰!      b !r /        ÆO ‰#      ¢ !r'       Æ� †	 9    â !r"1"       â!r,3,       ä† ;   â !r787       Æ ‰$      " !r9        â� !r.5.       ä ‰ %      â !r)!       Æ/ ‰&      b !r+#       ÆO ‰(      ¨ †	 9    è ‰*      " !r$       Æ ‰       ( † '    â !r%%        Æ� ‰!"      â !r&       Æ/ ‰#,      b !r(       ÆO ‰7 @    ¨ ‰.      b !r*       Æ†	 )    â	 !r        Æ † +   è	 †	 %    â	 !r!"!       â� !r#,#       ä/ †     â	 !r77       ÆO †	     â	 !r.       Æ†    è	 †	 
    è	 †     è	 †	 !    è	 † #   è	 ‰ 7      ( †	     â	 M™        € ê x   ÿàÿ â !r77        Ær   ÿàñ È rÿÿ   ÿä  ä z j  ÿ€ È z k    Ê †y 7    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y#      "   " ¹z  l      â Ey  €*    € â ™x?    â y$      !   b ¹z  F   
   â x&      ÿ  â r( ÿ       â y*      %   ¢ $r?ÿÿ   ÿ Ž â x      ÿ  â x      ÿ  â x      ÿ  â x       ÿ  â $r"ÿÿ   ÿ Ž â r #ÿ   pRð Ä r #ÿ   pRò â$x-#    $Ž â/ x**   ÿ   ÈO x ÿ   -  È r -   ÿ� ä r  ÿ       ä x%ÿ     È | %   pbô Ú G)  )    € ê y	      '   " $rÿÿ   ÿ Ž â r
 ÿ       â $z* p  %Ž â y      &   b xøÿÿÿÿÀŽ â $x:%   ÿ Ž â z7  X      â $r&ÿÿ   ÿ Ž â z3  Y      â $x--   
Ž â z9: q  ÿàÿ È x-   ÿ   ärD--   ÿ� ärÿ   ÿ3Ž ä xDÿ   D  ä x-   ÿ   â 'z	 y   Ž â xEÿ   D  Ä x;ÿ   -  â 'z s  
 Ž â/ zÿ z    ä x=   ÿÀŽ â %x   ÿŽ â zÿ t    Æ $z |  %Ž ä$z |  Ž â x*   ÿàÿ â $z v  Ž ä$z v  Ž ä $tÿn   ÿ Ž ä $z p  Ž Ä %xn   ÿŽ È %v  € Ž â x,J   áÙ Æ $z0	 }  ÿŽ â r.  €ÿà÷ â $z w  ÿŽ â r/ÿ   ¥~ â%x   Ž â r1  €ÿäÿ ä r00   ÿàõ â %xn   ÿŽ â r@-   ÿàù är>D   ÿà÷ â $vÿ q  ÿ Ž Ä $r<ÿÿ   ÿ â rB-   ÿàõ â $x5   :Ž âr+:   ÿŽ â$x2   :Ž âr4;   ÿä â $x6   :Ž ä$x8   :Ž ä %xn   DŽ â rDE   ÿäÿ Æ %x
   Ž È %x   Ž È %x   Ž È $xF   ; ä r7@   ÿàõ Ê $x3   F â r7   ÿàõ Ê �yB   ‘ ¢  r	3   ÿä Ê �yE   • è �y)   • ( �yA   • h �yC   • b r7B   ÿàõ Ê $x3   4 â r7   ÿàõ Ê �yB   ‘ b r3   ÿä Ê �yM   • h �yL   • h �yV   • h �y   • b r
7   ÿàõ Ê $x   3 Ê �yuJ   • b $tÿ   ÿ Ž â/ rH7,   ÿàõ È rI3/   ÿä Ê �yKH    • d rH7   ÿàõ Ê/ $xI   3 â rT7.   ÿàõ Ê $xU3   1 â xv $       ä x   ÿ   âO$x€   ÿ Ž âxG   ÿ   ä x   ÿÀŽ ä xJG   ÿÀŽ ä xÿ   EøŽ â��yGHJ   • ¢  xEÿ   E  Ä xSJÿ   )øŽ â%v   Ž âxEE   øŽ ä x)ÿ   )  â $x    ÿ Ž â �y]    ™ â %vPE   Ž Æ x))   øŽ â %vRS   Ž ä�yPP    ™ ¤ %vN)   Ž ä �yRR    ™ ¢  x   ÿ   Æ/ �yNN    ™ ¢ x   ÿÀŽ È xYÿ   AøŽ â$x   ÿ Ž âxAÿ   A  ä x   ÿ   ä xA   øŽ ä x   ÿÀŽ Æ %v   Ž â xÿ   CøŽ Æ %vXY   Ž â xCÿ   C  â �yQ    ™ b	x   ÿ   Æ �yXX    ™ ¢  xC   øŽ â %v   Ž È %v   Ž â �y\    ™ ¨  �yC    ™ ¢  $x   ÿ Ž Ê xJ   ÿÀŽ È xJÿ   MøŽ â$xJ€   ÿ Ž â xMÿ   M  Æ %v   Ž â	x)   ÿ   ä xM   JøŽ ä �yO    ™ "! x))   ÿÀŽ ä %v   Ž â xÿ   L  â �yMT    • ¢ x   ÿ   Ä x)ÿ   LøŽ ä �y)    ™ ¢  x   øŽ ä %v   Ž Ê �yL    ™ ¢ %vH   Ž â x   ÿ   äxÿ   V  â $x   ÿ Ž â x   ÿÀŽ â �yIH    ™ ¢  rA-0   :àµ ä x   øŽ Ä xÿ   VøŽ ä rE;<   ÿd~ â %vA Z  v Ž È $xA   ÿ Ž ä %v   Ž È/ %v   Ž â xA   ÿÀŽ â �yU    ™ ¤ $xEE$   ÿŽ ä �yA    ™ ¦  rE   ÿàÿ Ê �yY   ™ ¨  �yV   ™ ¢ xÿ   øŽ Æ �y[   ™ ¢ r`7>   ÿàõ â $xH   ÿ Ž â xEÿ     â �yZ   ™ ¢  %v   Ž â ra3D   ÿä ä xE   HøŽ â �yW   ™ ¢  rd7   ÿà÷ Ä r7   ÿàõ â/ �yS    ™ ¢ %v   Ž Æ �yH`j   ‘ ¢  r7   ÿàù Æ �yT   ™ ¢  $xe   3Ž â r3   ÿä ä r-0   9à§ â/ �y^   ™ ¢ $x   3 Æ �yE    ™ ¨ �y`    ™ ( �yJdj   ‘ " r-0   +àÙ ä/ r;<   ÿDþ â�y_    • " rc;<   ÿ¤~ â %v Z  v Ž Ä �yb    • $ $xe$   ÿŽ ä %v Z  v Ž â �ya    • $ re   ÿàÿ â $xcc$   ÿŽ â r        Æ/ $x   cŽ â �yf   ™ ( �yc   ™ ( �ye   ™ ( �yd   ™ ( �yo   ™ ( �yn   ™ ( �ym   ™ ( �yl   ™ ( �yk   ™ ( �yj   ™ ( �yi   ™ " rw-0   5àÙ Æ �ys   ™ " r-0   2àµ Æ �yh   ™ " rz;<   ÿ¤~ Æ �yr   ™ " ry;<   ÿd~ Æ �yg    ™ " r-0   8àµ Æ �yq    ™ ( �yp    • ( �yt    • " %vw Z  v Ž â/ rw;<   ÿd~ Æ $xz$   ÿŽ È $x   Ž ä %v Z  v Ž â r-0   6àÙ ä �y   ™ " $xww$   ÿŽ â rx;<   ÿ¤~ ä �y~   ™ $ rw   ÿàÿ Ä �yw   ™ " $x{x$   ÿŽ ä %v Z  v Ž â�y|   ™ & %v Z  v Ž â �yz   ™ & $xyy$   ÿŽ â �yx    ™ " $x   {Ž Æ �y{   ™ " ry   ÿàÿ Æ �yy   ™ ( �yv   ™ ( �y   ™ ( �y    • ( �y}   ™ ( �y„   ™ ( �yƒ   ™ ( �y‚   ™ ( �y�   ™ ( �y€    ™ ( �y‹    • ( �y…   ™ ( �yŠ   ™ ( �y   ™ "# xŽuÿ   ÿÀŽ Æ �y�   ™ ( �y‰   ™ " $x“u @  ŽŽ Æ �y�   ™ ( �yˆ   ™ ( �yŽ   ™ ( �y‡   ™ " x’ÿ   u  Æ �y�   ™ ( �y†   ™ " x‘“   ÿ   Æ �yu   ™ ( �y   ™ " x‘‘€€ÿÀŽ Æ �y   ™ " $x—’ @  ’Ž Æ �y    ™ " x““   ÿ   Æ �yŒ    • " x’‘€€€€ÿüŽ â $x”—€   ÿ Ž â x““€€ÿÀŽ â �y    ™ "! x’’ÿþþþÿàÿ Æ �y    • " x——   ÿ   ä x•”€€ÿÀŽ ä x”“€€€€ÿüŽ ä r’’‘   ÿüŽ ä x——€€ÿÀŽ ä x–•€€€€ÿüŽ Ä x””ÿþþþÿàÿ ä x’’˜º  ÿ    ä x—€€€€ÿüŽ ä x––ÿþþþÿàÿ ä r””“   ÿüŽ ä x‘’ÿÀŽ Ä xÿþþþÿàÿ ä x’’€€€€]ÞŽ ä� r––•   ÿüŽ ä x”˜º  ÿ    ä r——   ÿüŽ â $x’   ‘
Ž â x“–˜º  ÿ    Ä x’ÿÀŽ äx€€€€PÞŽ äO x–Kÿ   ÿÀŽ ä x”“ÿÀŽ äx““€€€€RÞŽ ä r’  €ÿàÿ Ä r–K–   ÿpŽ ä x——˜º  ÿ    â $x““   ”
Ž â x€€€€P´Ž â $x‘–   ÿ Ž â xP—ÿÀŽ äx——€€€€NÞŽ ä x’ÿ   K  Ä x€€€€]´Ž ä xR“€€€€R´Ž ä rK—P  €ÿàÿ ä x]–   ÿ   ä r“’’   ÿpŽ ä x‘‘€€ÿÀŽ Ä xKK€€€€N´Ž â $x’“€   ÿ Ž â x]]€€ÿÀŽ ä xN‘€€€€ÿüŽ ä x““   ÿ   ä xP]€€€€ÿüŽ ä xNNÿþþþÿàÿ Ä x”“€€ÿÀŽ ä xPPÿþþþÿàÿ ä rNN‘   ÿüŽ ä x’’€€ÿÀŽ ä x‘”€€€€ÿüŽ ä rPP]   ÿüŽ Ä xNN˜º  ÿ    ä x“’€€€€ÿüŽ ä x•‘ÿþþþÿàÿ ä x]P˜º  ÿ    ä x‘NÿÀŽ äxNN€€€€QÞŽ Äx““ÿþþþÿàÿ ä r••”   ÿüŽ ä x”Gÿ   ÿÀŽ ä xP]ÿÀŽ â$xNN   ‘
Ž â x]]€€€€XÞŽ ä r““’   ÿüŽ Ä r”G”   ÿpŽ ä r]]P  €ÿàÿ ä x““˜º  ÿ    ä xP•˜º  ÿ    ä xQN€€€€Q´Ž â $xN”€   ÿ Ž â xX]€€€€X´Ž Ä x’“ÿÀŽ äx]PÿÀŽ äx““€€€€\ÞŽ ä xPP€€€€CÞŽ ä xGÿ   G  ä xNN€€ÿÀŽ Ä r““’  €ÿàÿ â $x’G @  GŽ â rPP]  €ÿàÿ ä x]”   ÿ   ä xGN€€€€ÿüŽ ä x]]€€ÿÀŽ ä xGGÿþþþÿàÿ Ä xCP€€€€C´Ž ä x‘’   ÿ   äxP]€€€€ÿüŽ ä rGGN   ÿüŽ ä x’’   ÿ   ä x‘‘€€ÿÀŽ Ä xPPÿþþþÿàÿ ä xNG˜º  ÿ    ä x\“€€€€\´Ž ä x“’€€ÿÀŽ ä x’‘€€€€ÿüŽ ä rPP]   ÿüŽ Ä xGN€€€€OÞŽ äxNNÿÀŽ ä x’’ÿþþþÿàÿ ä xPP˜º  ÿ    â $xNG   N
Ž â x–Mÿ   ÿÀŽ ä x”“€€€€ÿüŽ Ä r’’‘   ÿüŽ ä xGP€€€€)ÞŽ äxPPÿÀŽ ä r–M–   ÿpŽ ä x””ÿþþþÿàÿ ä xMÿ   M  Ä x]’˜º  ÿ    ä rPGP  €ÿàÿ ä xG–   ÿ   ä r””“   ÿüŽ â $x“M @  MŽ â x‘]€€€€LÞŽ äx’]ÿÀŽ Ä xM–   ÿ   ä xGG€€ÿÀŽ ä r‘‘’  €ÿàÿ ä xNN€€€€O´Ž ä x)P€€€€)´Ž ä xP“   ÿ   Ä xOM€€ÿÀŽ ä xMG€€€€ÿüŽ â $x““   ÿ Ž â xL‘€€€€L´Ž ä x””˜º  ÿ    ä x‘P€€ÿÀŽ ä xPMÿþþþÿàÿ Ä x]”€€€€IÞŽ äx““€€ÿÀŽ ä x””ÿÀŽ ä rPPG   ÿüŽ ä xG“€€€€ÿüŽ â $x”]   ”
Ž â xPP˜º  ÿ    Ä x]O€€€€ÿüŽ ä x–Gÿþþþÿàÿ ä xGP€€€€AÞŽ äxPPÿÀŽ ä x’]ÿþþþÿàÿ ä rPGP  €ÿàÿ â &rGY   ÿ   â r’’O   ÿüŽ â &rYNY   ÿ   â xM‘€€€€ÿüŽ â &rGV   G   âx’’˜º  ÿ    â &rY)V   Y   â xI”€€€€I´Ž â &rGR[   G   â x”Mÿþþþÿàÿ Ä xM’€€€€UÞŽ â &rYL[   Y   â x’’ÿÀŽ â &rGKZ   G   âr””‘   ÿüŽ â &rYIZ   Y   â r’M’  €ÿàÿ ä xPP€€€€A´Ž â &rGXW   G   â x””˜º  ÿ    Ä xU’€€€€U´Ž â &rYPW   Y   â r––“   ÿüŽ â &rGQT   G   âxM”€€€€SÞŽ ärHÿ=   H  ä x””ÿÀŽ â &rYUT   Y   â x––˜º  ÿ    Ä xT        ä xAH   ÿ   â &rH\^   G   ä $xVM   ”
Ž â xM–€€€€EÞŽ â&rHC`   H   â x––ÿÀŽ ä xAA   TìŽ Ä rJÿ=   J  ä r–M–  €ÿàÿ â $tMÿ   ÿ Ž â xJJ   ÿ   â $rHHA   ÿŽ ä 0r_ÿ_    A   ä 0rbÿb    A   â xGJ   MìŽ Ä ErH H       ä  rM_b     A  È #r"MH   "   â &rHc   ÿ   ä&rJe   ÿ   ä &rHf   H   ä&rcNc   ÿ   ä &rMd   J   ä &rHRo   H   Ä &rc)f   c   ä &rMRn   M   ä &rHKm   H   ä&rJLo   c   ä &rMKl   M   ä &rHXk   H   Ä &rJIm   J   ä &rTXj   M   ä &rMQi   H   ä&rJPk   J   ä &rOQs   T   ä &rH\h   M   Ä &riUi   J   ä &rJ\r   O   ä &rHCg   H   ä&rJCq   J   ä $rHAH   ÿŽ ä 0rpÿp    A   Ä $rJAJ   ÿŽ â ErH H       â 0rtÿt    A   â  rM_p     A  äErJ J       ä  rO_t     A  â #r MH       â &rH   ÿ   ä#rOJ      â &rJN   ÿ   Ä &rw   H   È &rR~      È &rK|      È &rX{      ä &rMv   ÿ   ä &rQz      ä &rvNv   ÿ   ä &r\y      ä &rM   M   Ä &rv)   v   ä &rCx      ä 0rHÿ    A   ä $rA   ÿŽ Æ  r_H     A  ä Er        â &reNe   ÿ   È #r      â &r…   ÿ   ä&r…N…   ÿ   ä &r   ÿ   ä&rN   ÿ   ä &rŠ      ä &re)d   e   Ä &rw)w   J   ä&r…)Š   …   ä &r�      ä&r)�      ä &rR‰      ä &reLn   e   Ä &rwL~   w   ä &rMR}   M   ä&rvL}   v   ä&r…L‰   …   ä &rR�      ä&rL�      Ä &rKˆ      ä &reIl   e   ä&rwI|   w   ä &rMK„   M   ä&rvI„   v   ä&r…Iˆ   …   Ä &rKŽ      ä&rIŽ      ä &rX‡      ä &rjPj   e   ä&rwP{   w   ä &rMXƒ   M   Ä &rvPƒ   v   ä&r…P‡   …   ä &rX�      ä&rP�      â xVV€€€€S´Ž â &rQ†      ä &rsUs   j   Ä &rwUz   w   ä &rMQ‚   M   ä&rvU‚   v   ä&r…U†   …   ä &rQu      ä&rUu      â xE–€€€€E´Ž â &r\      â x%%   ÿàÿ â &rYV^   Y   Ä &rhVh   i   ä&rrVr   s   ä&rwVy   w   ä &rM\�   M   ä&rvV�   v   ä&r…V   …   Ä &r\      ä&rV      â | %   pbô â &rC      ä &r`E`   Y   ä&rhEg   h   ä&rrEq   r   Ä &rxEx   w   ä &rJC€   M   ä&rvE€   v   ä&rE   …   ä &rC      ä&rE      Ä $r``G   ÿŽ ä $rhGh   ÿŽ ä$rrGr   ÿŽ ä$rxGx   ÿŽ ä $rJAJ   ÿŽ ä $rvGv   ÿŽ Ä $rA   ÿŽ ä $rG   ÿŽ ä 0raÿa    A   ä 0rTÿ‹    A   ä 0rŒÿŒ    A   ä $rAA   ÿŽ Ä $rGG   ÿŽ ä 0rÿ    A   â Er` `       â  rbba     A  â  rpap     A  âErh h       â  rtat     A  âErr r       â  rHaH     A  âErx x       â  rOaT     A  âErJ J       â  raŒ     A  â Erv v       â  rM_T     A  âEr        â  r_Œ     A  âEr        â  r_     A  âErA A       â  raa     A  â ErG G       Ä x77p  ÿà÷ â #r!b`   !   â #rph      â #rtr      â #rHx      â #rMJ      â #r Ov       â #r?   ?   â #r(   (   â #r'A   '   â #r&aG   &   â x++@   ÿàÿ â $r3ÿÿ   3Ž â x55@   ÿàÿ Ä x22@   ÿàÿ ä x66@   ÿàÿ ä x88@   ÿàÿ ä x99@   ÿàÿ ä x::@   ÿàÿ â D)       À â Gy  °Úÿÿÿÿƒ ê Ay        € ê Ey      € â G‰  ð     € ê $x##À  $Ž Ê ˆs #" ùÿ H   è ˆs #!€ùÿ H   è ˆs #  úÿ H   è ˆs #€úÿ H   è ˆs # ûÿ H   è ˆs #€ûÿ H   è ˆs # üÿ H   è ˆs #€üÿ H   è ˆs # ýÿ H   è ˆs # €ýÿ H   è ˆs #? þÿ H   è ˆs #(€þÿ H   è ˆs #' ÿÿ H   è ˆs #&€ÿÿ H   ä Ay        € ê {          ì M        € ê „y$     H   b z * r  p`ð Ä z)  r      â „y$ €   H   ¢ r $ÿ   p"p Æ „y$ €  H   â $x/)   ÿŽ ä$x%)   ÿŽ â „y$    H   ( „y$    H   h „y$ €  H   ( „y	$    H   ( „y$ €  H   ( „y$    H   è „y$ €  H   â !r"       Æ/ „y$    H   â !r!!       Æ_ ‰      " !r       Æ� „y$ €  H   â !r        Æ„y$    H   â !r       Æ„y$ €  H   â !r       Æ y      &   b !r		       Æ ‰!      ¢ !r       â !r       ä y      '   " !r       Æ ‰      â !r        Æ ‰
      " !r??       Æ ‰      b $z x  *Ž Æ/ ‰      b x**   ÿàÿ â !r!       äO ‰      ¢ !r((       â z * r  p`ò â !r''       â ‰	      ¢ $z ~  Ž â x $   p"ò â !r&&       â !r       â� ‰      " 
  r  ÿàõ â !r

       Ä‰      â ÿÿ   ÿä â !r       ä‰      " 
  j  ÿ„ â!r       ä/ ‰ €    b 
  k   â !r       ÄO ‰ ?      ¢ !r	       Æ ‰      h ‰!(      b !r       Æ ‰"'      " !r       Æ� ‰#&      â !r       Æ‰	      " !r        â/ /   ÿà÷ ä ‰
      b !r??        ÆO ‰      ¢ !r       âr)   ÿ„ ä ‰      b !r!(!       Æ ‰      b !r''"       â !r#&#       ä� ‰       " !r	       Æ‰"      â 	%   ÿàû â !r
       ä/ ‰$      b r)   ÿˆ â !r       äO ‰&?      ¢ $ÿÿ   ÿŽ â 
 j  ÿ† â !r       â‰  @    " rÿÿ   ÿä â $r.ÿÿ   ÿ â !r       â ‰ €    b ÿÿ   ÿäÿ Ä 
	 j  ÿŠ â ‰(!      b !r         â z j  ÿ„ ä ‰*'      " !r""       â� z
 j  ÿˆ ä ‰,#      â !r$$       â/ 
 k  � Ä ‰ €    b !r&?&       âO 
		 k  � ä ‰ €    ¢ !r         â)   ÿàÿ ä ‰ €    " !r       âz k   Ä ‰ €    b !r(!(       â %   ÿàÿ ä ‰ €    b !r*'*       â z k  . â !r,#,       â� ‰!  €    " -   ÿàù â !r       Ä/ ‰#" €    b $x)   ÿŽ â 2ÿÿ   ÿä â !r       âO ‰%$ €    ¢ - j  ÿˆ â !r       ä‰       â 0   ÿà÷ â !r       Ä‰'& €    " - k  2 â $1ÿÿ   ÿŽ â !r       â ‰ @    b r   ÿàõ ä 0 j  ÿ† â ‰)( €    b r.ÿÿ   ÿä â !r! !       â z j  ÿ„ â ‰+* €    " !r#"#       Ä/ z k  . â ‰ @    b 0 k  1� â !r%$%       äO ‰ @    ¢ !r1        â� /   ÿàÿ ä ‰-, €    â !r'&'       Æ‰ @    " !r        â   ÿàõ ä ‰ @    b !r)()       â x        ä ‰ @    b !r+*+       â .ÿÿ   ÿä â %v j   Ž â ‰ ! @    " !r       â/  j  ÿ„ Ä ‰"# @    b !r       âO  k  . ä ‰$% @    ¢ !r-,-       Æ� ‰&' @    â !r       Æ‰       " !r       Æ‰() @    b !r       Æ ‰*+ @    b !r !        Æ ‰      " !r"#"       Æ/ ‰,- @    b !r$%$       ÆO ‰      ¢ !r&'&       Æ� ‰      â !r/        Æ‰      " !r()(       Æ‰      b !r*+*       Æ ‰!       b !r3       Æ ‰#"      " !r,-,       Æ/ ‰%$      b !r)       ÆO ‰&      ¢ !r       Æ� †	 1    â !r       Æ† /   â !r       Æ‰'(      â !r! !       Æ ‰*      " !r#"#       Æ ‰,      " !r%$%       Æ/ †	 3    â !r&       ÆO † )    è †	 
    è † 
   â !r'('       Æ� †	     â !r*       Æ† !    â !r,       Æ †	 #    è † %   è †	     è † '    è †	     â M™        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y       "   " ¹z  l      â Ey  ð'    € â ™x?    â y      !   b ¹z  F   
   â r" ÿ       â $rÿÿ   ÿ Ž â y#      %   ¢ $r5ÿÿ   ÿ Ž â r9 ÿ       â x      ÿ  â $r7ÿÿ   ÿ Ž â x      ÿ  â x       ÿ  â $r;ÿÿ   ÿ Ž â r  ÿ   pRð Ä r  ÿ   pRò â$x%     Ž â/ x##   ÿ   ÈO xÿ   %  È r%   ÿ� È xÿ     È |    pbô Ú G)  €&    € ê y	      '   " $rÿÿ   ÿ Ž â xøÿÿÿÿÀŽ â $r
ÿÿ   ÿ Ž â y      &   b $v/ÿ q  ÿ Ž â r%%  €ÿàÿ â $z# p  Ž â x2   ÿ   â $v-ÿ X  ÿ Ž â r<%%   ÿ� â$x%   ÿ Ž âx1ÿ   %  â $x%   ÿ Ž â x<ÿ   <  â $x*/   2Ž ârÿ   ÿ3Ž â $x./   2Ž âx=ÿ   <  â $v+ÿ Y  ÿ Ž â r/2   ÿŽ â %x   ÿŽ â x3   ÿÀŽ Ä z02 q  ÿàÿ â 'z	 y   Ž È 'z s  
 Ž â/ zÿ z    Æ $tÿn   ÿ Ž â zÿ t    â $z |  Ž ä$z |  Ž â x#   ÿàÿ â $z v  Ž ä$z v  Ž ä $z p  Ž Ä %xn   ÿŽ È %v  € Ž â x$J   áÙ Æ $z(	 }  ÿŽ â r&  €ÿà÷ â $z w  ÿŽ â r'ÿ   ¥~ â%x   Ž â r)  €ÿäÿ ä r((   ÿàõ â %xn   ÿŽ âr6<   ÿà÷ är4ÿÿ   ÿä â %xn   <Ž â r8%   ÿàù Ä r:%   ÿàõ â %x
   Ž Æ r,1   ÿä â $r"ÿÿ   ÿ Ž ä $x//   2Ž ä $x<   =Ž ä %x   Ž È %x   Ž È $x>   1 ä r-8   ÿàõ Ê $x+   > â r-   ÿàõ È �yCB   ‘ ¢ $x	   + Ê �y@   • è �y=   • ( �y?   • h �yB   • " r-:   ÿà÷ Ä r-   ÿàõ ä r+,   ÿäÿ Æ $x   + ä �yIB   ‘ h �yJ   • h �yE   • h �yG   • h �yH   • b rb
-   ÿàõ Ê $xc   + Ê �ybbJ   • b xx $       ä xC   ÿ   âO$xC€   ÿ Ž âxC   ÿ   ä x   ÿÀŽ â $tÿ   ÿ Ž â x   ÿÀŽ ä xÿ   @øŽ Ä� x@ÿ   @  Æ %v^   Ž â x@   øŽ ä/ r@-$   ÿàõ ä xÿ   =øŽ â�y^^    ™ ¢ x=ÿ   =  â $xA+   ' â xC   ÿ   â %v   Ž È %v   Ž â �y_@    • â  x=   øŽ Æ �y]    ™ $ %v   Ž ä �y`    ™ ¨  �ya    ™ ¢  xC   ÿ   È xD   ÿÀŽ ä x=ÿ   ?  äx?Dÿ   ?øŽ â $xFC   ÿ Ž âxC   ÿ   Æ %v?   Ž â/x==   FøŽ ä x   ÿÀŽ â $xCC   ÿ Ž â xÿ   B  â�yd    ™ b %v=   Ž â xÿ   BøŽ ä xA   CøŽ Ä �yc    ™ ¢ rB-   ÿàõ â %v   Ž È %v@A   Ž â rC+   ÿä â �ye    ™ ¨  �yf@    ™ ¨  �ygBJ   • ¢  $xI   ÿ Ž Ä $xI@   ÿ Ž Æ/ x   ÿÀŽ ä xÿ   J  äxI   ÿ   ä xÿ   JøŽ ä x=ÿ   E  ä x   ÿÀŽ Ä xI   ÿ   ä x   øŽ â %v   Ž âxÿ   EøŽ ä x==   øŽ â %v   Ž â r@-&   ÿàõ â �yh    ™ ¤  %v   Ž Ä �yi    ™ ¤ %vX=   Ž ä �yj    ™ ¤  $xA+   ) ä �yXX    ™ ¨ �yk@    • ¢ xI   ÿ   â$xBI   ÿ Ž Æ x   ÿÀŽ ä xI   ÿ   äx=ÿ   G  äxÿ   GøŽ â $xII   ÿ Ž â x   ÿÀŽ ä r%(   2àµ Ä/ xQ=   BøŽ ä xÿ   H  â%v   Ž â xÿ   HøŽ ä r14   ÿd~ â %vPQ   Ž â x   IøŽ â �y=    ™ ä  %v@ Z  x Ž â r?%(   0àÙ â �yPP    ™ ä %v   Ž È $x$   ÿŽ â r-   ÿà÷ â %v   Ž â r @       â �yC    ™ " r@14   ÿ¤~ â $xA   Ž â r+   ÿäÿ â �yB    ™ h �yA   ™ b %v? Z  x Ž Æ �yO   ™ b $x@$   ÿŽ Æ/ �y@j   ‘ b $x   Ž â rF-6   ÿàõ ä �y?   ™ h �yD   ™ b $xG+   < Æ �yM   ™ h �yL   ™ h �yE   ™ h �yJ   ™ b rT-   ÿàõ Æ �yNFj   ‘ b r%(   àÇ Æ/ �yK   ™ h �yH   ™ b rU+   ÿä Æ �yI   ™ b rV14   ÿ„þ Æ �yS   ™ b r-   ÿàõ Æ �yG    ™ b %v Z  x Ž Æ �yQ   ™ b $x   + Æ �yFT    • b r[%(   *àµ Æ �yR    • h �yW    • b $xUV$   ÿŽ Æ �yV   ™ b rZ14   ÿd~ ä rU   ÿàÿ â �yT    • h �yU    ™ b %v[ Z  x Ž Æ/ �yY   ™ b $x[Z$   ÿŽ â r%(   .àµ ä �yZ   ™ b $x   [Ž â r14   ÿd~ â %v Z  x Ž â �y[   ™ h �yl   ™ h �yt   ™ h �yr   ™ b r%(   /àµ Æ �yp   ™ b $x{$   ÿŽ Æ �yn   ™ b ry14   ÿd~ Æ �yw   ™ b %v Z  x Ž â r{   ÿàÿ ä �yv   ™ b $xyy$   ÿŽ Æ �y\   ™ b $x   yŽ Æ �ys   ™ h �yq   ™ h �yo   ™ h �ym    ™ h �yu    • h �yx    ™ h �y~   ™ h �y   ™ h# �y    • h �y}   ™ h �y€   ™ h �y   ™ h �y|   ™ h �y„   ™ b x�bÿ   ÿÀŽ Æ �y{   ™ h �yƒ   ™ b r‡b�   ÿpŽ Æ �yz   ™ h �y‚   ™ b x†ÿ   b  Æ �yy   ™ h �y�   ™ h �y    ™ h �yb   ™ h �y    • h �y    ™ h! �y    • b $x…‡€   ÿ Ž âr‹††   ÿpŽ â $x‡‡   ÿ Ž Æ x……€€ÿÀŽ ä xˆ‹   ÿ   ä x†…€€€€ÿüŽ â $x‹‹   ÿ Ž â x‰ˆ€€ÿÀŽ ä x‡‡€€ÿÀŽ ä x††ÿþþþÿàÿ Ä xŠ‰€€€€ÿüŽ ä xˆ‡€€€€ÿüŽ ä x‹‹€€ÿÀŽ ä r††…   ÿüŽ ä xŠŠÿþþþÿàÿ ä xˆˆÿþþþÿàÿ Ä x‹€€€€ÿüŽ ä x†˜º  ÿ    ä rŠŠ‰   ÿüŽ ä rˆˆ‡   ÿüŽ ä x†ÿþþþÿàÿ ä xÿÀŽ Ä x€€€€^ÞŽ äO x‡Š˜º  ÿ    ä xˆˆ˜º  ÿ    ä xŠ_ÿ   ÿÀŽ ä� r††‹   ÿüŽ ä r  €ÿàÿ â $xŠ_ @  ŠŽ â x…ˆÿÀŽ Ä x‡ÿÀŽ äxˆˆ€€€€]ÞŽ äx‡‡€€€€`ÞŽ ä x††˜º  ÿ    â $xˆˆ   …
Ž â r‡‡  €ÿàÿ ä x^€€€€^´Ž Ä xŠ   ÿ   ä x†ÿÀŽ äx…ÿ   _  ä x††€€€€aÞŽ â $x_Š   ÿ Ž â x€€ÿÀŽ ä x]ˆ€€€€]´Ž â $xˆ… @  …Ž â r††  €ÿàÿ Ä x€€€€ÿüŽ ä x__€€ÿÀŽ ä xa†€€€€a´Ž ä x†ˆ   ÿ   â$xˆˆ   ÿ Ž â xÿþþþÿàÿ ä x…_€€€€ÿüŽ Ä x`‡€€€€`´Ž ä x‡†€€ÿÀŽ ä r   ÿüŽ ä x†…ÿþþþÿàÿ ä x…ˆ€€ÿÀŽ ä x‡€€€€ÿüŽ Ä x˜º  ÿ    ä r††_   ÿüŽ ä x_…€€€€ÿüŽ ä xˆÿþþþÿàÿ ä xÿÀŽ äx€€€€dÞŽ ÄxŠ_ÿþþþÿàÿ ä x††˜º  ÿ    ä r  €ÿàÿ ä rˆˆ‡   ÿüŽ ä rŠŠ…   ÿüŽ ä x_†ÿÀŽ Ä x††€€€€cÞŽ ä xd€€€€d´Ž ä xˆˆ˜º  ÿ    ä xŠ˜º  ÿ    â $x†   _
Ž â x_ˆÿÀŽ ä x†ÿÀŽ Ä xˆˆ€€€€eÞŽ ä x€€€€fÞŽ ä xŠgÿ   ÿÀŽ ä rˆˆ_  €ÿàÿ â $x   †
Ž â x_ÿ   g  ä rŠgŠ   ÿpŽ Ä xf€€€€f´Ž â $x†_ @  _Ž â xŠ   ÿ   â$x_Š   ÿ Ž â x€€€€c´Ž ä xg†   ÿ   â$x††   ÿ Ž â x€€ÿÀŽ ä xc_€€ÿÀŽ Ä x_€€€€ÿüŽ ä x…g€€ÿÀŽ ä x‰†€€ÿÀŽ ä xgc€€€€ÿüŽ ä x†_ÿþþþÿàÿ ä x‡…€€€€ÿüŽ Ä x‹‰€€€€ÿüŽ ä xeˆ€€€€e´Ž ä xˆgÿþþþÿàÿ ä r††   ÿüŽ ä xŠ‡ÿþþþÿàÿ ä xŒ‹ÿþþþÿàÿ Ä rˆˆc   ÿüŽ ä x†˜º  ÿ    ä rŠŠ…   ÿüŽ ä rŒŒ‰   ÿüŽ ä xˆˆ˜º  ÿ    ä x_€€€€hÞŽ Ä x†ÿÀŽ ä xcŠ˜º  ÿ    ä xgŒ˜º  ÿ    ä xˆ€€€€iÞŽ är__†  €ÿàÿ ä xˆˆÿÀŽ Ä x…c€€€€jÞŽ äxŠcÿÀŽ ä xcg€€€€XÞŽ äxŒgÿÀŽ ä x†kÿ   ÿÀŽ ä xh_€€€€h´Ž â $xˆ   ˆ
Ž â r……Š  €ÿàÿ Ä x_ÿ   k  â $xc   Œ
Ž â r†k†   ÿpŽ ä xj…€€€€j´Ž â $x…_ @  _Ž â x€€€€X´Ž â $x_†   ÿ Ž âxX†   ÿ   ä xg…   ÿ   â$x……   ÿ Ž â xXX€€ÿÀŽ Ä xc_€€ÿÀŽ ä xkg€€ÿÀŽ ä x_X€€€€ÿüŽ ä x‡…€€ÿÀŽ ä xgc€€€€ÿüŽ ä x…k€€€€ÿüŽ Ä x__ÿþþþÿàÿ ä xiˆ€€€€i´Ž ä x†gÿþþþÿàÿ ä xˆ…ÿþþþÿàÿ ä r__X   ÿüŽ ä r††c   ÿüŽ Ä x‰‡€€€€ÿüŽ ä rˆˆk   ÿüŽ ä xX_˜º  ÿ    ä x_†˜º  ÿ    ä xŠ‰ÿþþþÿàÿ ä xˆˆ˜º  ÿ    Ä xg_€€€€PÞŽ äx†_ÿÀŽ ä rŠŠ‡   ÿüŽ ä x_ˆ€€€€CÞŽ äxˆˆÿÀŽ â $xgg   †
Ž â xkŠ˜º  ÿ    â &r†^A   ÿ   ârŠ_ˆ  €ÿàÿ â &rˆhA   ÿ   â r@ÿ3   @  â &r†]O   †   Ä &rˆiO   ˆ   ä $tAÿ   ÿ Ž ä $x@@   ÿ Ž â xcX€€€€=ÞŽ â &r†`?   †   âxXXÿÀŽ â &rˆj?   ˆ   â x?@   AìŽ â &r@^D   ÿ   â rXcX  €ÿàÿ â &r†aM   †   Ä &rA]L   @   â xXX€€€€=´Ž â &r†dE   †   â x…k€€€€BÞŽ â &rˆM   ˆ   â xŒkÿÀŽ â &rA`J   A   â rNÿ3   N  â &r†K   †   â xPg€€€€P´Ž â &rDhD   ÿ   â xCŠ€€€€C´Ž â &rˆXE   ˆ   â xŠ        â &r@aH   A   â x=N   ÿ   â $x……   Œ
Ž Ä &r†eI   †   ä &rEiL   D   ä &rˆPK   ˆ   â x==   ŠìŽ â &rDdS   @   â xB…€€€€B´Ž â &r†fG   †   ä &rˆCI   ˆ   Ä &rDQ   D   ä $r††=   ÿŽ ä 0rFÿF    A   ä 0rAÿR    A   ä &rˆBG   ˆ   ä &rDeV   D   â Er† †       â &rEjJ   E   â  r@FA     A  â $rˆˆ?   ÿŽ Ä 0rTÿT    A   ä &rDfU   D   ä &rEH   E   â #r;@†   ;   â Erˆ ˆ       â $rD=D   ÿŽ â  r@AT     A  â 0rWÿW    A   ä &rEXS   E   â #r@ˆ      â ErD D       â  r@FW     A  â &rEPQ   E   Æ #r!@D   !   â &rECV   E   ä &r@^l   ÿ   ä &rHBU   E   ä &rE]t   @   È &rE`r   E   ä &rA^Y   ÿ   ä &rEap   E   ä $rH?H   ÿŽ ä &rEdn   E   ä &rA]Z   A   Ä &rEw   E   â ErH H       â &rA`[   A   â  rWTW     A  â &rEev   E   ä &rAa\   A   â #r WH       â &rHfx   E   ä &rAds   A   Ä &rE^   ÿ   ä &rYhY   ÿ   ä&rlhl   ÿ   ä &r@q   A   ä &rE]~   E   ä &rh   ÿ   Ä &r^^   ÿ   ä&rh   ÿ   ä &rYiZ   Y   ä&rGit   l   ä &r@eo   @   ä &rE`}   E   Ä &ri~      ä &r]]€   ^   ä&rii€      ä &rYj[   Y   ä&rGjr   G   ä &r@fm   @   Ä &rEa|   E   ä &rj}      ä &r``„   ]   ä&rjj„   i   ä &rY\   Y   ä&rGp   G   Ä $r@=@   ÿŽ ä &rEd{   E   ä &r|      ä &raaƒ   `   ä0ruÿu    A   ä &rƒ   j   â Er@ @       â &rz   E   Ä &rad‚   a   â  rAFu     A  â &rYXs   Y   ä&rGXn   G   ä&rX{      ä&rX‚      ä &rey      Ä &ra�   a   â #rA@      â &rDPq   Y   ä&rGPw   G   ä&rzPz      ä&rP�      ä $rH=H   ÿŽ â x   ÿàÿ â 0rAÿ    A   Ä &rf      ä &raeb   a   ä &rDCo   D   ä&rGCv   G   ä&rzCy   z   ä&rCb      â ErH H       â  rFA     A  â $r=   ÿŽ â |    pbô â &rff   a   Ä &rDBm   D   ä&rxBx   G   ä0rÿ    A   ä &rzB   z   ä&rB      â #r7H   7   â $rD?D   ÿŽ âEr        â $rx?x   ÿŽ â rF     A  â $rz?z   ÿŽ Ä $rf=f   ÿŽ ä 0rÿ    A   ä $r?   ÿŽ â ErD D       â #r9   9   â Erx x       â  rFF     A  â Erz z       â  r@Tu     A  âErf f       â  rATA     A  âEr        â  rT     A  â rT     A  â x--p  ÿà÷ â #r@D      â #rAx      â #r5z   5   â #rFf      â #r"   "   â x@   ÿàÿ Ä x**@   ÿàÿ ä x..@   ÿàÿ ä x//@   ÿàÿ ä x00@   ÿàÿ ä x22@   ÿàÿ ä r+ÿ+   ÿäÿ â D)       À â Gy  0Ýÿÿÿÿƒ ê Ay        € ê Ey  ð     € â G‰  Ð     € ê $x  €  Ž Ê ˆs  ; úÿ H   è ˆs  €úÿ H   è ˆs  ! ûÿ H   è ˆs   €ûÿ H   è ˆs   üÿ H   è ˆs  €üÿ H   è ˆs  7 ýÿ H   è ˆs  €ýÿ H   è ˆs  9 þÿ H   è ˆs  5€þÿ H   è ˆs   ÿÿ H   è ˆs  "€ÿÿ H   ä Ay        € ê {          ì M        € ê „y      H   " z # r  p`ð Æ „y €   H   b r ÿ   p"p Æ „y    H   ¨ „y    H   è „y €  H   ( „y	    H   h „y €  H   h „y    H   h „y €  H   è „y
    H   è „y €  H   â !r  ;       â !r       È/ ‰       " !r       ÆO ‰      b !r!       Æ� „y €  H   â !r         Æ‰      ¢ !r	       Æ‰	       â !r       Æ ‰      " !r7       â !r        ä ‰      " !r       â !r       ä/ ‰       b !r

9       Æ ‰      b !r5       â !r       äO ‰      ¢ !r	 	       Æ� ‰      â !r       Æ‰
      " !r""       Æ ‰      " !r       Æ ‰      " !r        Æ/ y       &   b !r       â!r       äO ‰      ¢ x#   ÿàÿ â !r       ä� y      '   â z  r  p`ò â !r
       Ä‰	      " x    p"ò â !r       ä ‰      b !r       Æ ‰      " $z   x  #Ž Æ/ ‰      b $v#ÿ r  ÿ Ž Æ ‰      b !r       âO $x'#   ÿŽ ä ‰      ¢ $z  ~   Ž ä� $x)#   ÿŽ â ‰      â !r	       â
  r  ÿà÷ Ä ‰      " !r       Æ‰!"      b !r       Æ ‰ €    " !r       â/ ÿÿ   ÿäÿ ä ‰ €    b !r       â 
 j  ÿ† â!r       âO ‰ €    ä 
 k  � â !r       â� ‰ €    ¢ !r       Æ‰ €    â !r!"!       Æ‰ €    " !r
       Æ ‰ €    "  '   ÿàõ â !r       ä/ ‰ €    b r#    ÿˆ ä ÿÿ   ÿä â ‰ €    b 
 j  ÿ„ â $rÿÿ   ÿ â z j  ÿˆ â ‰  €    b !r       âO 
 k   Ä ‰" €    ¢ z k   â !r       â !r       â� ‰$!      â !r       âr#    ÿ„ ä ‰
 @    " !r       â  )   ÿàù â $r(ÿÿ   ÿ â ‰	 @    " !r       â/ &ÿÿ   ÿä â !r       â‰ @    b 
 j  ÿˆ â !r         Ä ‰ @    b !r""       ÆO ‰ @    ¢ !r$!$       Æ� ‰ @    â !r
       Æ‰ @    " #   ÿàÿ â !r		       ä ‰ @    " *    ÿà÷ ä z
 j  ÿ„ â‰ @    " !r       â/ $+ÿÿ   ÿŽ â * j  ÿ† â ‰!  @    b !r       âz k  ( Ä ‰#" @    b !r       âO 
 k  & ä ‰%$ €    ¢ !r       â� * k  +� ä ‰	      â !r       Æ‰      " !r       Æ ‰      " !r       â !r! !       ä/ ‰      b !r#"#       Æ‰      b !r%$%       ÆO ‰      ¢ !r+	       Æ� ‰      â 	'   ÿàÿ â !r-       ä‰      " x'        ä  	   ÿàõ â‰      " !r       â %v  j  ' Ž Ä ‰ !      " !r       â/ $&ÿÿ   ÿ â  j  ÿ„ â ‰$% @    b !r       ä k  & â ‰"#      b !r       âO !r       Ä� †	 +    â !r       Æ† -   â !r       Æ †	     â !r!!        Æ †     â !r$%$       Æ/ †	     â !r##"       Æ†    è †	     è †     è †	 
    è † 
!   è ‰$      ( †	 #    â M™        € ê x)   ÿàÿ âO !r$       Æ r    ÿàñ È r ÿÿ   ÿä  ä z j  ÿ€ È z k     Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      "   " ¹z  l      â Ey  `%    € â ™x?    â y      !   b ¹z  F   
   â r  ÿ       â $r%ÿÿ   ÿ Ž â y$      %   ¢ $r7ÿÿ   ÿ Ž â r; ÿ       â x"      ÿ  â $r9ÿÿ   ÿ Ž Ä $rNÿÿ   ÿ Ž ä $rÿÿ   ÿ Ž â r ÿ   pRð är ÿ   pRò â$x'    Ž ä/ $x$$   ÿ Ž ÆO x ÿ   '  È r '   ÿ� ä r  ÿ       ä x!ÿ     È | !   pbô Ú G)  ð#    € ê y	      '   " r ÿ       â $r
ÿÿ   ÿ Ž â xøÿÿÿÿÀŽ â y      &   b $z$ p  !Ž â z-  Y      â $x2!   ÿ Ž ä $x''   
Ž ä $vÿ X  ÿ Ž â z12 q  ÿàÿ â $x'   ÿ Ž âx'   ÿ   â$r ÿÿ   ÿ Ž â r<''   ÿ� Ä rÿ   ÿ3Ž ä x<ÿ   <  ä x3ÿ   '  â %x   ÿŽ â x=ÿ   <  ä x5   ÿÀŽ â 'z	 y   Ž È 'z s  
 Ž â/ zÿ z    Æ $z*	 }  ÿŽ â zÿ t    â $z |  !Ž ä$z |  Ž â x$   ÿàÿ â $z v  Ž ä$z v  Ž ä $tÿn   ÿ Ž Ä $z p  Ž ä %xn   ÿŽ È %v  € Ž â x&J   áÙ Æ $z w  ÿŽ â r(  €ÿà÷ â %x   Ž â r)ÿ   ¥~ ä r**   ÿàõ â $x+   Ž â z  q      â %xn   ÿŽ â r8'   ÿàù är6<   ÿà÷ â $r4ÿÿ   ÿ â r:'   ÿàõ â $x,   2Ž âr02   ÿŽ â$x/   2Ž Ä %xn   <Ž â r<=   ÿäÿ Æ %x
   Ž È %x   Ž È %x   Ž È $x>   3 ä$x.   3 ä r8   ÿàõ È r->   ÿä ä r   ÿàõ Æ �yDB   ‘ ¤ $x	   - Ê �y=   • è �y?   • ( �yB   • h �yC   • b r:   ÿàõ Ê $x-   . â r   ÿàõ È �yB   ‘ b r-   ÿä Ê �yO   • h �y   • h �y   • h �y   • b r@
   ÿàõ Ê $xA   - Ê �y@@J   • b x        â $xD   ÿ Ž äO$xD@   ÿ Ž ä$xD€   ÿ Ž â x   ÿÀŽ ä xF   ÿÀŽ ä xÿ   =øŽ Ä� x=ÿ   =  ä xMFÿ   ?øŽ â%v   Ž â/ r&   ÿàõ ä xK=   øŽ â %vLM   Ž â�y=    ™ ¢  r-)   ÿä ä $xD    ÿ Ž â x?ÿ   ?  â �yLL    ™ â %vJK   Ž Æ �yE    • ¢ x??   øŽ Æ �yJJ    ™ ¤ %vH?   Ž È $xD   ÿ Ž ä �yHH    ™ ¦  x   ÿÀŽ â $xD   ÿ Ž âxD   ÿ   ä xÿ   BøŽ äxBÿ   B  Æ %v   Ž âxB   øŽ ä xF   ÿÀŽ ä �yB    ™ b	 %v   Ž âxGFÿ   CøŽ är   ÿàõ ä xCÿ   C  â �yK    ™ ¢" xDD   ÿ   â %vFG   Ž È $x   - â xC   DøŽ ä �yCF    ™ ¨  �yAJ   • ¢  %v   Ž Ê�y?    ™ ¢ $x   ÿ Ž Ê xD   ÿÀŽ È xQDÿ   OøŽ äxOÿ   O  ä x   ÿ   â/%vPQ   Ž â r(   ÿàõ äxOO   øŽ â $x@   ÿ Ž â�yIP    ™ " $x-   + â xÿ     â$x    ÿ Ž â xD   ÿÀŽ â %vRO   Ž â �yG    • ä xDÿ   øŽ Ä x   øŽ â �yMR    ™ $ %v   Ž È %v   Ž â �yF    ™ " x   ÿ   Æ �yD    ™ " $xT   ÿ Ž â xÿ     ä x   ÿÀŽ ä rQ'*   2àµ ä xO   TøŽ ä xÿ   øŽ â $tPÿ$   ÿ Ž â rR34   ÿd~ Æ %v   Ž â x   ÿ   Æ/ %vQ Z  P Ž ä �yQ    ™ $ $xR$   ÿŽ â x   ÿÀŽ â %vO   Ž È $x   Ž â xÿ   øŽ â$x   ÿ Ž â xÿ     â �yO    ™ ( �yR   ™ " x   øŽ â %v   Ž Ä �yT   ™ " r6   ÿàõ â %v   Ž ä �yS   ™ " r-<   ÿä ä rY'*   1àµ â �yU    ™ ( �yW   ™ " r[34   ÿd~ Æ �yX   ™ " r   ÿàû â/$x[[$   ÿŽ ä �y    ™ ( �yVj   ‘ " r   ÿàù Æ �y^    • ( �y   ™ "! %vY Z  P Ž â r_'*   ,àµ ä �yY   ™ " $x   [Ž ä $x   -Ž â�y[    ™ " $x   - â r`34   ÿd~ Ä �ya   ™ ( �yZj   ‘ ( �y\    • ( �yh   ™ " %v_ Z  P Ž Æ �y_   ™ " $x`$   ÿŽ â/ r        ä �y`   ™ " $x   Ž Æ �y   ™ " r   ÿàû Æ �y   ™ " r-   ÿäÿ Æ �yi   ™ " r'*   /àµ Æ �yf   ™ ( �yg   ™ ( �yd   ™ ( �y]    • ( �ye   ™ ( �yb    ™ " r34   ÿd~ Ä r'*   0àµ â �yj    • & ro34   ÿd~ â %v Z  P Ž â�yc   ™ & $xk$   ÿŽ â �ym   ™ " %v Z  P Ž Æ �yl   ™ " $xo$   ÿŽ ä/ $x   kŽ â �yn    • " $x   Ž Æ �y   ™ ( �yk    ™ ( �yo   ™ ( �yv   ™ ( �y   ™ ( �yP   ™ ( �yu   ™ ( �y   ™ ( �ys   ™ ( �yt   ™ ( �yq   ™ ( �yr   ™ ( �yz   ™ ( �yp   ™ ( �yy   ™ " x}@ÿ   ÿÀŽ Æ �y    ™ ( �yx    ™ " r}@}   ÿpŽ ä x{ÿ   @  â �yw    • ( �y@    • " $x|}€   ÿ Ž Ä $x{{ @  {Ž Æ x||€€ÿÀŽ â $x{€   ÿ Ž â/ x}}   ÿ   ä x|€€€€ÿüŽ ä x€€ÿÀŽ ä x}}€€ÿÀŽ ä xÿþþþÿàÿ â $x{   ÿ Ž â x€€€€ÿüŽ Ä x{}€€€€ÿüŽ ä r||   ÿüŽ ä xÿþþþÿàÿ ä x{{ÿþþþÿàÿ ä x||˜º  ÿ    ä r   ÿüŽ Ä r}{}   ÿüŽ ä x€€ÿÀŽ ä x|ÿÀŽ äx{|€€€€=ÞŽ äO x˜º  ÿ    ä x€€€€ÿüŽ Ä r|{  €ÿàÿ ä xÿÀŽ äx{€€€€LÞŽ ä� xEÿ   ÿÀŽ ä xÿþþþÿàÿ ä x}}˜º  ÿ    Ä rE   ÿpŽ ä r   ÿüŽ â $x{{   
Ž â x}ÿÀŽ â $x€   ÿ Ž â x}}€€€€JÞŽ ä x˜º  ÿ    ä xEÿ   E  â $x}}   
Ž â xL{€€€€L´Ž â $x{   ÿ Ž â xÿÀŽ Ä x€€ÿÀŽ ä x€€€€HÞŽ â $xE @  EŽ â xE€€€€ÿüŽ ä r  €ÿàÿ ä x{€€ÿÀŽ â $x{€   ÿ Ž â xEEÿþþþÿàÿ Ä xH€€€€H´Ž ä x€€€€ÿüŽ ä x{{€€ÿÀŽ ä rEE   ÿüŽ ä xÿþþþÿàÿ ä x=|€€€€=´Ž Ä xJ}€€€€J´Ž ä x   ÿ   ä x|{€€€€ÿüŽ ä x}E˜º  ÿ    ä r   ÿüŽ ä x€€ÿÀŽ Ä x||ÿþþþÿàÿ ä x}ÿÀŽ äx}€€€€BÞŽ äx˜º  ÿ    ä xE€€€€ÿüŽ ä r||{   ÿüŽ â $x{   
Ž â xÿÀŽ Ä x€€€€KÞŽ ä xEEÿþþþÿàÿ ä x||˜º  ÿ    â $x   
Ž â rEE   ÿüŽ ä x|ÿÀŽ äx|€€€€CÞŽ Ä x{€€€€B´Ž ä x{Aÿ   ÿÀŽ â $x   
Ž â xBE˜º  ÿ    ä xK€€€€K´Ž ä rA{   ÿpŽ ä xBÿÀŽ Ä xBB€€€€?ÞŽ â $xE€   ÿ Ž â xC€€€€C´Ž ä xÿ   A  ä rBB  €ÿàÿ â $x   ÿ Ž â xAE€€ÿÀŽ â $x{ @  Ž â xB€€€€?´Ž Ä xA€€€€ÿüŽ ä x?€€ÿÀŽ ä x{   ÿ   â$x{{   ÿ Ž â xBÿþþþÿàÿ ä xE?€€€€ÿüŽ ä x€€ÿÀŽ Ä x{€€ÿÀŽ ä rBBA   ÿüŽ ä x|Eÿþþþÿàÿ ä xA€€€€ÿüŽ ä xBB˜º  ÿ    ä xE€€€€ÿüŽ Ä r?|?   ÿüŽ ä x|Aÿþþþÿàÿ ä xAB€€€€IÞŽ äxEEÿþþþÿàÿ ä xBBÿÀŽ ä rEE   ÿüŽ Ä r|   ÿüŽ â $xA   B
Ž â x|?˜º  ÿ    ä xBGÿ   ÿÀŽ ä x{E˜º  ÿ    ä x?|€€€€MÞŽ äxE|ÿÀŽ Ä rBGB   ÿpŽ ä r??E  €ÿàÿ ä xGÿ   G  â $xEB€   ÿ Ž â x}˜º  ÿ    ä xI€€€€I´Ž ä xA{€€€€FÞŽ Ä x?€€€€M´Ž â $x?G @  GŽ â x{{ÿÀŽ ä x}€€€€DÞŽ äx|}ÿÀŽ ä xGB   ÿ   ä xBE€€ÿÀŽ â $xAA   {
Ž Ä $x   |
Ž â xEB€€€€ÿüŽ â $xM?€   ÿ Ž â xFA€€€€F´Ž ä xEEÿþþþÿàÿ ä xAG€€ÿÀŽ ä xD€€€€D´Ž ä xM€€ÿÀŽ Ä rEEB   ÿüŽ ä xGA€€€€ÿüŽ ä xM€€€€ÿüŽ ä xBE˜º  ÿ    ä x|Gÿþþþÿàÿ ä xMMÿþþþÿàÿ Ä xEB€€€€QÞŽ ä r||A   ÿüŽ ä xBBÿÀŽ ä rM   ÿüŽ â $x??   ÿ Ž â xA|˜º  ÿ    ä rBEB  €ÿàÿ Ä xE˜º  ÿ    â &r=R   ÿ   â x|A€€€€OÞŽ äxAAÿÀŽ â &rJT      â x??€€ÿÀŽ Æ $x||   A
Ž â xG?€€€€ÿüŽ â &rLS      ä &rAIR   ÿ   â xRE€€€€UÞŽ â&rHW      â xEEÿÀŽ ä xGGÿþþþÿàÿ â &rAT   A   â xQB€€€€Q´Ž â $xRR   E
Ž â r?G?   ÿüŽ â &rBX      ârVÿ5   V  â &rAFS   A   â x??˜º  ÿ    â &rBK   B   â xUR€€€€U´Ž â $tRÿ   ÿ Ž â xV   ÿ   â &rADW   A   â xE?€€€€ÞŽ â &rBCY   B   â xT?ÿÀŽ â &rXQX   A   â x   RìŽ â &rB[   B   â rEET  €ÿàÿ â &rA=a   ÿ   Ä $rBB   ÿŽ â xE€€€€´Ž â 0r\ÿ\    A   ä 0r^ÿ^    A   ä &rEJ_   A   â ErB B       ä  r?\^     A  â &rEL`   E   Æ #r?B      â &rEH   E   ä &r?=   ÿ   ä&rI   ÿ   ä &rEh   E   ä &ri      ä &rEKf   E   Ä &rFg      ä &rECd   E   ä &raIa   ÿ   ä &rRDe      ä &rb   E   ä &ra_   a   Ä $r   ÿŽ ä &raF`   a   ä 0rjÿj    A   â Er        â &raD   a   ä &r?Ji   ?   â  r\j     A  Æ &r?Lg   ?   â #rN   N   â &r=   ÿ   ä&rI   ÿ   ä &r==   ÿ   ä&rII   ÿ   ä &rBHe   ?   Ä &r?Jo      ä&roo      ä &r=Jv   =   ä&rIv   I   ä &r?LP   ?   ä&roFP   o   Ä &r=Lu   =   ä&rIFu   I   ä &r?H   ?   ä&roD   o   ä &rHHs   =   ä&rDDs   I   â xO|€€€€O´Ž â &raQh   a   Ä &rBc   B   ä&rRQc   R   ä &r?t   ?   ä&roQt   o   â rZÿ5   Z  â &rHq   H   ä&rDQq   D   Ä &rXO   X   ä&raOf   a   ä &rBKm   B   ä&rROm   R   ä &r?Kr   ?   ä&roOr   o   Ä &rHKz   H   ä&rDOz   D   â x!!   ÿàÿ â $tTÿ   ÿ Ž ä $xZ   ÿ Ž ä &rXUY   X   ä&raUd   a   Ä &rBCl   B   ä&rRUl   R   ä &r?Cp   ?   ä&roUp   o   ä &rCCy   H   ä&rUUy   D   â x   TìŽ â &rX[   X   â| !   pbô â &rbb   a   Ä &rBk   B   ä&rRk   R   ä &r   ?   ä&r   o   ä &rx   C   ä&rx   U   Ä $rBB   ÿŽ ä$r   ÿŽ ä 0r]ÿ]    A   ä $rXX   ÿŽ ä $rbb   ÿŽ ä 0rnÿn    A   Ä $rRR   ÿŽ ä0rwÿw    A   ä $r   ÿŽ ä $r   ÿŽ ä $r   ÿŽ ä 0r@ÿ@    A   â Er        â  rA^]     A  â ErX X       â  rj]j     A  âErb b       â  r\n     A  âErB B       â  r]n     A  âErR R       â  r\w     A  â r]w     A  âEr        â  r\\@     A  âEr        â  r]]@     A  â Er        Ä xp  ÿà÷ â #r AX       â #r#jb   #   â #r9B   9   â #r"R   "   â #r;   ;   â #r7   7   â #r%\   %   â #r ]       â x,,@   ÿàÿ â $r-ÿÿ   -Ž â x//@   ÿàÿ Ä x00@   ÿàÿ ä x11@   ÿàÿ ä x22@   ÿàÿ â D)       À â Gy  °ßÿÿÿÿƒ ê Ay        € ê Ey  Ð     € â G‰  °     € ê $x@  Ž Ê ˆs  ûÿ H   è ˆs  €ûÿ H   è ˆs N üÿ H   è ˆs #€üÿ H   è ˆs 9 ýÿ H   è ˆs "€ýÿ H   è ˆs ; þÿ H   è ˆs 7€þÿ H   è ˆs % ÿÿ H   è ˆs  €ÿÿ H   ä Ay        € ê {          ì M        € ê „y     H   b z $ r  p`ð Ä x'        â „y €   H   ¢ r ÿ   p"p Æ „y    H   è „y €  H   ( „y    H   h „y	 €  H   ( „y    H   ( „y
 €  H   ( „y    H   è „y €  H   â !r       Æ/ y      &   b !r        ÆO ‰      ¢ !rN       Æ� ‰       " !r##       Æy      '   â !r99       Æ‰      " !r		"       Æ ‰#      b !r;;       Æ ‰9      b !r7
7       â !r       äO ‰
	      ¢ !r%%       â !r        ä ‰      " !r         Æ ‰;      â !r       Æ‰7      " !r#       Æ‰%      b z#  r      Æ ‰       b !r9       â/ $x!#   ÿŽ â !r
	
       âO ‰      b !r       Æ ‰	      " !r;       Æ� ‰      ¢ !r7       Æ‰  €    â !r%       Æ‰      " !r         Æ ‰      h ‰
      b !r       Æ/ ‰      b !r		       Æ ‰      " !r       ÆO ‰       ¢ !r        â� $z  x  $Ž â !r       â‰ €    ä $z  ~   Ž â !r       â‰ €    " !r
       Æ ‰	 €    b x
$   ÿàÿ â !r       ä/ ‰ @    b z 
 r  p`ò â !r       ä ‰ €    " x    p"ò â !r        ÄO ‰ €    ¨ ‰ €    ¢ !r       â�  !   ÿàõ ä ‰ €    â !r       â ÿÿ   ÿä Ä ‰ €    " !r	       â
 j  ÿ„ ä ‰ €    b !r
       â/ r#    ÿ† â!r       â 
 k    ä ‰
      " $rÿÿ   ÿŽ â !r       âO z j  ÿ† â ‰ @    b 
	  r  ÿàù â !r       â z k  � â ‰ @    ¢ r#    ÿŠ â !r       â� $$ÿÿ   ÿ â ‰ @    â 
	 j  ÿˆ â !r       â$r"ÿÿ   ÿŽ â ‰ @    " z j  ÿŠ â !r       â
		 k  $ â ‰ @    b z k  "� Æ ‰ @    b !r%
       â #   ÿàÿ ä ‰ @    " !r       Æ/ ‰ @    b !r       âO !   ÿàÿ ä ‰ @    ¢ !r       â� !    ÿàõ â%v
  j  ' Ž â !r       â    ÿà÷ â ‰      â $"ÿÿ   ÿ â !r       â! j  ÿ„ â ‰      "  ÿÿ   ÿäÿ â !r       â  j  ÿ† â ‰      b !r       Ä  k   � â ‰      " !r       Æ/ ‰      b !r       ÆO ‰      ¨ ‰       ¢ !r#       â� ! k  " ä ‰      â !r       Æ‰      " !r       â!r       ä †	 
%    â !r       Æ/ † 
#   â !r       ÆO †	     â !r        Æ †     â !r       Æ� †	     â !r       Æ†    è †	     è †     è †	     â M™        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      "   " ¹z  l      â Ey   "    € â ™x?    â y      !   b ¹z  F   
   â $r)ÿÿ   ÿ Ž â r3 ÿ       â $r"ÿÿ   ÿ Ž â x      ÿ  â $r5ÿÿ   ÿ Ž â r ÿ       Ä r ÿ   pRð är ÿ   pRò â$x#    Ž Ê/ x ÿ   #  È r #   ÿ� â $r ÿÿ   ÿ Ž Æ x7ÿ     È | 7   pbô Ú G)  p!    € ê y	      '   " $rÿÿ   ÿ Ž â xøÿÿÿÿÀŽ â $r
ÿÿ   ÿ Ž â y      %   b $x.7   ÿ Ž â r##  €ÿàÿ â $v!ÿ X  ÿ Ž â y      &   ¢ $v+ÿ Y  ÿ Ž â r8##   ÿ� â$x#   ÿ Ž âx/ÿ   #  â $x#   ÿ Ž â x8ÿ   8  Ä rÿ   ÿ3Ž ä x9ÿ   8  ä x1   ÿÀŽ ä r) ÿ       â 'z	 y   Ž â z-. q  ÿàÿ Æ $x   ÿ Ž â/ zÿ z    â $z(	 }  ÿŽ ä 'z s  
 Ž ÈO $z p  7Ž â zÿ t    â $z |  7Ž ä$z |  Ž â x   ÿàÿ â $z v  Ž ä$z v  Ž ä $tÿn   ÿ Ž Ä $z p  Ž ä %x   ÿŽ È %v  € Ž È %xn   ÿŽ â r&  €ÿà÷ Æ $z w  ÿŽ â x$J   áÙ â$x'   Ž â z  q      â %x   Ž â r((   ÿàõ ä r%ÿ   ¥~ â %xn   ÿŽ â r4#   ÿàù är28   ÿà÷ â $r0ÿÿ   ÿ â r6#   ÿàõ â $x   .Ž Ä $x,   .Ž ä %xn   8Ž â r89   ÿäÿ Æ %x
   Ž È %x   Ž È %x   Ž È $x:   / ä$x*   / ä r!4   ÿàõ Ê $x+   : â r!   ÿàõ È �yCB   ‘ ¢  r	+   ÿä Ê �y9   • è �y   • ( �y;   • h �yB   • b r>!6   ÿà÷ Ä r!   ÿàõ Æ $x?+   *Ž ä $x   + Æ �y>B   ‘ h �yL   • h �y   • h �y   • h �y   • b r<
!   ÿàõ Ê $x=   + Ê �y <J   • b rH!$   ÿàõ Ê $xI+   % Ê �yHH    • b rR#(   -àë â $tgÿ$   ÿ Ž â xC   ÿ   âO$xC@   ÿ Ž Æ/ x   ÿÀŽ â $tÿ   ÿ Ž â x   ÿÀŽ ä xKÿ   9øŽ â� $xC€   ÿ Ž â xGÿ   øŽ äx9ÿ   9  â %vJK   Ž Æ x99   øŽ Æ �yJJ    ™ ¢  %vFG   Ž âxÿ     ä xC   ÿ   â %vD9   Ž ä �yFF    ™ â x   øŽ Æ �yDD    ™ ¤ %v   Ž Ê �y9    ™ ¢  $xC   ÿ Ž â x<ÿ   ;  Èx   ÿÀŽ â $xC   ÿ Ž Æ xÿ   ;øŽ ä x;C   ÿ   ä x<   øŽ â %v   Ž â x;;   ÿÀŽ Æ %v@   Ž âxÿ   B  â�y>    ™ ¢
 x;ÿ   BøŽ â $xC   ÿ Ž â rB!   ÿàõ â �y?@    ™ ¦  x   øŽ â %v   Ž È $xC   + â �yK    ™ ¨  �yBBJ   • ¢  %v<   Ž Ì �y<<    ™ â  $x   ÿ Ž âxÿ   L  â$x;@   ÿ Ž âx   ÿ   ä x   ÿÀŽ ä x   øŽ ä xÿ   LøŽ â $x=    ÿ Ž â x;;   ÿÀŽ Ä xÿ     â%v   Ž â x;;ÿ   øŽ ä x=   =øŽ ä rN!&   ÿàõ â %v   Ž â�yI    ™ f %v@;   Ž â�yG    ™ f# %vL=   Ž â �yE@    ™ f	$xO+   ' â �y;L    ™ h �y=N    • b	 $x   ÿ Ž Ä $xP   ÿ Ž Æ x   ÿÀŽ ä xPP   ÿÀŽ â $x   ÿ Ž â/ xCÿ     äxÿ   øŽ ä xAPÿ   øŽ äxÿ     Ä x   ÿ   ä r!2   ÿàõ ä xCC   øŽ ä xQ   øŽ â %v   Ž â r+8   ÿä Æ %vC   Ž ârN!   ÿàù â �y@j   ‘ $ %vA   Ž ä�yC    ™ $ %vPQ   Ž â r#(   .àµ â �yA    ™ & r/0   ÿd~ â$xO   + â r/0   ÿÄþ â %v Z  g Ž â �yM    ™ & $x$   ÿŽ â �yLP    ™ " $x$   ÿŽ Ä/ $x   Ž â �yNNj   ‘ " %vR Z  g Ž Æ �y\   ™ $ r   ÿàÿ ä �yP   ™ ( �yO   ™ (! �yQ   ™ ( �yR   ™ ( �yZ   ™ ( �y[   ™ ( �yX   ™ ( �yY   ™ " r!   ÿàõ Æ �yV   ™ " r!   ÿà÷ Æ �yW   ™ ( �yT   ™ " $x   + Æ �yU   ™ " $x   +Ž â r#(   àµ ä�y_    ™ ( �yS   ™ " rb#(   ,àÙ ä rd/0   ÿd~ â�y]    • ( �y^    ™ ( �y`    • " rc/0   ÿ¤~ â %v Z  g Ž Ä �ya    • $ $xed$   ÿŽ ä %vb Z  g Ž â �y    • & $xcc$   ÿŽ ä $x   eŽ Æ rc   ÿàÿ ä �y   ™ (# �y   ™ ( �yl   ™ ( �yk   ™ ( �yj   ™ ( �yi   ™ ( �yh   ™ ( �yg   ™ ( �yf   ™ ( �ye   ™ ( �yd   ™ ( �yc   ™ ( �yb   ™ ( �y    ™ ( �y   ™ ( �y    ™ ( �ym    • ( �yn    • " xo ÿ   ÿÀŽ Ê $xp  @  oŽ â xoÿ      Æ $x p€   ÿ Ž âxpp   ÿ   â $xro @  oŽ Æ x  €€ÿÀŽ â $xqr€   ÿ Ž â xpp€€ÿÀŽ ä xo €€€€ÿüŽ â $xr   ÿ Ž â/ xqq€€ÿÀŽ ä xooÿþþþÿàÿ ä xp€€€€ÿüŽ Ä x€€ÿÀŽ ä roo    ÿüŽ ä xq€€€€ÿüŽ ä xÿþþþÿàÿ ä x€€€€ÿüŽ ä xoo˜º  ÿ    Ä xÿþþþÿàÿ ä x ÿþþþÿàÿ ä xoÿÀŽ äxoo€€€€JÞŽ äO rq   ÿüŽ ä r     ÿüŽ â $xoo   
Ž â x˜º  ÿ    Ä rp   ÿüŽ ä xÿÀŽ äx€€€€FÞŽ ä� xJo€€€€J´Ž ä x˜º  ÿ    ä xoHÿ   ÿÀŽ â $x   
Ž â xÿÀŽ Ä x€€€€DÞŽ â $xoH @  oŽ â x  ˜º  ÿ    ä r  €ÿàÿ ä xF€€€€F´Ž ä x ÿÀŽ ä xo   ÿ   Ä xHÿ   H  â $xo   ÿ Ž â x  €€€€9ÞŽ ä x€€ÿÀŽ ä rpHH   ÿpŽ â $x     
Ž â x€€€€D´Ž ä xD€€ÿÀŽ â $xHp€   ÿ Ž â x€€€€ÿüŽ â $xpp   ÿ Ž â x9 €€€€9´Ž Ä x D€€€€ÿüŽ ä xÿþþþÿàÿ ä xHH€€ÿÀŽ ä xpp€€ÿÀŽ ä xo ÿþþþÿàÿ ä r   ÿüŽ Ä xH€€€€ÿüŽ ä xp€€€€ÿüŽ ä rooD   ÿüŽ ä x˜º  ÿ    ä xqÿþþþÿàÿ ä xsÿþþþÿàÿ Ä xo˜º  ÿ    ä xÿÀŽ äx€€€€>ÞŽ ä rqqH   ÿüŽ ä xoÿÀŽ äx€€€€?ÞŽ Ä r  €ÿàÿ ä xq˜º  ÿ    â $x   o
Ž â x>€€€€>´Ž ä xÿÀŽ äx€€€€KÞŽ ä xoBÿ   ÿÀŽ Ä x?€€€€?´Ž â $x   
Ž â rssp   ÿüŽ â $xoB @  oŽ â xÿ   B  ä xss˜º  ÿ    Æ $xH @  Ž â xo   ÿ   â$xo   ÿ Ž â x sÿÀŽ äxss€€€€<ÞŽ â $xBH€   ÿ Ž â x€€ÿÀŽ ä rss   €ÿàÿ ä xHH   ÿ   Ä x €€ÿÀŽ ä x€€€€ÿüŽ ä xDB€€ÿÀŽ ä xpH€€ÿÀŽ ä xÿþþþÿàÿ ä xB €€€€ÿüŽ Ä xHD€€€€ÿüŽ ä xqp€€€€ÿüŽ ä x€€€€K´Ž ä r   ÿüŽ ä xKBÿþþþÿàÿ ä xoHÿþþþÿàÿ Ä xqqÿþþþÿàÿ ä x˜º  ÿ    ä rKK    ÿüŽ ä rooD   ÿüŽ ä rqqp   ÿüŽ ä x€€€€IÞŽ Äx K˜º  ÿ    ä xÿÀŽ ä xDo˜º  ÿ    ä xHq˜º  ÿ    â $x   
Ž â xB €€€€GÞŽ äxK ÿÀŽ Ä x D€€€€EÞŽ äxoDÿÀŽ ä xDH€€€€;ÞŽ äxqHÿÀŽ ä xH=ÿ   ÿÀŽ ä r  o  €ÿàÿ Ä xI€€€€I´Ž â $xH= @  HŽ â xÿ   =  â $xDD   q
Ž â xE €€€€E´Ž â $xH   ÿ Ž â r    ÿpŽ â $xBB   K
Ž â x;D€€€€;´Ž â $xH€   ÿ Ž â x=    ÿ   â$xD    ÿ Ž â x €€ÿÀŽ Ä xGB€€€€G´Ž ä x€€ÿÀŽ ä xB=€€ÿÀŽ ä xHD€€ÿÀŽ ä x= €€€€ÿüŽ ä x€€€€ÿüŽ Ä xDB€€€€ÿüŽ ä xoH€€€€ÿüŽ ä x==ÿþþþÿàÿ ä xÿþþþÿàÿ ä xKDÿþþþÿàÿ ä xooÿþþþÿàÿ Ä r==    ÿüŽ ä r   ÿüŽ ä rKKB   ÿüŽ ä rooH   ÿüŽ ä x =˜º  ÿ    ä x˜º  ÿ    Ä xBK˜º  ÿ    ä xoo˜º  ÿ    ä x= €€€€CÞŽ äx  ÿÀŽ ä x€€€€AÞŽ ä xDB€€€€MÞŽ Ä xKBÿÀŽ ä xÿÀŽ ä xBo€€€€LÞŽ äxooÿÀŽ ä r =   €ÿàÿ ä r@ÿ1   @  Ä rNÿ1   N  â $x   
Ž â xC €€€€C´Ž â $t=ÿ   ÿ Ž â x N   ÿ   â $xoB   o
Ž â xB        â $x@   ÿ Ž ä &r@JP   ÿ   â x     =ìŽ Ä x   BìŽ â &r=O   @   ä&rBJQ   ÿ   ä &r=F\   =   ä&rBR   B   ä &r=9Z   =   ä &rBF[   B   Ä &r=>X   =   ä&rB9Y   B   ä &r=?V   =   ä &rB>W   B   â x<s€€€€<´Ž â $xDD   K
Ž ä &r=T   =   Ä &rB?U   B   â xA€€€€A´Ž â &r@<_   =   â xD€€€€M´Ž â &rMS   B   ä $r@@   ÿŽ ä &rB<^   M   ä 0r]ÿ]    A   Ä 0r`ÿ`    A   â Er@ @       â $rBB   ÿŽ ä 0raÿa    A   â  r=]`     A  ä ErB B       ä #r=@      â  r=]a     A  È #r=B      â &r=J   ÿ   ä&rI   ÿ   ä &rJJ   ÿ   ä &r=l   =   ä &rPIP   ÿ   ä&rQIQ   ÿ   Ä &rI   ÿ   ä &rlGl      ä&rFk   =   ä &rPGO   P   ä&rRGR   Q   ä &rj   J   Ä &rGj      ä &r9i      ä &rPE\   P   ä&rRE[   R   ä&rlEk   l   ä &rFh      Ä &rEh      ä &r>g      ä &rP;Z   P   ä&rR;Y   R   ä&rl;i   l   ä &r9f      Ä &r;f      ä &r?e      ä &rPAX   P   ä&rRAW   R   ä&rlAg   l   ä &r>d      Ä &rAd      ä &rc      ä &rKCV   P   ä&rRCU   R   ä&rlCe   l   ä &r?b      Ä &rCb      â xLo€€€€L´Ž â &r<      â x77   ÿàÿ â &rKT   K   ä&rSS   R   ä&rlc   l   ä &r      Ä &r      â | 7   pbô â $r   ÿŽ ä &rKL_   K   ä&rSL^   S   ä&rL   l   ä &r<<      Ä 0rmÿm    A   ä &rL      â Er        â 0rÿ    A   â  r]m     A  â $rKK    ÿŽ ä $rS S   ÿŽ ä$r    ÿŽ Ä $r<<   ÿŽ ä $r    ÿŽ ä 0rnÿn    A   â  r``     A  â ErK K       â  raa     A  âErS S       â  rmm     A  â Er        â #r5   5   â Er< <       â  r]n     A  âEr        â  rn     A  â x!!à  ÿà÷ â #r`K      â #r aS       â #r"m   "   â #r3<   3   â #r)   )   â x€   ÿàÿ â $r+ÿÿ   +Ž â x,,€   ÿàÿ Ä x--€   ÿàÿ ä x..€   ÿàÿ â D)       À â Gy  @âÿÿÿÿƒ ê Ay        € ê Ey  °     € â G‰  �     € ê $x   Ž Ê ˆs  üÿ H   è ˆs €üÿ H   è ˆs  ýÿ H   è ˆs  €ýÿ H   è ˆs 5 þÿ H   è ˆs "€þÿ H   è ˆs 3 ÿÿ H   è ˆs )€ÿÿ H   ä Ay        € ê {          ì M        € ê „y €  H   h „y
    H   ¨ „y     H   è „y €  H   ( „y    H   h „y €   H   ( „y    H   ( „y    H   ( „y €	  H   ( „y €  H   ( „y    H   " !r        Æ/ „y €  H   b !r

5       ÆO „y    H   ¢ !r       Æ� „y    H   â !r       Æ„y	 €  H   " !r

       Æ„y €  H   b !r       â $vÿ r  ÿ Ž ä „y  	  H   " !r       Æ „y    H   " !r       Æ „y  
  H   " !r       Æ „y €
  H   " !r       Æ „y    H   " !r       Æ „y €  H   â !r"       Æ/ „y €  H   â !r       ÆO „y €  H   â !r3       â� !r	       ä‰      b !r       Æ‰	      ¢ !r       Æ ‰      " !r       Æ ‰      â !r

       Æ y      %   " !r       Æ ‰
      b !r       Æ ‰      " !r       â/ !r)       ä ‰      b !r		       âO !r       â !r       â ‰       $ !r       â !r       â� ‰      ¨ ‰	      â !r
       Æ‰      b !r       Æ‰      " !r       Æ/ ‰
      h ‰      b !r         Æ ‰      " !r       âO !r	       ä� ‰  €    ¢ !r       Æ‰	 €    â !r       Æ‰ €    " !r
       Æ/ ‰ €    b !r       Æ ‰ €    b !r       Æ ‰ €    ( ‰ €    " !r        ÆO ‰      ¢ !r
	       â� x   ÿ   ä y      &   â !r       âz  r  p`ð ä y      '   " !r       â/ r ÿ   p"p Ä ‰ @    b !r       â!r       ä ‰ @    " !r       Æ ‰ @    b !r       ÆO ‰
 @    ¢ $z  x  Ž Æ� ‰ @    â x   ÿàÿ â $z  ~   Ž ä‰ @    " z  r  p`ò â $x   ÿŽ â !r       â/ ‰ @    b 
  r  ÿàõ ä x    p"ò â ‰ €    b !r	       â $ÿÿ   ÿ â 
 j  ÿ„ â ‰      " !r       âr    ÿ† Ä 
 k   â !r
       âO ‰	      ¢ rÿÿ   ÿäÿ â !r       â�    ÿàÿ â ‰      â z j  ÿ† â!r       â    ÿàõ â ‰      " z k  � â !r       â/     ÿà÷ â ‰      b ÿÿ   ÿä â !r       â  j  ÿ„ â ‰      b $ÿÿ   ÿŽ â !r       â $tÿ   ÿ Ž â ‰ @    " 

 j  ÿ† Ä %v  j   Ž â ‰      " !r	       âO 	 k   â !r       â� 
 k  � â †	     â !r       Æ†    â !r       Æ/ †	     â !r       Æ†     â !r       Æ †	     â !r       Æ †    è ‰      ( †	 
    â M™        € ê x   ÿàÿ âO !r       Æ r    ÿàñ Ê $r ÿÿ   ÿ  â z j  ÿ€ È z k     Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      "   " ¹z  l      â Ey        € â ™x?    â y      !   b ¹z  F   
   â $rÿÿ   ÿ Ž â r6 ÿ       â $r8ÿÿ   ÿ Ž ä $r$ÿÿ   ÿ Ž Ä $r:ÿÿ   ÿ Ž â r ÿ   pRð är ÿ   pRò â$x    Ž Ê/ x ÿ     È r    ÿ� â $À   Ž ä $r ÿÿ   ÿ Ž â x+ÿ     È | +   pbô Ú G)  ð    € ê y      '   " $rÿÿ   ÿ Ž â x" n       â $rÿÿ   ÿ Ž â y      %   b $vÿ q  ÿ Ž ä $x(+   ÿ Ž â y	      &   ¢ $vÿ X  ÿ Ž ä $rÿÿ   ÿ Ž â r(   ÿŽ Ä z%( q  ÿàÿ â 'z y   Ž â xøÿÿÿÿÀŽ Æ $x   ÿ Ž â/ zÿ z    â $x   
Ž ä 'z	 s   Ž ÆO r2   ÿ� â$z p  +Ž â zÿ t    â $z |  +Ž âx2ÿ   2  â $x   ÿ Ž â x'ÿ     â $z |  Ž â x   ÿàÿ â $z v  Ž ârÿ   ÿ3Ž â $z v  Ž â x3ÿ   2  â $z p  Ž Ä %x   ÿŽ È %v"  € "Ž È $z  }  ÿŽ â r"  €ÿà÷ â $z		 w  ÿŽ ä %xn   ÿŽ Æ r 	    ÿàõ â %xn   ÿŽ â xJ   áÙ Æ $x"#   Ž â r!ÿ   ¥~ â $x   ÿ Ž âr,2   ÿà÷ â$r*ÿÿ   ÿ â r.   ÿàù â %xn   2Ž â r0   ÿàõ ä x)   ÿÀŽ â%x   Ž È %x
   Ž È $v#ÿ Y  ÿ Ž ä $x2   3Ž ä %x   Ž È %x   Ž È $x4   ' ä$x&   ' ä r.   ÿàõ Ê $x#   4 â r   ÿàõ È �y3B   ‘ ¢  r	#   ÿä Ê �y5   • è �y-   • ( �y/   • h �y1   • b r0   ÿàõ Ê $x#   & â r   ÿàõ È �yB   ‘ b $x   # Ê �y9   • h �y7   • h �y   • h �y   • b rH
   ÿàõ Ê $xI   # Ê �yHHJ   • b $tÿ   ÿ Ž â/ r>   ÿàõ Ê $x?#   ! Ê �yF>    • b xW $       â $x3   ÿ Ž âOx3   ÿ   â$x3@   ÿ Ž Æ x   ÿÀŽ ä x<   ÿÀŽ ä xÿ   5øŽ ä�x5ÿ   5  Æ %vJ   Ž â x5   øŽ ä x<ÿ   -øŽ Æ%v   Ž â�yJJ    ™ ¢  x-ÿ   -  ä $x<3    ÿ Ž â �yG    ™ " %v   Ž Æ x<-   <øŽ ä �yD    ™ ¦  %v<<   Ž Ê �yE<    ™ ¢  $x-3   ÿ Ž Ê x@-   ÿÀŽ â $xB3   ÿ Ž Æ x@@ÿ   /øŽ â$x-3   ÿ Ž â x/ÿ   /  Æ %v@@   Ž â x/   BøŽ ä xB-   ÿÀŽ ä r   ÿàõ â� �y>@    ™ â  %v   Ž â xBBÿ   1øŽ Æ $x   # â x1ÿ   1  â �y<    ™ b xL3   ÿ   â %vBB   Ž ä �y?J   • â xL1   LøŽ Æ �y;B    ™ ä %v@L   Ž È $x-   ÿ Ž â �y=@    ™ è  xL-   ÿÀŽ È xLLÿ   9øŽ â$x-@   ÿ Ž âx9ÿ   9  â $x€   ÿ Ž ä/ %vL   Ž ârL   ÿàõ ä xN-   ÿÀŽ â $xB    ÿ Ž â x9   øŽ â $xM#   " â x-ÿ   7  â�y1    ™ â x@Nÿ   7øŽ â %v   Ž â xB-   BøŽ Ä �y-L    • â  %v@@   Ž Æ �y/    ™ â  %vBB   Ž Æ �y5@    ™ b $x   ÿ Ž â/ r    (àµ ä �y3B    ™ d x   ÿÀŽ ä r'*   ÿd~ â $x   ÿ Ž â xNÿ   øŽ â%v Z  W Ž â xÿ     Æ $x9$   ÿŽ ä $x   ÿ Ž âx   øŽ â $x   ÿ Ž â xCÿ     â/ $x   9Ž â x   ÿÀŽ ä xC   øŽ â %v   Ž â xBÿ   øŽ Æ %v@N   Ž â�y9    ™ f %vT   Ž âr,   ÿàõ â �y7@    ™ d %vBB   Ž â rL    àë är    %àÇ â $x#   2 â rO'*   ÿÄþ â �yAB    ™ d! %v Z  W Ž Ä �y@T    ™ d $rÿÿ    Ž ä �yRj   ‘ b rB'*   ÿ„þ Æ �yI   ™ d $xUB$   ÿŽ ä/ �yK   ™ b %vL Z  W Ž Æ �yS   ™ b $xWO$   ÿŽ ä $x   UŽ â �yN   ™ b $x   WŽ Æ �yP   ™ h �yT   ™ h �yb   ™ h �y`   ™ b r   ÿàõ Æ �ya   ™ h �y^   ™ b r#   ÿä Æ �y_   ™ h �yQ   ™ h �y\   ™ h �yM   ™ h �yC   ™ h �yB    ™ h �yO    • h �y]   ™ h �yZ   ™ b r   ÿàõ Æ �y[   ™ d $x   # ä �yX   ™ h �yY   ™ h �yV   ™ h �yj   ‘ h �yW    ™ h �yU    ™ h �yL    • h �yc    • h �yd    • b r   ÿàõ Ê $x   # Ê �y    • b xeHÿ   ÿÀŽ Ê $xfH @  eŽ â xeÿ   H  Æ $xHf€   ÿ Ž âreee   ÿpŽ â $xf   ÿ Ž Æ xHH€€ÿÀŽ â $xe€   ÿ Ž â/ x€€ÿÀŽ ä xH€€€€ÿüŽ â $xe   ÿ Ž â x€€ÿÀŽ ä xÿþþþÿàÿ ä x€€€€ÿüŽ Ä x€€€€ÿüŽ ä xÿþþþÿàÿ ä x€€ÿÀŽ ä rH   ÿüŽ ä r   ÿüŽ ä xÿþþþÿàÿ Ä xH€€€€ÿüŽ ä x˜º  ÿ    ä x˜º  ÿ    ä r   ÿüŽ ä xHHÿþþþÿàÿ ä xÿÀŽ Ä x€€€€JÞŽ äO xÿÀŽ äx˜º  ÿ    ä x€€€€GÞŽ ärHH   ÿüŽ â $x   
Ž â xFÿ   ÿÀŽ â $x   
Ž â xÿÀŽ Ä x€€€€DÞŽ ä rF   ÿpŽ ä xHH˜º  ÿ    â $x   
Ž â xG€€€€G´Ž â $x€   ÿ Ž â xFÿ   F  ä xHÿÀŽ Ä xHH€€€€EÞŽ ä xD€€€€D´Ž â $x   ÿ Ž â x€€ÿÀŽ â $xFF @  FŽ ä $xHH   
Ž â x€€€€ÿüŽ â $xF€   ÿ Ž â x€€ÿÀŽ Ä xÿþþþÿàÿ ä x€€€€ÿüŽ â $xFF   ÿ Ž â x€€ÿÀŽ ä r   ÿüŽ ä xÿþþþÿàÿ ä x€€€€ÿüŽ Ä xEH€€€€E´Ž ä xHF€€ÿÀŽ ä x˜º  ÿ    ä r   ÿüŽ ä xFÿþþþÿàÿ ä x€€€€J´Ž Ä xJH€€€€ÿüŽ ä xÿÀŽ äx€€€€>ÞŽ ä� x˜º  ÿ    ä rFF   ÿüŽ ä xJÿþþþÿàÿ Ä r  €ÿàÿ ä xÿÀŽ äx€€€€<ÞŽ äxFF˜º  ÿ    ä rH   ÿüŽ ä x?ÿ   ÿÀŽ Ä x>€€€€>´Ž â $x   
Ž â xFÿÀŽ äxFF€€€€;ÞŽ ä x˜º  ÿ    â $x? @  Ž ä $xFF   
Ž ä $x€   ÿ Ž â x<€€€€<´Ž Ä x?ÿ   ?  ä xÿÀŽ äx€€€€=ÞŽ ä x€€ÿÀŽ â $x?? @  ?Ž â x   ÿ   â $x   
Ž â x€€€€ÿüŽ Ä xF€€€€;´Ž â $x;?€   ÿ Ž â x€€ÿÀŽ ä xFÿþþþÿàÿ ä x€€€€=´Ž â $x=?   ÿ Ž â x€€€€ÿüŽ ä x;;€€ÿÀŽ Ä rFF   ÿüŽ ä xHÿþþþÿàÿ ä x==€€ÿÀŽ ä x;€€€€ÿüŽ ä xFF˜º  ÿ    ä x=€€€€ÿüŽ Ä rHH   ÿüŽ ä xJÿþþþÿàÿ ä xF€€€€1ÞŽ äxFFÿÀŽ ä xfÿþþþÿàÿ ä rJJ;   ÿüŽ Ä xHH˜º  ÿ    ä xh-ÿ   ÿÀŽ â $xF   F
Ž â rff=   ÿüŽ ä xJJ˜º  ÿ    ä xH€€€€/ÞŽ â$xh- @  hŽ â xHHÿÀŽ Ä x-ÿ   -  ä xff˜º  ÿ    ä xJ€€€€5ÞŽ â$xH   H
Ž â xJJÿÀŽ â $xh€   ÿ Ž â xf€€€€3ÞŽ â$x;- @  -Ž â xffÿÀŽ Ä rJJ  €ÿàÿ â $xh   ÿ Ž â x€€ÿÀŽ â $xf   f
Ž â x-;   ÿ   ä x1F€€€€1´Ž ä x€€ÿÀŽ â $x;;   ÿ Ž â xFJ€€€€5´Ž Ä x€€€€ÿüŽ ä x5-€€ÿÀŽ ä x/H€€€€/´Ž ä x-€€€€ÿüŽ ä xHÿþþþÿàÿ ä x5€€€€ÿüŽ Ä x;;€€ÿÀŽ ä x3f€€€€3´Ž ä xJ-ÿþþþÿàÿ ä rHH   ÿüŽ ä xfÿþþþÿàÿ ä x;€€€€ÿüŽ Ä rJJ   ÿüŽ ä r5f5   ÿüŽ ä xfÿþþþÿàÿ ä xJJ˜º  ÿ    ä r;f;   ÿüŽ ä xJ€€€€9ÞŽ Ä xhJÿÀŽ ä x;;˜º  ÿ    ä xHH˜º  ÿ    â $xh   h
Ž â x;€€€€@ÞŽ äxl;ÿÀŽ ä xH€€€€7ÞŽ Ä xHHÿÀŽ â $x   l
Ž â xj5˜º  ÿ    Æ $xf   H
Ž ä &rHI   ÿ   â x@€€€€@´Ž â &rP   ÿ   â xj€€€€AÞŽ â&rHGK   H   ä &rT   ÿ   â xjjÿÀŽ â &rP1P   ÿ   Ä &rGb      ä&rHDS   H   ä &rG`      ä &rJ1I   ÿ   ä &rb/b   P   ä $xj   j
Ž Ä &r11T   ÿ   ä &rPDa      ä&rEN   H   ä &rD^      ä &rJ/K   J   ä&r1/`   1   Ä &rPE_   P   ârRÿ)   R  â &r>Q      ä &rE\      ä &rJFS   J   ä&rbFa   b   ä&r1F^   1   Ä &rP>]   P   â xHf€€€€7´Ž â &r<M      ä &r>Z      â x-        â &rN3N   J   ä$xRR   ÿ Ž ä &rb3_   b   Ä &r13\   1   ä &rP<[   P   â xJh€€€€9´Ž â &rC      ä &r<X      â xR   -ìŽ â &rQHQ   N   ä&rbH]   b   ârÿ)     â &r1HZ   1   Ä &rPY   P   â xNj€€€€A´Ž â &rRB      ä &rV      ä &rQJM   Q   ä&rbJ[   b   ä&r1JX   1   Ä &rPW   P   ä $rRR   ÿŽ ä &rU      â x++   ÿàÿ â 0rLÿL    A   ä 0rOÿO    A   ä $tÿ   ÿ Ž Ä $x   ÿ Ž ä &rQNC   Q   ä&rbNY   b   ä&r1NV   1   â ErR R       â &rB@B   Q   â  rLO     A  â &rb@W   b   â x   ìŽ â &r@@U   1   â | +   pbô â #r R       â $rPP   ÿŽ Ä $r   ÿŽ ä 0rcÿc    A   ä 0rÿd    A   ä $rBB   ÿŽ ä $rbb   ÿŽ ä$r@@   ÿŽ Ä 0rÿ    A   â ErP P       â  rLc     A  âEr        â  rL     A  â ErB B       â  rOO     A  â Erb b       â  rcc     A  âEr@ @       â  r     A  â xà  ÿà÷ â #r:P   :   â #r8   8   â #r6OB   6   â #r$cb   $   â #r@      â x€   ÿàÿ â $r#ÿÿ   #Ž â x%%€   ÿàÿ Ä x((€   ÿàÿ â D)       À â Gy  °äÿÿÿÿƒ ê Ay        € ê ˆ   ýÿ H   è ˆ 6€ýÿ H   è ˆ : þÿ H   è ˆ $€þÿ H   è ˆ 8 ÿÿ H   è ˆ €ÿÿ H   è {          ì M        € ê „y     H   ( „y    H   h „y €   H   ¨ „y    H   è „y €  H   ( „y    H   h „y    H   h „y €  H   h „y    H   h „y €  H   h „y €  H   b !r        Æ „y
    H   " !r       Æ/ „y	 €  H   b !r6       ÆO „y    H   ¢ !r:       Æ� „y €  H   â !r$       Æ„y €  H   â !r       Æ„y    H   â !r8       Æ „y €  H   â !r       â !r       ä ‰      " !r       Æ y      %   b !r       Æ y      &   b !r

       â z  r      â !r	       â/ !r       ÈO ‰      " !r       Æ� ‰      b !r       â!r       ä ‰	      ¢ !r

       Æ ‰       â !r       Æ ‰
      ( ‰      b !r       â !r       È/ ‰      " !r		       ÆO ‰      b !r         Æ� ‰	      ¢ !r
       Æ‰  €    â !r       Æ‰      ( ‰
      b !r       â !r       È/ ‰ €    " !r	       ÆO ‰ €    b !r        Æ� ‰	 €    ¢ !r       Æ‰  @    â !r

       Æ‰ €    ( ‰
 €    b !r       â !r       È/ ‰ @    " !r		       ÆO ‰ @    â !r        â� !r       È‰      b !r
       Æy      '   ¢ $x   ÿ Ž Æ ‰	 @    â $z  x  Ž â z  r  p`ð ä‰ @    " x
   ÿàÿ â !r       â r ÿ   p"p â ‰ @    " z 
 r  p`ò Æ ‰
      b x    p"ò â !r       â/ $tÿ   ÿ Ž ä $z  ~   Ž âO !r       â !r	       Æ� 
  r  ÿàõ â%v  j   Ž Æ 	   ÿàÿ â!r       âr    ÿ† â ‰      b $ÿÿ   ÿ â !r       â 
 j  ÿ„ â‰      " $rÿÿ   ÿŽ â !r
       â
 k   â ‰      ¢ 	 	   ÿàõ Ä z j  ÿ† â ‰      ä $ÿÿ   ÿ â 	 j  ÿ„ â †	     â	 z k  � ä 		 k   â †    â	 !r       â/ !r       È †	     â	 !r       ÆO †     â	 !r       Æ� †	     â	 M™        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      "   " ¹z  l      â Ey  P    € â ™x?    â y       !   b ¹z  F   
   â $rÿÿ   ÿ Ž ä $r(ÿÿ   ÿ Ž ä $rÿÿ   ÿ Ž Ä $r8ÿÿ   ÿ Ž â r ÿ   pRð är ÿ   pRò â$x!     Ž Ê/ xÿ   !  È r!   ÿ� â $€    Ž Æ x1ÿ     È | 1   pbô Ú G)  @    € ê y      '   " $rÿÿ   ÿ Ž â xøÿÿÿÿÀŽ â $rÿÿ   ÿ Ž â y      &   b $t&ÿn   ÿ Ž ä $x!!   
Ž â y      %   ¢ $x*1   ÿ Ž ä $x!   ÿ Ž âx+ÿ   !  â $x!   ÿ Ž â z* q  ÿàÿ â $v3ÿ X  ÿ Ž Ä $v)ÿ Y  ÿ Ž â x-   ÿÀŽ â $rÿÿ   ÿ Ž ä 'z y   Ž È 'z s   Ž â/ zÿ z    Æ $x   ÿ Ž âO zÿ t    â $z |  1Ž ä$z	 p  1Ž âx   ÿàÿ â $z v  Ž ä $z	 |  	Ž â rÿ   ÿ3Ž â $z p  Ž Ä $z	 v  	Ž ä $z$ }  ÿŽ ä %x   ÿŽ È %x	n   ÿŽ È %v&  € &Ž â x"J   áÙ Æ $z w  ÿŽ â r#&  €ÿà÷ â %x
n   ÿŽ â r!!   ÿ� ä r$$   ÿàõ â $x&'   Ž â xÿ     â %x   Ž â r%ÿ   ¥~ ä xÿ     â $r,ÿÿ   ÿ â r.   ÿà÷ â%x   
Ž â r/!   ÿàù Ä r0!
   ÿàõ â %xn   Ž È $x2   Ž ä %x   Ž È %x   
Ž È $x4   + ä$x'   + ä r3/   ÿàõ Ê $x)   4 â r3   ÿàõ È �yB   ‘ ¢  $x	   ) Ê �y5   • è �y    • ( �y9   • h �y   • b r30   ÿà÷ Ä r3   ÿàõ Æ $x)   'Ž ä $x   ) Æ �yEB   ‘ h �yA   • h �y@   • h �y   • h �yF   • b r3   ÿàõ Ê $x   ) Ê �y;J   • b $tÿ   ÿ Ž â r>3"   ÿàõ Ê $x?)   % Ê �y>>    • b $tJÿ$   ÿ Ž ä $x   ÿ Ž ÊO x   ÿÀŽ â $x@   ÿ Ž ä/$x€   ÿ Ž â xÿ   5øŽ ä�x5ÿ   5  ä x   ÿÀŽ ä x5   øŽ â %v   Ž â xÿ    øŽ Æ%v   Ž â�y=    ™ ¢ xÿ      ä $x    ÿ Ž â �y     ™ â  %v6   Ž Æ x   øŽ Æ �y66    ™ ä %v   Ž Ê/ �y5    ™ â $x   ÿ Ž âxÿ   9  â$x   ÿ Ž Æ x   ÿÀŽ ä x   øŽ ä xÿ   9øŽ â $x   ÿ Ž âxÿ     â/ $x   ÿ Ž ä %v   Ž â x   ÿÀŽ Æ %v   Ž â�y7    ™ "x   øŽ ä r3   ÿàõ â �y<    ™ b xÿ   øŽ â %v   Ž È $x   ) â �y9    ™ " %v   Ž Æ �y?J   • ( �y:    ™ " $xE   ÿ Ž âxÿ   A  â $xE€   ÿ Ž ä/$xBE@   ÿ Ž âx   ÿÀŽ â $xE    ÿ Ž â x   øŽ Ä xB   ÿÀŽ ä xÿ   @  äxÿ   AøŽ ä xÿ   @øŽ ä x   øŽ â %v   Ž â r3#   ÿàõ Æ %v   Ž â�yA    ™ f %v   Ž â�y@    ™ f %v   Ž â �yB    ™ f $x)   & â �yC    ™ h �yD    • b $xGE   ÿ Ž Ä $xE   ÿ Ž ä/$xE   ÿ Ž âxÿ     â $xEE   ÿ Ž â xHÿ   F  ä xGG   ÿÀŽ ä x   ÿÀŽ ä x   EøŽ Ä r!$   *àµ ä xIH   øŽ ä xGÿ   FøŽ ä xÿ   øŽ ä rG+,   ÿd~ ä r!$   àµ â %v   Ž È %vI   Ž â�yF    ™ f %v   Ž â�yE    ™ f %v   Ž â r+,   ÿd~ Æ %v Z  J Ž â�yH    ™ f $xIG$   ÿŽ ä %v Z  J Ž â/ �yG    ™ f $x$   ÿŽ Ä $x   IŽ ä $x   Ž Æ �yI   ™ h �yK   ™ b r3.   ÿàõ Æ �yJ   ™ h �yR   ™ h �yT   ™ b $x)   2 â r3   ÿà÷ Ä/ �yP   ™ h �yS   ™ b $x   )Ž Æ �yN   ™ h �yQ   ™ h �yj   ‘ b r3   ÿàõ Æ �yL   ™ h �yO   ™ h �yj   ‘ h �yW   ™ b $x   ) Æ �yM   ™ b r
3   ÿàõ Æ �yV   ™ h �yU    ™ b $x   ) Æ �y    ™ h �y    • h �yX    • h �yY    • h �y    • b xZ;ÿ   ÿÀŽ Ê $xZ; @  ZŽ â x;ÿ   ;  Æ $xZ€   ÿ Ž ä$xZ   ÿ Ž ä/ $x;; @  ;Ž â x€€ÿÀŽ ä x€€ÿÀŽ â $x;€   ÿ Ž â x€€€€ÿüŽ â $x;   ÿ Ž â x€€€€ÿüŽ Ä x€€ÿÀŽ ä xZÿþþþÿàÿ ä x€€ÿÀŽ ä x\ÿþþþÿàÿ ä x€€€€ÿüŽ ä rZZ   ÿüŽ Ä x€€€€ÿüŽ ä r\\   ÿüŽ ä xÿþþþÿàÿ ä xZZ˜º  ÿ    ä xÿþþþÿàÿ ä x\˜º  ÿ    Ä r   ÿüŽ ä xZÿÀŽ äxÿÀŽ ä x˜º  ÿ    ä r   ÿüŽ ä xZZ€€€€=ÞŽ ÄO x€€€€ ÞŽ Æ� $xZ   
Ž â xZÿÀŽ ä x>ÿ   ÿÀŽ â $x   
Ž â x€€€€6ÞŽ ä x˜º  ÿ    ä x€€€€ ´Ž â $x   Z
Ž Ä $x > @  Ž â xÿÀŽ äx€€€€5ÞŽ â $x €   ÿ Ž âx€€€€6´Ž â $x    ÿ Ž â x>ÿ   >  â $x   
Ž â x€€ÿÀŽ ä x€€ÿÀŽ â $x;> @  >Ž â x5€€€€5´Ž Ä x€€€€ÿüŽ â $x;€   ÿ Ž â x6€€€€ÿüŽ â $x;;   ÿ Ž â x ÿþþþÿàÿ ä x66ÿþþþÿàÿ ä x€€ÿÀŽ ä r     ÿüŽ Ä x;€€ÿÀŽ ä r66   ÿüŽ ä x€€€€ÿüŽ ä x  ˜º  ÿ    ä x;€€€€ÿüŽ ä x6˜º  ÿ    Ä x>ÿþþþÿàÿ ä x ÿÀŽ äx;;ÿþþþÿàÿ ä x  €€€€7ÞŽ äx6ÿÀŽ äx€€€€<ÞŽ Är>>   ÿüŽ â $x     
Ž â r;;   ÿüŽ â $x   6
Ž â x>˜º  ÿ    ä x;˜º  ÿ    ä x7 €€€€7´Ž ä x<€€€€<´Ž Ä x ÿÀŽ ä x6?ÿ   ÿÀŽ ä x€€€€:ÞŽ ä xÿÀŽ äx€€€€9ÞŽ ä xÿ   ?  â $x6? @  6Ž Ä $x    
Ž ä $x   
Ž ä $x @  Ž ä $x 6€   ÿ Ž âx:€€€€:´Ž â $x6   ÿ Ž ä $x€   ÿ Ž âx  €€ÿÀŽ â $x   ÿ Ž â x€€ÿÀŽ Ä x€€ÿÀŽ ä x€€ÿÀŽ ä x6 €€€€ÿüŽ ä x;€€€€ÿüŽ ä x€€€€=´Ž ä x€€€€9´Ž Ä x=€€€€ÿüŽ ä x?€€€€ÿüŽ ä x96ÿþþþÿàÿ ä x6;ÿþþþÿàÿ ä x>=ÿþþþÿàÿ ä xZ?ÿþþþÿàÿ Ä r66   ÿüŽ ä r99    ÿüŽ ä r>>   ÿüŽ ä rZZ   ÿüŽ ä x66˜º  ÿ    ä x99˜º  ÿ    Ä x>˜º  ÿ    ä xZZ˜º  ÿ    ä x6€€€€AÞŽ äx9€€€€@ÞŽ äx 9ÿÀŽ ä x66ÿÀŽ Ä x9€€€€BÞŽ äx>ÿÀŽ ä xZ€€€€CÞŽ äxZZÿÀŽ â $x6   6
Ž â xDÿ   ÿÀŽ â $x    
Ž ä $x99   >
Ž Ä $xZ   Z
Ž â xÿ   D  â $xDD @  Ž â x@€€€€@´Ž ä x9€€€€B´Ž â $x9 @  Ž ä $xD€   ÿ Ž ä$xD   ÿ Ž Ä $x 9€   ÿ Ž âxA6€€€€A´Ž â $x99   ÿ Ž â x€€ÿÀŽ ä x6€€ÿÀŽ ä x?9€€ÿÀŽ ä x€€€€ÿüŽ ä x96€€€€ÿüŽ Ä x; €€ÿÀŽ ä x ÿþþþÿàÿ ä x99ÿþþþÿàÿ ä r     ÿüŽ ä r996   ÿüŽ ä x ˜º  ÿ    Ä x 9˜º  ÿ    ä x€€€€FÞŽ äx6ÿÀŽ ä x €€€€EÞŽ äx=;€€€€ÿüŽ ä x  ÿÀŽ â $x   6
Ž â x>=ÿþþþÿàÿ Ä xB?€€€€ÿüŽ â $x     
Ž ä &rI   ÿ   â r>>;   ÿüŽ â &rI@I   ÿ   â xF€€€€F´Ž â &rK   ÿ   âxBBÿþþþÿàÿ â &r@@K   ÿ   ä &rJ      âx>>˜º  ÿ    â &rIAJ   I   â rBB?   ÿüŽ â &rR      Ä &r@AR   @   â xCZ€€€€C´Ž â &rT      âx;>€€€€GÞŽ â &rTT   I   â x9B˜º  ÿ    â &rP      âx>>ÿÀŽ â &r@P   @   ä &r5S      âx=9€€€€HÞŽ â &rTCS   T   â xB9ÿÀŽ â &r5N      Ä &rCCN   @   â x  €€€€E´Ž â $x>;   >
Ž â rÿ-     â &r7Q      ä&rTFQ   T   ä&r7L      ä&rCFL   C   â xG>€€€€G´Ž â $x==   B
Ž Ä &r<O      â rÿ-     â &rT O   T   ä $tÿ   ÿ Ž ä $x   ÿ Ž ä &r<W      ä&rC W   C   â xH=€€€€H´Ž â &r:M      âx   ìŽ â &rTGM   T   â x11   ÿàÿ â &r:V      Ä &rCGV   C   ä $tÿ   ÿ Ž ä $x   ÿ Ž ä &rU      ä&rTHU   T   â | 1   pbô â &r      Ä &rHH   C   â x   ìŽ â $r   ÿŽ ä 0rÿ    A   ä 0rÿX    A   â Er        â $r   ÿŽ ä 0rYÿY    A   â  r     A  â $rTT   ÿŽ Ä $rHH   ÿŽ ä 0rÿ    A   â #r8   8   â  rY     A  âEr        ä ErT T       â  r     A  â ErH H       â  rYY     A  â x33à  ÿà÷ â #r(   (   â #rT      â #rYH      â x€   ÿàÿ â $r)ÿÿ   )Ž â x**€   ÿàÿ â D)       À â Gy  @çÿÿÿÿƒ ê Ay        € ê ˆ 8 þÿ H   è ˆ €þÿ H   è ˆ ( ÿÿ H   è ˆ €ÿÿ H   è {          ì M        € ê „y      H   ( „y  €   H   h „y     H   ¨ „y     H   è „y  €  H   ( „y     H   h „y     H   h „y	  €  H   h „y     H   h „y  €  H   h „y
  €  H   b !r8       Æ „y  €  H   " !r       â/ !r(       âO !r       ä� !r       â!r       â!r       ä !r		       Æ ‰      b !r       Æ ‰	      ¢ !r       Æ ‰      â !r

       È !r
       â !r       È/ ‰      " !r	       ÆO ‰      b !r       Æ� ‰      ¨ ‰      â !r       â !r       È/ ‰      " !r       ÆO ‰ €    b !r       Æ� ‰ €    ¨ ‰
 €    è y      %   " !r       â !r       Ä/ y      &   " !r       ÆO ‰ @    b !r

       Æ� ‰	 @    ¢ $x   ÿ Ž Æ‰
 @    ä z  r  p`ð ä ‰ €    $ r  ÿ   p"p ä y      '   b !r       â/ x   ÿàÿ â $z x  Ž â !r		       ÄO ‰      " z  r  p`ò â $tÿ   ÿ Ž â !r
       ä� ‰
	      b x     p"ò â !r       ä‰      ¢ $z ~  Ž Æ‰ @    ä 
  r  ÿàõ Ê $ÿÿ   ÿ â 
  j  ÿ„ â !r       â %v j   Ž ä 
  k   â !r		
       ä/ †	     â !r       ÆO † 	   â !r       Æ� †	     è ‰      b  M™        € ê ¹z  r      â !r       â/ �x   ?àÿ Ì |   ÿàñ Ê $rÿÿ   ÿ  â z j  ÿ€ È z k    Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y       &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž ä $rÿÿ     Ž È % \   Ž Ê �	    ™ b ¹z  l      â Ey  P    € â ™x?    â y      "   b $r"ÿÿ   ÿ Ž Æ y      !   ¢ $ÿÿ   ÿ Ž ä $ÿÿ    Ž È '
 m   Ž È $‚ÿÿ   ÿ Ž â 
ÿ n    Ê $ÿÿ   
Ž â r ÿ   pRò ä/r ÿ   pRô â$
 o  Ž ä $x    Ž äO y      '   " $‚ÿÿ    Ž ä xÿ     Ä 'Š s   Ž ä r   ÿ� È xÿ     È |    pbö ä Šÿ t    Ö G9  Ð    € ê $rÿÿ   ÿ Ž â xøÿÿÿÿÀŽ â y	      %   " $z
 w  ÿŽ ä 'z y   Ž È $x   
Ž â zÿ z    â $z }  ÿŽ ä $x   ÿ Ž âr    ÿ� â$z |  Ž â r
   ÿà÷ â $tÿn   ÿ Ž â rÿ   ÿ3Ž â $z v  Ž âxÿ     â $x!   ÿ Ž â x ÿ      â %x   ÿŽ È $z	 p  Ž â xJ   ÿáñ â $tÿ   ÿ Ž ä %v X  Ž È $rÿÿ     ä %x!   ÿŽ âx!!   ÿÀŽ Æ $r#ÿÿ    Ž â x   ÿüŽ â $r"ÿÿ   ÿ Ž ä $x   ÿ Ž ä $rÿÿ   ÿŽ ä %v  € Ž È r#   ÿà÷ ä r#   ÿàñ Æ $x   Ž ä$x	     Æ �yB   ‘ ¨ �y    • è �y
   • ( �y   • h �y   • " r#   ÿàñ Ê $x     Ê �y$    • b $t)ÿ   ÿ Ž ä $t*ÿ$   ÿ Ž â r#   ÿàñ Ê/ $x     ä $x   ÿ Ž äO$x&€   ÿ Ž â x%ÿ     ä�x   ÿÀŽ ä x&%   &øŽ ä xÿ   øŽ â $x@   ÿ Ž Ä %v&  ) Ž Æ x   ÿÀŽ â %v&  ) Ž â r   àÇ â �y%    ™ ¢  xÿ   
  â$x    ÿ Ž â x
ÿ   
øŽ â �y&&    ™ â r(   ÿ„þ ä x   øŽ â %v

  ) Ž È %v Z  * Ž â �y'    • &+ $x($   ÿŽ ä %v  ) Ž â �y(
    ™ & $x		   Ž â �y*    ™ ( �y,   ™ ( �y+   ™ " $x   ÿ Ž Æ �y-   ™ $ x   ÿÀŽ ä �y.   ™ $ xÿ   øŽ â$x
   ÿ Ž âxÿ     â $x   ÿ Ž â �y0   ™ b %v  ) Ž â x
   
øŽ Ä x   ÿÀŽ â/ $x   ÿ Ž â xÿ     â�y    ™ b xÿ   øŽ â %v
  ) Ž âx   øŽ â �y/   ™ d %v
  ) Ž ä�y    ™ d %v  ) Ž â r #   ÿàñ â�y
    ™ f# r    ÿ  â �y    ™ h �y)   ™ h �yj   ‘ b $r
ÿÿ   # Ž Ä/ $rÿÿ    Ž â �y    ™ h �y1
    • h �y2    • b x3$ÿ   ÿÀŽ Ê $x3$ @  3Ž â x$ÿ   $  Æ $x3€   ÿ Ž ä$x33   ÿ Ž Æ x€€ÿÀŽ ä x3€€ÿÀŽ â $x3$ @  $Ž â x$€€€€ÿüŽ ä x
€€€€ÿüŽ â/ $x3€   ÿ Ž ä$x	3   ÿ Ž â x

ÿþþþÿàÿ Ä x€€ÿÀŽ ä x$$ÿþþþÿàÿ ä x		€€ÿÀŽ ä r

   ÿüŽ ä x€€€€ÿüŽ ä r$$   ÿüŽ Ä x	€€€€ÿüŽ ä xÿþþþÿàÿ ä x$˜º  ÿ    ä x$ÿþþþÿàÿ ä x4
˜º  ÿ    ä r   ÿüŽ Ä x
ÿÀŽ ä r$$	   ÿüŽ ä x˜º  ÿ    ä x$˜º  ÿ    ä x$ÿÀŽ ä x   ÿàÿ È |    pbð ä x##à  ÿà÷ ä x€   ÿàÿ Æ $rÿÿ   Ž â x3€€€€&ÞŽ ä� x4ÿÀŽ äx44€€€€%ÞŽ âO $x33   

Ž â x
'ÿ   ÿÀŽ Æ$x4   
Ž â x	3€€€€&´Ž ä x€€€€(ÞŽ â $x
' @  
Ž â x&ÿÀŽ äx€€€€*ÞŽ â $x   $
Ž â x€€€€%´Ž â $x
€   ÿ Ž Ä &r		,   ÿ   â x€€€€(´Ž â $x   &
Ž â x'ÿ   '  â &r+   	   â x€€ÿÀŽ ä x€€€€*´Ž â &r-      â x	€€€€ÿüŽ â $x'' @  'Ž Ä $x

   ÿ Ž ä &r..      â x	ÿþþþÿàÿ â $x'   ÿ Ž âx

€€ÿÀŽ â $x	'€   ÿ Ž â r   ÿüŽ ä x%€€ÿÀŽ ä x	€€ÿÀŽ Ä x
€€€€ÿüŽ ä x€€€€ÿüŽ ä xÿþþþÿàÿ ä x	˜º  ÿ    ä x&%€€€€ÿüŽ ä x$ÿþþþÿàÿ Ä r
   ÿüŽ ä x	ÿÀŽ äx		€€€€ÞŽ äx&&ÿþþþÿàÿ ä r$$   ÿüŽ ä x˜º  ÿ    â $x		   
Ž â r&&%   ÿüŽ Ä x$$˜º  ÿ    ä xÿÀŽ äx€€€€ÞŽ ä x&˜º  ÿ    ä x$ÿÀŽ â$x   
Ž â x$$€€€€ÞŽ Ä x		€€€€´Ž ä xÿÀŽ â$x$$   
Ž â x€€€€ÞŽ â &r		)   .   â x€€€€´Ž ä rÿ!     â $x   
Ž â x$$€€€€´Ž â &r00   	   Ä $t	ÿ   ÿ Ž â x€€€€´Ž â $x   ÿ Ž ä &r$$/   0   Æ x	   	ìŽ â &r   $   ä 0r1ÿ1    A   ä $r		   ÿŽ ä 0r2ÿ2    A   Æ Er	 	       ä  r112     A  È #r"1	   "   â G‰  0ôÿÿÿÿƒ ê Ay        € ê ˆ "€ÿÿ H   è {          ì M)        € ê „y     H   " r ÿ   pRð Æ „y €   H   h „y    H   ¨ y      %   â !r"       È !r       È/ !r       ÊO ‰      $ !r       Ê‰      $ !r       Ê ‰ €    $ !r       Ê ‰ @    $ !r	       â $z  x  Ž È� ‰
	      b  M	        € ê $tÿ   ÿ Ž â !r		
       â? $z ~  Ž È %v j   Ž Ê †y 	    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž È % \   Ž Ê �	    ™ ¢  $rÿÿ   ÿ Ž â Ey  Ð    € ä y
      "   b 'Š  s   Ž Æ y      !   b '
 m   Ž Æ y      %   â $rÿÿ   ÿ Ž â 
ÿ n    ä y	      '   " rÿ
   ÿü‚ â/ $x

    Ž Ê xÿ   
  È r
   ÿ� ä Šÿ t     â $r ÿÿ   ÿ Ž Æ r       âO G  À    € ê z ÿ `  pPô äŸ z ÿ ^  pPò ä z ÿ b  p ö ä z ÿ a   Sô ä z ÿ _  Sò ä z ÿ c  0't Ö $ 	 ~  ÿŽ âÿ     â$ª	 ~  ÿŽ ä $ x  ÿŽ â      àë â$ª x  ÿŽ â ¢   àÇ ä¨ÿ     ä ÿÿ   Äþ Ä      ÿàý ä ¢   ÿàû ä ¢ÿÿ   „þ â $ÿÿ    â   ^  ÿ† â$rÿÿ   ÿ Ž â ª b  ÿˆ â $¢ÿÿ   Ž â   _  � â $r ÿÿ   ÿ Ž Æ ª c   â �     h �©      d Ay        € êŸ ¹z  l      â xÿ     â ™x?    â $ÿÿ   
Ž â Ey  0    € â $rÿÿ   ÿ Ž Ä $
 o  Ž â|    pbò â $‚ÿÿ    Ž ä $rÿÿ   ÿ Ž Ô G  À    € ê xøÿÿÿÿÀŽ â $rÿÿ   ÿ Ž ä $rÿÿ    Ž ä 'z	 y   Ž È $x
   
Ž â zÿ z    â $z		 }  ÿŽ ä $x   ÿ Ž ä $z |  Ž ä $z w  ÿŽ â rÿ   ÿ3Ž â $z v  Ž Ä $rÿÿ   ÿ Ž â r		   ÿàõ â %x   ÿŽ È $z p  Ž â xJ   ÿáñ â $x
   ÿ Ž â r   ÿ� â %xn   ÿŽ â z X  ÿàó ä xÿ     â $rÿÿ    Ž â zÿ Y  þ  â $rÿÿ   ÿ Ž Ä $rÿÿ   ÿ ä $t*ÿ   ÿ Ž â x,ÿ     â$x)   ÿ Ž â z X  à‘ Æ %v) X  *Ž â z Y  ,$~  È r   ÿàñ â �yB   ‘ ¨ $x     Ê �y   • è  �y   • ( �y    • ( �y   • " %v X  *Ž Ê r   ÿàñ Ê $x     Ê �yJ   • " $t+ÿ   ÿ Ž â r   ÿàñ Ê/ $x     Ê �y.    • " $t$ÿ$   ÿ Ž ä $x   ÿ Ž äO$x€   ÿ Ž ä$x@   ÿ Ž â x   ÿÀŽ È xÿ   øŽ ä�xÿ     ä x   ÿÀŽ â %v  + Ž âx   øŽ ä xÿ   øŽ ä�y    ™ ¢  %v  + Ž â xÿ     Æ $x    ÿ Ž ä %v  + Ž â�y    ™ $ x   øŽ ä �y-    ™ & %v  + Ž Ê �y/    ™ " r	   
à‘ â $x   ÿ Ž Æ r",   ÿ$~  ä x   ÿÀŽ â� %v Z  $ Ž Æ xÿ    øŽ â/$x"$   ÿŽ â xÿ      â $x!   ÿ Ž ä $x   Ž ä %v  + Ž â x   !øŽ ä �y#   ™ â $x   ÿ Ž Æ �y0    ™ ä x   ÿÀŽ â %v  + Ž âxÿ     â�y&   ™ â  xÿ   øŽ â $x   ÿ Ž ä �y    ™ ä  %v$  + Ž âx   øŽ â �y'   ™ è �y   ™ â %v   + Ž â xÿ     Ä z X  à‘ â �y$$    ™ æ  z Y  $~  â �y      ™ è  �y(   ™ è  �yj   ‘ è  �y"   ™ â  z X  ÿàñ Æ �y   ™ â" z Y  ÿä  Æ �y    ™ è �y    • è �y%    • â x2ÿ   ÿÀŽ Ê $x2 @  2Ž â xÿ     Æ $x!2€   ÿ Ž ä$x2   ÿ Ž Æ x1!€€ÿÀŽ â $x! @  Ž â x€€ÿÀŽ ä x1€€€€ÿüŽ â $x2!€   ÿ Ž â x€€€€ÿüŽ ä/ xÿþþþÿàÿ ä x2€€ÿÀŽ Ä r1   ÿüŽ â $x!!   ÿ Ž â x2ÿþþþÿàÿ ä x€€€€ÿüŽ ä x˜º  ÿ    ä x4ÿþþþÿàÿ ä r22   ÿüŽ Ä x!€€ÿÀŽ ä xÿÀŽ ä r44   ÿüŽ ä x€€€€ÿüŽ ä x2˜º  ÿ    ä x44˜º  ÿ    Ä xÿþþþÿàÿ ä x2ÿÀŽ ä r   ÿüŽ ä z ÿ `  pPð È z ÿ a   Sð ä x   ÿàÿ ä x€€€€ÞŽ ÊO $x   
Ž â x4ÿÀŽ ä x€€€€ÞŽ äx€€€€´Ž ä x44€€€€-ÞŽ â $x   2
Ž â x.ÿ   ÿÀŽ ä x2˜º  ÿ    â $x44   
Ž â x€€€€´Ž â $x. @  Ž â xÿ   .  Ä x2ÿÀŽ â$x€   ÿ Ž â x22€€€€/ÞŽ â $x @  Ž ä $x   ÿ Ž ä $x.2   
Ž â x€€ÿÀŽ â $x!€   ÿ Ž â x€€ÿÀŽ Ä x€€€€ÿüŽ ä x!€€ÿÀŽ ä x!€€€€ÿüŽ ä x2ÿþþþÿàÿ ä x-4€€€€-´Ž ä x4!ÿþþþÿàÿ â $x   ÿ Ž â r22   ÿüŽ Ä r44   ÿüŽ ä x2˜º  ÿ    ä x€€€€ÿüŽ ä x€€ÿÀŽ ä x2ÿÀŽ ä xÿþþþÿàÿ Ä x€€€€0ÞŽ ä� x€€€€ÿüŽ ä x44˜º  ÿ    â $x   2
Ž â xÿþþþÿàÿ â &r#   ÿ   â r   ÿüŽ ä x4ÿÀŽ â &r      â x44€€€€ÞŽ Ä r   ÿüŽ ä x€€€€0´Ž ä x˜º  ÿ    â $x44   
Ž â x..€€€€/´Ž â &r-&      â x˜º  ÿ    ä xÿÀŽ Ä x€€€€$ÞŽ â &r.'      â x4€€€€´Ž ä x.ÿÀŽ â $x   
Ž â x€€€€ ÞŽ ä x)   ÿÀŽ â &r(      â x€€€€$´Ž â $x   .
Ž â rÿ     â &r"      Ä $tÿ   ÿ Ž â x €€€€ ´Ž â $x   ÿ Ž ä &r      Æ x   ìŽ â &r       ä 0rÿ    A   ä $r   ÿŽ ä 0rÿ%    A   â G‰   
    € ê z `  à‘ â %v) `  *Ž Æ z a  ,$~  ä r   ÿàñ Æ �y-B    ¤ $x     Ê �y     è �y%    ( �y$    ( �y.    " %v `  *Ž Ê r   ÿàñ Ê $x     Ê �y)J    " $x!-   ÿ Ž äO$x-€   ÿ Ž Æ x!!   ÿÀŽ ä xÿ      ä�x !ÿ    øŽ ä x   øŽ â $x-@   ÿ Ž ä/ %v   + Ž È %v   + Ž â xJ   ÿáõ â �y*    ™ ¦  z `  à‘ â $x0-    ÿ Ž â x   ÿÀŽ â $r2ÿÿ    â xÿ   %  â�y,     ™ â xÿ   %øŽ Ä x0   0øŽ ä z a  2$~  â %v  + Ž È %v0  + Ž â �y/     ( �y0    ™ ( �y1    ™ " $x -   ÿ Ž Ê/ x!    ÿÀŽ È x !ÿ   $øŽ â$x2-   ÿ Ž âx!ÿ   $  â $x3-   ÿ Ž ä %v$   + Ž âx !   2øŽ ä x33   ÿÀŽ ä �y2$    ™ " $x-   ÿ Ž â xÿ   .  â%v    + Ž â x3ÿ   .øŽ Ä x   øŽ Æ %v  + Ž â�y      ™ " z `  à‘ ä %v  + Ž ä �y    ™ " z a  $~  Æ �y    ™ ( �yj    " z$ `  ÿàñ È/ z% a  ÿä  Ê �y$$     " x)ÿ   ÿÀŽ Ê $x!) @  Ž â x)ÿ   )  Æ $x!€   ÿ Ž ä$x!!   ÿ Ž Æ x€€ÿÀŽ â $x)) @  )Ž â x!!€€ÿÀŽ ä x€€€€ÿüŽ â $x)€   ÿ Ž â x!€€€€ÿüŽ ä x.ÿþþþÿàÿ â $x%)   ÿ Ž â/ x4ÿþþþÿàÿ Ä x€€ÿÀŽ ä r..   ÿüŽ ä r44!   ÿüŽ ä x%€€ÿÀŽ ä x!€€€€ÿüŽ ä x.˜º  ÿ    Ä x%€€€€ÿüŽ ä x!!ÿþþþÿàÿ ä x.ÿÀŽ ä x4˜º  ÿ    ä x%%ÿþþþÿàÿ ä r!!   ÿüŽ Ä x4ÿÀŽ ä r%%   ÿüŽ ä x!!˜º  ÿ    ä x%˜º  ÿ    ä x€€€€*ÞŽ ÊO $x   .
Ž â x€€€€,ÞŽ È� x*€€€€*´Ž ä x.!ÿÀŽ â $x   4
Ž â x4ÿÀŽ äx/ÿ   ÿÀŽ äx!!€€€€0ÞŽ ä x€€€€1ÞŽ â $x/ @  Ž â x,€€€€,´Ž â $x!!   .
Ž Ä &r**#   ÿ   ä $x4   4
Ž â x!!€€€€0´Ž â $x€   ÿ Ž ä &r,   *   â x44€€€€1´Ž ä x€€ÿÀŽ â &r!&      â x/ÿ   /  Ä x€€€€ÿüŽ â &r'4'      ä $x!/ @  /Ž ä $x   ÿ Ž â xÿþþþÿàÿ â $x!€   ÿ Ž Æ x€€ÿÀŽ â $x!!   ÿ Ž â r   ÿüŽ ä x€€ÿÀŽ ä x€€€€ÿüŽ ä x˜º  ÿ    ä x#!€€ÿÀŽ Ä x!€€€€ÿüŽ ä xÿþþþÿàÿ ä x&ÿÀŽ äx€€€€2ÞŽ ä x%#€€€€ÿüŽ ä x*!ÿþþþÿàÿ Ä r   ÿüŽ â $x   &
Ž â x,%ÿþþþÿàÿ ä r**   ÿüŽ ä x˜º  ÿ    ä r,,#   ÿüŽ ä x*˜º  ÿ    Ä xÿÀŽ äx€€€€ ÞŽ ä x,˜º  ÿ    ä x&ÿÀŽ â $x   
Ž â x€€€€ÞŽ ä x€€€€2´Ž Ä xÿÀŽ â $x   &
Ž â x€€€€ÞŽ â &r(   '   â x€€€€ ´Ž ä rÿ     â $x   
Ž â x€€€€´Ž â &r"      ä $tÿ   ÿ Ž â x€€€€´Ž â $x   ÿ Ž Ä &r      Æ x   ìŽ â &r      ä 0r$ÿ$    A   ä $r   ÿŽ Æ  r$$     A  ä Er        Ê #r$      â |    pbð â  r     A  â Er        ä xà  ÿàó ä x

€   ÿàÿ â #r      ä $rÿÿ   Ž  È D	       À â Gy  àèÿÿÿÿƒ ê Ay        € ê y      "   " Ey        € â r ÿ   pRð är ÿ   pRò Ö G‰  `     € ê y      !   " z ÿ `  pPð È z ÿ a   Sð â $x    Ž Ê ˆs €ÿÿ H   î ˆ    H   ä Ay        € ê {          ì M        € ê y	      !   b z ÿ `  pPð È z ÿ a   Sð â „y	     H   X. „‰	 €   H   ¨ „‰	    H   " !r       È/ !‚       ÈO !‚       â G‰        € ê „y	 €  H   ( „y	 €   H   h „y	    H   ¨ „y
	    H   è „y	 €  H   " !r       â !r       Æ/ !r       âO !r
       Æ� !r       ä‰      " z ÿ `  pPð ä r 	ÿ   pRò ä z ÿ a   Sð â !r       Ê ‰      $ !r       Ê ‰ €    $ !r       Ê ‰ @    $ !r       Ê ‰      $ !r	       â G‰        € ê ‰      $ !r       Ê ‰      $ !r       Ê ‰ €    $ !r       Ê ‰ @    $ !r       Ê ‰      $ !r       ä M        € ê !r	       âG‰  `    € ê $tÿ   ÿ Ž â !r        È z  h  pð Ú G	  `    € ê $v ÿ h  ÿ Ž Ê x     pRð Ú G‰  �     € ê x     pRð Ú G	       € ê  x ;ª¸¿  A  Ì s          $ !t   €?    Ì s        $  r     A  È  r     A  â Gy  `    € ê  x '7=  A  â x*BL?  A  Æ #t   €?    È  r      A  â  x    ?  A  Ê s     $   $ !t  €?    È  r       A  È  r      A  â Gy  À    € ê $v ÿ h  ÿ Ž Ê x     pRð Ú G‰  à     € ê x     pRð Ú G	        € ê 	z  i    � ä 	z i    � Æ  x ;ª¸¿  A  â 	z i €  � Ê s        $ !t  €?    Ì s        $  r       A  È  r      A  â Gy  Ð     € ê  r     A  â Gy  °     € ê 	x   à@  � ä 	x  à@  � Æ  x #ÛÙ?  A  â 	x  àÀ  � Æ  x;ª¸¿  A  ä !t  €?    È s        $ !t  €?    Ì s        $  r       A  È  r      A  ä y       %   ( y      &   ( y      '   b $z  x   Ž ä $tÿ   ÿ Ž Ä $z  ~   Ž È/ %v  j   Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y       &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž ä $r	ÿÿ     Ž È %	 \   Ž Ê �	    ™ b ¹z  l      â Ey  À-    € â ™x?    â y#      "   b x       ÿ  â $r"ÿÿ   ÿ Ž ä y      !   ¢ $r6ÿÿ   ÿ Ž ä $ÿÿ   ÿ Ž â y%      %   " $ÿÿ   	 Ž È '
	 m   Ž È $‚ÿÿ   ÿ Ž â 
ÿ n    Ê $ÿÿ   
Ž â r #ÿ   pRò Æ/ $
	 o  	Ž ä $x$#    Ž äO $‚ÿÿ   	 Ž ä $x%%   ÿ Ž â xÿ   $  â 'Š	 s   Ž ä y      '   " r$   ÿ� È x4ÿ     È | 4   pbô ä Šÿ t    ä r #ÿ   pRð â$##€   Ž Ð G)  ð+    € ê $rÿÿ   ÿ Ž â xøÿÿÿÿÀŽ â $z		 w  ÿŽ ä 'z y   Ž È $x$$   
Ž â zÿ z    â $z% p  4Ž ä $x$   ÿ Ž â x-ÿ   $  â $z |  Ž ä $z( }  ÿŽ â rÿ   ÿ3Ž â $z v  Ž Ä$v'ÿ X  ÿ Ž â r(	(   ÿàõ â %x   ÿŽ È %xn   ÿŽ È $v+ÿ Y  ÿ Ž â x&J   áÇ â $r ÿÿ   ÿ Ž â r$$   ÿ� â $x4   ÿ Ž â r)ÿ   …þ â$x$   ÿ Ž âxÿ     â $r0ÿÿ   ÿ â r.$   ÿà÷ â%x   Ž â r,   ÿàù Ä x/   ÿÀŽ â%x   Ž â r2   ÿ Æ $x*   -Ž ä r'.   ÿàõ Ê $x+   * â r
'   ÿàõ È �y>B   ‘ ¢ $x   + Ê �y3
   • è  �y
   • ( �y<
   • h �y=
   • " $v_ÿ p  ÿ Ž È %x_n   Ž Ê �yB   ‘ b %x_n   
Ž Ê �y   • h �y   • h �y   • h �y   • b %x_n   Ž Ê �yXB   ‘ b %x_n   Ž Ê �y   • h �y   • b r
'   ÿàõ Æ �y^   • d $x	   + ä �yY   • h �y[
J   • b $t1ÿ   ÿ Ž ä %x_n   Ž Ê/ �yQ   • h �yK   • h �y7   • b %x
_n   
Ž È $x>   ÿ Ž ÊO x   ÿÀŽ â $x>@   ÿ Ž Æ xÿ   3øŽ â� $x>€   ÿ Ž È %v  1 Ž â x   ÿÀŽ ä x3ÿ   3  ä x8ÿ   øŽ â�yL    ™ ¢  r'&   ÿàõ ä x3   øŽ â $x:>    ÿ Ž Ä $x+   ) â xÿ     â %v  1 Ž Æ �yO    • " %v88  1 Ž âx   :øŽ ä �yM    ™ ¨  �yN8    ™ ¢  %v  1 Ž Ê �yP    ™ ¢ %x:_n   Ž Æ �y8   • ( �y3:B   ‘ " $x>   ÿ Ž Ê x   ÿÀŽ â $x>   ÿ Ž Æ xÿ   <øŽ â�y:
J   • b x<ÿ   <  È x<   øŽ â/ $x>   ÿ Ž ä %v  1 Ž È %v  1 Ž âx   ÿÀŽ â �y5    ™ h �y9    ™ b xÿ   =øŽ â$x>>   ÿ Ž â xÿ   =  ä r$(   àµ â %v  1 Ž â x   >øŽ Ä/ r-0   ÿd~ â $t>ÿ$   ÿ Ž â �y<    ™ b %v  1 Ž È %v Z  > Ž â �y;    ™ f $x$   ÿŽ È $x   Ž ä $x   ÿ Ž Æ �y=   ™ d x   ÿÀŽ ä �y>   ™ d xÿ   øŽ ä �y?   ™ b $x@   ÿ Ž ä/ %v  1 Ž â �y@   ™ d x   ÿÀŽ Ä �yA   ™ b $x€   ÿ Ž Æ �yB    ™ b xÿ     ä xÿ   øŽ â %x_n   Ž â� x   øŽ â �yC   ™ ä  %v  1 Ž ä�yF    • ä  %v  1 Ž Ä �yE    ™ è  �yD    ™ â  $x    ÿ Ž âxÿ     â %x
_n   
Ž â �yS   ™ ä  x   øŽ â/ $x   ÿ Ž â �yR
J   • è x   ÿÀŽ â %v  1 Ž â�yT    ™ ä xÿ   øŽ â$x   ÿ Ž â xÿ     â �yG    ™ ä  %v  1 Ž â x   øŽ â �yV    • ä $x   ÿ Ž Ä %v  1 Ž â �yH    ™ ä  x   ÿÀŽ ä �yI    ™ ä xÿ   øŽ â $x   ÿ Ž È %v  1 Ž â xÿ     È �yJ    ™ â  x   øŽ Ê %v  1 Ž Ê �yU    ™ â $xX   ÿ Ž Ê x   ÿÀŽ â $xX€   ÿ Ž Æ xÿ   øŽ â$xX@   ÿ Ž âxÿ     â $xX    ÿ Ž â xÿ     â/ %v  1 Ž âx   øŽ ä x   ÿÀŽ ä x   øŽ â �yW    ™ â  %v  1 Ž â xÿ   øŽ È �yZ    ™ â %v  1 Ž â r',   ÿàõ Æ %x_n   Ž â r'   ÿà÷ â �y    ™ ä  %v  1 Ž ä �ya    • ä $x+   2 ä �y    ™ â $x   +Ž Ä %x_n   Ž â xb[ÿ   ÿÀŽ â �y\j   ‘ ä  %x_n   Ž ä�y`j   ‘ ä  %x_n   Ž ä�y]    • ä  %x_n   Ž Ä/ �y    • ä $xX   ÿ Ž ä %x
_n   
Ž â�y    • æ  %x_n   Ž È %x_n   Ž È %x_n   Ž â x_   ÿÀŽ ä �y    • â $xc[ @  bŽ â xd_ÿ   ^øŽ äx[ÿ   [  â $xc€   ÿ Ž â x^ÿ   ^  â $x_X   ÿ Ž â �ybj   ‘ â %vd  1 Ž â x€€ÿÀŽ Ä/ xd^   _øŽ â �y
J   • â  x€€€€ÿüŽ Æ �y    ™ â xÿþþþÿàÿ â %vd  1 Ž ä �y    • â r^   ÿüŽ â $x
X   ÿ Ž ä $x[ @  [Ž â �y    ™ â  $xcc   ÿ Ž â x

   ÿÀŽ â $xXX   ÿ Ž â xÿ   Y  â$x€   ÿ Ž â/ x

ÿ   YøŽ â �yj   ‘ â xcc€€ÿÀŽ Ä xX   XøŽ â $x   ÿ Ž ä %v

  1 Ž âx€€ÿÀŽ ä/ xc€€€€ÿüŽ â %vX  1 Ž â x€€€€ÿüŽ â �y
    ™ â xÿþþþÿàÿ Ä xX€€ÿÀŽ ä xÿþþþÿàÿ â �y    ™ â  x^˜º  ÿ    ä rc   ÿüŽ ä xYX€€€€ÿüŽ ä r   ÿüŽ Ä x
€€€€LÞŽ äOxÿÀŽ ä x˜º  ÿ    ä xYYÿþþþÿàÿ ä x˜º  ÿ    ä xOÿ   ÿÀŽ â$x
   
Ž â xÿÀŽ Ä rYYX   ÿüŽ ä x€€€€MÞŽ â $xO @  Ž â x
ÿÀŽ äx€€€€NÞŽ â $x   
Ž â xYY˜º  ÿ    â $x€   ÿ Ž ä $x   

Ž â x
YÿÀŽ Ä xYY€€€€PÞŽ ä xÿ   O  ä x€€ÿÀŽ â $xOY   

Ž â xL€€€€L´Ž â $x   ÿ Ž â x
€€€€ÿüŽ â $x @  Ž Æ x

ÿþþþÿàÿ â $x€   ÿ Ž âx€€ÿÀŽ ä xM€€€€M´Ž â $x   ÿ Ž â xN€€€€N´Ž ä r
   ÿüŽ â $x3   ÿ Ž â x€€€€ÿüŽ Ä x€€ÿÀŽ ä xOO€€€€P´Ž ä xÿþþþÿàÿ ä x
€€€€ÿüŽ ä x€€ÿÀŽ ä xP   ÿÀŽ Ä r[   ÿüŽ â $x3@   ÿ Ž â x
ÿþþþÿàÿ ä x€€€€ÿüŽ ä x
Pÿ   QøŽ â$x^3€   ÿ Ž â xQÿ   Q  ä xÿþþþÿàÿ Ä rX   ÿüŽ â %v

  1 Ž â xd   ÿÀŽ ä xQ   ^øŽ ä rY   ÿüŽ ä xdÿ   KøŽ äxÿ   K  Ä �yK
    ™ ¢  %v  1 Ž È $xP3    ÿ Ž â�yQ    ™ " $x^3   ÿ Ž Æ xd   PøŽ â %v  1 Ž âxf^   ÿÀŽ ä x^˜º  ÿ    â %vd  1 Ž âxffÿ   7øŽ â�yP    ™ ¢ x7ÿ   7  â $xh3   ÿ Ž ä$xc3   ÿ Ž â �y    ™ ¢
 %v
f  1 Ž â x7   høŽ Ä/ xcc   ÿÀŽ ä �y7
    ™ ¢  %v  1 Ž â x[˜º  ÿ    ä xfcÿ   8øŽ â$xh3   ÿ Ž â x_^€€€€5ÞŽ â�y    ™ b x[ÿ   8  Ä x^^ÿÀŽ ä xdÿÀŽ äx3€€€€9ÞŽ â %vf  1 Ž â x
[   høŽ Æ $x^_   ^
Ž â xX˜º  ÿ    â $xd3   d
Ž â �y    ™ ¢  x3:ÿ   ÿÀŽ â %v

  1 Ž â x8^€€€€5´Ž ä x^ÿÀŽ â$x3: @  3Ž â x€€€€<ÞŽ â �y

    ™ ¢  xXY˜º  ÿ    â $x3€   ÿ Ž Ä/ $x   ^
Ž â x:ÿ   :  ä x1XÿÀŽ äxXX€€€€;ÞŽ ä x<€€€€<´Ž â $x3   ÿ Ž â x€€ÿÀŽ â $x3: @  :Ž Ä &rLL=   ÿ   â x€€€€ÿüŽ â $xXX   1
Ž ä $x13€   ÿ Ž â x:ÿþþþÿàÿ â &rLM>   L   â x€€ÿÀŽ ä x11€€ÿÀŽ â &rLN?   L   â r::   ÿüŽ Ä x€€€€ÿüŽ â &rLO@   L   â x1€€€€ÿüŽ ä x:˜º  ÿ    ä xNÿþþþÿàÿ â &r8A   L   â x:ÿþþþÿàÿ ä x€€€€BÞŽ Ä x8ÿÀŽ ä r::1   ÿüŽ Æ $x   8
Ž â rNN   ÿüŽ ä x::˜º  ÿ    ä xB€€€€B´Ž â $x3   ÿ Ž â xN˜º  ÿ    ä x:ÿÀŽ äx::€€€€EÞŽ Ä� x€€ÿÀŽ ä x8ÿÀŽ â$x::   
Ž â x€€€€DÞŽ ä x1Fÿ   ÿÀŽ äx€€€€ÿüŽ â $x   8
Ž ä $x1F @  1Ž â x8ÿþþþÿàÿ Æ $x1€   ÿ Ž âr88   ÿüŽ â $x1   ÿ Ž â xFÿ   F  â &rBB=   ÿ   â x€€€€D´Ž ä x€€ÿÀŽ â $xFF @  FŽ â x::€€€€E´Ž â &rB>   B   â x€€ÿÀŽ Ä x€€€€ÿüŽ â $x3F€   ÿ Ž â x88˜º  ÿ    â &r::?   B   â x1€€€€ÿüŽ ä xBÿþþþÿàÿ ä x8ÿÀŽ äx88€€€€GÞŽ Ä xD1ÿþþþÿàÿ ä x33€€ÿÀŽ ä rBB   ÿüŽ â $x88   
Ž â rDD   ÿüŽ ä x3€€€€ÿüŽ ä xB˜º  ÿ    Ä xDD˜º  ÿ    ä xBÿþþþÿàÿ â $xF   ÿ Ž â xG8€€€€G´Ž ä x€€€€HÞŽ äx8ÿÀŽ ä xDÿÀŽ Ä xDD€€€€IÞŽ ä rBB3   ÿüŽ ä x1€€ÿÀŽ â $xDD   
Ž â xB˜º  ÿ    â $x   8
Ž â x1€€€€ÿüŽ ä xBÿÀŽ Ä x€€€€JÞŽ ä x8ÿþþþÿàÿ â &r:G@   :   â x€€€€H´Ž ä x3Rÿ   ÿÀŽ â $x   B
Ž â r881   ÿüŽ â &rA   :   â xDD€€€€I´Ž â $x3R @  3Ž â x88˜º  ÿ    Ä x€€€€J´Ž â &rDDC      â x8ÿÀŽ â$x3€   ÿ Ž âx88€€€€UÞŽ ä xRÿ   R  â $x13   ÿ Ž â x€€ÿÀŽ â $x88   
Ž ä $xRR @  RŽ â x11€€ÿÀŽ Ä x€€€€ÿüŽ â $x5R€   ÿ Ž â xU8€€€€U´Ž ä x31€€€€ÿüŽ ä x8ÿþþþÿàÿ â $xR   ÿ Ž â xdd€€€€9´Ž ä x55€€ÿÀŽ Ä x:3ÿþþþÿàÿ ä r88   ÿüŽ â &rdC      â x3€€ÿÀŽ ä x5€€€€ÿüŽ ä r::1   ÿüŽ ä x88˜º  ÿ    â &r<S      â x13€€€€ÿüŽ Ä x<ÿþþþÿàÿ ä x:˜º  ÿ    ä x8€€€€WÞŽ äx88ÿÀŽ ä xB1ÿþþþÿàÿ ä r<<5   ÿüŽ Ä x:ÿÀŽ â $x8   8
Ž â x€€€€ZÞŽ ä rBB3   ÿüŽ ä x<˜º  ÿ    â $x   :
Ž â x1B˜º  ÿ    ä x:ÿÀŽ Ä x€€€€ÞŽ ä x88€€€€W´Ž ä xBaÿ   ÿÀŽ ä x<1ÿÀŽ â $x   :
Ž â x11€€€€ÞŽ â &r88=   ÿ   â x€€€€Z´Ž â $xBa @  BŽ â x€€€€´Ž â $x11   <
Ž Ä &r8>   8   ä $xB€   ÿ Ž âxaÿ   a  â &r?   8   â x11€€€€´Ž â $xB   ÿ Ž â x€€ÿÀŽ â $x5a @  aŽ ä &r1@      â x€€€€ÿüŽ â $x35€   ÿ Ž â x1€€ÿÀŽ Ä xÿþþþÿàÿ ä x1€€€€ÿüŽ ä x33€€ÿÀŽ ä r   ÿüŽ ä x8ÿþþþÿàÿ ä x3€€€€ÿüŽ Ä x˜º  ÿ    ä r881   ÿüŽ â $x55   ÿ Ž â x:ÿþþþÿàÿ ä x€€€€ÞŽ äxÿÀŽ ä x8˜º  ÿ    Ä x55€€ÿÀŽ â $x   
Ž â xÿÀŽ äx€€€€ÞŽ ä x95€€€€ÿüŽ ä x€€€€´Ž â $x   
Ž â x<9ÿþþþÿàÿ Ä r::3   ÿüŽ â &rA      â r<<5   ÿüŽ ä x€€€€´Ž ä x::˜º  ÿ    ä x<<˜º  ÿ    â &rC      â x1:ÿÀŽ Ä xÿ   ÿÀŽ ä x::€€€€ÞŽ ä x<ÿÀŽ äx<<€€€€ÞŽ â $x @  Ž â r\ÿ/   \  â $x::   1
Ž â xÿ     â $x<<   
Ž Ä $x€   ÿ Ž â x::€€€€´Ž â $tÿ   ÿ Ž â x<€€€€´Ž â $x\\   ÿ Ž â x€€ÿÀŽ â $x @  Ž ä $x   ÿ Ž â x\\   ìŽ â $x€   ÿ Ž â x€€€€ÿüŽ Ä x€€ÿÀŽ â &r:S      â xÿþþþÿàÿ ä x€€ÿÀŽ ä x€€€€ÿüŽ ä x8ÿ   ÿÀŽ â &rT      â x€€€€ÿüŽ Ä xÿþþþÿàÿ â $x8 @  8Ž â r   ÿüŽ ä xÿþþþÿàÿ â $x   ÿ Ž â r   ÿüŽ â $x8€   ÿ Ž âxÿ     ä r`ÿ/   `  â $x8   ÿ Ž â x€€ÿÀŽ â $x5 @  Ž â x€€ÿÀŽ â $t1ÿ   ÿ Ž â r   ÿüŽ â $x`   ÿ Ž â x€€€€ÿüŽ â $x35€   ÿ Ž â x€€€€ÿüŽ Ä x€€ÿÀŽ ä x   1ìŽ ä xÿþþþÿàÿ ä xÿþþþÿàÿ ä x1€€€€ÿüŽ ä x33€€ÿÀŽ Ä r   ÿüŽ ä r   ÿüŽ ä x1ÿþþþÿàÿ ä x3€€€€ÿüŽ ä x˜º  ÿ    ä r   ÿüŽ Ä x8ÿþþþÿàÿ ä xÿÀŽ äx€€€€KÞŽ äO x˜º  ÿ    Æ $x   
Ž â xÿÀŽ äx€€€€QÞŽ äx˜º  ÿ    Æ $x   
Ž â x˜º  ÿ    ä xÿÀŽ äx€€€€PÞŽ ä xQ€€€€Q´Ž â $x15   ÿ Ž â xÿÀŽ äx€€€€ÞŽ â $x   
Ž â x€€€€K´Ž Ä x˜º  ÿ    â $x   
Ž â x11€€ÿÀŽ â &r=   ÿ   â xÿÀŽ äx˜º  ÿ    ä x€€€€7ÞŽ ä x51€€€€ÿüŽ Ä r883   ÿüŽ ä x€€€€´Ž â &rQ>      â x€€€€P´Ž ä xÿÀŽ â$x   
Ž â x€€€€ÞŽ äx:5ÿþþþÿàÿ Ä x8˜º  ÿ    â &r?      â r::1   ÿüŽ â $x   
Ž â xÿÀŽ â &r@@      â x€€€€ÞŽ ä x:˜º  ÿ    ä x€€€€7´Ž â $x   
Ž â xÿÀŽ Ä x€€€€
ÞŽ â &rA   @   â x€€€€´Ž ä rbÿ/   b  â $x   
Ž â x€€€€´Ž â &rC      â rÿ/     â $tÿ   ÿ Ž â x€€€€
´Ž â $xbb   ÿ Ž â x;X€€€€;´Ž â &rS      â x44   ÿàÿ â $x   ÿ Ž Ä &rS   D   â xbb   ìŽ â&rT      âx   ìŽ â &r;T      â| 4   pbô â &rUT      ä $r\   ÿŽ ä $r   ÿŽ Ä $rb   ÿŽ ä $r   ÿŽ ä 0r]ÿ]    A   ä 0rVÿV    A   ä 0rÿ    A   ä 0rÿ    A   Ä 0rÿ    A   â x''à  ÿà÷ â  r]]V     A  â Er        â  rV     A  âEr        â  r
V     A  âEr        â  rV     A  â Er        â #r6]   6   â #r"   "   â #r!
   !   â x€   ÿàÿ â #r        â $r+ÿÿ   +Ž â D)       À â Gy   Öÿÿÿÿƒ ê Ay        € ê ˆ #6 þÿ H   è ˆ #"€þÿ H   è ˆ #! ÿÿ H   è ˆ # €ÿÿ H   è {          ì M	        € ê „y     H   "z % r  p`ô Æ „y	 €   H   b r ÿ   pVt Æ „y    H   ¨ „y    H   è „y €  H   ( „y    H   ( „y    H   ( „y €  H   ( „y €  H   ( „y    H   ( „y €  H   " !r6       Æ „y
 €  H   " !r		"       â/ !r!       âO !r       ä� !r	       â!r       â !r       â !r        È ‰      b !r       â !r       È ‰      ¢ !r       Æ ‰      â !r
       Ê ‰      " !r       â/ !r       ÈO ‰      b !r
       Æ� ‰      ¨ ‰	
      â !r       Ê ‰      " !r       â/ !r       ÈO ‰ €    b !r
	       Æ� ‰	 €    ¨ ‰ €    â !r       Ê ‰ €    " !r       â/ !r		       âO !r
       È� ‰	 @    h ‰ @    ¢ !r       Æ ‰
 @    ( ‰ @    â !r	       â/ x	%   ÿàÿ â !r       ÆO ‰      b $z  x  %Ž â x %   ÿàÿ â!r
       â x%   ÿàÿ â ‰      " z   r  p`ö â !r       â� z  r  p`ð â ‰      ¢ z 	 r  p`ò â $tÿ   ÿ Ž â x    pVö â‰
      â x    pVp  â$z ~  Ž â x    pVò  Æ %v j   Ž â !r       â/ !r       È †¹    â !r       ÆO †©     â !r
       Æ� †‰    â M        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      &   " z ÿ \  pPò â ¹z  F   
   Æ z ÿ ]  Sò Ú $ÿ   ÿ Ž È % \   Ž Ê �	    ™ ¢  $rÿÿ   ÿ Ž â Ey  0    € â $r.ÿÿ   ÿ Ž â y/      %   b ' m   Ž Æ y      "   â $r0ÿÿ   ÿ Ž â ÿ n    â 'š s   Ž â y       !   & $ÿÿ   
Ž â y      '   & $ o  Ž â ’        â $x//   ÿ Ž â/ r ÿ   pRð È� x     pDp  â$x     Ž â r
/    ÿàÿ È z 
 r  pdp  ä xÿ     ä š	ÿ t    È r	     € ÊO G	  À    € ê z ÿ `  pPò â $r0ÿÿ   ÿ Ž â z ÿ ^  pPð â $r.ÿÿ   ÿ Ž â z ÿ b  p ô ä z ÿ a  Sò ä z ÿ _   Sð ä z ÿ c   'ò Ö $
 ~  ÿŽ âÿ   /  â$š ~  ÿŽ ä $š x  ÿŽ â/    àë â$
 x  ÿŽ â ’
/    àÇ ä ˜ÿ   /  ä ÿ   ÿÄþ Ä ’
   ÿàû ä    ÿàõ ä ’
ÿ   ÿ„þ ä š b  ÿ† â $ÿÿ    â 
 ^  ÿ„ â $’ÿÿ   
Ž Æ 
 _   ä š c  � Æ �	0     h �™.     d Ay        € ê r   ÿ� â ¹z  l      â $z1 }  ÿŽ â ™x?    â x:ÿ     â $z w  ÿŽ â r ÿ       â Ey  ðY    € â x2      ÿ  â | :   pbð â $r]ÿÿ   ÿ Ž â r11   ÿàó â 'z y   Ž â x6      ÿ  Ä r9ÿÿ   ÿäÿ  â $raÿÿ   ÿ Ž ä $r5ÿÿ   ÿ Ž ä $r8ÿÿ   ÿ Ž ä G	  0Y    € ê zÿ z    â $z/ p  :Ž âxøÿÿÿÿÀŽ â $tBÿn   ÿ Ž â x/   ÿàÿ â $z |  :Ž âzK  Y      â $x   
Ž â x2      ÿ  â $z |  Ž â x6      ÿ  â $z;	 v  Ž â x/   ÿàÿ â $x   ÿ Ž ârV   ÿ� â $z	 v  Ž â x/   ÿàÿ â $z p  ;Ž ârÿ   ÿ3Ž â $z p  ;Ž âxVÿ   V  â $z	 p  ;Ž â xÿ     â %x   ÿŽ â xWÿ   V  Ä xD:   ÿ   â %v@  € BŽ È %v	  € BŽ â r<@  €ÿàñ Æ %vB  € BŽ â r=   ÿáó Æ %xn   ÿŽ â r>B   ÿáù Æ %x
n   ÿŽ â x?J   áµ Æ $x   ÿ Ž ä$x@A     â rA   ÿåÿ  â %x	n   VŽ ârH
   ÿàñ ärP   ÿàó â %x	n   ÿŽ â xF   ÿÀŽ ä rX   ÿäÿ  â %x   Ž È $xBC    â rCÿ   e~ â %x   ÿŽ â rN   ÿàõ ä rLV   ÿà÷ â $xJ     â xR   ÿàñ â %x   
Ž È %x   ÿŽ È %xn   Ž È %x
   
Ž È %xn   VŽ È %xn   VŽ È %x   Ž È $xT    ä $vEÿ `  ÿ Ž ä $vGÿ a  ÿ Ž ä $vIÿ X  ÿ Ž ä $r]ÿÿ   ÿ Ž ä $raÿÿ   ÿ Ž Ä $r5ÿÿ   ÿ Ž ä $r8ÿÿ   ÿ Ž ä %x   Ž È %x   Ž È $xV   WŽ ä $rZÿÿ   	  ä r IP   ÿàó ä rI   ÿàñ Æ $x!K   XŽ  ä $x	   K  Æ �y$ B   ‘ ¨  �y4   • è �y%   • ( �y)   • ( �y+   • " $vÿ p  ÿ Ž â rIR   ÿàñ Æ %x"n    Ž â rKZ   ÿä  È �y{"B   ‘ " %xn   Ž Ê �y(þÿÿ • ( �y*    • ( �yt   • ( �y-   • " %xn   "Ž È/ %x&n   Ž â �yvB   ‘ ( �y,&þÿÿ • ( �y}&    • " r I   ÿàñ Æ �yz&   • $ $x!   K  ä �yu&   • ( �yw J   • " $tMÿ   ÿ Ž ä %x&n   &Ž Ê/ �ym&    • " %x n    Ž È $x#$€   ÿ Ž âOxÿ   4  â� $x"$   ÿ Ž Æ x   #øŽ ä x"   ÿÀŽ Æ %vj  M Ž â rI?   ÿàñ ä xnÿ   4øŽ â $x4$    ÿ Ž âx#ÿ   %  â�yjj    ™ ¢  x"$   ÿ   â $xK   C  â x4#   4øŽ â %vnn  M Ž â x""   ÿÀŽ Ä �yg    • " %vb4  M Ž âx""ÿ   %øŽ ä �ynn    ™ ¦  %v^"  M Ž â �yhb    ™ ¨ �yi^    ™ " $x"$   ÿ Ž â r#1   Dà‘ Ä �yk&þÿÿ • $ x""   ÿÀŽ â $t4ÿ$   ÿ Ž â r%9   ÿ$~  ä xd"ÿ   )øŽ â %xbn   Ž â/ �y_&   • & %vdd  M Ž â �y`bB   ‘ & %v"# Z  4 Ž â �yld    ™ & $x%%$   ÿŽ Ê r##%   ÿàÿ â $x$   ÿ Ž â x)ÿ   )  â �ye&   • ( �yO"   ™ " x)   øŽ Æ �yQ"   ™ " $x%$   ÿ Ž ä %v  M Ž â �yS"   ™ $ x4%   ÿÀŽ ä �y^    ™ $ x44ÿ   +øŽ Ä �yU"   ™ & %vb4  M Ž â/ x+ÿ   +  â �yd J   • ( �ybb    ™ " $x$$   ÿ Ž Æ �yW"   ™ $ x$+   $øŽ ä �yY"   ™ & %v$$  M Ž â �y["   ™ ( �y4$    ™ " rIL   ÿàñ Ä x&{   ÿ   â$xf{€   ÿ Ž â �y\"    ™ " $xK   V  â x'&   ÿÀŽ È �ycj   ‘ " x&'ÿ   (øŽ äx'ÿ   (  Æ %v(&  M Ž â x$'   føŽ Æ/ $x&{@   ÿ Ž â �yf(    ™ " %v$$  M Ž Æ x'&   ÿÀŽ â %xn   Ž â� �yo$    ™ $ x&'ÿ   *øŽ â$x+{    ÿ Ž â x*ÿ   *  â �yp    • & x(*   +øŽ Ê/ %v((  M Ž È %v&&  M Ž â �yr(    ™ ( �yq&    ™ " $x*{   ÿ Ž Ê x%*   ÿÀŽ È� x*%ÿ   tøŽ â$x${   ÿ Ž â xtÿ   t  â %x n    Ž â x%{   ÿ   Æ %v**  M Ž â x&$   ÿÀŽ ä x$t   %øŽ ä �ys*    ™ â  x&&ÿ   -øŽ Æ �yt J   • â  %v$$  M Ž È %v&&  M Ž â �yx$    ™ è  �yy&    ™ â  $x({   ÿ Ž â/ x-ÿ   -  â $x*v   ÿ Ž â �y{"    • ä  x(-   (øŽ Ä x+*   ÿÀŽ â $x~v€   ÿ Ž â x-ÿ   ,  â%v((  M Ž âx,+ÿ   ,øŽ ä x*-   ~øŽ ä �y|(    ™ â %v,,  M Ž â x%v   ÿ   È x&%   ÿÀŽ â �y,,    ™ â $x~v    ÿ Ž â x#ÿ   }  â%v**  M Ž â x"&ÿ   }øŽ Æ %xn   Ž â x†#   ~øŽ â �y}*    ™ â  r$I   ÿàñ â %v""  M Ž ä �y-    • ä" $x%   K  ä �y€"    ™ â  %xn   Ž â x„wÿ   ÿÀŽ Ä �y~$    • â %x&n   $Ž Æ �y�j   ‘ â %x(n   Ž Æ �y‚&    • â %x*n   &Ž Æ �yƒ(j   ‘ â  %x$n   Ž È/ %v†  M Ž âx'v   ÿ   â �y$$    • ä $xˆw @  „Ž ä %x n    Ž â�y„*    • æ %x"n   (Ž â�y& J   • æ %xn   *Ž â x'   ÿÀŽ â �y(    ™ ä $x…ˆ€   ÿ Ž â x†ÿ   zøŽ â $x‡v   ÿ Ž âxwÿ   w  â $x+v   ÿ Ž â/ x……€€ÿÀŽ â %v †  M Ž â �y)    • ä x'…€€€€ÿüŽ Ä xzÿ   z  â �y%     ™ â x*'ÿþþþÿàÿ ä xzz   ‡øŽ â �y'"j   ‘ â r…*…   ÿüŽ â $x*ˆ   ÿ Ž ä %vz  M Ž â xz+   ÿÀŽ Æ $xww @  wŽ â x"*€€ÿÀŽ â/ $xvv   ÿ Ž â �y*    ™ â  xzzÿ   uøŽ â$x+w   ÿ Ž â x"€€€€ÿüŽ ä xuÿ   u  ä x!ÿþþþÿàÿ Ä xvu   vøŽ â %vz  M Ž âxu+€€ÿÀŽ ä r!!"   ÿüŽ ä x"u€€€€ÿüŽ â �y+    ™ â x#w   ÿ   â %vv  M Ž â xz"ÿþþþÿàÿ Ä x!!˜º  ÿ    ä x##€€ÿÀŽ ä rzzu   ÿüŽ ä �yu    ™ â  x"!ÿÀŽ äx!!€€€€jÞŽ äO x #€€€€ÿüŽ Ä x……˜º  ÿ    â $x!!   "
Ž â xv ÿþþþÿàÿ ä x"gÿ   ÿÀŽ äxz˜º  ÿ    ä/ x …€€€€nÞŽ ä rvv#   ÿüŽ Ä x……ÿÀŽ ä r#g"   ÿpŽ ä x"ÿÀŽ äxvv˜º  ÿ    ä x€€€€hÞŽ â $x…    …
Ž â x !€€€€j´Ž â $x#€   ÿ Ž â xgÿ   g  â $x   "
Ž â x!vÿÀŽ Ä xvv€€€€iÞŽ â $xgg @  gŽ â x€€ÿÀŽ ä xj€€€€h´Ž â $xv   !
Ž â x€€€€ÿüŽ ä x!#   ÿ   â $x#g€   ÿ Ž ä$xgg   ÿ Ž â x"ÿþþþÿàÿ Ä x€€€€i´Ž ä x!!€€ÿÀŽ ä xig€€ÿÀŽ ä r""   ÿüŽ ä x!€€€€ÿüŽ ä xhi€€€€ÿüŽ Ä x"˜º  ÿ    ä x"ÿþþþÿàÿ ä xvhÿþþþÿàÿ ä x€€€€lÞŽ äxhÿÀŽ â $x`   ÿ Ž â x##€€ÿÀŽ Ä x…€€€€n´Ž â $x   h
Ž â xg#€€€€ÿüŽ ä xh   ÿÀŽ â &rO   ÿ   â xngÿþþþÿàÿ ä r""!   ÿüŽ â &r  Q      â xhhÿ   køŽ â $x`€   ÿ Ž â rwn#   ÿüŽ Ä xkÿ   k  ä rnvi   ÿüŽ ä xl€€€€l´Ž â $x`@   ÿ Ž â x"˜º  ÿ    â %v"h  M Ž â xv`   ÿ   ä x!ÿ   m  Ä xk   øŽ â �yg"    ™ ¢  xz!   vøŽ ä xv   ÿÀŽ ä xhÿÀŽ äx€€€€^ÞŽ â %v  M Ž â xvvÿ   møŽ Æ $xk   h
Ž ä $x"`   ÿ Ž â �yh    ™ " xww˜º  ÿ    â &riS       ä x""   ÿÀŽ â %v v  M Ž â xvk€€€€^´Ž Æ %vz  M Ž â xwÿÀŽ äxww€€€€bÞŽ â $xk`   ÿ Ž âxz`   ÿ   ä x#ÿ   _  â &rjjU   i   â x""ÿ   _øŽ â �y^    ™ ¢  xz#   zøŽ â $xww   
Ž Ä �yi     ™ ¢ %v"  M Ž â xk   ÿÀŽ Æ $x``   ÿ Ž ä �y    ™ ¢  x ÿ   eøŽ â/%vz  M Ž âxeÿ   e  ä x#n˜º  ÿ    â %v    M Ž â x"e   `øŽ â �y    ™ ¢ xedÿ   ÿÀŽ Ä x`#ÿÀŽ â�y      ™ ¢  x_#€€€€4ÞŽ â %v""  M Ž È $xed @  eŽ ä �y""    ™ ¤ $xe€   ÿ Ž ä$x!e   ÿ Ž Æ x€€ÿÀŽ â/ &rjlW   j   â r__`  €ÿàÿ ä rcÿF   c  ä x€€€€ÿüŽ ä x!!€€ÿÀŽ ä xdÿ   d  â &rjvY   j   â xww€€€€b´Ž â $xcc   ÿ Ž â x__€€€€4´Ž â $t4ÿ   ÿ Ž â x`ÿþþþÿàÿ Ä x#!€€€€ÿüŽ ä rddd   ÿpŽ â &rjw[   j   â r``   ÿüŽ ä xb#ÿþþþÿàÿ â $x#d€   ÿ Ž â x4c   4ìŽ â &r_\   j   â x`˜º  ÿ    â $x_d   ÿ Ž â x##€€ÿÀŽ â $r4   ÿŽ â rbb!   ÿüŽ Ä x!€€€€fÞŽ äx4ÿÀŽ ä x#€€€€ÿüŽ ä x__€€ÿÀŽ ä xbb˜º  ÿ    â $x!!   4
Ž â x4ÿþþþÿàÿ Ä x`_€€€€ÿüŽ ä xbÿÀŽ äxbb€€€€oÞŽ ä x``ÿþþþÿàÿ Æ $xbb   
Ž â xpÿ   ÿÀŽ ä r``_   ÿüŽ ä r_p   ÿpŽ ä x`˜º  ÿ    ä r44#   ÿüŽ ä x`ÿÀŽ Ä x44˜º  ÿ    ä x€€€€rÞŽ â $x#_€   ÿ Ž â xpÿ   p  ä xf!€€€€f´Ž â $x   `
Ž â x!4ÿÀŽ äx44€€€€qÞŽ â $xpp @  pŽ â x##€€ÿÀŽ Ä xr€€€€r´Ž â $x44   !
Ž â x!#€€€€ÿüŽ â $xp€   ÿ Ž â x__   ÿ   ä x`!ÿþþþÿàÿ â &rffO   ÿ   â xbb€€€€o´Ž ä x!€€ÿÀŽ Ä x__€€ÿÀŽ ä r``#   ÿüŽ â &rbbQ   f   â x44€€€€q´Ž ä x#!€€€€ÿüŽ ä x_€€€€ÿüŽ ä x``˜º  ÿ    â &r44S   b   â xd#ÿþþþÿàÿ â $x#p   ÿ Ž â xbÿþþþÿàÿ Ä x`€€€€sÞŽ ä�x``ÿÀŽ ä rdd!   ÿüŽ ä x!tÿ   ÿÀŽ ä rbb_   ÿüŽ ä x##€€ÿÀŽ â $x`   `
Ž â xb˜º  ÿ    â $x_t @  !Ž â x!#€€€€ÿüŽ Ä xdd˜º  ÿ    ä xbÿÀŽ äx€€€€xÞŽ ä xf!ÿþþþÿàÿ ä x!dÿÀŽ äxdd€€€€yÞŽ â &r4rU   4   â x``€€€€s´Ž Ä rb  €ÿàÿ ä rff#   ÿüŽ â $x#_€   ÿ Ž â xx€€€€x´Ž â $xdd   !
Ž â xtÿ   t  â &r4`W   4   â x##€€ÿÀŽ â $x!_   ÿ Ž â xdd€€€€y´Ž â &r4xY   4   â x#€€€€ÿüŽ â $xtt @  tŽ â x!!€€ÿÀŽ â &rdd[   4   â x4ÿþþþÿàÿ Ä xf˜º  ÿ    ä x`t   ÿ   ä x_!€€€€ÿüŽ ä r#4#   ÿüŽ ä x4ÿÀŽ ä xb`€€ÿÀŽ Ä x€€€€|ÞŽ ä x`_ÿþþþÿàÿ ä x##˜º  ÿ    â $x   4
Ž â x_b€€€€ÿüŽ ä r``!   ÿüŽ ä x!#€€€€,ÞŽ Ä x4#ÿÀŽ ä x__ÿþþþÿàÿ ä x``˜º  ÿ    â $x!!   4
Ž â r__b   ÿüŽ â $x4t   ÿ Ž â x#`ÿÀŽ ä x__˜º  ÿ    Ä x``€€€€}ÞŽ ä x44€€ÿÀŽ ä xf-ÿ   ÿÀŽ ä x,!€€€€,´Ž ä xb_ÿÀŽ â $x``   #
Ž â x!4€€€€ÿüŽ Ä x__€€€€€ÞŽ ä rf-f   ÿpŽ ä x!!ÿþþþÿàÿ â $x__   b
Ž â x``€€€€}´Ž â $x#f€   ÿ Ž â r!!4   ÿüŽ â &r,,O   ÿ   â x__€€€€€´Ž Ä x##€€ÿÀŽ â &r,`Q   ,   â x!!˜º  ÿ    ä x4#€€€€ÿüŽ â &r__S   ,   â x,!ÿÀŽ äx!!€€€€(ÞŽ â $x`f   ÿ Ž â x44ÿþþþÿàÿ Ä x-ÿ   -  â $x!!   ,
Ž â r44#   ÿüŽ ä x``€€ÿÀŽ ä r---   ÿpŽ ä x(!€€€€(´Ž ä x44˜º  ÿ    â $x,-€   ÿ Ž âx!`€€€€ÿüŽ â $xb-   ÿ Ž â x#4€€€€%ÞŽ Ä x!!ÿþþþÿàÿ ä x44ÿÀŽ ä x-,€€ÿÀŽ ä xbb€€ÿÀŽ â $x4#   4
Ž â r!!`   ÿüŽ ä x#-€€€€ÿüŽ Ä x,b€€€€ÿüŽ ä x!!˜º  ÿ    ä x`#ÿþþþÿàÿ ä x#,ÿþþþÿàÿ ä x,!ÿÀŽ äx!!€€€€*ÞŽ Ä r``-   ÿüŽ Æ $x!!   ,
Ž â x``˜º  ÿ    ä r##b   ÿüŽ ä x*!€€€€*´Ž ä x!`ÿÀŽ äx``€€€€+ÞŽ ä x,#˜º  ÿ    Æ $x``   !
Ž â x!,ÿÀŽ äx,,€€€€uÞŽ Ê $x,,   !
Ž â x!&ÿ   ÿÀŽ â &r((U   _   â x44€€€€%´Ž ä r�ÿF   �  â $x#& @  !Ž â x&ÿ   &  â &r44W   (   â x€€€€|´Ž â $x��   ÿ Ž â x!#   ÿ   â0r~ÿ~    A   â x(        â 0r{ÿ{    A   â xu,€€€€u´Ž â $x##   ÿ Ž â x!!€€ÿÀŽ â $x,& @  &Ž â Er        â  r~~{     A  â x(�   (ìŽ â &r\   d   â x``€€€€+´Ž â $x+,€   ÿ Ž â x%#€€ÿÀŽ Ä x#!€€€€ÿüŽ â #ra~   a   â $r(   ÿŽ â rƒÿF   ƒ  ä x(#ÿþþþÿàÿ â $t-ÿ   ÿ Ž â x++€€ÿÀŽ â &r**Y   4   â x&ƒ   ÿ   ä r((!   ÿüŽ Ä x!+€€€€ÿüŽ â &r*`[   *   â x&&   -ìŽ â $x#,   ÿ Ž â x-$ÿ   ÿÀŽ äx,!ÿþþþÿàÿ â &ru\   *   â x!ÿ   $  â $x-$ @  -Ž â x*%€€€€ÿüŽ Ä x##€€ÿÀŽ â $x`! @  !Ž â x**ÿþþþÿàÿ ä x$-   ÿ   â $x4`€   ÿ Ž â x!#€€€€ÿüŽ ä r**%   ÿüŽ â $x%-   ÿ Ž â r,,+   ÿüŽ Ä x+$€€ÿÀŽ ä x$!ÿþþþÿàÿ ä x44€€ÿÀŽ ä x!+€€€€ÿüŽ ä x-%€€ÿÀŽ ä r%$#   ÿüŽ Ä x#4€€€€ÿüŽ ä x((˜º  ÿ    ä x$!ÿþþþÿàÿ ä x!-€€€€ÿüŽ ä x_#ÿþþþÿàÿ ä x#(ÿÀŽ Ä r+$+   ÿüŽ ä x((€€€€gÞŽ äO x$!ÿþþþÿàÿ ä x!*˜º  ÿ    ä r((#  €ÿàÿ â $x#`   ÿ Ž â r-$-   ÿüŽ Ä x$!ÿÀŽ äx!!€€€€hÞŽ äx*#€€ÿÀŽ ä x,,˜º  ÿ    â $x!!   $
Ž â x$*€€€€ÿüŽ ä x%%˜º  ÿ    Ä x#,ÿÀŽ äxh!€€€€h´Ž ä x,,€€€€iÞŽ ä x!$ÿþþþÿàÿ ä x$%ÿÀŽ ä x((€€€€g´Ž Ä x%%€€€€^ÞŽ ä x++˜º  ÿ    â $x,,   #
Ž â r%%$  €ÿàÿ â &r((O   ÿ   â r__4   ÿüŽ ä x--˜º  ÿ    ä x$+ÿÀŽ Ä x++€€€€ÞŽ â &r(hQ   (   â r!!*   ÿüŽ ä x,,€€€€i´Ž ä x*-ÿÀŽ ä x__˜º  ÿ    â $x++   $
Ž â x--€€€€ÞŽ â &r(,S   (   â x^%€€€€^´Ž Ä x$_ÿÀŽ â $x--   *
Ž â x__€€€€ ÞŽ ä x!!˜º  ÿ    â &r(^U   (   â x++€€€€´Ž â $x__   $
Ž â x!ÿÀŽ ä x-€€€€´Ž â &r(+W   (   â x!!€€€€"ÞŽ Ä r'ÿF   '  ä z ÿ `  pPð â &rY   (   â x__€€€€ ´Ž ä r!!  €ÿàÿ â $t ÿ   ÿ Ž â z ÿ a   Sð â $x''   ÿ Ž â x!!€€€€"´Ž â &r_[      Ä 0r‚ÿ‚    A   â x '    ìŽ â &r!!\      â Er        â 0rÿ„    A   â  r‚{‚     A  â 0r"ÿ)    A   ä $r&   ÿŽ ä $r !    ÿŽ â #r]‚   ]   â  r{     A  â r{"     A  â Er        Ä Er         â G‰  `
    € ê r$EP   ÿàõ ä r E   ÿàó Æ $x%G   X â r!	G   ÿäÿ  È �y_$B    ¨ �y&     è  �y(     ( �y4     ( �y^     " r"E   ÿàó Ê $x#   GŽ  Ê �y`"J    " r E?   ÿàó â $x'_   ÿ Ž âOx!_   ÿ   â$x*_€   ÿ Ž â x)ÿ   &  ä� x''   ÿÀŽ ä x$)   *øŽ ä x&'ÿ   &øŽ â $x#_    ÿ Ž â/ x!!   ÿÀŽ â %v$$  M Ž â x"ÿ   (  Æ%v&&  M Ž â x*!ÿ   (øŽ â �y$$    ™ ¢  x""   #øŽ Æ �yf&    ™ â $x!G   CŽ  ä %v**  M Ž Æ �y#      â %v,"  M Ž Æ �y(*    ™ è  �y),    ™ â $x"_   ÿ Ž Ê x%"   ÿÀŽ â $x"_   ÿ Ž Æ x&%ÿ   4øŽ ä/x4ÿ   4  ä x%_   ÿ   â%v&&  M Ž â x!"   ÿÀŽ äx44   %øŽ â $x__   ÿ Ž â �y &    ™ " x"ÿ   ^  Ä x,!ÿ   ^øŽ â %v*4  M Ž â x""   _øŽ Æ %v,,  M Ž â�y!*    ™ " rbEL   ÿàó ä %v^"  M Ž ä �y",    ™ $ $xcG   VŽ  ä �y%^    ™ ( �y'bj    " rdE   ÿàó È reG   ÿäÿ  Ê �y&d     " xg`ÿ   ÿÀŽ Ê $xg` @  gŽ â x`ÿ   `  Æ $x*g€   ÿ Ž ä/$x+g   ÿ Ž â r^``   ÿpŽ ä x**€€ÿÀŽ ä x,+€€ÿÀŽ â $x4^€   ÿ Ž â x+*€€€€ÿüŽ â $x^^   ÿ Ž â x-,€€€€ÿüŽ Ä x++ÿþþþÿàÿ ä x44€€ÿÀŽ ä x--ÿþþþÿàÿ ä r++*   ÿüŽ ä x^^€€ÿÀŽ ä x*4€€€€ÿüŽ Ä r--,   ÿüŽ ä x++˜º  ÿ    ä x`^€€€€ÿüŽ ä x_*ÿþþþÿàÿ ä x--˜º  ÿ    ä x*+ÿÀŽ Ä xc`ÿþþþÿàÿ ä r__4   ÿüŽ ä x,-ÿÀŽ ä rcc^   ÿüŽ ä x__˜º  ÿ    ä x--€€€€$ÞŽ ÄO x++€€€€fÞŽ ä� r--,  €ÿàÿ Æ $x++   *
Ž â x*c˜º  ÿ    ä x4#ÿ   ÿÀŽ ä x,_ÿÀŽ ä xf+€€€€f´Ž ä x__€€€€(ÞŽ ä x+*ÿÀŽ â$x4# @  4Ž â x**€€€€)ÞŽ â $x__   ,
Ž â x$-€€€€$´Ž â &rffO   ÿ   Ä $x**   +
Ž â x+4   ÿ   ä x__€€€€(´Ž â &r$$Q   f   â x(ÿ   #  ä x++€€ÿÀŽ â &r$_S   $   â x**€€€€)´Ž â $x,( @  (Ž â x#+€€€€ÿüŽ â $x)4   ÿ Ž Ä &r$*U   $   â x(#ÿþþþÿàÿ â $x#,€   ÿ Ž âx*,   ÿ   ä x))€€ÿÀŽ ä r((+   ÿüŽ ä x-*€€ÿÀŽ ä x+#€€ÿÀŽ Ä x*)€€€€ÿüŽ ä x#(˜º  ÿ    ä x,+€€€€ÿüŽ ä x**ÿþþþÿàÿ ä x(#ÿÀŽ ä x,,ÿþþþÿàÿ Ä r**)   ÿüŽ ä x##€€€€ ÞŽ äx4-€€€€ÿüŽ ä r,,+   ÿüŽ ä x**˜º  ÿ    â $x##   (
Ž â x44ÿþþþÿàÿ Ä x+,˜º  ÿ    ä x)*ÿÀŽ äx**€€€€!ÞŽ ä r44-   ÿüŽ ä x##€€€€ ´Ž ä x +ÿÀŽ â$x**   )
Ž â x++€€€€"ÞŽ Ä x44˜º  ÿ    â &r##W   $   â x**€€€€!´Ž â $x++    
Ž â x)4ÿÀŽ äx44€€€€%ÞŽ ä r'ÿF   '  â &r**Y   #   â r44)  €ÿàÿ Ä x++€€€€"´Ž â $t ÿ   ÿ Ž â x%4€€€€%´Ž â $x''   ÿ Ž ä &r*+[   *   Æ x '    ìŽ â &r%%\   *   ä 0r&ÿ&    A   ä $r %    ÿŽ Æ  r!{&     A  ä Er          Ê #r8!    8   ä G‰  �
    € ê r"EH   ÿàõ ä r 
E   ÿàó Æ $x#G   J â r!G   ÿäÿ  È �yb"B    ¨ �y%     è  �y$     ( �y4     ( �y`     " r&E   ÿàó Ê $x'   GŽ  Ê �yh&J    " r E<   ÿàó â $x(b   ÿ Ž âOx!b   ÿ   â$x*b€   ÿ Ž â x)ÿ   %  ä�x((   ÿÀŽ ä x")   *øŽ ä x((ÿ   %øŽ Æ %v""  M Ž È %v*(  M Ž â x!!   ÿÀŽ â �y%"    ™ ¨  �y**    ™ ¢ x(!ÿ   $øŽ â$x!G   @Ž  â x&ÿ   $  â/ $x'b    ÿ Ž Ä %v((  M Ž â�y       $ x&&   'øŽ ä �y'(    ™ & %v,&  M Ž Ê �y&,    ™ " $x"b   ÿ Ž Ê x#"   ÿÀŽ È x^#ÿ   4øŽ â$x"b   ÿ Ž â x4ÿ   4  Æ %v^^  M Ž â x#b   ÿ   È �y!^    ™ "! x44   #øŽ â $x)b   ÿ Ž â x#"   ÿÀŽ ä x"ÿ   `  â%v,4  M Ž âxb#ÿ   `øŽ ä xd"   )øŽ ä �y",    ™ " %vbb  M Ž â rfE   ÿàó Æ %vdd  M Ž â �y#b    ™ & $xg   GŽ  â �y$d    ™ " x(/   ÿàÿ Æ �y)fj    " r, E       â/ $z_( p  ;Ž ä $r-ÿÿ   G Ž È %x^_n   ,Ž Ê �y(^     " x+hÿ   ÿÀŽ Ê� $x,h @  +Ž â xhÿ   h  Æ $x+,€   ÿ Ž âx,,   ÿ   â $xhh @  hŽ Æ x++€€ÿÀŽ ä x-,€€ÿÀŽ ä x,+€€€€ÿüŽ ä x4-€€€€ÿüŽ â $x^h€   ÿ Ž â x,,ÿþþþÿàÿ ä x44ÿþþþÿàÿ Ä x^^€€ÿÀŽ â $x_h   ÿ Ž â r,,+   ÿüŽ ä r44-   ÿüŽ ä x-^€€€€ÿüŽ ä x+,˜º  ÿ    ä x44˜º  ÿ    Ä x`_€€ÿÀŽ ä x_-ÿþþþÿàÿ ä x,+ÿÀŽ ä x-4ÿÀŽ ä r__^   ÿüŽ ä xb`€€€€ÿüŽ È xcbÿþþþÿàÿ È rcc`   ÿüŽ È xcc˜º  ÿ    È x^cÿÀŽ ä x44€€€€%ÞŽ äO x++€€€€*ÞŽ Æ $x44   -
Ž â r++,  €ÿàÿ ä x,_˜º  ÿ    ä x*+€€€€*´Ž ä x+ ÿ   ÿÀŽ äx44€€€€%´Ž ä x%,ÿÀŽ Ä x,,€€€€'ÞŽ â $x+  @  +Ž ä &r**O   ÿ   ä $x,,   %
Ž ä $x%+€   ÿ Ž â xcc€€€€&ÞŽ â &r*4Q   *   â x,,€€€€'´Ž Ä x%%€€ÿÀŽ ä x'ÿ      â &r*,S   *   â rcc^  €ÿàÿ ä x %€€€€ÿüŽ â $x,' @  'Ž â x&c€€€€&´Ž â $x'+   ÿ Ž â x  ÿþþþÿàÿ Æ &r&&U   *   â x''€€ÿÀŽ ä r  %   ÿüŽ ä x*,   ÿ   â$x,,   ÿ Ž â x%'€€€€ÿüŽ ä x  ˜º  ÿ    ä x+*€€ÿÀŽ Ä x-,€€ÿÀŽ ä x*%ÿþþþÿàÿ ä x% ÿÀŽ äx,+€€€€ÿüŽ ä x  €€€€!ÞŽ ä x4-€€€€ÿüŽ Ä r**'   ÿüŽ â $x     %
Ž â x,,ÿþþþÿàÿ ä x44ÿþþþÿàÿ ä x%*˜º  ÿ    ä r,,+   ÿüŽ ä r44-   ÿüŽ Ä x*%ÿÀŽ äx%%€€€€"ÞŽ ä x,,˜º  ÿ    ä x  €€€€!´Ž ä x!4˜º  ÿ    â $x%%   *
Ž â x',ÿÀŽ Ä x,,€€€€#ÞŽ ä x*!ÿÀŽ äx!!€€€€$ÞŽ â &r  W   &   â r,,'  €ÿàÿ ä x"%€€€€"´Ž â $x!!   *
Ž â r)ÿF   )  Ä x,,€€€€#´Ž â &r "Y       â x!!€€€€$´Ž â $t"ÿ   ÿ Ž ä $x))   ÿ Ž ä &r ,[       Æ x))   "ìŽ â &r !\       ä 0r(ÿ(    A   ä $r  )   ÿŽ Æ  r({(     A  ä Er          Ê #r7(    7   ä G‰  �
    € ê r"EN   ÿàõ ä r E   ÿàó ä r#GT   ÿä Æ $x!   GŽ  ä �yb"B    ¨ �y%     è  �y$     ( �y4     ( �y`     " r&E   ÿàó Ê $x'   GŽ  Ê �yh&J    "	 r E=   ÿàó â $x(b   ÿ Ž âOx*b   ÿ   â$x!b@   ÿ Ž â x)ÿ   %  ä�x((   ÿÀŽ ä x")   *øŽ ä x((ÿ   %øŽ Æ %v""  M Ž È %v*(  M Ž â x!!   ÿÀŽ â �y%"    ™ ¢  x&ÿ   $  â$x'b    ÿ Ž ä �y**    ™ â x(!ÿ   $øŽ â $x!G   AŽ  â x&&   'øŽ Æ %v((  M Ž â�y       æ %v,&  M Ž â �y'(    ™ è �y&,    ™ â x"b   ÿ   È x#"   ÿÀŽ È x^#ÿ   4øŽ â$x#b   ÿ Ž âx4ÿ   4  â $x"b   ÿ Ž ä %v^^  M Ž â x44   #øŽ ä x#"   ÿÀŽ ä �y!^    ™ "$x)b   ÿ Ž â/ x"ÿ   `  Ä xb#ÿ   `øŽ â %v,4  M Ž â xd"   )øŽ Æ %vbb  M Ž â �y",    ™ " rfE   ÿàó ä %vdd  M Ž â �y#b    ™ " rgG   ÿäÿ  Æ �y$d    ™ " x(/   ÿàÿ Æ �y)fj    " $r,ÿÿ   E Ž ä/ $z_( p  ;Ž ä $r-ÿÿ   G Ž È %x^_n   ,Ž Ê �y(^     " x+hÿ   ÿÀŽ Ê $x,h @  +Ž â xhÿ   h  È x+,   ÿ   â$x,,   ÿ Ž Æ x++€€ÿÀŽ ä x-,€€ÿÀŽ â $xhh @  hŽ â x,+€€€€ÿüŽ ä x4-€€€€ÿüŽ â $x^h€   ÿ Ž â x,,ÿþþþÿàÿ ä x44ÿþþþÿàÿ Ä r,,+   ÿüŽ ä x_h   ÿ   ä x^^€€ÿÀŽ ä r44-   ÿüŽ ä x+,˜º  ÿ    ä x`_€€ÿÀŽ Ä x-^€€€€ÿüŽ ä x44˜º  ÿ    ä x,+ÿÀŽ ä xb`€€€€ÿüŽ ä x_-ÿþþþÿàÿ ä x-4ÿÀŽ Ä xcbÿþþþÿàÿ ä r__^   ÿüŽ ä rcc`   ÿüŽ È xcc˜º  ÿ    È x^cÿÀŽ ä x44€€€€%ÞŽ äO x++€€€€*ÞŽ Æ� $x44   -
Ž ä $x++   ,
Ž â x,_˜º  ÿ    È x*+€€€€*´Ž ä x+ ÿ   ÿÀŽ ä x44€€€€%´Ž ä x%,ÿÀŽ â $x+  @  +Ž â x,,€€€€'ÞŽ ä xcc€€€€&ÞŽ â &r**O   ÿ   â r,,%  €ÿàÿ â $x%+€   ÿ Ž Ä $xcc   ^
Ž â x,,€€€€'´Ž â &r*4Q   *   â x'ÿ      ä x%%€€ÿÀŽ â &r*,S   *   â x&c€€€€&´Ž â $x,' @  'Ž â x %€€€€ÿüŽ ä x'+   ÿ   â &r&&U   *   â x  ÿþþþÿàÿ â $x*,€   ÿ Ž âx''€€ÿÀŽ â $x,,   ÿ Ž â r  %   ÿüŽ Ä x+*€€ÿÀŽ ä x%'€€€€ÿüŽ ä x-,€€ÿÀŽ ä x  ˜º  ÿ    ä x,+€€€€ÿüŽ ä x*%ÿþþþÿàÿ Ä x% ÿÀŽ äx4-€€€€ÿüŽ ä x  €€€€!ÞŽ äx,,ÿþþþÿàÿ ä r**'   ÿüŽ â $x     %
Ž â x44ÿþþþÿàÿ Ä r,,+   ÿüŽ ä x%*˜º  ÿ    ä r44-   ÿüŽ ä x,,˜º  ÿ    ä x*%ÿÀŽ ä x  €€€€!´Ž Ä x%%€€€€"ÞŽ ä x!4˜º  ÿ    ä x',ÿÀŽ äx,,€€€€#ÞŽ ä r%%*  €ÿàÿ ä x*!ÿÀŽ â$x,,   '
Ž â x!!€€€€$ÞŽ â &r  W   &   â x"%€€€€"´Ž Ä x,,€€€€#´Ž â $x!!   *
Ž â r)ÿF   )  â &r "Y       ä $t"ÿ   ÿ Ž â x!!€€€€$´Ž â &r ,[       â x))   ÿ   Æ &r !\       â x))   "ìŽ â 0r(ÿ(    A   È $r  )   ÿŽ â  r!{(     A  È Er          Ê #r6!    6   â #r3   3   â #r2   2   â G‰  °
    € ê x/   ÿàÿ Ê $z- p  ;Ž È %x-n   Ž È %x-n   Ž â rE   ÿàó È rE   ÿàñ â $x   GŽ  È $x   G  â �y,B    ¨ �y     è  �y"    ( �y*    ( �y+    " r$E   ÿàñ Ê $x%   G  Ê �y4$J    " rE>   ÿàñ ä x!,   ÿ   âO$x&,€   ÿ Ž ä$x,@   ÿ Ž â x#ÿ      ä� x!!   ÿÀŽ ä x&#   &øŽ Ä x!ÿ    øŽ â $x%,    ÿ Ž ä/ %v &  M Ž â x   ÿÀŽ Æ %v  M Ž âx$ÿ   "  â�y#     ™ ¢  x&ÿ   "øŽ ä x$$   %øŽ â �y`    ™ â rGB   ÿä  â %v&&  M Ž È %v($  M Ž â �y     è �y$&    ™ è  �y%(    ™ â  $x,   ÿ Ž Ê/ x   ÿÀŽ â $x,   ÿ Ž Æ x ÿ   *øŽ â$x,   ÿ Ž â x*ÿ   *  Æ %v    M Ž âx**   øŽ ä x   ÿÀŽ ä �y     ™ "x,,   ÿ   â %v&*  M Ž â xÿ   +  äxÿ   +øŽ Ä x*   ,øŽ Æ %v(  M Ž ä �y&    ™ " r^E   ÿàñ â %v**  M Ž ä �y(    ™ $ $x_   G  ä �y *    ™ ( �y!^j    " $r&ÿÿ   E Ž Ä/ $r'ÿÿ   G Ž È %x,-n   &Ž Ê �y",     " xM4ÿ   ÿÀŽ È rM4M   ÿpŽ ä x4ÿ   4  Æ $x'M   ÿ Ž ä$x&M€   ÿ Ž ä $x+4 @  4Ž â x('€€ÿÀŽ ä x&&€€ÿÀŽ ä x)(€€€€ÿüŽ ä x'&€€€€ÿüŽ Ä x*+   ÿ   â$x++   ÿ Ž â x))ÿþþþÿàÿ ä x''ÿþþþÿàÿ ä x**€€ÿÀŽ ä x,+€€ÿÀŽ ä/ r))(   ÿüŽ Ä r''&   ÿüŽ ä x(*€€€€ÿüŽ ä x-,€€€€ÿüŽ ä x&)˜º  ÿ    ä x''˜º  ÿ    ä x+(ÿþþþÿàÿ Ä x)&ÿÀŽ ä x--ÿþþþÿàÿ ä x('ÿÀŽ ä r++*   ÿüŽ ä r--,   ÿüŽ ä x++˜º  ÿ    È x*+ÿÀŽ ä x&&€€€€#ÞŽ äO x''€€€€`ÞŽ Æ� $x&&   )
Ž ä $x''   (
Ž â x(-˜º  ÿ    ä x,ÿ   ÿÀŽ ä x&&€€€€#´Ž ä x#(ÿÀŽ ä x++€€€€$ÞŽ Ä x((€€€€%ÞŽ ä r,,   ÿpŽ â $x++   *
Ž â x`'€€€€`´Ž â $x((   #
Ž ä $x#,€   ÿ Ž â x++€€€€$´Ž â &r``O   ÿ   â x$ÿ     Ä x##€€ÿÀŽ â &r&&Q   `   â x((€€€€%´Ž â $x%,   ÿ Ž â x#€€€€ÿüŽ â &r&+S   &   â r'$$   ÿpŽ ä x$ÿþþþÿàÿ â &r&(U   &   â x%%€€ÿÀŽ â $x'€   ÿ Ž âr$$#   ÿüŽ â $x('   ÿ Ž â x#%€€€€ÿüŽ Ä x'€€ÿÀŽ ä x$˜º  ÿ    ä x+(€€ÿÀŽ ä x)'€€€€ÿüŽ ä x$#ÿþþþÿàÿ ä x(ÿÀŽ Ä x€€€€ÞŽ äx,+€€€€ÿüŽ ä x*)ÿþþþÿàÿ ä r$$%   ÿüŽ ä r#(  €ÿàÿ ä x,,ÿþþþÿàÿ Ä r**'   ÿüŽ ä x$˜º  ÿ    â 0r"ÿ"    A   â x##€€€€´Ž ä r,,+   ÿüŽ ä x**˜º  ÿ    â &r##W   &   â xÿÀŽ â  r"{"     A  â x€€€€ÞŽ Ä x%,˜º  ÿ    ä x'*ÿÀŽ â$x   
Ž â x**€€€€ÞŽ ä x%ÿÀŽ äx%%€€€€ ÞŽ â $x**   '
Ž â x€€€€´Ž Ä r!ÿF   !  â $x%%   
Ž â x**€€€€´Ž â &rY   #   â x        â $x!!   ÿ Ž â x%%€€€€ ´Ž â &r*[      Æ x!!   ìŽ â &r%\      È $r!   ÿŽ Ê Er        Ê #r5"   5   â x::   ÿàÿ ä xEEà  ÿàó ä | :   pbð ä xIIà  ÿàõ â $rGÿÿ   GŽ  â xDD€   ÿàÿ Æ $rKÿÿ   K Ì D	       À â Gy  @«ÿÿÿÿƒ ê Ay        € ê y      "   " Ey  0    € â r ÿ   pRð är ÿ   pRò Ö G‰  ð     € ê z ÿ `  pPð ä r    ÿ8Ž ä z ÿ a   Sð Æ ˆs a þÿ H   ô ˆƒ ]€þÿ H   è ˆƒ 3 ÿÿ H   è ˆƒ 2€ÿÿ H   â G‰  p     € ê ˆs 8   H   è ˆs ]€þÿ H   è ˆs 7€  H   è ˆs 3 ÿÿ H   è ˆs 6   H   è ˆs 2€ÿÿ H   è ˆs 5€  H   ä Ay        € ê {          ì M        € ê y      &   "> z ÿ `  pPð â $tÿ   ÿ Ž Ä y      '   b z ÿ a   Sð â $z x  /Ž È $z ~  Ž È/ %v j   Ž È G	  P    € ê „y      H   " z / r  p`ð Æ „y  €   H   b r  ÿ   pVp  Æ „y     H   ¨ „y	  €  H   è „y     H   ( „y  €  H   ( „y     H   ( „y     H   ( „y     H   ( „y  €  H   ( „y  €  H   " !ra       Æ „y  €  H   " !r]       â/ !r       ÆO !r	       â� !r       Æ!r       ä ‰      b !r3       Æ ‰      ¢ !r       È !r       â !r2       È ‰
      â !r       È !r       â !r       â/ !r       ÆO ‰	      ( ‰      h ‰      ¢ !r

       Ê� ‰
      â !r	       â !r       È/ ‰      " !r	       ÆO ‰ €    h ‰	 €    ¢ !r
       Ê� ‰ €    â !r       â !r       È/ ‰ €    " !r	       ÆO ‰ @    h ‰ @    ¢ !r       Ê� ‰ @    â !r       â !r       È/ ‰ @    " x/   ÿàÿ â!r       âO x/   ÿàÿ â ‰      b z  r  p`ò ä z  r  p`ô â !r       â� ‰	      ¢ x//   ÿàÿ Ä x     pVò  â‰
      â x     pVt â !r       â !r       È/ ‰      " !‚0       â!r	       ÈO †‰     â z / r  p`ð â !r	
       â� !’0       äx     pVp  â !¢		0       ä †™    è †© 	   â !r       Ê M	        € ê !r0       Ê/ †y    â My        € ê $vÿ h  ÿ Ž Ê x    pBð Ú G	  @    € ê x    pRð Ú G‰  À    € ê x    pRð Ú G	   $    € ê „y      H   " z / r  p`ð â Ey  p    € ä „y     H   b r  ÿ   pVp  Æ „y	     H   ¨ „y     H   è „y     H   ( „y   
  H   ( „y  €   H   ( „y  €  H   ( „y  €  H   è „y  €  H   " !ra       Æ „y  €
  H   â !r8       â/ !r	       ÆO !r       â� !r       Æ!r       ä ‰      " !r]       Æ ‰      b !r       È !r       â !r       ä „y  €  H   â !r       Æ/ ‰      ( ‰      b !r       â !r	       È/ ‰ €    ( ‰
	 €    b !r7       È !r       È !r       â !r       â !r
	
       È/ ‰ @    ( ‰
 @    b !r       â !r
       È/ ‰      ( ‰      b !r	       â !r       â/ G	  �     € ê !r.       â!r	0       Æ  x;ª¸¿  A  Ì s        $ !t  €?    Ì s        $  r     A  È  r     A  Ê †y     ä Ay        € ê ‰      b Ey  Ð    € æ ‰      ( „y     H   ¨ „y     H   è „y     H   â !r       â/ !r       È ‰      ( ‰      h „y     H   â !r6       âO !r       â !r	       È/ ‰ €    ( ‰
	 €    h „y   	  H   ¢ !r3       È� !r       â !r       ä „y     H   â !r       â !r
	
       â/ x	/   ÿàÿ â ‰ @    & z 	 r  p`ð â ‰
 @    f x     pVp  â !r       âO !r       â !r
       È/ ‰      ( ‰      b !r       â !r	       â !r       â/ G	  �     € ê !r.       â!r	0       Æ  x;ª¸¿  A  Ì s        $ !t  €?    Ì s
        $  r
     A  È  r     A  Ê †y    ä Ay        € ê ‰      b Ey       € æ ‰      ( „y  €  H   ¨ „y  €	  H   è „y  €  H   è „y  €  H   â !r       Æ/ „y  €  H   â !r	       Æ ‰      ( ‰	      b !r5       ÈO !r       â� !r       â !r
	       È/ ‰ €    ( ‰
 €    h „y	  €  H   ¢ !r       â !r
       È/ ‰ @    ( ‰ @    b !r		       âO !r       â x/   ÿàÿ â !r       Æ/ ‰      " z  r  p`ð Æ ‰      b x     pVp  â !r       â !r       Ö/ G	  �     € ê !r.       â!r0       Æ  x;ª¸¿  A  Ì s        $ !t  €?    Ì s        $  r     A  È  r     A  Ê †y    ä Ay        € ê ‰	      b !r2       â x//   ÿàÿ Æ !r       â z / r  p`ð Æ !r       â x     pVp  È ‰      ¢ !r	       Ê/ ‰      b !r       ÊO ‰      " !r
       Ê/ ‰
 €    b !r       Ê ‰ €    " !r	
       Ê/ ‰	 @    b !r       Ê ‰ @    " !r	       Ê/ ‰      b !r       Ê ‰       ¢  !r       â/ M	        € ì !r..       â!r        ÆO  x.;ª¸¿  A  â !r00       Ê s        $ !t  €?    Ì s        $  r.     A  È  r0     A  Ê †y    â My        € ê „y      H   " z / r  p`ð â Ey       € ä „y     H   b r  ÿ   pVp  Æ „y     H   ¨ „y     H   è „y	     H   ( „y   
  H   ( „y  €  H   ( „y  €  H   ( „y  €  H   è „y  €  H   â !ra       â !r8       Æ/ !r       âO !r       Æ� !r	       â!r       È ‰      " !r7       Æ ‰      b !r       â !r       â !r       È/ ‰      ( ‰      h „y  €   H   ¢ !r       â !r	       È/ ‰ €    ( ‰
	 €    b !r]       ÈO !r       â !r       ä „y  €
  H   â !r       â !r
	
       ä/ ‰ @    ( ‰
 @    b !r       â !r
       È/ ‰      ( ‰      b !r       â !r	       â !r       â/ G	  À     € ê !r.       È x'7=  A  â x*BL?  A  Æ #t  €?   È  r     A  â  x   ?  A  â !r	0       È s     $   $ !t  €?    È  r     A  È  r     A  Ê †y     ä Ay        € ê ‰      b Ey       € æ ‰      ( „y     H   ¨ „y     H   è „y     H   è „y     H   â !r       Æ/ „y     H   â !r       Æ ‰      ( ‰      b !r3       âO !r6       â� !r       ä !r       â !r	       Ä/ ‰ €    ( ‰
	 €    b !r       Æ „y   	  H   ¢ !r       â x/   ÿàÿ â !r
	
       Æ/ ‰ @    " z  r  p`ð Æ ‰
 @    b x     pVp  â !r       ÈO !r       â !r       â !r
       È/ ‰      ( ‰      b !r	       â !r       â/ G	  À     € ê !r.       È x'7=  A  â x*BL?  A  Æ #t  €?   È  r     A  â  x   ?  A  â !r	0       È s     $   $ !t  €?    È  r     A  È  r     A  Ê †y    ä Ay        € ê ‰      b Ey  Ð    € æ ‰      ( „y  €	  H   è „y  €  H   è „y  €  H   è „y  €  H   â !r       Æ/ „y  €  H   â !r       Æ ‰      ( ‰      h „y  €  H   ¢ !r       â !r	       È/ ‰ €    ( ‰
	 €    b !r5       ÈO !r       â !r       â x/   ÿàÿ ä !r       â !r
	
       â/ ‰ @    " z  r  p`ð Æ ‰
 @    b x     pVp  â !r       â !r
       È/ ‰      ( ‰      b !r	       â !r       â/ G	  À     € ê !r.       È x'7=  A  â x*BL?  A  Æ #t  €?   È  r     A  â  x   ?  A  â !r	0       È s     $   $ !t  €?    È  r     A  È  r     A  Ê †y    ä Ay        € ê ‰      " !r2       â x//   ÿàÿ Æ !r       â z / r  p`ð Æ !r       â x     pVp  È ‰      b !r       Ê ‰      " !r       Ê/ ‰      b !r	       Ê ‰
	 €    " !r       Ê/ ‰ €    b !r
	
       Ê ‰
 @    " !r       Ê/ ‰ @    b !r
       Ê ‰      " !r       Ê/ ‰       ¢ !r       â M	        € ì !r..       â/!r        ÆO  x.'7=  A  â x.*BL?  A  â!r00       ä #t.  €?   â x..   ?  A  Æ  r     A  Ì s     $   $ !t  €?    È  r.     A  È  r0     A  Ê †y    â My        € ê x    pRð Ú G‰  P    € ê x    pRð Ú G	  `    € ê „y      H   " z / r  p`ð â Ey       € ä „y     H   b r  ÿ   pVp  Æ „y     H   ¨ „y     H   è „y	     H   ( „y   
  H   ( „y  €  H   ( „y  €  H   ( „y  €  H   è „y  €  H   â !ra       â !r8       Æ/ !r       âO !r       Æ� !r	       â!r       È ‰      " !r7       Æ ‰      b !r       â !r       â !r       È/ ‰      ( ‰      h „y  €   H   ¢ !r       â !r	       È/ ‰ €    ( ‰
	 €    b !r]       ÈO !r       â !r       ä „y  €
  H   â !r       â !r
	
       ä/ ‰ @    ( ‰
 @    b !r       â !r
       È/ ‰      ( ‰      b !r       â !r	       â !r       â/ G	  À     € ê !r.       Ê	z i    � â !r	0       È  x;ª¸¿  A  â 	z i    � È 	z i €  � â s        $ !t  €?    Ì s        $  r     A  È  r     A  Ê †y     ä Ay        € ê ‰      b Ey       € æ ‰      ( „y     H   ¨ „y     H   è „y     H   è „y     H   â !r       Æ/ „y     H   â !r       Æ ‰      ( ‰      b !r3       âO !r6       â� !r       ä !r       â !r	       Ä/ ‰ €    ( ‰
	 €    b !r       Æ „y   	  H   ¢ !r       â x/   ÿàÿ â !r
	
       Æ/ ‰ @    " z  r  p`ð Æ ‰
 @    b x     pVp  â !r       ÈO !r       â !r       â !r
       È/ ‰      ( ‰      b !r	       â !r       â/ G	  À     € ê !r.       Ê	z i    � â !r	0       È  x;ª¸¿  A  â 	z i    � È 	z i €  � â s        $ !t  €?    Ì s        $  r     A  È  r     A  Ê †y    ä Ay        € ê ‰      b Ey  Ð    € æ ‰      ( „y  €	  H   è „y  €  H   è „y  €  H   è „y  €  H   â !r       Æ/ „y  €  H   â !r       Æ ‰      ( ‰      h „y  €  H   ¢ !r       â !r	       È/ ‰ €    ( ‰
	 €    b !r5       ÈO !r       â !r       â x/   ÿàÿ ä !r       â !r
	
       â/ ‰ @    " z  r  p`ð Æ ‰
 @    b x     pVp  â !r       â !r
       È/ ‰      ( ‰      b !r	       â !r       â/ G	  À     € ê !r.       Ê	z i    � â !r	0       È  x;ª¸¿  A  â 	z i    � È 	z i €  � â s        $ !t  €?    Ì s        $  r     A  È  r     A  Ê †y    ä Ay        € ê ‰      " !r2       â x//   ÿàÿ Æ !r       â z / r  p`ð Æ !r       â x     pVp  È ‰      b !r       Ê ‰      " !r       Ê/ ‰      b !r	       Ê ‰
	 €    " !r       Ê/ ‰ €    b !r
	
       Ê ‰
 @    " !r       Ê/ ‰ @    b !r
       Ê ‰      " !r       Ê/ ‰       ¢ !r       â M	        € ì !r..       â/!r        ÈO 	z.. i    � â !r00       È  x.;ª¸¿  A  â 	z00 i    � È 	z00 i €  � â s        $ !t  €?    Ì s        $  r.     A  È  r0     A  Ê †y    â My        € ê „y     H   " z / r  p`ð Æ „y
  €   H   b r  ÿ   pVp  Æ „y      H   ¨ „y  €  H   è „y     H   ( „y     H   ( „y     H   ( „y  €  H   ( „y     H   ( „y  €  H   ( „y     H   " !r8       Æ „y   
  H   " !r

]       Æ/ „y   	  H   b !ra       ÆO „y  €  H   ¢ !r7       Æ� „y	     H   â !r3       Æ„y  €
  H   " !r       Æ „y     H   " !r6       Æ „y     H   " !r

       Æ „y  €  H   " !r       Æ „y  €  H   " !r       Æ „y  €  H   " !r       Æ „y  €	  H   â !r       Æ „y!  €  H   â !r       â/ x/   ÿàÿ ä „y#  €  H   â !r

       âO z  r  p`ô â !r	       â� ‰      $ x     pVt â !r       â‰
      b !r       Æ ‰      ¢ !r       Æ ‰	      â !r2       Æ ‰      " !r5       Æ ‰      " !r       â !r       â !r       â !r
       Æ/ ‰      " !r!!       â !r       âO !r#       Æ ‰!      b !r
	       Æ� ‰      ¢ !r       â!r       ä ‰      è ‰	      ( ‰
      " !r       Æ ‰      " !r!       Æ/ ‰      b !r       ÆO ‰ €    ¢ x/   ÿàÿ äx//   ÿàÿ â ‰      ¢ !r       â� z  r  p`ò ä ‰      â !r		       âx     pVò  â !r
       â ‰ €    " z / r  p`ö â !r       Ä ‰	 €    " x     pVö â !r       ä/ ‰
 €    b !r       ÆO ‰ €    ¢ !r       Æ ‰ €    ¢ !r       Æ� ‰ €    è ‰ €    â !r       â!r	       ä ‰ @    " !r

       Æ/ ‰ @    b !r       ÆO ‰	 @    ¢ !r       Æ ‰
 @    " !r       Æ� ‰ @    â !r       Æ ‰ @    è ‰ @    â !r       Æ ‰ @    " !r       â/ !r		       äO ‰      b !r
       Æ‰      ¢ !r       Æ� ‰	      â !r       Æ ‰
      " !r       Æ ‰      " !r       Æ ‰      ( ‰      " !r       Æ/ ‰      b !r       âO !r		       Æ� !‚0       â!‚.       â !r
       â!’	0       ä  ‚     A  â !r       â !’.       Æ †‰     â !r       â !¢0       â  ’     A  ä !¢	.       â !r       ä †™    â  ¢		     A  â !r       È/ †© 	   â M9        € ê !r00       â !r.       ÈO  r0     A  Ê †y    â My        € ê „y      H   " z / r  p`ð â Ey  À    € ä „y     H   b r  ÿ   pVp  Æ „y     H   ¨ „y     H   è „y	     H   ( „y   
  H   ( „y  €  H   ( „y  €  H   ( „y  €  H   è „y  €  H   â !ra       â !r8       Æ/ !r       âO !r       Æ� !r	       â!r       È ‰      " !r7       Æ ‰      b !r       â !r       â !r       È/ ‰      ( ‰      h „y  €   H   ¢ !r       â !r	       È/ ‰ €    ( ‰
	 €    b !r]       ÈO !r       â !r       ä „y  €
  H   â !r       â !r
	
       ä/ ‰ @    ( ‰
 @    b !r       â !r
       È/ ‰      ( ‰      b !r       â !r	       â !r       â/ G	  à     € ê !r.       Ê	x  à@  � Ê  x#ÛÙ?  A  È  x;ª¸¿  A  â !r	0       Ê s        " 	x  à@  � È 	x  àÀ  � Ê !t  €?    â !t  €?    È s        $  r     A  È  r     A  Ê †y     ä Ay        € ê ‰      b Ey       € æ ‰      ( „y     H   ¨ „y     H   è „y     H   è „y     H   â !r       Æ/ „y     H   â !r       Æ ‰      ( ‰      b !r3       âO !r6       â� !r       ä !r       â !r	       Ä/ ‰ €    ( ‰
	 €    b !r       Æ „y   	  H   ¢ !r       â x/   ÿàÿ â !r
	
       Æ/ ‰ @    " z  r  p`ð Æ ‰
 @    b x     pVp  â !r       ÈO !r       â !r       â !r
       È/ ‰      ( ‰      b !r	       â !r       â/ G	  à     € ê !r.       Ê	x  à@  � Ê  x#ÛÙ?  A  È  x;ª¸¿  A  â !r	0       Ê s        " 	x  à@  � È 	x  àÀ  � Ê !t  €?    â !t  €?    È s        $  r     A  È  r     A  Ê †y    ä Ay        € ê ‰      b Ey  ð    € æ ‰      ( „y  €	  H   è „y  €  H   è „y  €  H   è „y  €  H   â !r       Æ/ „y  €  H   â !r       Æ ‰      ( ‰      h „y  €  H   ¢ !r       â !r	       È/ ‰ €    ( ‰
	 €    b !r5       ÈO !r       â !r       â x/   ÿàÿ ä !r       â !r
	
       â/ ‰ @    " z  r  p`ð Æ ‰
 @    b x     pVp  â !r       â !r
       È/ ‰      ( ‰      b !r	       â !r       â/ G	  à     € ê !r.       Ê	x  à@  � Ê  x#ÛÙ?  A  È  x;ª¸¿  A  â !r	0       Ê s        " 	x  à@  � È 	x  àÀ  � Ê !t  €?    â !t  €?    È s        $  r     A  È  r     A  Ê †y    ä Ay        € ê ‰      " !r2       â x//   ÿàÿ Æ !r       â z / r  p`ð Æ !r       â x     pVp  È ‰      b !r       Ê ‰      " !r       Ê/ ‰      b !r	       Ê ‰
	 €    " !r       Ê/ ‰ €    b !r
	
       Ê ‰
 @    " !r       Ê/ ‰ @    b !r
       Ê ‰      " !r       Ê/ ‰       ¢ !r       â M	        € ì !r..       â/!r        ÈO 	x..  à@  � â !r00       È  x.#ÛÙ?  A  â 	x00  à@  � Æ  x;ª¸¿  A  â 	x00  àÀ  � Ê s        " !t0  €?    â !t  €?    Ì s        $  r..     A  È  r.     A  Ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y       "   $ z   w  p`ð Ú M	        € ê y      !   " ¹z  l      â Ey       € â ™x?    â $rÿÿ   ÿ Ž â ¹z  F   
   â $rÿÿ   ÿ Ž â xÿ     È |    p`ð Ú G	       € ê y	      &   " $tÿ   ÿ Ž ä $z  x  	Ž È %v \   Ž Ì �y    ™ ¢  $rÿÿ   ÿ Ž â x   ÿÀŽ â$x
   ÿ Ž â y      %   b 'z	 m   Ž â x&ÿ     ä x
   ÿÀŽ â $z  r  ÿŽ â zÿ n    â $x(   ÿ Ž â x   ÿüŽ â $r'ÿÿ   ÿ Ž â x

   ÿÀŽ â $rÿÿ   
Ž â x((8   ÿÀŽ â $r)ÿÿ   ÿ Ž Ä $z	 o  	Ž ä $x   ÿ Ž ä $z	 u  ÿŽ ä $rÿÿ   ÿ Ž â x   ÿÀŽ â $x   ÿ Ž È/ $x   
Ž â r   ÿàó â $z q  Ž âx	   ÿàÿ â $x   ÿ Ž ä $vÿ Y  ÿ Ž ä $x   ÿ Ž â xþÿÿÿÿàñ È xÿÿÿÿÿ  â $rÿÿ   ÿŽ  ä $z t  Ž äO$z t  Ž ä $z		 q  Ž ä %x n   &Ž È %x"n   (Ž È %x,	n   Ž È $vÿ X  ÿ Ž ä $rÿÿ   ÿ Ž ä %x$n   Ž È %x&	n   &Ž È %x(	n   (Ž È %x*	n   Ž È %xn   Ž È r$   ÿàó är"   ÿàñ Æ $x%   Ž  ä$x#     Æ �yB   ‘ ¨  �y/   • è �y.   • ( �y   • h �y   • b r*   ÿàñ Ê $x+     â r(   ÿàñ È �y;B   ‘ b $x)     Ê �y8   • h �y7   • h �y6   • h �y:   • b r   ÿàñ Ê $x     Ê �yJ   • h �yL   • b r4   à‘ â $t9ÿ$   ÿ Ž ä $x   ÿ Ž äO$x€   ÿ Ž â/ xÿ   /  Ä� x   ÿÀŽ â $x   ÿ Ž â x2   øŽ â $x@   ÿ Ž âxÿ   /øŽ ä xÿ   .  â$x    ÿ Ž âxÿ     â$x   ÿ Ž â x   ÿÀŽ â %v22   Ž â x>   øŽ Æ %v   Ž âx<   øŽ ä xÿ   øŽ ä x   ÿÀŽ â �y2    ™ ¨  �y    ™ â %v   Ž â x/ÿ     Æ %v<<   Ž â�y5    ™ b	 x0ÿ   .øŽ ä %v>>   Ž â �y<    ™ b r.,   ÿàõ ä r3ÿ   /$~  â $x2   ÿ Ž â�y>    ™ b $x   ÿ Ž âxÿ     â %v4 Z  9 Ž È/ $x/-    ä $x33$   ÿŽ â x4   2øŽ â %v00   Ž âx   ÿÀŽ â �y.J   • $ $x   3Ž â xÿ   øŽ â �y0    ™ ( �y   ™ " %v2   Ž Æ �y   ™ " %v04   Ž Æ �y>2    ™ ( �y0    ™ ( �y4   ™ ( �y=   ™ ( �y?.L   • ( �y<   ™ " $x9;   ÿ Ž Æ �yF   ™ $ x99   ÿÀŽ ä �yG   ™ $ x@9ÿ   8øŽ â$x9;€   ÿ Ž â x8ÿ   8  â �yH    • $ %v2@   Ž â x88   9øŽ Æ $x0;@   ÿ Ž â �y@2    ™ " %v.8   Ž Æ/ x00   ÿÀŽ ä �yA.    ™ " $x8;    ÿ Ž âx00ÿ   7øŽ äx7ÿ   7  â $x9;   ÿ Ž Æ x87   8øŽ ä x99   ÿÀŽ Æ %v28   Ž âx89ÿ   6øŽ äx9ÿ   6  â $x.;   ÿ Ž â/ �yC2    ™ " %v68   Ž Æ x.9   .øŽ â %v00   Ž â�yD6    ™ & $x/;   ÿ Ž â �yB0    ™ " %v8.   Ž Æ x//   ÿÀŽ ä �yE8    ™ $ x./ÿ   :øŽ â$x0;   ÿ Ž â x/ÿ   :  Æ %v:.   Ž âx0/   0øŽ ä r2&   ÿàó är.    ÿàñ â %v00   Ž â �y::    ™ & $x3'   Ž  Ä $x/!     â �y00    ™ ( �yI.j   ‘ ( �y32j   ‘ ( �y;    ™ " $r.ÿÿ    Ž Ä/ $r/ÿÿ    Ž È %x6n   .Ž È %x8	n   .Ž ä �y66    • ( �y88    • " x1ÿ   ÿÀŽ Ê $x1 @  1Ž â xÿ     Æ $x.1€   ÿ Ž ä$x11   ÿ Ž ä $xJ @  Ž â x..€€ÿÀŽ ä x11€€ÿÀŽ â $x2J€   ÿ Ž â x.€€€€ÿüŽ ä x1€€€€ÿüŽ Ä x/ÿþþþÿàÿ ä x2€€ÿÀŽ ä r//.   ÿüŽ â $x.J   ÿ Ž â xÿþþþÿàÿ ä x€€€€ÿüŽ ä x2ÿ   ÿÀŽ Ä r11   ÿüŽ ä xÿþþþÿàÿ â $x2 @  2Ž â x.€€ÿÀŽ ä x.ÿ     â $x72   ÿ Ž â/r   ÿüŽ â $x2€   ÿ Ž â x€€€€ÿüŽ â $x.. @  .Ž â x77€€ÿÀŽ Ä xÿþþþÿàÿ â $x2.€   ÿ Ž âx€€ÿÀŽ ä r   ÿüŽ â $x..   ÿ Ž â x€€€€ÿüŽ ä x97€€€€ÿüŽ ä x22€€ÿÀŽ Ä xÿþþþÿàÿ ä xJ9ÿþþþÿàÿ ä x92€€€€ÿüŽ ä x..€€ÿÀŽ ä r   ÿüŽ ä rJ7   ÿüŽ Ä xK9ÿþþþÿàÿ ä xJ.€€€€ÿüŽ ä x7/˜º  ÿ    ä x91˜º  ÿ    ä r/K2   ÿüŽ ä xKJÿþþþÿàÿ Ä x27ÿÀŽ ä x19ÿÀŽ ä x˜º  ÿ    ä x˜º  ÿ    ä r.K.   ÿüŽ ä xJ9€€€€ÞŽ ÄO x77€€€€ÞŽ Æ� $x1J   1
Ž ä $x77   2
Ž â x2˜º  ÿ    ä x9ÿÀŽ äx1€€€€´Ž ä x€€€€5ÞŽ äx12ÿÀŽ Ä x22€€€€ÞŽ â $x   9
Ž â xÿÀŽ äx€€€€ÞŽ â $x22   1
Ž â x5€€€€5´Ž ä x7€€€€´Ž ä x2€€€€´Ž â $x   
Ž â xÿ   ÿÀŽ Äx..˜º  ÿ    ä x€€€€´Ž â $x @  Ž â x˜º  ÿ    â &r   ÿ   â x.ÿÀŽ ä x//˜º  ÿ    â &r      â x7ÿÀŽ â $x€   ÿ Ž â x€€€€ÞŽ Ä x..€€€€ÞŽ ä x/ÿÀŽ äx//€€€€>ÞŽ â $x..   
Ž â x€€ÿÀŽ â $x   7
Ž ä &r554      â x.€€€€´Ž â $x//   
Ž â x€€€€ÿüŽ â $x   ÿ Ž â xÿ     â &r=   5   â x€€€€´Ž Ä xÿþþþÿàÿ â $x @  Ž â x?ÿ   ÿÀŽ ä x€€ÿÀŽ â &r<      â x>/€€€€>´Ž â $x€   ÿ Ž â r   ÿüŽ â $x/? @  Ž â x€€€€ÿüŽ â $x   ÿ Ž Ä $x/€   ÿ Ž â xÿþþþÿàÿ ä x€€ÿÀŽ ä x€€ÿÀŽ ä r   ÿüŽ â $x//   ÿ Ž â x€€€€ÿüŽ ä x€€ÿÀŽ Ä x€€€€ÿüŽ ä xÿþþþÿàÿ ä x.€€€€ÿüŽ ä xÿþþþÿàÿ ä x/€€ÿÀŽ ä r   ÿüŽ Ä x..ÿþþþÿàÿ ä r   ÿüŽ ä x€€€€ÿüŽ ä x˜º  ÿ    ä r.   ÿüŽ ä x?ÿ   ?  Ä xÿþþþÿàÿ ä x.ÿÀŽ äx˜º  ÿ    ä x€€€€@ÞŽ â $x? @  ?Ž â r   ÿüŽ ä xÿÀŽ â$x   .
Ž â x€€€€AÞŽ â $x.€   ÿ Ž Æ x@€€€€@´Ž â $x   
Ž â x.€€ÿÀŽ Æ &r@@   ÿ   â xA€€€€A´Ž ä x€€€€ÿüŽ ä x˜º  ÿ    â &rAA   @   â xÿþþþÿàÿ ä xÿÀŽ äx€€€€CÞŽ Ä x˜º  ÿ    ä x˜º  ÿ    â $x   
Ž ä $x   ÿ Ž â xÿÀŽ äx˜º  ÿ    ä x€€€€DÞŽ Ä x.ÿÀŽ äx€€€€BÞŽ ä x€€ÿÀŽ ä xÿÀŽ â$x   
Ž â x€€€€EÞŽ â $x   .
Ž â x€€€€ÿüŽ Ä r   ÿüŽ â $x   
Ž â xÿþþþÿàÿ ä x€€€€B´Ž ä x˜º  ÿ    ä r   ÿüŽ â &r4   A   â x€€€€C´Ž Ä xÿÀŽ ä x˜º  ÿ    ä x€€€€:ÞŽ â &r=      â x€€€€D´Ž ä xÿÀŽ â $x   
Ž â x€€€€0ÞŽ â &r<      â xE€€€€E´Ž â &rF      â rIÿ
   I  â$x   
Ž â r3ÿ
   3  â &rFEF      â x€€€€:´Ž â &r>G      â x0€€€€0´Ž â $tÿ   ÿ Ž â x   ÿàÿ â $xII   ÿ Ž Ä $tÿ   ÿ Ž ä $x33   ÿ Ž ä &rG   F   â xI   ìŽ â &r;      âx3   ìŽ â &r0;      â |    p`ð â 0r6ÿ6    A   Ä 0rÿH    A   ä $r   ÿŽ ä $r   ÿŽ ä 0r8ÿ8    A   â  r6     A  â x¸  ÿàó ä Er        â  r8     A  â Er        â $rÿÿ   Ž  â x    ÿàÿ â #r      Ä #r      â D	       À â Gy   êÿÿÿÿƒ ê Ay        € ê ‰      ( ‰      h y      %   ¢ !r       â !r       È/ ‰      " $x   Ž ÆO ‰      d z  p  p`ð È x    pDp  â !r       â !r       È/ ‰ €    ( ‰	 €    b !r       â !r		       È/ ‰ @    ( ‰
	 @    b !r       â !r
	
       È/ ‰      h  ‰
      ¢  M	        € ê y	      &   â $x   ÿ Ž â !r       â/ $z   s  Ž â !r
       âO $tÿ   ÿ Ž â x    p"ò Ä r ÿ   p"ð Ö $ÿÿ    Ž ä $z 	 v   Ž È� %v  j   Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      "   $ z  w  p`ð Ú M	        € ê y"      !   ( y	      &   b ¹z  l      â $t ÿ   ÿ Ž â ™x?    â Ey  °.    € â ¹z  F   
   â x
ÿ   "  È | 
   p`ð â $z x  	Ž È/ %v \    Ž â x      ÿ  Æ $rÿÿ   ÿ Ž â �y    ™ b $rÿÿ   ÿ Ž ä $rÿÿ   ÿ Ž ä $rÿÿ   ÿ Ž â G	  à-    € ê $rÿÿ   ÿ Ž â y      %   " $x"   ÿ Ž âx"   ÿÀŽ â 'z	 m   Ž â Ey `-    € ä x   ÿÀŽ â $z r  ÿŽ â zÿ n    â $rÿÿ   ÿ Ž â x   ÿüŽ â $r#ÿÿ   ÿ Ž â xÿ     â $rÿÿ   
Ž Ä $rÿÿ   ÿ Ž ä $z o  	Ž ä $x"   ÿ Ž ä$r	ÿÿ   
 Ž ä $x""   ÿ Ž â x   ÿÀŽ â $x   ÿ Ž Ä $z t  	Ž âx""8   ÿÀŽ â $x   
Ž â x
   ÿàÿ â $z u  ÿŽ ä $x   ÿ Ž ä $z q  	Ž ä $z
 q  Ž â x*þÿÿÿÿàñ â $z t  Ž â r
   ÿàó â %xn   Ž â x+ÿÿÿÿÿ  Æ %x n   "Ž È %x(n   *Ž È $vÿ X  ÿ Ž ä $vÿ Y  ÿ Ž ä $vÿ `  ÿ Ž ä $vÿ a  ÿ Ž ä $rÿÿ   ÿ Ž ä $x	   ÿ Ž Ä %xn   Ž È %x"n   "Ž È %x$n   Ž È %x&n   Ž È $rÿÿ   ÿŽ  ä %x*n   *Ž È r&   ÿàó är    ÿàñ Æ $x'   Ž  ä$x!     Æ �yB   ‘ ¨ �y0   • è �y   • ( �y   • h �y@   • " r.$   ÿàñ Ê $x/%     â r,"   ÿàñ È �y9.B   ‘ b	 $x-#     Ê �y>,   • h �y7,   • h �y6,   • h �y8,   • b r*   ÿàñ Ê $x+     Ê �y4J   • h �yDL   • d xÿ     â/ $x   ÿ Ž ÊO x   ÿÀŽ â $x@   ÿ Ž ä$x€   ÿ Ž â xÿ   0øŽ ä�x0ÿ   0  ä x,   ÿÀŽ â %v    Ž â x;0   øŽ Ä x.ÿ     â$x-    ÿ Ž â �y    ™ ¢  x=,ÿ   øŽ â %v:;    Ž â r,
   à‘ Æ $x   ÿ Ž â x1.   -øŽ â �y::    ™ â rÿ   $~  â %v<=    Ž â x   ÿÀŽ Æ $t-ÿ$   ÿ Ž â xÿ   øŽ â%v01    Ž â �y<<    ™ b	 xÿ     ä $x/$   ÿŽ â �y;0    ™ b! %v,, Z  - Ž È $x   ÿ Ž ä$x   ÿ Ž ä %v    Ž â x   ÿÀŽ ä x   øŽ â $x--   /Ž â �y?    ™ b x3ÿ   @øŽ Ä %v    Ž â �y,   ™ b r.(   ÿàñ ä x@ÿ   @  â �y,   ™ b $x   ÿ Ž ä/ %v23    Ž â �y5    ™ f $x/)     â �y,   ™ b x1@   øŽ Æ �yC2    ™ d %v01    Ž ä �yB.J   • h �y,   ™ h �y,   ™ h# �y=0    ™ h	�y,   ™ h �yE.L   • b $x29   ÿ Ž Ä $xA9€   ÿ Ž â�y,   ™ d x32   ÿÀŽ â $x@9@   ÿ Ž â �yK,    • d x33ÿ   >øŽ äx>ÿ   >  Æ %v23    Ž âx1>   AøŽ äx@@   ÿÀŽ ä �yA2    ™ b	 %v01    Ž â x/@ÿ   7øŽ ä/x7ÿ   7  â $x@9   ÿ Ž â�y>0    ™ b $xF9    ÿ Ž Ä %v./    Ž âxG@   ÿÀŽ ä x77   FøŽ â $xI9   ÿ Ž â �y@.    ™ b x3Gÿ   6øŽ äx6ÿ   6  â %vF7    Ž Æ x76   IøŽ â $x09   ÿ Ž ä/ %v23    Ž â�yFF    ™ $ x10   ÿÀŽ â %v67    Ž â �yH2    ™ $ x11ÿ   8øŽ â$x/9   ÿ Ž â �yI6    ™ " x.ÿ   8  Ä %v81    Ž â �yG,    ™ "# x/.   /øŽ ä r0   ÿàñ â�yJ8    ™ " r2   ÿàó â %v./    Ž È $x1     ä$x3   Ž  â�y..    ™ ( �y00j   ‘ ( �y22j   ‘ " %x6n   Ž È %x8n   Ž ä �y66    • ( �y88    • " x/4ÿ   ÿÀŽ Ê/ $xM4 @  /Ž â x1ÿ   4  Æ $x/M€   ÿ Ž ä $x11 @  1Ž Æ xL/€€ÿÀŽ â $x4M   ÿ Ž ä $x-1€   ÿ Ž âx/L€€€€ÿüŽ â $x71   ÿ Ž â x1Dÿ   ÿÀŽ ä x3/ÿþþþÿàÿ ä x/4€€ÿÀŽ Ä x--€€ÿÀŽ ä r,3L   ÿüŽ ä x3/€€€€ÿüŽ â $x1D @  1Ž â x9-€€€€ÿüŽ ä x77€€ÿÀŽ ä x43ÿþþþÿàÿ Ä xL9ÿþþþÿàÿ â $x91€   ÿ Ž â x37€€€€ÿüŽ ä xDÿ   D  ä r/4/   ÿüŽ â $xM1   ÿ Ž â x43ÿþþþÿàÿ â $x3D @  DŽ â x99€€ÿÀŽ Ä r747   ÿüŽ â $x43€   ÿ Ž â xD9€€€€ÿüŽ â $x33   ÿ Ž â xMM€€ÿÀŽ ä r-L-   ÿüŽ ä xDDÿþþþÿàÿ ä xLM€€€€ÿüŽ Ä x33€€ÿÀŽ ä r9D9   ÿüŽ ä xLLÿþþþÿàÿ ä xD3€€€€ÿüŽ ä x,,˜º  ÿ    ä x44€€ÿÀŽ Ä rLLM   ÿüŽ ä xMDÿþþþÿàÿ ä xD,ÿÀŽ ä x14€€€€ÿüŽ ä x//˜º  ÿ    ä x11ÿþþþÿàÿ Ä x--˜º  ÿ    ä r114   ÿüŽ ä x47˜º  ÿ    ä rMM3   ÿüŽ ä xLL˜º  ÿ    ä x,,€€€€ÞŽ ÊO $xD,   D
Ž â x,/ÿÀŽ äx//€€€€:ÞŽ Ê� $x3/   ,
Ž â x,-ÿÀŽ äx/4ÿÀŽ ä x--€€€€<ÞŽ äxD€€€€´Ž ä x44€€€€;ÞŽ ä xD9˜º  ÿ    â $x--   ,
Ž â x:3€€€€:´Ž â $x,4   /
Ž â x/DÿÀŽ Ä x--€€€€<´Ž ä x4D€€€€?ÞŽ â &r   ÿ   â x<1˜º  ÿ    Æ $x44   /
Ž â x/LÿÀŽ â &r:      â x,,€€€€;´Ž ä xLL€€€€5ÞŽ ä x1<ÿÀŽ â &r-      â x44€€€€?´Ž â $xLL   /
Ž â x<<€€€€CÞŽ Ä x-Bÿ   ÿÀŽ Æ $x<<   1
Ž ä &r,      â x,M˜º  ÿ    â $x1B @  -Ž â xLL€€€€5´Ž â &r4      â xBÿ   B  ä x-,ÿÀŽ äx,,€€€€=ÞŽ â $x/1€   ÿ Ž Ä &rLL      ä $xB @  BŽ ä $x,,   -
Ž ä $x-1   ÿ Ž â x//€€ÿÀŽ â $x1€   ÿ Ž Æ x3-€€ÿÀŽ ä x-/€€€€ÿüŽ ä x51€€ÿÀŽ ä xC<€€€€C´Ž ä x4-ÿþþþÿàÿ ä x-5€€€€ÿüŽ Ä x<Eÿ   ÿÀŽ ä x13€€€€ÿüŽ ä r/4/   ÿüŽ â $x<E @  <Ž â x4-ÿþþþÿàÿ â $x   ÿ Ž â x:1ÿþþþÿàÿ â $x-<€   ÿ Ž â r145   ÿüŽ Ä x5ÿ   E  â $x4<   ÿ Ž â x€€ÿÀŽ ä x7-€€ÿÀŽ â $x55 @  5Ž â x,,€€€€=´Ž ä x-€€€€ÿüŽ â $x=5€   ÿ Ž â x94€€ÿÀŽ Ä x47€€€€ÿüŽ ä r3:3   ÿüŽ ä x:-ÿþþþÿàÿ ä x;9€€€€ÿüŽ ä x<4ÿþþþÿàÿ ä x4=€€ÿÀŽ Ä r-:   ÿüŽ ä xB;ÿþþþÿàÿ ä r<7   ÿüŽ ä x:/˜º  ÿ    ä x74€€€€ÿüŽ â $x55   ÿ Ž â r/B9   ÿüŽ Ä x9:ÿÀŽ äx;7ÿþþþÿàÿ ä x::€€€€AÞŽ ä x73˜º  ÿ    ä x55€€ÿÀŽ â $x::   9
Ž â r3;4   ÿüŽ Ä x47ÿÀŽ äx77€€€€>ÞŽ ä x95€€€€ÿüŽ ä x11˜º  ÿ    â $x77   4
Ž â xA:€€€€A´Ž ä x:9ÿþþþÿàÿ Ä x91ÿÀŽ äx41€€€€@ÞŽ ä x--˜º  ÿ    ä x˜º  ÿ    â $x94   9
Ž â x4-ÿÀŽ äx--€€€€FÞŽ â&rAA   ÿ   â r1:5   ÿüŽ Ä x>7€€€€>´Ž ä x<ÿÀŽ ä x:/˜º  ÿ    â $x--   4
Ž â x€€€€HÞŽ â &rA>   A   â x99€€€€@´Ž ä x/:ÿÀŽ â$x   <
Ž â x::€€€€IÞŽ â &r99   A   â xF-€€€€F´Ž Ä x33˜º  ÿ    â $x::   /
Ž â x€€€€H´Ž â &r9F   9   â x43ÿÀŽ äx33€€€€JÞŽ â &r   9   â x::€€€€I´Ž ä x11˜º  ÿ    â $x33   4
Ž Ä &r-:      â x41ÿÀŽ äx11€€€€.ÞŽ ä x   ÿÀŽ ä x33€€€€J´Ž â $x11   4
Ž â r0ÿ   0  är2ÿ   2  Ä z ÿ `  pPð â &r33   -   âx.1€€€€.´Ž â $t-ÿ   ÿ Ž â z ÿ a   Sð â $x00   ÿ Ž ä $t/ÿ   ÿ Ž ä $x22   ÿ Ž ä &rCC   L   â x-0   -ìŽ â &r..G   3   âx/2   /ìŽ â &r,,G   C   Ä 0r6ÿ6    A   ä 0rKÿK    A   ä 0r8ÿ8    A   ä $r,,-   ÿŽ ä $r../   ÿŽ â  r76K     A  â  r8K8     A  â Er9 ,       Ä Er6 .       â G‰  0
    € ê r,&   ÿàó Ê $x-'   Ž  â r.    ÿàó È �yA,B    ¢ $x/!   Ž  Ê �y2.    è �y>.    ( �y@.    h �yB.    " r0*   ÿàó Ê $x1+   Ž  Ê �y:0J    h �y;0L    b $x3A   ÿ Ž äO$x-A€   ÿ Ž Æ x33   ÿÀŽ ä x,ÿ   2  ä�x33ÿ   2øŽ ä x-,   -øŽ â $x,A@   ÿ Ž ä %v23    Ž È %v4-    Ž â �y<2    ™ â x-,   ÿÀŽ ä $x/A    ÿ Ž â�y=4    ™ â  x,ÿ   >  â$x.A   ÿ Ž â x--ÿ   >øŽ ä x/,   /øŽ Ä x1.   ÿÀŽ â/ %v,-    Ž Æ x11ÿ   @øŽ â%v./    Ž â�y>,    ™ " x@ÿ   @  ä $x3A   ÿ Ž âO�y?.    ™ b $x0A   ÿ Ž ä %v41    Ž â x3@   3øŽ Ä x10   ÿÀŽ â $xAA   ÿ Ž â x0ÿ   B  â�y@4    ™ b x-1ÿ   BøŽ â/ %v23    Ž â x/0   AøŽ ÆO %v,-    Ž ä �y22    ™ ¢ r0   ÿàó â %v./    Ž ä �y,,    ™ ¤ $x1   Ž  ä�y..    ™ ¨ �y00j    ¢ %x4n   Ž Ì �y44     ¢  xA:ÿ   ÿÀŽ äx/ÿ   :  ä/ xB;ÿ   ÿÀŽ â $xA: @  AŽ ä $x// @  /Ž ä $x3A€   ÿ Ž Ä $x5; @  BŽ â x;ÿ   ;  ä x33€€ÿÀŽ â $xC5€   ÿ Ž ä$x55   ÿ Ž â x-3€€€€ÿüŽ È x:-ÿþþþÿàÿ â $x-A   ÿ Ž ä $xA/   ÿ Ž â r3:3   ÿüŽ ä x:-€€ÿÀŽ â $x-/€   ÿ Ž Æ x/:€€€€ÿüŽ ä x1-€€ÿÀŽ ä x-A€€ÿÀŽ ä x//ÿþþþÿàÿ ä xB1€€€€ÿüŽ ä xAC€€ÿÀŽ Ä r:/:   ÿüŽ ä xBBÿþþþÿàÿ ä x/A€€€€ÿüŽ ä r1B1   ÿüŽ ä xC-€€€€ÿüŽ ä xB/ÿþþþÿàÿ â $x/; @  ;Ž â x;3˜º  ÿ    Ä xDCÿþþþÿàÿ â $xC/€   ÿ Ž âr3BA   ÿüŽ â $x//   ÿ Ž â xB;ÿÀŽ ä x::˜º  ÿ    ä r-D-   ÿüŽ ä x//€€ÿÀŽ Ä x33˜º  ÿ    ä xA;€€€€<ÞŽ ä� x;5€€ÿÀŽ ä x5C€€ÿÀŽ â $xAA   B
Ž â xC:ÿÀŽ äx::€€€€=ÞŽ Ä xB;€€€€ÿüŽ ä xAA€€€€<´Ž â $x::   C
Ž â xBBÿþþþÿàÿ ä xC1˜º  ÿ    â &rAA   ÿ   â r1B;   ÿüŽ ä xBCÿÀŽ Ä xCC€€€€>ÞŽ äx<-˜º  ÿ    ä x::€€€€=´Ž â $xCC   B
Ž â x-<ÿÀŽ ä xD5€€€€ÿüŽ â &rA:   A   â x<<€€€€?ÞŽ Äx;/€€€€ÿüŽ ä xDDÿþþþÿàÿ â $x<<   -
Ž â x:3ÿÀŽ ä xCC€€€€>´Ž ä x33€€€€@ÞŽ ä x>;ÿþþþÿàÿ â &rAC   A   â x11˜º  ÿ    â $x33   :
Ž â r5D5   ÿüŽ Ä x<<€€€€?´Ž ä r>>/   ÿüŽ ä x:1ÿÀŽ â &rA<   A   â x55˜º  ÿ    â 0r4ÿ4    A   âO x11€€€€2ÞŽ ä x->˜º  ÿ    Ä x<5ÿÀŽ â $x11   :
Ž â x55€€€€,ÞŽ ä x:-ÿÀŽ ä x33€€€€@´Ž â $x55   <
Ž â x--€€€€.ÞŽ ä x21€€€€2´Ž â &r33   A   â r0ÿ   0  â $x--   :
Ž â x55€€€€,´Ž â &r32   3   Ä $x00   ÿ Ž â x.-€€€€.´Ž â $t-ÿ   ÿ Ž ä &r35   3   Æ x-0   -ìŽ â &r..G   3   È $r-.-   ÿŽ Ê Er, -       â  r-K4     A  È #r-,      â #r79      â #r86      â G‰  0
    € ê r.$   ÿàñ Ê $x/%     â r0"   ÿàñ È �y2.B    ¢ $x1#     Ê �y80    è  �y50    ( �y40    h �y30    " r,(   ÿàñ Ê $x-)     Ê �y6,J    h	 �y7,L    b	 $x92   ÿ Ž äO$x02@   ÿ Ž Æ x99   ÿÀŽ È x/9ÿ   8øŽ â�$x92€   ÿ Ž â x8ÿ   8  Æ %v./    Ž âx18   9øŽ ä x:0   ÿÀŽ ä �y8.    ™ ¢  %v01    Ž â x-:ÿ   5øŽ äx5ÿ   5  â $x,2    ÿ Ž â�y90    ™ â $x:2   ÿ Ž Æ x55   ,øŽ â %v,-    Ž â x;:   ÿÀŽ Æ %v.5    Ž â�y:,    ™ " x5;ÿ   4øŽ ä$x=2   ÿ Ž â�y;.    ™ " $x02   ÿ Ž â/ x<ÿ   4  â %v45    Ž Æ x=<   =øŽ ä x00   ÿÀŽ â $x,2   ÿ Ž â x1ÿ   3  â�y<4    ™ " x-0ÿ   3øŽ â %v2=    Ž â x/1   ,øŽ Æ%v,-    Ž ä �y22    ™ b r0   ÿàñ â %v./    Ž ä �y,,    ™ d $x1     ä�y..    ™ h �y00j    b %x4n   Ž Ì �y44     b x=6ÿ   ÿÀŽ ä x/ÿ   6  Æ/ $x=6 @  =Ž È $x3=€   ÿ Ž Ê x33€€ÿÀŽ È x-3€€€€ÿüŽ È x6-ÿþþþÿàÿ â $x-=   ÿ Ž ä $x=/ @  /Ž â r/63   ÿüŽ ä x67ÿ   ÿÀŽ â $x3=€   ÿ Ž â x--€€ÿÀŽ â $x==   ÿ Ž ä $x67 @  6Ž â x1-€€€€ÿüŽ Ä x53€€ÿÀŽ â $x36€   ÿ Ž âx>1ÿþþþÿàÿ â $x66   ÿ Ž â x15€€€€ÿüŽ ä x==€€ÿÀŽ ä xA3€€ÿÀŽ ä r3>-   ÿüŽ Ä x>1ÿþþþÿàÿ ä x?=€€€€ÿüŽ ä x1A€€€€ÿüŽ ä x7ÿ   7  ä r5>5   ÿüŽ ä x@?ÿþþþÿàÿ Ä x>1ÿþþþÿàÿ â $x17 @  7Ž â r-@=   ÿüŽ ä x7/˜º  ÿ    â $x=1€   ÿ Ž âr/>A   ÿüŽ â $x11   ÿ Ž â x?6€€ÿÀŽ ä x>7ÿÀŽ Ä x@=€€ÿÀŽ ä x=?€€€€ÿüŽ ä x63˜º  ÿ    ä x3@€€€€ÿüŽ ä x11€€ÿÀŽ ä xA3ÿþþþÿàÿ Ä x//˜º  ÿ    ä x77€€€€8ÞŽ ÊO $x77   >
Ž â x>=ÿþþþÿàÿ ä x=6ÿÀŽ äx66€€€€9ÞŽ ä� x77€€€€8´Ž ä r3>?   ÿüŽ â $x66   =
Ž â x=5˜º  ÿ    â &r77   ÿ   â x-˜º  ÿ    Ä x66€€€€9´Ž ä x8=ÿÀŽ äx==€€€€:ÞŽ â&r76   7   â xÿÀŽ äx€€€€;ÞŽ â $x==   8
Ž â x-1€€€€ÿüŽ Ä x33˜º  ÿ    â $x   
Ž â x==€€€€:´Ž ä x8-ÿþþþÿàÿ ä x€€€€;´Ž â &r7=   7   â r5A@   ÿüŽ ä x6/ÿÀŽ â&r7   7   â x//€€€€<ÞŽ Ä r881   ÿüŽ ä x3ÿÀŽ â $x//   6
Ž â x55˜º  ÿ    ä x33€€€€2ÞŽ äx8˜º  ÿ    ä x5ÿÀŽ â $x33   
Ž â x55€€€€,ÞŽ â 0r4ÿ4    A   â xÿÀŽ Ä x//€€€€<´Ž â $x55   
Ž â x€€€€.ÞŽ â  r4K4     A  â x23€€€€2´Ž â &r//   7   â r0ÿ   0  â $x   
Ž â x55€€€€,´Ž â &r/2   /   ä $x00   ÿ Ž â x.€€€€.´Ž â $tÿ   ÿ Ž Ä &r/5   /   Æ x0   ìŽ â &r..G   /   È $r.   ÿŽ Ê Er        Ê #r4      â x		   ÿàÿ ä x¸  ÿàó ä | 	   p`ð ä x¸  ÿàõ â $rÿÿ   Ž  â x    ÿàÿ Æ $rÿÿ    Ì D	       À â Gy  PÕÿÿÿÿƒ ê Ay       € ê $rÿÿ    Ž ä $rÿÿ    Ž ä Ay        € ê z ÿ `  pPð È z ÿ a   Sð Ú G	  P    € ê ‰      ( ‰      b !r       â !r       È/ ‰      ( ‰      b !r       â !r	       È/ ‰ €    ( ‰
	 €    b !r       â !r
	
       È/ ‰ @    ( ‰
 @    b !r       â !r
       È/ ‰      ( ‰      b !r       â !r       â/ Gy  €    € ê ‰	      ( ‰      h ‰      ¨ ‰
      â !r		       â !r       â/ !r       ÈO ‰      " !r
       Æ� ‰
	      h ‰      ¨ ‰      â !r       â !r
	
       È/ ‰ €    " !r       ÆO ‰
 €    b !r       Æ� ‰ €    ¨ ‰ €    â !r       â !r
       â/ !r       ÈO ‰ @    " !r       Æ� ‰ @    h ‰ @    ¨ ‰ @    â !r       â !r       â/ !r       ÈO ‰	      â !r
       Æ� ‰      ( ‰      h ‰
      ¢ !r       â !r       â/ !r	       â !r
       âO y      !   ( y      %   $ $x   Ž Ê z  p  p`ð È x    pDp  Ú M	        € ê z ÿ ^  pPð â $z v  Ž Æz ÿ _   Sð Ú % ^    Ž Ì �	     ¢ $x   ÿ Ž Ê r ÿ   p"ò äx    p"ö äx    p"ø äx    p"ô Î $ÿÿ    Ž â z ÿ `  pPò â $2ÿÿ    Ž ä $Bÿÿ    Ž â z ÿ a  Sò â $"ÿÿ    Ž È !       ÐO G™      € ê z ÿ b  pPð È z ÿ c   Sð Ú % b    Ž Ì �	     ¢ $x   ÿ Ž ä $vÿ h  ÿ Ž Æ x øÿÿÿp"ò äx üÿÿÿp"ô är ÿ   p"ö äx    p"ø Î $ÿÿ    Ž â x    pBò â $"ÿÿ    Ž ä $2ÿÿ    Ž ä $Bÿÿ    Ž È !       ÈO G  P    € ê x    pRð Ú G‰  �     € ê x    pRð Ú G	  ð    € ê  x;ª¸¿  A  Ì s        $ !t  €?    Ì s        $  r     A  È  r     A  â Gy  P    € ê  x'7=  A  â x*BL?  A  Æ #t  €?   â  x   ?  A  Æ  r     A  Ì s     $   $ !t  €?    È  r     A  È  r     A  â Gy  °    € ê x    pRð Ú G‰  à     € ê x    pRð Ú G	        € ê 	z i    � ä 	z i    � Æ  x;ª¸¿  A  â 	z i €  � Ê s        $ !t  €?    Ì s        $  r     A  È  r     A  â Gy  Ð     € ê  r     A  â Gy  °     € ê 	x  à@  � ä 	x  à@  � Æ  x#ÛÙ?  A  â 	x  àÀ  � Æ  x;ª¸¿  A  ä !t  €?    È s        $ !t  €?    Ì s        $  r     A  È  r     A  ä y      "   ( y      &   b $z s  Ž È $z v  Ž È/ %v j    Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y#      "   " ¹z  l      â Ey   %    € â ™x?    â y$      !   b ¹z  F   
   â x&      ÿ  â r( ÿ       â x      ÿ  â r8 ÿ       â x      ÿ  â r  ÿ       â x      ÿ  â x      ÿ  â x       ÿ  â r" ÿ       Ä r #ÿ   pRð är #ÿ   pRò är+#$   ÿ(Ž È/ xÿ   +  È r+   ÿ� È x%ÿ     È | %   pbô Ú G)   #    € ê y      '   " r ÿ       ä xøÿÿÿÿÀŽ â y      %   b x5%   ÿ   ä r++  €ÿàÿ â y      &   ¢ z1  X      ä z/  Y      Ä r& ÿ       ä z45 q  ÿàÿ â 'z y   Ž â x   ÿ   Æ/ $z, }  ÿŽ â r ÿ       Ê 'z s   Ž âO zÿ z    Æ $z p  %Ž âzÿ t    â $z |  %Ž â x   ÿàÿ â $z |  Ž ä $z v  Ž ä$z v  Ž â r++   ÿ� â $z w  ÿŽ Ä $z p  Ž â xÿ     ä r,,   ÿàõ ä x	ÿ     äz  q      ä x+   ÿ   â %xˆ   Ž â r*5   ÿŽ Ä r05   ÿŽ â %xˆ   ÿŽ â x7   ÿÀŽ ä r6ÿÿ   ÿä â %xˆ   Ž È %x
ˆ   ÿŽ È $x.   5Ž ä$x-   5Ž ä$x2   5Ž ä$x3   5Ž ä %x   Ž È %x   
Ž È r1   ÿàõ È r/   ÿä Ê �yi   ™ ¨  �yh   ™ è  �yf   ™ ( �yg   ™ b r1   ÿàõ È r/   ÿä Ê �ye   ™ è �yd   ™ è �yc   ™ â r1   ÿà÷ är1   ÿàû â�ya   ™ â r1   ÿàõ Ä r	/   ÿäÿ är/   ÿäÿ är/   ÿä â �y`   ‘ â  xkÿ   +  ä r9+,   5àÇ â�yb   ‘ â r)+,   4àë Ä r
1   ÿàõ â �y^   • â xo $       ä r:k6   ÿ„þ â�yJ    • â rk6   ÿÄþ ä r/   ÿä â %v) Z  o Ž È/ %v9 Z  o Ž â �y_   • æ  $x:$   ÿŽ â r>+,   .àÙ â $x)$   ÿŽ â �yF    • ä  r   ÿàÿ ä r)   ÿàÿ Æ �yH   ™ è �yG   ™ è  �yI   ™ è �y;   ™ ( �yV   ™ ( �yU   ™ ( �yT   ™ ( �yS   ™ ( �yR   ™ ( �yQ   ™ ( �yP   ™ ( �yO   ™ " r+,   *àµ Æ �yN   ™ ( �yM   ™ b r<k6   ÿd~ är)k6   ÿ¤~ â �yL    ™ h �yK    ™ b %v Z  o Ž Æ �y9    • b $x=<$   ÿŽ ä %v>> Z  o Ž â �y:    • f $x))$   ÿŽ â r=   ÿàÿ ä r        Ä r?)   ÿàÿ ä r >       â �y<   ™ h �y=   ™ h �y>   ™ h �y)   ™ h �yE   ™ h �yD   ™ h �yC   ™ h �y\   ™ h �yB   ™ h �y[   ™ b r+,   0àÙ Æ/ �yA   ™ b r+,   -àµ Æ �yZ   ™ b rnk6   ÿ¤~ Æ �y@   ™ b rmk6   ÿd~ Æ �yY   ™ b r+,   2àË ärj+,   3àµ â �y?    ™ h �yX    ™ b %v Z  o Ž Æ �yW    • b %v Z  o Ž Æ �y]    • b %v Z  o Ž È %vj Z  o Ž È/ $xon$   ÿŽ ä $xmm$   ÿŽ â rlk6   ÿ„þ ä ro   ÿàÿ ä rm   ÿàÿ â $xql$   ÿŽ ä �y   ™ h �y   ™ b rq   ÿàÿ Æ �yj   ™ h �yq   ™ h �yp   ™ h �yx   ™ h �yo   ™ h �yw   ™ b rkk6   ÿd~ Æ �yn   ™ h �yv   ™ b $xkk$   ÿŽ Æ �ym   ™ h �yu   ™ b rk   ÿàÿ Æ �yl   ™ h �yt   ™ h �yk    ™ h �ys    ™ h �yr    • h �y    • b ‚x �˜­¿     Ä ‚x &     â �yy   ™ b |ˆ       Æ �y~   ™ b xÿ   i  ÆO �y}   ™ ¢ |‚       ä xww  ÿÀŽ â �y|   ™ ¢ xiww  ÿÀŽ Æ �y{   ™ ¨ �yz   ™ ¨ �y   ™ ¨ �y    ™ ¨ �y€    • ¢ |…       Ä |�       â �y„    • ¢  tˆˆÏÝêö    ät‰‚5EYq    ä �y   ™ ¨  �y   ™ ¢  t……ÏÝêö    Ä t†�5EYq    â �y   ™ ¨  �y   ™ ¢  xiiˆˆˆˆÿÀŽ Æ �y‚   ™ ¢  xƒi22ÿø� Æ �y�   ™ ¢  x‡ƒTv  ÿÀŽ Æ �yi   ™ ¢  xŠÿ   ƒ  Æ �yƒ    ™ ¢  r……‡   †    ä x‡ÿ   h  ä� x†hww  ÿÀŽ ä rˆˆŠ   ‰    ä xhhˆˆˆˆÿÀŽ ä |‰       Ä |Š       ä x‡‡ww  ÿÀŽ ä |Œ       ä |‹       ä xhh22ÿø� ä t‰ÏÝêö†    Ä t‰ŠÏÝêö‡    ätŠŒ5EYq‡    ä t†‹5EYq†    ä x‡hTv  ÿÀŽ ä x‹ÿ   h  ä x… d  ˆ    Ä xh…1u  ˆ    ä r†‡   †    ä x…ÿ   f  äxfww  ÿÀŽ äxffˆˆˆˆÿÀŽ ä r‰‰‹   Š    Ä |ˆ       ä |Š       ä |Œ       ä |Ž       ä x‡…ww  ÿÀŽ ä xff22ÿø� Ä t…ˆÏÝêö    ätˆŠ5EYq    ä tŠŒÏÝêö‡    ät‹Ž5EYq‡    ä x‡fTv  ÿÀŽ ä xŒÿ   f  Ä r……‡   ˆ    ä x† d  ‰    äxf†1u  ‰    ä x‡ÿ   g  äx†gww  ÿÀŽ ä rŠŠŒ   ‹    Ä xggˆˆˆˆÿÀŽ ä |‹       ä |Œ       ä x‰‡ww  ÿÀŽ ä |�       ä xgg22ÿø� Ä t‡‹ÏÝêö†    ät‹ŒÏÝêö‰    ätŒŽ5EYq‰    ä tˆ�5EYq†    ä x‰gTv  ÿÀŽ ä x�ÿ   g  Ä xg… d  Š    äx†…1u  Š    ä rˆ‡‰   ˆ    ä x‡ÿ   e  ä x…eww  ÿÀŽ äxeeˆˆˆˆÿÀŽ Ä r‹‹�   Œ    ä |Š       ä |Œ       ä xee22ÿø� ä x‰‡ww  ÿÀŽ ä |Ž       Ä t‡ŠÏÝêö…    ä|�       ä tŠŒ5EYq…    ä x…eTv  ÿÀŽ ä tŒŽÏÝêö‰    ät‰�5EYq‰    Ä x�ÿ   e  ä rŠ‡…   Š    ä x‡ÿ   d  ä x…dww  ÿÀŽ äxddˆˆˆˆÿÀŽ ä r‰Œ�   ‰    Ä |Œ       ä xeˆ d  ‹    ä|Ž       ä xˆˆ1u  ‹    ä xdd22ÿø� ä x‹‡ww  ÿÀŽ Ä |�       ä |’       ä t‡ŒÏÝêö…    ätŒŽ5EYq…    ä x…dTv  ÿÀŽ ä t��ÏÝêö‹    Ä tŽ’5EYq‹    ä x‹ÿ   d  ä rŒ‡…   Œ    ä x‡ÿ   c  ä x…cww  ÿÀŽ äxccˆˆˆˆÿÀŽ Ä r‹�‹   Ž    ä xdŠ d  ‰    ä|Ž       ä xŠŠ1u  ‰    ä |�       ä |’       Ä |”       ä x‰‡ww  ÿÀŽ ä xcc22ÿø� ä t‡ŽÏÝêö…    ätŽ�5EYq…    ä t�’ÏÝêö‰    Ä t�”5EYq‰    ä x…cTv  ÿÀŽ ä x‰ÿ   c  ä rŽ‡…   Ž    ä r‰�‰   �    ä x‡ÿ   a  Ä |�       ä x…aww  ÿÀŽ äxcŒ d  ‹    ä|’       ä xŒŒ1u  ‹    ä xaaˆˆˆˆÿÀŽ Ä x‹‡ww  ÿÀŽ ä t‡�ÏÝêö…    ät�’5EYq…    ä |”       ä |–       ä x…+   ÿ   Ä xaa22ÿø� ä t�”ÏÝêö‹    ät’–5EYq‹    ä rbÿ7   b  ä r^ÿ…   ^  ä x‹aTv  ÿÀŽ Ä x�ÿ   a  ä r…ÿ…   _  ä r`ÿ7   `  ä x_b   ÿÀŽ ä r‡‡‹   �    ä r’��   ’    Ä x……   ÿ   ä xb`   ÿÀŽ ä x�^   ÿ   ä x^‡ d  ’    äx`‡1u  ’    â &r‡G   ÿ   âxb…0   bâŽ â &r…H   ÿ   â x_�0   _âŽ â &r�h;   ‡   Ä &r…hI   …   ä &rHdH   ÿ   ä&rGdG   ÿ   ä &r…V   …   ä&r�U   �   ä &rHŠI   H   Ä &r’Š;   G   ä &r…fT   …   ä&r�fS   �   ä &r;cV   H   â xaŽ d  ‰    â &r…gR   …   ä&r�gQ   �   Ä &r;ŒT   ;   â xŽŽ1u  ‰    â &r…†P   …   ä&r�†O   �   ä &r;aR   ;   ä &r…eN   …   ä&r�eM   �   Ä &r;ŽP   ;   â x__àÿÿÿÿàÿ â &rHˆL   …   ä&r�ˆK   �   ä &r;^N   ;   â xbbàÿÿÿÿàÿ â 0rJÿJ    A   ä 0r9ÿ9    A   Ä 0rFÿF    A   ä $rHH_   ÿŽ ä $r�_�   ÿŽ ä &rG`L   ;   ä 0rIÿ:    A   â  r;J9     A  â r:9F     A  â $rGGb   ÿŽ â ErH H       â  r9JI     A  â Er� �       Ä ErG G       â #r";H   "   â &r;d<   ÿ   â #r 9�       â &r9<   ÿ   ä&r<=   ÿ   â #r!:G   !   â &r=d=   ÿ   ä &r:h>   9   Ä &r<h)   <   ä&r=Š)   =   ä&r:E   :   ä&r)D   <   ä &r>Š>   ;   ä &r:fC   :   Ä &r<f\   )   ä &r>cE   >   ä &r:gB   :   ä&r)g[   <   ä &r>ŒC   >   ä &r:†A   :   Ä &r<†Z   )   ä &r>aB   >   ä &r:e@   :   ä&r)eY   <   ä &rAŽA   >   ä &r:ˆ?   :   Ä &r>ˆX   )   ä $r:_:   ÿŽ ä$r>_>   ÿŽ ä 0rWÿW    A   ä 0r]ÿ]    A   â Er: :       ä Er> >       â  r)JW     A  â r9J]     A  Æ #r):      â #r9>      â &r)   ÿ   ä&r9d   ÿ   ä &r   ÿ   ä &r)hj   )   ä&rhq      Ä &rd   ÿ   ä &r)p   )   ä&rx      ä &rA^@   A   ä &rŠq      ä &r)fo   )   Ä &rfw      ä &rA`?   A   ä &rcx      ä &r)gn   )   ä&rgv      ä $rAbA   ÿŽ Ä &rŒw      ä &r)†m   )   ä&r†u      â ErA A       â  r<FW     A  â &rav      ä &r)el   )   ä&r:et      â #r<A      â &rŽu      Ä &rˆk   )   ä&r<ˆs   :   ä &rt^t      ä 0rÿr    A   ä $r_   ÿŽ ä$r<_<   ÿŽ Ä 0rÿ    A   â  r)J     A  â r:F     A  â Er        ä Er< <       â  rJ     A  ä #r)      â &r   ÿ   âO#r<      â &ry   ÿ   Ä &rydy   ÿ   ä&rd   ÿ   ä &rh~      ä&rh      ä &r9Šj   9   ä&r~Š~   y   Ä &rŠ      ä &r}      ä&r      ä &r’cU   ’   ä&r=cD   =   ä&r9cp   9   Ä &r}c}   ~   ä&rc      ä &rf|      ä&rf      ä &r’ŒS   ’   ä&r\Œ\   =   Ä &r9Œo   9   ä&r|Œ|   }   ä&rŒ      ä &rg{      ä&rg‚      ä &r’aQ   ’   Ä &r[a[   \   ä&r9an   9   ä&r{a{   |   ä&ra‚      ä &r†�      ä &r’ŽO   ’   Ä &r[ŽZ   [   ä&r9Žm   9   ä &r†z      ä&r{Žz   {   ä&rŽ�      â x%%   ÿàÿ â &rei      Ä &r’^M   ’   ä&r[^Y   [   ä&r9^l   9   ä&re      ä&r{^   {   ä&r^i      â | %   pbô â &rˆƒ      Ä &rK`K   ’   ä&r[`X   [   ä&r9`k   9   ä&rs`s   t   ä &rˆ      ä&r{`   {   Ä &rƒ`ƒ      ä 0rÿ€    A   ä $rKbK   ÿŽ ä$r[b[   ÿŽ ä$r9b9   ÿŽ ä$rsbs   ÿŽ Ä $r_   ÿŽ ä$r_   ÿŽ ä 0rÿ„    A   ä $r{b{   ÿŽ ä$rƒbƒ   ÿŽ â  rJ     A  âErK K       â  rJJ     A  â Er[ [       â  rLFI     A  âEr9 9       â  r@F]     A  âErs s       â  rF     A  âEr        â  rF     A  âEr{ {       â  rF     A  â Er        Ä Erƒ ƒ       ä x11@  ÿà÷ â #rLK      â #r@[      â #r:9      â #r s       â #r8   8   â #r({   (   â #r'J   '   â #r&ƒ   &   â x**@   ÿàÿ Ä x..@   ÿàÿ ä x00@   ÿàÿ ä x--@   ÿàÿ ä x22@   ÿàÿ ä x33@   ÿàÿ ä x44@   ÿàÿ Ä x55@   ÿàÿ ä r/ÿ/   ÿäÿ â D)       À â Gy  @ßÿÿÿÿƒ ê Ay        € ê Ey  0    € â G‰      € ê r##$   ÿHŽ Ê ˆs #" øÿ H   è ˆs #!€øÿ H   è ˆs #  ùÿ H   è ˆs #€ùÿ H   è ˆs # úÿ H   è ˆs #€úÿ H   è ˆs # ûÿ H   è ˆs #€ûÿ H   è ˆs # üÿ H   è ˆs #€üÿ H   è ˆs # ýÿ H   è ˆs # €ýÿ H   è ˆs #8 þÿ H   è ˆs #(€þÿ H   è ˆs #' ÿÿ H   è ˆs #&€ÿÿ H   ä Ay        € ê {          ì M        € ê „y$     H   b Ey  @    € æ „y$ €   H   ¨ „y$    H   â !r"       â/ !r!!       È_ ‰      " !r         Æ� ‰!      h ‰       ¢ !r       Æ y      %   " !r!       Æ/ ‰      b !r	        ÆO ‰      ¨ ‰	      â x   ÿ   â !r       Æ/ x   ÿàÿ äz  r  p`ð â !r       âO ‰ €    " z  r  p`ò â !r
	       â� r $ÿ   pVp  â‰ €    b x $   p"ò Æ ‰
 €    ° ‰      &   ä         Ä       &   " !r       Æ       '   " ˆ        â !r       ä/ ‰ @    b !r
       ÆO ‰	 @    ¨ ‰ @    h ‰      '   b $ x  Ž Æ„y
$ €  H   â $ ~  Ž â !r       â/ $Š x  Ž â� !r		       ÆO ‰      " !r       Æ‰	      b $Š ~  Ž Æ ‰      ¢ !r       â %† j   Ž â !r	       Æ/ % j   Ž â †‰     â !r       ÆO †    â G	  �     € ê y      &   ( y      '   b $z x  Ž È $z ~  Ž Ê/ z r  ÿàõ È rÿÿ   ÿä ä z j  ÿ„ È z k   Ê †y     è Ay        € ê !r

       â „y	$    H   â Ey  �    € æ ‰
      ( „y$ €  H   è „y$    H   â !r
       Ê ‰      $ !r       Ê ‰ €    $ !r       Ê ‰ @    $ !r       Ê ‰      $ !r       â G™  °     € ê y      &   " ¹z  r      ä �x   ?àÿ â y      '   b $z x  Ž È $z ~  Ž Ê/ |   ÿàõ È rÿÿ   ÿä ä z j  ÿ„ È z k   Ê †y     è Ay        € ê „y$ €  H   b !r		       â !r       â !r       â „y$    H   ¢ z-  r      Æ „y
$ €  H   ä $x+-   ÿŽ ä„y$    H   " $x#-   ÿŽ Æ „y$ €  H   è „y$    H   è „y$ €  H   è „y$    H   è ‰	      ( ‰      h ‰      b !r       â/ !r       âO !r
       Æ� ‰
      b !r       Æ‰      ¢ !r	       â !r        ä „y	$ €  H   â !r       â!r8       ä y       &   " !r(       â !r       Ä ‰      â !r''       Æ ‰      ( ‰      b !r

       Æ/ y      '   b !r       ÆO ‰      ¨ ‰      ¢ $z   x  Ž Æ ‰      " !r       Æ� ‰'      â !r       Æ‰      " !r       â!r&	&       ä ‰      b $z  ~   Ž â/ !r       äO ‰
      b !r       Æ ‰      ¢ !r       Æ ‰      " !r''       Æ� ‰      â !r       Æ‰      " r-    ÿ„ â !r       ä‰      b rÿÿ   ÿä ä z j  ÿ„ â ‰ €    b !r
       â/ ‚ +   ÿà÷ Ä ‰      b !r       âO z k   ä ‰'      ¢ !r       â ‚ÿÿ   ÿäÿ ä ‰!&      " !r       â� Š
 j  ÿ† Ä ‰ €    â !r       âŠ k  � ä ‰	 €    " !r       â!r       ä ‰ €    b ‚ #   ÿàù â !r       Ä/ ‰ €    b ‚*ÿÿ   ÿä â !r''       äO ‰ €    ¢ Š j  ÿˆ â!r!&!       ä ‰ €    " Š k  * â !r       Ä� ‰  €    â $x-   ÿŽ â!r	       ä‰ @    " r	-    ÿŠ â$x--   ÿŽ â r    ÿàõ â ‰" €    " !r       ârÿÿ   ÿäÿ Ä ‰$ €    b rÿÿ   ÿä â !r       â/ z j  ÿ„ â ‰&' €    b !r       âO z	 j  ÿŠ ä ‰(!      ¢ !r       â z k   Ä ‰ @    " !r         â� z		 k  � ä ‰ @    â !r       â‚ -   ÿà÷ â !r""       â ‰ @    " !r$$       Æ‰ @    b !r&'&       Æ/ ‰ @    b !r(!(       ÆO ‰ @    ¢ !+   ÿàÿ â !r       â #   ÿàÿ â ‰  @    " !r       ä�     ÿàõ â ‰#" @    â ‚ÿÿ   ÿäÿ ä Š j  ÿ† â ‰%$ @    â *ÿÿ   ÿä â !r       â j  ÿ„ â‰)( €    " Š k  � â !r       â k  * â ‰'& @    b ! !   ÿàõ â !r       â/ !r       âO ‰      b *ÿÿ   ÿä Ä ! j  ÿ„ â‰      ¢ !r         ä ! k  * â ‰      " !r#"#       â� !r%$%       ä ‰      â !r)()       Æ‰      " !r'&'       Æ‰      b !r!       Æ/ ‰      b !r       ÆO ‰       ¢ !r       Æ ‰#      ( ‰%      " !r       Æ� ‰$) @    â !r       Æ‰"'      " !r       â!r       ä/ †‰ !    â !r        ÆO †    â !r##       Æ †‰     â !r%%       Æ †     â !r$)$       Æ� †‰     â !r''"       Æ†    è †‰ 
    è †     è †‰ #    è † %   è ‰$      ( †‰ '    â M™        € ê x-   ÿàÿ â/ !r$       Æ r    ÿàñ È r ÿÿ   ÿä  ä z j  ÿ€ È z k     Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      "   " ¹z  l      â Ey  ð!    € â ™x?    â y       !   b ¹z  F   
   â x      ÿ  â x      ÿ  â r ÿ       â x      ÿ  â r- ÿ       â x       ÿ  â x"      ÿ  â r ÿ       â $r.ÿÿ   ÿ Ž â r ÿ   pRð Ä r ÿ   pRò är%    ÿ(Ž È/ xÿ   %  È r%   ÿ� È xÿ     È |    pbô Ú G)        € ê y      '   " r ÿ       â $v+ÿ X  ÿ Ž â r ÿ       â y      %   b xøÿÿÿÿÀŽ ä z/  q      â y      &   ¢ r%%  €ÿàÿ ä x,   ÿ   Ä r
%%   ÿ� ä r$/,   ÿŽ â$x(/   ,Ž âx
ÿ   
  â $x*/   ,Ž âr'/,   ÿŽ â$x//   ,Ž â xÿ   
  ä z)  Y      â 'z y   Ž â r ÿ       Ä z0, q  ÿàÿ â $z& }  ÿŽ â x   ÿ   ä/ z	ÿ z    â 'z s   Ž ÈO $z p  Ž â zÿ t    â $z	 |  Ž ä$z		 |  Ž â x   ÿàÿ â $z v  Ž ä$z w  ÿŽ ä $z	 v  	Ž Ä $z p  Ž â r&&   ÿàõ â %x	ˆ   
Ž Æ r1ÿÿ   ÿä â %x	ˆ   ÿŽ È $x%   ÿ Ž ä %xˆ   ÿŽ Æ x2   ÿÀŽ â%x   Ž È %x
ˆ   
Ž È %x   Ž È r+   ÿàõ È r)   ÿä Ê �ys   ™ ¨  �yJ   ™ è  �yG   ™ ( �yF   ™ b r+   ÿàõ È r)   ÿä Ê �yD   ™ è �y@   ™ è �y>   ™ â x^ÿ   %  ä r%&   ,àµ â �y3   ™ â xa $       Ä r^1   ÿd~ ä rH+   ÿàõ â %v Z  a Ž â rL
+   ÿàù Æ $x$   ÿŽ â r+   ÿàû â$rÿÿ    Ž â r+   ÿà÷ ä r   ÿàÿ ä rI)   ÿä är)   ÿäÿ â�y7   ™ " rM)   ÿä Ä r\%&   $àÙ â�y;H   ‘ " r	)   ÿäÿ ä r%&   0àµ â/ �y9   ™ " %v\\ Z  a Ž ä rK^1   ÿd~ â�y:   ™ " rH^1   ÿ¤~ â %v Z  a Ž Ä �y4   • $ $xIH$   ÿŽ ä �y5   • " $xKK$   ÿŽ ä $x]]   IŽ â �y6   ™ $ rK   ÿàÿ Ä �yK\   ™ ( �yI\   ™ ( �y8   ™ h �y=L   ‘ h �yV\   ™ h �y<   ™ h �yT\   ™ h �y?   ™ h �yR\   ™ h �yA    ™ h �yP\   ™ h �yB    • h �yE    • h �yN\   ™ b r%&   (àµ Æ �yC    • b r%&   'àÙ ä/ r^1   ÿd~ â�yX\    ™ h �yH   ™ b r^1   ÿ¤~ â $x$   ÿŽ Ä �yZ\    • b %v Z  a Ž Æ �yM   ™ b $x[$   ÿŽ Æ �yW   ™ d r[   ÿàÿ ä �yU   ™ h �y]   ™ h# �y\   ™ h �yS   ™ h �yh   ™ h �yQ   ™ h �yf   ™ h �yO   ™ h �yd   ™ h �yY    ™ b %v Z  a Ž Æ �yb   ™ b r%&   /àÙ Æ �yL    • b r   ÿàÿ Æ �y`   ™ h �yj    ™ b r%&   *àµ Æ �y[   ™ b r^1   ÿd~ är^^1   ÿ¤~ â %v Z  a Ž â�yl    • f $xk$   ÿŽ â �y_   ™ b %v Z  a Ž â ‚x &     Ä ‚x �˜­¿     â $xm^$   ÿŽ â rk   ÿàÿ â $~{ÿ   ÿ Ž â |u       â �yi   ™ b rm   ÿàÿ Æ �yg   ™ h �ye   ™ h �yc   ™ h �ya   ™ h �y^    • h �yk    ™ h �yp   ™ h �yo   ™ h �y   ™ h �yt   ™ h �yn   ™ h �ym   ™ h �y   ™ h �y   ™ h! �y    ™ b xrsww  ÿÀŽ ÄO xvsˆˆˆˆÿÀŽ â �yq    • ¢  tzuÏÝêör    ät{{5EYqr    â �yu   ™ ¨ �yr   ™ ¢ xwÿ   s  Æ �ys   ™ ¨ �y   ™ ¢ xyv22ÿø� Æ �y   ™ ¢ x}www  ÿÀŽ Æ �yv   ™ ¨ �yw    ™ ¨ �yx    • ¢ x|yTv  ÿÀŽ ä |~       Ä |€       ä rzz|   {    ä x{JˆˆˆˆÿÀŽ ä� t~~ÏÝêö}    ät}€5EYq}    ä xyÿ   y  Ä x|Jww  ÿÀŽ ä |       ä xJÿ   J  ä |�       ä x{{22ÿø� ä ry~y   }    Ä t}ÏÝêö|    äx~Jww  ÿÀŽ ä t|�5EYq|    ä x{Tv  ÿÀŽ ä/ xz d  y    äxJz1u  y    Ä tyÏÝêö~    â$~ÿ   ÿ Ž â t~�5EYq~    ä r|}   |    ä x{ÿ   {  ä |}       ä xzGww  ÿÀŽ ÄxGˆˆˆˆÿÀŽ ä xGÿ   G  ä ryy{   ~    ä t{}ÏÝêöz    ät~5EYqz    ä xz22ÿø� Ä |€       ä |‚       ä xGww  ÿÀŽ ä x}zTv  ÿÀŽ ä x| d  y    äxG|1u  y    Ä ty€ÏÝêö    ät|‚5EYq    ä xzÿ   z  ä r{{}   ~    ä |~       ä x}Fww  ÿÀŽ Ärzyz   |    ä |€       ä xyFˆˆˆˆÿÀŽ ä xFÿ   F  ä t~~ÏÝêö}    ät}€5EYq}    Ä x|y22ÿø� ä |�       ä |ƒ       ä x€Fww  ÿÀŽ ä x|Tv  ÿÀŽ ä xF{ d  z    Ä xy{1u  z    ä tz�ÏÝêö€    ät{ƒ5EYq€    ä x|ÿ   |  ä r}~   }    â $~ÿ   ÿ Ž â rzz|   {    Ä x{DˆˆˆˆÿÀŽ äx~Dww  ÿÀŽ ä |�       ä xDÿ   D  ä x{{22ÿø� ä t|ÏÝêö~    Ä t�5EYq~    ä |�       ä x€Dww  ÿÀŽ ä x~{Tv  ÿÀŽ ä xD} d  z    äxz}1u  z    Ä t}�ÏÝêö€    ät€ƒ5EYq€    ä r||~       ä x{ÿ   {  ä xÿ   @  ä x~@ww  ÿÀŽ Ä r{}{   €    ä |€       ä xww  ÿÀŽ ä t}�ÏÝêö~    ä |„       ä x�>ww  ÿÀŽ Ä |†       ä t€€ÏÝêö    ät~ƒ5EYq~    ä t‚5EYq    ä t‚„ÏÝêö�    ätƒ†5EYq�    Ä x@@ˆˆˆˆÿÀŽ ä x�ÿ   >  ä x>>ˆˆˆˆÿÀŽ â $~‡ÿ   ÿ Ž â x@@22ÿø� ä |…       ä x„�ww  ÿÀŽ Ä x�>22ÿø� ä x>@Tv  ÿÀŽ ä t……ÏÝêö„    ät†‡5EYq„    ä x@ÿ   @  ä x„�Tv  ÿÀŽ Ä x�ÿ   �  ä r}}>   ~    ä r€€@       ä r‚‚„   ƒ    ä r�…�   †    ä x>| d  {    Ä x@|1u  {    ä x{} d  €    äx}}1u  €    ä x|‚ d  �    äx‚1u  �    ä x€ÿ   3  Ä |�       ä x~3ww  ÿÀŽ ä|ƒ       ä x33ˆˆˆˆÿÀŽ ä x‚€ww  ÿÀŽ ä t€�ÏÝêö~    Ä t�ƒ5EYq~    ä x~322ÿø� â &r37   ÿ   ä&r7>7   ÿ   È &r7@9   7   ä&r3J9   3   ä &r9{:   7   âx7%   ÿ   â &r3:   3   Æ r4ÿ7   4  är5ÿ7   5  â &r7K   ÿ   ä&rK>K   ÿ   ä &r3G6   3   â r;ÿ2   ;  â &rK@I   K   â |…       â &r3F8   3   â r=ÿ2   =  â &rK{V   K   â x;;   ÿÀŽ â &r3y<   3   â x44   ÿ   Ä tƒ…ÏÝêö‚    ät„‡5EYq‚    â &r9}6   9   âx‚~Tv  ÿÀŽ â &rK}T   K   â x~ÿ   ~  â &r7JI   7   â x65   ÿ   â &r5D?   3   â x==   ÿÀŽ Ä x;40   ;âŽ â &rK|R   K   â r€€‚   �    ä rƒƒ~   „    â &r7V   7   â x460   =âŽ â &r6zA   5   â x3;àÿÿÿÿàÿ â &rKP   K   â x:€ d  ƒ    â &r7GT   7   Ä $r663   ÿŽ â x~€1u  ƒ    â 0rBÿB    A   ä 0rEÿE    A   ä &rK:N   K   â Er6 6       â &r7FR   7   â x44àÿÿÿÿàÿ â  r5BE     A  â &rK~X   K   Ä &r7yP   7   â #r.56   .   â &r9|8   9   ä &r5H   ÿ   ä$rK4K   ÿŽ ä 0rCÿC    A   ä &rH>H   ÿ   Ä &r7DN   7   ä 0rZÿZ    A   â ErK K       â &r9<   9   ä &r6JM   5   ä&rH@M   H   ä &r<zX   7   â  r7CZ     A  â &r6W   6   Ä &rH{W   H   â #r 7K       â &r7]   ÿ   ä &r6GU   6   ä&rH}U   H   ä &r7J\   7   ä &r6FS   6   Ä &rH|S   H   ä &r7h   7   ä &r6yQ   6   ä&rHQ   H   ä &r7Gf   7   ä &r9:?   9   Ä &r6DO   6   ä&rH:O   H   ä &r7Fd   7   ä &r9~A   9   ä &r6zY   6   ä $r<3<   ÿŽ Ä &rY~Y   H   ä &r7yb   7   â Er< <       â $r994   ÿŽ â  rHBZ     A  â $rY4Y   ÿŽ ä 0rLÿL    A   ä &r7D`   7   â Er9 9       â $r636   ÿŽ â  r8EC     A  â ErY Y       â #r!H<   !   â  r5CL     A  â &r<zj   7   â #r#89   #   â Er6 6       â  r8BL     A  â #r"5Y   "   â $r<3<   ÿŽ Ä &r5[   ÿ   ä 0rlÿl    A   â #r86      â &r6J_   5   â Er< <       ä  r5Bl     A  â &r[>[   ÿ   Æ #r5<      â &r5   ÿ   ä&r]>]   ÿ   ä&r7>   ÿ   ä &rr   ÿ   äO&r>>r   ÿ   ä &rJp   5   Ä &r8@_   [   ä&r]@\   ]   ä&rp@p   7   ä &rJu      ä&r>@u   >   ä &r6i   6   Ä &r8{i   8   ä&r]{h   ]   ä &ro      ä&rp{o   p   ä &rt      ä&r>{t   >   Ä &r6Gg   6   ä&r8}g   8   ä&r]}f   ]   ä &rGn      ä&rp}n   p   ä &rGs      Ä &r>}s   >   ä &r6Fe   6   ä&r8|e   8   ä&r]|d   ]   ä &rFm      ä&rp|m   p   Ä &rF      ä&r>|   >   ä &r6yc   6   ä&r8c   8   ä&r]b   ]   ä &ry      Ä &rp   p   ä &ryy      ä&r   >   â x   ÿàÿ â &r6Da   6   ä&rD      ä &r8:a   8   Ä &r]:`   ]   ä&r:   p   ä &ryDv   y   ä&r:v      â |    pbô â &r6zk   6   ä&rz      Ä &rk~k   8   ä&r]~j   ]   ä&r~      ä &rzzw   y   ä&r~w      ä 0r^ÿ^    A   Ä 0rqÿq    A   ä $r636   ÿŽ ä$rk4k   ÿŽ ä$r]4]   ÿŽ ä $r3   ÿŽ ä$rz3z   ÿŽ Ä $r4   ÿŽ ä$r4   ÿŽ ä 0rxÿx    A   â  r8B^     A  â rBq     A  â r^C^     A  âEr6 6       â  rlCl     A  âErk k       â  rqCq     A  âEr] ]       â  rBBx     A  â Er        â  rCCx     A  â Er        Ä Erz z       ä Er        ä x++@  ÿà÷ â #r-86   -   â #r^k      â #rl]      â #r      â #rq      â #rBz      â #rC      â x$$@   ÿàÿ â $r)ÿÿ   )Ž â x((@   ÿàÿ Ä x''@   ÿàÿ ä x**@   ÿàÿ ä x//@   ÿàÿ ä x00@   ÿàÿ ä x,,@   ÿàÿ â D)       À â Gy  0âÿÿÿÿƒ ê Ay        € ê Ey      € â G‰  ð     € ê $xÀ   Ž Ê ˆs . ùÿ H   è ˆs #€ùÿ H   è ˆs  úÿ H   è ˆs "€úÿ H   è ˆs ! ûÿ H   è ˆs  €ûÿ H   è ˆs - üÿ H   è ˆs €üÿ H   è ˆs  ýÿ H   è ˆs €ýÿ H   è ˆs  þÿ H   è ˆs €þÿ H   è ˆs  ÿÿ H   è ˆs €ÿÿ H   ä Ay        € ê {          ì M        € ê „y      H   b z5  r      Æ „y     H   ¤ $x15   ÿŽ ä„y  €   H   â $x35   ÿŽ Æ „y     H   ( „y  €  H   h „y  €  H   ( „y
     H   ( „y  €  H   ( „y     H   è „y  €  H   â !r.       Æ/ „y     H   â !r!!       Æ_ ‰      " !r#       Æ� „y  €  H   b !r       Æ„y$     H   ¢ !r""       Æy      %   â !r         Æ „y  €  H   â !r-
-       Æ y      &   " !r       Æ y      '   b !r       Æ ‰      " !r       Æ ‰      ( ‰      " x   ÿ   Æ� ‰	"      â x   ÿàÿ äz  r  p`ð â ‰!      â $z x  Ž â!r       â !r       â/ ‰       b $z ~  Ž â!r       â !r$       âO ‰-      ¢ !r
       â z  r  p`ò Ä ‰      " !r       â r  ÿ   pVp  ä ‰      " !r       â r5   ÿˆ ä ‰$      b !r""	       â� rÿÿ   ÿä â !r!       â ‰&      â !r!       Ä ‚1   ÿàõ â !r         â/ ‰(      b Š r  ÿà÷ â !r--       äO ‰
      ¢ ‚ÿÿ   ÿä ä Š j  ÿ„ â ‰*      ¢ !r       â z j  ÿˆ â !r#       â‰      " Š k   â !r%$       Ä‰	 €    " ‚ÿÿ   ÿäÿ ä z k   â ‰"      b !r'&       â� Š j  ÿ† ä ‰       â !r)(       â/ x     p"ò Ä ‰,!      b !r
       âO Š k  � â $x5   ÿŽ â ‰-      ¢ !r+*       Æ ‰      ¢ !r       â ‚3   ÿà÷ ä ‰      " !r 	       â	5   ÿàÿ â $‚ÿÿ   ÿŽ â ‰$#      " !r""       âr   ÿàû Ä ‰&%      b !r         â� r5   ÿ„ ä ‰('      â !r.!,       â/ 	   ÿà÷ ä ‰ €    b !r--       âO rÿÿ   ÿäÿ Ä ‰*)      ¢ !r       â r0ÿÿ   ÿä ä ‰,+      ¢ z
 j  ÿŠ â !r!       â z j  ÿ„ â ‰ €    " Š j  ÿˆ â !r##$       â2ÿÿ   ÿäÿ â ‰  @    "  j  ÿ† â !r%%&       âz k  � â ‰" €    b !r''(       â� z	 k  0 Ä ‰/.      â !r       â/ Š k   ä ‰- €    b !r))*       âO  k  2� ä ‰ €    ¢ !r++,       â 1   ÿàÿ Ä ‰  €    ¢ !r       â 3   ÿàÿ ä ‰$! €    " !r        â   ÿà÷ ä ‰&# €    " !r""       â   ÿàõ Ä ‰(% €    b !r/./       â� ÿÿ   ÿä ä ‰*' €    â !r--       â/  j  ÿ„ ä ‰ @    b !r       âO ÿÿ   ÿäÿ Ä ‰,) €    ¢ !r         â  j  ÿ† ä ‰.+ €    ¢ !r$!$       â  k  � ä ‰ @    " !r&#&       â k   Ä ‰       " !r(%(       â5        ä ‰" @    b !r*'*       Æ� ‰0/ €    â !r       â/ ˆ        ä ‰ @    b !r,),       äO %† j   Ž â ‰  @    ¢ !r.+.       Æ ‰- @    ¢ !r       Æ ‰!$ @    " !r1        Æ‰#& @    " !r"       Æ‰%( @    b !r0/0       Æ� ‰'* @    â !r       Æ/ ‰      b !r        ÆO ‰), @    ¢ !r--       Æ ‰+. @    ¢ !r!$!       Æ ‰      " !r#&#       Æ‰       " !r%(%       Æ‰/0 @    b !r'*'       Æ� ‰      â !r3       â/ % j  5 Ž ä ‰      b !r),)       ÆO ‰-      ¢ !r+.+       Æ ‰ !      ¢ !r       Æ ‰"#      " !r        Æ‰$%      " !r/0/       Æ†‰ 1    â !r       Æ� ‰'      â !r       Æ/ ‰)      b !r--       ÆO ‰ +      ¢ !r!!        Æ † 3   â !r##"       Æ ‰/      " !r%%$       Æ†‰     è	 †     â	 !r''       Æ� †‰     â	 !r))       Æ/ †    â	 !r++        ÆO †‰ -    è	 † !    â	 !r//       Æ †‰ #    è	 † %   è	 †‰ '    è	 † )    è	 †‰ 
+    â	 M™        € ê †y 
/   â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      "   " ¹z  l      â Ey  0    € â ™x?    â y      !   b ¹z  F   
   â $rÿÿ   ÿ Ž â x      ÿ  â y      %   ¢ x      ÿ  â x      ÿ  â x      ÿ  â x      ÿ  â r ÿ       Ä r ÿ   pRð är ÿ   pRò är	   ÿ(Ž È/ x ÿ   	  â $x   ÿ Ž ÆO r  	   ÿ� È xÿ      È |    pbô Ú G)  à    € ê y      '   " r ÿ       â $z p  Ž â x  øÿÿÿÿÀŽ â y      &   b r ÿ       ä z  X      â $x		    
Ž â z  Y      ä r ÿ       â $x	   ÿ Ž âr@		   ÿ� Ä x;	   ÿ   ä x@ÿ   @  ä x;   ÿÀŽ ä xAÿ   @  ä x ÿ   	  â 'z y   Ž È 'z s   Ž â/ zÿ z    Æ $z
 }  ÿŽ â zÿ t    â $z |  Ž ä$z |  Ž â x   ÿàÿ â $z v  Ž ä$z? v  Ž ä $z p  Ž â z  q      â $z w  ÿŽ Ä %x<?ˆ   @Ž Æ r

   ÿàõ â %x>?ˆ   ÿŽ Æ rÿÿ   ÿä â %x@ˆ   @Ž È $x   ÿ Ž ä %x~ˆ   ÿŽ Æ r   ÿŽ â$x   Ž âz q  ÿàÿ â $x   Ž ä %xB;   >Ž È $x   Ž ä %x:;   ~Ž È r B   ÿàõ È r!C   ÿä Ê �ye    ™ ¨  �yn    ™ è  �yl    ™ ( �yj    ™ b r":   ÿàõ Ê $x#;    Ê �yk"   ™ è �ym"   ™ è �y{"   ™ " r7	
   àµ Æ �yw"   ™ " xf $       ä r    ÿd~ ä r&@   ÿà÷ â %v67 Z  f Ž â r$<   ÿàõ Æ $x$   ÿŽ â rG	
   àÙ â $x'A   Ž âr%=   ÿä ä r77   ÿàÿ ä rD>   ÿàõ â�y$   ‘ " r8~   ÿà÷ ä r(    ÿ¤~ â �y 6   ™ " rE?   ÿä â $x9   Ž Ä �y"6   ™ b! %vFG Z  f Ž Æ �y!&   ‘ b $x%($   ÿŽ Æ �y#6   ™ h �y$D   • b rGG%   ÿàÿ Æ �y&8   • h# �y%6   ™ h �y'6   ™ h �y(F   ™ h �y)6   ™ h �y*F   ™ h �y+6   ™ h �y,F   ™ h �y-6    ™ b rU	
   àµ Æ �y.F   ™ b r3    ÿd~ Æ �y/D    • h �y06    • h �y1F   ™ b %vTU Z  f Ž Æ �y2F   ™ b $x53$   ÿŽ â r[	
   àµ ä �y4F   ™ d rUU5   ÿàÿ ä �y38    • b rD    ÿd~ Æ �y5F    ™ b %vZ[ Z  f Ž Æ �y6F    • b $xED$   ÿŽ Æ �y7T   ™ b $x[[   EŽ Æ �y8T   ™ b# ri	
   àµ Æ �yEZ   ™ h �y9T   ™ b rP    ÿd~ Æ �yGZ   ™ b rg	
   àµ Æ �yDT   ™ h �yIZ   ™ b rV    ÿd~ Æ �yFT   ™ h �yKZ   ™ b %vhi Z  f Ž Æ �yHT   ™ b $xQP$   ÿŽ ä %vfg Z  f Ž â �yJT   ™ h �yLT    ™ h �yMZ   ™ h �yNT    • b riiQ   ÿàÿ Æ �yOZ   ™ h �yPZ   ™ b $xUV$   ÿŽ Æ/ �ySh   ™ d rggU   ÿàÿ ä �yQZ    ™ h �yVf   ™ h �yTh   ™ h �yXf   ™ h �yRZ    • h �yUh   ™ h �yWh   ™ h �yZf   ™ h �y\f   ™ h �yYh   ™ h �y^f   ™ h �y[h   ™ h �y`f   ™ h �y]h   ™ h �ybf   ™ h �y_h    ™ h �ycf    ™ h �yah    • h �ydf    • b ‚x �˜­¿     Ä ‚x &     â $~rÿ   ÿ Ž â |t       â $~sÿ   ÿ Ž â xpÿ   e  äO xoeww  ÿÀŽ äxeeˆˆˆˆÿÀŽ ä |q       Ä xppww  ÿÀŽ ä xee22ÿø� ä thrÏÝêöo    ä/tgt5EYqo    ä tiqÏÝêöp    ä xfeTv  ÿÀŽ Ä tps5EYqp    ä xeÿ   e  ä rghf   g    ä rfie   p    ä xhÿ   n  ä� xenww  ÿÀŽ Ä xnnˆˆˆˆÿÀŽ â $~tÿ   ÿ Ž â |p       ä |v       ä xohww  ÿÀŽ ä |r       ä xnn22ÿø� Ä thpÏÝêöe    ä tptÏÝêöo    ätqv5EYqo    ä tir5EYqe    ä xonTv  ÿÀŽ ä xnÿ   n  Ä xeg d  f    äxgg1u  f    ä xflˆˆˆˆÿÀŽ ärhho   i    ä ripn   q    ä xoÿ   l  Ä |q       ä xnlww  ÿÀŽ â $~uÿ   ÿ Ž â xff22ÿø� ä xpoww  ÿÀŽ ä |y       ä tlqÏÝêön    Ä tos5EYqn    ä xnfTv  ÿÀŽ ä xrÿ   f  ä tquÏÝêöp    äxfh d  i    ä tpy5EYqp    Ä xhh1u  i    ä riln   o    ä xnÿ   j  äxojww  ÿÀŽ äxjjˆˆˆˆÿÀŽ ä rlqr   p    â $~tÿ   ÿ Ž â |r       Ä |q       ä xpnww  ÿÀŽ ä xjj22ÿø� ä tnrÏÝêöo    ä tqqÏÝêöp    ätrs5EYqp    Ä tot5EYqo    ä xpjTv  ÿÀŽ ä xsÿ   j  ä xji d  l    äxii1u  l    ä rlnp   o    Ä xnÿ   k  ä xpkww  ÿÀŽ äxkkˆˆˆˆÿÀŽ ä roqs   r    â $~sÿ   ÿ Ž â |q       ä xrnww  ÿÀŽ Ä xkk22ÿø� ä tnqÏÝêöp    ätqs5EYqp    ä tsuÏÝêör    äxpkTv  ÿÀŽ ä try5EYqr    Ä xtÿ   k  ä xkl d  o    äxll1u  o    ä ronp   q    ä rnst   r    ä xpÿ   m  Ä |s       ä xtmww  ÿÀŽ äxqmˆˆˆˆÿÀŽ ä xupww  ÿÀŽ â $~mÿ   ÿ Ž â tpsÏÝêöt    ä |r       Ä xsq22ÿø� ä tqy5EYqt    ä trrÏÝêöu    äxtsTv  ÿÀŽ ä tum5EYqu    ä xsÿ   s  Ä rppt   q    ä rsrs   u    ä |q       ä xr{ww  ÿÀŽ ä |u       ä tqqÏÝêör    Ä tru5EYqr    ä |u       ä xvwww  ÿÀŽ ä xmo d  n    äxoo1u  n    ä xnp d  s    Ä xpp1u  s    ä tuuÏÝêöv    äxsÿ   {  ä tvy5EYqv    ä xyÿ   w  â $~xÿ   ÿ Ž â x{{ˆˆˆˆÿÀŽ â $~zÿ   ÿ Ž â x}sww  ÿÀŽ Ä xyyww  ÿÀŽ ä x{{22ÿø� ä |t       ä tsxÏÝêö}    ätxxÏÝêöy    ätyz5EYqy    Ä xwwˆˆˆˆÿÀŽ ä xz{Tv  ÿÀŽ ä ttt5EYq}    ä x{ÿ   {  ä xww22ÿø� ä rqqz   r    â &rre    ÿ   ârts{   t    â &rsm    ÿ   â x|wTv  ÿÀŽ â &rrg"   r   â xwÿ   w  Ä x q d  t    äxqq1u  t    â &rto"   s   â ruu|   v    â &rrf#   r   â rxxw   y    ä rÿ     är$ÿ   $  â &r#n#   t   â r!ÿ   !  Ä r&ÿ   &  â &rrh%   r   âxsu d  x    äx"u1u  x    ä x   ÿÀŽ ä xt$   ÿ   â &r#p%   #   â x$!   ÿÀŽ Ä xu&   ÿ   â &rrj'   r   âx!t0   âŽ ä xu0   $âŽ â &r$ '   #   ä &r#e(   ÿ   ä &rri)   r   ä &r#g*   #   Ä &rrk+   r   â x!!àÿÿÿÿàÿ â &r#f,   #   ä &rrl-   r   ä &r$q)   $   ä &r(m(   ÿ   ä &r#h.   #   Ä $rrr!   ÿŽ ä &r$s+   $   ä 0r/ÿ/    A   ä 0r0ÿ0    A   ä &r%o*   (   ä &r#j1   #   â Err r       â &r&"-   $   â  r$/0     A  â &r%n,   %   Ä &r#i2   #   â xàÿÿÿÿàÿ â #r$r      â &r%p.   %   ä &r$k4   #   ä $r&&   ÿŽ ä &r% 1   %   ä 0r3ÿ3    A   Ä &r$l5   $   â Er& &       â &r%q2   %   â  r#03     A  â $r$!$   ÿŽ ä 0r6ÿ6    A   â #r#&      â &r%s4   %   â Er$ $       â &r#e7   ÿ   â  r&/6     A  â &r("5   %   Ä &r%g8   #   â #r&$      â &r$eE   ÿ   ä $r((   ÿŽ ä &r%f9   %   ä &r&mE   ÿ   ä &r$gG   $   â Er( (       â  r#36     A  â &r%hD   %   Ä &r&oG   &   ä &r$fI   $   â#r#(      â &r%jF   %   ä &r&nI   &   ä &r#hK   $   ä &r%iH   %   Ä &rKpK   &   ä &r7m7   ÿ   ä &r&jM   #   ä &r%kJ   %   ä &r7o8   7   ä &r&iO   &   Ä &r$lL   %   ä &r7n9   7   ä &r&kP   &   ä &r%eS   ÿ   ä$r$!$   ÿŽ ä &rSmS   ÿ   Ä &reeV   ÿ   ä &r7pD   7   ä 0rNÿN    A   ä &r&lQ   &   ä &rVmV   ÿ   â Er# $       â &r%gT   %   Ä &rSoT   S   â  r$/N     A  â &r7 F   7   ä $r&!&   ÿŽ ä &regX   e   ä&rooX   V   ä 0rRÿR    A   Ä &r%fU   %   ä&rSnU   S   â #r$#      â &r7qH   7   â Er& &       â &refZ   e   ä&ronZ   o   â  r$/R     A  â &r( M   K   Ä &r%hW   %   ä&rSpW   S   ä &r7sJ   7   ä &reh\   e   ä&rop\   o   â #r$&      â &r(qO   (   Ä &r$jY   %   ä&r& Y   S   ä &rL"L   7   ä &rej^   e   ä&ro ^   o   ä &r(sP   (   Ä &r$i[   $   ä&r&q[   &   ä $rLL   ÿŽ ä &rei`   e   ä&roq`   o   ä &r("Q   (   â x   ÿàÿ â &r$k]   $   Ä &r&s]   &   â ErL L       â &rekb   e   ä&rosb   o   â  r#3N     A  â $r((   ÿŽ â |    pbô â &r$l_   $   ä&r&"_   &   Ä &rllc   e   ä&r""c   o   â #r#L      â Er( (       â  r#3R     A  â $r$!$   ÿŽ ä 0r ÿa    A   ä $r&&   ÿŽ Ä $rl!l   ÿŽ ä $r""   ÿŽ ä 0rdÿd    A   â #r#(      â Er# $       â  r$/      A  â r3      A  â Er& &       â  r /d     A  â Erl l       â  r33d     A  â Er" "       Ä x@  ÿà÷ â #r$#      â #r&      â #r l      â #r3"      â x@   ÿàÿ â $rÿÿ   Ž â x@   ÿàÿ ä x@   ÿàÿ ä x@   ÿàÿ Ä x@   ÿàÿ ä x@   ÿàÿ â D)       À â Gy  àäÿÿÿÿƒ ê Ay        € ê Ey  ð     € â G‰  Ð     € ê $x€  Ž Ê ˆs  úÿ H   è ˆs €úÿ H   è ˆs  ûÿ H   è ˆs €ûÿ H   è ˆs  üÿ H   è ˆs €üÿ H   è ˆs  ýÿ H   è ˆs €ýÿ H   è ˆs  þÿ H   è ˆs €þÿ H   è ˆs  ÿÿ H   è ˆs €ÿÿ H   ä Ay        € ê {          ì M        € ê „y €   H   b z  r  p`ð Ä z+  r      â „y      H   ¢ r ÿ   p"p Æ „y    H   â $x'+   ÿŽ ä$x)+   ÿŽ â „y    H   ( „y €  H   h „y    H   ( „y	 €  H   ( „y €  H   ( „y €  H   ( „y    H   è „y
    H   â !r       â? !r         ÄO „y €  H   â !r       Æ� ‰      " !r       Æ‰       b !r       â!r       È ‰      ¢ !r		       Æ ‰      â !r       â !r       â !r       Æ ‰      " !r       â !r        ä/ ‰      b !r
       Æ ‰      " !r       ÆO ‰       ¢ !r       â !r       ä� ‰      è ‰
	      h ‰      b !r       Æ ‰      " !r       Æ/ ‰      b !r       Æy      &   " !r        âO !r       ä� ‰ €    ¢ x   ÿàÿ â !r		
       ä‰      â z  r  p`ò â !r       Ä ‰      b x    p"ò â !r       ä y      '   " !r       Æ/ ‰ €    b $z  x  Ž Æ‰      ( ‰	      " !r
       ÆO ‰      ¢ !r       Æ� ‰       è ‰      â $z  ~   Ž â !r       ä‰      " !r       â/ r+    ÿˆ Ä ‰"      b !r       â
  r  ÿà÷ â$rÿÿ   ÿ â ‰$      " !r	       â  '   ÿàõ â !r       âO ‰ €    ¢ ÿÿ   ÿäÿ â !r         Ä� ‰
 @    â 
 j  ÿ† â !r       ä ‰ €    b z j  ÿˆ â !r       ä ‰	 @    " ÿÿ   ÿä â !r""       Ä/ ‰ €    b 
 j  ÿ„ â !r$$       ä‰ €    " 
 k  � ä z k   â ‰ €    " 
 k   â !r       ÄO ‰!  €    ¢ !r
       â� r+    ÿ„ ä ‰ €    â !r       âr&ÿÿ   ÿä ä ‰ €    b !r	       â z
 j  ÿ„ Ä ‰#" €    " !r       â/ z k  & ä ‰%$      b !r       â	+   ÿàÿ â !r       â ‰ @    "  )   ÿàù â !r! !       ÄO ‰ @    ¢ ( 	   ÿà÷ â $ÿÿ   ÿ â !r       â� ‰      â 
 j  ÿˆ â !r       ä‰ @    b +ÿÿ   ÿäÿ â !r#"#       Ä ‰ @    " ( j  ÿ† â !r%$%       ä/ ‰ @    b 
	 k   ä ( k  +� â ‰"! @    b !r       â        Ä ‰ @    " !r       ÆO ‰  @    ¢ !r+       â� '   ÿàÿ ä ‰      â !r       Æ‰$# @    b !r       Æ ‰&% €    " !r       â/ !r"!"       â $!ÿ   ÿ Ž â ‰      b !r       Æ‰      " !r         ÆO ‰      ¢ !r-       â�     ÿàõ ä‰      â %  j   Ž â!r$#$       â(ÿÿ   ÿä â ‰      b %  j  ! Ž â !r&%&       â  j  ÿ„ â‰      &  k  ( â ‰!       " !r       Æ/ ‰#"      b !r       Æ‰'& @    " !r       ÆO ‰%$      ¢ !r       Æ� †	 +    â !r       Æ† -   â !r       Æ †	     â !r! !       â !r#"#       ä/ †     â !r'&'       Æ†	     â !r%$%       ÆO †    è †	     è †     è †	 
!    è † 
#   è ‰'      ( †	 %    â M™        € ê x)   ÿàÿ â� !r''       Æ r    ÿàñ È r ÿÿ   ÿä  ä z j  ÿ€ È z k     Ê †y '    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      "   " ¹z  l      â Ey  €    € â ™x?    â y      !   b ¹z  F   
   â r ÿ       â $r"ÿÿ   ÿ Ž â y!      %   ¢ $r-ÿÿ   ÿ Ž â x      ÿ  â x      ÿ  â y      &   â x      ÿ  â r ÿ   pRð Ä r ÿ   pRò â$x#    Ž ä/ $x!!   ÿ Ž ÆO x ÿ   #  È r #   ÿ� ä r  ÿ       ä x ÿ     È |     pbô Ú G)       € ê y      '   "Ž r ÿ       â $rÿÿ   ÿ Ž â xøÿÿÿÿÀŽ â $z	 w  ÿŽ âz  X      â 'z s   Ž â r##  €ÿàÿ ä z'  Y      â $vÿ q  ÿ Ž â r
##   ÿ� â $x*    ÿ Ž â zÿ t    â $rÿÿ   ÿ Ž â x
ÿ   
  â $x&   *Ž âr%*   ÿŽ Ä xÿ   
  ä r(*   ÿŽ ä z)* q  ÿàÿ â 'z y   Ž È $z$ }  ÿŽ â zÿ z    â $z! p   Ž Æ r$	$   ÿàõ â $z |   Ž ä$z |  Ž â x!   ÿàÿ â $z v  Ž ä$z	 v  Ž ä $z p  Ž ä %x	ˆ   
Ž È %x	ˆ   ÿŽ È %x
ˆ   
Ž È $x#   ÿ Ž ä %xˆ   ÿŽ Æ x,   ÿÀŽ â%x   Ž È $r+ÿÿ   ÿ ä %x   Ž È r   ÿàõ È r'   ÿä Ê �yV   ™ ¨  �yU   ™ è  �yT   ™ ( �yS   ™ b r   ÿàõ Ê $x   ' Ê �yR   ™ è �yQ   ™ ( �yP   ™ " xXÿ   #  ä r#$   *àµ â �yK   ™ " x[ $       Ä r.X+   ÿd~ Æ %v Z  [ Ž È $x.$   ÿŽ È $x   Ž â r   ÿàõ È �yL   ™ " r2   ÿà÷ är4   ÿàù â �yM   ™ b r'   ÿä â$x3	   'Ž â r5'   ÿä â �yN   ™ b r#$   )àÇ ä/ r6
   ÿàõ â �yA   ‘ h �yO   ™ b $x7   ' Æ �yE   ™ b rX+   ÿ„þ â/ %v Z  [ Ž ä �yD4   • d $x$   ÿŽ ä �y16   ‘ b r#$   %àµ Æ �yC   ™ b r   ÿàÿ ä r/X+   ÿd~ â �yB   ™ h �y?   ™ b $x7/$   ÿŽ Æ �y@    ™ b %v Z  [ Ž Æ �y.2   • h �y/   ™ b $x   7Ž Æ �y=2    • h �y>4    • h �y0    • h �y<   ™ h �y3   ™ h �y2   ™ h �y:   ™ h �y;   ™ h �y8   ™ b r#$   &àµ Æ/ �y9   ™ h �y6   ™ b rGX+   ÿd~ Æ �y7   ™ h �y4    ™ h �y5   ™ b %v Z  [ Ž â rY#$   (àµ Ä �yF    • b $xGG$   ÿŽ Æ �yI   ™ b rXX+   ÿd~ ä rG   ÿàÿ â �yH   ™ b %vY Z  [ Ž Æ �yG    ™ b $xYX$   ÿŽ Æ �yW   ™ b $x   YŽ Æ �yZ   ™ h �yJ    • h �yY   ™ h �yX   ™ h �y   ™ h# �ya   ™ h �y_   ™ h �y`   ™ h �y]   ™ h �y^   ™ h �y[   ™ h �y\   ™ h �y    ™ h �ye   ™ h �yd   ™ h �yc    ™ h �yb    • h �yf    • b ‚x �˜­¿     Ä ‚x &     â |i       â $~kÿ   ÿ Ž â |h       â $~jÿ   ÿ Ž â |l       ä xgÿ   V  äO xVww  ÿÀŽ äxVVˆˆˆˆÿÀŽ Ä xggww  ÿÀŽ ä xVV22ÿø� ä tiÏÝêö    ätk5EYq    ä xVTv  ÿÀŽ ä/ thÏÝêög    Ä thj5EYqg    ä xVÿ   V  ä r       ä xÿ   U  ä� xUww  ÿÀŽ ä rV   h    Ä xUUˆˆˆˆÿÀŽ ä |h       ä xVww  ÿÀŽ ä xUU22ÿø� ä thÏÝêö    ä thiÏÝêöV    Ä tik5EYqV    ä tgj5EYq    ä xVUTv  ÿÀŽ ä xUÿ   U  ä x d      äx1u      Ä rV   g    ä xVÿ   T  äxTww  ÿÀŽ â$~jÿ   ÿ Ž â xTTˆˆˆˆÿÀŽ ä rUhU   i    â $~iÿ   ÿ Ž â xhVww  ÿÀŽ Ä xTT22ÿø� ä tVjÏÝêö    ä tiiÏÝêöh    ätjk5EYqh    ä xhTTv  ÿÀŽ ä xTÿ   T  Ä tgl5EYq    ä x d  U    äx1u  U    ä riiT   j    ä xUÿ   S  äxTSww  ÿÀŽ Ä xSSˆˆˆˆÿÀŽ ä rVVh   g    ä |k       â $~mÿ   ÿ Ž â |j       ä xgUww  ÿÀŽ ä xSS22ÿø� Ä tUkÏÝêöT    ätjjÏÝêög    ätkl5EYqg    ä xgSTv  ÿÀŽ ä thm5EYqT    ä xSÿ   S  Ä rUUg   h    ä rjjS   k    â $~hÿ   ÿ Ž â xgÿ   R  ä xSRww  ÿÀŽ äxRRˆˆˆˆÿÀŽ â $~pÿ   ÿ Ž â xTV d  i    Ä xVV1u  i    ä |n       ä xigww  ÿÀŽ ä xRR22ÿø� ä tghÏÝêöS    äthl5EYqS    Ä tknÏÝêöi    ätlp5EYqi    ä xiRTv  ÿÀŽ ä xmÿ   R  ä xSU d  j    äxRU1u  j    Ä rhgi   h    ä xgÿ   Q  ä xUQww  ÿÀŽ ä rkkm   l    â $~lÿ   ÿ Ž â xQQˆˆˆˆÿÀŽ ä |j       Ä xigww  ÿÀŽ ä xQQ22ÿø� ä tgjÏÝêöU    ätjl5EYqU    ä tlnÏÝêöi    ätip5EYqi    Ä xUQTv  ÿÀŽ ä xmÿ   Q  ä rggU   j    ä rllm   i    ä xiÿ   P  ä xUPww  ÿÀŽ â$~nÿ   ÿ Ž â xPPˆˆˆˆÿÀŽ â $~rÿ   ÿ Ž â xQh d  k    Ä |j       ä xhh1u  k    ä |p       ä xkiww  ÿÀŽ ä xPP22ÿø� ä tijÏÝêöU    Ä tjn5EYqU    ä tmpÏÝêök    ätnr5EYqk    ä xkPTv  ÿÀŽ ä xPÿ   P  ä xUg d  l    Ä riik   j    ä rPmP   n    â $~pÿ   ÿ Ž â xlg1u  l    ä |n       ä xgKww  ÿÀŽ ä xji d  P    Ä xPi1u  P    ä tinÏÝêög    ätnp5EYqg    ä xgÿ   K  ä xKKˆˆˆˆÿÀŽ ä xkgww  ÿÀŽ Ä xgK22ÿø� â &rKL   ÿ   ä&rLQL   ÿ   ä &rKM   K   ä&rLhM   L   â xMgTv  ÿÀŽ â &rKN   K   â|p       â &rLUN   L   â riiM   n    â &rMO   K   â tmpÏÝêök    â &rNlO   L   â tpr5EYqk    â $xO#   ÿ Ž â xgÿ   g  â &rMTE   M   âr1ÿ,   1  â &rNjE   N   â rpmg   p    Ä rDÿO   D  â &rMVC   M   âxKi d  p    â &rNPC   N   â xC1   ÿÀŽ â $xDD   ÿ Ž â xLi1u  p    â &rMSB   M   ä&r1KB   N   â xCD0   CâŽ â &rB?   ÿ   â r.ÿO   .  â &rMR@   M   â rAÿ,   A  â &r@L@   1   â x1Càÿÿÿÿàÿ â &rB/   B   Ä &rCQ?   ÿ   â xAA   ÿÀŽ â 0r=ÿ=    A   â x..   ÿ   â 0r0ÿ0    A   ä 0r>ÿ>    A   ä &rB<   B   ä &rCh/   C   â x..0   AâŽ â $rA@1   ÿŽ â  r@=0     A  â &r/3   B   â  r?0>     A  â &rCU<   C   Ä &r02   ÿ   ä&r2Q2   ÿ   ä &r/T:   /   ä &rCl3   C   ä &r0;   0   ä&r2h;   2   Ä &r/V8   /   ä &rCj:   C   ä &r09   0   ä&r2U9   2   ä &r/S6   /   ä &rCP8   C   â x..àÿÿÿÿàÿ â &r07   0   Ä &r2l7   2   ä &r/R4   /   ä &rCK6   C   ä &r0T5   0   ä&r5j5   2   ä $r/./   ÿŽ Ä 0rFÿF    A   ä &r4L4   C   ä &r5PI   5   â Er/ /       â &r2VI   0   â  r0=F     A  â $r414   ÿŽ ä &r5KH   5   â #r0/      â  r/>F     A  â Er4 4       â &r0LG   5   Ä &r5W   ÿ   ä#r/4      ä &r4Z   5   ä &rWQW   ÿ   ä &r5Y   4   ä &r   ÿ   Ä &rQ   ÿ   ä &r5X   5   ä &rZhZ   W   ä &ra      ä&raha      ä &r5T_   5   Ä &rYUY   Z   ä &r`      ä&raU`   a   ä &rV]   5   ä &rYlX   Y   ä &r^      Ä &ral^   a   ä &rS[      ä &rYj_   Y   ä &rT\      ä&raj\   a   ä &rR      Ä &rP]   Y   ä &rVe      ä&raPe   a   â x     ÿàÿ â &r[K[      ä &r2SH   2   ä&rSd      Ä &raKd   a   â |     pbô â &rL   [   ä &r3RG   2   ä&rRc      ä&rLLc   a   ä $r.   ÿŽ Ä $rMM.   ÿŽ ä 0rJÿJ    A   ä $r3.3   ÿŽ ä $r210   ÿŽ ä0rbÿb    A   ä $r1   ÿŽ Ä $r.   ÿŽ ä $rL1L   ÿŽ ä 0rfÿf    A   â Er        â  r0=J     A  âErM M       â  r=b     A  âErA A       â  r/>J     A  âEr3 3       â  rb>b     A  âEr2 2       â  r==f     A  â Er        â  r>f     A  â Er        Ä ErL L       ä x@  ÿà÷ â #r@M      â #r?A      â #r03      â #r/2      â #r        â #r-b   -   â #r"=   "   â #rL      â x%%@   ÿàÿ Ä x&&@   ÿàÿ ä x((@   ÿàÿ ä x))@   ÿàÿ ä x**@   ÿàÿ ä r'ÿ'   ÿäÿ â D)       À â Gy  `çÿÿÿÿƒ ê Ay        € ê� Ey  Ð     € â G‰  °     € ê $x@  Ž Ê ˆs  ûÿ H   è ˆs €ûÿ H   è ˆs  üÿ H   è ˆs €üÿ H   è ˆs  ýÿ H   è ˆs €ýÿ H   è ˆs   þÿ H   è ˆs -€þÿ H   è ˆs " ÿÿ H   è ˆs €ÿÿ H   ä Ay        € ê {          ì M        € ê „y     H   b z ! r  p`ð Æ „y €   H   ¢ r ÿ   p"p Æ „y    H   è „y	 €  H   ( „y    H   h „y €  H   ( „y    H   ( „y    H   ( „y
 €  H   è „y €  H   â !r       â/ !r       ÈO ‰      " !r       â� !r	       È‰	      b !r       â!r       ä ‰      ¢ !r""       Æ ‰      â !r        â !r       ä ‰      " !r

-       Æ ‰      " !r       Æ ‰       b !r		       Æ/ ‰"      h ‰      b !r       ÆO ‰
      ¢ !r       Æ� ‰      â !r       â !r       ä‰	      " !r         Æ‰      " !r"       Æ/ ‰  €    b !r       Æ ‰      b !r
       ÆO ‰
      ¢ !r       Æ� ‰      â !r	       Æ ‰      " !r       Æ‰      " $z x  !Ž â !r        ä/ ‰      h ‰      b !r       â!r
       ÄO ‰ €    ¢ x
!   ÿàÿ â $v!ÿ r  ÿ Ž â !r       â� ‰	 €    â z 
 r  p`ò â !r       ä y      '   " x    p"ò â $x!   ÿŽ â !r       â‰ €    " !r       Æ/ ‰  @    b !r       Æ ‰ €    h ‰ €    b !r       ÆO ‰ €    ¢ !r	       Æ� ‰ €    â $z
 ~  Ž Æ ‰ €    " !r       âr!
   ÿ† ä‰ €    " !r         â/ 	
   ÿàõ â !r       â$rÿÿ   ÿŽ â ‰ @    b z j  ÿ† â !r       â  ÿÿ   ÿä â ‰ @    b z k  � â !r       âO 
	 j  ÿ„ â ‰ @    ¢ r!
   ÿŠ â !r       â� 
		 k    â ‰       â 

 r  ÿàù â !r       â $r"ÿÿ   ÿŽ â ‰ @    " z j  ÿŠ â!r       â$ÿÿ   ÿä â ‰ @    " z k  "� Ä 
 j  ÿˆ â ‰ @    " !r       â/    ÿàÿ ä ‰ @    b !r       â
   ÿà÷ ä ‰ @    b !r       âO  ÿÿ   ÿäÿ Ä ‰ @    ¢ !r#        â� !   ÿàÿ â !r       â ‰       $ !
   ÿàõ â!r       â‰      â 
 k  $ ä $"ÿÿ   ÿ â !r       â x        â ‰      " ! j  ÿ„ â !r       â/  j  ÿ† â ‰      b %v

 j   Ž â !r       â k   � â ‰      b !r       ÆO ‰      ¢ !r%        Æ ‰      " ! k  " â !r       ä� ‰      è ‰      â !r       â!r       Ä/ †	 
#    â !r       Æ† 
%   â !r       ÆO †	     â !r       Æ †     â !r       Æ� †	     â !r       Æ †    è †	     è †     è †	     â M™        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y       "   " ¹z  l      â Ey       € â ™x?    â y      !   b ¹z  F   
   â r% ÿ       â x,      ÿ  â y!      %   ¢ $r"ÿÿ   ÿ Ž â r/ ÿ       â $rÿÿ   ÿ Ž â y      &   â $r0ÿÿ   ÿ Ž Ä $r ÿÿ   ÿ Ž â y      '   " r  ÿ   pRð är  ÿ   pRò â$x#     Ž â/ x!!   ÿ   ÈO xÿ   #  È r#   ÿ� È x.ÿ     È | .   pbô Ú G)  �    € ê xøÿÿÿÿÀŽ â�$rÿÿ   ÿ Ž â z'  Y      â $z$ }  ÿŽ ä $x##   
Ž â r ÿ       â 'z y   Ž Æ r##   ÿ� â 'z s   Ž â zÿ z    ä xÿ     â $z! p  .Ž â zÿ t    â $z	 |  .Ž âxÿ     â $z |  Ž â x!   ÿàÿ â $z	 v  	Ž â x)#   ÿ   â $z w  ÿŽ Ä $z v  Ž ä $z p  	Ž â r$$   ÿàõ â %x
ˆ   Ž È %xˆ   Ž È %xˆ   ÿŽ È %xˆ   ÿŽ È $v&ÿ q  ÿ Ž ä $x+.   ÿ Ž ä $vÿ X  ÿ Ž ä $x&   +Ž âz(+ q  ÿàÿ â $r%ÿÿ   ÿ Ž ä $x&&   +Ž Ä $r*ÿÿ   ÿ ä %x)   Ž È %x)   Ž È r   ÿàõ È r'   ÿä Ê �y\   ™ ¨  �yP   ™ è  �yO   ™ " r   ÿàõ Æ �yN   ™ d $x   ' Ê �yC   ™ è �yB   ™ â xÿ   #  ä r#$   +àµ â �y1   ™ " $tUÿ$   ÿ Ž ä r*   ÿd~ â �y9   ™ " %v Z  U Ž â r#$   (àµ Æ $x$   ÿŽ Ê r   ÿàÿ Ê �y=   ™ " r*   ÿd~ Æ �y>   ™ " %v Z  U Ž Æ �y<   ™ " $x$   ÿŽ Æ �y:   ™ b $x   Ž Æ �y?   ™ b r@
   ÿàõ Æ �y4   ™ b r   ÿàù Æ/ �y5   ™ b $xA   ' âr2   ÿà÷ ä�y7   ™ b r   ÿàõ ä r'   ÿä â �y6   ™ b $x3	   'Ž Æ �y;   ™ b $x   ' Æ �y@@   ‘ h �yD   ‘ h �yJ   ™ h �yF2   • h �yI   • h �yH   ™ h �yG   ™ b r#$   àµ Æ �yE   ™ b r#$   &àÙ ä r*   ÿd~ â�yA    ™ b rQ*   ÿ¤~ â %v Z  U Ž ä �y8    ™ d $xS$   ÿŽ Ä �yK2    • b %v Z  U Ž Æ �yL    • b $xQQ$   ÿŽ Æ �y    • h �yM    • b rS   ÿàÿ â $x   QŽ È �y   ™ h �y   ™ h# �yY   ™ h �yW   ™ h �yX   ™ h �yU   ™ h �yV   ™ h �yS   ™ h �yT   ™ h �yQ   ™ h �yR   ™ h �y2   ™ h �y3   ™ h �y   ™ h �y    ™ h �y    ™ h �yZ    • h �y[    • b ‚x �˜­¿     Ä ‚x &     â $~_ÿ   ÿ Ž â |a       â $~ÿ   ÿ Ž ä $~ÿ   ÿ Ž â/ |b       â $~cÿ   ÿ Ž â |e       â $~gÿ   ÿ Ž â x]\ˆˆˆˆÿÀŽ ÄO x`ÿ   \  ä x^\ww  ÿÀŽ ä x]]22ÿø� ä x``ww  ÿÀŽ ä t\_ÏÝêö^    ät_a5EYq^    Ä x]Tv  ÿÀŽ ä tÏÝêö`    ät`5EYq`    ä x]ÿ   ]  ä r\\   _    ä |       Ä xPww  ÿÀŽ â�$~_ÿ   ÿ Ž â xPˆˆˆˆÿÀŽ ä r]   `    ä t]ÏÝêö    ä xPÿ   P  ä x22ÿø� Ä t^_5EYq    â $~`ÿ   ÿ Ž â x_Pww  ÿÀŽ ä xPTv  ÿÀŽ ä x\ d      äx\1u      ä rP]P   ^    â $~^ÿ   ÿ Ž â t`ÏÝêö_    Ä t\b5EYq_    ä xÿ     ä x]Oww  ÿÀŽ â$~`ÿ   ÿ Ž â r   \    ä t\^ÏÝêö]    ä xOˆˆˆˆÿÀŽ Ä x^ÿ   O  ä t]`5EYq]    ä xO22ÿø� ä |`       ä x_^ww  ÿÀŽ ä xP d      Ä xP1u      ä x^OTv  ÿÀŽ ä tP`ÏÝêö_    ät_b5EYq_    ä xOÿ   O  â $~`ÿ   ÿ Ž â r\\^   ]    â $~^ÿ   ÿ Ž â rOPO   _    Ä x]Nww  ÿÀŽ äxPNˆˆˆˆÿÀŽ ä xNÿ   N  ä t^^ÏÝêö]    ät_`5EYq]    ä x]P22ÿø� Ä |a       ä x`Nww  ÿÀŽ ä xN\ d  O    äxO\1u  O    ä xP]Tv  ÿÀŽ ä t\aÏÝêö`    Ä tac5EYq`    ä x]ÿ   ]  ä x`ÿ   C  ä rP^P   _    ä |_       ä x^Cww  ÿÀŽ Ä r]\]   a    â $~aÿ   ÿ Ž â x``ww  ÿÀŽ ä t\_ÏÝêö^    ät_c5EYq^    ä t^aÏÝêö`    ä xCCˆˆˆˆÿÀŽ Ä tac5EYq`    ä xbBww  ÿÀŽ äx`ÿ   B  ä xBBˆˆˆˆÿÀŽ ä xCC22ÿø� ä xd`ww  ÿÀŽ Ä x`B22ÿø� ä xBCTv  ÿÀŽ ä xCÿ   C  ä tceÏÝêöb    äteeÏÝêöd    ätfg5EYqd    Ä xd`Tv  ÿÀŽ ä tbg5EYqb    ä x`ÿ   `  ä r_\B   _    ä r^^C   a    ä rbcd   b    Ä ree`   f    ä x\_ d  ^    äx^_1u  ^    ä xBP d  ]    ä x`ÿ   1  ä x_1ww  ÿÀŽ Ä xPP1u  ]    ä x11ˆˆˆˆÿÀŽ â $~hÿ   ÿ Ž â xCb d  e    äx]b1u  e    ä |b       ä |f       Ä xa`ww  ÿÀŽ ä x1122ÿø� ä t`bÏÝêö_    â $~dÿ   ÿ Ž â tbfÏÝêöa    ätch5EYqa    ä xa1Tv  ÿÀŽ Ä x1ÿ   1  ä t_d5EYq_    ä rbb1   c    ä x19ˆˆˆˆÿÀŽ äxd9ww  ÿÀŽ ä |e       Ä r_`a   _    ä x9ÿ   9  ä x`122ÿø� â &r1=   ÿ   âtaeÏÝêöd    â &r=\=   ÿ   â xc9ww  ÿÀŽ â &r1>   1   â tdg5EYqd    Ä x9`Tv  ÿÀŽ â &r=^>   =   Æ raa9   d    â &r9<   1   ä&r=C<   =   â x1_ d  b    â &r<:   9   ä&r:]:   =   â t>fÏÝêöc    â &r=N?   <   â tch5EYqc    â &r?1?   :   â x`ÿ   `  Ä x9_1u  b    â &r=O4   =   â r>>`   c    Æ &r?94   ?   ä &r45   ÿ   âx:a d  >    â&r5\5   ÿ   â x<a1u  >    â &r47   4   ä&r>^7   5   â x5)   ÿÀŽ â &r=B6   =   Ä &r?:6   ?   ä &r6;   4   â r@ÿ5   @  â$x4#   ÿ Ž â rDÿ5   D  â &r5J   6   Æ rFÿ4   F  ärIÿ4   I  â &r4NH   5   â x@@   ÿÀŽ ä x7F   ÿ   â &r5OG   4   Æ x470   @âŽ â &r6BE   5   ä &r>C;   >   â x44àÿÿÿÿàÿ â &r;PA   6   ä 0rKÿK    A   ä 0rÿ    A   ä 0rLÿL    A   Ä $r64;   ÿŽ ä 0rMÿM    A   â  r7K     A  â r;L     A  â Er6 6       ä  rKM     A  È #r/6   /   â &r   ÿ   ä&r\   ÿ   ä &r   ÿ   ä&r\   ÿ   ä &rY      ä&r^Y      Ä &rW      ä&r^W      ä &rX      ä&rCX      ä &rU      ä&rCU      Ä &r>]J   >   ä &rV      ä&r]V      ä &rS      ä&r]S      ä &r>1H   >   Ä &rNT      ä&r1T      ä&rNQ      ä&r1Q      â xDD   ÿÀŽ â $xII   ÿ Ž ä &r>9G   >   Ä &rOR      ä&r9R      ä &rOO2      ä&r992      â x5I0   DâŽ â &r>:E   >   â x..   ÿàÿ â &r:3      Ä &rB3      ä&rOB   O   ä&r9:   9   â x55àÿÿÿÿàÿ â &r=P8   =   â| .   pbô â &r8<8   ?   ä&r><A   >   Ä &r<      ä&rP      ä&rOP   O   ä&r<<   9   ä $r==4   ÿŽ ä $r885   ÿŽ Ä $r>5>   ÿŽ ä$r4   ÿŽ ä0rZÿZ    A   ä $r5   ÿŽ ä$rO4O   ÿŽ ä $r<5<   ÿŽ Ä 0r[ÿ[    A   â Er= =       â  rMLM     A  âEr8 8       â  rKZ     A  âEr> >       â  rLZ     A  âEr        â  rKK[     A  âEr        â  rLL[     A  â ErO O       Ä Er< <       ä x€  ÿà÷ â #r 7=       â #r0;8   0   â #rM>      â #r-   -   â #r"   "   â #r,KO   ,   â #r%L<   %   â x€   ÿàÿ Ä x&&€   ÿàÿ ä x((€   ÿàÿ ä x++€   ÿàÿ ä r'ÿ'   ÿäÿ â D)       À â Gy  Àéÿÿÿÿƒ ê Ay        € ê�Ey  °     € â G‰  �     € ê $x    Ž Ê ˆs   üÿ H   è ˆs 0€üÿ H   è ˆs / ýÿ H   è ˆs €ýÿ H   è ˆs - þÿ H   è ˆs "€þÿ H   è ˆs , ÿÿ H   è ˆs %€ÿÿ H   ä Ay        € ê {          ì M        € ê „y     H   " z ! r  p`ð â $z x  !Ž â x!!   ÿàÿ â „y €  H   b r ÿ   p"p Ä z ! r  p`ò â „y    H   ¦ x    p"ò â „y €   H   è „y     H   ( „y €  H   h „y    H   h „y €  H   h „y	    H   h „y €	  H   h „y    H   b !r        Æ „y €  H   " !r       Æ/ „y €  H   b !r-       ÆO „y#    H   ¢ !r0       Æ� „y    H   â !r         Æ„y
 €  H   " !r       Æ„y    H   b !r       Æ „y  	  H   b !r       Æ „y  
  H   b !r  	       Æ „y €
  H   b !r       Æ „y'    H   b !r/       Æ „y €  H   â !r"       Æ „y €  H   â !r       Æ/ „y €  H   â !r##,       âO !r       ä� ‰       " !r

       Æ‰	      b z  r      â !r#       ä‰      ¢ !r       â !r       È ‰      â $x   ÿŽ â !r

       ä ‰      " !r'       Æ ‰
      b !r        â !r%       ä ‰      " !r		       â/ !r       ä ‰       b !r       âO !r       Ä ‰	      ¢ !r       Æ� ‰      â !r       Æ‰      " !r
       Æ‰      b !r       Æ ‰
      " !r         Æ/ ‰      b !r	       ÆO ‰      ¨ ‰  €    ¢ !r       â� !r       ä‰	 €    â !r       Æ‰ €    " !r
       Æ ‰ €    " !r       Æ/ ‰ €    b !r       ÆO ‰ €    ¢ !r        Æ ‰ €    b $z  ~  Ž Æ ‰      b !r		       â� r    ÿ† ä ‰ @    â !r       â!r
       ä ‰	 @    " !r       Æ/ ‰ @    b !r       ÆO ‰
 @    ¢ !r       Æ‰ @    " !r       Æ ‰ @    b !r       â� 
  r  ÿàõ ä ‰ @    â !r		       Æ ‰ €    " $ÿÿ   ÿ â 
 j  ÿ„ ä ‰      " !r       â/ 
 k   â !r
       âO ‰	      b    ÿàÿ â !r       Ä‰      ¢ rÿÿ   ÿäÿ â !r       ä‰      " z j  ÿ† â !r       ä� ‰      â     ÿàõ â !r       Ä ‰      " z k  � â $ÿÿ   ÿ â     ÿà÷ â ‰ @    b !r       â $tÿ   ÿ Ž â  j  ÿ„ â ‰      b ÿÿ   ÿäÿ â !r	       â/ 

 j  ÿ† â %v  j   Ž â 	 k   â !r       âO 
 k  � â !r       â†	     â !r       Æ� †    â !r       Æ †	     â !r       Æ†     â !r       Æ †	     è †    è ‰      ( †	 
    â M™        € ê x   ÿàÿ â/ !r       Æ r    ÿàñ È r ÿÿ   ÿä  ä z j  ÿ€ È z k     Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y!      "   " ¹z  l      â Ey  €    € â ™x?    â y      !   b ¹z  F   
   â $rÿÿ   ÿ Ž â r* ÿ       â y#      %   ¢ $r,ÿÿ   ÿ Ž ä $r ÿÿ   ÿ Ž â y      '   â $r.ÿÿ   ÿ Ž Ä $r ÿÿ   ÿ Ž â r !ÿ   pRð är !ÿ   pRò â$x!    Ž ä/ $x##   ÿ Ž ÆO xÿ     È r   ÿ� â $!!À   Ž â y      &   $ x(ÿ     È | (   pbô Ú G)      € ê $rÿÿ   ÿ Ž â� xøÿÿÿÿÀŽ â $rÿÿ   ÿ Ž â z  q      â 'z y   Ž È $x   
Ž â zÿ z    â 'z s   Ž Æ r   ÿ� â $z# p  (Ž â zÿ t    â $z	 |  (Ž âxÿ     â $z |  Ž â x#   ÿàÿ â $z	 v  	Ž â xÿ     â $z% }  ÿŽ ä $z
 w  ÿŽ Ä $z v  Ž ä $z p  	Ž â r%
%   ÿàõ â %x
ˆ   Ž È %xˆ   Ž È %xˆ   ÿŽ È %xˆ   ÿŽ È $x&(   ÿ Ž ä $x)   ÿ Ž ä $vÿ X  ÿ Ž â r&   ÿŽ â $v'ÿ Y  ÿ Ž â z"& q  ÿàÿ â $rÿÿ   ÿ Ž ä $r$ÿÿ   ÿ Ä %x)   Ž È %x)   Ž È r   ÿàõ Ê $x   ' Ê �yR   ™ ¨  �yS   ™ è  �yQ   ™ ( �y:   ™ b r   ÿàõ Ê $x   ' Ê �y6   ™ è �y5   ™ ( �y4   ™ " x7ÿ     â $t?ÿ$   ÿ Ž â r%   &àµ â �y2   ™ & r7$   ÿd~ â %v Z  ? Ž È $x$   ÿŽ â r   ÿà÷ Æ $x   Ž â r<
   ÿàõ â $x	   'Ž Æ �y0   ™ " r='   ÿä ä r%   "àµ â �y   ™ & r7$   ÿd~ â/ �y+   • b %v Z  ? Ž â r%   àµ Ä �y/<   ‘ b $x$   ÿŽ Æ �y;    • b $x   Ž Æ �y1   ™ b r<7$   ÿd~ â %v Z  ? Ž ä �y=   ™ d $x?<$   ÿŽ ä �y>   ™ d $x   ?Ž Ä �yJ   ™ b r   ÿà÷ Æ/ �y?   ™ h �yI   ™ h �y-   ™ h �y3   ™ h �y9   ™ h �y8   ™ h �y7    ™ h �y<    • h �yH   ™ b $x   'Ž Æ �yG   ™ b r   ÿàõ Æ �yF   ™ d $x   ' ä �yE   ™ h �yD   ™ h �yC   ™ h �yM   • h �y   ‘ h �yB   ™ h �yA    ™ h �y@   ™ h �yN    • h �yL   ™ h �yK    ™ h �yP    • h �yO    • b ‚x &     Ä ‚x �˜­¿     â $~Yÿ   ÿ Ž â |W       â $~ÿ   ÿ Ž ä $~ÿ   ÿ Ž â/ |V       â $~Xÿ   ÿ Ž ä $~Zÿ   ÿ Ž ä $~\ÿ   ÿ Ž â xRˆˆˆˆÿÀŽ ÄO xUÿ   R  ä xTRww  ÿÀŽ ä x22ÿø� ä xUUww  ÿÀŽ ä tRWÏÝêöT    ätY5EYqT    Ä xTv  ÿÀŽ ä tÏÝêöU    ätU5EYqU    â $~ÿ   ÿ Ž â rR       â $~Rÿ   ÿ Ž â xÿ     ä xSww  ÿÀŽ Ä� xSˆˆˆˆÿÀŽ ä xSÿ   S  ä r   U    ä tÏÝêö    â$~Tÿ   ÿ Ž â tR5EYq    ä x22ÿø� Ä |R       ä xSSww  ÿÀŽ ä xTv  ÿÀŽ ä tRRÏÝêöS    äx d      ä tST5EYqS    Ä xÿ     ä x1u      ä r       â $~ÿ   ÿ Ž â xQww  ÿÀŽ ärSR   S    ä xQˆˆˆˆÿÀŽ Ä tÏÝêö    ätT5EYq    ä xQÿ   Q  ä x22ÿø� â $~Uÿ   ÿ Ž â xRQww  ÿÀŽ â $~Wÿ   ÿ Ž â xQTv  ÿÀŽ Ä x d  S    äx1u  S    â $~Tÿ   ÿ Ž â tSUÏÝêöR    ätRW5EYqR    ä rQ       ä xÿ     Ä xQ:ww  ÿÀŽ äx:ˆˆˆˆÿÀŽ ä rSS   R    ä tTÏÝêöQ    ätRV5EYqQ    ä x:ÿ   :  Ä xQ22ÿø� ä xT:ww  ÿÀŽ ä xQTv  ÿÀŽ ä x: d  S    äx1u  S    ä tSUÏÝêöT    Ä tTW5EYqT    ä r   R    ä xQÿ   Q  ä xR6ww  ÿÀŽ äx6ˆˆˆˆÿÀŽ ä rTSQ   T    Ä tQUÏÝêöR    ätVW5EYqR    ä x6ÿ   6  ä xR22ÿø� ä xU6ww  ÿÀŽ ä xSRTv  ÿÀŽ Ä x d  T    äx61u  T    ä tTXÏÝêöU    ä rQS   V    â $~Sÿ   ÿ Ž â tUZ5EYqU    ä xRÿ   R  Ä xX5ww  ÿÀŽ ä xQÿ   5  ä rRTR   U    â $~Tÿ   ÿ Ž â tUSÏÝêöX    â$~Vÿ   ÿ Ž â tXW5EYqX    ä xQQww  ÿÀŽ Ä |Z       ä xW4ww  ÿÀŽ ä x[ÿ   4  ä tST5EYqQ    ätVVÏÝêöQ    ä tTZÏÝêöW    Ä tQ\5EYqW    â $~Wÿ   ÿ Ž â xZ5ˆˆˆˆÿÀŽ ä xY4ˆˆˆˆÿÀŽ ä x4[ww  ÿÀŽ â $~[ÿ   ÿ Ž â xZZ22ÿø� ä xYY22ÿø� Ä t5WÏÝêö4    ät4[5EYq4    ä x[ZTv  ÿÀŽ ä xWYTv  ÿÀŽ ä xYÿ   Y  ä xZÿ   Z  Ä rUU[   X    ä rQTW   Q    ä rX5Y   4    ä rVVZ   S    ä xS2ˆˆˆˆÿÀŽ â $~Zÿ   ÿ Ž â x4 d  R    Ä x51u  R    ä xTQ d  X    äxRQ1u  X    â $~Xÿ   ÿ Ž â xQ2ww  ÿÀŽ ä xWÿ   2  ä xSS22ÿø� Ä xU d  V    äxVU1u  V    ä t2Z5EYqQ    äxYWww  ÿÀŽ ä tUXÏÝêöQ    â &rX0   ÿ   â |Z       Ä xWSTv  ÿÀŽ ä tQZÏÝêöY    â &rZ   X   â rUUW   2    â $x2   ÿ Ž â tX\5EYqY    â &r\40   ÿ   â x0)   ÿÀŽ ä r+ÿ2   +  Ä r/ÿ0   /  â &r\5   \   ä $x+   ÿ Ž â x//   ÿÀŽ È x0   /âŽ â &r/=   ÿ   È &r/>   /   ä &r=4=   ÿ   ä &r/J   /   ä &r?   ÿ   ä&r44?   ÿ   ä &r/I   /   Ä &r=5>   =   ä &rH      ä&r45H   4   ä &r/:G   /   ä &rZ1   Z   ä&r\1   \   Ä &r=J   =   ä &rF      ä&rF   4   ä &rE   /   â xSÿ   S  â &rZ-   Z   ârMÿ2   M  â &r\V-   \   Ä &r=VI   =   ä &rD      ä&rVD      ä &rC      â rXQS   X    â &rZ:3   Z   ârÿ0     â &r\T3   \   âxàÿÿÿÿàÿ â &r=TG   =   Ä &r:B      ä&rTB      ä &r6A      â x+U d  X    â $xMM   ÿ Ž â x   ÿÀŽ â &rZ9   Z   ä&r\R9   \   Ä &rRE   =   ä &r@      ä&rR@      ä $r   ÿŽ â xXU1u  X    â 0r;ÿ;    A   â xM0   âŽ â 0rNÿN    A   â x((   ÿàÿ â &rZ8   Z   Ä &r\+8   \   ä&r+C      ä &rL      ä&r+L      â Er        â &rZ67   Z   â r;N     A  â xàÿÿÿÿàÿ â &r7X7   \   â | (   pbô â &rAXA      Ä &r66K      ä&rXK      â #r.   .   â $rZZ   ÿŽ ä 0r<ÿ<    A   ä $r66   ÿŽ ä 0rÿP    A   Ä $r77   ÿŽ ä $rAA   ÿŽ ä$r   ÿŽ ä 0rÿO    A   â ErZ Z       â  r-;<     A  â Er6 6       â  r<     A  â  rN     A  â Er7 7       â  r;;     A  âErA A       â  r     A  â Er        Ä x€  ÿà÷ â #r -Z       â #r,;6   ,   â #r*7   *   â #r A       â #r      â x€   ÿàÿ â $r'ÿÿ   'Ž â x""€   ÿàÿ ä x&&€   ÿàÿ â D)       À â Gy  0ìÿÿÿÿƒ ê Ay        € ê� ˆ !  ýÿ H   è ˆ !*€ýÿ H   è ˆ !. þÿ H   è ˆ ! €þÿ H   è ˆ !, ÿÿ H   è ˆ !€ÿÿ H   è {          ì M        € ê „y     H   b. z # r  p`ð â $z x  #Ž Ä „y    H   " r ÿ   p"p â $z ~  Ž ä „y	    H   ¨ „y €   H   è „y    H   ( „y €  H   h „y    H   h „y €  H   h „y    H   h „y €  H   h „y €  H   b !r        Æ/ „y
    H   b !r       Æ „y €  H   " !r	       ÆO „y    H   ¢ !r*       Æ� „y €  H   â !r.       Æ„y €  H   â !r        Æ„y    H   â !r,       Æ „y €  H   â !r       Æ ‰      " !r       â !r       â !r       â !r

       â/ !r       â !r       ÈO ‰	      " !r       Æ� ‰      b !r       â!r       ä ‰      ¢ !r

       Æ ‰       â !r       Æ ‰
      ( ‰      b !r		       â !r       È/ ‰	      " !r       ÆO ‰      b !r         Æ� ‰      ¢ !r
       Æ‰  €    â !r       Æ‰
      ( ‰      b !r	       â !r       È/ ‰	 €    " !r       ÆO ‰ €    b !r        Æ� ‰ €    ¢ !r

       Æ‰  @    â !r       Æ‰
 €    ( ‰ €    b !r	       â !r       È/ ‰ @    " !r       ÆO ‰ @    b !r         â� x#   ÿàÿ â !r
       Æ‰	       ¢ z  r  p`ò â !r       ä‰ @    â x    p"ò Æ ‰ @    ( ‰ @    b !r       â 
 r  ÿàõ â !r       Æ/ ‰
      " ÿÿ   ÿä â !r 	       âO !r       ä� ‰      b !r       â
 j  ÿ„ ä ‰      ¢ !r       â$vÿ r  ÿ Ž Ä ‰      â 
 k   Æ ‰      "    ÿàÿ â!r
       â r   ÿ„ ä 	   ÿà÷ â $tÿ   ÿ Ž ä $r ÿÿ   ÿ â z j  ÿ„ â $ÿÿ   ÿŽ â 	 j  ÿ† â %v j   Ž â !r       â/ z k    Ä 		 k  � â !r       âO †	     â !r       Æ� †    â !r       Æ†	     è †     è †	     â M™        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      "   " ¹z  l      â Ey  à    € â ™x?    â y
      !   b ¹z  F   
   â $rÿÿ   ÿ Ž ä y      %   ¢ $rÿÿ   ÿ Ž ä $rÿÿ   ÿ Ž â y!      &   â $r	ÿÿ   ÿ Ž Æ y#      '   " r ÿ   pRð är ÿ   pRò â$x    
Ž ä/ $x   ÿ Ž ÆO xÿ     È r   ÿ� â $€   
Ž Æ xÿ     È |    pbô Ú G)  �    € ê $r"ÿÿ   ÿ Ž â�xøÿÿÿÿÀŽ â $r ÿÿ   ÿ Ž ä 'z # y  " Ž È $x   
Ž â zÿ z     â 'z! s    Ž Æ r(   ÿ� â $z  p  Ž â zÿ t    â $z |  Ž âx(ÿ   (  â $z |   Ž â x    ÿàÿ â $z v  Ž â x)ÿ   (  â $z# }  ÿŽ ä $z! w  ÿŽ Ä $z' v  Ž ä $z  p  Ž â r   ÿàõ â %x$'ˆ   ÿŽ È %xNˆ   ÿŽ È $x   ÿ Ž ä $x    ÿ Ž ä %x&'ˆ   (Ž â z q  ÿàÿ Æ %x(ˆ   (Ž È $vÿ X  ÿ Ž ä $vÿ Y  ÿ Ž ä $rÿÿ   ÿ Ž ä $rÿÿ   ÿ ä %x,    $Ž È %x*    NŽ È r,   ÿàõ Ê $x-    Ê �yA   ™ ¨  �yD   ™ è  �yC   ™ ( �y@   ™ b r*   ÿàõ Ê $x+    Ê �yE   ™ è �yH   ™ " xÿ     ä r:   àµ â �yB   ™ " $t9ÿ$   ÿ Ž ä r   ÿd~ â �yK   ™ " %v:: Z  9 Ž È $x$   ÿŽ È $x;;   Ž Ê �yJ:   ™ b r8   àµ Æ �y:   ™ b r   ÿd~ Æ �y:   ™ b %v88 Z  9 Ž Æ �y:   ™ b $x$   ÿŽ â r<$   ÿà÷ ä�y:   ™ b! r.&   ÿàõ â$x99   Ž ä �y:   ™ b $x=%   Ž ä$x/'    â �y8   ™ b r6N   ÿà÷ Æ �y<   • b r>(   ÿàõ Æ �y8   ™ h �y.   ‘ b $x7O   Ž Æ �y8   ™ b $x?)    Æ �y8   ™ h �y:   ™ h �y/6   • h �y>   ‘ h �y8   ™ h �y:    ™ h �y.8   ™ h �y0<    • h �y1:    • h �y28   ™ h �y38    ™ h �y56    • h �y48    • b ‚x �˜­¿     Ä ‚x &     â $~?ÿ   ÿ Ž ä $~Gÿ   ÿ Ž ä $~8ÿ   ÿ Ž ä/ $~=ÿ   ÿ Ž ä $~Fÿ   ÿ Ž ä $~Lÿ   ÿ Ž â x<AˆˆˆˆÿÀŽ ÄO x>Aww  ÿÀŽ ä xAÿ   A  ä x<<22ÿø� ä t:?ÏÝêö>    ät;G5EYq>    â $~>ÿ   ÿ Ž â xAAww  ÿÀŽ Ä x6<Tv  ÿÀŽ ä t98ÏÝêöA    är6:6   ;    â $~:ÿ   ÿ Ž â t8>5EYqA    ä x<ÿ   <  ä x;Dww  ÿÀŽ Ä� x7DˆˆˆˆÿÀŽ ä r99<   8    ä t8:ÏÝêö;    äx:722ÿø� ä xDÿ   D  ä t;>5EYq;    Ä x<:Tv  ÿÀŽ â $~?ÿ   ÿ Ž â xDDww  ÿÀŽ ä r88<   ;    ä x<CˆˆˆˆÿÀŽ â$~>ÿ   ÿ Ž â t==ÏÝêöD    ätD?5EYqD    Ä x:ÿ   :  ä x;Cww  ÿÀŽ ä xCÿ   C  ä x<<22ÿø� ä r==:   D    â $~Dÿ   ÿ Ž â t:>ÏÝêö;    Ä x76 d  9    ät;F5EYq;    ä xCCww  ÿÀŽ ä x><Tv  ÿÀŽ ä x661u  9    ä x98 d  =    Ä x881u  =    ä t=DÏÝêöC    är;:>   ;    ä tDF5EYqC    ä x>ÿ   <  â $~?ÿ   ÿ Ž â x<@ww  ÿÀŽ Är>=>   D    â $~Aÿ   ÿ Ž â x=@ˆˆˆˆÿÀŽ ä x@ÿ   @  ä x==22ÿø� ä t??ÏÝêö<    â$~Dÿ   ÿ Ž â t<A5EYq<    Ä xA@ww  ÿÀŽ ä x@=Tv  ÿÀŽ ä x:; d  >    äx;;1u  >    ä t>DÏÝêöA    â $~Dÿ   ÿ Ž â r??@   <    Ä tAF5EYqA    ä x<ÿ   =  ä x=Eww  ÿÀŽ ä x@ÿ   E  ä r<><   A    ä t>D5EYq=    â $~Dÿ   ÿ Ž â xP@ww  ÿÀŽ â $~@ÿ   ÿ Ž â xAHww  ÿÀŽ Ä xEEˆˆˆˆÿÀŽ ä tDDÏÝêöA    ätA@5EYqA    ä x@ÿ   H  â $~Cÿ   ÿ Ž â tLLÏÝêö=    â $~=ÿ   ÿ Ž â xHHˆˆˆˆÿÀŽ Ä xEE22ÿø� ä x@@ww  ÿÀŽ ä xHH22ÿø� ä tFCÏÝêöP    äxIETv  ÿÀŽ ä t==5EYqP    Ä tCCÏÝêö@    äxEÿ   E  ä t@G5EYq@    ä xGHTv  ÿÀŽ ä r>LI   >    ä rEFE   =    Ä xHÿ   H  ä rDDG   A    ä x=? d  <    äx<?1u  <    ä xGBˆˆˆˆÿÀŽ â $~Iÿ   ÿ Ž â x?> d  E    Ä rCCH   @    ä x>>1u  E    â $~Eÿ   ÿ Ž â xFBww  ÿÀŽ ä xGG22ÿø� ä xAD d  C    äx@D1u  C    Ä xDÿ   B  ä tLIÏÝêöF    ätBE5EYqF    ä xIGTv  ÿÀŽ È rBLI   B    â &rI7J   ÿ   âxDDww  ÿÀŽ â $~Hÿ   ÿ Ž ä &rI6   I   ä$~Cÿ   ÿ Ž ä &rJ?J   ÿ   ä &rI9   I   â tHHÏÝêöD    â&rJ>   J   â tCC5EYqD    Ä xGÿ   G  â &rI8   I   ä &rJA   J   â rCHG   C    â &r:   I   ä &rJ@   J   â xB d  C    äxBB1u  C    â &rC;      Ä $x   ÿ Ž â xEKww  ÿÀŽ â $~Dÿ   ÿ Ž ä $~Fÿ   ÿ Ž ä &rJ   J   â x    ÿÀŽ â &r77   ÿ   âxHKˆˆˆˆÿÀŽ â &r?   ÿ   â xKÿ   K  Ä rÿ     â &r76   7   â tDDÏÝêöE    â &r>      â tEF5EYqE    â $~Fÿ   ÿ Ž â rÿ     â $~Gÿ   ÿ Ž â xHH22ÿø� â $x   ÿ Ž â xKKww  ÿÀŽ â &r79   7   âx   ÿÀŽ â &rA      â xLHTv  ÿÀŽ Ä tFFÏÝêöK    ätGG5EYqK    ä xHÿ   H  â &r78   7   âx0   âŽ â &r@      â r/ÿ   /  â &rC=   C   â rDDL   E    Ä rGFH   G    â &r7:   7   ârÿ     â &r      â xàÿÿÿÿàÿ â &rC<   C   â xD d  G    â $x/   ÿ Ž â x   ÿÀŽ â &rB   J   ä &r7;.   7   Ä &rB.      â xDD1u  G    â $rCC   ÿŽ â x0   âŽ â 0r0ÿ0    A   â x   ÿàÿ â 0r1ÿ1    A   ä &r      ä &r7=2   7   Ä &r2      â ErC C       â &rD      â  r01     A  â xàÿÿÿÿàÿ â &r7<3   7   â|    pbô â &rDD3      â #r	C   	   â 0rÿ5    A   ä $r77   ÿŽ Ä $r   ÿŽ ä $rDD   ÿŽ ä 0rÿ4    A   â  r1     A  â Er7 7       ä Er        â  r00     A  âErD D       â  r     A  â x€  ÿà÷ â #r07      â #r      â #rD      â x€   ÿàÿ â $rÿÿ   Ž â x€   ÿàÿ â D)       À â Gy  �îÿÿÿÿƒ ê Ay        € ê�ˆ 	 þÿ H   è ˆ €þÿ H   è ˆ  ÿÿ H   è ˆ €ÿÿ H   è {          ì M        € ê „y 
     H   " z  r  p`ð â $z!! x  Ž â x   ÿàÿ â „y
 €   H   b r 
ÿ   p"p â$z!# ~  !Ž â z  r  p`ò â „y
    H   ¦ x 
   p"ò â „y
    H   è „y
 €  H   ( „y
    H   h „y
    H   h „y
 €  H   h „y
    H   h „y
 €  H   h „y
 €  H   b !r  	       Æ „y
 €  H   " !r       â/ !r       âO !r         ä� !r       â!r       â!r         ä !r       Æ ‰       b !r       Æ ‰      ¢ !r       Æ ‰      â !r       È !r       Ê ‰      " !r        â/ !r       ÈO ‰      b !r       Æ� ‰      ¨ ‰	      â !r       Ê ‰      " !r       â/ !r       ÈO ‰ €    b !r	       Æ� ‰	 €    ¨ ‰  €    â !r       Ê ‰ €    " !r       â/ $tÿ   ÿ Ž â !r		       âO !r        È� ‰	 @    h ‰  @    ¨ ‰ @    â !r       Ê ‰ @    " !r	       â/ !r         ÈO ‰      b 
! r  ÿàõ â !r       ä� ‰       ¤ $ÿÿ   ÿ â 
 j  ÿ„ â ‰	      â !r       ä 
 k   Æ ‰      " !r       â/ !r        âO %v! j   Ž â !r		       È� †	     è †    è †	 	    â M™        € ê ¹z  r      â!r       â �x   ?àÿ Ì |!!   ÿàñ Ê $r ÿÿ   ÿ  â z! j  ÿ€ È z! k     Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y       &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž ä $r	ÿÿ     Ž È %	 \   Ž Ê �	    ™ b ¹z  l      â Ey  Ð    € â ™x?    â y      "   b $rÿÿ   ÿ Ž Æ y
      !   ¢ $ÿÿ   ÿ Ž ä $ÿÿ   	 Ž È '
	 m   Ž È $‚ÿÿ   ÿ Ž â 
ÿ n    Ê $ÿÿ   
Ž â r ÿ   pRò Æ/ $
	 o  	Ž â r ÿ   pRô â$x    
Ž äO $‚ÿÿ   	 Ž Æ xÿ     â 'Š	 s   Ž ä y      '   " r   ÿ� È xÿ     È |    pbö ä Šÿ t    Ö G9  P
    € ê y      %   " $rÿÿ   ÿ Ž â xøÿÿÿÿÀŽ â $z	 w  ÿŽ ä 'z y   Ž È $x   
Ž â zÿ z    â $z }  ÿŽ ä $x   ÿ Ž âr   ÿ� â $z |  Ž ä $rÿÿ   ÿ Ž ä $z v  Ž Ä$x   ÿ Ž ä $x   ÿ Ž ä $z p  Ž â xÿ     Æ %xˆ   ÿŽ â x	ÿ     È z X  àë â%x   Ž â z X  ÿàù ä z	 Y  	Äþ ä z X  ÿà÷ ä r   ÿàû ä x   ÿàñ Ä z Y  ÿä â $rÿÿ   ÿŽ â zÿ Y  d~  ä x   ÿàñ ä xÿ     ä x   ÿÀŽ â $rÿÿ   	  ä �y&üÿÿ ™ ¢  r   à± â $tÿ$   ÿ Ž Ä �y$    ™ â  r   ÿd~  â %v Z   Ž ä �y   ™ $ $x	$   ÿŽ ä �y   ™ d $x   	Ž Ê �y   ™ è �y   ™ è �y   ™ ( �y   ™ " $rÿÿ    Ž Ä $r	ÿÿ    Ž â �y   ™ ( �y     ‘ ( �y#   ™ ( �y"   ™ " $rÿÿ    Ž Ä $r	ÿÿ    Ž â �y%    • ( �y!   • ( �y    ™ ( �y    • b ‚x �˜­¿     Ä ‚x &     â $~'ÿ   ÿ Ž ä $~,ÿ   ÿ Ž ä $~(ÿ   ÿ Ž â x   ÿàÿ È |    pbð ä x€  ÿàù ä x€  ÿà÷ ä x€   ÿàÿ â $rÿÿ    ä $rÿÿ   Ž â x*ÿ   &  ÄO x+&ˆˆˆˆÿÀŽ äx)&ww  ÿÀŽ â $~&ÿ   ÿ Ž â x**ww  ÿÀŽ ä x+22ÿø� ä/ t''ÏÝêö*    ät*&5EYq*    Ä x&Tv  ÿÀŽ ä xÿ     ä t,ÏÝêö)    ät((5EYq)    ä r'   *    ä r&   (    Ä x*ÿ   $  ä� x&$ˆˆˆˆÿÀŽ â$~+ÿ   ÿ Ž â x,$ww  ÿÀŽ â $~(ÿ   ÿ Ž â x**ww  ÿÀŽ â $~)ÿ   ÿ Ž â x&&22ÿø� â $~'ÿ   ÿ Ž â t$+ÏÝêö,    Ä t((ÏÝêö*    ät))5EYq*    ä x*&Tv  ÿÀŽ ä x+ÿ   &  ä t''5EYq,    ä x& d      Ä x1u      ä r$$*   '    â &r&   ÿ   â r)(+   )    Æ &r      â x$ d  )    äx$1u  )    Æ &r      âxÿ     â $~$ÿ   ÿ Ž â xww  ÿÀŽ â$~'ÿ   ÿ Ž â xˆˆˆˆÿÀŽ â $~)ÿ   ÿ Ž â xww  ÿÀŽ â $~&ÿ   ÿ Ž â x22ÿø� ä t$ÏÝêö    Ä t$'ÏÝêö    ät')5EYq    ä xTv  ÿÀŽ ä t&5EYq    ä xÿ     â &r      â r       Ä r$   '    â $~$ÿ   ÿ Ž â xÿ     äxww  ÿÀŽ äxˆˆˆˆÿÀŽ â $~(ÿ   ÿ Ž â xww  ÿÀŽ ä t$ÏÝêö    Ä t&5EYq    ä x22ÿø� ä t$(ÏÝêö    äx d      ä t'&5EYq    ä xTv  ÿÀŽ Ä xÿ     â &r      â x1u      ä r!ÿ   !  ä r       ä r$$   '    ä r ÿ      â &r#      â x d  $    â $x!!   ÿ Ž â x     ÿÀŽ Ä x$1u  $    â &r"      â x !0    âŽ Æ &r$      â x àÿÿÿÿàÿ â 0rÿ    A   ä 0r%ÿ%    A   ä $r   ÿŽ Æ  r%%     A  â x€  ÿàû ä Er        Æ $rÿÿ   	Ž ä #r%      â G‰  À÷ÿÿÿÿƒ ê Ay        € ê ˆ €ÿÿ H   è {          ì M)        € ê „y
     H   "r 
ÿ   pRð Æ „y
 €   H   h „y
    H   ¨ y      %   â !r       È !r       È/ !r       ÊO ‰      $ !r       Ê ‰      $ !r       Ê ‰ €    $ !r       Ê ‰ @    $ !r	       â $z  x  Ž È� ‰	      b  M	        € ê $tÿ   ÿ Ž â !r		       â? $z ~  Ž È %v j   Ž Ê †y 	    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž È % \   Ž Ê �		    ™ ¢  $rÿÿ   ÿ Ž â Ey  Ð    € â $rÿÿ   ÿ Ž â y      "   b 'Š s   Ž Æ y      !   b '
  m   Ž Æ y
      %   â $rÿÿ   ÿ Ž â 
 ÿ n     ä y      '   " rÿ   ÿü‚ â/ $x    Ž Ê xÿ     È r   ÿ� ä Š	ÿ t    È r	       âO G  À    € ê z ÿ `  pPô â�$rÿÿ   ÿ Ž â z ÿ ^  pPò â $rÿÿ   ÿ Ž â z ÿ b  p ö ä z ÿ a   Sô ä z ÿ _  Sò ä z ÿ c  0't Ö $ ~  ÿŽ âÿ   
  â$ª ~  ÿŽ ä $ x  ÿŽ â    
àë â$ª x  ÿŽ â ¢   
àÇ ä¨ÿ   
  ä ÿÿ   Äþ Ä    ÿàý ä ¢   ÿàû ä ¢ÿÿ   „þ â $ÿÿ    â  ^  ÿ† ä ª b  ÿˆ â $¢ÿÿ   Ž â  _  � È ª c   â �     h �©     d Ay        € ê�¹z  l      â xÿ     â ™x?    â $ÿÿ    
Ž â Ey  0    € â $rÿÿ   ÿ Ž Ä $
  o  Ž â|    pbò â $‚ ÿÿ    Ž ä $rÿÿ   ÿ Ž Ô G  À    € ê $rÿÿ   ÿ Ž ä $z }  ÿŽ ä'z y   Ž È $z   w  ÿŽ â zÿ z    â $rÿÿ   ÿ Ž ä $rÿÿ   ÿ Ž â r    ÿàñ â $z |  Ž â xøÿÿÿÿÀŽ Æ $z	 v  Ž ä $x   
Ž ä $z
 p  Ž ä $x
   ÿ Ž â r   ÿ� â %xˆ   ÿŽ â xÿ     ä xÿ     â $rÿÿ    Ž Ä $r	ÿÿ   ÿ  â z X  ÿàó ä xÿ     È z  Y  ÿäÿ  ä $x   ÿ Ž ä $t ÿ   ÿ Ž È %v X   Ž Ê r   ÿàñ Ê $x     Ê �y'   ™ ¨  �y#   ™ â  r   
à‘ Æ �y"   ™ " r	   ÿ$~  â $tÿ$   ÿ Ž ä �y!   ™ ä  %v Z   Ž È $x$   ÿŽ È $x   Ž â r   ÿàñ È �y   ™ " z X  ÿàó Æ �y   ™ " z Y  ÿäÿ  â$x      ä �y   ™ ( �y   ™ ( �y$   • ( �y   ‘ ( �y   ™ ( �y   ™ ( �y   ™ ( �y    ™ ( �y%    • ( �y&    • " ‚x �˜­¿     Ä ‚x &     â $~)ÿ   ÿ Ž ä $~+ÿ   ÿ Ž ä $~-ÿ   ÿ Ž ä $~*ÿ   ÿ Ž â z ÿ `  pPð È z ÿ a   Sð ä x   ÿàÿ ä x(ÿ   '  äO x'ww  ÿÀŽ äx''ˆˆˆˆÿÀŽ ä x((ww  ÿÀŽ Ä x''22ÿø� ä t)ÏÝêö    ä/x'Tv  ÿÀŽ ä t+5EYq    ä t-ÏÝêö(    ät*5EYq(    Ä x'ÿ   '  ä r       ä r'       ä xÿ   #  ä� x#ww  ÿÀŽ äx##ˆˆˆˆÿÀŽ â $~+ÿ   ÿ Ž â x)ww  ÿÀŽ â $~'ÿ   ÿ Ž â x(#22ÿø� â $~-ÿ   ÿ Ž Ä $~ÿ   ÿ Ž â t+ÏÝêö    ät*'5EYq)    ät-5EYq    ä x'(Tv  ÿÀŽ ä t#ÏÝêö)    ä x(ÿ   (  Ä r'       ä r#(   *    ä x*ÿ   "  äx)"ˆˆˆˆÿÀŽ â$~#ÿ   ÿ Ž â x d      â $~'ÿ   ÿ Ž â x("ww  ÿÀŽ Ä x1u      â $~ÿ   ÿ Ž â x**ww  ÿÀŽ ä x))22ÿø� ä t"+ÏÝêö(    ä t+#5EYq*    ä t''5EYq(    Ä x#)Tv  ÿÀŽ ä t(ÏÝêö*    ä x)ÿ   )  ä r#"#   '    ä r(()   +    ä x d      Ä x1u      ä x# d  (    äx##1u  (    ä x'!ˆˆˆˆÿÀŽ â$~)ÿ   ÿ Ž â x(ÿ   !  â &r   ÿ   â x"!ww  ÿÀŽ â $~+ÿ   ÿ Ž â x''22ÿø� â $~!ÿ   ÿ Ž â x,(ww  ÿÀŽ â &r*      â t))ÏÝêö"    Ä t"+5EYq"    ä x('Tv  ÿÀŽ â &r   *   â t!ÏÝêö,    â $x!   ÿ Ž â r))(   "    â &r      â t+5EYq,    ä x"   ÿÀŽ Ä x'ÿ   '  ä r$ÿ!   $  â &r      â rÿ"     ä r'       â $x$$   ÿ Ž â x   ÿÀŽ â &r#      â x) d      Ä x))1u      ä x$0   âŽ â &r      Æ xàÿÿÿÿàÿ â &r)      ä 0r$ÿ%    A   ä $r#   ÿŽ ä 0r%ÿ&    A   â G‰  °    € ê %v `   Ž Ê r   ÿàñ Ê $x     Ê �y)    ¨  �y     è  �y    " z `  à‘ Æ �y&    â  z `  ÿàõ ä z a  $~  äz a  ÿä È �y    è �y(    è  �y'     " $~*ÿ   ÿ Ž ä $~ÿ   ÿ Ž Ä/ $~+ÿ   ÿ Ž ä $~ÿ   ÿ Ž â x-ÿ   )  äO x)ˆˆˆˆÿÀŽ äx,)ww  ÿÀŽ ä x)-ww  ÿÀŽ â $~-ÿ   ÿ Ž â x22ÿø� Ä t**ÏÝêö)    äxTv  ÿÀŽ ä tÏÝêö,    ät++5EYq,    ä t)5EYq)    ä xÿ     Ä r   +    ä r*   )    â $~)ÿ   ÿ Ž â x+ÿ      ä� x d      äx1u      ä x ˆˆˆˆÿÀŽ â &r   ÿ   â x++ww  ÿÀŽ â $~ÿ   ÿ Ž â x22ÿø� â &r      â x* ww  ÿÀŽ â $~ ÿ   ÿ Ž â tÏÝêö+    Ä t+5EYq+    ä xTv  ÿÀŽ ä xÿ     ä t-ÏÝêö*    ät))5EYq*    ä r   +    â $~ÿ   ÿ Ž â xÿ     Äxww  ÿÀŽ är   )    â $~ÿ   ÿ Ž â xˆˆˆˆÿÀŽ ä x)ww  ÿÀŽ ä x22ÿø� ä tÏÝêö    Ä tÏÝêö)    ä t 5EYq    ä x d      ä t) 5EYq)    ä x Tv  ÿÀŽ â &r      â xÿ     â $~ÿ   ÿ Ž â x1u      Ä r        â $~ÿ   ÿ Ž â r   )    â &r      â xÿ   &  â $~ÿ   ÿ Ž â x d      äx1u      â $~ÿ   ÿ Ž â x&ˆˆˆˆÿÀŽ â &r      â x&&ww  ÿÀŽ Ä x22ÿø� â &r      â xww  ÿÀŽ ä tÏÝêö&    ät5EYq&    ä xTv  ÿÀŽ ä tÏÝêö    Ä t5EYq    ä xÿ     ä r(ÿ!   (  ä rÿ"     ä r       â $x(   ÿ Ž â r       Ä x   ÿÀŽ ä x d      äx1u      ä x0   âŽ â &r      Æ xàÿÿÿÿàÿ â &r      ä 0rÿ'    A   ä $r   ÿŽ Æ  r%     A  ä Er        Ê #r      â |    pbð â  r$$%     A  â Er# #       ä x€  ÿàó ä x

€   ÿàÿ â #r$#      ä $rÿÿ   Ž  È G‰  °ðÿÿÿÿƒ ê Ay        € ê y       "   " Ey        € â r  ÿ   pRð är  ÿ   pRò Ö G‰  `     € ê y      !   " z ÿ `  pPð È z ÿ a   Sð â $x     Ž Ê ˆs €ÿÿ H   î ˆ    H   ä Ay        € ê {          ì M        € ê y      !   b z ÿ `  pPð È z ÿ a   Sð â „y      H   X. „‰ €   H   ¨ „‰    H   " !r        È/ !‚       ÈO !‚       â G‰        € ê „y  €  H   ( „y €   H   h „y    H   ¨ „y    H   è „y €  H   " !r         â !r       Æ/ !r         âO !r       Æ� !r        ä‰      " z ÿ `  pPð ä r ÿ   pRò ä z ÿ a   Sð â !r       Ê ‰       $ !r         Ê ‰  €    $ !r        Ê ‰ @    $ !r       Ê ‰      $ !r       â G‰        € ê ‰       $ !r         Ê ‰       $ !r        Ê ‰ €    $ !r       Ê ‰ @    $ !r       Ê ‰      $ !r       ä M        € ê !r       âG‰  @    € ê $v ÿ h  ÿ Ž â !r       È x     pBð Ú G	  P    € ê x     pRð Ú G‰  �     € ê x     pRð Ú G	  ð    € ê  x ;ª¸¿  A  Ì s          $ !t   €?    Ì s        $  r     A  È  r     A  â Gy  P    € ê  x '7=  A  â x*BL?  A  Æ #t   €?    È  r      A  â  x    ?  A  Ê s     $   $ !t  €?    È  r       A  È  r      A  â Gy  °    € ê x     pRð Ú G‰  à     € ê x     pRð Ú G	        € ê 	z  i    � ä 	z i    � Æ  x ;ª¸¿  A  â 	z i €  � Ê s        $ !t  €?    Ì s        $  r       A  È  r      A  â Gy  Ð     € ê  r     A  â Gy  °     € ê 	x   à@  � ä 	x  à@  � Æ  x #ÛÙ?  A  â 	x  àÀ  � Æ  x;ª¸¿  A  ä !t  €?    È s        $ !t  €?    Ì s        $  r       A  È  r      A  ä y       %   ( y      &   ( y      '   b $z  x   Ž ä $tÿ   ÿ Ž Ä $z  ~   Ž È/ %v  j   Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž ä $rÿÿ    Ž È % \   Ž Ê �	    ™ b ¹z  l      â Ey  ð    € â ™x?    â y      "   b $rÿÿ   ÿ Ž ä $rÿÿ   ÿ Ž â y
      !   ¢ $rÿÿ   ÿ Ž ä $ÿÿ   ÿ Ž â y      %   " $ÿÿ    Ž Ä $r	ÿÿ   ÿ Ž â y%      '   â '
  m   Ž È $‚ÿÿ   ÿ Ž â 
 ÿ n     Ê $ ÿÿ    
Ž â r ÿ   pRò Æ/ $
  o  Ž ä $x    
Ž äO $‚ÿÿ    Ž Æ xÿ     â 'Š s   Ž Æ r    ÿ� â $€   
Ž ä $x   ÿ Ž â xÿ      È |    pbô ä Šÿ t    ä r ÿ   pRð Ò G)      € ê $r$ÿÿ   ÿ Ž â� x øÿÿÿÿÀŽ â $z  p  Ž ä 'z% y  $ Ž È $x   
Ž â zÿ z    â $z w  ÿŽ ä $z% }  ÿŽ ä $z |   Ž ä $vNÿ p  ÿ Ž â r   ÿà÷ â $z& v  Ž âr   ÿ� â$x)   ÿ Ž Ä %x&&ˆ   ÿŽ â xÿ     ä x)   ÿÀŽ â %xNNˆ   ÿŽ â r&   ÿàõ äxÿ     â $vÿ X  ÿ Ž â xN   ÿüŽ â $vÿ Y  ÿ Ž â r'   ÿ â $rÿÿ   ÿ Ž Ä $x    ÿ Ž ä $x   ÿ Ž ä $rÿÿ   ÿŽ ä %x()   &Ž Ê r(   ÿàõ Ê $x)    Ê �yF   ™ ¨  �yE   ™ è  �yH   ™ ( �yJ   ™ b r,N   ÿàõ Ê $x-O    â rL    àµ â $tMÿ$   ÿ Ž Æ �yD,   ™ â r   ÿd~ â %vLL Z  M Ž ä �y,   ™ ä $x$   ÿŽ ä �y,   ™ ä $xMM   Ž Ä �y ,   ™ ( �yL   ™ ( �yL   ™ ( �yL   ™ ( �yL   ™ ( �yL   ™ ( �yL   ™ " r2N,   ÿàõ Æ �yL   ™ $ $x3O   - ä �yL    ™ h �y!2   ™ h �y"2   ™ b r0   ÿàõ Ä r4&   ÿà÷ â �y#2   ™ d $x1    ä $x5'   Ž â �y,2   ™ h! �y+0   ‘ h �y*4   • b r@N2   ÿàõ Ä x:N   4àÇ â�y.4    • d $xAO   3 â ‚x &     â �y2L    • h �y-@   ™ h �y4@   ™ h# �y7@   ™ b r8N0   ÿàõ Ä r;Oÿ   5„þ Æ $x9O   1 âr<N8   ÿàõ â �y0:    • b rB:   ÿà÷ ä r>N<   ÿàù â$x=O   9 â�y/8   ‘ b $xCO   ;Ž ä$x?O   = â �y1:þÿÿ • h �y3<   ‘ b x9Nþÿÿÿ:àµ Æ �y8@   ™ b x;Oÿÿÿÿ;d~ â/ ‚x �˜­¿     ä �y6Bþÿÿ • b $~=ÿ   ÿ Ž â r<9   ÿàõ â $~Gÿ   ÿ Ž â �y9>   ‘ h �y5Büÿÿ • d $~Cÿ   ÿ Ž Ä/ $~Lÿ   ÿ Ž ä $~Iÿ   ÿ Ž â x:Fww  ÿÀŽ ÈO tA=5EYq:    â$x=O   ; â t@GÏÝêö:    È �y:<    • ¨  �y;<þÿÿ • ¢  xBFˆˆˆˆÿÀŽ ä xFÿ   F  ä xBB22ÿø� â $~Gÿ   ÿ Ž â xFFww  ÿÀŽ Ä x?BTv  ÿÀŽ ä t>CÏÝêöF    är@@?   A    â $~?ÿ   ÿ Ž â tCG5EYqF    ä xAÿ   B  ä xFEww  ÿÀŽ Ä� xBEˆˆˆˆÿÀŽ ä xEÿ   E  ä rA>A   C    ä t??ÏÝêöF    ät>G5EYqF    â $~Fÿ   ÿ Ž â xBB22ÿø� Ä xEEww  ÿÀŽ ä xCBTv  ÿÀŽ ä x<@ d  A    äx=@1u  A    ä t@FÏÝêöE    ätEL5EYqE    Ä xBÿ   B  ä r??C   >    â $~>ÿ   ÿ Ž â xAHww  ÿÀŽ â$~Fÿ   ÿ Ž â r@@B   E    ä xBÿ   H  ä tK>ÏÝêöA    Ä tAF5EYqA    â $~Fÿ   ÿ Ž â xEBww  ÿÀŽ â $~Bÿ   ÿ Ž â xCJww  ÿÀŽ â$~Lÿ   ÿ Ž ä $~>ÿ   ÿ Ž â tFFÏÝêöC    ätCB5EYqC    Ä xBÿ   J  ä tGLÏÝêöE    ät>>5EYqE    â $~Eÿ   ÿ Ž â xJJˆˆˆˆÿÀŽ ä xHHˆˆˆˆÿÀŽ ä xBBww  ÿÀŽ Ä xJJ22ÿø� ä xHH22ÿø� ä tEEÏÝêöB    ätBI5EYqB    ä xIJTv  ÿÀŽ ä xJÿ   J  Ä xLHTv  ÿÀŽ ä xHÿ   H  ä rFFI   C    ä rEEJ   B    ä rAKL   A    ä rHGH   >    â $~Kÿ   ÿ Ž â x>? d  @    Ä xCF d  E    äxBF1u  E    â $~Eÿ   ÿ Ž â x??1u  @    ä x@A d  H    äxJDww  ÿÀŽ ä xAA1u  H    Ä xIDˆˆˆˆÿÀŽ ä xHÿ   D  â $~Fÿ   ÿ Ž â tDKÏÝêöJ    â$~Gÿ   ÿ Ž â tEE5EYqJ    ä xJI22ÿø� ä xHHww  ÿÀŽ Ä xIJTv  ÿÀŽ ä tFFÏÝêöH    ätGG5EYqH    ä xJÿ   J  È rGFJ   G    â &rF<   ÿ   â rDDI   E    ä xIÿ     â &r==   F   â xKww  ÿÀŽ äxˆˆˆˆÿÀŽ â $~Hÿ   ÿ Ž â x<Iww  ÿÀŽ â $~Eÿ   ÿ Ž Ä &r=>   =   â x22ÿø� â $~Lÿ   ÿ Ž ä $~Iÿ   ÿ Ž â tHHÏÝêöK    â&r=?   =   â tEEÏÝêö<    ätKL5EYqK    ä t<I5EYq<    Ä xFTv  ÿÀŽ ä x>ÿ     â &r@@   =   â xD d  G    ä rFHF   K    ä rEE>   <    â &r<   ÿ   â xGD1u  G    â &rAA   @   â x@ÿ     Ä x>ˆˆˆˆÿÀŽ â$~Hÿ   ÿ Ž â xF d  E    äx=F1u  E    â $~Fÿ   ÿ Ž â x?ww  ÿÀŽ â &r<G   <   â xD@ww  ÿÀŽ â $~Eÿ   ÿ Ž â x>>22ÿø� â $~Gÿ   ÿ Ž â tFÏÝêö?    Ä t@H5EYq?    ä tEEÏÝêöD    äx?>Tv  ÿÀŽ ä tDG5EYqD    ä x>ÿ   >  â &r   <   â r?   @    Ä r>E>   D    â &r=      Æ x< d  >    â &rAC   A   â xD ww  ÿÀŽ â $~=ÿ   ÿ Ž ä &r<      â x<ÿ   !  â &rBB   A   â x@!ww  ÿÀŽ â $~Aÿ   ÿ Ž ä $~?ÿ   ÿ Ž â x<<ww  ÿÀŽ â $~Cÿ   ÿ Ž Ä $~Eÿ   ÿ Ž â t==ÏÝêöD    ä tAAÏÝêö@    ä x1u  >    ä tD?5EYqD    ät@?5EYq@    ä t?CÏÝêö<    Ä t>E5EYq<    ä x<ÿ      ä x  ˆˆˆˆÿÀŽ ä x!!ˆˆˆˆÿÀŽ ä x<<ww  ÿÀŽ ä x  22ÿø� Ä xE!22ÿø� ä t!CÏÝêö<    ä xC Tv  ÿÀŽ ä xFETv  ÿÀŽ ä xEÿ   E  ä r==C   D    Ä xDÿ   "  ä r??E   >    â $~Eÿ   ÿ Ž â x>"ˆˆˆˆÿÀŽ â $~Cÿ   ÿ Ž â r@AF   @    â $~Hÿ   ÿ Ž â xFDww  ÿÀŽ â $~Jÿ   ÿ Ž â xA"ww  ÿÀŽ Ä xD>22ÿø� ä t>EÏÝêöF    ätEC5EYqF    ä t<G5EYq<    ä t"HÏÝêöA    ä x ÿ      Ä xCDTv  ÿÀŽ ä xFÿ   D  ä tAJ5EYqA    ä xD@ d  ?    är<!    <    ä x?@1u  ?    â &rDD   ÿ   â r""C   A    Ä rE>F   E    â &r      â x = d  <    â &rD?   D   â x>" d  E    äx=1u  <    ä x!"1u  E    â $~=ÿ   ÿ Ž â x<ÿ   #  Ä x"#ww  ÿÀŽ â&r       â x##ˆˆˆˆÿÀŽ â $~?ÿ   ÿ Ž ä &r >   D   â x><ww  ÿÀŽ â $~Aÿ   ÿ Ž ä $~Cÿ   ÿ Ž â x##22ÿø� Ä t<=ÏÝêö"    ät=?5EYq"    ä t?AÏÝêö>    ät@C5EYq>    ä x>#Tv  ÿÀŽ ä x#ÿ   #  Ä r+ÿ   +  ä r*ÿ   *  ä r<<>   =    ä r#?#   @    ä x"+   ÿÀŽ â $x+*   ÿ Ž ä &r      â x<1u  #    â&r !       â x!< d  #    Ä x#ÿ   ,  ä x"+0   "âŽ â $~+ÿ   ÿ Ž â x*,ww  ÿÀŽ â $~=ÿ   ÿ Ž â x<#ww  ÿÀŽ â &r !       â t!+ÏÝêö*    â$~?ÿ   ÿ Ž â t#+ÏÝêö<    Ä t*=5EYq*    ä x>-ww  ÿÀŽ ä t<=5EYq<    ä x=ÿ   -  â $~Aÿ   ÿ Ž â x--ˆˆˆˆÿÀŽ â $~@ÿ   ÿ Ž â t+?ÏÝêö>    â $~Dÿ   ÿ Ž â x?=ww  ÿÀŽ Ä x,,ˆˆˆˆÿÀŽ ä x=-22ÿø� ä t>A5EYq>    ä t@@ÏÝêö?    ätAD5EYq?    ä x,,22ÿø� Ä x?=Tv  ÿÀŽ ä x=ÿ   =  ä x-"àÿÿÿÿàÿ ä x",Tv  ÿÀŽ ä r++?   >    ä r@@=   A    Ä x,ÿ   ,  ä r"!"   *    ä x*+ d  @    ä r##,   <    â &r!       ä &r**   ÿ   â x" d  #    Ä x "1u  #    ä x"ÿ   4  ä x4ww  ÿÀŽ â $rBB-   ÿŽ â x44ˆˆˆˆÿÀŽ â $~<ÿ   ÿ Ž â x,"ww  ÿÀŽ â $~-ÿ   ÿ Ž ä $~=ÿ   ÿ Ž â x4422ÿø� â $~>ÿ   ÿ Ž â t"<ÏÝêö    Ä t--ÏÝêö,    ät<=5EYq,    ä x,4Tv  ÿÀŽ ä x++1u  @    ä t#>5EYq    ä x4ÿ   4  â &r+   *   â r"",   #    Ä r--4   <    â &r   !   â x#7ww  ÿÀŽ äx*ÿ   7  â $~,ÿ   ÿ Ž â x77ˆˆˆˆÿÀŽ â $~+ÿ   ÿ Ž â x!" d  -    äx"1u  -    â $~-ÿ   ÿ Ž â x**ww  ÿÀŽ â $~4ÿ   ÿ Ž â x7722ÿø� Ä t",ÏÝêö#    ät++ÏÝêö*    ät,-5EYq*    ä x*7Tv  ÿÀŽ ä t#45EYq#    ä x7ÿ   7  â &r!!      â r""*   #    Ä r++7   ,    â &r   !   Æ x" d  +    â &r       â x8ww  ÿÀŽ Æ &r      â xÿ   8  â $~ ÿ   ÿ Ž â x88ˆˆˆˆÿÀŽ ä x+"1u  +    â $~"ÿ   ÿ Ž â xww  ÿÀŽ â $~*ÿ   ÿ Ž â x8822ÿø� ä t ÏÝêö    Ä t  ÏÝêö    ät!"5EYq    ä x8Tv  ÿÀŽ ä t*5EYq    ä x8ÿ   8  ä r0ÿ   0  Ä r6ÿ   6  är:ÿ   :  äO r/ÿ   /  är3ÿ   3  ä r       ä r! 8   !    Ä r9ÿ   9  â $x00   ÿ Ž â x//   ÿÀŽ â $x66   ÿ Ž â x33   ÿÀŽ â $x::   ÿ Ž â x d  !    â &r++      â x99   ÿÀŽ ä x00   /âŽ â &r   +   â x360   3âŽ Ä x1u  !    ä x9:0   9âŽ ä x   ÿàÿ ä xàÿÿÿÿàÿ â &r      â x3àÿÿÿÿàÿ ä x9àÿÿÿÿàÿ Ä |    pbô â $r   ÿŽ ä $r   ÿŽ ä $r   ÿŽ ä 0r.ÿ.    A   ä 0rÿ2    A   ä 0rÿ1    A   Ä 0rÿ5    A   ä 0rÿ;    A   â x€  ÿà÷ â  r..     A  â ErB B       â  r     A  âEr        â  r     A  âEr        â  r     A  â Er        â #r	.B   	   â #r      â #r      â x  €   ÿàÿ â #r      â $rÿÿ   Ž â D)       À â Gy  Àãÿÿÿÿƒ ê Ay        € ê� ˆ 	 þÿ H   è ˆ €þÿ H   è ˆ  ÿÿ H   è ˆ €ÿÿ H   è {          ì M	        € ê „y 
     H   " z  r  p`ô Æ „y
 €   H   b r 
ÿ   pVt Æ „y
    H   ¨ „y
    H   è „y
 €  H   ( „y
    H   ( „y
 €  H   ( „y
    H   ( „y
 €  H   ( „y
    H   ( „y
 €  H   " !r  	       Æ „y
 €  H   " !r       â/ !r       âO !r         ä� !r       â!r       â !r       â !r         â !r       È ‰       b !r       Æ ‰      ¢ !r       Æ ‰      â !r       Ê ‰      "!r        â/ !r       ÈO ‰      b !r       Æ� ‰      ¨ ‰	      â !r       Ê ‰      " !r       â/ !r       ÈO ‰ €    b !r	       Æ� ‰	 €    ¨ ‰  €    â !r       Ê ‰ €    " !r       â/ x   ÿàÿ â !r	       ÆO z  r  p`ö â $tÿ   ÿ Ž â !r        â� ‰ @    b x 
   pVö Æ ‰  @    ¢ !r       Æ ‰	 @    ( ‰ @    â !r       â/ $z x  Ž â x   ÿàÿ â !r         äO ‰      b x   ÿàÿ â $z% ~  Ž â !r	       â z  r  p`ò â ‰	       " z  r  p`ð â !r       â� x 
   pVò  â‰      ¢ x 
   pVp  â %v j   Ž Ä ‰      â !r       â/ !r	 	       È †¹    â !r       ÆO †© 	    â !r       Æ� †‰    â M        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      &   " z ÿ \  pPò â ¹z  F   
   Æ z ÿ ]  Sò Ú $ÿ   ÿ Ž È % \   Ž Ê �    ™ ¢  $rÿÿ   ÿ Ž â Ey      € â x      ÿ  â y      %   b '  m   Ž Æ y	      "   ä  ÿ n     â 'š s   Ž â y      !   & $ÿÿ    
Ž â y      '   & $  o  Ž Ä $’ ÿÿ    Ž ä $x   ÿ Ž â/ r 	ÿ   pRð È� x    pDp  â$x		    Ž â r   ÿàÿ È z  r  pdp  ä xÿ   	  ä šÿ t    È r     € ÊO G	  °    € ê z ÿ `  pPò â x      ÿ  â z ÿ ^  pPð ä z ÿ b  p ô ä z ÿ a  Sò ä z ÿ _   Sð ä z ÿ c   'ò Ö $
 ~  ÿŽ âÿ     â$š ~  ÿŽ ä $
 x  ÿŽ â   àë â$š x  ÿŽ â ’
   àÇ ä ˜ÿ     ä    ÿàõ Ä ÿ   ÿÄþ ä ’
   ÿàû ä ’
ÿ   ÿ„þ ä ÿ   ÿä ä 
 ^  ÿ„ â $’
ÿÿ   
Ž â š b  ÿ† Ä 
 _   ä š c  
� Æ �	     h �™     d Ay        € ê r	   ÿ� â ¹z  l      â $z }  ÿŽ â ™x?    â xÿ     â $z   w  ÿŽ â Ey  0;    € â $rÿÿ   ÿ Ž â x      ÿ  â |    pbð â $rÿÿ   ÿ Ž â r    ÿàó â 'z y   Ž â x      ÿ  â x      ÿ  â r ÿ       â $r
ÿÿ   ÿŽ  Ì G	  p:    € ê zÿ z    â $z p  Ž âxøÿÿÿÿÀŽ â $v&ÿ p  ÿ Ž â x   ÿàÿ â $z |  Ž â x   ÿàÿ â $x		   
Ž â x   ÿàÿ â $z/ v  Ž â z  `      â $z  |  Ž â r4		   ÿ� â %x(/ˆ   ÿŽ â r ÿ       â x      ÿ  â x4ÿ   4  â $z v   Ž â x      ÿ  â $x	   ÿ Ž â x5ÿ   4  â $z1 p  Ž â x      ÿ  â %x^   ÿŽ È %x*   (Ž È $z7 p  Ž âx   ÿÀŽ â $z3 p  Ž â x*   ÿàñ â %x&&ˆ   ÿŽ Æ rÿ+   ÿä  â %x,1ˆ   ^Ž â x &   ÿüŽ Æ %x./ˆ   4Ž È %x01ˆ   4Ž È %x23ˆ   4Ž È %x47ˆ   4Ž È $vÿ a  ÿ Ž ä $vÿ X  ÿ Ž ä $vÿ Y  ÿ Ž ä $rÿÿ   ÿ Ž ä $x   ÿ Ž ä %x67ˆ   ^Ž È r8*   ÿàñ Ê $x9+     Ê �y[8   ™ ¨  �yY8   ™ â  xÿ   	  ä rH	   à‘ â �y#8   ™ " $tIÿ$   ÿ Ž ä r
   ÿ$~  â �y$8   ™ " %vHH Z  I Ž È $x$   ÿŽ È $xII   Ž Ê �yH   ™ ( �yH   ™ ( �yH   ™ " rF&   à‘ Æ �yH   ™ " rG'   $~  Æ �yH   ™ " r8(   ÿàó Æ �y H   ™ " rD.   ÿàñ Æ �y%Fôÿÿ ™ " rE/   ÿä  Æ �y!H   ™ " $x9)   Ž  Æ �y;D   ‘ ( �y:8   • ( �y"H    ™ ( �y<Føÿÿ ™ ( �y=Füÿÿ ™ ( �y>8    • ( �y?H    • " rN&F   ÿàñ Æ �y@F    ™ $ $xO'   G  Ê �yANôÿÿ ™ ( �yBNøÿÿ ™ " rL&D   ÿàñ Æ �yCNüÿÿ ™ " rT&8   ÿàó â$xM'   E  âr8&N   ÿàñ â �yGN    ™ " rU'9   ÿäÿ  Æ $x9'   O  â �yEL   ‘ ( �yF8ôÿÿ ™ ( �yDT   • " rQ&   (à‘ Æ �yN8üÿÿ ™ "# rI'   )$~  ärP Q   &àµ â ‚x &     â �yHT    • $ rQ'I   'd~ ä �yI8øÿÿ ™ " rRT    ÿàó Ä r\&L   ÿàñ â $~aÿ   ÿ Ž â ‚x �˜­¿     â $xS'   UŽ  â�yJP    • " $x]'   M  â |W       â $~Xÿ   ÿ Ž â �yLR    • ( �yMRþÿÿ • ( �yKPþÿÿ • " $~Zÿ   ÿ Ž Æ �yO\   ‘ ( �yP8    ™ $# r8&\   ÿàñ ä/ xT[ww  ÿÀŽ äO xUÿ   [  Ä tRa5EYqT    äxaUww  ÿÀŽ ä x9ÿ   Y  ä� tQWÏÝêöT    â $~Wÿ   ÿ Ž â |U       ä xVYww  ÿÀŽ Ä tSXÏÝêöa    â $~Xÿ   ÿ Ž â x99ww  ÿÀŽ ä tUUÏÝêöV    ätVW5EYqV    ä tWZÏÝêö9    ätXX5EYq9    Ä r9']   ÿä  Ì �y98   ‘ ¢  x[[ˆˆˆˆÿÀŽ ä xYYˆˆˆˆÿÀŽ â $~Tÿ   ÿ Ž â x[[22ÿø� ä xYY22ÿø� ä xZ[Tv  ÿÀŽ ä tTT5EYqa    Ä x[ÿ   [  ä x\YTv  ÿÀŽ ä xYÿ   Y  ä rQQZ   R    ä rTS[   T    ä rUU\   V    Ä rXWY   X    ä xW#ˆˆˆˆÿÀŽ äxRQ d  T    äxYÿ   #  ä xQQ1u  T    â $~Vÿ   ÿ Ž â xTU d  X    Ä xSU1u  X    â $~\ÿ   ÿ Ž â xU#ww  ÿÀŽ â $~Zÿ   ÿ Ž â xWW22ÿø� ä |X       ä x#Yww  ÿÀŽ ä x[WTv  ÿÀŽ Ä tVVÏÝêö#    ätXX5EYq#    â $~#ÿ   ÿ Ž â xWÿ   W  ä tY\ÏÝêöU    ätZZ5EYqU    â $~Uÿ   ÿ Ž â x8$ww  ÿÀŽ â &rRR   ÿ   â rVVW   X    Ä xWÿ   $  ä t##ÏÝêö8    ät8U5EYq8    â $~Uÿ   ÿ Ž â xX$ˆˆˆˆÿÀŽ â &rRQ   R   â x$Www  ÿÀŽ ä |W       Ä xQX22ÿø� â &rTT   R   â rYY[   Z    ä tUUÏÝêö$    ät$W5EYq$    ä xWQTv  ÿÀŽ â &rSS   T   â xQÿ   Q  Ä xRY d  V    ä r##W   8    ä r$UQ   $    â &rRR   S   â xYY1u  V    ä x8# d  $    äxQ#1u  $    â &r$Y    R   â xS%ww  ÿÀŽ â $~Rÿ   ÿ Ž â x#ÿ   %  â $~Tÿ   ÿ Ž Ä &r$8!   $   â t8RÏÝêöS    â$~Uÿ   ÿ Ž â x%%ˆˆˆˆÿÀŽ ä xR#ww  ÿÀŽ â $x#	   ÿ Ž â |W       ä tST5EYqS    ä xT%22ÿø� Ä t%UÏÝêöR    ätVW5EYqR    ä r;ÿ   ;  ä rRÿ#   :  ä xUTTv  ÿÀŽ ä x:;   ÿÀŽ â $x;R   ÿ Ž â xTÿ   T  â &r$Q"   $   â rS8U   S    Ä rT%T   V    ä x:;0   :âŽ ä xQÿ   <  ä x;<ˆˆˆˆÿÀŽ â $~Uÿ   ÿ Ž â x8S d  T    â$~Vÿ   ÿ Ž â x%S1u  T    Ä |T       ä x<<ww  ÿÀŽ ä xSQww  ÿÀŽ ä x;;22ÿø� ä tQUÏÝêö<    ä tTTÏÝêöS    Ä t<W5EYq<    ä xR;Tv  ÿÀŽ ä tSV5EYqS    ä xUÿ   ;  â &r88   ÿ   â x;:àÿÿÿÿàÿ ä r<QR   <    Ä rSTU   S    â &r8%   8   ä $r$$;   ÿŽ â x:< d  S    äx%<1u  S    ä x;=ww  ÿÀŽ äx<ÿ   =  â $~Rÿ   ÿ Ž â x==ˆˆˆˆÿÀŽ â $~Sÿ   ÿ Ž â |Q       â $~Tÿ   ÿ Ž â x<<ww  ÿÀŽ â &r8:   8   â x==22ÿø� â 0r>ÿ>    A   â t:RÏÝêö;    â 0r?ÿ?    A   â tQQÏÝêö<    Ä tRS5EYq<    ä x<=Tv  ÿÀŽ ä t;T5EYq;    ä x=ÿ   =  â  r>>?     A  â Er$ $       ä r::<   ;    Ä r=Q=   R    â #r>$      â &r$%   8   ä x8: d  =    â$~<ÿ   ÿ Ž â x;@ww  ÿÀŽ â $~>ÿ   ÿ Ž â x%:1u  =    â &r$8   $   â |:       Ä xQAww  ÿÀŽ ä x8ÿ   A  ä t<<ÏÝêö;    ät=>5EYq;    â $~;ÿ   ÿ Ž â t::ÏÝêöQ    ä x88ww  ÿÀŽ Ä tQ>5EYqQ    ä x>ÿ   @  ä xRAˆˆˆˆÿÀŽ â $~Tÿ   ÿ Ž â x@@ˆˆˆˆÿÀŽ ä t;;ÏÝêö8    ät8S5EYq8    Ä xA>ww  ÿÀŽ ä xSR22ÿø� ä x@@22ÿø� ä t>TÏÝêöA    ä xTSTv  ÿÀŽ ä xR@Tv  ÿÀŽ Ä xSÿ   S  ä r<<R   =    ä r;;S   8    ä x8ÿ   B  ä x=Bww  ÿÀŽ â$~Rÿ   ÿ Ž â xBBˆˆˆˆÿÀŽ â $~Xÿ   ÿ Ž â tAV5EYqA    â $~Vÿ   ÿ Ž â r::T   Q    Ä |T       ä xQ8ww  ÿÀŽ ä xBB22ÿø� ä t8RÏÝêö=    ät=T5EYq=    ä tRVÏÝêöQ    Ä tSX5EYqQ    ä x@ÿ   @  ä xQBTv  ÿÀŽ ä xTÿ   B  ä xB: d  ;    ä rA>@   A    Ä x;:1u  ;    â &rBB   ÿ   â r=8Q   =    ä rRRT   S    â &r$%    $   â x8< d  A    â&rB;   B   â x:= d  R    ä x%<1u  A    â &r$8!   $   â x<ÿ   C  â &r8:   B   â x:Cww  ÿÀŽ Ä x;=1u  R    â $~=ÿ   ÿ Ž â xCCˆˆˆˆÿÀŽ â $~Qÿ   ÿ Ž â |A       â $~Sÿ   ÿ Ž â xCC22ÿø� ä x><ww  ÿÀŽ ä t<=ÏÝêö:    Ä t=A5EYq:    ä x:CTv  ÿÀŽ ä t@QÏÝêö>    ätAS5EYq>    ä xCÿ   C  ä r:<:   =    Ä rA@C   A    â &r8;   8   â x=Fww  ÿÀŽ ä x<: d  A    äx;:1u  A    ä x:FˆˆˆˆÿÀŽ ä rDÿ#   D  Ä xFÿ   F  â $~@ÿ   ÿ Ž â rEÿ   E  â $~Aÿ   ÿ Ž â |>       â $~Cÿ   ÿ Ž â x::22ÿø� â $xDD   ÿ Ž â xFFww  ÿÀŽ ä xEE   ÿÀŽ Ä t>>ÏÝêö=    ät=@5EYq=    â &r$%"   $   â x@:Tv  ÿÀŽ ä tAAÏÝêöF    ätFC5EYqF    ä x:ÿ   :  â &r8<   8   â x%D0   EâŽ â $~Eÿ   ÿ Ž â |C       Ä x<Gww  ÿÀŽ ä r=>@   =    ä rFA:   F    â &r8;    8   â t:CÏÝêö<    ät;E5EYq<    ä x<= d  F    Ä xA=1u  F    ä x=IˆˆˆˆÿÀŽ â$~Bÿ   ÿ Ž â x@ÿ   I  â $~Dÿ   ÿ Ž â xIIww  ÿÀŽ ä x>=22ÿø� ä x=@ww  ÿÀŽ â $~Rÿ   ÿ Ž â t@BÏÝêöI    Ä tID5EYqI    ä xB>Tv  ÿÀŽ ä |F       ä r@@B   I    ä tCFÏÝêö=    ätDR5EYq=    Ä xBÿ   >  â &r><   ÿ   Æ rCCB   D    â &r>A   >   â x=ÿ   G  ä xB@ d  C    äxA@1u  C    ä xDNww  ÿÀŽ ä x@ÿ   N  Ä xNNˆˆˆˆÿÀŽ â $~Eÿ   ÿ Ž â xGGˆˆˆˆÿÀŽ â $~Iÿ   ÿ Ž â x==ww  ÿÀŽ ä x@@ww  ÿÀŽ ä xNN22ÿø� â &r>B   >   â xGG22ÿø� Ä t<FÏÝêö=    ä tBEÏÝêöD    ätCI5EYqD    ä tEEÏÝêö@    ätFI5EYq@    ä xDNTv  ÿÀŽ Ä xNÿ   N  ä x@GTv  ÿÀŽ â &r>A   >   â t=R5EYq=    ä xGÿ   G  ä rBBD   C    ä rEEN   F    Ä r::@   ;    ä xAÿ   P  ä x@Pww  ÿÀŽ äxPPˆˆˆˆÿÀŽ ä r=<G   =    â $~Gÿ   ÿ Ž â x<B d  E    Ä x;B1u  E    ä |C       â $~Eÿ   ÿ Ž â xBAww  ÿÀŽ ä xPP22ÿø� ä tACÏÝêö@    ä tCGÏÝêöB    Ä tDI5EYqB    ä xBPTv  ÿÀŽ ä t@E5EYq@    ä xPÿ   P  ä rLÿ#   L  â &r><   >   â r@AB   @    Ä rCCP   D    ä rOÿ   O  ärJÿ#   J  ä r9ÿ   9  âO $xLL   ÿ Ž â x<: d  =    äx=:1u  =    Ä z ÿ `  pPð â &r>;    >   â x:@ d  C    ä xOO   ÿÀŽ ä xJJ   ÿ   ä x99   ÿÀŽ ä x%%àÿÿÿÿàÿ Ä z ÿ a   Sð â &r8<!   8   âxC@1u  C    â &r::!   >   â xOL0   OâŽ ä x9J0   9âŽ â $r$$%   ÿŽ â xOOàÿÿÿÿàÿ â &r8="   8   âx99àÿÿÿÿàÿ â &r:C"   :   Ä 0rHÿH    A   â Er$ $       â 0r>ÿK    A   ä 0r<ÿM    A   â  r%?H     A  â$r88O   ÿŽ ä $r9:9   ÿŽ â  rA?>     A  â #r%$      â  r<?<     A  â Er> 8       Ä Er= 9       â G‰  €    € ê r:*   ÿàó Ê $x;+   Ž  Ê �yF:    ¨ �yE:    è �y@:    " r$.   ÿàó Æ �yB:    â r8(   ÿàõ â $x%/   Ž  È $x9)    ä �y%$    è  �yD8    è �yC8     " $~Iÿ   ÿ Ž â |G       â $~Hÿ   ÿ Ž â |K       â $~Jÿ   ÿ Ž Ä $~Mÿ   ÿ Ž â x;ÿ   F  äO x:Fww  ÿÀŽ äxFFˆˆˆˆÿÀŽ ä x;;ww  ÿÀŽ ä xFF22ÿø� ä t$GÏÝêö:    Ä t9I5EYq:    ä/ t:HÏÝêö;    â$~Hÿ   ÿ Ž â x8FTv  ÿÀŽ ä t;J5EYq;    ä xFÿ   F  ä r$$8   9    Ä r;:F   ;    ä x9ÿ   E  ä� x8$ d  ;    äx;$1u  ;    ä x$Eww  ÿÀŽ â&r88   ÿ   â xEEˆˆˆˆÿÀŽ Ä x:9ww  ÿÀŽ â &r8;   8   â xEE22ÿø� ä t9GÏÝêö$    ä tFKÏÝêö:    ätGM5EYq:    ä t$I5EYq$    Ä x:ETv  ÿÀŽ ä xEÿ   E  ä r$9:   $    â $~:ÿ   ÿ Ž â x9@ww  ÿÀŽ äx;ÿ   @  ä x@@ˆˆˆˆÿÀŽ Ä rEFE   G    ä |G       ä xF;ww  ÿÀŽ ä x@@22ÿø� ä t;H5EYq9    ä tGGÏÝêöF    Ä tHI5EYqF    ä xF@Tv  ÿÀŽ ä t::ÏÝêö9    ä x@ÿ   @  ä x9$ d  E    ä r::F   ;    â $~;ÿ   ÿ Ž â rGG@   H    â &r89   8   â xE$1u  E    â $~Fÿ   ÿ Ž â x$: d  G    â$~Hÿ   ÿ Ž â xG:1u  G    â &r8E   8   â x9BˆˆˆˆÿÀŽ Ä x:ÿ   B  â &r$$   8   â xBBww  ÿÀŽ ä x9922ÿø� â &r$G    $   â xE:ww  ÿÀŽ ä t:;ÏÝêöB    ät;I5EYqB    Ä x@9Tv  ÿÀŽ ä tBFÏÝêöE    ätEH5EYqE    ä x9ÿ   9  ä r%ÿ   %  ä r::@   ;    Ä rDÿ#   D  ä r9B9   E    ä x8%   ÿÀŽ ä x;D   ÿ   ä x%: d  9    äx9:1u  9    Ä x8;0   8âŽ â &r$%!   $   Æ x%8àÿÿÿÿàÿ â &r$9"   $   ä 0r8ÿC    A   ä $r$$%   ÿŽ Æ  r8?8     A  ä Er$ $       Ê #r8$      ä G‰  Ð    € ê r:6   ÿàó Ê $x;7   Ž  Ê �yF:    ¨ �yE:    è �y@:    " x8   ÿàÿ â $r9ÿÿ    Ž Ä �yB:    â r$4   ÿàó â $zC8 p  Ž ä $r8ÿÿ    Ž â r%5   ÿäÿ  Æ %x8Cˆ   8Ž Æ �y%$    è  �yD8    ( �yC8     " $~Iÿ   ÿ Ž ä $~Hÿ   ÿ Ž â |J       â $~Kÿ   ÿ Ž â |M       Ä x;ÿ   F  äO x:Fww  ÿÀŽ äxFFˆˆˆˆÿÀŽ ä xG;ww  ÿÀŽ ä xFF22ÿø� ä t;IÏÝêö:    Ä t8HÏÝêöG    ä/x$FTv  ÿÀŽ ä t:K5EYq:    ä tGJ5EYqG    ä xFÿ   F  ä r$;$   :    Ä rG8F   G    ä x9ÿ   E  ä� x8$ d  G    äxG$1u  G    ä x$Eww  ÿÀŽ â$~;ÿ   ÿ Ž â xEEˆˆˆˆÿÀŽ â $~Kÿ   ÿ Ž â x:9ww  ÿÀŽ â $~Iÿ   ÿ Ž â xEE22ÿø� Ä t9;ÏÝêö$    ä t;KÏÝêö:    ätFM5EYq:    ä x:ETv  ÿÀŽ ä t$I5EYq$    ä xEÿ   E  â &r88   ÿ   â r$9:   $    Ä r;;E   F    ä x9@ww  ÿÀŽ äxEÿ   @  â &r8G   8   â x@@ˆˆˆˆÿÀŽ â $~Hÿ   ÿ Ž â xFEww  ÿÀŽ â $~Gÿ   ÿ Ž â x@@22ÿø� â $~:ÿ   ÿ Ž â tEH5EYq9    Ä tGGÏÝêöF    ätHI5EYqF    ä xF@Tv  ÿÀŽ ä t::ÏÝêö9    ä x@ÿ   @  ä x9$ d  ;    Ä r::F   E    ä rGG@   H    â &r89   8   â x;$1u  ;    ä x$: d  G    äxG:1u  G    â &r8;   8   â x9BˆˆˆˆÿÀŽ Ä x:ÿ   B  â $~Fÿ   ÿ Ž â |;       ä xBBww  ÿÀŽ â $~Hÿ   ÿ Ž â x9922ÿø� ä xE:ww  ÿÀŽ ä t:;ÏÝêöB    Ä t;I5EYqB    ä x@9Tv  ÿÀŽ ä tBFÏÝêöE    ätEH5EYqE    ä x9ÿ   9  ä rDÿ#   D  Ä r%ÿ   %  â &r$$   8   â r::@   ;    ä r9B9   E    â $x;D   ÿ Ž â x8%   ÿÀŽ â &r$G    $   â x%: d  9    äx9:1u  9    Ä x8;0   8âŽ â &r$%!   $   Æ x%8àÿÿÿÿàÿ â &r$9"   $   ä 0r8ÿC    A   ä $r$$%   ÿŽ Æ  r%?8     A  ä Er$ $       Ê #r%$      ä G‰  à    € ê x$   ÿàÿ Ê $zC$ p  Ž È %x$Cˆ   ^Ž Ê r:$   ÿàó È r;%   ÿäÿ  Ê �yF:    ¨ �yE:    è �y@:    " r$2   ÿàó â $r8ÿÿ    Ž Ä �yB:    â $r9ÿÿ    Ž ä$x%3   Ž  ä %x8Cˆ   8Ž È �y%$    è  �yD8    è �yC8     " $~Hÿ   ÿ Ž â |I       â $~Kÿ   ÿ Ž Ä $~Jÿ   ÿ Ž ä $~Mÿ   ÿ Ž â x;ÿ   F  äO x:Fww  ÿÀŽ äxFFˆˆˆˆÿÀŽ ä xG;ww  ÿÀŽ ä xFF22ÿø� Ä t;IÏÝêö:    ät8HÏÝêöG    ä/x$FTv  ÿÀŽ ä t:K5EYq:    ä tGJ5EYqG    ä xFÿ   F  Ä r$;$   :    ä rG8F   G    ä x9ÿ   E  ä� x8$ d  G    äxG$1u  G    ä x$Eww  ÿÀŽ Ä xEEˆˆˆˆÿÀŽ â $~Kÿ   ÿ Ž â |;       â $~Iÿ   ÿ Ž â x:9ww  ÿÀŽ ä xEE22ÿø� ä t9;ÏÝêö$    ä t;KÏÝêö:    Ä tFM5EYq:    ä x:ETv  ÿÀŽ ä t$I5EYq$    ä xEÿ   E  â &r88   ÿ   â r$9:   $    ä r;;E   F    Ä x9@ww  ÿÀŽ äxEÿ   @  â &r8G   8   â x@@ˆˆˆˆÿÀŽ â $~Hÿ   ÿ Ž â xFEww  ÿÀŽ â $~Gÿ   ÿ Ž â x@@22ÿø� ä |:       Ä tEH5EYq9    ätGGÏÝêöF    ätHI5EYqF    ä xF@Tv  ÿÀŽ ä t::ÏÝêö9    ä x@ÿ   @  Ä x9$ d  ;    ä r::F   E    ä rGG@   H    â &r89   8   â x;$1u  ;    ä x$: d  G    äxG:1u  G    â &r8;   8   â x9BˆˆˆˆÿÀŽ â $~;ÿ   ÿ Ž â x:ÿ   B  Ä xBBww  ÿÀŽ â $~Hÿ   ÿ Ž â |F       ä x9922ÿø� ä xE:ww  ÿÀŽ ä t:;ÏÝêöB    ät;I5EYqB    Ä x@9Tv  ÿÀŽ ä tBFÏÝêöE    ätEH5EYqE    ä x9ÿ   9  ä rDÿ#   D  ä r%ÿ   %  â &r$$   8   â r::@   ;    Ä r9B9   E    â $x;D   ÿ Ž â x8%   ÿÀŽ â &r$G    $   â x%: d  9    äx9:1u  9    ä x8;0   8âŽ â &r$%!   $   Æ x%8àÿÿÿÿàÿ â &r$9"   $   ä 0r8ÿC    A   ä $r$$%   ÿŽ Æ  r8?8     A  ä Er$ $       Ê #r8$      â #r<>      â #rA=      â G‰  Ð    € ê r:,   ÿàñ Ê $x;-     Ê �yB:    ¨ �yA:    è �y<:    " x8   ÿàÿ â $r9ÿÿ    Ž Ä �y=:    â r$0   ÿàñ â$zC8 p  Ž â r8        Æ $x%1     ä %x8Cˆ   8Ž È �y%$    è  �y>8    è �y@8     â $~Eÿ   ÿ Ž ä $~Dÿ   ÿ Ž â |G       â $~Fÿ   ÿ Ž â x;ÿ   B  ÄO x:Bww  ÿÀŽ äxBBˆˆˆˆÿÀŽ ä xC;ww  ÿÀŽ ä xBB22ÿø� ä t;EÏÝêö:    ä t8DÏÝêöC    Ä/ x$BTv  ÿÀŽ ä t:G5EYq:    ä tCF5EYqC    ä xBÿ   B  ä r$;$   :    ä rC8B   C    È x8$ d  C    äxC$1u  C    ä x$ÿ   A  â� &r88   ÿ   â xAww  ÿÀŽ â$~:ÿ   ÿ Ž â xAAˆˆˆˆÿÀŽ ä x9$ww  ÿÀŽ Ä xAA22ÿø� ä t$:ÏÝêö    ä |B       ä t:DÏÝêö9    ät;F5EYq9    ä x9ATv  ÿÀŽ Ä xAÿ   A  ä tB5EYq    â &r8C   8   â r::A   ;    â $~;ÿ   ÿ Ž â r$9       ä x<ww  ÿÀŽ äx9ÿ   <  â $~Aÿ   ÿ Ž â x<<ˆˆˆˆÿÀŽ â $~Bÿ   ÿ Ž â t$;ÏÝêö    Ä x;9ww  ÿÀŽ ä |D       ä x<<22ÿø� ä t9A5EYq    ä tABÏÝêö;    ä x d  :    Ä tBD5EYq;    ä x;<Tv  ÿÀŽ ä x<ÿ   <  â &r   8   â x1u  :    ä r$$;   9    ä rAA<   B    â &r      â x=ˆˆˆˆÿÀŽ Ä x$ d  A    äxÿ   =  ä xA$1u  A    â $~$ÿ   ÿ Ž â x==ww  ÿÀŽ â $~9ÿ   ÿ Ž â x22ÿø� â $~:ÿ   ÿ Ž â x8ww  ÿÀŽ Ä |;       ä t$ÏÝêö=    äx$Tv  ÿÀŽ ä t99ÏÝêö8    ät=:5EYq=    ä t8;5EYq8    Ä xÿ     ä r>ÿ#   >  â &r      â r$   =    ä r%ÿ   %  ä r9   8    â $x>>   ÿ Ž â x%%   ÿÀŽ â &r A       â x d      Ä x1u      ä x%>0   %âŽ â &r!       Æ x%%àÿÿÿÿàÿ â &r"      ä 0r@ÿ@    A   ä $r%   ÿŽ Æ  r??@     A  ä Er        Ê #r?      â x   ÿàÿ ä x€  ÿàó ä |    pbð ä x€  ÿàõ â $rÿÿ   Ž  â x€   ÿàÿ Æ $rÿÿ    Ì D	       À â Gy  @Èÿÿÿÿƒ ê Ay        € ê y      "   " Ey  0    € â r ÿ   pRð är ÿ   pRò Ö G‰  ð     € ê z ÿ `  pPð ä r   ÿ8Ž ä z ÿ a   Sð Æ ˆs  þÿ H   ô ˆƒ €þÿ H   è ˆƒ  ÿÿ H   è ˆƒ €ÿÿ H   â G‰  p     € ê ˆs    H   è ˆs €þÿ H   è ˆs €  H   è ˆs  ÿÿ H   è ˆs    H   è ˆs €ÿÿ H   è ˆs €  H   ä Ay        € ê {          ì M        € ê y       &   ¢ z ÿ `  pPð â $tÿ   ÿ Ž Ä y      '   "> z ÿ a   Sð â $z   x  Ž ÈO $z  ~   Ž È %v  j   Ž È G	  P    € ê „y      H   " z  r  p`ð Æ „y	 €   H   b r ÿ   pVp  Æ „y    H   ¨ „y    H   è „y €  H   ( „y    H   ( „y    H   ( „y €  H   ( „y    H   ( „y €  H   („y
 €  H   " !r         Æ „y €  H   " !r		       â/ !r       âO !r         ä� !r	       â!r       â !r         ä !r       Æ ‰       b !r       Æ ‰      ¢ !r       Æ ‰      â !r

       È !r
       â !r        â/ !r       ÈO ‰      " !r
       Æ� ‰      h ‰      ¨ ‰	
      â !r       â !r       â/ !r       ÈO ‰      " !r
	       Æ� ‰ €    h ‰  €    ¨ ‰	 €    â !r       â !r       È/ ‰ €    " !r
        ÆO ‰  @    b !r		       â� x   ÿàÿ ä ‰
 @    ¤ z  r  p`ô ä ‰	 @    ä x    pVt â !r       â !r         È/ ‰ @    " x   ÿàÿ â!r
       âO x   ÿàÿ â ‰       b z  r  p`ò â !r	       ä� ‰      ¢ x    pVò  Æ ‰      â !r       â !r        È/ ‰       " !‚       â !r	       âO !r       Æ� †‰     â z  r  p`ð â !¢		       â!’       ä x    pVp  ä †© 	   è †™    â !r        Ì M	        € ê !r       Ê/ †y    â My        € ê $v ÿ h  ÿ Ž Ê x     pBð Ú G	  @    € ê x     pRð Ú G‰  À    € ê x     pRð Ú G	   $    € ê „y     H   " z  r  p`ð â Ey  p    € ä „y     H   b r ÿ   pVp  Æ „y	    H   ¨ „y    H   è „y    H   ( „y  
  H   ( „y €  H   è „y €  H   â !r       Æ „y €  H   " !r         â/ !r	       ÆO !r         â� !r       Æ!r         ä ‰      h ‰       ¢ !r       â !r       ä/ „y €   H   â !r        ÆO ‰      ( ‰      b !r       â !r	       È/ ‰ €    ( ‰
	 €    b !r       Æ „y €  H   ¢ !r       È !r       â !r       ä „y €
  H   â !r
	
       Æ/ ‰ @    ( ‰
 @    b !r       âO !r       â !r
       È/ ‰       ( ‰      b !r       â !r	        â !r       â/ G	  �     € ê !r       â!r 	       Æ  x;ª¸¿  A  Ì s        $ !t  €?    Ì s        $  r     A  È  r      A  Ê †y     ä Ay        € ê ‰       b Ey  Ð    € æ ‰      ¨ „y    H   è „y  	  H   è „y    H   â !r         Æ/ „y    H   â !r       ÆO ‰       h ‰      ( „y    H   ¢ !r       â� !r        â/ !r	       È ‰ €    " x   ÿàÿ Æ ‰
	 €    b z  r  p`ð â !r       Æ x    pVp  â !r       â !r       ÈO !r       â !r       ä „y    H   â !r
	
       Æ/ ‰ @    ( ‰
 @    b !r       â !r
       È/ ‰       ( ‰      b !r       â !r	        â !r       â/ G	  �     € ê !r       â!r 	       Æ  x;ª¸¿  A  Ì s        $ !t  €?    Ì s        $  r     A  È  r      A  Ê †y    ä Ay        € ê ‰       b Ey       € æ ‰      ¨ „y €  H   è „y €  H   è „y €  H   è „y €  H   â !r         Æ/ „y €  H   â !r       ÆO ‰       h ‰      ( „y €	  H   ¢ !r       â� !r        â/ !r	       È ‰ €    ( ‰
	 €    b !r       ÈO !r       â !r       â x   ÿàÿ â !r
	
       Æ/ ‰ @    " z  r  p`ð Æ ‰
 @    b x    pVp  â !r       â !r
       È/ ‰       ( ‰      b !r	        â !r       â/ G	  �     € ê !r       â!r 	       Æ  x;ª¸¿  A  Ì s        $ !t  €?    Ì s        $  r     A  È  r      A  Ê †y    ä Ay        € ê ‰      b !r       â x   ÿàÿ Æ !r       â z  r  p`ð Æ !r       â x    pVp  È ‰       ¢ !r       Ê/ ‰      " !r         ÊO ‰       b !r	       Ê ‰
	 €    " !r        Ê/ ‰ €    b !r
	
       Ê ‰
 @    " !r       Ê/ ‰ @    b !r
       Ê ‰      " !r       Ê/ ‰       ¢ !r       â M	        € ì !r       â/!r        ÆO  x;ª¸¿  A  â !r       Ê s        $ !t  €?    Ì s        $  r     A  È  r     A  Ê †y    â My        € ê „y      H   " z  r  p`ð â Ey       € ä „y    H   b r ÿ   pVp  Æ „y    H   ¨ „y    H   è „y	    H   ( „y  
  H   ( „y €  H   è „y €
  H   â !r         Æ „y €  H   " !r       Æ/ „y €   H   â !r         âO !r       Æ� !r	 	       â!r       È ‰ 	      h ‰      ¢ !r       â !r 	        â/ !r       ä !r       â !r       âO ‰       $ !r       Ä ‰      h „y €  H   â !r        â !r       È/ ‰ €    ( ‰ €    h „y €  H   ¢ !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       ÈO !r       â !r       â !r
	
       È/ ‰       ( ‰
      b !r	        â !r 
       â/ G	  À     € ê !r         È x '7=  A  â x *BL?  A  Æ #t   €?   È  r     A  â  x    ?  A  â !r 	       È s     $   $ !t  €?    È  r     A  È  r      A  Ê †y     ä Ay        € ê ‰       b Ey       € æ ‰      ¨ „y    H   è „y    H   ( „y    H   è „y    H   â !r         â/ !r       ÈO ‰       h ‰      ( „y  	  H   ¢ !r       â� !r       â!r        Ä/ !r       â !r       ä ‰ €    ( ‰ €    h „y    H   â !r       ÈO !r       â !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       â� !r       â x   ÿàÿ â !r
	
       Æ/ ‰       " z  r  p`ð Æ ‰
      b x    pVp  â !r	        â !r 
       Ö/ G	  À     € ê !r         È x '7=  A  â x *BL?  A  Æ #t   €?   È  r     A  â  x    ?  A  â !r 	       È s     $   $ !t  €?    È  r     A  È  r      A  Ê †y    ä Ay        € ê ‰       b Ey  Ð    € æ ‰      ¨ „y €  H   è „y €	  H   ( „y €  H   â !r         Æ/ „y €  H   â !r       ÆO ‰       h ‰      ( „y €  H   â !r       È� !r       â!r        Æ/ !r       â !r       â ‰ €    ( ‰ €    h „y €  H   â !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       â x   ÿàÿ â !r
	
       Æ/ ‰       " z  r  p`ð Æ ‰
      b x    pVp  â !r	        â !r 
       Ö/ G	  À     € ê !r         È x '7=  A  â x *BL?  A  Æ #t   €?   È  r     A  â  x    ?  A  â !r 	       È s     $   $ !t  €?    È  r     A  È  r      A  Ê †y    ä Ay        € ê ‰      b !r       â x   ÿàÿ Æ !r       â z  r  p`ð Æ !r       â x    pVp  È ‰       ¢ !r       Ê/ ‰      " !r         ÊO ‰       b !r       Ê ‰ €    " !r        Ê/ ‰ €    b !r	       Ê ‰
	 @    " !r       Ê/ ‰ @    b !r
	
       Ê ‰
      " !r       Ê/ ‰       ¢ !r
       â M	        € ì !r       â/!r        ÆO  x'7=  A  â x*BL?  A  â !r       ä #t  €?   â x   ?  A  Æ  r     A  Ì s     $   $ !t   €?    È  r      A  È  r     A  Ê †y    â My        € ê x     pRð Ú G‰  P    € ê x     pRð Ú G	  `    € ê „y      H   " z  r  p`ð â Ey       € ä „y    H   b r ÿ   pVp  Æ „y    H   ¨ „y    H   è „y	    H   ( „y  
  H   ( „y €  H   è „y €
  H   â !r         Æ „y €  H   " !r       Æ/ „y €   H   â !r         âO !r       Æ� !r	 	       â!r       È ‰ 	      h ‰      ¢ !r       â !r 	        â/ !r       ä !r       â !r       âO ‰       $ !r       Ä ‰      h „y €  H   â !r        â !r       È/ ‰ €    ( ‰ €    h „y €  H   ¢ !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       ÈO !r       â !r       â !r
	
       È/ ‰       ( ‰
      b !r	        â !r 
       â/ G	  À     € ê !r         Ê	z  i    � â !r 	       È  x;ª¸¿  A  â 	z   i    � È 	z   i €  � â s        $ !t  €?    Ì s        $  r     A  È  r      A  Ê †y     ä Ay        € ê ‰       b Ey       € æ ‰      ¨ „y    H   è „y    H   ( „y    H   è „y    H   â !r         â/ !r       ÈO ‰       h ‰      ( „y  	  H   ¢ !r       â� !r       â!r        Ä/ !r       â !r       ä ‰ €    ( ‰ €    h „y    H   â !r       ÈO !r       â !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       â� !r       â x   ÿàÿ â !r
	
       Æ/ ‰       " z  r  p`ð Æ ‰
      b x    pVp  â !r	        â !r 
       Ö/ G	  À     € ê !r         Ê	z  i    � â !r 	       È  x;ª¸¿  A  â 	z   i    � È 	z   i €  � â s        $ !t  €?    Ì s        $  r     A  È  r      A  Ê †y    ä Ay        € ê ‰       b Ey  Ð    € æ ‰      ¨ „y €  H   è „y €	  H   ( „y €  H   è „y €  H   â !r         Æ/ „y €  H   â !r       ÆO ‰       h ‰      " !r       È� !r       â!r        â/ !r       È ‰ €    ( ‰ €    h „y €  H   ¢ !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       âO !r       â x   ÿàÿ â !r
	
       Æ/ ‰       " z  r  p`ð Æ ‰
      b x    pVp  â !r	        â !r 
       Ö/ G	  À     € ê !r         Ê	z  i    � â !r 	       È  x;ª¸¿  A  â 	z   i    � È 	z   i €  � â s        $ !t  €?    Ì s        $  r     A  È  r      A  Ê †y    ä Ay        € ê ‰      b !r       â x   ÿàÿ Æ !r       â z  r  p`ð Æ !r       â x    pVp  È ‰       ¢ !r       Ê/ ‰      " !r         ÊO ‰       b !r       Ê ‰ €    " !r        Ê/ ‰ €    b !r	       Ê ‰
	 @    " !r       Ê/ ‰ @    b !r
	
       Ê ‰
      " !r       Ê/ ‰       ¢ !r
       â M	        € ì !r       â/!r        ÈO 	z i    � â !r       È  x;ª¸¿  A  â 	z i    � È 	z i €  � â s        $ !t  €?    Ì s        $  r     A  È  r     A  Ê †y    â My        € ê „y      H   " z  r  p`ð Æ „y    H   b r ÿ   pVp  Æ „y €   H   ¨ „y €  H   è „y
    H   ( „y    H   ( „y    H   ( „y    H   ( „y €  H   ( „y €  H   ( „y!    H   " !r         Æ „y#  	  H   " !r       Æ/ „y	    H   b !r       ÆO „y  
  H   ¢ !r       Æ� „y €  H   â !r

       Æ„y €
  H   " !r       Æ „y%    H   " !r         Æ „y'    H   " !r       Æ „y) €  H   " !r       Æ „y+ €  H   " !r       Æ „y €  H   " !r

!       Æ „y €	  H   â !r#       Æ „y- €  H   â !r 	       Æ/ „y/ €  H   â !r       âO !r       ä� ‰       " !r       Æ‰      b !r

%       Æ ‰      ¢ !r'       Æ ‰      â !r))       Æ ‰	
      " x   ÿàÿ â !r++       ä ‰      " z  r  p`ò â !r)       â !r         ä x    pVò  â !r+       â !r       â/ !r--       â !r       ÄO ‰       " !r//       â !r       ä� ‰-      b !r

	       Æ‰/      ¢ !r       â x   ÿàÿ ä‰      â x   ÿàÿ ä z  r  p`ô â ‰	      " z  r  p`ö ä x    pVt â‰      " x    pVö Æ ‰
      " !r        Æ ‰      " !r-       â/ !r/       äO ‰  €    b !r       Æ� ‰      ¢ !r		       Æ‰      â !r       Æ ‰ €    " !r
       Æ ‰	 €    " !r       Æ ‰ €    ( ‰
 €    " !r         Æ/ ‰ €    b !r       âO !r       ä� ‰  @    ¢ !r       Æ‰ €    â !r	       Æ ‰ €    " !r       Æ ‰ @    " !r

       Æ ‰	 @    " !r       Æ/ ‰ @    h ‰
 @    b !r        ÆO ‰ @    ¢ !r       â� !r       ä‰       â !r       Æ ‰ @    " !r		       Æ ‰ @    " !r       Æ/ ‰      b !r
       Æ ‰	      b !r       ÆO ‰      ¨ ‰
      ¢ !r        Æ� ‰      â !r       â !‚        â!r       Æ‰      " !r       Æ/ ‰      b !r		       â !‚       â!r       äO !’	       â  ‚      A  â !r
       â !’       Æ †‰     â !r       â� !¢       â  ’     A  ä !¢	       â !r       ä †™    â  ¢		     A  â !r       È/ †© 	   â M9        € ê !r       â !r       È  r     A  Ê †y    â My        € ê „y      H   " z  r  p`ð â Ey  À    € ä „y    H   b r ÿ   pVp  Æ „y    H   ¨ „y    H   è „y	    H   ( „y  
  H   ( „y €  H   è „y €
  H   â !r         Æ „y €  H   " !r       Æ/ „y €   H   â !r         âO !r       Æ� !r	 	       â!r       È ‰ 	      h ‰      ¢ !r       â !r 	        â/ !r       ä !r       â !r       âO ‰       $ !r       Ä ‰      h „y €  H   â !r        â !r       È/ ‰ €    ( ‰ €    h „y €  H   ¢ !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       ÈO !r       â !r       â !r
	
       È/ ‰       ( ‰
      b !r        â !r 
       â/ G	  à     € ê !r         Ê	x   à@  � Ê  x #ÛÙ?  A  È  x ;ª¸¿  A  â !r        Ê s        " 	x    à@  � È 	x    àÀ  � Ê !t   €?    â !t  €?    È s	        $  r	     A  È  r     A  Ê †y     ä Ay        € ê ‰       b Ey       € æ ‰      ¨ „y    H   è „y    H   ( „y    H   â !r         Æ/ „y    H   â !r       ÆO ‰       h ‰      ( „y  	  H   ¢ !r       â� !r       â!r        Ä/ !r       â !r       ä ‰ €    ( ‰ €    b !r       Æ „y    H   â !r       âO !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       â� !r       â x   ÿàÿ â !r
	
       Æ/ ‰       " z  r  p`ð Æ ‰
      b x    pVp  â !r        â !r 
       Ö/ G	  à     € ê !r         Ê	x   à@  � Ê  x #ÛÙ?  A  È  x ;ª¸¿  A  â !r        Ê s        " 	x    à@  � È 	x    àÀ  � Ê !t   €?    â !t  €?    È s	        $  r	     A  È  r     A  Ê †y    ä Ay        € ê ‰       b Ey  ð    € æ ‰      ¨ „y €  H   è „y €	  H   ( „y €  H   è „y €  H   â !r         Æ/ „y €  H   â !r       ÆO ‰       h ‰      ( „y €  H   â !r       È� !r       â!r        Æ/ !r       â !r       â ‰ €    ( ‰ €    b !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       â x   ÿàÿ â !r
	
       Æ/ ‰       " z  r  p`ð Æ ‰
      b x    pVp  â !r        â !r 
       Ö/ G	  à     € ê !r         Ê	x   à@  � Ê  x #ÛÙ?  A  È  x ;ª¸¿  A  â !r        Ê s        " 	x    à@  � È 	x    àÀ  � Ê !t   €?    â !t  €?    È s	        $  r	     A  È  r     A  Ê †y    ä Ay        € ê ‰      b !r       â x   ÿàÿ Æ !r       â z  r  p`ð Æ !r       â x    pVp  È ‰       ¢ !r       Ê/ ‰      " !r         ÊO ‰       b !r       Ê ‰ €    " !r        Ê/ ‰ €    b !r	       Ê ‰
	 @    " !r       Ê/ ‰ @    b !r
	
       Ê ‰
      " !r       Ê/ ‰       ¢ !r
       â M	        € ì !r       â/!r        ÈO 	x  à@  � â !r       È  x#ÛÙ?  A  â 	x  à@  � Æ  x;ª¸¿  A  â 	x  àÀ  � Ê s        " !t  €?    â !t  €?    Ì s         $  r      A  È  r     A  Ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y       "   $ z   w  p`ð Ú M	        € ê y      !   " ¹z  l      â Ey  �    € â ™x?    â y      %   b ¹z  F   
   â $rÿÿ   ÿ Ž ä $rÿÿ   ÿ Ž â x$ÿ     È | $   p`ð â $x   ÿ Ž Ø/ G	  ð    € ê y      &   " $tÿ   ÿ Ž ä $z  x  Ž È %v \   Ž Ì �y    ™ ¢  $rÿÿ   ÿ Ž ä $z  r  ÿŽ ä 'z m   Ž È $x	$   ÿ Ž â zÿ n    â $z q  $Ž ä $x   ÿ Ž ä$rÿÿ   
Ž ä $x   ÿ Ž â x   ÿÀŽ â $z o  Ž Ä $vÿ X  ÿ Ž â x   ÿÀŽ â $z u  ÿŽ ä $vÿ Y  ÿ Ž ä %x$   ÿ Ž È $rÿÿ   ÿ Ž ä %x$    Ž â x   ÿÀŽ Ê %x$    Ž â xÿ     Ê %x	$   Ž Ê z Z  ÿàñ ä x   ÿàÿ ä x   ÿàó È zÿ [  þ  â $z	 t  $Ž äO$z t  Ž ä $x   ÿ Ž ä $z q  	Ž Æ x	p   ÿÀŽ â %xˆ   ÿŽ È %xˆ   ÿŽ â r	   ÿàñ äx   þŽ är	   ÿàó â $r"ÿÿ     â x    þŽ Æ $rÿÿ   Ž  ä r   ÿàñ Ê $x	   "  Ê �y3   ™ ¨  �y2   ™ è  �y1   ™ " $rÿÿ    Ž â r   ÿàñ Ä �y%   ™ f $x     â �y(ôÿÿ ™ è �y+   ™ è �y*   ™ è  �y)øÿÿ ™ ( �y&   ™ ( �y   ™ ( �y!üÿÿ ™ ( �y   ™ ( �y'   ™ " r
   ÿàó Æ �y#    ™ " r   ÿà÷ Æ �y,   ™ " r    ÿàõ är   ÿàñ â $x   Ž  â�y-   ™ b $x   Ž ä$x    â�y.
   • b $x	     Æ �y   • h# �y    ‘ h �y    ‘ h �yðÿÿ • h �y/
    • h �y0    • b ‚x �˜­¿     Ä ‚x &     â $~6ÿ   ÿ Ž ä $~8ÿ   ÿ Ž ä $~5ÿ   ÿ Ž ä $~ÿ   ÿ Ž ä $~:ÿ   ÿ Ž â x$$   ÿàÿ È | $   p`ð ä x   ÿàó ä x€  ÿàõ Æ $rÿÿ   Ž  â x4ÿ   3  äO x	3ww  ÿÀŽ äx33ˆˆˆˆÿÀŽ ä x44ww  ÿÀŽ ä x3322ÿø� ä t
6ÏÝêö	    Ä t	85EYq	    ä x3Tv  ÿÀŽ ä t5ÏÝêö4    ä/t5EYq4    ä x3ÿ   3  ä r

   	    Ä r3       ä xÿ   2  ä� x	2ww  ÿÀŽ â$~4ÿ   ÿ Ž â x22ˆˆˆˆÿÀŽ â $~8ÿ   ÿ Ž â xww  ÿÀŽ â $~6ÿ   ÿ Ž â x2222ÿø� Ä t4ÏÝêö	    ä t38ÏÝêö    ät4:5EYq    ä t65EYq	    ä x2Tv  ÿÀŽ ä x2ÿ   2  Ä x	
 d      äx

1u      ä r       ä r32   4    â $~4ÿ   ÿ Ž â xÿ   1  äx1ww  ÿÀŽ Ä x11ˆˆˆˆÿÀŽ â $~3ÿ   ÿ Ž â x2ww  ÿÀŽ ä x1122ÿø� ä t4ÏÝêö    ät465EYq    ä t635EYq2    Ä x31Tv  ÿÀŽ â &r	(   ÿ   â t55ÏÝêö2    ä x1ÿ   1  â &r
+      â x2 d      ä r3   4    ä r651   6    Ä x	1u      â $~ÿ   ÿ Ž â x3*ˆˆˆˆÿÀŽ â $~1ÿ   ÿ Ž â x4%ww  ÿÀŽ â&r22)      â x d  6    â$~5ÿ   ÿ Ž â x
1u  6    â $~ÿ   ÿ Ž â xÿ   *  Ä x**ww  ÿÀŽ ä x3322ÿø� ä tÏÝêö4    ät415EYq4    ä x1ww  ÿÀŽ ä t655EYq*    Ä tÏÝêö*    ä x53Tv  ÿÀŽ ä x%ˆˆˆˆÿÀŽ ä t*8ÏÝêö1    ä r5   6    ä t1:5EYq1    Ä x22ÿø� ä x6ÿ   3  ä x3Tv  ÿÀŽ ä r6*6   1    ä x%ÿ   %  â &r		&   2   â r3   4    Ä x1 d  6    äx21u  6    ä x5%ww  ÿÀŽ ä xÿ     ä x3ww  ÿÀŽ â$~4ÿ   ÿ Ž â xˆˆˆˆÿÀŽ â $~6ÿ   ÿ Ž â t%8ÏÝêö5    Ä t*:5EYq5    ä x5ww  ÿÀŽ ä x22ÿø� ä t4ÏÝêö3    ät465EYq3    ä t68ÏÝêö5    Ä x3Tv  ÿÀŽ ä t5:5EYq5    ä xÿ     ä xÿ     â &r11(   ÿ   â r3   4    ä r66   5    â &r	!   	   â r*%   *    â &r22+   1   â x d  6    â &r	
   	   â x( d  *    Ä x1u  *    ä xÿ   '  â $~%ÿ   ÿ Ž â x
'ww  ÿÀŽ â &r)   2   â x''ˆˆˆˆÿÀŽ â $~*ÿ   ÿ Ž â xww  ÿÀŽ â $~2ÿ   ÿ Ž â x''22ÿø� â $~)ÿ   ÿ Ž â t%ÏÝêö
    Ä t%*ÏÝêö    äx1u  6    ä t*25EYq    ä t
)5EYq
    ä x'Tv  ÿÀŽ ä x'ÿ   '  â &r&&      â r
   
    â &r	(#   	   â x,ww  ÿÀŽ â $~(ÿ   ÿ Ž â xÿ   ,  Ä r%%'   *    â $~*ÿ   ÿ Ž â x,,ˆˆˆˆÿÀŽ â $~2ÿ   ÿ Ž â xww  ÿÀŽ ä x,,22ÿø� ä t(ÏÝêö    ät(*5EYq    Ä t2ÏÝêö    ät**5EYq    ä x,Tv  ÿÀŽ ä x
 d  %    ä x,ÿ   ,  ä r.ÿ   .  Ä rÿ     ä rÿ     â &r!   &   â x

1u  %    ä r   (    ä r*,   *    ä rÿ     â $x.   ÿ Ž â x   ÿÀŽ â $x   ÿ Ž â x   ÿÀŽ â &r

      â x d  *    Ä x0   âŽ ä x*1u  *    â &r
#   
   â x0   âŽ â &r	-   	   âxàÿÿÿÿàÿ â &r
*-   
   â xàÿÿÿÿàÿ â 0r/ÿ/    A   ä $r	   ÿŽ Ä $r
   ÿŽ ä 0r
ÿ    A   ä 0r	ÿ0    A   â Er        ä Er        â  r//
     A  â  r	
	     A  â $rÿÿ    Ä #r/      â #r	      â G‰  Ðñÿÿÿÿƒ ê Ay        € ê ‰      " $x   Ž Æ ‰      d z  p  p`ð È x    pDp  â !r       â !r       È/ ‰      ( ‰      b !r       â !r       È/ ‰ €    ( ‰ €    b !r       â !r       È/ ‰ @    ( ‰	 @    b !r       â !r		       È/ ‰      h  ‰	      ¢  M	        € ê y      &   â $x   ÿ Ž â !r       â/ $z   s  Ž â !r		       â_ $tÿ   ÿ Ž â x    p"ò Ä r ÿ   p"ð Ö $ÿÿ   	 Ž ä $z  v   Ž È� %v  j   Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      "   $ z  w  p`ð Ú M	        € ê y      &   " $tÿ   ÿ Ž â ¹z  F   
   â $z x  Ž È %v \   Ž Ê �y     ™ b ¹z  l      â Ey  À    € â ™x?    â y      !   b $r
ÿÿ   ÿ Ž â x      ÿ  â $rÿÿ   ÿ Ž â x      ÿ  â xÿ     È/ |    p`ð Ú G	       € ê $rÿÿ   ÿ Ž â x	   ÿÀŽ â$x   ÿ Ž â Ey À    € â 'z m   Ž Æ x   ÿÀŽ â $x   ÿ Ž â zÿ n    â $vÿ X  ÿ Ž ä $vÿ Y  ÿ Ž â x   ÿÀŽ â $rÿÿ   
Ž ä $vÿ `  ÿ Ž ä $z o  Ž Ä $z r  ÿŽ ä $z u  ÿŽ ä y      %   " $vÿ a  ÿ Ž ä %x$   ÿ Ž È $rÿÿ   ÿ Ž ä %x$    Ž È $x   ÿ Ž ä %x	$    Ž â x	ÿ   	  Ê %x$   Ž È $x   ÿ Ž â z
 Z  ÿàñ â $z  t  Ž Æx

   ÿàó ä x   ÿàÿ äzÿ [  þ  â $z q  Ž ä $z q  Ž ä $x   ÿ Ž ä $z  t  Ž Ä %xˆ   ÿŽ â xp   ÿÀŽ Æ %xˆ   ÿŽ â r   ÿàó äx 	   þŽ är   ÿàñ â $r$ÿÿ   Ž  â x	   þŽ Æ $r"ÿÿ     ä r   ÿàñ Ê $x	   "  Ê �y5   ™ ¨  �y4   ™ è  �y3   ™ " $rÿÿ   
 Ž ä $rÿÿ    Ž â r   ÿàñ â �y.   ™ h �y!ôÿÿ ™ â $x   $  Æ �y#   ™ è �y%øÿÿ ™ ( �y/   ™ ( �y+   ™ ( �y&   ™ ( �y'üÿÿ ™ ( �y(   ™ ( �y,   ™ " r   ÿàñ Æ �y)    ™ ( �y-   ™ b $x	     âr   ÿàñ är   ÿàõ â �y*   ™ b r
    ÿàó â $x     Ä $x    â�y	    ‘ b $x   Ž  Æ �y   • h# �y1   • h �y
    ‘ h �yðÿÿ • h �y0    • h �y2    • b ‚x �˜­¿     Ä ‚x &     â $~7ÿ   ÿ Ž ä $~9ÿ   ÿ Ž ä $~;ÿ   ÿ Ž ä $~=ÿ   ÿ Ž ä $~8ÿ   ÿ Ž ä $~:ÿ   ÿ Ž Ä $~<ÿ   ÿ Ž â z ÿ `  pPð È z ÿ a   Sð ä x
ÿ   5  äO x5ww  ÿÀŽ äx55ˆˆˆˆÿÀŽ ä x
ww  ÿÀŽ ä/ x5522ÿø� Ä t
7ÏÝêö    ät95EYq    ä x5Tv  ÿÀŽ ä t;ÏÝêö    ät=5EYq    ä x5ÿ   5  Ä r
       ä xÿ   4  ä� x
4ww  ÿÀŽ ä r5       â $~5ÿ   ÿ Ž â x44ˆˆˆˆÿÀŽ â $~ÿ   ÿ Ž â xww  ÿÀŽ Ä x4422ÿø� ä t55EYq
    ä t57ÏÝêö    ät695EYq    ä tÏÝêö
    ä x4Tv  ÿÀŽ Ä x4ÿ   4  ä x
 d      äx1u      ä r       ä xÿ   3  äx3ww  ÿÀŽ Ä x33ˆˆˆˆÿÀŽ ä r54   6    â $~6ÿ   ÿ Ž â x5ww  ÿÀŽ ä x3322ÿø� ä t7ÏÝêö    ä t66ÏÝêö5    Ä t785EYq5    ä x53Tv  ÿÀŽ ä x3ÿ   3  ä t495EYq    â &r

!   ÿ   â x d      är663   7    â $~3ÿ   ÿ Ž â x1u      Ä r5   4    â $~5ÿ   ÿ Ž â x.ww  ÿÀŽ â&r
#   
   Æ t33ÏÝêö    â&r4%   
   â t55EYq    â $~8ÿ   ÿ Ž â x5ÿ   /  ä x/ˆˆˆˆÿÀŽ â$~
ÿ   ÿ Ž â x//ww  ÿÀŽ ä x95ww  ÿÀŽ Ä x22ÿø� ä t58ÏÝêö/    äx d  6    ät8:5EYq/    ä x1u  6    â $~6ÿ   ÿ Ž â t/
ÏÝêö9    Ä x7Tv  ÿÀŽ ä x
.ˆˆˆˆÿÀŽ ä x.ÿ   .  ä r557   8    ä x

22ÿø� ä t665EYq9    Ä x7ÿ     ä x.ww  ÿÀŽ ä x.
Tv  ÿÀŽ ä r6/7   6    â &r&   4   â x7ÿ   +  â $~:ÿ   ÿ Ž â r3.       Ä x8+ˆˆˆˆÿÀŽ ä x45 d  6    äx.51u  6    â $~6ÿ   ÿ Ž â x++ww  ÿÀŽ â $~5ÿ   ÿ Ž â x77ww  ÿÀŽ ä x8822ÿø� â $~9ÿ   ÿ Ž â t3:ÏÝêö+    â$~;ÿ   ÿ Ž â t:<5EYq+    Ä t+6ÏÝêö7    äx68Tv  ÿÀŽ ä t<55EYq7    ä x8ÿ   8  ä t/9ÏÝêö    ät;5EYq    Ä x
ÿ   
  â &r44!   ÿ   â r336   :    ä r8+8   <    ä r//
       â &r
'      â x3 d  8    â &r..#   4   â x d  /    â&r(   
   â x1u  /    â &r
%   .   â xÿ   ,  Ä x,ww  ÿÀŽ â$~+ÿ   ÿ Ž â x,,ˆˆˆˆÿÀŽ â $~/ÿ   ÿ Ž â x.ww  ÿÀŽ â $~5ÿ   ÿ Ž â x,,22ÿø� â $~7ÿ   ÿ Ž â t+ÏÝêö    ät+/5EYq    Ä x,Tv  ÿÀŽ ä t/5ÏÝêö.    äx331u  8    ä t.75EYq.    ä x,ÿ   ,  â &r)      â r   +    Ä x+ÿ   -  â &r
3&   
   â r//,   .    â $~.ÿ   ÿ Ž â x-ww  ÿÀŽ â$~3ÿ   ÿ Ž â x--ˆˆˆˆÿÀŽ â $~5ÿ   ÿ Ž â x,+ww  ÿÀŽ â $~ÿ   ÿ Ž â x--22ÿø� Ä t+.5EYq    ät.3ÏÝêö,    ät355EYq,    ä x,-Tv  ÿÀŽ ä tÏÝêö    ä x-ÿ   -  Ä x d  /    ä r,   +    ä r-.-   3    ä rÿ     är1ÿ   1  â &r
'   
   â x/1u  /    Ä r	ÿ   	  ärÿ     ä x d  -    â$x,1   ÿ Ž â x-1u  -    â $x   ÿ Ž â x		   ÿÀŽ â &r
/(   
   â x   ÿÀŽ Ä x	0   	âŽ â &r
)   
   â x,0   âŽ â &r*      âx		àÿÿÿÿàÿ â &r
-*   
   â xàÿÿÿÿàÿ â 0rÿ    A   ä 0rÿ0    A   ä 0rÿ2    A   Ä $r	   ÿŽ ä $r

   ÿŽ â  r     A  â  r     A  â Er+        ä Er 
       â G‰  p    € ê r   ÿàó Ê $x   "Ž  Ê �y1    ¨ �y.    è �y-    " r   ÿàó ä r
   ÿàõ â �y,    d $x	   Ž  Ä $x    È �y	     è  �y/
    è �y0
     " $~3ÿ   ÿ Ž ä $~5ÿ   ÿ Ž Ä $~2ÿ   ÿ Ž ä $~4ÿ   ÿ Ž ä $~6ÿ   ÿ Ž ä $~8ÿ   ÿ Ž â xÿ   1  äO x1ww  ÿÀŽ äx11ˆˆˆˆÿÀŽ Ä xww  ÿÀŽ ä x1122ÿø� ä t3ÏÝêö    ät55EYq    ä/ t2ÏÝêö    ä x
1Tv  ÿÀŽ Ä t45EYq    ä x1ÿ   1  ä r
       ä r1       ä xÿ   .  ä� x.ww  ÿÀŽ Ä x..ˆˆˆˆÿÀŽ ä x
 d      äx1u      ä xww  ÿÀŽ â &r

!   ÿ   â x..22ÿø� ä t2ÏÝêö    â &r#   
   â t16ÏÝêö    Ä t285EYq    ä t45EYq    ä x.Tv  ÿÀŽ ä x.ÿ   .  ä r
       â $~ÿ   ÿ Ž â xÿ   -  Är11.   2    â $~2ÿ   ÿ Ž â x-ww  ÿÀŽ äx--ˆˆˆˆÿÀŽ ä x.ww  ÿÀŽ ä x--22ÿø� ä t25EYq    Ä t23ÏÝêö.    ät355EYq.    ä x.-Tv  ÿÀŽ ä tÏÝêö    ä x-ÿ   -  ä x
 d  1    Ä r.       ä r-2-   3    â &r%      â x1
1u  1    â $~3ÿ   ÿ Ž â x
 d  -    äx1u  -    â &r1&      â x,ˆˆˆˆÿÀŽ â$~-ÿ   ÿ Ž â xÿ   ,  â $~1ÿ   ÿ Ž â x,,ww  ÿÀŽ â &r
'      â x22ÿø� Ä x.ww  ÿÀŽ â &r(      â t-ÏÝêö,    ät,15EYq,    ä x-Tv  ÿÀŽ ä t13ÏÝêö.    ät.55EYq.    Ä xÿ     ä r/ÿ   /  ä r	ÿ   	  ä r-   ,    â $x//   ÿ Ž â r1   .    ä x
	   ÿÀŽ Ä x	 d      äx1u      ä x
/0   
âŽ â &r	)      Æ x	
àÿÿÿÿàÿ â &r*      È $r	   ÿŽ ä 0r	ÿ0    A   Æ Er
        ä  r	     A  È #r
      â #r+      â #r      â G‰  À    € ê r   ÿàñ Ê $x   $  Ê �y-    ¨ �y    è �y    " r   ÿàó Æ �y    b r
    ÿàñ â $x	   Ž  È $x     â �y+    è  �y,     è  �y
     " $~/ÿ   ÿ Ž Ä $~3ÿ   ÿ Ž ä $~1ÿ   ÿ Ž ä $~5ÿ   ÿ Ž ä $~2ÿ   ÿ Ž ä $~4ÿ   ÿ Ž ä $~0ÿ   ÿ Ž â xÿ   -  ÄO x-ww  ÿÀŽ äx--ˆˆˆˆÿÀŽ ä x.ww  ÿÀŽ ä x--22ÿø� ä t/ÏÝêö    ä t	3ÏÝêö.    Ä x-Tv  ÿÀŽ ä t15EYq    ä t.55EYq.    ä x-ÿ   -  ä r       ä r		-   .    Ä xÿ     ä� x
 d  	    ä/x1u  	    ä x	ww  ÿÀŽ â$~.ÿ   ÿ Ž â xˆˆˆˆÿÀŽ ä xww  ÿÀŽ Ä x22ÿø� â &r

!   ÿ   â t.ÏÝêö	    ä t!2ÏÝêö    ät.45EYq    ä xTv  ÿÀŽ ä xÿ     Ä t	05EYq	    â &r#   
   â r.!   .    â $~ÿ   ÿ Ž â r	   	    ä x
ww  ÿÀŽ äxÿ     â $~!ÿ   ÿ Ž â xˆˆˆˆÿÀŽ â $~0ÿ   ÿ Ž â tÏÝêö
    â $~2ÿ   ÿ Ž â xww  ÿÀŽ Ä x22ÿø� ä t!5EYq
    ä t!0ÏÝêö    äx
	 d  .    ä t025EYq    ä xTv  ÿÀŽ Ä xÿ     â &r
%      â x		1u  .    ä r       ä r!!   0    â &r	&      â x
ˆˆˆˆÿÀŽ äx	 d  !    Ä x!1u  !    ä xÿ     â $~&ÿ   ÿ Ž â xww  ÿÀŽ â $~ÿ   ÿ Ž â x

22ÿø� â $~.ÿ   ÿ Ž â xww  ÿÀŽ â $~#ÿ   ÿ Ž â t&ÏÝêö    Ä x
Tv  ÿÀŽ ä tÏÝêö    ät.5EYq    ä t#5EYq    ä x
ÿ   
  ä r+ÿ   +  â &r	'      â r       Ä rÿ     ä r
       â $x
+   ÿ Ž â x   ÿÀŽ â &r!(      â x	 d      äx1u      ä x
0   âŽ â &r	)      Æ xàÿÿÿÿàÿ â &r*      ä 0r	ÿ,    A   ä $r   ÿŽ Æ  r		     A  ä Er        Ê #r	      â x   ÿàÿ ä x   ÿàó ä |    p`ð ä x   ÿà÷ â $rÿÿ   Ž  â x
€  ÿàõ Æ $rÿÿ   Ž ä $rÿÿ    È D	       À â Gy  Ðãÿÿÿÿƒ ê Ay       € ê $r
ÿÿ    Ž ä $rÿÿ    Ž ä Ay        € ê z ÿ `  pPð È z ÿ a   Sð Ú G	  P    € ê ‰      ( ‰      b !r       â !r       È/ ‰      ( ‰      b !r       â !r       È/ ‰ €    ( ‰ €    b !r       â !r       È/ ‰ @    ( ‰	 @    b !r       â !r		       È/ ‰      ( ‰	      b !r       â !r	       â/ Gy  €    € ê ‰      ( ‰      h ‰      ¨ ‰      â !r       â !r       È/ ‰      " !r       ÆO ‰      b !r       Æ� ‰      ¨ ‰
      â !r       â !r       â/ !r	       ÈO ‰ €    " !r
       Æ� ‰ €    h ‰
	 €    ¨ ‰ €    â !r       â !r       È/ ‰ @    " !r
	
       ÆO ‰ @    b !r       Æ� ‰
 @    ¨ ‰ @    â !r       â !r       È/ ‰      " !r
       ÆO ‰      b !r	       Æ� ‰      ¨ ‰	      â !r       â !r
       â/ !r       âO !r	       â� y      !   ( y      %   $ $x   Ž Ê z  p  p`ð È x    pDp  Ú M	        € ê z ÿ ^  pPð â $z  v  Ž Æz ÿ _   Sð Ú $ÿ   ÿ Ž È % ^   Ž Ì �	     ¢ $x    ÿ Ž Ê r  ÿ   p"ò äx     p"ö äx     p"ø äx     p"ô Î $ ÿÿ    Ž â z ÿ `  pPò â $2 ÿÿ    Ž ä $B ÿÿ   
 Ž â z ÿ a  Sò â $" ÿÿ    Ž È !         ÐO G™       € ê z ÿ b  pPð È z ÿ c   Sð Ú $ÿ   ÿ Ž È % b   Ž Ì �	     ¢ $x   ÿ Ž Ê x øÿÿÿp"ò äx üÿÿÿp"ô är ÿ   p"ö äx    p"ø Î $ÿÿ    Ž ä $"ÿÿ    Ž ä $vÿ h  ÿ Ž ä $2ÿÿ   
 Ž ä $Bÿÿ    Ž â x    pBò Æ !       ÔO G  P    € ê x    pRð Ú G‰  �     € ê x    pRð Ú G	  ð    € ê  x;ª¸¿  A  Ì s        $ !t  €?    Ì s        $  r     A  È  r       A  â Gy  P    € ê  x'7=  A  â x*BL?  A  Æ #t  €?   â  x   ?  A  Æ  r     A  Ì s     $   $ !t  €?    È  r     A  È  r       A  â Gy  °    € ê x    pRð Ú G‰  à     € ê x    pRð Ú G	        € ê 	z i    � ä 	z   i    � Æ  x;ª¸¿  A  â 	z   i €  � Ê s        $ !t  €?    Ì s        $  r     A  È  r       A  â Gy  Ð     € ê  r       A  â Gy  °     € ê 	x  à@  � ä 	x    à@  � Æ  x#ÛÙ?  A  â 	x    àÀ  � Æ  x;ª¸¿  A  ä !t   €?    È s        $ !t  €?    Ì s        $  r      A  È  r       A  ä y      "   " $tÿ   ÿ Ž Æ y      &   b $z s  Ž È $z v  Ž È/ %v j   Ž Ê †y      â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      "   " ¹z  l      â Ey  À    € â ™x?    â y       !   b ¹z  F   
   â r^ ÿ       â x      ÿ  â y_      %   ¢ r ÿ       â x      ÿ  â x      ÿ  â x
      ÿ  â x      ÿ  â x      ÿ  â x      ÿ  â r ÿ   pRð Ä r ÿ   pRò är    ÿ(Ž È/ r   ÿ� ä x__   ÿ   äO x`ÿ     È | `   pbô Ú G)  `    € ê y      '   " xþÿÿÿÿÀŽ ä r ÿ       â y      &   b r ÿ       ä r  €ÿàÿ ä x        ä xb   ÿ   â$x   ÿ Ž â r^ ÿ       Ä z$` q  ÿàÿ â %xbb   ÿŽ â x&ÿ     È x!b   ÿüŽ â 'z y   Ž È 'z s   Ž â/ zÿ z    Æ $z_ p  `Ž â zÿ t    â $z }  ÿŽ ä $z |  Ž â x_   ÿàÿ â $z |  `Ž ä $z v  Ž ä$z v  Ž Ä $z w  ÿŽ ä $z p  Ž ä %v X  Ž â r   ÿàõ Æ %x   bŽ â x   ÿàù ä rÿÿ   ÿä â %v X  Ž â z X  ÿà÷ ä z  q      ä rÿ   ÿä ä x   ÿàù â $x   `Ž âr`   ÿŽ â$x   `Ž âr`   ÿŽ â$x    `Ž âzÿ Y  d~ â $x"   `Ž â x   &  Ä x   ÿ   ä r\        ä r]        Ê �yT\úÿÿ • ¨ �y\üÿÿ • â r(!   ÿàõ Æ �yX\þÿÿ • " r)c   ÿä Æ �yU\    • h �yZ(þÿÿ • ( �yY(    • " r9`   ÿàõ ä xf $       â �y#(   • " r%ÿ   ÿä Ä r5$   ÿà÷ â �y&(   • " %v89 Z  f Ž ä rÿ   ÿäÿ ä $x%%$   ÿŽ â r68   ÿàõ â %v45 Z  f Ž Æ r99%   ÿàÿ â $x'$   ÿŽ â r24   ÿà÷ â �y(    • b r79   ÿä ä r55'   ÿàÿ â �y+8    • b r;   ÿàõ ä r35   ÿäÿ â �y'6   ™ b $rÿÿ    Ž â r<ÿ   ÿä Ä �y)6   ™ b rA   ÿà÷ Æ �y*2   ™ b %v:; Z  f Ž Æ �y%üÿÿ • b $x9<$   ÿŽ Æ �y,6   ™ h �y-2   ™ b rÿ   ÿäÿ Æ/ �y.6   ™ b %v@A Z  f Ž â r>:   ÿàõ ä �y/2   ™ b $x=$   ÿŽ â r;;9   ÿàÿ ä �y12   ™ b r<@   ÿà÷ ä r?;   ÿä â �y04    • b rAA=   ÿàÿ Ä rE   ÿàõ â �y6>   ™ b r=A   ÿäÿ Æ �y3>   ™ b! rK   ÿà÷ ä rBÿ   ÿä â�y4<   ™ b rÿ   ÿäÿ â %vDE Z  f Ž ä �y5@    • h �y8>   ™ h �y7<   ™ b $xAB$   ÿŽ Æ �y2:    • b %vJK Z  f Ž â rHD   ÿàõ ä �y9<   ™ b $xC$   ÿŽ â rEEA   ÿàÿ ä �y:>   ™ b rFJ   ÿà÷ ä rKKC   ÿàÿ â �y;<   ™ b rIE   ÿä Ä rk    ÿàõ â$xGK   Ž â rg"   ÿà÷ â �y@H   ™ h �y=H   ™ b# rLÿ   ÿä ärÿ   ÿäÿ â �y>F   ™ b %vjk Z  f Ž Æ �y?J    • b %vfg Z  f Ž Æ �yBH   ™ b $xM$   ÿŽ Æ �yAF   ™ b $xKL$   ÿŽ Æ �y<D    • b rhj   ÿàõ ä rkkK   ÿàÿ â �yCF   ™ b rdf   ÿà÷ ä rggM   ÿàÿ â �yLj    • h �yDH   ™ b rik   ÿä Æ �yEF   ™ b reg   ÿäÿ Æ �yJh   ™ h �yHh   ™ h/ �yFh   ™ h �yGd   ™ h �yId   ™ h �yKd   ™ h �yMh   ™ h �yOd   ™ h �yNf    • b ‚x �˜­¿     Ä ‚x &     â |V       ä |P       ä |       ä |R       ä |d       ä xQTww  ÿÀŽ ÄO xTTT      È� xTTˆˆˆˆÿÀŽ ä xww  ÿÀŽ ä xTT22ÿø� ä tVVÏÝêöQ    ätPP5EYqQ    ä xSTTv  ÿÀŽ Ä tQÏÝêö    ätRR5EYq    ä xTÿ   T  ä rSVS   P    ä rRQT   R    â $~Tÿ   ÿ Ž â |V       Ä xWXww  ÿÀŽ äx[XT  U    äx\@  ÿàõ ä tQVÏÝêöW    ä x[[ˆˆˆˆÿÀŽ ä rÿ]   ÿä Ä tTT5EYqW    ä |V       ä xUUww  ÿÀŽ ä |\       ä xWZww  ÿÀŽ ä |X       Ä x]ZT  Y    ä xPS d  R    äxYYww  ÿÀŽ ä |Z       ä xRS1u  R    ä x[[22ÿø� Ä tSdÏÝêöU    ätVV5EYqU    ä tU\ÏÝêöW    ätXX5EYqW    ä x]]ˆˆˆˆÿÀŽ ä tW\ÏÝêöY    Ä tZZ5EYqY    ä xY[Tv  ÿÀŽ ä x]]22ÿø� ä rQQY   T    ä xY#ww  ÿÀŽ äx##T  &    Ä x[ÿ   [  ä x\]Tv  ÿÀŽ ä x]ÿ   ]  ä x##ˆˆˆˆÿÀŽ ä rVS[   V    ä rZW]   Z    Ä |S       ä x&&ww  ÿÀŽ ä rUU\   X    ä |T       ä |W       ä |X       Ä x##22ÿø� ä tSSÏÝêö&    ätTTÏÝêöY    ät&W5EYq&    ä tYX5EYqY    ä xW#Tv  ÿÀŽ Ä xXQ d  V    äx#ÿ   #  ä xVQ1u  V    ä xQU d  Z    ä rTTW   Y    ä r#S#   &    â &r&P'   ÿ   Ä &rSQ'   ÿ   â x'T d  #    äxTT1u  #    â &r#R)   &   ä &r&P*   ÿ   ä &r#X,   #   ä &r&R-   &   â xZU1u  Z    â &r#V.   #   Ä &r&X/   &   ä 0r%ÿ%    A   ä 0r+ÿ+    A   â Er# #       â &rSZ)   S   ä &r)V1   &   â  r&%+     A  â &r*Q*   ÿ   Æ #r&#      â &r#P3   ÿ   ä &r*Z-   *   ä &r#R6   #   ä &r&P4   ÿ   ä &r*'/   *   ä 0r0ÿ0    A   Ä &r#X8   #   ä &r&R7   &   â Er) )       â &r1T1   *   â  r*%0     A  â &r#V:   #   ä &r&X9   &   ä 0r2ÿ2    A   â #r*)      â &r*V;   &   â Er# #       â &r3Q3   ÿ   â  r&%2     A  Æ &r3Z6   3   â #r&#      â &r#P=   ÿ   ä&r=Q=   ÿ   ä &r3'8   3   ä &r#R@   #   ä &r&P>   ÿ   Ä &r=Z@   =   ä 0r5ÿ5    A   ä &r3T:   3   ä &r#XB   #   ä 0r(ÿ(    A   ä &r&RA   &   Ä &r='B   =   âEr* *       â &rS',   S   â  r,%5     A  â &r#VD   #   â Er3 3       â &r&XC   &   â  r)(2     A  â 0r<ÿ<    A   ä &r=TD   =   â #r,*      â &r*VE   &   â Er# #       â #r)3      â  r&%<     A  âEr= =       â  r)(<     A  Ä #r	&#   	   â &r#PF   ÿ   â#r)=      â &r4Q4   ÿ   ä&r>Q>   ÿ   ä&r)QF   ÿ   ä&rPPG   ÿ   ä&rGQG   ÿ   â x``    ÿàÿ â &r4Z7   4   Ä &r>ZA   >   ä&r)ZH   )   ä&r#RH   #   ä&rPRI   P   ä&rZZI   G   â | `   pbô â &r.T.   S   Ä &r4'9   4   ä&r>'C   >   ä&r#XJ   #   ä&r)'J   )   ä&rPXK   P   ä&rZ'K   Z   â Er. .       â &r4T;   4   â  r++(     A  â 0r?ÿ?    A   Ä &r>TE   >   ä 0rLÿL    A   ä &r#VM   #   ä&r)TM   )   ä &rPVO   P   ä&rZTO   Z   Ä 0rNÿN    A   â Er1 1       â #r+.      â Er4 4       â  r,%?     A  âEr* *       â  r&%L     A  âEr> >       â  r-(0     A  âEr# #       â  r5(5     A  âEr) )       â  r?(?     A  âErP P       â  r'(L     A  âErZ Z       â  r%%N     A  â r+(N     A  â x@  ÿà÷ Ä x@  ÿàù â #r-1      â #r
54   
   â #r,*      â #r?>      â #r&#      â #r')      â #r%P      â #r^+Z   ^   â x$$    ÿàÿ ä x    ÿàÿ Ä x    ÿàÿ ä x    ÿàÿ ä x    ÿàÿ ä x      ÿàÿ ä x""    ÿàÿ ä rÿ   ÿäÿ Ä rÿ   ÿä â D)       À â Gy  €ëÿÿÿÿƒ ê Ay        € ê Ey  0    € â G‰      € ê r    ÿHŽ Ê ˆs  øÿ H   è ˆs €øÿ H   è ˆs  ùÿ H   è ˆs €ùÿ H   è ˆs  úÿ H   è ˆs €úÿ H   è ˆs  ûÿ H   è ˆs 
€ûÿ H   è ˆs 	 üÿ H   è ˆs €üÿ H   è ˆs  ýÿ H   è ˆs €ýÿ H   è ˆs  þÿ H   è ˆs €þÿ H   è ˆs  ÿÿ H   è ˆs ^€ÿÿ H   ä Ay        € ê {          ì M        € ê „y      H   b z _ r  p`ð Æ „y  €   H   ¢ r  ÿ   p"p Æ „y     H   è „y  €  H   ( y      &   h „y  €  H   ( „y     H   h „y      H   h „y&     H   è „y#  €  H   â !r       Æ/ „y%  €  H   â !r       Æ_ ‰      " !r       Æ� y      '   b !r       Æ‰      ¨ ‰      â !r       Æ „y*     H   â $vÿ r  ÿ Ž â !r       ä„y.     H   â !r!        â !r       â $z x  _Ž â „y'  €  H   â !r&&	       â $x   ÿŽ â x__   ÿàÿ â ‰      " $z ~  Ž â/ !r       ÄO „y+     H   ä r   ÿ† â!r       â� „y)  €  H   â 
 r  ÿàõ â$x   ÿŽ â r$ÿÿ   ÿäÿ â ‰      b z j  ÿ† ä(ÿÿ   ÿä â ‰      ¢ z k  $� â !r$#
       â 
 j  ÿ„ â ‰      â    ÿàó Ä 
 k  ( â !r       â ‰"!      " !r(%       â !r**       â  ÿÿ   ÿäÿ  â ‰ €    " !r..       â 
 j  ÿˆ ä ‰      b r   ÿ„ â !r,'       Ä ‰
      b !r       â/ 
 k    â !r       âO „y7  €  H   â rÿÿ   ÿä â !r2+       ä ‰'$      b !r        â�    ÿàù â !r0)       â ‰+(      ¢ z j  ÿ„ â !r"!"       â z _ r  p`ò â !r       â‰)&      " z k   â !r       Ä‰	 @    â $x   ÿŽ â ÿÿ   ÿä â !r
       â ‰-*      " 
 j  ÿˆ ä r   ÿà÷ â ‰/,      b x     p"ò ä rÿÿ   ÿäÿ â ‰      b z j  ÿ† â!r'$'       â/ !r+(+       âO ‰1.      b z k  � Ä 
 k   â ‰30      ¢ !r)&)       â    ÿàÿ â !r77^       â !r	       â� ‰52      "    ÿàÿ â !r-*-       ä‰ €    â !r/,/       Æ‰      " !r       Æ ‰#       b !r1.1       Æ/ ‰      b !r303       ÆO ‰%"      ¢ !r525       Æ ‰('      " !r       Æ� ‰*)      â $x   ÿŽ â   ÿàÿ â !r!       â‰,+      $ 	   ÿà÷ â !r# #       â‰ €    b     ÿàù â !r9       Ä/ ‰.-      b    ÿàõ â !r%"%       äO ‰0/      ¢ ÿÿ   ÿä ä  j  ÿ„ â ‰  €    ¢ ÿÿ   ÿä â !r('(       â 
  j  ÿˆ â‰21      " ÿÿ   ÿäÿ â !r*)*       â� 
	 j  ÿ† â ‰43      â  k   â !r,+,       â  k   â ‰65      " !r       â
		 k  � Ä ‰7      b !r.-.       â/    ÿàõ ä ‰ @    b !r0/0       âO x        ä ‰"! €    ¢ !r         Æ ‰$# €    ¢ !r212       Æ ‰&% €    " !r434       â� !r656       ä‰'( €    â !r77       Æ‰)* €    " !r       Æ/ ‰+, €    b %v j   Ž â !r!"       äO ‰ @    ¢ !r#$       Æ ‰-. €    b $ÿÿ   ÿä â !r %&       ä ‰/0 €    "  j  ÿ„ Æ ‰!  @    " !r'('       â�  k  $ ä ‰12 €    â !r)*)       Æ‰34 €    " !r+,+       Æ/ ‰56 €    b !r       ÆO ‰87      ¢ !r-.-       Æ‰      b !r/0/       Æ ‰" @    " !r! !       Æ ‰# @    " !r121       Æ� ‰%  @    â !r343       Æ‰ ' @    " !r565       â/ !r878       äO ‰$) @    b !r;       Æ‰&+ @    ¢ !r""       Æ ‰      â !r##       Æ ‰(- @    " !r% %       Æ� ‰*/ @    â !r '        Æ‰!      ( ‰,1 @    b !r$)$       Æ/ ‰.3 @    b !r&+&       ÆO ‰78 €    ¨ ‰05 @    ¢ !r(-(       Æ ‰"      " !r*/*       Æ� ‰ #      â !r'!       Æ†	 9    â	 !r,1,       Æ‰%      b !r.3.       Æ/ † ;   â !r787       âO !r9       ä‰       b !r050       â !r)"       ä ‰$      " !r+#        Æ� ‰&      ¨ †	 9    â !r%%       Æ‰!(      è ‰*      ( † 
'    â !r        Æ/ ‰,      b !r$       Æ ‰#.      " !r&       ÆO ‰ 7 @    ¨ ‰0      b !r!(!       Æ� †	 )    â !r*       Æ† +   è †	 %    â !r,       â/ !r#.#       ä †     â !r 7        ÆO †	     â !r0       Æ†    è †	 !    è †     è †	     è † #   è ‰       ( †	     â M™        € ê x   ÿàÿ â !r        Æ r   ÿàñ È rÿÿ   ÿä  ä� z j  ÿ€ È z k    Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y       "   " ¹z  l      â Ey       € â ™x?    â y      !   b ¹z  F   
   â x      ÿ  â r ÿ       â y!      %   ¢ r1 ÿ       â x      ÿ  â x      ÿ  â x      ÿ  â x      ÿ  â r  ÿ       â $r(ÿÿ   ÿ Ž â r  ÿ   pRð Ä r  ÿ   pRò är    ÿ(Ž È/ r   ÿ� ä x!!   ÿ   äO xÿ     È |    pbô Ú G)  À    € ê y	      '   " r ÿ       â $z! p  Ž â xþÿÿÿÿÀŽ â y      &   b r
 ÿ       â $t+ÿ   ÿ Ž â r ÿ       ä z" q  ÿàÿ â 'z	 y   Ž â r  €ÿàÿ Æ 'z s  
 Ž â/ zÿ z    ä x   ÿ   â$z%	 }  ÿŽ â zÿ t    â $z |  Ž â x!   ÿàÿ â $z |  Ž â x   ÿ   â %x   ÿŽ â z	  q      Ä x-ÿ     â $z v  Ž â x/   ÿ   â $z v  Ž â x-   -  â $z* p  Ž â x$   ÿüŽ â %v X  +Ž È %x   Ž â x&   ÿà÷ Æ $z
 w  ÿŽ â z X  ÿàù â %v** X  +Ž â r'ÿ   ÿäÿ ä r%
%   ÿàõ â $x0	   Ž âx   ÿà÷ â $x#	   Ž âr) *       â $x,	   Ž âzÿ Y  „þ â $x.	   Ž âr*	   ÿŽ Ä r;ÿÿ   ÿä ä �yúÿÿ • ¨ �y6üÿÿ • â r)$   ÿàõ Æ �y8þÿÿ • " r	+   ÿä Æ �y3    • h �y2þÿÿ • ( �y5    • " r%   ÿàõ ä x@ $       â �y7   • $ $rÿÿ   ; Ä �y4   • " %v Z  @ Ž È $x$   ÿŽ â r/   ÿàõ È r   ÿàÿ ä r &       ä r	 '       â �y>    • b r-   ÿä ä r"%   ÿàõ â�y9üÿÿ • b r0%   ÿà÷ Ä rÿ;   ÿä â�y<   ™ b r
ÿ;   ÿäÿ â %v Z  @ Ž ä �y=   ™ b r )       â/ $r	ÿÿ   + Ž ä �y?   ™ b $x$   ÿŽ Æ �y:    • d r   ÿàÿ ä �yA   ™ h �yD    • b %v Z  @ Ž È $x
$   ÿŽ â r
/   ÿàõ È r		   ÿàÿ ä r-   ÿä ä r/   ÿà÷ â �yE    • b r#%   ÿàõ Æ �yB
   ™ b $x	   -Ž â r*%   ÿà÷ ä/ �yI
   ™ b rLÿ;   ÿä ärÿ;   ÿäÿ â �yH
   ™ b %v Z  @ Ž Æ �yJ
   ™ b $x	L$   ÿŽ ä %v Z  @ Ž â �yC   ™ f $x$   ÿŽ â r/   ÿàõ â �yF   ™ b r	   ÿàÿ Ä r/   ÿà÷ â �yG   ™ b r   ÿàÿ ä r-   ÿä â �yK   ™ d $x	   -Ž â r.%   ÿà÷ â/�yL   ™ h �yP   ™ b r,%   ÿàõ Ä r
ÿ;   ÿäÿ â�yN    • b rÿ;   ÿä Æ �yQ    • h �yS   ™ b %v Z  @ Ž Æ �yT   ™ b %v Z  @ Ž Æ/ �yM   ™ b $x$   ÿŽ ä $x
$   ÿŽ â r
/   ÿàõ â �yO   ™ b r   ÿàÿ ä r/   ÿà÷ â �yR   ™ b r   ÿàÿ Ä r-   ÿä â �yU   ™ d $x   -Ž ä �y
   ™ h �y   ™ h �yV
   ™ h �y@   ™ h �y	
   ™ h! �y   ™ h �yW
   ™ h �yX   ™ h �y    • h �y    • b	 ‚x &     Ä ‚x �˜­¿     â |       ä |Z       ä |\       ä |[       ä |       ä/ xYww  ÿÀŽ ÄO xT  6    È� xˆˆˆˆÿÀŽ ä x66ww  ÿÀŽ ä x22ÿø� ä t5EYq6    ä t
ZÏÝêöY    ä xTv  ÿÀŽ Ä tY\5EYqY    ä t[ÏÝêö6    ä xÿ     ä r

   Y    ä r       â $~6ÿ   ÿ Ž â |       Ä x8ww  ÿÀŽ äx
 d      äx
1u      ä t
ÏÝêö    ät65EYq    ä |6       Ä x88T  3    äx3ww  ÿÀŽ ä |Z       ä x2ww  ÿÀŽ äx22T  5    ä t66ÏÝêö    Ä x88ˆˆˆˆÿÀŽ ä tZ5EYq    ä xZ5ww  ÿÀŽ ä x52ˆˆˆˆÿÀŽ ä |3       ä x8822ÿø� Ä tÏÝêö    äxY522ÿø� â $~[ÿ   ÿ Ž â t35EYq    ä x58Tv  ÿÀŽ ä |3       ä r

5       Ä t23ÏÝêöZ    äx7ww  ÿÀŽ ä t3[5EYqZ    ä x77T  4    ä x8ÿ   8  ä xZYTv  ÿÀŽ Ä xYÿ   Y  ä x77ˆˆˆˆÿÀŽ ä r68       ä r22Y   3    ä |6       ä rZ       Ä |5       ä |Y       ä x44ww  ÿÀŽ ä |       ä x7722ÿø� ä t365EYq    Ä t65ÏÝêö4    ätYY5EYq4    ä tÏÝêö    ä x47Tv  ÿÀŽ ä x d  2    är54   3    â &r4<   ÿ   â x1u  2    Ä x2
 d      â&r3=   4   â x

1u      Æ &r32?   3   â x7ÿ   7  â 0r9ÿ9    A   ä &r3
A   3   â r667   Y    â 0r>ÿ>    A   Æ Er3 3       ä x5 d  6    â r9>     A  â x451u  6    â &r6B   ÿ   ä #r(3   (   ä &rI   6   È &r2H      ä &r<<   ÿ   ä &r
J      ä 0rDÿD    A   ä &r<=   <   â Er        â &rBB   ÿ   â  r69D     A  â &r<?   <   Ä &rII   B   â #r6      â &r3C   ÿ   ä&r<4A   <   ä &rL   ÿ   ä 0r:ÿ:    A   ä &rIH   I   Ä &r3F   3   âEr< <       â &rP      â  r5>:     A  â &rI4J   I   ä &r32G   3   ä&r2S      â #r 5<       â &r5
K   3   â ErI I       â  r3:D     A  â &r
T      Ä &r6M   ÿ   ä 0rNÿN    A   ä 0rEÿE    A   â #r3I      â &r3O   6   â Er        â  r69N     A  âEr5 5       â  r89E     A  Ä #r6      â &r   ÿ   â #r85      â &rCC   ÿ   ä&rLL   ÿ   ä&r8M   ÿ   ä&r   ÿ   ä &r   ÿ   Ä &r   ÿ   â x@  ÿàõ â &rV      â x    ÿàÿ â &rCF   C   ä&rLP   L   ä&rOO   8   ä&rV      Ä &r@      ä&r@      â rÿ   ÿä â &r2	      â |    pbô â &rCG   C   ä&rSS   L   ä&rOR   O   Ä &r	      ä &r32R   3   ä&r2      ä&r      ä &r
W      ä &rC4K   C   Ä &rS4T   S   ä&rO4U   O   ä&rW4W      ä 0rQÿQ    A   ä &r5
U   3   ä&r
X      Ä &r44X      ä 0rÿ    A   ä 0rÿ    A   â ErC C       â  r89Q     A  âErS S       â  rE:E     A  âEr5 5       â  r3:N     A  âErO O       â  rQ:Q     A  âEr        â  r9     A  âErW W       â  r:     A  âEr        â  r9     A  â Er4 4       â  r	:     A  â x))@  ÿà÷ Ä x&&@  ÿàù â #rEC      â #r3S      â #r85      â #rQO      â #r1   1   â #rW      â #r      â #r	4      â x""    ÿàÿ â $r+ÿÿ   +Ž â x00    ÿàÿ Ä x##    ÿàÿ ä x**    ÿàÿ ä x,,    ÿàÿ ä x..    ÿàÿ ä r'ÿ'   ÿä â D)       À â Gy   íÿÿÿÿƒ ê Ay        € ê Ey      € â G‰  ð     € ê $x À  Ž Ê ˆs ( ùÿ H   è ˆs  €ùÿ H   è ˆs  úÿ H   è ˆs €úÿ H   è ˆs  ûÿ H   è ˆs €ûÿ H   è ˆs  üÿ H   è ˆs €üÿ H   è ˆs  ýÿ H   è ˆs €ýÿ H   è ˆs 1 þÿ H   è ˆs €þÿ H   è ˆs  ÿÿ H   è ˆs €ÿÿ H   ä Ay        € ê {          ì M        € ê „y     H   " z ! r  p`ð Ä z-  r      â „y €  H   b r ÿ   p"p Æ „y €   H   ¢ $x+-   ÿŽ Æ „y    H   è „y    H   ( „y    H   h „y	 €  H   h „y €  H   h „y    H   è „y €  H   â !r(       Æ „y# €  H   â !r       Æ/ ‰       " !r        ÆO „y    H   b !r       Æ� „y"    H   â !r       Æ„y% €  H   â !r       Æy      '   ¢ !r		       Æ ‰      â !r       â !r        ä ‰      â !r       Æ y       &   " !r       Æ ‰      " !r #       Æ ‰      h ‰      b !r1       Æ/ ‰
      b !r       â� !r""       ä ‰	      â !r$%       â $z   x  !Ž ä ‰#      " !r       â$z  ~   Ž âO !r       â x!   ÿàÿ â !r       â‰      ¢ 
  r  ÿàõ â !r       Ä ‰      " z  r  p`ò â !r

       ä/ ‰!      b ÿÿ   ÿä â !r	       ä� ‰ €    â 
 j  ÿ„ ä x    p"ò â ‰      b 
 k   â !r##       â $x-   ÿŽ â ‰%       ( ‰'"      " !r       ÆO ‰      ¢ !r       Æ‰)$      " !r!!       Æ/ ‰	      b !r       â�     ÿà÷ ä ‰      â !r       âr-    ÿ„ ä ‰      b !r% %       â    ÿàÿ Ä ‰
      " !r'"'       â !r       äO ‰       ¢ !r)$)       Æ‰"      " !r	       Æ/ ‰$!      b 	 +   ÿàû â !r       ä� ‰&#      â ÿÿ   ÿäÿ â !r       ä‰ @    b $rÿÿ   ÿ â 
 j  ÿ† â !r
       â ‰ €    " r-    ÿˆ Ä z j  ÿ„ â ‰(%      " ÿÿ   ÿäÿ â !r         âO r.ÿÿ   ÿä â ‰*'      ¢ 
	 j  ÿŠ â !r""       â
 k  � â ‰ €    " !r$!$       â/ z
 j  ÿˆ Ä ‰ €    b z k   â !r&#&       â� 
		 k  � â ‰,)      â !r       â-   ÿàÿ ä ‰ €    b !r       â z k  . Ä ‰ €    " !r(%(       â +   ÿàÿ ä ‰ €    " !r*'*       âO /    ÿàù â !r       â‰#  €    ¤ 2ÿÿ   ÿä â !r       â/ ‰%" €    b $x-   ÿŽ â / j  ÿˆ â !r,),       â� ‰'$ €    â 0    ÿà÷ â !r       Ä‰)& €    " / k  2 â !r       ä ‰ @    " r    ÿàõ â !r       ä ‰+( €    b r.ÿÿ   ÿä ä z j  ÿ„ â ‰-* €    b !r# #       âO 1ÿÿ   ÿäÿ Ä ‰! @    ¢ !r%"%       â/ 0 j  ÿ† ä ‰ @    b !r'$'       â� z k  . ä ‰ @    â !r)&)       â0 k  1� Ä ‰/, €    " !r       â     ÿàõ ä ‰ @    " !r+(+       Æ‰      b !r-*-       Æ ‰ @    b !r!!       ÆO ‰ # @    ¢ !r       Æ/ ‰% @    b !r       Æ� ‰"' @    â !r/,/       â,ÿÿ   ÿä ä ‰$) @    " !r       Æ ‰      " !r1       âx        ä ‰&+ @    b !r       â  j  ÿ„ â %v  j   Ž â ‰(- @    b !r #        âO  k  , Ä ‰      ¢ !r%       Æ/ ‰*/ @    b !r"'"       Æ� ‰      â !r$)$       Æ‰       " !r3       Æ ‰      " !r&+&       Æ‰!      b !r(-(       Æ ‰       b !r)       ÆO ‰#      ¢ !r*/*       Æ/ ‰%"      b !r+       Æ� ‰$      â !r        Æ†	 1    â !r       Æ † 3   â !r!!       Æ‰'&      " !r        Æ ‰(      " !r##       ÆO ‰*      ¢ !r%"%       Æ/ †	 )    â !r$       Æ� † +    è †	 
    è † 
   â !r'&'       Æ †	 !    â !r(       Æ†     â !r*       ÆO †	 #    è † %   è †	     è † '    è †	     â M™        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y       "   " ¹z  l      â Ey      € â ™x?    â y      !   b ¹z  F   
   â x      ÿ  â $rÿÿ   ÿ Ž â y      %   ¢ r- ÿ       â x      ÿ  â x      ÿ  â r3 ÿ       â x      ÿ  â $r5ÿÿ   ÿ Ž â r  ÿ   pRð Ä r  ÿ   pRò är    ÿ(Ž È/ r   ÿ� ä x   ÿ   äO xÿ     È |    pbô Ú G)  À    € ê y      '   " r ÿ       â $t&ÿ   ÿ Ž â xþÿÿÿÿÀŽ â y	      &   b r ÿ       ä r ÿ       ä z, q  ÿàÿ â 'z y   Ž È $x   
Ž â zÿ z    â 'z	 s   Ž Æ/ x   ÿ   â $z p  Ž â zÿ t    â $z |  Ž ä$z |  Ž â x   ÿàÿ â %x   ÿŽ È $z v  Ž â x"   ÿüŽ â $z v  Ž ä %v$ X  &Ž È %x   Ž â x#$   ÿàù Æ $z! }  ÿŽ â z X  ÿà÷ â $z
	 w  ÿŽ â r$ÿ%   ÿä â $x   ÿ Ž â z  q      â $z p  Ž â r!
!   ÿàõ ä x   ÿàù â %v& X  &Ž â x)ÿ     Ä zÿ Y  d~ â $x    Ž âr   ÿŽ â$x(   Ž âx)   )  â$x*   Ž â x+   ÿ   ä rAÿÿ   ÿä ä �y0úÿÿ • ¨  �y%üÿÿ • â  r&"   ÿàõ Æ �y4þÿÿ • $ $x   ' ä �y/    • h �y.þÿÿ • ( �y1    • " r!   ÿàõ Ä x< $       â �y2   • " r,!   ÿà÷ ä rÿA   ÿä â�y   • " rÿA   ÿäÿ â %v
 Z  < Ž È $x$   ÿŽ ä $x	$   ÿŽ â r+
   ÿàõ â $x   Ž â r #       â/ %v Z  < Ž È $rÿÿ   $ Ž â r)   ÿä â �y;
    • b r+   ÿà÷ ä r	   ÿàÿ â �y7üÿÿ • b r!   ÿàõ ä r	)   ÿäÿ â �y6   ™ h �y9   ™ b r &       Ä r '       â �y:   ™ b rE !   ÿà÷ Æ �y8    • b %v Z  < Ž â rDÿA   ÿäÿ ä �y>   ™ h �y=   ™ b $rÿÿ   A Æ �y@   ™ b $x$   ÿŽ Æ/ �y?   ™ b %vE Z  < Ž â r   ÿàÿ ä �yC   ™ b $xED$   ÿŽ â r+   ÿàõ är
+   ÿà÷ â �yB    • b rE   ÿàÿ â $x   ) Ä �yD    • b r)   ÿäÿ ä r	*!   ÿà÷ â/�yE   ™ b r(!   ÿàõ Æ �yH   ™ b rÿA   ÿäÿ ärÿA   ÿä â �yF
   ™ b %v Z  < Ž Æ �yI    • b $x$   ÿŽ Æ �yK   ™ h �yG
   ™ b %v	 Z  < Ž Æ �yL   ™ b $x	$   ÿŽ â r+   ÿàõ ä �yJ
   ™ b $x   Ž â r+   ÿà÷ ä r	   ÿàÿ â �yM
   ™ b r	)   ÿä Ä r)   ÿäÿ â �y    • h �y   ™ h# �y   ™ h �yN   ™ h �y<   ™ h �y   ™ h �y
   ™ h �yO   ™ h �yP   ™ h �y    • b ‚x �˜­¿     Ä ‚x &     â $~Rÿ   ÿ Ž â |V       ä |T       ä |X       ä |       ä |Q       Ä x@  ÿàõ ä x    ÿàÿ ä x0ww  ÿÀŽ äOx00T  %    È� x00ˆˆˆˆÿÀŽ ä x%%ww  ÿÀŽ ä x0022ÿø� ä tRÏÝêö    ä/ tVÏÝêö%    ä x	0Tv  ÿÀŽ Ä tT5EYq    ä t%X5EYq%    ä x0ÿ   0  ä r	       ä r%0   %    ä |       Ä x	 d  %    äx4ww  ÿÀŽ äx1u  %    ä |0       ä x%4T  /    äx4.ww  ÿÀŽ Ä tÏÝêö    äx..T  1    â $~Rÿ   ÿ Ž â t05EYq    ä x%%ˆˆˆˆÿÀŽ ä |0       ä x//ww  ÿÀŽ Ä x..ˆˆˆˆÿÀŽ ä x%%22ÿø� ä t0ÏÝêö/    ät0R5EYq/    ä x/.22ÿø� ä x.%Tv  ÿÀŽ Ä |R       ä x11ww  ÿÀŽ ä tÏÝêö4    ä r.       ä t4Q5EYq4    ä x2ww  ÿÀŽ Ä tQRÏÝêö1    äx22T      ä tRT5EYq1    ä x%ÿ   %  ä x1/Tv  ÿÀŽ ä x/ÿ   /  Ä x22ˆˆˆˆÿÀŽ ä r0%   0    â $~%ÿ   ÿ Ž â rQ/   R    ä |/       ä x2222ÿø� ä r41   4    Ä xww  ÿÀŽ ä |R       ä t%ÏÝêö    ät./5EYq    ä x%2Tv  ÿÀŽ ä t/RÏÝêö    Ä tRT5EYq    ä x2ÿ   2  ä r%%   .    â &r.	6   ÿ   â x d  0    ä r2/2   R    â &r/9   .   ä &r.	:   ÿ   â x1u  0    â &r/>   /   Ä &r.=   .   â x4 d      â &r/@   /   ä &r.?   .   ä 0r7ÿ7    A   ä 0r;ÿ;    A   â x41u      â &r66   ÿ   â Er/ /       â &r1C   .   â  r.7;     A  â x% d  2    â &r69   6   Ä 0rBÿB    A   â #r5./   5   â &r/	E   ÿ   â x%%1u  2    â &r6>   6   â Er1 1       â  r97B     A  â &r/H   /   ä &r.	F   ÿ   ä &r6%@   6   â #r91      â 0r8ÿ8    A   Ä &r/K   /   ä &r1G   .   â Er6 6       â  r0;8     A  â &r/L   /   ä&r.J   1   ä 0rDÿD    A   â #r06      â &r0M   .   â Er/ /       Ä  r.7D     A  â &r::   ÿ   Æ #r3./   3   â &r/	   ÿ   ä &rEE   ÿ   ä&rFF   ÿ   ä&r   ÿ   ä &r		   ÿ   ä&r   ÿ   Ä &r/N   /   ä&r:=   :   ä&rHH   E   ä&rFG   F   ä&rNN      ä &r	<   	   Ä &r<      ä $rÿÿ    â |    pbô â &r:?   :   ä&rHK   H   ä&rFJ   F   ä&rN   N   Ä &r   /   ä&r	
   	   ä&r
      ä &r:%C   :   ä&rH%L   H   ä&rF%M   F   Ä &rN%O   N   ä0rIÿI    A   ä &rO      ä&r	P   	   ä&r%%P      ä 0rÿ    A   Ä 0rÿ    A   â Er: :       â  r27I     A  âErH H       â  rB8B     A  âEr0 0       â  r18D     A  âErF F       â  rI8I     A  âEr        â  r7     A  âErN N       â  r8     A  âEr	 	       â  r
7     A  âEr% %       â  r8     A  â x##@  ÿàù Ä x&&@  ÿà÷ â #rB:      â #r1H      â #r20      â #rIF      â #r-   -   â #rN      â #r
	      â #r%      â x    ÿàÿ â $r$ÿÿ   $ â x      ÿàÿ Ä x((    ÿàÿ ä x**    ÿàÿ ä x,,    ÿàÿ ä r'ÿ'   ÿäÿ â D)       À â Gy   ïÿÿÿÿƒ ê Ay        € ê Ey  ð     € â G‰  Ð     € ê $x  €  Ž Ê ˆs  5 úÿ H   è ˆs  €úÿ H   è ˆs   ûÿ H   è ˆs  €ûÿ H   è ˆs  3 üÿ H   è ˆs  €üÿ H   è ˆs   ýÿ H   è ˆs  €ýÿ H   è ˆs  - þÿ H   è ˆs  €þÿ H   è ˆs   ÿÿ H   è ˆs  €ÿÿ H   ä Ay        € ê {          ì M        € ê „y €   H   b z  r  p`ð Ä z%  r      â „y      H   " r ÿ   p"p Æ „y    H   ¢ $x#%   ÿŽ ä$x)%   ÿŽ â „y €  H   è „y    H   ( „y	 €  H   h „y    H   h „y €  H   h „y €  H   h „y
    H   è „y    H   â !r       â/ !r  5       Ä „y €  H   â !r       ÆO ‰      " !r       Æ� ‰       b !r33       Æ‰      ¢ !r		       Æ‰      â !r       Æ y      '   " !r       Æ ‰3      b !r       â !r       ä ‰      " !r-
-       â !r        ä/ ‰      b !r       â !r       ÄO ‰       ¢ !r       Æ� ‰
	      â !r       Æ ‰      " !r33       Æ‰      b x   ÿàÿ Æ ‰-      b !r       â z  r  p`ò ä ‰      " !r       â/ x    p"ò â !r        âO ‰      b !r		
       Æ� y       &   ¢ !r       Æ‰ €    â !r       Æ‰      " !r--       Æ ‰ €    b !r       Æ ‰3      ( ‰	      " !r       Æ/ ‰      b $z   x  Ž ÆO ‰      ¢ !r
       â� $z  ~   Ž ä ‰      â !r       är%    ÿ† â ‰-      " !r       â
  r  ÿàù Ä ‰      b !r33       â  #   ÿàõ ä ‰      " !r	       â ÿÿ   ÿä â $ÿÿ   ÿ â !r       â/ ‰	 @    b rÿÿ   ÿäÿ â !r       ÄO ‰ €    ¢ 
 j  ÿˆ â !r       ä� ‰
 @    â z j  ÿ† â !r--       ä‰ €    " 
 j  ÿ„ â !r       Ä‰3 €    b 
 k   â !r!       ä ‰ €    " z k  � ä 
 k   â ‰ €    " !r	       â/ 	%   ÿàÿ Ä ‰ €    b !r       ÆO ‰- €    ¢ !r
       â� r%    ÿ† ä ‰  €    â !r       âr&ÿÿ   ÿäÿ ä ‰" €    " !r33       âz
 j  ÿ† Ä ‰$!      b !r       â z k  &� â !r       â ‰ @    "  )   ÿàù â !r       ä/ ‰ @    b ÿÿ   ÿä â !r--       ÄO ‰      ¢ 
 j  ÿˆ â !r        ä� ‰3 @    â % 	   ÿàõ â !r"       ä‰ @    " 
	 k   â !r!!$       Ä‰ @    b (ÿÿ   ÿä ä % j  ÿ„ â‰ @    b !r       â $ÿ   ÿ Ž â % k  ( â ‰ - @    " !r       Æ/ ‰" @    b !r+       âO #   ÿàÿ ä ‰      ¢ !r3       â� #        ä ‰$ @    â !r       Æ‰&! €    " !r       â!r       ä ‰      b !r -        Æ ‰      " !r""       Æ/ ‰      b !r/       ÆO ‰      ¢     ÿàõ â%  j   Ž â!r$$       â� ‰      â (ÿÿ   ÿä â %  j  # Ž â !r&!&       â‰      "  j  ÿ„ Æ ‰!       " !r       â k  ( ä ‰#"      b !r       Æ ‰'& @    " !r       Æ/ ‰%$      b !r       ÆO †	 +    â !r       Æ� † /   â !r       Æ†	     â !r! !       â !r#"#       ä†     â !r'&'       Æ †	     â !r%$%       Æ/ †    è †	     è †     è †	 
!    è † 
#   è ‰'      ( †	 %    â M™        € ê x)   ÿàÿ âO !r''       Æ r    ÿàñ È r ÿÿ   ÿä  ä z j  ÿ€ È z k     Ê †y '    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y       "   " ¹z  l      â Ey      € â ™x?    â y      !   b ¹z  F   
   â x      ÿ  â $rÿÿ   ÿ Ž â x(      ÿ  â r, ÿ       â x      ÿ  â x      ÿ  â r  ÿ   pRð Ä r  ÿ   pRò är    ÿ(Ž È/ r   ÿ� È xÿ     È |    pbô Ú G)       € ê y	      '   " r ÿ       â $rÿÿ   ÿ Ž â xþÿÿÿÿÀŽ â y      %   b r
 ÿ       ä x"        â y      &   ¢ z& q  ÿàÿ â 'z	 y   Ž â r  €ÿàÿ Æ $x   ÿ Ž â/ zÿ z    â $x   ÿ Ž âx   ÿ   â 'z s  
 Ž ÆO x%ÿ     â $z p  Ž â zÿ t    â $z |  Ž âx%   %  â $z |  Ž â x   ÿàÿ â %x   ÿŽ È $z v  Ž â x   ÿüŽ â $z v  Ž ä $z p  Ž ä %v X  "Ž È %x   Ž â x   ÿàù Æ $z	 }  ÿŽ â z X  ÿà÷ â $z
 w  ÿŽ â r!ÿ   ÿä â %v" X  "Ž â z	  q      ä r
   ÿàõ â $rÿÿ   " Ž â x   ÿàù â $x"	   Ž âr	   ÿŽ â $x'   ÿ Ž â zÿ Y  d~ Ä r$	   ÿŽ ä r7ÿÿ   ÿä ä �y/úÿÿ • ¨  �y üÿÿ • â  r   ÿàõ Æ �y+þÿÿ • " r	#   ÿä Æ �y.    • h �y*þÿÿ • ( �y-    • " r   ÿàõ â $t8ÿ$   ÿ Ž Ä �y0   • " rÿ7   ÿä â %v
 Z  8 Ž ä �y1   • $ $x$   ÿŽ â r&   ÿà÷ ä r'
   ÿàõ ä r   ÿàÿ â $rÿÿ   7Ž Ä %v Z  8 Ž â r        â �y5
    • d $x   % ä $x	$   ÿŽ â/ r   ÿà÷ ä �y3   ™ b r'   ÿàõ Ä r	   ÿàÿ â �y6   ™ b r !       â $rÿÿ   7Ž â r	%   ÿä â �y4   ™ d $x=$   ÿŽ ä �y:   ™ h �y2üÿÿ • h �y9   ™ b r        Ä/ r #       â �y;   ™ b %v Z  8 Ž â r"   ÿà÷ ä �y<    • b $x   =Ž â r
'   ÿàõ ä r@ÿ7   ÿäÿ â �y=   ™ b r%   ÿä â %v Z  8 Ž â rE$   ÿà÷ â �y?   ™ d $x@$   ÿŽ â r'   ÿàõ â �y>    • d $x   Ž Ä �y@
   ™ h �yA
   ™ b rÿ7   ÿäÿ â %vE Z  8 Ž â/ r%   ÿä â �yB
   ™ d $x$   ÿŽ â r'   ÿàõ â �y8   ™ d $x		   Ž Ä �yC
   ™ h �y    • b r	%   ÿä Æ �yD   ™ h �yE   ™ h �y   ™ h �y
   ™ h �y    • h �yF   ™ h �y   ™ h �yG   ™ h �y    • b ‚x �˜­¿     Ä ‚x &     â |I       â $~ÿ   ÿ Ž â/ |K       ä |       â $~Hÿ   ÿ Ž ä $~Jÿ   ÿ Ž â x@  ÿàõ ä x    ÿàÿ Ä rÿ   ÿä ä |    pbô ä x@  ÿà÷ ä x@  ÿàù ä x    ÿàÿ ä x/ww  ÿÀŽ ÄO x//T       È� x//ˆˆˆˆÿÀŽ ä x  ww  ÿÀŽ ä x//22ÿø� ä tIÏÝêö    ätK5EYq    ä tÏÝêö     Ä x	/Tv  ÿÀŽ ä t5EYq     ä x/ÿ   /  ä r	       ä r/       ä |        Ä x+ww  ÿÀŽ äx d      äx1u      ä t	 ÏÝêö    ätH5EYq    ä x++T  .    Äx*ww  ÿÀŽ ä |       ä x..ww  ÿÀŽ ä |/       ä x**T  -    ä t  ÏÝêö    Ä x++ˆˆˆˆÿÀŽ ä tÏÝêö.    ätH5EYq    ä t./5EYq.    ä |H       ä xI-ww  ÿÀŽ Ä x/*ˆˆˆˆÿÀŽ ä x-+22ÿø� ä t*HÏÝêöI    äxH/22ÿø� ä x/-Tv  ÿÀŽ ä t+J5EYqI    Ä r		/       ä x0ww  ÿÀŽ äx00T  1    ä xIHTv  ÿÀŽ ä xHÿ   H  ä x-ÿ   -  Ä x00ˆˆˆˆÿÀŽ ä r**H   +    ä r.-   .    ä |+       ä |       ä |H       Ä x11ww  ÿÀŽ ä x0022ÿø� ä r I       ä t +5EYq    ätÏÝêö    ä t-HÏÝêö1    Ä x+0Tv  ÿÀŽ ä tHJ5EYq1    ä x0ÿ   0  ä x d  *    ä r++        â &r3   ÿ   âx1u  *    â &r33   ÿ   â r0-0   H    â &r6      âx 	 d  .    â &r36   3   â x*+ d  0    Ä x.	1u  .    â &r 4      âx0+1u  0    â &r3*4   3   ä &r49   ÿ   ä &r.:      ä&r30:   3   ä 0r2ÿ2    A   Ä 0r5ÿ5    A   ä 0r<ÿ<    A   ä &r+;   4   â Er        â  r	25     A  â Er3 3       â  r45<     A  â &r+ =   +   â #r	      â 0r	ÿ>    A   â #r43      â &r+.?   +   Ä &r4@   ÿ   â  r2	     A  ä Er+ +       â &r-A   4   È &r- B   -   â #r+      â &r+8   ÿ   ä &r-.C   -   ä 0rÿ    A   ä &r+D   +   â  r<	     A  â Er- -       â  r	2     A  â &r+    +   Ä &r69   ÿ   ä&r@@   ÿ   ä&r88   ÿ   ä &rE   ÿ   ä&rE   ÿ   â #r	-      â &r+.
   +   Ä 0r	ÿ    A   ä &r6;   6   ä&rAA   @   ä&r8D   8   ä &rF      ä&rF      â Er+ +       â &r6*=   6   â  r2	     A  â &rA*B   A   Ä &r*   8   ä &r       ä&r*      â #r,+   ,   â &r60?   6   ä&rA0C   A   ä&r0
      Ä &r.G      ä&r0G      ä 0rÿ    A   â Er6 6       â  r4<     A  âErA A       â  r	<	     A  â Er        â  r22     A  âEr        â  r<<     A  â Er        â #r)6   )   â #r4A      â #r(	   (   â #r2      â #r<      â x""    ÿàÿ â $r#ÿÿ   #Ž â x$$    ÿàÿ Ä x&&    ÿàÿ ä r!ÿ!   ÿä â G©  àðÿÿÿÿƒ ê Ay        € ê Ey  Ð     € â G‰  °     € ê $x  @  Ž Ê ˆs   ûÿ H   è ˆs  €ûÿ H   è ˆs   üÿ H   è ˆs  )€üÿ H   è ˆs   ýÿ H   è ˆs  €ýÿ H   è ˆs  , þÿ H   è ˆs  (€þÿ H   è ˆs   ÿÿ H   è ˆs  €ÿÿ H   ä Ay        € ê {          ì M        € ê „y     H   b Ey  @    € æ „y  €   H   ( „y    H   ¢ !r       â/ !r        È ‰       " !r       ÆO ‰      h ‰      ¢ !r         Æ y      %   " !r       Æ/ ‰       b !r       ÆO ‰      ¨ ‰	      â !r        â/ x    ÿ   â !r       ÆO ‰ €    " z   r  p`ð â !r
	       ä� ‰ €    b x    ÿàÿ ä r ÿ   pVp  â ‰	
 €    ¢ z  r  p`ò È x    p"ò Î ‰      &   â !r       Æ ‰      '   $ $ÿ   ÿ Ž â !r       â/ ‰ @    b !r
	       ÆO       &   ¨ ‰	 @    ( ‰ @    h       '   ( „y
 €  H   â !r       â/ $ x   Ž ÈO ‰      b !r		       â!r       È‰	      ¢ $Š x   Ž â� ˆ        â $ ~  Ž â ‰      " $Š ~  Ž â !r       Æ/ %† j   Ž È % j   Ž â †‰     â !r	       âO !r       È †    â G	  �     € ê y      &   (. y      '   b $z x   Ž È $z ~  Ž Ê/ z r  ÿàõ È rÿÿ   ÿä ä z j  ÿ„ È z k   Ê †y     è Ay        € ê !r

)       â „y    H   â Ey  À    € æ ‰
      (. „y €  H   è „y	    H   è „y €  H   è „y    H   è „y €  H   â !r
       Ê ‰      $ !r       Ê ‰ €    $ !r       Ê ‰ @    $ !r       Ê ‰      $ !r       â G™  °     € ê y      &   " ¹z  r      ä �x   ?àÿ â y      '   b $z x   Ž È $z ~  Ž Ê/ |   ÿàõ È rÿÿ   ÿä ä z j  ÿ„ È z k   Ê †y     è Ay        € ê !r       â !r       â !r		,       â !r(       â !r       â ‰      " !r       â z  r      ä ‰      h ‰	      ¨ ‰
      è ‰      ( ‰      b !r       Æ y      &   " !r       Æ/ ‰      b !r	       ÆO ‰      ¢ !r

       Æ� ‰      â !r       Æ‰	
      " !r       Æ‰      b $z x   Ž Æ ‰      " !r       â/ !r       ÈO ‰ €    b !r       Æ� ‰ €    ¢ !r	
	       Æ‰ €    â !r       Æ‰	 €    " !r       Æ ‰
 €    ( ‰ €    b !r       Æ/ y      '   b !r       ÆO ‰ @    ¢ !r       Æ� ‰ @    â !r	       Æ‰ @    " !r
       Æ ‰	 @    " !r       Æ‰ @    h ‰ @    b $z ~  Ž â/ !r        ÈO r   ÿ„ â!r
       â� $x   ÿŽ âr   ÿˆ â !r       ä‰
      b $rÿÿ   ÿ â z j  ÿ„ â!r	       â rÿÿ   ÿä â ‰	       " z k   â !r       â‚   ÿàõ â ‰      ¢    ÿàÿ â !r       â ‚ÿÿ   ÿä â ‰      â Š j  ÿ† Ä    ÿàõ â ‰      " Š k  � ä ÿÿ   ÿä â ‰      b  j  ÿ„ ä z j  ÿˆ â !r
       Æ/ z k   â !r 	       â 	 k   â †    â !r       ÆO †‰     â !r       Æ� †‰     â !r       Æ†     â !r       Æ†‰     â M™        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y       "   " ¹z  l      â Ey  `    € â ™x?    â y      !   b ¹z  F   
   â r ÿ       â x&      ÿ  â y      %   ¢ x      ÿ  â $r*ÿÿ   ÿ Ž â r ÿ       â $r,ÿÿ   ÿ Ž â r  ÿ   pRð Ä r  ÿ   pRò â$x     Ž â/ x   ÿ   ÈO r   ÿ� È x$ÿ     È | $   pbô Ú G)  0    € ê y      '   " $r
ÿÿ   ÿ Ž â xþÿÿÿÿÀŽ â $rÿÿ   ÿ Ž â y      &   b $z p  $Ž â r ÿ       â $tÿ   ÿ Ž â z"$ q  ÿàÿ â 'z y  
 Ž â r
  €ÿàÿ Æ 'z s   Ž â/ zÿ z    ä x
   ÿ   â$z }  ÿŽ â zÿ t    â $z |  $Ž âx

   ÿ   â $z	 |  Ž â x   ÿàÿ â %x   ÿŽ â x#ÿ   
  Ä x%
   ÿ   â $z v  Ž â x#
   #  â $z	 v  	Ž â x   ÿüŽ â $z p  Ž ä %v	 X  Ž È %x	   Ž â x   ÿàù Æ $z w  ÿŽ â z X  ÿà÷ â %v X  Ž â rÿ   ÿä ä r   ÿàõ â $vÿ q  ÿ Ž â x   ÿàù â $r!ÿÿ    Ž ä $x   $Ž âzÿ Y  	d~ â $x    $Ž Ä $r5ÿÿ   ÿ ä �y+úÿÿ • ¨  �yüÿÿ • â  r   ÿàõ Æ �y0    • " r!   ÿä Æ �y-þÿÿ • h �y)þÿÿ • ( �y(    • " r$   ÿàõ â $tÿ$   ÿ Ž Ä �y   • " rÿ5   ÿä â %v
 Z   Ž â r"   ÿàõ â �y.   • $ $x$   ÿŽ ä $r	ÿÿ   5 â r%
   ÿàõ ä r   ÿàÿ â %v Z   Ž È $x	$   ÿŽ â r        â $x	   # â r%   ÿàõ â/ $x   Ž â r        â �y9
    • h �y1   ™ b $x   # â r;   ÿàõ Ä �y3   ™ h �y/üÿÿ • b r    ÿà÷ Æ �y4   ™ b r<ÿ5   ÿä Æ �y6   ™ b r        ä r !       â �y7   ™ b $x=<$   ÿŽ Æ �y2    • h �y8   ™ h �y:   ™ b %v; Z   Ž â r;ÿ5   ÿäÿ Ä �y    • b %v Z   Ž â r%   ÿàõ ä/ �y?   ™ b $x;;$   ÿŽ ä $x   =Ž â r
%   ÿà÷ ä r;   ÿàÿ Ä r	#   ÿä â �y    • d $x   #Ž ä �y;   ™ h �y>   ™ h �y
   ™ h �y=   ™ h �y<
   ™ h �y   ™ h% �y
   ™ h �y@
   ™ h �y    • b ‚x �˜­¿     Ä ‚x &     â $~Aÿ   ÿ Ž â |B       ä |D       ä |       â $~Cÿ   ÿ Ž â x€  ÿàõ ä x$$@   ÿàÿ Ä rÿ   ÿä ä | $   pbô ä x€  ÿàù ä x€  ÿà÷ ä x@   ÿàÿ â $rÿÿ    â x  @   ÿàÿ Ä x""@   ÿàÿ ä r!ÿ!   ÿäÿ ä x	+ww  ÿÀŽ äOx++T      È� x++ˆˆˆˆÿÀŽ ä xww  ÿÀŽ ä x++22ÿø� ä tBÏÝêö	    ä tAÏÝêö    ä x
+Tv  ÿÀŽ Ä t	D5EYq	    ä t5EYq    ä x+ÿ   +  ä r	
   	    ä r+       â $~
ÿ   ÿ Ž â x)ww  ÿÀŽ Äx	 d      â$~Dÿ   ÿ Ž â x		1u      ä |       ä x0ww  ÿÀŽ ä/ x))T  (    ä x-ww  ÿÀŽ Ä|B       ä |+       ä x--T  0    ä t0ÏÝêö    ä x))ˆˆˆˆÿÀŽ ä t

ÏÝêö    Ä tDÏÝêö    ä tB5EYq    ätB5EYq    ä t+5EYq    ä |+       ä xB(ww  ÿÀŽ Ä xA)22ÿø� ä x--ˆˆˆˆÿÀŽ ä t(+ÏÝêöB    ät)C5EYqB    ä xBATv  ÿÀŽ ä x--22ÿø� Ä rB       ä x+-Tv  ÿÀŽ ä xww  ÿÀŽ äx-ÿ   -  ä xT  .    ä xAÿ   A  Ä r

+       ä r0-       â $~0ÿ   ÿ Ž â xˆˆˆˆÿÀŽ ä r((A   )    ä |+       ä x).ww  ÿÀŽ Ä |B       ä |       ä x22ÿø� ä t.+5EYq    ä t+0ÏÝêö)    ät0B5EYq)    Ä tÏÝêö    ä x)Tv  ÿÀŽ ä xÿ     ä x d  (    äx1u  (    â &r(1   ÿ   â r0+   0    Ä r))   .    â &r(	3   (   â x
 d      äx

1u      ä x) d  0    äx)1u  0    â &r)4   (   ä &r(6   ÿ   Ä &r66   ÿ   ä &r)
8   )   ä &r67   6   ä 0r/ÿ/    A   ä 0r9ÿ9    A   ä &r6:   6   â Er) )       â 0rÿ    A   â  r+/9     A  â &r6?   6   Ä 0r2ÿ2    A   â #r,+)   ,   â  r)/     A  â Er6 6       â &r+;   ÿ   â r2     A  â &r;;   ÿ   Æ #r6      â &r	>   +   ä&r>>   ;   ä &r   ÿ   ä &r.1   ÿ   ä&r   ÿ   ä &r(	7   (   Ä &r>=   >   ä &r.3   .   ä &r	<      ä&r<      ä &r(:   (   ä&r=      Ä &r>   >   ä &r.4   .   ä 0rÿ    A   ä &r      ä&r      â Er> >       â &r.8   .   Ä &r(
?   (   ä&r
      â  r2     A  â &r
@      ä&r@      ä 0rÿ    A   â Er. .       â  r/     A  â Er( (       â #r>      â Er        â  r992     A  â Er        â  r//     A  âEr        â  r2     A  â #r9.      â #r*)(   *   â #r'   '   â #r&/   &   â #r      â G©  €òÿÿÿÿƒ ê Ay        € ê Ey  °     € â G‰  �     € ê r    ÿ@Ž Ê ˆs , üÿ H   è ˆs €üÿ H   è ˆs * ýÿ H   è ˆs €ýÿ H   è ˆs ' þÿ H   è ˆs €þÿ H   è ˆs & ÿÿ H   è ˆs €ÿÿ H   ä Ay        € ê {          ì M        € ê „y     H   " z  r  p`ð Æ „y €  H   b r ÿ   pVp  Æ „y €   H   ¨ „y     H   è „y €  H   ( „y €  H   h „y    H   h „y €	  H   h „y    H   h „y    H   h „y €  H   b !r,       Æ „y	 €  H   " !r       Æ/ „y!    H   b !r       ÆO „y    H   ¢ !r         Æ� „y    H   â !r       Æ„y
 €  H   " !r       Æ„y    H   b !r         Æ „y  	  H   b !r       Æ „y  
  H   b !r*       Æ „y €
  H   b !r'       Æ „y#    H   b !r       Æ „y €  H   â !r	       Æ „y €  H   â !r!!&       Æ/ „y €  H   â !r       ÆO ‰       " !r       â� !r

       ä‰      b !r!       Æ‰	      ¢ !r       Æ y      &   â !r       Æ ‰      " !r

       Æ ‰      b !r#       Æ ‰
      â !r        â !r       ä ‰      " !r       â/ !r       ä ‰       b !r		       âO !r       Ä ‰      ¢ !r       Æ‰	      " !r       Æ‰      b !r
       Æ� ‰      â !r       Æ ‰
      " !r         Æ/ ‰      b !r       ÆO ‰      ¨ ‰  €    ¢ !r	       â!r       ä‰ €    " !r       Æ� ‰ €    â !r
       Æ ‰	 €    " !r       Æ/ ‰ €    b !r       ÆO ‰ €    ¢ !r        Æ ‰ €    b $z  x  Ž â x   ÿàÿ ä ‰      b !r       âz  r  p`ò ä y      '   " !r       â� x    p"ò Ä ‰ @    â !r
	       â !r       ä/ ‰ @    " !r       ÆO ‰	 @    b !r       Æ‰
 @    ¢ !r       Æ ‰ @    b $z  ~   Ž Æ‰ @    " z  r      â !r       ä� ‰ @    â Š  r  ÿàõ â $x   ÿŽ ä ‰ €    â !r       â $‚ÿÿ   ÿ â Š j  ÿ„ â !r	       â/ ‰      " r    ÿ† â !r
       âO Š k   â ‰      â 	   ÿàÿ â !r       ârÿÿ   ÿäÿ â ‰      b z j  ÿ† â!r       â 	   ÿàõ â ‰      ¢ z k  � â !r       â� ‚    ÿà÷ â ‰      â x	        â !r       â ÿÿ   ÿä â ‰      " 
 j  ÿ„ â %v  j  	 Ž â !r       â ‰ @    $ $‚ÿÿ   ÿŽ â Š j  ÿ† â‰      f Š k  � â !r       â/  k   â !r       â †‰     â !r       ÆO †    â !r       Æ� †‰     â !r       Æ† 
    â !r       Æ †‰     â !r       Æ†    è ‰      ( †‰     â M™        € ê x   ÿàÿ â/ !r       Æ r    ÿàñ È r ÿÿ   ÿä  ä z j  ÿ€ È z k     Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      "   " ¹z  l      â Ey  €    € â ™x?    â y       !   b ¹z  F   
   â r ÿ       â $r.ÿÿ   ÿ Ž â y      &   ¢ $r
ÿÿ   ÿ Ž â r0 ÿ       â $r(ÿÿ   ÿ Ž â y      '   â $r,ÿÿ   ÿ Ž â r ÿ   pRð Ä r ÿ   pRò â$x     Ž Ê/ r   ÿ� È xÿ     â $À    Ž Æ |    pbô Ú G)  @    € ê y	      %   "Î r ÿ       â $tÿ   ÿ Ž â xþÿÿÿÿÀŽ â $z }  ÿŽ âr ÿ       â 'z y   Ž È $x   
Ž â zÿ z    â 'z s   Ž Æ x   ÿ   â$z w  ÿŽ â zÿ t    ä x   ÿ   â %x   ÿŽ â r   ÿàõ ä xÿ     â $x		   ÿ Ž â rÿÿ   ÿä â $x   ÿ Ž âx     â $z	 p  Ž âx		   ÿàÿ Ä x   ÿüŽ â $z |  Ž ä$z |  Ž È $z v  Ž ä$z v  Ž ä $z		 p  Ž ä %v X  Ž È %x   Ž â x   ÿàù Æ %v	 X  Ž â z X  ÿà÷ ä rÿ   ÿä â $vÿ q  ÿ Ž â x   ÿàù â $r*ÿÿ    Ž â r ÿ       â $x   Ž â zÿ Y  d~ ä z q  ÿàÿ Æ �y8úÿÿ • ¨  �yüÿÿ • â  r *   ÿàõ Æ �y6þÿÿ • " r!   ÿä Æ �y	    • h �y) þÿÿ • ( �y2     • " r   ÿàõ â $t#ÿ$   ÿ Ž Ä �y-    • " rÿ   ÿä ä r&   ÿàõ â �y4    • " %v Z  # Ž Æ �y7üÿÿ • " $x$   ÿŽ â r   ÿà÷ â $rÿÿ    â r:   ÿàõ â %v&& Z  # Ž â r   ÿàÿ Æ $rÿÿ   Ž ä $x$   ÿŽ â r;   ÿä â %v  Z  # Ž â/ r"&   ÿàõ â�y9    • d $x$   ÿŽ ä $x''   Ž â �y/:   ™ b r$    ÿà÷ Ä $x#'    â r!!   ÿàÿ â �y1:   ™ f r%!   ÿäÿ â �y"   ™ h �y+:   ™ h �y"   ™ h �y$   ™ h �y:   ™ h �y3$   ™ b $rÿÿ   * Ž Æ/ �y<$   ™ b r        Æ �y5"   ™ h �y="   ™ h �y:$   ™ h �y&&    • h �y      • h �y;    • b ‚x �˜­¿     Ä ‚x &     â $~#ÿ   ÿ Ž â/ |"       â $~ÿ   ÿ Ž â |       ä x€  ÿàõ ä x@   ÿàÿ Æ $rÿÿ    â |    pbô ä x**€  ÿà÷ ä x€  ÿàù ä x@   ÿàÿ â $rÿÿ   Ž â x@   ÿàÿ ä rÿ   ÿä Ä x%8ww  ÿÀŽ äOx88T      È� x88ˆˆˆˆÿÀŽ ä x!ww  ÿÀŽ ä x8822ÿø� ä tÏÝêö%    äx8Tv  ÿÀŽ ä t5EYq%    Ä t##ÏÝêö!    ät""5EYq!    ä x8ÿ   8  ä r       ä r"#8   "    ä |       Ä x$6ww  ÿÀŽ â$~!ÿ   ÿ Ž â x d  "    äx#6T  	    äx%	ww  ÿÀŽ ä x"1u  "    ä |6       Ä x	)ww  ÿÀŽ ä tÏÝêö$    â$~ÿ   ÿ Ž â x')T  2    ä |8       ä t66ÏÝêö	    ä t$!5EYq$    Ä t	!5EYq	    â $~!ÿ   ÿ Ž â x##ˆˆˆˆÿÀŽ ä x''ˆˆˆˆÿÀŽ ä t88ÏÝêö%    ät5EYq%    ä x22ww  ÿÀŽ Ä |%       ä x##22ÿø� ä x''22ÿø� ä t!!ÏÝêö2    ät2%5EYq2    ä x%#Tv  ÿÀŽ Ä x)ÿ   #  ä x#'Tv  ÿÀŽ ä x'ÿ   '  ä r	6#   	    ä x#-ww  ÿÀŽ äx--T  4    Ä r2!'   2    ä x'-ˆˆˆˆÿÀŽ ä r88)       â $~ÿ   ÿ Ž â r%   $    â $~!ÿ   ÿ Ž â |$       ä x64ww  ÿÀŽ Ä x''22ÿø� ä |%       ä tÏÝêö#    ät4$5EYq#    ä t!!ÏÝêö6    ä x#'Tv  ÿÀŽ Ä t6%5EYq6    ä x'ÿ   '  ä x$	 d  2    ä r#   4    ä r!!'   6    ä x2	1u  2    â &r	/   ÿ   â x6 d  8    Ä x4 d  !    â &r	"1   	   â x81u  8    â &r   ÿ   âx1u  !    â &r!$   ÿ   ä &r	6+   	   ä &r"      ä&r!2   !   Ä &r   ÿ   ä &r/$/   ÿ   ä &r8   	   ä &r$   ÿ   ä &r	"3      ä 0rÿ7    A   Ä 0r9ÿ9    A   ä &r/21   /   ä&r323      ä &r6<   	   âEr	        â &r65      â  r9     A  â &r/4+   /   Ä &r45   !   ä&r34<   3   â #r,	   ,   â &r8=      ä&r88:      ä &r   /   ä&r==      Ä 0rÿ&    A   ä &r3:   3   ä 0r	ÿ     A   ä 0rÿ;    A   â Er        â  r     A  âEr8 8       â  r	     A  â Er        â  r99     A  â Er= =       â  r     A  âEr3 3       â  r		     A  â #r0   0   â #r.8   .   â #r(9   (   â #r
=   
   â #r	3      â G©  `ôÿÿÿÿƒ ê Ay        € êÏ ˆ , ýÿ H   è ˆ (€ýÿ H   è ˆ 0 þÿ H   è ˆ 
€þÿ H   è ˆ . ÿÿ H   è ˆ €ÿÿ H   è {          ì M        € ê „y      H   " z  r      Æ „y     H   h „y  €   H   ¨ „y     H   è „y     H   ( „y  €  H   h „y     H   h „y  €  H   h „y     H   h „y  €  H   h „y  €  H   b !r,       Æ „y     H   " !r       Æ/ „y	  €  H   b !r(       ÆO „y     H   ¢ !r0       Æ� „y  €  H   â !r       Æ„y  €  H   â !r
       Æ„y     H   â !r.       Æ „y  €  H   â !r       â !r       ä ‰      " !r       Æ y      %   b !r       â !r       â !r	       â/ !r       ÈO ‰      " !r       Æ� ‰      b !r       â!r       ä ‰	      ¢ $x   ÿ Ž â!r       ä ‰      â $z x  Ž â z  r  p`ð â !r       â ‰      " $z ~  Ž â r  ÿ   p"p Ä ‰      b !r       â !r       È/ ‰      " !r		       ÆO ‰      b !r       Æ� ‰	      ¢ !r       Æ‰ €    â !r       Æ‰
      ( ‰      b !r       â !r       È/ ‰ €    " !r	       ÆO ‰ €    b !r       Æ� ‰	 €    ¢ !r

       Æ‰ @    â !r       Æ‰
 €    ( ‰ €    b !r       â !r       È/ ‰ @    " !r		       ÆO ‰ @    b !r       Æ� ‰	 @    ¢ !r
       âx
   ÿàÿ ä ‰      â !r       âz 
 r  p`ò ä ‰ @    $ x     p"ò Ä ‰ @    b !r       â !r
       È/ ‰       " 
 r  ÿàõ â !r	       ÆO    ÿàÿ â $ÿÿ   ÿ â 
 j  ÿ„ â!r       â� r	   ÿ† ä 
 k   â !r       â   ÿàõ â ‰
      b x        â !r       â$rÿÿ   ÿŽ â ‰      ¢ z	 j  ÿ† â %v j   Ž â ÿÿ   ÿä â ‰      â  j  ÿ„ â !r        â z	 k  � â ‰      " 	 k   Æ †	     è	 †    â	 !r
       â/ !r       ÈO †	     â	 !r       Æ� †     â	 !r       Æ †	     â	 M™        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      "   " ¹z  l      â Ey  �    € â ™x?    â y       !   b ¹z  F   
   â $rÿÿ   ÿ Ž â r ÿ       â y      %   ¢ $rÿÿ   ÿ Ž ä $r$ÿÿ   ÿ Ž â r ÿ   pRð Ä r ÿ   pRò â$x     Ž â/ x   ÿ   ÈO r   ÿ� Ê $€    Ž â xÿ     È |    pbô Ú G)  p    € ê y      '   " $rÿÿ   ÿ Ž â x	þÿÿÿÿÀŽ â $rÿÿ   ÿ Ž â y      &   b $tÿ   ÿ Ž â z q  ÿàÿ â $x	   	
Ž ä $rÿÿ   ÿ Ž Æ x	   ÿ   â$x		   ÿ Ž È $x	   ÿ Ž ä 'z y   Ž È 'z s   Ž â/ zÿ z    Æ $z p  Ž â zÿ t    â $z |  Ž ä$z |  Ž ä $z v  Ž ä$z v  Ž â x   ÿàÿ â %x   ÿŽ È %v X  Ž È $z }  ÿŽ â x!   ÿàù â $z p  Ž ä %x   Ž È $z w  ÿŽ ä %v X  Ž â z X  ÿà÷ ä r   ÿàõ â $rÿÿ    â x   ÿàù ä xÿ   	  â $r,ÿÿ   ÿ â r        Ä zÿ Y  d~ ä x	     ä x   ÿüŽ ä �yúÿÿ • ¨  �yüÿÿ • â  r
   ÿàõ Æ �y%þÿÿ • " r   ÿä Æ �y"    • h �y#
þÿÿ • ( �y&
    • " r   ÿàõ â $tÿ$   ÿ Ž Ä �y 
   • $ $rÿÿ   , â r   ÿàõ â �y+
   • " %v Z   Ž È $x$   ÿŽ ä $rÿÿ   , â r   ÿàõ â %v Z   Ž â r   ÿàÿ Æ $x$   ÿŽ ä $x	    â r   ÿàõ â $x   Ž â r
 !       â/ $rÿÿ    Ž â �y*   ™ b $x    Æ �y)   ™ h �y'   ™ h �y(   ™ h �y-   ™ h �y.   ™ h �y

üÿÿ • h �y    • b $rÿÿ    Ž Æ/ �y/   ™ b $r	ÿÿ    Ž Æ �y    • h �y1   ™ h �y0    • b ‚x �˜­¿     ä ‚x &     â |2       â $~4ÿ   ÿ Ž Ä $~3ÿ   ÿ Ž ä $~ÿ   ÿ Ž ä $~	ÿ   ÿ Ž ä/ $~ÿ   ÿ Ž â x€  ÿàõ ä x@   ÿàÿ ä rÿ   ÿä Ä |    pbô ä x!!€  ÿàù ä x€  ÿà÷ ä x@   ÿàÿ â $rÿÿ    â rÿ   ÿäÿ ä xww  ÿÀŽ ÄO xT      È� xˆˆˆˆÿÀŽ ä xww  ÿÀŽ ä x22ÿø� ä t2ÏÝêö    äxTv  ÿÀŽ ä t45EYq    Ä t3ÏÝêö    ät5EYq    ä xÿ     ä r       ä r       â $~ÿ   ÿ Ž â x%ww  ÿÀŽ Äx d      äx1u      ä x#ww  ÿÀŽ äx##T  &    ä t		ÏÝêö    ät5EYq    Ä x%%T  "    äx"ww  ÿÀŽ â $~"ÿ   ÿ Ž â x##ˆˆˆˆÿÀŽ ä |       ä x##22ÿø� ä t""ÏÝêö    Ä t5EYq    â $~ÿ   ÿ Ž â x%%ˆˆˆˆÿÀŽ ä x#Tv  ÿÀŽ ä x&&ww  ÿÀŽ ä x%%22ÿø� ä r""       Ä tÏÝêö&    äx ww  ÿÀŽ ät&5EYq&    ä x  T  +    ä x%Tv  ÿÀŽ ä t22ÏÝêö    Ä x#ÿ   #  ä t45EYq    â $~4ÿ   ÿ Ž â x  ˆˆˆˆÿÀŽ ä r		       â $~ÿ   ÿ Ž â x++ww  ÿÀŽ â $~ÿ   ÿ Ž â x%ÿ   %  Ä r#   &    ä |&       ä x  22ÿø� ä t4ÏÝêö+    ä r22%       ä t45EYq+    Ä tÏÝêö    ät&&5EYq    ä x Tv  ÿÀŽ ä x ÿ      ä r&   &    â &r*   ÿ   â x" d      Ä r#    4    â &r)      â x	 d  2    â&r**   ÿ   â x"1u      â &r'   ÿ   âx		1u  2    â &r''   ÿ   â x& d  #    â &r(      ä &r*)   *   Ä &r-      ä&r-   '   â x&1u  #    â &r	.      ä 0r
ÿ
    A   ä 0rÿ    A   ä &r*(   *   Ä &r/      ä&r/      â Er        â &r*.   *   â  r
     A  â 0rÿ    A   ä &r		1      ä&r1      Ä 0r0ÿ0    A   â #r$   $   â  r
     A  â Er	 	       ä Er* *       â  r0     A  â Er        â  r0     A  â #r	      â #r*      Ä #r      â G©  öÿÿÿÿƒ ê Ay        € ê ˆ $ þÿ H   è ˆ €þÿ H   è ˆ  ÿÿ H   è ˆ €ÿÿ H   è {          ì M        € ê „y      H   " z  r  p`ð Æ „y  €   H   b r  ÿ   p"p Æ „y     H   ¨ „y     H   è „y  €  H   ( „y     H   h „y     H   h „y	  €  H   h „y     H   h „y  €  H   h „y
  €  H   b !r$       Æ „y  €  H   " !r       â/ !r       âO !r       ä� !r       â!r       â!r       ä !r		       Æ ‰      b !r       Æ ‰	      ¢ !r       Æ ‰      â !r

       Æ y      &   " !r
       â !r       È/ ‰      " !r	       ÆO ‰      b !r       Æ� ‰      ¨ ‰      â !r       â !r       È/ ‰      " !r       ÆO ‰ €    b !r       Æ� ‰ €    ¨ ‰
 €    â !r       â !r       È/ ‰ €    " x   ÿàÿ â $z x  Ž â!r       äO ‰ @    b z  r  p`ò â !r

       ä� ‰ @    ¢ x     p"ò Æ ‰	
 @    è y      '   " !r       â !r       â/ !r       ÈO ‰      " !r	
	       Æ� ‰      b $z ~  Ž Æ‰
 @    ¢ $tÿ   ÿ Ž â 
  r  ÿàõ ä ‰	      ä ÿÿ   ÿä ä 
  j  ÿ„ È 
  k   â !r       â %v j   Ž â !r       È/ †	     â !r

       ÆO †    â !r		       Æ� ‰
      h  †	 	    â M™        € ê ¹z  r      â !r
       â/ �x   ?àÿ Ì |   ÿàñ Ê $rÿÿ   ÿ  â z j  ÿ€ È z k    Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y       &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž ä $rÿÿ     Ž È % \   Ž Ê �	    ™ b ¹z  l      â Ey  à    € â ™x?    â y      "   b $rÿÿ   ÿ Ž Æ y      !   ¢ $ÿÿ   ÿ Ž ä $ÿÿ    Ž È '
 m   Ž È $‚ÿÿ   ÿ Ž â 
ÿ n    È ÿ   ÿáÿ ä r ÿ   pRò Æ/ $
 o  Ž â r ÿ   pRô â$x	    Ž äO $‚ÿÿ    Ž Æ r		   ÿ� â 'Š s   Ž Æ x
ÿ     È | 
   pbö ä Šÿ t    Ö G9  €    € ê y      '   " $rÿÿ   ÿ Ž â xþÿÿÿÿÀŽ â $rÿÿ   ÿ Ž â y      %   b 'z y   Ž È $z }  ÿŽ â zÿ z    Ê $z |  
Ž È $z v  Ž È$z p  Ž ä/ $x	   
Ž ä %x   ÿŽ Æ x   ÿ   â $z w  ÿŽ â z X  ÿàñ â $x   ÿ Ž ä %x   Ž â r   ÿàý ä x   ÿà÷ ä z X  ÿàù â $rÿÿ   ÿ â xÿ     Ä x   ÿàû ä zÿ Y  þ ä zÿ Y  „þ ä x     ä $rÿÿ    Ž Ê �yþÿÿ • ¨ �y    • â r
   ÿàñ â $tÿ$   ÿ Ž ä �y   • " rÿ   ÿä  â %v Z   Ž Ä �y   • d $x$   ÿŽ â r   ÿ€ Æ $x   Ž È $x     â �y    • ( �y   ™ ( �y   ™ h �y   ™ h �y   ™ d $rÿÿ    Ž Ä/ $rÿÿ    Ž Ê �yøÿÿ • b ‚x �˜­¿     ä ‚x &     â $~ÿ   ÿ Ž â |       â $~!ÿ   ÿ Ž ä $~#ÿ   ÿ Ž â |       â $~ÿ   ÿ Ž â x

@   ÿàÿ È | 
   pbð ä x€  ÿà÷ Ê $rÿÿ   Ž â xww  ÿÀŽ äOxT      È� xˆˆˆˆÿÀŽ ä xww  ÿÀŽ ä x22ÿø� ä tÏÝêö    ät5EYq    ä t!ÏÝêö    Ä/ xTv  ÿÀŽ ä t#5EYq    ä xÿ     ä r       ä r       ä xww  ÿÀŽ ÄxT      â$~ÿ   ÿ Ž ä $~ÿ   ÿ Ž â xˆˆˆˆÿÀŽ ä xww  ÿÀŽ ä x22ÿø� ä tÏÝêö    Ä t5EYq    ä tÏÝêö    ät5EYq    ä xTv  ÿÀŽ ä xÿ     ä x d      Ä x1u      ä r       â &r   ÿ   â r       Æ &r      â x d      äx1u      Æ &r      ä 0rÿ    A   ä &r      ä 0rÿ    A   â x€  ÿàù ä Er        ä  r     A  â $rÿÿ    Æ #r      â G‰  0ûÿÿÿÿƒ ê Ay        € ê ˆ 	€ÿÿ H   è {          ì M)        € ê „y     H   " r ÿ   pRð Æ „y €   H   h „y    H   ¢!r       È !r       È/ !r       ÊO ‰      $ !r       Ê ‰      $ !r       Ê ‰ €    $ !r       Ê ‰ @    $ !r       Ê ‰      b  M	        € ê y      %   ¨ y      '   â $z   x  Ž äO $tÿ   ÿ Ž ä $z  ~   Ž â� !r       Æ/ %v  j   Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž ä $ÿÿ    Ž È % \   Ž Ì �	    ™ b $ÿÿ   ÿ Ž â Ey  0    € â $rÿÿ   ÿ Ž â y      "   b '
  m   Ž â r ÿ       ä y      !   b $‚ÿÿ   ÿ Ž â 
 ÿ n     Ê $ ÿÿ    
Ž È $
   o  Ž ä $‚ÿÿ    Ž È 'Š s   Ž â rÿ   ÿü‚ Æ/ $x    Ž Ê r   ÿ� ä Šÿ t    Æ G       € ê z ÿ `  pPô â $rÿÿ   ÿ Ž â z ÿ ^  pPò â $rÿÿ   ÿ Ž â z ÿ b  p ö ä z ÿ a   Sô ä z ÿ _  Sò ä z ÿ c  0't Ö       '   "        ä¢       â ©	      '   h       %   ¢ $ª x  ÿŽ Æ ©      %   â $ ~  ÿŽ ä $ª
	 ~  ÿŽ ä/ $	 x  ÿŽ â    àë äOÿ     ä ¢

   àÇ Ä� ¨ÿ     ä ÿÿ   Äþ ä 	   ÿàý ä ¢

   ÿàû ä ¢ÿÿ   „þ â $ÿÿ    â  ^  ÿ† Æ $¢ÿÿ   Ž â ª
 b  ÿˆ ä 	 _  � ä ª
 c   Æ �     h �©     d Ay        € ê ¹z  l      â xÿ     â ™x?    â Ey  p
    € â $rÿÿ   ÿ Ž ä $‚ ÿÿ    Ž Ä |    pbò â $r
ÿÿ   ÿ Ž Ø G  
    € ê y      '   b r ÿ       â $z	  w  ÿŽ â xþÿÿÿÿÀŽ â y      %   " $tÿ   ÿ Ž â z ÿ `  pPð â $r
ÿÿ   ÿ Ž ä $x   
Ž â z ÿ a   Sð â $rÿÿ   ÿ Ž Ä $x   ÿ Ž ä$x   ÿ Ž ä 'z y   Ž È/ $z  }  ÿŽ â zÿ z    È r 	    ÿàù â $z |  Ž Æ rÿÿ   ÿä â $z v  Ž È$z p  Ž È %x   ÿŽ Ì %x   Ž È %v X  Ž â z X  ÿàõ Æ %v `  Ž â x   ÿàó ä x   ÿà÷ Æ $r	ÿÿ   Ž  â zÿ Y  Dþ Æ $rÿÿ    Ž ä $rÿÿ    Ž â r    ÿàó â $tÿ$   ÿ Ž â �yüÿÿ • ¨  �yúÿÿ • è �yüÿÿ • " $rÿÿ   Ž  Ä %v Z   Ž â �yþÿÿ • h �y    • ¢ $x$   ÿŽ È $x   Ž È %x   Ž â �y    • ¨ �y   ™ ¨  �y   ™ ¨  �y   ™ ¨  �y   ™ ¢  ‚x �˜­¿     Ä ‚x &     â $~"ÿ   ÿ Ž â |        â $~#ÿ   ÿ Ž ä $~$ÿ   ÿ Ž â x@   ÿàÿ ä x€  ÿàõ ä |    pbø Ä x!ww  ÿÀŽ ä�xT      ÈxˆˆˆˆÿÀŽ ä xww  ÿÀŽ ä/ x22ÿø� ä t"ÏÝêö!    ä t  ÏÝêö    ät#5EYq    Ä t!$5EYq!    ä xTv  ÿÀŽ ä xww  ÿÀŽ äxÿ     ä xT      äO r   !    â $~!ÿ   ÿ Ž â r        â $~ÿ   ÿ Ž â xˆˆˆˆÿÀŽ â $~ÿ   ÿ Ž â x ww  ÿÀŽ Ä x22ÿø� ä tÏÝêö    ät!5EYq    ä tÏÝêö     ät#5EYq     ä xTv  ÿÀŽ Ä x ÿ     ä x d      äx1u      ä r       â &r   ÿ   â r        Æ &r      â x d      äx1u      Æ &r      ä $rÿÿ    Ž â x€  ÿà÷ â 0rÿ    A   ä $rÿÿ    Ž â x€  ÿàó â 0rÿ    A   ä &r      â rÿ   ÿäÿ  â $rÿÿ   Ž â  r     A  Ä Er        â G‰  @    € ê %x   Ž ä �y     ¨  �y%    è �y    ( �y    h �y    ¢ $~ÿ   ÿ Ž Ä $~#ÿ   ÿ Ž ä $~!ÿ   ÿ Ž ä $~"ÿ   ÿ Ž â x %ww  ÿÀŽ ä�x%%T      Èx%%ˆˆˆˆÿÀŽ ä x$ww  ÿÀŽ ä x%22ÿø� ä/ tÏÝêö     ät#5EYq     ä xTv  ÿÀŽ Ä t!!ÏÝêö$    ät""5EYq$    ä xÿ     ä x ww  ÿÀŽ äxT      äO r!   "    Ä r       â $~ÿ   ÿ Ž â x"ˆˆˆˆÿÀŽ â $~!ÿ   ÿ Ž â x#ww  ÿÀŽ â $~ÿ   ÿ Ž â |       ä x"22ÿø� ä tÏÝêö     Ä t !5EYq     ä tÏÝêö#    ät#5EYq#    ä x!Tv  ÿÀŽ ä xÿ     ä x d      Ä x1u      ä r!        â &r   ÿ   â r   #    Æ &r      â x d      äx1u      Æ &r      ä 0rÿ    A   ä &r      Æ  r     A  ä Er        Ê #r      â #r
   
   â $r	ÿÿ   	 â GÉ  Ð÷ÿÿÿÿƒ ê Ay        € ê r ÿ   pRð âEy  €     € â r ÿ   pRò Ö G‰  P     € ê z ÿ `  pPð â $x    Ž Æ z ÿ a   Sð ä ˆs 
€ÿÿ H   ö ˆ    H   ä Ay        € ê {          ì M        € ê z ÿ `  pPð â „y     H   f. z ÿ a   Sð Ú „‰  €   H   ¨ „‰    H   â !r
       È/ !‚         ÈO !‚        â� G‰        € ê „y  €  H   h „y €   H   ¨ „y    H   è „y	    H   ( „y €  H   " !r         â/ !r       ÆO !r         â� !r	       Æ !r        ä‰      b z ÿ `  pPð ä r ÿ   pRò ä z ÿ a   Sð â !r       Ê/ ‰       d !r         Ê/ ‰  €    d !r        Ê/ ‰ @    d!r       Ê/ ‰      d !r       â/ G‰        € ê ‰       d !r         Ê/ ‰       d !r        Ê/ ‰ €    d !r       Ê/ ‰ @    d !r       Ê/ ‰      d !r       ä/ M        € ê !r       â G‰  `    € ê $v ÿ h  ÿ Ž â !r       È x     pBð Ú G	  p    € ê x  ÿÿÿÿÿàÿ È x       € Ê $x     ÿ Ž È ‚{   €     d xÿ     È/ Iy 0íÿÿÿÿƒ ê  x ;ª¸¿  A  Ì s          d !t   €?    Ì/ s        d  r     A  È/  r     A  â Gy  P    € ê  x '7=  A  â x*BL?  A  Æ #t   €?    È  r      A  â  x    ?  A  Ê s     $   d !t  €?    È/  r      A  È  r     A  â Gy  °    € ê x     pRð Ú G‰  à     € ê x     pRð Ú G	        € ê 	z  i    � ä 	z i    � Æ  x ;ª¸¿  A  â 	z i €  � Ê s        d !t  €?    Ì/ s        d  r      A  È/  r     A  â Gy  Ð     € ê  r     A  â Gy  °     € ê 	x   à@  � ä 	x  à@  � Æ  x #ÛÙ?  A  â 	x  àÀ  � Æ  x;ª¸¿  A  ä !t  €?    È s        d !t  €?    Ì/ s        d  r       A  È/  r      A  ä y       &   h y      %   h y      '   ¢ $z   x  Ž ä/ $tÿ   ÿ Ž Ä $z  ~   Ž ÈO %v  j   Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y       &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž ä $rÿÿ     Ž È % \   Ž Ê �	    ™ b ¹z  l      â Ey  à    € â ™x?    â y      "   b $rÿÿ   ÿ Ž Æ y      !   ¢ $ÿÿ   ÿ Ž ä $ÿÿ    Ž È '
 m   Ž È $‚ÿÿ   ÿ Ž â 
ÿ n    È ÿ   ÿáÿ ä r ÿ   pRò Æ/ $
 o  Ž â r ÿ   pRô â$x	    Ž äO $‚ÿÿ    Ž Æ r		   ÿ� â 'Š s   Ž Æ x
ÿ     È | 
   pbö ä Šÿ t    Ö G9  €    € ê y      '   " $rÿÿ   ÿ Ž â xþÿÿÿÿÀŽ â $rÿÿ   ÿ Ž â y      %   b 'z y   Ž È $z }  ÿŽ â zÿ z    Ê $z |  
Ž È $z v  Ž È$z p  Ž ä/ $x	   
Ž ä %x   ÿŽ Æ x   ÿ   â $z w  ÿŽ â z X  ÿàñ â $x   ÿ Ž ä %x   Ž â r   ÿàý ä x   ÿà÷ ä z X  ÿàù â $rÿÿ   ÿ â xÿ     Ä x   ÿàû ä zÿ Y  þ ä zÿ Y  „þ ä x     ä $rÿÿ    Ž Ê �yþÿÿ • ¨ �y    • â r
   ÿàñ â $tÿ$   ÿ Ž ä �y   • " rÿ   ÿä  â %v Z   Ž Ä �y   • d $x$   ÿŽ â r   ÿ€ Æ $x   Ž È $x     â �y    • ( �y   ™ ( �y   ™ h �y   ™ h �y   ™ d $rÿÿ    Ž Ä/ $rÿÿ    Ž Ê �yøÿÿ • b ‚x �˜­¿     ä ‚x &     â $~ÿ   ÿ Ž â |       â $~!ÿ   ÿ Ž ä $~#ÿ   ÿ Ž â |       â $~ÿ   ÿ Ž â x

@   ÿàÿ È | 
   pbð ä x€  ÿà÷ Ê $rÿÿ   Ž â xww  ÿÀŽ äOxT      È� xˆˆˆˆÿÀŽ ä xww  ÿÀŽ ä x22ÿø� ä tÏÝêö    ät5EYq    ä t!ÏÝêö    Ä/ xTv  ÿÀŽ ä t#5EYq    ä xÿ     ä r       ä r       ä xww  ÿÀŽ ÄxT      â$~ÿ   ÿ Ž ä $~ÿ   ÿ Ž â xˆˆˆˆÿÀŽ ä xww  ÿÀŽ ä x22ÿø� ä tÏÝêö    Ä t5EYq    ä tÏÝêö    ät5EYq    ä xTv  ÿÀŽ ä xÿ     ä x d      Ä x1u      ä r       â &r   ÿ   â r       Æ &r      â x d      äx1u      Æ &r      ä 0rÿ    A   ä &r      ä 0rÿ    A   â x€  ÿàù ä Er        ä  r     A  â $rÿÿ    Æ #r      â G‰  0ûÿÿÿÿƒ ê Ay        € ê ˆ 	€ÿÿ H   è {          ì M)        € ê „y     H   " r ÿ   pRð Æ „y €   H   h „y    H   ¢!r       È !r       È/ !r       ÊO ‰      $ !r       Ê ‰      $ !r       Ê ‰ €    $ !r       Ê ‰ @    $ !r       Ê ‰      b  M	        € ê y      %   ¨ y      '   â $z   x  Ž äO $tÿ   ÿ Ž ä $z  ~   Ž â� !r       Æ/ %v  j   Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž ä $ÿÿ    Ž È % \   Ž Ì �	    ™ b $ÿÿ   ÿ Ž â Ey  0    € â $rÿÿ   ÿ Ž â y      "   b '
  m   Ž â r ÿ       ä y      !   b $‚ÿÿ   ÿ Ž â 
 ÿ n     Ê $ ÿÿ    
Ž È $
   o  Ž ä $‚ÿÿ    Ž È 'Š s   Ž â rÿ   ÿü‚ Æ/ $x    Ž Ê r   ÿ� ä Šÿ t    Æ G       € ê z ÿ `  pPô â $rÿÿ   ÿ Ž â z ÿ ^  pPò â $rÿÿ   ÿ Ž â z ÿ b  p ö ä z ÿ a   Sô ä z ÿ _  Sò ä z ÿ c  0't Ö       '   "        ä¢       â ©	      '   h       %   ¢ $ª x  ÿŽ Æ ©      %   â $ ~  ÿŽ ä $ª
	 ~  ÿŽ ä/ $	 x  ÿŽ â    àë äOÿ     ä ¢

   àÇ Ä� ¨ÿ     ä ÿÿ   Äþ ä 	   ÿàý ä ¢

   ÿàû ä ¢ÿÿ   „þ â $ÿÿ    â  ^  ÿ† Æ $¢ÿÿ   Ž â ª
 b  ÿˆ ä 	 _  � ä ª
 c   Æ �     h �©     d Ay        € ê ¹z  l      â xÿ     â ™x?    â Ey  p
    € â $rÿÿ   ÿ Ž ä $‚ ÿÿ    Ž Ä |    pbò â $r
ÿÿ   ÿ Ž Ø G  
    € ê y      '   b r ÿ       â $z	  w  ÿŽ â xþÿÿÿÿÀŽ â y      %   " $tÿ   ÿ Ž â z ÿ `  pPð â $r
ÿÿ   ÿ Ž ä $x   
Ž â z ÿ a   Sð â $rÿÿ   ÿ Ž Ä $x   ÿ Ž ä$x   ÿ Ž ä 'z y   Ž È/ $z  }  ÿŽ â zÿ z    È r 	    ÿàù â $z |  Ž Æ rÿÿ   ÿä â $z v  Ž È$z p  Ž È %x   ÿŽ Ì %x   Ž È %v X  Ž â z X  ÿàõ Æ %v `  Ž â x   ÿàó ä x   ÿà÷ Æ $r	ÿÿ   Ž  â zÿ Y  Dþ Æ $rÿÿ    Ž ä $rÿÿ    Ž â r    ÿàó â $tÿ$   ÿ Ž â �yüÿÿ • ¨  �yúÿÿ • è �yüÿÿ • " $rÿÿ   Ž  Ä %v Z   Ž â �yþÿÿ • h �y    • ¢ $x$   ÿŽ È $x   Ž È %x   Ž â �y    • ¨ �y   ™ ¨  �y   ™ ¨  �y   ™ ¨  �y   ™ ¢  ‚x �˜­¿     Ä ‚x &     â $~"ÿ   ÿ Ž â |        â $~#ÿ   ÿ Ž ä $~$ÿ   ÿ Ž â x@   ÿàÿ ä x€  ÿàõ ä |    pbø Ä x!ww  ÿÀŽ ä�xT      ÈxˆˆˆˆÿÀŽ ä xww  ÿÀŽ ä/ x22ÿø� ä t"ÏÝêö!    ä t  ÏÝêö    ät#5EYq    Ä t!$5EYq!    ä xTv  ÿÀŽ ä xww  ÿÀŽ äxÿ     ä xT      äO r   !    â $~!ÿ   ÿ Ž â r        â $~ÿ   ÿ Ž â xˆˆˆˆÿÀŽ â $~ÿ   ÿ Ž â x ww  ÿÀŽ Ä x22ÿø� ä tÏÝêö    ät!5EYq    ä tÏÝêö     ät#5EYq     ä xTv  ÿÀŽ Ä x ÿ     ä x d      äx1u      ä r       â &r   ÿ   â r        Æ &r      â x d      äx1u      Æ &r      ä $rÿÿ    Ž â x€  ÿà÷ â 0rÿ    A   ä $rÿÿ    Ž â x€  ÿàó â 0rÿ    A   ä &r      â rÿ   ÿäÿ  â $rÿÿ   Ž â  r     A  Ä Er        â G‰  @    € ê %x   Ž ä �y     ¨  �y%    è �y    ( �y    h �y    ¢ $~ÿ   ÿ Ž Ä $~#ÿ   ÿ Ž ä $~!ÿ   ÿ Ž ä $~"ÿ   ÿ Ž â x %ww  ÿÀŽ ä�x%%T      Èx%%ˆˆˆˆÿÀŽ ä x$ww  ÿÀŽ ä x%22ÿø� ä/ tÏÝêö     ät#5EYq     ä xTv  ÿÀŽ Ä t!!ÏÝêö$    ät""5EYq$    ä xÿ     ä x ww  ÿÀŽ äxT      äO r!   "    Ä r       â $~ÿ   ÿ Ž â x"ˆˆˆˆÿÀŽ â $~!ÿ   ÿ Ž â x#ww  ÿÀŽ â $~ÿ   ÿ Ž â |       ä x"22ÿø� ä tÏÝêö     Ä t !5EYq     ä tÏÝêö#    ät#5EYq#    ä x!Tv  ÿÀŽ ä xÿ     ä x d      Ä x1u      ä r!        â &r   ÿ   â r   #    Æ &r      â x d      äx1u      Æ &r      ä 0rÿ    A   ä &r      Æ  r     A  ä Er        Ê #r      â #r
   
   â $r	ÿÿ   	 â GÉ  Ð÷ÿÿÿÿƒ ê Ay        € ê r ÿ   pRð âEy  €     € â r ÿ   pRò Ö G‰  P     € ê z ÿ `  pPð â $x    Ž Æ z ÿ a   Sð ä ˆs 
€ÿÿ H   ö ˆ    H   ä Ay        € ê {          ì M        € ê z ÿ `  pPð â „y     H   f. z ÿ a   Sð Ú „‰  €   H   ¨ „‰    H   â !r
       È/ !‚         ÈO !‚        â� G‰        € ê „y  €  H   h „y €   H   ¨ „y    H   è „y	    H   ( „y €  H   " !r         â/ !r       ÆO !r         â� !r	       Æ !r        ä‰      b z ÿ `  pPð ä r ÿ   pRò ä z ÿ a   Sð â !r       Ê/ ‰       d !r         Ê/ ‰  €    d !r        Ê/ ‰ @    d!r       Ê/ ‰      d !r       â/ G‰        € ê ‰       d !r         Ê/ ‰       d !r        Ê/ ‰ €    d !r       Ê/ ‰ @    d !r       Ê/ ‰      d !r       ä/ M        € ê !r       â G‰  `    € ê $v ÿ h  ÿ Ž â !r       È x     pBð Ú G	  p    € ê x  ÿÿÿÿÿàÿ È x       € Ê $x     ÿ Ž È ‚{   €     d xÿ     È/ Iy 0íÿÿÿÿƒ ê  x ;ª¸¿  A  Ì s          d !t   €?    Ì/ s        d  r     A  È/  r     A  â Gy  P    € ê  x '7=  A  â x*BL?  A  Æ #t   €?    È  r      A  â  x    ?  A  Ê s     $   d !t  €?    È/  r      A  È  r     A  â Gy  °    € ê x     pRð Ú G‰  à     € ê x     pRð Ú G	        € ê 	z  i    � ä 	z i    � Æ  x ;ª¸¿  A  â 	z i €  � Ê s        d !t  €?    Ì/ s        d  r      A  È/  r     A  â Gy  Ð     € ê  r     A  â Gy  °     € ê 	x   à@  � ä 	x  à@  � Æ  x #ÛÙ?  A  â 	x  àÀ  � Æ  x;ª¸¿  A  ä !t  €?    È s        d !t  €?    Ì/ s        d  r       A  È/  r      A  ä y       &   h y      %   h y      '   ¢ $z   x  Ž ä/ $tÿ   ÿ Ž Ä $z  ~   Ž ÈO %v  j   Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y       &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž ä $rÿÿ     Ž È % \   Ž Ê �	    ™ b ¹z  l      â Ey       € â ™x?    â y      "   b $rÿÿ   ÿ Ž Æ y      !   ¢ $ÿÿ   ÿ Ž ä $ÿÿ    Ž â y	      %   & '
 m   Ž È $‚ÿÿ   ÿ Ž â 
ÿ n    Ê $ÿÿ   
Ž â r ÿ   pRò Æ/ $
 o  Ž â r ÿ   pRô â$x    Ž äO $‚ÿÿ    Ž ä $rÿÿ   ÿ Ž â r   ÿ� â 'Š s   Ž Æ xÿ     â $€   Ž ä $x		   ÿ Ž â |    pbö ä Šÿ t    ä x      ÿ  Ò G9  ð    € ê y      '   " $r
ÿÿ   ÿ Ž â xþÿÿÿÿÀŽ â $z w  ÿŽ ä $vÿ X  ÿ Ž ä $x   
Ž È $x   ÿ Ž ä 'z y  
 Ž È $z
 }  ÿŽ â zÿ z    â $z	 p  Ž Æ r
   ÿàñ â $z |  Ž ä $x
   ÿ Ž ä $z v  Ž ä$vÿ Y  ÿ Ž â xÿ   
  â %x   ÿŽ Æ x
     â $rÿÿ   ÿ Ž ä $rÿÿ   ÿ  ä %x   Ž Ê r   ÿàñ Ê $x     Ê �y   • ¨  �y   • â  $v'ÿ p  ÿ Ž Æ �y   • " %x'   Ž Æ �y.   • ( �y-   • h �y2   • " r   ÿàñ â $tÿ$   ÿ Ž Ä �y*   • $ $rÿÿ     ä �y)   • " %v Z   Ž È $x$   ÿŽ â r
   ÿ€ Æ $x   Ž ä %x '   Ž Æ �y    • b $x     Æ �y0    • h �y&   ™ h �y    • h �y(   ™ b %x"'    Ž Æ �y+    • h �y4    • h �y,   ™ h �y/"   • h �y3"   • h �y1"   • h �y5"   • b r   ÿàñ Æ �y   ™ d $x     È %x'   Ž ä/�y    • h %x$'   Ž ä�y    • h %x '   $Ž â �y$    • j �y      • b ‚x &     Ä ‚x �˜­¿     â $~"ÿ   ÿ Ž ä $~'ÿ   ÿ Ž ä $~ÿ   ÿ Ž ä/ $~!ÿ   ÿ Ž ä $~%ÿ   ÿ Ž â x@   ÿàÿ È |    pbð ä x€  ÿà÷ Ê $rÿÿ   Ž â x#ww  ÿÀŽ äOxT      È� xˆˆˆˆÿÀŽ â $~ÿ   ÿ Ž â xww  ÿÀŽ ä x22ÿø� ä t$5EYq    ä xTv  ÿÀŽ ä t'ÏÝêö#    Ä t""5EYq#    ä tÏÝêö    ä xÿ     ä r   "    â $~"ÿ   ÿ Ž â r   $    â $~$ÿ   ÿ Ž â x#T  .    Äxww  ÿÀŽ ä x d      äx1u      ä x##ˆˆˆˆÿÀŽ ä t"ÏÝêö    ät$5EYq    â $~ÿ   ÿ Ž â x".ww  ÿÀŽ â $~.ÿ   ÿ Ž â x-ww  ÿÀŽ Äx--T  2    ä x'#22ÿø� ä t$!ÏÝêö"    ät5EYq"    ä t"%ÏÝêö    ä x--ˆˆˆˆÿÀŽ Ä x#'Tv  ÿÀŽ â $~%ÿ   ÿ Ž â t.5EYq    ä x.2ww  ÿÀŽ ä x--22ÿø� ä r#       ä x#*ww  ÿÀŽ Ä t!!ÏÝêö.    äx**T  )    ä t.%5EYq.    ä x'ÿ   '  ä x%-Tv  ÿÀŽ â $~2ÿ   ÿ Ž â x**ˆˆˆˆÿÀŽ Ä x-ÿ   -  ä r$'       â $~$ÿ   ÿ Ž â r"%       â $~"ÿ   ÿ Ž â x))ww  ÿÀŽ â $~ÿ   ÿ Ž â x**22ÿø� ä r.!-   .    Ä t!2ÏÝêö#    ä t2$5EYq)    ä t""5EYq#    ä x$*Tv  ÿÀŽ ä t#ÏÝêö)    ä x*ÿ   *  Ä r"!$   "    ä r##*   2    ä x d      äx% d  .    äx1u      â &r&   ÿ   â x!1u  .    Ä x" d  #    äx"1u  #    â &r%%&   ÿ   â x"0ww  ÿÀŽ äx00T      â &r(      ä&r!(   %   â x00ˆˆˆˆÿÀŽ â $~!ÿ   ÿ Ž Ä $~#ÿ   ÿ Ž â x0022ÿø� â $~$ÿ   ÿ Ž â xww  ÿÀŽ â $~*ÿ   ÿ Ž â t!!ÏÝêö"    ät"#5EYq"    ä x#0Tv  ÿÀŽ ä t$$ÏÝêö    Ä t%*5EYq    ä x0ÿ   0  ä r!!#   "    â $~#ÿ   ÿ Ž â r$$0   %    â $~%ÿ   ÿ Ž â x"+ww  ÿÀŽ äx++T  4    â &r,      âx! d  $    â &r,      â x!1u  $    Ä x++ˆˆˆˆÿÀŽ â $~$ÿ   ÿ Ž â t#ÏÝêö"    ät!%5EYq"    ä x"/ww  ÿÀŽ äx%4ww  ÿÀŽ ä x//T  3    Ä x++22ÿø� â $~'ÿ   ÿ Ž â t$$ÏÝêö%    äx//ˆˆˆˆÿÀŽ ä t%*5EYq%    â $~*ÿ   ÿ Ž â x)+Tv  ÿÀŽ â $~.ÿ   ÿ Ž â x33ww  ÿÀŽ Ä x//22ÿø� ä t##ÏÝêö"    är)   !    ä t"'5EYq"    ä x!1ww  ÿÀŽ ä t'*ÏÝêö3    Ä x+ÿ   +  ä x11T  5    ä x*/Tv  ÿÀŽ ä t..5EYq3    ä x/ÿ   /  ä r%$+   %    â $~$ÿ   ÿ Ž â x11ˆˆˆˆÿÀŽ Ä r"#*   "    â $~*ÿ   ÿ Ž â r''/   .    â $~.ÿ   ÿ Ž â x55ww  ÿÀŽ â $~0ÿ   ÿ Ž â x1122ÿø� ä t#$ÏÝêö!    ät$*5EYq!    Ä t*.ÏÝêö5    äx)1Tv  ÿÀŽ ä t505EYq5    ä x1ÿ   1  ä x!" d  '    â&r&   ÿ   â x'"1u  '    Ä r##)   $    â &r&!&   ÿ   â r**1   5    â &r(      âx! d  %    â&r''(   &   â x# d  *    äx%1u  %    â &r!,      âx##1u  *    â &r,   '   Ä &r      ä&r      ä&r%      ä&r#      ä 0rÿ    A   ä 0rÿ    A   Ä 0rÿ    A   ä 0rÿ    A   ä 0r ÿ     A   â  r     A  â  r     A  â r     A  âEr        â  r      A  â Er        Ä Er        ä Er        â #r      â #r      â #r      ä #r      â G‰  pñÿÿÿÿƒ ê Ay        € ê ˆ  þÿ H   è ˆ €þÿ H   è ˆ  ÿÿ H   è ˆ €ÿÿ H   è {          ì M)        € ê „y     H   " z 	 r  p`ô â $z   x  	Ž Ä „y €   H   b r ÿ   pVt Æ „y    H   ¨ „y    H   è „y €  H   („y    H   ( „y
    H   ( „y €  H   ( „y €  H   ( „y    H   ( „y €  H   " !r       Æ „y €  H   " !r       â/ !r       äO y      '   b !r       â� !r       â!r       ä !r
       â !r       È ‰      ¢ !r       â $z  ~   Ž â/ !r       Æ ‰      b !r       Æ ‰      â !r       Ê ‰      " !r       âO !r       â/ !r       È� ‰      h ‰      ¢ !r       Æ ‰      ( ‰      â !r       â/ !r       ÈO ‰
 €    b !r       Æ ‰ €    " !r       Æ� ‰ €    ¨ ‰ €    â !r

       â/ !r       È ‰
 @    " x	   ÿàÿ â !r       äO ‰ @    b x	   ÿàÿ â !r       ä� ‰ @    ¢ z  r  p`ö ä z  r  p`ò â ‰ @    â x    pVö Ä x    pVò  â !r
       â x
	   ÿàÿ â $t	ÿ   ÿ Ž â !r       ä/ ‰      " z 
 r  p`ð â !r       äO ‰      b x    pVp  â !r       Ä� ‰      ¨ ‰      â !r       â !r       â/ %v  j  	 Ž â !r       ÈO †©     â !r       Æ� †¹    è †‰    â M        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y	      &   " z ÿ \  pPò â ¹z  F   
   Æ z ÿ ]  Sò Ú         Ê %	 \   Ž Ê �    ™ ¢  r ÿ       â Ey       € ä y      "   d '	 m   Ž äy      %   æ ÿ n    â y       !   & ÿ   ÿáÿ â y      '   " 'š	 s   Ž È $ o  	Ž â ’ 	       ä r ÿ   pRð ä/ x   ÿ   ä� x     pDp  är    ÿàÿ är    ÿ(Ž Ä z  r  pdp  ä r ÿ       ä šÿ t    ä r ÿ       ä r	     € ÊO G	  À    € ê z ÿ `  pPò ä z ÿ ^  pPð ä z ÿ b  p ô ä z ÿ a  Sò ä z ÿ _   Sð ä z ÿ c   'ò Ö $
 ~  ÿŽ âÿ     â$š	 ~  ÿŽ ä $
 x  ÿŽ â     àë â$š
 x  ÿŽ â ’		    àÇ ä ˜ÿ     ä    ÿàõ Ä ÿ   ÿÄþ ä ’
	   ÿàû ä ’
ÿ   ÿ„þ ä 	ÿ   ÿä ä 
 ^  ÿ„ ä ’
ÿ
   ÿäÿ Ä š b  ÿ† ä
	 _  	 ä r ÿ       ä r ÿ       ä š c  
� ä �	     h �™     d Ay        € ê r   ÿ� â ¹z  l      â $z }  ÿŽ â ™x?    â xÿ     â $z w  ÿŽ â r ÿ       â Ey  p    € â r ÿ       â x
      ÿ  â |    pbð â 'z y   Ž â r   ÿàó â x      ÿ  â r. ÿ       â x      ÿ  â rÿÿ   ÿäÿ  Î G	  À    € ê xþÿÿÿÿÀŽ â $z
 p  Ž â zÿ z    ä r  €ÿàÿ ä x	   ÿàÿ â $z
 |  
Ž âx$   ÿ   â $z |  Ž â x   ÿàÿ â $z v  
Ž â x   ÿàÿ â %x$$   ÿŽ â x   ÿ   â x
      ÿ  â z ÿ `  pPð â $z v  Ž â xÿ     â %x   $Ž â z  `      Ä z  a      â $z p  Ž âx   ÿàó â $z	 p  Ž âz  X      â $z p  Ž â z  Y      â %x    $Ž âx      ÿ  â r ÿ       â x      ÿ  â %x"   $Ž â r. ÿ       Ä z ÿ a   Sð â %x$   $Ž â x     ä rÿ   ÿäÿ  ä r,   ÿàó È r-   ÿäÿ  Ê �y4,   • ¨  �y5,   • â  r&   ÿàó ä x' $       â �y/,   • " rÿ   ÿäÿ  Æ %v&& Z  ' Ž â �y6,   • & $x$   ÿŽ â r&   ÿ‚ ä r(   ÿàõ ä r''   ÿàÿ ä r)   ÿä Ä r'   ÿäÿ  ä z7  p      â %x,   Ž â �yA&    • & %x*7   (Ž â�y   ™ ( �y2   ™ ( �y8*úÿÿ • ( �y9*üÿÿ • ( �y3   ™ ( �y:*þÿÿ • ( �y;*    • " %x(7   *Ž Æ �y<   ™ ( �y=(úÿÿ • ( �y@(üÿÿ • ( �y>,    • " %x*7   (Ž Æ �y?(þÿÿ • ( �yB(    • ( �yD*úÿÿ • ( �yE*üÿÿ • ( �yC*þÿÿ • ( �y**    • " %x07   ,Ž Ì %x7   0Ž ä/�y00    • ( %x&7   Ž ä �y    • ( �y&&    • " ‚x �˜­¿     Ä ‚x &     â |-       ä |)       ä |(       ä |+       ä |1       ä/ x,4ww  ÿÀŽ ÄO x44T  5    È� x44ˆˆˆˆÿÀŽ È x4422ÿø� ä t-ÏÝêö,    ät,)5EYq,    ä x'4Tv  ÿÀŽ ä x55ww  ÿÀŽ ä r'   ,    Ä t((ÏÝêö5    ät++5EYq5    ä x4ÿ   4  ä x,/ww  ÿÀŽ äx//T  6    ä r4(4   +    Ä x//ˆˆˆˆÿÀŽ ä |+       ä x( d  4    ä |)       ä |'       ä x66ww  ÿÀŽ Ä x//22ÿø� ä x1u  4    ä |-       ä t4+5EYq,    ät))ÏÝêö,    â &r,(   ÿ   â t+'ÏÝêö6    Ä x'/Tv  ÿÀŽ ä t6-5EYq6    ä x/ÿ   /  â &r,2   ,   â r))'   4    ä r(+/   6    ä |       Ä x48ww  ÿÀŽ ä|+       ä x88T  9    ä tÏÝêö4    ät4+5EYq4    ä x88ˆˆˆˆÿÀŽ Ä x+) d  (    ä |'       ä x69ww  ÿÀŽ â &r,+3   ,   â x8822ÿø� ä t''ÏÝêö6    ä x+:ww  ÿÀŽ Ä t6-5EYq6    ä x::T  ;    ä x-8Tv  ÿÀŽ ä x8ÿ   8  ä x::ˆˆˆˆÿÀŽ ä x))1u  (    Ä r-   4    ä r('8   6    ä |4       ä |6       ä |8       ä x;;ww  ÿÀŽ Ä x::22ÿø� ä |/       ä t'4ÏÝêö+    ät665EYq+    ä x4 d  (    ä t-8ÏÝêö;    Ä x+:Tv  ÿÀŽ ä t8/5EYq;    ä x:ÿ   :  â &r44   ÿ   â x1u  (    â &r))<   ,   â r''+   6    ä r8-:   8    Ä x,=ww  ÿÀŽ ä|-       ä x==T  @    â &r42   4   â x+' d  8    äx'1u  8    â 0r(ÿA    A   â t'-ÏÝêö,    â 0r-ÿ>    A   â x==ˆˆˆˆÿÀŽ Ä x@@ww  ÿÀŽ â &r++3   4   â x==22ÿø� â  r--(     A  â Er) )       ä x4?ww  ÿÀŽ ä t11ÏÝêö@    ä x??T  B    Ä t,/5EYq,    ät@/5EYq@    ä x/=Tv  ÿÀŽ â #r.-)   .   â x??ˆˆˆˆÿÀŽ â &r<   +   â |)       ä |-       Ä x8Bww  ÿÀŽ ä r''/   ,    ä |+       ä |/       ä x6Dww  ÿÀŽ äxDDT  E    Ä x??22ÿø� ä t))ÏÝêö4    ä t--ÏÝêö8    ä x=ÿ   =  ä t4+5EYq4    ät8+5EYq8    Ä t+/ÏÝêö6    ä xDDˆˆˆˆÿÀŽ ä x/?Tv  ÿÀŽ ä x?ÿ   ?  ä r@1=   @    ä |1       Ä xDD22ÿø� ä r8-?   8    ä x-Cww  ÿÀŽ ä t615EYq6    ä |:       ä xEEww  ÿÀŽ Ä xCCT  *    ä |>       ä x1DTv  ÿÀŽ ä t5:ÏÝêöE    ä r))/   4    ä xCCˆˆˆˆÿÀŽ Ä x,' d  @    ä t:>5EYqE    ä xDÿ   D  ä r++1   6    ä |4       ä x''1u  @    Ä |6       ä |>       ä |@       ä x/*ww  ÿÀŽ ä xCC22ÿø� ä r:5D   :    Ä t*4ÏÝêö-    ät-65EYq-    ä t1>ÏÝêö/    ät6@5EYq/    ä x/CTv  ÿÀŽ ä xCÿ   C  Ä x4+ d  :    â&r,,   ÿ   â x++1u  :    ä r-*/   -    â &r44   ÿ   â r61C   6    â &r,'2   ,   âx*) d  8    â &r4+2   4   â x/- d  6    â 0r'ÿ0    A   â x))1u  8    â &r**3   ,   âx--1u  6    â &r//3   4   Ä 0r+ÿ    A   â  r('     A  â&r))<   *   âEr        â 0r'ÿ&    A   ä &r--<   /   â  r*(+     A  â Er, )       â #r      â  r+('     A  â Er) -       â G‰  `    € ê r   ÿàó È r   ÿäÿ  Ê �y/    ¨ �y4    è �y0    ( �y-    " %x&   Ž Ì �y&&     " |6       ä |8       ä |5       ä |7       ä |:       ä x1/ww  ÿÀŽ ÄO x//T  4    ä� t6ÏÝêö1    äx//ˆˆˆˆÿÀŽ ä t185EYq1    ä x//22ÿø� ä x44ww  ÿÀŽ Ä x/Tv  ÿÀŽ ä t'5ÏÝêö4    ä/r   1    ä x0ww  ÿÀŽ ät475EYq4    ä x/ÿ   /  Ä x00T  -    ä r''/   4    ä x00ˆˆˆˆÿÀŽ ä |4       ä |6       ä |8       Ä x/-ww  ÿÀŽ ä x0022ÿø� ä t-4ÏÝêö    ät465EYq    ä t18ÏÝêö/    ät6:5EYq/    Ä x0Tv  ÿÀŽ ä x/ÿ   0  ä x0 d  '    äx'1u  '    ä r-   4    â &r00   ÿ   â r61/   6    Æ &r0'2   0   â x d  6    äx1u  6    Æ &r3   0   È &r<      ä 0rÿ&    A   Æ Er&        ä  r(     A  È #r&      ä G‰  `    € ê r    ÿàó È r!   ÿäÿ  Ê �y/    ¨ �y4    è �y0    ( �y-    " %x&   Ž Ì �y&&     " |6       ä |8       ä |5       ä |7       ä |:       ä x1/ww  ÿÀŽ ÄO x//T  4    ä� t6ÏÝêö1    äx//ˆˆˆˆÿÀŽ ä t185EYq1    ä x//22ÿø� ä x44ww  ÿÀŽ Ä x/Tv  ÿÀŽ ä t'5ÏÝêö4    ä/r   1    ä x0ww  ÿÀŽ ät475EYq4    ä x/ÿ   /  Ä x00T  -    ä r''/   4    ä x00ˆˆˆˆÿÀŽ ä |4       ä |6       ä |8       Ä x/-ww  ÿÀŽ ä x0022ÿø� ä t-4ÏÝêö    ät465EYq    ä t18ÏÝêö/    ät6:5EYq/    Ä x0Tv  ÿÀŽ ä x/ÿ   0  ä x0 d  '    äx'1u  '    ä r-   4    â &r00   ÿ   â r61/   6    Æ &r0'2   0   â x d  6    äx1u  6    Æ &r3   0   È &r<      ä 0rÿ&    A   Æ Er        ä  r(     A  È #r
   
   ä G‰  `    € ê r&$   ÿàó È r'%   ÿäÿ  Ê �y/&    ¨ �y4&    è �y0&    ( �y-&    " %x   Ž Ì �y     " |6       ä |8       ä |5       ä |7       ä |:       ä x1/ww  ÿÀŽ ÄO x//T  4    ä� t&6ÏÝêö1    äx//ˆˆˆˆÿÀŽ ä t185EYq1    ä x//22ÿø� ä x44ww  ÿÀŽ Ä x/Tv  ÿÀŽ ä/ t'5ÏÝêö4    är&   1    ä x&0ww  ÿÀŽ ät475EYq4    ä x/ÿ   /  Ä x00T  -    ä r4'/   4    ä x00ˆˆˆˆÿÀŽ ä |/       ä |8       ä x--ww  ÿÀŽ Ä |'       ä x0022ÿø� ä t6/5EYq&    ät/8ÏÝêö-    ät8:5EYq-    ä t''ÏÝêö&    Ä x-0Tv  ÿÀŽ ä x0ÿ   0  ä x& d  4    äx1u  4    ä r''-   6    â &r&&   ÿ   â r0/0   8    Æ &r&2   &   â x-' d  0    â0rÿ    A   â x''1u  0    Æ &r&-3   &   â  r(     A  Æ &r&'<   &   Ê Er& &       Ê #r	&   	   â #r*,      â #r+)      â G‰  `    € ê r"   ÿàó È r#   ÿäÿ  Ê �y+    ¨ �y,    è �y)    ( �y*    " %x&   Ž Ì �y&&     " |0       ä |4       ä |/       ä |1       ä x-+ww  ÿÀŽ äOx++T  ,    Ä� t0ÏÝêö-    äx++ˆˆˆˆÿÀŽ ä t-45EYq-    ä x++22ÿø� ä x,,ww  ÿÀŽ ä x+Tv  ÿÀŽ Ä t'/ÏÝêö,    ä/r   -    ä x)ww  ÿÀŽ äx))T  *    ä t,15EYq,    ä x+ÿ   +  Ä x))ˆˆˆˆÿÀŽ ä r''+   ,    ä |,       ä |0       ä |-       ä |/       Ä x**ww  ÿÀŽ ä x))22ÿø� ä t+,ÏÝêö    ät,05EYq    ä t--ÏÝêö*    ät0/5EYq*    Ä x)Tv  ÿÀŽ ä x)ÿ   )  ä x* d  '    äx'1u  '    ä r+   ,    â &r**   ÿ   â r0-)   0    Æ &r*'2   *   â x d  0    äx1u  0    Æ &r3   *   È &r<      ä 0rÿ&    A   Æ Er        ä  r(     A  È #r      â x@   ÿàÿ ä x€  ÿàõ ä |    pbò ä x€  ÿà÷ ä rÿ   ÿä ä rÿ   ÿäÿ Î D       À â Gy  €ãÿÿÿÿƒ ê Ay        € ê y      "   " Ey  0    € â r ÿ   pRð är ÿ   pRò Ö G‰  ð     € ê z ÿ `  pPð ä r    ÿ8Ž ä z ÿ a   Sð Æ ˆs . þÿ H   ô ˆƒ €þÿ H   è ˆƒ  ÿÿ H   è ˆƒ €ÿÿ H   â G‰  p     € ê ˆs    H   è ˆs €þÿ H   è ˆs 
€  H   è ˆs  ÿÿ H   è ˆs 	   H   è ˆs €ÿÿ H   è ˆs €  H   ä Ay        € ê {          ì M        € ê y      &   ¢ z ÿ `  pPð Æ y      '   "> z ÿ a   Sð â $z x  Ž âO x        Æ $z ~  Ž È %v j   Ž È G	  P    € ê „y      H   " z  r  p`ð Æ „y
     H   b r  ÿ   pVp  Æ „y     H   ¨„y  €   H   è „y     H   ( „y	     H   ( „y  €  H   ( „y     H   ( „y  €  H   ( „y  €  H   ( „y  €  H   " !r.       Æ „y  €  H   " !r

       â/ !r       âO !r       ä� !r

       â!r	       â !r       È ‰      b !r

       â !r       È ‰
      ¢ !r       Æ ‰      â !r       È !r       â !r       È/ ‰      ( ‰      b !r
       âO !r	       È� ‰      ¨ ‰	      â !r       â !r       È/ ‰      " !r       ÆO ‰ €    b !r	       Æ� ‰ €    ¨ ‰ €    â !r       Ê ‰ €    " !r       â/ !r       ÈO ‰ @    b !r       â� x   ÿàÿ ä ‰ @    ¤ z  r  p`ô ä ‰
 @    ä x     pVt â !r       Ê ‰ @    " !r       â/ x   ÿàÿ äx   ÿàÿ â !r       âO z  r  p`ò â ‰      b !r

       ä� x     pVò  â ‰      è ‰	
      ¢ !r       â !r       È/ ‰      " !‚       â !r
	       ÈO †‰     â z  r  p`ð â !r	       â !’       äx     pVp  â !¢		       ä †™    è †© 	   â !r       Ê M	        € ê !r       Ê †y    â My        € ê z  h      È x    pBð Ú G	  @    € ê x    pRð Ú G‰  À    € ê x    pRð Ú G	   $    € ê „y      H   " z  r  p`ð â Ey  p    € ä „y     H   b r  ÿ   pVp  Æ „y     H   ¨ „y     H   è „y     H   ( „y   
  H   ( „y  €   H   ( „y  €  H   ( „y  €  H   ( „y  €  H   â !r.       â !r       Æ/ !r       âO !r       Æ� !r       â!r       È ‰      " !r       Æ ‰      b !r
       È !r       â !r       â !r       È/ ‰      ( ‰      b !r       â !r       È/ ‰ €    ( ‰ €    h „y  €  H   ¢ !r       Æ „y  €
  H   â !r       Æ/ ‰ @    ( ‰ @    b !r       ÈO !r       â !r       â !r       È/ ‰      ( ‰      b !r       â !r       â !r       â/ G	  �     € ê !r
       â!r       Æ  x
;ª¸¿  A  Ì s        $ !t  €?    Ì s        $  r

     A  È  r
     A  Ê †y     ä Ay        € ê ‰      " Ey  Ð    € æ ‰
      h „y     H   ¨ „y     H   è „y     H   è „y     H   â !r       â !r
       È/ ‰
      ( ‰      b !r       âO !r	       â� !r

       ä !r       â !r       Ä/ ‰
 €    " !r       Æ ‰ €    h „y   	  H   ¢ !r
       Æ „y
     H   â !r       â/ x   ÿàÿ ä ‰ @    $ z  r  p`ð ä ‰ @    d x     pVp  â !r       âO !r       â !r       È/ ‰      ( ‰      b !r
       â !r       â !r       â/ G	  �     € ê !r       â!r       Æ  x	;ª¸¿  A  Ì s	 	       $ !t
	  €?    Ì s
 
       $  r
     A  È  r     A  Ê †y    ä Ay        € ê ‰      " Ey       € æ ‰
      h „y  €  H   ¨ „y  €	  H   è „y  €  H   ( „y  €  H   â !r       Æ „y  €  H   â !r	
       Æ/ „y  €  H   è ‰
      ( ‰	      b !r       ÈO !r       È� !r       â!r

       â !r	       È/ ‰
 €    ( ‰ €    b !r
       â !r       È/ ‰ @    ( ‰ @    b !r       â x   ÿàÿ â !r       Æ/ ‰      " z  r  p`ð Æ ‰
      b x     pVp  â !r       â !r
       Ö/ G	  �     € ê !r       â!r       Æ  x;ª¸¿  A  Ì s        $ !t	  €?    Ì s
 	       $  r

     A  È  r
     A  Ê †y    ä Ay        € ê ‰      b !r       â x   ÿàÿ Æ !r       â z  r  p`ð Æ !r       â x     pVp  È ‰      " !r	       Ê/ ‰	      b !r       Ê ‰      " !r
	       Ê/ ‰
 €    b !r       Ê ‰ €    " !r
       Ê/ ‰ @    b !r       Ê ‰ @    " !r       Ê/ ‰	      b !r       Ê ‰      ¢  !r		       â/ M	        € ì !r	       â!r       ÆO  x ;ª¸¿  A  â !r       Ê s          $ !t   €?    Ì s        $  r     A  È  r     A  Ê †y    â My        € ê „y      H   " z  r  p`ð â Ey       € ä „y     H   b r  ÿ   pVp  Æ „y     H   ¨ „y     H   è „y     H   ( „y   
  H   ( „y  €  H   ( „y  €  H   ( „y  €  H   è „y  €
  H   " !r.       â !r       Æ/ !r       âO !r       Æ� !r       â!r       È ‰      " !r
       Æ ‰      b !r       È !r       â !r       â !r       È/ ‰      ( ‰      h „y  €   H   ¢ !r       â !r       È/ ‰ €    ( ‰ €    h „y  €  H   â !r       âO !r       â !r       È/ ‰ @    ( ‰ @    b !r       È� !r       â !r       â !r       È/ ‰      ( ‰      b !r       â !r       â/ G	  À     € ê !r       È x'7=  A  â x
*BL?  A  Æ #t  €?   È  r
     A  â  x
   ?  A  â !r       È s     $   $ !t  €?    È  r

     A  È  r
     A  Ê †y     ä Ay        € ê ‰      " Ey       € æ ‰
      h „y     H   ¨ „y     H   è „y   	  H   è „y     H   â !r       Æ „y     H   â !r
       Æ/ ‰
      ( ‰      b !r       âO !r	       â� !r

       Æ !r       â !r       â/ ‰
 €    $ !r       ä ‰ €    h „y     H   ¢ !r
       â x
   ÿàÿ â !r       Æ/ ‰ @    " z 
 r  p`ð Æ ‰ @    b x     pVp  â !r       ÈO !r       â !r       â !r       È/ ‰      ( ‰      b !r	       â !r       â/ G	  À     € ê !r       È x
'7=  A  â x*BL?  A  Æ #t
  €?
   È  r

     A  â  x   ?  A  â !r	       È s
 
    $   $ !t
  €?    È  r     A  È  r     A  Ê †y    ä Ay        € ê ‰      " Ey  Ð    € æ ‰      h „y  €  H   ¨ „y  €	  H   è „y  €  H   è „y  €  H   â !r       Æ „y  €  H   â !r	       Æ/ ‰
      ( ‰	      h „y  €  H   " !r       ÈO !r       â� !r

       â !r	       È/ ‰
 €    " x	   ÿàÿ Æ ‰ €    b z 	 r  p`ð È x     pVp  â !r       â!r
       â !r       È/ ‰ @    ( ‰ @    b !r       â !r       È/ ‰      ( ‰      b !r	       â !r       â/ G	  À     € ê !r       È x'7=  A  â x*BL?  A  Æ #t  €?   È  r     A  â  x   ?  A  â !r	       È s     $   $ !t
  €?    È  r

     A  È  r
     A  Ê †y    ä Ay        € ê ‰      b !r       â x   ÿàÿ Æ !r       â z  r  p`ð Æ !r       â x     pVp  È ‰      " !r	       Ê/ ‰	      b !r       Ê ‰      " !r
	       Ê/ ‰
 €    b !r       Ê ‰ €    " !r
       Ê/ ‰ @    b !r       Ê ‰ @    " !r       Ê/ ‰	      b !r       Ê ‰      ¢  !r		       â/ M	        € ì !r	       È x'7=  A  â x *BL?  A  Æ #t  €?   â x   ?  A  Æ  r       A  â !r       ÈO !r       â s       $   $ !t   €?    È  r     A  È  r     A  Ê †y    â My        € ê x    pRð Ú G‰  P    € ê x    pRð Ú G	  `    € ê „y      H   " z  r  p`ð â Ey       € ä „y     H   b r  ÿ   pVp  Æ „y     H   ¨ „y     H   è „y     H   ( „y   
  H   ( „y  €  H   ( „y  €  H   ( „y  €  H   è „y  €
  H   " !r.       â !r       Æ/ !r       âO !r       Æ� !r       â!r       È ‰      " !r
       Æ ‰      b !r       È !r       â !r       â !r       È/ ‰      ( ‰      h „y  €   H   ¢ !r       â !r       È/ ‰ €    ( ‰ €    h „y  €  H   â !r       âO !r       â !r       È/ ‰ @    ( ‰ @    b !r       È� !r       â !r       â !r       È/ ‰      ( ‰      b !r       â !r       â/ G	  À     € ê !r       Ê	z
 i    � â !r       È  x
;ª¸¿  A  â 	z i    � È 	z i €  � â s        $ !t  €?    Ì s        $  r

     A  È  r
     A  Ê †y     ä Ay        € ê ‰      " Ey       € æ ‰
      h „y     H   ¨ „y     H   è „y   	  H   â !r       Æ „y     H   â !r
       Æ/ ‰
      ( ‰      h „y     H   â !r       âO !r	       â� !r

       Æ !r       â !r       â/ ‰
 €    ( ‰ €    b !r       Æ „y     H   ¢ !r
       â x
   ÿàÿ â !r       Æ/ ‰ @    " z 
 r  p`ð Æ ‰ @    b x     pVp  â !r       ÈO !r       â !r       â !r       È/ ‰      ( ‰      b !r       â !r       â/ G	  À     € ê !r       Ê	z i    � â !r       È  x	;ª¸¿  A  â 	z i    � È 	z i €  � â s	 	       $ !t
	  €?    Ì s
 
       $  r
     A  È  r     A  Ê †y    ä Ay        € ê ‰      " Ey  Ð    € æ ‰      h „y  €  H   ¨ „y  €  H   è „y  €  H   è „y  €  H   â !r       Æ „y  €	  H   â !r	       Æ/ ‰
      ( ‰	      h „y  €  H   â !r       âO !r

       â !r	       È/ ‰
 €    " x	   ÿàÿ Æ ‰ €    b z 	 r  p`ð â !r       Æ x     pVp  â !r       â� !r
       â !r       È/ ‰ @    ( ‰ @    b !r       â !r       È/ ‰      ( ‰      b !r       â !r       â/ G	  À     € ê !r       Ê	z i    � â !r       È  x;ª¸¿  A  â 	z i    � È 	z i €  � â s        $ !t	  €?    Ì s
 	       $  r

     A  È  r
     A  Ê †y    ä Ay        € ê ‰      b !r       â x   ÿàÿ Æ !r       â z  r  p`ð Æ !r       â x     pVp  È ‰      " !r	       Ê/ ‰	      b !r       Ê ‰      " !r
	       Ê/ ‰
 €    b !r       Ê ‰ €    " !r
       Ê/ ‰ @    b !r       Ê ‰ @    " !r       Ê/ ‰	      b !r       Ê ‰      ¢  !r		       â/ M	        € ì !r	       â!r       ÈO 	z i    � â !r       È  x ;ª¸¿  A  â 	z i    � È 	z i €  � â s          $ !t   €?    Ì s        $  r     A  È  r     A  Ê †y    â My        € ê „y     H   " z  r  p`ð Æ „y     H   b r  ÿ   pVp  Æ „y  €   H   ¨ „y      H   è „y   
  H   ( „y  €  H   ( „y     H   ( „y     H   ( „y     H   ( „y  €  H   ( „y  €  H   " !r       Æ „y     H   " !r       Æ/ „y   	  H   b !r       ÆO „y     H   ¢ !r.       Æ� „y  €  H   â !r       Æ„y  €
  H   " !r
       Æ „y      H   " !r       Æ „y"     H   " !r	       Æ „y  €  H   " !r       Æ „y  €  H   " !r       Æ „y$  €  H   " !r       â x   ÿàÿ ä „y&  €	  H   â !r       â z  r  p`ô ä „y(  €  H   â !r       â/ x     pVt Ä „y*  €  H   â !r       ÆO ‰      " !r       â� !r       ä‰      b !r        Æ ‰	      ¢ !r"       Æ ‰
      â !r       Æ ‰      " !r       Æ ‰      " !r$       â !r       â !r&       â !r       Æ/ ‰      " !r((       â !r		       äO ‰      b !r*       â !r
       ä� ‰(      ¢ !r       Æ‰      â x   ÿàÿ â !r       ä ‰
	      " z  r  p`ò ä x   ÿàÿ â ‰      " x     pVò  ä z  r  p`ö â ‰      " !r       Ä x     pVö â ‰      " !r       â/ !r(       äO ‰ €    b !r       Æ� ‰ €    ¢ !r
	
       Æ‰      â !r       Æ ‰      " !r       Æ ‰	
 €    " !r       Æ ‰ €    ( ‰ €    " !r       Æ/ ‰ €    b !r       âO !r       ä� ‰ @    ¢ !r       Æ‰ €    â !r	
	       Æ ‰ €    " !r       Æ ‰ @    " !r       Æ ‰
	 @    " !r       Æ/ ‰ @    h ‰ @    b !r       ÆO ‰ @    ¢ !r       â� !r       ä‰      â !r       Æ ‰ @    " !r
	
       Æ ‰ @    " !r       Æ/ ‰      b !r       Æ ‰	
      b !r       ÆO ‰      ¨ ‰      ¢ !r       Æ� ‰      â !r       â !r       È‰      " !r       â/ !‚       ä‰      b !r
	       â !‚       â!r       äO !’       â  ‚     A  â !r       â !’       Æ †‰     â !r       â� !¢       â  ’     A  ä !¢
       â !r       ä †™    â  ¢
     A  â !r       È/ †©    â M9        € ê !r       â !r       ÈO  r     A  Ê †y    â My        € ê „y      H   " z  r  p`ð â Ey  À    € ä „y     H   b r  ÿ   pVp  Æ „y     H   ¨ „y     H   è „y     H   ( „y   
  H   ( „y  €  H   ( „y  €  H   ( „y  €  H   è „y  €
  H   " !r.       â !r       Æ/ !r       âO !r       Æ� !r       â!r       È ‰      " !r
       Æ ‰      b !r       È !r       â !r       â !r       È/ ‰      ( ‰      h „y  €   H   ¢ !r       â !r       È/ ‰ €    ( ‰ €    h „y  €  H   â !r       âO !r       â !r       È/ ‰ @    ( ‰ @    b !r       È� !r       â !r       â !r       È/ ‰      ( ‰      b !r       â !r       â/ G	  à     € ê !r       Ê	x
  à@  � Ê  x
#ÛÙ?  A  È  x;ª¸¿  A  â !r       Ê s        " 	x  à@  � È 	x  àÀ  � Ê !t  €?    â !t  €?    È s        $  r

     A  È  r
     A  Ê †y     ä Ay        € ê ‰      " Ey       € æ ‰
      h „y     H   ¨ „y     H   è „y   	  H   â !r       Æ „y     H   â !r
       Æ/ ‰
      ( ‰      h „y     H   â !r       âO !r	       â� !r

       Æ !r       â !r       â/ ‰
 €    ( ‰ €    b !r       Æ „y     H   ¢ !r
       â x
   ÿàÿ â !r       Æ/ ‰ @    " z 
 r  p`ð Æ ‰ @    b x     pVp  â !r       ÈO !r       â !r       â !r       È/ ‰      ( ‰      b !r       â !r       â/ G	  à     € ê !r       Ê	x  à@  � Ê  x#ÛÙ?  A  È  x	;ª¸¿  A  â !r       Ê s	 	       " 	x  à@  � È 	x  àÀ  � Ê !t  €?    â !t
	  €?    Ì s
 
       $  r
     A  È  r     A  Ê †y    ä Ay        € ê ‰      b Ey  ð    € æ ‰      ( „y  €  H   ¨ „y  €  H   è „y  €  H   è „y  €  H   â !r       Æ/ „y  €	  H   â !r	       Æ ‰
      ( ‰	      h „y  €  H   â !r       âO !r

       â !r	       È/ ‰
 €    " x	   ÿàÿ Æ ‰ €    b z 	 r  p`ð â !r       Æ x     pVp  â !r       â� !r
       â !r       È/ ‰ @    ( ‰ @    b !r       â !r       È/ ‰      ( ‰      b !r       â !r       â/ G	  à     € ê !r       Ê	x  à@  � Ê  x#ÛÙ?  A  È  x;ª¸¿  A  â !r       Ê s        " 	x  à@  � È 	x  àÀ  � Ê !t
  €?    â !t	  €?    È s 	       $  r     A  È  r
     A  Ê †y    ä Ay        € ê ‰      " !r       â x   ÿàÿ Æ !r       â z  r  p`ð Æ !r       â x     pVp  È ‰      b !r	       Ê ‰	      " !r       Ê/ ‰      b !r
	       Ê ‰
 €    " !r       Ê/ ‰ €    b !r
       Ê ‰ @    " !r       Ê/ ‰ @    b !r       Ê ‰	      " !r       Ê/ ‰      ¢ !r		       â M	        € ì !r	       â/!r       ÈO 	x  à@  � â !r       È  x #ÛÙ?  A  â 	x  à@  � Æ  x  ;ª¸¿  A  â 	x  àÀ  � Ê s          " !t  €?    â !t   €?    Ì s        $  r     A  È  r     A  Ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y       "   $ z   w  p`ð Ú M	        € ê y      !   " ¹z  l      â Ey  @    € â ™x?    â y      %   b ¹z  F   
   â $rÿÿ   ÿ Ž ä $rÿÿ   ÿ Ž â xÿ     È |    p`ð â $x   ÿ Ž Ø/ G	   
    € ê y      &   " $tÿ   ÿ Ž ä $z  x  Ž È %v \   Ž Ê �y    ™ ¢  $rÿÿ   ÿ Ž â x   ÿ   â $z	  r  ÿŽ â x   ÿàÿ â 'z m   Ž È $z q  Ž â zÿ n    â $rÿÿ   ÿ Ž ä $vÿ Y  ÿ Ž ä $rÿÿ   
Ž ä %x	$   ÿ Ž È $z o  Ž ä %x$    Ž â x   ÿÀŽ Æ $z u  ÿŽ â z Z  ÿàñ â %x   ÿ Ž Æ r   	à£ â$rÿÿ   ÿ Ž â z [  ä  â %x$   ÿ Ž â r	ÿÿ   ÿDþ  Æ %x$    Ž È $x	$   ÿŽ â z	 Z  ÿàñ â $rÿÿ    Ž ä $rÿÿ    Ž â z [  ä  â $vÿ X  ÿ Ž ä $z t  Ž äO$z t  Ž â x	   ÿàñ â $z q  Ž Ä %x   Ž È %x   Ž È $r#ÿÿ     ä r   ÿàñ Ê $x     Ê �y%   • ¨  �y&   • â  r   ÿàñ Æ �y   • $ $x     ä �y   • h �y   • ( �y   • ( �y   • ( �y!   • " $rÿÿ    Ž Ä $rÿÿ    Ž ä/ $rÿÿ   # Ž â r   ÿàñ È �y ðÿÿ ™ b $rÿÿ    Ž Æ �y"    ™ b %x
   Ž Æ �y#ôÿÿ ™ b $x	     ä %x   Ž â �y

    • h �y    • h �y$   ™ h �y    • b	 ‚x �˜­¿     Ä ‚x &     â $~'ÿ   ÿ Ž ä $~ÿ   ÿ Ž ä $~)ÿ   ÿ Ž ä $~+ÿ   ÿ Ž ä $~*ÿ   ÿ Ž ä $~,ÿ   ÿ Ž Ä $~-ÿ   ÿ Ž â x   ÿàÿ È |    p`ð ä x@  ÿàó ä x   ÿàõ Æ $rÿÿ   Ž  ä $rÿÿ    â x(%ww  ÿÀŽ äOx%%T  &    ä� x&&ww  ÿÀŽ ä x%%ˆˆˆˆÿÀŽ ä t''ÏÝêö&    Ä x%%22ÿø� ä t&5EYq&    ä t	)ÏÝêö(    ä/x%Tv  ÿÀŽ ä t(+5EYq(    ä x%ÿ   %  Ä r	   (    â $~(ÿ   ÿ Ž â r&'%   &    ä xww  ÿÀŽ äx)T      äx	 d  &    äx1u  &    â $~'ÿ   ÿ Ž â t%(ÏÝêö    Ä t&*5EYq    â $~*ÿ   ÿ Ž â xww  ÿÀŽ ä x))ˆˆˆˆÿÀŽ ä xww  ÿÀŽ ä xT      ä x))22ÿø� Ä t'ÏÝêö    ät(+5EYq    ä xˆˆˆˆÿÀŽ ä t*ÏÝêö    ät,5EYq    ä x*ww  ÿÀŽ Ä x,ÿ   )  ä x+22ÿø� ä t'ÏÝêö*    ä r(,   (    ä t-5EYq*    ä x'+Tv  ÿÀŽ Ä xww  ÿÀŽ äx*)Tv  ÿÀŽ ä xT  !    ä x)ÿ   +  ä r'       ä r%%*   &    â $~*ÿ   ÿ Ž â xˆˆˆˆÿÀŽ â $~'ÿ   ÿ Ž â r)       â $~ÿ   ÿ Ž â x&!ww  ÿÀŽ â $~,ÿ   ÿ Ž â x22ÿø� Ä t*ÏÝêö    ätÏÝêö&    ät*'5EYq&    ä x&Tv  ÿÀŽ ä t,5EYq    ä xÿ     Ä x! d      â&r		    ÿ   â x1u      ä r&       â &r!!    ÿ   â r   *    â &r"   	   âx	% d  (    â &r""   !   â x d      Ä x%%1u  (    â &r		#      âx1u      â &r"#   "   ä 0r
ÿ
    A   ä 0rÿ    A   ä &r	%$   	   ä&r$   "   Ä 0rÿ    A   â  r
     A  â x@  ÿà÷ ä Er	 	       â  r     A  â Er        â $r#ÿÿ   Ž ä #r	      ä #r      â G‰  ð÷ÿÿÿÿƒ ê Ay        € ê ‰      " $x   Ž Æ ‰      d z  p  p`ð È x    pDp  â !r       â !r       È/ ‰      ( ‰      b !r       â !r       È/ ‰ €    ( ‰ €    b !r       â !r       È/ ‰ @    ( ‰	 @    b !r       â !r		       È/ ‰      h  ‰	      ¢  M	        € ê y      &   â $x   ÿ Ž â !r       â/ $z   s  Ž â !r		       â_ $tÿ   ÿ Ž â x    p"ò Ä r ÿ   p"ð Ö $ÿÿ   	 Ž ä $z  v   Ž È� %v  j   Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      "   $ z  w  p`ð Ú M	        € ê y      &   " $tÿ   ÿ Ž â ¹z  F   
   â $z x  Ž È %v \   Ž Ê �y     ™ b ¹z  l      â Ey  Ð    € â ™x?    â y      !   b x
      ÿ  â $rÿÿ   ÿ Ž â x      ÿ  â $rÿÿ   ÿ Ž â x ÿ     È/ |     p`ð Ú G	  0    € ê $rÿÿ   ÿ Ž â y      %   " $z r  ÿŽ â x	   ÿ   â 'z m   Ž â Ey °    € æ $vÿ X  ÿ Ž â zÿ n    â $vÿ Y  ÿ Ž Ä $v#ÿ `  ÿ Ž ä $rÿÿ   
Ž ä %x$   ÿ Ž È $z o  Ž â x   ÿÀŽ â %x	$    Ž È $z u  ÿŽ â z Z  ÿàõ â %x   ÿ Ž Æ r	   à‘ â$r	ÿÿ   ÿ Ž ä %x$   ÿ Ž â rÿÿ   ÿ$~  ä z [  	ä â %x$    Ž È $rÿÿ    Ž â z
 Z  ÿàñ â $x   ÿ Ž ä $x$   ÿŽ â x

   ÿàó â $z! q   Ž ä $v%ÿ a  ÿ Ž â z [  ä  â $z!  t  !Ž âz ÿ `  pPð â $rÿÿ   ÿ Ž Ä %x!   Ž â z ÿ a   Sð Æ $rÿÿ   Ž  ä r   ÿàó Ê $x	   Ž  Ê �y/   • ¨  �y,   • â  x'ÿ     ä x   ÿàÿ â �y*   • $ $z'  t  'Ž Ä �y-   • d $z' q  'Ž È %x'   Ž Ê r   ÿàó Ê $x   Ž  ä $rÿÿ   
 Ž Æ �y(   • b $r	ÿÿ    Ž Æ �y)   • h �yðÿÿ ™ h �y    ™ h �y&   • h �y+   • b $rÿÿ    Ž â r
   ÿàó â $rÿÿ    Ž â �y"ôÿÿ ™ f %x!   Ž â�y$   ™ f %x'   Ž Ä �y    • d $x   Ž  ä �y    • h �y

    • b ‚x �˜­¿     ä ‚x &     â $~ÿ   ÿ Ž Ä $~1ÿ   ÿ Ž ä $~ÿ   ÿ Ž ä/ $~ÿ   ÿ Ž ä $~0ÿ   ÿ Ž ä $~2ÿ   ÿ Ž â x     ÿàÿ È |     p`ò ä x/ww  ÿÀŽ äOx//T  ,    È� x//ˆˆˆˆÿÀŽ È x//22ÿø� ä tÏÝêö    ät15EYq    ä x/Tv  ÿÀŽ ä x,,ww  ÿÀŽ ä r       Ä x*ww  ÿÀŽ äx**T  -    ätÏÝêö,    ät,5EYq,    ä x/ÿ   /  ä x*ˆˆˆˆÿÀŽ â $~ÿ   ÿ Ž â r,/   ,    Ä x22ÿø� ä t*ÏÝêö    ät/15EYq    ä x.Tv  ÿÀŽ ä x d  ,    äx1u  ,    Ä r*.   /    ä x/(ww  ÿÀŽ äx((T  )    â $~.ÿ   ÿ Ž Æ x,(ˆˆˆˆÿÀŽ â $~*ÿ   ÿ Ž â x1)ww  ÿÀŽ â &r   ÿ   â x,,22ÿø� â $~ÿ   ÿ Ž â t(.ÏÝêö/    â &r      â t.*5EYq1    ä x&ww  ÿÀŽ Ä x--ww  ÿÀŽ ä t/05EYq/    ä x*,Tv  ÿÀŽ ä x&&T  +    ä t)2ÏÝêö1    ä x,ÿ   ,  Ä tÏÝêö-    är((*   /    â $~*ÿ   ÿ Ž â x&&ˆˆˆˆÿÀŽ ä t-05EYq-    ä xÿ     ä r)),   .    â $~,ÿ   ÿ Ž â x++ww  ÿÀŽ â $~.ÿ   ÿ Ž â x&&22ÿø� Ä r   -    ä t*ÏÝêö    ät*,5EYq    ä t-.ÏÝêö+    ät.05EYq+    ä x,&Tv  ÿÀŽ Ä x&ÿ   &  ä x+( d  )    äx)(1u  )    ä r,   *    â &r++   ÿ   â r&-&   .    ä x d      Ä x1u      â &r))   +   â x d  &    â&r"      â x1u  &    Æ &r"   )   ä &r$      ä0rÿ    A   ä 0rÿ
    A   ä 0rÿ    A   ä &r$      â  r     A  â Er&        â  r     A  Ä Er        â G‰       € ê xÿ     Ê $z! q  Ž È $z!  t  !Ž È %x!   Ž Ê r
#   ÿàõ Ê $x   % Ê �y*
    ¨  �y+
    è  �y(
    ( �y)
    d r %       â $r
ÿÿ   # Ž È %x!   
Ž Ì �y     b $~-ÿ   ÿ Ž ä $~/ÿ   ÿ Ž ä $~.ÿ   ÿ Ž ä $~0ÿ   ÿ Ž â x,*ww  ÿÀŽ äOx**T  +    Ä� x++ww  ÿÀŽ ä x**ˆˆˆˆÿÀŽ ä t
-ÏÝêö,    äx**22ÿø� ä t/5EYq,    ä x*Tv  ÿÀŽ Ä t,.ÏÝêö+    ä r

       ä t+05EYq+    ä x*ÿ   *  ä x(ww  ÿÀŽ äx((T  )    Är,*   +    â $~*ÿ   ÿ Ž â x((ˆˆˆˆÿÀŽ â $~,ÿ   ÿ Ž â x+)ww  ÿÀŽ â 0rÿ    A   â x((22ÿø� ä t)*ÏÝêö    ät*,5EYq    Ä t,.ÏÝêö+    ät-05EYq+    ä x+(Tv  ÿÀŽ ä x(ÿ   (  ä x
 d      äx
1u      Ä r))+   *    â &r   ÿ   â r(,(   -    Æ &r      â x
) d  (    äx))1u  (    Æ &r

"      â  r     A  Æ &r
)$   
   Ê Er
 
       Ê #r
      â #r&      â #r      â G‰  `    € ê r#   ÿàõ Ê $x   % Ê �y    ¨  �y    è  �y    ( �y    d $rÿÿ   # Ž Ä $rÿÿ   % Ž È %x
'   Ž Ì �y

     b $~'ÿ   ÿ Ž ä $~)ÿ   ÿ Ž ä $~&ÿ   ÿ Ž ä $~(ÿ   ÿ Ž â xww  ÿÀŽ äOxT      Ä� xww  ÿÀŽ ä xˆˆˆˆÿÀŽ ä t'ÏÝêö    äx22ÿø� ä t)5EYq    ä xTv  ÿÀŽ Ä t&ÏÝêö    â $~&ÿ   ÿ Ž â r       ä t(5EYq    ä xÿ     ä xww  ÿÀŽ äxT      Är       â $~ÿ   ÿ Ž â xˆˆˆˆÿÀŽ ä xww  ÿÀŽ â 0r
ÿ
    A   â x22ÿø� ä tÏÝêö    ä t'ÏÝêö    â  r

     A  â t')5EYq    Ä t&5EYq    ä xTv  ÿÀŽ ä xÿ     ä x d      äx1u      ä r       â &r   ÿ   â r   '    Æ &r      â x d      äx1u      Æ &r"      È &r$      Ê Er        Ê #r
      â x@  ÿàõ ä x   ÿà÷ ä x
@  ÿàù â $rÿÿ    â x##   ÿàû â $rÿÿ   Ž ä $rÿÿ   	 â r%ÿ%   ÿäÿ â D       À â Gy  `ðÿÿÿÿƒ ê Ay       € ê $r
ÿÿ    Ž ä $rÿÿ    Ž Ä Ay        € ê z ÿ `  pPð È z ÿ a   Sð Ú G	  P    € ê ‰      ( ‰      b !r       â !r       È/ ‰      ( ‰      b !r       â !r       È/ ‰ €    ( ‰ €    b !r       â !r       È/ ‰ @    ( ‰	 @    b !r       â !r		       È/ ‰      ( ‰	      b !r       â !r	       â/ Gy  €    € ê ‰      ( ‰      h ‰      ¨ ‰      â !r       â !r       È/ ‰      " !r       ÆO ‰      b !r       Æ� ‰      ¨ ‰
      â !r	       â !r       â/ !r       ÈO ‰ €    " !r
       Æ� ‰
	 €    h ‰ €    ¨ ‰ €    â !r       â !r
	
       È/ ‰ @    " !r       âO !r       È� ‰ @    h ‰
 @    ¨ ‰ @    â !r       â !r       â/ !r
       ÈO ‰	      " !r       Æ� ‰      h ‰      ¨ ‰      â !r
	       â !r       â/ !r       âO !r       â� y      %   $ $x   Ž Ê z  p  p`ð È x    pDp  Ú M	        € ê z ÿ ^  pPð â $z  v  Ž Æz ÿ _   Sð Ú $ÿ   ÿ Ž È % ^   Ž Ì �	     ¢ $x    ÿ Ž Ê r  ÿ   p"ò äx     p"ô äx     p"ö äx     p"ø Î $ ÿÿ    Ž â z ÿ `  pPò â $" ÿÿ    Ž ä $2 ÿÿ   
 Ž â z ÿ a  Sò â $B ÿÿ    Ž È !         ÐO G™       € ê z ÿ b  pPð È z ÿ c   Sð Ú $ÿ   ÿ Ž È % b   Ž Ì �	     ¢ $x   ÿ Ž Ê x øÿÿÿp"ò äx üÿÿÿp"ô är ÿ   p"ö äx    p"ø Î $ÿÿ    Ž ä $"ÿÿ    Ž ä $vÿ h  ÿ Ž ä $2ÿÿ   
 Ž ä $Bÿÿ    Ž â x    pBò Æ !       ÔO G  P    € ê x    pRð Ú G‰  �     € ê x    pRð Ú G	  ð    € ê  x;ª¸¿  A  Ì s        $ !t  €?    Ì s        $  r     A  È  r       A  â Gy  P    € ê  x'7=  A  â x*BL?  A  Æ #t  €?   â  x   ?  A  Æ  r     A  Ì s     $   $ !t  €?    È  r     A  È  r       A  â Gy  °    € ê x    pRð Ú G‰  à     € ê x    pRð Ú G	        € ê 	z i    � ä 	z   i    � Æ  x;ª¸¿  A  â 	z   i €  � Ê s        $ !t  €?    Ì s        $  r     A  È  r       A  â Gy  Ð     € ê  r       A  â Gy  °     € ê 	x  à@  � ä 	x    à@  � Æ  x#ÛÙ?  A  â 	x    àÀ  � Æ  x;ª¸¿  A  ä !t   €?    È s        $ !t  €?    Ì s        $  r      A  È  r       A  ä y      "   ( y      &   b $z s  Ž ä $tÿ   ÿ Ž ä $z v  Ž È/ %v j   Ž Ê †y      â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y&      "   " ¹z  l      â Ey  �1    € â ™x?    â y'      !   b ¹z  F   
   â r) ÿ       â x*      ÿ  â r= ÿ       â x      ÿ  â x      ÿ  â x      ÿ  â x       ÿ  â x"      ÿ  â x$      ÿ  â r &ÿ   pRò Ä r &ÿ   pRð är/&'   ÿ(Ž È/ x ÿ   /  È r  /   ÿ� È x(ÿ      È | (   pbô Ú G)  @0    € ê y      '   " r ÿ       ä r ÿ       â y      %   b x  øÿÿÿÿÀŽ ä z7  q      â y      &   ¢ r//   €ÿàÿ ä x:(   ÿ   Ä r //   ÿ� ä r-7:   ÿŽ â$x37   :Ž âxÿ      â $x07   :Ž âr47:   ÿŽ â$x67   :Ž âz5  X      â $x77   :Ž â z1  Y      â 'z y   Ž â r) ÿ       Ä x9ÿ   /  â $z. }  ÿŽ â x   ÿ   ä/ z	ÿ z    â 'z s   Ž âO z8: q  ÿàÿ Æ $z p  (Ž â zÿ t    â $z	 |  (Ž â x   ÿàÿ â $z		 |  Ž ä $z v  Ž âx/   ÿ   â $z w  ÿŽ ä $z v  	Ž Ä $z p  Ž â r..   ÿàõ â %x8   ÿŽ â x þÿÿÿÿÀŽ ä r<ÿÿ   ÿä â %x8   ÿŽ â r/  €ÿàÿ ä xÿ     â %x
   Ž â r>   ÿàõ Ä r@   ÿà÷ â $x;   ÿŽ â rB   ÿä â%x/   Ž â r2	   ÿäÿ Æ %x   Ž È %x/   Ž È r5   ÿàõ ä r5>   ÿà÷ ä r1   ÿä ä r1B   ÿäÿ ä r5   ÿàõ â �y    ™ ¢  r5@   ÿà÷ Ä r1   ÿä â �yj    ‘ â r12   ÿäÿ Æ �y    ™ ( �yo    ‘ ( �yn!   ‘ ( �yl!   ‘ " r/.   :àµ Ä xs $       ä r9<   ÿd~ Æ %v Z  s Ž È/ $x$   ÿŽ â r        Èr   ÿàÿ Ê �yk   ™ h �yp   ™ h �y]   ™ h �yT   ™ h �yY   ™ h �yH   ™ h �yO    ™ h �y^    • b x ÿ   ÿÀŽ ÄO xj   ÿ   ä� xÿ   ÿÀŽ äx     âŽ ä x?o   ÿ   ä x        ä x?   âŽ Ä x`oð   ÿÀŽ â %v    Ž â xmjð   ÿÀŽ Æ %v   Ž â xÿ     â �y     ™ ¢ x`ÿ   `  ä xÿ     â �yq    ™ " xmÿ   m  ä xÿ   ÿÀŽ ä xÿ   ÿÀŽ Ä xm   ÿ   ä/ x`   ÿ   ä� x   âŽ ä x   âŽ Æ %v   Ž È %vV   Ž â �y\    ™ ê �yVV    ™ ¢  xÿ     ä xÿ     ä xn   ÿ   Ä xÿ   ÿÀŽ ä xinð   ÿÀŽ â �yW   ™ ¢ xl   ÿ   äxÿ   ÿÀŽ ä x_lð   ÿÀŽ ä x   âŽ Ä/ xiÿ   i  ä x   âŽ ä x_ÿ   _  â %v   Ž âxi   ÿ   ä x_   ÿ   â %v   Ž â x   ÿÀŽ â �yU    ™ â x   ÿÀŽ Ä r   ÿ@� â �y[    ™ â  r   ÿ@� Æ %v   Ž È %v   Ž â �yZ    ™ â
 r5   ÿàõ är5   ÿà÷ â/ �yX    ™ â  r1   ÿä ä r
5   ÿàõ ä r	1   ÿäÿ â�y?2   • â r1   ÿä Ä r/.   8àÇ â�yR4   • b r5   ÿàõ Æ �yK2   • â  r1   ÿä Æ �yD0   • â  r,9<   ÿ„þ Æ �yG6   • â %v Z  s Ž Æ �yF0   • â $x,$   ÿŽ Æ �yC4   • è  �yN0   • â  r   ÿàÿ Æ �yE0   • è  �y,6   • è  �yA   ™ h �yM   ™ h �yJ   ™ h �yP   ™ h �yS   ™ b r/.   -àµ Æ �yI   ™ h �yQ   ™ b r9<   ÿd~ Æ/ �yL    ™ b %v Z  s Ž Æ �yr    • b $x$   ÿŽ Ê r   ÿàÿ Ê �yb   ™ h �yd   ™ h �ya   ™ h �yg   ™ h �ye   ™ h �yf   ™ h �yh   ™ h �yc    ™ b r/.   3àµ È r9<   ÿd~ â %v Z  s Ž È $x$   ÿŽ â r/.   4àÙ är/.   0àµ ä r   ÿàÿ ä r{9<   ÿ¤~ äry9<   ÿd~ â �yt   ™ b r/.   6àÙ Ä rv/.   7àµ â �yu   ™ b %v Z  s Ž Æ �yx   ™ b %v Z  s Ž È %v Z  s Ž È %vv Z  s Ž ä �ys   ™ b rw9<   ÿ¤~ Æ �yv   ™ b rz9<   ÿd~ â $x}y$   ÿŽ ä $xw$   ÿŽ â �yy   ™ b $x�z$   ÿŽ Æ �yw   ™ h �yz    ™ b $x{{$   ÿŽ Ê r{   ÿàÿ ä �y{    • b r   ÿàÿ Æ �y„   ™ h �y€   ™ h �y   ™ b r}   ÿàÿ Æ �yƒ   ™ h �y}   ™ h �y‚   ™ h �y|   ™ h �y~    ™ h �y�    • b r�   ÿàÿ Æ �y�    • h �y‹   ™ h �yŠ   ™ h �y‡   ™ h �yŒ   ™ h �yˆ   ™ h �y†   ™ h �y‰   ™ h �y…    ™ h �yŽ   ™ h	 �y�    • h �y”   ™ h	 �y“   ™ b	 x�ÿ      ÄO x ÿÀŽ â �y’   ™ ¨ �y   ™ ¨ �y�   ™ ¨ �y   ™ ¨ �y‘    ™ ¨ �y•    • ¢ x��ÿÀŽ â &zk  € ÿ   Ä/ &ržk   ÿ   â �y–   ™ ¢  xjj   ÿÀŽ Æ �yœ   ™ ¢  xÿ   q  äxqqÿÀŽ â �y›   ™ " xÿÀŽ Æ &rkqk   ÿ   â �y™   ™ ¢  &zp  €    Æ �y   ™ ¢  &rž�p   ž   Æ �yš   ™ ¢  &rpp   k   Æ �y˜   ™ ¨  �y—    ™ ¨  �y    • ¢  xnn   ÿÀŽ ä Er  j       Ä x`   ÿÀŽ ä x\ÿÀŽ ä� x`VÿÀŽ ä Er         â &rž]   ž   âxoo   ÿÀŽ â &rj`]   p   â xmm   ÿÀŽ â &z]]  € ÿ   â xii   ÿÀŽ Ä xÿ   \  ä xk    =    ä xll   ÿÀŽ ä Er n       ä Er o       ä Ern m       Ä Erp i       â &ziT  € ]   â xÿÀŽ â #t    `¿k  â Er        ä Erl l       â #t  `¿k  âx__   ÿÀŽ â #t]n  `¿k  â xoÿ   V  â &ržT   ž   â Ern i       â #rV    ÿ   â #tl  `¿k  â Erl _       â #rm   ÿ   â xioÿÀŽ â #t_   `¿k  âEr ž       â #rV]n   V   â #t  `¿k  â#t\p  `¿k  â#tkl  `¿k  â &rjiT   j   â #rln_   m   â !rTV       â xVUÿÀŽ Ä xmÿ   U  ä xUÿ   [  ä x[[ÿÀŽ â &zpY  € ÿ   â xmmÿÀŽ â &rnVY   ÿ   âxUUÿÀŽ â &r [Y   ÿ   ä &zW  € p   Ä &rnmW   n   ä&r UW       â xWZÿÀŽ ä xZÿ   Z  ä xoÿ   X  ä xYXÿÀŽ â &zžH  € ÿ   â xZZÿÀŽ â &rnWH   n   âxXoÿÀŽ â &r YH       â xHÿ   ?  â &zžO  € ž   â Er        â &rnZO   n   âxRÿ   R  â&r XO       â xOHð   ÿÀŽ Ä xKÿ   K  ä rHÿ;   D  â #ro   ÿ   â Erž ž       ä xDG ð  ÿÀŽ ä xGR   ÿÀŽ ä xRKð   ÿÀŽ Ä rFFO   ÿ � ä xKH   ÿ   â #r?\ž   o   â xO        ä xHÿ   H  ä xCÿ   C  ä Ero j       Ä rNÿ;   N  ä xj        ä Ern n       ä rDDF   Gàÿ ä xHH   OìŽ ä rEER   ÿ � Ä xG, ð  ÿÀŽ ä xFC   ÿÀŽ ä xR        ä xCN   ÿ   ä xKK   jìŽ â !r??n       â Er, H       Ä rEGE   Fàÿ ä xGC   RìŽ ä xNÿ   N  ä ErC K       â &zKA  € ÿ   â  r??,     A  â xFN   OìŽ â &zOM  € ÿ   Ä &zKJ  € K   â #rTTC   ?   â Er? G       â &rGA   ÿ   ä &zOP  € O   â ErK K       â &rAqA   ÿ   ä &rG�J   G   ä&rAJ   A   â ErJ O       â #rNK   ÿ   â#rH K   ÿ   â &rGM   G   Ä &rM`M   A   â #rA_J   N   â&zNS  € ÿ   â #rH]J   H   â &rGP   G   ä &rJVS   ÿ   â#r   ÿ   â &zKI  € ÿ   ä &rR[S   ÿ   Ä &zNQ  € N   â ErO G       â &rJmQ   J   â Er          â &rPiP   M   â #ržk      â &rRUQ   R   â ErG N       â &zML  € K   â ErF F       â &rKWI   J   â!rHHO       â !r        â &rOYI   R   â ErJ M       â #rIG   ÿ   â#rNG   ÿ   â &rKZL   K   â ErP P       â 0r^ÿ^    A   â !rllo       â 0rEÿE    A   â  rF     A  â &rOXL   O   â #rI\J   I   â#rNkJ   N   â ErJ K       â !rPAP       â #rl?      â  rA^E     A  â ErO O       â 0rDÿD    A   â !rIIJ       â #r$A   $   â &rAb   ÿ   â !rONO       â  rJD^     A  â &zGb  € ÿ   â  rL,I     A  â &rA�d   A   â  rNFO     A  â #r%JT   %   â &zJa  € ÿ   â #rHCH   L   â &zGd  € G   Ä &zIg  € ÿ   â #r?P   N   â &rLa   A   ä &rAVg   ÿ   â ErG G       â &zJe  € J   ä &zPf  € ÿ   ä &zIh  € I   Ä &rAmh   A   â ErN J       â &zPc  € P   â ErI I       â #rJ G   ÿ   â#rRG   ÿ   â &rGWf   A   â ErP P       â #rJ]N   J   â#rR_N   R   â #rAI   ÿ   â#rNI   ÿ   â &rGZc   G   Ä &rg[g   ÿ   â #rA\P   A   â&rLe   L   â #rPkP   N   â ErN G       â &rbqb   ÿ   ä &rgUh   g   â ErI L       â 0rrÿr    A   â !rGAN       â &rbd   b   Ä &rgYf   g   â !rLJI       â  rJDr     A  â &rb`a   b   â  rN,G     A  â &rgXc   g   â #r#JH   #   â &rbie   b   â #rHCL   N   â &zLt  € ÿ   â Erg g       â  rAEr     A  â ErK b       â &zLu  € L   Ä !rgPg       â &rJt   ÿ   â #r"A   "   â &zAs  € ÿ   â !rRRK       â  rPFg     A  â &rJ�u   J   â ErL L       â &zGv  € A   â #r?R   P   â &zNx  € ÿ   Ä &rAs   J   â #rJ L   ÿ   â#rPL   ÿ   â &rLVx   ÿ   ä &zKw  € ÿ   ä &zIy  € N   â ErN G       â &rLmy   L   ä &zGz  € K   â ErI I       â &rKWw   L   â #rJ]N   J   â &rx[x   ÿ   â #rP_N   P   â ErL G       â #rNI   ÿ   â&rKZz   K   â #rGI   ÿ   â &rAv   A   â #rIkL   N   â &rxUy   x   â ErN K       â &rtqt   ÿ   â #rG\L   G   â ErA A       â &rwYw   x   Ä 0r{ÿ{    A   â !rGGN       â &rtu   t   â !rLJA       â &rwXz   w   â  rJD{     A  â &rt`s   t   â  rN,G     A  ä ErR w       â #r!JH   !   â &rtiv   t   â #rHCL   N   â &rJ„   ÿ   Ä &zL„  € ÿ   â !rIIR       â &zN€  € ÿ   â ErM t       â &rA�   J   ä &zL  € L   â  rRFI     A  â &zGƒ  € N   â  r{E{     A  â &rN€   A   â !rPPM       â &zI}  € ÿ   Ä &rAV}   ÿ   â ErL L       â #r {       â &zK|  € ÿ   â #r?P   R   â &zI‚  € I   ä &rAm‚   A   â ErP G       â #rJ L   ÿ   â#rLL   ÿ   â &zK~  € K   â ErI I       â &rAW|   A   â #rJ]P   J   â&rNƒ   N   â #rP_P   L   â &rGZ~   A   â ErL K       â #rAI   ÿ   â ErK N       Ä ErN G       â #rA\L   A   â &r}[}   ÿ   â #rRI   ÿ   â &r„q„   ÿ   â !rAAN       â &r}U‚   }   â #rRkL   R   â &r   „   â !rJJK       â  rL,A     A  â 0r�ÿ�    A   Ä &r}Y|   }   â #rACJ   L   â &r€`€      â  rLD�     A  â &r}X~   }   ä &r€iƒ   €   â #rLA      â 0r�ÿ�    A   â Er} }       â &zA‹  € ÿ   â ErG €       Ä  rJD�     A  â&zAŠ  € A   â !r}R}       â  r�E�     A  â !rPPG       â &zG‡  € ÿ   â  rNF}     A  â #r�      â ErA A       â &r‹   ÿ   â #rJH      â #rN?P   N   â &zGŒ  € G   â #rH A   ÿ   â&r�Š      â #rPA   ÿ   â &zIˆ  € ÿ   Ä &rAVˆ   ÿ   ä&rˆ[ˆ   ÿ   â ErJ G       â &r‡      ä &zL‰  € ÿ   ä &zI†  € I   ä &rˆU†   ˆ   â #rH]J   H   â &rŒ      â #rP_J   P   â &rJm†   A   â ErI I       â &zL…  € L   Ä &rˆY‰   ˆ   â Er        â &r‹q‹   ÿ   â ErL L       â &rJW‰   J   ä &rˆX…   ˆ   â #rRI   ÿ   â &r‹Š   ‹   â !rHH       â &rJZ…   J   â Er ˆ       â &r‹`‡   ‹   â #rAI   ÿ   â #rRkL   R   â ErJ J       â &r‹iŒ   ‹   â #rA\L   A   â !rR       Æ !rAAJ       â  rLF     A  â Er‹ ‹       â &zŽ  € ÿ   â  rJ,A     A  â 0r�ÿ�    A   ä &zG“  € ÿ   â !rPP‹       â &zA”  €    â #rJCH   J   â &rŽ   ÿ   â  rHD�     A  â&zG’  € G   âO #rL?P   L   â  r�E�     A  â &rŽqŽ   ÿ   â ErA A       â &r�”      â #rHJ      â #r�L      â &rŽ”   Ž   â ErJ G       â #rH A   ÿ   â#rLA   ÿ   â &r“      Ä &zG  € ÿ   ä &rAV   ÿ   â #rH]J   H   â&rŽ`“   Ž   â #rL_J   L   â &rI[   ÿ   ä &r’      ä &zJ  € ÿ   Ä &zG�  € G   ä &rm�   A   ä &rŽi’   Ž   ä &r�U�   I   â Er        â &zJ‘  € J   â ErG G       â &rW      Ä &r�Y   �   â Er Ž       â &rZ‘      â !rHH       â ErJ J       â #rG   ÿ   â !rLL       â Er        â &z–  € ÿ   â #r\J      Æ &zœ  €    â !r       â  r�E�     A  Æ  r,     A  â Er        â &rVV   ÿ   ä #rCH      ä #rH   ÿ   â &z  € ÿ   â #r�N      â &r[[   ÿ   â #rNG   ÿ   â &r–   ÿ   Ä &z˜  € ÿ   ä &zš  €    ä &rVmš   V   â #rNkJ   N   â &rqq–   ÿ   ä &r[Uš   [   ä &zJ›  € ÿ   Ä&r�œ      â Er        â &z—  €    ä &rVW˜   V   ä &rqœ   q   ä &r[Y˜   [   ä &zJ™  € J   Ä &r›      â Er        â &r�X‘   �   ä &rVZ—   V   â #r   ÿ   â &r``›   q   ä &r[X—   [   â #r   ÿ   â &r™      â ErJ J       â #r     ÿ   â x((   ÿàÿ â &r`i™   `   â ErA �       â #r\      â ErV V       â #rk      â Er[ [       â #r ]J       â Er        Ä | (   pbô â !rANA       â Er `       â !rV       â #rH_J   H   â !r[[       â !r        â 0rÿ    A   â  r FA     A  â  r,,     A  â 0r•ÿ•    A   â !rHH       â  r[F[     A  â #rL?L       â #r,C   ,   â  rD     A  â#rH?H   [   â  r D•     A  â rE•     A  â rE     A  â x55À  ÿà÷ â #r=    =   â #r+L   +   â #r*,   *   â #r)H   )   â x--@   ÿàÿ Ä x33@   ÿàÿ ä x44@   ÿàÿ ä x00@   ÿàÿ ä x66@   ÿàÿ ä x77@   ÿàÿ ä x88@   ÿàÿ Ä x::@   ÿàÿ ä r1ÿ1   ÿäÿ â D)       À â Gy  Óÿÿÿÿƒ ê Ay        € ê Ey  0    € â G™      € ê r&&'   ÿHŽ Ê ˆs &% øÿ H   è ˆs &$€øÿ H   è ˆs &# ùÿ H   è ˆs &"€ùÿ H   è ˆs &! úÿ H   è ˆs & €úÿ H   è ˆs & ûÿ H   è ˆs &€ûÿ H   è ˆs & üÿ H   è ˆs &€üÿ H   è ˆs & ýÿ H   è ˆs &€ýÿ H   è ˆs &= þÿ H   è ˆs &+€þÿ H   è ˆs &* ÿÿ H   è ˆs &)€ÿÿ H   ä Ay        € ê {          ì M	        € ê „y '     H   b z9  r      Æ „y' €   H   ¢ r599   ÿŽ är799   ÿŽ â „y' €  H   è „y	' €  H   ( „y'    H   h „y
'    H   ( „y'    H   ( „y' €  H   è y      %   ( „y'    H   â !r  %       Æ/ y      '   b !r$       ÆO ‰       ¢ !r""       ÆŸ ‰      " !r 	        Æ„y$'    H   â !r#       Æ‰       â !r

!       Æ „y'    H   â x!   ÿ   â !r       ä „y' €  H   â z ! r  p`ð â !r        äO „y'    H   â r 'ÿ   p"p â !r       Ä ‰      " !r%*       Æ y       &   ¢ !r         Æ� ‰      è „y' €  H   è „y&' €  H   â !r=$=       Æ ‰	      ( ‰      b !r       â !r       â $z   x  !Ž âO ‰"      " x!!   ÿàÿ â !r       â� $z  ~   Ž â/ ‰ €    b !r       â !r       Ä ‰
      ¢ 
  r  ÿàó Æ ‰ €    â ÿÿ   ÿäÿ  â !r	       â
 j  ÿ„ â ‰$      " z ! r  p`ò â !r       â
 k   â „y4' €  H   â r9    ÿ„ â !r!       â !r#       â ‰       b rÿÿ   ÿä â !r""       â !r+&+       â ‰      " z j  ÿ„ â !r       â/ x '   p"ò â ‰      b !r
       âO z k   Ä ‰&      ¢ !r       â�  5   ÿàû ä‰(!      â !r$$       â	 7   ÿà÷ ä ‰*      " ÿÿ   ÿäÿ ä 

 j  ÿŠ â ‰,#      " ÿÿ   ÿäÿ â !r         â
	 j  ÿ† â ‰.=      b r99  €ÿŽ Ä 
 k  � â ‰0+      b r9    ÿˆ â!r       â 
		 k  � â ‰"      "     ÿà÷ â!r       â/ 5   ÿàÿ â ‰2%      b rÿÿ   ÿä â !r&&       âO z j  ÿˆ â ‰ @    ¢ ÿÿ   ÿäÿ â !r(!(       â� 5    ÿà÷ â ‰      â !r**       âz k   â !r,#,       â ‰ @    " !r.=.       â
 j  ÿˆ â !r44)       â !r0+0       â 8ÿÿ   ÿäÿ â ‰+&      b 5 j  ÿ† â !r""       â 
 k   â ‰-(      " 5 k  8� â !r2%2       â/ $x9   ÿŽ â‰%$      b )        â !r       âO 9   ÿàÿ â ‰/*      ¢ r    ÿàõ â !r       â� 9    ÿàû â ‰ €    â r6ÿÿ   ÿä â !r       âz j  ÿ„ â ‰1,      " :ÿÿ   ÿäÿ Ä 9 j  ÿŠ â ‰3.      " !r+&+       âz k  6 ä ‰'0      b !r-(-       â 9 k  :� ä ‰" €    " !r$$%       Æ/ ‰!  €    b !r/*/       ÆO ‰52      ¢ !r       Æ� ‰4      â !r1,1       Æ‰      " !r3.3       Æ ‰ €    " !r00'       Æ‰# €    b !r""       â !r  !       ä/ ‰'$ €    " !r225       ÆO ‰&+ €    b !r44       Æ� ‰(- €    ¢ %  j  ) Ž â !r;       ä‰ @    â 7   ÿàÿ â !r       ä ‰*/ €    "     ÿàõ â !r#       Ä‰,1 €    b 6ÿÿ   ÿä ä  j  ÿ„ â‰" @    b !r'$'       ä  k  6 â ‰#  @    " !r&+&       Æ/ ‰.3 €    b !r(-(       ÆO ‰50 €    ¢ !r       Æ� ‰      â !r*/*       Æ‰72 €    " !r,1,       Æ‰94      b !r"       Æ ‰! @    b !r# #       Æ ‰% @    " !r.3.       â/ !r505       äO ‰ ' @    b !r=       Æ� ‰)( @    ¢ !r727       Æ‰& @    â !r949       Æ‰      " !r!!       Æ ‰+* @    b !r%%       Æ ‰-, @    ( ‰      " !r '        Æ/ ‰/. @    b !r)()       ÆO ‰"5 @    ¢ !r&&       Æ� ‰09 €    â !r1       Æ‰$7 @    " !r+*+       Æ‰!      b !r-,-       Æ †	 ;    â !r       Æ ‰#      b !r/./       Æ/ ‰%      b !r"5"       âO         ä ‰       " !r090       â� !r$7$       â%  j   Ž â ‰&      ¢ !r!!       Æ† =   è ‰)      â !r##       Æ †	 1    â !r%%       Æ/ ‰+      b !r        Æ ‰-      ( †     â !r&       ÆO ‰/      ¨ ‰"      " !r))       Æ� ‰0 @    è ‰$      b !r++       Æ/ †	 !    â !r--       Æ † #   è †	 
%    â !r//       âO !r"       ä†     â !r0       Æ� †	     â !r$       Æ† )   è †	 +    è † -    è †	 /    è †    è ‰      ( †	     â M™        € ê x   ÿàÿ â !r       Æ? r    ÿàñ È r ÿÿ   ÿä  ä z j  ÿ€ È z k     Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y"      "   " ¹z  l      â Ey  0-    € â ™x?    â y#      !   b ¹z  F   
   â r% ÿ       â x&      ÿ  â y)      %   ¢ r9 ÿ       â x      ÿ  â x      ÿ  â x      ÿ  â x      ÿ  â x       ÿ  â r "ÿ   pRð Ä r "ÿ   pRò är+"#   ÿ(Ž È/ x ÿ   +  ä x))   ÿ   äO r  +   ÿ� È x$ÿ      È | $   pbô Ú G)  Ð+    € ê y      '   " r ÿ       ä r ÿ       â y      &   b z3  q      ä x6$   ÿ   ä z1  X      ä r*36   ÿŽ â$x03   6Ž âr-36   ÿŽ â$x23   6Ž âz/  Y      â $x33   6Ž â r% ÿ       Ä z46 q  ÿàÿ â 'z y   Ž È 'z s   Ž â/ zÿ z    ä x øÿÿÿÿÀŽ â $z ) p  $Ž â zÿ t    â $z, }  ÿŽ â r++  €ÿàÿ â $z |  $Ž ä$z |   Ž â x )   ÿàÿ â $z v  Ž â x5ÿ   +  â $z w  ÿŽ Ä $z  p  Ž â r ++   ÿ� â $z v  Ž â r,,   ÿàõ â %x8   ÿŽ â x þÿÿÿÿÀŽ ä x+   ÿ   â %x8   ÿŽ â xÿ      ä r+  €ÿàÿ â%x+   Ž â r8ÿÿ   ÿä Ä xÿ     â %x
   Ž âr:   ÿàõ är<   ÿà÷ â $x7   ÿŽ â r>   ÿä â%x+   Ž â r.	   ÿäÿ Æ %x   Ž È r1   ÿàõ ä r1:   ÿà÷ ä r/   ÿä ä r/>   ÿäÿ ä r1   ÿàõ â �y    ™ ¢  r1<   ÿà÷ Ä r/   ÿä â �y|    ‘ â r/.   ÿäÿ Æ �y(    ™ ( �yt    ‘ ( �ys!   ‘ ( �y !   ‘ " r@+,   6àµ Ä xg $       ä r58   ÿd~ ä x;        â %v@ Z  g Ž È/ $x$   ÿŽ Ê r   ÿàÿ är        Ê �y‘   ™ h �y�   ™ h �yY   ™ h �yQ   ™ h �yT   ™ h �yU   ™ h �yI    ™ h �yZ    • b xÿ   ÿÀŽ ÄO xB|   ÿ   ä� x(ÿ   ÿÀŽ äxB   âŽ ä x=t   ÿ   ä xr|ð   ÿÀŽ â %v  ; Ž â x=   âŽ Ä xÿ     ä �y�    ™ ¢ %v  ; Ž â xmtð   ÿÀŽ ä xrÿ   r  ä �yŽ    ™ " xÿ   (  ä xmÿ   m  Ä xÿ   ÿÀŽ ä xr   ÿ   ä/ xÿ   ÿÀŽ ä xm   ÿ   ä� x   âŽ ä x   âŽ Æ %v  ; Ž â xÿ     Æ %v  ; Ž â �yX    ™ â xÿ   (  Æ �yO    ™ ¢  xs   ÿ   äxÿ   ÿÀŽ ä xnsð   ÿÀŽ ä x@    ÿ   äxÿ   ÿÀŽ ä x[ ð   ÿÀŽ Ä x   âŽ ä/ xnÿ   n  ä x@   âŽ ä x[ÿ   [  â %v  ; Ž âxn   ÿ   â �y@   ™ â  x[   ÿ   â %v  ; Ž â x   ÿÀŽ â �yR    ™ â x   ÿÀŽ Æ �yS    ™ â  r   ÿ@� ä r(   ÿ@� Æ %v  ; Ž È %v  ; Ž â �yV    ™ â
 r1   ÿàõ Æ �yW    ™ â  r/   ÿä ä rF
1   ÿàõ är1   ÿà÷ â/ �y;2   • â rG/   ÿä är	/   ÿäÿ â �y?0   • â r1   ÿàõ Ä r+,   4àÇ â�yPF0   • b r/   ÿä Æ �yB2   • b r(58   ÿ„þ Æ �yD4   • b %v Z  g Ž Æ �yM6   • b $xA($   ÿŽ Æ �yN0   • h �y=4   • h �yE0   • b rA   ÿàÿ Æ �y(6   • h �yF   ™ h �yA   ™ h �yL   ™ h �yK   ™ h �yC   ™ h �yG   ™ h �yH   ™ h �yJ    ™ b r+,   *àµ Æ/ �yd    • b r58   ÿd~ â %v Z  g Ž È $x$   ÿŽ Ê r   ÿàÿ Ê �y]   ™ h	 �y\   ™ h	 �y_   ™ h	 �y`   ™ h	 �yb   ™ h	 �yc   ™ h	 �ya   ™ h	 �y^    ™ b	 r+,   0àµ È r58   ÿd~ â %v Z  g Ž È $x$   ÿŽ Ê r   ÿàÿ Ê �yl   ™ h �yi   ™ h �yh   ™ b r+,   -àÙ Æ �yf   ™ h �yk   ™ b rq58   ÿ¤~ Æ �yj   ™ b r+,   2àµ är+,   3àÙ â �ye   ™ b %v Z  g Ž Æ �y{    • b %v Z  g Ž È %v Z  g Ž ä �yg    ™ b ro58   ÿ¤~ Ê $xoo$   ÿŽ ä $xqq$   ÿŽ Æ ro   ÿàÿ ä �yo    • b	 rq   ÿàÿ ä rp58   ÿd~ â �y‡   ™ h �yy   ™ b $xup$   ÿŽ Æ �yx   ™ d ru   ÿàÿ ä �yu   ™ h �yv   ™ h �yz   ™ h �yq   ™ h �yw   ™ h �yp    ™ h �y~    • h �y}   ™ h �y…   ™ h �y„   ™ h �yƒ   ™ h �y   ™ h �y�   ™ h �y€   ™ h �y‚    ™ b x�ÿÀŽ ÄO xÿ   Ž  â�y†    • ¢  x�ÿ   �  â &z‘  € ÿ   ä/ �y�   ™ " x��ÿÀŽ ä xtt   ÿÀŽ â �yŒ   ™ ¢ xŽÿÀŽ Ä xŽÿÀŽ â &r‘   ÿ   â�y   ™ ¢ &r‘‘   ÿ   â x    ÿÀŽ â &z�  €    â Ert t       â &r’��      â�yŠ   ™ ¢ &r”Ž�   ‘   â x�    =    Ä �y‹   ™ ¢ xss   ÿÀŽ Æ �y‰   ™ ¢ Er        Æ �yˆ    ™ ¨ �y    • ¢ x||   ÿÀŽ â #tt  `¿�  â Er s       ä xsm   ÿÀŽ ä xn   ÿÀŽ â/ &znY  € ÿ   â xmXÿÀŽ Ä� Ert        ä xÿ   X  ä xrr   ÿÀŽ ä Er  |       â &r’mY   ’   â xOÿÀŽ â &znQ  € n   â xÿÀŽ Ä Err r       â #t    `¿�  â Er        â &r|Y   ”   â #tXt  `¿�  â &r’Q   ’   â #tYr  `¿�  â xOÿ   O  â #rr    ÿ   â Ert n       ä Er” s       Ä xnOÿÀŽ â #rrYt   r   â Ers ’       ä x[[   ÿÀŽ â &r|nQ   |   â xQRÿÀŽ â !rsrs       â xrÿ   R  â #r   ÿ   â Er– [       â #t[”  `¿�  â xRÿ   S  Ä xSSÿÀŽ â #rtt[      â xRRÿÀŽ â &z’T  € ÿ   â xrrÿÀŽ â &rQT   ÿ   ä&r‘ST   ÿ   ä &zT@  € ÿ   â #t  `¿�  â#t  `¿�  â#tO–  `¿�  â &r“RU   ‘   Ä &z’U  € ’   ä &r�rU      â xUÿ   V  â &z‘I  € T   â xTVÿÀŽ ä x”ÿ   W  ä xVWÿÀŽ ä Er ’       Ä xWUÿÀŽ â &r’T@   �   âxU”ÿÀŽ â &r”V@   “   â Er@ ‘       â #r�   ÿ   â &r’WI   ’   â#r   ÿ   â &r”UI   ”   â xIÿ   ;  â #r;X@   �   â #r@O      â Er� |       Ä x@Ið   ÿÀŽ ä rPÿ7   P  äxBÿ   B  â !rtt�       â r??@   ÿ � ä xDÿ   D  ä x@P   ÿ   Ä x�        ä xPÿ   P  ä xBBð   ÿÀŽ ä xMM ð  ÿÀŽ ä xDD   ÿÀŽ ä rNÿ7   N  Ä Er’ ’       ä xPP   �ìŽ ä x=ÿ   =  ä rBEB   ÿ � ä r?M?   Dàÿ ä x|        Ä xEN   ÿ   ä xD( ð  ÿÀŽ â !r;;’       â x==   ÿÀŽ ä Er( P       ä xI        ä xNÿ   N  Ä xEE   |ìŽ ä r=DB   =àÿ â  rD;(     A  â xBN   IìŽ â &zNF  € ÿ   â Er; E       â &rEF   ÿ   ä&rFF   ÿ   ä &zNA  € N   Ä &rE�A   E   ä&rAŽA   F   ä &zPL  € ÿ   ä &rAL   A   â x@@   IìŽ â &zPK  € P   â ErN N       â &rEmL   E   Ä &zIC  € ÿ   ä &rLnK   A   âErP P       â &rEK   E   ä &rAQC   ÿ   â #rF N   ÿ   â &zKG  € ÿ   ä &zIH  € I   Ä &rCSC   ÿ   â ErE E       â &rArH   A   â #rFYP   F   â &zKJ  € K   â ErI I       â &rHRH   C   ä &rCTG   A   â!rEFE       â &rGVG   H   â ErF K       â #rAI   ÿ   â#rHI   ÿ   â &rCWJ   C   Ä #rAXF   A   â#rHOF   H   ä ErF C       â &rGUJ   G   â Er” ”       â 0rZÿZ    A   â Er@ @       â 0r?ÿ?    A   â !rCAF       â ErB B       â !r”       â #rDs@   D   â ErG G       â  rF?Z     A  â #rNN   ÿ   â  rC(C     A  â 0r=ÿ=    A   â  rB     A  â ErI L       â #r!FD   !   â #rN[P   N   â !rGHG       â &zF]  € ÿ   â #rD@E   C   â &zC\  € ÿ   â #rt;      â  rAZ=     A  â !rNNI       â  rHBG     A  â &zF_  € F   â #r A       â &zC`  € C   â #r;N   H   â &zHb  € ÿ   â ErF F       Ä ErJ C       â &rCQb   ÿ   ä &zEc  € H   â #rH F   ÿ   â&rA]   ÿ   â #rFF   ÿ   â &zGa  € ÿ   ä &rbSb   ÿ   ä &rCrc   C   â #rI[J   F   â &rA�_   A   â ErE E       â &r]]   ÿ   Ä &rbRc   b   ä &zG^  € G   ä &rFTa   C   â #rHYJ   H   â &rAm\   A   â ErJ G       â &r]Ž_   ]   ä &rbVa   b   Ä &rFW^   F   â #rCE   ÿ   â&rA`   A   ä &r]\   ]   â ErF F       â &rbU^   b   â #rLE   ÿ   â #rCXJ   C   â &r]n`   ]   â ErA A       â 0rdÿd    A   â ErG b       â #rLOJ   L   â !rCCF       â ErJ ]       â !rEHA       â  rF?d     A  â!rGLG       â  rA=d     A  â  rC(C     A  â #rFD      â !rJIJ       â  rHBG     A  â #rA      â #rD@E   C   â &zCl  € ÿ   Ä &rAl   ÿ   ä&rEl   ÿ   â #r;J   H   â &zHi  € ÿ   ä &zCh  € C   ä &rF�h   A   ä&rhŽh   E   Ä &zEf  € ÿ   â ErC C       â &zHk  € H   ä &zGj  € E   ä &rEQf   ÿ   â ErH H       â &rAmi   F   â #rF C   ÿ   â#rLC   ÿ   â &zJe  € ÿ   Ä &rErj   E   ä &rCSf   ÿ   â #rFYH   F   â#rL[H   L   â &zJg  € J   â ErG G       â &rHTe   E   ä &rCRj   C   ä &rAk   A   â ErJ J       â &rHWg   H   Ä &rNVe   C   â #rCG   ÿ   â ErA A       ä ErH H       â #rCXJ   C   ä !rAFA       â #rFG   ÿ   â !rCCH       Æ #rJOJ   F   â 0rFÿo    A   â  rC(C     A  â &rNUg   N   Æ #rC@A   C   â  rA=F     A  â ErG N       Æ #rA      â &zAy  € ÿ   â !rGJG       â 0rJÿ{    A   â  rH?F     A  â &rhi   h   ä &zAx  € A   ä &ry   ÿ   â  rF?J     A  â &rhnk   h   â #rHD      â &rD�x      â ErA A       â #rFC      â ErE h       â &zCu  € ÿ   Ä &rmu   D   â #rD A   ÿ   â#rHA   ÿ   â &zFv  € ÿ   ä &rAQv   ÿ   â !rLLE       â &zCz  € C   â  rNBG     A  â &zGw  € ÿ   ä &zEq  € F   Ä &rArq   A   â ErF C       â &ryy   ÿ   â ErE E       â &zGp  € G   ä &rATw   A   ä &rvSv   ÿ   â #rDYF   D   â &rxŽx   y   â #rI[F   H   â &rCWp   A   â ErF G       â &rz      â #rAE   ÿ   â &rvRq   v   Ä &ruu   x   â ErH C       â &rvVw   v   â Er        â #rAXF   A   â &runz   u   â #rL;L   N   â &rvUp   v   â !rAAH       â !rCD       â #rNE   ÿ   â ErD u       â 0r~ÿ~    A   â  rA(A     A  â ErE v       â #rNOF   N   â !rFID       â #rA@C   A   â  rD?~     A  â !rENE       â  r=J     A  â &zJ}  € ÿ   â #rDA      â &rA}   ÿ   â  rHBE     A  â #rL      â &zL„  € ÿ   â  r=~     A  â &zJ…  € J   â #rF;F   H   â &rA�…   A   Ä &zLƒ  € L   â ErJ J       â &rAm„   A   â #rF      â &zF  € ÿ   â ErL L       â &rAƒ   A   â #rD J   ÿ   â &zC€  € ÿ   ä &zF�  € F   Ä &rQ   ÿ   â ErG A       â #rDYL   D   â &zE‚  € C   â ErA F       â &rr�      â !rGDG       ä ErD E       â &rCT€      â #rA   ÿ   â#rHA   ÿ   Æ #rXD      â#rHOD   H   â &zD‡  € ÿ   ä &r‡   ÿ   ä &zD�  € D   ä&r��      ä &rS   ÿ   â ErD D       â &rmmŒ      ÄO &z  € ÿ   ä &rQQ   ÿ   ä&rS   ÿ   â #r  D   ÿ   â &r}}   ÿ   ä &rR�      â #rJJ   ÿ   â#rDD   ÿ   â &zAŒ  € ÿ   Ä &r‡   ÿ   ä &z‰  € ÿ   ä &z‹  €    ä &rQr‹   Q   ä&rR‹      ä &r}Ž…   }   Ä &rV€      ä &rmŠ   m   ä &zAŠ  € A   ä &r�Ž�      ä &zˆ  €    â Er        â &rQT‰   Q   Ä &rV‰      ä &r}„   }   ä &rU‚      âEr A       â &rCW‚   C   â Er        â &r�Œ   �   ä &rQWˆ   Q   Ä &rUˆ      â Er        â &r}nƒ   }   â #r   ÿ   â#r   ÿ   â x$$   ÿàÿ â &r�nŠ   �   â ErF C       â #r Y       â Erm m       â #rX      âEr Q       â #rO      â Er        â #rJ[L   J   â Er} }       â !rH       â | $   pbô â !rF       â Er� �       â #rD[   D   â !rm m       â !r       â !r       â !rJJ}       â  r B     A  â 0r†ÿ†    A   Ä 0rÿ    A   â  r(     A  â !rDD�       â  r(     A  â  rB     A  â #rJ;J       â  r ?†     A  â r??     A  â #r@G      â#r@m      â #r;D      â  r†=†     A  â r=     A  â x11À  ÿà÷ â #r9    9   â #r'†J   '   â #r&?   &   â #r%   %   â x**@   ÿàÿ Ä x00@   ÿàÿ ä x--@   ÿàÿ ä x22@   ÿàÿ ä x33@   ÿàÿ ä x44@   ÿàÿ ä x66@   ÿàÿ Ä r/ÿ/   ÿäÿ â D)       À â Gy  P×ÿÿÿÿƒ ê Ay        € ê Ey      € â G‰  ð     € ê $x""À  #Ž Ê ˆs "! ùÿ H   è ˆs " €ùÿ H   è ˆs " úÿ H   è ˆs "€úÿ H   è ˆs " ûÿ H   è ˆs "€ûÿ H   è ˆs " üÿ H   è ˆs "€üÿ H   è ˆs " ýÿ H   è ˆs "€ýÿ H   è ˆs "9 þÿ H   è ˆs "'€þÿ H   è ˆs "& ÿÿ H   è ˆs "%€ÿÿ H   ä Ay        € ê {          ì M        € ê „y #     H   b z ) r  p`ð Ä z3  r      â „y	# €  H   ¢ r #ÿ   p"p ä r133   ÿŽ â„y#    H   â r/33   ÿŽ Æ „y# €   H   ( „y#    H   h „y#    H   ( „y# €  H   ( „y# €  H   ( „y#    H   è „y# €  H   â !r  !       Æ/ „y#    H   â !r		       ÆO ‰       b !r       ÆŸ „y# €  H   â !r        Æ„y#    H   â !r       Æ„y# €  H   â !r       Æ ‰      " !r       Æ y      '   ¢ !r       â !r        ä/ ‰      b !r       Æ y       &   è ‰      " !r99       Æ ‰      b !r       Æ ‰      " !r       â !r''       ä ‰
      ( ‰	      " !r       â/ !r%%       â $z   x  )Ž â� ‰      b x))   ÿàÿ Ä $z  ~   Ž âO !r       â‰       ¢ z ) r  p`ò â !r       â!r&       â ‰"      â 
  r  ÿàõ â !r       â x #   p"ò â ‰ €    " ÿÿ   ÿä â !r

       â 
 j  ÿ„ â ‰$9      " !r	       Ä 
 k   â ‰      b !r       Æ/ ‰&'      h ‰(      b !r         ÆO ‰*%      ¢ !r""       Æ� ‰	      â !r       â  1   ÿà÷ ä ‰      " !r$9$       Æ‰      " !r       âr3    ÿ„ ä ‰
      b !r&'&       Æ/ ‰      b !r((       â !r*%*       äO ‰!      ¢ !r	       Æ� ‰#       â 	 /   ÿàû â !r       ä ‰%"      " ÿÿ   ÿäÿ â !r       ä‰ @    " rÿÿ   ÿä â !r
       Ä‰'$      b r3    ÿˆ â !r       ä/ ‰ €    b 
 j  ÿ† ä z j  ÿ„ â ‰)&      b !r!!       âO r.ÿÿ   ÿä Ä ‰+(      ¢ 
 k  � â !r# #       â� z
 j  ÿˆ â ‰ €    â !r%"%       â z k   ä ‰ €    " !r       â,ÿÿ   ÿäÿ Ä ‰-*      " !r'$'       â3   ÿàÿ ä ‰ €    b !r       â/ 
	 j  ÿŠ ä ‰ €    b !r)&)       â z k  . Ä ‰ €    b !r+(+       âO /   ÿàÿ ä ‰"! €    ¢ !r       â� .    ÿàù â $x3   ÿŽ â ‰$# €    â !r       â 
		 k  ,� Ä ‰&% €    " !r-*-       â/    ÿà÷ ä ‰(' €    " !r       âÿÿ   ÿä ä ‰ @    b . j  ÿˆ â !r       â/ r    ÿàõ â ‰*) €    b . k   â !r       â rÿÿ   ÿä â ‰,+ €    b z j  ÿ„ â !r"!"       âO 0ÿÿ   ÿäÿ â ‰ @    ¢ !r$#$       â� z k   Ä ‰ @    â !r&%&       â / j  ÿ† ä ‰.- €    " !r('(       â/ k  0� ä ‰ @    " !r       Æ‰      b !r*)*       Æ/ ‰ @    b !r,+,       Æ ‰  @    b !r       ÆO ‰!" @    ¢ !r       Æ� ‰#$ @    â !r.-.       Æ ‰%& @    " !r       Æ‰'( @    " !r/       â1   ÿàÿ ä ‰      b !r       â/     ÿàõ ä ‰)* @    b x        â !r         â 0ÿÿ   ÿä â ‰+, @    b !r!"!       âO %v  j   Ž â  j  ÿ„ â ‰      ¢ !r#$#       Ä�  k  0 â ‰-. @    â !r%&%       Æ ‰      " !r'('       Æ‰      " !r1       Æ‰      b !r)*)       Æ/ ‰       b !r+,+       Æ ‰ !      b !r       ÆO ‰#      ¢ !r-.-       Æ� ‰"%      â !r       Æ ‰'      " !r       Æ†	 /    â !r       Æ† 1   â !r        Æ/ ‰$)      b !r!!        Æ ‰+      " !r##       ÆO ‰-      ¢ !r%%"       Æ� †	     â !r''       Æ †     è †	 
    è † 
   â !r))$       Æ/ †	     â !r++       Æ† !    â !r--       ÆO †	 #    è † %   è †	 '    è † )    è †	 +    â M™        € ê †y -   â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y#      "   " ¹z  l      â Ey  à(    € â ™x?    â y$      !   b ¹z  F   
   â r* ÿ       â x      ÿ  â y)      %   ¢ r' ÿ       â x      ÿ  â r9 ÿ       â y      &   â x      ÿ  â x       ÿ  â r" ÿ       â y	      '   " r #ÿ   pRð Ä r #ÿ   pRò är+#$   ÿ(Ž È/ x ÿ   +  ä x))   ÿ   äO r  +   ÿ� È x&ÿ      È | &   pbô Ú G)  `'    € ê r ÿ       â�$z,	 }  ÿŽ â r ÿ       â $z w  ÿŽ â x6&   ÿ   â 'z	 y   Ž â z1  X      ä r,,   ÿàõ â 'z s   Ž â zÿ z    Ä x øÿÿÿÿÀŽ â $z ) p  &Ž â zÿ t    â $z |  &Ž â r++  €ÿàÿ â $z |   Ž â x )   ÿàÿ â $z v  Ž âr4ÿÿ   ÿä â $z
 v  Ž â z  q      â $z  p  Ž â r ++   ÿ� â %x

8   ÿŽ â r%6   ÿŽ Ä r06   ÿŽ â$x-   6Ž âx þÿÿÿÿÀŽ â $x2   6Ž â x+   ÿ   â %x8   ÿŽ â xÿ      ä r+  €ÿàÿ â%x+   
Ž â xÿ     Ä r8
   ÿàõ â %x   
Ž â r:   ÿà÷ ä z/  Y      â $x7   ÿŽ â r* ÿ       â %x   Ž â x5ÿ   +  ä z36 q  ÿàÿ â %x+   Ž â r<   ÿä Ä r.   ÿäÿ ä r1   ÿàõ ä r18   ÿà÷ ä r/   ÿä ä r/<   ÿäÿ ä r1   ÿàõ â �y(    ™ ¢  r1:   ÿà÷ Ä r/   ÿä â �y‚    ‘ â r/.   ÿäÿ Æ �y=    ™ ( �y„    ‘ ( �y !   ‘ ( �y…!   ‘ " r>+,   6àµ Ä xi $       ä r54   ÿd~ ä xA        â %v> Z  i Ž È/ $x$   ÿŽ â r        Èr   ÿàÿ Ê �y‰   ™ h �y‡   ™ h �y}   ™ h �yU   ™ h �yR   ™ h �yP   ™ h �y?   ™ h �yH    ™ b x;(ÿ   ÿÀŽ ÄO x@‚   ÿ   ä� x=ÿ   ÿÀŽ äx;@   ;âŽ ä x„   ÿ   ä x†‚ð   ÿÀŽ â %v~;  A Ž â x   âŽ Ä xƒ„ð   ÿÀŽ ä xÿ   (  â %v€  A Ž â xÿ   =  â �y~~    ™ ¢ x†ÿ   †  ä xƒÿ   ƒ  â �y€€    ™ " xÿ   ÿÀŽ Ä xÿ   ÿÀŽ ä x†   ÿ   ä xƒ   ÿ   ä x   âŽ ä x   âŽ Æ %v  A Ž È %v  A Ž â �y    ™ ¨" �yh    ™ ¢  xˆ ð   ÿÀŽ ä xÿ   (  ä xÿ   =  Ä x�…ð   ÿÀŽ ä�xˆÿ   ˆ  ä x;    ÿ   ä xÿ   ÿÀŽ ä x>…   ÿ   ä xÿ   ÿÀŽ Ä x�ÿ   �  ä xˆ   ÿ   ä x;   âŽ ä/ x>   âŽ ä x;�   ÿ   ä x   ÿÀŽ â %v  A Ž â x;   ÿÀŽ Ä r(   ÿ@� â %v  A Ž â�yO    ™ â  r@=   ÿ@� ä %v  A Ž â�yT    ™ æ %v@@  A Ž â �yS    ™ è  �yQ@    ™ â  r
1   ÿàõ È/ r/   ÿä ä rF1   ÿàõ är1   ÿà÷ â �y=2   • â  rG/   ÿä är/   ÿäÿ â �yC4   • â  rV1   ÿàõ Ä r+,   3àÇ â �yNF0   • â rW/   ÿä ä r(54   ÿ„þ â �y>0   • è  �yA2   • â  %v Z  i Ž Æ �yM6   • â  $xE($   ÿŽ Æ �y;4   • è  �yLV0   • â  rE   ÿàÿ Æ �yD0   • è  �y(6   • è  �yE   ™ è  �y@   ™ è  �yK   ™ è  �yJ   ™ è  �yB   ™ è  �yG   ™ è" �yF   ™ è �yI    ™ â r+,   %àµ Æ �yV    • â  r54   ÿd~ â %v Z  i Ž ä �y_    • ä $x$   ÿŽ Ê r   ÿàÿ Ê �yX   ™ è �yW   ™ è �yZ   ™ è �y[   ™ è �y]   ™ è �y\   ™ è �y^   ™ â r+,   -àµ Æ �yY    ™ â r54   ÿd~ â %v Z  i Ž È $x$   ÿŽ Ê r   ÿàÿ Ê �yg   ™ è  �yd   ™ è  �yc   ™ è  �yf   ™ è  �ya   ™ è  �ye   ™ è  �y`   ™ â  r+,   0àµ Æ �yb    ™ â  r+,   2àÙ ä/ r54   ÿd~ â�yr    • â  rj54   ÿ¤~ â %v Z  i Ž È $xk$   ÿŽ ä %v Z  i Ž È $xij$   ÿŽ Ê ri   ÿàÿ ä �yi    • â rk   ÿàÿ Æ �ys   ™ è �yl   ™ è  �yq   ™ è  �yn   ™ è  �yk   ™ è  �ym   ™ è  �yo   ™ è  �yj   ™ è  �yp    ™ è  �yt    • è  �y{   ™ è �yz   ™ è �yx   ™ è �yu   ™ è �yy   ™ è �yw   ™ è �yv    ™ â x|~ÿÀŽ ÆO �y    • ¢ x~ÿ   ~  ä xÿ   €  äx€ÿÀŽ ä x€~ÿÀŽ ä x~ÿÀŽ â &r|‰   ÿ   â
x‚‚   ÿÀŽ â &z‰  € ÿ   â x„„   ÿÀŽ â &r‰   ÿ   Ä &r‰€‡      âx    ÿÀŽ â &z‡  €    â Er  ‚       â &rŠ~‡      â x‡    =    ä Er„ „       ä x‚ÿÀŽ ä xhÿÀŽ â/ #t„  `¿‡  â xˆ   ÿÀŽ â &r„‚}   ‰   Ä &rŒ}   Š   ä &zˆ}  € ÿ   â x}ÿ     ä x††   ÿÀŽ ä xƒƒ   ÿÀŽ â &zˆU  € ˆ   â x}}ÿÀŽ â #t    `¿‡  â Er† †       Ä Er        ä ErŽ ƒ       â &r„}U   „   â xƒÿ   h  â #rh    ÿ   â x……   ÿÀŽ ä ErŠ        â #t†  `¿‡  â Erˆ ˆ       Ä xƒƒÿÀŽ ä x��   ÿÀŽ ä Er …       â #r…ˆ   h   â Er„ „       â &rŒƒU   Œ   â Er� �       â #t�Ž  `¿‡  â Er        â #r   ÿ   â xUOÿÀŽ â� !r……„       â x†ÿ   O  â &zOR  € ÿ   â x„ÿ   T  â #rˆˆ�      â xTTÿÀŽ â #t  `¿‡  âx††ÿÀŽ â #t  `¿‡  â#tŠ  `¿‡  â#th�  `¿‡  â &rUR   ÿ   âx„„ÿÀŽ â &z‡P  € O   â xOSÿÀŽ â &r‹TR   ÿ   â xSÿ   S  â &z‰?  € ÿ   â xŠÿ   Q  â &rR†P      âxQQÿÀŽ â &rP„P   ‹   â Er ‡       â &z‰H  € ‰   â x‡SÿÀŽ â &rRO?   R   âxSŠÿÀŽ â &r?Q?   P   â Er‰ ‰       â &rR‡H   R   â#rP   ÿ   â &r?SH   ?   â x=ÿ   =  â #rH   ÿ   â Er R       â #rP‰   P   â x==ð   ÿÀŽ Ä rNÿ7   N  ä Er� ?       â !r?P       â xCÿ   C  â #rH‰h   H   â r>=   ÿ � ä xAÿ   A  ä x=N   ÿ   Ä x‰        ä xNÿ   N  ä xMM ð  ÿÀŽ ä x>C   ÿÀŽ ä xP        ä xAAð   ÿÀŽ Ä x;ÿ   ;  ä xNN   ‰ìŽ ä rLÿ7   L  ä rM   >àÿ ä rADA   ÿ � ä x>=   PìŽ Ä x=( ð  ÿÀŽ ä xD;   ÿÀŽ ä x;L   ÿ   ä Er( N       ä xC        ä xLÿ   L  Ä r==A   Dàÿ â  rD?(     A  â xA;   PìŽ ä Er> >       ä x;L   CìŽ â &zCE  € ÿ   ä #r……>   D   â &zMK  € ÿ   Ä &zD@  € C   â Er; ;       â &rC|E   ÿ   â!rHH�       â &zMJ  € M   â ErD D       â &rEE   ÿ   â Er? A       â  rAH;     A  â Er‹ Œ       â &rC€@   C   âErH M       â &rE~@   E   â #r@ D   ÿ   â#rDD   ÿ   â !rˆˆ‹       â &rC‚K   C   Ä &rEK   E   â #rK�H   D   â &zDB  € ÿ   â #rˆˆ?   A   â &rC}J   C   â #r@H   @   â &rAUB   ÿ   ä &zDG  € D   ä &zHF  € ÿ   Ä &rBTB   ÿ   ä &rJƒJ   E   â ErE C       â &rA†G   A   âErC D       â &zHI  € H   ä &rG„G   B   ä &rBOF   A   â !rE@E       â ErA H       â #r@C   ÿ   â#rCC   ÿ   â &rB‡I   B   Ä #r@A   @   â#rDhA   C   ä ErA B       â 0rVÿV    A   ä 0rÿ    A   ä &rFQF   G   â !rC@A       ä  rAV     A  â &rFSI   F   â  rC(C     A  Ä #r"A…   "   ä #rA>E   C   â ErG F       â &zEX  € ÿ   â ErJ J       â &zBW  € ÿ   ä &zEZ  € E   â !rDDG       â &rC|X   ÿ   Ä &rGX   ÿ   â !rJKJ       â &zB[  € B   â  r@;D     A  â &rC€Z   C   âErE E       â &rZ~Z   G   ä &zG]  € ÿ   â #r@?J   @   â ErF B       â #rB E   ÿ   â#rJE   ÿ   â &zH\  € ÿ   Ä &zG^  € G   ä &rEU]   ÿ   ä &zHY  € H   â ErG G       â &rE†^   E   ä &rD‚W   C   â #rCF   B   â#rJ�F   J   â &rFO\   E   â ErE H       â &rD}[   D   â #rBG   ÿ   â#rIG   ÿ   â 0r=ÿ=    A   Ä #rGE   B   âErD D       â 0rBÿ_    A   â #rIhE   I   Æ !rECD       â  rCB     A  â rB=B     A  â &rF‡Y   F   Æ #rB@      â &zBg  € ÿ   â #r CA       â ErF F       â &zCd  € ÿ   ä &r@|g   ÿ   ä &zBc  € B   â !rGGF       â &r]T]   ÿ   Ä &zCf  € C   â ErB B       â &rA€c   @   ä &z@a  € ÿ   â  rG(G     A  â &r]„^   ]   â ErD C       â &rCUa   ÿ   ä &zFe  € @   â #r@ B   ÿ   â &r\Q\   ]   â #rE>E   G   â #rBB   ÿ   â &zG`  € ÿ   Ä &rC†e   C   â #rH�D   B   â &r\SY   \   ä &raTa   ÿ   ä &zGb  € G   ä &rBO`   C   â ErC F       â &rgg   ÿ   â Er\ \       â &ra„e   a   â #r@D   @   â &rA‚d   A   â ErG G       â &rD‡b   B   â #rBC   ÿ   â &rZW   Z   Ä &rg~c   g   ä &raQ`   a   â !r\I\       â &rA}f   A   â ErI D       â &rZƒ[   Z   â #rBG   B   â &rgd   g   ä &raSb   a   â ErA A       â #rCC   ÿ   â &rgƒf   g   â !rIBI       â ErK Z       â #rChG   C   â ErB a       â !rA@A       â 0r@ÿi    A   â  rI(I     A  â Erg g       â !rJJK       â  r\;\     A  â !rBCB       â 0rrÿr    A   â #rI>A   I   â  rA@     A  â #rJ?J   \   â !rHHg       â  r@=@     A  â  rB;B     A  â  rCr     A  â rr=r     A  â #r@J      â #rB?H   B   â &r@|l   ÿ   â #rAE      â &zDl  € ÿ   â #rrB      â &rll   ÿ   Ä &r@€q   @   ä &zBn  € ÿ   ä &rAUn   ÿ   â #rCI      â &rl~q   l   ä &r@‚k   @   ä &zCm  € B   Ä &rB†m   A   ä &zEo  € ÿ   ä &rlk   l   ä &rA}j   @   â ErC C       â &zDq  € D   ä &zkk  € ÿ   Ä &rBOo   B   ä &z@p  € E   â ErD D       â &rnTn   ÿ   ä &zkj  € k   â ErG @       â &rE‡p   B   â #rFC   ÿ   â &rn„m   n   â ErB k       â #r@ D   ÿ   â ErE E       â #rFG   F   â &rnQo   n   â ErA A       â #r@B   @   â !rEFE       â &rnSp   n   â #rCC   ÿ   â !rA@A       â  rE(E     A  â 0rtÿt    A   â Ern n       â #rChG   C   â &z@s  € ÿ   â #rE>A   E   â  rAt     A  â &r||s   ÿ   â!rnCn       â &rs   ÿ   â #rAE      â &z@{  € @   Ä &zCz  € ÿ   ä &rA€{   |   ä&r{~{      â Er@ @       â &zCx  € C   ä &r{z   {   ä &rUUu   ÿ   Ä &zu  € ÿ   ä &rTTu   ÿ   â Er C       â #r  @   ÿ   â #rDD   ÿ   â#r@@   ÿ   â &r††y   U   ä &zw  € ÿ   ä &zy  €    Ä &rT„y   T   â #r        â &r†Ow   †   ä&zv  €    â Er        â &rTQw   T   ä &rA‚z   A   â Er        â &r†‡v   †   Ä &rTSv   T   â #r   ÿ   â&rA}x   A   â #rC   ÿ   â x&&   ÿàÿ â &rlƒj   l   âEr †       â &r{ƒx   {   â ErT T       â #r      â#rCh   C   â ErA A       Ä | &   pbô â #rD�B   D   âErG l       â #r@�   @   â Er{ {       â !r       â !rTCT       â !rA A       â 0r ÿ    A   âO !rDDG       â  rn;n     A  â r(     A  â !r@@{       â  rT;T     A  â  rVV=     A  â  rt=t     A  â #rD?D   n   â #r>A      â  r      A  â #r@?@   T   â  r==      A  â x11À  ÿà÷ â #r!Vˆ   !   â #r9tD   9   â #r'   '   â #r*=@   *   â x%%@   ÿàÿ Ä x--@   ÿàÿ ä x00@   ÿàÿ ä x22@   ÿàÿ ä x33@   ÿàÿ ä x66@   ÿàÿ ä r/ÿ/   ÿäÿ â D)       À â Gy  �Ûÿÿÿÿƒ ê Ay        € ê�Ey  ð     € â G‰  Ð     € ê $x##€  $Ž Ê ˆs #" úÿ H   è ˆs #!€úÿ H   è ˆs #  ûÿ H   è ˆs #€ûÿ H   è ˆs # üÿ H   è ˆs #€üÿ H   è ˆs # ýÿ H   è ˆs #€ýÿ H   è ˆs # þÿ H   è ˆs #9€þÿ H   è ˆs #' ÿÿ H   è ˆs #*€ÿÿ H   ä Ay        € ê {          ì M        € ê „y $ €   H   b z ) r  p`ð â $z x  )Ž Ä „y$     H   ¢ r $ÿ   p"p Æ „y$ €  H   è „y$    H   ( „y$ €  H   h „y$    H   ( „y$    H   ( „y$ €  H   ( „y$    H   è „y
$ €  H   è „y$    H   â !r! !       â? !r"       ÈO ‰!      " !r       Æ� ‰       b !r         Æ„y$ €  H   â !r       Æ‰      ¢ !r       Æ ‰       â !r       Æ ‰      " !r       â !r!       ä ‰      " !r       â z!  r      â !r         â/ ‰      b !r

9       Æ ‰      b !r'       â r'!!   ÿŽ â !r       âO ‰      ¢ !r        Æ� ‰
      â !r       Æ‰      " !r**       Æ ‰      " !r       Æ ‰       " !r       â/ !r       ä‰      b x)   ÿàÿ â !r       äO ‰      ¢ z  r  p`ò â !r
       Ä� ‰      â x $   p"ò â !r       ä‰      " r)!!   ÿŽ â !r       ä ‰      b !r        Æ ‰      " $z 	 ~  Ž Æ ‰      " !r       â/ 
  r  ÿà÷ ä ‰      b !r       ÆO ‰      ¢ !r       Æ� ‰*      â !r       Æ‰ €    " !r       Æ‰ €    b !r       â !r       ä ‰ €    â !r       Æ/ ‰ €    " ÿÿ   ÿäÿ â !r       äO ‰ €    b 
 j  ÿ† â!r*       ä� ‰ €    ¢ 
 k  � â !r
       Ä‰ €    â  '   ÿàõ â !r       ä‰ €    " r!    ÿˆ ä ÿÿ   ÿä â ‰ €    b rÿÿ   ÿä ä z j  ÿˆ â ‰  €    b !r       â 
 j  ÿ„ Ä ‰" €    " z k   â !r       â !r       â/ ‰&      b !r       âO 
 k   ä ‰
 @    ¢ !r       â� r!    ÿ„ Ä ‰	 @    â !r       â )   ÿàù â !r       â‰ @    " r(ÿÿ   ÿä â !r         ä ‰ @    b $ÿÿ   ÿä â !r""       Ä ‰ @    " 
 j  ÿˆ â !r&&       ä/ ‰ @    b !r
       ÆO ‰ @    ¢ !   ÿàÿ â !r		       ä� ‰ @    â z
 j  ÿ„ ä*    ÿà÷ â ‰ @    â !r       âz k  ( Ä ‰!  @    " !r       â
 k  $ ä ‰#" @    b !r       â +ÿÿ   ÿäÿ ä ‰%& €    " !r       â/ * j  ÿ† Ä ‰	      b !r       âO * k  +� ä ‰      ¢ !r       Æ� ‰      â !r       â !r! !       ä‰      " !r#"#       Æ‰      b !r%&%       Æ ‰      " !r+	       Æ/ ‰      b 	'   ÿàÿ â !r-       äO ‰      ¢ x'        ä  	   ÿàõ â‰      ¢ !r       â� %v  j  ' Ž â &ÿÿ   ÿä â ‰ !      â  j  ÿ„ â !r       Ä‰$% @    " !r       â k  & ä ‰"#      b !r       â !r       ä/ †	 +    â !r       ÆO † -   â !r       Æ †	     â !r!!        Æ� †     â !r$%$       Æ†	     â !r##"       Æ†    è †	     è †     è †	 
    è † 
!   è ‰$      h  †	 #    â M™        € ê x)   ÿàÿ â !r$       Æ/ r    ÿàñ È r ÿÿ   ÿä  ä z j  ÿ€ È z k     Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      "   " ¹z  l      â Ey  �$    € â ™x?    â y      !   b ¹z  F   
   â r  ÿ       â x      ÿ  â y%      %   ¢ r& ÿ       ä r3 ÿ       â y      &   â r5 ÿ       Ä r7 ÿ       ä r# ÿ       ä r$ ÿ       ä r; ÿ       ä r ÿ   pRð är ÿ   pRò Ä r'   ÿ(Ž È/ x ÿ   '  ä x%%   ÿ   äO r  '   ÿ� È x"ÿ      È | "   pbô Ú G)  #    € ê y      '   "Ž r ÿ       â $z w  ÿŽ âr ÿ       ä z-  q      ä x0"   ÿ   â 'z s   Ž â z!  X      ä r)-0   ÿŽ â$x,-   0Ž â zÿ t    Ä z+  Y      ä r  ÿ       ä r--0   ÿŽ ä z.0 q  ÿàÿ â 'z y   Ž È $z( }  ÿŽ â zÿ z    ä x øÿÿÿÿÀŽ â $z % p  "Ž âr((   ÿàõ â $z	 |  "Ž â r''  €ÿàÿ â $z |   Ž â x %   ÿàÿ â $z	 v  	Ž âx'   ÿ   â$z v  Ž â r2ÿÿ   ÿä â $z
  p  	Ž â r ''   ÿ� â %x8   ÿŽ â xÿ     Ä x þÿÿÿÿÀŽ â %x

8   ÿŽ â xÿ      ä r'  €ÿàÿ â%x'   Ž âr4   ÿàõ är6
   ÿà÷ â %x   Ž â x/ÿ   '  Ä r8	   ÿä â$x1   ÿŽ â r*   ÿäÿ â %x   
Ž È %x'   
Ž È r!   ÿàõ ä r!4   ÿà÷ ä r+   ÿä ä r+8   ÿäÿ ä r!   ÿàõ â �y    ™ ¢  r!6   ÿà÷ Ä r+   ÿä â �yK    ‘ â r+*   ÿäÿ Æ �y    ™ ( �yP    ‘ h �yO!   ‘ h �yM!   ‘ b r'(   0àµ Ä xW $       ä r/2   ÿd~ ä xE        â %v Z  W Ž È/ $x$   ÿŽ Ê r   ÿàÿ Ê �yL   ™ h �yQ   ™ h �yC   ™ h �y<   ™ h �y?   ™ h �y>   ™ h �y9   ™ h �y:    ™ h �y^    • b x ÿ   ÿÀŽ ÄO xK   ÿ   ä� xÿ   ÿÀŽ äx     âŽ ä xP   ÿ   Æ%v   E Ž â x   âŽ È �y     ™ ¢ %v  E Ž Ê �yS    ™ " xNKð   ÿÀŽ ä xGPð   ÿÀŽ ä xÿ     ä xÿ     ä xNÿ   N  ä xGÿ   G  Ä xÿ   ÿÀŽ ä xÿ   ÿÀŽ ä xN   ÿ   ä/ xG   ÿ   ä� x   âŽ ä x   âŽ Æ %v  E Ž È %v  E Ž â �yB    ™ è �yA    ™ â
 xÿ     ä xÿ     ä xO   ÿ   Ä xÿ   ÿÀŽ ä xIOð   ÿÀŽ ä xDMð   ÿÀŽ äxM   ÿ   ä xÿ   ÿÀŽ ä x   âŽ Ä xIÿ   I  ä xDÿ   D  ä x   âŽ â/ %v  E Ž â
xI   ÿ   ä xD   ÿ   â %v  E Ž â x   ÿÀŽ â �y=    ™ b x   ÿÀŽ Ä r   ÿ@� â �y@    ™ b r   ÿ@� Æ %v  E Ž È %v  E Ž ä �y    ™ h �yH    ™ b r!   ÿàõ È/ r	+   ÿä ä r!   ÿàõ är
!   ÿà÷ â �yE2   • b r+   ÿä är+   ÿäÿ â �yF0   • b r!   ÿàõ Ä rX'(   .àÇ â �yJ0   • b r+   ÿä ä rY/2   ÿ„þ â �yR4   • h �y2   • b $xYY$   ÿŽ Æ �yT6   • b %vX Z  W Ž Æ/ �yV4   • h �y0   • b rY   ÿàÿ Æ �yU0   • h �yX6   • h �yZ   ™ h �y[   ™ h �yY   ™ h �y\   ™ h �y]   ™ h �y   ™ h �y   ™ b r'(   )àµ Æ/ �y    ™ b r_/2   ÿd~ â %v Z  W Ž È $x__$   ÿŽ â �y    • h r_   ÿàÿ Ê �yc   ™ h �ye   ™ h �yf   ™ h �y`   ™ h �yb   ™ h �y_   ™ h �yd   ™ h �ya    ™ b r'(   ,àµ Æ �yg    • b rh/2   ÿd~ â %v Z  W Ž È $xih$   ÿŽ Ê ri   ÿàÿ Ê �yj   ™ h �yn   ™ h �yl   ™ h �yh   ™ h �y   ™ h �yk   ™ b r'(   -àµ Æ/ �yi    ™ b ro/2   ÿd~ Æ �ym   ™ b %v Z  W Ž Æ �yq    • b $xoo$   ÿŽ Ê ro   ÿàÿ Ê �yW   ™ h �yp   ™ h �yo   ™ h �yv   ™ h �yr   ™ h �yw   ™ h �yu   ™ h �yt    ™ b xs ÿÀŽ ÄO xyÿ   S  äxSSÿÀŽ â &zxL  € ÿ   ä &rzsL   ÿ   ä&r|SL   ÿ   ä �yL    • ¢ x{ÿ      Ä xyyÿÀŽ ä x{{ÿÀŽ ä xP   ÿÀŽ â &r|yQ   |   âxKK   ÿÀŽ ä xN   ÿÀŽ â/ &rz{Q   z   â xMM   ÿÀŽ â &zxQ  € x   â Er  K       â &zKC  € ÿ   â Er M       Ä xPBÿÀŽ ä� xNAÿÀŽ Æ &rzPC   z   âxII   ÿÀŽ ä xQ    =    â &rMNC   |   â xDD   ÿÀŽ ä ErC        ä Er I       ä xBÿ   B  Ä Er        ä xGG   ÿÀŽ ä ErI D       â #tDC  `¿Q  âxOO   ÿÀŽ â &zK<  € K   â xCBÿÀŽ â #tB  `¿Q  â Er| G       â #t    `¿Q  âEr x       â #t  `¿Q  â Er O       â &rzC<   z   â xAÿ   A  â #tG|  `¿Q  â ErK K       â #rO    ÿ   â #rx   ÿ   â #t  `¿Q  â#t  `¿Q  âxAAÿÀŽ â #tQI  `¿Q  â Er z       â #rzDK   O   â #rIKG   x   â xK=ÿÀŽ Äx}ÿ   =  ä xOÿ   @  ä x=@ÿÀŽ â &rMA<   M   â !r<z       â x}}ÿÀŽ â &z?  € ÿ   â xOOÿÀŽ â &r@K?   ÿ   Ä &rx=?   ÿ   ä &z>  €    ä &r@}>   @   âxzÿ     â &r>O>   x   â xxÿÀŽ â &z?9  € ÿ   â xÿ   H  ä xHHÿÀŽ â &z?:  € ?   â xÿÀŽ â &r@x9   @   â Er        â &r9H9   >   â Er> ?       Æ &r9:   9   â #r?   ÿ   â xzzÿÀŽ â #r   ÿ   â xEÿ   E  â #r?B>   ?   â Er9 9       â #r>>Q      â &r@z:   @   â xEEð   ÿÀŽ ä Er: M       Ä rJÿ1   J  â !r>9       â rFFE   ÿ � ä xRÿ   R  ä x>ÿ     â !rII:       â xEJ   ÿ   ä x9        Ä xJÿ   J  ä Er@ @       ä x:        ä xTT ð  ÿÀŽ ä xR   ÿÀŽ ä x>>ð   ÿÀŽ Ä xVÿ   V  ä rÿ1     ä xJJ   9ìŽ â !r??@       â x:E   :ìŽ ä rTF   àÿ ä r>U>   ÿ � Ä xXX ð  ÿÀŽ ä xEV   ÿÀŽ ä x@   ÿ   ä xFÿ     ä Er J       ä xM        Ä r>X>   Eàÿ ä xE@   9ìŽ â  r??     A  â Er9 :       â &zJZ  € ÿ   â xFF   MìŽ â &zM[  € ÿ   ä #r@<9   ?   â Er: F       â &z?Y  € J   â Er< E       â &rJsZ   ÿ   Ä &zM\  € M   â  rF:     A  â Er? ?       â &rJ{Y   J   ä #rII<   F   â ErF M       â #r ?   ÿ   â#rR?   ÿ   â &rJP[   J   ä &zM]  € ÿ   â #r?DF      â&rEK]   ÿ   â #rGF   R   â &rJC\   J   Ä &zF  € ÿ   ä &zM  € M   ä &r]=]   ÿ   ä &rE}   E   â ErJ J       â &zF  € F   â ErM M       â &r]O   ]   Ä &rEx   E   ä&rZSZ   ÿ   â !r?J       â &r]H   ]   â ErF F       â #rM   ÿ   â#rJM   ÿ   â &rEz   E   ä &rZyY   Z   â #rBF      â#rJQF   J   â ErF E       â &rZN[   Z   Ä 0r^ÿ^    A   ä 0rÿ    A   ä &r]   ]   â !rF       â &r\A\   Z   â  rF^     A  ä Er] ]       â 0r>ÿ>    A   â Er\ \       â #rF@      â &z@c  € ÿ   â !r]J]       â  rJ     A  â !r\       â  r^>     A  â &z?e  € @   â  r]:]     A  â #r9   J   â #rI      â &zE`  € ÿ   â #r<   ]   â ErJ ?       â &zf  € ÿ   Ä &r?K`   ÿ   ä &rsf   ÿ   ä&rfSf   ÿ   ä &zb  €    ä &zR_  € ÿ   ä &zEd  € E   Ä &r?}d   ?   ä &r`=`   ÿ   ä &r{b      âEr        â &rfyb   f   â ErE E       â &zRa  € R   ä &r?x_   ?   Ä &r`Od   `   ä &rFPc      âErR R       â &r@Nc   f   ä &rIza   ?   â #r    ÿ   â &r`H_   `   â #r?E   ÿ   â #rT   ÿ   â &rFCe   F   Ä &reAe   @   â Er@ I       â #rDJ      â #r?BR   ?   â &r`a   `   â #rGJ   T   â #rJE   ÿ   â ErF F       â !r??@       â ErT e       â 0rÿ    A   â ErE `       â #rJQR   J   â !rFF       â  r@?     A  â !rT       â !rEJE       â  r>     A  â #rF9F   @   â 0rgÿg    A   â  r@     A  â  rE:E     A  â #r$   $   â &rsj   ÿ   â #r;@   ;   â #r<   E   â  r>g     A  â &r{n      Ä &z@l  € ÿ   â #r7   7   â &rPh      â  rg     A  â &zj  € ÿ   ä &z  € @   ä &rjSj   ÿ   ä &r@Kl   ÿ   â #r#F   #   â &zn  €    Ä &ryn   j   ä &zFk  € ÿ   ä &r@}   @   ä &zh  € ÿ   â Er        â &rhNh      â Er        â &zFi  € F   Ä &r@xk   @   ä &zm  €    ä &rl=l   ÿ   â #r    ÿ   â &r@zi   @   â #rJ   ÿ   â ErF F       â &rlO   l   â #r   ÿ   â Er        â &rCm      â Er@ @       â &rlHk   l   â #rBF      â #rD      â#rJG   J   â Er        â #r   ÿ   â &rli   l   â !r@       â &rhAm   h   â #rFQF      â !r       â  r     A  â Er l       â &zW  € ÿ   â Er h       â &zo  € ÿ   Ä &zp  €    â !rF       â !rJ       â &zv  €    ä  r:     A  â Er        â #r9      â Er@        â #r<      ä #r    ÿ   â#r    ÿ   Æ #rD@      â#rGG@       â &z@r  € ÿ   ä &r Kr   ÿ   ä&rr=r   ÿ   ä &rssW   ÿ   ä &zu  € ÿ   Ä &z@w  € @   ä &r }w       ä &rSSW   ÿ   ä &rrOw   r   ä &rs{p   s   ä &zt  €    â Er @       â &rxu       Ä &rSyp   S   ä &ruHu   r   â Er         â &rsPo   s   â #r   ÿ   â &rSNo   S   ä &rzt      ä&rut   u   â #r   ÿ   â &rsCv   s   âx""   ÿàÿ â &rSAv   S   â #rQQ       âEr        â #rB       â Er@ u       â 0rqÿq    A   â Er  s       Ä | "   pbô â !r       â Er S       â !rQQ@       â !r        â  r q     A  â 0rLÿL    A   âO !rGG       â  r     A  â  rQ:Q     A  â #r5    5   â  r >q     A  â #r9      â #rG<G   Q   â  rL     A  â r>L     A  â x!!À  ÿà÷ â #r3    3   â #r&   &   â #r G       â x))@   ÿàÿ Ä x,,@   ÿàÿ ä x--@   ÿàÿ ä x..@   ÿàÿ ä x00@   ÿàÿ ä r+ÿ+   ÿäÿ â D)       À â Gy  àßÿÿÿÿƒ ê Ay        € ê� Ey  Ð     € â G‰  °     € ê $x@  Ž Ê ˆs  ûÿ H   è ˆs €ûÿ H   è ˆs ; üÿ H   è ˆs $€üÿ H   è ˆs # ýÿ H   è ˆs 7€ýÿ H   è ˆs 5 þÿ H   è ˆs 3€þÿ H   è ˆs & ÿÿ H   è ˆs  €ÿÿ H   ä Ay        € ê {          ì M        € ê „y      H   b z % r  p`ð Ä z!  r      â „y €   H   ¢ r ÿ   p"p Æ „y    H   è „y	 €  H   ( „y    H   h „y €  H   ( „y    H   ( „y    H   ( „y
 €  H   è „y €  H   â !r         â/ !r       ÈO ‰       b !r;;       ÆŸ ‰      " !r		$       â!r##       ä‰;      ¢ !r77       Æ ‰	      â !r&       â !r55       â !r        â/ ‰#      b !r3
3       Æ ‰       " !r       â !r        ä ‰
7      ( ‰      b !r;       ÆO ‰5      ¢ !r	       Æ� ‰3      è ‰      â !r#       â/ !r         ä‰      b !r
7
       Æ ‰	      " !r       Æ‰  €    " !r5       ÆO ‰      ¢ !r3       Æ� ‰
      â !r       Æ ‰      h ‰      b !r       Æ/ ‰      b !r		       Æ ‰      " !r        Æ‰      " !r       âO $z x  %Ž â !r
       â� ‰ €    ¢ x
%   ÿàÿ â !r       ä‰	 €    â z 
 r  p`ò â !r       Ä ‰  @    b x    p"ò â !r       ä/ ‰ €    b r!!   ÿŽ â !r       ä y      '   " !r       Æ‰ €    ( ‰ €    " !r       ÆO ‰ €    ¢ !r	       Æ� ‰ €    â !r         Æ‰ €    h ‰ €    b $z
 ~  Ž â !r       â/ !r       â‰       $ r!
   ÿ† â !r       â ‰ @    d rÿÿ   ÿäÿ â !r       âO ‰ @    ¢ z j  ÿ† â !r       Ä� ‰ @    â 
   ÿàõ â !r       ä‰ @    " z k  � â !r       ä ‰ @    b  ÿÿ   ÿä ä 
 j  ÿ„ â ‰ @    b !r#        â !   ÿàÿ Ä ‰ @    " !r       â/ r!
   ÿŠ ä ‰ @    b !r       âO    ÿàÿ ä ‰ @    ¢ !r       â� 
 k    â !r       â‰       â 
	
 r  ÿàù â !r       Ä‰      " !
   ÿàõ â!r       â x        â ‰      b 
   ÿà÷ â!r       â r"ÿÿ   ÿäÿ â ‰      " %v

 j   Ž â !r       â/ z j  ÿŠ â ‰      b $ÿÿ   ÿä â !r       âO z k  "� â ‰      ¢ 
	 j  ÿˆ Ä "ÿÿ   ÿä â ‰      ¢ ! j  ÿ„ â !r%        â�  ÿÿ   ÿäÿ â ‰      â 
		 k  $ â !r       â j  ÿ† â ‰      " ! k  " â !r       â k   � â !r       â †	 
#    â !r       Æ/ † 
%   â !r       ÆO †	     â !r       Æ †     â !r       Æ� †	     â !r       Æ†    è †	     è †     è †	     â M™        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y!      "   " ¹z  l      â Ey        € â ™x?    â y"      !   b ¹z  F   
   â r& ÿ       â x      ÿ  â y'      %   ¢ r1 ÿ       â x      ÿ  â r5 ÿ       â y      &   â r  ÿ       Ä r !ÿ   pRò är !ÿ   pRð är)!"   ÿ(Ž È/ x ÿ   )  ä x''   ÿ   äO r  )   ÿ� È x8ÿ      È | 8   pbô Ú G)  Ð    € ê y      '   "Ž r ÿ       â $z w  ÿŽ âr ÿ       ä x.8   ÿ   ä z%  X      â 'z s   Ž â z+  Y      ä r& ÿ       Ä zÿ t    ä z,. q  ÿàÿ â 'z y   Ž È $z$ }  ÿŽ â zÿ z    ä x øÿÿÿÿÀŽ â $z ' p  8Ž âz  q      â $z	 |  8Ž â r))  €ÿàÿ â $z |   Ž â x '   ÿàÿ â $z	 v  	Ž âr$$   ÿàõ â $z v  Ž â r#.   ÿŽ â $z
  p  	Ž â r ))   ÿ� â $x*   .Ž â x)   ÿ   â %x8   ÿŽ â x þÿÿÿÿÀŽ Ä xÿ      â %x

8   ÿŽ â r)  €ÿàÿ ä r0ÿÿ   ÿä â %x   Ž â xÿ     ä r2   ÿàõ â$x/   ÿŽ â r4
   ÿà÷ â %x)   Ž â x-ÿ   )  Ä r6	   ÿä â%x   
Ž â r(   ÿäÿ Æ %x)   
Ž È r%   ÿà÷ ä r%2   ÿàõ ä r+   ÿäÿ ä r%   ÿà÷ ä r+6   ÿä â �y    ™ ¢  r%4   ÿàõ Ä r+   ÿäÿ â �yP    ‘ â r+(   ÿä Æ �yL!   ‘ ( �y    ™ h �yM    ‘ ( �yK!   ‘ " r)$   .àµ Ä xd $       ä r-0   ÿd~ ä xS        â %v Z  d Ž È/ $x$   ÿŽ Ê r   ÿàÿ Ê �yQ   ™ ( �yN   ™ ( �yF   ™ ( �yA   ™ ( �yB   ™ ( �y<   ™ ( �y>   ™ ( �y9    ™ ( �yZ    • " x ÿ   ÿÀŽ ÄO xP   ÿ   È� x     âŽ ä xÿ   ÿÀŽ äxM   ÿ   â%vT   S Ž Æ x   âŽ ä x Pð   ÿÀŽ â �yTT    ™ ¢ xIMð   ÿÀŽ â %v  S Ž â xÿ     ä xÿ     ä x ÿ      â �yO    ™ " xIÿ   I  Ä xÿ   ÿÀŽ ä xÿ   ÿÀŽ ä xI   ÿ   ä x    ÿ   ä� x   âŽ ä x   âŽ Æ %v  S Ž È %v  S Ž â �y@    ™ b xÿ     Æ �yE    ™ b xL   ÿ   ä xÿ   ÿÀŽ ä xJLð   ÿÀŽ ä xÿ     ä xGKð   ÿÀŽ ä x   âŽ Ä xJÿ   J  ä xK   ÿ   ä� xÿ   ÿÀŽ ä xGÿ   G  â %v  S Ž â xJ   ÿ   ä x   âŽ Ä xG   ÿ   â �yC    ™ â  x   ÿÀŽ â %v  S Ž â x   ÿÀŽ ä r   ÿ@� ä rR   ÿ@� â �y?    ™ ä  %v  S Ž È %vRR  S Ž â �y=    ™ è �yDR    ™ â r
%   ÿà÷ är%   ÿàõ ä r+   ÿäÿ Ä r	+   ÿä ä r%   ÿàõ â�y:2   • â  r%   ÿà÷ ä/ r+   ÿä â�y32   • â r+   ÿäÿ Æ �y0   • â  r)$   ,àµ Æ �y0   • â  rV-0   ÿd~ Æ �y;0   • â %v Z  d Ž Æ �yH4   • â $xWV$   ÿŽ Æ �yR4   • ä  rW   ÿàÿ ä �y70   • è  �yS6   • è �yU6   • è  �yW   ™ è �yV   ™ è �y   ™ è �y   ™ è  �yX   ™ è  �yY   ™ è  �y   ™ è  �y    ™ â  r)$   #àµ È/ r[-0   ÿd~ â %v Z  d Ž È $x][$   ÿŽ ä �y[    • æ  r]   ÿàÿ Ê �y^   ™ è �y\   ™ è �ya   ™ è �y`   ™ è �yc   ™ è �y_   ™ è �yb   ™ è �y]    ™ â r)$   *àµ Æ �yg    • â r-0   ÿd~ â %v Z  d Ž È $x$   ÿŽ Ê r   ÿàÿ Ê �y   ™ è  �yj   ™ è  �yi   ™ è  �yh    ™ è  �y   ™ è  �yd   ™ è  �yf   ™ è  �ye   ™ è  �yk    • â  &zQ  € ÿ   â/ xLL   ÿÀŽ Ä xPP   ÿÀŽ â &zN  €    â xMM   ÿÀŽ ä xK   ÿÀŽ ä ErK L       ä Er P       ä Ero        Ä ErM M       ä xpI   ÿÀŽ ä xL    ÿÀŽ ä xlTÿÀŽ äO xmÿ   T  ä xÿ   O  ÄxTOÿÀŽ ä xmmÿÀŽ ä xOÿÀŽ â &rlQ   ÿ   ä&rnTQ   ÿ   ä &rQmN      ä&rnON   n   â xN    =    Ê #tK  `¿N  âxI@ÿÀŽ â#t   `¿N  âxKEÿÀŽ â #tM  `¿N  â xPJ   ÿÀŽ â &rMIF   n   ä&rQKF   Q   ä &zJF  € ÿ   â xFÿ   E  â #to  `¿N  â xrG   ÿÀŽ Ä ErG L       â &zLA  € J   â Ero P       ä xFFÿÀŽ ä Er        â #tEG  `¿N  âx@ÿ   @  â #tGo  `¿N  â &rQFA   Q   â Erq p       Ä Ero L       â #rL    ÿ   â x@@ÿÀŽ â #tJq  `¿N  â Ers r       â #rP   ÿ   â Ern Q       â #rQEo   L   â &r@A   M   â xACÿÀŽ â� #tLs  `¿N  â #rPoJ   P   â !rMQn       â xCÿ   C  â &zoB  € ÿ   â xNÿ   ?  Ä xQ?ÿÀŽ â &zn<  € o   â x?CÿÀŽ â &roAB   ÿ   âxCNÿÀŽ â &rpQB   ÿ   â xB=ÿÀŽ â &ro?<   o   âxNDÿÀŽ â &rqC<   p   â x<ÿ   D  â &rDB>   o   âx=ÿ   =  â &rqN>   q   Ä &z>>  € ÿ   â x==ÿÀŽ ä x<<ÿÀŽ â &z>9  € >   â Ero        ä Er n       â &rD=9   D   ä&rq<9   q   â Er9 >       â #r>   ÿ   â xnÿ   :  â #r:   ÿ   â x3ÿ   3  â #rG9   >   â Er> q       â #r99L   :   â rÿ/     Ä x:3ð   ÿÀŽ â !r39>       â !rPPo       â r9ÿ/     ä xo        ä x   ÿ   ä xq        ä xÿ     Ä r;;:   ÿ � ä x:   oìŽ ä ErD D       ä x   qìŽ ä xs        â !rD       â x>9   ÿ   Ä Er        ä xnnð   ÿÀŽ ä xHÿ   H  ä xRÿ   R  ä xDÿ   9  ä x9>   sìŽ â  r>     A  â Er: :       Ä xu        ä r77n   ÿ � ä xSS ð  ÿÀŽ ä xUU ð  ÿÀŽ ä xHH   ÿÀŽ ä xRR   ÿÀŽ â #rMM:   >   â xD   uìŽ â &zDW  € ÿ   â r;S;   Hàÿ â &r>lW   ÿ   â r7U7   Ràÿ â &rRTW   ÿ   Ä &zHV  € ÿ   ä &zD  € D   ä &r>m   >   âEr        â &rO   R   ä &zH  € H   ä &rKV   >   âEr9 9       â &rVIV      â ErD D       â  r33     A  â Er H       â &zRX  € ÿ   Ä &rHAX   ÿ   ä&rXQX   ÿ   â #rPP9   3   â #r> D   ÿ   â#rSD   ÿ   â &zRY  € R   ä &rH?Y   H   ä&rXCY   X   Ä &z3  € ÿ   â #r>E   >   â#rSJ   S   â &rHB   H   ä&rN   X   â Er R       â &z3  € 3   ä &rF      ä&rV@   V   â Er3 3       â &rH=   H   Ä &r<      â #r   ÿ   â#r   ÿ   â Er H       ä #rG3      â Er        â #rL3      ä !r       â Er        â !r>       â 0r;ÿ;    A   â  r     A  Ä !r       â 0rÿ[    A   â #r:      â 0rZÿZ    A   Æ  r;     A  âErV V       â  r;Z     A  ä #r      â &z^  € ÿ   â !r>SV       â  rD     A  â #r M       â &z\  €    ä &rl^   ÿ   â #r>9>   D   â 0r7ÿ7    A   â Er        â &rDm\      Ä &zHa  € ÿ   â  rZZ7     A  â &rKa   D   â #rD    ÿ   â#r3   ÿ   â &z`  € ÿ   ä &rA`   ÿ   â #rZP      â &zHc  € H   ä &zP_  € ÿ   Ä &zb  €    ä &r?b      â Er H       â &r`Q`   ÿ   â Er        â &zP]  € P   ä &rHB_      â #rDE   D   â#r3J   3   â &r`Cb   `   â Er P       â &r=]   H   â #rH   ÿ   â &rFc      Ä &r`N_   `   â Er        â #rHG   H   â Er        â &r`<]   `   â #r   ÿ   â !rH       ä !rD       â Er` `       â #rL      â  r     A  â &rAA   ÿ   Ä !r`       â #r:      â &z  € ÿ   â  r7     A  â  rD     A  â &rA?j   A   ä &zi  € ÿ   ä &zj  €    â #r>      â &rABi   A   Ä &zh  €    â Er        â 0r>ÿg    A   ä &rA=h   A   â Er        â #r   ÿ   â  r;>     A  â &rQ   ÿ   â ErA A       â #rG      â #r      â &rll   ÿ   Ä &z  € ÿ   â !rAA       â &r^T^   ÿ   ä&rCj      ä &rT   ÿ   ä &rlmd   l   ä &zf  € ÿ   Ä &zd  €    ä &r^O\   ^   ä&riNi      ä &rOd      â Er        â &rlKf   l   ä &ze  €    Ä &r^Ia   ^   ä&ri<h   i   ä &rIf      â #r   ÿ   â &rlFe   l   â Er        â &r^@c   ^   â x88   ÿàÿ â #r     ÿ   â &r@e      â Er i       â #rL      â #r   ÿ   â Er l       â #r E       â| 8   pbô â !r       â Er^ ^       â #rJ      â Er        â !r        â 0r ÿk    A   â !r^3^       â  rAA     A  â  r     A  â !r       â  r>7>     A  â #r9^   D   â #r:   A   â  r;      A  â #r9      â  r77      A  â x%%€  ÿà÷ â #r5>   5   â #r1   1   â #r&7   &   â x##€   ÿàÿ Ä x**€   ÿàÿ ä x,,€   ÿàÿ ä x..€   ÿàÿ ä r+ÿ+   ÿäÿ â D)       À â Gy  äÿÿÿÿƒ ê Ay        € ê� Ey  °     € â G™  �     € ê r!!"   ÿ@Ž Ê ˆs !  üÿ H   è ˆs !€üÿ H   è ˆs ! ýÿ H   è ˆs !€ýÿ H   è ˆs ! þÿ H   è ˆs !5€þÿ H   è ˆs !1 ÿÿ H   è ˆs !&€ÿÿ H   ä Ay        € ê {          ì M	        € ê „y" €  H   b z ' r  p`ð Æ „y"     H   ¢ r "ÿ   p"p Æ „y" €  H   è „y" €   H   ( „y "    H   h „y" €	  H   ( „y"    H   ( „y	" €  H   ( „y"    H   ( „y"    H   ( „y
" €  H   " !r       Æ/ „y"    H   b !r        ÆO „y"    H   ¢ !r       Æ� „y" €  H   â !r       Æ„y"    H   " !r         Æ„y" €  H   b !r       Æ „y!"    H   " !r       Æ „y"  	  H   " !r	       Æ „y"  
  H   " !r         Æ „y" €
  H   " !r       Æ „y#"    H   " !r

5       Æ „y" €  H   â !r1       Æ/ „y" €  H   â !r       ÆO „y" €  H   â !r       â� !r       ä‰       b !r

       Æ‰      ¢ !r!       Æ ‰      " !r       Æ y      '   â !r       Æ ‰	      " !r

       Æ ‰      b z  r      â !r#       ä ‰
      â !r        â/ !r&       ä ‰      b !r       âO !r       Ä ‰       ¢ !r       â !r       ä ‰      " !r		       Æ‰      " !r       Æ‰	      b !r
       Æ� ‰      â !r       Æ/ ‰
      b !r         ÆO ‰      ¢ !r       Æ ‰      ( ‰  €    " !r       â!r	       ä‰ €    " !r       Æ� ‰	 €    â !r
       Æ/ ‰ €    b !r       ÆO ‰ €    ¢ !r       Æ ‰ €    " !r        Æ ‰ €    h ‰      b !r       Æ‰ @    " !r	       â� !r
       ä/ ‰ @    b $z x  'Ž â x''   ÿàÿ â !r       âO ‰ @    ¢ $z  ~  Ž â z ' r  p`ò â !r       â r   ÿŽ â ‰
 @    " 
  r  ÿà÷ â !r       âx "   p"ò â ‰ @    â ÿÿ   ÿäÿ â !r       â 
 j  ÿ† â ‰ @    b !r       Ä
 k  � â ‰ @    " !r	       Æ/ ‰ €    b r    ÿ„ â!r       äO ‰      ¢    ÿàÿ ä rÿÿ   ÿä â !r
       â ‰	      " z j  ÿ„ â !r       Ä� ‰      â     ÿà÷ â !r       ä‰      b z k   â !r       ä‰      "     ÿàõ â !r       Ä/ ‰      b ÿÿ   ÿäÿ â !r       äO ‰ @    ¢ x        ä  j  ÿ† â ‰      ¢ ÿÿ   ÿä â !r	       â 

 j  ÿ„ â %v  j   Ž â 	 k  � â !r       â� 
 k   â !r       â†	     â !r       Æ†    â !r       Æ/ †	     â !r       ÆO †     â !r       Æ †	     è †    è ‰      h  †	 
    â M™        € ê x   ÿàÿ â !r       Æ/ r    ÿàñ È r ÿÿ   ÿä  ä z j  ÿ€ È z k     Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      "   " ¹z  l      â Ey  À    € â ™x?    â y      !   b ¹z  F   
   â r ÿ       â x      ÿ  â y#      %   ¢ r$ ÿ       ä r4 ÿ       â y      &   â r ÿ       Ä r ÿ   pRð är ÿ   pRò är%   ÿ(Ž È/ x ÿ   %  ä x##   ÿ   äO r  %   ÿ� â $!À   Ž Æ x2ÿ      È | 2   pbô Ú G)  p    € ê y      '   "Ž r ÿ       â $z w  ÿŽ âr ÿ       ä x  øÿÿÿÿÀŽ ä x*2   ÿ   â 'z s   Ž â r%%   €ÿàÿ ä z/  X      Ä zÿ t    ä r %%   ÿ� ä z'  Y      ä xÿ      ä r ÿ       ä x)ÿ   %  â 'z y   Ž â z(* q  ÿàÿ Æ $z" }  ÿŽ â zÿ z    â $z# p  2Ž âz  q      ä r""   ÿàõ â $z	 |  2Ž â r*   ÿŽ â $z |  Ž â x#   ÿàÿ â $z	 v  	Ž âx%   ÿ   â $z v  Ž â r,ÿÿ   ÿä â $z
 p  	Ž â x þÿÿÿÿÀŽ â %x8   ÿŽ â xÿ     Ä r%  €ÿàÿ â %x

8   ÿŽ â r-   ÿàõ Æ %x   Ž âr.
   ÿà÷ ä r0	   ÿä â$x+   ÿŽ â r&   ÿäÿ â %x%   Ž È %x   
Ž È %x%   
Ž È r/   ÿàù ä r/-   ÿàõ ä r'   ÿä ä r/   ÿà÷ ä r/.   ÿàù â �y     ™ ¢ r'0   ÿä Ä r'   ÿäÿ ä r'&   ÿä â �y[    ‘ è  �y1    ™ ( �y     ‘ h �y\!   ‘ h �y]!   ‘ b r%"   *àµ Ä xG $       ä r3),   ÿd~ ä x?        â %v Z  G Ž È $x33$   ÿŽ Ê r3   ÿàÿ Ê �yU   ™ h �yT   ™ h �yQ   ™ h �yS   ™ h �y;   ™ h �y3   ™ h �y6   ™ h �y7    ™ b rB/   ÿà÷ È rC'   ÿäÿ Ê �y8B0   • h �y    • b x ÿ   ÿÀŽ äO xÿ      ä x5[   ÿ   ä�xZ[ð   ÿÀŽ Ä x1ÿ   ÿÀŽ äx    ÿ   äx5   âŽ ä x   âŽ ä xY ð   ÿÀŽ ä xZÿ   Z  Ä xÿ   ÿÀŽ â/ %v  ? Ž âxÿ   1  ä xYÿ   Y  â %v  ? Ž â xZ   ÿ   â �yV    ™ ¢ xÿ   ÿÀŽ ä xY   ÿ   â �yW    ™ " x   âŽ Ä x   âŽ Æ %v  ? Ž âxÿ      ä/ x_\ð   ÿÀŽ ä �yX    ™ b %v  ? Ž â xÿ   1  ä x^]ð   ÿÀŽ ä x\   ÿ   â� �yR    ™ b xÿ   ÿÀŽ Ä x_ÿ   _  ä x]   ÿ   ä xÿ   ÿÀŽ ä x^ÿ   ^  ä x   âŽ ä x_   ÿ   Ä/ x   âŽ ä x^   ÿ   â� %v  ? Ž âx   ÿÀŽ ä x   ÿÀŽ â %v  ? Ž â r    ÿ@� â �yA    ™ â r1   ÿ@� Æ �y=    ™ â  %v  ? Ž È %v>  ? Ž â r/   ÿàõ â �y     ™ æ  r	'   ÿä â �y5>    ™ â  r
/   ÿàõ ä/r/   ÿà÷ â �y:2   • â  r'   ÿä Ä r'   ÿäÿ â �y94   • â  r%"   (àµ Æ �y10   • è  �y>2   • è �y0   • è  �y<6   • â  rB),   ÿd~ Æ �y@0   • â %v Z  G Ž Æ �y?4   • â $xCB$   ÿŽ Æ �yB6   • ä rC   ÿàÿ Ê �yC   ™ è �y   ™ è �y   ™ è  �y   ™ è  �yF   ™ è  �yD   ™ è  �y   ™ â  r%"   àµ Æ/ �yE    ™ â  rH),   ÿd~ â %v Z  G Ž ä �yL    • ä  $xGH$   ÿŽ Ê rG   ÿàÿ Ê �yG   ™ è �yO   ™ è �yN   ™ è �yM    ™ è �yH   ™ è �yI   ™ è �yK   ™ è �yJ   ™ è �yP    • â x\\   ÿÀŽ Ä xZZ   ÿÀŽ ä xa    =    ä x[[   ÿÀŽ ä Er\ \       ä ErZ Z       ä x    ÿÀŽ Ä x]]   ÿÀŽ ä x^^   ÿÀŽ â #t\  `¿a  â/ Er  [       â #t[Z  `¿a  â Er        ä Er ]       ä x__   ÿÀŽ Ä xYY   ÿÀŽ ä Er\ ^       â #t  `¿a  â Er] _       â #t  `¿a  â ErY Y       â #t    `¿a  â#t\\  `¿a  â#t]]  `¿a  â&zbQ  € ÿ   â #tYY  `¿a  Æ &zbS  € b   Ê Erf b       ä r8ÿ+   8  ä xZVÿÀŽ äO x`ÿ   V  ä x^ÿ   W  äxVWÿÀŽ â &zWU  € ÿ   â x``ÿÀŽ â &rZU   ÿ   âx^^ÿÀŽ â &r_VU   ÿ   Ä &zUT  € W   ä &rW`T      ä&ra^T   _   â xTXÿÀŽ äx_ÿ   X  ä Er U       â &rdTQ   W   â xU_ÿÀŽ Ä xXRÿÀŽ â #rb    ÿ   â xRÿ   R  â &rdUS   d   â #r   ÿ   â &rhXQ   a   â #rb[f   b   â xQRÿÀŽ â #rffY      â ErW d       â &z;  € ÿ   Ä &zR3  € ÿ   ä &rhQS   h   â !rSbW       â &z_6  €    ä &za7  € R   â xÿ   A  ä� xbAÿÀŽ ä xAÿ   =  Ä xR=ÿÀŽ â &rWb;   ÿ   âx=ÿÀŽ ä Ere h       â &rcR;   ÿ   â xAAÿÀŽ â &rW=6   W   âEr _       â !r;fe       â xfÿ      â &rcA6   c   â x6 ÿÀŽ Ä x 5ÿÀŽ â #rj   ÿ   â Era a       ä x:ÿ   :  â #rh   ÿ   â xffÿÀŽ â &rW63   W   âxdÿ   5  â &rc 3   c   â #ra\   j   â x::ð   ÿÀŽ Ä x38   ÿ   ä xj        ä x5        ä x8ÿ   8  â &rWf7   W   â x9ÿ   9  ä r:1:   ÿ � Ä x33   jìŽ ä x18   5ìŽ ä x>ÿ   >  ä r8ÿ+     ä x<< ð  ÿÀŽ ä x99   ÿÀŽ â #rh]a   h   â ErW W       Ä xddÿÀŽ ä x5>ð   ÿÀŽ ä Er 3       ä x38   ÿ   ä r9<:   9àÿ â !rhhW       â Er1 1       â &r7d7   c   â r@@5   ÿ � Ä x:3   jìŽ â 0r3ÿ    A   â x5        ä x8ÿ   8  â 0rÿ9    A   â  rhh1     A  ä x58   5ìŽ ä Er8 7       â #rSS   h   â  r73     A  â Er5 5       Ä !r<8       â #r7S      â &r7ZC   ÿ   âEr8 :       â &r9VC   ÿ   â  r:<5     A  â &z<C  € ÿ   â x?ÿ   ?  â &r7`   7   ä&r9^   9   â #r:;8   :   â &z<  € <   â xBB ð  ÿÀŽ â &z  € ÿ   â x??   ÿÀŽ â &r7T   7   Ä &r;X   9   ä &z9  € ÿ   ä &rb   ÿ   ä&r>R   ÿ   â rB@   ?àÿ â &zF  €    â Er< <       â &z9D  € 9   Ä &r=D      ä&z@  € ÿ   ä &rDAD   >   â Er>        â &z@E  € @   â #r <   ÿ   â&r6      â Er9 9       â &rD    D   â #r<<   ÿ   â &rfE      â #r[>      â Er @       â &r7UF   7   â #r9   ÿ   â &rDdE   D   â #rC9   ÿ   â #r?Y>   <   â &r;QF   ;   â Er>        â #r9]      â Er< 7       â #rC\   C   â ErD D       â 0rÿ    A   â !r>9>       â Er@ ;       â !r<       â !rDCD       â  r<3     A  â  r1>     A  â !r??@       â  r5D     A  â #r<:      â &rbbG   ÿ   â #r      â &z:G  € ÿ   â #r8?      â 0rÿL    A   Ä &r==O   b   ä &z<N  € ÿ   ä &z:O  € :   â  r     A  â &r=6N   =   ä &z<M  € <   â Er3 :       â #r      â &r=fM   =   Ä &rRRG   ÿ   â Er <       â #r3   ÿ   â Er= =       â &rAAO   R   ä #r]      â #r33   ÿ   â &rA N   A   ä !r =       â &rZZH   ÿ   Ä &zH  € ÿ   â #r3\   3   â &rAdM   A   ä &rHVH   ÿ   ä &rZ`I   Z   ä &zK  € ÿ   ä &zI  €    â Er A       â &rH^I   H   Ä &rZTK   Z   âEr        â &zJ  €    ä &rHXK   H   â !r63       â &rZUJ   Z   â Er        â #r     ÿ   â x22   ÿàÿ â &rHQJ   H   â #r   ÿ   â Er Z       â #r [       â| 2   pbô â #rY      â Er H       â !r        â  r     A  â 0rÿP    A   â  r11      A  â  r556     A  â !r       â #r4   4   â #r   1   â  r     A  â #r8   5   â  r     A  â x//€  ÿà÷ â #r$   $   â x€   ÿàÿ â #r      â x((€   ÿàÿ Ä x**€   ÿàÿ ä r'ÿ'   ÿäÿ â D)       À â Gy  `èÿÿÿÿƒ ê Ay        € ê� ˆ ! ýÿ H   è ˆ !€ýÿ H   è ˆ ! þÿ H   è ˆ !4€þÿ H   è ˆ !$ ÿÿ H   è ˆ !€ÿÿ H   è {          ì M        € ê „y     H   " z # r  p`ð Æ „y     H   b r ÿ   p"p Æ „y €   H   ¨ „y    H   è „y    H   ( „y €  H   h „y	 €  H   h „y    H   h „y    H   h „y €  H   h „y €  H   b !r       Æ „y    H   " !r         Æ/ „y
 €  H   b !r       ÆO „y    H   ¢ !r       Æ� „y €  H   â !r         Æ„y    H   â !r4       Æ„y €  H   â !r	       Æ „y! €  H   â !r$       â !r       ä ‰       " !r       â !r       â !r       È ‰      " !r

       â/ !r       âO !r       È� ‰	      b !r        âz  r      â !r       â ‰      ¢ !r

       Æ ‰      â !r

!       Æ ‰       " !r       Æ ‰
      ( ‰      b !r		       â/ !r       ÈO ‰	      b !r       Æ� ‰      ¢ !r         Æ‰      â !r
       Æ ‰  €    " !r       Æ‰      ( ‰ €    b !r	       â/ !r       ÈO ‰	 €    b !r
       Æ� ‰ €    ¢ !r        Æ ‰
 €    " !r       Æ‰  @    â !r       Æ‰ €    ( ‰ @    b !r		       â/ !r       ÈO ‰	 @    b $z x  #Ž â !r
       ä y      '   " x
#   ÿàÿ â !r         ä� ‰ @    ¢ z 
 r  p`ò â !r       Ä‰       â x    p"ò â !r       ä‰ @    ( ‰ @    h ‰
      b $z ~  Ž â !r	       â/ 	   ÿàÿ â !r       ÄO 
 r  ÿàõ ä r   ÿ† â !r        â� ‰      " ÿÿ   ÿä â !r       â
 j  ÿ„ â ‰      b x        â !r       â
 k   â ‰      ¢ 		   ÿàõ â%v j   Ž â r ÿÿ   ÿäÿ â ‰      â ÿÿ   ÿä â !r
       â 	 j  ÿ„ â †	     â	 z j  ÿ† Ä 		 k   â †    â	 z k   � â !r       â !r       â/ !r       ÈO †	     â	 !r       Æ� †     è	 †	     â	 M™        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      "   " ¹z  l      â Ey  P    € â ™x?    â y      !   b ¹z  F   
   â x       ÿ  â r' ÿ       â y#      %   ¢ $r3ÿÿ   ÿ Ž Æ y	      &   è y      '   " r ÿ   pRð är ÿ   pRò â$x%    Ž ä/ $x##   ÿ Ž ÆO x ÿ   %  È r  %   ÿ� ä "   ÿ8Ž ä x2ÿ      È | 2   pbô Ú G)      € ê r ÿ       â�$rÿÿ   ÿ Ž â x  øÿÿÿÿÀŽ â $z$ }  ÿŽ âx*2   ÿ   â 'z y   Ž â r%%   €ÿàÿ ä z)  Y      â 'z	 s   Ž â zÿ z    Ä r %%   ÿ� â $z# p  2Ž â zÿ t    â $z |  2Ž âxÿ      â $z |  Ž â x#   ÿàÿ â $z v  Ž â x+ÿ   %  â $z	 w  ÿŽ â z* q  ÿàÿ â $z p  Ž â x þÿÿÿÿÀŽ â $z
 v  Ž â r$$   ÿàõ â $x%   ÿ Ž Ä %x8   ÿŽ â r(ÿÿ   ÿä ä xÿ     â %x

8   ÿŽ â r.   ÿà÷ Æ $x%   
Ž â r-
   ÿàõ â %x   
Ž Æ r0   ÿä â $v/ÿ X  ÿ Ž ä $r ÿÿ   ÿ Ž ä $x,   ÿŽ ä %x%   
Ž È %x   Ž È %x%   Ž È $x&   Ž ä r/   ÿàù ä r/-   ÿàõ ä r)   ÿä ä r/   ÿà÷ ä r/.   ÿàù â �y    ™ ¢ r)0   ÿä â $x   )Ž Ä $x)   & ä �yE    ‘ è �y    ™ ( �yH    ‘ h �yF!   ‘ h �y !   ‘ b r%$   *àµ Ä xQ $       ä r1+(   ÿd~ Æ %v Z  Q Ž È $x11$   ÿŽ Ê r1   ÿàÿ Ê �y>   ™ h �y;   ™ h �y8   ™ h �y9   ™ h �yP    • b xÿ   ÿÀŽ âO $x5E   ÿ Ž â�xDEð   ÿÀŽ Ä xÿ   ÿÀŽ äxH   ÿ   äx5   âŽ â $tÿ   ÿ Ž â x   âŽ â/ �y5    ™ ¢ xDÿ   D  â %v   Ž â xÿ     Ä xAHð   ÿÀŽ â %v   Ž â xÿ     â �y:    ™ â  xÿ   ÿÀŽ â $x1D   ÿ Ž â xAÿ   A  â �y=    ™ b	 xÿ   ÿÀŽ ä x4A   ÿ   Ä x1   âŽ ä x4   âŽ â $xF   ÿ Ž âxBFð   ÿÀŽ â %v   Ž âxÿ     â�y1   ™ $ %v6   Ž â xÿ     â �y<    ™ ¢  x@ ð   ÿÀŽ Ä xÿ   ÿÀŽ â �y66    ™ " x    ÿ   ä xÿ   ÿÀŽ â �y4   ™ " xBÿ   B  ä x   âŽ ä x@ÿ   @  â �y7   ™ " x   âŽ â $xB   ÿ Ž Ä %v   Ž â x@   ÿ   Æ %v   Ž â x   ÿÀŽ â �y?    ™ " x   ÿÀŽ ä r   ÿ@� â �yC    ™ " r   ÿ@� Æ %v   Ž È %v   Ž â �yG    ™ " r/   ÿà÷ Æ �yJ    ™ " r)   ÿäÿ ä r
/   ÿàõ È �y0   • " $x   ) â r/   ÿàõ È �yI2   • " $x   ) â r/   ÿàõ ä/ r%$   àÇ â �yL0   • " r)   ÿä ä rR+(   ÿ„þ â �yN2   • " %v Z  Q Ž Æ �yK4   • ( �yM6   • ( �y0   • ( �y0   • " $xR$   ÿŽ Æ �yO4   • " $x   Ž Æ �y6   • ( �y   ™ ( �yU   ™ ( �yV   ™ ( �yT    ™ ( �y   ™ (# �yS   ™ ( �yR   ™ ( �yQ   ™ ( �yW    • " xDD   ÿÀŽ Ä x    =    ä xAA   ÿÀŽ ä x@   ÿÀŽ ä/ Er@ D       ä ErD        ä xEE   ÿÀŽ Ä xFF   ÿÀŽ ä xHH   ÿÀŽ ä x    ÿÀŽ ä xBB   ÿÀŽ ä Er A       â #tAD  `¿  â Er  E       Ä Er F       ä ErH H       ä Er        ä ErB B       â #t    `¿  â#t  `¿  â#tH  `¿  â#t  `¿  â#t@@  `¿  â#tBB  `¿  â#t  `¿  â &zF>  € ÿ   È &zE;  € F   ä &zH8  € ÿ   È &zH9  € H   Ê ErY H       â 0rPÿP    A   â x:ÿÀŽ ä� xDÿ   :  ä x:ÿ   =  äx==ÿÀŽ â &r>   ÿ   â xDDÿÀŽ Ä x::ÿÀŽ â &r>=>   ÿ   ä &rFD;      ä&r>:;   >   â x;<ÿÀŽ äO x<ÿ   <  ä Er E       â &rF;8   F   â x<<ÿÀŽ Ä xE6ÿÀŽ â#r[    ÿ   â x6ÿ   6  â &rF<9   F   â #rX   ÿ   â &r>E8   >   â x66ÿÀŽ â #r8@Y   [   â Er F       â &zF7  € ÿ   â #rXY   X   â &r969   >   Ä !r88       â &z1  € F   ä &zH4  € ÿ   â xYÿ   ?  ä x>?ÿÀŽ ä xZÿ   C  ä xFCÿÀŽ â &zH5  € H   â ErC 9       Ä x?YÿÀŽ â &r\F7   ÿ   âx9ZÿÀŽ â &rZ>7   ÿ   â Er        â !r7XC       â ErX H       â &rC?1   Z   âxHGÿÀŽ â &rY91   \   â x1ÿ   G  Ä xZÿ   J  ä xJJÿÀŽ â #r[   ÿ   â #rG   ÿ   â x11ÿÀŽ ä xZÿÀŽ â &rZH4   C   ä&r4J4   Y   Ä &rZ15   Z   ä&r55   4   â r4ÿ,     â #r[BX   [   â x        â #rXXA   G   â x\ÿ   4  â $tCÿ   ÿ Ž â xIÿ   I  ä ErG Z       â $xZ4   ÿ Ž â x\   ìŽ Ä xIIð   ÿÀŽ â !r4[G       â xCZ   CìŽ ä Er        ä rLLI   ÿ � ä ErI 5       ä Er5 C       â  rC4     A  â xNÿ   N  Ä xKÿ   K  â #r885   C   â xGNð   ÿÀŽ ä rCÿ,     â $tNÿ   ÿ Ž â xMM ð  ÿÀŽ ä xKK   ÿÀŽ ä r4G   ÿ � Ä xGÿ   C  â !rXXI       â rKML   Kàÿ ä x        ä xCC   ÿ   â &zI  € ÿ   â xGG   NìŽ â 0rÿK    A   â xC   ìŽ â &zIV  € I   â ErC G       â &zGU  € ÿ   Ä &r>>   ÿ   â  rLP     A  â &zGT  € G   â ErI I       â &rF   ÿ   ä &r>?V   >   â #r3L8   3   â &r99V      â Er8 G       â #rI   ÿ   â &r>HU   >   â #rI   ÿ   â &r   ÿ   Ä &z  € ÿ   â #rB8      â &r=   ÿ   â #r8A8      â &r>1T   >   ä &rDS      ä &zR  € ÿ   ä &zS  €    Ä &r:S      â Er> >       â &r9JU   9   â Er        â &r;R      ä&zQ  €    ä &rER      â xOÿ   O  â &r9T   9   â !r>>       â &r<Q      âEr        â &r6Q      â #r    ÿ   âx ð  ÿÀŽ â #r   ÿ   â xOO   ÿÀŽ Ä x22   ÿàÿ â #r@      â Er9 9       ä Er         â #r      â Er        ä r4   Oàÿ ä | 2   pbô â !r889       â !r        â !r       â 0rÿ    A   â Er        â  r4XC     A  â  r>>     A  â 0r ÿW    A   â  rCC8     A  â  rPP     A  â #r>5   >   â #r47   4   â #r   C   â  r      A  â r      A  â x//€  ÿà÷ â #r'P4   '   â #r!>   !   â #r        â x€   ÿàÿ Ä x**€   ÿàÿ ä r)ÿ)   ÿäÿ â D)       À â Gy  �ìÿÿÿÿƒ ê Ay        € ê�ˆ "3 þÿ H   è ˆ "'€þÿ H   è ˆ "! ÿÿ H   è ˆ " €ÿÿ H   è {          ì M        € ê „y      H   " z # r  p`ð â $z		 x  #Ž â x##   ÿàÿ â „y €   H   b r ÿ   p"p â$z	 ~  	Ž â z # r  p`ò â „y    H   ¦ x    p"ò â „y    H   è „y €  H   ( „y    H   h „y    H   h „y €  H   h „y    H   h „y €  H   h „y €  H   b !r  3       Æ „y
 €  H   " !r'       â/ !r!       âO !r         ä� !r       â!r       â!r         ä !r       Æ ‰       b !r       Æ ‰      ¢ !r        Æ ‰      â !r       È !r
       â !r        È/ ‰       " !r       ÆO ‰      b !r       Æ� ‰      ¨ ‰      â !r        â !r       È/ ‰      " !r       ÆO ‰ €    b !r
       Æ� ‰ €    ¨ ‰
 €    â !r       â !r       â/ $tÿ   ÿ Ž â !r       ÆO ‰  @    " !r
       Æ� ‰ @    h ‰ @    ¨ ‰ €    â !r         â 
	 r  ÿàõ â !r       â/ !r
       ÄO ÿÿ   ÿä â ‰       " 
 j  ÿ„ â!r       ä� ‰      b 
 k   Æ ‰ @    ¨ ‰
      â !r        â %v	 j   Ž â !r       È/ †	     â !r       ÆO †    â !r
       Æ� ‰       h  †	     â M™        € ê ¹z  r      â !r        â? �x   ?àÿ Ì |		   ÿàñ È rÿÿ   ÿä  ä z	 j  ÿ€ È z	 k    Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž ä $rÿÿ    Ž È % \   Ž Ê �	    ™ b ¹z  l      â Ey  0    € â ™x?    â y      "   b $rÿÿ   ÿ Ž Æ y      !   ¢  ÿ       â $ÿÿ    Ž È '
  m   Ž â ‚ ÿ       È 
 ÿ n     Ê $ ÿÿ    
Ž â r ÿ   pRò Æ/ $
  o  Ž â r ÿ   pRô â$x    Ž äO $‚ÿÿ    Ž Æ x	ÿ     â 'Š  s   Ž ä y      '   " r		   ÿ� È xÿ   	  È |    pbö ä Šÿ t     Ö G9  °
    € ê y      %   " $rÿÿ   ÿ Ž â x		øÿÿÿÿÀŽ â $z }  ÿŽ âx   ÿ   â 'z  y   Ž â r	  €ÿàÿ Æ $tÿ8   ÿ Ž â zÿ z     â $x   ÿ Ž â xÿ     â $rÿÿ   ÿ Ž ä $z  |  Ž È $z  v   Ž âxÿ     Æ $z	 p   Ž ä $z  w  ÿŽ ä %x	8   Ž Æ r    ÿàù â %v	 X  Ž â z X  ÿàñ ä r    ÿ� â $rÿÿ    Ž â x!   ÿà÷ â $rÿÿ   ÿ â r	        â $tÿ   ÿ Ž â zÿ Y  þ Ä x þÿÿÿÿÀŽ ä xÿ      Æ $x   
Ž ä %v  € Ž È $x   ÿŽ ä $rÿÿ   	 Ž ä $r	ÿÿ    Ž ä %x&   Ž âr
   à± â �y,2   • ¨ �yÿÿÿ ‘ è �y&&    ™ ( �y    ‘ b r   ÿd~  â $tÿ$   ÿ Ž Ä $t%ÿ   ÿ Ž â �y*4   • ¢ %v

 Z   Ž Æ �y+0   • ¢ x   ÿ   ä�x#ð   ÿÀŽ ä x &ÿ   ÿÀŽ äx#ÿ   #  ä x     âŽ â $x$   ÿŽ È %v   % Ž â x ÿ   &  äx#   ÿ   â $x   Ž â �y"    ™ â  x ÿ   ÿÀŽ ä xÿ   &  â �y 
   ™ " x$ð   ÿÀŽ â$x   ÿ Ž â x   âŽ â �y!
   ™ b xÿ   ÿÀŽ Ä x$ÿ   $  â %v  % Ž âx   âŽ â �y)
   ™ d $x$   ÿ Ž ä %v  % Ž â �y    ™ d x   ÿÀŽ Ä �y'
   ™ d r&   ÿ@� ä �y    ™ f %v  % Ž â �y&
   ™ b r   ÿàñ Æ �y 
   ™ b r   ÿä  Æ �y%
    ™ h �y    ™ h �y(
   ™ h �y    • h �y6   • h# �y-
    • d x
   ÿÀŽ Ä x##   ÿÀŽ ä x$$   ÿÀŽ È Er$ $       ä x,ÿ   ,  äO x*ÿ   *  ä x,,ð   ÿÀŽ ä x   ÿàÿ ä r++,   ÿ � Ä x**   ÿÀŽ ä |    pbð ä x€  ÿà÷ ä x€  ÿàù ä/ x€   ÿàÿ ä x"ÿÀŽ Ä� x"ÿ   "  â &z  € ÿ   Æx""ÿÀŽ â &r    ÿ   â Er  
       â &z! €    ä&r!"!      â x   ÿÀŽ â $tÿ   =ÿ Ž â Er        ä Er        â #t    `¿  â Er" #       Æ #r
    ÿ   â #t   `¿  â &z#) € ÿ   â xÿÀŽ ä xÿ     â &z#& € #   Æ xÿÀŽ â &r)   ÿ   ä &z' € ÿ   â xÿÀŽ â &z  € ÿ   â Er# #       â &r&&      ä &z% €    â xÿÀŽ Ä xÿ     â &r!    !   ä &z( €    â #t"  `¿  âEr        â #t$  `¿  â xÿ     â #r  #   ÿ   â rÿ     â &r'   &   â xÿÀŽ Ä Er        â #r       â xÿÀŽ â $x    ÿ Ž â x        â &r%      â xÿ     â #r
   
   â $tÿ   ÿ Ž â x ð  ÿÀŽ â &r!(   !   â x   ìŽ Ä Er        ä x     ìŽ ä r+   *àÿ â !r       â Er! !       ä Er        â 0rÿ    A   â Er         â 0r ÿ-    A   â !r

!       â  r     A  Æ  r       A  â #r

      â $rÿÿ   	Ž â r	ÿ   ÿä ä #r 
      â G‰  `÷ÿÿÿÿƒ ê Ay        € ê ˆ €ÿÿ H   è {          ì M)        € ê „y      H   " r ÿ   pRð Æ „y €   H   h„y    H   ¢ !r         È !r         È/ !r         ÊO ‰       $ !r        Ê ‰      $ !r       Ê ‰ €    $ !r       ä y      %   ( ‰ @    b $z x  Ž â !r       Ê/ ‰	      b  M	        € ê $tÿ   ÿ Ž â !r		       â/ $z ~  Ž È %v j   Ž Ê †y 	    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž È % \   Ž Ê �	    ™ ¢ $rÿÿ   ÿ Ž â Ey  À    € â x*      ÿ  â y$      "   " 'Š s   Ž Æ y      !   " '
  m   Ž Ê 
 ÿ n     ä rÿ$   ÿü‚ â $x$$    Ž Ê x	ÿ   $  È r		$   ÿ� ä Šÿ t    ä y      %   $ r       âO G  Ð    € ê z ÿ `  pPô â x*      ÿ  â z ÿ ^  pPò ä z ÿ b  p ö Ä z ÿ a   Sô ä z ÿ _  Sò ä z ÿ c  0't Ö       '   b $ x  ÿŽ ä$ª x  ÿŽ â ©      '   ¢ $ ~  ÿŽ ä/ $ª
 ~  ÿŽ ÆO    àë äÿ     ä¢
   àÇ ä¨
ÿ     ä ÿÿ   Äþ ä    ÿàý Ä ¢   ÿàû ä ¢
ÿÿ   
„þ â $ÿÿ    â  ^  ÿ† ä ª b  ÿˆ â $¢
ÿÿ   
Ž â  _  � È ª c  
 â �*     h �©+     d Ay        € ê ¹z  l      â x&ÿ   	  â ™x?    â $ÿÿ    
Ž â Ey       € â x(      ÿ  Ä $
  o  Ž â| &   pbò â $‚ ÿÿ    Ž Ø G  @    € ê y      '   b $rÿÿ   ÿ Ž â x		øÿÿÿÿÀŽ â $z   w  ÿŽ â z ÿ `  pPð â x(      ÿ  ä $x$$   	
Ž â z ÿ a   Sð Æ $x $   ÿ Ž âr$$   ÿ� ä x!ÿ   $  ä xÿ      ä z  X  ÿàõ È z Y  ÿä â 'z y   Ž È/ $z' }  ÿŽ â zÿ z    ä xþÿÿÿÿÀŽ ä r' '   ÿàó â $z |  &Ž â xÿ     â $tÿ   ÿ Ž ä $z v  Ž Ä $x$   
Ž ä $z p  Ž ä $tÿ   ÿ Ž ä %x"8   ÿŽ È $r%ÿÿ   " Ž ä $x"&   ÿ Ž ä %v$ X  Ž È $rÿÿ   ÿŽ  ä %v X  Ž È $x   ÿŽ ä r%   ÿàó ä r%   ÿàõ Æ $x   #Ž  ä $x#    Æ �y    ™ ¨ �y     ‘ è �y!   ‘ " $tÿ$   ÿ Ž ä $tÿ   ÿ Ž â xÿ     ÄO x	ÿ   ÿÀŽ â $x    ÿ Ž â�x ð   ÿÀŽ ä x
ÿ   ÿÀŽ â $x   ÿ Ž âx   	âŽ ä r	$'   "à£ ä xÿ     Ä x   
âŽ â %v	 Z   Ž â x
ÿ     ä xð   ÿÀŽ â $x   ÿ Ž â r!   ÿDþ  â %v   Ž âx

ÿ   ÿÀŽ ä xÿ     â $x$   ÿŽ â x
   
âŽ â �y-    ™ ¢  %v   Ž È $x   Ž â �y3    ™ â %v

   Ž Æ �y   ™ " $x	   ÿ Ž Æ �y/
    ™ $ x	   ÿÀŽ ä/ �y   ™ $ r   ÿ@� ä �y   ™ " z% X  ÿàó ä %v   Ž â �y   ™ ( �y   ™ " z	# Y  ÿäÿ  Æ �y   ™ ( �y    ™ ( �y,2   • " r
%   ÿàó Æ �y   ™ $ $x   #Ž  ä �y    ™ ( �y.4   • ( �y
0   • ( �y10   • ( �y06   • ( �y2    • " x     ÿÀŽ Ä x   ÿÀŽ ä Er          ä Er        ä x   ÿÀŽ ä x   ÿÀŽ ä Er        Ä x&&   ÿàÿ ä x4-ÿÀŽ äO x-ÿ   -  ä x63ÿÀŽ ä� x--ÿÀŽ ä x3ÿ   3  â &r44   ÿ   âx/ÿÀŽ Ä/ x3ÿÀŽ â &z € ÿ   ä &r--   4   ä &r66   ÿ   ä &r	   -   ä $t-ÿ   =ÿ Ž ä &r   6   Ä &z € ÿ   ä &z €    â #t   `¿-  â&z €    â x
ÿÀŽ ä Er         ä x3ÿ   ,  ä xÿ     â #t  `¿-  â Er,        â #r    ÿ   â x/ÿ   /  â &r

      â xÿÀŽ â &z € ÿ   â x/ÿÀŽ â #r/,      â &z € ÿ   Ä &z €    ä &r
   
   â xÿ   .  â &z. €    â Er        â #t  `¿-  âEr        ä x43ð   ÿÀŽ ä rÿ     â &r	   	   â Er. .       â #t  `¿-  â #r   ÿ   â r114   ÿ � â $tÿ   ÿ Ž â x00 ð  ÿÀŽ Ä x   ÿÀŽ ä xÿ     â #r.      â Er
 
       â $tÿ   ÿ Ž â r01   àÿ â $x   ÿ Ž â x   ìŽ â !r0
       â Er
 	       â 0r2ÿ2    A   â Er	        â 0rÿ    A   â x   ìŽ â !r//
       Ä  r00	     A  â Er1        â  r2     A  â G‰  p    € ê %v
$ `  Ž â z% `   àµ È r

%   ÿàó ä z	# a  d~ Æ $x   #Ž  ä �y3     ¨ �y
     è  �y5!    " $t
ÿ   ÿ Ž Ä $x3   ÿ Ž âO x6ÿ   3  ä xÿ   ÿÀŽ ä� xÿ     äx   âŽ â $x45   ÿ Ž âxÿ   ÿÀŽ ä xÿ     â %v   Ž â x4   âŽ Ä x4ÿ   5  â $x6   ÿ Ž â xÿ   ÿÀŽ â �y7    ™ ¢  %v   Ž ä x   âŽ ä $x	4   ÿ Ž ä %v
 `  
Ž â �y    ™ ä x		   ÿÀŽ È r	   ÿ@� ä r

%   ÿàó Æ %v   Ž È %v   Ž ä �y    ™ $ $x   #Ž  â z% `  ÿàó â �y    ™ & z# a  ÿäÿ  â �y
0    ( �y83    ( �y4    (# �y90    ( �y:6    " x33   ÿÀŽ Ä x66   ÿÀŽ ä Er< 3       ä Er6 6       ä x55   ÿÀŽ â #t	<  `¿-  ä#t66  `¿-  â Er
 5       â #r	 	   ÿ   È #r,,6   	   â #t	
  `¿-  â x44   ÿÀŽ Æ #r		   ÿ   â Er4 4       Ê #t44  `¿-  â $tÿ   ÿ Ž Æ/ #r..4   	   â x 7ÿÀŽ äO x7ÿ   7  ä x
ÿÀŽ ä� xÿ     â &r     ÿ   â x77ÿÀŽ â &r

   ÿ   â xÿÀŽ Æ &r77       ä &r
   
   â xÿÀŽ äxÿ     ä x ÿÀŽ â &r   
   â xÿ     ä rÿ     Ä xÿÀŽ â &r     7   â x	ÿÀŽ ä x88ð   ÿÀŽ â &r      â xÿ     ä xÿ     â &r 	       â r898   ÿ � â $t
ÿ   ÿ Ž â x   ÿÀŽ â $x   ÿ Ž â x:: ð  ÿÀŽ Ä x   ìŽ ä Er        ä r:8   àÿ ä Er	         â !r.       â x   
ìŽ ä Er        â 0rÿ    A   â Er         â !r	,	       Ä  r     A  â  r22     A  Æ #r 	       È #r)2    )   â | &   pbò â #r//1   0   â x%%€  ÿàõ ä x""€   ÿàÿ â #r(/   (   ä $r#ÿÿ   # Ì G™  Ððÿÿÿÿƒ ê Ay        € ê y       "   b Ey        € â r  ÿ   pRð ä/r  ÿ   pRò Ö G‰  `     € ê y      !   b z ÿ `  pPð È z ÿ a   Sð â $x     Ž Ê/ ˆs (€ÿÿ H   î ˆ )   H   ä Ay        € ê {          ì M        € ê y      !   " z ÿ `  pPð È z ÿ a   Sð â „y     H   > „‰  €   H   h „‰    H   ¢ !r(       È !‚         È/ !‚        âO G‰        € ê „y  €  H   ( „y €   H   h „y    H   ¨ „y	    H   è „y €  H   " !r  )       â !r       Æ/ !r         âO !r	       Æ� !r)        ä‰      " z ÿ `  pPð ä r ÿ   pRò ä z ÿ a   Sð â !r       Ê ‰       $ !r         Ê ‰  €    $ !r        Ê ‰ @    $ !r       Ê ‰      $ !r       â G‰        € ê ‰ )      $ !r  )       Ê ‰       $ !r        Ê ‰ €    $ !r       Ê ‰ @    $ !r       Ê ‰      $ !r)       ä M        € ê !r*       âG‰  �    € ê $t ÿ   ÿ Ž â !r++)       È z   h  pð Ú G	  �    € ê ¹z  h      ä �xÿÿÿÿ?àÿ È Œx    pð È ‡x       È ™x   ?  Ø »z  €    ä ™x?    Ô Xy  æÿÿÿÿƒ ê  x +;ª¸¿  A  Ì s          $ !t   €?    Ì s        $  r+     A  È  r     A  â Gy  `    € ê  x +'7=  A  â x+*BL?  A  Æ #t +  €?    È  r      A  â  x +   ?  A  Ê s     $   $ !t  €?    È  r       A  È  r      A  â Gy  À    € ê $v ÿ h  ÿ Ž Ê x     pRð Ú G‰  à     € ê x     pRð Ú G	        € ê 	z + i    � ä 	z i    � Æ  x ;ª¸¿  A  â 	z i €  � Ê s        $ !t  €?    Ì s        $  r       A  È  r      A  â Gy  Ð     € ê  r+     A  â Gy  °     € ê 	x +  à@  � ä 	x  à@  � Æ  x #ÛÙ?  A  â 	x  àÀ  � Æ  x;ª¸¿  A  ä !t  €?    È s        $ !t  €?    Ì s        $  r       A  È  r      A  ä y       %   ( y      &   ( y      '   b $z  x   Ž ä $tÿ   ÿ Ž Ä $z  ~   Ž È/ %v  j   Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y       &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž ä $rÿÿ     Ž È % \   Ž Ê �	    ™ b ¹z  l      â Ey  Ð    € â ™x?    â y      "   b $rÿÿ   ÿ Ž Æ y      !   ¢ $ÿÿ   ÿ Ž ä $ÿÿ    Ž â y	      %   & '
 m   Ž È $‚ÿÿ   ÿ Ž â 
ÿ n    Ê $
ÿÿ   
Ž â r ÿ   pRò Æ/ $

 o  Ž â r ÿ   pRô â$x    Ž äO $‚ÿÿ    Ž Æ xÿ     â 'Š s   Ž Æ r   ÿ� â $rÿÿ   ÿ Ž ä $€   Ž â xÿ     â $x		   ÿ Ž Æ |    pbö ä Šÿ t    ä x      ÿ  Ò G9      € ê y      '   " $r
ÿÿ   ÿ Ž ä $vÿ X  ÿ Ž ä 'z y  
 Ž â x
øÿÿÿÿÀŽ Æ $z w  ÿŽ â zÿ z    â $z	 p  Ž ä $x
   

Ž ä $z |  Ž ä $z }  ÿŽ â r

   ÿ� â $z v  Ž âxÿ   
  â $x
   ÿ Ž â xþÿÿÿÿÀŽ â %x(8   ÿŽ â r   ÿà÷ Ä x+ÿ     â $x
   
Ž â r(   ÿàñ â$vÿ Y  ÿ Ž ä $rÿÿ   ÿ Ž â r)   ÿ  â $x   ÿ Ž ä $rÿÿ   ÿŽ ä $x   ÿŽ Ä %x*+   (Ž È %x,
   (Ž È r   ÿà÷ ä r,   ÿàñ Æ $x   Ž ä $x-     Æ �y     ‘ ¨  �y    ™ è �y&!   ‘ " $v]ÿ p  ÿ Ž È %x.]8   Ž È %x]8   Ž â�y!.    ‘ h �y    ™ h �yX.!   ‘ b %x0]8   .Ž È %x"]8   Ž â�y0    ‘ h �y"    ™ h# �yI0!   ‘ b %x2]8   0Ž È %x$]8   "Ž â �y2    ‘ h �y9$    ™ h �y2!   ‘ b r
   à± â $tÿ$   ÿ Ž Æ r   ÿd~  â %v0 Z   Ž È $x$   ÿŽ ä $t8ÿ   ÿ Ž ä $x11   Ž Ê �y0   ™ h �y>0   ™ h �y@0   ™ h �y:0   ™ h �y'0   ™ h �y%0   ™ b# r2*   ÿàñ Æ �y#0    ™ d $x3+     â r<(   ÿàñ Ê $x=)     Ê �yH<2   • h �yM<0   • h �yG<6   • b $x    ÿ Ž âO xÿ   ÿÀŽ È� x   âŽ ä xÿ     â $x5&   ÿ Ž âx&ð   ÿÀŽ â %v  8 Ž â xÿ   ÿÀŽ ä xÿ     ä �yA    ™ â x55   âŽ Ä x ð   ÿÀŽ â $x"   ÿ Ž â xÿ     ä xÿ     â %v45  8 Ž â x""   ÿÀŽ ä xÿ   ÿÀŽ â $x   ÿ Ž âO r/"   ÿ@� â�y?4    ™ & x   âŽ â %v./  8 Ž â�y0   ™ f %v  8 Ž â �y".    ™ b x[!ð   ÿÀŽ Ä $xW!   ÿ Ž â �y20   • h �y    ™ b x[ÿ   [  ä x$ÿ   ÿÀŽ ä xPXð   ÿÀŽ â$x.[   ÿ Ž â xWW   $âŽ â $x5X   ÿ Ž âO xSð   ÿÀŽ â �y00    • ¢  xÿ     Ä x$ÿ     ä xÿ   ÿÀŽ ä xPÿ   P  ä x$$ÿ   ÿÀŽ ä x.   âŽ â $x.P   ÿ Ž ä %vVW  8 Ž â x55   $âŽ Ä x..   ÿÀŽ â $x$   ÿ Ž ä �yVV    ™ ¢ %v45  8 Ž â r/.   ÿ@� ä xÿ   ÿÀŽ â %v6  8 Ž â �yZ4    ™ ¢  xSÿ   S  Ä xU$   âŽ â �yY6    ™ ¢  xÿ     â %v./  8 Ž Æ xÿ   ÿÀŽ â $x$S   ÿ Ž â �yW.    ™ ¢" %vTU  8 Ž Æ x5$   âŽ ä xÿ     â $x$I   ÿ Ž âxJIð   ÿÀŽ â �yTT    ™ ¢  xÿ   ÿÀŽ ä xJÿ   J  ä x$   âŽ â %v45  8 Ž È $xJ   ÿ Ž â �yQ4    ™ ¢ %vB  8 Ž Æ x   ÿÀŽ â %x2]8   2Ž â�yUB    ™ ¤ r   ÿ@� â %x4]8   <Ž â/ �y<4   • ¦ %v6  8 Ž È %x.]8   2Ž â �yR6    ™ ¨ �y20   • ¢ xBð   ÿÀŽ Æ �yO.0   • ¢  x9ÿ   ÿÀŽ â $x7   ÿ Ž â/ xBÿ   B  â �yN42   • ¦ x77   âŽ â �yL40   • ¢ xÿ   9  â$x<B   ÿ Ž â x$ÿ   9  â %x2]8   4Ž â x;ÿ   ÿÀŽ â �yD44   • ¨ �yK46   • ¢ %v67  8 Ž â x=<   ;âŽ Ä x<$ÿ   ÿÀŽ â $xC   ÿ Ž â�y6    ™ ¢  x4ð   ÿÀŽ â/ %x.]8   .Ž ä �yF22   • ¢  x$ÿ   4  â %v4=  8 Ž â x=C   <âŽ â �yE20   • ¤  $x<$   ÿ Ž Ä %v6=  8 Ž â �y44    ™ ¤ x<<   ÿÀŽ ä �y;24   • ¤  r99<   ÿ@� ä �y66    ™ ¦ %v89  8 Ž â �yC26   • ª  �y88    ™ ¢ %x<]8   2Ž Æ �y].0   • ¨ �y2<2   • ¨ �y3<4   • ¨  �y5<0   • ¨  �y\<6   • ¢  x.    ÿÀŽ â &z1>  € ÿ   â x   ÿÀŽ Ä Er. .       â &z=  € ÿ   ä &z<:  € 1   â x!!   ÿÀŽ ä x[[   ÿÀŽ ä xXX   ÿÀŽ ä x7AÿÀŽ Ä� xAÿ   A  Æ &r 7   ÿ   â x9AÿÀŽ â $t7ÿ   =ÿ Ž È #t/.  `¿7  âEr.        â &r99@       â x ?ÿÀŽ äx?ÿ   ?  â &z@  € =   â #t1.  `¿7  â Er. <       ä x<?ÿÀŽ â &r? '   ÿ   â Er        Ä x "ÿÀŽ â&r<<   ?   ä #r//   ÿ   ä &r  %   <   â x=ÿÀŽ ä x<ÿ   "  â #r11.   /   â &z"'  € ÿ   â x/&   ÿÀŽ â &r9=>   9   â x=<ÿÀŽ â &z&%  € ÿ   â x?   ÿÀŽ â &z  € "   â Er/ /       â &r"=#       â Er< ?       â &z=#  € &   â Er        â #t /  `¿7  âx&ÿ     â #t/<  `¿7  â Er =       â #r     ÿ   â rÿ     Ä x&&ÿÀŽ â #r /       â Er= "       â $t/ÿ   ÿ Ž â x"ÿ     â &r9&:   9   â Er< !       â $x   ÿ Ž â x""   /ìŽ ä Er& 9       Ä x   /ìŽ â #t/<  `¿7  â Er! "       â !r  =       â !r&1&       â Er        â #r1/   ÿ   â  r! !     A  â x/VÿÀŽ äO xVÿ   V  â #r&   !   â Er" [       â &r//   ÿ   â x!VÿÀŽ Ä x&ZÿÀŽ â #t "  `¿7  â xZÿ   Z  â &r!!@   /   â x"YÿÀŽ â &r&&'   ÿ   â xZZÿÀŽ ä xYÿ   Y  â &r!">   !   â x"WÿÀŽ â &r&Z   &   â xWÿ   W  Ä xYYÿÀŽ â &r&"%   &   â ErX X       ä xPP   ÿÀŽ ä x"WÿÀŽ â &r!Y:   !   â ErP P       â #r .    1   â #tXX  `¿7  â x1TÿÀŽ â &r""#   &   â x   ÿÀŽ Ä xTÿ   T  â #r/X   ÿ   â #tPP  `¿7  â Er! !       ä Er&        â &r11   ÿ   â x9TÿÀŽ â #r/P   /   â Er" "       â !r  !       â xSS   ÿÀŽ â #t!&  `¿7  â x<QÿÀŽ â &r19@   1   â x&UÿÀŽ â !r/"       â xUÿ   U  Ä Er" S       â &r1<>   1   â x<UÿÀŽ â &r&&'   ÿ   â #r!!   ÿ   â #t""  `¿7  â xII   ÿÀŽ â &r<<   &   ä #r&."   !   â Er" I       Ä xQÿ   Q  ä x/RÿÀŽ ä xRÿ   R  â #t""  `¿7  â xJJ   ÿÀŽ ä xQQÿÀŽ ä r!ÿ     â &r</%   <   â xRRÿÀŽ â $t9ÿ   ÿ Ž â ErJ J       â #r/"   ÿ   â &r1Q:   1   â x"ÿ   !  â &r<R#   <   â #tJJ  `¿7  Ä x""   9ìŽ ä Er 1       â #r/J   /   â rOÿ   O  â $tJÿ   ÿ Ž â Er< <       â $x!!   ÿ Ž â Er" "       â !r&       â $t1ÿ   ÿ Ž â x&ÿ   O  â !r</<       â x!!   JìŽ â  r/"     A  â $t"ÿ   ÿ Ž â x&&   1ìŽ â $xOO   ÿ Ž â Er !       Ä Er1 &       ä xOO   "ìŽ â #r    /   ä  r<<1     A  â Er  O       Ê #r     <   â xÿÀŽ ä xÿ     Æ &r   ÿ   â xÿÀŽ ä x4ÿÀŽ ä x4ÿ   4  â &r@      â x6ÿÀŽ ä x6ÿ   6  â &r>      â x4ÿÀŽ Ä x   ÿÀŽ â &r'   ÿ   â xHÿ   H  ä x66ÿÀŽ â &r:      â xNÿ   N  ä Er        ä x8ÿÀŽ â &r6      â xHHð   ÿÀŽ Ä xÿ     ä x8ÿ   8  ä x!Nð   ÿÀŽ ä xDÿ   D  ä x   ÿÀŽ ä x$$   ÿÀŽ â #t  `¿7  â rHMH   ÿ � â &r%      â xGG ð  ÿÀŽ Ä x   ÿÀŽ ä x88ÿÀŽ ä r!L!   ÿ � ä xKK ð  ÿÀŽ ä xDD   ÿÀŽ ä Er"        Ä Er$ $       ä r]ÿ   ]  â #r   ÿ   â xFÿ   F  ä x2ÿ   2  â $tÿ   ÿ Ž â rGH   àÿ â &r8#      â xBB   ÿÀŽ Ä rK!   Dàÿ â #t!"  `¿7  âxÿ   ]  â #t$$  `¿7  â xFFð   ÿÀŽ ä x;ÿ   ;  ä x22ð   ÿÀŽ ä x3ÿ   3  Ä ErB B       â #r!!   ÿ   â x   ìŽ â #r$      â Er        â $x]]   ÿ Ž â x   ÿàÿ ä rEEF   ÿ � ä xCC ð  ÿÀŽ Ä x;   ÿÀŽ ä r252   ÿ � ä x\\ ð  ÿÀŽ ä x33   ÿÀŽ â #tBB  `¿7  â Er        â !r       â |    pbð Ä rCE   àÿ ä r2\2   3àÿ â #r!.B   !   â Er        ä x]]   ìŽ â  r     A  â 0r0ÿ0    A   â !r!       â 0rÿ    A   â Er] ]       â 0rÿ    A   Ä 0rÿ    A   ä 0rÿ2    A   â x€  ÿà÷ â  r0     A  â  r0     A  â r0     A  â r0     A  â #r]      â #r      â #r      â #r       â #r      â x€   ÿàÿ â $rÿÿ   Ž â D	       À â Gy  Àäÿÿÿÿƒ ê Ay        € ê ˆ  þÿ H   è ˆ €þÿ H   è ˆ  ÿÿ H   è ˆ €ÿÿ H   è {          ì M)        € ê „y     H   " z 	 r  p`ô â $z   x  	Ž Ä „y €   H   b r ÿ   pVt Æ „y    H   ¨ „y    H   è„y €  H   ( „y    H   ( „y
    H   ( „y €  H   ( „y €  H   ( „y    H   ( „y €  H   " !r       Æ „y €  H   " !r       â/ !r       äO y      '   b !r       â� !r       â!r       ä !r
       â !r       È ‰      ¢ !r       â $z  ~   Ž â/ !r       Æ ‰      b !r       Æ ‰      â !r       Ê ‰      " !r       âO !r       â/ !r       È� ‰      h ‰      ¢ !r       Æ ‰      ( ‰      â !r       â/ !r       ÈO ‰
 €    b !r       Æ ‰ €    " !r       Æ� ‰ €    ¨ ‰ €    â !r

       â/ !r       È ‰
 @    " x	   ÿàÿ â !r       äO ‰ @    b x	   ÿàÿ â !r       ä� ‰ @    ¢ z  r  p`ö ä z  r  p`ò â ‰ @    â x    pVö Ä x    pVò  â !r
       â x
	   ÿàÿ â $t	ÿ   ÿ Ž â !r       ä/ ‰      " z 
 r  p`ð â !r       äO ‰      b x    pVp  â !r       Ä� ‰      ¨ ‰      â !r       â !r       â/ %v  j  	 Ž â !r       ÈO †©     â !r       Æ� †¹    è †‰    â M        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      &   " z ÿ \  pPò â ¹z  F   
   Æ z ÿ ]  Sò Ú         Ê % \   Ž Ê �	    ™ ¢  r ÿ       â Ey  0    € â r1 ÿ       â y<      "   b r2 ÿ       â '  m   Ž äy0      %   ä 'š s   Ž â  ÿ n     â y/      !   &  ÿ   ÿáÿ â y      '   ( $  o  Ž â ’         Ä r <ÿ   pRð ä/ x00   ÿ   ä� x /   pDp  är0/   ÿàÿ är<</   ÿ(Ž ä z  r  pdp  Ä xÿ   <  ä š	ÿ t    È r	     € ÊO G	  À    € ê z ÿ `  pPò ä z ÿ ^  pPð ä z ÿ b  p ô ä z ÿ a  Sò ä z ÿ _   Sð ä z ÿ c   'ò Ä r2 ÿ       ä r1 ÿ       Î $
 ~  ÿŽ âÿ   0  â$š ~  ÿŽ ä $
 x  ÿŽ â 0   /àë â$š x  ÿŽ â ’0   /àÇ ä ˜ÿ   0  ä    ÿàõ Ä ÿ   ÿÄþ ä ’   ÿàû ä ’ÿ   ÿ„þ ä ÿ   ÿä ä 
 ^  ÿ„ ä ’ÿ   ÿäÿ Ä š b  ÿ† ä 
 _   ä š c  � Æ �	2     h �™1     d Ay        € ê r<   ÿ� â ¹z  l      â $z3 }  ÿŽ â ™x?    â x;ÿ     â $z   w  ÿŽ â r ÿ       â Ey  @9    € â x4      ÿ  â | ;   pbð â x6      ÿ  â r33    ÿàó â 'z y   Ž â rI ÿ       â x8      ÿ  â rJ ÿ       Ä r:ÿÿ   ÿäÿ  Ê G	  �8    € ê xøÿÿÿÿÀŽ â $z0 p  ;Ž â zÿ z    â x4      ÿ  â r<<  €ÿàÿ â x6      ÿ  â x
0   ÿàÿ â $z |  Ž â r <<   ÿ� â $z |  ;Ž â x0   ÿàÿ â %x<   ÿŽ â xÿ      â x8      ÿ  â x0   ÿàÿ â $z=	 v  Ž âx<   ÿ   â $z	 v  Ž â x þÿÿÿÿÀŽ â $z
 p  =Ž â xÿ     â %x
   ÿŽ â r<  €ÿàÿ Ä z ÿ `  pPð â $z p  =Ž âz?  `      â $z p  =Ž â zA  a      â %x8   
Ž âzC  X      ä zE  Y      â %x8   
Ž â rI ÿ       Ä rJ ÿ       â %x8   Ž â z ÿ a   Sð ä x@;   ÿ   â %x
8   
Ž â xGÿ   <  Æ %x8   Ž È %x8   ÿŽ È %x8   Ž â r>   ÿàó Æ %x8   Ž â xB>!   ÿàõ ä rD   ÿäÿ  â $xF   ÿŽ ä %x8   Ž â rHÿD   ÿä Æ %x8   Ž È %x   Ž È %x<   Ž Ê rC   ÿàó ä rC>   ÿàõ ä rE   ÿäÿ  ä rED   ÿä Æ �y(    ™ ¨  �y !   ‘ è �yT    ‘ " rCB   ÿàó ä z_  p      Ä rEH   ÿäÿ  Æ %x _8   Ž È %x"_8   Ž â�yc     ™ ( �yK"ÿÿÿ ‘ ( �yL"    ‘ " %x_8    Ž È %x&_8   "Ž â�y%    ™ ( �yM&ÿÿÿ ‘ ( �yN&    ‘ " %x_8   Ž È/ %x_8   &Ž â �y$    ™ ( �yOÿÿÿ ‘ ( �yP    ‘ " r!<3   @à£ Ä x.        ä x" $       Ê %v! Z  " Ž â/ x ÿ   (  äO x(ÿ   ÿÀŽ ä x#    ÿ   ä� x  ÿ   ÿÀŽ ä xU ð   ÿÀŽ ä x#    âŽ Ä r G:   ÿDþ  ä x#T   ÿ   â%v  . Ž â xXTð   ÿÀŽ ä xUÿ   U  â $x $   ÿŽ â x##   âŽ â �yS    ™ ¢  xÿ   (  Ä xXÿ   X  â %v"#  . Ž â r   ÿàÿ ä xU   ÿ   ä xÿ   ÿÀŽ â �yY"    ™ â x!X   ÿ   ä x   ÿÀŽ â �y+   ™ â x!   âŽ Ä r(   ÿ@� â �y-   ™ ä %v  . Ž ä�y*   ™ ä %v   . Ž ä �yR    ™ è  �yV     ™ è  �y,   ™ è �y(   ™ è �y)   ™ â rC   ÿàó Æ �y'    ™ â rE   ÿäÿ  Æ �y&   ™ â x"cÿ   ÿÀŽ ä/ x!K   ÿ   â�yW0   • â  xQKð   ÿÀŽ ä x!!   "âŽ ä xÿ   c  ä xQÿ   Q  Ä x\ÿ   c  â %v !  . Ž â x"ÿ   ÿÀŽ ä x#Q   ÿ   ä rC   ÿàó â �y[     ™ â x]\ÿ   ÿÀŽ ä x\Lð   ÿÀŽ Ä x##   "âŽ ä rE   ÿäÿ  ä x^L   ÿ   ä x\ÿ   \  â %v"#  . Ž â�yZ2   • â  xa^   ]âŽ ä xb\   ÿ   â �y`"    ™ ä  %v a  . Ž â/ xdb   ÿÀŽ â �y^0   • æ rccd   ÿ@� â �yb     ™ è %v"c  . Ž â xd%ÿ   ÿÀŽ â �y]4   • â  xgM   ÿ   Æ �yc"    ™ â  %x_8   Ž â xgg   dâŽ ä xfÿ   %  â �ya6   • ä %v g  . Ž ä/ �ye0   • â xgN   ÿ   Ä x"fÿ   ÿÀŽ â �yd    • æ  x#g   "âŽ â %x_8   Ž âxgMð   ÿÀŽ â �yk     ™ â x"ÿ   %  ä xgÿ   g  â �yf2   • â xjNð   ÿÀŽ Ä x""ÿ   ÿÀŽ â �yi4   • â xg   ÿ   ä xjÿ   j  â �yl0   • â  x!   "âŽ äx"j   ÿ   â �yh6   • " %v#  . Ž Ä x""   ÿÀŽ ä %v !  . Ž â �ym    ™ " r#%"   ÿ@� Æ �yn     ™ $ %v"#  . Ž â x%$ÿ   ÿÀŽ ä xpO   ÿ   ä �yo"    ™ " %x_8   Ž â/ x%p   %âŽ ä xpOð   ÿÀŽ Ä x ÿ   $  â �yq0   • " xpÿ   p  ä x!ÿ   $  â %v%  . Ž â x  ÿ   ÿÀŽ ä x#p   ÿ   ä x%P   ÿ   â �yt    ™ " x"!ÿ   ÿÀŽ Ä xrPð   ÿÀŽ ä x!#    âŽ ä x#%   "âŽ ä xrÿ   r  â %v!  . Ž Æ/ x"r   ÿ   â %v #  . Ž â �yv    ™ $ x#"   ÿÀŽ Æ �y      ™ " r#$#   ÿ@� Ê %v"#  . Ž È %x_8   Ž ä�y""    ™ $ %x_8   Ž ä�ys2   • $ %x$_8   Ž ä �y0   • ( �y_$2   • ( �yw4   • ( �y!$4   • (# �yu0   • ( �yx6   • ( �yy$0   • ( �yz$6   • " x     ÿÀŽ Ä xTT   ÿÀŽ ä xXX   ÿÀŽ ä Er T       ä x%        ä/ x|SÿÀŽ äO xSÿ   S  Ä xYÿÀŽ ä� xYÿ   Y  Æ &r+   ÿ   â xYYÿÀŽ ä xSSÿÀŽ â &r||-   ÿ   ä &r#Y*      â xRÿÀŽ ä xVÿÀŽ ä xVÿ   V  â &r|S,   |   Æ x$VÿÀŽ â &r#(   #   â Er         â &zS+  € ÿ   â x U   ÿÀŽ â &r)   |   ä &z{-  € ÿ   â Er}         â &r$$'      â x    =    â &z *  € S   Ä &zY(  € ÿ   ä &zT)  € ÿ   ä &z,  € {   â Er          â #t  `¿  â xVÿ   R  â &z&  € Y   â ErU X       â &zS'  € T   â Er        â #t  `¿  â rWÿF   W  â #rR    ÿ   â xVVÿÀŽ â #tUU  `¿  â Er        â #tT}  `¿  â Er S       â #r   ÿ   â xXÿ   W  â &r#V&   #   â #rUU   R   â #rT      â x%X   %ìŽ Ä ErR $       ä xT        ä xWW   ÿ   ä Er$ #       â !rRR       â Er% %       ä xWW   TìŽ â !rU$       â xT[ÿÀŽ Ä x[ÿ   [  â  rRR%     A  â Er$ W       ä xZÿ   Z  â &rTT+   ÿ   â x[[ÿÀŽ â #r$   R   â xKK   ÿÀŽ ä x#Zð   ÿÀŽ Ä x$`ÿÀŽ â &r[[*   T   â x`ÿ   `  ä ErK K       ä xQQ   ÿÀŽ ä r^^#   ÿ � â &r$$(   [   â x#`ÿÀŽ Ä xRbÿÀŽ â #tKK  `¿  â ErQ Q       ä xbÿ   b  â &r##&   $   â #rK K   ÿ   â #tQQ  `¿  â x%bÿÀŽ â &rRR-   ÿ   â xLL   ÿÀŽ â #rKQ   K   â x$cÿÀŽ â &rR%,   R   â ErT #       Ä xcÿ   c  ä Er% L       ä x\\   ÿÀŽ â !r#KT       â xLcÿÀŽ â &r$$)   R   â ErK \       â #t%%  `¿  â x]ÿ   ]  Ä reÿF   e  â &rLL'   $   â xaa ð  ÿÀŽ â #r$%   ÿ   â x]]   ÿÀŽ â #tKK  `¿  â xT        ä x%ÿ   e  â #r$K   $   â r^a^   ]àÿ Ä xR        ä xee   ÿ   ä xK%   TìŽ ä ErQ L       â 0rdÿd    A   â xee   RìŽ â 0r%ÿ^    A   â ErL K       â !rQ$Q       â xfÿ   f  Ä Er$ e       â  r%%d     A  â  rLQL     A  â xKfð   ÿÀŽ ä xiÿ   i  â #rJ%   J   â #r$#$   L   â rlK   ÿ � ä xhh ð  ÿÀŽ ÄxLi   ÿÀŽ ä xRkÿÀŽ ä xMM   ÿÀŽ ä xkÿ   k  ä rh   Làÿ ä xLmÿÀŽ Ä xmÿ   m  â &rRR+   ÿ   â ErM M       ä xgg   ÿÀŽ ä xkkÿÀŽ â &r%L-   ÿ   â xmmÿÀŽ â #tMM  `¿  â Erg g       â &rkk*   R   â x#nÿÀŽ Ä xKN   ÿÀŽ ä xnÿ   n  â &rNm,   %   â #rM M   ÿ   â #tgg  `¿  â &rk#(   k   â x%oÿÀŽ ä ErK K       ä xjj   ÿÀŽ Ä xoÿ   o  ä x#nÿÀŽ â #rLg   M   â ErM j       â #tKK  `¿  â xRoÿÀŽ â &r%%)   N   ä &r##&   k   â rqÿF   q  â #rNK   ÿ   â #tMM  `¿  â &r%R'   %   â ErQ #       Ä xR        ä x#q   ÿ   â #rMM   N   â xKÿ   q  ä ErN %       ä x##   RìŽ ä xKK   TìŽ Ä xRtÿÀŽ ä xtÿ   t  â !r%MN       â xpp   ÿÀŽ ä ErN K       â !rLLQ       â xOO   ÿÀŽ â &rRR+   ÿ   â xKtÿÀŽ Ä ErQ p       ä xMvÿÀŽ â &rKK*   R   â ErO O       ä xvÿ   v  ä xR ÿÀŽ ä xSÿ      â #t Q  `¿  âxQvÿÀŽ â #tOO  `¿  â &rKM(   K   â xSSÿÀŽ Ä #rO O   ÿ   â &rKQ&   K   â xPP   ÿÀŽ â &rRR-   ÿ   â #r     O   â xM"ÿÀŽ ä ErO K       â &rRS,   R   â ErK P       ä x"ÿ   "  Ä xrr   ÿÀŽ â !rQ O       â x""ÿÀŽ â &rMM)   R   â #tKK  `¿  â ErO r       ä rÿF     â #r K   ÿ   â xsÿ   s  â &r""'   M   â x_ÿ   _  â  r%%N     A  â xR        â #tOO  `¿  â xKÿ     Ä xNsð   ÿÀŽ ä xwÿ   w  ä xP_ð   ÿÀŽ ä x!ÿ   !  â #r O       â xKK   RìŽ ä ErM "       Ä rNuN   ÿ � ä xxx ð  ÿÀŽ ä xww   ÿÀŽ ä rPyP   ÿ � ä xzz ð  ÿÀŽ ä x!!   ÿÀŽ â !r  M       â x   ÿ   â 0rÿ    A   â ErK K       Ä rNxN   wàÿ ä r!zP   !àÿ â  rP K     A  â x   RìŽ â  r d     A  â 0rÿN    A   â Er# #       â 0r!ÿ!    A   â Er"        â #rI $   I   â  rNd     A  â #rLL#   %   â  rd!     A  â #rQ"   P   â G‰  @    € ê r$?   ÿàó ä r ?>   ÿàõ Ä r%A   ÿäÿ  ä r!AD   ÿä Æ �yK$     ¨ �yT      è �yW !    " rP?   ÿàõ È rQA   ÿä Ì �yQP0    " x"Kÿ   ÿÀŽ äO xMÿ   K  äx#T   ÿ   ä� xVÿ   T  ä x##   "âŽ ä xOÿ   K  Ä xMMÿ   ÿÀŽ â %v"#  . Ž â x V   ÿ   ä xXÿ   W  äx$W   ÿ   â �yU"    ™ ¢  xOOÿ   ÿÀŽ ä xM    MâŽ Ä x X   ÿ   ä x%$   OâŽ ä x     ÿÀŽ â %v"M  . Ž È %v$%  . Ž â rSK    ÿ@� â �yM"    ™ è  �yO$    ™ â %vRS  . Ž â r ?   ÿàó È r!A   ÿäÿ  â �yRR    ™ è �yZ 3    è �y" 4    è �yY 0    è �y# 6    â xWW   ÿÀŽ Ä xXX   ÿÀŽ ä ErW W       ä xTT   ÿÀŽ ä Er! X       â/ #tWW  `¿  â Er% T       Æ#rPW   ÿ   â #t!!  `¿  â xVV   ÿÀŽ Æ #rP!   P   â ErK V       â #t%%  `¿  â rQÿF   Q  Æ #r% %   ÿ   â #tKK  `¿  È #r%K   %   â xK        ä x$UÿÀŽ äO xUÿ   U  Æ &r$$+   ÿ   â xUUÿÀŽ Ê &rUU*   $   â x$OÿÀŽ ä� xOÿ   O  Æ &r$$-   ÿ   â xOOÿÀŽ ä x!RÿÀŽ ä xRÿ   R  â &r$O,   $   â x MÿÀŽ ä xMÿ   M  â &r!!)   $   â xRRÿÀŽ â &r  (   U   â xMMÿÀŽ Ä xZZð   ÿÀŽ â &r!R'   !   â x"ÿ   "  ä x$ÿ   Q  â &r M&       â rYYZ   ÿ � ä x## ð  ÿÀŽ ä x""   ÿÀŽ Ä x$$   KìŽ ä xR        ä xQQ   ÿ   ä Er! !       ä r"#Y   "àÿ ä Er          â !rPP!       â xQQ   RìŽ Ä Er# $       â 0r!ÿ"    A   â Er" Q       â !r%%        ä  rPP#     A  â  r d!     A  Æ #r"%"   P   È #r9 "   9   ä G‰  p    € ê r$?   ÿàó är ?   ÿàõ ä r%A   ÿäÿ  är!A   ÿä Æ �y$$     ¨ �yO      è �yV !    " xX0   ÿàÿ ä x#$ÿ   ÿÀŽ ÄO x"ÿ   $  äxPO   ÿ   ä� xUÿ   O  ä x#P   #âŽ ä xK"ÿ   ÿÀŽ ä x%ÿ   $  â %v"#  . Ž â xPU   ÿ   Ä x V   ÿ   äx%%ÿ   ÿÀŽ â �yT"    ™ ¢  xKP   KâŽ ä xWÿ   V  ä x%    %âŽ â %vPK  . Ž â x W   ÿ   Æ %vR%  . Ž â x!    ÿÀŽ â �yPP    ™ æ r!$!   ÿ@� â �yRR    ™ " r ?   ÿàó Æ %v"!  . Ž â r$ ?       ä r% A       â$zKX p  =Ž â r!A   ÿäÿ  â �yX"    ™ $	 %x$K8   $Ž ä �y[ 0    ( �yZ$3    ( �yK$4    ( �yY$0    ( �yM$6    " xVV   ÿÀŽ Ä xOO   ÿÀŽ ä ErQ V       ä/ xWW   ÿÀŽ ä ErO O       â #tQQ  `¿  âxUU   ÿÀŽ ä ErW W       â #tOO  `¿  â ErU U       â #r$Q   ÿ   Ä #tWW  `¿  â#rO O   ÿ   â #tUU  `¿  ä #r$W   $   ä #rOU   O   â x TÿÀŽ äO xTÿ   T  Æ &r  +   ÿ   â x!TÿÀŽ Ê &r!!*       â x PÿÀŽ ä� xPÿ   P  ä x"RÿÀŽ äxRÿ   R  â &r  (   !   â x#PÿÀŽ â &r""-   ÿ   â x!RÿÀŽ Æ &r #&       â x#XÿÀŽ â &r"!,   "   â xXÿ   X  ä r[ÿF   [  â &r"#)   "   â xXXÿÀŽ ä xP        ä x!ÿ   [  â &r"X'   "   â xZZð   ÿÀŽ Ä xKÿ   K  ä x!!   PìŽ ä rYYZ   ÿ � ä xMM ð  ÿÀŽ ä xPK   ÿÀŽ ä xR        Ä x[[   ÿ   ä Er# "       ä rMMY   Pàÿ ä Er          â !r#$#       â x[[   RìŽ ä Er" !       â 0rMÿM    A   â Er[ [       â !r O        Ä  r##"     A  â  rMdM     A  Æ #r  [   #   È #r8M    8   ä G‰  p    € ê r$?   ÿàó är ?   ÿàõ ä r%A   ÿäÿ  är!A   ÿä Æ �y$$     ¨ �yO      è �yV !    " xX0   ÿàÿ ä x#$ÿ   ÿÀŽ ÄO x"ÿ   $  äxPO   ÿ   ä� xUÿ   O  ä x#P   #âŽ ä xK"ÿ   ÿÀŽ ä x%ÿ   $  â %v"#  . Ž â xPU   ÿ   Ä x V   ÿ   äx%%ÿ   ÿÀŽ â �yT"    ™ ¢  xKP   KâŽ ä xWÿ   V  ä x%    %âŽ â %vPK  . Ž â x W   ÿ   Æ %vR%  . Ž â x!    ÿÀŽ â �yPP    ™ æ r!$!   ÿ@� â �yRR    ™ " r ?   ÿàó Æ %v"!  . Ž â r$ ?       ä r% A       â$zKX p  =Ž â r!A   ÿäÿ  â �yX"    ™ $	 %x$K8   $Ž ä �y[ 0    ( �yZ$3    ( �yK$4    ( �yY$0    ( �yM$6    " xVV   ÿÀŽ Ä xOO   ÿÀŽ ä ErQ V       ä/ xWW   ÿÀŽ ä ErO O       â #tQQ  `¿  âxUU   ÿÀŽ ä ErW W       â #tOO  `¿  â ErU U       â #r$Q   ÿ   Ä #tWW  `¿  â#rO O   ÿ   â #tUU  `¿  ä #r$W   $   ä #rOU   O   â x TÿÀŽ äO xTÿ   T  Æ &r  +   ÿ   â x!TÿÀŽ Ê &r!!*       â x PÿÀŽ ä� xPÿ   P  ä x"RÿÀŽ äxRÿ   R  â &r  (   !   â x#PÿÀŽ â &r""-   ÿ   â x!RÿÀŽ Æ &r #&       â x#XÿÀŽ â &r"!,   "   â xXÿ   X  ä r[ÿF   [  â &r"#)   "   â xXXÿÀŽ ä xP        ä x!ÿ   [  â &r"X'   "   â xZZð   ÿÀŽ Ä xKÿ   K  ä x!!   PìŽ ä rYYZ   ÿ � ä xMM ð  ÿÀŽ ä xPK   ÿÀŽ ä xR        Ä x[[   ÿ   ä Er# "       ä rMMY   Pàÿ ä Er          â !r#$#       â x[[   RìŽ ä ErP !       â 0rMÿM    A   â Er" [       â !rOO        Ä  rP#P     A  â  r dM     A  Æ #r"O"   P   È #r7 "   7   â #r5NL   5   â #r4   4   â G‰  p    € ê r$?   ÿàó är"?   ÿàõ ä r%A   ÿäÿ  är#	A   ÿä Æ �y$     ¨ �yP"     è �yS"!    " xN0   ÿàÿ Ê $zON p  =Ž â xÿ   ÿÀŽ äO xKÿ     ä x P   ÿ   È� x    âŽ ä xTÿ   S  äxLS   ÿ   â %v   . Ž â xKKÿ   ÿÀŽ ä xÿ     ä xRÿ   P  â �yQ     ™ ¢  x"T   ÿ   Ä x#L   KâŽ ä xÿ   ÿÀŽ ä x$R   ÿ   ä x "   ÿÀŽ â %v"#  . Ž â x$   âŽ ä rM    ÿ@� Ä �yK"    ™ â  %v$  . Ž â r 
?   ÿàó Æ %vLM  . Ž ä �y$$    ™ " r!A   ÿäÿ  ä r" ?       â �yLL    ™ " r# A       Æ �yW 0    $ %xNO8   "Ž Ê �y.N3    ( �y%N4    (/ �y#N0    ( �yUN6    " xPP   ÿÀŽ Ä xRR   ÿÀŽ ä Er P       ä xSS   ÿÀŽ ä Er R       â #t  `¿  â ErS S       Æ #r    ÿ   â #t  `¿  â#tSS  `¿  â xTT   ÿÀŽ ä #r      â #rS   ÿ   â Er! T       Ê #t!!  `¿  â x         Æ #r!      â x"QÿÀŽ äO xQÿ   Q  Æ &r""+   ÿ   â xQQÿÀŽ ä xKÿÀŽ ä� xKÿ   K  â &rQQ*   "   ä &r-   ÿ   â xKKÿÀŽ ä x $ÿÀŽ Äx$ÿ   $  ä xLÿÀŽ â &rK,      â xLÿ   L  â &r  (   Q   â x$ÿÀŽ â &r)      â rWÿF   W  ä xLLÿÀŽ â &r &       â x        Ä xÿ   W  â &rL'      â x..ð   ÿÀŽ ä x%ÿ   %  ä x   ìŽ ä r##.   ÿ � ä xUU ð  ÿÀŽ Ä x%   ÿÀŽ ä xWW   ÿ   ä Er        ä rU#   àÿ ä Er          â !r       â xWW    ìŽ Ä Er        â 0rÿ    A   â ErW W       â !r         ä  r     A  â  rd     A  Æ #r  W      È #r6    6   â x;;   ÿàÿ ä x??€  ÿàõ ä | ;   pbò ä xCC€  ÿà÷ ä x@@€   ÿàÿ ä rAÿA   ÿä Ä rEÿE   ÿäÿ Ê D       À â Gy  °Êÿÿÿÿƒ ê Ay        € ê y       "   " Ey  0    € â r  ÿ   pRð är  ÿ   pRò Ö G‰  ð     € ê z ÿ `  pPð ä r  /   ÿ8Ž ä z ÿ a   Sð Æ ˆs  J þÿ H   ô ˆƒ  I€þÿ H   è ˆƒ  5 ÿÿ H   è ˆƒ  4€ÿÿ H   â G‰  p     € ê ˆs  9   H   è ˆs  I€þÿ H   è ˆs  8€  H   è ˆs  5 ÿÿ H   è ˆs  7   H   è ˆs  4€ÿÿ H   è ˆs  6€  H   ä Ay        € ê {          ì M        € ê y      &   ¢ z ÿ `  pPð Æ y      '   â z ÿ a   Sð â $z  x  0Ž â x        Æ $z  ~   Ž È� %v  j   Ž È G	  P    € ê „y/     H   " z 0 r  p`ð Æ „y/ €   H   b r /ÿ   pVp  Æ „y/    H   ¨ „y /    H   è „y	/ €  H   ( „y/    H   ( „y/    H   ( „y/ €  H   ( „y/    H   ( „y/ €  H   ( „y
/ €  H   " !rJ       Æ „y/ €  H   " !rI       â/ !r5       âO !r         ä� !r	       â!r       â !r         ä !r       Æ ‰       b !r       Æ ‰      ¢ !r4       Æ ‰      â !r

       È !r
       â !r        â/ !r       ÈO ‰      " !r
       Æ� ‰      h ‰      ¨ ‰	
      â !r       â !r       â/ !r       ÈO ‰      " !r
	       Æ� ‰ €    h ‰  €    ¨ ‰	 €    â !r       â !r       È/ ‰ €    " !r
        ÆO ‰  @    b !r		       â� x0   ÿàÿ ä ‰
 @    ¤ z  r  p`ô ä ‰	 @    ä x /   pVt â !r       â !r         È/ ‰ @    " x0   ÿàÿ â !r
       âO x00   ÿàÿ â ‰       b z  r  p`ò â !r	       ä� ‰      ¢ x /   pVò  Æ ‰      â !r       â !r        È/ ‰       " !‚2       â!r	       âO !r       Æ� †‰     â z 0 r  p`ð â !¢		2       â!’2       ä x /   pVp  ä †© 	   è †™    â !r        Ì M	        € ê !r2       Ê/ †y    â My        € ê z   h      È x     pBð Ú G	  @    € ê x     pRð Ú G‰  À    € ê x     pRð Ú G	   $    € ê „y	/     H   " z 0 r  p`ð â Ey  p    € ä „y /    H   b r /ÿ   pVp  Æ „y/    H   ¨ „y/    H   è „y/    H   ( „y/  
  H   ( „y/ €   H   ( „y/ €  H   ( „y/ €  H   ( „y/ €  H   " !r		J       â !r  9       Æ/ !r	       âO !r         Æ� !r       â!r         È ‰      " !rI       Æ ‰       b !r8       â !r       È !r       â !r       â !r        È/ ‰      ( ‰      b !r       â !r	       È/ ‰ €    ( ‰
	 €    h „y/ €  H   ¢ !r       Æ „y/ €
  H   â !r
	
       Æ/ ‰ @    ( ‰
 @    b !r       âO !r       â !r
       È/ ‰       ( ‰      b !r       â !r	        â !r       â/ G	  �     € ê !r1       â!r 	2       Æ  x;ª¸¿  A  Ì s        $ !t  €?    Ì s        $  r     A  È  r      A  Ê †y     ä Ay        € ê ‰       b Ey  Ð    € æ ‰      ¨ „y/    H   è „y/    H   ( „y/  	  H   â !r         Æ/ „y/    H   â !r       ÆO ‰       h ‰      ( „y/    H   ¢ !r5       È� !r       â!r        â/ !r	       È ‰ €    " x0   ÿàÿ Æ ‰
	 €    b z  r  p`ð â !r       Æ x /   pVp  â !r7       ÈO !r       â !r       ä „y/    H   â !r
	
       Æ/ ‰ @    ( ‰
 @    b !r       â !r
       È/ ‰       ( ‰      b !r       â !r	        â !r       â/ G	  �     € ê !r1       â!r 	2       Æ  x;ª¸¿  A  Ì s        $ !t  €?    Ì s        $  r     A  È  r      A  Ê †y    ä Ay        € ê ‰       b Ey       € æ ‰      ¨ „y/ €  H   è „y/ €  H   è „y/ €  H   è „y/ €  H   â !r         Æ/ „y/ €  H   â !r       ÆO ‰       h ‰      ( „y/ €	  H   ¢ !r6       â� !r        â/ !r	       È ‰ €    ( ‰
	 €    b !r       ÈO !r       â !r       â x0   ÿàÿ â !r
	
       Æ/ ‰ @    " z  r  p`ð Æ ‰
 @    b x /   pVp  â !r       â !r
       È/ ‰       ( ‰      b !r	        â !r       â/ G	  �     € ê !r1       â!r 	2       Æ  x;ª¸¿  A  Ì s        $ !t  €?    Ì s        $  r     A  È  r      A  Ê †y    ä Ay        € ê ‰      b !r4       â x00   ÿàÿ Æ !r       â z 0 r  p`ð Æ !r       â x /   pVp  È ‰       ¢ !r       Ê/ ‰      " !r         ÊO ‰       b !r	       Ê ‰
	 €    " !r        Ê/ ‰ €    b !r
	
       Ê ‰
 @    " !r       Ê/ ‰ @    b !r
       Ê ‰      " !r       Ê/ ‰       ¢ !r       â M	        € ì !r11       â/!r        ÆO  x1;ª¸¿  A  â !r22       Ê s        $ !t  €?    Ì s        $  r11     A  È  r121     A  Ê †y 1   â My        € ê „y/     H   " z 0 r  p`ð â Ey       € ä „y /    H   b r /ÿ   pVp  Æ „y/    H   ¨ „y/    H   è „y	/    H   ( „y/  
  H   ( „y/ €   H   ( „y/ €  H   ( „y/ €  H   ( „y/ €
  H   â !rJ       â !r  9       Æ/ !r       âO !r         Æ� !r		       â!r        È ‰ 	      " !rI       Æ ‰      b !r       È !r       â !r 	        â !r       È/ ‰       ( ‰      h „y/ €  H   ¢ !r        â !r       È/ ‰ €    ( ‰ €    h „y/ €  H   â !r8       âO !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       È� !r       â !r       â !r
	
       È/ ‰       ( ‰
      b !r	        â !r 
       â/ G	  À     € ê !r  1       È x '7=  A  â x *BL?  A  Æ #t   €?   È  r     A  â  x    ?  A  â !r 	2       È s     $   $ !t  €?    È  r     A  È  r      A  Ê †y     ä Ay        € ê ‰       b Ey       € æ ‰      ¨ „y/    H   è „y/    H   ( „y/    H   è „y/  	  H   â !r         Æ/ „y/    H   â !r       ÆO ‰       h ‰      " !r5       â� !r7       â!r        ä/ !r       â !r       Ä ‰ €    " !r       Æ ‰ €    b !r       Æ „y/    H   ¢ !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       âO !r       â x0   ÿàÿ â !r
	
       Æ/ ‰       " z  r  p`ð Æ ‰
      b x /   pVp  â !r	        â !r 
       Ö/ G	  À     € ê !r  1       È x '7=  A  â x *BL?  A  Æ #t   €?   È  r     A  â  x    ?  A  â !r 	2       È s     $   $ !t  €?    È  r     A  È  r      A  Ê †y    ä Ay        € ê ‰       b Ey  Ð    € æ ‰      ¨ „y/ €	  H   è „y/ €  H   è „y/ €  H   â !r         Æ/ „y/ €  H   â !r       ÆO ‰       h ‰      ( „y/ €  H   ¢ !r        â/ !r       È ‰ €    ( ‰ €    h „y/ €  H   â !r6       ÈO !r       â !r       Æ !r       â !r	       ä/ ‰ @    ( ‰
	 @    b !r       â x0   ÿàÿ â !r
	
       Æ/ ‰       " z  r  p`ð Æ ‰
      b x /   pVp  â !r	        â !r 
       Ö/ G	  À     € ê !r  1       È x '7=  A  â x *BL?  A  Æ #t   €?   È  r     A  â  x    ?  A  â !r 	2       È s     $   $ !t  €?    È  r     A  È  r      A  Ê †y    ä Ay        € ê ‰      b !r4       â x00   ÿàÿ Æ !r       â z 0 r  p`ð Æ !r       â x /   pVp  È ‰       ¢ !r       Ê/ ‰      " !r         ÊO ‰       b !r       Ê ‰ €    " !r        Ê/ ‰ €    b !r	       Ê ‰
	 @    " !r       Ê/ ‰ @    b !r
	
       Ê ‰
      " !r       Ê/ ‰       ¢ !r
       â M	        € ì !r11       â/!r        ÆO  x1'7=  A  â x1*BL?  A  â !r22       ä #t1  €?   â x11   ?  A  Æ  r     A  Ì s     $   $ !t   €?    È  r11      A  È  r121     A  Ê †y 1   â My        € ê x     pRð Ú G‰  P    € ê x     pRð Ú G	  `    € ê „y/     H   " z 0 r  p`ð â Ey       € ä „y /    H   b r /ÿ   pVp  Æ „y/    H   ¨ „y/    H   è „y	/    H   ( „y/  
  H   ( „y/ €   H   ( „y/ €  H   ( „y/ €  H   ( „y/ €
  H   â !rJ       â !r  9       Æ/ !r       âO !r         Æ� !r		       â!r        È ‰ 	      " !rI       Æ ‰      b !r       È !r       â !r 	        â !r       È/ ‰       ( ‰      h „y/ €  H   ¢ !r        â !r       È/ ‰ €    ( ‰ €    h „y/ €  H   â !r8       âO !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       È� !r       â !r       â !r
	
       È/ ‰       ( ‰
      b !r	        â !r 
       â/ G	  À     € ê !r  1       Ê	z  i    � â !r 	2       È  x;ª¸¿  A  â 	z   i    � È 	z   i €  � â s        $ !t  €?    Ì s        $  r     A  È  r      A  Ê †y     ä Ay        € ê ‰       b Ey       € æ ‰      ¨ „y/    H   è „y/    H   ( „y/    H   è „y/  	  H   â !r         Æ/ „y/    H   â !r       ÆO ‰       h ‰      " !r5       â� !r7       â!r        ä/ !r       â !r       Ä ‰ €    " !r       Æ ‰ €    b !r       Æ „y/    H   ¢ !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       âO !r       â x0   ÿàÿ â !r
	
       Æ/ ‰       " z  r  p`ð Æ ‰
      b x /   pVp  â !r	        â !r 
       Ö/ G	  À     € ê !r  1       Ê	z  i    � â !r 	2       È  x;ª¸¿  A  â 	z   i    � È 	z   i €  � â s        $ !t  €?    Ì s        $  r     A  È  r      A  Ê †y    ä Ay        € ê ‰       b Ey  Ð    € æ ‰      ¨ „y/ €	  H   è „y/ €  H   è „y/ €  H   â !r         Æ/ „y/ €  H   â !r       ÆO ‰       h ‰      ( „y/ €  H   ¢ !r        â/ !r       È ‰ €    ( ‰ €    h „y/ €  H   â !r6       ÈO !r       â !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       â� !r       â x0   ÿàÿ â !r
	
       Æ/ ‰       " z  r  p`ð Æ ‰
      b x /   pVp  â !r	        â !r 
       Ö/ G	  À     € ê !r  1       Ê	z  i    � â !r 	2       È  x;ª¸¿  A  â 	z   i    � È 	z   i €  � â s        $ !t  €?    Ì s        $  r     A  È  r      A  Ê †y    ä Ay        € ê ‰      b !r4       â x00   ÿàÿ Æ !r       â z 0 r  p`ð Æ !r       â x /   pVp  È ‰       ¢ !r       Ê/ ‰      " !r         ÊO ‰       b !r       Ê ‰ €    " !r        Ê/ ‰ €    b !r	       Ê ‰
	 @    " !r       Ê/ ‰ @    b !r
	
       Ê ‰
      " !r       Ê/ ‰       ¢ !r
       â M	        € ì !r11       â/!r        ÈO 	z11 i    � â !r22       È  x1;ª¸¿  A  â 	z22 i    � È 	z22 i €  � â s        $ !t  €?    Ì s        $  r11     A  È  r121     A  Ê †y 1   â My        € ê „y/     H   " x0   ÿàÿ äx0   ÿàÿ â„y /    H   b z 0 r  p`ð ä z  r  p`ò â „y/ €   H   ¢ z  r  p`ô Ä r /ÿ   pVp  â„y/ €  H   â x /   pVò  äx /   pVt â „y
/    H   " x00   ÿàÿ Æ „y/    H   " z 0 r  p`ö Æ „y/    H   " x /   pVö Æ „y/    H   ( „y/ €  H   ( „y/ €  H   ( „y/    H   " !rJ       Æ „y/  	  H   " !r  9       Æ/ „y	/    H   b !rI       ÆO „y/  
  H   ¢ !r88       Æ� „y/ €  H   â !r

5       Æ„y/ €
  H   " !r7       Æ „y/    H   " !r       Æ „y/    H   " !r         Æ „y/ €  H   " !r       Æ „y!/ €  H   " !r8       Æ „y/ €  H   " !r

       Æ „y/ €	  H   â !r       Æ „y#/ €  H   â !r	       Æ/ „y%/ €  H   â !r        âO !r       ä� ‰       " !r       Æ‰      b !r

       Æ ‰      ¢ !r       Æ ‰      â !r4       Æ ‰	
      " !r!!6       Æ ‰      " !r       â !r         â !r!       â !r       â/ !r##       â !r       äO ‰       " !r%%       â !r       Ä� ‰#      b !r

	       Æ‰%      ¢ !r       Æ ‰      è ‰	      ( ‰      ( ‰
      " !r        Æ ‰      " !r#       â/ !r%       äO ‰  €    b !r       Æ� ‰      ¢ !r		       Æ‰      â !r       Æ ‰ €    " !r
       Æ ‰	 €    " !r       Æ ‰ €    ( ‰
 €    " !r         Æ/ ‰ €    b !r       âO !r       ä� ‰  @    ¢ !r       Æ‰ €    â !r	       Æ ‰ €    " !r       Æ ‰ @    " !r

       Æ ‰	 @    " !r       Æ/ ‰ @    h ‰
 @    b !r        ÆO ‰ @    ¢ !r       â� !r       ä‰       â !r       Æ ‰ @    " !r		       Æ ‰ @    " !r       Æ/ ‰      b !r
       Æ ‰	      b !r       ÆO ‰      ¨ ‰
      ¢ !r        Æ� ‰      â !r       â !‚ 2       â!r       Æ‰      " !r       Æ/ ‰      b !r		       â !‚1       â!r       äO !’	2       â  ‚      A  â !r
       â !’1       Æ †‰     â !r       â� !¢2       â  ’     A  ä !¢	1       â !r       ä †™    â  ¢		     A  â !r       È/ †© 	   â M9        € ê !r22       â !r11       È  r121     A  Ê †y 1   â My        € ê „y/     H   " z 0 r  p`ð â Ey  À    € ä „y /    H   b r /ÿ   pVp  Æ „y/    H   ¨ „y/    H   è „y	/    H   ( „y/  
  H   ( „y/ €   H   ( „y/ €  H   ( „y/ €  H   ( „y/ €
  H   â !rJ       â !r  9       Æ/ !r       âO !r         Æ� !r		       â!r        È ‰ 	      " !rI       Æ ‰      b !r       È !r       â !r 	        â !r       È/ ‰       ( ‰      h „y/ €  H   ¢ !r        â !r       È/ ‰ €    ( ‰ €    h „y/ €  H   â !r8       âO !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       È� !r       â !r       â !r
	
       È/ ‰       ( ‰
      b !r        â !r 
       â/ G	  à     € ê !r  1       Ê	x   à@  � Ê  x #ÛÙ?  A  È  x ;ª¸¿  A  â !r 2       Ê s        " 	x    à@  � È 	x    àÀ  � Ê !t   €?    â !t  €?    È s	        $  r	     A  È  r     A  Ê †y     ä Ay        € ê ‰       b Ey       € æ ‰      ¨ „y/    H   è „y/    H   ( „y/    H   è „y/  	  H   â !r         Æ/ „y/    H   â !r       ÆO ‰       h ‰      " !r5       â� !r7       â!r        ä/ !r       â !r       Ä ‰ €    " !r       Æ ‰ €    b !r       Æ „y/    H   ¢ !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       âO !r       â x0   ÿàÿ â !r
	
       Æ/ ‰       " z  r  p`ð Æ ‰
      b x /   pVp  â !r        â !r 
       Ö/ G	  à     € ê !r  1       Ê	x   à@  � Ê  x #ÛÙ?  A  È  x ;ª¸¿  A  â !r 2       Ê s        " 	x    à@  � È 	x    àÀ  � Ê !t   €?    â !t  €?    È s	        $  r	     A  È  r     A  Ê †y    ä Ay        € ê ‰       b Ey  ð    € æ ‰      ¨ „y/ €  H   è „y/ €	  H   ( „y/ €  H   è „y/ €  H   â !r         Æ/ „y/ €  H   â !r       ÆO ‰       h ‰      ( „y/ €  H   â !r6       È� !r       â!r        Æ/ !r       â !r       â ‰ €    ( ‰ €    b !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       â x0   ÿàÿ â !r
	
       Æ/ ‰       " z  r  p`ð Æ ‰
      b x /   pVp  â !r        â !r 
       Ö/ G	  à     € ê !r  1       Ê	x   à@  � Ê  x #ÛÙ?  A  È  x ;ª¸¿  A  â !r 2       Ê s        " 	x    à@  � È 	x    àÀ  � Ê !t   €?    â !t  €?    È s	        $  r	     A  È  r     A  Ê †y    ä Ay        € ê ‰      b !r4       â x00   ÿàÿ Æ !r       â z 0 r  p`ð Æ !r       â x /   pVp  È ‰       ¢ !r       Ê/ ‰      " !r         ÊO ‰       b !r       Ê ‰ €    " !r        Ê/ ‰ €    b !r	       Ê ‰
	 @    " !r       Ê/ ‰ @    b !r
	
       Ê ‰
      " !r       Ê/ ‰       ¢ !r
       â M	        € ì !r11       â/!r        ÈO 	x11  à@  � â !r22       È  x1#ÛÙ?  A  â 	x22  à@  � Æ  x;ª¸¿  A  â 	x22  àÀ  � Ê s        " !t22  €?    â !t  €?    Ì s         $  r11      A  È  r112     A  Ê †y 1   â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      "   $ z  w  p`ð Ú M	        € ê y      !   " ¹z  l      â Ey  p    € â ™x?    â y      %   b ¹z  F   
   â $rÿÿ   ÿ Ž ä $rÿÿ   ÿ Ž â x#ÿ     È | #   p`ð â $x   ÿ Ž Ø/ G	  Ð    € ê y      &   " $tÿ   ÿ Ž ä $z x  Ž È %v \   Ž Ê �y     ™ ¢  $rÿÿ   ÿ Ž â x   ÿÀŽ â $z r  ÿŽ ä 'z m   Ž È $v%ÿ X  ÿ Ž â zÿ n    â $x#   ÿ Ž ä $vÿ Y  ÿ Ž ä $rÿÿ   
Ž ä $rÿÿ   ÿ Ž ä $z o  Ž Ä $x   ÿŽ ä $z u  ÿŽ È %x$   ÿ Ž Ì %x$    Ž â x   ÿÀŽ Ê %x$   Ž Ì %x$   Ž È $z q  #Ž âz Z  ÿàñ ä x   ÿàÿ ä x   ÿàó â $z  t  #Ž äO$z  t  Ž â zÿ [  þ  â $z q  Ž Ä $x    ÿ Ž ä$x   ÿ Ž ä %x8   ÿŽ â x    ÿÀŽ ä x   ÿÀŽ â %x8   ÿŽ â r   ÿàó ä r   ÿàñ Ä r   ÿàõ â $r$ÿÿ   	Ž  âr   ÿà÷ â $r"ÿÿ   	  â x    øŽ â $r&ÿÿ    â x!   øŽ â $rÿÿ   Ž ä r%   ÿàñ är%   ÿàó Æ $x   "  ä$x   $Ž  Æ �y    ™ ¨ �y2!   ‘ è �y6    ‘ " r
%   ÿàó är%   ÿàñ Æ $x   Ž  ä$x   &  Æ �y
    ‘ h �y    ™ h �y+
!   ‘ b x ÿ     äO x(2ð   ÿÀŽ â�$x2   ÿ Ž â x ÿ   ÿÀŽ â $x6   ÿ Ž âx(ÿ   (  Ä xÿ   ÿÀŽ ä x   âŽ â $x(   ÿ Ž â x   âŽ ä x-6ð   ÿÀŽ ä x ÿ     â %v
   Ž â/ x   ÿÀŽ Ä x-ÿ   -  â %v   Ž â x ÿ   ÿÀŽ ä x-   ÿ   â �y 
    ™ ¢  r*   ÿ@� Æ �y.    ™ â x   âŽ â $r
ÿÿ    Ž ä $rÿÿ    Ž ä %v*   Ž Æ �y'
ôÿÿ ™ " %v   Ž Æ �y*
   ™ ( �y/
üÿÿ ™ ( �y)
øÿÿ ™ ( �y5
   ™ ( �y3
   ™ " x7ð   ÿÀŽ Æ�y4
    ™ b $x   ÿ Ž Æ �y,    ™ b xÿ   ÿÀŽ Æ �y1
   ™ b x7ÿ   7  Æ �y0    ™ b x   âŽ â $x7   ÿ Ž â r%    ÿàñ â/ �yA
ðÿÿ • b xÿ     Æ $x	     â x:+ð   ÿÀŽ äxÿ   ÿÀŽ â %v   Ž â xÿ     â �y80   • b x   âŽ â $x+   ÿ Ž â xÿ   ÿÀŽ â �y9    ™ b x:ÿ   :  Ä x   âŽ Æ $x:   ÿ Ž ä %v   Ž Æ x   ÿÀŽ â %v   Ž â �y;    ™ d r   ÿ@� Æ �y    ™ d %v   Ž â r%   ÿàñ ä/ r%!   ÿàó Æ �y    ™ b $x	     ä$x   Ž  âr%   ÿàñ ä �y<2   • h �y0   • b $x     Æ �y>4   • h �y=2   • h �y4   • h �y0   • h �y?0   • h �y@6   • h �yB6   • b x66   ÿÀŽ â $tÿ   =ÿ Ž â/ x--   ÿÀŽ È Er- -       ä x++   ÿÀŽ È Er+ +       ä x77   ÿÀŽ Æ #t++  `¿  â Er7 7       ä x::   ÿÀŽ Æ #t77  `¿  â x##   ÿàÿ È | #   p`ð ä x%%à   ÿàó Ê $rÿÿ   Ž  â xD ÿÀŽ äO x ÿ      ä xC.ÿÀŽ ä� x.ÿ   .  ä x  ÿÀŽ ä x..ÿÀŽ â &rDD'   ÿ   Ä&rCC*   ÿ   ä &z/  € ÿ   ä &r  )   D   ä &r..5   C   â xC2   ÿÀŽ â &zD3  € ÿ   Æ ErC C       â &z4  €    âx,ÿÀŽ â &z1  € D   Æ Er2        â &r/       â x0ÿÀŽ â &z '  € ÿ   â Er 6       ä Er6        â #tC  `¿  â &r.3   .   â xC(   ÿÀŽ â &z*  € ÿ   â x,ÿ   ,  â &z )  €     â x(ÿ   0  â &z5  €    â x,,ÿÀŽ Ä x((ÿÀŽ â #t  `¿  â Er          â &r,4      â Er        â &r((1   .   â ErC C       â #r    ÿ   â #t0-  `¿  â r8ÿ   8  â #r   ÿ   â Er,        â #tDC  `¿  â Er (       â $t-ÿ   ÿ Ž â #r02      â x(ÿ   8  â #rDD6      Ä !r,       â x((   -ìŽ ä x,9ÿÀŽ ä x-   ÿÀŽ ä x9ÿ   9  â !rD       â Er (       â &r,,'   ÿ   â x(9ÿÀŽ Ä x'ÿÀŽ â  r     A  â xÿ     ä x;ÿÀŽ â &r(()   ,   â x;ÿ   ;  â &r''*   ÿ   â xÿÀŽ â #r++   ÿ   â Er- -       â &r/   (   â x;;ÿÀŽ Ä xÿÀŽ â &r5   '   â xÿ     â #t--  `¿  â &r;4      â Er' :       â &r3      â xÿÀŽ â #r- -   ÿ   â rÿ     â #t''  `¿  â Er        â &r1      â #r 27   -   â $t)ÿ   ÿ Ž Ä $x   ÿ Ž â xÿ     â !r         â x<ÿ   <  ä x=ÿ   =  â $x8   ÿ Ž â Er        â #r'6'   +   â $tÿ   ÿ Ž â x.        Ä x   )ìŽ ä x<<ð   ÿÀŽ ä x=ð   ÿÀŽ ä x>ÿ   >  ä xÿ     â !r'       â x   .ìŽ Ä x   ìŽ ä Er'        ä r<   ÿ � ä r?   ÿ � ä x@@ ð  ÿÀŽ ä xBB ð  ÿÀŽ Ä x>>   ÿÀŽ ä x   ÿÀŽ â  r''     A  â Er        ä Er        ä r@   >àÿ ä rB   àÿ â #r      â #r'    '   â 0r ÿA    A   Ä 0rÿ    A   ä 0rÿ    A   â x
€  ÿàõ ä  r      A  ä  r      A  â $rÿÿ    â #r      ä #r'      â G‰   ñÿÿÿÿƒ ê Ay        € ê ‰       " $x	   Ž Æ ‰      d z 	 p  p`ð È x    pDp  â !r         â !r       È/ ‰       ( ‰      b !r        â !r       È/ ‰ €    ( ‰ €    b !r       â !r       È/ ‰ @    ( ‰ @    b !r       â !r       È/ ‰       h  ‰      ¢  M	        € ê y      &   â $x   ÿ Ž â !r       â_ $z	 s  	Ž â !r        â/ $tÿ   ÿ Ž â x    p"ò Ä r ÿ   p"ð Ö $ÿÿ    Ž ä $z v  	Ž È� %v j   Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      "   $ z  w  p`ð Ú M	        € ê y      &   " $tÿ   ÿ Ž â ¹z  F   
   â $z x  Ž È %v \   Ž Ê �y    ™ b ¹z  l      â Ey  €    € â ™x?    â y      !   b $rÿÿ   ÿ Ž â x      ÿ  â $rÿÿ   ÿ Ž ä $rÿÿ   ÿ Ž ä $rÿÿ   ÿ Ž â x ÿ     È/ |     p`ð Ú G	  Ð    € ê $r
ÿÿ   ÿ Ž â Ey �    € â $x   ÿ Ž ä 'z m  
 Ž Æ x   ÿÀŽ â $vÿ a  ÿ Ž â zÿ n    â $rÿÿ   ÿ Ž È $r
ÿÿ   
Ž ä y      %   $ $z

 o  Ž ä $z r  ÿŽ â x   ÿÀŽ â $z
 u  ÿŽ È %x$   ÿ Ž Ì %x
$    Ž È $r	ÿÿ     Ž ä %x
$   
Ž È $x	   ÿ Ž ä $x    ÿ Ž ä %x
$   
Ž Æ x    ÿàÿ â $z t  	Ž âz
 Z  ÿàñ â $x   ÿ Ž ä $z  q  Ž â x   ÿàó â $v
ÿ Y  ÿ Ž â x   ÿÀŽ â %x  8   ÿŽ â zÿ [  þ  Ä x   ÿÀŽ â $z  q  	Ž â z  X      Æ $z" t  Ž ä $vÿ `  ÿ Ž ä %x""8   ÿŽ È $x   ÿŽ â r"   ÿàñ är"   ÿàó är    ÿàõ â $rÿÿ   #  âr    ÿà÷ â $rÿÿ   #Ž  â x"   øŽ â $rÿÿ   ! â x    øŽ â $rÿÿ   !Ž Ä r&   ÿàñ â�y3   ™ ¢ r$   ÿàó Æ $x'
     â�y.   ™ â $x%
   Ž  Æ �y&&    ™ ( �y $    ‘ h �y$!   ‘ ¢  r(   ÿàñ är   ÿàó â �y4üÿÿ ™ ä $x)
     âr
   ÿäÿ  â �y2   ™ è �y    ‘ è �yA(    ™ è  �y5!   ‘ è �y0   ™ è �y1    ™ â xÿ   &  Äx&ÿ   ÿÀŽ â $x'    ÿ Ž âx$ÿ   ÿÀŽ ä x ð   ÿÀŽ ä x/ð   ÿÀŽ âO$x%   ÿ Ž â x6'   âŽ ä x/ÿ   /  Ä xÿ     â %v66   Ž â xÿ   &  ä x$%   $âŽ â $x/   ÿ Ž â/ xÿ   ÿÀŽ â $x%   ÿ Ž â �y66    ™ ¤  x   ÿÀŽ Ä x*%   âŽ â %v$$   Ž â r(&   ÿ@� â �yôÿÿ ™ $ %v**   Ž ä�y$    ™ d %v((   Ž ä �yøÿÿ ™ h �y:*    ™ h �y>(    ™ b r$   ÿàó Ê/ $x%#   
Ž  âx'Aÿ   ÿÀŽ â� $x*   ÿ Ž âx9ð   ÿÀŽ ä r&"   ÿàñ â �y8$0   • â  x@5ð   ÿÀŽ ä x(*   'âŽ ä x9ÿ   9  â $x'#   
  â x*ÿ   A  Ä x$ÿ   A  â $x%5   ÿ Ž â x@ÿ   @  â %v((   Ž â x$$ÿ   ÿÀŽ â �y7&2   • â  x**ÿ   ÿÀŽ â $x+9   ÿ Ž â x$%   $âŽ â $x%@   ÿ Ž â �y=(    ™ ä x*+   *âŽ Ä �y<&4   • â  %v,$   Ž Æ �y;&0   • â  x(%   ÿÀŽ â/ %v**   Ž ä �y?&6   • â  r(A(   ÿ@� Æ �y,,    ™ ä %v((   Ž ä �y**    ™ â r&    ÿàñ ä r$   ÿàó â �y((    ™ ä  $x'!   
  âr%!
   ÿäÿ  â �y-ðÿÿ • è. �y+&2   • è �y$$0   • è  �yA&4   • è �yB&0   • è �yC&6   • â x   ÿÀŽ Ä x55   ÿÀŽ ä x   ÿÀŽ ä x99   ÿÀŽ ä x@@   ÿÀŽ È Er@ @       ä z ÿ `  pPð È z ÿ a   Sð ä xE6ÿÀŽ äO x6ÿ   6  È x)6ÿÀŽ â &rEE   ÿ   âx6ÿÀŽ äxDÿ     â &r%)   E   Æ x)DÿÀŽ â &r663   ÿ   â x:ÿÀŽ ä x'>ÿÀŽ â/ &r)).   6   â x>ÿ   >  â &r4   %   â x:ÿ   :  â &r&'2   )   â x%>ÿÀŽ Ä x::ÿÀŽ â &z64  € ÿ   ä &r&%0   &   ä &z%  € ÿ   â x>    ÿÀŽ â &r):1      â xD   ÿÀŽ â &z'3  € ÿ   ä &z:  € %   â Er> >       â &z%1  € 6   Ä $tÿ   =ÿ Ž â Er  :       â &zE2  € ÿ   â ErD D       â &z'.  € '   â Er %       â #t%>  `¿  â x:/   ÿÀŽ â &z60  € E   â Er>        ä Er '       â #t'D  `¿  âEr: :       â #r%%    ÿ   â #t>>  `¿  âEr/ 6       â #r6'   ÿ   â #t':  `¿  â Er: )       â #r%>   %   â Er) &       â #r&'/   6   â r8ÿ   8  â� !r6%:       â $t%ÿ   ÿ Ž Ä x:ÿ   8  ä x7ÿ   7  ä x%:   %ìŽ â !r&&)       â x)=ÿÀŽ â $x>8   ÿ Ž â Er% %       ä x=ÿ   =  Ä x87ð   ÿÀŽ ä x<ÿ   <  â  r%&%     A  â x7=ÿÀŽ â &r))   ÿ   â r8;8   ÿ � ä x?? ð  ÿÀŽ ä x<<   ÿÀŽ Ä Er& 5       â &r77   )   â r8?8   <àÿ ä x<,ÿÀŽ â #t5&  `¿  â x)*ÿÀŽ ä x,ÿ   ,  ä Er:        Ä x*ÿ   *  â #r&5   ÿ   â x5,ÿÀŽ â &r)4   7   â x**ÿÀŽ â &r<<3   ÿ   â #t):  `¿  â Er: 9       ä x,(ÿÀŽ â &r55.   <   â x(ÿ   (  â &r*1      â x*ÿ   +  â #r) )   ÿ   â #t::  `¿  â x+(ÿÀŽ â &r,,2   5   â r$ÿ   $  â #r):   )   â Er:        â $t9ÿ   ÿ Ž â xÿ   $  â &r++0   ,   â #t@@  `¿  â x5*ð   ÿÀŽ Ä xAÿ   A  â $t'ÿ   ÿ Ž â !r7):       â x(   9ìŽ â #r&/@   &   â Er+ +       â $t)ÿ   ÿ Ž â r5B5   ÿ � â $x$$   ÿ Ž â xCC ð  ÿÀŽ ä xA   ÿÀŽ â !r+&+       â x'>   'ìŽ Ä r5C5   àÿ ä Er( (       ä x&$   )ìŽ â 0rÿ-    A   â Er$ '       â 0r'ÿ8    A   â Er& &       â 0r5ÿ5    A   â  r(+(     A  â #r66$   %   â  r8'     A  Ä #r77&   (   â  r55     A  â G‰  @    € ê r&   ÿàõ är$   ÿàó Æ $x'    ä$x%   Ž  Æ �y9&     ¨  �y$$     è �y:&!    $ r&   ÿàõ Ê $x'#    Ê �y&&0    b $x)9   ÿ Ž âO x<ÿ   9  ä x($ÿ   ÿÀŽ ä� x*ÿ   $  ä x()   (âŽ ä x+ÿ   $  Ä x**ÿ   ÿÀŽ ä x%<   ÿ   â %v((   Ž âx=ÿ   :  äx++ÿ   ÿÀŽ â $x,:   ÿ Ž â x*%   *âŽ â $x%=   ÿ Ž â �y;(    ™ ¤  x,,   +âŽ â %v**   Ž â x%%   ÿÀŽ Æ %v,,   Ž ä �y**    ™ â r($%   ÿ@� Æ �y,,    ™ $ %v((   Ž â r$"   ÿàó Ê �y((    ™ " $x%#   Ž  Ê �y>$3    ( �y+$4    (# �y?$0    ( �y-$6    " x99   ÿÀŽ Ä x@:   ÿÀŽ ä Er: 9       ä x<<   ÿÀŽ ä Er@ @       â #t):  `¿  âEr< <       ä x==   ÿÀŽ â #t%@  `¿  â/ #r) )   ÿ   â #t$<  `¿  â Er: =       â #r<%   ÿ   Ä #r$$   )   ä #t):  `¿  â r&ÿ   &  äx';ÿÀŽ äO x;ÿ   ;  Æ &r''   ÿ   â x;;ÿÀŽ ä x%*ÿÀŽ ä� x*ÿ   *  â &r';   '   â x:,ÿÀŽ äx,ÿ   ,  â &r'%4   '   â x**ÿÀŽ â #r%/)   <   â x),ÿÀŽ â &r::3   ÿ   Ä &r'*1   '   â x*(ÿÀŽ â &r)).   :   â x(ÿ   (  Æ &r)*2   )   â x9(ÿÀŽ â $t;ÿ   ÿ Ž â x>>ð   ÿÀŽ ä x+ÿ   +  â &r)90   )   â x(ÿ   &  â $t9ÿ   ÿ Ž â r>?>   ÿ � â $x&&   ÿ Ž â x-- ð  ÿÀŽ Ä x++   ÿÀŽ ä x((   ;ìŽ ä Er* )       ä r+->   +àÿ ä Er' '       â !r%%*       â x&&   9ìŽ Ä Er( (       â 0r+ÿ+    A   â Er& &       â !r'$'       ä  r(%(     A  â  r$+     A  Æ #r&'&   (   È #r$&      â #r86      â #r57      â G‰  0    € ê r&   ÿàó är$   ÿàñ Æ $x'   Ž  ä$x%     Æ �y5&     ¨ �y$$     è �y7&!    " $x*5   ÿ Ž âO x6ÿ   5  Ä x)$ÿ   ÿÀŽ ä� x(ÿ   $  äx**   )âŽ â $x,6   ÿ Ž â x%ÿ   $  ä x+(ÿ   ÿÀŽ â %v(*   Ž â x%%ÿ   ÿÀŽ Æ $x*7   ÿ Ž âx8ÿ   7  â �y9(    ™ ¢  x,,   +âŽ ä x**   %âŽ ä x%8   ÿ   â %v,,   Ž Æ x%%   ÿÀŽ â %v**   Ž ä �y,,    ™ â r&$%   ÿ@� Æ �y**    ™ b	 r$   ÿàñ â %v&&   Ž È $x%!     âr(    ÿàñ â �y&&    ™ h $x)!     â �y$$0    h �y:(3    h �y-(4    h/ �y;(0    h �y+(6    "x55   ÿÀŽ È Er< 5       Ê #t5<  `¿  È #r5 5   ÿ   â x66   ÿÀŽ ä x77   ÿÀŽ ä x88   ÿÀŽ ä Er6 6       ä Er8 8       Æ #t66  `¿  ä#t88  `¿  ä #r6   5   â x'9ÿÀŽ äO x9ÿ   9  Æ &r''   ÿ   â x99ÿÀŽ ä x ,ÿÀŽ Æ� &r'9   '   â x,ÿ   ,  ä x*ÿÀŽ äx*ÿ   *  â &r  4   '   â x,,ÿÀŽ â &r3   ÿ   â x%*ÿÀŽ ä Er 7       â &r1,1       â x &ÿÀŽ â &r%%.      â x&ÿ   &  â #t'  `¿  â r$ÿ   $  â &r  2   %   â x&ÿÀŽ â $t%ÿ   ÿ Ž â x::ð   ÿÀŽ â #r'   ÿ   â xÿ   $  â &r 0       â x-ÿ   -  â $x$$   ÿ Ž â x   %ìŽ Ä r:;:   ÿ � ä x--   ÿÀŽ ä x++ ð  ÿÀŽ â#r/8      â Er         ä r++:   -àÿ ä x$$   %ìŽ â !r       â Er1 1       Ä Er%        â 0r+ÿ+    A   â Er $       â !r1       ä  r%%     A  â  r++     A  Æ #r   %   È #r+      â x		   ÿàÿ ä xà   ÿàó ä | 	   p`ð ä xà   ÿàõ â $r
ÿÿ   
Ž  â x€  ÿà÷ Æ $rÿÿ    ä $rÿÿ   Ž È D	       À â Gy  `äÿÿÿÿƒ ê Ay       € ê $rÿÿ    Ž ä $rÿÿ    Ž ä Ay        € ê z ÿ `  pPð È z ÿ a   Sð Ú G	  P    € ê ‰       ( ‰      b !r         â !r       È/ ‰       ( ‰	      b !r        â !r		       È/ ‰ €    ( ‰
	 €    b !r       â !r
	
       È/ ‰ @    ( ‰
 @    b !r       â !r
       È/ ‰       ( ‰      b !r        â !r       â/ Gy  €    € ê ‰       ( ‰      h ‰	      ¨ ‰      â !r         â !r       â/ !r
	       ÆO ‰	       " !r       Æ� ‰      h ‰
      ¨ ‰      â !r	 	       â !r       È/ ‰	 €    " !r
       ÆO ‰ €    b !r       Æ� ‰ €    ¨ ‰ €    â !r	       â !r       â/ !r       ÈO ‰ @    " !r       Æ� ‰ @    h ‰ @    ¨ ‰	 @    â !r       â !r       È/ ‰      " !r       ÆO ‰
      b !r		       Æ� ‰       ¨ ‰	      â !r       â !r
       â/ !r        âO !r	       â� y      %   ( y      !   $ $x   Ž Ê z  p  p`ð È x    pDp  Ú M	        € ê z ÿ ^  pPð â $z v  Ž Æz ÿ _   Sð Ú % ^   Ž Ì �	     ¢ $x    ÿ Ž Ê r  ÿ   p"ò äx     p"ö äx     p"ø äx     p"ô Î          ä z ÿ `  pPò â $2 ÿÿ    Ž ä $B ÿÿ    Ž â z ÿ a  Sò â $" ÿÿ    Ž È !         ÐO G™      € ê z ÿ b  pPð È z ÿ c   Sð Ú % b   Ž Ì �	     ¢ $x   ÿ Ž Ê x øÿÿÿp"ò äx üÿÿÿp"ô är ÿ   p"ö äx    p"ø Î $ÿÿ    Ž ä $"ÿÿ    Ž ä $vÿ h  ÿ Ž ä $2ÿÿ    Ž ä $Bÿÿ    Ž â x    pBò Æ !       ÔO G  P    € ê x    pRð Ú G‰  �     € ê x    pRð Ú G	  ð    € ê  x;ª¸¿  A  Ì s        $ !t  €?    Ì s        $  r	     A  È  r  	     A  â Gy  P    € ê  x'7=  A  â x*BL?  A  Æ #t  €?   â  x   ?  A  Æ  r     A  Ì s     $   $ !t  €?    È  r     A  È  r       A  â Gy  °    € ê x    pRð Ú G‰  à     € ê x    pRð Ú G	        € ê 	z i    � ä 	z   i    � Æ  x;ª¸¿  A  â 	z   i €  � Ê s        $ !t  €?    Ì s        $  r	     A  È  r  	     A  â Gy  Ð     € ê  r       A  â Gy  °     € ê 	x  à@  � ä 	x    à@  � Æ  x#ÛÙ?  A  â 	x    àÀ  � Æ  x;ª¸¿  A  ä !t	   €?    È s        $ !t  €?    Ì s        $  r      A  È  r  	     A  ä y      "   ( y      &   b $z s  Ž È $z v  Ž È/ %v j   Ž Ê †y      â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y       "   " ¹z  l      â Ey  °    € â ™x?    â y      !   b ¹z  F   
   â r ÿ       â x2      ÿ  â r  ÿ       â x      ÿ  â r ÿ       â x      ÿ  â r+ ÿ       â x      ÿ  â r- ÿ       Ä r% ÿ       ä r ÿ       ä r1 ÿ       ä r  ÿ   pRð är  ÿ   pRò är!    ÿ(Ž È/ xÿ   !  È r!   ÿ� È x'ÿ     È | '   pbô Ú G)  0    € ê yU      '   " rT ÿ       ä rV ÿ       â y      %   b xøÿÿÿÿÀŽ ä z	  q      â yW      &   ¢ r!!  €ÿàÿ ä x0'   ÿ   Ä x!   ÿ   ä r"	0   ÿŽ â$x$	   0Ž âr&	0   ÿŽ â$x(	   0Ž âz  X      â $x*	   0Ž âz  Y      â $x,	   0Ž â r ÿ       â 'zU y  T Ž â x#ÿ   !  Ä z.0 q  ÿàÿ ä x   ÿ   ä/ zÿ z    â 'zW s  V Ž ÈO $z p  'Ž â zÿ t    â $z |  'Ž ä$z |  Ž â x   ÿàÿ â $z v  Ž ä$z v  Ž ä $z p  Ž Ä %x2   ÿŽ È %x2   ÿŽ È %x   Ž È %x
!   Ž È %x   Ž È %x!   Ž È r   ÿàõ är
   ÿà÷ ä r	   ÿä är   ÿäÿ ä r4   ÿàõ â �y6   • ¨ �y/"   • â r5   ÿä Ä r   ÿàõ â �y<   • ( �y)4"   • b r   ÿä Ê �y9   • ( �y=   • " $zwW w  ÿŽ â x?        â $z:U }  ÿŽ Ê rw:w   ÿàõ È rvÿÿ   ÿä ä r4!w   0àµ ä/ xu $       Ê %v44 Z  u Ž â x7ÿ   6  äO x/   ÿ   ä�x8/   ÿ   ä x7   øŽ ä x8   ÿÀŽ Æ %v  ? Ž âx)   ÿ   äxÿ   6øŽ ä x   ÿÀŽ â �yq    ™ ¤  %v  ? Ž â r#v   ÿd~ ä xÿ   9øŽ Äx9ÿ   9  â $x7$   ÿŽ â x8)   ÿ   â %v  ? Ž â �ys    ™ â x6/   ÿ   ä x89   8øŽ â �yt    ™ b	 x)   ÿ   ä x6   ÿÀŽ Ä r57   ÿàÿ ä/ r 4       â %v88  ? Ž â x   ÿÀŽ Æ �yr   ™ ¢  xÿ   <øŽ äx:ÿ   =øŽ â �yo8    ™ ¨ �yp   ™ ¢  %v  ? Ž â x<ÿ   <  Ä �y5   ™ ¢  %v::  ? Ž â xÿ   /  ärD!w   .àµ â �y4    ™ ¢ x=ÿ   =  ä xÿ   )  â �y7   ™ ¢  x<   øŽ Ä r<#v   ÿd~ â �y6:    ™ ¢  x=   øŽ â %v  ? Ž ä�y8   ™ $# %vDD Z  u Ž ä �y9   ™ ¤ $x=<$   ÿŽ Ä %v  ? Ž â �y:    ™ $ rEE=   ÿàÿ ä �y;    ™ ( �y<   ™ ( �y=D   ™ " rB   ÿàõ Æ �yQD   ™ " rC   ÿä ä r!w   "àÅ â �yXD   ™ " rF   ÿà÷ ä r#v   ÿ„~ â �ySD   ™ " rG   ÿäÿ â %v Z  u Ž Ä �yRD   ™ ( �y>    ™ ( �yO    ™ ( �yPD   ™ ( �y@B    • " $x$   ÿŽ Æ/ �yAF    • " r!w   $àµ ä r   ÿàÿ â �yND   ™ " r#v   ÿd~ Æ �yM   ™ ( �yLD    ™ ( �yID    ™ " %v Z  u Ž Æ �y?   ™ " $x$   ÿŽ Æ �yK   ™ $ r   ÿàÿ ä �yJ   ™ ( �yC   ™ ( �yH   ™ ( �yG   ™ "# r!w   &àµ Æ �yD   ™ ( �yE   ™ " rZ#v   ÿd~ Æ �yF   ™ ( �y\   ™ " %v Z  u Ž Æ �yY    ™ " $x]Z$   ÿŽ Æ �yB    ™ ( �y[   ™ " rb!w   (àµ Æ �yZ   ™ " r]   ÿàÿ ä ra#v   ÿd~ â �y_   ™ " %vb Z  u Ž Æ �y^   ™ " $xaa$   ÿŽ Æ �yd    ™ ( �y]    ™ ( �y`   ™ " ra   ÿàÿ Æ �yc   ™ ( �yb   ™ ( �yg   ™ ( �ya   ™ " r!w   *àµ Æ/ �yj   ™ ( �ye   ™ " rl#v   ÿd~ Æ �yh   ™ " rz!w   ,àµ Æ �yi   ™ ( �yn   ™ " r{#v   ÿd~ Æ �ym    ™ " $xyl$   ÿŽ Æ �yf    ™ ( �yk   ™ " %v Z  u Ž Æ �yl   ™ " $x{{$   ÿŽ ä %vz Z  u Ž â �yx   ™ " ry   ÿàÿ Æ r{   ÿàÿ â �yy    ™ ( �yw    ™ ( �yv   ™ ( �yz   ™ ( �y   ™ (# �y€   ™ ( �y{   ™ ( �y~   ™ ( �yu   ™ ( �y|   ™ ( �y�   ™ ( �y†   ™ ( �y   ™ ( �y…   ™ ( �y}   ™ ( �yƒ    ™ ( �y„   ™ ( �y    ™ ( �y‡    ™ ( �y‚    ™ " x‹        Ä xˆÿ   )  ä xŠ        ä x‰ÿ   /  ä xˆ   ‹ìŽ ä xsÿÀŽ ä� xsÿ   s  Ä x‹ÿ   t  äxttÿÀŽ ä x‰‰   ŠìŽ ä xˆqÿÀŽ äO xŠÿ   q  ä xqsÿÀŽ â &rsr   ÿ   âx‹ÿÀŽ â/ &r‹tr   ÿ   â xoÿÀŽ â &rrqp   s   âxŒÿ   o  â &r‹p   ‹   â xoŠÿÀŽ â &rrˆ5   r   âxsŒÿÀŽ â &rŠ5   ‹   â x54ÿÀŽ â &rro7   r   âx4ÿ   4  â &r7s7   Š   â xp6ÿÀŽ Ä xŠÿ   6  ä x6/ €  ÿÀŽ ä x/4ÿÀŽ â &r7p8   7   â	x4ŠÿÀŽ â &rr58   r   â x)) €  ÿÀŽ Æ &rŠ49   7   âx7:ÿÀŽ â &rr/9   r   â x‹    7    ä Er6 6       ä Er8 )       ä x);ÿÀŽ â &r97<   r   â#t66  `¿‹  â &rr=   ÿ   â #t88  `¿‹  â &r‹)<   Š   Ä &rŠt=   ÿ   ä &r=qQ   r   âx:ÿ   :  â &rQQ   Š   ä &rŒˆX   =   âx<ÿ   ;  â &rXX   Q   â x::ÿÀŽ â &rŒoS   Œ   âx<<ÿÀŽ â &rSsS   X   Ä &rŒ5R   Œ   ä &r9:>   9   ä0r;ÿO  0 A   ä&r‹<>   ‹   ä &rŒ/P   Œ   â Err 9       â 0r>ÿ@    A   â Er@        â &rSpR   S   â  rŠ6;     A  â 0rOÿO    A   Ä 0rAÿA    A   ä &r97N   Œ   ä &rS4P   S   â #rrOr   Š   â  rA@     A  â  rŠ;8     A  â &r@M   ÿ   ä &r9:L   9   Ä &rS)N   S   ä 0r;ÿI  0 A   ä&rNtM   ÿ   â Er‰ ‰       â &r=q?   @   â Er@ 9       â &rS<L   S   â  rL6;     A  â 0rIÿI    A   Ä &r??   N   â  r>>‰     A  â &r=ˆK   =   â#r@I@   L   â &r?K   ?   ä &r=oJ   =   â #r>@      â &r?sJ   ?   ä &r@C   ÿ   Ä &r=5H   =   ä&r?pH   ?   ä &r@qG   @   ä &rHtC   ÿ   ä &r=/D   =   ä &r@ˆE   @   Ä &rHG   H   ä &r=7F   =   ä &r?4D   ?   ä &rDo\   @   â ErN S       â &rHE   H   â  r;8;     A  â &r=:Y   =   Ä 0r9ÿB  0 A   ä&rD5[   D   ä &rHs\   H   â #rNIN   ;   â &r?)F   ?   â Er= =       â 0rBÿB    A   â  r;69     A  â &rD/Z   D   Ä &r[p[   H   â  r@89     A  â &r?<Y   ?   â #r;B=   ;   â &r97_   D   ä &rZ4Z   [   â Er? ?       â &r=^   ÿ   â #r>;      â &r9:d   9   Ä &r_)_   Z   ä 0r;ÿ]  0 A   â#r?B?   @   â &r=q`   =   â Er@ 9       â &r_<d   _   â  rB6;     A  â 0r]ÿ]    A   ä &r^t^   ÿ   Ä &r=ˆc   =   â ErD _       â &r^`   ^   â #r@]@   B   â  r;8;     A  â &rBob   =   ä &r9g   ÿ   â #r;]D   ;   â &r^c   ^   ä &rB5a   B   Ä &r9qj   9   â #r;      â &r^sb   ^   ä &rB/e   B   ä &r;ˆh   9   â #r%>@   %   â &r^pa   ^   ä &rB7i   B   Ä &r@on   ;   ä &r^4e   ^   ä &rB:m   B   ä 0r9ÿf  0 A   ä&rA5k   @   ä &r^)i   ^   â Er; B       â 0rfÿf    A   â  r=69     A  â &rA/l   A   Ä &r^<m   ^   â #r;f;   =   â &rA7x   A   â #r?      â  r@89     A  â Er? ^       â &rA:y   A   â #r2>;   2   â 0r9ÿw  0 A   ä &r;v   ÿ   â #r?f?   @   â &rgtg   ÿ   Ä &rvtv   ÿ   ä &r   ÿ   âEr@ A       â &rt   ÿ   â  rB69     A  â 0rwÿw    A   ä &rgj   g   ä &r;qz   ;   Ä &rvz   v   ä &rqq€      ä&r€      â #r@w@   B   â &rgh   g   ä &rBˆ{   ;   ä&rv{   v   Ä &rqˆ~   q   ä&r~      ä &rgsn   g   ä &rBou   B   ä&rvsu   v   ä &rqo|   q   Ä &rss|      ä &rB5�   B   ä&rgpk   g   ä&rvp�   v   ä&r5†   q   ä&rsp†   s   Ä &rB/   B   ä&rg4l   g   ä&rv4   v   ä&r/…      ä&rs4…   s   â x''   ÿàÿ â &rB7}   B   Ä 0rÿƒ  0 A   ä &rg)x   g   ä&rv)}   v   ä &r77„      ä&rs)„   s   â | '   pbô â &rg<y   g   â  r6     A  â &rB:   B   â  r8     A  â &rv<   v   Ä &r7:‡   7   ä&rs<‡   s   ä 0rÿ‚  0 A   âErX ‹       â  r989     A  â ErD g       â 0rƒÿƒ    A   â ErB B       â 0r‚ÿ‚    A   â Erv v       â  r6     A  â Er7 7       â  r88     A  â Ers s       â #rXOX   Š   â #rDwD   9   â #rƒB      â#rƒv      â #r‚7      â#r8‚s   8   â x�  ÿà÷ â #r1>r   1   â#rX      â#rN      â#r?      â#r3>@   3   â#r-D   -   â#r+>   +   â#r      â#r >       â #r8      â rÿ   ÿäÿ Ä x""@   ÿàÿ ä x$$@   ÿàÿ ä x&&@   ÿàÿ ä x((@   ÿàÿ ä x**@   ÿàÿ ä x,,@   ÿàÿ Ä x..@   ÿàÿ ä x00@   ÿàÿ â D)       À â Gy  @äÿÿÿÿƒ ê Ay        € ê Ey  0    € â G‰      € ê r     ÿHŽ Ê ˆs  1 øÿ H   è ˆs  €øÿ H   è ˆs   ùÿ H   è ˆs  €ùÿ H   è ˆs   úÿ H   è ˆs  €úÿ H   è ˆs  % ûÿ H   è ˆs  €ûÿ H   è ˆs  2 üÿ H   è ˆs  €üÿ H   è ˆs  3 ýÿ H   è ˆs  -€ýÿ H   è ˆs  + þÿ H   è ˆs  €þÿ H   è ˆs    ÿÿ H   è ˆs  €ÿÿ H   ä Ay        € ê {          ì M        € ê „y      H   " Ey  @    € æ „y €   H   h „y    H   ¢ !r  1       â !r       È/ ‰       " !r       ÆO ‰      h ‰      ¢ !r        Æ y       %   " !r       Æ/ ‰      b !r       ÆO ‰      ¨ ‰      â x    ÿ   â !r       Æ/ z  r  p`ð â !r       ÆO r ÿ   pVp  â ‰ €    " !r	       Æ� ‰ €    h ‰
	 €    ¨ „y €  H   â !r       â x   ÿàÿ â !r       Æ/ ‰  @    " z  r  p`ò â !r
	
       äO ‰ @    b x    p"ò ä ˆ	        â ‰      &   ¨ ‰
 @    ì       &   "         Æ       '   b !r         â !r       È/ ‰       " $Š x  Ž ÆO ‰      '   b !r
       Æ� ‰      ¨ ‰      â $ x  Ž È$ ~  Ž ä$Š ~  Ž â/ !r        Æ %† j  	 Ž â !r	       ÆO % j   Ž â †‰     â !r       Æ� † 	   â G	  �     € ê y       &   h y      '   " $z   x  Ž È/ $z  ~   Ž Ê z   r  ÿàõ È rÿÿ   ÿä ä z  j  ÿ„ È z  k   Ê †y     è Ay        € ê !r       â „y    H   â Ey  �    € æ ‰      ( „y	 €  H   è „y    H   â !r       Ê ‰       $ !r         Ê ‰  €    $ !r        Ê ‰ @    $ !r       Ê ‰      $ !r       â G™  °     € ê y       &   " ¹z  r      ä �x   ?àÿ â y      '   b $z   x  Ž È $z  ~   Ž Ê/ |     ÿàõ È rÿÿ   ÿä ä z  j  ÿ„ È z  k   Ê †y     è Ay        € ê „y €  H   b !r		       â !r       â !r%       â „y    H   ¢ z#  r      Æ „y €  H   â r/##   ÿŽ är!##   ÿŽ â „y    H   ( „y €  H   ( „y
    H   è „y €  H   è „y    H   è ‰ 	      h ‰      h ‰      b !r       Æ/ „y €  H   â !r2       âO !r       È� ‰      b !r33       Æ ‰      " !r--       â!r	        ä‰3      ¢ !r+
+       â !r       ä ‰
      â !r       â !r       Ä ‰-      " !r        Æ ‰      h ‰+      b !r       Æ/ ‰      b !r       Æ ‰      " !r33       âO !r       ä ‰      ¢ !r

       Æ� ‰      â !r--       Æy       &   " !r       Æy	      '   b !r++       â !r       ä/ ‰
      b !r       Æ ‰      " !r       ÆO ‰      ¢ !r       Æ� ‰3      â $z   x  Ž Æ‰-      " $z 	 ~   Ž Æ‰ €    h ‰+      b r#    ÿ„ â !r
       â/ ‚ /   ÿà÷ â ‰      b rÿÿ   ÿä â !r       â z j  ÿ„ â ‰      " !r       âO ‚ÿÿ   ÿäÿ Ä ‰       ¢ !r33       â� z k   ä ‰ €    â !r--       âŠ
 j  ÿ† ä‰	 €    " !r       âŠ k  � â !r++       â ‰ €    b ‚ !   ÿàù â !r       Ä/ ‰ €    b ‚*ÿÿ   ÿä â !r       ä ‰ €    " Š j  ÿˆ â !r        äO ‰3 €    ¢ Š k  * â !r       Ä� ‰ - €    â $x#   ÿŽ â!r	       ä‰ @    " r	#    ÿŠ är    ÿàõ â ‰"+ €    " !r       ârÿÿ   ÿäÿ Ä ‰$ €    b rÿÿ   ÿä â !r       â/ z j  ÿ„ â ‰& €    b !r       â z	 j  ÿŠ ä ‰(      " !r3       âO z k   Ä ‰ @    ¢ !r -        â� z		 k  � ä ‰ @    â !r       â!r"+"       ä ‰ @    " r+##  €ÿŽ â !r$$       Ä‰ @    b ‚ +   ÿà÷ â !r&&       ä/ ‰ @    b !r((       Æ ‰ @    " !r       ÆO ‰  @    ¢ !   ÿàÿ â !r       ä� ‰#" @    â     ÿàõ ä ‚ÿÿ   ÿäÿ â ‰%$ @    â Š j  ÿ† â !r       â*ÿÿ   ÿä â ‰)( €    "  j  ÿ„ â !r       â!/   ÿàÿ â ‰'& @    b Š k  � â !r       â/  k  * â ‰      b ! !   ÿàõ â !r       Ä ‰      " *ÿÿ   ÿä â !r         âO ! j  ÿ„ â‰      ¢ !r#"#       ä� ! k  * â !r%$%       â ‰      â !r)()       Æ‰      " !r'&'       Æ‰      b !r!       Æ/ ‰      b !r       Æ ‰       " !r       ÆO ‰#      ¨ ‰%      ¢ !r       Æ� ‰$) @    â !r       Æ‰"'      " !r       â!r       ä/ †‰ !    â !r        Æ †    â !r##       ÆO †‰     â !r%%       Æ †     â !r$)$       Æ� †‰     â !r''"       Æ†    è †‰ 
    è †     è †‰ #    è † %   è ‰$      ( †‰ '    â M™        € ê x+   ÿàÿ â/ !r$       Æ r    ÿàñ È r ÿÿ   ÿä  ä z j  ÿ€ È z k     Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      "   " ¹z  l      â Ey       € â ™x?    â y       !   b ¹z  F   
   â x"      ÿ  â r$ ÿ       â x      ÿ  â r3 ÿ       â x      ÿ  â x      ÿ  â x      ÿ  â r ÿ       Ä r ÿ   pRð är ÿ   pRò är)    ÿ(Ž È/ x ÿ   )  È r )   ÿ� ä r  ÿ       ä x!ÿ     È | !   pbô Ú G)  °    € ê y      '   " r ÿ       ä xøÿÿÿÿÀŽ â y      %   b z/  q      ä r))  €ÿàÿ â y      &   ¢ x2!   ÿ   ä x)   ÿ   Ä r'/2   ÿŽ â$x,/   2Ž âr*/2   ÿŽ â$x./   2Ž âz-  X      â $x//   2Ž â z+  Y      ä r" ÿ       ä x1ÿ   )  â 'z y   Ž â z02 q  ÿàÿ Ä x   ÿ   â/ $z	 }  ÿŽ â r ÿ       â $z( w  ÿŽ ÈO 'z s   Ž â zÿ z    ä r(	(   ÿàõ â $z p  !Ž âzÿ t    â $z |  !Ž â x   ÿàÿ â $z |  Ž â r4ÿÿ   ÿä â $z v  Ž Ä $z v  Ž ä $z p  Ž ä %x2   ÿŽ È %x2   ÿŽ È %x   Ž È %x
)   Ž È %x   Ž È %x)   Ž È r
-   ÿàõ är-   ÿà÷ ä r+   ÿä är+   ÿäÿ ä r-   ÿàõ â�y9"   • ¢ r-   ÿà÷ Ä r	+   ÿä â�y7"   • â  r+   ÿäÿ Æ �y%   • ( �y8   • h �y:   • h �y&   • b r;)(   2àµ Ä x>        ä xl $       ä r14   ÿd~ Æ %v; Z  l Ž È/ $x$   ÿŽ Ê r   ÿàÿ Ê �yE   ™ h �yI   ™ h �yG   ™ h �yA   ™ h �yC   ™ b x59   ÿ   ÄO x7   ÿ   ä� x5   ÿÀŽ ä x   ÿÀŽ ä xÿ   %øŽ äxÿ   8øŽ äx%ÿ   %  â %v  > Ž â x9   ÿ   Ä x8ÿ   8  â %vJ  > Ž âx7   ÿ   â �y6    ™ ¢ x%   øŽ ä x8   øŽ â �yJJ    ™ " x57   ÿ   ä x%9   ÿ   â %v  > Ž âx55   ÿÀŽ â �y8    ™ b x%%   ÿÀŽ â %v  > Ž â x55ÿ   :øŽ â �yH    ™ b x%ÿ   &øŽ Æ/ �yF    ™ b x&ÿ   &  â %v<5  > Ž âx:ÿ   :  â �y5   ™ b xÿ   9  â %v  > Ž ä �yB<    ™ b xÿ   7  ä/ x&   øŽ â �yD    ™ b x:   øŽ Ä r-   ÿàõ â %v  > Ž â�y:   ™ h �y?    ™ b %v  > Ž â/ r+   ÿä ä rN-   ÿàõ â �y<    ™ h �y;    ™ b rO+   ÿä Ä rM)(   0àµ â�y=    • b rK)(   'àÙ ä� r14   ÿd~ â�y%N    • b r&14   ÿ¤~ â %vM Z  l Ž È/ $xM$   ÿŽ ä %vK Z  l Ž Æ rM   ÿàÿ â $x&$   ÿŽ È �yM   ™ b r   ÿàÿ Æ �yL   ™ h �yK   ™ h �y&   ™ h �y>   ™ h �y@   ™ h �yV   ™ h �yU   ™ h �yT   ™ h �yS   ™ h �yR   ™ b r)(   ,àµ Ä r)(   *àÙ â �yQ   ™ h �yP   ™ b r14   ÿd~ ärZ14   ÿ¤~ â �yO   ™ è‚ �yN    ™ è  �yX    ™ â  %v Z  l Ž Æ �yW    ™ â $x[$   ÿŽ ä %v Z  l Ž â �yY    ™ æ $x]Z$   ÿŽ â r[   ÿàÿ ä r        Ä r]   ÿàÿ ä r        â �yZ   ™ è  �y\   ™ è  �y]   ™ è  �y[   ™ è  �yh   ™ è  �yg   ™ è  �yf   ™ è  �ye   ™ è  �yd   ™ è  �yc   ™ è  �yb   ™ â  r)(   .àµ Æ/ �ya   ™ â r)(   /àÙ Æ �y`   ™ â  r14   ÿd~ Æ �y_   ™ â rm14   ÿ¤~ Æ �y^    ™ è  �yj    ™ è  �yi    ™ â $xo$   ÿŽ Æ �yk    ™ â %v Z  l Ž È %v Z  l Ž È $xmm$   ÿŽ â ro   ÿàÿ ä r        ä rm   ÿàÿ Æ �y   ™ è  �y   ™ è
 �yw   ™ è  �yv   ™ è
 �yu   ™ è  �yt   ™ è
 �ys   ™ è  �yr   ™ è
 �yq   ™ è  �yp   ™ è
 �yo   ™ è  �yn   ™ è
 �ym   ™ è  �yl   ™ è
 �y    ™ è  �y    ™ è
 �yx    ™ è  �yy    ™ â
 xz6ÿÀŽ ÄO x|ÿ   6  ä x6ÿ   J  äxJJÿÀŽ â &r{zE   ÿ   âx||ÿÀŽ ä x66ÿÀŽ â &r}JE   ÿ   â xHÿÀŽ â&rE|I   {   âxÿ   H  â &r{6I   }   â xÿ   F  Ä xFÿÀŽ â &rHG   E   âxI        ä xFÿ   9  ä xÿÀŽ â &r{G   {   â xÿÀŽ ä xEF   IìŽ â &rFA   H   âxIÿ   B  â &rHA   {   â xGDÿÀŽ Ä x{ÿ   D  ä xBBÿÀŽ ä xA        ä xDÿ   7  â &rFGC   F   âx{{ÿÀŽ â &rHBC   H   â xIIÿÀŽ Ä xAD   AìŽ ä xC?ÿÀŽ ä xDÿ   ?  ä x?ÿ   ;  â &rF{:   F   âx;;ÿÀŽ â &rHI:   H   â x99 €  ÿÀŽ Ä x77 €  ÿÀŽ â &rFC5   F   âx}DÿÀŽ â &rH;5   H   â x??ÿÀŽ ä x:    7    ä Er9 9       ä Er7 7       â &rF}8   F   Ä &rH?8   H   â #t99  `¿:  â0r8ÿ<  0 A   â#t77  `¿:  â 0r=ÿ=    A   â Er: E       â 0r%ÿ%    A   â  r598     A  â  rD87     A  â 0r<ÿ<    A   â ErF F       Ä Er8 A       â  r==:     A  ä #r5<F   5   ä  r8%8     A  â &r%zM   ÿ   ä&rMJM   ÿ   â #r=5      â &r5zK   ÿ   Ä &r%|L   %   ä&rM6L   M   ä &rKJK   ÿ   ä &r5|&   5   ä &r%>   %   ä&rM>   M   Ä &rK6&   K   ä &r5@   5   ä &r%V   %   ä&rMV   M   ä &rK@   K   ä &r&U   5   Ä &r%GT   %   ä&rMBT   M   ä &rKU   K   â ErE H       â &r&GS   &   ä &r%{R   %   ä&rMIR   M   Ä &r:BS   K   ä &r&{Q   &   ä &r%CP   %   â#rD<E   D   â &rM;P   M   ä &r<IQ   :   ä &r:CO   &   Ä� &r%}N   %   ä&rM?N   M   ä 0r&ÿX  0 A   âEr% %       â &r<;O   <   â ErM M       â 0rXÿX    A   â  r59&     A  â r&7&     A  â &r:}W   :   Ä &rW?W   <   â #r%X%   5   â0r<ÿY  0 A   â#r&XM   &   â 0rYÿY    A   â Er: :       â #r=%      â  r>9<     A  â #r8&      â &r%zZ   ÿ   ä&rZJZ   ÿ   Ä &r&z\   ÿ   â #r5Y:   >   â &r%|]   %   ä&rZ6]   Z   ä &r&|[   &   â #r=5      â &r%h   %   ä&r5h   Z   Ä &r&g   &   ä &r%f   %   ä&r5f   5   ä &r&e   &   ä &r%Gd   %   ä&r5Bd   5   Ä &r&Gc   &   ä &r%{b   %   ä&r5Ib   5   ä &r&{a   &   ä &r%C`   %   ä&r5;`   5   Ä &r:C_   &   â Er@ W       â &r%}^   %   â  rA7<     A  â &r^?^   5   ä 0r&ÿj  0 A   ä&r:}i   :   ä 0r<ÿk  0 A   â#r@Y@   A   â 0rjÿj    A   â Er% %       â 0rkÿk    A   â  r59&     A  â ErA ^       â  r&7&     A  â Er: :       â  r>9<     A  â #r%j%   5   â#r&jA   &   Ä #r5k:   >   â #r=%      â#r8&      â &r\J\   ÿ   â #r=5      â &r%z   ÿ   ä&r&J   ÿ   ä &r5z   ÿ   ä&rJ   ÿ   Ä &r\6[   \   ä &r|w   %   ä&r&6w   &   ä &r5|v   5   ä&r6v      ä &r\g   \   Ä &ru      ä&r&u   &   ä &r5t   5   ä&rt      ä &r\e   \   ä &rs      Ä &r&s   &   ä &r5r   5   ä&rr      ä &r\Bc   \   ä&rGq      ä&r&Bq   &   Ä &r5Gp   5   ä&rBp      ä &r\Ia   \   ä &r{o      ä&r&Io   &   ä &r5{n   5   Ä &rIn      â x!!   ÿàÿ â &r\;_   \   ä &rCm      ä&r&;m   &   ä &r5Cl   5   ä&r;l      â | !   pbô â &r\?i   \   Ä &r}      ä&r&?   &   ä &r5}   5   ä0rÿx  0 A   ä&r??      ä 0rÿy  0 A   âEr        â  rE7<     A  âEr &       â 0rxÿx    A   â Er\ \       â 0ryÿy    A   â  r7     A  âEr 5       â  r9     A  âEr& ?       â  r69     A  â r77     A  â #rEk\   E   â #rx      â#rx      â #ry   6   â#r7y&   7   â x--�  ÿà÷ â #r8D      â#r8@      â#r 8E       â#r3=   3   â#r$8   $   â#r#=   #   â #r"87   "   â x''@   ÿàÿ Ä x,,@   ÿàÿ ä x**@   ÿàÿ ä x..@   ÿàÿ ä x//@   ÿàÿ ä x00@   ÿàÿ ä x22@   ÿàÿ Ä r+ÿ+   ÿäÿ â D)       À â Gy  ðæÿÿÿÿƒ ê Ay        € ê Ey      € â G‰  ð     € ê $xÀ   Ž Ê ˆs  ùÿ H   è ˆs €ùÿ H   è ˆs  úÿ H   è ˆs €úÿ H   è ˆs  ûÿ H   è ˆs €ûÿ H   è ˆs  üÿ H   è ˆs €üÿ H   è ˆs  ýÿ H   è ˆs  €ýÿ H   è ˆs 3 þÿ H   è ˆs $€þÿ H   è ˆs # ÿÿ H   è ˆs "€ÿÿ H   ä Ay        € ê {          ì M        € ê „y      H   b Ey  @    € æ „y  €   H   ¨ „y     H   â !r       â/ !r       È_ ‰      " !r       Æ� ‰      h ‰      ¢ !r       Æ y      %   " !r       Æ/ ‰      b !r	       ÆO ‰      ¨ ‰	      â x   ÿ   â !r       Æ/ x   ÿàÿ äz  r  p`ð â !r       âO ‰ €    " z  r  p`ò â !r
	       â� r  ÿ   pVp  â‰ €    b x     p"ò Æ ‰
 €    ´       &   â         Æ       '   " !r       Æ ‰      &   " !r       Æ/ ‰ @    b !r
       ÆO ‰	 @    ¨ ‰ @    h ‰      '   b $ x  Ž Æ� „y
  €  H   â $ ~  Ž â!r       â/ $Š x  Ž â ˆ        â !r		       äO ‰      " !r       Æ‰	      b $Š ~  Ž Æ ‰      ¢ !r       â %† j   Ž â !r	       Æ/ % j   Ž â †‰     â !r       ÆO †    â G	  �     € ê y      &   ( y      '   b $z x  Ž È $z ~  Ž Ê/ z r  ÿàõ È rÿÿ   ÿä ä z j  ÿ„ È z k   Ê †y     è Ay        € ê !r

       â „y     H   â Ey  à    € æ ‰
      ( „y  €  H   è „y	     H   è „y  €  H   è „y     H   è „y  €  H   â !r
       Æ „y  €  H   è ‰      ( „y
     H   â !r       Ê ‰ €    $ !r       Ê ‰ @    $ !r       Ê ‰      $ !r       â G™  °     € ê y      &   " ¹z  r      ä �x   ?àÿ â y      '   b $z x  Ž È $z ~  Ž Ê/ |   ÿàõ È rÿÿ   ÿä ä z j  ÿ„ È z k   Ê †y     è Ay        € ê „y     H   â !r       â !r        â !r3
3       â „y  €  H   â !r       â !r		       â !r       â ‰      b !r       â !r$$       â z!  r      â ‰      ¨ ‰3      è ‰      ( ‰	      ( ‰      h ‰       b !r       â/ !r       ÄO ‰$      b !r#       â !r3       ä� y      &   ¢ !r""       â !r       ä ‰      " !r	       Æ‰      â !r
       Æ‰      " !r         Æ ‰"      b !r$       Æ/ ‰      h ‰      b $z x  Ž ÆO y      '   ¢ !r       Æ ‰       " !r       Æ� ‰	
      â !r       Æ‰      " !r""       Æ‰      b !r       Æ/ ‰ €    b !r       Æ ‰      b $z# ~  Ž ÆO ‰"      ¢ !r         â r!#   ÿ„ ä ‰ €    " !r
	       â� r
ÿÿ   ÿä ä ‰ €    â !r       âz j  ÿ„ â !r       â‰  €    $ z k  
 â !r       â/ ‰	 €    b !r       Æ ‰ €    b !r""       ÆO ‰ €    ¢ !r
       Æ ‰ €    " r!#   ÿˆ â !r       ä� ‰ €    â r!!   ÿŽ ä r ÿÿ   ÿä â ‰" €    â !r         Æ‰
 @    " !r	       â/ ‚	#   ÿàõ â !r       âz j  ÿˆ â ‰  @    b ‚ÿÿ   ÿä â !r       âO Š	 j  ÿ„ â ‰ @    ¢ z k    â !r       â $x !   ÿŽ â ‰ @    " Š		 k   â !r       â� r!!   ÿŽ â ‰ @    â !r"       Ä ‚#   ÿà÷ â ‰ @    b !r
       âr #   ÿàù ä ‰ @    " 
   ÿàÿ ä ‚$ÿÿ   ÿäÿ â ‰ @    " !r         â/    ÿàÿ Ä ‰ @    b !r       âO Š j  ÿ† ä‰  @    ¢ !r       â Š k  $� â !r       â� ‰       " !#
   ÿà÷ â !r       Ä‰      â rÿÿ   ÿä â !r       ä‰      " z
 j  ÿˆ â !r       ä ‰      b $ÿÿ   ÿäÿ â !r       Ä/ ‰      b ! j  ÿ† â !r         äO ‰      ¢ "#   ÿàõ ä z k   â ‰      ¢ ! k  $� â !r%        â #ÿÿ   ÿä â ‰      " !r       â� " j  ÿ„ Ä ‰      â !r       â" k  # ä ‰!       " !r       â!r       ä/ †‰ %    â !r       ÆO †    â !r       Æ †‰     â !r       Æ †     â !r       Æ� †‰     â !r! !       Æ†    è †‰     è †     è †‰ 
    â M™        € ê †y 
!   â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      "   " ¹z  l      â Ey  0    € â ™x?    â y      !   b ¹z  F   
   â r" ÿ       â x      ÿ  â y!      %   ¢ r ÿ       â x      ÿ  â r/ ÿ       â x      ÿ  â r1 ÿ       Ä r ÿ       ä r ÿ   pRð är ÿ   pRò är#   ÿ(Ž È/ x ÿ   #  ä x!!   ÿ   äO r #   ÿ� ä r  ÿ       ä xÿ     È |    pbô Ú G)  À    € ê y      '   " r ÿ       ä r ÿ       â y      &   b xøÿÿÿÿÀŽ ä z)  q      ä r##  €ÿàÿ ä x,   ÿ   Ä x#   ÿ   ä r),   ÿŽ â$x&)   ,Ž âr(),   ÿŽ â$x))   ,Ž â z'  X      ä z%  Y      ä r" ÿ       â 'z y   Ž â x+ÿ   #  Ä z*, q  ÿàÿ â 'z s   Ž â/ zÿ z    Æ $z! p  Ž â zÿ t    â $z	 }  ÿŽ ä $z |  Ž ä$z |  Ž â x!   ÿàÿ â $z v  Ž ä$z v  Ž Ä $z p  Ž ä $z$ w  ÿŽ ä %x2   ÿŽ Æ r$	$   ÿàõ â %x2   ÿŽ Æ r.ÿÿ   ÿä â %x   Ž È %x
#   Ž È %x   Ž È %x#   Ž È rB
'   ÿàõ är'   ÿà÷ ä rC%   ÿä är%   ÿäÿ ä r'   ÿà÷ â�yBB"   • ¢ r'   ÿàõ Æ �y3"   • â  r%   ÿäÿ är	%   ÿä Æ �y2   • ( �y    • h �y-   • h �y4   • b r#$   ,àµ Ä x_ $       ä xA        ä x0B   ÿ   äO x63   ÿ   ä� x50   ÿÀŽ ä x6   ÿÀŽ Ä r0+.   ÿd~ ä xLÿ   2øŽ â%v Z  _ Ž â x55ÿ    øŽ Æ$x0$   ÿŽ â/ x ÿ      â %v5  A Ž Æ r   ÿàÿ â %vLL  A Ž â xB   ÿ   â�yK    ™ ¢  x2ÿ   2  ä x53   ÿ   â �yLL    ™ â xB   ÿ   ä x    øŽ â �yI   ™ b	 x22   5øŽ Ä x   ÿÀŽ â �y5   ™ b	 %v>  A Ž âx 3   ÿ   ä xÿ   -øŽ â%v2  A Ž âx7    ÿÀŽ â �y>>    ™ b r2#$   *àµ â %v  A Ž â x77ÿ   4øŽ â �y<    ™ b x-ÿ   -  Ä x ÿ   B  â �y=   ™ b	 x0ÿ   4  ä r8+.   ÿd~ â %v67  A Ž â �y;    ™ b xÿ   3  Æ �y9   ™ b	 x00   øŽ â %v2 Z  _ Ž ä �y46    ™ b x-    øŽ â/ $x-8$   ÿŽ ä �y    ™ f	 r-   ÿàÿ â %v  A Ž â rP'   ÿàõ â �y2   ™ h	 �y6   ™ b %v@0  A Ž Æ �y7   ™ h �yE   ™ b rQ%   ÿä Æ �y0@    ™ h �y8    ™ b rN'   ÿàõ Æ �yH   ™ h �yF   ™ b	 r#$   àÇ Æ �yG   ™ b rO%   ÿä ä r@+.   ÿ„þ â/ �y?P    • h �yC    ™ h	 �y:    ™ b	 %v Z  _ Ž Æ �yD   ™ b $xM@$   ÿŽ Æ �y-N    • b r#$   &àµ Æ�yA   ™ b	 rM   ÿàÿ ä rM+.   ÿd~ â �yJ    ™ h	 �y@    ™ b	 %v Z  _ Ž Æ �yO   ™ b $xQM$   ÿŽ Æ/ �yM   ™ d rQ   ÿàÿ ä �yP   ™ h �yS   ™ h �yN   ™ h �yX   ™ h �yQ   ™ h �yV   ™ h �yW   ™ h �yT   ™ h �yU   ™ b r#$   (àµ Æ�y]   ™ ( �yZ    ™ " rY+.   ÿd~ Æ �yR    ™ ( �y\   ™ " r`#$   )àµ Æ �y[   ™ " %v Z  _ Ž â ra+.   ÿd~ ä �y^    ™ " $xcY$   ÿŽ Æ �yY    ™ " %v` Z  _ Ž â rc   ÿàÿ Æ $xaa$   ÿŽ ä �y`   ™ & ra   ÿàÿ â �y_   ™ ( �ya   ™ ( �y   ™ (# �yh   ™ ( �yi   ™ ( �yf   ™ ( �yg   ™ ( �yd   ™ ( �ye   ™ ( �yb   ™ ( �yc   ™ ( �ym   ™ ( �y    ™ ( �yj    ™ ( �yl   ™ ( �yk    ™ ( �yo    ™ " xt        Ä xqÿ   B  È xq   tìŽ ä/ xqÿ   3  ä xnKÿÀŽ äO xpÿ   K  ä xrÿ   L  ä� xKLÿÀŽ Ä xLpÿÀŽ ä xrÿÀŽ â &rn5   ÿ   ä
&r5K5   ÿ   ä &rpLI      ä&rII   5   â x>ÿÀŽ Ä x5ÿ   >  ä xÿ   <  ä x<<ÿÀŽ ä x55ÿÀŽ â &rp=   p   âxr        â &r><=   I   â xÿÀŽ Ä x=q   rìŽ â &rI59   p   âxp3 €  ÿÀŽ â &rq9   >   â x>;ÿÀŽ ä x9ÿ   4  ä x44ÿÀŽ ä xs    7    Ä Erp p       â &rI>    I   âxBB €  ÿÀŽ â &rq4    q   â x3ÿ   ;  â #t p  `¿s  â &rpn6   ÿ   â ErB B       â &rrK6   ÿ   â x33ÿÀŽ â &rpL7   p   â x99ÿÀŽ â #t;B  `¿s  â &rr7   r   Ä &rpE   p   ä&rI32   I   âxB0ÿÀŽ â &rq92   q   â x7ÿ   0  ä x28ÿÀŽ ä x8ÿ   8  â &rp5H   p   â x77ÿÀŽ â &rr<E   r   Ä &rqBF   q   ä&r62F   I   â xI8ÿÀŽ â &rp>G   p   â Er0        â &rrH   r   ä 0r?ÿ?    A   ä 0r8ÿC  0 A   Ä &rq7:   q   Æ  rF;8     A  â &r6I:   6   ä &rp3D   p   â  r?0     A  â  r:8      A  â 0rCÿC    A   â Er8 q       â &rr4G   r   â Er0 =       â 0r-ÿ-    A   Ä &rp2A   p   âEr6 6       â &rr9D   r   â #r8C8   :   â  r0-0     A  â &rpIJ   p   ä 0r-ÿ@  0 A   â#r6C6   F   â &rrBA   r   â #r08      â 0r@ÿ@    A   â Er= p       â &r8nO   ÿ   â  r?;-     A  â &rOKO   ÿ   â #r6      â &rr7J   r   â  r6 -     A  â &r-LM   8   â #r=@=   ?   â &rOM   O   â ErA r       â &r-P   -   â#r=      â &rO<P   O   Ä &r=nS   ÿ   â #r6@A   6   â &rSKS   ÿ   ä &r-5N   -   ä&rON   O   â #r06      â &rSX   S   ä &r6>Q   -   Ä &r84Q   O   ä &r=LX   =   ä &rS<V   S   ä&r63W   6   ä&r89W   8   ä &r=V   =   Ä &rST   S   ä&r62U   6   ä&r8BU   8   ä &r=5T   =   ä &rS4]   S   ä&r6IZ   6   Ä &r87Z   8   ä &r:>]   =   ä 0r-ÿR  0 A   ä&rS9\   S   â Er= 6       â &r:3\   :   â ErA 8       â 0rRÿR    A   â  r6 -     A  â &rSB[   S   â  r?;-     A  â &r:2[   :   â #r6RA   6   â 0r-ÿY  0 A   Ä &rS7^   S   ä&r:I^   :   â #r=R=   ?   â #r06      â 0rYÿY    A   â  r6;-     A  â Er8 S       â  r- -     A  â Er: :       â #r=      â &r=n`   ÿ   â #r-Y8   -   Ä #r6Y:   6   â &r=La   =   â #r 0-       â &r`K`   ÿ   â #r6      â &r-n   ÿ   ä&rKK   ÿ   ä &r6_   =   ä &r`a   `   Ä &r-Lh   -   ä&rKh   K   ä &r65i   6   ä &r`<_   `   ä &r-f   -   ä&rK<f   K   Ä &r6>g   6   ä &r`i   `   ä &r-5d   -   ä&rKd   K   ä &r63e   6   ä &r`4g   `   Ä &r>>b   -   ä&r44b   K   ä &r62c   6   ä &r`9e   `   ä &r33m   >   ä&r99m   4   â x   ÿàÿ â &r6I   6   Ä &r`Bc   `   ä 0rÿj  0 A   ä&r22l   3   ä&rBBl   9   â Er 6       â 0rjÿj    A   â |    pbô â  r-;     A  â &r`7   `   Ä &r2Ik   2   ä&rB7k   B   ä 0rÿo  0 A   â#rj   -   â 0roÿo    A   â  r      A  âEr `       â  r;     A  â Er2 2       â  r-      A  â ErB B       â #rj      â x''�  ÿà÷ â #ro2      Ä #r-oB   -   â #r1   1   â#r/0   /   â#r      â #r"0-   "   â x@   ÿàÿ ä x&&@   ÿàÿ ä x((@   ÿàÿ ä x))@   ÿàÿ Ä x**@   ÿàÿ ä x,,@   ÿàÿ ä r%ÿ%   ÿäÿ â D)       À â Gy  °éÿÿÿÿƒ ê Ay        € ê Ey  ð     € â G‰  Ð     € ê $x€  Ž Ê ˆs  úÿ H   è ˆs €úÿ H   è ˆs  ûÿ H   è ˆs €ûÿ H   è ˆs  üÿ H   è ˆs €üÿ H   è ˆs  ýÿ H   è ˆs  €ýÿ H   è ˆs 1 þÿ H   è ˆs /€þÿ H   è ˆs  ÿÿ H   è ˆs "€ÿÿ H   ä Ay        € ê {          ì M        € ê „y €   H   b z ! r  p`ð Ä z+  r      â „y     H   ¢ r ÿ   p"p ä r'++   ÿŽ â„y    H   â r)++   ÿŽ Æ „y    H   ( „y €  H   h „y €  H   ( „y €  H   ( „y	    H   ( „y €  H   è „y
    H   è „y    H   â !r       â? !r       ÈO ‰      " !r       Æ� ‰      b !r       â!r         È‰      ¢ !r       Æ ‰      â !r       Æ ‰       " !r	       Æ ‰      b !r       â !r/       ä „y €  H   â !r       â/ !r

1       ä ‰	      " !r       Æ ‰      b !r       ÆO ‰      ¢ !r       Æ� ‰      â !r        Æ‰
      " !r       âx!   ÿàÿ ä ‰      d z  r  p`ò ä ‰      b !r		       â x    p"ò Ä y       &   " !r       â/ !r       äO ‰      b !r       Æ� y      '   ¢ !r""       â !r
       ä‰ €    â !r        Æ‰      " !r       Æ ‰	      b $z   x  !Ž Æ ‰      ( ‰      " !r       Æ/ ‰      b $z  ~   Ž ÆO ‰      ¢ !r       â� 
  r  ÿàù ä ‰      â !r       âr+    ÿ† ä ‰ €    " !r	       ârÿÿ   ÿäÿ Ä ‰#       b !r       â ÿÿ   ÿä ä ‰%"      " !r       â 
 j  ÿˆ â !r!       â/ ‰ €    d 
 k   â !r       âO ‰ €    ¢ z j  ÿ† â !r       Ä� ‰	 @    â z k  � â !r
       ä‰ €    "  '   ÿàõ â !r# #       ä‰ €    b ÿÿ   ÿä â !r%"%       Ä ‰ €    " 
 j  ÿ„ Æ ‰"! €    " !r       â/ 
 k   ä ‰ €    b !r       âO  )   ÿàù ä ‰  €    ¢ !r	       â� 	+   ÿàÿ Ä ‰
 @    â !r       â* 	   ÿàõ ä ‰$# €    " !r       âÿÿ   ÿä ä ‰&%      b !r       â 
 j  ÿˆ â !r"!"       â ‰ @    $ 
	 k   â !r       â/ ‰ @    b !r         ÆO ‰      ¢ !r
       Æ� ‰ @    â r+    ÿ† â !r$#$       ä‰ @    " +ÿÿ   ÿä â !r&%&       ä‰ @    b * j  ÿ„ ä r(ÿÿ   ÿäÿ â ‰!" @    b !r       â * k  + Ä ‰ @    " !r       â/         ä ‰  @    b !r+       âO '   ÿàÿ ä ‰      ¢ !r       â� z
 j  ÿ† Ä ‰#$ @    â !r       âz k  (� ä ‰%& €    " !r       â!r""!       â !        â ‰      b !r       Æ ‰      " !r         Æ/ ‰      b !r-       âO     ÿàõ ä‰      ¢ %  j   Ž â !r$$#       â� (ÿÿ   ÿä â ‰      â %  j  ! Ž â !r&&%       â j  ÿ„ â‰      &  k  ( â ‰!       " !r       Æ‰#"      b !r       Æ ‰'& @    " !r       Æ/ ‰%$      b !r       ÆO †	 +    â !r       Æ� † -   â !r       Æ†	     â !r! !       â !r#"#       ä†     â !r'&'       Æ †	     â !r%$%       Æ/ †    è †	     è †     è †	 
!    è † 
#   è ‰'      ( †	 %    â M™        € ê x)   ÿàÿ âO !r''       Æ r    ÿàñ È r ÿÿ   ÿä  ä z j  ÿ€ È z k     Ê †y '    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      "   " ¹z  l      â Ey  �    € â ™x?    â y      !   b ¹z  F   
   â r ÿ       â x      ÿ  â y      %   ¢ r ÿ       â x      ÿ  â r) ÿ       Ä r+ ÿ       ä r ÿ       ä r ÿ   pRð är ÿ   pRò är!   ÿ(Ž È/ x ÿ   !  ä x   ÿ   äO r !   ÿ� ä r  ÿ       ä xÿ     È |    pbô Ú G)  0    € ê y      '   " r ÿ       â $z p  Ž â r ÿ       â y      &   b xøÿÿÿÿÀŽ ä z%  q      ä r!!  €ÿàÿ ä x(   ÿ   Ä x!   ÿ   ä r"%(   ÿŽ â$x$%   (Ž âz  X      ä z#  Y      ä r ÿ       ä r%%(   ÿŽ â 'z y   Ž â x'ÿ   !  Ä z&( q  ÿàÿ â 'z s   Ž â/ zÿ z    Æ $z  }  ÿŽ â zÿ t    â $z |  Ž ä$z |  Ž â x   ÿàÿ â $z v  Ž ä$z w  ÿŽ ä $z v  Ž Ä $z p  Ž â r     ÿàõ â %x2   ÿŽ Æ r*ÿÿ   ÿä â %x2   ÿŽ È %x   Ž È %x
!   Ž È %x   Ž È %x!   Ž È r   ÿàõ är8
   ÿà÷ ä r	#   ÿä ä r,   ÿàõ ä r9#   ÿäÿ â�y.   • ¢  r-#   ÿä Æ �y88"   • â r   ÿàõ Æ �y7,"   • " r#   ÿä Æ �y0   • h �y1   • h �y3   • b r!    (àµ ä xS $       Ä r4'*   ÿd~ Æ %v Z  S Ž È $x-4$   ÿŽ Ê/ r-   ÿàÿ ä x5        ä x8   ÿ   ä� x,7   ÿ   äx   ÿÀŽ ä x/ÿ   .  ÄO x28   ÿ   ä x,,   ÿÀŽ ä xÿ   .øŽ ä r        ä x//   2øŽ ä x,ÿ   1øŽ â%v@  5 Ž â x1ÿ   1  â �y;   ™ ¢  x.7   ÿ   Ä x,8   ÿ   â %v/  5 Ž âx.1   .øŽ â �y@@    ™ ä %v  5 Ž â x-,   ÿÀŽ â �y9    ™ b	 x/7   ÿ   Æ �yI    ™ ¢  x2/   ÿÀŽ Æ �yE   ™ ¢  x-ÿ   0øŽ â%v,.  5 Ž â x0ÿ   0  â �y>   ™ ¢ xÿ   8  Æ �y?,    ™ ¢ x00   øŽ â %v  5 Ž âx.2ÿ   3øŽ â�yG   ™ ¢ x3ÿ   3  ä xÿ   7  â �y=    ™ ¢  %v..  5 Ž äx13   øŽ â �yA   ™ ¢( %v0  5 Ž Æ �yC.    ™ ¨ �y:    ™ ¨  �yD   ™ ¢ r2!    &àÇ Æ �yF   ™ ¢ %v01  5 Ž â r-'*   ÿ„þ ä r   ÿàõ â�y,    ™ ¢ r.   ÿà÷ â/ %v2 Z  S Ž ä �y<0    ™ ¤  $x--$   ÿŽ â r#   ÿä â�yB    ™ ¢ r/#   ÿäÿ Ä r-   ÿàÿ â �yH    • ¢ r0!    "àµ Æ �yK   ™ ¢  r-'*   ÿd~ Æ �yL   ™ ¢  r!    $àµ â/ %v0 Z  S Ž ä�y6   ™ " rM'*   ÿd~ â $x--$   ÿŽ â rT!    %àµ â �yJ.    • & rU'*   ÿd~ â �y5   ™ " r-   ÿàÿ â %v Z  S Ž Ä �y-   ™ $ $xWM$   ÿŽ ä �y3   ™ " $xUU$   ÿŽ Æ �y1   ™ ( �y.    ™ (! �yN    ™ " rW   ÿàÿ Æ �y/   ™ ( �y4   ™ " %vT Z  S Ž Æ �y2   ™ $ rU   ÿàÿ ä �y0   ™ ( �yQ   ™ ( �yP   ™ ( �yO   ™ ( �yM    ™ ( �yR    ™ ( �yT   ™ ( �yU   ™ ( �y   ™ (! �y\   ™ ( �yS   ™ ( �yZ   ™ ( �y]   ™ ( �yX   ™ ( �y[   ™ ( �yV   ™ ( �yY   ™ ( �ya   ™ ( �yW   ™ ( �y`   ™ ( �y    ™ ( �y_    ™ ( �y^    ™ ( �yc    ™ " xn        Ä xb@ÿÀŽ ä� xdÿ   @  ä xe9ÿÀŽ â&rb;   ÿ   âO xfÿ   I  ä x@IÿÀŽ ä xIÿ   9  Ä x9dÿÀŽ â &rd@;   ÿ   â xfÿÀŽ Æ &r;9E      âx?ÿÀŽ â &rdE   d   â x?ÿ   ?  â &r;e>   ;   âxIÿÀŽ ä x?ÿÀŽ â &r?>   d   â xd=ÿÀŽ â &r;G   ;   âxfÿ   =  â &r?G   ?   Ä &r;dA   ;   âxffÿÀŽ ä x>CÿÀŽ ä x=ÿ   8  ä xh:ÿÀŽ â &r?>A   ?   â xlÿ   C  â &r;fD   ;   â xA=   nìŽ Æ &r=hF   ;   â x;ÿ   :  ä xllÿÀŽ ä x88 €  ÿÀŽ ä x;;ÿÀŽ â &r?lD   ?   â x:7 €  ÿÀŽ ä xDÿ   7  Ä xj<ÿÀŽ ä Er7 8       â &r8;,   =   â xn    7    ä ErE :       â &rFjF   ?   â xG        â 0r:ÿB  0 A   â Er? 8       â #t77  `¿n  â#t8E  `¿n  â xDD   GìŽ Ä  rC7:     A  â  rG:8     A  â &r:bK   ÿ   ä&rK@K   ÿ   ä &r:9L   :   ä&rKL   K   ä &r:e6   :   Ä&rK6   K   ä &r:5   :   ä&rK5   K   â x<ÿ   <  Æ &rK>-   K   ä&r:d-   :   ä &r5b/   ÿ   ä&r/@/   ÿ   ä &r594   5   ä&r/4   /   Ä &r5e2   5   ä &rKl3   K   ä&r50   5   â x=<ÿÀŽ â &r:f3   :   ä &rKj1   K   ä &r5dQ   5   Ä 0rBÿB    A   ä &r/2   /   ä &r:h1   :   ä &r4bT   ÿ   ä&rT@T   ÿ   ä &r49U   4   Ä &rbb   ÿ   ä&r@   ÿ   ä &rTU   T   ä &rb9\   b   ä&r\      ä &r4eS   4   Ä &rTS   T   ä &reeZ   b   ä&rZ      ä &rK=N   K   ä 0r-ÿ.  0 A   ä &r5fP   5   â #rCB?   C   â &r4]   4   Ä &rT]   T   ä&rF=,   F   â Er, A       â &r/0   /   ä &r:;N   :   ä &reX   e   ä&rX      â ErA D       â 0r?ÿH    A   â ErK K       â 0rJÿJ    A   â  r08-     A  â 0r.ÿ.    A   Ä &r5hO   5   ä &r4d[   4   ä&rT>[   T   ä&rQ>Q   /   â Er/ :       â &redV   e   ä&r>V      â  r??,     A  â  r17-     A  â &r5;R   5   â  r,JA     A  â #r0.K   0   â 0r-ÿM  0 A   Ä &r4fY   4   ä&rTlY   T   ä&rQlP   Q   ä &refa   e   ä&rla      â #r/./   1   â #r,0      â 0rMÿM    A   â Er. 5       â &r4hW   4   âx   ÿàÿ â &rTjW   T   â r07-     A  â &rQjO   Q   Ä &reh`   e   ä&rj`      â |    pbô â &r4;   4   â#r.M.   0   â &rT=   T   ä&rQ=R   Q   ä 0rÿ_  0 A   Ä &re;^   e   ä0rÿc  0 A   ä &r=^      â #r?.      â ErE F       â 0r_ÿ_    A   â Er2 Q       â  r7     A  âEr4 4       â 0rcÿc    A   â ErT T       â  r.7     A  â Er e       â  r-8-     A  âEr0        â  r8     A  â r8     A  â #rEBE   G   â #r-M2   -   â #r_4      â#r_T      â #rc   .   â#rc0      â x�  ÿà÷ â #r?C      â#r,E      â#r?/      â#r ,-       â#r+?   +   â#r),   )   â#r?      â #r,      â x""@   ÿàÿ Ä x$$@   ÿàÿ ä x%%@   ÿàÿ ä x&&@   ÿàÿ ä x((@   ÿàÿ ä r#ÿ#   ÿäÿ â D)       À â Gy  0ìÿÿÿÿƒ ê Ay        € ê Ey  Ð     € â G‰  °     € ê $x@  Ž Ê ˆs  ûÿ H   è ˆs €ûÿ H   è ˆs  üÿ H   è ˆs €üÿ H   è ˆs  ýÿ H   è ˆs  €ýÿ H   è ˆs + þÿ H   è ˆs )€þÿ H   è ˆs  ÿÿ H   è ˆs €ÿÿ H   ä Ay        € ê {          ì M        € ê „y     H   b z  r  p`ð Ä z#  r      â „y €   H   ¢ r ÿ   p"p ä r!##   ÿŽ â „y    H   â x'        Æ „y €  H   ( „y    H   h „y	 €  H   ( „y    H   ( „y
 €  H   ( „y    H   è „y €  H   â !r       â/ !r       ÈO ‰      " !r       Æ� y      &   b !r       Æ‰      ¢ !r       Æy      '   â !r 	        Æ ‰      " !r+       Æ ‰	      b !r

)       â !r       ä ‰      " !r       Æ ‰      b !r       â !r       äO ‰       ¨ ‰      â !r       Æ‰
      " !r		       Æ‰      h ‰      b !r       â !r       ä/ ‰      " !r        ÆO ‰      b !r       Æ� ‰	      ¢ !r
       Æ‰  €    â !r       Æ‰      " !r       Æ ‰
      h ‰      b !r       Æ ‰      " !r       Æ/ ‰      b !r		       ÆO ‰      ¢ !r        â� $z  x  Ž â !r       â‰ €    ä $z  ~   Ž â !r
       â‰ €    " x
   ÿàÿ â !r       Ä ‰	 €    b z 
 r  p`ò â !r       ä ‰ @    " x    p"ò â !r       ä/ ‰ €    b !r       ÆO ‰ €    ¨ ‰ €    ¢ !r       â�  !   ÿàõ ä ‰ €    â !r       âÿÿ   ÿä Ä ‰ €    " !r	       â
 j  ÿ„ ä ‰  €    b !r
       â r#    ÿ† â!r       â/ 
 k   â ‰ @    " rÿÿ   ÿäÿ â !r       âO z j  ÿ† â ‰ @    b 
	  r  ÿàù â !r       â z k  � â ‰
      ¢ r#    ÿŠ â !r       â� $ÿÿ   ÿä â ‰ @    â r"ÿÿ   ÿäÿ â !r       âz j  ÿŠ â ‰ @    " 
	 j  ÿˆ â !r         âz k  "� â ‰ @    b 
		 k  $ Æ ‰ @    b !r       Æ ‰ @    " !r       â/ !   ÿàÿ ä ‰ @    b !r%
       âO #   ÿàÿ ä ‰  @    ¢ !r       â� !    ÿàõ â%v
  j  ' Ž â !r       â‰      â     ÿà÷ â !r       Ä‰      " "ÿÿ   ÿä â !r       ä ‰      b ! j  ÿ„ â !r       ä ‰      " !r       Æ/ ‰      b !r        ÆO ‰      ¢  ÿÿ   ÿäÿ ä  j  ÿ† â ‰       ¢ !r       ä�  k   � â ‰      â !r#       â! k  " Ä ‰      " !r       â!r       ä †	 
%    â !r       Æ/ † 
#   â !r       ÆO †	     â !r        Æ †     â !r       Æ� †	     â !r       Æ†    è †	     è †     è †	     â M™        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      "   " ¹z  l      â Ey  À    € â ™x?    â y      !   b ¹z  F   
   â r( ÿ       â x0      ÿ  â y%      %   ¢ r. ÿ       â x      ÿ  â r$ ÿ       â y      &   â r ÿ   pRð Ä r ÿ   pRò är'   ÿ(Ž È/ x ÿ   '  ä x%%   ÿ   äO r '   ÿ� ä r  ÿ       ä x"ÿ     È | "   pbô Ú G)  p    € ê y      '   "Ž r ÿ       â $z& w  ÿŽ âr ÿ       ä xøÿÿÿÿÀŽ ä z*  q      â 'z s   Ž â r''  €ÿàÿ ä x-"   ÿ   Ä zÿ t    ä x'   ÿ   ä r *-   ÿŽ â$x**   -Ž â z!  X      ä z)  Y      ä r( ÿ       â 'z y   Ž â x,ÿ   '  Ä z+- q  ÿàÿ â $z }  ÿŽ â zÿ z    â $z% p  "Ž Æ r&&   ÿàõ â $z	 |  "Ž ä$z |  Ž â x%   ÿàÿ â $z	 v  	Ž âr/ÿÿ   ÿä â $z v  Ž ä $z	 p  	Ž ä %x2   ÿŽ È %x	2   ÿŽ È %x
   Ž È %x'   Ž È %x   Ž È %x'   Ž È r
!   ÿàõ är2!   ÿà÷ ä r)   ÿä är3)   ÿäÿ ä r>!   ÿàõ â �y   • ¨ �y22"   • â  r?)   ÿä Ä r!   ÿàõ â �y   • ( �y>>"   • " r)   ÿä Ê �y4   • ( �y   • " x        ä r3'&   -àµ ä xJ $       ä r5,/   ÿd~ Ê $x55$   ÿŽ â r?'&   +àµ äxÿ     äO x2   ÿ   ä�x2   ÿ   ä x   øŽ ä x   ÿÀŽ È xÿ   øŽ ä x>   ÿ   Èx   ÿÀŽ â/ %v   Ž Æ xÿ   4øŽ â%v   Ž â xÿ   4  â �y#    ™ ¢  x>   ÿ   â %v3 Z  J Ž ä �y=    ™ ä %v   Ž â x   øŽ È/ �y<    ™ " x2   ÿ   ä x>   ÿ   ä r5   ÿàÿ ä x   ÿÀŽ ä x   ÿÀŽ â %v   Ž â �y9   ™ b xÿ   øŽ Ä xÿ   øŽ â �y:    ™ b xÿ     â %v   Ž ä�y8   ™ b xÿ   2  â %v4   Ž â xÿ     â �y;   ™ b xÿ   >  Ä x   øŽ â �y3    ™ b x   øŽ Æ �y7   ™ b r,/   ÿd~ Æ/ �y54    ™ b %v   Ž Æ �y6   ™ b %v   Ž Æ �y@   ™ b %v? Z  J Ž Æ �y4   ™ b# $x$   ÿŽ Æ �y?    ™ d r   ÿàÿ ä �y    ™ h �yB   ™ b r'&    àÙ Æ �yC   ™ b r!   ÿàõ Æ �y   ™ b! r,/   ÿ¤~ Æ �yD    ™ b r)   ÿä Æ �yA    ™ b r!   ÿà÷ Æ �yE   ™ b r	)   ÿäÿ Æ �yG   ™ b %v Z  J Ž Æ/ �y    • b $x$   ÿŽ Æ �yF   ™ d r   ÿàÿ ä �y    • b r'&   *àµ Æ �yI   ™ b rK,/   ÿd~ â %v Z  J Ž ä �y   ™ d# $xMK$   ÿŽ ä �yL    ™ h �yH    ™ b rM   ÿàÿ Æ �y   ™ h �yK   ™ h �yM   ™ h �yJ   ™ h �yP   ™ h �yR   ™ h �yN   ™ h �yQ   ™ h �y   ™ h# �yO   ™ h �yW   ™ h �y    ™ h �yS    ™ h �yV   ™ h �yU   ™ h �yT    ™ h �yX    ™ b x[        Ä xZÿ   >  È xZ   [ìŽ ä x\        ä xYÿ   2  È xYY   \ìŽ ä x>> €  ÿÀŽ ä x22 €  ÿÀŽ È Er2 2       ä ErY Y       ä x#ÿÀŽ äO x#ÿ   #  È x#ÿÀŽ ä/ x[ÿ   <  äx<<ÿÀŽ ä x[ÿÀŽ ä xZ=ÿÀŽ ä� x=ÿ   =  â &r[9   ÿ   Ä&r]<9   ÿ   â x\:ÿÀŽ ä x#ÿ   :  â &r98   [   ä&r]8   ]   â x8=ÿÀŽ â &r=Z;   9   âx##ÿÀŽ â &r]\;   ]   â x:3ÿÀŽ Ä x3ÿ   3  â &r;87   =   ä&r]#7   ]   â x95ÿÀŽ ä x7ÿ   5  ä x53ÿÀŽ â &r;:6   ;   âx37ÿÀŽ â &r696   ]   â Er[ >       â &r;54   ;   Ä &r=34   6   â x6?ÿÀŽ ä x4ÿÀŽ Æ &r;6@   ;   âx>ÿ   ?  â &r?4@   =   ä &r@B   ÿ   âxÿ     â &rB<B   ÿ   ä &r@C   @   ä&rBC   B   â xCÿÀŽ Ä x=>ÿÀŽ â &r@Z   @   âx7    7    â &rB\   B   ä &r?CD   ?   ä&r;=D   ;   â #t22  `¿7  â 0r>ÿA  0 A   ä &r8E   @   Ä &r@#E   B   â ErB ?       â 0rAÿA    A   â Er; ;       â &r?:G      â r2>     A  â &r@9G   @   ä 0rÿ    A   â #t7[  `¿7  â &r?5F   ?   â#rA;      â &r@3F   @   â Er;        â 0rÿ    A   â  rY     A  â  r>>7     A  â &r?6I   ?   Ä &r@4I   @   â #r      â #r>AB   >   â  r;     A  â &r   ÿ   ä&r?=L   ?   ä 0rÿH  0 A   ä &r<   ÿ   Ä &r@CL   @   â #r>      â &r      â Er? ?       â &r>      â  r2     A  â 0rÿH    A   â Er @       â  r7     A  â &r;ZK      ä&r>\K   >   â #r?      â#r      â &rM   ÿ   Ä &rM<M   ÿ   ä &r;8J   ;   ä&r>#J   >   ä &rP      ä&rMP   M   ä &r:R   ;   Ä &r>9R   >   ä &rZN      ä&rM\N   M   ä &r5Q      ä&r>3Q   >   ä &r8      Ä &r#   M   ä &r6O      ä&r>4O   >   ä &r:W      ä&r9W      ä &r=      Ä &r>C   >   ä 0rÿS  0 A   ä&r5V      ä&r3V      â Er        â 0rSÿS    A   â  r2     A  â x""   ÿàÿ â &r6U      Ä &r4U      â #rS      â | "   pbô â  r7     A  â &r=T      ä&rCT      ä 0rÿX  0 A   â#r        âEr> >       â 0rXÿX    A   â Er        â #r$   $   â Er        â  r2     A  â r7     A  â #rS>      â x!!   ÿà÷ â #rX      â#rX      â #r1   1   â#r0   0   â #r.   .   â #r(   (   â x  €   ÿàÿ Ä x**€   ÿàÿ ä x++€   ÿàÿ ä x--€   ÿàÿ ä r)ÿ)   ÿäÿ â D)       À â Gy  Ðîÿÿÿÿƒ ê Ay        € ê� Ey  °     € â G‰  �     € ê r   ÿ@Ž Ê ˆs  üÿ H   è ˆs €üÿ H   è ˆs   ýÿ H   è ˆs 1€ýÿ H   è ˆs $ þÿ H   è ˆs 0€þÿ H   è ˆs . ÿÿ H   è ˆs (€ÿÿ H   ä Ay        € ê {          ì M        € ê „y     H   b z % r  p`ð Æ „y €  H   ¢ r ÿ   p"p Æ „y €  H   è „y	 €   H   ( „y    H   ( „y €	  H   h „y    H   h „y €  H   h „y    H   h „y    H   h „y €  H   b !r       Æ/ „y!    H   b !r1       ÆO „y    H   ¢ !r       Æ� „y €  H   â !r		       Æ„y
    H   " !r       Æ „y €  H   " !r       Æ„y    H   b !r        Æ „y  	  H   b !r	       Æ „y  
  H   b !r       Æ „y €
  H   b !r$       Æ „y#    H   b !r0       Æ „y €  H   â !r!!.       Æ/ „y €  H   â !r       ÆO „y €  H   â !r       â� !r

       ä‰      b !r       Æ ‰      " !r!       Æ‰      ¢ !r       Æ y      '   â !r

       Æ ‰	      " !r       Æ ‰
      b z  r      â !r#       ä ‰      â !r       â/ !r(       ä ‰      b !r       â !r       Ä ‰       " !r       âO !r       ä ‰      ¢ !r		       Æ‰      " !r
       Æ‰	      b !r       Æ� ‰      â !r       Æ/ ‰      b !r         Æ ‰
      " !r       ÆO ‰      ¨ ‰  €    ¢ !r       â!r		       ä‰ €    " !r       Æ� ‰      â !r       Æ/ ‰	 €    b !r
       Æ ‰ €    " !r       ÆO ‰ €    ¢ !r        Æ ‰ €    h ‰ €    b !r       â$z x  %Ž â x%%   ÿàÿ â ‰ @    " !r       â� $z  ~  Ž â z % r  p`ò â !r
	       â/ ‰ @    b r   ÿŽ â !r       Ä ‰ €    " 
  r  ÿàõ â !r       äO ‰
 @    ¢ ÿÿ   ÿä â !r       â
 j  ÿ„ â ‰ @    â x    p"ò â !r       â 
 k   â ‰ @    b !r       âr    ÿ† Ä ‰ @    " !r	       Æ/ ‰ @    b rÿÿ   ÿäÿ â !r       ä ‰      " z j  ÿ† â !r
       äO ‰ @    ¢    ÿàÿ ä z k  � â !r       â� ‰      â     ÿàõ â !r       Ä‰	      â     ÿà÷ â !r       ä‰      " ÿÿ   ÿä â !r       ä/ ‰      b  j  ÿ„ â !r       Ä ‰      " x        ä ÿÿ   ÿäÿ â ‰      b 

 j  ÿ† â %v  j   Ž â !r       äO 
 k  � ä †	     â !r	       â� 	 k   â !r       â !r       Ä†    â !r       Æ/ †	     â !r       Æ †     â !r       Æ†	     è †    è ‰      ( †	 
    â M™        € ê x   ÿàÿ âO !r       Æ r    ÿàñ È r ÿÿ   ÿä  ä z j  ÿ€ È z k     Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      "   " ¹z  l      â Ey       € â ™x?    â y      !   b ¹z  F   
   â r ÿ       ä y!      %   ¢ r  ÿ       ä r( ÿ       â y      &   â r* ÿ       Ä r ÿ       ä r ÿ   pRð är ÿ   pRò är#   ÿ(Ž È/ x ÿ   #  ä x!!   ÿ   äO r #   ÿ� â $À   Ž â r  ÿ       ä xÿ     È |    pbô Ú G)  À    € ê y      '   "Ž r ÿ       â $z w  ÿŽ âr ÿ       ä xøÿÿÿÿÀŽ ä x&   ÿ   â 'z s   Ž â r##  €ÿàÿ ä z)  X      Ä zÿ t    ä x#   ÿ   ä z%  Y      ä r ÿ       ä x$ÿ   #  ä z"& q  ÿàÿ â 'z y   Ž È $z }  ÿŽ â zÿ z    â $z! p  Ž Æ r   ÿàõ â $z	 |  Ž âz  q      â $z |  Ž â x!   ÿàÿ â $z	 v  	Ž âr&   ÿŽ â $z v  Ž â r'ÿÿ   ÿä â $z	 p  	Ž ä %x2   ÿŽ È %x	2   ÿŽ È %x
   Ž È %x#   Ž È %x   Ž È %x#   Ž È r
)   ÿàõ är)   ÿà÷ ä r%   ÿä ä r,)   ÿàõ ä r%   ÿäÿ â�y.   • ¢  r-%   ÿä Ä r)   ÿàõ â �y"   • ( �y+,"   • â r%   ÿä Æ �y/   • ( �y0   • ( �y1   • " r#   &àµ ä x7 $       Ä r2$'   ÿd~ Æ %v Z  7 Ž È $x-2$   ÿŽ Ê/ r-   ÿàÿ äx:        ä x   ÿ   ä x+   ÿ   ä� x   ÿÀŽ ä x   ÿÀŽ Ä xÿ   .  äOx   ÿ   ä xÿ   .øŽ ä xÿ   0øŽ äx   øŽ â %v  : Ž â x0ÿ   0  Ä x-+   ÿ   â %vN  : Ž â r        â �yM    ™ ¢  x   ÿ   â %v  : Ž â x-0   -øŽ â �yNN    ™ â x+   ÿ   ä x   ÿÀŽ â �yK    ™ b	 x.   ÿÀŽ Ä xÿ   /øŽ â �yI   ™ b x.ÿ   1øŽ Æ �yJ   ™ b %v-  : Ž â	x/ÿ   /  ä �y5   ™ " x,ÿ     â %v  : Ž ä�yL    ™ " x1ÿ   1  â %v  : Ž â x.ÿ   +  â �y3   ™ " x8/   ,øŽ Ä x.1   .øŽ â �y/    ™ ( �y,    ™ " %v88  : Ž Æ �y2   ™ " %v:.  : Ž â r#   "àÇ ä �y48    ™ $ r$'   ÿ„þ ä �y-:    ™ " r)   ÿàõ Æ/ �y0   ™ " %v Z  7 Ž Æ �y1   ™ " $x9$   ÿŽ â r%   ÿä ä r8#   àÅ â �y.    ™ " r9   ÿàÿ ä r9$'   ÿ„~ â �y6    ™ ( $x;9$   ÿŽ â �y    • ( �y9   ™ " %v8 Z  7 Ž Æ/ �y7   ™ $ r;   ÿàÿ ä r)   ÿà÷ â �y8   ™ ( �y   ™ ( �y?   ™ " r	%   ÿäÿ Æ �y=   ™ ( �yA   ™ ( �y;   ™ ( �y    • ( �y@   ™ ( �y>   ™ ( �y   ™ ( �yE   ™ ( �yC    ™ ( �y<   ™ ( �yD   ™ ( �y:    ™ ( �yB    ™ ( �yG    ™ " xR        Ä xOÿ   +  ä xQ    7    ä xOO   RìŽ ä x   ÿàÿ ä ErO O       ä |    pbô Ä x))   ÿà÷ ä x€   ÿàÿ ä x""€   ÿàÿ ä x&&€   ÿàÿ ä r%ÿ%   ÿäÿ ä xFMÿÀŽ ÄO xPÿ   M  ä xMÿ   N  ä� xNNÿÀŽ ä xPPÿÀŽ ä xMÿÀŽ â/ &rFI   ÿ   â
xHKÿÀŽ â &rINI   ÿ   â xKÿ   K  â &rPJ      Ä &rJ   I   ä &rIH5      â	xLÿÀŽ ä xMÿ   L  ä xLKÿÀŽ â &r55      â xJMÿÀŽ ä x/ÿÀŽ â &rIL3   I   âx/ÿ   /  â &r5J3   5   â x,ÿÀŽ Ä x3ÿ   ,  â &r,2   I   âxI/ÿÀŽ â &rM2   5   â x53ÿÀŽ ä x34ÿÀŽ ä x/-ÿÀŽ â &r2I0   ,   ä&r050   M   â xKÿ     â &r231   2   âx4ÿ   4  â &r1/1   0   â x0 €  ÿÀŽ Ä x,K   RìŽ ä x4ÿÀŽ ä xK+ €  ÿÀŽ ä Er+ 0       ä x-ÿ   -  â &r2.   2   ä 0r4ÿ6  0 A   âx--ÿÀŽ â #t++  `¿Q  â Er0 K       â 0r6ÿ6    A   â ErK 2       â  rM+4     A  â &r.-.   1   â #t00  `¿Q  Ä #rK6K   M   â Er1 ,       â  rQ40     A  â ErM .       â &r,F7   ÿ   ä&r7N7   ÿ   ä &rFF   ÿ   â #rM6M   Q   â &rN   ÿ   Ä &rQP9   ,   ä&r99   7   ä &rFP?   F   ä&r?      ä &r98   9   ä&rQH8   Q   Ä &rFH=   F   ä&r=      ä &r9JA   9   ä&rQLA   Q   ä&rFL;   F   ä&rJJ;      Ä &rQ@   Q   ä&r9@   9   ä &rF   F   ä&rJ   J   ä &rQI>   Q   ä&r95>   9   Ä &rFIE   F   ä&rJ5E   J   ä 0rÿ    A   ä 0rÿC  0 A   ä &rQ3<   Q   ä&r9/<   9   Ä &rF3D   F   ä&rJ/D   J   â  r7O     A  â 0rÿ    A   â  r+     A  â r0     A  â &rQ:   Q   ä&r9-:   9   ä 0rÿB  0 A   Ä &rFG   F   ä&rJ-G   J   â  r11     A  â 0rCÿC    A   â Er Q       â 0rBÿB    A   â Er 9       â  r++     A  â Er F       â  r0     A  â Er J       â #rC      â#rC      â #rB   +   Ä #rB      â #r1K      â #r 7M       â #r*1   *   â #r(7   (   â #r 1       â #r7      â G©  pñÿÿÿÿƒ ê Ay        € ê� ˆ  ýÿ H   è ˆ  €ýÿ H   è ˆ * þÿ H   è ˆ (€þÿ H   è ˆ   ÿÿ H   è ˆ €ÿÿ H   è {          ì M        € ê „y     H   " z ! r  p`ð Æ „y	 €   H   b r ÿ   p"p Æ „y    H   ¨ „y    H   è „y €  H   ( „y    H   h „y €  H   h „y    H   h „y    H   h „y €  H   h „y €  H   b !r       Æ „y
    H   " !r		        Æ/ „y €  H   b !r       ÆO „y    H   ¢ !r*       Æ� „y €  H   â !r(       Æ„y €  H   â !r        Æ„y    H   â !r	       Æ „y €  H   â !r       â !r       È ‰      " !r       â !r       â !r

       â !r       â/ !r       ÈO ‰      " !r       Æ� ‰	      b !r       â!r       ä ‰      ¢ !r

       Æ ‰       â !r       Æ ‰
      ( ‰      b !r       Æ y      '   " !r		       Æ/ ‰      b !r       ÆO ‰	      ¢ !r         Æ� ‰      â !r
       Æ‰  €    " !r       Æ‰      h ‰      " !r       â/ !r	       ÈO ‰ €    b !r
       Æ� ‰	 €    ¢ !r        Æ‰
 €    â !r       Æ‰  @    " !r       Æ ‰ €    ( ‰ €    b !r       â/ $z x  !Ž â !r	       ÆO ‰ @    b x	!   ÿàÿ â $z ~  Ž â !r
       â� z  r      â ‰ @    ¢ z 	 r  p`ò â !r         Æx    p"ò â !r       â ‰       " !r       Æ‰ @    è ‰ @    &    ÿàÿ â !r	       â/ ‰ @    b 
 r  ÿàõ Ä r   ÿ† â !r       âO ‰
	      ¢ ÿÿ   ÿä â !r        â r ÿÿ   ÿäÿ â !r       ä� ‰      " !r       â
 j  ÿ„ â !r       Æ/ ‰      b 
 k   ä    ÿàõ â ‰      â x        â !r	
       âO ÿÿ   ÿä â ‰      ¢  j  ÿ„ â %v j   Ž â z j  ÿ† Ä 	 k   ä z k   � â †	     â	 !r       Æ †    â	 !r       Æ/ †	     â	 !r       â� !r       ÈO †     è	 †	     â	 M™        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      "   " ¹z  l      â Ey  @    € â ™x?    â y       !   b ¹z  F   
   â $rÿÿ   ÿ Ž â r ÿ       â y      %   ¢ r  ÿ       â $r*ÿÿ   ÿ Ž Ä y      &   â r ÿ   pRð är ÿ   pRò är!    ÿ(Ž È/ xÿ   !  ä x   ÿ   äO r!   ÿ� ä     ÿ8Ž ä x&ÿ     È | &   pbô Ú G)       € ê y      '   "Ž $rÿÿ   ÿ Ž â r ÿ       â $z w  ÿŽ âxøÿÿÿÿÀŽ â $x$&   ÿ Ž â z#  Y      â 'z s   Ž â r!!  €ÿàÿ ä r ÿ       â $v'ÿ X  ÿ Ž â zÿ t    Ä x!   ÿ   ä x"ÿ   !  ä z$ q  ÿàÿ â 'z y   Ž È $z }  ÿŽ â zÿ z    â $z p  &Ž Æ r   ÿàõ â $z	 |  &Ž ä$z |  Ž â x   ÿàÿ â $z	 v  	Ž âr%ÿÿ   ÿä â $z v  Ž ä $z	 p  	Ž ä %x2   ÿŽ È %x	2   ÿŽ È %x
   Ž È %x!   Ž È %x   Ž È %x!   Ž È r'   ÿà÷ är('   ÿàõ Æ $x   #Ž âr
'   ÿà÷ ä r)#   ÿä ä�y+"   • ¢ r#   ÿäÿ ä r'   ÿàõ â �y(("   • è �y   • " $x   # Æ �y4   • h �y   • h �y5   • b r,!   $àµ â $t1ÿ$   ÿ Ž Æ r"%   ÿd~ â %v, Z  1 Ž â x7        Æ $x$   ÿŽ Ê r   ÿàÿ ä r8!   àµ Æ �y0   ™ b r9"%   ÿd~ Æ �y.   ™ d $x99$   ÿŽ ä �y:   ™ h �y=    ™ h �y;    ™ b x+   ÿ   ÈO x   ÿÀŽ ä x(   ÿ   ä�xÿ   øŽ äx   ÿÀŽ â $x(    ÿ Ž â xÿ     â %v  7 Ž â xÿ   øŽ Äx+   ÿ   ä�y/    ™ ¢ xÿ     â %v  7 Ž âx   øŽ ä x+   ÿ   ä x,   øŽ â �y-    ™ " %v  7 Ž â x   ÿÀŽ Æ/ %v,  7 Ž â x,(   ÿ   â �y)    ™ f x,   ÿÀŽ â� �y3    ™ f x6ÿ   5øŽ â %v8 Z  1 Ž â �y,   ™ f %v6  7 Ž â/ r9   ÿàÿ â �y1   ™ b xÿ   4øŽ Æ �y6    ™ h �y8   ™ b x5ÿ   5  â %v  7 Ž â xÿ   (  â� �y<   ™ b x4ÿ   4  ä xÿ   +  â �y9   ™ b x5   øŽ Ä x44   øŽ â �y2    ™ d %v  7 Ž ä�y?   ™ h �y5   ™ b %v4  7 Ž Æ �y>   ™ b r'   ÿàõ Æ/ �y4   ™ h �y    ™ h �y7    ™ b $x   # Æ �yA   ™ h �y@   ™ b r'   ÿàõ Æ �y    ™ b! r	#   ÿä Æ �y    ™ è€ �y    • è �y    • â xB        ä xEÿ   +  È xE   BìŽ ä x++ €  ÿÀŽ ä x&&   ÿàÿ È | &   pbô ä x''   ÿà÷ ä x€   ÿàÿ ä x$$€   ÿàÿ â $r#ÿÿ   #Ž â x/ÿÀŽ äO x/ÿ   /  Ä xCÿ   -  äx--ÿÀŽ ä x/ÿÀŽ ä/ x/CÿÀŽ â &rC0   ÿ   ä&r0-0   ÿ   ä &r.   C   Ä &rC/.   0   â x0ÿ   3  äx)ÿÀŽ ä x.ÿ   )  ä x33ÿÀŽ â &rD,      âx)ÿ   (  ä x..ÿÀŽ â &rC3,   C   â x00ÿÀŽ Ä x)   BìŽ â &r,.1   D   ä&rB01   C   â x1( €  ÿÀŽ ä x(ÿ   6  ä xC6ÿÀŽ â &r68   ÿ   ä&r--8   ÿ   Ä &r6<   6   ä&r</<   -   â ErD +       â &r9   6   âx+2ÿÀŽ â &r<39   <   â xEÿ   2  â &r.?      â Er2 1       â &r<0?   <   â x)    7    â &r,+5   ,   âx1EÿÀŽ â &rEC5   B   â x((ÿÀŽ â &r.+>      Ä &r<C>   <   â #tDD  `¿)  â#t22  `¿)  â &r514   ,   âxÿÀŽ â &r4(4   E   â x)7ÿÀŽ ä x,ÿ   7  ä xÿ     â &r.1A   .   Ä &r<(A   <   â x,,ÿÀŽ â &r5):   5   âxÿÀŽ â &r4:   4   ä &r.)@   .   ä&r<@   <   ä 0rÿ=  0 A   Ä &r5,;   5   ä&r4;   4   ä &r.,   .   ä&r<   <   ä 0rÿ  0 A   â� Er        â  r8D     A  â r2     A  â 0r=ÿ=    A   â Er6 5       â 0rÿ    A   â Er4 4       â 0rÿ    A   â Er        â 0rÿ    A   â Er .       â  rD     A  âEr) <       â  r22     A  â #r6=6   8   â r     A  â #r=4      â  r     A  â #r      â#r2)   2   â #r*6   *   â #r        â #r      â #r2      â G©  ôÿÿÿÿƒ ê Ay        € ê� ˆ * þÿ H   è ˆ  €þÿ H   è ˆ  ÿÿ H   è ˆ €ÿÿ H   è {          ì M        € ê „y      H   " z  r  p`ð Æ „y	  €   H   b r  ÿ   p"p Æ „y     H   ¨ „y     H   è „y  €  H   ( „y     H   h „y     H   h „y  €  H   h „y     H   h „y
  €  H   h „y  €  H   b !r*       Æ „y  €  H   " !r		        â/ !r       âO !r       ä� !r	       â!r       â!r       ä !r       Æ ‰      b !r       Æ ‰      ¢ !r

       Æ ‰      â !r
       È !r       ä y      '   ( ‰      " !r       â/ !r       ÈO ‰      b !r
       Æ� ‰      ¨ ‰	
      â !r       â!r       È/ ‰      b !r       ÆO ‰ €    ¢ !r
	       Æ� ‰	 €    è ‰ €    " !r       â/ !r       ÈO ‰ €    b !r		       â� !r       È‰	 @    ¨ ‰ @    è ‰ @    " !r       â/ x   ÿàÿ È ‰ @    b z  r  p`ò â !r
	       ÆO x     p"ò â !r       â� $z x  Ž â ‰
      ¢ x        â !r       â$z ~  Ž â ‰      & %v j   Ž â‰      â 
  r  ÿàõ È 	ÿÿ   ÿä ä 
  j  ÿ„ â !r       â/ !r
       ÈO ‰      ¢ !r       â 
  k  	 ä †    â !r       Æ� †	     è †	     â M™        € ê ¹z  r      â !r       âO �x   ?àÿ Ì |   ÿàñ È rÿÿ   ÿä  ä/ z j  ÿ€ È z k    Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y       &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú         â $rÿÿ     Ž È % \   Ž Ê �	    ™ b ¹z  l      â Ey  €    € â ™x?    â y      "   b $rÿÿ   ÿ Ž Æ y      !   ¢ $ÿÿ   ÿ Ž ä $ÿÿ    Ž È '
 m   Ž È $‚ÿÿ   ÿ Ž â 
ÿ n    Ê $ÿÿ   
Ž â r ÿ   pRò Æ/ $
 o  Ž â r ÿ   pRô är   ÿ(Ž äO ‚        ä xÿ     Æ 'Š s   Ž â r   ÿ� â y      '   & xÿ     È |    pbö ä Šÿ t    Ö G9       € ê y	      %   " $rÿÿ   ÿ Ž â xøÿÿÿÿÀŽ â $z w  ÿŽ â x 2       â 'z y   Ž È $x   
Ž â zÿ z    â $z }  ÿŽ ä $x   ÿ Ž â xÿ     â $z |  Ž â r   ÿàñ â $x   ÿ Ž ä $z v  Ž ârÿÿ   ÿä  â %x   ÿŽ È $z	 p  Ž â x   ÿüŽ Æ %v X  Ž È $tÿ   ÿ Ž â r        ä r ÿ       â %v  € Ž Ê r   ÿàñ Ê $x     â r
   ÿàñ È �y    • ¢ r   ÿä  Ê �y
    • è �y 
   • " r   à± â $tÿ   ÿ Ž â x $       ä r   ÿd~  Ê $x$   ÿŽ ä $x   ÿ Ž ÊO x	   ÿÀŽ È x	ÿ   øŽ â�%v Z   Ž â xÿ     ä x   ÿ   â %v
   Ž â r		   ÿàÿ ä x   øŽ â $x   ÿ Ž â �y
    ™ ¨  x   ÿÀŽ â %v   Ž â �y   ™ ä xÿ    øŽ Äx#ÿ      â �y    ™ b	 x
ÿ     â %v   Ž ä �y!   ™ b x
#   
øŽ Æ �y    ™ d %v

   Ž ä �y    ™ h �y"   ™ h �y#   ™ h �y

    ™ h �y   ™ (�y   ™ ( �y$    ™ ( �y%    ™ $ $rÿÿ    Ž â/ r	        Ê �y    • " x   ÿàÿ ä x	 €  ÿÀŽ ä xÿ     ä |    pbð ä x   ÿà÷ ä x€   ÿàÿ Æ $rÿÿ   Ž â xÿÀŽ äO xÿ     È xÿÀŽ â &r   ÿ   â� xÿÀŽ äxÿ     â &r!      Æ xÿÀŽ â &r       â xÿÀŽ ä xÿ     â &r""      Æ xÿÀŽ â &r#   "   â x
ÿÀŽ ä x
ÿ   
  â &r      âEr 	       â $tÿ   ÿ Ž â x	
ÿÀŽ â &r      â x    7    Ä x   ìŽ â &r	$      ä #t  `¿  â 0r
ÿ%  0 A   âEr        â 0r%ÿ%    A   â Er        ä  r

     A  È #r%   
   â 0rÿ    A   Ê  r     A  È #r      â G‰  €úÿÿÿÿƒ ê Ay        € ê ˆ €ÿÿ H   è {          ì M)        € ê „y     H   " r ÿ   pRð Æ „y €   H   h „y    H   ¨ y      %   â !r       È !r       È/ !r       ÊO ‰      $ !r       Ê‰      $ !r       Ê ‰ €    $ !r       Ê ‰ @    $ !r	       â $z  x  Ž È� ‰
	      b  M	        € ê x        â $z ~  Ž â !r		
       Æ? %v j   Ž Ê †y 	    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž È %
 \   Ž Ê �	
    ™ b r ÿ       â Ey       € ä y       "   d 'Š s   Ž äy      !   d '
 m   Ž Ê 
ÿ n    â $rÿÿ   ÿ Ž â rÿ    ÿü‚ â/ $x     Ž Ê x	ÿ     È r		   ÿ� ä Šÿ t    â $rÿÿ   ÿ Ž â G       € ê z ÿ ^  pPò ä z ÿ `  pPô ä z ÿ _  Sò ä z ÿ b  p ö Ä z ÿ a   Sô È z ÿ c  0't Ê       '   "        Æ
      %   h ©      '   ¢ ¢       Æ ©      %   ä $ª x  ÿŽ ä $ ~  ÿŽ Ê    
àë ä/
ÿ   
  â $ª ~  ÿŽ äO $ x  ÿŽ â 
ÿÿ   
Äþ â $rÿÿ   ÿ Ž â ¢   àÇ ä�¨ÿ     Ä    ÿàý ä ¢   ÿàû ä ¢ÿÿ   „þ â $ÿÿ   
 â 
 ^  ÿ† ä ¢ÿ   ÿäÿ ä ª b  ÿˆ Ä  _  � â $rÿÿ   ÿ Ž â ª c   Æ �
     h �©     d Ay        € ê ¹z  l      â xÿ   	  â ™x?    â ÿ   ÿáÿ â Ey  À    € â $rÿÿ   ÿ Ž Æ |    pbò â $
 o  Ž âr ÿ       â $‚ÿÿ    Ž Ô G  P    € ê y      '   " $r
ÿÿ   ÿ Ž â x		øÿÿÿÿÀŽ â $rÿÿ    Ž â y      %   b $z w  ÿŽ â r	  €ÿàÿ â $tÿ   ÿ Ž â z ÿ `  pPð â $x   ÿ Ž â r ÿ       â %v X  Ž â z ÿ a   Sð Ä xÿ     â 'z y  
 Ž Ê zÿ z    Ê $z |  Ž È $z v  Ž ä$z }  ÿŽ ä $z p  Ž Æ/ r	   ÿàó â $x   ÿ Ž ä %x
2   ÿŽ â rÿÿ   ÿäÿ  Æ $rÿÿ   ÿ Ž ä %v X  Ž È r
   ÿàó Ê $x   Ž  â r
   ÿàó È �y"   • ¢  $x   Ž  Ê �y   • è �y   • " r	   à£ â $t$ÿ   ÿ Ž ä $t!ÿ$   ÿ Ž â r   ÿDþ  Æ %v Z  ! Ž È $x$   ÿŽ È $x   Ž Ê �y    ™ h �y&   ™ h �y'   ™ h �y#   ™ h �y%   ™ h �y!    ™ h �y(    ™ b x   ÿ   âO$x   ÿ Ž â/x   ÿ   Ä x   ÿÀŽ ä xÿ     ä�xÿ   øŽ ä x   øŽ Æ %v  $ Ž â x   ÿÀŽ È �y    ™ ¢ %v  $ Ž âxÿ   øŽ äxÿ     ä �y    ™ " x"ÿ     â %v  $ Ž Æ x   "øŽ ä/ �y"   ™ ( �y    ™ " %v  $ Ž â z
 X  ÿàó ä� �y   ™ è  �y    ™ â z Y  ÿäÿ  Ê �y    • â x   ÿàÿ ä xÿÀŽ äO xÿ     Æ &r    ÿ   âx)ÿÀŽ ä xÿÀŽ äxÿ     â &r)&      Æ xÿÀŽ â &r'      â xÿÀŽ ä xÿ     â &r#      Æ xÿÀŽ â/ &r%      â xÿÀŽ ä� xÿ     â &r"      â x €  ÿÀŽ ä xÿÀŽ â &r      â x        â $tÿ   7ÿ Ž â xÿ     Ä Er        â &r!      â x   ìŽ â 0rÿ(  0 A   ä #t  `¿  â Er        â 0rÿ    A   â Er        â  r     A  â 0rÿ(    A   â G‰       € ê $tÿ   ÿ Ž È %v `  Ž È %v `  Ž â r
   ÿàó È r   ÿäÿ  ä r
   ÿàó Æ �y("    ¤ $x   Ž  Ê �y)    è �y,    " $x*(   ÿ Ž âOx-(   ÿ   â$x(   ÿ Ž Æ x**   ÿÀŽ È x+*ÿ   )øŽ ä�x)ÿ   )  Æ %v*+  $ Ž âx-)   -øŽ ä x   ÿÀŽ ä �y)*    ™ ¢  %v-  $ Ž âxÿ   ,øŽ äx-ÿ   (  ä x,ÿ   ,  â �y    ™ â %v  $ Ž Ä x-,   -øŽ È �y    ™ " %v,-  $ Ž Ì �y,,    ™ b z*
 `  ÿàó È z+ a  ÿäÿ  Ê �y**     b x)ÿÀŽ äO x)ÿ   )  Æ &r    ÿ   â x))ÿÀŽ ä x ÿÀŽ ä� xÿ     â &r&)&      Æ xÿÀŽ â &r  '   &   â xÿÀŽ äxÿ     â &r#       Æ xÿÀŽ â &r%      â x,ÿÀŽ äx,ÿ   ,  â &r"      â x( €  ÿÀŽ â $tÿ   ÿ Ž â x,,ÿÀŽ â &r      â x(ÿ   (  â 0r*ÿ*    A   â Er        â &r,!      â x((   ìŽ Æ #t  `¿  â Er (       ä Er        â  r     A  ä  r**     A  ä #r      È #r*      â |    pbò â #r      â  r     A  â x

   ÿàõ ä x€   ÿàÿ â #r      â rÿ   ÿä Ì G™  Pöÿÿÿÿƒ ê Ay        € ê r  ÿ   pRð âEy  €     € â r  ÿ   pRò Ö G‰  P     € ê z ÿ `  pPð â $x      Ž Æ z ÿ a   Sð ä ˆs  €ÿÿ H   ö ˆ     H   ä Ay        € ê {          ì M        € ê z ÿ `  pPð â „y     H   fz ÿ a   Sð Ú „‰  €   H   ( „‰    H   ¢ !r	       È/ !‚ 	        È !‚        âO G‰        € ê „y  €  H   ( „y €   H   h „y    H   ¨ „y    H   è „y €  H   " !r         â !r	       Æ/ !r         âO !r       Æ� !r        ä‰       " z ÿ `  pPð ä r ÿ   pRò ä z ÿ a   Sð â !r         Ê ‰       $ !r        Ê ‰ €    $ !r       Ê ‰	 @    $ !r		       Ê ‰	      $ !r	       â G‰        € ê ‰       $ !r         Ê ‰       $ !r        Ê ‰ €    $ !r       Ê ‰ @    $ !r       Ê ‰      $ !r       ä M        € ê !r       â G‰  �    € ê $t ÿ   ÿ Ž â !r       È z   h  pð Ú G	  �    € ê ¹z  h      ä �xÿÿÿÿ?àÿ È Œx    pð È ‡x       È ™x   ?  Ø »z  €    ä ™x?    Ô Xy ðëÿÿÿÿƒ ê  x ;ª¸¿  A  Ì s          $ !t   €?    Ì s        $  r     A  È  r     A  â Gy  `    € ê  x '7=  A  â x*BL?  A  Æ #t   €?    È  r      A  â  x    ?  A  Ê s     $   $ !t  €?    È  r       A  È  r      A  â Gy  À    € ê z   h      È x     pRð Ú G‰  à     € ê x     pRð Ú G	        € ê 	z  i    � ä 	z i    � Æ  x ;ª¸¿  A  â 	z i €  � Ê s        $ !t  €?    Ì s        $  r       A  È  r      A  â Gy  Ð     € ê  r     A  â Gy  °     € ê 	x   à@  � ä 	x  à@  � Æ  x #ÛÙ?  A  â 	x  àÀ  � Æ  x;ª¸¿  A  ä !t  €?    È s        $ !t  €?    Ì s        $  r       A  È  r      A  ä y       &   ( y      %   ( y      '   b $z   x  Ž ä $tÿ   ÿ Ž Ä $z  ~   Ž È/ %v  j   Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y       &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú         â $rÿÿ     Ž È % \   Ž Ê �	    ™ b ¹z  l      â Ey       € â ™x?    â y	      "   b $rÿÿ   ÿ Ž Æ y      !   ¢ $ÿÿ   ÿ Ž ä $ÿÿ    Ž â y      %   & '
 m   Ž â y      '   æ $‚ÿÿ   ÿ Ž â 
ÿ n    È ÿ   ÿáÿ ä r 	ÿ   pRò Æ/ $
 o  Ž ä $x
	    Ž äO $rÿÿ   ÿ Ž â ‚        ä xÿ   
  Æ 'Š s   Ž â r
   ÿ� ä 	   ÿ8Ž ä xÿ     È |    pbô ä Šÿ t    ä r 	ÿ   pRð â x      ÿ  â $x	   ÿ Ž Î G)  0    € ê $rÿÿ   ÿ Ž â� xøÿÿÿÿÀŽ â $z w  ÿŽ ä 'z y   Ž â r

  €ÿàÿ Æ $z }  ÿŽ â zÿ z    â $z	 p  Ž â x!
   ÿ   â $vÿ X  ÿ Ž â r   ÿàõ â $z |  Ž â xÿ   
  â $rÿÿ   ÿ Ž â rÿÿ   ÿä â $z v  Ž Ä$vÿ Y  ÿ Ž ä %x2   ÿŽ È $x   ÿ Ž ä %x !   Ž È %x
   Ž Ê r"   ÿàõ Ê $x#    â r    ÿàõ È �y+""   • ¢  $x!    Ê �y   • â z  p      Æ �y   • $ %x,2   "Ž È %x.2   Ž â�y,"   • ( �y.   • " %x(2   ,Ž Æ �y.   • " %x"2   .Ž Æ �y'("   • " %x02   (Ž Æ �y%"   • (	 �y0"   • " %x2   "Ž Æ/ �y<"   • (	 �y$   • ( �y-   • br>
   àµ â $t;ÿ   ÿ Ž â x? $       Ä r)   ÿd~ Æ %v>> Z  ? Ž È $x))$   ÿŽ Ê r??)   ÿàÿ Ê �y4>   ™ h �y5>   ™ h �y.>   ™ h �y0>   ™ h �y,>   ™ h �y7>    ™ b $x&+   ÿ Ž ÊO x&&   ÿÀŽ â $x(+    ÿ Ž Æ x&&ÿ   øŽ ä�xÿ     Æ %v&  ; Ž â/ x(   (øŽ Æ $x+   ÿ Ž â �y&    ™ ¢ %v((  ; Ž Æ x   ÿÀŽ ä x/ÿ   +  ä x"ÿ   øŽ äxÿ     â �y(    ™ & x   /øŽ â %v""  ; Ž È %v  ; Ž â/ �y:"    ™ & $x   ÿ Ž âx(ÿ     â� $x/    ÿ Ž â �y9    ™ $ x*   ÿÀŽ ä x((   /øŽ â �y>   ™ " x2*ÿ   øŽ Ä x#'   ÿ   â/ �y)>   ™ " %v(  ; Ž â�r"   ÿàõ ä �y*>    ™ â  %v22  ; Ž â x6#   ÿÀŽ ä �y    ™ â $x(   ÿ Ž Æ �y2    ™ ä x/(   ÿÀŽ ä x66ÿ   %øŽ â $x#    â x>/ÿ   øŽ äxDÿ   %  â %v26  ; Ž â/x%ÿ     â �y6"    • â  xÿ     â %v>>  ; Ž â x1'   ÿ   â �y82    ™ ä $x(   ÿ Ž â xB   %øŽ â �y>    ™ â  $x'   ÿ Ž â xDD   1øŽ Ä x%(   ÿÀŽ â %vBB  ; Ž Æ x@%ÿ   $øŽ â%vDD  ; Ž âx%   ÿÀŽ ä �yB    ™ â
 x$ÿ   $  ä x>%ÿ   <øŽ â �y2D    ™ â. x%   ÿ   â %v@@  ; Ž â x=ÿ   '  Ä x$$   %øŽ â $x(   ÿ Ž â �y1@    ™ â  %v>>  ; Ž âx<ÿ   <  ä x((   ÿÀŽ â %v$$  ; Ž â�y3>    ™ â x<<   =øŽ ä %x"2   "Ž â x((ÿ   -øŽ â�y$$    ™ è
 �y/"    • â  %v><  ; Ž â/ x-ÿ   -  Æ %v<(  ; Ž âx(ÿ     â �y>>    ™ æ x(-   (øŽ â �y<<    ™ è %vB(  ; Ž Ê �y%B    ™ b%x"2   "Ž Ì %x@2   "Ž ä �y""    • h �y@@    • b x   ÿàÿ È |    pbô ä x   ÿà÷ ä x€   ÿàÿ Æ $rÿÿ   Ž â x-&ÿÀŽ äO x&ÿ   &  Æ &r-4   ÿ   â x&&ÿÀŽ ä x-ÿÀŽ Æ&r&5      â xÿ     Æ &r-.      â xÿÀŽ ä x#:ÿÀŽ ä x:ÿ   :  â &r0      Æ x-:ÿÀŽ â &r#,      â x#9ÿÀŽ ä x9ÿ   9  â &r-      Æ x99ÿÀŽ â &r##)      â x&+ €  ÿÀŽ Æ &r#9*   #   â� Er& &       â $tÿ   7ÿ Ž â x        ä x-ÿÀŽ ä x(ÿ     ä x+ÿ   +  â &r--4   ÿ   â x((ÿÀŽ â 0rÿ7  0 A   â Er9 #       â #t&&  `¿  â x#ÿÀŽ Ä xÿ     â &r((5   -   â x+   ìŽ â 0r7ÿ7    A   â  r&&     A  â x+ÿÀŽ â 0rÿ6    A   â Er        â &r(#.   (   â #r979   &   â x#8ÿÀŽ Ä x&ÿÀŽ â &r++0   (   â  r     A  â xÿ     â &r(#4   ÿ   â x8ÿ   8  â &r#&,   +   â x&ÿÀŽ ä x88ÿÀŽ ä xÿÀŽ â &r#&   #   â xÿ     â &r+85   (   â x        Ä x(2ÿÀŽ ä x-ÿ     â &r#)   #   â x2ÿ   2  ä xÿÀŽ ä x €  ÿÀŽ ä x-   ìŽ â &r+(.   +   â x;2ÿÀŽ â &r#*   #   â x-1ÿÀŽ Ä x1ÿ   1  ä Er        â &r+;0   +   â x3ÿÀŽ â &r--4   ÿ   â x&1ÿÀŽ â #t  `¿  â x3ÿ   3  â &r+,   +   â Er        â 0r/ÿ/    A   â x$ÿÀŽ â &r-&5   -   â Er #       â  r     A  â x33ÿÀŽ Ä x$ÿ   $  â &r-.   -   â  r#/     A  â #r7      â x>ÿÀŽ â &r3   +   â x1$ÿÀŽ â $t/ÿ   ÿ Ž â xÿ   '  â &r)      â x$<ÿÀŽ â &r-10   -   â x<ÿ   <  Ä x   /ìŽ ä x €  ÿÀŽ â &r$$,   -   â x/<ÿÀŽ ä Er&        ä x%ÿÀŽ â&r$/   $   â x>ÿ   >  Ä x'' €  ÿÀŽ ä x%ÿ   %  â &r)   $   â x+>ÿÀŽ ä Er '       ä x%%ÿÀŽ ä x(        Ä xÿ     â &r+*      â#t  `¿  â#t&&  `¿  â &r%*      â x   (ìŽ â  r     A  â0r"ÿ"    A   â Er        â 0r@ÿ@    A   â Er        â  r&     A  â Er        Ä Er        â #r7      â r""     A  â #r7      ä  r@     A  â #r9      â #r#      â #r"      â #r      â D)       À â Gy   ïÿÿÿÿƒ ê Ay        € ê� ˆ  þÿ H   è ˆ €þÿ H   è ˆ  ÿÿ H   è ˆ €ÿÿ H   è {          ì M	        € ê „y     H   " z 	 r  p`ô Æ „y €   H   b r ÿ   pVt Æ „y    H   ¨ „y    H   è„y €  H   ( „y    H   ( „y
    H   ( „y €  H   ( „y €  H   ( „y    H   ( „y €  H   " !r       Æ „y €  H   " !r       â/ !r       âO !r       ä� !r       â!r       â !r
       â !r       È ‰      b !r       â !r       È ‰      ¢ !r       Æ ‰      â !r       Ê ‰      " !r       â/ !r       âO !r       È� ‰      h ‰      ¢ !r       Æ ‰      ( ‰      â !r       â/ !r       ÈO ‰
 €    b !r       Æ ‰ €    " !r       Æ� ‰ €    ¨ ‰ €    â !r

       â/ !r       È ‰
 @    " !r       ÆO ‰ @    b !r       Æ� ‰ @    ¨ ‰ @    â !r
       â x
	   ÿàÿ â !r       Æ/ ‰      " $z  x  	Ž â x 	   ÿàÿ â!r       âO ‰      b x	   ÿàÿ â $z ~  Ž â !r       â� ‰      ¢ z   r  p`ö ä z  r  p`ð â ‰      â z 
 r  p`ò Ä x    pVö äx    pVp  äx    pVò  ä x         â !r       È %v j    Ž â !r       È/ †¹    â !r       ÆO †©     â !r       Æ� †‰    â M        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      &   " z ÿ \  pPò â ¹z  F   
   Æ z ÿ ]  Sò Ú         Ê % \   Ž Ê �	    ™ ¢  r ÿ       â ¹z  l      â r ÿ       â y      "   b ™x?    â '  m   Ž âEy  `    € â y      %   â r ÿ       â 'š s   Ž â  ÿ n     â y
      !   &  ÿ   ÿáÿ â y      '   ( $  o  Ž â ’         Ä r ÿ       Æ $z   w  ÿŽ â r ÿ   pRð ä/ x   ÿ   ä� x 
   pDp  är
   ÿàÿ â$z }  ÿŽ â r
   ÿ(Ž â 'z y   Ž â z  r  pdp  Ä r    ÿàõ ä x ÿ     È r    ÿ� ä š	ÿ t    È r	     € âO G	  À    € ê z ÿ `  pPò ä z ÿ ^  pPð ä z ÿ b  p ö ä z ÿ a  Sò ä z ÿ _   Sð ä z ÿ c  0'ò Ä r ÿ       ä r ÿ       Î $
  ~  ÿŽ âÿ     â$š ~  ÿŽ â ˜ÿ     â $
 x  ÿŽ â     
àë â$š x  ÿŽ â ’   
àÇ ä ÿ   ÿÄþ ä      ÿàý Ä ’   ÿàû ä ’ÿ   ÿ„þ ä ÿ   ÿä ä 
  ^  ÿ† ä ’ÿ   ÿäÿ ä š b  ÿˆ Ä 
  _  � ä š c   Æ �	     h �™     d Ay        € ê xÿ     â Ey   $    € â r ÿ       â x      ÿ  â |    pbð â x      ÿ  â r# ÿ       Ä r% ÿ       ä r ÿ       ä rÿÿ   ÿä Ê G	  `#    € ê xøÿÿÿÿÀŽ â $z p  Ž â zÿ z    â x      ÿ  â r  €ÿàÿ â x      ÿ  â x   ÿàÿ â $z |  Ž â x4   ÿ   â $z  |  Ž â x   ÿàÿ â $z	 v  Ž â x   ÿàÿ â %x44   ÿŽ â z ÿ `  pPð Ä z  `      â $z 	 v   Ž â z  a      â %x&2   4Ž â z  X      ä z  Y      â %x.   ÿŽ â x!&   ÿàó ä r ÿ       â $z p   Ž âr# ÿ       â $z p   Ž âr% ÿ       â $z p   Ž â r ÿ       â %x(2   .Ž â z ÿ a   Sð Ä x    ÿ   â %x*2   .Ž â x"ÿ     ä r$ÿ'   ÿäÿ  â %x,2   .Ž È %x02   4Ž È %x22   4Ž È %x.2   .Ž È %x42   4Ž È r*   ÿàó äzO  p      ä r+   ÿäÿ  ä r8&   ÿàó Æ �yB"   • ¢  r9'   ÿäÿ  â %xO2   Ž â r!   ÿàõ Æ �y<8   • â r$   ÿä Æ �yC"   • ä %xO2   Ž ä�y78   • è �y6þÿÿ • è �y     • â %x:O2   Ž Ê �yD:"   • â %xO2   :Ž È %xO2   Ž â�yE"   • è  �yRþÿÿ • â  %xO2   Ž Æ�y@    • è �yAþÿÿ • è  �yP    • â r8    à£ ä/ x9 $       Ä r:"   ÿDþ  ä x	        â %v88 Z  9 Ž â xB   ÿ   ÈO x   ÿÀŽ È xÿ   <øŽ â�$x:$   ÿŽ â xC   ÿ   ä x<ÿ   <  â %v  	 Ž âxB   ÿ   ä r99   ÿàÿ ä x   ÿÀŽ â �yF    ™ ¢  x<   øŽ Ä xB   ÿ   â �y8   ™ â xHÿ   6øŽ â%v  	 Ž âx   ÿÀŽ â �y8   ™ " x6ÿ   6  â %vHH  	 Ž â xC   ÿ   â �yG    ™ " x:ÿ   7øŽ Ä x66   øŽ â �yHH    ™ " xÿ   7  â %v::  	 Ž ä�y8   ™ " xC   ÿ   â %v66  	 Ž â xÿ   B  â �y8   ™ " x=   ÿÀŽ Ä x>   øŽ â �yI:    ™ "# x<=ÿ    øŽ Æ �yJ6    ™ " %v>>  	 Ž Æ �y8   ™ " %v<<  	 Ž â x;ÿ   C  ä/ x ÿ      â �y8   ™ ( �yK>    ™ " x:    ;øŽ Æ �yL<    ™ $ %v::  	 Ž ä �y8   ™ " x>D   ÿ   â %x62   Ž ä �yQ:    ™ " x=>   ÿÀŽ Æ/ �y 8    ™ " x>E   ÿ   Æ �yN8    ™ " x<=ÿ   RøŽ äxT>   ÿÀŽ â �yM6    • " x?D   ÿ   â %v<<  	 Ž â xRÿ   R  ä xTTÿ   AøŽ ä xVE   ÿ   Ä xAÿ   A  ä x>R   ?øŽ ä x8D   ÿ   â/ �yR<    ™ " xVA   VøŽ â %v:T  	 Ž âxA8   ÿÀŽ ä xUE   ÿ   â %v8>  	 Ž â �yT:    ™ " xXAÿ   @øŽ Ä x=U   ÿÀŽ â/ %v>V  	 Ž â�yU8    ™ $ xZ=ÿ   PøŽ â %v<X  	 Ž â�yV>    ™ " xYÿ   D  ä xXÿ   @  â %v@Z  	 Ž â x;ÿ   E  â �yW<    ™ " xPÿ   P  Ä xXX   YøŽ â �y@@    ™ " x:P   ;øŽ Æ %v8X  	 Ž È/ %v::  	 Ž ä �y88    ™ $	 %x6O2   6Ž ä�y::    ™ ( �yS6    • $ %x6O2   6Ž Ì %x<O2   6Ž ä �y66    • ( �y<<    • " xAC €  ÿÀŽ È ErA A       ä xCÿ   C  ä x?FÿÀŽ äO xFÿ   F  Æ &r??   ÿ   â� xFFÿÀŽ ä x9GÿÀŽ Æ&r?F   ?   â xGÿ   G  ä x;HÿÀŽ ä/ xHÿ   H  ä x>GÿÀŽ â &r99   ?   â xHHÿÀŽ â &r;;   ÿ   Ä &r9>   9   â x7IÿÀŽ ä xIÿ   I  ä x?JÿÀŽ â &r=H   ;   â xJÿ   J  ä x>IÿÀŽ â &r97   9   â xF        Ä x;ÿ   B  â &r=?   =   â xJJÿÀŽ â &r9>   9   â x7;   FìŽ ä x;KÿÀŽ â &r=J   =   â x>LÿÀŽ ä xBB €  ÿÀŽ Ä xKÿ   K  â &r9;   9   â xLÿ   L  â &r=>   =   â x>    7    ä Er; B       ä xKKÿÀŽ ä xLLÿÀŽ â #tF;  `¿>  â x;QÿÀŽ â &r9K    9   â xQÿ   Q  â &r=L   =   Ä 0r?ÿN  0 A   âxQQÿÀŽ â 0rMÿM    A   â ErB 7       â 0rNÿN    A   â Er9 9       â &r;;   =   â  r7F?     A  â  rBMB     A  ä &r;Q    ;   â #r7N9   7   â #tFA  `¿>  â x9RÿÀŽ Ä #r%B7   %   â Er; ;       â  r7?F     A  â xRÿ   R  â &r99   ÿ   ä #rBN;   7   â xRRÿÀŽ ä x;TÿÀŽ ä xTÿ   T  Ä xH        ä x7UÿÀŽ â &r9R   9   â xUÿ   U  â &r;;   ÿ   â xTTÿÀŽ ä xCC   HìŽ â &r77   9   â xHUÿÀŽ Ä x=VÿÀŽ â &r;T   ;   â xVÿ   V  â &r7H   7   âx9WÿÀŽ â &r;=   ;   â xVVÿÀŽ ä xWÿ   W  â &r79   7   â x=@ÿÀŽ â &r;V   ;   â x9ÿ   @  Ä x@WÿÀŽ â &r==   ;   â xH9ÿÀŽ ä x98ÿÀŽ â &r7@   7   â ErC C       â 0rFÿS    A   â x@:ÿÀŽ â &r=H   =   â x8ÿ   8  Ä x;ÿ   D  â &r99   7   âxDD €  ÿÀŽ â  rFFC     A  â x78ÿÀŽ â &r@@   =   â x:ÿ   :  ä xC        ä x8ÿ   E  Ä Er= D       ä xEE €  ÿÀŽ â &r77    9   â xA:ÿÀŽ ä x98   CìŽ â #t8=  `¿>  â xJ        ä ErE E       â &r@A    @   â x;;   JìŽ â  rA?8     A  Ä #t8E  `¿>  â 0r6ÿ6    A   â Er; ;       â 0r<ÿ<    A   â Er7 7       â  r8?8     A  â Er9 9       ä Er= @       â #r#FB   #   â #rAN7   A   â  rC6;     A  â  r@<9     A  â #rBN=   8   â G‰  à    € ê r6*   ÿàó È r7+   ÿäÿ  ä r8&   ÿàó Æ �yF6"    ¢ r9'   ÿäÿ  Ê �y:8    è �y;8    " x<F   ÿ   äOx6F   ÿ   ä x=<   ÿÀŽ È xD=ÿ   :øŽ ä�x:ÿ   :  ä x=F   ÿ   â %vDD  	 Ž Æ x::   =øŽ ä x66   ÿÀŽ â �yDD    ™ ¤ %v<:  	 Ž âx:6ÿ   ;øŽ äx6ÿ   ;  ä x7ÿ   F  â �y<<    ™ â %v::  	 Ž Ä x66   7øŽ È �y::    ™ " %v86  	 Ž Ì �y98    ™ "	 %x62   Ž Ê �yG6     " x;        ä x=DÿÀŽ äO xDÿ   D  Æ &r==   ÿ   â xDDÿÀŽ ä xE<ÿÀŽ ä� x<ÿ   <  â &r=D   =   Æ x<<ÿÀŽ â &r=E   =   â x8:ÿÀŽ äx:ÿ   :  â &r=<   =   Æ x::ÿÀŽ â &r88   =   â x69ÿÀŽ ä/ x9ÿ   9  â &r7:   8   â x8F €  ÿÀŽ â 0rGÿG    A   â x99ÿÀŽ â &r66   7   â xFÿ   F  Ä Er7 8       â &r69    6   â xFF   ;ìŽ Æ #t87  `¿>  â ErF F       ä Er7 6       â  r9?8     A  ä  rGGF     A  ä #r7N7   9   È #rG7      ä G‰  à    € ê r6(   ÿàó È r7)   ÿäÿ  ä r80   ÿàó Æ �yF6"    ¢ r91   ÿäÿ  Ê �y:8    è �y;8    " x<F   ÿ   äOx6F   ÿ   ä x=<   ÿÀŽ È xD=ÿ   :øŽ ä�x:ÿ   :  ä x=F   ÿ   â %vDD  	 Ž Æ x::   =øŽ ä x66   ÿÀŽ â �yDD    ™ ¤ %v<:  	 Ž âx:6ÿ   ;øŽ äx6ÿ   ;  ä x7ÿ   F  â �y<<    ™ â %v::  	 Ž Ä x66   7øŽ È �y::    ™ " %v86  	 Ž Ì �y98    ™ "	 %x62   Ž Ê �yG6     " x;        ä x=DÿÀŽ äO xDÿ   D  Æ &r==   ÿ   â xDDÿÀŽ ä xE<ÿÀŽ ä� x<ÿ   <  â &r=D   =   Æ x<<ÿÀŽ â &r=E   =   â x8:ÿÀŽ äx:ÿ   :  â &r=<   =   Æ x::ÿÀŽ â &r88   =   â x69ÿÀŽ ä/ x9ÿ   9  â &r7:   8   â x8F €  ÿÀŽ â 0rGÿG    A   â x99ÿÀŽ â &r66   7   â xFÿ   F  Ä Er7 8       â &r69    6   â xFF   ;ìŽ Æ #t87  `¿>  â ErF F       ä Er7 6       â  r9?8     A  ä  rFGF     A  ä #r7N7   9   È #rF7      ä G‰  à    € ê r6,   ÿàó È r7-   ÿäÿ  ä r82   ÿàó Æ �yD6"    ¢ r93   ÿäÿ  Ê �y:8    è �y=8    " x;D   ÿ   äOxED   ÿ   ä x;;   ÿÀŽ ä x<ÿ   :  Ä� xF;ÿ   :øŽ ä x6D   ÿ   ä x<<   EøŽ â %v:F  	 Ž â xF6   ÿÀŽ È �yE:    ™ ¢  %v6<  	 Ž âx8Fÿ   =øŽ äx=ÿ   =  ä �yF6    ™ â x<ÿ   D  â %v88  	 Ž Æ x<=   <øŽ Æ �y88    ™ $ %v<<  	 Ž Ì �y<<    ™ " %x:2   Ž Ì �y::     " x=        ä/ x7EÿÀŽ äO xEÿ   E  Æ &r77   ÿ   â x6EÿÀŽ ä xEFÿÀŽ ä� xFÿ   F  â &r76   7   Æ xFFÿÀŽ â &r7E   7   â x68ÿÀŽ äx8ÿ   8  â &r7F   7   Æ x88ÿÀŽ â &r66   7   â x9<ÿÀŽ ä x<ÿ   <  â &r78   6   â x6D €  ÿÀŽ â 0r:ÿ:    A   â x;<ÿÀŽ â &r79   7   â xDÿ   D  Ä Er9 6       â &r7;    7   â xDD   =ìŽ Æ #t69  `¿>  â Er9 D       ä Er7 7       â  r;?6     A  ä  r9:9     A  ä #r7N7   ;   È #r97      â #rCA      â #r@B      â G‰  à    € ê r6.   ÿàó È r7/   ÿäÿ  ä r84   ÿàó Æ �y@6"    ¢ r95   ÿäÿ  Ê �y:8    è �y=8    " x;@   ÿ   äOxA@   ÿ   ä x;;   ÿÀŽ ä x<ÿ   :  Ä� xB;ÿ   :øŽ ä x<<   AøŽ ä x6@   ÿ   â %v:B  	 Ž Æ x66   ÿÀŽ ä �yA:    ™ ¢ %v8<  	 Ž âx66ÿ   =øŽ äx=ÿ   =  ä x<ÿ   @  â �y98    ™ â  %v66  	 Ž Ä x<=   <øŽ È �y76    ™ "	 %v<<  	 Ž Ì �y<<    ™ " %x:2   Ž Ì/ �y::     " x	AÿÀŽ äO xAÿ   A  Æ &r		   ÿ   â x8AÿÀŽ ä x9ÿÀŽ ä� x9ÿ   9  â &r	8   	   â x        ä x69ÿÀŽ â&r   	   â x7ÿÀŽ Ä x7ÿ   7  â &r6      Æ x7ÿÀŽ â &r      â x<ÿÀŽ ä x<ÿ   <  â &r      â x@ €  ÿÀŽ â 0r:ÿ:    A   â x<ÿÀŽ â &r      â x@ÿ   @  Ä Er        â &r       â x@@   ìŽ Æ #t>  `¿>  â Er @       ä Er        â  r??>     A  ä  r::     A  ä #rN   ?   È #r:      â x   ÿàÿ ä x   ÿàõ ä |    pbò ä x   ÿà÷ ä x  €   ÿàÿ ä rÿ   ÿä Ä rÿ   ÿäÿ Ê D       À â Gy  0ßÿÿÿÿƒ ê Ay        € ê y      "   " Ey  0    € â r ÿ   pRð är ÿ   pRò Ö G‰  ð     € ê z ÿ `  pPð ä r 
   ÿ8Ž ä z ÿ a   Sð Æ ˆs  % þÿ H   ô ˆƒ  #€þÿ H   è ˆƒ   ÿÿ H   è ˆƒ  €ÿÿ H   â G‰  p     € ê ˆs     H   è ˆs  #€þÿ H   è ˆs  €  H   è ˆs   ÿÿ H   è ˆs     H   è ˆs  €ÿÿ H   è ˆs  €  H   ä Ay        € ê {          ì M        € ê y       &   "> z ÿ `  pPð Ä x        â y      '   b z ÿ a   Sð â $z   x  Ž È $z  ~   Ž È/ %v  j   Ž È G	  P    € ê „y 
     H   " z  r  p`ð Æ „y
 €   H   b r 
ÿ   pVp  Æ „y
    H   ¨ „y
    H   è „y	
 €  H   ( „y
    H   ( „y
    H   ( „y
 €  H   („y
    H   ( „y
 €  H   ( „y
 €  H   " !r  %       Æ „y
 €  H   " !r#       â/ !r       âO !r         ä� !r	       â!r       â !r         ä !r       Æ ‰       b !r       Æ ‰      ¢ !r       Æ ‰      â !r       È !r       â !r        â/ !r       ÈO ‰      " !r       Æ� ‰      h ‰      ¨ ‰	      â !r       â !r       â/ !r       ÈO ‰      " !r	       Æ� ‰ €    h ‰	 €    ¨ ‰  €    â !r       â !r       È/ ‰ €    " !r		       âO !r        È� ‰	 @    h ‰  @    ¨ ‰ @    â !r       Ê ‰ @    " !r	       â/ x	   ÿàÿ â !r         ÆO ‰      b x   ÿàÿ â !r       â� z 	 r  p`ô â ‰       ¢ z  r  p`ò ä x   ÿàÿ â ‰      â x 
   pVò  äx 
   pVt â !r       â !r       È/ ‰      * !’       â !r        ÈO !‚       â!r	       â� †™    æ !¢		       â †‰     â z  r  p`ð Æ †© 	   â x 
   pVp  â !r       Ø M	        € ê !r       Ê/ †y    â My        € ê z   h      È x     pBð Ú G	  @    € ê x     pRð Ú G‰  À    € ê x     pRð Ú G	   $    € ê „y
     H   " z  r  p`ð â Ey  p    € ä „y
    H   b r 
ÿ   pVp  Æ „y	
    H   ¨ „y 
    H   è „y
    H   ( „y
  
  H   ( „y
 €   H   ( „y
 €  H   ( „y
 €  H   è „y
 €  H   " !r%       â !r       Ä/ „y
 €  H   â !r	       âO !r         Æ� !r       â!r         È ‰      " !r#       Æ ‰       b !r       È !r       â !r       â !r       â !r        Æ/ !r       â ‰      ( ‰      b !r       â !r	       È/ ‰ €    ( ‰	 €    b !r       Æ „y
 €
  H   â !r	       Æ/ ‰ @    ( ‰ @    b !r       â !r       È/ ‰       ( ‰      b !r       â !r	        â !r       â/ G	  �     € ê !r       â!r 	       Æ  x;ª¸¿  A  Ì s        $ !t  €?    Ì s        $  r     A  È  r      A  Ê †y     ä Ay        € ê ‰       b Ey  Ð    € æ ‰      ¨ „y
    H   è „y
    H   ( „y
    H   è „y
    H   â !r         â/ !r       ÈO ‰       h ‰      ( „y
    H   ¢ !r       â� !r       â!r        â/ !r       È ‰ €    ( ‰ €    h „y
  	  H   â !r       ÈO !r       â !r       â !r	       È/ ‰ @    ( ‰	 @    b !r       È� !r       â !r       â x   ÿàÿ â !r	       Æ/ ‰       " z  r  p`ð Æ ‰      b x 
   pVp  â !r	        â !r       Ö/ G	  �     € ê !r       â!r 	       Æ  x;ª¸¿  A  Ì s        $ !t  €?    Ì s        $  r     A  È  r      A  Ê †y    ä Ay        € ê ‰       b Ey       € æ ‰      ¨ „y
 €  H   è „y
 €	  H   ( „y
 €  H   è „y
 €  H   â !r         Æ/ „y
 €  H   â !r       ÆO ‰       h ‰      ( „y
 €  H   ¢ !r       È� !r       â!r        â/ !r       È ‰ €    ( ‰ €    b !r       âO !r       â !r	       È/ ‰ @    ( ‰	 @    b !r       â x   ÿàÿ â !r	       Æ/ ‰       " z  r  p`ð Æ ‰      b x 
   pVp  â !r	        â !r       Ö/ G	  �     € ê !r       â!r 	       Æ  x;ª¸¿  A  Ì s        $ !t  €?    Ì s        $  r     A  È  r      A  Ê †y    ä Ay        € ê ‰      b !r       â x   ÿàÿ Æ !r       â z  r  p`ð Æ !r       â x 
   pVp  È ‰       ¢ !r       Ê/ ‰      " !r         ÊO ‰       b !r       Ê ‰ €    " !r        Ê/ ‰ €    b !r	       Ê ‰	 @    " !r       Ê/ ‰ @    b !r	       Ê ‰      " !r       Ê/ ‰       ¢ !r       â M	        € ì !r       â/!r        ÆO  x;ª¸¿  A  â !r       Ê s        $ !t  €?    Ì s        $  r     A  È  r     A  Ê †y    â My        € ê „y 
     H   " z  r  p`ð â Ey       € ä „y
    H   b r 
ÿ   pVp  Æ „y
    H   ¨ „y
    H   è „y	
    H   ( „y
  
  H   ( „y
 €  H   ( „y
 €  H   ( „y
 €  H   è „y
 €
  H   " !r  %       â !r       Ä/ „y
 €   H   â !r         âO !r       Æ� !r	 	       â!r       È ‰ 	      " !r       Æ ‰      b !r       È !r       â !r 	        â !r#       â !r       Æ/ ‰       ( ‰      h „y
 €  H   ¢ !r        â !r       È/ ‰ €    ( ‰ €    b !r       ÈO !r       â !r       â !r	       È/ ‰ @    ( ‰	 @    b !r       â !r	       È/ ‰       ( ‰      b !r	        â !r        â/ G	  À     € ê !r         È x '7=  A  â x *BL?  A  Æ #t   €?   È  r     A  â  x    ?  A  â !r 	       È s     $   $ !t  €?    È  r     A  È  r      A  Ê †y     ä Ay        € ê ‰       b Ey       € æ ‰      ¨ „y
    H   è „y
    H   ( „y
  	  H   è „y
    H   â !r         Æ/ „y
    H   â !r       ÆO ‰       h ‰      " !r       â� !r       â!r        Æ/ !r       â !r       â ‰ €    ( ‰ €    b !r       Æ „y
    H   ¢ !r       â !r	       È/ ‰ @    ( ‰	 @    b !r       ÈO !r       â !r       â x   ÿàÿ â !r	       Æ/ ‰       " z  r  p`ð Æ ‰      b x 
   pVp  â !r	        â !r        Ö/ G	  À     € ê !r         È x '7=  A  â x *BL?  A  Æ #t   €?   È  r     A  â  x    ?  A  â !r 	       È s     $   $ !t  €?    È  r     A  È  r      A  Ê †y    ä Ay        € ê ‰       b Ey  Ð    € æ ‰      ¨ „y
 €  H   è „y
 €	  H   ( „y
 €  H   â !r         Æ/ „y
 €  H   â !r       ÆO ‰       h ‰      ( „y
 €  H   ¢ !r       È� !r       â!r        â/ !r       È ‰ €    ( ‰ €    h „y
 €  H   â !r       âO !r       â !r	       È/ ‰ @    ( ‰	 @    b !r       â x   ÿàÿ â !r	       Æ/ ‰       " z  r  p`ð Æ ‰      b x 
   pVp  â !r	        â !r        Ö/ G	  À     € ê !r         È x '7=  A  â x *BL?  A  Æ #t   €?   È  r     A  â  x    ?  A  â !r 	       È s     $   $ !t  €?    È  r     A  È  r      A  Ê †y    ä Ay        € ê ‰      b !r       â x   ÿàÿ Æ !r       â z  r  p`ð Æ !r       â x 
   pVp  È ‰       ¢ !r       Ê/ ‰      " !r         ÊO ‰       b !r       Ê ‰ €    " !r        Ê/ ‰ €    b !r	       Ê ‰	 @    " !r       Ê/ ‰ @    b !r	       Ê ‰      " !r       Ê/ ‰       ¢ !r       â M	        € ì !r       â/!r        ÆO  x'7=  A  â x*BL?  A  â !r       ä #t  €?   â x   ?  A  Æ  r     A  Ì s     $   $ !t   €?    È  r      A  È  r     A  Ê †y    â My        € ê x     pRð Ú G‰  P    € ê x     pRð Ú G	  `    € ê „y 
     H   " z  r  p`ð â Ey       € ä „y
    H   b r 
ÿ   pVp  Æ „y
    H   ¨ „y
    H   è „y	
    H   ( „y
  
  H   ( „y
 €  H   ( „y
 €  H   ( „y
 €  H   è „y
 €
  H   " !r  %       â !r       Ä/ „y
 €   H   â !r         âO !r       Æ� !r	 	       â!r       È ‰ 	      " !r       Æ ‰      b !r       È !r       â !r 	        â !r#       â !r       Æ/ ‰       ( ‰      h „y
 €  H   ¢ !r        â !r       È/ ‰ €    ( ‰ €    b !r       ÈO !r       â !r       â !r	       È/ ‰ @    ( ‰	 @    b !r       â !r	       È/ ‰       ( ‰      b !r	        â !r        â/ G	  À     € ê !r         Ê	z  i    � â !r 	       È  x;ª¸¿  A  â 	z   i    � È 	z   i €  � â s        $ !t  €?    Ì s        $  r     A  È  r      A  Ê †y     ä Ay        € ê ‰       b Ey       € æ ‰      ¨ „y
    H   è „y
    H   ( „y
  	  H   è „y
    H   â !r         Æ/ „y
    H   â !r       ÆO ‰       h ‰      " !r       â� !r       â!r        Æ/ !r       â !r       â ‰ €    ( ‰ €    b !r       Æ „y
    H   ¢ !r       â !r	       È/ ‰ @    ( ‰	 @    b !r       ÈO !r       â !r       â x   ÿàÿ â !r	       Æ/ ‰       " z  r  p`ð Æ ‰      b x 
   pVp  â !r	        â !r        Ö/ G	  À     € ê !r         Ê	z  i    � â !r 	       È  x;ª¸¿  A  â 	z   i    � È 	z   i €  � â s        $ !t  €?    Ì s        $  r     A  È  r      A  Ê †y    ä Ay        € ê ‰       b Ey  Ð    € æ ‰      ¨ „y
 €  H   è „y
 €	  H   ( „y
 €  H   è „y
 €  H   â !r         Æ/ „y
 €  H   â !r       ÆO ‰       h ‰      ( „y
 €  H   ¢ !r       È� !r       â!r        â/ !r       È ‰ €    ( ‰ €    b !r       âO !r       â !r	       È/ ‰ @    ( ‰	 @    b !r       â x   ÿàÿ â !r	       Æ/ ‰       " z  r  p`ð Æ ‰      b x 
   pVp  â !r	        â !r        Ö/ G	  À     € ê !r         Ê	z  i    � â !r 	       È  x;ª¸¿  A  â 	z   i    � È 	z   i €  � â s        $ !t  €?    Ì s        $  r     A  È  r      A  Ê †y    ä Ay        € ê ‰      b !r       â x   ÿàÿ Æ !r       â z  r  p`ð Æ !r       â x 
   pVp  È ‰       ¢ !r       Ê/ ‰      " !r         ÊO ‰       b !r       Ê ‰ €    " !r        Ê/ ‰ €    b !r	       Ê ‰	 @    " !r       Ê/ ‰ @    b !r	       Ê ‰      " !r       Ê/ ‰       ¢ !r       â M	        € ì !r       â/!r        ÈO 	z i    � â !r       È  x;ª¸¿  A  â 	z i    � È 	z i €  � â s        $ !t  €?    Ì s        $  r     A  È  r     A  Ê †y    â My        € ê „y 
     H   " z  r  p`ð Æ „y
    H   b r 
ÿ   pVp  Æ „y
 €   H   ¨ „y
 €  H   è „y
    H   ( „y!
    H   ( „y
    H   ( „y
    H   ( „y
 €  H   ( „y
 €  H   ( „y
    H   " !r  %       Æ „y
  	  H   " !r       Æ/ „y	
    H   b !r#       ÆO „y
  
  H   ¢ !r       Æ� „y
 €  H   â !r       Æ„y
 €
  H   " !r!!       Æ „y'
    H   " !r         Æ „y)
    H   " !r       Æ „y
 €  H   " !r       Æ „y
 €  H   " !r       Æ „y+
 €  H   " !r       Æ „y-
 €	  H   â !r!       Æ „y%
 €  H   â !r 	       Æ/ „y/
 €  H   â !r       âO !r       ä� ‰       " !r       Æ‰      b !r'       Æ ‰      ¢ !r))       Æ ‰      â x   ÿàÿ â !r       ä ‰	      " z  r  p`ò â !r       ä ‰)      " x 
   pVò  â !r+       â !r         â !r-       â !r       â/ !r%%       â !r       âO x   ÿàÿ â‰       " !r//       â !r       â� z  r  p`ô â ‰%      b x   ÿàÿ â !r	       âx 
   pVt â ‰/      ¢ z  r  p`ö â !r)       Ä ‰      â x 
   pVö Æ ‰	      ( ‰      ( ‰      " !r        Æ ‰      " !r%       â/ !r/       äO ‰  €    b !r       Æ� ‰      ¢ !r		       Æ‰      â !r       Æ ‰ €    " !r       Æ ‰	 €    " !r       Æ ‰ €    ( ‰ €    " !r         Æ/ ‰ €    b !r       âO !r       ä� ‰  @    ¢ !r       Æ‰ €    â !r	       Æ ‰ €    " !r       Æ ‰ @    " !r       Æ ‰	 @    " !r       Æ/ ‰ @    h ‰ @    b !r        ÆO ‰ @    ¢ !r       â� !r       ä‰       â !r       Æ ‰ @    " !r		       Æ ‰ @    " !r       Æ/ ‰      b !r       Æ ‰	      b !r       ÆO ‰      ¨ ‰      ¢ !r        Æ� ‰      â !r       â !‚        â!r       Æ‰      " !r       Æ/ ‰      b !r		       â !‚       â!r       äO !’	       â  ‚      A  â !r       â !’       Æ †‰     â !r       â� !¢       â  ’     A  ä !¢	       â !r       ä †™    â  ¢		     A  â !r       È/ †© 	   â M9        € ê !r       â !r       È  r     A  Ê †y    â My        € ê „y 
     H   " z  r  p`ð â Ey  À    € ä „y
    H   b r 
ÿ   pVp  Æ „y
    H   ¨ „y
    H   è „y	
    H   ( „y
  
  H   ( „y
 €  H   ( „y
 €  H   ( „y
 €  H   è „y
 €
  H   " !r  %       â !r       Ä/ „y
 €   H   â !r         âO !r       Æ� !r	 	       â!r       È ‰ 	      " !r       Æ ‰      b !r       È !r       â !r 	        â !r#       â !r       Æ/ ‰       ( ‰      h „y
 €  H   ¢ !r        â !r       È/ ‰ €    ( ‰ €    b !r       ÈO !r       â !r       â !r	       È/ ‰ @    ( ‰	 @    b !r       â !r	       È/ ‰       ( ‰      b !r        â !r        â/ G	  à     € ê !r         Ê	x   à@  � Ê  x #ÛÙ?  A  È  x ;ª¸¿  A  â !r        Ê s        " 	x    à@  � È 	x    àÀ  � Ê !t   €?    â !t  €?    È s	        $  r	     A  È  r     A  Ê †y     ä Ay        € ê ‰       b Ey       € æ ‰      ¨ „y
    H   è „y
    H   ( „y
  	  H   è „y
    H   â !r         Æ/ „y
    H   â !r       ÆO ‰       h ‰      " !r       â� !r       â!r        Æ/ !r       â !r       â ‰ €    ( ‰ €    b !r       Æ „y
    H   ¢ !r       â !r	       È/ ‰ @    ( ‰	 @    b !r       ÈO !r       â !r       â x   ÿàÿ â !r	       Æ/ ‰       " z  r  p`ð Æ ‰      b x 
   pVp  â !r        â !r        Ö/ G	  à     € ê !r         Ê	x   à@  � Ê  x #ÛÙ?  A  È  x ;ª¸¿  A  â !r        Ê s        " 	x    à@  � È 	x    àÀ  � Ê !t   €?    â !t  €?    È s	        $  r	     A  È  r     A  Ê †y    ä Ay        € ê ‰       b Ey  ð    € æ ‰      ¨ „y
 €  H   è „y
 €	  H   ( „y
 €  H   è „y
 €  H   â !r         Æ/ „y
 €  H   â !r       ÆO ‰       h ‰      ( „y
 €  H   ¢ !r       È� !r       â!r        â/ !r       È ‰ €    ( ‰ €    b !r       âO !r       â !r	       È/ ‰ @    ( ‰	 @    b !r       â x   ÿàÿ â !r	       Æ/ ‰       " z  r  p`ð Æ ‰      b x 
   pVp  â !r        â !r        Ö/ G	  à     € ê !r         Ê	x   à@  � Ê  x #ÛÙ?  A  È  x ;ª¸¿  A  â !r        Ê s        " 	x    à@  � È 	x    àÀ  � Ê !t   €?    â !t  €?    È s	        $  r	     A  È  r     A  Ê †y    ä Ay        € ê ‰      b !r       â x   ÿàÿ Æ !r       â z  r  p`ð Æ !r       â x 
   pVp  È ‰       ¢ !r       Ê/ ‰      " !r         ÊO ‰       b !r       Ê ‰ €    " !r        Ê/ ‰ €    b !r	       Ê ‰	 @    " !r       Ê/ ‰ @    b !r	       Ê ‰      " !r       Ê/ ‰       ¢ !r       â M	        € ì !r       â/!r        ÈO 	x  à@  � â !r       È  x#ÛÙ?  A  â 	x  à@  � Æ  x;ª¸¿  A  â 	x  àÀ  � Ê s        " !t  €?    â !t  €?    Ì s         $  r      A  È  r     A  Ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      "   $ z  w  p`ð Ú M	        € ê y      !   " ¹z  l      â Ey      € â ™x?    â y      %   b ¹z  F   
   â x      ÿ  ä y      &   ¢ x ÿ     È |     p`ð â $x   ÿ Ž Ø/ G	  p    € ê x        âO $z x  Ž È %v \   Ž Ì �y    ™ ¢  $rÿÿ   ÿ Ž â x   ÿàÿ â $z	 r  ÿŽ â r ÿ       â 'z m   Ž â z!  X      Æ $x
   ÿ Ž âzÿ n    â $x   ÿ Ž ä $v#ÿ Y  ÿ Ž â x

   ÿÀŽ â $rÿÿ   
Ž â x   ÿÀŽ â $rÿÿ   ÿ Ž ä $z o  Ž È $z u  ÿŽ È %x$   ÿ Ž Ì %x	$    Ž â x	   ÿÀŽ Ê %x	$   Ž â x	    ÿ   Ê %x	$   Ž Ê z Z  ÿàñ â $z q   Ž Æ x   ÿàó È zÿ [  þ  â $z t   Ž äO$z t  Ž ä $rÿÿ   ÿ Ž ä $z q  Ž ä %x2   
Ž È %x2   Ž È %x
2   
Ž È %x2   Ž Ê r$!   ÿàó â�y)    ™ ¢  r!   ÿàñ ä r%#   ÿäÿ  â �y.   ™ ä  $x	   #  â r
!   ÿàó â�y$"   • " r!   ÿàñ Æ �y"   • b $x   #Ž  ä$x   #  â �y-   ™ ¨  �y&   • ¨  �y    • è �y   • è  �y   • è  �y$ôÿÿ ™ è" �y0   ™ è �y/ðÿÿ ™ â $x'   ÿ Ž âx"   ÿ   Èx''   ÿÀŽ ä x"   ÿÀŽ ä x   ÿ   ä x''ÿ   &øŽ äOx&ÿ   &  ä xÿ    øŽ â�$x%    ÿ Ž â x ÿ      Ä x&   øŽ â %v   Ž â x%    %øŽ Æ $x&   ÿ Ž â �y     ™ ¢  %v'   Ž Æ x,&   ÿÀŽ â %v*%   Ž â�y"    ™ & $x'   ÿ Ž â �y%*    ™ " %v   Ž Æ �y&øÿÿ ™ " x,ÿ   øŽ ä�xÿ     â �y(    ™ " x,ÿ     ä x''   ÿÀŽ ä x   ,øŽ â %v   Ž âx'ÿ   øŽ â�y,   ™ " xÿ     â %v   Ž â�xÿ     â �y'üÿÿ ™ ä %v   Ž Ä �y*    ™ bx   øŽ Æ �y+    ™ d %v   Ž ä �y    ™ h �y    ™ b r !       â $rÿÿ   # Ž È %x2   Ž È %x2   Ž ä �y    • h �y    • b x     ÿàÿ È |     p`ð ä x€  ÿàõ ä/ x!!È   ÿàó Æ $rÿÿ    â r#ÿ#   ÿäÿ  ä x ÿÀŽ äO x ÿ      ä x1"ÿÀŽ äx"ÿ   "  â &r$   ÿ   âx ÿÀŽ â &r11$   ÿ   â x3"ÿÀŽ Ä x%ÿÀŽ â &r &      âx%ÿ   %  â &r&3&   1   â x(ÿÀŽ ä x"%ÿÀŽ ä x(ÿ   (  â &r'       ä�&r&'   &   â x*ÿÀŽ â&r")      â x((ÿÀŽ Ä x*ÿ   *  â &r,      âx+ÿÀŽ â &r)()   &   â x+ÿ   +  ä x*ÿÀŽ ä x"        ä xÿ     â &r),   )   â x++ÿÀŽ â &r .      â x   "ìŽ Ä xÿÀŽ â &r.+.   )   â xÿÀŽ ä xÿ     ä x €  ÿÀŽ ä xÿ     â &r-       âx%ÿÀŽ â &r-   .   â x €  ÿÀŽ Ä Er         ä xÿÀŽ â &r%0      â Er        â $t'ÿ   7ÿ Ž â x%        ä xÿ     â &r0      â #t  `¿'  â0rÿ/  0 A   âx   %ìŽ â #t    `¿'  â Er        â 0r/ÿ/    A   â  r     A  â Er        â 0rÿ    A   â Er        â 0rÿ    A   â Er        â  r      A  â #r/      â  r     A  Ä #r/      â  r     A  â #r      Æ #r      â G‰  ðöÿÿÿÿƒ ê Ay        € êO ‰      " $x   Ž Æ ‰       d z  p  p`ð È x    pDp  â !r       â !r        È/ ‰       ( ‰      b !r         â !r       È/ ‰  €    ( ‰	 €    b !r        â !r		       È/ ‰ @    ( ‰
	 @    b !r       â !r
	
       È/ ‰      h  ‰
      ¢  M	        € ê x   ÿ   â $z s  Ž â !r
       âO $tÿ   ÿ Ž â x    p"ò â $z v  Ž â !r       â/ r ÿ   p"ð Ä %v j   Ž Ð         Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      "   $ z  w  p`ð Ú M	        € ê y      &   " $t ÿ   ÿ Ž â ¹z  F   
   â $z	 x  Ž È %v	 \    Ž Ê �y    ™ b ¹z  l      â Ey  à    € â ™x?    â y      !   b r ÿ       â $rÿÿ   ÿ Ž â x      ÿ  â x      ÿ  â x,ÿ     È/ | ,   p`ð Ú G	  @    € ê $r
ÿÿ   ÿ Ž â x    ÿ   â$x6   ÿ Ž â z ÿ `  pPð â 'z m  
 Ž â x     ÿÀŽ â Ey °    € â x66   ÿÀŽ â $z t  ,Ž âzÿ n    â $r!ÿÿ   ÿ Ž â z ÿ a   Sð â $r7ÿÿ   ÿ Ž â r
ÿ   ÿáÿ â $vÿ `  ÿ Ž â y      %   & $z

 o  Ž Ä $z r  ÿŽ ä $z
 u  ÿŽ ä $v
ÿ X  ÿ Ž ä %x$   ÿ Ž È $vÿ a  ÿ Ž ä %x$    Ž â x   ÿÀŽ Ê %x$   Ž È $x,   ÿ Ž È %x$   Ž È $vÿ Y  ÿ Ž â z Z  ÿàó È x   ÿàõ È zÿ [  	$~ ä x	   ÿ   ä r ÿ       ä x	   ÿàÿ â$z	 q  ,Ž È $z q  Ž ä $z t  Ž ä %x2    Ž È %x2   6Ž È %x2    Ž È %x2   6Ž Ê r
   ÿàõ är
   ÿàó är&
   ÿà÷ â$x    â r   ÿäÿ  är'   ÿäÿ ä �y%"   • ¢ r
   ÿàó Æ �y"   • ä $x   Ž  ä�y&&"   • ( �y+   • h �y$   • h �y   • b %x:2   
Ž â x(%   ÿ   ÆO $x*%    ÿ Ž â x#ÿ   "  ä� x(   ÿÀŽ ä x&   ÿ   â$x&    ÿ Ž â x)#   *øŽ ä x#ÿ   "øŽ ä x"   ÿÀŽ â %v)    Ž â xÿ   +  Äx5"ÿ   +øŽ â %v#    Ž â/x)   øŽ â �y3    ™ ¢  x%   ÿ   â %v45    Ž ä�y/    ™ â x   ÿÀŽ â $x"&   ÿ Ž ä �yôÿÿ ™ " x#ÿ   $øŽ â %v()    Ž Ä �y44    ™ ¢
 x'"   ÿÀŽ Æ �yøÿÿ ™ ¢ x$ÿ   $  Æ �yüÿÿ ™ ¢. xÿ   %  â %v"#    Ž ä�y-(    ™ ¢  x''ÿ   øŽ äx$   øŽ â �y    ™ ¢ xÿ     â %v*'    Ž âx'ÿ   &  â �y."    ™ ¤ %v(    Ž âx1   'øŽ â �y   ™ ¨ �y+*    ™ ¢  %v01    Ž Æ �y'(    ™ ¨  �y   ™ ¨ �y"   ™ ¢. x#ÿ     Æ �y0    ™ ¤ $z5 t  #Ž ä�y#   ™ d $z$ q  5Ž ä �y)ðÿÿ ™ d %x8$2   
Ž Ä �y(:    • h �y*8    • b x,,   ÿàÿ È | ,   p`ò ä x1/ÿÀŽ ä� x2ÿ   /  Æ &r11   ÿ   â	x22ÿÀŽ ä x/4ÿÀŽ äO x0ÿ   3  ä x4ÿ   4  â &r12   1   â x33ÿÀŽ â &r//   ÿ   â x44ÿÀŽ Ä x00ÿÀŽ â &r33   1   â x1-ÿÀŽ â &r/4   /   â x-ÿ   -  â &r20   3   Æ x0-ÿÀŽ â &r/1   /   â x3.ÿÀŽ ä x1ÿ   .  â &r/0   /   ä &r.3   2   â x01ÿÀŽ ä x1+ÿÀŽ Ä x3ÿ   +  ä x+'ÿÀŽ â &r.0   .   âx03ÿÀŽ â &r/1   /   ä &r.+"   .   â x+ÿ   '  â &r/0   /   â x'ÿÀŽ Ä xÿ     ä x1+ÿÀŽ â &r/'"   /   â xÿÀŽ Æ &r'1#   .   â
x0ÿ   &  â &r.#   /   â x/% €  ÿÀŽ ä x&& €  ÿÀŽ â $t+ÿ   ÿ Ž â x    7    ä Er1 /       ä x-        Ä x2ÿ   %  ä Er3 &       â 0r%ÿ)    A   âx-2   -ìŽ â 0r)ÿ)  0 A   â #t&1  `¿  âx+0   +ìŽ â #t23  `¿  â 0r/ÿ*    A   â Er* '       â  r&&)     A  â Er. .       â 0r(ÿ(    A   â Er- -       â  r')2     A  â Er0 +       â #r&%*   &   Ä #r*%.   '   â  r((-     A  â  r'/0     A  â G‰  0    € ê xÿ     Ê $z	 q  Ž È $z t  Ž È %x2   6Ž È %x2    Ž â r.   ÿàõ Ê $x/    â r0   ÿàõ È �y.."    ¢ r1   ÿä Ê �y+0    è �y40    " $x-.   ÿ Ž âOx8.   ÿ   È x2-   ÿÀŽ â $x-.   ÿ Ž Æ x32ÿ   +øŽ ä�x+ÿ   +  Æ %v23    Ž âx9+   8øŽ ä x--   ÿÀŽ ä �y+2    ™ ¢  %v89    Ž â x1-ÿ   4øŽ È�y-8    ™ â %v01    Ž â x4ÿ   4  ä x/ÿ   .  Æ �y00    ™ " x/4   /øŽ Ê %v4/    Ž Ê �y/4    ™ b %x22   Ž Ì �y22     b x9+ÿÀŽ äO x8ÿ   +  Æ &r99   ÿ   â x88ÿÀŽ ä x+-ÿÀŽ ä� x-ÿ   -  â &r88   9   Æ x4-ÿÀŽ â/ &r8+   8   â x+0ÿÀŽ äx0ÿ   0  â &r44   8   Æ x00ÿÀŽ â &r-+   4   â x+/ÿÀŽ äx4ÿ   /  â &r-0   -   â x0. €  ÿÀŽ ä x44ÿÀŽ â &r++"   -   â x/        Ä x.ÿ   .  ä Er- 0       â &r+4#   +   â x..   /ìŽ Æ #t0-  `¿  â Er- .       â 0r2ÿ2    A   â Er. +       â  r0)0     A  Æ  r22-     A  â #r.%.   0   È #r2.      â #r(&      â #r'*      â G‰  à    € ê r&   ÿàõ Ê $x'    â r*   ÿàõ È �y(&"    ¢ r+   ÿä Ê �y-*    è �y2*    " $x.(   ÿ Ž äO$x&(   ÿ Ž Æ x..   ÿÀŽ È x1.ÿ   -øŽ ä�x-ÿ   -  ä x.(   ÿ   â %v01    Ž Æ x/-   .øŽ ä �y-0    ™ ¢  x'&   ÿÀŽ ä %v./    Ž äx''ÿ   2øŽ äx2ÿ   2  ä �y..    ™ â x3ÿ   (  â %v*'    Ž Æ x'2   3øŽ Æ �y**    ™ $ %v&'    Ž Ì �y&&    ™ b %x0$2   Ž Ì �y00     b x/-ÿÀŽ äO x-ÿ   -  Æ &r//   ÿ   â x$-ÿÀŽ ä x.ÿÀŽ ä� x.ÿ   .  â &r/$   /   Æ x..ÿÀŽ â &r   /   â x*ÿÀŽ äx*ÿ   *  â &r.      â x( €  ÿÀŽ ä x**ÿÀŽ â &r      â x&ÿÀŽ Äx&ÿ   &  â &r*      â x        ä x&&ÿÀŽ â &r"      â x(ÿ   (  â 0r0ÿ0    A   â Er        â &r&#      â x((   ìŽ Æ #t  `¿  â Er (       ä Er        â  r)     A  ä  r0     A  ä #r%      È #r      â xÈ   ÿà÷ ä x€  ÿàù ä x

È   ÿàõ â $rÿÿ   Ž ä $rÿÿ    â rÿ   ÿä â D       À â Gy  €îÿÿÿÿƒ ê Ay       € ê r        â $rÿÿ    Ž Ä Ay        € ê z ÿ `  pPð È z ÿ a   Sð Ú G	  P    € ê ‰      ( ‰      b !r       â !r	       È/ ‰      ( ‰	      b !r       â !r
	       È/ ‰ €    ( ‰
 €    b !r       â !r
       È/ ‰ @    ( ‰ @    b !r       â !r       È/ ‰      ( ‰	      b !r       â !r	       â/ Gy  €    € ê ‰      ( ‰
      h ‰	      ¨ ‰      â !r       â !r

       â/ !r	       ÆO ‰	      " !r       Æ� ‰
      h ‰      ¨ ‰      â !r		       â !r
       È/ ‰	 €    " !r       ÆO ‰ €    b !r       Æ� ‰ €    ¨ ‰ €    â !r	       â !r       â/ !r       ÈO ‰ @    " !r       Æ� ‰ @    h ‰ @    ¨ ‰	 @    â !r       â !r       â/ !r       ÈO ‰
      " !r	       Æ� ‰      h ‰	      ¨ ‰      â !r
       â !r       â/ !r	       âO !r       Ä� y      %   $ r   ÿŽ È z  p  p`ð È x    pDp  Ú M	        € ê z ÿ ^  pPð â $z	 v  Ž Æz ÿ _   Sð Ú %	 ^    Ž Ì �	     ¢ $x   ÿ Ž Ê r ÿ   p"ò äx    p"ö äx    p"ø äx    p"ô Î         ä z ÿ `  pPò â $2ÿÿ    Ž â B        ä z ÿ a  Sò â $"ÿÿ    Ž Ê r        Ê !       äO G™      € ê z ÿ b  pPð È z ÿ c   Sð Ú %	 b    Ž Ì �	     ¢ $x   ÿ Ž Ê x øÿÿÿp"ò äx üÿÿÿp"ô är ÿ   p"ö äx    p"ø â $vÿ h  ÿ Ž Ì         ä x    pBò â $"ÿÿ    Ž â 2        ä B        Ê !       ÈO G  P    € ê x    pRð Ú G‰  �     € ê x    pRð Ú G	  ð    € ê  x;ª¸¿  A  Ì s        $ !t  €?    Ì s        $  r     A  È  r     A  â Gy  P    € ê  x'7=  A  â x*BL?  A  Æ #t  €?   â  x   ?  A  Æ  r     A  Ì s     $   $ !t  €?    È  r     A  È  r     A  â Gy  °    € ê x    pRð Ú G‰  à     € ê x    pRð Ú G	        € ê 	z i    � ä 	z i    � Æ  x;ª¸¿  A  â 	z i €  � Ê s        $ !t  €?    Ì s        $  r     A  È  r     A  â Gy  Ð     € ê  r     A  â Gy  °     € ê 	x  à@  � ä 	x  à@  � Æ  x#ÛÙ?  A  â 	x  àÀ  � Æ  x;ª¸¿  A  ä !t  €?    È s        $ !t  €?    Ì s        $  r     A  È  r     A  ä y      "   ( y      &   b $z s  Ž È $z v  Ž È/ %v j    Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y&      "   " ¹z  l      â Ey  °.    € â ™x?    â y'      !   b ¹z  F   
   â x*      ÿ  â r( ÿ       â $r<ÿÿ   ÿ Ž â r? ÿ       â $r ÿÿ   ÿ Ž â x      ÿ  â x      ÿ  â x       ÿ  â x"      ÿ  â x$      ÿ  â r &ÿ   pRð Ä r &ÿ   pRò â$x.&    'Ž Ê/ xÿ   .  È r.   ÿ� È x)ÿ     È | )   pbô Ú G)  P-    € ê y      '   " r ÿ       â $v1ÿ Y  ÿ Ž â x;)   ÿ   â y      %   b z5  X      â $r*ÿÿ   ÿ Ž â z9; q  ÿàÿ â y      &   ¢ 'z y   Ž È $x   ÿ Ž â/ r ÿ       â $z/ }  ÿŽ ä $z	 p  )Ž âx   ÿàÿ â 'z s   Ž âOzÿ z    ä xøÿÿÿÿÀŽ â $z w  ÿŽ â zÿ t    â $z	 |  	Ž Ä $x..   
Ž â r//   ÿàõ â $z |  )Ž â x        â $z v  	Ž â x.!   ÿŽ â $z v  Ž â x:ÿ   .  â %xb   ÿŽ â r=ÿÿ   ÿä Æ %v  € Ž È $z	 p  Ž â r4   ÿà÷ â $vÿ q  ÿ Ž â x.   ÿ   Æ $x8   Ž âr-;   ÿŽ â %x	b   Ž È %x	b   ÿŽ È $x2   ;Ž ä$x3   ;Ž ä$x0   ;Ž ä$x6   ;Ž ä$x7   ;Ž ä %x
   Ž È %x   Ž È r
5   ÿàõ Ê $x   1 Ê �y   • ¨ �y   • è �y   • ( �y   • b r5   ÿàõ È r1   ÿä Ê �y,   • h	 �y   • h	 �y   • h	 �y   • b	 r54   ÿàõ Ê $x1   8 Ê �y‡   • h �y@   • b $tAÿ   ÿ Ž â x   ÿ   ÈO xü  ÿÀŽ â $x   ÿ Ž Æ� zb  ÿàõ ä xü  ÿÀŽ ä xÿ     ä zcÿ  ÿä ä z  ÿà÷ â %v  A Ž â xÿ     â �ybb    ™ ¢  zÿ  ÿäÿ â $x   ÿ Ž âr5   ÿàõ â �ya    ™ â
 %v  A Ž Ä xü  ÿÀŽ â �y`    ™ " r1   ÿä ä z  ÿàõ â �y_    ™ ¢  xÿ     Æ �y\   • ¤  %v  A Ž â x   ÿ   â�yO   • ¨
 �yG    ™ ¢ $vÿ  ÿ â xü  ÿÀŽ È �yC    ™ ¢ z  ÿàõ ä xÿ     Æ $vÿ  ÿ ä %v  A Ž Æ �yI    ™ ¨  �yK    ™ ¢  $x,   ÿ Ž â/ r./   ;àµ ä x   ÿ   äxü  ÿÀŽ Ä r:=   ÿd~ ä xü  ÿÀŽ ä z  ÿàõ ä xc $       ä zÿ  ÿä Æ %v Z  c Ž ä �y^    ™ d $x$   ÿŽ â z  ÿà÷ ä xÿ   ,  â $x   Ž â zÿ  ÿäÿ â $x   ÿ Ž â x   ÿ   â %v  A Ž Ä �y]    ™ ¢  xü  ÿÀŽ â $rÿÿ    Ž â xÿ     â �y,    ™ ¢  xü  ÿÀŽ ä z  ÿàõ â %v  A Ž â �y[   ™ ¢ xÿ     Ä xÿ     â $vÿ  ÿ â r>./   9à© â �yJ    ™ ¢ %v  A Ž Æ �yF   ™ ¢  %v  A Ž â/ rA:=   ÿD~ ä �yZ   ™ ¢  r5   ÿàû Æ �yD    ™ ¢ $xAA$   ÿŽ Æ �yH    ™ ¢  z  ÿà÷ Æ �yB   ™ ¨ �yX   ™ ¢ %v> Z  c Ž â zÿ  ÿäÿ ä �yV   ™ ¢  $x   AŽ ä $x   1Ž â r./   -àë â �yM    ™ ¨  �yA   ™ ¢  r:=   ÿÄþ Æ/ �yT   ™ ¢ %v Z  c Ž Æ �yE    ™ ¨  �yY   ™ ¢ r5   ÿà÷ Æ �yR    ™ ¢ $x$   ÿŽ Æ �yW   ™ ¢  r	1   ÿäÿ ä r   ÿàÿ â �yU   ™ ¢  r./   2àµ Æ �yS   ™ ¢  rP:=   ÿd~ ä r./   3àÙ â �yN    • ¨  �yQ   ™ â rf:=   ÿ¤~ Æ �y>   ™ â r./   0àµ är./   6àË â �yL    • " re:=   ÿd~ ä r./   7àµ â $xgP$   ÿŽ â �y    • "# %v Z  c Ž Æ �yP    ™ b %v Z  c Ž Æ �yh   ™ b %v Z  c Ž â rd:=   ÿ„þ ä �y    ™ b! %v Z  c Ž Æ �y    • b %v Z  c Ž â rc:=   ÿd~ ä �yk    • f $xcc$   ÿŽ ä $xid$   ÿŽ ä $x   cŽ â �yd   ™ h �yc   ™ b ri   ÿàÿ â $x   gŽ Ä �yi   ™ b $xgf$   ÿŽ Æ �yj   ™ d rg   ÿàÿ ä �yg   ™ b $xee$   ÿŽ Æ �yf   ™ b $x   eŽ Æ �yq   ™ h �ye   ™ h �yp   ™ h �yo   ™ h �yn   ™ h �ym   ™ h �yl   ™ h �y    ™ b# xy‡ÿ   ÿÀŽ Æ �y   ™ b 	sˆ y        & �yr    • h �y   ™ h �yw   ™ b xzÿ   ‡  Æ �yv   ™ b xˆˆ   ÿ   Æ �yu   ™ b x‰zÿ   ÿÀŽ Æ �yt   ™ b xˆˆ€   ÿÀŽ â 	sŠ ‰        $ �ys   ™ b xˆˆÿ   ‡xŽ Æ �y   ™ b# x‡‡T  @    Æ �y   ™ h �y~   ™ h �y}   ™ h �y|   ™ h �y{   ™ h �yy   ™ h �yz    ™ h �y†    • b x‡ÿ   ‡  Æ �y    ™ b x�     Æ �yx    • b xÿ   @  â/ $xŠŠ€   ÿ Ž ä �y…   ™ b x”ÿ   ÿÀŽ ä x’‡ÿ   ÿÀŽ â $xŽˆÿŽ â 	s• ”        b $tÿ @€ÿ Ž â �y   ™ b x‘‰€   ŠxŽ Æ �y„   ™ d 	s“ ’        $ �yƒ   ™ b xŒŽ€€€€ìŽ Æ �y‚   ™ h �y�   ™ h �y€   ™ h �y    ™ h �y    • b $x‘‘ÿŽ â x••   ÿ   â/ $t–ÿÿ Ž â x�Œÿþþþÿàÿ Ä x˜  @€    â $x““€   ÿ Ž â �y‹   ™ b xŽ€€€€�ìŽ ä x•”€   •xŽ â �y   ™ b x�ÿþþþÿàÿ Æ �yŠ   ™ b x��ŽøŽ ä x�‘€€€€–ìŽ â �y‰   ™ b xŽ� @€ŽøŽ ä x�‘€€€€˜ìŽ â �yˆ   ™ b x“’€   “xŽ Æ �y‡   ™ b x”�ÿþþþÿàÿ Æ �y   ™ h �y    ™ h �yŒ    • b $tšÿ @€ÿ Ž â x�ÿþþþÿàÿ â $x�•ÿŽ Æ x’‘øŽ äx” @€‘øŽ â $x‘“ÿŽ â x�€€€€–ìŽ äx��€€€€˜ìŽ ä x˜     ä x”ÿþþþÿàÿ Ä x–�ÿþþþÿàÿ ä x“‘€€€€˜ìŽ äx•‘€€€€šìŽ ä x�”�øŽ äx�˜º  ÿ    ä x�– @€�øŽ Ä x”“ÿþþþÿàÿ ä x–•ÿþþþÿàÿ ä x�Ž˜º  ÿ    ä xŽÿÀŽ äx€€€€bÞŽ äO x“’˜º  ÿ    Ä x””‘øŽ äx–– @€‘øŽ ä x‘�ÿÀŽ äx��€€€€aÞŽ ä� rŽ  €ÿàÿ ä xŽ“ÿÀŽ Ä x’˜º  ÿ    ä x““€€€€`ÞŽ â$x�   ‘
Ž â xb€€€€b´Ž ä x‘’ÿÀŽ är“Ž  €ÿàÿ ä x’’€€€€_ÞŽ Ä x�\ÿ   ÿÀŽ ä x€€€€`´Ž â $x`’   ‘
Ž â x––˜º  ÿ    â 	s’ �        b  xŽÿ   \  ä x‘–ÿÀŽ äx––€€€€GÞŽ Ä x””˜º  ÿ    ä x``€€€€_´Ž ä x_Žÿ   ÿÀŽ â $xŽ–   ‘
Ž â x€€€€a´Ž ä xa”ÿÀŽ ä x‘ÿ   O  Ä x””€€€€CÞŽ ä xŽŽ€€€€G´Ž ä xG‘ÿ   ÿÀŽ ä r””a  €ÿàÿ ä x�\T  O    ä x‘’   ÿ   Ä/ x’�˜º  ÿ    â 	sa _        " xC”€€€€C´Ž ä x�ÿ   �  ä x“‘€   ÿÀŽ ä x‘’ÿÀŽ ä x”�˜º  ÿ    Ä x’’€€€€IÞŽ â 	s� G        b x��ÿ   ÿÀŽ ä x““ÿ   \xŽ ä x�”ÿÀŽ â $x’’   ‘
Ž â x””€€€€KÞŽ â 	s‘ �        ¢ $x““ÿŽ â x\     Ä r””�  €ÿàÿ â $t–ÿ @€ÿ Ž â x\“€€€€\ìŽ ä xK”€€€€K´Ž â $x”a€   ÿ Ž â x\\ÿþþþÿàÿ ä x�“€€€€–ìŽ ä x��   ÿ   Ä/ x”_€   ”xŽ ä xI’€€€€I´Ž ä xa\“øŽ â $x\‘€   ÿ Ž âO x’�ÿþþþÿàÿ â $x””ÿŽ â xGG€   �xŽ â $t‘ÿ @€ÿ Ž â x_     Ä x�’ @€“øŽ â $x’GÿŽ â x��€   \xŽ ä xG”€€€€_ìŽ äx\”€€€€‘ìŽ ä x“  @€    ä xGGÿþþþÿàÿ Ä x�\ÿþþþÿàÿ â $x‘�ÿŽ â x\’€€€€_ìŽ â$t�ÿÿ Ž â x“’€€€€“ìŽ ä x_G”øŽ äxG\ÿþþþÿàÿ ä x�� @€”øŽ Ä x““ÿþþþÿàÿ ä x”  @€    ä x�‘€€€€�ìŽ ä x\G’øŽ äxaa˜º  ÿ    ä xG“ @€’øŽ Ä x’‘€€€€”ìŽ ä x”�ÿþþþÿàÿ ä x�aÿÀŽ äxaa€€€€^ÞŽ Ê$xaa   �
Ž â x�_˜º  ÿ    ä x��˜º  ÿ    ä x_�ÿÀŽ äx��€€€€]ÞŽ ä x–’ÿþþþÿàÿ ä x’�ÿÀŽ â$x��   _
Ž â x��€€€€,ÞŽ â &r_b[   ÿ   â x��˜º  ÿ    Ä x””‘øŽ ä r��’  €ÿàÿ â &r_F   _   â x’�ÿÀŽ ä x^a€€€€^´Ž ä x��€€€€JÞŽ ä xa”˜º  ÿ    Ä x–– @€‘øŽ â &r_Z   _   â r��’  €ÿàÿ ä x,�€€€€,´Ž ä x’aÿÀŽ äx�–˜º  ÿ    ä xaa€€€€DÞŽ â &r[^[   ÿ   Ä &r_`B   _   â x”�ÿÀŽ â $xaa   ’
Ž â x��€€€€HÞŽ â &r[,F   [   â x]�€€€€]´Ž â &r_CX   _   â x\\˜º  ÿ    ä xJ�€€€€J´Ž â &r[]Z   [   â r��”  €ÿàÿ â &r_ŽV   _   â x�G˜º  ÿ    Ä xGa€€€€D´Ž â &rDbA   ÿ   âxa\ÿÀŽ â &rA^A   ÿ   â xH�€€€€H´Ž â &r[JB   [   â x\\€€€€MÞŽ â &r_IT   _   â x��ÿÀŽ â &rDY   D   âx��€€€€EÞŽ â &rB,Y   A   â r\\a  €ÿàÿ â &r[GX   [   Ä &r_KR   _   ä &rDW   D   âxM\€€€€M´Ž â $x��   �
Ž ä &rB]W   B   â rA__   ÿ� â &r[HV   [   ä &rD`U   D   âxE�€€€€E´Ž â &rBJU   B   â x@ÿ   @  â &r[MT   [   â xAÿ   A  â &rDCS   D   Ä &rBGS   B   ä &rRER   [   ä &rDŽQ   D   ä&rQHQ   B   ä $rA@_   AŽ â rBRR   ÿ� â &rDI>   D   Ä &rQM>   Q   â rAAA   ÿ� â &rSKP   D   âxOÿ   O  â &rPEP   Q   â xBÿ   B  â 0rLÿL    A   â xAÿ   A  â 0rQÿ    A   ä $rBOR   BŽ â ErA A       Ä  rLQ     A  ä rBBB   ÿ� ä #r%A   %   â &rbc   ÿ   âxBÿ   B  â 0rNÿN    A   ä &rc^c   ÿ   ä 0rAÿ    A   â r>SS   ÿ� â &ri      âErB B       â &rc,i   c   â  rQQN     A  â &rh      âx>ÿ   >  â &rc]h   c   â #r$QB   $   â &rBbj   ÿ   Ä &r`g      ä $r>@S   >Ž ä &rcJg   c   ä &rBq   B   ä &rCf      â r>>>   ÿ� â &rcGf   c   Ä &rBp   B   ä &rŽe      âx>ÿ   >  â &rcHe   c   ä &rB`o   B   ä &rId      â Er> >       â &rdMd   c   â  rLA     A  â &rBCn   B   Ä &rQK      ä&rdE   d   â #r#>   #   â &rBŽm   B   â rQQ   ÿ� ä rdd   ÿ� â &rBIl   B   â xÿ     ä xÿ     â &rSK   B   Ä $r@Q   Ž ä $rOd   Ž â rSS   ÿ� ä r   ÿ� ä xÿ     ä r   ÿ� â &rj^j   ÿ   â xÿ     â 0rkÿk    A   â xÿ     â $rS@S   Ž â Er        â &rj,q   j   â  rLk     A  â Er        â  rkNk     A  â &rj]p   j   â rSSS   ÿ� â #r!   !   â &rb   ÿ   â #r k       â &rjJo   j   â xSÿ   S  â 0rÿr    A   Ä &r^   ÿ   ä &r      â ErS S       â &rjGn   j   â  rL     A  â &r,      ä &rw      â #rS      â &rjHm   j   Ä &r`v      ä &rjMl   j   ä &rw]w      ä &rCu      ä &rjE   j   ä &rvJv   w   Ä &rŽt      â rjj   ÿ� â &ruGu   v   ä &rIs      â xÿ     â &r>b   ÿ   ä &rtHt   u   ä &rK      Ä &r>   >   ä $rOj   Ž â r   ÿ� â &rtMs   t   ä &r^   ÿ   ä &r>~   >   â r   ÿ� â &rtE   t   â xÿ     â &r,      Ä &r>`}   >   â xÿ     â &r]~      â rtt   ÿ� â &r>C|   >   ä $r@   Ž â Er        â &rJ}      â xÿ     â &r>Ž{   >   â  rN     A  â r   ÿ� â &rG|      Ä &r>Iy   >   â #r      â $rOt   Ž â xÿ     â 0rÿx    A   ä &rH{      â r   ÿ� â &rKz   >   â Er        â  rL     A  â &rMy      Ä r   ÿ� â #r      â &rb   ÿ   ä&r^   ÿ   âxÿ     â &r^^   ÿ   ä &rEz      ä &rb   ÿ   Ä &r…      ä &r^,‹   ^   ä $r@   Ž â r   ÿ� â &r,…      ä &r‹      ä &r„      Ä &r^]Š   ^   âxÿ     â &r]„      ä &rŠ      â xÿ     â &r`ƒ      ä &r^J‰   ^   â r   ÿ� â &rJƒ      Ä &r`‰      ä $rO   Ž â Er        â &rC‚      ä &r^Gˆ   ^   â rN     A  â &rG‚      â xÿ     â &rCˆ      Ä 0rÿ†    A   â r   ÿ� â &rŽ�      ä &r^H‡   ^   â#r      â &rH�      â Er        â &rŽ‡      â  rL     A  â &rI€      â xÿ     â &r^M   ^   Ä &rM€      ä &rI      â #r        â &rK      â Er        â &r^E   ^   â rN     A  â &rE      ä &rK      â #r?   ?   â rDPP   ÿ� Ä r   ÿ� ä r   ÿ� ä r^^   ÿ� ä r   ÿ� ä xDÿ   D  ä xÿ     Ä xÿ     ä xÿ     ä xÿ     â $rDOP   DŽ â x))   ÿàÿ â $r@   Ž ä $rO   Ž ä$rO^   Ž Ä $r@   Ž â | )   pbô ä rDDD   ÿ� ä r   ÿ� ä r   ÿ� ä r   ÿ� ä r   ÿ� â 0rÿ    A   â xDÿ   D  â 0rÿŒ    A   â xÿ     Ä xÿ     ä xÿ     ä xÿ     â  rL     A  â rN     A  âErD D       â  rN     A  âEr        â  rANA     A  â Er        â  rL     A  â Er        Ä Er        ä x55  ÿà÷ â #r"AD   "   â #r<   <   â #r+   +   â #r*   *   â #r(   (   â x--@   ÿàÿ ä x22@   ÿàÿ Ä x33@   ÿàÿ ä x00@   ÿàÿ ä x66@   ÿàÿ ä x77@   ÿàÿ ä x99@   ÿàÿ ä x;;@   ÿàÿ Ä r1ÿ1   ÿäÿ â D)       À â Gy   Õÿÿÿÿƒ ê Ay        € ê Ey  0    € â G‰      € ê r&'   ÿHŽ Ê ˆs % øÿ H   è ˆs $€øÿ H   è ˆs # ùÿ H   è ˆs "€ùÿ H   è ˆs ! úÿ H   è ˆs  €úÿ H   è ˆs  ûÿ H   è ˆs €ûÿ H   è ˆs  üÿ H   è ˆs €üÿ H   è ˆs   ýÿ H   è ˆs ?€ýÿ H   è ˆs < þÿ H   è ˆs +€þÿ H   è ˆs ( ÿÿ H   è ˆs *€ÿÿ H   ä Ay        € ê {          ì M        € ê „y'     H   b Ey  @    € æ „y' €   H   ( „y'    H   ¨ y      %   ä $x   ÿ Ž â� !r%       È/ z  r  p`ð â !r$$       Æ r 'ÿ   pVp  â ‰      " !r##       ÆO ‰$      h ‰#      ¨ ‰      '   â ˆ        â !r       â !r$       È/ ‰      " !r#       ÆO ‰      h ‰	      ¢ !r       â !r       È/ ‰ €    " !r		       ÆO ‰ €    h ‰
	 €    ¢ !r       â x   ÿàÿ â !r       Æ/ z  r  p`ò â ‰ @    " !r
	
       äO x '   p"ò â ‰ @    h ‰      &   ¨ ‰
 @    (       &   b         Æ       '   b !r       Æ „y	' €  H   â !r       Æ/ ‰      " $Š x  Ž ÆO ‰      b $Š ~  Ž â� !r
       â$ x  Ž È‰      ¢ $ ~  Ž â !r       â %† j   Ž â !r       Æ/ % j   Ž â †‰     è †    â !r       âO G	  �     € ê y      &   ( y      '   b $z x  Ž È $z ~  Ž Ê/ z r  ÿàõ Ê $rÿÿ   ÿ â z j  ÿ„ È z k   Ê †y     è Ay        € ê !r		"       â „y'    H   â Ey  �    € æ ‰	      ( „y' €  H   è „y
'    H   â !r	       Ê ‰      $ !r       Ê ‰ €    $ !r       Ê ‰ @    $ !r       Ê ‰      $ !r	       â G™  °     € ê y      &   " ¹z  r      ä �x   ?àÿ â y      '   b $z x  Ž È $z ~  Ž Ê/ |   ÿàõ È rÿÿ   ÿä ä z j  ÿ„ È z k   Ê †y 	    è Ay        € ê „y' €  H   b !r        â !r

       â !r!       â „y'    H   ¢ z#  r      Æ „y' €  H   ä $x!#   ÿŽ ä„y'    H   " $x#   ÿŽ Æ „y' €  H   h „y'    H   è „y' €  H   è „y'    H   è ‰      ( ‰	
      ( ‰      " !r       Æ/ „y'' €  H   â !r       âO !r       â� !r        È‰      b !r??       â!r       ä ‰      " !r	
	       â !r<       ä ‰
      ¢ !r++       â !r       Ä ‰      â !r(       Æ ‰?      ( y       &   b !r       Æ/ ‰      h ‰      b !r       â !r''*       ä ‰+      " !r

       ÆO ‰      ¢ !r       Æ� ‰      â !r??       Æy      '   " $z   x  Ž Æ‰	      b !r       â/ !r       ä ‰
      b !r++       Æ ‰      " !r       ÆO ‰      ¢ !r       Æ� ‰      â $z  ~   Ž Æ‰?      " !r	       Æ‰ €    b r#    ÿ„ Æ ‰      b !r
       â/ $rÿÿ   ÿ â z j  ÿ„ â ‰+      b !r       â ‚ !   ÿà÷ ä ‰      " !r       âO z k   â $‚ÿÿ   ÿŽ â ‰ '      ¢ !r       â� Š
 j  ÿ† Ä ‰ €    â !r??       âŠ k  � ä ‰	 €    " !r       â‚    ÿàù â !r       â ‰ €    d ‚*ÿÿ   ÿä â !r++       â/ ‰ €    b Š j  ÿˆ â !r       Ä ‰ €    " Š k  * â !r''        äO ‰ €    ¢ !!   ÿàÿ â !r       ä� ‰ ? €    â $x#   ÿŽ â !r	       Ä‰ @    " r	#    ÿŠ är    ÿàõ â ‰" €    " !r       ârÿÿ   ÿäÿ ä ‰$+ €    b rÿÿ   ÿä â !r       â/ z j  ÿ„ â ‰& €    b !r       â z	 j  ÿŠ Ä ‰('      " !r       âO z k   ä ‰ @    ¢ !r ?        â� z		 k  � ä ‰ @    â !r       â!r""       Ä ‰ @    " !r$+$       Æ‰ @    b $x+#   ÿŽ â !r&&       ä/ ‰ @    d ‚ +   ÿà÷ â !r('(       â ‰ @    " !r       âO    ÿàÿ Ä ‰#" @    ¢ !r       â�     ÿàõ ä ‰  @    â ‚ÿÿ   ÿäÿ ä Š j  ÿ† â ‰%$ @    â $*ÿÿ   ÿ â  j  ÿ„ â!r       â‰)( €    " Š k  � â !r       â k  * â ‰'& @    b ! !   ÿàõ â !r       â/ !r       â ‰      " *ÿÿ   ÿä Ä ! j  ÿ„ â‰      b !r#"#       äO ! k  * â !r         â� ‰      ¢ !r%$%       Æ ‰      â !r)()       Æ‰      " !r'&'       Æ‰      b !r!       Æ ‰      " !r       Æ/ ‰       h ‰#      b !r       ÆO ‰%      ¢ !r       Æ� ‰$) @    â !r       Æ‰"'      " !r       â!r       ä †‰ !    â !r        Æ/ †    â !r##       Æ †‰     â !r%%       ÆO †     â !r$)$       Æ� †‰     â !r''"       Æ†    è †‰ 
    è †     è †‰ #    è † %   è ‰$      h  †‰ '    â M™        € ê x+   ÿàÿ â !r$       Æ/ r    ÿàñ È r ÿÿ   ÿä  ä z j  ÿ€ È z k     Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y       "   " ¹z  l      â Ey  À+    € â ™x?    â y!      !   b ¹z  F   
   â x$      ÿ  â r" ÿ       â $r4ÿÿ   ÿ Ž â x      ÿ  â $r6ÿÿ   ÿ Ž â x      ÿ  â x      ÿ  â x      ÿ  â r  ÿ   pRð Ä r  ÿ   pRò â$x'     !Ž Ê/ x ÿ   '  È r '   ÿ� ä r  ÿ       ä x#ÿ     È | #   pbô Ú G)  p*    € ê y      '   " $rÿÿ   ÿ Ž â xøÿÿÿÿÀŽ â $rÿÿ   ÿ Ž â y	      %   b $x3#   ÿ Ž â z+  Y      â $x''   
Ž â y      &   ¢ $v-ÿ X  ÿ Ž â z13 q  ÿàÿ â $r$ÿÿ   ÿ Ž â x'!   ÿŽ Ä x2ÿ   '  â 'z y   Ž â x	   ÿ   Æ/ $z( }  ÿŽ ä 'z s   Ž âO zÿ z    Æ $z w  ÿŽ â zÿ t    â $z p  #Ž âx   ÿàÿ â $z |  #Ž âr((   ÿàõ â $z |  Ž â z  q      â $z v  Ž âr5ÿÿ   ÿä â $z v  Ž â r)3   ÿŽ â $tÿ   ÿ Ž Ä $z	 p  Ž ä %v  € Ž È %xb   ÿŽ È $x'   ÿ Ž â r,   ÿà÷ â $x&   3Ž ä$x*   3Ž ä $x0   Ž ä%x	b   Ž È %x	b   ÿŽ È $x.   3Ž ä$x/   3Ž ä %x
   Ž È %x   Ž È r
-   ÿàõ Ê $x   + Ê �y   • ¨ �y   • è �y   • ( �y   • b r-   ÿàõ È r+   ÿä Ê �y:   • h �y   • h �y7   • h �y8   • b r-,   ÿàõ Ê $x+   0 Ê �yd   • h �y9   • b $t=ÿ   ÿ Ž ä $tMÿ$   ÿ Ž â x   ÿ   äO x`ÿ     Ä xü  ÿÀŽ Æ %v``  = Ž â z  ÿàõ Æ $x   ÿ Ž â� zÿ  ÿä â �y``    ™ â r-   ÿàõ ä xü  ÿÀŽ â �yc    ™ â  r+   ÿä â $x   ÿ Ž âz  ÿàõ ä/ xÿ     â �yW   • " xü  ÿÀŽ Ä zÿ  ÿä â %v  = Ž â z  ÿàõ â �y<   • ä $x   ÿ Ž ä�y]    ™ b zÿ  ÿä Æ �y_    ™ â xü  ÿÀŽ ä xÿ     â �y\    ™ â zZ  ÿàõ ä xÿ     â %v  = Ž È $v[ÿ  ÿ â �yb    ™ â  %v  = Ž Æ/ �yZZ    ™ è �ya    ™ âD x:   ÿ   È xü  ÿÀŽ È zX  ÿàõ ä x   ÿ   Æ $vYÿ  ÿ â xü  ÿÀŽ ä x:ÿ   :  ä �yXX    ™ â z  ÿàù ä rH'(   3àµ â %v:  = Ž Æ r25   ÿd~ âO $vÿ  ÿ â x7   ÿ   â �y[    ™ ¢  %vHH Z  M Ž Æ �yY    ™ ¢ $x$   ÿŽ â xü  ÿÀŽ È rII   ÿàÿ ä z:  ÿàõ Æ �yEH   ™ ¤  $v;ÿ  ÿ â xÿ     â �yDH   ™ ¢ x8   ÿ   Æ �y::    ™ ¢ %v  = Ž Æ �yBH   ™ ¢ xÿ   7  ä xü  ÿÀŽ â �yCH   ™ ¤  %v  = Ž â z  ÿàõ â �y^    ™ ¨  �y>H   ™ ¢ $vÿ  ÿ Æ �yAH   ™ ¢ xÿ   8  Æ �y7    ™ ¨  �y?H   ™ ¢ %v  = Ž Æ �y@    ™ ¨ �y=H    ™ ¢ r'(   1àÙ Æ �y;    ™ ¢  r25   ÿ¤~ â %v Z  M Ž â r-   ÿà÷ ä r'(   &àÙ â �yHH    • ¢ $x$   ÿŽ â r+   ÿäÿ ä r25   ÿ¤~ Ä r   ÿàÿ â �y8    • ¢  r        â %v Z  M Ž â rF-   ÿàõ Æ $x$   ÿŽ â �yU   ™ ¢  $rÿÿ    Ž ä $x   Ž â �yJ   ™ ¢  $xG	   + â r'(   *àÙ är'(   )àµ â �yK   ™ ¢  rO25   ÿ¤~ Æ �yI   ™ ¢" rN25   ÿd~ ä r'(   .àÙ â�yFF    • ¢ r'(   /àµ Æ �yS   ™ ¢  %v Z  M Ž Æ �yT   ™ ¨  �yG   ™ ¢  %v Z  M Ž Æ �yQ   ™ ¢ %v Z  M Ž Æ �yR   ™ ¢  %v Z  M Ž â rM25   ÿ¤~ ä�yP   ™ ¢  $xeO$   ÿŽ â rL25   ÿd~ ä �yO   ™ ¢ $xiM$   ÿŽ Æ �yM   ™ ¢ $xgN$   ÿŽ ä $xkL$   ÿŽ â �yN    ™ ¨  �yL   ™ ¨ �yV   ™ ¢ re   ÿàÿ Æ �ye    ™ ¨ �y    • ¢ $x   gŽ Æ �yf   ™ ¨  �yh   ™ ¨ �yn   ™ ¨ �y   ™ ¨ �ym   ™ ¨  �yq   ™ ¨  �yg    • ¢ $x   kŽ Æ �yl   ™ ¨  �yp   ™ ¨  �yk   ™ ¨  �y   ™ ¢" xudÿ   ÿÀŽ Æ �yj   ™ ¢  ri   ÿàÿ Æ �y   ™ ¢ 	s{ u        & �yi    ™ ¨  �yt   ™ ¨ �yr   ™ ¢ x|ÿ   9  Æ �yo    • ¢  xzdT  9    Æ �ys    ™ ¢ xˆ|ÿ   ÿÀŽ â $t‚ÿÿ Ž ä �yy    • ¢ xÿ   d  ä x|ÿ   z  â �y   ™ ¢  x„ÿ   ÿÀŽ ä x{{   ÿ   â 	s‰ ˆ        " x†|ÿ   ÿÀŽ â �yx   ™ ¢  x{{€   ÿÀŽ Ä xŠ  @€    â �yw   ™ ¤  	s€ „        b x{ÿ   dxŽ â �y~   ™ ¨ �yv   ™ ¤ 	s‡ †        " $x�ÿŽ â x‰‰   ÿ   â �yd   ™ ¦  x‚�€€€€‚ìŽ â�yu   ™ ¢  $x…€€   ÿ Ž â/ xƒ�€€€€ŠìŽ Ä x‚‚ÿþþþÿàÿ â �y   ™ ¢  x…„€   …xŽ â $x‡‡€   ÿ Ž âx‰ˆ€   ‰xŽ Æ $x……ÿŽ â x„ƒÿþþþÿàÿ â �y   ™ " xƒ‚�øŽ ä x‡†€   ‡xŽ â $t†ÿ @€ÿ Ž â x‚     â �y    ™ " $x‰‰ÿŽ Æ �yz    • " xˆ     Æ �y}   ™ ( �y|   ™ ( �y{   ™ ( �y   ™ ( �y   ™ ( �y    ™ ( �y€    • " x„„ @€�øŽ Ä x�‰€€€€ˆìŽ ä x…€€€€‚ìŽ ä/x…€€€€†ìŽ ä x†‰€€€€ŠìŽ ä xÿþþþÿàÿ ä xŠ†ÿþþþÿàÿ Ä x‚ÿþþþÿàÿ ä x†…øŽ äxˆ�ÿþþþÿàÿ ä xƒ˜º  ÿ    ä xƒ„˜º  ÿ    ä x…‚ @€…øŽ â $x‡‡ÿŽ â x‚ˆ‰øŽ â $tˆÿÿ Ž â xÿÀŽ Ä x€€€€cÞŽ ä� x�Š @€‰øŽ ä x„ƒÿÀŽ äxŠ  @€    ä xƒƒ€€€€`ÞŽ â $x   
Ž â xˆ‡€€€€ˆìŽ Ä x‰‡€€€€ŠìŽ ä rƒ„  €ÿàÿ ä xˆˆÿþþþÿàÿ ä xŠ‰ÿþþþÿàÿ ä x€€€€`´Ž ä x††˜º  ÿ    Ä x`…˜º  ÿ    ä xˆˆ‡øŽ ä x€€€€c´Ž ä xŠŠ @€‡øŽ ä xc†ÿÀŽ ä x‡Wÿ   ÿÀŽ Ä x††€€€€]ÞŽ äx„`ÿÀŽ äxƒ`€€€€_ÞŽ ä x…ˆ˜º  ÿ    â $x`†   c
Ž â r„ƒ„  €ÿàÿ â 	s‡ ‡        " xƒÿ   <  Ä x‰ÿ   W  ä x†…ÿÀŽ äx……€€€€\ÞŽ ä x‹ƒÿ   ÿÀŽ ä xc‰ÿ   ÿÀŽ ä x``€€€€]´Ž â $x]…   †
Ž â x„„€€€€_´Ž â 	s… ‹        b xƒ‚˜º  ÿ    Ä x_Š˜º  ÿ    ä xˆ�˜º  ÿ    Æ 	s‰ c        â x†ƒÿÀŽ ä x‚_ÿÀŽ ä xƒƒ€€€€ZÞŽ ä x__€€€€bÞŽ ä x�ˆÿÀŽ äxˆˆ€€€€aÞŽ â $xƒƒ   †
Ž â x]]€€€€\´Ž â $x‡‡€   ÿ Ž â r__‚  €ÿàÿ Ä x\WT  <    ä rˆˆ�  €ÿàÿ ä xb_€€€€b´Ž ä x\ÿ   \  ä x_ƒ€€€€Z´Ž ä x‚‡€   ÿÀŽ Ä xZˆ€€€€a´Ž â $xˆ…€   ÿ Ž â/ x†‰   ÿ   ä� x…\ÿ   ÿÀŽ ä x‚‚ÿ   WxŽ ä x†c€   †xŽ â 	s‡ …        " $taÿ @€ÿ Ž Ä $x‚‚ÿŽ â xW     â $x††ÿŽ â xc     ä x\‚€€€€aìŽ â$tƒÿ @€ÿ Ž â xW‚€€€€WìŽ ä xc†€€€€cìŽ ä xˆ‹€   ˆxŽ Ä xa\ÿþþþÿàÿ ä xWWÿþþþÿàÿ ä x�cÿþþþÿàÿ â $xˆˆÿŽ â xca @€‚øŽ â $taÿÿ Ž â xƒ†€€€€ƒìŽ ä x\‡   ÿ   Ä x‡  @€    ä xWW‚øŽ ä xƒƒÿþþþÿàÿ ä x\…€   \xŽ ä xaˆ€€€€aìŽ äx‚ˆ€€€€‡ìŽ Ä xWW˜º  ÿ    â $x…\ÿŽ â x��†øŽ äxƒƒ @€†øŽ â $t†ÿÿ Ž â x‡aÿþþþÿàÿ ä x‰‚ÿþþþÿàÿ ä x‚WÿÀŽ Ä xaW€€€€XÞŽ ä x\‡ˆøŽ ä x‡…€€€€†ìŽ â $xaa   ‚
Ž â x†�˜º  ÿ    ä x‚c˜º  ÿ    ä x�†ÿÀŽ Ä xc‚ÿÀŽ ä x††€€€€YÞŽ äO x‚‚€€€€[ÞŽ ä xW‰ @€ˆøŽ â $x††   �
Ž â xˆ‡ÿþþþÿàÿ ä r‚‚c  €ÿàÿ Ä xˆˆ…øŽ ä x[‚€€€€[´Ž ä x‚†€€€€Y´Ž â &rYE   ÿ   â xXa€€€€X´Ž ä xŠ  @€    ä xaˆ˜º  ÿ    â &rYD   Y   â x‰…€€€€ŠìŽ Ä xˆaÿÀŽ äxaa€€€€:ÞŽ â &rY`B   Y   â xŠ‰ÿþþþÿàÿ ä xƒƒ˜º  ÿ    â $xca   ˆ
Ž â xŠŠ @€…øŽ â &ra„C   Y   â x†ƒÿÀŽ â&rEXE   ÿ   â xƒƒ€€€€^ÞŽ â &ra]>   a   â xŠŠ˜º  ÿ    â &rE[D   E   â rƒƒ†  €ÿàÿ â &rDbA   a   â xYc€€€€:´Ž Ä x�ŠÿÀŽ ä xc\˜º  ÿ    ä xŠŠ€€€€7ÞŽ â &rE‚B   E   â x^ƒ€€€€^´Ž â &rD_?   D   â x\cÿÀŽ ä r:Š�  €ÿàÿ Ä xcc€€€€@ÞŽ â &rE^C   E   â xBW˜º  ÿ    â &rDZ=   D   â x::€€€€7´Ž â $x7c   \
Ž â xCBÿÀŽ â&rEY>   E   â xBB€€€€;ÞŽ ä r>DD   ÿ� Ä x77€€€€@´Ž ä r\BC  €ÿàÿ â &rB:A   E   â x9ÿ   9  ä x@ÿ   >  â &rB7?   B   â x>\€€€€;´Ž Æ $r@9D   @Ž ä &r?>=   B   Æ r@@@   ÿ� â 0r8ÿ8    A   â r=??   ÿ� â 0r;ÿH    A   â x@ÿ   @  â &rBU   ÿ   âxAÿ   =  â &rUXU   ÿ   â Er= @       â  r@8;     A  ä &rU[J   U   Ä #r@=      â &r@I   ÿ   ä &rU‚K   U   ä &rBJ   B   ä&rU^G   U   ä &r@S   @   ä &rB`K   B   Ä &rUYT   U   ä &r@`Q   @   â x<ÿ   <  â &rB„G   B   ä &rU:R   U   ä &r@„O   @   ä &r=]T   B   Ä $rA<?   AŽ ä &rU7P   U   ä &r?]M   @   ä &r=bR   =   ä &rU>N   U   ä &r?bL   ?   â rAAA   ÿ� â &r=_P   =   Ä &rB_V   ?   â r?UU   ÿ� â &r@ZN   =   âxAÿ   A  â 0rFÿF    A   â x?ÿ   ?  â &rBZe   B   â r=@@   ÿ� ä ErA A       â  r;;F     A  â $r?<U   ?Ž â x=ÿ   =  Ä #r;A      â r;BB   ÿ� ä r???   ÿ� â $r=9@   =Ž â x@ÿ   ;  â 0r;ÿ    A   â x?ÿ   ?  È  r8;     A  â Er? ?       â  r;F;     A  È #r;?      â &r;f   ÿ   âr===   ÿ� â $r@9B   @Ž ä &r;h   ;   âx=ÿ   =  â &rB   ÿ   â r@@@   ÿ� â &r;`n   ;   â Er= =       â &r?q   B   â x@ÿ   @  â &r;„m   ;   Ä 0rgÿg    A   â #r=      â &r?`p   ?   â Er@ @       â &r;]l   ;   â  r8g     A  â &r?„   ?   ä &r;bk   ;   â #r@      â &rX   ÿ   Ä &r?]   ?   ä &r;_j   ;   ä &rq[q      ä &r?bt   ?   ä &rZi   ;   ä &r?_r   ?   Æ r   ÿ� â &rq‚p   q   ä &r@Zs   ?   â xÿ     â &rfXf   ÿ   ä &rq^   q   â r@@   ÿ� â $r9   Ž ä &r=[h   f   â xÿ     Ä r   ÿ� â &r=‚n   =   ä $r9@   Ž â xÿ     â &r=^m   =   ä 0roÿo    A   â r   ÿ� â &r=Yl   =   â Er        Ä  r8o     A  â &r=:k   =   â xÿ     â 0ryÿy    A   â #r      â &r   ÿ   â Er        â &r=7j   =   â  r8y     A  â &rDXI   ÿ   ä&rX   ÿ   Ä &rx      ä &rXXd   ÿ   â#r        â &r=>i   =   ä &rd   ÿ   ä &rD[S   D   ä&r[x      Ä &r`w      ä &rX[~   X   ä &rqY   q   â r==   ÿ� â &r~      ä &rD‚Q   D   ä&rw‚w      Ä &r„v      ä&r‚}   X   â	xÿ     â &r}`}      ä &rD^O   D   ä&rv^v   w   ä &r]u      Ä &r^^|      ä&r|„|   }   ä &rMYM   D   ä$r<=   Ž ä &rvYu   v   ä &rb      Ä &r^Y{   ^   ä&r|]{   |   â r   ÿ� â &rM:L   M   ä&rq:t   q   ä&rv:   v   ä &r_      Ä &r^:   ^   ä&r|b   |   â xÿ     â &rM7V   M   ä&rq7r   q   ä&r7   v   ä &rZ      Ä &r77   ^   ä&r_   |   â Er        â &r7>   7   â r   ÿ� â &rM>e   M   â  roFo     A  â &rq>s   q   ä&r>      Ä &rZ      â xÿ     â #ro      â rAMM   ÿ� ä r;qq   ÿ� ä r   ÿ� ä r77   ÿ� ä r   ÿ� â $r9   Ž â xAÿ   A  Ä x;ÿ   ;  ä xÿ     ä xÿ     ä xÿ     â $rA<M   AŽ âx##   ÿàÿ â $r;<q   ;Ž ä$r<   Ž Ä $r<7   Ž ä $r9   Ž â | #   pbô ä rAAA   ÿ� ä r;;;   ÿ� ä r   ÿ� ä r   ÿ� Ä r   ÿ� ä r   ÿ� â 0rÿz    A   â xAÿ   A  â 0rÿ€    A   â x;ÿ   ;  ä xÿ     ä xÿ     Ä xÿ     ä xÿ     â  r8     A  â rF     A  âErA A       â  rF     A  âEr; ;       â  rgFg     A  âEr        â  ryFy     A  â Er        â  r8     A  â Er        Ä Er        ä x--  ÿà÷ â #rgA      â #r6y;   6   â #r4   4   â #r%   %   â #r$   $   â #r"   "   â x&&@   ÿàÿ â $r+ÿÿ   +Ž â x**@   ÿàÿ Ä x))@   ÿàÿ ä x..@   ÿàÿ ä x//@   ÿàÿ ä x11@   ÿàÿ ä x33@   ÿàÿ â D)       À â Gy  pØÿÿÿÿƒ ê Ay        € ê Ey      € â G‰  ð     € ê $x À  !Ž Ê ˆs  ùÿ H   è ˆs €ùÿ H   è ˆs  úÿ H   è ˆs €úÿ H   è ˆs  ûÿ H   è ˆs €ûÿ H   è ˆs  üÿ H   è ˆs €üÿ H   è ˆs   ýÿ H   è ˆs 6€ýÿ H   è ˆs 4 þÿ H   è ˆs %€þÿ H   è ˆs " ÿÿ H   è ˆs $€ÿÿ H   ä Ay        € ê {          ì M        € ê „y!     H   b Ey  @    € æ „y! €   H   ( „y!    H   ¨ y      %   ä x   ÿ   â� !r       Æ/ z  r  p`ð â !r       Æ r !ÿ   pVp  â ‰      " !r       ÆO ‰      h ‰      ¨ ‰      '   â $„ÿ   ÿ Ž â !r       â !r       È/ ‰      " !r       ÆO ‰      h ‰	      ¢ !r       â !r       È/ ‰ €    " !r		       ÆO ‰ €    h ‰
	 €    ¢ !r       â x   ÿàÿ â !r       Æ/ z  r  p`ò â ‰ @    " !r
	
       äO x !   p"ò â ‰ @    h ‰      &   ¨ ‰
 @    (       &   b         Æ       '   b !r       Æ „y	! €  H   â !r       Æ/ ‰      " $Š x  Ž ÆO ‰      b $Š ~  Ž â� !r
       â$ x  Ž È‰      ¢ $ ~  Ž â !r       â %† j   Ž â !r       Æ/ % j   Ž â †‰     è †    â !r       âO G	  �     € ê y      &   ( y      '   b $z x  Ž È $z ~  Ž Ê/ z r  ÿàõ Ê $rÿÿ   ÿ â z j  ÿ„ È z k   Ê †y     è Ay        € ê !r		       â „y!    H   â Ey  à    € æ ‰	      ( „y! €  H   è „y!    H   è „y! €  H   è „y! €  H   è „y!    H   â !r	       Æ „y! €  H   è ‰      ( „y	!    H   â !r       Ê ‰ €    $ !r       Ê ‰ @    $ !r       Ê ‰      $ !r       â G™  °     € ê y      &   " ¹z  r      ä �x   ?àÿ â y      '   b $z x  Ž È $z ~  Ž Ê/ |   ÿàõ È rÿÿ   ÿä ä z j  ÿ„ È z k   Ê †y     è Ay        € ê „y!    H   â !r
	        â !r       â !r66       â „y! €  H   â !r44       â !r       â !r       â ‰
      ¢ !r       â !r%%       Ä ‰      â $v!ÿ r  ÿ Ž Æ/ ‰6      b $x !   ÿŽ Æ ‰4      ( ‰      ( ‰       h ‰      b !r
       âO !r	       Ä� ‰%      ¢ !r"       â !r6       ä/ y      &   b !r4       â !r$       ä ‰      " !r       Æ‰      â !r        Æ‰      " !r       Æ ‰      b !r%%       ÆO ‰      ¨ ‰       ¢ $z x  Ž Æ/ ‰      b !r       Æ ‰      " !r       Æ� ‰
	      â !r       Æy      '   " !r       Æ‰%      b !r       ÆO ‰ €    ¢ !r         Æ ‰      ¢ !r       Æ/ ‰ €    b !r       Æ ‰      " !r	
       Æ� ‰  €    â $z# ~  Ž Æ‰ €    " !r%       âr!#   ÿ„ â !r       âO ‰	 €    ¤ rÿÿ   ÿä â !r       â ‰ €    b z j  ÿ„ â !r       Ä/ ‰ €    b z k   â !r       ä ‰ €    " ‚#    ÿàõ â !r         ä� ‰ €    â r!#   ÿˆ â $‚ÿÿ   ÿ â !r       â‰ €    " r"ÿÿ   ÿä Ä z j  ÿˆ â‰  @    " !r
	       äO z k  " â !r       âŠ j  ÿ„ â ‰ @    ¢ $x"!   ÿŽ â !r       â/ Š k   â ‰
 @    b $x!   ÿŽ â !r       Ä ‰ @    " !r       â� ‚	#   ÿà÷ ä ‰ @    â !r       Æ‰ @    " $‚$ÿÿ   ÿŽ â !r         â     ÿàÿ â ‰ @    b Š	 j  ÿ† Æ ‰ @    b !r       âO    ÿàÿ ä ‰ @    ¢ !r
       â/ Š		 k  $� ä ‰  @    b !r       â r"#   ÿàù â !r       â� ‰       " !#   ÿà÷ â !r       Ä‰      â rÿÿ   ÿä â $$ÿÿ   ÿŽ â !r       â‰      " z j  ÿˆ â !r       ä ‰      b 
! j  ÿ† â !r       ÄO ‰      ¢ "#   ÿàõ â !r         ä/ ‰      b z k   ä ! k  $� â ‰      b #ÿÿ   ÿä â !r%        â " j  ÿ„ â ‰      " !r       Ä� " k  # â ‰      â !r       Æ‰!       " !r       â!r       äO †‰ %    â !r       Æ/ †    â !r       Æ †‰     â !r       Æ † 
    â !r       Æ� †‰     â !r! !       Æ†    è †‰     è †     è †‰     â M™        € ê †y !   â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y       "   " ¹z  l      â Ey  Ð(    € â ™x?    â y      !   b ¹z  F   
   â $rÿÿ   ÿ Ž â r ÿ       â $rÿÿ   ÿ Ž â r! ÿ       â $rÿÿ   ÿ Ž â r	 ÿ       â $r(ÿÿ   ÿ Ž â rY ÿ       â $r,ÿÿ   ÿ Ž Ä $r
ÿÿ   ÿ Ž ä $rMÿÿ   ÿ Ž ä $r]ÿÿ   ÿ Ž â r  ÿ   pRð är  ÿ   pRò â$x     Ž Ê/ xÿ     È r   ÿ� È xÿ     È |    pbô Ú G)  P'    € ê y      '   " $rÿÿ   ÿ Ž â r ÿ       â $x   ÿ Ž â y      %   b xøÿÿÿÿÀŽ â $vÿ Y  ÿ Ž â xA        â y      &   ¢ z  X      â $x   
Ž â z q  ÿàÿ Æ $x?   ÿ Ž âx@!   ÿŽ ä xÿ     Æ %v@@  € AŽ È 'z y   Ž È $x   ÿ Ž ä/ $z }  ÿŽ â zÿ z    â 'z s   Ž ÈO $z p  Ž â zÿ t    â $z w  ÿŽ ä $z |  Ž ä$z |  Ž â x   ÿàÿ â $z v  Ž âr   ÿàõ â $z v  Ž Ä $z p  Ž ä %x<b   ÿŽ È %x~b   ÿŽ â r<@   ÿà÷ Æ $vÿ q  ÿ Ž â r=A   ÿäÿ â %xB?   <Ž Æ r   ÿŽ â$rÿÿ   ÿ Ž ä $x   Ž ä$x   Ž ä %x@b   @Ž È $x   Ž ä $rÿÿ   ÿ ä %x>?   ~Ž È rB   ÿàõ Ê $xC    Ê �y   • ¨ �y   • è �y   • ( �y   • b r$>   ÿàõ È r%?   ÿä Ê �y.$   • h �y $   • h �y2$   • h �y$   • b r"   ÿàõ Ê $x#    Ê �y"   • h �y)"   • b xT        ä xx $       â $x   ÿ Ž âO xÿ     È xü  ÿÀŽ â $x   ÿ Ž Æ� z&  ÿàõ â %v  T Ž â xü  ÿÀŽ Æ $v'ÿ  ÿ â r$@   ÿàõ â �yI    ™ ¢  x+ÿ     â $x   ÿ Ž âr%A   ÿä â �y-&    ™ " z"  ÿàõ ä/ xü  ÿÀŽ â �y;$   • ¤ $v#ÿ  ÿ Ä %v+  T Ž â �y$   • ¢ z&  ÿàõ ä� x   ÿ   â�yE"    ™ ä  $v'ÿ  ÿ ä �y9    ™ ¢
 xü  ÿÀŽ Ä xÿ     â �yw&    ™ ¢ x1ÿ     ä z*  ÿàõ â %v"  T Ž È %v01  T Ž â �yu"    ™ ¦  $v+ÿ  ÿ ä �y00    ™ ¨  �yK*    ™ ¢  x.   ÿ   È xü  ÿÀŽ È z&  ÿàõ â/ $x    ÿ Ž â xÿ   .  Æ$v'ÿ  ÿ â xü  ÿÀŽ ä r%   àµ ä �y7&    ™ b z*  ÿàù â $x2   ÿ Ž â r   ÿd~ â %v  T Ž Æ xü  ÿÀŽ â %v$% Z  x Ž â �y3    ™ ¦  $x$   ÿŽ â z.  ÿàõ È r%%   ÿàÿ â $v/ÿ  ÿ â x'ÿ   2  Æ/ �yq$   ™ ¢ x#ÿ      â %v&'  T Ž ä�y.    ™ ¤  $v+ÿ  ÿ ä �yr$   ™ ¢ %v"#  T Ž Æ �yn$   ™ ¢ $x   ÿ Ž Æ �yG*    ™ ¨  �y&&    ™ ¢ xü  ÿÀŽ Æ �yp$   ™ ¨ �yo"    ™ ¢ z*  ÿàõ ä xUÿ     â �ym$   ™ ¨  �yk$   ™ ¨  �y#$   ™ ¢  z+ÿ  ÿä â %vTU  T Ž â rO   àµ â �yi$    ™ è  �yl*    ™ â r   ÿd~ Æ �yTT    ™ â %vNO Z  x Ž Æ �y$$    • â $x$   ÿŽ â r<   ÿàõ Æ $xOO   Ž â r5   àÇ â $x=    Æ �ygN   ™ b r   ÿ„þ Æ �yjN   ™ b %v45 Z  x Ž Æ �y    • b $x$   ÿŽ Æ �yfN   ™ d r55   ÿàÿ ä �yhN   ™ h �yeN   ™ h �y 4   ™ b rP~   ÿàõ Æ �y\N   ™ h �yd4   ™ b $xQ    Æ �ycN   ™ b r/   àµ Æ �yb4   ™ b r   ÿd~ Æ �yaN    ™ h �y`4   ™ b %v./ Z  x Ž Æ �y4   ™ b $x$   ÿŽ Æ �y_P    • b $x//   Ž Æ �y^4   ™ h �y[4   ™ b	 r+   àµ Æ/ �yW.   ™ h �yR4    ™ b	 r   ÿd~ Æ �yZ.   ™ b ry   àÙ â $x$   ÿŽ ä �y.   ™ b r   ÿ¤~ Æ �yNN    • b %v*+ Z  x Ž Æ �yX.   ™ b %vxy Z  x Ž Æ �yV.   ™ b $x$   ÿŽ Æ �yU.   ™ b $xyy   Ž â xÿ   ÿÀŽ ä�yQ4    • b	 r++   ÿàÿ â 	s%         d  �yS.   ™ h �y'*   ™ h �y"*   ™ b xT  )    Æ �yO.    ™ b xÿ     Æ �yP*   ™ b xsÿ   ÿÀŽ Æ �yx   ™ h �yHx   ™ b $x%%€   ÿ Ž Æ/ �yL*   ™ h �yDx   ™ h �yJ*   ™ h �y8x   ™ h �y1.    • h �yF*   ™ b xÿ     Æ �y4x   ™ b	x/%€   ÿÀŽ Æ �y:*   ™ b x//ÿ   xŽ Æ �yx   ™ h	 �y6*    ™ h �y%*    • b x5ÿ   ÿÀŽ â $x+/ÿŽ Ä/ �y*x    ™ h	 �y/x   ™ b	 x2ÿ   )  â 	s 5        & x.2ÿ   ÿÀŽ Ê 	s2 .        b xz     Î 	st s        " $xv€   ÿ Ž ä �yx    • b $t|ÿ @€ÿ Ž â xz+€€€€zìŽ äx{2   ÿ   ä/ x|+€€€€|ìŽ ä xzzÿþþþÿàÿ Ä x||ÿþþþÿàÿ â $xtt€   ÿ Ž âx{.€   {xŽ ä x55€   vxŽ ä xzz+øŽ â$x{{ÿŽ â x+| @€+øŽ â $t|ÿÿ Ž â xvs€   txŽ â $tsÿÿ Ž Ä $x25ÿŽ ä $t.ÿ @€ÿ Ž â xy  @€    ä x|{€€€€|ìŽ äx52€€€€sìŽ äx.{€€€€.ìŽ ä xs2€€€€yìŽ Ä x||ÿþþþÿàÿ â $xvvÿŽ â x55ÿþþþÿàÿ ä x..ÿþþþÿàÿ ä xssÿþþþÿàÿ ä xzz˜º  ÿ    ä xy     Ä x||{øŽ äxx+˜º  ÿ    ä x552øŽ äx{. @€{øŽ ä x2s @€2øŽ â $tsÿ @€ÿ Ž â x.zÿÀŽ Ä x}z€€€€-ÞŽ ä x+v€€€€yìŽ ä xtxÿÀŽ äxyx€€€€IÞŽ âO $x.}   .
Ž â x++ÿþþþÿàÿ ä xsv€€€€sìŽ Ä rtyt  €ÿàÿ ä xz5˜º  ÿ    ä xx2˜º  ÿ    ä x++vøŽ ä xssÿþþþÿàÿ ä x..€€€€-´Ž Ä xIt€€€€I´Ž ä x€zÿÀŽ äx5z€€€€EÞŽ ä� xtÿ   ;  ä xzxÿÀŽ äx-x€€€€9ÞŽ Ä xy;ÿ   ÿÀŽ ä x‚+˜º  ÿ    ä xvs @€vøŽ â 	s2 y        b  xttÿ   ÿÀŽ â $xz-   z
Ž â xsÿ     ä x-‚€€€€wÞŽ Ä xx‚ÿÀŽ â 	s+ t        ¢ xssÿ   ÿÀŽ ä xvv˜º  ÿ    ä rx-x  €ÿàÿ ä x||˜º  ÿ    â 	s- s        â x{{˜º  ÿ    â $x€5   €
Ž â x9z€€€€9´Ž Ä xzvÿÀŽ äxyv€€€€uÞŽ ä x5|ÿÀŽ äxv{ÿÀŽ äx||€€€€KÞŽ ä x{{€€€€0ÞŽ â $xzy   z
Ž Ä $x||   5
Ž â r{{v  €ÿàÿ â $xv2€   ÿ Ž â/ x5;T      â $x++€   ÿ Ž âO xuz€€€€u´Ž ä xvv€   ÿÀŽ ä x5ÿ   5  ä xz-   ÿ   Ä� x-t€   +xŽ ä xvvÿ   ;xŽ ä x+5ÿ   ÿÀŽ â $ttÿÿ Ž â x;s€   zxŽ â $xsvÿŽ â 	s5 +        " $tzÿ @€ÿ Ž â xE€€€€€E´Ž â $x--ÿŽ â x2|€€€€K´Ž â $t€ÿ @€ÿ Ž â x|     Ä xKs€€€€tìŽ äxvs€€€€zìŽ ä xxx€€€€w´Ž ä xw-€€€€|ìŽ ä xtKÿþþþÿàÿ ä xy-€€€€€ìŽ Ä xvvÿþþþÿàÿ ä xzwÿþþþÿàÿ ä xttsøŽ äx|yÿþþþÿàÿ ä xsv @€søŽ â $xv5€   ÿ Ž â xKz-øŽ â $x;;ÿŽ â x€     Ä x-| @€-øŽ â $t|ÿ @€ÿ Ž â x++€   vxŽ ä x€;€€€€€ìŽ â$tvÿÿ Ž â x|;€€€€|ìŽ â $x++ÿŽ â xzt˜º  ÿ    ä x€€ÿþþþÿàÿ Ä x||ÿþþþÿàÿ ä xtzÿÀŽ äxwz€€€€7ÞŽ äxz  @€    ä x5€;øŽ ä x€+€€€€vìŽ Ä xvs˜º  ÿ    ä x;| @€;øŽ ä x|+€€€€zìŽ ä xzK˜º  ÿ    ä x€€ÿþþþÿàÿ ä xKvÿÀŽ Ä xvv€€€€3ÞŽ â $xtw   t
Ž â x||ÿþþþÿàÿ ä xw€+øŽ â $xvv   K
Ž â x+| @€+øŽ ä x|-˜º  ÿ    â &r-.q   ÿ   â xww˜º  ÿ    Ä x3v€€€€3´Ž â &r-Ir   -   â xvwÿÀŽ äxww€€€€ÞŽ ä x++˜º  ÿ    â &r-En   -   â xszÿÀŽ â$xww   v
Ž â xv+ÿÀŽ â &r-9p   -   â xzz€€€€GÞŽ Ä xtt€€€€7´Ž ä x++€€€€&ÞŽ ä x7|ÿÀŽ äx||€€€€oÞŽ â &rqtq   ÿ   â rzzs  €ÿàÿ ä r++v  €ÿàÿ â &rvx#   -   â xGz€€€€G´Ž â $x||   7
Ž Ä &rq3r   q   ä &rvum   v   â x55˜º  ÿ    â &rqGn   q   â x0{€€€€0´Ž ä xo|€€€€o´Ž ä x++€€€€&´Ž â &r&2k   v   â xr5ÿÀŽ Ä x;;˜º  ÿ    â &rqop   q   â xw€€€€´Ž â &r&0i   &   â x55€€€€lÞŽ ä xp;ÿÀŽ â&rn#   q   â x;;€€€€TÞŽ â $x55   r
Ž â r#&&   ÿ� â &rn+m   n   â x)ÿ   )  â $x;;   p
Ž â xl5€€€€l´Ž Ä x#ÿ   #  ä xT;€€€€T´Ž â &rnlk   n   ä $r#)&   #Ž ä &r-.g   ÿ   ä &riTi   n   â r###   ÿ� â &r-Ij   -   Ä 0rÿ    A   â r&ii   ÿ� â 0r$ÿ$    A   â x#ÿ   #  â &r5Ef   -   â x-ÿ   &  ä Er# #       â  r&$     A  â &r59h   5   Æ #r]&#   ]   â &r5xe   5   ä &r#.    ÿ   ä&r t    ÿ   â xÿ     â &r5u\   5   ä &r#Id   #   ä&r 3d       Ä &r52c   5   ä &r#Eb   #   ä$r-i   -Ž ä &rbGb       ä &r 0a   5   ä &r59`   #   â r---   ÿ� â 0r_ÿ_    A   Ä &r5x   5   â x-ÿ   -  Æ &r5u^   5   â Er- -       â &rbo`   b   â  r$$_     A  â r#     ÿ� â &r52[   5   ä &rb   b   â x#ÿ   #  â #rY$-   Y   â &rgtg   ÿ   Ä &r.W   ÿ   ä &r$0R   5   ä &rg3j   g   ä $r#)    #Ž ä &r5IZ      â r$$   ÿ� â &rgGf   g   â r###   ÿ� â &rWtW   ÿ   â xÿ     â &r5E   5   â x#ÿ   #  â &rgoh   g   Ä 0rNÿN    A   ä &rW3Z   W   ä &r59X   5   ä $r)$   Ž â Er# #       â &rge   g   â  r N     A  â &rWG   W   â r   ÿ� â &rxV   5   Ä &rg+\   g   â #rM #   M   â &r$uU      â xÿ     â 0r ÿQ    A   ä &rglc   g   ä &rb+^   b   â Er#        â &rgTa   g   â  r      A  â &r[l[   b   Æ #r	#   	   â r&gg   ÿ� â &r[TR   [   ä &r#.'   ÿ   ä &r$2S   $   â x&ÿ   &  â &r#I"   #   â r-[[   ÿ� â &r..   ÿ   Ä &r$0O   $   ä &rt   ÿ   ä&r't'   ÿ   â x-ÿ   -  â &rWoX   W   ä &r#EP   #   ä $r&g   &Ž Ä &r.IH   .   â r$$   ÿ� â &r3H      ä&r'3"   '   ä &rVV   W   ä &r#9L   #   â r&&&   ÿ� â &r.ED   .   â xÿ     â $r-[   -Ž Ä &rGD      ä&r'GP   '   ä &rV+U   V   ä &r#xJ   #   â x&ÿ   &  â &r998   .   âr---   ÿ� â &ro8      Ä &r'oL   '   ä &rVlS   V   ä $r)$   Ž ä &r#uF   #   â Er& &       â &r9x4   9   â  r_N     A  â &r44      âx-ÿ   -  â &r'J   '   â r   ÿ� â &rOTO   V   Ä &r#2:   #   ä &r9u   9   â #r
&   
   â &r4+   4   âEr- -       â &r'+F   '   â  r_      A  â &r"06   #   â r OO   ÿ� â 0rÿ1    A   â xÿ     â &r92/   9   Ä &r/l/   4   â#r,-   ,   â &r'l:   '   â x ÿ      â  r     A  â r""   ÿ� â &r/T*   /   âEr        â &r'T6   '   ä &r00*   9   â xÿ     â $r O    Ž â #r(   (   â r//   ÿ� Ä r#''   ÿ� ä r00   ÿ� â $r)"   Ž â r      ÿ� ä x"ÿ     ä x#ÿ   #  ä xÿ     Ä r   ÿ� ä x ÿ      â $r#'   #Ž â x   ÿàÿ â $r"/   "Ž ä $r)0   Ž â xÿ     â 0rÿ%    A   â Er          â  r_     A  â |    pbô Ä r###   ÿ� ä r"""   ÿ� ä r   ÿ� â  r     A  â Er        â #r!    !   â 0rÿ    A   â x#ÿ   #  ä x"ÿ   "  Ä xÿ     â #r      â  r_     A  âEr# #       â  r_     A  â Er" "       â  r     A  â Er        ä x  ÿà÷ â #r#      â #r"      â #r      â x@   ÿàÿ Ä x@   ÿàÿ ä x@   ÿàÿ ä x@   ÿàÿ ä x@   ÿàÿ ä x@   ÿàÿ ä rÿ   ÿäÿ â D)       À â Gy  €Ûÿÿÿÿƒ ê Ay        € ê Ey  ð     € â G‰  Ð     € ê $x  €  Ž Ê ˆs  ] úÿ H   è ˆs  Y€úÿ H   è ˆs  M ûÿ H   è ˆs  
€ûÿ H   è ˆs  	 üÿ H   è ˆs  ,€üÿ H   è ˆs  ( ýÿ H   è ˆs  !€ýÿ H   è ˆs   þÿ H   è ˆs  €þÿ H   è ˆs   ÿÿ H   è ˆs  €ÿÿ H   ä Ay        € ê {          ì M        € ê „y      H   " $v+ÿ r  ÿ Ž Æ „y €   H   b $x'+   ÿŽ ä$x)+   ÿŽ â „y €  H   ¨ „y    H   è „y    H   ( „y €  H   h „y    H   h „y €  H   h „y €  H   h „y    H   è „y    H   â !r  ]       â !rY       Ä/ y      %   " !r
       ÆO ‰       b !rM       Æ� ‰      ¢ !r	       Æ„y
 €  H   â !r,,       Æ‰	      â !r((       Æ ‰      " !r       Æ ‰      b $x   ÿ Ž â !r!       â !r        â/ ‰,      $ z  r  p`ð â !r       â !r       âO ‰       b r ÿ   p"p â !r       Ä ‰      ¢ !r		       Æ� ‰(      â !r       Æ‰      " !r       Æ‰      b x   ÿàÿ â !r
       ä ‰      b !r,       â z  r  p`ò ä ‰      " !r        â/ x    p"ò â !r       âO y       &   b !r(       Æ� ‰ €    ¢ !r       Æ‰	      â !r       Æ‰
      " !r       Æ y      '   b !r       Æ ‰ €    ( ‰      " $z   x  Ž Æ/ ‰      b !r       ÆO ‰      ¢ !r	       Æ� ‰      è ‰      â $z  ~   Ž â!r
       ä‰      " !r       â 
  r  ÿà÷ Ä ‰"      " !r       Æ ‰$      b !r       â/ r+    ÿˆ â !r       âO ‰
 @    b ÿÿ   ÿäÿ ä $rÿÿ   ÿ â !r        â� ‰ €    ¢ 
 j  ÿ† â !r       Ä ‰	 @    â  '   ÿàõ â !r       ä‰ €    " z j  ÿˆ â $ÿÿ   ÿ â !r""       â ‰ €    " 
 j  ÿ„ â !r$$       Ä‰ €    b 
 k  � ä z k   â ‰ €    b 
 k   â !r
       â/ r+    ÿ„ â ‰!  €    b !r       âO  )   ÿàù Ä ‰ €    ¢ !r	       â� r&ÿÿ   ÿä ä ‰ €    â !r       âz
 j  ÿ„ â $ÿÿ   ÿ â ‰#" €    " !r       â z k  & Ä ‰%$      " !r       â	+   ÿàÿ â !r       â ‰ @    d ( 	   ÿà÷ â !r! !       â/ ‰ @    b 
 j  ÿˆ ä $+ÿÿ   ÿŽ â !r       âO ‰      ¢ ( j  ÿ† â !r       Ä� ‰ @    â 
	 k   â !r#"#       ä‰ @    " ( k  +� â !r%$%       ä ‰ @    ( ‰"! @    " !r       â        Ä ‰ @    b !r       Æ/ ‰  @    b !r+       âO '   ÿàÿ ä ‰      ¢ !r       Æ� ‰$# @    â !r       Æ‰&% €    " !r       â !r"!"       â $!ÿ   ÿ Ž â ‰      " !r       Æ‰      b !r         Æ/ ‰      b !r-       âO     ÿàõ ä‰      ¢ %  j   Ž â!r$#$       ä� ‰      â %  j  ! Ž â !r&%&       Ä‰      " $(ÿÿ   ÿ â  j  ÿ„ ä‰!       " !r       â  k  ( ä ‰#"      " !r       Æ‰'& @    b !r       Æ/ ‰%$      b !r       ÆO †	 +    â !r       Æ� † -   â !r       Æ†	     â !r! !       â !r#"#       ä †     â !r'&'       Æ†	     â !r%$%       Æ/ †    è †	     è †     è †	 
!    è † 
#   è ‰'      ( †	 %    â M™        € ê x)   ÿàÿ âO !r''       Æ r    ÿàñ È r ÿÿ   ÿä  ä z j  ÿ€ È z k     Ê †y '    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      "   " ¹z  l      â Ey  P%    € â ™x?    â y      !   b ¹z  F   
   â $r ÿÿ   ÿ Ž â r- ÿ       â y!      %   ¢ $rÿÿ   ÿ Ž â x0      ÿ  â $r/ÿÿ   ÿ Ž â x      ÿ  â $r ÿÿ   ÿ Ž Ä $rÿÿ   ÿ Ž â r ÿ   pRð är ÿ   pRò â$x"    Ž ä/ $x!!   ÿ Ž ÆO xÿ   "  È r"   ÿ� È xÿ     È |    pbô Ú G)  ð#    € ê y      '   " r ÿ       â $vÿ q  ÿ Ž â x	øÿÿÿÿÀŽ â y      &   b $rÿÿ   ÿ Ž â z  X      â $x,   ÿ Ž ä $x""   	
Ž ä $x$   ,Ž âz), q  ÿàÿ â $x'   ,Ž âx"!   ÿŽ â $x(   ,Ž â x+ÿ   "  â $v%ÿ Y  ÿ Ž Ä $r ÿÿ   ÿ Ž ä 'z y   Ž È 'z s   Ž â/ zÿ z    Æ $z! p  Ž â zÿ t    â $z# }  ÿŽ ä $z |  Ž â x!   ÿàÿ â $z |  Ž ä $z w  ÿŽ ä $z v  Ž Ä $tÿ   ÿ Ž â r##   ÿàõ â $z v  Ž ä $z	 p  Ž ä %v  € Ž â x"   ÿ   Æ %xb   ÿŽ È $r.ÿÿ   ÿ â r&   ÿà÷ â %x
   Ž È $x*   Ž ä%x	b   Ž È %x	b   ÿŽ Ì %x   Ž È r
   ÿàõ Ê $x   % Ê �y4   • ¨ �y   • è �y2   • ( �y3   • b r   ÿàõ È r%   ÿä Ê �y<   • h �y:   • h �y>   • h �y   • b r&   ÿàõ Ê $x%   * Ê �yf   • h	 �y   • b	 $t[ÿ$   ÿ Ž ä $toÿÿ Ž â xs  @€    â $x4   ÿ Ž äO $x   ÿ Ž Æ� xü  ÿÀŽ È z  ÿàõ ä xü  ÿÀŽ â $x2   ÿ Ž âx        â $vÿ  ÿ â zP  ÿà÷ ä xÿ   4  ä �yY    ™ ¢ r   ÿàõ â $vQÿ  ÿŽ â xü  ÿÀŽ â %v   Ž Æ �yPP    ™ â  $x   % â z  ÿàõ ä/ �yX    ™ " xÿ   2  Æ �yK   • ¢ xNÿ     â $vÿ  ÿ ä %v   Ž â�y2   • ¢ x3   ÿ   ä%vNN   Ž â �y5    ™ ¢
 xü  ÿÀŽ Æ �y8    ™ ¢  z6  ÿàõ Æ �yNN    ™ ¤  $v7ÿ  ÿ â xÿ   3  È �y96    ™ ¢  %v   Ž Ê �y;    ™ ¢  $x<   ÿ Ž ä$x:   ÿ Ž â x<ÿ   <  ä xü  ÿÀŽ ä xü  ÿÀŽ â %v<<   Ž â z  ÿàõ Ä z  ÿà÷ â $x>   ÿ Ž â/ x:ÿ   :  â $vÿ  ÿ â zÿ  ÿäÿ â �y<<    ™ b r4"#   ,àµ â %v:   Ž â x@ü  ÿÀŽ â �yM    ™ â r3+.   ÿd~ Ä z@@  ÿàù â �y:    ™ â  $x   ÿ Ž â xÿ   >  â $x33$   ÿŽ â �yL    ™ â  $vAÿ  ÿ ä %v4 Z  [ Ž â/ xü  ÿÀŽ ä r"#   )àÇ â %v   Ž â r3   ÿàÿ â �y4@    ™ b r+.   ÿ„þ Ä z>  ÿàõ â �y7    ™ b %v Z  [ Ž Æ �y   ™ h �yI   ™ b $x$   ÿŽ Æ/ �yJ   ™ b $x   Ž â xÿ     ä �yH   ™ h �y@   ™ b $v?ÿ  ÿ Æ �yG   ™ b %v   Ž Æ �y=   ™ h �yE   ™ h �yF   ™ h �y>>    ™ h �yC   ™ h �y6    ™ h �yD   ™ h �yA   ™ b r"#   $àµ Æ �yB    ™ h �y3   ™ b r+.   ÿd~ Æ �y?   ™ b %v Z  [ Ž Æ �yS    ™ b $x$   ÿŽ â r   ÿàõ ä�yQ    • b $x   Ž ä $x   % â r   ÿàõ â �yZ    • h �yR   ™ b r	%   ÿä Ä x^fÿ   ÿÀŽ â �yO    • b r\"#   'àµ Æ �yT   ™ b 	sc ^        " r+.   ÿd~ ä/�yU   ™ b r"#   (àµ â %v\ Z  [ Ž ä �yV   ™ b r]+.   ÿd~ â $x_$   ÿŽ Ä �yW   ™ b %v Z  [ Ž Æ �y    • b $x   _Ž ä $x]]$   ÿŽ â �y\   ™ b $xcc€   ÿ Ž ä $x   ]Ž â �y   ™ b xdÿ   f  Æ �y[   ™ b xmdÿ   ÿÀŽ Æ �y    ™ b xec€   ÿÀŽ Æ �y]    • h �y   ™ b 	si m        & �yb   ™ h �y   ™ b# xeeÿ   fxŽ Æ �yd   ™ b xffT      Æ �ya   ™ b xfÿ   f  Æ �yh   ™ h �y`   ™ b xkfÿ   ÿÀŽ Æ �yg   ™ b $xleÿŽ Æ �y_   ™ h �yf   ™ b $xni€   ÿ Ž Æ �y^   ™ h �y    ™ h �yc    • b xnm€   nxŽ Æ �ye   ™ h �y   ™ b! xl€€€€oìŽ È xmÿþþþÿàÿ ä �y    ™ b xil€€€€sìŽ È xoiÿþþþÿàÿ ä �yi    • b xjÿ     È xjjÿ   ÿÀŽ È 	sq j        b xmmløŽ Î 	sp k        " xlo @€løŽ â $xnnÿŽ ä $toÿÿ Ž ä $xqq€   ÿ Ž Ê/ xqj€   qxŽ ä xjn€€€€oìŽ â $xpp€   ÿ Ž â xon€€€€sìŽ ä xjÿþþþÿàÿ â $xjqÿŽ â xooÿþþþÿàÿ â $tqÿ @€ÿ Ž â xpk€   pxŽ Ä xk     ä xnøŽ äxo @€nøŽ â $xnpÿŽ â xoj€€€€qìŽ â$tqÿÿ Ž â xkj€€€€kìŽ â $tuÿ @€ÿ Ž â xpm˜º  ÿ    Ä xsoÿþþþÿàÿ ä xkkÿþþþÿàÿ ä xon€€€€qìŽ äxqn€€€€uìŽ ä xmpÿÀŽ äxpp€€€€YÞŽ ÄO xkkjøŽ äxjs @€jøŽ ä xuqÿþþþÿàÿ â $xpp   m
Ž â xsoÿþþþÿàÿ ä xq˜º  ÿ    ä xol˜º  ÿ    Ä xsnøŽ äxlu @€nøŽ ä xrqÿÀŽ ä xnoÿÀŽ ä xqq€€€€PÞŽ ä� xoo€€€€XÞŽ ÄxYp€€€€Y´Ž ä xpKÿ   ÿÀŽ ä xm˜º  ÿ    â $xq   r
Ž â roon  €ÿàÿ ä x˜º  ÿ    â 	sp p        " xql˜º  ÿ    Ä xXo€€€€X´Ž ä x€€€€P´Ž ä xoÿÀŽ äxPqÿÀŽ äx€€€€5ÞŽ ä xqq€€€€8ÞŽ Ä xnmÿÀŽ â$x   o
Ž â xmm€€€€NÞŽ â $xqq   P
Ž â xlÿ   K  ä x€€€€5´Ž â $xmm   n
Ž â xllÿ   ÿÀŽ ä xnÿ   2  Ä x5q€€€€8´Ž ä x8k˜º  ÿ    ä xpp   ÿ   ä xNm€€€€N´Ž ä 	sm l        " xPKT  2    ä xnnÿ   ÿÀŽ Ä xq8ÿÀŽ ä xpp€   ÿÀŽ ä x88€€€€9ÞŽ â 	so n        b xPÿ   P  ä xjj˜º  ÿ    â $x88   q
Ž â xppÿ   KxŽ Ä xkPÿ   ÿÀŽ ä xqjÿÀŽ äxjj€€€€;ÞŽ â $xppÿŽ â 	sK k        ¢ $tsÿÿ Ž ä $tuÿ @€ÿ Ž â x88€€€€9´Ž â $xPj   q
Ž â x9p€€€€sìŽ Ä xjp€€€€uìŽ ä xmm   ÿ   â $xoo€   ÿ Ž â/ xPP€€€€;´Ž ä x99ÿþþþÿàÿ ä x;jÿþþþÿàÿ ä xml€   mxŽ â $trÿ @€ÿ Ž â xj9pøŽ Ä xl; @€pøŽ â $xmmÿŽ â xon€   oxŽ â $tpÿÿ Ž ä $xnK€   ÿ Ž âO x;m€€€€rìŽ â $xooÿŽ â x9m€€€€pìŽ â $ttÿ @€ÿ Ž â xr     Ä xKk€   nxŽ ä xn9ÿþþþÿàÿ ä xp;ÿþþþÿàÿ ä x;o€€€€rìŽ äxqo€€€€tìŽ ä x9nmøŽ Ä xkp @€møŽ â $xmKÿŽ â xn;ÿþþþÿàÿ ä xpqÿþþþÿàÿ ä xql˜º  ÿ    ä xKnoøŽ äx;p @€oøŽ Ä xom€€€€tìŽ ä xjj˜º  ÿ    ä xlqÿÀŽ äxqq€€€€<ÞŽ äxpoÿþþþÿàÿ ä xojÿÀŽ Ä xkk˜º  ÿ    ä xjj€€€€MÞŽ â $xqq   l
Ž â xlkÿÀŽ äxkk€€€€:ÞŽ â $xjj   o
Ž â xnm€€€€rìŽ ä x99˜º  ÿ    â $xkk   l
Ž â xnnÿþþþÿàÿ Ä xMj€€€€M´Ž ä xj9ÿÀŽ äx99€€€€LÞŽ ä xnnmøŽ äx:k€€€€:´Ž ä xk;˜º  ÿ    â &r;Y   ÿ   âxpp @€møŽ â &rM   ÿ   â x<q€€€€<´Ž Ä r99j  €ÿàÿ ä xmn˜º  ÿ    â &r;XI   ;   âx99€€€€L´Ž â &r<I      â xjmÿÀŽ äxmm€€€€4ÞŽ â &r;J   ;   âxpp˜º  ÿ    â &rI9J      Ä &rY@   ÿ   ä $xmm   j
Ž ä &r;NH   ;   â xopÿÀŽ â&r@M@   ÿ   â xpp€€€€7ÞŽ â &rXG      â xKK˜º  ÿ    â &r;=   ;   â x4m€€€€4´Ž â &rI:H   I   Ä &r@<G   @   ä &rE      â xLKÿÀŽ â$xpp   o
Ž â xKK€€€€>ÞŽ â &r;5F   ;   â xjkÿÀŽ â &rI4=   I   ä &r@9E   @   â xkk€€€€6ÞŽ â &r=NC      âx7p€€€€7´Ž â &r;8D   ;   â rKKL  €ÿàÿ â &rC:C   @   Ä &r@A   =   â x>K€€€€>´Ž â $xkk   j
Ž ä &rPB   ;   ä &rI7F   I   ä &r@53   @   â x6k€€€€6´Ž â &rI>D   I   â r;   ÿ� â &r@8?   @   â xÿ     â &rI6B   I   â x;ÿ   ;  â &r@PS   @   Ä &rB4A   C   ä $r;   ;Ž â r@@   ÿ� â &rB73   B   â r=II   ÿ� â &r3YR   ÿ   âr;;;   ÿ� â &rB>?   B   â xÿ     â 0rOÿO    A   â x2ÿ   2  â 0rQÿQ    A   â x=ÿ   =  â &rS6S   B   â x;ÿ   ;  â &r?MR   ÿ   Ä &rBXT   3   ä $r@   Ž â Er; ;       â $r=2I   =Ž â  r3OQ     A  â &rT<T   ?   â r   ÿ� â &rAU   B   â #r3;      â &rU9U   T   â r===   ÿ� â &rANV   A   âx3ÿ     â 0rZÿZ    A   â x=ÿ   =  â &rU:V   U   â Er; 3       â &rAW   A   â  r3OZ     A  â 0rÿ    A   â Er= =       â &rU4W   U   Ä &rA5\   A   â  rQ     A  â #r3;      â &r;Y   ÿ   ä &rU7\   U   ä &rA8[   A   â #r=      â &r=M   ÿ   ä &r;X   ;   Ä &rU>[   U   ä &rP   A   ä &rMM   ÿ   ä&rYY   ÿ   ä &r=<   =   ä &r;b   ;   Ä &rU6   U   â r   ÿ� â &rM<d   M   ä&rYXd   Y   ä &r=9b   =   ä&r;Na   ;   â xÿ     â &rM9h   M   Ä &rYh   Y   ä &r=:a   =   ä &r`   ;   ä &rM:g   M   ä&rYNg   Y   ä $r   Ž Ä &r`4`   =   ä&r5_      ä&rM4f   M   ä&rYf   Y   â r   ÿ� â &r_7_   `   ä &r8^      Ä &rM7   M   ä&rY5   Y   â xÿ     â &r_>^   _   ä &rP      ä 0r]ÿ]    A   ä &rM>e   M   Ä &rY8e   Y   â Er        â &r_6   _   â r   ÿ� â  rO]     A  â &rM6   M   â rUU   ÿ� â &rPP   Y   â xÿ     â #r1   1   â xÿ     Ä rMM   ÿ� ä r?SS   ÿ� ä r__   ÿ� ä rPP   ÿ� â $r   Ž â xÿ     â $r2U   Ž â x?ÿ   ?  Ä xÿ     ä xÿ     â $r?2S   ?Ž âr   ÿ� â $r2_   Ž â x   ÿàÿ â $r2M   Ž â r   ÿ� â $rP   Ž â xÿ     â 0rcÿc    A   â |    pbô Ä xÿ     ä r???   ÿ� ä r   ÿ� ä r   ÿ� ä r   ÿ� â  rOc     A  â Er        â  r]]     A  â Er        â 0riÿi    A   â x?ÿ   ?  Ä xÿ     ä xÿ     ä xÿ     â #r/   /   â Er? ?       â #r0]   0   â Er        â  r3Z     A  âEr        â  rc     A  âEr        â  ri     A  âx  ÿà÷ â  rOi     A  â #r 3?       â #r-   -   â #r        â #r      â x$$@   ÿàÿ â $r%ÿÿ   %Ž â x''@   ÿàÿ Ä x((@   ÿàÿ ä x))@   ÿàÿ ä x,,@   ÿàÿ â D)       À â Gy  °Þÿÿÿÿƒ ê Ay        € ê Ey  Ð     € â G‰  °     € ê $x@  Ž Ê ˆs  ûÿ H   è ˆs €ûÿ H   è ˆs  üÿ H   è ˆs  €üÿ H   è ˆs 1 ýÿ H   è ˆs 0€ýÿ H   è ˆs / þÿ H   è ˆs -€þÿ H   è ˆs  ÿÿ H   è ˆs  €ÿÿ H   ä Ay        € ê {          ì M        € ê „y     H   b z ! r  p`ð â $v#ÿ r  ÿ Ž â x'        â „y €   H   ¢ r ÿ   p"p Æ „y    H   è „y €  H   ( „y    H   h „y	 €  H   ( „y    H   ( „y
 €  H   ( „y    H   è „y €  H   â !r       â/ !r       ÈO ‰      b !r       ÆŸ y      '   " !r         Æ‰      ¢ !r1       Æ‰      â !r0	0       Æ ‰      " !r/       Æ ‰	       b !r

-       â !r       ä/ ‰0      b !r       Æ ‰      " !r        â !r       äO ‰      ¢ !r       Æ� ‰
      â !r       Æ‰      " !r	 	       Æ‰      b !r0       Æ/ ‰      b !r       Æ ‰	      " !r       ÆO ‰      ¢ !r
       Æ� ‰  €    â !r       Æ‰      " !r       Æ‰
      h ‰      b !r       Æ/ ‰      b !r		       Æ ‰      " !r       ÆO ‰      ¢ !r        â� !r       ä‰ €    â !r
       Æ‰ €    " x
!   ÿàÿ â !r       ä ‰	 €    b z 
 r  p`ò â !r       ä/ y       &   b x    p"ò â !r       Ä ‰ @    " !r       ÆO ‰ €    ¨ ‰ €    ¢ !r       Æ� ‰ €    â !r       Æ‰ €    " !r	       Æ‰ €    b $z   x  !Ž ä/ $x!#   ÿŽ â ‰ €    b !r
       â $z  ~   Ž â !r       äO ‰ @    $  !   ÿàõ â!r       â r#    ÿ† â‰
      ¢ 
	  r  ÿàù â !r       â� $ ÿÿ   ÿ â ‰ @    â 
 j  ÿ„ â !r       â$rÿÿ   ÿŽ â ‰ @    " z j  ÿ† â !r       â
 k    â ‰ @    b z k  � â !r       â/ r#    ÿŠ â ‰ @    b $ÿÿ   ÿä Ä 
	 j  ÿˆ â ‰ @    b !r       â $r"ÿÿ   ÿŽ â z j  ÿŠ â ‰ @    " !r%
       âO #   ÿàÿ ä ‰ @    ¢ !r       â� !   ÿàÿ Ä ‰ @    â !r       â!    ÿàõ â%v
  j  ' Ž â !r       â    ÿà÷ â ‰      " z k  "� â !r       â/ $"ÿÿ   ÿ â ‰      b ! j  ÿ„ â !r       â $ ÿÿ   ÿŽ â ‰      b 
		 k  $ â !r       â  j  ÿ† â ‰      " !r       ÄO  k   � â ‰      ¢ !r       Æ� ‰      è ‰       â !r#       â! k  " ä ‰      " !r       Æ/ ‰      b !r       â!r       ä †	 
%    â !r       ÆO † 
#   â !r       Æ� †	     â !r        Æ †     â !r       Æ†	     â !r       Æ/ †    è †	     è †     è †	     â M™        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y       "   " ¹z  l      â Ey  p"    € â ™x?    â y      !   b ¹z  F   
   â $r+ÿÿ   ÿ Ž â r  ÿ       â y      %   ¢ $r#ÿÿ   ÿ Ž ä $rÿÿ   ÿ Ž Ä $rÿÿ   ÿ Ž ä $r-ÿÿ   ÿ Ž ä $rÿÿ   ÿ Ž ä $r4ÿÿ   ÿ Ž â r  ÿ   pRð är  ÿ   pRò â$x!     Ž Ä/ $x   ÿ Ž ÆO xÿ   !  È r!   ÿ� È x.ÿ     È | .   pbô Ú G)  !    € ê y      '   " xøÿÿÿÿÀŽ â $z p  .Ž â r ÿ       â y      &   b $tÿ   ÿ Ž â x*.   ÿ   â $x!!   
Ž ä $rÿÿ   ÿ Ž â z'* q  ÿàÿ â $vÿ X  ÿ Ž â x)ÿ   !  â $v%ÿ Y  ÿ Ž Ä $r+ÿÿ   ÿ Ž ä 'z y   Ž È 'z s   Ž â/ zÿ z    ä x   ÿàÿ â $z" }  ÿŽ â zÿ t    â $z	 |  Ž â x!!   ÿŽ â $z |  .Ž ä $z v  	Ž Ä $z v  Ž ä $z w  ÿŽ ä $z	 p  Ž ä %v  € Ž â r""   ÿàõ Æ %xb   ÿŽ È $vÿ q  ÿ Ž â r$   ÿà÷ â $x!   ÿ Ž ä $x   *Ž â r(   ÿäÿ â %x	b   Ž È %x	b   ÿŽ È $x&   *Ž ä %x
   Ž È $r,ÿÿ   ÿ ä %x   Ž È r
   ÿàõ Ê $x   % Ê �y   • ¨ �y   • è �y   • ( �y   • b r   ÿàõ Ê $x   % Ê �y1   • ( �y/   • h �y0   • h �y   • b r$   ÿàõ Ê $x%   ( Ê �yT   • h	 �y7   • b	 $tYÿ$   ÿ Ž â x   ÿ   ÈO xü  ÿÀŽ È zP  ÿàõ â $x   ÿ Ž È� $vQÿ  ÿ â xÿ     â $tÿ   ÿ Ž â xü  ÿÀŽ ä �yPP    ™ ¤ z  ÿàõ â %v   Ž È $x   ÿ Ž â�yO    ™ â  r   ÿà÷ â $vÿ  ÿ ä xü  ÿÀŽ ä xÿ     â �yS    ™ " r%   ÿäÿ ä zL  ÿàõ â %v   Ž Ä �yI   • ¤  $vMÿ  ÿ ä �yQ    ™ ¢" $x   ÿ Ž Æ�yLL    ™ ¢
 xÿ     ä xü  ÿÀŽ â �y6   • ¢ xÿ     â %v   Ž â z  ÿàõ Æ %v   Ž â �yN    ™ ¦  $vÿ  ÿ â �yR    ™ ¨
 �yK    ™ ¢
 x1   ÿ   È/ xü  ÿÀŽ ä xÿ   /  â $x//   ÿ Ž â z  ÿàõ â $x0   ÿ Ž Æx//ü  ÿÀŽ â $vÿ  ÿ ä %v   Ž âxÿ   1  ä z/  ÿà÷ â �yJ    ™ b xü  ÿÀŽ â %v   Ž ä �yG    ™ $ $vÿ  ÿŽ â xÿ   0  â �yM    ™ h �yH    ™ b x   ÿ   Ä/ z  ÿàù ä r!"   *àµ ä x8ü  ÿÀŽ â $vÿ  ÿ â r),   ÿd~ ä z88  ÿàù â %v   Ž â r!"   'àë â �y2    ™ d $v9ÿ  ÿ Ä $x$   ÿŽ â xÿ     â %v Z  Y Ž â �y3    ™ b r),   ÿÄþ Æ �y18    ™ b $x   Ž ä %v Z  Y Ž Æ/ �y0   ™ b %v   Ž Æ �yD   ™ h �y5    ™ b $x$   ÿŽ ä $rÿÿ    Ž â �yC   ™ d r   ÿàÿ Ä �yB   ™ h �y   ™ h �yA   ™ h �y@   ™ h �y?   ™ h �y>   ™ h �y=   ™ h �y<   ™ h �y;   ™ h �y:    ™ b r   ÿàõ Ä r   ÿà÷ â �yE    • d $x   % ä$x	   %Ž â r!"   àµ â�y9   ™ b r!"   &àÙ ä r),   ÿd~ â�y    • b rU),   ÿ¤~ â %v Z  Y Ž Ä �y/   ™ d $xW$   ÿŽ ä %v Z  Y Ž â �y8    ™ f $xUU$   ÿŽ â �yF    • b $x   WŽ Æ rU   ÿàÿ â �y    • h �yX   ™ h �y   ™ h �y   ™ h �yW   ™ h �yV   ™ h �yU   ™ h �y   ™ h �y   ™ h# �y   ™ h �y   ™ b x_Tÿ   ÿÀŽ Æ �y^   ™ b xYÿ   T  Æ �y]   ™ h �y\   ™ b 	s_ _        & �y[   ™ b xbYÿ   ÿÀŽ Æ �yZ    ™ h �yY    ™ b x`ÿ   7  â $x__€   ÿ Ž Æ xd`ÿ   ÿÀŽ ä �y`    • b xa_€   ÿÀŽ Æ �y_    • b 	se d        0 	sc b        b xaaÿ   TxŽ Ä xTTT  7    â $xee€   ÿ Ž Æ xTÿ   T  â $xaaÿŽ â xfd€   exŽ â $xcc€   ÿ Ž â/ xeTÿ   ÿÀŽ ä xT  @€    ä xcb€   cxŽ ä 	sb e        " $tÿÿ Ž â xa€€€€TìŽ â $xccÿŽ Ä $tTÿÿ Ž â xa€€€€ìŽ â $tdÿ @€ÿ Ž â xÿþþþÿàÿ ä xc€€€€TìŽ äxT @€aøŽ ä xc€€€€dìŽ ä xÿþþþÿàÿ â $xgfÿŽ â xÿþþþÿàÿ â $thÿÿ Ž â xdÿþþþÿàÿ â $xfb€   ÿ Ž â xbcøŽ Ä x  @€    ä xaøŽ ä xd @€cøŽ ä xg€€€€hìŽ äxcg€€€€ìŽ ä xfe€   fxŽ Ä xÿþþþÿàÿ ä xa˜º  ÿ    ä xdcÿþþþÿàÿ â $teÿÿ Ž â xgøŽ â $xffÿŽ â xaÿÀŽ äxaa€€€€PÞŽ ÄO xd @€gøŽ â $tgÿ @€ÿ Ž â xcf€€€€eìŽ â $xaa   
Ž â xTT˜º  ÿ    ä xf€€€€gìŽ ä xgcÿþþþÿàÿ ä xcTÿÀŽ Ä xTT€€€€OÞŽ ä� xbb˜º  ÿ    Æ $xTT   c
Ž â xebÿÀŽ äxggføŽ ä xbb€€€€SÞŽ äxOT€€€€O´Ž ä xPa€€€€P´Ž ä xT˜º  ÿ    Ä xaIÿ   ÿÀŽ ä xiÿþþþÿàÿ ä xgg˜º  ÿ    ä rbe  €ÿàÿ ä xTÿÀŽ äxTT€€€€QÞŽ â 	sa a        " xbgÿÀŽ Ä x€€€€S´Ž ä xgg€€€€LÞŽ ä xSÿ   I  â $xTT   
Ž â xii @€føŽ â $xgg   b
Ž â xSÿ   ÿÀŽ ä xbÿ   6  Ä xTT€€€€Q´Ž â 	sS         b xQIT  6    ä xii˜º  ÿ    ä xebÿ   ÿÀŽ ä xcÿ   Q  ä xdiÿÀŽ â 	sf e        ¢ xb˜º  ÿ    Ä xQ˜º  ÿ    â $xaa€   ÿ Ž â xii€€€€NÞŽ ä xbÿÀŽ äxQÿÀŽ â $xii   d
Ž â xbb€€€€KÞŽ ä xQQ€€€€RÞŽ Ä xccÿ   ÿÀŽ ä xda€   ÿÀŽ ä rbb  €ÿàÿ â 	sa c        " $xQ   
Ž â xddÿ   IxŽ â $xS€   ÿ Ž ä/ $xff€   ÿ Ž ÄO $xddÿŽ â x€   xŽ â $tIÿÿ Ž â x€€€€R´Ž â $tSÿÿ Ž â xQ  @€    â $xRÿŽ â xfe€   fxŽ â $teÿ @€ÿ Ž â xd€€€€IìŽ ä xKb€€€€K´Ž â $xba€   ÿ Ž â xId€€€€QìŽ Ä xÿþþþÿàÿ ä xR€€€€SìŽ äxSR€€€€eìŽ ä xQIÿþþþÿàÿ â $xffÿŽ â xIdøŽ ä xbc€   bxŽ â $tcÿÿ Ž â xÿþþþÿàÿ Ä xaSÿþþþÿàÿ ä xSRøŽ äxaa @€RøŽ ä xf€€€€cìŽ â $xcbÿŽ â xRf€€€€eìŽ ä xb     Ä xQQ @€døŽ â $tdÿ @€ÿ Ž â xeRÿþþþÿàÿ ä xRc€€€€bìŽ ä xÿþþþÿàÿ ä xII˜º  ÿ    ä xe @€føŽ Ä xec€€€€dìŽ ä xdRÿþþþÿàÿ ä xRIÿÀŽ äxII€€€€JÞŽ äxføŽ ä xbQ˜º  ÿ    â $xII   R
Ž â xfeÿþþþÿàÿ Ä xSS˜º  ÿ    ä xedcøŽ ä xda˜º  ÿ    ä xQbÿÀŽ äxRSÿÀŽ ä xbb€€€€MÞŽ Ä xSS€€€€HÞŽ ä xJI€€€€J´Ž ä xIdÿÀŽ äxdd€€€€GÞŽ â $xbb   Q
Ž ä $xSS   R
Ž â rRdI  €ÿàÿ Ä xee˜º  ÿ    ä xd˜º  ÿ    ä xMb€€€€M´Ž ä xbeÿÀŽ äxdÿÀŽ ä xee€€€€2ÞŽ Ä xdd€€€€1ÞŽ ä x˜º  ÿ    â $xee   b
Ž ä $xdd   
Ž â xbÿÀŽ äx€€€€5ÞŽ â &rP0   ÿ   È &rOD      ä $xb   b
Ž â xff @€cøŽ â &rC      â xLg€€€€L´Ž â &rJ   ÿ   â xff˜º  ÿ    â &rTB      â xHS€€€€H´Ž â &r0J0   ÿ   â xNi€€€€N´Ž â &rMA      âxIfÿÀŽ â &rL@      â xff€€€€3ÞŽ â &r0MD   0   â xRR€€€€G´Ž â &rH?      Ä &rN>      â x2e€€€€2´Ž â &rCHC   0   ä $xff   I
Ž ä &rR=      ä&rK<      ä &rCRB   C   â x3f€€€€3´Ž â &r02;      Ä &r:      ä &rC2@   C   â x1d€€€€1´Ž ä r   ÿ� â &rC3>   C   â x5b€€€€5´Ž ä x7ÿ   7  â &rC1<   C   â xÿ     Æ &rC5:   C   ä $r7   Ž ä &rDP   ÿ   â rCC   ÿ� ä r   ÿ� â 0rÿ    A   â x6ÿ   6  â 0rEÿE    A   â xÿ     Ä xÿ     â &rDOA   D   ä $rC6C   Ž â Er        â  rE     A  â &rD?   D   Æ #r4   4   â &rP   ÿ   ä&rJ   ÿ   ä &rPP   ÿ   ä&rJJ   ÿ   ä &rDT=   D   ä &rOX      Ä &rMX      ä &rPOW   P   ä&rJMW   J   ä &rDL;   D   ä &rV      ä&rHV      Ä &rPU   P   ä&rJHU   J   ä &rDN9   D   ä&r:39   0   ä &rT      ä&rR      Ä &rPT   P   ä&rJR   J   ä &rDK/   D   ä&r:1/   :   ä &rL      ä&r2      Ä &rPL   P   ä&rJ2   J   ä &rD8   D   ä&r958   :   ä &rN^      ä&r3^      Ä &rPN]   P   ä&rJ3]   J   â r0DD   ÿ� â &rK\      âr/99   ÿ� â &r1\      ä &rPK[   P   ä&rJ1[   J   â x0ÿ   0  â &rZ      âx8ÿ   /  â &r5Z      Ä &rPY   P   ä&r55Y   J   â r   ÿ� â $r07D   0Ž â r   ÿ� â $r869   8Ž â r55   ÿ� ä rPP   ÿ� ä r000   ÿ� Ä xÿ     ä r888   ÿ� ä xÿ     ä xÿ     ä xÿ     â 0rÿF    A   â x0ÿ   0  â 0rÿ    A   â x8ÿ   8  â $r7   Ž â x..   ÿàÿ â $r6   Ž Ä $r65   Ž ä $r77P   Ž â Er/ 0       â  r0     A  âEr8 8       â  r     A  â | .   pbô ä rCCC   ÿ� ä r   ÿ� Ä r   ÿ� ä r   ÿ� ä r777   ÿ� â 0r_ÿ_    A   â #r8      â xCÿ   C  â 0rÿ`    A   â xÿ     ä xÿ     Ä xÿ     ä x7ÿ   7  â  r_     A  â  rEE     A  â ErC C       â  r__     A  âEr        â  r     A  âEr        â  r     A  â Er        Ä Er7 7       ä x   ÿà÷ â #rEC      â #r-0/   -   â #r        â #r_      â #r+   +   â #r#7   #   â x€   ÿàÿ ä x&&€   ÿàÿ Ä x''€   ÿàÿ ä x**€   ÿàÿ ä r%ÿ%   ÿäÿ â D)       À â Gy  €áÿÿÿÿƒ ê Ay        € ê Ey  °     € â G‰  �     € ê $x    Ž Ê ˆs 4 üÿ H   è ˆs €üÿ H   è ˆs - ýÿ H   è ˆs €ýÿ H   è ˆs   þÿ H   è ˆs €þÿ H   è ˆs # ÿÿ H   è ˆs +€ÿÿ H   ä Ay        € ê {          ì M        € ê „y    H   b z  r  p`ð Æ „y €  H   ¢ r ÿ   p"p Æ „y     H   ( „y €  H   è „y    H   ( „y €   H   h „y     H   h „y    H   h „y €	  H   h „y €  H   h „y    H   b !r        Æ/ „y €  H   b !r       ÆO „y    H   ¢ !r4       Æ „y    H   " !r       Æ� „y	 €  H   â !r       Æ„y
 €  H   " !r       Æ„y  	  H   b !r         Æ „y    H   b !r-       Æ „y  
  H   b !r       Æ „y €
  H   b !r       Æ „y!    H   b !r         Æ „y €  H   â !r       Æ/ „y €  H   â !r       ÆO „y €  H   â !r#       â !r	       ä� ‰       " !r

       Æ‰      b !r       Æ‰	      ¢ !r       Æ ‰      â !r       Æ y      &   " !r

       Æ ‰      b !r!       Æ ‰
      " !r        â !r+       ä ‰      " !r       â/ !r       ä ‰       b !r		       âO !r       â !r       â� ‰      ¨ ‰      â !r       Æ‰	      b !r
       Æ‰      " !r       Æ ‰
      " !r         Æ/ ‰      h ‰      b !r       ÆO ‰  €    ¢ !r       â� !r	       ä‰ €    â !r       Æ‰ €    " !r
       Æ ‰	 €    " !r       Æ/ ‰ €    b !r       Æ ‰ €    b !r        ÆO ‰ €    ¢ $z  x  Ž â x   ÿàÿ ä ‰      ¢ !r       â� z  r  p`ò ä y      '   â !r       âx    p"ò Ä ‰ @    " !r
	       â !r       ä/ ‰ @    " !r       Æ‰ @    b !r       ÆO ‰
 @    ¢ !r       Æ ‰ @    b $z  ~   Ž Æ� ‰ @    â $vÿ r  ÿ Ž â !r       ä‰ @    " 
  r  ÿàõ â $x   ÿŽ ä ‰ €    " !r	       â $ÿÿ   ÿ â 
 j  ÿ„ â ‰      " !r       â/ r    ÿ† Ä 
 k   â !r
       âO ‰	      b    ÿàÿ â !r       â$rÿÿ   ÿŽ â ‰      ¢ z j  ÿ† â!r       â�     ÿàõ â ‰      â z k  � â !r       â    ÿà÷ â ‰      " $ÿÿ   ÿ â !r       â  j  ÿ„ â ‰      b $ÿÿ   ÿŽ â !r       â x        â ‰ @    " 

 j  ÿ† Æ ‰      " !r	       â/ %v  j   Ž â 	 k   â !r       âO 
 k  � ä †	     â !r       Æ� †    â !r       Æ†	     â !r       Æ†     â !r       Æ †	     â !r       Æ †    è ‰      ( †	 
    â M™        € ê x   ÿàÿ â/ !r       Æ r    ÿàñ Ê $r ÿÿ   ÿ  â z j  ÿ€ È z k     Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      "   " ¹z  l      â Ey  P    € â ™x?    â y      !   b ¹z  F   
   â $r!ÿÿ   ÿ Ž ä y      %   ¢ $rÿÿ   ÿ Ž ä $r)ÿÿ   ÿ Ž Ä $rÿÿ   ÿ Ž ä $r+ÿÿ   ÿ Ž â r ÿ   pRð är ÿ   pRò â$x    Ž ä/ $x   ÿ Ž ÆO x ÿ     È r    ÿ� â $r ÿÿ   ÿ Ž ä $À   Ž â x*ÿ     È | *   pbô Ú G)       € ê y      '   " xøÿÿÿÿÀŽ â $rÿÿ   ÿ Ž â z#  Y      â y      &   b $t	ÿ   ÿ Ž ä $x   
Ž ä $rÿÿ   ÿ Ž ä $x&*   ÿ Ž â x!   ÿŽ â $v'ÿ X  ÿ Ž â x%ÿ     â $r!ÿÿ   ÿ Ž â z$& q  ÿàÿ â 'z y   Ž È 'z s   Ž â/ zÿ z    Æ $z p  *Ž â zÿ t    â $z }  ÿŽ ä $z |  Ž ä$z |  *Ž ä $z v  Ž ä$z v  Ž Ä %v  € 	Ž È %xb   ÿŽ È $z w  ÿŽ â x   ÿàÿ ä r    ÿà÷ ä r   ÿàõ â $z	 p  Ž ä $x"   Ž ä%x	b   Ž È $vÿ q  ÿ Ž ä %x	b   ÿŽ È $x   ÿ Ž ä $x   &Ž ä %x
   Ž È $r(ÿÿ   ÿ ä %x   Ž È r
'   ÿàõ Ê $x   # Ê �y   • ¨ �y   • è �y0   • ( �y8   • b r'   ÿàõ Ê $x   # Ê �y<   • h	 �y   • h	 �y9   • h	 �y   • b	 r'    ÿàõ Ê $x#   " Ê �y4   • h �y/   • b $tÿ   ÿ Ž ä $tEÿ$   ÿ Ž ä $x   ÿ Ž âO xÿ     È xü  ÿÀŽ â $x   ÿ Ž ä� %v   Ž â z,  ÿàõ ä xü  ÿÀŽ â $x0   ÿ Ž â�y    ™ ¢  $v-ÿ  ÿ â r'   ÿàõ È �y,,    ™ â r#   ÿä ä z  ÿàõ ä xÿ     â �y7   • ¢ xü  ÿÀŽ â $vÿ  ÿ ä %v   Ž â z2  ÿàõ â �y.   • ¤ $x8   ÿ Ž Ä�y5    ™ b $v3ÿ  ÿ Æ �y6    ™ ¢ x:ü  ÿÀŽ ä xÿ   0  â �y12    ™ ¢ z::  ÿàõ Æ %v   Ž È $v;ÿ  ÿ â �y    ™ h �y::    ™ b x-ÿ   8  Ê/ %v-   Ž Ê �y0    ™ b $x<   ÿ Ž Êxü  ÿÀŽ È z>  ÿàõ â $x   ÿ Ž È $v?ÿ  ÿ â xü  ÿÀŽ ä x=ÿ   <  ä �y3>    ™ b	 z  ÿàù ä/ %v<=   Ž â r8   &àµ Æ $vÿ  ÿ â �y2<    ™ b %v8 Z  E Ž â r%(   ÿd~ ä �y8    ™ b $x9   ÿ Ž â xÿ     â $x-$   ÿŽ â x=ÿ   9  Æ/ %v   Ž âxü  ÿÀŽ ä x>   ÿ   â$x   -Ž ä %v=   Ž â z  ÿàõ â �y-    ™ " x@>ü  ÿÀŽ Æ �y9   ™ " $vÿ  ÿ Æ �y>    ™ " z@  ÿàõ Æ �y;   ™ ( �y<    ™ " $vÿ  ÿ Æ �y=   ™ " r   $àÇ ä/ xÿ     â �yD    ™ " rG%(   ÿ„þ Æ �y?   ™ " %v   Ž Æ �yA   ™ " %v Z  E Ž â rF   àÇ ä �y@   ™ " $xG$   ÿŽ â/ rG%(   ÿ„þ ä �y    ™ " $x   Ž Æ �yC   ™ " %vF Z  E Ž Æ �yE   ™ " $xGG$   ÿŽ Æ �yB    ™ " $x   GŽ Æ �yF   ™ ( �yR   ™ ( �yG   ™ ( �yQ   ™ ( �yP   ™ " r'   ÿàõ Æ �yO   ™ ( �yN   ™ " $x   # Æ �yM   ™ ( �yL   ™ ( �y    • (# �yK   ™ ( �yJ   ™ " r'   ÿàõ Æ/ �yI    ™ $ $x	   # ä �yH   ™ ( �y    • ( �y    • ( �yS    ™ ( �y   ™ ( �yT    • ( �yU    • " xX4ÿ   ÿÀŽ Ì 	sX X        b xVÿ   4  ä xYÿ   /  äxVVÿ   ÿÀŽ ä x4T  /    ä xYÿ   ÿÀŽ â/ 	sW V        " xÿ     â $xYX€   ÿ Ž Æ xÿ   ÿÀŽ Æ 	s         b xYY€   ÿÀŽ È xYÿ   4xŽ Æ 	s         b $t]ÿ @€ÿ Ž ä $xÿŽ ä $t[ÿÿ Ž ä $xYW€   ÿ Ž â xX€€€€]ìŽ äxW€€€€[ìŽ Ä x4V€   YxŽ ä xYXÿþþþÿàÿ â $xX€   ÿ Ž â/ xWWÿþþþÿàÿ â $x€   ÿ Ž â xVY @€øŽ äxWøŽ â $x44ÿŽ â xW€   xŽ â $tÿÿ Ž Ä $tYÿ @€ÿ Ž â x€   XxŽ ä x4€€€€ìŽ äx4€€€€YìŽ â $xÿŽ â xÿþþþÿàÿ â $tZÿ @€ÿ Ž â xX     ä xYÿþþþÿàÿ Ä x4øŽ ä xX€€€€XìŽ ä xY @€4øŽ â $x4WÿŽ â xZ€€€€ZìŽ ä xW˜º  ÿ    â $tYÿÿ Ž â xXÿþþþÿàÿ â $t[ÿ @€ÿ Ž â xZZÿþþþÿàÿ Ä xXV˜º  ÿ    ä xVWÿÀŽ äxWW€€€€,ÞŽ ä� xøŽ äxZ @€øŽ ä xY4€€€€YìŽ Ä xZ4€€€€[ìŽ â $xWW   V
Ž â xXÿÀŽ äxXX€€€€ÞŽ äO xYYÿþþþÿàÿ ä x[Zÿþþþÿàÿ â $xXX   
Ž â x,W€€€€,´Ž Ä xVY4øŽ äx[ @€4øŽ ä xW˜º  ÿ    ä x4˜º  ÿ    ä xX€€€€´Ž ä xZWÿÀŽ Ä x4ÿÀŽ äxX4€€€€5ÞŽ äxWW€€€€6ÞŽ ä xVV˜º  ÿ    â $xXX   
Ž â xÿ   7  â $xYW   Z
Ž â xWVÿÀŽ Ä xZV€€€€1ÞŽ ä x\ÿ   .  ä xÿ   ÿÀŽ ä x5X€€€€5´Ž â $xXZ   W
Ž â x[7ÿ   ÿÀŽ ä xW˜º  ÿ    â 	s         " xV\ÿ   ÿÀŽ Ä x6Y€€€€6´Ž ä xY˜º  ÿ    ä xWÿÀŽ â 	s4 [        ¢ x]˜º  ÿ    ä xYÿÀŽ äx\Y€€€€:ÞŽ Æ 	s V        â x[W€€€€ÞŽ Ê/ $x[[   
Ž â xX€€€€1´Ž â $x1\   
Ž Æ x[€€€€´Ž ä x7T  .    ä x   ÿ   â $x44€   ÿ Ž âO xÿ     â $x€   ÿ Ž â� x€   xŽ ä xÿ   ÿÀŽ Ä x44€   ÿÀŽ ä xVV€   xŽ ä 	s         " x:1€€€€:´Ž ä xW]ÿÀŽ äxZ]€€€€0ÞŽ ä x14ÿ   7xŽ â $tÿÿ Ž Ä $xWZ   W
Ž ä $x11ÿŽ ä $xÿŽ ä $tZÿ @€ÿ Ž ä $t4ÿ @€ÿ Ž â x0W€€€€0´Ž â $tXÿÿ Ž â x1€€€€ìŽ Ä xW€€€€ZìŽ ä x41€€€€4ìŽ â $x€   ÿ Ž â xÿþþþÿàÿ â $xVVÿŽ â x7€€€€XìŽ â $tYÿ @€ÿ Ž â x\Wÿþþþÿàÿ â $tWÿÿ Ž â xX4ÿþþþÿàÿ Ä x41øŽ ä x€   xŽ ä x1X @€1øŽ ä xWV€€€€WìŽ äxXV€€€€YìŽ ä x4˜º  ÿ    â $x4ÿŽ â xWÿþþþÿàÿ Ä xYXÿþþþÿàÿ â $t[ÿ @€ÿ Ž â xXÿÀŽ äxW€€€€3ÞŽ ä xZ7ÿþþþÿàÿ â $t7ÿÿ Ž â xVøŽ äxY @€VøŽ â $xVW   X
Ž â xX4€€€€[ìŽ Ä xZøŽ ä x11˜º  ÿ    ä xW4€€€€7ìŽ ä x[Xÿþþþÿàÿ ä x7˜º  ÿ    ä xX1ÿÀŽ Ä x11€€€€2ÞŽ ä xYWÿþþþÿàÿ ä x7ÿÀŽ äxW7€€€€8ÞŽ â $x11   X
Ž â x\ @€øŽ ä x7Y4øŽ Ä rWW  €ÿàÿ ä x4[ @€4øŽ ä xV€€€€3´Ž ä xV˜º  ÿ    ä x1€€€€2´Ž ä x1W€€€€8´Ž Ä xW4˜º  ÿ    ä x3VÿÀŽ â &r2,9   ÿ   âxVV€€€€-ÞŽ ä x4WÿÀŽ äxWW€€€€>ÞŽ ä x77˜º  ÿ    â &r2;   2   Ä $xVV   3
Ž â x87ÿÀŽ â $xWW   4
Ž â x77€€€€<ÞŽ â &r49   ÿ   ä &r35=   2   â x2V€€€€-´Ž â &r4;   4   â x-˜º  ÿ    â $x77   8
Ž Ä &r41=   4   â x8-ÿÀŽ â&r36?   3   âx--€€€€DÞŽ â &r42?   4   â x<7€€€€<´Ž â &r3A   3   âxW€€€€>´Ž â $x--   8
Ž â x7˜º  ÿ    â &r4<A   4   Ä &r3@   3   âxD-€€€€D´Ž â &r4@   4   â x87ÿÀŽ â&r3:C   3   âx77€€€€ÞŽ â &rCDC   4   ä &r4,E   ÿ   ä &r0B   3   Ä $x77   8
Ž ä &r4F   4   â r-   ÿ� ä x7€€€€´Ž â &r45R   4   â x/ÿ   /  â &rEE   ÿ   âx-ÿ   -  â &rG   ÿ   Ä &rCB   C   ä &r,,G   ÿ   ä &r46Q   4   ä &rEF   E   ä&rP      ä $r-/   -Ž â rCC   ÿ� â &r,P   ,   Ä &r4O   4   ä &rE1R   E   ä&r1N      âx.ÿ   .  â &r55N   ,   â xÿ     â &r4M   4   ä &rE2Q   E   Ä &r2L      ä&r56L   5   ä &r4:K   4   ä $r.C   Ž ä &r<O   E   ä&r<J      Ä &r5J   5   ä &r40I   4   â r   ÿ� â &rM      ä&rH      ä&r5H   5   â r44   ÿ� â 0rÿ    A   â xÿ     â 0rÿ    A   Ä &rD      ä&rDK      ä &r5:   5   â xÿ     â &rS      âEr        â &rII      â  r     A  â &r00S   5   â r   ÿ� Ä #r+   +   â rII   ÿ� ä r00   ÿ� â $r/4   Ž â xÿ     ä xÿ     ä xÿ     â $r.   Ž âr   ÿ� â $r.I   Ž â x**   ÿàÿ â $r/0   Ž â xÿ     â 0rÿ    A   â | *   pbô â 0rÿT    A   â r---   ÿ� Ä r   ÿ� ä r   ÿ� ä r   ÿ� â  r     A  â Er        â 0rUÿU    A   â x-ÿ   -  ä xÿ     Ä xÿ     ä xÿ     â #r      â Er3 -       â  r-     A  â Er        â  r     A  âEr        â  rU     A  â Er        â  rU     A  â x''   ÿà÷ â #r -3       â #r)   )   â #r!   !   â #r      â x€   ÿàÿ â $r#ÿÿ   #Ž â x$$€   ÿàÿ Ä x&&€   ÿàÿ â D)       À â Gy  €äÿÿÿÿƒ ê Ay        € ê ˆ   ýÿ H   è ˆ +€ýÿ H   è ˆ  þÿ H   è ˆ )€þÿ H   è ˆ  ÿÿ H   è ˆ !€ÿÿ H   è {          ì M        € ê „y     H   " z  r  p`ð Æ „y    H   b r ÿ   p"p Æ „y €   H   ¨ „y    H   è „y €  H   ( „y    H   h „y
    H   h „y €  H   h „y    H   h „y €  H   h „y    H   b !r        Æ „y	 €  H   " !r       Æ/ „y    H   b !r+       ÆO „y €  H   ¢ !r       Æ� „y €  H   â !r)       Æ„y    H   â !r       Æ„y €  H   â !r

       Æ „y €  H   â !r       â !r       ä ‰      " !r       Æ y      &   b !r

       Æ y      '   b !r	       â !r       È/ ‰      " !r       ÆO ‰      b !r!       Æ� ‰	      ¢ !r       â!r

       È ‰       â !r       Æ ‰
      " !r       Ê ‰      b !r       â !r       È/ ‰      " !r		       ÆO ‰      h ‰	      ¢ !r         â� !r
       È‰  €    è ‰      " !r       Ê‰
      b !r       â !r       È/ ‰ €    " !r	       ÆO ‰ €    h ‰	 €    ¢ !r        â� !r       È‰  @    è ‰ €    " !r

       Ê‰
 €    b !r       â !r       È/ ‰ @    " !r		       ÆO ‰ @    h ‰	 @    ¢ !r         â� !r       È‰       è ‰ @    " !r
       âx
   ÿàÿ â $z x  Ž Æ ‰ @    b z 
 r  p`ò â !r       â $z ~  Ž â !r       â/ x    p"ò ä ‰
      " !r	       âO 
 r  ÿàõ È ‰      b $ÿÿ   ÿ â 
 j  ÿ„ â !r        â� !r       ä
 k   â ‰      ¨ ‰      â !r       â$vÿ r  ÿ Ž È ‰      "    ÿàÿ â!r
       â r	   ÿ† ä    ÿàõ â $tÿ   ÿ Ž â !r       â/ $r ÿÿ   ÿŽ â z	 j  ÿ† â $ÿÿ   ÿ â  j  ÿ„ â %v j   Ž â z	 k   � â !r       âO 	 k   Ä †	     â !r       Æ� †    â !r       Æ†	     è †     è †	     â M™        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      "   " ¹z  l      â Ey      € â ™x?    â y      !   b ¹z  F   
   â $rÿÿ   ÿ Ž ä y	      %   ¢ $rÿÿ   ÿ Ž ä $rÿÿ   ÿ Ž Ä $rÿÿ   ÿ Ž â r ÿ   pRð är ÿ   pRò â$x    Ž ä/ $x		   ÿ Ž ÆO xÿ     È r    ÿ� â $
€   Ž Æ xÿ      È |    pbô Ú G)  à    € ê y      '   " $rÿÿ   ÿ Ž â x øÿÿÿÿÀŽ â $rÿÿ   ÿ Ž â y      &   b $z 	 p  Ž ä $x   
Ž ä $t)ÿ   ÿ Ž ä $x'   ÿ Ž âx(!   ÿŽ Ä xÿ     Æ %v((  € )Ž È 'z y   Ž È 'z s   Ž â/ zÿ z    Æ $x   ÿ Ž â zÿ t    â $z |  Ž ä$z |   Ž â x 	   ÿàÿ â $z v  Ž â z q  ÿàÿ â $z$ v  Ž ä $z }  ÿŽ Ä $z w  ÿŽ ä %x$$b   ÿŽ Æ r   ÿàõ â $z  p  Ž â r$(   ÿà÷ â %x*'   $Ž È %xNb   ÿŽ È $x %   )Ž ä%x(b   (Ž È $vÿ X  ÿ Ž ä $vÿ Y  ÿ Ž ä $rÿÿ   ÿ Ž ä $rÿÿ   ÿ ä %x&'   NŽ È r*   ÿàõ Ê $x+    Ê �y   • ¨ �y   • è �y   • ( �y,   • b r&   ÿàõ Ê $x'    Ê �y    • ( �y!   • ( �y#   • h �y"   • b r   ÿàõ Ê $x     Ê �yI   • h �y   • b $t0ÿ   ÿ Ž ä $t=ÿ$   ÿ Ž ä $x   ÿ Ž âO xÿ     â $x   ÿ Ž Æ� xü  ÿÀŽ È zJ  ÿàõ ä xü  ÿÀŽ Æ $vKÿ  ÿ â z  ÿà÷ â %v  0 Ž Æ/ �yJJ    ™ ¢  $vÿ  ÿŽ â r(   ÿàõ ä �yH    ™ â xÿ     Æ �yE    ™ ¢
 $x   ÿ Ž ä$x)    Æ xü  ÿÀŽ â %v  0 Ž â �yG   • ä z  ÿàõ ä �yD    ™ â  xÿ     ä $vÿ  ÿ ä $x,   ÿ Ž Æ�yC    ™ â
 %v  0 Ž â xü  ÿÀŽ ä �y   • ä z  ÿàõ ä �yB    ™ æ  $vÿ  ÿ â xÿ   ,  È/ �y    ™ â %v  0 Ž Ì �y    ™ â  $x    ÿ Ž ä$x,!   ÿ Ž Æ xü  ÿÀŽ ä xÿ      ä z   ÿàõ ä x,,ü  ÿÀŽ ä xÿ   !  â$v!ÿ  ÿ â z,  ÿà÷ â $x,#   ÿ Ž Ä %v  0 Ž â�y     ™ $ x,,ü  ÿÀŽ â %v  0 Ž â/ �y    ™ b x#ÿ   #  ä $vÿ  ÿŽ â �y    ™ " r>   àµ Ä z,  ÿàù â/ �y    ™ " %v #  0 Ž â r-   ÿd~ Æ $vÿ  ÿ â x1ÿ   "  â %v>> Z  = Ž Æ �y    ™ " $x--$   ÿŽ â r<   àÇ ä �y     ™ "# $x,"   ÿ Ž â r"   ÿ„þ â %v01  0 Ž È $x??   -Ž â x,,ü  ÿÀŽ â �y0    ™ " %v<< Z  = Ž Æ �y>   ™ " $x!"$   ÿŽ â/ z,,  ÿàõ ä �y >   ™ " $x==   !Ž ä $v-ÿ  ÿ â �y!>   ™ ( �y"<   ™ ( �y#>   ™ ( �y,    ™ ( �y.<   ™ ( �y/>   ™ ( �y,<   ™ ( �y->   ™ ( �y0<   ™ ( �y1>   ™ ( �y2<   ™ ( �y3>    ™ ( �y4<   ™ ( �y5<   ™ h �y6<    ™ b r@$   ÿàõ Æ �y8>    • d $xA%    âr:N   ÿàõ â �y9<    • h $x;O    â �y7@    • h �y::    • b xMIÿ   ÿÀŽ Ì 	sM M        b xKÿ   I  ä x@ÿ     â $xFM€   ÿ Ž Ê/ xFF€   ÿÀŽ È xAFÿ   IxŽ ä xIIT      ä xKKÿ   ÿÀŽ ä xF@ÿ   ÿÀŽ ä x>ÿ   I  â 	sL K        & x>>ÿ   ÿÀŽ Ê 	s= F        b $xAAÿŽ Î 	s; >        b $t<ÿÿ Ž ä $t@ÿ @€ÿ Ž ä $xLL€   ÿ Ž â x<A€€€€<ìŽ â $x==€   ÿ Ž â/ x@A€€€€@ìŽ ä x<<ÿþþþÿàÿ Ä x@@ÿþþþÿàÿ ä xKK€   LxŽ â $x;;€   ÿ Ž â xFF€   =xŽ ä xM<AøŽ â$xKKÿŽ â xA@ @€AøŽ â $t@ÿÿ Ž â x>>€   ;xŽ â $tIÿÿ Ž Ä $xFFÿŽ ä $t<ÿ @€ÿ Ž ä $t;ÿ @€ÿ Ž â x@K€€€€@ìŽ äxIF€€€€IìŽ â $x>>ÿŽ â x<K€€€€<ìŽ â $t?ÿÿ Ž â x;F€€€€;ìŽ â $t=ÿ @€ÿ Ž â x@@ÿþþþÿàÿ Ä xIIÿþþþÿàÿ ä x<<ÿþþþÿàÿ ä x;;ÿþþþÿàÿ ä xMM˜º  ÿ    ä x?>€€€€?ìŽ ä x@@KøŽ Ä x=>€€€€=ìŽ ä xIIFøŽ äxK< @€KøŽ ä xF; @€FøŽ ä x;MÿÀŽ äx<M€€€€JÞŽ ÄO x??ÿþþþÿàÿ ä x==ÿþþþÿàÿ ä xAA˜º  ÿ    ä x@@˜º  ÿ    â $x;<   ;
Ž â x??>øŽ äx>= @€>øŽ Ä x<AÿÀŽ äx=@ÿÀŽ äxAA€€€€HÞŽ ä� x@@€€€€EÞŽ Æ $xAA   <
Ž ä $x<@   =
Ž â xKK˜º  ÿ    È x<<€€€€E´Ž ä xEGÿ   ÿÀŽ ä x@KÿÀŽ äxKK€€€€DÞŽ ä x;;€€€€J´Ž â 	sE E        " xJ?˜º  ÿ    Ä x=A€€€€H´Ž â $xAK   @
Ž â x@JÿÀŽ äxKJ€€€€CÞŽ ä xM>˜º  ÿ    ä x?ÿ   G  â $xJK   @
Ž â xHMÿÀŽ Ä xMM€€€€BÞŽ ä x>A€€€€D´Ž ä x??ÿ   ÿÀŽ ä xDÿ     ä xAJ€€€€C´Ž â $xJE€   ÿ Ž ä $xMM   H
Ž â xHGT      â 	s@ ?        " xCDÿ   ÿÀŽ Ä xLI˜º  ÿ    ä xJJ€   ÿÀŽ ä xEÿ   H  â 	sD C        b xHLÿÀŽ äxKL€€€€ÞŽ ä xJJÿ   GxŽ Ä xEEÿ   ÿÀŽ ä xIF˜º  ÿ    â $xFK   H
Ž â xBM€€€€B´Ž â 	sG E        ¢ $xJJÿŽ ä $tKÿÿ Ž ä $tMÿ @€ÿ Ž â xHIÿÀŽ â$x@@€   ÿ Ž â xII€€€€ÞŽ Ä xKJ€€€€KìŽ äxMJ€€€€MìŽ â $xDD€   ÿ Ž â/ xFF€€€€´Ž â $xII   H
Ž â xKKÿþþþÿàÿ ä xMÿþþþÿàÿ ä x@?€   @xŽ Ä xMKJøŽ â$tHÿÿ Ž â xJ @€JøŽ â $x@ÿŽ â x?C€   DxŽ ä xII€€€€´Ž â $xG€   ÿ Ž âO xH€€€€HìŽ â $x??ÿŽ ä $t@ÿÿ Ž â xE€   xŽ â $tDÿ @€ÿ Ž â xHHÿþþþÿàÿ â $tLÿ @€ÿ Ž â x@?€€€€@ìŽ â$xÿŽ â xC?€€€€DìŽ â $tKÿÿ Ž Ä $tGÿ @€ÿ Ž â xL€€€€LìŽ ä xDHøŽ ä x@@ÿþþþÿàÿ ä xHCÿþþþÿàÿ ä xMM˜º  ÿ    ä xK€€€€KìŽ Ä xG€€€€GìŽ ä xLLÿþþþÿàÿ ä x@@?øŽ äx?H @€?øŽ ä xJJ˜º  ÿ    ä xCMÿÀŽ Ä xHM€€€€ÞŽ äxKKÿþþþÿàÿ ä xGGÿþþþÿàÿ ä xL @€øŽ ä xEJÿÀŽ â $xMH   C
Ž â xJJ€€€€ÞŽ ÄxKKøŽ äxGG @€øŽ ä xCD˜º  ÿ    ä x˜º  ÿ    â $xJJ   E
Ž â xDCÿÀŽ ä xÿÀŽ Ä xEC€€€€ÞŽ ä xH€€€€ÞŽ ä xKK˜º  ÿ    ä xG˜º  ÿ    â $xEE   D
Ž â xCKÿÀŽ â $xHH   
Ž â xÿÀŽ Ä xDK€€€€ÞŽ ä x€€€€ÞŽ Æ $xCD   C
Ž ä $x   
Ž â x?˜º  ÿ    ä xC€€€€´Ž ä xDÿÀŽ äxC€€€€ÞŽ â &r;   ÿ   â xM€€€€´Ž Æ &r=       â xJ€€€€´Ž â &r   ÿ   ä &r<!      ä &r;;"   ÿ   ä&r""   ÿ   â xE€€€€´Ž â &r>#      â x@@˜º  ÿ    â &r       Ä &r;=,   ;   ä&r,   "   â xH€€€€´Ž â &rA-      â x?@ÿÀŽ â &r !      â x@@€€€€ÞŽ â &r;<.   ;   ä&r.      Ä &rB/      ä &r #       âx€€€€´Ž â &r;>0   ;   ä&r0      ä $x@   ?
Ž ä &rF1      Ä &r -       ä&r;A2   ;   ä&r2      â x€€€€´Ž â $xCC   D
Ž ä &rI3      ä &r /       Ä &r;B4   ;   ä&r4      â xC€€€€´Ž â &r 1       â r   ÿ� â &r;F5   ;   ä&r5      â xÿ     â &r33       â xÿ     â &r;I6   ;   Ä &r6      â xÿ     â $r   Ž â r33   ÿ� ä r;;   ÿ� ä r   ÿ� ä xÿ     ä xÿ     Ä xÿ     â $r3   Ž â x   ÿàÿ â $r;   Ž ä $r   Ž â |    pbô ä r   ÿ� ä r   ÿ� Ä r   ÿ� ä r   ÿ� â 0r7ÿ7    A   â xÿ     â 0r8ÿ8    A   â xÿ     â 0rÿ9    A   â xÿ     â 0rÿ:    A   â xÿ     â  r78     A  âEr        â  r7     A  â Er        â  r88     A  â Er        â  r     A  â Er        Ä x   ÿà÷ â #r      â #r      â #r8      â #r      â x€   ÿàÿ â $rÿÿ   Ž â x€   ÿàÿ â D)       À â Gy  €çÿÿÿÿƒ ê Ay        € ê ˆ 
 þÿ H   è ˆ 
€þÿ H   è ˆ 
 ÿÿ H   è ˆ 
€ÿÿ H   è {          ì M        € ê „y      H   " z 	 r  p`ð Æ „y €   H   b r ÿ   p"p Æ „y    H   ¨ „y    H   è „y €  H   ( „y    H   h „y    H   h „y €  H   h „y    H   h „y €  H   h „y €  H   b !r         Æ „y
 €  H   " !r       â/ !r       äO y      &   b !r         â� !r       â!r       ä!r         â !r       È ‰       ¢ !r       Æ ‰      â !r       Æ ‰      " !r       È !r
       â !r        ÈO ‰       " !r       Æ� ‰      ¢ !r       Æ‰      è ‰      " !r        â !r       ÈO ‰       " !r       Æ� ‰ €    ¢ !r
       Æ‰ €    è ‰
 €    " !r        â !r       ÈO ‰ €    " !r       Æ� ‰  @    ¢ !r
       Æ‰ @    è ‰ @    ( y      '   b !r
       â !r         âO !r       È� ‰       " x	   ÿàÿ â $z	 x  	Ž â/ !r       â‰      b z  r  p`ò â $z	 ~  	Ž ä‰
 @    ¢ x    p"ò â $tÿ   ÿ Ž Ä ‰      â 
	 r  ÿàõ Ê $ÿÿ   ÿ â 
 j  ÿ„ â !r        â %v	 j   Ž â !r       â/ 
 k   Æ †	     â !r
       ÆO †    â !r       Æ� ‰       h  †	     â M™        € ê ¹z  r      â !r        â? �x   ?àÿ Ì |		   ÿàñ Ê $rÿÿ   ÿ  â z	 j  ÿ€ È z	 k    Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y       &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž ä $rÿÿ     Ž È % \   Ž Ê �	    ™ b ¹z  l      â Ey  €    € â ™x?    â y      "   b $rÿÿ   ÿ Ž Æ y      !   ¢ $ÿÿ   ÿ Ž ä $ÿÿ    Ž È '
 m   Ž È $‚ÿÿ   ÿ Ž â 
ÿ n    Ê $ÿÿ   
Ž â r ÿ   pRò Æ/ $
 o  Ž â r ÿ   pRô â$x    Ž äO $‚ÿÿ    Ž Æ xÿ     â 'Š s   Ž ä y      '   " r   ÿ� È xÿ     È |    pbö ä Šÿ t    Ö G9       € ê y	      %   " $rÿÿ   ÿ Ž â xøÿÿÿÿÀŽ â $z w  ÿŽ ä 'z y   Ž È $x   
Ž â zÿ z    â $z }  ÿŽ ä $tÿb   ÿ Ž â xÿ     â $z |  Ž â r   ÿàñ â $x   ÿ Ž ä $z v  Ž Ä%x   ÿŽ È $z	 p  Ž ä $tÿ   ÿ Ž ä %v X  Ž â x    ÿŽ Æ $rÿÿ    Ž â x   ÿüŽ â $rÿÿ   ÿ Ž ä $x   ÿ Ž ä $rÿÿ   ÿ  ä %v  € Ž È r   ÿàñ Ê $x	     Ê �yþÿÿ • ¨ �y    • è �y   • ( �y(   • b r
   ÿàñ Ê $x     Ê �y
þÿÿ • ( �y 
    • " $t)ÿ   ÿ Ž ä $t$ÿ$   ÿ Ž ä $x   ÿ Ž âO xÿ     È xü  ÿÀŽ È z  ÿàñ â $x   ÿ Ž È� $vÿ  ÿ  â xü  ÿÀŽ â %v  ) Ž â r	   à± ä �y    ™ ¢  z
  ÿàù ä/ r"   ÿd~  â �y!    ™ â
 %v	 Z  $ Ž È $x#"$   ÿŽ â xÿ     â $x   ÿ Ž ä$x		   #Ž ä $vÿ  ÿ â xü  ÿÀŽ â %v  ) Ž â �y#   ™ " xÿ     Ä $x(   ÿ Ž â�y"
    ™ h �y$    ™ b xü  ÿÀŽ Æ �y&   ™ b z
  ÿàñ â/ %v  ) Ž ä �y%   ™ d $vÿ  ÿ  â z  ÿàñ â �y    ™ b xÿ   (  Æ �y'
    ™ b $vÿ  ÿ  ä %v  ) Ž â �y(   ™ h �y    ™ h �y    ™ h �y+   ™ h �y*   ™ h �y)   ™ h �y    ™ b# $r
ÿÿ    Ž Ä $rÿÿ    Ž â �y,    • h �y

    • b x-ÿ   ÿÀŽ ä xÿ     È 	s- -        " xÿ   ÿÀŽ ä xÿ      Ê 	s/         " x1ÿ   ÿÀŽ ä xT       â/ $x-€   ÿ Ž Æ xÿ     ä x.€   ÿÀŽ â 	s2 1        " $x0/€   ÿ Ž ä x.ÿ   xŽ ä xÿ   ÿÀŽ ä x€   0xŽ â $t.ÿÿ Ž â 	s         b $xÿŽ Ä $t0ÿ @€ÿ Ž Æ x	€€€€.ìŽ äx-€€€€0ìŽ ä x.	ÿþþþÿàÿ â $x22€   ÿ Ž â x0-ÿþþþÿàÿ ä x..øŽ äx0 @€øŽ â $x0€   ÿ Ž â/ x1€   2xŽ â $xÿŽ Ä $t/ÿÿ Ž â x€   0xŽ â $xÿŽ ä $t2ÿÿ Ž â x/€€€€/ìŽ â $t0ÿ @€ÿ Ž â x..˜º  ÿ    ä x	/ÿþþþÿàÿ â $t1ÿ @€ÿ Ž â x-€€€€2ìŽ Ä x/€€€€0ìŽ ä x.ÿÀŽ ä x0˜º  ÿ    ä x2-ÿþþþÿàÿ ä x4/ÿþþþÿàÿ ä x0ÿÀŽ Ä x1€€€€1ìŽ â $xÿŽ ä $t-ÿ @€ÿ Ž â x11ÿþþþÿàÿ ä x		øŽ äx-€€€€-ìŽ ä x1 @€øŽ Ä x		˜º  ÿ    ä x ÿ      ä x   ÿàÿ È |    pbð ä x   ÿà÷ ä x€   ÿàÿ Æ $rÿÿ   Ž â x..€€€€ÞŽ äO x00€€€€!ÞŽ Æ� $x..   
Ž â x2øŽ äx4 @€øŽ â $t4ÿÿ Ž ä $x00   
Ž â x..€€€€´Ž ä x€€€€4ìŽ ä x!0€€€€!´Ž Ä x0-ÿþþþÿàÿ â &r#.#   ÿ   âx˜º  ÿ    ä x.ÿþþþÿàÿ ä x00 @€øŽ ä xÿÀŽ ä x2	ÿÀŽ Ä x€€€€$ÞŽ äx..øŽ ä x		€€€€"ÞŽ ä x0˜º  ÿ    â $x   
Ž â x..˜º  ÿ    â $x		   2
Ž â xÿÀŽ Ä x€€€€ÞŽ ä x.ÿÀŽ äx..€€€€'ÞŽ ä x˜º  ÿ    â &r&!&   #   â x		€€€€"´Ž â $x   
Ž â xÿÀŽ â$x..   
Ž â x€€€€ÞŽ â &r		%   &   â x€€€€$´Ž Ä x˜º  ÿ    â $x   
Ž â x..€€€€'´Ž â &r((   	   â xÿÀŽ äx€€€€ÞŽ â &r(.+   (   â x€€€€´Ž ä x€€€€´Ž â $x   
Ž Ä &r(*   (   Æ x€€€€´Ž â &r)   (   È &r      Ê r   ÿ� È xÿ     Ê $r    Ž Ê r   ÿ� â 0r
ÿ
    A   ä 0r	ÿ,    A   â xÿ     È Er        â  r	
	     A  È #r	      â G‰  �ôÿÿÿÿƒ ê Ay        € ê ˆ €ÿÿ H   è {          ì M)        € ê „y     H   " r ÿ   pRð Æ „y €   H   h „y    H   ¨ y      %   â !r       È !r       È/ !r       ÊO ‰      $ !r       Ê‰      $ !r       Ê ‰ €    $ !r       Ê ‰ @    $ !r	       â $z  x  Ž È� ‰
	      b  M	        € ê $tÿ   ÿ Ž â !r		
       â? $z ~  Ž È %v j   Ž Ê †y 	    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $	ÿ   ÿ Ž È % \  	 Ž Ê �	    ™ b $rÿÿ   ÿ Ž â Ey  ð    € â $rÿÿ   ÿ Ž â y      "   b 'Š  s   Ž Æ y      !   b $rÿÿ   ÿ Ž â rÿ   ÿü‚ â/ $x    Ž ä '
 m   Ž Æ xÿ     ä 
ÿ n    ä r   ÿ� ä Šÿ t     ä y       %   b G       € ê z ÿ `  pPô ä z ÿ ^  pPò Ä z ÿ b  p ö ä z ÿ a   Sô ä z ÿ _  Sò ä z ÿ c  0't Ö       '   "        ä¢       â ©	      '   ¤ $ x  ÿŽ â ¨ÿ      â/ $ª x  ÿŽ ä $ ~  ÿŽ Ä $ª
	 ~  ÿŽ âO 	ÿ      ä    àë ä¢

    àÇ ä 	ÿÿ   	Äþ ä    ÿàý ä ¢
   ÿàû Ä ¢ÿÿ   „þ â $
ÿÿ   	 â ª b  ÿˆ ä ^  ÿ† â$¢ÿÿ   Ž ä $rÿÿ   ÿ Ž â 	 _  
� ä ª c   â $rÿÿ   ÿ Ž Ä �     h �©     d Ay        € ê ¹z  l      â xÿ     â ™x?    â $ÿÿ   
Ž â Ey       € â $r
ÿÿ   ÿ Ž ä $
 o  Ž â|    pbò â $‚ÿÿ    Ž Ä $r	ÿÿ   ÿ Ž Ô G  0    € ê y      '   " $rÿÿ   ÿ Ž â xøÿÿÿÿÀŽ â $r	ÿÿ   ÿ Ž ä $r
ÿÿ   ÿ Ž ä $x   
Ž ä 'z y   Ž Ê zÿ z    Ê $z |  Ž ä $z }  ÿŽ â xÿ     â $z v  Ž ä$x   ÿ Ž ä $z  p  Ž ä/ $z  w  ÿŽ Ä %xb   ÿŽ Æ r    ÿàñ Ê $r ÿÿ   ÿ  ä $x#   ÿ Ž ä $tÿ   ÿ Ž È %v# X  Ž Ê r   ÿàñ Ê $x     Ê �y   • ¨ �y   • è �y   • ( �y   • b x    ÿŽ Ê %v X  Ž Ê r   ÿàñ Ê $x     Ê �y%   • h �y    • b $t!ÿ   ÿ Ž ä $tÿ$   ÿ Ž ä $x   ÿ Ž äO $x   ÿ Ž Æ� x&ü  ÿÀŽ È z&&  ÿàñ ä xü  ÿÀŽ Æ $v'ÿ  ÿ  â z  ÿàó ä xÿ     â $x   ÿ Ž â�y&&    ™ ¢  $vÿ  ÿŽ  â r   à‘ â %v  ! Ž â/xÿ     ä xü  ÿÀŽ â �y"    ™ â r    ÿ$~  â %v  ! Ž Ä �y$    ™ ä %v Z   Ž ä �y'    ™ â z  ÿàñ â/ $x$   ÿŽ â xÿ     â $x   ÿ Ž ä$x   Ž Ä %v  ! Ž â xü  ÿÀŽ ä �y   ™ b $vÿ  ÿ  â z  ÿàñ ä �y    ™ b xÿ     Æ �y(    ™ b $vÿ  ÿ  Æ �y   ™ b %v  ! Ž Æ �y   ™ h �y    ™ h �y   ™ h# �y    ™ h �y   ™ h �y   ™ (�y   ™ ( �y    ™ " z X  ÿàñ Æ �y    • " z Y  ÿä  Ê �y    • " x*%ÿ   ÿÀŽ Ì 	s* *        " x)ÿ   %  È x))ÿ   ÿÀŽ â $x*€   ÿ Ž Ê x,€   ÿÀŽ ä xÿ      äx-,ÿ   %xŽ ä x%%T       â 	s+ )        " x,ÿ   ÿÀŽ ä xÿ   %  È xÿ   ÿÀŽ â 	s ,        b. $x%-ÿŽ ä $tÿÿ Ž ä $t*ÿ @€ÿ Ž Æ 	s         b x%€€€€ìŽ â $x++€   ÿ Ž â x*%€€€€*ìŽ ä xÿþþþÿàÿ ä x**ÿþþþÿàÿ â $x€   ÿ Ž â/ x))€   +xŽ ä x+%øŽ Ä x%* @€%øŽ â $x))ÿŽ â x,,€   xŽ â $t*ÿÿ Ž ä $x€   ÿ Ž ä $tÿ @€ÿ Ž â x*)€€€€*ìŽ â $x,,ÿŽ â x€   xŽ â $tÿÿ Ž â x)€€€€ìŽ â $t-ÿ @€ÿ Ž â x**ÿþþþÿàÿ â $xÿŽ â xÿþþþÿàÿ â $tÿÿ Ž â x,€€€€ìŽ Ä x-,€€€€-ìŽ ä x**)øŽ äx) @€)øŽ â $tÿ @€ÿ Ž â xÿþþþÿàÿ ä x--ÿþþþÿàÿ ä x€€€€ìŽ Ä x++˜º  ÿ    ä x,øŽ äx-- @€,øŽ ä x€€€€ìŽ ä x,ÿþþþÿàÿ ä x+ÿÀŽ Ä xÿþþþÿàÿ ä x%%˜º  ÿ    ä x**˜º  ÿ    ä x,,øŽ äx @€øŽ ä x*ÿÀŽ Ä x))˜º  ÿ    ä x˜º  ÿ    ä x˜º  ÿ    ä x--˜º  ÿ    ä z ÿ `  pPð ä x ÿ      Ä z ÿ a   Sð ä x   ÿàÿ ä x++€€€€&ÞŽ ÊO $x++   
Ž â x%ÿÀŽ äx**€€€€"ÞŽ ä� x%%€€€€$ÞŽ ä x&+€€€€&´Ž Æ $x%   
Ž ä $x%*   
Ž â x)ÿÀŽ äx*)€€€€'ÞŽ ä x),˜º  ÿ    ä x€€€€$´Ž â $x**   
Ž â x$)ÿÀŽ â &r&&   ÿ   âxÿÀŽ Ä x€€€€ÞŽ ä x"%€€€€"´Ž ä x))€€€€(ÞŽ â $x   
Ž ä &r   &   â x'*€€€€'´Ž â $x))   $
Ž ä &r"      â x"ÿÀŽ Ä x€€€€ÞŽ ä x()€€€€(´Ž ä x€€€€´Ž â &r'      â x-ÿÀŽ â $x   "
Ž â x--€€€€ÞŽ â &r(      Æ x€€€€´Ž â $x--   
Ž ä &r      Æx--€€€€´Ž â &r      È &r-      Ê r   ÿ� È xÿ     Ê $r    Ž Ê r    ÿ� â 0rÿ    A   ä 0r"ÿ    A   â x ÿ      â G‰  
    € ê %v# `  Ž Ê r   ÿàñ Ê $x     Ê �y%    ¨ �y    è  �y    ( �y$    b %v `  Ž Ê r   ÿàñ Ê $x     Ê �y#    h	 �y&    b	 $x%   ÿ Ž âO xÿ   %  È xü  ÿÀŽ È z  ÿàñ â $x   ÿ Ž È� $vÿ  ÿ  â xü  ÿÀŽ â %v  ! Ž Æ �y%    ™ ¢  z  ÿàñ â $x   ÿ Ž âxÿ     â �y'    ™ ä $vÿ  ÿ  â xü  ÿÀŽ â %v  ! Ž â xÿ     ä �y(    ™ â z  ÿàñ â $x$   ÿ Ž Ä�y)    ™ d $vÿ  ÿ  â xü  ÿÀŽ â/ %v  ! Ž Æ z  ÿàñ â �y    ™ b xÿ   $  Æ�y    ™ b	 $vÿ  ÿ  ä %v  ! Ž Æ �y    ™ h �y    ™ b z `  ÿàñ È z a  ÿä  Ê �y     b x#ÿ   ÿÀŽ ä/ x!ÿ   #  È 	s         " x!!ÿ   ÿÀŽ ä xÿ   &  Ê 	s$ !        b xÿ   ÿÀŽ ä x#T  &    â $x€   ÿ Ž È	s         " x*€   ÿÀŽ ä xÿ     â $x$$€   ÿ Ž â/ x*ÿ   #xŽ ä xÿ   ÿÀŽ â $t*ÿÿ Ž â x!€   $xŽ â $xÿŽ â 	s!         " $t,ÿ @€ÿ Ž Æ x#€€€€*ìŽ äx*€€€€,ìŽ â $t$ÿÿ Ž â x##ÿþþþÿàÿ â $xÿŽ ä $x,€   ÿ Ž âx+*ÿþþþÿàÿ â $t*ÿ @€ÿ Ž â x#øŽ Ä x#€€€€$ìŽ ä x€   ,xŽ ä x+ @€øŽ â $x!!€   ÿ Ž â x+€€€€*ìŽ ä x*#ÿþþþÿàÿ â $x#ÿŽ â x!€   !xŽ È x#€€€€$ìŽ ä x$˜º  ÿ    â $x!ÿŽ â x,+ÿþþþÿàÿ ä x!$ÿÀŽ ä x*øŽ â$t*ÿÿ Ž â x, @€øŽ â $t,ÿ @€ÿ Ž â x˜º  ÿ    Ä x*€€€€*ìŽ äx,€€€€,ìŽ ä x**ÿþþþÿàÿ ä x,,ÿþþþÿàÿ ä x˜º  ÿ    ä x˜º  ÿ    Ä x**øŽ äx,, @€øŽ ä xÿÀŽ â $t+ÿ @€ÿ Ž â xÿþþþÿàÿ È x+#€€€€+ìŽ ä x#øŽ äx++ÿþþþÿàÿ ä x˜º  ÿ    ä x#+ @€#øŽ È x##˜º  ÿ    ä x&ÿ   &  ä x$$€€€€%ÞŽ ÊO $x$$   !
Ž â x!ÿÀŽ äx€€€€'ÞŽ ä� x$$€€€€%´Ž Æ $x   !
Ž â x€€€€(ÞŽ â &r$$   ÿ   â xÿÀŽ ä x€€€€'´Ž ä x€€€€)ÞŽ â$x   
Ž ä &r$   $   â x*˜º  ÿ    â $x   
Ž â x,˜º  ÿ    Ä x€€€€(´Ž ä x(ÿÀŽ äx)€€€€)´Ž ä x€€€€ÞŽ ä xÿÀŽ äx€€€€ÞŽ â $x   (
Ž Ä &r   $   ä $x   
Ž â xÿÀŽ äx€€€€ÞŽ â &r)      â x€€€€´Ž ä x#ÿÀŽ â $x   
Ž â x##€€€€ÞŽ â &r      â x€€€€´Ž Ä x€€€€´Ž â $x##   
Ž ä &r      Æ x##€€€€´Ž â &r      È &r##      Ê r##   ÿ� È xÿ     Ê $r&#   Ž Ê r   ÿ� â 0rÿ    A   Æ xÿ     ä  r"     A  ä Er        Ê #r
   
   â |    pbð â  r""     A  â Er          ä x   ÿàó ä x€   ÿàÿ â #r	"    	   ä $rÿÿ   Ž  È D	       À â Gy  ðéÿÿÿÿƒ ê Ay        € ê y       "   ". Ey        € â r  ÿ   pRð är  ÿ   pRò Ö G‰  `     € ê y      !   " z ÿ `  pPð È z ÿ a   Sð â $x     Ž Ê ˆs 	€ÿÿ H   î ˆ 
   H   ä Ay        € ê {          ì M        € ê y      !   b z ÿ `  pPð È z ÿ a   Sð â „y      H   X. „‰ €   H   ¨ „‰    H   " !r	 	       È/ !‚	       ÈO !‚       â G‰        € ê „y  €  H   ( „y €   H   h „y    H   ¨ „y    H   è „y €  H   "!r  
       â !r	       Æ/ !r         âO !r       Æ� !r
        ä‰      " z ÿ `  pPð ä r ÿ   pRò ä z ÿ a   Sð â !r       Ê ‰       $ !r         Ê ‰  €    $!r        Ê ‰ @    $ !r       Ê ‰      $ !r       â G‰        € ê ‰ 
      $ !r  
       Ê ‰       $ !r        Ê ‰ €    $ !r       Ê ‰ @    $ !r       Ê ‰      $ !r
       ä M        € ê !r       â G‰  `    € ê $t ÿ   ÿ Ž â !r
       È z   h  pð Ú G	  `    € ê $v ÿ h  ÿ Ž Ê x     pRð Ú G‰  �     € ê x     pRð Ú G	       € ê  x ;ª¸¿  A  Ì s          $ !t   €?    Ì s        $  r     A  È  r     A  â Gy  `    € ê  x '7=  A  â x*BL?  A  Æ #t   €?    È  r      A  â  x    ?  A  Ê s     $   $ !t  €?    È  r       A  È  r      A  â Gy  À    € ê $v ÿ h  ÿ Ž Ê x     pRð Ú G‰  à     € ê x     pRð Ú G	        € ê 	z  i    � ä 	z i    � Æ  x ;ª¸¿  A  â 	z i €  � Ê s        $ !t  €?    Ì s        $  r       A  È  r      A  â Gy  Ð     € ê  r     A  â Gy  °     € ê 	x   à@  � ä 	x  à@  � Æ  x #ÛÙ?  A  â 	x  àÀ  � Æ  x;ª¸¿  A  ä !t  €?    È s        $ !t  €?    Ì s        $  r       A  È  r      A  ä y       %   ( y      &   ( y      '   b $z  x   Ž ä $tÿ   ÿ Ž Ä $z  ~   Ž È/ %v  j   Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y       &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž ä $rÿÿ     Ž È % \   Ž Ê �	    ™ b ¹z  l      â Ey  0,    € â ™x?    â y	      "   b $rÿÿ   ÿ Ž Æ y      !   ¢ $ÿÿ   ÿ Ž ä $ÿÿ    Ž â y      %   & '
 m   Ž â y)      '   æ $‚ÿÿ   ÿ Ž â 
ÿ n    Ê $ÿÿ   
Ž â r 	ÿ   pRò Æ/ $
 o  Ž ä $x
	    Ž äO $‚ÿÿ    Ž ä $rÿÿ   ÿ Ž â xÿ   
  â 'Š s   Ž Æ r
   ÿ� â $	€   Ž Æ xÿ     È |    pbô ä Šÿ t    ä r 	ÿ   pRð â x      ÿ  â $x	   ÿ Ž Î G)  `*    € ê $r(ÿÿ   ÿ Ž â� xøÿÿÿÿÀŽ â $z w  ÿŽ ä 'z) y  ( Ž È $z) }  ÿŽ â zÿ z    â $z	 p  Ž ä $x

   
Ž â r   ÿàõ â $z |  Ž ä $x/
   ÿ Ž âx+
!   ÿŽ â $z, v  Ž âxÿ   
  â $vÿ X  ÿ Ž Ä %x,,b   ÿŽ È $vÿ Y  ÿ Ž ä $rÿÿ   ÿ Ž ä $x   ÿ Ž ä $rÿÿ   ÿ ä %x./   ,Ž È %x*+   ,Ž È r.   ÿàõ Ê $x/    Ê �y   • ¨ �y   • è  �y"   • ( �y   • b $vEÿ p  ÿ Ž È %xEb   Ž Ê �y   • ( �y   • ( �y   • ( �y   • b %xEb   Ž â r4*   ÿàõ È �yM   • b $x5+    Æ �yO   • h �yQ4    • h �yP   • h �y4   • h �y@   • d %xEb   Ž Ê �yG   • b $t7ÿ$   ÿ Ž Æ �yK   • h �yA   • b %x4Eb   4Ž Ê �y&4    • b $x   ÿ Ž ÊO xü  ÿÀŽ È z  ÿàõ Ê/ $vÿ  ÿ â xÿ     È �yL    ™ ¢ $x   ÿ Ž ä� $tÿ   ÿ Ž Æ xü  ÿÀŽ â %v   Ž Æ z   ÿàõ â $x"   ÿ Ž â�yN    ™ & $v!ÿ  ÿ â xü  ÿÀŽ ä/ xÿ   "  ä �yF     ™ " z  ÿàù Ä� r6
   àµ Æ $vÿ  ÿ ä %v   Ž â x$ÿ     ä �yJ    ™ " r    ÿd~ â/ %v66 Z  7 Ž ä �yI    ™ $ %v$   Ž È $x! $   ÿŽ â �yH    ™ & $x77   !Ž Ê �y"6   ™ " $x   ÿ Ž Æ�y#6   ™ $ xü  ÿÀŽ ä �y!6   ™ $ z  ÿàõ ä� �y   • æ  $vÿ  ÿ â x$ÿ     â �y 6   ™ è
 �y    ™ â  %v$$   Ž Æ �y6   ™ â* $x   ÿ Ž Æ �y$$    ™ ä xü  ÿÀŽ ä �y4   • ä z>  ÿàõ ä �y6   ™ æ
 $v?ÿ  ÿ â xÿ     â $x   ÿ Ž â x0ÿ     â �y6   ™ è
 �y>>    ™ â  %v   Ž â xü  ÿÀŽ Ê �y    ™ â z2  ÿàõ â %v00   Ž â xBÿ     â �y?6    ™ ä $v3ÿ  ÿ ä $x   ÿ Ž â �y=0    ™ è  �y<2    ™ â  xü  ÿÀŽ â %x4Eb   4Ž Æ z8  ÿàõ ä �y4    • æ" $v9ÿ  ÿ â x2ÿ     â $x   ÿ Ž Æ �y'8    ™ ä  xü  ÿÀŽ â %v22   Ž Æ z0  ÿàõ ä �y%2    ™ æ  $v1ÿ  ÿ Ê �y0    ™ â  %vBB   Ž â r:,   ÿàõ Æ $xM   ÿ Ž â �yB    ™ â $x;-    Æ x2ü  ÿÀŽ ä �y4   • â  %x8Eb   :Ž âz22  ÿàõ ä �y:    • æ  %x0Eb   8Ž â�y8    • æ  $v3ÿ  ÿ â xRÿ   Q  â �yC0    • è" �yB6    • â
 %x8Eb   4Ž È $x5O   ÿ Ž â x4ÿ   M  â %x:Eb   0Ž â xRRÿ   ÿÀŽ â �yE2    ™ â  x55ü  ÿÀŽ ä x6ÿ   O  â�yD8    • â
 z05  ÿàõ â/ %v44   Ž â 	sT R        b �yM8   • ä
 $v1ÿ  ÿ Ä �yO4    ™ â xSQÿ   ÿÀŽ Æ �y::    • è �y90    ™ â
%v26   Ž â 	sS S        & $x7T€   ÿ Ž â/ �y82    ™ â x6ÿ     Æ x7R€   7xŽ â $xRP   ÿ Ž â x0QT      â$tTÿÿ Ž â x66ÿ   ÿÀŽ â $x77ÿŽ â xRRü  ÿÀŽ ä x4ÿ   0  â 	s2 6        b. z0R  ÿàõ â $xSS€   ÿ Ž â x37€€€€TìŽ Ä x44ÿ   ÿÀŽ â $v1ÿ  ÿ â xT3ÿþþþÿàÿ ä xRS€   ÿÀŽ â 	s3 4        " �y;0    ™ ä
 xQRÿ   QxŽ â $tRÿ @€ÿ Ž ä $xU@   ÿ Ž Ä $xS2€   ÿ Ž â/ xR7€€€€RìŽ ä xUUü  ÿÀŽ ä xRRÿþþþÿàÿ ä xTT7øŽ ä x66€   SxŽ â $xS3€   ÿ Ž â z2U  ÿàõ â $tWÿÿ Ž â x7R @€7øŽ â $xRQÿŽ â x5T˜º  ÿ    Ä xTÿ   P  â $v3ÿ  ÿ â xU4€   SxŽ â $tSÿ @€ÿ Ž â x4R€€€€WìŽ â %v0T   Ž âxW7˜º  ÿ    â �yQ2    ™ d $tXÿÿ Ž ä $x7UÿŽ â xTR€€€€SìŽ â $x66ÿŽ â xS4ÿþþþÿàÿ Ä xUTÿþþþÿàÿ â $t3ÿÿ Ž â x4SRøŽ ä xP5€€€€LÞŽ äOx55ÿÀŽ Ê $x5P   5
Ž ä �yP0    ™ ¦  xT5€€€€L´Ž â $t5ÿ @€ÿ Ž â x17€€€€XìŽ È x21ÿþþþÿàÿ ä x16€€€€3ìŽ ä x227øŽ ä x0U @€RøŽ â $tRÿ @€ÿ Ž â xL6€€€€5ìŽ ä x51ÿþþþÿàÿ Ä xVWÿÀŽ äx12˜º  ÿ    ä xWW€€€€NÞŽ äx37€€€€RìŽ ä x21ÿÀŽ â $xUW   V
Ž â x11€€€€FÞŽ Ä xV3ÿþþþÿàÿ ä x34˜º  ÿ    â $x11   2
Ž â xR56øŽ â $x5G   ÿ Ž â x20˜º  ÿ    ä x03ÿÀŽ äx33€€€€JÞŽ Ä xVV @€7øŽ ä x45ü  ÿÀŽ ä x52ÿÀŽ â$x33   0
Ž â x22€€€€IÞŽ ä xSLÿþþþÿàÿ ä x7V˜º  ÿ    â $x22   5
Ž â xLS @€6øŽ Ä x67ÿÀŽ ä xF1€€€€F´Ž â $x1K   ÿ Ž â x33€€€€J´Ž ä x77€€€€HÞŽ ä z44  ÿàõ â &r33"   ÿ   â x22€€€€I´Ž â $x77   6
Ž â x0ÿ   G  â $v5ÿ  ÿ â xG1ü  ÿÀŽ â &r32#   3   â xS7€€€€H´Ž â %v00   Ž â z2G  ÿàõ â �yH4    ™ " x6ÿ   K  â $x7A   ÿ Ž Ä &rTT!   3   â �yG0    ™ ¢ $v3ÿ  ÿ â x77ü  ÿÀŽ â %v46   Ž Æ �yJ2    ™ ¢  z67  ÿà÷ â $xX   ÿ Ž ä� �yI4    ™ ¢ xUU€€€€N´Ž â $v7ÿ  ÿŽ â xN&ÿ   ÿÀŽ ä xXXü  ÿÀŽ ä xVÿ   A  â �yK6    ™ ¢ z2X  ÿàõ â 	sN N        " xWR˜º  ÿ    â %v0V   Ž â/ xRÿ   @  Ä x@WÿÀŽ â$v3ÿ  ÿ â xAW€€€€ÞŽ â %v4R   Ž â� �y00    ™ â x6ÿ   &  ä $xAA   @
Ž â �y22    ™ ¢ x@ÿ     ä x6ÿ   ÿÀŽ â �y44    ™ ¢ $xNN€   ÿ Ž Ä %v6@   Ž Æ xWN€   ÿÀŽ Æ �y66    ™ ¢  xWÿ   &xŽ â &rTU    T   â xUL˜º  ÿ    â 	s5         b. $tLÿÿ Ž ä $xÿŽ â x@UÿÀŽ äxUU€€€€$ÞŽ Ä x1€€€€LìŽ â $tLÿ @€ÿ Ž ä $xUU   @
Ž â x@1ÿþþþÿàÿ â &rFF   T   â x&&T      ä x1€€€€LìŽ â $x55€   ÿ Ž â/ x@@øŽ â &rSS   F   â x&ÿ   &  Ä xF1ÿþþþÿàÿ ä x1@˜º  ÿ    ä x5€   5xŽ ä x&&ÿ   ÿÀŽ ä x1€€€€>ÞŽ äx11ÿÀŽ â 	s7 &        " xFF @€øŽ â $x5ÿŽ Ä $t@ÿÿ Ž â x3F˜º  ÿ    â $x1   1
Ž â xA€€€€´Ž ä x€€€€@ìŽ ä xAÿ     ä x@3ÿÀŽ äx53€€€€ÞŽ Ä x31€€€€>´Ž ä x>ÿþþþÿàÿ ä xAÿ   ÿÀŽ ä x>>øŽ ä 	s1         b $x55   @
Ž ä $xA7€   ÿ Ž â x7>˜º  ÿ    â $t@ÿ @€ÿ Ž Æ xA&€   AxŽ ä x@€€€€@ìŽ ä x&7ÿÀŽ äx77€€€€<ÞŽ â &r33"   ÿ   â x5€€€€´Ž ä x@@ÿþþþÿàÿ â $x77   &
Ž Ä $xAAÿŽ ä $t>ÿ @€ÿ Ž â x@@ @€øŽ â &r#   3   ä $t&ÿÿ Ž ä $x31€   ÿ Ž â/ x1A€€€€>ìŽ äx&A€€€€&ìŽ Ä x@@˜º  ÿ    ä x3€   3xŽ ä x7€€€€<´Ž ä x1ÿþþþÿàÿ ä x&&ÿþþþÿàÿ ä x1@ÿÀŽ â$x33ÿŽ â x@@€€€€=ÞŽ â $t<ÿ @€ÿ Ž â x&&AøŽ â &r!      Ä $tÿÿ Ž â x5ÿ   ÿÀŽ â $x@@   1
Ž â x13€€€€<ìŽ äx&&˜º  ÿ    ä x3€€€€ìŽ ä x<1ÿþþþÿàÿ â 	s5 5        " x1&ÿÀŽ Ä xÿþþþÿàÿ ä x&&€€€€'ÞŽ ä x @€AøŽ ä x@@€€€€=´Ž â $x&&   1
Ž â x3øŽ äx<< @€3øŽ Ä x3ÿ     â &r@@       â x˜º  ÿ    ä x3ÿ   ÿÀŽ ä x&&€€€€'´Ž ä x'ÿÀŽ â	s1         b x€€€€%ÞŽ Ä x˜º  ÿ    â $x55€   ÿ Ž ä $x   '
Ž â x'ÿÀŽ äx€€€€ÞŽ ä x>5€   ÿÀŽ â &r&&   @   â x%€€€€%´Ž â $x   '
Ž â x>>ÿ   xŽ Ä x<<˜º  ÿ    â &r%   &   â x€€€€´Ž â $x&1€   ÿ Ž â/ x'<ÿÀŽ â $x>>ÿŽ â x<<€€€€ÞŽ â $tÿÿ Ž â x&€   &xŽ ä xT      â $x<<   '
Ž â x>€€€€ìŽ â $t%ÿ @€ÿ Ž â xÿ     Ä xÿþþþÿàÿ â $x&ÿŽ â x<€€€€´Ž â $t<ÿÿ Ž â x1ÿ   ÿÀŽ ä x>øŽ â $t&ÿ @€ÿ Ž â x>€€€€%ìŽ â &r      â x˜º  ÿ    Ä x%€€€€<ìŽ ä xÿþþþÿàÿ ä x'€€€€&ìŽ â 	s3 1        " x€€€€EÞŽ äx&%ÿþþþÿàÿ ä xÿÀŽ Ä x @€>øŽ ä x%&øŽ â $x   
Ž â x&˜º  ÿ    È x&ÿÀŽ äx&&€€€€OÞŽ ä xE€€€€E´Ž ä x%˜º  ÿ    â $x&&   
Ž â x<'ÿþþþÿàÿ ä xÿ     â $x>3€   ÿ Ž â xÿÀŽ Ä x€€€€9ÞŽ ä x<< @€øŽ ä x%ÿ   ÿÀŽ â $x   
Ž â x>1€   >xŽ ä 	s' %        " x<˜º  ÿ    â $tÿÿ Ž Ä $x>>ÿŽ â x<ÿÀŽ äx€€€€8ÞŽ ä x>€€€€ìŽ â &rEE"   ÿ   â x&&€€€€O´Ž â $x   <
Ž â xÿþþþÿàÿ ä x€€€€9´Ž â &r&&#   E   â x€€€€8´Ž â $x8'€   ÿ Ž â x>øŽ â &r!   &   Æ x&˜º  ÿ    ä x8%€   8xŽ â &r       â x&ÿÀŽ äx&&€€€€;ÞŽ â $x8ÿŽ ä $t8ÿÿ Ž ä $tÿ @€ÿ Ž Ä $t<ÿ @€ÿ Ž â x'Dÿ   ÿÀŽ â $x&&   
Ž â x€€€€8ìŽ äx>€€€€ìŽ ä x%€€€€<ìŽ ä x<ÿþþþÿàÿ â 	s' '        " x1ÿ   D  Ä xÿþþþÿàÿ ä x<<øŽ ä x11ÿ   ÿÀŽ ä x @€>øŽ ä x>%ÿþþþÿàÿ ä x<<˜º  ÿ    â 	s% 1        b x&&€€€€;´Ž Ä x<ÿÀŽ äx<<€€€€QÞŽ â&r&      â x&ÿ   M  â $x''€   ÿ Ž â x˜º  ÿ    â $x<<   
Ž â xDT  M    ä x5&ÿ   ÿÀŽ Ä x8ÿÀŽ äx€€€€PÞŽ âO 	s7 5        " x''€   ÿÀŽ ä xÿ     â $x&%€   ÿ Ž â/ x''ÿ   DxŽ â $x   8
Ž â x3ÿ   ÿÀŽ Ä x&1€   &xŽ â $x''ÿŽ â 	s1 3        b x€€€€P´Ž â $t8ÿÿ Ž â xQ<€€€€Q´Ž â $t<ÿ @€ÿ Ž â x> @€øŽ â &r      â x'€€€€8ìŽ â$t>ÿÿ Ž â x'€€€€<ìŽ â $x%&ÿŽ Ä $t@ÿ @€ÿ Ž â x&ÿþþþÿàÿ â $x<7€   ÿ Ž â x8ÿþþþÿàÿ ä x%€€€€>ìŽ äx%€€€€@ìŽ ä x55€   <xŽ â $x@1€   ÿ Ž â/ x<ÿþþþÿàÿ Ä x>ÿþþþÿàÿ â $tDÿÿ Ž â x<<%øŽ â$x55ÿŽ â x%> @€%øŽ â $t>ÿ @€ÿ Ž â x@3€   @xŽ ä x5€€€€DìŽ â$t1ÿÿ Ž â x&&'øŽ â$t3ÿ @€ÿ Ž â x5€€€€>ìŽ â $x@@ÿŽ â x88 @€'øŽ Ä x>ÿþþþÿàÿ ä x&˜º  ÿ    ä xDÿþþþÿàÿ ä x1@€€€€1ìŽ äx3@€€€€3ìŽ ä x88˜º  ÿ    Ä x'ÿÀŽ äx&€€€€HÞŽ äx>5øŽ äxD @€5øŽ ä x51ÿþþþÿàÿ ä x73ÿþþþÿàÿ Ä x3<˜º  ÿ    ä x18ÿÀŽ â $x'&   '
Ž â x88€€€€GÞŽ ä x&3ÿÀŽ ä x<%˜º  ÿ    â $x88   1
Ž â x33€€€€JÞŽ Ä x''€€€€H´Ž ä x%<ÿÀŽ â$x33   &
Ž â x55@øŽ ä x<<€€€€IÞŽ â &r''"   ÿ   â x88€€€€G´Ž ä x"5˜º  ÿ    â $x<<   %
Ž â x33€€€€J´Ž â &r88#   '   â x#"ÿÀŽ Ä x77 @€@øŽ ä x""€€€€KÞŽ â &r33!   8   â xI<€€€€I´Ž ä x˜º  ÿ    â $x&"   #
Ž â x77˜º  ÿ    â &r"I    3   â x ÿÀŽ Ä x87ÿÀŽ ä x€€€€2ÞŽ ä x77€€€€0ÞŽ ä� x˜º  ÿ    â $x    
Ž ä $x77   8
Ž â x ÿÀŽ Ä x€€€€4ÞŽ ä x˜º  ÿ    ä x&&€€€€K´Ž â $x    
Ž â x77€€€€0´Ž ä x0ÿÀŽ â &r   S   â x$U€€€€$´Ž Ä x€€€€6ÞŽ â &r"&   "   â x4€€€€4´Ž â &r$?      âx€€€€2´Ž â $x   0
Ž ä &r"7   "   ä &r?      â x6€€€€6´Ž â &rQ      â r   ÿ� â &r   "   â xÿ     â &r4?      âxÿ     â &r66?      â r   ÿ� Ä r   ÿ� â $r   Ž â r66   ÿ� ä xÿ     ä xÿ     ä xÿ     ä xÿ     Ä xMÿ   M  ä xÿ     â $r   Ž â x   ÿàÿ â $r   Ž ä $rM6   Ž â |    pbô ä r   ÿ� Ä r   ÿ� ä r   ÿ� ä r   ÿ� â 0rÿ    A   â xÿ     â 0rBÿB    A   â xÿ     â 0rÿ    A   â xÿ     â 0rCÿC    A   â xÿ     â 0rÿ:    A   â x   ÿà÷ â  rB     A  â Er        â  rB     A  âEr        â  rBC     A  âEr        â  rB     A  â Er        â #r      â #r      â #r      â x€   ÿàÿ â #r      â $rÿÿ   Ž â D)       À â Gy   ×ÿÿÿÿƒ ê Ay        € ê� ˆ  þÿ H   è ˆ €þÿ H   è ˆ  ÿÿ H   è ˆ €ÿÿ H   è {          ì M	        € ê „y     H   " z 	 r  p`ô Æ „y €   H   b r ÿ   pVt Æ „y    H   ¨ „y    H   è„y €  H   ( „y    H   ( „y
    H   ( „y €  H   ( „y €  H   ( „y    H   ( „y €  H   " !r       Æ „y €  H   " !r       â/ !r       âO !r       ä� !r       â!r       â !r
       â !r       È ‰      b !r       â !r       È ‰      ¢ !r       Æ ‰      â !r       Ê ‰      " !r       â/ !r       âO !r       È� ‰      h ‰      ¢ !r       Æ ‰      ( ‰      â !r       â/ !r       ÈO ‰
 €    b !r       Æ ‰ €    " !r       Æ� ‰ €    ¨ ‰ €    â !r

       â/ !r       È ‰
 @    " !r       ÆO ‰ @    b !r       Æ� ‰ @    ¨ ‰ @    â !r
       â x
	   ÿàÿ â !r       Æ/ ‰      " $z  x  	Ž â x 	   ÿàÿ â!r       âO ‰      b x	   ÿàÿ â $z) ~  Ž â !r       â� ‰      ¢ z   r  p`ö â $t ÿ   ÿ Ž â z  r  p`ð â ‰      â z 
 r  p`ò Ä x    pVö äx    pVp  äx    pVò  â %v j    Ž â !r       â !r       Ì/ †¹    â !r       ÆO †©     â !r       Æ� †‰    â M        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      &   " z ÿ \  pPò â ¹z  F   
   Æ z ÿ ]  Sò Ú $ÿ   ÿ Ž È % \   Ž Ê �+    ™ ¢  $rÿÿ   ÿ Ž â ¹z  l      â Ey  €    € â y      %   b ' m   Ž â™x?    â x       ÿ  â y)      "   ä ÿ n    â 'š s   Ž â y       !   & $	ÿÿ   
Ž â y      '   & $	 o  Ž â ’        Ê $z w  ÿŽ ä $rÿÿ   ÿ Ž ä $x   ÿ Ž â/ r )ÿ   pRð Æ� $x    Ž âx     pDp  â $x))     Ž ä $z }  ÿŽ âz  r  pdp  â 'z y   Ž Æ r   ÿàõ ä š+ÿ t    ä xÿ   )  ä r+     € äO r)   ÿ� â G	  °    € ê z ÿ `  pPò â x       ÿ  â z ÿ ^  pPð Ä z ÿ b  p ö ä z ÿ a  Sò ä z ÿ _   Sð ä z ÿ c  0'ò Ö $
 ~  ÿŽ âÿ     â$š ~  ÿŽ ä $
	 x  ÿŽ â    àë â$š
 x  ÿŽ â ’    àÇ ä ˜ÿ     ä ÿ   ÿÄþ Ä 	   ÿàý ä ’
   ÿàû ä ’ÿ   ÿ„þ â $ÿÿ    â 
 ^  ÿ† ä ’
ÿ   ÿäÿ ä š b  ÿˆ Ä 
 _  � ä š c  
 Æ �	!     h �™      d Ay        € ê x*ÿ     â Ey  ÀT    € â x"      ÿ  â $r@ÿÿ   ÿ Ž â | *   pbð â $rCÿÿ   ÿ Ž â x$      ÿ  â x&      ÿ  â $r(ÿÿ   ÿ Ò G	  0T    € ê xøÿÿÿÿÀŽ â $v/ÿ `  ÿ Ž â zÿ z    â $z p  *Ž âx-   ÿàÿ â $x))   
Ž â x   ÿàÿ â $z |  *Ž âx        â $z |  Ž â x)   ÿ   â $z+ v  Ž â x   ÿàÿ â $z++ v  Ž â x)!   ÿŽ â %x   ÿŽ â x"      ÿ  â x$      ÿ  â x&      ÿ  â $z, p  Ž âx7ÿ   )  â $z-- p  Ž Ä $z. p  Ž ä %x+b   Ž È %v  € Ž â x2   ÿàñ Æ %x,b   Ž È %x-b   Ž È %x
,b   Ž È %x+b   Ž È %x-b   Ž È $v1ÿ a  ÿ Ž ä $v3ÿ X  ÿ Ž ä $v5ÿ Y  ÿ Ž ä $r@ÿÿ   ÿ Ž ä $rCÿÿ   ÿ Ž ä $x0*   ÿ Ž Ä %x.b   Ž È %x.b   Ž È $r4ÿÿ     ä r3   ÿàñ È r5   ÿä  Ê �y:   • ¨ �y8   • è �y9   • ( �y;   • â r32   ÿàñ â $veÿ p  ÿ Ž È $x5   4  È %xeb   Ž Ê �yúÿÿ • ( �yþÿÿ • ( �yK    • ( �yJüÿÿ • " r3   ÿàñ Ê $x   5  Ê �y`   • " %xeb   Ž Æ �yE   • ( �ycúÿÿ • ( �yaþÿÿ • ( �y[üÿÿ • ( �yY    • " $t6ÿ   ÿ Ž Ä %xeb   Ž Ê �yOüÿÿ • " $t?ÿ$   ÿ Ž Æ �yIúÿÿ • ( �yFþÿÿ • " %xeb   Ž Æ/ �yH    • ( �yB   • " $x:   ÿ Ž ÊO xü  ÿÀŽ È z  ÿàñ È zÿ  ÿä  Ê �y]    ™ ¢ $x8   ÿ Ž â� xÿ   8  È xü  ÿÀŽ â $x9   ÿ Ž ä%v  6 Ž â zR  ÿàñ ä xü  ÿÀŽ ä �yT    ™ $ z  ÿàõ â $vSÿ  ÿ  â xÿ   :  Ä r8)   0à‘ ä zÿ  ÿä ä x=ÿ   9  â %v  6 Ž â r:7(   ÿ$~  â �yQ    ™ $ %v<=  6 Ž ä �yU    ™ $ $x9:$   ÿŽ Ä �yN<    ™ ( �yRR    ™ " %v8 Z  ? Ž È $x   9Ž ä $x;   ÿ Ž Æ� �y8   ™ è  �y9   ™ â  xü  ÿÀŽ Æ �y:   ™ â  z  ÿàñ ä xÿ   ;  â �y<   ™ ä  $vÿ  ÿ  ä �y;   ™ è  �yD    ™ â %v  6 Ž Æ �y=   ™ è  �yA    ™ è  �y>   ™ â  xG   ÿ   Æ �y?    ™ â  xGü  ÿÀŽ Æ/ �yG   • â z  ÿàñ ä xÿ     Æ $vÿ  ÿ  â x_ÿ   K  â %v  6 Ž â xb`ÿ   ÿÀŽ â �y    • è  �yL    ™ è  �yM    ™ â  $x   ÿ Ž Ä $xK   ÿ Ž Æ xü  ÿÀŽ ä xPü  ÿÀŽ ä z  ÿàñ ä xÿ     â %xeb   Ž â/ xJ   ÿ   Æ $vÿ  ÿ  â zP  ÿàñ â %v  6 Ž â xü  ÿÀŽ â �yV   • ä  $vÿ  ÿ  â z  ÿàñ â �yP    ™ è �yS    ™ â  $vÿ  ÿ  â xKÿ   J  Ä �yW    ™ è �yX    ™ â %vK  6 Ž Ê �y\    ™ â  %v_  6 Ž â/ 	sb b        d �y_   • è  �yZ    ™ ä  $rÿÿ   3 Ž Ä $rÿÿ   5 Ž È %xJ+b   Ž â xc   ÿ   Ê %xeb   JŽ âx^ÿ   `  â �yJJ    • â  xü  ÿÀŽ â $xb€   ÿ Ž ä/ %xeb   Ž âxf^ÿ   ÿÀŽ ä z  ÿàñ â %xeb   Ž â	sg f        b xiÿ   c  â �yb    • â %xeb   Ž â xe€   ÿÀŽ Ä �y^    • â $vÿ  ÿ  â xeeÿ   `xŽ ä �yK   • è �yc    ™ â  $xa   ÿ Ž ä/ %vi  6 Ž â xi     â �yd   • ä  $xeÿŽ Ä $xgg€   ÿ Ž â �ye    ™ è �y    • â xü  ÿÀŽ ä x€€€€iìŽ ä xgf€   gxŽ â $tiÿ @€ÿ Ž â xÿþþþÿàÿ â $tÿÿ Ž Ä $xggÿŽ â xøŽ â $tfÿ @€ÿ Ž â x€€€€iìŽ ä/ xg€€€€ìŽ äxg€€€€fìŽ ä x˜º  ÿ    ä xÿþþþÿàÿ Ä z  ÿàñ ä xfÿþþþÿàÿ ä xhÿþþþÿàÿ ä x @€øŽ â $x[   ÿ Ž â xfgøŽ â $vÿ  ÿ  â x``T  E    Ä xgh @€gøŽ ä xhÿ   E  â �yf    ™ â  xÿ   a  ä x`ÿ   `  ä xaü  ÿÀŽ ä xhhÿ   ÿÀŽ Ä xm`ÿ   ÿÀŽ ä za  ÿàñ â 	so h        " xiÿ   [  Æ $vÿ  ÿ  â xk˜º  ÿ    Æ 	sn m        b �y[    ™ â  x€€€€]ÞŽ äOxÿÀŽ È rj  €ÿàÿ â $xY   ÿ Ž ä %v  6 Ž Æ xü  ÿÀŽ ä xaj€€€€]´Ž ä z  ÿàó â �y]    ™ â xj˜º  ÿ    â %vi  6 Ž È $vÿ  ÿŽ  ä �y    ™ è �y`    ™ â $xoo€   ÿ Ž ä $xnn€   ÿ Ž Æ/ xoh€   oxŽ ä xÿ   Y  äO xnm€   nxŽ â $tÿÿ Ž ä %v  6 Ž â x     Æ $xooÿŽ â �y    ™ ¢  $xnnÿŽ Æ xo€€€€ìŽ â $tYÿ @€ÿ Ž â xn€€€€ìŽ äxgg˜º  ÿ    ä xÿþþþÿàÿ ä xÿþþþÿàÿ ä xn€€€€YìŽ Ä xYoøŽ ä xgÿÀŽ ä xnøŽ ä xgg€€€€TÞŽ äxÿþþþÿàÿ ä x˜º  ÿ    â $xgg   
Ž â xn @€nøŽ â $xO   ÿ Ž â xÿÀŽ Ä x€€€€QÞŽ ä xn˜º  ÿ    ä xü  ÿÀŽ ä xijÿÀŽ äxhg€€€€T´Ž ä xjj€€€€UÞŽ Ä r  €ÿàÿ â $xI   ÿ Ž â xTÿÀŽ äx€€€€NÞŽ ä z  ÿàñ ä rji  €ÿàÿ â $x   T
Ž â xlkÿÀŽ Ä xg€€€€Q´Ž ä xkk€€€€RÞŽ ä xü  ÿÀŽ â $vÿ  ÿ  â x€€€€U´Ž â &raa8   ÿ   â� xT€€€€N´Ž â $xkk   l
Ž â �yN    ™ â  &ra9   a   â z  ÿàñ â $xQF   ÿ Ž â xUk€€€€R´Ž Ä xÿ   O  ä zÿ  ÿä  â $tRÿ @€ÿ Ž ä &rUU:   a   â xaQü  ÿÀŽ â %v  6 Ž â �yQ    ™ " xIÿ   I  ä xYY˜º  ÿ    â �yO    ™ ¢ xRo€€€€RìŽ Ä za  ÿàñ ä xjYÿÀŽ äxRRÿþþþÿàÿ ä xYY€€€€DÞŽ â %vI  6 Ž â/ xRR @€oøŽ Æ $vÿ  ÿ  ä $xYY   j
Ž â �y    ™ ¢  xRR˜º  ÿ    Æ �yI    ™ ¢ xYY€€€€D´Ž â $xDH   ÿ Ž â xRÿÀŽ äxRR€€€€AÞŽ ä xaBÿ   ÿÀŽ ä xiÿ   F  ä xDDü  ÿÀŽ â &rUh;   U   Ä $xRR   
Ž â 	sa a        " %vi  6 Ž â zD  ÿàñ Æ/ &rUg<   U   â �y    ™ ¢ $vÿ  ÿ  ä &rUT=   U   Æ �y    ™ ¢ &rUY>   U   â xYÿ   H  ä xaa   ÿ   Æ %vY  6 Ž â xaa€   ÿÀŽ Ê �y    ™ ¢  xaaÿ   BxŽ â$tDÿÿ Ž â xRR€€€€A´Ž ä xAÿ   B  â $xaaÿŽ Æ xAAÿ   ÿÀŽ ä xDa€€€€DìŽ ä 	sF A        $ xDDÿþþþÿàÿ â $thÿ @€ÿ Ž â xÿ   G  ä/xBBT  G    Ä xDDaøŽ ä xÿ   ÿÀŽ ä xBÿ   B  ä xD˜º  ÿ    ä xDa€€€€hìŽ â &rRR?   U   â xHBÿ   ÿÀŽ Ä xDDÿþþþÿàÿ â $xFF€   ÿ Ž â 	sT H        " rRR   ÿ� ä xDD @€aøŽ ä xEÿ   E  ä xÿ     â 	sU         b xFA€   FxŽ Ä xDD˜º  ÿ    â $rER   Ž â x€€€€LÞŽ äxBÿÀŽ â $xFFÿŽ â xDÿÀŽ â$tEÿ @€ÿ Ž â xDD€€€€MÞŽ ä rB  €ÿàÿ Ä xBF€€€€EìŽ â $xDD   
Ž â xA€€€€L´Ž â $xET€   ÿ Ž â xBÿþþþÿàÿ â $tÿÿ Ž â xBU   ÿ   â/ &rAA8   ÿ   â xDD€€€€M´Ž ä xEH€   ExŽ Ä xF€€€€ìŽ ä xB€   BxŽ â &rD9   A   ä $xEEÿŽ ä $tHÿÿ Ž ä $tDÿ @€ÿ Ž â xÿþþþÿàÿ Ä xAE€€€€HìŽ äxDE€€€€DìŽ â $tMÿÿ Ž â xFøŽ â$tUÿ @€ÿ Ž â x @€FøŽ â $xFBÿŽ â xBAÿþþþÿàÿ ä xDDÿþþþÿàÿ Ä xAF€€€€MìŽ äxHF€€€€UìŽ ä xBBEøŽ äxDD @€EøŽ ä xMAÿþþþÿàÿ ä xEHÿþþþÿàÿ Ä xAB˜º  ÿ    ä xDD˜º  ÿ    ä xLVÿ   ÿÀŽ ä xHE @€FøŽ ä xBA€€€€PÞŽ äxEDÿÀŽ â 	sL L        " xAAÿÀŽ Ä xDD€€€€SÞŽ ä x˜º  ÿ    ä rABA  €ÿàÿ â $xBD   E
Ž â xMMFøŽ ä xDÿÀŽ äx€€€€XÞŽ Ä x˜º  ÿ    ä xFM˜º  ÿ    â $x   D
Ž â xDÿÀŽ äxEFÿÀŽ äx€€€€\ÞŽ ä xFF€€€€WÞŽ â $xLL€   ÿ Ž â x€€€€X´Ž â $x   D
Ž â xDÿ   V  â $xFF   E
Ž â xEH˜º  ÿ    â &r:      â xML€   ÿÀŽ Ä xDÿ   ÿÀŽ ä xHEÿÀŽ äxEE€€€€ZÞŽ ä x\€€€€\´Ž ä xMMÿ   VxŽ â 	sD         " xAA€€€€P´Ž â &r\;      â rEEH  €ÿàÿ â $xMMÿŽ â xVVT  _    â $tHÿÿ Ž â xBB€€€€S´Ž â &rA<      â xVÿ   V  Ä xM€€€€HìŽ â &rB=      â xFF€€€€W´Ž ä xZE€€€€Z´Ž ä xEÿ   _  â $tLÿ @€ÿ Ž â xVVÿ   ÿÀŽ ä xBÿþþþÿàÿ â &rF>      â xHEÿ   ÿÀŽ â 	sA V        b xBBMøŽ â $xFD€   ÿ Ž â xM€€€€LìŽ Ä xBB˜º  ÿ    Æ 	sE H        " xF€   FxŽ ä xDÿþþþÿàÿ ä xB€€€€cÞŽ â $xFFÿŽ â xBBÿÀŽ â $tÿÿ Ž â xDD @€MøŽ â $tMÿ @€ÿ Ž Ä $xB   B
Ž â xF€€€€ìŽ äxAA   ÿ   ä/ xF€€€€MìŽ â $xME€   ÿ Ž â xÿþþþÿàÿ ä xVV€   AxŽ ä xDD˜º  ÿ    Ä xEÿþþþÿàÿ ä xAFøŽ â$xVVÿŽ â xDÿÀŽ äxE @€FøŽ â $tEÿÿ Ž â xDD€€€€eÞŽ ä xMH€   MxŽ Ä xcB€€€€c´Ž ä rDD  €ÿàÿ â $tSÿ @€ÿ Ž â xBV€€€€EìŽ â $xFMÿŽ â xDD€€€€e´Ž â &rc8   ÿ   â xEBÿþþþÿàÿ â $tMÿÿ Ž â xBV€€€€SìŽ â &rD9      â xEEVøŽ Ä xDF€€€€MìŽ ä xMBÿþþþÿàÿ ä xBF€€€€UìŽ ä xEE˜º  ÿ    ä xUBÿþþþÿàÿ ä xBE€€€€fÞŽ Ä xMM @€VøŽ ä xEEÿÀŽ ä xSDÿþþþÿàÿ ä xDM˜º  ÿ    â $xEB   E
Ž â xSSFøŽ ä xBA˜º  ÿ    Ä xMDÿÀŽ äxDD€€€€]ÞŽ ä xSS˜º  ÿ    ä xABÿÀŽ äx˜º  ÿ    ä xBB€€€€[ÞŽ Ä xUU @€FøŽ ä rFDM  €ÿàÿ ä xLÿ   K  ä xHSÿÀŽ ä xDÿÀŽ â $xBB   A
Ž â xSS€€€€`ÞŽ Ä x€€€€ÞŽ ä xLLÿ   ÿÀŽ â $xSS   H
Ž â xHKÿ   ÿÀŽ â $xA   D
Ž â xBB€€€€[´Ž â 	sD L        & xA€€€€´Ž â &rB:      â xAKT  d    Ä xBU˜º  ÿ    â 	sH H        b xMÿ   A  ä xABÿÀŽ äxBB€€€€ÞŽ âO &r;      â xEE€€€€f´Ž ä xMMÿ   ÿÀŽ Ä rBA  €ÿàÿ â $xAD€   ÿ Ž ä &rE<      ä 	sE M        " xAL€   AxŽ â $xHH€   ÿ Ž â/ xS€€€€`´Ž â $tSÿÿ Ž ä $xBAÿŽ â xHH€   ÿÀŽ Ä x€€€€´Ž ä xÿ   d  ä xHHÿ   KxŽ ä xDB€€€€SìŽ â $tUÿ @€ÿ Ž â xFF€€€€]´Ž â $xLE€   ÿ Ž â xPÿ   ÿÀŽ â $tKÿÿ Ž â xEDÿþþþÿàÿ â $xHHÿŽ â 	sR P        " &rF=      â xFB€€€€UìŽ Ä xEEBøŽ äxH€€€€KìŽ ä xKFÿþþþÿàÿ ä xEE˜º  ÿ    ä xÿþþþÿàÿ ä xKK @€BøŽ Ä xLM€   LxŽ ä xBEÿÀŽ äxEE€€€€NÞŽ â� $xLLÿŽ â xA  @€    ä xHøŽ â $xEE   B
Ž â xAH€€€€AìŽ â $xMR€   ÿ Ž â xB˜º  ÿ    Ä xDL€€€€SìŽ ä xAAÿþþþÿàÿ ä xBÿÀŽ ä xPP€   MxŽ ä xBB€€€€QÞŽ äxMDÿþþþÿàÿ Ä xAA @€HøŽ â $xBB   
Ž â xFL€€€€UìŽ ä xDK˜º  ÿ    ä xMMLøŽ ä xAA˜º  ÿ    ä xSFÿþþþÿàÿ Ä xKDÿÀŽ äxFM˜º  ÿ    ä xDD€€€€OÞŽ ä xQB€€€€Q´Ž ä xBAÿÀŽ äxAA€€€€ÞŽ Ä xFÿÀŽ â $xPPÿŽ â rDDK  €ÿàÿ â $tKÿÿ Ž â xFF€€€€IÞŽ â $xAA   B
Ž â xSS @€LøŽ Æ $xFF   
Ž â xP€€€€KìŽ â &rQQ8   ÿ   â xA€€€€´Ž ä xÿþþþÿàÿ â $tAÿ @€ÿ Ž â xBS˜º  ÿ    â &rQ9   Q   â xPøŽ ä xBÿÀŽ Ä xBB€€€€ÞŽ ä xHP€€€€AìŽ ä xA˜º  ÿ    ä rBB  €ÿàÿ ä xHÿþþþÿàÿ ä xAÿÀŽ Ä xAA€€€€ÞŽ ä xEE€€€€N´Ž ä x @€PøŽ â $xAA   
Ž â xDD€€€€O´Ž â &rEE:   Q   â x˜º  ÿ    ä xFF€€€€I´Ž â &rED;   E   â xAA€€€€´Ž Ä xÿÀŽ äx€€€€ÞŽ â &rEF<   E   â xBB€€€€´Ž â &rZ?      ä &r>      ä $x   
Ž ä &rEB=   E   Ä &r?      â r   ÿ� â &rAA>   E   â x€€€€´Ž â 0rÿ    A   â xÿ   G  ä xÿ     â 0rÿJ    A   â r   ÿ� â &rA?   A   Ä $r   Ž â xÿ     â  r     A  â r   ÿ� ä rAA   ÿ� ä Er        ä z ÿ `  pPð ä x_ÿ   _  Ä xÿ     ä xdÿ   d  ä r   ÿ� ä xÿ     â #rC   C   â z ÿ a   Sð â 0rÿ^    A   â xÿ     â $r_   Ž Ä $rddA   Ž â Er        â  r     A  â r   ÿ� ä rddd   ÿ� â 0rÿb    A   â #r@   @   â 0rÿ    A   â xÿ     ä xÿ   d  â  r     A  â r     A  â Er        Ä Er        â G‰  
    € ê r/   ÿàó Ê $x   1Ž  Ê �y    ¨ �yB    è �y    ( �yA    " r/   ÿàó Ê $x   1Ž  Ê �yH    (	 �y    "	 xD   ÿ   äO xMÿ     ä xDDü  ÿÀŽ È zND  ÿàó â $xDB   ÿ Ž È� $vOÿ  ÿŽ  â xDDü  ÿÀŽ È zDD  ÿàó â �yN    ™ ¢ xKÿ   B  â $x   ÿ Ž ä$vEÿ  ÿŽ  ä %vJK  6 Ž âxü  ÿÀŽ ä �yD    ™ â %vLM  6 Ž â zF  ÿàó Ä �yJ    ™ â xA   ÿ   ä xÿ     â �yL    ™ â xNü  ÿÀŽ â $vGÿ  ÿŽ  Æ zNN  ÿàó ä �yF    ™ â  %vL  6 Ž â/ xEÿ   A  È �yL    ™ â $vOÿ  ÿŽ  ä %vJE  6 Ž Æ�yAN    ™ è �yBJ    ™ â  $rDÿÿ   / Ž ä $rEÿÿ   1 Ž È %xD+b   DŽ Ì �yDD     â xFÿ   H  È xGFÿ   ÿÀŽ È 	sI G        " xFHÿ   ÿÀŽ äxEHT      Ê/ 	sF F        b xEÿ   E  ä xLI   ÿ   ä xIÿ     ä xLG€   LxŽ ä xNIÿ   ÿÀŽ â$tIÿÿ Ž â xJEÿ   ÿÀŽ â $xLLÿŽ â 	sO N        " $tEÿ @€ÿ Ž Ä $xFF€   ÿ Ž â/ xGL€€€€IìŽ È 	sM J        b xIL€€€€EìŽ ä xEF€   ÿÀŽ È xEEÿ   HxŽ â $xOO€   ÿ Ž â xGGÿþþþÿàÿ â $tHÿ @€ÿ Ž â xF     â $xEEÿŽ â xKIÿþþþÿàÿ ä xIGLøŽ â $xMM€   ÿ Ž â/ xON€   OxŽ Ä xFE€€€€FìŽ äxGE€€€€HìŽ â $xOOÿŽ â xMJ€   MxŽ â $tJÿÿ Ž â xFFÿþþþÿàÿ ä xHGÿþþþÿàÿ ä xKK @€LøŽ â $tLÿ @€ÿ Ž â xFFEøŽ Ä xHH @€EøŽ ä xEO€€€€JìŽ â $xMMÿŽ â xGO€€€€LìŽ â $tPÿ @€ÿ Ž â xJEÿþþþÿàÿ ä xN     ä xLGÿþþþÿàÿ Ä xGM€€€€PìŽ ä xPJOøŽ äxJI˜º  ÿ    ä xEM€€€€NìŽ ä xNGÿþþþÿàÿ ä xOL @€OøŽ Ä xGJÿÀŽ ä xLEÿþþþÿàÿ ä xEF˜º  ÿ    ä xHH˜º  ÿ    ä xFEÿÀŽ ä xKK˜º  ÿ    Ä xILMøŽ äxNN @€MøŽ ä xLKÿÀŽ ä xII˜º  ÿ    ä xNN˜º  ÿ    ä xPP˜º  ÿ    Ä xOO˜º  ÿ    ä xÿ     ä xJJ€€€€ÞŽ ÊO $xJJ   G
Ž â xEE€€€€ÞŽ È� xJJ€€€€´Ž ä xHÿÀŽ â$xEE   F
Ž â xHH€€€€ÞŽ äxKK€€€€ÞŽ ä xEE€€€€´Ž ä rHH  €ÿàÿ â $xKK   L
Ž â xIÿÀŽ â &rEE8   ÿ   â xHH€€€€´Ž Ä xII€€€€ÞŽ ä xNÿÀŽ â &rEH9   E   â xNN€€€€ÞŽ â $xII   
Ž â xK€€€€´Ž â &rEJ:   E   ä $xNN   
Ž â xPÿÀŽ Ä xPP€€€€AÞŽ â &rE;   E   â xII€€€€´Ž ä xOÿÀŽ â$xPP   
Ž â xOO€€€€BÞŽ â &rEI<   E   â xNN€€€€´Ž ä xPP€€€€A´Ž â $xOO   
Ž Ä &rEN=   E   Æ xBO€€€€B´Ž â &rEP>   E   È &rBB?   E   Ê rBB   ÿ� È xÿ     Ê $rB   Ž Ê r   ÿ� â 0rÿD    A   Æ xÿ     ä  r     A  ä Er        Ê #r'   '   ä G‰  
    € ê r/   ÿàó È r1   ÿäÿ  Ê �y    ¨ �yB    è �y    ( �yA    " r
/   ÿàó Ê $x   1Ž  Ê �yH    (	 �y    "	 $xD   ÿ Ž âO xMÿ     È xDDü  ÿÀŽ È zND  ÿàó â $xDB   ÿ Ž È� $vOÿ  ÿŽ  â xDDü  ÿÀŽ È zDD  ÿàó â �yN    ™ ¢ xKÿ   B  â %vLM  6 Ž â x   ÿ   Æ$vEÿ  ÿŽ  ä %vJK  6 Ž â �yL    ™ â xü  ÿÀŽ Æ �yD    ™ â zF  ÿàó Æ �yJ    ™ â $xA   ÿ Ž â xÿ     â $vGÿ  ÿŽ  Æ xNü  ÿÀŽ â %vL  6 Ž â/�yF    ™ ä  zNN  ÿàó ä xEÿ   A  â�yL    ™ $ $vOÿ  ÿŽ  ä %vJE  6 Ž Æ �yAN    ™ ( �yBJ    ™ " rD /       â $rEÿÿ   1 Ž È %xD,b   DŽ Ì �yDD     " xFÿ   H  â $tPÿ @€ÿ Ž Æ xGFÿ   ÿÀŽ ä xFHÿ   ÿÀŽ ä	sI G        " xEHT      È/ xEÿ   E  Æ 	sF F        b xJEÿ   ÿÀŽ â $tEÿ @€ÿ Ž ä $xLI€   ÿ Ž â xIÿ     Æ 	sM J        " xÿ     ä xLG€   LxŽ ä xNIÿ   ÿÀŽ â $tIÿÿ Ž â xFF   ÿ   â/ $xLLÿŽ â 	sO N        h xGL€€€€IìŽ Ä xIL€€€€EìŽ â $xMM€   ÿ Ž â xEF€   ÿÀŽ â $tFÿÿ Ž â xGGÿþþþÿàÿ ä xEEÿ   HxŽ â $tHÿ @€ÿ Ž â xKIÿþþþÿàÿ â $xOO€   ÿ Ž â/ xIGLøŽ â $xEEÿŽ â xMJ€   MxŽ Ä xON€   OxŽ â $tNÿÿ Ž â xFE€€€€FìŽ â $xMMÿŽ â xGE€€€€HìŽ â $xOOÿŽ â xFFÿþþþÿàÿ ä xHGÿþþþÿàÿ ä xKK @€LøŽ â $tLÿ @€ÿ Ž â xJ     Ä xFFEøŽ äxHH @€EøŽ ä xEO€€€€JìŽ äxGO€€€€LìŽ ä xJEÿþþþÿàÿ ä xLGÿþþþÿàÿ Ä xGM€€€€PìŽ ä xPJOøŽ ä xJI˜º  ÿ    ä xEM€€€€NìŽ ä xNGÿþþþÿàÿ ä xGJÿÀŽ Ä xOL @€OøŽ ä xLEÿþþþÿàÿ ä xEF˜º  ÿ    ä xHH˜º  ÿ    ä xFEÿÀŽ ä xILMøŽ Ä xKK˜º  ÿ    ä xNN @€MøŽ ä xII˜º  ÿ    ä xLKÿÀŽ ä xNN˜º  ÿ    ä xPP˜º  ÿ    Ä xOO˜º  ÿ    ä xJJ€€€€ÞŽ ÊO $xJJ   G
Ž â xKK€€€€ÞŽ È� xJJ€€€€´Ž ä xEE€€€€ÞŽ ä xHÿÀŽ äxHH€€€€ÞŽ â $xEE   F
Ž â rKKL  €ÿàÿ Æ $xHH   
Ž â xEE€€€€´Ž ä xIÿÀŽ ä xHH€€€€´Ž â &rEE8   ÿ   â xII€€€€ÞŽ ä xNÿÀŽ â &rEH9   E   â xNN€€€€ÞŽ â$xII   
Ž â xK€€€€´Ž â &rEJ:   E   Ä $xNN   
Ž â xPÿÀŽ â &rE;   E   â xII€€€€´Ž ä xPP€€€€AÞŽ ä xOÿÀŽ â &rEI<   E   â xOO€€€€BÞŽ ä xNN€€€€´Ž Ä rPP  €ÿàÿ â $xOO   
Ž ä &rEN=   E   â xPP€€€€A´Ž â 0rÿD    A   â xBO€€€€B´Ž Æ &rEP>   E   â  r     A  Æ &rBB?   E   Ê rBB   ÿ� È xÿ     Ê $rB   Ž Ê r   ÿ� È xÿ     È Er        Ê #r&   &   ä G‰  
    € ê r/   ÿàó Ê $x	   1Ž  Ê �y    ¨ �yB    è �y    ( �yA    " r/   ÿàó Ê $x   1Ž  Ê �yH    (	 �y    "	 $xD   ÿ Ž âO xMÿ     È xDDü  ÿÀŽ È zND  ÿàó ä xDB   ÿ   Æ� $vOÿ  ÿŽ  â xDDü  ÿÀŽ È zDD  ÿàó â �yN    ™ ¢ xKÿ   B  â $x   ÿ Ž ä$vEÿ  ÿŽ  ä %vJK  6 Ž âxü  ÿÀŽ ä �yD    ™ â %vLM  6 Ž â zF  ÿàó Ä �yJ    ™ â $xA   ÿ Ž â xÿ     ä �yL    ™ ä xNü  ÿÀŽ ä zGÿ  ÿäÿ  ä zNN  ÿàó Ä xEÿ   A  â/ �yF    ™ â  %vL  6 Ž Ê�yL    ™ " $vOÿ  ÿŽ  ä %vJE  6 Ž Æ �yAN    ™ ( �yBJ    ™ " $rDÿÿ   / Ž ä $rEÿÿ   1 Ž È %xD.b   DŽ Ì �yDD     " xFÿ   H  È xGFÿ   ÿÀŽ È 	sI G        " xFHÿ   ÿÀŽ äxEHT      Ê/ 	sF F        b xEÿ   E  â $xLI€   ÿ Ž â xIÿ     È xLG€   LxŽ ä xNIÿ   ÿÀŽ ä xJEÿ   ÿÀŽ â $xLLÿŽ â 	sO N        " xI     â $tEÿ @€ÿ Ž ä $xFF€   ÿ Ž â/ xGL€€€€IìŽ Æ 	sM J        b xIL€€€€EìŽ ä xEF€   ÿÀŽ È xEEÿ   HxŽ â $xOO€   ÿ Ž â xGGÿþþþÿàÿ â $tFÿÿ Ž â xH  @€    â $xEEÿŽ â xKIÿþþþÿàÿ ä xIGLøŽ â $xMM€   ÿ Ž â/ xON€   OxŽ Ä xFE€€€€FìŽ äxGE€€€€HìŽ â $xOOÿŽ â xMJ€   MxŽ â $tJÿÿ Ž â xFFÿþþþÿàÿ ä xHGÿþþþÿàÿ ä xKK @€LøŽ â $tLÿ @€ÿ Ž â xFFEøŽ Ä xHH @€EøŽ ä xEO€€€€JìŽ â $xMMÿŽ â xGO€€€€LìŽ ä xP  @€    ä xJEÿþþþÿàÿ â $tNÿÿ Ž â xLGÿþþþÿàÿ Ä xGM€€€€PìŽ ä xPJOøŽ ä xJI˜º  ÿ    ä xEM€€€€NìŽ ä xNGÿþþþÿàÿ ä xGJÿÀŽ Ä xOL @€OøŽ ä xLEÿþþþÿàÿ ä xEF˜º  ÿ    ä xHH˜º  ÿ    ä xFEÿÀŽ ä xKK˜º  ÿ    Ä xNN @€MøŽ äxILMøŽ ä xLKÿÀŽ ä xNN˜º  ÿ    ä xII˜º  ÿ    ä xPP˜º  ÿ    Ä xOO˜º  ÿ    ä xÿ     ä xJJ€€€€ÞŽ ÊO $xJJ   G
Ž â xEE€€€€ÞŽ È� xJJ€€€€´Ž ä xHÿÀŽ â$xEE   F
Ž â xHH€€€€ÞŽ ä xKK€€€€ÞŽ ä xEE€€€€´Ž â $xHH   
Ž â xNÿÀŽ â $xKK   L
Ž Ä &rEE8   ÿ   â xHH€€€€´Ž ä xIÿÀŽ äxII€€€€ÞŽ â &rEH9   E   â xNN€€€€ÞŽ äxK€€€€´Ž â &rEJ:   E   â rII  €ÿàÿ â $xNN   
Ž â xPÿÀŽ Ä xPP€€€€AÞŽ â &rE;   E   â xII€€€€´Ž ä xOÿÀŽ â$xPP   
Ž â xOO€€€€BÞŽ â &rEI<   E   â xNN€€€€´Ž ä xPP€€€€A´Ž â $xOO   
Ž Ä &rEN=   E   Æ xBO€€€€B´Ž â &rEP>   E   È &rBB?   E   Ê rBB   ÿ� È xÿ     Ê $rB   Ž Ê r   ÿ� â 0rÿD    A   Æ xÿ     ä  r     A  ä Er        Ê #r%   %   â #r#   #   â #r"   "   â G‰   
    € ê r/   ÿàñ Ê $x   1  Ê �y    ¨ �y    è �yA    ( �yB    "	 r/   ÿàñ È r1   ÿä  Ê �yL    ( �y    " $x   ÿ Ž âO xÿ     â $x   ÿ Ž Æ� xü  ÿÀŽ È z  ÿàñ Ê $vÿ  ÿ  ä %v  6 Ž Æ �y    ™ ¢ xü  ÿÀŽ â $xA   ÿ Ž ä�y    ™ â zF  ÿàñ ä xÿ     ä xHü  ÿÀŽ ä zGÿ  ÿä  â %v  6 Ž â zHH  ÿàñ Æ $xB   ÿ Ž â/ �y    ™ " xEÿ   A  â $vIÿ  ÿ  ä �yF    ™ " xü  ÿÀŽ â %vDE  6 Ž ä �yH    ™ " zJ  ÿàñ Æ �yD    ™ " xGÿ   B  â $vKÿ  ÿ  È %vFG  6 Ž â �yJ    ™ ( �yF    ™ " r 1       â/ $rÿÿ   / Ž È %x-b   Ž Ì �y     " x6Lÿ   ÿÀŽ ä xAÿ   L  È 	s6 6        " xBAÿ   ÿÀŽ ä xEÿ     Ê 	sD B        " xLT      ä/ xGEÿ   ÿÀŽ â $xA6€   ÿ Ž â xÿ     Æ 	sH G        " xAA€   ÿÀŽ ä xFÿ   ÿÀŽ ä xAAÿ   LxŽ â $xD€   ÿ Ž ä $tDÿÿ Ž ä $xAAÿŽ Ä $tJÿ @€ÿ Ž â x6B€   xŽ â 	sE F        b xA€€€€DìŽ äxBA€€€€JìŽ â $xD6ÿŽ â xI     â $tKÿ @€ÿ Ž â x6ÿþþþÿàÿ â $xHH€   ÿ Ž â xBBÿþþþÿàÿ Ä x66AøŽ äxBB @€AøŽ ä xD€€€€IìŽ äxAD€€€€KìŽ ä xHG€   HxŽ ä xÿþþþÿàÿ Ä xAAÿþþþÿàÿ â $xGE€   ÿ Ž â/ xEDøŽ â$xHHÿŽ â xDA @€DøŽ â $tAÿ @€ÿ Ž â xGF€   GxŽ ä xH€€€€IìŽ äxFH€€€€AìŽ Ä xAÿþþþÿàÿ ä x6˜º  ÿ    ä xIAHøŽ ä x6ÿÀŽ ä xAB˜º  ÿ    â $xGGÿŽ â xJ     â $tLÿ @€ÿ Ž â xKFÿþþþÿàÿ Ä xFG€€€€JìŽ ä xKK @€HøŽ ä xHG€€€€LìŽ ä xFFÿþþþÿàÿ ä xDD˜º  ÿ    ä xEE˜º  ÿ    Ä xHHÿþþþÿàÿ ä xFFGøŽ äxHH @€GøŽ ä xFF˜º  ÿ    ä xHH˜º  ÿ    ä xII˜º  ÿ    Ä xÿ     ä x€€€€ÞŽ ÊO $x   6
Ž â x6AÿÀŽ äxAA€€€€ÞŽ Ê� $xAA   6
Ž â x€€€€´Ž È xA€€€€´Ž ä xADÿÀŽ äxDD€€€€ÞŽ â&r8   ÿ   â xEÿÀŽ äxEE€€€€ÞŽ â $xDD   A
Ž ä &r9      â rEE  €ÿàÿ Ä x9FÿÀŽ äxFF€€€€ÞŽ ä xEE€€€€´Ž ä xDD€€€€´Ž ä xHÿÀŽ â $xFF   9
Ž â xHH€€€€ÞŽ â &rE:      â xK˜º  ÿ    Ä xIÿÀŽ â$xHH   
Ž â xFF€€€€´Ž â &rD;      â xII€€€€ÞŽ ä xÿÀŽ äx€€€€ÞŽ â &rF<      â xHH€€€€´Ž Ä rII  €ÿàÿ â $x   
Ž ä &rH=      â xII€€€€´Ž ä x€€€€´Ž Æ &rI>      È &r?      Ê r   ÿ� È xÿ     Ê $r   Ž Ê r   ÿ� â 0rÿ    A   Æ xÿ     ä  r     A  ä Er        Ê #r$   $   â x**   ÿàÿ ä x//   ÿàó ä | *   pbð ä x33   ÿàõ â $r1ÿÿ   1Ž  â x00€   ÿàÿ Æ $r5ÿÿ   5 Ì D	       À â Gy  P®ÿÿÿÿƒ ê Ay        € ê y      "   " Ey  0    € â r ÿ   pRð är ÿ   pRò Ö G‰  ð     € ê z ÿ `  pPð ä r    ÿ8Ž ä z ÿ a   Sð Æ ˆs C þÿ H   ô ˆƒ @€þÿ H   è ˆƒ # ÿÿ H   è ˆƒ "€ÿÿ H   â G‰  p     € ê ˆs '   H   è ˆs @€þÿ H   è ˆs &€  H   è ˆs # ÿÿ H   è ˆs %   H   è ˆs "€ÿÿ H   è ˆs $€  H   ä Ay        € ê {          ì M        € ê y      &   ¢ z ÿ `  pPð â $tÿ   ÿ Ž Ä y      '   "> z ÿ a   Sð â $z x  Ž ÈO $z ~  Ž È %v j   Ž È G	  P    € ê „y      H   " z  r  p`ð Æ „y	  €   H   b r  ÿ   pVp  Æ „y     H   ¨ „y  €  H   è „y     H   ( „y  €  H   ( „y     H   ( „y     H   ( „y     H   ( „y  €  H   ( „y  €  H   " !rC       Æ „y  €  H   " !r		@       â/ !r       ÆO !r	       â� !r       Æ!r       ä ‰      b !r#       Æ ‰      ¢ !r       È !r       â !r"       È ‰
      â !r       È !r       â !r       â/ !r       ÆO ‰	      ( ‰      h ‰      ¢ !r

       Ê� ‰
      â !r	       â !r       È/ ‰      " !r	       ÆO ‰ €    h ‰	 €    ¢ !r
       Ê� ‰ €    â !r       â !r       È/ ‰ €    " !r	       ÆO ‰ @    h ‰ @    ¢ !r       Ê� ‰ @    â !r       â !r       È/ ‰ @    " x   ÿàÿ â!r       âO x   ÿàÿ â ‰      b z  r  p`ò ä z  r  p`ô â !r       â� ‰	      ¢ x   ÿàÿ Ä x     pVò  â‰
      â x     pVt â !r       â !r       È/ ‰      " !‚!       â!r	       ÈO †‰     â z  r  p`ð â !r

       â� !’!       äx     pVp  â !¢	
!       ä †™    è †© 	   â !r        Ê M	        € ê !r! !       Ê †y !   â My        € ê $vÿ h  ÿ Ž Ê x    pBð Ú G	  @    € ê x    pRð Ú G‰  À    € ê x    pRð Ú G	   $    € ê „y      H   " z  r  p`ð â Ey  p    € ä „y     H   b r  ÿ   pVp  Æ „y	     H   ¨ „y     H   è „y     H   ( „y   
  H   ( „y  €   H   ( „y  €  H   ( „y  €  H   è „y  €
  H   â !rC       â !r'       Æ/ !r	       âO !r       Æ� !r       â!r       È ‰      " !r@       Æ ‰      b !r&       â !r       ä „y  €  H   â !r       Æ/ ‰      ( ‰      b !r       â !r	       È/ ‰ €    ( ‰
	 €    b !r       Æ „y  €  H   ¢ !r       â !r       â !r
	
       È/ ‰ @    ( ‰
 @    b !r       ÈO !r       â !r       â !r
       È/ ‰      ( ‰      b !r       â !r       â/ G	  �     € ê !r        â!r!       Æ  x;ª¸¿  A  Ì s        $ !t  €?    Ì s
        $  r
     A  È  r     A  Ê †y     ä Ay        € ê ‰      b Ey  Ð    € æ ‰      ( „y     H   ¨ „y     H   è „y     H   è „y     H   â !r       Æ/ „y     H   â !r	       Æ ‰      ( ‰	      b !r#       âO !r%       â� !r       ä !r       â !r
	       Ä/ ‰ €    " !r       Æ ‰
 €    h „y	   	  H   ¢ !r       â !r
       È/ ‰ @    ( ‰ @    b !r		       ÈO !r		       â !r       â x   ÿàÿ â !r       Æ/ ‰      " z  r  p`ð Æ ‰      b x     pVp  â !r       â !r       Ö/ G	  �     € ê !r        â!r!       Æ  x;ª¸¿  A  Ì s        $ !t  €?    Ì s        $  r     A  È  r     A  Ê †y    ä Ay        € ê ‰      b Ey       € æ ‰	      ¨ „y  €	  H   è „y  €  H   è „y  €  H   â !r       Æ/ „y  €  H   â !r	       ÆO ‰      ( ‰      h „y	  €  H   ¢ !r       â !r
       È/ ‰ €    ( ‰
 €    h „y  €  H   â !r		$       ÈO !r		       â !r       â !r
       È/ ‰ @    ( ‰ @    b !r	       â� !r       â x   ÿàÿ â !r       Æ/ ‰      " z  r  p`ð Æ ‰      b x     pVp  â !r       â !r       Ö/ G	  �     € ê !r        â!r!       Æ  x;ª¸¿  A  Ì s        $ !t  €?    Ì s
        $  r
     A  È  r     A  Ê †y    ä Ay        € ê ‰      " !r"       â x   ÿàÿ Æ !r       â z  r  p`ð Æ !r       â x     pVp  È ‰      b !r	       Ê ‰	      " !r       Ê/ ‰      b !r
	       Ê ‰
 €    " !r       Ê/ ‰ €    b !r
       Ê ‰ @    " !r       Ê/ ‰ @    b !r       Ê ‰      " !r       Ê/ ‰       ¢ !r       â M	        € ì !r         â/!r         ÆO  x ;ª¸¿  A  â !r! !       Ê s        $ !t  €?    Ì s        $  r       A  È  r!!      A  Ê †y !   â My        € ê „y      H   " z  r  p`ð â Ey       € ä „y     H   b r  ÿ   pVp  Æ „y     H   ¨ „y     H   è „y	     H   ( „y   
  H   ( „y  €  H   ( „y  €  H   â !rC       â !r'       Æ/ !r       âO !r       Æ� !r	       â!r       È ‰      " !r&       Æ ‰      h „y  €   H   ¢ !r       â !r       È/ ‰      ( ‰      h „y  €  H   â !r@       âO !r       â !r	       È/ ‰ €    ( ‰
	 €    h „y  €  H   ¢ !r       â� !r       Ä „y  €
  H   â !r
	
       Æ/ ‰ @    ( ‰
 @    b !r       ÈO !r       â !r       â !r
       È/ ‰      ( ‰      b !r       â !r
       â !r       â/ G	  À     € ê !r        È x'7=  A  â x*BL?  A  Æ #t  €?   È  r     A  â  x   ?  A  â !r
!       È s     $   $ !t  €?    È  r     A  È  r     A  Ê †y     ä Ay        € ê ‰      b Ey       € æ ‰      ( „y     H   ¨ „y     H   è „y   	  H   è „y     H   â !r       Æ/ „y     H   â !r       Æ ‰      ( ‰      h „y     H   â !r%       âO !r       Æ !r       â !r	       â/ ‰ €    $ !r       ä ‰
	 €    b !r#       È� !r       â !r       â x   ÿàÿ ä !r       â !r
	
       â/ ‰ @    " z  r  p`ð Æ ‰
 @    b x     pVp  â !r       â !r
       È/ ‰      ( ‰      b !r
       â !r       â/ G	  À     € ê !r        È x'7=  A  â x*BL?  A  Æ #t  €?   È  r     A  â  x   ?  A  â !r
!       È s     $   $ !t  €?    È  r     A  È  r     A  Ê †y    ä Ay        € ê ‰      b Ey  Ð    € æ ‰      ( „y  €  H   ¨ „y  €  H   è „y  €  H   è „y  €  H   â !r       Æ/ „y  €  H   â !r       Æ ‰      ( ‰      h „y  €	  H   â !r$       âO !r       â !r	       È/ ‰ €    ( ‰
	 €    b !r       È� !r       â !r       â x   ÿàÿ â !r
	
       Æ/ ‰ @    " z  r  p`ð Æ ‰
 @    b x     pVp  â !r       â !r
       È/ ‰      ( ‰      b !r
       â !r       â/ G	  À     € ê !r        È x'7=  A  â x*BL?  A  Æ #t  €?   È  r     A  â  x   ?  A  â !r
!       È s     $   $ !t  €?    È  r     A  È  r     A  Ê †y    ä Ay        € ê ‰      " !r"       â x   ÿàÿ Æ !r       â z  r  p`ð Æ !r       â x     pVp  È ‰      b !r       Ê ‰      " !r       Ê/ ‰      b !r	       Ê ‰
	 €    " !r       Ê/ ‰ €    b !r
	
       Ê ‰
 @    " !r       Ê/ ‰ @    b !r
       Ê ‰      " !r       Ê/ ‰       ¢ !r       â M	        € ì !r         â/!r         ÆO  x '7=  A  â x *BL?  A  â !r! !       ä #t   €?   â x     ?  A  Æ  r     A  Ì s     $   $ !t  €?    È  r       A  È  r!!      A  Ê †y !   â My        € ê x    pRð Ú G‰  P    € ê x    pRð Ú G	  `    € ê „y      H   " z  r  p`ð â Ey       € ä „y     H   b r  ÿ   pVp  Æ „y     H   ¨ „y     H   è „y	     H   ( „y   
  H   ( „y  €  H   ( „y  €  H   â !rC       â !r'       Æ/ !r       âO !r       Æ� !r	       â!r       È ‰      " !r&       Æ ‰      h „y  €   H   ¢ !r       â !r       È/ ‰      ( ‰      h „y  €  H   â !r@       âO !r       â !r	       È/ ‰ €    ( ‰
	 €    h „y  €  H   ¢ !r       â� !r       Ä „y  €
  H   â !r
	
       Æ/ ‰ @    ( ‰
 @    b !r       ÈO !r       â !r       â !r
       È/ ‰      ( ‰      b !r       â !r
       â !r       â/ G	  À     € ê !r        Ê	z i    � â !r
!       È  x;ª¸¿  A  â 	z i    � È 	z i €  � â s        $ !t  €?    Ì s        $  r     A  È  r     A  Ê †y     ä Ay        € ê ‰      b Ey       € æ ‰      ( „y     H   ¨ „y     H   è „y   	  H   è „y     H   â !r       Æ/ „y     H   â !r       Æ ‰      ( ‰      h „y     H   â !r%       âO !r       Æ !r       â !r	       â/ ‰ €    $ !r       ä ‰
	 €    b !r#       È� !r       â !r       â x   ÿàÿ ä !r       â !r
	
       â/ ‰ @    " z  r  p`ð Æ ‰
 @    b x     pVp  â !r       â !r
       È/ ‰      ( ‰      b !r
       â !r       â/ G	  À     € ê !r        Ê	z i    � â !r
!       È  x;ª¸¿  A  â 	z i    � È 	z i €  � â s        $ !t  €?    Ì s        $  r     A  È  r     A  Ê †y    ä Ay        € ê ‰      b Ey  Ð    € æ ‰      ( „y  €  H   ¨ „y  €  H   è „y  €  H   è „y  €  H   â !r       Æ/ „y  €  H   â !r       Æ ‰      ( ‰      h „y  €	  H   â !r$       âO !r       â !r	       È/ ‰ €    ( ‰
	 €    b !r       È� !r       â !r       â x   ÿàÿ â !r
	
       Æ/ ‰ @    " z  r  p`ð Æ ‰
 @    b x     pVp  â !r       â !r
       È/ ‰      ( ‰      b !r
       â !r       â/ G	  À     € ê !r        Ê	z i    � â !r
!       È  x;ª¸¿  A  â 	z i    � È 	z i €  � â s        $ !t  €?    Ì s        $  r     A  È  r     A  Ê †y    ä Ay        € ê ‰      " !r"       â x   ÿàÿ Æ !r       â z  r  p`ð Æ !r       â x     pVp  È ‰      b !r       Ê ‰      " !r       Ê/ ‰      b !r	       Ê ‰
	 €    " !r       Ê/ ‰ €    b !r
	
       Ê ‰
 @    " !r       Ê/ ‰ @    b !r
       Ê ‰      " !r       Ê/ ‰       ¢ !r       â M	        € ì !r         â/!r         ÈO 	z   i    � â !r! !       È  x ;ª¸¿  A  â 	z!! i    � È 	z!! i €  � â s        $ !t  €?    Ì s        $  r       A  È  r!!      A  Ê †y !   â My        € ê „y      H   " x   ÿàÿ äz  r  p`ð â „y     H   b z  r  p`ò ä r  ÿ   pVp  â„y  €   H   ¢ x     pVò  Æ „y  €  H   è „y     H   ( „y     H   ( „y     H   ( „y     H   ( „y  €  H   ( „y  €  H   ( „y     H   " !rC       Æ „y   	  H   " !r'       Æ/ „y	     H   b !r@       ÆO „y   
  H   ¢ !r&&       Æ� „y  €  H   â !r#       Æ„y
  €
  H   " !r%       Æ „y     H   " !r       Æ „y     H   " !r       Æ „y  €  H   " !r       Æ „y)  €  H   " !r&       Æ „y  €  H   " !r       â x   ÿàÿ ä „y  €	  H   â !r       â z  r  p`ô ä „y+  €  H   â !r	       â/ x     pVt Ä „y-  €  H   â !r       âO x   ÿàÿ â !r       â� ‰      $ z  r  p`ö â !r

       â‰      d x     pVö â !r       â ‰      ¢ !r       Æ ‰	
      â !r"       Æ ‰      " !r))$       Æ ‰      " !r       â !r       â !r)       â !r       Æ/ ‰      " !r++       â !r       äO ‰      b !r-       â !r

	       ä� ‰+      ¢ !r       Æ‰      â !r       Æ ‰	      ( ‰
      ( ‰      " !r       Æ ‰      " !r       â/ !r+       äO ‰ €    b !r       Æ� ‰ €    ¢ !r		       Æ‰      â !r
       Æ ‰      " !r       Æ ‰	 €    " !r       Æ ‰
 €    ( ‰ €    " !r       Æ/ ‰ €    b !r       âO !r       ä� ‰ @    ¢ !r       Æ‰ €    â !r	       Æ ‰ €    " !r

       Æ ‰ @    " !r       Æ ‰	 @    " !r       Æ/ ‰
 @    h ‰ @    b !r       ÆO ‰ @    ¢ !r       â� !r       ä‰      â !r       Æ ‰ @    " !r		       Æ ‰ @    " !r
       Æ/ ‰      b !r       Æ ‰	      b !r       ÆO ‰
      ¨ ‰      ¢ !r       Æ� ‰      â !r       â !‚!       â!r       Æ‰      " !r       Æ/ ‰      b !r	       â !‚        â !r
       äO !’!       â  ‚     A  â !r       â !’        Æ †‰     â !r	       â� !¢!       â  ’     A  ä !¢		        â !r       ä †™    â  ¢		     A  â !r       È/ †© 	   â M9        € ê !r!!       â !r         È  r!!      A  Ê †y !   â My        € ê „y      H   " z  r  p`ð â Ey  À    € ä „y     H   b r  ÿ   pVp  Æ „y     H   ¨ „y     H   è „y	     H   ( „y   
  H   ( „y  €  H   ( „y  €  H   â !rC       â !r'       Æ/ !r       âO !r       Æ� !r	       â!r       È ‰      " !r&       Æ ‰      h „y  €   H   ¢ !r       â !r       È/ ‰      ( ‰      h „y  €  H   â !r@       âO !r       â !r	       È/ ‰ €    ( ‰
	 €    h „y  €  H   ¢ !r       â� !r       Ä „y  €
  H   â !r
	
       Æ/ ‰ @    ( ‰
 @    b !r       ÈO !r       â !r       â !r
       È/ ‰      ( ‰      b !r       â !r
       â !r       â/ G	  à     € ê !r        Ê	x  à@  � Ê  x#ÛÙ?  A  È  x;ª¸¿  A  â !r
!       Ê s        " 	x  à@  � È 	x  àÀ  � Ê !t  €?    â !t  €?    È s        $  r     A  È  r     A  Ê †y     ä Ay        € ê ‰      b Ey       € æ ‰      ( „y     H   ¨ „y     H   è „y   	  H   è „y     H   â !r       Æ/ „y     H   â !r       Æ ‰      ( ‰      h „y     H   â !r%       âO !r       Æ !r       â !r	       â/ ‰ €    $ !r       ä ‰
	 €    b !r#       È� !r       â !r       â x   ÿàÿ ä !r       â !r
	
       â/ ‰ @    " z  r  p`ð Æ ‰
 @    b x     pVp  â !r       â !r
       È/ ‰      ( ‰      b !r
       â !r       â/ G	  à     € ê !r        Ê	x  à@  � Ê  x#ÛÙ?  A  È  x;ª¸¿  A  â !r
!       Ê s        " 	x  à@  � È 	x  àÀ  � Ê !t  €?    â !t  €?    È s        $  r     A  È  r     A  Ê †y    ä Ay        € ê ‰      b Ey  ð    € æ ‰      ( „y  €  H   ¨ „y  €  H   è „y  €  H   è „y  €  H   â !r       Æ/ „y  €  H   â !r       Æ ‰      ( ‰      h „y  €	  H   â !r$       âO !r       â !r	       È/ ‰ €    ( ‰
	 €    b !r       È� !r       â !r       â x   ÿàÿ â !r
	
       Æ/ ‰ @    " z  r  p`ð Æ ‰
 @    b x     pVp  â !r       â !r
       È/ ‰      ( ‰      b !r
       â !r       â/ G	  à     € ê !r        Ê	x  à@  � Ê  x#ÛÙ?  A  È  x;ª¸¿  A  â !r
!       Ê s        " 	x  à@  � È 	x  àÀ  � Ê !t  €?    â !t  €?    È s        $  r     A  È  r     A  Ê †y    ä Ay        € ê ‰      " !r"       â x   ÿàÿ Æ !r       â z  r  p`ð Æ !r       â x     pVp  È ‰      b !r       Ê ‰      " !r       Ê/ ‰      b !r	       Ê ‰
	 €    " !r       Ê/ ‰ €    b !r
	
       Ê ‰
 @    " !r       Ê/ ‰ @    b !r
       Ê ‰      " !r       Ê/ ‰       ¢ !r       â M	        € ì !r         â/!r         ÈO 	x    à@  � â !r! !       È  x #ÛÙ?  A  â 	x!!  à@  � Æ  x;ª¸¿  A  â 	x!!  àÀ  � Ê s        " !t!!  €?    â !t  €?    Ì s        $  r       A  È  r! !     A  Ê †y !   â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y       "   $ z   w  p`ð Ú M	        € ê y      !   " ¹z  l      â Ey  @    € â ™x?    â $rÿÿ   ÿ Ž â ¹z  F   
   â $r#ÿÿ   ÿ Ž â x%ÿ     È | %   p`ð Ú G	  À    € ê y      &   " $tÿ   ÿ Ž ä $z  x  Ž È %v \   Ž Ì �y    ™ ¢  $rÿÿ   ÿ Ž â x   ÿÀŽ â $z  r  ÿŽ â y      %   b 'z m   Ž È $x   ÿ Ž â zÿ n    â $r	ÿÿ   ÿ Ž ä $vÿ X  ÿ Ž â x   ÿÀŽ â $rÿÿ   
Ž ä $v!ÿ Y  ÿ Ž ä $z o  Ž Ä $rÿÿ   ÿ Ž ä $z u  ÿŽ ä $x%   ÿ Ž ä $x   ÿ Ž â/ r   ÿàó Æ $z q  %Ž âx   ÿàÿ â $rÿÿ   ÿŽ  ä $z t  %Ž äO$z t  Ž ä $z q  Ž ä %xb   ÿŽ È %xb   Ž â r    ÿàñ Æ %xb   ÿŽ È %x    Ž È $r"ÿÿ     ä %x
    Ž È r   ÿàñ Ê $x	   !  Ê �y   • ¨ �y   • è �y   • ( �y/   • b r
   ÿàñ Ê $x   !  Ê �y   • h �y   • h �y:   • h �y2   • b r    ÿàñ Ê $x!   "  Ê �y'B   • h �y$D   • d xÿ     ä/ r   ÿàõ â $t1ÿ$   ÿ Ž ä $x   ÿ Ž âO x,ÿ     È� xü  ÿÀŽ È z(  ÿàñ â $x   ÿ Ž ä $x   ÿ Ž ä$v)ÿ  ÿ  â xü  ÿÀŽ ä xÿ     ä z  ÿàñ â �y)(    ™ ¢ xü  ÿÀŽ â %v   Ž È $vÿ  ÿ  â z  ÿàñ â %v,,   Ž â �y*    ™ æ  $vÿ  ÿ  â r.   à‘ â $x/   ÿ Ž â�y+,    ™ h	 xü  ÿÀŽ â �y&    ™ b xÿ     Æ �y(    ™ b# r,ÿ   $~  â$x   ! â z  ÿàñ Æ $x-,$   ÿŽ â �y3D   • ( �y,B   • " %v. Z  1 Ž È $vÿ  ÿ  ä %v   Ž â xÿ   /  ä/ �y.    ™ " $x   -Ž Æ �y-    ™ " %v   Ž Æ �y1   ™ ( �y/   ™ ( �y0   ™ ( �y4    ™ ( �y5   ™ ( �y6   ™ " $x   ÿ Ž Æ/ �y=   ™ $ xü  ÿÀŽ ä �y<   ™ $ z  ÿàñ Ê $vÿ  ÿ  â xÿ     È �y7    ™ " $x   ÿ Ž ä %v   Ž Æ xü  ÿÀŽ â $x:   ÿ Ž â �y8    ™ " xÿ     ä z  ÿàñ ä xü  ÿÀŽ â %v   Ž È $vÿ  ÿ  â z  ÿàñ â $x2   ÿ Ž â xÿ   :  ä �y9    ™ " $vÿ  ÿ  â xü  ÿÀŽ ä �y:    ™ " %v   Ž â z  ÿàñ Ä �y    ™ " x2ÿ   2  ä $vÿ  ÿ  â �y;    ™ $ %v2   Ž ä �y    ™ ( �y    ™ ( �y2    ™ " r   ÿàñ Ä r   ÿàó â/ �y    • $ $x   !  ä$x   !Ž  Æ �y    • ( �y    • " x>'ÿ   ÿÀŽ Ì 	s> >        b xÿ   '  È xÿ   ÿÀŽ È 	sA         " xÿ   $  â$x?>€   ÿ Ž â/ x'T  $    È x@?€   ÿÀŽ ä xÿ   ÿÀŽ ä xÿ     ä x@ÿ   'xŽ â 	s         b xÿ   ÿÀŽ â $x>A€   ÿ Ž ä $t@ÿÿ Ž Ä $xÿŽ â x€   >xŽ â $tBÿ @€ÿ Ž â 	s         $ x'€€€€@ìŽ â $xÿŽ â x?€€€€BìŽ â $tBÿÿ Ž â x>'ÿþþþÿàÿ ä x@?ÿþþþÿàÿ Ä x'>øŽ äx>€€€€BìŽ â $xB€   ÿ Ž â/ x@ @€øŽ â $tDÿ @€ÿ Ž â x>>ÿþþþÿàÿ â $x@€   ÿ Ž â xB€   BxŽ ä x>øŽ â $t>ÿÿ Ž â x€   @xŽ â $xBÿŽ â x?€€€€DìŽ â $t@ÿ @€ÿ Ž Ä $xÿŽ â x?ÿþþþÿàÿ â $tBÿ @€ÿ Ž â x>€€€€>ìŽ äx@€€€€@ìŽ ä x @€øŽ â $tÿÿ Ž â xA>ÿþþþÿàÿ ä x?€€€€BìŽ Ä x>@ÿþþþÿàÿ ä x@'˜º  ÿ    ä x€€€€ìŽ ä xD?ÿþþþÿàÿ ä x?@ÿÀŽ ä xBÿþþþÿàÿ Ä xAøŽ äx> @€øŽ ä x'BøŽ äxD @€øŽ ä xD˜º  ÿ    ä xA˜º  ÿ    Ä x''˜º  ÿ    ä xAÿÀŽ ä x@@€€€€)ÞŽ ÊO $x>@   ?
Ž â x?˜º  ÿ    È x?ÿÀŽ äx@?€€€€*ÞŽ ä� x>>€€€€)´Ž Æ $x@   
Ž â x)'ÿÀŽ ä xBA€€€€&ÞŽ äx*€€€€*´Ž ä x''€€€€(ÞŽ ä xDÿÀŽ â $xB   
Ž â xDD€€€€+ÞŽ â $x''   )
Ž â x,ÿ   ÿÀŽ Æ$xDD   
Ž â x)˜º  ÿ    ä x˜º  ÿ    â 	s         " x€€€€&´Ž ä x'€€€€(´Ž ä xÿ   ,  ä x()ÿÀŽ Ä x'ÿÀŽ ä x))€€€€.ÞŽ â &r>>1   ÿ   â x€€€€-ÞŽ ä xÿ   ÿÀŽ â $x))   (
Ž â x&D€€€€+´Ž â $x   '
Ž â x+˜º  ÿ    â &r**/   >   â 	s'         b x)€€€€.´Ž Ä x(+ÿÀŽ â &r0   *   â x++€€€€4ÞŽ â $x)€   ÿ Ž â x*ÿ   3  â &r&&5      ä $x+   (
Ž â x+)€   ÿÀŽ ä x*ÿ   ÿÀŽ Ä x++ÿ   ,xŽ â &r6   &   â x,,T  3    â 	s)         " $x&'€   ÿ Ž ä/ x,ÿ   ,  â $t*ÿÿ Ž ä x(€   &xŽ â $x+ÿŽ â x&,ÿ   ÿÀŽ â $t.ÿ @€ÿ Ž â x€€€€-´Ž â $t-ÿÿ Ž â 	s' &        b $x((ÿŽ â x*€€€€*ìŽ â$t?ÿ @€ÿ Ž â x+€€€€.ìŽ â $x.)€   ÿ Ž â x**ÿþþþÿàÿ Ä x)(€€€€-ìŽ äx-(€€€€?ìŽ ä x**øŽ ä x,+ÿþþþÿàÿ ä x--ÿþþþÿàÿ ä x+)ÿþþþÿàÿ Ä x.€   .xŽ ä x**˜º  ÿ    ä x), @€øŽ ä x- @€(øŽ â$x-'€   ÿ Ž â/ x+(øŽ â $x'.ÿŽ â x+*ÿÀŽ â$t.ÿ @€ÿ Ž â x(*€€€€7ÞŽ â $t,ÿÿ Ž â x*&€   -xŽ Æ $x((   +
Ž â x+'€€€€.ìŽ ä x))˜º  ÿ    â $x**ÿŽ â x.+ÿþþþÿàÿ â $t+ÿÿ Ž â x&'€€€€,ìŽ ä x,)ÿÀŽ â$t-ÿ @€ÿ Ž â x))€€€€8ÞŽ Ä x+*€€€€+ìŽ ä x&&ÿþþþÿàÿ â $x))   ,
Ž â x,˜º  ÿ    ä x-*€€€€-ìŽ ä x++ÿþþþÿàÿ ä x˜º  ÿ    Ä x((€€€€7´Ž ä x&&'øŽ äx'. @€'øŽ ä x,ÿÀŽ ä x--ÿþþþÿàÿ ä x++*øŽ Ä x.ÿÀŽ äx,,€€€€9ÞŽ ä x€€€€:ÞŽ â &r((1   ÿ   â x))€€€€8´Ž ä x-- @€*øŽ ä x++˜º  ÿ    â $x,,   
Ž Ä $x   .
Ž â x*-˜º  ÿ    â &r))/   (   â x(+ÿÀŽ äx++€€€€ÞŽ ä x,,€€€€9´Ž ä x:€€€€:´Ž ä x*ÿÀŽ Ä x&˜º  ÿ    â $x++   (
Ž â x**€€€€;ÞŽ â &r),0   )   â x&ÿÀŽ ä x''˜º  ÿ    â $x**   
Ž â x€€€€ÞŽ â &r):5   )   â x++€€€€´Ž Ä x'ÿÀŽ â $x   &
Ž â x''€€€€ÞŽ â &r)+6   )   â x**€€€€;´Ž â &r=      âx€€€€´Ž â $x''   
Ž ä &r)*=   )   â x€€€€4´Ž â &r<      âx''€€€€´Ž â &r<   )   Ä &r2      ä&r2'2      Æ r   ÿ� ä r22   ÿ� ä x$ÿ   $  ä x3ÿ   3  ä xÿ     ä xÿ     Ä x%%   ÿàÿ â $r$   Ž ä $r32   Ž â | %   p`ð ä r   ÿ� ä r   ÿ� â 0rÿ    A   â xÿ     â 0rÿ    A   â xÿ     â 0rÿ    A   â xˆ  ÿàó Ä  r     A  â Er        â  r     A  â Er        â $r!ÿÿ   !Ž  â x    ÿàÿ â #r#   #   ä #r      â D	       À â Gy  €êÿÿÿÿƒ ê Ay        € ê ‰#      ( ‰      h y      %   ¢ !r#       â !r       È/ ‰      " $x   Ž ÆO ‰      d z  p  p`ð È x    pDp  â !r       â !r       È/ ‰ €    ( ‰ €    b !r       â !r       È/ ‰ @    ( ‰	 @    b !r       â !r		       È/ ‰      h  ‰	      ¢  M	        € ê y      &   â $x   ÿ Ž â !r	       âO $z   s  Ž â !r       â? $tÿ   ÿ Ž â x    p"ò Ä r ÿ   p"ð Ö $ÿÿ    Ž ä $z  v   Ž È� %v  j   Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      "   $ z  w  p`ð Ú M	        € ê y      &   " $t ÿ   ÿ Ž â ¹z  F   
   â $z x  Ž È %v \    Ž Ê �y    ™ b ¹z  l      â Ey  p,    € â ™x?    â y      !   b $r
ÿÿ   ÿ Ž â x      ÿ  â $rÿÿ   ÿ Ž ä $rÿÿ   ÿ Ž ä $rÿÿ   ÿ Ž â xÿ     È/ |    p`ð Ú G	  À+    € ê y      %   " $rÿÿ   ÿ Ž â Ey p+    € â $z r  ÿŽ ä 'z m   Ž È $x   ÿ Ž âzÿ n    â $r	ÿÿ   ÿ Ž â x   ÿÀŽ â $vÿ X  ÿ Ž â x   ÿÀŽ â $rÿÿ   
Ž ä $vÿ Y  ÿ Ž ä $z o  Ž Ä $z t  Ž ä$z u  ÿŽ ä $x   ÿ Ž ä $vÿ `  ÿ Ž â r   ÿàó â $vÿ a  ÿ Ž â x   ÿàÿ â $rÿÿ   ÿ Ž Ä $x   ÿ Ž ä $z q  Ž ä $z q  Ž ä %xb   Ž È $z t  Ž ä %xb   ÿŽ È %xb   ÿŽ È %x
    Ž â r   ÿàñ Æ %x    Ž È $rÿÿ     ä $rÿÿ   ÿŽ  ä r   ÿàñ Ê $x	     Ê �y#   • ¨ �y*   • è �y+   • ( �y.   • b r
   ÿàñ Ê $x     Ê �y;   • ( �y<   • h �y1   • h �y0   • b r(   ÿàñ Ê $x)     Ê �y%(B   • h �y&(D   • b x/ÿ     â $t5ÿ$   ÿ Ž ä $x#   ÿ Ž äO $x*   ÿ Ž Æ� xü  ÿÀŽ ä xü  ÿÀŽ ä z   ÿàñ ä z"  ÿàó ä xÿ   #  â $v!ÿ  ÿ  â r   ÿàñ â $v#ÿ  ÿŽ  Ä %v(    Ž â/�y'     ™ ¢  xÿ   *  ä $x*+   ÿ Ž â�y$"    ™ â $x,.   ÿ Ž ä$x     â �y((    ™ b	 %v    Ž â x**ü  ÿÀŽ Ä x-ÿ   +  â �y B   • â x+,ü  ÿÀŽ ä z**  ÿàñ â �y!    ™ â  z"+  ÿàó â/ %v,-    Ž È $v+ÿ  ÿ  â �y),    ™ "$v#ÿ  ÿŽ  â r4   à‘ ä �y2*    ™ $ r/ÿ   /$~  ä �y3"    ™ " %v4 Z  5 Ž Æ �y5D   • " $x7/$   ÿŽ â x/ÿ   .  Æ $x   7Ž ä %v*/    Ž Æ/ �y4   ™ ( �y:*    ™ ( �y6   ™ ( �y7   ™ ( �y8   ™ " $x";   ÿ Ž Æ �y9   ™ " $x<   ÿ Ž â x""ü  ÿÀŽ ä x#ÿ   <  â �y?   ™ " z."  ÿàñ ä xü  ÿÀŽ â �y@   ™ $ $v/ÿ  ÿ  â z  ÿàñ Ä x+ÿ   ;  â $x"1   ÿ Ž â �y;.    ™ " $vÿ  ÿ  ä %v*+    Ž âx,"ü  ÿÀŽ ä �y=    ™ " %v"#    Ž â z,,  ÿàñ Ä �y<*    ™ " $x.0   ÿ Ž Æ �y>"    ™ $ x..ü  ÿÀŽ â $v-ÿ  ÿ  â x1ÿ   1  ä z..  ÿàñ ä �y,,    ™ " %v*1    Ž â/ x1ÿ   0  Æ $v/ÿ  ÿ  ä �y**    ™ " %v01    Ž Æ �y..    ™ ( �y00    ™ ( �y-    ™ " r   ÿàñ är"   ÿàó â �y+    • $! $x     Ä $x#   Ž  Æ �y    • ( �y""    • " xA%ÿ   ÿÀŽ Ì 	sA A        " x/ÿ   %  È x//ÿ   ÿÀŽ â $x1A€   ÿ Ž Ê xB1€   ÿÀŽ ä xCÿ   &  äxBBÿ   %xŽ â 	s1 /        " x%%T  &    ä xCÿ   ÿÀŽ ä xÿ   %  Æ/ 	s#         b xÿ   ÿÀŽ â $xBÿŽ ä $tBÿÿ Ž ä $tDÿ @€ÿ Ž ä 	s%         b $x11€   ÿ Ž â xA€€€€BìŽ Ä xC€€€€DìŽ ä xBAÿþþþÿàÿ ä x1/€   1xŽ â $x/#€   ÿ Ž â/ xDCÿþþþÿàÿ ä xBBøŽ â$x#1ÿŽ â xD @€øŽ â $t1ÿÿ Ž Ä $xD%€   ÿ Ž â x%€   /xŽ â $tAÿ @€ÿ Ž â x1#€€€€1ìŽ â $tÿÿ Ž â x/€   DxŽ â $x%ÿŽ â xA#€€€€AìŽ â $tDÿ @€ÿ Ž â x11ÿþþþÿàÿ ä x€€€€ìŽ Ä xCAÿþþþÿàÿ ä x%€€€€DìŽ â $x//ÿŽ â xA1#øŽ â $t1ÿÿ Ž â xÿþþþÿàÿ ä x%%ÿþþþÿàÿ ä xBB˜º  ÿ    Ä x1/€€€€1ìŽ äxD/€€€€DìŽ ä xøŽ äx% @€øŽ ä x%BÿÀŽ ä x11ÿþþþÿàÿ Ä xDDÿþþþÿàÿ ä x˜º  ÿ    ä xAA˜º  ÿ    ä x#C @€#øŽ ä x11/øŽ äxDD @€/øŽ Ä x/ÿÀŽ ä x##˜º  ÿ    ä x˜º  ÿ    ä xBB€€€€'ÞŽ ÊO $xCB   %
Ž â x%AÿÀŽ ä xB€€€€(ÞŽ äxAA€€€€$ÞŽ Æ� $xB   /
Ž â x/ÿ      â $xAA   %
Ž â x%#ÿÀŽ äxB#€€€€!ÞŽ ä x#/ÿ   ÿÀŽ ä x€€€€(´Ž â $x(B   %
Ž â x'C€€€€'´Ž â 	s% #        " x/1˜º  ÿ    Ä xBD˜º  ÿ    ä xC˜º  ÿ    ä x$A€€€€$´Ž ä x1/ÿÀŽ äxBÿÀŽ ä xD/€€€€2ÞŽ ÄxAB€€€€)ÞŽ ä x/CÿÀŽ äxBC€€€€3ÞŽ ä x!(€€€€!´Ž ä x( ÿ   ÿÀŽ â $xBB   /
Ž ä $xA   
Ž Æ x3B€€€€3´Ž â 	s( (        b $xB%€   ÿ Ž ä $x1D   1
Ž â x€€€€)´Ž ä x/ÿ   5  äx% T  5    ä x)#€   BxŽ â &r''4   ÿ   â xBÿÀŽ Ä x21€€€€2´Ž ä x€€€€:ÞŽ ä x#/ÿ   ÿÀŽ ä x1ÿ   %  â &r/6   '   â 	s% #        " $x   B
Ž â x1ÿ   ÿÀŽ â $tBÿÿ Ž Ä $x))ÿŽ ä $tDÿ @€ÿ Ž â 	s'         ¢ $x((€   ÿ Ž â/ x1)€€€€BìŽ â&r$$7   /   â xA)€€€€DìŽ ä x/(€   ÿÀŽ â &r!!8   $   â x$1ÿþþþÿàÿ Ä x(Aÿþþþÿàÿ ä x//ÿ    xŽ ä x $)øŽ äx)( @€)øŽ â $x(%€   ÿ Ž ä $x%/ÿŽ ä $x$'€   ÿ Ž âO x'#€   (xŽ â &r229   !   â x!%€€€€BìŽ Ä x#%€€€€DìŽ ä x€   $xŽ ä x$!ÿþþþÿàÿ â $x!'ÿŽ â x(#ÿþþþÿàÿ ä x$$%øŽ äx#( @€%øŽ â $x%ÿŽ â x!€€€€BìŽ â$t(ÿÿ Ž â x'!€€€€DìŽ â $t/ÿ @€ÿ Ž â xBÿþþþÿàÿ Ä xD'ÿþþþÿàÿ ä x'%€€€€(ìŽ äx  ˜º  ÿ    ä x(%€€€€/ìŽ ä xB!øŽ äx!D @€!øŽ Ä xD'ÿþþþÿàÿ ä xB ÿÀŽ äx' €€€€;ÞŽ ä x/(ÿþþþÿàÿ ä x($˜º  ÿ    ä x$D%øŽ â$xD'   B
Ž â x / @€%øŽ Ä x))˜º  ÿ    ä xB(ÿÀŽ äx%(€€€€=ÞŽ ä x##˜º  ÿ    ä x')ÿÀŽ â $xB%   B
Ž â x))€€€€<ÞŽ Ä x(D€€€€;´Ž ä xD#ÿÀŽ äx%#€€€€>ÞŽ â $x')   '
Ž â x)$˜º  ÿ    ä x=B€€€€=´Ž â $x%%   D
Ž â x$)ÿÀŽ Ä x#'€€€€<´Ž â &r==4   ÿ   â x))€€€€,ÞŽ ä x%%€€€€>´Ž ä x' ˜º  ÿ    ä x˜º  ÿ    â $x))   $
Ž â x 'ÿÀŽ â &r%%6   =   â x''€€€€*ÞŽ Ä x$ÿÀŽ äx€€€€.ÞŽ â &r%(7   %   â x!!˜º  ÿ    â $x''    
Ž â x))€€€€,´Ž â $x   $
Ž â x$!ÿÀŽ â &r #8   %   â x!!€€€€0ÞŽ Ä x*'€€€€*´Ž â &r )9       â x.€€€€.´Ž â $x!!   $
Ž ä &r?   2   ä&r*?       â x:€€€€:´Ž â &r3@      âx0!€€€€0´Ž â &r..@      Ä &r:-      ä&r.0-   .   Æ r   ÿ� ä r..   ÿ� ä z ÿ `  pPð ä x&ÿ   &  ä x5ÿ   5  ä xÿ     Ä xÿ     ä z ÿ a   Sð â $r&   Ž ä $r5.   Ž Æ r   ÿ� ä r   ÿ� â 0rÿ    A   â xÿ     â 0r+ÿ+    A   â xÿ     â 0r"ÿ"    A   â Er)        ä  r'+     A  â Er&        â  r(+"     A  â G‰  à	    € ê r   ÿàó Ê $x	   Ž  Ê �y     ¨ �y"    è �y3    ( �y2    b r   ÿàó Ê $x   Ž  Ê �y*B    h	 �y/D    b	 $x!    ÿ Ž âO x#ÿ      È x!!ü  ÿÀŽ È z$!  ÿàó â $x!"   ÿ Ž ä� $x3   ÿ Ž ä$v%ÿ  ÿŽ  â x!!ü  ÿÀŽ È z!  ÿàó â �y,$    ™ ¢
 xÿ   "  â %v #    Ž â x"ü  ÿÀŽ Æ $vÿ  ÿŽ  â z""  ÿàó â %v    Ž â x3ÿ   3  â �y.     ™ ä  $x$2   ÿ Ž ä�y0    ™ " $v#ÿ  ÿŽ  Æ �y1    ™ â
 x$$ü  ÿÀŽ Æ �y""    ™ â
 z$$  ÿàó â %v 3    Ž â xÿ   2  Æ/ $v%ÿ  ÿŽ  ä �y      ™ â  %v    Ž Æ �y$$    ™ è �y    ™ â  r   ÿàó Ê$x   Ž  Ê �y     â
 x#ÿ   *  â $t:ÿÿ Ž â x!*ÿ   ÿÀŽ â$t<ÿ @€ÿ Ž â x2#ÿ   ÿÀŽ â $t>ÿ @€ÿ Ž â x%ÿ   /  â/$t;ÿ @€ÿ Ž â 	s# 2        " x*T  /    Ä x/ÿ   /  ä xÿ     Æ 	s! !        b xÿ   ÿÀŽ â $x##€   ÿ Ž Ì 	s         âx32€   #xŽ ä x#%ÿ   ÿÀŽ â $x!!€   ÿ Ž ä/ $x33ÿŽ ä 	s% #        " x!!€   ÿÀŽ ä x23€€€€:ìŽ Ä x53€€€€<ìŽ ä x:2ÿþþþÿàÿ ä x<5ÿþþþÿàÿ â $t5ÿÿ Ž â x2!ÿ   *xŽ ä x!:3øŽ äx*< @€3øŽ â $x:%€   ÿ Ž â x!!˜º  ÿ    â $x%2ÿŽ Ä $t<ÿÿ Ž â x3#€   :xŽ â $x2€   ÿ Ž â x#%€€€€>ìŽ äx%€€€€<ìŽ ä x€   2xŽ ä x2ÿþþþÿàÿ â $x3ÿŽ â x:#ÿþþþÿàÿ Ä x#2%øŽ â $x2ÿŽ â x3€€€€>ìŽ äx€€€€<ìŽ ä x<3ÿþþþÿàÿ ä x32€€€€5ìŽ äx52€€€€;ìŽ Ä x;*˜º  ÿ    ä xA5ÿþþþÿàÿ ä x5!ÿÀŽ ä x=3ÿþþþÿàÿ ä x%: @€%øŽ ä x3#˜º  ÿ    Ä x#=2øŽ ä x:ÿþþþÿàÿ È x:øŽ äx:;ÿÀŽ ä x˜º  ÿ    ä x< @€øŽ È x˜º  ÿ    ä x*!€€€€,ÞŽ äO x!A @€2øŽ ä x2%˜º  ÿ    â $x5*   5
Ž â x*3ÿÀŽ ä x%2ÿÀŽ Ä x,5€€€€,´Ž ä x5#˜º  ÿ    ä x33€€€€0ÞŽ äx;;€€€€.ÞŽ ä� x22€€€€1ÞŽ â $x33   *
Ž â x*5ÿÀŽ â $x;;   :
Ž â x55€€€€"ÞŽ â $x22   %
Ž â x33€€€€0´Ž Ä x!!˜º  ÿ    â $x55   *
Ž â x12€€€€1´Ž â &r334   ÿ   â x*!ÿÀŽ ä x55€€€€"´Ž â &r116   3   â x!!€€€€ ÞŽ ä x"ÿÀŽ â&r,,7   1   â x€€€€$ÞŽ â $x!!   *
Ž â x#;€€€€.´Ž Æ $x   "
Ž â x"ÿÀŽ â &r,#8   ,   â x€€€€ÞŽ ä x !€€€€ ´Ž â &r,59   ,   â x$€€€€$´Ž â $x   "
Ž ä &r ?   ,   Ä 0rÿ    A   â x€€€€´Ž â &r$@      È &r-      Ê r   ÿ� È xÿ     Ê $r/   Ž Ê r   ÿ� È xÿ     È Er        â  r+     A  È #r      â #r')      â #r(&      â G‰   
    € ê r 
   ÿàñ Ê $x!     Ê �y#     ¨ �y$     è �y*     ( �y,     b r   ÿàñ Ê $x     Ê �y/B    (	 �y&D    "	 $x#   ÿ Ž äO $x"$   ÿ Ž Æ� xü  ÿÀŽ È z  ÿàñ Ê $vÿ  ÿ  â x!ÿ   #  È �y(    ™ ¢
 %v !    Ž â x""ü  ÿÀŽ È �y)     ™ â  z""  ÿàñ ä x%ÿ   $  â $x*   ÿ Ž ä$v#ÿ  ÿ  ä %v$%    Ž â xü  ÿÀŽ ä �y'"    ™ " xÿ   *  â $x,   ÿ Ž Ä�y*$    ™ " z  ÿàñ â %v    Ž â x"ü  ÿÀŽ Æ/ $vÿ  ÿ  â x#ÿ   ,  â �y    ™ " z "  ÿàñ Æ �y    ™ $ $v!ÿ  ÿ  ä %v"#    Ž Æ �y      ™ ( �y""    ™ " r$   ÿàñ Ê$x%     Ê �y$$     " x,/ÿ   ÿÀŽ ä xÿ   /  È/ 	s, ,        b xÿ   ÿÀŽ ä xÿ   &  Ê 	s!         " x1ÿ   ÿÀŽ Î 	s# 1        b $x,€   ÿ Ž Ê/ x.€   ÿÀŽ â $x0!€   ÿ Ž Æ x..ÿ   /xŽ ä x//T  &    ä x0€   0xŽ ä x/ÿ   /  â $x2#€   ÿ Ž ä$tÿÿ Ž ä $x..ÿŽ â x/ÿ   ÿÀŽ â $x00ÿŽ â x%1€   2xŽ â $t/ÿ @€ÿ Ž â x.€€€€ìŽ â $t1ÿÿ Ž Ä $t#ÿ @€ÿ Ž â x/0€€€€/ìŽ â	s!         " xÿþþþÿàÿ ä x,0€€€€1ìŽ ä x#.€€€€#ìŽ ä x5/ÿþþþÿàÿ ä x/.øŽ Ä x3,ÿþþþÿàÿ ä x1#ÿþþþÿàÿ ä x//˜º  ÿ    â $x%%ÿŽ â x#30øŽ â$t2ÿ @€ÿ Ž â x5 @€0øŽ â $t0ÿÿ Ž â x,1 @€.øŽ Ä x./ÿÀŽ ä x1%€€€€0ìŽ äx2%€€€€2ìŽ â $x0!€   ÿ Ž â x,,˜º  ÿ    ä x22ÿþþþÿàÿ ä x0€   0xŽ Ä x2 @€%øŽ â $t3ÿ @€ÿ Ž ä $x00ÿŽ â x˜º  ÿ    ä x˜º  ÿ    ä x&ÿ   &  ä x//€€€€(ÞŽ ÊO $x//   .
Ž â x.1ÿþþþÿàÿ È x!.%øŽ ä x%,ÿÀŽ äx,,€€€€)ÞŽ â� $t1ÿÿ Ž â x//€€€€(´Ž Æ $x,,   %
Ž â x(#˜º  ÿ    ä x.0€€€€3ìŽ äx%0€€€€1ìŽ â &r#/4   ÿ   â x,,€€€€)´Ž ä x/(ÿÀŽ ä x).ÿþþþÿàÿ Ä x((€€€€'ÞŽ äx%%ÿþþþÿàÿ ä x.ÿÀŽ äx€€€€*ÞŽ â $x((   /
Ž â x%%0øŽ Æ $x   .
Ž â x)) @€0øŽ ä x%%˜º  ÿ    â &r,,6   #   â x((€€€€'´Ž ä x€€€€*´Ž ä x))˜º  ÿ    ä x*%ÿÀŽ Ä x%%€€€€ÞŽ â &r((7   ,   â x,)ÿÀŽ ä x!!˜º  ÿ    â $x%%   *
Ž â x))€€€€ÞŽ ä x*!ÿÀŽ äx!!€€€€ ÞŽ â $x))   ,
Ž â x%%€€€€´Ž â &r(8   (   â xÿÀŽ â $x!!   *
Ž â x€€€€"ÞŽ â &r%%9   (   â x)€€€€´Ž Ä x!!€€€€ ´Ž â $x   
Ž ä &r%?   %   Æ x"€€€€"´Ž â &r!!@   %   È &r!"-   !   Ê r!!   ÿ� È xÿ     Ê $r&!   Ž Ê r   ÿ� â 0r$ÿ$    A   Æ xÿ     ä  r$+$     A  ä Er        Ê #r$      â x   ÿàÿ ä xˆ  ÿàó ä |    p`ð ä xˆ  ÿàõ â $rÿÿ   Ž  â x    ÿàÿ Æ $rÿÿ    Ì D	       À â Gy  �Öÿÿÿÿƒ ê Ay       € ê $r
ÿÿ    Ž ä $rÿÿ    Ž ä Ay        € ê z ÿ `  pPð È z ÿ a   Sð Ú G	  P    € ê ‰      ( ‰      b !r       â !r       È/ ‰      ( ‰      b !r       â !r       È/ ‰ €    ( ‰ €    b !r       â !r       È/ ‰ @    ( ‰	 @    b !r       â !r		       È/ ‰      ( ‰	      b !r       â !r	       â/ Gy  €    € ê ‰      ( ‰      h ‰      ¨ ‰      â !r       â !r       â/ !r       ÈO ‰      " !r       Æ� ‰      h ‰      ¨ ‰
      â !r       â !r       â/ !r	       ÈO ‰ €    " !r
       Æ� ‰ €    h ‰
	 €    ¨ ‰ €    â !r       â !r       È/ ‰ @    " !r
	
       ÆO ‰ @    b !r       Æ� ‰
 @    ¨ ‰ @    â !r       â !r       â/ !r
       ÈO ‰	      " !r       Æ� ‰      h ‰      ¨ ‰      â !r
	       â !r       â/ !r       âO !r       â� y      !   ( y      %   $ $x   Ž Ê z  p  p`ð È x    pDp  Ú M	        € ê z ÿ ^  pPð â $z v  Ž Æz ÿ _   Sð Ú % ^    Ž Ì �	     ¢ $x   ÿ Ž Ê r ÿ   p"ò äx    p"ö äx    p"ø äx    p"ô Î $ÿÿ    Ž â z ÿ `  pPò â $2ÿÿ    Ž ä $Bÿÿ   
 Ž â z ÿ a  Sò â $"ÿÿ    Ž È !       ÐO G™      € ê z ÿ b  pPð È z ÿ c   Sð Ú % b    Ž Ì �	     ¢ $x   ÿ Ž ä $vÿ h  ÿ Ž Æ x øÿÿÿp"ò äx üÿÿÿp"ô är ÿ   p"ö äx    p"ø Î $ÿÿ    Ž â x    pBò â $"ÿÿ    Ž ä $2ÿÿ   
 Ž ä $Bÿÿ    Ž È !       ÈO G  P    € ê x    pRð Ú G‰  �     € ê x    pRð Ú G	  ð    € ê  x;ª¸¿  A  Ì s        $ !t  €?    Ì s        $  r     A  È  r     A  â Gy  P    € ê  x'7=  A  â x*BL?  A  Æ #t  €?   â  x   ?  A  Æ  r     A  Ì s     $   $ !t  €?    È  r     A  È  r     A  â Gy  °    € ê x    pRð Ú G‰  à     € ê x    pRð Ú G	        € ê 	z i    � ä 	z i    � Æ  x;ª¸¿  A  â 	z i €  � Ê s        $ !t  €?    Ì s        $  r     A  È  r     A  â Gy  Ð     € ê  r     A  â Gy  °     € ê 	x  à@  � ä 	x  à@  � Æ  x#ÛÙ?  A  â 	x  àÀ  � Æ  x;ª¸¿  A  ä !t  €?    È s        $ !t  €?    Ì s        $  r     A  È  r     A  ä y      "   ( y      &   b $z s  Ž È $z v  Ž È/ %v j    Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y5      "   " ¹z  l      â Ey  P/    € â ™x?    â y6      !   b ¹z  F   
   â r9 ÿ       â $r7ÿÿ   ÿ Ž â r: ÿ       â $rIÿÿ   ÿ Ž â r  ÿ       â x*      ÿ  â x,      ÿ  â x.      ÿ  â x0      ÿ  â x2      ÿ  â $r4ÿÿ   ÿ Ž â r 5ÿ   pRð Ä r 5ÿ   pRò â$x5    6Ž Ê/ xÿ     È r   ÿ� È x8ÿ     È | 8   pbô Ú G)  à-    € ê y      '   " r ÿ       â $xF8   ÿ Ž â r ÿ       â y	      %   b xøÿÿÿÿÀŽ â $v?ÿ Y  ÿ Ž â zC  q      â y      &   ¢ zA  X      â $x   
Ž â r@CF   ÿŽ â$x;C   FŽ âr9 ÿ       â $x>C   FŽ âxÿ     â $x<C   FŽ âzEF q  ÿàÿ â $xBC   FŽ Ä $xCC   FŽ ä 'z y   Ž È $x		   ÿ Ž â/ zÿ z    â $z= }  ÿŽ ä 'z s   Ž ÈO $z	 p  8Ž â zÿ t    â $z |  8Ž ä$z |  Ž â x	   ÿàÿ â $z v  Ž ä$z v  Ž ä $z p  Ž Ä %x
R   ÿŽ È $z w  ÿŽ â rD
   ÿà÷ â %xR   ÿŽ Æ r==   ÿàõ â $x	   ÿ Ž ä %xR   Ž â rGÿÿ   ÿä Æ %x	   
Ž È $xH   Ž ä %x	   Ž È rTA   ÿàõ È rU?   ÿä ä rA   ÿàõ Æ �yTB   ‘ ¤  $x	   ? Ê �y   • â rXAD   ÿàõ Æ �y   • " rY?H   ÿä ä rA   ÿàõ â �y‰"   • h �yXB   ‘ b  $x   ? Æ �yŽ$   • ( �y   • ( �y   • " $t&ÿ   ÿ Ž Æ �yy"   • ( �yQ$   • " $taÿ$   ÿ Ž Æ �yXXJ   ‘ b rl=   Eàë Æ �yTTJ   ‘ b r=   Fàµ â %vll Z  a Ž â x   ÿ   ÈO x   ÿÀŽ È xÿ   øŽ ä�xÿ     Æ %v$ 
 & Ž â x   ÿ   È x   øŽ â �y$$    › ¨ %v 
 & Ž Ì �y    › â $x   ÿ Ž Ê/ x   ÿÀŽ È xÿ   øŽ Ê%v  
 & Ž â x   ÿ   ä xÿ     Æ �y      › " x   øŽ Ê %v" 
 & Ž Ì �y""    › ¢ $x   ÿ Ž Ê x   ÿÀŽ È xÿ   øŽ äx   ÿ   ä xÿ     â %v 
 & Ž È $x   ÿ Ž â x   øŽ â �y    › ¦ x   ÿÀŽ â %v 
 & Ž Æ xÿ   øŽ Æ �y    › ¤ %v 
 & Ž â x   ÿ   ä xÿ     ä rG   ÿd~ â �y    › ¢ x   øŽ â %v Z  a Ž È $x$   ÿŽ ä %v 
 & Ž Æ r   ÿàÿ Æ �y    › ¨ �yP   ™ ¨ �yL   ™ ¨ �y_   ™ ¨ �y]   ™ ¨ �yO   ™ ¨ �yM   ™ ¨ �y^   ™ ¨ �yN    ™ ¢ rG   ÿÄþ Ä r=   ;àÇ â �yK    • ¤ $x'$   ÿŽ â rG   ÿ„þ ä rA   ÿàû â $xmm   'Ž ä $x)$   ÿŽ â r?   ÿäÿ Ä �yVl   ™ ¢  r'=   @àÙ ä r
A   ÿàõ â �yZl   ™ ¢  rcG   ÿ¤~ Æ �yJl   ™ ¢  rW=   BàË â$x   ? â r(=   >àµ â �y[l   ™ ¨
 �y\l   ™ â
 %vW Z  a Ž â/ rdG   ÿd~ Ä �yUl   ™ â r&=   <àµ Æ �yWl   ™ â  %v Z  a Ž â rbG   ÿd~ ä �yYl    ™ è  �yS    • â r)   ÿàÿ â %v(( Z  a Ž Ä �yR    • ä  $xed$   ÿŽ ä $xcc$   ÿŽ â �yh   ™ " r=   Càµ â/ $xgb$   ÿŽ ä �yf   ™ " %v' Z  a Ž Æ �yd    ™ " %v&& Z  a Ž âxŠ‰ÿ   ÿÀŽ ä�yk    • b %v Z  a Ž â raG   ÿ„þ Æ $x))   eŽ â r`G   ÿd~ â $xaa$   ÿŽ â �ye   ™ " rc   ÿàÿ Æ ra   ÿàÿ â �yc   ™ " $x''   gŽ Æ �ya   ™ ( �yg   ™ " $xi`$   ÿŽ Æ �y`   ™ ( �ybl    • ( �yn(   ™ ( �yj(   ™ ( �yl(   ™ h �yp(   ™ b $x   iŽ Æ �yq(   ™ h �yi(   ™ h �yo(   ™ h �ym(    ™ b r�‰Š   ÿpŽ Ä x�ÿ   ‰  â �ys   ™ d $x‹�€   ÿ Ž ä �yv   ™ b r”��   ÿpŽ Æ �y   ™ b# x‘‹€€ÿÀŽ Æ �yw   ™ b x’”   ÿ   â $x��   ÿ Ž ä �yx   ™ b x�‘€€€€ÿüŽ Æ �y   ™ b x’’€€ÿÀŽ Æ �yt   ™ b x��ÿþþþÿàÿ Æ �yu    ™ b x��€€ÿÀŽ ä x“’€€€€ÿüŽ â �yr(    • b r‘�‘   ÿüŽ â $x””   ÿ Ž â x��€€€€ÿüŽ â �y{    • b x—Žÿ   ÿÀŽ ä x•“ÿþþþÿàÿ â �y}&   ™ b x��ÿþþþÿàÿ Ä r—Ž—   ÿpŽ â �y)&   ™ b x””€€ÿÀŽ Æ �y€&   ™ h �y~&   ™ h �y(&   ™ h �yz&   ™ h �y&   ™ h �y|&    ™ h �y‡&    • h �y   ™ h# �y†   ™ h �y‚   ™ h �y„   ™ h �y�   ™ h �y…   ™ h �yƒ   ™ h �y    ™ h �yˆ    • h �y&   ™ h �yŒ   ™ h �yŠ   ™ h �y‰   ™ h �y'   ™ h �y‹   ™ h �y   ™ h! �y    ™ h �y�    • b r“��   ÿüŽ â $x�—€   ÿ Ž â r•’   ÿüŽ Ä x•ÿ   Ž  ä x”€€€€ÿüŽ ä x’—   ÿ   â $x•• @  •Ž â x�ÿþþþÿàÿ ä x��€€ÿÀŽ â $x˜•   ÿ Ž â xŽ’€€ÿÀŽ Ä r’�”   ÿüŽ ä x�€€€€ÿüŽ ä x”Ž€€€€ÿüŽ ä x˜˜€€ÿÀŽ ä x–•   ÿ   ä x�ÿþþþÿàÿ Ä x•”ÿþþþÿàÿ ä x”˜€€€€ÿüŽ ä x‘‘˜º  ÿ    ä r���   ÿüŽ ä r�•Ž   ÿüŽ ä x•”ÿþþþÿàÿ Ä x”‘ÿÀŽ ä x‘$€€€€‘ÞŽ äO x“˜º  ÿ    ä r‘‘”  €ÿàÿ ä x”yÿ   ÿÀŽ ä x—–€€ÿÀŽ Ä x˜º  ÿ    ä x–ÿÀŽ ä x“%€€€€ÞŽ ä r•˜   ÿüŽ ä x’’˜º  ÿ    ä r•y”   ÿpŽ Ä x”ÿÀŽ ä x$‘€€€€$´Ž ä x€€€€ÞŽ ä� x‘’ÿÀŽ ä x’€€€€’ÞŽ â $x–“   –
Ž ä $x“   ”
Ž â x•   ÿ   â$x”’   ‘
Ž â x‘•   ÿ   Ä x’ÿ   y  ä xy€€ÿÀŽ ä x‘‘€€ÿÀŽ â $x•’ @  ’Ž â x“€€€€´Ž ä xy€€€€ÿüŽ ä x’‘€€€€ÿüŽ Ä x”€€€€´Ž â $x”•   ÿ Ž âxÿþþþÿàÿ ä x“•   ÿ   ä x’’ÿþþþÿàÿ ä ry   ÿüŽ ä x•”€€ÿÀŽ Ä x““€€ÿÀŽ ä r’’‘   ÿüŽ ä xy˜º  ÿ    ä x‘•€€€€ÿüŽ ä x”“€€€€ÿüŽ ä x’’˜º  ÿ    Ä x%–€€€€%´Ž ä xyÿÀŽ ä xy €€€€yÞŽ äx–‘ÿþþþÿàÿ ä x””ÿþþþÿàÿ ä x‘’ÿÀŽ Ä x’!€€€€’ÞŽ ä ryy  €ÿàÿ ä r””“   ÿüŽ â $x’’   ‘
Ž â x y€€€€ ´Ž ä xy”˜º  ÿ    ä x!’€€€€!´Ž Ä x’Qÿ   ÿÀŽ ä xyÿÀŽ ä xy"€€€€yÞŽ ä r––•   ÿüŽ ä r“Q’   ÿpŽ ä x’ÿ   Q  â $xQy   
Ž â x––˜º  ÿ    â $x“€   ÿ Ž â r’’’   ÿpŽ Ä x‘–ÿÀŽ ä xQQ€€€€"´Ž â $x"“   ÿ Ž â x–#€€€€–ÞŽ ä x€€ÿÀŽ ä r––‘  €ÿàÿ ä xy’   ÿ   Ä x‘"€€ÿÀŽ ä x"€€€€ÿüŽ â $x”’   ÿ Ž â x“y€€ÿÀŽ ä x’‘€€€€ÿüŽ ä xy"ÿþþþÿàÿ ä x””€€ÿÀŽ Ä x’’ÿþþþÿàÿ ä ryy   ÿüŽ ä x"“€€€€ÿüŽ ä x”€€€€ÿüŽ ä r’’‘   ÿüŽ ä xyy˜º  ÿ    Ä x""ÿþþþÿàÿ ä x‘ÿþþþÿàÿ ä x’’˜º  ÿ    ä xyÿÀŽ ä xy€€€€yÞŽ ä r""“   ÿüŽ Ä r”‘”   ÿüŽ ä x‘’ÿÀŽ ä x’€€€€’ÞŽ ä ryy  €ÿàÿ ä x“"˜º  ÿ    ä x””˜º  ÿ    â $x’   ‘
Ž â xyy€€€€´Ž Ä x“ÿÀŽ ä x‘”ÿÀŽ ä x“€€€€“ÞŽ ä x”€€€€”ÞŽ ä x��˜º  ÿ    ä x��˜º  ÿ    Ä r““  €ÿàÿ â $x””   ‘
Ž â x�ÿÀŽ ä x�€€€€�ÞŽ ä xŽ—€€€€ÿüŽ ä x‘�ÿÀŽ ä x�€€€€�ÞŽ Ä x€€€€´Ž ä r�  €ÿàÿ ä xŽŽÿþþþÿàÿ â $x��   ‘
Ž â x€€€€´Ž ä rŽŽ—   ÿüŽ ä x˜º  ÿ    Ä x�€€€€´Ž ä x�Ž˜º  ÿ    ä x"ÿÀŽ ä x€€€€ÞŽ â &r P   ÿ   â x�ÿÀŽ ä x�€€€€�ÞŽ â $x"   "
Ž Ä &ryL   ÿ   ä&r!_      â x“€€€€´Ž â &rLL   ÿ   â r��  €ÿàÿ â &r]      â x#–€€€€#´Ž â &rQO      â x”€€€€´Ž â &rL]   L   â x�€€€€´Ž â &rM      Ä &r#^      â xXð   ÿÀŽ â &rN      â x"€€€€´Ž â &r"M   L   â xXX   ÿÀŽ â &rP$P   ÿ   â xÿ     â &rNN   "   â r"   ÿàÿ â $rMX   ÿŽ Ä &rP%_   P   â r"""   ÿ� â $rM   MŽ ä &rPO   P   â xLTð   ÿÀŽ äxTT   ÿÀŽ â &rO^   P   â r""M   ÿú� ä xÿ   L  â $rTO   ÿŽ â xMÿ   "  â $xOO   NŽ Ä &rL V   ÿ   â rMM"   ÿ� â &rV$V   ÿ   â rOOO   ÿ� â $rN   Ž ä &r"yZ   ÿ   ä &rL!J   L   ä &rZZ   ÿ   Ä &rV%J   V   â rO   ÿú� â &rLQ\   L   ä&rZ[   Z   ä&rV\   V   â xOÿ     â &r"[   "   ä &r[#W   L   Ä &rZU   Z   ä &rWW   V   â rOO   ÿ� â &r"U   "   ä &rZY   Z   ä $rTW   ÿŽ ä &r"Y   "   Ä $xWW   ZŽ ä $rUX[   ÿŽ ä $rZZ   Ž â xÿ   O  â 0rKÿK    A   ä 0rSÿS    A   â r[["   ÿàÿ â$rU"   UŽ â Er"        Ä  rKS     A  â 0rRÿR    A   â xMÿ   M  ä #r3"   3   â &r e   ÿ   â ErM M       â  rORK     A  â &r"yc   ÿ   â rWWW   ÿ� â &r!a      â #r4OM   4   â &rMg   "   â rJ[[   ÿ� â &rQh      â rWWZ   ÿú� â &r"f   M   â rJJU   ÿú� â &rM#`      â xÿ   W  â &rL$e   ÿ   â xKÿ   J  â &rcc   ÿ   â rW   ÿ� â &rL%a   L   â rKKJ   ÿ� â &r"d   "   â xÿ     â &rcg   c   Ä 0rbÿb    A   â rJM"   ÿàÿ â &rLh   L   â xKÿ   K  â &rcf   c   â Er        â $rOXM   ÿŽ â  rMRb     A  â rbSb     A  â &rU`   L   â ErK K       â &rdd   c   â rJJJ   ÿ� â $rO"   OŽ â #r1b   1   â $rLTU   ÿŽ Ä &r n   ÿ   â#r2MK   2   â $xUU   dŽ ä &rn$n   ÿ   â rJJO   ÿú� â &rMyl   ÿ   ä&r!j      ä &rOl   ÿ   â rUUU   ÿ� â $rLd   LŽ Ä &rn%j   n   ä &rMp   M   ä&rQq      ä&rpp   O   â rLUL   ÿú� â &rnq   n   ä &rO#i      Ä &rpo   p   ä&r"o   M   â xMÿ   L  â &rii   n   ä &rpm   p   â xKÿ   J  â $rTi   ÿŽ â rMML   ÿ� â &r"m   "   Ä $xii   pŽ â rKKJ   ÿ� â $rUXO   ÿŽ ä $rpp   Ž â xÿ   M  â 0rkÿk    A   â rOO"   ÿàÿ â$rU"   UŽ â xKÿ   K  ä Er"        â  rSk     A  â ErK K       â  rMRk     A  Ä #r/"   /   â &r s   ÿ   ä &r"yv   ÿ   ârJOO   ÿ� â &rL$s   ÿ   â riii   ÿ� â #r0MK   0   â &r!      ä &rvv   ÿ   Ä &rMw   "   â rJJU   ÿú� â &rL%   L   â riip   ÿú� â &rQx      ä &rww   v   ä &r"   M   â xKÿ   J  â &rLx   L   â xÿ   i  â &r#t      Ä &r   w   ä &r"u   "   â rKKJ   ÿ� â &rOt   L   â ri   ÿ� â &rJu      â r"   ÿàÿ â $rMX   ÿŽ â xÿ     â 0rrÿr    A   Ä $rTO   ÿŽ â xKÿ   K  â $xOO   JŽ â Er        â $rJJ   Ž â  rRr     A  âErK K       â  rrSr     A  È #r-r   -   â #r.K   .   â &r }   ÿ   ä &ry(   ÿ   ä &r!)      ä &rz      ä $rM"   MŽ Ä &r"$}   ÿ   â r   ÿ� â &rK(   ÿ   ä &rQ€      ä &r      ä &r"%)   "   â rM   ÿú� â &rKz   K   Ä &r)#~      ä &r|      â rOOO   ÿ� â &r   K   â xÿ     â &r"€   "   ä $rKX)   ÿŽ â r))   ÿàÿ ä rJOJ   ÿú� â &rM~   "   â r   ÿ� â $rK   KŽ â r))   ÿ� â &r||      â x)ÿ   J  â $rTM   ÿŽ â rKK   ÿú� â $xMM   |Ž â r))J   ÿ� â $r||   Ž â xÿ   K  â 0r{ÿ{    A   â xÿ   )  Ä rKK   ÿ� ä Er        â  rS{     A  â rMMM   ÿ� ä xÿ     â #r+   +   â &r    ÿ   ârMM|   ÿú� â &r$   ÿ   â Er        â &r!†      â r)R{     A  â 0r‡ÿ‡    A   â xKÿ   K  â &r†%†      â xÿ   M  â &ry�   ÿ   Ä &r��   ÿ   ä &r$$&   ÿ   ä&ryy'   ÿ   â #r,)   ,   â &r  &   ÿ   â ErK K       â &r'   ÿ   â rM   ÿ� â  rR‡     A  â &r…      Ä &r$%Œ   $   ä&ry‹   y   â xÿ     â &r�…   �   ä&r !Œ       ä &r‹      â #r*K   *   â &r†‚   †   Ä &rQ‚      ä &rƒ      ä&r$Š   $   ä &ry   y   â Er        â &r�ƒ   �   ä &r QŠ       Ä &r      â  r‡S‡     A  â &r)„   †   ä&r#„      ä &r      ä &r‰   $   ä &ry   y   Ä &r   �   ä &r##‰       ä &r      â #r ‡       â $rT)   ÿŽ ä $rX   ÿŽ ä $rTT   ÿŽ â r   ÿàÿ â $x))   Ž Ä $rXX#   ÿŽ â r##y   ÿàÿ â $x   Ž â r   ÿ� â $r   Ž âr)))   ÿ� â $r   Ž â r   ÿ� â $rXy   XŽ â r###   ÿ� â $rT   TŽ â r   ÿú� Ä r)   ÿú� ä rT   ÿú� ä r##X   ÿú� ä x88   ÿàÿ ä xÿ     ä xÿ     Ä xÿ     ä xÿ   #  ä | 8   pbô ä r   ÿ� ä r   ÿ� ä r   ÿ� Ä r##   ÿ� â 0rˆÿˆ    A   â xÿ     â 0r�ÿ�    A   â xÿ     ä xÿ     ä x#ÿ   #  â  rRˆ     A  âEr        â  rSˆ     A  âEr        â  rS�     A  âEr        â  rRR�     A  â Er# #       Ä xAA�  ÿà÷ â #rI   I   â #r:   :   â #r9   9   â #r7R#   7   â x;;@   ÿàÿ ä x>>@   ÿàÿ ä x@@@   ÿàÿ ä x<<@   ÿàÿ Ä xBB@   ÿàÿ ä xCC@   ÿàÿ ä xEE@   ÿàÿ ä xFF@   ÿàÿ ä r?ÿ?   ÿäÿ â D)       À â Gy  àÔÿÿÿÿƒ ê Ay        € ê Ey  0    € â G‰      € ê $x55   6Ž Ê ˆs 54 øÿ H   è ˆs 53€øÿ H   è ˆs 52 ùÿ H   è ˆs 51€ùÿ H   è ˆs 50 úÿ H   è ˆs 5/€úÿ H   è ˆs 5. ûÿ H   è ˆs 5-€ûÿ H   è ˆs 5, üÿ H   è ˆs 5+€üÿ H   è ˆs 5* ýÿ H   è ˆs 5 €ýÿ H   è ˆs 5I þÿ H   è ˆs 5:€þÿ H   è ˆs 57 ÿÿ H   è ˆs 59€ÿÿ H   ä Ay        € ê {          ì M        € ê „y6     H   b Ey  @    € æ „y6 €   H   ¨ „y6    H   â !r4       â/ !r33       È_ ‰      " !r22       Æ� ‰3      h ‰2      ¢ !r       Æ y      %   " !r3       Æ/ ‰      b !r	2       ÆO ‰      ¨ ‰	      â x   ÿ   â !r       Æ/ x   ÿàÿ äz  r  p`ð â !r       âO ‰ €    " z  r  p`ò â !r
	       â� r 6ÿ   pVp  â‰ €    b x 6   p"ò Æ ‰
 €    ° ‰      &   ä         Ä       &   " !r       Æ       '   " $„ÿ   ÿ Ž â !r       ä/ ‰ @    b !r
       ÆO ‰	 @    ¨ ‰ @    h ‰      '   b $ x  Ž Æ„y
6 €  H   â $ ~  Ž â !r       â/ $Š x  Ž â� !r		       ÆO ‰      " !r       Æ‰	      b $Š ~  Ž Æ ‰      ¢ !r       â %† j   Ž â !r	       Æ/ % j   Ž â †‰     â !r       ÆO †    â G	  �     € ê y      &   ( y      '   b $z x  Ž È $z ~  Ž Ê/ z r  ÿàõ Ê $rÿÿ   ÿ â z j  ÿ„ È z k   Ê †y     è Ay        € ê !r

1       â „y	6    H   â Ey  �    € æ ‰
      ( „y6 €  H   è „y6    H   â !r
       Ê ‰      $ !r       Ê ‰ €    $ !r       Ê ‰ @    $ !r       Ê ‰      $ !r       â G™  °     € ê y      &   " ¹z  r      ä �x   ?àÿ â y      '   b $z x  Ž È $z ~  Ž Ê/ |   ÿàõ È rÿÿ   ÿä ä z j  ÿ„ È z k   Ê †y     è Ay        € ê „y6 €  H   b !r/       â !r		0       â !r.       â „y6    H   ¢ z#  r      Æ „y
6 €  H   ä $x!#   ÿŽ ä „y6    H   ( „y6 €  H   h „y6    H   è „y6 €  H   è „y6    H   è ‰      ( ‰	      ( ‰      " !r--       Æ/ „y66 €  H   â !r,       âO !r+
+       ä� ‰
-      b !r*       â!r        â!r       Æ ‰      " !rII       â !r	       ä ‰      ¢ !r:       â !r       ä ‰      â !r77       Æ ‰+      ( ‰      b !r
-
       â/ $x-#   ÿŽ ä ‰I      h ‰      b !r       â !r669       Ä ‰7      " !r       ÆO ‰      ¢ !r       Æ� ‰      â !r++       Æy       &   " !r       Æy	      '   b !rII       â/ !r       ä ‰
      b !r77       Æ ‰      " !r       ÆO ‰+      ¢ !r       Æ� ‰      â $z   x  Ž Æ‰      " $z 	 ~   Ž Æ‰ €    h ‰I      b r#    ÿ„ â !r
       â/ ‚ -   ÿà÷ â ‰      b !r       â $rÿÿ   ÿ â z j  ÿ„ â ‰7      " !r++       âO $‚ÿÿ   ÿŽ â z k   â ‰6      ¢ !r       â� Š
 j  ÿ† Ä ‰ €    â !r       âŠ k  � ä ‰	 €    " !r       â‚ !   ÿàù â !rII       â ‰ €    d ‚(ÿÿ   ÿä â !r       â/ ‰ €    b Š j  ÿˆ â !r77       Ä ‰+ €    " Š k  ( â !r66       äO ‰ €    ¢ !r       Æ� ‰  €    â $x#   ÿŽ â!r	       ä‰ @    " r	#    ÿŠ är    ÿàõ â ‰"I €    " !r       ârÿÿ   ÿäÿ Ä ‰$ €    b rÿÿ   ÿä â !r       â/ z j  ÿ„ â ‰&7 €    b !r+       â $x+#   ÿŽ â z k   â ‰6      " !r       âO z	 j  ÿŠ Ä ‰ @    ¢ !r         â� z		 k  � ä ‰ @    â !r       â‚ +   ÿà÷ â !r"I"       â ‰ @    " !r$$       Æ‰ @    b !r&7&       Æ/ ‰ @    b !r66       Æ ‰ @    " !r       ÆO ‰  @    ¢ !   ÿàÿ â !r       ä� ‰#" @    â     ÿàõ ä ‚ÿÿ   ÿäÿ â ‰%$ @    â Š j  ÿ† â $(ÿÿ   ÿ â  j  ÿ„ â ‰)6 €    â !-   ÿàÿ â !r       âŠ k  � â ‰'& @    "  k  ( â !r       â! !   ÿàõ â ‰      b !r       â/ !r       â (ÿÿ   ÿä â ‰      " ! j  ÿ„ â !r         ÄO ‰      b !r#"#       â� ! k  ( â !r%$%       â ‰      ¢ !r)6)       Æ ‰      â !r'&'       Æ‰      " !r!       Æ‰      b !r       Æ ‰       " !r       Æ/ ‰#      h ‰%      b !r       ÆO ‰$) @    ¢ !r       Æ� ‰"'      â !r       â!r       ä†‰ !    â	 !r        Æ †    â	 !r##       Æ/ †‰     â	 !r%%       Æ †     â	 !r$)$       ÆO †‰     â	 !r''"       Æ� †    è	 †‰ 
    è	 †     è	 †‰ #    è	 † %   è	 ‰$      ( †‰ '    â	 M™        € ê x+   ÿàÿ â!r$       Æ r    ÿàñ È r ÿÿ   ÿä  ä z j  ÿ€ È z k     Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y1      "   " ¹z  l      â Ey  Ð+    € â ™x?    â y2      !   b ¹z  F   
   â r5 ÿ       â $r3ÿÿ   ÿ Ž â rD ÿ       â $r6ÿÿ   ÿ Ž â x(      ÿ  â $r ÿÿ   ÿ Ž â x*      ÿ  â x,      ÿ  â x.      ÿ  â $r0ÿÿ   ÿ Ž â r 1ÿ   pRð Ä r 1ÿ   pRò â$x1    2Ž Ê/ xÿ     È r   ÿ� È x4ÿ     È | 4   pbô Ú G)  p*    € ê y      '   " r ÿ       â $rÿÿ   ÿ Ž â xøÿÿÿÿÀŽ â y      &   b $v>ÿ q  ÿ Ž â z=  X      â $xA4   ÿ Ž â y	      %   ¢ $v;ÿ Y  ÿ Ž ä $x:>   AŽ âr7>A   ÿŽ â$x8>   AŽ âz?A q  ÿàÿ â $x<>   AŽ Ä $r5ÿÿ   ÿ Ž ä $x>>   AŽ ä 'z y   Ž È 'z s   Ž â/ zÿ z    Æ $x		   ÿ Ž âO zÿ t    â $z9 }  ÿŽ ä $z	 p  4Ž âx	   ÿàÿ â $z |  4Ž ä$z |  Ž â r  €ÿàÿ â $z v  Ž Ä $z v  Ž ä $z p  Ž â xÿ     â $z
 w  ÿŽ ä %xR   ÿŽ Æ r9
9   ÿàõ â %xR   Ž âr@   ÿà÷ ä rCÿÿ   ÿä â %x
R   ÿŽ È $x   ÿ Ž ä $xB   Ž ä %x   Ž È %x   
Ž È r=   ÿàõ Ê $x	   ; â r=   ÿàõ È �y"B   ‘ ¢ r;   ÿä â $t&ÿ   ÿ Ž â xK $       â �ybJ   ‘ è  �y   • " r =@   ÿàõ Æ �y   • d $x!;   B â r=   ÿàõ â �yf"   • è �y B   ‘ â  $x   ; Æ �ye$   • è �y   • h �y   • h �yd"   • h �yc$   • b r#9   Aàµ Æ �ya J   ‘ b r$C   ÿd~ Ê $x%$$   ÿŽ â rh9   ?àµ È r$C   ÿd~ â %vhh Z  K Ž â rN=   ÿàõ È rO;   ÿä ä r 9   :àÙ Æ �yNN    • b rMC   ÿ¤~ â %v   Z  K Ž È $xMM$   ÿŽ Ê r!!M   ÿàÿ Ê �yk    ™ h �ym    ™ h �yj    ™ h �yn    ™ h �yo    ™ h �yl     ™ h �yq     • b x"   ÿ   ÈO x   ÿÀŽ È xÿ   øŽ â$x"@   ÿ Ž â xÿ     Æ %v 
 & Ž â x   øŽ Ê �y    › ¢ %v 
 & Ž Ì/ �y    › " x   ÿ   È� x   ÿÀŽ È xÿ   øŽ â$x@   ÿ Ž â xÿ     Æ %v 
 & Ž â x   øŽ Ê �y    › â %v 
 & Ž Ì �y    › b $x   ÿ Ž Ê x   ÿÀŽ È xÿ   øŽ äx   ÿ   ä xÿ     â %v 
 & Ž È $x"   ÿ Ž â x   øŽ â �y    › ¦ x   ÿÀŽ â $x"   ÿ Ž Æ xÿ   øŽ â%v 
 & Ž â xÿ     Æ %v 
 & Ž âx   øŽ â �y    › ¨ �y    › â %v 
 & Ž È %v"# Z  K Ž ä �y    › ä $x##   %Ž Ê �y_"   ™ è �y`"   ™ è �yI"   ™ è �yJ"   ™ è �y["   ™ è �y]"   ™ è �y\"   ™ è �y^"    ™ â $x%$$   ÿŽ Æ �yG"    • â $xii   %Ž â r$
=   ÿà÷ È �yUh   ™ â  $x%   ;Ž Æ �yXh   ™ â  r9   7àµ Æ �yHh   ™ â  rFC   ÿd~ Æ �yYh   ™ â  r'9   8àµ Æ �yZh   ™ è  �yTh   ™ â rLC   ÿd~ är&9   >àµ â �yVh   ™ è �yWh    ™ â $xOF$   ÿŽ â rFC   ÿd~ â %v Z  K Ž â�yE$    • f %v"' Z  K Ž â/ �yMh    • f $x   OŽ Ä %v&& Z  K Ž â r$9   <àÙ ä �yR   ™ b $xOF$   ÿŽ ä %v$$ Z  K Ž â �yF   ™ b rKC   ÿ¤~ ä r''O   ÿàÿ â $xQL$   ÿŽ â �yO   ™ b $xKK$   ÿŽ Æ �yL   ™ b $x##   QŽ Æ �yS   ™ b $x%%   KŽ Æ �yQ   ™ h �yK   ™ h �yP    ™ h �yh    ™ h# �yi    ™ h �yg    • b xfÿ   ÿÀŽ Æ �ys"   ™ h �yp"   ™ h �y"   ™ h �y"   ™ b $x…f @  Ž Æ �yv"   ™ b x‚ÿ   f  Æ �yu"   ™ b $x€…€   ÿ Ž Æ �yt"   ™ b rˆ‚‚   ÿpŽ Æ �yr"    ™ b xƒ€€€ÿÀŽ â $x……   ÿ Ž ä $x‡ˆ€   ÿ Ž â x‚ƒ€€€€ÿüŽ â �y}"    • b x†…€€ÿÀŽ ä x„‚ÿþþþÿàÿ â �y!$   ™ b# x‡‡€€ÿÀŽ Ä r…„ƒ   ÿüŽ â �y|$   ™ b xƒ†€€€€ÿüŽ ä x„‡€€€€ÿüŽ â �yx$   ™ b xŠeÿ   ÿÀŽ ä xˆˆ   ÿ   â �yz$   ™ b xƒƒÿþþþÿàÿ â $xŠe @  ŠŽ â x„„ÿþþþÿàÿ â �yw$   ™ b xˆˆ€€ÿÀŽ Æ �y{$   ™ b r†ƒ†   ÿüŽ Æ �yy$   ™ b $xƒŠ€   ÿ Ž Æ �y $    ™ h �y~$    • h �y"&   ™ h �y�&   ™ h �y&   ™ h �yf&   ™ h �y#&   ™ h �y€&   ™ h �y%&   ™ h! �y$&    ™ h �y‚&    • b xƒƒ€€ÿÀŽ Ä r'„‡   ÿüŽ ä x‡ÿ   e  ä x&ˆ€€€€ÿüŽ ä x„Š   ÿ   â $x‰‡ @  ‡Ž â xe&ÿþþþÿàÿ ä x„„€€ÿÀŽ â $x‡‰€   ÿ Ž â x&ƒ€€€€ÿüŽ Ä rŠeˆ   ÿüŽ ä xe„€€€€ÿüŽ ä xˆ‰   ÿ   ä x&&ÿþþþÿàÿ ä x‡‡€€ÿÀŽ ä xeeÿþþþÿàÿ Ä x‹ˆ€€ÿÀŽ ä rƒ&ƒ   ÿüŽ ä x……˜º  ÿ    ä x&‡€€€€ÿüŽ ä r„e„   ÿüŽ ä xe‹€€€€ÿüŽ Ä xˆ…ÿÀŽ ä x&&ÿþþþÿàÿ ä x…€€€€…ÞŽ äO x††˜º  ÿ    ä xŒeÿþþþÿàÿ ä re&‡   ÿüŽ â $x‡…   ˆ
Ž â x‰†ÿÀŽ Ä x''˜º  ÿ    ä x†€€€€†ÞŽ ä x…dÿ   ÿÀŽ ä xŠŠ˜º  ÿ    ä xˆ'ÿÀŽ ä r††‰  €ÿàÿ â $x‰d @  …Ž â x'€€€€'ÞŽ Äx…ŠÿÀŽ ä x‡€€€€´Ž ä xŠ€€€€ŠÞŽ ä r‡'ˆ  €ÿàÿ â $x'‰€   ÿ Ž â x†€€€€´Ž ä x†ÿ   d  Ä rŠŠ…  €ÿàÿ â $x…‰   ÿ Ž â xd'€€ÿÀŽ ä r‰††   ÿpŽ ä x'‡€€€€´Ž ä x†…€€ÿÀŽ â $xˆ‰€   ÿ Ž â xd€€€€ÿüŽ Ä x‡†€€€€ÿüŽ ä x…ÿþþþÿàÿ ä x‰   ÿ   ä xˆˆ€€ÿÀŽ ä x‡‡ÿþþþÿàÿ ä xŠ€€€€´Ž Ä xŠ€€ÿÀŽ ä xˆ€€€€ÿüŽ ä r……d   ÿüŽ ä r‡‡†   ÿüŽ ä x‰ÿþþþÿàÿ ä x……˜º  ÿ    Ä x†Š€€€€ÿüŽ ä x‡˜º  ÿ    ä r&Œ‹   ÿüŽ ä xd…ÿÀŽ ä x‹†ÿþþþÿàÿ ä x‡ÿÀŽ Ä x…€€€€…ÞŽ ä� x€€€€ÞŽ ä r‰‰ˆ   ÿüŽ ä r‹‹Š   ÿüŽ â $x……   d
Ž â r‡  €ÿàÿ ä x‰‰˜º  ÿ    Ä xd‹˜º  ÿ    ä x†cÿ   ÿÀŽ ä x€€€€´Ž ä x‰ÿÀŽ ä x…€€€€´Ž ä x‰€€€€‰ÞŽ â$x‡c @  †Ž â x…dÿÀŽ Ä xd€€€€dÞŽ ä x†ÿ   c  ä rc‰  €ÿàÿ â $xdd   …
Ž â x‡   ÿ   â $x†† @  †Ž â xcc€€€€´Ž ä x‡   ÿ   Ä x€€ÿÀŽ â $x…†€   ÿ Ž â xd€€€€´Ž ä xd€€ÿÀŽ ä x€€€€ÿüŽ ä x‡†   ÿ   ä x†d€€€€ÿüŽ Ä xˆ…€€ÿÀŽ ä x…ÿþþþÿàÿ ä xŠ‡€€ÿÀŽ ä x‡†ÿþþþÿàÿ ä r……   ÿüŽ ä xŠ€€€€ÿüŽ Ä r‡‡d   ÿüŽ ä x‹ÿþþþÿàÿ ä x‡˜º  ÿ    ä xˆ€€€€ÿüŽ ä x‡ÿÀŽ ä x€€€€ÞŽ Ä x……˜º  ÿ    ä x‰ÿþþþÿàÿ ä r‡  €ÿàÿ ä x…ÿÀŽ ä xƒƒ˜º  ÿ    ä x…€€€€…ÞŽ Ä x€€€€´Ž ä r‹‹Š   ÿüŽ ä xƒÿÀŽ â $x……   
Ž â xƒ€€€€ƒÞŽ ä xee˜º  ÿ    ä x‹˜º  ÿ    â $xƒƒ   
Ž â r‰‰ˆ   ÿüŽ Ä x…€€€€´Ž ä xeÿÀŽ ä x…ÿÀŽ ä xe€€€€eÞŽ ä x‰‰˜º  ÿ    ä x€€€€ÞŽ Ä x„„˜º  ÿ    ä x&&˜º  ÿ    â $xee   
Ž â xd‰ÿÀŽ â &r_   ÿ   â r…  €ÿàÿ â &r__   ÿ   â x‰€€€€‰ÞŽ ä x…„ÿÀŽ Ä xƒ€€€€´Ž ä x„€€€€„ÞŽ ä xƒ&ÿÀŽ ä x&€€€€&ÞŽ â $x‰‰   d
Ž â r„„…  €ÿàÿ â &r`      âx€€€€´Ž â &r_`   _   â r&&ƒ  €ÿàÿ â &rI   ÿ   âx„€€€€´Ž â &rcJ      âx‰€€€€´Ž â &r_'J   _   â x&€€€€´Ž â &rJI   ÿ   Ä &r&[      âxe€€€€´Ž â &rJ[   J   ä &r&]   &   ä&r\      â xað   ÿÀŽ â&rJ]   J   â xaa   ÿÀŽ â &rI^   &   âxÿ     â &r^^   J   â x[bð   ÿÀŽ â$r&a   ÿŽ â xbb   ÿÀŽ â $xJ   IŽ Ä &r_\   _   â xÿ   [  â $r&I   &Ž â rIJJ   ÿ� â $r[b_   ÿŽ â r__^   ÿàÿ ä r&I&   ÿú� â $r[^   [Ž â rJ__   ÿ� ä xIÿ   &  Ä rJJ[   ÿú� â &r[U   ÿ   ârII&   ÿ� â &rUU   ÿ   ä &r&X   ÿ   ä &r[H   [   ä&rHH   U   ä &rXX   ÿ   Ä &rUY   &   ä&r[cZ   [   ä&rZ'Z   H   ä &rYY   X   ä &rHT   U   âxUÿ   J  â &rTT   Y   Ä &r&V   [   â rUUJ   ÿ� â &rYW   H   ä $rHa&   ÿŽ ä $xJ&   YŽ â x&ÿ   U  â 0rGÿG    A   ä 0rEÿE    A   Ä &r[V   Z   ä &rTW   T   ä $rYY   HŽ â ErH &       â  r&GE     A  â $rWb[   ÿŽ â r[[T   ÿàÿ â 0rNÿN    A   â #r/&H   /   â xIÿ   I  â &r&F   ÿ   Ä &rHF   ÿ   ä $rWT   WŽ â rT[[   ÿ� â &rFO   ÿ   â ErI I       â &r&L   &   â rUNG     A  â &rHL   H   â rTTW   ÿú� â &rFR   F   Ä &rOO   ÿ   â rJJJ   ÿ� â &r&cS   &   â #r0UI   0   â &rH'S   H   â xIÿ   T  â &rFQ   F   ä &rRR   O   â rJJY   ÿú� â &r&K   &   Ä &rOK   H   â rIIT   ÿ� â &rKP   F   â xGÿ   J  â $rFa&   ÿŽ ä $xH&   KŽ â x&ÿ   I  â 0rMÿM    A   â rGGJ   ÿ� â &rRQ   R   Ä $rKK   FŽ â ErF &       â  r&EM     A  â &rRP   R   â xGÿ   G  ä #r-&F   -   â &r&k   ÿ   â ErG G       â $rQbO   ÿŽ â rOOR   ÿàÿ â  rINM     A  â &rFm   ÿ   Ä &r&j   &   â rHHH   ÿ� â $rQR   QŽ â rJOO   ÿ� â &rkk   ÿ   â #r.IG   .   â &rmm   ÿ   ä &rIn   F   ä &r&co   &   â rHHK   ÿú� â &rjj   k   â rJJQ   ÿú� â &rmn   m   Ä &rFh   I   ä&r&i   &   â xGÿ   H  â &rj'o   j   â xIÿ   J  â &rmh   m   ä &rKl   F   ä $rFa&   ÿŽ â rGGH   ÿ� â &rii   j   â rIIJ   ÿ� â &rml   m   Ä $xH&   KŽ ä $rKK   FŽ â x&ÿ   I  â $rFbi   ÿŽ ä 0rgÿg    A   â riim   ÿàÿ â$rmm   FŽ â ErF &       ä  r&Eg     A  È #r+&F   +   â &r&s   ÿ   ä &rF   ÿ   â xGÿ   G  â &r&   &   ä &rFp   F   â ErG G       â &r   ÿ   â  rINg     A  â &rLs   ÿ   Ä &r&cv   &   ä &rFu   F   â rHHH   ÿ� â &rp      â #r,IG   ,   â &rL   L   ä &r&t   &   ä &rGr   F   â rJii   ÿ� â &ruu      â rHHK   ÿú� â &rL'v   L   Ä $ra&   ÿŽ ä $x&&   GŽ â rJJm   ÿú� â &rKt   L   â xÿ   H  â &rur   u   â rI&&   ÿ� â $rG   Ž â rH   ÿ� â $rFbK   ÿŽ â xGÿ   J  Ä rKKu   ÿàÿ â0rqÿq    A   â rII   ÿú� â $rFu   FŽ â rGGJ   ÿ� ä xÿ     ä rKKK   ÿ� ä x&ÿ   I  Ä xÿ   G  â  rGNq     A  â Er        ä rFKF   ÿú� ä rI&I   ÿ� â #r*G   *   â Er&        â  rEq     A  â xGÿ   F  Æ #r)&   )   â &r!   ÿ   ârGGF   ÿ� â &r!!   ÿ   ä &r&w   ÿ   ä&rFw   ÿ   ä &r"   ÿ   âxIÿ   I  â &r"   ÿ   Ä &r#   ÿ   ä&r#   ÿ   ä 0r}ÿ}    A   ä &r|      ä&r!|   !   ä &r&{   &   Ä &r�      â ErI I       â &rF{   F   â xGÿ   G  â &r�      ä &r€      ä&r€      â  rN}     A  â &rcx      Ä &r!'x   !   ä &r&y   &   ä &rc      â ErG G       â &rFy   F   ä &r'      ä &r%      Ä &r%      â #r(I   (   â  r}E}     A  â &rz      ä&rIz   !   ä &r    &   ä &rf      ä &rF    F   Ä &rf      ä &r$      ä&r$      â #r }G       â $r!a   ÿŽ ä$rGbI   ÿŽ ä $raa   ÿŽ â rIIF   ÿàÿ â $x   Ž Ä $rb   ÿŽ â r   ÿàÿ â $x   Ž ä $rGF   GŽ â r   ÿ� â $r!   !Ž ârII   ÿ� â $ra   aŽ â r   ÿ� â $r   Ž â r   ÿ� Ä r!   ÿú� ä rG   ÿú� ä r   ÿú� ä ra   ÿú� ä x44   ÿàÿ ä xÿ     Ä xÿ     ä xÿ     ä xÿ     ä | 4   pbô ä r   ÿ� ä r   ÿ� Ä r   ÿ� ä r   ÿ� â 0r~ÿ~    A   â xÿ     â 0r‚ÿ‚    A   â xÿ     ä xÿ     ä xÿ     â  rN~     A  âEr        â  rE~     A  âEr        â  rE‚     A  âEr        â  rNN‚     A  â Er        Ä x==�  ÿà÷ â #rD   D   â #r6   6   â #r5   5   â #r3N   3   â x77@   ÿàÿ â $r;ÿÿ   ;Ž â x::@   ÿàÿ ä x88@   ÿàÿ ä x<<@   ÿàÿ Ä x>>@   ÿàÿ ä x??@   ÿàÿ ä xAA@   ÿàÿ â D)       À â Gy  @Øÿÿÿÿƒ ê Ay        € ê Ey      € â G‰  ð     € ê $x11À  2Ž Ê ˆs 10 ùÿ H   è ˆs 1/€ùÿ H   è ˆs 1. úÿ H   è ˆs 1-€úÿ H   è ˆs 1, ûÿ H   è ˆs 1+€ûÿ H   è ˆs 1* üÿ H   è ˆs 1)€üÿ H   è ˆs 1( ýÿ H   è ˆs 1 €ýÿ H   è ˆs 1D þÿ H   è ˆs 16€þÿ H   è ˆs 13 ÿÿ H   è ˆs 15€ÿÿ H   ä Ay        € ê {          ì M        € ê „y2     H   b $v'ÿ r  ÿ Ž Æ „y2 €  H   ¨ „y2 €   H   è „y2    H   ( „y2 €  H   h „y2    H   ( „y	2    H   ( „y2 €  H   ( „y2    H   è „y2 €  H   â !r0       Æ/ „y2    H   â !r++       Æ_ ‰      " !r/       Æ� „y2 €  H   â !r..       Æ„y2    H   â !r-       Æ„y2 €  H   â $x-'   ÿŽ â !r,,       ä y      %   b !r*	*       Æ y      &   ¢ !r))       â !r       ä ‰      " !r((       Æ y      '   è ‰.      " !rDD       Æ ‰      b x   ÿ   Æ/ ‰	      b x   ÿàÿ Æ ‰,      b $z x  Ž âO z  r  p`ð â !r        â ‰+      ¢ !r
       â !r66       â !r33       â ‰*      " !r55       â $z ~  Ž â� z  r  p`ò â ‰)      â !r..       âr 2ÿ   pVp  Ä ‰(      " !r       âr'   ÿˆ ä ‰      b !r	       â/ x 2   p"ò â $rÿÿ   ÿ â !r,,       â ‰D      b z j  ÿˆ â !r+       ÄO ‰
      ¢ $x+'   ÿŽ â Š r  ÿà÷ â !r**       â ‰6      " z k   ä ‚ +   ÿàõ â ‰3      " !r))       â� ‚ÿÿ   ÿäÿ â !r((       â‰5      â ‚ÿÿ   ÿä â !r        Ä‰.      " Š j  ÿ† äŠ  j  ÿ„ â ‰	 €    b !rDD       â/ Š k  � ä ‰      b !r
       âO Š  k   â $x'   ÿŽ â ‰      ¢ !r66       Æ ‰,      " !r33       â r   ÿàû ä ‰*      " !r55       ä� $rÿÿ   ÿŽ â ‰)      â !r..       â‚-   ÿà÷ Ä ‰(      " !r 	       â	'   ÿàÿ ä‰!       b !r       â/ r'   ÿ„ â !r       âO ‰#D      b ‚ÿÿ   ÿäÿ ä $rÿÿ   ÿ â !r,,       â ‰ €    " &	   ÿà÷ â !r**       Ä ‰%6      ¢ z
 j  ÿŠ ä z j  ÿ„ â ‰"3      ¢ Š j  ÿˆ â !r)       â� 'ÿÿ   ÿäÿ â ‰. €    â & j  ÿ† â !r((       â!r  !       â‰  @    " z k  � Ä z	 k   â ‰ €    b !rDD#       â/ Š k   ä ‰, €    b !r       â & k  '� ä ‰$5      " !r66%       âO +   ÿàÿ Ä ‰ €    ¢ !r33"       â    ÿàõ ä ‰* €    ¢ !r..       â� -   ÿàÿ â $"ÿÿ   ÿ â ‰# €    â !r        â   ÿà÷ Ä ‰%( €    " !r       â j  ÿ„ ä‰'  €    b !r,,       â/  k  " ä ‰)D €    b !r55$       Æ ‰ @    " !r       âO ÿÿ   ÿäÿ ä ‰+6 €    ¢  j  ÿ† â!r**       ä ‰03 €    ¢ !r##       â�  k  � Ä ‰. @    â !r%(%       Æ‰       " !r' '       Æ‰25 €    b !r)D)       Æ/ ‰ @    b !r       â $„ÿ   ÿ Ž ä ‰, @    " !r+6+       âO %† j   Ž ä ‰ @    ¢ !r030       Æ ‰!* @    ¢ !r.       Æ� ‰ # @    â !r1        Æ‰"% @    " !r252       â5        ä ‰$' @    b !r       Æ/ ‰&) @    b !r,       Æ ‰      " !r       ÆO ‰(+ @    ¢ !r!*!       Æ ‰-0 @    ¢ !r #        Æ� ‰       â !r"%"       Æ‰/2 @    " !r$'$       Æ‰      b !r&)&       Æ/ ‰      b !r3       â % j  5 Ž ä ‰      " !r(+(       ÆO ‰!      ¢ !r-0-       Æ ‰#       ¢ !r)        Æ� ‰%"      â !r/2/       Æ‰'$      " !r       Æ†‰ 1    â !r       Æ/ ‰&      b !r       Æ ‰(      " !r!!       ÆO ‰ -      ¢ !r# #       Æ † 3   â !r%"%       Æ� ‰/      â !r'$'       Æ†‰ )    è	 †     â	 !r&       Æ/ †‰     â	 !r(       Æ †    â	 !r--        ÆO †‰ !    è	 † #    â	 !r//       Æ� †‰ %    è	 † '   è	 †‰     è	 †     è	 †‰ 
-    â	 M™        € ê †y 
/   â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y       "   " ¹z  l      â Ey  P(    € â ™x?    â y      !   b ¹z  F   
   â $rÿÿ   ÿ Ž â r ÿ       â x      ÿ  â $rPÿÿ   ÿ Ž â x      ÿ  â $rTÿÿ   ÿ Ž â rV ÿ       â $rÿÿ   ÿ Ž â x      ÿ  â r  ÿ   pRð Ä r  ÿ   pRò â$x~     Ž Ê/ xÿ   ~  È r
~   ÿ� È xÿ   
  È |    pbô Ú G)   '    € ê y      '   " $rÿÿ   ÿ Ž â x
øÿÿÿÿÀŽ â $vÿ q  ÿ Ž â y      %   b $x   ÿ Ž â r~~  €ÿàÿ â $vÿ X  ÿ Ž â y      &   ¢ $x   Ž â xÿ   ~  Ä x3~   ÿ   ä z q  ÿàÿ â 'z y   Ž È $x   ÿ Ž ä/ $rÿÿ   ÿ Ž ä $z p  Ž âx   ÿàÿ â 'z s   Ž âO zÿ z    Æ $z }  ÿŽ â zÿ t    â $z |  Ž ä$z |  Ž ä $z v  Ž ä$z4 v  Ž â z  Y      â $z9 p  Ž Ä $z
 w  ÿŽ ä %x44R   ÿŽ Æ r
   ÿàõ â %x69R   ~Ž â r~4   ÿà÷ Æ %x89R   ÿŽ È $x   Ž ä$x   Ž ä %x:3   4Ž È $rÿÿ   ÿ Ž ä $x   Ž ä $rÿÿ   ÿ ä $x5   Ž ä %x23   8Ž È rl6   ÿàõ Ê $xm7    â r2   ÿàõ È �y!lB   ‘ ¢  r3   ÿä Ê �y   • â rp   ÿàõ Æ �y    • $ $xq    â rL:   ÿàõ â �y"   • h �ypB   ‘ b  $xM;    Æ �y*$   • ( �y
L   • ( �yL   • " $t$ÿ   ÿ Ž Æ �y+L"   • " xN $       Æ �yppJ   ‘ ( �yllJ   ‘ ( �yLL$   • b x"!   ÿ   ÈO x""   ÿÀŽ È x""ÿ   øŽ ä�xÿ     Æ %v@" 
 $ Ž È $x"!@   ÿ Ž ä �y@@    › ¦ x   "øŽ Ê %v< 
 $ Ž Ì �y<<    › â x   ÿ   È/ x   ÿÀŽ È xÿ   
øŽ Ê%v> 
 $ Ž â x
ÿ   
  Æ $x@   ÿ Ž ä �y>>    › & x
   øŽ Ê %vB 
 $ Ž Ì �yBB    › ¢ $x   ÿ Ž Ê x   ÿÀŽ È xÿ   øŽ äx   ÿ   ä xÿ     â %vD 
 $ Ž Æ x   øŽ â $x!   ÿ Ž ä �yDD    › ¤ %vF 
 $ Ž â x   ÿÀŽ ä r~   àµ â $x!!   ÿ Ž â xÿ    øŽ â�yFF    › ¢ x ÿ      Ä r   ÿd~ â %vH 
 $ Ž â xJ    !øŽ Æ %v  Z  N Ž ä �yHH    › ¤ $x$   ÿŽ ä %vJJ 
 $ Ž È $x!!   Ž ä �yJJ    › ¨ �y,    ™ ¨ �yw    ™ è �y#    ™ è �y(    ™ è �yu    ™ è �y    ™ â r1~   àµ Æ �yv    ™ è �y%     ™ â r   ÿd~ â %v01 Z  N Ž È $x$   ÿŽ È $x11   Ž â �y      • (# �yn0   ™ ( �yr0   ™ ( �y0   ™ ( �ys0   ™ ( �yt0   ™ " r~   àÙ Æ �ym0   ™ " rd4   ÿàõ Æ �yo0   ™ " r
   ÿ¤~ Æ �yq0    ™ " re5   ÿä â %v Z  N Ž â r&8   ÿà÷ â �yb0    • $ $x)
$   ÿŽ ä �ydd    • " $x'9   Ž ä $x   )Ž â r)~   àµ Ä �yg&    • $ r
   ÿd~ ä �yh   ™ ( �ye   ™ " %v&) Z  N Ž Æ/ �yj   ™ " $x)
$   ÿŽ Æ �y
   ™ ( �yk   ™ " r/~   àÙ Æ �yi   ™ ( �yc   ™ " r"   ÿ¤~ Æ �yf    ™ " %v./ Z  N Ž Æ �y]    • " $x!"$   ÿŽ Ê r//!   ÿàÿ Ê �yQ.   ™ ( �yZ.   ™ ( �ya.   ™ ( �y[.   ™ ( �y^.   ™ ( �y`.   ™ ( �y\.   ™ ( �y_.    ™ " rO~   àÙ â $x''   )Ž â x!ÿ   ÿÀŽ â�y.    • b r   ÿ¤~ â %vNO Z  N Ž Ä �y&   ™ d $x-$   ÿŽ ä �yS&   ™ d $xOO   -Ž â r}!   ÿpŽ â �y$&   ™ h �yY&   ™ h �yU&   ™ h �yX&   ™ h �yW&   ™ h �yR&    ™ h �y"&    • h �y!N   ™ h �yMN   ™ h �y'N   ™ h �y0N   ™ h �y/N   ™ b! $xy}€   ÿ Ž Æ �y-N   ™ b xÿ     Æ �y&N   ™ h �y)N    ™ b xyy€€ÿÀŽ â $x. @  Ž â x}}   ÿ   ä xy€€€€ÿüŽ ä x1*ÿ   ÿÀŽ â $x.€   ÿ Ž â xxÿþþþÿàÿ â $x..   ÿ Ž â x}}€€ÿÀŽ Ä r1*1   ÿpŽ ä x*ÿ   *  ä ryxy   ÿüŽ ä xx}€€€€ÿüŽ ä x..€€ÿÀŽ ä r***   ÿpŽ Ä x€€ÿÀŽ ä �yN    • b xxxÿþþþÿàÿ â $x|*€   ÿ Ž ä$x**   ÿ Ž â r}x}   ÿüŽ â $xx1   ÿ Ž âxO.€€€€ÿüŽ â $xN1€   ÿ Ž Æ xOOÿþþþÿàÿ ä x1N€€ÿÀŽ ä rOO.   ÿüŽ ä x**€€ÿÀŽ ä x.1€€€€ÿüŽ ä x{€€€€ÿüŽ Ä xz*€€€€ÿüŽ ä x..ÿþþþÿàÿ ä x}}˜º  ÿ    ä x{{ÿþþþÿàÿ ä xyy˜º  ÿ    ä xzzÿþþþÿàÿ Ä xxx€€ÿÀŽ ä r1.1   ÿüŽ ä x.}ÿÀŽ ä x}A€€€€}ÞŽ äO r{   ÿüŽ ä xNyÿÀŽ Ä rzz*   ÿüŽ ä x{x€€€€ÿüŽ ä xy@€€€€yÞŽ ä x*+ÿ   ÿÀŽ â $x.}   .
Ž â x{{ÿþþþÿàÿ ä rNyN  €ÿàÿ â $x*+ @  *Ž â x||€€ÿÀŽ Ä x˜º  ÿ    ä rx{x   ÿüŽ ä xNN€€€€@´Ž ä x..€€€€A´Ž â $xA*€   ÿ Ž â x{|€€€€ÿüŽ ä x@ÿÀŽ Ä xy<€€€€ÞŽ ä� xOO˜º  ÿ    ä x{{ÿþþþÿàÿ ä r@y@  €ÿàÿ ä x+ÿ   +  ä xAA€€ÿÀŽ Ä xOÿÀŽ ä x**   ÿ   ä xO=€€€€OÞŽ ä r{{|   ÿüŽ ä x<@€€€€<´Ž â $x@+ @  +Ž â x|A€€€€ÿüŽ Ä x+*€€ÿÀŽ â $xO   
Ž â x||ÿþþþÿàÿ ä xy+€€€€ÿüŽ â $x*@€   ÿ Ž â x=€€€€=´Ž ä r||A   ÿüŽ ä x@   ÿ   Ä xyyÿþþþÿàÿ ä x**€€ÿÀŽ ä x||˜º  ÿ    ä x€€ÿÀŽ ä ryy+   ÿüŽ ä x@*€€€€ÿüŽ Ä x+|ÿÀŽ ä xA€€€€ÿüŽ ä xyy˜º  ÿ    ä x|>€€€€|ÞŽ äx@@ÿþþþÿàÿ ä xAAÿþþþÿàÿ â $x+|   +
Ž â xOyÿÀŽ Ä xy?€€€€yÞŽ ä r@@*   ÿüŽ ä rAA   ÿüŽ â $xy   O
Ž â x>+€€€€>´Ž ä x+@˜º  ÿ    ä x@A˜º  ÿ    Ä xOLÿ   ÿÀŽ ä x€€€€?´Ž ä x?+ÿÀŽ ä x*@ÿÀŽ ä x|B€€€€+ÞŽ ä xAC€€€€@ÞŽ Ä r+LO   ÿpŽ ä xLÿ   L  â $x@A   *
Ž â xA+   ÿ   â $x?|   ?
Ž ä $x*L @  LŽ â xAA€€ÿÀŽ â $x++   ÿ Ž â xB?€€€€B´Ž Ä x?*   ÿ   äx|A€€€€ÿüŽ â $x**   ÿ Ž â x++€€ÿÀŽ ä x??€€ÿÀŽ ä x||ÿþþþÿàÿ ä xC@€€€€C´Ž Ä x@+€€€€ÿüŽ ä xL?€€€€ÿüŽ ä x**€€ÿÀŽ ä rO|A   ÿüŽ ä x@@ÿþþþÿàÿ ä xLLÿþþþÿàÿ Ä xA*€€€€ÿüŽ ä xOO˜º  ÿ    ä r@@+   ÿüŽ ä rLL?   ÿüŽ ä x?Aÿþþþÿàÿ ä x+OÿÀŽ Ä x|D€€€€OÞŽ ä x@@˜º  ÿ    ä r??*   ÿüŽ â $xO|   +
Ž â x*@ÿÀŽ ä xAE€€€€@ÞŽ ä x+L˜º  ÿ    Ä x@?˜º  ÿ    ä rLA*  €ÿàÿ ä x?+ÿÀŽ ä x|F€€€€+ÞŽ ä x*@ÿÀŽ ä x+G€€€€@ÞŽ Ä x11˜º  ÿ    ä xxx˜º  ÿ    â $x*+   *
Ž â x@1ÿÀŽ ä x+xÿÀŽ ä x1H€€€€1ÞŽ ä xxI€€€€xÞŽ Ä x{{˜º  ÿ    ä r11@  €ÿàÿ â $xxx   +
Ž â x@{ÿÀŽ â &r+>,   ÿ   âx{J€€€€{ÞŽ â &r,N,   ÿ   â xDO€€€€D´Ž â $x?|   ?
Ž â xEL€€€€E´Ž â &r+w   +   âxH1€€€€H´Ž â $x{{   @
Ž â xzz˜º  ÿ    â &rw.w   ,   Ä &r@D#   ÿ   âxF?€€€€F´Ž â &r+B(   +   âxIx€€€€I´Ž â &rw<(   w   â x1zÿÀŽ â &r(H#   ÿ   â xzK€€€€zÞŽ â &r@Eu   @   âx**€€€€G´Ž â &ruIu   (   â xJ{€€€€J´Ž â &r@F   @   â rzz1  €ÿàÿ Ä x(pð   ÿÀŽ â&r+Cv   +   âxpp   ÿÀŽ â &r@*%   @   â x,z€€€€K´Ž â &ruJ   u   â xÿ   (  â $r(p+   ÿŽ â x1lð   ÿÀŽ â$x++   @Ž â xll   ÿÀŽ â &r#=v   w   Ä &rL,%   u   â x%ÿ   1  â $r(@   (Ž â r+++   ÿ� â $r1l#   ÿŽ ä $x##   LŽ â r++(   ÿú� â $r1%L   1Ž Æ r(##   ÿ� ä x@ÿ   +  ä r((1   ÿú� â &r1>n   ÿ   âr#@+   ÿ� â &rnNn   ÿ   ä &r@Dr   ÿ   ä &r1   1   Ä &r+.   n   ä &rEs   @   ä &rrHr   ÿ   ä &r@Bt   1   ä&r1Fm      â xÿ   (  â &rt<t   +   Ä &rrIs   r   ä &r+Co   @   â r(   ÿ� â &r@*q   1   ä &rrJm   r   ä $r(p+   ÿŽ ä &ro=o   t   â r++@   ÿàÿ â &rr,q   r   Ä $r@@   (Ž â x#ÿ   #  â $r(lo   ÿŽ ä 0rdÿd    A   ä 0r ÿ     A   ä $xoo   rŽ â Er# #       â $rr%r   (Ž â x(ÿ     â  rd      A  â 0rgÿg    A   Æ #r#      â &r>h   ÿ   ä&r1Nh   ÿ   â Er( (       â &r#De   ÿ   â  r  g     A  â &r
      âr+++   ÿ� â &r1.
   1   â rooo   ÿ� â &r#Ej   #   â #r (      â &r
Bk      âr++@   ÿú� â &r(<k   1   â roor   ÿú� â &r#Fi   #   â x ÿ   +  â &rCc   
   âx
ÿ   o  â &rc=c   (   Ä &r(*f   #   â r  +   ÿ� â &reHe   ÿ   â r

o   ÿ� â $r#p   ÿŽ ä $x+   (Ž â x ÿ      â 0rbÿb    A   â x
ÿ   
  â &reIj   e   Ä $r#(   #Ž â r(++   ÿ� â &reJi   e   â  rdb     A  â Er          â  rbgb     A  â Er
 
       ä r#(#   ÿú� â &rf,f   e   â #rV    V   â #rb
      â x
ÿ   #  â &r>Q   ÿ   Ä &r DZ   ÿ   ä $r1lc   ÿŽ â rccf   ÿàÿ â&ra      â r

#   ÿ� â &rQNQ   ÿ   ä &r#E[       â r@cc   ÿ� â $r1%f   1Ž Ä &rZHZ   ÿ   ä &rB^      ä &rQ.a   Q   ä &r#F`   #   â r1@1   ÿú� â &r[I[   Z   ä &rC\      Ä &rQ<^   Q   ä &r(*_   #   â x ÿ   1  â &r[J`   [   ä $r#p   ÿŽ ä &rQ=\   Q   ä $x   (Ž â r  1   ÿ� â &r@,_   [   Ä $r#(   #Ž â r(   ÿ� â $r+lQ   ÿŽ â x ÿ      â $xQQ   @Ž ä 0r]ÿ]    A   â r((#   ÿú� â $r+%@   +Ž â r@QQ   ÿ� ä Er#         â  r g]     A  â x
ÿ   
  Ä r@@+   ÿú� â #r	 #   	   â x+ÿ   (  â &r#>   ÿ   ä
&r1N   ÿ   â Er
 
       â &rDS   ÿ   â  rd]     A  â r++(   ÿ� â &r#$   #   Ä &r1.$   1   ä &rSHS   ÿ   ä &rEX      ä &r>>!   ÿ   ä &rDD'   ÿ   â #r
      â &r!N!   ÿ   â x+ÿ   +  â &r'H'   ÿ   Ä 0rÿ    A   ä &r
BY   #   ä&r <Y   1   ä &rSIX   S   ä &r1FW      ä &r>M   >   â Er+ +       â &r!.M   !   Ä &rDE0   D   ä&r'I0   '   â xÿ   @  â  rd     A  â &r#CU   
   ä &rSJW   S   ä &r
*R   1   ä &r>B/   >   Ä &r!</   !   ä &rDF-   D   ä&r'J-   '   â r@   ÿ� â #rT+   T   â &rU=U       ä $rp#   ÿŽ â r##
   ÿàÿ â &rR,R   S   Ä &rCC&   >   ä&r!=&   !   ä &rD*)   D   ä&r,,)   '   â xÿ     â $r

   Ž â r###   ÿ� â $r lU   ÿŽ Ä $rppC   ÿŽ ä $xUU   RŽ ä $rll!   ÿŽ â r!!,   ÿàÿ â $xCC   DŽ â Er        â  rg     A  â $r %R    Ž â r
#
   ÿú� â $rpD   pŽ â rUUU   ÿ� â $rl%,   lŽ â rCCC   ÿ� Ä r!!!   ÿ� â #rP   P   â xÿ   
  ä r U    ÿú� ä rCCp   ÿú� ä r!!l   ÿú� ä x   ÿàÿ Ä r
   ÿ� ä xÿ      ä xÿ   !  ä x
ÿ   C  â 0r"ÿ"    A   â |    pbô ä xÿ     Ä r    ÿ� ä r!   ÿ� ä r

C   ÿ� â  rd"     A  â Er        â 0rÿ    A   â xÿ     ä xÿ     Ä x
ÿ   
  â #r      â Er        â  rg"     A  âEr        â  rg     A  â Er
 
       â  rd     A  â x�  ÿà÷ â #r      â #r      â #r
      â x@   ÿàÿ â $rÿÿ   Ž â x@   ÿàÿ Ä x@   ÿàÿ ä x@   ÿàÿ ä x@   ÿàÿ ä x@   ÿàÿ â D)       À â Gy   Ûÿÿÿÿƒ ê Ay        € ê Ey  ð     € â G‰  Ð     € ê $x €  Ž Ê ˆs  úÿ H   è ˆs €úÿ H   è ˆs V ûÿ H   è ˆs €ûÿ H   è ˆs  üÿ H   è ˆs 	€üÿ H   è ˆs T ýÿ H   è ˆs P€ýÿ H   è ˆs  þÿ H   è ˆs €þÿ H   è ˆs  ÿÿ H   è ˆs €ÿÿ H   ä Ay        € ê {          ì M        € ê „y     H   " z+  r      Æ „y  €   H   d $x'+   ÿŽ ä„y €  H   ¢ $x)+   ÿŽ Æ „y    H   è „y €  H   ( „y    H   h „y
 €  H   h „y    H   h „y €  H   h „y    H   è „y    H   â !r       â !r        Ä/ y      %   " !r       ÆO ‰       b !rVV       Æ� ‰      ¢ !r       Æ„y €  H   â !r       Æ‰      â !r
	       Æ ‰      " !rT       Æ ‰	V      b !rP       â !r         ä/ ‰      b !r       â !r       äO ‰       ¢ !r       Æ ‰      " !r
       Æ� ‰      â !r       Æ‰      " !r	V	       Æ‰      b !r       Æ ‰      b !r       Æ/ ‰      b !r        âO !r       ä y       &   " !r       Æ� y      '   ¢ $x   ÿ Ž â !r       ä‰ €    ä x   ÿàÿ â !r       â‰	      " z  r  p`ð â !r       Ä ‰
      b z  r  p`ò â !r       ä/ ‰ €    b r ÿ   p"p äx    p"ò â ‰      b $z   x  Ž Æ ‰      " $z  ~   Ž ÆO ‰      ¢ !r       â�  '   ÿàõ ä‰      â !r	       â
	  r  ÿà÷ ä ‰      " !r
       âr+    ÿˆ â $&ÿÿ   ÿŽ â ‰      b !r       â/ 
	 j  ÿ† Ä ‰      b !r       â 
		 k  &� ä ‰$      b !r       â $rÿÿ   ÿ â z j  ÿˆ â!r       âO ‰ @    $ z k   â !r        â� ‰ €    ¢ !r       Æ‰
 @    â ÿÿ   ÿä â !r       ä‰ €    " 
 j  ÿ„ â!r"       ä/ ‰ €    b 
 k   â !r$$       Ä ‰ €    h ‰ €    b !r       â r+    ÿ„ ä ‰!  €    " !r       âO  )   ÿàù â $r&ÿÿ   ÿ â ‰ €    ¢ !r
       â� z
 j  ÿ„ Ä ‰ €    â !r       âz k  & ä ‰#" €    " !r       â/ +   ÿàÿ â $ÿÿ   ÿ â ‰%$      b !r       â(    ÿà÷ â !r       â ‰ @    b 
 j  ÿˆ â !r! !       Ä ‰ @    " +ÿÿ   ÿäÿ â !r       äO ‰      ¢ ( j  ÿ† â !r       ä� ‰ @    â 
 k   â !r#"#       Ä‰ @    " ( k  +� â !r%$%       ä/ ‰ @    h ‰"! @    b !r       â$ÿ   ÿ Ž Ä ‰ @    b !r       Æ ‰  @    " !r+       âO '   ÿàÿ ä ‰      ¢ !r       Æ� ‰$# @    â !r       Æ‰&% €    " !r       â/ !r"!"       â $!ÿ   ÿ Ž â ‰      b !r       Æ‰      b !r         Æ ‰      " !r-       âO     ÿàõ ä‰      ¢ %  j   Ž â!r$#$       â� (ÿÿ   ÿä â ‰      â %  j  ! Ž â !r&%&       â j  ÿ„ â‰      &  k  ( â ‰!       " !r       Æ/ ‰#"      b !r       Æ‰'& @    b !r       Æ ‰%$      " !r       ÆO †	 +    â !r       Æ� † -   â !r       Æ†	     â !r! !       â !r#"#       ä/ †     â !r'&'       Æ†	     â !r%$%       Æ †    è †	     è †     è †	 
!    è † 
#   è ‰'      ( †	 %    â M™        € ê x)   ÿàÿ âO !r''       Æ r    ÿàñ Ê $r ÿÿ   ÿ  â z j  ÿ€ È z k     Ê †y '    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y/      "   " ¹z  l      â Ey  À$    € â ™x?    â y0      !   b ¹z  F   
   â $r4ÿÿ   ÿ Ž â x@      ÿ  â y5      %   ¢ $r1ÿÿ   ÿ Ž â x*      ÿ  â $r ÿÿ   ÿ Ž â y      &   â x,      ÿ  â r. ÿ       Ä r /ÿ   pRð är /ÿ   pRò â$x/    0Ž ä/ $x55   ÿ Ž ÆO xÿ     È r   ÿ� È x2ÿ     È | 2   pbô Ú G)  `#    € ê y      '   "Ž $rÿÿ   ÿ Ž â xøÿÿÿÿÀŽ â $rÿÿ   ÿ Ž â x
5   ÿàÿ â $z w  ÿŽ âz:  q      â 'z s   Ž È $x=2   ÿ Ž â zÿ t    â $v3ÿ X  ÿ Ž ä $x8:   =Ž âr6:=   ÿŽ â$v9ÿ Y  ÿ Ž â z;= q  ÿàÿ â $r4ÿÿ   ÿ Ž ä $x::   =Ž Ä 'z y   Ž È $z7 }  ÿŽ â z	ÿ z    â $z5 p  2Ž Æ r77   ÿàõ â $z	 |  2Ž ä$z		 |  Ž ä $x   
Ž ä $z v  Ž ä$z	 v  	Ž â xÿ     â $z
 p  Ž Ä %x	R   ÿŽ È %x
R   Ž â r<   ÿà÷ Æ %xR   ÿŽ È $x   ÿ Ž ä $r?ÿÿ   ÿ ä %x   Ž È $x>	   Ž ä %x   Ž È rL
3   ÿàõ È rM9   ÿä ä r3   ÿàõ Æ �y#LB   ‘ ¤  $x   9 Ê �y   • â rP3<   ÿàõ Æ �y"   • $ $xQ9   > â r3   ÿàõ â �yi"   • h �yPB   ‘ " $x   9 Æ �yG$   • ( �y   • ( �y   • " x'        Æ �yH"   • ( �yK$   • b r%7   =àµ â $t`ÿ$   ÿ Ž ä �yPPJ   ‘ h �yLLJ   ‘ b r?   ÿd~ â/ %v$% Z  ` Ž È $x#   ÿ Ž ÊO x   ÿÀŽ È xÿ   øŽ ä�xÿ     Æ %v  
 ' Ž È $x#@   ÿ Ž ä �y      › ¦ x   øŽ Ê %v 
 ' Ž Ì �y    › â $x   ÿ Ž Êx   ÿÀŽ È xÿ   øŽ äxÿ     Æ %v 
 ' Ž È $x@   ÿ Ž ä �y    › & x   øŽ Ê %v 
 ' Ž Ì �y    › ¢ $x   ÿ Ž Ê x   ÿÀŽ È xÿ   øŽ Ê %v 
 ' Ž â x   ÿ   Æ $x#   ÿ Ž â xÿ     â �y    › ¦ x   ÿÀŽ ä x   øŽ ä xÿ   "øŽ â$x##   ÿ Ž â x"ÿ   "  â %v 
 ' Ž È %v 
 ' Ž âx"   #øŽ â �y    › ¨ �y    › â %v 
 ' Ž È $x$   ÿŽ ä �y    › ä $x%%   Ž Ê �yI$   ™ è �yD$   ™ " r7   ;àµ Æ �yJ$   ™ ( �yU$   ™ " r"?   ÿd~ Æ �yF$   ™ ( �yB$   ™ " %v Z  ` Ž Æ �yV$   ™ " $x#"$   ÿŽ Æ �yE$    ™ $ r#   ÿàÿ ä �yX$    • ( �yN   ™ ( �yR   ™ ( �yC   ™ " r&3   ÿàõ Æ �yS   ™ ( �yT   ™ " $x'	   9 Æ �yM   ™ " r#7   6àÅ Æ �yO   ™ " rY?   ÿ„~ Æ �yQ    ™ " %v"# Z  ` Ž Æ �yW&    • " $xYY$   ÿŽ â r(3   ÿà÷ Æ $x##   YŽ â r%7   :àÙ â/$x)   9Ž â �y    • ( �yY"   ™ ( �yZ"   ™ ( �y_"   ™ ( �y["   ™ ( �y((    • ( �y\"   ™ ( �y^"   ™ ( �y)"   ™ ( �y]"    ™ " r'7   8àµ Èrb?   ÿd~ ära?   ÿ¤~ â %v&' Z  ` Ž È $xcb$   ÿŽ â xjiÿ   ÿÀŽ â %v$% Z  ` Ž ä �y`"    • d $xaa$   ÿŽ ä $x''   cŽ ä $xoi @  jŽ â r%%a   ÿàÿ Ä �ya&   ™ b $xko€   ÿ Ž Æ �yc&   ™ b xmÿ   i  Æ �yb&   ™ h �y"$   ™ h# �yf&   ™ b xqk€€ÿÀŽ Æ �yi$   ™ h �yg&   ™ h �yl$   ™ h �ye&   ™ h �yk$   ™ h �yd&   ™ h �yj$   ™ h �y&    ™ h �yh&    • h �y#$   ™ h �y'$   ™ h �y&$    ™ b $xtm @  mŽ â xmq€€€€ÿüŽ È xnmÿþþþÿàÿ ä �ym$    • b $xoo   ÿ Ž â xpt   ÿ   È xoo€€ÿÀŽ ä xpp€€ÿÀŽ ä rqnq   ÿüŽ â $xtt   ÿ Ž â xno€€€€ÿüŽ ä xvGÿ   ÿÀŽ ä xrp€€€€ÿüŽ Ä xnnÿþþþÿàÿ â $xvG @  vŽ â xsrÿþþþÿàÿ ä xtt€€ÿÀŽ ä rrno   ÿüŽ â $xnv€   ÿ Ž ârosp   ÿüŽ â $xpv   ÿ Ž â x$t€€€€ÿüŽ Ä/ xGÿ   G  ä x%$ÿþþþÿàÿ ä xnn€€ÿÀŽ ä rGGG   ÿpŽ ä xsp€€ÿÀŽ ä rp%t   ÿüŽ â $xuG€   ÿ Ž âx$n€€€€ÿüŽ â $xvG   ÿ Ž â x%s€€€€ÿüŽ Ä xG$ÿþþþÿàÿ ä xt%ÿþþþÿàÿ ä x%u€€ÿÀŽ ä x$v€€ÿÀŽ ä rGGn   ÿüŽ ä rnts   ÿüŽ Ä xs%€€€€ÿüŽ ä xt$€€€€ÿüŽ ä xqq˜º  ÿ    ä xvsÿþþþÿàÿ ä xstÿþþþÿàÿ ä xrr˜º  ÿ    Ä xtqÿÀŽ ä xq €€€€qÞŽ äO xurÿÀŽ ä xr!€€€€rÞŽ ä r$s$   ÿüŽ â $xsq   t
Ž â xoo˜º  ÿ    Ä xqHÿ   ÿÀŽ ä xpp˜º  ÿ    â $xrr   u
Ž â xtoÿÀŽ ä ruHq   ÿpŽ ä xo€€€€oÞŽ ä� xqpÿÀŽ Ä xp€€€€pÞŽ ä x s€€€€ ´Ž â $xso   t
Ž ä $xou€   ÿ Ž ä$xtp   q
Ž â xqÿ   H  â $xpu   ÿ Ž â xHo€€ÿÀŽ Ä ruqq   ÿpŽ ä x!r€€€€!´Ž ä xrp€€ÿÀŽ ä xos€€€€´Ž ä xpH€€€€ÿüŽ ä xt€€€€´Ž â $xu€   ÿ Ž â xsr€€€€ÿüŽ Ä xqpÿþþþÿàÿ â $xpu   ÿ Ž â x€€ÿÀŽ ä xssÿþþþÿàÿ ä rqqH   ÿüŽ ä xHp€€ÿÀŽ ä xp€€€€ÿüŽ Ä rssr   ÿüŽ ä xtpÿþþþÿàÿ ä xqq˜º  ÿ    ä xps˜º  ÿ    ä xuH€€€€ÿüŽ ä xrqÿÀŽ Ä xspÿÀŽ ä xq€€€€qÞŽ äxp€€€€pÞŽ ä xuuÿþþþÿàÿ ä rtt   ÿüŽ â $xqq   r
Ž â ruuH   ÿüŽ â $xpp   s
Ž â xrKÿ   ÿÀŽ Ä xt˜º  ÿ    ä xHu˜º  ÿ    ä xp€€€€´Ž ä rsKr   ÿpŽ ä xpÿÀŽ ä xrÿ   K  Ä xq€€€€´Ž ä x€€€€ÞŽ ä xqHÿÀŽ ä xH€€€€HÞŽ â $xKs€   ÿ Ž ä $xtr @  rŽ ä $x   p
Ž â xps   ÿ   â $xHH   q
Ž â xKK€€ÿÀŽ â $xqt€   ÿ Ž â xpp€€ÿÀŽ Ä x€€€€´Ž ä xH€€€€´Ž ä xK€€€€ÿüŽ ä xrq€€ÿÀŽ ä xqp€€€€ÿüŽ â $xtt   ÿ Ž â xHÿþþþÿàÿ Ä xqqÿþþþÿàÿ ä rHHK   ÿüŽ ä xt€€ÿÀŽ ä rqqp   ÿüŽ ä xsr€€€€ÿüŽ ä xKH˜º  ÿ    Ä xt€€€€ÿüŽ ä xHq˜º  ÿ    ä xssÿþþþÿàÿ ä xpKÿÀŽ ä xttÿþþþÿàÿ ä xK€€€€KÞŽ Ä xqHÿÀŽ ä xH€€€€HÞŽ ä rssr   ÿüŽ ä rtt   ÿüŽ â $xK   p
Ž â xss˜º  ÿ    â $xHH   q
Ž â xnn˜º  ÿ    Ä x€€€€´Ž ä xpsÿÀŽ ä xtt˜º  ÿ    ä xs€€€€sÞŽ ä xH€€€€´Ž ä xGG˜º  ÿ    Ä xnÿÀŽ ä xn€€€€nÞŽ ä xKtÿÀŽ ä rssp  €ÿàÿ ä x$$˜º  ÿ    ä r%v%   ÿüŽ Ä xt€€€€tÞŽ ä xpGÿÀŽ â $xnn   
Ž â xG€€€€GÞŽ ä x$ÿÀŽ â $xHt   K
Ž â x%%˜º  ÿ    ä x$€€€€$ÞŽ â $xGG   p
Ž â xp%ÿÀŽ Ä r$$  €ÿàÿ ä x%€€€€%ÞŽ ä xHH€€€€´Ž ä xG€€€€´Ž ä xn€€€€´Ž â &rI   ÿ   â x$$€€€€´Ž â &rD   ÿ   Ä $x%%   p
Ž â xs€€€€´Ž â &rI I   ÿ   ä &rDD   ÿ   ä &rJ      ä&rU      â x%€€€€´Ž â &rI!J   I   Ä &rUU   D   ä &rF      ä &rB      â xPð   ÿÀŽ â&rIoF   I   â xPP   ÿÀŽ â &rUB   U   â xDLð   ÿÀŽ â &rV      Ä &rBHE      â xLL   ÿÀŽ â &rIV   I   â xÿ     â $rP   ÿŽ ä $x%   BŽ â xÿ   D  â &rF$E   U   ä $rDLI   ÿŽ Ä $rB   Ž â rB%%   ÿ� â $xII   FŽ ä $rDF   DŽ â rB   ÿú� â &rFN   ÿ   â rIII   ÿ� â &r%R   ÿ   â xBÿ     â &rFC   F   Ä &rN N   ÿ   â rDID   ÿú� â &r%S   %   â rB   ÿ� â &rRR   ÿ   ä &rFT   F   ä &rN!C   N   ä &rBM   %   â x%ÿ   D  â &rRS   R   Ä &rCO   F   ä &rNoT   N   â r%%D   ÿ� â &rRM   R   ä &rBHQ   B   ä $rDPC   ÿŽ ä &rOO   N   Ä &rR$Q   R   ä $xCC   BŽ ä $rDB   DŽ â xÿ     â $rBLO   ÿŽ ä 0rWÿW    A   ä 0rXÿX    A   Ä $xOO   RŽ â Er        â $rRR   BŽ â xBÿ   %  â  r%WX     A  â rCCC   ÿ� Æ #r.%   .   â rCCD   ÿú� â &rDY   ÿ   ä&rZ   ÿ   ä &rF Y   ÿ   ä &rD_   D   ä &r%[      Ä 0r(ÿ(    A   ä &rZZ   ÿ   ä &rF!_   F   â ErB B       â &r\   D   ä &r%^   %   â  rXX(     A  â &r[[   Z   Ä &r\o\   F   ä &r)      ä &rFH]   %   â #r-XB   -   â &r[^   [   â xBÿ   C  â $r%P   ÿŽ â rF   ÿàÿ â &r))   \   Ä &rJ$]   [   â rBBC   ÿ� â $r%F   %Ž â rOOO   ÿ� â $rCL)   ÿŽ â rF   ÿ� â $x))   JŽ â rOOR   ÿú� â 0rÿ    A   â r%F%   ÿú� â $rCJ   CŽ â xBÿ   B  Ä rF))   ÿ� â  rW     A  â xDÿ   O  ä ErB B       ä rCFC   ÿú� ä xFÿ   %  â #r,B   ,   â rDDO   ÿ� â &ra   ÿ   ârFF%   ÿ� â &r%c   ÿ   â xDÿ   D  â &ra a   ÿ   Ä &rBb      ä &r"   ÿ   â ErD D       â &rcc   ÿ   â xFÿ   F  â &rf   %   ä &r  "   ÿ   â  r(     A  â 0r`ÿ`    A   Ä &ri   ÿ   ä&ri   ÿ   ä &rb!b   a   ä &r%g   B   ä &rl      â ErF F       â &rcf   c   Ä &r)e      ä &r !l       â #r+D   +   â &rk      â rW`     A  â &rk      â xÿ   C  â &rgog   b   ä &r%d   %   Ä &rj      ä&rce   c   ä &rBH   )   ä &r oj       â #r*F   *   â &r'      ä&r'      â rC   ÿ� â &rgd   g   Ä $rP%   ÿŽ ä &r#      ä &rD$   c   ä $x%%   BŽ ä &r##       ä &rHH&      Ä &r$$&      â xÿ     â $r)Lg   ÿŽ ä$rB   Ž â rB%%   ÿ� â $rLL#   ÿŽ ä $xgg   DŽ ä $rPP   ÿŽ â rH   ÿàÿ â $x##   $Ž â Er        â  r(`     A  â $r)D   )Ž â rB   ÿú� â $rPH   PŽ â rDgg   ÿ� â $rL$   LŽ â r   ÿ� Ä r###   ÿ� â #r        â xÿ     ä r)D)   ÿú� ä rP   ÿú� ä r##L   ÿú� ä x22   ÿàÿ Ä r   ÿ� ä xÿ   )  ä xÿ   #  ä xÿ     â 0rhÿh    A   â | 2   pbô ä xÿ     Ä r)   ÿ� ä r#   ÿ� ä r   ÿ� â  rWh     A  â Er        â 0rmÿm    A   â xÿ     ä xÿ     Ä xÿ     â #rA   A   â Er        â  r(h     A  âEr        â  r(m     A  âEr        â  rWm     A  â x33�  ÿà÷ â #r@   @   â #r4   4   â #r1   1   â x66@   ÿàÿ â $r9ÿÿ   9Ž â x88@   ÿàÿ Ä x::@   ÿàÿ ä x;;@   ÿàÿ ä x==@   ÿàÿ â D)       À â Gy   ßÿÿÿÿƒ ê Ay        € ê� Ey  Ð     € â G‰  °     € ê $x//@  0Ž Ê ˆs /. ûÿ H   è ˆs /-€ûÿ H   è ˆs /, üÿ H   è ˆs /+€üÿ H   è ˆs /* ýÿ H   è ˆs / €ýÿ H   è ˆs /A þÿ H   è ˆs /@€þÿ H   è ˆs /1 ÿÿ H   è ˆs /4€ÿÿ H   ä Ay        € ê {          ì M        € ê „y0     H   b z 5 r  p`ð â $v!ÿ r  ÿ Ž Ä „y0 €   H   ¢ r 0ÿ   p"p â $x!   ÿŽ ä „y0    H   è „y0 €  H   ( „y	0    H   h „y0 €  H   ( „y
0    H   ( „y0    H   ( „y0 €  H   è „y0 €  H   â !r.       â/ !r-       ÈO ‰      b !r,,       ÆŸ ‰      " !r+       â!r*	*       ä‰	,      ¢ !r         Æ ‰      â !r

1       â !rA       â !r       Æ/ ‰
      b !r@@       Æ ‰*      " !r       â !r44       ä ‰      ( ‰       b !r	,	       ÆO ‰      ¢ !r       Æ� ‰@      è ‰4      â !r
       â/ !r*       ä‰	      b !r       Æ ‰
      " !r        Æ‰      " !r       ÆO ‰  €    ¢ !r@       Æ� ‰      â !r4       Æ ‰      h ‰      b !r	       Æ/ ‰      b !r
       â x
5   ÿàÿ ä ‰      " !r       â$z x  5Ž â z 
 r  p`ò â ‰      " !r         äO x 0   p"ò â !r       â� ‰ €    ¢ !r       Æ‰	 €    â !r       Æ y      '   b !r       Æ/ ‰ €    b !r       Æ ‰  @    " !r       Æ‰ €    ( ‰ €    " !r       ÆO ‰ €    ¢ !r	       Æ� ‰ €    â $z
 ~  Ž Æ‰ €    b !r       â/ r!
   ÿ† ä‰ €    b !r         â 	
   ÿàõ â !r       â$rÿÿ   ÿŽ â ‰ @    " z j  ÿ† â !r       â $ ÿÿ   ÿ â ‰ @    " z k  � â !r       âO 
	 j  ÿ„ â ‰ @    ¢ r!
   ÿŠ â !r       â� 
		 k    â ‰       â 

 r  ÿàù â !r       â$r"ÿÿ   ÿŽ â ‰ @    b z j  ÿŠ â!r       â/ $ÿÿ   ÿä â ‰ @    b z k  "� Ä 
 j  ÿˆ â ‰ @    b !r       â    ÿàÿ ä ‰ @    " !r       â
   ÿà÷ ä ‰ @    " !r       âO 
 k  $ â $ ÿÿ   ÿŽ â ‰ @    ¢ !r#        â� !   ÿàÿ â !r       â‰       ä !
   ÿàõ â !r       â/ ‰      b  j  ÿ† Ä $"ÿÿ   ÿ â !r       â x        â ‰      b ! j  ÿ„ â !r       â  k   � â ‰      " %v

 j   Ž â !r       ä‰      " !r       ÆO ‰      ¢ !r%        Æ� ‰      â ! k  " â !r       ä/ ‰      h ‰      b !r       â!r       Ä †	 
#    â !r       Æ† 
%   â !r       ÆO †	     â !r       Æ� †     â !r       Æ/ †	     â !r       Æ †    è †	     è †     è †	     â M™        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y)      "   " ¹z  l      â Ey  P!    € â ™x?    â y*      !   b ¹z  F   
   â $r9ÿÿ   ÿ Ž â r/ ÿ       â y.      %   ¢ $r0ÿÿ   ÿ Ž â x&      ÿ  â $r,ÿÿ   ÿ Ž Ä $r ÿÿ   ÿ Ž ä $r(ÿÿ   ÿ Ž â r )ÿ   pRð är )ÿ   pRò â$x)    *Ž ä/ $x..   ÿ Ž ÆO xÿ     È r   ÿ� È x;ÿ     È | ;   pbô Ú G)        € ê y      '   " $rÿÿ   ÿ Ž â xøÿÿÿÿÀŽ â $rÿÿ   ÿ Ž â y      &   b $z	. p  ;Ž âx.   ÿàÿ â $vÿ q  ÿ Ž ä $x6;   ÿ Ž ä $v+ÿ X  ÿ Ž Ä $x2   6Ž âr-6   ÿŽ â $v3ÿ Y  ÿ Ž â z56 q  ÿàÿ â $r9ÿÿ   ÿ Ž ä 'z y   Ž È 'z s   Ž â/ zÿ z    Æ $z1 }  ÿŽ â zÿ t    â $z |  ;Ž ä$z	 |  	Ž â r  €ÿàÿ â $z v  Ž ä$z v  	Ž ä $z p  Ž â xÿ     â $z
 w  ÿŽ Ä %xR   ÿŽ Æ r1
1   ÿàõ â %xR   Ž â r4   ÿà÷ Æ %x
R   ÿŽ È $x   ÿ Ž ä $r7ÿÿ   ÿ ä %x   Ž È $x8   Ž ä %x   
Ž È rB+   ÿàõ Ê $xC	   3 â r+   ÿàõ È �y BB   ‘ ¢  $x   3 Ê �y   • â r+4   ÿàõ Æ �y   • " r38   ÿä ä r+   ÿàõ â �y_"   • h �yB   ‘ b $x   3 Æ �y^$   • ( �y   • ( �y   • " $t"ÿ   ÿ Ž Æ �y]"   • ( �y\$   • b r$1   6àµ â $tWÿ$   ÿ Ž ä �yEJ   ‘ d %v$$ Z  W Ž Ä �yBBJ   ‘ d $x    ÿ Ž ÊO x   ÿÀŽ È xÿ   øŽ â�$x @   ÿ Ž â xÿ     Æ %v 
 " Ž â x   øŽ Ê �y    › ¢ %v 
 " Ž Ì �y    › â $x   ÿ Ž Ê/ x   ÿÀŽ È xÿ   øŽ äxÿ     Æ %v 
 " Ž â x   ÿ   È x   øŽ â �y    › ( %v 
 " Ž Ì �y    › â $x   ÿ Ž Ê x   ÿÀŽ â $x   ÿ Ž â xÿ     äxÿ   øŽ ä x   øŽ Æ %v 
 " Ž È $x    ÿ Ž ä �y    › " %v 
 " Ž ä x   ÿÀŽ È �y    › " xÿ   øŽ â$x    ÿ Ž â xÿ     Æ %v 
 " Ž â x   øŽ ä r 7   ÿd~ Æ �y    › " %v 
 " Ž È $x! $   ÿŽ ä �y    › & r%%!   ÿàÿ ä r 1   5àµ Æ �y<$   ™ ( �y=$   ™ " r"7   ÿd~ Æ �yA$   ™ ( �y@$   ™ ( �y:$   ™ " %v   Z  W Ž Æ �y>$   ™ " $x#"$   ÿŽ Æ �y?$   ™ " $x!!   #Ž Æ �yJ$    ™ ( �yH    ™ ( �yD    ™ ( �yL    ™ ( �yI    ™ ( �yF    ™ ( �yM    ™ ( �yK    ™ ( �yC     ™ " rR+   ÿà÷ Ä r"
+   ÿàõ â �y$$    • $! $xS   3Ž â r1   -àÇ â $x#   3 â �yN     • $ rO7   ÿ„þ â %v Z  W Ž â �yGR    • & $xOO$   ÿŽ â �y""    • & $x   OŽ â rU1   2àµ È �yO   ™ ( �yQ   ™ " rV7   ÿd~ Æ �yP   ™ ( �yS   ™ "# %v U Z  W Ž Æ�yT   ™ b $xUV$   ÿŽ Æ �y%   ™ h �yR   ™ b $x!!   UŽ Æ �y#    ™ h �yV    ™ b xX_ÿ   ÿÀŽ Æ �yU    ™ h �y[    ™ b rb_X   ÿpŽ Æ �yZ    ™ h �yY    ™ b xaÿ   _  Æ �yX    ™ h �yW    ™ h �y_     ™ b $x`b€   ÿ Ž Æ �y    • d# xc`€€ÿÀŽ â $xga @  aŽ Æ x`c€€€€ÿüŽ â $xab   ÿ Ž Æ xb`ÿþþþÿàÿ ä �y`     • b $xdg€   ÿ Ž âxi^ÿ   ÿÀŽ â $xgg   ÿ Ž Æ xed€€ÿÀŽ ä xaa€€ÿÀŽ ä xde€€€€ÿüŽ ä ri^i   ÿpŽ ä xgg€€ÿÀŽ ä rcbc   ÿüŽ â $x!i€   ÿ Ž â xba€€€€ÿüŽ Ä xfdÿþþþÿàÿ ä x^ÿ   ^  ä x g€€€€ÿüŽ ä xbbÿþþþÿàÿ ä rfe   ÿüŽ â/ $xf^ @  ^Ž â x  ÿþþþÿàÿ Ä x!!€€ÿÀŽ ä rdba   ÿüŽ â $xbf€   ÿ Ž â rg g   ÿüŽ â $xff   ÿ Ž â x !€€€€ÿüŽ â $xai   ÿ Ž Æ x  ÿþþþÿàÿ ä xhf€€ÿÀŽ ä xeb€€ÿÀŽ ä rb !   ÿüŽ ä x!h€€€€ÿüŽ ä xcc˜º  ÿ    Ä xdd˜º  ÿ    ä xaa€€ÿÀŽ ä xi!ÿþþþÿàÿ ä xfcÿÀŽ ä x!dÿÀŽ ä xc€€€€cÞŽ ÄO x^a€€€€ÿüŽ ä xd€€€€dÞŽ ä rccf  €ÿàÿ ä x^^ÿþþþÿàÿ â $xdd   !
Ž â x e€€€€ÿüŽ ä x!˜º  ÿ    Ä xfg˜º  ÿ    ä xc€€€€´Ž ä ra^a   ÿüŽ ä x  ÿþþþÿàÿ ä x]ÿ   ÿÀŽ ä r^ih   ÿüŽ Ä xh!ÿÀŽ ä xcfÿÀŽ ä x!€€€€!ÞŽ ä� xf€€€€fÞŽ ä r  e   ÿüŽ â $xe] @  Ž ä $x!!   h
Ž Ä $xff   c
Ž â xcÿ   ]  â $xe€   ÿ Ž âx]e   ÿ   ä x!!€€€€´Ž â $xec @  cŽ â x€€ÿÀŽ ä x]]€€ÿÀŽ ä xd€€€€´Ž â $xde€   ÿ Ž â xf€€€€´Ž Ä x€€€€ÿüŽ ä xc]€€€€ÿüŽ â $xfe   ÿ Ž â xÿþþþÿàÿ ä xed€€ÿÀŽ ä xdcÿþþþÿàÿ ä r   ÿüŽ Ä xcf€€ÿÀŽ ä rdd]   ÿüŽ ä xe€€€€ÿüŽ ä x˜º  ÿ    ä x]c€€€€ÿüŽ ä xdd˜º  ÿ    Ä xfÿþþþÿàÿ ä xÿÀŽ ä xh]ÿþþþÿàÿ ä x€€€€ÞŽ äx]dÿÀŽ ä xd€€€€dÞŽ Ä rffe   ÿüŽ â $x   
Ž â rdd]  €ÿàÿ ä rhhc   ÿüŽ ä x]f˜º  ÿ    ä xe\ÿ   ÿÀŽ ä x€€€€´Ž Ä xhh˜º  ÿ    ä x]ÿÀŽ ä x]€€€€]ÞŽ â $xe\ @  eŽ â xchÿÀŽ ä x\ÿ   \  â $x]]   
Ž â xh€€€€hÞŽ â $xe€   ÿ Ž âxe   ÿ   â $x\\ @  \Ž Ä $xhh   c
Ž â x€€ÿÀŽ â $xe\   ÿ Ž â x€€ÿÀŽ â $xc\€   ÿ Ž â x]]€€€€´Ž ä xh€€€€´Ž ä x€€€€ÿüŽ ä x\€€€€ÿüŽ Ä xd€€€€´Ž ä xfe€€ÿÀŽ ä xdc€€ÿÀŽ ä xÿþþþÿàÿ ä xc\ÿþþþÿàÿ ä xgf€€€€ÿüŽ Ä x\d€€€€ÿüŽ ä r   ÿüŽ ä rcc   ÿüŽ ä xggÿþþþÿàÿ ä xe\ÿþþþÿàÿ ä x˜º  ÿ    Ä xc˜º  ÿ    ä rggf   ÿüŽ ä reed   ÿüŽ ä xÿÀŽ ä xcÿÀŽ ä x€€€€ÞŽ Ä x€€€€ÞŽ ä x\g˜º  ÿ    ä xee˜º  ÿ    â $x   
Ž â rc  €ÿàÿ ä xc\ÿÀŽ ä xeÿÀŽ Ä x\€€€€\ÞŽ ä xe€€€€eÞŽ ä x€€€€´Ž â $x\   c
Ž ä $xee   
Ž â xa˜º  ÿ    ä x€€€€´Ž Ä xÿÀŽ ä x€€€€ÞŽ ä xa ˜º  ÿ    ä x^^˜º  ÿ    â $x   
Ž â x€€€€´Ž ä xb˜º  ÿ    Ä x aÿÀŽ ä x^ÿÀŽ ä xa€€€€aÞŽ ä x^€€€€^ÞŽ ä x\ÿÀŽ ä x€€€€ÞŽ Ä x€€€€´Ž â $x^   
Ž â raa   €ÿàÿ â &r <   ÿ   ä &r=   ÿ   â xe€€€€´Ž â $x   \
Ž ä &r A       ä &r@      â x€€€€´Ž â &r<<   ÿ   Ä &r ]:       ä &r>      â xEð   ÿÀŽ â &r<A   <   â xEE   ÿÀŽ â &r ?       ä &r==   ÿ   â x€€€€´Ž â &r<!:   <   â xÿ     â &rJ      â xa€€€€´Ž â $r:E    ÿŽ Ä &r=@   =   ä $x     Ž ä $r:   :Ž ä &r=>   =   ä &rH   ÿ   ä &r>D   ÿ   Ä &rL      ä &r>I   >   ä &r??   <   â x<Bð   ÿÀŽ â &rJJ   =   â xBB   ÿÀŽ â &r=]F      ä &r>M   >   â r     ÿ� â &r K   =   â xÿ   <  â $r<B?   ÿŽ Ä &r>C   >   â r:   ÿú� â $x??   JŽ ä $r:E    ÿŽ ä $x     >Ž â r???   ÿ� â $r<J   <Ž ä $r:>   :Ž â r=     ÿ� Ä r<?<   ÿú� ä x ÿ     ä r:=:   ÿú� ä x=ÿ   <  ä r     ÿ� â &rHH   ÿ   â r==<   ÿ� â &rDD   ÿ   â x?ÿ   :  â 0rGÿG    A   â x ÿ      â 0r$ÿ$    A   â x=ÿ   =  â 0r"ÿ"    A   â r??:   ÿ� â &rHL   H   â Er          â &rDI   D   â  rG$     A  â Er= =       â &rH!F   H   â  r$$"     A  â &rDM   D   â x?ÿ   ?  â 0rNÿN    A   â #r(    (   â &r O   ÿ   â #r'$=   '   â &rKK   H   â Er? ?       â &rDC   D   â  rGN     A  â &r$Q   ÿ   Ä &r:Q   ÿ   ä &r P       â #r&?   &   â $r>BK   ÿŽ â rKKD   ÿàÿ â &r$S   $   ä&r:S   :   ä &r]T       ârKKK   ÿ� â $r>D   >Ž Ä &r$%   $   ä&r:%   :   ä &r R      â r>K>   ÿú� â &r$#   $   ä&r:#   :   ä $r#E    ÿŽ â xAÿ   >  â $x     $Ž Ä $r#$   #Ž ä &rOO   ÿ   ä&r$V   ÿ   â rAA>   ÿ� â &rVV   ÿ   ä &rU   ÿ   ä&rU   ÿ   Ä &rOP   O   âxAÿ   A  â &r$[   $   ä&rV[   V   ä &rZ      ä&rZ      ä &rO!T   O   âErA A       â &r$!Y   $   Ä &rV]Y   V   â  r"N     A  â &rX      ä&rX      ä &rOR   O   ä&rW   $   ä&rVW   V   â #r A       â &r_      Ä &r_      ä $rBO   ÿŽ ä$xOO   :Ž ä $rBB   ÿŽ ä $rEV   ÿŽ ä $x   Ž Ä $xVV   Ž â rOO   ÿ� â $r:   Ž âr      ÿ� â $rB   BŽ â r   ÿ� â $r   Ž â rVV   ÿ� ä r   ÿú� ä r  #   ÿú� Ä rBB   ÿú� ä r   ÿú� ä x;;   ÿàÿ ä xÿ     ä xÿ      ä xÿ   B  Ä xÿ     ä | ;   pbô ä r   ÿ� ä r    ÿ� ä rB   ÿ� ä r   ÿ� â 0rÿ    A   â xÿ     â 0r`ÿ`    A   â xÿ     Ä xÿ     ä xÿ     â  rG     A  âEr        â  r"     A  âEr        â  r""`     A  âEr        â  rGG`     A  â Er        Ä x++   ÿà÷ â #r/   /   â #r,   ,   â #r9"   9   â #r0G   0   â x--€   ÿàÿ ä x22€   ÿàÿ ä x55€   ÿàÿ ä x66€   ÿàÿ Ä r3ÿ3   ÿäÿ â D)       À â Gy  `âÿÿÿÿƒ ê Ay        € ê Ey  °     € â G‰  �     € ê $x))   *Ž Ê ˆs )( üÿ H   è ˆs )'€üÿ H   è ˆs )& ýÿ H   è ˆs ) €ýÿ H   è ˆs )/ þÿ H   è ˆs ),€þÿ H   è ˆs )0 ÿÿ H   è ˆs )9€ÿÿ H   ä Ay        € ê {          ì M        € ê „y* €  H   b z . r  p`ð Æ „y
*    H   ¢ r *ÿ   p"p Æ „y*     H   è „y* €  H   ( „y*    H   h „y* €   H   ( „y*    H   ( „y*    H   ( „y* €	  H   ( „y* €  H   ( „y*    H   " !r        Æ/ „y* €  H   b !r

/       ÆO „y*    H   ¢ !r(       Æ� „y*    H   â !r       Æ„y	* €  H   " !r

       Æ„y* €  H   b !r'       Æ „y*  	  H   " !r       Æ „y*    H   " !r&       Æ „y*  
  H   " !r       Æ „y* €
  H   " !r       Æ „y*    H   " !r       Æ „y* €  H   â !r,       Æ/ „y* €  H   â !r       ÆO „y* €  H   â !r0       â� !r	       ä‰      b !r       Æ‰      ¢ !r       Æ ‰	      " !r       Æ ‰      â $vÿ r  ÿ Ž â !r

       â !r       È ‰
      " !r       Æ ‰      b !r       â/ !r9       ä ‰      b !r       âO !r       ä ‰       ¢ !r		       â !r       â !r       â� ‰      ( ‰      â !r
       Æ‰	      " !r       Æ‰      b !r       Æ/ ‰
      b !r         ÆO ‰      ¨ ‰      ¢ !r       Æ ‰  €    " !r       â� !r	       ä‰ €    â !r       Æ‰ €    " !r
       Æ/ ‰	 €    b !r       ÆO ‰ €    ¢ !r       Æ ‰ €    b !r        Æ ‰ €    ( ‰      " !r       Æ� y      &   â !r       Æy      '   " !r
	       Æ/ ‰ @    b !r       âO !r       ä‰ @    ¢ !r       Æ ‰ @    " !r       Æ ‰
 @    b $z  x  .Ž Æ� ‰ @    â x..   ÿàÿ â $z  ~   Ž ä‰ @    " z . r  p`ò â $x   ÿŽ â !r       â/ ‰ @    b 
  r  ÿàõ ä x *   p"ò â ‰ €    b !r	       âO $ÿÿ   ÿ â 
 j  ÿ„ â ‰      ¢ !r       â r    ÿ† Ä 
 k   â !r
       â‰	      " $rÿÿ   ÿŽ â !r       â�    ÿàÿ â ‰      â z j  ÿ† â!r       â    ÿàõ â ‰      " z k  � â !r       â/     ÿà÷ â ‰      b $ÿÿ   ÿ â !r       â  j  ÿ„ â ‰      b $ÿÿ   ÿŽ â !r       âO x        â ‰ @    ¢ 

 j  ÿ† Æ ‰      ¢ !r	       â %v  j   Ž â 	 k   â !r       â� 
 k  � ä †	     â !r       Æ†    â !r       Æ/ †	     â !r       Æ†     â !r       ÆO †	     â !r       Æ †    è ‰      h  †	 
    â M™        € ê x   ÿàÿ â !r       Æ/ r    ÿàñ Ê $r ÿÿ   ÿ  â z j  ÿ€ È z k     Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      "   " ¹z  l      â Ey  à    € â ™x?    â y)      !   b ¹z  F   
   â $r0ÿÿ   ÿ Ž ä y.      %   ¢ $r*ÿÿ   ÿ Ž ä $r8ÿÿ   ÿ Ž Ä $r,ÿÿ   ÿ Ž ä $r ÿÿ   ÿ Ž ä $r(ÿÿ   ÿ Ž â r ÿ   pRð är ÿ   pRò â$x    )Ž ä/ $x..   ÿ Ž ÆO xÿ     È r   ÿ� â $/À   )Ž Æ x;ÿ     È | ;   pbô Ú G)  �    € ê y      '   " $rÿÿ   ÿ Ž â xøÿÿÿÿÀŽ â $rÿÿ   ÿ Ž â y      &   b $z	. p  ;Ž âx.   ÿàÿ â $x5;   ÿ Ž ä $v9ÿ X  ÿ Ž ä $v3ÿ Y  ÿ Ž â z45 q  ÿàÿ â $r0ÿÿ   ÿ Ž Ä 'z y   Ž È 'z s   Ž â/ zÿ z    Æ $z1 }  ÿŽ â zÿ t    â $z |  ;Ž ä$z	 |  	Ž ä $z v  Ž ä $x   
Ž ä $z v  	Ž Ä $z p  Ž â xÿ     â $z
 w  ÿŽ ä %xR   ÿŽ Æ r1
1   ÿàõ â %xR   Ž â r2   ÿà÷ Æ %x
R   ÿŽ È $vÿ q  ÿ Ž ä $x   ÿ Ž ä $x+   5Ž ä %x   Ž È $r7ÿÿ   ÿ ä $x6   Ž ä %x   
Ž È r$9   ÿàõ Ê $x%	   3 â r9   ÿàõ È �y$B   ‘ ¢ $x   3 ä $tÿ   ÿ Ž â �yG$J   ‘ è  �y   • " r"92   ÿàõ Æ �y   • d $x#3   6 â r@9   ÿàõ â �y?"   • è �y"B   ‘ b $xA   3 Æ �y=$   • è �y@   • è  �y@   • è  �y<@"   • â  $t&ÿ$   ÿ Ž Æ �yE"J   ‘ è �y@@$   • b r1   5àµ È r'7   ÿd~ Ê $x''$   ÿŽ ä $x   ÿ Ž ÊO x   ÿÀŽ È xÿ   øŽ äxÿ     Æ %v 
  Ž È $x@   ÿ Ž ä �y    › ¦ x   øŽ Ê %v 
  Ž Ì �y    › " $x   ÿ Ž Ê/ x   ÿÀŽ È xÿ   øŽ ä�xÿ     Æ %v 
  Ž È $x@   ÿ Ž ä �y    › æ x   øŽ Ê %v 
  Ž Ì �y    › â x   ÿ   È x   ÿÀŽ â $x   ÿ Ž Æ xÿ   øŽ äxÿ     È x   øŽ â $x   ÿ Ž ä %v 
  Ž Æ x   ÿÀŽ â %v 
  Ž ä�y    › â x!ÿ   øŽ â$x   ÿ Ž â xÿ     â �y    › d %v ! 
  Ž â x   øŽ È �y     › â %v 
  Ž È %v Z  & Ž ä �y    › d $x   'Ž Ê �y>   ™ h �yD   ™ h �y-   ™ h �yC   ™ h �yA   ™ h �yB   ™ h �y:   ™ h �yF    ™ b r!1   4àµ È/ r'7   ÿd~ â %v ! Z  & Ž È $x''$   ÿŽ â r"9   ÿà÷ â �y    • d $x!!   'Ž ä $x#   3Ž Æ �yI    ™ b r'1   +àÇ Æ �yH    ™ b rP7   ÿ„þ Æ �yL    ™ h �yK    ™ h �yO    ™ b %v&' Z  & Ž Æ �yN    ™ b $xQP$   ÿŽ Æ �yM    ™ b $x''   QŽ Æ �yJ     ™ h �yQ&   ™ h �yP&   ™ h �y""    • h �yU&   ™ h �yT&   ™ h �yS&   ™ h �yR&   ™ h �y&   ™ h �y#&    ™ b r$
9   ÿàõ Æ �y      • d# $x%   3 Ê �y$$    • h �y%&    • b xV?ÿ   ÿÀŽ Ê $xX? @  VŽ â x?ÿ   ?  Æ $xVX€   ÿ Ž ä $xZ? @  ?Ž Æ xVV€€ÿÀŽ â $xXX   ÿ Ž ä $xYZ€   ÿ Ž â x?V€€€€ÿüŽ ä xXX€€ÿÀŽ ä xW?ÿþþþÿàÿ ä x?Y€€ÿÀŽ Ä rWWV   ÿüŽ â $xVZ   ÿ Ž â x!X€€€€ÿüŽ ä/ x&?€€€€ÿüŽ ä xZ=ÿ   ÿÀŽ ä x!!ÿþþþÿàÿ ä x&&ÿþþþÿàÿ â $xZ= @  ZŽ â xVV€€ÿÀŽ Ä rX!X   ÿüŽ ä r?&?   ÿüŽ â $x&Z€   ÿ Ž â x!V€€€€ÿüŽ â $xYZ   ÿ Ž â x=ÿ   =  ä x'!ÿþþþÿàÿ ä x!&€€ÿÀŽ Ä x&Y€€ÿÀŽ â $x\= @  =Ž â rV'V   ÿüŽ ä x=&€€€€ÿüŽ â $x[\€   ÿ Ž â x'!€€€€ÿüŽ ä xY=ÿþþþÿàÿ ä xZ'ÿþþþÿàÿ Ä x\\   ÿ   ä x'[€€ÿÀŽ ä r&Y&   ÿüŽ ä xYW˜º  ÿ    ä x]X˜º  ÿ    ä x=\€€ÿÀŽ Ä x['€€€€ÿüŽ ä r!Z!   ÿüŽ ä xWYÿÀŽ ä xZ€€€€YÞŽ äO xX]ÿÀŽ ä xY€€€€]ÞŽ Ä x\=€€€€ÿüŽ ä x[[ÿþþþÿàÿ â $xXY   X
Ž â x\\ÿþþþÿàÿ ä r'['   ÿüŽ â $x[Z   W
Ž â xY?˜º  ÿ    ä xW<ÿ   ÿÀŽ Ä r=\=   ÿüŽ ä x\YÿÀŽ â $x?< @  WŽ â xZV˜º  ÿ    ä xY€€€€YÞŽ äx[€€€€´Ž â $xW?€   ÿ Ž â xVZÿÀŽ â $xYY   \
Ž â x[€€€€ZÞŽ Ä x<ÿ   <  â $x??   ÿ Ž â xX€€€€´Ž â $xX[   V
Ž â xVW€€ÿÀŽ ä xY€€€€´Ž â $xY< @  <Ž â x??€€ÿÀŽ ä xWV€€€€ÿüŽ Ä x<X€€€€´Ž â $xY€   ÿ Ž â xX?€€€€ÿüŽ â $xYY   ÿ Ž â xWWÿþþþÿàÿ ä x€€ÿÀŽ ä xXXÿþþþÿàÿ ä rWWV   ÿüŽ Ä xVY€€ÿÀŽ ä xY€€€€ÿüŽ ä r?X?   ÿüŽ ä xWW˜º  ÿ    ä x[V€€€€ÿüŽ ä xXYÿþþþÿàÿ Ä xZ?˜º  ÿ    ä xYWÿÀŽ ä x\€€€€WÞŽ ä� xW[ÿþþþÿàÿ ä rXX   ÿüŽ ä x?ZÿÀŽ Ä xZ€€€€ZÞŽ ä rWWV   ÿüŽ â $xY\   Y
Ž â xX˜º  ÿ    â $xZZ   ?
Ž â xVW˜º  ÿ    ä xXÿÀŽ ä x?@ÿ   ÿÀŽ Ä x€€€€ÞŽ ä xY€€€€´Ž ä xYVÿÀŽ ä xV€€€€VÞŽ â $x?@ @  ?Ž â xZ€€€€´Ž â $xW   X
Ž â xZÿ   @  â $x@V   Y
Ž Ä $xV?€   ÿ Ž â xW€€€€´Ž â $xZ @  ZŽ ä $xW?   ÿ Ž â x?V€€ÿÀŽ â $xX€   ÿ Ž â x@@€€€€´Ž ä xVW€€ÿÀŽ â $x   ÿ Ž â xW?€€€€ÿüŽ Ä xX€€ÿÀŽ ä xYV€€€€ÿüŽ ä xXWÿþþþÿàÿ ä xW€€€€ÿüŽ ä x€€ÿÀŽ ä xYYÿþþþÿàÿ Ä xZWÿþþþÿàÿ ä xW€€€€ÿüŽ ä rYYV   ÿüŽ ä rXX?   ÿüŽ ä rZ   ÿüŽ ä xZWÿþþþÿàÿ Ä xVY˜º  ÿ    ä x?X˜º  ÿ    ä rZZ   ÿüŽ ä xWVÿÀŽ ä xX?ÿÀŽ ä xV€€€€VÞŽ Ä x?€€€€?ÞŽ ä xZZ˜º  ÿ    â $xVV   W
Ž â x˜º  ÿ    ä r?X  €ÿàÿ ä x?ZÿÀŽ ä x!!˜º  ÿ    Ä xXÿÀŽ ä xZ€€€€ZÞŽ äx€€€€´Ž ä x€€€€ÞŽ ä xV€€€€´Ž ä x&&˜º  ÿ    Ä xV!ÿÀŽ â $xZZ   ?
Ž â x!€€€€!ÞŽ â $x   X
Ž â x?&ÿÀŽ ä x&€€€€&ÞŽ â $x!   V
Ž â x''˜º  ÿ    ä x€€€€´Ž â $x&   ?
Ž â x&'ÿÀŽ Ä x€€€€´Ž ä x'€€€€'ÞŽ ä x=˜º  ÿ    ä x€€€€´Ž â $x'   &
Ž â xVÿÀŽ ä x!€€€€ÞŽ â &r>   ÿ   â x€€€€´Ž â &r-   ÿ   Ä &rD      ä $xV!   V
Ž ä &r!C      ä &rA      â xZ€€€€´Ž â &r>>   ÿ   ä &r@B   !   Ä &r--   ÿ   ä &r!:      â xEð   ÿÀŽ â &r>D   >   â xEE   ÿÀŽ â &r-A   -   ä &r!F   !   â xV€€€€´Ž â &rCC   >   â xÿ     â $r&E   ÿŽ â x'Gð   ÿÀŽ â &r-:   -   Ä $x   !Ž â xGG   ÿÀŽ â &rB<B   C   ä &r-F   -   ä $r&!   &Ž â r!   ÿ� ä xÿ   '  â $r'GB   ÿŽ Ä $xBB   -Ž â r!!&   ÿú� â $r&-   'Ž Æ r'BB   ÿ� È r&'&   ÿú� â &r'I   ÿ   ä&rII   ÿ   ä &r-H   ÿ   ä &r'L   '   ä &rHH   ÿ   â x:ÿ   !  â &rIL   I   Ä &r-K   -   ä &r'O   '   ä &rHK   H   â r!:!   ÿ� â &rIO   I   ä &r-N   -   ä&r:@M   '   â x'ÿ   &  â &rHN   H   Ä &r><M   I   â r&'&   ÿ� â &r-J   -   ä&r?J   H   ä $r'G>   ÿŽ ä $r=E:   ÿŽ ä $x::   -Ž Ä $x>>   ?Ž ä &rQ   ÿ   ä&rP   ÿ   ä&rQ   ÿ   â x!ÿ   !  â $r??   'Ž ä &rP   ÿ   Ä 0r"ÿ"    A   ä 0r'ÿ    A   â r:::   ÿ� â $r=-   =Ž ä &rU      ä&rT      ä&rU      â Er !       â &rT      â  r!"'     A  â r>>>   ÿ� â &rS      â r::=   ÿú� â &rR      Ä &rS      ä &rRR      â #r(!   (   â r>>?   ÿú� â &r@      âx!ÿ   :  â &r<      ä &r#      ä&r#   R   â xÿ   >  â $rEE   ÿŽ â r!!:   ÿ� â $rGG   ÿŽ Ä $x   Ž ä $x   Ž â r>>   ÿ� â 0rÿ     A   â x!ÿ   !  â $rE   EŽ â r   ÿ� â $rG   GŽ â r   ÿ� â  r"     A  â Er! !       Ä rG   ÿú� ä rE   ÿú� â #r,!   ,   â x;;   ÿàÿ ä xÿ     ä xÿ     â 0r$ÿ$    A   â x>ÿ   >  Ä | ;   pbô ä r   ÿ� ä r   ÿ� â  r$     A  â Er> >       â 0r%ÿ%    A   â x&ÿ   &  ä xÿ     Ä xÿ     â #r8>   8   â Er& &       â  r$%     A  â Er        â  r''$     A  â Er        â  r%"%     A  â x99   ÿà÷ â #r '&       â #r0   0   â #r*%   *   â x++€   ÿàÿ â $r3ÿÿ   3Ž â x44€   ÿàÿ Ä x55€   ÿàÿ â D)       À â Gy  Àåÿÿÿÿƒ ê Ay        € ê ˆ /( ýÿ H   è ˆ / €ýÿ H   è ˆ /, þÿ H   è ˆ /8€þÿ H   è ˆ /* ÿÿ H   è ˆ /0€ÿÿ H   è {          ì M        € ê „y)     H   " z . r  p`ð Æ „y)    H   b r )ÿ   p"p Æ „y) €   H   ¨ „y)    H   è „y) €  H   ( „y)    H   h „y)    H   h „y) €  H   h „y)    H   h „y) €  H   h „y
)    H   b !r(       Æ „y	) €  H   " !r       Æ/ „y)    H   b !r        ÆO „y) €  H   ¢ !r,       Æ� „y) €  H   â !r8       Æ„y)    H   â !r       Æ„y) €  H   â !r*       Æ „y) €  H   â !r       â !r       ä ‰      " !r       â !r

       ä y      &   b !r	       â !r       È/ ‰      " !r       ÆO ‰      b !r0       Æ� ‰	      ¢ !r       â!r

       È ‰       â !r       Æ ‰
      " !r       Ê ‰      b !r       â !r       È/ ‰      " !r		       ÆO ‰      h ‰	      ¢ !r         â� !r
       È‰  €    è ‰      " !r       Ê‰
      b !r       â !r       È/ ‰ €    " !r	       ÆO ‰ €    h ‰	 €    ¢ !r        â� !r       È‰  @    è ‰ €    " !r

       Ê‰
 €    b !r       â !r       È/ ‰ @    " !r		       ÆO ‰ @    è ‰	 @    b !r         â� !r       äy      '   ¨ ‰ @    â !r
       âx
.   ÿàÿ â $z. x  .Ž â ‰       " $vÿ r  ÿ Ž â z 
 r  p`ò â !r       â ‰ @    $ x )   p"ò Ä ‰
      b !r	       Ê/ ‰      b $z. ~  .Ž âO !r       Æ 	   ÿàÿ â!r       â� 
. r  ÿàõ â ‰      ¢ r.   ÿ† â !r        â$tÿ   ÿ Ž â ‰      â $ÿÿ   ÿ â !r       â 
 j  ÿ„ â$r ÿÿ   ÿŽ â z j  ÿ† â !r
       â‰      " 
 k   â %v. j   Ž â	.	   ÿàõ Ä z k   � â !r       â/ †	     â $ÿÿ   ÿ â 	 j  ÿ„ ä †    â !r       âO 		 k   â !r       Æ� †	     è †     â !r       Æ †	     â M™        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y)      "   " ¹z  l      â Ey  P    € â ™x?    â y       !   b ¹z  F   
   â $r(ÿÿ   ÿ Ž ä y*      %   ¢ $r,ÿÿ   ÿ Ž ä $r&ÿÿ   ÿ Ž Ä $r2ÿÿ   ÿ Ž â r )ÿ   pRð är )ÿ   pRò â$x)     Ž ä/ $x**   ÿ Ž ÆO xÿ     È r   ÿ� â $))€    Ž Æ x5ÿ     È | 5   pbô Ú G)       € ê y      '   " $rÿÿ   ÿ Ž â xøÿÿÿÿÀŽ â $rÿÿ   ÿ Ž â y      &   b $z	* p  5Ž ä $x/5   ÿ Ž ä $v3ÿ X  ÿ Ž ä $v-ÿ Y  ÿ Ž â z'/ q  ÿàÿ â $r(ÿÿ   ÿ Ž Ä 'z y   Ž È 'z s   Ž â/ zÿ z    Æ $z+ }  ÿŽ â zÿ t    â $z |  5Ž ä $z
 w  ÿŽ â x*   ÿàÿ â $z |  	Ž ä $z v  Ž â r+
+   ÿàõ â $x   
Ž Ä $z v  Ž ä $z p  Ž â xÿ     â %xR   ÿŽ È %x
R   Ž â r.   ÿà÷ Æ %xR   ÿŽ È $x   ÿ Ž ä $r1ÿÿ   ÿ ä %x   Ž È $x0	   Ž ä %x   Ž È r"
3   ÿàõ Ê $x#   - â r3   ÿàõ È �y"B   ‘ ¢ $x   - ä $t?ÿ$   ÿ Ž â �y="J   ‘ è  �y   • " r$3.   ÿàõ Æ �y   • d $x%-   0 â r3   ÿàõ â �yK"   • è �y$B   ‘ b $x   - Æ �yJ$   • è �y   • è �y   • è �yH"   • è  �yG$   • â  r6+   /àµ Æ �y;$J   ‘ â  r1   ÿd~ Ê $x9$   ÿŽ ä %v66 Z  ? Ž È $x   ÿ Ž ÊO x   ÿÀŽ â $tÿ   ÿ Ž Æ xÿ   øŽ äxÿ     Æ %v 
  Ž È $x@   ÿ Ž ä �y    › ¦ x   øŽ Ê %v 
  Ž Ì �y    › " $x   ÿ Ž Ê/ x   ÿÀŽ È xÿ   øŽ â�$x@   ÿ Ž â xÿ     Æ %v 
  Ž â x   øŽ Ê �y    › â %v  
  Ž Ì �y      › â $x   ÿ Ž Ê x   ÿÀŽ È xÿ   øŽ â$x   ÿ Ž â xÿ     Æ %v 
  Ž â x   øŽ Æ $x   ÿ Ž ä �y    › æ x   ÿÀŽ â %v 
  Ž Æ xÿ   øŽ Æ�y    › b $x   ÿ Ž ä %v 
  Ž â xÿ     È �y    › b x   øŽ Ê %v 
  Ž Ì �y    › b $x7   9Ž ä $rÿÿ   6 Ž Ê �y7   ™ h �y4   ™ b r+   'àµ Æ �y:   ™ h �y8   ™ b r>1   ÿd~ Æ �y6   ™ b %v Z  ? Ž Æ �y<   ™ b $xA>$   ÿŽ Æ �y9   ™ h �y?    ™ b $x   AŽ Ê �y>   ™ h �y@   ™ h �yF   ™ h �yE   ™ h �yD   ™ h �yC   ™ h �yB   ™ h �yA    ™ b r"3   ÿàõ Æ �y    • d# $x#	   - âr$3   ÿàõ â �yI    • h �y""    • b $x%   - Ê �y$$    • b xLKÿ   ÿÀŽ Ê $xMK @  LŽ â xKÿ   K  Æ $xLM€   ÿ Ž ä $xNK @  KŽ Æ xLL€€ÿÀŽ â $xKN€   ÿ Ž Æ x#L€€€€ÿüŽ â $xM   ÿ Ž â/ xKK€€ÿÀŽ ä x##ÿþþþÿàÿ ä x€€ÿÀŽ ä r#L   ÿüŽ ä x#K€€€€ÿüŽ Ä xMJÿ   ÿÀŽ â $x%N   ÿ Ž â x€€€€ÿüŽ ä xN#ÿþþþÿàÿ â $xMJ @  MŽ â x#%€€ÿÀŽ ä xLÿþþþÿàÿ ä r%NK   ÿüŽ â $xKM€   ÿ Ž â xNÿ   J  Ä rL   ÿüŽ ä x#€€€€ÿüŽ â $xNN @  NŽ â xKK€€ÿÀŽ â $xLM   ÿ Ž â xJÿþþþÿàÿ â $xON€   ÿ Ž â xK€€€€ÿüŽ ä xML€€ÿÀŽ Ä xLÿþþþÿàÿ â $xQN   ÿ Ž â rJJ#   ÿüŽ ä xPO€€ÿÀŽ ä x#M€€€€ÿüŽ ä rLK   ÿüŽ ä xLP€€€€ÿüŽ Ä xN#ÿþþþÿàÿ ä xK˜º  ÿ    ä xQQ€€ÿÀŽ ä x˜º  ÿ    ä xLÿþþþÿàÿ ä r#NM   ÿüŽ Ä xLKÿÀŽ ä xMQ€€€€ÿüŽ ä xOÿÀŽ ä xRMÿþþþÿàÿ ä x%%˜º  ÿ    ä xJJ˜º  ÿ    Ä xK€€€€KÞŽ äO xN€€€€ÞŽ Æ $xMK   L
Ž â xKHÿ   ÿÀŽ â $xNN   O
Ž â xL%ÿÀŽ Æ $xOH @  KŽ â xKJÿÀŽ ä x%€€€€%ÞŽ äxJ€€€€JÞŽ ä xM€€€€´Ž â $xM%   L
Ž ä $x%O€   ÿ Ž Ä $xLJ   K
Ž â xKÿ   H  â $xJO   ÿ Ž â xH%€€ÿÀŽ ä xN€€€€´Ž â $xNK @  KŽ â x%M€€€€´Ž ä xJJ€€ÿÀŽ ä xH€€€€ÿüŽ â $xMN€   ÿ Ž â xL€€€€´Ž Ä xLJ€€€€ÿüŽ ä xKÿþþþÿàÿ â $xN   ÿ Ž â xNM€€ÿÀŽ ä xMLÿþþþÿàÿ ä xL€€ÿÀŽ ä xN€€€€ÿüŽ Ä rMMJ   ÿüŽ ä xOÿþþþÿàÿ ä xM˜º  ÿ    ä rKKH   ÿüŽ ä xMÿÀŽ ä xKK˜º  ÿ    Ä xJL€€€€ÿüŽ ä x€€€€ÞŽ ä� rRQ   ÿüŽ ä rOON   ÿüŽ â $x   M
Ž â xHKÿÀŽ ä xQJÿþþþÿàÿ Ä xK€€€€KÞŽ ä xOO˜º  ÿ    ä rQQL   ÿüŽ â $xKK   H
Ž â x€€€€´Ž ä xJGÿ   ÿÀŽ ä xOÿÀŽ Ä xHQ˜º  ÿ    ä xO €€€€OÞŽ â $xLG @  JŽ â xK€€€€´Ž ä xKHÿÀŽ ä xJÿ   G  â $xGO   
Ž â xH!€€€€HÞŽ â $xL€   ÿ Ž Æ xGG€€€€ ´Ž â $xHH   K
Ž ä $x L   ÿ Ž â x€€ÿÀŽ â $xMJ @  JŽ â x!H€€€€!´Ž ä xH €€ÿÀŽ â $xJM€   ÿ Ž âx €€€€ÿüŽ â $xMM   ÿ Ž Æ xK ÿþþþÿàÿ ä xLJ€€ÿÀŽ ä xJH€€€€ÿüŽ ä rKK   ÿüŽ ä xNM€€ÿÀŽ ä xMJÿþþþÿàÿ Ä x L€€€€ÿüŽ ä xKK˜º  ÿ    ä xN€€€€ÿüŽ ä rMMH   ÿüŽ ä xO ÿþþþÿàÿ ä x KÿÀŽ Ä xQÿþþþÿàÿ ä xK€€€€KÞŽ ä xM˜º  ÿ    ä rOOL   ÿüŽ â $xKK    
Ž â rQQN   ÿüŽ ä xMÿÀŽ Ä x€€€€ÞŽ ä xOO˜º  ÿ    ä x Q˜º  ÿ    â $x   M
Ž â xKK€€€€´Ž ä xOÿÀŽ ä xM ÿÀŽ Ä xO€€€€OÞŽ äx €€€€ ÞŽ ä x˜º  ÿ    â $xOO   
Ž ä $x    M
Ž â x ÿÀŽ ä x€€€€ÞŽ Ä rP   ÿüŽ ä x€€€€´Ž â $x    
Ž â x#˜º  ÿ    ä x€€€€´Ž ä x˜º  ÿ    ä x#ÿÀŽ Ä x€€€€ÞŽ ä x€€€€´Ž ä xÿÀŽ ä x€€€€ÞŽ â $x   #
Ž â x˜º  ÿ    Æ $x   
Ž â x€€€€´Ž ä xÿÀŽ ä x€€€€ÞŽ â &r7   ÿ   â x€€€€´Ž â &rK4   ÿ   â xO€€€€´Ž â $x   
Ž Ä &r:      ä &r8      ä &rG6      â x;ð   ÿÀŽ â&r<      â x;;   ÿÀŽ â &r!9      ä &r?      â x€€€€´Ž â $r;   ÿŽ â xÿ     â $x   Ž Ä &r77   ÿ   ä &r44   ÿ   ä $r   Ž â r   ÿ� â &r>   ÿ   ä&r>   ÿ   ä &rK@   ÿ   Ä &r7:   7   ä &r@@   ÿ   ä &r8   4   ä &rF      ä&rF      ä &r6%6   7   Ä &rE      ä&r@E   @   ä &r<<      â x=ð   ÿÀŽ â &rGD      âx==   ÿÀŽ â &r%D      ä &r69   6   Ä &rC      ä&r@C   @   ä &r<?   <   â xÿ     â &r!B      ä&rB      ä $r =6   ÿŽ Ä &rA      ä&r@A   @   ä $x66   <Ž ä $r;;   ÿŽ ä $r==   ÿŽ ä $x   Ž Ä $x   @Ž â r66   ÿ� â $r <    Ž â r   ÿú� â $r;   ;Ž â r   ÿ� â $r=@   =Ž â r   ÿ� ä r    ÿú� ä xÿ     Ä r;   ÿú� ä r=   ÿú� ä xÿ     ä r   ÿ� ä x55   ÿàÿ ä xÿ     Ä xÿ     â 0r"ÿ"    A   â r   ÿ� â 0rÿ    A   â xÿ     ä | 5   pbô ä r   ÿ� ä r   ÿ� â 0rIÿI    A   â Er        â  r"     A  â 0r$ÿ$    A   â xÿ     Ä xÿ     ä xÿ     â #r2   2   â  r"I     A  â Er        â  r$     A  â Er        â  rI$I     A  â Er        ä x33   ÿà÷ â #r,   ,   â #r&   &   â #r(I   (   â x''€   ÿàÿ â $r-ÿÿ   -Ž â x//€   ÿàÿ â D)       À â Gy  éÿÿÿÿƒ ê Ay        € ê ˆ )2 þÿ H   è ˆ )&€þÿ H   è ˆ ), ÿÿ H   è ˆ )(€ÿÿ H   è {          ì M        € ê „y      H   " z * r  p`ð Æ „y  €   H   b r  ÿ   p"p Æ „y     H   ¨ „y     H   è „y  €  H   ( „y     H   h „y     H   h „y	  €  H   h „y     H   h „y  €  H   h „y
  €  H   b !r2       Æ „y  €  H   " !r&       â/ !r,       âO !r       ä� !r       â!r       â!r       ä !r		       Æ ‰      b !r       Æ ‰	      ¢ !r(       Æ ‰      â !r

       Æ y      &   " !r
       â !r       È/ ‰      " !r	       ÆO ‰      b !r       Æ� ‰      ¨ ‰      â !r       â !r       È/ ‰      " !r       ÆO ‰ €    b !r       Æ� ‰ €    ¨ ‰
 €    â !r       â !r       È/ ‰ €    " x*   ÿàÿ â $z* x  *Ž â!r       äO ‰ @    b z  r  p`ò â !r

       ä� ‰ @    ¢ x     p"ò Æ ‰	
 @    è y      '   " !r       â !r       â/ !r       ÈO ‰      " !r	
	       Æ� ‰      b $z* ~  *Ž Æ‰
 @    ¢ $tÿ   ÿ Ž â 
 * r  ÿàõ ä ‰	      æ $ÿÿ   ÿ â 
  j  ÿ„ È 
  k   â !r       â %v* j   Ž â !r       È/ †	     â !r

       ÆO †    â !r		       Æ� ‰
      h  †	 	    â M™        € ê ¹z  r      â !r
       â/ �x   ?àÿ Ì |**   ÿàñ Ê $rÿÿ   ÿ  â z* j  ÿ€ È z* k    Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y       &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž ä $rÿÿ     Ž È % \   Ž Ê �	    ™ b ¹z  l      â Ey  P    € â ™x?    â y      "   b $rÿÿ   ÿ Ž Æ y      !   ¢ $ÿÿ   ÿ Ž ä $ÿÿ    Ž È '
 m   Ž È $‚ÿÿ   ÿ Ž â 
ÿ n    Ê $ÿÿ   
Ž â r ÿ   pRò Æ/ $
 o  Ž â r ÿ   pRô â$x    Ž äO $‚ÿÿ    Ž Æ xÿ     â 'Š s   Ž Æ r   ÿ� È xÿ     È |    pbö ä Šÿ t    Ö G9  à    € ê y      '   " $rÿÿ   ÿ Ž â xøÿÿÿÿÀŽ â $z w  ÿŽ â y	      %   b $tÿR   ÿ Ž ä $x   
Ž ä $rÿÿ   ÿ Ž ä $x   ÿ Ž â xÿ     â 'z y   Ž È $z }  ÿŽ â zÿ z    â $tÿ   ÿ Ž Æ r   ÿàñ â $z |  Ž ä $x   ÿ Ž ä $z v  Ž ä%v  € Ž È $z	 p  Ž È/ %v X  Ž È $rÿÿ    Ž ä $rÿÿ   ÿ  Æ r   ÿàñ Ê $x     â r
   ÿàñ È �yB   ‘ ¢  $x     ä $tÿ   ÿ Ž â �y%J   ‘ è  �y
àÿÿ • ( �y
âÿÿ • h �y
    • è
 �y
   • â
 $rÿÿ    Ž Ä $rÿÿ    Ž Ê �y&    • â  $x   ÿ Ž ÊO x   ÿÀŽ È xÿ   øŽ äxÿ     Æ %v 
  Ž È $x@   ÿ Ž ä �y    › ¦ x   øŽ ä r   à± Æ %v 
  Ž Ì �y    › " r   ÿd~  â $tÿ$   ÿ Ž ä $x   ÿ Ž ä %v Z   Ž È $x$   ÿŽ â x   ÿÀŽ Æ $x   Ž â xÿ   øŽ â$x   ÿ Ž Æ �y   ™ b xÿ     â %v
 
  Ž ä �y"   ™ b x   øŽ Æ �y

    › d %v 
  Ž ä �y#   ™ h �y    › h �y!   ™ h �y    ™ h �y   ™ h �y   ™ h �y$    ™ h �y'    • b x(ÿ   ÿÀŽ Ä� x)ÿ     Æ $x* @  (Ž È $x*   ÿ Ž ä$x(*€   ÿ Ž ä $x*) @  )Ž â x€€ÿÀŽ ä x((€€ÿÀŽ â $x*€   ÿ Ž âx€€€€ÿüŽ â/ $x*   ÿ Ž â x)(€€€€ÿüŽ Ä x€€ÿÀŽ ä xÿþþþÿàÿ ä x))ÿþþþÿàÿ ä x€€ÿÀŽ ä x€€€€ÿüŽ ä r   ÿüŽ Ä r))(   ÿüŽ ä x€€€€ÿüŽ ä xÿþþþÿàÿ ä x(˜º  ÿ    ä x,ÿþþþÿàÿ ä x(ÿÀŽ Ä r   ÿüŽ ä r,,   ÿüŽ ä x˜º  ÿ    ä x,,˜º  ÿ    ä x))˜º  ÿ    È x*)ÿÀŽ ä x   ÿàÿ È |    pbð â 0r&ÿ&    A   â x   ÿà÷ ä x€   ÿàÿ Æ $rÿÿ   Ž â x€€€€(ÞŽ äO x(ÿ   ÿÀŽ Æ $x   
Ž â xÿÀŽ â $x( @  (Ž Æ x€€€€´Ž ä x,ÿÀŽ ä x€€€€ÞŽ äx,	€€€€,ÞŽ ä xÿ     â $x   
Ž â x)€€€€)ÞŽ â $x(€   ÿ Ž Ä $x,,   
Ž ä $x @  Ž â x€€ÿÀŽ â $x(   ÿ Ž â x€€€€´Ž â $x))   *
Ž ä $x€   ÿ Ž â x€€ÿÀŽ ä x€€€€ÿüŽ Ä x)€€€€´Ž â $x   ÿ Ž â x€€ÿÀŽ ä x€€€€ÿüŽ ä xÿþþþÿàÿ â &r   ÿ   âx€€€€ÿüŽ ä x(ÿþþþÿàÿ Ä r   ÿüŽ ä x€€ÿÀŽ ä xÿþþþÿàÿ ä r((   ÿüŽ ä x˜º  ÿ    ä x€€€€ÿüŽ Ä r   ÿüŽ ä x((˜º  ÿ    ä xÿÀŽ ä xÿþþþÿàÿ ä x
€€€€ÞŽ ä x(ÿÀŽ Ä r   ÿüŽ ä x˜º  ÿ    â $x   
Ž â x(€€€€(ÞŽ ä x˜º  ÿ    ä xÿÀŽ â $x((   
Ž â x€€€€ÞŽ Ä x€€€€
´Ž ä xÿÀŽ â $x   
Ž â x€€€€ÞŽ â &r"   ÿ   â x((€€€€´Ž â &r#      â x	,€€€€	´Ž â $x   
Ž â x€€€€´Ž â &r((!      Ä &r       â x€€€€´Ž â &r   (   ä &r	      â x%ð   ÿÀŽ äx%%   ÿÀŽ â &r$      â xÿ     Æ $r%%   ÿŽ ä $x   Ž ä $r   %Ž Æ r   ÿ� È r   ÿú� È xÿ     È r   ÿ� â 0r'ÿ'    A   Æ xÿ     ä  r'&'     A  ä Er        Ê #r'      â G‰  �õÿÿÿÿƒ ê Ay        € ê ˆ €ÿÿ H   è {          ì M)        € ê „y     H   " r ÿ   pRð Æ „y €   H   h „y    H   ¢ !r       È !r       È/ !r       ÊO ‰      $ !r       Ê ‰      $ !r       Ê ‰ €    $ !r       Ê ‰ @    $!r       Ê ‰      b  M	        € ê y      %   ¨ y      '   â $z   x  Ž äO $tÿ   ÿ Ž ä $z  ~   Ž â� !r       Æ/ %v  j   Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž È % \   Ž Ê �	    ™ ¢ $rÿÿ   ÿ Ž â Ey  à    € â $rÿÿ   ÿ Ž â y      "   " 'Š  s   Ž Æ y
      !   " '
 m   Ž Æ y	      %   d 
ÿ n    ä rÿ
   ÿü‚ â $x    
Ž Ê xÿ     È r   ÿ� ä Šÿ t     â $r ÿÿ   ÿ Ž Æ r       âO G  à    € ê z ÿ `  pPô â/ $rÿÿ   ÿ Ž â z ÿ ^  pPò ä z ÿ b  p ö ä z ÿ a   Sô ä z ÿ _  Sò ä z ÿ c  0't Ö        '   " $ x  ÿŽ ä$ª x  ÿŽ â ©      '   b ¨ÿ   	  â$   ~  ÿŽ ä $ª ~  ÿŽ â/ ÿ   	  ä  
   	àë Ä ¢
   	àÇ ä ÿÿ   Äþ ä      ÿàý ä ¢   ÿàû ä ¢
ÿÿ   „þ â $ÿÿ    â   ^  ÿ† Ä ª b  ÿˆ â $¢
ÿÿ   
Ž â   _  � â $r ÿÿ   ÿ Ž Æ ª c  
 â �     h �©      d Ay        € ê/ ¹z  l      â xÿ     â ™x?    â $ÿÿ   
Ž â Ey       € â $rÿÿ   ÿ Ž Ä $
 o  Ž â|    pbò â $‚ÿÿ    Ž ä $rÿÿ   ÿ Ž Ô G  0    € ê y      '   " $rÿÿ   ÿ Ž â xøÿÿÿÿÀŽ â $rÿÿ   ÿ Ž ä $rÿÿ   ÿ Ž ä $x   
Ž ä $x   ÿ Ž Æ xÿ     â 'z y   Ž È $z }  ÿŽ â zÿ z    Ê $z |  Ž ä $z w  ÿŽ ä $z v  Ž Æ r   ÿàñ â $z	 p  Ž È %xR   ÿŽ È $rÿÿ    Ž â z X  ÿàó â $rÿÿ   ÿ  Æ z Y  ÿäÿ  ä $x&   ÿ Ž â r   ÿàñ â $t'ÿ   ÿ Ž È %v& X  'Ž È $x     â r   ÿàñ È �y	B   ‘ ¢  $x     Ê �y   • è �y   • ( �y*"   • ( �y)$   • " $t%ÿ   ÿ Ž â r   à‘ â $tÿ$   ÿ Ž â �yJ   ‘ $ r   ÿ$~  Ê $x$   ÿŽ â z X  ÿàñ â $x	   ÿ Ž äO$x	@   ÿ Ž Æ x   ÿÀŽ È xÿ   øŽ ä�xÿ     Æ %v 
 % Ž â x   øŽ Æ $x	   ÿ Ž ä �y    › ¢ %v 
 % Ž ä/ x   ÿÀŽ È xÿ   øŽ â�y    › â $x		   ÿ Ž â xÿ     ä %v 
 % Ž ä x
   	øŽ È �y    › " %v Z   Ž È %v

 
 % Ž È $x		   Ž ä �y

    › ( �y   ™ ( �y#   ™ ( �y   ™ ( �y"   ™ ( �y!   ™ ( �y    ™ ( �y   ™ ( �y$    ™ " z Y  ÿä  Æ �y(    • ( �y    • " x+*ÿ   ÿÀŽ Ê $x,* @  +Ž â x*ÿ   *  Æ $x+,€   ÿ Ž ä $x.* @  *Ž ä $x,,   ÿ Ž â x++€€ÿÀŽ â $x-.€   ÿ Ž Æ x+€€€€ÿüŽ ä x,,€€ÿÀŽ ä x--€€ÿÀŽ ä x*ÿþþþÿàÿ â $x.   ÿ Ž â x,€€€€ÿüŽ ä/ x	-€€€€ÿüŽ Ä x.)ÿ   ÿÀŽ är**+   ÿüŽ ä x+ÿþþþÿàÿ â $x.) @  .Ž â x	ÿþþþÿàÿ ä x	€€ÿÀŽ â $x.€   ÿ Ž â r-   ÿüŽ Ä x	€€€€ÿüŽ ä x-ÿ   )  â $x).   ÿ Ž â r++,   ÿüŽ ä x,ÿþþþÿàÿ â $x/- @  -Ž â x€€ÿÀŽ ä x)€€ÿÀŽ Ä r	,	   ÿüŽ â $x,/€   ÿ Ž â x)€€€€ÿüŽ â $x//   ÿ Ž â x-€€€€ÿüŽ ä x.)ÿþþþÿàÿ ä x,,€€ÿÀŽ ä r.   ÿüŽ Ä x0-ÿþþþÿàÿ ä x)/€€ÿÀŽ ä x.,€€€€ÿüŽ ä x-*˜º  ÿ    ä x/)€€€€ÿüŽ ä x1.ÿþþþÿàÿ Ä x*+˜º  ÿ    ä x.-ÿÀŽ ä r0   ÿüŽ ä x2/ÿþþþÿàÿ ä x0*ÿÀŽ ä x˜º  ÿ    Ä r)2)   ÿüŽ ä z ÿ `  pPð È z ÿ a   Sð ä x   ÿàÿ ä x-€€€€-ÞŽ äO x/€€€€*ÞŽ Æ $x+-   .
Ž â x-˜º  ÿ    â $x0/   0
Ž â r*1,   ÿüŽ ä x	˜º  ÿ    ä x,-ÿÀŽ ä x-€€€€-ÞŽ ä� x++€€€€´Ž Ä x0€€€€´Ž ä xÿÀŽ â $x--   ,
Ž â x€€€€ÞŽ ä x,˜º  ÿ    Æ $x   
Ž â x,ÿÀŽ ä x,€€€€,ÞŽ äx.ÿÀŽ ä x€€€€ÞŽ â $x,,   
Ž â x*˜º  ÿ    Æ $x   .
Ž â x,€€€€´Ž ä x,ÿÀŽ ä x*)˜º  ÿ    ä x
€€€€ÞŽ ä x€€€€´Ž â &r++   ÿ   â x€€€€´Ž â $x   ,
Ž â x*ÿÀŽ Ä x*€€€€*ÞŽ â &r#   ÿ   â x-€€€€´Ž â &r   +   â x€€€€
´Ž â $x**   
Ž ä &r"      ä &r!      â x*€€€€´Ž â &r       Ä &r      â xð   ÿÀŽ äx   ÿÀŽ â &r$      â xÿ     Æ $r   ÿŽ ä $x   Ž ä $r   Ž Æ r   ÿ� È r   ÿú� È xÿ     È r   ÿ� â 0rÿ    A   ä 0r)ÿ(    A   â xÿ     â G‰   	    € ê z `  à‘ â %v& `  'Ž Æ z	 a  $~  ä r   ÿàñ Æ �y
B    ¤  $x     Ê �y    è �y    ( �y&"    ( �y'$    ( �yJ    " z `  ÿàñ È/ z a  ÿä  Ê �y     " $x
   ÿ Ž ÊO x   ÿÀŽ È xÿ   øŽ Ê� %v 
 % Ž â xÿ     Æ $x
@   ÿ Ž ä �y    › ¦ x   øŽ Ê %v 
 % Ž Ì �y    › â $x
   ÿ Ž Ê x   ÿÀŽ â $x

   ÿ Ž Æ xÿ   øŽ äxÿ     Æ %v 
 % Ž â x
   
øŽ Ê �y    › " %v

 
 % Ž Ì �y

    › " x%&ÿ   ÿÀŽ Ê $x(& @  %Ž È $x%(€   ÿ Ž ä$x((   ÿ Ž â x&ÿ   &  ä x%%€€ÿÀŽ ä x((€€ÿÀŽ ä x	%€€€€ÿüŽ â $x+& @  &Ž â x*(€€€€ÿüŽ Ä x&	ÿþþþÿàÿ â $x	+€   ÿ Ž âx*ÿþþþÿàÿ ä/ r&&%   ÿüŽ â $x++   ÿ Ž â r((   ÿüŽ ä x*	€€ÿÀŽ ä x&˜º  ÿ    Ä x	+€€ÿÀŽ ä x%*€€€€ÿüŽ ä x&ÿÀŽ ä x((˜º  ÿ    ä x,	€€€€ÿüŽ ä x+%ÿþþþÿàÿ Ä x%(ÿÀŽ ä x,,ÿþþþÿàÿ ä r++*   ÿüŽ ä r,,	   ÿüŽ ä x++˜º  ÿ    ä x	ÿ   '  Ä x,,˜º  ÿ    â 0rÿ    A   â x€€€€ÞŽ äO x(€€€€(ÞŽ Æ $x   &
Ž â x&'ÿ   ÿÀŽ â $x((   %
Ž Æ x€€€€´Ž â $x&' @  &Ž â x(€€€€´Ž ä x€€€€+ÞŽ â� &r   ÿ   â x+ÿÀŽ â $x+	 @  	Ž ä $x	&€   ÿ Ž ä &r      Ä $x&   ÿ Ž â x	€€ÿÀŽ â $x   
Ž â x	,ÿÀŽ â $x&+€   ÿ Ž â x€€€€,ÞŽ ä x%€€ÿÀŽ ä x€€€€ÿüŽ â $x   	
Ž â x'&€€ÿÀŽ â $x++   ÿ Ž â x	%€€€€ÿüŽ Ä x&ÿþþþÿàÿ ä x'€€€€ÿüŽ ä x(	ÿþþþÿàÿ ä r&&   ÿüŽ ä x++€€ÿÀŽ ä x*ÿþþþÿàÿ Ä r((%   ÿüŽ ä x	&˜º  ÿ    ä x+€€€€ÿüŽ ä r**'   ÿüŽ ä x((˜º  ÿ    ä x&	ÿÀŽ Ä x,ÿþþþÿàÿ ä x	€€€€	ÞŽ äx(ÿÀŽ ä r,,+   ÿüŽ ä x*˜º  ÿ    â $x		   &
Ž â x(€€€€(ÞŽ Ä x,,˜º  ÿ    ä x&ÿÀŽ â $x((   
Ž â x
€€€€ÞŽ ä x	€€€€´Ž ä x	,ÿÀŽ â $x   &
Ž â x,€€€€,ÞŽ â &r#   ÿ   â x(€€€€´Ž Ä x€€€€´Ž â $x,,   	
Ž â x€€€€
´Ž â &r"      â x€€€€´Ž â &r!      â x,€€€€´Ž â &r       ä &r      â x   ÿÀŽ â &r$      â xÿ     Æ $r   ÿŽ ä $x   Ž ä $r   Ž Æ r   ÿ� È r   ÿú� È xÿ     È r   ÿ� È xÿ     È Er        â  r)     A  È #r      â |    pbð â  r)     A  â Er        ä x   ÿàó ä x€   ÿàÿ â #r      ä $rÿÿ   Ž  È D	       À â Gy   ìÿÿÿÿƒ ê Ay        € ê y      "   " Ey        € â r ÿ   pRð är ÿ   pRò Ö G‰  `     € ê y      !   " z ÿ `  pPð È z ÿ a   Sð â $x    Ž Ê ˆs €ÿÿ H   î ˆ    H   ä Ay        € ê {          ì M        € ê y      !   b z ÿ `  pPð È z ÿ a   Sð â „y     H   X. „‰ €   H   ( „‰    H   ¢ !r       È/ !‚       È !‚       âO G‰        € ê „y €  H   ( „y €   H   h „y    H   ¨ „y    H   è „y €  H   " !r       â !r       Æ/ !r       âO !r       Æ� !r       ä‰      " z ÿ `  pPð ä r ÿ   pRò ä z ÿ a   Sð â !r       Ê ‰      $ !r       Ê ‰ €    $ !r       Ê ‰ @    $ !r       Ê ‰      $ !r	       â G‰        € ê ‰      $ !r       Ê ‰      $ !r       Ê ‰ €    $ !r       Ê ‰ @    $ !r       Ê ‰      $ !r       ä M        € ê !r	       âG‰  �    € ê $tÿ   ÿ Ž â !r        È z  h  pð Ú G	  �    € ê ¹z  h      ä �xÿÿÿÿ?àÿ È Œx    pð È ‡x       È ™x   ?  Ø »z  €    ä ™x?    Ô Xy  âÿÿÿÿƒ ê  x ;ª¸¿  A  Ì s          $ !t   €?    Ì s        $  r     A  È  r     A  â Gy  `    € ê  x '7=  A  â x*BL?  A  Æ #t   €?    È  r      A  â  x    ?  A  Ê s     $   $ !t  €?    È  r       A  È  r      A  â Gy  À    € ê $v ÿ h  ÿ Ž Ê x     pRð Ú G‰  à     € ê x     pRð Ú G	        € ê 	z  i    � ä 	z i    � Æ  x ;ª¸¿  A  â 	z i €  � Ê s        $ !t  €?    Ì s        $  r       A  È  r      A  â Gy  Ð     € ê  r     A  â Gy  °     € ê 	x   à@  � ä 	x  à@  � Æ  x #ÛÙ?  A  â 	x  àÀ  � Æ  x;ª¸¿  A  ä !t  €?    È s        $ !t  €?    Ì s        $  r       A  È  r      A  ä y       %   ( y      &   ( y      '   b $z  x   Ž ä $tÿ   ÿ Ž Ä $z  ~   Ž È/ %v  j   Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y       &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž ä $rÿÿ     Ž È % \   Ž Ê �	    ™ b ¹z  l      â Ey  0(    € â ™x?    â y      "   b $rÿÿ   ÿ Ž Æ y      !   ¢ $ÿÿ   ÿ Ž ä $ÿÿ    Ž â y	      %   & '
 m   Ž È $‚ÿÿ   ÿ Ž â 
ÿ n    Ê $
ÿÿ   
Ž â r ÿ   pRò Æ/ $

 o  Ž â r ÿ   pRô â$x    Ž äO $‚ÿÿ    Ž Æ xÿ     â 'Š s   Ž Æ r   ÿ� â $rÿÿ   ÿ Ž ä $€   Ž â xÿ     â $x		   ÿ Ž Æ |    pbö ä Šÿ t    ä x      ÿ  Ò G9  p&    € ê y      '   " $r
ÿÿ   ÿ Ž ä $vÿ Y  ÿ Ž ä 'z y  
 Ž â x
øÿÿÿÿÀŽ Æ $z }  ÿŽ â zÿ z    â $z	 p  Ž ä $x
   

Ž ä $z |  Ž ä $z w  ÿŽ â xÿ   
  â $z v  Ž Ä$x!
   ÿ Ž â r   ÿà÷ â %xR   ÿŽ È $vÿ X  ÿ Ž â r
   ÿàñ â $rÿÿ   ÿ Ž ä $x   ÿ Ž ä $rÿÿ   ÿŽ ä %x !   Ž È $x     ä r8   ÿàñ Ê $x9     â r    ÿàñ È �y8B   ‘ ¢ $x!     Ê �y   • è �y   • " $vYÿ p  ÿ Ž Æ �yP"   • b %x>YR   8Ž Æ �yO$   • ( �y>B   ‘ " %xYR   Ž È $tHÿ   ÿ Ž â �y4   • ( �y   • " %x"YR   >Ž Æ �yG"   • ( �yR"B   ‘ " %xBYR   Ž âr$
   à± ä �yF$   • h �yB   • b %x*YR   "Ž Æ �yQB   • h �yM*B   ‘ b %x&YR   BŽ Æ �yB"   • h# �yS&   • h �yL&   • b $t%ÿ$   ÿ Ž Æ �yE>J   ‘ b %v$$ Z  % Ž Æ �yK"J   ‘ h �yN*J   ‘ b $x   ÿ Ž ÊO x   ÿÀŽ È xÿ   øŽ Ê� %v0 
 H Ž â xÿ     Æ $x@   ÿ Ž ä �y00    › ¦ x   øŽ Ê %v, 
 H Ž È $x   ÿ Ž ä �y,,    › æ x   ÿÀŽ È x)ÿ   øŽ Ê%v() 
 H Ž Ì �y((    › " $x   ÿ Ž â xÿ     È x   øŽ ä r   ÿd~  Æ %v. 
 H Ž È $x$   ÿŽ ä �y..    › $ $x%%   Ž ä $x   ÿ Ž Æ �yD$   ™ ( �y@$   ™ " x   ÿÀŽ Æ �y$   ™ " xÿ   4øŽ Æ �yA$   ™ $ %v2 
 H Ž ä �y$   ™ ( �y$   ™ ( �y$   ™ ( �y22    › " $x5@   ÿ Ž Æ �y$    ™ " x4ÿ   4  Æ �y8J   ‘ " x54   5øŽ â $x6   ÿ Ž ä $xR   ÿ Ž â rV   ÿàñ â %v45 
 H Ž â x[Pÿ   ÿÀŽ â�yI$    • h �y44    › b x76   ÿÀŽ â $x   ÿ Ž Æ x77ÿ   øŽ äxÿ     Æ %v67 
 H Ž âx   øŽ ä x   ÿÀŽ â �yB$   • d %v8 
 H Ž ä�y66    › h �y88    › b x;ÿ   øŽ â$xR@   ÿ Ž â xÿ     â �y&"   • d %v:; 
 H Ž â x   øŽ Ê �y::    › b %v< 
 H Ž Ì �y<<    › b $xW     ä $x[P @  [Ž ä %xTYR   VŽ â �yV    • f $x#M   ÿ Ž â �yT    • b %x>YR   TŽ È $xV[€   ÿ Ž âxWÿ   P  â $x%[   ÿ Ž â/ �yJ>    • d xVV€€ÿÀŽ â $xWW @  WŽ â �yP&$   • b %xBYR   >Ž â x"V€€€€ÿüŽ Æ $x*W   ÿ Ž â x>#   ÿÀŽ â $x$W€   ÿ Ž â x%%€€ÿÀŽ â $xTM@   ÿ Ž â x#"ÿþþþÿàÿ â �yBB    • b x?>ÿ   SøŽ ä x>*€€ÿÀŽ ä x&$€€ÿÀŽ Ä/ x'%€€€€ÿüŽ ä r$#V   ÿüŽ â %v"? 
 H Ž â x?>€€€€ÿüŽ ä x*'ÿþþþÿàÿ ä x+&€€€€ÿüŽ â �y""    › b x??ÿþþþÿàÿ Ä r%*%   ÿüŽ ä xSÿ   S  ä x'$˜º  ÿ    ä x++ÿþþþÿàÿ ä r*?>   ÿüŽ â $x?R   ÿ Ž â x>%˜º  ÿ    Ä xSS   TøŽ ä xT'ÿÀŽ ä r&+&   ÿüŽ ä x+>ÿÀŽ â %v$S 
 H Ž â xX?   ÿÀŽ ä xVOÿ   ÿÀŽ Ä xSXÿ   QøŽ â �y$$    › b xC&˜º  ÿ    ä x**˜º  ÿ    ä x'0€€€€'ÞŽ äO x>1€€€€>ÞŽ Æ $x''   T
Ž ä $x>>   +
Ž ä $x+O @  VŽ â x?'€€€€0´Ž â %v&S 
 H Ž â x1>€€€€1´Ž Æ $xSR   ÿ Ž â xRÿ   Q  â $x>+€   ÿ Ž â x0CÿÀŽ â �y&&    › ¢ xC,€€€€CÞŽ ä� xUR   SøŽ ä xR>€€ÿÀŽ â $xQC   0
Ž â xC*ÿÀŽ â $xT+   ÿ Ž â xSR€€€€ÿüŽ Ä x0-€€€€*ÞŽ â %v*U 
 H Ž â xSSÿþþþÿàÿ ä xOÿ   O  â $x>0   C
Ž â rSSR   ÿüŽ â �y**    › â xCT€€ÿÀŽ â $xOO @  OŽ â x0Q€€€€,´Ž Ä xSS˜º  ÿ    ä x,C€€€€ÿüŽ ä x>>€€€€-´Ž â $x-O€   ÿ Ž â xRSÿÀŽ ä x,,ÿþþþÿàÿ ä xS(€€€€SÞŽ â$xQO   ÿ Ž â r,,C   ÿüŽ Ä xO-€€ÿÀŽ â $xCS   R
Ž ä $xRM   ÿ Ž â xTQ€€ÿÀŽ ä x-O€€€€ÿüŽ ä x,,˜º  ÿ    ä xUR   ÿÀŽ Ä xQT€€€€ÿüŽ ä xR-ÿþþþÿàÿ ä x-,ÿÀŽ ä x,)€€€€,ÞŽ ä xUUÿ   LøŽ ä xSQÿþþþÿàÿ Ä rQRO   ÿüŽ â $xR,   -
Ž â rOST   ÿüŽ â %v,U 
 H Ž â xLÿ   L  Æ $xSM   ÿ Ž â xCC€€€€(´Ž â �y,,    › " xMR€€€€)´Ž ä x(Gÿ   ÿÀŽ ä x)L   SøŽ Æ $xSG @  (Ž ä %v() 
 H Ž Ì �y((    › ¢ xQQ˜º  ÿ    È xRQÿÀŽ ä xQ.€€€€QÞŽ â $xHS€   ÿ Ž â xLO˜º  ÿ    Æ $xQQ   R
Ž ä $xRS   ÿ Ž â xOLÿÀŽ â &r??D   ÿ   â xL/€€€€LÞŽ ä xHH€€ÿÀŽ ä xRR€€ÿÀŽ â &r11@   ?   â x?H€€€€ÿüŽ â $xLL   O
Ž â xOR€€€€ÿüŽ Ä x??ÿþþþÿàÿ â &rCC   ÿ   â xOOÿþþþÿàÿ ä xGÿ   G  â &r00A   1   â r??H   ÿüŽ â &rCM   C   â x.Q€€€€.´Ž ä rOOR   ÿüŽ â $xGG @  GŽ â x??˜º  ÿ    â &r>>   0   â x0O˜º  ÿ    â &r..   C   â xH?ÿÀŽ â $xCG€   ÿ Ž â x10ÿÀŽ Ä x?2€€€€?ÞŽ ä x03€€€€0ÞŽ â $xGG   ÿ Ž â xCC€€ÿÀŽ â $x??   H
Ž â x/L€€€€/´Ž â $x00   1
Ž â x1C€€€€ÿüŽ ä xGG€€ÿÀŽ â &r//   .   â x??€€€€2´Ž Ä x.ð   ÿÀŽ ä x21ÿþþþÿàÿ ä x   ÿÀŽ ä x1G€€€€ÿüŽ ä x.ÿ   .  â $r>   ÿŽ â xH1ÿþþþÿàÿ Ä r22C   ÿüŽ â $r./   Ž â rHHG   ÿüŽ â $x>>   /Ž â x30€€€€3´Ž â &r.?D   ÿ   â x1Fÿ   ÿÀŽ ä x/2˜º  ÿ    ä xHH˜º  ÿ    â &r.3@   .   â x0/ÿÀŽ â $x3F @  1Ž â x/4€€€€/ÞŽ Äx1HÿÀŽ ä xFÿ   F  ä xH5€€€€HÞŽ â $x23€   ÿ Ž ä $x//   0
Ž ä $xCF @  FŽ ä $x0H   1
Ž Ä $x13   ÿ Ž â x22€€ÿÀŽ â $x3C€   ÿ Ž â x4/€€€€4´Ž ä x11€€ÿÀŽ ä x/2€€€€ÿüŽ â $xCC   ÿ Ž â x?3€€ÿÀŽ ä x31€€€€ÿüŽ Ä x//ÿþþþÿàÿ ä x00€€€€5´Ž ä x5?€€€€ÿüŽ ä xF3ÿþþþÿàÿ ä r//2   ÿüŽ ä xCC€€ÿÀŽ Ä x25ÿþþþÿàÿ ä rFF1   ÿüŽ ä x//˜º  ÿ    ä x1C€€€€ÿüŽ ä r?2?   ÿüŽ ä xFF˜º  ÿ    Ä x2/ÿÀŽ ä xH1ÿþþþÿàÿ ä x/6€€€€/ÞŽ ä x1FÿÀŽ ä rHHC   ÿüŽ ä x??˜º  ÿ    â $x//   2
Ž â xF7€€€€FÞŽ Ä xHH˜º  ÿ    ä x2?ÿÀŽ â $xFF   1
Ž â x?8€€€€?ÞŽ ä x//€€€€6´Ž â &r.4A   .   â x1HÿÀŽ â $x??   2
Ž â xH9€€€€HÞŽ Ä x4ÿ   ÿÀŽ â &r//   ÿ   â x7F€€€€7´Ž â $xHH   1
Ž â x??€€€€8´Ž â $x4 @  4Ž ä &r27   /   â x9H€€€€9´Ž â $x/4€   ÿ Ž â x1ÿ     â &r2?   2   Ä $x4   ÿ Ž â x//€€ÿÀŽ â &r00   .   â x.Eð   ÿÀŽ â &r99   2   â x2€€ÿÀŽ â $x51 @  1Ž â xEE   ÿÀŽ ä x1/€€€€ÿüŽ ä x32€€€€ÿüŽ â $rEE0   ÿŽ Ä $x0   9Ž â x01ÿþþþÿàÿ â $x45€   ÿ Ž â x33ÿþþþÿàÿ â $x15   ÿ Ž â r00/   ÿüŽ ä x44€€ÿÀŽ ä r332   ÿüŽ ä x51€€ÿÀŽ Ä x14€€€€ÿüŽ ä x/0˜º  ÿ    ä x03˜º  ÿ    ä x31ÿþþþÿàÿ ä x65€€€€ÿüŽ ä x10ÿÀŽ Ä x0;€€€€0ÞŽ ä x2/ÿÀŽ ä x66ÿþþþÿàÿ ä r334   ÿüŽ â $x00   1
Ž â x/:€€€€/ÞŽ ä r665   ÿüŽ Ä x33˜º  ÿ    â $x//   2
Ž â x1ÿ   ÿÀŽ ä x;0€€€€;´Ž ä x66˜º  ÿ    ä x03ÿÀŽ ä x3<€€€€3ÞŽ â $x4 @  1Ž â x//€€€€:´Ž Ä x16ÿÀŽ â $x33   0
Ž â x6=€€€€6ÞŽ â &r//D   ÿ   ä $x24€   ÿ Ž â x33€€€€<´Ž â $x66   1
Ž ä $x14   ÿ Ž â x4ÿ   ÿÀŽ â &r0;@   /   â x22€€ÿÀŽ Æ &r/3A   0   â x02€€€€ÿüŽ â $x5 @  4Ž â x4ÿ     ä x31€€ÿÀŽ ä x10ÿþþþÿàÿ â $x84 @  4Ž â x03€€€€ÿüŽ â $x45   ÿ Ž â r12   ÿüŽ â $x15€   ÿ Ž â x00ÿþþþÿàÿ Ä x44€€ÿÀŽ ä x21€€ÿÀŽ ä x66€€€€=´Ž ä r003   ÿüŽ â $x38€   ÿ Ž â x54€€€€ÿüŽ ä x12€€€€ÿüŽ â &r/6   /   â x55ÿþþþÿàÿ Ä x63€€ÿÀŽ ä x31ÿþþþÿàÿ ä xÿ     ä r554   ÿüŽ â $x48   ÿ Ž â x76€€€€ÿüŽ ä r332   ÿüŽ â $x @  Ž â x.ÿ   .  Ä x77ÿþþþÿàÿ ä x33˜º  ÿ    ä x84€€ÿÀŽ â $x1€   ÿ Ž â r776   ÿüŽ â $r..9   EŽ â x63ÿÀŽ ä x98€€€€ÿüŽ Ä x3"€€€€3ÞŽ ä x11€€ÿÀŽ ä x45˜º  ÿ    ä x99ÿþþþÿàÿ â $x33   6
Ž â x21€€€€ÿüŽ ä x54ÿÀŽ Ä r998   ÿüŽ ä x4#€€€€4ÞŽ ä x22ÿþþþÿàÿ ä x33€€€€"´Ž ä x"9˜º  ÿ    â $x44   5
Ž â r221   ÿüŽ â $x   ÿ Ž â x6"ÿÀŽ Ä x1%€€€€"ÞŽ â &r33D   ÿ   â x#4€€€€#´Ž ä x"€€ÿÀŽ â $x61   6
Ž â x00˜º  ÿ    â &r@#@   3   â x"€€€€ÿüŽ ä x77˜º  ÿ    Ä x˜º  ÿ    ä x#0ÿÀŽ ä x66€€€€%´Ž ä x0'€€€€0ÞŽ äO x%ÿþþþÿàÿ ä x47ÿÀŽ Ä xÿÀŽ ä x7$€€€€7ÞŽ ä x&€€€€ÞŽ â $x00   #
Ž â r%%"   ÿüŽ ä x#2˜º  ÿ    â $x77   4
Ž ä $x   
Ž â x%˜º  ÿ    Ä x"#ÿÀŽ ä x1Pÿ   ÿÀŽ ä x#*€€€€#ÞŽ ä� x77€€€€$´Ž ä x€€€€&´Ž ä x%ÿÀŽ Ä x$+€€€€ÞŽ â $x##   "
Ž â x'0€€€€'´Ž â $x1P @  1Ž ä &r   ÿ   ä $x$$   %
Ž â x##€€€€*´Ž â $x"1   ÿ Ž â xPÿ   P  â &r'      â x+$€€€€+´Ž â $x1€   ÿ Ž â x$"€€ÿÀŽ â &r#      Ä $x&P @  PŽ â x€€ÿÀŽ â &r++      â x#$€€€€ÿüŽ â $x&€   ÿ Ž âx"€€€€ÿüŽ ä x##ÿþþþÿàÿ â $x&&   ÿ Ž â x""ÿþþþÿàÿ ä x%€€ÿÀŽ Ä r##$   ÿüŽ ä r""   ÿüŽ ä x$%€€€€ÿüŽ ä x&&€€ÿÀŽ ä x#˜º  ÿ    ä x"˜º  ÿ    Ä x$$ÿþþþÿàÿ ä x'&€€€€ÿüŽ ä x#ÿÀŽ ä x-€€€€ÞŽ äx"ÿÀŽ ä r$$%   ÿüŽ Ä x,€€€€ÞŽ ä x''ÿþþþÿàÿ â $x   #
Ž â x#$˜º  ÿ    â $x   "
Ž â r''&   ÿüŽ ä x-€€€€-´Ž ä x"#ÿÀŽ Ä x'˜º  ÿ    ä x#(€€€€#ÞŽ ä x€€€€,´Ž ä x%ÿÀŽ â $x##   "
Ž â x$)€€€€ÞŽ â &r   ÿ   Æ x##€€€€(´Ž â $x$$   %
Ž ä &r-      ä &r77A   @   â x)$€€€€)´Ž â &r#      â xNð   ÿÀŽ â &r6   7   â xK   ÿÀŽ Ä xNN   ÿÀŽ â &r)      â xKð   ÿÀŽ â $r/   ÿŽ â xÿ     â $rNN   ÿŽ â xÿ     â $x//   +Ž ä $x   Ž â r>>>   ÿ� â $r+   Ž â r   ÿ� â $r   NŽ â r///   ÿ� Ä r   ÿ� ä r>   ÿú� ä r.   ÿú� ä r//   ÿú� ä r   ÿú� ä x   ÿàÿ Ä xÿ     ä xÿ     ä xÿ   /  ä xÿ     ä |    pbð ä r   ÿ� Ä r   ÿ� ä r/   ÿ� ä r   ÿ� â 0rÿ    A   â xÿ     â 0rÿI    A   â xÿ     â 0rÿ    A   â xÿ     â 0rÿJ    A   â xÿ     â 0rÿB    A   â x   ÿà÷ â  r     A  â Er        â  r     A  âEr        â  r     A  âEr        â  r     A  â Er        â #r      â #r      â #r      â x€   ÿàÿ â #r      â $rÿÿ   Ž â D	       À â Gy   Ûÿÿÿÿƒ ê Ay        € ê ˆ  þÿ H   è ˆ €þÿ H   è ˆ  ÿÿ H   è ˆ €ÿÿ H   è {          ì M)        € ê „y     H   " z 	 r  p`ô â $z   x  	Ž Ä „y €   H   b r ÿ   pVt Æ „y    H   ¨ „y    H   è„y €  H   ( „y    H   ( „y
    H   ( „y €  H   ( „y €  H   ( „y    H   ( „y €  H   " !r       Æ „y €  H   " !r       â/ !r       äO y      '   b !r       â� !r       â!r       ä !r
       â !r       È ‰      ¢ !r       â $z  ~   Ž â/ !r       Æ ‰      b !r       Æ ‰      â !r       Ê ‰      " !r       âO !r       â/ !r       È� ‰      h ‰      ¢ !r       Æ ‰      ( ‰      â !r       â/ !r       ÈO ‰
 €    b !r       Æ ‰ €    " !r       Æ� ‰ €    ¨ ‰ €    â !r

       â/ !r       È ‰
 @    " x	   ÿàÿ â !r       äO ‰ @    b x	   ÿàÿ â !r       ä� ‰ @    ¢ z  r  p`ö ä z  r  p`ò â ‰ @    â x    pVö Ä x    pVò  â !r
       â x
	   ÿàÿ â $t	ÿ   ÿ Ž â !r       ä/ ‰      " z 
 r  p`ð â !r       äO ‰      b x    pVp  â !r       Ä� ‰      ¨ ‰      â !r       â !r       â/ %v  j  	 Ž â !r       ÈO †©     â !r       Æ� †¹    è †‰    â M        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      &   " z ÿ \  pPò â ¹z  F   
   Æ z ÿ ]  Sò Ú $ÿ   ÿ Ž È % \   Ž Ê �Q    ™ ¢  $rÿÿ   ÿ Ž â Ey  0    € â $rCÿÿ   ÿ Ž â yB      %   b '  m   Ž Æ y      "   â $rDÿÿ   ÿ Ž â  ÿ n     â 'š s   Ž â yA      !   $  ÿ   ÿáÿ ä y      '   & $ o  Ž Ä $’ÿÿ    Ž ä $xBB   ÿ Ž â/ r ÿ   pRð Æ� $xB   AŽ âx A   pDp  â $x     AŽ Æ z  r  pdp  ä x	ÿ      ä šQÿ t    È rQ     € ÊO G	  À    € ê z ÿ `  pPò â $rDÿÿ   ÿ Ž â z ÿ ^  pPð â $rCÿÿ   ÿ Ž â z ÿ b  p ô ä z ÿ a  Sò ä z ÿ _   Sð ä z ÿ c   'ò Ö $
 ~  ÿŽ âÿ   B  â$š ~  ÿŽ ä $
 x  ÿŽ âB   Aàë â$š x  ÿŽ â ’
B   AàÇ ä ˜ÿ   B  ä    ÿàõ Ä ÿ   ÿÄþ ä ’
   ÿàû ä ’
ÿ   ÿ„þ â $ÿÿ    â 
 ^  ÿ„ ä ’
ÿ
   ÿäÿ ä š b  ÿ† Ä 
 _   ä š c  
� Æ �	D     h �™C     d Ay        € ê r		    ÿ� â ¹z  l      â $zE }  ÿŽ â ™x?    â xOÿ   	  â $z w  ÿŽ â Ey  M    € â $rÿÿ   ÿ Ž â xF      ÿ  â | O   pbð â $rVÿÿ   ÿ Ž â rEE   ÿàó â 'z y   Ž â xH      ÿ  â xJ      ÿ  â rMÿÿ   ÿäÿ  â $rXÿÿ   ÿ Ž Ì G	  PL    € ê x	øÿÿÿÿÀŽ â $zB p  OŽ âzÿ z    â $vSÿ `  ÿ Ž â xB   ÿàÿ â $x    
Ž â x	B   ÿàÿ â $zL |  OŽ â xF      ÿ  â $x
   ÿ Ž â xÿ     â $z |  Ž â xB   ÿàÿ â %x

   ÿŽ â rV ÿ       â xH      ÿ  â xJ      ÿ  â $zLQ v  LŽ Ä $zQQ v  Ž ä $z p  LŽ ä %xQR   
Ž È $z p  LŽ âxP$   ÿàñ â $z	 p  LŽ ä %xR   
Ž È %xR   
Ž È %x
R   
Ž È $vUÿ a  ÿ Ž ä $vWÿ X  ÿ Ž ä $vYÿ Y  ÿ Ž ä $rXÿÿ   ÿ Ž ä $xNO   ÿ Ž ä %xR   Ž È %xR   Ž È %xR   Ž È %xQR   Ž È $rRÿÿ     ä rW   ÿàñ Ê $x   Y  â rW   ÿàñ È �yB   ‘ ¢ rY   ÿä  â $vgÿ p  ÿ Ž ä $tTÿ   ÿ Ž â �yZJ   ‘ è  �y    • ( �y   • â r WP   ÿàñ â %x8gR   Ž Ä �yd"   • ä $x!Y   R  ä �y"8B   ‘ ä  %x gR    Ž ä�y`$   • è �y Þÿÿ • è �y# àÿÿ • â %x0gR   8Ž Æ �y[ þÿÿ • è  �yc0B   ‘ â %x.gR    Ž Æ �y\     • è  �y&.Þÿÿ • è  �ye.àÿÿ • â  %x,gR   .Ž Æ �y].þÿÿ • è  �ya,Þÿÿ • è �y_,àÿÿ • è �y88J   ‘ è �yb.    • â  $x   ÿ Ž ÊO x   ÿÀŽ È xÿ    øŽ äx ÿ      Æ %v( 
 T Ž È $x@   ÿ Ž ä �y((    › ¦ x    øŽ Ê %v* 
 T Ž È %xgR   0Ž ä �y**    › ( �y^B   ‘ " x    ÿ   È x    ÿÀŽ È xÿ   øŽ â� $x"@   ÿ Ž â r E   Nà‘ â $x   ÿ Ž â xhc   ÿ   â %v$ 
 T Ž â xÿ     â �yJ   ‘ â rM   ÿ$~  â $tÿ$   ÿ Ž â x   øŽ â/ �y$$    › ä %v  Z   Ž È $x$   ÿŽ È $x   Ž ä %v 
 T Ž Æ �y@   ™ è �y?   ™ è �y    › è �y=   ™ è �y>   ™ è �y<   ™ è �y;   ™ è �y:   ™ è �y     ™ â $x"   ÿ Ž Ê x   ÿÀŽ È xÿ   øŽ â$x'c@   ÿ Ž ä $r.ÿÿ   W Ž ä $r/ÿÿ   Y Ž â xhh   ÿÀŽ â %v 
 T Ž â xÿ     â �y    • è" �y    › â x   øŽ Ê %v 
 T Ž Ì �y    › â x"   ÿ   È x   ÿÀŽ È xÿ   #øŽ Ê %v 
 T Ž Ì �y    › â $x""   ÿ Ž â x#ÿ   #  È x!#   "øŽ â $x"c   ÿ Ž È %v ! 
 T Ž â x#"   ÿÀŽ Ê �y      › â x##ÿ   &øŽ Ê %v"# 
 T Ž Ì �y""    › â x&ÿ   &  È x'&   'øŽ Ê %v&' 
 T Ž È %x2QR   .Ž ä �y&&    › è %x4gR   2Ž âxihÿ   eøŽ â �y92    • è  %x6gR   4Ž âxhÿ   e  â �y44    • è %x.gR   6Ž â xgdÿ   ÿÀŽ â �yf6    • â xeÿ   d  â %v2i 
 T Ž Ä �y5,þÿÿ • ä  $xdd @  gŽ ä $xkc   ÿ Ž â �y22    › â $x6d€   ÿ Ž ä$xdd   ÿ Ž â xkh   køŽ â $xje @  eŽ â x76€€ÿÀŽ â �yc0J   ‘ â xed€€ÿÀŽ Ä xgj   ÿ   â �y6,    • â  xd7€€€€ÿüŽ Æ �y..    • â  %v0k 
 T Ž â/ xhg€€ÿÀŽ ä xddÿþþþÿàÿ â $xjj   ÿ Ž â xge€€€€ÿüŽ â �y00    › â rdd7   ÿüŽ ä xjj€€ÿÀŽ Ä x,gÿþþþÿàÿ ä xih€€€€ÿüŽ ä x7j€€€€ÿüŽ ä x-d˜º  ÿ    ä r,,e   ÿüŽ ä xiiÿþþþÿàÿ Ä x77ÿþþþÿàÿ ä xd-ÿÀŽ ä x,,˜º  ÿ    ä riih   ÿüŽ ä rj7j   ÿüŽ ä x7,ÿÀŽ Ä xii˜º  ÿ    ä xjj˜º  ÿ    ä x-(€€€€-ÞŽ äO x,)€€€€,ÞŽ Æ $x--   d
Ž ä $x,,   7
Ž â x7jÿÀŽ ä x--€€€€(´Ž ä x(iÿÀŽ ä xi*€€€€iÞŽ äxj+€€€€jÞŽ â $xe^   ÿ Ž â x,,€€€€)´Ž â $xii   (
Ž â rdj7  €ÿàÿ Ä x)`ÿ   ÿÀŽ ä x(e   ÿÀŽ ä xdd€€€€+´Ž â $x+` @  )Ž â x)(ÿ   aøŽ ä x`ÿ   `  Æ %v() 
 T Ž â xhi€€€€*´Ž Æ $xe` @  `Ž â xaÿ   a  â $xj^@   ÿ Ž â �y((    › ¢ $x*+€   ÿ Ž ä$x++   ÿ Ž â x`e   ÿ   â$xee   ÿ Ž â xma   jøŽ ä x7*€€ÿÀŽ Ä xa+€€ÿÀŽ ä x*7€€€€ÿüŽ ä xke€€ÿÀŽ ä xea€€€€ÿüŽ ä xg`€€ÿÀŽ ä x`*ÿþþþÿàÿ â %v*m 
 T Ž â xjeÿþþþÿàÿ Ä xig€€€€ÿüŽ ä rjja   ÿüŽ â �y**    › " r``7   ÿüŽ ä x7k€€€€ÿüŽ ä xjj˜º  ÿ    ä xliÿþþþÿàÿ Ä xn7ÿþþþÿàÿ ä xajÿÀŽ ä x7`˜º  ÿ    ä xj%€€€€jÞŽ ä� rllg   ÿüŽ â $xg^   ÿ Ž â x`7ÿÀŽ â $xjj   a
Ž â x7$€€€€7ÞŽ Ä xel˜º  ÿ    ä xlg   ÿÀŽ â $x77   `
Ž â rnnk   ÿüŽ ä xaj€€€€%´Ž â &r--@   ÿ   â x%lÿ   _øŽ ä xnn˜º  ÿ    â &r,,?   -   â x77€€€€$´Ž â %v$% 
 T Ž â x^^   ÿ   Ä x_ÿ   _  ä x-nÿÀŽ â �y$$    › â xn€€€€nÞŽ ä x__   ^øŽ â &rhh=   ,   â x`eÿÀŽ â $xnn   -
Ž â xe€€€€eÞŽ â %v,_ 
 T Ž È $xee   `
Ž ä �y,,    › ¢ &r77>   ÿ   ä xee€€€€´Ž ä &r7a<   7   â xZð   ÿÀŽ ä x_n€€€€´Ž â &rd;   h   â xZZ   ÿÀŽ â &r7e:   7   â x^[ÿ   ÿÀŽ Ä xÿ     â &r7_    7   ä $rZZ   ÿŽ ä $x^[ @  ^Ž ä $x   7Ž ä $r7   ZŽ ä $x7^€   ÿ Ž Ä $xZ^   ÿ Ž â x[ÿ   [  ä x77€€ÿÀŽ ä x^Z€€ÿÀŽ ä r   ÿ� ä x7€€€€ÿüŽ ä r[[[   ÿpŽ Ä x/^€€€€ÿüŽ ä r   ÿú� ä xZÿþþþÿàÿ â $x[€   ÿ Ž â x//ÿþþþÿàÿ ä xÿ     ä rZZ7   ÿüŽ Ä r//^   ÿüŽ ä r   ÿ� â $x[   ÿ Ž â x€€ÿÀŽ ä xZ˜º  ÿ    ä xZ/˜º  ÿ    ä x^ÿÀŽ Ä x7€€ÿÀŽ ä x€€€€ÞŽ ä xZÿÀŽ ä x/€€€€ÿüŽ ä xZ€€€€ZÞŽ â $x   ^
Ž â x[7€€€€ÿüŽ Ä x^/ÿþþþÿàÿ â $xZZ   
Ž â x\ÿ   ÿÀŽ ä x`[ÿþþþÿàÿ ä r^^   ÿüŽ ä x€€€€´Ž ä xZZ€€€€´Ž â $x\ @  Ž â r``7   ÿüŽ Ä x^˜º  ÿ    â &r@   ÿ   â x``˜º  ÿ    ä xÿÀŽ ä x€€€€ÞŽ â &rZZ?      â x   ÿ   â$x   ÿ Ž â x`ÿÀŽ â $x   
Ž â x`€€€€`ÞŽ Ä x€€ÿÀŽ ä x€€ÿÀŽ â $x``   
Ž â x€€€€ÿüŽ ä x€€€€´Ž ä x€€€€ÿüŽ ä x\ÿ   \  Ä xÿþþþÿàÿ ä xÿþþþÿàÿ ä x``€€€€´Ž â $x\ @  \Ž â r   ÿüŽ ä r   ÿüŽ â &r=   Z   â x˜º  ÿ    â $x€   ÿ Ž â x˜º  ÿ    Ä x   ÿ   ä xÿÀŽ ä x€€ÿÀŽ ä x€€€€ÞŽ ä xÿÀŽ ä x€€€€ÞŽ Ä x€€ÿÀŽ ä xZ]ÿ   ÿÀŽ â $x   
Ž â x€€€€ÿüŽ â $x   
Ž â x€€€€ÿüŽ â $xZ] @  ZŽ â x€€€€´Ž ä xÿþþþÿàÿ Ä xÿþþþÿàÿ â $xZ€   ÿ Ž â r   ÿüŽ â &r>   ÿ   â x€€€€´Ž â $xZ   ÿ Ž â x€€ÿÀŽ ä x˜º  ÿ    â &r<      â r   ÿüŽ Ä x€€ÿÀŽ ä x€€€€ÿüŽ ä xÿÀŽ ä x €€€€ÞŽ ä x˜º  ÿ    ä x/€€€€ÿüŽ Ä xÿþþþÿàÿ ä r  €ÿàÿ ä xÿÀŽ ä x/ÿþþþÿàÿ ä r   ÿüŽ ä x]ÿ   ]  Ä x!€€€€ÞŽ ä r   ÿüŽ â $x]] @  ]Ž â x˜º  ÿ    â $x   
Ž â x€€€€ ´Ž â $x]€   ÿ Ž â x˜º  ÿ    ä xÿÀŽ Ä x"€€€€ÞŽ â &r:      â x!€€€€!´Ž ä xÿÀŽ ä x#€€€€ÞŽ â $x   
Ž â x€€ÿÀŽ ä x]   ÿ   â $x   
Ž â x€€€€ÿüŽ â &r!       â x€€ÿÀŽ Ä x8ð   ÿÀŽ â &r`;      â x88   ÿÀŽ ä xÿþþþÿàÿ ä x€€€€ÿüŽ â $r88   ÿŽ â xÿ     ä x ÿþþþÿàÿ Ä x€€€€"´Ž ä r   ÿüŽ â $x   Ž â r     ÿüŽ â $r   8Ž â x€€€€#´Ž â &r@   ÿ   â x˜º  ÿ    ä xbÿ   ÿÀŽ Ä x  ˜º  ÿ    â &r?      â xÿÀŽ â $xb @  Ž â x&€€€€ÞŽ ä xbÿ   b  ä x ÿÀŽ ä x '€€€€ ÞŽ â $x   
Ž Ä $x€   ÿ Ž âr    €ÿàÿ â $x"b @  bŽ ä $x   ÿ Ž â x€€ÿÀŽ â $x"€   ÿ Ž Æ x€€ÿÀŽ ä x€€€€ÿüŽ â $x""   ÿ Ž â x€€ÿÀŽ ä x€€€€ÿüŽ ä xÿþþþÿàÿ ä xÿþþþÿàÿ Ä x!€€€€ÿüŽ ä x""€€ÿÀŽ ä r   ÿüŽ ä r   ÿüŽ ä x!ÿþþþÿàÿ ä x"€€€€ÿüŽ Ä x˜º  ÿ    ä r   ÿüŽ ä x!ÿþþþÿàÿ ä x˜º  ÿ    ä xÿÀŽ ä x2€€€€ÞŽ Ä r!!"   ÿüŽ ä xÿÀŽ â $x   
Ž â x3€€€€ÞŽ ä x!˜º  ÿ    Æ $x   
Ž â xÿÀŽ ä x1€€€€ÞŽ ä x˜º  ÿ    Æ $x   
Ž â x€€€€&´Ž ä x"ÿÀŽ ä x1€€€€1´Ž ä x€€€€2´Ž ä x0€€€€ÞŽ ä x5ÿ   ÿÀŽ â &r=      â x  €€€€'´Ž â &r>   ÿ   â x3€€€€3´Ž â $x5 @  Ž â r"  €ÿàÿ â &r  ;      â r   ÿ� â &r3<      â x€€€€0´Ž â $x€   ÿ Ž âxcð   ÿÀŽ â$x   ÿ Ž â xcc   ÿÀŽ â &r:      â x€€ÿÀŽ Ä r   ÿú� â &r1       â xÿ     â $rc    ÿŽ â x5ÿ   5  ä x€€ÿÀŽ ä x€€€€ÿüŽ â $r   Ž â x€€€€ÿüŽ â $x55 @  5Ž â xÿþþþÿàÿ Ä r    ÿàÿ â $x5€   ÿ Ž â xÿþþþÿàÿ ä r   ÿüŽ ä r   ÿüŽ ä x˜º  ÿ    ä x€€ÿÀŽ Ä x˜º  ÿ    ä xÿÀŽ â $x 5   ÿ Ž â x€€€€ÿüŽ ä x(€€€€ÞŽ äO xÿÀŽ ä xÿþþþÿàÿ â $x   
Ž â x)€€€€ÞŽ Ä x  €€ÿÀŽ ä r   ÿüŽ â $x   
Ž â x! €€€€ÿüŽ ä x€€€€(´Ž ä x6ÿ   ÿÀŽ ä x˜º  ÿ    â &r@   ÿ   â x!!ÿþþþÿàÿ Ä x€€€€)´Ž ä xÿ   6  â $x6 @  Ž â xÿÀŽ ä x*€€€€ÞŽ â&r?      â r!!    ÿüŽ â $x# @  Ž â r  €ÿàÿ â $x€   ÿ Ž â x!˜º  ÿ    â $x   ÿ Ž Ä $x!#€   ÿ Ž â x€€ÿÀŽ ä xÿÀŽ ä x+€€€€ÞŽ ä x €€ÿÀŽ ä x!!€€ÿÀŽ ä x€€€€ÿüŽ â $x   
Ž â x"!€€€€ÿüŽ Ä x €€€€ÿüŽ ä xÿþþþÿàÿ ä x##   ÿ   ä x""ÿþþþÿàÿ ä xÿþþþÿàÿ ä r   ÿüŽ Ä x##€€ÿÀŽ ä r""!   ÿüŽ ä r    ÿüŽ ä x˜º  ÿ    ä x!#€€€€ÿüŽ ä x˜º  ÿ    Ä x ÿÀŽ ä x&!ÿþþþÿàÿ ä x$€€€€ÞŽ ä� xÿÀŽ ä r&&#   ÿüŽ ä x!"˜º  ÿ    â $x    
Ž â x%€€€€ÞŽ Ä x&&˜º  ÿ    ä x !ÿÀŽ â $x   
Ž â x!,€€€€!ÞŽ ä x€€€€$´Ž ä x&ÿÀŽ â $x!!    
Ž â x&-€€€€&ÞŽ â &r>   ÿ   â x%€€€€%´Ž Ä x€€€€*´Ž â $x&&   
Ž â x!!€€€€,´Ž â &r%<      â x€€€€+´Ž â &r=      â x-&€€€€-´Ž â &r!:      â xð   ÿÀŽ â &r;      â x   ÿÀŽ â &r-       â xÿ     Æ $r   ÿŽ ä $x   Ž â xÿ     â 0rÿ    A   â xÿ     â 0r9ÿ9    A   â r   ÿ� â $r   Ž â r   ÿ� ä z ÿ `  pPð â  r99     A  â Er        Ä r   ÿ� ä r   ÿú� ä r   ÿú� â 0rÿ4    A   â z ÿ a   Sð â #rX9   X   â xÿ     ä xÿ     Ä xÿ     â  r     A  â Er        ä r   ÿ� ä r   ÿ� â 0rÿf    A   â #rV   V   â 0rÿ.    A   â xÿ     Ä xÿ     â  r     A  â r     A  â Er        ä Er        â G‰   	    € ê rS   ÿàó È rU   ÿäÿ  ä r&S   ÿàó Æ �y#B    ¤  $x'   UŽ  Ê �y&    è �y"&    ( �y%&"    ( �y$&$    ( �yJ    " r& S       â/ $r'ÿÿ   U Ž È %x&QR   &Ž È $x#   ÿ Ž äO$x#@   ÿ Ž Æ x   ÿÀŽ È xÿ   øŽ Ê� %v 
 T Ž â xÿ     Ê �y    › ¢ x   øŽ Ê %v 
 T Ž Ì �y    › â $x#   ÿ Ž Ê x   ÿÀŽ È xÿ   "øŽ â$x##   ÿ Ž â x"ÿ   "  Æ %v  
 T Ž âx#"   #øŽ â �y&     ( �y      › " %v"# 
 T Ž Ì �y""    › " x(%ÿ   ÿÀŽ Ê $x)% @  (Ž â x%ÿ   %  Æ $x()€   ÿ Ž ä$x))   ÿ Ž ä $x*% @  %Ž â x((€€ÿÀŽ ä x))€€ÿÀŽ ä x(€€€€ÿüŽ ä x%)€€€€ÿüŽ Ä x'*   ÿ   ä/x&%ÿþþþÿàÿ ä xÿþþþÿàÿ ä x''€€ÿÀŽ â $x**   ÿ Ž â r&&)   ÿüŽ ä r(   ÿüŽ Ä x%'€€€€ÿüŽ ä x&&˜º  ÿ    ä x+*€€ÿÀŽ ä x˜º  ÿ    ä x*%ÿþþþÿàÿ ä x%&ÿÀŽ Ä x(ÿÀŽ ä r**'   ÿüŽ ä x)+€€€€ÿüŽ ä x'ÿ   $  ä x&€€€€&ÞŽ äO x€€€€ÞŽ Æ $x&&   %
Ž ä $x   (
Ž â x%*˜º  ÿ    ä x&&€€€€´Ž ä x$ÿ   ÿÀŽ ä x€€€€´Ž ä x%ÿÀŽ Ä x%€€€€%ÞŽ â� $x$ @  Ž â x()ÿþþþÿàÿ â &r@   ÿ   ä $x%%   
Ž â r((+   ÿüŽ â $x€   ÿ Ž ä &r&&?      â r*''   ÿpŽ â $x   ÿ Ž â x((˜º  ÿ    Ä x€€ÿÀŽ â $x'*€   ÿ Ž â x(ÿÀŽ ä x$€€ÿÀŽ ä x(€€€€(ÞŽ ä x€€€€ÿüŽ â $x**   ÿ Ž â x''€€ÿÀŽ â $x((   
Ž â xÿþþþÿàÿ Ä x$€€€€ÿüŽ ä r   ÿüŽ ä x)'€€€€ÿüŽ ä xÿþþþÿàÿ ä x**€€ÿÀŽ ä x˜º  ÿ    Ä x)ÿþþþÿàÿ ä r$   ÿüŽ ä x)*€€€€ÿüŽ ä x$ÿÀŽ ä r''   ÿüŽ ä x €€€€ÞŽ Äx˜º  ÿ    ä x))ÿþþþÿàÿ ä r$  €ÿàÿ ä xÿÀŽ ä r))*   ÿüŽ ä x''˜º  ÿ    Ä x!€€€€ÞŽ ä x€€€€ ´Ž ä x )˜º  ÿ    ä x$'ÿÀŽ â $x   
Ž â x'"€€€€'ÞŽ ä x ÿÀŽ Ä x #€€€€ ÞŽ â $x''   $
Ž â x!€€€€!´Ž â &r>   ÿ   â x%%€€€€´Ž â $x     
Ž â x''€€€€"´Ž â &r!<      â x((€€€€´Ž â &r%%=   &   â x# €€€€#´Ž â &r':      â x   ÿÀŽ â &r%(;   %   Ä &r#       â xÿ     â $r%   ÿŽ ä $x%%   Ž ä $r   Ž Æ r%%%   ÿ� È r%   ÿú� È xÿ     È r   ÿ� â 0rÿ    A   Æ xÿ     ä  r     A  ä Er        Ê #rK   K   ä G‰   	    € ê rS   ÿàó Ê $x   UŽ  â r&S   ÿàó È �y#B    ¢  r'U   ÿäÿ  Ê �y&    è �y"&    ( �y%&"    ( �y$&$    ( �yJ    " r' U       â/ $r&ÿÿ   S Ž Ä $x#   ÿ Ž äO$x#@   ÿ Ž Æ x   ÿÀŽ È xÿ   øŽ Ê� %v 
 T Ž â xÿ     Ê �y    › ¢ x   øŽ Ê %v 
 T Ž Ì �y    › â $x#   ÿ Ž Ê x   ÿÀŽ È xÿ   "øŽ â$x##   ÿ Ž â x"ÿ   "  Æ %v  
 T Ž â x#"   #øŽ Ê �y      › " %v"# 
 T Ž Ì �y""    › " xB   ÿàÿ Ê $z p  LŽ È %x&R   &Ž Ê �y&     " x(%ÿ   ÿÀŽ Ê $x)% @  (Ž â x%ÿ   %  Æ $x()€   ÿ Ž ä$x))   ÿ Ž ä $x*% @  %Ž Æ x))€€ÿÀŽ ä x((€€ÿÀŽ ä x%)€€€€ÿüŽ â $x'*€   ÿ Ž â/ x(€€€€ÿüŽ ä x&%ÿþþþÿàÿ ä xÿþþþÿàÿ Ä x''€€ÿÀŽ ä r&&)   ÿüŽ ä x**   ÿ   ä r(   ÿüŽ ä x%'€€€€ÿüŽ ä x&&˜º  ÿ    Ä x+*€€ÿÀŽ ä x˜º  ÿ    ä x*%ÿþþþÿàÿ ä x%&ÿÀŽ ä x(ÿÀŽ ä r**'   ÿüŽ Ä x)+€€€€ÿüŽ ä x'ÿ   $  ä x&€€€€&ÞŽ äO x€€€€ÞŽ Æ $x&&   %
Ž ä $x   (
Ž â x%*˜º  ÿ    ä x&&€€€€´Ž ä x$ÿ   ÿÀŽ äx€€€€´Ž ä x%ÿÀŽ â $x$ @  Ž â x%€€€€%ÞŽ â� &r@   ÿ   â x()ÿþþþÿàÿ Æ $x%%   
Ž â r((+   ÿüŽ â &r&&?      â x   ÿ   â$x   ÿ Ž â x((˜º  ÿ    ä x€€ÿÀŽ â $x*' @  'Ž â x$€€ÿÀŽ ä x(ÿÀŽ Ä x€€€€ÿüŽ â $x'*€   ÿ Ž âx(€€€€(ÞŽ ä xÿþþþÿàÿ â $x**   ÿ Ž â x''€€ÿÀŽ â $x((   
Ž â x$€€€€ÿüŽ ä r   ÿüŽ Ä x)'€€€€ÿüŽ ä xÿþþþÿàÿ ä x**€€ÿÀŽ ä x˜º  ÿ    ä x)ÿþþþÿàÿ ä r$   ÿüŽ Ä x)*€€€€ÿüŽ ä x$ÿÀŽ ä x €€€€ÞŽ är''   ÿüŽ ä x))ÿþþþÿàÿ â $x   $
Ž â x˜º  ÿ    Ä r))*   ÿüŽ ä x''˜º  ÿ    ä xÿÀŽ ä x!€€€€ÞŽ ä x€€€€ ´Ž ä x )˜º  ÿ    Ä x$'ÿÀŽ ä x'"€€€€'ÞŽ ä r  €ÿàÿ ä x ÿÀŽ â $x''   $
Ž â x #€€€€ ÞŽ â &r>   ÿ   â x!€€€€!´Ž Ä x%%€€€€´Ž â $x     
Ž â x''€€€€"´Ž â &r!<      â x((€€€€´Ž â &r%%=   &   â x# €€€€#´Ž â &r':      â x   ÿÀŽ â &r%(;   %   ä &r#       â xÿ     â $r%   ÿŽ Ä $x%%   Ž ä $r   Ž Æ r%%%   ÿ� È r%   ÿú� È xÿ     È r   ÿ� â 0rÿ    A   Æ xÿ     ä  r     A  ä Er        Ê #rJ   J   ä G‰   	    € ê rS   ÿàó Ê $x   UŽ  â r"
S   ÿàó È �y&B    ¢  $x#   UŽ  Ê �y"    è �y%"    ( �y*""    ( �y$"$    ( �yJ    " x&   ÿ   âO$x&@   ÿ Ž Æ x   ÿÀŽ È xÿ   øŽ Ê� %v 
 T Ž â xÿ     Ê �y    › ¢ x   øŽ Ê %v 
 T Ž Ì �y    › â $x&   ÿ Ž Ê x    ÿÀŽ È x! ÿ   %øŽ â$x&&   ÿ Ž È %v ! 
 T Ž â x%ÿ   %  Ê �y      › " x%%   &øŽ Ê %v&% 
 T Ž Ê �y"&    › "# xB   ÿàÿ Ê $z p  LŽ ä $r&ÿÿ   S Ž ä/ $r'ÿÿ   U Ž È %x(R   &Ž Ê �y(     " x%*ÿ   ÿÀŽ È r+*%   ÿpŽ Ê $x%+€   ÿ Ž ä$x++   ÿ Ž â x*ÿ   *  ä x%%€€ÿÀŽ ä x++€€ÿÀŽ â $x** @  *Ž â x%€€€€ÿüŽ ä x'+€€€€ÿüŽ Ä x&ÿþþþÿàÿ â $x*€   ÿ Ž â x('ÿþþþÿàÿ È/ x'€€ÿÀŽ ä r((+   ÿüŽ ä r&&%   ÿüŽ ä x%'€€€€ÿüŽ ä x((˜º  ÿ    ä x)*   ÿ   Ä x&˜º  ÿ    ä x*%ÿþþþÿàÿ ä x%(ÿÀŽ ä x&ÿÀŽ ä r**'   ÿüŽ ä x))€€ÿÀŽ È x+)€€€€ÿüŽ ä x'ÿ   $  ä x(€€€€(ÞŽ äO x€€€€ÞŽ Æ $x((   %
Ž ä $x   &
Ž â x%*˜º  ÿ    ä x((€€€€´Ž ä x$ÿ   ÿÀŽ ä x€€€€´Ž ä x%ÿÀŽ Ä x%€€€€%ÞŽ â� $x$ @  Ž â x&+ÿþþþÿàÿ â &r@   ÿ   ä $x%%   
Ž â r&&)   ÿüŽ â $x€   ÿ Ž ä&r((?      ä $x   ÿ Ž â x&&˜º  ÿ    Ä x€€ÿÀŽ ä r*''   ÿpŽ ä x$€€ÿÀŽ ä x&ÿÀŽ ä x€€€€ÿüŽ â $x'*€   ÿ Ž â x&€€€€&ÞŽ Ä xÿþþþÿàÿ ä x''€€ÿÀŽ â $x&&   
Ž â x$€€€€ÿüŽ ä x**   ÿ   ä r   ÿüŽ ä x)'€€€€ÿüŽ Ä xÿþþþÿàÿ ä x**€€ÿÀŽ ä x˜º  ÿ    ä x)ÿþþþÿàÿ ä r$   ÿüŽ ä x)*€€€€ÿüŽ Ä x$ÿÀŽ ä x €€€€ÞŽ är''   ÿüŽ ä x˜º  ÿ    ä x))ÿþþþÿàÿ â $x   $
Ž â xÿÀŽ Ä r))*   ÿüŽ ä x''˜º  ÿ    ä x!€€€€ÞŽ ä x€€€€ ´Ž ä x )˜º  ÿ    ä x$'ÿÀŽ â $x   
Ž â x'"€€€€'ÞŽ Ä x ÿÀŽ ä x #€€€€ ÞŽ â $x''   $
Ž â x!€€€€!´Ž â &r>   ÿ   â x%%€€€€´Ž â $x     
Ž â x''€€€€"´Ž â &r!<      â x&&€€€€´Ž â &r%%=   (   â x# €€€€#´Ž â &r':      â x   ÿÀŽ â &r%&;   %   Ä &r#       â xÿ     â $r%   ÿŽ Æ r%%   ÿàÿ â$r   Ž Æ r%%%   ÿ� È r%   ÿú� È xÿ     È r   ÿ� â 0rÿ    A   Æ xÿ     ä  r     A  ä Er        Ê #rI   I   â #rG   G   â #rF   F   â G‰  	    € ê rS   ÿàñ Ê $x   U  â r"S   ÿàñ È �yB    ¢  $x#	   U  Ê �y"    è �y"    ( �y!""    ( �y "$    ( �yJ    " $x   ÿ Ž ÊO x   ÿÀŽ È xÿ   øŽ â�$x@   ÿ Ž â xÿ     Æ %v 
 T Ž â x   øŽ Ê �y    › ¢ %v 
 T Ž Ì �y    › â x   ÿ   È x   ÿÀŽ È xÿ   øŽ â$x   ÿ Ž â xÿ     Æ %v 
 T Ž â x   øŽ Ê �y    › " %v 
 T Ž Ì �y    › " xB   ÿàÿ â $rTÿÿ   S Ž È $z p  LŽ È %x"R   TŽ Ê/ �y"     " x$!ÿ   ÿÀŽ Ê $x%! @  $Ž â x!ÿ   !  Æ $x$%€   ÿ Ž âx%%   ÿ   â $x&! @  !Ž Æ x$$€€ÿÀŽ ä x%%€€ÿÀŽ ä x$€€€€ÿüŽ ä x!%€€€€ÿüŽ â $x#&€   ÿ Ž â/ xÿþþþÿàÿ ä x"!ÿþþþÿàÿ Ä x##€€ÿÀŽ â $x&&   ÿ Ž â r$   ÿüŽ ä r""%   ÿüŽ ä x!#€€€€ÿüŽ ä x˜º  ÿ    ä x""˜º  ÿ    Ä x'&€€ÿÀŽ ä x&!ÿþþþÿàÿ ä x$ÿÀŽ ä x!"ÿÀŽ ä r&&#   ÿüŽ ä x%'€€€€ÿüŽ Ä x#ÿ      ä x€€€€ÞŽ äO x"€€€€"ÞŽ Æ $x   $
Ž â r""!  €ÿàÿ ä x!&˜º  ÿ    ä x""€€€€´Ž ä x ÿ   ÿÀŽ ä x€€€€´Ž ä x!ÿÀŽ Ä x!€€€€!ÞŽ â� $x  @  Ž â x$%ÿþþþÿàÿ â &r@   ÿ   ä $x!!   
Ž â r$$'   ÿüŽ â $x€   ÿ Ž ä&r""?      â x   ÿ   Ä x$$˜º  ÿ    ä x€€ÿÀŽ â $x&# @  #Ž â x €€ÿÀŽ ä x$ÿÀŽ ä x€€€€ÿüŽ â $x#&€   ÿ Ž â x$€€€€$ÞŽ Ä xÿþþþÿàÿ â $x&&   ÿ Ž â x##€€ÿÀŽ â $x$$   
Ž â x €€€€ÿüŽ ä r   ÿüŽ ä x%#€€€€ÿüŽ ä xÿþþþÿàÿ Ä x&&€€ÿÀŽ ä x˜º  ÿ    ä x%ÿþþþÿàÿ ä r    ÿüŽ ä x%&€€€€ÿüŽ ä x ÿÀŽ Ä x€€€€ÞŽ är##   ÿüŽ ä x%%ÿþþþÿàÿ â $x    
Ž â x˜º  ÿ    ä r%%&   ÿüŽ ä x##˜º  ÿ    Ä xÿÀŽ ä x€€€€ÞŽ ä x€€€€´Ž ä x%˜º  ÿ    ä x #ÿÀŽ ä x#€€€€#ÞŽ Ä r  €ÿàÿ ä xÿÀŽ â $x##    
Ž â x€€€€ÞŽ â &r>   ÿ   â x€€€€´Ž ä x!!€€€€´Ž â $x   
Ž â x##€€€€´Ž â &r<      â x$$€€€€´Ž â &r!!=   "   â x€€€€´Ž â &r#:      â x   ÿÀŽ â &r!$;   !   Ä &r       â xÿ     â $r!   ÿŽ ä $x!!   Ž ä $r   Ž Æ r!!!   ÿ� È r!   ÿú� È xÿ     È r   ÿ� â 0rÿ    A   Æ xÿ     ä  r     A  ä Er        Ê #rH   H   â xOO   ÿàÿ ä xSS   ÿàó ä | O   pbð ä xWW   ÿàõ â $rUÿÿ   UŽ  â xNN€   ÿàÿ ä rYÿY   ÿä Î D	       À â Gy   ¶ÿÿÿÿƒ ê Ay        € ê y       "   " Ey  0    € â r  ÿ   pRð är  ÿ   pRò Ö G‰  ð     € ê z ÿ `  pPð â $x €   AŽ Æ z ÿ a   Sð ä ˆs X þÿ H   ö ˆƒ V€þÿ H   è ˆƒ G ÿÿ H   è ˆƒ F€ÿÿ H   â G‰  p     € ê ˆs K   H   è ˆs V€þÿ H   è ˆs J€  H   è ˆs G ÿÿ H   è ˆs I   H   è ˆs F€ÿÿ H   è ˆs H€  H   ä Ay        € ê {          ì M        € ê y      &   ¢ z ÿ `  pPð Æ y      '   "> z ÿ a   Sð â $z  x  BŽ äO $tÿ   ÿ Ž ä $z  ~   Ž È %v  j   Ž È G	  P    € ê „yA     H   " z B r  p`ð Æ „y	A €   H   b r Aÿ   pVp  Æ „yA    H   ¨ „y A    H   è „yA €  H   ( „yA    H   ( „yA    H   ( „yA €  H   ( „yA    H   ( „yA €  H   ( „y
A €  H   " !rX       Æ „yA €  H   " !r		V       â/ !rG       âO !r         ä� !r	       â!r       â !r         ä !r       Æ ‰       b !r       Æ ‰      ¢ !rF       Æ ‰      â !r

       È !r
       â !r        â/ !r       ÈO ‰      " !r
       Æ� ‰      h ‰      ¨ ‰	
      â !r       â !r       â/ !r       ÈO ‰      " !r
	       Æ� ‰ €    h ‰  €    ¨ ‰	 €    â !r       â !r       È/ ‰ €    " !r
        ÆO ‰  @    b !r		       â� xB   ÿàÿ ä ‰
 @    ¤ z  r  p`ô ä ‰	 @    ä x A   pVt â !r       â !r         È/ ‰ @    " xB   ÿàÿ â !r
       âO xBB   ÿàÿ â ‰       b z  r  p`ò â !r	       ä� ‰      ¢ x A   pVò  Æ ‰      â !r       â !r        È/ ‰       " !‚D       â!r	       âO !r       Æ� †‰     â z B r  p`ð â !¢		D       â!’D       ä x A   pVp  ä †© 	   è †™    â !r        Ì M	        € ê !rD       Ê/ †y    â My        € ê $v ÿ h  ÿ Ž Ê x     pBð Ú G	  @    € ê x     pRð Ú G‰  À    € ê x     pRð Ú G	   $    € ê „y	A     H   " z B r  p`ð â Ey  p    € ä „y A    H   b r Aÿ   pVp  Æ „yA    H   ¨ „yA    H   è „yA    H   ( „yA  
  H   ( „yA €   H   ( „yA €  H   ( „yA €  H   ( „yA €  H   " !r		X       â !r  K       Æ/ !r	       âO !r         Æ� !r       â!r         È ‰      " !rV       Æ ‰       b !rJ       â !r       È !r       â !r       â !r        È/ ‰      ( ‰      b !r       â !r	       È/ ‰ €    ( ‰
	 €    h „yA €  H   ¢ !r       Æ „yA €
  H   â !r
	
       Æ/ ‰ @    ( ‰
 @    b !r       âO !r       â !r
       È/ ‰       ( ‰      b !r       â !r	        â !r       â/ G	  �     € ê !rC       â!r 	D       Æ  x;ª¸¿  A  Ì s        $ !t  €?    Ì s        $  r     A  È  r      A  Ê †y     ä Ay        € ê ‰       b Ey  Ð    € æ ‰      ¨ „yA    H   è „yA    H   ( „yA  	  H   â !r         Æ/ „yA    H   â !r       ÆO ‰       h ‰      ( „yA    H   ¢ !rG       È� !r       â!r        â/ !r	       È ‰ €    " xB   ÿàÿ Æ ‰
	 €    b z  r  p`ð â !r       Æ x A   pVp  â !rI       ÈO !r       â !r       ä „yA    H   â !r
	
       Æ/ ‰ @    ( ‰
 @    b !r       â !r
       È/ ‰       ( ‰      b !r       â !r	        â !r       â/ G	  �     € ê !rC       â!r 	D       Æ  x;ª¸¿  A  Ì s        $ !t  €?    Ì s        $  r     A  È  r      A  Ê †y    ä Ay        € ê ‰       b Ey       € æ ‰      ¨ „yA €  H   è „yA €  H   è „yA €  H   è „yA €  H   â !r         Æ/ „yA €  H   â !r       ÆO ‰       h ‰      ( „yA €	  H   ¢ !rH       â� !r        â/ !r	       È ‰ €    ( ‰
	 €    b !r       ÈO !r       â !r       â xB   ÿàÿ â !r
	
       Æ/ ‰ @    " z  r  p`ð Æ ‰
 @    b x A   pVp  â !r       â !r
       È/ ‰       ( ‰      b !r	        â !r       â/ G	  �     € ê !rC       â!r 	D       Æ  x;ª¸¿  A  Ì s        $ !t  €?    Ì s        $  r     A  È  r      A  Ê †y    ä Ay        € ê ‰      b !rF       â xBB   ÿàÿ Æ !r       â z B r  p`ð Æ !r       â x A   pVp  È ‰       ¢ !r       Ê/ ‰      " !r         ÊO ‰       b !r	       Ê ‰
	 €    " !r        Ê/ ‰ €    b !r
	
       Ê ‰
 @    " !r       Ê/ ‰ @    b !r
       Ê ‰      " !r       Ê/ ‰       ¢ !r       â M	        € ì !rCC       â/!r        ÆO  xC;ª¸¿  A  â !rDD       Ê s        $ !t  €?    Ì s        $  rCC     A  È  rCDC     A  Ê †y C   â My        € ê „yA     H   " z B r  p`ð â Ey       € ä „y A    H   b r Aÿ   pVp  Æ „yA    H   ¨ „yA    H   è „y	A    H   ( „yA  
  H   ( „yA €  H   ( „yA €  H   ( „yA €  H   è „yA €
  H   " !rX       â !r  K       Æ/ !r       âO !r         Æ� !r		       â!r        È ‰ 	      " !rJ       Æ ‰      b !r       È !r       â !r 	        â !r       È/ ‰       ( ‰      h „yA €   H   ¢ !r        â !r       È/ ‰ €    ( ‰ €    h „yA €  H   â !rV       âO !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       È� !r       â !r       â !r
	
       È/ ‰       ( ‰
      b !r	        â !r 
       â/ G	  À     € ê !r  C       È x '7=  A  â x *BL?  A  Æ #t   €?   È  r     A  â  x    ?  A  â !r 	D       È s     $   $ !t  €?    È  r     A  È  r      A  Ê †y     ä Ay        € ê ‰       b Ey       € æ ‰      ¨ „yA    H   è „yA    H   ( „yA    H   è „yA  	  H   â !r         Æ/ „yA    H   â !r       ÆO ‰       h ‰      " !rG       â� !rI       â!r        ä/ !r       â !r       Ä ‰ €    " !r       Æ ‰ €    b !r       Æ „yA    H   ¢ !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       âO !r       â xB   ÿàÿ â !r
	
       Æ/ ‰       " z  r  p`ð Æ ‰
      b x A   pVp  â !r	        â !r 
       Ö/ G	  À     € ê !r  C       È x '7=  A  â x *BL?  A  Æ #t   €?   È  r     A  â  x    ?  A  â !r 	D       È s     $   $ !t  €?    È  r     A  È  r      A  Ê †y    ä Ay        € ê ‰       b Ey  Ð    € æ ‰      ¨ „yA €	  H   è „yA €  H   è „yA €  H   â !r         Æ/ „yA €  H   â !r       ÆO ‰       h ‰      ( „yA €  H   ¢ !r        â/ !r       È ‰ €    ( ‰ €    h „yA €  H   â !rH       ÈO !r       â !r       Æ !r       â !r	       ä/ ‰ @    ( ‰
	 @    b !r       â xB   ÿàÿ â !r
	
       Æ/ ‰       " z  r  p`ð Æ ‰
      b x A   pVp  â !r	        â !r 
       Ö/ G	  À     € ê !r  C       È x '7=  A  â x *BL?  A  Æ #t   €?   È  r     A  â  x    ?  A  â !r 	D       È s     $   $ !t  €?    È  r     A  È  r      A  Ê †y    ä Ay        € ê ‰      b !rF       â xBB   ÿàÿ Æ !r       â z B r  p`ð Æ !r       â x A   pVp  È ‰       ¢ !r       Ê/ ‰      " !r         ÊO ‰       b !r       Ê ‰ €    " !r        Ê/ ‰ €    b !r	       Ê ‰
	 @    " !r       Ê/ ‰ @    b !r
	
       Ê ‰
      " !r       Ê/ ‰       ¢ !r
       â M	        € ì !rCC       â/!r        ÆO  xC'7=  A  â xC*BL?  A  â !rDD       ä #tC  €?   â xCC   ?  A  Æ  r     A  Ì s     $   $ !t   €?    È  rCC      A  È  rCDC     A  Ê †y C   â My        € ê x     pRð Ú G‰  P    € ê x     pRð Ú G	  `    € ê „yA     H   " z B r  p`ð â Ey       € ä „y A    H   b r Aÿ   pVp  Æ „yA    H   ¨ „yA    H   è „y	A    H   ( „yA  
  H   ( „yA €  H   ( „yA €  H   ( „yA €  H   è „yA €
  H   " !rX       â !r  K       Æ/ !r       âO !r         Æ� !r		       â!r        È ‰ 	      " !rJ       Æ ‰      b !r       È !r       â !r 	        â !r       È/ ‰       ( ‰      h „yA €   H   ¢ !r        â !r       È/ ‰ €    ( ‰ €    h „yA €  H   â !rV       âO !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       È� !r       â !r       â !r
	
       È/ ‰       ( ‰
      b !r	        â !r 
       â/ G	  À     € ê !r  C       Ê	z  i    � â !r 	D       È  x;ª¸¿  A  â 	z   i    � È 	z   i €  � â s        $ !t  €?    Ì s        $  r     A  È  r      A  Ê †y     ä Ay        € ê ‰       b Ey       € æ ‰      ¨ „yA    H   è „yA    H   ( „yA    H   è „yA  	  H   â !r         Æ/ „yA    H   â !r       ÆO ‰       h ‰      " !rG       â� !rI       â!r        ä/ !r       â !r       Ä ‰ €    " !r       Æ ‰ €    b !r       Æ „yA    H   ¢ !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       âO !r       â xB   ÿàÿ â !r
	
       Æ/ ‰       " z  r  p`ð Æ ‰
      b x A   pVp  â !r	        â !r 
       Ö/ G	  À     € ê !r  C       Ê	z  i    � â !r 	D       È  x;ª¸¿  A  â 	z   i    � È 	z   i €  � â s        $ !t  €?    Ì s        $  r     A  È  r      A  Ê †y    ä Ay        € ê ‰       b Ey  Ð    € æ ‰      ¨ „yA €	  H   è „yA €  H   è „yA €  H   â !r         Æ/ „yA €  H   â !r       ÆO ‰       h ‰      ( „yA €  H   ¢ !r        â/ !r       È ‰ €    ( ‰ €    h „yA €  H   â !rH       ÈO !r       â !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       â� !r       â xB   ÿàÿ â !r
	
       Æ/ ‰       " z  r  p`ð Æ ‰
      b x A   pVp  â !r	        â !r 
       Ö/ G	  À     € ê !r  C       Ê	z  i    � â !r 	D       È  x;ª¸¿  A  â 	z   i    � È 	z   i €  � â s        $ !t  €?    Ì s        $  r     A  È  r      A  Ê †y    ä Ay        € ê ‰      b !rF       â xBB   ÿàÿ Æ !r       â z B r  p`ð Æ !r       â x A   pVp  È ‰       ¢ !r       Ê/ ‰      " !r         ÊO ‰       b !r       Ê ‰ €    " !r        Ê/ ‰ €    b !r	       Ê ‰
	 @    " !r       Ê/ ‰ @    b !r
	
       Ê ‰
      " !r       Ê/ ‰       ¢ !r
       â M	        € ì !rCC       â/!r        ÈO 	zCC i    � â !rDD       È  xC;ª¸¿  A  â 	zDD i    � È 	zDD i €  � â s        $ !t  €?    Ì s        $  rCC     A  È  rCDC     A  Ê †y C   â My        € ê „yA     H   " xB   ÿàÿ äxB   ÿàÿ â„y A    H   b z B r  p`ð ä z  r  p`ò â „yA €   H   ¢ z  r  p`ô Ä r Aÿ   pVp  â„yA €  H   â x A   pVò  äx A   pVt â „y
A    H   " xBB   ÿàÿ Æ „yA    H   " z B r  p`ö Æ „yA    H   " x A   pVö Æ „yA    H   ( „yA €  H   ( „yA €  H   ( „yA    H   " !rX       Æ „yA  	  H   " !r  K       Æ/ „y	A    H   b !rVV       ÆO „yA  
  H   ¢ !rJJ       Æ� „yA €  H   â !r

G       Æ„yA €
  H   " !rI       Æ „yA    H   " !r       Æ „yA    H   " !r         Æ „yA €  H   " !rV       Æ „yA €  H   " !rJ       Æ „yA €  H   " !r

       Æ „yA €	  H   â !r       Æ „y!A €  H   â !r	       Æ/ „y#A €  H   â !r        âO !r       ä� ‰       " !r       Æ‰      b !r

       Æ ‰      ¢ !r       Æ ‰      â !rF       Æ ‰	
      " !rH       Æ ‰      " !r       â !r         â !r       â !r       â/ !r!!       â !r       äO ‰       " !r##       â !r       Ä� ‰!      b !r

	       Æ‰#      ¢ !r       Æ ‰      è ‰	      ( ‰      ( ‰
      " !r        Æ ‰      " !r!       â/ !r#       äO ‰  €    b !r       Æ� ‰      ¢ !r		       Æ‰      â !r       Æ ‰ €    " !r
       Æ ‰	 €    " !r       Æ ‰ €    ( ‰
 €    " !r         Æ/ ‰ €    b !r       âO !r       ä� ‰  @    ¢ !r       Æ‰ €    â !r	       Æ ‰ €    " !r       Æ ‰ @    " !r

       Æ ‰	 @    " !r       Æ/ ‰ @    h ‰
 @    b !r        ÆO ‰ @    ¢ !r       â� !r       ä‰       â !r       Æ ‰ @    " !r		       Æ ‰ @    " !r       Æ/ ‰      b !r
       Æ ‰	      b !r       ÆO ‰      ¨ ‰
      ¢ !r        Æ� ‰      â !r       â !‚ D       â!r       Æ‰      " !r       Æ/ ‰      b !r		       â !‚C       â!r       äO !’	D       â  ‚      A  â !r
       â !’C       Æ †‰     â !r       â� !¢D       â  ’     A  ä !¢	C       â !r       ä †™    â  ¢		     A  â !r       È/ †© 	   â M9        € ê !rDD       â !rCC       È  rCDC     A  Ê †y C   â My        € ê „yA     H   " z B r  p`ð â Ey  À    € ä „y A    H   b r Aÿ   pVp  Æ „yA    H   ¨ „yA    H   è „y	A    H   ( „yA  
  H   ( „yA €  H   ( „yA €  H   ( „yA €  H   è „yA €
  H   " !rX       â !r  K       Æ/ !r       âO !r         Æ� !r		       â!r        È ‰ 	      " !rJ       Æ ‰      b !r       È !r       â !r 	        â !r       È/ ‰       ( ‰      h „yA €   H   ¢ !r        â !r       È/ ‰ €    ( ‰ €    h „yA €  H   â !rV       âO !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       È� !r       â !r       â !r
	
       È/ ‰       ( ‰
      b !r        â !r 
       â/ G	  à     € ê !r  C       Ê	x   à@  � Ê  x #ÛÙ?  A  È  x ;ª¸¿  A  â !r D       Ê s        " 	x    à@  � È 	x    àÀ  � Ê !t   €?    â !t  €?    È s	        $  r	     A  È  r     A  Ê †y     ä Ay        € ê ‰       b Ey       € æ ‰      ¨ „yA    H   è „yA    H   ( „yA    H   è „yA  	  H   â !r         Æ/ „yA    H   â !r       ÆO ‰       h ‰      " !rG       â� !rI       â!r        ä/ !r       â !r       Ä ‰ €    " !r       Æ ‰ €    b !r       Æ „yA    H   ¢ !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       âO !r       â xB   ÿàÿ â !r
	
       Æ/ ‰       " z  r  p`ð Æ ‰
      b x A   pVp  â !r        â !r 
       Ö/ G	  à     € ê !r  C       Ê	x   à@  � Ê  x #ÛÙ?  A  È  x ;ª¸¿  A  â !r D       Ê s        " 	x    à@  � È 	x    àÀ  � Ê !t   €?    â !t  €?    È s	        $  r	     A  È  r     A  Ê †y    ä Ay        € ê ‰       b Ey  ð    € æ ‰      ¨ „yA €  H   è „yA €	  H   ( „yA €  H   è „yA €  H   â !r         Æ/ „yA €  H   â !r       ÆO ‰       h ‰      ( „yA €  H   â !rH       È� !r       â!r        Æ/ !r       â !r       â ‰ €    ( ‰ €    b !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       â xB   ÿàÿ â !r
	
       Æ/ ‰       " z  r  p`ð Æ ‰
      b x A   pVp  â !r        â !r 
       Ö/ G	  à     € ê !r  C       Ê	x   à@  � Ê  x #ÛÙ?  A  È  x ;ª¸¿  A  â !r D       Ê s        " 	x    à@  � È 	x    àÀ  � Ê !t   €?    â !t  €?    È s	        $  r	     A  È  r     A  Ê †y    ä Ay        € ê ‰      b !rF       â xBB   ÿàÿ Æ !r       â z B r  p`ð Æ !r       â x A   pVp  È ‰       ¢ !r       Ê/ ‰      " !r         ÊO ‰       b !r       Ê ‰ €    " !r        Ê/ ‰ €    b !r	       Ê ‰
	 @    " !r       Ê/ ‰ @    b !r
	
       Ê ‰
      " !r       Ê/ ‰       ¢ !r
       â M	        € ì !rCC       â/!r        ÈO 	xCC  à@  � â !rDD       È  xC#ÛÙ?  A  â 	xDD  à@  � Æ  x;ª¸¿  A  â 	xDD  àÀ  � Ê s        " !tDD  €?    â !t  €?    Ì s         $  rCC      A  È  rCCD     A  Ê †y C   â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y       "   $ z   w  p`ð Ú M	        € ê y      !   " ¹z  l      â Ey  0    € â ™x?    â y      %   b ¹z  F   
   â $rÿÿ   ÿ Ž ä $rÿÿ   ÿ Ž â xÿ     È |    p`ð â $x   ÿ Ž Ø/ G	  �    € ê y	      &   " $tÿ   ÿ Ž ä $z  x  	Ž È %v \   Ž Ì �y    ™ ¢  $rÿÿ   ÿ Ž â x   ÿÀŽ â$x   ÿ Ž ä 'z	 m   Ž Æ x   ÿÀŽ â $z
 q  Ž â zÿ n    â $rÿÿ   ÿ Ž ä $rÿÿ   ÿ Ž ä $rÿÿ   
Ž ä $vÿ X  ÿ Ž ä $z o  	Ž â x	   ÿàÿ â $z  r  ÿŽ Ä $z u  ÿŽ ä $vÿ Y  ÿ Ž Æ r   ÿàñ â $x   ÿ Ž ä $z t  Ž äO$z t  
Ž ä $z		 q  Ž ä %xR   Ž È $rÿÿ   ÿ Ž ä %xR   Ž È %x	R   Ž È %x	R   Ž È $r
ÿÿ   ÿ  ä r0   ÿàñ Ê $x1     â r   ÿàñ È �y0B   ‘ ¢  $x     ä $t4ÿ   ÿ Ž â �y>0J   ‘ è  �y   • ( �y   • b r2   ÿàñ Æ �y"   • ä $x3     âr    ÿàñ â �y$   • è �y62B   ‘ â  $x!     Ê �y"    • è �y5    • è �y< $   • è �y22J   ‘ â $r0ÿÿ    Ž Ä $r1ÿÿ    Ž ä $x   ÿ Ž ÊO x$   ÿÀŽ È x/$ÿ   øŽ â$x$@   ÿ Ž â xÿ     Æ %v./ 
 4 Ž â x   $øŽ Æ $x   ÿ Ž ä �y..    › ¢ %v, 
 4 Ž ä x   ÿÀŽ È/ x)ÿ   øŽ â�y,,    › ( %v() 
 4 Ž â xÿ     Æ $x   ÿ Ž â xÿ     â�y((    › b r   à‘ ä x   øŽ ä rÿ
   $~  â $tÿ$   ÿ Ž ä %v* 
 4 Ž È $x$   ÿŽ ä �y "   • b %v Z   Ž Æ �y**    › b $x   Ž Ê �y:   ™ b $x6   ÿ Ž Æ� �y   ™ ä  x   ÿÀŽ ä �y9   ™ ä  x'ÿ   "øŽ ä�y   ™ æ  %v&' 
 4 Ž â �y8   ™ è  �y   ™ è  �y&&    › â $x#6@   ÿ Ž â x"ÿ   "  Ä �y;   ™ ä  x#"   #øŽ ä �y=    ™ æ  %v$# 
 4 Ž â �y?    • ê  �y$$    › â $x"6   ÿ Ž Ê x""   ÿÀŽ È x#"ÿ   5øŽ â$x66   ÿ Ž â x5ÿ   5  Æ %v"# 
 4 Ž â x55   6øŽ Ê �y""    › â %v 5 
 4 Ž Ì/ �y      › â %x4R   0Ž È %x6	R   0Ž ä �y44    • è �y66    • â  x@ÿ   ÿÀŽ Ê $xA @  @Ž â x0ÿ     Æ $x@A€   ÿ Ž ä$xA   ÿ Ž ä $xA0 @  0Ž â x@@€€ÿÀŽ ä x1€€ÿÀŽ â $x0A€   ÿ Ž â x@€€€€ÿüŽ ä x1€€€€ÿüŽ Ä xÿþþþÿàÿ ä x00€€ÿÀŽ ä xBÿþþþÿàÿ ä r@   ÿüŽ ä x3ÿ   ÿÀŽ ä x0€€€€ÿüŽ Ä rB1   ÿüŽ â $x1A   ÿ Ž â xÿþþþÿàÿ â $x7 @  3Ž â x5ÿ     ä/ x1€€ÿÀŽ â $x37€   ÿ Ž â r00   ÿüŽ â $x5 @  5Ž â x1€€€€ÿüŽ â $x57   ÿ Ž â x33€€ÿÀŽ â $xA€   ÿ Ž â x71ÿþþþÿàÿ Ä x15€€ÿÀŽ ä x@3€€€€ÿüŽ ä x5A€€ÿÀŽ ä r7   ÿüŽ ä x71€€€€ÿüŽ â $xC   ÿ Ž â x@@ÿþþþÿàÿ Ä xA5€€€€ÿüŽ ä xB7ÿþþþÿàÿ ä x7˜º  ÿ    ä r@3   ÿüŽ ä x@Aÿþþþÿàÿ ä x3C€€ÿÀŽ Ä r1B1   ÿüŽ ä xB7ÿÀŽ ä r@5   ÿüŽ ä x53€€€€ÿüŽ ä x@˜º  ÿ    ä xD5ÿþþþÿàÿ Ä x5@ÿÀŽ ä x˜º  ÿ    ä r3D3   ÿüŽ ä x7.€€€€7ÞŽ äO x@/€€€€@ÞŽ Æ $x7   B
Ž â x70˜º  ÿ    â $x@@   5
Ž Æ x07ÿÀŽ ä x7,€€€€7ÞŽ äx€€€€.´Ž Æ $x77   0
Ž â x01˜º  ÿ    ä x.@€€€€/´Ž ä x/˜º  ÿ    ä xÿÀŽ ä x-€€€€ÞŽ ä x10ÿÀŽ Ä x0)€€€€0ÞŽ äx@/ÿÀŽ ä x/(€€€€/ÞŽ â $x   
Ž ä $x00   1
Ž ä $x/   @
Ž â x-€€€€-´Ž Ä x/˜º  ÿ    ä x0€€€€)´Ž ä x3˜º  ÿ    ä x0ÿ   ÿÀŽ ä x€€€€(´Ž ä x(/ÿÀŽ Ä x/*€€€€/ÞŽ ä x)ÿÀŽ â $x0 @  0Ž â x+€€€€ÞŽ â $x//   (
Ž ä $x(0€   ÿ Ž ä $x   )
Ž ä &r:   ÿ   Ä $x)0   ÿ Ž â x((€€ÿÀŽ â &r.      Æ� x.)€€ÿÀŽ ä x€€€€+´Ž ä x)(€€€€ÿüŽ ä x,7€€€€,´Ž ä x+.€€€€ÿüŽ ä xÿ     Ä x))ÿþþþÿàÿ â &r,9      â x++ÿþþþÿàÿ â $x @  Ž â r))(   ÿüŽ â &r--      â r++.   ÿüŽ â $x€   ÿ Ž â x))˜º  ÿ    ä x(+˜º  ÿ    â $x+   ÿ Ž â x€€ÿÀŽ â &r8   ÿ   â x,)ÿÀŽ Ä x(ÿÀŽ ä x)&€€€€)ÞŽ ä x('€€€€(ÞŽ â &r      â x++€€ÿÀŽ â $x))   ,
Ž â x€€€€ÿüŽ â $x((   
Ž â x+€€€€ÿüŽ Ä x,ÿþþþÿàÿ ä x))€€€€&´Ž ä x&ÿþþþÿàÿ ä x*/€€€€*´Ž ä r,,   ÿüŽ ä x<ÿ   ÿÀŽ Ä r&&+   ÿüŽ â &r**;      â x,˜º  ÿ    â &r)):   ÿ   â x(€€€€'´Ž â $x< @  Ž â x&&˜º  ÿ    ä x(ÿÀŽ â &r   )   â x$€€€€ÞŽ Ä x<ÿ   <  â $x€   ÿ Ž â x&ÿÀŽ ä x&%€€€€&ÞŽ â $x   (
Ž ä $x)< @  <Ž â x€€ÿÀŽ â $x&&   
Ž ä $x   ÿ Ž â x$€€€€$´Ž â $x)€   ÿ Ž â x€€€€ÿüŽ Ä x€€ÿÀŽ ä x'€€ÿÀŽ ä x%&€€€€%´Ž ä x&ÿþþþÿàÿ â $x))   ÿ Ž â x€€€€ÿüŽ ä x'€€€€ÿüŽ Ä r&&   ÿüŽ ä x(ÿþþþÿàÿ ä x,ÿþþþÿàÿ ä x))€€ÿÀŽ ä x&˜º  ÿ    ä r((   ÿüŽ Ä x)€€€€ÿüŽ ä x&ÿÀŽ ä x"€€€€ÞŽ ä r,,'   ÿüŽ ä x((˜º  ÿ    ä x.ÿþþþÿàÿ â $x   &
Ž â x(ÿÀŽ Ä r..)   ÿüŽ ä x,˜º  ÿ    ä x(#€€€€(ÞŽ ä x€€€€"´Ž ä x..˜º  ÿ    ä x"ÿÀŽ â $x((   
Ž â x €€€€ÞŽ Ä x.ÿÀŽ ä x.!€€€€.ÞŽ â $x   "
Ž â x((€€€€#´Ž â &r8   ÿ   ä $x..   
Ž â x€€€€ ´Ž â &r((      ä &r$9      â x..€€€€!´Ž â &r;   (   â x>ð   ÿÀŽ Ä x2ð   ÿÀŽ â &r%%      â x>>   ÿÀŽ ä x22   ÿÀŽ â &r=   *   ä&r..=      â xÿ     â $r>>-   ÿŽ â xÿ     â $r22%   ÿŽ Ä $x--   Ž ä $x%%   .Ž ä $r   >Ž â r--   ÿ� â $r.   2Ž â r%%   ÿ� ä r   ÿú� ä r   ÿú� Ä x   ÿàÿ ä xÿ     ä xÿ     ä |    p`ð ä r   ÿ� ä r   ÿ� â 0r4ÿ4    A   â xÿ     â 0r?ÿ?    A   â xÿ     â 0r6ÿ6    A   â xH  ÿàó Ä  r4?     A  â Er        â  r??6     A  â Er        â $rÿÿ   Ž  â x    ÿàÿ â #r      ä #r?      â D	       À â Gy  pìÿÿÿÿƒ ê Ay        € ê ‰      " $x   Ž Æ ‰      d z  p  p`ð È x    pDp  â !r       â !r       È/ ‰      ( ‰      b !r       â !r	       È/ ‰ €    ( ‰
	 €    b !r       â !r
	
       È/ ‰ @    ( ‰
 @    b !r       â !r
       È/ ‰      h  ‰      ¢  M	        € ê y      &   â $x   ÿ Ž â !r       â/ $z   s  Ž â !r       âO $t	ÿ   ÿ Ž â r ÿ   p"ð Ä x    p"ò Ö $ÿÿ    Ž ä $ÿÿ    Ž ä $z  v   Ž È� %v  j  	 Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      "   $ z  w  p`ð Ú M	        € ê y	      &   " $tÿ   ÿ Ž â ¹z  F   
   ä y      !   b ¹z  l      ä ™x?    â $z x  	Ž È %v \   Ž Ê �y     ™ b Ey  @(    € â $rÿÿ   ÿ Ž â x      ÿ  â $r
ÿÿ   ÿ Ž â x      ÿ  â xÿ     È/ |    p`ð Ú G	  À'    € ê y      %   " $rÿÿ   ÿ Ž â x   ÿÀŽ â$x   ÿ Ž â Ey P'    € â 'z	 m   Ž Æ x   ÿÀŽ â $z r  ÿŽ â zÿ n    â $rÿÿ   ÿ Ž ä $rÿÿ   ÿ Ž ä $rÿÿ   
Ž ä $v
ÿ X  ÿ Ž ä $z o  	Ž Ä $z	  t  Ž ä$z u  ÿŽ ä $x   ÿ Ž ä $vÿ `  ÿ Ž â r   ÿàñ â $vÿ Y  ÿ Ž â x   ÿàÿ â $vÿ a  ÿ Ž Ä $x   ÿ Ž ä $z q  	Ž ä $z	 q  Ž ä $rÿÿ   ÿ Ž ä $z	  t  	Ž ä %xR   Ž È %x	R   Ž È %xR   Ž È %x	R   Ž È $rÿÿ   ÿ  ä r
   ÿàñ Ê $x     â r
   ÿàñ È �y!B   ‘ ¢  $x     Ê �y   • è �y    • " r"
   ÿàñ Æ �y<"   • d $x#     âr2
   ÿàñ â �y>$   • ( �y4"B   ‘ " $x3     Æ �yJ   ‘ ( �y(2   • h �y52   • b r   à‘ Æ/ �y=2"   • h �y?2$   • h �y""J   ‘ b %x6R   
Ž Ì �y66    • b $x!   ÿ Ž ÊO x$   ÿÀŽ â $tÿ   ÿ Ž ä $x&!@   ÿ Ž â x%$ÿ   øŽ ä�xÿ     Æ %v$% 
  Ž â x   &øŽ Æ $x&!   ÿ Ž ä �y$$    › ¦ x&   ÿÀŽ â %v& 
  Ž Æ xÿ    øŽ Æ�y&&    › ä %v. 
  Ž â x ÿ      Æ $x!!   ÿ Ž ä �y..    › & x1    !øŽ ä xÿ     Æ %v01 
  Ž â rÿ   $~  Æ $tÿ$   ÿ Ž ä �y00    › " $x$   ÿŽ ä %v  Z   Ž È $x!!   Ž ä $x)4   ÿ Ž Æ �y    ™ ( �y    ™ " x))   ÿÀŽ Æ �y    ™ " x))ÿ   (øŽ Æ�y8    ™ $ %v,) 
  Ž ä �y9    ™ ( �y:    ™ ( �y,,    › " $x)4@   ÿ Ž â x(ÿ   (  Ä �y;    ™ $ x)(   )øŽ ä �y@     ™ & %v*) 
  Ž â �yA     • * �y**    › " $x(4   ÿ Ž Ê x((   ÿÀŽ â $x44   ÿ Ž Æ x)(ÿ   5øŽ äx5ÿ   5  Æ %v() 
  Ž â x55   4øŽ Ê �y((    › " %v25 
  Ž Ì/ �y22    › " %x4	R   
Ž Ì �y44    • " xC<ÿ   ÿÀŽ Ê $xC< @  CŽ â x<ÿ   <  Æ $xBC€   ÿ Ž ä$xC   ÿ Ž ä $x#< @  <Ž â xBB€€ÿÀŽ ä x€€ÿÀŽ â $x!#€   ÿ Ž âx5B€€€€ÿüŽ ä/ x €€€€ÿüŽ Ä x55ÿþþþÿàÿ ä x7!€€ÿÀŽ ä x< ÿþþþÿàÿ ä r 5B   ÿüŽ â $x##   ÿ Ž â x!7€€€€ÿüŽ ä x5>ÿ   ÿÀŽ Ä r<   ÿüŽ ä x<!ÿþþþÿàÿ â $x!> @  5Ž â x##€€ÿÀŽ ä xDÿ   >  â $x>!€   ÿ Ž â x5#€€€€ÿüŽ â $x!!   ÿ Ž â r7<7   ÿüŽ Ä x<5ÿþþþÿàÿ â $xDD @  DŽ â x5!€€ÿÀŽ ä r#<#   ÿüŽ â $xBD€   ÿ Ž â x<5€€€€ÿüŽ ä x>>€€ÿÀŽ ä x<<ÿþþþÿàÿ Ä xC>€€€€ÿüŽ ä x!B€€ÿÀŽ ä r<<5   ÿüŽ ä xCCÿþþþÿàÿ ä x5 ˜º  ÿ    ä xB!€€€€ÿüŽ Ä x ˜º  ÿ    â $xDD   ÿ Ž â r>C>   ÿüŽ ä xCBÿþþþÿàÿ ä x ÿÀŽ ä xBD€€ÿÀŽ ä r!C!   ÿüŽ Ä xC5ÿÀŽ ä xEB€€€€ÿüŽ ä x77˜º  ÿ    ä x!!˜º  ÿ    ä x %€€€€ ÞŽ äO xD$€€€€5ÞŽ Æ $x     
Ž â x5Eÿþþþÿàÿ â $xCD   C
Ž â xD7ÿÀŽ ä x&€€€€7ÞŽ ä� x% €€€€%´Ž ä x #˜º  ÿ    ä x7>˜º  ÿ    Ä x><˜º  ÿ    â $x   D
Ž â rB5B   ÿüŽ ä x< ÿÀŽ ä x5'€€€€ ÞŽ ä x#7ÿÀŽ ä xD.€€€€7ÞŽ Äx >ÿÀŽ ä x7/€€€€>ÞŽ â $x<5   <
Ž È $x 7    
Ž â x'<€€€€'´Ž ä x<=ÿ   ÿÀŽ ä x/ €€€€/´Ž ä xBB˜º  ÿ    ä x !ÿÀŽ ä x!0€€€€!ÞŽ â $x<= @  <Ž â x&€€€€&´Ž Ä xBÿÀŽ â $x!!    
Ž â xB1€€€€BÞŽ â $x <€   ÿ Ž â x$C€€€€$´Ž Æ $xBB   
Ž â x  €€ÿÀŽ â &r$$   ÿ   ä $x<   ÿ Ž â x0!€€€€0´Ž â $x#D   #
Ž â x! €€€€ÿüŽ â &r$%   $   â x€€ÿÀŽ ä x.#€€€€.´Ž â &r$&   $   â x#€€€€ÿüŽ Ä x=ÿ   =  ä x!!ÿþþþÿàÿ â &r''8   $   â x$#ÿþþþÿàÿ â &r..9   ÿ   â r!!    ÿüŽ â $x== @  =Ž â r$$   ÿüŽ â &r//:   .   â x!!˜º  ÿ    â $x=€   ÿ Ž Ä $x.=   ÿ Ž â x !ÿÀŽ ä x!,€€€€!ÞŽ ä x$$˜º  ÿ    ä x%€€ÿÀŽ ä x.€€ÿÀŽ â $x!!    
Ž â x#$ÿÀŽ Ä x$-€€€€$ÞŽ ä x €€€€ÿüŽ ä x&%€€€€ÿüŽ â $x$$   #
Ž â x. ÿþþþÿàÿ ä x&&ÿþþþÿàÿ ä r..   ÿüŽ Ä r&&%   ÿüŽ ä x-$€€€€-´Ž ä x$?ÿ   ÿÀŽ ä x..˜º  ÿ    ä x,!€€€€,´Ž ä x&˜º  ÿ    â $x%? @  $Ž â x!.ÿÀŽ â &r ,   ÿ   â x$ÿÀŽ Ä x.+€€€€.ÞŽ ä x*€€€€ÞŽ ä x?ÿ   ?  â $x#%€   ÿ Ž ä &r -       ä $x..   !
Ž ä $x   $
Ž Ä $x-? @  ?Ž ä $x!%   ÿ Ž â x##€€ÿÀŽ â $x$-€   ÿ Ž â x*€€€€*´Ž ä x!!€€ÿÀŽ ä x#€€€€ÿüŽ â $x--   ÿ Ž â x,$€€ÿÀŽ Ä x%!€€€€ÿüŽ ä x$ÿþþþÿàÿ ä x,€€€€ÿüŽ ä x&%ÿþþþÿàÿ ä r$$#   ÿüŽ ä x--€€ÿÀŽ Ä x#ÿþþþÿàÿ ä r&&!   ÿüŽ ä x$˜º  ÿ    ä x!-€€€€ÿüŽ ä r##,   ÿüŽ ä x&&˜º  ÿ    Ä x$ÿÀŽ ä x,!ÿþþþÿàÿ ä x(€€€€ÞŽ ä x!&ÿÀŽ ä r,,-   ÿüŽ ä x##˜º  ÿ    â $x   $
Ž â x&)€€€€&ÞŽ Ä x,,˜º  ÿ    ä x$#ÿÀŽ â $x&&   !
Ž â x#2€€€€#ÞŽ ä x(€€€€(´Ž ä x,ÿÀŽ â $x##   $
Ž â x,3€€€€,ÞŽ â &r((9   ÿ   â x)&€€€€)´Ž Ä x##€€€€2´Ž â $x,,   
Ž ä &r():   (   â x+.€€€€+´Ž â &r**       â x1B€€€€1´Ž â &r/0;   /   âx3,€€€€3´Ž â &r##;   (   â xð   ÿÀŽ Ä x!"ð   ÿÀŽ â &r++8   *   â x   ÿÀŽ ä x""   ÿÀŽ â &r 1@   /   ä&r$3@   #   â xÿ     â $r'   ÿŽ â x!ÿ   !  â $r""+   ÿŽ Ä $x''    Ž ä $x++   $Ž ä $r    Ž â r'''   ÿ� â $r!!$   "Ž â r ++   ÿ� ä z ÿ `  pPð ä r'   ÿú� Ä r  !   ÿú� ä z ÿ a   Sð ä xÿ     ä x!ÿ      ä r   ÿ� ä r!!    ÿ� â 0r*ÿ4    A   â xÿ     â 0rAÿA    A   â x!ÿ   !  â 0r6ÿ6    A   â Er,        Ä  r**A     A  â Er+ !       â  r-A6     A  â G‰  à    € ê r    ÿàó Ê $x!   Ž  â r(   ÿàó È �y' B    ¢  $x)   Ž  Ê �y(    è �y&(    ( �y.("    h �y/($    ( �y  J    " %x(	R   Ž Ì/ �y((     " $x'   ÿ Ž ÊO x   ÿÀŽ È xÿ   øŽ Ê� %v$ 
  Ž â xÿ     Æ $x'@   ÿ Ž ä �y$$    › ¦ x   øŽ Ê %v 
  Ž Ì �y    › â $x"'   ÿ Ž Ê x#"   ÿÀŽ â $x''   ÿ Ž Æ x##ÿ   &øŽ äx&ÿ   &  Æ %v"# 
  Ž â x'&   'øŽ Ê �y""    › " %v&' 
  Ž Ì �y&&    › " x1.ÿ   ÿÀŽ Ê$x1. @  1Ž È $x01€   ÿ Ž ä$x11   ÿ Ž â x.ÿ   .  ä x00€€ÿÀŽ ä x11€€ÿÀŽ ä x!0€€€€ÿüŽ â $x3. @  .Ž â x.1€€€€ÿüŽ Ä x!!ÿþþþÿàÿ â $x)3€   ÿ Ž â/ x..ÿþþþÿàÿ ä r0!0   ÿüŽ ä x2)€€ÿÀŽ â $x33   ÿ Ž â r..1   ÿüŽ ä x)0˜º  ÿ    Ä x12€€€€ÿüŽ ä x..˜º  ÿ    ä x0)ÿÀŽ ä x!3€€ÿÀŽ ä x31ÿþþþÿàÿ ä x1.ÿÀŽ Ä x4!€€€€ÿüŽ ä r332   ÿüŽ ä x44ÿþþþÿàÿ ä x33˜º  ÿ    ä r44!   ÿüŽ È x44˜º  ÿ    â 0r(ÿ(    A   Ê  r(A(     A  â x)$€€€€)ÞŽ äO x.%€€€€.ÞŽ Æ $x))   0
Ž ä $x..   1
Ž â x0/ÿ   ÿÀŽ ä x$)€€€€$´Ž ä x%.€€€€%´Ž ä x)ÿ   /  â &r$$   ÿ   Ä $x// @  0Ž ä &r$%   $   â x.3ÿÀŽ â $x%/€   ÿ Ž â x!€€€€3ÞŽ â� $x5) @  )Ž ä $x//   ÿ Ž â x)%€€ÿÀŽ â $x!!   .
Ž â x%4ÿÀŽ â $x05€   ÿ Ž â x.€€€€4ÞŽ Ä x1/€€ÿÀŽ ä x/)€€€€ÿüŽ â $x..   %
Ž â x30€€ÿÀŽ â $x55   ÿ Ž â x%1€€€€ÿüŽ ä x0/ÿþþþÿàÿ ä x/3€€€€ÿüŽ Ä x2%ÿþþþÿàÿ ä r00)   ÿüŽ ä x55€€ÿÀŽ ä x4/ÿþþþÿàÿ ä r221   ÿüŽ ä x%0˜º  ÿ    Ä x)5€€€€ÿüŽ ä r443   ÿüŽ ä x22˜º  ÿ    ä x0%ÿÀŽ ä x6)ÿþþþÿàÿ ä x%"€€€€%ÞŽ Äx)2ÿÀŽ ä r665   ÿüŽ ä x/4˜º  ÿ    â $x%%   0
Ž â x2#€€€€2ÞŽ ä x66˜º  ÿ    ä x0/ÿÀŽ â $x22   )
Ž â x/&€€€€/ÞŽ Ä x"%€€€€"´Ž ä x%6ÿÀŽ â $x//   0
Ž â x6'€€€€6ÞŽ â &r""9   ÿ   â x#2€€€€#´Ž ä x!!€€€€´Ž â $x66   %
Ž â x//€€€€&´Ž â &r"#:   "   â x.€€€€´Ž â &r!!   $   â x'6€€€€'´Ž â &r"/;   "   â x    ÿÀŽ â &r8   !   Ä &r"'@   "   â x ÿ      â $r!   ÿŽ ä $x   "Ž ä $r  "   !Ž Æ r   ÿ� È r    ÿú� È xÿ     È r   ÿ� È xÿ     È Er        Ê #r(      â #r*,      â #r-+      â G‰  à    € ê r$   ÿàñ Ê $x%     â r&   ÿàñ È �y)$B    ¢  $x'     Ê �y &    è �y(&    ( �y,&"    h �y-&$    ( �y$$J    " %x*R   Ž Ì �y**     " $x)   ÿ Ž ÊO x   ÿÀŽ È xÿ    øŽ Ê� %v 
  Ž â x ÿ      Æ $x!)@   ÿ Ž ä �y    › ¦ x!    !øŽ Ê %v ! 
  Ž Ì �y      › â $x")   ÿ Ž Ê x#"   ÿÀŽ È x##ÿ   (øŽ â$x))   ÿ Ž â x(ÿ   (  Æ %v"# 
  Ž â x)(   )øŽ Ê �y""    › " %v() 
  Ž Ê �y&(    › "# x/,ÿ   ÿÀŽ Ê$x/, @  /Ž È $x/€   ÿ Ž ä$x./   ÿ Ž Æ x€€ÿÀŽ ä x,ÿ   ,  ä x%€€€€ÿüŽ ä x..€€ÿÀŽ â $x,, @  ,Ž â x%%ÿþþþÿàÿ ä x(.€€€€ÿüŽ â/ $x+,€   ÿ Ž â r%%   ÿüŽ Ä x)(ÿþþþÿàÿ â $x,,   ÿ Ž â x%%˜º  ÿ    ä x++€€ÿÀŽ ä r)).   ÿüŽ ä x/,€€ÿÀŽ ä x(%ÿÀŽ Ä x,+€€€€ÿüŽ ä x)˜º  ÿ    ä x,,ÿþþþÿàÿ ä x)ÿÀŽ ä r,,+   ÿüŽ ä x./€€€€ÿüŽ È x..ÿþþþÿàÿ È r../   ÿüŽ È x..˜º  ÿ    â 0r*ÿ*    A   Ê  rAA*     A  â x%€€€€%ÞŽ äO x€€€€ÞŽ Æ $x%%   (
Ž ä $x   )
Ž Æ x%€€€€´Ž ä x%,˜º  ÿ    ä x(-ÿ   ÿÀŽ â &r   ÿ   â x€€€€´Ž ä x%ÿÀŽ ä x €€€€%ÞŽ â� $x(- @  (Ž â x%ÿ   -  Æ $x   
Ž ä $x(€   ÿ Ž ä$x,% @  %Ž ä $x%(   ÿ Ž ä &r      â x€€ÿÀŽ â $x),€   ÿ Ž â x.ÿÀŽ Ä x!€€€€.ÞŽ ä x(%€€ÿÀŽ ä x%€€€€ÿüŽ â $x   
Ž â x))€€ÿÀŽ â $x,,   ÿ Ž â x(€€€€ÿüŽ ä x%%ÿþþþÿàÿ Ä x+)€€€€ÿüŽ ä xÿþþþÿàÿ ä r%%   ÿüŽ ä x,,€€ÿÀŽ ä x+ÿþþþÿàÿ ä r(   ÿüŽ Ä x%%˜º  ÿ    ä x+,€€€€ÿüŽ ä r))   ÿüŽ ä x˜º  ÿ    ä x(%ÿÀŽ ä x++ÿþþþÿàÿ Ä x%"€€€€%ÞŽ äxÿÀŽ ä r++,   ÿüŽ ä x))˜º  ÿ    â $x%%   (
Ž â x#€€€€ÞŽ ä x(+˜º  ÿ    Ä x,)ÿÀŽ â $x   
Ž â x)&€€€€)ÞŽ ä x"%€€€€"´Ž ä x(ÿÀŽ â $x))   ,
Ž â x('€€€€(ÞŽ â &r""9   ÿ   â x#€€€€#´Ž Ä x€€€€ ´Ž â $x((   
Ž â x))€€€€&´Ž â &r"#:   "   â x!€€€€!´Ž â &r      â x'(€€€€'´Ž â &r");   "   â x$   ÿÀŽ â &r!8      ä &r"'@   "   â x$ÿ   $  â $r   ÿŽ Ä $x   "Ž ä $r$"   Ž Æ r   ÿ� È r   ÿú� È xÿ     È r   ÿ� È xÿ     È Er        Ê #rA      â x   ÿàÿ ä xH  ÿàó ä |    p`ð ä x

H  ÿàõ â $rÿÿ   Ž  â x    ÿàÿ Æ $rÿÿ    Ì D	       À â Gy  pÚÿÿÿÿƒ ê Ay       € ê $rÿÿ    Ž ä $r
ÿÿ    Ž ä Ay        € ê z ÿ `  pPð È z ÿ a   Sð Ú G	  P    € ê ‰      ( ‰      b !r       â !r       È/ ‰      ( ‰      b !r       â !r       È/ ‰ €    ( ‰	 €    b !r       â !r		       È/ ‰ @    ( ‰	 @    b !r       â !r	       È/ ‰      ( ‰      b !r       â !r       â/ Gy  €    € ê ‰      ( ‰      h ‰      ¨ ‰	      â !r       â !r       â/ !r
       ÈO ‰      " !r	       Æ� ‰	      h ‰
      ¨ ‰      â !r       â !r		       È/ ‰ €    " !r
       ÆO ‰	 €    b !r       Æ� ‰ €    ¨ ‰ €    â !r       â !r	       â/ !r       ÈO ‰ @    " !r
       Æ� ‰ @    h ‰ @    ¨ ‰
 @    â !r       â !r       â/ !r       ÈO ‰      â !r	
       Æ� ‰      ( ‰      h ‰	      ¢ !r       â !r       â/ !r       â !r
	       âO y      !   ( y      %   $ $x   Ž Ê z  p  p`ð È x    pDp  Ú M	        € ê z ÿ ^  pPð â $z  v  Ž Æz ÿ _   Sð Ú $ÿ   ÿ Ž È % ^   Ž Ì �	     ¢ $x    ÿ Ž Ê r  ÿ   p"ò äx     p"ö äx     p"ø äx     p"ô Î $ ÿÿ    Ž â z ÿ `  pPò â $2 ÿÿ    Ž ä $B ÿÿ    Ž â z ÿ a  Sò â $" ÿÿ   
 Ž È !         ÐO G™       € ê z ÿ b  pPð È z ÿ c   Sð Ú $ÿ   ÿ Ž È % b   Ž Ì �	     ¢ $x   ÿ Ž ä $vÿ h  ÿ Ž Æ x øÿÿÿp"ò äx üÿÿÿp"ô är ÿ   p"ö äx    p"ø Î $ÿÿ    Ž â x    pBò â $"ÿÿ    Ž ä $2ÿÿ    Ž ä $Bÿÿ   
 Ž È !       ÈO G  P    € ê x    pRð Ú G‰  �     € ê x    pRð Ú G	  ð    € ê  x;ª¸¿  A  Ì s        $ !t  €?    Ì s        $  r     A  È  r       A  â Gy  P    € ê  x'7=  A  â x*BL?  A  Æ #t  €?   â  x   ?  A  Æ  r     A  Ì s     $   $ !t  €?    È  r     A  È  r       A  â Gy  °    € ê x    pRð Ú G‰  à     € ê x    pRð Ú G	        € ê 	z i    � ä 	z   i    � Æ  x;ª¸¿  A  â 	z   i €  � Ê s        $ !t  €?    Ì s        $  r     A  È  r       A  â Gy  Ð     € ê  r       A  â Gy  °     € ê 	x  à@  � ä 	x    à@  � Æ  x#ÛÙ?  A  â 	x    àÀ  � Æ  x;ª¸¿  A  ä !t   €?    È s        $ !t  €?    Ì s        $  r      A  È  r       A  ä y      "   " $tÿ   ÿ Ž Æ y      &   b $z s  Ž È $z v  Ž È/ %v j   Ž Ê †y      â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y7      "   " ¹z  l      â Ey  �/    € â ™x?    â y8      !   b ¹z  F   
   â r; ÿ       â $r9ÿÿ   ÿ Ž â r< ÿ       â $rHÿÿ   ÿ Ž â r  ÿ       â x,      ÿ  â x.      ÿ  â x0      ÿ  â x2      ÿ  â x4      ÿ  â $r6ÿÿ   ÿ Ž â r 7ÿ   pRð Ä r 7ÿ   pRò â$x7    8Ž Ê/ xÿ     È r   ÿ� È x:ÿ     È | :   pbô Ú G)   .    € ê y      '   " r ÿ       â $vCÿ X  ÿ Ž â r ÿ       â y      &   b xøÿÿÿÿÀŽ â $vAÿ Y  ÿ Ž â xG:   ÿ   â y	      %   ¢ r; ÿ       ä zFG q  ÿàÿ â 'z y   Ž È 'z s   Ž â/ zÿ z    Æ $x		   ÿ Ž âO zÿ t    â $z? }  ÿŽ ä $z	 p  :Ž âx	   ÿàÿ â $z |  :Ž ä$z |  Ž ä $z v  Ž Ä $x   
Ž ä $z p  Ž ä $z
 w  ÿŽ â xÿ     â$z v  Ž â z  q      â $x   ÿ Ž â r?
?   ÿàõ â %xJ   Ž â rBG   ÿŽ Ä rIÿÿ   ÿä â %xJ   ÿŽ È %x
J   Ž È %xJ   ÿŽ È $x=   GŽ ä$x@   GŽ ä$x>   GŽ ä$xD   GŽ ä$xE   GŽ ä %x   Ž È %x   Ž È rC   ÿàõ Ê $x   A Ê �y„   • ¨ �y}   • è �yƒ   • ( �yŠ   • b	 rC   ÿàõ È rA   ÿä Ê �y_   • h �ya   • h �y`   • h �y^   • b rd?   GàÝ Ä x+ $       ä rI   ÿ¤~ Ê $x'$   ÿŽ ä %vdd Z  + Ž È $xee   'Ž Ê �yMd   ™ h �yKd   ™ h �yNd   ™ h �yJd   ™ h �y]d   ™ h �y\d   ™ h �yLd   ™ b r$
C   ÿàû Æ �yQd    ™ b r%A   ÿäÿ Ê �yZ$B   ‘ h �ydd    • b $x„   ÿ Ž âO x}   ÿ   È� xø  ÿÀŽ ä xø  ÿÀŽ ä z  ÿàõ Ê $vÿ 	 ÿ â xƒ   ÿ   äz  ÿàõ Æ �y    › ¢ xø  ÿÀŽ â $vÿ 	 ÿ Æ z  ÿàõ â $xŠ   ÿ Ž ä�y    › â zÿ 	 ÿä ä xø  ÿÀŽ È �y    › " z  ÿàõ È zÿ 	 ÿä Ì �y    › b $x_   ÿ Ž Ê x ø  ÿÀŽ ä xa   ÿ   ä z    ÿàõ ä xø  ÿÀŽ Æ $v!ÿ 	 ÿ â z  ÿàõ ä x`   ÿ   Æ $vÿ 	 ÿ â �y      › ¢ xø  ÿÀŽ â $x^   ÿ Ž â rC   ÿà÷ ä �y    › ¢ z  ÿàù ä xø  ÿÀŽ â $x   AŽ â zÿ 	 ÿä Ä z  ÿà÷ ä �y[B   ‘ ¢  r"C   ÿàõ ä zÿ 	 ÿäÿ â �y    › ª �y    › ¢ rb?   FàÇ â $x#	   A â r&C   ÿàõ Ä r(I   ÿ„þ ä r?   =àë â �yO"    • ¢  r'A   ÿä â %vbb Z  + Ž â rI   ÿÄþ ä r$?   @àµ â/ $x%($   ÿŽ â �yP&    • ¢ %v" Z  + Ž â r?   BàÙ Ä rYI   ÿd~ â $xcc   %Ž â r)?   >àµ ärXI   ÿ¤~ är(?   DàË â �ySb   ™ ¢  rVI   ÿd~ Æ �yWb   ™ ¢  r*?   Eàµ â $xiX$   ÿŽ â rRI   ÿ„þ â�yUb   ™ ¢  rTI   ÿd~ Æ �yXb   ™ â  $xgY$   ÿŽ ä $xmR$   ÿŽ â �yYb   ™ â  $xkV$   ÿŽ Æ �yRb   ™ â  $xoT$   ÿŽ Æ �yTb   ™ è  �yVb    ™ â  $x$   ÿŽ Ê r##   ÿàÿ â %v$$ Z  + Ž â�ybb    • & %v&) Z  + Ž â/�yj"   ™ & %v(( Z  + Ž â �yh"   ™ & $x%%   gŽ Ä �yg"   ™ b %v Z  + Ž Æ �yl"   ™ b $x''   kŽ â r))m   ÿàÿ ä �ym"   ™ b ri   ÿàÿ Æ �yk"   ™ h �yf"   ™ h �yi"    ™ b %v** Z  + Ž Æ �yq$   ™ b $x++   oŽ Æ �yo$   ™ h �ye$   ™ h �ys$   ™ h �yt$   ™ h �yr$   ™ b x‹ÿ	   }  Æ �yc$   ™ h �yp$    ™ b 	sŒ ‹        & �yn"    • b x•ÿ	   ƒ  Æ �yv   ™ b 	s� •        " x’ÿ	   „  ä �yy   ™ h �y#   ™ b# xŒŒ   ÿ   â 	s� ’        $ �yz   ™ b x�‹€   ŒøŽ Æ �y{   ™ b x—ÿ	   Š  â $x”�ÿŽ ä �y"   ™ b x“  @€    Æ �yw   ™ b $x–�€   ÿ Ž Æ �yx    ™ b 	s˜ —        b x“”€€€€“ìŽ â $t‘ÿÿ Ž â x–•€   –øŽ â �yu$    • b x•“ÿþþþÿàÿ ä x“�   ÿ   â �y|    • b x‘”€€€€‘ìŽ Ä x“’€   “øŽ â �y&   ™ b x‘‘ÿþþþÿàÿ Æ �y%&   ™ b# $x––ÿŽ Æ �y‚&   ™ h �y€&   ™ h �y$&   ™ h �y}&   ™ h �y�&   ™ h �y~&    ™ h �y&    • b $x““ÿŽ Æ �yƒ(   ™ b $tšÿ @€ÿ Ž Æ �y‰(   ™ h �y…(   ™ h �y‡(   ™ h �y„(   ™ h �yˆ(   ™ h �y†(   ™ h �y(    ™ h �y&(    • h! �y'*   ™ h �yŽ*   ™ h �yŒ*   ™ h �y‹*   ™ h �yŠ*   ™ h �y�*   ™ h �y)*   ™ h �y(*    ™ h �y�*    • b $x˜˜€   ÿ Ž â x‘‘”øŽ Ä x�• @€”øŽ â $t+ÿ @€ÿ Ž â/ x*     ä x•     ä x”–€€€€+ìŽ ä x*“€€€€*ìŽ äx+“€€€€šìŽ Ä x˜—€   ˜øŽ ä x’–€€€€•ìŽ ä x—”ÿþþþÿàÿ ä x**ÿþþþÿàÿ ä x”+ÿþþþÿàÿ ä x•’ÿþþþÿàÿ Ä x**“øŽ äx”” @€“øŽ â $x˜˜ÿŽ â x’— @€–øŽ â$t—ÿ @€ÿ Ž â x+•–øŽ ä x•     ä x“*˜º  ÿ    Ä x””˜º  ÿ    ä x–˜€€€€—ìŽ äx•˜€€€€•ìŽ ä x*“ÿÀŽ äx—”ÿÀŽ ä x““€€€€ÞŽ ÄO x””€€€€ÞŽ ä x™•ÿþþþÿàÿ ä r•“*  €ÿàÿ â $x*”   —
Ž â x›–ÿþþþÿàÿ ä x–�˜º  ÿ    ä x‘‘˜º  ÿ    Ä x**€€€€´Ž ä xaÿ	   a  ä x–ÿÀŽ äx––€€€€ÞŽ ä� x•€€€€´Ž ä x�‘ÿÀŽ â	s• a        " x‘‘€€€€ÞŽ Ä x’’˜º  ÿ    ä x++˜º  ÿ    â $x––   
Ž â r‘‘�  €ÿàÿ ä x’ÿÀŽ äx�+ÿÀŽ ä x’’€€€€ÞŽ Äx++€€€€ÞŽ ä x”™˜øŽ â $x’’   
Ž â r++�  €ÿàÿ ä x`ÿ	   `  ä x+€€€€´Ž ä x+”˜º  ÿ    Ä x’€€€€´Ž ä x’•   ÿ   â 	s `        " x“› @€˜øŽ ä x‘€€€€´Ž ä x‘+ÿÀŽ äx�+€€€€ÞŽ Äxaa€   ’øŽ ä x_ÿ	   _  ä x’“˜º  ÿ    â $x‘�   ‘
Ž â x^ÿ	   ^  â $x�aÿŽ â 	s+ _        b xa’ÿÀŽ â$t•ÿ @€ÿ Ž â x’’€€€€ÞŽ Ä x“     Æ 	s” ^        ¢ r’’a  €ÿàÿ ä xa�€€€€“ìŽ äx“�€€€€•ìŽ â $x•€   ÿ Ž â xaaÿþþþÿàÿ ä x““ÿþþþÿàÿ ä x•`€   •øŽ Ä x‘€€€€´Ž ä xa�øŽ äx`“ @€�øŽ ä x�+   ÿ   â/ $x••ÿŽ â x–€€€€´Ž â $xa”€   ÿ Ž âO x–  @€    Ä x’€€€€´Ž â $t’ÿÿ Ž â x�_€   �øŽ ä x_•€€€€–ìŽ â$t—ÿ @€ÿ Ž â xa^€   aøŽ â $x��ÿŽ â x+•€€€€’ìŽ ä x’_ÿþþþÿàÿ â $x“aÿŽ â x‘     Ä x^+ÿþþþÿàÿ ä xa�€€€€—ìŽ â $t”ÿ @€ÿ Ž â x+’ @€•øŽ ä x_�€€€€‘ìŽ ä x’     ä xaaÿþþþÿàÿ Ä x__ÿþþþÿàÿ ä x‘“€€€€’ìŽ äx’“€€€€”ìŽ ä xaa @€�øŽ äx__�øŽ ä x�‘ÿþþþÿàÿ Ä x”’ÿþþþÿàÿ ä x’a˜º  ÿ    ä x‘_˜º  ÿ    ä x^^•øŽ ä xa�“øŽ äx_” @€“øŽ Ä x“’ÿÀŽ ä x•˜º  ÿ    ä x’’€€€€!ÞŽ ä x”`˜º  ÿ    ä x`•ÿÀŽ ä r’“  €ÿàÿ Ä x••€€€€ÞŽ ä x€€€€!´Ž ä x!^˜º  ÿ    â $x••   `
Ž â x`+˜º  ÿ    ä x�‘ÿÀŽ äx‘‘€€€€ ÞŽ Ä x^!ÿÀŽ äx+`ÿÀŽ äx!!€€€€ÞŽ ä x``€€€€ÞŽ ä xaa˜º  ÿ    ä r‘‘�  €ÿàÿ Ä x�_˜º  ÿ    ä r!!^  €ÿàÿ â $x``   +
Ž â x^aÿÀŽ äxaa€€€€ÞŽ ä x+�ÿÀŽ äx��€€€€ÞŽ Ä x ‘€€€€ ´Ž ä x!€€€€´Ž ä r!a^  €ÿàÿ ä r��+  €ÿàÿ â &r+M   ÿ   âx!!€€€€´Ž â &rM M   ÿ   ä &rK   ÿ   Ä &r+*N   +   ä&rNN   M   â x•€€€€´Ž â &rMJ      ä &r+]   +   âx`€€€€´Ž â &r^\   M   â x[ð   ÿÀŽ â&r]]   N   â x[[   ÿÀŽ â &rKK   ÿ   Ä &rNL   +   ä &r+Q   ^   â x“”ÿÀŽ â&rKJ   K   â x””€€€€ÞŽ â $rJ[N   ÿŽ â xÿ     â $xNN   +Ž â r””“  €ÿàÿ Æ $rJ+   JŽ â r+NN   ÿ� â &r\!\   K   â x”€€€€´Ž ä x�€€€€´Ž ä xKZð   ÿÀŽ ärJ+J   ÿú� â &r]L   ]   â xZZ   ÿÀŽ â &r\Q   \   â x+ÿ   K  Ä xKÿ   J  â $rLZ]   ÿŽ â r]]\   ÿàÿ â &rMW   ÿ   â rKKJ   ÿ� â &rJS   ÿ   â r]]]   ÿ� â $rL+\   LŽ ä &rS S   ÿ   ä &rJ*U   J   Ä &rMX   M   ä &rWW   ÿ   â rL]L   ÿú� â &rSU   S   ä &rJY   J   ä&rNR   M   ä &rWX   W   â xMÿ   L  â &rSY   S   Ä &rJT   J   ä &rQV   N   ä &rW!R   W   â rMML   ÿ� â &rST   S   â rNJQ   ÿàÿ â $rL[J   ÿŽ ä &rVV   W   â xKÿ   K  â $rJZS   ÿŽ Ä 0rOÿO    A   ä 0rdÿd    A   ä $rLQ   LŽ â rQNN   ÿ� â $xSS   VŽ â ErK K       â $rV+V   JŽ â xJÿ   M  â 0rPÿP    A   â  rMOd     A  â rLQL   ÿú� Ä ErJ J       â #r6MK   6   â  rddP     A  â xKÿ   L  Æ #r5dJ   5   â rKKL   ÿ� â &rJj   ÿ   ä &rLh   ÿ   ä &rj j   ÿ   ä &rhh   ÿ   â rSSS   ÿ� â &rJ*g   J   Ä &rLl   L   ä&rjg   j   ä &rhl   h   â rSSV   ÿú� â &rJm   J   ä &rNk   L   ä &rjm   j   Ä &rh!k   h   ä &rLf   J   âxJÿ   S  â &rMi   N   ä&rQf   j   ä &rhi   h   â rJJS   ÿ� â $rN[L   ÿŽ â rLLM   ÿàÿ â 0rbÿb    A   â xKÿ   K  â $rRZQ   ÿŽ â rSQh   ÿàÿ â $rNM   NŽ â xJÿ   J  Ä rQLL   ÿ� â  rMOb     A  âErK K       â  rbPb     A  â ErJ J       ä rNQN   ÿú� â #r4MK   4   â &rLo   ÿ   â #r3bJ   3   â xKÿ   N  â &rJq   ÿ   Æ rKKN   ÿ� â &rNo   ÿ   ä &rM*e   J   ä &rLs   L   â rSSS   ÿ� â $rR+h   RŽ ä &rNs   N   ä &rJt   M   Ä &rLr   L   â rRSR   ÿú� â &rN!r   N   ä &rJc   J   ä &rQp   L   âxMÿ   R  â &rpp   N   ä &rq q   ÿ   Ä $rL[J   ÿŽ ä $xNJ   QŽ â rMMR   ÿ� â 0rnÿn    A   â xKÿ   K  â &rqe   q   ä $rLQ   LŽ â rQNN   ÿ� ä xJÿ   M  â &rtt   q   â ErK K       â  rMOn     A  â rLQL   ÿú� â  rnPn     A  â ErJ J       â &rcc   t   â #r2MK   2   â xKÿ   L  Ä #r1nJ   1   â $rRZc   ÿŽ â rccp   ÿàÿ â &rJv   ÿ   â rKKL   ÿ� â &rLy   ÿ   â rccc   ÿ� â $rR+p   RŽ ä &rv v   ÿ   ä &rJ*#   J   Ä &rLz   L   ä &ryy   ÿ   â rRcR   ÿú� â &rv#   v   ä &rJ{   J   ä&rL"   L   ä &ryz   y   â x#ÿ   R  â &rv{   v   Ä &rJw   J   ä&rMx   L   ä &ry!"   y   â r##R   ÿ� â $r"[J   ÿŽ ä &rww   v   â rLJM   ÿàÿ â &rxx   y   Ä $rMM   "Ž â x"ÿ   #  â $rJZw   ÿŽ ä 0ruÿu    A   ä $xww   xŽ â xKÿ   K  â $rx+x   JŽ â ErJ "       â  r"Pu     A  â ErK K       â  r#Ou     A  Ä #r/"J   /   â &r"   ÿ   ä &rJ$   ÿ   â#r0#K   0   â &r#*%   "   ä &rJ}   J   ä &r    ÿ   ä &r$$   ÿ   Ä &r#‚   #   ä &rJ�   J   ä &r%      ä &r$}   $   â rLLL   ÿ� â &r"€   #   ä &r%~   J   Ä &r�!�   $   â rLLM   ÿú� â $r$["   ÿŽ â r""%   ÿàÿ â&r‚      â x#ÿ   L  â $r$%   $Ž â r%""   ÿ� â &r€      ä &r~~   �   â r##L   ÿ� â $rJZ   ÿŽ â r$%$   ÿú� Ä rwww   ÿ� ä r~   ÿàÿ â0r|ÿ|    A   â x%ÿ   $  â $rJ+~   JŽ â x#ÿ   #  ä rwwx   ÿú� ä r   ÿ� Ä r$%$   ÿ� â  r%O|     A  â Er# #       ä x"ÿ   w  ä rJJ   ÿú� â 0rÿ    A   â x$ÿ   $  â #r.%#   .   â r""w   ÿ� Ä x%ÿ   J  â  r#O     A  â Er$ $       ä x"ÿ   "  ä r%%J   ÿ� â #r,#$   ,   â Er" "       â &r#„   ÿ   â x%ÿ   %  â  r|P|     A  Ä &r$ˆ   #   â Er% %       â #r-|"   -   â  r#P     A  â &r"ƒ   ÿ   ä&rƒ ƒ   ÿ   â #r #%       â &r„„   ÿ   ä &r"*‰   "   Ä &r#'   ÿ   ä &rƒ‰   ƒ   ä &r' '   ÿ   ä &rŠ   ÿ   ä&rŠ   ÿ   ä &r„ˆ   „   Ä &r"…   "   ä&rJ…   ƒ   ä &r$†   $   ä &r'Ž   '   ä&r#*Ž   #   ä &r�      Ä &r�      ä &r„!†   „   ä &r"‡   "   ä&r‡‡   J   ä &rK   $   ä &rŒ   '   Ä &rŒ   #   ä &r)      ä&r!)      ä &r„   „   ä $r["   ÿŽ ä $rZ‡   ÿŽ Ä $x""   KŽ ä &r‹      ä&r‹      ä &r(      ä&r(      â r‡‡„   ÿàÿ â $rK   Ž Ä $r„+„   Ž â r""   ÿ� â $rZZ   ÿŽ ä $r[[   ÿŽ â r   ÿàÿ â$x   Ž â r   ÿú� â $r[   [Ž â r‡‡‡   ÿ� â $rZ+   ZŽ â r   ÿ� Ä r   ÿ� ä xÿ     ä r„‡„   ÿú� ä rZ   ÿú� ä r[   ÿú� ä x::   ÿàÿ Ä r   ÿ� ä xÿ   „  ä xÿ     ä xÿ     ä | :   pbô ä r„„   ÿ� Ä r   ÿ� ä r   ÿ� â 0r&ÿ&    A   â xÿ     â 0r�ÿ�    A   â x„ÿ   „  ä xÿ     ä xÿ     â  rP&     A  âEr        â  rO&     A  âEr„ „       â  rPP�     A  âEr        â  rO�     A  â Er        Ä xCCP  ÿà÷ â #rH   H   â #r<„   <   â #r;P   ;   â #r9   9   â x==@   ÿàÿ ä x@@@   ÿàÿ ä xBB@   ÿàÿ ä x>>@   ÿàÿ Ä xDD@   ÿàÿ ä xEE@   ÿàÿ ä xFF@   ÿàÿ ä xGG@   ÿàÿ ä rAÿA   ÿäÿ â D)       À â Gy  �Ôÿÿÿÿƒ ê Ay        € ê Ey  0    € â G‰      € ê r778   ÿHŽ Ê ˆs 76 øÿ H   è ˆs 75€øÿ H   è ˆs 74 ùÿ H   è ˆs 73€ùÿ H   è ˆs 72 úÿ H   è ˆs 71€úÿ H   è ˆs 70 ûÿ H   è ˆs 7/€ûÿ H   è ˆs 7. üÿ H   è ˆs 7-€üÿ H   è ˆs 7, ýÿ H   è ˆs 7 €ýÿ H   è ˆs 7H þÿ H   è ˆs 7<€þÿ H   è ˆs 79 ÿÿ H   è ˆs 7;€ÿÿ H   ä Ay        € ê {          ì M        € ê „y8     H   b Ey  @    € æ „y8 €   H   ¨ „y8    H   â !r6       â/ !r55       È_ ‰      " !r44       Æ� ‰5      h ‰4      ¢ !r       Æ y      %   " !r5       Æ/ ‰      b !r	4       ÆO ‰      ¨ ‰	      â $x   ÿ Ž â !r       È/ x   ÿàÿ äz  r  p`ð â !r       âO ‰ €    " z  r  p`ò â !r
	       â� r 8ÿ   pVp  â‰ €    b x 8   p"ò Æ ‰
 €    ° ‰      &   ä         Ä       &   " !r       Æ       '   " ˆ        â !r       ä/ ‰ @    b !r
       ÆO ‰	 @    ¨ ‰ @    h ‰      '   b $ x  Ž Æ„y
8 €  H   â $ ~  Ž â !r       â/ $Š x  Ž â� !r		       ÆO ‰      " !r       Æ‰	      b $Š ~  Ž Æ ‰      ¢ !r       â %† j   Ž â !r	       Æ/ % j   Ž â †‰     â !r       ÆO †    â G	  �     € ê y      &   ( y      '   b $z x  Ž È $z ~  Ž Ê/ z r  ÿàõ Ê $rÿÿ   ÿ â z j  ÿ„ È z k   Ê †y     è Ay        € ê !r

3       â „y	8    H   â Ey  �    € æ ‰
      ( „y8 €  H   è „y8    H   â !r
       Ê ‰      $ !r       Ê ‰ €    $ !r       Ê ‰ @    $ !r       Ê ‰      $ !r       â G™  °     € ê y      &   " ¹z  r      ä �x   ?àÿ â y      '   b $z x  Ž È $z ~  Ž Ê/ |   ÿàõ È rÿÿ   ÿä ä z j  ÿ„ È z k   Ê †y     è Ay        € ê „y8 €  H   b !r1       â !r		2       â !r0       â „y8    H   ¢ z#  r      Æ „y
8 €  H   ä $x+#   ÿŽ ä„y8    H   " $x!#   ÿŽ Æ „y8 €  H   h „y8    H   è „y8 €  H   è „y8    H   è ‰      ( ‰	      ( ‰      " !r//       Æ/ „y88 €  H   â !r.       âO !r-
-       ä� ‰
/      b !r,       Æ‰-      ¢ !r        â!r       ä ‰      " !rH       â !r	       ä ‰      â !r<       â !r       Ä ‰      " !r99       Æ ‰      h ‰      b !r
/
       Æ/ ‰      b !r--       ÆO ‰9      ¢ !r       â !r88;       ä ‰      " !r       Æ� ‰      â !r       Æy       &   " !r       Æy	      '   b !r       â !r       ä/ ‰
      b !r99       ÆO ‰      ¢ !r       Æ ‰-      " !r       Æ� ‰      â $z   x  Ž Æ‰      " $z 	 ~   Ž Æ‰ €    h ‰      b r#    ÿ„ â !r
       â/ ‚ +   ÿà÷ â ‰      b !r       âO $rÿÿ   ÿ â z j  ÿ„ â ‰9      ¢ !r--       â $‚ÿÿ   ÿŽ â z k   â ‰8      " !r       â� Š
 j  ÿ† Ä ‰ €    â !r       âŠ k  � ä ‰	 €    " !r       â‚ !   ÿàù â !r       â ‰ €    d ‚*ÿÿ   ÿä â !r       â/ ‰ €    b Š j  ÿˆ â !r99       ÄO ‰- €    ¢ Š k  * â !r8       ä ‰ €    " !r       Æ� ‰  €    â $x#   ÿŽ â!r	       ä‰ @    " r	#    ÿŠ är    ÿàõ â ‰" €    " !r       ârÿÿ   ÿäÿ Ä ‰$ €    b rÿÿ   ÿä â !r       â/ z j  ÿ„ â ‰&9 €    b !r-       âO $x-#   ÿŽ â z k   â ‰(      ¢ !r       â z	 j  ÿŠ Ä ‰ @    " !r         â� z		 k  � ä ‰ @    â !r       â‚ -   ÿà÷ â !r""       â ‰ @    " !r$$       Æ‰ @    b !r&9&       Æ/ ‰ @    b !r((       ÆO ‰ @    ¢ !r       Æ ‰  @    " !   ÿàÿ â !r       ä� ‰#" @    â     ÿàõ ä ‚ÿÿ   ÿäÿ â ‰%$ @    â Š j  ÿ† â $*ÿÿ   ÿ â  j  ÿ„ â ‰)( €    â !+   ÿàÿ â !r       âŠ k  � â ‰'& @    "  k  * â !r       â! !   ÿàõ â ‰      b !r       â/ !r       âO *ÿÿ   ÿä â ‰      b ! j  ÿ„ â !r         Ä ‰      " !r#"#       â� ! k  * â !r%$%       â ‰      ¢ !r)()       Æ ‰      â !r'&'       Æ‰      " !r!       Æ‰      b !r       Æ/ ‰       b !r       Æ ‰#      ( ‰%      " !r       ÆO ‰$) @    ¢ !r       Æ� ‰"'      â !r       â!r       ä†‰ !    â	 !r        Æ/ †    â	 !r##       Æ †‰     â	 !r%%       Æ †     â	 !r$)$       ÆO †‰     â	 !r''"       Æ� †    è	 †‰ 
    è	 †     è	 †‰ #    è	 † %   è	 ‰$      ( †‰ '    â	 M™        € ê x-   ÿàÿ â!r$       Æ r    ÿàñ È r ÿÿ   ÿä  ä z j  ÿ€ È z k     Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y3      "   " ¹z  l      â Ey  ,    € â ™x?    â y4      !   b ¹z  F   
   â $r7ÿÿ   ÿ Ž â r5 ÿ       â $r8ÿÿ   ÿ Ž â rD ÿ       â $r ÿÿ   ÿ Ž â x*      ÿ  â x,      ÿ  â x.      ÿ  â x0      ÿ  â r2 ÿ       Ä r 3ÿ   pRð är 3ÿ   pRò är34   ÿ(Ž È/ xÿ     È r   ÿ� È x6ÿ     È | 6   pbô Ú G)  °*    € ê y      '   " $rÿÿ   ÿ Ž â xøÿÿÿÿÀŽ â $rÿÿ   ÿ Ž â y      &   b $vÿ q  ÿ Ž â z?  X      â $xB6   ÿ Ž â y	      %   ¢ z=  Y      â $r7ÿÿ   ÿ Ž Ä r9B   ÿŽ â$x<   BŽ âzAB q  ÿàÿ â $x:   BŽ ä$x>   BŽ ä$x@   BŽ ä 'z y   Ž È 'z s   Ž â/ zÿ z    ä x		   ÿ   âO $z; }  ÿŽ â zÿ t    â $z |  6Ž âx	   ÿàÿ â$z	 p  6Ž ä $z v  Ž Ä $z |  Ž â r  €ÿàÿ â $z p  Ž ä $z
 w  ÿŽ â xÿ     â$z v  Ž â x   ÿ   ä r;
;   ÿàõ â %xJ   Ž È %xJ   ÿŽ È %x
J   Ž È %xJ   ÿŽ È %x   Ž È $rCÿÿ   ÿ ä %x   Ž È r?   ÿàõ È r=   ÿä Ê �yy   • ¨ �yo   • è �yx   • ( �yw   • b	 r?   ÿàõ Ê $x   = Ê �yn   • h �yu   • h �yv   • h �ym   • b xY $       Ä xy   ÿ   ÈO xø  ÿÀŽ â $xo   ÿ Ž Æ� z  ÿàõ ä xø  ÿÀŽ ä zÿ 	 ÿä ä z  ÿàõ È �y    › ¢ $xx   ÿ Ž âzÿ 	 ÿä È xø  ÿÀŽ ä �y    › ä z  ÿàõ ä xw   ÿ   Æ$vÿ 	 ÿ â xø  ÿÀŽ È z   ÿàõ â �y    › ( $v!ÿ 	 ÿ Ì �y      › b xn   ÿ   È x"ø  ÿÀŽ â $xu   ÿ Ž Æ z""  ÿàõ ä xø  ÿÀŽ ä z#ÿ 	 ÿä ä z  ÿàõ â $xv   ÿ Ž Æ �y""    › ¤ xø  ÿÀŽ â $xm   ÿ Ž â zÿ 	 ÿä ä z  ÿàù ä r;   Bàµ ä xø  ÿÀŽ â �y    › ¢ $vÿ 	 ÿ â rC   ÿd~ Ä z  ÿàõ Æ $x$   ÿŽ â �y    › ¢ zÿ 	 ÿä â %v Z  Y Ž Ê �y    › ¢ $x   Ž â r?   ÿàõ È �yk   ™ ¨  �yG   ™ ¢  r=   ÿä Æ �yl   ™ ¨  �yF   ™ ¨  �yW   ™ ¨  �yV   ™ ¨  �yUB   ‘ ¢ r&
?   ÿàõ Æ �yH   ™ ¢  rb;   AàÇ Æ �yK    ™ ¢  $x'   = â r)C   ÿ„þ â %vbb Z  Y Ž â �yE    • ¦  $x)$   ÿŽ â/ �yT&B   ‘ ¢ r(?   ÿà÷ Æ rcc   ÿàÿ ä r$?   ÿàõ â $x)   =Ž ä �yMb   ™ ¢  r%	=   ÿä ä rL;   9àµ â �yQb   ™ è  �yOb   ™ â  r\C   ÿd~ Æ �yRb   ™ â  rP;   :àµ Æ �yJ(    • è  �ySb   ™ è �yI$    • " %v(L Z  Y Ž Æ �yLb   ™ " r&;   <àÙ Æ/ �yNb   ™ " %v$P Z  Y Ž Æ �yPb    ™ " r[C   ÿ¤~ är;   >àÙ ärZC   ÿd~ ä r;   @àµ â %v Z  Y Ž È $x]\$   ÿŽ ä %v&& Z  Y Ž È %v Z  Y Ž â rYC   ÿ¤~ Æ $x))   ]Ž ä $x[[$   ÿŽ ä $xYY$   ÿŽ â �y](   ™ b $x_Z$   ÿŽ â r''[   ÿàÿ ä rY   ÿàÿ â �y[(   ™ b rXC   ÿd~ â $x%%   _Ž Ä �yY(   ™ h �y_(   ™ b $xaX$   ÿŽ Æ �y`(   ™ h �y^(   ™ h �yX(   ™ h �y\(    ™ h �yZb    • h �yd&   ™ h �yg&   ™ h �yc&   ™ h �yh&   ™ h �yi&   ™ h �yb&   ™ h �ye&   ™ h �yf&    ™ b $x   aŽ â x~ÿ	   o  Ä �ya(    • d 	s ~        " x‰ÿ	   x  â �yq$   ™ h �yo$   ™ h �y($   ™ b# 	sƒ ‰        f �y)$   ™ b x†ÿ	   y  Ä x   ÿ   â �yt$   ™ h �ys$   ™ b 	s„ †        " x�~€   øŽ ä �yr$   ™ h �yp$    ™ b $xˆ�ÿŽ â x‡  @€    â $xŠƒ€   ÿ Ž â/ x‹ÿ	   w  â �yj&    • b x‡ˆ€€€€‡ìŽ Ä 	sŒ ‹        b xŠ‰€   ŠøŽ â $t…ÿÿ Ž â x‰‡ÿþþþÿàÿ â �y}   ™ b x‡„   ÿ   ä x…ˆ€€€€…ìŽ â �y&$    • b! x‡†€   ‡øŽ â $xŠŠÿŽ Ä �y'   ™ b x……ÿþþþÿàÿ â $x‡‡ÿŽ ä �yy   ™ b $tŽÿ @€ÿ Ž Æ �y{   ™ h �yx   ™ h �y|   ™ h �yz   ™ h �y$    ™ h �y%    • h �yw   ™ h �y‚   ™ h �y€   ™ h �y   ™ h �y~   ™ h �y�   ™ h �y   ™ h! �y    ™ h �yƒ    • b $xŒŒ€   ÿ Ž â x„‰ @€ˆøŽ Ä x……ˆøŽ ä x‰     â $tÿ @€ÿ Ž â x     È xˆŠ€€€€ìŽ äx‡€€€€ìŽ äx‡€€€€ŽìŽ ä xŒ‹€   ŒøŽ ä x†Š€€€€‰ìŽ ä x‹ˆÿþþþÿàÿ Ä xÿþþþÿàÿ ä xˆÿþþþÿàÿ ä x‰†ÿþþþÿàÿ ä x‡øŽ äxˆˆ @€‡øŽ â $xŒŒÿŽ â x†‹ @€ŠøŽ â$t‹ÿ @€ÿ Ž â x‰ŠøŽ Ä x‰     ä x‡˜º  ÿ    ä xˆˆ˜º  ÿ    ä x‰Œ€€€€‰ìŽ äxŠŒ€€€€‹ìŽ ä x‡ÿÀŽ Ä x‹ˆÿÀŽ äx‡‡€€€€ÞŽ äO xˆˆ€€€€ÞŽ ä x�‰ÿþþþÿàÿ ä r‰‡  €ÿàÿ â $xˆ   ‹
Ž â x……˜º  ÿ    Ä x‰€€€€´Ž ä xˆ„˜º  ÿ    ä x‰ÿ	   u  ä x„…ÿÀŽ äx€€€€´Ž ä x……€€€€ÞŽ â� 	su ‰        " xˆÿÀŽ Ä xˆˆ€€€€ÞŽ ä r……„  €ÿàÿ Æ $xˆˆ   
Ž â x…€€€€´Ž ä x††˜º  ÿ    ä x…ÿ	   v  ä xˆ€€€€´Ž ä x˜º  ÿ    â 	sˆ …        b x†ÿÀŽ Ä x††€€€€ÞŽ äx��ŒøŽ ä x�Šÿþþþÿàÿ ä x„ÿÀŽ â$x††   
Ž â x€€€€ÞŽ ä xvu   ÿ   Ä x��˜º  ÿ    ä xuÿ	   n  ä x‡� @€ŒøŽ ä r„  €ÿàÿ ä xn�ÿÀŽ â 	s„ u        " x‰‰€   vøŽ Ä x†€€€€´Ž ä x��€€€€ ÞŽ äxv‡˜º  ÿ    ä x†ÿ	   m  ä x€€€€´Ž â $x�   n
Ž â xvÿÀŽ â$x‰‰ÿŽ â xvv€€€€!ÞŽ â 	s‡ †        ¢ $xˆˆ€   ÿ Ž â/ xn     â $tŠÿ @€ÿ Ž â rvv  €ÿàÿ Ä x‰€€€€nìŽ ä xˆ…€   ˆøŽ ä xm‰€€€€ŠìŽ ä x€€€€ ´Ž â $xˆˆÿŽ â x ÿþþþÿàÿ ä xnmÿþþþÿàÿ Ä x„„   ÿ   ä x…  @€    ä x ‰øŽ äx n @€‰øŽ ä x„u€   „øŽ â $x‡‡€   ÿ Ž âO xnˆ€€€€…ìŽ â $tmÿÿ Ž â x…     â $x„„ÿŽ â xunÿþþþÿàÿ â $t‰ÿ @€ÿ Ž â x‡†€   ‡øŽ Ä xnu @€ˆøŽ ä xmˆ€€€€mìŽ ä xu„€€€€…ìŽ â $x‡‡ÿŽ â x!v€€€€!´Ž ä xv„€€€€‰ìŽ ä x†     Ä xmmÿþþþÿàÿ ä xuuÿþþþÿàÿ ä x…vÿþþþÿàÿ ä xv‡€€€€†ìŽ ä xmmˆøŽ â $tˆÿ @€ÿ Ž â xuu„øŽ Ä x…… @€„øŽ ä xvvÿþþþÿàÿ ä xuu˜º  ÿ    ä x†‡€€€€ˆìŽ ä x„…˜º  ÿ    ä xˆv‡øŽ Ä xvuÿÀŽ äxuu€€€€"ÞŽ ä x˜º  ÿ    ä x††ÿþþþÿàÿ ä x…„ÿÀŽ äx„„€€€€#ÞŽ Ä ruuv  €ÿàÿ ä xvÿÀŽ äx‡† @€‡øŽ ä x€€€€ÞŽ ä x† ˜º  ÿ    â $x „   …
Ž â xmm˜º  ÿ    Ä rv  €ÿàÿ ä xvn˜º  ÿ    ä xnmÿÀŽ äx  €€€€#´Ž ä xmm€€€€ÞŽ ä x#ˆ˜º  ÿ    Ä x€€€€´Ž ä xvÿÀŽ äxvv€€€€ÞŽ ä rmmn  €ÿàÿ ä xn#ÿÀŽ äx##€€€€ÞŽ â $xvv   
Ž â x"u€€€€"´Ž Ä r#n  €ÿàÿ â &r#k   ÿ   ä&rk"k   ÿ   â x€€€€´Ž â &rG   ÿ   ä &r#l   #   ä&rl l   k   ä &rkF      â xm€€€€´Ž â &r#W   #   Ä &rnV   k   â xv€€€€´Ž â &rWW   l   â xUð   ÿÀŽ â &rGG   ÿ   â x„‡˜º  ÿ    â &rlH   #   â xUU   ÿÀŽ â &r#!K   n   â x…†ÿÀŽ â&rGF   G   â x††€€€€ÞŽ â $rFUl   ÿŽ â xu„ÿÀŽ Ä xÿ     ä x„„€€€€ÞŽ ä rll#   ÿàÿ â$x††   …
Ž ä $rF#   FŽ â r#ll   ÿ� â $x„„   u
Ž ä &rVV   G   â xGTð   ÿÀŽ Ä x†€€€€´Ž ä rF#F   ÿú� ä x„€€€€´Ž ä x#ÿ   G  â &rWH   W   â xTT   ÿÀŽ ä xGÿ   F  â &rVK   V   Ä $rHTW   ÿŽ â rGGF   ÿ� â &rFM   ÿ   ä$xWW   VŽ ä &rM"M   ÿ   ä &rKQ   ÿ   ä &rFO   F   â rWWW   ÿ� â $rH#V   HŽ Ä &rM O   M   ä &rKR   K   ä &rQQ   ÿ   ä &rFS   F   ârHWH   ÿú� â &rSS   M   ä &rML   K   Ä &rQR   Q   â xKÿ   H  â &rFN   F   ä &rM!P   M   ä &rQL   Q   ä &rSN   S   â rKKH   ÿ� â $rHUF   ÿŽ â rLFM   ÿàÿ â &rPP   Q   Ä $rFTS   ÿŽ ä $xSS   PŽ ä $rP#P   FŽ â xFÿ   K  â 0rEÿE    A   ä 0rJÿJ    A   â xGÿ   G  â $rMM   HŽ â ErH F       â 0rIÿI    A   â  rFEJ     A  â ErG G       Æ  rKIE     A  â #r1FH   1   â &rF]   ÿ   ä #r2KG   2   â &rH[   ÿ   â rSSS   ÿ� â &rGY   F   â rLLL   ÿ� â &rF[   ÿ   ä &rH_   H   â rPSP   ÿú� â &rG`   G   â rLLM   ÿú� â &r__   F   Ä &rH^   H   ä &rFX   G   â xGÿ   P  â &r]"]   ÿ   â xEÿ   L  â &rK!\   H   â rGGP   ÿ� â &r] Y   ]   â rEEL   ÿ� â $rHUF   ÿŽ â rLFK   ÿàÿ â 0rZÿZ    A   â xFÿ   G  â &r]`   ]   â xEÿ   E  â &r_^   _   â  rGIZ     A  â ErF F       â  rZJZ     A  â &r]X   ]   â ErE E       â $rHK   HŽ â rKLL   ÿ� â &r\\   _   â #r/ZF   /   â $rMT]   ÿŽ â rHKH   ÿú� â &rFd   ÿ   â #r0GE   0   â $x]]   \Ž Ä &rGg   ÿ   â xEÿ   H  â &rFc   F   â rL]]   ÿ� â $rM#\   MŽ ä &rd"d   ÿ   ä &rGh   G   â rEEH   ÿ� â &rgg   ÿ   Ä &rFi   F   â rLLM   ÿú� â &rd c   d   ä &rHb   G   ä &rgh   g   ä &rFe   F   â xGÿ   L  â &rdi   d   Ä &rK!f   H   ä &rgb   g   ä $rHUF   ÿŽ â rGGL   ÿ� â &ree   d   â rLFK   ÿàÿ â&rff   g   ä $rKK   HŽ â xEÿ   E  â $rHTe   ÿŽ â xFÿ   G  â 0raÿa    A   Ä $xee   fŽ â ErE E       â $rf#f   HŽ â ErH F       â  rGIa     A  â rFJa     A  Æ #r.GE   .   â #r-FH   -   â &rEq   ÿ   ä &rF(   ÿ   ä&r((   ÿ   ä &rE)   E   ä &rFo   F   Ä &rq"q   ÿ   ä &rHo   (   ä &rEt   E   ä &r(s   F   ä &rq )   q   â rLLL   ÿ� â &rFr   E   Ä &r)!p   (   ä &rqt   q   â rKLK   ÿú� â &rHs   H   ä $rEUF   ÿŽ â rFF)   ÿàÿ â&rqr   q   â x(ÿ   K  â &rpp   H   â rHFF   ÿ� â $rE)   EŽ Ä $r)Tq   ÿŽ â r((K   ÿ� â $xqq   pŽ â rEHE   ÿú� â $r)#p   )Ž â x(ÿ   (  â 0rjÿj    A   â rLqq   ÿ� ä xHÿ   E  ä rLL)   ÿú� â  r)Ij     A  â Er( (       Ä rHHE   ÿ� ä reee   ÿ� â #r,)(   ,   â xEÿ   L  â 0r&ÿ&    A   â xHÿ   H  â &r('   ÿ   âreef   ÿú� â &r'"'   ÿ   â rEEL   ÿ� â &r(}   (   âErH H       â  r)I&     A  â &r} }   '   â xFÿ   e  â &r'x   ÿ   â xEÿ   E  â #r*)H   *   â rFFe   ÿ� â &r)|   '   â ErE E       â  r'J&     A  â xFÿ   F  â &rxx   ÿ   Ä #r 'E       â &r'w   ÿ   âErF F       â &rw"w   ÿ   â  rjJj     A  â &r~   ÿ   ä&r~   ÿ   ä &r(y   (   ä &r)z   )   Ä &rw ‚   w   â #r+jF   +   â &rx|   x   ä &r'‚   '   ä &r�      ä&r�      ä &rFy   }   Ä &r({   (   ä &r)!$   )   ä &r€   w   ä &rxz   x   ä &r€   '   ä &r      Ä &r      ä &r{{   F   ä $r&U(   ÿŽ â r(()   ÿàÿ â &r      ä&rx$   x   ä &r      Ä &r!!      ä&r      ä $r$T{   ÿŽ â r((   ÿ� â $r&)   &Ž ä $rTT   ÿŽ ä $x{{   xŽ Ä $rUU   ÿŽ â r!   ÿàÿ â $x   Ž â r&   ÿú� â $r$#x   $Ž â r{{{   ÿ� â $rU!   UŽ â r   ÿ� â $rT#   TŽ â r   ÿ� ä xÿ     Ä r${$   ÿú� ä rT   ÿú� ä rU   ÿú� ä x66   ÿàÿ ä r   ÿ� ä xÿ   $  Ä xÿ     ä xÿ     ä | 6   pbô ä r$$   ÿ� ä r   ÿ� ä r   ÿ� â 0r%ÿ%    A   â xÿ     â 0rƒÿƒ    A   â x$ÿ   $  Ä xÿ     ä xÿ     â  rI%     A  â r%J%     A  âEr        â  rJJƒ     A  âEr$ $       ä Er        â  rIƒ     A  â Er        Ä x??P  ÿà÷ â #rD   D   â #r8%$   8   â #r7J   7   â #r5   5   â x99@   ÿàÿ ä x<<@   ÿàÿ ä x::@   ÿàÿ ä x>>@   ÿàÿ Ä x@@@   ÿàÿ ä xAA@   ÿàÿ ä xBB@   ÿàÿ ä r=ÿ=   ÿäÿ â D)       À â Gy  ð×ÿÿÿÿƒ ê Ay        € ê Ey      € â G‰  ð     € ê $x33À  4Ž Ê ˆs 32 ùÿ H   è ˆs 31€ùÿ H   è ˆs 30 úÿ H   è ˆs 3/€úÿ H   è ˆs 3. ûÿ H   è ˆs 3-€ûÿ H   è ˆs 3, üÿ H   è ˆs 3+€üÿ H   è ˆs 3* ýÿ H   è ˆs 3 €ýÿ H   è ˆs 3D þÿ H   è ˆs 38€þÿ H   è ˆs 35 ÿÿ H   è ˆs 37€ÿÿ H   ä Ay        € ê {          ì M        € ê „y4     H   b Ey  @    € æ „y4 €   H   ¨ „y4    H   â !r2       â/ !r11       È_ ‰      " !r00       Æ� ‰1      h ‰0      ¢ !r       Æ y      %   " !r1       Æ/ ‰      b !r	0       ÆO ‰      ¨ ‰	      â $x   ÿ Ž â !r       È/ x   ÿàÿ äz  r  p`ð â !r       âO ‰ €    " z  r  p`ò â !r
	       â� r 4ÿ   pVp  â‰ €    b x 4   p"ò Æ ‰
 €    ° ‰      &   ä $ÿ   ÿ Ž Ä       &   " !r       Æ       '   " ˆ        â !r       ä/ ‰ @    b !r
       ÆO ‰	 @    ¨ ‰ @    h ‰      '   b $ x  Ž Æ„y
4 €  H   â $ ~  Ž â !r       â/ $Š x  Ž â� !r		       ÆO ‰      " !r       Æ‰	      b $Š ~  Ž Æ ‰      ¢ !r       â %† j   Ž â !r	       Æ/ % j   Ž â †‰     â !r       ÆO †    â G	  �     € ê y      &   ( y      '   b $z x  Ž È $z ~  Ž Ê/ z r  ÿàõ È rÿÿ   ÿä ä z j  ÿ„ È z k   Ê †y     è Ay        € ê !r

/       â „y4    H   â Ey  à    € æ ‰
      ( „y4 €  H   è „y	4    H   è „y4 €  H   è „y4    H   è „y4 €  H   â !r
       Æ „y4    H   è ‰      ( „y4 €  H   â !r       Ê ‰ €    $ !r       Ê ‰ @    $ !r       Ê ‰      $ !r       â G™  °     € ê y      &   " ¹z  r      ä �x   ?àÿ â y      '   b $z x  Ž È $z ~  Ž Ê/ |   ÿàõ Ê $rÿÿ   ÿ â z j  ÿ„ È z k   Ê †y     è Ay        € ê „y4    H   â !r*       â !r-       â !r		,       â „y4 €  H   â !r        â !rD       â !r.       â ‰      b !r+       â !r88       â z!  r      â ‰      ( ‰	      ¢ $x !   ÿŽ Æ ‰      è ‰      ( ‰       h ‰      b !r       â/ !r       Ä ‰8      " !r5       â !r	       äO y      '   b !r       â� !r7       ä y      &   ¢ !r       Æ‰      â !r         Æ‰      " !r
       Æ ‰      b !r8       Æ ‰      ( ‰      ( ‰       " $z x  Ž ÆO ‰      ¢ !r       â� $z# ~  Ž ä/ ‰      b !r       Æ‰	
      â !r       Æ‰      " !r       â !r       ä ‰ €    " !r         Æ ‰      b r!#   ÿ„ â !r       äO ‰ €    ¢ !r       Æ/ ‰      b $rÿÿ   ÿ â z j  ÿ„ â!r
	       â� ‰  €    ä z k   â !r       â‰ €    " !r       Æ ‰	 €    " ‚#    ÿàõ â !r       ä‰ €    b ‚ÿÿ   ÿä â !r       äO ‰ €    ¢ !r       Æ/ ‰ €    b !r         Æ� ‰ €    â r!#   ÿˆ â !r       ä‰ €    $ $r"ÿÿ   ÿ â z j  ÿˆ â‰  @    " !r
	       ä z k  " â !r       âŠ j  ÿ„ â ‰ @    " $x"!   ÿŽ â !r       âO Š k   â ‰
 @    ¢ $x!   ÿŽ â !r       Ä/ ‰ @    b !r       â� ‚	#   ÿà÷ ä ‰ @    â !r       â‚$ÿÿ   ÿäÿ ä ‰ @    " Š	 j  ÿ† â!r         â     ÿàÿ â ‰ @    b Š		 k  $� Æ ‰ @    b !r       â    ÿàÿ ä ‰ @    " !r
       âO r"#   ÿàù ä ‰  @    ¢ !r       â/ !#   ÿà÷ â $rÿÿ   ÿ â !r       â� ‰       b z j  ÿˆ â !r       Ä‰      â $ÿÿ   ÿäÿ â !r       ä‰      " 
! j  ÿ† â !r       ä ‰      b "#   ÿàõ â !r       Ä ‰      " z k   â $#ÿÿ   ÿ â !r         âO ‰      ¢ ! k  $� ä " j  ÿ„ â ‰      ¢ !r%        ä/ " k  # â ‰      b !r       Æ� ‰      â !r       Æ‰!       " !r       â!r       ä †‰ %    â !r       ÆO †    â !r       Æ †‰     â !r       Æ/ † 
    â !r       Æ� †‰     â !r! !       Æ†    è †‰     è †     è †‰     â M™        € ê †y !   â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y       "   " ¹z  l      â Ey   (    € â ™x?    â y      !   b ¹z  F   
   â r ÿ       â $rÿÿ   ÿ Ž â r ÿ       â $rÿÿ   ÿ Ž â x      ÿ  â $r0ÿÿ   ÿ Ž â r ÿ       â $rNÿÿ   ÿ Ž â x      ÿ  â $rPÿÿ   ÿ Ž â r  ÿ   pRò Ä r  ÿ   pRð â$x€     Ž Ê/ xÿ   €  È r
€   ÿ� È xÿ   
  È |    pbô Ú G)  @'    € ê y      '   " $rÿÿ   ÿ Ž â x�
øÿÿÿÿÀŽ â $rÿÿ   ÿ Ž â y      %   b $x   ÿ Ž ä $x€€   �
Ž â y      &   ¢ $vÿ X  ÿ Ž â z q  ÿàÿ Ä x�ÿ   €  ä x3€   ÿ   â 'z y   Ž â x   ÿ   È/ zÿ z    â $z }  ÿŽ ä 'z s   Ž ÈO $z
 p  Ž â zÿ t    â $z |  Ž ä$z |  
Ž â x
   ÿàÿ â $z v  Ž ä$z7 v  Ž â z  Y      â $z w  ÿŽ Ä $z
 p  Ž ä %x47J   €Ž â r   ÿàõ ä z  q      â %x67J   ÿŽ È %x8J   €Ž È %x~J   ÿŽ È %x:3   6Ž È $rÿÿ   ÿ Ž ä $x   Ž ä$x   Ž ä$x   Ž ä$x   Ž ä $rÿÿ   ÿ Ä %x23   ~Ž È r:   ÿàõ È r;   ÿä Ê �y{   • ¨ �y    • è �y   • ( �y&   • b	 r2   ÿàõ Ê $x3    Ê �yt   • h �y+   • h �yv   • h �ys   • b $t„ÿ$   ÿ Ž Ä $x
{   ÿ Ž ÊO x

ø  ÿÀŽ È z>
  ÿàõ ä x
    ÿ   Æ� $v?ÿ 	 ÿ â x

ø  ÿÀŽ â $x   ÿ Ž ÆzD
  ÿàõ â �y>>    › ¢ xø  ÿÀŽ ä zEÿ 	 ÿä ä z@  ÿàõ È �yDD    › â $vAÿ 	 ÿ ä $x&   ÿ Ž È�y@@    › " xø  ÿÀŽ È zH  ÿàõ È zIÿ 	 ÿä Ì �yHH    › b $xt   ÿ Ž Ê xJø  ÿÀŽ â $x+   ÿ Ž Æ zJJ  ÿàõ ä xø  ÿÀŽ ä zKÿ 	 ÿä ä zB  ÿàõ â $xv   ÿ Ž Æ �yJJ    › ¢ $vCÿ 	 ÿ â xFø  ÿÀŽ ä xs   ÿ   ä zFF  ÿàù â �yBB    › ¢ r)€   àµ ä xø  ÿÀŽ â $vGÿ 	 ÿ â r�   ÿd~ â %v() Z  „ Ž â z<  ÿàõ Æ �yFF    › ¢ z=ÿ 	 ÿä â $x$   ÿŽ È $x))   Ž â �y<<    › ¢ r4   ÿàõ Æ �y$(   ™ ¤  $x5    ä �y(   ™ ¨  �y'(   ™ ¨  �y(   ™ ¢  r8   ÿàõ Æ �yr(   ™ è  �yq(   ™ è  �ypB   ‘ â r9   ÿä Æ �y!(   ™ è  �yf(    ™ â  r#€   àÇ Æ �yoB   ‘ â  r
�   ÿ„þ â %v"# Z  „ Ž È $x
$   ÿŽ È $x##   Ž â r^6   ÿàõ â�y(    • ( �yh"   ™ ( �yl"   ™ ( �yj"   ™ h �ym"   ™ h �yn"   ™ b r,~   ÿà÷ Æ �yg"   ™ b r€   àÙ Æ/ �yi"   ™ b $x_7    â r-   ÿäÿ ä �yk"    ™ b r
�   ÿ¤~ ä r€   àµ â �y^^    • b %v Z  „ Ž Æ �ya,    • b r*�   ÿd~ är…€   àµ â �y\"    • b $x-
$   ÿŽ ä r
�   ÿd~ ä $x   -Ž Æ $x-
$   ÿŽ ä �yb   ™ h �y_   ™ h �y
   ™ h �yd   ™ h �ye   ™ h �yc   ™ h �y]   ™ h �y`    ™ b %v Z  „ Ž â r‡€   àÙ Ä �yW    • b $x)*$   ÿŽ È $x   )Ž Ê �y1   ™ h �yT   ™ h �y[   ™ h �yU   ™ b r%�   ÿ¤~ Æ �yX   ™ h �yZ   ™ h �yV   ™ b %v†‡ Z  „ Ž Æ �yY    ™ b $x%%$   ÿŽ Æ �y    • d r‡‡%   ÿàÿ ä x%ÿ	      â %v„… Z  „ Ž ä �y†   ™ b 	s" %        b. $x……   -Ž ä �yM†   ™ h �y †   ™ h �y„   ™ h �yR†   ™ h �y#„   ™ b $x""€   ÿ Ž Æ/ �yS†   ™ h �y.„   ™ b x{ÿ	   {  Æ �yQ†   ™ h �y,„   ™ b x/%€   "øŽ Æ �yO†   ™ h �y*„   ™ b 	sx {        & �yL†    ™ h �y(„   ™ b x-ÿ	     Æ �y"„   ™ h �y%„    ™ b 	s| -        b x&ÿ	   &  â $x//ÿŽ â x‚     â $tzÿ @€ÿ Ž â xxx   ÿ   â �y†    • d 	s) &        " x‚/€€€€‚ìŽ â�y„    • b xz/€€€€zìŽ Ä x‚‚ÿþþþÿàÿ ä xzzÿþþþÿàÿ ä x{{€   xøŽ â $x||€   ÿ Ž â/ xu‚/øŽ äx/z @€/øŽ â $x{{ÿŽ ä $tzÿÿ Ž â x|-€   |øŽ â $x))€   ÿ Ž Ä $txÿ @€ÿ Ž â xz{€€€€zìŽ â$x||ÿŽ â x&&€   )øŽ â $t)ÿÿ Ž â x-  @€    ä xzzÿþþþÿàÿ ä xx{€€€€xìŽ ä x)|€€€€)ìŽ Ä x-|€€€€-ìŽ ä x}z{øŽ â $x&&ÿŽ â xxxÿþþþÿàÿ â $tyÿ @€ÿ Ž â x))ÿþþþÿàÿ ä x--ÿþþþÿàÿ ä x}}˜º  ÿ    Ä xw     ä x{x @€{øŽ ä x))|øŽ äx-- @€|øŽ ä xx&€€€€yìŽ ä xy}ÿÀŽ Ä x|}€€€€>ÞŽ äO xw&€€€€wìŽ ä x{{˜º  ÿ    â $xy|   y
Ž â xwwÿþþþÿàÿ ä x}xÿþþþÿàÿ ä xuu˜º  ÿ    Ä xz{ÿÀŽ äx{{€€€€?ÞŽ ä xww&øŽ äx}} @€&øŽ ä xyy€€€€>´Ž ä x&/˜º  ÿ    Ä x/uÿÀŽ äx>u€€€€DÞŽ ä� rz{z  €ÿàÿ ä x{)˜º  ÿ    â $x/>   /
Ž â x+ÿ	   +  ä xu&ÿÀŽ Ä x&&€€€€EÞŽ ä x>-˜º  ÿ    â 	s) +        " x-{ÿÀŽ äxx{€€€€@ÞŽ â$x&&   u
Ž â xu/€€€€D´Ž ä r-x-  €ÿàÿ Ä x/>ÿÀŽ äx>>€€€€AÞŽ ä xvÿ	   v  ä x@-€€€€@´Ž â $x>>   /
Ž â 	s- v        b xtÿ	   t  Æ xA>€€€€A´Ž â $x>)€   ÿ Ž â xsÿ	   s  ä 	s) t        " xww˜º  ÿ    ä x/+€   >øŽ ä x&&€€€€E´Ž ä xEwÿÀŽ â	s+ s        ¢ xDw€€€€HÞŽ Äx>}˜º  ÿ    â $x--€   ÿ Ž â/ rEDE  €ÿàÿ â $x//ÿŽ â xx>ÿÀŽ â$tDÿÿ Ž â xw>€€€€IÞŽ â $t>ÿ @€ÿ Ž â x-v€   -øŽ â $x))€   ÿ Ž â xD/€€€€DìŽ Ä x>/€€€€>ìŽ â $x--ÿŽ â x?z€€€€?´Ž â $t|ÿ @€ÿ Ž â xz     â $xv+€   ÿ Ž âO xDDÿþþþÿàÿ ä x>>ÿþþþÿàÿ ä x)t€   )øŽ Ä x+-€€€€zìŽ äxDD/øŽ äx>> @€/øŽ ä x/-€€€€|ìŽ â $x))ÿŽ â xss€   vøŽ â $tvÿ @€ÿ Ž â xz+ÿþþþÿàÿ Ä xt     ä x|/ÿþþþÿàÿ ä xHE€€€€H´Ž ä x/z-øŽ ä xE)€€€€tìŽ äxz)€€€€vìŽ Ä x+| @€-øŽ â $x-sÿŽ â rxwx  €ÿàÿ â $t|ÿÿ Ž â xt  @€    ä xvEÿþþþÿàÿ ä xzzÿþþþÿàÿ ä xxx€€€€I´Ž Ä xE-€€€€|ìŽ äxI-€€€€tìŽ ä xsv)øŽ äxtz @€)øŽ ä xvEÿþþþÿàÿ ä xzIÿþþþÿàÿ Ä xIs˜º  ÿ    ä xtt˜º  ÿ    ä x)v-øŽ ä xEIÿÀŽ äxstÿÀŽ ä xvI€€€€JÞŽ Ä xtt€€€€KÞŽ ä xID˜º  ÿ    ä x>>˜º  ÿ    â $xtt   s
Ž â xDIÿÀŽ äxII€€€€BÞŽ ä xs>ÿÀŽ Ä x>>€€€€CÞŽ ä rIID  €ÿàÿ ä x//˜º  ÿ    â $x>>   s
Ž â xD+˜º  ÿ    ä x+I€€€€B´Ž ä xB/ÿÀŽ Ä x//€€€€FÞŽ â $xEv   E
Ž â x>>€€€€C´Ž ä x))˜º  ÿ    ä xCDÿÀŽ â$x//   B
Ž â xDD€€€€GÞŽ ä xB)ÿÀŽ Ä x))€€€€<ÞŽ ä xEE€€€€J´Ž ä rDDC  €ÿàÿ â $x))   B
Ž â xKt€€€€K´Ž â &rBy$   ÿ   âxGD€€€€G´Ž â &rDE$   ÿ   â x-z @€-øŽ â &r$@   ÿ   Ä &rB?'   B   âxJ-˜º  ÿ    â &rDK'   D   â xF/€€€€F´Ž â &r'A   $   â x-JÿÀŽ â&rBur   B   â xJJ€€€€=ÞŽ â &r'Hq   '   â x<)€€€€<´Ž â &r)F   ÿ   â x$pð   ÿÀŽ â &r&!   B   â xpp   ÿÀŽ â &r'xf   '   Ä $xJJ   -
Ž ä &r)G   )   â xÿ   $  â &rD+r   D   ä $r$p   ÿŽ â r'   ÿàÿ â &rD>!   D   â x=J€€€€=´Ž â &r)<q   )   â x!oð   ÿÀŽ â $r$'   $Ž â r   ÿ� Ä xoo   ÿÀŽ â &r)=f   )   â r$$   ÿú� ä xÿ   !  â $rBoD   ÿŽ ä $xDD   )Ž â x!ÿ   $  â $rB)   BŽ Æ r'DD   ÿ� ä r!!$   ÿ� â &r$yh   ÿ   â r''B   ÿú� â &rB@l   ÿ   ä &r$?j   $   ä &rhEh   ÿ   ä &rBAm   B   Ä &rlFl   ÿ   ä &r)un   $   ä &rhKj   h   â x$ÿ   '  â &rBHg   B   ä &rlGm   l   ä &r)&i   )   Ä &rh+n   h   â r$$'   ÿ� â &rBxk   B   ä &rl<g   l   ä $r'p)   ÿŽ ä &rh>i   h   â r))B   ÿàÿ â&rl=k   l   â x!ÿ   !  â $rBB   'Ž Ä $r'oh   ÿŽ ä 0r^ÿ^    A   ä 0rÿ    A   â Er! !       â $xhh   lŽ ä $rll   'Ž â  r'^     A  â 0raÿa    A   â x$ÿ   $  Ä r-))   ÿ� â #r'!      â &r!yb   ÿ   â Er) $       â &rbEb   ÿ   â  r$a     A  â &r'@_   ÿ   âr--B   ÿú� â &rB?
   !   â rhh   ÿ� â &r_F_   ÿ   Ä &rbK
   b   ä &r'Ad   '   â #r$)      â rl   ÿú� â &rBue   B   âx$ÿ   -  â &r_Gd   _   ä &rb+e   b   ä &r'Hc   '   â x
ÿ     â &r!&]   B   â r$$-   ÿ� â &r_<c   _   Ä &rb>]   b   ä &r'x`   '   âr

   ÿ� â $rp!   ÿŽ â x$ÿ   $  â &r_=`   _   ä 0r\ÿ\    A   â r!!'   ÿàÿ â$rBob   ÿŽ â Er$ $       â $r''   Ž â  r^\     A  â $xbb   _Ž Ä $r__   BŽ â rB!!   ÿ� â #rP$   P   â &r$y1   ÿ   â x
ÿ   
  ä r'B'   ÿú� â &rB@T   ÿ   â Er
 
       â &r$?[   $   â  r\a\     A  â &rDE1   ÿ   â rbbb   ÿ� â &r!AU   B   Ä &rTFT   ÿ   ä &ruX   $   â #r\
      â &rDK[   D   â r_b_   ÿú� â &r!HZ   !   â x
ÿ   '  â &rUGU   T   ä &r&V      â x$ÿ   _  â &r)+X   D   Ä &r!xY   !   â r

'   ÿ� â &rZ<Z   U   ä $r'p   ÿŽ â r$$_   ÿ� â &rV>V   )   â r!   ÿàÿ â &rZ=Y   Z   ä $r'!   'Ž â rB   ÿ� â $r!oV   ÿŽ â x
ÿ   
  â 0rWÿW    A   â x$ÿ   $  â $xVV   ZŽ â r'B'   ÿú� â $rZZ   !Ž â Er
 
       â  r^W     A  âEr! $       â  r$aW     A  â xBÿ   '  â #r
      â &ry   ÿ   â #r	$!   	   â &r!E   ÿ   â rBB'   ÿ� â &r@M   ÿ   â r)VV   ÿ� â &r$?       Ä &ryy   ÿ   â xBÿ   B  â &rAR      ä &rE   ÿ   ä &r@@#   ÿ   ä 0rÿ    A   ä &rMFM   ÿ   Ä &r#F#   ÿ   â r))Z   ÿú� â &r K    !   ä &r$uS   $   ä &ry?.   y   â ErB B       â &r!HQ      ä &rK.      Ä &rAA,   @   â  r^     A  â &rMGR   M   ä&r#G,   #   â x
ÿ   )  â &r +S       ä &r&O   $   ä &ryu*   y   Ä &r!xL   !   ä &r++*      ä &rAH(   A   â #rNB   N   â &rM<Q   M   ä&r(<(   #   â r

)   ÿ� â &r >O       Ä $rp   ÿŽ ä &ry&"   y   âr!   ÿàÿ â &r">"   +   ä &rM=L   M   ä&rAx%   A   ä&r(=%   (   â x
ÿ   
  â $r!!   Ž â  ra     A  â $r$o    ÿŽ â r   ÿ� â $roo"   ÿŽ Ä $x     MŽ ä $rppy   ÿŽ â ryyA   ÿàÿ â $x""   (Ž â Er
 
       â $r$M   $Ž âr!   ÿú� â $rpA   pŽ â r     ÿ� â $ro(   oŽ â ryyy   ÿ� Ä r"""   ÿ� â #r0
   0   â xÿ     ä r$   ÿú� ä rpyp   ÿú� ä r""o   ÿú� ä x   ÿàÿ Ä r   ÿ� ä x
ÿ     ä x!ÿ   "  ä xÿ   p  â 0rÿ    A   â |    pbô ä xÿ     Ä r

   ÿ� ä r!!"   ÿ� ä rp   ÿ� â 0rÿ    A   â Er        â  r^     A  â x
ÿ   
  ä x!ÿ   !  Ä xÿ     â #r      â Er
 
       â  ra     A  âEr! !       â  ra     A  â Er        â  r^     A  â xP  ÿà÷ â #r
      â #r!      â #r      â x@   ÿàÿ Ä x@   ÿàÿ ä x@   ÿàÿ ä x@   ÿàÿ ä x@   ÿàÿ ä x@   ÿàÿ ä rÿ   ÿäÿ â D)       À â Gy  PÛÿÿÿÿƒ ê Ay        € ê Ey  ð     € â G™  Ð     € ê $x  €  Ž Ê ˆs   úÿ H   è ˆs  €úÿ H   è ˆs  P ûÿ H   è ˆs  €ûÿ H   è ˆs   üÿ H   è ˆs  	€üÿ H   è ˆs  N ýÿ H   è ˆs  0€ýÿ H   è ˆs   þÿ H   è ˆs  €þÿ H   è ˆs   ÿÿ H   è ˆs  €ÿÿ H   ä Ay        € ê {          ì M	        € ê „y     H   b z+  r      Æ „y  €   H   $ $x'+   ÿŽ ä„y    H   ¢ $x)+   ÿŽ Æ „y €  H   è „y    H   ( „y €  H   h „y    H   h „y €  H   h „y €  H   h „y    H   è „y    H   â !r       â/ !r        Ä „y €  H   â !rP       ÆO ‰       " !r       Æ� ‰      b !r       Æ‰      ¢ !r
	       Æ‰      â !rN       Æ ‰      " !r       Æ ‰      b !r0       â !r         ä ‰      " !r       â !r       ä/ ‰       b !r       â !r       ÄO ‰	      ¢ !r       Æ� ‰
      â !r       â !r       ä‰      " !r       Æ‰      h ‰      b !r       Æ y      %   " !r        â/ !r		       äO y       &   b !r
       Æ� ‰	 €    ¢ !r       Æy      '   â !r       Æ‰      " !r       Æ ‰      b $x   ÿ Ž Æ ‰ €    ( ‰      " $z   x  Ž â/ x   ÿàÿ ä ‰      b !r
	       âO z  r  p`ð Ä ‰      ¢ $z  ~   Ž â� z  r  p`ò ä ‰      â !r       âr ÿ   p"p ä ‰      " !r       âx    p"ò Ä ‰      b !r       Æ ‰"      " !r       ä 
  r  ÿà÷ â ‰$      " !r       â/ r+    ÿˆ â !r       âO ‰
 @    b  '   ÿàõ â $ÿÿ   ÿŽ â !r        â� ‰ €    ¢ rÿÿ   ÿä â $ÿÿ   ÿ â !r       â‰	 @    â 
 j  ÿ† â !r       Ä‰ €    " z j  ÿˆ â!r""       ä ‰ €    " 
 j  ÿ„ â !r$$       ä ‰ €    b 
 k  � ä z k   â ‰ €    b 
 k   â !r
       â/ r+    ÿ„ â ‰!  €    b !r       ÄO r&ÿÿ   ÿä â ‰ €    ¢ !r	       â� z
 j  ÿ„ ä ‰ €    â !r       âz k  & ä ‰#" €    " !r       â 	+   ÿàÿ Ä ‰%$      " !r       â )   ÿàù â !r       â ‰ @    b ( 	   ÿà÷ ä $ÿÿ   ÿ â !r! !       â/ ‰ @    b 
 j  ÿˆ â !r       ÄO ‰      ¢ +ÿÿ   ÿäÿ â !r       ä� ‰ @    â ( j  ÿ† â !r#"#       ä‰ @    " 
	 k   â !r%$%       Ä ‰ @    " ( k  +� Æ ‰"! @    " !r       â$ÿ   ÿ Ž ä ‰ @    b !r       Æ/ ‰  @    b !r+       âO '   ÿàÿ ä ‰      ¢ !r       Æ� ‰$# @    â !r       Æ‰&% €    " !r       â !r"!"       â !        â ‰      " !r       Æ‰      b !r         Æ/ ‰      b !r-       âO     ÿàõ ä‰      ¢ %  j   Ž â!r$#$       ä� ‰      â %  j  ! Ž â !r&%&       Ä‰      " $(ÿÿ   ÿ â  j  ÿ„ ä‰!       " !r       â  k  ( ä ‰#"      " !r       Æ‰'& @    b !r       Æ/ ‰%$      b !r       ÆO †	 +    â !r       Æ� † -   â !r       Æ†	     â !r! !       â !r#"#       ä †     â !r'&'       Æ†	     â !r%$%       Æ/ †    è †	     è †     è †	 
!    è † 
#   è ‰'      ( †	 %    â M™        € ê x)   ÿàÿ âO !r''       Æ r    ÿàñ È r ÿÿ   ÿä  ä z j  ÿ€ È z k     Ê †y '    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y0      "   " ¹z  l      â Ey  ð$    € â ™x?    â y1      !   b ¹z  F   
   â $r6ÿÿ   ÿ Ž â r2 ÿ       â y5      %   ¢ $r>ÿÿ   ÿ Ž â x*      ÿ  â $r ÿÿ   ÿ Ž â x,      ÿ  â x.      ÿ  â r 0ÿ   pRð Ä r 0ÿ   pRò â$x0    1Ž ä/ $x55   ÿ Ž ÆO xÿ     È r   ÿ� È x4ÿ     È | 4   pbô Ú G)   #    € ê y      '   " r ÿ       â $rÿÿ   ÿ Ž â xøÿÿÿÿÀŽ â y      &   b x5   ÿàÿ â $x=4   ÿ Ž â z3  X      â $v9ÿ Y  ÿ Ž ä $r6ÿÿ   ÿ Ž â z<= q  ÿàÿ â 'z y   Ž È 'z s   Ž â/ z	ÿ z    Æ $z7 }  ÿŽ â zÿ t    â $z5 p  4Ž ä$z	 |  4Ž ä$z		 |  Ž ä $z v  Ž ä $x   
Ž Ä $z p  Ž ä $z
 w  ÿŽ â xÿ     â$z	 v  	Ž ä $vÿ q  ÿ Ž â r7
7   ÿàõ â %x	J   Ž Æ r8=   ÿŽ â %x	J   ÿŽ È %x
J   Ž È %xJ   ÿŽ È $x   ÿ Ž ä $x:   =Ž ä$x;   =Ž ä %x   Ž È $r?ÿÿ   ÿ ä %x   Ž È r3   ÿàõ È r9   ÿä Ê �y`   • ¨ �yZ   • è �ya   • ( �yf   • b	 r3   ÿàõ Ê $x   9 Ê �yT   • h �yE   • h �yU   • h �yS   • b x[ $       â $x`   ÿ Ž ÄO $xZ   ÿ Ž Æ� xø  ÿÀŽ È z  ÿàõ ä xø  ÿÀŽ ä zÿ 	 ÿä ä z  ÿàõ È �y    › ¢ $xa   ÿ Ž ä$vÿ 	 ÿ Æ xø  ÿÀŽ Æ �y    › â z  ÿàõ ä xf   ÿ   Æ$vÿ 	 ÿ â xø  ÿÀŽ Ê �y    › " z  ÿàõ Ê $vÿ 	 ÿ Ì �y    › b $xT   ÿ Ž Ê x ø  ÿÀŽ â $xE   ÿ Ž Æ z    ÿàõ ä xø  ÿÀŽ ä z!ÿ 	 ÿä ä z  ÿàõ â $xU   ÿ Ž Æ �y      › ¢ $vÿ 	 ÿ â xø  ÿÀŽ â $xS   ÿ Ž Æ z  ÿàù â �y    › â r$7   =àµ ä xø  ÿÀŽ â $vÿ 	 ÿ â r?   ÿd~ ä z  ÿàõ â %v$$ Z  [ Ž ä �y    › ä $x$   ÿŽ Ä $vÿ 	 ÿ ä $x%%   Ž â r3   ÿàõ Æ �y    › è �yC$   ™ è  �yA$   ™ " r9   ÿä Æ �yD$   ™ ( �y@$   ™ " r"
3   ÿàõ Æ �yR$   ™ ( �yQ$   ™ h �yPB   ‘ b r&7   <àÇ Æ �yB$   ™ b $x#   9 Æ �yF$    ™ b r(?   ÿ„þ â %v&& Z  [ Ž ä �yO"B   ‘ d $x)($   ÿŽ ä $rÿÿ   & Ž â/ �yW$    • b $x'   )Ž â r&3   ÿàõ È �yH   ™ h �yL   ™ b r'	9   ÿä Æ �yJ   ™ b r"7   8àÅ Æ �yM   ™ h �yN   ™ b rX?   ÿ„~ Æ �yG   ™ h �yI   ™ b %v"" Z  [ Ž Æ �yK    ™ b $xYX$   ÿŽ â r(3   ÿà÷ ä �yV&    • b $x##   YŽ ä $x)   9Ž â r^7   :àµ ä�yX"   ™ b r\7   ;àÙ Æ �y((    • b %v&^ Z  [ Ž â r]?   ÿd~ ä �yY"   ™ b %v$\ Z  [ Ž Æ/ �y^"   ™ b $xc]$   ÿŽ Æ �y)"   ™ b xgÿ	   Z  Æ �y["   ™ h �y]"   ™ h �yZ"   ™ h �y\"    ™ b 	sh g        f �y    • b $x''   cŽ â r_?   ÿ¤~ Ä xoÿ	   `  ä �y`&   ™ b $x__$   ÿŽ â xnÿ	   a  ä �yd&   ™ h �ya&   ™ b r%%_   ÿàÿ â $xhh€   ÿ Ž Ä/ �ye&   ™ h �yb&   ™ b xpÿ	   f  Æ �y_"    • b xqg€   høŽ Æ �y&   ™ h �yc&   ™ h �yf$   ™ h �yh$   ™ h �y"&    ™ h! �yk$   ™ h �yj$   ™ h �y#&    • h �yi$   ™ h �yg$   ™ h �y'$   ™ h �y&$    ™ b 	sl n        " $trÿÿ Ž Î 	sm o        b $xqqÿŽ â xul   ÿ   ä �yl$    • d xrq€€€€rìŽ â $ttÿ @€ÿ Ž â xun€   uøŽ â 	sv p        " xnrÿþþþÿàÿ â $xrm€   ÿ Ž â/ xsq€€€€tìŽ È xtsÿþþþÿàÿ ä xro€   røŽ ä xmt @€qøŽ â $ttÿ @€ÿ Ž ä $x%rÿŽ â xx  @€    â $xuuÿŽ ä $t$ÿÿ Ž â xo%€€€€tìŽ â $xsv€   ÿ Ž â xru€€€€xìŽ â $tvÿÿ Ž â x$%€€€€$ìŽ Ä xtoÿþþþÿàÿ ä xpp€   søŽ ä xxrÿþþþÿàÿ ä xnnqøŽ ä xr$ÿþþþÿàÿ ä xqu€€€€vìŽ Ä xtt @€%øŽ â$xppÿŽ â xrr%øŽ â $tsÿÿ Ž â xvqÿþþþÿàÿ ä xtt˜º  ÿ    ä x%r˜º  ÿ    ä xsp€€€€sìŽ Ä xovuøŽ äx$x @€uøŽ ä xqtÿÀŽ äxu  @€    ä xtt€€€€ÞŽ äO xr%ÿÀŽ Ä xssÿþþþÿàÿ ä x%%€€€€ÞŽ â $xtt   q
Ž â xup€€€€uìŽ ä xqspøŽ â $x%%   r
Ž â xsn˜º  ÿ    ä xuuÿþþþÿàÿ Ä xnm˜º  ÿ    ä xmÿ	   E  ä xpu @€pøŽ ä x%%€€€€´Ž ä xunÿÀŽ äxrn€€€€ÞŽ Ä� xsÿÀŽ â	sn m        " xss€€€€ÞŽ ä xoo˜º  ÿ    Æ $xEs   
Ž â rru  €ÿàÿ ä xr$˜º  ÿ    ä x$oÿÀŽ äxoo€€€€ÞŽ äxUÿ	   U  ä xEE€€€€´Ž â $xoo   $
Ž â 	s U        b xsrÿÀŽ â $x$n€   ÿ Ž â xrr€€€€ÞŽ Ä xTÿ	   T  ä xqq˜º  ÿ    â $xrr   s
Ž â xmm€   $øŽ ä 	s$ T        " xnqÿÀŽ ä xpp˜º  ÿ    ä x€€€€´Ž Ä xqq€€€€ÞŽ äxo€€€€´Ž ä xr€€€€´Ž ä xSÿ	   S  ä xpÿÀŽ ärqqn  €ÿàÿ â $xnmÿŽ â xpp€€€€ÞŽ â 	sm S        ¢ x   ÿ   â/ $toÿÿ Ž â xq€€€€´Ž â $tsÿ @€ÿ Ž â xqU€   øŽ â $xpp   
Ž â xn€€€€oìŽ â $xU$€   ÿ Ž â xon€€€€sìŽ â $xqqÿŽ â xp€€€€´Ž â $t$ÿÿ Ž â xÿþþþÿàÿ â $tpÿ @€ÿ Ž â xooÿþþþÿàÿ Ä xUT€   UøŽ ä x$q€€€€$ìŽ ä xnøŽ ä xTq€€€€pìŽ ä xo @€nøŽ ä xnm   ÿ   âO $tpÿÿ Ž â x$$ÿþþþÿàÿ â $xUUÿŽ â xTTÿþþþÿàÿ â $trÿ @€ÿ Ž â xnS€   nøŽ Ä xS$qøŽ äx$T @€qøŽ ä xTU€€€€pìŽ â$xonÿŽ â xmU€€€€rìŽ ä xTTÿþþþÿàÿ ä xnmÿþþþÿàÿ Ä xmo€€€€pìŽ äxTTUøŽ äxpo€€€€rìŽ ä xnn @€UøŽ ä xrmÿþþþÿàÿ ä xmT˜º  ÿ    Ä xt€€€€´Ž ä xtpÿþþþÿàÿ ä xpn˜º  ÿ    ä xnmÿÀŽ äxmm€€€€ ÞŽ ä xTroøŽ Ä xr˜º  ÿ    â $xm   n
Ž â xq˜º  ÿ    ä xmrÿÀŽ äxrr€€€€ÞŽ ä xqÿÀŽ ä xUt @€oøŽ Ä xqq€€€€ÞŽ â $xrr   m
Ž â xopÿÀŽ äxpp€€€€!ÞŽ ä xmS˜º  ÿ    ä x$$˜º  ÿ    â $xqq   
Ž â rppo  €ÿàÿ Ä xmÿÀŽ äxo$ÿÀŽ äxmm€€€€ÞŽ ä x$$€€€€ÞŽ ä xSq€€€€´Ž ä xr€€€€´Ž Ä xqT˜º  ÿ    ä rm  €ÿàÿ â $x$   o
Ž â x€€€€ ´Ž ä x qÿÀŽ äxqq€€€€ÞŽ ä x€€€€´Ž â &r%C   ÿ   â x€€€€´Ž â &rA   ÿ   Ä $xqq    
Ž ä &rD      ä &r @      â xTU˜º  ÿ    â &rER      â x!p€€€€!´Ž â &r Q       â xPð   ÿÀŽ â&rCC   ÿ   â xPP   ÿÀŽ â &rAA   ÿ   â xUTÿÀŽ â &rB      â xTT€€€€ÞŽ â &r F       â xÿ     â &rC!D   C   Ä &r@@   A   ä $rAP   ÿŽ â rTTU  €ÿàÿ â $x    Ž â xq€€€€´Ž â &rCSR   C   â x$Oð   ÿÀŽ â $rA    AŽ â xT€€€€´Ž â &rBB   C   â r    ÿ� â &r@Q   @   â xOO   ÿÀŽ Ä r  A   ÿú� â &r@F   @   â xÿ   $  â $r$OB   ÿŽ â xAÿ      â $xBB   @Ž ä $r$@   $Ž ä &r@%H   ÿ   â r A    ÿ� â &rHH   ÿ   â rCBB   ÿ� â &rAL   ÿ   Ä &rLL   ÿ   ä &rH!J   H   â r$C$   ÿú� â &rAM   A   ä&rLM   L   ä &rHSN   H   ä &rBG   A   âxAÿ   $  â &rLG   L   Ä &r@J   @   ä &rHI   H   â rAA$   ÿ� â &rLK   L   â x ÿ      â &r@EN   @   ä $r$OH   ÿŽ ä 0rVÿV    A   Ä 0rWÿW    A   â Er          â $xHH   LŽ ä &rCI   @   ä &rBK   B   ä $rLL   $Ž â  r$VW     A  â $rGPC   ÿŽ â rCCB   ÿàÿ Ä #r/$    /   â &r %X   ÿ   â xAÿ   A  â $rGB   GŽ â r@CC   ÿ� â 0r(ÿ(    A   ä &r$Y   ÿ   ä &r ^       â ErA A       â &rXX   ÿ   â rCHH   ÿ� â &r$)   $   â r@@G   ÿú� â &rYY   ÿ   â  rWW(     A  â &rGE[       â rCCL   ÿú� â &rX!^   X   â #r.WA   .   â &r$]   $   Ä &rY)   Y   â xAÿ   @  â &rGZ   G   â x ÿ   C  â &r[S[   X   ä &r$\   $   ä &rY]   Y   â rAA@   ÿ� â $r)PG   ÿŽ â r  C   ÿ� â &rZZ   [   Ä $xGG   $Ž ä &rY\   Y   â xAÿ   A  â $r$$   )Ž ä 0r)ÿ    A   â x ÿ      â $r@OZ   ÿŽ â rZZY   ÿàÿ ä rGGG   ÿ� â  rV)     A  â ErA A       â $r@Y   @Ž â Er          â  r)()     A  â rCZZ   ÿ� â #r-A   -   â r$G$   ÿú� â &r%`   ÿ   â r@C@   ÿú� â #r,)    ,   â xCÿ   $  â &r d   ÿ   Ä &rdd   ÿ   ä &r``   ÿ   ä &ra      â rCC$   ÿ� â &r$e       ä&rde   d   ä &r`!a   `   Ä &r Eb      âxCÿ   C  â 0r_ÿ_    A   ä &r$   $   ä&rd   d   ä &r`Sb   `   ä &rc       â ErC C       â &r %f   ÿ   â rV_     A  â &rff   ÿ   Ä &rh   ÿ   ä&rh   ÿ   ä &r$"   $   ä &r`c   `   â #r+C   +   â &rf!k   f   ä $rAP   ÿŽ Ä &rd"   d   ä &r k       ä &rj      ä&rj      ä $x   $Ž ä $rO`   ÿŽ Ä &rfSi   f   ä$x``   dŽ ä &r Ei       ä &r'      ä&r'      ä $rA$   AŽ Ä $rdd   Ž â r   ÿ� â &rfg   f   ä&rg       ä &r&      ä&r&      â rA   ÿú� â $rOf   ÿŽ Ä $rPP   ÿŽ â r   ÿàÿ â$xff   Ž â xÿ     â $rP   PŽ â r``   ÿ� â $r   Ž â r   ÿ� ä rff   ÿ� ä r   ÿ� â 0r#ÿ#    A   â rd   ÿú� Ä rP   ÿú� ä r   ÿú� ä xÿ     ä x44   ÿàÿ ä x)ÿ   @  ä xÿ     Ä xÿ     ä xÿ     â  rV#     A  â Er        ä | 4   pbô ä r))@   ÿ� ä r   ÿ� Ä r   ÿ� ä r   ÿ� â #r        â 0rÿl    A   â x)ÿ   )  ä xÿ     ä xÿ     ä xÿ     â  r(     A  âEr) )       â  r_(_     A  âEr        â  r#(#     A  â Er        â  rV     A  â Er        Ä x33P  ÿà÷ â #r*_)   *   â #r>#   >   â #r6   6   â #r2   2   â x88@   ÿàÿ â $r9ÿÿ   9Ž â x::@   ÿàÿ ä x;;@   ÿàÿ ä x<<@   ÿàÿ Ä x==@   ÿàÿ â D)       À â Gy  ÀÞÿÿÿÿƒ ê Ay        € ê Ey  Ð     € â G‰  °     € ê $x0@  1Ž Ê ˆs / ûÿ H   è ˆs .€ûÿ H   è ˆs - üÿ H   è ˆs ,€üÿ H   è ˆs + ýÿ H   è ˆs *€ýÿ H   è ˆs   þÿ H   è ˆs >€þÿ H   è ˆs 2 ÿÿ H   è ˆs 6€ÿÿ H   ä Ay        € ê {          ì M        € ê „y1     H   b z 5 r  p`ð â $t#ÿ   ÿ Ž â z%  r      â „y1 €   H   ¢ r 1ÿ   p"p Æ „y1    H   â $x!%   ÿŽ Æ „y1 €  H   ( „y1    H   h „y	1 €  H   ( „y1    H   ( „y1 €  H   ( „y1    H   è „y1 €  H   â !r/       Æ/ y      '   b !r.       ÆO ‰      " !r-       â� !r,,       È‰      ¢ !r+       â!r*	*       ä ‰	,      â !r        Æ ‰      " !r>>       â !r       ä ‰*      " !r22       Æ ‰      b !r6       Æ ‰       b !r       ÆO ‰      ¨ ‰>      ¢ !r	,	       Æ� ‰2      â !r       Æ‰
      " !r*       â !r       ä‰      " !r         Æ/ ‰      b !r       ÆO ‰	      ¢ !r>       Æ ‰  €    b !r2       Æ� ‰      â !r
       Æ‰      ( ‰
      " !r       Æ ‰      " !r       Æ/ ‰      b !r		       ÆO ‰      ¢ !r        â!r       ä� ‰ €    â !r       Æ‰ €    " !r
       Æ ‰	 €    b x
5   ÿàÿ â !r       ä y       &   " z 
 r  p`ò â !r       ä/ ‰ @    b x 1   p"ò â !r       ÄO ‰ €    ¨ ‰ €    ¢ !r       Æ� ‰ €    â !r       Æ‰ €    " !r	       Æ‰ €    b $z   x  5Ž Æ ‰ €    " !r
       â/ $z  ~   Ž â !r       äO ‰ @    d 	 !   ÿàõ â!r       â r%    ÿ† â‰
      ¢  ÿÿ   ÿä â !r       â� 
	 j  ÿ„ â ‰ @    â $rÿÿ   ÿŽ â !r       âz j  ÿ† â ‰ @    " 
		 k    â !r       âz k  � â ‰ @    b 
  r  ÿàù â !r       â r%    ÿŠ â ‰ @    $ $$ÿÿ   ÿ â 
 j  ÿˆ â ‰ @    " !r       â/ $r"ÿÿ   ÿŽ â z j  ÿŠ â ‰ @    b !r'
       âO %   ÿàÿ Ä ‰ @    ¢ !r       â� !   ÿàÿ ä ‰ @    â !r       â!    ÿàõ â%v
  j  # Ž â !r       â    ÿà÷ â ‰      " z k  "� â !r       â "ÿÿ   ÿä â ‰      " ! j  ÿ„ â !r       â $ ÿÿ   ÿŽ â ‰      b 
 k  $ â !r       â/  j  ÿ† â ‰      b !r       ÄO  k   � â ‰      ¢ !r       Æ� ‰      è ‰       â !r#       â! k  " ä ‰      " !r       Æ ‰      " !r       â!r       ä/ †	 
'    â !r       ÆO † 
#   â !r       Æ� †	     â !r        Æ †     â !r       Æ†	     â !r       Æ †    è †	     è †     è †	     â M™        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y       "   " ¹z  l      â Ey  �!    € â ™x?    â y+      !   b ¹z  F   
   â $r8ÿÿ   ÿ Ž â r* ÿ       â y0      %   ¢ $r2ÿÿ   ÿ Ž â r, ÿ       â $r1ÿÿ   ÿ Ž Ä $r<ÿÿ   ÿ Ž ä $rBÿÿ   ÿ Ž ä $r@ÿÿ   ÿ Ž â r  ÿ   pRð är  ÿ   pRò â$x     +Ž ä/ $x00   ÿ Ž ÆO xÿ     È r   ÿ� È x;ÿ     È | ;   pbô Ú G)  0     € ê y      '   " $rÿÿ   ÿ Ž â xøÿÿÿÿÀŽ â $rÿÿ   ÿ Ž â y      &   b $z	0 p  ;Ž âx0   ÿàÿ â $v-ÿ X  ÿ Ž â z4  q      â $v5ÿ Y  ÿ Ž â x7;   ÿ   â $r8ÿÿ   ÿ Ž Æ z67 q  ÿàÿ â $x/4   7Ž ä$x44   7Ž ä 'z y   Ž È 'z s   Ž â/ zÿ z    Æ $z3 }  ÿŽ â zÿ t    â $z |  ;Ž ä$z	 |  	Ž ä $z v  Ž ä $x   
Ž ä $z p  Ž Ä $z
 w  ÿŽ â xÿ     â$z	 v  	Ž ä $x   ÿ Ž â r3
3   ÿàõ â %x	J   Ž È %x	J   ÿŽ È %x
J   Ž È %xJ   ÿŽ È %x   Ž È $r9ÿÿ   ÿ ä %x   Ž È r-   ÿàõ È r5   ÿä Ê �yW   • ¨ �yT   • è �yX   • ( �yZ   • b r-   ÿàõ Ê $x   5 Ê �yP   • h �ya   • h �y`   • h �yA   • b xY $       â $xW   ÿ Ž ÊO xø  ÿÀŽ È z  ÿàõ â $xT   ÿ Ž Æ� zÿ 	 ÿä ä xø  ÿÀŽ â $xX   ÿ Ž ä�y    › ¢ z  ÿàõ ä xø  ÿÀŽ Æ $vÿ 	 ÿ â z  ÿàõ È �y    › â" $vÿ 	 ÿ â xZ   ÿ   Ê�y    › " xø  ÿÀŽ È z  ÿàõ Ê $vÿ 	 ÿ Ì �y    › b $xP   ÿ Ž Ê xø  ÿÀŽ â $xa   ÿ Ž Æ z  ÿàõ ä/ xø  ÿÀŽ ä zÿ 	 ÿä â $x`   ÿ Ž â z  ÿàõ È �y    › " $vÿ 	 ÿ â xø  ÿÀŽ â $xA   ÿ Ž â r&3   7àµ ä z  ÿàù â �y    › " xø  ÿÀŽ â %v&& Z  Y Ž â r9   ÿd~ Ä z  ÿàõ â $vÿ 	 ÿ ä $x!$   ÿŽ ä $vÿ 	 ÿ ä �y    › b $x''   !Ž â r -   ÿàõ ä �y    › d r!5   ÿä Ä �yG&   ™ h �yM&   ™ b r"3   6àµ Æ �y?&   ™ h �yO&   ™ b r$9   ÿd~ Æ �y=&   ™ h �yN&   ™ h �yJ B   ‘ b %v"" Z  Y Ž Æ �yL&   ™ b $x%$$   ÿŽ Æ �yK&    ™ b $x##   %Ž â rR
-   ÿàõ È �yE"   ™ b $xS   5 Æ �yC"   ™ h �y>"   ™ h �yD"   ™ h �y:RB   ‘ h �yF"   ™ h �yI"   ™ h �yH"   ™ h �y."    ™ b r(-   ÿà÷ Ä r$-   ÿàõ â �y&&    • d $x)	   5Ž â r 3   /àÇ â/�yQ"    • b r%5   ÿä ä rR9   ÿ„þ â �y((    • b r\3   4àµ Ä x[ÿ	   T  â %v   Z  Y Ž â �y$$    • b r]9   ÿd~ ä $xSR$   ÿŽ ä %v"\ Z  Y Ž ä/ 	sY [        d $x!!   SŽ Ä $x]]$   ÿŽ Æ �yR    ™ b $x##   ]Ž Æ �yS    ™ b xfÿ	   X  ä xcÿ	   W  â �y)    ™ b $x\Y€   ÿ Ž Æ/ �y'    ™ h �yX"   ™ h �yW"   ™ h �y%    ™ h �yU    ™ b x_[€   \øŽ Æ �yY"   ™ h �y^"   ™ h �yV    ™ h �yT     ™ b xjÿ	   Z  Æ �y]"   ™ h �y\"   ™ h �y["   ™ h �yZ"    ™ h �y      • b $xe_ÿŽ Æ �y_"    • b 	sd f        " xh     Î 	sb c        b $tlÿ @€ÿ Ž ä $xid€   ÿ Ž â xde€€€€hìŽ È 	sk j        " xif€   iøŽ ä xfdÿþþþÿàÿ â $xdb€   ÿ Ž â/ xge€€€€lìŽ È xhgÿþþþÿàÿ ä xdc€   døŽ â $xiiÿŽ â x!h @€eøŽ â$thÿ @€ÿ Ž ä $x"dÿŽ ä $t#ÿÿ Ž â xbfeøŽ â $tfÿÿ Ž â xc  @€    Ä xei€€€€hìŽ ä x#"€€€€#ìŽ äxkk   ÿ   ä xc"€€€€cìŽ ä xdi€€€€fìŽ ä xfeÿþþþÿàÿ Ä x##ÿþþþÿàÿ ä xkj€   køŽ ä xecÿþþþÿàÿ ä xddÿþþþÿàÿ ä x##"øŽ â$xkkÿŽ â xee @€"øŽ â $t"ÿÿ Ž â xcdiøŽ Ä xdf @€iøŽ ä x##˜º  ÿ    ä xfe˜º  ÿ    ä xgk€€€€"ìŽ ä x"#ÿÀŽ ä xefÿÀŽ Ä x##€€€€ÞŽ äO xik€€€€hìŽ ä xff€€€€ÞŽ ä xhgÿþþþÿàÿ â $xg#   "
Ž â xjiÿþþþÿàÿ â $x"f   e
Ž â xib˜º  ÿ    Ä xf!˜º  ÿ    ä xbiÿÀŽ äx""€€€€´Ž ä xii€€€€ÞŽ ä� xaÿ	   a  ä xfÿÀŽ Ä xcc˜º  ÿ    ä xff€€€€ÞŽ â 	s! a        " riib  €ÿàÿ ä xdd˜º  ÿ    ä xbcÿÀŽ â$xff   
Ž â xcc€€€€ÞŽ ÄxdÿÀŽ äxdd€€€€ÞŽ â $xcc   b
Ž â x`ÿ	   `  ä xehkøŽ â $xdd   
Ž â xc€€€€´Ž ä 	sc `        b xee˜º  ÿ    Ä xd€€€€´Ž ä xd!   ÿ   ä x#j @€køŽ ä xPÿ	   P  ä xbeÿÀŽ äxee€€€€ÞŽ Äxaa€   døŽ â 	s! P        " xd#˜º  ÿ    ä xg€€€€´Ž ä xgÿ	   A  â $xee   b
Ž â xdÿÀŽ â$xaaÿŽ â xdd€€€€ÞŽ â $tbÿÿ Ž â 	sA g        ¢ $xcc€   ÿ Ž â/ rdd  €ÿàÿ Ä xa€€€€bìŽ ä xf€€€€´Ž â $tfÿ @€ÿ Ž â xc`€   cøŽ ä xbÿþþþÿàÿ â $x!!€   ÿ Ž â xd€€€€´Ž â $xccÿŽ â xbaøŽ â $tbÿÿ Ž â x#a€€€€fìŽ Ä xd  @€    â $x`A€   ÿ Ž âO x!P€   !øŽ ä xf#ÿþþþÿàÿ ä xPc€€€€dìŽ äxAc€€€€bìŽ ä x#f @€aøŽ â $taÿÿ Ž â xgg€   `øŽ Ä xbPÿþþþÿàÿ â $xP!ÿŽ â xe€€€€´Ž â $teÿ @€ÿ Ž â x`Aÿþþþÿàÿ ä xAb @€cøŽ äx!`cøŽ ä x`P€€€€aìŽ â$xggÿŽ â xbP€€€€eìŽ Ä xd     ä xa`ÿþþþÿàÿ ä xcbÿþþþÿàÿ â $tfÿ @€ÿ Ž â x`g€€€€dìŽ ä xaaPøŽ äxcc @€PøŽ Ä xP`ÿþþþÿàÿ ä xbg€€€€fìŽ ä xaa˜º  ÿ    ä x`c˜º  ÿ    ä xePgøŽ ä xbbÿþþþÿàÿ Ä xPaÿÀŽ äxc`ÿÀŽ äxaa€€€€ÞŽ ä x``€€€€ÞŽ ä xgb @€gøŽ ä x˜º  ÿ    â $xaa   P
Ž â xb#˜º  ÿ    â $xP`   c
Ž â x`ÿÀŽ Ä x#bÿÀŽ äx€€€€ÞŽ ä xbb€€€€ÞŽ ä r`  €ÿàÿ ä xdA˜º  ÿ    â $xbb   #
Ž â x!!˜º  ÿ    Ä x€€€€´Ž ä xdÿÀŽ äxdd€€€€ÞŽ ä xb€€€€´Ž ä x`!ÿÀŽ äx!!€€€€ÞŽ Ä xbg˜º  ÿ    â $xdd   
Ž â xee˜º  ÿ    â $x!!   `
Ž â xbÿÀŽ äxbb€€€€ÞŽ ä x`eÿÀŽ äxee€€€€ÞŽ â $xbb   
Ž Ä &rG   ÿ   â ree`  €ÿàÿ â &r`M   ÿ   â xi€€€€´Ž ä xPP€€€€´Ž â &r"?      â x!€€€€´Ž â &r!O   `   ä &r=      â xJð   ÿÀŽ â &r`N   !   â xa€€€€´Ž â &rL      â xJJ   ÿÀŽ â &r`K   `   â xd€€€€´Ž Ä xb€€€€´Ž â &rbM   ÿ   â xe€€€€´Ž â &rGG   ÿ   ä $r!J   ÿŽ ä &rOO   b   ä $x   `Ž ä &rGP?   G   Ä &rbN   O   â rN   ÿ� â &rE   ÿ   ä &rG=   G   ä &rEE   ÿ   ä &r=C   ÿ   ä &r#">      Ä &rEP>   E   ä &r>D   =   â x:ð   ÿÀŽ â &rCC   ÿ   â x::   ÿÀŽ â &rGL   G   ä &r=F   #   ä &r>I   >   â xÿ     â &rbK   b   Ä &rCD   C   â xÿ     â $r#:G   ÿŽ ä &r=H   =   ä &r>.   >   ä &rCI   C   ä $xGG   bŽ Ä $r!`   !Ž ä $r?J=   ÿŽ â r==>   ÿàÿ â $r#b   #Ž ä &rD.   C   â r.GG   ÿ� â $r?>   ?Ž â r!N!   ÿú� ä r>==   ÿ� Ä r#.#   ÿú� ä x.ÿ   !  ä r?>?   ÿú� ä x>ÿ   #  ä r..!   ÿ� â &rEF   E   â r>>#   ÿ� â 0r(ÿ(    A   â x.ÿ   .  â 0r!ÿ&    A   â x>ÿ   >  â 0r$ÿ$    A   â Er. .       â &rEH   E   â  r#(!     A  â Er> >       â  r!!$     A  â $rA:E   ÿŽ â #r,#.   ,   â $xEE   DŽ â #r@!>   @   â &r#R   ÿ   Ä &rRR   ÿ   â rFEE   ÿ� â $rAD   AŽ ä &r>S   ÿ   â xDÿ   ?  â &rRP)   R   â rAFA   ÿú� â &r>'   >   â rDD?   ÿ� â &r.S   ÿ   Ä &rW   ÿ   ä &rX   ÿ   ä&r#")   #   ä &rX   ÿ   ä &rW   ÿ   ä &rR%   R   â x&ÿ   A  â &r>U   >   â xDÿ   D  â &r.'   .   Ä &r^      ä&r"Y      ä 0rQÿQ    A   ä &r#%   #   ä &rPY      ä &r^      Ä &r%V   R   â r&&A   ÿ� â &r>T   >   â ErD D       â &r\      ä &r]      â  r!(Q     A  â &r.U   .   ä &r]      Ä &r\      ä $r':%   ÿŽ â x&ÿ   &  â &r#V   #   ä $x%%   >Ž ä &rZ      ä &r[      â #rB!D   B   â &r.T   .   Ä &r[      ä &rZZ      â Er& &       â $r!J#   ÿŽ â r%%   ÿ� â $r'>   'Ž â  rQ$Q     A  â $rJJ   ÿŽ ä $x##   .Ž Ä $r::   ÿŽ â r   ÿàÿ â $x   ZŽ â r''   ÿú� â #r<Q&   <   â $r!.   !Ž â r&##   ÿ� â $r:   :Ž â r   ÿ� â $rJZ   JŽ â r   ÿ� Ä xÿ   '  ä r!&!   ÿú� ä r:   ÿú� ä rJJ   ÿú� ä r'   ÿ� ä x;;   ÿàÿ â 0rÿ     A   â xÿ   !  Ä xÿ     ä xÿ   J  ä xÿ     ä | ;   pbô ä r!   ÿ� ä r   ÿ� Ä rJ   ÿ� â  r(     A  â Er        â  r$     A  â0r_ÿ_    A   â xÿ     ä xÿ     ä xÿ     â #r*   *   â Er        â  r$_     A  âEr        â  r_(_     A  â Er        Ä x--   ÿà÷ â #r1   1   â #r8   8   â #r2_   2   â x//€   ÿàÿ â $r5ÿÿ   5Ž â x44€   ÿàÿ ä x66€   ÿàÿ ä x77€   ÿàÿ â D)       À â Gy   âÿÿÿÿƒ ê Ay        € ê Ey  °     € â G‰  �     € ê $x    +Ž Ê ˆs , üÿ H   è ˆs @€üÿ H   è ˆs B ýÿ H   è ˆs <€ýÿ H   è ˆs 1 þÿ H   è ˆs *€þÿ H   è ˆs 2 ÿÿ H   è ˆs 8€ÿÿ H   ä Ay        € ê {          ì M        € ê „y+ €  H   b z 0 r  p`ð Æ „y+     H   " r +ÿ   p"p Æ „y+ €  H   ¨ „y+ €   H   è „y +    H   ( „y+    H   h „y+ €	  H   h „y+ €  H   h „y+    H   h „y+    H   h „y+ €  H   b !r<       Æ/ „y+    H   b !r,       Æ „y+    H   " !r       ÆO „y	+ €  H   ¢ !r@       Æ� „y+    H   â !r         Æ„y
+ €  H   " !rB       Æ„y+  	  H   b !r       Æ „y+    H   b !r       Æ „y+  
  H   b !r         Æ „y+ €
  H   b !r1       Æ „y+    H   b !r*       Æ „y+ €  H   â !r       Æ/ „y+ €  H   â !r2       Æ „y+ €  H   â !r	       âO !r       ä� ‰       " !r

       Æ‰      b !r       Æ‰	      ¢ !r       Æ ‰      â z  r      â !r       ä y      &   " !r

       Æ ‰      b !r       Æ ‰
      " !r        â !r8       ä ‰      " !r       â/ !r       ä ‰       b !r		       âO !r       â !r       â� ‰      ¨ ‰      â !r       Æ‰	      b !r
       Æ‰      " !r       Æ ‰
      " !r         Æ/ ‰      h ‰      b !r       ÆO ‰  €    ¢ !r       â� !r	       ä‰ €    â !r       Æ‰ €    " !r
       Æ ‰	 €    " !r       Æ/ ‰ €    b !r       Æ ‰ €    b !r        ÆO ‰ €    ¢ $z  x  0Ž â x00   ÿàÿ ä ‰      ¢ !r       â� z 0 r  p`ò ä y      '   â !r       âx +   p"ò Ä ‰ @    " !r
	       â !r       ä/ ‰ @    " !r       Æ‰ @    b !r       ÆO ‰
 @    ¢ !r       Æ ‰ @    b $z  ~   Ž Æ� ‰ @    â $x   ÿŽ â !r       ä‰ @    " 
  r  ÿàõ Æ ‰ €    " !r	       â $ÿÿ   ÿ â 
 j  ÿ„ â‰      " !r       â/ r    ÿ† ä
 k   â !r
       âO ‰	      b    ÿàÿ â !r       â$rÿÿ   ÿŽ â ‰      ¢ z j  ÿ† â!r       â�     ÿàõ â ‰      â z k  � â !r       â    ÿà÷ â ‰      " $ÿÿ   ÿ â !r       â  j  ÿ„ â ‰      b ÿÿ   ÿäÿ â !r       â $tÿ   ÿ Ž â ‰ @    " 

 j  ÿ† Ä %v  j   Ž â ‰      " !r	       â/ 	 k   â !r       âO 
 k  � â †	     â !r       Æ� †    â !r       Æ†	     â !r       Æ†     â !r       Æ †	     â !r       Æ †    è ‰      ( †	 
    â M™        € ê x   ÿàÿ â/ !r       Æ r    ÿàñ Ê $r ÿÿ   ÿ  â z j  ÿ€ È z k     Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y-      "   " ¹z  l      â Ey      € â ™x?    â y       !   b ¹z  F   
   â $r0ÿÿ   ÿ Ž ä y.      %   ¢ $r*ÿÿ   ÿ Ž ä $r4ÿÿ   ÿ Ž Ä $r,ÿÿ   ÿ Ž ä $rDÿÿ   ÿ Ž ä $rFÿÿ   ÿ Ž â r -ÿ   pRð är -ÿ   pRò â$x-     Ž â/ x..   ÿ   ÈO xÿ     Ê $--À    Ž â r   ÿ� È x9ÿ     È | 9   pbô Ú G)  À    € ê y      '   " $rÿÿ   ÿ Ž â xøÿÿÿÿÀŽ â $rÿÿ   ÿ Ž â y      &   b $z	. p  9Ž âx.   ÿàÿ â $x39   ÿ Ž â r0 ÿ       â $v7ÿ X  ÿ Ž ä $v1ÿ Y  ÿ Ž â z23 q  ÿàÿ â 'z y   Ž È 'z s   Ž â/ zÿ z    Æ $z/ }  ÿŽ â zÿ t    â $z |  9Ž ä$z	 |  	Ž ä $z v  Ž ä $x   
Ž ä $z p  Ž Ä $z
 w  ÿŽ â xÿ     â$z	 v  	Ž ä $vÿ q  ÿ Ž â r/
/   ÿàõ â %x	J   Ž È %x	J   ÿŽ È %x
J   Ž È %xJ   ÿŽ È $x   ÿ Ž ä $x+   3Ž ä %x   Ž È $r5ÿÿ   ÿ ä %x   Ž È r7   ÿàõ Ê $x   1 Ê �y>   • ¨ �y=   • è �y8   • ( �y;   • b r7   ÿàõ Ê $x   1 Ê �y@   • h �yB   • h �y?   • h �yA   • b $t(ÿ$   ÿ Ž Ä $x>   ÿ Ž ÊO xø  ÿÀŽ È z  ÿàõ â $x=   ÿ Ž È� $vÿ 	 ÿ â xø  ÿÀŽ â $x8   ÿ Ž Æz  ÿàõ â �y    › ¢ xø  ÿÀŽ ä zÿ 	 ÿä ä z  ÿàõ Æ �y    › â" $x;   ÿ Ž ä$vÿ 	 ÿ Æ xø  ÿÀŽ Æ �y    › " z  ÿàõ Ê $vÿ 	 ÿ Ì �y    › b $x@   ÿ Ž â x?   ÿ   â $xB   ÿ Ž Æ xø  ÿÀŽ ä xø  ÿÀŽ ä z  ÿàõ ä/ z  ÿà÷ Æ $vÿ 	 ÿ ä $vÿ 	 ÿŽ ä $xA   ÿ Ž â xø  ÿÀŽ â �y    › h �y    › b xø  ÿÀŽ Ä z  ÿàù ä z  ÿàû ä r!/   3àµ â $vÿ 	 ÿ ä $vÿ 	 ÿŽ â r"5   ÿd~ â %v ! Z  ( Ž ä �y    › d $x#"$   ÿŽ Ä �y    › d $x!!   #Ž â r$7   ÿàõ È �yC    ™ h �y:    ™ b $x%   1 Æ �y6    ™ h �y<    ™ b r&
7   ÿàõ Æ �yG    ™ h �yE    ™ h �yH$B   ‘ b $x'   1 Æ �yI    ™ b r#/   2àµ Æ �yJ     ™ b r)5   ÿd~ â %v"# Z  ( Ž ä �yK&B   ‘ d $x))$   ÿŽ ä �y      • f r##)   ÿàÿ Ä r$7   ÿàõ Æ/ �yN"   ™ d $x%	   1 ä �yM"   ™ b r)/   +àÅ Æ �yP"   ™ b rT5   ÿ„~ Æ �yL"   ™ h �yS"   ™ h �yR"   ™ b %v() Z  ( Ž Æ �yQ"   ™ b $xUT$   ÿŽ Æ �yO"    ™ b $x))   UŽ Æ �y$$    • h �yU(   ™ h �yT(   ™ h �yY(   ™ h �yX(   ™ h �yW(   ™ h �yV(   ™ h �y!(   ™ h �y%(    ™ b r&7   ÿà÷ Æ �y""    • d# $x'   1Ž Ê �y&&    • h �y'(    • b xZÿ	   =  È 	s= Z        " x[ÿ	   8  ä x>ÿ	   >  Ê 	s\ [        b x8ÿ	   ;  Î 	s] >        b $x==€   ÿ Ž Ê x=Z€   =øŽ ä 	s; 8        " $x\\€   ÿ Ž ä/ $x#=ÿŽ ä $t=ÿÿ Ž â xZ  @€    â $x]]€   ÿ Ž â x[[€   \øŽ Ä x)#€€€€=ìŽ äx\>€   ]øŽ ä x>#€€€€ZìŽ â $t(ÿÿ Ž â xZ)ÿþþþÿàÿ â $x)[ÿŽ â x]>ÿþþþÿàÿ â $x=;€   ÿ Ž ä $t>ÿ @€ÿ Ž â x()€€€€(ìŽ â$x\\ÿŽ â x88€   =øŽ â $t=ÿÿ Ž â x>)€€€€>ìŽ â $t[ÿ @€ÿ Ž â x((ÿþþþÿàÿ Ä x>>ÿþþþÿàÿ ä x=\€€€€=ìŽ ä x(()øŽ äx)> @€)øŽ ä x>\€€€€[ìŽ ä x[=ÿþþþÿàÿ Ä x;Z#øŽ äx[[\øŽ â $x88ÿŽ â x#] @€#øŽ â $t=ÿÿ Ž â x>>ÿþþþÿàÿ â $t]ÿ @€ÿ Ž â x[[˜º  ÿ    ä x=8€€€€=ìŽ Ä x\> @€\øŽ ä x]8€€€€]ìŽ ä x>[ÿÀŽ äx[[€€€€ÞŽ äO x==ÿþþþÿàÿ ä x]]ÿþþþÿàÿ Ä x;;˜º  ÿ    ä r[[>  €ÿàÿ ä x>#˜º  ÿ    ä x==8øŽ äx]] @€8øŽ ä x8;ÿÀŽ Ä x;;€€€€ÞŽ ä� x[€€€€´Ž ä x#>ÿÀŽ äx>>€€€€ÞŽ ä x[(˜º  ÿ    â $x;;   8
Ž â xBÿ	   B  â $x(>   #
Ž â x8[ÿÀŽ Ä x[[€€€€ÞŽ â	s# B        " x>)˜º  ÿ    Æ $x[[   8
Ž â x?ÿ	   ?  ä x((€€€€´Ž ä x[€€€€´Ž ä x)>ÿÀŽ â 	s ?        b x==˜º  ÿ    ä x\\˜º  ÿ    Ä x>>€€€€ÞŽ ä x8=ÿÀŽ â $x[#€   ÿ Ž â xZ\ÿÀŽ â $x>>   )
Ž â x==€€€€ÞŽ äx\\€€€€ÞŽ ä xAÿ	   A  â $x#=   8
Ž â x>€€€€´Ž â $xZ\   Z
Ž â x=B€   [øŽ Ä x@ÿ	   @  ä x>]˜º  ÿ    â 	s) A        " x;€€€€´Ž â $tBÿÿ Ž â xZ€€€€´Ž â $x€   ÿ Ž â/ x;>ÿÀŽ â$x==ÿŽ ä 	s8 @        b x>>€€€€ÞŽ â $tZÿ @€ÿ Ž â x??€   øŽ Ä r>>;  €ÿàÿ ä x##€€€€´Ž ä xB=€€€€BìŽ äx=€€€€ZìŽ ä x>€€€€´Ž â $x?ÿŽ â xBBÿþþþÿàÿ â $t>ÿÿ Ž â xÿþþþÿàÿ Ä x;B=øŽ â$xB)€   ÿ Ž â x= @€=øŽ â $x)8€   ÿ Ž â/ x€€€€>ìŽ äx>€€€€ZìŽ ä xÿþþþÿàÿ ä x>>ÿþþþÿàÿ Ä x8A€   BøŽ ä x)@€   )øŽ ä xøŽ â$t?ÿÿ Ž â x> @€øŽ â $x88ÿŽ ä $x>)ÿŽ â xA  @€    â $t)ÿÿ Ž Ä $t[ÿ @€ÿ Ž â x?8€€€€?ìŽ äxA8€€€€AìŽ ä x)>€€€€)ìŽ äx@>€€€€[ìŽ ä x[?ÿþþþÿàÿ ä x]Aÿþþþÿàÿ Ä x?)ÿþþþÿàÿ ä xA@ÿþþþÿàÿ ä x??>øŽ äxAA @€>øŽ ä x>=˜º  ÿ    ä x)[8øŽ Ä x[?˜º  ÿ    ä xZA˜º  ÿ    ä x=>ÿÀŽ äx>>€€€€ÞŽ ä xA[ÿÀŽ äxB[€€€€ÞŽ Ä x@ZÿÀŽ äx[Z€€€€ÞŽ â $x>>   =
Ž È $x@[   @
Ž â x>>€€€€´Ž ä x˜º  ÿ    ä x@€€€€´Ž ä x8] @€8øŽ ä x;;˜º  ÿ    ä x@˜º  ÿ    Ä x)˜º  ÿ    â $xAB   A
Ž â xÿÀŽ äx€€€€ÞŽ ä x?;ÿÀŽ äxZ;€€€€ÞŽ ä xB8˜º  ÿ    Ä x;ÿÀŽ äx8€€€€ÞŽ â $x   
Ž â xA€€€€´Ž Æ $x8   ;
Ž â x€€€€´Ž â &rC   ÿ   âx=BÿÀŽ â &rCC   ÿ   â x€€€€´Ž â &r:   ÿ   â xBB€€€€ÞŽ â &r6      âx)@ÿÀŽ â&rC6   C   â x@@€€€€ÞŽ â &r6<      Ä $xB   =
Ž â r@@)  €ÿàÿ â &rG      â xHð   ÿÀŽ â &r)#E   6   â x€€€€´Ž â $x?Z   ?
Ž â xHH   ÿÀŽ â &r(I      ä &r)J   )   â x@€€€€´Ž â &r::   ÿ   â x?€€€€´Ž â $rH   ÿŽ â xÿ     â $x   )Ž Ä &r:<   :   ä &rCG   C   â r6   ÿ� â $r))   Ž â x8Kð   ÿÀŽ â &rEE   :   â xKK   ÿÀŽ â &r:>I   C   â r)6)   ÿú� â &rEJ   E   â xÿ   8  â $rK:   ÿŽ â x6ÿ   )  â $x::   EŽ Ä $rEE   Ž â r6)   ÿ� â &r)N   ÿ   ä&rNN   ÿ   ä &r6M   ÿ   âr:::   ÿ� â &r)P   )   ä &rMM   ÿ   Ä &rNP   N   ä &r6L   6   âr::E   ÿú� â &r)S   )   ä&rML   M   ä &rNS   N   ä &r6#R   6   Ä &r;(Q   )   âx)ÿ   :  â &rMR   M   ä &rN>Q   N   ä &r6O   6   âr:):   ÿ� â &rMO   M   ä $r=H;   ÿŽ Ä $r)KN   ÿŽ ä $x;;   6Ž ä $xNN   MŽ ä $rMM   )Ž ä $r=6   =Ž ä 0r)ÿ     A   â r ;;   ÿ� â &rU   ÿ   â xÿ     â &rT   ÿ   Ä &rU   ÿ   ä &rT   ÿ   ä 0r$ÿ$    A   â r= =   ÿú� â &rY      â Er         â &rX      ä &rY      Ä &rX      â  r$)     A  â &rW      ä&r#V      ä&rW      ä &rV      â #rF    F   â x ÿ   =  â &r(!      Ä &r%      ä &r>!      ä &r%      â rNNN   ÿ� â $rHH   ÿŽ â r  =   ÿ� â $x   Ž ä $rKK   ÿŽ â r   ÿàÿ â 0rÿ"    A   â rMNM   ÿú� â $rH   HŽ â x ÿ      â $rK   KŽ â r   ÿ� Ä r   ÿ� â  r$     A  â xÿ   M  ä Er          ä rK   ÿú� ä rH   ÿú� â #r,    ,   â rM   ÿ� Ä x99   ÿàÿ ä xÿ     ä xÿ     â 0r&ÿ&    A   â xÿ     ä | 9   pbô ä r   ÿ� Ä r   ÿ� â  r&     A  âEr        â 0r'ÿ'    A   â x:ÿ   :  ä xÿ     ä xÿ     â #r4   4   â Er: :       â  r&'     A  â Er        â  r))&     A  â Er        â  r'$'     A  â x77   ÿà÷ â #rD):   D   â #r0   0   â #r*'   *   â x++€   ÿàÿ â $r1ÿÿ   1Ž â x22€   ÿàÿ Ä x33€   ÿàÿ â D)       À â Gy  €åÿÿÿÿƒ ê Ay        € ê ˆ -F ýÿ H   è ˆ -D€ýÿ H   è ˆ -, þÿ H   è ˆ -4€þÿ H   è ˆ -* ÿÿ H   è ˆ -0€ÿÿ H   è {          ì M        € ê „y      H   " z . r  p`ð Æ „y     H   b r  ÿ   p"p Æ „y  €   H   ¨ „y     H   è „y  €  H   ( „y     H   h „y     H   h „y  €  H   h „y     H   h „y  €  H   h „y  €  H   b !rF       Æ „y
     H   " !r       Æ/ „y	  €  H   b !rD       ÆO „y     H   ¢ !r,       Æ� „y  €  H   â !r4       Æ„y  €  H   â !r       Æ„y     H   â !r*       Æ „y  €  H   â !r       â !r       ä ‰      " !r       Æ y      '   b !r0       â !r

       â $vÿ r  ÿ Ž â !r	       â/ !r       ÈO ‰      " !r       Æ� ‰      b !r       â!r       ä ‰	      ¢ !r

       Æ ‰      â !r       Æ ‰
      ( ‰      b !r       Æ y      &   " !r       Æ/ ‰      b !r		       ÆO ‰      ¢ !r       Æ� ‰	      â !r
       Æ‰ €    " !r       Æ‰
      h ‰      " !r       â/ !r       ÈO ‰ €    b !r	       Æ� ‰ €    ¢ !r       Æ‰	 €    â !r

       Æ‰ @    " !r       Æ ‰
 €    ( ‰ €    b !r       â/ !r       ÈO ‰ @    b !r		       Æ� ‰ @    ¢ !r       Æ‰	 @    â !r
       â x
.   ÿàÿ ä ‰      " $z. x  .Ž â !r       âz 
 r  p`ò â ‰ @    " $z. ~  .Ž ä x     p"ò â ‰ @    d r.   ÿ† â !r       â/ 
. r  ÿàõ â !r
       ÆO ‰       b $ÿÿ   ÿ â 
 j  ÿ„ â!r	       â� 	   ÿàÿ ä 
 k   ä .	   ÿàõ â !r       â $tÿ   ÿ Ž â !r       â$r	ÿÿ   ÿŽ â ‰
      " z j  ÿ† â !r       â$ÿÿ   ÿ â ‰      ¢  j  ÿ„ â %v. j   Ž Ä ‰      â !r        â/ z k  	� ä ‰      b 	 k   Æ †	     è	 †    â	 !r
       â !r       ÈO †	     â	 !r       Æ� †     â	 !r       Æ/ †	     â	 M™        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y-      "   " ¹z  l      â Ey  €    € â ™x?    â y       !   b ¹z  F   
   â $r,ÿÿ   ÿ Ž ä y.      %   ¢ $r0ÿÿ   ÿ Ž ä $r*ÿÿ   ÿ Ž Ä $r4ÿÿ   ÿ Ž â r -ÿ   pRð är -ÿ   pRò â$x-     Ž ä/ $x..   ÿ Ž ÆO xÿ     È r   ÿ� â $--€    Ž Æ x7ÿ     È | 7   pbô Ú G)  P    € ê y      '   " $rÿÿ   ÿ Ž â xøÿÿÿÿÀŽ â $rÿÿ   ÿ Ž â y      &   b $x27   ÿ Ž ä $v5ÿ X  ÿ Ž ä $v1ÿ Y  ÿ Ž â z+2 q  ÿàÿ â $r,ÿÿ   ÿ Ž Ä 'z y   Ž È 'z s   Ž â/ zÿ z    Æ $z/ }  ÿŽ â zÿ t    â $z
 w  ÿŽ ä $z. p  7Ž ä$z |  7Ž â r/
/   ÿàõ â $x   
Ž â x.   ÿàÿ â $z |  Ž Ä $z v  Ž âxÿ     â $z v  Ž ä $z p  Ž ä %xJ   ÿŽ È %x
J   Ž È %xJ   ÿŽ È $x   ÿ Ž ä %xJ   Ž È %x   Ž È $r3ÿÿ   ÿ ä %x   Ž È r5   ÿàõ Ê $x   1 Ê �yO   • ¨ �yN   • è �yM   • ( �yP   • b	 r5   ÿàõ Ê $x   1 Ê �y?   • h �yI   • h �yB   • h �y>   • b $t&ÿ$   ÿ Ž Ä $xO   ÿ Ž ÊO xø  ÿÀŽ È z  ÿàõ â $xN   ÿ Ž È� $vÿ 	 ÿ â xø  ÿÀŽ â $xM   ÿ Ž Æz   ÿàõ â �y    › ¢ xø  ÿÀŽ Æ $v!ÿ 	 ÿ â z  ÿàõ Ê �y      › â $vÿ 	 ÿ ä $xP   ÿ Ž È�y    › " xø  ÿÀŽ È z  ÿàõ Ê $vÿ 	 ÿ Ì �y    › b $x?   ÿ Ž Ê xø  ÿÀŽ â $xI   ÿ Ž Æ z  ÿàõ ä xø  ÿÀŽ Æ $vÿ 	 ÿ â z  ÿàõ â $xB   ÿ Ž È �y    › b $vÿ 	 ÿ â xø  ÿÀŽ â $x>   ÿ Ž â r/   2àµ ä z  ÿàù â �y    › b xø  ÿÀŽ â %v Z  & Ž â r3   ÿd~ Ä z  ÿàõ â $vÿ 	 ÿ ä $x#$   ÿŽ ä $vÿ 	 ÿ ä �y    › b $x   #Ž Æ �y    › b r"/   +àµ Æ �y6   ™ b r%3   ÿd~ ä r$5   ÿàõ â �y8   ™ b %v"" Z  & Ž Æ �y=   ™ b $x'%$   ÿŽ ä $x%   1 â �y<   ™ b $x##   'Ž Æ �y9   ™ h �y:   ™ h �y@$B   ‘ h �y;   ™ h �yA    ™ h �yD"   ™ h �yF"   ™ h �yG"   ™ b r$
5   ÿàõ Æ/ �yK"   ™ d $x%   1 ä �yJ"   ™ h �yC"    ™ h �yH"   ™ h �y$$B   ‘ h �yE"   ™ b r&5   ÿàõ Æ �y    • d $x'	   1 âr(5   ÿàõ â �yL"    • h �y&&    • b $x)   1 Ê �y((    • b xNÿ	   N  È 	sQ N        b xMÿ	   M  ä xOÿ	   O  ä x'ÿ	   P  Æ 	sR M        0 	s% O        b. $xQQ€   ÿ Ž Ê xQN€   QøŽ ä 	sS '        b $tNÿ @€ÿ Ž ä $xQQÿŽ ä $xRR€   ÿ Ž Æ x#Q€€€€NìŽ â $xN%€   ÿ Ž â/ xRM€   RøŽ È xNO€   NøŽ â $xTS€   ÿ Ž ä $xRRÿŽ ä $t)ÿ @€ÿ Ž â xT'€   TøŽ â $xNNÿŽ â xP#ÿþþþÿàÿ â $t%ÿ @€ÿ Ž ä $t'ÿÿ Ž Ä $t#ÿÿ Ž â x)R€€€€)ìŽ â $t"ÿÿ Ž â x%N€€€€%ìŽ ä x'R€€€€'ìŽ ä x#N€€€€#ìŽ ä xO)ÿþþþÿàÿ ä x)%ÿþþþÿàÿ Ä xM'ÿþþþÿàÿ ä xQ€€€€"ìŽ ä x'#ÿþþþÿàÿ ä x)) @€NøŽ â $xTTÿŽ â x#MRøŽ â $tMÿÿ Ž â x"ÿþþþÿàÿ Ä x''NøŽ ä xP @€QøŽ ä xP)˜º  ÿ    ä x''˜º  ÿ    ä xMT€€€€MìŽ ä x)PÿÀŽ Ä x%O @€RøŽ ä xN'ÿÀŽ ä xOMÿþþþÿàÿ ä x""QøŽ ä xIÿ	   I  ä x##˜º  ÿ    â $tQÿ @€ÿ Ž â xBÿ	   B  Ä xPP€€€€ÞŽ äO x''€€€€ÞŽ Æ $xPP   )
Ž â x)OTøŽ â $xM'   N
Ž â xO"˜º  ÿ    ä xN˜º  ÿ    ä x"OÿÀŽ äxOO€€€€ ÞŽ ä� xNÿÀŽ Ä xNN€€€€!ÞŽ ä xM€€€€´Ž ä xP€€€€´Ž â 	sM I        " xP%˜º  ÿ    â $xOO   "
Ž ä $x"N   
Ž â xN#ÿÀŽ Ä x%PÿÀŽ äx##€€€€ÞŽ äxPP€€€€ÞŽ ä xQT€€€€QìŽ â $x#   N
Ž ä $xPP   %
Ž â xQQÿþþþÿàÿ Ä x€€€€´Ž ä xP€€€€´Ž â 	s B        b $xPM€   ÿ Ž â x))˜º  ÿ    ä x'Q @€TøŽ ä x?ÿ	   ?  ä xN)ÿÀŽ Ä xII€   PøŽ ä x))€€€€ÞŽ äxP'˜º  ÿ    ä x""€€€€!´Ž ä 	s! ?        " x>ÿ	   >  â $x))   N
Ž â x#PÿÀŽ â$xIIÿŽ â xPP€€€€ÞŽ â $tNÿÿ Ž Ä 	sM >        ¢ $x'€   ÿ Ž ä/ $xPP   #
Ž â x#I€€€€NìŽ ä x'B€   'øŽ ä xN#ÿþþþÿàÿ â $xB!€   ÿ Ž â xP€€€€´Ž â $x''ÿŽ â x#NIøŽ â $tNÿÿ Ž Ä $tPÿ @€ÿ Ž â xB?€   BøŽ â $xMM€   ÿ Ž âO x'€€€€NìŽ äx!'€€€€PìŽ â $xBBÿŽ â x)€€€€´Ž â $t)ÿÿ Ž â xM>€   MøŽ ä x>ÿþþþÿàÿ Ä xN!ÿþþþÿàÿ â $t?ÿ @€ÿ Ž â x!>'øŽ äxN @€'øŽ ä x'B€€€€)ìŽ â$xMMÿŽ â x)B€€€€?ìŽ â $t>ÿÿ Ž ä $tNÿ @€ÿ Ž â x''ÿþþþÿàÿ â $tRÿ @€ÿ Ž â x))ÿþþþÿàÿ Ä x>M€€€€>ìŽ äx?M€€€€NìŽ ä x%I€€€€RìŽ ä x''BøŽ äx)) @€BøŽ ä x>>ÿþþþÿàÿ Ä xN?ÿþþþÿàÿ ä xR%ÿþþþÿàÿ ä x?'˜º  ÿ    ä xB)˜º  ÿ    ä x)>MøŽ ä x%R @€IøŽ Ä x>?ÿÀŽ äx??€€€€ÞŽ ä x'N @€MøŽ ä x##˜º  ÿ    ä xN%˜º  ÿ    â $x??   >
Ž â xIBÿÀŽ Ä xBB€€€€ÞŽ ä x>#ÿÀŽ äx%NÿÀŽ äx##€€€€ÞŽ ä xNN€€€€ÞŽ â $xBB   I
Ž â x!!˜º  ÿ    â $x##   >
Ž Ä $xNN   %
Ž â xB€€€€´Ž ä x>!ÿÀŽ ä xB˜º  ÿ    ä x!!€€€€ÞŽ ä x#€€€€´Ž ä xN€€€€´Ž Ä x))˜º  ÿ    ä xN'˜º  ÿ    â $x!   >
Ž â x>)ÿÀŽ äx!NÿÀŽ äx))€€€€ÞŽ ä xNN€€€€ÞŽ Æ $x))   >
Ž â x#BÿÀŽ â $xNN   !
Ž â xBB€€€€ÞŽ ä x€€€€´Ž ä x)€€€€´Ž â &r>6   ÿ   â xN€€€€´Ž â &r8   ÿ   Ä $xBB   #
Ž â x O€€€€ ´Ž â &r>=   >   ä &r<      â xB€€€€´Ž â &r88   ÿ   ä &r> 9   >   ä &r#:      â x@ð   ÿÀŽ â&r%<   8   â x@@   ÿÀŽ â &r!";   >   Ä &r8A   #   ä &r#D   ÿ   â xÿ     â $r@!   ÿŽ ä &r#F   #   ä $x!!   8Ž â x?€€€€´Ž â $r8   Ž Ä &r#G   #   â r!!   ÿ� â &rK   ÿ   ä &r66   ÿ   â r   ÿú� â &rK   ÿ   ä &rD   ÿ   ä &rJ      Ä &r6=   6   ä&rC   #   â xÿ     â &rJ      ä &rF      â r   ÿ� â &r69   6   ä &r H      â x$ð   ÿÀŽ â &r%:   %   Ä &rH      â x$$   ÿÀŽ â &rG      ä &r;;   6   ä &r"E      ä &r6A   %   ä &rEE      Ä &rC      â xÿ     â $r$;   ÿŽ ä $r@@   ÿŽ ä $r$$E   ÿŽ ä $x   Ž ä $x;;   6Ž Ä $xEE   Ž â r   ÿ� â $r@   @Ž ä $r6   Ž â rEEE   ÿ� â $r$   $Ž â r;;   ÿ� ä r@   ÿú� ä r   ÿú� Ä r$E$   ÿú� ä x77   ÿàÿ ä xÿ     ä xÿ     ä xÿ   $  â 0r&ÿ&    A   â xÿ     â 0rÿ    A   â | 7   pbô Ä r   ÿ� ä r   ÿ� ä r$   ÿ� â 0rÿL    A   â Er        â  r&     A  â 0r(ÿ(    A   â xÿ     ä xÿ     Ä xÿ     â #r4   4   â  r&     A  â Er        â  r(     A  â Er        â  r(     A  â Er        ä x55   ÿà÷ â #r0   0   â #r*   *   â #r,   ,   â x++€   ÿàÿ â $r1ÿÿ   1Ž â x22€   ÿàÿ â D)       À â Gy  Ðèÿÿÿÿƒ ê Ay        € ê ˆ -4 þÿ H   è ˆ -*€þÿ H   è ˆ -0 ÿÿ H   è ˆ -,€ÿÿ H   è {          ì M        € ê „y      H   " z . r  p`ð Æ „y  €   H   b r  ÿ   p"p Æ „y     H   ¨ „y     H   è „y  €  H   ( „y     H   h „y     H   h „y	  €  H   h „y     H   h „y  €  H   h „y
  €  H   b !r4       Æ „y  €  H   " !r*       â/ !r0       âO !r       ä� !r       â!r       â!r       ä !r		       Æ ‰      b !r       Æ ‰	      ¢ !r,       Æ ‰      â !r

       Æ y      &   " !r
       â !r       È/ ‰      " !r	       ÆO ‰      b !r       Æ� ‰      ¨ ‰      â !r       â !r       È/ ‰      " !r       ÆO ‰ €    b !r       Æ� ‰ €    ¨ ‰
 €    â !r       â !r       È/ ‰ €    " x.   ÿàÿ â $z. x  .Ž â!r       äO ‰ @    b z  r  p`ò â !r

       ä� ‰ @    ¢ x     p"ò Æ ‰	
 @    è y      '   " !r       â !r       â/ !r       ÈO ‰      " !r	
	       Æ� ‰      b $z. ~  .Ž Æ‰
 @    ¢ $tÿ   ÿ Ž â 
 . r  ÿàõ ä ‰	      æ $ÿÿ   ÿ â 
  j  ÿ„ È 
  k   â !r       â %v. j   Ž â !r       È/ †	     â !r

       ÆO †    â !r		       Æ� ‰
      h  †	 	    â M™        € ê ¹z  r      â !r
       â/ �x   ?àÿ Ì |..   ÿàñ Ê $rÿÿ   ÿ  â z. j  ÿ€ È z. k    Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y
      &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž ä $ÿÿ   
 Ž È % \   Ž Ê �	    ™ b $ÿÿ   ÿ Ž â ¹z  l      â Ey  ð    € â y	      "   b '
 m   Ž â ™x?    ä y       !   ¢ $rÿÿ   ÿ Ž â 
ÿ n    â $‚ÿÿ   ÿ Ž È $ÿÿ   
Ž È $
 o  Ž ä $‚ÿÿ   
 Ž È $‚ÿÿ    Ž ä $‚ÿÿ    Ž â r 	ÿ   pRò â/ 'Š s   Ž â r 	ÿ   pRô Æ $x	     Ž ÊO xÿ     È r   ÿ� È xÿ     È |    pbö ä Šÿ t    Ö G9  €    € ê y      '   " $rÿÿ   ÿ Ž â xøÿÿÿÿÀŽ â $z w  ÿŽ â y
      %   b $x   ÿ Ž ä $x   
Ž Ê xÿ     â 'z y   Ž È $z }  ÿŽ â zÿ z    È r   ÿàý â $z |  Ž È $z v  Ž ä$rÿÿ   ÿ ä $z
 p  Ž È/ %xJ   ÿŽ È $x   ÿ Ž â z X  àÙ äz X  ÿà÷ â %x   Ž â z Y  ¤~ ä xB   ÿàû ä z X  ÿàñ ä z Y  ÿäÿ â $rÿÿ   ÿ Ž â x   ÿàù â $rÿÿ   Ž Æ zÿ Y  ~ Ê �yþÿÿ • ¨ �y!    • è �y   • ( �y   • b $tÿ$   ÿ Ž Ä $rÿÿ    Ž ä $rÿÿ    Ž Ê �y)    ‘ $ $rÿÿ    Ž ä $rÿÿ    Ž Ê �y*    • " $x   ÿ Ž ÊO xø  ÿÀŽ È z  ÿàñ Ê $vÿ 	 ÿ  â r   à± Ê �y    › ¢ $x!   ÿ Ž â� r	   ÿd~  â %v Z   Ž Æ xø  ÿÀŽ â $x		$   ÿŽ Æ z  ÿàñ â $x   	Ž È $v	ÿ 	 ÿ  â �y    ™ è �y"   ™ è �y    › â $x
   ÿ Ž Ä$x   ÿ Ž â�y'   ™ $ x

ø  ÿÀŽ ä �y&   ™ d z

  ÿàñ ä xø  ÿÀŽ â �y%   ™ $ $vÿ 	 ÿ  â z  ÿàñ â �y$   ™ ( �y

    › " $vÿ 	 ÿ  Æ �y#   ™ ( �y    › ( �y(    ™ ( �y+    • " x-ÿ	     È 	s -        b x!ÿ	   !  â $tÿÿ Ž Ì 	s, !        " $x€   ÿ Ž Ê/ x-€   øŽ Ê $xÿŽ ä $tÿ @€ÿ Ž Æ x€€€€ìŽ È xÿþþþÿàÿ ä x€€€€ìŽ â $x,,€   ÿ Ž â xøŽ äx-ÿþþþÿàÿ ä x˜º  ÿ    ä x-- @€øŽ Ä x!€   ,øŽ ä xÿÀŽ ä xÿ	     â $xÿŽ â x,-˜º  ÿ    â $t-ÿÿ Ž â xÿ	     ä x,ÿÀŽ Ä x-€€€€-ìŽ È x--ÿþþþÿàÿ È x--øŽ È x--˜º  ÿ    ä x   ÿàÿ È |    pbð â 0r*ÿ*    A   â x   ÿàù ä x   ÿà÷ ä x   ÿàû â $rÿÿ    â x€   ÿàÿ â $rÿÿ   Ž ä $rÿÿ   Ž â x€€€€ÞŽ âO $tÿ @€ÿ Ž â x!,€€€€ÞŽ Æ $x   
Ž ä 	s         " $x!   
Ž â x€€€€ìŽ Ì 	s!         b xÿþþþÿàÿ ä x€€€€´Ž ä x @€øŽ â $x€   ÿ Ž â x€€€€´Ž â &r    ÿ   â� x-ÿÀŽ äx--€€€€ÞŽ Ä x˜º  ÿ    ä x€   øŽ â &r""      â xÿÀŽ â$x-   
Ž â x€€€€	ÞŽ â $x!!€   ÿ Ž ä/ $xÿŽ ä $tÿÿ Ž â x!€   !øŽ â $tÿ @€ÿ Ž Ä $x   
Ž â x€€€€ìŽ â$tÿÿ Ž â x€€€€´Ž â $x!ÿŽ â x€€€€ìŽ ä xÿþþþÿàÿ â $tÿ @€ÿ Ž â xÿþþþÿàÿ ä xøŽ Ä x€€€€ìŽ äx€€€€ìŽ ä x @€øŽ ä x˜º  ÿ    ä xÿþþþÿàÿ ä xÿþþþÿàÿ Ä x˜º  ÿ    ä xÿÀŽ äxøŽ äx€€€€
ÞŽ äx @€øŽ ä xÿÀŽ Ä x˜º  ÿ    â $x   
Ž â x€€€€ÞŽ ä x˜º  ÿ    ä xÿÀŽ â$x   
Ž â x€€€€ÞŽ ä x
€€€€
´Ž Ä xÿÀŽ â$x   
Ž â x€€€€ÞŽ â &r

'   ÿ   â x€€€€´Ž â &r%   "   â x€€€€	´Ž â $x   
Ž â x€€€€´Ž â &r&   
   ä&r#      â x€€€€´Ž â &r$      â x)ð   ÿÀŽ Ä x))   ÿÀŽ â &r(      â xÿ     Æ $r))   ÿŽ ä $x   Ž ä $r   )Ž Æ r   ÿ� È r   ÿú� È xÿ     È r   ÿ� â 0r+ÿ+    A   Æ xÿ     ä  r**+     A  ä Er        Ê #r*      â G‰  Põÿÿÿÿƒ ê Ay        € ê ˆ €ÿÿ H   è {          ì M)        € ê „y      H   " r  ÿ   pRð Æ „y  €   H   h „y     H   ¢ !r       È !r       È/ !r       ÊO ‰      $ !r       Ê ‰      $ !r       Ê ‰ €    $ !r       Ê ‰ @    $!r       Ê ‰      b  M	        € ê y       %   ¢ $tÿ   ÿ Ž â !r       ä/ y      '   b $z x   Ž ÈO $z ~  Ž È/ %v j   Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž È % \   Ž Ê �	    ™ b $rÿÿ   ÿ Ž â Ey  ð    € ä y      "   b 'Š  s   Ž Æ y      !   b '
	 m   Ž È $rÿÿ   ÿ Ž â 
	ÿ n  	  â $x    Ž â/ rÿ   ÿü‚ ä y      %   $ xÿ     È r   ÿ� ä Šÿ t     â $r ÿÿ   ÿ Ž ä G       € ê z ÿ `  pPô ä z ÿ ^  pPò ä z ÿ b  p ö ä z ÿ a   Sô Ä z ÿ _  Sò ä z ÿ c  0't Ö       '   "         ä¢       â ©      '   d $  x  ÿŽ ä $ª x  ÿŽ ä $ ~  ÿŽ Ä $ª
 ~  ÿŽ â/ ÿ     ä    àë ä¢

   àÇ ä¨ÿ     ä ÿÿ   Äþ ä      ÿàý Ä ¢
   ÿàû ä ¢ÿÿ   „þ â $ÿÿ    â 
  ^  ÿ† â$rÿÿ   ÿ Ž â ª b  ÿˆ â $¢ÿÿ   Ž â   _  � â $r ÿÿ   ÿ Ž Æ ª c   â �
     h �©      d Ay        € ê ¹z  l      â xÿ     â ™x?    â $	ÿÿ   	
Ž â Ey  à    € â $rÿÿ   ÿ Ž Ä $
		 o  Ž â|    pbò â $‚	ÿÿ    Ž ä $rÿÿ   ÿ Ž Ô G  p    € ê y      '   " $rÿÿ   ÿ Ž ä $x   ÿ Ž ä 'z y   Ž Ê zÿ z    Ê $z |  Ž â xøÿÿÿÿÀŽ Æ $z v  Ž ä$z }  ÿŽ ä $z p  Ž ä $z	 w  ÿŽ ä %xJ   ÿŽ È $x   
Ž â r   ÿàñ â $rÿÿ   	 Ž ä $rÿÿ   ÿ Ž â z	 X  ÿàó â $rÿÿ   ÿ Ž â x
ÿ     â $rÿÿ   ÿ  Æ z
 Y  ÿäÿ  ä $x.   ÿ Ž ä $tÿ   ÿ Ž È %v. X  Ž Ê r   ÿàñ Ê $x     Ê �y*   • ¨ �y+   • è �y#   • ( �y,   • b $t%ÿ$   ÿ Ž Ä $x*   ÿ Ž ÊO xø  ÿÀŽ È z  ÿàñ â $x+   ÿ Ž È� $vÿ 	 ÿ  â xø  ÿÀŽ â $x#   ÿ Ž Æz  ÿàñ â �y    › ¢ xø  ÿÀŽ Æ $vÿ 	 ÿ  â z  ÿàõ â $x,   ÿ Ž âr$   à‘ Æ �y    › â $vÿ 	 ÿ â xø  ÿÀŽ ä r
   ÿ$~  Æ �y    › " z  ÿàñ â %v$$ Z  % Ž È $vÿ 	 ÿ  ä $x$   ÿŽ È �y    › b $x%%   Ž â r(	   ÿàñ È �y$   ™ h �y$   ™ b $x)     Æ �y$   ™ h �y$   ™ h �y$   ™ h �y$   ™ h �y (B   ‘ h �y$   ™ h �y$    ™ b z& X  ÿàñ Æ �y"$    • b z' Y  ÿä  Ê �y!&    • b x-ÿ	   +  ä x+ÿ	   #  ä 	s0 -        " x*ÿ	   *  Î 	s/ +        " x,ÿ	   ,  Î 	s( *        b. $x00€   ÿ Ž Î 	s# ,        " x0-€   0øŽ â $x&/€   ÿ Ž ä $t%ÿÿ Ž ä $x$0ÿŽ â x&+€   &øŽ â $x)(€   ÿ Ž ä/ $t'ÿ @€ÿ Ž â x($€€€€%ìŽ â $x&&ÿŽ â x0*€   )øŽ â $t)ÿ @€ÿ Ž â x+$€€€€'ìŽ â $x'#€   ÿ Ž â x#(ÿþþþÿàÿ Ä x++ÿþþþÿàÿ ä x)&€€€€)ìŽ ä x##$øŽ â$x00ÿŽ â x$+ @€$øŽ â $t+ÿÿ Ž â x-)ÿþþþÿàÿ â $t)ÿ @€ÿ Ž Æ x+0€€€€+ìŽ äx)0€€€€)ìŽ ä x++ÿþþþÿàÿ ä x))ÿþþþÿàÿ ä x%&€€€€%ìŽ ä x++0øŽ Ä x',€   'øŽ ä x0) @€0øŽ ä x%%ÿþþþÿàÿ ä x++˜º  ÿ    â $x''ÿŽ â x00˜º  ÿ    â $t(ÿÿ Ž â x%%&øŽ â$t,ÿ @€ÿ Ž â x&- @€&øŽ Ä x)+ÿÀŽ ä x('€€€€(ìŽ äx,'€€€€,ìŽ ä x((ÿþþþÿàÿ ä x,,ÿþþþÿàÿ ä x##˜º  ÿ    Ä x$$˜º  ÿ    ä x(('øŽ äx,, @€'øŽ ä x'#ÿÀŽ ä x%%˜º  ÿ    ä x&&˜º  ÿ    Ä x,,˜º  ÿ    ä z ÿ `  pPð È z ÿ a   Sð ä x   ÿàÿ ä x*+€€€€ÞŽ äO x+0ÿÀŽ äx-0€€€€ÞŽ â $x)*   )
Ž È $x*-   +
Ž â x))€€€€´Ž ä x$ÿÀŽ ä x*€€€€´Ž ä x*#€€€€ÞŽ ä� x+$€€€€ÞŽ Æ $x#*   '
Ž â x'%€€€€ÞŽ â$x$+   
Ž â x%ÿÀŽ ä x%&ÿÀŽ äx&&€€€€ÞŽ â $x''   
Ž È $x&   %
Ž â x%(˜º  ÿ    ä x##€€€€´Ž ä x%ÿÀŽ äx%%€€€€ÞŽ äx'€€€€´Ž ä x€€€€´Ž Ä x,ÿÀŽ â$x%%   
Ž â x,,€€€€ÞŽ â &r))   ÿ   ä &r   ÿ   â x%%€€€€´Ž â $x,,   
Ž ä &r   )   ä &r      â x$€€€€´Ž â &r#      â x,,€€€€´Ž â &r%%      â x ð   ÿÀŽ â&r      â x     ÿÀŽ â &r%,   %   â xÿ     Æ $r     ÿŽ ä $x   %Ž ä $r%    Ž Æ r   ÿ� È r   ÿú� È xÿ     È r   ÿ� â 0r$ÿ!    A   ä 0r'ÿ"    A   â x%ÿ     â G‰  0	    € ê %v.. `  Ž Ê r.   ÿàñ Ê $x/     Ê �y    ¨ �y)    è �y(    ( �y&    b $x   ÿ Ž ÊO xø  ÿÀŽ È z  ÿàñ Ê $vÿ 	 ÿ  Ì �y    › ¢ $x)   ÿ Ž ä� $x(   ÿ Ž Æxø  ÿÀŽ È z  ÿàñ ä xø  ÿÀŽ Æ $vÿ 	 ÿ  â z  ÿàñ â $x&   ÿ Ž È�y    › â xø  ÿÀŽ â $vÿ 	 ÿ  Æ z  ÿàñ Æ �y    › $ $vÿ 	 ÿ  Ì �y    › b z" `  à‘ È z# a  
$~  Ê �y""B    â  z  `  ÿàñ È z! a  ÿä  Ê �y       " x,ÿ	     â $t*ÿÿ Ž â x(ÿ	   (  â $t.ÿ @€ÿ Ž â 	s ,        " x)ÿ	   )  â $x€   ÿ Ž Ê x+,€   øŽ ä 	s (        & $x++ÿŽ Ê x!+€€€€*ìŽ ä/x#+€€€€.ìŽ â 	s* )        b x,!ÿþþþÿàÿ ä x!ÿ	   &  ä x&#ÿþþþÿàÿ â $x€   ÿ Ž â x,,+øŽ â	s# !        " x&& @€+øŽ Ä x-,˜º  ÿ    ä x.&˜º  ÿ    ä x+-ÿÀŽ â $x**€   ÿ Ž â/ x&.ÿÀŽ È x*)€   *øŽ â $t)ÿ @€ÿ Ž â x,-€€€€ÞŽ äO x-.€€€€ÞŽ Æ $x+,   +
Ž ä $x,#€   ÿ Ž â x#(€   øŽ â $x&-   &
Ž â x+€€€€´Ž â $x(*ÿŽ â x+!€   ,øŽ â $x!#ÿŽ â x&€€€€´Ž â $tÿÿ Ž â x)(€€€€)ìŽ â $t-ÿÿ Ž Ä $t*ÿ @€ÿ Ž â x!€€€€ìŽ â $x++ÿŽ â x#(€€€€-ìŽ ä x*!€€€€*ìŽ ä xÿþþþÿàÿ ä x##ÿþþþÿàÿ ä x**ÿþþþÿàÿ Ä x))ÿþþþÿàÿ ä x!øŽ äx##(øŽ äx!* @€!øŽ â $t*ÿÿ Ž â x)) @€(øŽ â $t(ÿ @€ÿ Ž â x-#˜º  ÿ    Ä x*+€€€€*ìŽ äx(+€€€€(ìŽ ä x#-ÿÀŽ äx.-€€€€ÞŽ ä� x**ÿþþþÿàÿ ä x((ÿþþþÿàÿ Ä x,)˜º  ÿ    â $x).   #
Ž â x**+øŽ â0r ÿ     A   âx(( @€+øŽ ä x+˜º  ÿ    ä x#,ÿÀŽ â  r '      A  â x.!˜º  ÿ    Ä x,,€€€€ÞŽ ä x))€€€€´Ž ä x+ÿÀŽ â$x,,   #
Ž â x++€€€€ÞŽ â &r))   ÿ   â x.ÿÀŽ äx!.€€€€ÞŽ â $x++   
Ž â x#*˜º  ÿ    Ä x((˜º  ÿ    â $x!   
Ž â x#ÿÀŽ äx##€€€€ÞŽ äx+€€€€´Ž ä x€€€€´Ž â $x##   
Ž â x(ÿÀŽ â &r   ÿ   â x((€€€€ÞŽ Ä x,€€€€´Ž â &r      â x##€€€€´Ž â $x((   
Ž â x"   ÿÀŽ â &r   )   â x"ÿ   "  â &r#      â x((€€€€´Ž â &r      È &r      ä &r(      ä $r   ÿŽ ä $x   Ž ä $r"   Ž Æ r   ÿ� È r   ÿú� È xÿ     È r   ÿ� È xÿ     È Er        Ê #r       â |    pbð â  r'$'     A  â Er% %       ä x   ÿàó ä x€   ÿàÿ â #r'%      ä $rÿÿ   Ž  È D	       À â Gy  àëÿÿÿÿƒ ê Ay        € ê y      "   " Ey  °     € â $r
ÿÿ    Ž â r ÿ   pRð är ÿ   pRò Ö G‰  `     € ê y      !   " z ÿ `  pPð È z ÿ a   Sð â $x    Ž Ê ˆs €ÿÿ H   î ˆ 
   H   ä Ay        € ê {          ì M        € ê y	      !   b z ÿ `  pPð È z ÿ a   Sð â „y	     H   X. „‰	 €   H   ( „‰	    H   ¢ !r       È/ !‚       È !‚       âO G‰        € ê „y	 €  H   ( „y	 €   H   h „y	    H   ¨ „y	    H   è „y	 €  H   " !r
       â !r       Æ/ !r       âO !r       Æ� !r
       ä‰      " z ÿ `  pPð ä r 	ÿ   pRò ä z ÿ a   Sð â !r       Ê ‰      $ !r       Ê ‰ €    $ !r       Ê ‰ @    $ !r       Ê ‰      $!r	       â G‰        € ê ‰
      $ !r
       Ê ‰      $ !r       Ê ‰ €    $ !r       Ê ‰ @    $ !r       Ê ‰      $ !r
       ä M        € ê !r	       â G‰  `    € ê $tÿ   ÿ Ž â !r 
       È z  h  pð Ú G	  `    € ê $v ÿ h  ÿ Ž Ê x     pRð Ú G‰  �     € ê x     pRð Ú G	       € ê  x ;ª¸¿  A  Ì s          $ !t   €?    Ì s        $  r     A  È  r     A  â Gy  `    € ê  x '7=  A  â x*BL?  A  Æ #t   €?    È  r      A  â  x    ?  A  Ê s     $   $ !t  €?    È  r       A  È  r      A  â Gy  À    € ê $v ÿ h  ÿ Ž Ê x     pRð Ú G‰  à     € ê x     pRð Ú G	        € ê 	z  i    � ä 	z i    � Æ  x ;ª¸¿  A  â 	z i €  � Ê s        $ !t  €?    Ì s        $  r       A  È  r      A  â Gy  Ð     € ê  r     A  â Gy  °     € ê 	x   à@  � ä 	x  à@  � Æ  x #ÛÙ?  A  â 	x  àÀ  � Æ  x;ª¸¿  A  ä !t  €?    È s        $ !t  €?    Ì s        $  r       A  È  r      A  ä y       %   ( y      &   ( y      '   b $z  x   Ž ä $tÿ   ÿ Ž Ä $z  ~   Ž È/ %v  j   Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y       &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž ä $rÿÿ     Ž È % \   Ž Ê �	    ™ b ¹z  l      â Ey  à(    € â ™x?    â y	      "   b $rÿÿ   ÿ Ž Æ y      !   ¢ $ÿÿ   ÿ Ž ä $ÿÿ    Ž â y      %   & '
 m   Ž â y      '   æ $‚ÿÿ   ÿ Ž â 
ÿ n    Ê $ÿÿ   
Ž â r 	ÿ   pRò Æ/ $
 o  Ž ä $x
	    Ž äO $‚ÿÿ    Ž ä $rÿÿ   ÿ Ž â xÿ   
  â 'Š s   Ž Æ r
   ÿ� â $	€   Ž Æ xÿ     È |    pbô ä Šÿ t    ä r 	ÿ   pRð â x      ÿ  â $x	   ÿ Ž Î G)  '    € ê $rÿÿ   ÿ Ž â� xøÿÿÿÿÀŽ â $vÿ Y  ÿ Ž ä 'z y   Ž È $x

   
Ž â zÿ z    â $z	 p  Ž ä $x#
   ÿ Ž â xÿ   
  â $z |  Ž ä $z w  ÿŽ ä $z v  Ž Ä$z }  ÿŽ ä %x J   ÿŽ Æ r   ÿà÷ â $vÿ X  ÿ Ž â r
    ÿàõ â $rÿÿ   ÿ Ž ä $x   ÿ Ž ä $rÿÿ   ÿŽ ä %x"#    Ž È $x!    ä r"   ÿàõ Ê $x#    Ê �y   • ¨  �yW   • è  �yV   • ( �yT   • b $vUÿ p  ÿ Ž È %xUJ   Ž Ê �yS   • ( �yR   • ( �y   • h �yI   • b %xUJ   Ž Ê �yO   • h �yN   • h �yL   • h �yM   • d %xUJ   Ž Ê �y   • h �yJ   • h �yK   • h �yB   • b $tÿ$   ÿ Ž Ä $x   ÿ Ž ÊO xø  ÿÀŽ È z*  ÿàõ Ê $v+ÿ 	 ÿ Ì �y**    › ¢ $xW   ÿ Ž Ê� x&ø  ÿÀŽ â $xV   ÿ Ž Æz&&  ÿàõ ä xø  ÿÀŽ Æ $v'ÿ 	 ÿ â z2  ÿàõ Ê �y&&    › â $v3ÿ 	 ÿ Ì �y22    › " $xT   ÿ Ž Êxø  ÿÀŽ È z$  ÿàõ Ê $v%ÿ 	 ÿ Ì �y$$    › b $xS   ÿ Ž ä $xR   ÿ Ž Æ xø  ÿÀŽ È z4  ÿàù ä xø  ÿÀŽ Æ $v5ÿ 	 ÿ â z6  ÿàû ä r
   àµ Æ �y44    › b r   ÿd~ â $v7ÿ 	 ÿŽ ä %v Z   Ž È $x$   ÿŽ â �y66    › f $x   Ž Ê �yG   ™ b $x   ÿ Ž Æ �yE   ™ h �yF   ™ b xø  ÿÀŽ Æ �y   ™ b! z0  ÿàõ Æ �yD   ™ d $v1ÿ 	 ÿ ä �y   ™ h �y00    › b $xI   ÿ Ž Æ �y    ™ d xø  ÿÀŽ ä �y   ™ d z.  ÿàõ â $xO   ÿ Ž È $v/ÿ 	 ÿ â xø  ÿÀŽ ä xÿ	     â �y    • h �y..    › b z,  ÿàõ Ê $v-ÿ 	 ÿ ä $xN   ÿ Ž È �y,,    › b x(ø  ÿÀŽ ä 	s         d z((  ÿàõ Ê $v)ÿ 	 ÿ Ì �y((    › b $x€   ÿ Ž â/ r8   ÿàõ È xZ€   øŽ â $xL   ÿ Ž ä $x9    â r>    ÿà÷ ä xXø  ÿÀŽ â %x<UJ   8Ž â �yH8B   ‘ b xWÿ	   W  ä $xM   ÿ Ž â �y<B   ‘ b $tCÿÿ Ž Ä $xZZÿŽ â z8X  ÿàõ â/ $x?!   Ž â 	sX W        b $tYÿ @€ÿ Ž ä %xPUJ   <Ž â x<ø  ÿÀŽ â �y>    • d $v9ÿ 	 ÿ â xVÿ	   V  â %x:UJ   >Ž â xZ€€€€CìŽ Ä z<<  ÿàõ â �y88    › b x>Z€€€€YìŽ ä xÿþþþÿàÿ â 	sY V        " x]>ÿþþþÿàÿ â %x@UJ   :Ž â x[ZøŽ â�y:    • b x]] @€ZøŽ â $v=ÿ 	 ÿ â x[[˜º  ÿ    â $xXX€   ÿ Ž â/ x\]˜º  ÿ    â �y@    • b xZ[ÿÀŽ â %x>UJ   @Ž Ä �y<<    › b xXW€   XøŽ â %x:UJ   PŽ ä �yCPB   ‘ b xA\ÿÀŽ â/ $xYY€   ÿ Ž ä �y>>    • h �yP:B   ‘ b xYV€   YøŽ â $tQÿ @€ÿ Ž Ä $x:XÿŽ ä $t;ÿÿ Ž ä $tVÿÿ Ž â xQ:€€€€QìŽ â $tXÿ @€ÿ Ž Æ xQQÿþþþÿàÿ È xQQ @€:øŽ ä x[[€€€€*ÞŽ äO x\\€€€€+ÞŽ Æ $x[[   Z
Ž ä $x@\   A
Ž Æ xA[€€€€*´Ž ä x*:€€€€;ìŽ ä x@@€€€€+´Ž â $x+YÿŽ â x;*ÿþþþÿàÿ È x?+€€€€VìŽ ä/ x;;:øŽ ä xU+€€€€XìŽ ä x*ÿ	   T  ä xT?ÿþþþÿàÿ ä x;;˜º  ÿ    Ä xXUÿþþþÿàÿ â 	s: *        " xVQ˜º  ÿ    ä xQT+øŽ äxT;ÿÀŽ äx;;€€€€&ÞŽ ä� xXX @€+øŽ Ä x?VÿÀŽ äxVV€€€€'ÞŽ â $x;;   T
Ž â xXX˜º  ÿ    ä xQQ˜º  ÿ    â $xVV   ?
Ž â x+XÿÀŽ äxXX€€€€3ÞŽ ÄxT;€€€€&´Ž ä x&QÿÀŽ äxQQ€€€€2ÞŽ ä x?V€€€€'´Ž â $x':€   ÿ Ž ä $xVX   +
Ž ä $x+   ÿ Ž Ä $xQQ   &
Ž â x&*€   'øŽ ä x++ø  ÿÀŽ ä xQQ€€€€2´Ž â $x2J   ÿ Ž â z*+  ÿàõ â $t;ÿÿ Ž ä $x&&ÿŽ â x:2ø  ÿÀŽ â $tWÿ @€ÿ Ž Ä $v+ÿ 	 ÿ â x2&€€€€;ìŽ â$x'K   ÿ Ž â xW&€€€€WìŽ ä x;2ÿþþþÿàÿ â �y**    › ¢ xWWÿþþþÿàÿ ä xX'ø  ÿÀŽ ä z::  ÿàõ Ä x';&øŽ äx2W @€&øŽ â $v;ÿ 	 ÿ â xY'˜º  ÿ    ä z&X  ÿàõ ä xX2˜º  ÿ    â �y::    › â x'YÿÀŽ Ä x2Y€€€€$ÞŽ Ê$xY2   '
Ž â x2V€€€€3´Ž â $x3B   ÿ Ž â x[XÿÀŽ â$v'ÿ 	 ÿ â xXX€€€€%ÞŽ ä x33ø  ÿÀŽ ä xVY€€€€$´Ž â $xXX   [
Ž â �y&&    › " z$3  ÿàõ Ä xRÿ	   R  ä xSÿ	   S  ä xXX€€€€%´Ž â $v%ÿ 	 ÿ â 	sU R        * �y$$    › f 	sW S        b $xUU€   ÿ Ž Ä $xZW€   ÿ Ž Æ/ x3R€   UøŽ ä xRS€   ZøŽ â $t\ÿÿ Ž ä $tZÿ @€ÿ Ž ä $tSÿÿ Ž ä $x33ÿŽ ä $xRRÿŽ Ä $tUÿ @€ÿ Ž â x\3€€€€\ìŽ äxZ3€€€€ZìŽ ä xSR€€€€SìŽ äxUR€€€€UìŽ ä x\\ÿþþþÿàÿ ä xZZÿþþþÿàÿ Ä xSSÿþþþÿàÿ ä xÿ	     ä xUUÿþþþÿàÿ ä x\\3øŽ äxZZ @€3øŽ ä xSSRøŽ â	s3         " xRU @€RøŽ Ä xSS˜º  ÿ    ä x[\˜º  ÿ    ä xRR˜º  ÿ    ä xWSÿÀŽ äx\S€€€€4ÞŽ ä xSRÿÀŽ Ä xRR€€€€5ÞŽ â $xW\   W
Ž â xY[ÿÀŽ â $x\3€   ÿ Ž â x[[€€€€6ÞŽ â $xRR   S
Ž â x4W€€€€4´Ž ä x3€   \øŽ â $xY[   Y
Ž â xRR€€€€5´Ž â &r44G   ÿ   Ä $x33ÿŽ â xYY€€€€6´Ž â &r4RE   4   ä $t6ÿÿ Ž â xIÿ	   I  â $tRÿ @€ÿ Ž ä &rYYF   4   â x43€€€€6ìŽ â	s I        " x53€€€€RìŽ Ä xZZ˜º  ÿ    ä x44ÿþþþÿàÿ ä x65ÿþþþÿàÿ ä xUZÿÀŽ ä x543øŽ ä xZZ€€€€7ÞŽ â &rTT   ÿ   â x66 @€3øŽ Ä xOÿ	   O  ä x55˜º  ÿ    â $xZZ   U
Ž â xR6˜º  ÿ    â &r??D   T   â 	s4 O        b x65ÿÀŽ â $xT€   ÿ Ž â x55€€€€0ÞŽ ä xZZ€€€€7´Ž Ä x7RÿÀŽ â$x55   6
Ž â xRR€€€€1ÞŽ ä xTI€   TøŽ â &rVVG   ÿ   â x05€€€€0´Ž â $xRR   7
Ž ä &r3Q   ?   ä $xTTÿŽ Ä $t7ÿÿ Ž ä $t5ÿ @€ÿ Ž ä &rVXE   V   ä &r32   3   â x2T€€€€7ìŽ â $x64€   ÿ Ž â/ x4T€€€€5ìŽ â &rAAF   V   â x52ÿþþþÿàÿ Ä x74ÿþþþÿàÿ ä xNÿ	   N  â &r@@   A   â x1R€€€€1´Ž â &r00   ÿ   â x6O€   6øŽ â 	sA N        " x55TøŽ äx77 @€TøŽ â &r11D   0   â x55˜º  ÿ    â $x06ÿŽ â x47˜º  ÿ    â $t?ÿÿ Ž â x25ÿÀŽ Ä x55€€€€.ÞŽ ä x74ÿÀŽ ä x60€€€€?ìŽ ä x44€€€€/ÞŽ â $tIÿ @€ÿ Ž â x?6ÿþþþÿàÿ â $x55   2
Ž ä $x44   7
Ž â x20€€€€IìŽ â $xAA€   ÿ Ž â x??0øŽ Ä x55€€€€.´Ž ä x.4€€€€/´Ž ä x/2ÿþþþÿàÿ ä x??˜º  ÿ    ä xAN€   AøŽ ä x// @€0øŽ â $t2ÿÿ Ž â x0?ÿÀŽ â $xAAÿŽ â x??€€€€,ÞŽ â $t4ÿ @€ÿ Ž â xLÿ	   L  â &r15   1   â x2A€€€€2ìŽ â $x??   0
Ž â 	s5 L        " x0/˜º  ÿ    Ä x4A€€€€4ìŽ ä x22ÿþþþÿàÿ ä x/0ÿÀŽ ä x44ÿþþþÿàÿ ä x00€€€€-ÞŽ ä xMÿ	   M  Ä x22AøŽ â $x00   /
Ž â x44 @€AøŽ â 	s6 M        b x/2˜º  ÿ    ä x44˜º  ÿ    â $x55€   ÿ Ž â x,?€€€€,´Ž ä x2/ÿÀŽ Ä x00€€€€-´Ž ä x//€€€€(ÞŽ ä x-4ÿÀŽ äx44€€€€)ÞŽ â &r,,G   ÿ   â x5L€   5øŽ â $x//   2
Ž ä $x44   -
Ž Ä &r-0E   ,   â x//€€€€(´Ž â $x55ÿŽ ä $t,ÿÿ Ž ä $x66€   ÿ Ž â/ x(4€€€€)´Ž â &r-/F   -   â x)5€€€€,ìŽ â $t0ÿ @€ÿ Ž â x6M€   6øŽ â &r((   -   â x,)ÿþþþÿàÿ â &r..   1   â x-5€€€€0ìŽ â $x66ÿŽ Ä $t)ÿÿ Ž â x,,5øŽ â $t1ÿ @€ÿ Ž â x0-ÿþþþÿàÿ ä x-6€€€€)ìŽ ä x),˜º  ÿ    ä x16€€€€1ìŽ ä x,)ÿÀŽ Ä x11ÿþþþÿàÿ ä x))€€€€8ÞŽ ä x00 @€5øŽ ä x/-ÿþþþÿàÿ ä x11 @€6øŽ â $x))   ,
Ž â x00˜º  ÿ    Ä x//6øŽ ä x21˜º  ÿ    ä x8)€€€€8´Ž ä x-0ÿÀŽ ä x//˜º  ÿ    ä xJÿ	   J  Ä x)2ÿÀŽ ä x00€€€€9ÞŽ ä x22€€€€=ÞŽ â 	s1 J        " x,/ÿÀŽ â$x00   -
Ž â x//€€€€<ÞŽ â $x22   )
Ž â x)Hð   ÿÀŽ Ä x-H   ÿÀŽ â $x//   ,
Ž â x,ÿ   )  ä xKÿ	   K  â $r)-@   ÿŽ â x-ÿ	     â &r88   ÿ   â x90€€€€9´Ž â 	s4 K        b x//€€€€<´Ž â $x11€   ÿ Ž Ä &r99D   8   È 	s0 -        " $x@@   3Ž â x22€€€€=´Ž â &r9/   9   â x1J€   1øŽ â $r,3   )Ž â r/@@   ÿ� â &r)2   9   â x,ð   ÿÀŽ â &rZ   Y   â x   ÿÀŽ â $t2ÿÿ Ž Ä $x11ÿŽ â r/   ÿú� â $t6ÿ @€ÿ Ž â x/ÿ   ,  â $x44€   ÿ Ž ä/ $x00€   ÿ Ž â x,1€€€€2ìŽ â$r   ÿŽ â x21€€€€6ìŽ â $x   .Ž â xBÿ	   B  â $r./.   Ž â x4K€   4øŽ Ä x0-€   0øŽ ä x,ÿþþþÿàÿ â 	s3 B        " x,2ÿþþþÿàÿ â $x44ÿŽ ä $t/ÿÿ Ž â x1øŽ â$x00ÿŽ â x,, @€1øŽ â $t-ÿÿ Ž Ä $t1ÿ @€ÿ Ž â x/4€€€€/ìŽ â $t5ÿ @€ÿ Ž â x-0€€€€-ìŽ ä x24€€€€1ìŽ ä x1/ÿþþþÿàÿ ä x/0€€€€5ìŽ ä x--ÿþþþÿàÿ â $x33€   ÿ Ž â x//ÿþþþÿàÿ Ä x--0øŽ äx// @€0øŽ ä x--˜º  ÿ    ä xBB€   3øŽ ä x32ÿþþþÿàÿ ä x2/˜º  ÿ    Ä x0-ÿÀŽ äx--€€€€*ÞŽ äO x/2ÿÀŽ äx22€€€€+ÞŽ â $x--   0
Ž È $x22   /
Ž â x*-€€€€*´Ž ä x,,˜º  ÿ    â $xBBÿŽ â x22€€€€+´Ž â $t5ÿÿ Ž â x33 @€4øŽ â &r**G   ÿ   â x-˜º  ÿ    ä x+,ÿÀŽ â$t7ÿ @€ÿ Ž â x,,€€€€;ÞŽ â� &rE2E   *   â x114øŽ Ä x4B€€€€5ìŽ ä x*3˜º  ÿ    ä x-ÿÀŽ â$x,,   +
Ž â x--€€€€:ÞŽ ä x6B€€€€7ìŽ ä x54ÿþþþÿàÿ Ä x11˜º  ÿ    ä x+*ÿÀŽ äx**€€€€'ÞŽ â$x--   
Ž â x76ÿþþþÿàÿ ä x55BøŽ ä x1ÿÀŽ â$x**   +
Ž â x11€€€€&ÞŽ Ä x77 @€BøŽ ä x55˜º  ÿ    â $x11   
Ž â x'*€€€€'´Ž ä x07˜º  ÿ    ä x*5ÿÀŽ äx55€€€€$ÞŽ Äx&1€€€€&´Ž ä x+0ÿÀŽ â$x55   *
Ž â x00€€€€%ÞŽ â &r&&   ÿ   â x--€€€€:´Ž ä x,,€€€€;´Ž â $x00   +
Ž â x55€€€€$´Ž â &r--F   E   Ä &r''D   &   â x00€€€€%´Ž â &r-,   -   â xCð   ÿÀŽ â &r'5   '   â xPð   ÿÀŽ äxCC   ÿÀŽ ä xPP   ÿÀŽ â &r'0   '   â xÿ     â $rCC(   ÿŽ â xÿ     â $rPP-   ÿŽ Ä $x((   )Ž ä $x--   'Ž â r   ÿ� â $r)   CŽ â r((   ÿ� â $rP'   PŽ â r---   ÿ� ä r.   ÿú� ä r   ÿú� Ä r--P   ÿú� ä x   ÿàÿ ä xÿ     ä xÿ     ä xÿ     ä x$ÿ   -  Ä |    pbô ä r   ÿ� ä r   ÿ� ä r   ÿ� ä r$$-   ÿ� â 0rÿ    A   â xÿ     â 0rÿ    A   â xÿ     â 0rÿ    A   â xÿ     â 0rÿ    A   â x$ÿ   $  â 0r>ÿ>    A   â  r     A  â  r     A  â r     A  âx   ÿà÷ â  r>     A  â Er        Ä Er        ä Er        ä Er$ $       â #r      â #r      â #r      â x€   ÿàÿ â #r$      â $rÿÿ   Ž â D)       À â Gy  PÚÿÿÿÿƒ ê Ay        € ê� ˆ  þÿ H   è ˆ €þÿ H   è ˆ  ÿÿ H   è ˆ €ÿÿ H   è {          ì M	        € ê „y     H   " z 	 r  p`ò Æ „y €   H   b r ÿ   pVò  Æ „y    H   ¨ „y    H   è„y €  H   ( „y    H   ( „y
    H   ( „y €  H   ( „y €  H   ( „y    H   ( „y €  H   " !r       Æ „y €  H   " !r       â/ !r       âO !r       ä� !r       â!r       â !r
       â !r       È ‰      b !r       â !r       È ‰      ¢ !r       Æ ‰      â !r       Ê ‰      " !r       â/ !r       âO !r       È� ‰      h ‰      ¢ !r       Æ ‰      ( ‰      â !r       â/ !r       ÈO ‰
 €    b !r       Æ ‰ €    " !r       Æ� ‰ €    ¨ ‰ €    â !r

       â/ !r       È ‰
 @    " !r       ÆO ‰ @    b !r       Æ� ‰ @    ¨ ‰ @    â !r
       â x
	   ÿàÿ â !r       Æ/ ‰      " $z  x  	Ž â x 	   ÿàÿ â!r       âO ‰      b x	   ÿàÿ â $z ~  Ž â !r       â� ‰      ¢ z   r  p`ô â $t ÿ   ÿ Ž â z  r  p`ö â ‰      â z 
 r  p`ð Ä x    pVt äx    pVö äx    pVp  â %v j    Ž â !r       â !r       Ì/ †©    â !r       ÆO †™     â !r       Æ� †¹    â M	        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      &   " z ÿ \  pPò â ¹z  F   
   Æ z ÿ ]  Sò Ú $ÿ   ÿ Ž È % \   Ž Ê �    ™ ¢  $rÿÿ   ÿ Ž â ¹z  l      â Ey  €    € â y      %   b ' m   Ž â™x?    ä y      "   â 'š s   Ž â ÿ n    Ä y       !   & $ÿÿ   
Ž â y	      '   f $ o  Ž ä$’ÿÿ    Ž È $z w  ÿŽ â x   ÿ   â/ $rÿÿ   ÿ Ž â r ÿ   pRð Æ� $x
    Ž âx     pDp  ä r(    ÿ(Ž â $z	 }  ÿŽ âz 
 r  pdp  â 'z	 y   Ž â x      ÿ  ä r   ÿàõ ä xÿ   (  È r(   ÿ� ä šÿ t    È r     € âO G	  °    € ê z ÿ `  pPò ä z ÿ ^  pPð ä z ÿ b  p ö ä z ÿ a  Sò ä z ÿ _   Sð ä z ÿ c  0'ò Ö $
	 ~  ÿŽ âÿ     â$š	 ~  ÿŽ ä $
 x  ÿŽ â     àë â$š	 x  ÿŽ â ’    àÇ ä ˜ÿ     ä ÿ   ÿÄþ Ä    ÿàý ä ’	   ÿàû ä ’ÿ   ÿ„þ â $ÿÿ    â 
 ^  ÿ† ä š
 b  ÿˆ â $’	ÿÿ   Ž â 
 _  � Ä x      ÿ  ä š c  	 È �	     h �™
     d Ay        € ê xÿ     â Ey  PM    € â x      ÿ  â $r#ÿÿ   ÿ Ž â |    pbð â $r	ÿÿ   ÿ Ž â r$ ÿ       â x
      ÿ  â $rÿÿ   ÿ Ž Ä $rÿÿ   ÿ Î G	  °L    € ê xøÿÿÿÿÀŽ â $z p  Ž âzÿ z    â $vÿ `  ÿ Ž â r((  €ÿàÿ â $vÿ a  ÿ Ž â x   ÿàÿ â $z |  Ž â x	   ÿàÿ â $x8(   ÿ Ž â x)ÿ   (  â $z |  Ž â x   ÿàÿ â %x88   ÿŽ â z  Y      Æ $z v  Ž ä$z v  Ž ä $z p  Ž ä %x*J   8Ž È $z p  Ž âx*   ÿàñ â $z	 p  Ž â r	 ÿ       â %x4J   8Ž âx      ÿ  â x
      ÿ  ä %x6J   8Ž È $vÿ X  ÿ Ž ä $r#ÿÿ   ÿ Ž ä $r$ÿÿ   ÿ Ž ä $rÿÿ   ÿ Ž ä $x   ÿ Ž ä %x,J   (Ž È %x.J   (Ž È %x0J   (Ž È %x2J   (Ž È %x8J   8Ž È $rÿÿ   +  ä r*   ÿàñ È r+   ÿä  Ê �yg   • ¨ �yS   • è �yZ   • ( �yf   • " r   ÿàñ â $vmÿ p  ÿ Ž È $x     È %xmJ   Ž Ê �yeúÿÿ • ( �y`üÿÿ • ( �y"þÿÿ • ( �ya    • " %xmJ   Ž Ê �yúÿÿ • ( �y_üÿÿ • ( �yWþÿÿ • ( �yV    • " %xFmJ   Ž Ê �y'Fþÿÿ • ( �y^Füÿÿ • ( �y&Fúÿÿ • " $tCÿ$   ÿ Ž Æ �y%F    • " $xg   ÿ Ž ÊO xø  ÿÀŽ È z:  ÿàñ â $xS   ÿ Ž Æ� z;ÿ 	 ÿä  ä xø  ÿÀŽ È �y::    › ¢ z<  ÿàñ â $xZ   ÿ Ž È$v=ÿ 	 ÿ  â x>ø  ÿÀŽ ä xf   ÿ   Æ �y<<    › â z>>  ÿàñ ä xø  ÿÀŽ Æ $v?ÿ 	 ÿ  â z@  ÿàõ Ê �y>>    › " $vAÿ 	 ÿ Ì �y@@    › " rB(   à‘ È r)   ÿ$~  â %vBB Z  C Ž È $x$   ÿŽ Ê rCC   ÿàÿ Ê �yB   ™ ( �yB   ™ (! �yB   ™ ( �yB   ™ ( �yB   ™ " $xHe   ÿ Ž Æ �yB   ™ " rD.   ÿàñ ä xHHø  ÿÀŽ â �y B   ™ $ $xE/     â zFH  ÿàñ â �y!B    ™ ( �ybDB   ‘ " $vGÿ 	 ÿ  Ä $xI"   ÿ Ž â xLa   ÿ   â $xN   ÿ Ž â xR_   ÿ   â �yFF    › " xH`   ÿ   ä xnÿ	   S  â %xPJ   Ž â xHHø  ÿÀŽ â �yhB    • & zHH  ÿàñ Ä xJIø  ÿÀŽ ä xLLø  ÿÀŽ â $vIÿ 	 ÿ  â xNNø  ÿÀŽ ä xRRø  ÿÀŽ â 	so n        " �ycP    • ( �yHH    › " zJJ  ÿàñ Ê $vKÿ 	 ÿ  Ì �yJJ    › " zLL  ÿàñ Ê $vMÿ 	 ÿ  Ì �yLL    › " zNN  ÿàñ Ê $vOÿ 	 ÿ  Ì �yNN    › " zRR  ÿàñ â %xTmJ   PŽ È $vSÿ 	 ÿ  â �ydT    • " $xPW   ÿ Ž ä %xDmJ   DŽ ä/�yRR    › $ %xXmJ   TŽ â xlÿ	   Z  â �yiDB   ‘ " xTPø  ÿÀŽ â $xoo€   ÿ Ž Ä %x\mJ   DŽ â	sp l        b xgÿ	   g  â �yjX    • " xon€   oøŽ ä zDT  ÿàñ â %xZmJ   XŽ â xfÿ	   f  â �yk\B   ‘ " zEÿ 	 ÿä  â $tBÿÿ Ž â 	sC g        " $xooÿŽ â �yZZ    • " $tXÿ @€ÿ Ž Æ/ �yDD    › " xBo€€€€BìŽ äxUo€€€€XìŽ â $x[V   ÿ Ž â xBBÿþþþÿàÿ â %xPmJ   \Ž â xYp   ÿ   ä x[[ø  ÿÀŽ ä xXUÿþþþÿàÿ â 	s\ f        " xUBoøŽ â �yTPB   ‘ " zB[  ÿàñ Ä xYl€   YøŽ â $t[ÿÿ Ž ä $xPC€   ÿ Ž â/ xQX @€oøŽ â $vCÿ 	 ÿ  ä $xXYÿŽ â xPg€   PøŽ ä xg  @€    â �yBB    › " xYX€€€€[ìŽ Ä x[X€€€€gìŽ â $x]\€   ÿ Ž â xYYÿþþþÿàÿ â $xPPÿŽ â x[[ÿþþþÿàÿ â $tgÿÿ Ž â xYYXøŽ â$tmÿ @€ÿ Ž â xX[ @€XøŽ ä x]f€   ]øŽ Ä x[P€€€€gìŽ äx\P€€€€mìŽ â $xl]ÿŽ â x[[ÿþþþÿàÿ ä x]\ÿþþþÿàÿ ä x[[PøŽ ä x\l€€€€gìŽ Ä xP] @€PøŽ ä x][˜º  ÿ    ä xg\ÿþþþÿàÿ ä xfl€€€€mìŽ ä x\P˜º  ÿ    ä xP]ÿÀŽ Ä xn'   ÿ   ä xmfÿþþþÿàÿ ä x[gløŽ ä xg\ÿÀŽ ä xnnø  ÿÀŽ ä xUU˜º  ÿ    Ä xYY˜º  ÿ    ä x]]€€€€:ÞŽ äO xf\€€€€;ÞŽ Æ $x]]   P
Ž â zPn  ÿàñ â $xff   g
Ž â x\m @€løŽ â $xg^   ÿ Ž â x]]€€€€:´Ž ä xlQ˜º  ÿ    ä x:UÿÀŽ ä zQÿ 	 ÿä  Ä xUU€€€€<ÞŽ ä� xmgø  ÿÀŽ ä xglÿÀŽ â$xUU   :
Ž â xll€€€€=ÞŽ â �yPP    › ¢ z:m  ÿàñ ä xff€€€€;´Ž â $xnl   g
Ž â xlU€€€€<´Ž â $v;ÿ 	 ÿ  â x<YÿÀŽ Ä xYY€€€€>ÞŽ äx[[˜º  ÿ    â �y::    › â xpX˜º  ÿ    ä rYY<  €ÿàÿ ä x<[ÿÀŽ ä xXn€€€€=´Ž Ä x[[€€€€@ÞŽ ä x=pÿÀŽ äxpp€€€€?ÞŽ â $x[[   <
Ž â x\\˜º  ÿ    â $x<&   ÿ Ž â xeÿ	   e  â $xpp   =
Ž â x=\ÿÀŽ Ä xYY€€€€>´Ž ä x\\€€€€AÞŽ â	s> e        " x<<ø  ÿÀŽ ä r\\=  €ÿàÿ â $x=%   ÿ Ž â z<<  ÿàñ ä x\\€€€€A´Ž Ä xA=ø  ÿÀŽ â $v=ÿ 	 ÿ  â x?p€€€€?´Ž â &rYY   ÿ   â x[[€€€€@´Ž â $x@>€   ÿ Ž ä �y<<    › " z>A  ÿàñ â &rY?   Y   Æ z?ÿ 	 ÿä  Ì �y>>    › ¢ &r]]   ÿ   È &r]f   ]   â x@e€   @øŽ Æ &r]l   ]   ä $xA@ÿŽ ä &r]X   ]   ä $tXÿÿ Ž ä $t@ÿ @€ÿ Ž â x`ÿ	   `  ä xUA€€€€XìŽ â &rY[    Y   â x[A€€€€@ìŽ â 	se `        " xXUÿþþþÿàÿ â &rY\!   Y   â x\[ÿþþþÿàÿ Ä x"ÿ	   "  ä xXXAøŽ äx@bð   ÿÀŽ â	s[ "        b xbb   ÿÀŽ ä x\\ @€AøŽ ä xAX˜º  ÿ    â $rbb]   ÿŽ â x@ÿ   @  â $x]]   YŽ â x\\˜º  ÿ    Ä xXAÿÀŽ äxAA€€€€FÞŽ â $r@@Y   bŽ â xU\ÿÀŽ äx\\€€€€GÞŽ â $xAA   X
Ž â xee   ÿ   ä r]]]   ÿ� â $x\\   U
Ž â xe`€   eøŽ â $x[[€   ÿ Ž â/ r@]@   ÿú� Ä xUA€€€€F´Ž â $xeeÿŽ â xAÿ   @  ä x\\€€€€G´Ž â &rUU   ÿ   â xF     ä x["€   [øŽ ä r"A@   ÿ� â &r@\   U   â xAe€€€€FìŽ â $x[[ÿŽ Ä $t\ÿÿ Ž â xFAÿþþþÿàÿ â $tXÿ @€ÿ Ž â xaÿ	   a  ä xA[€€€€\ìŽ ä xFFeøŽ ä x\Aÿþþþÿàÿ â 	sY a        " xAF˜º  ÿ    Ä xGe€€€€XìŽ ä xFAÿÀŽ ä xXGÿþþþÿàÿ ä xAA€€€€HÞŽ â $t`ÿ @€ÿ Ž â xXX @€eøŽ ä rAAF  €ÿàÿ Ä xU[€€€€`ìŽ ä xXX˜º  ÿ    ä x\\[øŽ ä xAA€€€€H´Ž ä xHÿ	     â $xY€   ÿ Ž â x`Uÿþþþÿàÿ Ä xGXÿÀŽ â 	sY H        " xU\˜º  ÿ    ä xXX€€€€IÞŽ ä xFUÿÀŽ äxUU€€€€JÞŽ â $xXX   G
Ž â xaa€   øŽ â &r@A   @   â x`` @€[øŽ â $xUU   F
Ž â xX€€€€I´Ž â $xaaÿŽ Ä $tFÿÿ Ž â x``˜º  ÿ    â &r   @   â xUU€€€€J´Ž â $tJÿ @€ÿ Ž â x@a€€€€FìŽ â $xYY€   ÿ Ž â xG`ÿÀŽ äx``€€€€KÞŽ ä xFa€€€€JìŽ Ä x@@ÿþþþÿàÿ ä xYH€   YøŽ ä r``G  €ÿàÿ ä xFFÿþþþÿàÿ ä x@@aøŽ ä x_ÿ	   _  â $xYYÿŽ â xA`€€€€K´Ž Ä xH     â 	sK _        " xFF @€aøŽ ä xG@˜º  ÿ    ä xHY€€€€HìŽ ä xFF˜º  ÿ    ä x@GÿÀŽ Ä xGG€€€€LÞŽ ä xHHÿþþþÿàÿ ä xIFÿÀŽ â$xGG   @
Ž â xFF€€€€MÞŽ ä xJY€€€€JìŽ ä xHHYøŽ Ä xWÿ	   W  â $xFF   I
Ž â xJJÿþþþÿàÿ ä xGG€€€€L´Ž ä xIH˜º  ÿ    â 	sL W        b $xHK€   ÿ Ž â xJJ @€YøŽ â &rUU   ÿ   â x@IÿÀŽ Ä xJJ˜º  ÿ    ä xH_€   HøŽ ä xII€€€€NÞŽ â &rAA   U   â xFF€€€€M´Ž â $xHHÿŽ â xKJÿÀŽ â$tMÿÿ Ž â xJJ€€€€OÞŽ Ä rII@  €ÿàÿ â &rAG    A   â xVÿ	   V  â $xJJ   K
Ž â xGH€€€€MìŽ â $xLL€   ÿ Ž â/ xII€€€€N´Ž â 	sK V        " xM  @€    ä xGGÿþþþÿàÿ â &rII   ÿ   â x@J€€€€O´Ž â &rAF!   A   â xLW€   LøŽ Ä xFH€€€€MìŽ ä xGGHøŽ â&r@@   I   â xIFÿþþþÿàÿ â $xLLÿŽ â xGG˜º  ÿ    â $tMÿÿ Ž ä $tOÿ @€ÿ Ž â xII @€HøŽ Ä xFGÿÀŽ äxHL€€€€MìŽ äxGG€€€€RÞŽ ä xJL€€€€OìŽ ä xOK   ÿ   ä xKHÿþþþÿàÿ â $xGG   F
Ž â xMJÿþþþÿàÿ Ä xFI˜º  ÿ    ä xKKLøŽ äxMM @€LøŽ ä xIFÿÀŽ äxFF€€€€SÞŽ ä xOV€   OøŽ Ä xKK˜º  ÿ    ä xJM˜º  ÿ    â $xFF   I
Ž â xHKÿÀŽ â $xOOÿŽ â xKK€€€€DÞŽ â $tLÿÿ Ž â xIJÿÀŽ â$tNÿ @€ÿ Ž â xJJ€€€€EÞŽ Ä rKKH  €ÿàÿ ä xHO€€€€LìŽ â$xJJ   I
Ž â xIO€€€€NìŽ ä xKK€€€€D´Ž ä xHHÿþþþÿàÿ ä xDIÿþþþÿàÿ â &rKK   ÿ   â xHHOøŽ Ä xEJ€€€€E´Ž ä xDD @€OøŽ ä xJÿ	   '  ä xIH˜º  ÿ    â &rEE   K   â xGG€€€€R´Ž â 	sK J        " xHD˜º  ÿ    Ä xDIÿÀŽ â &rGG   @   â xII€€€€BÞŽ ä x@ið   ÿÀŽ äx'HÿÀŽ äxii   ÿÀŽ â $xII   D
Ž â xHH€€€€CÞŽ Ä x^ÿ	   ^  â $rii   ÿŽ â x@ÿ   @  â $xHH   '
Ž â 	sD ^        b xFF€€€€S´Ž ä xII€€€€B´Ž ä r'A   ÿàÿ â$r@A   iŽ â x@kð   ÿÀŽ â $xKK€   ÿ Ž â xHH€€€€C´Ž â &rEI    E   â xkk   ÿÀŽ â &rFF   G   â r'''   ÿ� â &rEH!   E   â xAÿ   @  â $rBkF   ÿŽ â xKJ€   KøŽ Ä r'   ÿú� â $r'AE   BŽ â xB     â $xAD€   ÿ Ž ä/ $xKKÿŽ ä $tDÿ @€ÿ Ž â xHÿ	   %  ä x%K€€€€BìŽ Ä xA^€   AøŽ ä xBK€€€€DìŽ â $x@F   EŽ â xD%ÿþþþÿàÿ â $tCÿÿ Ž â xFBÿþþþÿàÿ â $xBAÿŽ â x&ÿ	   &  ä xDDKøŽ Ä 	sA &        " x%B€€€€CìŽ â $tGÿ @€ÿ Ž â xED˜º  ÿ    ä x%%ÿþþþÿàÿ ä xDEÿÀŽ â	sJ H        b xEE€€€€PÞŽ äO xCB€€€€GìŽ Ä x%%BøŽ ärGED  €ÿàÿ ä xECÿþþþÿàÿ ä xFF @€KøŽ ä xC%˜º  ÿ    â $xAA€   ÿ Ž â xEE @€BøŽ Ä xFF˜º  ÿ    ä xBCÿÀŽ äxCC€€€€:ÞŽ ä� xA&€   AøŽ ä xIFÿÀŽ â$xCC   B
Ž â xFF€€€€QÞŽ Ä x%G€€€€P´Ž ä xGJ   ÿ   â/ $xAAÿŽ ä $t&ÿÿ Ž â xGH€   GøŽ â $xFF   I
Ž â xBC€€€€:´Ž â $t:ÿ @€ÿ Ž â x&A€€€€&ìŽ â $xGGÿŽ â xQF€€€€Q´Ž Ä xDE˜º  ÿ    ä xF     ä x&&ÿþþþÿàÿ ä x:A€€€€:ìŽ ä xEDÿÀŽ äxCG€€€€FìŽ â $tHÿ @€ÿ Ž â xDD€€€€;ÞŽ Ä x&&AøŽ ä x::ÿþþþÿàÿ ä xFCÿþþþÿàÿ â $xDD   E
Ž â xC&˜º  ÿ    ä x:: @€AøŽ ä xEG€€€€HìŽ Ä x&CÿÀŽ äx::˜º  ÿ    ä xCC€€€€<ÞŽ äxHEÿþþþÿàÿ ä xFFGøŽ ä xA:ÿÀŽ â$xCC   &
Ž â x::€€€€=ÞŽ Ä xHH @€GøŽ ä xEF˜º  ÿ    â $x::   A
Ž â xHH˜º  ÿ    ä x&EÿÀŽ ä xCC€€€€<´Ž ä xEE€€€€>ÞŽ Ä xAHÿÀŽ â&rCC   ÿ   â xHH€€€€?ÞŽ ä x::€€€€=´Ž â &r%%   ÿ   â rEE&  €ÿàÿ â $xHH   A
Ž ä &r::   C   â xEE€€€€>´Ž â &r%Q   %   â xDD€€€€;´Ž â &r:B   :   â xHH€€€€?´Ž â &rEE    %   â x%Tð   ÿÀŽ â &r;D   :   â xTT   ÿÀŽ â &rHH!   E   â x%ÿ   %  Æ $rTT;   ÿŽ ä $x;;   HŽ ä $rT%H   TŽ â x%ÿ     ä r@@@   ÿ� ä r;;;   ÿ� ä z ÿ `  pPð Ä r%%   ÿ� ä r@@'   ÿú� ä rT;T   ÿú� â 0rcÿc    A   â z ÿ a   Sð â 0rhÿh    A   â x"ÿ   "  â 0r'ÿd    A   â x%ÿ   %  Ä x;ÿ   @  ä x=ÿ   T  â  rcch     A  â Er" "       â  rh'     A  â Er% %       ä r;;@   ÿ� ä r==T   ÿ� â 0r'ÿj    A   â #r$c"   $   â 0r?ÿZ    A   â #r#%   #   â x"ÿ   ;  Ä x%ÿ   =  â  rh'     A  â r'h?     A  â Er" "       ä Er% %       â G‰  	    € ê r>*   ÿàó Ê $x?+   Ž  Ê �yG>    ¨ �yF>    è �yI>    ( �yH>    â %xBJ   Ž Ì �yBB     â  x&G   ÿ   ÈO x&&ø  ÿÀŽ È z:&  ÿàó Ê $v;ÿ 	 ÿŽ  Ì �y::    › ¢ $x&F   ÿ Ž Ê� x&&ø  ÿÀŽ È z<&  ÿàó Ê $v=ÿ 	 ÿŽ  Ì �y<<    › â $x&I   ÿ Ž âx>H   ÿ   È x&&ø  ÿÀŽ ä x>>ø  ÿÀŽ ä z@&  ÿàó ä z>>  ÿàõ Æ $vAÿ 	 ÿŽ  ä $v?ÿ 	 ÿ È �y@@    › ( �y>>    › " rD.   ÿàó È rE/   ÿäÿ  Ê �y&DB    " xGÿ	   G  È 	sJ G        " $tKÿÿ Ž â xFÿ	   F  â $xJJ€   ÿ Ž Ê xJG€   JøŽ Ê $xJJÿŽ â 	sD F        (. xEJ€€€€KìŽ â $tKÿ @€ÿ Ž Æ xEEÿþþþÿàÿ ä xCJ€€€€KìŽ ä xEEJøŽ äxCCÿþþþÿàÿ ä xEE˜º  ÿ    ä xJC @€JøŽ Ä xCD   ÿ   ä xDEÿÀŽ ä xJJ˜º  ÿ    ä xIÿ	   I  ä xFF€   CøŽ ä xHÿ	   H  Ä xGJÿÀŽ â 	sC I        â $xFFÿŽ ä $tMÿÿ Ž Ê 	sK H        $ $xKK€   ÿ Ž Ê xKH€   KøŽ â $tHÿÿ Ž È $xKKÿŽ â xEE€€€€:ÞŽ äO xJJ€€€€;ÞŽ Æ $xEE   D
Ž ä $xJJ   G
Ž Æ xEE€€€€:´Ž ä x:F€€€€MìŽ ä xM  @€    ä xGJ€€€€;´Ž ä x;:ÿþþþÿàÿ ä x:F€€€€MìŽ â &rDE   ÿ   â x;;FøŽ Ä xE:ÿþþþÿàÿ â $x:C€   ÿ Ž â x;;˜º  ÿ    â &rGG   D   Æ x:I€   :øŽ ä xD;ÿÀŽ äx;;€€€€<ÞŽ â� $x::ÿŽ â xI     ä xFE @€FøŽ â $xE;   D
Ž â x;:€€€€IìŽ â $tIÿ @€ÿ Ž â xCK€€€€HìŽ â $tHÿ @€ÿ Ž â x;;ÿþþþÿàÿ Ä xDCÿþþþÿàÿ ä xC:€€€€IìŽ ä x;;:øŽ ä xCCÿþþþÿàÿ ä xHK€€€€HìŽ ä xDDKøŽ Ä x;;˜º  ÿ    ä xCC @€:øŽ ä xHHÿþþþÿàÿ ä xID˜º  ÿ    ä x:;ÿÀŽ ä xDC˜º  ÿ    Ä x;;€€€€@ÞŽ äxKH @€KøŽ ä xHIÿÀŽ äxII€€€€>ÞŽ ä xCDÿÀŽ â$x;;   :
Ž â xDD€€€€AÞŽ Ä rIIH  €ÿàÿ ä xHK˜º  ÿ    â $xDD   C
Ž â xFF˜º  ÿ    ä x;;€€€€@´Ž ä xKHÿÀŽ äxHH€€€€?ÞŽ Ä xCFÿÀŽ â&r;;   ÿ   â xFF€€€€=ÞŽ ä xDD€€€€A´Ž â $xHH   K
Ž â rFFC  €ÿàÿ ä xEE€€€€<´Ž â &rDD   ;   â xII€€€€>´Ž Ä xFF€€€€=´Ž â &rDE   D   â xHH€€€€?´Ž â &rGI    G   â x:&   ÿÀŽ â &r;F   D   ä &rGH!   G   â x&ÿ   &  â $r::;   ÿŽ ä $x;;   GŽ Ä $r&&G   :Ž Æ r;;;   ÿ� È r&;&   ÿú� È x;ÿ   &  È r;;&   ÿ� â 0r=ÿB    A   Æ x;ÿ   ;  ä  r=h=     A  ä Er; ;       Ê #r=;      ä G‰  	    € ê r>4   ÿàó Ê $x?5   Ž  Ê �yG>    ¨ �yF>    è �yH>    ( �yI>    â %xBJ   Ž Ì �yBB     â  $x&G   ÿ Ž ÊO x&&ø  ÿÀŽ È z:&  ÿàó È z;ÿ 	 ÿäÿ  Ì �y::    › ¢ $x&F   ÿ Ž Ê� x&&ø  ÿÀŽ È z<&  ÿàó Ê $v=ÿ 	 ÿŽ  Ì �y<<    › â $x>H   ÿ Ž âx&I   ÿ   È x>>ø  ÿÀŽ ä x&&ø  ÿÀŽ ä z>>  ÿàõ ä z@&  ÿàó Æ $v?ÿ 	 ÿ ä $vAÿ 	 ÿŽ  È �y>>    › ( �y@@    › " rD,   ÿàó Ê $xE-   Ž  Ê �y&DB    " xGÿ	   G  È 	sJ G        " xFÿ	   F  â $tKÿÿ Ž â xJJ   ÿ   È xJG€   JøŽ â 	sD F        (. $xJJÿŽ Ê xEJ€€€€KìŽ â $tKÿ @€ÿ Ž Æ xEEÿþþþÿàÿ ä xCJ€€€€KìŽ ä xEEJøŽ äxCCÿþþþÿàÿ ä xEE˜º  ÿ    ä xJC @€JøŽ â $xCD€   ÿ Ž â xDEÿÀŽ Ä xJJ˜º  ÿ    ä xFF€   CøŽ ä xIÿ	   I  ä xHÿ	   H  ä xGJÿÀŽ â 	sC I        â $xFFÿŽ Ä $tMÿÿ Ž Ê 	sK H        $ xKK   ÿ   È xKH€   KøŽ â $tHÿÿ Ž È $xKKÿŽ â xEE€€€€:ÞŽ äO xJJ€€€€;ÞŽ ä rEED  €ÿàÿ Æ $xJJ   G
Ž â xEE€€€€:´Ž ä x:F€€€€MìŽ â $tMÿ @€ÿ Ž â xGJ€€€€;´Ž ä x;:ÿþþþÿàÿ ä x:F€€€€MìŽ â &rDE   ÿ   â x;;FøŽ Ä xE:ÿþþþÿàÿ â $x:C€   ÿ Ž â x;;˜º  ÿ    â &rGG   D   Æ x:I€   :øŽ â $tIÿÿ Ž â xD;ÿÀŽ äx;;€€€€<ÞŽ â� $x::ÿŽ â xFE @€FøŽ ä rE;D  €ÿàÿ ä x;:€€€€IìŽ â $tIÿ @€ÿ Ž â xCK€€€€HìŽ â $tHÿ @€ÿ Ž Æ xDCÿþþþÿàÿ ä xC:€€€€IìŽ ä x;;ÿþþþÿàÿ ä xHK€€€€HìŽ ä xCCÿþþþÿàÿ ä xDDKøŽ Ä x;;:øŽ äxHHÿþþþÿàÿ ä xCC @€:øŽ ä xID˜º  ÿ    ä x;;˜º  ÿ    ä xKH @€KøŽ Ä xDC˜º  ÿ    ä xHIÿÀŽ äxII€€€€>ÞŽ äx:;ÿÀŽ äx;;€€€€@ÞŽ ä xCDÿÀŽ Ä xDD€€€€AÞŽ â $xII   H
Ž â r;;:  €ÿàÿ ä xHK˜º  ÿ    ä xFF˜º  ÿ    â $xDD   C
Ž â x;;€€€€@´Ž ä xKHÿÀŽ Ä xCFÿÀŽ ä xHH€€€€?ÞŽ ä xFF€€€€=ÞŽ â &r;;   ÿ   â xDD€€€€A´Ž â $xHH   K
Ž â xII€€€€>´Ž â $xFF   C
Ž â xEE€€€€<´Ž â &rDD   ;   â xHH€€€€?´Ž â &rGI    G   â xFF€€€€=´Ž â &rDE   D   â x:&   ÿÀŽ â &rGH!   G   Ä &r;F   D   â x&ÿ   &  Æ $r::;   ÿŽ â r;;G   ÿàÿ Æ $r&&G   :Ž â r;;;   ÿ� È r&;&   ÿú� È x;ÿ   &  È r;;&   ÿ� â 0r=ÿB    A   Æ x;ÿ   ;  ä  r&h=     A  ä Er; ;       Ê #r&;      ä G‰  	    € ê r>8   ÿàó Ê $x?9   Ž  Ê �yG>    ¨ �yF>    è �yI>    ( �yH>    â %xBJ   Ž Ì �yBB     â  $x&G   ÿ Ž ÊO x&&ø  ÿÀŽ È z:&  ÿàó Ê $v;ÿ 	 ÿŽ  Ì �y::    › ¢ x&F   ÿ   È� x&&ø  ÿÀŽ È z<&  ÿàó Ê $v=ÿ 	 ÿŽ  ä $x&I   ÿ Ž È�y<<    › â $x>H   ÿ Ž â x&&ø  ÿÀŽ È x>>ø  ÿÀŽ ä z@&  ÿàó ä z>>  ÿàõ Æ $vAÿ 	 ÿŽ  â z?ÿ 	 ÿä Ê �y@@    › ( �y>>    › " rD0   ÿàó Ê $xE1   Ž  Ê �y&DB    " xGÿ	   G  È 	sJ G        " xFÿ	   F  ä xK     â $xJJ€   ÿ Ž Ê xJG€   JøŽ â 	sD F        (. $xJJÿŽ Ê xEJ€€€€KìŽ â $tKÿ @€ÿ Ž Æ xEEÿþþþÿàÿ ä xCJ€€€€KìŽ ä xEEJøŽ äxCCÿþþþÿàÿ ä xEE˜º  ÿ    ä xJC @€JøŽ â $xCD€   ÿ Ž â xDEÿÀŽ Ä xJJ˜º  ÿ    ä xFF€   CøŽ ä xHÿ	   H  ä xIÿ	   I  ä xGJÿÀŽ â 	sK H        " $xFFÿŽ â xM     Ì 	sC I        b $xKK€   ÿ Ž Ê xKH€   KøŽ â $tHÿÿ Ž È $xKKÿŽ â xEE€€€€:ÞŽ äO xJJ€€€€;ÞŽ Æ $xEE   D
Ž ä $xJJ   G
Ž Æ xEE€€€€:´Ž ä x:F€€€€MìŽ â $tMÿ @€ÿ Ž â xGJ€€€€;´Ž ä x;:ÿþþþÿàÿ ä x:F€€€€MìŽ â &rDE   ÿ   â x;;FøŽ Ä xE:ÿþþþÿàÿ ä x:C   ÿ   ä/ x;;˜º  ÿ    ä x:I€   :øŽ â &rGG   D   â xD;ÿÀŽ â$tIÿÿ Ž â x;;€€€€<ÞŽ â� $x::ÿŽ â xFE @€FøŽ Æ $xE;   D
Ž â x;:€€€€IìŽ â$tIÿ @€ÿ Ž â xCK€€€€HìŽ ä x;;ÿþþþÿàÿ ä xH  @€    ä xDCÿþþþÿàÿ ä xC:€€€€IìŽ Ä x;;:øŽ ä xHK€€€€HìŽ ä xDDKøŽ ä xCCÿþþþÿàÿ ä x;;˜º  ÿ    ä xHHÿþþþÿàÿ Ä xID˜º  ÿ    ä xCC @€:øŽ ä x:;ÿÀŽ äxKH @€KøŽ ä x;;€€€€@ÞŽ äxHIÿÀŽ Ä xDC˜º  ÿ    ä xII€€€€>ÞŽ â $x;;   :
Ž â xCDÿÀŽ äxDD€€€€AÞŽ â $xII   H
Ž â xFF˜º  ÿ    ä xHK˜º  ÿ    Ä rDDC  €ÿàÿ ä x;;€€€€@´Ž ä xCFÿÀŽ ä xKHÿÀŽ ä xFF€€€€=ÞŽ ä xHH€€€€?ÞŽ â &r;;   ÿ   â xDD€€€€A´Ž â $xFF   C
Ž â xEE€€€€<´Ž â $xHH   K
Ž â xII€€€€>´Ž â &rDD   ;   â xFF€€€€=´Ž Ä xHH€€€€?´Ž â &rDE   D   ä &rGI    G   â x:&   ÿÀŽ â &r;F   D   ä &rGH!   G   â x&ÿ   &  â $r::;   ÿŽ ä $x;;   GŽ Ä $r&&G   :Ž Æ r;;;   ÿ� È r&;&   ÿú� È x;ÿ   &  È r;;&   ÿ� â 0r=ÿB    A   Æ x;ÿ   ;  ä  r=h=     A  ä Er; ;       Ê #r
=;   
   â #r"      â #r'%      â G‰  	    € ê r<6   ÿàñ È r=7   ÿä  Ê �yD<    ¨ �y%<    è �yF<    ( �yE<    â %x@J   Ž È $xD   ÿ Ž ÊO xø  ÿÀŽ È z&  ÿàñ Ê $v'ÿ 	 ÿ  Ì �y&&    › ¢ $x%   ÿ Ž Ê� xø  ÿÀŽ È z:  ÿàñ È z;ÿ 	 ÿä  â $xF   ÿ Ž Ê�y::    › â $x"E   ÿ Ž â xø  ÿÀŽ È x""ø  ÿÀŽ ä z>  ÿàñ ä z<"  ÿàó ä z?ÿ 	 ÿä  â �y"@     $ $v=ÿ 	 ÿŽ  Æ �y>>    › ( �y<<    › " rB2   ÿàñ Ê $xC3     Ê �yBB    " xDÿ	   D  È 	sG D        " x%ÿ	   %  â $tHÿÿ Ž ä $xGG€   ÿ Ž Ê xGD€   GøŽ â 	sB %        (. $xGGÿŽ Ê xCG€€€€HìŽ ä xH  @€    ä x@Cÿþþþÿàÿ ä xAG€€€€HìŽ ä x@@GøŽ ä xDAÿþþþÿàÿ â $xBB€   ÿ Ž â xA@˜º  ÿ    Ä xDD @€GøŽ ä xEÿ	   E  ä x@AÿÀŽ ä xFÿ	   F  â 	sG E        â xDD˜º  ÿ    ä xB%€   BøŽ â $tIÿÿ Ž â xCDÿÀŽ Æ 	s% F        " $xBBÿŽ ä $x%%€   ÿ Ž Ê x%F€   %øŽ Ê $x%%ÿŽ â xAA€€€€&ÞŽ äO xDD€€€€'ÞŽ Æ $xAA   @
Ž â rDDC  €ÿàÿ È xAA€€€€&´Ž ä x&B€€€€IìŽ ä xCD€€€€'´Ž ä x'&ÿþþþÿàÿ â $x@G€   ÿ Ž ä $tIÿ @€ÿ Ž â x''BøŽ â &rA   ÿ   â x@E€   @øŽ Ä x''˜º  ÿ    ä x&B€€€€IìŽ â &rCC      â x'ÿÀŽ â$x@@ÿŽ â xA&ÿþþþÿàÿ â $t&ÿÿ Ž â x''€€€€:ÞŽ ä� xE     Ä xBA @€BøŽ â $xA'   
Ž â x@€€€€EìŽ ä x%€€€€&ìŽ ä xD  @€    ä x'ÿþþþÿàÿ â $tEÿ @€ÿ Ž â xÿþþþÿàÿ Ä x%€€€€DìŽ ä x''@øŽ ä x%øŽ ä xÿþþþÿàÿ ä x&@€€€€EìŽ ä xE'˜º  ÿ    Ä x'˜º  ÿ    ä x @€%øŽ ä x'ÿÀŽ äx˜º  ÿ    ä x''€€€€>ÞŽ äxG&ÿþþþÿàÿ Ä x&EÿÀŽ ä x%ÿÀŽ â$x''   
Ž â xEE€€€€<ÞŽ ä x€€€€?ÞŽ ä xGG @€@øŽ ä xBB˜º  ÿ    â $xEE   &
Ž â x&G˜º  ÿ    â $x   %
Ž â x''€€€€>´Ž Ä x%BÿÀŽ äxBB€€€€;ÞŽ â &r''   ÿ   â xG&ÿÀŽ ä x€€€€?´Ž ä x&&€€€€=ÞŽ â $xBB   %
Ž â xAA€€€€:´Ž â &r   '   â r&&G  €ÿàÿ Ä xEE€€€€<´Ž â &rA      â xBB€€€€;´Ž ä x&&€€€€=´Ž â &rCE    C   â x   ÿÀŽ â &rB      ä &r&&!   C   â xÿ     â $r   ÿŽ Ä $x   &Ž ä $r&   Ž Æ r   ÿ� È r   ÿú� È xÿ     È r   ÿ� â 0rÿ"    A   Æ xÿ     ä  rhh     A  ä Er        Ê #r	h   	   â x   ÿàÿ ä x   ÿàõ ä |    pbð ä x   ÿàó â $rÿÿ    â x€   ÿàÿ ä rÿ   ÿäÿ  Î D	       À â Gy  °µÿÿÿÿƒ ê Ay        € ê y      "   " Ey  0    € â r ÿ   pRð är ÿ   pRò Ö G‰  ð     € ê z ÿ `  pPð â $x€    Ž Æ z ÿ a   Sð ä ˆs $ þÿ H   ö ˆƒ #€þÿ H   è ˆƒ  ÿÿ H   è ˆƒ €ÿÿ H   â G‰  p     € ê ˆs    H   è ˆs #€þÿ H   è ˆs €  H   è ˆs  ÿÿ H   è ˆs 
   H   è ˆs €ÿÿ H   è ˆs 	€  H   ä Ay        € ê {          ì M        € ê y      &   "> z ÿ `  pPð Ä x        â y      '   b z ÿ a   Sð â $z x  Ž È $z ~  Ž È/ %v j   Ž È G	  P    € ê „y	      H   " z  r  p`ð Æ „y  €   H   br  ÿ   pVp  Æ „y     H   ¨ „y     H   è „y  €  H   ( „y     H   ( „y     H   ( „y  €  H   ( „y     H   ( „y  €  H   ( „y
  €  H   " !r		$       Æ „y  €  H   " !r#       â/ !r       âO !r	       ä� !r       â!r       â !r       ä !r       Æ ‰      b !r       Æ ‰      ¢ !r       Æ ‰	      â !r

       È !r
       Ê ‰      " !r       â/ !r       ÈO ‰      b !r
	       Æ� ‰	      ¨ ‰
      â !r       Ê ‰      " !r       â/ !r	       ÈO ‰	 €    b !r
       Æ� ‰ €    ¨ ‰ €    â !r       Ê ‰ €    " !r		       â/ !r       ÈO ‰	 @    b x   ÿàÿ â !r
       ä� ‰ @    ¢ z  r  p`ò Æ ‰
 @    â x     pVò  â !r       Ê ‰ @    " !r	       â/ x	   ÿàÿ äx   ÿàÿ â !r       âO z 	 r  p`ô â !r
       Æ� ‰      b x     pVt Æ ‰      ¨ ‰      â !r       â !r       â/ !r       ÆO !’	       â‰      " !‚       â !r       ä� †™ 	   ä !¢       ä †‰     â z  r  p`ð Æ †©    â x     pVp  â !r       Ø M	        € ê !r       Ê/ †y    â My        € ê $vÿ h  ÿ Ž Ê x    pBð Ú G	  @    € ê x    pRð Ú G‰  À    € ê x    pRð Ú G	   $    € ê „y      H   " z  r  p`ð â Ey  p    € ä „y     H   b r  ÿ   pVp  Æ „y     H   ¨ „y     H   è „y     H   ( „y   
  H   ( „y  €   H   ( „y  €  H   ( „y  €  H   ( „y  €  H   " !r$       Æ „y  €
  H   â !r       â/ !r       ÆO !r       â� !r       Æ!r       ä ‰      " !r#       Æ ‰      b !r       â !r       È !r       â !r       ä „y  €  H   â !r       Æ/ ‰      ( ‰      b !r       â !r       È/ ‰ €    ( ‰ €    b !r       È !r       â !r       â !r       È/ ‰ @    ( ‰ @    b !r       â !r       È/ ‰      ( ‰      b !r       â !r       â/ G	  �     € ê !r       â!r       Æ  x;ª¸¿  A  Ì s        $ !t  €?    Ì s        $  r     A  È  r     A  Ê †y     ä Ay        € ê ‰      b Ey  Ð    € æ ‰      ¨ „y     H   è „y     H   è „y     H   è „y     H   â !r       â/ !r       ÈO ‰      h ‰      ( „y     H   ¢ !r
       â� !r       â/ !r       È ‰ €    ( ‰ €    h „y   	  H   â !r       ÈO !r       â !r       Æ !r       â !r       ä/ ‰ @    ( ‰ @    b !r       È� !r       â !r       â x   ÿàÿ â !r       Æ/ ‰      " z  r  p`ð Æ ‰      b x     pVp  â !r       â !r       Ö/ G	  �     € ê !r       â!r       Æ  x;ª¸¿  A  Ì s        $ !t
  €?    Ì s
 
       $  r
     A  È  r     A  Ê †y    ä Ay        € ê ‰      b Ey       € æ ‰      ( „y  €	  H   è „y  €  H   è „y  €  H   è „y  €  H   â !r       Æ/ „y  €  H   â !r
       Æ ‰      ( ‰
      h „y  €  H   ¢ !r       â !r
       È/ ‰ €    ( ‰ €    b !r	       ÈO !r       â !r       â x   ÿàÿ ä !r       â !r       â/ ‰ @    " z  r  p`ð Æ ‰ @    b x     pVp  â !r       â !r       È/ ‰      ( ‰      b !r       â !r       â/ G	  �     € ê !r       â!r       Æ  x;ª¸¿  A  Ì s        $ !t	  €?    Ì s
 	       $  r
     A  È  r     A  Ê †y    ä Ay        € ê ‰      " !r       â x   ÿàÿ Æ !r       â z  r  p`ð Æ !r       â x     pVp  È ‰      b !r       Ê ‰	      " !r       Ê/ ‰      b !r
	       Ê ‰	
 €    " !r       Ê/ ‰ €    b !r
	       Ê ‰ @    " !r       Ê/ ‰	 @    b !r       Ê ‰      " !r		       Ê/ ‰ 	      ¢ !r       â M	        € ì !r       â/!r 	        ÆO  x;ª¸¿  A  â !r        Ê s        $ !t  €?    Ì s        $  r     A  È  r     A  Ê †y    â My        € ê „y      H   " z  r  p`ð â Ey       € ä „y     H   b r  ÿ   pVp  Æ „y     H   ¨ „y     H   è „y     H   ( „y   
  H   ( „y  €   H   ( „y  €  H   ( „y  €  H   ( „y  €  H   â !r$       â !r       Æ/ !r       âO !r       Æ� !r       â!r       È ‰      " !r#       Æ ‰      b !r       È !r       â !r       â !r       È/ ‰      ( ‰      h „y  €  H   ¢ !r       â !r       È/ ‰ €    ( ‰ €    b !r       ÈO !r       â !r       ä „y  €
  H   â !r       Æ/ ‰ @    ( ‰ @    b !r       â !r       È/ ‰      ( ‰      b !r       â !r       â !r       â/ G	  À     € ê !r       È x'7=  A  â x*BL?  A  Æ #t  €?   È  r     A  â  x   ?  A  â !r       È s     $   $ !t  €?    È  r     A  È  r     A  Ê †y     ä Ay        € ê ‰      b Ey       € æ ‰      ¨ „y     H   è „y     H   ( „y     H   è „y     H   â !r       â/ !r       ÈO ‰      h ‰      ( „y     H   ¢ !r       â� !r
       â!r       â/ !r       È ‰ €    ( ‰ €    h „y   	  H   â !r       ÈO !r       â !r       â !r       È/ ‰ @    ( ‰ @    b !r       È� !r       â !r       â x   ÿàÿ â !r       Æ/ ‰      " z  r  p`ð Æ ‰      b x     pVp  â !r       â !r       Ö/ G	  À     € ê !r       È x'7=  A  â x*BL?  A  Æ #t  €?   È  r     A  â  x   ?  A  â !r       È s     $   $ !t
  €?    È  r
     A  È  r     A  Ê †y    ä Ay        € ê ‰      b Ey  Ð    € æ ‰      ( „y  €  H   ¨ „y  €	  H   è „y  €  H   è „y  €  H   â !r       Æ/ „y  €  H   â !r
       Æ ‰      ( ‰
      h „y  €  H   â !r	       ÈO !r       â� !r       Æ !r       â !r
       â/ ‰ €    ( ‰ €    b !r       â !r       È/ ‰ @    ( ‰ @    b !r       â x   ÿàÿ â !r       Æ/ ‰      " z  r  p`ð Æ ‰      b x     pVp  â !r       â !r       Ö/ G	  À     € ê !r       È x	'7=  A  â x*BL?  A  Æ #t  €?	   È  r     A  â  x   ?  A  â !r       È s     $   $ !t
  €?    È  r
     A  È  r     A  Ê †y    ä Ay        € ê ‰      " !r       â x   ÿàÿ Æ !r       â z  r  p`ð Æ !r       â x     pVp  È ‰      b !r       Ê ‰	      " !r       Ê/ ‰      b !r
	       Ê ‰	
 €    " !r       Ê/ ‰ €    b !r
	       Ê ‰ @    " !r       Ê/ ‰	 @    b !r       Ê ‰      " !r		       Ê/ ‰ 	      ¢ !r       â M	        € ì !r       â/!r 	        ÆO  x'7=  A  â x*BL?  A  â !r        ä #t  €?   â x   ?  A  Æ  r     A  Ì s     $   $ !t  €?    È  r     A  È  r     A  Ê †y    â My        € ê x    pRð Ú G‰  P    € ê x    pRð Ú G	  `    € ê „y      H   " z  r  p`ð â Ey       € ä „y     H   b r  ÿ   pVp  Æ „y     H   ¨ „y     H   è „y     H   ( „y   
  H   ( „y  €   H   ( „y  €  H   ( „y  €  H   ( „y  €  H   â !r$       â !r       Æ/ !r       âO !r       Æ� !r       â!r       È ‰      " !r#       Æ ‰      b !r       È !r       â !r       â !r       È/ ‰      ( ‰      h „y  €  H   ¢ !r       â !r       È/ ‰ €    ( ‰ €    b !r       ÈO !r       â !r       ä „y  €
  H   â !r       Æ/ ‰ @    ( ‰ @    b !r       â !r       È/ ‰      ( ‰      b !r       â !r       â !r       â/ G	  À     € ê !r       Ê	z i    � â !r       È  x;ª¸¿  A  â 	z i    � È 	z i €  � â s        $ !t  €?    Ì s        $  r     A  È  r     A  Ê †y     ä Ay        € ê ‰      b Ey       € æ ‰      ¨ „y     H   è „y     H   ( „y     H   è „y     H   â !r       â/ !r       ÈO ‰      h ‰      ( „y     H   ¢ !r       â� !r
       â!r       â/ !r       È ‰ €    ( ‰ €    h „y   	  H   â !r       ÈO !r       â !r       â !r       È/ ‰ @    ( ‰ @    b !r       È� !r       â !r       â x   ÿàÿ â !r       Æ/ ‰      " z  r  p`ð Æ ‰      b x     pVp  â !r       â !r       Ö/ G	  À     € ê !r       Ê	z i    � â !r       È  x;ª¸¿  A  â 	z i    � È 	z i €  � â s        $ !t
  €?    Ì s
 
       $  r
     A  È  r     A  Ê †y    ä Ay        € ê ‰      b Ey  Ð    € æ ‰      ( „y  €  H   ¨ „y  €	  H   è „y  €  H   è „y  €  H   â !r       Æ/ „y  €  H   â !r
       Æ ‰      ( ‰
      h „y  €  H   â !r	       ÈO !r       â� !r       Æ !r       â !r
       â/ ‰ €    ( ‰ €    b !r       â !r       È/ ‰ @    ( ‰ @    b !r       â x   ÿàÿ â !r       Æ/ ‰      " z  r  p`ð Æ ‰      b x     pVp  â !r       â !r       Ö/ G	  À     € ê !r       Ê	z i    � â !r       È  x;ª¸¿  A  â 	z i    � È 	z i €  � â s        $ !t	  €?    Ì s
 	       $  r
     A  È  r     A  Ê †y    ä Ay        € ê ‰      " !r       â x   ÿàÿ Æ !r       â z  r  p`ð Æ !r       â x     pVp  È ‰      b !r       Ê ‰	      " !r       Ê/ ‰      b !r
	       Ê ‰	
 €    " !r       Ê/ ‰ €    b !r
	       Ê ‰ @    " !r       Ê/ ‰	 @    b !r       Ê ‰      " !r		       Ê/ ‰ 	      ¢ !r       â M	        € ì !r       â/!r 	        ÈO 	z i    � â !r        È  x;ª¸¿  A  â 	z i    � È 	z i €  � â s        $ !t  €?    Ì s        $  r     A  È  r     A  Ê †y    â My        € ê „y     H   " z  r  p`ð Æ „y  €   H   b r  ÿ   pVp  Æ „y      H   ¨ „y     H   è „y  €  H   ( „y     H   ( „y     H   ( „y  €  H   ( „y     H   ( „y!     H   ( „y  €  H   " !r       Æ „y   	  H   " !r#       Æ/ „y   
  H   b !r$       ÆO „y  €  H   ¢ !r       Æ� „y     H   â !r       Æ„y%     H   " !r
       Æ „y  €
  H   " !r       Æ „y'     H   " !r       Æ „y)  €  H   " !r       Æ „y  €  H   " !r!       Æ „y  €  H   " !r       Æ „y+  €	  H   â !r       Æ „y-  €  H   â !r       Æ/ „y/  €  H   â !r       âO !r       ä� ‰      " !r%       Æ‰      b !r       Æ ‰      ¢ !r'       Æ ‰      â !r))       Æ ‰
      " !r	       Æ ‰      " !r)       â !r       â !r+       â !r       Æ/ ‰	      " !r--       â !r       äO ‰      b !r//       â !r       ä� ‰-      ¢ x   ÿàÿ â !r

       Ä‰/      â z  r  p`ô â !r       ä ‰      " x   ÿàÿ äx     pVt â ‰
      " z  r  p`ò ä x   ÿàÿ â ‰      " !r		       â x     pVò  Ä ‰      " !r       â/ z  r  p`ö â !r-       âO ‰	 €    d x     pVö â !r/       â� ‰      ¢ !r       Æ‰      â !r
       Æ ‰ €    " !r       Æ ‰ €    " !r       Æ ‰
 €    ( ‰ €    " !r	       Æ/ ‰ €    b !r       âO !r       ä� ‰	 @    ¢ !r       Æ‰ €    â !r       Æ ‰ €    " !r

       Æ ‰ @    " !r       Æ ‰ @    " !r       Æ/ ‰
 @    h ‰ @    b !r		       ÆO ‰ @    ¢ !r       â� !r       ä‰	      â !r       Æ ‰ @    " !r       Æ ‰ @    " !r
       Æ/ ‰      b !r       Æ ‰      b !r       ÆO ‰
      ¨ ‰      ¢ !r	       Æ� ‰      â !r       â !‚       â!r       Æ‰      " !r       Æ/ ‰      b !r       â !‚       â!r
       äO !’       â  ‚     A  â !r       â !’       Æ †‰     â !r	       â� !¢       â  ’     A  ä !¢		       â !r       ä †™    â  ¢		     A  â !r       È/ †© 	   â M9        € ê !r       â !r       È  r     A  Ê †y    â My        € ê „y      H   " z  r  p`ð â Ey  À    € ä „y     H   b r  ÿ   pVp  Æ „y     H   ¨ „y     H   è „y     H   ( „y   
  H   ( „y  €   H   ( „y  €  H   ( „y  €  H   ( „y  €  H   â !r$       â !r       Æ/ !r       âO !r       Æ� !r       â!r       È ‰      " !r#       Æ ‰      b !r       È !r       â !r       â !r       È/ ‰      ( ‰      h „y  €  H   ¢ !r       â !r       È/ ‰ €    ( ‰ €    b !r       ÈO !r       â !r       ä „y  €
  H   â !r       Æ/ ‰ @    ( ‰ @    b !r       â !r       È/ ‰      ( ‰      b !r       â !r       â !r       â/ G	  à     € ê !r       Ê	x  à@  � Ê  x#ÛÙ?  A  È  x;ª¸¿  A  â !r       Ê s        " 	x  à@  � È 	x  àÀ  � Ê !t  €?    â !t  €?    È s        $  r     A  È  r     A  Ê †y     ä Ay        € ê ‰      b Ey       € æ ‰      ¨ „y     H   è „y     H   ( „y     H   è „y     H   â !r       â/ !r       ÈO ‰      h ‰      ( „y     H   ¢ !r       â� !r
       â!r       â/ !r       È ‰ €    ( ‰ €    h „y   	  H   â !r       ÈO !r       â !r       â !r       È/ ‰ @    ( ‰ @    b !r       È� !r       â !r       â x   ÿàÿ â !r       Æ/ ‰      " z  r  p`ð Æ ‰      b x     pVp  â !r       â !r       Ö/ G	  à     € ê !r       Ê	x  à@  � Ê  x#ÛÙ?  A  È  x;ª¸¿  A  â !r       Ê s        " 	x  à@  � È 	x  àÀ  � Ê !t  €?    â !t
  €?    Ì s
 
       $  r
     A  È  r     A  Ê †y    ä Ay        € ê ‰      b Ey  ð    € æ ‰      ( „y  €  H   ¨ „y  €	  H   è „y  €  H   è „y  €  H   â !r       Æ/ „y  €  H   â !r
       Æ ‰      ( ‰
      h „y  €  H   â !r	       ÈO !r       â� !r       Æ !r       â !r
       â/ ‰ €    ( ‰ €    b !r       â !r       È/ ‰ @    ( ‰ @    b !r       â x   ÿàÿ â !r       Æ/ ‰      " z  r  p`ð Æ ‰      b x     pVp  â !r       â !r       Ö/ G	  à     € ê !r       Ê	x  à@  � Ê  x#ÛÙ?  A  È  x;ª¸¿  A  â !r       Ê s        " 	x  à@  � È 	x  àÀ  � Ê !t  €?    â !t	  €?    È s
 	       $  r
     A  È  r     A  Ê †y    ä Ay        € ê ‰      " !r       â x   ÿàÿ Æ !r       â z  r  p`ð Æ !r       â x     pVp  È ‰      b !r       Ê ‰	      " !r       Ê/ ‰      b !r
	       Ê ‰	
 €    " !r       Ê/ ‰ €    b !r
	       Ê ‰ @    " !r       Ê/ ‰	 @    b !r       Ê ‰      " !r		       Ê/ ‰ 	      ¢ !r       â M	        € ì !r       â/!r 	        ÈO 	x  à@  � â !r        È  x#ÛÙ?  A  â 	x  à@  � Æ  x;ª¸¿  A  â 	x  àÀ  � Ê s        " !t  €?    â !t  €?    Ì s        $  r     A  È  r     A  Ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y       "   $ z   w  p`ð Ú M	        € ê y      !   " ¹z  l      â Ey  p    € â ™x?    â y      %   b ¹z  F   
   â $rÿÿ   ÿ Ž ä $rÿÿ   ÿ Ž â xÿ     È |    p`ð â $x   ÿ Ž Ø/ G	  Ð    € ê y	      &   " $tÿ   ÿ Ž ä $z  x  	Ž È %v \   Ž Ì �y    ™ ¢  $rÿÿ   ÿ Ž â x   ÿÀŽ â$x   ÿ Ž ä 'z	 m   Ž Æ x8   ÿÀŽ â $z
 q  Ž â zÿ n    â $rÿÿ   ÿ Ž ä $rÿÿ   ÿ Ž ä $rÿÿ   
Ž ä $vÿ X  ÿ Ž ä $z o  	Ž â x	   ÿàÿ â $z  r  ÿŽ Ä $z u  ÿŽ ä $vÿ Y  ÿ Ž Æ r   ÿàñ â $x   ÿ Ž ä $z t  Ž äO$z t  
Ž ä $z		 q  Ž ä %xJ   Ž È $rÿÿ   ÿ Ž ä %xJ   Ž È %x	J   Ž È %x	J   Ž È $r
ÿÿ   ÿ  ä r   ÿàñ Ê $x     Ê �y   • ¨  �y   • è  �y?   • ( �y   • b r    ÿàñ Ê $x!     Ê �y    • h �y=    • h �y    • h �y9    • b xÿ     â $t!ÿ$   ÿ Ž Ä/ $x   ÿ Ž ÊO xø  ÿÀŽ È z,  ÿàñ â $x   ÿ Ž È� $v-ÿ 	 ÿ  â xø  ÿÀŽ Ê �y,,    › ¢ z.  ÿàñ â $x?   ÿ Ž È$v/ÿ 	 ÿ  â xø  ÿÀŽ È z*  ÿàñ â $x   ÿ Ž â�y..    › æ $v+ÿ 	 ÿ  â r   à‘ ä xø  ÿÀŽ ä r ÿ
   $~  â �y**    › " z(  ÿàñ â %v Z  ! Ž È $v)ÿ 	 ÿ  ä $x $   ÿŽ È �y((    › b $x   Ž Ê �y@   ™ h �y;   ™ h �y>   ™ h �y:   ™ h �y<   ™ h �y   ™ h �y8   ™ b $x    ÿ Ž â r2   ÿàó Ä �yA    ™ d x  ø  ÿÀŽ ä �yC    • d z&   ÿàñ â $x =   ÿ Ž È $v'ÿ 	 ÿ  â x$ ø  ÿÀŽ â $x    ÿ Ž È �y&&    › b z$$  ÿàñ ä x  ø  ÿÀŽ Æ $v%ÿ 	 ÿ  â z"   ÿàñ â $x 9   ÿ Ž È �y$$    › b x  ø  ÿÀŽ â $v#ÿ 	 ÿ  Æ z    ÿàñ Æ �y""    › d $v!ÿ 	 ÿ  Ì �y      › b r0   ÿàñ â $x3   Ž  È $x1     â�y22B   ‘ h �yB0B   ‘ d $r0ÿÿ    Ž ä $r1ÿÿ    Ž È %x4J   0Ž È %x6	J   0Ž ä �y44    • h �y66    • b xDÿ	     È 	s D        b x?ÿ	   ?  ä xÿ	     Ê 	s1 ?        " $x3€   ÿ Ž Î/ 	s0         b xÿ	     ä xD€   3øŽ Ê 	s         b $xÿŽ ä $tÿÿ Ž ä $t3ÿ @€ÿ Ž ä $xD1€   ÿ Ž â x1€€€€ìŽ â $x0€   ÿ Ž â/ x3€€€€3ìŽ Ä x11ÿþþþÿàÿ ä x33ÿþþþÿàÿ â $x0€   ÿ Ž â x€   øŽ ä x??€   DøŽ ä x1øŽ â$t1ÿÿ Ž â x3 @€øŽ â $xÿŽ â x€   0øŽ â $x??ÿŽ Ä $t0ÿÿ Ž ä $tDÿ @€ÿ Ž ä $t7ÿ @€ÿ Ž â x€€€€1ìŽ äx3?€€€€0ìŽ äx5?€€€€DìŽ ä x0€€€€7ìŽ Ä x1ÿþþþÿàÿ ä x3ÿþþþÿàÿ ä xD5ÿþþþÿàÿ ä x30ÿþþþÿàÿ ä x11øŽ ä x0D @€?øŽ Ä xD3 @€øŽ â $xÿŽ â x31˜º  ÿ    â $tÿÿ Ž ä $t5ÿ @€ÿ Ž â x13ÿÀŽ ä x€€€€ìŽ äx5€€€€5ìŽ Ä xÿþþþÿàÿ ä xDD˜º  ÿ    ä x55ÿþþþÿàÿ ä xøŽ äx5 @€øŽ ä x5˜º  ÿ    Ä x?øŽ ä x00˜º  ÿ    ä x=ÿ	   =  ä xÿ	     ä x33€€€€,ÞŽ ÊO $x13   1
Ž â x3DÿÀŽ äxDD€€€€-ÞŽ ä x1€€€€,´Ž ä x,˜º  ÿ    â $xDD   3
Ž â x5ÿÀŽ ä x3,ÿÀŽ Ä x,,€€€€/ÞŽ ä� x15€€€€.ÞŽ ä x5˜º  ÿ    â $x,   3
Ž â x30ÿÀŽ â $x11   
Ž â x,5ÿÀŽ â	s =        " x55€€€€*ÞŽ Äx00€€€€+ÞŽ ä x/€€€€/´Ž â $x55   ,
Ž ä $x00   3
Ž â x3˜º  ÿ    ä x˜º  ÿ    ä x*5€€€€*´Ž Ä x3ÿÀŽ äx33€€€€(ÞŽ äx5ÿÀŽ äx€€€€)ÞŽ â 	s         b xÿ	     ä x9ÿ	   9  â $x33   
Ž Ä $x,   5
Ž ä 	s         ¢ $x€   ÿ Ž â x-D€€€€-´Ž â &r@   ÿ   Ê 	s 9        " x==€   øŽ â &r**;   ÿ   â x+0€€€€+´Ž â &r->      â x11€€€€.´Ž â $x==ÿŽ â x(3€€€€(´Ž â $t0ÿ @€ÿ Ž â x,€€€€)´Ž â &r*+:   *   Ä &r1<      ä $t.ÿÿ Ž ä $x,€   ÿ Ž â/ x)=€€€€0ìŽ â &r(   *   ä &r//8      â x=€€€€.ìŽ â $x(€   ÿ Ž âO x+€   ,øŽ â $x€   ÿ Ž â x,)ÿþþþÿàÿ Ä x*ÿþþþÿàÿ ä x(€   (øŽ â $x++ÿŽ â x99€   øŽ â $tÿÿ Ž â x, @€=øŽ â$t,ÿ @€ÿ Ž â x*=øŽ â $x(ÿŽ ä $t*ÿÿ Ž â x+€€€€ìŽ Ä x(+€€€€,ìŽ ä x€€€€*ìŽ äxÿþþþÿàÿ ä x*(ÿþþþÿàÿ ä x)€€€€,ìŽ ä x(ÿþþþÿàÿ Ä x+øŽ äx* @€+øŽ ä x*)ÿþþþÿàÿ ä x)(øŽ äx** @€øŽ â $x(9ÿŽ â x))˜º  ÿ    â $t+ÿÿ Ž â x,*˜º  ÿ    â $t-ÿ @€ÿ Ž â x*)ÿÀŽ Ä x+(€€€€+ìŽ ä x))€€€€&ÞŽ ä x,ÿÀŽ äx,,€€€€'ÞŽ ä x-(€€€€-ìŽ â $x))   *
Ž â x++ÿþþþÿàÿ â $x,,   
Ž â x--ÿþþþÿàÿ Ä x+(øŽ ä x+˜º  ÿ    ä x))€€€€&´Ž ä x&- @€(øŽ ä x(˜º  ÿ    ä x+ÿÀŽ Ä x++€€€€$ÞŽ ä x,€€€€'´Ž ä x˜º  ÿ    ä x'(ÿÀŽ â$x++   
Ž â x((€€€€%ÞŽ ä x˜º  ÿ    Ä xÿÀŽ äx€€€€"ÞŽ â $x((   '
Ž â x'ÿÀŽ äx˜º  ÿ    â $x   
Ž â x€€€€#ÞŽ ä x&&˜º  ÿ    Ä xÿÀŽ â$x   '
Ž â x€€€€ ÞŽ ä x"€€€€"´Ž ä x&ÿÀŽ â$x   
Ž â x&&€€€€!ÞŽ â &r))@   ÿ   â x#€€€€#´Ž â &r"";   ÿ   â x++€€€€$´Ž â $x&&   
Ž â x€€€€ ´Ž â &r>   )   Ä &r"#:   "   â x%(€€€€%´Ž â &r+<      â x&&€€€€!´Ž â &rA      â xBð   ÿÀŽ â &r   "   â x2ð   ÿÀŽ â &r%8      â xBB   ÿÀŽ ä x22   ÿÀŽ â &r&&A      â xÿ     â $rBB/   ÿŽ â xÿ     â $r22   ÿŽ Ä $x//   Ž ä $x   &Ž ä $r   BŽ â r//   ÿ� â $r&   2Ž â r   ÿ� ä r   ÿú� ä r   ÿú� Ä x   ÿàÿ ä xÿ     ä xÿ     ä |    p`ð ä r   ÿ� ä r   ÿ� â 0r4ÿ4    A   â xÿ     â 0rCÿC    A   â xÿ     â 0r6ÿ6    A   â x(  ÿàó Ä  r4C     A  â Er        â  rCC6     A  â Er        â $rÿÿ   Ž  â x    ÿàÿ â #r      ä #rC      â D	       À â Gy  0ìÿÿÿÿƒ ê Ay        € ê ‰      " $x   Ž Æ ‰      d z  p  p`ð È x    pDp  â !r       â !r       È/ ‰      ( ‰      b !r       â !r	       È/ ‰ €    ( ‰
	 €    b !r       â !r
	
       È/ ‰ @    ( ‰
 @    b !r       â !r
       È/ ‰      h  ‰      ¢  M	        € ê y      &   â $x   ÿ Ž â !r       â/ $z   s  Ž â !r       âO $t	ÿ   ÿ Ž â r ÿ   p"ð Ä x    p"ò Ö $ÿÿ    Ž ä $ÿÿ    Ž ä $z  v   Ž È� %v  j  	 Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      "   $ z  w  p`ð Ú M	        € ê y	      &   " $tÿ   ÿ Ž â ¹z  F   
   ä y      !   b ¹z  l      ä ™x?    â $z x  	Ž È %v \   Ž Ê �y     ™ b Ey  à(    € â $rÿÿ   ÿ Ž â x      ÿ  â $r
ÿÿ   ÿ Ž â x      ÿ  â xÿ     È/ |    p`ð Ú G	  `(    € ê y      %   " $rÿÿ   ÿ Ž â x   ÿÀŽ â$x   ÿ Ž â Ey ð'    € â 'z	 m   Ž Æ x8   ÿÀŽ â $z r  ÿŽ â zÿ n    â $rÿÿ   ÿ Ž ä $rÿÿ   ÿ Ž ä $rÿÿ   
Ž ä $v
ÿ X  ÿ Ž ä $z o  	Ž Ä $z	  t  Ž ä$z u  ÿŽ ä $x   ÿ Ž ä $vÿ `  ÿ Ž â r   ÿàñ â $z q  Ž âx   ÿàÿ â $vÿ Y  ÿ Ž Ä $z  t  Ž ä $z q  	Ž ä $vÿ a  ÿ Ž ä %xJ   Ž È $rÿÿ   ÿ Ž ä $x	   ÿ Ž ä %xJ   Ž È %xJ   Ž È %xJ   Ž È $rÿÿ   ÿ  ä r
   ÿàñ Ê $x     Ê �y<   • ¨ �y;   • è �yB   • ( �y@   • b	 r
   ÿàñ Ê $x     Ê �y?   • h �y=   • h �y>   • h �yA   • b $tÿ$   ÿ Ž Ä %x4J   
Ž È %x6J   
Ž ä �y44    • h �y66    • b $x<   ÿ Ž ÊO xø  ÿÀŽ È z,  ÿàñ â $x;   ÿ Ž È� $v-ÿ 	 ÿ  â xø  ÿÀŽ â $xB   ÿ Ž Æz.  ÿàñ â �y,,    › ¢ xø  ÿÀŽ Æ $v/ÿ 	 ÿ  â z(  ÿàñ â $x@   ÿ Ž È$v)ÿ 	 ÿ  â �y..    › â x*ø  ÿÀŽ È �y((    › " z**  ÿàõ ä xÿ   	  är   	à‘ â $v+ÿ 	 ÿ Æ rÿ   $~  â %v Z   Ž ä �y**    › d $x$   ÿŽ ä $x?   ÿ Ž ä $x   Ž Æ x"ø  ÿÀŽ ä �y   ™ d z""  ÿàñ ä �y   ™ f $v#ÿ 	 ÿ  â �y   ™ h �y   ™ h �y""    › b $x =   ÿ Ž Æ �y8   ™ h �y9   ™ b x  ø  ÿÀŽ Æ �y:   ™ b z    ÿàñ â $x%>   ÿ Ž ä $x$A   ÿ Ž â �yC    ™ b $v!ÿ 	 ÿ  â x%%ø  ÿÀŽ ä �yD    • h �y      › b x&$ø  ÿÀŽ Ä z$%  ÿàó ä z&&  ÿàñ Æ $v%ÿ 	 ÿŽ  ä $v'ÿ 	 ÿ  È �y$$    › h �y&&    › b r0
   ÿàñ är2
   ÿàó Æ $x1     ä$x3   Ž  Æ �y00B   ‘ h �y22B   ‘ b x;ÿ	   ;  È 	s1 ;        " xBÿ	   B  ä x<ÿ	   <  Ê 	sE B        0 	s3 <        b. $x1€   ÿ Ž â x@ÿ	   @  È x;€   øŽ ä 	s1 @        " $xEE€   ÿ Ž ä $xÿŽ ä $t5ÿÿ Ž ä $x33€   ÿ Ž ä/ $tÿ @€ÿ Ž â xBB€   EøŽ Ä x<<€   3øŽ ä x3€€€€5ìŽ äx5€€€€ìŽ â $tÿÿ Ž â x33ÿþþþÿàÿ â $xBBÿŽ â x;5ÿþþþÿàÿ â $x11€   ÿ Ž â x73øŽ â$t5ÿ @€ÿ Ž â x; @€øŽ â $x<<ÿŽ â xB€€€€ìŽ â $tEÿÿ Ž â x3@€   1øŽ â $t;ÿ @€ÿ Ž â x1B€€€€5ìŽ Ä x5ÿþþþÿàÿ ä x<€€€€EìŽ äx;<€€€€;ìŽ ä xEÿþþþÿàÿ ä x;ÿþþþÿàÿ ä xEE<øŽ Ä x< @€<øŽ ä xEE˜º  ÿ    ä x11ÿþþþÿàÿ ä x<<˜º  ÿ    ä x;EÿÀŽ â $x33ÿŽ â x55BøŽ â$t@ÿ @€ÿ Ž â x11 @€BøŽ â $tBÿÿ Ž â x<ÿÀŽ Ä x?ÿ	   ?  ä xB3€€€€BìŽ äx@3€€€€@ìŽ ä x77˜º  ÿ    ä xBBÿþþþÿàÿ ä x@@ÿþþþÿàÿ Ä x55˜º  ÿ    ä xBB3øŽ äx@@ @€3øŽ ä x3˜º  ÿ    ä x5ÿÀŽ ä x@@˜º  ÿ    Ä x=ÿ	   =  ä x>ÿ	   >  ä xAÿ	   A  ä xEE€€€€,ÞŽ äO x<<€€€€-ÞŽ Æ $x;E   ;
Ž ä $x<<   
Ž ä 	s ?        " x;;€€€€,´Ž ä x,<€€€€-´Ž ä x<7ÿÀŽ äx-7€€€€.ÞŽ Ä� x71˜º  ÿ    ä x55€€€€(ÞŽ â$x--   <
Ž â x17ÿÀŽ äx77€€€€)ÞŽ â $x55   
Ž â x<3ÿÀŽ äx33€€€€/ÞŽ â $x7   1
Ž â x--€€€€.´Ž â $x.€   ÿ Ž Ä $x<3   <
Ž â x€€€€)´Ž ä x)?€   .øŽ ä xB˜º  ÿ    ä x/<€€€€/´Ž â $t<ÿÿ Ž â x.ÿÀŽ â $x))ÿŽ â x€€€€*ÞŽ Äx1@ÿÀŽ äx@@€€€€+ÞŽ â $tBÿ @€ÿ Ž â x3)€€€€<ìŽ â $x   .
Ž ä $x@@   1
Ž â 	s. =        " x<3ÿþþþÿàÿ ä x1)€€€€BìŽ â &r;;   ÿ   â x€€€€*´Ž Ä x*@€€€€+´Ž ä x<<)øŽ ä x@1ÿþþþÿàÿ â &r,,   ;   â x+<˜º  ÿ    ä x@@ @€)øŽ â &r--   ,   â x,+ÿÀŽ Ä x@@˜º  ÿ    â $x..€   ÿ Ž â x++€€€€"ÞŽ â 	s) >        " x(5€€€€(´Ž ä x3@ÿÀŽ â$x1+   ,
Ž â x@@€€€€#ÞŽ â &r((   ÿ   â x.=€   .øŽ â 	s+ A        b &r//8   -   â x"1€€€€"´Ž â $x@@   3
Ž Ä &r(9   (   ä $x.ÿŽ ä $t-ÿÿ Ž ä $t1ÿ @€ÿ Ž â x@@€€€€#´Ž â &r#"   ÿ   â x"€€€€-ìŽ â&r:   (   â x(€€€€1ìŽ â $x1)€   ÿ Ž â x-"ÿþþþÿàÿ Ä x)(ÿþþþÿàÿ ä x(>€   1øŽ ä x--øŽ â $x"+€   ÿ Ž ä/ $t1ÿÿ Ž â x--˜º  ÿ    â $x((ÿŽ â x"A€   "øŽ Ä x) @€øŽ ä x,-ÿÀŽ äx)-€€€€ ÞŽ ä x-(€€€€1ìŽ â $t1ÿ @€ÿ Ž ä $x""ÿŽ â x--ÿþþþÿàÿ â $t+ÿÿ Ž Ä $x))   ,
Ž â x,(€€€€1ìŽ ä x+"€€€€+ìŽ ä x--(øŽ ä x3,ÿþþþÿàÿ ä x,"€€€€1ìŽ ä x++ÿþþþÿàÿ Ä x1-˜º  ÿ    ä x.3 @€(øŽ ä x++"øŽ ä x-,ÿþþþÿàÿ ä x(1ÿÀŽ äx11€€€€$ÞŽ Ä x++˜º  ÿ    ä x-- @€"øŽ â $x31   (
Ž â x(˜º  ÿ    ä x"+ÿÀŽ ä x-˜º  ÿ    ä x++€€€€&ÞŽ Ä x-ÿÀŽ äx€€€€'ÞŽ â $x++   "
Ž â x..˜º  ÿ    Æ $x   -
Ž â x++€€€€&´Ž ä x1(ÿÀŽ äx((€€€€!ÞŽ ä x5.ÿÀŽ äx..€€€€%ÞŽ â &r++   ÿ   â x€€€€'´Ž â $x((   1
Ž â x))€€€€ ´Ž â $x..   5
Ž â x33€€€€$´Ž â &r+   +   Ä &r#@9   #   â x!(€€€€!´Ž â &r))   +   â x..€€€€%´Ž â &r#3:   #   â x0ð   ÿÀŽ ä x 2ð   ÿÀŽ â &r!!8   )   â x00   ÿÀŽ ä x22   ÿÀŽ â &r*C      Ä &r#.C   #   â xÿ     â $r00/   ÿŽ â x ÿ      â $r22!   ÿŽ ä $x//   Ž ä $x!!   #Ž ä $r   0Ž â r///   ÿ� â $r  #   2Ž â r!!!   ÿ� Ä z ÿ `  pPð ä r/   ÿú� ä r !    ÿú� ä z ÿ a   Sð ä xÿ     ä x!ÿ      Ä r   ÿ� ä r !    ÿ� â 0r!ÿ4    A   â xÿ     â 0rDÿD    A   â x ÿ      â 0r-ÿ6    A   â Er,        ä  r*!D     A  â Er+         â  r-D-     A  â G‰  	    € ê r   ÿàó Ê $x   Ž  Ê �y0    ¨ �y1    è �y.    ( �y/    b	 %x(J   Ž Ì �y((     â  $x 0   ÿ Ž ÊO x  ø  ÿÀŽ È z"   ÿàó Ê $v#ÿ 	 ÿŽ  Ì �y""    › ¢ $x 1   ÿ Ž Ê� x  ø  ÿÀŽ È z    ÿàó Ê $v!ÿ 	 ÿŽ  ä $x.   ÿ Ž È�y      › â x$ø  ÿÀŽ â $x/   ÿ Ž Æz$$  ÿàó ä xø  ÿÀŽ Æ $v%ÿ 	 ÿŽ  â z  ÿàõ Ê �y$$    › " $vÿ 	 ÿ Ì �y    › b r&   ÿàó Ê $x'   Ž  Ê �y&&B    " x0ÿ	   0  È 	s2 0        " x1ÿ	   1  â $t4ÿ @€ÿ Ž ä $x32€   ÿ Ž Ê x30€   3øŽ â $t2ÿÿ Ž â 	s' 1        &. $x33ÿŽ Ê x23€€€€2ìŽ äx)3€€€€4ìŽ ä x22ÿþþþÿàÿ ä x0)ÿþþþÿàÿ ä x223øŽ ä x/ÿ	   /  Ä x52˜º  ÿ    ä x30 @€3øŽ â 	s) /        b $x0'€   ÿ Ž â x25ÿÀŽ ä x.ÿ	   .  ä x43˜º  ÿ    ä x'1€   0øŽ Ä 	s0 .        " x14ÿÀŽ â$t6ÿÿ Ž ä $x''ÿŽ ä $t<ÿ @€ÿ Ž ä $t3ÿ @€ÿ Ž â x55€€€€"ÞŽ äO x44€€€€#ÞŽ Æ $x55   2
Ž ä $x44   1
Ž Æ x25€€€€"´Ž ä x"'€€€€6ìŽ â $x6)€   ÿ Ž â/ x44€€€€#´Ž ä x""ÿþþþÿàÿ ä x#'€€€€<ìŽ ä x6/€   6øŽ â &r/2   ÿ   â x""'øŽ â $x)0€   ÿ Ž â x2#ÿþþþÿàÿ â &r#49   /   â x/"˜º  ÿ    Ä x..€   )øŽ ä x"2 @€'øŽ â $x)6ÿŽ â x0/ÿÀŽ â$t2ÿÿ Ž â x'/€€€€ ÞŽ â� $x..ÿŽ ä $t/ÿÿ Ž ä $x''   0
Ž â x0)€€€€2ìŽ â $t4ÿ @€ÿ Ž â x/.€€€€/ìŽ Ä x20ÿþþþÿàÿ ä x1)€€€€4ìŽ ä x//ÿþþþÿàÿ ä x0.€€€€3ìŽ ä x41ÿþþþÿàÿ ä x//.øŽ Ä x10ÿþþþÿàÿ ä x//˜º  ÿ    ä x11 @€.øŽ ä x./ÿÀŽ ä x01˜º  ÿ    ä x22)øŽ Ä x//€€€€$ÞŽ äx44 @€)øŽ ä x)0ÿÀŽ â $x//   .
Ž â x32˜º  ÿ    ä x00€€€€%ÞŽ ä x""˜º  ÿ    Ä x23ÿÀŽ â$x00   )
Ž â x44˜º  ÿ    ä x33€€€€ÞŽ äx//€€€€$´Ž ä x)"ÿÀŽ ä x14ÿÀŽ Ä x""€€€€!ÞŽ â $x33   2
Ž â x44€€€€ÞŽ â &r//   ÿ   â x00€€€€%´Ž â $x""   )
Ž â x''€€€€ ´Ž â $x44   1
Ž â x33€€€€´Ž â &r/0   /   â x!"€€€€!´Ž Ä x44€€€€´Ž â &r''   /   ä &r#3:   #   â x&   ÿÀŽ â &r!!8   '   ä &r#4C   #   â x&ÿ   &  â $r!   ÿŽ ä $x!!   #Ž Ä $r&#   Ž Æ r!!!   ÿ� È r!   ÿú� È xÿ     È r   ÿ� â 0r!ÿ(    A   Æ xÿ     ä  rD!     A  ä Er        Ê #r      â #r*,      â #r-+      â G‰  	    € ê r   ÿàñ Ê $x     Ê �y-    ¨ �y,    è �y*    ( �y+    b	 %x&J   Ž Ì �y&&     â  $x -   ÿ Ž ÊO x  ø  ÿÀŽ È z"   ÿàñ Ê $v#ÿ 	 ÿ  Ì �y""    › ¢ $x ,   ÿ Ž Ê� x  ø  ÿÀŽ È z    ÿàñ Ê $v!ÿ 	 ÿ  ä $x*   ÿ Ž ä$x+   ÿ Ž ä�y      › â x$ø  ÿÀŽ ä xø  ÿÀŽ ä z$$  ÿàñ Ä z  ÿàó Æ $v%ÿ 	 ÿ  ä $vÿ 	 ÿŽ  È �y$$    › ( �y    › b r(   ÿàñ Ê $x)     Ê �y((B    " x-ÿ	   -  È 	s. -        " x,ÿ	   ,  â $t/ÿÿ Ž ä $t1ÿ @€ÿ Ž ä $x..€   ÿ Ž Ê x.-€   .øŽ â 	s) ,        (. $x..ÿŽ Ê x/.€€€€/ìŽ äx'.€€€€1ìŽ ä x//ÿþþþÿàÿ ä x''ÿþþþÿàÿ ä x//.øŽ äx'' @€.øŽ â $x))€   ÿ Ž â x//˜º  ÿ    Ä x*ÿ	   *  ä x0'˜º  ÿ    ä x+ÿ	   +  â 	s' *        " x./ÿÀŽ ä x),€   )øŽ ä x-0ÿÀŽ Æ 	s1 +        b $x))ÿŽ ä $t2ÿ @€ÿ Ž ä $x''€   ÿ Ž Ê x'*€   'øŽ Ê $x''ÿŽ â x//€€€€"ÞŽ äO x00€€€€#ÞŽ Æ $x//   .
Ž ä $t.ÿÿ Ž ä $x00   -
Ž â x,/€€€€"´Ž ä x")€€€€.ìŽ ä x00€€€€#´Ž ä x""ÿþþþÿàÿ Ä x#)€€€€2ìŽ â &r,   ÿ   â x"")øŽ â $x.1€   ÿ Ž â/ x,#ÿþþþÿàÿ â &r909      â x"˜º  ÿ    ä x.+€   .øŽ ä x", @€)øŽ â $t,ÿÿ Ž â x*ÿÀŽ â$x..ÿŽ â x€€€€ ÞŽ â� $t)ÿÿ Ž â x#'€€€€,ìŽ â $t,ÿ @€ÿ Ž Ä $x   *
Ž â x).€€€€)ìŽ â $t+ÿ @€ÿ Ž â x*#ÿþþþÿàÿ ä x#'€€€€,ìŽ ä x))ÿþþþÿàÿ ä x**'øŽ ä x,#ÿþþþÿàÿ Ä x+.€€€€+ìŽ ä x)).øŽ ä x#*˜º  ÿ    ä x,, @€'øŽ ä x++ÿþþþÿàÿ ä x))˜º  ÿ    Ä x*#ÿÀŽ ä x,,˜º  ÿ    ä x++ @€.øŽ ä x##€€€€$ÞŽ äx.)ÿÀŽ äx))€€€€ÞŽ Äx',ÿÀŽ â$x##   *
Ž â x,,€€€€%ÞŽ â $x))   .
Ž â x""˜º  ÿ    ä x.+˜º  ÿ    â $x,,   '
Ž â x##€€€€$´Ž ä x'"ÿÀŽ Ä x+.ÿÀŽ ä x""€€€€!ÞŽ ä x..€€€€ÞŽ â &r##   ÿ   â x,,€€€€%´Ž â $x""   '
Ž â x€€€€ ´Ž â $x..   +
Ž â x))€€€€´Ž â &r#,   #   â x!"€€€€!´Ž Ä x..€€€€´Ž â &r   #   ä &r)):   9   â x(   ÿÀŽ â &r!8      ä &r).C   )   â x(ÿ   (  â $r   ÿŽ ä $x   )Ž Ä $r()   Ž Æ r   ÿ� È r   ÿú� È xÿ     È r   ÿ� â 0rÿ&    A   Æ xÿ     ä  rD     A  ä Er        Ê #r      â x   ÿàÿ ä x(  ÿàó ä |    p`ð ä x

(  ÿàõ â $rÿÿ   Ž  â x		    ÿàÿ Æ $rÿÿ    Ì D	       À â Gy  ÐÙÿÿÿÿƒ ê Ay       € ê $rÿÿ    Ž ä $r
ÿÿ    Ž ä Ay        € ê z ÿ `  pPð È z ÿ a   Sð Ú G	  P    € ê ‰      ( ‰      b !r       â !r       È/ ‰      ( ‰      b !r       â !r       È/ ‰ €    ( ‰	 €    b !r       â !r		       È/ ‰ @    ( ‰	 @    b !r       â !r	       È/ ‰      ( ‰      b !r       â !r       â/ Gy  €    € ê ‰      ( ‰      h ‰      ¨ ‰	      â !r       â !r       â/ !r
       ÈO ‰      " !r	       Æ� ‰	      h ‰
      ¨ ‰      â !r       â !r		       È/ ‰ €    " !r
       ÆO ‰	 €    b !r       Æ� ‰ €    ¨ ‰ €    â !r       â !r	       â/ !r       ÈO ‰ @    " !r
       Æ� ‰ @    h ‰ @    ¨ ‰
 @    â !r       â !r       â/ !r       ÈO ‰      â !r	
       Æ� ‰      ( ‰      h ‰	      ¢ !r       â !r       â/ !r       â !r
	       âO y      !   ( y      %   $ $x   Ž Ê z  p  p`ð È x    pDp  Ú M	        € ê z ÿ ^  pPð â $z  v  Ž Æz ÿ _   Sð Ú $ÿ   ÿ Ž È % ^   Ž Ì �	     ¢ $x    ÿ Ž Ê r  ÿ   p"ò äx     p"ö äx     p"ø äx     p"ô Î $ ÿÿ    Ž â z ÿ `  pPò â $2 ÿÿ    Ž ä $B ÿÿ    Ž â z ÿ a  Sò â $" ÿÿ   
 Ž È !         ÐO G™       € ê z ÿ b  pPð È z ÿ c   Sð Ú $ÿ   ÿ Ž È % b   Ž Ì �	     ¢ $x   ÿ Ž ä $vÿ h  ÿ Ž Æ x øÿÿÿp"ò äx üÿÿÿp"ô är ÿ   p"ö äx    p"ø Î $ÿÿ    Ž â x    pBò â $"ÿÿ    Ž ä $2ÿÿ    Ž ä $Bÿÿ   
 Ž È !       ÈO G  P    € ê x    pRð Ú G‰  �     € ê x    pRð Ú G	  ð    € ê  x;ª¸¿  A  Ì s        $ !t  €?    Ì s        $  r     A  È  r       A  â Gy  P    € ê  x'7=  A  â x*BL?  A  Æ #t  €?   â  x   ?  A  Æ  r     A  Ì s     $   $ !t  €?    È  r     A  È  r       A  â Gy  °    € ê x    pRð Ú G‰  à     € ê x    pRð Ú G	        € ê 	z i    � ä 	z   i    � Æ  x;ª¸¿  A  â 	z   i €  � Ê s        $ !t  €?    Ì s        $  r     A  È  r       A  â Gy  Ð     € ê  r       A  â Gy  °     € ê 	x  à@  � ä 	x    à@  � Æ  x#ÛÙ?  A  â 	x    àÀ  � Æ  x;ª¸¿  A  ä !t   €?    È s        $ !t  €?    Ì s        $  r      A  È  r       A  ä y      "   " $tÿ   ÿ Ž Æ y      &   b $z s  Ž È $z v  Ž È/ %v j   Ž Ê †y      â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y       "   " ¹z  l      â Ey  `+    € â ™x?    â y#      !   b ¹z  F   
   â x$      ÿ  â x&      ÿ  â y-      %   ¢ r@ ÿ       â xB      ÿ  â $rFÿÿ   ÿ Ž â x(      ÿ  â xH      ÿ  â xJ      ÿ  â r, ÿ       â $r6ÿÿ   ÿ Ž â r  ÿ   pRð Ä r  ÿ   pRò â$x/     #Ž â/ x--   ÿ   ÈO xÿ   /  È r/   ÿ� È x"ÿ     È | "   pbô Ú G)  à)    € ê y      '   " xøÿÿÿÿÀŽ â $rÿÿ   ÿ Ž â x?"   ÿ   â y      &   b $rÿÿ   ÿ Ž â r//  €ÿàÿ â $z- p  "Ž â r$ ÿ       â $t<ÿB   ÿ Ž â x>ÿ   /  â 'z y   Ž È 'z s   Ž â/ zÿ z    ä x/   ÿ   â $z2 }  ÿŽ â zÿ t    â $z |  Ž â x-   ÿàÿ â $z	 |  "Ž ä %x   ÿŽ È $z v  Ž â x1   ÿüŽ â $z	 v  	Ž ä %v X  <Ž È $z w  ÿŽ â x4   ÿà÷ â %xB   Ž Æ r22   ÿàõ â $z p  	Ž â z X  ÿàù â $vÿ q  ÿ Ž â r7ÿ   ÿäÿ â %v< X  <Ž â x   ÿà÷ ä r9?   ÿŽ â$r5ÿÿ   = Ž â zÿ Y  „þ â $x.   ?Ž âz=? q  ÿàÿ â $x8   ?Ž Ä $x:   ?Ž ä$x0   ?Ž ä$x;   ?Ž ä $rAÿÿ   ÿ ä r<1   ÿàõ â �yúÿÿ • ¦ r5   ÿä â �y	üÿÿ • è �y
þÿÿ • ( �y    • h �yŠ   • ¨ �y+   • ¨ �yˆþÿÿ • ¢ $tÿ   ÿ Ž Æ �yE    • ¢ x_ $       ä rh/2   =àÙ È r>A   ÿ¤~ â %vhh Z  _ Ž È $x!$   ÿŽ Ê rii!   ÿàÿ ä r/2   8àÙ Æ �yGh   ™ â  r 4       Æ �yPh   ™ â  r>A   ÿ¤~ är /2   9àÙ â�yRh   ™ d $x!$   ÿŽ â rO>A   ÿ¤~ â �yQh   ™ b r*/2   0àÙ â %v Z  _ Ž Ä �yMh   ™ b rL>A   ÿ¤~ â $x[O$   ÿŽ â r!   ÿàÿ â �yOh   ™ b %v   Z  _ Ž Æ �yg   ™ b $x!!   [Ž Æ �yl   ™ h �yk   ™ h �yj   ™ h �ye   ™ h �yf    ™ h �yn    ™ h �ys    ™ h �ym    • h �yr    ™ h �yq    ™ h �yp    ™ h �yo    ™ b $t�ÿÿ Ž Æ �yt     • b $x
   ÿ Ž Êxø  ÿÀŽ È z  ÿàõ È zÿ  ÿä Ì �y    › " xÿ   
  Ê %v   Ž Ì �y    › " x   ÿ   ÈO xø  ÿÀŽ È z  ÿàõ ä xÿ     Æ $vÿ  ÿ â x	   ÿ   â� %v   Ž Æ xø  ÿÀŽ â �y    › ¦ z  ÿàõ â �y    › â xÿ   	  Æ $vÿ  ÿ â r/2   ?àµ ä xÿ     â$x   ÿ Ž â r
>A   ÿd~ â �y    › b %v   Ž ä xø  ÿÀŽ ä %v Z  _ Ž Ä �y    › ¤ $x
$   ÿŽ ä %v   Ž â z
  ÿàõ ä r		   ÿàÿ Æ �y    › ¢ $vÿ  ÿ Æ �yV   ™ ¨  �y

    › ¨ �yY   ™ ¨  �yZ   ™ ¨  �yX   ™ ¨  �yW   ™ ¨  �yU   ™ ¨  �yT   ™ ¢  $rÿÿ   7 Ž â r/2   .àµ Ä �yS    ™ ¤  r>A   ÿd~ ä �y3üÿÿ • ¢  r/2   :àµ â %v Z  _ Ž ä �yD    • ¤ $rÿÿ   < Ž â r 5       â $x$   ÿŽ È �yN    • ¢  $x   Ž ä %v* Z  _ Ž Æ/ �y[   ™ ¢ r>A   ÿd~ â %v Z  _ Ž â r/2   ;àµ â �yc   ™ ¤ $x]$   ÿŽ â r*>A   ÿd~ â %v Z  _ Ž â �yb   ™ ¦ $x_L$   ÿŽ â �y`   ™ ¢ $xa*$   ÿŽ Æ �yLh   ™ ¨  �y*h    ™ ¢  r]   ÿàÿ ä ra   ÿàÿ â �y]h    • ¨  �ya   ™ ¨ �y\   ™ ¢ $x		   _Ž Æ �y_   ™ ¨ �y^    ™ ¨ �yi   ™ ¨ �yh   ™ ¨  �yd    • ¨ �yu   ™ ¢  x{Šÿ   ÿÀŽ Æ �yz   ™ ¨  �y    ™ ¢" 	s‚ {        " x|ÿ   Š  ä �yy   ™ ¤  x‹|ÿ   ÿÀŽ Ä �yx   ™ ¤  	sŒ ‹        d �yw   ™ ¨ �yv   ™ è �y     ™ â $x‚‚€   ÿ Ž Æ �y   ™ ä  x‰‚€   ÿÀŽ ä �y    ™ b xŒŒ   ÿ   ä/ x‰‰ÿ   ŠxŽ â �y!    • b xŽ‹€   ŒxŽ ä xŒÿ   +  â$x‹‰ÿŽ â xŠŠT  +    â �y    ™ b x’  @€    Æ �y�   ™ b x‰Œÿ   ÿÀŽ Æ �y€   ™ b xŒ‹€€€€�ìŽ Æ �y   ™ h �y~   ™ h �y}   ™ h �y|   ™ h �y{    ™ h �y    • h �y   ™ h �y‡   ™ h �y†   ™ h �y…   ™ h �y„   ™ h �yƒ   ™ h �y‚   ™ h �y    ™ h! �y	    • b x�‹€€€€’ìŽ Ä xŒŒÿþþþÿàÿ ä xÿ   Š  È xÿ   ÿÀŽ â 	s“ ‰        â xŠ�ÿþþþÿàÿ ä xŒŒ‹øŽ Ê 	s‘         " xŠŠ @€‹øŽ â $t”ÿÿ Ž â xŒ˜º  ÿ    â $x�ŽÿŽ â xŽ  @€    ä xŒŠ˜º  ÿ    ä xŠ€€€€ÞŽ äx‹�€€€€”ìŽ Ä xÿÀŽ ä x��€€€€ŽìŽ ä x�‹ÿþþþÿàÿ â $x‹Š   
Ž â x’�ÿþþþÿàÿ â $x‘‘€   ÿ Ž â xŠ“   ÿ   ä x�Œ€€€€ÞŽ Ä xŽŒÿÀŽ ä xŒ��øŽ ä x�ˆÿ   ÿÀŽ ä xŠ‰€   ŠxŽ ä x‘€   ‘xŽ ä 	s� �        " $x“ŠÿŽ â x’ @€�øŽ â $x‘‘ÿŽ â x     Ä x’  @€    â $t–ÿ @€ÿ Ž â rŽ�Ž  €ÿàÿ ä x�ÿ   ˆ  ä xŠ“€€€€”ìŽ ä x‘€€€€ìŽ äx‰‘€€€€’ìŽ Ä x‹€€€€´Ž ä x��ÿ   ÿÀŽ ä x”Šÿþþþÿàÿ ä xŽ€€€€´Ž ä x‹“€€€€–ìŽ ä xŠÿþþþÿàÿ Ä x’‰ÿþþþÿàÿ ä xŽÿ   E  â 	s� �        b x–‹ÿþþþÿàÿ ä xŠŠ‘øŽ äx‰’ @€‘øŽ ä xŽŽÿ   ÿÀŽ Ä x‘ˆT  E    ä x•Œ˜º  ÿ    ä x��   ÿ   ä x‹”“øŽ äx– @€“øŽ ä xŒÿ   ‘  â 	s“ Ž        " x’•ÿÀŽ Ä x•€€€€•ÞŽ ä x‘�€   ÿÀŽ ä x”˜º  ÿ    ä xŒŒÿ   ÿÀŽ ä r•’  €ÿàÿ ä x‘‘ÿ   ˆxŽ â 	s’ Œ        " x€€€€´Ž â $tˆÿÿ Ž â x�   ÿ   â/ $x‘‘ÿŽ â x�  @€    Ä x—”ÿÀŽ äx””€€€€ÞŽ ä x��€   xŽ ä x‘€€€€ˆìŽ ä x““   ÿ   ä xˆ‘€€€€�ìŽ â $x””   —
Ž â xÿþþþÿàÿ â $x��ÿŽ â xŽŽ€   “xŽ â $t—ÿ @€ÿ Ž â x•     Ä xˆˆÿþþþÿàÿ ä x�‘øŽ ä x”€€€€´Ž â $x”ŽÿŽ â x“’   ÿ   äx�ˆ @€‘øŽ ä x�€€€€•ìŽ Ä xˆ�€€€€—ìŽ ä x’Œ€   “xŽ ä x‘ÿþþþÿàÿ ä x“ˆÿþþþÿàÿ ä x”€€€€•ìŽ ä xŽ‘�øŽ â$x‘’ÿŽ â xŒ“ @€�øŽ â $t’ÿ @€ÿ Ž â xˆ”€€€€—ìŽ Ä x“ÿþþþÿàÿ ä x�     ä x•ˆÿþþþÿàÿ ä xˆ“”øŽ ä x�‘€€€€�ìŽ äx“‘€€€€’ìŽ Ä x• @€”øŽ ä x��ÿþþþÿàÿ ä x”“ÿþþþÿàÿ ä x��˜º  ÿ    ä x“�‘øŽ äx”” @€‘øŽ Ä x’�˜º  ÿ    ä x��ÿÀŽ ä x‘Ž˜º  ÿ    ä x�€€€€�ÞŽ äO xŽŒ˜º  ÿ    ä x�’ÿÀŽ Ä xŒ‘ÿÀŽ ä x’’€€€€ÞŽ ä r���  €ÿàÿ ä x‘€€€€‘ÞŽ ä� r’’�  €ÿàÿ ä x�€€€€´Ž â $x�‘   Œ
Ž â x�ŽÿÀŽ Ä xŽŽ€€€€ÞŽ ä x““˜º  ÿ    ä x””˜º  ÿ    ä rŽŽ�  €ÿàÿ ä x��€€€€´Ž ä x“ÿÀŽ Ä x�”ÿÀŽ ä x“€€€€“ÞŽ äx””€€€€ÞŽ ä x‘ˆ˜º  ÿ    ä xˆ˜º  ÿ    ä r““  €ÿàÿ â $x”   �
Ž â x‘ÿÀŽ Ä x�ˆÿÀŽ ä x‘€€€€‘ÞŽ ä xˆˆ€€€€ÞŽ ä x‹‹˜º  ÿ    ä x€€€€´Ž ä r‘  €ÿàÿ Ä rˆ�  €ÿàÿ ä x�‹ÿÀŽ ä xˆ€€€€‹ÞŽ ä x‹Š˜º  ÿ    ä x€€€€´Ž â &rV   ÿ   âx’€€€€´Ž â &rVV   ÿ   â xŠ‰˜º  ÿ    â $x�ˆ   �
Ž â xˆ‹ÿÀŽ â &rY      âx‹
€€€€‹ÞŽ â &rVY   V   â x‰ŠÿÀŽ â &r�Z      âxŽ€€€€´Ž â &rVZ   V   â xŠŠ€€€€ÞŽ Ä r‹‹ˆ  €ÿàÿ â &rVX   V   âx“€€€€´Ž ä rŠŠ‰  €ÿàÿ ä xZ˜º  ÿ    â &rX      â x
‹€€€€
´Ž ä xŠ€€€€´Ž â &rW      Ä &rV
W   V   â x�€€€€´Ž â &rU      ä&rVU   V   â x€€€€´Ž â &rT      ä&rVT   V   ä &rTG   ÿ   â xYZÿÀŽ Ä xZZ€€€€ÞŽ â &rTP   T   â xÿ   E  â &rGG   ÿ   â xÿ   +  â &rT�R   T   â x   ÿüŽ â $xZZ   Y
Ž ä &r+S      ä &rTQ   T   â xZ€€€€´Ž â &rGP   G   Ä $r++   ÿŽ ä &rTM   T   â x   ÿüŽ â &rRR   G   â xÿ   +  â &rSS   V   ä &rTO   T   â r+   ÿ� â &rRQ   R   Ä $rSS   ÿŽ ä &rTL   T   â xÿ     â &rR
M   R   ä 0r3ÿ3    A   ä 0rDÿD    A   â xPÿ   S  â &rG*   T   Ä 0rNÿN    A   â Er        â &rRO   R   â  r+3D     A  â $rGG   ÿŽ â rPPS   ÿ� â  rEDN     A  â &rRL   R   â #r6+   6   â &rD[   ÿ   âxPÿ   P  â &rL[   ÿ   â xÿ   G  â &r+*   R   Ä &rDc   D   âErP P       â &rLc   L   â rG   ÿ� â $r*+   ÿŽ ä &rD�b   D   âxÿ     â &rLb   L   â #r,EP   ,   â 0r]ÿ]    A   â xEÿ   *  â &rGa   D   âEr        â &raa   L   â  r+3]     A  â rEE*   ÿ� â &rG`   G   Ä &ra
`   a   ä &r*g   ÿ   â #rJ+   J   â &r\   G   ä&r\\   a   ä &r*l   *   ä &r\_   \   Ä &r*�k   *   ä &r_      â xEÿ   E  â &rgg   ÿ   ä &rG^   \   ä &r*j   *   ä &r+^      â ErE E       â &rll   g   Ä &r*i   *   ä $rGG   ÿŽ â  rN]     A  â $r++   ÿŽ ä &rlk   l   ä &rDe   *   â x*ÿ   G  â #rKE   K   â xÿ   +  â &rlj   l   â r**G   ÿ� â &rDh   D   â r+   ÿ� â &rl
i   l   â x*ÿ   *  â 0rdÿd    A   Ä &r+f   D   â Er* *       â &rle   l   ä $rD+   ÿŽ â  r+3d     A  â rdNd     A  â &rlh   l   â xÿ     ä #r)d*   )   â &r*n   ÿ   â xEÿ   D  â &rGf   l   Ä &rnn   ÿ   â Er        â &r*   *   â rEED   ÿ� â $rGG   ÿŽ ä &rn   n   ä &ru   ÿ   â #rB+   B   â &r*�s   *   â xÿ   G  â 0rmÿm    A   â xEÿ   E  â &rz      Ä &rns   n   ä &r*r   *   ârGG   ÿ� â  r3m     A  â ErE E       â &r�y      ä &rnr   n   ä &r+q   *   â #rIE   I   â &rx      Ä &rq
q   n   ä &rp   +   ä&r+w      ä &rpp   q   ä &ro      ä &rv   +   Ä &rpo   p   ä &r+      ä&r   p   ä $r++   ÿŽ ä &r      ä $r   ÿŽ â xÿ   +  â &ruu   ÿ   Ä &rE      â xÿ     ä r+   ÿ� â &rzz   u   â r   ÿ� â 0rtÿt    A   â xÿ     â &rzy   z   â xÿ     ä Er        â  r3t     A  â &rzx   z   Æ #r'   '   â &rw
w   z   ä &r    ÿ   ä&r     ÿ   ä &r   ÿ   ä&r   ÿ   ä &rvv   w   Ä &r�      ä&r �       ä &r‡      ä&r‡      ä &rv   v   ä &r�€      Ä &r €       ä &r�†      ä&r†      ä &r+   v   ä &r      ä&r       Ä &r…      ä&r…      ä $r++   ÿŽ ä &r~      ä&r
~      ä &r„      Ä &r
„      â xÿ   +  â &r}      ä&r}      ä &rƒ      ä&rƒ      â r+   ÿ� â &r|      Ä &r||      ä &r‚      ä&r‚      â xÿ     â 0r!ÿ!    A   ä &r{      ä&r{{   |   Ä &r      ä&r      â Er        â $rEE   ÿŽ â  rN!     A  â $r   ÿŽ ä $r{{   ÿŽ ä $r   ÿŽ Ä $r   ÿŽ â x  ÿàõ â #rC   C   â x""   ÿàÿ ä x*ÿ   E  ä xÿ     ä xÿ   {  ä x
ÿ     Ä xÿ     ä rÿ   ÿä ä | "   pbô ä r**E   ÿ� ä r   ÿ� ä r{   ÿ� Ä r

   ÿ� ä r   ÿ� â 0rÿ    A   â Er        â 0r	ÿ	    A   â xGÿ   G  â  rNt     A  â x*ÿ   *  ä xÿ     Ä xÿ     ä x
ÿ   
  ä xÿ     â #r(   (   â ErG G       â  r3     A  âEr* *       â  r3	     A  âEr        â  rmNm     A  âEr        â  r3!     A  â Er
 
       â  rN     A  âEr        â  r	N	     A  â x44  ÿàù Ä x<<  ÿà÷ â #rHmG   H   â #rF*   F   â #r@   @   â #r&   &   â #r%
   %   â #r$	   $   â x..@   ÿàÿ â $r7ÿÿ   7 â x88@   ÿàÿ ä x99@   ÿàÿ Ä x::@   ÿàÿ ä x00@   ÿàÿ ä x;;@   ÿàÿ ä x==@   ÿàÿ ä x??@   ÿàÿ ä r5ÿ5   ÿäÿ â D)       À â Gy  ðØÿÿÿÿƒ ê Ay        € ê Ey  0    € â G‰      € ê r #   ÿHŽ Ê ˆs 6 øÿ H   è ˆs ,€øÿ H   è ˆs J ùÿ H   è ˆs K€ùÿ H   è ˆs B úÿ H   è ˆs )€úÿ H   è ˆs I ûÿ H   è ˆs H€ûÿ H   è ˆs ' üÿ H   è ˆs (€üÿ H   è ˆs F ýÿ H   è ˆs C€ýÿ H   è ˆs @ þÿ H   è ˆs &€þÿ H   è ˆs % ÿÿ H   è ˆs $€ÿÿ H   ä Ay        € ê {          ì M        € ê „y#     H   " z+  r      Ä z - r  p`ð â „y# €   H   d $x!+   ÿŽ â r #ÿ   p"p â „y#    H   ¢ $x/+   ÿŽ Æ „y# €  H   è y      '   ( „y# €  H   h „y	#    H   ( „y#    H   ( „y# €  H   è „y# €  H   â !r6       Æ „y# €  H   â !r,,       Æ/ ‰       " !rJJ       ÆO „y#    H   â !r((       Æ� ‰J      h „y#    H   ¢ !rKK       Æ„y# €  H   â !rB	B       Æ„y#    H   â !rII       â !r        ä „y#    H   â !r))       Æ y       &   " !rHH       â !rJJ       ä/ ‰,      b !r&&       Æ ‰      ( ‰J      b !r''       Æ ‰K      h ‰B      b !rFF       âO $z   x  -Ž ä ‰)      " x--   ÿàÿ â !rCC       â� !r
,       â/ $z  ~   Ž â ‰I      â !r@@       â !r       Ä‰
      b r+    ÿ† â!rJJ       â !   ÿàù â ‰	 €    ¢ rÿÿ   ÿäÿ â !rKK       â z j  ÿ† â‰(      â ÿÿ   ÿä â !r%%       â z k  � â „y# €  H   â 
 j  ÿˆ â!rBB       â 
  r  ÿàó â‰H      " 
 k   â !r))       â  /   ÿàù â $ÿÿ   ÿŽ  â 
 j  ÿ„ â ‰F      " ÿÿ   ÿä â !r
       â/ 
 j  ÿˆ â ‰&      b 
 k   â !r	       âO 
 k   â ‰ €    ¢ !rII       â !r((       â� z - r  p`ò â ‰'      â $x	+   ÿŽ âr+    ÿ† Ä ‰C      b x #   p"ò â !rHH       âr		    ÿàõ â ‰K      " $r"ÿÿ   ÿŽ â z
 j  ÿ† â $x+   ÿŽ â ‰@      " rÿÿ   ÿä â !rFF       â z	 j  ÿ„ â ‰ @    " !r&&       â/ z k  "� â !r$$       â !r       âO ‰ %      b z		 k   â !r''       Ä� ‰B      ¢ !   ÿàÿ â !rCC       ä‰)      â *    ÿàõ â!rKK       ä‰ @    " ,    ÿà÷ â !r@@       Ä ‰I      b 1ÿÿ   ÿäÿ â !r       ä ‰H      " +   ÿàÿ ä , j  ÿ† â !r"%        â/ ‰'      b .    ÿàù â $ÿÿ   ÿ â !rBB       âO ‰!(      ¢ 
* j  ÿ„ â !r))       Ä� ‰J €    â ÿÿ   ÿä â !r       ä‰#F      " . j  ÿˆ â !rII       ä‰+&      b 
* k   â !rHH       Ä ‰ C      " . k   ä , k  1� â ‰K €    " !r''       Æ/ ‰%@      b !r((!       ÆO ‰-"      ¢ !rJJ       Æ� ‰$      â !rFF#       Æ‰B €    " !r*&+       â$t+ÿ   ÿ Ž ä ‰      b !rCC        Æ ‰) €    " !rKK       Æ ‰I €    " !r@@%       â/ !r,"-       äO ‰!H €    b !r.$       Æ� ‰ ' €    ¢ %v  j  + Ž â !rBB       ä‰#( €    â !r3       Æ‰J @    " /   ÿàÿ â !r))       ä ‰%F €    "     ÿàõ â !rII       ä ‰"C €    b ÿÿ   ÿä ä  j  ÿ„ â ‰K @    b !r!H!       Ä/  k   â ‰+@ €    b !r''        ÆO ‰-* €    ¢ !r#(#       Æ� ‰/, €    â !rJ       Æ‰1.      " !r%F%       Æ ‰      " !rCC"       Æ‰B @    b !rK       Æ ‰) @    b !r+@+       Æ/ ‰I @    b !r-*-       ÆO ‰ ! @    ¢ !r/,/       â� !r1.1       ä‰"' @    â !r5       Æ ‰$# @    " !rB       Æ‰      â !r))       Æ ‰&% @    " !rII       Æ/ ‰(C @    b !r !        ÆO ‰      ¨ ‰*+ @    b !r"'"       Æ� ‰,- @    â !r$#$       Æ ‰01 €    ( ‰./ @    " !r&%&       Æ‰      â !r(C(       Æ/ †	 3    â !r'       ÆO ‰)      ¢ !r*+*       â!r,-,       ä� † 5   â !r3       Æ/ ‰       b !r010       â !r./.       ä ‰I      ( ‰"      è ‰$      " !r))       ÆO †	 3    è ‰!&      h ‰(      b !r        â/ !r3       ÄO † '    â !rII       â !r"       ä� ‰*      " !r$       Æ‰#,      h ‰%0 @    ¢ !r!&!       Æ‰.      â !r(       Æ †	 3    è	 † )   è	 †	 I    â	 !r*       â !r#,#       Ä/ †     â	 !r%0%       ÆO †	 
    â	 !r.       Æ� † 
   è	 †	 !    è	 †     è	 †	     è	 † #   è	 ‰%      ( †	     â	 M™        € ê x   ÿàÿ â !r%%       Ær    ÿàñ È r ÿÿ   ÿä  ä z j  ÿ€ È z k     Ê †y %    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y       "   " ¹z  l      â Ey  p(    € â ™x?    â y      !   b ¹z  F   
   â x       ÿ  â x"      ÿ  â r6 ÿ       â x8      ÿ  â x:      ÿ  â x$      ÿ  â x>      ÿ  â $r&ÿÿ   ÿ Ž â r  ÿ   pRò Ä r  ÿ   pRð â$x(     Ž Ê/ xÿ   (  È r(   ÿ� È xÿ     È |    pbô Ú G)  0'    € ê y      '   " r ÿ       â $rÿÿ   ÿ Ž â xøÿÿÿÿÀŽ â y	      %   b $t2ÿB   ÿ Ž â x5   ÿ   ä y      &   ¢ r((  €ÿàÿ ä r  ÿ       Ä x4ÿ   (  â 'z y   Ž È $x		   ÿ Ž â/ zÿ z    â $z+ }  ÿŽ â x(   ÿ   â 'z s   Ž âO x	   ÿàÿ Æ $z	 p  Ž â zÿ t    â $z |  Ž ä$z |  Ž ä %x   ÿŽ È $z v  Ž â x*   ÿüŽ â $z v  Ž ä %v X  2Ž È $z
 w  ÿŽ â x,   ÿà÷ â %xB   Ž Æ r+
+   ÿàõ â $z p  Ž â z X  ÿàù â $vÿ q  ÿ Ž â r/ÿ   ÿäÿ â %v2 X  2Ž â x   ÿà÷ ä r05   ÿŽ â$r-ÿÿ   3 Ž â zÿ Y  „þ â $x'   5Ž âz35 q  ÿàÿ â $x.   5Ž Ä $x)   5Ž ä$x1   5Ž ä $r7ÿÿ   ÿ ä r2*   ÿàõ â �yúÿÿ • ¦ r-   ÿä â �y	üÿÿ • è �y
þÿÿ • ( �y    • h �yY   • ¨ �yN   • ¨ �ySþÿÿ • ¢ $tÿ   ÿ Ž Æ �yR    • ¢ xI $       ä r(+   .àÙ È rJ47   ÿ¤~ ä rA(+   )àÙ â %v Z  I Ž È $x]J$   ÿŽ Ê r]   ÿàÿ Ê �yd   ™ h �yb   ™ h �yh   ™ h �yg   ™ h �yf   ™ h �ye   ™ h �yi    • b $x
   ÿ Ž Êxø  ÿÀŽ È z  ÿàõ È zÿ  ÿä Ì �y    › " xÿ   
  Ê %v   Ž Ì �y    › " x   ÿ   ÈO xø  ÿÀŽ È z  ÿàõ ä x	   ÿ   Æ� $vÿ  ÿ â xÿ     ä xø  ÿÀŽ Æ �y    › ¢ %v   Ž â z  ÿàõ Ê $vÿ  ÿ â �y    › â r(+   5àµ â $x   ÿ Ž âxÿ   	  ä r47   ÿd~ â �y    › b %v Z  I Ž â xø  ÿÀŽ ä xÿ     â $x$   ÿŽ Ä %v   Ž â z  ÿàõ ä r		   ÿàÿ â %v   Ž ä �y    › ¤ $vÿ  ÿ ä �y    › ¨ �yZ   ™ ¨ �yX   ™ ¨ �y    › ¨ �y[   ™ ¨ �y\   ™ ¨ �yT   ™ ¢ rV(+   3àµ Æ �yU   ™ ¢ r
47   ÿd~ Æ �yO   ™ ¢ %vVV Z  I Ž â r<(+   'àµ ä �yQ    ™ ¢ $x
$   ÿŽ â r
47   ÿd~ ä �yB    • ¤ rWW   ÿàÿ â $x
$   ÿŽ â r /       â $r
ÿÿ   , Ž Ä �yPV   ™ ¢  r(+   0àµ Æ �yCV   ™ ¢  %v<< Z  I Ž Æ �yD
üÿÿ • ¢  rK47   ÿd~ â $x==   Ž â r@(+   1àµ â �yFV   ™ ¨ �yGV   ™ ¢ r
 2       â $rÿÿ   - Ž Ä �yHV   ™ ¢  %v Z  I Ž Æ �yE
    • ¢  %v@ Z  I Ž Æ/ �yJV   ™ ¢ $x_K$   ÿŽ Æ �yK<   ™ ¢ %v
A Z  I Ž â rA47   ÿ¤~ ä�yI<   ™ ¢ r@47   ÿd~ ä $xAA$   ÿŽ â �yM<   ™ ¤ $xa@$   ÿŽ Ä �yLV   ™ ¢  rA   ÿàÿ Æ �yA<   ™ ¨ �y@V    ™ ¢  $x   _Ž Æ �y]<   ™ ¢ $x		   aŽ Æ �y_<   ™ ¨ �ya<   ™ ¨ �y`<    ™ ¨ �y^V    • ¨  �yc<    • ¨ �yn   ™ ¨  �ym   ™ ¨  �y=   ™ ¨" �yW   ™ ¨ �yl   ™ è �yV    ™ â  xqÿ   Y  Ä xpYÿ   ÿÀŽ â �yk   ™ â xqÿ   ÿÀŽ ä 	sw p        b �yj   ™ è �y
   ™ f 	s~         " �y<   ™ h �y    ™ h �yo    • h �yv
   ™ h �y|   ™ h �y   ™ b# x}w   ÿ   Æ �yu
   ™ h �y{   ™ b $x€~€   ÿ Ž â x~}€   ÿÀŽ ä �yt
   ™ h �yz   ™ b x~~ÿ   YxŽ Æ �ys
   ™ h �yy   ™ b x€€   €xŽ Æ �yr
   ™ b $x~~ÿŽ Æ �yx   ™ b x     ä xYYT  N    â�yq
   ™ b x�ÿ   N  Æ �yw   ™ b xYÿ   Y  â $t…ÿ @€ÿ Ž ä �yp
    ™ h �y
    • b $x}€ÿŽ â xƒ�ÿ   ÿÀŽ Ä x†     â �y
    ™ h! �y    • b x�Yÿ   ÿÀŽ â $t€ÿ @€ÿ Ž â xY~€€€€…ìŽ äx~€€€€ìŽ È x	ÿþþþÿàÿ â 	s„ ƒ        " x}€€€€†ìŽ ä x		~øŽ ä xYYÿþþþÿàÿ ä x}€€€€€ìŽ â 	s‚ �        b xÿþþþÿàÿ Ä x		˜º  ÿ    ä xYY @€~øŽ ä x~ÿþþþÿàÿ ä x}øŽ ä x€€€€	ÞŽ äx		ÿÀŽ Ä x€Y˜º  ÿ    Æ $x	   	
Ž â x„„   ÿ   ä xSÿ   ÿÀŽ ä x‚‚   ÿ   ä/ xY€€€€€ÞŽ äx		€€€€´Ž ä x€€ÿÀŽ â 	s         " x„ƒ€   „xŽ Ä x‚�€   ‚xŽ â $x€Y   €
Ž â x~~ @€}øŽ â $x„„ÿŽ ä $t}ÿ @€ÿ Ž â x�     â $x‚‚ÿŽ â x€€€€€´Ž â $tƒÿ @€ÿ Ž â x}„€€€€}ìŽ Ä x€‚€€€€�ìŽ äx}}ÿþþþÿàÿ ä x�‚€€€€ƒìŽ ä xƒ€ÿþþþÿàÿ ä x€   ÿ   ä xY     Ä x} @€„øŽ ä x}ÿ   S  ä x€€€   ÿÀŽ ä xY„€€€€YìŽ ä x…�ÿþþþÿàÿ ä x�}ÿ   ÿÀŽ Ä x}ƒ‚øŽ ä x€€ÿ   SxŽ ä xYYÿþþþÿàÿ ä xƒÿ   R  äxSST  R    ä x… @€‚øŽ Ä x˜º  ÿ    â 	s‚ �        " xYY„øŽ ä x…ƒÿ   ÿÀŽ ä xSÿ   S  ä x„ÿÀŽ ä x€€€€ÞŽ â 	s† …        b xƒSÿ   ÿÀŽ Æ $x   „
Ž È 	s„ ƒ        " x~~˜º  ÿ    â $x‚‚€   ÿ Ž â x‡     â $x€€ÿŽ â xS~ÿÀŽ â$t‰ÿ @€ÿ Ž â x~~€€€€ÞŽ ä x‚�€   ‚xŽ ä x††   ÿ   â/ $t�ÿÿ Ž â rˆ~S  €ÿàÿ â $x‚‚ÿŽ â xS€€€€€‡ìŽ â $x„„€   ÿ Ž âx~€€€€€‰ìŽ Ä x†…€   †xŽ ä x€€€€´Ž ä xSSÿþþþÿàÿ â $x††ÿŽ â x~ÿþþþÿàÿ ä x‡  @€    ä x~‚€€€€�ìŽ Ä x„ƒ€   „xŽ ä x…     ä xSS€øŽ äx @€€øŽ â $x„„ÿŽ â x€‚€€€€‡ìŽ â $t‰ÿÿ Ž â x�~ÿþþþÿàÿ Ä x~†€€€€…ìŽ ä xƒ€ÿþþþÿàÿ ä x€†€€€€‡ìŽ ä x…~ÿþþþÿàÿ ä x‹  @€    ä x~„€€€€‰ìŽ Ä xSS˜º  ÿ    ä xˆ€€€€´Ž ä x‡€ÿþþþÿàÿ ä xˆ…†øŽ ä x��‚øŽ ä x€„€€€€‹ìŽ Ä x…~ÿþþþÿàÿ ä x~SÿÀŽ ä xS€€€€SÞŽ äO x†‡ @€†øŽ ä x‡€ÿþþþÿàÿ ä x��˜º  ÿ    Ä x€˜º  ÿ    â $xSS   ~
Ž â xƒƒ @€‚øŽ ä x~�ÿÀŽ ä x……„øŽ ä x€ÿÀŽ ä x�€€€€�ÞŽ Ä� x€€€€€€ÞŽ ä x‚ƒ˜º  ÿ    â $x��   ~
Ž â x‡‡ @€„øŽ ä x……˜º  ÿ    ä r€€  €ÿàÿ ä x‚ÿÀŽ Ä x‚‚€€€€ÞŽ ä xSS€€€€´Ž ä x~‡˜º  ÿ    ä x…ÿÀŽ ä x…€€€€…ÞŽ äx�€€€€´Ž Ä r‚‚  €ÿàÿ ä x�ˆ˜º  ÿ    ä x~ÿÀŽ â$x……   
Ž â x~~€€€€ÞŽ ä x††˜º  ÿ    ä x�ÿÀŽ Ä x�€€€€�ÞŽ ä xYY˜º  ÿ    ä r~~  €ÿàÿ ä x†ÿÀŽ ä x€€€€€´Ž ä x††€€€€ÞŽ Ä r��  €ÿàÿ â &rSZ   ÿ   â x€YÿÀŽ ä xY€€€€YÞŽ ä x˜º  ÿ    â $x††   
Ž â rY€  €ÿàÿ â &rX      â x}}˜º  ÿ    â &rZ	Z   ÿ   â xYÿÀŽ Ä x€€€€ÞŽ â &r[      â x‚€€€€´Ž ä x~€€€€´Ž â &rZX   Z   â x~}ÿÀŽ â $x   Y
Ž â x}€€€€}ÞŽ â &rY\      â x…€€€€´Ž Ä x˜º  ÿ    â &rZ[   Z   â r}}~  €ÿàÿ ä x[ÿÀŽ äx€€€€ÞŽ â &rT   Y   â x}€€€€´Ž â &rY\   Z   â x�€€€€´Ž â &rU      â x€€€€´Ž â &rTT   Y   â x†€€€€´Ž â &rO      â xRÿ   R  â &rTU   T   Ä &rUQ      â xR   ÿüŽ ä r[  €ÿàÿ Æ $rUU   ÿŽ â x€€€€´Ž ä x€€€€´Ž ä xÿ   U  ä xÿ   N  â &rNSP   ÿ   ârU   ÿ� â &rP	P   ÿ   Ä &rTO   T   ä &rOC   N   âxÿ     â &rCC   P   ä 0rDÿD    A   ä 0rBÿB    A   ä &rPF   C   â ErC        â &rNF   O   â  rDB     A  Æ &rOG   N   â #r%C   %   â &rGG   P   ä &rSI   ÿ   â x   ÿüŽ â &rOH   O   ä&rGH   G   ä &rK      Ä &rQQ   T   ä &rH	I   ÿ   ä &rOJ   O   ä&rGJ   G   ä &rM      ä $rQQ   ÿŽ Ä &rHK   H   ä &rOL   O   ä&rGL   G   ä &rA      â xFÿ   Q  â &rHM   H   ä &rO@   O   Ä &rG@   G   ä &r@]      â rFFQ   ÿ� â &rHA   H   ä $rGG   ÿŽ ä &rA_   @   â xFÿ   F  â 0rEÿE    A   â x@ÿ   G  â &rAa   A   â ErF F       â $rOO   ÿŽ â  rBBE     A  â &rA`   A   â r@@G   ÿ� â &rH]   H   â #r?BF   ?   â xÿ   O  â $rBA   ÿŽ â x@ÿ   @  â &rH_   H   Ä 0r^ÿ^    A   â rO   ÿ� â &rHa   H   â Er@ @       ä xGÿ   B  â  rCE^     A  â xÿ     â &rI`   H   â rGGB   ÿ� â #r&C@   &   â &r@Sd   ÿ   â Er        â  rAD^     A  â $rII   ÿŽ â xGÿ   G  â 0rcÿc    A   Ä &rCb   @   â #r8A   8   â ErG G       â  rADc     A  â xÿ   I  â &rd	d   ÿ   ä &r@h   C   â rI   ÿ� â #r>AG   >   â &rIb   d   Ä &rAg   @   ä &rBS=   ÿ   ä&r=	=   ÿ   ä &rhh   I   ä &r@f   A   ä &rBn   B   Ä &r=n   =   ä &rgg   h   ä &r@e   @   ä &rBm   B   ä&r=m   =   ä &rff   g   Ä &r@W   @   ä &rBl   B   ä&rll   =   ä &rfe   f   ä &r=V   @   ä &rfW   f   Ä $r==   ÿŽ â xÿ     â &rAV   f   Æ x@ÿ   =  â &rBk   B   âEr        â &rlk   l   â  rcEc     A  â $rAA   ÿŽ â r@@=   ÿ� â &r=S   ÿ   â#r:c   :   â &r	   ÿ   â xÿ   A  â &rCj   B   Ä &rGj   l   ä &rSS   ÿ   ä&r		   ÿ   â rA   ÿ� â &rC<   C   ä&rG<   G   ä &r=v   =   Ä &rv      ä &rS|   S   ä&r	|   	   ä &rC   C   ä0riÿi    A   ä &rG   G   â xÿ     â &r=u   =   Ä &ru      ä &rS{   S   ä&r	{   	   â  rDi     A  â &r=t   =   â Er        â &rt      ä &rSz   S   Ä &r	z   	   â  riEi     A  â $rGG   ÿŽ ä &r=s   =   ä&rs      ä &rSy   S   ä&r	y   	   â #r$i   $   â xÿ   G  â &r=r   =   Ä &rr      ä &rSx   S   ä&r	x   	   â rG   ÿ� â &r=q   =   ä&rq      ä &rSw   S   Ä &r	w   	   â xÿ     â 0roÿo    A   ä &r=p   =   ä&rp      ä &rS
   S   ä&r	
   	   â Er        â $rCC   ÿŽ â rEo     A  â $r==   ÿŽ Ä $r   ÿŽ ä $rSS   ÿŽ ä $r		   ÿŽ â x  ÿàõ â #r9   9   â x   ÿàÿ ä x<ÿ   C  ä x
ÿ   =  Ä xÿ     ä xÿ   S  ä xÿ   	  ä rÿ   ÿä ä x@ÿ   @  ä |    pbô Ä r<<C   ÿ� ä r

=   ÿ� ä r   ÿ� ä rS   ÿ� ä r	   ÿ� â 0rÿ    A   â Er@ @       â 0rÿ    A   â x<ÿ   <  Ä x
ÿ   
  ä xÿ     ä xÿ     ä xÿ     â #r#@   #   â  rDo     A  â rD     A  âEr< <       â  rDD     A  â Er
 
       â  rE     A  âEr        â  rE     A  â Er        Ä Er        ä x22  ÿà÷ ä x,,  ÿàù â #r;<   ;   â #r6
   6   â #r"   "   â #r!D   !   â #r        â x''@   ÿàÿ â $r-ÿÿ   -Ž â x..@   ÿàÿ Ä x00@   ÿàÿ ä x))@   ÿàÿ ä x11@   ÿàÿ ä x33@   ÿàÿ ä x55@   ÿàÿ ä r/ÿ/   ÿä â D)       À â Gy  °Ûÿÿÿÿƒ ê Ay        € ê Ey      € â G™  ð     € ê $x À  Ž Ê ˆs % ùÿ H   è ˆs ?€ùÿ H   è ˆs 8 úÿ H   è ˆs &€úÿ H   è ˆs > ûÿ H   è ˆs :€ûÿ H   è ˆs # üÿ H   è ˆs $€üÿ H   è ˆs ; ýÿ H   è ˆs 9€ýÿ H   è ˆs 6 þÿ H   è ˆs "€þÿ H   è ˆs ! ÿÿ H   è ˆs  €ÿÿ H   ä Ay        € ê {          ì M	        € ê „y      H   b Ey  @    € æ „y €   H   ¨ „y    H   â !r  %       â/ !r??       È_ ‰       " !r88       Æ� ‰?      h ‰8      ¢ !r        Æ y       %   " !r?       Æ/ ‰      b !r8       ÆO ‰      ¨ ‰      â x    ÿ   â !r       Æ/ z  r  p`ð â !r       ÆO r ÿ   pVp  â ‰ €    " !r	       Æ� ‰ €    h ‰
	 €    ¨ ‰      '   â !r       â x   ÿàÿ â !r       Æ/ ‰  @    " z  r  p`ò â !r
	
       äO ‰ @    b x    p"ò Æ ‰	      &   ¨ ‰
 @    ,       &   b         Æ       '   b !r         â !r       È/ ‰       " $Š	 x  Ž ÆO ‰      b $„	ÿ   ÿ Ž ä $Š ~  Ž â� „y €  H   â !r
       â$ x  Ž È‰      ¢ $ ~  Ž â !r        â %† j  	 Ž â !r	       Æ/ % j   Ž â †‰     è † 	   â !r       âO G	  �     € ê y      &   ( y       '   b $z x  Ž È $z   ~  Ž Ê/ z   r  ÿàõ È rÿÿ   ÿä ä z  j  ÿ„ È z  k   Ê †y     è Ay        € ê !r&       â „y	    H   â Ey  à    € æ ‰       ( „y €  H   è „y    H   è „y €  H   è „y    H   è „y €  H   â !r         Æ „y €  H   è ‰       ( „y    H   â !r        Ê ‰ €    $ !r       Ê ‰ @    $ !r       Ê ‰       $ !r        â G™  °     € ê y      &   " ¹z  r      ä �x   ?àÿ â y       '   b $z x  Ž È $z   ~  Ž Ê/ |     ÿàõ Ê $rÿÿ   ÿ â z  j  ÿ„ È z  k   Ê †y     è Ay        € ê „y    H   â !r;;       â !r		>       â !r$       â „y €  H   â !r6       â !r99       â !r:       â ‰
;      b !r#       â !r""       Ä ‰ 	      ¨ ‰      è ‰      ( ‰      ( ‰      b !r;
       Æ/ ‰9      b !r 	        ÆO ‰"      ¢ !r!       â !r	       ä� y      '   â z!  r      â !r       â!r         â ‰      " !r       Æ ‰      " !r       Æ‰      b !r9       Æ/ ‰       b !r"       ÆO y      &   ¨ ‰      è ‰       â !r       Æ‰      " !r       Æ ‰
	      " !r       Æ‰      b !r         Æ/ ‰      b $z x  Ž ÆO ‰ €    ¢ !r       â� $z# ~  Ž ä ‰      â !r         ä r!#   ÿ„ â ‰       â !r       ärÿÿ   ÿä â ‰ €    " !r	
       â z j  ÿ„ â!r       â‰ €    $ z k   â !r       â/ ‰  €    b !r       ÆO ‰	 €    ¢ !r       Æ� ‰ €    â !r         Æ ‰ €    b !r       Æ‰ €    " $x!   ÿŽ Æ ‰ €    " !r       â ‚#   ÿàõ ä ‰  €    " !r         â/ ‚ÿÿ   ÿä â !r
	       âO ‰ @    b r!#   ÿˆ â !r       Ä� ‰  @    ¢ Š j  ÿ„ â !r       ä‰ @    â Š k   â $x!   ÿŽ â !r       â‰
 @    " r"ÿÿ   ÿä â !r       Ä ‰ @    b z j  ÿˆ â !r        ä ‰ @    " ‚	#   ÿà÷ ä z k  " â ‰ @    " $x"!   ÿŽ â !r       â/ $‚$ÿÿ   ÿŽ â ‰ @    b !r         âO    ÿàÿ Ä ‰ @    ¢ !r       â�    ÿàÿ ä ‰  @    â !r
       âŠ	 j  ÿ† â !r       â‰       " r"#   ÿàù â !r       Ä ‰      " Š		 k  $� â $rÿÿ   ÿ â !r       â ‰      b !#   ÿà÷ â !r       ä/ ‰      b z j  ÿˆ â !r       ÄO ‰      ¢ $ÿÿ   ÿäÿ â !r         ä� ‰      â 
! j  ÿ† ä "#   ÿàõ â ‰      â z k   â !r%        â! k  $� â ‰      " #ÿÿ   ÿä â !r       â " j  ÿ„ â ‰      " !r       Ä" k  # â ‰!       b !r       â/ !r       äO †‰ %    â !r       Æ� †    â !r       Æ †‰     â !r       Æ† 
    â !r       Æ †‰     â !r! !       Æ†    è †‰     è †     è †‰     â M™        € ê †y !   â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y)      "   " ¹z  l      â Ey   %    € â ™x?    â y*      !   b ¹z  F   
   â x,      ÿ  â r. ÿ       â $rAÿÿ   ÿ Ž â x"      ÿ  â $r ÿÿ   ÿ Ž â x$      ÿ  â x&      ÿ  â r( ÿ       Ä r )ÿ   pRð är )ÿ   pRò â$x2)    *Ž Ê/ xÿ   2  È r2   ÿ� È x+ÿ     È | +   pbô Ú G)  `$    € ê y      '   " $rÿÿ   ÿ Ž â xøÿÿÿÿÀŽ â $rÿÿ   ÿ Ž â y      %   b x6 B       â $v=ÿ q  ÿ Ž â r22  €ÿàÿ â y      &   ¢ x@+   ÿ   â $r,ÿÿ   ÿ Ž â x?ÿ   2  Ä r1=@   ÿŽ â$x;=   @Ž âz>@ q  ÿàÿ â $x3=   @Ž ä$x==   @Ž ä 'z y   Ž È $x   ÿ Ž â/ zÿ z    â $z5 }  ÿŽ ä 'z s   Ž ÈO $z	 p  +Ž â zÿ t    â $x2   ÿ Ž â x   ÿàÿ â $z |  	Ž ä$z	 |  +Ž ä %x   ÿŽ È $z v  Ž â x4   ÿüŽ â $z	 v  	Ž ä %v X  6Ž È $z
 w  ÿŽ â x8   ÿà÷ â %xB   Ž Æ r5
5   ÿàõ â $z p  	Ž â z X  ÿàù â $r9ÿÿ   Ž ä %v6 X  6Ž â x   ÿà÷ Æ $r<ÿÿ   6 Ž â zÿ Y  „þ â $rBÿÿ   ÿ Æ r<4   ÿàõ â �yúÿÿ • ¦ r	7   ÿä â �yüÿÿ • è �y
þÿÿ • ( �y    • h �yI   • ¨  �y:   • ¨  �yCþÿÿ • ¢ x        Æ �y6    • ¢ x[ $       â $rÿÿ   8 Ž ä $rÿÿ   9 Ž Ê �yRüÿÿ • ¤ r <       â/ $rÿÿ   7 Ž Ê �yT    • b rW25   =àÙ È r?B   ÿ¤~ â/ $x
   ÿ Ž Êxø  ÿÀŽ È z  ÿàõ Ê $vÿ  ÿ Ì �y    › " xÿ   
  Ê %v   Ž Ê �y    › " $x
   ÿ Ž ÊO x

ø  ÿÀŽ È z

  ÿàõ ä x   ÿ   Æ� $vÿ  ÿ â xÿ     ä xø  ÿÀŽ Æ �y

    › ¢ %v   Ž â z  ÿàõ Ê �y    › â $vÿ  ÿ ä $x   ÿ Ž âxÿ     ä �y    › b r25   @àµ ä xø  ÿÀŽ â %v   Ž â r?B   ÿd~ Ä xÿ     ä z  ÿàõ â �y    › ¤ %v   Ž È $vÿ  ÿ ä %v Z  [ Ž â �y    › ¦ $x$   ÿŽ â �y    › ¦ $x   Ž â r25   >àµ È �yF   ™ ¢  r?B   ÿd~ Æ �yJ   ™ ¨  �yG   ™ ¢  %v Z  [ Ž Æ �yH   ™ ¢  $x$   ÿŽ Æ �y/   ™ ¤  r   ÿàÿ ä �yQ   ™ ¨  �yK   ™ ¨ �yE   ™ ¨  �yP   ™ ¨ �yO    ™ ¨  �yN   ™ ¢ r25   1àµ Æ �yM   ™ ¢ r?B   ÿd~ Æ �yD   ™ ¨ �yL   ™ ¢ %v Z  [ Ž â r 25   ;àµ ä �y0   ™ ¢ $x$   ÿŽ Æ �yS    ™ ¢ $x   Ž â r?B   ÿd~ ä�yU    • ¢  %v   Z  [ Ž â r25   3àµ ä �yV    • ¢ $xY$   ÿŽ â rX?B   ÿd~ Ä �y^   ™ ¤ r!Y   ÿàÿ ä �y!   ™ ¢ %vW Z  [ Ž Æ �y   ™ ¢  %v Z  [ Ž Æ �y]   ™ ¢  $xX$   ÿŽ ä $xW$   ÿŽ â �yX   ™ ¢  $rÿÿ     Ž â r   ÿàÿ ä �y    ™ ¢  $x   WŽ Æ �y   ™ ¨ �yW   ™ ¨ �y[   ™ ¨  �yY    ™ ¨  �y_    • ¢  xdÿ   I  Æ �y\   ™ â xsdÿ   ÿÀŽ Æ �yd   ™ â  xfIÿ   ÿÀŽ Æ �yZ   ™ è �yc   ™ h �yb   ™ h �ya   ™ h �y`    ™ h �ye    • b 	sk f        & �yj   ™ h �yp   ™ h �y   ™ b# 	sr s        & �yi   ™ h �yo   ™ h �yh   ™ h �yn   ™ h �yg   ™ h �ym   ™ b xqk   ÿ   Æ �yf   ™ h �yl   ™ b $xtr€   ÿ Ž â xrq€   ÿÀŽ ä �y   ™ h �yk   ™ b xrrÿ   IxŽ Æ �y    ™ b xuÿ   :  Æ �y    • b xIIT  :    â $xrrÿŽ â xss€   txŽ â �y    ™ b! xquÿ   ÿÀŽ ä xtÿ   I  â �y    • b xu     â $twÿ @€ÿ Ž â 	sI q        b $xssÿŽ â xx     Ä xtÿ   ÿÀŽ ä xtr€€€€uìŽ â$tzÿ @€ÿ Ž â xvr€€€€wìŽ ä xutÿþþþÿàÿ â 	s         " xts€€€€xìŽ ä xwvÿþþþÿàÿ Ä xuurøŽ ä xvs€€€€zìŽ ä xttÿþþþÿàÿ ä xww @€røŽ ä xuu˜º  ÿ    ä xxvÿþþþÿàÿ Ä xrtsøŽ ä xvw˜º  ÿ    ä xt€€€€uÞŽ äxsx @€søŽ ä xuuÿÀŽ ä xxI   ÿ   Ä/ xwv€€€€ÞŽ â $xut   u
Ž â xvvÿÀŽ ä x   ÿ   ä xxq€   xxŽ ä xtCÿ   ÿÀŽ â $xvw   v
Ž â xI€   xŽ â $xxÿŽ â xu€€€€´Ž â $tuÿ @€ÿ Ž â xq     â 	st t        " xv€€€€´Ž Ä x€€€€qìŽ äxu€€€€uìŽ â $xIIÿŽ â xv     â $txÿ @€ÿ Ž â xqÿþþþÿàÿ ä xuuÿþþþÿàÿ ä xI€€€€vìŽ Ä xqqøŽ ä xwI€€€€xìŽ ä xu @€øŽ ä xuÿ   C  ä xvÿþþþÿàÿ ä xxwÿþþþÿàÿ Ä xtt   ÿ   ä xuuÿ   ÿÀŽ ä xvIøŽ äxIx @€IøŽ â 	sw u        " xtt€   ÿÀŽ ä xxÿ   6  Ä xyr˜º  ÿ    ä xrCT  6    ä xvtÿ   CxŽ ä xCxÿ   ÿÀŽ ä xxyÿÀŽ ä xtÿ   r  Ä xy€€€€yÞŽ â 	sr C        b xzs˜º  ÿ    ä xstÿ   ÿÀŽ â $xyy   x
Ž â x}zÿÀŽ â $xvvÿŽ â xxz€€€€	ÞŽ â 	st s        " xz     â $x|w€   ÿ Ž Æ xzv€€€€zìŽ â $t{ÿ @€ÿ Ž â rxx}  €ÿàÿ ä xwzÿþþþÿàÿ ä x|u€   |xŽ ä x	x€€€€	´Ž ä xuwvøŽ â $twÿÿ Ž â xxr   ÿ   â/ $xr|ÿŽ â x{v€€€€{ìŽ Ä xy€€€€´Ž ä xy  @€    â $xzt€   ÿ Ž âxxC€   xxŽ ä x{{ÿþþþÿàÿ ä xtr€€€€wìŽ äxyr€€€€yìŽ â $xxxÿŽ â xv{ @€vøŽ â $t{ÿ @€ÿ Ž â xws€   zxŽ Ä xC     ä xstÿþþþÿàÿ ä xyyÿþþþÿàÿ ä xssrøŽ äxty @€røŽ ä xCx€€€€CìŽ Ä xrx€€€€{ìŽ â $xwwÿŽ â xz     ä xCCÿþþþÿàÿ â $t|ÿ @€ÿ Ž â x{rÿþþþÿàÿ ä xzw€€€€zìŽ ä xrCxøŽ Ä xC{ @€xøŽ ä x{u˜º  ÿ    ä xyw€€€€|ìŽ ä x|zÿþþþÿàÿ ä xzv˜º  ÿ    ä xv{ÿÀŽ Ä x{
€€€€{ÞŽ äO xss˜º  ÿ    ä xu|wøŽ ä r{{v  €ÿàÿ ä xvt˜º  ÿ    ä xtsÿÀŽ Ä xs€€€€sÞŽ ä� xuu˜º  ÿ    ä xxyÿþþþÿàÿ ä rsst  €ÿàÿ ä xtuÿÀŽ ä xu€€€€uÞŽ Äxwx @€wøŽ ä ruut  €ÿàÿ ä xtC˜º  ÿ    ä xyzÿÀŽ äxxw˜º  ÿ    ä xzz€€€€ÞŽ Ä xwr˜º  ÿ    ä xCtÿÀŽ äxtt€€€€ÞŽ â $xzz   y
Ž â xrwÿÀŽ ä xw€€€€wÞŽ ä xyvÿÀŽ Ä x˜º  ÿ    â $xtt   C
Ž â xvv€€€€ÞŽ ä rwwr  €ÿàÿ ä xrÿÀŽ ä xC€€€€ÞŽ â $xvv   y
Ž â x
{€€€€
´Ž Ä xt€€€€´Ž ä xqq˜º  ÿ    ä xtI˜º  ÿ    â $xIC   r
Ž â xz€€€€´Ž â &rC
F   ÿ   âxv€€€€´Ž â &rFF   ÿ   â xvqÿÀŽ Ä xq€€€€qÞŽ â &rCJ   C   âxs€€€€´Ž â &rFJ   F   â xsxÿÀŽ äxxx€€€€ÞŽ ä x˜º  ÿ    ä rqv  €ÿàÿ â &rCG   C   âxqtÿÀŽ â &rFG   F   â xtt€€€€ÞŽ â $xxx   s
Ž â xGÿÀŽ Ä x€€€€ÞŽ â &rCH   C   âxu€€€€´Ž â &rF	H   F   â rttq  €ÿàÿ ä xI€€€€´Ž â $xH   G
Ž â xx€€€€´Ž â &r/   C   âxt€€€€´Ž â &rF/   F   â xw€€€€´Ž â &rQ      âx€€€€´Ž â &rFQ   F   Ä &r
K   ÿ   â xH€€€€´Ž â &rE      ä&rHE   F   ä &rP      ä &rKK   ÿ   ä &rFO      âxÿ   :  â &rOO   H   Ä &rHN      ä &rKP   K   â xÿ   6  â &r/M   H   â x   ÿüŽ â &rNN   K   â x   ÿüŽ â &r/D   /   ä $rOO   ÿŽ Ä &rM	M   N   ä &rCL   /   â x6ÿ   O  â $rFF   ÿŽ ä &rMD   M   ä &r:0   C   â x/ÿ   F  â &rML   M   â r66O   ÿ� â &r:S   :   â r//F   ÿ� â &r00   M   â x6ÿ   6  â 0rUÿU    A   Ä 0rTÿT    A   ä $r::   ÿŽ â x/ÿ   /  â &rSS   0   â Er6 6       â 0rRÿR    A   â  r0UT     A  â xEÿ   :  ä Er/ /       â  rCRU     A  â #r'06   '   â &r6
!   ÿ   ârEE:   ÿ� â &r!!   ÿ   â #r(C/   (   â &rC   6   â xEÿ   E  â 0r/ÿV    A   Ä &rG   !   â ErE E       â &rC    C   â r!R/     A  â &rG    G   ä &r
   ÿ   ä &rCX   C   â#r&!E   &   â &rG	X   G   Ä $rSS   ÿŽ ä &r!W      ä &r^   C   ä&r^^   G   â x0ÿ   S  â &r]      ä&r^]   ^   â r00S   ÿ� â &r    ÿ   Ä &r[      ä&r^[   ^   â x0ÿ   0  â &rWW       ä &r Y      ä&rYY   ^   â Er0 0       â  rT/     A  Ä $rYY   ÿŽ ä #r%0   %   Æ xÿ   Y  â $r     ÿŽ Æ rY   ÿ� â 0r_ÿ_    A   â xÿ      ä xÿ     ä r    ÿ� â  r T_     A  â Er        Ê #r#    #   â &r 
d   ÿ   ä&rdd   ÿ   ä &r

   ÿ   ä&r   ÿ   ä &rdj   d   ä&r j       Ä &r
p   
   ä&rp      ä &rW\   W   ä&rdi   d   ä&r!\   !   ä&r i       Ä &r
o   
   ä&ro      ä &rW	Z   W   ä&rd	h   d   ä &r!Z   !   ä&r h       Ä &r
n   
   ä&r	n      ä &rWc   W   ä&rdg   d   ä &r!c   !   ä&r g       Ä &r
m   
   ä&rm      ä &rWb   W   ä&rdf   d   ä &r!b   !   ä&r f       Ä &r
l   
   ä&rl      ä &rWa   W   ä&rd   d   ä &r!a   !   ä&r        Ä &r
k   
   ä&rk      ä &rW`   W   ä&r   d   ä &r6`   !   ä&r        Ä &r
   
   ä&r      ä $r66   ÿŽ ä$rWW   ÿŽ ä$r     ÿŽ ä $r   ÿŽ Ä $r

   ÿŽ ä $r   ÿŽ â x  ÿàõ ä x++   ÿàÿ ä x!ÿ   6  ä xÿ   W  ä x	ÿ      Ä xÿ     ä xÿ   
  ä xÿ     ä rÿ   ÿä ä xÿ     ä | +   pbô Ä r6!6   ÿ� ä rW   ÿ� ä r		    ÿ� ä r   ÿ� ä r
   ÿ� ä r   ÿ� â  r!R_     A  â Er        â 0reÿe    A   Ä 0rÿ    A   â x6ÿ   6  â 0rÿ    A   â xÿ     ä x	ÿ   	  ä xÿ     ä xÿ     ä xÿ     â #r$!   $   â  rRe     A  â rR     A  âEr6 6       â  rRR     A  â Er        â  reTe     A  âEr	 	       â  rT     A  âEr        â  rT     A  â Er        Ä Er        ä x<<  ÿà÷ ä x88  ÿàù â #r"6   "   â #r e       â #rA	   A   â #r.   .   â #r-R   -   â #r,   ,   â x11@   ÿàÿ â $r7ÿÿ   7Ž â x;;@   ÿàÿ Ä x33@   ÿàÿ ä x==@   ÿàÿ ä x>>@   ÿàÿ ä x@@@   ÿàÿ ä r9ÿ9   ÿä â D)       À â Gy  pÞÿÿÿÿƒ ê Ay        € ê Ey  ð     € â G‰  Ð     € ê $x))€  *Ž Ê ˆs )( úÿ H   è ˆs )'€úÿ H   è ˆs )& ûÿ H   è ˆs )%€ûÿ H   è ˆs )$ üÿ H   è ˆs )#€üÿ H   è ˆs )" ýÿ H   è ˆs ) €ýÿ H   è ˆs )A þÿ H   è ˆs ).€þÿ H   è ˆs )- ÿÿ H   è ˆs ),€ÿÿ H   ä Ay        € ê {          ì M        € ê „y*     H   b Ey  @    € æ „y* €   H   ¨ „y*    H   â !r(       â/ !r''       È_ ‰      " !r&&       Æ� ‰'      h ‰&      ¢ !r       Æ y      %   " !r'       Æ/ ‰      b !r	&       ÆO ‰      ¨ ‰	      â !r       â/ $x   ÿ Ž â !r       ÆO ‰ €    " z  r  p`ð â !r
	       ä� ‰ €    b r *ÿ   pVp  Æ ‰
 €    ´ ‰      '   â ˆ        â !r       â x   ÿàÿ â !r       Æ/ ‰ @    " z  r  p`ò â !r
       äO ‰	 @    b x *   p"ò Æ ‰ @    ¨ ‰      &   ( „y
* €  H   è       &   b !r       Æ       '   " !r		       Æ/ ‰      b !r       ÆO ‰	      ¢ $Š x  Ž Æ‰      " $Š ~  Ž ä� $ÿ   ÿ Ž ä $ x  Ž È$ ~  Ž â !r       â/ %† j   Ž â !r	       ÆO % j   Ž â †‰     â !r       Æ†    â G	  �     € ê y      &   ( y      '   b $z x  Ž È $z ~  Ž Ê/ z r  ÿàõ È rÿÿ   ÿä ä z j  ÿ„ È z k   Ê †y     è Ay        € ê !r

%       â „y*    H   â Ey  Ð    € æ ‰
      ( „y* €  H   è „y	*    H   è „y* €  H   è „y* €  H   è „y*    H   â !r
       Æ „y
*    H   è ‰      $ !r       Ê ‰ €    $ !r       Ê ‰ @    $ !r       Ê ‰      $ !r       â G™  °     € ê y      &   " ¹z  r      ä �x   ?àÿ â y      '   b $z x  Ž È $z ~  Ž Ê/ |   ÿàõ Ê $rÿÿ   ÿ â z j  ÿ„ È z k   Ê †y     è Ay        € ê „y* €  H   â !r$       â !r#       â !r		"       â !r        â !rA
A       â !r.       â !r-       â ‰       b z  r      Æ ‰      ¤ $x   ÿŽ ä ‰	      è ‰
      ( ‰A      h ‰      ( ‰      " !r         â/ !r       ÄO y      &   b !r	       â� !r,       ä ‰       ¢ !r

       Æ‰      â !rA       Æ‰      " !r       Æ ‰	
      b !r       Æ ‰      " $z x  Ž Æ/ ‰      h ‰      b !r        ÆO ‰      ¢ !r       â� !r       ä‰  €    â !r	
	       Æ‰ €    " !r       Æ ‰ €    " !r       Æ/ ‰
	 €    b !r       Æ ‰ €    b !r       ÆO ‰ €    ¨ ‰ €    ¢ !r         Æ� ‰      â !r       Æy      '   " !r       â !r
	
       ä/ ‰  @    " !r       Æ‰	 @    â !r       ÆO ‰ @    â !r       Æ ‰
 @    b !r       Æ� ‰ @    ¢ $z ~  Ž Æ‰ @    â $x   ÿŽ Æ ‰ @    " r   ÿ„ â !r         â ‚   ÿà÷ â ‰ €    $ $rÿÿ   ÿ â ‚ÿÿ   ÿäÿ ä Š j  ÿ† ä z j  ÿ„ â !r
       â/ Š k  � â !r	       â z k   â !r       â !r       âO ‰	       b r   ÿ„ â !r       â�    ÿàÿ â ‰      ¢ ‚   ÿàù â !r       â   ÿà÷ â ‰      â $rÿÿ   ÿ â !r       â z j  ÿ„ â ‰      " $ ÿÿ   ÿŽ â  j  ÿ† Ä ‰      " ‚ÿÿ   ÿä ä Š
 j  ÿˆ â ‰      b z k   ä Š k   â ‰ @    b !r 	       â/ 	 k   � Ä ‰      b !r       âO !r       ä� †‰     â !r       Æ †    â !r       Æ†‰     â !r       Æ†     â !r       Æ †‰     â !r       Æ/ †    è ‰       ( †‰ 
    â M™        € ê x   ÿàÿ âO !r        Æ r   ÿàñ È rÿÿ   ÿä  ä z j  ÿ€ È z k    Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y"      "   " ¹z  l      â Ey  p"    € â ™x?    â y#      !   b ¹z  F   
   â $r%ÿÿ   ÿ Ž â x&      ÿ  â r8 ÿ       â $r:ÿÿ   ÿ Ž â x      ÿ  â x       ÿ  â r "ÿ   pRð Ä r "ÿ   pRò â$x+"    #Ž Ê/ x ÿ   +  È r +   ÿ� â $r ÿÿ   ÿ Ž Æ x$ÿ     È | $   pbô Ú G)  @!    € ê y      '   " $rÿÿ   ÿ Ž â xøÿÿÿÿÀŽ â $rÿÿ   ÿ Ž â y      %   b x
 B       â $v4ÿ q  ÿ Ž â x7$   ÿ   â y	      &   ¢ $x++   
Ž â r% ÿ       Ä $x*4   7Ž âz57 q  ÿàÿ â $x,4   7Ž âx6ÿ   +  â $x44   7Ž ä 'z y   Ž â x   ÿ   Æ/ $z. }  ÿŽ â zÿ z    â $x+   ÿ Ž ä 'z	 s   Ž ÈO $z p  $Ž â zÿ t    â $z |  $Ž ä$z |  Ž ä %x   ÿŽ È $z v  Ž â x-   ÿüŽ â $z v  Ž â x   ÿàÿ â %v X  
Ž È $z	 w  ÿŽ â x0   ÿà÷ â %xB   Ž Æ r..   ÿàõ â $z p  Ž â z X  ÿàù â $r3ÿÿ   Ž â r9ÿÿ   ÿä â %v X  
Ž â x   ÿà÷ Æ $r2ÿÿ    Ž â zÿ Y  	„þ â $r1ÿÿ    Ž Æ r2-   ÿàõ â �yúÿÿ • ¨ $x   1 â �y	üÿÿ • è �y
þÿÿ • ( �y    • h �ye   • ¨ �y<   • ¨ �ycþÿÿ • ¢ $tÿ   ÿ Ž Æ �y>    • ¢ xY $       â $rÿÿ   0 Ž ä $rÿÿ   3 Ž Ê �y)üÿÿ • ¢  rU+.   4àÙ È rV69   ÿ¤~ â $x
   ÿ Ž Êxø  ÿÀŽ È z  ÿàõ È zÿ  ÿä Ì �y    › " xÿ   
  Ê %v   Ž Ì �y    › " $x   ÿ Ž ÊO xø  ÿÀŽ È z
  ÿàõ â $x	   ÿ Ž Æ� zÿ  ÿä ä xø  ÿÀŽ ä xÿ     ä �y

    › ¢ z  ÿàõ ä %v   Ž â xÿ   	  Æ $vÿ  ÿ â r+.   7àµ â �y    › â $x   ÿ Ž är	69   ÿd~ â �y    › b %v   Ž â xø  ÿÀŽ ä xÿ     â %v Z  Y Ž Ä �y    › ¤ $x		$   ÿŽ È $x   	Ž ä %v   Ž â z  ÿàõ ä �y=   ™ ¨ �y    › ¢ zÿ  ÿä Æ �yM   ™ ¨ �y    › ¨ �yO   ™ ¨ �yN   ™ ¢ r+.   5àµ Æ �yL   ™ ¢ r69   ÿd~ Æ �yK   ™ ¢ r+.   *àµ â %v Z  Y Ž ä �yJ   ™ ¤ $x$   ÿŽ â r(69   ÿd~ â �yI    ™ ¦ r   ÿàÿ â %v Z  Y Ž â �y;    • ¦ $x/($   ÿŽ â �y?   ™ ¦  $x   /Ž â �y(   ™ ¨  �y@   ™ ¨  �y/   ™ ¨  �yF   ™ ¨  �yH   ™ ¨  �yD   ™ ¨  �yG   ™ ¨  �yB   ™ ¢  $rÿÿ   2 Ž Æ/ �yE   ™ ¢ r 1       Æ �yA   ™ ¨  �yC   ™ ¨ �yS   ™ ¨  �yP    • ¨  �yT    ™ ¨ �yR   ™ ¨ �yQ    ™ ¢ r+.   ,àµ â %vU Z  Y Ž Ä �y    • â  rW69   ÿd~ â $xUV$   ÿŽ ä %v Z  Y Ž â �y    • æ $xWW$   ÿŽ ä $x   UŽ â xUÿ   e  â $x   WŽ Æ xgUÿ   ÿÀŽ â �y   ™ è" �yU   ™ h �y[   ™ b xVeÿ   ÿÀŽ Æ �yb   ™ b 	s] V        & �yZ   ™ h �ya   ™ h �yY   ™ h �y`   ™ h �yX   ™ h �y_   ™ h �yW   ™ b xd]   ÿ   Æ �y^   ™ h �yV   ™ h �y]   ™ b 	sf g        & �y    ™ h �y\    • h �y    ™ b# $xhf€   ÿ Ž â xfd€   ÿÀŽ Ä �y    • d xffÿ   exŽ ä xeeT  <    äxiÿ   <  ä xhg€   hxŽ ä xgÿ   e  â $xffÿŽ â xdiÿ   ÿÀŽ â $tiÿÿ Ž â xgÿ   ÿÀŽ Ä/ xk  @€    â 	se d        " xf€€€€iìŽ â $xhhÿŽ â xgf€€€€kìŽ â $toÿ @€ÿ Ž È 	sl         b xÿþþþÿàÿ ä xggÿþþþÿàÿ â $tmÿÿ Ž â xjh€€€€oìŽ ä xføŽ äxfg @€føŽ ä xkjÿþþþÿàÿ Ä xih€€€€mìŽ ä xjf˜º  ÿ    ä xfk @€høŽ ä xke   ÿ   â $xel€   ÿ Ž â/ xiiÿþþþÿàÿ ä x˜º  ÿ    Ä xgihøŽ ä xh€€€€ÞŽ äxe€   exŽ ä xÿÀŽ ä xcÿ   ÿÀŽ ä xkd€   kxŽ Ä xij€€€€ÞŽ äxjjÿÀŽ â $xh   
Ž â 	s         " $tdÿÿ Ž â rjij  €ÿàÿ â $xkkÿŽ ä $thÿ @€ÿ Ž â x€€€€´Ž â $tiÿ @€ÿ Ž â xj€€€€´Ž â $xjeÿŽ â xk€€€€dìŽ Ä xdk€€€€hìŽ ä xe     ä xhÿþþþÿàÿ ä xldÿþþþÿàÿ ä xdj€€€€eìŽ äxij€€€€iìŽ Ä xehkøŽ äxl @€køŽ ä xhÿ   c  ä xkdÿþþþÿàÿ ä xmiÿþþþÿàÿ â $xi€   ÿ Ž â xhhÿ   ÿÀŽ Ä xdkjøŽ ä xkg˜º  ÿ    ä xli€   ÿÀŽ ä xgcT  >    ä xm @€jøŽ ä xiÿ   >  â 	sj h        " xllÿ   cxŽ Ä xgÿ   g  ä xciÿ   ÿÀŽ ä xpf˜º  ÿ    ä xnkÿÀŽ â 	si c        b xk€€€€kÞŽ ä xfgÿ   ÿÀŽ Ä xgpÿÀŽ â $xllÿŽ â xpp€€€€ÞŽ â $tmÿÿ Ž â rqkn  €ÿàÿ ä 	sk f        " $xpp   g
Ž â xgl€€€€mìŽ äxml€€€€oìŽ Ä xoj   ÿ   â $xji€   ÿ Ž Æ/ xoh€   oxŽ ä xggÿþþþÿàÿ ä xmmÿþþþÿàÿ â $xooÿŽ â xp€€€€´Ž â $tnÿÿ Ž â xp  @€    â $xkk€   ÿ Ž âxggløŽ Ä xhm @€løŽ ä xlc€   jxŽ ä xco€€€€nìŽ äxio€€€€pìŽ â $xmlÿŽ â xkf€   kxŽ ä xfcÿþþþÿàÿ Ä xjiÿþþþÿàÿ â $xlkÿŽ â xifoøŽ äxjj @€oøŽ â $toÿÿ Ž â xcm€€€€nìŽ ä xq€€€€´Ž ä xkm€€€€pìŽ Ä xq  @€    ä xfcÿþþþÿàÿ ä xpkÿþþþÿàÿ ä xkl€€€€oìŽ äxnl€€€€qìŽ ä xcfmøŽ Ä xfp @€møŽ ä xmkÿþþþÿàÿ ä xonÿþþþÿàÿ ä xkg˜º  ÿ    ä xnh˜º  ÿ    ä xhmløŽ Ä xgo @€løŽ ä xlkÿÀŽ ä xk
€€€€kÞŽ äO xmnÿÀŽ äxnn€€€€ÞŽ â $xkk   l
Ž â xii˜º  ÿ    Ä xlj˜º  ÿ    â $xnn   m
Ž â x
k€€€€
´Ž ä xkh˜º  ÿ    ä xjiÿÀŽ ä xmlÿÀŽ ä xi€€€€iÞŽ Ä� xll€€€€ÞŽ ä xhkÿÀŽ ä xk€€€€kÞŽ â$xll   m
Ž â riij  €ÿàÿ ä xjg˜º  ÿ    ä xcc˜º  ÿ    â $xkk   h
Ž â xgjÿÀŽ Ä xhcÿÀŽ ä xl€€€€´Ž ä xjj€€€€ÞŽ ä xc€€€€cÞŽ ä xlf˜º  ÿ    ä xee˜º  ÿ    â $xcc   h
Ž â rfjg  €ÿàÿ Ä xglÿÀŽ äxll€€€€ÞŽ ä xjeÿÀŽ ä xe€€€€eÞŽ ä xff€€€€´Ž â $xll   g
Ž â xc€€€€´Ž Ä xn€€€€´Ž ä xcd˜º  ÿ    â &rd
=   ÿ   â reej  €ÿàÿ ä xj˜º  ÿ    â &rdM   d   â xi€€€€´Ž ä xl€€€€´Ž â &r=   ÿ   â xjÿÀŽ Ä xhcÿÀŽ ä xjj€€€€ÞŽ â &rdO   d   â xc€€€€cÞŽ â &rM      â xk€€€€´Ž â $xjj   
Ž ä $xcc   h
Ž ä &rN   d   Ä &rO      â xh˜º  ÿ    â &rL      â xc€€€€´Ž â &rN      â x=hÿÀŽ â &rfK      â xhh€€€€	ÞŽ â &rL      â xj€€€€´Ž â &rJ      â x>ÿ   >  Ä rhh=  €ÿàÿ â &rK      â xe€€€€´Ž â &rI      âx>   ÿüŽ ä x>h€€€€	´Ž â &rJ      ä $r   ÿŽ ä &rJ>I      â x	ÿ   <  â &r
?   ÿ   âx<ÿ     â &r??   ÿ   Ä &r(      âr<<   ÿ� â &r?(   ?   ä &r=
@   ÿ   ä &r/      âx		   ÿüŽ â &r?/   ?   â x<ÿ   <  â 0r)ÿ)    A   Ä 0r;ÿ;    A   ä &r=F   =   ä &rH      â Er< <       â &r?H   ?   â  r();     A  â &r=D   =   ä $rJJ	   ÿŽ Ä &rG      ä&r?G   ?   â #r!(<   !   â &r=B   =   â xÿ   J  â &r(fE      ä&r<E   ?   ä &r=A   =   â rJ   ÿ� â &rC   (   Ä &rCC   <   ä &r=fS   =   â xÿ     â 0rPÿP    A   ä &rT      ä&rT>T   C   ä &r=R   =   â Er        â  r;;P     A  â $rT	T   ÿŽ Ä &r=Q   =   â #r ;       â $r   ÿŽ âxÿ   T  â $r==   ÿŽ Æ rT   ÿ� ä x<ÿ   =  â &r@@   ÿ   â x(ÿ     â 0rÿ    A   â xÿ     ä r=<=   ÿ� â &r/F   @   â r((   ÿ� â  rP     A  â Er        â 0rÿ    A   â x=ÿ   =  â &r/D   /   â x(ÿ   (  â #r      â Er= =       â  r<)     A  âEr( (       â  r)     A  â &r/B   /   Ä &r
U   ÿ   â #r =       â &r/A   /   â #r<(      â &r[      ä &r(U   ÿ   ä &r
   ÿ   ä&r   ÿ   Ä &r/S   /   ä &r[[   (   ä &rb      ä&rb      ä &r/R   /   ä &r
Z      Ä &rZZ   [   ä &ra      ä&ra      ä &r@>Q   /   ä &rY   
   ä&rYY   Z   Ä &r`      ä&r`      ä $r@	@   ÿŽ ä &rX      ä&rYX   Y   ä &r_      Ä &r__      â xÿ   @  â &rfW      ä&rYW   Y   ä &rff^      ä&r^   _   â r@   ÿ� â &rV      Ä &rYV   Y   ä &r]   f   ä&r]      â xÿ     â &r      ä&r      ä&r
>   Y   Ä &r>      â Er        â $r   ÿŽ â  rP     A  â $r
	
   ÿŽ âx  ÿàõ â $r   ÿŽ ä $r	   ÿŽ â x$$   ÿàÿ â #r:   :   â xÿ     â $rÿÿ    â xÿ   
  Ä xÿ     ä x	ÿ     ä | $   pbô ä r   ÿ� ä r

   ÿ� ä r   ÿ� Ä r		   ÿ� â 0r\ÿ\    A   â xÿ     â 0rÿ    A   â x
ÿ   
  ä xÿ     ä x	ÿ   	  â  r)\     A  â r\P\     A  âEr        â  r))     A  âEr
 
       â  rPP     A  â Er        Ä Er	 	       ä x22  ÿà÷ ä x00  ÿàù â #r8   8   â #r'\
   '   â #r&)   &   â #r%P	   %   â x**@   ÿàÿ â $r1ÿÿ   1Ž â x,,@   ÿàÿ Ä x44@   ÿàÿ ä x55@   ÿàÿ ä x77@   ÿàÿ ä r3ÿ3   ÿä â D)       À â Gy  �áÿÿÿÿƒ ê Ay        € ê Ey  Ð     € â G‰  °     € ê $x"@  #Ž Ê ˆs ! ûÿ H   è ˆs  €ûÿ H   è ˆs  üÿ H   è ˆs €üÿ H   è ˆs   ýÿ H   è ˆs :€ýÿ H   è ˆs 8 þÿ H   è ˆs '€þÿ H   è ˆs & ÿÿ H   è ˆs %€ÿÿ H   ä Ay        € ê {          ì M        € ê „y#     H   b Ey  @    € æ „y# €   H   ( „y#    H   ¨ y      %   ä $x   ÿ Ž â� !r!       È/ z  r  p`ð â !r         Æ r #ÿ   pVp  â ‰      " !r       ÆO ‰       h ‰      ¨ ‰      &   è ‰      '   " !r       â !r        È/ ‰      " !r       ÆO ‰      h ‰	      ¢ !r       â !r       È/ ‰ €    " !r		       ÆO ‰ €    h ‰
	 €    ¢ !r       â x   ÿàÿ â !r       Æ/ z  r  p`ò â ‰ @    " $Š x  Ž â� !r
	
       âO x #   p"ò â ‰ @    b $„ÿ   ÿ Ž ä $Š ~  Ž â‰
 @    ¨ „y	# €  H   è       &   â         Æ       '   " !r       â !r       È/ ‰      " !r
       ÆO ‰      h ‰      ¢ $ x  Ž È� $ ~  Ž â!r       â %† j   Ž â !r       Æ/ % j   Ž â †‰     â !r       ÆO †    â G	  �     € ê y      &   ( y      '   b $z x  Ž È $z ~  Ž Ê/ z r  ÿàõ Ê $rÿÿ   ÿ â z j  ÿ„ È z k   Ê †y     è Ay        € ê !r		       â „y#    H   â Ey  À    € æ ‰	      ( „y# €  H   è „y#    H   è „y# €  H   è „y# €  H   â !r	       Æ „y	#    H   è ‰      $ !r       Ê ‰ €    $ !r       Ê ‰ @    $ !r       Ê ‰      $ !r       â G™  °     € ê y      &   " ¹z  r      ä �x   ?àÿ â y      '   b $z x  Ž È $z ~  Ž Ê/ |   ÿàõ Ê $rÿÿ   ÿ â z j  ÿ„ È z k   Ê †y     è Ay        € ê !r        â !r:       â !r8       â !r'       â !r		&       â ‰       b !r%%       â z  r      ä ‰      ( ‰      ¨ ‰      è ‰
	      ( ‰%      b !r         Æ/ y      &   b !r       â !r       ÈO ‰      " !r       Æ� ‰      ¢ !r
	
       Æ‰       â !r%       Æ‰	      ( ‰
      b $z x  Ž Æ/ ‰      b !r       â !r       ÈO ‰ €    " !r        Æ� ‰ €    ¢ !r		       Æ‰  €    â !r
       Æ‰	 €    " !r       Æ/ ‰ €    h ‰ €    b !r       Æ y
      '   " !r       ÆO ‰ @    ¢ !r         Æ� ‰ @    â !r	       Æ‰  @    " !r       Æ/ ‰ @    b !r       Æ‰ @    b $z
 ~  Ž Æ ‰ @    " !r
       âO r   ÿˆ â!r       â� $x   ÿŽ â !r	        âr   ÿ„ ä rÿÿ   ÿä â !r       â/ ‰ 	      b $rÿÿ   ÿ â z j  ÿ„ â!r       â‰
      ¤ z k   â !r       â ‰      " ‚   ÿàõ Ä    ÿàÿ â ‰      ä $‚ÿÿ   ÿ â    ÿà÷ â ‰      " Š j  ÿ„ ä z j  ÿˆ â ‰      b Š k   â $ÿÿ   ÿŽ â  j  ÿ„ â !r	        â/ z k   Ä 	 k   â !r
       âO †‰     â !r       Æ †    â !r       Æ� †‰     â !r       Æ†     â !r       Æ†‰     â M™        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y       "   " ¹z  l      â Ey  p    € â ™x?    â y      !   b ¹z  F   
   â x       ÿ  â $rÿÿ   ÿ Ž â x.      ÿ  â x      ÿ  â r0 ÿ       Ä r  ÿ   pRð är  ÿ   pRò â$x"     Ž Ê/ xÿ   "  È r"   ÿ� È xÿ     È |    pbô Ú G)  `    € ê y      '   " $rÿÿ   ÿ Ž â xøÿÿÿÿÀŽ â $t(ÿB   ÿ Ž â y	      %   b r ÿ       â $x,   ÿ Ž ä y      &   ¢ $x""   
Ž ä $r!ÿÿ   ÿ Ž Æ x*ÿ   "  â 'z y   Ž È $x		   ÿ Ž â/ zÿ z    â $x"   ÿ Ž ä 'z s   Ž âO x	   ÿàÿ Æ $z	 p  Ž â zÿ t    â $z |  Ž ä$z |  Ž ä %x   ÿŽ È $z v  Ž â x'   ÿüŽ â $z v  Ž ä $z% }  ÿŽ ä %v X  (Ž È $z
 w  ÿŽ â x$   ÿàù â %xB   Ž Æ r%
%   ÿàõ â $z p  Ž â z X  ÿà÷ â $r+ÿÿ    â z  q      â %v( X  (Ž â x   ÿàù ä r#,   ÿŽ â$rÿÿ   ( Ž â zÿ Y  d~ â $x&   ,Ž â z(, q  ÿàÿ â $r-ÿÿ   ÿ Ä r'   ÿàõ â �yúÿÿ • ¨ $x   ) â �y	üÿÿ • è �yþÿÿ • ( �y    • h �yI   • ¨ �y<   • ¨ �yDþÿÿ • ¢ $t1ÿ   ÿ Ž Æ �y;    • ¢ $tQÿ$   ÿ Ž â rM +       â $rLÿÿ   $ Ž Ê �y3Lüÿÿ • ¤  $rLÿÿ    Ž ä $rMÿÿ   ) Ž Ê �y8L    • â  $x   ÿ Ž Êxø  ÿÀŽ È z  ÿàõ È zÿ  ÿä Ì �y    › " xÿ     Ê %v  1 Ž Ì �y    › " $x   ÿ Ž ÊO xø  ÿÀŽ È z  ÿàõ ä xÿ     Æ $vÿ  ÿ â x	   ÿ   â� %v  1 Ž È �y    › ¢ xø  ÿÀŽ Æ �y    › â z
  ÿàõ ä xÿ   	  ä xÿ     â$vÿ  ÿ â r"%   ,àµ â $x   ÿ Ž ä %v  1 Ž â r*-   ÿd~ â �y

    › b xø  ÿÀŽ â %v Z  Q Ž Ä �y    › ¢ z  ÿàõ â %v  1 Ž È $x$   ÿŽ ä $v	ÿ  ÿ â �y    › ¢ r>"%   (àµ ä r   ÿàÿ ä r*-   ÿd~ â �y    › ¨ �yF   ™ ¢ %v>> Z  Q Ž Æ �yH   ™ ¢ $x$   ÿŽ Æ �yJ   ™ ¢ $x??   Ž Æ �yE   ™ ¨ �y=>   ™ ¨  �yG   ™ ¨ �yA>   ™ ¨  �yC   ™ ¨ �y:>   ™ ¨  �y@   ™ ¨ �y2>   ™ ¨  �y6>   ™ ¨  �y5>   ™ ¢  r"%   #àµ Æ �yB    ™ ¨ �y1>   ™ ¢  r9*-   ÿd~ ä rN"%   &àµ â �y4>    ™ ¢  %v Z  Q Ž ä rO*-   ÿd~ â �y7    • ¢ $x99$   ÿŽ Æ $xOO$   ÿŽ â �yK>    • ¢  $x   9Ž ä %vN Z  Q Ž Æ/ �y9   ™ ¤ rO   ÿàÿ ä �yL   ™ è �yR   ™ è  �yN   ™ è �yW   ™ h �yM   ™ h �yU   ™ h �yX   ™ h �yT   ™ h �yV   ™ h �yS   ™ b xZIÿ   ÿÀŽ Æ �yP   ™ b x>ÿ   I  Æ �yQ   ™ h �y?    ™ b x]>ÿ   ÿÀŽ Æ �yO   ™ b 	sZ Z        & �y>    ™ h �yY    • b $x[Z€   ÿ Ž Æ �yZ    • b 	s^ ]        " x\[€   ÿÀŽ È x\\ÿ   IxŽ ä xIIT  <    äx[ÿ   <  â $x\\ÿŽ â x_  @€    â $x^^€   ÿ Ž â xIÿ   I  ä xa[ÿ   ÿÀŽ Ä x^]€   ^xŽ â $t]ÿÿ Ž â x[\€€€€_ìŽ äx_Iÿ   ÿÀŽ â 	sb a        " x\€€€€]ìŽ â/ $tcÿÿ Ž Ì 	s` _        b xÿþþþÿàÿ â $x^^ÿŽ â x[[ÿþþþÿàÿ ä x\øŽ â$tÿ @€ÿ Ž â x^€€€€cìŽ ä x˜º  ÿ    ä x[[ @€\øŽ â $xbb€   ÿ Ž â x]ÿþþþÿàÿ Ä x^€€€€ìŽ â $x``€   ÿ Ž â/ x[˜º  ÿ    ä xba€   bxŽ ä x[ÿþþþÿàÿ ä x`_€   `xŽ â $xbbÿŽ ä $taÿ @€ÿ Ž â xc  @€    â $x``ÿŽ â x€€€€ÞŽ ÄxÿÀŽ ä xI€€€€ÞŽ äxÿÀŽ â $x   
Ž â xDÿ   ÿÀŽ ä rI  €ÿàÿ ä x€€€€´Ž â 	s_         b  $tÿÿ Ž â x€€€€´Ž â $tIÿÿ Ž â xab€€€€aìŽ Ä x[ @€^øŽ ä x`€€€€cìŽ ä x[ÿ   D  ä xb€€€€ìŽ ä xcaÿþþþÿàÿ ä x\]^øŽ Ä xaÿþþþÿàÿ ä x[[ÿ   ÿÀŽ ä x^ÿ   ;  ä x`€€€€IìŽ ä xIÿþþþÿàÿ â 	s] [        " xa @€`øŽ Ä xÿþþþÿàÿ ä xa^ÿ   ÿÀŽ ä xIIbøŽ äxc @€bøŽ â $x__€   ÿ Ž â/ x^DT  ;    ä xc\˜º  ÿ    Ä xb˜º  ÿ    â 	s\ a        b x`øŽ ä x^ÿ   ^  ä x`cÿÀŽ ä xebÿÀŽ ä xc€€€€cÞŽ Ä xbb€€€€ÞŽ ä x__€   ÿÀŽ ä x^^ÿ   ÿÀŽ â $xbb   e
Ž â rc`  €ÿàÿ ä x__ÿ   DxŽ â 	s` ^        " x€€€€´Ž â $x]€   ÿ Ž â xb€€€€´Ž â $x__ÿŽ â xD     â $tbÿ @€ÿ Ž â x[[€   xŽ â $x\\€   ÿ Ž â/ x_€€€€DìŽ Ä xD_€€€€bìŽ â $tcÿÿ Ž â xaa€   \xŽ â $x\[ÿŽ â xÿþþþÿàÿ â $teÿ @€ÿ Ž â xDDÿþþþÿàÿ ä x]`   ÿ   äx[_øŽ Ä x_D @€_øŽ ä x\€€€€cìŽ äxD\€€€€eìŽ â $xcaÿŽ â x^^€   ]xŽ ä x]ÿþþþÿàÿ â $tÿÿ Ž â xaDÿþþþÿàÿ â $tDÿ @€ÿ Ž Æ xc€€€€ìŽ äxDc€€€€DìŽ â $xb^ÿŽ â x`]\øŽ ä xÿþþþÿàÿ ä x]     ä x^Dÿþþþÿàÿ Ä xaa @€\øŽ ä xDcøŽ ä x\b€€€€]ìŽ ä x^ @€cøŽ ä xcb€€€€eìŽ ä x[[˜º  ÿ    Ä x]\ÿþþþÿàÿ ä x^_˜º  ÿ    ä x_cÿþþþÿàÿ ä x\[ÿÀŽ ä xc]bøŽ ä x[€€€€[ÞŽ ÄO xb_ @€bøŽ ä x_`˜º  ÿ    â $x[[   \
Ž â x]^ÿÀŽ äx^^€€€€ÞŽ ä x\_ÿÀŽ ä x`a˜º  ÿ    Ä x_€€€€_ÞŽ â� $x^^   ]
Ž â x[€€€€´Ž ä x]`ÿÀŽ är[_\  €ÿàÿ ä x``€€€€ÞŽ ä xcc˜º  ÿ    Ä xbb˜º  ÿ    ä x[[€€€€´Ž ä x_D˜º  ÿ    â $x``   ]
Ž â xcÿÀŽ ä xc
€€€€cÞŽ äx]bÿÀŽ Ä xD˜º  ÿ    ä xbb€€€€ÞŽ ä x_ÿÀŽ ä x_€€€€_ÞŽ â $xcc   
Ž â xII˜º  ÿ    ä rb]  €ÿàÿ â $x__   
Ž â x]DÿÀŽ Ä xDD€€€€ÞŽ ä xIÿÀŽ ä xI€€€€IÞŽ ä x˜º  ÿ    â $xDD   ]
Ž â x˜º  ÿ    ä x^€€€€´Ž Ä x€€€€´Ž â &rF   ÿ   â r]I  €ÿàÿ â &rFF   ÿ   â xÿÀŽ ä xIÿÀŽ ä x€€€€ÞŽ â &rH      âx€€€€	ÞŽ â &rFH   F   â x`€€€€´Ž â $x   
Ž â x˜º  ÿ    â $x   I
Ž Ä &r[J      ä&rFJ   F   â x
c€€€€
´Ž â &rE      âx€€€€´Ž â &rFE   F   â xHÿÀŽ äx€€€€ÞŽ ä x€€€€	´Ž â &r=   ÿ   Ä &r
G      ä&rFG   F   ä &rA      â x_€€€€´Ž â &r	C      âx]€€€€´Ž â &rCC   F   ä &r[:      Ä &r	@   	   ä&rC@   C   ä &r@=   ÿ   ä &r2      â xD€€€€´Ž â &r@A   @   ä &r
6      Ä &r@:   @   ä &r5      ä &rB   	   â x	ÿ   ;  â &r=2   @   ä &r1      â x		   ÿüŽ â &r=6   =   Ä &r4      ä &r=5   =   ä $r	   ÿŽ ä $r	   ÿŽ ä &r=1   =   â x1ÿ     ä xÿ     Ä r11   ÿ� ä r   ÿ� â &r9   ÿ   âx1ÿ   1  â 0r3ÿ3    A   ä 0rKÿK    A   ä &r99   ÿ   ä &rL      â Er1 1       â &rR   ÿ   Ä &rR   ÿ   â  r3K     A  â &r[N      ä &r9L   9   ä &rW      ä&rW      â #r1      â &r1M      Ä &rNN   9   ä &r[[U      ä&rU      ä &r1
X   1   ä &rMM   N   ä &r[T   [   Ä &rT      ä &r1V   1   ä &rMX   M   ä &r[
S   [   ä&rS      â rHH  €ÿàÿ â &rP   1   Ä &rMV   M   ä &r[Q   [   ä&rQ      â xH€€€€´Ž â &r?      â xÿ   <  â &rPP   M   ä &r[O   [   Ä &rO      â x   ÿüŽ â 0r7ÿ7    A   â xÿ     â $r	   ÿŽ ä &rBB   C   ä&r44   =   ä&rP?   P   Ä &r>   [   ä&r>      â Er        â $rBB   ÿŽ â  r537     A  â $r44   ÿŽ âxÿ     â $rPP   ÿŽ â x   ÿàõ â $r		   ÿŽ Ä $r   ÿŽ â x   ÿàÿ â #r05   0   â r   ÿ� â $rÿÿ    â x;ÿ   B  ä xÿ   4  ä xÿ   P  ä xÿ   	  Ä xÿ     â 0rYÿY    A   â |    pbô ä xÿ     ä r;;B   ÿ� ä r4   ÿ� ä rP   ÿ� Ä r	   ÿ� ä r   ÿ� â 0r8ÿ8    A   â Er        â  r3Y     A  â x;ÿ   ;  â 0rZÿZ    A   â xÿ     ä xÿ     Ä xÿ     ä xÿ     â  r778     A  â Er; ;       â  rK8K     A  â Er        â #r/   /   â  rY8Y     A  âEr        â  r3Z     A  â Er        â  r88Z     A  â Er        Ä x   ÿà÷ ä x$$   ÿàù â #r 7;       â #rK      â #r.Y   .   â #r      â #r!8   !   â x##€   ÿàÿ â $r)ÿÿ   )Ž â x&&€   ÿàÿ Ä x((€   ÿàÿ ä x,,€   ÿàÿ ä r+ÿ+   ÿä â D)       À â Gy  Päÿÿÿÿƒ ê Ay        € ê Ey  °     € â G‰  �     € ê $x     Ž Ê ˆs  0 üÿ H   è ˆs   €üÿ H   è ˆs   ýÿ H   è ˆs  €ýÿ H   è ˆs  / þÿ H   è ˆs  .€þÿ H   è ˆs   ÿÿ H   è ˆs  !€ÿÿ H   ä Ay        € ê {          ì M        € ê „y     H   b Ey       € æ „y €   H   ¨ „y     H   ( „y €  H   è „y    H   ( „y	 €  H   h „y    H   h „y    H   h „y  	  H   h „y €  H   â !r0       Æ/ „y €	  H   â !r        âO !r         Æ !r       â� !r         Æ!r		       ä‰       " !r       Æ ‰	      b !r       È !r       Ê ‰      ¢ !r        Æ y       %   " !r	       Æ/ ‰      h ‰      â !r       ÊO ‰      ¢ $x     ÿ Ž â !r       È/ z   r  p`ð â !r       Æ� r ÿ   pVp  â ‰ €    ( ‰ €    b !r       ÎO ‰      '   ¨ ‰
 €    â !r       â x    ÿàÿ â !r       Æ/ ‰ @    " z  r  p`ò Æ ‰ @    b x    p"ò â !r

       ä� ‰      &   è ‰	
 @    , $ÿ   ÿ Ž â       &   b !r       Æ       '   " !r       Æ/ ‰      h ‰      h „y €  H   â $Š x   Ž â� ˆ        â !r	
	       Ä$Š ~  Ž äO $ x   Ž â‰	      ¦ $ ~  Ž â !r       â/ %† j   Ž â !r       Æ % j   Ž â †‰     è †    â !r		       âO G	  �     € ê y      &   ( y      '   b $z x   Ž È $z ~  Ž Ê/ z r  ÿàõ Ê $rÿÿ   ÿ â z j  ÿ„ È z k   Ê †y 	    è Ay        € ê !r       â „y    H   â Ey  0    € ä !r       â „y	    H   æ !r       â „y  
  H   è ‰      ( „y €  H   è „y €  H   è „y
    H   è „y    H   è „y    H   â !r       Æ „y €  H   è ‰      ( „y €
  H   è „y €  H   â !r       Ê ‰ €    $ !r       Ê ‰ @    $ !r       Ê ‰      $ !r       â G™  °     € ê y      &   " ¹z  r      ä �x   ?àÿ â y      '   b $z x   Ž È $z ~  Ž Ê/ |   ÿàõ È rÿÿ   ÿä ä z j  ÿ„ È z k   Ê †y     è Ay        € ê „y €  H   â !r/       â !r.       â !r

       â !r!       â !r	       â !r       â !r

       â !r       â !r       â !r       â !r

       â y      &   h ‰      ( ‰      ¨ ‰
      â $z x   Ž â/ !r       â !r       â !r       ÆO ‰      " !r
       Æ� ‰      h ‰	      ¨ ‰      â !r       â !r       È/ ‰ €    " !r	       âO !r	       È� ‰      h ‰ €    ¨ ‰
	 €    â !r       Æ y      '   ( ‰ @    " !r       â/ !r       ÈO ‰ €    b !r
	
       Æ� ‰ @    ¨ ‰
 @    â $z ~  Ž â !r       â!r       â/ !r       ÈO ‰ @    " $vÿ r  ÿ Ž â !r
       ä� ‰	      b $x
   ÿŽ âr   ÿ„ ä ‰      ¤ ‚
   ÿà÷ â $rÿÿ   ÿ â ‰       â z j  ÿ„ Ä ‚ÿÿ   ÿäÿ ä Š j  ÿ† ä z k   ä Š k  � â !r       â !r		       â/ !r       ÈO † 	   â !r        Æ� †‰     è ‰      h  †‰     â M™        € ê x 
   ÿàÿ â !r       Æ/ r     ÿàñ Ê $rÿÿ   ÿ  â z  j  ÿ€ È z  k    Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y!      "   " ¹z  l      â Ey  @    € â ™x?    â y      !   b ¹z  F   
   â $r$ÿÿ   ÿ Ž ä $r.ÿÿ   ÿ Ž ä $rÿÿ   ÿ Ž Ä $r0ÿÿ   ÿ Ž ä $r2ÿÿ   ÿ Ž ä $r ÿÿ   ÿ Ž â r !ÿ   pRð är !ÿ   pRò â$x"!    Ž Ê/ xÿ   "  È r"   ÿ� â $!!À   Ž Æ x,ÿ     È | ,   pbô Ú G)      € ê y      '   " $rÿÿ   ÿ Ž â xøÿÿÿÿÀŽ â $rÿÿ   ÿ Ž â y	      %   b $tÿB   ÿ Ž â x+,   ÿ   â $x""   
Ž â y      &   ¢ $r$ÿÿ   ÿ Ž â z)+ q  ÿàÿ Ä x*ÿ   "  â 'z y   Ž È $x		   ÿ Ž â/ zÿ z    â $x"   ÿ Ž ä 'z s   Ž âO x	   ÿàÿ Æ $z	 p  ,Ž â zÿ t    â $z |  ,Ž ä$z |  Ž ä %x   ÿŽ È $z v  Ž â x%   ÿüŽ â $z v  Ž ä $z# }  ÿŽ ä %v X  Ž È $z
 w  ÿŽ â x&   ÿà÷ â %xB   Ž Æ r#
#   ÿàõ â $r'ÿÿ   Ž â z X  ÿàù â $z p  Ž ä $vÿ q  ÿ Ž â x   ÿà÷ â %v X  Ž Æ zÿ Y  „þ â $x   +Ž ä $r-ÿÿ   ÿ ä r%   ÿàõ â �y
úÿÿ • ¨ $x    â �yüÿÿ • è �yþÿÿ • ( �y    • h �yN   • ¨ �y/   • ¨ �y;þÿÿ • ¢ $tÿ   ÿ Ž Æ �y6    • ¢ $tAÿ$   ÿ Ž â r"#   àÙ ä �y(&üÿÿ • ¤ r:*-   ÿ¤~ ä �yI    • ¢ $x	   ÿ Ž Êx		ø  ÿÀŽ È z	  ÿàõ Ê $vÿ  ÿ Ì �y    › " xÿ     Ê %v   Ž Ì �y    › " $x
   ÿ Ž ÊO xø  ÿÀŽ È z  ÿàõ â $x   ÿ Ž È� $vÿ  ÿ â x
ÿ   
  ä xø  ÿÀŽ Æ �y    › ¢ %v

   Ž â z  ÿàõ ä xÿ     Æ �y

    › â $vÿ  ÿ â r>"#   +àµ â %v   Ž Æ r*-   ÿd~ â �y    › ¨ $x$   ÿŽ â xÿ     â$x   ÿ Ž â �y    › b %v>> Z  A Ž Æ xø  ÿÀŽ ä r??   ÿàÿ â %v   Ž â z  ÿàõ Æ �y9>   ™ ¤  $vÿ  ÿ ä �y    › ¨ �y    › ¨ �y1>   ™ ¢  r"#   )àµ Æ �y8>   ™ ¢  r*-   ÿd~ Æ �y7>   ™ ¨  �y3>   ™ ¢  %v Z  A Ž Æ �y5>   ™ ¢  $x=$   ÿŽ ä %v Z  A Ž â �y4>   ™ ¦  $xA:$   ÿŽ â �y>    ™ ¢  $x   =Ž ä $x   AŽ â �y >    • ¨  �y:   ™ ¨ �y<   ™ ¨  �y=   ™ ¨ �yG   ™ ¨  �yH   ™ ¨ �yF   ™ ¨ �yE   ™ ¨  �yD   ™ ¨ �yC   ™ ¨  �yB   ™ ¨ �yA   ™ ¨ �y@   ™ ¨ �y?   ™ è �yK    ™ è �y>   ™ è  �yJ    ™ â  xMÿ   N  Æ �yL    • b xRMÿ   ÿÀŽ Æ �yM    • b xONÿ   ÿÀŽ Ì 	sO O        0 	sS R        b $xPO€   ÿ Ž Ê xQP€   ÿÀŽ È xQQÿ   NxŽ â $tPÿÿ Ž ä $tTÿ @€ÿ Ž ä $xQQÿŽ â xNNT  /    â$xSS€   ÿ Ž â/ xÿ   /  ä xQ€€€€PìŽ Ä xQ€€€€TìŽ ä xSR€   SxŽ ä xOÿ   N  ä xPÿþþþÿàÿ ä xRÿþþþÿàÿ ä xÿ   ÿÀŽ Ä xOÿ   ÿÀŽ â $xNSÿŽ â xPQøŽ äxRR @€QøŽ â $tQÿ @€ÿ Ž â 	s         â $tUÿÿ Ž â xTR˜º  ÿ    ä xQN€€€€QìŽ È 	sP         " xON€€€€UìŽ ä x˜º  ÿ    ä xUQÿþþþÿàÿ ä xSOÿþþþÿàÿ ä xRÿÀŽ â $xPP€   ÿ Ž Ä $tVÿ @€ÿ Ž ä $tWÿ @€ÿ Ž â xQT€€€€ÞŽ äxTTÿÀŽ ä xO€€€€ÞŽ Æ $xTQ   T
Ž ä $xO   R
Ž ä $xQ€   ÿ Ž â xT€€€€´Ž ä x€€€€´Ž ä x€   QxŽ ä x€   PxŽ â $tÿÿ Ž â xT;ÿ   ÿÀŽ â $xÿŽ â x     â $xÿŽ â xOSNøŽ Ä xNU @€NøŽ â 	sT T        " xS€€€€ìŽ äxU€€€€VìŽ ä x€€€€ìŽ äx€€€€WìŽ ä xVSÿþþþÿàÿ Ä xPÿ   ;  ä xXUÿþþþÿàÿ ä xSÿþþþÿàÿ ä xUÿþþþÿàÿ ä xRÿ   6  ä xPPÿ   ÿÀŽ Ä xSøŽ äxU @€øŽ ä xS;T  6    ä xUO˜º  ÿ    â 	sQ P        b xVøŽ ä xRRÿ   ÿÀŽ Ä xVN˜º  ÿ    â $xTT€   ÿ Ž â xOÿ   S  ä xNUÿÀŽ â 	sS R        " xU€€€€UÞŽ ä xWVÿÀŽ äxVV€€€€	ÞŽ â $xUU   N
Ž â xTT€   ÿÀŽ Ä xNOÿ   ÿÀŽ â $xVV   W
Ž â xTTÿ   ;xŽ ä 	s; N        " xOU€€€€´Ž ä xV€€€€	´Ž â $x	Q€   ÿ Ž ä/ $xSS€   ÿ Ž Ä $xQTÿŽ â x	P€   	xŽ â $tTÿÿ Ž ä $tVÿ @€ÿ Ž â xRR€   SxŽ â $x		ÿŽ â xX @€øŽ â $tXÿ @€ÿ Ž â xPQ€€€€TìŽ äxSQ€€€€VìŽ â $tVÿÿ Ž â xPPÿþþþÿàÿ Ä xU;   ÿ   äxTSÿþþþÿàÿ ä x;	€€€€VìŽ äxS	€€€€XìŽ ä xPPQøŽ ä xUN€   UxŽ â $xNRÿŽ â xQT @€QøŽ Ä xR;ÿþþþÿàÿ ä xTSÿþþþÿàÿ â $tSÿ @€ÿ Ž ä $t;ÿÿ Ž â xRR	øŽ äx	T @€	øŽ ä xTN€€€€SìŽ â$xSUÿŽ â x;N€€€€;ìŽ Ä xUTÿþþþÿàÿ ä x;;ÿþþþÿàÿ ä xTS€€€€VìŽ ä x;;NøŽ äxNU @€NøŽ ä xTTÿþþþÿàÿ Ä xUP˜º  ÿ    ä xVS€€€€XìŽ ä xPQ˜º  ÿ    ä xQTSøŽ ä xTUÿÀŽ ä xU€€€€UÞŽ ÄO xXVÿþþþÿàÿ ä xWPÿÀŽ â$xUU   T
Ž â xVP€€€€ÞŽ ä xPX @€SøŽ ä xSR˜º  ÿ    ä xT	˜º  ÿ    Ä xRSÿÀŽ ä x	U€€€€´Ž ä xS
€€€€SÞŽ ä� xUTÿÀŽ äxTT€€€€ÞŽ ä xQQ˜º  ÿ    â $xVV   W
Ž Ä $xSS   R
Ž â xRQÿÀŽ â $xTT   U
Ž â xQ€€€€QÞŽ ä xPP˜º  ÿ    ä xV€€€€´Ž ä x;;˜º  ÿ    ä xS€€€€
´Ž â $xQQ   R
Ž â x
T€€€€´Ž Ä xNN˜º  ÿ    ä xPÿÀŽ äxR;ÿÀŽ ä xPP€€€€ÞŽ ä x;€€€€;ÞŽ äxTNÿÀŽ Ä xSN€€€€ÞŽ â $xNP   
Ž ä $x;   R
Ž ä $xTS   T
Ž â x˜º  ÿ    ä x€€€€´Ž ä xT€€€€´Ž Ä xPÿÀŽ ä x€€€€ÞŽ ä x˜º  ÿ    ä xNN€€€€´Ž â &r	9   ÿ   âx˜º  ÿ    â &r99   ÿ   â xÿÀŽ Ä x€€€€ÞŽ â &r1      âx;ÿÀŽ â &r91   9   â x€€€€ÞŽ â $x   
Ž ä &r8      âr;  €ÿàÿ â &r9O8   9   â x˜º  ÿ    â &r
7      âxQ€€€€´Ž â &r97   9   â x€€€€´Ž Ä x€€€€´Ž â $x   P
Ž â x8ÿÀŽ äx€€€€ÞŽ â &r3      ä&r93   9   â x€€€€´Ž â $x8   8
Ž ä &rN5      Ä &r95   9   â x8€€€€´Ž â &r4      ä&r94   9   ä &r      ä&r4   9   ä &r	:   ÿ   Ä &r::   ÿ   ä &r		<   ÿ   ä&r:=   :   ä&r=      ä&r<   ÿ   ä &rG   	   Ä &r	OH   :   ä&rH      ä&rG      ä &r	F   	   ä&rE      ä &r
F      Ä &rOE      ä &r	D   	   ä&r
C      ä&rD      ä &rC      ä &r	B   	   Ä &rA      ä&rNB      â xÿ   /  â &rA      â x6ÿ   6  â &r	@   	   ä&rN?      ä&r@      â x   ÿüŽ â &r?      â x6   ÿüŽ â &rK   	   Ä &r>      ä&r
K      ä&r>      ä $r   ÿŽ ä &rJ      ä $r

   ÿŽ Ä &rJ      â xÿ     â $r   ÿŽ â x	ÿ   
  â $r   ÿŽ â x   ÿàõ â $r44   ÿŽ ä $r   ÿŽ â r   ÿ� ä x,,   ÿàÿ Ä r		
   ÿ� â 0r(ÿ(    A   â xÿ     â $rÿÿ    â xÿ   4  â 0rÿL    A   â xÿ     ä x
ÿ     â 0rÿI    A   â xÿ     Ä | ,   pbô ä r   ÿ� ä r4   ÿ� ä r   ÿ� ä r

   ÿ� â  r(     A  âEr        â 0rÿ     A   â  r     A  â 0rMÿM    A   â xÿ     Ä xÿ     ä x	ÿ   	  ä xÿ     ä x
ÿ   
  â #r      â Er        â  r(     A  âEr        â  r(M     A  â Er	 	       â  r     A  â Er        â  rMM     A  â Er
 
       Ä x   ÿà÷ ä x&&   ÿàù â #r        â #r2   2   â #r0	   0   â #r.   .   â #r$M
   $   â x€   ÿàÿ â $rÿÿ   Ž â x))€   ÿàÿ â $r'ÿÿ   ' â x++€   ÿàÿ â D)       À â Gy  €çÿÿÿÿƒ ê Ay        € ê ˆ !  ýÿ H   è ˆ !2€ýÿ H   è ˆ !0 þÿ H   è ˆ !€þÿ H   è ˆ !. ÿÿ H   è ˆ !$€ÿÿ H   è {          ì M        € ê „y     H   ( „y €   H   h „y    H   ¨ „y €  H   è „y    H   ( „y	 €  H   h „y    H   h „y €  H   h „y    H   h „y    H   h „y €  H   b !r        Æ „y €  H   " !r2       Æ/ „y
    H   b !r       ÆO „y    H   ¢ !r       Æ� „y €  H   â !r       Æ„y €  H   â !r	       Æ„y    H   â !r0       Æ „y €  H   â !r       Æ ‰      " !r.       Æ ‰      b !r       â !r       ä y      %   b !r$       Æ y      '   " !r

       â/ !r       âO !r       È� ‰      b !r       â!r       ä ‰	      ¢ !r       â!r

       ä ‰       â !r       Æ ‰
      ( ‰      h ‰      " !r       â/ !r		       ÈO ‰      b !r         Æ� ‰	      ¢ !r
       Æ‰  €    â !r       Æ‰      " !r       Æ ‰
      ( ‰ €    b !r       Æ/ y      &   b !r	       ÆO ‰ €    ¢ !r        Æ� ‰	 €    â !r       Æ‰  @    " !r

       Æ ‰ €    " !r       Æ‰
 €    h ‰ @    b !r       âO !r		       È� ‰ @    ¢ !r        â!r       ä ‰	 @    " !r
       â$x   ÿ Ž ä ‰
 @    â !r       â/ $z  x  Ž Ä ‰ @    b z  r  p`ð äx   ÿàÿ â ‰      " r ÿ   pVp  â$z  ~   Ž â z  r  p`ò â ‰      b !r       äO x    p"ò Æ ‰      ¢ !r	       â $v	ÿ r  ÿ Ž ä Š  r  ÿàõ â!r
       â� ‰      " $„ÿ   ÿ Ž â !r       â/ $‚ÿÿ   ÿ â ‰      b Š j  ÿ„ â !r       â	   ÿàÿ â ‰      â Š k   â !r       âr	    ÿ† â $ÿ   ÿ Ž â 	    ÿàõ â%†  j   Ž â !r       ÆO $ÿÿ   ÿ â 	 j  ÿ„ â $rÿÿ   ÿŽ â z j  ÿ† â %
  j   Ž â 		 k   â !r       â z k  � â †‰     â !r       Æ/ † 
   â !r       Æ� †‰     è †     è †‰     â M™        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      "   " ¹z  l      â Ey  p    € â ™x?    â y       !   b ¹z  F   
   â $r$ÿÿ   ÿ Ž ä y       %   ¢ $rÿÿ   ÿ Ž ä $rÿÿ   ÿ Ž Ä $r*ÿÿ   ÿ Ž â r ÿ   pRð är ÿ   pRò â$x"     Ž ä/ $x     ÿ Ž ÆO xÿ   "  È r"   ÿ� â $€    Ž Æ x-ÿ     È | -   pbô Ú G)  @    € ê y	      '   " $rÿÿ   ÿ Ž â xøÿÿÿÿÀŽ â $r
ÿÿ   ÿ Ž â y      &   b $z  p  -Ž âx    ÿàÿ â $x""   
Ž ä $t'ÿB   ÿ Ž ä $x"   ÿ Ž â x(ÿ   "  â $x)-   ÿ Ž Ä $r$ÿÿ   ÿ Ž Æ z) q  ÿàÿ â 'z	 y   Ž È 'z s  
 Ž â/ zÿ z    Æ $z!	 }  ÿŽ â zÿ t    â $z |  Ž ä$z |  -Ž ä %x   ÿŽ È $z v  Ž âx#   ÿüŽ â $z v  Ž ä $z& p  Ž ä %v X  'Ž È %xB   Ž â x   ÿàù Æ $z
 w  ÿŽ â z X  ÿà÷ â $rÿÿ    ä %v&& X  'Ž â r!
!   ÿàõ ä x   ÿàù â $r,ÿÿ   & Ž ä $r+ÿÿ   ÿ â zÿ Y  d~ È r,#   ÿàõ â �yúÿÿ • ¨ $x   ' â �y	üÿÿ • è �y
þÿÿ • ( �y    • h �yC   • ¨ �y2   • ¨ �y>þÿÿ • ¢ $tÿ   ÿ Ž Æ �y%    • ¢ $t/ÿ$   ÿ Ž Æ �y?üÿÿ • ¢ $x
   ÿ Ž Êxø  ÿÀŽ È z  ÿàõ Ê $vÿ  ÿ Ì �y    › " xÿ   
  Ê %v   Ž Ì �y    › " $x   ÿ Ž ÊO xø  ÿÀŽ È z  ÿàõ Ê $vÿ  ÿ â xÿ     Ê �y    › ¢ %v   Ž È $x	   ÿ Ž ä� �y    › æ xø  ÿÀŽ ä x
ÿ     äz  ÿàõ â $x   ÿ Ž Ä %v

   Ž È $vÿ  ÿ â r"!   )àµ â �y

    › b xÿ   	  ä xø  ÿÀŽ â �y    › ¢ r(+   ÿd~ ä r"!   àÇ â %v   Ž â z  ÿàõ Ä r&(+   ÿ„þ â %v Z  / Ž ä�y    › ¤ $x$   ÿŽ ä $v	ÿ  ÿ ä $x   Ž ä %v Z  / Ž Ä �y    › ¤ $x/&$   ÿŽ ä �y<   ™ ¤  $x   /Ž ä �y0   ™ ¨  �y&   ™ ¨ �y9   ™ ¨ �y=   ™ ¨  �y7   ™ ¨ �y;   ™ ¨  �y5   ™ ¨ �y:   ™ ¨  �y/   ™ ¨ �y8   ™ ¨  �y3   ™ ¨ �y6   ™ ¨  �y.   ™ ¨ �y4    ™ è  �y1    ™ è �y@    • è  �yB    • â $rÿÿ   , Ž Ä $rÿÿ   ' Ž Ê �yA    • b xDCÿ   ÿÀŽ ä xEÿ   C  È 	sD D        " xGEÿ   ÿÀŽ Î 	sF G        b $xED€   ÿ Ž â xÿ   2  â $xHF€   ÿ Ž Æ/ xFE€   ÿÀŽ È xFFÿ   CxŽ ä xCCT  2    ä xÿ   ÿÀŽ â $tJÿÿ Ž â xÿ   C  â $xCFÿŽ â xHG€   HxŽ â $tFÿ @€ÿ Ž Æ xC€€€€JìŽ â$tEÿÿ Ž â xFC€€€€FìŽ â $xHHÿŽ â 	s         " xÿ   ÿÀŽ â $tGÿ @€ÿ Ž â xDÿþþþÿàÿ ä xFFÿþþþÿàÿ ä xEH€€€€EìŽ â 	s         b xDDCøŽ Ä xFF @€CøŽ ä xGH€€€€GìŽ ä xEEÿþþþÿàÿ ä xCD˜º  ÿ    ä xJF˜º  ÿ    ä xGGÿþþþÿàÿ Ä xDEHøŽ äxFCÿÀŽ ä xHG @€HøŽ â $xG€   ÿ Ž ä $x€   ÿ Ž Ê/ x€   xŽ â $tÿ @€ÿ Ž È $xÿŽ â xKD˜º  ÿ    ä xE€€€€CÞŽ äxCJ€€€€ÞŽ äxJJÿÀŽ â $xEE   F
Ž È $xJC   J
Ž â xC€   GxŽ ä xE€€€€´Ž â $tÿ @€ÿ Ž â xF>ÿ   ÿÀŽ â $xCCÿŽ ä $tEÿÿ Ž ä $tÿÿ Ž â xC€€€€ìŽ â 	sF F        " xJ€€€€´Ž Ä xJÿþþþÿàÿ ä xE€€€€EìŽ äx€€€€ìŽ ä xC€€€€ìŽ ä xEÿþþþÿàÿ ä xIÿþþþÿàÿ Ä xGÿ   >  ä xÿþþþÿàÿ ä xEÿ   %  ä xøŽ äxI @€øŽ ä xCøŽ Ä xGGÿ   ÿÀŽ ä xI>T  %    ä xCJ @€CøŽ â $xJF€   ÿ Ž â xEEÿ   ÿÀŽ â 	s G        " xDÿ   I  ä xIJ€   ÿÀŽ Ä xDDÿ   ÿÀŽ Æ 	sF E        b xMKÿÀŽ ä xJ€€€€KÞŽ ä xIIÿ   >xŽ ä xHH˜º  ÿ    â 	s> D        " $xMJ   M
Ž ä xJHÿÀŽ Ä xLH€€€€ÞŽ â $xHIÿŽ ä $x€   ÿ Ž ä $tIÿÿ Ž ä $xJL   J
Ž ä $xLF€   ÿ Ž â/ xG€   xŽ Ä xIH€€€€IìŽ ä xLE€   LxŽ â $xE>€   ÿ Ž âxJJ€€€€´Ž â $xÿŽ â xIIÿþþþÿàÿ â $t>ÿ @€ÿ Ž â xED€   ExŽ â $tKÿ @€ÿ Ž â xFIHøŽ â $xDLÿŽ â x>€€€€>ìŽ â $tGÿÿ Ž Ä $tIÿ @€ÿ Ž â xKH€€€€KìŽ ä xL>ÿþþþÿàÿ ä xGD€€€€GìŽ äx>D€€€€IìŽ ä xKKÿþþþÿàÿ â $tÿÿ Ž â xGGÿþþþÿàÿ Ä xI>ÿþþþÿàÿ ä xHK @€HøŽ ä x>GDøŽ ä x€€€€ìŽ ä xDI @€DøŽ ä xIF˜º  ÿ    Ä xHH˜º  ÿ    ä xÿþþþÿàÿ ä xFIÿÀŽ ä xGHÿÀŽ ä xI€€€€IÞŽ äO xHH€€€€ÞŽ Ä xøŽ äxL @€øŽ â $xII   F
Ž ä $xHH   G
Ž â xG˜º  ÿ    ä x˜º  ÿ    ä xI€€€€´Ž Ä xFGÿÀŽ ä xÿÀŽ ä xG€€€€GÞŽ ä� xI€€€€ÞŽ â $xEEÿŽ ä $xGG   F
Ž ä $xI   
Ž Ä $tKÿÿ Ž ä $tLÿ @€ÿ Ž â xG€€€€´Ž ä x€€€€´Ž ä x>˜º  ÿ    ä xG˜º  ÿ    ä xKE€€€€KìŽ Ä x>C˜º  ÿ    ä xLE€€€€LìŽ ä xGÿÀŽ ä xKKÿþþþÿàÿ ä xC>ÿÀŽ ä xG
€€€€GÞŽ ÄxLLÿþþþÿàÿ ä x>>€€€€ÞŽ â $xGG   
Ž â xKKEøŽ äxEL @€EøŽ â $x>>   C
Ž â xC˜º  ÿ    ä xE˜º  ÿ    Ä xG€€€€
´Ž ä x
CÿÀŽ ä xC€€€€CÞŽ ä xKK˜º  ÿ    ä xEÿÀŽ äx€€€€ÞŽ â &r<   ÿ   â xH€€€€´Ž â $xCC   

Ž â xFKÿÀŽ â &r
&   ÿ   â xK€€€€KÞŽ â $x   E
Ž â xDD˜º  ÿ    â &r0      Ä &r
9   
   â xEDÿÀŽ â $xKK   F
Ž â x€€€€´Ž â &r=      â xDD€€€€ÞŽ â &r<<   ÿ   â xÿÀŽ â &r
7   
   â x€€€€ÞŽ â &r&&   ÿ   â x˜º  ÿ    â $xDD   E
Ž â xK€€€€´Ž â &r;      â xM€€€€´Ž â &r<0   <   â xEÿÀŽ â &r
5   
   â x€€€€	ÞŽ â $x   
Ž Ä &r&9   &   ä &r:      ä &r<=   <   ä &r
/   
   â x€€€€´Ž â &r&7   &   â xC€€€€´Ž â $x   E
Ž Ä &r8      ä&r;J;   <   ä &r
3   
   â xD€€€€´Ž â &r&J5   &   â x	€€€€	´Ž â &r6      â x%ÿ   %  â &r;:   ;   Ä &r
.   
   ä &r&/   &   â x%   ÿüŽ â &r4      ä &r;	8   ;   ä &r1   
   ä &r&	3   &   â x>€€€€´Ž â &r;6   ;   â x2ÿ   2  â $r   ÿŽ Ä &r&.   &   â x22   ÿüŽ â $r   ÿŽ â xÿ     â &r;4   ;   ä&r1   &   â xÿ     â $r;;2   ÿŽ â r   ÿ� â $r2   ÿŽ â x   ÿàõ Ä r   ÿ� ä x--   ÿàÿ ä x	ÿ     â $rÿÿ    â xÿ   ;  ä x
ÿ     â 0r?ÿ?    A   â xÿ     â 0r@ÿ@    A   â | -   pbô Ä r;   ÿ� ä r

   ÿ� â 0rAÿA    A   â Er        â  r?@     A  â 0rBÿB    A   â x
ÿ   
  ä xÿ     â #r*   *   â Er	 	       â  r@A     A  â Er
 
       â  r??B     A  â Er        â  rAAB     A  â x,,   ÿà÷ Ä x   ÿàù â #r?	      â #r$A
   $   â #r      â x€   ÿàÿ â $r'ÿÿ   'Ž â x))€   ÿàÿ â $rÿÿ    â D)       À â Gy   êÿÿÿÿƒ ê Ay        € ê ˆ * þÿ H   è ˆ €þÿ H   è ˆ  ÿÿ H   è ˆ $€ÿÿ H   è {          ì M        € ê „y      H   " z   r  p`ð Æ „y  €   H   b r  ÿ   p"p Æ „y     H   ¨ „y     H   è „y  €  H   ( „y     H   h „y     H   h „y	  €  H   h „y     H   h „y  €  H   h „y
  €  H   b !r*       Æ „y  €  H   " !r       â/ !r       âO !r       ä� !r       â!r       â!r       ä !r		       Æ ‰      b !r       Æ ‰	      ¢ !r$       Æ ‰      â !r

       Æ y      &   " !r
       â !r       È/ ‰      " !r	       ÆO ‰      b !r       Æ� ‰      ¨ ‰      â !r       â !r       È/ ‰      " !r       ÆO ‰ €    b !r       Æ� ‰ €    ¨ ‰
 €    â !r       â !r       È/ ‰ €    " x    ÿàÿ â $z  x   Ž â!r       äO ‰ @    b z  r  p`ò â !r

       ä� ‰ @    ¢ x     p"ò Æ ‰	
 @    è y      '   " !r       â !r       â/ !r       ÈO ‰      " !r	
	       Æ� ‰      b $z  ~   Ž Æ‰
 @    ¢ $tÿ   ÿ Ž â 
   r  ÿàõ ä ‰	      æ $ÿÿ   ÿ â 
  j  ÿ„ È 
  k   â !r       â %v  j   Ž â !r       È/ †	     â !r

       ÆO †    â !r		       Æ� ‰
      h  †	 	    â M™        € ê ¹z  r      â !r
       â/ �x   ?àÿ Ì |     ÿàñ Ê $rÿÿ   ÿ  â z  j  ÿ€ È z  k    Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y       &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž ä $r	ÿÿ     Ž È %	 \   Ž Ê �	    ™ b ¹z  l      â Ey      € â ™x?    â y      "   b $rÿÿ   ÿ Ž Æ y      !   ¢ $ÿÿ   ÿ Ž ä $ÿÿ   	 Ž È '
	 m   Ž È $‚ÿÿ   ÿ Ž â 
ÿ n    Ê $ÿÿ   
Ž â r ÿ   pRò Æ/ $
	 o  	Ž â r ÿ   pRô â$x    Ž äO $‚ÿÿ   	 Ž Æ xÿ     â 'Š	 s   Ž Æ r   ÿ� È xÿ     È |    pbö ä Šÿ t    Ö G9       € ê y      '   " $rÿÿ   ÿ Ž â xøÿÿÿÿÀŽ â $z		 w  ÿŽ â y      %   b $rÿÿ   ÿ Ž ä $x   
Ž ä $x   ÿ Ž Æ xÿ     â 'z y   Ž È $z }  ÿŽ â zÿ z    È r	   ÿàû â $z |  Ž È $z v  Ž ä$x   ÿ Ž ä $z p  Ž ä/ $rÿÿ   ÿŽ ä %xB   ÿŽ Ê z X  ÿàù â %x   Ž Æ z Y  ÿä ä z X  ÿàñ È x   ÿà÷ È zÿ Y  þ Ê �yþÿÿ • ¨ �y    • è �y$   • ( �y   • b $tÿ$   ÿ Ž Ä $tÿ   ÿ Ž â �y"    • â $x   ÿ Ž ÊO xø  ÿÀŽ È z  ÿàù ä r   à± Æ $vÿ  ÿ â r	   ÿd~  Ê �y    › ¢ %v Z   Ž È $x		$   ÿŽ È $x   	Ž Ê �y   ™ ¢
 xÿ     Æ �y   ™ ¢
 $x
   ÿ Ž ä� %v   Ž â �y   ™ ä
 x

ø  ÿÀŽ ä �y   ™ ä
 z

  ÿàñ Ä �y    › â xÿ     ä $vÿ  ÿ  â �y!   ™ ä
 %v   Ž ä �y    ™ è
 �y

    › è �y    › è �y   ™ è
 �y    ™ è
 �y#    • â
 x&$ÿ   ÿÀŽ Ì	s& &        " x(ÿ   $  â $x%&€   ÿ Ž Ê x%%€   ÿÀŽ È x'%ÿ   $xŽ ä x%(ÿ   ÿÀŽ â $t(ÿÿ Ž ä $x''ÿŽ â xÿ     â$t*ÿ @€ÿ Ž Æ x&'€€€€(ìŽ â	s %        â x$$T      ä x(&ÿþþþÿàÿ ä x)'€€€€*ìŽ ä xÿ   ÿÀŽ ä x(('øŽ Ä x*ÿ   $  â 	s&         â x$)ÿþþþÿàÿ ä x)(˜º  ÿ    ä x'$ @€'øŽ ä x$*ÿ   ÿÀŽ ä x*'˜º  ÿ    Ä 	s' $        $ $x''€   ÿ Ž Ê x$$€   'xŽ ä xÿ     È x   ÿüŽ ä x   ÿàÿ È |    pbð â 0r"ÿ"    A   â x   ÿàù ä x   ÿà÷ ä x€   ÿàÿ â $rÿÿ    ä $rÿÿ   Ž â x(€€€€)ÞŽ ÄO x))ÿÀŽ Ê $x)(   )
Ž ä $x(€   ÿ Ž ä $x&€   ÿ Ž â x+*€€€€ÞŽ äx%%€   (xŽ ä x**ÿÀŽ ä x€   xŽ â $x%%ÿŽ Ä $tÿÿ Ž ä $x*+   *
Ž ä $t&ÿ @€ÿ Ž â x))€€€€´Ž â $xÿŽ â x%€€€€ìŽ â$t(ÿÿ Ž â x*€€€€´Ž ä x%€€€€&ìŽ â &r)   ÿ   â xÿþþþÿàÿ Ä x'€€€€(ìŽ ä x&ÿþþþÿàÿ â &r      â x%øŽ â$x$ÿŽ â x'ÿþþþÿàÿ â $t'ÿÿ Ž â x˜º  ÿ    â $t)ÿ @€ÿ Ž â x%& @€%øŽ Ä x€€€€'ìŽ ä x$%˜º  ÿ    ä x'ÿþþþÿàÿ ä xÿÀŽ ä x&€€€€)ìŽ ä x€€€€ÞŽ â� $t*ÿ @€ÿ Ž â x%$ÿÀŽ Ä x)&ÿþþþÿàÿ â $x   
Ž â x''øŽ ä x$$€€€€	ÞŽ ä x(€€€€*ìŽ ä x)) @€øŽ ä x''˜º  ÿ    â $x$$   %
Ž â x€€€€´Ž Ä x((ÿþþþÿàÿ ä xøŽ ä x)˜º  ÿ    ä x'ÿÀŽ ä x'
€€€€'ÞŽ â &r      â x$$€€€€	´Ž Ä x( @€øŽ ä xÿÀŽ ä x	˜º  ÿ    â $x''   
Ž â x€€€€ÞŽ ä x˜º  ÿ    ä x	ÿÀŽ â $x   
Ž â x	€€€€	ÞŽ â &r$      â x''€€€€
´Ž Ä xÿÀŽ â$x		   
Ž â x€€€€ÞŽ â &r!'!      â x€€€€´Ž ä x		€€€€´Ž â $x   
Ž ä &r    !   Æ x€€€€´Ž â &r		      È &r   	   È $r   ÿŽ Ê xÿ     È r   ÿ� â 0r#ÿ#    A   Æ xÿ     ä  r#"#     A  ä Er        Ê #r#      â G‰  ðõÿÿÿÿƒ ê Ay        € ê ˆ €ÿÿ H   è {          ì M)        € ê „y     H   " r ÿ   pRð Æ „y €   H   h „y    H   ¢!r       È !r       È/ !r       ÊO ‰      $ !r       Ê ‰      $ !r       Ê ‰ €    $ !r       Ê ‰ @    $ !r       Ê ‰      b  M	        € ê y      %   ¨ y      '   â $z   x  Ž äO $tÿ   ÿ Ž ä $z  ~   Ž â� !r       Æ/ %v  j   Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž È % \   Ž Ê �	    ™ b $rÿÿ   ÿ Ž â Ey       € â $rÿÿ   ÿ Ž â y       "   b 'Š s   Ž Æ y      !   b '
 m   Ž È $rÿÿ   ÿ Ž â 
ÿ n    ä rÿ    ÿü‚ â/ $x     Ž Ê x
ÿ     È r

   ÿ� ä Šÿ t    â G       € ê z ÿ `  pPô â $rÿÿ   ÿ Ž â z ÿ ^  pPò â $rÿÿ   ÿ Ž â z ÿ b  p ö ä z ÿ a   Sô ä z ÿ _  Sò Ä z ÿ c  0't Ö       '   "        ä¢       â ©	      '   d $ x  ÿŽ ä       %   ¢ $ª x  ÿŽ Æ ©      %   â $ ~  ÿŽ ä $ª	 ~  ÿŽ Æ/    àë äOÿ     ä ¢   àÇ ä�¨ÿ     ä ÿÿ   Äþ ä    ÿàý Ä ¢   ÿàû ä ¢ÿÿ   „þ â $ÿÿ    â  ^  ÿ† äª b  ÿˆ â $¢ÿÿ   Ž â  _  � È ª c   â �     h �©     d Ay        € ê ¹z  l      â xÿ   
  â ™x?    â $ÿÿ   
Ž â Ey  à    € â $rÿÿ   ÿ Ž Ä $
 o  Ž â|    pbò â $‚ÿÿ    Ž ä $rÿÿ   ÿ Ž Ô G  p    € ê y      '   " $rÿÿ   ÿ Ž ä $tÿB   ÿ Ž â y      %   b $rÿÿ   ÿ Ž ä $rÿÿ   ÿ Ž ä $x   ÿ Ž ä 'z y   Ž È $z }  ÿŽ â zÿ z    Ê $z |  Ž ä $z w  ÿŽ ä $z v  Ž Ær   ÿàñ â $z p  Ž â/ x
øÿÿÿÿÀŽ Æ %v `  Ž È %v
 X  Ž È $x   
Ž ä $rÿÿ   ÿ  ä $x#   ÿ Ž ä $rÿÿ   
 Ž ä $rÿÿ    Ž È %x#   Ž â xÿ     â �y$    • ¨ �y
   • è  �y   • ( �y'   • h �y&   • ¢  $t"ÿ   ÿ Ž Ä $tÿ$   ÿ Ž ä $x
   ÿ Ž Ê� xø  ÿÀŽ È z  ÿàñ Ê $vÿ  ÿ  Ì �y    › â $x   ÿ Ž âr   à‘ È xø  ÿÀŽ ä r   ÿ$~  ä z  ÿàõ ä xÿ   
  Æ $vÿ  ÿ ä %v
  " Ž È �y    › " %v Z   Ž Æ �y

    › " $x$   ÿŽ È $x		   Ž â xÿ     È �y   ™ " %v  " Ž Æ �y!   ™ ( �y   ™ ( �y    › ( �y   ™ ( �y    ™ ( �y   ™ ( �y   ™ ( �y    ™ ( �y%    • " x('ÿ   ÿÀŽ Äx)ÿ   '  È 	s( (        " x+)ÿ   ÿÀŽ Î 	s, +        b x'T  &    âO $x)(€   ÿ Ž Æ x-ÿ     ä x*)€   ÿÀŽ â $x,,€   ÿ Ž Æ/ x**ÿ   'xŽ â $t/ÿÿ Ž â x	-ÿ   ÿÀŽ ä x)+€   ,xŽ â $x**ÿŽ â 	s' 	        " $t+ÿ @€ÿ Ž â xÿ   &  ä x(*€€€€/ìŽ Ä x,*€€€€+ìŽ ä xÿ   ÿÀŽ â $t.ÿÿ Ž â x+(ÿþþþÿàÿ â $x))ÿŽ â x-,ÿþþþÿàÿ â $t0ÿ @€ÿ Ž â 	s(         b x++*øŽ ä x,)€€€€.ìŽ Ä x-- @€*øŽ ä x*)€€€€0ìŽ ä x++˜º  ÿ    â $x0'€   ÿ Ž â x,,ÿþþþÿàÿ ä x.*ÿþþþÿàÿ ä x',)øŽ Ä x,-˜º  ÿ    ä x0	€   0xŽ â $x-(€   ÿ Ž È/ $x(0ÿŽ â x-€   -xŽ ä x). @€)øŽ ä x''˜º  ÿ    ä x&ÿ   &  ä z ÿ `  pPð È z ÿ a   Sð â 0r$ÿ$    A   â x   ÿàÿ È |    pbò ä x*€€€€+ÞŽ ä� x++ÿÀŽ Ê $x	*   +
Ž â x*,€€€€ÞŽ äx+,ÿÀŽ Ê $x**   +
Ž â x+(€€€€/ìŽ â $t/ÿ @€ÿ Ž â x		€€€€´Ž ä x++ÿþþþÿàÿ â $x-ÿŽ â x*€€€€´Ž â $t-ÿÿ Ž â x(€€€€/ìŽ ä x++(øŽ Ä xÿþþþÿàÿ ä x,€€€€-ìŽ äx++˜º  ÿ    ä x-€€€€/ìŽ ä x* @€(øŽ ä x,ÿþþþÿàÿ Ä x(+ÿÀŽ ä x--ÿþþþÿàÿ ä x+€€€€+ÞŽ äx**˜º  ÿ    ä xøŽ äx- @€øŽ â $x-+   (
Ž â x,*ÿÀŽ Ä x+*€€€€ÞŽ ä x*)˜º  ÿ    ä x('ÿÀŽ ä x'
€€€€'ÞŽ â $x,+   ,
Ž â x+*ÿÀŽ äx**€€€€ÞŽ â $x)'   (
Ž Ä &r'	   ÿ   ä $x**   +
Ž â x))€€€€
´Ž â &r!   '   â x˜º  ÿ    ä x	-€€€€´Ž â &r))      â x**€€€€´Ž ä x˜º  ÿ    Ä xÿÀŽ ä x€€€€ÞŽ â &r)*   )   â x,,€€€€´Ž ä xÿÀŽ â$x   
Ž â x€€€€ÞŽ â &r		    )   Æ x€€€€´Ž â $x   
Ž ä &r,,   	   Æ x€€€€´Ž â &r,   ,   â x&   ÿüŽ Æ &r   ,   È $r   ÿŽ Ê xÿ     È r   ÿ� â $rÿÿ    Ž â x
   ÿà÷ â 0r%ÿ%    A   â xÿ     â $rÿÿ    Ž â x   ÿàõ â $rÿÿ   Ž â  r$$%     A  â Er        ä $rÿÿ    â G‰  �    € ê %x#   Ž ä �y     ¨  �y    è �y#    ( �y&    h �y    ¢ $x   ÿ Ž Ê� xø  ÿÀŽ È z  ÿàñ Ê $vÿ  ÿ  Ì �y    › â xÿ     Ê %v  " Ž Ì �y    › â $x'#   ÿ Ž âx#ÿ   #  È x''ø  ÿÀŽ È z'  ÿàñ Ê/ $v	ÿ  ÿ  ä %v"#  " Ž È �y    › ( �y""    › " x(&ÿ   ÿÀŽ Ì	s( (        b x'ÿ   &  È x*'ÿ   ÿÀŽ È 	s) *        " $x'(€   ÿ Ž Ê/ x''€   ÿÀŽ È x''ÿ   &xŽ â $t,ÿÿ Ž â x&&T      âO $t.ÿ @€ÿ Ž ä $x''ÿŽ ä $x)€   ÿ Ž â x(ÿ   &  ä x&'€€€€,ìŽ Ä x)'€€€€.ìŽ ä x+*€   xŽ ä x(ÿ   ÿÀŽ ä x(&ÿþþþÿàÿ ä x*)ÿþþþÿàÿ ä x)ÿ     Ä x(('øŽ äx** @€'øŽ ä x')ÿ   ÿÀŽ â $x)+ÿŽ â x+(˜º  ÿ    â 	s&         " $t.ÿÿ Ž â x,*˜º  ÿ    â $t0ÿ @€ÿ Ž Æ x.)€€€€.ìŽ äx/)€€€€0ìŽ ä x..ÿþþþÿàÿ â 	s( '        f x..)øŽ â $x&&€   ÿ Ž Ê x&€   &xŽ â $x((€   ÿ Ž È/ $x&&ÿŽ â x('€   (xŽ â $t'ÿ @€ÿ Ž È $x((ÿŽ Ê x'(€€€€'ìŽ È x''ÿþþþÿàÿ È x'' @€(øŽ ä xÿ     È x   ÿüŽ â 0rÿ    A   Ê  r%     A  â x*€€€€+ÞŽ ä� x++ÿÀŽ ä x-,€€€€ÞŽ äx,,ÿÀŽ â $x+*   +
Ž â x*/ÿþþþÿàÿ Æ $x,-   ,
Ž â x** @€)øŽ ä x++€€€€´Ž ä x,,€€€€´Ž ä x.˜º  ÿ    ä x**˜º  ÿ    â &r+   ÿ   â xÿÀŽ Ä x€€€€ÞŽ ä x*ÿÀŽ äx**€€€€ÞŽ â $x   
Ž È $x*   
Ž ä $tÿÿ Ž â x€€€€´Ž â $tÿ @€ÿ Ž Æ x&€€€€ìŽ â &r,!      â x&€€€€ìŽ â $t!ÿÿ Ž â xÿþþþÿàÿ ä xÿþþþÿàÿ ä x!(€€€€!ìŽ ä x&øŽ Ä x!!ÿþþþÿàÿ ä x˜º  ÿ    ä x @€&øŽ â &r      â x!!(øŽ ä xÿÀŽ ä x˜º  ÿ    Ä x€€€€ÞŽ äx€€€€´Ž ä x!!˜º  ÿ    ä xÿÀŽ â$x   
Ž â x€€€€	ÞŽ â &r      â x!ÿÀŽ Ä x'˜º  ÿ    â $x   
Ž â x!"€€€€!ÞŽ ä x€€€€´Ž ä xÿÀŽ â$x!!   
Ž â x€€€€#ÞŽ â &r       â x€€€€	´Ž Ä x!!€€€€"´Ž â $x   
Ž ä &r      Æ x€€€€#´Ž â &r!      È &r      È $r   ÿŽ Ê xÿ     È r   ÿ� È xÿ     È Er        Ê #r      â #r$      â x€   ÿàÿ â D       À â Gy  Ðìÿÿÿÿƒ ê Ay        € ê r  ÿ   pRð âEy  €     € â r  ÿ   pRò Ö G‰  P     € ê z ÿ `  pPð â $x      Ž Æ z ÿ a   Sð ä ˆs  €ÿÿ H   ö ˆ     H   ä Ay        € ê {          ì M        € ê z ÿ `  pPð â „y      H   & z ÿ a   Sð Ú „‰ €   H   h „‰    H   ¢ !r        È !‚       È/ !‚       âO G‰        € ê „y  €  H   ( „y €   H   h „y    H   ¨ „y    H   è„y €  H   " !r         â !r       Æ/ !r         âO !r       Æ� !r        ä‰      " z ÿ `  pPð ä r ÿ   pRò ä z ÿ a   Sð â !r       Ê ‰       $ !r         Ê ‰  €    $ !r        Ê ‰ @    $ !r       Ê ‰      $!r       â G‰        € ê ‰       $ !r         Ê ‰       $ !r        Ê ‰ €    $ !r       Ê ‰ @    $ !r       Ê ‰      $ !r       ä M        € ê !r       â G‰  `    € ê $v ÿ h  ÿ Ž â !r       È x     pBð Ú G	  p    € ê x  ÿÿÿÿÿàÿ È x       € Ê $x     ÿ Ž È ‚{   €     $ xÿ     È Iy ðâÿÿÿÿƒ ê  x ;ª¸¿  A  Ì s          $ !t   €?    Ì s        $  r     A  È  r     A  â Gy  P    € ê  x '7=  A  â x*BL?  A  Æ #t   €?    È  r      A  â  x    ?  A  Ê s     $   $ !t  €?    È  r       A  È  r      A  â Gy  °    € ê x     pRð Ú G‰  à     € ê x     pRð Ú G	        € ê 	z  i    � ä 	z i    � Æ  x ;ª¸¿  A  â 	z i €  � Ê s        $ !t  €?    Ì s        $  r       A  È  r      A  â Gy  Ð     € ê  r     A  â Gy  °     € ê 	x   à@  � ä 	x  à@  � Æ  x #ÛÙ?  A  â 	x  àÀ  � Æ  x;ª¸¿  A  ä !t  €?    È s        $ !t  €?    Ì s        $  r       A  È  r      A  ä y       &   ( y      %   ( y      '   b $z   x  Ž ä $tÿ   ÿ Ž Ä $z  ~   Ž È/ %v  j   Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y       &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž ä $rÿÿ     Ž È % \   Ž Ê �	    ™ b ¹z  l      â Ey  @'    € â ™x?    â y-      "   b x*      ÿ  â $r,ÿÿ   ÿ Ž ä y(      !   ¢ $r6ÿÿ   ÿ Ž ä $ÿÿ   ÿ Ž â y/      %   " $ÿÿ    Ž È '
 m   Ž È $‚ÿÿ   ÿ Ž â 
ÿ n    Ê $ÿÿ   
Ž â r -ÿ   pRò Æ/ $
 o  Ž ä $x.-    (Ž äO $‚ÿÿ    Ž ä $x//   ÿ Ž â xÿ   .  â 'Š s   Ž ä y      '   " r.   ÿ� È x4ÿ     È | 4   pbô ä Šÿ t    ä r -ÿ   pRð â$--€   (Ž Ð G)  p%    € ê $rÿÿ   ÿ Ž â xøÿÿÿÿÀŽ â $z w  ÿŽ ä 'z y   Ž È $z0 }  ÿŽ â z	ÿ z    â $z/ p  4Ž ä $x..   
Ž â r00   ÿàõ â $z		 |  Ž ä $x.   ÿ Ž â x2ÿ   .  â $z v  	Ž Ä$v1ÿ X  ÿ Ž ä %xB   ÿŽ È $v3ÿ Y  ÿ Ž ä $r*ÿÿ   ÿ Ž ä $x)4   ÿ Ž ä $r5ÿÿ   ÿ ä %x   Ž Ê r1   ÿàõ Ê $x	   3 Ê �y
   • ¨ �y   • â $vÿ p  ÿ Ž Æ �yL   • " %xB   Ž Æ �y7   • h �y   • è �y   • â %xB   Ž Æ �y8   • ( �y   • ( �yK   • " %xB   Ž Æ �yA   • ( �yG   • " $tDÿ   ÿ Ž ä $t!ÿ$   ÿ Ž â �yE   • ( �yB   • ( �yC   • " $tRÿ @€ÿ Ž Æ �yM   • " $x
   ÿ Ž ÊO xø  ÿÀŽ È z  ÿàõ Ê $vÿ  ÿ Ì �y    › ¢ $x   ÿ Ž â� r .0   )àµ ä xÿ   
  ä xø  ÿÀŽ ä r25   ÿd~ â %v  D Ž â z  ÿàù Æ %v   Z  ! Ž ä �y    › ä $x	$   ÿŽ ä $vÿ  ÿ ä $x!!   	Ž È �y    › è �y9    ™ è  �y:    ™ â  x	ÿ     Æ �y;    ™ ä  %v	  D Ž ä �y<    ™ è  �y    › è �y=    ™ â  $x
   ÿ Ž Æ �y>    ™ ä  x

ø  ÿÀŽ ä �y?    ™ ä  z

  ÿàõ ä �y@     ™ æ  $vÿ  ÿ Ä $x   ÿ Ž â xÿ     â/ �yH     • è  �y

    › â xÿ     Ê %v  D Ž â xø  ÿÀŽ Ê �y    › â z  ÿàõ Ê $vÿ  ÿ Ì �y    › â %v  D Ž Ì �y    › â $x   ÿ Ž Êxø  ÿÀŽ È z  ÿàõ ä xNLÿ   ÿÀŽ Æ $vÿ  ÿ â x#ÿ     ä x"ÿ   L  Æ �y    › " 	sN N        b %v#  D Ž â xP"ÿ   ÿÀŽ Ê �y    › " 	sJ P        " r"1   ÿàõ â $xON€   ÿ Ž È/ $x#   3 â xOO€   ÿÀŽ â $xNK   ÿ Ž Æ �yF"    • " $xQJ€   ÿ Ž â xOOÿ   LxŽ â %x$B   "Ž â �yJ   • " x Nø  ÿÀŽ ä xQP€   QxŽ â $tPÿÿ Ž ä $xOOÿŽ â x"ÿ   7  â%x&B   $Ž â xLLT  7    â �y$$    • b z   ÿàõ Ä x O€€€€PìŽ â�yI&    • " x!O€€€€RìŽ ä x""ÿ   ÿÀŽ â %xB   &Ž â x%ÿ   L  ä/ xL ÿþþþÿàÿ â $tNÿÿ Ž â xKÿ   K  â $vÿ  ÿ â xP!ÿþþþÿàÿ â $x'QÿŽ â 	s# "        " $t&ÿ @€ÿ Ž â x%%ÿ   ÿÀŽ â �y    • " xLLOøŽ â%v K  D Ž â xPP @€OøŽ â �y    › " xN'€€€€NìŽ â	sK %        b xOL˜º  ÿ    â �y      › " xQ'€€€€&ìŽ Ä xPP˜º  ÿ    ä x&Nÿþþþÿàÿ ä xLQÿþþþÿàÿ â $x##€   ÿ Ž â x&&'øŽ â$xRK€   ÿ Ž â/ xLL @€'øŽ â $tKÿÿ Ž Æ xLL˜º  ÿ    ä xN€€€€OÞŽ äO xOOÿÀŽ ä xQP€€€€ÞŽ äxPPÿÀŽ â $xON   O
Ž È $xPQ   P
Ž â xNO€€€€´Ž ä x"€   #xŽ ä x#P€€€€´Ž â $xG   ÿ Ž â x"%€   RxŽ â $x'ÿŽ ä $tPÿÿ Ž â xø  ÿÀŽ â $x""ÿŽ Ä $tRÿ @€ÿ Ž â x'€€€€PìŽ ä z  ÿàõ ä xPÿþþþÿàÿ ä xK"€€€€KìŽ â $vÿ  ÿ â x%'€€€€RìŽ â $tOÿ @€ÿ Ž â xKKÿþþþÿàÿ Ä xR%ÿþþþÿàÿ â �y    › ¢ x%P'øŽ ä xP"€€€€OìŽ ä xOK"øŽ ä xK&˜º  ÿ    ä xSPÿþþþÿàÿ Ä xQO˜º  ÿ    ä x&KÿÀŽ ä xOLÿÀŽ ä xK€€€€KÞŽ ä� xLL€€€€ÞŽ ä xSS @€"øŽ â $xKK   &
Ž â xPQÿÀŽ â $x&L   O
Ž â xQ€€€€QÞŽ â &rNN9   ÿ   â x"S˜º  ÿ    Ä xOÿ   G  â &rN#:   N   â xK€€€€´Ž â $xQQ   P
Ž â xL"ÿÀŽ äxG"€€€€ÞŽ â %v"O  D Ž â x&&€€€€´Ž Æ $xPG   L
Ž â xLQ€€€€´Ž â &rO;   N   â xÿ   8  â �y"    › â  $xNE   ÿ Ž â x'R @€'øŽ â &rO&<   O   â x&ÿ   ÿÀŽ ä xNø  ÿÀŽ Ä x%%˜º  ÿ    ä xN'˜º  ÿ    ä z  ÿàõ ä x"%ÿÀŽ ä x'NÿÀŽ ä x%€€€€%ÞŽ Ä xNN€€€€	ÞŽ ä xGP€€€€´Ž â $vÿ  ÿ â xEÿ   E  â $x%%   "
Ž ä $xNN   '
Ž â xK8ÿ   ÿÀŽ â �y    › ¢ x%%€€€€´Ž Ä xNN€€€€	´Ž â %vE  D Ž â 	sK K        * �y    › ¦ 	s# &        b $xKK€   ÿ Ž Ê xKK€   ÿÀŽ È x"Kÿ   8xŽ â $tEÿÿ Ž È $x""ÿŽ ä $x'#€   ÿ Ž Æ/ x#"€€€€EìŽ â &rLL=   O   â xDÿ   A  ä x##ÿþþþÿàÿ â &rLG>   L   â x7ÿ   7  â $tEÿ @€ÿ Ž â xG&€   'xŽ ä x'#"øŽ Ä x#Dÿ   ÿÀŽ ä x&7   ÿüŽ ä x7"€€€€EìŽ â 	sK #        " x88T  A    â &r%%?   L   â xE7ÿþþþÿàÿ ä x8ÿ   8  â &r%N@   %   â xEE @€"øŽ Ä x"8ÿ   ÿÀŽ â $r%&   ÿŽ â x''˜º  ÿ    ä x&E˜º  ÿ    ä 	sE "        b x7
€€€€'ÞŽ â $xLK€   ÿ Ž â xD'ÿÀŽ ä x%&€€€€ÞŽ Ä x&&ÿÀŽ â $x77   D
Ž â xL#€   LxŽ â $xGGÿŽ ä $t8ÿÿ Ž â x
7€€€€
´Ž â $x&%   &
Ž ä $xLLÿŽ ä $t#ÿÿ Ž Ä $tDÿ @€ÿ Ž ä $t%ÿ @€ÿ Ž â x'G€€€€8ìŽ â $xEE€   ÿ Ž â/ x&&€€€€´Ž â &r
9   ÿ   â x
L€€€€#ìŽ äxDG€€€€DìŽ ä x#L€€€€%ìŽ Ä x8'ÿþþþÿàÿ ä xE"€   ExŽ â &r"&:      â x
ÿþþþÿàÿ ä xDDÿþþþÿàÿ ä x##ÿþþþÿàÿ ä x88GøŽ â$t&ÿÿ Ž â x
LøŽ â $xEEÿŽ â xDD @€GøŽ Ä x# @€LøŽ ä x#8˜º  ÿ    ä x'E€€€€&ìŽ ä xDD˜º  ÿ    ä x&#ÿÀŽ ä x#€€€€#ÞŽ Ä x%DÿÀŽ äxDD€€€€ÞŽ â $x##   &
Ž â x&'ÿþþþÿàÿ Æ $xDD   %
Ž â x&&EøŽ â $t8ÿ @€ÿ Ž â x'Bÿ   ÿÀŽ ä xDD€€€€´Ž ä x&˜º  ÿ    ä x##€€€€´Ž â 	s' '        " x7E€€€€8ìŽ Ä xÿÀŽ â &r##;   "   â x€€€€ÞŽ ä x"ÿ   B  ä x87ÿþþþÿàÿ â $x   
Ž â x"ÿ   ÿÀŽ ä x88 @€EøŽ Ä 	s"         b &rDD<   #   â x88˜º  ÿ    ä x#
˜º  ÿ    â $x''€   ÿ Ž â x€€€€´Ž ä x%8ÿÀŽ ä x˜º  ÿ    Ä x88€€€€ÞŽ ä x
#ÿÀŽ ä x#€€€€#ÞŽ ä xÿÀŽ â $x88   %
Ž â x€€€€ÞŽ ä x''€   ÿÀŽ â $x##   

Ž â x88€€€€´Ž â $x   
Ž â x''ÿ   BxŽ â $x"€   ÿ Ž â/ x##€€€€´Ž â &r=   D   Ä $x''ÿŽ ä $tÿ @€ÿ Ž â x€€€€´Ž â &r
8>      â x€   xŽ â $tÿÿ Ž â x'€€€€ìŽ â &r
#?   
   ä $t#ÿÿ Ž â x'€€€€ìŽ â $xÿŽ â xÿþþþÿàÿ â $tÿ @€ÿ Ž â xBBT  C    Ä x€€€€#ìŽ ä xÿþþþÿàÿ ä x @€'øŽ ä xÿ   B  ä x"€€€€ìŽ ä xÿþþþÿàÿ Ä x'øŽ ä x˜º  ÿ    ä xÿ   ÿÀŽ ä x#"ÿþþþÿàÿ ä x"øŽ ä x˜º  ÿ    Ä x€€€€ÞŽ â	s%         " xÿÀŽ ä x&ÿ   C  ä x€€€€ÞŽ â $x   
Ž â xÿÀŽ ä x"˜º  ÿ    Ä x## @€øŽ ä x&&ÿ   ÿÀŽ â $x   
Ž â xÿÀŽ ä 	s" &        b x#˜º  ÿ    ä x€€€€ÞŽ ä x€€€€´Ž Ä x€€€€´Ž â $x   
Ž â xÿÀŽ â$x%€   ÿ Ž â x€€€€ÞŽ â &r9   ÿ   â x€€€€´Ž ä x€   xŽ â $x   
Ž ä &r:      Ä $x"€   ÿ Ž â/ x€€€€´Ž â &r;      ä $xÿŽ ä $tÿ @€ÿ Ž ä $tÿÿ Ž â x&&€   xŽ â &r<      â x€€€€ìŽ Ä xJÿ   ÿÀŽ ä x€€€€ìŽ â &r
@   
   â xÿþþþÿàÿ â 	s         " xÿþþþÿàÿ â $tÿÿ Ž ä $x&&ÿŽ â x @€øŽ â$tÿ @€ÿ Ž â xøŽ Ä x&€€€€ìŽ ä x˜º  ÿ    ä x˜º  ÿ    ä x&€€€€ìŽ ä xÿþþþÿàÿ ä xÿÀŽ Ä xÿÀŽ ä x€€€€ÞŽ ä x€€€€ÞŽ â $x€   ÿ Ž â xÿþþþÿàÿ â $x   
Ž â x&øŽ â $x   
Ž â xÿ   J  Ä x @€&øŽ ä x€   ÿÀŽ ä xÿ   ÿÀŽ ä x˜º  ÿ    ä x€€€€´Ž ä xÿ   JxŽ Ä x˜º  ÿ    ä 	s         " xÿÀŽ â &r=      â x €€€€ÞŽ â $xÿŽ ä $tÿÿ Ž â xÿÀŽ â$x   
Ž â x€€€€!ÞŽ â $tÿ @€ÿ Ž â xJJT  M    Ä x€€€€ìŽ â$x   
Ž â xJÿ   J  ä x€€€€ìŽ ä xÿþþþÿàÿ ä x€€€€´Ž â $x€   ÿ Ž â xJJÿ   ÿÀŽ Ä xÿþþþÿàÿ ä xøŽ â &r>      â x!€€€€!´Ž ä 	s J        " x€€€€ ´Ž ä x @€øŽ ä x˜º  ÿ    Ä x€   xŽ ä xÿ   M  â &r?      â x˜º  ÿ    ä x€€€€ÞŽ âO $tÿÿ Ž â xÿ   ÿÀŽ â $xÿŽ â xÿÀŽ Ä x€€€€ÞŽ â	s         b xÿÀŽ â $x   
Ž â x€€€€ìŽ â $tÿ @€ÿ Ž ä $x   
Ž â xÿþþþÿàÿ â $x€   ÿ Ž â x€€€€ìŽ Ä xøŽ ä xÿþþþÿàÿ ä xJ€   xŽ ä x€€€€´Ž ä x˜º  ÿ    â $tÿÿ Ž â x @€øŽ â $xÿŽ Ä $x€   ÿ Ž â/ xÿÀŽ â &r99   ÿ   â x€€€€ÞŽ â� $tÿ @€ÿ Ž â x€€€€ìŽ ä x€   xŽ ä x˜º  ÿ    â $x   
Ž â x€€€€´Ž â $xÿŽ â xÿþþþÿàÿ â $tÿÿ Ž â x€€€€ìŽ Ä xÿÀŽ äx€€€€ÞŽ â &r:   9   â x€€€€´Ž â $tÿ @€ÿ Ž â xÿþþþÿàÿ â $x   
Ž â xøŽ äx€€€€ìŽ â &r;      â x @€øŽ Ä x€€€€ìŽ ä x˜º  ÿ    ä xÿþþþÿàÿ ä x€€€€´Ž ä x˜º  ÿ    ä xÿþþþÿàÿ Ä xÿÀŽ ä xøŽ ä x€€€€ÞŽ â &r<      â xÿÀŽ ä x @€øŽ ä x˜º  ÿ    â $x   
Ž â x€€€€ÞŽ Ä x˜º  ÿ    ä xÿÀŽ â $x   
Ž â x€€€€ÞŽ ä x€€€€´Ž ä xÿÀŽ â$x   
Ž â x€€€€	ÞŽ â &r=      â x€€€€´Ž Ä x€€€€´Ž â $x   
Ž ä &r>      â xAÿ   A  ä xCÿ   C  â &r?      â xMÿ   M  ä x€€€€	´Ž â &r!@      â xAA   ÿüŽ Ä xC   ÿüŽ â &r@      â xM   ÿüŽ â $r

A   ÿŽ ä $r   ÿŽ ä $r   ÿŽ â x44   ÿàÿ ä xÿ     Ä x	ÿ   
  ä xÿ     ä xÿ     ä | 4   pbô ä r   ÿ� ä r
	
   ÿ� Ä r   ÿ� ä r   ÿ� â 0rFÿF    A   â xÿ     â 0r	ÿH    A   â x
ÿ   
  â 0r$ÿ$    A   âxÿ     â 0rÿI    A   â xÿ     â 0rÿ    A   â  rF	     A  â  r	$     A  â r	     A  â x11   ÿà÷ â  r		     A  â Er        Ä Er
 
       ä Er        ä Er        â #r6   6   â #r,
   ,   â #r+   +   â x))€   ÿàÿ â #r*	   *   â $r3ÿÿ   3Ž â D)       À â Gy  ÐÛÿÿÿÿƒ ê Ay        € ê ˆ -6 þÿ H   è ˆ -,€þÿ H   è ˆ -+ ÿÿ H   è ˆ -*€ÿÿ H   è {          ì M	        € ê „y(     H   " z / r  p`ô Æ „y	( €   H   b r (ÿ   pVt Æ „y(    H   ¨„y(    H   è „y( €  H   ( „y(    H   ( „y(    H   ( „y( €  H   ( „y( €  H   ( „y(    H   ( „y( €  H   " !r6       Æ „y
( €  H   " !r		,       â/ !r+       âO !r       ä� !r	       â!r       â !r       â !r*       È ‰      b !r       â !r       È ‰      ¢ !r       Æ ‰      â !r
       Ê ‰      " !r       â/ !r       ÈO ‰      b !r
       Æ� ‰      ¨ ‰	
      â !r       Ê ‰      " !r       â/ !r       ÈO ‰ €    b !r
	       Æ� ‰	 €    ¨ ‰ €    â !r       Ê ‰ €    " !r       â/ !r		       âO !r
       È� ‰	 @    h ‰ @    ¢ !r       Æ ‰
 @    ( ‰ @    â !r	       â/ x	/   ÿàÿ â !r       ÆO ‰      b $z  x  /Ž â x /   ÿàÿ â!r
       â x/   ÿàÿ â ‰      " z   r  p`ö â !r       â� z  r  p`ð â ‰      ¢ z 	 r  p`ò â $tÿ   ÿ Ž â x (   pVö â‰
      â x (   pVp  â$z ~  Ž â x (   pVò  Æ %v j   Ž â !r       â/ !r       È †¹    â !r       ÆO †©     â !r
       Æ� †‰    â M        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      &   " z ÿ \  pPò â ¹z  F   
   Æ z ÿ ]  Sò Ú $ÿ   ÿ Ž È % \   Ž Ê �    ™ ¢  $rÿÿ   ÿ Ž â Ey  0    € ä y      %   b ' m   Ž Æ y      "   ä ÿ n    ä y       !   & $ÿÿ   
Ž â y	      '   " 'š s   Ž È $ o  Ž â ’        â $x   ÿ Ž â/ r ÿ   pRð È� x     pDp  â$x     Ž â r    ÿàÿ È z  r  pdp  â $rÿÿ   ÿ Ž â xÿ     ä šÿ t    â $rÿÿ   ÿ Ž Æ r     € ÊO G	  À    € ê z ÿ `  pPò ä z ÿ ^  pPð ä z ÿ b  p ô ä z ÿ a  Sò ä z ÿ _   Sð ä z ÿ c   'ò Ö $
	 ~  ÿŽ âÿ     â$š	 ~  ÿŽ ä $
 x  ÿŽ â     àë â$š x  ÿŽ â ’    àÇ ä ˜ÿ     ä    ÿàõ Ä ÿ   ÿÄþ ä ’   ÿàû ä ’
ÿ   ÿ„þ â $ÿÿ    â 
 ^  ÿ„ â $rÿÿ   ÿ Ž â š b  ÿ† â $’ÿÿ   
Ž â 
 _   â $rÿÿ   ÿ Ž Æ š c  � â �	     h �™     d Ay        € ê r   ÿ� â ¹z  l      â $z	 }  ÿŽ â ™x?    â xÿ     â $z w  ÿŽ â r ÿ       â Ey   J    € â $rGÿÿ   ÿ Ž â |    pbð â x
      ÿ  â r   ÿàó â 'z		 y   Ž â x      ÿ  â x      ÿ  Ä $rÿÿ   ÿ Ž ä $rÿÿ   ÿŽ  È G	  pI    € ê xøÿÿÿÿÀŽ â $z p  Ž âz	ÿ z  	  â $vÿ `  ÿ Ž â x   ÿàÿ â$x   
Ž â x   ÿàÿ â $z	 |  Ž â x   ÿàÿ â $z	 |  Ž â x&   ÿ   â $z v  Ž ârG ÿ       â $z v  Ž â x
      ÿ  â %x&&   ÿŽ â x      ÿ  â xÿ     Ä $z p  Ž ä %x B   &Ž È $z p  Ž âx    ÿàñ â $z p  Ž ä %x"B   &Ž â x      ÿ  Æ %x$B   &Ž È $vÿ a  ÿ Ž ä $vÿ X  ÿ Ž ä $vÿ Y  ÿ Ž ä $rÿÿ   ÿ Ž ä $x   ÿ Ž ä %x&B   &Ž È $rÿÿ   !  ä r    ÿàñ Ê $x!     Ê �y,   • ¨ �y-   • â r*   ÿàñ â $vWÿ p  ÿ Ž ä �yZ   • " r+   ÿä  Æ �yM   • ä  %x*WB   *Ž Ê �y	*úÿÿ • è �y)*üÿÿ • " %xHWB   *Ž Æ �yQ*þÿÿ • ( �y(Húÿÿ • ( �ySHüÿÿ • " %x0WB   HŽ Æ �yP*    • ( �yT0úÿÿ • " $tYÿ$   ÿ Ž Æ �yR0üÿÿ • ( �yNH    • " %x`B   Ž Æ �yO0þÿÿ • " $x,   ÿ Ž âO x>ÿ   ,  ä �yK`    • ¤ xø  ÿÀŽ È z.  ÿàñ Ê $v/ÿ  ÿ  Ì �y..    › ¢ $x-   ÿ Ž Ê� xø  ÿÀŽ È z<  ÿàñ È z=ÿ  ÿä  ä rX   à‘ â $tÿ   ÿ Ž Æ �y<<    › â r   ÿ$~  â %v>>   Ž â x:ÿ   -  Æ %vXX Z  Y Ž ä �y>>    › ä $x$   ÿŽ È $xYY   Ž ä %v::   Ž Æ �yX   ™ è  �y::    › è �yX   ™ è  �y@X   ™ è  �yAX   ™ è  �yBX   ™ è  �yCX   ™ è  �yDX   ™ â  $x,	   ÿ Ž Æ �yEX    ™ ä  x,,ø  ÿÀŽ ä x4ÿ   	  ä x6ÿ   )  äx2Zÿ   ÿÀŽ ä x3ÿ   Z  â %x^WB   `Ž â z8,  ÿàñ â �yLX    • & z9ÿ  ÿä  â %v44   Ž È $x	)   ÿ Ž â �y88    › ( �y44    › " x		ø  ÿÀŽ â %v66   Ž â 	s2 2        " x33ÿ   ÿÀŽ â �yJ^    • " z,	  ÿàñ Æ �y	Hþÿÿ • $ $v-ÿ  ÿ  ä �y66    › ( �y,,    › " $x)(   ÿ Ž Ê x))ø  ÿÀŽ È z*)  ÿàñ Ê/ $v+ÿ  ÿ  Ì �y**    › " 	sF 3        b x(ÿ   (  â $x[2€   ÿ Ž â xUS   ÿ   â %x\WB   ^Ž È %v((   Ž â x[[€   ÿÀŽ â �yI\    • " xUUø  ÿÀŽ â %xVWB   \Ž â x[[ÿ   ZxŽ â �y((    › " z2U  ÿàñ â $x`F€   ÿ Ž ä/ �yHV    • " $x][ÿŽ Ä xU3€   `xŽ â $v3ÿ  ÿ  â xX  @€    â �yF0    • " $tVÿÿ Ž ä/ xX]€€€€XìŽ â�y22    › $ x`]€€€€VìŽ ä x1ZT  M    â $t\ÿÿ Ž â x``ÿþþþÿàÿ â $xUUÿŽ â x1ÿ   1  Ä xZXÿþþþÿàÿ ä xWÿ   M  ä xX1ÿ   ÿÀŽ ä x[`]øŽ äx]Z @€]øŽ ä xZU€€€€\ìŽ â 	sY X        " xWWÿ   ÿÀŽ â $t0ÿ @€ÿ Ž â xZZÿþþþÿàÿ Ä x[[˜º  ÿ    Æ 	sV W        b x^ÿ   S  ä x]]˜º  ÿ    ä xSZUøŽ ä x0U€€€€0ìŽ â %v^^   Ž Æ x00ÿþþþÿàÿ È xU0 @€UøŽ ä �y0^    › $ $t^ÿ @€ÿ Ž â xZ.€€€€[ÞŽ äO x[[ÿÀŽ ä x\]€€€€/ÞŽ äx]]ÿÀŽ â $x[Z   [
Ž È $x\\   ]
Ž â x]Y   ÿ   ä xY[€€€€.´Ž â $x.V€   ÿ Ž â/ x]X€   ]xŽ â $xZT   ÿ Ž â xV\€€€€/´Ž â $t\ÿÿ Ž â xXW€   .xŽ â $xW]ÿŽ â xZZø  ÿÀŽ È x\W€€€€\ìŽ äz.Z  ÿàñ ä x^W€€€€^ìŽ ä x\\ÿþþþÿàÿ â $xXXÿŽ â x]     â $v/ÿ  ÿ  â x^^ÿþþþÿàÿ Ä x\\WøŽ ä xZX€€€€]ìŽ â �y..    › ¢ x^^ @€WøŽ ä xW\˜º  ÿ    â $t[ÿ @€ÿ Ž â x]Zÿþþþÿàÿ ä xZWÿÀŽ Ä xW<€€€€WÞŽ ä� x[X€€€€[ìŽ Æ $xaW   Z
Ž â x[[ÿþþþÿàÿ ä xWS˜º  ÿ    ä x^^˜º  ÿ    ä xZU˜º  ÿ    ä xSa€€€€<´Ž ä x]]XøŽ Ä x[[ @€XøŽ ä x<WÿÀŽ ä xX^ÿÀŽ äx_^€€€€=ÞŽ ä xW>€€€€WÞŽ ä xUZÿÀŽ Ä xZZ€€€€?ÞŽ â $xX_   X
Ž â rWW<  €ÿàÿ â &rYY   ÿ   â x]]˜º  ÿ    â $xZZ   U
Ž â xWW€€€€>´Ž â &rYV   Y   â xU]ÿÀŽ ä x<:€€€€]ÞŽ Ä xXX€€€€=´Ž â $x=R   ÿ Ž â x>ÿ   T  â &rYW@   Y   â xZZ€€€€?´Ž â $xU<   U
Ž â x<ÿ   Q  â %v>>   Ž â x==ø  ÿÀŽ Æ &rYZA   Y   â xZ[˜º  ÿ    â �y>>    › â xT<ÿ   ÿÀŽ ä z<=  ÿàñ ä x[ZÿÀŽ äxZZ€€€€;ÞŽ â $v=ÿ  ÿ  â xUU€€€€:´Ž Ä rZZ[  €ÿàÿ ä x:ÿ   R  â �y<<    › ¢ xZZ€€€€;´Ž Æ %v::   Ž â xVQÿ   ÿÀŽ Ê �y::    › ¢ 	sV V        " &rSSB   Y   Î 	sW T        b &rSXC   S   ä $xVV€   ÿ Ž ä &rRUD   S   Æ xVV€   ÿÀŽ â $xUW€   ÿ Ž Æ/ xSVÿ   QxŽ â $tVÿ @€ÿ Ž â xMÿ   M  ä xUT€   UxŽ â $tTÿÿ Ž ä $xSSÿŽ â xQQT  P    ä xWM   ÿüŽ Ä xMS€€€€TìŽ äxVS€€€€VìŽ ä xQÿ   Q  ä xTMÿþþþÿàÿ ä xXVÿþþþÿàÿ ä xQQÿ   ÿÀŽ Ä xYÿ   P  â &rRZE   R   â xVTSøŽ â	sM Q        " xXX @€SøŽ â $xUUÿŽ â xSYÿ   ÿÀŽ ä xZ     ä x[V˜º  ÿ    â 	sT S        b x\X˜º  ÿ    â $t^ÿ @€ÿ Ž â xYU€€€€ZìŽ â $rRRW   ÿŽ â xX8€€€€[ÞŽ Äx[[ÿÀŽ ä xV\€€€€9ÞŽ äxW\ÿÀŽ ä x\Yÿþþþÿàÿ ä xZU€€€€^ìŽ â $x[X   [
Ž â x\\UøŽ â $xVV   W
Ž â xZZÿþþþÿàÿ â $xXM€   ÿ Ž â xM\˜º  ÿ    Ä x[[€€€€8´Ž ä xZZ @€UøŽ ä x8V€€€€9´Ž ä xVT   ÿ   ä/ xTMÿÀŽ ä xM4€€€€MÞŽ Ä xQQ€   XxŽ ä xZZ˜º  ÿ    â $xMM   T
Ž â xVS€   VxŽ â &r9[   ÿ   â xUZÿÀŽ â$xQQÿŽ â xZZ€€€€5ÞŽ â $tTÿÿ Ž ä $tXÿ @€ÿ Ž â xMM€€€€4´Ž â &r98   9   â xSQ€€€€TìŽ â$x8Z   U
Ž â xUQ€€€€XìŽ â $x4VÿŽ â xW     â $tYÿ @€ÿ Ž â xTSÿþþþÿàÿ Ä xVUÿþþþÿàÿ ä xU4€€€€WìŽ äxW4€€€€YìŽ ä xSTQøŽ äxVV @€QøŽ ä xUUÿþþþÿàÿ Ä xWWÿþþþÿàÿ ä xSS˜º  ÿ    â &rTM@   9   â x58€€€€5´Ž ä xVV˜º  ÿ    ä x8U4øŽ äx9W @€4øŽ Ä x4SÿÀŽ â &r55A   T   â xS,€€€€SÞŽ ä xMVÿÀŽ äxT	ÿ   ÿÀŽ ä xVV€€€€-ÞŽ â $xSS   4
Ž ä 	s4 T        " xQ8˜º  ÿ    â $xVV   M
Ž â x89˜º  ÿ    Ä xSS€€€€,´Ž ä xVV€€€€-´Ž ä x,QÿÀŽ ä x-8ÿÀŽ ä xQ6€€€€QÞŽ ä x88€€€€7ÞŽ Ä x9ÿ   	  â $xQQ   ,
Ž â r88-  €ÿàÿ ä x-9ÿ   ÿÀŽ â &r,SB   5   â x88€€€€7´Ž â $x44€   ÿ Ž â 	s7 -        " xQQ€€€€6´Ž â &r5VC   ,   Ä x44€   ÿÀŽ ä &r,QD   5   â x5ÿ   R  ä xPÿ   P  ä x44ÿ   	xŽ ä rR5R   ÿ� â &r,8E   ,   â x5P   ÿüŽ â $x44ÿŽ â x		T  N    â $t9ÿÿ Ž Ä $tMÿ @€ÿ Ž ä $x87€   ÿ Ž â x7ÿ   	  â $r,,5   ÿŽ â x54€€€€9ìŽ äx64€€€€MìŽ ä x8-€   8xŽ ä x55ÿþþþÿàÿ Ä x-7ÿ   ÿÀŽ ä x	6ÿþþþÿàÿ ä x554øŽ â	sP -        " x4	 @€4øŽ ä x	ÿ   N  ä x75˜º  ÿ    Ä x		ÿ   ÿÀŽ â $x58ÿŽ â x6*€€€€7ÞŽ ä x77ÿÀŽ ä x84˜º  ÿ    â $tVÿ @€ÿ Ž â xT     â 	s4 	        b x98€€€€+ÞŽ â $x76   7
Ž â x88ÿÀŽ Ä x65€€€€TìŽ â$xTP€   ÿ Ž â xM5€€€€VìŽ â $x89   8
Ž â x77€€€€*´Ž ä xPMÿþþþÿàÿ ä xT-€   TxŽ ä x66ÿþþþÿàÿ â &r*7   ÿ   â xPP @€5øŽ â $xTTÿŽ â x+8€€€€+´Ž â $t-ÿÿ Ž â x665øŽ â $t5ÿ @€ÿ Ž â x44   ÿ   â/ &r++   *   â xPP˜º  ÿ    Ä x		€   4xŽ ä x*T€€€€-ìŽ äx4T€€€€5ìŽ ä x5PÿÀŽ ä x-6˜º  ÿ    ä xPP€€€€)ÞŽ Ä x7*ÿþþþÿàÿ ä x*-ÿÀŽ â $xPP   5
Ž â x-(€€€€-ÞŽ ä x77TøŽ ä x94ÿþþþÿàÿ ä x5Oÿ   ÿÀŽ â $x--   *
Ž â x77˜º  ÿ    Ä x99 @€TøŽ ä 	s5 5        " x*7ÿÀŽ ä x72€€€€7ÞŽ ä x49˜º  ÿ    ä x--€€€€(´Ž ä xPP€€€€)´Ž Ä r77*  €ÿàÿ â $x*	ÿŽ â x)4ÿÀŽ â&r(-@   +   â x44€€€€3ÞŽ â $t	ÿÿ Ž ä $t+ÿ @€ÿ Ž ä $x44   )
Ž â x)*€€€€	ìŽ Ä x+*€€€€+ìŽ â $x55€   ÿ Ž â x))ÿþþþÿàÿ ä x++ÿþþþÿàÿ â &r	PA   (   â x77€€€€2´Ž ä x))*øŽ ä x25€   ÿÀŽ Ä x++ @€*øŽ â &r(7B   	   â x44€€€€3´Ž ä x))˜º  ÿ    ä x-2ÿ   OxŽ äx*ÿ   O  ä x2+˜º  ÿ    â &r	4C   (   â x()ÿÀŽ Ä x+0€€€€)ÞŽ ä x**ÿ   ÿÀŽ ä x)2ÿÀŽ â$x3-ÿŽ â x22€€€€1ÞŽ â $t4ÿÿ Ž â 	s5 *        " $t6ÿ @€ÿ Ž â r++(  €ÿàÿ â $x22   )
Ž â x)3€€€€4ìŽ Ä x(ÿ   F  äxOOT  F    ä x-3€€€€6ìŽ ä x4)ÿþþþÿàÿ ä x((ÿ   ÿÀŽ ä x)ÿ   O  Ä x6-ÿþþþÿàÿ ä 	s- (        b x443øŽ äx))ÿ   ÿÀŽ ä x66 @€3øŽ â $x55€   ÿ Ž â x34˜º  ÿ    â 	s8 )        " x66˜º  ÿ    Ä x4.€€€€3ÞŽ äO x7*€   5xŽ ä x33ÿÀŽ ä x56€€€€/ÞŽ äx66ÿÀŽ ä r343  €ÿàÿ Ä x*2€€€€1´Ž â $x65   6
Ž â x33€€€€.´Ž â $x1-€   ÿ Ž ä/ $x77ÿŽ ä $t4ÿ @€ÿ Ž â x88   ÿ   â $t2ÿÿ Ž â x66€€€€/´Ž Ä x1(€   1xŽ â &r(3   ÿ   â x-7€€€€4ìŽ ä x0+€€€€0´Ž ä x+7€€€€2ìŽ ä x8)€   8xŽ â &r)6   (   â x2-ÿþþþÿàÿ â $t(ÿÿ Ž â x.+ÿþþþÿàÿ â $x-1ÿŽ Ä $x88ÿŽ ä $t+ÿÿ Ž ä $t3ÿ @€ÿ Ž â x/-€€€€(ìŽ äx1-€€€€4ìŽ ä x(8€€€€+ìŽ äx+8€€€€3ìŽ Ä x4/ÿþþþÿàÿ ä x61ÿþþþÿàÿ ä x..7øŽ ä x/(ÿþþþÿàÿ ä x1+ÿþþþÿàÿ ä x+4-øŽ Ä x(6 @€-øŽ ä x22 @€7øŽ ä x-.˜º  ÿ    ä x//8øŽ ä x22˜º  ÿ    ä x.-ÿÀŽ Ä x->€€€€-ÞŽ ä� x81 @€8øŽ ä x1/˜º  ÿ    ä x/2ÿÀŽ â$x--   .
Ž â x22€€€€?ÞŽ ä x88˜º  ÿ    Ä x41ÿÀŽ â $x22   /
Ž â x1<€€€€1ÞŽ ä x--€€€€>´Ž ä x38ÿÀŽ äx++˜º  ÿ    ä x88€€€€=ÞŽ Ä r114  €ÿàÿ â &r.-@   )   â x22€€€€?´Ž ä x4(˜º  ÿ    ä x(+ÿÀŽ â $x88   3
Ž â x+:€€€€+ÞŽ â &r)2A   .   â x11€€€€<´Ž Ä x-4ÿÀŽ â$x++   (
Ž â x44€€€€;ÞŽ â &r(1B   )   â x88€€€€=´Ž â &r	0D   	   âx++€€€€:´Ž â $x44   -
Ž â xNÿ   N  â &r)8C   (   â xFÿ   F  â &r	*E   	   â x44€€€€;´Ž â &r(+D   )   â xNN   ÿüŽ Ä x)F   ÿüŽ â &r(4E   (   ä $rN	N   ÿŽ â x	ÿ   ,  â $r(()   ÿŽ â z ÿ `  pPð ä r		,   ÿ� â 0rLÿL    A   â z ÿ a   Sð â 0r)ÿJ    A   â x-ÿ   (  Ä x	ÿ   	  ä x+ÿ   N  â 0rKÿK    A   â r--(   ÿ� â  r(L)     A  â Er	 	       ä xRÿ   R  ä r++N   ÿ� â 0rIÿI    A   â ErR R       â 0r3ÿH    A   â #rG(	   G   â  rKKL     A  â x+ÿ   +  Ä x	ÿ   -  â #rKR      â  r2LI     A  â r3L3     A  â Er4 +       ä Er	 	       â G‰  °    € ê r(    ÿàó È r)!   ÿäÿ  Ê �y*(    ¨ �y0(    è �y6(    ( �y5(    â  $x+*   ÿ Ž ÊO x++ø  ÿÀŽ È z.+  ÿàó Ê $v/ÿ  ÿŽ  Ì �y..    › ¢ x,ÿ   *  â $x*0   ÿ Ž È� %v,,   Ž â x**ø  ÿÀŽ Ê �y,,    › â z**  ÿàó Ê $v+ÿ  ÿŽ  â x(ÿ   0  Ê �y**    › â %v((   Ž Ì �y((    › â %x0B   Ž Ì �y00     â  x76ÿ   ÿÀŽ â$t<ÿÿ Ž â x:ÿ   6  â $tFÿ @€ÿ Ž Æ 	s7 7        b x::ÿ   ÿÀŽ Î 	s; :        " x87   ÿ   È/ x98€   ÿÀŽ ä x8ÿ   5  äx99ÿ   6xŽ â $x;;€   ÿ Ž âx66T  5    ä x>8ÿ   ÿÀŽ â $x99ÿŽ â x8ÿ   6  Ä 	s? >        b x;:€   ;xŽ ä x19€€€€<ìŽ ä x<8ÿ   ÿÀŽ â $x;;ÿŽ â x61ÿþþþÿàÿ ä x79€€€€FìŽ â 	s= <        " x669øŽ Ä x87ÿþþþÿàÿ ä x16˜º  ÿ    â $x??€   ÿ Ž â/ x88 @€9øŽ ä x:     ä x88˜º  ÿ    ä x?>€   ?xŽ Ä x5ÿ   5  â $x==€   ÿ Ž ä $x??ÿŽ â x55   ÿüŽ ä x=<€   =xŽ â $t<ÿ @€ÿ Ž È $x==ÿŽ â x6.€€€€1ÞŽ äO x11ÿÀŽ ä x78€€€€/ÞŽ äx88ÿÀŽ â $x16   1
Ž ä $t6ÿ @€ÿ Ž â r878  €ÿàÿ Ä x11€€€€.´Ž â $t.ÿÿ Ž â x88€€€€/´Ž ä x/;€€€€6ìŽ äx.;€€€€.ìŽ ä x7?€€€€:ìŽ ä x6.ÿþþþÿàÿ â &r.1   ÿ   â x1?€€€€<ìŽ â $t<ÿÿ Ž â x:/ÿþþþÿàÿ â &r/8   .   â x.7ÿþþþÿàÿ Ä x81ÿþþþÿàÿ ä x66;øŽ äx:: @€;øŽ ä x;.?øŽ äx.8 @€?øŽ â $t8ÿ @€ÿ Ž â x16˜º  ÿ    Ä x7=€€€€<ìŽ äx::˜º  ÿ    ä x9=€€€€8ìŽ ä x61ÿÀŽ ä x87ÿþþþÿàÿ ä x1,€€€€1ÞŽ Ä� x7:ÿÀŽ äx::€€€€-ÞŽ â $x11   6
Ž â x<9ÿþþþÿàÿ ä x88=øŽ är::7  €ÿàÿ ä x<< @€=øŽ Ä x78˜º  ÿ    ä x11€€€€,´Ž ä x<<˜º  ÿ    ä x67ÿÀŽ â &r,1@   /   â x7*€€€€7ÞŽ ä x/<ÿÀŽ Ä x<<€€€€+ÞŽ â $x77   6
Ž â x;;˜º  ÿ    ä x::€€€€-´Ž â $x<<   /
Ž â x77€€€€*´Ž ä x..˜º  ÿ    â &r-:A   ,   â x*;ÿÀŽ Ä x;(€€€€;ÞŽ ä x<<€€€€+´Ž ä x+.ÿÀŽ â$x;;   *
Ž â x..€€€€)ÞŽ â &r*7B   -   Æ x;;€€€€(´Ž â $x..   +
Ž ä &r+<C   *   Æ x..€€€€)´Ž â &r(;D   +   ä 0r+ÿ0    A   ä &r(.E   (   È $r((5   ÿŽ Ê x)ÿ   (  È r))(   ÿ� â  r(L+     A  Æ x)ÿ   )  È Er) )       Ê #r()      ä G‰  °    € ê r("   ÿàó È r)#   ÿäÿ  Ê �y*(    ¨ �y0(    è �y6(    ( �y5(    â  $x+*   ÿ Ž ÊO x++ø  ÿÀŽ È z.+  ÿàó Ê $v/ÿ  ÿŽ  Ì �y..    › ¢ x,ÿ   *  â $x*0   ÿ Ž È� %v,,   Ž â x**ø  ÿÀŽ È z**  ÿàó â �y,,    › è $v+ÿ  ÿŽ  â x(ÿ   0  Ê �y**    › â %v((   Ž Ì �y((    › â %x0B   Ž Ì �y00     â  x76ÿ   ÿÀŽ â$t<ÿÿ Ž â x:ÿ   6  â $tFÿ @€ÿ Ž Æ 	s7 7        b x::ÿ   ÿÀŽ Î 	s; :        " x87   ÿ   È/ x98€   ÿÀŽ ä x8ÿ   5  äx99ÿ   6xŽ â $x;;€   ÿ Ž âx66T  5    ä x>8ÿ   ÿÀŽ â $x99ÿŽ â x8ÿ   6  Ä 	s? >        b x;:€   ;xŽ ä x19€€€€<ìŽ ä x<8ÿ   ÿÀŽ â $x;;ÿŽ â x61ÿþþþÿàÿ ä x79€€€€FìŽ â 	s= <        " x669øŽ Ä x87ÿþþþÿàÿ ä x16˜º  ÿ    â $x??€   ÿ Ž â/ x88 @€9øŽ ä x:     ä x88˜º  ÿ    ä x?>€   ?xŽ Ä x5ÿ   5  â $x==€   ÿ Ž ä $x??ÿŽ â x55   ÿüŽ ä x=<€   =xŽ â $t<ÿ @€ÿ Ž È $x==ÿŽ â x6.€€€€1ÞŽ äO x11ÿÀŽ ä x78€€€€/ÞŽ äx88ÿÀŽ â $x16   1
Ž ä $t6ÿ @€ÿ Ž â r878  €ÿàÿ Ä x11€€€€.´Ž â $t.ÿÿ Ž â x88€€€€/´Ž ä x/;€€€€6ìŽ äx.;€€€€.ìŽ ä x7?€€€€:ìŽ ä x6.ÿþþþÿàÿ â &r.1   ÿ   â x1?€€€€<ìŽ â $t<ÿÿ Ž â x:/ÿþþþÿàÿ â &r/8   .   â x.7ÿþþþÿàÿ Ä x81ÿþþþÿàÿ ä x66;øŽ äx:: @€;øŽ ä x;.?øŽ äx.8 @€?øŽ â $t8ÿ @€ÿ Ž â x7=€€€€<ìŽ Ä x16˜º  ÿ    ä x::˜º  ÿ    ä x9=€€€€8ìŽ ä x87ÿþþþÿàÿ ä x61ÿÀŽ ä x7:ÿÀŽ Ä x1,€€€€1ÞŽ ä� x::€€€€-ÞŽ ä x88=øŽ â $x11   6
Ž â x<9ÿþþþÿàÿ ä r::7  €ÿàÿ ä x78˜º  ÿ    Ä x<< @€=øŽ ä x11€€€€,´Ž ä x67ÿÀŽ ä x7*€€€€7ÞŽ â &r,1@   /   â x<<˜º  ÿ    ä x;;˜º  ÿ    â $x77   6
Ž â x/<ÿÀŽ Ä x<<€€€€+ÞŽ ä x::€€€€-´Ž ä x77€€€€*´Ž â $x<<   /
Ž â x*;ÿÀŽ â &r-:A   ,   â x;(€€€€;ÞŽ ä x..˜º  ÿ    Ä x<<€€€€+´Ž â $x;;   *
Ž â x+.ÿÀŽ â&r*7B   -   â x..€€€€)ÞŽ ä x;;€€€€(´Ž ä r..+  €ÿàÿ â &r+<C   *   Æ x..€€€€)´Ž â &r(;D   +   ä 0r+ÿ0    A   ä &r(.E   (   Æ  r+L+     A  â $r((5   ÿŽ Ê x)ÿ   (  È r))(   ÿ� È x)ÿ   )  È Er) )       Ê #r+)      ä G‰  °    € ê r(&   ÿàó Ê $x)'   Ž  Ê �y*(    ¨ �y0(    è �y6(    ( �y5(    â  $x+*   ÿ Ž ÊO x++ø  ÿÀŽ È z.+  ÿàó Ê $v/ÿ  ÿŽ  Ì �y..    › ¢ x,ÿ   *  ä x*0   ÿ   Æ� %v,,   Ž â x**ø  ÿÀŽ Ê �y,,    › â z**  ÿàó Ê $v+ÿ  ÿŽ  â x(ÿ   0  Ê �y**    › â %v((   Ž Ì �y((    › â %x0B   Ž Ì �y00     â  x76ÿ   ÿÀŽ â$t<ÿÿ Ž â x:ÿ   6  ä xF  @€    ä 	s7 7        b x::ÿ   ÿÀŽ Î 	s; :        " $x87€   ÿ Ž Ê/ x98€   ÿÀŽ ä x8ÿ   5  â$x;;€   ÿ Ž âx99ÿ   6xŽ ä x66T  5    ä x>8ÿ   ÿÀŽ â $x99ÿŽ â x8ÿ   6  Ä 	s? >        b x;:€   ;xŽ ä x19€€€€<ìŽ äx79€€€€FìŽ â $x;;ÿŽ â x61ÿþþþÿàÿ ä x<8ÿ   ÿÀŽ ä x669øŽ Ä x87ÿþþþÿàÿ â 	s= <        " x16˜º  ÿ    ä x88 @€9øŽ ä x??   ÿ   ä/ x88˜º  ÿ    ä x?>€   ?xŽ Ä x:     ä x5ÿ   5  â $x??ÿŽ ä $x==€   ÿ Ž â x55   ÿüŽ È x=<€   =xŽ â $t<ÿ @€ÿ Ž È $x==ÿŽ â x6.€€€€1ÞŽ äO x11ÿÀŽ ä x78€€€€/ÞŽ äx88ÿÀŽ â $x16   1
Ž ä $t6ÿ @€ÿ Ž Ä $x87   8
Ž â x11€€€€.´Ž â $t.ÿÿ Ž â x7?€€€€:ìŽ ä x88€€€€/´Ž ä x.;€€€€.ìŽ äx/;€€€€6ìŽ ä x6.ÿþþþÿàÿ â &r.1   ÿ   â x1?€€€€<ìŽ â $t<ÿÿ Ž â x:/ÿþþþÿàÿ â &r/8   .   â x.7ÿþþþÿàÿ Ä x81ÿþþþÿàÿ ä x66;øŽ äx:: @€;øŽ ä x;.?øŽ äx.8 @€?øŽ â $t8ÿ @€ÿ Ž â x16˜º  ÿ    Ä x7=€€€€<ìŽ äx::˜º  ÿ    ä x9=€€€€8ìŽ ä x61ÿÀŽ ä x87ÿþþþÿàÿ ä x1,€€€€1ÞŽ Ä� x7:ÿÀŽ äx::€€€€-ÞŽ ä x<9ÿþþþÿàÿ ä x88=øŽ â $x::   7
Ž â r116  €ÿàÿ ä x<< @€=øŽ Ä x78˜º  ÿ    ä x11€€€€,´Ž ä x<<˜º  ÿ    ä x67ÿÀŽ â &r,1@   /   â x7*€€€€7ÞŽ ä x/<ÿÀŽ Ä x<<€€€€+ÞŽ â $x77   6
Ž â x;;˜º  ÿ    ä x::€€€€-´Ž â $x<<   /
Ž â x..˜º  ÿ    ä x77€€€€*´Ž â &r-:A   ,   â x*;ÿÀŽ Ä x;(€€€€;ÞŽ ä x<<€€€€+´Ž ä x+.ÿÀŽ äx..€€€€)ÞŽ ä r;;*  €ÿàÿ â &r*7B   -   ä $x..   +
Ž â x;;€€€€(´Ž â &r+<C   *   Æ x..€€€€)´Ž â &r(;D   +   ä 0r+ÿ0    A   ä &r(.E   (   È $r((5   ÿŽ Ê x)ÿ   (  È r))(   ÿ� â  r(L+     A  Æ x)ÿ   )  È Er) )       Ê #r()      â #r24      â #r3	      â G‰       € ê r($   ÿàñ Ê $x)%     Ê �y*(    ¨ �y.(    è �y2(    ( �y	(    â  %x0B   Ž Ì �y00     â $x+*   ÿ Ž ÊO x++ø  ÿÀŽ È z,+  ÿàñ È z-ÿ  ÿä  ä x*ÿ   *  È �y,,    › ¢ %v**   Ž Ì �y**    › ¢ $x/.   ÿ Ž â� x.ÿ   .  È x//ø  ÿÀŽ È z(/  ÿàñ Ê $v)ÿ  ÿ  ä %v..   Ž È �y((    › è �y..    › â x2ÿ   ÿÀŽ Ì	s         " x3ÿ   2  È x43ÿ   ÿÀŽ È 	s5 4        " $x3€   ÿ Ž Ê x33€   ÿÀŽ È x33ÿ   2xŽ â $t8ÿ @€ÿ Ž â x22T  	    â $x55€   ÿ Ž âx6     â $x33ÿŽ â x2ÿ   2  ä x54€   5xŽ ä x3€€€€6ìŽ Ä x13€€€€8ìŽ ä/ x72ÿ   ÿÀŽ ä x2ÿþþþÿàÿ ä x41ÿþþþÿàÿ â $x55ÿŽ â xÿ   	  â $t<ÿ @€ÿ Ž â x223øŽ â	s8 7        " x44 @€3øŽ Ä x9ÿ   ÿÀŽ ä x2˜º  ÿ    ä x35€€€€6ìŽ äx44˜º  ÿ    ä x65€€€€<ìŽ ä x2ÿÀŽ Ä x66ÿþþþÿàÿ â 	s: 9        f x66 @€5øŽ â $x88€   ÿ Ž Æ x66˜º  ÿ    ä x87€   8xŽ Ê $x88ÿŽ ä $x::€   ÿ Ž Ê/ x:9€   :xŽ Ê $x::ÿŽ â x	ÿ   	  È x		   ÿüŽ ä x1,€€€€ÞŽ äO x4€€€€-ÞŽ äx44ÿÀŽ Ê $x4   4
Ž â r112  €ÿàÿ ä x23ÿþþþÿàÿ ä x44€€€€-´Ž ä x-6ÿÀŽ äx66€€€€+ÞŽ ä x11€€€€,´Ž Ä x225øŽ â $x66   -
Ž ä &r1   ÿ   ä $t-ÿÿ Ž â x2˜º  ÿ    â &r4      ä $t1ÿ @€ÿ Ž â x8€€€€-ìŽ Ä x,ÿÀŽ ä x*€€€€ÞŽ ä x-ÿþþþÿàÿ â $t3ÿÿ Ž â x8€€€€1ìŽ ä r,  €ÿàÿ ä x5  @€    Ä x--8øŽ ä x1ÿþþþÿàÿ ä x€€€€*´Ž ä x*:€€€€3ìŽ äx,:€€€€5ìŽ ä x--˜º  ÿ    Ä x11 @€8øŽ ä x3*ÿþþþÿàÿ ä x5,ÿþþþÿàÿ ä x-ÿÀŽ ä x*1˜º  ÿ    ä x-(€€€€-ÞŽ â� &r@@      â x33:øŽ Ä x55 @€:øŽ ä x*ÿÀŽ â$x--   
Ž â x**€€€€)ÞŽ ä x33˜º  ÿ    ä x66€€€€+´Ž ä x,5˜º  ÿ    â $x**   
Ž â x3ÿÀŽ â &rA6A   @   â x3.€€€€3ÞŽ Ä x--€€€€(´Ž ä x,ÿÀŽ äx,,€€€€/ÞŽ â &rB-B   A   â r33  €ÿàÿ ä x**€€€€)´Ž â $x,,   
Ž â x33€€€€.´Ž Æ &rC*C   B   â x,,€€€€/´Ž Æ &rD3D   C   È &r,,E   D   È $r	,	   ÿŽ Ê xÿ   	  È r	   ÿ� â 0r	ÿ0    A   Æ xÿ     ä  r	L	     A  ä Er        Ê #r
	   
   â x   ÿàÿ ä x   ÿàó ä |    pbð ä x   ÿàõ â $rÿÿ   Ž  â x€   ÿàÿ Æ $rÿÿ    Ì D	       À â Gy   ¸ÿÿÿÿƒ ê Ay        € ê y      "   " Ey  0    € â r ÿ   pRð är ÿ   pRò Ö G‰  ð     € ê z ÿ `  pPð ä r    ÿ8Ž ä z ÿ a   Sð Æ ˆs  þÿ H   ô ˆƒ G€þÿ H   è ˆƒ  ÿÿ H   è ˆƒ €ÿÿ H   â G‰  p     € ê ˆs    H   è ˆs G€þÿ H   è ˆs €  H   è ˆs  ÿÿ H   è ˆs    H   è ˆs €ÿÿ H   è ˆs 
€  H   ä Ay        € ê {          ì M        € ê y      &   ¢ z ÿ `  pPð Æ y      '   "> z ÿ a   Sð â $z x  Ž äO $tÿ   ÿ Ž ä $z ~  Ž È %v j   Ž È G	  P    € ê „y      H   " z  r  p`ð Æ „y
  €   H   b r  ÿ   pVp  Æ „y     H   ¨ „y     H   è„y  €  H   ( „y     H   ( „y	     H   ( „y  €  H   ( „y     H   ( „y  €  H   ( „y  €  H   " !r       Æ „y  €  H   " !r

G       â/ !r       âO !r       ä� !r

       â!r       â !r	       ä !r

       Æ ‰      b !r       Æ ‰
      ¢ !r       Æ ‰	      â !r       È !r       â !r       â/ !r
       ÈO ‰      " !r	       Æ� ‰      è ‰	      h ‰      ¢ !r       â !r	       â/ !r	       Æ ‰ €    " !r
       ÆO ‰      h ‰	 €    ¨ ‰
 €    â !r       â !r       â/ !r	       ÈO ‰ €    " !r
       Æ� ‰ @    h ‰ @    ¨ ‰ @    â !r       â !r       È/ ‰ @    " x   ÿàÿ â!r       âO x   ÿàÿ â ‰      b z  r  p`ò â !r       â� z  r  p`ô â ‰	      ¢ x   ÿàÿ ä x     pVò  â‰      â x     pVt â !r       â !r       È/ ‰      " !‚       â !r	       âO !r	       Æ� †‰     â z  r  p`ð â !’       â!¢		       ä x     pVp  ä †™    è †© 	   â !r       Ì M	        € ê !r       Ê †y    â My        € ê $vÿ h  ÿ Ž Ê x    pBð Ú G	  @    € ê x    pRð Ú G‰  À    € ê x    pRð Ú G	   $    € ê „y	      H   " z  r  p`ð â Ey  p    € ä „y     H   b r  ÿ   pVp  Æ „y     H   ¨ „y     H   è „y     H   ( „y   
  H   ( „y  €  H   ( „y  €  H   â !r		       â !r       Æ/ !r		       äO „y  €  H   â !r       â� !r	       ä„y  €   H   â !r       Æ ‰      " !r       Æ ‰      b !r       â !r       È/ ‰	      " !r       Æ ‰      b !rG       â !r		       â !r       È/ ‰	 €    ( ‰ €    h „y  €  H   ¢ !r	       Æ „y	  €
  H   â !r       Æ/ ‰ @    ( ‰ @    b !r       ÈO !r       â !r       â !r       È/ ‰      ( ‰      b !r	       â !r       â !r       â/ G	  �     € ê !r       â!r       Æ  x	;ª¸¿  A  Ì s	 	       $ !t	  €?    Ì s        $  r     A  È  r     A  Ê †y     ä Ay        € ê ‰      " Ey  Ð    € æ ‰	      h „y     H   ¨ „y     H   â !r       Æ „y     H   â !r	       Æ/ ‰      ( ‰      h „y     H   " !r       ÈO !r       â� !r       â !r       È/ ‰	 €    ( ‰ €    b !r       Æ „y   	  H   ¢ !r       â!r		       ä „y     H   â !r       â/ x   ÿàÿ ä ‰	 @    $ z  r  p`ð Ä ‰ @    d x     pVp  â !r       âO !r	       â !r       È/ ‰      ( ‰	      b !r       â !r       â !r	       â/ G	  �     € ê !r       â!r       Æ  x;ª¸¿  A  Ì s        $ !t	  €?    Ì s 	       $  r     A  È  r     A  Ê †y    ä Ay        € ê ‰      " Ey       € æ ‰      h „y  €  H   ¨ „y  €	  H   è „y  €  H   ( „y  €  H   â !r       Æ „y  €  H   â !r	       Æ/ „y  €  H   è ‰      ( ‰	      b !r
       ÈO !r       È� !r       â!r       â !r	       È/ ‰ €    ( ‰ €    b !r       â !r       È/ ‰ @    ( ‰ @    b !r       â x   ÿàÿ â !r       Æ/ ‰      " z  r  p`ð Æ ‰      b x     pVp  â !r       â !r       Ö/ G	  �     € ê !r       â!r       Æ  x;ª¸¿  A  Ì s        $ !t	  €?    Ì s
 	       $  r

     A  È  r
     A  Ê †y    ä Ay        € ê ‰      b !r       â x   ÿàÿ Æ !r       â z  r  p`ð Æ !r       â x     pVp  È ‰      " !r	       Ê/ ‰	      b !r       Ê ‰      " !r
	       Ê/ ‰
 €    b !r       Ê ‰ €    " !r
       Ê/ ‰ @    b !r       Ê ‰ @    " !r       Ê/ ‰	      b !r       Ê ‰      ¢  !r		       â/ M	        € ì !r	       â!r       ÆO  x ;ª¸¿  A  â !r       Ê s          $ !t   €?    Ì s        $  r     A  È  r     A  Ê †y    â My        € ê „y      H   " z  r  p`ð â Ey       € ä „y     H   b r  ÿ   pVp  Æ „y     H   ¨ „y	     H   è „y     H   ( „y   
  H   ( „y  €  H   ( „y  €  H   è „y  €
  H   â !r       â !r       Æ/ !r       âO !r		       Æ� !r       â!r	       È ‰      " !r       Æ ‰      h „y  €  H   ¢ !r       Æ „y  €   H   â !r       Æ/ ‰      ( ‰      b !r       ÈO !r       â !r       â !r       È/ ‰	 €    ( ‰ €    b !rG       Æ „y  €  H   ¢ !r		       â !r       È/ ‰	 @    ( ‰ @    b !r       ÈO !r       â !r	       â !r       È/ ‰      ( ‰      b !r       â !r       â/ G	  À     € ê !r       È x'7=  A  â x*BL?  A  Æ #t	  €?   È  r		     A  â  x   ?  A  â !r       È s	 	    $   $ !t	  €?    È  r     A  È  r     A  Ê †y     ä Ay        € ê ‰      " Ey       € æ ‰      h „y     H   ¨ „y     H   è „y   	  H   è „y     H   â !r       Æ „y     H   â !r       Æ/ ‰      ( ‰      b !r       âO !r       â� !r       Æ !r       â !r       â/ ‰	 €    $ !r       ä ‰ €    h „y     H   ¢ !r		       â x   ÿàÿ â !r       Æ/ ‰	 @    " z  r  p`ð Æ ‰ @    b x     pVp  â !r       ÈO !r       â !r	       â !r       È/ ‰      ( ‰	      b !r       â !r	       â/ G	  À     € ê !r       È x'7=  A  â x*BL?  A  Æ #t  €?   È  r     A  â  x   ?  A  â !r       È s     $   $ !t  €?    È  r     A  È  r     A  Ê †y    ä Ay        € ê ‰      " Ey  Ð    € æ ‰      h „y  €  H   ¨ „y  €	  H   è „y  €  H   è „y  €  H   â !r       Æ „y  €  H   â !r	       Æ/ ‰      ( ‰	      h „y  €  H   " !r
       ÈO !r       â� !r       â !r	       È/ ‰ €    ( ‰ €    b !r       â!r       â x   ÿàÿ â !r       Æ/ z  r  p`ð â ‰ @    & x     pVp  â ‰ @    b !r       â !r       È/ ‰      ( ‰      b !r	       â !r       â/ G	  À     € ê !r       È x'7=  A  â x*BL?  A  Æ #t  €?   È  r     A  â  x   ?  A  â !r	       È s     $   $ !t
  €?    È  r

     A  È  r
     A  Ê †y    ä Ay        € ê ‰      b !r       â x   ÿàÿ Æ !r       â z  r  p`ð Æ !r       â x     pVp  È ‰      " !r	       Ê/ ‰	      b !r       Ê ‰      " !r
	       Ê/ ‰
 €    b !r       Ê ‰ €    " !r
       Ê/ ‰ @    b !r       Ê ‰ @    " !r       Ê/ ‰	      b !r       Ê ‰      ¢  !r		       â/ M	        € ì !r	       È x'7=  A  â x *BL?  A  Æ #t  €?   â x   ?  A  Æ  r       A  â !r       ÈO !r       â s       $   $ !t   €?    È  r     A  È  r     A  Ê †y    â My        € ê x    pRð Ú G‰  P    € ê x    pRð Ú G	  `    € ê „y      H   " z  r  p`ð â Ey       € ä „y     H   b r  ÿ   pVp  Æ „y     H   ¨ „y	     H   è „y     H   ( „y   
  H   ( „y  €  H   ( „y  €  H   è „y  €
  H   â !r       â !r       Æ/ !r       âO !r		       Æ� !r       â!r	       È ‰      " !r       Æ ‰      h „y  €  H   ¢ !r       Æ „y  €   H   â !r       Æ/ ‰      ( ‰      b !r       ÈO !r       â !r       â !r       È/ ‰	 €    ( ‰ €    b !rG       Æ „y  €  H   ¢ !r		       â !r       È/ ‰	 @    ( ‰ @    b !r       ÈO !r       â !r	       â !r       È/ ‰      ( ‰      b !r       â !r       â/ G	  À     € ê !r       Ê	z i    � â !r       È  x	;ª¸¿  A  â 	z i    � È 	z i €  � â s	 	       $ !t	  €?    Ì s        $  r     A  È  r     A  Ê †y     ä Ay        € ê ‰      " Ey       € æ ‰      h „y     H   ¨ „y     H   è „y   	  H   è „y     H   â !r       Æ „y     H   â !r       Æ/ ‰      ( ‰      b !r       âO !r       â� !r       Æ !r       â !r       â/ ‰	 €    $ !r       ä ‰ €    h „y     H   ¢ !r		       â x   ÿàÿ â !r       Æ/ ‰	 @    " z  r  p`ð Æ ‰ @    b x     pVp  â !r       ÈO !r       â !r	       â !r       È/ ‰      ( ‰	      b !r       â !r	       â/ G	  À     € ê !r       Ê	z i    � â !r       È  x;ª¸¿  A  â 	z i    � È 	z i €  � â s        $ !t	  €?    Ì s 	       $  r     A  È  r     A  Ê †y    ä Ay        € ê ‰      " Ey  Ð    € æ ‰      h „y  €  H   ¨ „y  €	  H   è „y  €  H   è „y  €  H   â !r       Æ „y  €  H   â !r	       Æ/ ‰      ( ‰	      h „y  €  H   " !r
       ÈO !r       â� !r       â !r	       È/ ‰ €    ( ‰ €    b !r       â!r       â x   ÿàÿ â !r       Æ/ z  r  p`ð â ‰ @    & x     pVp  â ‰ @    b !r       â !r       È/ ‰      ( ‰      b !r       â !r       â/ G	  À     € ê !r       Ê	z i    � â !r       È  x;ª¸¿  A  â 	z i    � È 	z i €  � â s        $ !t	  €?    Ì s
 	       $  r

     A  È  r
     A  Ê †y    ä Ay        € ê ‰      b !r       â x   ÿàÿ Æ !r       â z  r  p`ð Æ !r       â x     pVp  È ‰      " !r	       Ê/ ‰	      b !r       Ê ‰      " !r
	       Ê/ ‰
 €    b !r       Ê ‰ €    " !r
       Ê/ ‰ @    b !r       Ê ‰ @    " !r       Ê/ ‰	      b !r       Ê ‰      ¢  !r		       â/ M	        € ì !r	       â!r       ÈO 	z i    � â !r       È  x ;ª¸¿  A  â 	z i    � È 	z i €  � â s          $ !t   €?    Ì s        $  r     A  È  r     A  Ê †y    â My        € ê „y      H   " z  r  p`ð Æ „y     H   b r  ÿ   pVp  Æ „y     H   ¨ „y     H   è „y  €   H   ( „y  €  H   ( „y      H   ( „y"     H   ( „y	     H   ( „y  €  H   ( „y  €  H   " !r       Æ „y     H   " !r       Æ/ „y   	  H   b !r       ÆO „y   
  H   ¢ !r       Æ� „y  €  H   â !rG       Æ„y  €
  H   " !r       Æ „y$     H   " !r         Æ „y&     H   " !r""       Æ „y  €  H   " !r		       Æ „y  €  H   " !r       â x   ÿàÿ ä „y(  €  H   " !r       â z  r  p`ô ä „y*  €	  H   â !r        â x     pVt Ä „y,  €  H   â !r"       Æ/ „y.  €  H   â !r	       ÆO ‰      " !r	       â� !r       ä‰      b !r$       Æ ‰	      ¢ !r&       Æ ‰      â !r       Æ ‰      " !r
       Æ ‰      " !r(       â !r       â !r*       â !r       Æ/ ‰      " !r,,       â !r		       äO ‰      b !r.       â !r       ä� ‰,      ¢ !r       Æ‰      â !r       â x   ÿàÿ ä‰
	      " x   ÿàÿ ä z  r  p`ò â ‰      " z  r  p`ö â !r       Ä ‰      " x     pVò  äx     pVö â ‰      " !r       â/ !r,       äO ‰ €    b !r       Æ� ‰ €    ¢ !r
	
       Æ‰      â !r       Æ ‰      " !r       Æ ‰	
 €    " !r       Æ ‰ €    ( ‰ €    " !r       Æ/ ‰ €    b !r       âO !r       ä� ‰ @    ¢ !r       Æ‰ €    â !r	
	       Æ ‰ €    " !r       Æ ‰ @    " !r       Æ ‰
	 @    " !r       Æ/ ‰ @    h ‰ @    b !r       ÆO ‰ @    ¢ !r       â� !r       ä ‰      " !r       Æ‰ @    â !r
	
       Æ ‰ @    " !r       Æ/ ‰      b !r       Æ ‰	
      b !r       ÆO ‰      ¨ ‰      ¢ !r       Æ ‰      " !r       â� !‚       â!r       Æ‰      â !r       Æ/ ‰      b !r
	       â !‚       â !r       äO !’       â  ‚     A  â !r       â !’       Æ †‰     â !r       â !¢       â  ’     A  ä !¢
       â !r       ä� †™    â  ¢
     A  â !r       È/ †©    â M9        € ê !r       â !r       ÈO  r     A  Ê †y    â My        € ê „y      H   " z  r  p`ð â Ey  À    € ä „y     H   b r  ÿ   pVp  Æ „y     H   ¨ „y	     H   è „y     H   ( „y   
  H   ( „y  €  H   ( „y  €  H   è „y  €
  H   â !r       â !r       Æ/ !r       âO !r		       Æ� !r       â!r	       È ‰      " !r       Æ ‰      h „y  €  H   ¢ !r       Æ „y  €   H   â !r       Æ/ ‰      ( ‰      b !r       ÈO !r       â !r       â !r       È/ ‰	 €    ( ‰ €    b !rG       Æ „y  €  H   ¢ !r		       â !r       È/ ‰	 @    ( ‰ @    b !r       ÈO !r       â !r	       â !r       È/ ‰      ( ‰      b !r       â !r       â/ G	  à     € ê !r       Ê	x  à@  � Ê  x#ÛÙ?  A  È  x	;ª¸¿  A  â !r       Ê s	 	       " 	x  à@  � È 	x  àÀ  � Ê !t  €?    â !t	  €?    È s        $  r     A  È  r     A  Ê †y     ä Ay        € ê ‰      " Ey       € æ ‰      h „y     H   ¨ „y     H   è „y   	  H   è „y     H   â !r       Æ „y     H   â !r       Æ/ ‰      ( ‰      b !r       âO !r       â� !r       Æ !r       â !r       â/ ‰	 €    $ !r       ä ‰ €    h „y     H   ¢ !r		       â x   ÿàÿ â !r       Æ/ ‰	 @    " z  r  p`ð Æ ‰ @    b x     pVp  â !r       ÈO !r       â !r	       â !r       È/ ‰      ( ‰	      b !r       â !r	       â/ G	  à     € ê !r       Ê	x  à@  � Ê  x#ÛÙ?  A  È  x;ª¸¿  A  â !r       Ê s        " 	x  à@  � È 	x  àÀ  � Ê !t  €?    â !t	  €?    È s 	       $  r     A  È  r     A  Ê †y    ä Ay        € ê ‰      b Ey  ð    € æ ‰      ( „y  €  H   ¨ „y  €	  H   è „y  €  H   è „y  €  H   â !r       Æ/ „y  €  H   â !r	       Æ ‰      ( ‰	      h „y  €  H   " !r
       ÈO !r       â� !r       â !r	       È/ ‰ €    ( ‰ €    b !r       â!r       â x   ÿàÿ â !r       Æ/ z  r  p`ð â ‰ @    & x     pVp  â ‰ @    b !r       â !r       È/ ‰      ( ‰      b !r       â !r       â/ G	  à     € ê !r       Ê	x  à@  � Ê  x#ÛÙ?  A  È  x;ª¸¿  A  â !r       Ê s        " 	x  à@  � È 	x  àÀ  � Ê !t
  €?    â !t	  €?    È s 	       $  r     A  È  r
     A  Ê †y    ä Ay        € ê ‰      " !r       â x   ÿàÿ Æ !r       â z  r  p`ð Æ !r       â x     pVp  È ‰      b !r	       Ê ‰	      " !r       Ê/ ‰      b !r
	       Ê ‰
 €    " !r       Ê/ ‰ €    b !r
       Ê ‰ @    " !r       Ê/ ‰ @    b !r       Ê ‰	      " !r       Ê/ ‰      ¢ !r		       â M	        € ì !r	       â/!r       ÈO 	x  à@  � â !r       È  x #ÛÙ?  A  â 	x  à@  � Æ  x  ;ª¸¿  A  â 	x  àÀ  � Ê s          " !t  €?    â !t   €?    Ì s        $  r     A  È  r     A  Ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y       "   $ z   w  p`ð Ú M	        € ê y      !   " ¹z  l      â Ey  À    € â ™x?    â y      %   b ¹z  F   
   â $rÿÿ   ÿ Ž ä $r!ÿÿ   ÿ Ž â x$ÿ     È | $   p`ð â $x   ÿ Ž Ø/ G	       € ê y      &   " $tÿ   ÿ Ž ä $z  x  Ž È %v \   Ž Ì �y    ™ ¢ $rÿÿ   ÿ Ž â x   ÿÀŽ â $z q  $Ž ä 'z m   Ž â x   ÿàÿ Æ $t#ÿB   ÿ Ž â z	ÿ n    â %x   ÿ Ž È $rÿÿ   	
Ž â x%   ÿüŽ â $z  r  ÿŽ ä $rÿÿ   ÿ Ž ä $x$   ÿ Ž ä $z	 t  Ž äO$z t  $Ž Ä $z o  Ž ä $z" q  Ž ä %x	B   Ž È %v	 X  #Ž â z X  ÿàó Æ $z u  ÿŽ â x   ÿàñ â %v"" X  #Ž â x   ÿàõ ä r   ÿà÷ â $rÿÿ     â zÿ Y  $~ Æ $r ÿÿ   ÿŽ ä �y
úÿÿ • ¨ �yüÿÿ • â r"%   ÿàñ Æ �y&þÿÿ • $ $x   #  ä �y'    • h �yþÿÿ • è �y1    • â xÿ     â $t0ÿ   ÿ Ž Ä �y,   • " $tÿ$   ÿ Ž Æ �y/   • ( �y2üÿÿ • " $x
   ÿ Ž ÊO xø  ÿÀŽ È z  ÿàñ Ê $vÿ  ÿ  Ì �y    › ¢ $x   ÿ Ž Ê� xø  ÿÀŽ È z  ÿàõ ä r   à‘ Æ $v	ÿ  ÿ â rÿ    $~  ä xÿ   
  Æ �y    › ä %v
  0 Ž È %v Z   Ž ä �y

    › ä $x$   ÿŽ È $x   Ž â xÿ     È �y)   ™ è �y-   ™ â %v  0 Ž Æ �y+   ™ â $x   ÿ Ž Æ �y    › ä xø  ÿÀŽ ä �y*   ™ ä z  ÿàñ ä �y(   ™ æ $vÿ  ÿ  â �y.   ™ â $x1   ÿ Ž â xÿ     Ä xÿ   1  â �y    › â xø  ÿÀŽ â %v  0 Ž ä �y1   ™ â z  ÿàñ Æ �y    › ä $vÿ  ÿ  ä %v  0 Ž â �y3    • è �y    › è �y    › è �y0    ™ ä $rÿÿ   " Ž Ä/ $rÿÿ   # Ž Ê �y4    • â  x8&ÿ   ÿÀŽ Ì	s8 8        b x6ÿ   &  È x66ÿ   ÿÀŽ â $x58€   ÿ Ž Ê/ x55€   ÿÀŽ È x55ÿ   &xŽ â 	s7 6        b $t:ÿÿ Ž Æ $x55ÿŽ â x&&T  '    Èx5€€€€:ìŽ ä xÿ   &  ä x&ÿþþþÿàÿ â $t:ÿ @€ÿ Ž â xÿ   '  ä x8&5øŽ â $x;7€   ÿ Ž â/ xÿ   ÿÀŽ Ä x95€€€€:ìŽ ä x78˜º  ÿ    â 	s&         " xÿ   ÿÀŽ ä x:9ÿþþþÿàÿ ä x97ÿÀŽ ä x66€   ;xŽ Ä x:: @€5øŽ â $t=ÿ @€ÿ Ž ä $x66ÿŽ â x::˜º  ÿ    È x;:ÿÀŽ â $x&&€   ÿ Ž ä $t<ÿ @€ÿ Ž â x8€€€€7ÞŽ äO 	s7         & $x58   9
Ž ä $t9ÿÿ Ž â x8:€€€€ÞŽ È x:6€€€€9ìŽ äx96€€€€=ìŽ â $x88   ;
Ž â x;:ÿþþþÿàÿ ä x=9ÿþþþÿàÿ â $x77€   ÿ Ž â x9€   &xŽ ä x5€€€€´Ž Ä x;6øŽ â$x99ÿŽ â x&= @€6øŽ â $t6ÿÿ Ž â x5€   7xŽ ä x8€€€€´Ž â $t8ÿ @€ÿ Ž â x9€€€€6ìŽ â $x55ÿŽ ä $t:ÿÿ Ž â x79€€€€8ìŽ Ä x6ÿþþþÿàÿ ä x:5€€€€:ìŽ ä x87ÿþþþÿàÿ ä x5€€€€<ìŽ ä x669øŽ ä x::ÿþþþÿàÿ Ä x88 @€9øŽ ä x<ÿþþþÿàÿ ä x96˜º  ÿ    ä x:5øŽ äx68˜º  ÿ    ä x5< @€5øŽ Ä x<,ÿ   ÿÀŽ ä x86ÿÀŽ äx;6€€€€	ÞŽ ä� 	s6 <        " x79ÿÀŽ ä x:€€€€9ÞŽ ä x˜º  ÿ    Æ $x9:   7
Ž â x:&˜º  ÿ    ä x7ÿÀŽ ä x&
€€€€ÞŽ â &r)   ÿ   È $x7&   7
Ž ä $x66€   ÿ Ž â x9€€€€´Ž â &r-      â x
7€€€€
´Ž ä x9:ÿÀŽ äx::€€€€ÞŽ ä xÿ   ,  Ä x76€   ÿÀŽ â &r

+      â xÿ   ÿÀŽ â $x::   9
Ž â x˜º  ÿ    ä x67ÿ   ,xŽ â 	s         " x&5˜º  ÿ    ä x5ÿÀŽ Ä x€€€€ÞŽ â $x6ÿŽ â x:€€€€´Ž â $t6ÿÿ Ž â x,,T  /    Æ &r
*   
   â x€€€€6ìŽ â $x5   5
Ž â x7&ÿÀŽ â&r(   
   â x&&€€€€ÞŽ â $tÿ @€ÿ Ž â x
ÿþþþÿàÿ â $x8;   8
Ž â x,ÿ   ,  ä x5€€€€´Ž â $x5€   ÿ Ž â x

øŽ â $x&&   7
Ž â x,ÿ   ÿÀŽ Ä x€€€€ìŽ ä x	8€€€€	´Ž ä x
˜º  ÿ    ä x5€   5xŽ ä x&€€€€´Ž â 	s         " xÿþþþÿàÿ Ä x&ÿ   /  â &r		.      â x
€€€€ÞŽ ä xÿÀŽ ä x @€øŽ ä x&ÿ   ÿÀŽ â $x
   
Ž ä $x5ÿŽ â 	s
         b $t5ÿÿ Ž Ä $t7ÿ @€ÿ Ž â x&˜º  ÿ    â $x9€   ÿ Ž â x,€€€€5ìŽ äx6€€€€7ìŽ ä x&€€€€ÞŽ äx&&ÿÀŽ ä x5,ÿþþþÿàÿ Ä x76ÿþþþÿàÿ ä x9€   9xŽ â $x&   &
Ž â x5øŽ äx7 @€øŽ â $x
€   ÿ Ž â/ x€€€€´Ž â $x99ÿŽ ä $t6ÿ @€ÿ Ž â x&&€€€€´Ž â $t,ÿÿ Ž â x€   xŽ â &r
)   ÿ   â x9€€€€6ìŽ Ä x9€€€€,ìŽ â &r
&-   
   â xÿþþþÿàÿ â $t&ÿÿ Ž ä $xÿŽ ä $t,ÿ @€ÿ Ž â xÿþþþÿàÿ ä x @€9øŽ Ä x€€€€&ìŽ äx&€€€€,ìŽ ä x˜º  ÿ    ä xÿþþþÿàÿ ä x&&ÿþþþÿàÿ ä x9øŽ Ä x˜º  ÿ    ä xÿÀŽ ä x€€€€ÞŽ ä xøŽ äx&& @€øŽ ä xÿÀŽ Ä x˜º  ÿ    â $x   
Ž â x€€€€ÞŽ ä x˜º  ÿ    ä xÿÀŽ â $x   
Ž â x€€€€ÞŽ ä x€€€€´Ž Ä xÿÀŽ ä x˜º  ÿ    â $x   
Ž â x€€€€ÞŽ â &r+   
   â x€€€€´Ž ä x&&˜º  ÿ    ä xÿÀŽ â $x   
Ž â x€€€€ÞŽ â &r*      â x€€€€´Ž Ä x&ÿÀŽ â$x   
Ž â x&&€€€€ÞŽ â &r(      â x€€€€´Ž ä x€€€€´Ž â $x&&   
Ž ä &r.      â x'ÿ   '  â &r	1   	   âx&&€€€€´Ž â &r1      â x/ÿ   /  â &r	0   	   âx'   ÿüŽ â &r&0      â x/   ÿüŽ Æ $r	   ÿŽ ä $r   ÿŽ â x$$   ÿàÿ ä xÿ     ä x	ÿ     ä | $   p`ð ä r   ÿ� Ä r		   ÿ� â 0r2ÿ2    A   â xÿ     â 0r3ÿ3    A   â x	ÿ   	  â 0r4ÿ4    A   â x""  ÿàó ä  r223     A  â x  ÿà÷ â  r334     A  â Er        Ä Er	 	       ä x  ÿàõ â $rÿÿ   Ž â #r!2   !   â #r3	      â x    ÿàÿ â $r#ÿÿ   #Ž  ä $rÿÿ    â D	       À â Gy  ðìÿÿÿÿƒ ê Ay        € ê ‰!      " $x   Ž Æ ‰      d z  p  p`ð È x    pDp  â !r!       â !r       È/ ‰      ( ‰      b !r       â !r       È/ ‰ €    ( ‰ €    b !r       â !r       È/ ‰ @    ( ‰	 @    b !r       â !r		       È/ ‰      h  ‰	      ¢  M	        € ê y      &   â $x   ÿ Ž â !r	       âO $z   s  Ž â !r       â? $tÿ   ÿ Ž â x    p"ò Ä r ÿ   p"ð Ö $ÿÿ    Ž ä $z  v   Ž È� %v  j   Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      "   $ z  w  p`ð Ú M	        € ê y      &   " $tÿ   ÿ Ž â ¹z  F   
   â $z x  Ž È %v \   Ž Ê �y     ™ b ¹z  l      â Ey  P'    € â ™x?    â y'      !   b $r
ÿÿ   ÿ Ž â x       ÿ  â $rÿÿ   ÿ Ž â x"      ÿ  â x%ÿ   '  È/ | %   p`ð Ú G	  °&    € ê y      %   " $rÿÿ   ÿ Ž â x''   ÿÀŽ â $z r  ÿŽ â Ey @&    € â 'z m   Ž È $v)ÿ X  ÿ Ž â zÿ n    â $v+ÿ Y  ÿ Ž ä $v-ÿ `  ÿ Ž ä $rÿÿ   
Ž ä $v/ÿ a  ÿ Ž ä $z o  Ž Ä $z  t  %Ž ä$z u  ÿŽ ä $x   ÿ Ž ä $r ÿÿ   ÿ Ž â r$   ÿàñ â $z q  %Ž âx   ÿàÿ â $x&%   ÿ Ž Ä $z  t  Ž ä $z q  Ž ä %xB   ÿŽ È %xB   ÿŽ È %x'    Ž È %x'    Ž È $r(ÿÿ   ÿ  ä r)   ÿàñ Ê $x   +  Ê �y   • ¨ �y   • â r)   ÿàñ Æ �y6   • $ $x   +  ä �y7   • h �y   • è �y   • â $tÿ$   ÿ Ž Ä $t*ÿ   ÿ Ž â �y8   • ( �y9   • " $x
   ÿ Ž ÊO x

ø  ÿÀŽ È z

  ÿàñ Ê $vÿ  ÿ  Ì �y

    › ¢ $x   ÿ Ž Ê� xø  ÿÀŽ È z  ÿàõ ä xÿ   &  är'$   &à‘ â $vÿ  ÿ â xÿ     ä rÿ(   $~  â %v Z   Ž â �y    › æ %v  * Ž È $x$   ÿŽ ä �y    › ä $x   Ž Ê �y,   ™ â xÿ     Æ �y.   ™ è �y0   ™ â %v  * Ž Æ �y1   ™ â  $x   ÿ Ž Æ �y    › ä xø  ÿÀŽ ä �y2   ™ ä  z  ÿàñ ä �y3   ™ æ  $vÿ  ÿ  â xÿ     â $x   ÿ Ž â xÿ     â/ �y4   ™ è  �y    › â %v  * Ž â xø  ÿÀŽ Ä �y5    ™ ä  z  ÿàñ ä �y    › æ $vÿ  ÿ  â r)   ÿàó â%v  * Ž â �y:    • è  �y    › è �y    › â r)   ÿàñ â $x   +Ž  È $x	   +  â �y    • è �y    • â  x=6ÿ   ÿÀŽ Ì	s= =        " x<ÿ   6  È x<<ÿ   ÿÀŽ â $x;=€   ÿ Ž Êx;;€   ÿÀŽ â 	s> <        & x;;ÿ   6xŽ â $tÿÿ Ž È $x;;ÿŽ â x66T  7    Èx;€€€€ìŽ ä xÿ   6  â $t@ÿ @€ÿ Ž â x6ÿþþþÿàÿ â $x=>€   ÿ Ž âxÿ   ÿÀŽ ä x66;øŽ ä x;€€€€@ìŽ Ä x?ÿ   7  â 	s         " x6˜º  ÿ    ä/ x><€   =xŽ ä x<ÿþþþÿàÿ ä x?ÿ   ÿÀŽ ä x=ÿÀŽ Ä x<< @€;øŽ â $t?ÿÿ Ž ä $tAÿ @€ÿ Ž â x<<˜º  ÿ    ä x6
€€€€ÞŽ äO 	s         f $x;6   =
Ž Ä $x6>ÿŽ Ê x>6€€€€?ìŽ äx?6€€€€AìŽ â $xA€   ÿ Ž â x=<€€€€ÞŽ äx<<ÿÀŽ ä xA€   AxŽ â $x€   ÿ Ž â/ x>ÿþþþÿàÿ â $x<=   <
Ž â x;;€€€€
´Ž â $x
AÿŽ â x€   xŽ â $tÿÿ Ž â x6øŽ â $t=ÿ @€ÿ Ž â x<€€€€´Ž â $xÿŽ â x
€€€€ìŽ â$t>ÿÿ Ž â x<
€€€€=ìŽ Ä xÿþþþÿàÿ ä x>€€€€>ìŽ äx=<ÿþþþÿàÿ ä x€€€€@ìŽ ä x
øŽ ä x>>ÿþþþÿàÿ Ä x== @€
øŽ ä x??ÿþþþÿàÿ ä x<ÿþþþÿàÿ ä x˜º  ÿ    ä x
>øŽ ä x=˜º  ÿ    Ä x6? @€6øŽ ä x< @€øŽ ä x?8ÿ   ÿÀŽ ä x<ÿÀŽ ä x€€€€ÞŽ ä� x>ÿÀŽ Ä x=€€€€ÞŽ ä 	s ?        " x˜º  ÿ    â $x   <
Ž â x<6˜º  ÿ    ä x6ÿÀŽ ä x€€€€ÞŽ â $x>=   >
Ž â x=<ÿÀŽ â &r;;,   ÿ   Ä $x   6
Ž â x<<€€€€ÞŽ â &r.   ;   Æ x€€€€´Ž â $x€   ÿ Ž â x€€€€´Ž â $x<<   =
Ž â x>€€€€´Ž â &r0      â xÿ   8  ä x€   ÿÀŽ ä x<€€€€´Ž Ä x
˜º  ÿ    ä x˜º  ÿ    ä xÿ   ÿÀŽ ä xÿ   8xŽ ä x6ÿÀŽ â 	s
         " x<ÿÀŽ â $t;ÿÿ Ž â x€€€€ÞŽ Ä x€€€€ÞŽ â $xÿŽ ä $x   6
Ž ä &r1      â x€€€€;ìŽ ä x88T  9    â &r2      â xÿþþþÿàÿ â $tÿ @€ÿ Ž â x8ÿ   8  â $x<   <
Ž â xøŽ â $x6
€   ÿ Ž â x€€€€´Ž Ä x
8ÿ   ÿÀŽ ä x€€€€ìŽ ä x˜º  ÿ    â 	s 
        " x<€€€€´Ž ä x6€   6xŽ ä xÿþþþÿàÿ Ä x€€€€ÞŽ ä xÿÀŽ ä xÿ   9  Æ $x   
Ž â xÿ   ÿÀŽ ä x @€øŽ ä 	s         d x˜º  ÿ    â $x;€   ÿ Ž ä &r3      â x€€€€ÞŽ â $x66ÿŽ â xÿÀŽ â $tÿÿ Ž Ä $tÿ @€ÿ Ž â x;
€   ;xŽ â $x   
Ž â x6€€€€ìŽ äx6€€€€ìŽ â $x;;ÿŽ â x€€€€´Ž â $t8ÿ @€ÿ Ž â xÿþþþÿàÿ â $x€   ÿ Ž â/ xÿþþþÿàÿ â &r,   ÿ   â x€€€€´Ž Ä x
6øŽ äx @€6øŽ â $t6ÿÿ Ž â x;€€€€8ìŽ ä x€   xŽ â &r.      â xÿþþþÿàÿ ä x
˜º  ÿ    â $xÿŽ â x;€€€€6ìŽ Ä x˜º  ÿ    ä x
ÿÀŽ ä xÿþþþÿàÿ ä x€€€€ÞŽ ä xÿÀŽ äx€€€€ÞŽ â $x   

Ž â x;øŽ Ä x€€€€6ìŽ â$x   
Ž â x€€€€8ìŽ ä x @€;øŽ ä x
ÿþþþÿàÿ ä x˜º  ÿ    ä x€€€€´Ž Ä x6ÿþþþÿàÿ ä x
øŽ ä x˜º  ÿ    ä xÿÀŽ ä x€€€€ÞŽ â &r
0      â x€€€€´Ž Ä x66 @€øŽ ä xÿÀŽ äx˜º  ÿ    â $x   
Ž â x€€€€ÞŽ â &r1   
   â x66˜º  ÿ    ä x
ÿÀŽ â $x   
Ž â x€€€€ÞŽ Ä x€€€€´Ž ä x6ÿÀŽ â$x   

Ž â x66€€€€ÞŽ â &r
2      â x€€€€´Ž ä x€€€€´Ž â $x66   
Ž ä &r3   
   â x7ÿ   7  â &r4      âx66€€€€´Ž â &r4      â x9ÿ   9  â &r5      âx
7   ÿüŽ â &r65      â x99   ÿüŽ Ä z ÿ `  pPð â $r

   ÿŽ ä $r9   ÿŽ â z ÿ a   Sð ä xÿ   
  ä xÿ     ä r
   ÿ� Ä r   ÿ� â 0rÿ    A   â xÿ     â 0r:ÿ:    A   â xÿ     â 0rÿ    A   â Er        ä  r:     A  â Er        â  r:     A  â G‰  À    € ê r-   ÿàó Ê $x   /Ž  Ê �y    ¨ �y    è �y    ( �y    b $x
   ÿ Ž ÊO x

ø  ÿÀŽ È z

  ÿàó Ê $vÿ  ÿŽ  Ì �y

    › ¢ xÿ     â $x   ÿ Ž È� %v  * Ž â xø  ÿÀŽ Ê �y    › â z  ÿàó ä xÿ     Æ $vÿ  ÿŽ  ä %v  * Ž È �y    › è �y    › â r-   ÿàó Ê $x	   /Ž  Ê �y     â  xÿ   ÿÀŽ â$t7ÿÿ Ž â xÿ     â $t9ÿ @€ÿ Ž Æ 	s         d $x€   ÿ Ž Ê/ x€   ÿÀŽ ä xÿ   ÿÀŽ ä xÿ   xŽ ä xT      â	s6         b xÿ     â $xÿŽ â xÿ     Ä xÿ   ÿÀŽ ä x€€€€7ìŽ ä	s         " x€€€€9ìŽ ä xÿþþþÿàÿ ä xÿ   ÿÀŽ ä xøŽ â$x76€   ÿ Ž â/ xÿþþþÿàÿ â 	s8         b x˜º  ÿ    â $t6ÿ @€ÿ Ž â x @€øŽ Ä x7€   7xŽ â $x€   ÿ Ž â x˜º  ÿ    ä xÿ     â $x77ÿŽ â x€   xŽ Ê $xÿŽ ä $tÿÿ Ž ä $x88€   ÿ Ž Ê/ x8€   8xŽ Ê $x88ÿŽ â x
€€€€ÞŽ äO xÿÀŽ ä x€€€€ÞŽ äxÿÀŽ â $x   
Ž ä $tÿ @€ÿ Ž Ä $x   
Ž â x€€€€
´Ž â $t
ÿÿ Ž ä $tÿ @€ÿ Ž â x€€€€´Ž â &r,   ÿ   â x
7€€€€
ìŽ â$tÿÿ Ž â x7€€€€ìŽ â &r.      â x

ÿþþþÿàÿ Ä x€€€€ìŽ ä xÿþþþÿàÿ ä x€€€€6ìŽ ä x
7øŽ äxÿþþþÿàÿ ä x @€7øŽ Ä x
ÿþþþÿàÿ ä x˜º  ÿ    ä xøŽ äx˜º  ÿ    ä x

 @€øŽ ä xÿÀŽ Ä x8€€€€ìŽ äx8€€€€ìŽ ä x€€€€ÞŽ ä� xÿþþþÿàÿ ä xÿþþþÿàÿ â $x   
Ž â xÿÀŽ Ä x€€€€ÞŽ ä x8øŽ äx @€8øŽ â $x   
Ž â x€€€€´Ž ä x˜º  ÿ    ä x˜º  ÿ    â &r0      â xÿÀŽ Ä xÿÀŽ ä x€€€€´Ž ä x€€€€ÞŽ ä x€€€€ÞŽ â &r1      â x˜º  ÿ    â $x   
Ž â x
˜º  ÿ    â $x   
Ž â x
ÿÀŽ Ä x€€€€ÞŽ ä x€€€€´Ž ä x€€€€´Ž â $x   

Ž â xÿÀŽ â&r
2      â x€€€€ÞŽ ä x€€€€´Ž â &r3   
   Ä $x   
Ž ä &r4      â x   ÿüŽ ä x€€€€´Ž Ê &r5      È $r   ÿŽ Ê x
ÿ     È r

   ÿ� â 0rÿ    A   Æ x
ÿ   
  È Er 
       â  r
:     A  È #r!
   !   â #r#   #   â #r"   "   â G‰  À    € ê r-   ÿàñ Ê $x   /  Ê �y
    ¨ �y    è �y    ( �y    b $x
   ÿ Ž ÊO xø  ÿÀŽ È z  ÿàñ Ê $vÿ  ÿ  â xÿ   
  Ê �y    › ¢ %v
  * Ž Ì �y

    › ¢ $x   ÿ Ž Ê� xø  ÿÀŽ È z  ÿàñ ä xÿ     Æ $vÿ  ÿ  ä %v  * Ž È �y    › è �y    › â r-   ÿàñ Ê $x   /  Ê �y     â  xÿ   ÿÀŽ â$tÿÿ Ž â xÿ     â $tÿÿ Ž Æ 	s         b xÿ   ÿÀŽ â $tÿ @€ÿ Ž Ì 	s         " $x€   ÿ Ž Ê/ x€   ÿÀŽ â $x€   ÿ Ž Æ xÿ   xŽ ä xT      äx€   xŽ â $xÿŽ â xÿ     Æ $xÿŽ â x€€€€ìŽ äx€€€€ìŽ ä xÿþþþÿàÿ ä xÿ   ÿÀŽ ä xÿþþþÿàÿ ä xøŽ â 	s         " xÿ     Ä x @€øŽ ä x˜º  ÿ    ä x€€€€ìŽ ä xÿ   ÿÀŽ â $tÿ @€ÿ Ž â x˜º  ÿ    ä xÿþþþÿàÿ â 	s         b x€€€€ìŽ â $x€   ÿ Ž â xøŽ Ä xÿþþþÿàÿ ä x˜º  ÿ    ä x @€øŽ ä x€   xŽ ä x˜º  ÿ    ä xÿ     â $x€   ÿ Ž Ä/ $xÿŽ Æ x€   xŽ â $tÿ @€ÿ Ž È $xÿŽ Ê x€€€€ìŽ ä x€€€€ÞŽ äO xÿÀŽ ä x€€€€ÞŽ äxÿÀŽ â $x   
Ž È $x   
Ž â x€€€€´Ž ä xÿÀŽ ä x
€€€€ÞŽ â &r,   ÿ   â x€€€€´Ž ä xÿÀŽ â$x   
Ž â x€€€€ÞŽ â $tÿ @€ÿ Ž â xÿþþþÿàÿ â &r.      â x€€€€
´Ž â $t
ÿÿ Ž â x€€€€ìŽ â $x   
Ž â x @€øŽ â $tÿÿ Ž â x
€€€€
ìŽ â &r00      â xÿþþþÿàÿ Ä x€€€€ìŽ ä x

ÿþþþÿàÿ ä xÿþþþÿàÿ ä x

øŽ â 0rÿ    A   â� xøŽ ä x
˜º  ÿ    Ä x @€øŽ â  r:     A  â x
ÿÀŽ ä x˜º  ÿ    ä x€€€€ÞŽ ä x€€€€´Ž ä x˜º  ÿ    â $x   

Ž â xÿÀŽ â &r11   0   â x€€€€ÞŽ Ä x
ÿÀŽ ä x˜º  ÿ    â $x   
Ž â x€€€€ÞŽ ä x€€€€´Ž ä xÿÀŽ â$x   

Ž â x€€€€ÞŽ â &r22   1   â x€€€€´Ž Ä x€€€€´Ž â $x   
Ž â x   ÿüŽ â &r33   2   Æ x€€€€´Ž â &r44   3   È &r5   4   È $r   ÿŽ Ê x
ÿ     È r

   ÿ� È x
ÿ   
  È Er
 
       Ê #r 
       â x%%   ÿàÿ ä x--  ÿàó ä | %   p`ð ä x))  ÿàõ â $r/ÿÿ   /Ž  â x&&    ÿàÿ Æ $r+ÿÿ   + Ì D	       À â Gy  @Ûÿÿÿÿƒ ê Ay       € ê $r
ÿÿ   ! Ž ä $rÿÿ     Ž ä Ay        € ê z ÿ `  pPð È z ÿ a   Sð Ú G	  P    € ê ‰#      ( ‰"      b !r#       â !r"       È/ ‰      ( ‰      b !r       â !r       È/ ‰ €    ( ‰ €    b !r       â !r       È/ ‰ @    ( ‰	 @    b !r       â !r		       È/ ‰      ( ‰	      b !r       â !r	       â/ Gy  €    € ê ‰!      ( ‰#      h ‰"      ¨ ‰       â !r!       â !r#       È/ ‰      " !r"       ÆO ‰      b !r        Æ� ‰      ¨ ‰
      â !r       â !r       â/ !r	       ÈO ‰ €    " !r
       Æ� ‰ €    h ‰
	 €    ¨ ‰ €    â !r       â !r       È/ ‰ @    " !r
	
       ÆO ‰ @    b !r       Æ� ‰
 @    ¨ ‰ @    â !r       â !r       È/ ‰      " !r
       ÆO ‰      b !r	       Æ� ‰      ¨ ‰	      â !r       â !r
       â/ !r       âO !r	       â� y      !   ( y      %   $ $x   Ž Ê z  p  p`ð È x    pDp  Ú M	        € ê z ÿ ^  pPð â $z  v  Ž Æz ÿ _   Sð Ú $ÿ   ÿ Ž È % ^   Ž Ì �	     ¢ $x    ÿ Ž Ê r  ÿ   p"ò äx     p"ö äx     p"ø äx     p"ô Î $ ÿÿ    Ž â z ÿ `  pPò â $2 ÿÿ    Ž ä $B ÿÿ   
 Ž â z ÿ a  Sò â $" ÿÿ    Ž È !         ÐO G™       € ê z ÿ b  pPð È z ÿ c   Sð Ú $ÿ   ÿ Ž È % b   Ž Ì �	     ¢ $x   ÿ Ž Ê x øÿÿÿp"ò äx üÿÿÿp"ô är ÿ   p"ö äx    p"ø Î $ÿÿ    Ž ä $"ÿÿ    Ž ä $vÿ h  ÿ Ž ä $2ÿÿ   
 Ž ä $Bÿÿ    Ž â x    pBò Æ !       ÔO G  P    € ê x    pRð Ú G‰  �     € ê x    pRð Ú G	  ð    € ê  x;ª¸¿  A  Ì s        $ !t  €?    Ì s        $  r     A  È  r       A  â Gy  P    € ê  x'7=  A  â x*BL?  A  Æ #t  €?   â  x   ?  A  Æ  r     A  Ì s     $   $ !t  €?    È  r     A  È  r       A  â Gy  °    € ê x    pRð Ú G‰  à     € ê x    pRð Ú G	        € ê 	z i    � ä 	z   i    � Æ  x;ª¸¿  A  â 	z   i €  � Ê s        $ !t  €?    Ì s        $  r     A  È  r       A  â Gy  Ð     € ê  r       A  â Gy  °     € ê 	x  à@  � ä 	x    à@  � Æ  x#ÛÙ?  A  â 	x    àÀ  � Æ  x;ª¸¿  A  ä !t   €?    È s        $ !t  €?    Ì s        $  r      A  È  r       A  ä y      "   " $tÿ   ÿ Ž Æ y      &   b $z s  Ž È $z v  Ž È/ %v j   Ž Ê †y      â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y       "   " ¹z  l      â Ey  Ð    € â ™x?    â y      !   b ¹z  F   
   â x      ÿ  â r ÿ       â $r>ÿÿ   ÿ Ž â r% ÿ       â x
      ÿ  â r( ÿ       â $rBÿÿ   ÿ Ž â r	 ÿ       Ä rA ÿ       ä r ÿ       ä r ÿ       ä rL ÿ       ä r ÿ       ä r  ÿ   pRð Ä r  ÿ   pRò är    ÿ(Ž ä/ rN ÿ       ä xÿ     È r   ÿ(� È xÿ     È |    pbô Ú G)  0    € ê y      '   " r ÿ       â $rÿÿ   ÿ Ž â xàÿÿÿÿÀŽ â y      &   b $vÿ Y  ÿ Ž ä r  €ÿàÿ â y      %   ¢ r        ä xÿ     Ä x"   ÿ   ä r   ÿ � ä x1   ÿ   ä z " q  ÿàÿ â 'z y   Ž È 'z s   Ž â/ x   ÿ   ÆO $z w  ÿŽ â zÿ z    â $z }  ÿŽ â zÿ t    â $z p  Ž âx   ÿàÿ â $z |  Ž âxðÿÿÿÿÀŽ â $z |  Ž â xÿ     â $z v  Ž â r  €ÿàÿ â $z v  Ž â z  q      â $za p  Ž â r   ÿ� â %xfÒ   ÿŽ â xÿ     Ä xøÿÿÿÿÀŽ â $x   "Ž âr&   ÿ� â $x   "Ž âr  €ÿàÿ â $x   "Ž âx   ÿ   â $x   "Ž â r   ÿ� â %xd1   fŽ â r&&   ÿò� Ä xÿ     ä x'ÿ   &  ä r   ÿàõ â $x   "Ž â r   ÿàÿ â %x^aÒ   &Ž â r"   ÿŽ ä r$ÿÿ   ÿä â %x`aÒ   ÿŽ â x   ÿ   Ä xÿ     â $x   Ž â r&&f   ÿàõ â %x,   fŽ â x   ÿ   ä z  X      â %x.   `Ž â r ÿ       ä x     â%x01   `Ž â x   ÿ   Ä r!g'   ÿä ä xÿ     ä r*,   ÿà÷ är6d   ÿàõ är+-   ÿäÿ är7e   ÿä ä r8.   ÿàõ â�yV*€   • ¢  r:0   ÿà÷ Æ �yU*‚   • ¢  $x9/    Æ �yZ6    • è �yY6   • è �y28€   • h	 �y)8‚   • b	 r;1   ÿäÿ Ê �yX:    • h �yW:   • b	 r   "àµ ä x* $       ä r$   ÿd~ Æ %v6 Z  * Ž â/ r4&   ÿàù Æ $x$   ÿŽ â rD6   ÿàõ ä r5!   ÿä ä r77   ÿàÿ ä rG    àÇ â �y=4À   ‘ b rE7   ÿä ä r$   ÿ„þ Ä rJ^   ÿàû â�ySD   ™ b rHf   ÿàõ â %vFG Z  * Ž â rK_   ÿäÿ â �y:4Ä   ‘ $$x$   ÿŽ ä �y8DL   ™ " $xIg    Ä rGG   ÿàÿ â �y;JÀ   ‘ " r4F   ÿàõ Æ �y9JÄ   ‘ " r   àÇ Æ �yT6    • " r5G   ÿä Æ �yR6H   • " r$   ÿ„þ Æ �y#4   ™ " rm   àÇ â $x$   ÿŽ â rD`   ÿàõ â/ �yOHÐ   • " %v6 Z  * Ž Æ �y'4L   ™ " r$   ÿ„þ ä rEa   ÿä â �yPF    • " r77   ÿàÿ â %vlm Z  * Ž ä �yQFH   • " r46   ÿàõ â $x$   ÿŽ Ä �yDÐ   • " r57   ÿä â$xmm   Ž â rhl   ÿàõ â �y6    • ( �yI4   ™ "# rim   ÿä ä rk   àµ â �yM4L   ™ & r$   ÿd~ â �yh   ™ " %vjk Z  * Ž Æ �yK6H   • " $x$   ÿŽ â rbj   ÿàõ ä �yJl    • $ rkk   ÿàÿ ä �yGhL   ™ $ rck   ÿä Ä r7   àµ â�yDlH   • & r4$   ÿd~ â/ %v67 Z  * Ž â r]   àµ â �yCb   ™ $ $x44$   ÿŽ â r$   ÿd~ â �yHj    • " rl6   ÿàõ Ä r574   ÿàÿ â %v\] Z  * Ž â �yFbL   ™ $ rm5   ÿä â $x$   ÿŽ â rh\   ÿàõ â �yEjH   • " r+   àÇ ä r]]   ÿàÿ â �y?l   ™ " $r4ÿÿ   6 Ž Ä ri]   ÿä â �y@lL   ™ " r$   ÿ„þ â %vb+ Z  * Ž ä/ �y<4    • $ $x$   ÿŽ ä �y6h   ™ " rjb   ÿàõ Æ �y4H   • " rcc   ÿàÿ Æ �y\    • " rkc   ÿä Æ �y+hL   ™ ( �y*j   ™ ( �y5jL   ™ (/ �y4b    • ( �y3\H   • ( �y7bH   • " xVVT  U    ÈO rVÿ   V  ä xYZT  Y    ä� xUV0000ÿÀŽ ä xZÿ   Y  ä x22T  )    äxUUZøŽ Ä xZV   ÿ   ä r)ÿ   2  ä xZZ0000ÿÀŽ Æ $x2)   ÿ Ž â xZZYøŽ ä xWXT  W    ä x\U€€€€ÿüŽ ä x220000ÿÀŽ ä x]Z€€€€ÿüŽ ä xVÿ   W  Ä x))0000ÿÀŽ ä x\\àßßßÿàÿ ä x22WøŽ ä x]]àßßßÿàÿ ä x))VøŽ ä x[U    \Ž Ä xW2€€€€ÿüŽ ä xYZ    ]Ž ä xX)€€€€ÿüŽ ä x[[˜º  ÿ    ä xWWàßßßÿàÿ ä xVU˜º  ÿ    Ä xYY˜º  ÿ    ä xZZ˜º  ÿ    ä xXXàßßßÿàÿ ä xU[€€€€\Ž äx\2    WŽ ä xV[V`Ž Ä x]Y€€€€]Ž äxZYZ`Ž ä x[)    XŽ äxb)˜º  ÿ    ä/ x\\˜º  ÿ    ä x)2˜º  ÿ    Ä rZZ]   ÿüŽ ä x[[˜º  ÿ    ä xW\€€€€WŽ â &r2ZS   ÿ   â x\\)`Ž ä rVVU   ÿüŽ ä x==€ˆ  ÿ    Ä xX[€€€€XŽ äx[[b`Ž â &r)V8   ÿ   âr\\W   ÿüŽ â $r2=2   ÿŽ â x::€ˆ  ÿ    ä rX[X   ÿüŽ â &rb\S   ÿ   â x;;€ˆ  ÿ    â $r):)   ÿŽ â x99€ˆ  ÿ    â 0rTÿT    A   â ErS 2       â &r8X8   ÿ   â Er) )       â $rb;b   ÿŽ Ä 0rRÿR    A   â #rSTS   ÿ   â $r898   ÿŽ â ErU b       â &r2Z#   ÿ   â #r)R)   S   â 0rSÿO    A   â ErW 8       â &rOV'   ÿ   ä &r8\#   ÿ   Ä $r2=2   ÿŽ â #rTTU   ÿ   â $r8;8   ÿŽ ä $rO:O   ÿŽ â Er# 2       â 0rPÿP    A   â #rTRW   T   â &rRX'   ÿ   â Er' 8       â 0rQÿQ    A   â Er2 O       â #r8P#   ÿ   â $rR9R   ÿŽ Ä 0r#ÿ    A   â #rOQ2   8   â &rZI   ÿ   âErR R       â &r2\I   ÿ   â #r'P'   ÿ   â &r8XM   ÿ   ä $r=   ÿŽ ä $rI;2   ÿŽ â #rRQR   '   â $r'98   ÿŽ â Er8        â 0rÿ    A   â Er I       â &r2Z   ÿ   Ä &r\   ÿ   â #rNS)   N   â 0rKÿK    A   â ErP '       â #r88   ÿ   â $r2=2   ÿŽ â #r   ÿ   â &r)VM   ÿ   ä $r;   ÿŽ â #rPKP      â $r):)   ÿŽ â Er' 2       â 0rJÿJ    A   â Er        â #r#T      â ErT )       Ä #rJ'   ÿ   â#rJJ   ÿ   â &rVG   ÿ   â#r8KT   8   â &rXG   ÿ   ä $r:   ÿŽ â #rS8      â &r2ZC   ÿ   ä&r8\C   ÿ   Ä $r)9   ÿŽ â Er        â 0r'ÿD    A   ä $r2=2   ÿŽ ä $r8;8   ÿŽ â #r'      â 0rHÿH    A   â ErC 2       ä Er 8       â #rAS   A   â Er) )       â &rZ?   ÿ   â #r2HC   ÿ   â &r\?   ÿ   â #rHH   ÿ   â #rJ')   J   â &rV@   ÿ   Ä $r=   ÿŽ ä &r@X@   ÿ   ä $r);   ÿŽ â Er'        â $r:   ÿŽ ä $r@9@   ÿŽ â Er )       â 0r<ÿ<    A   Ä &rZ6   ÿ   ä&r6\6   ÿ   â Er        â 0rÿ    A   â Er@ @       â #r'<'   ÿ   â#r<   ÿ   â $r=   ÿŽ ä $r6;6   ÿŽ â #r   '   â#r@      â 0rÿ    A   â Er        Ä Er 6       â #r>S   >   â &rZZ*   ÿ   â #r   ÿ   ä#r   ÿ   â &rV+   ÿ   â #r(#   (   â &r*\*   ÿ   ä &rGVF   ÿ   â x   ÿàÿ â &rV5   ÿ   Ä &rFXF   ÿ   ä$rZ=Z   ÿŽ ä &rX+   ÿ   ä $r:   ÿŽ ä $r*;*   ÿŽ ä &rXX5   ÿ   Ä $rG:G   ÿŽ â |    pbô â $r::   ÿŽ â Er' Z       â $r9   ÿŽ âEr        â $rF9F   ÿŽ âEr *       â $rX9X   ÿŽ ä 0r4ÿ4    A   â ErG G       â 0rEÿE    A   â ErI F       â 0r3ÿ3    A   â Er        â 0r7ÿ7    A   â Er: :       â #r*4'   ÿ   âEr) X       â #r44   ÿ   â #rGEG   2   â#rHEI   H   â #r3      â#r3      â #r*7:   *   â#r)7)   4   â x¤  ÿà÷ â #rLSO   L   â #rB#R   B   â#r#P      â#r#J      â #r
SG   
   â #r	#H   	   â #r%S   %   â #r#      â #rS*      â #r#)      â x   ÿàÿ Ä x   ÿàÿ ä x   ÿàÿ ä x   ÿàÿ ä x   ÿàÿ ä x   ÿàÿ ä x     ÿàÿ Ä x""   ÿàÿ ä rÿ   ÿäÿ â D)       À â Gy   êÿÿÿÿƒ ê Ay        € ê Ey  0    € â G‰      € ê r    ÿHŽ Ê ˆs N øÿ H   è ˆs €øÿ H   è ˆs L ùÿ H   è ˆs B€ùÿ H   è ˆs  úÿ H   è ˆs €úÿ H   è ˆs A ûÿ H   è ˆs €ûÿ H   è ˆs 
 üÿ H   è ˆs 	€üÿ H   è ˆs > ýÿ H   è ˆs (€ýÿ H   è ˆs % þÿ H   è ˆs €þÿ H   è ˆs  ÿÿ H   è ˆs €ÿÿ H   ä Ay        € ê {          ì M        € ê „y     H   " z9  r      Æ „y €   H   d $x59   ÿŽ ä „y    H   ¨ „y €  H   è „y    H   è „y €  H   ( y      %   h „y    H   ( y      &   è y      '   " !rN       Æ „y     H   â !r       Æ/ ‰       " x   ÿ   Æ‰      b !rL       âO z  r  p`ð ä „y €  H   â !r       Æ� „y    H   ¢ r ÿ   p"p â !rB       ä‰      è „y €  H   â !rAA       â !r         â !r       â „y#    H   â !r       Æ/ ‰       ( ‰      b !r% %       Æ ‰      è „y) €  H   è „y' €  H   è „y"    H   â !r
       âO !r        Ä ‰      " $z  x  Ž â x   ÿàÿ â !r       â/ ‰      b $z  ~   Ž â !r!	       â !r       â� ‰ €    ¤ 
  r  ÿàó â !r##>       â ‰A      â r9    ÿˆ Ä ÿÿ   ÿäÿ  â ‰       " 
 j  ÿ„ â $rÿÿ   ÿ â z  r  p`ò â „y7 €  H   â 
 k   â !r       â !r       â ‰ €    b !r       â r	9    ÿ„ â !r))       â !r       â/ ‰      â x    p"ò â !r''(       â !r       âO ‰      "  5   ÿàû â !r+"       â !rAA       â� ‰ @    b $x9   ÿŽ â ÿÿ   ÿäÿ â !r         â‰$      ¢ 
 j  ÿŠ Ä     ÿà÷ â ‰&      â rÿÿ   ÿä â !r       ä‰(!      " z	 j  ÿ„ ä 
 k  � â ‰*#      b ÿÿ   ÿäÿ Ä 

 j  ÿ† â ‰0)      b !r       â $x9   ÿŽ â z		 k   â ‰,'      " !r       â 
 k  � â $x9   ÿŽ â ‰.%      " !r       â/     ÿà÷ â!r$$       âO ‰      b 5   ÿàÿ â !r&&       â� z j  ÿˆ â ‰"A      ¢ r    ÿàõ â !r(!(       â4ÿÿ   ÿäÿ â ‰2+      â 5    ÿà÷ â!r*#*       âz k   â ‰ @    " !r0)0       â 9   ÿàÿ â !r77       â !r,',       â r6ÿÿ   ÿä â ‰'$      " z j  ÿ„ â !r.%.       â 8ÿÿ   ÿäÿ â ‰)&      b 9    ÿàû â !r       â/ 5 j  ÿ† â ‰-*      b z k  6 â !r"A"       âO 5 k  8� â ‰30      ¢    ÿàÿ â !r2+2       â� 
 j  ÿˆ â ‰+(      â     ÿàõ â !r       â$:ÿÿ   ÿŽ â ‰ €    " 9 j  ÿŠ Ä 
 k  4 â ‰/,      " !r'$'       â 9 k  :� ä ‰1.      " !r)&)       Æ‰ €    b !r-*-       Æ/ ‰#  €    b !r303       âO 0ÿÿ   ÿä ä ‰%" €    ¢ !r+(+       â�  j  ÿ„ ä‰52      â !r       â k  0 Ä ‰7      " !r/,/       Æ ‰! €    " !r1.1       Æ ‰      " !r       â!r# #       ä/ ‰$' €    b !r%"%       ÆO ‰&) €    ¢ !r525       Æ� ‰(+ €    â !r77       Æ‰ @    " !r!!       Æ ‰*- €    b !r9       Æ ‰,/ €    "         Æ ‰ @    " !r$'$       â/ %  j   Ž ä ‰ # @    b !r&)&       ÆO ‰.1 €    ¢ !r(+(       Æ� ‰03 €    â !r       Æ‰"% @    " !r*-*       Æ‰25 €    b !r,/,       Æ ‰47      " !r       Æ ‰      " !r #        Æ/ ‰! @    b !r.1.       âO !r030       ä� ‰#$ @    ¢ !r"%"       Æ‰'& @    â !r252       Æ‰%( @    " !r474       Æ ‰)* @    " !r;       Æ ‰+, @    b !r!       Æ/ ‰      h ‰      b !r#$#       ÆO ‰-. @    ¢ !r'&'       Æ� ‰/0 @    â !r%(%       Æ‰34 €    " !r)*)       Æ ‰12 @    " !r+,+       Æ‰      b !r5       Æ/ ‰!       b !r       Æ †	 9    â !r-.-       ÆO ‰"      ¢ !r/0/       Æ� ‰#      â !r343       â        â !r121       â ‰'      " !r       ä%  j   Ž â ‰)      " !r! !       Æ/ † ;   è ‰%      b !r"       ÆO †	 5    â !r##       Æ� ‰+      ¨ †     â !r''       Æ ‰-      " !r))       Æ‰ /      è ‰$3 @    " !r%%       Æ/ ‰1      h †	     â !r++       ÆO † !   è †	     â !r--       â !r//        ä� † #    â !r$3$       Æ†	 '    â !r11       Æ/ † %   è †	 
)    è † +    è †	 -    è † /   è ‰$      ( †	 1    â M™        € ê x   ÿàÿ â !r$       Æ_ r    ÿàñ È r ÿÿ   ÿä  ä z j  ÿ€ È z k     Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y       "   " ¹z  l      â Ey  À    € â ™x?    â y      !   b ¹z  F   
   â x      ÿ  â r ÿ       â y      %   ¢ $r$ÿÿ   ÿ Ž â r& ÿ       â x
      ÿ  â y+      &   â rB ÿ       â $rEÿÿ   ÿ Ž â r	 ÿ       â $rMÿÿ   ÿ Ž â r ÿ       Ä r ÿ       ä rF ÿ       ä r  ÿ   pRð är  ÿ   pRò är    ÿ(Ž È/ xÿ     ä x   ÿ   äO r   ÿ(� È xÿ     È |    pbô Ú G)      € ê y      '   "Ž xàÿÿÿÿÀŽ â $rÿÿ   ÿ Ž â r* ÿ       â $z+ w  ÿŽ âr  €ÿàÿ ä r        â 'z+ s  * Ž â xÿ     ä z  q      â $z p  Ž â r   ÿ � Ä zÿ t    ä r   ÿ� ä x   ÿ   ä xøÿÿÿÿÀŽ ä z q  ÿàÿ â 'z y   Ž È $z }  ÿŽ â zÿ z    â $x   Ž âxðÿÿÿÿÀŽ â $x   Ž â xÿ     â $z |  Ž â r  €ÿàÿ â $z |  Ž â x   ÿàÿ â $z v  Ž âxÿ     â $z` v  Ž â r   ÿàõ â $x   
Ž â r"   ÿ� â$z/ p  Ž â r   ÿ� â %x``Ò   ÿŽ â x   ÿ   Ä x!ÿ     â $x   Ž âr""   ÿò� â $x   Ž â r   ÿàÿ â $x   ÿ Ž â x#ÿ   "  â$vÿ X  ÿ Ž â r   ÿŽ ä x   ÿ   â %x,/Ò   "Ž â xÿ     Ä rÿÿ   ÿä â %x./Ò   ÿŽ â x   ÿ   ä r "`   ÿàõ â %x^   `Ž â r!   ÿŽ ä x     â %x0   `Ž â z  Y      Ä r ÿ       â %x2   .Ž âx!!   ÿ   ä r"a#   ÿä â %x4   .Ž â xÿ     Æ r:0   ÿà÷ är(^   ÿàõ Æ $x;1   Ž â r)_   ÿä ä r82   ÿàõ ä�yV:€   • ¢  r<4   ÿà÷ Æ �yU:‚   • ¢  r93   ÿä Æ �yY(    • è �yZ(   • â r=5   ÿäÿ Æ �y'8€   • h	 �y68‚   • h	 �yX<    • h �yW<   • b r   àµ â $t\ÿ$   ÿ Ž Æ r   ÿd~ â %v( Z  \ Ž â/ rH    ÿàù Æ $x$   ÿŽ â r:,   ÿàû ä rI"   ÿä ä r8(   ÿàõ är))   ÿàÿ â �yAHÀ   ‘ " r;-   ÿäÿ Æ �y>HÄ   ‘ " $x9)    â r   àÇ ä �y?:À   ‘ ( �y=:Ä   ‘ " r   ÿ„þ Æ �yS8   ™ " rJ`   ÿàõ â %vH Z  \ Ž ä �y<8L   ™ $ $x$   ÿŽ â rKa   ÿä â �yT(    • " r:H   ÿàõ ä/ rII   ÿàÿ â �y%(H   • " r   àÇ Ä r;I   ÿä â �yPJÐ   • " r   ÿ„þ â %v8 Z  \ Ž ä �yQ:   ™ " re   àÙ â $x$   ÿŽ ä �y:L   ™ $ $x99   Ž â r   ÿ¤~ â �y#H    • " rb8   ÿà÷ â %vde Z  \ Ž â r(.   ÿàõ â/ �yRHH   • " rc9   ÿäÿ â $x$   ÿŽ â r)/   ÿä â �yN8    • " r:d   ÿàõ Ä ree   ÿàÿ â �yOb   ™ ( �y(Ð   • " $x;e    Æ �yH8H   • (# �yKbL   ™ " r)   àµ Æ �yG:   ™ " r   ÿd~ â %v() Z  \ Ž ä �yJ:L   ™ " r   àµ â $x7$   ÿŽ ä �yLd    • " r   ÿd~ ä r8(   ÿàõ â/ �yIdH   • " r))7   ÿàÿ â %vb Z  \ Ž Æ r9)   ÿä â $x$   ÿŽ â r]   àÇ â �y@(    • " rfb   ÿàõ ä rcc   ÿàÿ â �yC8   ™ " r   ÿ„þ Æ �yD8L   ™ " $xgc    ä %v\] Z  \ Ž â �y(H   • & $x$   ÿŽ â �y:f   ™ " rd\   ÿàõ Æ/ r]]   ÿàÿ â �y7bH   • & re]   ÿä â �yb    • ( �y(d   ™ ( �y)fL   ™ ( �y9dL   ™ ( �y8\    • ( �y;\H   • " xVVT  U    ÈO rVÿ!   V  ä xYYT  Z    ä� xUV0000ÿÀŽ ä xZÿ   Y  ä x6'T  6    äxUUZøŽ Ä xZV   ÿ   ä rVÿ!   6  ä xWXT  W    ä xZZ0000ÿÀŽ ä x6V   ÿ   ä x\U€€€€ÿüŽ Ä x'ÿ   W  ä xVV0000ÿÀŽ ä xZZYøŽ ä x660000ÿÀŽ ä x\\àßßßÿàÿ ä x'V'øŽ Ä x]Z€€€€ÿüŽ ä x66WøŽ ä x[U    \Ž ä x]]àßßßÿàÿ ä xX'€€€€ÿüŽ ä xW6€€€€ÿüŽ Ä xVU˜º  ÿ    ä x[[˜º  ÿ    ä xYZ    ]Ž ä xXXàßßßÿàÿ ä xWWàßßßÿàÿ ä xU[€€€€\Ž Ä xV[V`Ž ä xYY˜º  ÿ    ä xZZ˜º  ÿ    ä x['    XŽ ä x\6    WŽ ä x]Y€€€€]Ž Ä xZYZ`Ž ä x[[˜º  ÿ    ä x''˜º  ÿ    ä x\\˜º  ÿ    ä xY6˜º  ÿ    ä xX[€€€€XŽ Ä x'['`Ž ä xW\€€€€WŽ äx\\Y`Ž ä rZZ]   ÿüŽ ä rX'X   ÿüŽ ä xAA€ˆ  ÿ    Äx>>€ˆ  ÿ    ä r\\W   ÿüŽ ä x'?€ˆ  ÿ    ä x6=€ˆ  ÿ    â $r?ÿÿ   A Ž â rVVU   ÿüŽ â &rbZS   ÿ   âr= >       â &rS\S   ÿ   â r> '       â &rdV<   ÿ   Ä/ $rb?b   ÿŽ ä &r'X<   ÿ   ä $rS>S   ÿŽ â ErA b       â $rd=d   ÿŽ ä 0rTÿT    A   ä $r<6'   ÿŽ â Er' S       â 0rSÿ%    A   â Erd d       â #rbTA   ÿ   âEr< <       â #r'T'   ÿ   Ä #rbSd   b   ä#rSS<   '   â &r<ZQ   ÿ   ä&rQ\Q   ÿ   ä 0r%ÿP    A   ä &rPV   ÿ   ä $r<?<   ÿŽ Ä $rQ>Q   ÿŽ ä $r'=P   ÿŽ â ErP <       â 0r#ÿ#    A   â Er< Q       â &rAX   ÿ   â Er' '       â 0rRÿR    A   ä $rA6A   ÿŽ â #r#<   ÿ   â&r<ZO   ÿ   â#rQ#P   ÿ   â &rO\O   ÿ   â ErA A       â $r<?<   ÿŽ â #r'R'   Q   â $rO>O   ÿŽ Ä #rRRA      â 0rNÿN    A   â ErA <       â #r%'      â 0r'ÿH    A   ä &r<VK   ÿ   â Er O       â &rHZG   ÿ   ä 0r#ÿ    A   â #rNA   ÿ   â &rG\G   ÿ   Ä $r<=<   ÿŽ ä &rPVJ   ÿ   ä&rAXJ   ÿ   ä $rH?H   ÿŽ â #rNN   ÿ   â $r>G   ÿŽ â Er< <       â $rP=P   ÿŽ Ä $rG6A   ÿŽ â ErA H       â 0rLÿL    A   â Er        â 0rIÿI    A   â ErP P       â &rKXK   ÿ   â #r'<      â #rHLA   ÿ   â#rLL   ÿ   â $rK6K   ÿŽ â #rE%   E   â #rIP   H   â &rZC   ÿ   Ä &rC\C   ÿ   â ErK K       â #r
%   
   â &r<VD   ÿ   ä $r?   ÿŽ ä &rAXD   ÿ   ä $r>C   ÿŽ â #rN'K   N   â $r<=<   ÿŽ â Er'        â $rA6A   ÿŽ â Er        â 0r@ÿ@    A   Ä &rC\:   ÿ   â Er< <       â &rDZ:   ÿ   â Er A       â 0rÿ    A   â #r'@'   ÿ   â#r@   ÿ   â $rC>C   ÿŽ ä $rD?D   ÿŽ â #r'<   '   â#r      â 0rÿ    A   â Er C       â &rZZ(   ÿ   â ErD D       â #r&#   &   â &r\(   ÿ   â #r   ÿ   â&rV)   ÿ   â #rDD   ÿ   â $rZ?Z   ÿŽ â x   ÿàÿ â &rVV9   ÿ   Ä $r>   ÿŽ ä &r)X)   ÿ   ä&r9X9   ÿ   â |    pbô â $r=   ÿŽ âEr Z       â $rV=V   ÿŽ â Er        â $r)6)   ÿŽ Ä $r969   ÿŽ ä 0r8ÿ8    A   â ErG G       â 0r7ÿ7    A   â Er        â 0r;ÿ;    A   â Er) )       â #r(8   ÿ   âErV V       â #r88   ÿ   â Er9 9       â #rLIG   L   â #r7   D   â#r7)      â #r(;V   (   â#r9;9   8   â x¤  ÿà÷ â #rM%b   M   â #rF#S   F   â#r#R      â#r#N      â#r	#L   	   â #rB%'   B   â#r$%   $   â #r#      â #r%(      â #r#9      â x   ÿàÿ Ä x   ÿàÿ ä x   ÿàÿ ä x   ÿàÿ ä x   ÿàÿ ä x   ÿàÿ ä x   ÿàÿ Ä rÿ   ÿäÿ â D)       À â Gy   ìÿÿÿÿƒ ê Ay        € ê� Ey      € â G‰  ð     € ê $x À  Ž Ê ˆs M ùÿ H   è ˆs F€ùÿ H   è ˆs  úÿ H   è ˆs €úÿ H   è ˆs E ûÿ H   è ˆs €ûÿ H   è ˆs 
 üÿ H   è ˆs 	€üÿ H   è ˆs B ýÿ H   è ˆs &€ýÿ H   è ˆs $ þÿ H   è ˆs €þÿ H   è ˆs  ÿÿ H   è ˆs €ÿÿ H   ä Ay        € ê {          ì M        € ê „y      H   b z  r  p`ð â $v-ÿ r  ÿ Ž Ä „y €   H   ¢ r ÿ   p"p Æ „y    H   è „y €  H   ( „y    H   h „y €  H   ( „y    H   ( „y €  H   ( „y    H   è „y €  H   â !r  M       Æ/ „y    H   â !rF       ÆO ‰       " !r       Æ� „y €  H   â !r       Æ„y    H   â !r
       Æ„y! €  H   â !r       Æ ‰
      b !rEE       Æ y      '   ¢ !r	       â !r        ä ‰      " $z + x  Ž â !rBB       â $x+-   ÿŽ â ‰      â !r&&       Æ ‰      " !r$$       Æ ‰      b !r

       Æ/ ‰      b $z  ~   Ž ÆO ‰E      ¢ !r       â !r       â !r       â ‰      " 
  r  ÿàõ â !r!       â !r       â� x   ÿàÿ â ‰&      â !r       Äz  r  p`ò â ‰B      " !r       äx    p"ò â ‰ €    b !r       â/ ÿÿ   ÿä ä ‰!$      b 
 j  ÿ„ â !rE       ÄO ‰
      ¢ 
 k   â !r       ä ‰#      ( ‰%      " !r&&       Æ� ‰      â !rBB       Æ‰      " !r       Æ‰      b !r$$!       Æ/ ‰)      b !r
       âO r-    ÿ„ ä ‰      ¢ !r"#       Æ ‰	      " !r'%       â !r       ä� ‰       â $x-   ÿŽ â !r       ä‰B      " r-    ÿˆ â !r       Ä‰!&      b     ÿà÷ â !r))       ä/ ‰ @    b r(ÿÿ   ÿä â !r       äO ‰#$      ¢ rÿÿ   ÿä â !r	       Ä ‰ €    " 	 +   ÿàû ä ÿÿ   ÿäÿ â ‰%"      " 
 j  ÿ† â !r         â� z
 j  ÿˆ â ‰*'      â z j  ÿ„ â !rBB       âÿÿ   ÿäÿ â ‰ €    " 
	 j  ÿŠ â !r&&!       â
 k  � â ‰ €    b z k  ( â !r       â/ z k   â ‰,)      b !r$$#       âO -   ÿàÿ Ä ‰ €    ¢ !r       â 
		 k  � ä ‰ €    " !r("%       â +   ÿàÿ ä ‰ €    " !r*'*       â� /    ÿà÷ â !r       â‰#  €    ä 0ÿÿ   ÿäÿ â !r       â‰%B €    " $x-   ÿŽ â / j  ÿ† â !r,),       â/ ‰'& €    b .    ÿàù â !r       ÄO ‰)$ €    ¢ / k  0� â !r       ä ‰ @    " r    ÿàõ â !r       ä ‰+( €    b 1ÿÿ   ÿä â $r"ÿÿ   ÿ â z j  ÿ„ â ‰-* €    b !r# #       â� . j  ÿˆ Ä ‰! @    â !r%B%       âz k  " ä ‰ @    " !r'&'       â/ . k  1 ä ‰ @    b !r)$)       âO    ÿàÿ Ä ‰/, €    ¢ !r       â     ÿàõ ä ‰ @    " !r+(+       Æ‰      b !r-*-       Æ ‰ @    b !r!!       Æ� ‰ # @    â !r       Æ‰% @    " !r       Æ/ ‰"' @    b !r/,/       âO ,ÿÿ   ÿä ä ‰$) @    ¢ !r       Æ ‰      " !r1       âx        ä ‰&+ @    b !r       â  j  ÿ„ â %v  j   Ž â ‰(- @    b !r #        â�  k  , Ä ‰      â !r%       Æ‰*/ @    " !r"'"       Æ/ ‰      b !r$)$       ÆO ‰       ¢ !r3       Æ ‰      " !r&+&       Æ‰!      b !r(-(       Æ ‰       b !r)       Æ� ‰#      â !r*/*       Æ‰%"      " !r+       Æ/ ‰$      b !r        ÆO †	 1    â !r       Æ † 3   â !r!!       Æ‰'&      " !r        Æ ‰(      ¢ !r##       Æ� ‰*      â !r%"%       Æ†	 )    â	 !r$       Æ/ † +    è	 †	 
    è	 † 
   â	 !r'&'       Æ †	 !    â	 !r(       ÆO †     â	 !r*       Æ� †	 #    è	 † %   è	 †	     è	 † '    è	 †	     â	 M™        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y       "   " ¹z  l      â Ey  `    € â ™x?    â y      !   b ¹z  F   
   â r ÿ       â $rÿÿ   ÿ Ž â y      %   ¢ r" ÿ       â $r>ÿÿ   ÿ Ž â r ÿ       â y%      &   â rK ÿ       â $rLÿÿ   ÿ Ž â x
      ÿ  â r ÿ       â $rÿÿ   ÿ Ž â rP ÿ       Ä r  ÿ   pRò är  ÿ   pRð â$x     Ž â/ x   ÿ   ÈO xÿ     È r   ÿ(� È xÿ     È |    pbô Ú G)  À    € ê y	      '   "Ž xàÿÿÿÿÀŽ â $r$ÿÿ   ÿ Ž â r ÿ       â $z p  Ž â r  €ÿàÿ â $z% w  ÿŽ â x   ÿ   ä xÿ     ä z  X      Ä r   ÿ � är   ÿ� ä xÿ     ä z q  ÿàÿ â 'z	 y   Ž È 'z% s  $ Ž â zÿ z    ä xðÿÿÿÿÀŽ â $z	 }  ÿŽ â zÿ t    â $z |  Ž â r	  €ÿàÿ â $z |  Ž â x   ÿàÿ â $z v  Ž âxÿ   	  â $z& v  Ž â x   ÿ   â $z+ p  Ž â r	   ÿ� â%x&&Ò   ÿŽ â r	   ÿ� Ä xøÿÿÿÿÀŽ â $v	ÿ q  ÿ Ž â r   ÿò� â $rÿÿ   ÿ Ž â r   ÿàõ â $x	   Ž âr   €ÿàÿ â $x	   Ž âr	   ÿŽ â$x	   Ž â x	ÿ     â $x   ÿ Ž â r    ÿàÿ Ä xÿ      â %x(+Ò   Ž âx4ÿ     ä rÿÿ   ÿä â %x*+Ò   ÿŽ â x      äx    ÿ   â $x   ÿ Ž â r &   ÿàõ â %x,   &Ž â r4   ÿŽ Ä z  Y      â %x.   &Ž â r'	   ÿä ä xÿ     â %x0   *Ž È %x2   *Ž È $x4   ÿ Ž ä r.   ÿà÷ är8,   ÿàõ är	/   ÿäÿ är62   ÿà÷ â$x9-    â r40   ÿàõ â �y?€   • ¨  �yB‚   • ¢  r51   ÿä Ä r73   ÿäÿ â �yD8    • è �y8   • è �yN4€   • h	 �yA4‚   • h	 �yC6    • h �yH6   • b rW   àÇ â $t:ÿ$   ÿ Ž Æ r	   ÿ„þ â %vVW Z  : Ž â r<    ÿàõ Æ $x	$   ÿŽ â r=   ÿä ä r(   ÿà÷ â $xWW   Ž â r8V   ÿàõ â/ �yJ<À   ‘ b r	)   ÿäÿ ä r9W   ÿä â �yG<Ä   ‘ b r7   àÇ Æ �yE8   ™ b r   ÿ„þ ä r4&   ÿàõ â�yIÀ   ‘ b	 %v67 Z  : Ž Æ �yFÄ   ‘ b	 $x$   ÿŽ Æ �y@8L   ™ b r5'   ÿä â $x77   Ž â r!   àµ â �yOV    • b rT6   ÿàù ä r   ÿd~ â�y	VH   • b	 rU7   ÿä â %v8! Z  : Ž Ä/ �y4Ð   • d $x;$   ÿŽ ä �yQT   ™ b	 r<*   ÿàù â $x99   ;Ž ä �yRTL   ™ b	 r48   ÿàõ ä/ r=+   ÿä â �y#6    • b r59   ÿä Ä r;   àµ â �y<Ð   • h �y!6H   • h �yS4   ™ b r<   ÿd~ Æ/ �yT8    • "%v6; Z  : Ž â r;   àÇ ä �yM8H   • " $x=<$   ÿŽ â rV   ÿ„þ ä �yU4L   ™ " r<6   ÿàõ ä r77=   ÿàÿ â %v8; Z  : Ž È/ �yX6    • " $x;V$   ÿŽ ä $x=7    â �yY6H   • $ r99;   ÿàÿ ä r;   àÇ â �yV<   ™ & r[   ÿ„þ â �yW<L   ™ " r48   ÿàõ â %v:; Z  : Ž Ä �y\8    • $ $x][$   ÿŽ â r59   ÿä â�y68H   • $! $x;;   ]Ž â r<:   ÿàõ â �yZ4   ™ & r=;   ÿä â �y[4L   ™ ( �y7<   ™ ( �y]<L   ™ ( �y5:    • (# �y::H   • "# x   ÿàÿ Ä xB?T  B    ÈO rBÿ   B  ä xDT      ä� x?B0000ÿÀŽ ä x4ÿ     ä xNNT  A    äx4?4øŽ Ä rNÿ   N  ä xBB   ÿ   ä xCCT  H    â $x8N   ÿ Ž â xBB0000ÿÀŽ ä x94€€€€ÿüŽ ä x;ÿ   C  Ä/ xNN0000ÿÀŽ ä x=BøŽ ä x99àßßßÿàÿ ä x880000ÿÀŽ ä xN;øŽ ä x;=€€€€ÿüŽ Ä x<4    9Ž ä x88CøŽ ä xA4˜º  ÿ    ä x4;àßßßÿàÿ ä xB<˜º  ÿ    ä xC8€€€€ÿüŽ Ä xD€€€€ÿüŽ ä x?=    4Ž ä x<B€€€€9Ž äx9Càßßßÿàÿ ä x;BA`Ž ä xBDàßßßÿàÿ Ä xA?˜º  ÿ    ä xD=˜º  ÿ    ä x?8    9Ž ä x=AD`Ž ä xC    BŽ äxH˜º  ÿ    Ä xD?˜º  ÿ    ä x8˜º  ÿ    ä x4A€€€€4Ž ä xCC˜º  ÿ    ä x9D€€€€9Ž äxDD`Ž Ä r8=4   ÿüŽ ä xBC€€€€BŽ äxCCH`Ž ä rDD9   ÿüŽ â &r48E   ÿ   âr<;<   ÿüŽ ä x?J€ˆ  ÿ    â &rHDE   ÿ   â rBCB   ÿüŽ Ä x=I€ˆ  ÿ    â $r4?4   ÿŽ â x;G€ˆ  ÿ    ä x9F€ˆ  ÿ    â &rF<@   ÿ   ä&rJB@   ÿ   ä $rH=H   ÿŽ â Er@ 4       â $rF;F   ÿŽ Ä 0rOÿO    A   â Er4 H       â $rJ9J   ÿŽ â ErF F       â 0r	ÿ	    A   â #r@O@   ÿ   â#rHO4   ÿ   â Er J       ä #r4	F   @   â &r@8Q   ÿ   Ä #r	   H   â &rHDQ   ÿ   ä &rF<R   ÿ   ä $r@?@   ÿŽ ä $rH=H   ÿŽ ä $rF;F   ÿŽ â Er@ @       â 0r#ÿ#    A   Ä 0r	ÿ    A   â Er H       â 0rÿ    A   â ErF F       â 0r!ÿ!    A   â #r@#@   ÿ   â &rRBR   ÿ   â #r4      â #r#   ÿ   â #r#!F   @   â &r48S   ÿ   Ä $rR9R   ÿŽ ä &rFDS   ÿ   â #r	      â &r@<U   ÿ   â	ErR R       â $r4?4   ÿŽ ä &rHBU   ÿ   ä $r=F   ÿŽ â ErA 4       â $r@;@   ÿŽ â #r!R      â 0rTÿT    A   â Er        â $rH9H   ÿŽ â Er@ @       â 0rMÿM    A   â #r4TA   ÿ   â#r	      â Er! H       â #rTT   ÿ   â &r8V   ÿ   â #rM@   4   â&r4<W   ÿ   â #r!M!   T   â $r?   ÿŽ Ä $r4;4   ÿŽ â #rL	!   L   â &rVDV   ÿ   â Er!        â 0rXÿX    A   â #r
   
   â $r=V   ÿŽ â Er4 4       â 0rYÿY    A   â #rX!   ÿ   â &rF8Z   ÿ   âEr        â &rZDZ   ÿ   â #rY4      â &r887   ÿ   Ä &rDD7   ÿ   â #rK   K   â &r@BW   ÿ   â #rXX   ÿ   â &r4<[   ÿ   ä$r8?8   ÿŽ ä&rB[   ÿ   ä $rF?F   ÿŽ Ä $r=Z   ÿŽ ä&r<<]   ÿ   ä&rBB]   ÿ   ä $rD=D   ÿŽ â #rP#   P   â |    pbô â $r@9@   ÿŽ â Er8 8       â $r4;4   ÿŽ âEr! F       â $r<;<   ÿŽ â Er        â $r#9   ÿŽ âEr D       â $rB9B   ÿŽ Ä 0r5ÿ5    A   ä 0r\ÿ\    A   â Er@ @       â 0r6ÿ6    A   â Er4 4       â 0r:ÿ:    A   â Er# #       â #r758   ÿ   âEr< <       â #r!\!   ÿ   âErB B       â #r\   ÿ   â #r55   ÿ   â #r@Y@   X   â #r!64   !   â#r6#      â #r<:<   7   â#r5:B   5   â x¤  ÿà÷ â #r	@      â #r>!   >   â #r"	   "   â #r<      â #r	5      â x   ÿàÿ Ä x   ÿàÿ ä x   ÿàÿ ä x   ÿàÿ ä x   ÿàÿ ä x   ÿàÿ ä rÿ   ÿäÿ â D)       À â Gy  0îÿÿÿÿƒ ê Ay        € ê� Ey  ð     € â G™  Ð     € ê $x €  Ž Ê ˆs  úÿ H   è ˆs €úÿ H   è ˆs P ûÿ H   è ˆs €ûÿ H   è ˆs 
 üÿ H   è ˆs L€üÿ H   è ˆs K ýÿ H   è ˆs €ýÿ H   è ˆs > þÿ H   è ˆs "€þÿ H   è ˆs  ÿÿ H   è ˆs €ÿÿ H   ä Ay        € ê {          ì M	        € ê „y      H   b z  r  p`ð Ä z!  r      â „y €   H   " r ÿ   p"p Æ „y    H   ¢ $x'!   ÿŽ ä$x)!   ÿŽ â „y €  H   è „y    H   ( „y €  H   h „y    H   h „y €  H   h „y    H   h „y    H   è „y €  H   â !r         â/ !r       Ä „y €  H   â !rP       ÆO ‰       " !r       Æ� ‰	      b !r
       Æy      '   ¢ !rL       Æ‰      â !rKK       Æ ‰
      " !r       Æ ‰K      b !r       Æ ‰      ¢ !r        â !r>       ä ‰      " !r		       â/ !r"       ä ‰      h ‰      b !r       â� !r       â x   ÿàÿ â ‰      â !r
       Äz  r  p`ò â ‰      " !rKK       äx    p"ò â ‰       b !r       ÆO ‰	      ¢ !r       â !r       ä/ ‰      " !r       Æ ‰      b !r       Æ� ‰      â !r       Æ‰      " !r        Æ‰K      b $z % x  Ž â !r		       äO ‰      ¢ $z  ~   Ž Æ ‰      ¢ !r       Æ ‰      " !r       Æ/ ‰      b !r       Æ� ‰      â !r       âr!    ÿ† ä ‰ €    " !rKK       ârÿÿ   ÿäÿ ä ‰	 €    b !r       âO !r       Ä ‰ €    ¢ !r       Æ ‰ €    " !r       Æ/ ‰ €    b  '   ÿàõ â!r       ä� ‰ €    â ÿÿ   ÿä â !r       ä‰K €    " 
  r  ÿàù â !r
	       Ä‰ €    b 
 j  ÿ„ â $ÿÿ   ÿ â 
 j  ÿˆ â‰ €    b z j  ÿ† ä
 k   â ‰  €    b !r       âO !r       â 
 k   â ‰" €    " !r       â/ z k  � Ä ‰$      b !r       â� r!    ÿ„ ä ‰	 @    ¢ !rK       ä$r(ÿÿ   ÿ â ‰
 @    â !r       â!r       Ä ‰ @    " !r         Æ ‰ @    b !r""       Æ ‰ @    " !r$$       Æ/ ‰ @    b !r	       ÆO ‰ @    ¢ 	!   ÿàÿ â !r
       ä� ‰ @    â * 	   ÿàõ ä z j  ÿ† â ‰ @    â !r       â+ÿÿ   ÿä Ä ‰!  @    " !r       â* j  ÿ„ ä ‰#" @    b !r       â z	 k  (� ä ‰%$ €    " !r       â/  )   ÿàù Ä ‰      b !r       âO * k  + ä ‰      ¢ !r       â� &ÿÿ   ÿä ä ‰      â !r       â 

 j  ÿˆ â!r! !       â‰      $ 
 k  & â !r#"#       â‰      b !r%$%       Æ ‰      " !r+       Æ/ ‰      b '   ÿàÿ â $t'ÿ   ÿ Ž â !r-       âO ‰      ¢     ÿàõ â%v  j  ' Ž â !r       â� ‰      â &ÿÿ   ÿä ä  j  ÿ„ â ‰ !      â !r       Ä k  & â ‰$% @    " !r       Æ‰"#      b !r       â !r       ä/ †	 +    â !r       ÆO † -   â !r       Æ� †	     â !r!!        Æ †     â !r$%$       Æ†	     â !r##"       Æ†    è †	     è †     è †	     è † !   è ‰$      h  †	 
#    â M™        € ê x)   ÿàÿ â !r$       Æ/ r    ÿàñ È r ÿÿ   ÿä  ä z j  ÿ€ È z k     Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y       "   " ¹z  l      â Ey  0    € â ™x?    â y      !   b ¹z  F   
   â r ÿ       â $rÿÿ   ÿ Ž â y      %   ¢ r ÿ       â x       ÿ  â $r9ÿÿ   ÿ Ž â y#      &   â r: ÿ       â $r	ÿÿ   ÿ Ž â r ÿ       â $r>ÿÿ   ÿ Ž â r  ÿ   pRð Ä r  ÿ   pRò â$x     Ž â/ x   ÿ   ÈO xÿ     È r   ÿ(� È xÿ     È |    pbô Ú G)  °    € ê y      '   "Ž xàÿÿÿÿÀŽ â $rÿÿ   ÿ Ž â r" ÿ       â $z# w  ÿŽ âz  q      â $x
   
Ž â r        â 'z# s  " Ž Æ xÿ   
  â $z p  Ž â zÿ t    ä x+
   ÿ   ä x   ÿ   È z q  ÿàÿ â 'z y   Ž â r
   ÿ � är
   ÿ� â $z }  ÿŽ â xðÿÿÿÿÀŽ ä xÿ     ä zÿ z    â $x
   
Ž â xøÿÿÿÿÀŽ Ä r   ÿàõ â $z |  Ž â xÿ     â $x
   
Ž â r   ÿŽ â $z |  Ž â r   ÿ� â$z v  Ž â r   ÿ� â $x   ÿ Ž â x   ÿàÿ â $zT v  Ž â xÿ     â $x   Ž â r   ÿò� â $zO p  Ž â r
   ÿàÿ â %xTTÒ   ÿŽ â xÿ     Ä xÿ     â $x   Ž â x)
   ÿ   â %x$OÒ   Ž âx     ä r   ÿŽ â %xNOÒ   ÿŽ â z  Y      ä xÿ     â $rÿÿ   ÿ â rT   ÿàõ â %xP+   TŽ Æ rU   ÿä â %x&)   TŽ È $vÿ X  ÿ Ž ä $rÿÿ   ÿ Ž ä $x   ÿ Ž ä %x()   NŽ È %x*+   NŽ È $x   ÿ Ž ä r0&   ÿà÷ ärP   ÿàõ Æ $x1'   Ž ârQ   ÿä är.(   ÿàõ ä�yE0€   • ¢  r2*   ÿà÷ ä r/)   ÿä â �y60‚   • ¨  �yH    • è �yK   • è �yI.€   • h	 �y,.‚   • b	 $x3+   Ž Ê �yF2    • h �yG2   • b r   àµ â $tLÿ$   ÿ Ž Æ r   ÿd~ â %v Z  L Ž È/ $x$   ÿŽ â r<   ÿàù ä r0$   ÿàû ä r.   ÿàõ är   ÿàÿ â $x1%   Ž â r=   ÿä Æ $x/    â rY   àÇ â �y8<À   ‘ & r   ÿ„þ â �y70À   ‘ ( �yC.   ™ " %vXY Z  L Ž Æ �y40Ä   ‘ " $x$   ÿŽ Æ �y3.L   ™ $ rYY   ÿàÿ ä �yD    • " rRT   ÿàõ Æ �y5<Ä   ‘ ( �yBH   • ( �y@X    • " r<X   ÿà÷ â $xSU    Ä �yAXH   • " r=Y   ÿäÿ ä r/   àµ â �y?RÐ   • ( �y<   ™ " r   ÿd~ ä r0N   ÿàõ â/ �y<L   ™ " %v./ Z  L Ž â rW   àÇ Æ $x$   ÿŽ ä $x1O    â rR.   ÿàõ ä r//   ÿàÿ ä �y
0Ð   • " r   ÿ„þ ä rS/   ÿä â %vVW Z  L Ž â �y.    • & $x$   ÿŽ â �y=R   ™ "# rXV   ÿàõ Ä $xWW   Ž â rM   àÇ â �yRL   ™ & rYW   ÿä â �y<.H   • " r   ÿ„þ â %vLM Z  L Ž ä �y0X   ™ $ $x$   ÿŽ â rRL   ÿàõ â/ �y-XL   ™ & rMM   ÿàÿ â �y;V    • ( $xSM    â �yVH   • ( �y/R   ™ ( �y1RL   ™ ( �y.L    • ( �y2LH   • " x   ÿàÿ È |    pbô ä xEET  6    ÈO r6ÿ   E  ä xKHT  K    ä� xE60000ÿÀŽ ä xHÿ   K  ä xIIT  ,    Èr,ÿ   I  ä xEEHøŽ ä xJ6   ÿ   ä x6,   ÿ   ä xGFT  G    ä xJJ0000ÿÀŽ Ä xLE€€€€ÿüŽ ä/ x660000ÿÀŽ ä xIÿ   G  ä x,,0000ÿÀŽ ä xJJKøŽ ä xLLàßßßÿàÿ Ä x66GøŽ ä x,,IøŽ ä xMJ€€€€ÿüŽ ä xKE    LŽ ä xI6€€€€ÿüŽ ä xMMàßßßÿàÿ Ä xKK˜º  ÿ    ä xG,€€€€ÿüŽ ä xIIàßßßÿàÿ ä xFE˜º  ÿ    ä xHJ    MŽ ä xEK€€€€LŽ Ä xGGàßßßÿàÿ ä xL6    IŽ ä xFKF`Ž ä xHH˜º  ÿ    ä xKJ˜º  ÿ    ä xJ,    GŽ Ä xLL˜º  ÿ    ä xS6˜º  ÿ    ä xMH€€€€MŽ äxHHK`Ž ä xJJ˜º  ÿ    ä xIL€€€€IŽ Ä xK,˜º  ÿ    ä xLLS`Ž ä xGJ€€€€GŽ ä rHHM   ÿüŽ ä xJJK`Ž ä rLLI   ÿüŽ Ä rFFE   ÿüŽ â &rEHC   ÿ   â	x88€ˆ  ÿ    â &r,LC   ÿ   â rJJG   ÿüŽ ä x77€ˆ  ÿ    â &r6F3   ÿ   â x44€ˆ  ÿ    â $rE8E   ÿŽ ä &r3J3   ÿ   Ä $r,7,   ÿŽ â ErE E       â 0rDÿD    A   ä $r343   ÿŽ â ErC ,       â 0rBÿB    A   â x55€ˆ  ÿ    â #rEDE   ÿ   âEr3 3       â #rDDC   ÿ   Ä $r656   ÿŽ ä #rDB3   D   â &r3H   ÿ   ä ErG 6       â &rL   ÿ   ä &r6F   ÿ   ä $r383   ÿŽ Ä &rJ   ÿ   ä $r7   ÿŽ â Er3 3       â $r656   ÿŽ ä 0r@ÿ@    A   ä 0r,ÿ?    A   ä $r?4   ÿŽ â Er        â 0rAÿA    A   â Er6 6       â #r@3   ÿ   â0r3ÿ
    A   â Er
 ?       â #r@   ÿ   â #rA6      Æ #r
A
      â #r	,   	   â &rH=   ÿ   ä&rL=   ÿ   ä $r8   ÿŽ ä $r7   ÿŽ ä &r?J   ÿ   â Er=        â 0rÿ    A   â Er        â $r?4?   ÿŽ Ä &rAH0   ÿ   â#r==   ÿ   â#r   ÿ   â &r0L0   ÿ   â Er? ?       â &rF-   ÿ   ä $rA8A   ÿŽ ä 0r<ÿ<    A   â #r:3
   :   â $r670   ÿŽ â Er A       â $r05   ÿŽ â #r<?      â &r
F   ÿ   Ä 0r;ÿ;    A   â Er 6       â $r
5
   ÿŽ â Er0 0       â 0rÿ    A   â #r;   ÿ   â #r 3       â &rH/   ÿ   ä&rLL/   ÿ   â #r;;   ÿ   â Er
 
       â #r0      â &rFF1   ÿ   Ä $r8   ÿŽ ä &r-J-   ÿ   ä&r1J1   ÿ   ä $r7L   ÿŽ â #r
<
   =   â $rF5F   ÿŽ â Er        â $r-4-   ÿŽ âEr        â $r141   ÿŽ Ä 0r.ÿ.    A   â #r9,
   9   â 0r2ÿ2    A   â Er- -       ä ErF F       â #r.   ÿ   âEr
 1       â #r.   ÿ   â x¤  ÿà÷ â #rEBG   E   â #r--   ;   â #r2F      â#r2
      â #r>,E   >   â#r3D      â#r!,   !   â#r3-      â#r,      â #r3      â x   ÿàÿ â $rÿÿ   Ž â x   ÿàÿ Ä x   ÿàÿ ä x   ÿàÿ ä x   ÿàÿ â G©  @ðÿÿÿÿƒ ê Ay        € ê� Ey  Ð     € â G‰  °     € ê $x  @  Ž Ê ˆs  > ûÿ H   è ˆs  €ûÿ H   è ˆs  	 üÿ H   è ˆs  :€üÿ H   è ˆs  9 ýÿ H   è ˆs   €ýÿ H   è ˆs  ! þÿ H   è ˆs  €þÿ H   è ˆs   ÿÿ H   è ˆs  €ÿÿ H   ä Ay        € ê {          ì M        € ê „y     H   b z  r  p`ð Ä z%  r      â „y €   H   ¢ r ÿ   p"p Æ „y    H   è „y €  H   ( „y    H   h „y €  H   ( „y
    H   ( „y €  H   ( „y    H   è „y €  H   â !r>       â/ !r       ÈO ‰       " !r	       â� !r::       È‰	      b !r9       Æ‰      ¢ !r        Æ ‰:      â !r

!       â $x!%   ÿŽ â !r       â !r         ä ‰
      " !r       Æ ‰       " !r       â !r		       ä/ ‰      b !r       ÆO ‰      ¢ !r:       Æ� ‰      è ‰      h ‰      b !r
       â !r        Ä‰	      " !r       Æ/ ‰  €    b !r       ÆO ‰      ¢ !r       Æ� ‰      â !r       Æ‰
      " !r       Æ ‰      h ‰      b !r	       Æ ‰      " !r        â/ $z # x  Ž ä ‰      b !r       âO $t#ÿ   ÿ Ž ä ‰      ¢ !r       â� !r
       Ä‰	 €    â x
   ÿàÿ â !r       ä‰ €    " z 
 r  p`ò â !r       ä ‰ €    b x    p"ò â !r       Ä ‰ @    " !r       Æ/ ‰ €    b !r       ÆO ‰ €    ¨ y      '   ¢ !r	       Æ� ‰ €    â !r       Æ‰ €    " !r       Æ‰ €    b !r
       Æ ‰ €    " !r       â/ !r       äO ‰ @    b $z  ~   Ž Æ ‰
      ¢ !r       Æ� ‰ @    â  !   ÿàõ â !r       ä‰ @    " r%    ÿ† â !r       ä‰ @    b  ÿÿ   ÿä â !r       Ä ‰ @    " 
 j  ÿ„ ä rÿÿ   ÿäÿ â ‰ @    " !r       â/ z j  ÿ† ä ‰ @    b !r'
       âO %   ÿàÿ Ä ‰ @    ¢ !r       â� !   ÿàÿ ä ‰ @    â !r       â
 k    â !r       â‰      " z k  � â !r       Ä ‰      " !    ÿàõ â%v
  j  # Ž â !r       â ‰      b 
	  r  ÿàù â !r       ä/ ‰      b r%    ÿŠ â !r       ÄO ‰      ¢     ÿà÷ â $r"ÿÿ   ÿŽ â !r       â� ‰      â z j  ÿŠ ä$ÿÿ   ÿä â ‰       â z k  "� â !r#       â
	 j  ÿˆ â ‰      " "ÿÿ   ÿä â !r       â ! j  ÿ„ â ‰      "  ÿÿ   ÿäÿ â !r       â
		 k  $ â !r       â/  j  ÿ† â †	 
'    â ! k  " â !r       âO  k   � â † 
#   â !r       Æ� †	     â !r        Æ †     â !r       Æ†	     â !r       Æ †    è †	     è †     è †	     â M™        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y       "   " ¹z  l      â Ey      € â ™x?    â y      !   b ¹z  F   
   â r ÿ       â $rÿÿ   ÿ Ž â y
      %   ¢ r3 ÿ       â $rÿÿ   ÿ Ž â r: ÿ       â y      &   â $rAÿÿ   ÿ Ž â r ÿ       â $rÿÿ   ÿ Ž â r  ÿ   pRð Ä r  ÿ   pRò â$x     Ž â/ x

   ÿ   ÈO xÿ     È r   ÿ(� È xÿ     È |    pbô Ú G)       € ê y      '   "Ž xàÿÿÿÿÀŽ â $rÿÿ   ÿ Ž â r ÿ       â $z w  ÿŽ ä$x	   
Ž â r        â 'z s   Ž Æ xÿ   	  â $z
 p  Ž â zÿ t    â $x+	   ÿ Ž â x   ÿ   È z q  ÿàÿ â 'z y   Ž âr	   ÿ � är	   ÿ� â $z }  ÿŽ â xðÿÿÿÿÀŽ ä zÿ z    ä xÿ     â $x	   
Ž â xøÿÿÿÿÀŽ â $z |  Ž â r   ÿàõ â $z |  Ž â xÿ     â $z v  Ž â x
   ÿàÿ â $z v  Ž â r   ÿ� â$x	   
Ž â x   ÿ   â $z# p  Ž â r   ÿ� â %xÒ   ÿŽ â r   ÿò� Ä z  q      â $x   Ž â xÿ     â %x$+   Ž â x	ÿ     ä r   ÿŽ â$x   Ž â rÿÿ   ÿä â %x #Ò   Ž â xÿ     Ä r   ÿàõ â %x"#Ò   ÿŽ â x     äz  Y      â $x)   ÿ Ž â x   ÿ   â $x   	Ž â x	   ÿ   â %x&)   Ž Æ xÿ     â $vÿ X  ÿ Ž ä $rÿÿ   ÿ Ž ä %x()   "Ž È %x*+   "Ž È $x    ä r,&   ÿà÷ är4$   ÿàõ är-'   ÿäÿ ä rD*   ÿà÷ â$x5%    ä �y7,€   • ¢  r<(   ÿàõ Æ �y,‚   • ¢  rE+   ÿäÿ Æ �yB4    • â $x=)    Æ �y;4   • è �y9<€   • h	 �yF<‚   • h	 �y8D    • h �y?D   • b r	   àµ Ä x0 $       ä r   ÿd~ ä r.   ÿàù â %v4	 Z  0 Ž È/ $x$   ÿŽ ä $x/    â rM   àÇ â $x55   Ž â r,    ÿàû ä rD4   ÿàõ â �yC.À   ‘ b r	   ÿ„þ ä r-!   ÿäÿ â �y>.Ä   ‘ b rE5   ÿä â %vLM Z  0 Ž â rJ"   ÿà÷ â �y@,À   ‘ d $x	$   ÿŽ Ä �y6D   ™ b r.   ÿàõ Æ �y=,Ä   ‘ "rMM   ÿàÿ â $x/    ä�y<DL   ™ " $xK#   Ž Æ �y	4    • " r,L   ÿàõ Æ/ �y24H   • " r-M   ÿä ä r1   àµ â �yE.Ð   • ( �y,   ™ " r4   ÿd~ Æ/ �yD,L   ™ " %v.1 Z  0 Ž Æ �yGJÐ   • " $x14$   ÿŽ â r4.   ÿàõ ä �yHL    • " $x//   1Ž â r1   àÇ ä �yILH   • $ r5/   ÿä Ä r,   ÿ„þ â/ %v01 Z  0 Ž ä �yJ4   ™ $ $x-,$   ÿŽ â r,0   ÿàõ â �yK4L   ™ & r11-   ÿàÿ â �yL.    • ( $x-1    â �y40    • ( �y..H   • (# �yM,   ™ ( �y50H   • ( �y,,L   ™ " x   ÿàÿ È |    pbô ä xH  ÿà÷ ä x    ÿàÿ ä x    ÿàÿ â $rÿÿ   Ž â x    ÿàÿ ä x    ÿàÿ Ä x77T      ÈO r7ÿ   7  ä x;BT  ;    ä� x/70000ÿÀŽ ä/xÿ   ;  ä x77   ÿ   ä x//øŽ Ä x99T  F    äx0/€€€€ÿüŽ ä r9ÿ   9  ä x?8T  ?    ä x870000ÿÀŽ ä x9   ÿ   Ä x00àßßßÿàÿ ä x88;øŽ ä xFÿ   ?  ä x990000ÿÀŽ ä x0000ÿÀŽ ä x-/    0Ž Ä x78€€€€ÿüŽ ä xF9FøŽ ä x?øŽ ä x9-˜º  ÿ    ä x//˜º  ÿ    ä x77àßßßÿàÿ Ä x;F€€€€ÿüŽ ä x1€€€€ÿüŽ ä x-9€€€€0Ž äxB8    7Ž ä x09/`Ž ä x/;àßßßÿàÿ Ä x11àßßßÿàÿ ä xBB˜º  ÿ    ä x?8˜º  ÿ    ä x9F    /Ž ä x;    1Ž ä x7B€€€€7Ž Ä x8B?`Ž ä xB9˜º  ÿ    ä x9F˜º  ÿ    ä xF;˜º  ÿ    ä x;˜º  ÿ    ä x1F€€€€1Ž Ä xFF;`Ž ä x/B€€€€/Ž äxBB9`Ž ä r-0-   ÿüŽ ä r787   ÿüŽ ä r0F1   ÿüŽ Ä r/B/   ÿüŽ â &r76   ÿ   âx;C€ˆ  ÿ    â &r606   ÿ   â x9@€ˆ  ÿ    ä x8>€ˆ  ÿ    â &r>/<   ÿ   â	x1=€ˆ  ÿ    â &r=-<   ÿ   ä $r;   ÿŽ Ä $r<96   ÿŽ ä $r=8=   ÿŽ â Er6        â 0r	ÿ	    A   â Er <       â $r>1>   ÿŽ â Er< =       â 0r2ÿ2    A   â #r=	6   ÿ   â#r		   ÿ   â Er> >       â &r67   ÿ   â#r=2<   =   â &r<0   ÿ   Ä #r22>   	   â &r	-D   ÿ   ä $r6;6   ÿŽ ä $r>9<   ÿŽ ä 0rGÿG    A   â Er 6       â $r<8	   ÿŽ ä 0rHÿH    A   â Er	 >       â 0rIÿI    A   â Er< <       â #rG2      â #r6H   ÿ   â &r27J   ÿ   Ä &rJ0J   ÿ   â #rHH	   ÿ   â 0rÿE    A   â #r6I<   6   â &rD/D   ÿ   ä &r	-K   ÿ   ä&r</K   ÿ   ä $rJ9J   ÿŽ â #rA6   A   â $rD1D   ÿŽ Ä $r2;2   ÿŽ ä $r68	   ÿŽ â Er	 J       â $r<1<   ÿŽ ä 0rLÿL    A   â #r=      â Er D       â 0r.ÿ.    A   â Er= 2       â #r	L	   ÿ   â Er< <       â &r27M   ÿ   Ä &r00M   ÿ   â #rI   H   â #r<.<   	   â &r--,   ÿ   ä $r2;2   ÿŽ ä $r	90   ÿŽ ä &r,/,   ÿ   â #r:G   :   â $r-8-   ÿŽ â Er 2       â 0r4ÿ4    A   â Er	 	       â $r11,   ÿŽ â Er6 6       â 0r5ÿ5    A   â Er- -       â #r,4   ÿ   â Er1 1       â #r=L=   ÿ   â #r44	   ÿ   â #r,5-   ,   Ä #r6.6   =   â #r151   4   â #r3G<   3   â#r,      â#r6      â #rG1      â G©  @òÿÿÿÿƒ ê Ay        € ê� Ey  °     € â G‰  �     € ê r     ÿ@Ž Ê ˆs   üÿ H   è ˆs  €üÿ H   è ˆs  A ýÿ H   è ˆs  :€ýÿ H   è ˆs   þÿ H   è ˆs  3€þÿ H   è ˆs   ÿÿ H   è ˆs  €ÿÿ H   ä Ay        € ê {          ì M        € ê „y    H   b z 
 r  p`ð â $z x  
Ž Ä „y    H   ¢ r ÿ   p"p Æ „y     H   è „y  	  H   ( „y €   H   h „y     H   ( „y €  H   ( „y	    H   ( „y €  H   ( „y €  H   ( „y    H   " !rA       Æ/ „y €  H   b !r       ÆO „y    H   ¢ !r       Æ� „y €  H   â !r       Æ„y    H   " !r       Æ„y! €  H   b !r         Æ „y%    H   " !r       Æ „y €	  H   " !r  	       Æ „y  
  H   " !r       Æ „y# €
  H   " !r:       Æ „y'    H   " !r       Æ „y €  H   â !r3       Æ/ „y €  H   â !r       âO z  r      ä „y €  H   â !r       Æ� ‰       b !r       â!r!       ä‰      ¢ !r%       Æ ‰	      " !r       Æ y      '   â !r       Æ ‰      " !r#       Æ ‰      b !r'       Æ ‰      b !r        â/ !r       ä ‰      b !r       âO !r       ä ‰       ¢ !r		       â !r       Ä ‰      " $z ~  Ž ä� $x   ÿŽ â !r       â‰	      â !r       Æ‰      " !r       Æ ‰      b !r       Æ/ ‰      b !r         ÆO ‰      ¢ !r       Æ ‰      ( ‰  €    " !r	       Æ� ‰ €    â !r       â!r       ä‰	 €    " !r       Æ/ ‰ €    b !r       ÆO ‰ €    ¢ !r       Æ ‰ €    " !r        Æ ‰ €    b !r       Æ� ‰      è ‰  @    â !r	       âx	
   ÿàÿ ä ‰ @    " !r       â/ z 	 r  p`ò â !r       âO ‰ @    d x    p"ò â !r       â 
 r  ÿàõ â ‰ @    " !r
       Æ‰ @    ¢ $ÿÿ   ÿ â !r       ä� ‰ @    â !r         â r   ÿ† ä ‰
 @    b !r	       â
 j  ÿ„ Ä ‰ €    " rÿÿ   ÿäÿ â !r       ä/ ‰       b z j  ÿ† ä 
 k   â !r       â ‰	      " z k  � â !r       ÄO ‰      ¢    ÿàÿ â !r       ä� ‰      â    ÿà÷ â !r
       ä‰      b !r       Æ‰      " !r        Æ/ ‰      b     ÿàõ â $ÿÿ   ÿŽ â x        â ‰ @    b ÿÿ   ÿä â !r	       â   j  ÿ„ â %v j   Ž â 

 j  ÿ† ä 	  k   â !r       âO !r       â� 
 k  � â †	     â !r       Æ†    â !r       Æ†	     â !r       Æ/ †     â !r       Æ †	     è †    è ‰      h  †	 
    â M™        € ê x   ÿàÿ â !r       Æ/ r   ÿàñ È r ÿÿ   ÿä  ä z j  ÿ€ È z k     Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      "   " ¹z  l      â Ey       € â ™x?    â y       !   b ¹z  F   
   â $r
ÿÿ   ÿ Ž â r2 ÿ       â y	      %   ¢ $r0ÿÿ   ÿ Ž ä $rÿÿ   ÿ Ž â y      &   â $r:ÿÿ   ÿ Ž â r ÿ   pRð Ä r ÿ   pRò â$x     Ž â/ x		   ÿ   ÈO xÿ     Ê $À    Ž â r   ÿ(� ä r ÿ       ä xÿ     È |    pbô Ú G)       € ê y      '   "Ž xàÿÿÿÿÀŽ â $rÿÿ   ÿ Ž ä $rÿÿ    Ž ä $x   
Ž â r ÿ       â 'z s   Ž Æ xÿ     â $z w  ÿŽ â zÿ t    â $x'   ÿ Ž â r   ÿ � â $v/ÿ X  ÿ Ž Æ xÿ     â 'z y   Ž È $z }  ÿŽ â zÿ z    â $z	 p  Ž âxðÿÿÿÿÀŽ ä r   ÿàõ â $z |  Ž ä$x   
Ž ä $z |  Ž â x	   ÿàÿ â $z v  Ž âxÿ     â $z v  Ž â z  q      â $z p  Ž â r   ÿ� â$x   ÿ Ž â r   ÿ� â $x   ÿ Ž â r   ÿ� â %xÒ   ÿŽ â r   ÿò� Ä xøÿÿÿÿÀŽ â $x   Ž â xÿ     â %x '   Ž â r  €ÿàÿ ä xÿ     â %xÒ   Ž âr   ÿàÿ ä xÿ     â %xÒ   ÿŽ â x   ÿ   Ä x     â $rÿÿ   ÿ â r   ÿàõ â $x%   ÿ Ž â r ÿ       â $x   Ž â z q  ÿàÿ â %x"%   Ž â r   ÿä ä xÿ     â $vÿ Y  ÿ Ž Ä %x$%   Ž È %x&'   Ž È $x   ÿ Ž ä r("/   ÿà÷ är6 /   ÿàõ Æ $x)#   Ž â r7!   ÿä ä r,$/   ÿàõ ä�y1(€   • ¢  r4&/   ÿà÷ Æ �y(‚   • ¢  r-%   ÿä Æ �y>6    • è �y6   • è �yB,€   • h	 �yA,‚   • b	 $x5'   Ž Ê �y34    • h �y84   • b r=   àµ â $t(ÿ$   ÿ Ž Æ r   ÿd~ â %v<= Z  ( Ž â rD/   ÿàù Æ $x$   ÿŽ â r*/   ÿàû ä rE   ÿä ä r,<   ÿàõ är==   ÿàÿ â $x+   Ž â �y9DÀ   ‘ " r   àÇ Ä $x-=    â �y5*Ä   ‘ "# r)   ÿ„þ Æ �y6DÄ   ‘ ( �y?,   ™ " $x))$   ÿŽ Æ �y4,L   ™ " %vD Z  ( Ž Æ �y7*À   ‘ $ rEE)   ÿàÿ ä �y@<    • " r,D   ÿà÷ ä/ r*/   ÿàõ ä r-E   ÿäÿ â �y=<H   • $ $x+    â r)   àµ â �y,   ™ & rC   ÿd~ â �y;*Ð   • ( �y.,L   ™ ( �y<D    • " r*/   ÿàõ â/ %v,) Z  ( Ž È $x)C$   ÿŽ ä $x+    â r(,   ÿàõ â �yCDH   • " r--)   ÿàÿ Æ �y**Ð   • $ $x)-    ä �yD,    • ( �y+(   ™ (# �y,,H   • ( �y((L   ™ "# x   ÿàÿ È |    pbô ä x//H  ÿà÷ ä x    ÿàÿ ä x    ÿàÿ â $rÿÿ   Ž â x    ÿàÿ ä x1T      ÈO rÿ     ä x>>T      ä� x10000ÿÀŽ ä xÿ   >  ä xBBT  A    äx   ÿ   Ä x1øŽ ä rBÿ   B  ä xE0000ÿÀŽ ä x€€€€ÿüŽ â $x)B   ÿ Ž â/ x33T  8    ä xàßßßÿàÿ Ä xAÿ   3  ä x8B0000ÿÀŽ ä xB)0000ÿÀŽ ä xEE>øŽ ä x-    Ž ä xA8AøŽ Ä xBB3øŽ ä x)E€€€€ÿüŽ ä x3-˜º  ÿ    ä x1A€€€€ÿüŽ ä x>˜º  ÿ    ä x-B€€€€ÿüŽ Ä x)àßßßÿàÿ ä x)3€€€€Ž ä x1àßßßÿàÿ ä x--àßßßÿàÿ ä x8E    Ž ä x>3>`Ž Ä x3A    Ž ä x1B    -Ž ä x88˜º  ÿ    ä xEE˜º  ÿ    ä x33˜º  ÿ    ä xAA˜º  ÿ    Ä x11˜º  ÿ    ä xBB˜º  ÿ    ä x8€€€€Ž äx88E`Ž ä x3€€€€Ž äxA3A`Ž Ä x-1€€€€-Ž äxB1B`Ž ä r88   ÿüŽ ä x99€ˆ  ÿ    ärAA   ÿüŽ ä rB-   ÿüŽ Ä r)>)   ÿüŽ ä x-5€ˆ  ÿ    ä x16€ˆ  ÿ    â &r68?   ÿ   â r5 9       â &r>)4   ÿ   ä&r4A4   ÿ   ä $r656   ÿŽ â x37€ˆ  ÿ    â $r>1>   ÿŽ Ä $r4-4   ÿŽ â Er7 6       â 0r@ÿ@    A   ä &rB?   ÿ   â Er> >       â 0r=ÿ=    A   â Er9 4       â #r4@7   ÿ   â $r3B   ÿŽ Æ #r7=>   4   â &r48   ÿ   â Er        â &r>   ÿ   ä &r6).   ÿ   ä $r454   ÿŽ â #r@@   ÿ   â $r3>   ÿŽ ä $r616   ÿŽ â Er 4       â 0r<ÿ<    A   Ä &r.A.   ÿ   â Er        â 0rCÿC    A   â Er6 6       â $r4-.   ÿŽ â #r.<   ÿ   â0r;ÿ;    A   â #r<<   ÿ   â 0rÿ*    A   â #rC6   .   â &r88+   ÿ   Ä &r*+   ÿ   â #r2;   2   â &r)(   ÿ   ä&r(A(   ÿ   ä $r858   ÿŽ ä $r3*   ÿŽ ä $r1   ÿŽ â Er) 8       â $r(-(   ÿŽ â Er        â 0rDÿD    A   â #r9=9   @   â Er= 4       â 0r,ÿ,    A   â Er        â #r)D)   ÿ   âEr( (       â #rD   ÿ   â #rCC=   <   â #r,   )   Ä #r(,(      â #r:;7   :   â #r9      â#r0C   0   â #r
;   
   â #r(      â G©  0ôÿÿÿÿƒ ê Ay        € ê� ˆ : ýÿ H   è ˆ €ýÿ H   è ˆ 2 þÿ H   è ˆ 0€þÿ H   è ˆ 
 ÿÿ H   è ˆ €ÿÿ H   è {          ì M        € ê „y      H   " z 	 r  p`ð Æ „y     H   b r  ÿ   p"p Æ „y  €   H   ¨ „y     H   è „y     H   ( „y  €  H   h „y     H   h „y  €  H   h „y     H   h „y  €  H   h „y     H   b !r:       Æ „y  €  H   " !r       Æ/ „y  €  H   b !r       ÆO „y     H   ¢ !r2       Æ� „y  €  H   â !r       Æ„y     H   â !r0       Æ„y  €  H   â !r
       Æ „y  €  H   â !r       â !r       ä ‰      " !r       Æ y      '   b !r       â !r       â !r       â/ !r       ÈO ‰      " !r       Æ� ‰      b !r       â!r       ä ‰      ¢ !r       Æ ‰      â !r       Æ ‰      ( ‰      b !r       â !r
       È/ ‰      " !r       ÆO ‰
      b !r       Æ� ‰      ¢ !r       Æ‰      â !r       Æ‰ €    ( ‰      b !r       â !r
       â/ !r       ÈO ‰ €    " !r
       Æ� ‰ €    b !r       Æ‰ €    ¢ !r       Æ‰
 €    è ‰ €    ( ‰ @    b !r       â !r       È/ ‰ @    " !r       ÆO ‰ @    b !r
       â� x
	   ÿàÿ ä ‰ @    ¢ $z	 x  	Ž â !r       âz 
 r  p`ò â ‰ @    â $z	 ~  	Ž â !r       âx     p"ò â ‰ @    " z  r      Æ ‰      b !r
       â 
	 r  ÿàõ â !r       Ê/    ÿàÿ är	   ÿ† â !r       âO ‰       " !r       Æ� ‰
      b $ÿÿ   ÿ â 
 j  ÿ„ â !r       â‰      ¢ $rÿÿ   ÿŽ â 
 k   â !r       â‰      â 	   ÿàõ ä x        â ‰      " ÿÿ   ÿä Ä  j  ÿ„ â %v	 j   Ž â z j  ÿ† ä 	 k   â !r        â z k  � â †	     â !r
       Æ/ †    â !r       ÆO †	     â !r       Æ� †     â !r       Æ†	     â M™        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      "   " ¹z  l      â Ey       € â ™x?    â y       !   b ¹z  F   
   â $rÿÿ   ÿ Ž ä y      %   ¢ $rÿÿ   ÿ Ž ä $r+ÿÿ   ÿ Ž â y      &   â r ÿ   pRð Ä r ÿ   pRò â$x     Ž ä/ $x   ÿ Ž ÆO xÿ     È r   ÿ(� â $rÿÿ   ÿ Ž â     ÿ8Ž ä xÿ     È |    pbô Ú G)  `    € ê y      '   "Ž x	àÿÿÿÿÀŽ â $rÿÿ   ÿ Ž â x   ÿàÿ â $rÿÿ   ÿ Ž ä $x   	
Ž ä $z w  ÿŽ Æ xÿ     ä x%   ÿ   ä r
   ÿ � är   ÿ� È xøÿÿÿÿÀŽ â 'z y   Ž È 'z s   Ž â zÿ z    Æ $z	 }  ÿŽ â x
ðÿÿÿÿÀŽ â $z p  Ž â zÿ t    â $z |  Ž âx
ÿ   
  â $x   
Ž â r		   ÿàõ â $z |  Ž ä $z v  Ž âxÿ     â $z v  Ž Ä $z p  Ž â r   ÿ� â$x
   ÿ Ž â r   ÿ� â $x   
Ž â xÿ     â %xÒ   ÿŽ â r&   ÿò� ä xÿ     â$x   Ž â x'ÿ   &  â$rÿÿ   ÿ â r&   ÿàõ â $x   ÿ Ž â x     â %xÒ   &Ž â z  X      Ä z q  ÿàÿ â %xÒ   ÿŽ È $x#   ÿ Ž ä $x
   Ž â x   ÿ   â %x%   Ž Æ x
ÿ     â %x #   Ž È $vÿ Y  ÿ Ž ä $rÿÿ   ÿ Ž ä $x   ÿ Ž ä %x"#   Ž È %x$%   Ž È $x   ' ä r0    ÿà÷ är4   ÿàõ är("   ÿàù â$x1!   Ž ä$x5    ä$x)#    â �y:0€   • ¨  �y0‚   • ¨  �y=4    • è �y,4   • è �y;(€   • h	 �y.(‚   • b	 r&$   ÿàõ Ê $x'%    Ê �y3&    • h �y8&   • b rA	   àµ ä x* $       ä rD   ÿàù ä r-
   ÿd~ â %v@A Z  * Ž â r>   ÿàû Æ $xE    â r'	   àÇ â $x--$   ÿŽ â rB@   ÿàõ ä r)
   ÿ„þ â$xAA   -Ž â �y5DÄ   ‘ "/ $x?   Ž ä $xCA    â �y1DÀ   ‘ " %v&' Z  * Ž Æ �y7>À   ‘ " $x))$   ÿŽ â r(&   ÿàõ ä �y4B   ™ " $x''   )Ž Æ �y/BL   ™ $ r)'   ÿä ä �y*>Ä   ‘ ( �y-(   ™ ( �y2(L   ™ ( �y0&    • ( �y<&H   • " r(   ÿàõ Æ �y9@    • ( �y6@H   • " r&   ÿà÷ â/ $x)    È $x'   Ž â �y@(Ð   • ( �yA&Ð   • " x   ÿàÿ È |    pbô ä xH  ÿà÷ ä x    ÿàÿ ä x    ÿàÿ ä rÿ   ÿäÿ ä x::T      ÈO r:ÿ   :  ä x=T  ,    È� x=ÿ     ä x,;T  .    äx.:0000ÿÀŽ ä r,ÿ   ,  ä x=.=øŽ Æ $x(,   ÿ Ž â x.=€€€€ÿüŽ â $x::   ÿ Ž Æ x((0000ÿÀŽ ä x33T  8    ä x..àßßßÿàÿ ä x((3øŽ ä x':0000ÿÀŽ ä/ x;=    .Ž Ä x&ÿ   3  ä x3(€€€€ÿüŽ ä x'øŽ ä x;;˜º  ÿ    ä x8=˜º  ÿ    ä x33àßßßÿàÿ Ä x'€€€€ÿüŽ ä x.;€€€€.Ž äx;;8`Ž ä x8(    3Ž ä x),0000ÿÀŽ ä x''àßßßÿàÿ Ä x88˜º  ÿ    ä x,    'Ž ä x&)&øŽ ä x=(˜º  ÿ    ä x(8€€€€3Ž ä x,,˜º  ÿ    Ä x3&€€€€ÿüŽ ä x)˜º  ÿ    ä x=8=`Ž ä x',€€€€'Ž äx33àßßßÿàÿ ä x,,)`Ž Ä r(=(   ÿüŽ ä x&    3Ž ä r','   ÿüŽ ä x)5€ˆ  ÿ    â&r5(4   ÿ   â x,7€ˆ  ÿ    ä x˜º  ÿ    Ä x7&˜º  ÿ    â &r:'4   ÿ   â x&€€€€3Ž â $r4,5   ÿŽ â x77`Ž ä r.;.   ÿüŽ ä Er 4       â &r4'-   ÿ   âr&7&   ÿüŽ â &r-(-   ÿ   â x11€ˆ  ÿ    Ä x'*€ˆ  ÿ    â &r(&/   ÿ   ä&r&&2   ÿ   ä $r-,-   ÿŽ ä &r8./   ÿ   ä $r31:   ÿŽ ä$r114   ÿŽ Ä &r..2   ÿ   ä $r('(   ÿŽ ä$r&'&   ÿŽ â Er' -       â $r5)8   ÿŽ â Er8 3       â 0r0ÿ0    A   â Er1 1       â 0r9ÿ9    A   Ä $r.).   ÿŽ â Er& &       â 0r<ÿ<    A   â #r'0'   ÿ   â0r6ÿ6    A   â Er5 5       â #r398   ÿ   âEr( (       â #r99   ÿ   â Er. .       â #r101   ÿ   â #r'<&   '   â 0r@ÿ@    A   â #r65   3   â0r&ÿA    A   â #r(6(   9   â #r.<.   1   â #r+@   +   Ä #r&(      â #r@.      â #r&'      â G©  @öÿÿÿÿƒ ê Ay        € ê� ˆ + þÿ H   è ˆ €þÿ H   è ˆ  ÿÿ H   è ˆ €ÿÿ H   è {          ì M        € ê „y      H   " z  r  p`ð Æ „y  €   H   b r  ÿ   p"p Æ „y     H   ¨ „y     H   è „y  €  H   ( „y     H   h „y	     H   h „y  €  H   h „y     H   h „y  €  H   h „y  €  H   b !r+       Æ „y  €  H   " !r       â/ !r       âO !r       ä� !r       â!r       â!r	       ä !r       Æ ‰      b !r       Æ ‰      ¢ !r       Æ ‰
      â !r       È !r       Ê ‰      " !r       â/ !r       ÈO ‰      b !r

       Æ� ‰      ¨ ‰	
      â !r       â $tÿ   ÿ Ž È ‰      " !r       â/ !r       ÈO ‰ €    b !r
	       Æ� ‰	 €    ¨ ‰ €    â !r       Ê ‰ €    " !r       Æ/ y      '   b !r		       âO !r
       È� ‰	 @    ¢ x   ÿàÿ â $z x  Ž ä ‰ @    â z  r  p`ò Æ ‰
 @    " x     p"ò â !r       â $z ~  Ž Ê/ 
  r  ÿàõ â !r	       ÈO $ÿÿ   ÿ â !r       â� 
  j  ÿ„ â ‰      " !r
       Æ‰      h ‰
 @    ¨ ‰	      â !r       â 
  k   â !r       â/ %v j   Ž â !r

       ÈO †	     â !r	       Æ� †    è ‰	
      h  †	     â M™        € ê ¹z  r      â !r	
	       â/ �x   ?àÿ Ì |   ÿàñ Ê $r ÿÿ   ÿ  â z j  ÿ€ È z k     Ê †y 	    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž ä $rÿÿ    Ž È % \   Ž Ê �	    ™ b $ÿÿ   ÿ Ž â ¹z  l      â Ey  à    € â y      "   b '
  m   Ž â ™x?    ä y      !   ¢ $‚ÿÿ   ÿ Ž â 
 ÿ n     â $rÿÿ   ÿ Ž â y      '   & $ ÿÿ    
Ž È $
  o  Ž È 'Š  s   Ž â r ÿ   pRò ä/r ÿ   pRô â$x    Ž ÊO x	ÿ     È r		   ÿ(� È xÿ   	  È |    pbö ä Šÿ t     Ö G9  �    € ê x	àÿÿÿÿÀŽ â y      %   " $rÿÿ   ÿ Ž â x   ÿ   â $z }  ÿŽ ä$x   
Ž ä 'z y   Ž Æ x ÿ     â $tÿÒ   ÿ Ž â zÿ z    â $rÿÿ   ÿ Ž â r    ÿ � är     ÿ� ä xðÿÿÿÿÀŽ ä xÿ     Æ $x
   
Ž ä $x   ÿ Ž ä $z |  Ž â xÿ   
  â $z w  ÿŽ ä $z v  Ž âr	
   ÿ� är

   ÿ� â $z p  Ž â r	   ÿò� Ä r   ÿàù ä x	ÿ     Æ $rÿÿ   ÿ ä %xÒ   Ž È %v X  Ž â z X  ÿàñ ä x øÿÿÿÿÀŽ â $rÿÿ    Ž â x Ä   ÿà÷ ä xÿ   
  â $x   
Ž â zÿ Y  þ â $tÿ   ÿ Ž Ä $x   Ž âxÿ     â $x   Ž ä $x   ÿ Ž â x     â $x   ÿ Ž â xÿ     â %v  € Ž È $x   ÿ Ž ä r   ÿàñ Ê $x     Ê �y    • ¨  �y!   • ¤  $rÿÿ    Ž ä $rÿÿ    Ž È %x   Ž â r   à± â �yÐ   • è  �y    • ( �y#   • " r
   ÿd~  â $tÿ$   ÿ Ž È %v Z   Ž È $x%
$   ÿŽ â r
   ÿ€ â $rÿÿ     Ž ä/ $r	ÿÿ    Ž â r%   ÿàÿ È �y"üÿÿ ‘ b $x     Æ �y     ‘ è �y 
   ™ è �y$
L   ™ è �y%    • è  �yH   • â x   ÿàÿ È |    pbð ä xH  ÿàù ä x    ÿàÿ Æ $rÿÿ    â x!T  !    ÈO r!ÿ   !  Ê $x!   ÿ Ž âx
!0000ÿÀŽ È/ x0000ÿÀŽ ä xT  #    Èx!øŽ äx#ÿ     ä x!€€€€ÿüŽ ä x
#øŽ ä xàßßßÿàÿ ä x€€€€ÿüŽ Ä x
!    Ž ä xàßßßÿàÿ ä x

˜º  ÿ    ä x    Ž ä x!˜º  ÿ    ä x
€€€€Ž Ä x˜º  ÿ    ä x˜º  ÿ    ä x

`Ž ä x€€€€Ž äx`Ž ä r
   ÿüŽ Ä x""€ˆ  ÿ    är   ÿüŽ â &r     ÿ   ä� $rÿÿ   " Ž â x
 €ˆ  ÿ    â &r$   ÿ   ä $r    ÿŽ ä $r

   ÿŽ Ä 0r ÿ%    A   â Er        â 0rÿ    A   â Er
 
       Æ #r    ÿ   â x H  ÿà÷ â 0rÿ    A   ä #r
      ä $rÿÿ   	Ž ä #r      â G‰  `ûÿÿÿÿƒ ê Ay        € ê ˆ €ÿÿ H   è {          ì M)        € ê „y      H   " r ÿ   pRð Æ „y €   H   h „y    H   ¢!r         È !r         È/ !r         ÊO ‰       $ !r        Ê ‰      $ !r       Ê ‰ €    $ !r       ä y      %   ( ‰ @    b $z x  Ž â !r       Ê/ ‰	      b  M	        € ê $tÿ   ÿ Ž â !r		       â/ $z ~  Ž È %v j   Ž Ê †y 	    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž È % \   Ž Ê �		    ™ ¢  $rÿÿ   ÿ Ž â Ey  °    € ä y       "   b 'Š s   Ž Æ y      !   b '

 m   Ž â x      ÿ  ä y      '   $ 
ÿ n  
  ä rÿ    ÿü‚ â/ $x     Ž Ê xÿ     È r   ÿ(� ä Š	ÿ t    ä y      %   d r
	       âO G  °    € ê z ÿ `  pPô ä z ÿ ^  pPò ä z ÿ b  p ö Ä z ÿ a   Sô ä z ÿ _  Sò ä z ÿ c  0't Ö $ ~  ÿŽ âÿ     â/$ª ~  ÿŽ ä $
 x  ÿŽ â    àë â$ª
 x  ÿŽ â ¢   àÇ ä¨
ÿ     ä ÿÿ   Äþ Ä    ÿàý ä ¢   ÿàû ä ¢ÿÿ   
„þ â $ÿÿ    â 
 ^  ÿ† ä ª b  ÿˆ â $¢ÿÿ   Ž â  _  � Ä x      ÿ  ä ª c   È �
     h �©     d Ay        € ê ¹z  l      â xÿ     â ™x?    â $ÿÿ   
Ž â Ey  @    € â $r ÿÿ   ÿ Ž ä $

 o  Ž â|    pbò â $‚
ÿÿ    Ž Ä $rÿÿ   ÿ Ž Ô G  Ð
    € ê xàÿÿÿÿÀŽ â $rÿÿ   ÿ Ž ä $rÿÿ    Ž ä $x   
Ž ä 'z y   Ž Æ xÿ     â $z
 w  ÿŽ â zÿ z    â $r ÿÿ   ÿ Ž â r   ÿ � Æ $z |  Ž â xðÿÿÿÿÀŽ ä xÿ     â $z	 v  Ž ä $x	   
Ž ä $z }  ÿŽ ä $x   ÿ Ž â xÿ   	  â $z p  Ž â/ r   ÿàñ Ä r	   ÿ� â%x
Ò   ÿŽ â r	   ÿ� ä r   ÿ� â $r	ÿÿ    Ž â xÿ     â $rÿÿ   ÿ Ž â xøÿÿÿÿÀŽ â $rÿÿ   ÿ  â r   ÿò� â $x   Ž Ä $x   
Ž â xÀ   ÿàñ â$x   ÿ Ž â xÿ     â $x   Ž âxÿ     â $x   ÿ Ž â rÿ   ÿ  â $x   ÿ Ž â x     â $x   ÿ Ž Ä $t"ÿ   ÿ Ž È %v X  "Ž È %v X  "Ž â r
   ÿàñ Ê $x   	  â r
   ÿàñ È �y$€   • ¨  �y'‚   • ¢  $x   	  Ê �y&    • è �y)   • â r   à‘ â $tÿ$   ÿ Ž Æ r   ÿ$~  â %v Z   Ž â z
 X  à‘ Æ $x!$   ÿŽ â z	 Y  $~  ä r   ÿ€ â $x   !Ž ä �y(    ‘ $ $x     ä �y*   ‘ ( �y!   ™ ( �y#L   ™ ( �y,    • ( �y+H   • " z
 X  ÿàñ È/ z	 Y  ÿä  Ê �y%Ð   • bz ÿ `  pPð È z ÿ a   Sð ä x   ÿàÿ ä x$$T  '    ÈO r$ÿ   $  Ê $x$   ÿ Ž â x&&T  )    È� x0000ÿÀŽ ä xÿ   &  äx$0000ÿÀŽ ä x&øŽ ä xøŽ ä x€€€€ÿüŽ Ä x€€€€ÿüŽ ä xàßßßÿàÿ ä xàßßßÿàÿ ä x    Ž ä x    Ž ä/ x˜º  ÿ    Ä x˜º  ÿ    ä x€€€€Ž äx˜º  ÿ    ä x`Ž ä x˜º  ÿ    ä x€€€€Ž Ä r   ÿüŽ ä x`Ž ä x(€ˆ  ÿ    â&r!   ÿ   â r   ÿüŽ ä x**€ˆ  ÿ    â $r   ÿŽ ä &r#   ÿ   Ä $rÿÿ   * Ž â Er)        â 0r$ÿ,    A   ä $r   ÿŽ ä 0r&ÿ+    A   â #r)$)   ÿ   ä Er'        â G‰  @    € ê %v `  "Ž È %v `  "Ž â r
   ÿàñ Ê $x   	  â r
   ÿàñ È �y"€    ¨ �y+‚    ¢ $x   	  Ê �y*     è  �y-    â  z
 `  à‘ È z	 a  $~  Ê �y(     " z
 `  ÿàñ Æ �y,    " z	 a  ÿä  Ê �yÐ    " x""T  +    ÈO r"ÿ   "  È x"0000ÿÀŽ â$x"   ÿ Ž â x**T  -    È� x0000ÿÀŽ ä xÿ   *  äx*øŽ ä xøŽ ä x€€€€ÿüŽ ä/ x€€€€ÿüŽ Ä xàßßßÿàÿ ä xàßßßÿàÿ ä x    Ž ä x    Ž ä x˜º  ÿ    ä x˜º  ÿ    Ä x˜º  ÿ    ä x€€€€Ž ä x"˜º  ÿ    ä x`Ž ä x€€€€Ž äx"`Ž Ä r   ÿüŽ â 0rÿ    A   âr   ÿüŽ Æ &r!!   ÿ   ä &r##   ÿ   ä $r!(!   ÿŽ ä $r#,#   ÿŽ Æ Er! !       ä Er# #       Æ #r!$!   ÿ   È #r!&#   !   È #r !       â |    pbð â 0r%ÿ%    A   âx

H  ÿàó â #r'&'   )   â x    ÿàÿ Æ #r%'      â $r	ÿÿ   	Ž  Ê G‰  Ð÷ÿÿÿÿƒ ê Ay        € ê r  ÿ   pRð âEy  €     € â r  ÿ   pRò Ö G‰  P     € ê z ÿ `  pPð â $x     Ž Æ z ÿ a   Sð ä ˆs €ÿÿ H   ö ˆ     H   ä Ay        € ê {          ì M        € ê z ÿ `  pPð â „y     H   & z ÿ a   Sð Ú „‰  €   H   ¨ „‰	    H   â !r       È !‚         ÈO !‚ 	       â� G‰        € ê „y  €  H   ( „y €   H   ¨ „y    H   è „y    H   ( „y	 €  H   b !r          â !r       ÆO !r         â� !r       Æ!r  	       ä/ ‰      " z ÿ `  pPð ä r ÿ   pRò ä z ÿ a   Sð â !r       Ê ‰       $ !r         Ê ‰	  €    $ !r	 	       Ê ‰	 @    $ !r	       Ê ‰      $ !r
       â G‰        € ê ‰        $ !r          Ê ‰       $ !r        Ê ‰ €    $ !r       Ê ‰ @    $ !r       Ê ‰      $ !r        ä M        € ê !r
       âG‰  �    € ê $t ÿ   ÿ Ž â !r        È z   h  pð Ú G	  �    € ê ¹z  h      ä �xÿÿÿÿ?àÿ È Œx    pð È ‡x       È ™x   ?  Ø »z  €    ä ™x?    Ô Xy Àìÿÿÿÿƒ ê  x ;ª¸¿  A  Ì s          $ !t   €?    Ì s        $  r     A  È  r     A  â Gy  `    € ê  x '7=  A  â x*BL?  A  Æ #t   €?    È  r      A  â  x    ?  A  Ê s     $   $ !t  €?    È  r       A  È  r      A  â Gy  À    € ê $v ÿ h  ÿ Ž Ê x     pRð Ú G‰  à     € ê x     pRð Ú G	        € ê 	z  i    � ä 	z i    � Æ  x ;ª¸¿  A  â 	z i €  � Ê s        $ !t  €?    Ì s        $  r       A  È  r      A  â Gy  Ð     € ê  r     A  â Gy  °     € ê 	x   à@  � ä 	x  à@  � Æ  x #ÛÙ?  A  â 	x  àÀ  � Æ  x;ª¸¿  A  ä !t  €?    È s        $ !t  €?    Ì s        $  r       A  È  r      A  ä y      &   $ $z x  Ž ä? $tÿ   ÿ Ž ä $z ~  Ž È %v j   Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž ä $rÿÿ    Ž È % \   Ž Ê �	    ™ b $ÿÿ   ÿ Ž â ¹z  l      â Ey  à    € â y      "   b '
  m   Ž â ™x?    ä y      !   ¢ $‚ÿÿ   ÿ Ž â 
 ÿ n     â $rÿÿ   ÿ Ž â y      '   & $ ÿÿ    
Ž È $
  o  Ž È 'Š  s   Ž â r ÿ   pRò ä/r ÿ   pRô â$x    Ž ÊO x	ÿ     È r		   ÿ(� È xÿ   	  È |    pbö ä Šÿ t     Ö G9  �    € ê x	àÿÿÿÿÀŽ â y      %   " $rÿÿ   ÿ Ž â x   ÿ   â $z }  ÿŽ ä$x   
Ž ä 'z y   Ž Æ x ÿ     â $tÿÒ   ÿ Ž â zÿ z    â $rÿÿ   ÿ Ž â r    ÿ � är     ÿ� ä xðÿÿÿÿÀŽ ä xÿ     Æ $x
   
Ž ä $x   ÿ Ž ä $z |  Ž â xÿ   
  â $z w  ÿŽ ä $z v  Ž âr	
   ÿ� är

   ÿ� â $z p  Ž â r	   ÿò� Ä r   ÿàù ä x	ÿ     Æ $rÿÿ   ÿ ä %xÒ   Ž È %v X  Ž â z X  ÿàñ ä x øÿÿÿÿÀŽ â $rÿÿ    Ž â x Ä   ÿà÷ ä xÿ   
  â $x   
Ž â zÿ Y  þ â $tÿ   ÿ Ž Ä $x   Ž âxÿ     â $x   Ž ä $x   ÿ Ž â x     â $x   ÿ Ž â xÿ     â %v  € Ž È $x   ÿ Ž ä r   ÿàñ Ê $x     Ê �y    • ¨  �y!   • ¤  $rÿÿ    Ž ä $rÿÿ    Ž È %x   Ž â r   à± â �yÐ   • è  �y    • ( �y#   • " r
   ÿd~  â $tÿ$   ÿ Ž È %v Z   Ž È $x%
$   ÿŽ â r
   ÿ€ â $rÿÿ     Ž ä/ $r	ÿÿ    Ž â r%   ÿàÿ È �y"üÿÿ ‘ b $x     Æ �y     ‘ è �y 
   ™ è �y$
L   ™ è �y%    • è  �yH   • â x   ÿàÿ È |    pbð ä xH  ÿàù ä x    ÿàÿ Æ $rÿÿ    â x!T  !    ÈO r!ÿ   !  Ê $x!   ÿ Ž âx
!0000ÿÀŽ È/ x0000ÿÀŽ ä xT  #    Èx!øŽ äx#ÿ     ä x!€€€€ÿüŽ ä x
#øŽ ä xàßßßÿàÿ ä x€€€€ÿüŽ Ä x
!    Ž ä xàßßßÿàÿ ä x

˜º  ÿ    ä x    Ž ä x!˜º  ÿ    ä x
€€€€Ž Ä x˜º  ÿ    ä x˜º  ÿ    ä x

`Ž ä x€€€€Ž äx`Ž ä r
   ÿüŽ Ä x""€ˆ  ÿ    är   ÿüŽ â &r     ÿ   ä� $rÿÿ   " Ž â x
 €ˆ  ÿ    â &r$   ÿ   ä $r    ÿŽ ä $r

   ÿŽ Ä 0r ÿ%    A   â Er        â 0rÿ    A   â Er
 
       Æ #r    ÿ   â x H  ÿà÷ â 0rÿ    A   ä #r
      ä $rÿÿ   	Ž ä #r      â G‰  `ûÿÿÿÿƒ ê Ay        € ê ˆ €ÿÿ H   è {          ì M)        € ê „y      H   " r ÿ   pRð Æ „y €   H   h „y    H   ¢!r         È !r         È/ !r         ÊO ‰       $ !r        Ê ‰      $ !r       Ê ‰ €    $ !r       ä y      %   ( ‰ @    b $z x  Ž â !r       Ê/ ‰	      b  M	        € ê $tÿ   ÿ Ž â !r		       â/ $z ~  Ž È %v j   Ž Ê †y 	    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž È % \   Ž Ê �		    ™ ¢  $rÿÿ   ÿ Ž â Ey  °    € ä y       "   b 'Š s   Ž Æ y      !   b '

 m   Ž â x      ÿ  ä y      '   $ 
ÿ n  
  ä rÿ    ÿü‚ â/ $x     Ž Ê xÿ     È r   ÿ(� ä Š	ÿ t    ä y      %   d r
	       âO G  °    € ê z ÿ `  pPô ä z ÿ ^  pPò ä z ÿ b  p ö Ä z ÿ a   Sô ä z ÿ _  Sò ä z ÿ c  0't Ö $ ~  ÿŽ âÿ     â/$ª ~  ÿŽ ä $
 x  ÿŽ â    àë â$ª
 x  ÿŽ â ¢   àÇ ä¨
ÿ     ä ÿÿ   Äþ Ä    ÿàý ä ¢   ÿàû ä ¢ÿÿ   
„þ â $ÿÿ    â 
 ^  ÿ† ä ª b  ÿˆ â $¢ÿÿ   Ž â  _  � Ä x      ÿ  ä ª c   È �
     h �©     d Ay        € ê ¹z  l      â xÿ     â ™x?    â $ÿÿ   
Ž â Ey  @    € â $r ÿÿ   ÿ Ž ä $

 o  Ž â|    pbò â $‚
ÿÿ    Ž Ä $rÿÿ   ÿ Ž Ô G  Ð
    € ê xàÿÿÿÿÀŽ â $rÿÿ   ÿ Ž ä $rÿÿ    Ž ä $x   
Ž ä 'z y   Ž Æ xÿ     â $z
 w  ÿŽ â zÿ z    â $r ÿÿ   ÿ Ž â r   ÿ � Æ $z |  Ž â xðÿÿÿÿÀŽ ä xÿ     â $z	 v  Ž ä $x	   
Ž ä $z }  ÿŽ ä $x   ÿ Ž â xÿ   	  â $z p  Ž â/ r   ÿàñ Ä r	   ÿ� â%x
Ò   ÿŽ â r	   ÿ� ä r   ÿ� â $r	ÿÿ    Ž â xÿ     â $rÿÿ   ÿ Ž â xøÿÿÿÿÀŽ â $rÿÿ   ÿ  â r   ÿò� â $x   Ž Ä $x   
Ž â xÀ   ÿàñ â$x   ÿ Ž â xÿ     â $x   Ž âxÿ     â $x   ÿ Ž â rÿ   ÿ  â $x   ÿ Ž â x     â $x   ÿ Ž Ä $t"ÿ   ÿ Ž È %v X  "Ž È %v X  "Ž â r
   ÿàñ Ê $x   	  â r
   ÿàñ È �y$€   • ¨  �y'‚   • ¢  $x   	  Ê �y&    • è �y)   • â r   à‘ â $tÿ$   ÿ Ž Æ r   ÿ$~  â %v Z   Ž â z
 X  à‘ Æ $x!$   ÿŽ â z	 Y  $~  ä r   ÿ€ â $x   !Ž ä �y(    ‘ $ $x     ä �y*   ‘ ( �y!   ™ ( �y#L   ™ ( �y,    • ( �y+H   • " z
 X  ÿàñ È/ z	 Y  ÿä  Ê �y%Ð   • bz ÿ `  pPð È z ÿ a   Sð ä x   ÿàÿ ä x$$T  '    ÈO r$ÿ   $  Ê $x$   ÿ Ž â x&&T  )    È� x0000ÿÀŽ ä xÿ   &  äx$0000ÿÀŽ ä x&øŽ ä xøŽ ä x€€€€ÿüŽ Ä x€€€€ÿüŽ ä xàßßßÿàÿ ä xàßßßÿàÿ ä x    Ž ä x    Ž ä/ x˜º  ÿ    Ä x˜º  ÿ    ä x€€€€Ž äx˜º  ÿ    ä x`Ž ä x˜º  ÿ    ä x€€€€Ž Ä r   ÿüŽ ä x`Ž ä x(€ˆ  ÿ    â&r!   ÿ   â r   ÿüŽ ä x**€ˆ  ÿ    â $r   ÿŽ ä &r#   ÿ   Ä $rÿÿ   * Ž â Er)        â 0r$ÿ,    A   ä $r   ÿŽ ä 0r&ÿ+    A   â #r)$)   ÿ   ä Er'        â G‰  @    € ê %v `  "Ž È %v `  "Ž â r
   ÿàñ Ê $x   	  â r
   ÿàñ È �y"€    ¨ �y+‚    ¢ $x   	  Ê �y*     è  �y-    â  z
 `  à‘ È z	 a  $~  Ê �y(     " z
 `  ÿàñ Æ �y,    " z	 a  ÿä  Ê �yÐ    " x""T  +    ÈO r"ÿ   "  È x"0000ÿÀŽ â$x"   ÿ Ž â x**T  -    È� x0000ÿÀŽ ä xÿ   *  äx*øŽ ä xøŽ ä x€€€€ÿüŽ ä/ x€€€€ÿüŽ Ä xàßßßÿàÿ ä xàßßßÿàÿ ä x    Ž ä x    Ž ä x˜º  ÿ    ä x˜º  ÿ    Ä x˜º  ÿ    ä x€€€€Ž ä x"˜º  ÿ    ä x`Ž ä x€€€€Ž äx"`Ž Ä r   ÿüŽ â 0rÿ    A   âr   ÿüŽ Æ &r!!   ÿ   ä &r##   ÿ   ä $r!(!   ÿŽ ä $r#,#   ÿŽ Æ Er! !       ä Er# #       Æ #r!$!   ÿ   È #r!&#   !   È #r !       â |    pbð â 0r%ÿ%    A   âx

H  ÿàó â #r'&'   )   â x    ÿàÿ Æ #r%'      â $r	ÿÿ   	Ž  Ê G‰  Ð÷ÿÿÿÿƒ ê Ay        € ê r  ÿ   pRð âEy  €     € â r  ÿ   pRò Ö G‰  P     € ê z ÿ `  pPð â $x     Ž Æ z ÿ a   Sð ä ˆs €ÿÿ H   ö ˆ     H   ä Ay        € ê {          ì M        € ê z ÿ `  pPð â „y     H   & z ÿ a   Sð Ú „‰  €   H   ¨ „‰	    H   â !r       È !‚         ÈO !‚ 	       â� G‰        € ê „y  €  H   ( „y €   H   ¨ „y    H   è „y    H   ( „y	 €  H   b !r          â !r       ÆO !r         â� !r       Æ!r  	       ä/ ‰      " z ÿ `  pPð ä r ÿ   pRò ä z ÿ a   Sð â !r       Ê ‰       $ !r         Ê ‰	  €    $ !r	 	       Ê ‰	 @    $ !r	       Ê ‰      $ !r
       â G‰        € ê ‰        $ !r          Ê ‰       $ !r        Ê ‰ €    $ !r       Ê ‰ @    $ !r       Ê ‰      $ !r        ä M        € ê !r
       âG‰  �    € ê $t ÿ   ÿ Ž â !r        È z   h  pð Ú G	  �    € ê ¹z  h      ä �xÿÿÿÿ?àÿ È Œx    pð È ‡x       È ™x   ?  Ø »z  €    ä ™x?    Ô Xy Àìÿÿÿÿƒ ê  x ;ª¸¿  A  Ì s          $ !t   €?    Ì s        $  r     A  È  r     A  â Gy  `    € ê  x '7=  A  â x*BL?  A  Æ #t   €?    È  r      A  â  x    ?  A  Ê s     $   $ !t  €?    È  r       A  È  r      A  â Gy  À    € ê $v ÿ h  ÿ Ž Ê x     pRð Ú G‰  à     € ê x     pRð Ú G	        € ê 	z  i    � ä 	z i    � Æ  x ;ª¸¿  A  â 	z i €  � Ê s        $ !t  €?    Ì s        $  r       A  È  r      A  â Gy  Ð     € ê  r     A  â Gy  °     € ê 	x   à@  � ä 	x  à@  � Æ  x #ÛÙ?  A  â 	x  àÀ  � Æ  x;ª¸¿  A  ä !t  €?    È s        $ !t  €?    Ì s        $  r       A  È  r      A  ä y      &   $ $z x  Ž ä? $tÿ   ÿ Ž ä $z ~  Ž È %v j   Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y       &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž ä $rÿÿ     Ž È % \   Ž Ê �	    ™ b ¹z  l      â Ey       € â ™x?    â y      "   b $rÿÿ   ÿ Ž Æ y      !   ¢ $ÿÿ   ÿ Ž ä $ÿÿ    Ž â y	      %   & '
 m   Ž È $‚ÿÿ   ÿ Ž â 
ÿ n    Ê $
ÿÿ   
Ž â r ÿ   pRò Æ/ $

 o  Ž â r ÿ   pRô â$x    Ž äO $‚ÿÿ    Ž Æ xÿ     â 'Š s   Ž Æ r
   ÿ(� â $rÿÿ   ÿ Ž ä $€   Ž â xÿ   
  â $x		   ÿ Ž Æ |    pbö ä Šÿ t    ä x      ÿ  Ò G9  `    € ê y      '   " x

àÿÿÿÿÀŽ â $z w  ÿŽ È $x   

Ž ä $r
ÿÿ   ÿ Ž ä $x   ÿ Ž â xÿ     È r   ÿ � är   ÿ� ä xðÿÿÿÿÀŽ ä xøÿÿÿÿÀŽ â 'z
 y  
 Ž È $z }  ÿŽ â zÿ z  
  â $z
	 p  Ž ä $x   
Ž ä $z |  
Ž â r
   ÿàñ ä xÿ     â $z v  Ž âxÿ     â $vÿ X  ÿ Ž Ä $x   ÿ Ž â r   ÿ� â%xÒ   ÿŽ â r   ÿ� ä xÿ     â $x   
Ž â r   ÿò� â $rÿÿ   ÿ  ä $x   Ž âxÿ     â $x   Ž Ä $x   ÿ Ž â x     â $x   ÿ Ž â r   ÿàñ â$vÿ Y  ÿ Ž â xÿ     â $rÿÿ   ÿ Ž â r   ÿ  â $x   ÿ Ž ä %x   Ž È %x   Ž Ê r&   ÿàñ Ê $x'     â r    ÿàñ È �y<&€   • ¨  �y&‚   • ¢  $v#ÿ p  ÿ Ž ä $x!     ä %xD#Ò   &Ž Æ �y)     • â %x.#Ò    Ž Æ �y:    • è �y9D€   • h	 �y$D‚   • h	 �y7.    • h �y,.   • b rB
   à± â $tCÿ$   ÿ Ž Æ r   ÿd~  â %vBB Z  C Ž â r0   ÿàñ Æ $x$   ÿŽ ä $x1     â r@B   ÿ€ â %xD#Ò   DŽ Æ�y+0Ä   ‘ b	 $xCC   Ž ä %x.#Ò   .Ž â�y5D€   • f %x>#Ò   0Ž â �y4D‚   • f $xAC     â �y6.    • b %x #Ò   DŽ Æ/ �y3.   • h �y(>À   ‘ h �y%@   ™ h �y'>Ä   ‘ h �y*@L   ™ h �y2 €   • h �y80À   ‘ b	 %x@#Ò   .Ž Æ �y/ ‚   • h �yB    • h �y1@    • (�y0@   • " %x #Ò   >Ž Æ/ �y&BH   • & %xD#Ò    Ž â �y. À   ‘ " r@   ÿàñ Æ �y Ä   ‘ $ $xA     ä �yDÀ   ‘ $ %x>#Ò   @Ž ä�y-DÄ   ‘ ( �y;@Ð   • " %x #Ò   >Ž Æ/ �y>>Ð   • & %x"#Ò    Ž ä �y  Ð   • ( �y#"Ð   • " x   ÿàÿ È |    pbð ä xH  ÿà÷ ä x    ÿàÿ Æ $rÿÿ   Ž â x<T      ÈO rÿ     Ê $x<   ÿ Ž â x))T  :    È� x<<0000ÿÀŽ ä x:ÿ   )  äx)<)øŽ ä x99T  $    äx0000ÿÀŽ ä r9ÿ   9  Ä x")€€€€ÿüŽ ä/ x,7T  ,    ä x:øŽ ä x:90000ÿÀŽ â$x99   ÿ Ž â x!ÿ   ,  ä x""àßßßÿàÿ Ä x$€€€€ÿüŽ ä x!:!øŽ ä x7)    "Ž ä x990000ÿÀŽ ä x$$àßßßÿàÿ ä x:!€€€€ÿüŽ Ä x77˜º  ÿ    ä x<)˜º  ÿ    ä x99,øŽ ä x)    $Ž ä x::àßßßÿàÿ ä x"7€€€€"Ž Ä x,9€€€€ÿüŽ ä x77<`Ž ä x))˜º  ÿ    ä x<˜º  ÿ    ä x!    :Ž ä x,,àßßßÿàÿ Ä x$)€€€€$Ž äx))<`Ž ä x˜º  ÿ    ä x<!˜º  ÿ    ä x!9    ,Ž ä x:€€€€:Ž Ä x<`Ž ä x!!˜º  ÿ    ä x<9˜º  ÿ    ä x,!€€€€,Ž äx!!<`Ž ä x55T  4    Ä r"7"   ÿüŽ ä r,!,   ÿüŽ ä r:   ÿüŽ ä r))$   ÿüŽ ä x36T  3    ä r$ÿ   5  â &r7"%   ÿ   âx!(€ˆ  ÿ    â &r,,%   ÿ   â x''€ˆ  ÿ    â &r(*   ÿ   â x"+€ˆ  ÿ    Ä x+ÿ   3  ä x4$0000ÿÀŽ ä x/2T  /    â $r!,   ÿŽ ä $r!'(   ÿŽ â x(4+øŽ â $x$$   ÿ Ž â r/ÿ   /  â &r))*   ÿ   â x01T  0    Äx'(€€€€ÿüŽ â $x+/   ÿ Ž âx$$0000ÿÀŽ â $r"")   ÿŽ â x,ÿ   0  ä x//0000ÿÀŽ ä x)'àßßßÿàÿ ä x3$3øŽ Ä x'+0000ÿÀŽ ä x$/,øŽ ä x+(    )Ž ä x,3€€€€ÿüŽ ä x''0øŽ ä x0+˜º  ÿ    Ä x1(˜º  ÿ    ä x,,àßßßÿàÿ ä x+'€€€€ÿüŽ ä x(0€€€€)Ž ä x2$€€€€ÿüŽ ä x/3    ,Ž Ä x)01`Ž ä x0+àßßßÿàÿ ä x+2àßßßÿàÿ ä x//˜º  ÿ    ä x23˜º  ÿ    ä x1'    0Ž Ä x,/€€€€,Ž äx//2`Ž ä x11˜º  ÿ    ä x2'˜º  ÿ    ä x3$    +Ž ä x01€€€€0Ž Ä x112`Ž ä x43˜º  ÿ    ä x'$˜º  ÿ    ä r,/,   ÿüŽ ä r010   ÿüŽ ä x+4€€€€+Ž Ä r))(   ÿüŽ ä x..€ˆ  ÿ    ä x44'`Ž â &r,,%   ÿ   âx88€ˆ  ÿ    â &r00%   ÿ   â x€ˆ  ÿ    â &r$)*   ÿ   â x%€ˆ  ÿ    â $rÿÿ   . Ž â r+4+   ÿüŽ â $r887   ÿŽ â x'-€ˆ  ÿ    â $r,,   ÿŽ Ä $r$   ÿŽ ä &r*+*   ÿ   ä $r$%0   ÿŽ â Er8 8       â 0rÿ    A   â Er "       â $r''*   ÿŽ â Er(        ä Er" ,       Ä Er$ $       â 0r&ÿ&    A   â Er,        â #r8   ÿ   âEr* !       â #r(   ÿ   âEr( '       â #r!"   ÿ   â#r$   ÿ   â #r&      â0r;ÿ;    A   â #r&*      â0rÿ>    A   â #r!&,   !   â0r ÿ     A   â #r&(      â 0r#ÿ#    A   â #r;      â #r      â #r !      Ä #r#      â G‰  `òÿÿÿÿƒ ê Ay        € ê ˆ  þÿ H   è ˆ €þÿ H   è ˆ  ÿÿ H   è ˆ €ÿÿ H   è {          ì M)        € ê „y     H   " z 	 r  p`ô â $z   x  	Ž Ä „y €   H   b r ÿ   pVt Æ „y    H   ¨ „y    H   è„y €  H   ( „y    H   ( „y
    H   ( „y €  H   ( „y €  H   ( „y    H   ( „y €  H   " !r       Æ „y €  H   " !r       â/ !r       äO y      '   b !r       â� !r       â!r       ä !r
       â !r       È ‰      ¢ !r       â $z  ~   Ž â/ !r       Æ ‰      b !r       Æ ‰      â !r       Ê ‰      " !r       âO !r       â/ !r       È� ‰      h ‰      ¢ !r       Æ ‰      ( ‰      â !r       â/ !r       ÈO ‰
 €    b !r       Æ ‰ €    " !r       Æ� ‰ €    ¨ ‰ €    â !r

       â/ !r       È ‰
 @    " x	   ÿàÿ â !r       äO ‰ @    b x	   ÿàÿ â !r       ä� ‰ @    ¢ z  r  p`ö ä z  r  p`ò â ‰ @    â x    pVö Ä x    pVò  â !r
       â x
	   ÿàÿ â $t	ÿ   ÿ Ž â !r       ä/ ‰      " z 
 r  p`ð â !r       äO ‰      b x    pVp  â !r       Ä� ‰      ¨ ‰      â !r       â !r       â/ %v  j  	 Ž â !r       ÈO †©     â !r       Æ� †¹    è †‰    â M        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      &   " z ÿ \  pPò â ¹z  F   
   Æ z ÿ ]  Sò Ú         Ê % \   Ž Ê �    ™ ¢  r ÿ       â Ey  0    € ä y      "   d ' m   Ž äy      %   æ ÿ n    â y       !   & ÿ   ÿáÿ â y	      '   " 'š s   Ž È $ o  Ž â ’        ä r ÿ   pRð ä/ x   ÿ   ä� x     pDp  är    ÿàÿ är    ÿ(Ž Ä z  r  pdp  ä r ÿ       ä x
ÿ     ä šÿ t    ä r ÿ       ä r     € ÆO G	  À    € ê z ÿ `  pPò ä z ÿ ^  pPð ä z ÿ b  p ô ä z ÿ a  Sò ä z ÿ _   Sð ä z ÿ c   'ò Ö $
	 ~  ÿŽ âÿ     â$š	 ~  ÿŽ ä $
 x  ÿŽ â     àë â$š x  ÿŽ â ’    àÇ ä ˜ÿ     ä    ÿàõ Ä ÿ   ÿÄþ ä ’   ÿàû ä ’ÿ   ÿ„þ ä ÿ   ÿä ä 
 ^  ÿ„ ä ’ÿ   ÿäÿ Ä š b  ÿ† ä 
 _   ä r ÿ       ä r ÿ       ä š c  � ä �	     h �™     d Ay        € ê r
   ÿ(� â ¹z  l      â $z	 }  ÿŽ â ™x?    â xÿ     â $z w  ÿŽ â r ÿ       â Ey  @     € â rE ÿ       â x
      ÿ  â |    pbð â 'z	 y   Ž â r   ÿàó â x      ÿ  â x      ÿ  â r	 ÿ       Ä rÿÿ   ÿäÿ  Ì G	  �    € ê xàÿÿÿÿÀŽ ä zÿ z    ä r  €ÿàÿ ä r        ä x	ÿ     ä x7   ÿ   â $z
 p  Ž âr	   ÿ � â$z |  Ž â r		   ÿ� â $z |  
Ž â xðÿÿÿÿÀŽ â $z v  Ž â x		øÿÿÿÿÀŽ â $z v  Ž â xÿ     â %x&7   ÿŽ â r  €ÿàÿ Ä r	  €ÿàÿ â %x4Ò   ÿŽ â x   ÿ   â x      ÿ  â x
ÿ     ä r   ÿàÿ â %x67   4Ž â r
   ÿ� är
   ÿ� Ä x9   ÿ   ä x
   ÿàÿ âx      ÿ  â x   ÿàÿ â %x,9   ÿŽ â x   ÿàÿ ä r   ÿò� â $z3 p  Ž â x!ÿ     â $z/
 p  Ž âxÿ     â $z1 p  Ž â r4   ÿà÷ â %x89   4Ž â z ÿ `  pPð â x
      ÿ  â xÿ     â %x"3Ò   &Ž â x6   ÿàó Ä r!   ÿŽ â %x$/Ò   &Ž â x8‚   ÿàõ ä xÄ   ÿàù â %x(/Ò   ,Ž âz  `      ä z  a      â %x*1Ò   ,Ž â z  X      Ä z  Y      â %x&1Ò   &Ž â r	 ÿ       ä rE ÿ       â %x,3Ò   ,Ž â z ÿ a   Sð ä x   ÿ   â %x./Ò   Ž â x     Ä rÿ7   ÿäÿ  â %x01Ò   Ž â r ÿ9   ÿä Æ %x23Ò   Ž â r5   ÿäÿ ä x!   ÿ   ä x!ÿ     ä r@ÿ   ÿä ä rB8   ÿàõ är:6   ÿàó ä rC9   ÿä Ä r<   ÿàõ äz  p      â �y\B€   • ¢  r=    ÿä ä r;7   ÿäÿ  â �yFB‚   • ¢  r>   ÿà÷ â %x<Ò   <Ž ä �y]:    • â r?   ÿäÿ Æ �yV:   • è �y[<þÿÿ • è �yZ<    • â %x>Ò   >Ž Ê �yD>þÿÿ • è �yM>    • â rA   à£ ä xH $       ä rB!   ÿDþ  ä rR   ÿàó â%v:A Z  H Ž â/ rN   ÿà÷ Æ $xAB$   ÿŽ â rS   ÿäÿ  ärO@   ÿäÿ ä rP:   ÿ‚ â �yIRÀ   ‘ â  r;;A   ÿàÿ â %xNÒ   NŽ ä �yWRÄ   ‘ â  rQ;   ÿäÿ  Æ �yYNüÿÿ ‘ è �yAP   ™ â  %x<Ò   <Ž Æ�yBPL   ™ â %x>Ò   >Ž Æ �yXN    ‘ è �yU<þÿÿ • è �yJ<    • è �yS:    • è �yC:H   • è  �yT>þÿÿ • è �yK>    • â %x:Ò   <Ž Ê �yQ:þÿÿ • "%x>Ò   >Ž Æ/ �yP:    • ( �yR>þÿÿ • ( �yG>    • " %xNÒ   NŽ â r<4   ÿàó È r=5   ÿäÿ  â �yHN    ‘ " %x>Ò   NŽ È/ %x:Ò   <Ž â�yL>üÿÿ ‘ ( �yNNüÿÿ ‘ ( �y<<Ð   • ( �yO>    ‘ (# �y=:Ð   • " %x>Ò   :Ž Ì/ %x:Ò   >Ž ä �y>Ð   • ( �y::Ð   • " x\\T  F    ÈO r\ÿ   \  ä xV]T  V    ä� x]\0000ÿÀŽ äxFÿ   V  ä xZ[T  Z    ä x\\   ÿ   Ä rZÿ   Z  ä xF]FøŽ ä x;\0000ÿÀŽ ä/ x>Z   ÿ   ä x?F€€€€ÿüŽ ä xDDT  M    Ä x\;VøŽ ä x[>0000ÿÀŽ ä x??àßßßÿàÿ ä x]ÿ   D  ä xZZ0000ÿÀŽ ä x;\€€€€ÿüŽ Ä x[[DøŽ ä x>F    ?Ž ä x]Z]øŽ ä xMF˜º  ÿ    ä x;;àßßßÿàÿ ä xF[€€€€ÿüŽ Ä xV>˜º  ÿ    ä x>]€€€€ÿüŽ ä xZ\    ;Ž ä xFFàßßßÿàÿ ä xDV€€€€?Ž ä x>>àßßßÿàÿ Ä xZZ˜º  ÿ    ä x\\˜º  ÿ    ä x?[    FŽ ä xMVM`Ž ä xV]    >Ž ä x;Z€€€€;Ž Ä x\Z\`Ž ä x??˜º  ÿ    ä xZ[˜º  ÿ    ä xVV˜º  ÿ    ä x]]˜º  ÿ    ä xF?€€€€FŽ Ä x??Z`Ž ä x>V€€€€>Ž äx]V]`Ž ä r\\;   ÿüŽ ä rMMD   ÿüŽ ä rF?F   ÿüŽ â &r\\A   ÿ   â r]]>   ÿüŽ Ä x;I€ˆ  ÿ    â &rMMB   ÿ   â x>W€ˆ  ÿ    â &rFFA   ÿ   â x?Y€ˆ  ÿ    â $r\;\   ÿŽ â xDX€ˆ  ÿ    â $r;>M   ÿŽ ä &r]]B   ÿ   ä $r>?F   ÿŽ â xJUT  J    â $rDD]   ÿŽ Ä 0rSÿS    A   â Er> >       ä xKTT  K    ä rJÿ   J  â 0rCÿC    A   â Er\ \       ä Er? D       â #rDS>   ÿ   â xFÿ   K  Ä xIJ0000ÿÀŽ äxPQT  P    â#r\S\   ÿ   â Er; ;       â #rDC?   D   â xJJ   ÿ   ä x?IFøŽ ä rPÿ   P  Ä xRRT  G    â #r>C;   \   â xJJ0000ÿÀŽ ä x;?€€€€ÿüŽ ä xFP   ÿ   äxGÿ   R  ä xPP0000ÿÀŽ Ä xKJKøŽ ä xJ;àßßßÿàÿ ä xMF0000ÿÀŽ ä x;PGøŽ ä xGK€€€€ÿüŽ ä xFMRøŽ Ä xI?    JŽ ä xPGàßßßÿàÿ ä xMF€€€€ÿüŽ ä xII˜º  ÿ    ä xR?˜º  ÿ    ä xGK    PŽ Ä xQ;€€€€ÿüŽ ä xMMàßßßÿàÿ ä x?I€€€€JŽ äxJIR`Ž ä xGG˜º  ÿ    ä xRK˜º  ÿ    Ä xTQàßßßÿàÿ ä xIF    MŽ ä xPG€€€€PŽ äxGGR`Ž ä xK;    TŽ äxV;˜º  ÿ    Ä xRI˜º  ÿ    ä x;F˜º  ÿ    ä xKK˜º  ÿ    ä xMR€€€€MŽ äxRR;`Ž ä xTK€€€€TŽ Ä xKKV`Ž ä rPGP   ÿüŽ ä rRRM   ÿüŽ ä rIJ?   ÿüŽ â &rPPA   ÿ   ârKKT   ÿüŽ â &rRRA   ÿ   â x;N€ˆ  ÿ    Ä xGL€ˆ  ÿ    â &rFIB   ÿ   âx?H€ˆ  ÿ    â &rKKB   ÿ   â xHO€ˆ  ÿ    â $r;;P   ÿŽ ä $rGGR   ÿŽ ä $r??F   ÿŽ â ErJ ;       â $rKHK   ÿŽ Ä 0rFÿ<    A   â Er< G       â 0rHÿ=    A   â Er? ?       â #rJSJ   ÿ   âErI K       â #r<S<   ÿ   â #r	HD   	   â #rEF>   E   â #rDC?   J   â#rICI   <   â 0rÿ    A   Ä 0rKÿ:    A   â G‰       € ê r>8   ÿàó È r?9   ÿäÿ  ä rF6   ÿàó Æ �yN>€    ¨ �yO>‚    ¢ rG7   ÿäÿ  Ê �yLF     è �yMF    â r<   ÿàó È r=   ÿäÿ  Ê �yH<À    " r:4   ÿàó Æ �yJ<Ä    " r;5   ÿäÿ  Ê �y::Ð    " xNNT  O    ÈO rNÿ   N  È x>N   ÿ   ä xLLT  M    ä� x?>0000ÿÀŽ ä x>ÿ   L  äxGN0000ÿÀŽ ä x??LøŽ Ä xGG>øŽ ä x<?€€€€ÿüŽ ä x;G€€€€ÿüŽ ä/ x<<àßßßÿàÿ ä xF;àßßßÿàÿ ä x;?    <Ž â 0r:ÿ:    A   âx=G    FŽ Ä x;;˜º  ÿ    ä x>?˜º  ÿ    ä x==˜º  ÿ    ä x<;€€€€<Ž ä xLG˜º  ÿ    ä x;;>`Ž Ä xF=€€€€FŽ äx==L`Ž ä r<;<   ÿüŽ ä r==F   ÿüŽ Æ &r;<A   ÿ   ä &r==B   ÿ   ä $r;H;   ÿŽ ä $r=J=   ÿŽ Æ Er< ;       ä Er= =       Æ #r<S<   ÿ   È #r<C=   <   È #r:<      ä G‰  0    € ê r:(   ÿàó È r;)   ÿäÿ  ä r<$   ÿàó Æ �yJ:€    ¨  �yO:‚    ¢  r=%   ÿäÿ  Ê �yH<     è �yM<    â r>.   ÿàó ä xL   ÿàÿ ä r?/   ÿäÿ  Ê �yF>À    "	 $zQL p  Ž Æ �yG>Ä    "	 %x:QÒ   Ž Ì �y::Ð    " xJJT  O    ÈO rJÿ   J  È x<J   ÿ   ä/ xHHT  M    ä� x=<0000ÿÀŽ ä x<ÿ   H  äx?J0000ÿÀŽ äx==HøŽ Ä x??<øŽ ä x<=€€€€ÿüŽ ä x;?€€€€ÿüŽ ä x<<àßßßÿàÿ ä xH;àßßßÿàÿ ä x;=    <Ž Ä xJ?    HŽ ä x;;˜º  ÿ    ä x>=˜º  ÿ    ä x=J˜º  ÿ    ä x<;€€€€<Ž ä xL?˜º  ÿ    Ä x;;>`Ž ä xH=€€€€HŽ äx==L`Ž ä r<;<   ÿüŽ ä r==H   ÿüŽ Æ &r;<A   ÿ   ä &r<=B   ÿ   ä $r;F;   ÿŽ ä $rGG<   ÿŽ Æ Er< ;       â 0r;ÿ:    A   â ErG G       Æ #r<S<   ÿ   È #r<CG   <   È #r;<      ä G‰  0    € ê r:*   ÿàó È r;+   ÿäÿ  ä r<&   ÿàó Æ �yJ:€    ¨  �yO:‚    ¢  r='   ÿäÿ  Ê �yH<     è �yM<    â r>0   ÿàó ä xL   ÿàÿ ä r?1   ÿäÿ  Ê �yF>À    "	 $zQL p  Ž Æ �yG>Ä    "	 %x:QÒ   Ž Ì �y::Ð    " xJJT  O    ÈO rJÿ   J  È x<J   ÿ   ä/ xHHT  M    ä� x=<0000ÿÀŽ ä x<ÿ   H  äx?J0000ÿÀŽ äx==HøŽ Ä x??<øŽ ä x<=€€€€ÿüŽ ä x;?€€€€ÿüŽ ä x<<àßßßÿàÿ ä xH;àßßßÿàÿ ä x;=    <Ž â 0r:ÿ:    A   â xJ?    HŽ Ä x;;˜º  ÿ    ä x>=˜º  ÿ    ä x=J˜º  ÿ    ä x<;€€€€<Ž ä xL?˜º  ÿ    ä x;;>`Ž Ä xH=€€€€HŽ äx==L`Ž ä r<;<   ÿüŽ ä r==H   ÿüŽ Æ &r;<A   ÿ   ä &r<=B   ÿ   ä $r;F;   ÿŽ ä $rGG<   ÿŽ Æ Er< ;       ä ErG G       Æ #r<S<   ÿ   È #r<CG   <   È #r:<      â #rD      â #rKI      â G‰  0    € ê r<,   ÿàó È r=-   ÿäÿ  ä r:"   ÿàó Æ �yH<€    ¨  �yI<‚    ¢  r;#   ÿäÿ  Ê �yF:     è �yG:    â r>2   ÿàó ä xJ   ÿàÿ ä r?3   ÿäÿ  Ê �y>À    "	 $zKJ p  Ž Æ �yD>Ä    "	 %x<KÒ   Ž Ì �y<<Ð    " xHHT  I    ÈO rHÿ   H  È x:H   ÿ   ä/ xFFT  G    ä� x;:0000ÿÀŽ ä x:ÿ   F  äx?H0000ÿÀŽ äx;;FøŽ Ä x??:øŽ ä x:;€€€€ÿüŽ ä x=?€€€€ÿüŽ ä x::àßßßÿàÿ ä xF=àßßßÿàÿ ä x=;    :Ž Ä xG?    FŽ ä x==˜º  ÿ    ä x>;˜º  ÿ    â 0r;ÿ<    A   â xGG˜º  ÿ    ä x:=€€€€:Ž ä xH?˜º  ÿ    Ä x==>`Ž ä xFG€€€€FŽ äxGGH`Ž ä r:=:   ÿüŽ ä rGGF   ÿüŽ Æ &rA:A   ÿ   ä &rGGB   ÿ   ä $rA   ÿŽ ä $rDDG   ÿŽ Æ Er        ä ErD D       Æ #rS   ÿ   È #rCD      È #r
;   
   â x   ÿàÿ ä xH  ÿàõ ä |    pbò ä xH  ÿà÷ ä x    ÿàÿ ä rÿ   ÿä Ä rÿ   ÿäÿ Ê D       À â Gy  ðäÿÿÿÿƒ ê Ay        € ê y      "   " Ey  0    € â r ÿ   pRð är ÿ   pRò Ö G‰  ð     € ê z ÿ `  pPð ä r    ÿ8Ž ä z ÿ a   Sð Æ ˆs E þÿ H   ô ˆƒ 	€þÿ H   è ˆƒ  ÿÿ H   è ˆƒ €ÿÿ H   â G‰  p     € ê ˆs    H   è ˆs 	€þÿ H   è ˆs €  H   è ˆs  ÿÿ H   è ˆs    H   è ˆs €ÿÿ H   è ˆs 
€  H   ä Ay        € ê {          ì M        € ê y      &   ¢ z ÿ `  pPð Æ y      '   "> z ÿ a   Sð â $z x  Ž âO x        Æ $z ~  Ž È %v j   Ž È G	  P    € ê „y      H   "z  r  p`ð Æ „y
     H   b r  ÿ   pVp  Æ „y     H   ¨ „y  €   H   è „y     H   ( „y     H   ( „y  €  H   ( „y     H   ( „y  €  H   ( „y  €  H   ( „y  €  H   " !rE       Æ „y  €  H   " !r

       â/ !r       âO !r	       ä� !r

       â!r       â !r       È ‰      b !r

       â !r       È ‰
      ¢ !r       Æ ‰      â !r       È !r       â !r       È/ ‰      ( ‰      b !r
       âO !r	       È� ‰      ¨ ‰	      â !r       â !r       È/ ‰      " !r       ÆO ‰ €    b !r	       Æ� ‰ €    ¨ ‰ €    â !r       Ê ‰ €    " !r       â/ !r       ÈO ‰ @    b !r       â� x   ÿàÿ ä ‰ @    ¤ z  r  p`ô ä ‰
 @    ä x     pVt â !r       Ê ‰ @    " !r       â/ x   ÿàÿ äx   ÿàÿ â !r       âO z  r  p`ò â ‰      b !r

       ä� x     pVò  â ‰      è ‰	
      ¢ !r       â !r       È/ ‰      " !‚       â !r
	       ÈO †‰     â z  r  p`ð â !r	       â !’       äx     pVp  â !¢		       ä †™    è †© 	   â !r       Ê M	        € ê !r       Ê †y    â My        € ê z  h      È x    pBð Ú G	  @    € ê x    pRð Ú G‰  À    € ê x    pRð Ú G	   $    € ê „y      H   " z  r  p`ð â Ey  p    € ä „y     H   b r  ÿ   pVp  Æ „y     H   ¨ „y     H   è „y     H   ( „y   
  H   ( „y  €   H   ( „y  €  H   è „y  €  H   è „y  €
  H   â !rE       â !r       Æ/ !r       âO !r       Æ� !r       â!r       È ‰      " !r	       Æ ‰      b !r       Æ „y  €  H   â !r       Æ/ ‰      ( ‰      b !r       â !r       È/ ‰ €    ( ‰ €    b !r       Æ „y  €  H   ¢ !r       È !r       â !r       â !r       È/ ‰ @    ( ‰ @    b !r       ÈO !r       â !r       â !r       È/ ‰      ( ‰      b !r       â !r       â/ G	  �     € ê !r       â!r       Æ  x	;ª¸¿  A  Ì s	 	       $ !t	  €?    Ì s        $  r     A  È  r     A  Ê †y     ä Ay        € ê ‰      " Ey  Ð    € æ ‰      h „y     H   ¨ „y     H   è „y     H   è „y     H   â !r       â !r       È/ ‰      ( ‰      b !r       âO !r       â� !r       ä !r       â !r       Ä/ ‰	 €    " !r       Æ ‰ €    h „y   	  H   ¢ !r		       Æ „y     H   â !r       â/ x   ÿàÿ ä ‰	 @    $ z  r  p`ð ä ‰ @    d x     pVp  â !r       âO !r	       â !r       È/ ‰      ( ‰	      b !r       â !r       â !r	       â/ G	  �     € ê !r       â!r       Æ  x;ª¸¿  A  Ì s        $ !t	  €?    Ì s 	       $  r     A  È  r     A  Ê †y    ä Ay        € ê ‰      " Ey       € æ ‰      h „y  €  H   ¨ „y  €	  H   è „y  €  H   ( „y  €  H   â !r       Æ „y  €  H   â !r	       Æ/ „y  €  H   è ‰      ( ‰	      b !r
       ÈO !r       È� !r       â!r       â !r	       È/ ‰ €    ( ‰ €    b !r       â !r       È/ ‰ @    ( ‰ @    b !r       â x   ÿàÿ â !r       Æ/ ‰      " z  r  p`ð Æ ‰      b x     pVp  â !r       â !r       Ö/ G	  �     € ê !r       â!r       Æ  x;ª¸¿  A  Ì s        $ !t	  €?    Ì s
 	       $  r

     A  È  r
     A  Ê †y    ä Ay        € ê ‰      b !r       â x   ÿàÿ Æ !r       â z  r  p`ð Æ !r       â x     pVp  È ‰      " !r	       Ê/ ‰	      b !r       Ê ‰      " !r
	       Ê/ ‰
 €    b !r       Ê ‰ €    " !r
       Ê/ ‰ @    b !r       Ê ‰ @    " !r       Ê/ ‰	      b !r       Ê ‰      ¢  !r		       â/ M	        € ì !r	       â!r       ÆO  x ;ª¸¿  A  â !r       Ê s          $ !t   €?    Ì s        $  r     A  È  r     A  Ê †y    â My        € ê „y      H   " z  r  p`ð â Ey       € ä „y     H   b r  ÿ   pVp  Æ „y     H   ¨ „y     H   è „y     H   ( „y   
  H   ( „y  €   H   ( „y  €  H   è „y  €  H   â !rE       â !r       Æ/ !r       âO !r       Æ� !r       â!r       È ‰      " !r	       Æ ‰      b !r       â !r       È/ ‰      ( ‰      h „y  €  H   ¢ !r       â !r       È/ ‰ €    ( ‰ €    h „y  €  H   â !r       ÈO !r       â !r       ä „y  €
  H   â !r       Æ/ ‰ @    ( ‰ @    b !r       È� !r       â !r       â !r       È/ ‰      ( ‰      b !r       â !r       â !r       â/ G	  À     € ê !r       È x'7=  A  â x*BL?  A  Æ #t	  €?   È  r		     A  â  x   ?  A  â !r       È s	 	    $   $ !t	  €?    È  r     A  È  r     A  Ê †y     ä Ay        € ê ‰      " Ey       € æ ‰      h „y     H   ¨ „y   	  H   è „y     H   è „y     H   â !r       Æ „y     H   â !r       Æ/ ‰      ( ‰      b !r       âO !r       â !r       È/ ‰	 €    ( ‰ €    b !r       Æ „y     H   ¢ !r       È !r       â !r		       â x   ÿàÿ â !r       Æ/ ‰	 @    " z  r  p`ð Æ ‰ @    b x     pVp  â !r       ÈO !r       â !r	       â !r       È/ ‰      ( ‰	      b !r       â !r	       â/ G	  À     € ê !r       È x'7=  A  â x*BL?  A  Æ #t  €?   È  r     A  â  x   ?  A  â !r       È s     $   $ !t  €?    È  r     A  È  r     A  Ê †y    ä Ay        € ê ‰      " Ey  Ð    € æ ‰      h „y  €  H   ¨ „y  €	  H   è „y  €  H   è „y  €  H   â !r       Æ „y  €  H   â !r	       Æ/ ‰      ( ‰	      h „y  €  H   " !r
       ÈO !r       â� !r       â !r	       È/ ‰ €    ( ‰ €    b !r       â!r       â x   ÿàÿ â !r       Æ/ z  r  p`ð â ‰ @    & x     pVp  â ‰ @    b !r       â !r       È/ ‰      ( ‰      b !r	       â !r       â/ G	  À     € ê !r       È x'7=  A  â x*BL?  A  Æ #t  €?   È  r     A  â  x   ?  A  â !r	       È s     $   $ !t
  €?    È  r

     A  È  r
     A  Ê †y    ä Ay        € ê ‰      b !r       â x   ÿàÿ Æ !r       â z  r  p`ð Æ !r       â x     pVp  È ‰      " !r	       Ê/ ‰	      b !r       Ê ‰      " !r
	       Ê/ ‰
 €    b !r       Ê ‰ €    " !r
       Ê/ ‰ @    b !r       Ê ‰ @    " !r       Ê/ ‰	      b !r       Ê ‰      ¢  !r		       â/ M	        € ì !r	       È x'7=  A  â x *BL?  A  Æ #t  €?   â x   ?  A  Æ  r       A  â !r       ÈO !r       â s       $   $ !t   €?    È  r     A  È  r     A  Ê †y    â My        € ê x    pRð Ú G‰  P    € ê x    pRð Ú G	  `    € ê „y      H   " z  r  p`ð â Ey       € ä „y     H   b r  ÿ   pVp  Æ „y     H   ¨ „y     H   è „y     H   ( „y   
  H   ( „y  €   H   ( „y  €  H   è „y  €  H   â !rE       â !r       Æ/ !r       âO !r       Æ� !r       â!r       È ‰      " !r	       Æ ‰      b !r       â !r       È/ ‰      ( ‰      h „y  €  H   ¢ !r       â !r       È/ ‰ €    ( ‰ €    h „y  €  H   â !r       ÈO !r       â !r       ä „y  €
  H   â !r       Æ/ ‰ @    ( ‰ @    b !r       È� !r       â !r       â !r       È/ ‰      ( ‰      b !r       â !r       â !r       â/ G	  À     € ê !r       Ê	z i    � â !r       È  x	;ª¸¿  A  â 	z i    � È 	z i €  � â s	 	       $ !t	  €?    Ì s        $  r     A  È  r     A  Ê †y     ä Ay        € ê ‰      " Ey       € æ ‰      h „y     H   ¨ „y   	  H   è „y     H   è „y     H   â !r       Æ „y     H   â !r       Æ/ ‰      ( ‰      b !r       âO !r       â !r       È/ ‰	 €    ( ‰ €    b !r       Æ „y     H   ¢ !r       È !r       â !r		       â x   ÿàÿ â !r       Æ/ ‰	 @    " z  r  p`ð Æ ‰ @    b x     pVp  â !r       ÈO !r       â !r	       â !r       È/ ‰      ( ‰	      b !r       â !r	       â/ G	  À     € ê !r       Ê	z i    � â !r       È  x;ª¸¿  A  â 	z i    � È 	z i €  � â s        $ !t	  €?    Ì s 	       $  r     A  È  r     A  Ê †y    ä Ay        € ê ‰      " Ey  Ð    € æ ‰      h „y  €  H   ¨ „y  €	  H   è „y  €  H   è „y  €  H   â !r       Æ „y  €  H   â !r	       Æ/ ‰      ( ‰	      h „y  €  H   " !r
       ÈO !r       â� !r       â !r	       È/ ‰ €    ( ‰ €    b !r       â!r       â x   ÿàÿ â !r       Æ/ z  r  p`ð â ‰ @    & x     pVp  â ‰ @    b !r       â !r       È/ ‰      ( ‰      b !r       â !r       â/ G	  À     € ê !r       Ê	z i    � â !r       È  x;ª¸¿  A  â 	z i    � È 	z i €  � â s        $ !t	  €?    Ì s
 	       $  r

     A  È  r
     A  Ê †y    ä Ay        € ê ‰      b !r       â x   ÿàÿ Æ !r       â z  r  p`ð Æ !r       â x     pVp  È ‰      " !r	       Ê/ ‰	      b !r       Ê ‰      " !r
	       Ê/ ‰
 €    b !r       Ê ‰ €    " !r
       Ê/ ‰ @    b !r       Ê ‰ @    " !r       Ê/ ‰	      b !r       Ê ‰      ¢  !r		       â/ M	        € ì !r	       â!r       ÈO 	z i    � â !r       È  x ;ª¸¿  A  â 	z i    � È 	z i €  � â s          $ !t   €?    Ì s        $  r     A  È  r     A  Ê †y    â My        € ê „y     H   " z  r  p`ð Æ „y     H   b r  ÿ   pVp  Æ „y  €   H   ¨ „y      H   è „y   
  H   ( „y  €  H   ( „y     H   ( „y      H   ( „y     H   ( „y  €  H   ( „y  €  H   " !r       Æ „y     H   " !r       Æ/ „y   	  H   b !r		       ÆO „y     H   ¢ !rE       Æ� „y  €  H   â !r       Æ„y  €
  H   " !r       Æ „y"     H   " !r       Æ „y$     H   " !r         Æ „y  €  H   " !r       Æ „y  €  H   " !r		       Æ „y&  €  H   " !r       â x   ÿàÿ ä „y(  €	  H   â !r       â z  r  p`ô ä „y*  €  H   â !r        â/ x     pVt Ä „y,  €  H   â !r       ÆO ‰      " !r		       â� !r       ä‰      b !r"       Æ ‰	      ¢ !r$       Æ ‰      â !r       Æ ‰      " !r
       Æ ‰      " !r&       â !r       â !r(       â !r       â/ !r**       â !r		       ÆO ‰      " !r,       â !r       ä� ‰*      b !r       Æ‰      ¢ x   ÿàÿ â !r       ä ‰      â z  r  p`ò ä x   ÿàÿ â ‰
	      " x     pVò  äz  r  p`ö â ‰      & x     pVö â ‰      " !r       Æ ‰      " !r*       â/ !r       äO ‰ €    b !r       Æ� ‰      ¢ !r
	
       Æ‰      â !r       Æ ‰ €    " !r       Æ ‰	
 €    " !r       Æ ‰ €    ( ‰ €    " !r       Æ/ ‰ €    b !r       âO !r       ä� ‰ @    ¢ !r       Æ‰ €    â !r	
	       Æ ‰ €    " !r       Æ ‰ @    " !r       Æ ‰
	 @    " !r       Æ/ ‰ @    h ‰ @    b !r       ÆO ‰ @    ¢ !r       â� !r       ä‰      â !r       Æ ‰ @    " !r
	
       Æ ‰ @    " !r       Æ/ ‰      b !r       Æ ‰	
      b !r       ÆO ‰      ¨ ‰      ¢ !r       Æ� ‰      â !r       â !‚       â!r       Æ‰      " !r       Æ/ ‰      b !r
	       â !‚       â !r       äO !’       â  ‚     A  â !r       â !’       Æ †‰     â !r       â� !¢       â  ’     A  ä !¢
       â !r       ä †™    â  ¢
     A  â !r       È/ †©    â M9        € ê !r       â !r       ÈO  r     A  Ê †y    â My        € ê „y      H   " z  r  p`ð â Ey  À    € ä „y     H   b r  ÿ   pVp  Æ „y     H   ¨ „y     H   è „y     H   ( „y   
  H   ( „y  €   H   ( „y  €  H   è „y  €  H   â !rE       â !r       Æ/ !r       âO !r       Æ� !r       â!r       È ‰      " !r	       Æ ‰      b !r       â !r       È/ ‰      ( ‰      h „y  €  H   ¢ !r       â !r       È/ ‰ €    ( ‰ €    h „y  €  H   â !r       ÈO !r       â !r       ä „y  €
  H   â !r       Æ/ ‰ @    ( ‰ @    b !r       È� !r       â !r       â !r       È/ ‰      ( ‰      b !r       â !r       â !r       â/ G	  à     € ê !r       Ê	x  à@  � Ê  x#ÛÙ?  A  È  x	;ª¸¿  A  â !r       Ê s	 	       " 	x  à@  � È 	x  àÀ  � Ê !t  €?    â !t	  €?    È s        $  r     A  È  r     A  Ê †y     ä Ay        € ê ‰      " Ey       € æ ‰      h „y     H   ¨ „y   	  H   è „y     H   è „y     H   â !r       Æ „y     H   â !r       Æ/ ‰      ( ‰      b !r       âO !r       â !r       È/ ‰	 €    ( ‰ €    b !r       Æ „y     H   ¢ !r       È !r       â !r		       â x   ÿàÿ â !r       Æ/ ‰	 @    " z  r  p`ð Æ ‰ @    b x     pVp  â !r       ÈO !r       â !r	       â !r       È/ ‰      ( ‰	      b !r       â !r	       â/ G	  à     € ê !r       Ê	x  à@  � Ê  x#ÛÙ?  A  È  x;ª¸¿  A  â !r       Ê s        " 	x  à@  � È 	x  àÀ  � Ê !t  €?    â !t	  €?    È s 	       $  r     A  È  r     A  Ê †y    ä Ay        € ê ‰      b Ey  ð    € æ ‰      ( „y  €  H   ¨ „y  €	  H   è „y  €  H   è „y  €  H   â !r       Æ/ „y  €  H   â !r	       Æ ‰      ( ‰	      h „y  €  H   " !r
       ÈO !r       â� !r       â !r	       È/ ‰ €    ( ‰ €    b !r       â!r       â x   ÿàÿ â !r       Æ/ z  r  p`ð â ‰ @    & x     pVp  â ‰ @    b !r       â !r       È/ ‰      ( ‰      b !r       â !r       â/ G	  à     € ê !r       Ê	x  à@  � Ê  x#ÛÙ?  A  È  x;ª¸¿  A  â !r       Ê s        " 	x  à@  � È 	x  àÀ  � Ê !t
  €?    â !t	  €?    È s 	       $  r     A  È  r
     A  Ê †y    ä Ay        € ê ‰      " !r       â x   ÿàÿ Æ !r       â z  r  p`ð Æ !r       â x     pVp  È ‰      b !r	       Ê ‰	      " !r       Ê/ ‰      b !r
	       Ê ‰
 €    " !r       Ê/ ‰ €    b !r
       Ê ‰ @    " !r       Ê/ ‰ @    b !r       Ê ‰	      " !r       Ê/ ‰      ¢ !r		       â M	        € ì !r	       â/!r       ÈO 	x  à@  � â !r       È  x #ÛÙ?  A  â 	x  à@  � Æ  x  ;ª¸¿  A  â 	x  àÀ  � Ê s          " !t  €?    â !t   €?    Ì s        $  r     A  È  r     A  Ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y       "   $ z   w  p`ð Ú M	        € ê y      !   " ¹z  l      â Ey  0    € â ™x?    â y      %   b ¹z  F   
   â $rÿÿ   ÿ Ž ä $rÿÿ   ÿ Ž â xÿ     È |    p`ð â $x   ÿ Ž Ø/ G	  �    € ê y      &   " $t	ÿ   ÿ Ž ä $z  x  Ž È %v \  	 Ž Ê �y    ™ ¢  $rÿÿ   ÿ Ž â x!   ÿÀŽ â$tÿ$   ÿ Ž ä 'z m   Ž â x   ÿÀŽ ä xÿ   !  ä zÿ n    â $z  r  ÿŽ È $rÿÿ   
Ž â xÿ     Æ $z
 o  Ž â xÿ     â $x   Ž ä $z	
 u  ÿŽ â x
øÿÿ?ÿÀŽ ä x   ÿàÿ ä r	
   à‘ È r   ÿàõ är	ÿÿ   ÿ$~  ä r
   ÿàñ Æ $rÿÿ   	 ä$rÿÿ   	  ä %v Z   Ž È $x$   ÿŽ ä $x$   ÿŽ ä %v

 Z   Ž È $x	   Ž â x   ÿÀŽ â $z q  Ž âx	   ÿàÿ â $x   Ž â xÿ     â $rÿÿ   ÿ Ž ä $z t  Ž äO$z t  Ž Ä $x   ÿ Ž ä$x   Ž ä $z	 q  Ž â xøÿÿÿâŽ â %xÒ   ÿŽ È %xÒ   Ž â r   ÿàñ Æ %xÒ   ÿŽ È $x   ÿ Ž ä $x   ÿ Ž ä %x!    Ž â x   ÿÀŽ Æ %x   Ž È $vÿ X  ÿ Ž ä $v	ÿ Y  ÿ Ž ä $rÿÿ   ÿ Ž ä %x   Ž È %x !    Ž È $rÿÿ     ä r$   ÿàó är"   ÿàñ Æ $x%   	Ž  âr.   ÿàó â $x#   	  Æ �y$€   • ¨  �y*$‚   • ¢  $x/   	Ž  Æ �y"    • è �y,"   • â r(    ÿàñ Æ �y1.€   • d	 $x)!   	  ä �y2.‚   • h	 �y0(    • h �y-(   • b	 r&
   ÿàñ Ä r4   ÿàó ä r"   ÿàõ â/ $r'ÿÿ     â r$   ÿàñ â $x5	   Ž  ä $x#   	 â �y&   ™ b $r%ÿÿ     Æ �y+4À   ‘ h �y/"À   ‘ h	�y.4Ä   ‘ h �y3"Ä   ‘ h	 �y$$   ™ b	 r&   ÿàñ Ä r(   ÿàó â �y4
    • b! $r"ÿÿ    Ž ä$r#ÿÿ    Ž ä $x'   	  ä$x)   	Ž  â �y%"    • ( �y&&Ð   • ( �y((Ð   • " x5ÿ     Ä x   ÿàÿ È |    p`ð ä x   ÿàó ä x

   ÿàõ ä xÒ   ÿà÷ â $rÿÿ   Ž  ä $rÿÿ    ä $r	ÿÿ   	Ž â x*T  *    ÈO r*ÿ5   *  ä x,T  ,    Æ� $x*   ÿ Ž âx#*0000ÿÀŽ ä/ x"ÿ   ,  ä x0000ÿÀŽ ä x21T  2    äx#"øŽ ä r2ÿ5   2  Ä x"€€€€ÿüŽ ä x',øŽ â $x2   ÿ Ž â x-0T  -    ä x""àßßßÿàÿ ä x#ÿ   -  ä x220000ÿÀŽ Ä x*0000ÿÀŽ ä x)    "Ž ä x0'€€€€ÿüŽ ä x2#øŽ ä x#*-øŽ ä x-)˜º  ÿ    Ä x00àßßßÿàÿ ä x*˜º  ÿ    ä x,#€€€€ÿüŽ ä x-€€€€"Ž ä x)'    0Ž ä x"-*`Ž Ä x-€€€€ÿüŽ ä x*,àßßßÿàÿ ä x))˜º  ÿ    ä x2'˜º  ÿ    ä x,-àßßßÿàÿ ä x-#    *Ž Ä x')€€€€0Ž äx0)2`Ž ä x1    ,Ž ä x)-˜º  ÿ    ä x2#˜º  ÿ    ä x#1˜º  ÿ    Ä x˜º  ÿ    ä x*)€€€€*Ž äx))2`Ž ä x,#€€€€,Ž är00'   ÿüŽ ä x##`Ž Ä r*)*   ÿüŽ â &r0   ÿ   â r"   ÿüŽ ä r##,   ÿüŽ â &r*   ÿ   â x"+€ˆ  ÿ    ä x,/€ˆ  ÿ    â &r*$   ÿ   âx'.€ˆ  ÿ    â &r$#$   ÿ   â x)3€ˆ  ÿ    â $r"   ÿŽ Ä $r,   ÿŽ ä $r''*   ÿŽ ä $r))$   ÿŽ â Er$        â 0rÿ4    A   â Er"        â 0r%ÿ%    A   âEr' '       ä Er) )       â #r$$   ÿ   â#r""   ÿ   â 0rÿ&    A   Ä 0rÿ(    A   â #r$%'   $   â#r"%)   "   Æ #r$      â #r"      â G‰   øÿÿÿÿƒ ê Ay        € ê ‰      " $x   Ž Æ ‰      d z  p  p`ð È x    pDp  â !r       â !r       È/ ‰      ( ‰      b !r       â !r	       È/ ‰ €    ( ‰
	 €    b !r       â !r
	
       È/ ‰ @    ( ‰
 @    b !r       â !r
       È/ ‰      h  ‰      ¢  M	        € ê y      &   â $x   ÿ Ž â !r       â/ $z   s  Ž â !r       âO $t	ÿ   ÿ Ž â r ÿ   p"ð Ä x    p"ò Ö $ÿÿ    Ž ä $ÿÿ    Ž ä $z  v   Ž È� %v  j  	 Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      "   $ z  w  p`ð Ú M	        € ê y      &   " $tÿ   ÿ Ž â ¹z  F   
   â $z x  Ž È %v \   Ž Ì �y    ™ b ¹z  l      â Ey  p    € â ™x?    â y      !   b $rÿÿ   ÿ Ž â x      ÿ  â $rÿÿ   ÿ Ž â x
      ÿ  â x	ÿ     È/ | 	   p`ð Ú G	  Ð    € ê $rÿÿ   ÿ Ž â y      %   " x   ÿÀŽ â$z r  ÿŽ â x   ÿÀŽ â'z  m   Ž â Ey @    € ä xÿ     â $tÿ$   ÿ Ž â z ÿ n     â $x   ÿ Ž È $rÿÿ    
Ž â x ÿ     äxÿ     â $z o  Ž â xÿ     â $x    Ž ä $z u  ÿŽ â xøÿÿ?ÿÀŽ ä x    ÿàÿ Ä r   à‘ È r    ÿàõ â$x    ÿ Ž â rÿÿ   ÿ$~  ä r   ÿàñ â %v Z   Ž â x    ÿàÿ Æ $rÿÿ    ä $z t  	Ž ä$x$   ÿŽ ä $rÿÿ    Ž â x   ÿÀŽ â $rÿÿ     â z ÿ `  pPð â $z q  Ž â xÿ     â $x$   ÿŽ â z ÿ a   Sð â %v Z   Ž È $x   Ž ä $x   Ž ä$x   ÿ Ž ä $z  q  	Ž ä $x   Ž â xøÿÿÿâŽ â $z6 t  Ž Ä $rÿÿ   ÿ Ž ä %x66Ò   ÿŽ È %x Ò   Ž â x   ÿÀŽ Æ $x   ÿ Ž â r6   ÿàó â %xÒ   ÿŽ È $rÿÿ    Ž ä %x   6Ž È %x   6 Ž È $vÿ X  ÿ Ž ä $vÿ Y  ÿ Ž ä $vÿ `  ÿ Ž ä $v ÿ a  ÿ Ž ä $rÿÿ   ÿ Ž ä %x   Ž È %x    Ž È $rÿÿ   7Ž  ä r&   ÿàõ är"   ÿàó Æ $x'    âr$   ÿàõ â$x#   Ž  âr   ÿàó ä �y2&€   • ¨  �y-&‚   • ¢  $x   Ž  Ä $x%    â �y3"    • è �y."   • è �y1€   • h	 �y0‚   • h	 �y,$    • h �y/$   • b r*   ÿàó Ä r(    ÿàõ Æ $x+   Ž  â r:   ÿàó â$x)!    â r8   ÿàõ ä �y"*À   ‘ b# $r;ÿÿ   Ž  ä $r9ÿÿ    â �y$(À   ‘ h �y%(Ä   ‘ h �y:   ™ (�y8   ™ ( �y#*Ä   ‘ " $r8ÿÿ    Ž Ä $r9ÿÿ    Ž â r46   ÿàó är*   ÿàõ ä/ �y&8    • " r57   ÿäÿ  ä $x+    Æ �y(4Ð   • " $r8ÿÿ    Ž Æ �y)*Ð   • " $r9ÿÿ    Ž Ê �y'8    • " x22T  -    ÈO r2ÿ   2  ä x33T  .    ä� x.20000ÿÀŽ â$x22   ÿ Ž â x-ÿ   3  ä x01T  0    Èr0ÿ   0  ä x-.-øŽ ä x,,T  /    ä x220000ÿÀŽ â $x+0   ÿ Ž âx*ÿ   ,  ä x/00000ÿÀŽ Ä x.-€€€€ÿüŽ ä x323øŽ ä x++0000ÿÀŽ ä x*/*øŽ ä x2.àßßßÿàÿ ä x/3€€€€ÿüŽ Ä x++,øŽ ä x5-    2Ž ä/ x,/àßßßÿàÿ ä x/*€€€€ÿüŽ ä x0+€€€€ÿüŽ ä x55˜º  ÿ    Ä x13    ,Ž ä x//àßßßÿàÿ ä x.-˜º  ÿ    ä x00àßßßÿàÿ ä x-5€€€€2Ž ä x11˜º  ÿ    Ä x23˜º  ÿ    ä x3*    /Ž ä x4+    0Ž ä x,1€€€€,Ž äx112`Ž ä x.5.`Ž Ä x23˜º  ÿ    ä x3*˜º  ÿ    ä x54˜º  ÿ    ä x*+˜º  ÿ    ä x05€€€€0Ž äx55*`Ž Ä x/2€€€€/Ž ä x*$€ˆ  ÿ    ä x223`Ž ä r11,   ÿüŽ ä r550   ÿüŽ ä x""€ˆ  ÿ    Ä x+%€ˆ  ÿ    â $r%ÿÿ   * Ž â r..-   ÿüŽ â $r$ÿÿ   " Ž â r22/   ÿüŽ â &r11   ÿ   ä	&r*5   ÿ   â x##€ˆ  ÿ    â $r"ÿÿ   + Ž ä &r..   ÿ   Ä &r-2   ÿ   ä $r$$1   ÿŽ ä $r+%*   ÿŽ ä $r.#.   ÿŽ â Er% $       â $r""-   ÿŽ â Er# +       â 0r*ÿ&    A   â Er- .       â 0r+ÿ'    A   â Er' "       Ä #r"*%   ÿ   â#r,*#   ÿ   â 0r/ÿ(    A   ä #r-+-   "   â#r,+'   ,   â 0r.ÿ)    A   â G‰       € ê r&   ÿàó Ê $x'    Ž  â r(   ÿàó È �y0&€    ¨ �y1&‚    ¢ $x)    Ž  Ê �y2(     è  �y5(    â  r$   ÿàó Ê $x%    Ž  â r"6   ÿàó È �y3$À    ( �y4$Ä    b $x#7    Ž  Ê �y""Ð    " x00T  1    ÈO r0ÿ   0  È x)00000ÿÀŽ â$x&0   ÿ Ž â x22T  5    È� x'&0000ÿÀŽ ä x(ÿ   2  äx&'2øŽ ä x$)(øŽ ä/ x%&€€€€ÿüŽ ä x#$€€€€ÿüŽ Ä x%%àßßßÿàÿ ä x##àßßßÿàÿ ä x(&    %Ž ä x)$    #Ž ä x((˜º  ÿ    ä x'&˜º  ÿ    Ä x%(€€€€%Ž ä x$$˜º  ÿ    ä x))˜º  ÿ    ä x(('`Ž ä x#)€€€€#Ž äx$)$`Ž Ä r%(%   ÿüŽ ä r&$#   ÿüŽ Æ &r$%   ÿ   ä &r#&   ÿ   ä $r$3$   ÿŽ ä$r44#   ÿŽ ÆEr# $       ä Er4 4       Æ #r$*#   ÿ   â 0r#ÿ"    A   Æ #r$+4   $   È #r#$      â #r
/-   
   â #r.,      â G‰       € ê r&   ÿàó Ê $x'    Ž  â r(   ÿàó È �y0&€    ¨ �y1&‚    ¢ $x)    Ž  Ê �y,(     è  �y/(    â  r$    ÿàó Ê $x%!    Ž  â r"   ÿàó È �y-$À    ( �y.$Ä    b $x#    Ž  Ê �y""Ð    " x00T  1    ÈO r0ÿ   0  È x)00000ÿÀŽ â$x&0   ÿ Ž â x,,T  /    È� x'&0000ÿÀŽ ä x&ÿ   ,  äx'',øŽ ä x))&øŽ ä x$'€€€€ÿüŽ ä/ x#)€€€€ÿüŽ Ä x$$àßßßÿàÿ ä x(#àßßßÿàÿ ä x#'    $Ž â 0r"ÿ"    A   âx%)    (Ž ä x##˜º  ÿ    ä x&'˜º  ÿ    Ä x%%˜º  ÿ    ä x$#€€€€$Ž ä x,)˜º  ÿ    ä x##&`Ž ä x(%€€€€(Ž äx%%,`Ž Ä r##$   ÿüŽ ä r(%(   ÿüŽ Æ &r#   ÿ   ä &r(   ÿ   ä $r-   ÿŽ ä $r..   ÿŽ ÆEr        ä Er. .       Æ #r**   ÿ   È #r*+.   *   È #r"*      â x		   ÿàÿ ä x   ÿàõ ä | 	   p`ò ä x   ÿà÷ â $rÿÿ    â xÒ   ÿàù ä xÒ   ÿàû â $rÿÿ   Ž Ä $rÿÿ    â r ÿ    ÿäÿ Æ G™  `ñÿÿÿÿƒ ê Ay       € ê $rÿÿ    Ž ä $rÿÿ    Ž ä Ay        € ê z ÿ `  pPð È z ÿ a   Sð Ú G	  P    € ê ‰
      ( ‰       b !r
       â !r        È/ ‰       ( ‰      b !r         â !r       È/ ‰  €    ( ‰ €    b !r        â !r       È/ ‰ @    ( ‰ @    b !r       â !r       È/ ‰      ( ‰	      b !r       â !r		       â/ Gy  €    € ê ‰      ( ‰
      h ‰       ¨ ‰      â !r       â !r
       È/ ‰      " !r        ÆO ‰       b !r       Æ� ‰      ¨ ‰
      â !r       â !r         â/ !r       ÈO ‰  €    " !r
       Æ� ‰ €    h ‰	 €    ¨ ‰ €    â !r        â !r       È/ ‰ @    " !r
	       ÆO ‰ @    b !r       Æ� ‰
 @    ¨ ‰	 @    â !r       â !r       â/ !r 
       ÆO ‰      " !r	       Æ� ‰	      h ‰       ¨ ‰      â !r       â !r	       â/ !r	        âO !r       â� y      %   $ $x   Ž Ê z  p  p`ð È x    pDp  Ú M	        € ê z ÿ ^  pPð â $z v  Ž Æz ÿ _   Sð Ú $ÿ   ÿ Ž È % ^   Ž Ì �	     ¢ $x    ÿ Ž Ê r  ÿ   p"ò äx     p"ô äx     p"ö äx     p"ø Î $ ÿÿ    Ž â z ÿ `  pPò â $" ÿÿ   	 Ž ä $2 ÿÿ    Ž â z ÿ a  Sò â $B ÿÿ    Ž È !         ÐO G™       € ê z ÿ b  pPð È z ÿ c   Sð Ú $ÿ   ÿ Ž È % b   Ž Ì �	     ¢ $x   ÿ Ž ä $vÿ h  ÿ Ž Æ x øÿÿÿp"ò äx üÿÿÿp"ô är ÿ   p"ö äx    p"ø Î $ÿÿ    Ž â x    pBò â $"ÿÿ   	 Ž ä $2ÿÿ    Ž ä $Bÿÿ    Ž È !       ÈO G  P    € ê x    pRð Ú G‰  �     € ê x    pRð Ú G	  ð    € ê  x;ª¸¿  A  Ì s        $ !t  €?    Ì s        $  r     A  È  r       A  â Gy  P    € ê  x'7=  A  â x*BL?  A  Æ #t  €?   â  x   ?  A  Æ  r     A  Ì s     $   $ !t  €?    È  r     A  È  r       A  â Gy  °    € ê x    pRð Ú G‰  à     € ê x    pRð Ú G	        € ê 	z i    � ä 	z   i    � Æ  x;ª¸¿  A  â 	z   i €  � Ê s        $ !t  €?    Ì s        $  r     A  È  r       A  â Gy  Ð     € ê  r       A  â Gy  °     € ê 	x  à@  � ä 	x    à@  � Æ  x#ÛÙ?  A  â 	x    àÀ  � Æ  x;ª¸¿  A  ä !t   €?    È s        $ !t  €?    Ì s        $  r      A  È  r       A  ä y      "   ( y      &   b $z s  Ž ä $tÿ   ÿ Ž ä $z v  Ž È/ %v j   Ž Ê †y      â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y       "   " ¹z  l      â Ey  ð(    € â ™x?    â y      !   b ¹z  F   
   â x      ÿ  â r( ÿ       â y      %   ¢ x      ÿ  â x
      ÿ  â x      ÿ  â x      ÿ  â x      ÿ  â r* ÿ       Ä r  ÿ   pRð är  ÿ   pRò är    ÿ(Ž È/ xÿ     ä x   ÿ   äO r   ÿ � ä x      ÿ  ä xÿ     È |    pbô Ú G)  €'    € ê y      '   " xðÿÿÿÿÀŽ ä r ÿ       â y      &   b r  €ÿàÿ ä r        ä r ÿ       ä xÿ     Ä x   ÿàÿ ä r   ÿ� ä z  q      ä x,ÿ     ä xüÿÿÿÿÀŽ ä z  X      â 'z y   Ž â z$  Y      Æ 'z s   Ž â/ zÿ z    Æ $z p  Ž â zÿ t    â $z w  ÿŽ ä $z |  Ž ä$z |  Ž ä $z: v  Ž â r  €ÿàÿ â $z v  Ž â x   ÿ   â $z }  ÿŽ â x,   ÿ   â $z p  Ž â xÿ     â $x   Ž âr   àµ â $x   Ž â x     â $x    Ž âx   ÿ   â $x"   Ž âr&   ÿŽ â%x::°   ÿŽ â r   ÿŽ Ä xÿ     â %x<°   ÿŽ â r%,   ÿ(Ž ä rÿÿ   d~ ä xÿ     ä r!:   ÿà÷ ä x/ÿ   %  Ä r#%:   ÿàõ äx':   øŽ är%%<   ÿàû är)<   ÿàù ä x1<   øŽ ä r+;   ÿäÿ Ä r-;/   ÿä är ÿ       ä z q  ÿàÿ ä x''   ÿüŽ ä r/=/   ÿäÿ ä x11   ÿüŽ Ä r=   ÿä ä r2'   ÿàõ È r3;$   ÿä Ê �y]2øÿÿ • ¢  r41   ÿàõ Æ �yX2    • â  r5=$   ÿä ä r6#   ÿàõ â�yZ2üÿÿ • " r8!   ÿà÷ ä r7$-   ÿä â�yU4   • b r9$+   ÿäÿ Æ �y[4üÿÿ • " r?   ÿàõ ä rB)   ÿàù â�yR4    • " rD%   ÿà÷ Æ �y,60   ™ " xH $       Æ �y.6@   ™ b rC$   ÿä Æ �y08   ™ b r5ÿ   ÿä Æ/ �y28    ™ b rE$/   ÿäÿ â %v>? Z  H Ž ä �y6B   ™ d $x55$   ÿŽ ä �y3D0   ™ h �y8B    ™ b r@>   ÿàõ Æ �y4D@   ™ b r??5   ÿàÿ È rA?   ÿä â �yK>    • h �y5@   ™ b rC   ÿàõ Æ/ �y7@8   ™ b rDÿ   ÿä Æ �yJ@(   ™ h �y9@   ™ b %vBC Z  H Ž Æ �yL>$   • b $xED$   ÿŽ â rDB   ÿàõ È rCCE   ÿàÿ È rEC   ÿä â �yQB    • h �yMD   ™ h �yND8   ™ h �yOD   ™ h �yPD(   ™ b r>:   ÿàõ Ä rG&   ÿàù â �ySB$   • b r?;$   ÿä ä r@<   ÿà÷ ä/ rFÿ   ÿä â �yT>    ™ b rA=$   ÿäÿ â %vDG Z  H Ž È $xGF$   ÿŽ â �yY@    ™ b rBD   ÿàõ Æ rEEG   ÿàÿ È rCE   ÿä â �y_D    • b rA   ÿàõ Æ/ �y\B   ™ b r>ÿ   ÿä Æ �yVB8   ™ h �yWB   ™ b %v@A Z  H Ž Æ �y^B(   ™ b $x?>$   ÿŽ â r>@   ÿàõ ä �yaD$   • d rAA?   ÿàÿ È r?A   ÿä â �ye@    • h �yb>   ™ b rC   ÿàõ Æ/ �yc>8   ™ h �y`>   ™ b rDÿ   ÿä Æ �yd>(   ™ b %vBC Z  H Ž Æ �yf@$   • b $xGD$   ÿŽ â rDB   ÿàõ È rGCG   ÿàÿ È rEG   ÿä ä rC   ÿàõ Æ �ygD   ™ b r>ÿ   ÿä ä/ rA"   ÿàõ â �yhD8   ™ b rF B       ä r@ÿ   ÿä â �yiD   ™ b %vBC Z  H Ž Æ �yjD(   ™ b $xm>$   ÿŽ Æ �ykF    • d rCCm   ÿàÿ ä �ylF$   • b %vDA Z  H Ž â r?    ÿà÷ ä �ymB    • b $xA@$   ÿŽ â r@B   ÿàõ Ä �ynB$   • d rEEA   ÿàÿ ä rAC   ÿä ä rIÿ   ÿäÿ â %v>? Z  H Ž â �yrD    • h �yq@   ™ h �yt@8   ™ h �yu@   ™ h �yv@(   ™ b $xII$   ÿŽ â rF>   ÿà÷ Ä/ �ysD$   • d r??I   ÿàÿ È rG?   ÿäÿ â �yp>    • h �yyF   ™ h �yxF   ™ h �yzF8   ™ b rHD   ÿàù Æ �ywF(   ™ b rIE   ÿä Æ �yo>$   • h �yDH   ™ h �yEH   ™ h �y|H8   ™ b x{]??  ÿÀŽ ÄO xC]ÀÀ  ÿÀŽ ä �y]H(   ™ ¢  x    pBô ä x?X  ÿÀŽ Ö� "> ÿ       Ê '({C   @> Ž â x>ÿ   X  ä x}Z??  ÿÀŽ äx?>  ÿÀŽ ä xZZÀÀ  ÿÀŽ ä "> ÿ       ä xBÿ   U  ÄxF[??  ÿÀŽ â/'(}Z   @> Ž â x[[ÀÀ  ÿÀŽ ä xAU  ÿÀŽ ä "@ ÿ       ä xCB  ÿÀŽ ä xIRÀÀ  ÿÀŽ Ä "B ÿ       ä x>,ÿÀŽ ä x?ÿ   ,  ä x,.ÿÀŽ ä r0ÿ   0  â '(F[   @@ Ž â xGR??  ÿÀŽ Ä x.ÿ   .  ä r2ÿ   2  â '(GI   @B Ž â xA?ÿÀŽ ä x?0   ÿ   ä x..ÿÀŽ ä xC2   ÿ   Ä xI2   ÿ   ä x>>?øŽ ä r8ÿ   8  ä x2,CøŽ ä x?.IøŽ ä x00   ÿ   Ä x,3ÿÀŽ ä x.4ÿÀŽ ä x3ÿ   3  ä x4ÿ   4  ä r6ÿ   6  ä xC8   ÿ   Ä xAA0øŽ ä x03ÿÀŽ ä x@4ÿÀŽ ä x36   ÿ   ä x4.CøŽ ä xI6   ÿ   â &z.5  € ÿ   â xU8   ÿ   â &rB45   ÿ   âx6,3øŽ â &r,25   ÿ   Ä &z57  € ÿ   â x3@UøŽ ä xC0IøŽ â &z0J  € 5   â x5{?   ÿÀŽ â &r@?7   ÿ   â x8}?   ÿÀŽ â &r,>9   ,   ä &z.9  € .   Ä &rB69   B   â x9G?   ÿÀŽ â &rI37   ÿ   â x{ÿ   {  â &r@AJ   @   âx7F?   ÿÀŽ â $r,,5   ÿŽ â x}ÿ   }  â &rICJ   I   â xGÿ   G  â $rX.8   ÿŽ âxFÿ   F  â $rJ.9   ÿŽ Ä $rHB7   ÿŽ ä $rR@{   ÿŽ â Er@ ,       â $rU0}   ÿŽ âEr, X       â 0rKÿK    A   ä $rB0G   ÿŽ â Er0 J       â $rIIF   ÿŽ â Er. H       â 0rLÿL    A   â ErH U       â #r,K,   ÿ   âErU B       â #r0K0   ÿ   âEr[ I       â #rIK@   ÿ   â#r.K.   ÿ   â #rKLH   ,   â#rULU   0   â &r,2M   ÿ   Ä &z0M  € ÿ   â #rHL[   .   â &rB4M   ÿ   ä &r@?N   ÿ   ä&zMN  € ÿ   ä &r.>O   ,   ä &rN3N   ÿ   Ä &z,O  € 0   ä &r06O   B   â ErR R       â &rBAP   @   ä $r.5.   ÿŽ ä &z@P  € M   ä &rOCP   N   Ä $rM70   ÿŽ ä $r08,   ÿŽ â #rILR   I   â $rJ{B   ÿŽ â ErB .       â 0rQÿQ    A   ä $rN9,   ÿŽ ä $rOFO   ÿŽ â Er. M       â $rL}@   ÿŽ â Er, 0       â 0rSÿS    A   â ErJ J       â #rBQB   ÿ   â Er0 N       â $r@G@   ÿŽ â Er[ O       Ä ErO L       â #rLQ.   ÿ   â#r.Q,   ÿ   â #rMSJ   B   â #r0Q0   ÿ   â 0rBÿT  0 A   âErQ @       â #rOSO   .   â 0rTÿT    A   ä 0r@ÿY  0 A   â r.KB     A  â #rQSQ   0   â  rOBO     A  â 0rYÿY    A   â #rIIT   .  â #r,S[   L   â  r0@Q     A  â #r.TM   O  â  rUU@     A  Ä #rKY,   0  â !r.       â &r,2\   ÿ   ä&z.\  € ÿ   â !r*I*       â &r\4\   ÿ   â #rHHY   U  â &r0?V   ÿ   ä &zIV  € ÿ   Ä &rV3V   ÿ   ä &r,>W   ,   ä&z.W  € .   ä &r\6W   \   â !rH       â !rK       â &rHA^   0   ä&rKC^   V   Ä $r,5,   ÿŽ ä &zJ^  € I   ä $r\7\   ÿŽ ä $r08.   ÿŽ ä$rL9.   ÿŽ ä $rI{H   ÿŽ â ErH ,       â $rMFK   ÿŽ â Er. \       â 0r_ÿ_    A   â Er, 0       â $rK}J   ÿŽ â Er0 L       â $rJGJ   ÿŽ â ErI I       â 0raÿa    A   â ErM M       â #rH_H   ÿ   âErK K       â #r._.   ÿ   âErO J       â #r,_,   ÿ   â#r0_0   ÿ   â #rIaI   H   â#rHaM   .   â#rKaK   ,   â#raaO   0   â &rM2b   ÿ   Ä &zOb  € ÿ   ä &zQc  € ÿ   ä &r,>`   M   ä &zO`  € O   ä &rS4b   ÿ   ä &zJd  € Q   Ä $r,5,   ÿŽ ä &r.?c   ÿ   ä $r08O   ÿŽ ä&r`6`   S   ä $rM}J   ÿŽ ä$rQGJ   ÿŽ â ErJ ,       â &r.Ad   .   â Er, 0       â $rO9O   ÿŽ Ä 0reÿe    A   ä $r`7`   ÿŽ â Er0 O       â $rL{.   ÿŽ â ErN M       â 0rfÿf    A   â #rOe,   ÿ   âEr. `       â  r,@a     A  â ErL L       â #rMeJ   ÿ   â #rOfN   O   â ErR Q       â  rKBK     A  â &rc3c   ÿ   â #rQe.   ÿ   â #r,YH   ,  â #rMfL   M   â  r.BO     A  â #rITI   K  â &rcCd   c   â !r,       â #rMTM   .  â &r,2g   ÿ   â#ree0   ÿ   â &z.g  € ÿ   â !rI       â &rg4g   ÿ   Ä $rcFc   ÿŽ ä &r0?h   ÿ   ä&zIh  € ÿ   ä &rh3h   ÿ   ä &r,>i   ,   ä &z.i  € .   Ä &rL6i   g   â ErP c       â &rHAj   0   â #refR   e   â &rKCj   h   ä $r,5,   ÿŽ ä &zJj  € I   ä $r08.   ÿŽ â !rM       â $rL7L   ÿŽ â #rPfP   Q   â $rM9.   ÿŽ â  re@e     A  â $rI{H   ÿŽ â ErH ,       â $rNFK   ÿŽ â Er, 0       â $rK}J   ÿŽ â Er. L       â 0rkÿk    A   â #rPYP   e  â $rOGJ   ÿŽ â Er0 M       Ä ErL K       â #rKk,   ÿ   â0rlÿl    A   â ErJ I       â #rMk.   ÿ   âErN N       â &r,2q   ÿ   â !rP       â #rIkH   ÿ   â &z.q  € ÿ   â ErP O       â #rkk0   ÿ   â &rq4q   ÿ   Ä &r0?t   ÿ   ä &zOt  € ÿ   â #rHlN   M   â &rt3t   ÿ   ä &r,>u   ,   â#rKlL   K   â&z.u  € .   â #rIlJ   I   â &rN6u   q   Ä &zLv  € O   â #rMlP   k   â &rJAv   0   ä&rOCv   t   ä $r,5,   ÿŽ ä $r08.   ÿŽ ä $rP7N   ÿŽ Ä $rN{J   ÿŽ â ErJ ,       â $rRFO   ÿŽ â Er, 0       â $rQ9.   ÿŽ ä $rO}L   ÿŽ â Er. P       â 0rmÿm    A   ä $rSGL   ÿŽ â Er0 Q       â 0rnÿn    A   â ErL O       â #rQm,   ÿ   âErP S       â #rSm.   ÿ   â rKBK     A  â  r.@M     A  â ErN N       â #rOmJ   ÿ   â #rQnL   Q   â #r,TI   K  â #rIYH   .  â #rmm0   ÿ   â #rOnN   O   â  r0BQ     A  â !r
I
       â &rI2y   ÿ   Ä &zKy  € ÿ   â #r0TO   0  â &ry4y   ÿ   â !r,       â &r,>x   I   â!r	0	       â &zKx  € K   ä &rx6x   y   ä &r.?z   ÿ   Ä &z0z  € ÿ   ä &rz3z   ÿ   ä $rx7x   ÿŽ ä &r.Aw   .   ä &z0w  € 0   ä $r,5,   ÿŽ Ä &rwCw   z   ä $rI8K   ÿŽ ä$rM9K   ÿŽ â ErK x       â 0rpÿp    A   â ErO ,       â $r.{.   ÿŽ â ErI I       â $rwFw   ÿŽ Ä $rJ}0   ÿŽ âErM M       â $r0G0   ÿŽ â ErH .       â 0roÿo    A   â #rLpK   ÿ   âErw w       â #r,pO   ÿ   âErK J       â #r.pI   ÿ   âEr0 0       â #rppM   ÿ   â &r22D   ÿ   Ä &r44D   ÿ   â #rIoH   ,   â&zDD  € ÿ   â #r,ow   L   â#rKoK   .   â#roo0   p   â &r2>E   2   ä&r46E   4   ä &zDE  € D   Ä &r.?|   ÿ   ä&r03|   ÿ   ä &z||  € ÿ   ä $r252   ÿŽ ä $r474   ÿŽ ä &r.A]   .   ÄO $r38D   ÿŽ ä&z|]  € |   ä $r79D   ÿŽ â Er9 2       â $r.{.   ÿŽ â Er3 3       â &rCC]   0   â Er5 4       â $r}}|   ÿŽ âEr7 7       â $rGG|   ÿŽ Ä 0rrÿr    A   â x   ÿàÿ â $rCFC   ÿŽ â Er2 .       â 0rsÿs    A   â Er} }       â #r.r9   ÿ   âErG G       â #r0r3   ÿ   â#r4r5   ÿ   â#rrr7   ÿ   â |    pbô â #rmnP   m   â ErR R       â #r}s}   0   âErC C       â #rGsG   r   â#r5s2   .   â rKBK     A  â #rRnR   S   â  rm@m     A  â#r4sC   4   â  r.@o     A  â  rBB}     A  â  rG@G     A  â #rmYR   m  â#rITI   K  â#r3Y,   .  â#rBT5   B  â #rGY4   G  â xÀ  ÿà÷ â !rm       â !r(I(       â !r3       â !rB       â !rG       â x    ÿàÿ Ä x    ÿàÿ ä x&&    ÿàÿ ä x    ÿàÿ ä x    ÿàÿ ä x    ÿàÿ ä x      ÿàÿ Ä x""    ÿàÿ ä r$ÿ$   ÿäÿ â D)       À â Gy  PÜÿÿÿÿƒ ê Ay        € ê Ey  0    € â G‰      € ê r    ÿHŽ Ê ˆs * øÿ H   è ˆs €øÿ H   è ˆs  ùÿ H   è ˆs €ùÿ H   è ˆs  úÿ H   è ˆs €úÿ H   è ˆs  ûÿ H   è ˆs €ûÿ H   è ˆs  üÿ H   è ˆs 
€üÿ H   è ˆs 	 ýÿ H   è ˆs €ýÿ H   è ˆs ( þÿ H   è ˆs €þÿ H   è ˆs  ÿÿ H   è ˆs €ÿÿ H   ä Ay        € ê {          ì M        € ê „y     H   " z  r  p`ð Ä z3  r      â „y €   H   b r ÿ   p"p ä r133   ÿŽ â„y    H   ¢ r733   ÿŽ Æ „y €  H   è „y    H   ( „y €  H   h „y    H   h y      '   è „y% €  H   è „y"    H   â !r*       Æ „y) €  H   â !r       Æ/ ‰       " !r       ÆO „y+    H   â !r       Æ� „y- €  H   â !r       Æ„y$    H   â !r       Æ‰      b !r        Æ ‰      ¢ !r        Æ ‰      " !r%%
       Æ ‰      â !r'"	       Æ ‰      " !r))       Æ y       &   b !r       Æ/ ‰!       b !r++(       Æ „y5 €  H   â !r       âO !r--       â !r       â ‰(%      " !r/$       â !r       ä� ‰.+      ¢ $x3   ÿŽ â !r       Ä„y €  H   â $z   x  Ž Æ„y    H   â x   ÿàÿ â $z  ~   Ž ä ‰ €    â !r! !       Æ/ ‰      b 
  r  ÿàó ä 1   ÿàù â ‰      " r3    ÿ† â !r(%(       â ,ÿÿ   ÿäÿ  â ‰0-      " z  r  p`ò â !r.+.       âO &ÿÿ   ÿä â ‰2/      ¢ r*ÿÿ   ÿäÿ Ä 
 j  ÿˆ â ‰       b z j  ÿ† â !r55       â 
 j  ÿ„ â !r       â� 
 k  & â ‰"!      â z k  *� â !r       â/ 
 k  , â ‰*'      b r	3    ÿ† â!r       â !r       â‰,)      " r    ÿàõ â !r#       â  7   ÿàù â ‰$      " x    p"ò â !r0-0       â ÿÿ   ÿä â ‰&#      " 

 j  ÿˆ â !r2/2       âO rÿÿ   ÿäÿ â ‰      ¢ z	 j  ÿ† â !r         ârÿÿ   ÿä â ‰      b z j  ÿ„ â !r"!"       â� 
 k   â ‰ @    â !r*'*       â/ z		 k  � Ä ‰ €    b !r,),       âz k   ä ‰      " !r$$       â 1   ÿàÿ â !r&#&       â ‰)(      " 3   ÿàÿ â !r       ÄO ‰%$      ¢ 8    ÿàù â !r       ä‰'&      b 9ÿÿ   ÿä â !r       ä� ‰ €    â 8 j  ÿˆ â !r       Ä/ ‰+*      b r33  €ÿŽ â !r       ä‰-,      "     ÿàõ ä    ÿà÷ â ‰ €    " 4ÿÿ   ÿä â !r(()       â 
 j  ÿ„ â ‰/.      " !r$$%       âO 6ÿÿ   ÿäÿ Ä ‰10      ¢ !r&&'       â
 k  4 ä ‰      b !r       â� 8 k  9 ä ‰32      â !r**+       â/  j  ÿ† Ä ‰5      b !r,,-       â k  6� ä ‰ @    " !r       Æ ‰!  €    " !r../       Æ ‰ €    " !r001       ÆO ‰#" €    ¢ !r9       â7   ÿàÿ â !r223       â� ‰'$ €    ä     ÿàõ â !r45       â/ ‰)& €    d 6ÿÿ   ÿä â !r       â‰+( €    " x        â !r  !       Ä ‰ @    b  j  ÿ„ â%v  j   Ž â !r       â ‰-* €    "  k  6 â !r""#       äO ‰/, €    ¨ ‰ @    ¢ !r'$'       Æ� ‰1. €    â !r)&)       Æ/ ‰30 €    b !r+(+       Æ‰52 €    " !r       Æ‰74      b !r-*-       Æ ‰      " !r/,/       ÆO ‰#  @    ¢ !r       Æ ‰! @    ¢ !r1.1       Æ� ‰%" @    â !r303       Æ/ ‰$' @    b !r525       â!r747       ä‰&) @    " !r;       Æ ‰(+ @    " !r# #       ÆO ‰      â !r!!       Æ ‰*- @    ¢ !r%"%       Æ� ‰ / @    â !r$'$       Æ/ ‰      h ‰"1 @    b !r&)&       Æ‰,3 @    " !r(+(       Æ ‰07 €    ( ‰.5 @    " !r*-*       ÆO ‰!      ¢ !r /        Æ� ‰#      â !r)       Æ/ †	 9    â !r"1"       â!r,3,       ä† ;   â !r070       Æ ‰$      " !r9       â/ !r.5.       ä ‰%      b !r+!       ÆO ‰&      ¢ !r-#       Æ� ‰(      è †	 9    è ‰*      " !r$       Æ ‰       ( † )    â !r%%       Æ/ ‰!"      b !r&       ÆO ‰#,      ¢ !r(       Æ� ‰'0 @    è ‰.      b !r*       Æ†	 +    â	 !r        Æ † -   è	 †	 %    â	 !r!"!       â/ !r#,#       äO †     â	 !r'0'       Æ� †	     â	 !r.       Æ†    è	 †	 
    è	 †     è	 †	 !    è	 † #   è	 ‰'      ( †	     â	 M™        € ê x   ÿàÿ â !r''       Ær    ÿàñ È r ÿÿ   ÿä  ä z j  ÿ€ È z k     Ê †y '    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      "   " ¹z  l      â Ey  @%    € â ™x?    â y      !   b ¹z  F   
   â r ÿ       â x      ÿ  â y      %   ¢ r/ ÿ       â x      ÿ  â x      ÿ  â r1 ÿ       â x      ÿ  â x      ÿ  â r3 ÿ       Ä r ÿ   pRð är ÿ   pRò är   ÿ(Ž È/ xÿ     ä x   ÿ   äO r    ÿ � È xÿ      È |    pbô Ú G)  Ð#    € ê y      '   " x ðÿÿÿÿÀŽ ä r ÿ       â y      &   b r  €ÿàÿ ä r ÿ       ä x	ÿ     ä x   ÿàÿ Ä r		   ÿ� ä x    ÿ   ä z-  X      ä z'  Y      ä r ÿ       ä z$  q  ÿàÿ â 'z  y   Ž È 'z s   Ž â/ zÿ z     Æ $z  p  Ž â zÿ t    â $z# }  ÿŽ â z  q      â $z |   Ž âx ÿ   	  â $z |  Ž â x		üÿÿÿÿÀŽ â $z w  ÿŽ â x!    ÿ   â $z v  Ž â r		  €ÿàÿ â $z v  Ž â r##   !àµ â $z p  Ž â xÿ   	  â %x
°   ÿŽ â x+	   ÿ   Ä x)	     â $x&    Ž âx.ÿ   !  â $x*    Ž âr"    ÿŽ â$x,    Ž âr(    ÿŽ â %x°   ÿŽ â r +   ÿ(Ž ä r.ÿÿ   .d~ Ä x%ÿ   +  ä r0+
   ÿà÷ äx6
   !øŽ ä xÿ     ä r2
   ÿàõ är4   ÿàû Ä r8+   ÿàù ä x@   !øŽ ä r:%   ÿäÿ äx66   ÿüŽ ä r<   ÿä är>	   ÿäÿ Ä x@@   ÿüŽ ä r%	%   ÿä ä r-6   ÿàõ È r'   ÿä Ê �yWøÿÿ • ¢  r-@   ÿàõ är-0   ÿà÷ â �yO    • â  r	'   ÿä ä r-2   ÿàõ â �yQüÿÿ • ( �yK   • b r'<   ÿä Æ �yA    • " r':   ÿäÿ Æ �yRüÿÿ • " r-4   ÿà÷ Æ �y 0   ™ " r-8   ÿàù ä x\ $       â �yD    ™ " r #   ÿàõ Æ/ �y5@   ™ " rÿ.   ÿä ä r'>   ÿäÿ â�yB   ™ " r'%   ÿä Æ �y70   ™ " %v Z  \ Ž Æ/ �yH    ™ " $x$   ÿŽ â r+   ÿàõ ä �y9@   ™ $ r   ÿàÿ ä �yF   ™ $ r)   ÿä Ä r$#   ÿàõ â �yG    • ( �y=8   ™ ( �y;   ™ " rÿ.   ÿä Æ/ �yC(   ™ ( �y?   ™ " %v Z  \ Ž Æ �yL$   • " $x$   ÿŽ â r+   ÿàõ È r   ÿàÿ È r)   ÿä â �yE    • ( �yN   ™ ( �yJ8   ™ ( �yI   ™ " r
-   ÿàõ Æ �yP(   ™ " r'   ÿä ä r"#   ÿàõ â/ �yM$   • & rÿ.   ÿä â �yV    ™ " %v Z  \ Ž È $x$   ÿŽ â/ r+   ÿàõ È r   ÿàÿ È r)   ÿä â �yY    • ( �yS   ™ ( �yT8   ™ ( �yU   ™ ( �yX(   ™ " r-   ÿàõ Ä r&#   ÿà÷ â �yZ$   • " r	'   ÿä ä rÿ.   ÿäÿ â %v Z  \ Ž ä �y]    ™ $ $x$   ÿŽ â r+   ÿàõ È r   ÿàÿ È r)   ÿä â �ya    • " r(#   ÿàõ Æ/ �y^   ™ " rÿ.   ÿä Æ �y_8   ™ ( �y[   ™ " %v Z  \ Ž Æ �y`(   ™ " $x$   ÿŽ â r+   ÿàõ ä �ye$   • $ r   ÿàÿ È r)   ÿä Ê �yd   ™ ( �yc8   ™ ( �yb   ™ " r        Æ �yf(   ™ " r*#   ÿàõ Æ �yg    • " rÿ.   ÿä â/ %v Z  \ Ž â r,#   ÿàõ â �yh$   • $ $x$   ÿŽ â rÿ.   ÿä ä r+   ÿàõ ä r   ÿàÿ â %v Z  \ Ž Æ r)   ÿä Ê �y\   ™ " $x$   ÿŽ Æ �yj8   ™ ( �yi   ™ " r        Æ �ym(   ™ " r   ÿàÿ ä r+   ÿà÷ â �y    • &' r)   ÿäÿ â �y$   • ( �yp8   ™ ( �yo   ™ ( �yn(   ™ ( �yk    • ( �yl$   • " xqW??  ÿÀŽ ÄO xsWÀÀ  ÿÀŽ ä �yW   ™ ¢ x    pBô ä xO  ÿÀŽ ä� xÿ   O  ä xÿ   K  äxOA??  ÿÀŽ Ê" ÿ       ä xAAÀÀ  ÿÀŽ ä x  ÿÀŽ ä/ " ÿ       â '(qs   @ Ž â x  ÿÀŽ ä xQÀÀ  ÿÀŽ â '(OA   @ Ž â " ÿ       Ä x ÿÀŽ ä x ÿ      ä xrQ??  ÿÀŽ ä rDÿ!   D  â '(r   @ Ž â xA ÿÀŽ ä x5ÿÀŽ Ä x D   ÿ   ä x5ÿ   5  ä x øŽ ä rBÿ!   B  äx ÿ   7  ä rHÿ!   H  Ä x7ÿÀŽ ä x55ÿÀŽ ä xDD   ÿ   ä x79ÿÀŽ ä xÿ   9  ä xB   ÿ   Ä x9 ÿÀŽ ä rFÿ!   F  ä x H   ÿ   ä xK  ÿÀŽ ä xsRÀÀ  ÿÀŽ ä " ÿ       Ä xBB   ÿ   ä x55DøŽ ä xÿÀŽ ä xKH   ÿ   ä xøŽ ä xQR??  ÿÀŽ Ä xF   ÿ   äx77 øŽ â '(Qs   @ Ž â xAABøŽ ä xFF   ÿ   â &rB5=   ÿ   â xKKøŽ â &r ;   ÿ   âxøŽ â &z;  € ÿ   â x99FøŽ â &r;7;   ÿ   Ä &zF=  € ÿ   ä &rHK=   ÿ   ä &rDAC   B   â xBQ?   ÿÀŽ â &r?   ;   â x=r?   ÿÀŽ â &z?  €    ä &zFC  € F   Ä &rH9C   H   â xCO?   ÿÀŽ â $rB   ÿŽ â x;q?   ÿÀŽ â &r?       â xrÿ   r  â $r=   ÿŽ âxqÿ   q  â $rC   ÿŽ ä $r;   ÿŽ â Er        â $rFr   ÿŽ â x?ÿ   Q  â 0r ÿG    A   â xOÿ   O  â $rDDq   ÿŽ â Er        â 0rLÿL    A   â ErG        â #r    ÿ   â Er        â $rHH?   ÿŽ â Er        â $rFFO   ÿŽ â ErD D       â #rs    ÿ   â#rG G   ÿ   â #rL      â &rN   ÿ   â Er H       â #rQ    ÿ   â ErF F       â #rLD   G   â 0r ÿE    A   Ä &zN  € ÿ   ä &rE5J   ÿ   ä &rDI      â #rL   Q   â#rLLF   s   â &zI  €    ä &zGJ  € ÿ   ä &rFAP   E   Ä $rD;D   ÿŽ ä $r=   ÿŽ ä &zEP  € G   â ErG D       â $rFqF   ÿŽ ä &rR7N   ÿ   â Er        â 0rMÿM    A   â ErH F       â &rJKJ   ÿ   â #rD G   ÿ   â &rII   R   Ä $rNrE   ÿŽ â #rF    ÿ   â &rJ9P   J   â #rGMH   D   â $rIBI   ÿŽ ä $rC   ÿŽ â ErQ N       â 0rDÿV    A   ä $rOE   ÿŽ Ä 0rVÿV  0 A   â ErI I       â $rJ?J   ÿŽ â Er        â #rQMQ   F   â Ers        â  rV     A  â ErE J       â #rH I   ÿ   â#rJ    ÿ   â  rQVQ     A  â #r D     â #rME   H   Ä #rFDG   Q  â !r3 3       â &r S   ÿ   ä&zHS  € ÿ   â !rF       â &rS7S   ÿ   ä &r5T   ÿ   ä &rFU       â #rMMs   J   â &zT  € ÿ   Ä &zHU  € H   ä &rSU   S   ä &rJAX      ä $rF;F   ÿŽ ä &zX  €    ä $rSBS   ÿŽ Ä $r=H   ÿŽ ä$rCH   ÿŽ â ErE F       â $rJqJ   ÿŽ ä 0r ÿY    A   â Er        â $rGr   ÿŽ â Er        â $rHO   ÿŽ â Er S       â 0rZÿZ    A   â ErJ J       â #rE E   ÿ   â#r    ÿ   â#rI    ÿ   â#rQ    ÿ   â Er  G       â #rZJ   E   â 0rEÿ]  0 A   Ä 0r]ÿ]    A   â #rZ       â &rTKT   ÿ   â  rLE     A  â  r EM     A  â  rV     A  â &rT9X   T   â #r]     â #r]      â #r D     â $rT?T   ÿŽ â ErH H       â !r       â !r        â &r ^   ÿ   Ä &z^  € ÿ   ä &r^7^   ÿ   â ErT T       â !r       â #rHZH   Q   â &r5_   ÿ   ä &z_  € ÿ   ä &r [       Ä &z[  €    ä &r_K_   ÿ   ä &r[[   ^   â #rTZT   I   â  rGEH     A  â &rA`      ä &z`  €    ä $r ;    ÿŽ Ä &r`9`   _   ä $r[B[   ÿŽ â #rG]T   G  â $r=   ÿŽ â ErF         â $rq   ÿŽ ä 0raÿa    A   ä $rHC   ÿŽ â Er [       â $r`?`   ÿŽ â !rG       â Er         â 0reÿe    A   â Er        â $rGr   ÿŽ âEr H       â $rIO   ÿŽ â Er `       â #rFaF   ÿ   â#ra   ÿ   âEr G       â #r a    ÿ   â ErI I       â #re   F   â #ra   ÿ   â #rFe      â &rd   ÿ   Ä &zGd  € ÿ   â #re       â#rHeI      â &r5c   ÿ   ä &r b      ä &zIc  € ÿ   ä &zb  € G   ä &rM7d   ÿ   Ä &rJAf      ä $r ;    ÿŽ ä &zIf  € I   ä $r=   ÿŽ ä &rMb   M   ä $rLC   ÿŽ Ä $rqJ   ÿŽ â ErJ         â $rGrI   ÿŽ â Er         â 0rgÿg    A   â Er L       â $rMBM   ÿŽ â ErL G       â 0rhÿh    A   â #rGg    ÿ   â &rcKc   ÿ   â Er M       â $rNOI   ÿŽ â ErI        â &rf9f   c   â #rgJ   ÿ   â#rGhL   G   â  rV     A  â #rMg   ÿ   â $rf?f   ÿŽ â #rhI      â  rVG     A  â ErN N       â #r D     â#rgg   ÿ   â #rD     â Erf f       â !r1 1       â #rNhN   g   â &r \   ÿ   â  rHEH     A  â!r       â &z\  € ÿ   â #rfhf   M   â  rGEN     A  â &r5j   ÿ   Ä &r i       â #r]F   H  â &r\7\   ÿ   â #rG]f   G  â &zi  €    ä &rAm      ä $r ;    ÿŽ â !r       â &rii   \   â !rG       â $r=   ÿŽ â ErG         â &zFj  € ÿ   Ä $rq   ÿŽ ä 0rÿ    A   ä $rBi   ÿŽ â Er        â $rC   ÿŽ â Er        â &zFm  € F   â #r G   ÿ   â 0rÿ    A   â Er        â $rHrF   ÿŽ â Er        â #r   ÿ   â&rjKj   ÿ   â #r       â &rW   ÿ   âO #rJ   ÿ   â &r77W   ÿ   Ä &z W  € ÿ   â #rL   ÿ   â Er H       â &rj9m   j   ä &zp  € ÿ   ä &ro      ä&r7o   7   ä &z o  €     Ä &r55p   ÿ   â #r      â $rj?j   ÿŽ ä $rFOF   ÿŽ ä &zn  €    ä $r;   ÿŽ ä $r7B7   ÿŽ Ä &rAn   5   ä $r==    ÿŽ ä &rKKp   ÿ   ä $rC    ÿŽ â ErI j       â $rrr   ÿŽ âErF F       â $rqq   ÿŽ â Er        â $rO   ÿŽ â Er  =       â &r9n   K   â Er 7       â 0rkÿk    A   â Er        â #rI   J   âx   ÿàÿ â #rFF   L   â $r?   ÿŽ â ErB        â 0rlÿl    A   â Err r       â #rk   ÿ   â#rk    ÿ   â#rk   ÿ   â#rkk   ÿ   â |    pbô â #rlr      âErq q       â #rBlB   k   â Er        â  rV     A  â  rFEF     A  â#rlq      â  rVV     A  â #rl      â  rEEB     A  â #r D     â#r]   F  â#rD   V  â #r]   E  â x--À  ÿà÷ â !r/ /       â !r       â !r       â !r       â x      ÿàÿ Ä x$$    ÿàÿ ä x""    ÿàÿ ä x&&    ÿàÿ ä x((    ÿàÿ ä x**    ÿàÿ ä x,,    ÿàÿ Ä r'ÿ'   ÿäÿ â D)       À â Gy  àßÿÿÿÿƒ ê Ay        € ê Ey      € â G‰  ð     € ê $xÀ  Ž Ê ˆs 3 ùÿ H   è ˆs €ùÿ H   è ˆs  úÿ H   è ˆs €úÿ H   è ˆs  ûÿ H   è ˆs €ûÿ H   è ˆs 1 üÿ H   è ˆs €üÿ H   è ˆs  ýÿ H   è ˆs €ýÿ H   è ˆs / þÿ H   è ˆs €þÿ H   è ˆs  ÿÿ H   è ˆs €ÿÿ H   ä Ay        € ê {          ì M        € ê „y      H   b z  r  p`ð Æ „y €  H   ¢ r ÿ   p"p Æ „y €  H   è „y    H   ( „y    H   h „y €   H   ( „y    H   ( „y	 €  H   ( „y    H   è „y €  H   â !r  3       Æ/ „y#    H   â z3  r      â !r       äO ‰       " !r       Æ� „y    H   b !r       Æ„y €  H   ¢ !r       Æ„y €  H   â !r       Æ y      '   â !r11       Æ ‰      â !r	       â !r        ä ‰      ( y       &   ( ‰      b !r##       Æ ‰      h ‰
      b !r//       Æ/ ‰      b !r!       ÆO ‰1      ¢ !r       â !r       â !r%       â $z   x  Ž â‰$/      " x   ÿàÿ ä $z  ~   Ž â� !r       â !r       â !r       â!r       â ‰      â 
  r  ÿàõ â !r

       Ä ‰       " ÿÿ   ÿä â !r       ä/ ‰"      b 
 j  ÿ„ â !r1       äO ‰ €    ¢ 
 k   ä z  r  p`ò â ‰      b !r$/$       â r133   ÿŽ Ä ‰&!      " x    p"ò ä r/33   ÿŽ â ‰(#      " !r       Æ� ‰*%      â !r         Æ‰	      " !r""       Æ/ ‰      b !r       âO r3    ÿˆ ä ‰      ¢ !r       â /   ÿàû ä ‰
      b !r&!&       â r.ÿÿ   ÿä Ä ‰      " !r(#(       â z j  ÿˆ â !r*%*       â� ‰!      â ,ÿÿ   ÿäÿ â !r	       ä‰#       " r	3    ÿ„ â !r       Ä/ ‰%"      b 
 j  ÿŠ â !r       äO ‰ @    ¢ rÿÿ   ÿä â !r
       ä‰'$      b  1   ÿà÷ â !r       Ä ‰ €    " ÿÿ   ÿäÿ ä 

 j  ÿ† â ‰)&      " z	 j  ÿ„ â !r!!       â� 
 k  � â ‰+(      â !r# #       âz		 k   Ä ‰ €    " !r%"%       â/ z k  . ä ‰ €    b !r       âO 3   ÿàÿ ä ‰-*      ¢ !r'$'       â    ÿàù Ä ‰ €    b !r       â 
 k  ,� ä ‰ €    " !r)&)       â .ÿÿ   ÿä ä ‰ €    " !r+(+       â�  j  ÿˆ Ä ‰"! €    â !r       â k  . â $x3   ÿŽ â ‰$# €    " !r       â/ /   ÿàÿ ä ‰&% €    b !r-*-       âO r    ÿàõ Ä ‰(' €    ¢ !r       ârÿÿ   ÿä ä ‰ @    b !r       â z j  ÿ„ ä ‰*) €    " !r       â /    ÿà÷ Ä ‰,+ €    " !r"!"       â� z k   ä ‰ @    â !r$#$       â0ÿÿ   ÿäÿ ä ‰ @    " !r&%&       â/ / j  ÿ† Ä ‰.- €    b !r('(       âO / k  0� ä ‰ @    ¢ !r       âx        ä ‰ @    b !r*)*       ä %v  j   Ž â ‰      " !r,+,       Æ ‰  @    " !r       Æ� ‰!" @    â !r       Æ‰#$ @    " !r.-.       Æ/ ‰%& @    b !r       ÆO ‰'( @    ¢ !r       Æ‰      b !r/       â 1   ÿàÿ ä ‰)* @    " !r         â     ÿàõ ä ‰+, @    " !r!"!       â� 0ÿÿ   ÿä Ä ‰      â !r#$#       â j  ÿ„ ä‰-. @    " !r%&%       â/  k  0 ä ‰      b !r'('       ÆO ‰      ¢ !r1       Æ‰      b !r)*)       Æ ‰       " !r+,+       Æ ‰ !      " !r       Æ� ‰#      â !r-.-       Æ‰"%      " !r       Æ/ ‰'      b !r       ÆO †	 /    â !r       Æ† 1   â !r        Æ ‰$)      " !r!!        Æ ‰+      ¢ !r##       Æ� ‰-      â !r%%"       Æ†	     â	 !r''       Æ/ †     è	 †	     è	 †    â	 !r))$       Æ †	     â	 !r++       ÆO † !    â	 !r--       Æ� †	 #    è	 † %   è	 †	 
'    è	 † )    è	 †	 +    â	 M™        € ê †y -   â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      "   " ¹z  l      â Ey  p!    € â ™x?    â y      !   b ¹z  F   
   â r ÿ       â x      ÿ  â y      %   ¢ r ÿ       â x      ÿ  â r' ÿ       â y	      &   â r) ÿ       â x      ÿ  â r ÿ       Ä r+ ÿ       ä r ÿ   pRð är ÿ   pRò är   ÿ(Ž È/ xÿ     ä x   ÿ   äO r   ÿ � È xÿ     È |    pbô Ú G)  ð    € ê y      '   "Ž r ÿ       â $z
 p  Ž â xðÿÿÿÿÀŽ ä r ÿ       ä r  €ÿàÿ ä z  X      ä xÿ     Ä z!  Y      ä r   ÿ� ä r ÿ       ä xÿ     ä xüÿÿÿÿÀŽ ä x   ÿ   â 'z  y   Ž â r  €ÿàÿ Ä x&ÿ     â 'z	 s   Ž â zÿ z     ä x%   ÿ   â $z 	 w  ÿŽ â zÿ t    â $z |  
Ž â x
   ÿàÿ â $z |  Ž ä $z v  Ž Ä $z v  Ž â x   ÿ   â $z }  ÿŽ ä $z
 p  Ž â z  q      â %x
°   ÿŽ â r    àµ ä r$   ÿŽ â$x   Ž âr    ÿŽ â$x"   Ž â r%   ÿ(Ž â %x°   ÿŽ â x ÿ     Ä r&ÿÿ   &d~ ä x#      ä x.
   øŽ ä xÿ   %  ä r(%
   ÿà÷ ä xÿ     Ä r*
   ÿàõ är,   ÿàû är0%   ÿàù ä x8   øŽ ä z q  ÿàÿ ä x..   ÿüŽ Ä r2   ÿäÿ är4   ÿä är6   ÿäÿ äx88   ÿüŽ ä r   ÿä ä r:.   ÿàõ È r;!   ÿä ä r<8   ÿàõ Æ �yP:øÿÿ • ¢  r=!   ÿä Æ �yQ:    • â  r>*   ÿàõ Æ �yI<   • " r@(   ÿà÷ ä r?!4   ÿä â�yO:üÿÿ • b rA!2   ÿäÿ Æ �yR<üÿÿ • b rB,   ÿà÷ Æ �y1>0   ™ b r   ÿàõ Æ �y/>@   ™ b r:0   ÿàù Æ �yD@    ™ b rC!6   ÿäÿ ä x` $       â �yE<    • b r ÿ&   ÿä Æ �yF@   ™ b r;!   ÿä â %v> Z  ` Ž ä/ �y-B0   ™ d $x $   ÿŽ ä �yB@   ™ h �yJ:    ™ b r@%>   ÿàõ Ä r??   ÿàÿ â �yH:   ™ f rA?#   ÿä â �yG>    • b r=   ÿàõ Æ �y @8   ™ h �y@   ™ b r3ÿ&   ÿä Æ �y9@(   ™ h �y@   ™ b %v<= Z  ` Ž Æ �yL>$   • b $x33$   ÿŽ â r:%<   ÿàõ È/ r==3   ÿàÿ È r;=#   ÿä â �yN<    • h �y7:   ™ h �y5:8   ™ h �y3:   ™ h �yK:(   ™ b r>
   ÿàõ Ä rC$   ÿàù â �yS<$   • b r@   ÿà÷ ä r?!   ÿä ärTÿ&   ÿä â %vBC Z  ` Ž â rA!   ÿäÿ â �yM>    ™ d $x;T$   ÿŽ â/ r:%B   ÿàõ â �yU@    ™ f rCC;   ÿàÿ È r;C#   ÿä â �y\B    • b r?   ÿàõ Æ �yW:   ™ b r<ÿ&   ÿä Æ �yV:8   ™ h �yT:   ™ b %v>? Z  ` Ž Æ �yX:(   ™ b $x=<$   ÿŽ â r<%>   ÿàõ ä �y_B$   • d r??=   ÿàÿ È r=?#   ÿä â �yb>    • h �yY<   ™ h �yZ<8   ™ b rA    ÿàõ Æ/ �y[<   ™ h �yB<(   ™ b r:ÿ&   ÿä â %v@A Z  ` Ž ä �yc>$   • d $x;:$   ÿŽ â r:%@   ÿàõ È rAA;   ÿàÿ È r;A#   ÿä â �yd@    • b r?"   ÿàõ Æ �yC:   ™ b r<ÿ&   ÿä Æ/ �y]:8   ™ h �y^:   ™ b %v>? Z  ` Ž Æ �ya:(   ™ b $x=<$   ÿŽ â r<%>   ÿàõ ä �ye@$   • d r??=   ÿàÿ È r=?#   ÿä â �yA>    • h �y`<   ™ h	 �yg<   ™ h	 �yf<8   ™ h	 �y@<(   ™ b	 xhP??  ÿÀŽ ÄO xiPÀÀ  ÿÀŽ ä �yP>$   • ¢  x    pBô ä x;Q  ÿÀŽ ä� x<ÿ   I  Ò": ÿ       Ê '(hi   @: Ž â x:ÿ   Q  ä xkOÀÀ  ÿÀŽ äx;:  ÿÀŽ ä ": ÿ       ä x?<  ÿÀŽ ä xiO??  ÿÀŽ Ä xmRÀÀ  ÿÀŽ äx=I  ÿÀŽ ä "< ÿ       â '(ik   @: Ž â xQR??  ÿÀŽ ä x:1ÿÀŽ â '(Qm   @< Ž â x1ÿ   1  Ä x</ÿÀŽ ä rDÿ   D  ä x/ÿ   /  ä x;1ÿÀŽ ä x1D   ÿ   ä xOE??  ÿÀŽ Ä rFÿ   F  ä x=/ÿÀŽ ä xDD   ÿ   ä xEEÀÀ  ÿÀŽ ä "> ÿ       ä x<<1øŽ Ä x/F   ÿ   äx1=DøŽ â '(OE   @> Ž â xFF   ÿ   ä xD-ÿÀŽ ä x-ÿ   -  ä x>ÿÀŽ Ä rJÿ   J  äx::/øŽ ä xÿ     ä x/;FøŽ ä rHÿ   H  ä x;-ÿÀŽ Ä x-J   ÿ   äx=ÿÀŽ ä xJJ   ÿ   ä xH   ÿ   äxHH   ÿ   ä x>>-øŽ Ä x-=JøŽ â &zj   € ÿ   â x;;HøŽ â &rH1    ÿ   âxDDøŽ â &r -    ÿ   ä &r<   ÿ   ä&zF  € ÿ   ä &r>   ÿ   â xEi?   ÿÀŽ â &rR/9   H   Ä &zj9  € j   ä &rk;9       â x9h?   ÿÀŽ â &rJ:      âx?Q?   ÿÀŽ â &z  € F   â xHO?   ÿÀŽ â &rlD      â x=ÿ   h  â 0r ÿG    A   â xFÿ   Q  â $rJJ9   ÿŽ â xGÿ   i  â $rE   ÿŽ â xIÿ   O  â $rll?   ÿŽ Ä $rH   ÿŽ â ErO J       â $rRR=   ÿŽ â Er        â $rkkF   ÿŽ â Er l       â $rQjG   ÿŽ âEr        â $rjjI   ÿŽ â ErJ R       â 0rLÿL    A   â Erh k       â #rO O   ÿ   âErR Q       â #r    ÿ   âErj j       â #r    ÿ   â#r    ÿ   â #rJLJ   O   â &r <7   ÿ   â#rOLh      â#rQLR      â&zh7  € ÿ   â #rLLj      â &r7>7   ÿ   Ä &rj15   ÿ   ä &rR:3       ä &zl5  € ÿ   ä &zh3  € h   ä &rp-5   ÿ   ä &rnD3   7   Ä &rj/K   j   ä$rR9R   ÿŽ ä &zlK  € l   ä $rEh   ÿŽ â Er3 R       â &rK;K   p   ä $r?n   ÿŽ Ä $rj=j   ÿŽ ä 0r ÿN    A   ä $rHh   ÿŽ â Er        â $r5Gl   ÿŽ â Er        â $rKFK   ÿŽ â Erj j       â 0rSÿS    A   â Er        â $rlIl   ÿŽ â #rN 3   ÿ   âEr5 5       â #rR    ÿ   âErK K       â #rh    ÿ   âEr l       â #rjSj   N   â#r     ÿ   â 0rNÿM  0 A   â #r5S5   R   â#rKSK   h   â#rhS       â 0r3ÿU  0 A   â  r N5     A  â 0rMÿM    A   â  rQQN     A  â 0rRÿU    A   â  r3h     A  Ä #rMj      â #rJJM   Q  â #r RK     â &zW  € ÿ   â !r       â &r<W   ÿ   ä&rW>W   ÿ   â  rLL3     A  â !r+J+       â !r        â &r 1V   ÿ   Ä &zV  € ÿ   ä &r5-V   ÿ   ä &rJ:T      ä&zT  €    ä &rVDT   W   â #rOOR   L  â &rL/X       Ä &zTX  €    ä &r5;X   5   ä $rJ9J   ÿŽ ä $rV?V   ÿŽ ä $rE   ÿŽ â!rO       â $rH   ÿŽ Ä $rL=L   ÿŽ ä $rOF5   ÿŽ â Er5 J       â $rKGT   ÿŽ â Er V       â 0r ÿ\    A   â Er        â $rTIT   ÿŽ â Er        â 0r_ÿ_    A   â Er7 L       â #rJ 5   ÿ   âErK K       â #rL    ÿ   âErQ T       â #rT    ÿ   â#r     ÿ   â #rJ_7   J   â#r7_K   L   âErO O       â #rL_Q       â &r <Y   ÿ   Ä &zVY  € ÿ   ä &rY>Y   ÿ   â #r5_O   T   â &r1Z   ÿ   ä&zZ  € ÿ   ä &r-Z   ÿ   ä &rT:[       Ä &zV[  € V   ä &r\D[   Y   ä &rX/B      ä&zZB  €    ä &rK;B      ä $rT9T   ÿŽ Ä $r?\   ÿŽ ä $rEV   ÿŽ ä$rHV   ÿŽ ä $rOFK   ÿŽ â ErK T       â $rBGZ   ÿŽ âEr        â 0r ÿb    A   â Er        â $rZIZ   ÿŽ â Er        â 0rcÿc    A   â ErQ O       â #rV    ÿ   âErO B       â #rB K   ÿ   âErS Z       â #rT    ÿ   â$rX=X   ÿŽ â #r     ÿ   Ä #rOcO   T   ä#rTcS       â ErX X       â  r N7     A  â r3L     A  â  rONO     A  ä #rMJ      â #rBcX   B   â #r R5     â &zC  € ÿ   â !r       â #rOMB   O  â &r<C   ÿ   Ä &rB1]   ÿ   ä &zJ]  € ÿ   ä &z^  €    â !r        â &rC>C   ÿ   ä &r :^      ä &rB/a   B   Ä &zJa  € J   ä $rE   ÿŽ ä &r^D^   C   ä $r 9    ÿŽ â Er        â $rB=B   ÿŽ ä $rCGJ   ÿŽ Ä 0rdÿd    A   ä $r^?^   ÿŽ â Er5         â 0reÿe    A   â Er7 B       â $rH   ÿŽ â ErC C       â $rJIJ   ÿŽ â Er ^       â #rBd   ÿ   â#r d5   ÿ   â &r<`   ÿ   â ErK J       â #rCeC   B   â#rJd   ÿ   â &r>`   ÿ   â #r e7       â &z``  € ÿ   â  r<NC     A  â &rL-]   ÿ   Ä &r::g      ä&r11f   ÿ   ä&r--f   ÿ   â #r>M    <  â &rDDg      ä &z`g  € `   ä &zf  € ÿ   Ä $r 9:   ÿŽ ä &r:/@   1   ä&r-;@   -   ä $rD?D   ÿŽ ä &z@@  €    ä $rEE`   ÿŽ Ä $r<H`   ÿŽ â Er        â $r==:   ÿŽ â Er          â $rGG@   ÿŽ â Er: D       â $rII@   ÿŽ â Er@ E       â &ra;a   L   â Er< <       â 0rAÿA    A   â x   ÿàÿ â $raFa   ÿŽ â #rdd   ÿ   â $r-F-   ÿŽ â ErG G       â 0rPÿP    A   âO ErD I       â #rA@   ÿ   â#rA    ÿ   â#rA:   ÿ   â#rAA<   ÿ   â |    pbô Ä ErB =       â #rdeK   d   â Era a       â #rPG      âEr  -       â #rDPD   A   â#rBPB      â #rQcQ   V   â  rT3T     A  â #raea   J   â  r3d     A  â rNN     A  â #rP       â  rD3D     A  â #rTRQ   T  â#r Ra     â#rMMB   N  â #rDR   D  â xÀ  ÿà÷ â !rO       â !rT       â !r)>)       â !r' '       â !rM       â !rD       â x    ÿàÿ Ä x    ÿàÿ ä x$$    ÿàÿ ä x    ÿàÿ ä x      ÿàÿ ä x""    ÿàÿ ä r!ÿ!   ÿäÿ â D)       À â Gy   ãÿÿÿÿƒ ê Ay        € ê� Ey  ð     € â G‰  Ð     € ê $x€  Ž Ê ˆs + úÿ H   è ˆs €úÿ H   è ˆs  ûÿ H   è ˆs €ûÿ H   è ˆs  üÿ H   è ˆs €üÿ H   è ˆs  ýÿ H   è ˆs €ýÿ H   è ˆs ) þÿ H   è ˆs '€þÿ H   è ˆs  ÿÿ H   è ˆs €ÿÿ H   ä Ay        € ê {          ì M        € ê „y      H   b z  r  p`ð Ä z  r      â „y €   H   ¢ r ÿ   p"p Æ „y    H   è „y €  H   ( „y    H   ( „y €  H   h „y    H   h „y €  H   h „y    H   h „y
 €  H   è „y    H   â !r  +       â/ r+   ÿŽ â !r       âO „y €  H   â !r       Æ� ‰       b !r       Æ ‰      " !r       â!r       ä‰      ¢ !r       â !r       È ‰      â !r))       Æ ‰      " !r        â/ !r'
'       ä ‰      b !r       â !r       ä ‰
      ( ‰      b !r       âO !r       Ä ‰)      ¨ ‰'      ¢ !r       Æ� ‰      â !r       Æ‰       " !r       â/ x   ÿàÿ ä ‰      b !r
       â z  r  p`ò â !r       â‰      $ x    p"ò â !r))       âO y      '   ¢ !r''       Æ ‰      b !r       Æ� ‰      â !r        Æ‰      " $z 	 x  Ž â !r       ä/ ‰      h ‰      b !r       Æ ‰)      " $z  ~   Ž ÆO ‰'      ¢ !r       â +   ÿàõ ä ‰      b !r       â� ÿÿ   ÿä ä ‰      â !r       â
 j  ÿ„ Ä ‰ €    " !r       â/ 
 k   ä ‰ €    b !r       â !r))       ä ‰ €    " !r''       ÆO ‰ €    ¢ !r       Æ‰ €    b !r       Æ� ‰ €    â !r       Æ‰ €    " 
  r  ÿà÷ â !r
       ä/ ‰ €    b ÿÿ   ÿäÿ ä 
 j  ÿ† â ‰) €    b r    ÿˆ ä 
 k  � â ‰ ' €    b !r       â !r       âO rÿÿ   ÿä â ‰" €    " !r       âz j  ÿˆ Ä ‰$      ¢ !r       â� z k   ä ‰	 @    â !r       Æ‰
 @    " !r       â/ !r)       ä ‰ @    b !r '        Æ ‰ @    b r'   ÿŽ â !r""       ä ‰ @    "  '   ÿàù â !r$$       äO ‰ @    ¢ &ÿÿ   ÿä â !r	       Ä� ‰ @    â 	   ÿàÿ â!r
       âr    ÿ„ â ‰ @    " ) 	   ÿà÷ â !r       ä/ ‰ @    b 
 j  ÿˆ â !r       Ä‰!  @    b *ÿÿ   ÿäÿ â !r       ä ‰#" @    " r(ÿÿ   ÿä â !r       äO ‰%$ €    ¢ ) j  ÿ† â !r       Ä� ‰      â z
 j  ÿ„ ä 
	 k  & â ‰      â !r       â) k  *� ä ‰      " !r       â/ z k  ( â !r! !       â‰      b x)        â !r#"#       Ä ‰      " !r%$%       ÆO ‰      ¢ !r-       Æ� ‰      â +   ÿàÿ â !r/       ä ‰      b     ÿàõ â%v  j  ) Ž â !r       â‰      " &ÿÿ   ÿä ä  j  ÿ„ â ‰ !      " !r       Ä/  k  & â ‰$% @    b !r       Æ ‰"#      " !r       âO !r       ä� †	 -    â !r       Æ† /   â !r       Æ†	     â !r!!        Æ †     â !r$%$       Æ/ †	     â !r##"       Æ †    è †	     è †     è †	 
    è † 
!   è ‰$      ( †	 #    â M™        € ê x'   ÿàÿ âO !r$       Æ r    ÿàñ È r ÿÿ   ÿä  ä z j  ÿ€ È z k     Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      "   " ¹z  l      â Ey       € â ™x?    â y      !   b ¹z  F   
   â r ÿ       â x
      ÿ  â y      %   ¢ r ÿ       â xJ      ÿ  â r& ÿ       â y/      &   â r) ÿ       Ä r ÿ       ä r ÿ       ä r ÿ   pRð är ÿ   pRò är   ÿ(Ž È/ x ÿ     ä x   ÿ   äO r     ÿ � È xÿ      È |    pbô Ú G)  0    € ê y	      '   "Ž x  ðÿÿÿÿÀŽ â $z p  Ž â r ÿ       â $z/ w  ÿŽ â r   €ÿàÿ ä r. ÿ       ä xÿ     ä x   ÿ   â 'z/ s  . Ž â r   ÿ� Ä z q  ÿàÿ ä zÿ t    â 'z 	 y   Ž â xÿ     Æ $z		 }  ÿŽ â z ÿ z     ä x   ÿ   Æ $z  |  Ž â x   ÿàÿ â $z   |  Ž â r	   àµ â $z v  Ž âxÿ     â $z v   Ž â x üÿÿÿÿÀŽ â %x0°   ÿŽ â z  q      ä r    €ÿàÿ â $z^ p  Ž â r   ÿŽ â$x   Ž âx    ÿ   â %x^^°   ÿŽ â r   ÿŽ Ä r!   ÿ(Ž ä rÿÿ   d~ ä xÿ   !  ä r!0   ÿàõ ä xÿ      ä x#0   øŽ Ä x	ÿ     ä r0   ÿà÷ ä r!!^   ÿàû är%^   ÿàù ä x'^   øŽ ä z  X      Ä z  Y      ä r ÿ       ä x      ä x##   ÿüŽ ä r 1	   ÿäÿ är"1   ÿä Ä r$_   ÿäÿ ä x''   ÿüŽ ä r_	   ÿä ä r,#   ÿàõ È r-1   ÿä ä r6'   ÿàõ Æ �y+,øÿÿ • ¢  r7_   ÿä Æ �y;,    • â  r<   ÿàõ är8   ÿà÷ â �yG,üÿÿ • " r="   ÿä är9    ÿäÿ â �yO6   • b r   ÿàõ Æ �yE6üÿÿ • b r2!   ÿà÷ Æ �yA<0   ™ b r,%   ÿàù Æ �y<@   ™ b xb $       Æ �yT6    • b r ÿ   ÿä Æ �y58    ™ b r3$   ÿäÿ är-   ÿä â �yC8   ™ b %v< Z  b Ž Æ/ �y620   ™ b $x $   ÿŽ Æ �y?,    ™ b rH<   ÿàõ Æ �y72@   ™ b r==   ÿàÿ Æ �y9,   ™ b rI=   ÿä Æ �ya<    • h �yH   ™ b r-   ÿàõ Æ/ �y4H8   ™ b r ÿ   ÿä Æ �yBH   ™ h �y@H(   ™ b %v,- Z  b Ž Æ �y*<$   • b $x	 $   ÿŽ â r2,   ÿàõ È r--	   ÿàÿ È r3-   ÿä â �y ,    • h �yY2   ™ h �y\28   ™ h �yZ2   ™ h �yX2(   ™ b r<0   ÿàõ Æ/ �yS,$   • b r=1   ÿä ä rI   ÿàõ Æ �yU<    ™ b r	ÿ   ÿä â %vHI Z  b Ž È $x		$   ÿŽ â r,H   ÿàõ È rII	   ÿàÿ È r-I   ÿä â �yMH$   • h �yQ,   ™ h �y3,8   ™ h �yP,   ™ h �y,(   ™ b r=   ÿà÷ Æ/ �yNH    • b rV^   ÿàõ ä r	ÿ   ÿäÿ ä rW_   ÿä â %v<= Z  b Ž È $x		$   ÿŽ â �yLV    ™ b r,<   ÿàõ Æ r==	   ÿàÿ È r-=   ÿä â �y(<    • h �yI,   ™ b! rc   ÿàõ Æ �y2,8   ™ h �y	,   ™ b r8ÿ   ÿä Æ �yF,(   ™ b %vbc Z  b Ž Æ �yH<$   • b $xW8$   ÿŽ â rdb   ÿàõ È rccW   ÿàÿ È rec   ÿä â �y8b    • h �y-d   ™ h �y,d8   ™ h �yDd   ™ h �y>d(   ™ h �y:b$   • b x    pBô Ú "V ÿ       ä "b ÿ       ä/ "d ÿ       ä x=+??  ÿÀŽ äOx++ÀÀ  ÿÀŽ ä xW;  ÿÀŽ Ä� x;ÿ   ;  Æ '(=+   @V Ž â x+G??  ÿÀŽ äxGGÀÀ  ÿÀŽ ä xW;  ÿÀŽ ä xRÿ   O  äx<E??  ÿÀŽ â '(+G   @V Ž â xcO  ÿÀŽ Ä xEEÀÀ  ÿÀŽ ä xeR  ÿÀŽ ä xGAÿÀŽ ä xOÿ   A  ä xAÿÀŽ ä xRÿ     Ä x;T??  ÿÀŽ ärÿ   5  äxTTÀÀ  ÿÀŽ â '(<E   @b Ž â x5RÿÀŽ ä rEÿ   C  ä xR   ÿ   â '(;T   @d Ž â xT   ÿ   Ä xVE   ÿ   äx`E   ÿ   ä xARøŽ ä xE6ÿÀŽ ä xAÿ   6  ä r?ÿ   ?  Ä x67ÿÀŽ ä x7ÿ   7  ä rRÿ   9  ä x9?   ÿ   ä x55TøŽ ä x77ÿÀŽ Ä xT?   ÿ   ä xGGVøŽ ä xVR   ÿ   ä x669øŽ ä xCOÿÀŽ ä xAAÿÀŽ Ä xRR   ÿ   ä x77TøŽ â &rW   ÿ   âxEEVøŽ â &z?  € ÿ   â xCC`øŽ â &r6   ÿ   â xAARøŽ â &rb54   ÿ   ä &z94  € ÿ   Ä &rO74   ÿ   ä &rWGB   W   ä&z4B  € ?   ä &rEB      â xB=?   ÿÀŽ â &rbC@   b   ä&z`@  € 9   â x9+?   ÿÀŽ â &rOA@   O   â x@<?   ÿÀŽ Ä x?;?   ÿÀŽ â $rVWB   ÿŽ â x=ÿ   =  â $rR@   ÿŽ â x<ÿ   <  â $rT49   ÿŽ âx+ÿ   +  â $r44?   ÿŽ â x;ÿ   ;  â $rWb=   ÿŽ â ErV V       â 0raÿa    A   â ErT T       â $rOO<   ÿŽ â ErR R       â $r[`+   ÿŽ âEr4 4       â $r`;   ÿŽ â ErW W       â 0r*ÿ*    A   â #r]aV   ÿ   âEr[ [       â #rTaT   ÿ   âErO O       â #rRaR   ÿ   âEr        â #r4a4   ÿ   â #rW*W   ]   â&z]Y  € ÿ   â #rT*[   T   â#rO*O   R   â#r**   4   â &z\  € ÿ   Ä &z4Z  € ]   ä &r[Y   ÿ   ä &zRX  €    ä $r94   ÿŽ ä &r[GZ   [   ä 0r ÿ     A   â Er        â $r[B[   ÿŽ Ä &r`5\   ÿ   ä &rV6Y   ÿ   â #rb    ÿ   â&r`CX   `   â Er[ [       â &rEZ   V   ä &r\7\   ÿ   ä $r]=`   ÿŽ â #r` [   ÿ   â $r@   ÿŽ Ä $r4?4   ÿŽ ä &r[AX   \   â Era        â $rY+R   ÿŽ â Er 4       â $rR;R   ÿŽ ä 0rVÿU    A   ä $r[<[   ÿŽ Ä 0rUÿU  0 A   â ErR R       â 0rSÿS    A   â Er] ]       â #r    ÿ   â ErY Y       â  rTTU     A  â Er4 [       â #rX a   ÿ   â #rRSR      â#r]S]   `   â#rYSY   b   â #rTWV   T  â #r4S4   X   â &rQ   ÿ   Ä &zSQ  € ÿ   ä &rW6Q   ÿ   ä &r 53   ÿ   ä&zQ3  € ÿ   â !rKTK       â &r373   ÿ   ä &rGP      Ä &zTP  € S   â  rYUY     A  â &zX  € Q   ä &rQA   3   ä &rWEP   W   ä $r3B   ÿŽ ä $r9T   ÿŽ â #rYV]   Y  â &rPC       Ä 0r ÿM    A   ä $rW@W   ÿŽ ä $rM?T   ÿŽ â Er        â $r+X   ÿŽ â !rY       â 0rNÿN    A   â ErY 3       â $rS=P   ÿŽ â Er3 W       â $rP<Q   ÿŽ â ErM M       â $rQ;X   ÿŽ â Er        â #rN   ÿ   âErS S       â #rYNY   ÿ   â#r3N3   ÿ   â#rWNM   ÿ   â ErN Q       â #r       âErP P       â 0rMÿL  0 A   â#rS S   Y   â rQU     A  â #rN N   W   â0rLÿL    A   â #r3 P   3   â  rMR     A  â#r VS   Q  â  rNMN     A  Ä #r4L4     â !r        â #r3L3   N  â &r I   ÿ   â !r4       â &zI  € ÿ   â  r**M     A  â &r46I   ÿ   â !r
3
       â &rI72   ÿ   ä &rG	       Ä &z32  € ÿ   ä &z	  €    â #rOOL   *  â &r*52   ÿ   ä &rIAF   I   ä $rB   ÿŽ ä &z2F  € 3   Ä $r 9   ÿŽ ä&r	E	   4   ä $r3?   ÿŽ ä $r4<I   ÿŽ â ErI        â &r*CF   *   â Er         â $r+2   ÿŽ Ä 0r(ÿ(    A   ä $r*=*   ÿŽ â Er        â 0rHÿH    A   â #r(   ÿ   â#rI(I   ÿ   â Er* *       ä #rH      â $r	@	   ÿŽ Ä #rIH*   I   â  rU     A  â &r-   ÿ   ä&r66-   ÿ   ä &z--  € ÿ   â Er	 	       â #r VI     â &r55,   ÿ   ä&r77,   ÿ   Ä &rGGD      ä&rEED   6   ä &z,  € ÿ   â Er3 3       â &zDD  € -   â Er4 4       â &rC>   5   â#r	(	   ÿ   â &r7A>   7   Ä &z>>  €    â !rJOJ       â $r9D   ÿŽ ä$r??D   ÿŽ â #rO(3   ÿ   â $r2;2   ÿŽ â #r3H4   	   â $rGBG   ÿŽ â Er        â $r	@E   ÿŽ Ä $r+>   ÿŽ âEr+ ?       â $r;;>   ÿŽ ä 0r8ÿ8    A   â x   ÿàÿ â $r=   ÿŽ â Er2 2       â $r7<7   ÿŽ â ErG G       â 0r:ÿ:    A   â Er	 	       â #r8   ÿ   âEr        â #r+8+   ÿ   â Er* ;       Ä |    pbô â #rHH2   O   â Er        â #rG8G   ÿ   âEr( 7       â #r	8	   ÿ   â #r:      â#r*:*   +   â#rG:   G   â rMH     A  â #r	:(   	   â  rUU     A  â  r*M*     A  â #rL3     â xÀ  ÿà÷ â #rVVG   U  â #r	L	   *  â !r) )       â !r&&       â !rV       â !r	       â x    ÿàÿ Ä x    ÿàÿ ä x    ÿàÿ ä x    ÿàÿ ä x    ÿàÿ ä rÿ   ÿäÿ â D)       À â Gy  Pçÿÿÿÿƒ ê Ay        € ê� Ey  Ð     € â G‰  °     € ê $x@  Ž Ê ˆs K ûÿ H   è ˆs J€ûÿ H   è ˆs  üÿ H   è ˆs €üÿ H   è ˆs  ýÿ H   è ˆs 
€ýÿ H   è ˆs ) þÿ H   è ˆs &€þÿ H   è ˆs  ÿÿ H   è ˆs €ÿÿ H   ä Ay        € ê {          ì M        € ê „y      H   b $z./ x  Ž â z  r  p`ð Ä „y €   H   ¢ z!  r      ä r ÿ   p"p â „y	    H   â r!!   ÿŽ Æ „y €  H   ( „y    H   ( „y €  H   h „y    H   h „y    H   h „y €  H   è „y €  H   â !r  K       â/ !rJ       ÈO ‰       b !r		       â� !r       ä ‰      " !r       Æ‰	      ¢ !r
       â!r       È ‰
      â !r))       â !r        ä/ ‰      b !r&       Æ ‰       " !r       Æ ‰      b !r       Æ ‰      " !r	       ÆO ‰)      ¢ !r

       Æ� ‰      è ‰      â !r       â/ !r         ä‰      b !r       Æ‰	      " !r       Æ ‰      " !r)       ÆO ‰  €    ¢ !r       Æ� ‰      â !r       Æ ‰
      h ‰      b !r       Æ/ ‰      b !r		       Æ‰      " !r       Æ ‰      " !r        âO !r       ä� ‰ €    ¢ !r
       Æ‰	 €    â x
   ÿàÿ â !r       ä ‰ €    b z 
 r  p`ò â !r       ä/ ‰  @    b x    p"ò â !r       Ä‰ €    " !r       Æ y      '   ( ‰ €    " !r
       ÆO ‰ €    ¢ !r	       Æ� ‰ €    â !r       Æ‰ €    b !r         Æ/ ‰ €    b $z. ~  .Ž â !r       ä‰       " !r       â r!.   ÿ† ä‰
 @    " .   ÿàõ â !r       âO rÿÿ   ÿäÿ â ‰ @    ¢ ÿÿ   ÿä â !r       â� z j  ÿ† â ‰ @    â 
 j  ÿ„ â !r       âz k  � â ‰ @    b 
 k   â !r       â/ r!.   ÿŠ â ‰ @    b 
	. r  ÿàù Ä r ÿÿ   ÿäÿ â ‰ @    b !r#        â !   ÿàÿ ä ‰ @    " !r
       â!.   ÿàõ ä ‰ @    " !r       âO    ÿàÿ Ä ‰ @    ¢ !r       â� "ÿÿ   ÿä â !r       â‰       â 
	 j  ÿˆ â !r       ä/ ‰      b z j  ÿŠ â !r       â x        â ‰      b 
		 k  " â !r       â .   ÿà÷ â‰      " %v
. j   Ž â !r       â"ÿÿ   ÿä â ‰      " ! j  ÿ„ â !r       âO z k   � â ‰      ¢  ÿÿ   ÿäÿ â !r%        Ä� ‰      â  j  ÿ† â !r       ä/ ‰      b ! k  " ä  k   � â ‰      b !r       â!r       Ä †	 
#    â !r       Æ† 
%   â !r       ÆO †	     â !r       Æ� †     â !r       Æ/ †	     â !r       Æ †    è †	     è †     è †	     â M™        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      "   " ¹z  l      â Ey       € â ™x?    â y      !   b ¹z  F   
   â r ÿ       â x      ÿ  â y      %   ¢ r' ÿ       ä r) ÿ       Ä r ÿ       ä r ÿ       ä r+ ÿ       ä r ÿ   pRð är ÿ   pRò är   ÿ(Ž È/ xÿ     ä x   ÿ   äO r    ÿ � È xÿ      È |    pbô Ú G)  °    € ê y      '   " x ðÿÿÿÿÀŽ ä r ÿ       â y      &   b r  €ÿàÿ ä r ÿ       ä x	ÿ     ä x   ÿàÿ Ä r		   ÿ� ä x$   ÿ   ä z  X      ä z  Y      ä r ÿ       ä z"$ q  ÿàÿ â 'z  y   Ž È 'z s   Ž â/ zÿ z     Æ $z  p  Ž â zÿ t    â $z }  ÿŽ â z  q      â $z |   Ž âx ÿ   	  â $z |  Ž â x		üÿÿÿÿÀŽ â $z w  ÿŽ â x    ÿ   â $z v  Ž â r		  €ÿàÿ â $z v  Ž â r   àµ â $z p  Ž â x#	   ÿ   â $x    $Ž â xÿ   	  â %x°   ÿŽ â r$   ÿŽ Ä x%ÿ     â %x°   ÿŽ â r #   ÿ(Ž ä x!	     ä r%ÿÿ   %d~ ä x,   øŽ ä xÿ   #  Ä r&#   ÿà÷ äx	ÿ     ä r(   ÿàõ är*   ÿàû är.#   ÿàù ä x6   øŽ Ä x,,   ÿüŽ ä r0   ÿäÿ är2	   ÿä är4	   ÿäÿ äx66   ÿüŽ ä r   ÿä Ä r,   ÿàõ È r   ÿä ä r6   ÿàõ Æ �y;øÿÿ • ¢  r	   ÿä Æ �y?    • â  r(   ÿàõ är
&   ÿà÷ â �y=üÿÿ • " r2   ÿä Æ �y5   • b r0   ÿäÿ Æ �y7üÿÿ • h �y1    • b r*   ÿà÷ Æ �y 0   ™ b r.   ÿàù Æ �y-@   ™ b r$   ÿàõ Æ/ �y:
    ™ b r4   ÿäÿ äxK $       â �y8
   ™ b r9ÿ%   ÿä ä r   ÿä â �y/0   ™ b %v Z  K Ž Æ �y3@   ™ b $x99$   ÿŽ Æ �y>    ™ b r
#   ÿàõ ä/ r		9   ÿàÿ â �y<   ™ f r	!   ÿä â �yD    • h �y9
8   ™ b r"   ÿàõ Æ �y@
   ™ b rÿ%   ÿä Æ/ �yB
(   ™ h �yA
   ™ b %v Z  K Ž Æ �yG$   • b $x$   ÿŽ â r#   ÿàõ È r   ÿàÿ È r!   ÿä â �yI    • h �yF   ™ h �yE8   ™ h �yC   ™ b r
   ÿàù Æ �yH(   ™ b r   ÿàõ ä rÿ%   ÿä â �yJ$   • b r   ÿà÷ ä/ r   ÿä â$x$   ÿŽ â r	   ÿäÿ â %v
 Z  K Ž Ä �y    ™ f r
#   ÿàõ â �yO    ™ b r   ÿàÿ È r!   ÿä Ê �yM
   ™ h �yN
8   ™ h �yL
   ™ b r        ä r    ÿàõ â/ �yP
(   ™ f rÿ%   ÿä â �y    • b %v Z  K Ž Æ �yQ$   • b $x$   ÿŽ â r#   ÿàõ È r		   ÿàÿ È r	!   ÿä â �y    • h �yK   ™ h �yS   ™ h �yR8   ™ h �y(   ™ h �yT$   • b x    pBô Ú "
 ÿ       ä/ " ÿ       ä xU;??  ÿÀŽ äOx;;ÀÀ  ÿÀŽ ä x?  ÿÀŽ ä� xÿ   ?  Ä xW=ÀÀ  ÿÀŽ â'(U;   @
 Ž â x  ÿÀŽ ä " ÿ       ä x
ÿ   5  äx;=??  ÿÀŽ ä x=7??  ÿÀŽ â '(;W   @ Ž â x?7ÀÀ  ÿÀŽ Ä x71??  ÿÀŽ äx
  ÿÀŽ ä x11ÀÀ  ÿÀŽ ä "
 ÿ       ä x ÿÀŽ ä xÿ      Ä x -ÿÀŽ ä x-ÿ   -  ä r:ÿ   :  â '(71   @
 Ž â x	5  ÿÀŽ ä x
-ÿÀŽ ä x-:   ÿ   â'(=?   @ Ž â x	:   ÿ   Ä x- -øŽ ä r8ÿ   8  äx:/ÿÀŽ ä x ÿ   /  ä x/3ÿÀŽ ä r>ÿ   >  Ä x3ÿ   3  ä x1ÿÀŽ ä x8   ÿ   ä x5 ÿÀŽ ä r<ÿ   <  ä x >   ÿ   Ä x33ÿÀŽ ä x>>   ÿ   ä x8   ÿ   ä x8øŽ ä x

	øŽ ä x<   ÿ   Ä x<<   ÿ   ä x33>øŽ ä x// øŽ â &rV
9   ÿ   âx::øŽ â &z9  € ÿ   â x55<øŽ â &r939   ÿ   â x11øŽ â &r -@   ÿ   Ä &z@  € ÿ   ä &r/@   ÿ   â x<=?   ÿÀŽ â &rW5B   9   âx>ÿ   =  â &r@8A       âx9U?   ÿÀŽ â &zA  €    â x=;?   ÿÀŽ â &rA:A      â x?7?   ÿÀŽ â &zB  €    â x;ÿ   ;  â &rV1B   V   â xUÿ   U  â $r@@9   ÿŽ â x7ÿ   7  â $r	A<   ÿŽ Ä $r=   ÿŽ ä$r?   ÿŽ ä $r;   ÿŽ â ErA @       â $rVVU   ÿŽ â Er        â $rBW>   ÿŽ â Er	 	       â $r7   ÿŽ â Er        â 0r ÿD    A   â ErV V       â 0rGÿG    A   â ErW        Ä ErY B       â #r A   ÿ   âEr[        â #r@    ÿ   â#r 	   ÿ   â#r     ÿ   â #rGV      â#r@GW   @   â#rGY      â#rGG[       â &r -F   ÿ   Ä &zDF  € ÿ   ä &rF/F   ÿ   ä &r
E   ÿ   ä &z	E  € ÿ   ä &rB8C       ä &zDC  € D   Ä &rE3E   ÿ   ä &rC:C   F   ä &zVH  € 	   ä &rF1H      ä&rE5H   E   ä $r	<C   ÿŽ Ä $rB9B   ÿŽ ä $r=D   ÿŽ ä$r?D   ÿŽ â Er	 	       â $rE>E   ÿŽ â ErA B       â 0r ÿI    A   â Er        â $rFUF   ÿŽ â Er        â $rC;V   ÿŽ âErB E       â $rV7V   ÿŽ â #rE 	   ÿ   â0rJÿJ    A   â #rA A   ÿ   â#r	    ÿ   â#r    ÿ   â ErF F       Ä Er  C       â 0rÿ  0 A   âErV V       â #rDJB   E   â0rBÿO  0 A   â#rFJF   A   â#r J    	   â#rAJV      â 0r	ÿ    A   â  r@     A  â 0rOÿO    A   â  rGGB     A  â  r       A  â #r	     â  rBA     A  â &zAM  € ÿ   â #rGO   G  â &r
N   ÿ   â #r 	F      â &zCN  € ÿ   â !r++       â &rN3N   ÿ   Ä &r-M   ÿ   ä&zL  € A   â !r        â &rM/M   ÿ   â #rOD     â &r@1P      ä&rA5P   N   ä &r 8L      Ä &zDP  € C   ä $r=   ÿŽ ä &rM:L   M   ä $rE>A   ÿŽ ä $r 9    ÿŽ ä $rU@   ÿŽ â Er@        â $rA;D   ÿŽ Ä 0rÿ    A   ä $rM<M   ÿŽ ä $rC?   ÿŽ â Er          â 0rQÿQ    A   â ErA A       â #r@@   ÿ   â Er M       â $rD7D   ÿŽ â Er C       â #r     ÿ   â Er        â #r@QA   @   â #r   ÿ   â#r   ÿ   â Er D       â #r Q       â  r@@     A  â ErE E       â &r--K   ÿ   â #rQ      â #r	    @  â &r//K   ÿ   Ä &z K  € ÿ   â !r       â #rQE      â  rB     A  â &r88S   -   ä&r/:S   /   ä &z S  €     ä &r

R   ÿ   Ä &r33R   ÿ   ä &zR  € ÿ   â #rO     â $r898   ÿŽ ä $r</   ÿŽ ä &r
1   
   ä&r35   3   Ä $r==    ÿŽ ä&z  €    ä $r?    ÿŽ â Er  8       â $rUU
   ÿŽ â Er
 =       â $r;;   ÿŽ â Er        â $r77   ÿŽ â Er        â 0rÿ    A   â x   ÿàÿ â $r3>3   ÿŽ â Er; ;       â 0rTÿT    A   â Er8 7       â #r-
   ÿ   â#r    ÿ   â#r/   ÿ   â#r   ÿ   â |    pbô â #r T;   -   âEr U       â #rT8      â Er 3       â  r       A  Ä #rT      â  rBB     A  â #rT   /   â x€  ÿà÷ â #r 	      ä #rOO   B  â !rG       â !r       â !r))       â !r' '       â !rO       â x@   ÿàÿ Ä x  @   ÿàÿ ä x""@   ÿàÿ ä x$$@   ÿàÿ ä rÿ   ÿäÿ â D)       À â Gy  Ðêÿÿÿÿƒ ê Ay        € ê Ey  °     € â G‰  �     € ê r   ÿ@Ž Ê ˆs + üÿ H   è ˆs €üÿ H   è ˆs  ýÿ H   è ˆs €ýÿ H   è ˆs  þÿ H   è ˆs )€þÿ H   è ˆs ' ÿÿ H   è ˆs €ÿÿ H   ä Ay        € ê {          ì M        € ê „y €  H   b z  r  p`ð Æ „y      H   ¢ r ÿ   p"p Æ „y €  H   è „y €   H   ( „y    H   h „y €	  H   ( „y    H   ( „y €  H   ( „y    H   ( „y    H   ( „y
 €  H   " !r       Æ/ „y    H   b !r  +       ÆO „y    H   ¢ !r       Æ� „y	 €  H   â !r       Æ„y    H   " !r         Æ„y €  H   b !r       Æ „y#    H   " !r       Æ „y  	  H   " !r       Æ „y  
  H   " !r         Æ „y! €
  H   " !r       Æ „y%    H   " !r

)       Æ „y €  H   â !r'       Æ/ „y €  H   â !r       ÆO „y €  H   â !r	       Æ� ‰       b !r       â!r

       ä‰      ¢ !r#       Æ ‰	      " !r       Æ y      &   â !r       Æ ‰      " !r

!       Æ ‰      b !r%       Æ ‰
      â !r        â/ !r       ä ‰      b !r       âO !r       ä ‰       ¢ !r		       â !r       Ä ‰      " !r       Æ‰	      " !r       Æ‰      b !r
       Æ� ‰      â !r       Æ/ ‰
      b !r         ÆO ‰      ¢ !r       Æ ‰      ( ‰  €    " !r	       â!r       ä‰ €    " !r       Æ� ‰ €    â !r
       Æ/ ‰	 €    b !r       ÆO ‰ €    ¢ !r       Æ ‰ €    " !r        Æ ‰ €    b $z  x  Ž â x   ÿàÿ ä ‰      b !r       âz  r  p`ò ä y      '   " !r       â� x    p"ò Ä ‰ @    â !r
	       â/ !r       äO ‰ @    b !r       Æ ‰ @    " !r       Æ‰
 @    ¢ !r       Æ ‰ @    b $z  ~   Ž Æ‰ @    " z  r      â !r       ä� ‰ @    â 
  r  ÿàõ ä r   ÿŽ â ‰ €    â ÿÿ   ÿä â !r	       â/ 
 j  ÿ„ â ‰      b !r       â r    ÿ† â !r
       âO 
 k   Ä ‰      â    ÿàÿ â !r       ârÿÿ   ÿäÿ â ‰	      " z j  ÿ† â!r       â    ÿàõ â ‰      ¢ z k  � â !r       â�     ÿà÷ â ‰      â ÿÿ   ÿä â !r       â  j  ÿ„ â ‰      " ÿÿ   ÿäÿ â !r       â/ x        â ‰ @    b 

 j  ÿ† Æ ‰      b %v  j   Ž â 
 k  � â !r	       â 	 k   â !r       â †	     â !r       ÆO †    â !r       Æ� †	     â !r       Æ†     â !r       Æ/ †	     â !r       Æ†    è ‰      h  †	 
    â M™        € ê x   ÿàÿ â !r       Æ/ r    ÿàñ È r ÿÿ   ÿä  ä z j  ÿ€ È z k     Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      "   " ¹z  l      â Ey  @    € â ™x?    â y       !   b ¹z  F   
   â r ÿ       â $rÿÿ   ÿ Ž â y	      %   ¢ r# ÿ       ä r! ÿ       Ä r ÿ       ä r ÿ   pRð är ÿ   pRò är    ÿ(Ž È/ xÿ     È r   ÿ � â $À    Ž â r ÿ       ä xÿ     ä x	   ÿ   äO |    pbô Ú G)  ð    € ê y      '   " x	ðÿÿÿÿÀŽ â $vÿ q  ÿ Ž â r ÿ       â y      &   b r
 ÿ       â $x   	
Ž â x   ÿàÿ ä x   ÿ   ä xÿ     Ä z  X      ä r   ÿ� È xÿ     ä xüÿÿÿÿÀŽ â 'z y   Ž Æ r  €ÿàÿ â 'z
 s  
 Ž â/ z	ÿ z    ä xÿ     â $z p  Ž â z
ÿ t  
  â $z	 |  Ž â x   ÿ   â $z	 |  Ž â x	   ÿ   â $z
 v  Ž â x     â $z }  ÿŽ â xÿ   	  â $z w  ÿŽ Ä $z
 v  Ž â z
  Y      â $z p  Ž â r   	àµ â %x$°   ÿŽ â r   ÿ(Ž ä rÿÿ   d~ â %x&°   ÿŽ â xÿ     ä r$   ÿàõ â $x   Ž â xÿ     Ä r$   ÿà÷ ä r&   ÿàû äx$   	øŽ är&   ÿàù ä x&   	øŽ ä r%   ÿäÿ Ä r%   ÿä â $x'   Ž âr ÿ       ä z q  ÿàÿ ä x   ÿüŽ ä x   ÿüŽ ä r'   ÿä Ä r(   ÿàõ È r)%
   ÿä ä r4   ÿàõ Æ �y,(øÿÿ • ¢  r5'
   ÿä Æ �yI(    • â  r0   ÿàõ är2   ÿà÷ â �yD4   • " r1
   ÿä Æ �yG(üÿÿ • b $x3
   Ž â r<   ÿàõ ä �yH4üÿÿ • b r*   ÿà÷ Æ �y4    • b r(   ÿàù ä x- $       â �y00   ™ b r+
   ÿäÿ är)
   ÿä â �y"2   ™ b r4ÿ   ÿä Æ/ �y 0@   ™ b %v<< Z  - Ž Æ �y.2    ™ h �y/*0   ™ b $x14$   ÿŽ Æ �y0*@   ™ b r:<   ÿàõ â $x==   1Ž ä �y4(    ™ h �y2(   ™ b! r;=   ÿä Æ �y7<    • h �y1:   ™ h �y3:8   ™ b r(   ÿàõ Æ �y5:   ™ b r*ÿ   ÿä Æ �y6:(   ™ b %v(( Z  - Ž Æ �y8<$   • b $x+*$   ÿŽ â r*(   ÿàõ È r))+   ÿàÿ È r+)   ÿä â �y=(    • h# �y9*   ™ b r@   ÿàù Æ �y:*8   ™ h �y;*   ™ b rJ$   ÿàõ ärAÿ   ÿä â �y<*(   ™ b rL&   ÿà÷ â $xK%   
 Ä $xAA$   ÿŽ â �y>($   • b rM'
   ÿäÿ Æ �y?J    ™ b %v*@ Z  - Ž Æ/ �y@L    ™ d r(*   ÿàõ ä r++A   ÿàÿ È r)+   ÿä â �yF*    • h �yA(   ™ h �yC(8   ™ h �yB(   ™ h �yE((   ™ h �yJ*$   • b x    pBô Ú "* ÿ       ä x   ÿàÿ ä x€  ÿà÷ ä x@   ÿàÿ ä x@   ÿàÿ â $r
ÿÿ   
Ž â x@   ÿàÿ Ä xK,??  ÿÀŽ äOxM,ÀÀ  ÿÀŽ â/ $",ÿÿ   ÿ Ž â x-I  ÿÀŽ ä� x(ÿ   I  Æ '(KM   @, Ž â x,ÿ   D  äxLG??  ÿÀŽ äxMGÀÀ  ÿÀŽ ä xGH??  ÿÀŽ äxIHÀÀ  ÿÀŽ ä x)(  ÿÀŽ Ä xH??  ÿÀŽ ä x-,  ÿÀŽ ä "( ÿ       ä xÀÀ  ÿÀŽ ä ", ÿ       â '(LM   @( Ž â x+D  ÿÀŽ Ä x(ÿÀŽ â '(H   @, Ž â r"ÿ	   "  äxÿ     ä x ÿÀŽ ä x ÿ      ä r.ÿ	   .  â '(GI   @* Ž â x)"   ÿ   Ä xÿÀŽ ä x-"   ÿ   â $x+.   ÿ Ž â x  ÿÀŽ ä xI.   ÿ   ä x*0ÿÀŽ ä r4ÿ	   4  Ä x0ÿ   0  ä x(()øŽ ä x)-øŽ ä x./ÿÀŽ ä r2ÿ	   2  ä x  IøŽ Ä x/ÿ   /  ä x-4   ÿ   ä x,0ÿÀŽ ä xI4   ÿ   ä x"+øŽ ä x+2   ÿ   Ä x//ÿÀŽ ä x22   ÿ   ä x**-øŽ ä x,,IøŽ â &r"1   ÿ   âx..+øŽ â &z1  € ÿ   â x+/2øŽ â &r1*1   ÿ   Ä &r/ 3   ÿ   ä&z-3  € ÿ   ä &r3,3   ÿ   ä &r(5      ä&z5  €    ä &r5.5   1   â x1K?   ÿÀŽ â &r/)6   /   Ä &z-6  € -   â x2G?   ÿÀŽ â &r6+6   3   â x3L?   ÿÀŽ ä xIH?   ÿÀŽ â $r1   ÿŽ â xGÿ   G  â $r552   ÿŽ â xLÿ   L  â $rD3   ÿŽ âxHÿ   H  â $r4I   ÿŽ â x0ÿ   K  â $r66G   ÿŽ â ErM        â $rK-L   ÿŽ âEr D       â 0r7ÿ7    A   â Er 5       â $r--H   ÿŽ â Er4 4       â 0r8ÿ8    A   â ErK K       â #r7   ÿ   âEr5 6       â #r7   ÿ   âEr- -       â #r474   ÿ   â #rK8K      â#r585      â &r"9   ÿ   â #r-8-   4   â &z9  € ÿ   Ä &r4*9   ÿ   â #rM7M   ÿ   â &r9 :   ÿ   ä&z7:  € ÿ   ä &r(;      ä &z;  €    ä &r:,:   ÿ   Ä &r;.;   4   ä $r//0   ÿŽ ä $r1   ÿŽ ä $r43   ÿŽ ä &r9)<   9   ä&r6+<   :   Ä $r;2;   ÿŽ â Er/ /       â &z7<  € 7   â Er<        â $r:I   ÿŽ â Er 4       â 0r=ÿ=    A   â Er ;       â $r909   ÿŽ ä $r6G6   ÿŽ â #r/8/   M   â $r8L7   ÿŽ â Er4 :       â 0r>ÿ>    A   â ErM 9       â #r;=   ÿ   âEr6 6       â #r9=   ÿ   â0rÿ?  0 A   â#r:=<   ÿ   â#r==4   ÿ   â Er8 8       â #r9>6   9   â0rÿ?    A   â  r6K     A  â 0r4ÿ@  0 A   â #r;>8   ;   Ä #r6/   6  â &r/*A   ÿ   â r8-4     A  â &r-"A   ÿ   ä &zAA  € ÿ   ä &r  C   ÿ   ä&r,,C   ÿ   ä &r((B   -   Ä &r/.B   /   ä &z*C  € ÿ   ä &zBB  € A   ä &r))E       ä&r,+E   ,   ä &zEE  € *   Ä $r33B   ÿŽ ä$r-IB   ÿŽ ä $r7H7   ÿŽ ä $rLLE   ÿŽ â Er3 3       â $r(1(   ÿŽ â Er- -       â $r/2/   ÿŽ Ä $rEHE   ÿŽ ä 0rFÿF    A   â Er7 7       â $r00)   ÿŽ â Er) (       â $r,G,   ÿŽ â Er+ /       â 0rJÿJ    A   â ErL L       â #r3F3   ÿ   âEr  E       â #r-F-   ÿ   â |    pbô â #r7>7   =   â Er0 0       â #r)F)   ÿ   âEr, ,       â #r+F+   ÿ   â #r3JL   3   â#r-J    -   â 0r@ÿ@    A   â #rM>M   :   â  r;;     A  â  r747     A  â #r)J0   )   â#r+J,   +   â  r33     A  â  r-4-     A  â #r:M   ;  â #r85@   8  â #r"@9   7  â #r)   3  â #r@@+   -  â !r6       â !r#:#       â !r!8!       â !r"       â !r       â !r@       â D)       À â Gy  €îÿÿÿÿƒ ê Ay        € ê ˆ  ýÿ H   è ˆ !€ýÿ H   è ˆ # þÿ H   è ˆ €þÿ H   è ˆ  ÿÿ H   è ˆ €ÿÿ H   è {          ì M        € ê „y      H   " z  r  p`ð Æ „y	     H   b r  ÿ   p"p Æ „y  €   H   ¨ „y     H   è „y     H   ( „y  €  H   h „y
     H   h „y  €  H   h „y     H   h „y  €  H   h „y  €  H   b !r       Æ „y     H   " !r	       Æ/ „y  €  H   b !r!       ÆO „y     H   ¢ !r#       Æ� „y  €  H   â !r       Æ„y%  €  H   â !r       Æ„y     H   â !r

       Æ „y'  €  H   â !r       â !r       ä ‰      " z  r      â !r       ä y      &   b !r       â !r

       â !r       â/ !r       ÈO ‰      " !r       â� !r       ä‰      b !r%       Æ ‰      ¢ !r
       Æ ‰
      â !r''       Æ ‰      ( ‰'      b !r       â !r       È/ ‰	      " !r       ÆO ‰      b !r

       Æ� ‰ €    ¢ !r       Æ‰
      â !r'       Æ‰      ( ‰      b !r		       â !r       È/ ‰	 €    " !r       ÆO ‰ €    b !r
       Æ� ‰ @    ¢ !r       Æ‰
 €    â !r       Æ‰ €    ( ‰ €    b !r	       â !r       È/ ‰ @    " !r       âO x   ÿàÿ ä ‰	 @    b $z x  Ž â !r
       â� z  r  p`ò â y
      '   ¢ !r       Æ‰ @    â x     p"ò â !r       ä‰ @    ( ‰ @    h ‰      b !r       Ê ‰       " $z
 ~  Ž âO !r
	       â/ 	   ÿàÿ â!r       ä� 
 r  ÿàõ â ‰
      b r   ÿ† â !r       âÿÿ   ÿä â ‰      ¢ 
 j  ÿ„ â !r       Ä‰      â 
 k   â !r       â 	   ÿàõ â ‰      " x        ä r	ÿÿ   ÿäÿ â !r        â z j  ÿ† â %v j   Ž â ÿÿ   ÿä Ä  j  ÿ„ äz k  	� â !r
       â/ 	 k   â †	     â !r       ÆO †    â !r       Æ� †	     â !r       Æ†     è †	     â M™        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y	      "   " ¹z  l      â Ey  P    € â ™x?    â y       !   b ¹z  F   
   â r ÿ       â $rÿÿ   ÿ Ž â y      %   ¢ r ÿ       â $rÿÿ   ÿ Ž â r 	ÿ   pRð Ä r 	ÿ   pRò â$x	     Ž ä/ $x   ÿ Ž ÆO xÿ     È r   ÿ � ä 	    ÿ8Ž ä xÿ     È |    pbô Ú G)       € ê y      '   " xðÿÿÿÿÀŽ â $r
ÿÿ   ÿ Ž â r ÿ       â y      &   b r  €ÿàÿ â $z	 p  Ž â x   ÿ   â $vÿ X  ÿ Ž ä 'z y  
 Ž â x
ÿ     Æ 'z s   Ž â/ zÿ z    ä r

   ÿ� â $z }  ÿŽ â zÿ t    â $z |  	Ž â x   ÿàÿ â $z |  Ž â x
üÿÿÿÿÀŽ â $z w  ÿŽ â x	ÿ   
  â $z v  Ž âr  €ÿàÿ â $z v  Ž â z  Y      â $z" p  Ž â xÿ     â $x
	   ÿ Ž Ä $x   ÿ Ž âx     â %x""°   ÿŽ â r   
àµ äxÿ   
  â %x °   ÿŽ â r		   ÿ(Ž ä rÿÿ   d~ â $rÿÿ   ÿ Ž â xÿ   	  Ä r	"   ÿàû ä xÿ     ä r    ÿà÷ â$x#   Ž â r	    ÿàõ ä x    
øŽ â $x!   Ž â r"   ÿàù Ä x	"   
øŽ ä r!   ÿä ä z q  ÿàÿ ä x   ÿüŽ ä x	   ÿüŽ ä r#   ÿä Ä r$   ÿàõ Ê $x%!    â r0   ÿàõ È �yA$øÿÿ • ¢  r1#   ÿä Æ �y=$    • â  r*   ÿà÷ är(   ÿàõ â �y;0   • " r+   ÿäÿ Æ �y?$üÿÿ • b $x)    â r&   ÿà÷ ä �y,0    • b r	   ÿàõ Æ �y5*   ™ b r$   ÿàù â $x'   Ž ä�y7(0   ™ b x2 $       â $x%    ä �y@0üÿÿ • b rÿ   ÿä Æ �y/*    ™ d $x$   ÿŽ ä �y9(@   ™ h �y1&@   ™ h# �y-$    ™ b %v(	 Z  2 Ž Æ �y8&0   ™ b $x))   Ž â r*(   ÿàõ ä �y$   ™ d r+)   ÿä ä �y	($   • b r&   ÿàõ Æ/ �y0*8   ™ d $rÿÿ    ä �y4*   ™ b %v&& Z  2 Ž Æ �y.*(   ™ b $x$   ÿŽ Æ �y6*   ™ b r$&   ÿàõ ä r''   ÿàÿ â �y3(    • h $x%'    â �y>&    • h �y$   ™ h �y:$   ™ h �y2$8   ™ b r(    ÿàõ Æ �y<$(   ™ b r)!   ÿä ä r*"   ÿàõ â/ �yB&$   • h $x+#    â �yC(    ™ h	 �y**    ™ b x    pBô Ú "$ ÿ       ä "& ÿ       ä/ x   ÿàÿ ä x€  ÿà÷ ä x@   ÿàÿ ä x@   ÿàÿ Ä rÿ   ÿäÿ ä xDA??  ÿÀŽ äOxAAÀÀ  ÿÀŽ ä x%=  ÿÀŽ ä� x(ÿ   ;  Æ'(DA   @$ Ž â x$ÿ   =  ä xA???  ÿÀŽ äx';  ÿÀŽ ä x??ÀÀ  ÿÀŽ ä x%$  ÿÀŽ â $"$ÿÿ   ÿ Ž â x;,ÀÀ  ÿÀŽ Ä x)(  ÿÀŽ â $"(ÿÿ   ÿ Ž â x+,??  ÿÀŽ ä r5ÿ
   5  â '(A?   @$ Ž â xE@ÀÀ  ÿÀŽ ä x$5   ÿ   â '(+;   @( Ž â x)7ÿÀŽ Ä x7ÿ   7  â $x(5   ÿ Ž â r/ÿ
   /  ä x=@??  ÿÀŽ ä x))$øŽ ä x77ÿÀŽ ä x$1ÿÀŽ â '(=E   @& Ž â x1ÿ   1  Ä r-ÿ
   -  â $x,/   ÿ Ž â x%9ÿÀŽ ä x9ÿ   9  ä x&/   ÿ   ä x/7(øŽ ä x(1ÿÀŽ â $x1-   ÿ Ž â x'9ÿÀŽ â $x--   ÿ Ž â x58ÿÀŽ Ä x8ÿ   8  ä rÿ
     ä x'',øŽ ä x88ÿÀŽ ä x7   ÿ   ä x-(-øŽ Ä x%%&øŽ ä x&   ÿ   â &r('0   ÿ   âx$1øŽ â &z90  € ÿ   â x187øŽ â &r0-0   ÿ   â x55&øŽ â &r$%4   ÿ   ä &z74  € ÿ   Ä &r(/.   (   ä&z,.  € 9   ä &r01.   0   ä &r44   ÿ   ä 0r.ÿ	    A   â x	D?   ÿÀŽ â &r$)6   $   âx;A?   ÿÀŽ â &z&6  € 7   â x9=?   ÿÀŽ â &r656   4   â x?+?   ÿÀŽ â $r$$	   ÿŽ â x7ÿ   D  â $rD&;   ÿŽ â xAÿ   A  â $r669   ÿŽ â x+ÿ   +  â $r4&?   ÿŽ â ErE $       â 0r3ÿ3    A   â Er$ D       â $r8(7   ÿŽ â Er& 6       â $r@,A   ÿŽ âEr( 4       â $r,,+   ÿŽ â #r43E   ÿ   â &r%%   ÿ   â#rE3$   ÿ   â &r   ÿ   â #r$3&   ÿ   â &z  € ÿ   â #r33(   ÿ   â Er, ,       â &r&):   %   Ä &r(5:      â x=ÿ   =  â &z::  €    ä &r''2   ÿ   ä&r--2   ÿ   ä &z2  € ÿ   â #r3.,   3   â $r		&   ÿŽ Ä &r&/<   '   ä&r,1<   -   ä &z<<  €    ä $r;;:   ÿŽ ä $r00=   ÿŽ â Er% 	       â $r(9(   ÿŽ Ä $r?:   ÿŽ â Er	 ;       â $r&7&   ÿŽ ä $rAA<   ÿŽ â Er8 8       â 0r>ÿ>    A   â Er@ @       â $r==,   ÿŽ â Er0 0       â $r++<   ÿŽ â Er (       â 0rBÿB    A   â Er        â #r%>%   ÿ   âEr& &       â #r'>	   ÿ   â Er, A       â #r8.8   4   â#r@.@   E   â#r$.0   $   â Er. =       â #r>   ÿ   âEr+ +       â #r>   ÿ   â 0r	ÿC  0 A   â#r(B&   %   â#r,B,   '   â|    pbô â 0r&ÿ*  0 A   â#r.B.      â#rB+      â  r@	     A  â  r,	,     A  â 0r	ÿ*    A   â  r33&     A  â 0rCÿC    A   â  r&&     A  Ä #r$$	   3  ä #r88C     â #r(C(   ,  â #r		.   &  â !r$       â !r8       â !r(       â !r	       â G©  0òÿÿÿÿƒ ê Ay        € ê ˆ  þÿ H   è ˆ €þÿ H   è ˆ  ÿÿ H   è ˆ €ÿÿ H   è {          ì M        € ê „y      H   " z  r  p`ð Æ „y  €   H   b r  ÿ   p"p Æ „y
     H   ¨ „y     H   è „y	  €  H   ( „y     H   h „y     H   h „y  €  H   h „y     H   h „y  €  H   h „y  €  H   b !r       Æ „y  €  H   " !r       â/ !r

       âO !r       ä� !r	       â!r

       â!r       ä !r       Æ ‰      b !r
       Æ ‰      ¢ !r       Æ ‰
      â !r       È !r       â !r       â/ !r       ÈO ‰      " !r
       Æ� ‰      h ‰
      ¨ ‰      è y      &   " !r       â !r       È/ ‰ €    " !r

       âO !r       È� ‰	
 €    h ‰ €    ¨ ‰      è y      '   b !r       â !r	
	       â/ !r       ÈO ‰	 @    " !r       Æ� ‰ @    h ‰ @    ¨ ‰ €    â !r	       â !r       â/ $tÿ   ÿ Ž â !r       ÄO ‰      â !r
       â� x   ÿàÿ ä ‰      " $z x  Ž âz  r  p`ò ä ‰	
 @    b $z ~  Ž âx     p"ò Ä ‰      ¤ 
  r  ÿàõ È ÿÿ   ÿä ä 
  j  ÿ„ â !r       â %v j   Ž â !r       â 
  k   â !r	
	       ä/ †	     â !r       ÆO †    è ‰ 	      h  †	     â M™        € ê ¹z  r      â !r		        â? �x   ?àÿ Ì |   ÿàñ È rÿÿ   ÿä  ä z j  ÿ€ È z k    Ê †y 	    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y       &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž ä $rÿÿ     Ž È % \   Ž Ê �	    ™ b ¹z  l      â Ey  `	    € â ™x?    â y	      "   b $rÿÿ   ÿ Ž Æ y      !   ¢  ÿ       â $ÿÿ    Ž ä y      '   $ '
 m   Ž â ‚ ÿ       È 
ÿ n    Ê $ÿÿ   
Ž â r 	ÿ   pRò Æ/ $
 o  Ž â r 	ÿ   pRô â$x	    Ž äO $‚ÿÿ    Ž Æ x
ÿ     â 'Š s   Ž Æ r

   ÿ � È xÿ   
  È |    pbö ä Šÿ t    Ö G9  à    € ê y      %   " $rÿÿ   ÿ Ž â x
ðÿÿÿÿÀŽ â $tÿ°   ÿ Ž ä 'z y   Ž â r  €ÿàÿ Æ $rÿÿ   ÿ Ž â zÿ z    ä xÿ     Æ $z	 |  Ž â r   ÿ� â $z w  ÿŽ ä $z v  	Ž âx	üÿÿÿÿÀŽ ä xÿ     â $z }  ÿŽ ä $z p  Ž Ä $x   	
Ž ä %v X  Ž Æ x
ÿ     â $x   ÿ Ž â r        ä x
   
  â$x   ÿ Ž â r	   à± äxÿ     â $x    Ž â x   ÿÀŽ Ä x   ÿ   ä rÿÿ   d~  ä x   ÿüŽ È r"   ÿàù ä r$   ÿà÷ är    ÿàñ â $r#ÿÿ    â r%   ÿ� Ä r	   ÿà÷ ä �y"üÿÿ • ¢  r!   ÿ  Æ �y$   ™ â x $       Æ �y"   • " $rÿÿ   Ž Æ �y 0   ™ " %v Z   Ž Æ �y @   ™ " $x$   ÿŽ Æ �y$    ™ " $x   Ž â r   ÿàñ ä �y"    • $ r
   ÿä  ä �y$$   • (% �y 8   ™ h�y   ™ h �y"(   ™ h �y!   ™ h �y#    • b $rÿÿ    Ž Ä/ $rÿÿ    Ž Ê �y%    ™ b x    pBð ä x   ÿàÿ ä x€  ÿà÷ ä x@   ÿàÿ ä rÿ   ÿäÿ ä x??  ÿÀŽ ÄO xÀÀ  ÿÀŽ ä rÿ     â� $ÿÿ   ÿ Ž â x  ÿÀŽ äx   ÿ   ä xÿÀŽ â '   @ Ž â xÿ     Ä xÿÀŽ ä xÿ     ä rÿ     ä xøŽ â $x   ÿ Ž â xÿ     ä xÿÀŽ â $x   ÿ Ž âx   ÿ   Ä xÿÀŽ ä xøŽ ä xøŽ ä x??  ÿÀŽ ä x  ÿÀŽ â $ÿÿ   ÿ Ž â xøŽ Ä xÀÀ  ÿÀŽ â &z  € ÿ   ä&r    ÿ   ä '   @ Ž â x?   ÿÀŽ Æ &r   ÿ   ä &r"      â x?   ÿÀŽ â &z!  €    ä &z   € ÿ   â xÿ     â &r!      â xÿ     â &z"  €    Ä $r   ÿŽ ä $r   ÿŽ ä $r   ÿŽ â Er        â $r   ÿŽ ä 0r#ÿ#    A   â Er        â 0r$ÿ$    A   â Er        Ä #r#   ÿ   â |    pbð â #r##   ÿ   â Er        â 0rÿ%  0 A   â#r$      â0r%ÿ%    A   ä #r$   #   â  r     A  È #r%     È !r       â G‰  úÿÿÿÿƒ ê Ay        € ê ˆ €ÿÿ H   è {          ì M)        € ê „y     H   " r ÿ   pRð Æ „y €   H   h „y    H   ¨ y	      %   â !r       È !r       È/ !r       ÊO ‰      $ !r       â $z  x  	Ž È� ‰      $ !r       Ê ‰ €    $ !r       Ê ‰ @    $ !r       Ê ‰      b M	        € ê $tÿ   ÿ Ž â !r       â? $z ~  Ž È %v j   Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y	      &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž È %	 \   Ž Ê �	    ™ ¢  $rÿÿ   ÿ Ž â Ey  Ð    € â $rÿÿ   ÿ Ž â y      "   b 'Š	 s   Ž Æ y      !   b '

	 m   Ž Æ y      '   â $rÿÿ   ÿ Ž â 
ÿ n  
  ä rÿ   ÿü‚ â/ $x     Ž Ê xÿ      È r    ÿ � ä Šÿ t    ä y      %   $ r
	       âO G  À    € ê z ÿ `  pPô ä� z ÿ ^  pPò ä z ÿ b  p ö Ä z ÿ a   Sô ä z ÿ _  Sò ä z ÿ c  0't Ö $ ~  ÿŽ âÿ     â$ª ~  ÿŽ ä $
 x  ÿŽ â   àë â$ª
 x  ÿŽ â ¢   àÇ ä¨
ÿ     ä ÿÿ   Äþ Ä    ÿàý ä ¢   ÿàû ä ¢ÿÿ   
„þ â $ÿÿ    â 
 ^  ÿ† äª b  ÿˆ â $¢ÿÿ   Ž â  _  � â $rÿÿ   ÿ Ž â r ÿ       Ä ª c   È �
     h �©     d Ay        € ê� ¹z  l      â x
ÿ     â ™x?    â $ÿÿ   
Ž â Ey  ð    € â $r$ÿÿ   ÿ Ž ä $
 o  	Ž â | 
   pbò â $rÿÿ   ÿ Ž Ä $‚ÿÿ   	 Ž Ô G  €    € ê $rÿÿ   ÿ Ž â xðÿÿÿÿÀŽ â $r$ÿÿ   ÿ Ž â r	 
       â 'z y   Ž â z ÿ `  pPð ä x	   ÿ   â $x    
Ž â z
ÿ z    â $z w  ÿŽ â z ÿ a   Sð Ä x ÿ     â $z

 |  	Ž â x    pBò ä r     ÿ� â $z v  
Ž ä $z
 }  ÿŽ â x üÿÿÿÿÀŽ â $z p  Ž â x ÿ      Ä r

   ÿàõ â $x   
Ž ä $x    ÿ Ž ä$x   ÿ Ž ä %x°   ÿŽ â x   ÿÀŽ ä xÿ     â$x     Ž â z X  ÿàû â $rÿÿ   ÿ Ž â z X  ÿàù â $rÿÿ   ÿ â z X  ÿà÷ â $vÿ Y  ÿŽ â xÿ     Ä x#ÿ     ä z Y  ÿä ä z Y  ÿäÿ ä r4   ÿà÷ â$t-ÿ$   ÿ Ž â r2   ÿàù är.   ÿàõ â $x5   Ž â r,
   àë â $x3    âr/   ÿä Ä �y"4    ™ ¢  r#   ÿÄþ Æ �y2   • è �y2   • ( �y .@   ™ " %v,, Z  - Ž Æ �y%2   • " $x$   ÿŽ Æ �y(4   ™ " $x--   Ž Æ �y'.0   ™ " %x0   ,Ž â z  X  ÿàõ ä �y3,    • (# �y0   ™ ( �y&0   ™ ( �y+08   ™ ( �y)0(   ™ " z! Y  ÿä Æ �y4,$   • ( �y2     ™ " $.ÿÿ   ÿ Ž ä $ ÿÿ   ÿ Ž â x		   ÿàÿ ä r"ÿ   "  ÄO x5??  ÿÀŽ Æ� $x1"   ÿ Ž â/ xÀÀ  ÿÀŽ ä x/  ÿÀŽ äx* ÿÀŽ Æ '5   @. Ž â x*1øŽ ä x*ÿ     ä x%??  ÿÀŽ är(ÿ   (  ä x%%ÀÀ  ÿÀŽ ä x!*  ÿÀŽ Ä x*'ÿÀŽ ä x,ÿ   '  â $x'(   ÿ Ž â x-ÿ      â '%   @  Ž â x ,ÿÀŽ Æ $x%(   ÿ Ž â x'*'øŽ â &r   ÿ   â x--ÿÀŽ ä x""   ÿ   ä x  %øŽ â &z% € ÿ   â x"-"øŽ â &r'&      â x'?   ÿÀŽ â &z(+ € ÿ   â x!5?   ÿÀŽ â &z*& € %   â x%ÿ     â &r""+   ÿ   Ä &z() € (   ä $r'*'   ÿŽ â x5ÿ   5  â $r!!   ÿŽ ä &r" )   "   ä $r(%   ÿŽ â Er% '       â 0r ÿ3    A   â Er! !       â $r""5   ÿŽ â Er         â 0r'ÿ4    A   â #r %   ÿ   â#r !   ÿ   â Er% "       â 0r!ÿ2  0 A   â #r'       Æ #r%'%      â  r"!     A  â 0rÿ2    A   â G‰  @    € ê z2 `  àÙ äz4 `  àµ ä z3 a  ¤~ äz5 a  d~ Æ �y2     ¢  z. `  àµ Æ �y04@    è �y 40    ( �y22    " z/ a  ÿd~ Ê �y!.    ( �y3.    ( �y1.    " z, `  ÿàõ È z- a  ÿä Ê �y,,     $ r-ÿ     äO x40ÿÀŽ Æ� $x5-   ÿ Ž âx.ÿ   0  ä/ rÿ   2  äx045øŽ ä x2 ÿÀŽ â $x/   ÿ Ž ä $x5-   ÿ Ž â x..ÿÀŽ Ä x-2/øŽ â &r/0   ÿ   â xÿ   !  ä x..5øŽ â $x5   ÿ Ž â x23??  ÿÀŽ â $ÿÿ   ÿ Ž â x ÿ      ä x  ÿÀŽ Ä x33ÀÀ  ÿÀŽ ä x01??  ÿÀŽ äx11ÀÀ  ÿÀŽ â '23   @ Ž â x!!  ÿÀŽ ä x4 ÿÀŽ â $ ÿÿ   ÿ Ž â x2?   ÿÀŽ Æ '01   @  Ž â x445øŽ ä x!ÿ   2  â &r..+   ÿ   â x0?   ÿÀŽ â &r--&   /   ä $r*   ÿŽ ä &r.4)   .   â x)ÿ   0  â $r-   ÿŽ Ä $r((!   ÿŽ â Er!        â $r).)   ÿŽ â Er        ä Er( (       â #r !   ÿ   âEr) )       â 0r!ÿ,  0 A   â#r     ÿ   â #r'(      â 0rÿ,    A   Ä #r ')       â  r!!     A  È #r    !  È !r$$       â | 	   pbô â #r%   "  â x€  ÿà÷ ä x@   ÿàÿ â !r       â rÿ   ÿäÿ Î G©  àõÿÿÿÿƒ ê Ay        € ê r ÿ   pRð âEy  €     € â r ÿ   pRò Ö G‰  P     € ê z ÿ `  pPð â $x    Ž Æ z ÿ a   Sð ä ˆs €ÿÿ H   ö ˆ $   H   ä Ay        € ê {          ì M        € ê z ÿ `  pPð â „y      H   f z ÿ a   Sð Ú „‰ €   H   ( „‰	    H   ¢ !r        È/ !‚       È !‚	       âO G‰        € ê „y  €  H   ( „y €   H   h „y    H   ¨ „y    H   è „y	 €  H   " !r  $       â !r       Æ/ !r         âO !r       Æ� !r$ 	       ä‰      " z ÿ `  pPð ä r ÿ   pRò ä z ÿ a   Sð â !r       Ê ‰       $ !r         Ê ‰	  €    $ !r	 	       Ê ‰	 @    $ !r	       Ê ‰      $ !r
       â G‰        € ê ‰ $      $ !r  $       Ê ‰       $ !r        Ê ‰ €    $ !r       Ê ‰ @    $ !r       Ê ‰      $ !r$       ä M        € ê !r
       âG‰  `    € ê $v ÿ h  ÿ Ž â !r$       È x     pBð Ú G	  p    € ê x  ÿÿÿÿÿàÿ È x       € Ê $x     ÿ Ž È ‚{   €     $ xÿ     È Iy ëÿÿÿÿƒ ê  x ;ª¸¿  A  Ì s          $ !t   €?    Ì s        $  r     A  È  r     A  â Gy  P    € ê  x '7=  A  â x*BL?  A  Æ #t   €?    È  r      A  â  x    ?  A  Ê s     $   $ !t  €?    È  r       A  È  r      A  â Gy  °    € ê x     pRð Ú G‰  à     € ê x     pRð Ú G	        € ê 	z  i    � ä 	z i    � Æ  x ;ª¸¿  A  â 	z i €  � Ê s        $ !t  €?    Ì s        $  r       A  È  r      A  â Gy  Ð     € ê  r     A  â Gy  °     € ê 	x   à@  � ä 	x  à@  � Æ  x #ÛÙ?  A  â 	x  àÀ  � Æ  x;ª¸¿  A  ä !t  €?    È s        $ !t  €?    Ì s        $  r       A  È  r      A  ä y      &   " $tÿ   ÿ Ž ä $z x  Ž È $z ~  Ž È %v j   Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y       &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž ä $rÿÿ     Ž È % \   Ž Ê �	    ™ b ¹z  l      â Ey  `	    € â ™x?    â y	      "   b $rÿÿ   ÿ Ž Æ y      !   ¢  ÿ       â $ÿÿ    Ž ä y      '   $ '
 m   Ž â ‚ ÿ       È 
ÿ n    Ê $ÿÿ   
Ž â r 	ÿ   pRò Æ/ $
 o  Ž â r 	ÿ   pRô â$x	    Ž äO $‚ÿÿ    Ž Æ x
ÿ     â 'Š s   Ž Æ r

   ÿ � È xÿ   
  È |    pbö ä Šÿ t    Ö G9  à    € ê y      %   " $rÿÿ   ÿ Ž â x
ðÿÿÿÿÀŽ â $tÿ°   ÿ Ž ä 'z y   Ž â r  €ÿàÿ Æ $rÿÿ   ÿ Ž â zÿ z    ä xÿ     Æ $z	 |  Ž â r   ÿ� â $z w  ÿŽ ä $z v  	Ž âx	üÿÿÿÿÀŽ ä xÿ     â $z }  ÿŽ ä $z p  Ž Ä $x   	
Ž ä %v X  Ž Æ x
ÿ     â $x   ÿ Ž â r        ä x
   
  â$x   ÿ Ž â r	   à± äxÿ     â $x    Ž â x   ÿÀŽ Ä x   ÿ   ä rÿÿ   d~  ä x   ÿüŽ È r"   ÿàù ä r$   ÿà÷ är    ÿàñ â $r#ÿÿ    â r%   ÿ� Ä r	   ÿà÷ ä �y"üÿÿ • ¢  r!   ÿ  Æ �y$   ™ â x $       Æ �y"   • " $rÿÿ   Ž Æ �y 0   ™ " %v Z   Ž Æ �y @   ™ " $x$   ÿŽ Æ �y$    ™ " $x   Ž â r   ÿàñ ä �y"    • $ r
   ÿä  ä �y$$   • (% �y 8   ™ h�y   ™ h �y"(   ™ h �y!   ™ h �y#    • b $rÿÿ    Ž Ä/ $rÿÿ    Ž Ê �y%    ™ b x    pBð ä x   ÿàÿ ä x€  ÿà÷ ä x@   ÿàÿ ä rÿ   ÿäÿ ä x??  ÿÀŽ ÄO xÀÀ  ÿÀŽ ä rÿ     â� $ÿÿ   ÿ Ž â x  ÿÀŽ äx   ÿ   ä xÿÀŽ â '   @ Ž â xÿ     Ä xÿÀŽ ä xÿ     ä rÿ     ä xøŽ â $x   ÿ Ž â xÿ     ä xÿÀŽ â $x   ÿ Ž âx   ÿ   Ä xÿÀŽ ä xøŽ ä xøŽ ä x??  ÿÀŽ ä x  ÿÀŽ â $ÿÿ   ÿ Ž â xøŽ Ä xÀÀ  ÿÀŽ â &z  € ÿ   ä&r    ÿ   ä '   @ Ž â x?   ÿÀŽ Æ &r   ÿ   ä &r"      â x?   ÿÀŽ â &z!  €    ä &z   € ÿ   â xÿ     â &r!      â xÿ     â &z"  €    Ä $r   ÿŽ ä $r   ÿŽ ä $r   ÿŽ â Er        â $r   ÿŽ ä 0r#ÿ#    A   â Er        â 0r$ÿ$    A   â Er        Ä #r#   ÿ   â |    pbð â #r##   ÿ   â Er        â 0rÿ%  0 A   â#r$      â0r%ÿ%    A   ä #r$   #   â  r     A  È #r%     È !r       â G‰  úÿÿÿÿƒ ê Ay        € ê ˆ €ÿÿ H   è {          ì M)        € ê „y     H   " r ÿ   pRð Æ „y €   H   h „y    H   ¨ y	      %   â !r       È !r       È/ !r       ÊO ‰      $ !r       â $z  x  	Ž È� ‰      $ !r       Ê ‰ €    $ !r       Ê ‰ @    $ !r       Ê ‰      b M	        € ê $tÿ   ÿ Ž â !r       â? $z ~  Ž È %v j   Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y	      &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž È %	 \   Ž Ê �	    ™ ¢  $rÿÿ   ÿ Ž â Ey  Ð    € â $rÿÿ   ÿ Ž â y      "   b 'Š	 s   Ž Æ y      !   b '

	 m   Ž Æ y      '   â $rÿÿ   ÿ Ž â 
ÿ n  
  ä rÿ   ÿü‚ â/ $x     Ž Ê xÿ      È r    ÿ � ä Šÿ t    ä y      %   $ r
	       âO G  À    € ê z ÿ `  pPô ä� z ÿ ^  pPò ä z ÿ b  p ö Ä z ÿ a   Sô ä z ÿ _  Sò ä z ÿ c  0't Ö $ ~  ÿŽ âÿ     â$ª ~  ÿŽ ä $
 x  ÿŽ â   àë â$ª
 x  ÿŽ â ¢   àÇ ä¨
ÿ     ä ÿÿ   Äþ Ä    ÿàý ä ¢   ÿàû ä ¢ÿÿ   
„þ â $ÿÿ    â 
 ^  ÿ† äª b  ÿˆ â $¢ÿÿ   Ž â  _  � â $rÿÿ   ÿ Ž â r ÿ       Ä ª c   È �
     h �©     d Ay        € ê� ¹z  l      â x
ÿ     â ™x?    â $ÿÿ   
Ž â Ey  ð    € â $r$ÿÿ   ÿ Ž ä $
 o  	Ž â | 
   pbò â $rÿÿ   ÿ Ž Ä $‚ÿÿ   	 Ž Ô G  €    € ê $rÿÿ   ÿ Ž â xðÿÿÿÿÀŽ â $r$ÿÿ   ÿ Ž â r	 
       â 'z y   Ž â z ÿ `  pPð ä x	   ÿ   â $x    
Ž â z
ÿ z    â $z w  ÿŽ â z ÿ a   Sð Ä x ÿ     â $z

 |  	Ž â x    pBò ä r     ÿ� â $z v  
Ž ä $z
 }  ÿŽ â x üÿÿÿÿÀŽ â $z p  Ž â x ÿ      Ä r

   ÿàõ â $x   
Ž ä $x    ÿ Ž ä$x   ÿ Ž ä %x°   ÿŽ â x   ÿÀŽ ä xÿ     â$x     Ž â z X  ÿàû â $rÿÿ   ÿ Ž â z X  ÿàù â $rÿÿ   ÿ â z X  ÿà÷ â $vÿ Y  ÿŽ â xÿ     Ä x#ÿ     ä z Y  ÿä ä z Y  ÿäÿ ä r4   ÿà÷ â$t-ÿ$   ÿ Ž â r2   ÿàù är.   ÿàõ â $x5   Ž â r,
   àë â $x3    âr/   ÿä Ä �y"4    ™ ¢  r#   ÿÄþ Æ �y2   • è �y2   • ( �y .@   ™ " %v,, Z  - Ž Æ �y%2   • " $x$   ÿŽ Æ �y(4   ™ " $x--   Ž Æ �y'.0   ™ " %x0   ,Ž â z  X  ÿàõ ä �y3,    • (# �y0   ™ ( �y&0   ™ ( �y+08   ™ ( �y)0(   ™ " z! Y  ÿä Æ �y4,$   • ( �y2     ™ " $.ÿÿ   ÿ Ž ä $ ÿÿ   ÿ Ž â x		   ÿàÿ ä r"ÿ   "  ÄO x5??  ÿÀŽ Æ� $x1"   ÿ Ž â/ xÀÀ  ÿÀŽ ä x/  ÿÀŽ äx* ÿÀŽ Æ '5   @. Ž â x*1øŽ ä x*ÿ     ä x%??  ÿÀŽ är(ÿ   (  ä x%%ÀÀ  ÿÀŽ ä x!*  ÿÀŽ Ä x*'ÿÀŽ ä x,ÿ   '  â $x'(   ÿ Ž â x-ÿ      â '%   @  Ž â x ,ÿÀŽ Æ $x%(   ÿ Ž â x'*'øŽ â &r   ÿ   â x--ÿÀŽ ä x""   ÿ   ä x  %øŽ â &z% € ÿ   â x"-"øŽ â &r'&      â x'?   ÿÀŽ â &z(+ € ÿ   â x!5?   ÿÀŽ â &z*& € %   â x%ÿ     â &r""+   ÿ   Ä &z() € (   ä $r'*'   ÿŽ â x5ÿ   5  â $r!!   ÿŽ ä &r" )   "   ä $r(%   ÿŽ â Er% '       â 0r ÿ3    A   â Er! !       â $r""5   ÿŽ â Er         â 0r'ÿ4    A   â #r %   ÿ   â#r !   ÿ   â Er% "       â 0r!ÿ2  0 A   â #r'       Æ #r%'%      â  r"!     A  â 0rÿ2    A   â G‰  @    € ê z2 `  àÙ äz4 `  àµ ä z3 a  ¤~ äz5 a  d~ Æ �y2     ¢  z. `  àµ Æ �y04@    è �y 40    ( �y22    " z/ a  ÿd~ Ê �y!.    ( �y3.    ( �y1.    " z, `  ÿàõ È z- a  ÿä Ê �y,,     $ r-ÿ     äO x40ÿÀŽ Æ� $x5-   ÿ Ž âx.ÿ   0  ä/ rÿ   2  äx045øŽ ä x2 ÿÀŽ â $x/   ÿ Ž ä $x5-   ÿ Ž â x..ÿÀŽ Ä x-2/øŽ â &r/0   ÿ   â xÿ   !  ä x..5øŽ â $x5   ÿ Ž â x23??  ÿÀŽ â $ÿÿ   ÿ Ž â x ÿ      ä x  ÿÀŽ Ä x33ÀÀ  ÿÀŽ ä x01??  ÿÀŽ äx11ÀÀ  ÿÀŽ â '23   @ Ž â x!!  ÿÀŽ ä x4 ÿÀŽ â $ ÿÿ   ÿ Ž â x2?   ÿÀŽ Æ '01   @  Ž â x445øŽ ä x!ÿ   2  â &r..+   ÿ   â x0?   ÿÀŽ â &r--&   /   ä $r*   ÿŽ ä &r.4)   .   â x)ÿ   0  â $r-   ÿŽ Ä $r((!   ÿŽ â Er!        â $r).)   ÿŽ â Er        ä Er( (       â #r !   ÿ   âEr) )       â 0r!ÿ,  0 A   â#r     ÿ   â #r'(      â 0rÿ,    A   Ä #r ')       â  r!!     A  È #r    !  È !r$$       â | 	   pbô â #r%   "  â x€  ÿà÷ ä x@   ÿàÿ â !r       â rÿ   ÿäÿ Î G©  àõÿÿÿÿƒ ê Ay        € ê r ÿ   pRð âEy  €     € â r ÿ   pRò Ö G‰  P     € ê z ÿ `  pPð â $x    Ž Æ z ÿ a   Sð ä ˆs €ÿÿ H   ö ˆ $   H   ä Ay        € ê {          ì M        € ê z ÿ `  pPð â „y      H   f z ÿ a   Sð Ú „‰ €   H   ( „‰	    H   ¢ !r        È/ !‚       È !‚	       âO G‰        € ê „y  €  H   ( „y €   H   h „y    H   ¨ „y    H   è „y	 €  H   " !r  $       â !r       Æ/ !r         âO !r       Æ� !r$ 	       ä‰      " z ÿ `  pPð ä r ÿ   pRò ä z ÿ a   Sð â !r       Ê ‰       $ !r         Ê ‰	  €    $ !r	 	       Ê ‰	 @    $ !r	       Ê ‰      $ !r
       â G‰        € ê ‰ $      $ !r  $       Ê ‰       $ !r        Ê ‰ €    $ !r       Ê ‰ @    $ !r       Ê ‰      $ !r$       ä M        € ê !r
       âG‰  `    € ê $v ÿ h  ÿ Ž â !r$       È x     pBð Ú G	  p    € ê x  ÿÿÿÿÿàÿ È x       € Ê $x     ÿ Ž È ‚{   €     $ xÿ     È Iy ëÿÿÿÿƒ ê  x ;ª¸¿  A  Ì s          $ !t   €?    Ì s        $  r     A  È  r     A  â Gy  P    € ê  x '7=  A  â x*BL?  A  Æ #t   €?    È  r      A  â  x    ?  A  Ê s     $   $ !t  €?    È  r       A  È  r      A  â Gy  °    € ê x     pRð Ú G‰  à     € ê x     pRð Ú G	        € ê 	z  i    � ä 	z i    � Æ  x ;ª¸¿  A  â 	z i €  � Ê s        $ !t  €?    Ì s        $  r       A  È  r      A  â Gy  Ð     € ê  r     A  â Gy  °     € ê 	x   à@  � ä 	x  à@  � Æ  x #ÛÙ?  A  â 	x  àÀ  � Æ  x;ª¸¿  A  ä !t  €?    È s        $ !t  €?    Ì s        $  r       A  È  r      A  ä y      &   " $tÿ   ÿ Ž ä $z x  Ž È $z ~  Ž È %v j   Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y       &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž ä $rÿÿ     Ž È % \   Ž Ê �	    ™ b ¹z  l      â Ey  À    € â ™x?    â y	      "   b $rÿÿ   ÿ Ž Æ y      !   ¢  ÿ       â $ÿÿ    Ž ä y      %   $ '
 m   Ž â ‚ ÿ       È 
ÿ n    Ê $ÿÿ   
Ž â r 	ÿ   pRò Æ/ $
 o  Ž â r 	ÿ   pRô â$rÿÿ   ÿ Ž ä $x	    Ž äO $‚ÿÿ    Ž Æ xÿ     â 'Š s   Ž Æ r
   ÿ � ä 	   ÿ8Ž â $x	   ÿ Ž â xÿ   
  È |    pbö ä Šÿ t    ä x      ÿ  Ò G9       € ê y      '   " x
ðÿÿÿÿÀŽ â $r
ÿÿ   ÿ Ž â z  Y      â $vÿ X  ÿ Ž ä $x   
Ž Ê xÿ     È r   ÿ� È xüÿÿÿÿÀŽ È r  €ÿàÿ â 'z y  
 Ž Æ xÿ     â $z
 }  ÿŽ â zÿ z    â $z	 p  Ž È $z |  Ž â xÿ     â $z w  ÿŽ â x     â$z v  Ž âx   ÿ   â $x   ÿ Ž ä $x   ÿ Ž ä %x °   ÿŽ Æ r
   à± â$x    Ž â xÿ     â$rÿÿ   ÿ Ž â x    øŽ ä r    ÿàû är    ÿàù ä rÿÿ   d~  Ä x   ÿüŽ ä r!   ÿ� är!   ÿ ä r*   ÿà÷ är&   ÿàñ ä x- $       â $x+   Ž âr,   ÿà÷ â $x'     â r"   ÿàñ Ä �y*   ™ ¢  $r
ÿÿ   Ž Æ �y:*    ™ â  %v,, Z  - Ž Æ �yD&0   ™ " r#!   ÿä  â $x
$   ÿŽ ä �y&@   ™ b r,   ÿà÷ â $x--   Ž ä �yC"øÿÿ • h �yB"    • b $x-   Ž Æ �y1"üÿÿ • h �y5   ™ h �y
8   ™ h �y4   ™ h �y3(   ™ b $vEÿ p  ÿ Ž Æ �y0,    • b %x*E°   *Ž Æ �y=,$   • b %x&E°   &Ž Æ/ �y>*    ™ b %x"E°   "Ž Æ �yA&@   ™ h �y?"øÿÿ • h �y"    • h �y7"üÿÿ • h �y2*   ™ h �y@&0   ™ b %x(E°   *Ž â r$    ÿàñ Æ %x,E°   &Ž â �y(   ™ b r%!   ÿä  Æ �y8(    ™ b %x&E°   "Ž Æ/ �y<,@   ™ h �y$    ™ h �y;,0   ™ h �y9&øÿÿ • h �y6&    • h �y&üÿÿ • b %x.E°   &Ž È %x*E°   (Ž È %x"E°   $Ž â�y)*   ™ f %x$E°   ,Ž â�y"    ™ h �y-.    • h# �y**    ™ h �y($0   ™ h �y,$@   ™ h �y$.üÿÿ • h �y..øÿÿ • b! %x"E°   "Ž Ê �y%"    ™ b	 %x&E°   "Ž Ì �y&&    ™ b x    pBð ä rEÿ     äOrÿ   :  Æ� $x+E   ÿ Ž â x"DÿÀŽ äx:ÿ   D  ä xDÿÀŽ äxÿ     â $x/   ÿ Ž ä$x   ÿ Ž â x#ÿÀŽ Ä x+"+øŽ ä x"C??  ÿÀŽ äxB  ÿÀŽ ä xCCÀÀ  ÿÀŽ ä x'#øŽ â/ $ÿÿ   ÿ Ž â xEE   ÿ   Ä x::ÿÀŽ ä xDD/øŽ ä xBÿ   B  â '"C   @ Ž â x#1??  ÿÀŽ ä x::EøŽ â &rDD5   ÿ   â xB  ÿÀŽ â &r''
   ÿ   â x11ÀÀ  ÿÀŽ Ä xB"?   ÿÀŽ â &r++4   D   â x"ÿ   "  â &r':3   '   ä '#1   @ Ž È &z5  € ÿ   ä $r++B   ÿŽ ä $r'"   ÿŽ â x'#?   ÿÀŽ â &z:
  € ÿ   ä &z"4  €    â xBÿ   #  â 0r0ÿ0    A   â Er+ +       â &z#3  € :   Ä $r"'   ÿŽ â Er'        â 0r=ÿ=    A   â #r0+   ÿ   â$rB#B   ÿŽ â Er        ä r>ÿ   >  â #r'='      â Er1 B       â #r:0   ÿ   â xÿ   A  â $xE>   ÿ Ž Ä #r+=1   :   â xÿÀŽ ä x:AÿÀŽ ä x1???  ÿÀŽ äxA?ÀÀ  ÿÀŽ ä x?EøŽ â $xC>   ÿ Ž â x  ÿÀŽ Ä  ÿ       ä x>ÿ     ä x7??  ÿÀŽ â '1A   @ Ž â r2ÿ   2  ä x>  ÿÀŽ ä x77ÀÀ  ÿÀŽ Ä x/@ÿÀŽ ä xB2   ÿ   â '7   @ Ž â x::CøŽ ä x@ÿ   @  â $xA2   ÿ Ž â x//BøŽ â &r::5   ÿ   â x@@ÿÀŽ Ä x?   ÿÀŽ â &r//4   :   â x@@AøŽ â &r??
   ÿ   â x1?   ÿÀŽ ä x:ÿ     â $r"   ÿŽ â x2ÿ   1  â &r?@3   ?   ä $r/   ÿŽ â Er        â $r:#:   ÿŽ Ä $r??2   ÿŽ â Er/        ä rÿ     â #r0   ÿ   âEr1 :       â #r0/   ÿ   â Er? ?       ä r8ÿ   8  â #r/=1      âxÿ   <  â 0r2ÿ  0 A   â #r=?      â $x:   ÿ Ž â x;ÿÀŽ â $x?8   ÿ Ž â x<<ÿÀŽ Ä x78   ÿ   ä xÿÀŽ â  r2+2     A  â x+:øŽ ä x19??  ÿÀŽ ä x<<7øŽ ä x6  ÿÀŽ Ä x99ÀÀ  ÿÀŽ ä x7?øŽ â $ÿÿ   ÿ Ž â x8ÿ   6  Æ '19   @ Ž â x8  ÿÀŽ ä x9ÀÀ  ÿÀŽ äx;ÿ   ;  ä x6??  ÿÀŽ ä x   ÿ   â '69   @ Ž â x;;ÿÀŽ Æ &r<<5   ÿ   â x;øŽ ä x1?   ÿÀŽ â &r++4   <   â x6?   ÿÀŽ â &r77
   ÿ   â x6ÿ   6  â $r++   ÿŽ â x1ÿ   1  â $r"   ÿŽ Ä 0rÿ    A   ä &r3   7   â Er+ +       â $r6#6   ÿŽ â Er        â $r11   ÿŽ â #r''   2  â xÿ   -  â 0rÿ  0 A   â Er 6       â #r60+   ÿ   âr*ÿ   *  â #r80   ÿ   â Er1 1       Ä x  ÿÀŽ â $ÿÿ   ÿ Ž â x+$ÀÀ  ÿÀŽ ä x2,ÿÀŽ â $x7*   ÿ Ž â x,ÿ   ,  ä r)ÿ   )  ä x$??  ÿÀŽ â  r//     A  â #r=1   6   â '+   @ Ž â x6,ÿÀŽ Ä x9*   ÿ   â $x,)   ÿ Ž â x-  ÿÀŽ ä x+.ÀÀ  ÿÀŽ ä x1(ÿÀŽ ä x227øŽ ä x$.??  ÿÀŽ Ä x969øŽ â '$+   @ Ž â x(ÿ   (  ä x11,øŽ â $x))   ÿ Ž â x((ÿÀŽ â &r225   ÿ   â x?   ÿÀŽ â &r99
   ÿ   â x
$?   ÿÀŽ â &r114   2   â x(()øŽ â $r""   ÿŽ â xÿ     â $r1
   ÿŽ â x)ÿ   $  â &r((3   9   â Er "       â $r#   ÿŽ â Er        â $r(()   ÿŽ â x   ÿàÿ Ä Er"        â #r0   ÿ   â |    pbð â #r00   ÿ   â 0r
ÿ%  0 A   â Er (       â 0rÿ&  0 A   â#r=   8   â#r="      â0rÿ    A   â #r=   0   â 0r%ÿ%    A   â  r
     A  â 0r&ÿ&    A   â  r     A  â #r   /  â #r%     â x€  ÿà÷ â #r&     â !r'       â !r       â !r       â !r       â x@   ÿàÿ Ä rÿ   ÿäÿ â D	       À â Gy  0îÿÿÿÿƒ ê Ay        € ê ˆ  þÿ H   è ˆ €þÿ H   è ˆ  ÿÿ H   è ˆ €ÿÿ H   è {          ì M)        € ê „y     H   " z 	 r  p`ò â $z   x  	Ž Ä „y €   H   b r ÿ   pVò  Æ „y    H   ¨ „y    H   è „y €  H   („y    H   ( „y
    H   ( „y €  H   ( „y €  H   ( „y    H   ( „y €  H   " !r       Æ „y €  H   " !r       â/ !r       äO y      '   b !r       â� !r       â!r       ä !r
       â !r       È ‰      ¢ !r       â $z  ~   Ž â/ !r       Æ ‰      b !r       Æ ‰      â !r       Ê ‰      " !r       âO !r       â/ !r       È� ‰      h ‰      ¢ !r       Æ ‰      ( ‰      â !r       â/ !r       ÈO ‰
 €    b !r       Æ ‰ €    " !r       Æ� ‰ €    ¨ ‰ €    â !r

       â/ !r       È ‰
 @    " x	   ÿàÿ â !r       äO ‰ @    b x	   ÿàÿ â !r       ä� ‰ @    ¢ z  r  p`ô ä z  r  p`ð â ‰ @    â x    pVt Ä x    pVp  â !r
       â x
	   ÿàÿ â $t	ÿ   ÿ Ž â !r       ä/ ‰      " z 
 r  p`ö â !r       äO ‰      b x    pVö â !r       Ä� ‰      ¨ ‰      â !r       â !r       â/ %v  j  	 Ž â !r       ÈO †™     â !r       Æ� †©    è †¹    â M	        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      &   " z ÿ \  pPò â ¹z  F   
   Æ z ÿ ]  Sò Ú         Ê % \   Ž Ê �    ™ ¢  r ÿ       â Ey      € ä y      "   d '  m   Ž äy      %   ä 'š s   Ž â  ÿ n     â y      !   &  ÿ   ÿáÿ â y	      '   ( $  o  Ž â ’         Ä r ÿ   pRð ä/ x   ÿ   ä� x    pDp  är
   ÿàÿ är   ÿ(Ž ä z 
 r  pdp  Ä x
ÿ     ä šÿ t    ä x      ÿ  ä r     € ÊO G	  °    € ê z ÿ `  pPò ä z ÿ ^  pPð ä z ÿ b  p ô ä z ÿ a  Sò ä z ÿ _   Sð ä z ÿ c   'ò Ö $
	 ~  ÿŽ âÿ     â$š	 ~  ÿŽ ä $
 x  ÿŽ â    àë â$š x  ÿŽ â ’   àÇ ä ˜ÿ     ä    ÿàõ Ä ÿ   ÿÄþ ä ’   ÿàû ä ’ÿ   ÿ„þ ä ÿ   ÿä ä 
 ^  ÿ„ ä ’ÿ   ÿäÿ Ä š b  ÿ† ä
 _   ä x      ÿ  â š c  � Ê �	     h �™     d Ay        € ê r
   ÿ � â ¹z  l      â $z	 }  ÿŽ â ™x?    â xÿ     â $z   w  ÿŽ â r ÿ       â Ey  ð(    € â x
      ÿ  â |    pbð â x      ÿ  â r    ÿàó â 'z	 y   Ž â r ÿ       â x      ÿ  â r6 ÿ       Ä rÿÿ   ÿäÿ  Ê G	  @(    € ê zÿ z    â $z p  Ž âxðÿÿÿÿÀŽ â x
      ÿ  â x   ÿàÿ â $z  |  Ž â r  €ÿàÿ â $z |  Ž â x   ÿàÿ â $z v   Ž âx ÿ     â $z8 v  Ž â x   ÿàÿ â $z p  Ž âr     ÿ� â $z p  Ž âz  `      â %x:°   ÿŽ â x üÿÿÿÿÀŽ â x      ÿ  â x ÿ      â $z p  Ž â r  €ÿàÿ â %x88°   ÿŽ â x    ÿ   â x      ÿ  â xÿ     â %x<°   ÿŽ â x:   øŽ Ä x     â%x>°   ÿŽ â x   ÿ   ä x8   øŽ ä r3    ÿ(Ž ä x5ÿ     ä r%8   ÿàñ Ä r':   ÿàù är)<   ÿàû är+>   ÿàý ä xAÿ   3  ä r-38   ÿà÷ är/3:   ÿàõ Ä r13<   ÿàó ä r95   ÿä  är;5   ÿä är=5   ÿäÿ är ?5   ÿä ä x!>   øŽ Ä x#<   øŽ ä r33>   ÿàñ ä x5%    ÿàû ä x7-@   ÿàù ä r"9A   ÿäÿ ä z  a      Ä z  X      ä z  Y      ä r ÿ       ä r6 ÿ       ä x   ÿ   ä xÿ     Ä x   ÿüŽ ä x   ÿüŽ ä x!!   ÿüŽ ä x##   ÿüŽ ä r$;A   ÿä är&=A   ÿäÿ  Ä r(?A   ÿä  ä r*ÿ   ÿäÿ ä r,ÿ"   ÿä ä rB-   ÿàñ ärD%   ÿàó ä rC"   ÿä  ä rH   à¡ Ä rE   ÿäÿ  â �yQB@   ™ ¢  xI $       ä r    ÿD~  â �yUD   ™ â rJ   ÿàñ â %vHH Z  I Ž ä �ySD    ™ $ $xA $   ÿŽ â r@5   ÿàó â �yPB0   ™ " rK9   ÿä  Ä rIIA   ÿàÿ ä zY  p      â �yOJøÿÿ • " rLH   ÿàñ ä rA*   ÿäÿ  â �y J    • " rB7   ÿàõ ä rMI   ÿä  â %x@Y°   @Ž â rC,   ÿä â �ycJüÿÿ • ( %xBY°   BŽ â�y2L8   ™ ( �yW@    ™ ( �y.L   ™ ( �y4L(   ™ ( �y^B    ™ ( �y0L   ™ ( �ya@ðÿÿ ™ " %xDY°   JŽ Æ/ �yTBðÿÿ ™ ( �yND    • ( �yRDøÿÿ • " rF8   ÿàñ Æ �y`H    • " rG9   ÿä  Æ �ybH$   • ( �yLF    ™ ( �ydDüÿÿ • " %x@Y°   @Ž È %xBY°   BŽ â/�y]@    ™ & %xJY°   DŽ â�y\B    ™ ( �y[@ðÿÿ ™ ( �y_Jøÿÿ • ( �yXJ    • ( �yZBðÿÿ ™ " %xFY°   FŽ Æ �yMJüÿÿ • ( �yVF    ™ " %xHY°   @Ž È %xDY°   BŽ â�ykH    ™ & %x@Y°   JŽ â/�yfD    ™ ( �ygHðÿÿ ™ ( �yeDðÿÿ ™ ( �yh@    • ( �yj@üÿÿ • ( �yi@øÿÿ • " %xFY°   FŽ Ê �yKF    ™ " %xBY°   FŽ Ê �yYB    ™ " x    pBò ä xEÿ   Q  äO rUÿ   U  ä�rSÿ   S  äxQQÿÀŽ ä xDPÿÀŽ Ä xAU   ÿ   ä/ x@S   ÿ   ä xEEÿÀŽ ä xFS   ÿ   ä xPÿ   P  ä xCDAøŽ Ä xQQ@øŽ ä xBO??  ÿÀŽ äxA   ÿÀŽ ä xOOÀÀ  ÿÀŽ ä @ ÿ       ä xEEFøŽ Ä xPPÿÀŽ ä xUU   ÿ   â 'BO   @@ Ž â rWÿ   W  ä xUPUøŽ â &rEE2   ÿ   â x@B?   ÿÀŽ â &rQQ.   ÿ   â xAÿ   B  â &rBU4   E   â xD^ÿÀŽ Ä xEW   ÿ   â &rCC0   Q   â x ÿ      â $rBBA   ÿŽ â raÿ   a  â $rCC@   ÿŽ â xGDEøŽ ä xA   ÿÀŽ ä xDc??  ÿÀŽ Ä x TÿÀŽ ä xEa   ÿ   ä xccÀÀ  ÿÀŽ ä @ ÿ       ä xE EøŽ â &r G.   ÿ   â x^ÿ   ^  â 'Dc   @@ Ž â xAN  ÿÀŽ Ä xGRÀÀ  ÿÀŽ â &r E0       â xTÿ   T  â &zE.  € ÿ   â xWW   ÿ   ä x^^ÿÀŽ ä xFR??  ÿÀŽ â &zJ0  € E   â xaa   ÿ   â 'FG   @@ Ž â xTTÿÀŽ Ä xW^WøŽ â &zG2  € ÿ   â xAD?   ÿÀŽ ä xaTaøŽ â &rWW2   ÿ   â xEF?   ÿÀŽ â &zG4  € G   â xDÿ   D  â $rAJA   ÿŽ â xFÿ   F  â &rWa4   W   Ä $r@ E   ÿŽ â ErE A       â $rDGD   ÿŽ ä 0r ÿ`    A   â ErA C       â $rWWF   ÿŽ â ErD D       â 0rFÿb    A   â #rE E   ÿ   âErC @       â 0rIÿL  0 A   â ErB B       â #rA A   ÿ   â #rDFD   E   â ErH W       â #rE C   ÿ   â x@ÿ   N  â #rBFB   A   â  rCDI     A  â #rDFH   E   â xA@  ÿÀŽ Ä xEdÀÀ  ÿÀŽ ä@ ÿ       ä xHd??  ÿÀŽ Æ 'HE   @@ Ž â r]ÿ   ]  È xAH?   ÿÀŽ ä xN\ÿÀŽ ä xO]   ÿ   â $r@JA   ÿŽ â r[ÿ   [  ä xONOøŽ ä xN_??  ÿÀŽ Ä xHÿ   H  ä ErE @       ä xAX  ÿÀŽ ä x__ÀÀ  ÿÀŽ ä @ ÿ       ä xIZÿÀŽ Ä xP[   ÿ   â $rHGH   ÿŽ ä 'N_   @@ Ž â xIIPøŽ Æ &rOO.   ÿ   â ErH H       â #rE E   ÿ   â x@N?   ÿÀŽ â &rII0   O   ä #rEFH   E   â x\ÿ   \  â $rHI@   ÿŽ ä 0r@ÿV  0 A   â xAÿ   X  Ä xZÿ   Z  ä x]]   ÿ   â  rIE@     A  â x\\ÿÀŽ ä xPÿ   N  ä xNM??  ÿÀŽ ä xAA  ÿÀŽ Ä xMMÀÀ  ÿÀŽ ä @ ÿ       ä x[[   ÿ   ä xZZÿÀŽ ä x]\]øŽ â 'NM   @@ Ž â x[Z[øŽ Æ &r]]2   ÿ   â xMN?   ÿÀŽ â 0rEÿL    A   ä &r[[4   ]   ä 0rAÿV    A   â xNÿ   N  â $rP[P   ÿŽ â rkÿ   k  â $rMJM   ÿŽ â ErO H       â #rEBE   C  â #rDDA   I  â xAfÿÀŽ â $rNGN   ÿŽ â rgÿ   g  Ä xBk   ÿ   â #rO O   ÿ   â ErH P       ä ErM M       ä x@eÿÀŽ ä xCg   ÿ   ä xQABøŽ Ä xBÿ   h  â #rHFH   O   â ErN N       â #rM M   ÿ   â xO@CøŽ ä xCB  ÿÀŽ â #rMFN   M   â xIjÀÀ  ÿÀŽ ä B ÿ       Ä xfÿ   f  ä xLi??  ÿÀŽ äxNj??  ÿÀŽ ä xAh  ÿÀŽ ä xiiÀÀ  ÿÀŽ ä @ ÿ       â 'NI   @B Ž â xeÿ   e  Ä xffÿÀŽ ä xkk   ÿ   â 'Li   @@ Ž â xeeÿÀŽ ä x@g   ÿ   â &rQQ.   ÿ   â xkfkøŽ ä xAN?   ÿÀŽ Ä xee@øŽ â &rkk2   ÿ   â x@L?   ÿÀŽ â &rOO0   Q   â xNÿ   N  â $rAJA   ÿŽ â xLÿ   L  â &ree4   k   â z ÿ `  pPð â $rOO@   ÿŽ â ErA A       â $rNGN   ÿŽ â z ÿ a   Sð â 0r@ÿK  0 A   â ErO O       â $rLeL   ÿŽ â ErN N       â #rA A   ÿ   â rMM@     A  â 0r@ÿY  0 A   â ErL L       â #rO O   ÿ   â #rAFN   A   â 0rKÿK    A   Ä 0rYÿY    A   â #rLFL   O   â  rAA@     A  â !r
E
       â !r	D	       â #rKHK   M  â #rLLY   A  â G‰  @    € ê r@-   ÿàõ ärB%   ÿà÷ ä rA"   ÿä Ä rC   ÿäÿ ä rD   ÿàõ â �yN@@    ¦ rE9   ÿä â �yOB    è �yQB     ( �yM@0    ( �ySD     ( �yRDüÿÿ  ( �yPDøÿÿ  " rH8   ÿàõ È rI9   ÿä Ê �yVH     " x@ÿ   N  äO rOÿ   O  ä�xB@ÿÀŽ ä rQÿ   Q  äxAO   ÿ   ä xTMÿÀŽ Ä xCQ   ÿ   äx@ÿ   S  ä xEQ   ÿ   ä/ xQTAøŽ ä xNNÿÀŽ ä xA@  ÿÀŽ Ä xIRÀÀ  ÿÀŽ ä@ ÿ       ä xHR??  ÿÀŽ ä xUNCøŽ ä xWBEøŽ â 'HI   @@ Ž â xEPÀÀ  ÿÀŽ Ä xCS  ÿÀŽ ä B ÿ       ä xMÿ   M  ä xDP??  ÿÀŽ ä xMMÿÀŽ â 'DE   @B Ž â x@O   ÿ   Ä xAH?   ÿÀŽ â &rUU.   ÿ   â xMM@øŽ â &rWW2   ÿ   â x@D?   ÿÀŽ â &rQQ0   U   â xHÿ   H  â $rAJA   ÿŽ â xDÿ   D  â &rMM4   W   ä $r@Q@   ÿŽ â ErC A       â $rHGH   ÿŽ Ä $rDMD   ÿŽ â ErA @       â #rC C   ÿ   âErH H       â 0r@ÿV  0 A   âErD D       â #rA A   ÿ   ä #rCFH   C   â0rVÿV    A   â #rAFD   A   Ä  r@C@     A  È #r@AV   @  È !r@       ä G‰  P    € ê r@/   ÿàõ ärB'   ÿà÷ ärA$   ÿä ärC   ÿäÿ ä rD   ÿàõ â �yN@@    ¦ rE;   ÿä â �yOB    è �yQB     ( �yM@0    ( �ySD    ( �yRD     ( �yPDüÿÿ  " xH   ÿàÿ Ê $zIH p  Ž È %xHI°   Ž Ê �yTH     " x@ÿ   N  äO rOÿ   O  ä�xD@ÿÀŽ ä rQÿ   Q  äxAO   ÿ   ä xBMÿÀŽ Ä xCQ   ÿ   äx@ÿ   S  ä xEQ   ÿ   ä xQBAøŽ ä xNNÿÀŽ ä xA@  ÿÀŽ Ä xIRÀÀ  ÿÀŽ ä/@ ÿ       ä xHR??  ÿÀŽ ä xUNCøŽ ä xWDEøŽ â 'HI   @@ Ž â xEPÀÀ  ÿÀŽ Ä xCS  ÿÀŽ ä B ÿ       ä xMÿ   M  ä xDP??  ÿÀŽ ä xMMÿÀŽ â 'DE   @B Ž â x@O   ÿ   Ä xAH?   ÿÀŽ â &rUU.   ÿ   â xMM@øŽ â &rWW2   ÿ   â x@D?   ÿÀŽ â &rQQ0   U   â xHÿ   H  â $rAJA   ÿŽ â xDÿ   D  â &rMM4   W   ä $r@Q@   ÿŽ â ErC A       â $rHGH   ÿŽ Ä $rDMD   ÿŽ â ErA @       â #rC C   ÿ   âErH H       â 0r@ÿT  0 A   âErD D       â #rA A   ÿ   ä #rCFH   C   â0rTÿT    A   â #rAFD   A   Ä  r@C@     A  È #rAAT   @  È !rA       ä G‰  P    € ê r@3   ÿàõ ärB+   ÿà÷ ärA(   ÿä ärC    ÿäÿ ä rD!   ÿàõ â �yN@@    ¦ rE?   ÿä â �yOB    è �yQB     ( �yM@0    ( �ySD    ( �yRD     ( �yPDüÿÿ  " xH   ÿàÿ Ê $zIH p  Ž È %xHI°   Ž Ê �yTH     " x@ÿ   N  äO rOÿ   O  ä�xD@ÿÀŽ ä rQÿ   Q  äxAO   ÿ   ä xBMÿÀŽ Ä xCQ   ÿ   äx@ÿ   S  ä xEQ   ÿ   ä xQBAøŽ ä xNNÿÀŽ ä xA@  ÿÀŽ Ä xIRÀÀ  ÿÀŽ ä/@ ÿ       ä xHR??  ÿÀŽ ä xUNCøŽ ä xWDEøŽ â 'HI   @@ Ž â xEPÀÀ  ÿÀŽ Ä xCS  ÿÀŽ ä B ÿ       ä xMÿ   M  ä xDP??  ÿÀŽ ä xMMÿÀŽ â 'DE   @B Ž â x@O   ÿ   Ä xAH?   ÿÀŽ â &rUU.   ÿ   â xMM@øŽ â &rWW2   ÿ   â x@D?   ÿÀŽ â &rQQ0   U   â xHÿ   H  â $rAJA   ÿŽ â xDÿ   D  â &rMM4   W   ä $r@Q@   ÿŽ â ErC A       â $rHGH   ÿŽ Ä $rDMD   ÿŽ â ErA @       â #rC C   ÿ   âErH H       â 0r@ÿT  0 A   âErD D       â #rA A   ÿ   ä #rCFH   C   â0rTÿT    A   â #rAFD   A   Ä  r@C@     A  È #rAAT   @  È !r66A       â !rK       â !rL       â G‰  P    € ê r@)   ÿàõ ärB1   ÿàñ ä rA   ÿä ärC&   ÿä  Æ �yO@     ¢ rH#   ÿàñ Æ �yLB@    è  �yM@    " rI=   ÿä  Æ �yKB0    ( �yQH     ( �yPHüÿÿ  ( �yNHøÿÿ  " xD   ÿàÿ Ê $zED p  Ž È %xDE°   Ž Ê �yRD     " B ÿ       ä rOÿ   O  äOx@LÿÀŽ ä� xAO   ÿ   ä rMÿ   M  äxC@AøŽ Ä x@KÿÀŽ ä xAM   ÿ   ä xLÿ   L  ä xIO   ÿ   ä/ xO@AøŽ ä xLLÿÀŽ Ä x@ÿ   Q  ä xSLIøŽ ä xA@  ÿÀŽ â &r.C.   ÿ   â xIPÀÀ  ÿÀŽ ä @ ÿ       ä xENÀÀ  ÿÀŽ Ä xHP??  ÿÀŽ ä xCQ  ÿÀŽ â 'HI   @@ Ž â xDN??  ÿÀŽ ä xKÿ   K  â 'DE   @B Ž â xLM   ÿ   ä xKKÿÀŽ Ä xCH?   ÿÀŽ â &r.O0   .   â xAD?   ÿÀŽ â &rSS2   ÿ   â xKKLøŽ â $rCJC   ÿŽ â xHÿ   H  â $rA.A   ÿŽ â xDÿ   D  â &rKK4   S   â ErC C       â $rHGH   ÿŽ â ErA A       â $rDKD   ÿŽ Ä #rC C   ÿ   âErH H       â #rA A   ÿ   â ErD D       â 0r ÿR  0 A   ä#rCFH   C   â0rRÿR    A   â #rAFD   A   ä  r C      A  È #r AR      È !r        â x   ÿàÿ ä x€  ÿàó ä |    pbð ä x€  ÿàõ ä x@   ÿàÿ ä rÿ   ÿäÿ  Ä rÿ   ÿä Ê D	       À â Gy  Üÿÿÿÿƒ ê Ay        € ê y      "   " Ey  0    € â r ÿ   pRð är ÿ   pRò Ö G‰  ð     € ê z ÿ `  pPð ä r    ÿ8Ž ä z ÿ a   Sð Æ ˆs  
 þÿ H   ô ˆƒ  	€þÿ H   è ˆƒ   ÿÿ H   è ˆƒ  €ÿÿ H   â G‰  p     € ê ˆs     H   è ˆs  	€þÿ H   è ˆs  €  H   è ˆs   ÿÿ H   è ˆs  6   H   è ˆs  €ÿÿ H   è ˆs  €  H   ä Ay        € ê {          ì M        € ê y       &   "> z ÿ `  pPð Ä x        â y      '   b z ÿ a   Sð â $z   x  Ž È $z  ~   Ž È/ %v  j   Ž È G	  P    € ê „y     H   "z  r  p`ð Æ „y €   H   b r ÿ   pVp  Æ „y    H   ¨ „y     H   è „y €  H   ( „y    H   ( „y    H   ( „y €  H   ( „y    H   ( „y €  H   ( „y €  H   " !r
       Æ „y €  H   " !r	       â/ !r       âO !r         ä� !r       â!r       â !r         ä !r       Æ ‰       b !r       Æ ‰      ¢ !r       Æ ‰	      â !r       È !r       â !r        â/ !r       ÈO ‰      " !r
	       Æ� ‰	      h ‰      ¨ ‰
      â !r       â !r	       â/ !r	       ÈO ‰      " !r
       Æ� ‰ €    h ‰ 	 €    ¨ ‰ €    â !r       â !r       È/ ‰ €    " !r	        ÆO ‰  @    b !r       Æ� ‰ @    ¨ ‰ @    â !r       â !r         È/ ‰ @    " x   ÿàÿ â!r       âO x   ÿàÿ â ‰       b z  r  p`ò â !r       â� z  r  p`ô â ‰	      ¢ x   ÿàÿ ä x    pVò  â‰
      â x    pVt â !r       â !r        È/ ‰      " !‚       â !r	       âO !r	
       Æ� †‰     â z  r  p`ð â !’       â!¢		       ä x    pVp  ä †™    è †© 	   â !r       Ì M	        € ê !r       Ê †y    â My        € ê z   h      È x     pBð Ú G	  @    € ê x     pRð Ú G‰  À    € ê x     pRð Ú G	   $    € ê „y     H   " z  r  p`ð â Ey  p    € ä „y     H   b r ÿ   pVp  Æ „y    H   ¨ „y    H   è „y    H   ( „y  
  H   ( „y €   H   ( „y €  H   ( „y €  H   " !r
       â !r         Æ/ !r       äO „y €  H   â !r         â� !r       Æ!r         ä ‰      " !r	       Æ ‰       b !r       È !r       â !r       ä „y €
  H   â !r        â/ !r       ä ‰      ( ‰      b !r       â !r       È/ ‰
 €    ( ‰ €    h „y €  H   ¢ !r

       â !r       È/ ‰
 @    ( ‰  @    b !r       ÈO !r       â !r
       â !r        È/ ‰       ( ‰      b !r        â !r        â/ G	  �     € ê !r        â!r        Æ  x	;ª¸¿  A  Ì s	 	       $ !t
	  €?    Ì s 
       $  r     A  È  r      A  Ê †y     ä Ay        € ê ‰       b Ey  Ð    € æ ‰	      ¨ „y    H   è „y    H   ( „y    H   è „y    H   â !r         Æ/ „y    H   â !r
	       ÆO ‰	       h ‰
      " !r       â� !r6       â!r	 	       ä/ !r       â !r
       Ä ‰	 €    " !r       Æ ‰ €    h „y
  	  H   ¢ !r	       â !r       È/ ‰ @    ( ‰ @    b !r

       ÈO !r

       â !r       â x   ÿàÿ â !r       Æ/ ‰       " z  r  p`ð Æ ‰	      b x    pVp  â !r        â !r	       Ö/ G	  �     € ê !r       â!r        Æ  x;ª¸¿  A  Ì s        $ !t	  €?    Ì s	 	       $  r	     A  È  r      A  Ê †y    ä Ay        € ê ‰       b Ey       € æ ‰	
      ¨ „y €	  H   è „y €  H   è „y €  H   è „y €  H   â !r         Æ/ „y €  H   â !r
	       ÆO ‰	       h ‰      ( „y
 €  H   ¢ !r	 	       â/ !r       È ‰	 €    " x   ÿàÿ Æ ‰ €    b z  r  p`ð È x    pVp  â !r

       ÈO !r

       â !r	       Æ !r

       â !r       ä/ ‰ @    ( ‰ @    b !r       â !r       È/ ‰       ( ‰      b !r        â !r       â/ G	  �     € ê !r       â!r        Æ  x;ª¸¿  A  Ì s        $ !t	  €?    Ì s	 	       $  r	     A  È  r      A  Ê †y    ä Ay        € ê ‰	
      b !r       â x   ÿàÿ Æ !r       â z  r  p`ð Æ !r       â x    pVp  È ‰       ¢ !r
	       Ê/ ‰	      b !r         ÊO ‰       ¢ !r	       Ê? ‰ €    " !r        ÊO ‰ €    b !r       Ê ‰ @    " !r       Ê/ ‰	 @    b !r       Ê ‰      " !r		       Ê/ ‰ 	      ¢ !r       â M	        € ì !r       â/!r		        ÆO  x;ª¸¿  A  â !r	       Ê s        $ !t  €?    Ì s        $  r     A  È  r     A  Ê †y    â My        € ê „y     H   " z  r  p`ð â Ey       € ä „y     H   b r ÿ   pVp  Æ „y    H   ¨ „y    H   è „y    H   ( „y  
  H   ( „y €   H   ( „y €  H   â !r
       â !r         Æ/ !r       âO !r         Æ� !r       â!r        È ‰       " !r	       Æ ‰      b !r         Æ „y €  H   â !r
       Æ/ ‰       ( ‰
      h „y €  H   ¢ !r        â !r
       È/ ‰ €    ( ‰ €    b !r       Æ „y
 €  H   â !r       ÈO !r       â !r       ä „y €
  H   â !r       Æ/ ‰ @    ( ‰ @    b !r

       â� !r       â !r       È/ ‰       ( ‰      b !r
       â !r        â !r        â/ G	  À     € ê !r         È x	 '7=  A  â x *BL?  A  Æ #t	   €?	   È  r		     A  â  x    ?  A  â !r        È s	 	    $   $ !t	  €?    È  r     A  È  r      A  Ê †y     ä Ay        € ê ‰       b Ey       € æ ‰      ¨ „y    H   è „y    H   ( „y  	  H   â !r         Æ/ „y    H   â !r
       ÆO ‰	       h ‰
      ( „y    H   â !r       â� !r6       â!r	 	       Æ/ !r       â !r
       â ‰	 €    ( ‰ €    b !r       Æ „y
    H   ¢ !r	       â x	   ÿàÿ â !r       Æ/ z 	 r  p`ð â ‰ @    & x    pVp  â ‰ @    b !r

       ÈO !r
       â !r       â !r       È/ ‰       ( ‰      b !r        â !r        â/ G	  À     € ê !r         È x	 '7=  A  â x *BL?  A  Æ #t	   €?	   È  r		     A  â  x    ?  A  â !r        È s	 	    $   $ !t	  €?    È  r     A  È  r      A  Ê †y    ä Ay        € ê ‰       b Ey  Ð    € æ ‰      ¨ „y €  H   è „y €	  H   ( „y €  H   è „y €  H   â !r         Æ/ „y €  H   â !r       ÆO ‰	       h ‰      ( „y €  H   ¢ !r       È� !r       â!r	 	       â/ !r
       È ‰	 €    " x   ÿàÿ Æ ‰
 €    b z  r  p`ð È x    pVp  â !r       âO !r	       â !r
       È/ ‰ @    ( ‰ @    b !r       â !r       È/ ‰       ( ‰      b !r        â !r        â/ G	  À     € ê !r         È x	 '7=  A  â x *BL?  A  Æ #t	   €?	   È  r		     A  â  x    ?  A  â !r        È s	 	    $   $ !t	  €?    È  r     A  È  r      A  Ê †y    ä Ay        € ê ‰      b !r       â x   ÿàÿ Æ !r       â z  r  p`ð Æ !r       â x    pVp  È ‰       ¢ !r       Ê/ ‰	      b !r         ÊO ‰       ¢ !r
	       Ê/ ‰	
 €    b !r        ÊO ‰ €    ¢ !r
	       Ê? ‰ @    " !r       ÊO ‰	 @    b !r       Ê ‰      " !r		       Ê/ ‰ 	      ¢ !r       â M	        € ì !r       â/!r		        ÆO  x'7=  A  â x*BL?  A  â !r	       ä #t  €?   â x   ?  A  Æ  r     A  Ì s     $   $ !t   €?    È  r      A  È  r     A  Ê †y    â My        € ê x     pRð Ú G‰  P    € ê x     pRð Ú G	  `    € ê „y     H   " z  r  p`ð â Ey       € ä „y     H   b r ÿ   pVp  Æ „y    H   ¨ „y    H   è „y    H   ( „y  
  H   ( „y €   H   ( „y €  H   â !r
       â !r         Æ/ !r       âO !r         Æ� !r       â!r        È ‰       " !r	       Æ ‰      b !r         Æ „y €  H   â !r
       Æ/ ‰       ( ‰
      h „y €  H   ¢ !r        â !r
       È/ ‰ €    ( ‰ €    b !r       Æ „y
 €  H   â !r       ÈO !r       â !r       ä „y €
  H   â !r       Æ/ ‰ @    ( ‰ @    b !r

       â� !r       â !r       È/ ‰       ( ‰      b !r
       â !r        â !r        â/ G	  À     € ê !r         Ê	z  i    � â !r        È  x	;ª¸¿  A  â 	z   i    � È 	z   i €  � â s	 	       $ !t
	  €?    Ì s 
       $  r     A  È  r      A  Ê †y     ä Ay        € ê ‰       b Ey       € æ ‰      ¨ „y    H   è „y    H   ( „y  	  H   â !r         Æ/ „y    H   â !r
       ÆO ‰	       h ‰
      ( „y    H   â !r       â� !r6       â!r	 	       Æ/ !r       â !r
       â ‰	 €    ( ‰ €    b !r       Æ „y
    H   ¢ !r	       â x	   ÿàÿ â !r       Æ/ z 	 r  p`ð â ‰ @    & x    pVp  â ‰ @    b !r

       ÈO !r
       â !r       â !r       È/ ‰       ( ‰      b !r        â !r        â/ G	  À     € ê !r         Ê	z  i    � â !r        È  x;ª¸¿  A  â 	z   i    � È 	z   i €  � â s        $ !t	  €?    Ì s	 	       $  r	     A  È  r      A  Ê †y    ä Ay        € ê ‰       b Ey  Ð    € æ ‰      ¨ „y €  H   è „y €	  H   ( „y €  H   è „y €  H   â !r         Æ/ „y €  H   â !r       ÆO ‰	       h ‰      ( „y €  H   ¢ !r       È� !r       â!r	 	       â/ !r
       È ‰	 €    " x   ÿàÿ Æ ‰
 €    b z  r  p`ð È x    pVp  â !r       âO !r	       â !r
       È/ ‰ @    ( ‰ @    b !r       â !r       È/ ‰       ( ‰      b !r        â !r        â/ G	  À     € ê !r         Ê	z  i    � â !r        È  x;ª¸¿  A  â 	z   i    � È 	z   i €  � â s        $ !t	  €?    Ì s	 	       $  r	     A  È  r      A  Ê †y    ä Ay        € ê ‰      b !r       â x   ÿàÿ Æ !r       â z  r  p`ð Æ !r       â x    pVp  È ‰       ¢ !r       Ê/ ‰	      b !r         ÊO ‰       ¢ !r
	       Ê/ ‰	
 €    b !r        ÊO ‰ €    ¢ !r
	       Ê? ‰ @    " !r       ÊO ‰	 @    b !r       Ê ‰      " !r		       Ê/ ‰ 	      ¢ !r       â M	        € ì !r       â/!r		        ÈO 	z i    � â !r	       È  x;ª¸¿  A  â 	z i    � È 	z i €  � â s        $ !t  €?    Ì s        $  r     A  È  r     A  Ê †y    â My        € ê „y     H   " z  r  p`ð Æ „y     H   b r ÿ   pVp  Æ „y €   H   ¨ „y €  H   è „y!    H   ( „y#    H   ( „y    H   ( „y    H   ( „y €  H   ( „y €  H   ( „y    H   " !r
       Æ „y  	  H   " !r         Æ/ „y    H   b !r	       ÆO „y  
  H   ¢ !r       Æ� „y €  H   â !r!!       Æ„y €
  H   " !r##6       Æ „y%    H   " !r       Æ „y'    H   " !r         Æ „y €  H   " !r       Æ „y €  H   " !r       Æ „y) €  H   " !r!       Æ „y+ €	  H   â !r#       Æ „y- €  H   â !r       Æ/ „y/ €  H   â !r        âO !r       ä� ‰       " !r
       Æ‰      b !r%%       Æ ‰      ¢ !r'       Æ ‰	
      â !r       Æ ‰%      " !r       Æ ‰      " !r)       â !r         â !r+       â !r       Æ/ ‰       " !r--       â !r       âO x   ÿàÿ â !r//       ä ‰-      b !r

	       â� z  r  p`ô ä ‰/      ¢ !r%       âx    pVt â !r       â ‰	      â x   ÿàÿ Ä x   ÿàÿ â ‰      " z  r  p`ò ä z  r  p`ö â ‰
      " !r        â x    pVò  ä‰      " x    pVö â !r-       Ä/ ‰      b !r/       ÆO ‰      ¨ ‰      ¢ !r		       â� !r       ä‰  €    â !r
       Æ ‰	 €    " !r       Æ ‰ €    " !r       Æ/ ‰
 €    b !r       ÆO ‰ €    ¢ !r       Æ ‰ €    ¨ ‰ €    ¢ !r         Æ� ‰ €    â !r	       â!r       ä ‰  @    " !r

       Æ/ ‰	 @    b !r       ÆO ‰ @    ¢ !r       Æ ‰
 @    " !r       Æ ‰ @    " !r       Æ� ‰ @    è ‰ @    â !r        Æ ‰ @    " !r		       â/ !r       äO ‰       b !r
       Æ‰	      ¢ !r       Æ ‰      " !r       Æ� ‰
      â !r       Æ ‰      â !r       Æ ‰      ( ‰      " !r        Æ/ ‰      b !r	       âO !‚        â!r       ä!‚       â!r

       â� !’       ä  ‚      A  â !r       â !’
       Æ †‰     â !r       â !¢       â  ’     A  ä !¢	       â !r       ä †™    â  ¢		     A  â !r       È/ †© 	   â M9        € ê !r       â !r       È  r     A  Ê †y    â My        € ê „y     H   " z  r  p`ð â Ey  À    € ä „y     H   b r ÿ   pVp  Æ „y    H   ¨ „y    H   è „y    H   ( „y  
  H   ( „y €   H   ( „y €  H   è „y €
  H   â !r
       â !r         Æ/ !r       âO !r         Æ� !r       â!r        È ‰       " !r	       Æ ‰      b !r         Æ „y €  H   â !r
       Æ/ ‰       ( ‰
      h „y €  H   ¢ !r        â !r
       È/ ‰ €    ( ‰ €    b !r       Æ „y
 €  H   â !r       ÈO !r       â !r       â !r       È/ ‰ @    ( ‰ @    b !r

       È� !r
       â !r       â !r       È/ ‰       ( ‰      b !r        â !r        â/ G	  à     € ê !r         Ê	x   à@  � Ê  x #ÛÙ?  A  È  x	 ;ª¸¿  A  â !r        Ê s	 	       " 	x    à@  � È 	x    àÀ  � Ê !t   €?    â !t
	  €?    È s 
       $  r     A  È  r     A  Ê †y     ä Ay        € ê ‰       b Ey       € æ ‰      ¨ „y    H   è „y    H   ( „y  	  H   è „y    H   â !r         Æ/ „y    H   â !r
       ÆO ‰	       h ‰
      " !r       â� !r6       â!r	 	       Æ/ !r       â !r
       â ‰	 €    ( ‰ €    b !r       Æ „y
    H   ¢ !r	       â x	   ÿàÿ â !r       Æ/ z 	 r  p`ð â ‰ @    & x    pVp  â ‰ @    b !r

       ÈO !r
       â !r       â !r       È/ ‰       ( ‰      b !r        â !r        â/ G	  à     € ê !r         Ê	x   à@  � Ê  x #ÛÙ?  A  È  x ;ª¸¿  A  â !r        Ê s        " 	x    à@  � È 	x    àÀ  � Ê !t   €?    â !t	  €?    Ì s	 	       $  r	     A  È  r     A  Ê †y    ä Ay        € ê ‰       b Ey  ð    € æ ‰      ¨ „y €  H   è „y €	  H   ( „y €  H   è „y €  H   â !r         Æ/ „y €  H   â !r       ÆO ‰	       h ‰      ( „y €  H   ¢ !r       È� !r       â!r	 	       â/ !r
       È ‰	 €    " x   ÿàÿ Æ ‰
 €    b z  r  p`ð È x    pVp  â !r       âO !r	       â !r
       È/ ‰ @    ( ‰ @    b !r       â !r       È/ ‰       ( ‰      b !r        â !r        â/ G	  à     € ê !r         Ê	x   à@  � Ê  x #ÛÙ?  A  È  x ;ª¸¿  A  â !r        Ê s        " 	x    à@  � È 	x    àÀ  � Ê !t   €?    â !t	  €?    Ì s	 	       $  r	     A  È  r     A  Ê †y    ä Ay        € ê ‰      b !r       â x   ÿàÿ Æ !r       â z  r  p`ð Æ !r       â x    pVp  È ‰       ¢ !r       Ê/ ‰	      b !r         ÊO ‰       ¢ !r
	       Ê/ ‰	
 €    b !r        ÊO ‰ €    ¢ !r
	       Ê? ‰ @    " !r       ÊO ‰	 @    b !r       Ê ‰      " !r		       Ê/ ‰ 	      ¢ !r       â M	        € ì !r       â/!r		        ÈO 	x  à@  � â !r	       È  x#ÛÙ?  A  â 	x  à@  � Æ  x;ª¸¿  A  â 	x  àÀ  � Ê s        " !t  €?    â !t  €?    Ì s         $  r      A  È  r     A  Ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y       "   $ z   w  p`ð Ú M	        € ê y      !   " ¹z  l      â Ey  p    € â ™x?    â y      %   b ¹z  F   
   â $rÿÿ   ÿ Ž ä $rÿÿ   ÿ Ž â x"ÿ     È | "   p`ð â $x   ÿ Ž Ø/ G	  Ð    € ê y      &   " x        â $z
  x  Ž È %v

 \   Ž Ê �y
    ™ ¢  $rÿÿ   ÿ Ž â xÿ     â $vÿ X  ÿ Ž â x   ÿàÿ â 'z m   Ž â x   ÿÀŽ È zÿ n    â $tÿ$   ÿ Ž â xÿ     Æ $r	ÿÿ   
Ž ä $z  r  ÿŽ ä $z		 o  Ž ä $x   ÿ Ž ä $z		 u  ÿŽ Æ xøÿÿøŽ È r	   à‘ È r	ÿÿ   ÿ$~  â %v
 Z   Ž È $x	$   ÿŽ â x	   ÿ   â $z q  "Ž â x
8   ÿàñ ä x		   ÿÀŽ Æ $x     ä $z
 t  "Ž äO$z t  Ž ä $z
 q  
Ž ä %x°   ÿŽ â r ÿ       Æ $x    	Ž â x   øŽ â%x
°   ÿŽ â r	   ÿüŽ är   ÿàñ â$vÿ Y  ÿ Ž â r   ÿàó äx   øŽ â $rÿÿ     â x   ÿüŽ â $r
ÿÿ   Ž  â r	   ÿüŽ Ä x   ÿüŽ ä r.   ÿàñ â�y2ìÿÿ • ¢  r$   ÿàó Æ $x/     â r%   ÿäÿ  ä r   ÿàñ ä �y+.@   ™ è �y,$    ™ b	 r	   ÿàó Æ �y*$   ™ ¢ $x     Æ �y).0   ™ ¢ $rÿÿ   Ž  â r    ÿàó ä�y#øÿÿ • ¨  �y'    • ¢  $x!   Ž  â r   ÿàñ Ä �y&Üÿÿ ™ ¨  �y(    ™ ¢  r
   ÿä  Æ �yÌÿÿ ™ ¨  �y1üÿÿ • ¨  �y.     ™ ¨" �y-ðÿÿ ™ ¨  �y0@   ™ ¢  r$   ÿàñ Æ�y/Èÿÿ • ¤ $x%     ä �y     ™ ¨" �y5$   • ¨  �y30   ™ ¨ �y!$    • ¨" �y4$üÿÿ • ¢ r   ÿàñ Ê $x     â r   ÿàñ È �y    ™ "$x     Ê �y    ™ " x    p@ð ä x""   ÿàÿ ä x@  ÿàõ ä x`  ÿàó Æ $rÿÿ    â rÿ   ÿäÿ  ä xÿ   +  ä� r,ÿ   ,  ä
r*ÿ   *  äO x%ÿÀŽ â/ $x,   ÿ Ž â x)ÿÀŽ Ä x$*   ÿ   â $x,,   ÿ Ž â x++ÿÀŽ ä x)ÿ   )  ä x$$øŽ â $x**   ÿ Ž â x+øŽ ä x#??  ÿÀŽ Ä x'  ÿÀŽ ä x##ÀÀ  ÿÀŽ ä  ÿ       ä x%%,øŽ ä x))ÿÀŽ â &r&   ÿ   ä '#   @ Ž â x*)*øŽ Ä xÿ   '  â &r%%(   ÿ   â r.ÿ   .  â &r$$      â x1??  ÿÀŽ â&r#*-   %   â x11ÀÀ  ÿÀŽ ä x  ÿÀŽ â $ÿÿ   ÿ Ž â x%?   ÿÀŽ â $x,.   ÿ Ž â x'0ÿÀŽ â '1   @ Ž È $r$%   ÿŽ ä &z$&  € ÿ   â x',øŽ â &z%(  € ÿ   â x*ÿ     â &z  € $   â x,?   ÿÀŽ ä x$ÿ     â &z-  € %   Ä $r,,   ÿŽ ä $r##*   ÿŽ ä $r'$   ÿŽ â Er$        â 0r/ÿ/    A   â Er ,       ä r ÿ      â 0r2ÿ2    A   â Er) #       â #r#/$   ÿ   âEr, '       â #r'/   ÿ   â $x+    ÿ Ž â xÿ   5  â #r#2)   #   âx*3ÿÀŽ â &r)&   ÿ   â x%ÿ   0  â #r$2,   '   â x  ÿÀŽ â $ÿÿ   ÿ Ž â x1!ÀÀ  ÿÀŽ â $x,    ÿ Ž â x**+øŽ Ä x'!??  ÿÀŽ â $ ÿÿ   ÿ Ž â x+4ÀÀ  ÿÀŽ ä x!5  ÿÀŽ â ''1   @ Ž â x%%ÿÀŽ ä x..   ÿ   ä x&4??  ÿÀŽ Ä x3ÿ   3  â '&+   @  Ž â x%%.øŽ ä x33ÿÀŽ ä x '?   ÿÀŽ â &r%%(   ÿ   â x&?   ÿÀŽ â &r)*   )   â x,3,øŽ â $r     ÿŽ â x(ÿ   '  â $r)   ÿŽ â x&ÿ   &  â &r%,-   %   â Er          â 0rÿ  0 A   Ä$r((   ÿŽ â Er        â $r%%&   ÿŽ â  r$$     A  â #r/    ÿ   âEr( (       â #r//   ÿ   â | "   p`ð â 0rÿ  0 A   â Er  %       â #r2(      â 0rÿ    A   Ä 0rÿ    A   â #r 2    /   â  r     A  â #r##   $  Æ #r      â !r#       Æ !r       â G‰   öÿÿÿÿƒ ê Ay        € ê ‰      " $x   Ž Æ ‰      d z  p  p`ð È x    pDp  â !r       â !r	       È/ ‰      ( ‰
	      b !r       â !r
	
       È/ ‰ €    ( ‰
 €    b !r       â !r
       È/ ‰ @    ( ‰ @    b !r       â !r       È/ ‰      h  ‰      ¢  M	        € ê y      &   â $x   ÿ Ž â !r       â/ $z   s  Ž â x	        â !r       â_ r ÿ   p"ð Ä x    p"ò Ö $ÿÿ    Ž ä $ÿÿ    Ž ä $z  v   Ž È� %v  j  	 Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      "   $ z  w  p`ð Ú M	        € ê y      !   ( y	      &   b ¹z  l      â $tÿ   ÿ Ž â ™x?    â Ey  Ð    € â ¹z  F   
   â x ÿ     È |     p`ð â $z x  	Ž È/ %v \   Ž â r ÿ       â x      ÿ  ä $rÿÿ   ÿ Ž â �y    ™ b $rÿÿ   ÿ Ž ä $rÿÿ   ÿ Ž â G	       € ê $rÿÿ   ÿ Ž â y      %   " x   ÿÀŽ â Ey      € â 'z	 m   Ž Ä xÿ     Æ zÿ n    ä $x   ÿ Ž ä rÿ   ÿáÿ Ê $z
 o  	Ž â xÿ     â $z	 r  ÿŽ ä $z

 u  ÿŽ â xøÿÿøŽ â $tÿ$   ÿ Ž Æ r

   	à‘ â $r	ÿÿ     Ž â x    ÿ   ä rÿÿ   ÿ$~  â %v

 Z   Ž â x    ÿàÿ Æ $x   ÿ Ž â x

8   ÿàñ â $x$   ÿŽ â z  Y      â $z t  	Ž âx   ÿÀŽ â $z  q  	Ž ä $z q  Ž ä $z t  Ž Ä $x    Ž ä %x°   ÿŽ â r ÿ       Æ %x°   ÿŽ â x   øŽ Æ $x     â r   ÿàñ â$vÿ X  ÿ Ž â r   ÿàó â $vÿ `  ÿ Ž â x   øŽ â $vÿ a  ÿ Ž â r   ÿüŽ â$rÿÿ     â r   ÿüŽ â $rÿÿ   Ž  â x   ÿüŽ Ä x   ÿüŽ ä r   ÿàó â�y.
Èÿÿ • ¢ r    ÿàñ Æ $x   Ž  â r"   ÿàó â �y1
ìÿÿ • â r!   ÿä  Æ �y/    ™ " $x#   Ž  â r(
   ÿàñ ä �y* @   ™ h �y'   ™ ¢  $r)ÿÿ     Æ �y-"øÿÿ • è  �y+"    • è  �y  0   ™ è �y(Üÿÿ ™ è  �y$(Ìÿÿ ™ è  �y%(    ™ è  �y&(ðÿÿ ™ è  �y,"üÿÿ • â  r    ÿàó Ä/ r   ÿàñ ä r!   ÿäÿ  Æ $x     â r"   ÿàñ â �y7     ™ è
 �y2@   ™ â  $x#     Æ �y3    ™ è
 �y00   ™ è  �y5"   • è �y4"üÿÿ • è �y6"    • â r(   ÿàñ Ê $x)     â r   ÿàñ È �y((    ™ â  r   ÿä  Ê �y8    ™ â  x    p@ò ä z ÿ `  pPð È z ÿ a   Sð ä r/ÿ   /  äx *ÿÀŽ Æ$x!/   ÿ Ž â r'ÿ   '  ÈO x# !øŽ ä/ x"-??  ÿÀŽ â� $x'   ÿ Ž â x--ÀÀ  ÿÀŽ â $ ÿÿ   ÿ Ž â x!+  ÿÀŽ ä x*ÿ   *  ä x ÿÀŽ Ä x ÿ      â '"-   @  Ž â x**ÿÀŽ ä x//   ÿ   â $x''   ÿ Ž â xøŽ â &r##   ÿ   â x ÿÀŽ ä x**/øŽ â &r$   #   â x "?   ÿÀŽ Ä x''øŽ â &r**%   ÿ   â xÿ   +  â $r#    ÿŽ â x"ÿ   "  â &r''&   *   â x  ÿÀŽ ä x!,ÀÀ  ÿÀŽ ä  ÿ       Ä x ,??  ÿÀŽ â $r"'"   ÿŽ â Er# #       â 0r ÿ.    A   ä ' !   @ Ž â Er "       Æ &z)  € ÿ   â #r" #   ÿ   â 0r1ÿ1    A   â x ?   ÿÀŽ â &z'%  € ÿ   ä &z)$  € )   â #r"1   "   â x ÿ      â &z'&  € '   ä $r)   ÿŽ Ä $r '    ÿŽ â r7ÿ   7  äEr        ä x2ÿÀŽ ä r3ÿ   3  ä Er#         â #r     ÿ   â x2ÿ   2  â $x7   ÿ Ž âx!7   ÿ   â $x*3   ÿ Ž â x22ÿÀŽ â #r#1#       â x+0ÿÀŽ Ä x ÿ   5  ä x7øŽ â $ÿÿ   ÿ Ž â x22!øŽ ä x5  ÿÀŽ ä x-4ÀÀ  ÿÀŽ ä x++*øŽ Ä x!   ÿÀŽ ä x0ÿ   0  ä x*4??  ÿÀŽ ä x/6ÀÀ  ÿÀŽ ä  ÿ       â '*-   @ Ž â x,6??  ÿÀŽ Ä x00ÿÀŽ â $x33   ÿ Ž ä ',/   @  Ž â x*?   ÿÀŽ Æ &r77   ÿ   â x303øŽ â &r22%   ÿ   â x ,?   ÿÀŽ â &r++$   7   â x!ÿ   *  â &r23&   2   â x,ÿ   ,  â $r++   ÿŽ ä $r)    ÿŽ Ä 0rÿ(  0 A   â Er+ +       â $r!2!   ÿŽ â Er        â $r,',   ÿŽ â  r##     A  â #r +   ÿ   âEr! !       â #r     ÿ   â Er+ ,       Æ #r1!      â0r!ÿ8  0 A   â #r 1+       â 0r-ÿ(    A   ä 0rÿ8    A   â  r! !     A  ä #r-"-   #  ä #r(   !  â G‰  @    € ê r    ÿà÷ är   ÿàõ Æ $x!   Ž ä $x    â r"   ÿàõ ä �y7      ¤ r#   ÿä ä �y5     è �y3@    ( �y.0    h �y,"     h �y/"üÿÿ  h �y0"øÿÿ  b r*   ÿàõ Ê $x+    Ê �y**     b r7ÿ   7  äOr2ÿ   5  ä� x7   ÿ   ä x ÿ   3  äx33ÿÀŽ â $x2   ÿ Ž â x4.ÿÀŽ Äx53øŽ ä xÿ   ,  â $x77   ÿ Ž â x#4øŽ ä/ x  ÿÀŽ ä x  ÿÀŽ â $ÿÿ   ÿ Ž â x3/ÀÀ  ÿÀŽ Ä x"/??  ÿÀŽ ä x4 7øŽ â '"3   @ Ž â x/0ÀÀ  ÿÀŽ ä x!,  ÿÀŽ â $ ÿÿ   ÿ Ž â x.ÿ   .  ä x+0??  ÿÀŽ â &r55   ÿ   â x..ÿÀŽ â '+/   @  Ž â x2   ÿ   Ä x "?   ÿÀŽ â &r44%   ÿ   â x.øŽ â &r##$   5   â x+?   ÿÀŽ â $r )    ÿŽ â x,ÿ   "  â &r&   4   â x"ÿ   +  â $r#   ÿŽ â Er!         â $r,',   ÿŽ Ä $r""   ÿŽ â Er        â #r  !   ÿ   âEr, ,       â 0r#ÿ*  0 A   âEr! "       â #r    ÿ   ä #r 1,       â0rÿ*    A   â #r1!      Ä  r# #     A  È #r   #  È !r       â !r-       â !r(       â G‰  @    € ê r   ÿàõ är    ÿàñ Æ $x    â r!   ÿä  È �y5     ¢ r*   ÿàñ Æ �y/ @    è �y3    " $x+     Æ �y. 0    h �y(*    h �y,*     h �y-*üÿÿ  b r"   ÿàñ Ê $x#     Ê �y""     b r5ÿ   5  äOx/ÿÀŽ ä� x 5   ÿ   ä rÿ   3  äx! øŽ ä x/ÿ   /  â $x    ÿ Ž â x.ÿÀŽ â$x*5   ÿ Ž â/ x//ÿÀŽ Ä x# øŽ ä xÿ   (  ä x0/*øŽ ä x ÿ   .  ä x  ÿÀŽ ä x+,ÀÀ  ÿÀŽ Ä x/   ÿ   â $ÿÿ   ÿ Ž â x.,??  ÿÀŽ â &r*!   ÿ   â x-??  ÿÀŽ ä x!(  ÿÀŽ â '.+   @ Ž â x--ÀÀ  ÿÀŽ ä x( ÿÀŽ â $ ÿÿ   ÿ Ž È '-   @  Ž â x .?   ÿÀŽ ä x/(/øŽ â &r00%   ÿ   â x?   ÿÀŽ â &r##$   *   â x.ÿ   .  â $r )    ÿŽ â xÿ     â &r//&   0   Ä $r#   ÿŽ â Er!         â $r.'.   ÿŽ ä $r//   ÿŽ â Er        â #r !   ÿ   âEr. .       â 0r!ÿ"  0 A   â Er/ /       â #r     ÿ   Ä #r1.      â0rÿ"    A   â #r 1/       ä  r!!     A  È #r     !  È !r        â x		   ÿàÿ ä x

@  ÿàõ ä | 	   p`ð ä x`  ÿà÷ â $rÿÿ    â x`  ÿàó Æ $rÿÿ   Ž â rÿ   ÿäÿ  Ê D	       À â Gy  0íÿÿÿÿƒ ê Ay       € ê r        â $rÿÿ    Ž ä Ay        € ê z ÿ `  pPð È z ÿ a   Sð Ú G	  P    € ê ‰      ( ‰       b !r       â !r        È/ ‰       ( ‰      b !r         â !r	       È/ ‰  €    ( ‰	 €    b !r        â !r	       È/ ‰ @    ( ‰ @    b !r       â !r       È/ ‰      ( ‰      b !r       â !r       â/ Gy  €    € ê ‰      ( ‰	      h ‰       ¨ ‰      â !r       â !r		       â/ !r
        ÆO ‰       " !r       Æ� ‰	      h ‰
      ¨ ‰      â !r         â !r	       È/ ‰  €    " !r
       ÆO ‰ €    b !r       Æ� ‰ €    ¨ ‰ €    â !r        â !r       â/ !r	       ÈO ‰ @    " !r       Æ� ‰ @    h ‰
	 @    ¨ ‰ @    â !r        â !r       â/ !r
	
       ÈO ‰      " !r       Æ� ‰
      h ‰       ¨ ‰      â !r       â !r
       â/ !r        âO !r       â� y      %   ( y      !   $ $x   Ž Ê z  p  p`ð È x    pDp  Ú M	        € ê z ÿ ^  pPð â $z v  Ž Æz ÿ _   Sð Ú         Ê % ^   Ž Ì �	     ¢ $x    ÿ Ž Ê r  ÿ   p"ò äx     p"ö äx     p"ø äx     p"ô Î $ ÿÿ    Ž â z ÿ `  pPò ä 2         â $B ÿÿ    Ž â z ÿ a  Sò â $" ÿÿ    Ž È !         ÐO G™       € ê z ÿ b  pPð È z ÿ c   Sð Ú         Ê % b   Ž Ì �	     ¢ $x   ÿ Ž ä $vÿ h  ÿ Ž Æ x øÿÿÿp"ò äx üÿÿÿp"ô är ÿ   p"ö äx    p"ø Î $ÿÿ    Ž â x    pBò ä "        â $2ÿÿ    Ž ä B        Ê !       ÆO G  P    € ê x    pRð Ú G‰  �     € ê x    pRð Ú G	  ð    € ê  x;ª¸¿  A  Ì s        $ !t  €?    Ì s        $  r     A  È  r       A  â Gy  P    € ê  x'7=  A  â x*BL?  A  Æ #t  €?   â  x   ?  A  Æ  r     A  Ì s     $   $ !t  €?    È  r     A  È  r       A  â Gy  °    € ê x    pRð Ú G‰  à     € ê x    pRð Ú G	        € ê 	z i    � ä 	z   i    � Æ  x;ª¸¿  A  â 	z   i €  � Ê s        $ !t  €?    Ì s        $  r     A  È  r       A  â Gy  Ð     € ê  r       A  â Gy  °     € ê 	x  à@  � ä 	x    à@  � Æ  x#ÛÙ?  A  â 	x    àÀ  � Æ  x;ª¸¿  A  ä !t   €?    È s        $ !t  €?    Ì s        $  r      A  È  r       A  ä y      "   " $tÿ   ÿ Ž Æ y      &   b $z s  Ž È $z v  Ž È/ %v j   Ž Ê †y      â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      "   " ¹z  l      â Ey   '    € â ™x?    â y      !   b ¹z  F   
   â r ÿ       â x      ÿ  â y#      %   ¢ r3 ÿ       â x      ÿ  â x      ÿ  â r5 ÿ       â x       ÿ  â r7 ÿ       â x      ÿ  â x      ÿ  â r ÿ   pRð Ä r ÿ   pRò är"   ÿ(Ž È/ xÿ   "  ä x##   ÿ   äO r "   ÿ � È xÿ      È |    pbô Ú G)  €%    € ê y      '   " x ðÿÿÿÿÀŽ ä r ÿ       â y      &   b r""  €ÿàÿ ä r ÿ       ä x	ÿ   "  ä x#   ÿàÿ Ä r		"   ÿ� ä x$   ÿ   ä z)  X      ä z+  Y      ä r ÿ       ä z($ q  ÿàÿ â 'z  y   Ž È 'z s   Ž â/ zÿ z     Æ $z # p  Ž â zÿ t    â $z% w  ÿŽ â z  q      â $z |   Ž âx ÿ   	  â $z |  Ž â x		üÿÿÿÿÀŽ â $z v  Ž âr2$   ÿŽ â$z v  Ž â r	"	  €ÿàÿ â $x&   $Ž âr*$   ÿŽ â$z p  Ž â xÿ   	  â $x,   $Ž âx/	   ÿ   â $x.   $Ž â x-	     â $x0   $Ž â r 	   ÿŽ â$z }  ÿŽ â x    ÿ   â %x
�   ÿŽ â x   ÿ   Ä r%%   àµ â%x�   ÿŽ â x1ÿ     äx8
   øŽ ä xÿ     ä r4
   ÿàù är6   ÿàû Ä x'   øŽ ä r1ÿÿ   1d~ ä x88   ÿüŽ ä r:   ÿä är<	   ÿäÿ ä x''   ÿüŽ Ä r)8   ÿàõ È r+   ÿä Ê �yZøÿÿ • ¢  r)'   ÿàõ ä xd $       â �yR    • â  r	+   ÿä ä r$%   ÿàõ â �ySüÿÿ • " r)4   ÿà÷ är ÿ1   ÿä â �yQ   • b r)6   ÿàù Ä r+:   ÿäÿ â �yTüÿÿ • " %v Z  d Ž â r+<   ÿä ä �yO    • " $x $   ÿŽ Æ �y    ™ $ r   ÿàÿ ä �y@    ™ " r/   ÿàõ Æ/ �yA   ™ " r-   ÿä Æ �yB    ™ " r        Æ �yC   ™ " r(%   ÿàõ Æ �yD(   ™ ( �yE   ™ ( �yG    • " rÿ1   ÿä Æ �y?$   • ( �yF8   ™ " %v Z  d Ž È $x$   ÿŽ â r/   ÿàõ È r   ÿàÿ È r-   ÿä â �yH    • ( �y=   ™ ( �y;(   ™ ( �y9   ™ " r2%   ÿàù Æ/ �y>8   ™ " r
)   ÿàõ ä rÿ1   ÿä â �yJ$   • " r)   ÿà÷ ä r+   ÿä â%v Z  d Ž â/ r	+   ÿäÿ Æ �yI    ™ " $x$   ÿŽ â r/   ÿàõ ä �yL    ™ $ r   ÿàÿ È r-   ÿä â �yU    • " r&%   ÿàõ Æ �yN   ™ " rÿ1   ÿä Æ/ �yM(   ™ ( �yK   ™ " %v Z  d Ž Æ �yP8   ™ " $x$   ÿŽ â r/   ÿàõ ä �yX$   • $ r   ÿàÿ È r-   ÿä â �y\    • ( �yV   ™ ( �yW(   ™ ( �yY   ™ " r*%   ÿàõ Æ/ �y[8   ™ " rÿ1   ÿä â %v Z  d Ž ä �y`$   • $ $x$   ÿŽ â r/   ÿàõ È r   ÿàÿ È r-   ÿä â �yb    • " r,%   ÿàõ Æ �y_   ™ " rÿ1   ÿä â/ %v Z  d Ž ä �y^(   ™ $ $x$   ÿŽ ä �y]   ™ " r/   ÿàõ Æ �ya8   ™ " r   ÿàÿ Æ �ye$   • " r-   ÿä ä r0%   ÿà÷ â�yi    • ( �yc   ™ ( �yg(   ™ ( �yf   ™ " r.%   ÿàõ Æ �yh8   ™ " rÿ1   ÿä ärÿ1   ÿäÿ â �yj$   • " %v Z  d Ž È/ $x$   ÿŽ ä %v Z  d Ž È $x$   ÿŽ â r   ÿàÿ ä r/   ÿàõ är   ÿàÿ ä r-   ÿä ä r        â �y    • ( �yl   ™ ( �yk(   ™ ( �yd   ™ ( �ym8   ™ " r/   ÿà÷ Æ �y$   • " r-   ÿäÿ Æ �y    • ( �yn(   ™ ( �yo   ™ ( �yp8   ™ ( �y$   • " xqZ??  ÿÀŽ ÄO xsZÀÀ  ÿÀŽ ä �yZ   ™ ¢ x "   pBô ä xR  ÿÀŽ ä� xÿ   R  ä xÿ   Q  Î" ÿ       ä x  ÿÀŽ ä x  ÿÀŽ â/ '(qs   @ Ž â xSÀÀ  ÿÀŽ äxsTÀÀ  ÿÀŽ ä xQ  ÿÀŽ Ä " ÿ       ä " ÿ       ä xrS??  ÿÀŽ ä xST??  ÿÀŽ ä xRO??  ÿÀŽ â '(r   @ Ž â xOOÀÀ  ÿÀŽ Ä " ÿ       â '(Ss   @ Ž â x ÿÀŽ ä x@ÿÀŽ ä x ÿ      ä x@ÿ   @  â '(RO   @ Ž â xÿ   A  Ä xAÿÀŽ ä xBÿÀŽ ä xÿ   B  ä xB ÿÀŽ ä xA@ÿÀŽ ä x@ÿÀŽ â &z C  € ÿ   â xÿÀŽ â &rTC   ÿ   Ä &rCC   ÿ   ä &rtBD   ÿ   ä&zD  € ÿ   ä &rD@D   ÿ   ä &zE  €     ä 0r ÿG    A   Ä &rTE   T   ä&rvE   C   â xES?   ÿÀŽ â 0rGÿ?    A   â x?q?   ÿÀŽ â &rtAF   t   âxCr?   ÿÀŽ â &zF  €    ä &rOF   D   â xFR?   ÿÀŽ â $rTT?   ÿŽ â xqÿ   q  â $rvvE   ÿŽ â xDÿ   S  â $rC   ÿŽ â xrÿ   r  Ä xRÿ   R  â $rQF   ÿŽ â Er        â $rttq   ÿŽ ä $rsOD   ÿŽ â ErO T       â $rSr   ÿŽ ä$ruR   ÿŽ â Er v       Ä Er Q       â #rT O   ÿ   âErQ t       â #rt    ÿ   âErs s       â #rv    ÿ   âEru u       â #r     ÿ   ä #rOGs   v   â &zz=  € ÿ   â #rvGu       â &r =   ÿ   Ä &r==   ÿ   ä &rB;   ÿ   ä&z;  € ÿ   ä &r;@;   ÿ   ä &zz9  € z   ä &rx9       Ä &r=9   =   â ErS S       â &r|A>      ä&r;>   ;   ä &z>  €    ä $rCz   ÿŽ ä $rx?x   ÿŽ Ä $r=E=   ÿŽ ä $rFz   ÿŽ â #rTGQ   T   â$r>D;   ÿŽ â #rtGS   t   â $r;r   ÿŽ â Er        â 0r ÿH    A   â Er9 x       â $r|q|   ÿŽ â Er        â $rGR   ÿŽ â Er =       â 0rJÿJ    A   â ErH >       â #r    ÿ   âEr> ;       â #r9 9   ÿ   âEr| |       â #r    ÿ   â#r    ÿ   â Er  G       â 0r;ÿI  0 A   â#r>J>      â#r|J|   9   â0r9ÿL  0 A   â#r J       â 0rIÿI    A   â  rt;     A  â  r>;>     A  â 0rLÿL    A   â #rJH      â  r9      A  â #rTTI     â #rI|   >  â &r N   ÿ   â#rL     â &zHN  € ÿ   â !r       â &rNN   ÿ   Ä &rBM   ÿ   ä &r>K       â!r       â &zM  € ÿ   ä &zHK  € H   ä &rNK   N   ä &rJAP      Ä $r>?>   ÿŽ ä &rM@M   ÿ   ä &zP  €    ä $rNEN   ÿŽ ä $rCH   ÿŽ ä$rFH   ÿŽ â Er= >       â $rJqJ   ÿŽ Ä 0r ÿU    A   â Er        â &rMP   M   â Er        â $rGr   ÿŽ ä$rKR   ÿŽ â Er N       â 0rXÿX    A   â ErJ J       â #r= =   ÿ   â $rMDM   ÿŽ â Er> G       â #r    ÿ   â#r    ÿ   â#r    ÿ   â #r XJ   =   â &zJV  € ÿ   â ErH M       â #r>X>      â &rV   ÿ   â ErN K       â &zPW  € ÿ   Ä &zJY  € J   â #r=XH      â &rHY      ä &rVV   ÿ   ä &z[  € P   ä $rCJ   ÿŽ â #rXXN      â !rT       â &rNBW   ÿ   Ä $rFJ   ÿŽ â Er        â &rT@W   ÿ   ä &rVY   V   ä $rJr   ÿŽ ä 0r\ÿ\    A   ä &rNA[   N   Ä $rH?H   ÿŽ â ErJ J       â &r[[   T   ä $rVEV   ÿŽ â #r\   ÿ   â 0r`ÿ`    A   â ErG H       â $rNqN   ÿŽ ä $r[D[   ÿŽ Ä $rKR   ÿŽ â Er V       â #rJ`J      â  r>;>     A  â  r9X     A  â ErN N       â #rG\G   ÿ   âErH [       â #r\   ÿ   â #r I    >  â #r>L=     â #rN`N   G   âEr        â #rG`H      â  r;J     A  â !r7 7       â !r!>!       â &r _   ÿ   Ä &z>_  € ÿ   â ErP K       â &r__   ÿ   â #r\   ÿ   â #rIN     â &rHB^   ÿ   ä&zJ^  € ÿ   ä &r^@^   ÿ   ä &rN]   _   â #rP`P      â &z>]  € >   â !r         â &rHAa   H   Ä &za  € J   ä &r ]       ä &raa   ^   ä $rNEN   ÿŽ â  rP9P     A  â $rC>   ÿŽ ä$rF>   ÿŽ Ä $r ?    ÿŽ ä $raDa   ÿŽ ä $r>r   ÿŽ ä$rJR   ÿŽ â Er N       â 0rbÿb    A   â #rPLG   P  â Er        Ä Er        â 0reÿe    A   â Er=         â !rP       â Era a       â #r b   ÿ   âErG >       â $rHqH   ÿŽ â #rNb   ÿ   â#rPb   ÿ   â &rc   ÿ   â ErK J       â &zc  € ÿ   â #rJb=   ÿ   â #rea       â#rNeG   N   â &r=Bg   ÿ   â ErH H       â &r f      Ä &zGg  € ÿ   ä &zf  €    ä &rcc   ÿ   ä &r>Ah   =   ä $r ?    ÿŽ â #rJeH   J   â &zGh  € G   Ä $rC   ÿŽ â #rPeK   P   â $rHF   ÿŽ â ErT         â &rff   c   ä $rq>   ÿŽ â Er         â 0riÿi    A   ä $rKrG   ÿŽ â Er=        â $rfEf   ÿŽ â #riT   ÿ   â&rg@g   ÿ   â ErV K       â 0rjÿj    A   â #ri    ÿ   â $rGRG   ÿŽ â Er> f       â &rgh   g   â ErH H       â #r j=      â#rVjV      â  r9P     A  â $rgDg   ÿŽ â Er\ G       â #rGi>   ÿ   â#riiH   ÿ   â  rN;N     A  â r;V     A  â #rL     â ErX g       â #r\j\   i   â #rJIJ   N  â#r I      â !r       â &rl   ÿ   Ä &zl  € ÿ   ä &rl   ÿ   â #rGjX   G   â  r\9\     A  â !r5J5       â !r        â &r>Bk   ÿ   ä&zHk  € ÿ   ä &rJ@k   ÿ   Ä &r d      ä&zd  €    ä &rdd      â #rGLG   \  â &r>Am   >   ä&z=m  € H   ä &rmm   J   Ä 0rÿ    A   ä $r ?    ÿŽ ä $rEd   ÿŽ ä $rC   ÿŽ ä$rJF   ÿŽ â !rG       â $rGr=   ÿŽ â ErH         â $rq>   ÿŽ â Er        â $rmDm   ÿŽ â Er         â $r=R=   ÿŽ â Er> J       â 0rÿ    A   â Er        Ä Erm m       ä ErG G       â #rHH   ÿ   âEr= =       â #r   ÿ   â#r     ÿ   â#r>>   ÿ   â &rZ   ÿ   äO&rZ   ÿ   Ä &zZZ  € ÿ   â #rH   H   â#rmm      â#r G       â#r>=   >   â &zn  € ÿ   ä &ro      ä&ro      ä &zZo  € Z   Ä &r@@n   ÿ   ä &zp  €    ä $r?   ÿŽ ä $rE   ÿŽ ä &rp   @   ä $rCZ   ÿŽ Ä &rBBn   ÿ   ä $rFZ   ÿŽ â Er        â $rDD   ÿŽ â Er        â $rrr   ÿŽ ä$rRR   ÿŽ â Er        â &rBAp   B   â Er        â 0rÿ    A   â x   ÿàÿ â $rBqB   ÿŽ â Er r       â 0rÿ    A   â Er? R       â #r   ÿ   â#r   ÿ   â#r   ÿ   â#r   ÿ   â |    pbô â #r      âErB B       â #r?      â Er= D       â  r ;      A  â  rvv9     A  â #rBB      â  r9>     A  â #r==      â  r;;     A  â  r9     A  â #rvOL   v  â #r IH      â#rLm     â#r;IB   ;  â #rL=     â x))@  ÿà÷ â !rv       â !r3 3       â !r       â !r;       â !r       â x$$    ÿàÿ Ä x((    ÿàÿ ä x22    ÿàÿ ä x&&    ÿàÿ ä x**    ÿàÿ ä x,,    ÿàÿ ä x..    ÿàÿ Ä x00    ÿàÿ ä r+ÿ+   ÿäÿ â D)       À â Gy   Þÿÿÿÿƒ ê Ay        € ê Ey  0    € â G‰      € ê r   ÿHŽ Ê ˆs  øÿ H   è ˆs €øÿ H   è ˆs  ùÿ H   è ˆs €ùÿ H   è ˆs 7 úÿ H   è ˆs !€úÿ H   è ˆs   ûÿ H   è ˆs €ûÿ H   è ˆs 5 üÿ H   è ˆs €üÿ H   è ˆs  ýÿ H   è ˆs €ýÿ H   è ˆs 3 þÿ H   è ˆs €þÿ H   è ˆs  ÿÿ H   è ˆs €ÿÿ H   ä Ay        € ê {          ì M        € ê „y      H   b z/  r      Ä z # r  p`ð â „y €   H   ¢ r-//   ÿŽ är ÿ   p"p â „y €  H   â r9//   ÿŽ Æ „y €  H   ( „y	    H   h „y    H   ( y      &   ( y      '   ( „y    H   è „y' €  H   â !r         Æ/ „y, €  H   â !r       Æ_ ‰       " !r       Æ� ‰      b !r!!       Æ‰      ¢ !r 	        Æ‰!      â !r       Æ „y$ €  H   è ‰       " !r        â $z  x  #Ž ä „y&    H   â !r77       â !r
       â/ $z  ~   Ž â ‰      " !r''       â !r       âO x##   ÿàÿ â ‰
      b r/    ÿ† â !r!!       â�  -   ÿàù â „y €  H   ¢ rÿÿ   ÿäÿ â !r),       â z j  ÿ† â„y(    H   â ÿÿ   ÿä â !r"        âz k  � â „y*    H   â 
 j  ÿˆ Æ „y+    H   â 
 k   â !r       ä ‰      " 
  r  ÿàó â !r$$       â !r
       â/ ‰      b ÿÿ   ÿäÿ  ä 
 j  ÿ„ â ‰	 €    â z # r  p`ò Ä 
 k   â ‰7      " !r&&5       â r/    ÿ† ä ‰ €    b x    p"ò Æ ‰"      b !r#       âO !r       ä ‰&      " !r%(       â !r       ä/ „y €  H   â  9   ÿàù â !r3*3       â !r	       â� ‰      b $x	/   ÿŽ â !r++       â !r77       â‰ @    ¢ ÿÿ   ÿä Ä 
 j  ÿˆ â ‰$      â r		    ÿàõ ä 
 k   â ‰(#      " !r       ârÿÿ   ÿäÿ â !r""       â ‰*%      b z
 j  ÿ† â !r&&       â rÿÿ   ÿä â ‰7      " z	 j  ÿ„ Ä r//  €ÿŽ â ‰,'      " z k  � â !r       â/ z		 k   â ‰.3      b     ÿàõ â !r       âO /   ÿàÿ â ‰0)      ¢ 4ÿÿ   ÿä â !r$$       â� 
 j  ÿ„ â ‰2+      â !r(#(       â-   ÿàÿ Ä ‰ !      " !r*%*       â    ÿà÷ ä‰ @    b !r7       â 7    ÿàù â !r,',       â ‰      " 
 k  4 â !r4       â !r.3.       â/ ‰'$      b 8ÿÿ   ÿä â !r0)0       ÄO ‰ €    ¢ 7 j  ÿˆ â !r2+2       ä� ‰)&      â 7 k  8 â !r !        ä‰+(      " x7        â !r       Ä‰-*      b 6ÿÿ   ÿäÿ ä  j  ÿ† â ‰/,      b !r;       â 9   ÿàÿ ä ‰ €    " !r'$'       â/     ÿàõ Ä ‰1.      b !r       âO  k  6� ä ‰30      ¢ !r)&)       Æ� ‰52      â !r+(+       Æ‰4      " !r-*-       Æ‰%" €    b !r/,/       Æ ‰#  €    b !r       Æ ‰! €    " !r1.1       â/ !r303       äO ‰$' €    b !r525       Æ� ‰&) €    ¢ !r44       Æ‰(+ €    â %v  j  7 Ž â !r%"%       ä‰ @    " "ÿÿ   ÿä â !r# #       ä ‰*- €    b  j  ÿ„ â !r!!       Ä ‰,/ €    "  k  " Æ ‰ @    " !r$'$       Æ/ ‰.1 €    b !r&)&       ÆO ‰03 €    ¢ !r(+(       Æ� ‰25 €    â !r       Æ‰74      " !r*-*       Æ‰      b !r,/,       Æ ‰ # @    " !r       Æ ‰! @    " !r.1.       Æ/ ‰"% @    b !r030       ÆO ‰'$ @    ¢ !r252       â� !r747       ä‰)& @    â !r9       Æ‰+( @    " !r #        Æ ‰      " !r!       Æ ‰-* @    b !r"%"       Æ/ ‰#, @    b !r'$'       ÆO ‰      ¨ ‰%. @    ¢ !r)&)       Æ� ‰/0 @    â !r+(+       Æ‰47 €    " !r3       Æ ‰12 @    " !r-*-       Æ‰      b !r#,#       Æ/ ‰!       b !r       ÆO ‰"      ¢ !r%.%       â !r/0/       ä� †	 ;    â !r474       Æ† 9   â !r121       Æ ‰'      " !r       Æ‰)      â !r! !       Æ/ ‰+      b !r"       ÆO †	 3    è ‰-      ¨ ‰#      ( †     â !r''       Æ ‰%      " !r))       Æ� ‰ /      â !r++       Æ/ ‰4 @    h ‰1      b !r--       ÆO †	     â !r##       Æ† !   è †	     â !r%%       â !r//        ä� † '    â !r44       Æ/ †	 
)    â !r11       Æ† 
+   è †	 -    è † #    è †	 %    è † /   è ‰4      ( †	 1    â M™        € ê x   ÿàÿ â !r4       Æ_ r    ÿàñ È r ÿÿ   ÿä  ä z j  ÿ€ È z k     Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      "   " ¹z  l      â Ey  0#    € â ™x?    â y      !   b ¹z  F   
   â r ÿ       â x      ÿ  â y      %   ¢ r- ÿ       â x      ÿ  â x      ÿ  â r1 ÿ       â x      ÿ  â x      ÿ  â r/ ÿ       Ä r ÿ   pRð är ÿ   pRò är   ÿ(Ž È/ xÿ     ä x   ÿ   äO r    ÿ � È xÿ      È |    pbô Ú G)  À!    € ê y      '   " x ðÿÿÿÿÀŽ ä r ÿ       â y      &   b r  €ÿàÿ ä r ÿ       ä x	ÿ     ä x   ÿàÿ Ä r		   ÿ� ä x   ÿ   ä z%  X      ä z#  Y      ä r ÿ       ä z" q  ÿàÿ â 'z  y   Ž È 'z s   Ž â/ zÿ z     Æ $z  p  Ž â zÿ t    â $z w  ÿŽ â z  q      â $z |   Ž âx ÿ   	  â $z |  Ž â x		üÿÿÿÿÀŽ â $z v  Ž âr    ÿŽ â$z v  Ž â r)	  €ÿàÿ â $x$   Ž âr&   ÿŽ â$z p  Ž â xÿ   )  â $x(   Ž Ä $x*   Ž â r )   ÿŽ â$z }  ÿŽ â x    ÿ   â %x�   ÿŽ â x   ÿ   ä r   àµ â%x
�   ÿŽ â x+ÿ     äx0   øŽ Ä xÿ     ä r,   ÿàù är.
   ÿàû ä x!
   øŽ ä x')     äx))   ÿ   Ä r+ÿÿ   +d~ ä x00   ÿüŽ ä r2	   ÿä är4   ÿäÿ ä x!!   ÿüŽ ä r%0   ÿàõ È r	#   ÿä Ê �ybøÿÿ • ¢  r%!   ÿàõ ä xc $       â �y\    • â  r#   ÿä ä r6   ÿàõ â �y]üÿÿ • & r ÿ+   ÿä â %v66 Z  c Ž â �yX   • b r%,   ÿà÷ Ä r%.   ÿàù â �y^üÿÿ • " $x $   ÿŽ â r#2   ÿäÿ ä r)6   ÿàõ â �yR    • " r77   ÿàÿ ä r#4   ÿä â �y    ™ " r7'   ÿä Æ �yK   ™ ( �yJ    ™ ( �yL    ™ ( �yM   ™ " r"   ÿàõ Æ/ �yO   ™ ( �yF6$   • " rÿ+   ÿä Æ �yN(   ™ ( �yQ6    • ( �yP8   ™ " %v Z  c Ž È $x$   ÿŽ â r)   ÿàõ È r   ÿàÿ È r'   ÿä â �y@    • ( �y>   ™ " r%   ÿàõ Æ/ �y9(   ™ ( �y:   ™ " r	#   ÿä Æ �y;8   ™ " r    ÿàõ Æ �y?$   • " r3ÿ+   ÿä â %v Z  c Ž ä �y<    ™ $ $x33$   ÿŽ â r)   ÿàõ È r3   ÿàÿ È r'   ÿä â �yB    • ( �y5   ™ ( �y8   ™ ( �y3(   ™ ( �y=8   ™ " r$   ÿà÷ Ä/ r6
%   ÿàõ â �yE$   • " rÿ+   ÿäÿ â %v Z  c Ž â r7#   ÿä Æ $x$   ÿŽ â r)   ÿàõ â �y66    ™ & r   ÿàÿ È r'   ÿä ä r&   ÿàõ Æ/ �yD   ™ " r        Æ �yC(   ™ " rÿ+   ÿä Æ �yA   ™ " %v Z  c Ž Æ �yG8   ™ " $x$   ÿŽ â r)   ÿàõ ä �yT    • $ r   ÿàÿ ä �yY$   • $ r'   ÿä Ä �y[    • ( �y7   ™ ( �yH(   ™ ( �yI   ™ " r(   ÿàõ Æ/ �yS8   ™ " rÿ+   ÿä â %v Z  c Ž ä �y_$   • $ $x$   ÿŽ â r)   ÿàõ È r   ÿàÿ È r'   ÿä â �y`    • " r*   ÿàõ Æ �yU   ™ " rÿ+   ÿä Æ �yV(   ™ ( �yW   ™ " %v Z  c Ž Æ/ �yZ8   ™ " $x$   ÿŽ â r)   ÿà÷ ä �ya$   • $ r   ÿàÿ È r'   ÿäÿ â �yc    • ( �y   ™ ( �y(   ™ ( �ye8   ™ ( �yd$   • " xfb??  ÿÀŽ ÄO xgbÀÀ  ÿÀŽ ä �yb   ™ ¢  x    pBô ä x\  ÿÀŽ ä� xÿ   X  Ò" ÿ       ä xk^ÀÀ  ÿÀŽ Æ'(fg   @ Ž â xÿ   \  ä xX  ÿÀŽ ä " ÿ       ä x  ÿÀŽ ä xh^??  ÿÀŽ ä x  ÿÀŽ Ä xmRÀÀ  ÿÀŽ ä" ÿ       ä xg]??  ÿÀŽ äxi]ÀÀ  ÿÀŽ ä " ÿ       â '(hk   @ Ž â x]R??  ÿÀŽ Ä x ÿÀŽ â '(]m   @ Ž â x ÿ      ä xXÿ   K  â '(gi   @ Ž â xKÿÀŽ ä xJÿÀŽ â &z^M  € ÿ   â xRÿ   J  Ä xJLÿÀŽ ä x\ÿ   L  ä xL ÿÀŽ â &r M   ÿ   âxXXÿÀŽ â &rMM   ÿ   â xRRÿÀŽ â &zO  € ^   â x\\ÿÀŽ â 0r^ÿF    A   â xFf?   ÿÀŽ â &rLN   ÿ   Ä &zKN  € ÿ   ä &rkXN   ÿ   â xNg?   ÿÀŽ â &rO       ä&rlJO   M   â xMh?   ÿÀŽ â 0r ÿQ    A   â xQ]?   ÿÀŽ â &rjRP      Ä &ziP  € K   â xKÿ   f  â &rn\P   k   â xPÿ   g  â $rF   ÿŽ â xOÿ   h  â $rN   ÿŽ âx]ÿ   ]  â $rllM   ÿŽ ä $rhQ   ÿŽ â Erg        â $rjjK   ÿŽ â Er        â $rfiP   ÿŽ â Er l       â $rnnO   ÿŽ Ä $rii]   ÿŽ â Er h       â #r    ÿ   âErh j       ä Erf f       â #rg g   ÿ   âErn n       ä Erj i       â #ri    ÿ   â#r    ÿ   â #r^f      â#rh^h   g   â#rf^n   i   â#r^^j      â &r>   ÿ   Ä &z>  € ÿ   ä 0r ÿ@    A   ä &r@L9   ÿ   ä &r:      ä &zl9  € ÿ   ä &z:  €    Ä &rjR;   @   ä $rF   ÿŽ ä &zl;  € l   ä $r@N   ÿŽ â Erg        â &ri>   ÿ   ä $r>Pl   ÿŽ â Er @       â &rkX9   ÿ   Ä 0r?ÿ?    A   â Er9 >       â &r:J:   i   â #r@    ÿ   â $rjKj   ÿŽ ä &rn\;   k   â #r;?9   @   â $r:M:   ÿŽ ä $rQ   ÿŽ â Erj j       â 0r9ÿ<    A   â#r> g   ÿ   â 0r<ÿ<  0 A   â Er :       â $rnOn   ÿŽ â Er        â #r>?j   >   â $rl]l   ÿŽ â  r<     A  â  r:<;     A  â #r@    ÿ   â#r     ÿ   â Ern n       â #rhh9     â #r>9>   :  â &r5   ÿ   â Erg l       â &z5  € ÿ   â !r>       â &r8      â #r:?n   @   â &z>3  € ÿ   â #r??g       â &z8  €    Ä &r55   ÿ   ä &r L3   ÿ   ä&r3X3   ÿ   ä $rF   ÿŽ ä &z>=  € >   ä $rN   ÿŽ â !r/h/       â &r@J8   5   Ä &r8R=       ä&rh\=   3   â Er3        â $r;P>   ÿŽ â Er        â 0r ÿB    A   ä $r5Q   ÿŽ ä $r@M@   ÿŽ Ä $r8K8   ÿŽ â Er; ;       â 0rEÿE    A   â Er 5       â $r=]>   ÿŽ â #r>    ÿ   â$rhOh   ÿŽ â Er @       ä Er5 8       â #r8 3   ÿ   â#r;E;   >   â#r@    ÿ   âEr h       â #r8E5   8   â Er= =       â #r     ÿ   â 0r3ÿ6  0 A   Ä 0r5ÿ6    A   â  r6<;     A  â #r@E   @   â#r>E=       â  r^3     A  â  r 3?     A  â #r698   6  â  r3>     A  â #rf5     â #r5:      â !r6       â &r D   ÿ   Ä &z6D  € ÿ   ä &rDD   ÿ   â !r       â !r       â #r5@     â &rLC   ÿ   ä &z:C  € ÿ   ä &rA       Ä &rCXC   ÿ   ä &z6A  € 6   ä &rDJA   D   â !r       â &r8RG      ä &z:G  € :   ä $rF   ÿŽ Ä &r>\G   C   ä $rMD   ÿŽ ä $rN6   ÿŽ ä$r=Q6   ÿŽ â Er;        â $r6P:   ÿŽ â Er        â $r8K8   ÿŽ â Er        â $r>O>   ÿŽ Ä $r:]:   ÿŽ â Er =       â 0r ÿT    A   â Er8 8       â 0rYÿY    A   â Er? >       ä Er= 6       â #r6 ;   ÿ   âErA :       â #r>    ÿ   â#r:    ÿ   â#r     ÿ   â #r6Y8   6   â#r8Y?   >   â#r=Y=   :   â#r:YA       â &r 7   ÿ   Ä &z>7  € ÿ   ä &r77   ÿ   ä &rLH   ÿ   ä &zBH  € ÿ   ä &rI       ä &z>I  € >   Ä &rDJI   7   ä &r@RS      ä &zBS  € B   ä $rF   ÿŽ ä $rMD   ÿŽ ä $rN>   ÿŽ Ä $r;Q>   ÿŽ â Er7        â $r>PB   ÿŽ â Er        â $r@K@   ÿŽ â Er        â $rB]B   ÿŽ â Er ;       â 0r ÿ[    A   ä &rHXH   ÿ   â Er; @       â 0r_ÿ_    A   â Er? >       â &rH\S   H   â ErC B       â #r> 7   ÿ   â#rB    ÿ   â#r@    ÿ   â#r     ÿ   â $rHOH   ÿŽ Ä #r?_?   @   â#r@_C       â  r <=     A  â ErA H       Æ #r 96      â  r:3:     A  â #rB_A   B   â  r3@     A  â !r1 1       â #r>_;   >   â  r?<?     A  â &r U   ÿ   ä&z6U  € ÿ   ä &rUU   ÿ   â #r858   :  â#rB5B     â #r?9>   ?  â &rLV   ÿ   Ä &zV  € ÿ   ä &rW       ä&z6W  € 6   ä &r>JW   U   â !r8       â &r7XV   ÿ   ä &r8RZ      Ä &z:Z  €    ä $rF   ÿŽ ä $rM>   ÿŽ ä $rN6   ÿŽ ä$r6Q6   ÿŽ ä &rZ\Z   7   â Er7        â $r8K8   ÿŽ â Er        â $r=P:   ÿŽ â Er        â 0r ÿ`    A   â Er 6       â $r:]:   ÿŽ â Er; 8       â 0raÿa    A   â #r6 7   ÿ   â#r>    ÿ   â#r8    ÿ   âEr: :       â #r     ÿ   â &r   ÿ   Ä &r   ÿ   ä &z  € ÿ   â #r:a:       â &rLL   ÿ   ä&rXX   ÿ   ä &z   € ÿ   ä &rb      ÄO &rJJb      ä &zb  €    ä &rLRe   L   ä&rX\e   X   ä &zee  €     ä $rF   ÿŽ Ä $rMJ   ÿŽ ä $r N   ÿŽ ä$rQ   ÿŽ â Er        â $rPPe   ÿŽ âEr          â $r]]e   ÿŽ â Er        â 0rcÿc    A   â Er        â $rZOZ   ÿŽ â x   ÿàÿ â $rLKL   ÿŽ â Er= =       â $rXOX   ÿŽ â ErP P       â 0rdÿd    A   â Er ]       â #rc   ÿ   â#rc    ÿ   â#rc   ÿ   â#rcc   ÿ   â |    pbô â !r?       â Er? Z       â #r=a=   8   â ErL L       â #rdP      âErX X       â #rd   c   â r:3:     A  â #r6a;   6   â#r>a?   >   â #rLdL      â#rXdX      â  r <=     A  â r<     A  â  r33     A  â #r:5>   :  â #r 96      â#r9L     â #rX5X   3  â x%%@  ÿà÷ â !rB       â !r- -       â !r:       â !r       â !rX       â x    ÿàÿ Ä x""    ÿàÿ ä x      ÿàÿ ä x$$    ÿàÿ ä x&&    ÿàÿ ä x((    ÿàÿ ä x**    ÿàÿ Ä r#ÿ#   ÿäÿ â D)       À â Gy  °áÿÿÿÿƒ ê Ay        € ê Ey      € â G‰  ð     € ê $xÀ  Ž Ê ˆs / ùÿ H   è ˆs €ùÿ H   è ˆs  úÿ H   è ˆs €úÿ H   è ˆs  ûÿ H   è ˆs €ûÿ H   è ˆs 1 üÿ H   è ˆs €üÿ H   è ˆs  ýÿ H   è ˆs €ýÿ H   è ˆs - þÿ H   è ˆs €þÿ H   è ˆs  ÿÿ H   è ˆs €ÿÿ H   ä Ay        € ê {          ì M        € ê „y      H   b z  r  p`ð Ä z3  r      â „y €  H   ¢ r ÿ   p"p Æ „y    H   è „y    H   ( „y €   H   h „y    H   ( „y	 €  H   ( „y    H   è „y €  H   è „y    H   â !r  /       Æ/ „y  €  H   â r/33   ÿŽ â !r       äO ‰       " !r       Æ� „y €  H   â !r       Æ„y#    H   b !r       Æ„y" €  H   â !r11       Æ y      '   ¢ !r		       Æ ‰      â !r        Æ ‰      " !r       Æ y       &   â !r--       Æ ‰      " !r!        Æ ‰      h ‰
      h ‰	      b !r##       Æ/ ‰1      b !r       â !r       â !r%"       â $z   x  Ž â� ‰"      " x   ÿàÿ ä $z  ~   Ž âO !r       â !r       â !r       â!r       â‰       ¢ 
  r  ÿàõ â !r

       Ä ‰      â ÿÿ   ÿä â !r	       ä ‰ €    " 
 j  ÿ„ â !r1       ä/ ‰$-      b 
 k   ä z  r  p`ò â ‰      b !r""       â r133   ÿŽ Ä ‰&!      " x    p"ò Æ ‰(#      " !r         ÆO ‰*%      ¢ !r       Æ� ‰	      â !r       âr3    ÿˆ ä ‰      " !r$-$       â/ r.ÿÿ   ÿä ä ‰
      b !r       â /   ÿàû Ä ‰      b !r&!&       â z j  ÿˆ ä ‰      " !r(#(       â ,ÿÿ   ÿäÿ â !r*%*       âO ‰!      ¢ 
 j  ÿŠ â !r	       Ä� ‰#       â r	3    ÿ„ â !r       ä‰%"      " z k  . â !r
       ä/ ‰ @    b  1   ÿà÷ â !r       Ä‰'$      b ÿÿ   ÿäÿ â !r       ä ‰ €    " rÿÿ   ÿä ä 

 j  ÿ† â ‰)&      " z	 j  ÿ„ â !r!!       âO 
 k  � â ‰+(      ¢ !r# #       â� z		 k   Ä ‰ €    â !r%"%       â
 k  ,� ä ‰ €    " !r       â/ 3   ÿàÿ ä ‰-*      b !r'$'       â    ÿàù Ä ‰ €    b !r       â .ÿÿ   ÿä ä ‰ €    " !r)&)       â  j  ÿˆ ä‰ €    " !r+(+       âO  k  . Ä ‰"! €    ¢ !r       â� $x3   ÿŽ â /   ÿàÿ â ‰$# €    â !r       är    ÿàõ â ‰&% €    " !r-*-       â/ /    ÿà÷ Ä ‰(' €    b !r       ârÿÿ   ÿä ä ‰ @    b !r       â z j  ÿ„ ä‰*) €    " !r       â z k   Ä ‰,+ €    " !r"!"       âO 0ÿÿ   ÿäÿ ä ‰ @    ¢ !r$#$       â� / j  ÿ† ä ‰ @    â !r&%&       â/ k  0� Ä ‰.- €    " !r('(       Æ/ ‰ @    b !r       âx        ä ‰ @    b !r*)*       ä %v  j   Ž â ‰      " !r,+,       Æ ‰  @    " !r       ÆO ‰!" @    ¢ !r       Æ� ‰#$ @    â !r.-.       Æ‰%& @    " !r       Æ/ ‰'( @    b !r       Æ‰      b !r/       â 1   ÿàÿ ä ‰)* @    " !r         â     ÿàõ ä ‰+, @    " !r!"!       âO 0ÿÿ   ÿä Ä ‰      ¢ !r#$#       â�  j  ÿ„ ä‰-. @    â !r%&%       â k  0 ä ‰      " !r'('       Æ/ ‰      b !r1       Æ‰      b !r)*)       Æ ‰       " !r+,+       Æ ‰ !      " !r       ÆO ‰#      ¢ !r-.-       Æ� ‰"%      â !r       Æ‰'      " !r       Æ/ †	 /    â !r       Æ† 1   â !r        Æ ‰$)      " !r!!        Æ ‰+      b !r##       ÆO ‰-      ¢ !r%%"       Æ� †	     â !r''       Æ†     è †	     è †    â !r))$       Æ †	     â !r++       Æ/ † !    â !r--       ÆO †	 #    è † %   è †	 
'    è † )    è †	 +    â M™        € ê †y -   â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      "   " ¹z  l      â Ey  @    € â ™x?    â y      !   b ¹z  F   
   â x      ÿ  â x      ÿ  â x      ÿ  â x      ÿ  â x      ÿ  â x      ÿ  â r ÿ   pRð Ä r ÿ   pRò är   ÿ(Ž È/ x ÿ     È r     ÿ � È xÿ      È |    pbô Ú G)       € ê y      '   " x  ðÿÿÿÿÀŽ ä r ÿ       â y      %   b r ÿ       ä r   €ÿàÿ â y      &   ¢ z  X      ä z
  Y      Ä r ÿ       â 'z  y   Ž â x   ÿ   Æ/ $z }  ÿŽ â z	ÿ z     â 'z s   Ž âO xÿ     ä x   ÿàÿ â$z  p  Ž â r   ÿ� â $z w  ÿŽ â zÿ t    â $z	 |   Ž âx ÿ     â $z		 |  Ž â xüÿÿÿÿÀŽ â $z^ v  Ž â x    ÿ   â $z	 v  	Ž â r  €ÿàÿ â %x^^�   ÿŽ â z  q      Ä r#    ÿŽ â $z` p  	Ž â x   ÿ   ä x##   ÿ   â %x``�   ÿŽ â xÿ     ä x)ÿ   #  â $x   Ž âr!#^   ÿàù â $x   Ž â r   àµ Ä x ÿ     äx%^   øŽ är##`   ÿàû ä x+`   øŽ ä r	   ÿŽ är   ÿŽ Ä x     ä r'_)   ÿä ä z q  ÿàÿ ä x   ÿ   ä r ÿÿ    d~ ä x%%   ÿüŽ Ä r)a)   ÿäÿ ä x++   ÿüŽ ä r,%   ÿàõ È r-_
   ÿä ä r2+   ÿàõ Æ �yH,øÿÿ • ¢  r3a
   ÿä ä r   ÿàõ â �yS,    • â  xL $       ä r ÿ    ÿä â �y[,üÿÿ • " r0!   ÿà÷ Æ �yK2   • b r.#   ÿàù Æ �yY2üÿÿ • " %v, Z  L Ž Æ �yJ2    • " $x $   ÿŽ â r1
'   ÿäÿ är/
)   ÿä ä r--   ÿàÿ â �yP0   ™ " r2,   ÿàõ Æ/ �yG.   ™ " r3-   ÿä Æ �yN0    ™ ( �y.    ™ ( �y2   ™ " r1   ÿàõ Æ �y2   ™ " r"ÿ    ÿä Æ �yR2(   ™ " %v01 Z  L Ž Æ �y\28   ™ " $x/"$   ÿŽ â/ r.0   ÿàõ ä �y ,    • $ r11/   ÿàÿ ä �yO,$   • $ r/1   ÿä Ä �y*0    • ( �y".   ™ ( �y$.(   ™ ( �y&.   ™ ( �y(.8   ™ " r4`   ÿà÷ Ä r9	   ÿàù â �y,0$   • "# r:^   ÿàõ ä r5a
   ÿäÿ ä r2ÿ    ÿä â %v89 Z  L Ž â r;_
   ÿä â �y.4    ™ $ $x/2$   ÿŽ â r68   ÿàõ â �y-:    ™ & r99/   ÿàÿ È r79   ÿä â �y38    • " r?   ÿàõ Æ �y/6   ™ " r5ÿ    ÿä â %v>? Z  L Ž ä �y06(   ™ $! $x55$   ÿŽ ä �y16   ™ " r<>   ÿàõ Æ �y268   ™ " r??5   ÿàÿ Æ �y48$   • " r=?   ÿä Æ �y:>$   • ( �y5<   ™ ( �y6<(   ™ ( �y7<   ™ " rE   ÿàõ Æ �y8<8   ™ "! r;ÿ    ÿä Æ �y9>    • " %vDE Z  L Ž È $x;;$   ÿŽ â rBD   ÿàõ È rEE;   ÿàÿ È rCE   ÿä â �y?D    • "# rM   ÿàõ Æ �y;B   ™ " rAÿ    ÿä Æ �y<B(   ™ ( �y=B   ™ " %vLM Z  L Ž Æ �y>B8   ™ " $xAA$   ÿŽ â rVL   ÿàõ ä �y@D$   • $ rMMA   ÿàÿ È rWM   ÿä â �yFL$   • ( �yAV   ™ ( �yCV(   ™ (# �yBV   ™ ( �yDV8   ™ ( �yEL    • " x    pBô Ú "L ÿ       ä xQH??  ÿÀŽ äOxUHÀÀ  ÿÀŽ ä xIS  ÿÀŽ ä� "H ÿ       Ê '(QU   @H Ž â xIÿ   S  ä xHÿ   K  äxMI  ÿÀŽ ä xIH  ÿÀŽ ä xU[??  ÿÀŽ äxSY??  ÿÀŽ Ä xWJ??  ÿÀŽ ä/x]JÀÀ  ÿÀŽ ä "H ÿ       ä x[[ÀÀ  ÿÀŽ ä xYYÀÀ  ÿÀŽ ä xKK  ÿÀŽ Ä "J ÿ       â '(W]   @H Ž â xHPÿÀŽ ä xPÿ   P  â '(U[   @L Ž â xLGÿÀŽ Æ '(SY   @J Ž â xKÿ   G  ä xJNÿÀŽ â &rYH   ÿ   âxIÿ   N  â &zT  € ÿ   â xNÿÀŽ ä xGPÿÀŽ â &rPL   ÿ   â xKKÿÀŽ Ä xMÿ     â &rYJ   Y   âxIIÿÀŽ â &z  € T   â xMMÿÀŽ â &rXGR   ÿ   âxTU?   ÿÀŽ â &z[R  € ÿ   ä &rKR   ÿ   â xRS?   ÿÀŽ â &rN   P   â xPQ?   ÿÀŽ Ä xVW?   ÿÀŽ â &rXI\   X   ä&zb\  € [   â xQÿ   Q  â &r\M\      â xSÿ   S  â $rYYP   ÿŽ â xUÿ   U  â $rR   ÿŽ â xWÿ   W  â $rT   ÿŽ Ä $rV   ÿŽ â ErY Y       â 0r ÿ     A   â Er        â $rXXQ   ÿŽ â Er        â $r\\S   ÿŽ â Er        â $rZbU   ÿŽ ä$r[bW   ÿŽ â ErX X       â 0rOÿO    A   â ErZ Z       â #rb Y   ÿ   âEr\ \       â #r    ÿ   âEr[ [       â #r    ÿ   â#r     ÿ   â #rXOX   b   â#rOZ      â#rO\      â#r O[       â &z"  € ÿ   Ä &rOH"   ÿ   ä&rYL"   ÿ   ä &r"G$   ÿ   ä&z[$  € ÿ   ä &r$K$   ÿ   ä &z&  €    Ä &rOJ&   O   ä&rYN&   Y   ä &r"I(   "   ä&z&(  € [   ä &r$M(   $   ä $r(T   ÿŽ Ä $rV   ÿŽ ä $rZU&   ÿŽ ä$r&W&   ÿŽ â Er        â 0r*ÿ*    A   ä $rYRY   ÿŽ â Er[ (       â $r$S$   ÿŽ â Er& &       â 0r,ÿ,    A   â #r*   ÿ   â $r"Q"   ÿŽ â ErY Y       â $rOPO   ÿŽ â Erc $       â #r$*[   ÿ   â Ere Z       â #r&,&      â Er] "       â 0rÿ.  0 A   âErO O       â #r"*Y   ÿ   â #re,e   $   â 0r$ÿ-  0 A   â r       A  â 0r.ÿ.    A   â #r",c   "   â  rY&     A  â #rO*O   ÿ   â 0r-ÿ-    A   â  r$     A  â #r .      â #r"."   Y  â #rO,]   O   â  r($e     A  â #rX-     â !r        â !r"       â #rO-O   (  â &r H/   ÿ   Ä &z"/  € ÿ   ä &r(L/   ÿ   â !r       â &r&G0   ÿ   ä &z/0  € ÿ   ä &rJ1       ä &z1  € "   Ä &r0K0   ÿ   ä &r1N1   (   ä &r"I2   &   ä$rP   ÿŽ ä &z 2  € /   ä &r0M2   0   Ä $r1R1   ÿŽ ä $r&T   ÿŽ ä$r*V   ÿŽ â Er(        â $rQ"   ÿŽ â Er" 1       â 0r3ÿ3    A   ä $r0S0   ÿŽ Ä $r,U    ÿŽ ä$r/W    ÿŽ â Er  &       â 0r4ÿ4    A   â Er& *       ä Er*        â #r3(   ÿ   âEr0 0       â #r3    ÿ   âEr2 /       â #r/3"   ÿ   â&r H5   ÿ   â #r(4*      â#r33&   ÿ   â #r40   /   â &z"5  € ÿ   Ä &r&G6   ÿ   ä &r/J7       â Er, ,       â &z17  € "   ä &r&I8   &   ä $r/P/   ÿŽ â #r4,      â $r T1   ÿŽ Ä $r,Q&   ÿŽ â Er* /       â &r5L5   ÿ   ä 0r9ÿ9    A   â #r442   3   â $r1V1   ÿŽ â Er          â &z36  € ÿ   â Er, ,       â &r5N7   5   â #r/9*   ÿ   â0r:ÿ:    A   â Er& 1       â &z"8  € 3   â #r19    ÿ   â $r5R5   ÿŽ â #r,:,   /   â  r $     A  â  r/4     A  â $r0U"   ÿŽ Ä $r2W"   ÿŽ â Er" 5       â #r-(      â #r .   /  â Er0 0       ä #r39"   ÿ   â !r        â &r H;   ÿ   ä &z";  € ÿ   Ä &r;L;   ÿ   â #r99&   ÿ   â #r1:0   1   â &r&G<   ÿ   â!r       â &z/<  € ÿ   ä &rJ=       ä&r<K<   ÿ   ä &r;N=   ;   Ä &z=  € "   â  r1$1     A  â &r&I>   &   ä$rP   ÿŽ ä &z">  € /   ä &r<M>   <   ä $r;R;   ÿŽ Ä $r T   ÿŽ â#r1-,   1  â $r*V   ÿŽ â Er(        â $rQ&   ÿŽ ä 0r?ÿ?    A   ä $r<S<   ÿŽ ä $r,U"   ÿŽ Ä $r/W"   ÿŽ â Er" ;       ä Er          â 0r@ÿ@    A   â Er& *       â #r?(   ÿ   âEr*        ä Er< <       ä Er0 /       â #r/?"   ÿ   â Er, ,       â &r6K6   ÿ   â #r?    ÿ   â #r*@*      â!r1       â #r@<   /   â&r6M8   6   â #r@,      â &r/HA   ÿ   Ä &r1LA   ÿ   ä &zAA  € ÿ   â  r $     A  â $r6S6   ÿŽ â Er2 2       â &rGGC   ÿ   ä&rKKC   ÿ   ä &r/JB   /   Ä &r1NB   1   ä &zAB  € A   ä &zCC  € ÿ   â #r-*      â &r"ID   G   â Er6 6       â &r(MD   K   â #r??&   ÿ   â $r P/   ÿŽ â #r2:2   9   â $r1R1   ÿŽ Ä &zDD  € C   ä $rTTA   ÿŽ ä$r&VA   ÿŽ â #r3:6   3   â $rQQ"   ÿŽ â  r22     A  â $rUUD   ÿŽ â Er          â $rDWD   ÿŽ â ErT T       â 0rEÿE    A   â Er" 1       Ä Er& &       â #r2.3   2  â x   ÿàÿ â $rSS(   ÿŽ â ErU U       â 0rFÿF    A   â ErD D       â #r1ET   ÿ   â#r/E    ÿ   â#r3E"   ÿ   â#rEE&   ÿ   â |    pbô Ä Er( Q       â #r@@0   ?   â Er  S       â #r1FU   1   â#rDFD   E   â#r(F(   /   â  r/@     A  â r$$1     A  â #r3F    3   â  rDD     A  â #r .   /  â #r--(   $  â x@  ÿà÷ â #rD.3   D  â !rO       â !r2       â !r       â !r        â !r-       â !rD       â x    ÿàÿ Ä x    ÿàÿ ä x		    ÿàÿ ä x    ÿàÿ ä x    ÿàÿ ä x    ÿàÿ ä r
ÿ
   ÿäÿ â D)       À â Gy  `åÿÿÿÿƒ ê Ay        € ê Ey  ð     € â G‰  Ð     € ê $x€  Ž Ê ˆs  úÿ H   è ˆs €úÿ H   è ˆs  ûÿ H   è ˆs €ûÿ H   è ˆs  üÿ H   è ˆs €üÿ H   è ˆs  ýÿ H   è ˆs €ýÿ H   è ˆs  þÿ H   è ˆs €þÿ H   è ˆs  ÿÿ H   è ˆs €ÿÿ H   ä Ay        € ê {          ì M        € ê „y     H   b z+  r      Æ „y  €   H   ¢ r'++   ÿŽ är)++   ÿŽ â „y    H   è „y €  H   ( „y    H   h „y €  H   ( „y	    H   ( „y
 €  H   ( „y €  H   ( „y    H   è „y    H   â !r       â/ !r        ÄO „y €  H   â !r       Æ� ‰       " !r       Æ‰      b !r       Æ‰      ¢ !r       â !r	       È ‰      â !r

       Æ ‰	      " !r       â !r         ä ‰      " !r       â !r       ä/ ‰       b !r       Æ ‰      b !r       âO !r"       ä ‰
      ¢ !r       Æ� ‰      â !r		       Æ‰      ( ‰      " !r       Æ ‰      " !r        Æ/ y      %   b !r       Æy       &   b !r
       âO !r       ä� y      '   ¢ !r       Æ‰      â !r       Æ ‰	      " !r        Æ ‰ €    " x   ÿ   Æ/ ‰      b x   ÿàÿ Æ ‰      b $z   x  Ž âz  r  p`ð ä ‰      b $z  ~   Ž âO z  r  p`ò ä ‰      ¢ !r       â� r ÿ   p"p Ä ‰      â !r	       âx    p"ò ä ‰      " !r       Æ ‰ €    " !r       ä/ 
  r  ÿà÷ â‰#       b !r       â  '   ÿàõ ä ‰%"      b !r       âÿÿ   ÿäÿ â !r!       âO ‰ €    ¢ 
 j  ÿ† â !r       Ä� ‰	 @    â ÿÿ   ÿä â !r       ä‰ €    " 
 j  ÿ„ â !r
       ä ‰ €    " r+    ÿˆ â !r# #       Ä/ ‰ €    b 
 k  � â !r%"%       ä ‰ €    b rÿÿ   ÿä ä z j  ÿˆ â ‰"! €    b !r       âO 
 k   Ä ‰ €    ¢ !r	       â� z k   ä ‰  €    â !r       â	+   ÿàÿ ä ‰
 @    " !r       â  )   ÿàù Ä ‰$# €    " !r       â/ * 	   ÿà÷ ä ‰&%      b !r       âÿÿ   ÿä â !r"!"       â ‰ @    b 
 j  ÿˆ â !r       ÄO ‰ @    ¢ 
	 k   â !r         ä� ‰      â !r
       Æ‰ @    " r+    ÿ„ â !r$#$       ä ‰ @    " +ÿÿ   ÿäÿ â !r&%&       ä/ ‰ @    b * j  ÿ† ä r(ÿÿ   ÿä â ‰!" @    b !r       â* k  +� Ä ‰ @    b !r       âO z
 j  ÿ„ ä ‰  @    ¢ !r+       â�         ä ‰      â !r       â'   ÿàÿ Ä ‰#$ @    " !r       â z k  ( ä ‰%& €    " !r       â/ !r""!       â !        â ‰      b !r       Æ‰      b !r         ÆO ‰      ¢ !r-       â�     ÿàõ ä‰      â %  j   Ž â !r$$#       â(ÿÿ   ÿä â ‰      " %  j  ! Ž â !r&&%       â  j  ÿ„ â‰      &  k  ( â ‰!       " !r       Æ/ ‰#"      b !r       Æ‰'& @    b !r       ÆO ‰%$      ¢ !r       Æ� †	 +    â !r       Æ† -   â !r       Æ †	     â !r! !       â !r#"#       ä/ †     â !r'&'       Æ†	     â !r%$%       ÆO †    è †	     è †     è †	 
!    è † 
#   è ‰'      ( †	 %    â M™        € ê x)   ÿàÿ â� !r''       Æ r    ÿàñ È r ÿÿ   ÿä  ä z j  ÿ€ È z k     Ê †y '    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y       "   " ¹z  l      â Ey  °    € â ™x?    â y      !   b ¹z  F   
   â r ÿ       â x      ÿ  â y)      &   ¢ r ÿ       â x
      ÿ  â r ÿ       â y+      '   â r ÿ       â x      ÿ  â r  ÿ   pRð Ä r  ÿ   pRò är    ÿ(Ž È/ xÿ     È r   ÿ � È xÿ     È |    pbô Ú G)  `    € ê y      %   "Î xðÿÿÿÿÀŽ â $z) w  ÿŽ â r* ÿ       ä r  €ÿàÿ ä r        â 'z+ y  * Ž â r( ÿ       ä xÿ     Ä zÿ z    ä r   ÿ� ä z  q      ä xÿ     ä xüÿÿÿÿÀŽ ä x   ÿ   Ä x   ÿ   â 'z) s  ( Ž â x   ÿ   ä r  €ÿàÿ â $z p  Ž âzÿ t    ä x   ÿàÿ â $z |  Ž â r   ÿŽ â $z |  Ž â z  Y      â $z+ }  ÿŽ â z q  ÿàÿ â $z v  Ž Ä $z, v  Ž â r   àµ â $z. p  Ž â r   ÿŽ â$x   Ž âr   ÿŽ â %x,,�   ÿŽ â xÿ     ä z  X      â %x..�   ÿŽ â x     Ä x   ÿ   ä x   ÿ   ä xÿ     äx,   øŽ äx!.   øŽ ä r,   ÿàù Ä r.   ÿàû ä xÿ     ä r ÿ       ä r#-   ÿä ä rÿÿ   d~ ä x   ÿüŽ Ä x!!   ÿüŽ ä r/   ÿäÿ ä r$   ÿàõ È r%-   ÿä ä r&!   ÿàõ Æ �yB$øÿÿ • ¢  r'/   ÿä Æ �yA$    • â  r5   ÿàõ ä xL $       â �y=&   • " rÿ   ÿä ä r2   ÿà÷ â�y@$üÿÿ • b r0   ÿàù â %v45 Z  L Ž â r3#   ÿäÿ â�y?&üÿÿ • b r1   ÿä Æ �y<&    • b $x%$   ÿŽ â r84   ÿàõ ä �y2   ™ d r55%   ÿàÿ ä �y2    ™ h �y 0   ™ b r95   ÿä Æ �y"0    ™ h �y$8   ™ b r3   ÿàõ Æ �y%8(   ™ h �y&8   ™ b! r:ÿ   ÿä Æ �y'88   ™ h �y74$   • b %v23 Z  L Ž Æ �y64    • b $x1:$   ÿŽ â r02   ÿàõ È r331   ÿàÿ È r13   ÿä â �y>2    • h �yF0   ™ h �yE0(   ™ h �yC0   ™ b r8,   ÿàõ Æ �yG08   ™ b r9-   ÿä ä r5   ÿàõ â/ �yD2$   • f r:ÿ   ÿä â �yH8    ™ b %v45 Z  L Ž È $x3:$   ÿŽ â/ r24   ÿàõ È r553   ÿàÿ È r35   ÿä â �yN4    • h �yI2   ™ h �yJ2(   ™ h �yK2   ™ h �yM28   ™ b r:.   ÿàõ Ä r9   ÿà÷ â �yO4$   • b r;/   ÿä ä r0ÿ   ÿäÿ â %v89 Z  L Ž ä �y::    ™ d $x10$   ÿŽ â r08   ÿàõ È r991   ÿàÿ È r19   ÿä â �y;8    • b r5   ÿàõ Æ/ �yR0   ™ b r2ÿ   ÿä â %v45 Z  L Ž ä �yQ0(   ™ d $x32$   ÿŽ ä �yP0   ™ b r24   ÿàõ Æ �yS08   ™ b r553   ÿàÿ Æ �yT8$   • b r35   ÿä Æ �yW4$   • h	 �yU2   ™ h �yV2   ™ h �yL2(   ™ h �y828   ™ h �y94    • b	 x    pBô Ú "0 ÿ       ä/ "2 ÿ       ä xYBÀÀ  ÿÀŽ äOxXB??  ÿÀŽ ä x1A  ÿÀŽ Ê� '(XY   @0 Ž â x0ÿ   A  ä x4ÿ   =  äx10  ÿÀŽ ä xY@ÀÀ  ÿÀŽ äxA?ÀÀ  ÿÀŽ ä x3=  ÿÀŽ Ä x54  ÿÀŽ ä "0 ÿ       ä xB@??  ÿÀŽ ä x=<ÀÀ  ÿÀŽ ä "4 ÿ       ä x@???  ÿÀŽ Ä x?<??  ÿÀŽ â '(BY   @0 Ž â x1ÿÀŽ Æ '(@A   @2 Ž â x3ÿÀŽ ä xÿ     â '(?=   @4 Ž â xÿ     ä x5 ÿÀŽ ä x ÿ      ä x="ÿÀŽ Ä x0ÿ   "  ä xÿÀŽ ä x"ÿÀŽ ä x ÿÀŽ â &zA$  € ÿ   â x 0ÿÀŽ â &r01$   ÿ   ä&r25$   ÿ   Ä &r4%   ÿ   ä&z<%  € ÿ   ä &rY%   ÿ   ä &r$3&   0   ä&z%&  € A   ä &r2=&   2   Ä &r0"'   4   ä&z&'  € <   â x<B?   ÿÀŽ â &r' '   Y   â xY@?   ÿÀŽ â 0r4ÿ7    A   â x7X?   ÿÀŽ ä x[ÿ   B  ä xB??   ÿÀŽ â $rZ%<   ÿŽ âxAÿ   X  â $rX$7   ÿŽ â x@ÿ   @  â $r22Y   ÿŽ â x?ÿ   ?  â $r$%B   ÿŽ â Er% Z       â 0r6ÿ6    A   â ErX X       â $r00A   ÿŽ Ä $r\'@   ÿŽ â Er' 2       â $r]&[   ÿŽ âEr$ $       â $r&&?   ÿŽ â Er0 0       â #r26%   ÿ   âEr] ]       â #rX6X   ÿ   âEr\ \       â #r%6'   ÿ   âEr& &       â #r66$   ÿ   â #r'40   X   â#r]4]   2   â#r$4\   %   â#r&4&   6   â &r01F   ÿ   Ä &z4F  € ÿ   ä &zXE  € ÿ   ä &r%E   ÿ   ä &r23C   0   ä &z0C  € 4   ä &z4G  € X   Ä &r6"G   %   ä $rX72   ÿŽ ä $r%<0   ÿŽ ä $r2A6   ÿŽ â ErX X       â 0r>ÿ>    A   â Er% %       â $r6[4   ÿŽ Ä &rF5F   ÿ   â Er2 2       â 0rDÿD    A   â Er6 6       â &rEE   ÿ   â #rX>X   ÿ   â#r%>%   ÿ   â &rF=C   F   ä &rG G   E   â #rCD2   X   â#rED6   %   â 0r2ÿH    A   Ä $rFYF   ÿŽ ä $r%B0   ÿŽ ä 0rHÿH  0 A   â ErF F       â $r0@G   ÿŽ â Er% %       â $rG?4   ÿŽ â  r4]H     A  â  rEHE     A  â Er0 0       â #r]>F   ÿ   âErG G       â #r4'2   4  â #r>>%   ÿ   â #r62C   E  â #r0D0   ]   â !r4       â #rGDG   >   â !r6       â &r41I   ÿ   Ä &z6I  € ÿ   ä &rD5I   ÿ   ä &r'J   ÿ   ä&z>J  € ÿ   ä &r43K   4   ä &rJJ   ÿ   Ä &z%K  € 6   ä &rF=K   D   ä &r6"M   '   ä&zDM  € >   ä $r'74   ÿŽ ä &rM M   J   Ä $rIYF   ÿŽ ä $rC<%   ÿŽ ä$rJB%   ÿŽ â ErE '       â $rFA6   ÿŽ â Er' I       â 0rNÿN    A   â Er% C       â $r4[D   ÿŽ â ErC J       â $r>@M   ÿŽ â ErI F       â $r6?D   ÿŽ â #rDNE   ÿ   â#rJN'   ÿ   â#rFN%   ÿ   â0rOÿO    A   â #rNNC   ÿ   â Er' 4       â 0r%ÿ:  0 A   â ErC >       Æ #rEO'   F   â0r'ÿ:    A   â #rJOC   J   â  rC&%     A  â  r&%G     A  â #rIOI   D   â  rGHE     A  â #rC$'   C  Æ #r$2I   G  â !rC       â &zCR  € ÿ   ä !r	$	       â &r$1R   ÿ   ä&rR5R   ÿ   â #rE'0   &  â &r&Q   ÿ   ä &rR=P   R   Ä &r&"S   &   ä $rRYR   ÿŽ â ErK 6       â &z0Q  € ÿ   ä &r$3P   $   ä &zCP  € C   ä $r6A&   ÿŽ â Er& R       â 0r;ÿ;    A   Ä &z4S  € 0   ä $r$7$   ÿŽ ä $r0<C   ÿŽ â#rG;&   ÿ   â $rCBC   ÿŽ ä $r:[4   ÿŽ ä$rD?4   ÿŽ â Er4 $       â &r&1U   ÿ   â Er$ 0       Ä Er0 C       â &r&3V   &   â !r
E
       â &r55U   ÿ   â #rC;4   ÿ   â#rE;$   ÿ   â &zUU  € ÿ   â #r;;0   ÿ   â $r&7&   ÿŽ ä &r0=V   5   Ä &r3L   ÿ   ä&r7L   ÿ   ä&rQQ   ÿ   ä &z5L  € ÿ   ä &zUV  € U   ä &r""8   3   Ä &r7 8   7   ä&rQ S   Q   ä &z88  € 5   ä $r<U   ÿŽ ä$r BU   ÿŽ ä $rQ@Q   ÿŽ Ä $r[[8   ÿŽ â Er        â $rY0   ÿŽ â Er          â $r8?8   ÿŽ ä 0r9ÿ9    A   â x   ÿàÿ â 0rTÿT    A   â Er6 6       â $r"A"   ÿŽ â Er> Q       â $r7@7   ÿŽ â Er: :       â 0rWÿW    A   â ErD D       â #r9   ÿ   âEr& &       â #r 9    ÿ   â Er        Ä Er[ [       ä Er5 8       ä |    pbô â #rCT6   C   â#rET:   E   â#r$T>   G   âEr" "       â #rNOK   N   â Er7 7       â #rTTD   ;   â #r&9&   ÿ   â#r9   ÿ   â #r[W[      â#r W5       â  r1HE     A  â rK%N     A  â #r3W"   &   â  rT%T     A  â #rW7      â  r[H[     A  â  r5%      A  â #r12C   1  â#rK'J   K  â#r%'$   T  â#r223   [  â #r5'   5  â x@  ÿà÷ â !rK       â !r1       â !r%       â !r2       â !r5       â x    ÿàÿ Ä x    ÿàÿ ä x    ÿàÿ ä x    ÿàÿ ä x    ÿàÿ ä rÿ   ÿäÿ â D)       À â Gy   éÿÿÿÿƒ ê Ay        € êÏ Ey  Ð     € â G‰  °     € ê $x @  Ž Ê ˆs  ûÿ H   è ˆs €ûÿ H   è ˆs  üÿ H   è ˆs 
€üÿ H   è ˆs 	 ýÿ H   è ˆs €ýÿ H   è ˆs  þÿ H   è ˆs €þÿ H   è ˆs  ÿÿ H   è ˆs €ÿÿ H   ä Ay        € ê {          ì M        € ê „y      H   b z&  r      Æ „y €   H   ¢ r &&   ÿŽ Æ „y    H   è „y €  H   ( „y    H   h „y €  H   ( „y    H   ( „y €  H   ( „y €  H   è „y    H   â !r         â/ !r       ÈO ‰       " !r       Æ� ‰      b !r
       â!r	       È‰      ¢ !r       Æ ‰      â !r       Æ ‰
      " !r       â !r        ä y      %   " !r       Æ ‰      b !r       â/ !r       ä ‰       b !r       ÆO ‰      ¢ !r       Æ� ‰      â !r

       Æ‰      ( ‰      " !r       Æ‰      b !r         Æ/ ‰	      b !r       ÆO ‰      ¢ !r       Æ� ‰
      â !r       Æ‰  €    " !r       Æ ‰      ( ‰      " !r       Æ‰      b !r		       Æ/ ‰      b !r       ÆO ‰      ¢ !r
       â� !r        ä‰ €    â !r       Æ ‰	 €    " !r       Æ ‰ €    " !r       Æ‰ €    b x   ÿ   â !r       ä/ ‰ €    b x
   ÿàÿ â $z)) x  Ž â !r       âO ‰  @    ¢ z 
 r  p`ò â $z)+ ~  )Ž â z  r  p`ð â ‰ €    ¢ !r
       â� x    p"ò Ä ‰ €    â !r	       â r ÿ   p"p ä ‰ €    " !r       âr&)   ÿ† ä‰ €    " !r       âr&)   ÿŠ â !r       â/ ‰
 @    b rÿÿ   ÿäÿ â !r         ÄO ‰ @    ¢ )    ÿàõ â !r       ä ‰       b ÿÿ   ÿä â !r       ä� ‰ @    â z j  ÿ† â !r       Ä ‰ @    " 
 j  ÿ„ â !r       ä‰ @    " z k  � ä 
 k   â ‰ @    " !r
       â/ 
&   ÿàÿ Ä ‰ @    b !r       âO r"ÿÿ   ÿäÿ ä ‰ @    ¢ !r#        â
	) r  ÿàù ä ‰ @    b !r       â� z j  ÿŠ â !r       â ‰       "     ÿàÿ â !r       Ä‰      â !)
   ÿàõ â !r       ä ‰      " x        â !r       ä/ ‰      b $ÿÿ   ÿä â %v
) j   Ž â !r       âO ‰      ¢ z k  "� â !r       Ä‰      b )   ÿà÷ ä 
	 j  ÿˆ â ‰      b "ÿÿ   ÿä â !r%        â ! j  ÿ„ â ‰      "  ÿÿ   ÿäÿ â !r       â� 
		 k  $ â ‰      â  j  ÿ† â !r       â! k  " â !r       â/  k   � â †	 
#    â !r       ÆO † 
%   â !r       Æ†	     â !r       Æ †     â !r       Æ †	     â !r       Æ� †    è †	     è †     è †	     â M™        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      "   " ¹z  l      â Ey      € â ™x?    â y      !   b ¹z  F   
   â r ÿ       ä y      %   ¢ r ÿ       ä r! ÿ       â y	      &   â r ÿ       Ä r% ÿ       ä r ÿ       ä r ÿ       ä r# ÿ       ä r ÿ   pRð är ÿ   pRò Ä r   ÿ(Ž È/ xÿ     ä x   ÿ   äO r   ÿ � È xÿ     È |    pbô Ú G)       € ê y      '   "Ž xðÿÿÿÿÀŽ â $z	 w  ÿŽ â r ÿ       ä r  €ÿàÿ ä r ÿ       ä xÿ     ä x   ÿàÿ Ä r   ÿ� ä x   ÿ   ä z  X      ä z  Y      ä r ÿ       ä z q  ÿàÿ â 'z  y   Ž È 'z	 s   Ž â zÿ z     Æ $z  p  Ž â zÿ t    â $z |  Ž ä$z |   Ž â x ÿ     â $z
 }  ÿŽ â xüÿÿÿÿÀŽ â $z v  Ž âz  q      â $z v  Ž â r  €ÿàÿ â $z p  Ž â x    ÿ   â $x   Ž âr   ÿŽ Ä r    ÿŽ ä r
   àµ â %x
�   ÿŽ â xÿ     ä x   ÿ   â %x�   ÿŽ â x     äx   ÿ   Ä xÿ     äx$
   øŽ ä xÿ     ä r 
   ÿàù är"   ÿàû ä x   øŽ Ä rÿÿ   d~ ä x$$   ÿüŽ ä r&   ÿä är(   ÿäÿ ä x   ÿüŽ ä r,$   ÿàõ È r-   ÿä ä r.   ÿàõ Æ �yB,øÿÿ • ¢  r/   ÿä ä r    ÿàõ â �y@,    • â  xI $       ä rÿ   ÿä â �y?,üÿÿ • " r2    ÿà÷ Æ �y<.   • b $x$   ÿŽ â r*"   ÿàù ä �yA.üÿÿ • " %v,  Z  I Ž â r3&   ÿäÿ ä�y6.    • " r+(   ÿä ä r0,   ÿàõ â �y2   ™ " r--   ÿàÿ Æ �y*   ™ " r1-   ÿä Æ �y 2    ™ " r.   ÿàõ Æ/ �y*    ™ ( �y)0   ™ " r2ÿ   ÿä Æ �y'0(   ™ ( �y40   ™ ( �y;,    • " %v.. Z  I Ž Æ �y708   ™ " $x+2$   ÿŽ â/ r*.   ÿàõ ä �y=,$   • $ r//+   ÿàÿ È r+/   ÿä â �y>.    • ( �y9*   ™ ( �y8*(   ™ ( �y5*   ™ " r2   ÿàù Æ �y:*8   ™ " r3ÿ   ÿä ä r0
   ÿàõ â�yC.$   • " r,   ÿà÷ â/ $x33$   ÿŽ â r1   ÿä â %v*2 Z  I Ž â r-   ÿäÿ Æ �y00    ™ " r.*   ÿàõ Æ �yD,    ™ " r3+3   ÿàÿ È r/3   ÿä Ê �yF.   ™ ( �yG.(   ™ ( �yE.   ™ " r2 *       ä r,   ÿàõ â �yH.8   ™ & r*ÿ   ÿä â �yK2    • " %v,, Z  I Ž Æ �yL2$   • " $x+*$   ÿŽ â r*,   ÿàõ È r--+   ÿàÿ È r+-   ÿä â �y3,    • ( �y1*   ™ (# �yJ*   ™ ( �yI*(   ™ ( �y2*8   ™ ( �yM,$   • " x    pBô Ú ". ÿ       ä ", ÿ       ä xOBÀÀ  ÿÀŽ äOxNB??  ÿÀŽ ä x/@  ÿÀŽ ä� x*ÿ   @  Æ/ '(NO   @. Ž â x.ÿ   <  äx+*  ÿÀŽ ä xO?ÀÀ  ÿÀŽ äx@A??  ÿÀŽ ä x/.  ÿÀŽ ä "* ÿ       Ä xAAÀÀ  ÿÀŽ ä x-<  ÿÀŽ ä xQ6ÀÀ  ÿÀŽ ä". ÿ       ä xB???  ÿÀŽ ä x?6??  ÿÀŽ â '(BO   @* Ž â x*ÿÀŽ Ä x6ÿ     â '(@A   @, Ž â x,ÿÀŽ Æ '(?Q   @. Ž â x.ÿ     ä x+ ÿÀŽ â &z)  € ÿ   â xAÿ      â &r *)   ÿ   âx-ÿÀŽ â &r),)   ÿ   â x..ÿÀŽ ä x66ÿÀŽ Ä x/ÿ     â &rP.'   ÿ   âxAAÿÀŽ â &zR'  € ÿ   â x//ÿÀŽ â &rV6'   ÿ   â x'N?   ÿÀŽ â &r<+4       ä&z4  €    ä &rT-4   )   â x4B?   ÿÀŽ â 0r ÿ;    A   â x)@?   ÿÀŽ Ä x;??   ÿÀŽ â &rOA7   P   âxNÿ   N  â &z7  € R   â xBÿ   B  â &rV/7   V   â x7ÿ   @  â $r<<'   ÿŽ â x?ÿ   ?  â $r4   ÿŽ ä $rTT)   ÿŽ Ä $rP;   ÿŽ â Er        â $r@ON   ÿŽ â ErO <       â $rQB   ÿŽ ä$rVV7   ÿŽ ä $rR?   ÿŽ â Er T       ä Er P       â 0r=ÿ=    A   â ErP @       â #r< O   ÿ   âErQ Q       â #r@    ÿ   â ErS V       Ä ErU R       â #rR    ÿ   â#r     ÿ   â #r<=P   <   â#rO=Q   @   â#r@=S   R   â &zP9  € ÿ   â #r==U       â &r *9   ÿ   ä&rR,9   ÿ   Ä &r.8   ÿ   ä&z8  € ÿ   ä &r968   ÿ   ä &zP5  € P   ä &r8+5       ä &rA:      Ä &z:  €    ä &r:/:   9   ä $r4P   ÿŽ ä &rR-5   R   ä $r8'8   ÿŽ â Er        â $r:7:   ÿŽ Ä 0r ÿ>    A   ä $rR)R   ÿŽ ä $rP;P   ÿŽ â Er5 8       â $r9N   ÿŽ ä $r>B   ÿŽ âErQ :       â $rS?   ÿŽ â Er R       â #r:    ÿ   âEr P       â 0rCÿC    A   â Er9 9       â #r8 5   ÿ   âEr >       â #rP    ÿ   â#rR    ÿ   â #r C9   8   â0r8ÿ0  0 A   â #rC   :   â 0r:ÿD  0 A   Ä 0r5ÿ0    A   â  r08     A  â 0r9ÿD    A   â  r==:     A  â  rOO8     A  â &z>F  € ÿ   â #r5    0  â #r=@9   =  â #r<<5   O  â &zG  € ÿ   ä &r *F   ÿ   Ä &z>E  € >   â ErS S       â !r       â !r=       â !r#<#       â &rF,F   ÿ   ä &r.G   ÿ   ä &z=H  €    ä &r<+E       Ä $r4>   ÿŽ â #rPCQ   P   â#rSCS   R   â &rF-E   F   ä &rAH      ä $r@;>   ÿŽ ä $r'<   ÿŽ ä $r>B=   ÿŽ Ä $rC?=   ÿŽ â Er=        â 0r ÿK    A   ä $rF)F   ÿŽ â Er        â $r<N   ÿŽ â Er> >       â 0rLÿL    A   â #r= =   ÿ   â Er F       â &rG6G   ÿ   â Er @       â #r    ÿ   â Er< <       â #r=L>   =   â &rH/H   G   â #r    ÿ   â#r    ÿ   â #r<L<      â  r 8=     A  â $rH7H   ÿŽ â Er> C       â &r,,1   ÿ   â#r=5<      â &r *1   ÿ   Ä &z11  € ÿ   â ErH H       â #rL>      â &r +J       ä&r,-J   ,   ä &z1J  € 1   ä &r..I   ÿ   ä&r66I   ÿ   Ä &zII  € ÿ   â #rHLH      â  r*:     A  â$r'    ÿŽ â  r0:S     A  â $r),   ÿŽ ä $r441   ÿŽ ä&zI2  € I   ä $r;1   ÿŽ â #r09P   0  â &rAA2   .   âEr        â &r6/2   6   â Er' 4       â $rBBI   ÿŽ â Er        â $r??I   ÿŽ â Er        â 0r ÿ3    A   â x   ÿàÿ â !r%0%       â $rANA   ÿŽ â ErB B       â $r676   ÿŽ â Er? ?       â 0rMÿM    A   â #r. '   ÿ   â#r,    ÿ   â#r0    ÿ   â#r     ÿ   â |    pbô â #rMB   .   âErA A       â #r?M?       â Er 6       â  r88     A  Ä #r,MA   ,   â r::?     A  â #r0M   0   â #r*9H   *  â #r,5,   8  â x€  ÿà÷ â #r990   :  â !r=       â !r!*!       â !r,       â !r9       â x@   ÿàÿ Ä x@   ÿàÿ ä x@   ÿàÿ ä x@   ÿàÿ ä rÿ   ÿäÿ â D)       À â Gy  �ìÿÿÿÿƒ ê Ay        € ê� Ey  °     € â G‰  �     € ê r   ÿ@Ž Ê ˆs # üÿ H   è ˆs €üÿ H   è ˆs  ýÿ H   è ˆs %€ýÿ H   è ˆs  þÿ H   è ˆs !€þÿ H   è ˆs  ÿÿ H   è ˆs €ÿÿ H   ä Ay        € ê {          ì M        € ê „y      H   b z  r  p`ð Æ „y €  H   ¢ r ÿ   p"p Æ „y €   H   è „y €  H   ( „y    H   h „y €	  H   ( „y    H   ( „y €  H   ( „y    H   ( „y    H   ( „y
 €  H   " !r  #       Æ/ „y    H   b !r%       ÆO „y    H   ¢ !r       Æ� „y €  H   â !r       Æ„y    H   " !r         Æ„y €  H   b !r       Æ „y)    H   b !r       Æ „y  	  H   b !r       Æ „y  
  H   " !r         Æ „y' €
  H   " !r       Æ „y+    H   " !r

!       Æ „y# €  H   â !r       Æ/ „y €  H   â !r       ÆO „y €  H   â !r       â� !r       ä‰       b !r

       Æ‰      ¢ !r)       Æ ‰      â !r       â !r       È ‰      " !r

'       Æ ‰      " !r+       Æ ‰
      b !r        â/ !r##       ä ‰      b !r       âO !r#       ä ‰       ¢ !r       â� !r       Ä ‰      â !r       Æ ‰      " !r       Æ‰      " !r
       Æ‰      b !r       Æ/ ‰
      b !r         ÆO ‰      ¢ !r       Æ� ‰      è ‰  €    â !r       â !r       ä‰ €    " !r       Æ‰ €    " !r
       Æ/ ‰ €    b !r       ÆO ‰ €    ¢ !r       Æ� ‰ €    â !r        Æ ‰ €    b x    ÿàÿ Æ ‰      b !r       â z   r  p`ò ä y      '   " !r
       âx    p"ò ä ‰ @    " !r       â/ !r       ÄO ‰ @    b $z	 x  Ž â !r       ä� ‰
 @    ¢ !r       Æ‰ @    â !r       Æ ‰ @    b $z  ~  Ž Æ ‰ @    " z  r      â !r       ä‰ @    " 
  r  ÿàõ ä r   ÿŽ â ‰ €    " !r	       â/ ÿÿ   ÿä Ä ‰      b 
 j  ÿ„ â !r
       âO r    ÿ† â !r       â� 
 k   â ‰	      ¢    ÿàÿ â !r       ârÿÿ   ÿäÿ â ‰      â z j  ÿ† â!r       â     ÿàõ â ‰      " z k  � â !r       â    ÿà÷ â ‰      " ÿÿ   ÿä â !r       â  j  ÿ„ â ‰      b ÿÿ   ÿäÿ â !r       â/ x        â ‰ @    b 

 j  ÿ† Æ ‰      b !r	       âO %v  j   Ž â 	 k   â !r       â� 
 k  � ä †	     â !r       Æ †    â !r       Æ†	     â !r       Æ†     â !r       Æ/ †	     â !r       Æ †    è ‰      ( †	 
    â M™        € ê x   ÿàÿ âO !r       Æ r    ÿàñ È r ÿÿ   ÿä  ä z j  ÿ€ È z k     Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      "   " ¹z  l      â Ey  0    € â ™x?    â y      !   b ¹z  F   
   â r ÿ       ä y      %   ¢ r ÿ       ä r& ÿ       â y	      &   â r* ÿ       Ä r( ÿ       ä r ÿ       ä r ÿ   pRð är ÿ   pRò är   ÿ(Ž È/ x ÿ     ä x   ÿ   äO r     ÿ � â $À   Ž Æ xÿ      È |    pbô Ú G)  Ð    € ê y      '   "Ž x  ðÿÿÿÿÀŽ â $z p  Ž âr ÿ       â $z	 w  ÿŽ â r   €ÿàÿ ä r ÿ       ä xÿ     ä x   ÿàÿ Ä r   ÿ� ä x   ÿ   ä z#  X      ä z  Y      ä r ÿ       ä z q  ÿàÿ â 'z  y   Ž È 'z	 s   Ž â z ÿ z     È zÿ t    â $z  |  Ž âxÿ     â $z  |  Ž â xüÿÿÿÿÀŽ â $z  }  ÿŽ â x   ÿ   â $z v  Ž âr  €ÿàÿ â $z v  Ž â r    àµ â $z p  Ž â r   ÿŽ â %x
�   ÿŽ â z   q      Ä x   ÿ   â %x�   ÿŽ â xÿ     ä xÿ     äx 
   øŽ ä xÿ     ä r
   ÿàù Ä r   ÿàû ä x   øŽ ä r    ÿŽ ä x     äx!   ÿ   ä rÿÿ   d~ Ä x     ÿüŽ ä r"   ÿä är$   ÿäÿ ä x   ÿüŽ ä r#    ÿàõ È r   ÿä ä r.#   ÿàõ Æ �y5øÿÿ • ¢  r/   ÿä ä r   ÿàõ â �y8    • â  xA $       ä r ÿ   ÿä â �y3.   • " r,#   ÿà÷ â%v Z  A Ž â r#   ÿàù â �y6üÿÿ • b r-"   ÿäÿ Æ �y4.üÿÿ • b r$   ÿä Æ �y%.    • b $x $   ÿŽ â r!   ÿàõ ä �y ,   ™ d r   ÿàÿ ä �y',    ™ d r   ÿä Ä �y)   ™ h �y+    ™ h �y1   ™ h �y0(   ™ b r,   ÿàõ Æ �y2   ™ h �y78   ™ b r.ÿ   ÿä Æ/ �y<    • b %v,, Z  A Ž Æ �y>$   • b $x.$   ÿŽ â r!,   ÿàõ È r--   ÿàÿ È r-   ÿä â �y?,    • h �y9   ™ h �y;   ™ b r   ÿàù Æ/ �y:(   ™ b rBÿ   ÿä Æ �y=8   ™ b r.
#   ÿàõ â %v Z  A Ž â r#   ÿà÷ â �y@,$   • b r/   ÿä â $xB$   ÿŽ â/ r!   ÿàõ Æ �y..    ™ b r   ÿàÿ ä r   ÿäÿ ä r   ÿä â �yD$   • h �y,   ™ h �yB    ™ h �yA   ™ h	 �y-(   ™ h	 �yC8   ™ h	 �y/    • b# x    pBô Ú " ÿ       ä " ÿ       ä/ x   ÿàÿ ä x##€  ÿà÷ ä x@   ÿàÿ ä x@   ÿàÿ Ä x@   ÿàÿ ä rÿ   ÿäÿ ä xE5??  ÿÀŽ äOx55ÀÀ  ÿÀŽ ä x8  ÿÀŽ ä� xÿ   3  Æ'(E5   @ Ž â x56??  ÿÀŽ äxG6ÀÀ  ÿÀŽ ä x64??  ÿÀŽ äxI4ÀÀ  ÿÀŽ ä x4%??  ÿÀŽ ä x  ÿÀŽ Ä x%%ÀÀ  ÿÀŽ ä " ÿ       ä xÿ   8  ä x8'ÿÀŽ â '(4%   @ Ž â x  ÿÀŽ ä " ÿ       Ä x%ÿ      ä xLÿ   '  ä x3  ÿÀŽ ä x ÿÀŽ ä x'ÿ   )  ä xF)ÿÀŽ â '(5G   @ Ž â x%%ÿÀŽ Ä xJÿ   +  â '(6I   @ Ž â x''ÿÀŽ ä xH+ÿÀŽ â &r1   ÿ   âxLLÿÀŽ â &z1  € ÿ   â xJJÿÀŽ â &r1F1   ÿ   ä &r %0   ÿ   Ä &z0  € ÿ   ä &r0'0   ÿ   ä &r82      ä&z2  €    ä &rH2   1   ä &r2L7       Ä &z7  €    â x)E?   ÿÀŽ â &r7J7   0   â x+6?   ÿÀŽ ä x05?   ÿÀŽ ä x14?   ÿÀŽ â 0r ÿ<    A   â x6ÿ   6  ä x5ÿ   5  â $r3)   ÿŽ â xEÿ   E  â $r+   ÿŽ â x4ÿ   4  â $r<0   ÿŽ Ä $r1   ÿŽ ä $r776   ÿŽ ä $r5   ÿŽ â Er3 3       â $r22E   ÿŽ â Er <       â $r4   ÿŽ â Er        ä Er        â 0r>ÿ>    A   â ErG 7       â #r3 3   ÿ   â ErK 2       Ä ErI        â #r    ÿ   âEr7        â #r    ÿ   â#r     ÿ   â #r>K   3   â#r>I      â#r>G      â#r>>7       â &r 9   ÿ   ä &z9  € ÿ   Ä &r9F9   ÿ   ä &r8;       ä&z ;  €    ä &r2%:   ÿ   ä &rH;   9   ä &z<:  € ÿ   Ä &r3':   ÿ   ä &r:L=   2   ä $r)   ÿŽ ä $r;+   ÿŽ ä &z2=  € <   ä $r0    ÿŽ Ä $r<1    ÿŽ ä $r9E:   ÿŽ â Er:        â 0r?ÿ?    A   â Er ;       ä Er         â $r752   ÿŽ â Er <       â &r3J=   3   â Er< 9       â #r9?:   ÿ   â#r=?   ÿ   â#r;?    ÿ   â#r??   ÿ   â 0r@ÿ@    A   â Er 7       â $r42   ÿŽ Ä 0r2ÿ.  0 A   ä$r363   ÿŽ â #r;@   ;   â 0r.ÿ.    A   â Er        â  r2     A  â &r,   ÿ   ä&rFF,   ÿ   â Er  3       â &z,,  € ÿ   â #r.     â 0rÿB  0 A   â #r@   ?   â &z,A  € ,   â !r       â &r''-   ÿ   Ä &z-  € ÿ   â #r @    =   â 0rBÿB    A   â  r     A  â &r8A      ä&rFHA   F   ä &r%%-   ÿ   ä &r'JC   '   Ä &zC  €    ä $r00,   ÿŽ ä$r1,   ÿŽ â #rB      â $r )   ÿŽ â  r>>     A  â $rF+F   ÿŽ â Er0 0       â &rLLC   %   â Er        â $r'6'   ÿŽ Ä $r55   ÿŽ ä$r44   ÿŽ ä 0r/ÿ/    A   â #rB   >  â $rLEL   ÿŽ â Er          â 0rDÿD    A   â Er F       â #r%/0   ÿ   â Er5 5       Ä Er '       â #r'/   ÿ   â Er4 4       â !r((       â |    pbô â #r/    ÿ   âErL L       â #r//   ÿ   â #r%D5   %   â#r4D4   '   â #r3@<   9   â  r;2;     A  â #rDL      â#r/D   /   â  r%2%     A  â  r4     A  â #r3.3   ;  Ä #r.   %  â #rB/     â !r*3*       â !r&&       â !r       â !r       â G©  Pðÿÿÿÿƒ ê Ay        € ê� ˆ  ýÿ H   è ˆ (€ýÿ H   è ˆ * þÿ H   è ˆ &€þÿ H   è ˆ  ÿÿ H   è ˆ €ÿÿ H   è {          ì M        € ê „y     H   " z  r  p`ð Æ „y     H   b r ÿ   p"p Æ „y €   H   ¨ „y    H   è „y    H   ( „y €  H   h „y €  H   h „y    H   h „y    H   h „y €  H   h „y €  H   b !r       Æ „y    H   " !r         Æ/ „y €  H   b !r(       ÆO „y    H   ¢ !r*       Æ� „y €  H   â !r         Æ„y    H   â !r&       Æ„y
 €  H   â !r       Æ „y €  H   â !r       â !r       ä ‰       " !r       â !r       â !r       È ‰      " !r       â/ !r       âO !r       È� ‰      b !r        âz  r      â !r       â ‰      ¢ !r

       Æ ‰      â !r

       Æ ‰       " !r       Æ ‰
      ( ‰      b !r       â/ !r       ÈO ‰      b !r       Æ� ‰      ¢ !r         Æ‰      â !r
       Æ ‰  €    " !r       Æ‰
      ( ‰ €    b !r       â/ !r       ÈO ‰ €    b !r       Æ� ‰ €    ¢ !r        Æ ‰ €    " !r

       Æ‰  @    â !r       Æ‰
 €    ( ‰ @    b !r       â/ !r       ÈO ‰ @    b !r       Æ y      '   " $z	 x  Ž â !r         ä� ‰ @    ¢ !r
       Æ‰ @    â x
   ÿàÿ â !r       ä‰ @    " z 
 r  p`ò Æ ‰       b x    p"ò Æ ‰
      b $z ~  Ž â !r       â/ !r       ÆO 
 r  ÿàõ â ‰      & 	   ÿàÿ â !r       â� ÿÿ   ÿä â ‰      b 
 j  ÿ„ â!r       âr   ÿ† â ‰      ¢ 
 k   â !r        â		   ÿàõ â ‰      â x        â !r
       â ÿÿ   ÿä Ä r ÿÿ   ÿäÿ â %v j   Ž â 	 j  ÿ„ ä z j  ÿ† ä 		 k   â !r       â z k   � â †	     â !r       Æ/ †    â !r       ÆO †	     â !r       Æ� †     è †	     â M™        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      "   " ¹z  l      â Ey  P    € â ™x?    â y      !   b ¹z  F   
   â $rÿÿ   ÿ Ž â r ÿ       â y      %   ¢ r ÿ       â $r#ÿÿ   ÿ Ž Ä y	      &   â r ÿ   pRð är ÿ   pRò är   ÿ(Ž È/ xÿ     ä x   ÿ   äO r    ÿ � ä    ÿ8Ž ä xÿ      È |    pbô Ú G)      € ê y      '   "Ž x ðÿÿÿÿÀŽ â $z p  Ž â r ÿ       â $z	 w  ÿŽ â r ÿ       â $x   
Ž â x
   ÿàÿ â $x   ÿ Ž â z!  X      ä xÿ     Ä z  Y      ä r   ÿ� ä r ÿ       ä xÿ     ä xüÿÿÿÿÀŽ ä z q  ÿàÿ â 'z  y   Ž â r  €ÿàÿ Æ 'z	 s   Ž â z ÿ z     ä xÿ     â$x   Ž â zÿ t    â $z  |  Ž âx     â $z  |  Ž â x   ÿ   â $z v  Ž âx   ÿ   â $z v  Ž â xÿ     â $z  }  ÿŽ Ä $z
 p  Ž ä %x
�   ÿŽ È $x   ÿ Ž â r
   ÿàù â %x�   ÿŽ Æ r    àµ â $x     â xÿ     äx
   øŽ är   ÿàû ä x   øŽ ä rÿÿ   d~ Ä x   ÿüŽ ä r"   ÿäÿ ä x   ÿüŽ ä r!   ÿàõ È r   ÿä ä r*!   ÿàõ Æ �y3øÿÿ • ¤  $x+    â r$   ÿàõ â �y/    • â  x; $       ä r ÿ   ÿä â �y.*   • " r&!   ÿà÷ â%v$$ Z  ; Ž â r!   ÿàù â �y5üÿÿ • b r'    ÿäÿ Æ �y(*    • b $x   " Æ �y1*üÿÿ • b $x $   ÿŽ â r$   ÿàõ ä �y &   ™ d r%%   ÿàÿ ä �y)&    ™ d r%   ÿä Ä �y   ™ h �y*    ™ h# �y,   ™ b r&   ÿàõ Æ �y+(   ™ b r0ÿ   ÿä Æ �y-   ™ h �y28   ™ b %v&& Z  ; Ž Æ �y7$    • b $x0$   ÿŽ â/ r&   ÿàõ ä �y9$$   • b $x''   Ž Ê r'   ÿä â �y:&    • h �y0   ™ h �y6   ™ h �y4(   ™ h �y88   ™ b r
!   ÿàõ Æ �y<&$   • b r   ÿä ä r$!   ÿàõ Æ/ �y;    ™ b r%   ÿä Ê �y=$    ™ b	 x    pBô Ú $"ÿÿ   ÿ Ž â " ÿ       ä x   ÿàÿ ä x!!€  ÿà÷ ä x@   ÿàÿ ä x@   ÿàÿ â $rÿÿ   Ž â x>3??  ÿÀŽ ÄO x33ÀÀ  ÿÀŽ ä x/  ÿÀŽ ä� x$ÿ   .  Æ'(>3   @ Ž â xÿ   /  ä x&5??  ÿÀŽ äx  ÿÀŽ ä x%$  ÿÀŽ ä x55ÀÀ  ÿÀŽ ä x3(ÀÀ  ÿÀŽ Ä " ÿ       ä "$ ÿ       ä x/(??  ÿÀŽ â '(&5   @ Ž â x'1??  ÿÀŽ äx11ÀÀ  ÿÀŽ â '(/3   @$ Ž â x.  ÿÀŽ Ä x(ÿ      ä x)ÿÀŽ ä x3ÿ   )  ä x% ÿÀŽ ä x)ÿÀŽ ä x$ÿ     â '('1   @ Ž â x((ÿÀŽ Ä x*ÿÀŽ â &r.%,   ÿ   âx1ÿ   *  â &z ,  € ÿ   â x$$ÿÀŽ â &r),   ÿ   â x33ÿÀŽ â &r(+   ÿ   âx11ÿÀŽ â &z5+  € ÿ   â x,'?   ÿÀŽ â &r+$+   ÿ   Ä &r.-   .   ä&z-  €     ä &r-      â x-ÿ   '  â &r32      âx'&?   ÿÀŽ â &z*2  € 5   â x5ÿ   &  â &r212   +   â x+>?   ÿÀŽ â $rA,   ÿŽ â x>ÿ   >  â $r'   ÿŽ â x&/?   ÿÀŽ â $r..+   ÿŽ Ä $r@>   ÿŽ â Er A       â $rB&   ÿŽ â Er        â $r?*5   ÿŽ ä 0r ÿ7    A   ä $r22-   ÿŽ â Er7 .       â 0r9ÿ9    A   â Er B       â #r    ÿ   âEr. ?       â #rB    ÿ   â Er 2       Ä x/ÿ   /  â #r.9.      â &r%0   ÿ   â #r%9   B   â #rD 7   ÿ   â &z0  € ÿ   â #r     ÿ   â $r*/   ÿŽ ä &r6      ä &r((4   ÿ   Ä &r$$4   ÿ   ä &z4  € ÿ   ä &r))0   ÿ   ä &z6  €    â Er@ @       â &r338   (   âEr        â &r$18   $   Ä &z88  €    ä &r)6   )   ä $r'   ÿŽ â #r*9@   D   â $r+   ÿŽ â #r99       â $r),)   ÿŽ â Er        â $r558   ÿŽ Ä 0r ÿ:    A   ä $r&   ÿŽ â Er'        â $r3>3   ÿŽ â Er )       â 0r<ÿ<    A   â Er 5       â $r//8   ÿŽ â #r)    ÿ   â$r$-$   ÿŽ â Er        â 0rÿ;  0 A   âEr3 3       â #r' '   ÿ   â #r<   )   â Er/ /       â #r    ÿ   âEr$ $       â #r    ÿ   â 0r;ÿ;    A   â |    pbô â #r <3   '   â r     A  â 0rÿ=  0 A   â#r</      â#r'<$      â #r$;      â 0r ÿ=    A   â  r.     A  â  r9     A  â  r     A  Ä #r**;     â #r%      â #r '     â !r$       â !r#*#       â !r       â !r       â G©  ðóÿÿÿÿƒ ê Ay        € ê� ˆ # þÿ H   è ˆ €þÿ H   è ˆ  ÿÿ H   è ˆ €ÿÿ H   è {          ì M        € ê „y      H   " z  r  p`ð Æ „y €   H   b r ÿ   p"p Æ „y    H   ¨ „y    H   è „y €  H   ( „y    H   h „y    H   h „y €  H   h „y    H   h „y €  H   h „y
 €  H   b !r  #       Æ „y €  H   " !r       â/ !r       âO !r         ä� !r       â!r       â!r         ä !r       Æ ‰       b !r       Æ ‰      ¢ !r       Æ ‰      â !r

       È !r
       â !r        È/ ‰       " !r       ÆO ‰      b !r       Æ� ‰      ¨ ‰      è y      '   " !r        â !r       È/ ‰      " !r       ÆO ‰ €    b !r       Æ� ‰  €    ¨ ‰
 €    â !r       â !r       È/ ‰ €    " x   ÿàÿ â !r        äO ‰  @    b z  r  p`ò â !r
       ä� ‰ @    ¢ x    p"ò Æ ‰ @    â !r       â !r         È/ ‰ @    " !r       ÆO ‰       b !r       â� $z	 x  Ž ä ‰
      ¢ x	        â $z ~  Ž ä‰      æ 
 r  ÿàõ È ÿÿ   ÿä ä 
 j  ÿ„ â !r       Æ 
 k   â !r        â/ %v j  	 Ž â !r	
       ÈO †	     â !r       Æ� † 	   è ‰      h  †	     â M™        € ê ¹z  r      â !r       â/ �x   ?àÿ Ì |   ÿàñ È rÿÿ   ÿä  ä z j  ÿ€ È z k    Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y       &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž ä $r	ÿÿ     Ž È %	 \   Ž Ê �	    ™ b  ÿ       â $ÿÿ   	 Ž â ¹z  l      â y      "   b ™x?    â '
	 m   Ž â Ey       € â y      !   b ‚ ÿ       â $rÿÿ   ÿ Ž â 
ÿ n    â y      '   ¨ $ÿÿ   
Ž Ä y      %   ä $
	 o  	Ž È $‚ÿÿ   	 Ž È 'Š	 s   Ž È $x    Ž Ê/ xÿ     È r   ÿ � È xÿ     È |    pbò ä Šÿ t    Ö G       € ê xðÿÿÿÿÀŽ âß $z		 w  ÿŽ â rÿ   ÿ3Ž â $z
 }  ÿŽ âr  €ÿàÿ â $rÿÿ   ÿ Ž â |   ÿàÿ â Ey P    € â xÿ     â 'z y   Ž â xÿ   ÿÀ‚ Ä r   ÿ� â $z p  ÿŽ â xÿøÿÿÿÿÀ€ â $rÿÿ   ÿ Ž â xüÿÿÿÿÀŽ ä xÿ     ä r  €ÿàÿ ä zÿ z    â $x   ÿ Ž âxÿ     â$x   Ž âx   ÿÀŽ Ä r		
   àµ äx
ÿ     ä x   ÿ   ä r
ÿÿ   
d~ ä xÿ     â G  ð    € ê $z v  Ž âr	   ÿ‚ â $tÿ�   ÿ Ž â x $       â $z |  Ž Ä $rÿÿ   
Ž  ä $x   Ž ä %v Z   Ž È %v X  Ž â r   ÿ„ Æ $x$   ÿŽ â r$   ÿàó â�y     ™ ¤  $x   Ž â r%   ÿäÿ  ä r   ÿ‚ ä r    â �y#$   ™ ä $rÿÿ   Ž  Ä �y$    ™ ( �y   • ¨ �y'   • ( �y&   • (	 �y(   ™ ( �y(   ™ ( �y)   ™ ( �y8   ™ ( �y!    • ( �y"$   • " x    pBò Ú $ÿÿ   ÿ Ž â  ÿ       ä/ x   ÿàÿ ä x%#ÿÀŽ ä� xÿ     äO x'??  ÿÀŽ äx  ÿÀŽ Ä x''ÀÀ  ÿÀŽ ä x&ÀÀ  ÿÀŽ ä x  ÿÀŽ â ''   @ Ž â x&??  ÿÀŽ ä x$ÿÀŽ â &z*(  € ÿ   â x#ÿ   #  â '   @ Ž â xÿ     Æ &z  € ÿ   â x##ÿÀŽ â &r%%(   ÿ   â xÿÀŽ â &z)  € *   â x?   ÿÀŽ â &z  €    â x?   ÿÀŽ â &r$$)   %   â xÿ     â &r##   ÿ   Ä $r   ÿŽ â xÿ     â $r$   ÿŽ ä &r   #   â Er        â $r   ÿŽ ä 0rÿ!    A   â Er        â $r   ÿŽ â Er        â 0r"ÿ"    A   â #r   ÿ   â#r   ÿ   â Er        â 0rÿ   0 A   â#r"      â 0r ÿ     A   Ä #r"      â  r     A  È #r      È !rÿ       ä Ay       € ê $vÿ X  ÿ Ž ä $vÿ Y  ÿ Ž â G‰   
    € ê $z |  Ž ä $x   ÿ Ž ä $z v  Ž âx@   ÿŽ È r   ÿàÿ Ê %x�   ÿŽ Ê r   ÿàñ ä x   þŽ Æ $x     ä r   ÿàó â $t%ÿ$   ÿ Ž â xÀÿÿÿÿàÿ ä r   ÿäÿ  ä r	   ÿàó â $x1   ÿ Ž â r   ÿàñ â �y*üÿÿ • ¢  rÿ
   ÿäÿ  â %v  Z  % Ž Ä �y'   • ä  $x$   ÿŽ â x#     â $x     â r1    ÿàñ â $x!!   Ž â �y)    • ( �y(   ™ b r#!   ÿä  Æ �y&    ™ ( �y   ™ ( �y   ™ ( �y(   ™ ( �y+8   ™ " r$	   ÿàñ Æ �y,     • $ $r"ÿÿ   
  ä �y- $   • " %v$$ Z  % Ž Æ �y/„  • " $x3"$   ÿŽ â r"1$   ÿàñ ä �y.|  • " $x%%   3Ž Æ �y0€  • $ r#%#   ÿä  ä �y1�  ™ ( �y    ™ ( �y2"   ™ ( �y!"   ™ " r   ÿàñ Æ �y3"(   ™ ( �y4"8   ™ " $x     Æ �y5$    • ( �y    ™ () �y€  ™ (	 �y$$$   • " x    pBð Ú $ÿÿ   ÿ Ž â x   ÿàÿ ä x 	  ÿàó ä x€   ÿàÿ ä rÿ   ÿäÿ  ä x#*ÀÀ  ÿÀŽ äOx"*??  ÿÀŽ Ä x'  ÿÀŽ Ê� '"#   @ Ž â xÿ   '  ä x(ÿÀŽ äx(ÿ   (  ä x)??  ÿÀŽ äx))ÀÀ  ÿÀŽ â &r   ÿ   â x  ÿÀŽ Ä  ÿ       â &z  € ÿ   â x#&ÿÀŽ ä x&ÿ   &  ä x((ÿÀŽ â ')   @ Ž â x&&ÿÀŽ Æ &z  €    ä &r#      â x"?   ÿÀŽ â &r((   ÿ   â x?   ÿÀŽ ä xÿ   "  â &r&&+   (   ä &z(  € ÿ   Ä $r   ÿŽ â x#ÿ     â $r   ÿŽ ä &z(+  € (   â Er        â $r&   ÿŽ ä 0r,ÿ,    A   â Er        â $r#(#   ÿŽ â Er        â 0r-ÿ-    A   â #r,   ÿ   â#r,   ÿ   â Er# #       Ä #r-      âxÿ   /  ä #r-#      â x#.ÀÀ  ÿÀŽ â $ÿÿ   ÿ Ž â x/  ÿÀŽ ä x  ÿÀŽ â $ÿÿ   ÿ Ž â x%0ÀÀ  ÿÀŽ Ä x.??  ÿÀŽ ä x"0??  ÿÀŽ ä x'1ÿÀŽ â '#   @ Ž È '"%   @ Ž â x ÿÀŽ Æ &r''2   ÿ   â x?   ÿÀŽ â &z22  € ÿ   â x1ÿ   1  â &r!   '   â x ÿ      â &z2!  € 2   â x1ÿÀŽ â $r   ÿŽ â x"?   ÿÀŽ â &z!3  € ÿ   â x ÿÀŽ â &r3   ÿ   â x"ÿ   "  â &z!4  € !   Ä $r2   ÿŽ â x ÿ     â &r4      ä $r!!"   ÿŽ â Er        â 0r5ÿ5    A   â Er        â $r     ÿŽ â Er! !       â 0r$ÿ$    A   â #r5   ÿ   â0rÿ  0 A   â|    pbð â #r55   ÿ   â Er          â 0rÿ  0 A   â#r$!      â 0rÿ    A   â  r     A  â 0rÿ    A   â #r $    5   â  r     A  â #r     Æ #r       â !r       È !r        â G‰  �öÿÿÿÿƒ ê Ay        € êß r ÿ   pRð är ÿ   pRò Ö ˆ €ÿÿ H   è {          ì M        € ê „y     H   " r ÿ   pRð â $z  x  Ž Ä „y €   H   h „y	    H   ¢ !r       È !r       È/ !r	       ÊO ‰      $ !r       Ê ‰      $ !r       Ê ‰	 €    $ !r		       Ê ‰	 @    $ !r	       Ê ‰      b  M	        € ê $tÿ   ÿ Ž â !r       â/ $z ~  Ž È %v j   Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y	      &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž È %	 \   Ž Ê �	    ™ ¢  $rÿÿ   ÿ Ž â Ey  Ð    € â $rÿÿ   ÿ Ž â y      "   b 'Š	 s   Ž Æ y      !   b '

	 m   Ž Æ y-      '   â $rÿÿ   ÿ Ž â 

ÿ n  
  ä rÿ   ÿü‚ ä/ r    ÿ(Ž È xÿ      È r    ÿ � ä Šÿ t    ä y      %   $ r	       âO G  À    € ê z ÿ ^  pPò ä� z ÿ `  pPô ä z ÿ _  Sò Ä z ÿ b  p ö ä z ÿ a   Sô È z ÿ c  0't Ê $- ~  ÿŽ â ÿ     â $ x  ÿŽ Æ    àë Æ $ª- ~  ÿŽ â ÿÿ   Äþ â $ª x  ÿŽ â ¨ÿ     ä¢   àÇ ä    ÿàý ä ¢   ÿàû ä ¢ÿÿ   „þ â $ÿÿ    â  ^  ÿ† Ä ¢ÿ   ÿäÿ ä ª b  ÿˆ ä  _  � â $rÿÿ   ÿ Ž â ª c   â $rÿÿ   ÿ Ž È �     h �©     d Ay        € ê� ¹z  l      â xÿ     â ™x?    â $ÿÿ   

Ž â Ey  ð
    € â $rÿÿ   ÿ Ž ä $

 o  	Ž â|    pbò â $‚
ÿÿ   	 Ž â r ÿ       Ö G  €
    € ê $r,ÿÿ   ÿ Ž â xðÿÿÿÿÀŽ â $z
 w  ÿŽ â r	        â 'z- y  , Ž â z ÿ `  pPð ä x	   ÿ   â $x    
Ž â zÿ z    â $z
- }  ÿŽ â z ÿ a   Sð Ä x ÿ     â $z |  	Ž â r

   ÿàõ ä r     ÿ� â $z v  Ž â x    pBò â $rÿÿ   ÿ â x üÿÿÿÿÀŽ â $z p  Ž â x ÿ      Æ %x�   ÿŽ È $x   
Ž ä $x    ÿ Ž ä$rÿÿ   ÿ Ž â r    ÿŽ â $rÿÿ   ÿ Ž â x   ÿÀŽ ä xÿ     â $x   ÿ Ž â z X  ÿàù È z X  ÿà÷ â $vÿ Y  ÿ â xÿ     È z Y  ÿäÿ ä r 
   àµ â $t!ÿ$   ÿ Ž Æ r    ÿd~ â %v   Z  ! Ž â r   ÿà÷ är   ÿàõ â $x $   ÿŽ â r   ÿäÿ Æ $x!!   Ž ä $x    â �y%   • ¨  �y&   • â  %x"    Ž Æ �y    ™ ( �y$   • ( �y"   ™ ( �y    ™ (# �y"(   ™ ( �y"8   ™ ( �y"   ™ "! z X  ÿàõ È z Y  ÿä â �y"     • ( �y#    ™ ( �y  $   • " $ÿÿ   ÿ Ž â/ x		   ÿàÿ Ä x'%??  ÿÀŽ äOx%%ÀÀ  ÿÀŽ ä x&  ÿÀŽ ä� x&ÿ   &  ä x) ÿÀŽ â''%   @ Ž â x&  ÿÀŽ Ä x%$ÀÀ  ÿÀŽ äx!$??  ÿÀŽ ä x ÿ      â '!%   @ Ž â x(ÿÀŽ ä x$ÿ     ä x ÿÀŽ â &r))   ÿ   Ä &z € ÿ   ä &r((   )   â x)'?   ÿÀŽ â &z& € ÿ   â x%!?   ÿÀŽ â &z €    â x $ÿÀŽ â $r()   ÿŽ â x(ÿ   !  â &r$   ÿ   Ä &z € &   ä $r%%   ÿŽ â x'ÿ   '  â &r$    $   ä $r((   ÿŽ â Er% %       â 0r ÿ"    A   â Er        â $r$$'   ÿŽ â Er( (       â 0r ÿ     A   â #r% %   ÿ   â0r'ÿ#  0 A   âEr$ $       â #r    ÿ   â #r" (   %   â 0r#ÿ#    A   Ä #r! $      â  r""'     A  â G‰  `    € ê z( `  àµ äz* `  àÙ ä z) a  d~ äz+ a  ÿ¤~ Æ �y%(    ¨  �y$*    è �y&*    ( �y'*    ( �y((     " z* `  ÿàõ È/ z+ a  ÿä Ê �y)*     $ x*%ÿÀŽ äO xÿ   $  Æ� &r**   ÿ   â  ÿ       ä x+&??  ÿÀŽ äx  ÿÀŽ ä x&&ÀÀ  ÿÀŽ Ê '+&   @ Ž â xÿ   %  ä x'??  ÿÀŽ äx%$  ÿÀŽ â $$ÿÿ   ÿ Ž â x''ÀÀ  ÿÀŽ ä xÿÀŽ ä x&ÿ   (  â ''   @$ Ž â x((ÿÀŽ Æ &r   ÿ   â x+?   ÿÀŽ â &r((   *   â x&&ÿÀŽ ä x?   ÿÀŽ â $r   ÿŽ â x$ÿ   +  â &r&      â xÿ     â $r(   ÿŽ â Er        â $r$$   ÿŽ Ä $r   ÿŽ â Er        â #r    ÿ   âEr$ $       â 0rÿ)  0 A   âEr        â #r    ÿ   ä #r $      â0r ÿ)    A   â #r       Ä  r     A  È #r      È !r       â | 	   pbô â #r"!#   "  â x€  ÿà÷ ä x@   ÿàÿ â !r"       ä $rÿÿ   Ž Ì G©  °÷ÿÿÿÿƒ ê Ay        € ê r ÿ   pRð âEy  �     € â r ÿ   pRò ä r        Ò G‰  P     € ê z ÿ `  pPð â $x    Ž Æ z ÿ a   Sð ä ˆs €ÿÿ H   ö ˆ    H   ä Ay        € ê {          ì M        € ê z ÿ `  pPð â „y      H   f z ÿ a   Sð Ú „‰ €   H   ( „‰	    H   ¢ !r        È/ !‚       È !‚	       âO G‰        € ê „y  €  H   ( „y €   H   h „y    H   ¨ „y    H   è „y	 €  H   " !r         â !r       Æ/ !r         âO !r       Æ� !r 	       ä‰      " z ÿ `  pPð ä r ÿ   pRò ä z ÿ a   Sð â !r       Ê ‰       $ !r         Ê ‰	  €    $ !r	 	       Ê ‰	 @    $ !r	       Ê ‰      $ !r       â G‰        € ê ‰       $ !r         Ê ‰       $ !r        Ê ‰ €    $ !r       Ê ‰ @    $ !r       Ê ‰      $ !r       ä M        € ê !r       âG‰  `    € ê $v ÿ h  ÿ Ž â !r       È x     pBð Ú G	  p    € ê x  ÿÿÿÿÿàÿ È x       € È x     ÿ   È ‚{   €     $ xÿ     È Iy  íÿÿÿÿƒ ê  x ;ª¸¿  A  Ì s          $ !t   €?    Ì s        $  r     A  È  r     A  â Gy  P    € ê  x '7=  A  â x*BL?  A  Æ #t   €?    È  r      A  â  x    ?  A  Ê s     $   $ !t  €?    È  r       A  È  r      A  â Gy  °    € ê x     pRð Ú G‰  à     € ê x     pRð Ú G	        € ê 	z  i    � ä 	z i    � Æ  x ;ª¸¿  A  â 	z i €  � Ê s        $ !t  €?    Ì s        $  r       A  È  r      A  â Gy  Ð     € ê  r     A  â Gy  °     € ê 	x   à@  � ä 	x  à@  � Æ  x #ÛÙ?  A  â 	x  àÀ  � Æ  x;ª¸¿  A  ä !t  €?    È s        $ !t  €?    Ì s        $  r       A  È  r      A  ä y      &   $ $z x  Ž ä $tÿ   ÿ Ž ä $z- ~  Ž È %v j   Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y       &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž ä $r	ÿÿ     Ž È %	 \   Ž Ê �	    ™ b  ÿ       â $ÿÿ   	 Ž â ¹z  l      â y      "   b ™x?    â '
	 m   Ž â Ey       € â y      !   b ‚ ÿ       â $rÿÿ   ÿ Ž â 
ÿ n    â y      '   ¨ $ÿÿ   
Ž Ä y      %   ä $
	 o  	Ž È $‚ÿÿ   	 Ž È 'Š	 s   Ž È $x    Ž Ê/ xÿ     È r   ÿ � È xÿ     È |    pbò ä Šÿ t    Ö G       € ê xðÿÿÿÿÀŽ âß $z		 w  ÿŽ â rÿ   ÿ3Ž â $z
 }  ÿŽ âr  €ÿàÿ â $rÿÿ   ÿ Ž â |   ÿàÿ â Ey P    € â xÿ     â 'z y   Ž â xÿ   ÿÀ‚ Ä r   ÿ� â $z p  ÿŽ â xÿøÿÿÿÿÀ€ â $rÿÿ   ÿ Ž â xüÿÿÿÿÀŽ ä xÿ     ä r  €ÿàÿ ä zÿ z    â $x   ÿ Ž âxÿ     â$x   Ž âx   ÿÀŽ Ä r		
   àµ äx
ÿ     ä x   ÿ   ä r
ÿÿ   
d~ ä xÿ     â G  ð    € ê $z v  Ž âr	   ÿ‚ â $tÿ�   ÿ Ž â x $       â $z |  Ž Ä $rÿÿ   
Ž  ä $x   Ž ä %v Z   Ž È %v X  Ž â r   ÿ„ Æ $x$   ÿŽ â r$   ÿàó â�y     ™ ¤  $x   Ž â r%   ÿäÿ  ä r   ÿ‚ ä r    â �y#$   ™ ä $rÿÿ   Ž  Ä �y$    ™ ( �y   • ¨ �y'   • ( �y&   • (	 �y(   ™ ( �y(   ™ ( �y)   ™ ( �y8   ™ ( �y!    • ( �y"$   • " x    pBò Ú $ÿÿ   ÿ Ž â  ÿ       ä/ x   ÿàÿ ä x%#ÿÀŽ ä� xÿ     äO x'??  ÿÀŽ äx  ÿÀŽ Ä x''ÀÀ  ÿÀŽ ä x&ÀÀ  ÿÀŽ ä x  ÿÀŽ â ''   @ Ž â x&??  ÿÀŽ ä x$ÿÀŽ â &z*(  € ÿ   â x#ÿ   #  â '   @ Ž â xÿ     Æ &z  € ÿ   â x##ÿÀŽ â &r%%(   ÿ   â xÿÀŽ â &z)  € *   â x?   ÿÀŽ â &z  €    â x?   ÿÀŽ â &r$$)   %   â xÿ     â &r##   ÿ   Ä $r   ÿŽ â xÿ     â $r$   ÿŽ ä &r   #   â Er        â $r   ÿŽ ä 0rÿ!    A   â Er        â $r   ÿŽ â Er        â 0r"ÿ"    A   â #r   ÿ   â#r   ÿ   â Er        â 0rÿ   0 A   â#r"      â 0r ÿ     A   Ä #r"      â  r     A  È #r      È !rÿ       ä Ay       € ê $vÿ X  ÿ Ž ä $vÿ Y  ÿ Ž â G‰   
    € ê $z |  Ž ä $x   ÿ Ž ä $z v  Ž âx@   ÿŽ È r   ÿàÿ Ê %x�   ÿŽ Ê r   ÿàñ ä x   þŽ Æ $x     ä r   ÿàó â $t%ÿ$   ÿ Ž â xÀÿÿÿÿàÿ ä r   ÿäÿ  ä r	   ÿàó â $x1   ÿ Ž â r   ÿàñ â �y*üÿÿ • ¢  rÿ
   ÿäÿ  â %v  Z  % Ž Ä �y'   • ä  $x$   ÿŽ â x#     â $x     â r1    ÿàñ â $x!!   Ž â �y)    • ( �y(   ™ b r#!   ÿä  Æ �y&    ™ ( �y   ™ ( �y   ™ ( �y(   ™ ( �y+8   ™ " r$	   ÿàñ Æ �y,     • $ $r"ÿÿ   
  ä �y- $   • " %v$$ Z  % Ž Æ �y/„  • " $x3"$   ÿŽ â r"1$   ÿàñ ä �y.|  • " $x%%   3Ž Æ �y0€  • $ r#%#   ÿä  ä �y1�  ™ ( �y    ™ ( �y2"   ™ ( �y!"   ™ " r   ÿàñ Æ �y3"(   ™ ( �y4"8   ™ " $x     Æ �y5$    • ( �y    ™ () �y€  ™ (	 �y$$$   • " x    pBð Ú $ÿÿ   ÿ Ž â x   ÿàÿ ä x 	  ÿàó ä x€   ÿàÿ ä rÿ   ÿäÿ  ä x#*ÀÀ  ÿÀŽ äOx"*??  ÿÀŽ Ä x'  ÿÀŽ Ê� '"#   @ Ž â xÿ   '  ä x(ÿÀŽ äx(ÿ   (  ä x)??  ÿÀŽ äx))ÀÀ  ÿÀŽ â &r   ÿ   â x  ÿÀŽ Ä  ÿ       â &z  € ÿ   â x#&ÿÀŽ ä x&ÿ   &  ä x((ÿÀŽ â ')   @ Ž â x&&ÿÀŽ Æ &z  €    ä &r#      â x"?   ÿÀŽ â &r((   ÿ   â x?   ÿÀŽ ä xÿ   "  â &r&&+   (   ä &z(  € ÿ   Ä $r   ÿŽ â x#ÿ     â $r   ÿŽ ä &z(+  € (   â Er        â $r&   ÿŽ ä 0r,ÿ,    A   â Er        â $r#(#   ÿŽ â Er        â 0r-ÿ-    A   â #r,   ÿ   â#r,   ÿ   â Er# #       Ä #r-      âxÿ   /  ä #r-#      â x#.ÀÀ  ÿÀŽ â $ÿÿ   ÿ Ž â x/  ÿÀŽ ä x  ÿÀŽ â $ÿÿ   ÿ Ž â x%0ÀÀ  ÿÀŽ Ä x.??  ÿÀŽ ä x"0??  ÿÀŽ ä x'1ÿÀŽ â '#   @ Ž È '"%   @ Ž â x ÿÀŽ Æ &r''2   ÿ   â x?   ÿÀŽ â &z22  € ÿ   â x1ÿ   1  â &r!   '   â x ÿ      â &z2!  € 2   â x1ÿÀŽ â $r   ÿŽ â x"?   ÿÀŽ â &z!3  € ÿ   â x ÿÀŽ â &r3   ÿ   â x"ÿ   "  â &z!4  € !   Ä $r2   ÿŽ â x ÿ     â &r4      ä $r!!"   ÿŽ â Er        â 0r5ÿ5    A   â Er        â $r     ÿŽ â Er! !       â 0r$ÿ$    A   â #r5   ÿ   â0rÿ  0 A   â|    pbð â #r55   ÿ   â Er          â 0rÿ  0 A   â#r$!      â 0rÿ    A   â  r     A  â 0rÿ    A   â #r $    5   â  r     A  â #r     Æ #r       â !r       È !r        â G‰  �öÿÿÿÿƒ ê Ay        € êß r ÿ   pRð är ÿ   pRò Ö ˆ €ÿÿ H   è {          ì M        € ê „y     H   " r ÿ   pRð â $z  x  Ž Ä „y €   H   h „y	    H   ¢ !r       È !r       È/ !r	       ÊO ‰      $ !r       Ê ‰      $ !r       Ê ‰	 €    $ !r		       Ê ‰	 @    $ !r	       Ê ‰      b  M	        € ê $tÿ   ÿ Ž â !r       â/ $z ~  Ž È %v j   Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y	      &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž È %	 \   Ž Ê �	    ™ ¢  $rÿÿ   ÿ Ž â Ey  Ð    € â $rÿÿ   ÿ Ž â y      "   b 'Š	 s   Ž Æ y      !   b '

	 m   Ž Æ y-      '   â $rÿÿ   ÿ Ž â 

ÿ n  
  ä rÿ   ÿü‚ ä/ r    ÿ(Ž È xÿ      È r    ÿ � ä Šÿ t    ä y      %   $ r	       âO G  À    € ê z ÿ ^  pPò ä� z ÿ `  pPô ä z ÿ _  Sò Ä z ÿ b  p ö ä z ÿ a   Sô È z ÿ c  0't Ê $- ~  ÿŽ â ÿ     â $ x  ÿŽ Æ    àë Æ $ª- ~  ÿŽ â ÿÿ   Äþ â $ª x  ÿŽ â ¨ÿ     ä¢   àÇ ä    ÿàý ä ¢   ÿàû ä ¢ÿÿ   „þ â $ÿÿ    â  ^  ÿ† Ä ¢ÿ   ÿäÿ ä ª b  ÿˆ ä  _  � â $rÿÿ   ÿ Ž â ª c   â $rÿÿ   ÿ Ž È �     h �©     d Ay        € ê� ¹z  l      â xÿ     â ™x?    â $ÿÿ   

Ž â Ey  ð
    € â $rÿÿ   ÿ Ž ä $

 o  	Ž â|    pbò â $‚
ÿÿ   	 Ž â r ÿ       Ö G  €
    € ê $r,ÿÿ   ÿ Ž â xðÿÿÿÿÀŽ â $z
 w  ÿŽ â r	        â 'z- y  , Ž â z ÿ `  pPð ä x	   ÿ   â $x    
Ž â zÿ z    â $z
- }  ÿŽ â z ÿ a   Sð Ä x ÿ     â $z |  	Ž â r

   ÿàõ ä r     ÿ� â $z v  Ž â x    pBò â $rÿÿ   ÿ â x üÿÿÿÿÀŽ â $z p  Ž â x ÿ      Æ %x�   ÿŽ È $x   
Ž ä $x    ÿ Ž ä$rÿÿ   ÿ Ž â r    ÿŽ â $rÿÿ   ÿ Ž â x   ÿÀŽ ä xÿ     â $x   ÿ Ž â z X  ÿàù È z X  ÿà÷ â $vÿ Y  ÿ â xÿ     È z Y  ÿäÿ ä r 
   àµ â $t!ÿ$   ÿ Ž Æ r    ÿd~ â %v   Z  ! Ž â r   ÿà÷ är   ÿàõ â $x $   ÿŽ â r   ÿäÿ Æ $x!!   Ž ä $x    â �y%   • ¨  �y&   • â  %x"    Ž Æ �y    ™ ( �y$   • ( �y"   ™ ( �y    ™ (# �y"(   ™ ( �y"8   ™ ( �y"   ™ "! z X  ÿàõ È z Y  ÿä â �y"     • ( �y#    ™ ( �y  $   • " $ÿÿ   ÿ Ž â/ x		   ÿàÿ Ä x'%??  ÿÀŽ äOx%%ÀÀ  ÿÀŽ ä x&  ÿÀŽ ä� x&ÿ   &  ä x) ÿÀŽ â''%   @ Ž â x&  ÿÀŽ Ä x%$ÀÀ  ÿÀŽ äx!$??  ÿÀŽ ä x ÿ      â '!%   @ Ž â x(ÿÀŽ ä x$ÿ     ä x ÿÀŽ â &r))   ÿ   Ä &z € ÿ   ä &r((   )   â x)'?   ÿÀŽ â &z& € ÿ   â x%!?   ÿÀŽ â &z €    â x $ÿÀŽ â $r()   ÿŽ â x(ÿ   !  â &r$   ÿ   Ä &z € &   ä $r%%   ÿŽ â x'ÿ   '  â &r$    $   ä $r((   ÿŽ â Er% %       â 0r ÿ"    A   â Er        â $r$$'   ÿŽ â Er( (       â 0r ÿ     A   â #r% %   ÿ   â0r'ÿ#  0 A   âEr$ $       â #r    ÿ   â #r" (   %   â 0r#ÿ#    A   Ä #r! $      â  r""'     A  â G‰  `    € ê z( `  àµ äz* `  àÙ ä z) a  d~ äz+ a  ÿ¤~ Æ �y%(    ¨  �y$*    è �y&*    ( �y'*    ( �y((     " z* `  ÿàõ È/ z+ a  ÿä Ê �y)*     $ x*%ÿÀŽ äO xÿ   $  Æ� &r**   ÿ   â  ÿ       ä x+&??  ÿÀŽ äx  ÿÀŽ ä x&&ÀÀ  ÿÀŽ Ê '+&   @ Ž â xÿ   %  ä x'??  ÿÀŽ äx%$  ÿÀŽ â $$ÿÿ   ÿ Ž â x''ÀÀ  ÿÀŽ ä xÿÀŽ ä x&ÿ   (  â ''   @$ Ž â x((ÿÀŽ Æ &r   ÿ   â x+?   ÿÀŽ â &r((   *   â x&&ÿÀŽ ä x?   ÿÀŽ â $r   ÿŽ â x$ÿ   +  â &r&      â xÿ     â $r(   ÿŽ â Er        â $r$$   ÿŽ Ä $r   ÿŽ â Er        â #r    ÿ   âEr$ $       â 0rÿ)  0 A   âEr        â #r    ÿ   ä #r $      â0r ÿ)    A   â #r       Ä  r     A  È #r      È !r       â | 	   pbô â #r"!#   "  â x€  ÿà÷ ä x@   ÿàÿ â !r"       ä $rÿÿ   Ž Ì G©  °÷ÿÿÿÿƒ ê Ay        € ê r ÿ   pRð âEy  �     € â r ÿ   pRò ä r        Ò G‰  P     € ê z ÿ `  pPð â $x    Ž Æ z ÿ a   Sð ä ˆs €ÿÿ H   ö ˆ    H   ä Ay        € ê {          ì M        € ê z ÿ `  pPð â „y      H   f z ÿ a   Sð Ú „‰ €   H   ( „‰	    H   ¢ !r        È/ !‚       È !‚	       âO G‰        € ê „y  €  H   ( „y €   H   h „y    H   ¨ „y    H   è „y	 €  H   " !r         â !r       Æ/ !r         âO !r       Æ� !r 	       ä‰      " z ÿ `  pPð ä r ÿ   pRò ä z ÿ a   Sð â !r       Ê ‰       $ !r         Ê ‰	  €    $ !r	 	       Ê ‰	 @    $ !r	       Ê ‰      $ !r       â G‰        € ê ‰       $ !r         Ê ‰       $ !r        Ê ‰ €    $ !r       Ê ‰ @    $ !r       Ê ‰      $ !r       ä M        € ê !r       âG‰  `    € ê $v ÿ h  ÿ Ž â !r       È x     pBð Ú G	  p    € ê x  ÿÿÿÿÿàÿ È x       € È x     ÿ   È ‚{   €     $ xÿ     È Iy  íÿÿÿÿƒ ê  x ;ª¸¿  A  Ì s          $ !t   €?    Ì s        $  r     A  È  r     A  â Gy  P    € ê  x '7=  A  â x*BL?  A  Æ #t   €?    È  r      A  â  x    ?  A  Ê s     $   $ !t  €?    È  r       A  È  r      A  â Gy  °    € ê x     pRð Ú G‰  à     € ê x     pRð Ú G	        € ê 	z  i    � ä 	z i    � Æ  x ;ª¸¿  A  â 	z i €  � Ê s        $ !t  €?    Ì s        $  r       A  È  r      A  â Gy  Ð     € ê  r     A  â Gy  °     € ê 	x   à@  � ä 	x  à@  � Æ  x #ÛÙ?  A  â 	x  àÀ  � Æ  x;ª¸¿  A  ä !t  €?    È s        $ !t  €?    Ì s        $  r       A  È  r      A  ä y      &   $ $z x  Ž ä $tÿ   ÿ Ž ä $z- ~  Ž È %v j   Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y       &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú         â $rÿÿ     Ž È % \   Ž Ê �	    ™ b ¹z  l      â Ey  p    € â ™x?    â y	      "   b r ÿ       Æ y      !   ¢         â $ÿÿ   ÿ Ž ä y      %   $ '
 m   Ž È $‚ÿÿ   ÿ Ž â 
ÿ n    È ÿ   ÿáÿ ä r 	ÿ   pRò Æ/ $
 o  Ž â r 	ÿ   pRô â$rÿÿ   ÿ Ž ä $x	    Ž âO ‚        È xÿ     â 'Š s   Ž Æ r
   ÿ � â $	€   Ž â x	   ÿ   ä xÿ   
  È |    pbö ä Šÿ t    ä x      ÿ  Ò G9  °    € ê y      '   " x
ðÿÿÿÿÀŽ â $z w  ÿŽ â r
 ÿ       â $vÿ Y  ÿ Ž â z  X      â $x   
Ž Ê xÿ     È r   ÿ� â 'z y  
 Ž Ê zÿ z    â $z	 p  Ž È $z |  Ž â xÿ     â $z }  ÿŽ â xüÿÿÿÿÀŽ â $z v  Ž ä$x   ÿ Ž â r
  €ÿàÿ â %x�   ÿŽ Æ r   à± â$x   
Ž âxÿ     â$rÿÿ   ÿ Ž â xÿ   
  ä x   ÿ   ä x   øŽ ä r   ÿàù Ä x   ÿ   ä rÿÿ   d~  ä x
     ä x   ÿüŽ ä r   ÿ ä r   ÿàñ ä r6   ÿà÷ Ä x7 $       â $x     â r   ÿàñ â $rÿÿ   Ž ä %v66 Z  7 Ž â �y:    • ¦  $x$   ÿŽ â �y9øÿÿ • â  r   ÿä  â $v=ÿ p  ÿ Ž â r"
6   ÿ€ â �y,üÿÿ • " r77   ÿàÿ â %x =�   Ž Ä �y5   ™ d $x#7     ä �y' øÿÿ • h �y+     • h �y3    ™ h �y*"   ™ h �y"   ™ h �y."(   ™ h �y-"8   ™ b %x=�   Ž Æ/ �y4 üÿÿ • h �y   ™ h �y1    ™ b r$   ÿàñ Æ �y26    • b r%   ÿä  Æ �y&6$   • h �y$    ™ b %x"=�    Ž Ê �y"øÿÿ • b %x(=�   Ž Æ �y0"    • h �y(   ™ h �y/"üÿÿ • h �y(    ™ b %x =�   $Ž È/ %x$=�   "Ž â�y     ™ h �y8$    • b %x=�   (Ž Æ �y7$üÿÿ • h �y;$øÿÿ • h �y6   ™ h �y<    ™ b %x(=�    Ž Ê �y$(    ™ b %x"=�   (Ž Ì �y""    ™ b x    pBð Ú  ÿ       ä/ x   ÿàÿ ä x€  ÿà÷ ä x@   ÿàÿ ä x ÿ   :  äO x!:  ÿÀŽ Ä x=9??  ÿÀŽ ä�x99ÀÀ  ÿÀŽ ä x   ÿÀŽ â $ ÿÿ   ÿ Ž â x%,??  ÿÀŽ äx,,ÀÀ  ÿÀŽ â '=9   @  Ž â x!5ÿÀŽ Äx)'ÀÀ  ÿÀŽ â'%,   @ Ž â x#'??  ÿÀŽ ä x+  ÿÀŽ ä x 3ÿÀŽ Æ '#)   @ Ž È &r!!*   ÿ   â x5ÿ   5  Æ &r)    !   â xÿ   +  â &z .  € ÿ   â x+ÿ   %  â &z*  € ÿ   â x3ÿ   3  â &z -  €     â x55ÿÀŽ â &z!  €    â x,%?   ÿÀŽ â $r +   ÿŽ â x(3ÿÀŽ â &r'5.   ÿ   Ä $r,!,   ÿŽ â x%4ÀÀ  ÿÀŽ â &r'(-   '   â x  ÿÀŽ ä Er3        â $ÿÿ   ÿ Ž â x(4??  ÿÀŽ ä Er+ ,       â '(%   @ Ž â x,=?   ÿÀŽ Ä xÿÀŽ ä xÿ     â $r)),   ÿŽ â x4ÿ   =  ä x,1ÿÀŽ ä x1ÿ   1  ä x%ÿÀŽ â $r''4   ÿŽ â x41ÿÀŽ â 0r2ÿ2    A   Ä &r%%.   ÿ   ä &r*   ÿ   â #r+2+   ÿ   â 0r&ÿ&    A   â Er) )       â &r%4-   %   â x4(?   ÿÀŽ â &r,      â x,#?   ÿÀŽ ä x:ÿ   #  â 0r#ÿ  0 A   â#r&3   +   â Er '       â #r)2)   ÿ   â x'ÿ   (  â $r4!4   ÿŽ â  r##     A  â $r,   ÿŽ â #r&   )   â 0rÿ    A   Ä $r( '   ÿŽ â Er' 4       â $r:%:   ÿŽ â Er        â #r#   #  â x%??  ÿÀŽ â#r'2'   ÿ   â Er( (       ä x0  ÿÀŽ ä xÀÀ  ÿÀŽ Ä  ÿ       â #r'&(   '   â xÿ   0  ä x(ÿ     â '%   @ Ž â xÿÀŽ ä x/??  ÿÀŽ äx  ÿÀŽ â $ÿÿ   ÿ Ž â x//ÀÀ  ÿÀŽ Ä xÿÀŽ ä xÿ     ä x)(ÿÀŽ â '/   @ Ž â xÿÀŽ Æ &r)).   ÿ   ä &r*   ÿ   â x%?   ÿÀŽ â &r)-   )   â x?   ÿÀŽ â &r      â x(ÿ   %  ä x%ÿ     â $r   ÿŽ Ä $r!   ÿŽ ä $r()(   ÿŽ â Er        â $r% %   ÿŽ â Er        â 0rÿ  0 A   â Er (       â #r2   ÿ   âEr %       â #r%2   ÿ   â  r'     A  â x'ÿ   8  â #r&      âx7??  ÿÀŽ â #r&   %   â x'  ÿÀŽ Ä x77ÀÀ  ÿÀŽ ä  ÿ       ä x)6ÿÀŽ Æ '7   @ Ž â x;??  ÿÀŽ äx98  ÿÀŽ â $8ÿÿ   ÿ Ž â x;;ÀÀ  ÿÀŽ ä x6ÿ   6  â &r))*   ÿ   â x<ÿÀŽ â ';   @8 Ž â x%6ÿÀŽ Ä x<ÿ   <  â &r   )   â x(?   ÿÀŽ â &r%%.   ÿ   â x<<ÿÀŽ ä x?   ÿÀŽ â $r(!(   ÿŽ â x'ÿ     â &r%<-   %   â xÿ     â $r   ÿŽ â Er (       â $r' '   ÿŽ Ä $r%   ÿŽ â Er        â 0rÿ$  0 A   â Er' '       â #r2   ÿ   âEr        â #r2   ÿ   â |    pbð â  r     A  â Er: :       â 0rÿ"  0 A   â #r2   ÿ   â #r&'      â#r&      â0rÿ    A   â #r&:      â 0rÿ$    A   â  r     A  â 0r!ÿ"    A   â #r     Ä #r     ä #r!     â !r#       â !r       â !r       â !r       â rÿ   ÿäÿ â G‰  `ñÿÿÿÿƒ ê Ay        € ê ˆ  þÿ H   è ˆ €þÿ H   è ˆ  ÿÿ H   è ˆ €ÿÿ H   è {          ì M)        € ê „y     H   " z 	 r  p`ô â $z   x  	Ž Ä „y €   H   b r ÿ   pVt Æ „y    H   ¨ „y    H   è „y €  H   („y    H   ( „y
    H   ( „y €  H   ( „y €  H   ( „y    H   ( „y €  H   " !r       Æ „y €  H   " !r       â/ !r       äO y      '   b !r       â� !r       â!r       ä !r
       â !r       È ‰      ¢ !r       â $z  ~   Ž â/ !r       Æ ‰      b !r       Æ ‰      â !r       Ê ‰      " !r       âO !r       â/ !r       È� ‰      h ‰      ¢ !r       Æ ‰      ( ‰      â !r       â/ !r       ÈO ‰
 €    b !r       Æ ‰ €    " !r       Æ� ‰ €    ¨ ‰ €    â !r

       â/ !r       È ‰
 @    " x	   ÿàÿ â !r       äO ‰ @    b x	   ÿàÿ â !r       ä� ‰ @    ¢ z  r  p`ö ä z  r  p`ò â ‰ @    â x    pVö Ä x    pVò  â !r
       â x
	   ÿàÿ ä x	        â !r       â/ ‰      " z 
 r  p`ð â !r       äO ‰      b x    pVp  â !r       Ä� ‰      ¨ ‰      â !r       â !r       â/ %v  j  	 Ž â !r       ÈO †©     â !r       Æ� †¹    è †‰    â M        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y	      &   " z ÿ \  pPò â ¹z  F   
   Æ z ÿ ]  Sò Ú         Ê %	 \   Ž Ê �    ™ ¢  r ÿ       â Ey      € ä y      "   d ' 	 m   Ž äy      %   ä 'š	 s   Ž â  ÿ n     â y      !   &  ÿ   ÿáÿ â y      '   ( $ o  	Ž â ’ 	       Ä r ÿ   pRð ä/ x   ÿ   ä� x    pDp  är   ÿàÿ är    ÿ(Ž ä z  r  pdp  Ä x      ÿ  â xÿ      ä šÿ t    È r		     € ÈO G	  °    € ê z ÿ `  pPò ä z ÿ ^  pPð ä z ÿ b  p ô ä z ÿ a  Sò ä z ÿ _   Sð ä z ÿ c   'ò Ö $
 ~  ÿŽ âÿ     â$š ~  ÿŽ â ˜ÿ     â $
	 x  ÿŽ â   àë â$š	 x  ÿŽ â ’	   àÇ ä    ÿàõ ä ÿ   ÿÄþ Ä ’		   ÿàû ä ’ÿ   ÿ„þ ä ÿ   ÿä ä 
 ^  ÿ„ ä ’ÿ   ÿäÿ ä š	 b  ÿ† Ä 
 _   ä x      ÿ  â š	 c  � Ê �	     h �™     d Ay        € ê r    ÿ � â ¹z  l      â $z	 }  ÿŽ â™x?    â xÿ     â $z w  ÿŽ â r
 ÿ       â Ey   !    € â r' ÿ       â x      ÿ  â |    pbð â 'z y  
 Ž â r	   ÿàó â x
      ÿ  â r	 ÿ       â x      ÿ  â rÿÿ   ÿäÿ  Î G	  p     € ê x	ðÿÿÿÿÀŽ â $z
 p  Ž âzÿ z    ä r	 	  €ÿàÿ ä x   ÿàÿ â $z |  Ž âx ÿ   	  â $z |  
Ž â x
   ÿàÿ â $z v  Ž âr  	   ÿ� â $z, v  Ž â x   ÿàÿ â $z.
 p  Ž âx üÿÿÿÿÀŽ â $z0 p  Ž âx ÿ      â $z2 p  Ž â r	  €ÿàÿ â %x,,�   ÿŽ â x    ÿ   âx
      ÿ  â r    ÿŽ â %x..�   ÿŽ â z ÿ `  pPð â x      ÿ  â x   ÿ   â %x22�   ÿŽ â x.   øŽ â x      ÿ  â x#ÿ     â %x00�   ÿŽ â r",   ÿàõ Ä r$.   ÿàù är&2   ÿàû ä x,   øŽ äx0   øŽ äx 2   øŽ ä x ÿ     Ä r(0   ÿàý ä x*"    ÿà÷ ä r-#   ÿä äx 	   pBò ä r/#   ÿä är!3#   ÿäÿ Ä z  `      ä z  a      ä z  X      ä z  Y      ä r	 ÿ       ä r' ÿ       Ä z ÿ a   Sð ä x   ÿ   ä xÿ     ä x   ÿüŽ ä x   ÿüŽ ä x   ÿüŽ Ä x     ÿüŽ ä x      ä r#1#   ÿä ä r%ÿ   ÿäÿ ä r:   àÅ ä r4   ÿà÷ ä x; $       Ä r    ÿ„~ ä r>"   ÿàõ â %v:: Z  ; Ž â r5-   ÿäÿ ä r?   ÿä â $x $   ÿŽ â r6:   ÿ„ â �yE4øÿÿ • ¦  r;;   ÿàÿ â �y@4    • è  �yD>   ™ " r7;   ÿä Æ �yC>    ™ " zI  p      ä r8*   ÿàõ â �y6   ™ $ %x<I�   4Ž âr9%   ÿä â �y 6(   ™ ( �y6   ™ " %x8I�   8Ž Æ �y+4üÿÿ • ( �y)68   ™ ( �yG<    • ( �yH<üÿÿ • ( �yA8ðÿÿ ™ ( �yK<øÿÿ • ( �yB8    ™ ( �yF:    • ( �yJ:$   • " r>,   ÿàõ È/ r?-   ÿä â %x6I�   <Ž È �yL>    ™ " %x<I�   8Ž Æ �yQ6øÿÿ • " %x4I�   >Ž Æ �yM6    • ( �yN<ðÿÿ ™ ( �yP<    ™ ( �yR4    ™ " %x8I�   6Ž Æ �yO6üÿÿ • " %x:I�   <Ž Æ �y>8øÿÿ • ( �y?8    • ( �yT:ðÿÿ ™ ( �yS8üÿÿ • ( �yU:    ™ " %x4I�   4Ž Ê/ �yV4    ™ " %x<I�   4Ž Ì �y<<    ™ " 6 ÿ       ä xIE??  ÿÀŽ äOxEEÀÀ  ÿÀŽ ä x7@  ÿÀŽ ä� xXDÿÀŽ äxDÿ   D  â 'IE   @6 Ž â x7CÿÀŽ Ä x4ÿ   @  ä/ xCÿ   C  ä x9DÿÀŽ â &rXX   ÿ   â x54  ÿÀŽ ä x4CÿÀŽ â &r99    ÿ   â x6I?   ÿÀŽ â &r77   X   â x8+??  ÿÀŽ Ä x++ÀÀ  ÿÀŽ â &r:4)   9   â 4 ÿ       â $r976   ÿŽ â x6ÿ   G  Æ '8+   @4 Ž â x76  ÿÀŽ ä x+HÀÀ  ÿÀŽ äx@AÿÀŽ ä 6 ÿ       ä x;K??  ÿÀŽ ä x=H??  ÿÀŽ Ä xKKÀÀ  ÿÀŽ ä x5G  ÿÀŽ â &r@@   ÿ   â xCBÿÀŽ â '=+   @6 Ž â x68?   ÿÀŽ Æ &z+  € ÿ   ä ';K   @4 Ž È &r4C   @   ä &z@   € ÿ   ä &z+  € +   â xAÿ   A  ä x5ÿ   8  â &z8)  € @   ä $r6+6   ÿŽ â xBÿ   B  â $r785   ÿŽ â xAAÿÀŽ â 0rFÿF    A   â Er5 6       Ä xBBÿÀŽ â &r6A    ÿ   â Er7 7       â 0rJÿJ    A   â xA;?   ÿÀŽ â #r5F5   ÿ   â &rBB)   6   â xIÿ   I  ä xCÿ   ;  â #r;J7   5   â x6=?   ÿÀŽ â $r54A   ÿŽ â x=ÿ   =  â $r::I   ÿŽ â Er9 9       â $r7+6   ÿŽ â Er5 5       â $rBBC   ÿŽ â Er6 :       â $r=8=   ÿŽ â Er7 7       â 0r4ÿL  0 A   â ErB B       â #r9F9   ÿ   â#r5F5   ÿ   âEr: =       â #r=F7   ÿ   â #r6J6   9   â r7;4     A  â #r9JB   5   â xAQ??  ÿÀŽ Ä xQQÀÀ  ÿÀŽ ä x5M  ÿÀŽ ä 4 ÿ       ä x@NÿÀŽ â #r:J:   =   â x=PÿÀŽ â 'AQ   @4 Ž È 0r;ÿR  0 A   ä &r@@   ÿ   â xBA?   ÿÀŽ ä  r4:;     A  â &r==   @   ä 0r:ÿL    A   ä $r==B   ÿŽ Ä 0r@ÿR    A   â #r:6:   7  â x6ÿ   M  ä Er= =       ä xDÿ   A  â #r@9@   4  â xAO??  ÿÀŽ äx76  ÿÀŽ â #r9F=   ÿ   â xOOÀÀ  ÿÀŽ Ä 6 ÿ       ä x5?  ÿÀŽ ä x=>ÀÀ  ÿÀŽ ä4 ÿ       â 'AO   @6 Ž â x;>??  ÿÀŽ ä x6TÿÀŽ â ';=   @4 Ž â x4ÿ   ?  Ä x>S??  ÿÀŽ â &r=6   ÿ   â xTÿ   T  ä x54  ÿÀŽ ä x6A?   ÿÀŽ ä xSSÀÀ  ÿÀŽ ä 4 ÿ       Ä x7UÿÀŽ â $r6+6   ÿŽ â xUÿ   U  ä x?TÿÀŽ â '>S   @4 Ž â xNÿ   N  ä xBUÿÀŽ â &r??    ÿ   â xAÿ   A  Ä Er5 6       ä xPÿ   P  ä xCNÿÀŽ ä x6>?   ÿÀŽ â &r?B)   ?   â x4;?   ÿÀŽ â $rA8A   ÿŽ â xBÿ   ;  â &rCC    ÿ   â xPPÿÀŽ â &r77   =   â x;ÿ   >  â $r6+6   ÿŽ â ErA A       â &rCP)   C   â #r5F5   ÿ   â $r474   ÿŽ â Er7 6       â $r;8;   ÿŽ â #rAJA   5   â $rCCD   ÿŽ â Er5 4       â $r??B   ÿŽ â Er; ;       â #r7F7   ÿ   âErC C       â 0r4ÿV  0 A   â Er6 ?       â 0rBÿ<  0 A   â #r5F5   ÿ   â #r7J;   7   â0rVÿV    A   â #r9JC   9   â0r>ÿ<    A   â  r4A4     A  â #r5J6   5   â  rB7B     A  â !r':'       â !r@       â #r99V   4  â #r>5>   B  â G‰  `    € ê r6   ÿàõ È r7-   ÿä ä r4"   ÿàõ Æ �y:6øÿÿ  ¢ r5   ÿä Æ �yB6     è �y?4    ( �yA6üÿÿ  ( �y@4     "	 r<,   ÿàõ È r=-   ÿä Ê �y<<     " xD:??  ÿÀŽ äOxC:ÀÀ  ÿÀŽ ä x;B  ÿÀŽ ä� : ÿ       ä x4ÿ   ?  äxBÿ   B  â 'DC   @: Ž â x7A??  ÿÀŽ Ä x;4ÿÀŽ ä x5B  ÿÀŽ ä xAAÀÀ  ÿÀŽ ä 4 ÿ       ä x6?ÿÀŽ ä x:ÿ   @  â '7A   @4 Ž â x5@ÿÀŽ Ä x@:ÿÀŽ â &r66   ÿ   â x:7?   ÿÀŽ â &r;;    ÿ   â x4D?   ÿÀŽ â &r55   6   â x7ÿ   7  â $r:+:   ÿŽ â xDÿ   D  â &r;@)   ;   ä $r554   ÿŽ Ä $r487   ÿŽ â Er7 :       â $r;;D   ÿŽ â Er5 5       ä Er6 4       â #r7F7   ÿ   âEr4 ;       â #r5F5   ÿ   â 0r;ÿ<  0 A   â #r6J6   7   Ä #r4J4   5   â 0r5ÿ<    A   â  r;6;     A  È #r445   ;  È !r4       ä G‰  `    € ê r6   ÿàõ È r7/   ÿä ä r4$   ÿàõ Æ �y:6üÿÿ  ¢ r5   ÿä Æ �yB6    è �y?4    ( �yA6     ( �y@4     "	 r<.   ÿàõ È r=/   ÿä Ê �y<<     " xD:??  ÿÀŽ äOxC:ÀÀ  ÿÀŽ ä x;B  ÿÀŽ ä� : ÿ       ä x4ÿ   ?  äxBÿ   B  â 'DC   @: Ž â x7A??  ÿÀŽ Ä x;4ÿÀŽ ä x5B  ÿÀŽ ä xAAÀÀ  ÿÀŽ ä 4 ÿ       ä x6?ÿÀŽ ä x:ÿ   @  â '7A   @4 Ž â x5@ÿÀŽ Ä x@:ÿÀŽ â &r66   ÿ   â x:7?   ÿÀŽ â &r;;    ÿ   â x4D?   ÿÀŽ â &r55   6   â x7ÿ   7  â $r:+:   ÿŽ â xDÿ   D  â &r;@)   ;   ä $r554   ÿŽ Ä $r487   ÿŽ â Er7 :       â $r;;D   ÿŽ â Er5 5       ä Er6 4       â #r7F7   ÿ   âEr4 ;       â #r5F5   ÿ   â 0r;ÿ<  0 A   â #r6J6   7   Ä #r4J4   5   â 0r5ÿ<    A   â  r;6;     A  È #r545   ;  È !r5       ä G‰  `    € ê r6   ÿàõ È r71   ÿä ä r4(   ÿàõ Æ �y:6üÿÿ  ¢ r5#   ÿä Æ �yB6    è �y?4    ( �yA6     ( �y@4     "	 r<0   ÿàõ È r=1   ÿä Ê �y<<     " xD:??  ÿÀŽ äOxC:ÀÀ  ÿÀŽ ä x;B  ÿÀŽ ä� : ÿ       ä x4ÿ   ?  äxBÿ   B  â 'DC   @: Ž â x7A??  ÿÀŽ Ä x;4ÿÀŽ ä x5B  ÿÀŽ ä xAAÀÀ  ÿÀŽ ä 4 ÿ       ä x6?ÿÀŽ ä x:ÿ   @  â '7A   @4 Ž â x5@ÿÀŽ Ä x@:ÿÀŽ â &r66   ÿ   â x:7?   ÿÀŽ â &r;;    ÿ   â x4D?   ÿÀŽ â &r55   6   â x7ÿ   7  â $r:+:   ÿŽ â xDÿ   D  â &r;@)   ;   ä $r554   ÿŽ Ä $r487   ÿŽ â Er7 :       â $r;;D   ÿŽ â Er5 5       ä Er6 4       â #r7F7   ÿ   âEr4 ;       â #r5F5   ÿ   â 0r;ÿ<  0 A   â #r6J6   7   Ä #r4J4   5   â 0r5ÿ<    A   â  r;6;     A  È #r445   ;  È !r4       â !r
9
       â !r	>	       â G‰  `    € ê r6    ÿà÷ är4&   ÿàõ ä r73   ÿäÿ ä r5!   ÿä Æ �y@6     ¨ �y:6üÿÿ  è �y94    ( �y?6øÿÿ  ( �y>4     " r<2   ÿàõ È r=3   ÿä Ê �y<<     " 4 ÿ       ä x;ÿ   @  äO xC:??  ÿÀŽ ä�xA:ÀÀ  ÿÀŽ ä x;;  ÿÀŽ ä : ÿ       Ä xB9ÿÀŽ äx6???  ÿÀŽ â 'CA   @: Ž â x9ÿ   9  ä x5@  ÿÀŽ ä x??ÀÀ  ÿÀŽ â &rBB   ÿ   â x99ÿÀŽ Ä x7ÿ   >  â '6?   @4 Ž â x>ÿÀŽ ä x4C?   ÿÀŽ â &r99    ÿ   â x:7ÿÀŽ â &r   B   â x 6?   ÿÀŽ â $r4+4   ÿŽ â xCÿ   C  â &r9:)   9   â x6ÿ   6  â $r    ÿŽ â Er) 4       â $rC8C   ÿŽ Ä $r696   ÿŽ â Er        â #r)F)   ÿ   âErC C       â 0r ÿ<  0 A   âEr6 6       â #rF   ÿ   ä #r)JC   )   â0r<ÿ<    A   â #rJ6      Ä  r )      A  È #r<      È !r       â x   ÿàÿ ä x€  ÿà÷ ä |    pbô ä x€  ÿàù ä x@   ÿàÿ ä rÿ   ÿäÿ Ä rÿ   ÿä Ê D)       À â Gy  `ãÿÿÿÿƒ ê Ay        € ê y      "   " Ey  0    € â r ÿ   pRð är ÿ   pRò Ö G‰  ð     € ê z ÿ `  pPð ä r    ÿ8Ž ä z ÿ a   Sð Æ ˆs  ' þÿ H   ô ˆƒ  €þÿ H   è ˆƒ  
 ÿÿ H   è ˆƒ  	€ÿÿ H   â G‰  p     € ê ˆs     H   è ˆs  €þÿ H   è ˆs  €  H   è ˆs  
 ÿÿ H   è ˆs     H   è ˆs  	€ÿÿ H   è ˆs  €  H   ä Ay        € ê {          ì M        € ê y       &   "> z ÿ `  pPð Ä x        â y      '   b z ÿ a   Sð â $z   x  Ž È $z  ~   Ž È/ %v  j   Ž È G	  P    € ê „y      H   " z  r  p`ð Æ „y €   H   br ÿ   pVp  Æ „y    H   ¨ „y    H   è „y €  H   ( „y    H   ( „y    H   ( „y €  H   ( „y    H   ( „y €  H   ( „y €  H   " !r  '       Æ „y €  H   " !r       â/ !r
       âO !r         ä� !r       â!r       â !r         ä !r       Æ ‰       b !r       Æ ‰      ¢ !r	       Æ ‰
      â !r       È !r       â !r        â/ !r       ÈO ‰      " !r
       Æ� ‰
      h ‰	      ¨ ‰      â !r       â !r
       â/ !r
	       ÈO ‰      " !r       Æ� ‰	 €    h ‰ €    ¨ ‰
 €    â !r       â !r		       È/ ‰ €    " !r       ÆO ‰ 	 @    b !r
       Æ� ‰ @    ¢ x
   ÿàÿ Æ ‰ @    â z 
 r  p`ô È x    pVt â !r       â !r 	        È/ ‰ @    " x	   ÿàÿ â!r       âO x   ÿàÿ â ‰       b z 	 r  p`ò â !r       ä� ‰      ¢ x    pVò  Æ ‰      â !r       â !r
        È/ ‰       " !‚	
       â !r       âO !r       Æ� †‰ 	    â z  r  p`ð â !¢       â!’       ä x    pVp  ä †©    è †™    â !r         Ì M	        € ê !r        Ê †y    â My        € ê z   h      È x     pBð Ú G	  @    € ê x     pRð Ú G‰  À    € ê x     pRð Ú G	   $    € ê „y     H   " z  r  p`ð â Ey  p    € ä „y     H   b r ÿ   pVp  Æ „y    H   ¨ „y    H   è „y    H   ( „y  
  H   ( „y €   H   ( „y €  H   ( „y €  H   " !r'       â !r         Æ/ !r       âO !r         Æ� !r       â!r         È ‰      " !r       Æ ‰       b !r       È !r       â !r       ä „y €  H   â !r        Æ/ ‰      ( ‰      b !r       â !r       È/ ‰ €    ( ‰ €    b !r       Æ „y €  H   ¢ !r       Æ „y €
  H   â !r       Æ/ ‰ @    ( ‰ @    b !r       âO !r       â !r       È/ ‰       ( ‰      b !r       â !r         â !r       â/ G	  �     € ê !r       â!r         Æ  x;ª¸¿  A  Ì s        $ !t  €?    Ì s        $  r     A  È  r      A  Ê †y     ä Ay        € ê ‰       b Ey  Ð    € æ ‰      ¨ „y    H   è „y    H   ( „y    H   è „y  	  H   â !r         Æ/ „y    H   â !r       ÆO ‰       h ‰      " !r
       â� !r       â!r        ä/ !r       â !r       Ä ‰ €    " !r       Æ ‰ €    b !r       Æ „y    H   ¢ !r       â !r       È/ ‰ @    ( ‰ @    b !r       âO !r       â x   ÿàÿ â !r       Æ/ ‰       " z  r  p`ð Æ ‰      b x    pVp  â !r         â !r       Ö/ G	  �     € ê !r       â!r         Æ  x
;ª¸¿  A  Ì s
 
       $ !t
  €?    Ì s        $  r     A  È  r      A  Ê †y    ä Ay        € ê ‰       b Ey       € æ ‰      ¨ „y €  H   è „y €  H   è „y €  H   è „y €  H   â !r         Æ/ „y €  H   â !r
       ÆO ‰       h ‰
      ( „y €	  H   ¢ !r       â� !r        â/ !r
       È ‰ €    " x
   ÿàÿ Æ ‰ €    b z 
 r  p`ð È x    pVp  â !r       ÈO !r       â !r       â !r       È/ ‰ @    ( ‰ @    b !r       â !r       È/ ‰       ( ‰      b !r         â !r       â/ G	  �     € ê !r       â!r         Æ  x
;ª¸¿  A  Ì s
 
       $ !t
  €?    Ì s        $  r     A  È  r      A  Ê †y    ä Ay        € ê ‰      b !r	       â x   ÿàÿ Æ !r       â z  r  p`ð Æ !r       â x    pVp  È ‰       ¢ !r
       Ê/ ‰
      b !r         ÊO ‰	       ¢ !r
       Ê/ ‰ €    b !r	 	       ÊO ‰	 €    ¢ !r       Ê? ‰ @    " !r	       ÊO ‰ @    b !r       Ê ‰      " !r       Ê/ ‰       ¢ !r       â M	        € ì !r       â/!r         ÆO  x;ª¸¿  A  â !r        Ê s        $ !t  €?    Ì s        $  r     A  È  r     A  Ê †y    â My        € ê „y      H   " z  r  p`ð â Ey       € ä „y    H   b r ÿ   pVp  Æ „y    H   ¨ „y    H   è „y    H   ( „y  
  H   ( „y €   H   ( „y €  H   è „y €
  H   â !r  '       â !r       Æ/ !r         âO !r       Æ� !r        â!r       È ‰       " !r       Æ ‰      b !r         Æ „y €  H   â !r       Æ/ ‰       ( ‰      h „y €  H   ¢ !r        â !r       È/ ‰ €    ( ‰ €    b !r       Æ „y €  H   â !r       ÈO !r       â !r       â !r       È/ ‰ @    ( ‰ @    b !r       È� !r       â !r       â !r       È/ ‰       ( ‰      b !r        â !r       â/ G	  À     € ê !r        È x '7=  A  â x *BL?  A  Æ #t   €?   È  r     A  â  x    ?  A  â !r        È s     $   $ !t  €?    È  r     A  È  r      A  Ê †y     ä Ay        € ê ‰       b Ey       € æ ‰      ¨ „y    H   è „y    H   ( „y  	  H   è „y    H   â !r         Æ/ „y    H   â !r       ÆO ‰       h ‰      " !r
       â� !r       â!r        Æ/ !r       â !r       â ‰ €    ( ‰ €    b !r       Æ „y    H   ¢ !r       â !r       È/ ‰ @    ( ‰ @    b !r       ÈO !r       â !r       â x   ÿàÿ â !r       Æ/ ‰       " z  r  p`ð Æ ‰      b x    pVp  â !r
        â !r       Ö/ G	  À     € ê !r        È x '7=  A  â x *BL?  A  Æ #t   €?   È  r     A  â  x    ?  A  â !r 
       È s     $   $ !t  €?    È  r     A  È  r      A  Ê †y    ä Ay        € ê ‰       b Ey  Ð    € æ ‰      ¨ „y €  H   è „y €	  H   ( „y €  H   è „y €  H   â !r         Æ/ „y €  H   â !r
       ÆO ‰       h ‰
      ( „y €  H   â !r       È� !r       â!r        Æ/ !r       â !r
       â ‰ €    ( ‰ €    b !r       â !r       È/ ‰ @    ( ‰ @    b !r       â x   ÿàÿ â !r       Æ/ ‰       " z  r  p`ð Æ ‰      b x    pVp  â !r
        â !r       Ö/ G	  À     € ê !r        È x '7=  A  â x *BL?  A  Æ #t   €?   È  r     A  â  x    ?  A  â !r 
       È s     $   $ !t  €?    È  r     A  È  r      A  Ê †y    ä Ay        € ê ‰      b !r	       â x   ÿàÿ Æ !r       â z  r  p`ð Æ !r       â x    pVp  È ‰       ¢ !r
       Ê/ ‰
      b !r         ÊO ‰	       ¢ !r
       Ê/ ‰ €    b !r	 	       ÊO ‰	 €    ¢ !r       Ê? ‰ @    " !r	       ÊO ‰ @    b !r       Ê ‰      " !r       Ê/ ‰       ¢ !r       â M	        € ì !r       â/!r         ÆO  x'7=  A  â x*BL?  A  â !r        ä #t  €?   â x   ?  A  Æ  r     A  Ì s     $   $ !t  €?    È  r     A  È  r     A  Ê †y    â My        € ê x     pRð Ú G‰  P    € ê x     pRð Ú G	  `    € ê „y      H   " z  r  p`ð â Ey       € ä „y    H   b r ÿ   pVp  Æ „y    H   ¨ „y    H   è „y    H   ( „y  
  H   ( „y €   H   ( „y €  H   è „y €
  H   â !r  '       â !r       Æ/ !r         âO !r       Æ� !r        â!r       È ‰       " !r       Æ ‰      b !r         Æ „y €  H   â !r       Æ/ ‰       ( ‰      h „y €  H   ¢ !r        â !r       È/ ‰ €    ( ‰ €    b !r       Æ „y €  H   â !r       ÈO !r       â !r       â !r       È/ ‰ @    ( ‰ @    b !r       È� !r       â !r       â !r       È/ ‰       ( ‰      b !r        â !r       â/ G	  À     € ê !r        Ê	z  i    � â !r        È  x;ª¸¿  A  â 	z   i    � È 	z   i €  � â s        $ !t  €?    Ì s        $  r     A  È  r      A  Ê †y     ä Ay        € ê ‰       b Ey       € æ ‰      ¨ „y    H   è „y    H   ( „y    H   è „y    H   â !r         â/ !r       ÈO ‰       h ‰      ( „y  	  H   ¢ !r
       â� !r       â!r        â/ !r       È ‰ €    ( ‰ €    h „y    H   â !r       ÈO !r       â !r       â !r       È/ ‰ @    ( ‰ @    b !r       È� !r       â !r       â x   ÿàÿ â !r       Æ/ ‰       " z  r  p`ð Æ ‰      b x    pVp  â !r        â !r       Ö/ G	  À     € ê !r        Ê	z  i    � â !r        È  x
;ª¸¿  A  â 	z   i    � È 	z   i €  � â s
 
       $ !t
  €?    Ì s        $  r     A  È  r      A  Ê †y    ä Ay        € ê ‰       b Ey  Ð    € æ ‰      ¨ „y €  H   è „y €	  H   ( „y €  H   è „y €  H   â !r         Æ/ „y €  H   â !r
       ÆO ‰       h ‰
      " !r       È� !r       â!r        â/ !r
       È ‰ €    ( ‰ €    h „y
 €  H   ¢ !r       â !r       È/ ‰ @    ( ‰ @    b !r
       âO !r       â x   ÿàÿ â !r       Æ/ ‰       " z  r  p`ð Æ ‰      b x    pVp  â !r        â !r       Ö/ G	  À     € ê !r        Ê	z  i    � â !r        È  x
;ª¸¿  A  â 	z   i    � È 	z   i €  � â s
 
       $ !t
  €?    Ì s        $  r     A  È  r      A  Ê †y    ä Ay        € ê ‰      b !r	       â x   ÿàÿ Æ !r       â z  r  p`ð Æ !r       â x    pVp  È ‰       ¢ !r
       Ê/ ‰
      b !r         ÊO ‰	       ¢ !r
       Ê/ ‰ €    b !r	 	       ÊO ‰	 €    ¢ !r       Ê? ‰ @    " !r	       ÊO ‰ @    b !r       Ê ‰      " !r       Ê/ ‰       ¢ !r       â M	        € ì !r       â/!r         ÈO 	z i    � â !r        È  x;ª¸¿  A  â 	z i    � È 	z i €  � â s        $ !t  €?    Ì s        $  r     A  È  r     A  Ê †y    â My        € ê „y      H   " z  r  p`ð Æ „y    H   b r ÿ   pVp  Æ „y €   H   ¨ „y €  H   è „y!    H   ( „y    H   ( „y    H   ( „y    H   ( „y €  H   ( „y €  H   ( „y    H   " !r  '       Æ „y#  	  H   " !r       Æ/ „y  
  H   b !r       ÆO „y    H   ¢ !r       Æ� „y €  H   â !r!!
       Æ„y €
  H   " !r       Æ „y%    H   " !r       Æ „y)    H   " !r         Æ „y €  H   " !r       Æ „y+ €  H   " !r       Æ „y- €  H   " !r!       Æ „y €	  H   â !r#       Æ „y' €  H   â !r       Æ/ „y/ €  H   â !r        âO !r       ä� ‰      " !r       Æ‰       b !r%       Æ ‰
      ¢ !r)       Æ ‰      â !r	       Æ ‰      " !r++       Æ ‰      " !r-       â !r       â !r+       â !r         â/ !r''       â !r

       ÆO ‰	       " x   ÿàÿ â!r//       â !r       â� ‰'      b z  r  p`ô â !r       ä‰/      ¢ x   ÿàÿ â !r       Ä ‰      â z  r  p`ò ä x    pVt â‰
      " x    pVò  ä x   ÿàÿ â ‰      & z  r  p`ö â ‰      " !r	 	       Ä x    pVö â ‰      " !r'       â/ !r/       äO ‰ 	 €    b !r       Æ� ‰      ¢ !r
       Æ‰      â !r       Æ ‰ €    " !r       Æ ‰
 €    " !r       Æ ‰ €    ( ‰ €    " !r 	        Æ/ ‰ €    b !r       âO !r       ä� ‰	  @    ¢ !r       Æ‰ €    â !r

       Æ ‰ €    " !r       Æ ‰ @    " !r       Æ ‰
 @    " !r       Æ/ ‰ @    h ‰ @    b !r	 	       ÆO ‰ @    ¢ !r       â� !r       ä‰ 	      â !r       Æ ‰ @    " !r
       Æ ‰ @    " !r       Æ/ ‰      b !r       Æ ‰
      b !r       ÆO ‰      ¨ ‰      ¢ !r 	        Æ� ‰      â !r       â !‚         â!r       Æ‰      " !r       Æ/ ‰      b !r
       â !‚       â !r	       äO !’       â  ‚      A  â !r
       â !’		       Æ †‰     â !r       â� !¢

       â  ’		     A  ä !¢       â !r       ä †™ 	   â  ¢
     A  â !r       È/ †©    â M9        € ê !r       â !r       È  r     A  Ê †y    â My        € ê „y      H   " z  r  p`ð â Ey  À    € ä „y    H   b r ÿ   pVp  Æ „y    H   ¨ „y    H   è „y    H   ( „y  
  H   ( „y €   H   ( „y €  H   è „y €
  H   â !r  '       â !r       Æ/ !r         âO !r       Æ� !r        â!r       È ‰       " !r       Æ ‰      b !r         Æ „y €  H   â !r       Æ/ ‰       ( ‰      h „y €  H   ¢ !r        â !r       È/ ‰ €    ( ‰ €    b !r       Æ „y €  H   â !r       ÈO !r       â !r       â !r       È/ ‰ @    ( ‰ @    b !r       È� !r       â !r       â !r       È/ ‰       ( ‰      b !r        â !r       â/ G	  à     € ê !r        Ê	x   à@  � Ê  x #ÛÙ?  A  È  x ;ª¸¿  A  â !r        Ê s        " 	x    à@  � È 	x    àÀ  � Ê !t   €?    â !t  €?    È s        $  r     A  È  r     A  Ê †y     ä Ay        € ê ‰       b Ey       € æ ‰      ¨ „y    H   è „y    H   ( „y    H   è „y    H   â !r         â/ !r       ÈO ‰       h ‰      ( „y  	  H   ¢ !r
       â� !r       â!r        â/ !r       È ‰ €    ( ‰ €    h „y    H   â !r       ÈO !r       â !r       â !r       È/ ‰ @    ( ‰ @    b !r       È� !r       â !r       â x   ÿàÿ â !r       Æ/ ‰       " z  r  p`ð Æ ‰      b x    pVp  â !r        â !r       Ö/ G	  à     € ê !r        Ê	x   à@  � Ê  x #ÛÙ?  A  È  x
 ;ª¸¿  A  â !r        Ê s
 
       " 	x    à@  � È 	x    àÀ  � Ê !t   €?    â !t
  €?    Ì s        $  r     A  È  r     A  Ê †y    ä Ay        € ê ‰       b Ey  ð    € æ ‰      ¨ „y €  H   è „y €	  H   ( „y €  H   è „y €  H   â !r         Æ/ „y €  H   â !r
       ÆO ‰       h ‰
      " !r       È� !r       â!r        â/ !r
       È ‰ €    ( ‰ €    h „y
 €  H   ¢ !r       â !r       È/ ‰ @    ( ‰ @    b !r
       âO !r       â x   ÿàÿ â !r       Æ/ ‰       " z  r  p`ð Æ ‰      b x    pVp  â !r        â !r       Ö/ G	  à     € ê !r        Ê	x   à@  � Ê  x #ÛÙ?  A  È  x
 ;ª¸¿  A  â !r        Ê s
 
       " 	x    à@  � È 	x    àÀ  � Ê !t   €?    â !t
  €?    Ì s        $  r     A  È  r     A  Ê †y    ä Ay        € ê ‰      b !r	       â x   ÿàÿ Æ !r       â z  r  p`ð Æ !r       â x    pVp  È ‰       ¢ !r
       Ê/ ‰
      b !r         ÊO ‰	       ¢ !r
       Ê/ ‰ €    b !r	 	       ÊO ‰	 €    ¢ !r       Ê? ‰ @    " !r	       ÊO ‰ @    b !r       Ê ‰      " !r       Ê/ ‰       ¢ !r       â M	        € ì !r       â/!r         ÈO 	x  à@  � â !r        È  x#ÛÙ?  A  â 	x  à@  � Æ  x;ª¸¿  A  â 	x  àÀ  � Ê s        " !t  €?    â !t  €?    Ì s        $  r     A  È  r     A  Ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y       "   $ z   w  p`ð Ú M	        € ê y      !   " ¹z  l      â Ey  �    € â ™x?    â $rÿÿ   ÿ Ž â ¹z  F   
   â r ÿ       ä xÿ     È |    p`ð Ú G	      € ê y      &   " $t	ÿ   ÿ Ž ä $z  x  Ž È %v \  	 Ž Ì �y    ™ ¢  $r
ÿÿ   ÿ Ž â x   ÿÀŽ â $tÿ$   ÿ Ž â y      %   b 'z m  
 Ž â xÿ     ä xÿ     â $vÿ X  ÿ Ž â zÿ n    â $x   ÿ Ž Æ rÿ   ÿáÿ Ê $z
 o  Ž â xøÿÿøŽ â $z  r  ÿŽ ä $z

 u  ÿŽ ä $x   ÿ Ž Æ r
   à‘ ä x
   ÿÀŽ ä x   ÿ   ä/ rÿÿ   ÿ$~  â $x	    
Ž ä $z q  Ž âx   ÿàÿ â %v
 Z   Ž È $x$   ÿŽ Ê r   ÿàÿ â $z t  Ž äO$z t  Ž ä $z q  Ž ä %x�   ÿŽ â z  Y      Æ %x�   ÿŽ â r	   ÿàñ äx   øŽ â$x   ÿ Ž â r		   ÿàó â $rÿÿ   ÿ Ž â x   øŽ â $rÿÿ     â x   ÿÀŽ ä x   ÿüŽ Ä x   ÿüŽ ä rÿ   ÿäÿ  ä r   ÿàó â�y
    • ¢  r*   ÿàñ Æ $x   Ž  â r
   ÿàó â $x+     â �y,
$   • è  �y øÿÿ • ( �y#    • b rÿ   ÿäÿ  Æ �y*   ™ ¨  �y$*    ™ è  �y%   ™ è  �y"   ™ è  �y((   ™ è  �y&üÿÿ • è �y'8   ™ â  r   ÿàó Ä r	   ÿàñ Æ $x   Ž  ä $x     Æ �y)   • è �y+   ™ è �y*    • è �y-üÿÿ • â r   ÿàñ Æ �y    ™ â( r   ÿä  ä r   ÿàñ Æ �y    ™ ä $x     Ê �y    ™ â  x    p@ð ä x   ÿàÿ ä x   ÿàó ä x

@  ÿàõ Æ $rÿÿ   Ž  â rÿ   ÿä ä x ??  ÿÀŽ äx ÀÀ  ÿÀŽ â $ ÿÿ   ÿ Ž â x!#  ÿÀŽ äxÿÀŽ ÆO '   @  Ž â xÿ     ä x$ÿÀŽ ä� xÿÀŽ â &r%   ÿ   â xÿ   #  Æ &r"      â x ?   ÿÀŽ â &r(   ÿ   â x$ÿ   $  ä x&??  ÿÀŽ äx&ÀÀ  ÿÀŽ ä x&ÿ     ä x  ÿÀŽ Ä  ÿ       ä x$$ÿÀŽ Æ '   @ Ž È &z%  € ÿ   ä &r$'      ä $r    ÿŽ â x ?   ÿÀŽ â &z$(  € ÿ   ä &z"  €    ä $r!&   ÿŽ â xÿ     â &z'  € $   Ä $r    ÿŽ â Er$        â $r   ÿŽ ä 0rÿ    A   â Er        â 0r,ÿ,    A   â Er        ä xÿ   )  â #r   ÿ   â x +ÿÀŽ Ä x*??  ÿÀŽ â #r,      â x  ÿÀŽ â $ÿÿ   ÿ Ž â x**ÀÀ  ÿÀŽ â &r% %   ÿ   â Er! !       â #r#$   ÿ   â x -??  ÿÀŽ â '*   @ Ž â x)  ÿÀŽ Ä x--ÀÀ  ÿÀŽ â #r!,!   #   â x+ÿ   +  ä x#ÿ     â ' -   @ Ž â xÿÀŽ ä x++ÿÀŽ ä x?   ÿÀŽ â &r"   %   â x ?   ÿÀŽ Ä x$#ÿÀŽ â &r++(   ÿ   â x#ÿ     â $r   ÿŽ â xÿ      â $r   ÿŽ ä &r$$'   +   ä $r##   ÿŽ â Er        â $r$   ÿŽ â Er        Ä Er# #       â #r   ÿ   â |    p`ð â #r   ÿ   â 0rÿ  0 A   â Er        â 0rÿ  0 A   â #r,#      â0rÿ    A   â  r     A  â 0rÿ    A   â #r,      â  r     A  â #r!!     Æ #r     â !r!       Æ !r       â G‰  ð÷ÿÿÿÿƒ ê Ay        € ê ‰      ( ‰      h y      %   ¢ !r       â !r       È/ ‰      " $x   Ž ÆO ‰
      d z  p  p`ð È x    pDp  â !r       â !r

       È/ ‰ €    ( ‰	
 €    b !r       â !r	
	       È/ ‰ @    ( ‰	 @    b !r       â !r	       È/ ‰      h  ‰      ¢  M	        € ê y	      &   â x   ÿ   â $z   s  Ž â !r       â/ !r       âO x    p"ò â$tÿ   ÿ Ž â r ÿ   p"ð Ö         â $z 	 v   Ž È� %v  j   Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      "   $ z  w  p`ð Ú M	        € ê y      !   ( y	      &   b ¹z  l      â $tÿ   ÿ Ž â ™x?    â Ey  `    € â ¹z  F   
   â x(ÿ     È | (   p`ð â $z x  	Ž È/ %v \   Ž â r ÿ       â x      ÿ  ä $rÿÿ   ÿ Ž â �y     ™ b $rÿÿ   ÿ Ž â r ÿ       â G	  �    € ê $rÿÿ   ÿ Ž â y
      %   " $z r  ÿŽ â x   ÿ   â'z	 m   Ž â x   ÿÀŽ â Ey      € â x   ÿÀŽ â $vÿ a  ÿ Ž â zÿ n    Ä xÿ     ä z  `      â $rÿÿ   
Ž â xÿ     ä x   ÿÀŽ â $z o  	Ž â xøÿÿÿÀŽ ä x	 $       â $z u  ÿŽ â x   øŽ â $x   ÿ Ž Æ r   à‘ â $x
   ÿ Ž â x`   øŽ ä rÿÿ   ÿ$~  â %v Z  	 Ž â z ÿ `  pPð ä x    p@ò â $x$   ÿŽ â z
  X      â $vÿ Y  ÿ Ž â z ÿ a   Sð â $x		   Ž Ä $z q  (Ž È $z  t  Ž È%x�   ÿŽ â r ÿ       È r   ÿàõ är   ÿüŽ Æ $rÿÿ    ä r
   ÿà÷ â�y'    • ¢ r
   ÿàõ ä r   ÿäÿ â �y-$   • â r    ÿàù â $x    ä �y/   • " r!ÿ	   ÿä Æ �y.   ™ b xÿ     Æ �y,   • ¢  x   ÿàÿ Æ �y&    ™ â $z  t  Ž Æ �y    ™ â  $z q  Ž Æ �y    ™ è  �y$   • â  %x�   ÿŽ Æ �y (   ™ ä  r%   ÿüŽ ä �y 8   ™ ä  r
%   ÿàõ Ê/ $x    â r+   ÿàõ È r*ÿ   ÿä â �y0   • â r
+   ÿàõ Æ �y)   • ä $x   * ä �y2   • è �y1   ™ â r 
   ÿàõ Æ �y3    ™ â r!   ÿä ä r"
   ÿàõ Æ �y      ™ ä  $x#    Ê �y""    ™ â   ÿ       ä/ x((   ÿàÿ È | (   p`ô ä x4/??  ÿÀŽ äx//ÀÀ  ÿÀŽ ä x5.ÿÀŽ äx,  ÿÀŽ äO x&ÿÀŽ Æ� '4/   @ Ž â xÿ   ,  Æ &r5   ÿ   â x.ÿ   .  ä x  ÿÀŽ â &r      ä $ÿÿ   ÿ Ž â x$??  ÿÀŽ äx!$ÀÀ  ÿÀŽ ä x$4?   ÿÀŽ Ä x&ÿ   &  â '!   @ Ž â x#.ÿÀŽ Æ $r$   ÿŽ â x&&ÿÀŽ â &z$  € ÿ   â x!?   ÿÀŽ â &r##   ÿ   ä &z  € ÿ   ä &z$  € $   â xÿ     â &r&   #   Ä &z#  €    ä $r$!   ÿŽ â x4ÿ   4  â $r!#   ÿŽ ä 0r'ÿ'    A   â Er        â $r4   ÿŽ â Er! !       â 0r-ÿ-    A   Ä #r'   ÿ   â Er&        ä Er        â #r!-!      â x)  ÿÀŽ ä x/0ÀÀ  ÿÀŽ ä ÿ       ä x0??  ÿÀŽ Ä x)ÿ   )  â '/   @ Ž â xÿ   1  â #r&'&   ÿ   â x)  ÿÀŽ ä x)2ÀÀ  ÿÀŽ â #r-   &   â x/ÿÀŽ â $ÿÿ   ÿ Ž â x&2??  ÿÀŽ Ä x11ÿÀŽ â '&)   @ Ž â x,ÿ   3  ä x3ÿÀŽ â &r11   ÿ   â x)&?   ÿÀŽ â &r//   ÿ   â x,,ÿÀŽ â &r   1   â x?   ÿÀŽ â $r)$)   ÿŽ â x&ÿ   &  â &r,,   /   â x/ÿ     â $r   ÿŽ Ä 0rÿ   0 A   ä $r#&   ÿŽ â Er& )       â $r,,/   ÿŽ â  r!     A  â Er        ä Er!        â #r&'&   ÿ   âEr ,       â #r'   ÿ   Ä #r!-!   &   â0r&ÿ"  0 A   â #r-      â 0r ÿ     A   ä 0rÿ"    A   â  r!!&     A  ä #r)      ä #r"   !  â G‰  °    € ê xÿ     Ê $z q  Ž È $z  t  Ž È %x �   ÿŽ Ê r    ÿüŽ È r   ÿà÷ È r!   ÿäÿ ä r   ÿà÷ Æ �y1    ¤ $x   Ž ä �y0    è �y,    ( �y&    h �y2     b r.    ÿà÷ È r/!   ÿäÿ Ê �y..     b $ ÿÿ   ÿ Ž â  ÿ       ä x!1  ÿÀŽ äO x1ÿ   1  ä x50ÀÀ  ÿÀŽ ä�x30??  ÿÀŽ Ä x1  ÿÀŽ ä x,ÀÀ  ÿÀŽ â'35   @  Ž â x,??  ÿÀŽ ä x ÿ   &  â'   @ Ž â x!&ÿÀŽ ä x&ÿ   2  Ä x/ ÿÀŽ â/ &r!!   ÿ   â x22ÿÀŽ ä x?   ÿÀŽ â &r//   ÿ   â x&&ÿÀŽ â &r!2   !   â x3?   ÿÀŽ â $r$   ÿŽ â xÿ     â &r&&   /   â x3ÿ   3  â $r!   ÿŽ Ä $r#   ÿŽ â Er        â $r&&3   ÿŽ â Er        ä Er        â #r'   ÿ   âEr &       â #r'   ÿ   â 0r!ÿ.  0 A   â #r-      Ä #r-      â 0rÿ.    A   â  r!!     A  È #r   !  È !r       â !r)       â !r"       â G‰  P    € ê r%   ÿàù är+   ÿà÷ Æ $x    â r*   ÿäÿ È �y+    ¨ �y*    è �y&    ( �y)    h �y%     b r    ÿà÷ Ê $x!   Ž Ê �y       b xÿ   +  ÈO x  ÿÀŽ ä x/*ÀÀ  ÿÀŽ ä� ÿ       ä x"*??  ÿÀŽ ä x1&ÿÀŽ â'"/   @ Ž â x)??  ÿÀŽ Äx+  ÿÀŽ ä x))ÀÀ  ÿÀŽ â &r11   ÿ   â x&ÿ   &  ä x%ÿÀŽ â ')   @ Ž â xÿ   %  ä x&ÿÀŽ â &r   1   â x"?   ÿÀŽ Ä xÿÀŽ â &r   ÿ   â x?   ÿÀŽ â $r$   ÿŽ â x"ÿ   "  â &r      â xÿ     â $r   ÿŽ â Er        â $r"#"   ÿŽ ä $r   ÿŽ â Er        â #r'   ÿ   â Er" "       Ä Er        â #r'   ÿ   â 0rÿ   0 A   â#r-"      ä#r-      â 0rÿ     A   â  r     A  È #r     È !r       â x

   ÿà÷ ä x   ÿàû ä x@  ÿàù â $rÿÿ   Ž ä $rÿÿ   Ž â r	ÿ	   ÿä â D)       À â Gy   ðÿÿÿÿƒ ê Ay       € ê r        â $rÿÿ    Ž Ä Ay        € ê z ÿ `  pPð È z ÿ a   Sð Ú G	  P    € ê ‰      ( ‰      b !r       â !r       È/ ‰      ( ‰      b !r       â !r	       È/ ‰ €    ( ‰
	 €    b !r       â !r
	
       È/ ‰ @    ( ‰
 @    b !r       â !r
       È/ ‰      ( ‰      b !r       â !r       â/ Gy  €    € ê ‰	      ( ‰      h ‰      ¨ ‰
      â !r		       â !r       â/ !r       ÈO ‰      " !r
       Æ� ‰
	      h ‰      ¨ ‰      â !r       â !r
	
       È/ ‰ €    " !r       ÆO ‰
 €    b !r       Æ� ‰ €    ¨ ‰ €    â !r       â !r
       â/ !r       ÈO ‰ @    " !r       Æ� ‰ @    h ‰ @    ¨ ‰ @    â !r       â !r       â/ !r       ÈO ‰	      â !r
       Æ� ‰      ( ‰      h ‰
      ¢ !r       â !r       â/ !r	       â !r
       ÄO y      %   $ r   ÿŽ È z  p  p`ð È x    pDp  Ú M	        € ê z ÿ ^  pPð â $z  v  Ž Æz ÿ _   Sð Ú $ÿ   ÿ Ž È % ^   Ž Ì �	     ¢ x    ÿ   È r  ÿ   p"ò äx     p"ô äx     p"ö äx     p"ø Î $ ÿÿ    Ž â z ÿ `  pPò ä "         â $2 ÿÿ    Ž â z ÿ a  Sò ä B         Ê !         ÌO G™       € ê z ÿ b  pPð È z ÿ c   Sð Ú $ÿ   ÿ Ž È % b   Ž Ì �	     ¢ x   ÿ   ä z  h      ä x øÿÿÿp"ò äx üÿÿÿp"ô är ÿ   p"ö äx    p"ø Î $ÿÿ    Ž â x    pBò ä "        â $2ÿÿ    Ž ä $Bÿÿ    Ž È !       ÈO G  P    € ê x    pRð Ú G‰  �     € ê x    pRð Ú G	  ð    € ê  x;ª¸¿  A  Ì s        $ !t  €?    Ì s        $  r     A  È  r       A  â Gy  P    € ê  x'7=  A  â x*BL?  A  Æ #t  €?   â  x   ?  A  Æ  r     A  Ì s     $   $ !t  €?    È  r     A  È  r       A  â Gy  °    € ê x    pRð Ú G‰  à     € ê x    pRð Ú G	        € ê 	z i    � ä 	z   i    � Æ  x;ª¸¿  A  â 	z   i €  � Ê s        $ !t  €?    Ì s        $  r     A  È  r       A  â Gy  Ð     € ê  r       A  â Gy  °     € ê 	x  à@  � ä 	x    à@  � Æ  x#ÛÙ?  A  â 	x    àÀ  � Æ  x;ª¸¿  A  ä !t   €?    È s        $ !t  €?    Ì s        $  r      A  È  r       A  ä y      "   ( y      &   b $z s  Ž â x        Æ $z v  Ž È/ %v j   Ž Ê †y      â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y8      "   " ¹z  l      â Ey  �3    € â ™x?    â y/      !   b ¹z  F   
   â r. ÿ       â x0      ÿ  â x6      ÿ  â x2      ÿ  â x4      ÿ  â $r]ÿÿ   ÿ Ž â x:      ÿ  â $rYÿÿ   ÿ Ž â r[ ÿ       â $r<ÿÿ   ÿ Ž Ä $r_ÿÿ   ÿ Ž â r 8ÿ   pRð är 8ÿ   pRò â$x 8    /Ž Ê/ xÿ      È r    ÿ � È x9ÿ     È | 9   pbô Ú G)   2    € ê y      %   " xðÿÿÿÿÀŽ â $rÿÿ   ÿ Ž â zA  Y      â y      '   b r   €ÿàÿ â $xO9   ÿ Ž ä y      &   ¢ x ÿ     â $x   ÿ Ž â zMO q  ÿàÿ â $vGÿ X  ÿ Ž â r    ÿ� â $r.ÿÿ   ÿ Ž Æ xøÿÿÿÿÀŽ ä xDÿ     Æ $xS   
Ž â xDD   ÿ   ä x    ÿ   â $rÿÿ   ÿ Ž â x
ÿ   S  â 'z y   Ž â/xÿ   S  ä xUÿ   D  â $z> }  ÿŽ â zÿ z    â $z w  ÿŽ âO xQS     â 'z s   Ž Æ r>>   ÿàõ â $z |  9Ž âr
S   ÿ� â $z	  p  9Ž â zÿ t    â $rBÿÿ   ÿ â r   ÿò� â $z	 |  	Ž â x     ÿàÿ â $z v  Ž âxp   ÿàÿ â $z v  	Ž â xÿ     â $z  p  Ž â xÿ   p  â $xS   ÿ Ž â xl   ÿàÿ â $xSS   ÿ Ž â x    ÿàÿ Ä r    ÿ� är#p   ÿ� är$p   ÿ� ä r"   ÿ� ä xÿ   l  ä x	ÿ      Ä xj øÿÿÿÿÀŽ ä xf"üÿÿÿÿÀŽ ä r%l   ÿ� â $xj   j
Ž â r'	    ÿ� ä r&l   ÿ� â %xn   ÿŽ â r(	    ÿ� Ä xf   fèÿ ä xo%øÿÿÿÿÀŽ â %x   Ž â xc'øÿÿÿÿÀŽ ä x#øÿÿÿÿÀŽ â $xnl   o
Ž â xq$üÿÿÿÿÀŽ â $xb    c
Ž â xm&üÿÿÿÿÀŽ â %x   Ž â x!(üÿÿÿÿÀŽ Ä rp  €ÿàÿ äxpp   qèÿ ä xll   mèÿ ä x     !èÿ ä z  q      ä xgÿ   f  Ä xkÿ   j  â $xC   OŽ âx!ÿ      â $xE   OŽ âxcÿ   b  â $x?   OŽ âxmÿ   l  â $xI   OŽ âxoÿ   n  â $xK   OŽ â xqÿ   p  â %xn   fŽ â xÿ     Ä rTf   ÿàû â%x
n    Ž ârV    ÿà÷ ärZb   ÿàý â %xn   bŽ âr=O   ÿŽ ä xHÿ   "  â %xn   jŽ â xJÿ   #  Ä xNÿ   %  â %xn   lŽ âxPÿ   &  ä rXj   ÿàõ â %xn   nŽ â rbc   ÿä ä xFÿ      â %xn   pŽ â xLÿ   $  Ä xRÿ   '  â %xn   Ž âxWÿ   (  ä r\l   ÿàù â%xn   ÿŽ â rdp   ÿàý äxFF   ÿ   â $x^   !Ž âr`n   ÿà÷ â$xf   gŽ â rh   ÿàû â %x   Ž â xLL   ÿ   Ä xRR   ÿ   â %x   Ž â xWW   ÿ   ä rlm   ÿä â$xHH   ÿ Ž â rpq   ÿä â $xJJ   ÿ Ž ä $xNN   ÿ Ž ä $xPP   ÿ Ž Ä $xj   k ä$xn   oŽ ä$x@   Ž ä rG   ÿàõ är"G   ÿà÷ ä rA   ÿä Æ $x#   AŽ ä �yk"   • ¨  �yt    • ¨  �yq"    • è �yv"   • â r G   ÿà÷ Ä r$G   ÿàõ Æ $x!   AŽ â r%A   ÿä ä r&GT   ÿà÷ ä�ye     • b	 rGX   ÿàõ Æ �y $"   • ¢  $x'A   fŽ Æ �yc$    • ¨  �yr    • b	 r(Gd   ÿà÷ ä rAj   ÿä â �ym&`   ‘ ¢  r"Gh   ÿàõ â/ $x)A   pŽ Ä �yg`   ‘ ¢ r G\   ÿà÷ är#A@   ÿä â �yi(`   ‘ ¢ r$G`   ÿàõ â $x!A   lŽ ä �ya"`   ‘ â  rGV   ÿà÷ ä/ r%An   ÿä â �ys `   ‘ b r*GZ   ÿàõ â $xA   ^Ž Ä �yo$`   ‘ b r+Ab   ÿä ä r"G   ÿà÷ â$t�ÿ$   ÿ Ž â r&G   ÿàõ â �yx`   ‘ d $x#   AŽ â r'	A   ÿä â �yu*`   ‘ b r(G   ÿàõ Är G   ÿà÷ â/ �yw&`   ‘ " r)A   ÿä är!A   ÿäÿ â �yy"`   ‘ " r,D>   Oàµ ä r$G   ÿàù â �y{(`   ‘ " r-UB   ÿd~ Æ $x%   A â rG   ÿàù â %v&, Z  � Ž â/ r–
G   ÿàõ â �y� `   ‘ $ $x--$   ÿŽ â rA   ÿä â�y}$`   ‘ " r(D>   MàÇ â $x''   -Ž â r—A   ÿä â �y‡`   ‘ " r*G   ÿàû Ä r UB   ÿ„þ â %v(( Z  � Ž â �y––`   ‘ " r$S&   ÿàõ Æ/ �y˜&    • " r%Q'   ÿä â $x+   AŽ ä $x $   ÿŽ â �yƒ&H   • " rS(   ÿàõ Æ �y‹$   ™ " $x))   Ž Æ �y‰*`   ‘ $ rQ)   ÿä ä �yŠ($   • ( �y„   ™ ( �y…$(   ™ ( �y�(    • ( �y~(   ™ ( �y—&$   • ( �y$L   ™ " r"G   ÿàõ Ä r*G   ÿà÷ â �y‚L   ™ $ $x#   A â r+A   ÿäÿ â �y•$p   ™ " r D>   =àµ är,D>   CàÙ â �y™p   ™ " r-UB   ÿd~ â %v   Z  � Ž Ä �y€(H   • $ $x�-$   ÿŽ â r-D>   Eàµ â �yz"l   • $ $x!!   �Ž ä �yˆ&l   • " %v$- Z  � Ž Æ �y†(l   • " %v", Z  � Ž Æ/ �y|*l   • " r&UB   ÿd~ ä r,S    ÿà÷ â �yœ     • & r-Q!   ÿäÿ â �y“ $   • ( �y¬,   ™ " rŒUB   ÿ¤~ Æ �yª,(   ™ " r*D>   ?àë ä/r(D>   IàÇ â $x+Œ$   ÿŽ â �y’ H   • " $x'&$   ÿŽ â rUB   ÿÄþ ä �y” l   • " $x%%   'Ž â r##+   ÿàÿ â $x'$   ÿŽ â �y¦,L   ™ " %v** Z  � Ž Æ �y©,p   ™ " r!UB   ÿ„þ är&D>   KàÇ â �y›"    • & rUB   ÿ„þ â $x++   'Ž â rS"   ÿàõ â %v&& Z  � Ž â�y�"$   • & $x$   ÿŽ â r S$   ÿàû â %v(( Z  � Ž â �y‘"H   • & $x''   Ž â rQ#   ÿä â $x!!$   ÿŽ â �yŒ"l   • ( �y­   ™ ( �y¯(   ™ ( �y®L   ™ " r))!   ÿàÿ â $x!Q   %Ž Ä �y±p   ™ ( �y²    ™ " r"S*   ÿà÷ Æ �y¡$    • " r#Q+   ÿäÿ Æ �y° (   ™ ( �y´"   ™ ( �y¥*    • " r,S&   ÿàõ Æ/ �yš$$   • ( �y�$H   • ( �yŽ$l   • ( �y *$   • ( �yŸ*H   • " r$S(   ÿàù Æ �yž*l   • ( �y¨(    • ( �y§($   • ( �y¤(H   • ( �y�(l   • ( �y«&    • ( �y£&$   • ( �y*&H   • (# �y¢&l   • ( �y( L   ™ " $x%Q   ) Æ �y)"L   ™ ( �y&"(   ™ (# �y+ p   ™ ( �y³"p   ™ " xttT  k    ÄO xqqT  v    â� �y$L   ™ è �y $   ™ è �y!$(   ™ â xÿ   t  Ä rqÿq   ÿ3Ž ä r-Q'   ÿä ä x¶ÿÀŽ â �y'$p   ™ â rÿD   q  ä xµ¶€€€€ÿüŽ â �y",   ™ â  xvÿÀŽ Æ �y#,L   ™ ä  $x¹µ   v
Ž â xµ     â �yk,(   ™ â  x$ÿ   t  äOx·ÿ   t  â �yq,p   ™ ¢  x%¶øŽ ä xvt€€€€µìŽ Ä xµ$ÿÀŽ â $x$   ÿ Ž â xº·ÿÀŽ ä r¸%¹   ¶`Ž ä x-ÿ     ä x%   ÿ   ä x$ÿÀŽ Ä x$º€€€€ÿüŽ ä x½-ÿÀŽ ä x·µ€€€€ÿüŽ ä x,%ÿÀŽ â $x-v   
Ž ä $xv$   ½
Ž â r·,·   ÿáÿ È r,µ,   ·XŽ ä r½º½   vXŽ ä xtøŽ äx$-txŽ ä x,,˜º  ÿ    ä xµµ˜º  ÿ    Ä x½½˜º  ÿ    ä r$   ÿÀŽ ä x$,€€€€·Ž äxµ,µ`Ž ä x,½€€€€vŽ ä xvcT       Ä xeeT  r    äxttÿÀŽ ä x ˜º  ÿ    ä xcÿ   v  ä reÿe   ÿ3Ž ä xt˜º  ÿ    Ä x»¶˜º  ÿ    ä x- €€€€-Ž äx¶cÿÀŽ ä reÿD   e  ä x  `Ž ä xºº˜º  ÿ    Ä xr¶€€€€ÿüŽ ä xeÿÀŽ ä r- -   ÿüŽ ä xcÿ   v  äx ÿ   v  â $t·ÿÿ Ž â x½½º`Ž Ä r$µ$   ÿüŽ ä x¸¸˜º  ÿ    ä rµr   ÿáÿ ä xºcÿÀŽ â $xce   ÿ Ž â x¶eøŽ ä xt ÿÀŽ Ä x e   ÿ   ä x%¸€€€€¹Ž ä xrv€€€€·ìŽ ä r·µ   ¶`Ž ä x¹¶˜º  ÿ    ä xeÿ   e  Ä x¶t€€€€ÿüŽ ä x ÿÀŽ ä xccÿÀŽ ä x¸¸»`Ž ä x»eÿÀŽ â $x¶¶   c
Ž â rer   ÿáÿ Ä xr·˜º  ÿ    ä r%¸%   ÿüŽ ä xvøŽ äx evxŽ ä xµr€€€€µŽ äx¸r¹`Ž Ä rrtc   ¶XŽ ä r     ÿÀŽ ä xr˜º  ÿ    ä xtt˜º  ÿ    ä xvvÿÀŽ ä rc¸µ   ÿüŽ Ä x¶€€€€¶Ž äxµt`Ž ä x  ˜º  ÿ    ä xv˜º  ÿ    ä rmÿH   m  ä rgÿF   g  Ä ryÿH   y  â$xmm   ÿ Ž â xt €€€€eŽ äx `Ž ä rwÿF   w  ä raÿJ   a  ä riÿL   i  Ä x g   ÿÀŽ ä reµ¶   ÿüŽ â $x¶y   ÿ Ž â rgt   ÿüŽ ä roÿN   o  ä rsÿP   s  ä xww   ÿÀŽ Ä xaa   ÿÀŽ ä xti   ÿ   ä x m0    âŽ â $xvs   ÿ Ž â xo   ÿÀŽ ä xi¶0   wâŽ ä xot0   aâŽ Ä xa àÿÿÿÿàÿ â &r -‹   ÿ   âruÿR   u  â &rtg‹   ÿ   â rxÿW   x  ä xiiàÿÿÿÿàÿ ä xsv0   âŽ â $ra    ÿŽ â r‰ÿR   ‰  Ä r–ÿW   –  ä r{ÿJ   {  ä r}ÿL   }  â $rtit   ÿŽ â xuu   ÿÀŽ ä xxx   ÿ   ä r‡ÿP   ‡  Ä r�ÿN   �  ä x‰‰   ÿÀŽ ä x––   ÿ   â 0r ÿ˜    A   â x{{   ÿÀŽ ä xv}   ÿ   ä xux0   uâŽ â $xx‡   ÿ Ž â Erm        Ä x��   ÿÀŽ ä Er t       ä x‰–0   ‰âŽ â &r–-„   ÿ   âx{v0   {âŽ â &r„g„   ÿ   â #rt m   ÿ   â x�x0   �âŽ â #rx    ÿ   â xmoàÿÿÿÿàÿ â 0ryÿŠ    A   â xo{àÿÿÿÿàÿ â &rv$…   ÿ   Ä &rŠe…   ÿ   ä $r–a–   ÿŽ ä $ri„   ÿŽ ä $rvmv   ÿŽ â Er{ –       â $rŠoŠ   ÿŽ â Er        â 0r ÿ�    A   Ä &r˜$~   ÿ   â Erv v       â 0r—ÿ—    A   â Erw Š       â $r˜m˜   ÿŽ â #r„ {   ÿ   â#rŠ    ÿ   â xssàÿÿÿÿàÿ â &r %   ÿ   â#rv—v   t   â Er˜ ˜       â &rc   ÿ   â xt�àÿÿÿÿàÿ â &rc‚   ÿ   Ä $r s    ÿŽ â x¼º€€€€ÿüŽ â #rx—w   x   â #rwy˜   „   â $rt   ÿŽ âEr          â $rt   ÿŽ â r,½,   ÿüŽ â 0rƒÿƒ    A   ä &r„e~   ÿ   â r¼»¼   ÿáÿ â &r~%‚   ÿ   â Er        â $r„o„   ÿŽ â Er‚        â #rƒ    v   â xuuàÿÿÿÿàÿ â &r ,•   ÿ   â r»º»   ¼XŽ â #rƒ   x   â Er„ „       â $r~s~   ÿŽ â x»»˜º  ÿ    â &rx,™   ÿ   â xºº˜º  ÿ    â $r u    ÿŽ â xr»€€€€¼Ž â 0r€ÿ€    A   â #ryy„   Š   â $rxux   ÿŽ â x»»º`Ž â 0rˆÿˆ    A   â Er~ ~       Ä Er          â #r{€‚   y   ârr»r   ÿüŽ â 0r†ÿ†    A   â Ery x       â 0rzÿz    A   â #rw€~   w   â #r ˆ       â xv‰àÿÿÿÿàÿ â &r™r™   ÿ   â #rw†y   w   â #r_z    _   â &r -¬   ÿ   Ä &r•r•   ÿ   ä &r¬g¬   ÿ   â #r<zw   <   â $r™v™   ÿŽ ä0rwÿ|    A   ä $ra    ÿŽ ä $r•v•   ÿŽ Ä &r|$ª   ÿ   ä $r¬i¬   ÿŽ â Er~ ™       â $r|m|   ÿŽ â Ery        â 0r ÿœ    A   â Erx •       ä Er ¬       â #r~†~   {   â &rªeª   ÿ   â Er{ |       â #rxˆx      â #r| y   ÿ   â#r€    ÿ   â &r %¦   ÿ   â #r;wx   ;   â $rªoª   ÿŽ Ä &rc¦   ÿ   ä $r s    ÿŽ ä &rx,©   ÿ   ä 0r“ÿ“    A   â Er} ª       â $rt   ÿŽ â Er          â $rxux   ÿŽ â #r{“{   |   â0r’ÿ’    A   â Er|        â #r}“}   €   â &r©r©   ÿ   â Erx x       â 0r”ÿ”    A   â #r{’    {   â#r}’|   }   â $r©v©   ÿŽ Ä &r|-­   ÿ   â#rx”x   {   â #r[w~   [   â &r~g­   ÿ   â Er  ©       â $r|a|   ÿŽ â #rYzx   Y   â &rx$¯   ÿ   ä$ri~   ÿŽ â Ery |       â &r~e¯   ÿ   â #r}”    }   â &r|%®   ÿ   Ä $rxmx   ÿŽ ä 0r ÿ›    A   â Er        â $r~o~   ÿŽ â Erx x       â $r|s|   ÿŽ â #ry y   ÿ   â0r�ÿ�    A   â Er~ ~       â #r    ÿ   â 0r‘ÿ‘    A   â Er| |       â #rx�x   y   â &r ,±   ÿ   â #r~�~      Ä #r‘|   x   â $r u    ÿŽ ä &rx-²   ÿ   ä&r²g²   ÿ   â Er          â 0rŒÿŒ    A   ä $ryax   ÿŽ ä $r²i²   ÿŽ â #rxŒ       â &r{c®   ÿ   â Ery y       â 0r ÿ¡    A   â Er ²       â $r{t{   ÿŽ Ä &r±r±   ÿ   â #r:w}   :   â #r]zx   ]   â #ry y   ÿ   â#r}    ÿ   â $r±v±   ÿŽ â Er{ {       â &r $°   ÿ   ä &rx-´   ÿ   ä&r´g´   ÿ   â Er± ±       â $rm    ÿŽ â #r{‘{   ~   â $r|ax   ÿŽ Ä $r´i´   ÿŽ â Erx        â #r{Œ±   {   â 0r ÿ¥    A   â Er |       ä Er ´       â #r6w{   6   ä #r    ÿ   ä#r{    ÿ   â &r %(   ÿ   Ä &r|$&   ÿ   ä&r&e&   ÿ   ä &rc(   ÿ   ä $r s    ÿŽ ä 0ršÿš    A   ä $r|m|   ÿŽ Ä &r(%)   ÿ   ä $r&o&   ÿŽ â Er          â &r°e°   ÿ   ä $rt   ÿŽ â #rxšx   y   â 0r�ÿ�    A   â Er| |       â 0r ÿ     A   â Er& &       â $r(s(   ÿŽ â Ery        â $r°o°   ÿŽ â #r�    x   â 0rŸÿŸ    A   â Er( (       â #r| |      â &r ,+   ÿ   â#r  &   {   â &r+r+   ÿ   â Er° °       â &r&,³   ÿ   â #r{Ÿ(   |   â $r+v+   ÿŽ Ä &r)c)   ÿ   â #r°š°   }   â $r(u&   ÿŽ â Er& +       â 0rŽÿŽ    A   â #ry�y   °   â $r)t)   ÿŽ â Er( (       â &r³r³   ÿ   ä 0ržÿž    A   â Er) )       â $r u    ÿŽ â #r&Ž&   y   â $r³v³   ÿŽ â #r(ž(   {   â #r)Ÿ)       â Er          â #r4w&   4   â Erx ³       â #r3z(   3   â &r&-    ÿ   Ä� &r(g    ÿ   â #r Ž       â #r)žx   )   â &r $!   ÿ   ä$r&a&   ÿŽ ä $ri(   ÿŽ ä &rxe!   ÿ   â #r5z    5   â $r m    ÿŽ â Er! &       â 0r ÿ¨    A   â #r2w)   2   â $rxox   ÿŽ â Er        â &r(%   ÿ   Ä &r)c   ÿ   â Er          â 0r§ÿ§    A   â #r !   ÿ   â$r(s(   ÿŽ â Erx x       â #r     ÿ   â $r)t)   ÿŽ â #r§       â0r¤ÿ¤    A   â Er( (       â #r§§x       â Er )       â &r -"   ÿ   Ä &r"g"   ÿ   â #r(¤(      â$raa    ÿŽ â #r¤¤   §   â &r %#   ÿ   ä $ri"   ÿŽ ä &r$$k   ÿ   ä&rek   ÿ   â Era a       â $rss    ÿŽ â Er        â &r#c#   ÿ   Ä $r$m$   ÿŽ ä $roo   ÿŽ ä 0r ÿ«    A   â x99   ÿàÿ â &r,'   ÿ   ä&r,,q   ÿ   âO Er$ $       â $r#t#   ÿŽ â #r  a   ÿ   â &r'r'   ÿ   âEro o       â &rqrq   ÿ   â #r     ÿ   â 0r£ÿ£    A   â | 9   pbô â $ru   ÿŽ âErs s       â $r,u,   ÿŽ â Er# #       â $r'v'   ÿŽ Ä $rqvq   ÿŽ â #r£$       â0r*ÿ*    A   â #ro£o       â 0r�ÿ�    A   â Er        â 0r¢ÿ¢    A   â Er' '       â #r*s      âEr, ,       â #r#*#   o   â Er  q       â #r�   (   â xGG¸  ÿà÷ â #r¤�'   ¤   â #r¢,      â#r#¢    #   â #r7z   7   â#r1w¤   1   â#r0z   0   â #r.w#   .   â x==    ÿàÿ â $rAÿÿ   AŽ â xCC    ÿàÿ Ä xEE    ÿàÿ ä x??    ÿàÿ ä xII    ÿàÿ ä xKK    ÿàÿ ä xMM    ÿàÿ ä xOO    ÿàÿ â D)       À â Gy  PÖÿÿÿÿƒ ê Ay        € ê Ey  0    € â G‰      € ê r88/   ÿHŽ Ê ˆs 8_ øÿ H   è ˆs 8;€øÿ H   è ˆs 8< ùÿ H   è ˆs 8[€ùÿ H   è ˆs 8Y úÿ H   è ˆs 8:€úÿ H   è ˆs 8] ûÿ H   è ˆs 86€ûÿ H   è ˆs 85 üÿ H   è ˆs 84€üÿ H   è ˆs 83 ýÿ H   è ˆs 82€ýÿ H   è ˆs 87 þÿ H   è ˆs 81€þÿ H   è ˆs 80 ÿÿ H   è ˆs 8.€ÿÿ H   ä Ay        € ê {          ì M        € ê „y /     H   b z?  r      Æ „y/ €   H   ¤ $x9?   ÿŽ ä„y/    H   â $x=?   ÿŽ Æ „y/    H   ( „y/ €  H   h y	      %   ( „y/ €  H   h „y
/ €  H   h „y/    H   è „y/    H   â !r  _       Æ/ „y/    H   â !r;;       Æ_ ‰       " $x	   ÿ Ž â !r<<       ä� ‰;      d z  r  p`ð â !rY       ây      '   ¤ r /ÿ   p"p â !r::       â‰<      è ‰      â !r66       Æ ‰:      " !r

[       â !r        ä „y/ €  H   â !r]]       â !r;       ä/ y       &   " $x;?   ÿŽ â !r33       Ä ‰      b !r<<       Æ� ‰	      â !r55       Æ „y/ €  H   b !r::       Æ„y/    H   è „y/ €  H   â $z   x  Ž Æ „y/    H   â x   ÿàÿ â $z  ~   Ž âO !r       â/ ‰<      & 
  r  ÿàó â ‰ €    b r?    ÿ† â !r44       â 
 j  ÿ„ â $ÿÿ   ÿŽ  â ‰
      ¢ z  r  p`ò â !r       â rÿÿ   ÿäÿ â ‰]      " 
 k   â !r	       â� z j  ÿ† â ‰      â x /   p"ò Ä z k  � â ‰ €    â  ;   ÿà÷ â !r22       â!r<<       â ‰ 5      " !r77       â $ÿÿ   ÿŽ â !r       â/ „y/ €  H   â !r11       â $x?   ÿŽ âr	?    ÿŠ â !r
       âO ‰ @    b  9   ÿàù â !r00       â 

 j  ÿ† â ‰6      ¢ !r]]       â$ÿÿ   ÿ â 
 j  ÿˆ â ‰4      " 
 k  � â !r       Ä� ‰"3      â 
 k   â !r       â r    ÿàõ â ‰2      b rÿÿ   ÿäÿ â !r 5        â rÿÿ   ÿä â ‰$7      " z j  ÿ„ ä =   ÿà÷ â ‰&1      " z k   â !r       â/ ?   ÿàÿ â ‰      b z	 j  ÿŠ â!r66       âO *    ÿàõ â ‰!0      ¢ !r44       âz		 k  � â $ÿÿ   ÿŽ â ‰:      " !r"3"       â� 
 j  ÿˆ â !r..       â ‰]      â !r22       â
 k   Ä ‰ @    b !r$7$       â 3ÿÿ   ÿä â !r&1&       â ‰#6      " 9   ÿàÿ â !r       ä/ ‰%       b * j  ÿ„ â !r00!       ÄO ‰'4      ¢ * k  3 â !r::       ä‰)"      "     ÿà÷ â !r]       ä� ‰< €    ä $(ÿÿ   ÿŽ â !r       â‰+2      b  j  ÿ† Æ ‰-$      b !r66#       â  k  (� ä ‰ €    " !r  %       Æ/ ‰/&      b !r44'       ÆO ‰10      ¢ !r"")       Æ‰.      " !r<<       Æ� ‰      â !r22+       Æ‰! €    b !r*$-       Æ ‰ €    b !r       Æ ‰: €    " !r,&/       â/ !r001       äO ‰%6 €    b !r..       Æ‰'  €    ¢ !r7       Æ� ‰)4 €    â ;   ÿàÿ â !r!       ä‰< @    " x!        â !r       ä ‰+" €    b     ÿàõ â%v  j  ! Ž â !r::       â ‰-2 €    $ $$ÿÿ   ÿ â  j  ÿ„ â ‰ @    " !r%6%       Ä/  k  $ â ‰/* €    b !r' '       ÆO ‰1, €    ¢ !r)4)       Æ� ‰30 €    â !r<       Æ‰5.      " !r+"+       Æ‰      b !r-2-       Æ ‰ @    " !r       Æ ‰!: @    " !r/*/       Æ/ ‰# @    b !r1,1       ÆO ‰ % @    ¢ !r303       â� !r5.5       ä‰"' @    â !r9       Æ‰$) @    " !r       Æ ‰      â !r!:!       Æ ‰&+ @    " !r##       Æ/ ‰(- @    b !r %        ÆO ‰      ¨ ‰*/ @    b !r"'"       Æ� ‰,1 @    â !r$)$       Æ‰05 €    ( ‰.3 @    " !r&+&       Æ ‰      " !r(-(       Æ/ ‰!      b !r'       ÆO †	 7    â !r*/*       â!r,1,       ä� † 9   â !r050       Æ‰       â !r7       âO !r.3.       ä ‰#      ¢ !r)       Æ ‰"      " !r+!       Æ/ ‰$      h †	 7    è ‰&      " !r        Æ� ‰(      è † '    â !r##       ÆO ‰*      ¢ !r"       Æ ‰!,      " !r$       Æ/ ‰%0 @    h ‰.      b !r&       Æ†	 )    â	 !r(       Æ� † +   è	 †	 #    â	 !r*       âO !r!,!       ä †     â	 !r%0%       Æ/ †	     â	 !r.       Æ†    è	 †	 
    è	 †     è	 †	     è	 † !   è	 ‰%      ( †	     â	 M™        € ê x=   ÿàÿ â!r%%       Æ r    ÿàñ È r ÿÿ   ÿä  ä z j  ÿ€ È z k     Ê †y %    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y       "   " ¹z  l      â Ey  00    € â ™x?    â y      !   b ¹z  F   
   â x      ÿ  â r ÿ       â $r,ÿÿ   ÿ Ž â x      ÿ  â x
      ÿ  â $rjÿÿ   ÿ Ž â rt ÿ       â $ryÿÿ   ÿ Ž â r ÿ       â $r~ÿÿ   ÿ Ž Ä $r*ÿÿ   ÿ Ž â r  ÿ   pRð är  ÿ   pRò â$x     Ž Ê/ xÿ     È r   ÿ � È xÿ     È |    pbô Ú G)  À.    € ê y      %   " xðÿÿÿÿÀŽ â $vÿ X  ÿ Ž â r ÿ       â y      '   b z  q      â $x   
Ž ä y      &   ¤ $x_   ÿ Ž â xÿ     È r   ÿ� â $rÿÿ    Ž Æ xøÿÿÿÿÀŽ â $x   ÿ Ž â xÿ     Æ $x   
Ž â x   ÿ   â $rÿÿ   ÿ Ž â x   ÿ   â 'z y   Ž â/xÿ     äxÿ     â $z }  ÿŽ â zÿ z    â 'z s   Ž âO r   ÿ� Ä x     â $z w  ÿŽ â zÿ t    â $z p  Ž âr   ÿò� â $z |  Ž â x   ÿàÿ â $z |  Ž â xÿ     â $z v  Ž âx@   ÿàÿ â $zN v  Ž â r   ÿ� â$z± p  Ž â r   ÿ� â %xNNn   ÿŽ â x:üÿÿÿÿÀŽ Ä x>øÿÿÿÿÀŽ â $x   Ž âx2   ÿàÿ â$x   Ž âxD   ÿàÿ â$x>   >
Ž âx:   :èÿ â $x   Ž â xÿ   @  â $x]   ÿ Ž â xÿ   2  â %xZ_   NŽ â xÿ   D  Ä r%@   ÿ� â %xX]   NŽ â r'2   ÿ� ä r#D   ÿ� â$rÿÿ   ÿ Ž â r$D   ÿ� â $x   ÿ Ž â r@   ÿ� ä r(2   ÿ� Ä xC%øÿÿÿÿÀŽ ä x7'øÿÿÿÿÀŽ ä x#øÿÿÿÿÀŽ â $xB@   C
Ž â xE$üÿÿÿÿÀŽ â $x62   7
Ž â xAüÿÿÿÿÀŽ ä x3(üÿÿÿÿÀŽ Ä rD  €ÿàÿ äxDD   Eèÿ ä x@@   Aèÿ ä x22   3èÿ ä x;ÿ   :  ä x?ÿ   >  Ä x3ÿ   2  â %xP±n   :Ž âx7ÿ   6  ä xAÿ   @  â %xR±n   >Ž âxCÿ   B  ä xEÿ   D  â %x®±n   2Ž â xÿ     Ä r   ÿàõ â $x   Ž â r+2N   ÿà÷ â%x¬±n   6Ž âr/6N   ÿàý är):N   ÿàû â %xª±n   @Ž âr   ÿŽ ä x"ÿ     â %x¦±n   BŽ â x#ÿ   #  Ä x%ÿ   %  â %xT±n   DŽ âx&ÿ     ä x(ÿ   (  â %xV±n   Ž â r;O;   ÿäÿ ä x ÿ     â $rÿÿ   ÿ â r->N   ÿàõ â %x°±n   ÿŽ â x$ÿ   $  Ä x'ÿ   '  â $x3O   3Ž âr5BN   ÿà÷ â$x7O   7 â r9DN   ÿàý â%x\]   °Ž âr1@N   ÿàù är=N   ÿàû â %x^_   °Ž â z  Y      Ä z q  ÿàÿ â $x""   ÿ Ž â x!ÿ     â $x##   ÿ Ž â x     ÿ   â $x%%   ÿ Ž â x$$   ÿ   â $x&&   ÿ Ž â x''   ÿ   â $x((   ÿ Ž â rAOA   ÿä â$x?O   ? ârO   ÿäÿ â$xCO   CŽ Ä $xEO   E ä rJX   ÿàõ Ê $xKY    â rZ   ÿàõ È �y@J    • ¨  �yGJ   • ¢  $x[    Ê �y¥"   • è �y0    • â rH\   ÿàõ È rI]   ÿä ä rd^   ÿàõ Æ �yaH    • h	 �y6H   • b	 $xe_    â rJ-   ÿàõ È �yŸd"   • â  rL)   ÿà÷ Æ �y4d    • b rM;   ÿäÿ â$xK   ? â rH9   ÿà÷ är–=   ÿàõ ä �ykJ`   ‘ b	 $xI   EŽ â r1   ÿà÷ â/ $x—    â rb5   ÿàõ â $tÿ$   ÿ Ž â �yML`   ‘ b $x   AŽ ârcC   ÿä Ä rdR   ÿàõ â�y�H`   ‘ b r”V   ÿà÷ Æ $xeS    ârJP   ÿàù â$x•W   Ž â rh   àµ â �y›`   ‘ b	 rKQ   ÿä ä r!   ÿd~ â %vhh Z   Ž â �ygd`   ‘ f $x$   ÿŽ â �y™J`   ‘ b rª   ÿàõ Ærii   ÿàÿ â �y––`   ‘ b	 $x«    â rrh   ÿàõ ä rHT   ÿàù â �y””`   ‘ d $xs   i â rIU   ÿä â �ybb`   ‘ b r†/   ÿàõ Ä r�¦   ÿàû â �y—H`   ‘ h	�y•r   ™ b $x‡   7 Æ �yJr(   ™ b! rd+   ÿà÷ â $x‘§   Ž â rl®   ÿàõ â �yHh    • "re3   ÿäÿ Æ �y�`   ‘ " $xm¯    Æ �y��`   ‘ ( �yLh$   • " r   àµ Æ �y††`   ‘ " r!   ÿd~ Æ �yed`   ‘ " %v Z   Ž Æ �yŽrL   ™ " $x$   ÿŽ â rp   ÿàõ ä rn¬   ÿàù â �y‘hH   • "# r   ÿàÿ Æ $xo­    â �y“l`   ‘ " $xq    Æ �y’n`   ‘ ( �y…p   ™ " r   àÇ Æ �yFp(   ™ ( �y.hl   • " rnN   ÿàõ Æ �y‰rp   ™ " r!   ÿ„þ Æ �yŒpL   ™ ( �yŠ    • " $xoO    â rl°   ÿàõ â %vh Z   Ž â/ �y‹$   • " r2   àÇ ä $xI$   ÿŽ â �y‡pp   ™ " rm±   ÿä Ä r!   ÿ„þ â �yˆH   • " rvh   ÿàõ â $xii   IŽ ä �ynl   • " $x$   ÿŽ ä %vp2 Z   Ž â �y�l   • & $xw   i â rqq   ÿàÿ â �y|h    • ( �yv   ™ " rr   àë Æ �y„v(   ™ " r!   ÿÄþ ä/ r²p   ÿàõ â �y�h$   • ( �yƒhH   • ( �yhl   • ( �y‚vL   ™ " %vrr Z   Ž Æ �y}vp   ™ " rh   àë â $xI$   ÿŽ ä �y€ll   • " $x³   q â r!   ÿÄþ â $xss   IŽ â r¨r   ÿàõ â �yp    • $ $xI$   ÿŽ Ä �y²   ™ " $x©   s Æ �yz²(   ™ ( �y{¨   ™ ( �yop$   • ( �ylr    • ( �ynpH   • ( �ympl   • ( �yx²L   ™ " r    àÇ Æ �ywr$   • (# �yp¨(   ™ " r2!   ÿ„þ â %v   Z   Ž ä�yvrH   • $ %vhh Z   Ž Ä �yqrl   • $ $x2$   ÿŽ â r¢    ÿà÷ â�yu¨p   ™ ( �ys²p   ™ (# �yr¨L   ™ " r¡¡   ÿàÿ â $xii   IŽ â r´h   ÿàù Ä r£¡   ÿäÿ â�y     • $ $xµ   i ä �y>h    • ( �y<h$   • ( �y8hH   • ( �y:hl   • " x@@T  G    ÆO �yd¢(   ™ ( �yf $   • ( �yi¢   ™ ( �yB H   • ( �y2 l   • ( �yh¢L   ™ " x¥0T  ¥    Æ� �yG´   ™ â r¡ÿ@   ÿ3Ž â/ $t0ÿÿ Ž ä �yc¢p   ™ â r¡ÿ   ¡  Æ �yD´(   ™ â x@¡   ÿ   ä xI¥€€€€0ìŽ ä x@@ÿÀŽ È xK@¥øŽ â$x0I   @
Ž ä �y@´L   ™ è �yI´p   ™ â x`ÿ   ¥  È x``ÿÀŽ ä x˜¡ÿÀŽ äx³`€€€€ÿüŽ ä x©ÿ   ¥  ä x£¡   ÿ   âO $x³³   ˜
Ž â x©©ÿÀŽ Ä x˜ÿ   ¥  äxž0¥xŽ ä xš`¡øŽ ä xœ©€€€€ÿüŽ ä x££ÿÀŽ ä x˜˜ÿÀŽ Ä x¡ÿ   ¡  ä ržžK   ÿÀŽ ä ršš³   ``Ž ä xK`˜º  ÿ    â $xœœ   £
Ž â x`˜€€€€ÿüŽ ä x¡¡ÿÀŽ Ä xšš˜º  ÿ    ä r£©£   œXŽ â $x``   ¡
Ž â x6aT  6    äx š€€€€³Ž äxKšK`Ž ä xš©˜º  ÿ    Ä x££˜º  ÿ    ä r©˜¡   `XŽ ä r¡ÿ6   ÿ3Ž ä rKK    ÿüŽ ä x £€€€€œŽ ä rœÿ   ¡  Ä xa¥ÿÀŽ ä x©©˜º  ÿ    ä xŸ4T  Ÿ    â $x4œ   ÿ Ž â x6˜˜º  ÿ    ä xžž˜º  ÿ    ä xaa˜º  ÿ    Ä x˜     ä x££š`Ž ä x¥©€€€€`Ž äx`ÿ   Ÿ  ä x6©6`Ž ä x¡ž€€€€0Ž Ä x0ža`Ž ä x©Ÿ€€€€˜ìŽ ä x44ÿÀŽ ä ra£    ÿüŽ ä x£`ÿÀŽ ä x¢ÿ   Ÿ  â $xš©   4
Ž â r66¥   ÿüŽ Ä x¥4ŸøŽ â $x4œ   ÿ Ž âx £€€€€ÿüŽ ä x©œÿÀŽ ä x¢¢ÿÀŽ ä r ©    ÿáÿ ä x©¢€€€€ÿüŽ Ä x44ÿÀŽ Ê $xž©   4
Ž â x`šŸxŽ È r4¢4   žXŽ ä r00¡   ÿüŽ ä r``¥   ÿÀŽ ä x¡ÿ   Ÿ  ä x¢¢˜º  ÿ    ä x44˜º  ÿ    Ä xŸŸÿÀŽ ä xž4€€€€žŽ äx©4¢`Ž ä x³`˜º  ÿ    ä x4Ÿ˜º  ÿ    ä rMÿ"   M  Ä r™ÿ"   ™  ä rkÿ    k  ärgÿ    g  ä xš³€€€€šŽ â $x`™   ÿ Ž â x³³4`Ž â $x4M   ÿ Ž â r�ÿ$   �  Ä r–ÿ#   –  ä xkk   ÿÀŽ ä xgg   ÿÀŽ â $x��   ÿ Ž â rš³š   ÿüŽ ä x––   ÿÀŽ ä x440   kâŽ Ä x``0   gâŽ â &rk0•   ÿ   âx–�0   –âŽ â &r•š•   ÿ   â x44àÿÿÿÿàÿ ä x``àÿÿÿÿàÿ ä rMÿ#   ”  â &rgaJ   ÿ   â r”ÿ$   —  â $rk4k   ÿŽ â x––àÿÿÿÿàÿ â $r—`•   ÿŽ â x¥£œøŽ â 0rHÿH    A   âErk k       â $r•–g   ÿŽ â r¥¥    £`Ž Ä Erg —       ä rbÿ%   b  ä r›ÿ&   ›  är�ÿ&   �  ä xMM   ÿÀŽ ä x””   ÿ   â #rkHk   ÿ   âx¥¥˜º  ÿ    â #rgHg   ÿ   â x˜£˜º  ÿ    â 0rLÿL    A   â xbb   ÿÀŽ Ä x››   ÿ   ä r�ÿ%   �  ä Er• •       ä rHÿ'   †  â $x��   ÿ Ž â r©©ž   ÿüŽ ä xM”0   MâŽ Ä x ¥€€€€ Ž äx¥¥˜`Ž ä xb›0   bâŽ ä x��   ÿÀŽ ä r†ÿ(   e  â #rkL•   k   â xeH   ÿÀŽ â &r•©J   ÿ   â xHMàÿÿÿÿàÿ â &rMKŽ   ÿ   â r¥¥    ÿüŽ Ä x��0   �âŽ ä xbbàÿÿÿÿàÿ â $r•H•   ÿŽ â x��àÿÿÿÿàÿ â &rŽ¥Ž   ÿ   ä $rMbM   ÿŽ â Er• •       â $r��Ž   ÿŽ â x¡¡ÿÀŽ â 0rŽÿ‘    A   â xœÿ   œ  Ä x††   ÿ   ä ErM M       â #r•L•   g   â x£¡€€€€ÿüŽ ä xœœÿÀŽ ä xg†0   eâŽ â #r†ŽM   k   â r“ÿ(   “  â &rk0…   ÿ   â r’ÿ'   ’  â &r‘š…   ÿ   â r£œ£   ÿáÿ â $x““   ÿ Ž â xJ’   ÿÀŽ â &r…aF   ÿ   Ä $rk4k   ÿŽ â rœ¡œ   £XŽ â 0reÿ.    A   â x.gàÿÿÿÿàÿ â &rM6‰   ÿ   â Er� �       â &rgKŒ   ÿ   â xJ“0   JâŽ â $r…–…   ÿŽ â Erk k       â 0rŠÿŠ    A   â xœœ˜º  ÿ    â &r“©F   ÿ   â x¡¡˜º  ÿ    â $r‘`‘   ÿŽ Ä $rF.M   ÿŽ â #rLŽ�   •   â $rgbg   ÿŽ â Er… …       â 0r‹ÿ‹    A   â x£œ€€€€£Ž â #rŽŠk   ÿ   â xœœ¡`Ž ä ErM ‘       â 0rˆÿˆ    A   â ErF F       â &rk6‡   ÿ   â Erg g       â #r…‹…   Ž   â rœœ£   ÿüŽ â $r“H“   ÿŽ â #rŠŠM   ÿ   â &rŒ¥Œ   ÿ   â #rMeF   †   â $rk.k   ÿŽ â #rFˆg   …   â xJJàÿÿÿÿàÿ â 0rgÿ    A   â Er“ “       â $rŒ�Œ   ÿŽ Ä &r‡œ‡   ÿ   â Erk k       â 0r�ÿ�    A   â #r*gM   *   â $r‡J‡   ÿŽ â ErŒ Œ       â &rM0   ÿ   â #r‹‹“   Š   â #rk�k   F   â Er‡ ‡       â $rM4M   ÿŽ â #rŒˆŒ   ‹   â &rš   ÿ   â #r~gk   ~   â &rka„   ÿ   â #r�‡   Œ   â 0r|ÿ|    A   â ErM M       â $r`   ÿŽ Ä &r‡©„   ÿ   ä &r…K‚   ÿ   ä$rk–k   ÿŽ â #rF|M   ÿ   â &r‰œ‰   ÿ   â Er        â $r‡H‡   ÿŽ ä &r‚¥‚   ÿ   â Erk k       â $r…b…   ÿŽ Ä &rM6}   ÿ   ä 0r�ÿ�    A   ä $r‰J‰   ÿŽ â Er‡ ‡       â $r‚�‚   ÿŽ ä &r}œ}   ÿ   â #r||   ÿ   â $rM.M   ÿŽ â Er… …       â 0rƒÿƒ    A   â #rF�k   F   âEr‰ ‰       â &rk0   ÿ   â Er‚ ‚       â $r}J}   ÿŽ â #r|�‡   |   â &rš   ÿ   â ErM M       â 0rÿ    A   â #rFƒ…   F   â #rLe‰   L   â $r`   ÿŽ â Er} }       â 0reÿ€    A   â #r|ƒ‚   |   â $rk4k   ÿŽ â #rFM   F   â &raz   ÿ   Ä &rM0{   ÿ   âEr        â &r{š{   ÿ   â #r|}   |   â 0rÿ    A   â #reL      â #rygF   y   â &r©z   ÿ   â ErL k       â $r–   ÿŽ ä $rF4M   ÿŽ â #re      â $r{`{   ÿŽ â #r   ÿ   â$rH   ÿŽ â Er        â 0roÿo    A   â #rLL   ÿ   â 0rlÿl    A   â Erk F       Ä Er {       ä ErM        â #ro   L   â #rlk   ÿ   â#rll   ÿ   â &rkap   ÿ   ä &rKx   ÿ   ä $rk–k   ÿŽ Ä $rb   ÿŽ ä 0rwÿw    A   â Erk k       â 0rnÿn    A   â Er        Æ #rkwk      ä #rn      â &r6s   ÿ   â #rMoM      â &roKr   ÿ   ä $r.   ÿŽ ä $robo   ÿŽ ä &r6u   ÿ   â Er        â 0rmÿm    A   â Ero o       â &rx¥x   ÿ   Ä $r.   ÿŽ ä 0rvÿv    A   â #rm      â $rx�x   ÿŽ â Er        â 0rqÿq    A   â #rvo   k   â &r0i   ÿ   ä&riši   ÿ   â Erx x       â #rjg   j   â #rq      â $r4   ÿŽ Ä $r`i   ÿŽ ä &ri©d   ÿ   â #rxnx   M   â #r	g   	   â 0rÿ    A   â Er        â &rad   ÿ   â Er        â &rMKh   ÿ   ä $riHi   ÿŽ Ä &rh¥h   ÿ   â #r   ÿ   â$r–   ÿŽ â Eri i       â 0rfÿf    A   â #r   ÿ   â $rh�h   ÿŽ â Er        â 0rBÿB    A   â #rfi      â &ri0G   ÿ   â�Erh h       â &rGšG   ÿ   â #rf      â &r{©p   ÿ   Ä #rBh      â $ri4i   ÿŽ ä $r``G   ÿŽ ä &raaD   ÿ   ä&r©©D   ÿ   ä &r6c   ÿ   ä &rKK@   ÿ   â Eri i       â $r{H{   ÿŽ Ä &rr¥r   ÿ   ä&r@¥@   ÿ   ä $ra–a   ÿŽ ä $r©H©   ÿŽ ä $r.   ÿŽ â Er `       â 0r>ÿ>    A   â x   ÿàÿ â &rcœc   ÿ   Ä &rG6I   ÿ   â Er{ {       â &rsœs   ÿ   âEra a       â $rr�r   ÿŽ â Er ©       â &ruœu   ÿ   ä $rMbM   ÿŽ ä$rKbK   ÿŽ Ä $r@�@   ÿŽ ä &rIœI   ÿ   â #ri>i   ÿ   â0r<ÿ<    A   â #r>   ÿ   â $rcJc   ÿŽ â |    pbô â $rsJs   ÿŽ âErr r       â $ruJu   ÿŽ â ErM M       â $rG.G   ÿŽ â ErK K       â $rIJI   ÿŽ â Er@ @       â 0r8ÿ8    A   â #r{w{   l   â #ra<a   i   â#r<      â 0r2ÿ2    A   â Er c       â 0r:ÿ:    A   â Ers s       â #rrvr   {   â Eru u       â #rBM      â Er        â #rK8K   a   âErG G       â #r8@      â Er. I       â #r2      â#rsms   x   â #ruqu   r   â #r2      â #r:G   K   â#r:.      â x¸  ÿà÷ â #rte|   t   â#r
es   
   â#reu      â #r,g   ,   â #re      â #rg      â #re      â x    ÿàÿ Ä x    ÿàÿ ä x    ÿàÿ ä x    ÿàÿ ä x    ÿàÿ ä x    ÿàÿ ä x    ÿàÿ Ä rÿ   ÿäÿ â D)       À â Gy  °Ùÿÿÿÿƒ ê Ay        € ê Ey      € â G‰  ð     € ê $x À  Ž Ê ˆs * ùÿ H   è ˆs €ùÿ H   è ˆs ~ úÿ H   è ˆs €úÿ H   è ˆs y ûÿ H   è ˆs t€ûÿ H   è ˆs j üÿ H   è ˆs 
€üÿ H   è ˆs 	 ýÿ H   è ˆs €ýÿ H   è ˆs , þÿ H   è ˆs €þÿ H   è ˆs  ÿÿ H   è ˆs €ÿÿ H   ä Ay        € ê {          ì M        € ê „y     H   " $v3ÿ r  ÿ Ž Æ „y €   H   b $x13   ÿŽ ä$x/3   ÿŽ â „y    H   ¨ „y €  H   è „y €  H   ( „y    H   h „y    H   h „y €  H   h „y €  H   è y      %   b !r*       Æ „y    H   â !r       Æ/ ‰       " !r~       ÆO „y    H   b !r       Æ� „y €  H   ¢ !rtt       Æ„y    H   â !ry       Æ„y# €  H   " $x%   ÿ Ž â !rjj       ä y      '   d z % r  p`ð â !r       â !r        â ‰      " r ÿ   p"p â !r
       Ä y       &   ( ‰      b !r		       Æ ‰      h ‰      b !r,       Æ/ ‰      b !r       ÆO ‰j      ¢ !r!       â� $z   x  %Ž ä ‰t      " x%%   ÿàÿ â !r##       â$z  ~   Ž â!r       â !r       â ‰      â !r       â 
  r  ÿàõ Ä ‰      " !r       â z % r  p`ò ä ‰
      b !r       â/ ÿÿ   ÿä ä ‰ €    b !rj       âO 
 j  ÿ„ Ä ‰ 	      ¢ x    p"ò â !rt       â 
 k   â ‰"      ( ‰$      " !r       Æ� ‰      â !r       Æ‰&      " !r
       Æ‰(!      b !r       â/ r3    ÿ„ ä ‰*#      b !r 	        âO 	 /   ÿàû ä ‰      ¢ !r""       ä $,ÿÿ   ÿŽ â ‰      " !r$$       â 
	 j  ÿŠ Ä ‰      " !r       â� r3    ÿˆ ä ‰ €    â !r&&       âr.ÿÿ   ÿä â !r(!(       â‰ @    â z
 j  ÿˆ â !r*#*       Ä/ ‰!      b z k  . â !r       äO ‰#       ¢ 
		 k  ,� â !r       ä ‰%"      " !r       Æ ‰'$      "  1   ÿà÷ ä rÿÿ   ÿä â ‰ €    b !r       â� $ÿÿ   ÿŽ â z j  ÿ„ â ‰)&      â 
 j  ÿ† â $x3   ÿŽ â z k   â ‰ €    â 
 k  � â !r!!       â/ !r       â ‰+(      b !r# #       âO r    ÿàõ Ä ‰ €    ¢ !r%"%       â 3   ÿàÿ ä ‰-*      " !r'$'       âz j  ÿˆ ä ‰ €    " !r       â/   ÿàÿ Ä ‰ €    b !r)&)       â� /    ÿà÷ â!r       â rÿÿ   ÿä â ‰"! €    â .    ÿàõ â !r+(+       â/ z k   â ‰$# €    b 0ÿÿ   ÿäÿ â !r       âO $ÿÿ   ÿ â ‰&% €    ¢ . j  ÿ„ â !r-*-       â / j  ÿ† â‰(' €    " . k   â !r       â/ k  0� â ‰ @    " !r       Æ‰*) €    h ‰,+ €    b !r"!"       Æ� ‰ @    â !r$#$       Æ/ ‰ @    b !r&%&       ÆO ‰.- €    ¢ !r('(       Æ ‰ @    " !r       Æ‰      " !r*)*       Æ‰ @    b !r,+,       Æ ‰  @    b !r       Æ� ‰!" @    â !r       Æ/ ‰#$ @    b !r.-.       ÆO ‰%& @    ¢ !r       Æ ‰'( @    " !r/       â1   ÿàÿ ä ‰      " !r       â    ÿàõ â$tÿ   ÿ Ž â ‰)* @    b !r         â 0ÿÿ   ÿä â %v  j   Ž â ‰+, @    b !r!"!       â�  j  ÿ„ Ä ‰      â !r#$#       â/  k  0 ä ‰-. @    b !r%&%       ÆO ‰      ¢ !r'('       Æ ‰      " !r1       Æ‰      " !r)*)       Æ‰       b !r+,+       Æ ‰ !      b !r       Æ� ‰#      â !r-.-       Æ/ ‰"%      b !r       ÆO ‰'      ¢ !r       Æ †	 /    â !r       Æ† 1   â !r        Æ‰$)      " !r!!        Æ ‰+      " !r##       Æ� ‰-      â !r%%"       Æ/ †	     â !r''       ÆO †     è †	 
    è † 
   â !r))$       Æ †	     â !r++       Æ† !    â !r--       Æ� †	 #    è † %   è †	 '    è † )    è †	 +    â M™        € ê †y -   â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y"      "   " ¹z  l      â Ey  €,    € â ™x?    â y#      !   b ¹z  F   
   â x,      ÿ  â x&      ÿ  â $rGÿÿ   ÿ Ž â rK ÿ       â $rIÿÿ   ÿ Ž â x(      ÿ  â $rEÿÿ   ÿ Ž â x*      ÿ  â r "ÿ   pRð Ä r "ÿ   pRò â$x "    #Ž Ê/ xÿ      È r    ÿ � È x%ÿ     È | %   pbô Ú G)  @+    € ê y      %   " xðÿÿÿÿÀŽ â $rÿÿ   ÿ Ž â r- ÿ       â y      '   b $x8%   ÿ Ž ä $x    
Ž â y      &   ¢ $v3ÿ X  ÿ Ž â z68 q  ÿàÿ â $x   ÿ Ž â x ÿ     â $v1ÿ Y  ÿ Ž Æ r    ÿ� È xøÿÿÿÿÀŽ ä x7ÿ     ä x    ÿ   â $x:   
Ž â x77   ÿ   â $rÿÿ   ÿ Ž ä 'z y   Ž â/ x9ÿ   :  Ä x<ÿ   7  â $z/ }  ÿŽ â xÿ   :  â 'z s   Ž âO zÿ z    ä r:   ÿ� â $z
 w  ÿŽ â zÿ t    â $z |  %Ž âr   ÿò� â $z	  p  %Ž âx     ÿàÿ â $z v  Ž â x`   ÿàÿ â$z	 |  	Ž â xÿ     â $z  p  Ž â xÿ   `  â $z v  	Ž â x\   ÿàÿ â $x:   ÿ Ž â r    ÿ� Ä xN   ÿàÿ ä r    ÿ� ä r!`   ÿ� är0`   ÿ� ä xÿ   \  ä xZ øÿÿÿÿÀŽ Ä x	ÿ   N  ä xV üÿÿÿÿÀŽ â $xZ   Z
Ž â r@\   ÿ� ä rB	N   ÿ� ä xV   Vèÿ ä rA\   ÿ� â %xn   ÿŽ â rC	N   ÿ� Ä x!øÿÿÿÿÀŽ ä x_@øÿÿÿÿÀŽ â %x   Ž â xSBøÿÿÿÿÀŽ ä xa0üÿÿÿÿÀŽ â $x`   
Ž â x]AüÿÿÿÿÀŽ â $x^\   _
Ž â xOCüÿÿÿÿÀŽ â $vÿ q  ÿ Ž â rRNS  €ÿàÿ â %x   Ž â x``   aèÿ Ä x\\   ]èÿ â $x$   8Ž âxNN   Oèÿ â $x.   8Ž âxWÿ   V  â $x2   8Ž âx[ÿ   Z  â $x4   8Ž â xOÿ   N  â %xn   VŽ â xSÿ   R  Ä x]ÿ   \  â %xn   ZŽ âx_ÿ   ^  ä xaÿ   `  â %xn   RŽ âxÿ     ä r/
/   ÿàõ â %x
n   NŽ â rDV   ÿàû Ä rJR   ÿàý â%xn   \Ž ârFN   ÿà÷ ä x;ÿ      â %xn   ^Ž â rNO   ÿäÿ ä x=ÿ      â %xn   `Ž â x?ÿ   0  Ä x@ÿ   @  â %xn   Ž âxAÿ   A  ä xBÿ   B  â $r5ÿÿ   ÿ â xCÿ   C  â %xn   ÿŽ â rHZ   ÿàõ ärP^   ÿà÷ â$xR   S ârT`   ÿàý â$xV   WŽ â rX   ÿàû â %x   Ž â x>ÿ   !  Ä rL\   ÿàù â %x   Ž â x9:   9  ä x>>   ÿ   â $x::   ÿ Ž â r\]   ÿä â $x;;   ÿ Ž ä $x==   ÿ Ž ä $x??   ÿ Ž Ä $x@@   ÿ Ž ä $xAA   ÿ Ž ä $xBB   ÿ Ž ä $xCC   ÿ Ž ä $xZ   [ ä$x^   _Ž Ä $x`   a ä$x0   Ž ä rb3   ÿàõ Ê $xc   1 â r3   ÿàõ È �ytb    • ¨  �yMb   • ¢  r1   ÿä Ê �yS"   • è �yv    • â r 3   ÿàõ Ê $x!   1 â rd3   ÿàõ È �yO     • h	 �yz    • b	 $xe   1 Ê �y d"   • h �yQd    • b rp3D   ÿà÷ ärr3H   ÿàõ Æ $xq1   VŽ â rx3T   ÿà÷ â $xs1   Z â rj3X   ÿàõ ä rf3   ÿàù â $xy1   `Ž ârk10   ÿä â $tuÿ$   ÿ Ž â r3L   ÿà÷ â/�ypp`   ‘ b rb3P   ÿàõ Æ $x1   \Ž ârh3   ÿà÷ â$xc1   ^ â r 3   ÿàõ â$xg   1 â�y|`   ‘ " $xi   1Ž â r!	1   ÿä ä rl7/   8àµ â �yWb`   ‘ & r[<5   ÿd~ â �y~ `   ‘ " rd3   ÿàõ Æ �y€f`   ‘ " $x[[$   ÿŽ â rb3   ÿàû ä �yrr`   ‘ " %v l Z  u Ž Æ/ �yUj`   ‘ " $xe   1 â r:    ÿàõ â $x!!   [Ž â �yxx`   ‘ " $xc   1Ž Æ �yYh`   ‘ " rj3   ÿàù ä/ r9!   ÿä â �y]b`   ‘ " rf3J   ÿàõ ä rk1   ÿä â �y_   ™ " rh3F   ÿà÷ â $xg1   R ârl
3   ÿàõ â �y‚j`   ‘ $ $xi1   NŽ Ä �y[(   ™ " rm1   ÿä ä rb7/   6àµ â/ �y„d`   ‘ ( �yq     • " rj3   ÿàù â %vbb Z  u Ž Ä �ya $   • " rd<5   ÿd~ Æ/ �ysf`   ‘ " $xk   1 ä $xed$   ÿŽ â �yll`   ‘ " rn:b   ÿàõ ä $xcc   eŽ â �y{L   ™ ( �y†h`   ‘ " $xo9   c Æ �ywj`   ‘ ( �yy H   • ( �y‡n   ™ ( �yˆn(   ™ ( �y�p   ™ ( �y…b    • ( �ymnL   ™ " rh7/   $àµ Æ �y}b$   • ( �y l   • " re<5   ÿd~ ä rd3   ÿàõ â �yƒbH   • $ $xie$   ÿŽ Ä �y‰np   ™ " %v h Z  u Ž â rf3   ÿà÷ ä/ �yŠbl   • " $xe   1 â rh:    ÿàõ â $x!!   iŽ â rg1   ÿäÿ ä �y‹dl   • " $xi9   ! Æ �y�fl   • " rj7/   .àµ Æ �yoh   ™ " r<5   ÿd~ Æ �ynh(   ™ " rb7/   2àµ â %vjj Z  u Ž ä �y�     • $ $xc$   ÿŽ â r<5   ÿd~ â �yŽ $   • " r:j   ÿàõ â $xkk   cŽ Ä �yŒhL   ™ " $xg$   ÿŽ ä/ r9k   ÿä â �y‘hp   ™ ( �y� H   • ( �y˜   ™ " rd7/   4àÙ Æ �y’ l   • " %vbb Z  u Ž Æ �y—j    • " %vdd Z  u Ž â rf<5   ÿ¤~ ä �yu(   ™ ( �yœL   ™ " $x•f$   ÿŽ â rf:b   ÿàõ â $xcc   gŽ â �y�p   ™ ( �yhj$   • " $xg9   c Æ �yijH   • " $xee   •Ž Æ �y“jl   • ( �y™b    • ( �y•b$   • ( �y”bH   • ( �y–bl   • " r :d   ÿà÷ Æ/ �y›d    • ( �yjd$   • ( �ybf(   ™ ( �ycfp   ™ " xMtT  M    ÆO �ykdH   • ¨ �ytf   ™ ¢  ržÿM   ÿ3Ž Æ �yMfL   ™ ¢  ržÿ7   ž  ä xvvT  S    â� $tSÿÿ Ž â r!9e   ÿäÿ â $xž   ÿ Ž â �yšdl   • ä x¢v€€€€SìŽ ä xSÿÀŽ Ä �y    ™ ¤
 xfSvøŽ â$x¢   S
Ž â �yd (   ™ ¨* �yS L   ™ ¨
 �ye p   ™ ¢
 x¡ÿ   v  È x¡¡ÿÀŽ ä xŸžÿÀŽ äx¢¡€€€€ÿüŽ ä x ÿ   v  ä rŸŸ¢   ÿáÿ â $x¢ž   ÿ Ž â x  ÿÀŽ Ä xgvxŽ ä x£ €€€€ÿüŽ ä x¢¢ÿÀŽ ä x¤¡žøŽ ä rfgf   ÿÀŽ â $x!£   ¢
Ž âxgÿ   v  Ä r¤¤Ÿ   ¡`Ž ä xggÿÀŽ ä xžÿ   ž  ä x ¡˜º  ÿ    ä x£¤˜º  ÿ    ä r¢ ¢   !XŽ Ä xOOT  z    ä x¡g€€€€ÿüŽ ä xžžÿÀŽ ä xŸ£€€€€ŸŽ äx £ `Ž ä x¢¢˜º  ÿ    Ä x£ ˜º  ÿ    ä rzÿO   ÿ3Ž â $x¡¡   ž
Ž â xvvÿÀŽ ä x!¢€€€€!Ž äx ¢£`Ž ä rzÿ7   z  Ä x¢f˜º  ÿ    ä r  Ÿ   ÿüŽ ä xOv˜º  ÿ    ä ržgž   ¡XŽ äxŸg˜º  ÿ    ä xvQT       â $x z   ÿ Ž â xg¢€€€€Ž Ä x     ä xžž˜º  ÿ    ä x¢¢O`Ž ä rO !   ÿüŽ ä x v€€€€ìŽ ä x! ÿÀŽ Ä xfž€€€€¡Ž äxŸžŸ`Ž ä xžÿ   v  â$x    !
Ž â x¡ÿ   v  äxžžÿÀŽ ä x !vøŽ Ä r!¢g   ÿüŽ ä x¡¡ÿÀŽ ä x¢z   ÿ   ä rfŸf   ÿüŽ ä xQvxŽ ä x ž€€€€ÿüŽ Ä xŸzÿÀŽ ä x£¡€€€€ÿüŽ ä x¢¢ÿÀŽ ä r Q    ÿÀŽ â $xg    Ÿ
Ž â xQÿ   v  ä xŸžzøŽ Ä x ÿ   z  â $xz£   ¢
Ž â xQQÿÀŽ ä rŸŸg   ž`Ž ä x£Q€€€€ÿüŽ ä x  ÿÀŽ ä r¢¡¢   zXŽ Ä xžž˜º  ÿ    ä xŸŸ˜º  ÿ    â $x££    
Ž â x¢¢˜º  ÿ    ä x¡¡˜º  ÿ    ä xgŸ€€€€gŽ äxžŸž`Ž Ä xvvÿÀŽ ä xŸ¢€€€€zŽ äx¢¢¡`Ž ä r Q    £XŽ äx¡Q˜º  ÿ    ä x  ˜º  ÿ    Ä xQv˜º  ÿ    ä x €€€€Ž äx  Q`Ž ä rpÿ=   p  ä r|ÿA   |  är€ÿ=   €  Ä rzžg   ÿüŽ ä rrÿ;   r  ä rWÿ@   W  ä rg    ÿüŽ â $xp   ÿ Ž â r~ÿ;   ~  â $xp|   ÿ Ž â rQ¢Ÿ   ÿüŽ â $xŸ€   ÿ Ž â xrr   ÿÀŽ Ä xWW   ÿÀŽ ä rUÿ>   U  ärxÿ?   x  ä x~~   ÿÀŽ ä rYÿ>   Y  ä x 0   râŽ Ä xpp0   WâŽ ä xUU   ÿÀŽ ä xxx   ÿ   ä xWŸ0   ~âŽ ä r~ÿ@   ]  ä x]Y   ÿÀŽ Ä xY àÿÿÿÿàÿ â &r !_   ÿ   âxUx0   UâŽ â &r|g_   ÿ   â xWWàÿÿÿÿàÿ â $rY    ÿŽ â rrÿ?   ‚  â &rxO[   ÿ   â xUUàÿÿÿÿàÿ â $r|W|   ÿŽ â r„ÿA   „  â $xrr   ÿ Ž â Er_        â $rxUx   ÿŽ â Er |       â 0r ÿq    A   â x~~   ÿÀŽ Ä xŸ„   ÿ   â 0raÿa    A   â Erq x       â #rx _   ÿ   â x]r0   ]âŽ â #rr    ÿ   â x~Ÿ0   ~âŽ ä rÿB   s  ä rsÿC   l  â &rlQ[   ÿ   â #r aq   x   â xppàÿÿÿÿàÿ â &r_ {   ÿ   âx]]àÿÿÿÿàÿ â &rxz{   ÿ   â x[~àÿÿÿÿàÿ Ä r†ÿC   †  â $rl]l   ÿŽ â rwÿB   w  â $r_p_   ÿŽ ä $rx[x   ÿŽ â x   ÿÀŽ â $x††   ÿ Ž â xqw   ÿÀŽ â $x|s   ÿ Ž â Erl l       â 0ryÿy    A   â Er_ _       Ä Erw x       â &rx!‡   ÿ   â x†0   âŽ â #rlal   r   â xq|0   qâŽ â #rsy_       â &r|Oˆ   ÿ   â x_àÿÿÿÿàÿ â $rxYx   ÿŽ ä &r~g‡   ÿ   Ä &rrf�   ÿ   â x  ˜º  ÿ    â #rayw   l   â $r|U|   ÿŽ â Erw x       â 0r ÿ…    A   ä $rW~   ÿŽ ä $rr_r   ÿŽ â xv €€€€£Ž â &ry m   ÿ   â x¡ ¡`Ž â 0r}ÿ}    A   â Er| |       â #rx w   ÿ   â Er        â 0rlÿ    A   â Err r       â $rypy   ÿŽ â rv¡v   ÿüŽ â #rx}|   x   â &rˆQˆ   ÿ   â #r~    ÿ   â &r|zm   ÿ   â #r lr   s   â xqqàÿÿÿÿàÿ â 0rƒÿƒ    A   â Ery y       â &rrv�   ÿ   Ä $rˆ]ˆ   ÿŽ ä $r[|   ÿŽ ä &r|f‰   ÿ   â #ryƒy   x   â $rrqr   ÿŽ â Er{ ˆ       â &rxv‰   ÿ   â Erm        â $r|_|   ÿŽ â Er r       â $rxqx   ÿŽ â #r~}{   ~   â 0rrÿ‹    A   â Er| |       â 0rŠÿŠ    A   â #rl   a   â Erx x       â 0rlÿ�    A   â #rmƒm   ~   â #r+r    +   â #ryŠ|   y   â &r !o   ÿ   Ä &r|go   ÿ   â #rmŠx   m   â #r'l   '   â&rxOn   ÿ   ä $rY    ÿŽ ä $r|W|   ÿŽ ä &rnQn   ÿ   â Era        â 0r ÿ�    A   â Er |       â $rn]n   ÿŽ â #r*lm   *   â $rxUx   ÿŽ Ä 0rŽÿŽ    A   â Ern n       â &rm Œ   ÿ   â#r    ÿ   â &rŒzŒ   ÿ   â Erx x       â #ra a   ÿ   â #rŽn      â $rmpm   ÿŽ ä $rŒ[Œ   ÿŽ Ä &r f‘   ÿ   ä&rnv‘   ÿ   â #raŽx   a   â $r _    ÿŽ â Erm m       â $rnqn   ÿŽ â ErŒ Œ       â 0r�ÿ�    A   ä &rx!˜   ÿ   Ä &r˜g˜   ÿ   â Er          â 0r’ÿ’    A   â Ern n       â #ra�m   a   â#r�Œ      â $r˜W˜   ÿŽ ä #ra’    a   â$rxYx   ÿŽ â #rn’n      â 0r ÿ—    A   â Er ˜       â &r|Ou   ÿ   âErm x       â &rxQu   ÿ   Ä #ro    ÿ   â $r|U|   ÿŽ ä &r œ   ÿ   â #rm m   ÿ   â $rx]x   ÿŽ ä &rœzœ   ÿ   â Er| |       â $rp   ÿŽ ä &r f�   ÿ   Ä 0rhÿh    A   â Erx x       â $rœ[œ   ÿŽ ä &r~v�   ÿ   â Er        â $r _    ÿŽ â #r|h|   m   â0riÿi    A   â Erœ œ       â $r~q~   ÿŽ â #rxhx   o   â 0r“ÿ“    A   â Er          â #r|i   |   â &rh!t   ÿ   âO Er ~       â #rxiœ   x   â #r)ln   )   â &rtgt   ÿ   â #r “    |   â &rnOb   ÿ   Ä $rhYh   ÿŽ â #r“   x   â $rtWt   ÿŽ â #r(ra   (   â#rEr    E   â $rnUn   ÿŽ â Era h       â 0r ÿ™    A   â #rKl   K   â Er t       â 0r•ÿ•    A   â Ern n       â &rtQb   ÿ   â #rb a   ÿ   â#rh    ÿ   â &ra M   ÿ   Ä #r•n   b   â $rt]t   ÿŽ ä &rbzM   ÿ   ä $rapa   ÿŽ ä &r fc   ÿ   â Ert t       â $rb[b   ÿŽ â Era a       â $r _    ÿŽ Ä 0r”ÿ”    A   â Erb b       â #r••t   h   â 0r–ÿ–    A   â Er          â #r”a      â&rh!   ÿ   â #r•”b   •   â &rbg   ÿ   â #r–       â $rhYh   ÿŽ Ä &r Od   ÿ   ä $rbWb   ÿŽ ä &rdQd   ÿ   ä &r S   ÿ   â Er! h       â $rUU    ÿŽ ä 0r ÿ›    A   Ä &rzzS   ÿ   ä $rd]d   ÿŽ ä $rpp   ÿŽ â Er b       â &r vc   ÿ   â ErU U       â 0rjÿj    A   â x%%   ÿàÿ â &rffe   ÿ   â #r! !   ÿ   â $rz[z   ÿŽ â Erd d       â &rvve   ÿ   â #r    ÿ   â | %   pbô â $rf_f   ÿŽ â #r jU   !   â$r q    ÿŽ âErp p       â $rvqv   ÿŽ â Er! z       â 0rkÿk    A   â #rdjd      â 0ršÿš    A   â� Er          Ä Erf f       â #rkp       âErv v       â #r!k!   d   â x33¸  ÿà÷ â #r –    •   â #ršf      â#rvšv   !   â #r,ry   ,   â#rIr   I   â #rGl    G   â #r&r   &   â #r-lv   -   â x$$    ÿàÿ â $r1ÿÿ   1Ž â x..    ÿàÿ Ä x22    ÿàÿ ä x44    ÿàÿ ä x66    ÿàÿ ä x88    ÿàÿ â D)       À â Gy  Ýÿÿÿÿƒ ê Ay        € ê Ey  ð     € â G‰  Ð     € ê $x""€  #Ž Ê ˆs "+ úÿ H   è ˆs "'€úÿ H   è ˆs ", ûÿ H   è ˆs "*€ûÿ H   è ˆs "( üÿ H   è ˆs ")€üÿ H   è ˆs "E ýÿ H   è ˆs "K€ýÿ H   è ˆs "I þÿ H   è ˆs "G€þÿ H   è ˆs "& ÿÿ H   è ˆs "-€ÿÿ H   ä Ay        € ê {          ì M        € ê „y #     H   b $vÿ r  ÿ Ž Æ „y# €   H   ¨ „y#    H   è „y	# €  H   ( „y#    H   h „y# €  H   ( „y#    H   ( „y# €  H   ( „y#    H   ( „y
#    H   è „y# €  H   â !r  +       â/ !r'       ÄO „y# €  H   â $x'   ÿŽ â !r,       ä� ‰       " !r		*       Æ‰      b !r(       Æy      %   ¢ !r))       Æ y      '   â !rEE       Æ ‰      " !rKK       Æ ‰	      b !r&       â !r        ä ‰      " !rI
I       â !r       ä/ ‰       b !rGG       â x   ÿ   ÄO ‰
)      ¢ !r--       Æ ‰E      â z  r  p`ð â !r       ä‰K      " r #ÿ   p"p â !r		       ä‰      h ‰I      b !r       â x   ÿàÿ Ä ‰G      " !r        â/ z  r  p`ò ä ‰      b !r))
       âO x #   p"ò ä y       &   ¢ !rEE       â� !rKK       Ä‰      â !r       Æ‰	      " !rII       Æ ‰E      b !rGG       Æ ‰      " !r       Æ/ ‰)      b $z   x  Ž ÆO ‰K      ¢ $z  ~   Ž Æ ‰      ¢ !r       Æ� ‰I      â !r	       â
  r  ÿàõ ä ‰G      " !rEE       ä$ÿÿ   ÿ â ‰-      b !r       â 
 j  ÿ„ Ä ‰ €    " !r))       â/ 
 k   ä ‰ €    b !rKK       âO !r       ä ‰ €    ¢ !rII       Æ� ‰ €    â !rGG       Æ‰ €    " !r--       Æ‰) €    b !r
       Æ ‰E €    "  '   ÿàù â !r       ä/ ‰K €    b r    ÿ† ä ÿÿ   ÿä â ‰" €    b !r       âO $rÿÿ   ÿŽ â z j  ÿ† â ‰I €    ¢ 
 j  ÿˆ Ä z k  � â ‰ G €    ¢ !r       â� !r       â
 k   â ‰$-      â !r)       â$x)   ÿŽ âr    ÿ„ â ‰
 @    " !rE       Ä  )   ÿàù â ‰	 @    " !rK       â/ $r(ÿÿ   ÿ â !r""       â ‰ @    b $&ÿÿ   ÿ â 
 j  ÿˆ â !rI       âO ‰ @    ¢ !r G        Æ ‰ @    b !r$-$       Æ� ‰ @    â !r
       Æ‰ @    "    ÿàÿ â !r		       ä ‰ @    " *    ÿàõ ä z
 j  ÿ† â‰ @    " !r       â/ $+ÿÿ   ÿ â * j  ÿ„ â ‰!  @    b !r       âO z k  (� Ä ‰#" @    ¢ !r       â
 k  & ä ‰%$ €    b !r       â� * k  + ä ‰	      â !r       Æ‰      " !r       Æ ‰      " !r       â !r! !       ä/ ‰      b !r#"#       ÆO ‰      ¢ !r%$%       Æ‰      b !r+	       Æ� ‰      â 	'   ÿàÿ â !r-       ä‰      " x'        ä  	   ÿàõ â‰      " !r       â %v  j  ' Ž Ä ‰ !      " !r       â/ $&ÿÿ   ÿ â  j  ÿ„ â ‰$% @    b !r       äO  k  & â ‰"#      ¢ !r       â!r       Ä� †	 +    â !r       Æ† -   â !r       Æ †	     â !r!!        Æ †     â !r$%$       Æ/ †	     â !r##"       ÆO †    è †	     è †     è †	 
    è † 
!   è ‰$      ( †	 #    â M™        € ê x)   ÿàÿ â� !r$       Æ r    ÿàñ Ê $r ÿÿ   ÿ  â z j  ÿ€ È z k     Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y,      "   " ¹z  l      â Ey   )    € â ™x?    â y-      !   b ¹z  F   
   â $r.ÿÿ   ÿ Ž â x2      ÿ  â $rKÿÿ   ÿ Ž â r1 ÿ       â $rMÿÿ   ÿ Ž Ä $rQÿÿ   ÿ Ž ä $rOÿÿ   ÿ Ž ä $rSÿÿ   ÿ Ž ä $rUÿÿ   ÿ Ž â r ,ÿ   pRð är ,ÿ   pRò â$x ,    -Ž Ê/ xÿ      È r    ÿ � È x0ÿ     È | 0   pbô Ú G)  Ð'    € ê xðÿÿÿÿÀŽ â y      '   " $rÿÿ   ÿ Ž ä $x?0   ÿ Ž â y      &   b $x    
Ž ä $v/ÿ X  ÿ Ž â y      %   ¢ $v7ÿ Y  ÿ Ž â x ÿ     â $r.ÿÿ   ÿ Ž â z=? q  ÿàÿ Ä r    ÿ� È xøÿÿÿÿÀŽ ä x:ÿ     Æ $xC   
Ž ä 'z  y   Ž Æ xÿ   C  â $z4 }  ÿŽ ä $z	 w  ÿŽ ä/ $x   ÿ Ž äO $rÿÿ   ÿ Ž â r4	4   ÿàõ â $x::   ÿ Ž â r	C   ÿ� â 'z s   Ž â zÿ z     Ä r		   ÿò� â $z  p  0Ž â zÿ t    â $z |  0Ž â x   ÿàÿ â $z |   Ž â x ÿ   	  â $z v  Ž âxd	   ÿàÿ â $z v  Ž â r" 	   ÿ� â$z p  Ž â r  	   ÿ� â %xn   ÿŽ â x^"øÿÿÿÿÀŽ Ä xZ üÿÿÿÿÀŽ â $r8ÿÿ   ÿ â x`	   ÿàÿ â$x^	   ^
Ž âxZ	   Zèÿ äx		   ÿàÿ ä xÿ   `  ä xÿ   	  Ä xÿ   d  ä r%`   ÿ� ä r'	   ÿ� ä r#d   ÿ� är$d   ÿ� ä r&`   ÿ� Ä r(	   ÿ� â $vÿ q  ÿ Ž â xc%øÿÿÿÿÀŽ ä xV'øÿÿÿÿÀŽ â $x5   ?Ž â x#øÿÿÿÿÀŽ â $xb`   c
Ž â xe$üÿÿÿÿÀŽ â $xV	   V
Ž â xa&üÿÿÿÿÀŽ â $x9   ?Ž âx (üÿÿÿÿÀŽ â $x;   ?Ž â rd  €ÿàÿ Ä xdd   eèÿ ä x``   aèÿ ä x 	    èÿ ä x[ÿ   Z  ä x_ÿ   ^  ä x!ÿ      â %xn   ZŽ â xWÿ   V  Ä xaÿ   `  â %xn   ^Ž âxcÿ   b  ä xeÿ   d  â %xn    Ž âxÿ     ä rJ    ÿà÷ â%xn   VŽ â rNV   ÿàý Ä rHZ   ÿàû â %xn   `Ž â x   ÿ   ä rZ[   ÿäÿ â %xn   bŽ âxÿ   C  ä x<ÿ   "  â %xn   dŽ â x@ÿ   #  Ä xBÿ   $  â %xn   Ž âxDÿ   %  ä xFÿ   &  â %xn   ÿŽ â xGÿ   '  ä xIÿ   (  â $xR   !Ž ârL^   ÿàõ â$xV   W â rP`   ÿàù â$xC   ÿ Ž â rTb   ÿà÷ â%x   Ž â rXd   ÿàý Ä r\   ÿàû â %x
   Ž â x>ÿ      ä xAC     â$xCC   ÿ Ž â xEÿ   :  â %x   Ž â x>>   ÿ   Æ %x   Ž È $x<<   ÿ Ž ä $x@@   ÿ Ž ä $xBB   ÿ Ž ä $xDD   ÿ Ž ä $xFF   ÿ Ž ä $xGG   ÿ Ž Ä $xII   ÿ Ž ä $x^   _ ä$x`   a ä$xb   cŽ ä$xd   e ä$x6   Ž Ä r"
/   ÿà÷ är/   ÿàõ Ä r#7   ÿäÿ Æ $x	   7 ä �yh"    • ¨  �y["   • ¨  �y_"   • è �y     • â r$/   ÿàõ Ä r /   ÿà÷ Æ $x%   7 ä$x!   7Ž Æ �yj$"   • b	 r/L   ÿàõ Æ/ �yY$    • h	 �yW     • h �yn    • b r&/H   ÿà÷ â $x7   ^ â r(/\   ÿàõ Æ $x'7   ZŽ â r*/X   ÿà÷ â $x)7   6 â rp/T   ÿàõ â �yi`   ‘ d $x+7   dŽ â rr/P   ÿà÷ â $xq7   b â rl/N   ÿàõ â �y](`   ‘ b rs7`   ÿäÿ Ä r$/J   ÿà÷ â$xm7   V âr"/   ÿàõ â�ye&`   ‘ $ $x%7   RŽ â r/   ÿà÷ â$x#   7 âr /   ÿàõ â/�ycp`   ‘ $ $x   7Ž âr(/   ÿà÷ â �yf$`   ‘ " $x!   7 Æ $x)   7Ž â r&/   ÿàù â�yo `   ‘ " x‚ $       ä r%:4   ?àµ â/ �yg*`   ‘ & rpE8   ÿd~ â �yll`   ‘ " r$/   ÿàõ â $x'   7 Ä %v % Z  ‚ Ž â �yk"`   ‘ " r*/   ÿàù Æ/ �yar`   ‘ " $x%   7 Æ �ym`   ‘ " $x+   7 Æ �yq&`   ‘ " r"/   ÿàû Æ �ys(`   ‘ "! $xp$   ÿŽ â rC    ÿàõ ä �yt$`   ‘ " $x!!   Ž â r':4   =àÇ ä �yu*`   ‘ " $xA   ! â r(E8   ÿ„þ â $x#   7Ž â �y     • " %v&' Z  ‚ Ž Æ �yr   ™ " $x)($   ÿŽ â r(C&   ÿàõ ä �yp"`   ‘ " $x''   )Ž Æ �y|(   ™ $ r)A'   ÿä ä �y€&    • ( �y{(   ™ ( �y� $   • ( �yw((   ™ ( �yyL   ™ ( �yz(L   ™ ( �y}&$   • " r#:4   5àµ Æ �yv H   • ( �yx&H   • " r%E8   ÿd~ Æ/ �yƒp   ™ " r$/   ÿàõ â %v"# Z  ‚ Ž ä �y„(p   ™ $ $x+%$   ÿŽ ä $x%   7 â �y~ l   • " r*C"   ÿàõ â $x##   +Ž Ä �y…&l   • $ $x+A   # ä �yˆ$l   • ( �y‡*   ™ ( �y(*(   ™ " r:4   9àÇ Æ/ �y‹"    • " rE8   ÿ„þ Æ �y‰"$   • " r /   ÿàõ Æ �y)*L   ™ " %v$ Z  ‚ Ž Æ �yŠ*p   ™ " $x$   ÿŽ ä $x!   7 â �y†"H   • " rC$   ÿàõ â $x%%   Ž ä �y�"l   • $ $xA   % Ä �yŒ l   • ( �yŽ   ™ " r':4   ;àµ Æ �y�$    • " r+E8   ÿd~ Æ �y*(   ™ " %v&' Z  ‚ Ž Æ �y‚L   ™ " $x‘+$   ÿŽ â r C&   ÿàõ ä/ �y�p   ™ $ r''‘   ÿàÿ ä �y+$$   • ( �y"$H   • " $x!A   ' Æ �y#$l   • ( �y’ p   ™ ( �y&l   • ( �y$&H   • "! xhhT  [    ÆO �y[&$   • ¢  r%ÿh   ÿ3Ž Æ �yh L   ™ ¢ x  T  _    ä� rÿ:   %  â �y_&    • â  x‘ÿ      Æ �y%    ™ ¢ x–‘ÿÀŽ Æ �y‘ (   ™ ¢ x”–€€€€ÿüŽ äx“ÿÀŽ ä x•–øŽ Æ $x“”   “
Ž â x&ÿ      â $x    ÿ Ž Æ/ r˜•“   –`Ž â $x•   ÿ Ž â x&&ÿÀŽ ä xÿ     â $t—ÿÿ Ž â x'ÿ      ä x™–˜º  ÿ    ä x–•ÿÀŽ Ä xš&€€€€ÿüŽ ä x•ÿÀŽ ä x''ÿÀŽ ä x” €€€€—ìŽ ä x! ÿÀŽ â $xšš   •
Ž â x—'€€€€ÿüŽ Ä x˜˜˜º  ÿ    â $x”   !
Ž â r—–—   ÿáÿ ä r•&•   šXŽ ä x ! øŽ äx! xŽ ä x“˜€€€€“Ž Ä x˜˜™`Ž ä r–'–   —XŽ ä x••˜º  ÿ    ä x&&˜º  ÿ    ä xjYT  j    äxWWT  n    Ä r !    ÿÀŽ ä r!˜“   ÿüŽ ä x––˜º  ÿ    ä xš•€€€€šŽ ä x“'˜º  ÿ    ä x••&`Ž Ä x&ÿ   j  ä rYÿW   ÿ3Ž ä x  ÿÀŽ ä x'–€€€€—Ž äx––“`Ž ä x  ˜º  ÿ    Ä x“&ÿÀŽ ä rYÿ:   Y  ä xW ˜º  ÿ    ä x €€€€Ž ä x—“€€€€ÿüŽ ä xnYÿÀŽ Ä x  W`Ž ä rW•š   ÿüŽ â $t•ÿÿ Ž â r     ÿüŽ â $xn—   n
Ž â x“YøŽ ä x”j€€€€•ìŽ ä x&ÿ   j  Ä r•n   “`Ž â $xY   ÿ Ž â x ÿ   j  ä x˜“˜º  ÿ    â $x“Y   ÿ Ž â x&&ÿÀŽ ä x—ÿ   Y  ä x  ÿÀŽ Ä xYÿÀŽ ä r'–'   ÿüŽ ä x–&€€€€ÿüŽ ä x““ÿÀŽ â $x”   Y
Ž â xš €€€€ÿüŽ ä x——ÿÀŽ Ä r–“–   ÿáÿ ä x••˜º  ÿ    â $xšš   —
Ž â xYYjøŽ äx”jxŽ ä r“&“   –XŽ ä xn•€€€€nŽ Ä x••˜`Ž ä r””Y   ÿÀŽ ä r— —   šXŽ ä x““˜º  ÿ    ä x&&˜º  ÿ    ä xjjÿÀŽ Ä rY•n   ÿüŽ ä x——˜º  ÿ    ä x–“€€€€–Ž äx•“&`Ž ä x””˜º  ÿ    ä reÿ>   e  ÄrgÿB   g  ä x  ˜º  ÿ    ä x“j˜º  ÿ    ä rmÿ>   m  ä riÿ<   i  ä r]ÿ@   ]  Ä rkÿ<   k  â $xjg   ÿ Ž â xš—€€€€šŽ ä x&”€€€€Ž â$xe   ÿ Ž â x—— `Ž ä x“”“`Ž â $x”m   ÿ Ž â x i   ÿÀŽ Ä x]]   ÿÀŽ ä raÿF   a  ä xkk   ÿÀŽ ä rfÿI   f  ä rcÿD   c  ä x 0    âŽ Ä rlÿG   l  ä xgj0   ]âŽ â $xna   ÿ Ž â r&“&   ÿüŽ ä x]”0   kâŽ â $xf   ÿ Ž â xcc   ÿÀŽ ä xll   ÿÀŽ Ä xa àÿÿÿÿàÿ â &r  r   ÿ   âx]]àÿÿÿÿàÿ â &rr&r   ÿ   â xfn0   câŽ ä xc0   lâŽ â $ra    ÿŽ â roÿ@   o  ä rqÿB   q  â $rr]r   ÿŽ â rtÿI   t  Ä rpÿG   p  ä xoo   ÿÀŽ ä xjq   ÿ   â $xtt   ÿ Ž â Erm        â 0r ÿ    A   â xkp   ÿÀŽ ä Er r       Ä ri•–   ÿüŽ ä xjj0   oâŽ â #rn m   ÿ   â xkt0   kâŽ â #rr    ÿ   â xggàÿÿÿÿàÿ â &rt {   ÿ   â xjjàÿÿÿÿàÿ â 0r ÿ€    A   ä &rp'|   ÿ   Ä &ri|   ÿ   â ruÿF   u  â &r€&{   ÿ   â rsÿD   s  â $rtat   ÿŽ ä $rpgp   ÿŽ ä $rj   ÿŽ ä $r€]€   ÿŽ â xls   ÿÀŽ â $xuu   ÿ Ž â Erm t       â 0r�ÿ�    A   â Erp p       Ä Ero        ä Er €       â #rt m   ÿ   âxlu0   lâŽ â #rm�p   n   â &r|'w   ÿ   â #rn    ÿ   â xllàÿÿÿÿàÿ â &r!y   ÿ   ä &rwiw   ÿ   Ä &ryYy   ÿ   â xffàÿÿÿÿàÿ â $r|g|   ÿŽ ä &rs!z   ÿ   ä&ruYz   ÿ   ä $rwjw   ÿŽ ä $ryly   ÿŽ â Er| |       â $rsfs   ÿŽ â Erw w       â 0r}ÿ}    A   â Er  y       â $rulu   ÿŽ â #ro�o   r   â 0rvÿv    A   â Ers s       â 0rxÿx    A   â #rq}|   t   âErr u       â #rw}w   n   â #rov    o   â xccàÿÿÿÿàÿ â $rf   ÿŽ Ä &r Wƒ   ÿ   â#rpxs   q   â#rxxr   w   â &rrW„   ÿ   â Er        â $r c    ÿŽ â re—š   ÿüŽ â $rrcr   ÿŽ ä 0r~ÿ~    A   â Er          â #rv   m   â xkkàÿÿÿÿàÿ â &rneƒ   ÿ   â Err r       â 0r…ÿ…    A   â #r ~       â 0rˆÿˆ    A   Ä $rnkn   ÿŽ â #rp…r   p   ä #rUˆ    U   â &r  ‡   ÿ   âErn n       â &rr&‡   ÿ   â #rSˆp   S   â &rp'(   ÿ   ä $ra    ÿŽ â #rn~n   o   â $rr]r   ÿŽ Ä $rpgp   ÿŽ â Ero        â 0r ÿ‹    A   â Er r       â 0r‰ÿ‰    A   â Erp p       â &rsi(   ÿ   â #r( o   ÿ   â&rq!)   ÿ   â #r     ÿ   â $rsjs   ÿŽ â #r‰p   (   â $rqfq   ÿŽ Ä &r(WŠ   ÿ   â Ers s       â &r)Y)   ÿ   â Erq q       â $r(c(   ÿŽ ä 0r†ÿ†    A   â #rs‰s       â $r)l)   ÿŽ â Er( (       â &rŠeŠ   ÿ   â#r †q      â 0r�ÿ�    A   â Er) )       â &r„e„   ÿ   Ä $rŠkŠ   ÿŽ â #r �(       â &r( Ž   ÿ   ä $r„k„   ÿŽ â Er Š       â #r††)   s   â &rŽ&Ž   ÿ   ä $r(a(   ÿŽ ä 0rŒÿŒ    A   â Erm „       â #r�   †   â #rOˆ    O   â $rŽ]Ž   ÿŽ â Er) (       â 0r ÿ�    A   â #r1Œn   1   â&rn'*   ÿ   â #rm…m   x   â #rQŒ   Q   â Er Ž       â #r( )   ÿ   â $rngn   ÿŽ â #r2Œm   2   â &r)!‚   ÿ   Ä &rmi*   ÿ   â #r*    ÿ   â $r)f)   ÿŽ â Ern n       â &r W�   ÿ   ä 0r+ÿ+    A   â Er) )       â $r c    ÿŽ ä 0r"ÿ"    A   â #r+n   (   â $rmjm   ÿŽ â Er(         â &roY‚   ÿ   â #r ")      â 0r#ÿ#    A   â Erm m       â &r  %   ÿ   ÄO $rolo   ÿŽ ä &r&&%   ÿ   â #r(#(       â $r a    ÿŽ â #rm+m   *   â &r '‘   ÿ   ä &r!!h   ÿ   â Er* o       â $r]&   ÿŽ Ä &rii‘   ÿ   ä $rgg    ÿŽ ä $rff!   ÿŽ â Er!         â &rYYh   ÿ   â Er        â $riji   ÿŽ ä 0r ÿ_    A   â� #r*"*   m   â &r"W’   ÿ   â x00   ÿàÿ â $rYlY   ÿŽ â Erg g       â &r�e�   ÿ   â#r  !   ÿ   â&r’e’   ÿ   â Eri i       â 0r[ÿ[    A   â #r     ÿ   â | 0   pbô â $r"c"   ÿŽ â Erf f       â $r�k�   ÿŽ âErY Y       â $r’k’   ÿŽ â #rg[g       â0r$ÿ$    A   â #ri[i       â Er� �       â 0rÿ    A   â Er" "       â #rg$f   g   âEr’ ’       â #rY$Y   i   â x//¸  ÿà÷ â #r*#�   *   â #r""   g   â#rY’   Y   â #rMˆ(   M   â #rKŒ*   K   â #r3ˆ"   3   â #r.ŒY   .   â x55    ÿàÿ â $r7ÿÿ   7Ž â x99    ÿàÿ Ä x;;    ÿàÿ ä x==    ÿàÿ ä x??    ÿàÿ â D)       À â Gy  pàÿÿÿÿƒ ê Ay        € ê Ey  Ð     € â G‰  °     € ê $x,,@  -Ž Ê ˆs ,U ûÿ H   è ˆs ,1€ûÿ H   è ˆs ,S üÿ H   è ˆs ,2€üÿ H   è ˆs ,O ýÿ H   è ˆs ,Q€ýÿ H   è ˆs ,M þÿ H   è ˆs ,K€þÿ H   è ˆs ,3 ÿÿ H   è ˆs ,.€ÿÿ H   ä Ay        € ê {          ì M        € ê „y -     H   b $v%ÿ r  ÿ Ž Ä $t#ÿ   ÿ Ž â „y- €   H   ¢ $x!%   ÿŽ Æ „y-    H   è „y- €  H   ( „y-    H   h „y- €  H   ( „y
-    H   ( „y- €  H   ( „y-    H   è „y- €  H   â !r  U       Æ/ y      %   b !r1       ÆO ‰       ¢ !rS       Æ� y      '   â !r22       Æ‰      " !rO       Æ‰      " !rQ       Æ ‰	2      b !r

M       Æ ‰      â $x   ÿ Ž â/ !rK       â !r        âO ‰      d z  r  p`ð â !r3       â ‰       ¤ r -ÿ   p"p â !r.       â !r       â ‰
      " !r       Æ‰      " !r	2	       Æ‰      b !r       Æ� ‰      â !r       Æ/ ‰      b !r         ÆO ‰      ¢ !r
       Æ ‰	      " !r       Æ‰  €    " !r       Æ‰      b !r       Æ� ‰
      è ‰      â !r       Æ/ ‰      b !r       ÆO ‰      ¢ !r		       Æ ‰      " !r        â!r       ä‰ €    " !r
       Æ� ‰ €    â x
   ÿàÿ â !r       ä ‰	 €    b z 
 r  p`ò â !r       ä/ y       &   b x -   p"ò â !r       ÄO ‰ @    ¢ !r       Æ ‰ €    ( ‰ €    " !r
       Æ‰ €    " !r       Æ� ‰ €    â !r	       Æ‰ €    b $z   x  Ž Æ/ ‰ €    b !r       âO $z  ~   Ž â !r       ä ‰
 @    $ r%    ÿ† â!r       â  !   ÿàõ â‰      ¢ 
	  r  ÿàù â !r       â$rÿÿ   ÿŽ â ‰ @    " $ ÿÿ   ÿ â !r       â� 
 j  ÿ„ â ‰ @    â z j  ÿ† â !r       â
 k    â ‰ @    b z k  � â !r       â/ r%    ÿŠ â ‰ @    b $$ÿÿ   ÿ â 
	 j  ÿˆ Ä ‰ @    b !r
       â r"ÿÿ   ÿäÿ ä ‰ @    " !r'       âO %   ÿàÿ ä ‰ @    ¢ !r       â!   ÿàÿ Ä ‰ @    " !r       â� !    ÿàõ â%v  j  # Ž â !r       â    ÿà÷ â ‰      â z j  ÿŠ â !r       â/ 
		 k  $ â ‰      b z k  "� â !r       â $"ÿÿ   ÿ â ‰      b 
! j  ÿ„ â !r       â $ ÿÿ   ÿŽ â ‰      "  j  ÿ† â !r       ÄO ‰      ¢  k   � â !r       ä‰      ( ‰       " !r#       â� ! k  " Ä ‰      â !r       Æ/ ‰      b !r       â!r       ä †	 '    â !r       ÆO † #   â !r       Æ†	     â !r        Æ † 
    â !r       Æ� †	     â !r       Æ/ †    è †	     è †     è †	     â M™        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y.      "   " ¹z  l      â Ey  À%    € â ™x?    â y/      !   b ¹z  F   
   â $r6ÿÿ   ÿ Ž ä $rMÿÿ   ÿ Ž ä $rOÿÿ   ÿ Ž Ä $r4ÿÿ   ÿ Ž ä $rSÿÿ   ÿ Ž ä $r3ÿÿ   ÿ Ž ä $rQÿÿ   ÿ Ž ä $rKÿÿ   ÿ Ž â r .ÿ   pRð är .ÿ   pRò â$x .    /Ž Ê/ xÿ      È r    ÿ � È x2ÿ     È | 2   pbô Ú G)  €$    € ê xðÿÿÿÿÀŽ â y      '   " $rÿÿ   ÿ Ž ä $x>2   ÿ Ž â y      %   b $x    
Ž ä $v1ÿ X  ÿ Ž â y      &   ¢ $v7ÿ Y  ÿ Ž â x ÿ     â $r6ÿÿ   ÿ Ž â z<> q  ÿàÿ Ä r    ÿ� È xøÿÿÿÿÀŽ ä x;ÿ     Æ $x?   
Ž ä 'z  y   Ž Æ xÿ   ?  â $z5 }  ÿŽ â x   ÿ   â/ $rÿÿ   ÿ Ž â r	?   ÿ� â $z
 w  ÿŽ äO'z s   Ž â zÿ z     ä r		   ÿò� â $z  p  2Ž â zÿ t    â $z |  2Ž â x   ÿàÿ â $z |   Ž â x ÿ   	  â $z v  Ž âxd	   ÿàÿ â $z v  Ž â r" 	   ÿ� â$z p  Ž â r  	   ÿ� â %xn   ÿŽ â x^"øÿÿÿÿÀŽ Ä xZ üÿÿÿÿÀŽ â $x;;   ÿ Ž â x`	   ÿàÿ â$x^	   ^
Ž âxZ	   Zèÿ äx		   ÿàÿ ä xÿ   d  ä xÿ   `  Ä xÿ   	  ä r#d   ÿ� ä r%`   ÿ� ä r'	   ÿ� är$d   ÿ� ä r&`   ÿ� â $x   ÿ Ž â r(	   ÿ� â $x?   ÿ Ž â x#øÿÿÿÿÀŽ Ä xc%øÿÿÿÿÀŽ ä xV'øÿÿÿÿÀŽ â $xd   
Ž â xe$üÿÿÿÿÀŽ â $xb`   c
Ž â xa&üÿÿÿÿÀŽ â $xV	   V
Ž â x (üÿÿÿÿÀŽ â $vÿ q  ÿ Ž â xdd   eèÿ Ä x``   aèÿ â $x0   >Ž â x 	    èÿ â $x:   >Ž â x[ÿ   Z  â %x   Ž â x_ÿ   ^  ä x!ÿ      â %xn   ZŽ â xWÿ   V  Ä xaÿ   `  â %xn   ^Ž âxcÿ   b  ä xeÿ   d  â %xn    Ž âxÿ     ä rJ    ÿà÷ â%xn   VŽ â rNV   ÿàý Ä r5
5   ÿàõ â %xn   `Ž ârHZ   ÿàû ä r9ÿÿ   ÿä â %xn   bŽ â rZ[   ÿäÿ ä xÿ   ?  â %xn   dŽ â xAÿ   "  Ä xBÿ      â %xn   Ž âxCÿ   #  ä xDÿ   $  â %xn   ÿŽ â xEÿ   %  ä xFÿ   &  â $xR   !Ž âxGÿ   '  â $xV   W â xIÿ   (  â %x
   Ž â rL^   ÿàõ Ä rP`   ÿàù â%x   Ž â rTb   ÿà÷ ärXd   ÿàý â$xAA   ÿ Ž â r\   ÿàû â $xBB   ÿ Ž â x=?     â$x??   ÿ Ž â x@ÿ   ;  â %x   Ž È $xCC   ÿ Ž ä $xDD   ÿ Ž ä $xEE   ÿ Ž ä $xFF   ÿ Ž ä $xGG   ÿ Ž ä $xII   ÿ Ž Ä $x^   _ ä$x`   a ä$xb   cŽ ä$xd   e ä$x8   Ž ä r"
1   ÿà÷ är1   ÿàõ Æ $x#   7Ž ä$x	   7 Æ �yh"    • ¨  �yW"   • ¨  �yU"   • è �yf    • â r 1   ÿàõ Ê $x!   7 â r$1   ÿàõ È �y      • h	 �y[    • b	 $x%   7 Ê �yY$"   • h �yj$    • b	 r"1H   ÿà÷ ärp1L   ÿàõ Æ $x#7   ZŽ â r1X   ÿà÷ â/ $xq7   ^ â rt1\   ÿàõ â $t…ÿ$   ÿ Ž â �y_"`   ‘ b $x7   dŽ âru78   ÿä ä rl1T   ÿàõ â�ypp`   ‘ b r 1P   ÿà÷ Æ$xm7   b ârn1   ÿàõ â$x!7   `Ž â rr1   ÿà÷ â�ya`   ‘ d	 $xo   7 ä$xs   7Ž â �yc `   ‘ b r&1   ÿàù Æ �ytt`   ‘ b r;5   >àµ är(1   ÿàû â �yll`   ‘ b	 r@9   ÿd~ â %v  Z  … Ž â r*1   ÿàõ â �ynn`   ‘ d $x$   ÿŽ Ä $x+   7 â r?    ÿàõ â $x!!   Ž â �yrr`   ‘ b $x'   7 âr$1   ÿàù ä r=!   ÿä â $x)   7Ž â�y]&`   ‘ b $x%   7 â r"1N   ÿàõ Ä �yk   ™ h �yg(   ™ b $x#7   V Æ �ye$`   ‘ h �yv(`   ‘ h	 �yo     • b r$1J   ÿà÷ Æ/ �yi*`   ‘ b r(1   ÿàõ â$x%7   RŽ ä �ym $   • $ $x)   7 âr,;5   <àµ â �yx"`   ‘ & r*@9   ÿd~ â/ %v,, Z  … Ž â r&1   ÿàù â �yz$`   ‘ $ $x+*$   ÿŽ â r*?,   ÿàõ â �yqL   ™ " $x'   7 Ä $x#-   +Ž â �y~(`   ‘ & $x+=   # â �y|&`   ‘ ( �ys H   • ( �y-*   ™ " $r"ÿÿ   , Ž Æ �y{p   ™ ( �y*(   ™ ( �y}"    • ( �yw l   • ( �yu"$   • " r(;5   0àµ Æ �y,*L   ™ " r$@9   ÿd~ Æ/ �y€*p   ™ " r 1   ÿàõ â %v(( Z  … Ž ä �yy"H   • $ $x%$$   ÿŽ ä $x!   7 â �y�"l   • " r&?(   ÿàõ â $x)   %Ž â r$1   ÿà÷ â �y‚ l   • $ $x'=    Ä $x%   7Ž Æ �yƒ&   ™ " r (       Æ �y$$l   • ( �y)&(   ™ " r";5   :àµ Æ/ �y„    • " r#@9   ÿd~ Æ �y($   • " %v " Z  … Ž Æ �y+&L   ™ " $x##$   ÿŽ â r"?    ÿàõ ä �y%&p   ™ " $x!!   #Ž Æ �y*H   • " $x#=   ! Æ �yˆl   • ( �y†"   ™ ( �y‡"(   ™ ( �yŠ     • ( �y…"L   ™ ( �y‰"p   ™ ( �y& $   • ( �y' H   • ( �y l   • "! xWhT  W    ÈO rhÿW   ÿ3Ž È rhÿ;   h  ä xUfT  U    â� $tfÿÿ Ž ä $xh   ÿ Ž â x"ÿ   U  ä x#U€€€€fìŽ ä xfÿÀŽ Ä xW"ÿÀŽ Æ $x"#   f
Ž â xfUøŽ äxŒW€€€€ÿüŽ ä x "UxŽ äx#hÿÀŽ ä x‹ÿ   U  äx!ÿ   U  Ä r    ÿÀŽ â $x h   ÿ Ž â xfWhøŽ â$xŒŒ   #
Ž â x‹‹ÿÀŽ ä xhÿ   h  ä x!!ÿÀŽ ä x�‹€€€€ÿüŽ Ä xhhÿÀŽ ä x#!€€€€ÿüŽ ä x  ÿÀŽ ä rffŒ   W`Ž â $x��   h
Ž â xŽW˜º  ÿ    â $x##    
Ž â xWf˜º  ÿ    Ä x  T  [    äxŒW€€€€ŒŽ ä rh‹h   �XŽ ä xWWŽ`Ž ä r !    #XŽ ä xUUÿÀŽ Ä rŽÿ    ÿ3Ž ä xhh˜º  ÿ    ä x‹‹˜º  ÿ    ä x˜º  ÿ    ä xf ˜º  ÿ    ä x U˜º  ÿ    Ä rŽÿ;   Ž  ä r WŒ   ÿüŽ ä x�h€€€€�Ž äxŒh‹`Ž ä xh€€€€"Ž ä xYjT  Y    â $tjÿÿ Ž â x!!˜º  ÿ    Ä x `Ž â $x Ž   ÿ Ž â x#f€€€€#Ž äxff!`Ž ä x"ÿ   Y  ä x!Y€€€€jìŽ ä xj ÿÀŽ Ä xW"ÿÀŽ ä r j!   ÿáÿ ä rff#   ÿüŽ ä r"Œ�   ÿüŽ ä x#ÿ   Y  ä xŒW€€€€ÿüŽ Ä xUŽÿÀŽ ärhh   ÿüŽ â $xŽ   ÿ Ž â xjjYøŽ äx! YxŽ äx‹ÿ   Y  â $xŒŒ   U
Ž â x##ÿÀŽ Ä r!!j   ÿÀŽ ä x[WŽøŽ äxjÿÀŽ ä x‹‹ÿÀŽ ä xÿ   Ž  ä xU#€€€€ÿüŽ Ä r[[Œ   W`Ž ä xŽW˜º  ÿ    ä xW‹€€€€ÿüŽ ä x�ÿÀŽ â $xUU   j
Ž È $xWW   �
Ž â rj#j   UXŽ ä x[[˜º  ÿ    ä r�‹�   WXŽ ä x##˜º  ÿ    ä xjj˜º  ÿ    ä xYYÿÀŽ Ä xŒ[€€€€ŒŽ äx[[Ž`Ž ä x��˜º  ÿ    ä x‹‹˜º  ÿ    ä xŽj€€€€UŽ äx#j#`Ž Ä x!!˜º  ÿ    ä xjY˜º  ÿ    ä xW�€€€€WŽ äx’�‹`Ž ä x�!€€€€ Ž äx!!j`Ž Ä r_ÿB   _  ä rpÿA   p  ä r�!�   ÿüŽ â $x!_   ÿ Ž â raÿD   a  ä rtÿC   t  ä xpp   ÿÀŽ Ä rŽ#Ž   ÿüŽ â $x#a   ÿ Ž â rcÿF   c  ä rlÿE   l  ä xtt   ÿÀŽ ä x!!0   pâŽ ä rŒ[Œ   ÿüŽ â $x[c   ÿ Ž â r]ÿB   ]  â &r hk   ÿ   â xll   ÿÀŽ Ä x##0   tâŽ ä x!!àÿÿÿÿàÿ ä rnÿA   n  â $xU]   ÿ Ž â x[[0   lâŽ â &rlfg   ÿ   â x##àÿÿÿÿàÿ â $r!    ÿŽ â xnn   ÿÀŽ Ä reÿD   e  â $rl#l   ÿŽ â rrÿC   r  â 0r ÿo    A   â rj’W   ÿüŽ â $xee   ÿ Ž â xUU0   nâŽ ä ErW        ä riÿF   i  Ä xrr   ÿÀŽ â 0rmÿm    A   ârvÿE   v  â &rn�k   ÿ   â xUUàÿÿÿÿàÿ ä ErY l       â #rl W   ÿ   â $xii   ÿ Ž â xre0   râŽ ä xvv   ÿÀŽ â $rnUn   ÿŽ â #rlmY   l   â xYràÿÿÿÿàÿ â &rpŽg   ÿ   â xvi0   vâŽ Ä Er n       â $rpYp   ÿŽ â xW[àÿÿÿÿàÿ â &rrŒq   ÿ   â x[vàÿÿÿÿàÿ ä rzÿI   z  â #rn    ÿ   â rxÿG   x  â &r  q   ÿ   â Erp p       â $rr[r   ÿŽ â xxx   ÿÀŽ â $x]z   ÿ Ž â r|ÿG   |  â $r W    ÿŽ â r~ÿI   ~  â 0rsÿs    A   â Er_ r       â #rnmp   n   â x]]0   xâŽ â &rph-   ÿ   âx||   ÿÀŽ â &rv�-   ÿ   â x~   ÿ   â #r_s_   n   â Er          â $rp!p   ÿŽ â x]]àÿÿÿÿàÿ â &rn"{   ÿ   â x|0   |âŽ â &rxŽ   ÿ   Ä $rUv   ÿŽ â #ras    l   â 0r ÿ}    A   â Er- p       â $rn]n   ÿŽ â Er        â $rxYx   ÿŽ ä 0rlÿw    A   â Ern n       â #rp -   ÿ   â0ruÿu    A   â Er- x       â #r     ÿ   â &rrf   ÿ   â #raln   a   Ä #rnu-       â &r Œ,   ÿ   ä&rt ,   ÿ   ä $rr#r   ÿŽ ä $r[    ÿŽ ä $rtWt   ÿŽ â Err r       â &r "€   ÿ   Ä 0ryÿy    A   â Ert t       â $r ]    ÿŽ â #rpur   p   â x,|àÿÿÿÿàÿ â &r{j{   ÿ   ä&r-j€   ÿ   â Er          â 0r�ÿ�    A   â #rpyt   p   â $r{,{   ÿŽ âEr        â $r-,-   ÿŽ â #r �    p   â 0r‚ÿ‚    A   â Er{ {       â #rny   n   â Er- -       Ä #r3‚    3   â &r hƒ   ÿ   â #r_l{   _   â #r-�-   n   â 0r$ÿ$    A   ä &rn�ƒ   ÿ   ä &rlf)   ÿ   ä $r!    ÿŽ â #rS$-   S   â $rnUn   ÿŽ Ä $rl#l   ÿŽ â Er-        â 0r ÿ„    A   â Er n       â 0r(ÿ(    A   â Erl l       ä #r- -   ÿ   â &rnŽ)   ÿ   â #r    ÿ   â &r  +   ÿ   â#r-(l   -   â &rlŒ+   ÿ   Ä $rnYn   ÿŽ ä $r W    ÿŽ ä $r)[l   ÿŽ ä &rl"%   ÿ   âErn n       â &r%j%   ÿ   â Er          â 0r*ÿ*    A   Ä $rl]l   ÿŽ â Er) )       â $r%,%   ÿŽ â #r(n      â #r-*    -   â 0rˆÿˆ    A   â Erl l       â &rhh†   ÿ   âEr  %       â #r*)      â &r†�†   ÿ   â #r-ˆl   -   â$rh!h   ÿŽ â #rˆˆ       â &rff‡   ÿ   Ä $rU†   ÿŽ ä &rŽŽ‡   ÿ   â Er! h       â $rf#f   ÿŽ ä 0r ÿŠ    A   â Er        â &r  …   ÿ   ä&rŒŒ…   ÿ   Ä $rŽYŽ   ÿŽ â Erf f       â &r""‰   ÿ   â x22   ÿàÿ â 0r&ÿ&    A   â #r! !   ÿ   â $r W    ÿŽ â ErŽ Ž       â $rŒ[Œ   ÿŽ ä &r‰j‰   ÿ   â #r    ÿ   â | 2   pbô â $r"]"   ÿŽ â #rf&f   !   â $r‰,‰   ÿŽ â Er          â 0r'ÿ'    A   â Er! Œ       â #rŽ&Ž      â 0rÿ    A   â Er" "       â #r'    f   âEr  ‰       â #r!'!   Ž   â x11p  ÿà÷ â #r"      Ä #r!    !   â #rK‚a   K   â #rQ$_   Q   â #r4‚-   4   â #rO$ˆ   O   â #rM‚   M   â #r6$!   6   â x00@   ÿàÿ â $r7ÿÿ   7Ž â x::@   ÿàÿ ä x<<@   ÿàÿ Ä x>>@   ÿàÿ â D)       À â Gy  °ãÿÿÿÿƒ ê Ay        € ê Ey  °     € â G‰  �     € ê $x..   /Ž Ê ˆs .K üÿ H   è ˆs .Q€üÿ H   è ˆs .3 ýÿ H   è ˆs .S€ýÿ H   è ˆs .4 þÿ H   è ˆs .O€þÿ H   è ˆs .M ÿÿ H   è ˆs .6€ÿÿ H   ä Ay        € ê {          ì M        € ê „y/ €  H   h „y /     H   ¨ „y/ €  H   è „y/ €   H   ( „y/    H   h „y/    H   ( „y/ €	  H   ( „y/ €  H   ( „y/    H   ( „y/    H   ( „y
/ €  H   " !rS       Æ/ „y/    H   b !r  K       ÆO „y/    H   ¢ !r       Æ� „y	/ €  H   â !rQ       Æ„y/    H   " !r         Æ„y/ €  H   b !r3       Æ „y/  	  H   " !r       Æ „y/    H   " !r       Æ „y/  
  H   " !r         Æ „y/ €
  H   " !r4       Æ „y/    H   " !r

O       Æ „y/ €  H   â !r       Æ/ „y/ €  H   â !rM       ÆO „y/ €  H   â !r	       â� !r       ä‰       b !r

       Æ‰	      ¢ !r       Æ ‰      " !r       Æ ‰      â $vÿ r  ÿ Ž â !r       ä y      %   " !r

       Æ ‰      b !r       Æ ‰
      " !r        â/ !r6       ä ‰      b !r		       âO !r       â !r       â ‰       $ !r       â !r       â� ‰	      ¨ ‰      â !r       Æ‰      b !r
       Æ‰      " !r       Æ/ ‰
      h ‰      b !r         Æ ‰      " !r	       âO !r       ä� ‰  €    ¢ !r       Æ‰	 €    â !r       Æ‰ €    " !r
       Æ/ ‰ €    b !r       Æ ‰ €    b !r       Æ ‰ €    ( ‰ €    " !r        ÆO ‰      ¢ !r
	       â� $x   ÿ Ž ä y      &   â !r       äz  r  p`ð â y      '   " !r       ä/ r /ÿ   p"p â ‰ @    b !r       â!r       Ä ‰ @    " !r       Æ ‰ @    b !r       ÆO ‰
 @    ¢ $z  x  Ž Æ� ‰ @    â x   ÿàÿ â $z  ~   Ž ä‰ @    " z  r  p`ò â $x   ÿŽ â !r       â/ ‰ @    b 
  r  ÿàõ ä x /   p"ò â ‰ €    b !r	       â $ÿÿ   ÿ â 
 j  ÿ„ â ‰      " !r       âr    ÿ† Ä 
 k   â !r
       âO ‰	      ¢ $rÿÿ   ÿŽ â !r       â�    ÿàÿ â ‰      â z j  ÿ† â!r       â    ÿàõ â ‰      " z k  � â !r       â/     ÿà÷ â ‰      b $ÿÿ   ÿ â !r       â  j  ÿ„ â ‰      b $ÿÿ   ÿŽ â !r       â $tÿ   ÿ Ž â ‰ @    " 

 j  ÿ† Ä %v  j   Ž â ‰      " !r	       âO 	 k   â !r       â� 
 k  � â †	     â !r       Æ†    â !r       Æ/ †	     â !r       Æ†     â !r       Æ †	     â !r       Æ †    è ‰      ( †	 
    â M™        € ê x   ÿàÿ âO !r       Æ r    ÿàñ Ê $r ÿÿ   ÿ  â z j  ÿ€ È z k     Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      "   " ¹z  l      â Ey  0"    € â ™x?    â y       !   b ¹z  F   
   â $rÿÿ   ÿ Ž ä $r-ÿÿ   ÿ Ž ä $rSÿÿ   ÿ Ž Ä $rÿÿ   ÿ Ž ä $r)ÿÿ   ÿ Ž â r ÿ   pRð är ÿ   pRò â$x     Ž Ê/ xÿ     È r   ÿ � â $rÿÿ   ÿ Ž ä $À    Ž â xÿ     È |    pbô Ú G)   !    € ê x	ðÿÿÿÿÀŽ â y      '   " $rÿÿ    Ž ä $rÿÿ   ÿ Ž â y      %   b $x	   	
Ž ä $v0ÿ X  ÿ Ž â y      &   ¤ xÿ   	  È r
	   ÿ� â $rÿÿ   ÿ Ž Æ x
øÿÿÿÿÀŽ Ê $x	   
Ž ä 'z y   Ž Æ xÿ     â$z }  ÿŽ ä $x   ÿ Ž â/ xÿ     â $z w  ÿŽ âO x     â 'z s   Ž â r   ÿ� ä zÿ z    â $z p  Ž â r   ÿò� â $x   ÿ Ž â zÿ t    â $z |  Ž â x   ÿàÿ â $z |  Ž â xÿ     â $z v  Ž âx.   ÿàÿ â $z4 v  Ž â r   ÿ� â$zƒ p  Ž â r   ÿ� â %x44n   ÿŽ â xJøÿÿÿÿÀŽ Ä x&üÿÿÿÿÀŽ â $x		   ÿ Ž â xÿ   .  â $xJ   J
Ž âx2   ÿàÿ â %xD	   4Ž â x&   &èÿ äx   ÿàÿ â $xG   ÿ Ž â r.   ÿ� â$x   ÿ Ž â r.   ÿ� â %xBG   4Ž â xÿ   2  Ä xÿ     ä r2   ÿ� är   ÿ� är2   ÿ� â $rÿÿ   ÿ Ž â r   ÿ� ä xøÿÿÿÿÀŽ Ä x+øÿÿÿÿÀŽ ä x$øÿÿÿÿÀŽ â $x.   
Ž â x/üÿÿÿÿÀŽ â $x*2   +
Ž â x3üÿÿÿÿÀŽ â $x$   $
Ž â x"üÿÿÿÿÀŽ â $vÿ q  ÿ Ž â x..   /èÿ Ä x22   3èÿ â $x   Ž â x"   "èÿ â $vÿ Y  ÿ Ž â x'ÿ   &  ä xKÿ   J  ä x#ÿ   "  â %x6ƒn   &Ž â x%ÿ   $  Ä x3ÿ   2  â %x8ƒn   JŽ âx+ÿ   *  ä x/ÿ   .  â %x€ƒn   "Ž âxÿ     ä r   ÿàõ â %x~ƒn   $Ž â r"4   ÿà÷ Ä r$4   ÿàý â %x:ƒn   2Ž âxÿ   
  ä xÿ     â %x<ƒn   *Ž âxÿ     ä xÿ     â %x>ƒn   .Ž â xÿ     Ä xÿ     â %x@ƒn   Ž âxÿ     ä xÿ     â %x‚ƒn   ÿŽ â r 24   ÿàù äz q  ÿàÿ â $rÿÿ   ÿ â rJ4   ÿàõ â %xH	   ‚Ž â x	ÿ     Ä r&4   ÿàû â$x"5   #Ž âr!*4   ÿà÷ â$x$5   % âr#.4   ÿàý â$x&5   'Ž â r%4   ÿàû â $x   ÿ Ž ä %xFG   ‚Ž Æ xÿ     â $x   ÿ Ž ä $x   ÿ Ž ä $x   ÿ Ž ä $x   ÿ Ž ä $x   ÿ Ž ä $x   ÿ Ž Ä $x   ÿ Ž ä $x'	   ÿ Ž ä $x(5   K ä$x*5   3 ä$x,5   +Ž ä$x.5   / Ä $x
5   Ž ä r2B0   ÿà÷ ärPD0   ÿàõ Æ $x3C   Ž ä$xQE    Æ �yt2    • ¨  �yM2   • ¨  �yP"   • è �yzP    • â rNF0   ÿàõ Ê $xOG    â rTH0   ÿàõ È �yXN    • h	 �y1N   • b	 $xUI    Ê �yuT"   • h �yRT    • b	 rJ0   ÿàõ är20   ÿà÷ Æ $xK   ( ârr0%   ÿàõ â$x3   &Ž ârP0#   ÿà÷ ä/�y]J`   ‘ b $xs   
 ârZ0!   ÿàõ â$xQ   .Ž ârV0    ÿà÷ â �yO2`   ‘ $$x[   , ârT80   ÿàõ â$xW   *Ž â rJ60   ÿàù â$tÿ$   ÿ Ž â �yZZ`   ‘ " r2@0   ÿà÷ â/ $xU9    Ä �yoV`   ‘ $ $x3A   Ž ârw   àµ â $xK7    â �yYT`   ‘ $ r	   ÿd~ â %vvw Z   Ž â rb:0   ÿàõ â �ymJ`   ‘ $ $x		$   ÿŽ Ä �yqP`   ‘ " $xc;    â rh<0   ÿàû â $xww   	Ž â �yrr`   ‘ " rJv   ÿàõ Æ �yp2`   ‘ " rP>0   ÿàù â/ $xK   w ä �ygb`   ‘ " $xi=   Ž ä$xQ?    â �ynJ   ™ " r20   ÿàõ Æ �ykP`   ‘ ( �yhh`   ‘ " $x3   $ Æ �yVJ(   ™ " r`0   ÿà÷ ä rP€0   ÿàõ â �yUv    • $ $xa   "Ž ä $xQ�    â �yWv$   • " r‡   àµ Ä r^~0   ÿàù â �yT``   ‘ " r	   ÿd~ â %v†‡ Z   Ž ä �yfJL   ™ $ $x_    ä $x		$   ÿŽ â �y\2`   ‘ ( �yj^`   ‘ " $x‡‡   	Ž Æ �yivH   • "# r2†   ÿàõ Æ �y+vl   • $ $x3   ‡ â r	   àµ â �ylP`   ‘ & r/   ÿd~ â �y^2   ™ " r„40   ÿàõ â %vv	 Z   Ž Ä/ �ybJp   ™ $ $x//$   ÿŽ ä �yP2(   ™ " $x…5    â rxv   ÿàõ â $xww   /Ž â �ya†    • ( �yd2L   ™ " $xy   w Æ �yc†$   • ( �yx   ™ ( �y_†H   • ( �yJx(   ™ (# �y`2p   ™ ( �y[„l   • ( �yKv    • ( �y2xL   ™ " r|‚0   ÿà÷ Æ �yLxp   ™ ( �y/v$   • ( �yNvH   • " $x}ƒ   Ž Æ �ye†l   • ( �y	vl   • ( �y3|l   • " xMtT  M    ÈO rtÿM   ÿ3Ž È rtÿ   t  ä xzzT      â� $tÿÿ Ž ä $xQt   ÿ Ž â xMÿ   z  ä xxz€€€€ìŽ ä/ xQQÿÀŽ Ä xMMÿÀŽ ä x{ÿ   z  â$xx   Q
Ž â x}M€€€€ÿüŽ â $xst   ÿ Ž âxxtÿÀŽ ä x{{ÿÀŽ ä xvQzøŽ Ä xyzxŽ â $x}}   x
Ž â xQÿ   z  ä xx{€€€€ÿüŽ ä xssÿÀŽ ä ryyv   ÿÀŽ ä xvMtøŽ â $xxx   s
Ž â xQQÿÀŽ Ä xtÿ   t  ä rvv}   M`Ž ä xwQ€€€€ÿüŽ ä xttÿÀŽ ä xMM˜º  ÿ    ä xvv˜º  ÿ    Ä rs{s   xXŽ â $xww   t
Ž â x|v€€€€}Ž äxMvM`Ž ä xss˜º  ÿ    ä xv{˜º  ÿ    ä x1XT  1    Äxxs€€€€xŽ ä rtQt   wXŽ ä xzzÿÀŽ ä xssv`Ž ä rvÿ1   ÿ3Ž ä x1Q˜º  ÿ    Ä xXt˜º  ÿ    ä xQz˜º  ÿ    ä xzy˜º  ÿ    ä rvÿ   v  ä xtX€€€€wŽ äx1X1`Ž Ä xXz€€€€Ž äxzQ`Ž â $tzÿÿ Ž â xuRT  u    â $xRv   ÿ Ž Æ xwu€€€€zìŽ ä xRRÿÀŽ ä r11t   ÿüŽ Æ $xtw   R
Ž â xwRuøŽ ärMM|   ÿüŽ ä xRtuxŽ äx|ÿ   u  ä rX   ÿüŽ ä rXRw   ÿÀŽ â $xRv   ÿ Ž â x||ÿÀŽ Ä xyÿ   u  ä xw|€€€€ÿüŽ ä xRRÿÀŽ ä rQsx   ÿüŽ ä xsyÿÀŽ â $xww   R
Ž â xyvÿÀŽ Ä xxs€€€€ÿüŽ ä rR|R   wXŽ äx||˜º  ÿ    â $xxx   y
Ž â xyÿ   u  ä xRR˜º  ÿ    ä xuuÿÀŽ Ä xwR€€€€wŽ äx|R|`Ž ä x}X˜º  ÿ    ä xRu˜º  ÿ    ä rOÿ   O  ä	rmÿ   m  Ä rZÿ   Z  ä xt}€€€€tŽ är]ÿ   ]  ärYÿ   Y  ä x}}R`Ž â $xRO   ÿ Ž â roÿ   o  â $xmm   ÿ Ž â rqÿ   q  Ä xOZ   ÿÀŽ ä rrÿ   r  ä x]]   ÿÀŽ ä xZY   ÿÀŽ â $xXo   ÿ Ž â r}}t   ÿüŽ â $xqq   ÿ Ž â xrr   ÿÀŽ Ä xRR0   ]âŽ ä xZm0   ZâŽ ä xOX0   OâŽ â &rmn   ÿ   âxzsvøŽ â &ro}n   ÿ   â xYq0   râŽ ä rpÿ   p  Ä xXRàÿÿÿÿàÿ ä xZZàÿÿÿÿàÿ ä r]ÿ   k  ä rkÿ   h  â &rRQV   ÿ   â rzzx   s`Ž â $rmXm   ÿŽ â xhp   ÿÀŽ â $rpZo   ÿŽ â xYYàÿÿÿÿàÿ â $x]]   ÿ Ž â rgÿ   g  Ä xss˜º  ÿ    â $roYR   ÿŽ â xzz˜º  ÿ    â 0rUÿU    A   â Ern m       â $xrg   ÿ Ž â ErR p       ä xxz€€€€xŽ äxszs`Ž Ä rw|w   ÿüŽ ä xkk   ÿÀŽ ä x]]0   hâŽ â 0rWÿW    A   â Ero o       â #rnUn   ÿ   â#rhUR   ÿ   â rssx   ÿüŽ ä xkr0   kâŽ Ä xR]àÿÿÿÿàÿ â &r]wV   ÿ   â #rUWo   n   â rgÿ'   T  â &rnMf   ÿ   âxOOàÿÿÿÿàÿ â $r]R]   ÿŽ â xTkàÿÿÿÿàÿ â &rksf   ÿ   â r\ÿ   \  â$rnOn   ÿŽ â rjÿ   j  â $rkTk   ÿŽ â xyyÿÀŽ Ä xvÿ   v  ä Er] ]       â $xgg   ÿ Ž â x\\   ÿÀŽ â 0rfÿi    A   â xVj   ÿÀŽ ä Ern n       â #r]W]   h   â x{y€€€€ÿüŽ Ä xvvÿÀŽ ä Erj k       â #rWfn   U   âxgg0   \âŽ â $x{{   v
Ž ä #rUfj   ]   â 0r\ÿ+    A   â x+gàÿÿÿÿàÿ â &r]^   ÿ   Ä &rg}^   ÿ   â rvyv   {XŽ â &r^QP   ÿ   ä$r]X]   ÿŽ ä &riwP   ÿ   ä $rPZg   ÿŽ â xvv˜º  ÿ    â &rh1b   ÿ   â xyy˜º  ÿ    â $rgY^   ÿŽ â Er^ ]       â 0raÿa    A   â ErP P       â &rfMd   ÿ   Ä $riRi   ÿŽ ä &r]sd   ÿ   â x{v€€€€{Ž â$rh+h   ÿŽ â xvvy`Ž â 0rcÿc    A   â Erg g       â #r^a^   ÿ   â rlÿ'   l  â $rfOf   ÿŽ â Eri i       â $r]T]   ÿŽ â Erh h       â #rjaP   ÿ   â r{v{   ÿüŽ â #rgcg   ^   â &r   ÿ   â Erf f       â $xml   ÿ Ž â Er] ]       â 0r^ÿ_    A   â #rcci   j   â &r}}   ÿ   â #rP\h   W   â &rW{b   ÿ   Ä $rX   ÿŽ ä &rQJ   ÿ   â xVm0   VâŽ â &rb1`   ÿ   â #rf^f   g   â&r_{`   ÿ   â #r`^]   c   â $r}Z}   ÿŽ ä &rwwJ   ÿ   Ä 0r^ÿ[    A   ä &rJM2   ÿ   ä&rss2   ÿ   â Er2        â $rYY   ÿŽ â xVVàÿÿÿÿàÿ â $rwRw   ÿŽ â Er }       â 0rKÿK    A   â #r)^P   )   â $rJOJ   ÿŽ â x   ÿàÿ â $rsTs   ÿŽ â ErY Y       â &rP1L   ÿ   â #r2K2   ÿ   â &r{{L   ÿ   â Erw w       â 0r/ÿ/    A   â #rK   ÿ   â $rWVW   ÿŽ â |    pbô â $rb+b   ÿŽ â ErJ J       â $r_V_   ÿŽ â #rY/Y   2   â $rP+P   ÿŽ â Ers s       â 0rNÿN    A   Ä $r{V{   ÿŽ â #rw/w      â ErW W       â 0reÿe    A   â Erb b       â #rJNJ   Y   âEr_ _       â 0r	ÿ	    A   â ErP P       â #rNNs   w   â Er{ {       â #rW\W   U   â 0r\ÿ3    A   â x00p  ÿà÷ â #rbeb   f   â#rUe_   `   â #rJ	P   J   â#rN	{   N   â #rS^b   S   â #r\W      â#r-\U   -   â #r^J      â #r\N      â x@   ÿàÿ â $rÿÿ   Ž â x@   ÿàÿ Ä x@   ÿàÿ â D)       À â Gy  0çÿÿÿÿƒ ê Ay        € ê ˆ ) ýÿ H   è ˆ €ýÿ H   è ˆ S þÿ H   è ˆ -€þÿ H   è ˆ  ÿÿ H   è ˆ €ÿÿ H   è {          ì M        € ê „y      H   ( „y     H   h „y  €   H   ¨ „y     H   è „y  €  H   ( „y	     H   h „y
     H   h „y  €  H   h „y     H   h „y  €  H   h „y  €  H   b !r)       Æ „y     H   " !r       Æ/ „y  €  H   b !r       ÆO „y     H   ¢ !rS       Æ� „y  €  H   â !r-       Æ„y  €  H   â !r	       Æ„y     H   â !r

       Æ „y  €  H   â !r       â !r       ä ‰      " !r       Æ y      %   b !r       Æ y      &   b !r

       â !r       â/ !r       ÈO ‰      " !r       â� !r       ä‰      b !r       â $x   ÿ Ž ä‰      ¢ !r
       Æ ‰
      â z  r  p`ð â !r       ä ‰      " r  ÿ   p"p Æ ‰      b !r       â !r       È/ ‰      " $vÿ r  ÿ Ž â !r       äO ‰	      b !r

       Æ� ‰ €    ¢ !r       Æ‰
      â !r       Æ‰      ( ‰      b !r       â !r		       È/ ‰ €    " !r       ÆO ‰	 €    b !r
       Æ� ‰ @    ¢ !r       Æ‰ €    â !r       Æ‰
 €    ( ‰ €    b !r       â !r	       È/ ‰ @    " !r       âO $z x  Ž â !r	       Æ� ‰      b !r
       Æ‰ @    ¢ x
   ÿàÿ â !r       ä‰ @    â z 
 r  p`ò Æ ‰	 @    " x     p"ò Æ y      '   b !r       Æ ‰ @    ( ‰       " !r       â/ $tÿ   ÿ Ž â !r
       âO !r       È� ‰
      b !r	       â	   ÿàÿ â $z ~  Ž Æ‰      ¢ !r       â 
 r  ÿàõ ä ‰      " r   ÿ† â !r        â $ÿÿ   ÿ â 
 j  ÿ„ â‰      æ 
 k   Ä 	   ÿàõ â $r	ÿÿ   ÿŽ â z j  ÿ† â %v j   Ž â !r
       ä/ z k  	� â $ÿÿ   ÿ â  j  ÿ„ â!r       âO †	     ä 	 k   â !r       â †    è †	     â !r       Æ� †     è †	     â M™        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      "   " ¹z  l      â Ey  €    € â ™x?    â y       !   b ¹z  F   
   â $rÿÿ   ÿ Ž ä $rÿÿ   ÿ Ž ä $r*ÿÿ   ÿ Ž â r ÿ   pRð Ä r ÿ   pRò â$x     Ž Ê/ xÿ     È r   ÿ � â $rÿÿ   ÿ Ž ä $€    Ž â x)ÿ     È | )   pbô Ú G)  p    € ê x	ðÿÿÿÿÀŽ â y      '   " $rÿÿ   ÿ Ž Æ y
      %   b $x   	
Ž Æ y      &   ¢ $xI   ÿ Ž â xÿ     È r   ÿ� È xøÿÿÿÿÀŽ ä xÿ     Æ $x   
Ž ä 'z y   Ž È $z	 }  ÿŽ â xÿ     â $x
   ÿ Ž â/ zÿ z    â $r
ÿÿ   ÿ Ž â r   ÿ� â $z p  )Ž âx   ÿàÿ â 'z
 s  
 Ž âO r   ÿò� Æ $z w  ÿŽ â z
ÿ t  
  â $z |  )Ž âxJ   ÿàÿ â$z |  Ž â xÿ     â $z
 v  Ž âxN   ÿàÿ â $z
 v  Ž â r
   ÿ� â$zA p  Ž â r   ÿ� â %x.n   ÿŽ â xP
øÿÿÿÿÀŽ Ä xÿ   J  â $x)   ÿ Ž â x&üÿÿÿÿÀŽ â $xP   P
Ž âx"   ÿàÿ â $x   ÿ Ž â rJ   ÿ� â$xG   ÿ Ž â rJ   ÿ� â %xDI   .Ž â x&   &èÿ Ä xÿ   N  â %xBG   .Ž â xÿ   "  ä rN   ÿ� â $rÿÿ   ÿ Ž â r"   ÿ� ä rN   ÿ� ä r"   ÿ� â $vÿ Y  ÿ Ž â x-øÿÿÿÿÀŽ Ä xMøÿÿÿÿÀŽ ä x%øÿÿÿÿÀŽ â $x,J   -
Ž â xKüÿÿÿÿÀŽ â $xLN   M
Ž â xOüÿÿÿÿÀŽ â $x$"   %
Ž â x#üÿÿÿÿÀŽ ä xJJ   Kèÿ Ä xNN   Oèÿ ä x""   #èÿ ä x'ÿ   &  ä xQÿ   P  ä x#ÿ   "  â %x0An   &Ž â x%ÿ   $  Ä xOÿ   N  â %x2An   PŽ âxMÿ   L  ä xKÿ   J  â %x4An   "Ž âx-ÿ   ,  ä x ÿ     â %x6An   $Ž â r		   ÿàõ Ä r&.   ÿàû â%x8An   NŽ âr".   ÿà÷ är$.   ÿàý â %x:An   LŽ âxÿ     ä xÿ   
  â %x<An   JŽ â xÿ     Ä xÿ     â %x>An   ,Ž âxÿ     ä xÿ     â $rÿÿ   ÿ â xÿ     â %x@An   ÿŽ â xÿ     ä rP.   ÿàõ â$x/   #Ž âr!N.   ÿàù â$x/   % âr#L.   ÿà÷ â$x/   'Ž â r%J.   ÿàý â$vÿ X  ÿ Ž â r',.   ÿàû â %xFG   @Ž â x     Ä z q  ÿàÿ â $x   ÿ Ž â xÿ     â %xHI   @Ž È $x   ÿ Ž ä $x   ÿ Ž ä $x   ÿ Ž ä $x   ÿ Ž ä $x   ÿ Ž ä $x   ÿ Ž Ä $x   ÿ Ž ä $x     ÿ Ž ä $x"/   Q ä$x$/   O ä$x&/   MŽ ä$x(/   K Ä $x
/   -Ž ä rJD   ÿàõ ärLB   ÿà÷ Æ $xKE    ä$xMC   Ž Æ �yJ"   • ¨  �y,J    • ¨  �y+L    • è �yVL   • â rXF   ÿà÷ Ä rRH   ÿàõ Æ $xYG   Ž ä$xSI    Æ �y-X    • h	 �yZX   • h	 �yWR"   • h �y\R    • b rT   ÿàõ Ä rP   ÿà÷ Æ $xU   " ârN'   ÿàõ â$xQ   Ž ârL%   ÿà÷ ä/�y^T`   ‘ b $xO   
 ârJ#   ÿàõ â$xM   (Ž ârR!   ÿà÷ â�y`P`   ‘ d $xK   & ârX   ÿàõ â$xS   $Ž ârT   ÿà÷ â/ �y[N`   ‘ $ $xY    Ä $xU   Ž â rP2   ÿàõ â�ybL`   ‘ " rN0   ÿà÷ Æ/ �ydR`   ‘ " $xQ3    ä$xO1   Ž â rL>   ÿàõ â�y]J`   ‘ " rR:   ÿà÷ Æ/ $xM?    â�yYX`   ‘ " $xS;   Ž â re	   àµ â $tkÿ$   ÿ Ž â rJ<   ÿàù â �yaL`   ‘ " rg   ÿd~ Æ �yXN`   ‘ "# $xK=    ä $xgg$   ÿŽ â �y_T`   ‘ " %vLe Z  k Ž Æ �yhJ`   ‘ " rN4   ÿàõ Æ/ �yfP`   ‘ " rT8   ÿàù â $xO5    ä �ycR`   ‘ " $xMM   gŽ â rJL   ÿàõ â $xU9    â �yiN`   ‘ & $xK   M â �yTT`   ‘ ( �yjJ   ™ ( �yeJ(   ™ " rP6   ÿàû Æ/ �ymL    • " rR	   àµ â $xQ7   Ž ä �ylL$   • " rU   ÿd~ â %vRR Z  k Ž ä �ygP`   ‘ ( �ykJL   ™ " rNR   ÿàõ Æ �ynLH   • " $xQU$   ÿŽ Æ �yUJp   ™ " $xSS   QŽ Æ �yqLl   • " $xO   S Æ �ysR    • ( �yoN   ™ ( �yrN(   ™ ( �ypNL   ™ ( �ytNp   ™ " rJ.   ÿàõ Æ/ �yuR$   • " rP@   ÿà÷ Æ �yvRH   • " $xK/    ä$xQA   Ž â �yLRl   • ( �yJJl   • ( �yPPl   • " x,T      ÈO x,ÿ     ä x++T  V    È� r+ÿ+   ÿ3Ž ä xN,ÿÀŽ ä rOÿ   +  ä x,N€€€€ÿüŽ äx+OÿÀŽ äxKÿ     â $xQO   ÿ Ž â xRNOøŽ â/ $x+,   +
Ž â xKKÿÀŽ â $xwO   ÿ Ž â xOÿ   O  â $t,ÿÿ Ž â xMÿ     Ä rRR+   N`Ž ä xxK€€€€ÿüŽ ä xOOÿÀŽ ä x,€€€€,ìŽ ä xMMÿÀŽ ä xSQÿÀŽ Ä xQwÿÀŽ ä xwR˜º  ÿ    â $xRx   O
Ž â xNN˜º  ÿ    â $x,,   S
Ž â xVM€€€€ÿüŽ ä x+w€€€€+Ž äxzwN`Ž â $xVV   Q
Ž â rOKO   RXŽ Ä xSSøŽ äxx,xŽ ä r+z+   ÿüŽ ä xOO˜º  ÿ    ä xzK˜º  ÿ    ä x--T  Z    ÄrQMQ   VXŽ ärNxS   ÿÀŽ ä xxM˜º  ÿ    ä xRO€€€€RŽ äxMOz`Ž ä rOÿ-   ÿ3Ž Ä xQQ˜º  ÿ    ä x-ÿÀŽ ä rÿ   O  ä xKQ€€€€VŽ äxVQx`Ž ä xW\T  W    â $t\ÿÿ Ž â xQN˜º  ÿ    Ä xN-˜º  ÿ    â $x-   ÿ Ž â xSÿ   W  ä xOQ€€€€,Ž äxZQN`Ž ä xQW€€€€\ìŽ ä x\-ÿÀŽ Ä xNSÿÀŽ ä r-VK   ÿüŽ â $x,Q   \
Ž â xVN€€€€ÿüŽ ä xSÿÀŽ ä rKMR   ÿüŽ ä xRÿ   W  â $xSV   S
Ž â x\\WøŽ Ä xQ,WxŽ â $xV   ÿ Ž â rMZO   ÿüŽ ä xRRÿÀŽ ä rOQ\   ÿÀŽ ä xZNøŽ ä xQÿ   W  Ä xwR€€€€ÿüŽ ä xVVÿÀŽ ä rZZS   N`Ž ä xyN˜º  ÿ    ä xQQÿÀŽ ä xNÿ     â $xw   V
Ž â x\Z˜º  ÿ    Ä x{Q€€€€ÿüŽ ä xZNÿÀŽ ä rNRV   XŽ Æ $xV{   Z
Ž â xw\€€€€SŽ äxx\y`Ž ä x\N˜º  ÿ    ä xSR˜º  ÿ    ä rZQZ   VXŽ ä x\€€€€Ž Ä xWWÿÀŽ ä x\\S`Ž ä xSZ˜º  ÿ    ä xRQ˜º  ÿ    ä xOO˜º  ÿ    ä xZW˜º  ÿ    Ä r`ÿ   `  ä xQS€€€€VŽ äxVSR`Ž ä r^ÿ   ^  ä xRO€€€€,Ž äxWOZ`Ž Ä r_ÿ    _  â$x``   ÿ Ž â rbÿ   b  ä rYÿ   Y  ä r,ÿ   X  ä r[ÿ   [  ä xS^   ÿÀŽ Ä rfÿ   f  ä rQVQ   ÿüŽ â $xV_   ÿ Ž â rRWR   ÿüŽ â $xbb   ÿ Ž â rdÿ   d  ä xWY   ÿÀŽ â $xY,   ÿ Ž â rO\   ÿüŽ Ä x[   ÿÀŽ ä xS`0   SâŽ ä r]ÿ   ]  ä xff   ÿÀŽ â $x[d   ÿ Ž â rTÿ   T  ä xWV0   WâŽ â &r,Mj   ÿ   â xb0   âŽ Ä xX]   ÿÀŽ ä xVY0   fâŽ ä xSSàÿÿÿÿàÿ â $x^T   ÿ Ž â rhÿ   h  â &rY-e   ÿ   â xX[0   XâŽ â &r[Rj   ÿ   â xTàÿÿÿÿàÿ â $r,S,   ÿŽ â xVVàÿÿÿÿàÿ Ä raÿ   a  â $x]h   ÿ Ž â rcÿ   c  â $r[V[   ÿŽ â xZa   ÿÀŽ â $rYTY   ÿŽ â Er\ ,       â 0rÿm    A   â xZ]0   ZâŽ â 0rlÿl    A   â xcc   ÿÀŽ Ä Er, [       â #r[\   ÿ   â Er] Y       ä rNxw   ÿüŽ ä xc^0   câŽ â #r[l]   [   â xXXàÿÿÿÿàÿ â &r]+k   ÿ   â#r_,   ÿ   â xcàÿÿÿÿàÿ â &r\Nk   ÿ   â riÿ    i  â $r]X]   ÿŽ â xWWàÿÿÿÿàÿ â &r,KU   ÿ   â rgÿ   g  â $r\\   ÿŽ â Er] ]       â $x`i   ÿ Ž â xgg   ÿÀŽ â $r,W,   ÿŽ Ä 0rnÿn    A   â Er^ \       â 0rqÿq    A   â xY`0   gâŽ â &r`Mo   ÿ   â Er\ ,       â #r,n]   [   â &roRo   ÿ   â xZZàÿÿÿÿàÿ â &reOe   ÿ   â #rMq\   ,   â $r`S`   ÿŽ Ä $r,Vo   ÿŽ ä &r--r   ÿ   ä&rOOr   ÿ   ä &rNNp   ÿ   â Er` `       â $reZe   ÿŽ â Er, ,       â &r++p   ÿ   Ä $r-T-   ÿŽ ä $rOZO   ÿŽ ä $rSN   ÿŽ ä 0rÿs    A   â xYYàÿÿÿÿàÿ â &rTKt   ÿ   âx))   ÿàÿ â $r+X+   ÿŽ â Ere e       â &rNQU   ÿ   âEr- -       â &rtQt   ÿ   â ErO O       â 0ruÿu    A   â #r``   ÿ   â#r,,   ÿ   â | )   pbô â $rTWT   ÿŽ â Er+ +       â $rNYN   ÿŽ âErS S       â $rtYt   ÿŽ â #r_le   _   â 0rvÿv    A   â #r-u-   `   â#rOuO   ,   â ErN N       â 0rLÿL    A   â ErT T       â #rRn^   _   â Ert t       â #r+v+   -   â#rOvS   O   â 0rÿJ    A   â xp  ÿà÷ â 0rPÿP    A   â #rNqN   R   â #r+LT   +   â#rOLt   O   â #r*M   *   â#rPN      â#r+      â #rPO      â x@   ÿàÿ â $rÿÿ   Ž â x@   ÿàÿ â D)       À â Gy  �êÿÿÿÿƒ ê Ay        € ê ˆ * þÿ H   è ˆ €þÿ H   è ˆ  ÿÿ H   è ˆ €ÿÿ H   è {          ì M        € ê „y      H   ( „y  €   H   h „y     H   ¨ „y     H   è „y	  €  H   ( „y
     H   h „y     H   h „y  €  H   h „y     H   h „y  €  H   h „y  €  H   b !r*       Æ „y  €  H   " !r       â/ !r       äO y      &   b !r       â� !r	       â!r

       ä!r       â !r       È ‰      ¢ !r

       Æ ‰      â !r       Æ ‰	
      " !r       Æ y      '   b !r       â !r       âO !r       È� ‰      " !r
	       Æ‰      ¨ ‰      è ‰      " !r       â !r       ÄO y      %   " !r
       Æ� ‰	 €    ¢ !r       Æ‰ €    è ‰
 €    ( ‰      b $x   ÿ Ž â !r		       ÈO z  r  p`ð â !r       â� ‰	 @    $ r  ÿ   p"p â !r
       â‰ @    ¢ !r       Æ/ ‰ @    h ‰ €    â !r	       â $t	ÿ   ÿ Ž â !r       ÆO ‰
      â $z x  Ž â !r       ä/ ‰      " x   ÿàÿ â $z ~  Ž â!r       ä� ‰      b z  r  p`ò Æ ‰ @    ¢ x     p"ò ä 
  r  ÿàõ Ê $ÿÿ   ÿ â 
  j  ÿ„ È 
  k   â !r       â %v j  	 Ž â !r	
       â !r       Æ/ †	     â !r       ÆO † 	   è ‰      h  †	     â M™        € ê ¹z  r      â !r       â/ �x   ?àÿ Ì |   ÿàñ Ê $rÿÿ   ÿ  â z j  ÿ€ È z k    Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž ä $rÿÿ    Ž È % \   Ž Ê �	    ™ b ¹z  l      â Ey  à    € â ™x?    â y      "   b $r3ÿÿ   ÿ Ž Æ y      !   ¢ $ÿÿ   ÿ Ž ä $ÿÿ    Ž â y      %   & '
  m   Ž È $‚ÿÿ   ÿ Ž â 
 ÿ n     Ê $ÿÿ    
Ž â r ÿ   pRò Æ/ $
 o  Ž â r ÿ   pRô â$x    Ž äO $‚ÿÿ    Ž Æ xÿ     â 'Š s   Ž Æ r    ÿ � È xÿ      È |    pbö ä Šÿ t    Ö G9  `    € ê y      '   " x) ðÿÿÿÿÀŽ â $rÿÿ   ÿ Ž ä $tÿn   ÿ Ž ä $x)   )
Ž ä $r3ÿÿ   ÿ Ž ä $x   ÿ Ž â x ÿ   )  È r  )   ÿ� È x'ÿ      â 'z y   Ž â x øÿÿÿÿÀŽ Æ $z }  ÿŽ â zÿ z    â $x)   
Ž ä $x''   ÿ Ž ä $z |  Ž â xÿ     â $x))   ÿ Ž â x+ÿ   '  â $z v  Ž âr	   ÿ� â $x-   ÿ Ž Ä $z p  Ž â r		   ÿò� â $z w  ÿŽ ä %v X  Ž â xÿ   	  ä x"	   ÿàÿ ä r	   ÿ� ä xÿ   "  Ä r	   ÿ� ä x$	   ÿàÿ äx&	   ÿàÿ ä r   ÿà÷ ä xøÿÿÿÿÀŽ ä r"   ÿ� â $r(ÿÿ   ÿŽ â x üÿÿÿÿÀŽ â $x	   
Ž â xÿ   $  Ä xÿ   &  ä r"   ÿ� ä x!øÿÿÿÿÀŽ ä x 	    èÿ ä r$   ÿ� â $x!"   !
Ž â r
&   ÿ� Ä x	üÿÿÿÿÀŽ ä r$   ÿ� ä r&   ÿ� ä x#øÿÿÿÿÀŽ ä x%
øÿÿÿÿÀŽ ä xl   ÿàñ â $x#$   #
Ž â x""   	èÿ â $x%&   %
Ž â x*ÿ     â $rÿÿ     â x,ÿ     Ä x.ÿ     â $x**   ÿ Ž â x/ÿ     â $x,,   ÿ Ž â x0ÿ     â $x..   ÿ Ž â x1ÿ     â $x//   ÿ Ž â x2ÿ   
  â $x00   ÿ Ž â xÿ     â $x11   ÿ Ž â x	üÿÿÿÿÀŽ â $x22   ÿ Ž â xüÿÿÿÿÀŽ â $x   ÿ Ž â x$$   	èÿ Ä x&&   èÿ ä %x)   Ž È %x-   Ž â �y ¶ÿÿ • ¨  �y5´ÿÿ • ¨  �y4”ÿÿ • è �y7–ÿÿ • â r   ÿàñ Ä r
'   àÇ â $tÿ$   ÿ Ž â r	   ÿ  â �y<    • b	 r    ÿàñ ä r+(   ÿ„þ â %v
 Z   Ž â r"   ÿàù â�y9ôÿÿ ‘ b r!   ÿà÷ â $x$   ÿŽ â r    ÿ  Ä r"   ÿ är#   ÿàù ä/r!   ÿ� â�yôÿÿ ‘ b r
$   ÿà÷ â$x   Ž â r#   ÿ â �yôÿÿ ‘ b r&   ÿàù Ä r$   ÿ� â�yôÿÿ ‘ b r%   ÿàñ är&   ÿ â�y

ôÿÿ ‘ b r%   ÿ  â %x   Ž ä �yôÿÿ ‘ h �yôÿÿ ‘ h �yôÿÿ ‘ h �y   ™ h# �y(   ™ h �yL   ™ h �y;p   ™ h �y:    • h �y8$   • h �y6H   • h �yl   • b $tÿÿ Ž â x   ÿàÿ È |    pbð ä xp  ÿà÷ äx@   ÿàÿ Æ $rÿÿ   Ž â x 5T       äO x44T  7    ä� x7ÿ      ä rÿ4   ÿ3Ž ä x77ÿÀŽ ä rÿ'     Ä x5ÿ      ä x €€€€ìŽ â/ $x4   ÿ Ž âx7€€€€ÿüŽ ä xÿÀŽ äx55ÿÀŽ â $x=   ÿ Ž â x7øŽ â $x   
Ž â x	5€€€€ÿüŽ Ä x44ÿÀŽ ä r   7`Ž ä x==ÿÀŽ â $x	   4
Ž â x	ÿ      ä x˜º  ÿ    â $x   =
Ž â x77˜º  ÿ    Ä r454   XŽ ä x		ÿÀŽ ä xÿ     ä x€€€€Ž äx7`Ž ä x= øŽ Ä x74˜º  ÿ    ä x= xŽ ä x4	€€€€ÿüŽ ä xÿÀŽ ä r=   ÿÀŽ ä x55˜º  ÿ    â $x44   
Ž â x  ÿÀŽ Ä x7€€€€Ž äx575`Ž ä x˜º  ÿ    ä x7 ˜º  ÿ    ä r	   4XŽ ä rÿ,     Är9ÿ*   9  ä x€€€€Ž äx7`Ž ä x7˜º  ÿ    â $x   ÿ Ž â rÿ/     ä x99   ÿÀŽ Ä rÿ.     â $x   ÿ Ž â r   ÿüŽ ä r
ÿ1   
  ä x	˜º  ÿ    ä r	   ÿüŽ ä x90   9âŽ Ä x   ÿÀŽ ä rÿ0     â $x

   ÿ Ž â rÿ     â &r	   ÿ   â x 7€€€€4Ž äx7`Ž ä x99àÿÿÿÿàÿ Ä r5   ÿüŽ ä x0   âŽ ä rÿ2     ä x   ÿÀŽ â $x   ÿ Ž â r     ÿüŽ â $r9   ÿŽ â xàÿÿÿÿàÿ â &r   ÿ   â x   ÿÀŽ Ä x
0   âŽ â $r   ÿŽ â x0   âŽ â &r   ÿ   â xàÿÿÿÿàÿ â &r
 ;   ÿ   â Er        â 0r ÿ:    A   â xàÿÿÿÿàÿ â $r   ÿŽ â Er        â 0r8ÿ8    A   â #r    ÿ   â $r

   ÿŽ â Er        â 0r6ÿ6    A   â #r8      â 0rÿ    A   â Er
 
       Ä #r6      â 0r<ÿ<    A   Æ #r
      È #r3<   3   â G‰  €öÿÿÿÿƒ ê Ay        € ê ˆ 3€ÿÿ H   è {          ì M)        € ê „y      H   " r ÿ   pRð â $z x  Ž Ä „y €   H   h „y    H   ¨ y	      '   â !r  3       È !r         È/ !r         ÊO ‰       $ !r        Ê ‰      $ !r       Ê ‰ €    $ !r       Ê ‰ @    $ !r       Ê ‰      b  M	        € ê $tÿ   ÿ Ž â� !r       â/ $z	 ~  Ž È %v j   Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž È % \   Ž Ê �		    ™ ¢  $rÿÿ   ÿ Ž â Ey  Ð    € â $r<ÿÿ   ÿ Ž â y       "   b 'Š s   Ž Æ y      !   b '
 m   Ž Æ y      %   â $r;ÿÿ   ÿ Ž â 
ÿ n    ä rÿ    ÿü‚ â/ $x      Ž Ê xÿ      È r
    ÿ � ä Š	ÿ t    ä y      '   $ r	       âO G  À    € ê z ÿ `  pPô â� $r;ÿÿ   ÿ Ž â z ÿ ^  pPò â $r<ÿÿ   ÿ Ž â z ÿ b  p ö Ä z ÿ a   Sô ä z ÿ _  Sò ä z ÿ c  0't Ö $ ~  ÿŽ âÿ     â$ª ~  ÿŽ ä $ x  ÿŽ â   àë â$ª x  ÿŽ â ¢   àÇ ä¨ÿ     ä ÿÿ   Äþ Ä    ÿàý ä ¢   ÿàû ä ¢ÿÿ   „þ â $ÿÿ    â  ^  ÿ† ä ª b  ÿˆ â $¢ÿÿ   Ž â  _  � È ª c   â �;     h �©<     d Ay        € ê� ¹z  l      â x8ÿ   
  â ™x?    â $ÿÿ   
Ž â Ey  p    € â $r:ÿÿ   ÿ Ž Ä $
 o  Ž â| 8   pbò â $‚ÿÿ    Ž ä $r9ÿÿ   ÿ Ž Ô G       € ê x
ðÿÿÿÿÀŽ â $rÿÿ   ÿ Ž ä $z6 }  ÿŽ ä$x    
Ž ä 'z y   Ž Æ x ÿ     â $z w  ÿŽ â zÿ z    â $t$ÿ   ÿ Ž â r     ÿ� â $x   ÿ Ž â r66   ÿàñ â $z |  8Ž â x øÿÿÿÿÀŽ â %v" `  $Ž â xÿ      Æ $x7   
Ž ä $z	 v  Ž ä $x   ÿ Ž â xÿ   7  â $z p  Ž ä $rÿÿ   ÿ  â r7   ÿ� â %x4n   ÿŽ â xÿ     Ä r   ÿò� â $x7   ÿ Ž ä $r9ÿÿ   ÿ Ž â xÿ     â $r:ÿÿ   ÿ Ž â x   ÿàÿ â $x8   ÿ Ž â r   ÿ� â %v$ `  $Ž â x	ÿ     Ä r   ÿ� ä x   ÿàÿ äx   ÿàÿ ä xøÿÿÿÿÀŽ ä r	   ÿ� ä x üÿÿÿÿÀŽ â $x!   
Ž â xÿ     Ä xÿ     ä r		   ÿ� ä xøÿÿÿÿÀŽ ä x     èÿ ä r   ÿ� â $x   
Ž â r   ÿ� Ä x	üÿÿÿÿÀŽ ä r   ÿ� ä r   ÿ� ä xøÿÿÿÿÀŽ ä xøÿÿÿÿÀŽ ä x   èÿ â $x   
Ž â xüÿÿÿÿÀŽ â $x   
Ž â xüÿÿÿÿÀŽ Ä x   èÿ ä x   èÿ ä xÿ   	  ä x	ÿ     ä xÿ     â $x   ÿ Ž â xÿ     â $x		   ÿ Ž â xÿ     â $x   ÿ Ž â xÿ     â $x   ÿ Ž â xÿ     â $x   ÿ Ž â xÿ     â $x   ÿ Ž â z `  ÿàñ â $x   ÿ Ž â z `  ÿàó â $x   ÿ Ž â xÿ     Ä x
ÿ     ä z a  ÿä  ä z
 a  ÿäÿ  ä $tÿ   ÿ Ž È %v X  Ž È %v X  Ž â r4   ÿàñ È r4   ÿàó â $x   5  È $x   5Ž  â �y/"   • ¨  �y     • ¨  �y1    • è �yB   • â x2ÿ      Ä z&4 X   àµ äx0ÿ   !  äzD4 X  !à‘ äx.ÿ     äz*4 X  àÙ ä z'5 Y  2d~ Ä r6   àµ ä/ zE5 Y  0$~  äxJÿ     â $t,ÿ$   ÿ Ž â z+5 Y  .¤~ â �y3&`   ‘ " z4 X  àƒ ä r)   ÿd~ â �yH*`   ‘ " xFÿ     Ä z(4 X  àµ â�yDD`   ‘ " z4 X  àÙ ä z5 Y  Jþ  â %v& Z  , Ž â/ z5 Y  
¤~ Æ $x+)$   ÿŽ â z)5 Y  Fd~ â �y=`   ‘ " z,4 X  àµ â$x''   +Ž â x>ÿ     â�yC`   ‘ " z@4 X  à‘ ä z-5 Y  d~ â �yL(`   ‘ " %x*7   &Ž â zA5 Y  >$~  Ä �yK&    • ( �y-,`   ‘ ( �y*   ™ ( �y@@`   ‘ ( �y*(   ™ ( �y*L   ™ (% �yG&$   • ( �y*p   ™ ( �yI&H   • ( �yE&l   • " z(4 X  ÿàñ È z)5 Y  ÿä  Ê �y?(l   • d$t)ÿÿ Ž â z ÿ `  pPð È z ÿ a   Sð ä x88   ÿàÿ ä x* T  /    ÈO x,ÿ   *  ä x11T  B    È� r ÿ1   ÿ3Ž ä x,,ÿÀŽ ä r/ÿ      ä x+,€€€€ÿüŽ ä x /ÿÀŽ Ê $x'+    
Ž â/ x ,/øŽ È r( '   ,`Ž â $x /   ÿ Ž âx&*€€€€)ìŽ ä x((˜º  ÿ    ä x+ ÿÀŽ ä x)ÿ   *  ä x (€€€€'Ž â $x'&   +
Ž â x1,˜º  ÿ    â $x&/   ÿ Ž â x))ÿÀŽ Ä x1(1`Ž ä x(+*øŽ äx,)€€€€ÿüŽ ä x+&ÿÀŽ ä x&ÿ   *  äxA'*xŽ â $x,,   +
Ž â r 1    ÿüŽ Ä x&&ÿÀŽ ä x1ÿ   /  ä r(A(   ÿÀŽ ä x**ÿÀŽ ä r/)+   ,XŽ ä xB&€€€€ÿüŽ Ä x+1ÿÀŽ ä r3ÿ   3  äx**˜º  ÿ    ä x((˜º  ÿ    ä rDÿ   D  â $x33   ÿ Ž â x1)˜º  ÿ    â $x)B   +
Ž â r=ÿ   =  Ä x//˜º  ÿ    ä x'(€€€€'Ž äx*(*`Ž ä xDD   ÿÀŽ ä rHÿ   H  â $x==   ÿ Ž â rCÿ   C  Ä x,/€€€€,Ž äx//1`Ž ä r+&+   )XŽ ä r**'   ÿüŽ ä xD30   DâŽ ä rLÿ   L  Ä xHH   ÿÀŽ â $xCC   ÿ Ž â r-ÿ   -  â &r'*   ÿ   â x(+˜º  ÿ    ä r//,   ÿüŽ ä xLL   ÿÀŽ ä xDDàÿÿÿÿàÿ Ä xH=0   HâŽ ä x+&˜º  ÿ    ä r@ÿ	   @  â $x--   ÿ Ž â xLC0   LâŽ â $r'D'   ÿŽ â xHHàÿÿÿÿàÿ â &r//   ÿ   â x)(€€€€)Ž Ä x&(+`Ž ä x@@   ÿÀŽ â &r+    ÿ   â xLLàÿÿÿÿàÿ â $r/H/   ÿŽ â r)&)   ÿüŽ â 0r ÿK    A   â x@-0   @âŽ ä Er' '       â $r+L+   ÿŽ â x@@àÿÿÿÿàÿ â 0rDÿG    A   â Er/ /       â &r))   ÿ   â #r' '   ÿ   â 0rBÿI    A   â Er+ +       â $r)@)   ÿŽ â #r'D/   '   â 0r=ÿE    A   Æ #r@B+   '   â ErA )       â G‰  0    € ê r&$4   ÿàó är("4   ÿàñ ä z,4 `   àÇ â$x'%   5Ž  â z*4 `  !à£ â $x)#   5  Æ �yC&     ¨  �yH&    ¢  z+5 a  0Dþ  äz-5 a  2„þ â �yE("    è �yL(     è �yG*`    â z04 `  àµ Ä z24 `  à‘ â�yI,`    â z15 a  Jd~ är&4   ÿàõ äz35 a  .$~  äz.4 `  à‘ â$x'5    âr*4   ÿàõ â�y10`    " z/5 a  F$~  Ä z(4 `  à‘ â/ �y22`    " $x+5    Æ �y'&`    " z)5 a  >$~  Æ �y..`    ( �y+*`    ( �y((`    " z,4 `  ÿàñ È z-5 a  ÿä  Ê �y,,l    " $t3ÿÿ Ž â/ xCCT  H    ÈO r0ÿC   ÿ3Ž ä xLLT  E    ä� r/ÿ   0  ä x0ÿ   L  Æ $x)/   ÿ Ž â x*0ÿÀŽ ä x0L€€€€3ìŽ ä x3)ÿÀŽ ä x>ÿ   L  ä rGÿ   G  â $x)0   3
Ž â xF*€€€€ÿüŽ â $xC/   ÿ Ž âx-/ÿÀŽ Ä x03LøŽ äx>>ÿÀŽ ä x&G   ÿÀŽ ä x3)LxŽ â$x-F   -
Ž â xGÿ   L  ä xF>€€€€ÿüŽ Ä xECÿÀŽ ä r030   ÿÀŽ ä xC*/øŽ äx3GÿÀŽ ä xHÿ   /  â $x/F   E
Ž â rGC-   *`Ž Ä xF3€€€€ÿüŽ ä xCHÿÀŽ ä xHG˜º  ÿ    ä xG*˜º  ÿ    â $x*F   C
Ž â xLLÿÀŽ ä rIÿ   I  Ä xFH€€€€-Ž äxGHG`Ž ä rC3C   *XŽ äxH3˜º  ÿ    ä x00˜º  ÿ    ä rE>E   /XŽ Ä x3L˜º  ÿ    â $xII   ÿ Ž â r1ÿ   1  äx)0€€€€)Ž ä x>>˜º  ÿ    ä xEE˜º  ÿ    ä x003`Ž Ä r2ÿ   2  ä r'ÿ   '  â $x11   ÿ Ž â r.ÿ   .  ä x&I0   &âŽ ä x/E€€€€/Ž äx>E>`Ž Ä r00)   ÿüŽ ä x22   ÿÀŽ â $x''   ÿ Ž â r+ÿ   +  ä x..   ÿÀŽ â &r0   ÿ   â xCC˜º  ÿ    ä r/>/   ÿüŽ Ä x210   2âŽ ä r(ÿ	   (  ä x&&àÿÿÿÿàÿ â $x++   ÿ Ž â x.'0   .âŽ â &r/   ÿ   â rFGF   ÿüŽ â $r&   ÿŽ â x*C€€€€*Ž Ä x-CH`Ž ä x((   ÿÀŽ ä x'2àÿÿÿÿàÿ â &rF   ÿ   â r--*   ÿüŽ ä x(+0   (âŽ â $r'   ÿŽ â x..àÿÿÿÿàÿ Ä Er        â &r--   ÿ   â x((àÿÿÿÿàÿ â $r.   ÿŽ â Er        â #r    ÿ   ä $r((-   ÿŽ â Er        â #rD      Æ Er( (       â #rBB      â 0rÿ,    A   Æ #r(=(   B   È #r:(   :   â | 8   pbð â 0r ÿ?    A   âx44p  ÿàó â #r@=A   @   â x@   ÿàÿ Æ #r9 @   9   â $r5ÿÿ   5Ž  Ê D	       À â Gy  @íÿÿÿÿƒ ê Ay        € ê y       "   " Ey        € â r  ÿ   pRð är  ÿ   pRò Ö G‰  `     € ê y      !   " z ÿ `  pPð È z ÿ a   Sð â $x     Ž Ê ˆs 9€ÿÿ H   î ˆ :   H   ä Ay        € ê {          ì M        € ê y      !   b z ÿ `  pPð È z ÿ a   Sð â „y      H   X. „‰ €   H   ¨ „‰    H   " !r9 9       È/ !‚9       ÈO !‚       â G‰        € ê „y  €  H   ( „y €   H   h „y    H   ¨ „y    H   è „y €  H   " !r  :       â !r9       Æ/ !r         âO !r       Æ� !r:        ä‰      " z ÿ `  pPð ä r ÿ   pRò ä z ÿ a   Sð â !r       Ê ‰       $ !r         Ê ‰  €    $ !r        Ê ‰ @    $ !r       Ê ‰      $ !r       â G‰        € ê ‰ :      $ !r  :       Ê ‰       $ !r        Ê ‰ €    $ !r       Ê ‰ @    $ !r       Ê ‰      $ !r:       ä M        € ê !r;;       âG‰  `    € ê $t ÿ   ÿ Ž â !r<:       È z   h  pð Ú G	  `    € ê $v ÿ h  ÿ Ž Ê x     pRð Ú G‰  �     € ê x     pRð Ú G	       € ê  x ;ª¸¿  A  Ì s          $ !t   €?    Ì s        $  r     A  È  r;;     A  â Gy  `    € ê  x '7=  A  â x*BL?  A  Æ #t   €?    È  r      A  â  x    ?  A  Ê s     $   $ !t  €?    È  r       A  È  r;;      A  â Gy  À    € ê $v ÿ h  ÿ Ž Ê x     pRð Ú G‰  à     € ê x     pRð Ú G	        € ê 	z  i    � ä 	z;; i    � Æ  x ;ª¸¿  A  â 	z;; i €  � Ê s        $ !t  €?    Ì s        $  r       A  È  r;;      A  â Gy  Ð     € ê  r;;     A  â Gy  °     € ê 	x   à@  � ä 	x;;  à@  � Æ  x #ÛÙ?  A  â 	x;;  àÀ  � Æ  x;ª¸¿  A  ä !t;;  €?    È s        $ !t  €?    Ì s        $  r       A  È  r; ;     A  ä y       %   ( y      &   ( y      '   b $z  x   Ž ä $tÿ   ÿ Ž Ä $z  ~   Ž È/ %v  j   Ž Ê †y ;    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y       &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž ä $r	ÿÿ     Ž È %	 \   Ž Ê �	    ™ b ¹z  l      â Ey   )    € â ™x?    â y      "   b $rÿÿ   ÿ Ž ä $r+ÿÿ   ÿ Ž â y      !   ¢ $ÿÿ   ÿ Ž ä $ÿÿ   	 Ž È '
	 m   Ž È $‚ÿÿ   ÿ Ž â 
ÿ n    Ê $ÿÿ   
Ž â r ÿ   pRò Æ/ $
	 o  	Ž â r ÿ   pRô â$x    Ž äO $‚ÿÿ   	 Ž Æ xÿ     â 'Š	 s   Ž Æ r   ÿ � â $€   Ž Æ x,ÿ     È | ,   pbö ä Šÿ t    ä x      ÿ  Ò G9  `'    € ê xðÿÿÿÿÀŽ â y      '   " $r
ÿÿ   ÿ Ž ä $z	 w  ÿŽ â y      %   b $x   
Ž ä $v6ÿ p  ÿ Ž ä $x9   ÿ Ž â xÿ     È r   ÿ� È xøÿÿÿÿÀŽ ä xÿ     Æ $x   
Ž ä 'z
 y  
 Ž Æ xÿ     â $x   ÿ Ž â/ z
ÿ z  
  â $z }  ÿŽ â r   ÿ� â $z p  ,Ž â xÿ     â $x   ÿ Ž â r   ÿò� â $z
 |  Ž â r
   ÿàñ â $x;   ÿ Ž â xÿ     â $z v  Ž âx	   ÿàÿ â $x,   ÿ Ž â r   ÿ� â%x4n   ÿŽ â r   ÿ� Ä xøÿÿÿÿÀŽ â $rÿÿ   ÿ  â xüÿÿÿÿÀŽ â %x89   4Ž â xÿ   	  ä x   ÿàÿ â$x   
Ž âx   èÿ â %x:;   4Ž â r	   ÿ� Ä r	   ÿ� ä x   ÿàÿ ä xÿ     ä xüÿÿÿÿÀŽ ä xÿ     ä r   ÿ� Ä x-	   èÿ ä r   ÿ� ä xüÿÿÿÿÀŽ ä r   ÿ� ä xøÿÿÿÿÀŽ ä xøÿÿÿÿÀŽ Ä x)   èÿ â $x1   
Ž â r   ÿ� â $x#	   
Ž â x8"   ÿàñ â $v	ÿ X  ÿ Ž â xüÿÿÿÿÀŽ ä x     â $rÿÿ   9  â xøÿÿÿÿÀŽ Ä xÿ     ä xÿ     â $x/   
Ž â r4   ÿàñ â$tÿn   ÿ Ž â r4   ÿà÷ â $rÿÿ   ÿ Ž â x%   èÿ â $vÿ Y  ÿ Ž â x&ÿ     â %v66  € Ž â r5   ÿ  Ä x'ÿ     â $x   ÿ Ž â r5   ÿ� â $x   ÿ Ž â r#4   ÿàù â$x&&   ÿ Ž â r%4   ÿàû â $x''   ÿ Ž â xÿ     ä xÿ     Ä xÿ     â $x   ÿ Ž â xÿ     â $x   ÿ Ž â r"#5   ÿ â$x   ÿ Ž â r$%5   ÿ� â $x   ÿ Ž â r 14   ÿà÷ är!)4   ÿàñ Ä r#/4   ÿàù är%-4   ÿàû ä xÿ     ä r(15   ÿ� är))5   ÿ  är*/5   ÿ Ä r-5   ÿ� ä r<:	   ÿà÷ är28	   ÿàñ Æ $x=;   Ž ä$x39     Æ �y/<    • ¨  �yT<   • ¨  �yg2"   • è �yX2    • â %zt p  <Ž È %z> p  2Ž â �y\t    • h	 �yQt   • h	 �yS>"   • h �y->    • b r0	   ÿàû â$twÿ$   ÿ Ž â rF	   ÿàñ Ä rv
   àÇ â $x1   Ž â rJ	%   ÿàû â $xG     â rH	   ÿàý â %vvv Z  w Ž â r2   ÿ„þ â/ �y^0`   ‘ b r<	#   ÿàñ â $xK   Ž ä$xI   " â �yWF`   ‘ b r>	!   ÿàý â $x32$   ÿŽ Ä �ybJ`   ‘ b rjv   ÿ† Æ �y.H`   ‘ b %zD p  0Ž È %zJ p  JŽ â�yhD`   ‘ f %zH p  HŽ â/ �y1J`   ‘ f $xww   3Ž â �yRH`   ‘ b $x?   ) Ä $x=   *  â �yiv    • b $xk   wŽ â rn	   ÿàñ â %zF p  FŽ â �yU>`   ‘ h �yN<`   ‘ b $xo     Æ �yYj   ™ h	 �ycF`   ‘ b	 %z> p  >Ž Æ �yVj(   ™ b %z< p  <Ž Æ/ �y2>`   ‘ b %zn p  nŽ Æ �yM<`   ‘ d rZ6n   ÿàù ä �yOjL   ™ b %zt p  tŽ Æ�y`v$   • b	 $x[7   o Æ �yPvH   • b	 rB	   ÿà÷ Æ �y0Zþÿÿ • h �yeZüÿÿ • h �yft    • b r@	    ÿàñ Æ �yLjp   ™ h �y[t   • b $xC   $Ž ä %zF p  FŽ â �yvvl   • h	�y3B`   ‘ b $xA   (  Ä %zD p  DŽ â�yaF`   ‘ h �y_@`   ‘ b %zB p  BŽ Æ �y]D`   ‘ h �yZB`   ‘ b %z@ p  @Ž Ê/ �yd@`   ‘ b %zJ p  JŽ È %zH p  HŽ â�ykJ`   ‘ b xrnþÿÿÿÿàñ ä %z> p  >Ž â�yjH`   ‘ f %z< p  <Ž â �ym>`   ‘ b xsoÿÿÿÿÿä  Æ �yl<`   ‘ b rp4	   ÿàñ â %zt p  tŽ È %zr p  rŽ â �ynt    • f $xq5     â �yot   • b rr6r   ÿàñ Ê $xs7   s  ä %zt p  pŽ ä�yppl   • h �yqr    • h �yrrþÿÿ • b %zF p  FŽ È %zD p  DŽ ä�yFF`   ‘ d %zJ p  JŽ ä�yED`   ‘ d %zB p  BŽ ä�yJJ`   ‘ d %zH p  HŽ È %z@ p  @Ž ä/�yIH`   ‘ d %z> p  >Ž ä�ys@`   ‘ d %zz p  tŽ ä�yttl   • d %z< p  <Ž Ä �y>>`   ‘ d %z@ p  @Ž ä�y=<`   ‘ h �yuB`   ‘ h# �yA@`   ‘ h �ywzl   • b	%zB p  BŽ Ì/ �yBB`   ‘ b %zx p  zŽ Ì �yxxl   • b x//T  T    ÈO r/ÿ/   ÿ3Ž ä xXXT  g    È� xTÿ   X  ärDÿ   /  ä xKTÿÀŽ ä x<DÿÀŽ ä xGK€€€€ÿüŽ ä x/ÿ   X  Ä xHÿ   X  â $x<G   <
Ž â x?/ÿÀŽ â $xTD   ÿ Ž â x/HÿÀŽ â $tgÿÿ Ž â xHKDøŽ â$x@D   ÿ Ž â xCÿ   D  ä/ rHH<   K`Ž Ä xy?€€€€ÿüŽ ä xTTÿÀŽ ä xGX€€€€gìŽ ä xD@ÿÀŽ ä xg/€€€€ÿüŽ ä x@CÿÀŽ Ä x{H˜º  ÿ    â$xHy   T
Ž â xKK˜º  ÿ    â $xCG   D
Ž â xGDXøŽ â $xDg   @
Ž â x<{€€€€<Ž äxK{K`Ž ä rT?T   HXŽ Ä xQ\T  Q    är<K<   ÿüŽ ä rg/@   DXŽ ä xKT˜º  ÿ    ä x@?˜º  ÿ    ä rQÿQ   ÿ3Ž Ä xzCXxŽ ä xT/˜º  ÿ    ä x/K€€€€HŽ ä xXXÿÀŽ ä xgg˜º  ÿ    ä x@K@`Ž Ä rHÿ   Q  ä rGzG   ÿÀŽ ä xKX˜º  ÿ    â $tXÿÿ Ž â xDg€€€€DŽ äx?gT`Ž ä r/@/   ÿüŽ â $x@H   ÿ Ž â xS-T  S    Ä xTG˜º  ÿ    ä r-?D   ÿüŽ ä x?S€€€€XìŽ ä xGT€€€€CŽ ä xX@ÿÀŽ ä xCÿ   S  Ä xTTK`Ž â $xD?   X
Ž â xCCÿÀŽ ä xKÿ   S  äx\C€€€€ÿüŽ ä xQHÿÀŽ ä r@TG   ÿüŽ â $xGH   ÿ Ž â x?XSøŽ Ä xXDSxŽ ä xKKÿÀŽ â $xT\   Q
Ž â r?X?   ÿÀŽ ä xQK€€€€ÿüŽ ä xXGÿÀŽ ä x\ÿ   S  Ä xgCHøŽ äxzÿ   H  â $xHQ   X
Ž â xG\ÿÀŽ ä rygT   C`Ž ä x\G€€€€ÿüŽ ä xQzÿÀŽ Ä rgKX   HXŽ ä xyy˜º  ÿ    ä xzC˜º  ÿ    â $x\\   Q
Ž â xgg˜º  ÿ    ä xTy€€€€TŽ äxyyz`Ž Ä xSSÿÀŽ ä xCg€€€€HŽ ä rHyT   ÿüŽ ä rWÿ   W  ä rQGQ   \XŽ äxTG˜º  ÿ    Ä rbÿ   b  ä xG?˜º  ÿ    ä xzS˜º  ÿ    ä r^ÿ   ^  ä r.ÿ   .  â $xWW   ÿ Ž â xXK˜º  ÿ    â $xbb   ÿ Ž â xDG€€€€DŽ Ä xKGz`Ž ä x^^   ÿÀŽ ä xG.   ÿÀŽ ä r1ÿ   1  ä rRÿ   R  ä r.KD   ÿüŽ Ä x^W0   ^âŽ ä xDb0   GâŽ â $xG1   ÿ Ž â rUÿ   U  ä xXgX`Ž ä xRR   ÿÀŽ â &r1@Y   ÿ   â x^^àÿÿÿÿàÿ Ä rNÿ   N  â $xUU   ÿ Ž â rCXC   ÿüŽ ä rcÿ   c  ä xRG0   RâŽ â $r1^1   ÿŽ â xNN   ÿÀŽ â &r//V   ÿ   â xDDàÿÿÿÿàÿ Ä rhÿ   h  â $xcc   ÿ Ž â xRRàÿÿÿÿàÿ â &rCCV   ÿ   â xNU0   NâŽ â &rG.Y   ÿ   â xhh   ÿÀŽ â $r/D/   ÿŽ â Er. 1       â 0riÿi    A   â r2ÿ   2  â $rRRC   ÿŽ â xNNàÿÿÿÿàÿ â &rC<O   ÿ   â xhc0   hâŽ Ä rMÿ   M  ä Er1 /       â #r/i.   ÿ   â $x.2   ÿ Ž â xhhàÿÿÿÿàÿ â $rCNC   ÿŽ â xMM   ÿÀŽ â 0r`ÿ`    A   ä $rGhG   ÿŽ â xM.0   MâŽ â 0rPÿP    A   â Er. C       â #r/`1   /   â x[fT  [    Ä Er2 G       â #r.P.   /   â r/ÿ[   ÿ3Ž ä ErR R       â #r1i2   ÿ   â r/ÿ   /  â &rHHO   ÿ   â xMMàÿÿÿÿàÿ â #rC`R   1   â r3ÿ'   3  â $tGÿÿ Ž â x0eT  0    â $x1/   ÿ Ž Ä $rHMH   ÿŽ â x20€€€€GìŽ â $x<3   ÿ Ž â x31ÿÀŽ ä ErH H       ä r_ÿ&   _  â $xG2   3
Ž â x@ÿ   0  â&r2-L   ÿ   â x__   ÿÀŽ â #r-PH   C   â xDÿ   0  â $xC/   ÿ Ž â x@@ÿÀŽ Ä rZÿ'   Z  ä x_<0   _âŽ ä xHDÿÀŽ ä rdÿ&   d  ä xCCÿÀŽ ä xD@€€€€ÿüŽ Ä x330øŽ äx<G0xŽ â $xKZ   ÿ Ž â x1_àÿÿÿÿàÿ â $xDD   C
Ž â xdd   ÿÀŽ ä r<<3   ÿÀŽ â $r112   ÿŽ â x30ÿÀŽ Ä x2K0   dâŽ ä xK/ÿÀŽ ä xNH€€€€ÿüŽ ä x<<˜º  ÿ    ä x33˜º  ÿ    ä rC@C   DXŽ Ä x0ÿ   0  â $xKN   K
Ž â xG<€€€€GŽ äraÿ   a  ä x<<3`Ž ä xCC˜º  ÿ    ä r]ÿ   ]  Ä x3H/øŽ äx@@˜º  ÿ    ä x/ÿ   /  ä x00ÿÀŽ â $xRa   ÿ Ž â rkÿ   k  ä xDC€€€€DŽ Ä x]]   ÿÀŽ ä r33K   H`Ž ä xCC@`Ž ä rjÿ   j  ä x//ÿÀŽ ä x@0€€€€ÿüŽ â $xkk   ÿ Ž â rmÿ   m  Ä xN3˜º  ÿ    ä x]R0   ]âŽ ä r<<G   ÿüŽ ä xjj   ÿÀŽ â $x@@   /
Ž â x3H˜º  ÿ    ä rlÿ   l  â $xmm   ÿ Ž â xKN€€€€KŽ â &r<<Y   ÿ   â rCCD   ÿüŽ Ä x]]àÿÿÿÿàÿ ä xjk0   jâŽ ä xNN3`Ž ä xll   ÿÀŽ ä r/0/   @XŽ â $r<]<   ÿŽ â xQQ˜º  ÿ    â &rCCV   ÿ   â xjjàÿÿÿÿàÿ Ä xnnT  o    ä rNNK   ÿüŽ ä xlm0   lâŽ ä x3/˜º  ÿ    â $rCjC   ÿŽ â x?Q€€€€\Ž äxTQT`Ž Ä x00˜º  ÿ    ä rGÿn   ÿ3Ž â &r/NO   ÿ   â xllàÿÿÿÿàÿ ä Er< <       ä x@3€€€€@Ž ä r?T?   ÿüŽ Ä x330`Ž ä rGÿ   G  â $rll/   ÿŽ â ErC C       â #r/i<   ÿ   â x02àÿÿÿÿàÿ â &r??L   ÿ   â r33@   ÿüŽ â $x2G   ÿ Ž â xqrT  q    â $t@ÿÿ Ž â #r/`C   /   â $r00?   ÿŽ â xCÿ   q  Ä x<2ÿÀŽ ä x?q€€€€@ìŽ ä xHCÿÀŽ ä xCGÿÀŽ â $x@?   <
Ž â xDH€€€€ÿüŽ ä x<<qøŽ Ä x?@qxŽ â $xMD   C
Ž â xCÿ   q  â $xDG   ÿ Ž â r??<   ÿÀŽ ä x<qÿÀŽ ä xCCÿÀŽ ä x??˜º  ÿ    Ä x<<˜º  ÿ    ä xDDÿÀŽ ä xqÿ   q  ä xKC€€€€ÿüŽ ä x@?€€€€@Ž äx??<`Ž â $xKK   D
Ž â x<ÿ   G  Ä xqqÿÀŽ ä r@?@   ÿüŽ ä xN<ÿÀŽ ä x?q€€€€ÿüŽ ä rFÿ   F  ä xGHGøŽ Ä rDCD   KXŽ ä rEÿ   E  â $xQ?   N
Ž â rJÿ   J  â $xFF   ÿ Ž â rGGM   H`Ž ä xDD˜º  ÿ    ä xCC˜º  ÿ    Ä xEE   ÿÀŽ ä rIÿ   I  â $xJJ   ÿ Ž â r>ÿ   >  ä x<G˜º  ÿ    ä rNqN   QXŽ ä xKD€€€€KŽ Ä x?H˜º  ÿ    ä xDDC`Ž ä xEF0   EâŽ ä xII   ÿÀŽ ä r=ÿ   =  â $x>>   ÿ Ž â ruÿ'   u  Ä rBÿ'   B  ä xM<€€€€MŽ ä xNN˜º  ÿ    â &r@@Y   ÿ   â rsÿ&   s  ä x<<?`Ž ä xqq˜º  ÿ    Ä rKDK   ÿüŽ ä xEEàÿÿÿÿàÿ ä xIJ0   IâŽ ä rAÿ&   A  ä x==   ÿÀŽ â $xuu   ÿ Ž â xQN€€€€QŽ â $xBB   ÿ Ž â x2s   ÿÀŽ â &rVKV   ÿ   â xNNq`Ž â $r@E@   ÿŽ â r<<M   ÿüŽ Ä xIIàÿÿÿÿàÿ ä xAA   ÿÀŽ ä x=>0   =âŽ â &r<<O   ÿ   â x2u0   2âŽ â $rIIV   ÿŽ â rQNQ   ÿüŽ ä xAB0   AâŽ Ä x==àÿÿÿÿàÿ ä Er@ @       â &r33L   ÿ   âx22àÿÿÿÿàÿ â &rLQL   ÿ   â xAAàÿÿÿÿàÿ â $r<=<   ÿŽ â x,,   ÿàÿ â #rii@   ÿ   â ErI I       â $r223   ÿŽ â | ,   pbð â $rAAL   ÿŽ â Erl l       â #rI`I   i   â Er< <       â 0r3ÿv    A   â Er1 1       Ä Er0 0       â #r/Pl   /   âEr2 2       â #r<P<   I   â ErA A       â #r.31   .   â#r-30   -   â x		p  ÿà÷ â 0rpÿp    A   â #r/32   /   â0rtÿt    A   â #r<3A   <   â 0r0ÿw    A   Ä 0r1ÿx    A   â #r+p.   +   â #rt-      â #r0/      â x@   ÿàÿ â #r1<      â $rÿÿ   Ž â D	       À â Gy   ßÿÿÿÿƒ ê Ay        € ê ˆ + þÿ H   è ˆ €þÿ H   è ˆ  ÿÿ H   è ˆ €ÿÿ H   è {          ì M)        € ê „y     H   ( „y
 €   H   h „y    H   ¨ „y    H   è „y	 €  H   ( „y    H   ( „y    H   ( „y €  H   ( „y €  H   („y    H   ( „y €  H   " !r+       Æ „y €  H   " !r

       â/ !r       âO !r       ä� !r	
	       â!r       â !r       â !r       È ‰      b !r		       â !r       È ‰	      ¢ !r       Æ ‰      â !r       Ê ‰      " !r       Ê/ ‰      b !r	       âO !r
       È� ‰      ¨ ‰
      â !r       Ê ‰      " !r       â/ !r       ÈO ‰ €    b !r
       Æ� ‰ €    ¨ ‰ €    â !r       Æ y      '   ( ‰ €    " !r       â/ !r       ÈO ‰ @    b !r       Æ� ‰ @    ¨ ‰	 @    â !r       Æy      %   ( ‰
 @    " !r       â/ !r       ÈO ‰      b !r		       Æ� ‰      ¢ $x   ÿ Ž Æ‰	      â !r
       â x   ÿàÿ â$z   x  Ž â x
   ÿàÿ ä‰      " x   ÿàÿ â $z  ~   Ž â z  r  p`ð ä z  r  p`ò Ä z 
 r  p`ô ä z  r  p`ö â $tÿ   ÿ Ž â r ÿ   pVp  â!r       â/ x    pVò  â!r       âO x    pVt äx    pVö â %v  j   Ž â !r		       Ê� †‰     â !r       Æ †™    è †© 	   â M9        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      &   " z ÿ \  pPò â ¹z  F   
   Æ z ÿ ]  Sò Ú $ÿ   ÿ Ž È % \   Ž Ê �	    ™ ¢  $rÿÿ   ÿ Ž â r ÿ       â ¹z  l      â y›      %   b '  m   Ž â™x?    â Ey  @    € â y      "   â x˜      ÿ  â  ÿ n     â 'š s   Ž â yœ      !   & $ÿÿ    
Ž â y      '   & $  o  Ž Ä $’ ÿÿ    Ž È $zš  w  ÿŽ ä $x››   ÿ Ž â/ r ÿ   pRð Æ� $x›   œŽ âx œ   pDp  â $x     œŽ ä $z }  ÿŽ âz  r  pdp  â 'z y   Ž Æ ršš   ÿàõ ä xÿ      È r    ÿ � ä š	ÿ t    È r	     € âO G	  °    € ê z ÿ `  pPò â x˜      ÿ  â z ÿ ^  pPð ä z ÿ b  p ö ä z ÿ a  Sò ä z ÿ _   Sð ä z ÿ c  0'ò Ö $
 ~  ÿŽ âÿ   ›  â$š ~  ÿŽ ä $
 x  ÿŽ â›   œàë â$š
 x  ÿŽ â ’›   œàÇ ä ˜ÿ   ›  ä ÿ   ÿÄþ Ä    ÿàý ä ’
   ÿàû ä ’ÿ   ÿ„þ â $ÿÿ    â 
 ^  ÿ† ä š b  ÿˆ â $’
ÿÿ   Ž â 
 _  � È š c  
 â �	˜     h �™™     d Ay        € ê xŠÿ     â Ey  L    € â x–      ÿ  â x”      ÿ  â | Š   pbð â x’      ÿ  â $r�ÿÿ   ÿ Ž Ä $rŽÿÿ   ÿ Ž ä $r‘ÿÿ   ÿ Î G	  €K    € ê xðÿÿÿÿÀŽ â $z› p  ŠŽ â z ÿ `  pPð â $v�ÿ `  ÿ Ž â z�  a      â $x     
Ž â x–      ÿ  â $v‹ÿ X  ÿ Ž â x”      ÿ  â $xE    ÿ Ž â xÿ      â $v‰ÿ Y  ÿ Ž â x’      ÿ  â $r�ÿÿ   ÿ Ž â r    ÿ� â $rŽÿÿ   ÿ Ž â zÿ z    â $xˆŠ   ÿ Ž â xøÿÿÿÿÀŽ Ä x‡ÿ     â $z |  ŠŽ ârŒ   €ÿàÿ â$z |  Ž â x›   ÿàÿ â $z	 v  Ž âxÿ   Œ  â $z	 v  Ž â x	›   ÿàÿ â $z5 p  Ž ârŒ   ÿ� â $xKŒ   ÿ Ž â x‡‡   ÿ   â $z;	 p  Ž â r   ÿò� Ä x›   ÿàÿ ä xÿ     ä xN   ÿàÿ â $z7 p  Ž â r\   ÿ� â%xn   ÿŽ â r]   ÿ� ä xQ\øÿÿÿÿÀŽ â %x7n   ÿŽ â xS]üÿÿÿÿÀŽ Ä xÿ   N  â $xP   Q
Ž âxV   ÿàÿ â%x@K   Ž â xR   Sèÿ äx   ÿàÿ â %x<E   Ž â r^N   ÿ� är_N   ÿ� â %xBE   Ž â xÿ   V  Ä xÿ     â %xHK   Ž â r`V   ÿ� ä rb   ÿ� ä raV   ÿ� â $tÿn   ÿ Ž â rc   ÿ� â %x5n   ÿŽ â x„@   ÿàó Ä xU^øÿÿÿÿÀŽ ä xM`øÿÿÿÿÀŽ â $rƒÿÿ   AŽ  â xYbøÿÿÿÿÀŽ â $xTN   U
Ž â xO_üÿÿÿÿÀŽ â $xLV   M
Ž â xWaüÿÿÿÿÀŽ â $xX   Y
Ž â x[cüÿÿÿÿÀŽ â %x>E   Ž â xQÿ   P  Ä r}P   ÿàó â %xFK   Ž â xNN   Oèÿ ä xVV   Wèÿ â $xy   QŽ  â xZ   [èÿ â %x;n   PŽ âxSÿ   R  ä x[ÿ   Z  â $vÿ p  ÿ Ž â xYÿ   X  â %x;n   RŽ â xWÿ   V  Ä xMÿ   L  â %x;n   ZŽ âxOÿ   N  ä xUÿ   T  â %x ;n   XŽ âr|R   ÿàõ ärwT   ÿàó â %x&;n   VŽ â x‚ÿ   \  Ä x�ÿ   ]  â %x,;n   LŽ âx€ÿ   ^  ä xÿ   _  â %x2;n   NŽ âx~ÿ   `  ä xzÿ   b  â %x8;n   TŽ â xpÿ   c  Ä rtX   ÿàù â%x:;n   ÿŽ â rsV   ÿàû äx…ÿ   Œ  â $xx   S ârvN   ÿàõ â$xr   UŽ  â rqL   ÿàó â %x
5n   RŽ â x{ÿ   a  Ä ruZ   ÿà÷ â %x7n   RŽ â z ÿ a   Sð ä x†ÿ   ‡  â %x5n   PŽ â x…Œ   …  ä x{{   ÿ   â %x7n   PŽ â rn[   ÿäÿ Æ %x5n   ZŽ È %x7n   ZŽ È %x5n   XŽ È %x7n   XŽ È %x"5n   VŽ È %x$7n   VŽ È %x(5n   LŽ È %x*7n   LŽ È %x.5n   NŽ È %x07n   NŽ È %v  € Ž È %x45n   TŽ È %x67n   TŽ È %xDE   :Ž È %xJK   :Ž È $x‚‚   ÿ Ž ä $x��   ÿ Ž ä $x€€   ÿ Ž ä $x   ÿ Ž ä $x~~   ÿ Ž ä $xzz   ÿ Ž Ä $xpp   ÿ Ž ä $xo   O ä$xm   Y ä$xl   WŽ ä$xk   MŽ  ä rL@‹   ÿàó är\<‹   ÿàõ Æ $xMA   ‰Ž  ä$x]=   ‰ Æ �y L    • ¨  �y¥L   • ¨  �y¤\"   • è �y£\    • â rR‹|   ÿàù â $taÿ$   ÿ Ž â r`‡š   ˆà£ Ä rP‹}   ÿà÷ ä rS‰x   ÿä ärL†‘   ÿDþ  â $xQ‰   yŽ â rV‹v   ÿàù â%v`` Z  a Ž â rT‹w   ÿà÷ â �y¦R`   ‘ $ $xML$   ÿŽ Ä $xW‰   o â�y§P`   ‘ " $xU‰   rŽ â rbŒ`   ÿ‚ â $xaa   MŽ â �y¨V`   ‘ ( �y©T`   ‘ " rc…a   ÿäÿ  Ä r^‹„   ÿàó â$v±ÿ p  ÿ Ž â rZ‹s   ÿàý â �yjb   ™ $ $x_‰   ƒŽ  â rX‹q   ÿàû â �yib(   ™ $ %x^±n   ^Ž ä �y¬`    • $ $x[‰   l Ä $xY‰   kŽ â �y�^þÿÿ • " %x\±n   \Ž Æ/ �yž^    • ( �yªZ`   ‘ ( �y«X`   ‘ ( �y®`$   • ( �yŸ\"   • ( �y \    • ( �yhbL   ™ " %xR±n   RŽ â rN‹u   ÿà÷ Ä rL‹t   ÿàõ â %xP±n   PŽ â �y¢R`   ‘ & $xO‰   nŽ â �y¡P`   ‘ " %xT±n   TŽ Æ �y¯N`   ‘ " %xV±n   VŽ Æ �yµT`   ‘ " $xM‰   m ä %xZ±n   ZŽ â�y²V`   ‘ & %xX±n   XŽ â�y­L`   ‘ & %xN±n   NŽ â�y°Z`   ‘ ( �y³X`   ‘ " %xL±n   LŽ Æ/ �y¹N`   ‘ ( �y·L`   ‘ " %x^±n   ^Ž Æ �y´bp   ™ " rd\   ÿàó Æ �y»^þÿÿ • ( �yº^    • ( �y¶`H   • " $xe   ]Ž  Æ �y¼`l   • ( �y¸d    • ( �y½d   • " %xR±n   RŽ È %xP±n   PŽ â�y¾R`   ‘ & %xV±n   VŽ â�y¿P`   ‘ & %xT±n   TŽ â �yÀV`   ‘ " x\\    ÿàó Æ �yÁT`   ‘ " %xb±n   ^Ž È $r]ÿÿ   ]Ž  â r`‹   ÿàó â �yÄbþÿÿ • $ %x\±n   \Ž ä �yÅb    • " ra‰   ÿäÿ  ä r^\   ÿàó â %xN±n   NŽ È/ $x_   ]Ž  ä %xd±n   `Ž ä�y``l   • ( �yÇ^    • ( �yÈ^þÿÿ • " %xL±n   LŽ Æ �yaN`   ‘ (# �ycL`   ‘ " %xZ±n   ZŽ Æ �yÆdl   • " %xP±n   PŽ Æ �yÂZ`   ‘ " %xR±n   RŽ Æ �yQP`   ‘ " %xX±n   XŽ Æ �yRR`   ‘ " %xV±n   VŽ Æ �yÃX`   ‘ " %xT±n   TŽ Æ �yVV`   ‘ " %xZ±n   ZŽ Æ �yUT`   ‘ " %x^±n   LŽ Æ �yZZ`   ‘ " %xX±n   XŽ Æ/ �y_^`   ‘ " %xL±n   NŽ Æ �yYX`   ‘ ( �yML`   ‘ " %xf±n   dŽ Ê �ybfl   • " %x\±n   fŽ Ì �y\\l   • " $tTÿÿ Ž â x  T  ¥    ÈO r ÿ    ÿ3Ž ä x££T  ¤    È� xNÿ   £  är ÿ‡      ä x]NÿÀŽ ä/ xO ÿÀŽ ä xP]€€€€ÿüŽ ä xNÿ   £  Ä xX] øŽ â $xdP   O
Ž â xSNÿÀŽ â $xN    ÿ Ž ä$xP    ÿ Ž â xO£€€€€TìŽ ä xNNÿÀŽ ä rXXd   ]`Ž Ä xWS€€€€ÿüŽ ä xTPÿÀŽ â $xPO   N
Ž â x[X˜º  ÿ    ä xL]˜º  ÿ    ä rWTW   ÿáÿ ä xd[€€€€dŽ Ä rTST   WXŽ ä x[[L`Ž ä xNN£øŽ äxOP£xŽ äxL£ÿÀŽ ä x£ÿ   £  Ä xTT˜º  ÿ    ä xSS˜º  ÿ    ä x££ÿÀŽ ä x ÿ      ä rNON   ÿÀŽ ä xWT€€€€WŽ Ä xTTS`Ž ä r¦ÿ�   ¦  äxS£€€€€ÿüŽ ä x  ÿÀŽ ä xON˜º  ÿ    ä xLL˜º  ÿ    Ä r§ÿ‚   §  â $x¦¦   ÿ Ž â r¨ÿ   ¨  â $xSS    
Ž â xPO€€€€PŽ äxOOL`Ž ä x§§   ÿÀŽ ä r©ÿ€   ©  â $x¨¨   ÿ Ž â rOOP   ÿüŽ Ä x§¦0   §âŽ ä x©©   ÿÀŽ ä r £    SXŽ ä rTTW   ÿüŽ ä x©¨0   ©âŽ ä x§§àÿÿÿÿàÿ Ä xL ˜º  ÿ    â &r Oj   ÿ   â x££˜º  ÿ    â &rTTi   ÿ   â x©©àÿÿÿÿàÿ â $r§§    ÿŽ â xSL€€€€SŽ äxPL£`Ž â $rT©T   ÿŽ â x��T  ž    â 0r ÿ¬    A   â Er§ §       Ä rªÿ{   ª  ä rSPS   ÿüŽ ä rPÿ�   ÿ3Ž â 0r®ÿ®    A   â r«ÿ~   «  â #rO §   ÿ   â ErT T       â $xªª   ÿ Ž â xŸ T  Ÿ    Ä rPÿ‡   P  ä x««   ÿÀŽ â #rO®T   O   â $tXÿÿ Ž â xWÿ   Ÿ  ä xTP   ÿ   ä r[[d   ÿüŽ ä x«ª0   «âŽ Ä xeWÿÀŽ ä xWŸ€€€€XìŽ ä xTTÿÀŽ â &r[[h   ÿ   â xL«àÿÿÿÿàÿ ä x]PÿÀŽ â $xXW   T
Ž â x^e€€€€ÿüŽ â $rLL[   ÿŽ â x[ÿ   Ÿ  Ä xTTŸøŽ â $xd^   ]
Ž â xWXŸxŽ â $x]P   ÿ Ž â x[[ÿÀŽ ä rTWT   ÿÀŽ ä xW[€€€€ÿüŽ ä x^]ÿÀŽ Ä x]ePøŽ Æ $xWW   ^
Ž â xPÿ   P  ä xfÿ   Ÿ  ä r]]d   e`Ž ä r^[^   WXŽ ä xgPÿÀŽ ä xffÿÀŽ Ä x]]˜º  ÿ    ä xPe˜º  ÿ    ä x^^˜º  ÿ    ä xŸŸÿÀŽ ä xžf€€€€ÿüŽ ä xd]€€€€dŽ Ä xe]P`Ž ä r¢ÿ�   ¢  ä x]^€€€€WŽ ä xWT˜º  ÿ    ä r¡ÿ‚   ¡  ä xPŸ˜º  ÿ    â $xžž   g
Ž â xXW€€€€XŽ â $x¢¢   ÿ Ž â x¡¡   ÿÀŽ Ä x[[˜º  ÿ    ä xWWP`Ž ä rµÿ€   µ  ä r²ÿ   ²  ä r¯ÿp   ¯  ä r°ÿ{   °  Ä rgfg   žXŽ ä r­ÿz   ­  ä x¡¢0   ¡âŽ ä x^^[`Ž ä r³ÿ~   ³  ä xµµ   ÿÀŽ Ä x²²   ÿ   ä rWWX   ÿüŽ â $x¯¯   ÿ Ž â r¹ÿp   ¹  â $x°°   ÿ Ž â xgg˜º  ÿ    â &rPWj   ÿ   â xN­   ÿÀŽ ä xff˜º  ÿ    Ä r^^]   ÿüŽ ä x³³   ÿÀŽ ä r·ÿz   ·  ä xµ²0   µâŽ ä x¡¡àÿÿÿÿàÿ â $xX¹   ÿ Ž â xžg€€€€žŽ â &r^^i   ÿ   â xN¯0   NâŽ â $rP¡P   ÿŽ â reed   ÿüŽ Ä xggf`Ž ä x³°0   ³âŽ ä x··   ÿÀŽ ä xµµàÿÿÿÿàÿ â &rTeh   ÿ   â rggž   ÿüŽ â &rSS´   ÿ   â x³³àÿÿÿÿàÿ â $r^µ^   ÿŽ â x·X0   ·âŽ Ä xNNàÿÿÿÿàÿ ä ErW P       ä xº»T  º    â $rT³T   ÿŽ â x··àÿÿÿÿàÿ â &rPg´   ÿ   â Er^ ^       â $rNNS   ÿŽ â ErL L       â 0r¶ÿ¶    A   â #rW W   ÿ   â rºÿº   ÿ3Ž â $rP·P   ÿŽ â ErT T       â 0r¼ÿ¼    A   â ErN N       â #rW®^   W   â #rO¶L   O   ârºÿ‡   º  â $tSÿÿ Ž â ErL P       â #rW¶T   W   â #rO¼N   O   â x¸½T  ¸    â $xNº   ÿ Ž â #rL¼L   W   â $xWº   ÿ Ž â xP¸€€€€SìŽ Ä xSNÿÀŽ ä xXÿ   ¸  äx[WÿÀŽ â $xWP   S
Ž â x]XÿÀŽ ä xSS¸øŽ äxPW¸xŽ Ä x^ºÿÀŽ ä xe]€€€€ÿüŽ ä xN¸ÿÀŽ ä xTÿ   ¸  ä rPPS   ÿÀŽ â $x^e   ^
Ž â xSN˜º  ÿ    Ä xTTÿÀŽ ä xN]ºøŽ ä xPP˜º  ÿ    ä xXT€€€€ÿüŽ ä rNN^   ]`Ž ä xWP€€€€WŽ Ä xPPS`Ž ä x¸ÿ   ¸  ä rX[X   ÿáÿ ä xSN˜º  ÿ    ä xN]˜º  ÿ    ä rWPW   ÿüŽ Ä r¾ÿ�   ¾  ä xºÿ   º  ä x¸¸ÿÀŽ ä r[T[   XXŽ ä x^S€€€€^Ž ä r¿ÿ‚   ¿  Ä xSSN`Ž â &rNWj   ÿ   â xWºÿÀŽ â $x¾¾   ÿ Ž â xP¸€€€€ÿüŽ ä rÀÿ   À  ä x[[˜º  ÿ    ä xTT˜º  ÿ    Ä x¿¿   ÿÀŽ ä rÁÿ€   Á  â $xPP   W
Ž â xX[€€€€XŽ â$xÀÀ   ÿ Ž â x[[T`Ž ä x¿¾0   ¿âŽ ä xÁÁ   ÿÀŽ Ä rX[X   ÿüŽ ä x¿¿àÿÿÿÿàÿ ä rW¸W   PXŽ ä xÁÀ0   ÁâŽ ä xÄÄT  Å    â $rN¿N   ÿŽ â xWW˜º  ÿ    â &rXXi   ÿ   â x¸¸˜º  ÿ    Ä xÁÁàÿÿÿÿàÿ ä rÄÿÄ   ÿ3Ž ä xPW€€€€PŽ â $rXÁX   ÿŽ â x[W¸`Ž ä rÄÿ‡   Ä  ä rWS^   ÿüŽ Ä ErS N       ä r[[P   ÿüŽ â $xPÄ   ÿ Ž â ErX X       â #rS S   ÿ   â xT     ä xÇÈT  Ç    â &rNWh   ÿ   â xPPÿÀŽ â #rS®X   S   â xWÇ€€€€TìŽ Ä raÿp   a  ä xXÿ   Ç  â $xTW   P
Ž â rcÿz   c  â $x^a   ÿ Ž â xaXÿÀŽ ä xcc   ÿÀŽ ä xPPÇøŽ Ä xWTÇxŽ ä xXÄÿÀŽ ä x]a€€€€ÿüŽ ä xc^0   câŽ ä rWWP   ÿÀŽ â $x^]   X
Ž â xPÇÿÀŽ Ä xXÿ   Ç  â $x]Ä   ÿ Ž â xWW˜º  ÿ    ä xPP˜º  ÿ    ä xXXÿÀŽ ä xTW€€€€TŽ äxWWP`Ž Ä x]]ÿÀŽ ä xÇÿ   Ç  ä xPX€€€€ÿüŽ ä xdÿ   Ä  äxÄaÄøŽ â $xPP   ]
Ž â xÇÇÿÀŽ Ä rÄÄ^   a`Ž ä rWWT   ÿüŽ ä xddÿÀŽ ä xeÇ€€€€ÿüŽ ä xTa˜º  ÿ    ä xaÄ˜º  ÿ    Ä r]X]   PXŽ ä rQÿ‚   Q  ä rRÿ�   R  â $xee   d
Ž â x^a€€€€^Ž ä rVÿ   V  ä x]]˜º  ÿ    Ä xaaT`Ž ä xXX˜º  ÿ    ä xQQ   ÿÀŽ ä xTR   ÿ   ä rUÿ€   U  â $xVV   ÿ Ž â rÂÿ{   Â  Ä rZÿ{   Z  ä rdÇd   eXŽ ä xP]€€€€PŽ ärÃÿ~   Ã  ä x]]X`Ž ä xQT0   QâŽ Ä rYÿ~   Y  ä xUU   ÿÀŽ â $xÂÂ   ÿ Ž â rMÿp   M  â $xZZ   ÿ Ž â xdd˜º  ÿ    ä xÃÃ   ÿÀŽ ä xÇÇ˜º  ÿ    Ä rR]P   ÿüŽ â &rPWj   ÿ   â xYY   ÿÀŽ ä r_ÿz   _  ä xUV0   UâŽ ä xQQàÿÿÿÿàÿ â $xVM   ÿ Ž â xed€€€€eŽ â &rRRi   ÿ   â xÃÂ0   ÃâŽ â $rPQP   ÿŽ â xddÇ`Ž Ä raa^   ÿüŽ ä xYZ0   YâŽ ä x__   ÿÀŽ ä xUUàÿÿÿÿàÿ â &rTah   ÿ   â xÃÃàÿÿÿÿàÿ ä rede   ÿüŽ â $rRUR   ÿŽ â xYYàÿÿÿÿàÿ Ä x_V0   _âŽ ä ErM P       â $rNÃN   ÿŽ â xccàÿÿÿÿàÿ â $rTYT   ÿŽ â x__àÿÿÿÿàÿ â &rP[´   ÿ   â ErR R       â &rVe´   ÿ   â #rM M   ÿ   â $rPcP   ÿŽ â ErN N       â $rV_V   ÿŽ â ErT T       â #rQ®R   M   â 0rMÿ`    A   â ErP P       â 0rÆÿÆ    A   â ErV V       â #rS¶N   S   â#rQ¶T   Q   â #r”MO   ”   â #r•ÆL   •   â #rM¼P   S   â#rL¼V   Q   â 0rOÿb    A   Ä 0r\ÿ\    A   â G‰  0    € ê rP@�   ÿàõ ärR<�   ÿàó Æ $xQA   � ä$xS=   �Ž  Æ �yNP     ¨  �y]P    ¨  �y`R"    è �yaR     â rT�}   ÿàó Ä rZ�v   ÿàõ Æ $xU�   yŽ  â rV�|   ÿàó â $x[�   o Æ rW�x   ÿäÿ  â �ycT`    â rX�w   ÿàó ärR�s   ÿàõ â/ �ybV`    ä $xY�   rŽ  â r^�q   ÿàó â $xS�   l â�yZZ`    æ  $x_�   kŽ  â �yYX`    â  rV�u   ÿàõ Ä/ rT�t   ÿàó â�ydR`    $ $xW�   n ä$xU�   mŽ  â �y_^`    ( �yWV`    ( �yeT`    " rP�   ÿàó È rQ�   ÿäÿ  Ê �yPPl    " $tSÿÿ Ž â/ xNNT  ]    ÈO rNÿN   ÿ3Ž ä x`aT  `    ä� rNÿ‡   N  ä xXÿ   `  Æ $xRN   ÿ Ž â x[XÿÀŽ ä xT`€€€€SìŽ ä xV[€€€€ÿüŽ ä xSNÿÀŽ ä xQRÿÀŽ Æ $xVV   S
Ž ä $xST   Q
Ž â xT[NøŽ ä xRQ`øŽ äxUS`xŽ ä rTTV   [`Ž ä rRUR   ÿÀŽ Ä xQ`ÿÀŽ ä xUT˜º  ÿ    ä xT[˜º  ÿ    ä xRR˜º  ÿ    ä xQQ˜º  ÿ    ä xVU€€€€VŽ Ä xUUT`Ž ä xTÿ   `  ä xSR€€€€SŽ äxRRQ`Ž â $xQN   ÿ Ž â xTTÿÀŽ ä xXÿ   N  Ä xNT€€€€ÿüŽ ä xQQÿÀŽ ä x`ÿ   `  Æ $xNN   Q
Ž â x``ÿÀŽ ä rbÿ�   b  ä x^`€€€€ÿüŽ ä x[XÿÀŽ ä rQTQ   NXŽ ä rcÿ‚   c  â $xbb   ÿ Ž â rZÿ   Z  Ä r^[^   ÿáÿ ä xQQ˜º  ÿ    ä xcc   ÿÀŽ ä xTT˜º  ÿ    ä rYÿ€   Y  â $xZZ   ÿ Ž â rSRS   ÿüŽ Ä rdÿ{   d  är[`[   ^XŽ ä xRQ€€€€NŽ äxcb0   câŽ ä xQQT`Ž ä r_ÿ~   _  Ä xYY   ÿÀŽ â $xdd   ÿ Ž â rWÿp   W  ä x[[˜º  ÿ    â &rNSj   ÿ   â x``˜º  ÿ    ä xccàÿÿÿÿàÿ ä rRQR   ÿüŽ Ä x__   ÿÀŽ ä reÿz   e  ä xYZ0   YâŽ ä rUUV   ÿüŽ â $xVW   ÿ Ž â x^[€€€€^Ž â $rNcN   ÿŽ â x[[``Ž â &rRRi   ÿ   â x_d0   _âŽ Ä xee   ÿÀŽ ä xYYàÿÿÿÿàÿ â &rTUh   ÿ   â r[[^   ÿüŽ ä x__àÿÿÿÿàÿ â $rRYR   ÿŽ â xeV0   eâŽ ä ErQ N       â $rT_T   ÿŽ â xeeàÿÿÿÿàÿ â &rN[´   ÿ   â ErR R       â #rQ Q   ÿ   â ErT T       â $rNeN   ÿŽ Ä #rQ®R   Q   Æ ErN N       â #rS¶T   Q   â 0rQÿP    A   Æ #rN¼N   S   È #rŽQN   Ž   ä G‰  0    € ê rPF�   ÿàõ ärR>�   ÿàó Æ $xQG   � ä$xS?   �Ž  Æ �yNP     ¨  �y]P    ¨  �y`R"    è �yaR     â rT�   ÿàó È rU�   ÿäÿ  ä rV
�   ÿàó ärZ.�   ÿàõ â�ycT`    ä $xW   �Ž  ârX4�   ÿàó â$x[/   � â rR"�   ÿàõ Ä/ �ybV`    â $xY5   �Ž  âr^(�   ÿàó â$xS#   � â�yZZ`    æ  $x_)   �Ž  â rT�   ÿàó â	�yYX`    â rV�   ÿàõ Æ/ �ydR`    â rW�   ÿä â $xU   �Ž  ä �y_^`    è �yWV`    è �yeT`    â rP�   ÿàó Ê $xQ   �Ž  Ê �yPPl    â  $tSÿÿ Ž â/ xNNT  ]    ÈO rNÿN   ÿ3Ž ä x`aT  `    ä� rNÿ‡   N  ä xXÿ   `  Æ$xRN   ÿ Ž â x[XÿÀŽ ä xT`€€€€SìŽ ä xV[€€€€ÿüŽ ä xSNÿÀŽ ä xQRÿÀŽ Æ $xVV   S
Ž ä $xST   Q
Ž â xT[NøŽ ä xRQ`øŽ äxUS`xŽ ä rTTV   [`Ž ä rRUR   ÿÀŽ Ä xQ`ÿÀŽ ä xUT˜º  ÿ    ä xT[˜º  ÿ    ä xRR˜º  ÿ    ä xQQ˜º  ÿ    ä xVU€€€€VŽ Ä xUUT`Ž ä xTÿ   `  äxSR€€€€SŽ äxRRQ`Ž â $xQN   ÿ Ž â xTTÿÀŽ ä x`ÿ   `  Ä xXÿ   N  ä xNT€€€€ÿüŽ ä xQQÿÀŽ ä x``ÿÀŽ ä rNQN   ÿáÿ ä x^`€€€€ÿüŽ Ä x[XÿÀŽ ä rbÿ�   b  ä rQTQ   NXŽ ä rcÿ‚   c  â $x^^   [
Ž â rZÿ   Z  â $xbb   ÿ Ž â xQQ˜º  ÿ    Ä xcc   ÿÀŽ ä xTT˜º  ÿ    ä rYÿ€   Y  â $xZZ   ÿ Ž â rSRS   ÿüŽ ä rdÿ{   d  ä r[`[   ^XŽ Ä xRQ€€€€NŽ äxcb0   câŽ ä xQQT`Ž ä r_ÿ~   _  ä xYY   ÿÀŽ â $xdd   ÿ Ž â rWÿp   W  Ä x[[˜º  ÿ    â &rNSj   ÿ   â x``˜º  ÿ    ä xccàÿÿÿÿàÿ ä rRQR   ÿüŽ ä x__   ÿÀŽ ä reÿz   e  Ä xYZ0   YâŽ ä rUUV   ÿüŽ â $xVW   ÿ Ž â x^[€€€€^Ž â $rNcN   ÿŽ â x[[``Ž â &rRRi   ÿ   â x_d0   _âŽ ä xee   ÿÀŽ Ä xYYàÿÿÿÿàÿ â &rTUh   ÿ   â r[[^   ÿüŽ ä x__àÿÿÿÿàÿ â $rRYR   ÿŽ â xeV0   eâŽ ä ErQ N       â $rT_T   ÿŽ â xeeàÿÿÿÿàÿ â &rN[´   ÿ   â ErR R       â #rQ Q   ÿ   â ErT T       â $rNeN   ÿŽ Ä #rQ®R   Q   Æ ErN N       â #rS¶T   Q   â 0rQÿP    A   Æ #rN¼N   S   È #r�QN   �   ä G‰  0    € ê rPH�   ÿàõ ärRB�   ÿàó ä rQI�   ÿä Æ $xSC   �Ž  ä �yNP     ¨  �y]P    ¨  �y`R"    è �yaR     â rT�   ÿàó Ä rZ0�   ÿàõ Æ $xU   �Ž  ârV�   ÿàó â $x[1   � Æ �ycT`    â $xW   �Ž  ârX6�   ÿàó ärR$�   ÿàõ â/�yZZ`    ä $xY7   �Ž  âr^*�   ÿàó â �ybV`    â  $xS%   � Æ $x_+   �Ž  â rT�   ÿàó â	�yYX`    â rV�   ÿàõ Æ �ydR`    â  rU�   ÿäÿ  â $xW   � ä �y_^`    è  �yeT`    è  �yWV`    â  rP�   ÿàó Ê $xQ	   �Ž  Ê �yPPl    â $tSÿÿ Ž â xNNT  ]    ÈO rNÿN   ÿ3Ž ä x`aT  `    ä� rNÿ‡   N  ä xXÿ   `  Æ$xRN   ÿ Ž â x[XÿÀŽ ä/ xT`€€€€SìŽ ä xV[€€€€ÿüŽ ä xSNÿÀŽ ä xQRÿÀŽ Æ $xVV   S
Ž ä $xST   Q
Ž â xT[NøŽ ä xRQ`øŽ äxUS`xŽ ä rTTV   [`Ž ä rRUR   ÿÀŽ Ä xUT˜º  ÿ    ä xQ`ÿÀŽ ä xT[˜º  ÿ    ä xRR˜º  ÿ    ä xVU€€€€VŽ ä xQQ˜º  ÿ    Ä xUUT`Ž ä xTÿ   `  ä xSR€€€€SŽ äxRRQ`Ž ä xTTÿÀŽ ä xQN   ÿ   Ä xXÿ   N  ä x`ÿ   `  ä xNT€€€€ÿüŽ ä xQQÿÀŽ ä x``ÿÀŽ Æ $xNN   Q
Ž â x^`€€€€ÿüŽ ä x[XÿÀŽ ä rbÿ�   b  ä rQTQ   NXŽ ä rcÿ‚   c  â $x^^   [
Ž â rZÿ   Z  â $xbb   ÿ Ž â xQQ˜º  ÿ    Ä xcc   ÿÀŽ ä xTT˜º  ÿ    ä rYÿ€   Y  â $xZZ   ÿ Ž â rSRS   ÿüŽ ä rdÿ{   d  ä r[`[   ^XŽ Ä xRQ€€€€NŽ äxcb0   câŽ ä xQQT`Ž ä r_ÿ~   _  ä xYY   ÿÀŽ â $xdd   ÿ Ž â rWÿp   W  Ä x[[˜º  ÿ    â &rNSj   ÿ   â x``˜º  ÿ    ä xccàÿÿÿÿàÿ ä rRQR   ÿüŽ ä x__   ÿÀŽ ä reÿz   e  Ä xYZ0   YâŽ ä rUUV   ÿüŽ â $xVW   ÿ Ž â x^[€€€€^Ž â $rNcN   ÿŽ â x[[``Ž â &rRRi   ÿ   â x_d0   _âŽ ä xee   ÿÀŽ Ä xYYàÿÿÿÿàÿ â &rTUh   ÿ   â r[[^   ÿüŽ ä x__àÿÿÿÿàÿ â $rRYR   ÿŽ â xeV0   eâŽ ä ErQ N       â $rT_T   ÿŽ â xeeàÿÿÿÿàÿ â &rN[´   ÿ   â ErR R       â #rQ Q   ÿ   â ErT T       â $rNeN   ÿŽ Ä #rQ®R   Q   Æ ErN N       â #rS¶T   Q   â 0rQÿP    A   Æ #rN¼N   S   È #r’QN   ’   â #r–OM   –   â #r—\L   —   â G‰  0    € ê rLJ�   ÿàõ ärND�   ÿàó Æ $xMK   � â rOE�   ÿäÿ  È �y\L     ¨  �y]L    ¨  �y^N"    è �y_N     â rP�   ÿàó Ä rR�   ÿàõ Æ $xQ   �Ž  ârT8�   ÿàó â$xS   � ârV2�   ÿàõ ä�yaP`    â $xU9   �Ž  ârN,�   ÿàó â/$xW3   � ârX&�   ÿàõ â�yRR`    ä $xO-   �Ž  â rZ �   ÿàó â �yVV`    â  rY'�   ÿä Ä rP�   ÿàõ â�yUT`    â $x[!   �Ž  Æ �yXX`    â  $xQ   � Æ �ycN`    è �y[Z`    è �yeP`    â rL:�   ÿàó Ê $xM;   �Ž  Ê �yLLl    â  $tOÿÿ Ž â/ x\\T  ]    ÈO r\ÿ\   ÿ3Ž ä x^_T  ^    ä� r\ÿ‡   \  ä xNÿ   ^  È xWNÿÀŽ â $xN\   ÿ Ž âxQ\ÿÀŽ ä xTW€€€€ÿüŽ äxP^€€€€OìŽ ä xOÿ   ^  ä xMNÿÀŽ â $xTT   Q
Ž â xQOÿÀŽ Ä rOMP   ÿáÿ ä xYW\øŽ â $xP\   ÿ Ž â xMM^øŽ äxNO^xŽ ärYYT   W`Ž ä xZÿ   ^  Ä rMNM   ÿÀŽ ä xYY˜º  ÿ    ä xNW˜º  ÿ    ä xSQ€€€€ÿüŽ ä xPPÿÀŽ ä xZZÿÀŽ Ä x\ÿ   \  ä xTY€€€€TŽ äxYYN`Ž â $xSS   P
Ž â xNZ€€€€ÿüŽ ä xW\ÿÀŽ ä x^^ÿÀŽ Ä rRÿ�   R  â $x\N   W
Ž â rPQP   SXŽ ä xNM˜º  ÿ    ä xM^˜º  ÿ    ä raÿ‚   a  â $xRR   ÿ Ž â rVÿ   V  Ä xPP˜º  ÿ    ä xON€€€€OŽ äxQQ˜º  ÿ    ä xNNM`Ž ä rUÿ€   U  ä xaa   ÿÀŽ â $xVV   ÿ Ž â rXÿ{   X  Ä xSP€€€€SŽ ä rWZW   \XŽ ä xPPQ`Ž ä rONO   ÿüŽ ä xUU   ÿÀŽ ä rcÿ~   c  Ä xaR0   aâŽ â $xXX   ÿ Ž â xWW˜º  ÿ    â &rjOj   ÿ   â xZZ˜º  ÿ    ä rPPS   ÿüŽ ä xUV0   UâŽ ä xcc   ÿÀŽ Ä r[ÿz   [  ä reÿp   e  ä xaaàÿÿÿÿàÿ ä x\W€€€€\Ž â &rPPi   ÿ   â rYYT   ÿüŽ â $raaj   ÿŽ â xWWZ`Ž Ä xcX0   câŽ ä x[[   ÿÀŽ ä xNe   ÿ   ä xUUàÿÿÿÿàÿ â &rhYh   ÿ   â rWW\   ÿüŽ ä x[N0   [âŽ â $rPUP   ÿŽ â xccàÿÿÿÿàÿ Ä Era a       â &r´W´   ÿ   â x[[àÿÿÿÿàÿ â $rcch   ÿŽ â ErP P       â #ra a   ÿ   ä $r[[´   ÿŽ â Erc c       â #ra®P   a   Æ Er[ [       â #ra¶c   a   â 0rLÿL    A   Æ #r[¼[   a   È #r“L[   “   â xŠŠ   ÿàÿ ä x��p  ÿàõ ä | Š   pbò ä x‹‹p  ÿà÷ â $r�ÿÿ   � â xˆˆ@   ÿàÿ Æ $r‰ÿÿ   ‰Ž Ì D       À â Gy   ¾ÿÿÿÿƒ ê Ay        € ê y      "   " Ey  0    € â r ÿ   pRð är ÿ   pRò Ö G‰  ð     € ê z ÿ `  pPð â $x€   œŽ Æ z ÿ a   Sð ä ˆs ” þÿ H   ö ˆƒ •€þÿ H   è ˆƒ – ÿÿ H   è ˆƒ —€ÿÿ H   â G‰  p     € ê ˆs Ž   H   è ˆs •€þÿ H   è ˆs �€  H   è ˆs – ÿÿ H   è ˆs ’   H   è ˆs —€ÿÿ H   è ˆs “€  H   ä Ay        € ê {          ì M        € ê y       &   ¢ z ÿ `  pPð â $tÿ   ÿ Ž Ä y      '   "> z ÿ a   Sð â $z   x  ›Ž ÈO $z  ~   Ž È %v  j   Ž È G	  P    € ê „yœ     H   " z › r  p`ð Æ „yœ €   H   b r œÿ   pVp  Æ „yœ    H   ¨ „y œ    H   è „y	œ €  H   ( „yœ    H   ( „yœ    H   ( „yœ €  H   ( „yœ    H   ( „y
œ €  H   ( „yœ €  H   " !r”       Æ „yœ €  H   " !r•       â/ !r–       âO !r         ä� !r	       â!r       â !r         ä !r       Æ ‰       b !r       Æ ‰      ¢ !r

—       Æ ‰      â !r

       È !r
       â !r        â/ !r       ÈO ‰      " !r
       Æ� ‰      h ‰      ¨ ‰	
      â !r       â !r       â/ !r       ÈO ‰      " !r
	       Æ� ‰ €    h ‰  €    ¨ ‰	 €    â !r       â !r       È/ ‰ €    " !r
        ÆO ‰  @    b !r		       â� x›   ÿàÿ ä ‰
 @    ¤ z  r  p`ô ä ‰	 @    ä x œ   pVt â !r       â !r         È/ ‰ @    " x›   ÿàÿ â!r
       âO x››   ÿàÿ â ‰       b z  r  p`ò â !r	       ä� ‰      ¢ x œ   pVò  Æ ‰      â !r       â !r        È/ ‰       " !‚˜       â!r	       âO !r       Æ� †‰     â z › r  p`ð â !¢		˜       â!’˜       ä x œ   pVp  ä †© 	   è †™    â !r        Ì M	        € ê !r˜       Ê/ †y    â My        € ê $v ÿ h  ÿ Ž Ê x     pBð Ú G	  @    € ê x     pRð Ú G‰  À    € ê x     pRð Ú G	   $    € ê „y	œ     H   " z › r  p`ð â Ey  p    € ä „yœ    H   b r œÿ   pVp  Æ „yœ    H   ¨ „y œ    H   è „yœ    H   ( „yœ  
  H   ( „yœ €   H   ( „yœ €  H   ( „yœ €  H   ( „yœ €  H   " !r		”       â !rŽ       Æ/ !r	       âO !r         Æ� !r       â!r         È ‰      " !r•       Æ ‰       b !r�       â !r       È !r       â !r       â !r        È/ ‰      ( ‰      b !r       â !r	       È/ ‰ €    ( ‰
	 €    h „yœ €  H   ¢ !r       Æ „yœ €
  H   â !r
	
       Æ/ ‰ @    ( ‰
 @    b !r       âO !r       â !r
       È/ ‰       ( ‰      b !r       â !r	        â !r       â/ G	  �     € ê !r™       â!r 	˜       Æ  x;ª¸¿  A  Ì s        $ !t  €?    Ì s        $  r     A  È  r      A  Ê †y     ä Ay        € ê ‰       b Ey  Ð    € æ ‰      ¨ „yœ    H   è „yœ    H   ( „yœ    H   â !r         Æ/ „yœ    H   â !r       âO x›   ÿàÿ ä ‰       d z  r  p`ð ä ‰      $ x œ   pVp  â !r–       â� !r’       â!r        Ä/ !r       â !r	       ä ‰ €    ( ‰
	 €    b !r       Æ „yœ  	  H   ¢ !r       Æ „yœ    H   â !r
	
       Æ/ ‰ @    ( ‰
 @    b !r       âO !r       â !r
       È/ ‰       ( ‰      b !r       â !r	        â !r       â/ G	  �     € ê !r™       â!r 	˜       Æ  x;ª¸¿  A  Ì s        $ !t  €?    Ì s        $  r     A  È  r      A  Ê †y    ä Ay        € ê ‰       b Ey       € æ ‰      ¨ „yœ €	  H   è „yœ €  H   è „yœ €  H   è „yœ €  H   â !r         Æ/ „yœ €  H   â !r       ÆO ‰       h ‰      ( „yœ €  H   ¢ !r        â/ !r	       È ‰ €    ( ‰
	 €    b !r“       ÈO !r       â !r       â x›   ÿàÿ ä !r       â !r
	
       â/ ‰ @    " z  r  p`ð Æ ‰
 @    b x œ   pVp  â !r       â !r
       È/ ‰       ( ‰      b !r	        â !r       â/ G	  �     € ê !r™       â!r 	˜       Æ  x;ª¸¿  A  Ì s        $ !t  €?    Ì s        $  r     A  È  r      A  Ê †y    ä Ay        € ê ‰      b !r—       â x››   ÿàÿ Æ !r       â z › r  p`ð Æ !r       â x œ   pVp  È ‰       ¢ !r       Ê/ ‰      " !r         ÊO ‰       b !r	       Ê ‰
	 €    " !r        Ê/ ‰ €    b !r
	
       Ê ‰
 @    " !r       Ê/ ‰ @    b !r
       Ê ‰      " !r       Ê/ ‰       ¢ !r       â M	        € ì !r™™       â/!r        ÆO  x™;ª¸¿  A  â !r˜˜       Ê s        $ !t  €?    Ì s        $  r™™     A  È  r™˜™     A  Ê †y ™   â My        € ê „yœ     H   " z › r  p`ð â Ey       € ä „yœ    H   b r œÿ   pVp  Æ „y œ    H   ¨ „yœ    H   è „y	œ    H   ( „yœ  
  H   ( „yœ €   H   ( „yœ €  H   ( „yœ €  H   ( „yœ €
  H   â !r”       â !rŽ       Æ/ !r         âO !r       Æ� !r	 	       â!r       È ‰ 	      " !r•       Æ ‰      b !r       È !r       â !r 	        â !r       È/ ‰       ( ‰      h „yœ €  H   ¢ !r        â !r       È/ ‰ €    ( ‰ €    h „yœ €  H   â !r�       âO !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       È� !r       â !r       â !r
	
       È/ ‰       ( ‰
      b !r	        â !r 
       â/ G	  À     € ê !r  ™       È x '7=  A  â x *BL?  A  Æ #t   €?   È  r     A  â  x    ?  A  â !r 	˜       È s     $   $ !t  €?    È  r     A  È  r      A  Ê †y     ä Ay        € ê ‰       b Ey       € æ ‰      ¨ „yœ    H   è „yœ    H   ( „yœ  	  H   â !r         Æ/ „yœ    H   â !r       ÆO ‰       h ‰      ( „yœ    H   ¢ !r–       â� !r’       â!r        Æ/ !r       â !r       â ‰ €    ( ‰ €    h „yœ    H   â !r       âO !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       È� !r       â !r       â x›   ÿàÿ â !r
	
       Æ/ ‰       " z  r  p`ð Æ ‰
      b x œ   pVp  â !r	        â !r 
       Ö/ G	  À     € ê !r  ™       È x '7=  A  â x *BL?  A  Æ #t   €?   È  r     A  â  x    ?  A  â !r 	˜       È s     $   $ !t  €?    È  r     A  È  r      A  Ê †y    ä Ay        € ê ‰       b Ey  Ð    € æ ‰      ¨ „yœ €  H   è „yœ €	  H   ( „yœ €  H   â !r         Æ/ „yœ €  H   â !r       ÆO ‰       h ‰      ( „yœ €  H   ¢ !r“       È� !r       â!r        â/ !r       È ‰ €    ( ‰ €    h „yœ €  H   â !r       âO !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       â x›   ÿàÿ â !r
	
       Æ/ ‰       " z  r  p`ð Æ ‰
      b x œ   pVp  â !r	        â !r 
       Ö/ G	  À     € ê !r  ™       È x '7=  A  â x *BL?  A  Æ #t   €?   È  r     A  â  x    ?  A  â !r 	˜       È s     $   $ !t  €?    È  r     A  È  r      A  Ê †y    ä Ay        € ê ‰      b !r—       â x››   ÿàÿ Æ !r       â z › r  p`ð Æ !r       â x œ   pVp  È ‰       ¢ !r       Ê/ ‰      " !r         ÊO ‰       b !r       Ê ‰ €    " !r        Ê/ ‰ €    b !r	       Ê ‰
	 @    " !r       Ê/ ‰ @    b !r
	
       Ê ‰
      " !r       Ê/ ‰       ¢ !r
       â M	        € ì !r™™       â/!r        ÆO  x™'7=  A  â x™*BL?  A  â !r˜˜       ä #t™  €?   â x™™   ?  A  Æ  r     A  Ì s     $   $ !t   €?    È  r™™      A  È  r™˜™     A  Ê †y ™   â My        € ê x     pRð Ú G‰  P    € ê x     pRð Ú G	  `    € ê „yœ     H   " z › r  p`ð â Ey       € ä „yœ    H   b r œÿ   pVp  Æ „y œ    H   ¨ „yœ    H   è „y	œ    H   ( „yœ  
  H   ( „yœ €   H   ( „yœ €  H   ( „yœ €  H   ( „yœ €
  H   â !r”       â !rŽ       Æ/ !r         âO !r       Æ� !r	 	       â!r       È ‰ 	      " !r•       Æ ‰      b !r       È !r       â !r 	        â !r       È/ ‰       ( ‰      h „yœ €  H   ¢ !r        â !r       È/ ‰ €    ( ‰ €    h „yœ €  H   â !r�       âO !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       È� !r       â !r       â !r
	
       È/ ‰       ( ‰
      b !r	        â !r 
       â/ G	  À     € ê !r  ™       Ê	z  i    � â !r 	˜       È  x;ª¸¿  A  â 	z   i    � È 	z   i €  � â s        $ !t  €?    Ì s        $  r     A  È  r      A  Ê †y     ä Ay        € ê ‰       b Ey       € æ ‰      ¨ „yœ    H   è „yœ    H   ( „yœ  	  H   â !r         Æ/ „yœ    H   â !r       ÆO ‰       h ‰      ( „yœ    H   ¢ !r–       â� !r’       â!r        Æ/ !r       â !r       â ‰ €    ( ‰ €    h „yœ    H   â !r       âO !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       È� !r       â !r       â x›   ÿàÿ â !r
	
       Æ/ ‰       " z  r  p`ð Æ ‰
      b x œ   pVp  â !r	        â !r 
       Ö/ G	  À     € ê !r  ™       Ê	z  i    � â !r 	˜       È  x;ª¸¿  A  â 	z   i    � È 	z   i €  � â s        $ !t  €?    Ì s        $  r     A  È  r      A  Ê †y    ä Ay        € ê ‰       b Ey  Ð    € æ ‰      ¨ „yœ €  H   è „yœ €	  H   ( „yœ €  H   è „yœ €  H   â !r         Æ/ „yœ €  H   â !r       ÆO ‰       h ‰      " !r“       È� !r       â!r        â/ !r       È ‰ €    ( ‰ €    h „yœ €  H   ¢ !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       âO !r       â x›   ÿàÿ â !r
	
       Æ/ ‰       " z  r  p`ð Æ ‰
      b x œ   pVp  â !r	        â !r 
       Ö/ G	  À     € ê !r  ™       Ê	z  i    � â !r 	˜       È  x;ª¸¿  A  â 	z   i    � È 	z   i €  � â s        $ !t  €?    Ì s        $  r     A  È  r      A  Ê †y    ä Ay        € ê ‰      b !r—       â x››   ÿàÿ Æ !r       â z › r  p`ð Æ !r       â x œ   pVp  È ‰       ¢ !r       Ê/ ‰      " !r         ÊO ‰       b !r       Ê ‰ €    " !r        Ê/ ‰ €    b !r	       Ê ‰
	 @    " !r       Ê/ ‰ @    b !r
	
       Ê ‰
      " !r       Ê/ ‰       ¢ !r
       â M	        € ì !r™™       â/!r        ÈO 	z™™ i    � â !r˜˜       È  x™;ª¸¿  A  â 	z˜˜ i    � È 	z˜˜ i €  � â s        $ !t  €?    Ì s        $  r™™     A  È  r™˜™     A  Ê †y ™   â My        € ê „yœ     H   " x›   ÿàÿ äx›   ÿàÿ â„yœ    H   b z › r  p`ð ä z  r  p`ò â „yœ €   H   ¢ z  r  p`ô Ä r œÿ   pVp  â„yœ €  H   â x œ   pVò  äx œ   pVt â „yœ    H   " x››   ÿàÿ Æ „yœ    H   " z › r  p`ö Æ „y œ    H   " x œ   pVö Æ „yœ    H   ( „yœ €  H   ( „yœ €  H   ( „y
œ    H   " !r”       Æ „yœ  	  H   " !rŽ       Æ/ „y	œ    H   b !r•       ÆO „yœ  
  H   ¢ !r��       Æ� „yœ €  H   â !r–       Æ„yœ €
  H   " !r’       Æ „yœ    H   " !r         Æ „yœ    H   " !r       Æ „yœ €  H   " !r       Æ „yœ €  H   " !r�       Æ „yœ €  H   " !r

       Æ „y!œ €	  H   â !r       Æ „y#œ €  H   â !r 	       Æ/ „y%œ €  H   â !r       âO !r       ä� ‰       " !r       Æ‰      b !r

       Æ ‰      ¢ !r       Æ ‰      â !r—       Æ ‰	
      " !r“       Æ ‰      " !r       â !r         â !r!       â !r       â/ !r##       â !r       äO ‰       " !r%%       â !r       Ä� ‰#      b !r

	       Æ‰%      ¢ !r       Æ ‰      è ‰	      ( ‰      ( ‰
      " !r        Æ ‰      " !r#       â/ !r%       äO ‰  €    b !r       Æ� ‰      ¢ !r		       Æ‰      â !r       Æ ‰ €    " !r
       Æ ‰	 €    " !r       Æ ‰ €    ( ‰
 €    " !r         Æ/ ‰ €    b !r       âO !r       ä� ‰  @    ¢ !r       Æ‰ €    â !r	       Æ ‰ €    " !r       Æ ‰ @    " !r

       Æ ‰	 @    " !r       Æ/ ‰ @    h ‰
 @    b !r        ÆO ‰ @    ¢ !r       â� !r       ä‰       â !r       Æ ‰ @    " !r		       Æ ‰ @    " !r       Æ/ ‰      b !r
       Æ ‰	      b !r       ÆO ‰      ¨ ‰
      ¢ !r        Æ� ‰      â !r       â !‚ ˜       â!r       Æ‰      " !r       Æ/ ‰      b !r		       â !‚™       â!r       äO !’	˜       â  ‚      A  â !r
       â !’™       Æ †‰     â !r       â� !¢˜       â  ’     A  ä !¢	™       â !r       ä †™    â  ¢		     A  â !r       È/ †© 	   â M9        € ê !r˜˜       â !r™™       È  r™˜™     A  Ê †y ™   â My        € ê „yœ     H   " z › r  p`ð â Ey  À    € ä „yœ    H   b r œÿ   pVp  Æ „y œ    H   ¨ „yœ    H   è „y	œ    H   ( „yœ  
  H   ( „yœ €   H   ( „yœ €  H   ( „yœ €  H   ( „yœ €
  H   â !r”       â !rŽ       Æ/ !r         âO !r       Æ� !r	 	       â!r       È ‰ 	      " !r•       Æ ‰      b !r       È !r       â !r 	        â !r       È/ ‰       ( ‰      h „yœ €  H   ¢ !r        â !r       È/ ‰ €    ( ‰ €    h „yœ €  H   â !r�       âO !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       È� !r       â !r       â !r
	
       È/ ‰       ( ‰
      b !r        â !r 
       â/ G	  à     € ê !r  ™       Ê	x   à@  � Ê  x #ÛÙ?  A  È  x ;ª¸¿  A  â !r ˜       Ê s        " 	x    à@  � È 	x    àÀ  � Ê !t   €?    â !t  €?    È s	        $  r	     A  È  r     A  Ê †y     ä Ay        € ê ‰       b Ey       € æ ‰      ¨ „yœ    H   è „yœ    H   ( „yœ  	  H   â !r         Æ/ „yœ    H   â !r       ÆO ‰       h ‰      ( „yœ    H   ¢ !r–       â� !r’       â!r        Æ/ !r       â !r       â ‰ €    ( ‰ €    h „yœ    H   â !r       âO !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       È� !r       â !r       â x›   ÿàÿ â !r
	
       Æ/ ‰       " z  r  p`ð Æ ‰
      b x œ   pVp  â !r        â !r 
       Ö/ G	  à     € ê !r  ™       Ê	x   à@  � Ê  x #ÛÙ?  A  È  x ;ª¸¿  A  â !r ˜       Ê s        " 	x    à@  � È 	x    àÀ  � Ê !t   €?    â !t  €?    È s	        $  r	     A  È  r     A  Ê †y    ä Ay        € ê ‰       b Ey  ð    € æ ‰      ¨ „yœ €  H   è „yœ €	  H   ( „yœ €  H   è „yœ €  H   â !r         Æ/ „yœ €  H   â !r       ÆO ‰       h ‰      ( „yœ €  H   ¢ !r“       È� !r       â!r        â/ !r       È ‰ €    ( ‰ €    b !r       âO !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       â x›   ÿàÿ â !r
	
       Æ/ ‰       " z  r  p`ð Æ ‰
      b x œ   pVp  â !r        â !r 
       Ö/ G	  à     € ê !r  ™       Ê	x   à@  � Ê  x #ÛÙ?  A  È  x ;ª¸¿  A  â !r ˜       Ê s        " 	x    à@  � È 	x    àÀ  � Ê !t   €?    â !t  €?    È s	        $  r	     A  È  r     A  Ê †y    ä Ay        € ê ‰      b !r—       â x››   ÿàÿ Æ !r       â z › r  p`ð Æ !r       â x œ   pVp  È ‰       ¢ !r       Ê/ ‰      " !r         ÊO ‰       b !r       Ê ‰ €    " !r        Ê/ ‰ €    b !r	       Ê ‰
	 @    " !r       Ê/ ‰ @    b !r
	
       Ê ‰
      " !r       Ê/ ‰       ¢ !r
       â M	        € ì !r™™       â/!r        ÈO 	x™™  à@  � â !r˜˜       È  x™#ÛÙ?  A  â 	x˜˜  à@  � Æ  x;ª¸¿  A  â 	x˜˜  àÀ  � Ê s        " !t˜˜  €?    â !t  €?    Ì s         $  r™™      A  È  r™™˜     A  Ê †y ™   â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y6      "   $ z 6 w  p`ð Ú M	        € ê y7      !   " ¹z  l      â Ey      € â ™x?    â $r8ÿÿ   ÿ Ž â ¹z  F   
   â $r<ÿÿ   ÿ Ž â xHÿ   7  È | H   p`ð Ú G	  �    € ê y      &   " $tÿ   ÿ Ž ä $z6 x  Ž È %v \   Ž Ê �y?    ™ ¢  x:7   ÿÀŽ â$rÿÿ   ÿ Ž â x;7   ÿÀŽ â‚| ?       â x 7   ÿÀŽ ây      %   b x:   ÿ   â $v=ÿ X  ÿ Ž â x	7   ÿÀŽ â 'z m   Ž â x7   ÿ   Æ $x ;    
Ž â zÿ n    â %x�   ÿ Ž Æ r 	    ÿð� â $rÿÿ   
Ž â |   ÿàÿ â $v9ÿ Y  ÿ Ž â x(    ÿàÿ â $z o  Ž â x     ÿàÿ â %x    Ž â xÿ   (  ä xÿ      â $x7   ÿ Ž â r(   ÿ� â$z u  ÿŽ â r	(   ÿ� â $r8ÿÿ   ÿ Ž â r    ÿ� Ä r    ÿ� ä xøÿÿÿÿÀŽ ä xüÿÿÿÿÀŽ ä xÿ      â $x    
Ž âx    èÿ â $z6 r  ÿŽ â x     ÿàÿ Ä |   ÿàÿ ä r
    ÿ� är    ÿ� â %x$    Ž â xÿ      ä xøÿÿÿÿÀŽ ä r    ÿ� â%x$    Ž â r    ÿ� Ä x%
øÿÿÿÿÀŽ â $x   ÿ Ž â/ xøÿÿÿÿÀŽ â $x(   
Ž â x)	üÿÿÿÿÀŽ â $x$    %
Ž â x!üÿÿÿÿÀŽ â $x    
Ž â xüÿÿÿÿÀŽ ä x1   ÿÀŽ Ä x((   )èÿ ä x     !èÿ ä x    èÿ â $z  q  HŽ âz1 Z  1à‘ äz Z  ÿàõ ä xA   ÿàÿ Ä xFÿ     ä zJ [  ÿ$~  ä xÿ     â $xFF   ÿ Ž â xÿ     ä xÿ     ä xÿ     Ä xÿ     ä x!ÿ      ä x%ÿ   $  ä x)ÿ   (  ä x>ÿ     ä x@ÿ     Ä xBÿ     â $x>>   ÿ Ž â xCÿ   	  â $x@@   ÿ Ž â xDÿ   
  â $xBB   ÿ Ž â xEÿ     â $xCC   ÿ Ž â xGÿ     â $xDD   ÿ Ž â x11L   ÿàñ â $xEE   ÿ Ž â z [  ÿä â $xGG   ÿ Ž Ä $rJÿÿ   J  ä $z? t  HŽ äO$z?? t   Ž ä $zAA q  Ž ä %x
?n   Ž È %x?n   Ž È %x?n   Ž È %xAn   Ž È %xAn   Ž È %xAn    Ž È %x"An   $Ž È %x&An   (Ž È %x?n   ÿŽ È %xAn   ÿŽ È %xAn   Ž È %xAn   Ž È %xAn   Ž È %x?n   Ž È %x?n   Ž È %x ?n    Ž È %x$?n   $Ž È %x(?n   (Ž È $x 7   ÿ Ž ä %x;    Ž È %x,;    Ž â x     ÿÀŽ â �yP    • ¦ r*=    à‘ â �yZ$   • æ r+9ÿ   $~  â �ybH   • " r=   ÿàñ Æ �yR*    • h �yY*   • b $x   9  Ê �yV"   • ¨ �y[    • ¢ r.=    à‘ È r/9ÿ   	$~  ä r,,=   ÿàñ Æ �yT.    • è  �yW.   • â  $x--   9  Ê �yX,"   • è �y],    • â r*
=   ÿàñ är=   ÿàó ä/r2=   ÿàõ â$x+   9  âr4=   ÿàñ â$x   9Ž  âr0(=   ÿàó â$x3   9 â�yS*`   ‘ â  $x5   9  â r,=   ÿàñ Ä �yK`   ‘ æ $x-   9  â �yQ2`   ‘ â $r*ÿÿ   1 Ž ä $x1)   9Ž  â r`&=   ÿàó â�yO4`   ‘ â  r^=   ÿàñ â $r+ÿÿ   J Ž Ä $xa'   9Ž  âr. =   ÿàó â$x_   9  â �y 0`   ‘ â r4$=   ÿàñ Æ �yM,`   ‘ " $x/!   9Ž  ä$x5%   9  â r"=   ÿàñ â/�yL``   ‘ " r,=   ÿàó Æ �yI^`   ‘ " $x#   9  âr0=   ÿàñ â$x-   9Ž  âr2=   ÿàó â �yN*¸ÿÿ ™ $ $x1   9  ä�yU4`   ‘ " $x3   9Ž  Æ �y\.`   ‘ ( �yJ*Üÿÿ ™ " r4=   ÿàó Æ �y`,`   ‘ "! r.=   ÿàñ â $x5   9Ž  ä�y_`   ‘ $ $x/   9  ä �y22`   ‘ ( �y44`   ‘ ( �y10`   ‘ ( �y/.`   ‘ ( �y^*    ™ " $r,ÿÿ   = Ž Æ �yc*$   ™ " $r-ÿÿ   9 Ž Æ �yal   • " %x?n   ,Ž È/ %x,An   ,Ž ä �yl   • ( �y,,l   • " x.ÿ   :  â $t5ÿÿ Ž â xHH   ÿàÿ È | H   p`ð ä xRRT  Y    Èr3ÿR   ÿ3Ž È r3ÿ.   3  ä xV[T  V    ÆO $x03   ÿ Ž â xRV€€€€5ìŽ ä xYÿ   V  ä x50ÿÀŽ ä xdYÿÀŽ Æ $xR   5
Ž â xYd€€€€ÿüŽ ä xR3ÿÀŽ ä x05VøŽ äx-VxŽ ä/x5ÿ   V  â $x[Y   R
Ž â r0-0   ÿÀŽ â $x-3   ÿ Ž â xRÿ   V  Ä x55ÿÀŽ ä xfRÿÀŽ ä xed3øŽ äxY5€€€€ÿüŽ ä xR-ÿÀŽ ä x3ÿ   3  Ä xhf€€€€ÿüŽ â $xYY   R
Ž â x33ÿÀŽ ä ree[   d`Ž ä rR5R   YXŽ â $xhh   3
Ž â x-d˜º  ÿ    ä xde˜º  ÿ    Ä xRR˜º  ÿ    ä x55˜º  ÿ    ä r3f3   hXŽ ä xTTT  W    ä� x[d€€€€[Ž äxdd-`Ž Ä xYR€€€€YŽ äxRR5`Ž ä x33˜º  ÿ    ä xff˜º  ÿ    ä r5ÿT   ÿ3Ž ä xVVÿÀŽ Ä r-d[   ÿüŽ ä xh3€€€€hŽ äx[3f`Ž ä r3ÿ.   5  ä x00˜º  ÿ    ä x5V˜º  ÿ    Ä xW0€€€€Ž äx005`Ž â $t5ÿÿ Ž â xX]T  X    â $x.3   ÿ Ž â rRY   ÿüŽ ä xTÿ   X  ä xRX€€€€5ìŽ Ä x5.ÿÀŽ ä xVTÿÀŽ ä xT3ÿÀŽ â $xYR   5
Ž â x]V€€€€ÿüŽ ä x55XøŽ äxRYXxŽ â $x]]   T
Ž â xTÿ   X  Ä rW0W   ÿüŽ ä x0XÿÀŽ ä r5R5   ÿÀŽ â $xR3   ÿ Ž â xdV3øŽ ä xTTÿÀŽ ä r.[h   ÿüŽ Ä x[0˜º  ÿ    ä rdd]   V`Ž ä xeV˜º  ÿ    ä x05˜º  ÿ    ä xVT€€€€ÿüŽ ä x5RÿÀŽ Ä rKÿ@   K  ä rSÿ>   S  â $xVV   5
Ž â rOÿ@   O  ä r ÿC      ä xXÿ   X  â $xRK   ÿ Ž â rQÿ>   Q  Ä rMÿB   M  â$xfO   ÿ Ž â r5T5   VXŽ â $x     ÿ Ž â xSS   ÿÀŽ ä xY0€€€€YŽ ä xXXÿÀŽ ä x3ÿ   3  Ä x00[`Ž ä xQQ   ÿÀŽ ä xMM   ÿÀŽ ä rLÿC   L  ä x55˜º  ÿ    ä xTT˜º  ÿ    Ä xSR0   SâŽ ä rIÿB   I  ä xRX€€€€ÿüŽ ä x33ÿÀŽ ä xQf0   QâŽ ä rY0Y   ÿüŽ Ä xM 0   MâŽ â $x L   ÿ Ž â xV5€€€€VŽ äxK5T`Ž ä xII   ÿÀŽ â &r0WN   ÿ   âxSSàÿÿÿÿàÿ â $xRR   3
Ž â xQQàÿÿÿÿàÿ â &rNYN   ÿ   â rKKV   ÿüŽ Ä xI 0   IâŽ â $r0S0   ÿŽ â r3X3   RXŽ â $rNQN   ÿŽ â xIIàÿÿÿÿàÿ â &rLJ   ÿ   âr\ÿE   \  â&rJKJ   ÿ   â r`ÿE   `  â 0r ÿP    A   â xdd˜º  ÿ    Ä xK3˜º  ÿ    â $rIIJ   ÿŽ â Er3 0       ä Er N       ä rUÿD   U  är_ÿD   _  â $x\\   ÿ Ž â r2ÿG   2  â$x``   ÿ Ž â r4ÿG   4  Ä x]d€€€€]Ž â #r3 3   ÿ   â xdde`Ž â #r    ÿ   â xXX˜º  ÿ    ä xUU   ÿÀŽ ä r1ÿF   1  ä x__   ÿÀŽ Ä r/ÿF   /  â 0rZÿZ    A   â Er  I       â $x22   ÿ Ž â xRK€€€€RŽ â $x44   ÿ Ž â r5d]   ÿüŽ ä xMMàÿÿÿÿàÿ ä xKKX`Ž Ä xU\0   UâŽ ä x11   ÿÀŽ ä x_`0   _âŽ ä x//   ÿÀŽ â #rZ       â rRKR   ÿüŽ â &r -^   ÿ   â xUUàÿÿÿÿàÿ â $rLML   ÿŽ â x120   1âŽ â &r^5^   ÿ   â x__àÿÿÿÿàÿ Ä x/40   /âŽ â $r U    ÿŽ â x11àÿÿÿÿàÿ â &r..c   ÿ   âx//àÿÿÿÿàÿ â $r^_^   ÿŽ â ErL L       â &rRRc   ÿ   â Er          â $r.1.   ÿŽ â Er^ ^       â $r//R   ÿŽ â #r3ZL   3   â 0rbÿb    A   â Er. .       â 0raÿa    A   â Er/ /       Ä #r b    3   â#r^b^      â x1*@  ÿàõ â 0rÿ    A   â x@  ÿà÷ â 0r-ÿ,    A   â x==Ü   ÿàó â #r a.       â#r/a/   ^   â $rJÿÿ   + ä $rÿÿ   Ž â #r<    <   â #r8-/   8   â $r9ÿÿ   9Ž  â D	       À â Gy  Ðíÿÿÿÿƒ ê Ay        € ê ‰<      ( ‰8      h y
      %   ¢ !r<       â !r8       È/ ‰       " $x
   7Ž ÆO ‰      d z  p  p`ð È x 7   pDp  â !r         â !r       È/ ‰  €    ( ‰ €    b !r        â !r       È/ ‰ @    ( ‰ @    b !r       â !r       È/ ‰      h  ‰	      ¢  M	        € ê y       &   â $x77   ÿ Ž â !r       â/ $z6 s  Ž â !r		       âO $tÿ   ÿ Ž â x 7   p"ò Ä r 7ÿ   p"ð Ö $ÿÿ   	 Ž ä $z  v  Ž ÈŸ %v j   Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      "   $ z  w  p`ð Ú M	        € ê y      &   " $t9ÿ   ÿ Ž â ¹z  F   
   â $z8 x  Ž È %v88 \  9 Ž Ì �y88    ™ b ¹z  l      â Ey  -    € â ™x?    â y6      !   b $r
ÿÿ   ÿ Ž ä $rÿÿ   ÿ Ž ä $r4ÿÿ   ÿ Ž ä $r2ÿÿ   ÿ Ž Ä $r0ÿÿ   ÿ Ž ä $r.ÿÿ   ÿ Ž â x)ÿ   6  È/ | )   p`ð Ú G	  P,    € ê xt6   ÿÀŽ â$rÿÿ   ÿ Ž â x,6   ÿÀŽ â‚| ?       â x6   ÿÀŽ â 'z m   Ž â xt   ÿ   â y*      %   " x 6   ÿÀŽ â $x,   
Ž â zÿ n    â %x�   ÿ Ž â Ey `+    € â xtÿ   t  Ä r     ÿð� â $r	ÿÿ   
Ž â |   ÿàÿ â $x#6   ÿ Ž âx6   ÿ   â $z		 o  Ž â x7    ÿàÿ â $v(ÿ X  ÿ Ž â xÿ      â %x    Ž â xÿ   7  Ä x/    ÿàÿ â $z r  ÿŽ â |   ÿàÿ â $z		 u  ÿŽ â r7   ÿ� â$x**   ÿ Ž â r:7   ÿ� â %x$    Ž â xøÿÿÿÿÀŽ ä x+    ÿàÿ â $x6   ÿ Ž â x##   ÿÀŽ â %x	$    Ž â x
ÿ   +  Ä x   ÿÀŽ â $x97   
Ž â r    ÿ� â$v'ÿ Y  ÿ Ž â r    ÿ� â $v$ÿ `  ÿ Ž â zJ Z  à‘ â$v%ÿ a  ÿ Ž â zH Z  ÿàõ â $r4ÿÿ   ÿ Ž â x5øÿÿÿÿÀŽ Ä x3üÿÿÿÿÀŽ ä zK [  ÿ$~  â$x5    5
Ž â zI [  ÿä â $z* q  )Ž â x3    3èÿ ä x ÿ   /  â $z&8 t  Ž âr=
+   ÿ� Ä r; /   ÿ� â%x€&n   ÿŽ â r< /   ÿ� ä x ;øÿÿÿÿÀŽ ä x:üÿÿÿÿÀŽ ä r>
+   ÿ� â $x1/    
Ž â r"€#   ÿàó Ä x
=øÿÿÿÿÀŽ ä x77   èÿ â $rÿÿ   �Ž  â x<üÿÿÿÿÀŽ â $x-+   

Ž â x >üÿÿÿÿÀŽ ä x!ÿ   5  ä r 5€   ÿàõ Ä xÿ   3  ä r3€   ÿà÷ â$x�   ! â xÿ   9  ä r9€   ÿàó â $x�   Ž â x//   èÿ ä x++    èÿ â $x�   Ž  â xÿ     Ä xÿ     ä xÿ   7  â $x   ÿ Ž â r7€   ÿàõ â$x   ÿ Ž â xÿ   1  ä r1€   ÿà÷ â $x�    â xÿ   /  Ä r/€   ÿàù â$x�   Ž â xÿ   -  ä r-€   ÿàû â$x�    â xÿ   +  ä r+€   ÿàý â $x
�   Ž â xÿ     Ä xÿ   :  â $x	�    â xÿ   ;  â %x€,   € Ž â xÿ   <  ä xÿ   =  â $x   ÿ Ž â x ÿ   >  â $x   ÿ Ž â xJJL   ÿàó â $x   ÿ Ž â z ÿ `  pPð â $x   ÿ Ž Ä $x   ÿ Ž â z ÿ a   Sð â $x     ÿ Ž ä $rKÿÿ   KŽ  ä rB("   ÿàõ ä r>€(   ÿàó Æ $xC'    ä $x?�   'Ž  Æ �yxB    • ¨  �yuB   • ¨  �yv>"   • è �yw>    • â x;ÿ   6  Ä r<(    ÿàó ä xq*   ÿàÿ â $z:8 t  ;Ž â r@(   ÿàõ â$x='   Ž  ä $zqq q  :Ž â r:(   ÿàó â$xA'    âr>(   ÿàõ â/�yW<`   ‘ $ $x;'   Ž  ârB(   ÿàó â�yV@`   ‘ b $x?'    â rF(   ÿàõ Ä $xC'   Ž  â rD(   ÿàó â%x~qn   ÿŽ â r<(   ÿà÷ â/ �yX>`   ‘ $ $xG'    ä$x='   	Ž â�yY:`   ‘ " $xE'   
Ž  Æ �y[B`   ‘ ( �yZF`   ‘ " r;~#   ÿàó Æ �y_<`   ‘ " r@(;   ÿàõ â $r:ÿÿ   Ž  ä �y]D`   ‘ " $r<ÿÿ   J Ž ä $r=ÿÿ   K Ž ä $xA'   : Æ �yB<¸ÿÿ ™ "/ %x>,   ~ Ž Æ �yC<Üÿÿ ™ ( �y\@    • " rJ>(   ÿàó Æ �ya@   • $ $xK?   'Ž  â rG3~   ÿàõ â�yD<    ™ " rE5~   ÿàó Æ �y`J"   • " $r@ÿÿ   H Ž ä $rAÿÿ   I Ž â �ycJ    • " $xH    Æ �y^@    • " rP(G   ÿàõ â$xF   !Ž  â rR(E   ÿàó â�yb@$   • $ $xQ'   H â rK7~   ÿàõ â/$xS'   FŽ  â rI9~   ÿàó â�yl@l   • $ $xL    ârj(K   ÿàõ â$xJ   Ž  â rT(I   ÿàó â�ydP`   ‘ " rO/~   ÿàù â$xk'   L â rM1~   ÿà÷ â $xU'   JŽ  â rn(O   ÿàõ â�yeR`   ‘ " rr(M   ÿàó â $xN   Ž Ä $xP    â �ygT`   ‘ " $xs'   NŽ  ä$xo'   P â �yfj`   ‘ " rS+~   ÿàõ ä/rQ-~   ÿàó â �yhn`   ‘ $ $xT    ârz(S   ÿàõ â$xR   Ž  â r|(Q   ÿàó â �yir`   ‘ $ $x{'   T Ä $x}'   RŽ  â �yU<$   ™ ( �ymz`   ‘ ( �yk|`   ‘ " $rnÿÿ   ( Ž Ä/ $roÿÿ   ' Ž â �yj@H   • & %xr&n   nŽ È %xpqn   nŽ ä �ynrl   • ( �yopl   • " $tyÿÿ Ž â xuxT  u    ÈO ruÿu   ÿ3Ž ä xpwT  v    È� xvÿ   p  äruÿt   u  ä xtvÿÀŽ ä xvuÿÀŽ äxwt€€€€ÿüŽ ä xsÿ   p  â $x{u   ÿ Ž Ä $xqw   v
Ž â xssÿÀŽ ä xvÿ   u  â$xwu   ÿ Ž â xxtuøŽ ä xus€€€€ÿüŽ ä xvvÿÀŽ ä xrÿ   p  Ä xzp€€€€yìŽ ä x{{ÿÀŽ â $xuu   v
Ž â xrrÿÀŽ Æ $xyz   {
Ž â x|r€€€€ÿüŽ ä xwwÿÀŽ ä rvsv   uXŽ ä rxxq   t`Ž ä xz{pøŽ äx}ypxŽ â $x{|   w
Ž â xvv˜º  ÿ    Ä xtt˜º  ÿ    ä xxx˜º  ÿ    ä xss˜º  ÿ    ä rz}z   ÿÀŽ ä xppÿÀŽ ä rVÿ   V  Äxuv€€€€uŽ ä rWÿ   W  äxqx€€€€qŽ äxtxt`Ž ä xvvs`Ž ä rwrw   {XŽ Ä xzz˜º  ÿ    ä xsp˜º  ÿ    â $xpV   ÿ Ž â xWW   ÿÀŽ ä rXÿ   X  ä rqtq   ÿüŽ ä xtw˜º  ÿ    Ä xyz€€€€yŽ ä rYÿ   Y  ä xwr˜º  ÿ    ä xzzs`Ž ä xWp0   WâŽ â $xpX   ÿ Ž â r[ÿ   [  Ä x{t€€€€{Ž ärZÿ   Z  ä xYY   ÿÀŽ ä xttw`Ž ä ryzy   ÿüŽ ä xX[   ÿÀŽ â $xsZ   ÿ Ž â r_ÿ    _  Ä x[p0   YâŽ â &rZyB   ÿ   â rtt{   ÿüŽ ä xYWàÿÿÿÿàÿ ä r]ÿ   ]  â $xp_   ÿ Ž â x[[àÿÿÿÿàÿ â &rttC   ÿ   â x]]   ÿÀŽ â $rYYZ   ÿŽ â xZ6   ÿÀŽ Ä x\\T  a    â $r[[t   ÿŽ â xWp0   ]âŽ ä xtÿ   Z  ä r]ÿ\   ÿ3Ž È r]ÿt   ]  â $t_ÿÿ Ž â ErY Y       â 0rpÿ^    A   â x`cT  `    â $xZ]   ÿ Ž â xXs0   XâŽ â 0rrÿb    A   â Er[ [       â #rYpY   ÿ   â x\`€€€€_ìŽ Ä x_ZÿÀŽ â &rqqD   ÿ   â x^ÿ   `  äxXXàÿÿÿÿàÿ â #rYr[   Y   â xaÿ   `  â $x[\   _
Ž â x^^ÿÀŽ â $rXXq   ÿŽ â xaaÿÀŽ â $xq]   ÿ Ž â xs^€€€€ÿüŽ Ä x\]ÿÀŽ ä xZ_`øŽ äxc[`xŽ ä xba€€€€ÿüŽ ä xqqÿÀŽ â $x_s   \
Ž â rZcZ   ÿÀŽ Ä xc^]øŽ â $xbb   q
Ž â x\ÿ   `  ä x]ÿ   ]  ä rsc_   ^`Ž ä x\\ÿÀŽ ä xc]ÿÀŽ Ä rVvu   ÿüŽ ä r]aq   bXŽ ä xus˜º  ÿ    ä xv\€€€€ÿüŽ ä x``ÿÀŽ ä xs]˜º  ÿ    Ä rdÿ   d  ä x]u€€€€_Ž ä xqa˜º  ÿ    â $xav   c
Ž â xZZ˜º  ÿ    ä x_`˜º  ÿ    ä reÿ   e  â $xdd   ÿ Ž â rfÿ   f  Ä x^^˜º  ÿ    ä x[Z€€€€[Ž äx`Z_`Ž ä rgÿ   g  ä xee   ÿÀŽ â $xff   ÿ Ž â rc\c   aXŽ Ä rhÿ   h  ä x^u^`Ž ä xbs€€€€bŽ äxqsq`Ž ä r[`[   ÿüŽ ä riÿ   i  Ä xgg   ÿÀŽ ä xed0   eâŽ â $xhh   ÿ Ž â xvc˜º  ÿ    ä rmÿ    m  ä xc\˜º  ÿ    ä r]^]   ÿüŽ â &r^[B   ÿ   â r\qb   ÿüŽ Ä xii   ÿÀŽ ä rkÿ   k  ä xgf0   gâŽ ä xeeàÿÿÿÿàÿ â $xbm   ÿ Ž â xav€€€€aŽ â &r\\C   ÿ   â xvvc`Ž â $r^e^   ÿŽ â xih0   iâŽ Ä xkk   ÿÀŽ ä xggàÿÿÿÿàÿ â &r`]D   ÿ   â rZva   ÿüŽ ä xiiàÿÿÿÿàÿ â $r\g\   ÿŽ â xkb0   kâŽ ä Er[ ^       â $r`i`   ÿŽ â xWWàÿÿÿÿàÿ â &rVVU   ÿ   âxkkàÿÿÿÿàÿ â &rZZU   ÿ   â Er\ \       â #r[p[   ÿ   â $rVWV   ÿŽ â ErX X       â $rZkZ   ÿŽ â Er` `       â 0rjÿj    A   â #r[r\   [   â 0rlÿl    A   â ErV V       Ä ErZ Z       â #rYjX   Y   â#r[j`   [   â x))   ÿàÿ â 0rmÿn    A   â #rklV   Y   â#rilZ   [   â | )   p`ò â 0roÿo    A   â G‰  �    € ê xWÿ   6  Ê $zW* q  WŽ È $z&8 t  WŽ È %xV&n   ÿŽ Ê rX$#   Vàµ â%x€,   V Ž Æ rY%ÿ   Wd~ ä rf€$   ÿàõ Æ �yhX     ¨  �yqX    ¢  rd$V   5àÇ â$xg�   % â rb$V   3àë äre%W   !„þ ärc%W   Äþ â �ynf"    è �ysf     è �yed`    h	 �yvb`    b r`$V   9àµ Ä r^$V   7àÙ ära%W   d~ är\$V   1àµ är_%W   ¤~ ärZ$V   /àÙ â �ya``    b r]%W   d~ Ä rX$V   -à§ â�y^^`    b r[%W   ¤~ ärV$V   +àÙ â �y]\`    b rY%W   Dþ ärW%W   ¤~ â �yZZ`    h �yYX`    h �yWV`    b %xb&n   $Ž Ì �ybbl    b xd6   ÿÀŽ Èxgÿ   d  â/ $t_ÿÿ Ž â xhhT  q    ÈO rhÿh   ÿ3Ž È r\ÿg   h  ä xnsT  n    Æ� $xV\   ÿ Ž â reÿ   e  äx`n€€€€_ìŽ ä rvÿ   v  ä xcVÿÀŽ ä xXe   ÿÀŽ â $xev   ÿ Ž â x[ÿ   n  â$x_`   c
Ž â xccnøŽ Ä x[[ÿÀŽ ä x`_nxŽ ä xXe0   XâŽ ä xeÿ   n  â $xf\   ÿ Ž âxV\ÿÀŽ ä xg[€€€€ÿüŽ Ä rd`c   ÿÀŽ ä xceÿÀŽ â $xVg   V
Ž â x`nÿÀŽ ä xgfÿÀŽ ä xhc€€€€ÿüŽ ä xnÿ   n  Ä xfd˜º  ÿ    â $xdh   g
Ž â xe`˜º  ÿ    ä xqÿ   \  ä x`nÿÀŽ ä x_f€€€€_Ž ä xqqÿÀŽ Ä xh`€€€€ÿüŽ ä xffe`Ž ä rgcg   dXŽ ä xe[\øŽ ä x\c˜º  ÿ    â $xch   q
Ž â xgg˜º  ÿ    Ä rheV   [`Ž ä r^ÿ   ^  ä raÿ   a  ä xdg€€€€dŽ äxeg\`Ž â $x^^   ÿ Ž â xgh˜º  ÿ    Ä x\[˜º  ÿ    ä rZÿ   Z  ä rq`q   cXŽ ä r_f_   ÿüŽ ä r]ÿ   ]  ä xaa   ÿÀŽ â $xZZ   ÿ Ž â x[g€€€€VŽ Ä x\g\`Ž ä rWÿ    W  ä xfq˜º  ÿ    â &r__B   ÿ   â xq`˜º  ÿ    ä xXXàÿÿÿÿàÿ ä rded   ÿüŽ Ä x]]   ÿÀŽ ä rYÿ   Y  ä xa^0   aâŽ ä r[\[   ÿüŽ â $x\W   ÿ Ž â xVf€€€€cŽ â $rXX_   ÿŽ â xqfq`Ž â &rddC   ÿ   â x]Z0   ]âŽ Ä xYY   ÿÀŽ ä xaaàÿÿÿÿàÿ â &rZ[D   ÿ   â rVqV   ÿüŽ ä x]]àÿÿÿÿàÿ â $raad   ÿŽ â xY\0   YâŽ ä ErW X       â $rZ]Z   ÿŽ â xYYàÿÿÿÿàÿ â &rVVU   ÿ   â Era a       â #rWpW   ÿ   â ErZ Z       â $rVYV   ÿŽ Ä #rWra   W   Æ ErV V       â #rYjZ   W   â 0rWÿb    A   Æ #rVlV   Y   È #r2WV   2   â #r.mk   .   â #r0oi   0   â G‰  0    € ê rV$;   ÿà÷ ä r>>$   ÿàõ Æ $xW%   :Ž ä $x??   % Æ �yZV     ¨  �y[V    ¨  �y\>"    è �y]>     â rX$G   ÿà÷ Ä r:$E   ÿàõ Æ $xY%   HŽ ârH$K   ÿà÷ â$x;%   F ârF$I   ÿàõ ä�yXX`    " $xI%   LŽ âr>$O   ÿà÷ â/$xG%   J ârJ$M   ÿàõ â �yE:`    d $x?%   PŽ Ä $xK%   N â rL$Q   ÿàõ â�yHH`    " r:$S   ÿà÷ Æ/ �yWF`    " $xM%   R ä$x;%   TŽ â �y>>`    ( �yKJ`    ( �y;:`    ( �yML`    " rF~$   ÿàõ Ê $xG   % Ê �yFFl    " $t?ÿÿ Ž â/ xZZT  [    ÈO rIÿZ   ÿ3Ž ä x\]T  \    ä� rIÿt   I  ä xJÿ   \  Æ $x:I   ÿ Ž âxPJÿÀŽ ä xLIÿÀŽ ä xOP€€€€ÿüŽ ä xJ\€€€€?ìŽ ä xGÿ   \  ä x?:ÿÀŽ â $xQO   L
Ž â xLGÿÀŽ Ä xRPIøŽ â $xGJ   ?
Ž â x??\øŽ â $xJI   ÿ Ž â rRRQ   P`Ž ä x:G\xŽ äxSÿ   \  ä r::?   ÿÀŽ Ä xRR˜º  ÿ    ä xNL€€€€ÿüŽ ä xOJÿÀŽ ä x?P˜º  ÿ    ä xSSÿÀŽ ä xIÿ   I  Ä xQR€€€€QŽ â $xNN   O
Ž â xRR?`Ž ä x?S€€€€ÿüŽ ä xJIÿÀŽ ä x\\ÿÀŽ ä rXÿ   X  â$xI?   J
Ž â rOLO   NXŽ Ä x::˜º  ÿ    ä x?\˜º  ÿ    ä rEÿ   E  â$xXX   ÿ Ž â rHÿ   H  ä xOO˜º  ÿ    ä xG:€€€€GŽ Ä xLL˜º  ÿ    ä x::?`Ž ä rWÿ   W  ä xEE   ÿÀŽ â $xHH   ÿ Ž â r>ÿ   >  ä xNO€€€€NŽ Ä rJSJ   IXŽ ä xOOL`Ž ä rG:G   ÿüŽ ä xWW   ÿÀŽ ä rKÿ   K  ä xEX0   EâŽ â $x>>   ÿ Ž â r;ÿ    ;  Ä xJJ˜º  ÿ    â &rBGB   ÿ   â xSS˜º  ÿ    ä rNON   ÿüŽ ä xWH0   WâŽ ä xKK   ÿÀŽ ä rMÿ   M  Ä xEEàÿÿÿÿàÿ â $x:;   ÿ Ž â xIJ€€€€IŽ â &rNNC   ÿ   â rQRQ   ÿüŽ â $rBEB   ÿŽ â xJJS`Ž ä xK>0   KâŽ ä xMM   ÿÀŽ Ä xWWàÿÿÿÿàÿ â &rDQD   ÿ   â rJJI   ÿüŽ ä xM:0   MâŽ â $rNWN   ÿŽ â xKKàÿÿÿÿàÿ ä Er; B       â &rJJU   ÿ   â xMMàÿÿÿÿàÿ â $rDKD   ÿŽ â ErN N       â #r;p;   ÿ   Ä $rJMJ   ÿŽ â ErD D       â #r;rN   ;   Æ ErJ J       â #r?jD   ;   â 0r;ÿF    A   Æ #r?lJ   ?   È #r4;?   4   â x((Ü   ÿàõ ä x$$Ü   ÿàû ä xJ<@  ÿà÷ â $r'ÿÿ   ' â xH@@  ÿàù â $r%ÿÿ   %Ž ä $rKÿÿ   =Ž ä $rIÿÿ   A â D       À â Gy  `Ûÿÿÿÿƒ ê Ay       € ê $r
ÿÿ   2 Ž ä $rÿÿ   4 Ž Ä Ay        € ê z ÿ `  pPð È z ÿ a   Sð Ú G	  P    € ê ‰.      ( ‰0      b !r.       â !r0       È/ ‰       ( ‰      b !r         â !r       È/ ‰  €    ( ‰ €    b !r        â !r       È/ ‰ @    ( ‰ @    b !r       â !r       È/ ‰      ( ‰	      b !r       â !r		       â/ Gy  €    € ê ‰2      ( ‰.      h ‰0      ¨ ‰	4      â !r2       â !r.       È/ ‰      " !r0       ÆO ‰       b !r
	4       Æ� ‰      ¨ ‰	
      â !r       â !r         â/ !r       ÈO ‰  €    " !r
	       Æ� ‰ €    h ‰	 €    ¨ ‰ €    â !r        â !r       È/ ‰ @    " !r
	       ÆO ‰ @    b !r       Æ� ‰
 @    ¨ ‰	 @    â !r       â !r       â/ !r 
       ÆO ‰      " !r	       Æ� ‰	      h ‰       ¨ ‰      â !r       â !r
	       â/ !r	        âO !r       Ä� y      %   $ $x   6Ž Ê z  p  p`ð È x 6   pDp  Ú M	        € ê z ÿ ^  pPð â $z88 v  Ž Æz ÿ _   Sð Ú $ÿ   ÿ Ž È %8 ^   Ž Ì �	     ¢ $x 6   ÿ Ž Ê r  ÿ   p"ò äx     p"ô äx     p"ö äx     p"ø Î $ ÿÿ    Ž â z ÿ `  pPò â $" ÿÿ   	 Ž ä $2 ÿÿ   
 Ž â z ÿ a  Sò â $B ÿÿ    Ž È !         ÐO G™       € ê z ÿ b  pPð È z ÿ c   Sð Ú $ÿ   ÿ Ž È %8 b   Ž Ì �	     ¢ $x66   ÿ Ž ä $vÿ h  ÿ Ž Æ x 6øÿÿÿp"ò äx 6üÿÿÿp"ô är 6ÿ   p"ö äx 6   p"ø Î $ÿÿ    Ž â x    pBò â $"ÿÿ   	 Ž ä $2ÿÿ   
 Ž ä $Bÿÿ    Ž È !       ÈO G  P    € ê x    pRð Ú G‰  �     € ê x    pRð Ú G	  ð    € ê  x;ª¸¿  A  Ì s        $ !t  €?    Ì s        $  r     A  È  r       A  â Gy  P    € ê  x'7=  A  â x*BL?  A  Æ #t  €?   â  x   ?  A  Æ  r     A  Ì s     $   $ !t  €?    È  r     A  È  r       A  â Gy  °    € ê x    pRð Ú G‰  à     € ê x    pRð Ú G	        € ê 	z i    � ä 	z   i    � Æ  x;ª¸¿  A  â 	z   i €  � Ê s        $ !t  €?    Ì s        $  r     A  È  r       A  â Gy  Ð     € ê  r       A  â Gy  °     € ê 	x  à@  � ä 	x    à@  � Æ  x#ÛÙ?  A  â 	x    àÀ  � Æ  x;ª¸¿  A  ä !t   €?    È s        $ !t  €?    Ì s        $  r      A  È  r       A  ä y      "   " $tÿ   ÿ Ž Æ y      &   b $z s  Ž È $z v  Ž È/ %v j   Ž Ê †y      â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y(      "   " ¹z  l      â Ey  `3    € â ™x?    â y)      !   b ¹z  F   
   â r+ ÿ       â x,      ÿ  â y/      %   ¢ r? ÿ       â x      ÿ  â x      ÿ  â x       ÿ  â x"      ÿ  â x$      ÿ  â x&      ÿ  â r (ÿ   pRð Ä r (ÿ   pRò är ()   ÿ(Ž È/ xÿ      ä x//   ÿ   äO r    ÿ � È x*ÿ     È | *   pbô Ú G)  ð1    € ê y      '   " xðÿÿÿÿÀŽ ä r ÿ       â y      &   b r   €ÿàÿ ä r ÿ       ä x ÿ     ä x>*   ÿ   Ä r    ÿ� ä z5  X      ä x9ÿ     ä z3  Y      ä x99   ÿ   ä r+ ÿ       â 'z  y   Ž â z<> q  ÿàÿ Ä xBÿ   9  â 'z s   Ž â/ xøÿÿÿÿÀŽ Æ $z1 }  ÿŽ â r=  €ÿàÿ â$z
 w  ÿŽ â zÿ z     â $z / p  *Ž â xÿ   =  ä zÿ t    â $z |  *Ž âr1
1   ÿàõ â $z	 |   Ž â r=   ÿ� â $z v  Ž âx /   ÿàÿ â $z	 v  	Ž â r   ÿò� Ä z  q      â $z  p  Ž â r7ÿÿ   ÿä â %x	T   ÿŽ â r.>   ÿŽ är6>   ÿŽ â$x4   >Ž âx ÿ   =  â $x0   >Ž âr@   ÿàõ â $x8   >Ž â x;=      â $x:   >Ž â xÿ     â %xT   ÿŽ â x==   ÿ   Ä r2   ÿä â %x
T   Ž È %x   Ž È %x   Ž È rh5   ÿàõ är5   ÿà÷ ä r5@   ÿàù ä ri3   ÿä är3   ÿäÿ ä r91   >àµ â �yhh   ™ ¢  r
5   ÿàû Ä r32   ÿä â �ym   ™ â x $       ä r B7   ÿd~ â �y}    ‘ b	 r3   ÿäÿ â %v Z   Ž ä �yy   ‘ ¤ $x $   ÿŽ Ä �y{    ‘ ¢  r=   ÿàõ ä r   ÿàÿ â �y„   ‘ æ  r;   ÿä â �yz$   • è  �y…   ™ è  �yr(   ™ è  �y     • è  �y‚   ‘ è �yjH   • è  �yn   ‘ è �ylL   ™ è �yi   ‘ è �yq   ‘ â  r91   <àµ Æ/ �yol   • â rB7   ÿd~ â %v Z   Ž ä �ykp   ™ ä $x$   ÿŽ â r=   ÿàõ È r   ÿàÿ È r;   ÿä â �yp    • è  �yg   ™ è  �yc(   ™ è  �yf$   • è  �yaL   ™ â  r91   .àÙ Æ�ybH   • â r5   ÿàõ Æ/ �y`p   ™ â r5   ÿà÷ ä rB7   ÿ¤~ â �ydl   • â r3   ÿä är	3   ÿäÿ â %v Z   Ž ä �y_P   ™ ä  $x$   ÿŽ Ä/ �y^P   ™ â r=   ÿàõ ä r   ÿàÿ È r;   ÿä â �y]    • è  �y[   ™ è  �yS(   ™ è  �y\$   • è �yUL   ™ â r91   4àµ Æ�yTH   • â rB7   ÿd~ ä r91   6àµ â/ %v Z   Ž â �yYp   ™ æ  $x$   ÿŽ â rB7   ÿd~ â �yVl   • æ r   ÿàÿ â $x$   ÿŽ â r91   0àë Ä r=   ÿàõ â �yW    • ä %v Z   Ž â r;   ÿä â �yX$   • è �ye   ™ è
 �yZ(   ™ è
 �yvL   ™ è
 �yNH   • â r91   8àÇ Ä rB7   ÿÄþ â�y€p   ™ ( �yMl   • " %v Z   Ž È $x$   ÿŽ â rB7   ÿ„þ ä/ r91   :àÇ È rB7   ÿ„þ ä r   ÿàÿ â %v Z   Ž â r=   ÿàû Æ %v Z   Ž â �yQ    • & $x$   ÿŽ â �yP$   • ( r   ÿàÿ â �yFH   • " r;   ÿäÿ Ä r   ÿàÿ â �yHl   • ( �yƒ   ™ ( �yu(   ™ ( �ysL   ™ ( �yp   ™ " r=   ÿàõ â$x$   ÿŽ â r=   ÿà÷ â �yR    • " r;   ÿä Æ �yO$   • ( �y�   ™ " r   ÿàÿ Æ �yt(   ™ " r;   ÿäÿ Æ �yK    • " r=   ÿàù Æ �yJ$   • ( �yCH   • ( �yDl   • ( �y~p   ™ " r;   ÿä Æ �yEH   • ( �yL   ™ ( �yGl   • ( �yL    • ( �yI$   • ( �yAH   • " x}ð   ÿÀŽ Æ�yl   • b! xÿ     Æ �y(   ™ " x{ð   ÿÀŽ ÆO �yL   ™ ¢ xyð   ÿÀŽ Æ �y   ™ ¢ xÿ   h  Æ �yp   ™ ¢ r   ÿ@Ž ä xÿÀŽ â �yw(   ™ ¢  xÿ     ä xÿ     ä/ xxhÿÀŽ â 0r ÿ     A   â� r|   ÿ@Ž â �y   ™ â  r   ÿ€Ž Ä xmÿÀŽ â &rˆr   ÿ   â xyy   ÿÀŽ â &r…   ÿ   â r   ÿ@Ž â 0rÿz    A   â r|||   ÿ€Ž â &r†x…   ÿ   âx}}   ÿÀŽ â &rŠ…   ÿ   â x{{   ÿÀŽ â $rˆˆy   ÿŽ â rz   ÿ€Ž â &r‹|…   ÿ   â Er        â $r††}   ÿŽ â �yL   ™ ¢  $rŠŠ{   ÿŽ â Er… ˆ       â &r‡zr   ÿ   â #rˆ    ÿ   â Er ‹       Ä Er †       ä Er‰ Š       â �yp   ™ ¢ Er‡ ‡       â #r�    ÿ   â#r†    ÿ   â#rŽ ‰   ÿ   â x„ð   ÿÀŽ â #r‡‡   ˆ   â x ‚ð   ÿÀŽ Ä xˆÿ   m  â#r…   †   â xŠÿ   m  ä x‰ÿ     ä x…‚   ÿÀŽ ä xÿ      ä x‚ˆÿÀŽ Ä xˆŠÿÀŽ ä rŠ‰‰   ÿ@Ž ä r   ÿ@Ž â 0r ÿj    A   â x†ÿ   h  ä rjŠŠ   ÿ€Ž â &r‰‚r   ÿ   â x„„   ÿÀŽ Ä rŠ   ÿ€Ž â &rrjr   ÿ   â x††ÿÀŽ â &r’ˆl   ÿ   â xn   ÿÀŽ â $r‰‰„   ÿŽ ä &rŠl   ÿ   â Er‹ r       â &rŒ†l   ÿ   â Er‰ ‰       â $r’’   ÿŽ â Err        â $rŒŒ…   ÿŽ â xnnð   ÿÀŽ â #r‰‰   Ž   âEr’ ’       â #r r   ‡   â #r‹   �   â x‡ið   ÿÀŽ Ä ErŒ Œ       ä x‹qð   ÿÀŽ ä xnÿ   n  â #rr ’   ‰   âx‰ÿ   h  ä x‡ÿ   ‡  â #r Œ      â x‹ÿ   ‹  Ä rnnn   ÿ@Ž â 0rŒÿo    A   â xhq   ÿÀŽ ä xq‰ÿÀŽ ä ro‡‡   ÿ@Ž ä r‡‹‹   ÿ@Ž ä rnnn   ÿ€Ž â &rŽqk   ÿ   â xmÿ   m  Ä xii   ÿÀŽ ä rooo   ÿ€Ž â &r‰nl   ÿ   â r‡‡‡   ÿ€Ž â $rŽŽi   ÿŽ â xmmÿÀŽ â &rlok   ÿ   â Er‰ ‰       â &r�‡k   ÿ   ä&r‹mk   ÿ   â Erk Ž       Ä Erl l       â #r ‰      â ErŽ �       â $r‹‹h   ÿŽ ä #r‰Œl      â0r ÿp    A   â #rlŒŽ      â &rŽg   ÿ   â Er‹ ‹       â &rpxg   ÿ   â #rkŒk      â $r{Ž   ÿŽ Ä &rg   ÿ   ä$rp}p   ÿŽ ä &r|g   ÿ   â #rrŒ‹   r   â &rŒc   ÿ   â Er        â &r‹zc   ÿ   â Er        ä Erg p       Ä Er        â $rŒyŒ   ÿŽ â Erp ‹       â 0rfÿf    A   â #r    ÿ   â#r‹    ÿ   â#rg g   ÿ   â#r    ÿ   â &r‚c   ÿ   â ErŒ Œ       â &r †a   ÿ   â #rfp      â $rp„   ÿŽ Ä $r …    ÿŽ ä &rŠa   ÿ   â #rgfŒ   g   â &rŒjc   ÿ   â Erc p       â 0rbÿb    A   â Er          â &rŽˆa   ÿ   â Erp        â &rana   ÿ   â ErŒ Œ       â &ro`   ÿ   â #rcfc   ‹   â #rgb    g   â#rbp      â $rŽŽ   ÿŽ â Era a       â &r q`   ÿ   Ä &r‹m`   ÿ   ä&rp‡`   ÿ   â Er`        â 0rdÿd    A   â #rfŒ      â $r i    ÿŽ â ErŽ Ž       â $r‹h‹   ÿŽ â Erp p       â #rba      â #rd`      â 0r`ÿ_  0 A   âEr          â 0raÿ^  0 A   âErf ‹       â #rcbŽ   c   â #rpdp      â 0rbÿ_    A   â  r`     A  â 0r_ÿ^    A   â  rla     A  â #rgd    g   â#rfdf   c   â  rpap     A  â  r ‰`     A  â #rrr_     â #rbg     â #r_f   p  â &rf[   ÿ   â#r kb      â &r^x[   ÿ   â !r&&       â !r$$       â $r{f   ÿŽ Ä &r[   ÿ   â!r' '       â $r^}^   ÿŽ ä &r|[   ÿ   ä 0r ÿ]    A   â Er        â &r]zS   ÿ   â Er        ä Er[ ^       Ä Er        â &rdS   ÿ   â Er^ ]       â 0r\ÿ\    A   â #r    ÿ   â#r]    ÿ   â#r[ [   ÿ   â#r    ÿ   â &r‚S   ÿ   ä $rdyd   ÿŽ Ä &r †U   ÿ   â #r\^      â $r^„   ÿŽ â Erd d       â $r …    ÿŽ ä &rcjS   ÿ   â ErS ^       â 0rTÿT    A   â Er          â #r[\d   [   â &rdˆU   ÿ   â #r]\S   ]   â &rŠU   ÿ   â #rST    [   â $rfd   ÿŽ Ä &r qY   ÿ   ä&r[mY   ÿ   â Erd c       â &rUnU   ÿ   â Er^        â $r i    ÿŽ â Erf f       â &roY   ÿ   ä $r[h[   ÿŽ â ErU U       â &rY‡Y   ÿ   â #r\d      â #rT^      â0rVÿV    A   â Er          â #r]Tf   ]   â Er\        Ä Er^ [       â #rTU      â Erd Y       â #rSV    S   â#rUV\      â#rTV^   ]   ä#rVVd      â &rde   ÿ   ä&r\xe   ÿ   Ä $r{d   ÿŽ ä &re   ÿ   ä $r\}\   ÿŽ â Er        â &rYzZ   ÿ   â Er        â 0r ÿW    A   â ErW \       â 0rXÿX    A   â Er\ Y       Ä #r    ÿ   â#rY    ÿ   â &r‚Z   ÿ   ä&r^Z   ÿ   ä &r|e   ÿ   â #rX\      â &r\ˆv   ÿ   ä $r[„   ÿŽ Ä &rZjZ   ÿ   â Er        â $r^y^   ÿŽ ä &rŠv   ÿ   ä $r]\   ÿŽ â Er\ Z       â 0rNÿN    A   â Er^ ^       â #rW W   ÿ   âEr[ [       â #r    ÿ   â ErZ        â &r †v   ÿ   â #rWX^   W   â#rYX[   Y   â#rX\      â #rZNZ      â $r …    ÿŽ Ä &rXq€   ÿ   ä	&ro€   ÿ   ä &rvnv   ÿ   â Er          â &rm€   ÿ   â Er        â $rXiX   ÿŽ ä 0rMÿM    A   â Er^ ]       â $rh   ÿŽ â Erv v       â &r€‡€   ÿ   â ErX X       â #r N    W   â#rM   Z   â#r^N^   Y   âErW        â #rvNv      â ErY €       â #rMX       â  rN`     A  â  rVaV     A  Ä #rvMY   v   â#rNb   N  â  rU`U     A  â #r^MW   ^   â  rav     A  â #r_T   V  â!r!N!       â &rNxƒ   ÿ   ä&rVƒ   ÿ   â #r bS   U  â #r_^     â !r""       â &rTu   ÿ   Ä $rN}N   ÿŽ ä $r{V   ÿŽ ä &rƒ   ÿ   â!r# #       â !r         â 0r ÿQ    A   â ErM N       â &r|ƒ   ÿ   â Er        â $rTyT   ÿŽ â Er        â &rQzu   ÿ   â Er        â 0rPÿP    A   â ErT T       â #rM M   ÿ   âErN Q       â #r    ÿ   â#rQ    ÿ   â &r‚u   ÿ   â #r    ÿ   â #rMPT   M   â#rPN      â &r †s   ÿ   Ä &rTˆs   ÿ   ä $rN„   ÿŽ ä $r …    ÿŽ ä $rUT   ÿŽ â ErS N       â 0rFÿF    A   â Er          ä ErV U       â #rQPS   Q   â &ruju   ÿ   Ä &rŠs   ÿ   â #rMF    M   â#rSFV   Q   â &r q   ÿ   â ErT u       â &rQm   ÿ   ä &rsns   ÿ   â ErN        â &ro   ÿ   ä $r i    ÿŽ â Ers s       â $rQhQ   ÿŽ Ä &r‡   ÿ   â #rPT      â #rFN      â0rHÿH    A   â Er          â #rFs      â ErN        ä ErP Q       ä ErT        â #rMH    M   â#rQHN      â#rFHP   S   â &rP�   ÿ   â#rTHT      â &rHx�   ÿ   Ä &rNt   ÿ   ä &r�   ÿ   ä$rH}H   ÿŽ ä $r{P   ÿŽ ä &r|�   ÿ   â Er        â 0r ÿR    A   Ä $rNyN   ÿŽ â ErS H       â &rRzt   ÿ   â Er        ä Er        ä ErP N       â #rN    ÿ   âErU R       â #rH S   ÿ   â#rR    ÿ   â#rV    ÿ   â &r‚t   ÿ   Ä &r †   ÿ   ä $r„   ÿŽ ä $r …    ÿŽ ä &rŠ   ÿ   ä 0rOÿO    A   â Er        â &rtjt   ÿ   â Er          â 0rEÿE    A   â Er        â #rHOP   H   â#rNOU   N   â&rPˆ   ÿ   âEr t       â &rn   ÿ   â #rRO   R   â #r E    H   â#rNE   N   â $rPP   ÿŽ Ä &rHq~   ÿ   ä&rm~   ÿ   ä&ro~   ÿ   âEr        â &r~‡~   ÿ   â #rVO   V   â $rh   ÿŽ â ErS P       â 0rGÿG    A   â ErO ~       â $rHiH   ÿŽ â #rVE   V   âEr        â #rRES   R   â Er        â #rVGO   V   âErH H       â #rG   N   â  rTaT     A  â #rRG   R   â  raV     A  â #rGH       â  rH`     A  â#r_F   T  â#r_R     â  rQ`Q     A  â &rFx   ÿ   ÄO &rN   ÿ   â #rHb   H  â!r       â !r       â #r bM   Q  â $rF}F   ÿŽ ä $r{N   ÿŽ ä &r   ÿ   ä&r|   ÿ   â !r        â !rH       â 0r ÿK    A   â Er F       â &rH   ÿ   â Er        Ä Er        ä Er        â &rz   ÿ   â #r    ÿ   â &rE‚   ÿ   â #r    ÿ   â$rHyH   ÿŽ â #r    ÿ   â#r    ÿ   â &r †   ÿ   â Er        â 0rJÿJ    A   Ä $rE„E   ÿŽ â ErH H       â $r …    ÿŽ â #rJ      â &rFˆ   ÿ   âErE E       â &rŠ   ÿ   â #rHJH      â 0rCÿC    A   â Er          â &rj   ÿ   Ä $rF   ÿŽ â #rFJE      â Er        â #rC    H   â ErN        â &r q   ÿ   â Er        â &rn   ÿ   ä #rNJN      â $r i    ÿŽ â #rC   F   â &rm   ÿ   â Er        â 0rDÿD    A   â Er          â &ro   ÿ   Ä $rh   ÿŽ â #rC      â#rCC   N   â Er        â #rD       â Er         â &rxx   ÿ   ä�&r   ÿ   â #rD      â#rD       â &rw   ÿ   Ä &r   ÿ   ä &r q   ÿ   ä $rx}x   ÿŽ ä $r{   ÿŽ ä &r‚w   ÿ   ä &rˆˆ   ÿ   Ä &r|   ÿ   â Er        â $ry   ÿŽ ä &rzzw   ÿ   ä $rii    ÿŽ ä $r„„   ÿŽ â Er        â &r††   ÿ   Ä $rˆˆ   ÿŽ â Er x       â 0r ÿL    A   â Er        â &rjjw   ÿ   ä &r‡   ÿ   ä &rmm   ÿ   ä &rŠŠ   ÿ   Ä &rnn   ÿ   â Er z       â &roo   ÿ   ä&r‡‡   ÿ   â Er        â $r†…†   ÿŽ â #r    ÿ   â0rIÿI    A   â Er„ „       â #r    ÿ   â#r    ÿ   âEr j       â #r     ÿ   â 0rAÿA    A   â Er Š       â #rI      âEr† †       â #rI      âEr ˆ       â #rI„      âErn n       â #rI       â x**   ÿàÿ â $rmhm   ÿŽ â Er        â #rA      â#rA†      â#rA      â0rÿ    A   âEr  o       â #rAAn      â Er ‡       Ä | *   pbô â #rD   C   â Eri i       â #r       âEr m       â #r   A   â  r`     A  â  ra     A  â#ri      â#r      â  r``     A  â  raa     A  â #r b     â#r_     â#rb   `  â #r_   a  â x55P  ÿà÷ â !r%r%       â !r? ?       â !r--       â !r,,       â !r++       â x..    ÿàÿ Ä x44    ÿàÿ ä x66    ÿàÿ ä x00    ÿàÿ ä x88    ÿàÿ ä x::    ÿàÿ ä x<<    ÿàÿ Ä x>>    ÿàÿ ä r3ÿ3   ÿäÿ â D)       À â Gy  pÑÿÿÿÿƒ ê Ay        € ê Ey  0    € â G‰      € ê r(()   ÿHŽ Ê ˆs (' øÿ H   è ˆs (%€øÿ H   è ˆs (& ùÿ H   è ˆs ($€ùÿ H   è ˆs (# úÿ H   è ˆs ("€úÿ H   è ˆs (! ûÿ H   è ˆs ( €ûÿ H   è ˆs ( üÿ H   è ˆs (€üÿ H   è ˆs ( ýÿ H   è ˆs (€ýÿ H   è ˆs (? þÿ H   è ˆs (-€þÿ H   è ˆs (, ÿÿ H   è ˆs (+€ÿÿ H   ä Ay        € ê {          ì M        € ê „y )     H   b z3  r      Ä z / r  p`ð â „y) €   H   ¢ r133   ÿŽ är )ÿ   p"p â „y	) €  H   â r733   ÿŽ Æ y      &   ( „y)    H   h „y)    H   ( y      '   ( „y)    H   ( „y) €  H   è „y) €  H   â !r  '       Æ/ „y)    H   â !r%%       Æ_ ‰       " !r"	"       Æ� ‰%      h ‰"      ¢ !r&&       Æ„y)    H   â !r!!       Æ „y) €  H   è „y)    H   â !r##       â!r        â $z  x  /Ž â „y) €  H   â !r$$       â !r
%       â/ $z  ~   Ž â ‰      " !r         â !r""       âO x//   ÿàÿ â ‰
      b r3    ÿ† â !r',       â  1   ÿàù â „y)    H   â rÿÿ   ÿäÿ Ä z j  ÿ† â„y) €  H   â ÿÿ   ÿä ä z k  � â ‰&      ¢ 
 j  ÿˆ Æ ‰!      " 
 k   â !r       â� !r       â ‰#      " !r       â 
  r  ÿàó â !r
       â/ ‰$      b !r       â (ÿÿ   ÿäÿ  Ä ‰	 €    â 
 j  ÿ„ â !r%       ä ‰ €    b 
 k  ( â !r&&       äO ‰       ¢  7   ÿà÷ â !r!!       Ä„y4) €  H   â ÿÿ   ÿäÿ â !r??       â !r##       â 
 j  ÿ† â‰!      " !r--       â !r$$       â/ 
 k  � â ‰      b z / r  p`ò â !r	       â� r3    ÿ„ â‰      â $x	3   ÿŽ â !r       ârÿÿ   ÿä â ‰&      " x )   p"ò â !r         âO z
 j  ÿ„ â ‰*      ¢ r		    ÿàù Ä z k   â ‰,%      b rÿÿ   ÿä ä z	 j  ÿˆ â ‰$      b !r!!       â 1   ÿàÿ ä ‰.?      " !r       â/ z		 k   Ä ‰0-      b !r(       â� !r44+       â 6    ÿà÷ â ‰2'      â !r&&       ä9ÿÿ   ÿäÿ â ‰#      " !r**       ÆO ‰ @    ¢ !r,%,       Æ‰"      b !r$$       Æ ‰ @    b !r.?.       â !r0-0       ä/ ‰! €    â !r2'2       Æ� ‰%       " !r##       Æ‰'      b 6 j  ÿ† â !r       äO ‰)(      ¢ 3   ÿàÿ â!r""       ä‰& €    â r33  €ÿŽ â !r       Ä ‰-*      "     ÿàù ä8    ÿàõ â ‰/,      b ÿÿ   ÿä ä 
 j  ÿˆ â ‰$ €    b !r  %       â ÿÿ   ÿä Ä ‰1.      " 
 k   â !r!       â !r'       â/ ‰30      b !r(()       âO 6 k  9� ä ‰52      ¢ !r&&       â� x!        Ä ‰4      â !r-*-       â8 j  ÿ„ ä ‰# €    " !r/,/       â8 k   ä ‰      b !r$$       Æ ‰" €    b !r1.1       â !r303       ä/ ‰'  €    " !r525       ÆO ‰) €    b !r44       Æ� ‰+( €    ¢ !r##       Æ‰& @    â %v  j  ! Ž â !r9       ä‰*- €    " 7   ÿàÿ â !r""       ä ‰,/ €    b     ÿàõ Æ ‰$ @    b !r' '       â 6ÿÿ   ÿä ä ‰.1 €    " !r))       â/  j  ÿ„ ä‰03 €    b !r+(+       âO  k  6 Ä ‰25 €    ¢ !r&       Æ� ‰74      â !r*-*       Æ‰      " !r,/,       Æ‰# @    b !r$       Æ ‰!" @    b !r.1.       Æ ‰% @    " !r030       Æ/ ‰ ' @    b !r252       âO !r747       ä� ‰$) @    ¢ !r;       Æ‰&+ @    â !r#       Æ‰      " !r!"!       Æ ‰#* @    b !r%%       Æ ‰-, @    " !r '        Æ/ ‰      h ‰/. @    b !r$)$       ÆO ‰10 @    ¢ !r&+&       Æ� ‰"7 €    â !r5       Æ‰32 @    " !r#*#       Æ‰      b !r-,-       Æ ‰!      " !r'       Æ/ ‰%      b !r/./       â !r101       äO †	 9    â !r"7"       Æ� † ;   â !r323       Æ‰       ¢ !r       Æ‰$      â !r!!       Æ ‰&      " !r%%       Æ/ †	 5    è ‰#      h ‰-      ( † '    â !r        ÆO ‰/      ¢ !r$       Æ� ‰1      â !r&       Æ ‰" @    ( ‰3      b !r##       Æ/ †	     â !r--       Æ† !   è †	 %    â !r//       âO !r11       ä� †     â !r""       Æ †	 
    â !r33       Æ† 
   è †	 #    è † -    è †	 /    è † 1   è ‰"      ( †	 3    â M™        € ê x   ÿàÿ â !r"       Æ? r    ÿàñ È r ÿÿ   ÿä  ä z j  ÿ€ È z k     Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y       "   " ¹z  l      â Ey  `.    € â ™x?    â y      !   b ¹z  F   
   â x      ÿ  â r ÿ       â y      %   ¢ r! ÿ       â x      ÿ  â r
 ÿ       â x      ÿ  â rU ÿ       Ä r ÿ       ä rf ÿ       ä rd ÿ       ä rp ÿ       ä r  ÿ   pRð är  ÿ   pRò Ä r    ÿ(Ž È/ xÿ     ä x   ÿ   äO r   ÿ � È xÿ     È |    pbô Ú G)  Ð,    € ê y      '   " xðÿÿÿÿÀŽ ä r ÿ       â y      &   b rA  €ÿàÿ ä r ÿ       ä xÿ   A  ä z  X      Ä rA   ÿ� ä r        ä x"øÿÿÿÿÀŽ ä xÿ     ä rA"  €ÿàÿ ä x    ÿ   â 'z y   Ž â xÿ     Ä x   ÿ   â 'z s   Ž â/ zÿ z    ä r#   ÿ� â $z p  Ž â zÿ t    â $z |  Ž â r"#"   ÿò� â $z |  Ž â x   ÿàÿ â $z }  ÿŽ â x#ÿ   "  â $z w  ÿŽ â z  q      â $z v  Ž â z  q  ÿàÿ â $z v  Ž â r   ÿàõ â $z p  Ž â xÿ     â %x~T   ÿŽ â rÿÿ   ÿä Ä r    ÿŽ â%x€T   ÿŽ â r    ÿŽ ä x     â %x<T   "Ž â r""~   ÿàõ ä z  Y      â $x    Ž âr ÿ       â $x    Ž âx   ÿ   â $x    Ž â x$ÿ     â %x>A   ~Ž â r#   ÿä Æ %x@A   €Ž È r;    àÇ ä r."   ÿàû ä x| $       ä r$   ÿ„þ ä r/   ÿäÿ â %v:; Z  | Ž â r6>   ÿàõ Æ $x$   ÿŽ â �yP.    ‘ ¢  r7?   ÿä ä r4:   ÿàõ â �yM.   ‘ â  r;;   ÿàÿ Æ �y66   ™ " r5;   ÿä ä r*<   ÿà÷ â�yƒ:    • h �y24   ™ h �y'4(   ™ b r0@   ÿàõ Ä r+=   ÿäÿ â�yF:$   • b r1A   ÿä Æ �yN*    ‘ h �y00   ™ h �yK*   ‘ h �yJ.   ‘ h �yB4L   ™ h �y8:H   • h �yI*   ‘ h �yG.   ‘ h �yE*   ‘ h �y34p   ™ h �ys:l   • b r)   àÇ Ä r,~   ÿàõ ä r$   ÿ„þ ä r-   ÿä â %v() Z  | Ž È $x$   ÿŽ â �y-,P   ™ b r.(   ÿàõ Æ r))   ÿàÿ È r/)   ÿä â �yq(    • h �y{.(   ™ h �y.   ™ h �yy($   • h �yt.L   ™ b r+   àµ Æ �ym.p   ™ b r$   ÿd~ Æ �yv(H   • h �yx(l   • b %v*+ Z  | Ž È $x$   ÿŽ â rR*   ÿàõ È r++   ÿàÿ È rS+   ÿä â �y*    • h �y.R   ™ h �yoR(   ™ h �yn*$   • h �yl*H   • h �yiRL   ™ b r5   àµ Æ �yh*l   • b r$   ÿd~ ä r(€   ÿàõ â �ykRp   ™ f r)�   ÿä â %v45 Z  | Ž È $x$   ÿŽ â �yg(P   ™ b r:4   ÿàõ Æ/ r55   ÿàÿ È r;5   ÿä â �y4    • h �y`:   ™ h �yc:(   ™ b r   àë Æ �yW:L   ™ b r%$   ÿÄþ Æ �y\4$   • b %v( Z  | Ž Æ �y[4H   • b $x%%$   ÿŽ Æ �yb4l   • d r))%   ÿàÿ ä �ye:p   ™ b r+   àÇ är}   àë â �yV(    • b r4(   ÿàõ Æ/ �yZ($   • b r5)   ÿä Æ �yY(H   • h �y^4   ™ b r#$   ÿ„þ â%v*+ Z  | Ž â r$   ÿÄþ â �y]4(   ™ d $x##$   ÿŽ Ä %v|} Z  | Ž â �yX4L   ™ d r++#   ÿàÿ â $x$   ÿŽ â �y_(l   • b r†*   ÿà÷ Æ r}}   ÿàÿ â �yT*$   • h �y*    • h �yQ*H   • h �yR*l   • h �ya4p   ™ b r‡+   ÿäÿ Ä x%6ÿÀŽ Æ�y,†   ™ b	 x#Pð   ÿÀŽ äO x)Mð   ÿÀŽ â� 0rƒÿƒ    A   âx#ÿ   #  â �yS†(   ™ ¢ x*ÿ   6  ä x)ÿ   )  â �y9†L   ™ â r###   ÿ@Ž Ä xPP   ÿÀŽ â �y+†p   ™ b	 x**ÿÀŽ ä r)))   ÿ@Ž â 0rFÿF    A   â r7##   ÿ€Ž â &r#%2   ÿ   â xMM   ÿÀŽ â &rH*'   ÿ   â r:))   ÿ€Ž â $r##P   ÿŽ â xONð   ÿÀŽ â &r472   ÿ   â/ xLKð   ÿÀŽ â $r;HM   ÿŽ â ErH #       â &r):'   ÿ   â Er4 4       â �y1|    • ¢  Er; ;       â #r#ƒH   ÿ   âEr) )       â #r4ƒ4   ÿ   â x/ÿ   0  â #r;F;   #   â x#0ÿÀŽ â #r)F)   4   â x4ÿ   6  â 0r8ÿ8    A   â xOÿ   O  â &rC#2   ÿ   â xNN   ÿÀŽ â �yD|$   • ¢  x44ÿÀŽ Ä rOOO   ÿ@Ž â �y(|H   • ¢  xLÿ   L  ä x5Jð   ÿÀŽ â$rHCN   ÿŽ â xJJ   ÿÀŽ â &rw4B   ÿ   â rOOO   ÿ€Ž â �y&|l   • ¢  rLLL   ÿ@Ž â $rzwJ   ÿŽ â x//ÿÀŽ â &r2O2   ÿ   â x5ÿ   5  Ä rLLL   ÿ€Ž ä ErH H       â &rj/'   ÿ   âxKK   ÿÀŽ â &r'L'   ÿ   â r555   ÿ@Ž â #ruƒH   ÿ   â Erz z       ä Er2 2       Ä xHIð   ÿÀŽ â $rCjK   ÿŽ â r555   ÿ€Ž â #rz8z   ;   â Er' '       â #r2ƒ2   ÿ   â x;ÿ   0  ä xHÿ   H  â &rj5B   ÿ   â ErC C       â #r'F'   2   â x;;ÿÀŽ Ä rHHH   ÿ@Ž ä x2Gð   ÿÀŽ â #rCFC   u   â Erj j       â &rF;B   ÿ   â x6ÿ   6  ä xII   ÿÀŽ ä rHHH   ÿ€Ž Ä x2ÿ   2  â #rj8j   )   â x66ÿÀŽ â $r)FI   ÿŽ â r222   ÿ@Ž â &rBHB   ÿ   â xFEð   ÿÀŽ ä x0ÿ   0  â &rr63   ÿ   â xGG   ÿÀŽ Ä r222   ÿ€Ž ä r„|   ÿàù ä Er) )       ä ErB B       ä xFÿ   F  ä x00ÿÀŽ â $rrrG   ÿŽ â r…}   ÿä â &r‚23   ÿ   â #r|8)   C   â rFFF   ÿ@Ž â #r'8B   '   â xEE   ÿÀŽ â &r803   ÿ   â Err r       â 0rsÿs    A   â Er‚ ‚       â �yB„   ™ ¢  rFFF   ÿ€Ž â $r88E   ÿŽ Ä �yC„(   ™ ¢  #rrsr   z   â#r‚s‚   j   â &rzF3   ÿ   â Erj 8       â �y3„p   ™ ¨  �y8„L   ™ ¢  0r)ÿ-    A   Ä 0r-ÿ-  0 A   â Erz z       È  ru‚-     A  â &r‚7   ÿ   Æ #rur)   u  â &rr*{   ÿ   â#rjsj   |   â#r'sz   '   â &rs#   ÿ   ä $r„Mr   ÿŽ â Er‚ ‚       â &rr:{   ÿ   ä 0rqÿq    A   Ä &rw%   ÿ   ä$r|Ns   ÿŽ ä &rO   ÿ   ä &rz/{   ÿ   âErr r       â 0ryÿy    A   â #r‚q‚   ÿ   â $rzKz   ÿŽ â Er| |       â &r{L{   ÿ   â Er        â $r†Pw   ÿŽ â#rwyr   ‚   â &rr;t   ÿ   â Erz z       â #r|q|   ÿ   âEr{ {       â #rq   ÿ   â &r}4t   ÿ   â Er† †       â &rs5t   ÿ   Ä $rrIr   ÿŽ â #rƒyz   |   â#r{y{      â $r}J}   ÿŽ â Er„ „       â &rz6m   ÿ   â Ers s       â &r2m   ÿ   â #r†q†   ÿ   â 0rvÿv    A   â Err r       â $rzGz   ÿŽ â Er} }       â 0rxÿx    A   â Er        â #rqy„   †   â #rsvs   w   â#rƒvr   ƒ   â &rr0m   ÿ   â Erz z       â &rtHt   ÿ   â #rqv}   q   â #rx   s   â $rrEr   ÿŽ Ä &rmFm   ÿ   â Ert t       â #rzxz   q   â  r-     A  â Err r       ä Erm m       â #rq)z     â #rtvt   {   â #rƒxr   ƒ   ä!rfqf       â #rxxm   t   â &rq#.   ÿ   Ä &rm%.   ÿ   ä$rsNq   ÿŽ ä &r*o   ÿ   ä $rmPm   ÿŽ ä &rq7.   ÿ   ä&r.O.   ÿ   â !rpup       â $rM   ÿŽ â Ery m       â &ru:o   ÿ   â Erq q       â 0rÿ    A   â Erm .       â &r./o   ÿ   â Erw        â 0rnÿn    A   â Eru u       â #rq   ÿ   â $r.K.   ÿŽ â Ers s       â &rrLo   ÿ   â #rnu      â 0rqÿl    A   Ä &rv;i   ÿ   âEro .       â &ru4i   ÿ   ä&rl5i   ÿ   â #ryy   ÿ   â#rss   ÿ   âErr r       â #rmm   ÿ   â $r.Iv   ÿŽ ä $rtJu   ÿŽ â Eru l       â #rnw   y   â#rlno   s   â#rnnr   m   â &roHi   ÿ   â Erm .       â 0riÿh    A   Ä &r.6k   ÿ   ä&rh2k   ÿ   âErt t       â &rr0k   ÿ   ä$r.G.   ÿŽ ä &rsFk   ÿ   â Erk h       â #rqu      â0rhÿg  0 A   âEr. .       â #rqt      â #rik      â $rrEr   ÿŽ â Ero o       â 0rgÿg    A   â  r''h     A  â #ri.      â  r-     A  â #rlqm   l   âErm r       â #rnqo   n   â Ers s       â #r'jg   '  â #r)     â #rlim   l   Ä #rsis   n   â !rd'd       â !r       â &r'%`   ÿ   ä&ri#`   ÿ   ä&r7`   ÿ   â  rshs     A  â &r*c   ÿ   ä $r'P'   ÿŽ Ä $r.Ni   ÿŽ â Er        â &rjO`   ÿ   ä &rk:c   ÿ   ä 0rÿ    A   â #rsgl   s  â $riM   ÿŽ â Er` '       ä Er .       Ä Er. j       ä Erl k       â #rk   ÿ   â &r/c   ÿ   â Erj i       â #ri`   ÿ   â#r'   ÿ   â#r.   ÿ   â &r.;W   ÿ   ä $rK   ÿŽ Ä 0r\ÿ\    A   ä &rcLc   ÿ   â Er        â $r`I.   ÿŽ â #rj\j   i   â &ri4W   ÿ   â #rl\l   k   âEr. c       â 0r[ÿ[    A   â Erk `       â &r5W   ÿ   â #r\   '   â $riJi   ÿŽ â #r\\.      â &rWHW   ÿ   â Erc        â #rk[k      â &r2e   ÿ   â Eri i       â &r6e   ÿ   Ä &r0e   ÿ   â ErW W       â &reFe   ÿ   â Er.        â &r%^   ÿ   â#ri[i   j   â &r'#^   ÿ   â #rc[c   l   â0rbÿb    A   â Er` e       â #r[[W   \   â $rP   ÿŽ Ä $rWN'   ÿŽ â #r[b`   [   â&r'7^   ÿ   â #r.b.   c   â &r`O^   ÿ   â Erc        ä Er '       â 0rVÿV    A   â Er' W       ä ErW `       â $rG   ÿŽ Ä $rE   ÿŽ â #rcVc   ÿ   â &r^:]   ÿ   â #rV   ÿ   â#reV'   ÿ   â#rWVW   ÿ   â &rV/]   ÿ   âEr        â 0rZÿZ    A   â Er        â &r\*]   ÿ   â Er^ ^       â &r`;X   ÿ   Ä $rVKV   ÿŽ â #rb   i   â#rb   k   â $r\M\   ÿŽ â #r^Z^      â &r]L]   ÿ   â ErV V       â $rbI`   ÿŽ ä &r4X   ÿ   â Er\ \       â 0rYÿY    A   â Er` ]       â $rJ   ÿŽ â Er] b       â #rVZV   e   â#r\Z\   c   â#r`Z`   W   â Er        â #rZY]   V   â &rV6a   ÿ   Ä &r'5X   ÿ   â #r\Y   \   â &rb0a   ÿ   ä $rVGV   ÿŽ â Er' '       â &r2a   ÿ   ä $rbEb   ÿŽ â ErV V       â 0r_ÿ_    A   â Er        â #r^Y'   ^   â Er' b       Ä #r\_V   \   â#r^_   ^   â  rVh[     A  â  r.-.     A  â #rZ_'   Z   â  r'-^     A  â #rg   V  â #r)   .  â&rXHX   ÿ   â #r\)\   '  â !r

       â &r%,   ÿ   â!rUU       â &r'#,   ÿ   â ErX X       â &raFa   ÿ   Ä &r*S   ÿ   äO &r7,   ÿ   ä$rP   ÿŽ ä $rVN'   ÿŽ ä &r,O,   ÿ   â Era a       â $r'M   ÿŽ â Er        â 0rÿ    A   â #rXYX   `   â &rW:S   ÿ   â Er.        Ä Er V       ä Er, ,       â #ra_a   X   â ErV '       â #r'   ÿ   â ErX W       â &r/S   ÿ   â #r.   ÿ   â#rW   ÿ   â#rY,   ÿ   â &r49   ÿ   Ä� $rK   ÿŽ ä &rSLS   ÿ   ä $rJ   ÿŽ ä 0rTÿT    A   â Er        â &r;9   ÿ   â Er, S       â 0rQÿQ    A   â Er        â #rVTV      â#rXTX   '   â$r'I   ÿŽ â #rT   W   â#rTT,   Y   â #r,Q   V   â 0rÿR    A   Ä &r59   ÿ   ä &rR0+   ÿ   âEr' '       â &r6+   ÿ   â Er        â $rRER   ÿŽ ä $rG   ÿŽ â #rQ'      â#r.Q   X   â Er' R       â &r2+   ÿ   â Er        â &r##B   ÿ   Ä &r%%B   ÿ   â #rR'      â&r**C   ÿ   â Er        â $rN#   ÿŽ â #r,   ,   â &rN/C   ÿ   ä $r%P%   ÿŽ ä &r7B   ÿ   Ä $rMM*   ÿŽ ä &r003   ÿ   ä &rBOB   ÿ   ä &r*;8   ÿ   â #r.   .   â $rNKN   ÿŽ â Er        â &r::C   ÿ   â Er        â $rEE0   ÿŽ â Er0 %       â 0r1ÿ1    A   Ä &r48   ÿ   ä &rCLC   ÿ   ä $rII*   ÿŽ â Er* B       â &r558   ÿ   â ErN N       â 0rDÿD    A   â #r%1   ÿ   â&r9H9   ÿ   â Er: :       â $rJJ   ÿŽ â #r10   ÿ   â&r8H8   ÿ   â #r#1   ÿ   âEr C       â #r11*   ÿ   â &r663   ÿ   â Er* I       â &r223   ÿ   â #r%DN   %   â 0r(ÿ(    A   â Er9 9       â &r+F+   ÿ   âErM M       â &r3F3   ÿ   â Er 5       â #r#D:   #   âEr8 8       â #r1D   1   â x   ÿàÿ â $r6G6   ÿŽ â #r%(*   %   â Er+ +       â 0r&ÿ&    A   â ErJ J       â #rTQ9   T   â Er2 2       â #rDM      â Er* 3       â #r#(   #   â#r1(8   1   â |    pbô Ä Er E       â #r+   T   â Er6 6       â #r(J      â #r2&2   #   â#r1&*   1   â#r&   %   â  r.-.     A  â rxhx     A  â raha     A  â #r6&6      â  rh     A  â r--2     A  â  r1h1     A  â #rxgƒ   x  â#ragZ   a  â #r,),   .  â #rgR     â #r6)6   -  â #r1g   1  â xP  ÿà÷ â !rx       â !rs       â !r	\	       â !ra       â !r!,!       â !r       â !r6       â !r1       â x    ÿàÿ Ä x    ÿàÿ ä x    ÿàÿ ä x    ÿàÿ ä x    ÿàÿ ä x    ÿàÿ ä x      ÿàÿ Ä rÿ   ÿäÿ â D)       À â Gy  �Öÿÿÿÿƒ ê Ay        € ê Ey      € â G‰  ð     € ê $x À  Ž Ê ˆs p ùÿ H   è ˆs d€ùÿ H   è ˆs f úÿ H   è ˆs €úÿ H   è ˆs  ûÿ H   è ˆs €ûÿ H   è ˆs U üÿ H   è ˆs 
€üÿ H   è ˆs 	 ýÿ H   è ˆs €ýÿ H   è ˆs ! þÿ H   è ˆs €þÿ H   è ˆs  ÿÿ H   è ˆs €ÿÿ H   ä Ay        € ê {          ì M        € ê „y     H   " z  r  p`ð Ä z3  r      â „y €   H   b r ÿ   p"p ä r133   ÿŽ â„y    H   ¢ r/33   ÿŽ Æ „y    H   è „y €  H   ( „y €  H   h „y    H   h „y €  H   h „y    H   è „y €  H   â !rp       Æ „y    H   â !rd       Æ/ ‰       " !rf       ÆO „y# €  H   b !r       Æ� „y    H   ¢ !r       Æ„y% €  H   â !r       Æy      '   â !rU       Æ ‰      â !r
       â !r        ä ‰      " !r		       Æ y       &   " !r       Æ ‰      b !r!!       Æ ‰      h ‰      b !r##       Æ/ ‰      b !r       ÆO ‰      ¢ !r       â !r%%       â $z   x  Ž â‰      " x   ÿàÿ ä $z  ~   Ž â� !r       â!r       â ‰ 	      â !r       â 
  r  ÿàõ Ä ‰"      " !r       â z  r  p`ò ä ‰ €    b !r       â/ ÿÿ   ÿä ä ‰$!      b !r       âO 
 j  ÿ„ Ä ‰      ¢ x    p"ò â !r       â 
 k   â ‰&#      "  1   ÿà÷ Æ ‰(      " !r 	        â� r	3    ÿ„ ä ‰*%      â !r""       Æ‰      " !r       â /   ÿàû ä ‰
      b !r$!$       â/ ,ÿÿ   ÿäÿ ä ‰      b !r       âO 
 j  ÿŠ Ä ‰      ¢ !r&#&       â 
 k  ,� ä ‰      " !r((       â r3    ÿˆ â !r*%*       â� ‰!      ä r.ÿÿ   ÿä â !r       â‰#       " z j  ÿˆ â !r
       Ä‰%"      b z k  . â !r       ä/ ‰ @    b ÿÿ   ÿäÿ â !r       äO ‰'$      ¢ 

 j  ÿ† â !r       Ä ‰ €    " rÿÿ   ÿä ä z	 j  ÿ„ â ‰)&      " !r!!       â� 
 k  � ä ‰+(      â !r# #       âz		 k   Ä ‰ €    " !r%"%       Æ‰ €    b !r       â/ 3   ÿàÿ ä ‰-*      b !r'$'       âO     ÿàõ ä ‰ €    ¢ !r       â .ÿÿ   ÿä Ä ‰ €    " !r)&)       â  j  ÿ„ ä‰ €    " !r+(+       â�  k  . ä ‰"! €    â !r       â$x3   ÿŽ â /   ÿàÿ â ‰$# €    " !r       Är    ÿàù â ‰&% €    b !r-*-       â/ /    ÿà÷ ä ‰(' €    b !r       âO rÿÿ   ÿä ä ‰ @    ¢ !r       â z j  ÿˆ Ä ‰*) €    " !r       â z k   ä ‰,+ €    " !r"!"       â� 0ÿÿ   ÿäÿ ä ‰ @    â !r$#$       â/ j  ÿ† Ä ‰ @    " !r&%&       â/ k  0� ä ‰.- €    b !r('(       Æ/ ‰ @    b !r       âO x        ä ‰ @    ¢ !r*)*       ä %v  j   Ž â ‰      " !r,+,       Æ ‰  @    " !r       Æ� ‰!" @    â !r       Æ‰#$ @    " !r.-.       Æ‰%& @    b !r       Æ/ ‰'( @    b !r       ÆO ‰      ¢ !r/       â 1   ÿàÿ ä ‰)* @    " !r         â     ÿàõ ä ‰+, @    " !r!"!       â� 0ÿÿ   ÿä Ä ‰      â !r#$#       â j  ÿ„ ä‰-. @    " !r%&%       â k  0 ä ‰      b !r'('       Æ/ ‰      b !r1       ÆO ‰      ¢ !r)*)       Æ ‰       " !r+,+       Æ ‰ !      " !r       Æ� ‰#      â !r-.-       Æ‰"%      " !r       Æ‰'      b !r       Æ/ †	 /    â !r       ÆO † 1   â !r        Æ ‰$)      " !r!!        Æ ‰+      b !r##       Æ� ‰-      ¢ !r%%"       Æ†	     â !r''       Æ†     è †	     è †    â !r))$       Æ †	     â !r++       Æ/ † !    â !r--       ÆO †	 #    è † %   è †	 
'    è † )    è †	 +    â M™        € ê †y -   â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      "   " ¹z  l      â Ey   )    € â ™x?    â y      !   b ¹z  F   
   â r$ ÿ       â x       ÿ  â r3 ÿ       â x"      ÿ  â r5 ÿ       â x      ÿ  â r6 ÿ       â x      ÿ  â r ÿ   pRð Ä r ÿ   pRò är    ÿ(Ž È/ xÿ      È r    ÿ � È xÿ     È |    pbô Ú G)  À'    € ê y      '   " xðÿÿÿÿÀŽ ä r ÿ       â y      %   b r   €ÿàÿ ä r ÿ       â y      &   ¢ x ÿ     ä x0   ÿ   Ä r    ÿ� ä z)  X      ä x-ÿ     ä z'  Y      ä x--   ÿ   ä r$ ÿ       â 'z y   Ž â z.0 q  ÿàÿ Ä x4ÿ   -  â $z% }  ÿŽ â x    ÿ   ä/ xøÿÿÿÿÀŽ â 'z s   Ž âO zÿ z    ä r1  €ÿàÿ â$z w  ÿŽ â zÿ t    â $z  p  Ž â x
ÿ   1  â $z |  Ž â x     ÿàÿ â $z	 |  Ž â r%%   ÿàõ â $z v  Ž â r
1   ÿ� â $z	 v  	Ž â r+ÿÿ   ÿä â $z  p  Ž â r   ÿò� â %x	T   ÿŽ â z  q      Ä x ÿ   1  â %xT   ÿŽ â r0   ÿŽ är*0   ÿŽ â$x&   0Ž âr2   ÿàõ â $x,   0Ž â xÿ     â %x   Ž â x/1      Ä x11   ÿ   â %x
T   Ž â r(	   ÿä Æ %x   Ž È r)   ÿàõ ärL)   ÿà÷ ä r'   ÿä är)2   ÿàõ ä rM'   ÿäÿ â �yC   ™ ¢  r-%   0àÇ Ä r
)   ÿàû â �yLL   ™ â r'(   ÿä ä r 4+   ÿ„þ ä r'   ÿäÿ â �y]    ‘ d	 $x $   ÿŽ ä �y_   ‘ ¨ �y     ‘ â  xE $       Æ �yH   ‘ ä  %v Z  E Ž ä �yI   ‘ æ r   ÿàÿ â �yM   ‘ â  r        är1   ÿàõ â �yJ   ‘ æ  r/   ÿä â �y^    • è �yW   ™ è  �yO(   ™ è  �yY$   • è �yBH   • è �yKL   ™ â  r-%   .àµ Æ �yF   ‘ â r4+   ÿd~ Æ �yNl   • â %v Z  E Ž Æ �yGp   ™ â  $x$   ÿŽ â r1   ÿàõ È r   ÿàÿ È r/   ÿä â �yP    • è  �yA   ™ è  �y7(   ™ è  �y@$   • è  �y;L   ™ è �y8p   ™ â r)   ÿà÷ Ä/ r-%   àÙ â �y9H   • â  r)   ÿàõ Æ �y:l   • â  r'   ÿäÿ ä r4+   ÿ¤~ är	'   ÿä â �yDP   ™ " %v Z  E Ž Æ �y<P   ™ " $x$   ÿŽ â r1   ÿàõ È r   ÿàÿ È r/   ÿä â �yX    • ( �yV   ™ ( �y?(   ™ " r-%   &àµ Æ �yU$   • " r4+   ÿd~ Æ/ �y=L   ™ " %v Z  E Ž Æ �yQp   ™ " $x$   ÿŽ â r1   ÿàõ ä �y>H   • $ r   ÿàÿ ä �y\l   • $ r/   ÿä Ä �yf    • ( �yc   ™ ( �yT(   ™ ( �yb$   • ( �yZL   ™ " r-%   *àµ Æ �ySH   • " r-%   ,àÙ ä r4+   ÿd~ â�y[l   • ( �yp   ™ "' r4+   ÿ¤~ â %v Z  E Ž È $x$   ÿŽ ä %v Z  E Ž â r1   ÿàõ Æ $x$   ÿŽ â r   ÿàÿ È r   ÿàÿ â �yd    • " r/   ÿä Æ �y`$   • ( �yh   ™ ( �yg(   ™ ( �yEH   • ( �yRl   • " r1   ÿà÷ Æ �ye    • ( �ya$   • ( �yL   ™ (# �yp   ™ " r/   ÿäÿ Æ �yH   • ( �yl   • ( �y(   ™ "! xi]ð   ÿÀŽ äx_ð   ÿÀŽ ÄO x ð   ÿÀŽ ä�xk    ÿÀŽ ä x ÿ   i  ä �yi   ™ ¢  xÿ     Æ �yL   ™ è  �yp   ™ b xmÿ     ä xlÿ   C  ä r     ÿ@Ž â 0r ÿ^    A   â xllÿÀŽ Ä r^   ÿ€Ž ä xjCÿÀŽ ä r   ÿ@Ž ä rmmm   ÿ@Ž â &rtlO   ÿ   â x__   ÿÀŽ â &r^W   ÿ   âx]]   ÿÀŽ â &rrjW   ÿ   â xLÿÀŽ Ä rn   ÿ€Ž ä rpmm   ÿ€Ž â $rtt_   ÿŽ â Er        â &rvW   ÿ   ä $rrr]   ÿŽ ä &rmnO   ÿ   ä &rqpW   ÿ   â ErW t       â $rvvk   ÿŽ â Er r       â #rt    ÿ   â 0rYÿY    A   â Erm m       Ä Er q       â #rr    ÿ   âEro v       â #rmYm   t   âxtÿ   L  â #r|    ÿ   â xHð   ÿÀŽ â #rYW   r   â #rz o   ÿ   â xrÿ   C  ä x Ið   ÿÀŽ Ä xvÿ   L  ä xÿ     ä xttÿÀŽ ä xrrÿÀŽ ä xÿ      ä xHH   ÿÀŽ Ä xvvÿÀŽ ä rx   ÿ@Ž â &rtO   ÿ   â xWI   ÿÀŽ â 0r ÿB    A   â rB   ÿ@Ž â &rrK   ÿ   âxIM   ÿÀŽ â &rqvK   ÿ   â rxxx   ÿ€Ž â $roH   ÿŽ â rBBB   ÿ€Ž â $rW   ÿŽ Ä $rqqI   ÿŽ â Ero o       â &rOxO   ÿ   â Er        â &rBK   ÿ   â Erq q       â #roYo   z   â ErO O       â #r       âEr        â #r q   o   â xMMð   ÿÀŽ â #rYYO   |   â xCÿ   C  Ä xoÿ   L  â #rm    m   â xOJð   ÿÀŽ äxLJ   ÿÀŽ ä xMÿ   M  ä xFð   ÿÀŽ ä xJCÿÀŽ Ä xCoÿÀŽ ä xoÿ   O  ä rOMM   ÿ@Ž ä xÿ     â 0rzÿN    A   â rooo   ÿ@Ž ä rNOO   ÿ€Ž Ä rM   ÿ@Ž â &rqCG   ÿ   â rOoo   ÿ€Ž â &roNK   ÿ   â rMMM   ÿ€Ž â &rJG   ÿ   â xFF   ÿÀŽ â $rqqL   ÿŽ â Ero o       â &rKMG   ÿ   ä $rF   ÿŽ â Erq q       â &rGOG   ÿ   â Er| K       â #ro o   Y   â Er        â 0r ÿP    A   â Ers G       â #rYzq      â &rPjA   ÿ   Ä &r^A   ÿ   â #rKz|   m   â#rGz      â#rmzs   o   â &r|A   ÿ   ä $rP]P   ÿŽ â Er        â &rzl7   ÿ   ä&ron7   ÿ   Ä $rk|   ÿŽ ä &rpA   ÿ   â ErA P       â $rz_z   ÿŽ â ErP o       â 0r@ÿ@    A   â #r    ÿ   âEr        ä Er        â #rA A   ÿ   â Erz z       â #rP@P      â &rt7   ÿ   â #r    ÿ   â#r    ÿ   â #r @z   A   â &r|v;   ÿ   Ä $rAH   ÿŽ ä $r|I|   ÿŽ ä &r7x7   ÿ   â ErA A       â &rzr;   ÿ   âEro |       â &rB;   ÿ   â Er7 7       â #r|@A      â &rAC8   ÿ   Ä $rzWz   ÿŽ ä &r;N;   ÿ   â Er        â 0r9ÿ9    A   â #r@@7      â $rALA   ÿŽ â Erz z       â &rJ8   ÿ   â Er; ;       â 0rÿ:    A   Ä &r:O8   ÿ   â#rP9   P   â &r7M8   ÿ   â Er8 A       â $rF   ÿŽ â #r9o   |   â#r 9z       â#r@9;   @   â Er9 :       â 0r:ÿD  0 A   âEr        â #r8      â0r8ÿ<  0 A   â Er7 7       â #r99   @   â 0rDÿD    A   â  r@m:     A  â #r       â 0r<ÿ<    A   â  r K8     A  â  r9:9     A  â #r77   P   â #r@YD   @  â #r G<      â #rD   9  â  r787     A  â !r@       â &r@jV   ÿ   â!r        â !r       â 0r ÿX    A   â #r<   7  â &rXV   ÿ   Ä &rPl?   ÿ   ä &r^V   ÿ   ä $r@]@   ÿŽ â !r       â $rkX   ÿŽ â Er        â $rP_P   ÿŽ â Er7 @       â &rpV   ÿ   â Er        â 0rUÿU    A   â Er9 P       â #rP    ÿ   âEr        â #r@ 7   ÿ   â &rt?   ÿ   â #rV    ÿ   â &r;n?   ÿ   â #rX    ÿ   â #r9U9   @   â $rH   ÿŽ Ä &r r=   ÿ   ä&r@v=   ÿ   ä&rB=   ÿ   â Er; ;       â $r W    ÿŽ ä $rI@   ÿŽ â Er@        â &r?x?   ÿ   Ä &rJQ   ÿ   â #r;U;   P   â 0r>ÿ>    A   â Er          â &r=N=   ÿ   â ErP        â &rCQ   ÿ   â Er        â $rF   ÿŽ â Er? ?       â #r7>    9   â $rL   ÿŽ â Er= =       â #rU   V   â#r?U?   X   â 0r\ÿ\    A   â Er        â &r MQ   ÿ   â #r;>@   ;   â#r>P      âEr@        â #r=>=   ?   â Er          â #r7\   7   â &r>jc   ÿ   Ä &r^c   ÿ   ä &rQOQ   ÿ   â #r9\@      â$r>]>   ÿŽ â #r;\    ;   â &r@lT   ÿ   âEr        â &rAnT   ÿ   â ErP Q       â 0r ÿf    A   â Er? >       â $r@_@   ÿŽ â Er> A       â 0rbÿb    A   â #r    ÿ   â #r=\P   =   â &rPc   ÿ   â Er@ @       â #r? ?   ÿ   â #r>b>      â &rtT   ÿ   Ä $rkP   ÿŽ ä &rpc   ÿ   â #r@b@   ?   â $r?H   ÿŽ â Er        â &rPvZ   ÿ   âEr        â &rBZ   ÿ   â Er? ?       â &rTxT   ÿ   Ä $rPIP   ÿŽ â Er        â 0rSÿS    A   â #r    ÿ   â#r    ÿ   â &r rZ   ÿ   â ErT T       â #r>S   >   â ErA P       â #rPb?      â $r W    ÿŽ Ä &rJ   ÿ   ä&rM   ÿ   ä &rZNZ   ÿ   â #rTbT      â &rC   ÿ   â Er          â $rF   ÿŽ â Er        â 0r[ÿ[    A   â ErZ Z       â &rO   ÿ   â Er        â $rL   ÿŽ â #r S    @   â #r[   >   â Er?        â #rTSZ   T   â Er        â #r[       â  r8     A  â #rPSA   P   â #r?[?   T   â r=:=     A  â #r<     â #r[   P   â  r>:?     A  â  r;8;     A  â &rjh   ÿ   Ä &r@h   ÿ   â!r""       â #r D9   =  â#r>D   >  â &r^h   ÿ   â #r7<7   ;  â $r]   ÿŽ ä $rk@   ÿŽ ä &rph   ÿ   â !r# #       â &r9ng   ÿ   â !r676       â 0r ÿd    A   â Er        â !r!>!       â Er7        Ä Er        ä Er        â &r>lg   ÿ   â Er 9       â 0r`ÿ`    A   â #r    ÿ   â#r7 7   ÿ   â#r9    ÿ   â#r    ÿ   â $r>_>   ÿŽ Ä &r r   ÿ   â #r`      â &rtg   ÿ   âEr> >       â &rgxg   ÿ   ä $r W    ÿŽ ä $rH   ÿŽ â ErP g       â &rB   ÿ   â Er          â 0rEÿE    A   â #r>`>   7   âEr7        â #rP`P      â Er        â #rE    >   â &rJ   ÿ   Ä &r M   ÿ   â #rE      â &r@v   ÿ   ä $rF   ÿŽ â Er         â 0rRÿR    A   ä $r@I@   ÿŽ â Er         ä #rR      Ä Er; @       â #r@`7   9   â &r7C   ÿ   â #rR       â &rN   ÿ   â  r8     A  â $r7L7   ÿŽ ä &rO   ÿ   â #r<     â Er        â &rt   ÿ   Ä &rjji   ÿ   äO&ri   ÿ   â Er> 7       â $rHH   ÿŽ â Er         â #r7E   P   â $rj]j   ÿŽ ä $rk   ÿŽ ä &r^i   ÿ   Ä &rJ   ÿ   ä&rll   ÿ   ä &rpi   ÿ   â #r7R    7   â &rrr   ÿ   â� Er        â &rvv   ÿ   ä&rBB   ÿ   Ä &rNN   ÿ   â Er j       â $rFF   ÿŽ â Er        â $rl_l   ÿŽ â Er        â 0r ÿe    A   ä &rnn   ÿ   ä&rxx   ÿ   â Erl l       â $rrWr   ÿŽ â #r    ÿ   â&rCC   ÿ   â#r    ÿ   â &rMM   ÿ   Ä &rOO   ÿ   â #r    ÿ   â0raÿa    A   â #r     ÿ   â Ern n       ä Er x       â 0rÿ    A   â Er H       â $rvIv   ÿŽ â Err r       â #ral      âEr B       â #ran      âEr N       â #r a       â#ra      â x   ÿàÿ â $rCLC   ÿŽ â #rr      âErv v       â 0rÿ    A   â ErM M       â #r      âEr O       â #r       â |    pbô â #r;E;   @   â ErF F       â #rv      â Er  C       â #rM      â#r      â  r7:7     A  â #rR>   ;   â #rF      â  r8     A  â #r       â  r::     A  â #r D   7  â #r<     â x))P  ÿà÷ â #rD   :  â !r55       â !r3 3       â !r         â !r$$       â x    ÿàÿ Ä x&&    ÿàÿ ä x**    ÿàÿ ä x,,    ÿàÿ ä x..    ÿàÿ ä x00    ÿàÿ ä r'ÿ'   ÿäÿ â D)       À â Gy   Ûÿÿÿÿƒ ê Ay        € ê Ey  ð     € â G‰  Ð     € ê $x€  Ž Ê ˆs  úÿ H   è ˆs €úÿ H   è ˆs  ûÿ H   è ˆs €ûÿ H   è ˆs 6 üÿ H   è ˆs #€üÿ H   è ˆs " ýÿ H   è ˆs !€ýÿ H   è ˆs 5 þÿ H   è ˆs 3€þÿ H   è ˆs   ÿÿ H   è ˆs $€ÿÿ H   ä Ay        € ê {          ì M        € ê „y €   H   b z+  r      Æ „y      H   ¢ r'++   ÿŽ är)++   ÿŽ â „y    H   è „y	 €  H   ( „y    H   h „y €  H   ( „y    H   ( „y €  H   ( „y    H   ( „y
 €  H   è „y    H   â !r       â/ !r         ÄO „y €  H   â !r       Æ� ‰      b !r		       Æ‰       ¢ !r6       Æy      %   â !r##       Æ y      '   " !r"       Æ ‰      " !r!!       Æ ‰      b !r55       â !r       ä/ ‰	      b !r3
3       â !r        äO ‰      ¢ !r        Æ ‰       " !r$$       Æ ‰      â !r       Æ‰
#      " !r       Æ‰!      h ‰5      b !r		       Æ/ ‰3      b !r       ÆO ‰      ¢ !r        â !r       ä� y       &   " !r##
       Æ‰ €    â !r!!       Æ‰      " !r55       Æ ‰	      b !r33       Æ/ ‰ €    b !r       ÆO ‰      ¢ x   ÿ   Æ ‰#      ¢ x   ÿàÿ â$z   x  Ž â z  r  p`ð â ‰      " !r
       â� z  r  p`ò â $z  ~   Ž â ‰!      â !r       âr ÿ   p"p Ä ‰5      " !r	       âx    p"ò ä ‰3      b !r       â/ r+    ÿˆ ä ‰      b !r       âO rÿÿ   ÿä Ä ‰$      ¢ !r##       â 
  r  ÿà÷ â!r       â ‰
 @    "  '   ÿàõ â !r!!       ä� ‰ €    â ÿÿ   ÿäÿ â !r55       Ä‰ €    " 
 j  ÿ† â !r33       ä‰	 @    b z j  ÿˆ â !r       ä/ ‰ €    b ÿÿ   ÿä â !r$$       ÄO ‰# €    ¢ 
 j  ÿ„ ä 
 k  � â ‰ €    ¢ z k   â !r
       â 
 k   â ‰! €    " !r       â� r+    ÿ„ Ä ‰5 €    â !r       âr&ÿÿ   ÿä ä ‰ 3 €    " !r	       âz
 j  ÿ„ ä ‰" €    b !r       â/ z k  & Ä ‰$      b !r##       âO  )   ÿàù â !r       â ‰ @    ¢ 	+   ÿàÿ â !r!!       ä ‰ @    " ÿÿ   ÿä â !r55       Ä� ‰      â 
 j  ÿˆ â !r33        ä‰ @    "  	   ÿà÷ â !r"       ä‰# @    b 
	 k   â !r%$       Ä/ ‰ @    b (ÿÿ   ÿäÿ ä  j  ÿ† â ‰! @    b !r       äO  k  (� â ‰ 5 @    ¢ !r       â         Ä ‰"3 @    " !r+       â� '   ÿàÿ ä ‰      â !r       Æ‰$ @    " !r#       Æ‰&% €    b !r       â/ !r!       â !        â ‰      b !r 5        ÆO ‰      ¢ !r"3"       Æ ‰      " !r-       â�     ÿàõ ä‰      â %  j   Ž â!r$$       â(ÿÿ   ÿä â ‰      " %  j  ! Ž â !r&%&       â j  ÿ„ â‰      f  k  ( â ‰!       b !r       Æ/ ‰#"      b !r       ÆO ‰'& @    ¢ !r       Æ ‰%$      " !r       â� !r       ä†	 +    â !r       Æ† -   â !r! !       Æ †	     â !r#"#       Æ/ †     â !r'&'       ÆO †	     â !r%$%       Æ †    è †	     è †     è †	 
!    è † 
#   è ‰'      ( †	 %    â M™        € ê x)   ÿàÿ â� !r''       Æ r    ÿàñ È r ÿÿ   ÿä  ä z j  ÿ€ È z k     Ê †y '    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y       "   " ¹z  l      â Ey  Ð#    € â ™x?    â y!      !   b ¹z  F   
   â r" ÿ       â x      ÿ  â y      &   ¢ r% ÿ       â x      ÿ  â r0 ÿ       â y	      '   â r2 ÿ       â x      ÿ  â r  ÿ   pRð Ä r  ÿ   pRò är  !   ÿ(Ž È/ xÿ      È r    ÿ � È x$ÿ     È | $   pbô Ú G)  €"    € ê y      %   "Î xðÿÿÿÿÀŽ â $z&	 }  ÿŽ â r ÿ       â $z w  ÿŽ â r   €ÿàÿ ä r
 ÿ       â 'z 	 y   Ž â xÿ     ä x1$   ÿ   Ä r   ÿ� â 'z s  
 Ž â r&&   ÿàõ ä xøÿÿÿÿÀŽ ä zÿ t    ä r5  €ÿàÿ ä x,ÿ     Ä x   ÿ   ä zÿ z     ä x ÿ   5  â $z p  $Ž âx   ÿàÿ â $z |  $Ž âr*ÿÿ   ÿä â $z |  Ž â r 5   ÿ� â $z v  Ž âz   q      â $z v  Ž â r   ÿò� â $z p  Ž â r) 1   ÿŽ â %xT   ÿŽ â xÿ     Ä r- 1   ÿŽ â$x+    1Ž â x ÿ   5  â %xT   Ž âx,,   ÿ   ä r.   ÿàõ â %xT   ÿŽ â x35      ä z#  X      â %x   Ž â z'  Y      Ä r" ÿ       â %x   Ž â z/1 q  ÿàÿ ä x55   ÿ   ä x7ÿ   ,  ä r(   ÿä ä r#   ÿàõ ä r#.   ÿàù Ä r'   ÿä ä r,&   1àµ ä xC $       â �yI   ™ ¢  r 7*   ÿd~ ä r#   ÿàý â %v Z  C Ž â r'(   ÿä Ä r'   ÿä â $x $   ÿŽ â r5   ÿàõ â�yd    ‘ â r8#   ÿàû ä r   ÿàÿ â �yc   ‘ " r        ä r3   ÿä â �yb    ‘ b r9'   ÿäÿ Æ �yf   ™ h �yg$   • h �yG8   ™ h �yJ(   ™ h �y     • h �yH   ‘ h �yK   ‘ h �yLL   ™ h �yM   ‘ h �yNp   ™ h �yOH   • h �yPl   • h �yQ   ‘ b r8,&   /àÇ Æ �yR   ‘ b r#   ÿàõ ä r47*   ÿ„þ â %v88 Z  C Ž â r'   ÿä Æ $x4$   ÿŽ â/ r58   ÿàõ â �ySP   ™ f r9   ÿàÿ È r3   ÿä Ê �yE   ™ b r 8       Æ �y<(   ™ h �yT    • h �y=L   ™ h �yF$   • h �y>H   • b r,&   )àµ Æ �y?p   ™ b r7*   ÿd~ Æ �y@l   • b %v Z  C Ž È $x$   ÿŽ â r5   ÿàõ È/ r   ÿàÿ È r3   ÿä â �yA    • h �y:   ™ h �y6(   ™ h �y4L   ™ h �y;$   • b r#   ÿàõ Æ �y8H   • b r,&   +àÇ Æ �y9p   ™ b r'   ÿä ä r7*   ÿ„þ â �yBl   • b %v Z  C Ž Æ �yUP   ™ b $x$   ÿŽ â r5   ÿàõ È r   ÿàÿ È r3   ÿä â �y[    • h �yZ   ™ h �y(   ™ b# r,&   -àµ Æ �yX$   • d %v Z  C Ž ä �yCL   ™ b r7*   ÿd~ Æ �yH   • d $x$   ÿŽ ä �y`p   ™ b r5   ÿàõ ä r   ÿàÿ â �yVl   • f r3   ÿä â �y\    • h �y^   ™ h �y_(   ™ h �y]L   ™ h �yp   ™ h �yY$   • h �yDH   • h �yWl   • b xaIÿÀŽ ÄO xhÿ   I  ä xdð   ÿÀŽ ä� xecð   ÿÀŽ äxÿ     ä xibð   ÿÀŽ äxÿ   e  Ä/ xÿ   i  ä r   ÿ@Ž ä xdd   ÿÀŽ ä xhÿÀŽ ä ri   ÿ@Ž â &raf   ÿ   â xc   ÿÀŽ â 0rhÿg    A   â xcGÿÀŽ Ä xjÿ   G  ä r   ÿ@Ž ä rg   ÿ€Ž â &rJ   ÿ   ä $rd   ÿŽ â xbb   ÿÀŽ â &rmcf   ÿ   â xejÿÀŽ Ä rk   ÿ€Ž â &rgf   ÿ   ä $r   ÿŽ â Er        â 0r ÿ     A   â xH   ÿÀŽ â &roeJ   ÿ   â riii   ÿ€Ž â $rlmb   ÿŽ â Er        â &rfkf   ÿ   â Er        â #rm    ÿ   â$rno   ÿŽ â Er l       â &rjiJ   ÿ   â Er f       â #ro    ÿ   â #rh   m   â xfHð   ÿÀŽ Ä xÿ   I  â#rm    ÿ   âErn n       ä Erj j       ä xHKð   ÿÀŽ ä xfÿ   f  â #r    ÿ   â xIÿ   I  Ä xÿÀŽ â #r hn   m   âxHÿ   H  â #rhj   o   â rmff   ÿ@Ž â &rjL   ÿ   â xKK   ÿÀŽ ä xIIÿÀŽ ä rfHH   ÿ@Ž Ä rmmm   ÿ€Ž â &roIN   ÿ   â xHM   ÿÀŽ â $rljK   ÿŽ ä &rnmJ   ÿ   ä $rooH   ÿŽ â ErJ l       â 0rjÿO    A   â rfff   ÿ€Ž â 0rlÿP    A   â Ern n       Ä Ero o       â #rjJ      â &rOfL   ÿ   â xMMð   ÿÀŽ â #rhn      â #rhlo      â xQð   ÿÀŽ ä xMÿ   M  ä ErP O       Ä xJÿ   G  äxOÿ   G  ä xÿ     ä rGMM   ÿ@Ž â #rjP      â xJJÿÀŽ ä rP   ÿ@Ž Ä rGGG   ÿ€Ž ä xMRð   ÿÀŽ â &rnJL   ÿ   â xQQ   ÿÀŽ ä rPPP   ÿ€Ž â &rGN   ÿ   â xoÿ   M  â $rnnQ   ÿŽ â xMOÿÀŽ â &rLPL   ÿ   â rOoo   ÿ@Ž Ä Er        ä Ero n       â &rqMN   ÿ   â xRR   ÿÀŽ ä rOOO   ÿ€Ž ä Ern L       â 0rLÿS  0 A   â #rl      â $rqqR   ÿŽ Ä &rON   ÿ   â  rL     A  â 0rSÿS    A   â #rojo       â ErN q       ä Er         â #rhhS     â &raE   ÿ   â#rjn      â &rcE   ÿ   â #rNlN   o   â $rd   ÿŽ â #rll       â $rnb   ÿŽ Ä &rgE   ÿ   ä&r<   ÿ   ä &rokE   ÿ   â ErE        â 0r ÿT    A   â Er        â $rT   ÿŽ â Er n       ä Er o       â &rji<   ÿ   â#rE E   ÿ   â#r    ÿ   â#ro    ÿ   â#rq    ÿ   â &re<   ÿ   Ä &r =   ÿ   â Erj j       â 0rFÿF    A   â ErT T       â $r   ÿŽ ä $r K    ÿŽ â #rFj      â0r>ÿ>    A   â Er        â #rEFT   E   â Er          â &rf=   ÿ   Ä #rsF   o   â &roI?   ÿ   â#rE>    E   â &r G?   ÿ   â Er        â &r<m<   ÿ   ä $roHo   ÿŽ â Er          â 0r@ÿ@    A   â #r>      â &rTJ=   ÿ   âEr< <       â &r=P=   ÿ   â Ero o       â #r@       â $rTQT   ÿŽ â #rF<   q   â &rM?   ÿ   â Er= =       â #r @o   E   â  r<L     A  â $rR   ÿŽ â ErT T       â &r?O?   ÿ   â #r=>=      â#rS    <  â Er        â #rs>T   s   â ErF ?       â !r       â &ra:   ÿ   Ä &rc:   ÿ   â #r<@   s   â#r=@F   =   â $rd   ÿŽ ä $r@b   ÿŽ ä &rg:   ÿ   ä 0r ÿA    A   ä &r6   ÿ   Ä &rAk:   ÿ   â Er?        â $r:   ÿŽ â Er        â &r>i6   ÿ   â Er @       ä Er A       ä Er@ :       â #rF    ÿ   âErA >       â #r: ?   ÿ   â#r>    ÿ   â#rT    ÿ   â &re6   ÿ   Ä &r 4   ÿ   ä 0r;ÿ;    A   ä $r   ÿŽ ä &r6m6   ÿ   ä $r K    ÿŽ ä &rf4   ÿ   â #r;@   :   â &r:J4   ÿ   â Er? 6       â 0r8ÿ8    A   â Er          â &r4P4   ÿ   â Er6        â &rI9   ÿ   â Er        â #r8       â Er4 4       â $rH   ÿŽ â #r?;?   T   â &r G9   ÿ   â #rA;A   >   â $r:Q:   ÿŽ â #r;   F   â 0rBÿB    A   â Er        â &r;M9   ÿ   â#r?84   ?   â&r9O9   ÿ   â Er          â 0r4ÿU  0 A   â#rA86   A   â Er: :       â $r;R;   ÿŽ â #r6B      â0rUÿU    A   â Er> 9       â #rAB    A   â  rl4     A  â  r 4=     A  â #r8:      â Er8 ;       â #rNNU     â #r?B>   ?   â#rU<      â &rcZ   ÿ   â #r:B8      â  r?4?     A  â !r       â &raZ   ÿ   â  r8LA     A  â #r?U:   ?  â $r:b   ÿŽ Ä $rd   ÿŽ ä &rgZ   ÿ   â #r6S6   8  â &r8i   ÿ   âEr9        â 0r ÿ[    A   â Er        â &r   ÿ   ä &rZkZ   ÿ   â !r6       â 0rXÿX    A   â Er8 8       â $r6   ÿŽ â Er :       â #r    ÿ   âEr Z       â #r9 9   ÿ   â Er6 6       â #r8X8      â #r;    ÿ   â#r    ÿ   â &re   ÿ   Ä &r C   ÿ   â #r6X6   9   â $r   ÿŽ ä &r:JC   ÿ   ä &r9m   ÿ   ä $r K    ÿŽ â Er        â &rfC   ÿ   Ä $r<Q:   ÿŽ â Er: 9       â 0rÿ    A   â Er          â #rX   ;   âEr        â #r:X:      â Er9 <       â #r    6   â &r;M`   ÿ   ä&rI`   ÿ   Ä &r G`   ÿ   â #r99      â$r;R;   ÿŽ ä $rH   ÿŽ â Er         â 0rVÿV    A   â #r   8   â &rCPC   ÿ   â Er6 ;       ä Er         â #rV      â &raa^   ÿ   Ä &rcc^   ÿ   â ErC C       â #rV6   9   â#r V       â  r6L     A  â $rada   ÿŽ ä $rbc   ÿŽ ä &rg^   ÿ   ä&rk^   ÿ   â #rCC   :   â #rS    6  â &r_   ÿ   â Era a       â &ree_   ÿ   â Er        â &rii_   ÿ   â Er        â &r]   ÿ   â Er        â &r_m_   ÿ   Ä 0r ÿ\    A   ä &r`O`   ÿ   ä $r   ÿŽ â Eri i       â $ree   ÿŽ â #r a   ÿ   â &rJJ]   ÿ   â Er_ _       â $rKK   ÿŽ â #r    ÿ   â &rII   ÿ   Ä &rMM   ÿ   ä&rff]   ÿ   ä&rPP]   ÿ   ä &rGG   ÿ   ä&rOO   ÿ   â #r    ÿ   â0rYÿY    A   â #r     ÿ   â $rJQJ   ÿŽ â Er        â 0rDÿD    A   â Ere e       â #riYi      âErf f       â #r_Y_       âErP P       â $rIHI   ÿŽ â x$$   ÿàÿ â $rMRM   ÿŽ â ErK K       â #rY      âEr` `       â 0rWÿW    A   â ErJ J       â #reYe      â ErG G       â #rfDf   i   âErO O       â #rPDP   _   â | $   pbô â #r DK      â ErI I       â #r9V`   C   â ErM M       â #rJDJ   e   â #rGWG   f   â#rOWO   P   â#rWI       â r 49     A  â #rJWM   J   â  rLLG     A  â  rO4O     A  â #rU      â x##P  ÿà÷ â #rS   L  â #rOUJ   O  â !rh       â !rN       â !r?       â !r22       â !r00       â !r%%       â !r"O"       â x))    ÿàÿ Ä x++    ÿàÿ ä x--    ÿàÿ ä x//    ÿàÿ ä x11    ÿàÿ ä r'ÿ'   ÿäÿ â D)       À â Gy  °àÿÿÿÿƒ ê Ay        € êÏ Ey  Ð     € â G‰  °     € ê $x @  !Ž Ê ˆs  ûÿ H   è ˆs €ûÿ H   è ˆs  üÿ H   è ˆs €üÿ H   è ˆs  ýÿ H   è ˆs €ýÿ H   è ˆs 2 þÿ H   è ˆs 0€þÿ H   è ˆs % ÿÿ H   è ˆs "€ÿÿ H   ä Ay        € ê {          ì M        € ê „y !     H   b z&  r      Æ „y! €   H   ¨ „y!    H   è „y! €  H   ( „y!    H   h „y! €  H   ( „y!    H   ( „y! €  H   ( „y!    H   è „y! €  H   â !r         â/ !r       ÈO ‰       " !r       Æ� ‰      b !r       Æ‰      ¢ !r       Æy      %   â !r       Æ ‰      " !r22       Æ ‰      b !r00       â !r        ä ‰      " !r%       Æ ‰       " !r       â/ !r""       ä ‰2      b !r       ÆO ‰0      ¢ x   ÿ   Æ� ‰      â !r       â$z x  Ž â z  r  p`ð â ‰"      " !r       â$z	 ~  Ž â r !ÿ   p"p â !r       â ‰      " r&&   ÿŽ â !r         Ä ‰      b !r2       Æ/ ‰      b !r0       ÆO ‰
      ¢ 	   ÿàõ â !r       ä� ‰  €    â  ÿÿ   ÿä â !r"       ä‰      ( ‰      " !r       Æ ‰      " !r       Æ‰      b !r       Æ/ ‰      b !r
       âO x
   ÿàÿ â !r        â� ‰ €    ¤ z 
 r  p`ò â !r       â‰ €    ä x !   p"ò â !r       â ‰ €    " !r       Æ ‰ €    " !r       Æ‰  @    b !r       Æ/ ‰ €    h ‰ €    b !r
       âO 
 r  ÿàù ä ‰ €    ¢ !r       â� r&   ÿŠ Ä ‰ €    â !r       âr"ÿÿ   ÿäÿ ä ‰ €    " !r       â 
	 j  ÿ„ â !r         â‰
 @    " r&   ÿ† â !r       Ä/ ‰       b rÿÿ   ÿäÿ â !r       ä ‰ @    b z j  ÿ† â !r       äO ‰ @    ¢ z k  � â !r       Ä� ‰ @    â z j  ÿŠ â !r       ä‰ @    " 
		 k    ä $ÿÿ   ÿä â ‰ @    " !r
       â 
&   ÿàÿ Ä ‰ @    " !r#        â/ 
   ÿàõ ä ‰ @    b !r       âx        ä ‰ @    b !r       âO    ÿàÿ â !r       â� ‰       ¢ z k  "� â !r       Ä‰      â     ÿà÷ â%v
 j   Ž â !r       â ‰      " 
 j  ÿˆ â !r       ä ‰      " "ÿÿ   ÿä â !r       Ä/ ‰      b  j  ÿ„ â !r       ä‰      b !ÿÿ   ÿäÿ ä 
 k  $ â ‰      b !r%        âO   j  ÿ† Ä ‰      ¢ !r       â�  k  " ä ‰      â !r       â  k  !� â !r       â †	 
#    â !r       Æ/ † 
%   â !r       Æ†	     â !r       Æ †     â !r       ÆO †	     â !r       Æ� †    è †	     è †     è †	     â M™        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      "   " ¹z  l      â Ey  p    € â ™x?    â y      !   b ¹z  F   
   â r  ÿ       â x.      ÿ  â y      %   ¢ r- ÿ       â x      ÿ  â r ÿ       â y	      &   â r ÿ       Ä r ÿ   pRð är ÿ   pRò är    ÿ(Ž È/ xÿ      ä x   ÿ   äO r    ÿ � È xÿ     È |    pbô Ú G)       € ê y      '   "Ž xðÿÿÿÿÀŽ â $z
 p  Ž â r ÿ       â $z	 w  ÿŽ â r   €ÿàÿ ä r ÿ       ä x ÿ     ä x(   ÿ   Ä r    ÿ� ä z  X      ä xøÿÿÿÿÀŽ ä x'ÿ     ä r+  €ÿàÿ ä x''   ÿ   â 'z  y   Ž â z#  Y      Ä r  ÿ       â 'z	 s   Ž â zÿ z     ä x ÿ   +  â $z! }  ÿŽ â zÿ t    â $z |  Ž âz&( q  ÿàÿ â $z |  
Ž â x   ÿàÿ â $z
 v  Ž â r +   ÿ� â $z v  Ž â r!!   ÿàõ â $z p  
Ž â r   ÿò� â %x
T   ÿŽ â z  q      Ä r%ÿÿ   ÿä â %xT   ÿŽ â r(   ÿŽ äx ÿ   +  â $x$   (Ž â xÿ     â %x   
Ž â r*
   ÿàõ ä x)+      â %xT   Ž â x++   ÿ   Ä x,ÿ   '  â %x   Ž â r"   ÿä Æ rR   ÿàõ ä x; $       ä rS#   ÿä ä r'!   (àµ ä r*   ÿàù â �yRR   ™ ¢  r ,%   ÿd~ â %v Z  ; Ž â r   ÿàý Ä r#"   ÿä â $x $   ÿŽ â r#   ÿä ä r0   ÿàû â �yJ    ‘ â r   ÿàÿ ä r+   ÿàõ â �yH    ‘ b	 r1#   ÿäÿ Ä r)   ÿä â �yG   ‘ h �yM$   • h �y\   ™ h �yU0   ™ h �yX(   ™ h �y     • h �yZ   ‘ h	 �yY   ‘ h �yWH   • h �yV   ‘ h	 �yNL   ™ h �yS   ‘ b r0'!   &àµ Æ �yT   ‘ b	 r2,%   ÿd~ Æ �yPl   • b %v00 Z  ; Ž Æ �yLp   ™ b $x32$   ÿŽ â r2+0   ÿàõ È r113   ÿàÿ È r3)1   ÿä â �yF0    • h �yD2   ™ h �y@2(   ™ h �y>2L   ™ h �yC0$   • h �y?0H   • h	 �y92p   ™ b	 r   ÿà÷ Ä r'!   àÙ â/ �yA0l   • b r
   ÿàõ ä r#   ÿäÿ ä r,%   ÿ¤~ är#   ÿä â �y7P   ™ " %v Z  ; Ž Æ �y8P   ™ " $x$   ÿŽ â r+   ÿàõ È r   ÿàÿ È r)   ÿä â �y6    • ( �y5   ™ ( �y2(   ™ ( �y4$   • " r'!   $àµ Æ/ �y1L   ™ " r,%   ÿd~ Æ �y0H   • " %v Z  ; Ž Æ �y=p   ™ " $x$   ÿŽ â r+   ÿàõ ä �y3l   • $ r   ÿàÿ È r)   ÿä â �yE    • ( �y<   ™ ( �y;L   ™ ( �y:(   ™ ( �yB$   • ( �yp   ™ (+ �yH   • ( �yl   • " xIRÿÀŽ ÄO xKÿ   R  ä xJð   ÿÀŽ ä� xHð   ÿÀŽ äxÿ     ä xÿ     ä r   ÿ@Ž Ä xOGð   ÿÀŽ â 0rÿM    A   â xJJ   ÿÀŽ ä xKKÿÀŽ â &rI\   ÿ   â r   ÿ@Ž ä rM   ÿ€Ž ä xOÿ   O  Ä xG   ÿÀŽ â &rKX   ÿ   â xGUÿÀŽ â $rJ   ÿŽ â rQ   ÿ€Ž â &rM\   ÿ   ârOOO   ÿ@Ž â &r_G\   ÿ   â xHH   ÿÀŽ â $r[   ÿŽ â rOOO   ÿ€Ž â 0r ÿ     A   â Er        â &r^Q\   ÿ   â Er        â $r__H   ÿŽ Ä &r]OX   ÿ   â Er[ [       â #r\    ÿ   âEr ^       â #r    ÿ   âEra _       ä Er] ]       â #r[      â #rd    ÿ   â#rb a   ÿ   â xÿ   R  Ä x[Zð   ÿÀŽ â #r_]   \   â x Yð   ÿÀŽ äx]Y   ÿÀŽ ä xYÿÀŽ ä x`ÿ   [  ä x^ÿ   U  Ä xÿ      ä xaÿ   U  ä r```   ÿ@Ž ä x[^ÿÀŽ ä r   ÿ@Ž â 0r ÿW    A   â xZ   ÿÀŽ Ä xZaÿÀŽ ä rW``   ÿ€Ž â &ra[X   ÿ   â r^   ÿ€Ž â &rcZN   ÿ   â x\V   ÿÀŽ â &rXWX   ÿ   ä $raa   ÿŽ ä &r^N   ÿ   Ä &r`YN   ÿ   ä $rec\   ÿŽ â Erc X       â $r``]   ÿŽ â Era a       ä ErX        ä Er` `       â #ra   b   â xaÿ   U  â #r_ X   _   â xXVð   ÿÀŽ Ä xUSð   ÿÀŽ â #r `      â xXÿ   X  ä xRÿ   R  ä x`Tð   ÿÀŽ ä xUÿ   U  ä rXXX   ÿ@Ž Ä xVRÿÀŽ ä x`ÿ   `  ä rUUU   ÿ@Ž â 0rbÿP    A   â xRaÿÀŽ â #rc   d   â rPXX   ÿ€Ž â &rdVL   ÿ   â xSS   ÿÀŽ Ä r```   ÿ@Ž ä rXUU   ÿ€Ž â &raRL   ÿ   â xTT   ÿÀŽ â &rUPN   ÿ   â r```   ÿ€Ž â $rddS   ÿŽ ä &rNXL   ÿ   â Erf e       â $raaT   ÿŽ Ä &rc`L   ÿ   â ErL d       ä ErN N       â #r f      â ErU U       ä Erd a       â #r_bN   _   â Erf c       â #rU U      â &rGD   ÿ   â #rNbd      â&rID   ÿ   â #rLbL      â#rUbf   U   â $rbH   ÿŽ Ä $rJ   ÿŽ ä &rMD   ÿ   ä 0r ÿF    A   â Era        â &rK@   ÿ   ä &rcQD   ÿ   â Er        â &rFO@   ÿ   Ä $rD   ÿŽ â Er c       ä Er b       ä Erc F       â #rF    ÿ   â&r[@   ÿ   â #rd    ÿ   Æ $r   ÿŽ â #rf    ÿ   â Erb D       â &rZ>   ÿ   â #rD a   ÿ   â &r@W@   ÿ   â Er        â &r^>   ÿ   ä0rCÿC    A   ä &r Y>   ÿ   â Er        â $ra\   ÿŽ â Er @       â 0r?ÿ?    A   â #rFCc   F   â $r ]    ÿŽ â #rdC   d   â &r@V9   ÿ   Ä &rR9   ÿ   ä &r>P>   ÿ   â #rF?   F   â $r@S@   ÿŽ â Er          â &rX9   ÿ   â #rDCb   D   â$rT   ÿŽ â Era a       â &r9`9   ÿ   â Er> >       â #rfC   f   â 0rAÿA    A   â Er@ @       â #r ?    D   âEr        â #rd?a   d   âEr        â #rf?>   f   â Er9 9       â #r A@       â#rA   F   â#rdA   d   â0r@ÿ7  0 A   â#rAA9   f   â 0r>ÿ8  0 A   Ä 0r7ÿ7    A   â  rD@A     A  â 0r9ÿ8    A   â  r8>     A  ä #r7d   D  â &rG5   ÿ   Æ !r       â &rI5   ÿ   â #r 9    8  â $r?H   ÿŽ ä $rJ   ÿŽ ä &rK2   ÿ   ä &rM5   ÿ   â !r        â 0r ÿ6    A   Ä &rAQ5   ÿ   â Er5        â $r6   ÿŽ â Er        ä Er ?       â 0r4ÿ4    A   â Er6 6       â &r8O2   ÿ   â#r5 5   ÿ   â#r?    ÿ   â&r[2   ÿ   â #r    ÿ   â Er8 8       â #r546   5   â &r6W2   ÿ   â Er A       â $r2   ÿŽ Ä &r^1   ÿ   â#r48      â &rAP1   ÿ   â Er8 6       â #r    ÿ   â Er6 2       â &r Y1   ÿ   â Er2        â 0r0ÿ0    A   ä &r1Z1   ÿ   â ErD A       â $r ]    ÿŽ â #r48      â#r?46   ?   â #r202      â &rR=   ÿ   Ä $r1\1   ÿŽ ä &r4V=   ÿ   ä&rX=   ÿ   â #rD0D      â Er          â $rT   ÿŽ â Er1 1       â $r4S4   ÿŽ â Er        â 0r3ÿ3    A   â #r 0    5   âEr        â #r001   ?   â Er 4       â #r3   2   Ä #r23   0   â &rII<   ÿ   â#r 3       â  r0>     A  â &rGG<   ÿ   ä &r=`=   ÿ   â #r19    0  â $rIJI   ÿŽ ä $rHG   ÿŽ Ä &rM<   ÿ   ä&rQ<   ÿ   ä &r Y;   ÿ   ä &r[[:   ÿ   â Er= =       â $r]]    ÿŽ â ErI I       â 0r ÿE    A   â Er        â $r[[   ÿŽ â Er        Ä Er        â &rKK:   ÿ   ä&rOO:   ÿ   ä&r:W:   ÿ   â #r=3=   D   â #rI I   ÿ   â#r    ÿ   â#r    ÿ   â#r    ÿ   â &r5Z;   ÿ   â Er  [       â &r^^;   ÿ   Ä 0rBÿB    A   â ErO O       â $rKK   ÿŽ â Er: :       â &r;P;   ÿ   â  r4@=     A  â &rVV   ÿ   â#rB       â &r=R   ÿ   ä&rXX   ÿ   â ErK K       â $r5\5   ÿŽ â Er^ ^       â &r`   ÿ   â Er  ;       â 0rÿ    A   â #rBO      â#rB:      â x   ÿàÿ â $rVSV   ÿŽ â Er] ]       â $r=T=   ÿŽ â Er5 5       â 0rÿ    A   â ErX X       â #rIBK   I   â Er        â #r^      â#r       â |    pbô â #rI]   I   âErV V       â #r5      â Er  =       â #rX      â#r      â r__>     A  â  rUU@     A  â #rVV   I   â  r>>     A  â #r        â  r@@     A  â #rNN7   U  â #r372   4  â #rLL9   _  â #r>9V   >  â #r77    @  â x   ÿà÷ â !r/L/       â !rN       â !r1       â !r.3.       â !r->-       â !r 7        â x@   ÿàÿ Ä x$$@   ÿàÿ ä x&&@   ÿàÿ ä x((@   ÿàÿ ä r#ÿ#   ÿäÿ â D)       À â Gy  ðåÿÿÿÿƒ ê Ay        € ê� Ey  °     € â G‰  �     € ê r   ÿ@Ž Ê ˆs / üÿ H   è ˆs €üÿ H   è ˆs  ýÿ H   è ˆs €ýÿ H   è ˆs  þÿ H   è ˆs .€þÿ H   è ˆs - ÿÿ H   è ˆs  €ÿÿ H   ä Ay        € ê {          ì M        € ê „y €   H   b z  r  p`ð Æ „y €  H   ¢ r ÿ   p"p Æ „y      H   è „y €  H   ( „y    H   h „y    H   ( „y €	  H   ( „y €  H   ( „y    H   ( „y    H   ( „y! €  H   " !r       Æ/ „y    H   b !r       ÆO „y    H   ¢ !r  /       Æ� „y €  H   â !r       Æ„y    H   " !r         Æ„y
 €  H   b !r       Æ „y  	  H   " !r       Æ „y%    H   " !r       Æ „y  
  H   " !r         Æ „y# €
  H   " !r       Æ „y'    H   " !r!!.       Æ „y €  H   â !r       Æ/ „y €  H   â !r-       ÆO „y €  H   â !r       â� !r       ä‰       b !r
!
       Æ‰      ¢ !r       Æ ‰      " !r%       Æ ‰      â !r       â !r

#       È ‰      " !r'       Æ ‰
      b !r        â/ !r        ä ‰      b !r       âO !r       ä ‰       ¢ !r       â !r       â !r       â� ‰      ( ‰      â !r       Æ‰      " !r
       Æ‰      b !r       Æ/ ‰
      b !r         ÆO ‰      ¨ ‰      ¢ !r       Æ ‰  €    " !r       â� !r       ä‰ €    â !r       Æ‰ €    " !r

       Æ/ ‰ €    b !r       ÆO ‰ €    ¢ !r       Æ ‰
 €    b !r        Æ ‰ €    ( ‰      " !r       Æ� y      '   â !r       Æ‰ @    " !r       â/ !r       äO ‰ @    b $z	 x  Ž â x   ÿàÿ â !r
       â‰ @    ä z  r  p`ò â !r       â ‰
 @    $ x    p"ò â !r       â ‰ @    ¢ $z  ~  Ž Æ� ‰ @    â z  r      â !r       ä‰ @    " 
  r  ÿàõ ä r   ÿŽ â ‰ €    b ÿÿ   ÿä â !r	       â/ 
 j  ÿ„ â‰      b r    ÿ† â !r
       â 
 k   â !r       â    ÿàÿ â !r       âO ‰	      " rÿÿ   ÿäÿ â !r       Ä� ‰      ¢ z j  ÿ† â!r       ä‰      â     ÿàõ â !r       ä‰      " z k  � â !r       Ä/ ‰      b     ÿà÷ ä x        â ‰ @    b ÿÿ   ÿä ä  j  ÿ„ â ‰      b ÿÿ   ÿäÿ â !r	       â 

 j  ÿ† â %v  j   Ž â 	 k   â !r       âO 
 k  � â !r       â� †	     â !r       Æ†    â !r       Æ/ †	     â !r       Æ†     â !r       Æ †	     è †    è ‰      h  †	 
    â M™        € ê x   ÿàÿ â !r       Æ/ r    ÿàñ È r ÿÿ   ÿä  ä z j  ÿ€ È z k     Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      "   " ¹z  l      â Ey  `    € â ™x?    â y      !   b ¹z  F   
   â r ÿ       ä y!      %   ¢ r0 ÿ       ä r" ÿ       â y      &   â r ÿ       Ä r, ÿ       ä r2 ÿ       ä r ÿ   pRð är ÿ   pRò är    ÿ(Ž È/ xÿ      ä x!!   ÿ   äO r    ÿ � â $À   Ž Æ x.ÿ     È | .   pbô Ú G)       € ê y      '   "Ž xðÿÿÿÿÀŽ â $z! p  .Ž â r ÿ       â $z	 w  ÿŽ â r   €ÿàÿ ä r ÿ       ä x ÿ     ä x'.   ÿ   Ä r    ÿ� ä z  X      ä xøÿÿÿÿÀŽ ä x(ÿ     ä r+  €ÿàÿ ä x((   ÿ   â 'z  y   Ž â z#  Y      Ä r ÿ       â 'z s   Ž â zÿ z     ä x ÿ   +  â $z  }  ÿŽ â zÿ t    â $z |  .Ž â z%' q  ÿàÿ â $z |  Ž â x!   ÿàÿ â $z v  Ž â r +   ÿ� â $z v  Ž â z   q      â $z p  Ž â r 	    ÿàõ â %xT   ÿŽ â r   ÿò� Ä r '   ÿŽ â %xT   ÿŽ â xÿ     ä r&ÿÿ   ÿä â %x   Ž â x ÿ   +  ä r*   ÿàõ â %x
T   Ž â x)+      Ä x++   ÿ   â %x   Ž â x-ÿ   (  ä r$   ÿä ä r   ÿàõ ä xI $       ä r#   ÿä ä r(    'àµ Ä r*   ÿàù â �y/   ™ ¢  r -&   ÿd~ â %v Z  I Ž â r
   ÿàý ä r#$   ÿä â $x $   ÿŽ â r#   ÿä ä r   ÿàû â �y     ‘ â r   ÿàÿ Ä r+   ÿàõ â�y>   ‘ " r        Æ �y<    ‘ b r)   ÿä ä r#   ÿäÿ â �y:$   • h �y;   ™ h �y5   ™ h �y7(   ™ h �y=    • h �y8   ‘ h �y1   ‘ h �y3H   • h �y6   ‘ h �y4L   ™ h �y9   ‘ b r(    %àµ Æ �y?   ‘ b r-&   ÿd~ â %v Z  I Ž ä �yEp   ™ d $x$   ÿŽ â r+   ÿàõ â �yAl   • f r   ÿàÿ È r)   ÿä â �yH    • h �yG   ™ h �yB(   ™ h �yF$   • h �yCL   ™ b r(    àµ Æ/ �y@H   • b r-&   ÿd~ â %v Z  I Ž ä �yDp   ™ d $x$   ÿŽ â r+   ÿàõ â �yIl   • f r   ÿàÿ È r)   ÿä â �yM    • h �yJ   ™ h �yK(   ™ h �yL   ™ h �yL$   • b r   ÿàõ Æ �yH   • h! �yp   ™ b r#   ÿä ä r   ÿàõ â �yl   • f r	#   ÿä â �yP   ™ h �yP   ™ b xPÿ   /  ÄO xN/ÿÀŽ ä x ð   ÿÀŽ ä�x    ÿÀŽ ä x >ð   ÿÀŽ äxÿ     ä x<ð   ÿÀŽ ÄxO<   ÿÀŽ ä x<PÿÀŽ ä xÿ      ä xÿ     ä/ rP   ÿ@Ž ä r   ÿ@Ž â &rN;   ÿ   â r   ÿ@Ž Ä rPPP   ÿ€Ž â 0rÿ:    A   â x>>   ÿÀŽ â $r   ÿŽ â x5ÿÀŽ ä r:   ÿ€Ž â &r<7   ÿ   â rR   ÿ€Ž â &rP;   ÿ   ä 0r ÿ=    A   Ä &rT;   ÿ   ä $r=>   ÿŽ â Er        â &r;R;   ÿ   â Er        â $rTTO   ÿŽ â Er= =       Æ #rV    ÿ   âEr ;       ä Er T       â #rT    ÿ   â &rQ:7   ÿ   â #rZ    ÿ   âxÿ   /  â #rX    ÿ   â #r=   T   â x 8ð   ÿÀŽ Ä xTÿ   5  ä x1ð   ÿÀŽ äx=1   ÿÀŽ ä ErQ Q       ä x1ÿÀŽ ä xÿ      Ä xSÿ     ä xTTÿÀŽ â #rQQ   V   â r   ÿ@Ž ä xUÿ   5  ä rVSS   ÿ@Ž â &rST7   ÿ   ä 0r ÿ3    A   â r3   ÿ€Ž â $rSS=   ÿŽ â xUÿÀŽ Ä rVVV   ÿ€Ž â &r34   ÿ   âx;8   ÿÀŽ â &rU4   ÿ   âx86   ÿÀŽ â &r7V7   ÿ   â ErS S       â &r\14   ÿ   â Er        â $rWU8   ÿŽ â ErU 7       â #r7S   X   â xXÿ   5  â $r\\;   ÿŽ â x56ð   ÿÀŽ â #rQ    Q   â x9ð   ÿÀŽ Ä x/ÿ   /  ä xS?ð   ÿÀŽ ä x5ÿ   5  ä Er\ \       ä xÿ     ä x6/ÿÀŽ Ä xSÿ   S  ä r/55   ÿ@Ž â #r \      â r   ÿ@Ž ä xXXÿÀŽ â #rU   Z   â r\SS   ÿ@Ž ä r///   ÿ€Ž Ä x99   ÿÀŽ ä rZ   ÿ€Ž â &r6E   ÿ   âx??   ÿÀŽ â &rSXE   ÿ   â r\\\   ÿ€Ž â &r5/4   ÿ   ä $r9   ÿŽ â ErW W       â &r4ZE   ÿ   Ä $rSS?   ÿŽ â Er5 5       â &rE\E   ÿ   â Er        â 0rAÿA    A   â Er4 4       â #r7 W   7   âErS S       â #r  5      â ErE E       â #r5A      â#r4A4   Q   â#r7AS   7   â&rNG   ÿ   â #rAAE       â 0r ÿH    A   Ä &rHG   ÿ   ä &r<B   ÿ   ä $r   ÿŽ ä &rPG   ÿ   ä$rHOH   ÿŽ ä &rSRG   ÿ   â ErE        â $rG>   ÿŽ â Er        â &rQ:B   ÿ   â Er H       â 0rFÿF    A   â ErH G       â #rE E   ÿ   âEr S       â #rG    ÿ   â ErQ Q       â #rFH   E   â &rHTB   ÿ   â #r    ÿ   â#r    ÿ   â #rEFQ   G   â &r 1C   ÿ   Ä &rQC   ÿ   ä $rH=H   ÿŽ ä &rGVB   ÿ   ä $r ;    ÿŽ â ErH H       â $rQ8Q   ÿŽ â ErG G       â 0r@ÿ@    A   â Er          â &rB3C   ÿ   â ErQ Q       â #rFH      â#rFG      â #r@       â &rF6D   ÿ   â ErB B       â #r @Q      â &rZD   ÿ   Ä &rC/C   ÿ   ä &rHXD   ÿ   ä $rF9F   ÿŽ â #rE@B   E   â 0rBÿI    A   â Er        â $rH?H   ÿŽ â ErC C       â &rD\D   ÿ   â ErF F       â #rEB   E   â Er H       â #r@C      â ErD D       â #r@BF      â &rNJ   ÿ   Ä &rFJ   ÿ   â #rB       â#rBBD      â $r   ÿŽ ä &rTTK   ÿ   ä$rFOF   ÿŽ ä &r<K   ÿ   ä &rPPJ   ÿ   Ä &rRJ   ÿ   ä $rT=T   ÿŽ â Er=        â $r>>   ÿŽ â Er P       â 0r ÿM    A   â Er F       ä Er        ä #r= =   ÿ   â#r    ÿ   â#r    ÿ   â#r    ÿ   â &r 1   ÿ   Ä &r::K   ÿ   ä&rKVK   ÿ   â Er> >       â $r ;    ÿŽ ä 0rLÿL    A   ä &r   ÿ   âEr K       â &r33   ÿ   â Er: :       â $r8   ÿŽ â Er          â 0rÿ    A   â #r=L>   =   â ErT T       â &r//   ÿ   Ä &r66   ÿ   â#rL      â &rZZ   ÿ   â Er 3       â #rL:      â #r    =   â Er        â #rLT      â $r696   ÿŽ â Er  /       â 0rÿ    A   â ErZ Z       â &rXX   ÿ   â#r      â &r\   ÿ   â #r8      âEr 6       â #r       â 0r ÿ  0 A   â#rZ      â x..   ÿàÿ â $rX?X   ÿŽ â Er        â #r      â0rÿ    A   â  r44      A  â  rE E     A  â r      A  â | .   pbô â 0r ÿ  0 A   âEr X       â #r      â #r     â 0rÿ    A   â  rA      A  â  r B     A  â#r   8   â  r       A  â #r55   4  â #rE@   E  â #r77     â #r     â#r      â x   ÿà÷ â !r252       â !rE       â !r00       â !r,7,       â !r""       â !r       â x@   ÿàÿ Ä x%%@   ÿàÿ ä x''@   ÿàÿ ä r#ÿ#   ÿäÿ â D)       À â Gy   ëÿÿÿÿƒ ê Ay        € ê� ˆ 2 ýÿ H   è ˆ ,€ýÿ H   è ˆ  þÿ H   è ˆ "€þÿ H   è ˆ 0 ÿÿ H   è ˆ €ÿÿ H   è {          ì M        € ê „y     H   " z ! r  p`ð â $z x  !Ž â z  r      â „y     H   b r ÿ   p"p Æ „y €   H   ¨ „y    H   è „y    H   ( „y €  H   h „y    H   h „y €  H   h „y    H   h „y €  H   h „y    H   b !r2       Æ „y	 €  H   " !r         Æ/ „y    H   b !r,       ÆO „y €  H   ¢ !r       Æ� „y €  H   â !r         Æ„y    H   â !r"       Æ„y
 €  H   â !r0       Æ „y €  H   â !r       â !r       ä ‰       " !r       â !r       â !r	       â !r       È/ ‰      " !r       ÆO ‰      b !r       Æ� ‰	      ¢ !r        â!r       È ‰       â !r

       Æ ‰      " !r

       Ê ‰
      b !r       â !r       È/ ‰      " !r		       ÆO ‰      h ‰	      ¢ !r         â� !r       È‰  €    è ‰      " !r
       Ê‰
      b !r       â !r       È/ ‰ €    " !r	       ÆO ‰ €    h ‰	 €    ¢ !r        â� !r       È‰  @    è ‰ €    " !r

       Ê‰
 €    b !r       â !r       È/ ‰ @    " !r		       ÆO ‰ @    è ‰	 @    b !r         â� !r       äy      '   ¨ ‰ @    â !r
       âx
!   ÿàÿ Ä ‰       " !r       â z 
 r  p`ò ä ‰ @    $ x    p"ò ä ‰
      b !r	       Ê/ ‰      b $z ~  Ž âO !r       â !r       Æ� 
 r  ÿàõ â ‰      ¢ 	   ÿàÿ â!r        âÿÿ   ÿä â ‰      â 
 j  ÿ„ â !r       â r   ÿ† ä 
 k   â !r
       â‰      " 		   ÿàõ Ä x        ä ÿÿ   ÿä â !r       â/ r ÿÿ   ÿäÿ â %v j   Ž â 	 j  ÿ„ ä z j  ÿ† â !r       âO 		 k   â †	     â z k   � â !r       Ä� †    è †	     â !r       Æ †     è †	     â M™        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      "   " ¹z  l      â Ey  Ð    € â ™x?    â y      !   b ¹z  F   
   â x      ÿ  â r  ÿ       â y      %   ¢ $r(ÿÿ   ÿ Ž Æ y      &   â r ÿ   pRð är ÿ   pRò är    ÿ(Ž È/ xÿ      ä x   ÿ   äO r    ÿ � ä    ÿ8Ž ä xÿ     È |    pbô Ú G)       € ê y      '   "Ž xðÿÿÿÿÀŽ â $z
 p  Ž â r ÿ       â $z w  ÿŽ â r ÿ       â $x    
Ž â x#   ÿ   â $rÿÿ   ÿ Ž â z-  X      ä x ÿ     Ä z!  Y      ä r    ÿ� ä z# q  ÿàÿ ä xøÿÿÿÿÀŽ ä x%ÿ     â 'z  y   Ž Æ x%%   ÿ   â 'z s   Ž â zÿ z     ä x+ÿ   %  â $x)   
Ž â zÿ t    â $z }  ÿŽ ä $z |  Ž âx ÿ   )  â $z	 |  
Ž â x   ÿàÿ â $z
 v  Ž â r )   ÿ� â $z	 v  	Ž â r   ÿàõ â $z p  
Ž â r   ÿò� â %x	T   ÿŽ â r$ÿÿ   ÿä Ä xÿ     ä x ÿ   )  â %x   Ž â r&   ÿàõ ä x')      â %xT   Ž â x))   ÿ   Æ %x
T   ÿŽ È $x"    ä %x   
Ž È r-   ÿàõ ä r-&   ÿàù ä r!   ÿä ä r%   #àµ ä r-   ÿàý â �y*   ™ ¢  r!"   ÿä Ä x? $       ä r +$   ÿd~ â �y5    ‘ â r	!   ÿä â %v Z  ? Ž ä �y6   ‘ $ $x1 $   ÿŽ ä �y     ‘ ¢
 r)   ÿàõ Ä r1   ÿàÿ â �y.   ‘ ¢
 r-   ÿàû ä r'   ÿä â �y7    • ¤  $x   !Ž ä �y4   ™ ¨  �y1$   • ¨  �y/   ™ ¨  �y,(   ™ ¨  �y2   ‘ ¨ �y0   ‘ ¨ �y9l   • ¨  �y3H   • ¨  �y8L   ™ ¨ �y:   ‘ ¨
 �y<   ‘ ¢
 r%   àµ Æ �y;p   ™ ¢  r=+$   ÿd~ â %v Z  ? Ž È $x==$   ÿŽ â r)   ÿàõ Æ/ $x   =Ž Ê r'   ÿä â �y$   • h�y=   ™ h �y>(   ™ h �y?    • h �yL   ™ b r-   ÿàõ Æ �yAp   ™ h �y@H   • b r!   ÿä Æ �yBl   • b r
-   ÿàõ Æ �yP   ™ b r!   ÿä Ê �yP   ™ b x   ÿàÿ È |    pbô ä x--   ÿà÷ ä x@   ÿàÿ ä x##@   ÿàÿ â $r!ÿÿ   !Ž â xD5ð   ÿÀŽ È� xDÿ   D  ä xE6ð   ÿÀŽ äx ð   ÿÀŽ äO xFÿ   *  ä xÿ     â $xD   DŽ â xEÿ   E  Ä xC*ÿÀŽ ä xFÿÀŽ ä rEE   ÿ@Ž ä r   ÿ@Ž ä r   ÿ€Ž â &rDC4   ÿ   â x55   ÿÀŽ Ä x    ÿÀŽ â/ 0r ÿ7    A   â x66   ÿÀŽ â 0r7ÿ1    A   â x/ÿÀŽ â $x   Ž â r1   ÿ€Ž â &rE,   ÿ   ä &r4   ÿ   ä $rDD5   ÿŽ Ä &rJ4   ÿ   â Er        â $rFE6   ÿŽ ä &rH,   ÿ   â ErE D       â $rJJ   ÿŽ â ErG F       â &r414   ÿ   â ErI H       â #rF    ÿ   â#rD E   ÿ   â ErK J       Ä xH.ð   ÿÀŽ â #rE7I   F   âEr 4       â #r47G   D   â xI2ð   ÿÀŽ ä xGÿ   /  ä xF0ð   ÿÀŽ â #rD K   ÿ   â xJÿ   H  Ä xKÿ   I  ä xHÿ   F  ä xGGÿÀŽ ä rFJJ   ÿ@Ž â 0rIÿ9    A   â x..   ÿÀŽ â $xMK   KŽ â rJHH   ÿ@Ž â &rKG,   ÿ   â r9FF   ÿ€Ž â 0rHÿ3    A   â r3JJ   ÿ€Ž â $rKK.   ÿŽ â rFMM   ÿ€Ž â &rM9,   ÿ   Ä &rJ38   ÿ   â ErK K       â #r     ÿ   â ErM M       ä ErJ J       â #rK7K   D   âxÿ   *  â #rM7M       â xDÿ   *  ä x ÿ   /  Ä x*ÿ   /  â #rJHJ   E   â x/:ð   ÿÀŽ ä xE<ð   ÿÀŽ äx7<   ÿÀŽ â &r,F;   ÿ   â xÿÀŽ ä x<DÿÀŽ Ä xDÿ   /  â &rL8   ÿ   â x00   ÿÀŽ ä x/ ÿÀŽ ä x ÿ   E  â $xED   DŽ â Er, ,       â $rLL0   ÿŽ â x::   ÿÀŽ Ä rEEE   ÿ€Ž â #r,I,   J   â &rJ/8   ÿ   â ErL L       ä x**ÿÀŽ â $rJJ:   ÿŽ â rD     ÿ@Ž â &r8E8   ÿ   â #r4HL   4   â x22   ÿÀŽ â &rL<;   ÿ   ârDDD   ÿ€Ž â &r *;   ÿ   â ErJ J       Ä Er8 8       â $rLL2   ÿŽ ä $r  7   ÿŽ â #rJHJ   K   â &r;D;   ÿ   â #rMH8   M   â &rH=   ÿ   â
ErL L       â &r8C=   ÿ   â ErK         â $rH   ÿŽ â Er  ;       â &r;>   ÿ   Ä $r558   ÿŽ ä &rGG>   ÿ   ä &r=   ÿ   â#r4IL   4   â &r11=   ÿ   â #rIK   J   â#rII    M   â &r>   ÿ   ä0rÿ    A   â Er        â $r;6;   ÿŽ â Er5 5       â $r..G   ÿŽ â Er        â 0r ÿ?    A   â Er 1       â &r99>   ÿ   â Er6        Ä Er; ;       â #r    ÿ   âEr. .       â #r5 5   ÿ   â#r    ÿ   âEr8 9       â &r   ÿ   â#r    ÿ   â &r//   ÿ   ä&r33   ÿ   â #r ;   5   â#r6      â#r.      â $r0   ÿŽ â #r8      â $r/:/   ÿŽ Ä &rEE   ÿ   â Er 3       â &r**A   ÿ   ä&rFFA   ÿ   ä&r<A   ÿ   ä 0r@ÿ@    A   â Er        â $r*7*   ÿŽ â Er/ /       â $r2   ÿŽ â ErF F       â 0rBÿB    A   â #r@      â &rADA   ÿ   â #r@       âErE E       â #r@/      â Er* *       â #rFBF      â Er         â 0rÿ  0 A   âErA A       â #rE@E      â #r*B*      â#rB       â 0rÿ    A   â  r,     A  â  rFF     A  â 0r ÿ  0 A   â#rABA   E   â #r4     â #r   F  â 0rÿ    A   â  rII      A  â  r  A     A  Æ #rI   I  â #r *      â !r((       â !r         â !rI       â !r        â G©  @ðÿÿÿÿƒ ê Ay        € ê� ˆ ( þÿ H   è ˆ €þÿ H   è ˆ   ÿÿ H   è ˆ €ÿÿ H   è {          ì M        € ê „y     H   " z  r  p`ð Æ „y	 €   H   b r ÿ   p"p Æ „y    H   ¨ „y     H   è „y €  H   ( „y    H   h „y    H   h „y €  H   h „y    H   h „y €  H   h „y €  H   b !r(       Æ „y
 €  H   " !r		       â/ !r        âO !r         ä� !r	       â!r       â!r         ä !r       Æ ‰       b !r       Æ ‰      ¢ !r       Æ ‰      â !r       È !r
       â !r        È/ ‰       " !r       ÆO ‰      b !r       Æ� ‰      ¨ ‰	      è y      '   " !r        â !r       È/ ‰      " !r       ÆO ‰ €    b !r
	       Æ� ‰	 €    ¨ ‰
 €    â !r       â !r       È/ ‰ €    " !r		       ÆO ‰  @    b !r
       Æ� ‰	 @    ¨ ‰ @    â !r       â x        â !r        â/ x    ÿàÿ â !r	       ÆO ‰      " $z x  Ž â z   r  p`ò â !r       â� ‰      b $z  ~  Ž âx    p"ò ä ‰ @    ¤ 
  r  ÿàõ Ä ‰	      ä 
ÿÿ   ÿä ä 
 j  ÿ„ È 
 k  
 â !r       â %v  j   Ž â !r       È/ †	     â !r       ÆO †    â !r		       Æ� ‰      h  †	 	    â M™        € ê ¹z  r      â !r       â? �x   ?àÿ Ì |     ÿàñ È rÿÿ   ÿä  ä z  j  ÿ€ È z  k    Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y
      &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž ä $rÿÿ   
 Ž È % \   Ž Ê �	    ™ b $ÿÿ   ÿ Ž â ¹z  l      â Ey  ð	    € â y      "   b '
  m   Ž â ™x?    ä y      !   ¢ $‚ÿÿ   ÿ Ž â 
 ÿ n     â $rÿÿ   ÿ Ž È $ ÿÿ    
Ž È $
  o  Ž È 'Š  s   Ž â r ÿ   pRò ä/r ÿ   pRô är   ÿ(Ž ÈO xÿ     È r   ÿ � È xÿ     È |    pbö ä Šÿ t     Ö G9  °    € ê y      '   " xðÿÿÿÿÀŽ â $rÿÿ   ÿ Ž â r ÿ       â y      %   d $x   
Ž Ê x ÿ     È r     ÿ� â 'z y   Ž È $z }  ÿŽ â z	ÿ z    ä x øÿÿÿÿÀŽ ä x ÿ      â $z		 |  Ž ä $x   
Ž ä $z v  	Ž Æxÿ     â $z p  Ž ä/ $z w  ÿŽ â r	   ÿ� â %xT   ÿŽ Æ r		   ÿò� ä z X  ÿàñ â%x   Ž â xÿ   	  äz	 X  	àë ä xP   ÿàù ä z X  ÿà÷ Ä z Y  Äþ ä zÿ Y  ~ â $x    ÿ Ž â x   ÿàý ä x	   ÿàñ ä r   ÿàû â $x   ÿ Ž â x ÿ     Ä zÿ Y  d~ â $rÿÿ     â rÿÿ   ÿäÿ ä xÿ     ä x      ä r   à± â $t	ÿ$   ÿ Ž â �yþÿÿ ‘ ¤  r    ÿd~  â %v Z  	 Ž â �y    ™ æ $x $   ÿŽ â �y    ‘ ( �y üÿÿ ‘ b $x		   Ž â r   ÿ€ Ä �y   ‘ ä  r	   ÿä  ä �y     • ( �y   ™ ( �y(   ™ ( �y$   • ( �yL   ™ ( �y"p   ™ ( �y!H   • ( �y#l   • " $rÿÿ    Ž Ä $rÿÿ    Ž Ê �y$    ™ " x   ÿàÿ È |    pbð ä x   ÿàù ä x   ÿàû ä x   ÿà÷ â/ $rÿÿ    â x@   ÿàÿ â $rÿÿ   Ž â rÿ   ÿäÿ Ä xð   ÿÀŽ ÈO x	ÿ     ä x% ð   ÿÀŽ Èx%ÿ   %  ä xÿÀŽ Æ� $x%%   %Ž â r		   ÿ@Ž ä xÿ     â $x	%   %Ž â x%    ÿÀŽ â &r    ÿ   â	xÿÀŽ â $x   Ž ä &r		   ÿ   Ä $r% %   ÿŽ â x   ÿÀŽ â &r   ÿ   âEr	 	       â 0r ÿ     A   ä &r   ÿ   â Er %       â 0rÿ    A   â Er        â $r   ÿŽ â #r	 	   ÿ   âxð   ÿÀŽ â #r    ÿ   Ä #r    	   â x	ð   ÿÀŽ ä xÿ     ä xÿ     ä Er        â $x   Ž â x	ÿ   	  ä xÿÀŽ Ä xÿ     â #r      â r			   ÿ@Ž â $x   Ž â x   ÿÀŽ â &r   ÿ   âxÿÀŽ â $x		   	Ž â x   ÿÀŽ â &r   ÿ   ä $r   ÿŽ Ä &r"   ÿ   â Er        â &r		"   ÿ   ä 0r!ÿ!    A   ä $r   ÿŽ â Er        â 0r#ÿ#    A   â Er 	       â #r !       âEr	        â #r!      â 0rÿ$  0 A   â#r #       â0r$ÿ$    A   â #r	#	      Ä  r       A  È #r 	$      È !r        â G‰  àùÿÿÿÿƒ ê Ay        € ê ˆ €ÿÿ H   è {          ì M)        € ê „y      H   " r ÿ   pRð Æ „y €   H   h „y    H   ¢ !r         È !r         È/ !r         ÊO ‰       $ !r        Ê ‰      $ !r       Ê ‰ €    $ !r       Ê ‰ @    $ !r       Ê ‰      b  M	        € ê y      %   ¨ y      '   â $z

 x  Ž äO $tÿ   ÿ Ž ä $z ~  
Ž â� !r       Æ/ %v j   Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž È % \   Ž Ê �	    ™ b $rÿÿ   ÿ Ž â Ey  ð    € â $rÿÿ   ÿ Ž â y      "   b 'Š s   Ž Æ y      !   b '
  m   Ž Æ y      %   ¢ $rÿÿ   ÿ Ž â 
 ÿ n     ä rÿ   ÿü‚ ä/ r   ÿ(Ž È x	ÿ     È r		   ÿ � ä Šÿ t    â G       € ê z ÿ `  pPô â_ $rÿÿ   ÿ Ž â z ÿ ^  pPò ä z ÿ b  p ö ä z ÿ a   Sô ä z ÿ _  Sò Ä z ÿ c  0't Ö       '   "        ä¢
       â ©      '   d $ x  ÿŽ â ¨ÿ     â $ª
 x  ÿŽ ä $ ~  ÿŽ Ä $ª ~  ÿŽ â/ ÿ     ä   àë ä¢   àÇ ä ÿÿ   Äþ ä    ÿàý ä ¢
   ÿàû â $rÿÿ   ÿ Ž â ¢ÿÿ   „þ â $ÿÿ    â  ^  ÿ† Ä ¢ÿ   ÿäÿ ä ª
 b  ÿˆ ä  _  � ä ª
 c   Æ �     h �©     d Ay        € ê_ ¹z  l      â xÿ   	  â ™x?    â $ÿÿ    
Ž â Ey  P    € â r! ÿ       â $rÿÿ   ÿ Ž Ä |    pbò â $
  o  Ž ä$‚ ÿÿ    Ž Ô G  à    € ê y      '   " x		ðÿÿÿÿÀŽ â $rÿÿ   ÿ Ž â z ÿ `  pPð â $z  w  ÿŽ â r	  €ÿàÿ â $rÿÿ   ÿ Ž â r! ÿ       â $x   ÿ Ž â xÿ     ä z ÿ a   Sð Ä r   ÿ� È x	øÿÿÿÿÀŽ Ê $x   	
Ž ä 'z y   Ž Æ x ÿ     â $z }  ÿŽ â zÿ z    â $tÿ   ÿ Ž â x      ä r   ÿàó â $z |  Ž Æ rÿÿ   ÿäÿ  â $z v  Ž âxÿ     Æ $z p  Ž â r   ÿ� Æ %xT   ÿŽ â r	   ÿò� Æ %v X  Ž â r        ä xÿ     ä z X  ÿàõ ä xÿ     â $x   ÿ Ž Æ z  Y  ÿä ä xÿ     ä r
   à£ â $tÿ$   ÿ Ž Æ r    ÿDþ  â %v

 Z   Ž â r   ÿàó ä r   ÿàõ â $x $   ÿŽ ä $x   Ž  â r
   ÿ‚ ä r   ÿàÿ â $x     â �y&   ™ ¦  $x	   Ž  â �y     ‘ è �y#   ™ ( �y'   ‘ h �y"(   ™ h �y*
    • h �y)
$   • h �y(   ‘ h �y+   ‘ h �y$L   ™ b z X  ÿàó Æ �y%p   ™ b z Y  ÿäÿ  Æ �y,
H   • h �y-
l   • h �yP   ™ b x   ÿàÿ ä xÿ   &  ÄO x&ÿÀŽ ä x	ÿÀŽ ä/ x    ÿÀŽ â� &r#   ÿ   âx'   ÿÀŽ Æ$r   ÿŽ â x ð   ÿÀŽ â &r		"   ÿ   â x''ð   ÿÀŽ ä xÿ     â $r	   ÿŽ â Er        â 0r ÿ*    A   â x'ÿ   '  â 0r)ÿ)    A   â Er        Ä #r    ÿ   â r   ÿ@Ž â $x''   'Ž â x
(ð   ÿÀŽ â #r)      â xÿ   &  â $x   Ž â r	''   ÿ€Ž ä x
ÿ   
  â &r#   ÿ   â x+ð   ÿÀŽ â &r		"   ÿ   â xÿÀŽ â $x

   
Ž â xÿ     Ä Er        ä Er 	       â $x   Ž â x&ÿ   &  â #r*    ÿ   â r	

   ÿ€Ž â &r$   ÿ   âx((   ÿÀŽ â $x   Ž â x&&ÿÀŽ â &r		$   ÿ   â #r)   *   â $r(   ÿŽ â x++   ÿÀŽ â &r&&%   ÿ   â Er	 	       â &r%   ÿ   â Er        â 0r(ÿ,    A   Ä $r&&+   ÿŽ â Er
        â 0r'ÿ-    A   â #r(	      â#r(      â Er &       â 0rÿ  0 A   â#r'
      â0r
ÿ    A   ä #r'      â  r&     A  â G‰  `    € ê $tÿ   ÿ Ž â z `  à£ Æ %v `  Ž â z a  Dþ  È r   ÿàó â �y+     ¦ r   ÿäÿ  â �y*    è �y    ( �y,    h �y-    b z `  ÿàó È z	 a  ÿäÿ  Ê �yP    b xÿ   +  äO xÿ   *  Æ� $x	   Ž â xÿÀŽ â$x   Ž Æ r			   ÿ€Ž ä x++   ÿÀŽ â &r#   ÿ   âxÿ     â $x   Ž ä &r		#   ÿ   â x#ÿÀŽ â $r+   ÿŽ ä &r"   ÿ   âEr	 	       â &r##"   ÿ   â Er        Ä Er        â #r" 	   ÿ   âxÿ   ,  â#r     ÿ   â x	ÿ     â #r)   "   â xÿ   -  â $x   Ž â x**   ÿÀŽ ä x		ÿÀŽ â $x   Ž â xÿ     Ä r   ÿ€Ž â $r##*   ÿŽ â x,,   ÿÀŽ â &r		$   ÿ   âxÿÀŽ â &r$   ÿ   ä $x   Ž â x--   ÿÀŽ â $r,	,   ÿŽ â Er# #       â &r%   ÿ   â Er        â &r%   ÿ   â Er, ,       â $r-   ÿŽ â #r#)#       â #r (      âEr        â 0rÿ  0 A   â Er        â #r,(,   #   Ä #r '       â0rÿ    A   â #r'   ,   ä  r       A  È #r       È !r!!        â |    pbò â #r

   &  â x   ÿàõ ä x@   ÿàÿ â !r
       â rÿ   ÿä Î G™  põÿÿÿÿƒ ê Ay        € ê r ÿ   pRð âEy  €     € â r ÿ   pRò Ö G‰  P     € ê z ÿ `  pPð â $x    Ž Æ z ÿ a   Sð ä ˆs €ÿÿ H   ö ˆ !   H   ä Ay        € ê {          ì M        € ê z ÿ `  pPð â „y      H   f z ÿ a   Sð Ú „‰ €   H   ¨ „‰    H   â !r        È? !‚       ÈO !‚       â� G‰        € ê „y  €  H   ( „y €   H   h „y    H   ¨ „y    H   è „y €  H   "!r  !       â !r       Æ/ !r         âO !r       Æ� !r!        ä‰      " z ÿ `  pPð ä r ÿ   pRò ä z ÿ a   Sð â !r       Ê ‰       $ !r         Ê ‰  €    $!r        Ê ‰ @    $ !r       Ê ‰      $ !r       â G‰        € ê ‰ !      $ !r  !       Ê ‰       $ !r        Ê ‰ €    $ !r       Ê ‰ @    $ !r       Ê ‰      $ !r!       ä M        € ê !r       â G‰  `    € ê $v ÿ h  ÿ Ž â !r!       È x     pBð Ú G	  p    € ê x  ÿÿÿÿÿàÿ È x       € È x     ÿ   È ‚{   €     $ xÿ     È Iy �êÿÿÿÿƒ ê  x ;ª¸¿  A  Ì s          $ !t   €?    Ì s        $  r     A  È  r     A  â Gy  P    € ê  x '7=  A  â x*BL?  A  Æ #t   €?    È  r      A  â  x    ?  A  Ê s     $   $ !t  €?    È  r       A  È  r      A  â Gy  °    € ê x     pRð Ú G‰  à     € ê x     pRð Ú G	        € ê 	z  i    � ä 	z i    � Æ  x ;ª¸¿  A  â 	z i €  � Ê s        $ !t  €?    Ì s        $  r       A  È  r      A  â Gy  Ð     € ê  r     A  â Gy  °     € ê 	x   à@  � ä 	x  à@  � Æ  x #ÛÙ?  A  â 	x  àÀ  � Æ  x;ª¸¿  A  ä !t  €?    È s        $ !t  €?    Ì s        $  r       A  È  r      A  ä y      &   ( y      '   b $z x  Ž ä $tÿ   ÿ Ž ä $z ~  Ž È/ %v j   Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú         â $rÿÿ    Ž È % \   Ž Ê �	    ™ b ¹z  l      â Ey       € â ™x?    â y      "   b $rÿÿ   ÿ Ž â x      ÿ  â r! ÿ       â y      !   ¢         â $ÿÿ   ÿ Ž Ä y      %   $ '
  m   Ž È $‚ÿÿ   ÿ Ž â 
 ÿ n     Ê $ ÿÿ    
Ž â r ÿ   pRò Æ/ $
  o  Ž â r    ÿ(Ž ÈO ‚        ä xÿ      ä x   ÿ   â 'Š s   Ž â r    ÿ � ä y      '   $ x(ÿ     È | (   pbô ä Šÿ t    ä r ÿ   pRð â$€   Ž Ð G)  0    € ê xðÿÿÿÿÀŽ â $z
 w  ÿŽ â r ÿ       â $z }  ÿŽ âz)  X      â $x    
Ž â r ÿ       â 'z y   Ž â r
   ÿà÷ ä x ÿ     â $v#ÿ Y  ÿ Ž â z	ÿ z    â $z p  (Ž â r     ÿ� â $x(   ÿ Ž Ä $z		 |  Ž â x øÿÿÿÿÀŽ â $r%ÿÿ   ÿŽ â x"ÿ      â $z v  	Ž ä$x   
Ž â x""   ÿ   È xÿ     äx'ÿ     ä r   ÿ� ä x&ÿ   "  ä r   ÿò� â %xT   ÿŽ â x'   '  È r$   ÿàõ â %x   Ž Æ r    ÿ ä r"   àÇ ä r
)$   ÿàû ä r &%   ÿ„þ ä r#    ÿäÿ ä x $       â $x $   ÿŽ â r)   ÿàõ â �y 
    ‘ ¤  %v Z   Ž Ä �y4
   ‘ ä  $x	   # â r   ÿ„ â $x   Ž â �y1
   ‘ ( r'   ÿä â �y-   ™ h �y,   ™ ( �y3    • ( �y+(   ™ ( �y/$   • ( �y*L   ™ ( �y0
   ‘ " $v5ÿ p  ÿ Ž Æ �y2p   ™ " %x5T   
Ž Æ �y.H   • " %x5T   Ž Æ �y7l   • ( �y:    ‘ ( �y6   ™ ( �y9   ‘ ( �y8   ‘ " %x
5T   Ž Æ �y;   ‘ " %x5T   Ž Æ/ �y@
   ™ " r)   ÿàõ Æ �y=    ‘ " r#   ÿä Æ �y>   ‘ ( �yAP   ™ ( �y?   ‘ ( �y<   ‘ " %x5T   Ž È %x5T   
Ž â�yP   ™ & %x
5T   Ž ä/�y   ™ ( �yC
    ‘ ( �yB
   ‘ ( �y
   ‘ ( �y
   ‘ " %x5T   Ž Ê �yP   ™ " %x5T   Ž Ì �yP   ™ " x ð   ÿÀŽ ÈO xÿ     ä x54ð   ÿÀŽ ä� x    ÿÀŽ â $x    Ž â x5ÿ   5  Æ $x     Ž â/ r555   ÿ@Ž ä x ÿ   -  är55   ÿ€Ž â &r,   ÿ   â	x-ÿÀŽ ä x
 ÿÀŽ â 0r ÿ3    A   â Er        â &r,   ÿ   Ä &r+   ÿ   âxÿ   -  â &r
+   ÿ   â #r
    ÿ   â 0r/ÿ/    A   â Er        â $r   ÿŽ â xÿÀŽ ä x1ð   ÿÀŽ â #r/   
   â x44   ÿÀŽ â &r
*   ÿ   â xÿ     Ä x0ð   ÿÀŽ â $r44   ÿŽ ä $x   Ž â xÿ     È r   ÿ€Ž â $x   Ž â x11   ÿÀŽ ä x-ÿ   -  â &r*   ÿ   â r   ÿ€Ž ä Er        â $r

1   ÿŽ â x--ÿÀŽ â &r2   ÿ   â Er4 4       â 0r.ÿ.    A   â Er        â #r    ÿ   â x00   ÿÀŽ â &r--2   ÿ   â Er
 
       â #r/4      â 0r7ÿ7    A   â Er        â #r.      â $r--0   ÿŽ â x0:   ÿÀŽ â #r.
      â xÿ   6  â #r
7      â x::ð   ÿÀŽ Ä x9ð   ÿÀŽ ä Er- -       ä x6ÿÀŽ ä x:ÿ   :  ä xÿÀŽ ä xÿ   6  Ä x8ð   ÿÀŽ ä xÿ     â #r7-      â &r,   ÿ   â xÿÀŽ â $x::   :Ž â xÿ     â &r-*   ÿ   â r   ÿ@Ž â $r0   ÿŽ â x99   ÿÀŽ â $x:   :Ž â r0   ÿ€Ž â &r+   ÿ   Ä $x   Ž ä &r,   ÿ   â x88   ÿÀŽ â $r9   ÿŽ â Er        â &r00+   ÿ   â r   ÿ€Ž ä Er        â $r-8   ÿŽ â Er        â #r    ÿ   â &r*   ÿ   â x6ÿ   6  Ä Er 0       â #r0    ÿ   â #r/      â x;ð   ÿÀŽ ä Er        â #r/   0   â x-6ÿÀŽ ä Er        ä xÿ     â #r0.      âx4;   ÿÀŽ â &r--2   ÿ   â #r.      â x@ÿÀŽ Ä xÿ   @  â $x   Ž ä $r--4   ÿŽ â x4=   ÿÀŽ â &r,   ÿ   â x6ÿÀŽ ä r   ÿ€Ž â $r44   ÿŽ â x>   ÿÀŽ â &r66+   ÿ   Ä &r2   ÿ   ä 0rÿA  0 A   ä$r66   ÿŽ â Er 4       ä x==ð   ÿÀŽ â 0rAÿA    A   â Er        â  r
     A  â Er6 6       â #r
    ÿ   â x>>ð   ÿÀŽ Ä x=ÿ   =  â #rA     â #r7      â #r/6   
   â x>ÿ   >  â $x==   =Ž â xÿ   @  ä x
?ð   ÿÀŽ â $x==   =Ž â r>>>   ÿ@Ž Ä x@ÿ   @  ä xÿÀŽ ä xÿ   
  ä Er -       â $x>>   >Ž â x
?   ÿÀŽ â &r==,   ÿ   â r   ÿ@Ž â &r*   ÿ   â x@ÿÀŽ â #r7   0   â r   ÿ€Ž â &r>>+   ÿ   â x0<   ÿÀŽ â $r
   ÿŽ â Er= =       â &r2   ÿ   â Er> >       â &r-*   ÿ   Ä $r10   ÿŽ â #r
 =   ÿ   â Er        â 0rÿ  0 A   â x<<ð   ÿÀŽ â #r/>   
   â Er0 -       â #r
.      â xCð   ÿÀŽ ä x<ÿ   <  â  r     A  â #r0.0      â xÿÀŽ Ä xÿ     â $x<<   <Ž â x-ÿ     â &r,   ÿ   â x4C   ÿÀŽ â $x   Ž â xBð   ÿÀŽ ä r<<   ÿ€Ž ä x6-ÿÀŽ â $r44   ÿŽ â xÿ     Ä r   ÿ€Ž â &r2   ÿ   â xB   ÿÀŽ â &r66+   ÿ   ä &r,,   ÿ   â Er 4       â $x-   Ž â Er        â $r6   ÿŽ â Er ,       â #r,    ÿ   â r8--   ÿ€Ž â #r07   0   â Er        Ä xð   ÿÀŽ â #r     ÿ   ä #r/   ,   â xð   ÿÀŽ â &r88+   ÿ   â xÿ     ä xÿ     ä xÿ     â $x   Ž â Er8 8       Ä xÿÀŽ â $x   Ž â xÿ     ä r+   ÿ@Ž â #r//8       â x    ÿÀŽ â &r*   ÿ   âxÿÀŽ â &r*   ÿ   â r+++   ÿ€Ž â $r     ÿŽ â x   ÿÀŽ â &r2   ÿ   âEr        â &r++2   ÿ   â x((   ÿàÿ â $r   ÿŽ â Er          â #r.   /   â Er+ +       Ä | (   pbô â 0rÿ  0 A   âEr1 1       â 0rÿ  0 A   â Er        â #r .       â #r7+      â0rÿ    A   â #r
71   
   â0rÿ    A   â  r0     A  â 0rÿ    A   â #r 7       â  r     A  â x))   ÿà÷ â #r     â #r
     â #r       â !r!!       â !r       â !r       â !r        â x@   ÿàÿ â $r#ÿÿ   #Ž â D)       À â Gy  Ðìÿÿÿÿƒ ê Ay        € ê ˆ ! þÿ H   è ˆ €þÿ H   è ˆ  ÿÿ H   è ˆ €ÿÿ H   è {          ì M	        € ê „y      H   " z  r  p`ò Æ „y €   H   b r ÿ   pVò  Æ „y    H   ¨ „y    H   è „y	 €  H   ( „y    H   („y €  H   ( „y    H   ( „y €  H   ( „y    H   ( „y €  H   " !r  !       Æ „y €  H   " !r       â/ !r       âO !r         ä� !r	       â!r       â !r       â !r         â !r       È ‰       b !r       Æ ‰      ¢ !r       Æ ‰      â !r       Ê ‰
      " !r        â/ !r       ÈO ‰      b !r       Æ� ‰      ¨ ‰	      â !r 
       Ê ‰       " !r       â/ !r       ÈO ‰ €    b !r
	       Æ� ‰	 €    ¨ ‰
 €    â !r        Ê ‰ €    " !r       â/ x   ÿàÿ â !r	       ÆO ‰  @    b z  r  p`ô â !r
       ä� ‰ @    ¢ x
   ÿàÿ ä x    pVt â ‰	 @    â !r       â z 
 r  p`ð Ä x        ä ‰ @    " x    pVp  â !r         â/ x   ÿàÿ â !r       ÆO z  r  p`ö â $z x  Ž â !r	       â� ‰      b x    pVö â $z ~  Ž ä ‰	       ¢ !r       â %v j   Ž ä ‰      ( ‰      â !r       â/ !r	 	       ÈO †©    â !r       Æ †™ 	    â !r       Æ� †¹    â M	        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      &   " z ÿ \  pPò â ¹z  F   
   Æ z ÿ ]  Sò Ú         Ê % \   Ž Ê �	    ™ ¢  r ÿ       â Ey      € â x       ÿ  â y      "   d '  m   Ž äy      %   ä 'š s   Ž â  ÿ n     â y      !   &  ÿ   ÿáÿ â y      '   ( $ o  Ž â ’        Ä r ÿ   pRð ä/ x   ÿ   ä� x    pDp  är   ÿàÿ är    ÿ(Ž ä z  r  pdp  Ä xÿ      ä š	ÿ t    È r	     € ÊO G	  °    € ê z ÿ `  pPò â x       ÿ  â z ÿ ^  pPð ä z ÿ b  p ô ä z ÿ a  Sò ä z ÿ _   Sð ä z ÿ c   'ò Ö $
 ~  ÿŽ âÿ     â$š ~  ÿŽ ä $
 x  ÿŽ â   àë â$š x  ÿŽ â ’
   àÇ ä ˜ÿ     ä ÿ   ÿÄþ Ä    ÿàõ ä ’
   ÿàû ä ’
ÿ   ÿ„þ ä ÿ   ÿä ä 
 ^  ÿ„ ä ’ÿ
   ÿäÿ Ä š b  ÿ† ä
 _   ä š c  � Æ �	!     h �™      d Ay        € ê r    ÿ � â ¹z  l      â $z }  ÿŽ â™x?    â x,ÿ     â $z w  ÿŽ â r ÿ       â Ey  Ð(    € â r# ÿ       â x$      ÿ  â | ,   pbð â 'z y   Ž â r"   ÿàó â x&      ÿ  â x(      ÿ  â r> ÿ       Ä r*ÿÿ   ÿäÿ  Ì G	   (    € ê xðÿÿÿÿÀŽ â $z p  ,Ž âzÿ z    â x$      ÿ  â r   €ÿàÿ â x&      ÿ  â x-   ÿàÿ â $z |  Ž âx ÿ     â $z |  ,Ž â x   ÿàÿ â $z	 v  Ž âr     ÿ� â $z	 v  Ž â x/   ÿàÿ â %x
T   ÿŽ â x øÿÿÿÿÀŽ â x(      ÿ  â x0ÿ      â $z-- p  Ž âr+  €ÿàÿ â $z. p  Ž âz ÿ `  pPð â $z// p  Ž â xÿ   +  â %x   ÿŽ â x00   ÿ   Ä r+   ÿ� â %x   
Ž â x ÿ   +  ä r   ÿò� â %x-T   Ž â z1  `      ä xÿ     â %x/T   Ž â r2
   ÿàó Ä z3  a      â %x-T   Ž âx42   ÿàõ ä r6   ÿäÿ  â %x/T   Ž â z5  X      ä z7  Y      â %x.T   Ž â r# ÿ       Ä r> ÿ       â %x.T   Ž â z ÿ a   Sð ä x9,   ÿ   ä x;ÿ   0  ä x=+      ä r8ÿ6   ÿä Ä r0"   9àµ ä r52   ÿàù Ä x $       ä r ;*   ÿd~ ä r76   ÿä â %v Z   Ž â r5   ÿàó Æ $x $   ÿŽ â �yF    ‘ ¢  r7   ÿäÿ  ä r+   ÿ‚ â �y    ‘ â  r   ÿàÿ Æ �yD   ™ " r=   ÿäÿ  Æ �yA   ‘ ( �y:   ™ " r54   ÿàó Æ �y<(   ™ " zK  p      Æ �yH    • " r78   ÿäÿ  â %xKT   Ž ä/�yE$   • $ %xKT   Ž ä �yJ   ™ ( �yG   ‘ ( �y?L   ™ ( �yIúÿÿ ‘ ( �y@p   ™ ( �yMH   • ( �yLüÿÿ ‘ ( �yQl   • ( �yPþÿÿ ‘ ( �yN    ‘ " rB
5   ÿàó È rC7   ÿäÿ  Ê %xKT   BŽ â�yOBP   ™ ( �yRP   ™ " %xKT   Ž Ê �yUúÿÿ ‘ "	 %xKT   Ž Æ �yVüÿÿ ‘ (	 �yW   ™ " %xKT   Ž Æ �yTþÿÿ ‘ "	 %xKT   Ž Æ/ �y   ™ ( �yXúÿÿ ‘ ( �yS    ‘ (	 �yCüÿÿ ‘ ( �yBþÿÿ ‘ ( �yY    ‘ " %xKT   Ž Ê �yZP   ™ " %xKT   Ž Ì �yP   ™ " xFð   ÿÀŽ äO x ð   ÿÀŽ ä� xÿ     ä xÿ   D  äxDÿÀŽ ä/ xÿ     Ä r   ÿ@Ž ä xFF   ÿÀŽ â &r:   ÿ   â xÿÀŽ ä r   ÿ@Ž ä r   ÿ€Ž â &r<   ÿ   â x     ÿÀŽ â $rF   ÿŽ â r   ÿ€Ž â &r:   ÿ   â xAð   ÿÀŽ â $r    ÿŽ â Er        â &r<   ÿ   â Er        â 0r ÿH    A   â xÿ     Ä xÿ   D  â0rEÿE    A   â Er        â #r    ÿ   âEr        â #r    ÿ   â r   ÿ@Ž ä xÿÀŽ ä xDÿ   D  â #rE      âr   ÿ€Ž â #rE      â xJÿÀŽ â &r?   ÿ   â xAA   ÿÀŽ Ä xGð   ÿÀŽ ä xDÿÀŽ ä xÿ   J  â &r?   ÿ   â xDI   ÿÀŽ â &r:   ÿ   â xGG   ÿÀŽ â $rA   ÿŽ â xÿ     â &r@   ÿ   â xÿÀŽ â $rD   ÿŽ â r   ÿ@Ž â 0rÿM    A   â Er        â $rG   ÿŽ â xDL   ÿÀŽ â &r<   ÿ   â Er        â #r      âr   ÿ€Ž â $rD   ÿŽ â Er        â 0rÿQ    A   â Er        â #r      â &r@   ÿ   â Er        â #r    ÿ   â #r      â xÿ   J  Ä Er        â #rE      â xIIð   ÿÀŽ ä xÿÀŽ â #r      â xLLð   ÿÀŽ ä xIÿ   I  â &r?   ÿ   â xPð   ÿÀŽ Ä xP   ÿÀŽ ä xLÿ   L  ä rIII   ÿ@Ž ä xÿ     â $r   ÿŽ â xNð   ÿÀŽ ä rLLL   ÿ@Ž Ä rIII   ÿ€Ž ä r   ÿ@Ž ä xJÿ   J  ä xÿ     â &rII:   ÿ   â rLL   ÿ€Ž ä r   ÿ€Ž Ä xJÿÀŽ ä rA   ÿ@Ž â &r<   ÿ   â ErI I       ä rGAA   ÿ€Ž â &rA@   ÿ   âxNN   ÿÀŽ â &r?   ÿ   â Er        â #r I   ÿ   â &rGG@   ÿ   â Er        â $rAAN   ÿŽ â Er        â #rE      â ErG G       â #r      âErA A       â 0rÿO  0 A   â #r      â 0rÿR  0 A   Ä #rAA      â#rG      â 0rÿO    A   â  r     A  ä  r     A  ä #r     â xUð   ÿÀŽ ä xVð   ÿÀŽ Ä xÿ     â 0rRÿR    A   â xWÿÀŽ ä xÿ     ä r   ÿ@Ž â #rAR     â xU   ÿÀŽ â &r:   ÿ   â r   ÿ@Ž Ä r   ÿ€Ž ä xÿ   W  â$r   ÿŽ â r   ÿ€Ž â &r:   ÿ   â xÿÀŽ ä xÿ   W  â &r<   ÿ   âxVV   ÿÀŽ â &r<   ÿ   â xÿÀŽ Ä Er        ä Er        â $rV   ÿŽ â xDT   ÿÀŽ â &rA?   ÿ   â Er        â #r    ÿ   âxWÿ   W  â #r    ÿ   â $rAAD   ÿŽ â xÿÀŽ â #rE      â Er        Ä xWWÿÀŽ ä xÿ     â &r:   ÿ   â xDX   ÿÀŽ â #rE      â ErA A       â &rWW@   ÿ   â xS   ÿÀŽ ä xÿÀŽ â $rD   ÿŽ â #rA      â $rW   ÿŽ â xDC   ÿÀŽ â &r<   ÿ   â xXXð   ÿÀŽ Ä Er        â $rD   ÿŽ â Er        ä xÿ     ä xXÿ   X  â #r      â Er        â #r    ÿ   â xÿÀŽ Ä xCCð   ÿÀŽ ä rXXX   ÿ@Ž â #rE      â xTTð   ÿÀŽ â &r?   ÿ   â xAB   ÿÀŽ äxBð   ÿÀŽ ä xCÿ   C  Ä rXX   ÿ€Ž â $rA   ÿŽ â xSSð   ÿÀŽ ä xTÿ   T  â &r:   ÿ   â xÿ     ä rCCC   ÿ@Ž ä xAYð   ÿÀŽ Ä xSÿ   S  ä rTTT   ÿ@Ž ä r   ÿ@Ž ä rCCC   ÿ€Ž ä xBÿ   A  ä rSSS   ÿ@Ž Ä rTTT   ÿ€Ž ä xÿ     ä Er        â &rCC<   ÿ   â r   ÿ€Ž ä rBBB   ÿ@Ž â &rTT?   ÿ   â rSSS   ÿ€Ž Ä xAÿÀŽ â #r    ÿ   â rBB   ÿ€Ž â &r?   ÿ   â ErC C       â &rSS@   ÿ   â ErT T       â &rAA@   ÿ   âxBY   ÿÀŽ â &r@   ÿ   â Er        â #rCEC      â ErS S       â #rTT      â$rAAB   ÿŽ â Er        â #r   C   âEr        â 0rÿZ  0 A   â #rS   T   âErA A       â 0rBÿ  0 A   â #r      â #r      â  r     A  â 0rZÿZ    A   â #rA      â 0rÿ    A   â  rB     A  â !r&&       â !r%%       â #rZ     â #r     â G‰  @    € ê r12   ÿàó È r36   ÿäÿ  ä r1   ÿàó Æ �y     ¢ r3   ÿäÿ  Æ �yB    è  �y    ( �yF    ( �y    " r
1   ÿàó È r3   ÿäÿ  Ê �yP    " xAÿ     äO x   ÿÀŽ ä xCÿ   B  ä� rAAA   ÿ@Ž ä rDCC   ÿ@Ž ä rAAA   ÿ€Ž Ä xCÿ     ärDD   ÿ€Ž â &rAA:   ÿ   â xCCÿÀŽ ä xÿÀŽ â &r<   ÿ   âErA A       â &rCC<   ÿ   â xDB   ÿÀŽ â &r:   ÿ   â xGÿ     â #rB A   ÿ   â xÿ   F  â/ $rCCD   ÿŽ â Er        â $r   ÿŽ â xDGÿÀŽ Ä xAÿ     â #rE   B   â r   ÿ@Ž â &rDD?   ÿ   â xÿ     ä xF   ÿÀŽ ä rAAA   ÿ@Ž ä r   ÿ€Ž â $rDD   ÿŽ â xÿÀŽ Ä rAAA   ÿ€Ž â &r?   ÿ   â Er        â &r@   ÿ   âx   ÿÀŽ â &rAA@   ÿ   â ErC C       â #r    ÿ   â Er        â $r   ÿŽ â ErD D       â #rCEC      â ErA A       â #r      âEr        â 0rÿ  0 A   â#rDD   C   â #rA      â 0rÿ    A   Ä #r   D   â  r     A  È #r     È !r))       ä G‰  P    € ê r1   ÿàó È r3   ÿäÿ  ä r1   ÿàó Æ �y     ¢  r	3   ÿäÿ  Æ �yB    è  �y    ( �yF    ( �y    $ r 1       Ä r 3       Ê %x-T   Ž Ì �yP    " xAÿ     äO x   ÿÀŽ ä xCÿ   B  ä� rAAA   ÿ@Ž ä rDCC   ÿ@Ž ä rAAA   ÿ€Ž Ä xCÿ     ärDD   ÿ€Ž â &rAA:   ÿ   â xCCÿÀŽ ä xÿÀŽ â &r<   ÿ   âErA A       â &rCC<   ÿ   â xDB   ÿÀŽ â &r:   ÿ   â xGÿ     â #rB A   ÿ   â xÿ   F  â $rCCD   ÿŽ â Er        â $r   ÿŽ â xDGÿÀŽ Ä xAÿ     â #rE   B   â r   ÿ@Ž â &rDD?   ÿ   â xÿ     ä xF   ÿÀŽ ä rAAA   ÿ@Ž ä r   ÿ€Ž â $rDD   ÿŽ â xÿÀŽ Ä rAAA   ÿ€Ž â &r?   ÿ   â Er        â &r@   ÿ   âx   ÿÀŽ â &rAA@   ÿ   â ErC C       â #r    ÿ   â Er        â $r   ÿŽ â ErD D       â #rCEC      â ErA A       â #r      âEr        â 0rÿ  0 A   â#rDD   C   â #rA      â 0rÿ    A   Ä #r   D   â  r     A  È #r     È !r>>       ä G‰  P    € ê r1   ÿàó È r3   ÿäÿ  ä r1   ÿàó Æ �y     ¢  r3   ÿäÿ  Æ �yB    è  �y    ( �yF    ( �y    $ r 1       Ä r 3       Ê %x.T   Ž Ì �yP    " xAÿ     äO x   ÿÀŽ ä xCÿ   B  ä� rAAA   ÿ@Ž ä rDCC   ÿ@Ž ä rAAA   ÿ€Ž Ä xCÿ     ärDD   ÿ€Ž â &rAA:   ÿ   â xCCÿÀŽ ä xÿÀŽ â &r<   ÿ   âErA A       â &rCC<   ÿ   â xDB   ÿÀŽ â &r:   ÿ   â xGÿ     â #rB A   ÿ   â xÿ   F  â $rCCD   ÿŽ â Er        â $r   ÿŽ â xDGÿÀŽ Ä xAÿ     â #rE   B   â r   ÿ@Ž â &rDD?   ÿ   â xÿ     ä xF   ÿÀŽ ä rAAA   ÿ@Ž ä r   ÿ€Ž â $rDD   ÿŽ â xÿÀŽ Ä rAAA   ÿ€Ž â &r?   ÿ   â Er        â &r@   ÿ   âx   ÿÀŽ â &rAA@   ÿ   â ErC C       â #r    ÿ   â Er        â $r   ÿŽ â ErD D       â #rCEC      â ErA A       â #r      âEr        â 0rÿ  0 A   â#rDD   C   â #rA      â 0rÿ    A   Ä #r   D   â  r     A  È #r     È !r((       â !r$$       â !r##       â G‰  P    € ê r1   ÿàõ är1   ÿàó ä r3   ÿä är3   ÿäÿ  Æ �y     ¨  �y    è �yB    ( �y    ( �y    $ r 1       Ä r 3       Ê %x/T   Ž Ì �yP    " xAÿ     äO x   ÿÀŽ ä xÿÀŽ ä� rAAA   ÿ@Ž ä xÿ     â &r:   ÿ   â rAAA   ÿ€Ž Ä xÿ   B  â$r   ÿŽ â xÿÀŽ â &rA:   ÿ   â r   ÿ@Ž ä Er        â &r<   ÿ   â Er        ä r   ÿ€Ž â #r    ÿ   âxÿ     â #r     ÿ   â xÿ     â &r<   ÿ   â xÿ     Ä r   ÿ@Ž ä x:ÿ     ä r<   ÿ@Ž ä xBB   ÿÀŽ ä xÿÀŽ ä r   ÿ€Ž â $rB   ÿŽ â x   ÿÀŽ â &r?   ÿ   âx:ÿÀŽ â &r?   ÿ   â r<<   ÿ€Ž â $r   ÿŽ â Er        â &r@   ÿ   âx   ÿÀŽ â &r@   ÿ   â Er        â #rE       â Er        â $r   ÿŽ â Er        â #rE      â Er        â #r      âEr        â 0r ÿ  0 A   â#r      â #r      â 0rÿ    A   Ä #r      â  r       A  È #r       È !r''        â x,,   ÿàÿ ä x11   ÿàõ ä | ,   pbò ä x55   ÿà÷ ä x99@   ÿàÿ ä r3ÿ3   ÿä Ä r7ÿ7   ÿäÿ Ê D       À â Gy   Ûÿÿÿÿƒ ê Ay        € ê y      "   " Ey  0    € â r ÿ   pRð är ÿ   pRò Ö G‰  ð     € ê z ÿ `  pPð ä r    ÿ8Ž ä z ÿ a   Sð Æ ˆs  & þÿ H   ô ˆƒ  %€þÿ H   è ˆƒ  $ ÿÿ H   è ˆƒ  #€ÿÿ H   â G‰  p     € ê ˆs  )   H   è ˆs  %€þÿ H   è ˆs  >€  H   è ˆs  $ ÿÿ H   è ˆs  (   H   è ˆs  #€ÿÿ H   è ˆs  '€  H   ä Ay        € ê {          ì M        € ê y       &   "> z ÿ `  pPð Ä x        â y      '   b z ÿ a   Sð â $z   x  Ž È $z  ~   Ž È/ %v  j   Ž È G	  P    € ê „y     H   " z  r  p`ð Æ „y €   H   b r ÿ   pVp  Æ „y    H   ¨ „y     H   è „y	 €  H   ( „y    H   ( „y    H   ( „y €  H   ( „y    H   ( „y
 €  H   ( „y €  H   " !r&       Æ „y €  H   " !r%       â/ !r$       âO !r         ä� !r	       â!r       â !r         ä !r       Æ ‰       b !r       Æ ‰      ¢ !r

#       Æ ‰      â !r

       È !r
       â !r        â/ !r       ÈO ‰      " !r
       Æ� ‰      h ‰      ¨ ‰	
      â !r       â !r       â/ !r       ÈO ‰      " !r
	       Æ� ‰ €    h ‰  €    ¨ ‰	 €    â !r       â !r       È/ ‰ €    " !r
        ÆO ‰  @    b !r		       â� x   ÿàÿ ä ‰
 @    ¤ z  r  p`ô ä ‰	 @    ä x    pVt â !r       â !r         È/ ‰ @    " x   ÿàÿ â!r
       âO x   ÿàÿ â ‰       b z  r  p`ò â !r	       ä� ‰      ¢ x    pVò  Æ ‰      â !r       â !r        È/ ‰       " !‚!       â!r       âO !r       Æ� †‰     â z  r  p`ð â !¢	!       â!’!       ä x    pVp  ä †© 	   è †™    â !r         Ì M	        € ê !r! !       Ê †y !   â My        € ê z   h      È x     pBð Ú G	  @    € ê x     pRð Ú G‰  À    € ê x     pRð Ú G	   $    € ê „y	     H   " z  r  p`ð â Ey  p    € ä „y     H   b r ÿ   pVp  Æ „y    H   ¨ „y    H   è „y    H   ( „y  
  H   ( „y €  H   ( „y €   H   ( „y €  H   ( „y €  H   " !r		&       â !r  )       Æ/ !r	       âO !r         Æ� !r       â!r         È ‰      " !r>       Æ ‰       b !r%       È !r       È !r       â !r       â !r        È/ ‰      ( ‰      b !r       â !r	       È/ ‰ €    ( ‰
	 €    h „y €  H   ¢ !r       Æ „y €
  H   â !r
	
       Æ/ ‰ @    ( ‰
 @    b !r       âO !r       â !r
       È/ ‰       ( ‰      b !r       â !r         â !r       â/ G	  �     € ê !r        â!r  !       Æ  x;ª¸¿  A  Ì s        $ !t  €?    Ì s        $  r     A  È  r      A  Ê †y     ä Ay        € ê ‰       b Ey  Ð    € æ ‰      ¨ „y    H   è „y    H   ( „y    H   â !r         Æ/ „y    H   â !r       âO x   ÿàÿ ä ‰       d z  r  p`ð ä ‰      $ x    pVp  â !r$       â� !r(       â!r        Ä/ !r       â !r	       ä ‰ €    ( ‰
	 €    b !r       Æ „y  	  H   ¢ !r       Æ „y    H   â !r
	
       Æ/ ‰ @    ( ‰
 @    b !r       âO !r       â !r
       È/ ‰       ( ‰      b !r       â !r         â !r       â/ G	  �     € ê !r        â!r  !       Æ  x;ª¸¿  A  Ì s        $ !t  €?    Ì s        $  r     A  È  r      A  Ê †y    ä Ay        € ê ‰       b Ey       € æ ‰      ¨ „y €	  H   è „y €  H   è „y €  H   è „y €  H   â !r         Æ/ „y €  H   â !r       ÆO ‰       h ‰      ( „y €  H   ¢ !r        â/ !r	       È ‰ €    ( ‰
	 €    b !r'       ÈO !r       â !r       â x   ÿàÿ ä !r       â !r
	
       â/ ‰ @    " z  r  p`ð Æ ‰
 @    b x    pVp  â !r       â !r
       È/ ‰       ( ‰      b !r         â !r       â/ G	  �     € ê !r        â!r  !       Æ  x;ª¸¿  A  Ì s        $ !t  €?    Ì s        $  r     A  È  r      A  Ê †y    ä Ay        € ê ‰      b !r#       â x   ÿàÿ Æ !r       â z  r  p`ð Æ !r       â x    pVp  È ‰       ¢ !r       Ê/ ‰      " !r         ÊO ‰       b !r	       Ê ‰
	 €    " !r        Ê/ ‰ €    b !r
	
       Ê ‰
 @    " !r       Ê/ ‰ @    b !r
       Ê ‰      " !r       Ê/ ‰       ¢ !r       â M	        € ì !r         â/!r         ÆO  x ;ª¸¿  A  â !r! !       Ê s        $ !t  €?    Ì s        $  r       A  È  r!!      A  Ê †y !   â My        € ê „y     H   " z  r  p`ð â Ey       € ä „y     H   b r ÿ   pVp  Æ „y    H   ¨ „y    H   è „y	    H   ( „y  
  H   ( „y €   H   ( „y €  H   ( „y €  H   ( „y €
  H   â !r&       â !r  )       Æ/ !r       âO !r         Æ� !r		       â!r        È ‰ 	      " !r%       Æ ‰      b !r       È !r       â !r 	        â !r       È/ ‰       ( ‰      h „y €  H   ¢ !r        â !r       È/ ‰ €    ( ‰ €    h „y €  H   â !r>       âO !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       È� !r       â !r       â !r
	
       È/ ‰       ( ‰
      b !r        â !r
       â/ G	  À     € ê !r         È x '7=  A  â x *BL?  A  Æ #t   €?   È  r     A  â  x    ?  A  â !r !       È s     $   $ !t  €?    È  r     A  È  r      A  Ê †y     ä Ay        € ê ‰       b Ey       € æ ‰      ¨ „y    H   è „y  	  H   ( „y    H   â !r         Æ/ „y    H   â !r       ÆO ‰       h ‰      ( „y    H   ¢ !r(       È� !r       â!r        â/ !r       È ‰ €    ( ‰ €    b !r       Æ „y    H   â !r$       âO !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       È� !r       â !r       â x   ÿàÿ â !r
	
       Æ/ ‰       " z  r  p`ð Æ ‰
      b x    pVp  â !r        â !r
       Ö/ G	  À     € ê !r         È x '7=  A  â x *BL?  A  Æ #t   €?   È  r     A  â  x    ?  A  â !r !       È s     $   $ !t  €?    È  r     A  È  r      A  Ê †y    ä Ay        € ê ‰       b Ey  Ð    € æ ‰      ¨ „y €  H   è „y €	  H   ( „y €  H   è „y €  H   â !r         Æ/ „y €  H   â !r       ÆO ‰       h ‰      ( „y €  H   â !r'       È� !r       â!r        Æ/ !r       â !r       â ‰ €    ( ‰ €    b !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       â x   ÿàÿ â !r
	
       Æ/ ‰       " z  r  p`ð Æ ‰
      b x    pVp  â !r        â !r
       Ö/ G	  À     € ê !r         È x '7=  A  â x *BL?  A  Æ #t   €?   È  r     A  â  x    ?  A  â !r !       È s     $   $ !t  €?    È  r     A  È  r      A  Ê †y    ä Ay        € ê ‰      b !r#       â x   ÿàÿ Æ !r       â z  r  p`ð Æ !r       â x    pVp  È ‰       ¢ !r       Ê/ ‰      " !r         ÊO ‰       b !r       Ê ‰ €    " !r        Ê/ ‰ €    b !r	       Ê ‰
	 @    " !r       Ê/ ‰ @    b !r
	
       Ê ‰
      " !r       Ê/ ‰       ¢ !r
       â M	        € ì !r         â/!r         ÆO  x '7=  A  â x *BL?  A  â !r! !       ä #t   €?   â x     ?  A  Æ  r     A  Ì s     $   $ !t  €?    È  r       A  È  r!!      A  Ê †y !   â My        € ê x     pRð Ú G‰  P    € ê x     pRð Ú G	  `    € ê „y     H   " z  r  p`ð â Ey       € ä „y     H   b r ÿ   pVp  Æ „y    H   ¨ „y    H   è „y	    H   ( „y  
  H   ( „y €   H   ( „y €  H   ( „y €  H   ( „y €
  H   â !r&       â !r  )       Æ/ !r       âO !r         Æ� !r		       â!r        È ‰ 	      " !r%       Æ ‰      b !r       È !r       â !r 	        â !r       È/ ‰       ( ‰      h „y €  H   ¢ !r        â !r       È/ ‰ €    ( ‰ €    h „y €  H   â !r>       âO !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       È� !r       â !r       â !r
	
       È/ ‰       ( ‰
      b !r        â !r
       â/ G	  À     € ê !r         Ê	z  i    � â !r !       È  x;ª¸¿  A  â 	z   i    � È 	z   i €  � â s        $ !t  €?    Ì s        $  r     A  È  r      A  Ê †y     ä Ay        € ê ‰       b Ey       € æ ‰      ¨ „y    H   è „y    H   ( „y  	  H   â !r         Æ/ „y    H   â !r       ÆO ‰       h ‰      ( „y    H   ¢ !r$       â� !r(       â!r        Æ/ !r       â !r       â ‰ €    ( ‰ €    h „y    H   â !r       âO !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       È� !r       â !r       â x   ÿàÿ â !r
	
       Æ/ ‰       " z  r  p`ð Æ ‰
      b x    pVp  â !r        â !r
       Ö/ G	  À     € ê !r         Ê	z  i    � â !r !       È  x;ª¸¿  A  â 	z   i    � È 	z   i €  � â s        $ !t  €?    Ì s        $  r     A  È  r      A  Ê †y    ä Ay        € ê ‰       b Ey  Ð    € æ ‰      ¨ „y €  H   è „y €	  H   ( „y €  H   è „y €  H   â !r         Æ/ „y €  H   â !r       ÆO ‰       h ‰      " !r'       È� !r       â!r        â/ !r       È ‰ €    ( ‰ €    h „y €  H   ¢ !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       âO !r       â x   ÿàÿ â !r
	
       Æ/ ‰       " z  r  p`ð Æ ‰
      b x    pVp  â !r        â !r
       Ö/ G	  À     € ê !r         Ê	z  i    � â !r !       È  x;ª¸¿  A  â 	z   i    � È 	z   i €  � â s        $ !t  €?    Ì s        $  r     A  È  r      A  Ê †y    ä Ay        € ê ‰      b !r#       â x   ÿàÿ Æ !r       â z  r  p`ð Æ !r       â x    pVp  È ‰       ¢ !r       Ê/ ‰      " !r         ÊO ‰       b !r       Ê ‰ €    " !r        Ê/ ‰ €    b !r	       Ê ‰
	 @    " !r       Ê/ ‰ @    b !r
	
       Ê ‰
      " !r       Ê/ ‰       ¢ !r
       â M	        € ì !r         â/!r         ÈO 	z   i    � â !r! !       È  x ;ª¸¿  A  â 	z!! i    � È 	z!! i €  � â s        $ !t  €?    Ì s        $  r       A  È  r!!      A  Ê †y !   â My        € ê „y     H   " x   ÿàÿ äx   ÿàÿ â„y     H   b z  r  p`ò ä z  r  p`ð â „y €   H   ¢ z  r  p`ô Ä r ÿ   pVp  â„y €  H   â x    pVò  äx    pVt â „y    H   " x   ÿàÿ Æ „y    H   " z  r  p`ö Æ „y    H   " x    pVö Æ „y    H   ( „y €  H   ( „y €  H   ( „y
    H   " !r&       Æ „y  	  H   " !r  )       Æ/ „y	    H   b !r%       ÆO „y  
  H   ¢ !r>>       Æ� „y €  H   â !r$       Æ„y €
  H   " !r(       Æ „y    H   " !r       Æ „y    H   " !r         Æ „y €  H   " !r       Æ „y €  H   " !r>       Æ „y+ €  H   " !r

       Æ „y- €	  H   â !r       Æ „y/ €  H   â !r	       Æ/ „y1 €  H   â !r        âO !r       ä� ‰       " !r       Æ‰      b !r

       Æ ‰      ¢ !r       Æ ‰      â !r#       Æ ‰	
      " !r'       Æ ‰      " !r+       â !r         â !r-       â !r       â/ !r//       â !r       äO ‰       " !r11       â !r       Ä� ‰/      b !r

	       Æ‰1      ¢ !r       Æ ‰      è ‰	      ( ‰      ( ‰
      " !r        Æ ‰      " !r/       â/ !r1       äO ‰  €    b !r       Æ� ‰      ¢ !r		       Æ‰      â !r       Æ ‰ €    " !r
       Æ ‰	 €    " !r       Æ ‰ €    ( ‰
 €    " !r         Æ/ ‰ €    b !r       âO !r       ä� ‰  @    ¢ !r       Æ‰ €    â !r	       Æ ‰ €    " !r       Æ ‰ @    " !r

       Æ ‰	 @    " !r       Æ/ ‰ @    h ‰
 @    b !r        ÆO ‰ @    ¢ !r       â� !r       ä‰       â !r       Æ ‰ @    " !r		       Æ ‰ @    " !r       Æ/ ‰      b !r
       Æ ‰	      b !r       ÆO ‰      ¨ ‰
      ¢ !r         Æ� ‰      â !r       â !‚  !       â!r       Æ‰      " !r       Æ/ ‰      b !r	       â !‚        â !r       äO !’!       â  ‚      A  â !r
       â !’        Æ †‰     â !r	       â� !¢!       â  ’     A  ä !¢		        â !r       ä †™    â  ¢		     A  â !r       È/ †© 	   â M9        € ê !r!!       â !r         È  r!!      A  Ê †y !   â My        € ê „y     H   " z  r  p`ð â Ey  À    € ä „y     H   b r ÿ   pVp  Æ „y    H   ¨ „y    H   è „y	    H   ( „y  
  H   ( „y €   H   ( „y €  H   ( „y €  H   ( „y €
  H   â !r&       â !r  )       Æ/ !r       âO !r         Æ� !r		       â!r        È ‰ 	      " !r%       Æ ‰      b !r       È !r       â !r 	        â !r       È/ ‰       ( ‰      h „y €  H   ¢ !r        â !r       È/ ‰ €    ( ‰ €    h „y €  H   â !r>       âO !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       È� !r       â !r       â !r
	
       È/ ‰       ( ‰
      b !r        â !r
       â/ G	  à     € ê !r         Ê	x   à@  � Ê  x #ÛÙ?  A  È  x ;ª¸¿  A  â !r !       Ê s        " 	x    à@  � È 	x    àÀ  � Ê !t   €?    â !t  €?    È s	        $  r	     A  È  r     A  Ê †y     ä Ay        € ê ‰       b Ey       € æ ‰      ¨ „y    H   è „y    H   ( „y  	  H   â !r         Æ/ „y    H   â !r       ÆO ‰       h ‰      ( „y    H   ¢ !r$       â� !r(       â!r        Æ/ !r       â !r       â ‰ €    ( ‰ €    h „y    H   â !r       âO !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       È� !r       â !r       â x   ÿàÿ â !r
	
       Æ/ ‰       " z  r  p`ð Æ ‰
      b x    pVp  â !r        â !r
       Ö/ G	  à     € ê !r         Ê	x   à@  � Ê  x #ÛÙ?  A  È  x ;ª¸¿  A  â !r !       Ê s        " 	x    à@  � È 	x    àÀ  � Ê !t   €?    â !t  €?    È s	        $  r	     A  È  r     A  Ê †y    ä Ay        € ê ‰       b Ey  ð    € æ ‰      ¨ „y €  H   è „y €	  H   ( „y €  H   è „y €  H   â !r         Æ/ „y €  H   â !r       ÆO ‰       h ‰      " !r'       È� !r       â!r        â/ !r       È ‰ €    ( ‰ €    h „y €  H   ¢ !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       âO !r       â x   ÿàÿ â !r
	
       Æ/ ‰       " z  r  p`ð Æ ‰
      b x    pVp  â !r        â !r
       Ö/ G	  à     € ê !r         Ê	x   à@  � Ê  x #ÛÙ?  A  È  x ;ª¸¿  A  â !r !       Ê s        " 	x    à@  � È 	x    àÀ  � Ê !t   €?    â !t  €?    È s	        $  r	     A  È  r     A  Ê †y    ä Ay        € ê ‰      b !r#       â x   ÿàÿ Æ !r       â z  r  p`ð Æ !r       â x    pVp  È ‰       ¢ !r       Ê/ ‰      " !r         ÊO ‰       b !r       Ê ‰ €    " !r        Ê/ ‰ €    b !r	       Ê ‰
	 @    " !r       Ê/ ‰ @    b !r
	
       Ê ‰
      " !r       Ê/ ‰       ¢ !r
       â M	        € ì !r         â/!r         ÈO 	x    à@  � â !r! !       È  x #ÛÙ?  A  â 	x!!  à@  � Æ  x;ª¸¿  A  â 	x!!  àÀ  � Ê s        " !t!!  €?    â !t  €?    Ì s        $  r       A  È  r! !     A  Ê †y !   â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y       "   $ z   w  p`ð Ú M	        € ê y      !   " ¹z  l      â Ey  À    € â ™x?    â $rÿÿ   ÿ Ž â ¹z  F   
   â r ÿ       ä xÿ     È |    p`ð Ú G	  @    € ê y      &   " $tÿ   ÿ Ž ä $z  x  Ž È %v \   Ž Ê �y    ™ ¢  $rÿÿ   ÿ Ž â xÿ     â $x   ÿ Ž â y      %   b 'z m   Ž â xøÿÿÿÀŽ ä x	   ÿÀŽ â $vÿ X  ÿ Ž â zÿ n    â $z  r  ÿŽ â x   ÿÀŽ â $rÿÿ   ÿ Ž â rÿ   ÿáÿ Ä x   ÿÀŽ ä z  Y      â $z o  Ž â xÿ     â $x   
Ž ä $z u  ÿŽ â x   þŽ È r   à‘ â $tÿ$   ÿ Ž ä $rÿÿ   ÿ Ž â rÿÿ   ÿ$~  â %v Z   Ž â x   ÿ   ä/ r	   ÿð� â $x$   ÿŽ â x
˜ÿÿÿÿàñ â $z q  Ž âx   ÿàÿ Ä x<   ÿÀŽ ä x	ÿ     ä xÿÿÿÿä  ä x   ÿÀŽ â $z t  Ž äO$z t  Ž ä $z q  Ž Ä %xT   Ž È %xT   Ž È %xT   Ž È %xT   Ž È r   ÿàñ â�y%
üÿÿ • ¢  r   ÿàó ä r   ÿä  â �y$
    • â  r
   ÿàñ â $x   Ž  ä �y!   ™ $ $rÿÿ     Ä �y    ‘ ¨
 �y   ‘ è
 �y#    ™ è  �y$   ™ è  �y"   ‘ â
 r   ÿàñ Æ/ �y H   ™ è  �y'   ‘ â
 r	   ÿä  ä r   ÿàñ â �y&
D   • è �y)l   ™ â  $x     Æ �y(    ‘ è  �y   ™ è �y*
h   • è �y,   ‘ è �y+   ‘ è �y-   ‘ â r        â $rÿÿ    Ž È %xT   Ž È%xT   Ž ä �yP   ™ h �yP   ™ b	 x   ÿàÿ È |    p`ð ä x

@  ÿàõ ä/ x¨   ÿàó Æ $rÿÿ    â rÿ   ÿäÿ  ä xÿ   !  äx!ÿÀŽ ä xÿÀŽ ä x   ÿÀŽ äOxð   ÿÀŽ Ä x   ÿÀŽ â� &r#   ÿ   â xð   ÿÀŽ ä xÿ     â $r   ÿŽ â xÿ   !  ä xÿ     â $x   Ž â xÿÀŽ â &r   ÿ   â r   ÿ@Ž â $x   Ž â x"   ÿÀŽ â $r   ÿŽ â x""ð   ÿÀŽ â 0r%ÿ%    A   â r   ÿ€Ž â &r#   ÿ   â Er        Ä x"ÿ   "  â &r    ÿ   â Er        â 0r$ÿ$    A   â x!ÿ   !  â &r   ÿ   â x'ð   ÿÀŽ â #r%   ÿ   â Er        â $x""   "Ž â x!!ÿÀŽ â $r   ÿŽ â xÿ     â #r$      â Er        â #r%   ÿ   â r""   ÿ€Ž â 0r&ÿ&    A   â x''   ÿÀŽ â &r!)   ÿ   â Er        â $x   Ž â #r$      â x(ð   ÿÀŽ â &r    ÿ   â #r&      â $r'   ÿŽ â r   ÿ€Ž Ä xÿ     ä Er        â &r)   ÿ   â Er        â $x   Ž â xÿÀŽ â #r&      â 0r*ÿ*    A   â x(   ÿÀŽ ä Er        â &r#   ÿ   â r"   ÿ€Ž â #r*      â x!,ð   ÿÀŽ â $r   ÿŽ â #r*      â xÿ     â &r"#   ÿ   â x!ÿ   !  Ä Er        ä xÿÀŽ ä Er        ä x+ð   ÿÀŽ â $x!!   !Ž â #r%   ÿ   â x"-ð   ÿÀŽ â #r%%   ÿ   â xÿ     Ä xÿ     ä r!!!   ÿ€Ž â $x   Ž â x"ÿ   "  â &r   ÿ   âx,,   ÿÀŽ ä xÿ     ä xÿÀŽ â &r!   ÿ   â x+   ÿÀŽ â $x""   "Ž â r!   ÿ€Ž â $r,   ÿŽ â xÿÀŽ â &r    ÿ   âx--   ÿÀŽ â &r!!    ÿ   â r"""   ÿ€Ž â $r   ÿŽ â Er        â &r)   ÿ   âEr        â &r"")   ÿ   â Er        â $r--   ÿŽ â Er !       â #r$      â#r%$   %   â Er" "       Ä #r&      â#r%&   %   â 0rÿ  0 A   âEr- -       â 0rÿ  0 A   â#r"*"   %   â0rÿ    A   â  r     A  â 0rÿ    A   â #r*-      â  r""     A  â #r     Æ #r   "  â !r       Æ !r       â G‰  °õÿÿÿÿƒ ê Ay        € ê ‰      ( ‰      h y      %   ¢ !r       â !r       È/ ‰      " $x   Ž ÆO ‰      d z  p  p`ð È x    pDp  â !r       â !r       È/ ‰ €    ( ‰ €    b !r       â !r       È/ ‰ @    ( ‰	 @    b !r       â !r		       È/ ‰      h  ‰	      ¢  M	        € ê y      &   â x   ÿ   â $z   s  Ž â !r       â/ !r		       â_ x    p"ò â$tÿ   ÿ Ž â r ÿ   p"ð Ö  	       â $z  v   Ž È� %v  j   Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      "   $ z  w  p`ð Ú M	        € ê y	      &   " $tÿ   ÿ Ž â ¹z  F   
   â $z x  	Ž È %v \   Ž Ê �y     ™ b ¹z  l      â Ey       € â ™x?    â y      !   b r ÿ       â $rÿÿ   ÿ Ž â r ÿ       â $r(ÿÿ   ÿ Ž â x      ÿ  â xÿ     È/ |    p`ð Ú G	  p    € ê r ÿ       â $z r  ÿŽ â xÿ     â $x   ÿ Ž âx   ÿÀŽ â 'z	 m   Ž â r ÿ       â Ey Ð    € â x<   ÿÀŽ ä zÿ n    Ä z  a      ä x   ÿÀŽ â $rÿÿ   
Ž â x   ÿÀŽ Æ $z o  	Ž â x   ÿø� â y	      %   $ $z u  ÿŽ â xøÿÿøŽ ä x $       ä r   à‘ ä x   ÿÀŽ Ä r
ÿÿ   ÿ$~  â %v Z   Ž â x   ÿÀŽ Æ $x
$   ÿŽ â x˜ÿÿÿÿàñ â $x   
Ž â r        â $v
ÿ `  ÿ Ž â xÿÿÿÿä  ä r   ÿð� â $x	   ÿ Ž â z ÿ `  pPð â $rÿÿ   ÿ Ž â xÿ     â $z q  Ž â z  X      Ä z	  Y      â $z  t  Ž âz ÿ a   Sð Æ %xT   Ž È %xT   Ž È r   ÿàó â�y,üÿÿ • ¢ r"   ÿàõ Æ $x   	Ž  â r&   ÿàó â �y-    • â r#	   ÿä Æ �y)   ™ " $r'ÿÿ   Ž  Æ �y*"    ‘ h �y&    ™ ¨  �y+"   ‘ è �y&$   ™ â  xÿ     Ä x   ÿàÿ â �y."   ‘ ä $z  t  Ž ä �y&H   ™ ä $z q  Ž ä �y1"   ‘ ä %xT   Ž Ä �y0&l   ™ ä  %xT   Ž â r$   ÿàó â�y/D   • æ r    ÿàõ â �y2h   • â r%	   ÿäÿ  Æ $x!   	 ä�y$$   ™ è �y3     ‘ è  �y4    ‘ â  %x"T   Ž Æ/ �y5    ‘ è �y""P   ™ è  �y#    ‘ â %x&T   Ž Ì �y&&P   ™ â  x   ÿàÿ È |    p`ò ä x%ÿ   )  äx')ÿÀŽ ä x %ÿÀŽ ä/ x%*   ÿÀŽ â&r!'   ÿ   âO x'+   ÿÀŽ Æ� $r%!%   ÿŽ ä &r     ÿ   ä 0r,ÿ,    A   â Er! %       â $r  '   ÿŽ â x%ÿ   )  â 0r-ÿ-    A   â x**ð   ÿÀŽ â #r!,!   ÿ   â Er          Ä x%%ÿÀŽ ä x++ð   ÿÀŽ â #r -    !   â x*ÿ   *  â &r%%   ÿ   â x!.   ÿÀŽ ä x+ÿ   +  ä r***   ÿ@Ž â $r!%!   ÿŽ â x..ð   ÿÀŽ â $x++   +Ž â r%**   ÿ€Ž Ä x.ÿ   .  ä x)ÿ   )  â &r'%   ÿ   â r*++   ÿ€Ž â $x..   .Ž â x%1ð   ÿÀŽ ä x))ÿÀŽ â &r**   ÿ   â Er' '       â $x+.   .Ž â x11   ÿÀŽ â &r))0   ÿ   â x%ÿ   %  â #r.,'   ÿ   â Er* *       â &r++   ÿ   â r%%%   ÿ@Ž â 0r/ÿ/    A   â Er! !       â $r))1   ÿŽ â x1$ÿÀŽ â #r*-*   .   â x.ÿ   $  â $x'%   %Ž â Er+ +       â #r!/!       â x%3   ÿÀŽ â &r 1   ÿ   â x..ÿÀŽ â #r+/+   *   â &r*'0   ÿ   Ä $r% %   ÿŽ â x 4   ÿÀŽ â &r'.   ÿ   â Er* *       â 0r2ÿ2    A   â Er% %       â $r''    ÿŽ â Er) )       ä x33ð   ÿÀŽ â #r 2*   +   â x44ð   ÿÀŽ Ä Er* '       â #r',%   ÿ   â x3ÿ   3  â #r!2)   !   â x%5   ÿÀŽ äx)ÿ   $  ä x55ð   ÿÀŽ ä x4ÿ   4  Ä r333   ÿ@Ž â #r*-*   '   â x+)ÿÀŽ â 0r)ÿ"  0 A   â x5ÿ   5  â $x44   4Ž â x'#ð   ÿÀŽ ä r333   ÿ€Ž ä x$ÿ   $  â $x55   5Ž â  r  )     A  â x'ÿ   '  â &r33   ÿ   â r444   ÿ€Ž Ä x)$ÿÀŽ â &r.+   ÿ   â r555   ÿ€Ž â $x''   'Ž â x##   ÿÀŽ â &r44   ÿ   â Er3 3       â &r$)0   ÿ   ä $r%.%   ÿŽ â Er4 4       â $x''   'Ž Ä &r55   ÿ   ä $r#$#   ÿŽ â #r$,3   ÿ   â Er% %       â &r''0   ÿ   â Er5 5       â #r$-4   $   â Er# #       â #r%/%   *   âEr* '       â #r5/5   $   Ä #r#2#   %   â0r%ÿ&  0 A   â #r*2*   5   â 0r"ÿ"    A   ä 0r$ÿ&    A   â  r**%     A  ä #r&!"      ä #r'#$   *  â G‰  €    € ê xÿ     Ê $z q  Ž È $z  t  Ž È %xT   Ž È %xT   Ž â r"
   ÿàõ È r#   ÿä ä r 
   ÿàõ Æ �y*"     ¨ �y)"    â $x!    Æ �y+"    ( �y      h �y."    " %x$T   
Ž Ì �y$$P    " x1ÿ   *  äO x3ÿ   )  ä� r111   ÿ@Ž Æ $x53   3Ž â r311   ÿ€Ž ä x4ÿ      â$x"5   5Ž ä &r#3   ÿ   â x44ÿÀŽ â &r""   ÿ   â x!ÿ      ä Er# #       â &r%4   ÿ   â/ Er" "       Ä x1 ÿÀŽ â #r4,#   ÿ   â x!!ÿÀŽ ä x))   ÿÀŽ â #r"-"   4   â x#ÿ   +  â&r11   ÿ   â x**   ÿÀŽ â &r!!   ÿ   â x4+   ÿÀŽ â $r)%)   ÿŽ â x+ÿ   .  Ä r###   ÿ@Ž â $r*1*   ÿŽ â x%ÿ      â $r !4   ÿŽ â r!##   ÿ€Ž â $x1+   +Ž â x+%ÿÀŽ ä Er# *       â $x11   1Ž â x..   ÿÀŽ â &r%!   ÿ   â Er) )       â &r++0   ÿ   â#r*,#   ÿ   â &r!10   ÿ   â Er% %       â $r++.   ÿŽ â Er#         â #r*-)   *   â Er! !       â #r%/%   "   âEr  +       â #r#/#   *   â 0r"ÿ$  0 A   â #r!2!   %   Ä #r 2    #   â 0r#ÿ$    A   â  r!!"     A  È #r! #   !  È !r((!       â !r&       â !r'       â G‰  0    € ê r
   ÿà÷ är
   ÿàõ ä r   ÿäÿ Æ $x    ä�y%     ¨ �y    è �y$    ( �y"    h	 �y#    b	 %x T   
Ž Ì �y  P    b x&ÿ   %  äO xÿÀŽ ä� r'&&   ÿ@Ž ä x&%   ÿÀŽ â &r   ÿ   â xÿ     â $x''   'Ž â xÿ   $  â$r&   ÿŽ â x$   ÿÀŽ â &r''   ÿ   â xÿÀŽ Ä r!   ÿ@Ž ä Er        â &r   ÿ   â r$!!   ÿ€Ž ä Er '       â #r,   ÿ   â xÿ   "  â&r$$   ÿ   â xÿ     â #r,,   ÿ   â $r   ÿŽ â xÿ   #  â $x   Ž â xÿÀŽ Ä xÿ     ä r   ÿ@Ž â &r   ÿ   âr   ÿ€Ž ä x""   ÿÀŽ â $x!   Ž â xÿÀŽ â &r   ÿ   â Er$ $       â $r"   ÿŽ â x#   ÿÀŽ â &r0   ÿ   âEr        â &r!!0   ÿ   â Er        â #r$-$   ,   â Er        â $r   ÿŽ â #r-      â Er! !       â #r/   $   Ä #r/      â Er        â 0rÿ   0 A   â#r2!      â0r ÿ     A   ä #r2      â  r     A  È #r      È !r       â x

¨   ÿà÷ ä x¨   ÿàõ ä x@  ÿàù â $rÿÿ   Ž â r	ÿ	   ÿä ä rÿ   ÿä â D       À â Gy  pìÿÿÿÿƒ ê Ay       € ê r (       ä r        Ä Ay        € ê z ÿ `  pPð È z ÿ a   Sð Ú G	  P    € ê ‰      ( ‰      b !r       â !r       È/ ‰      ( ‰      b !r       â !r	       È/ ‰ €    ( ‰
	 €    b !r       â !r
	
       È/ ‰ @    ( ‰
 @    b !r       â !r
       È/ ‰      ( ‰      b !r       â !r       â/ Gy  €    € ê ‰      ( ‰(      h ‰	      ¨ ‰
      â !r       â !r(       È/ ‰      " !r	       ÆO ‰	      b !r
       Æ� ‰
      ¨ ‰      â !r       â !r		       È/ ‰ €    " !r
       ÆO ‰
	 €    b !r       Æ� ‰ €    ¨ ‰ €    â !r       â !r
	
       â/ !r       ÆO ‰ @    " !r       Æ� ‰
 @    h ‰ @    ¨ ‰ @    â !r       â !r
       â/ !r       ÆO ‰      " !r	       Æ� ‰      h ‰      ¨ ‰	      â !r       â !r       â/ !r       âO !r	       â� y      %   $ $x   Ž Ê z  p  p`ð È x    pDp  Ú M	        € ê z ÿ ^  pPð â $z  v  Ž Æz ÿ _   Sð Ú         Ê % ^   Ž Ì �	     ¢ x    ÿ   È r  ÿ   p"ò äx     p"ö äx     p"ø äx     p"ô Î $ ÿÿ    Ž â z ÿ `  pPò ä 2         ä z ÿ a  Sò ä B         â $" ÿÿ    Ž È !         ÌO G™       € ê z ÿ b  pPð È z ÿ c   Sð Ú         Ê % b   Ž Ì �	     ¢ x   ÿ   â $vÿ h  ÿ Ž Æ x øÿÿÿp"ò äx üÿÿÿp"ô är ÿ   p"ö äx    p"ø Î $ÿÿ    Ž â x    pBò ä "        ä 2        ä B        Ê !       äO G  P    € ê x    pRð Ú G‰  �     € ê x    pRð Ú G	  ð    € ê  x;ª¸¿  A  Ì s        $ !t  €?    Ì s        $  r     A  È  r       A  â Gy  P    € ê  x'7=  A  â x*BL?  A  Æ #t  €?   â  x   ?  A  Æ  r     A  Ì s     $   $ !t  €?    È  r     A  È  r       A  â Gy  °    € ê x    pRð Ú G‰  à     € ê x    pRð Ú G	        € ê 	z i    � ä 	z   i    � Æ  x;ª¸¿  A  â 	z   i €  � Ê s        $ !t  €?    Ì s        $  r     A  È  r       A  â Gy  Ð     € ê  r       A  â Gy  °     € ê 	x  à@  � ä 	x    à@  � Æ  x#ÛÙ?  A  â 	x    àÀ  � Æ  x;ª¸¿  A  ä !t   €?    È s        $ !t  €?    Ì s        $  r      A  È  r       A  ä y      "   ( y      &   b $z s  Ž â x        Æ $z v  Ž È/ %v j   Ž Ê †y      â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y,      "   " ¹z  l      â Ey  `-    € â ™x?    â y-      !   b ¹z  F   
   â r/ ÿ       â x0      ÿ  â rB ÿ       â x2      ÿ  â rD ÿ       â x$      ÿ  â rF ÿ       â x&      ÿ  â x(      ÿ  â x*      ÿ  â r ,ÿ   pRð Ä r ,ÿ   pRò är4,-   ÿ(Ž È/ r 44   ÿ� È x.ÿ      È | .   pbô Ú G)  ,    € ê y      %   " x þÿÿÿÿÀŽ â $xK.   ÿ Ž â r ÿ       â y      '   b r44  €ÿàÿ ä zC  X      â y      &   ¢ x4   ÿ   ä r/ ÿ       Ä x>   ÿàÿ ä x@ÿ   4  ä x:ÿ   >  äx>ÿ   >  ä x;ÿ   :  ä zIK q  ÿàÿ Ä x   ÿ   ä r ÿ       â 'z  y   Ž â/ x<ÿ   >  Æ 'z s   Ž âO z	ÿ z     Æ $z  p  .Ž â zÿ t    â $z	 |  .Ž ä$z		 |   Ž â x    ÿàÿ â $z v  Ž ä$z	 v  	Ž â x4   ÿ   â $z  p  Ž â z   q      â %x	$   ÿŽ Æ r5 K   ÿŽ â $z6 }  ÿŽ â xÿ     â$z w  ÿŽ â r7   ÿà÷ â%x
$   ÿŽ â r8:   ÿàù ä r66   ÿàõ â %x$   Ž â r9	   ÿäÿ Ä rA K   ÿŽ â %x$   :Ž â r:	;   ÿä ä z;  Y      â %x   Ž â r‚ÿÿ   ÿä Æ %x
   
Ž È $x?    KŽ ä$x=    KŽ ä$xE    KŽ ä$xG    KŽ ä rC   ÿàõ È r	;   ÿä Ê �y_   ™ ¨  �yf   ™ è  �yc   ™ ( �y^   ™ b r
C   ÿàõ È r;   ÿä Ê �yh   ™ ¢ rC7   ÿàõ ärTC   ÿàù â �ye   ™ ¢ rC8   ÿà÷ ä rK6   ÿàû â �y]   ™ ¢ r;9   ÿä ä rU;   ÿä â �ya   ™ ¢ rC   ÿàý Ä r#I6   ÿàõ â �y\    ‘ ¢  r;:   ÿäÿ ä r4   ÿàù â�yTT    ‘ ¢  r>   ÿà÷ ä rÿ‚   ÿäÿ â$x   ; â r4#   ÿàû â �yY    ‘ ¢  rÿ‚   ÿä Ä/ xH $       â �yW    ‘ ¢ r#>#   ÿàý ä r<   ÿäÿ är@   ÿä är@   ÿäÿ â%v Z  H Ž â r <   ÿä Æ $x$   ÿŽ ä %v Z  H Ž È %v Z  H Ž â r   ÿàÿ Æ $x$   ÿŽ ä $x$   ÿŽ ä %v"# Z  H Ž â r56   ÿàù â �yX   ™ ä  $x $   ÿŽ â r   ÿàÿ â �yU   ™ â  r   ÿàÿ Ä r?6   ÿàû â �y`   ™ â  r##   ÿàÿ ä r4   ÿà÷ â�yb   ™ â  rÿ‚   ÿä ärÿ‚   ÿäÿ â/ �yd   ™ â  r4   ÿàù Ä r>   ÿàû â �yg   ™ â  r@   ÿäÿ â$x@    â r <   ÿäÿ â %v Z  H Ž â�yV   ™ & $x$   ÿŽ â �yZ    ™ ( r   ÿàÿ â $x $   ÿŽ â �y[    • ( �y     • " r>   ÿàõ â %v Z  H Ž â r!A6   ÿàù â �yP   ™ " r<   ÿä â $x$   ÿŽ Ä %v Z  H Ž â�yN"   ™ $ r   ÿàÿ â %v Z  H Ž â r=6   ÿàû â �yQ   ™ $ $x$   ÿŽ â rM4!   ÿà÷ â �yR    • " rÿ‚   ÿä Ä r   ÿàÿ â �yl   ™ " r   ÿàÿ ä r!>!   ÿàõ â %vM Z  H Ž â rÿ‚   ÿäÿ â �yt   ™ b r4   ÿàù är>   ÿàû â �yn    • b rJ@   ÿäÿ Ä r <   ÿä â �yp    • b r@   ÿä är<   ÿäÿ â %v! Z  H Ž â �yr   ™ f $x!J$   ÿŽ â �ys   ™ b $xM $   ÿŽ Æ r!   ÿàÿ â �yq   ™ b rM   ÿàÿ ä r!E6   ÿàù â �yM   ™ b %v Z  H Ž â rSG6   ÿàû ä �y}   ™ b $x$   ÿŽ â rk4!   ÿà÷ â%v Z  H Ž â r!>!   ÿàõ â �yo   ™ d $xO$   ÿŽ â riÿ‚   ÿä â$x   Ž â rÿ‚   ÿäÿ â/ �yy    • b r4S   ÿàù Ä rO   ÿàÿ â %v ! Z  H Ž ârS>S   ÿàû â �yO   ™ b rJ<i   ÿä ärj@i   ÿäÿ â%v Z  H Ž âri@   ÿä â�yu   ™ b rL<   ÿäÿ â %vk Z  H Ž Ä �yv   ™ d $xkJ$   ÿŽ ä �yJ"   ™ b %vS Z  H Ž Æ �yH"    ™ b $xSj$   ÿŽ ä $xmL$   ÿŽ â �yz    • h �yL"   ™ b rS   ÿàÿ Æ �yS"    • b $xii$   ÿŽ â rm   ÿàÿ ä �ym   ™ d ri   ÿàÿ ä �yi   ™ b r!!k   ÿàÿ Æ �y"   ™ h# �yj   ™ h �yk   ™ h �y#    ™ h �y    ™ h �y|   ™ h �y{   ™ h �yx   ™ h �yw   ™ h �y€    • h �y„   ™ h �y   ™ h# �y   ™ h �y   ™ h �y    ™ h �y~    • h �yƒ   ™ h �y�   ™ h �y   ™ h �y‹    • h �y   ™ h �y   ™ h �yˆ   ™ h �y   ™ h �yŠ   ™ h �y    ™ h �y‡    • h �y‰   ™ h �y   ™ h# �yŽ    • b x_ww  ÿÀŽ âO ‚x       Ä �y   ™ ¢ ‚x  ÿþý     Æ �y�   ™ ¢  xÿ   _  ä/ x__ˆˆˆˆÿÀŽ â �y    ™ ¢ |       ä |‘       â �y†    ™ ¢ x__22ÿø� â $~“ÿ   ÿ Ž Ä �y…    ™ ¢ |•       Æ �y     ™ ¢ xww  ÿÀŽ Æ �y     • ¢ x�_Tv  ÿÀŽ Æ �yŒ   ™ ¢  t’•üúøô    Æ �y   ™ ¢  t!    Æ/ �y    ™ ¢ t‘üúøô    Æ �y�    • ¢  t‘“    ä x_ÿ   _  â �y   ™ ¢ r!�       ä x�ÿ   f  â� �y    ™ â x!fww  ÿÀŽ Æ �y   ™ ¢ xffˆˆˆˆÿÀŽ ä r_‘_   ’    â �y    • ¢ |’       ä |–       ä |˜       ä x‘�ww  ÿÀŽ Ä |”       ä xff22ÿø� ä t�’!    ä t’–‘    ät“˜üúøô‘    ä t!”üúøô!    Ä x‘fTv  ÿÀŽ ä x d  _    ä/x1u  _    ä xfÿ   f  ä x_ÿ   c  äxcww  ÿÀŽ Ä xccˆˆˆˆÿÀŽ ä r!�‘   !    ä |‘       ä rf’f   “    ä |•       ä |—       Ä x�_ww  ÿÀŽ ä |“       ä xcc22ÿø� ä t_‘    ä t‘•�    ät’—üúøô�    Ä t“üúøô    ä x�cTv  ÿÀŽ ä x“ÿ   c  ä xc! d  f    äxf!1u  f    ä r_�       Ä x_ÿ   ^  äx!^ww  ÿÀŽ äx^^ˆˆˆˆÿÀŽ ä r“‘“   ’    ä |�       ä |’       Ä x^^22ÿø� ä |”       ä |–       ä x‘_ww  ÿÀŽ ä t_�!    ät�’üúøô!    Ä x’^Tv  ÿÀŽ ä t””‘    ät•–üúøô‘    ä x^ÿ   ^  ä x‘ d  “    äx!1u  “    Ä r•”^   •    ä x^ÿ   h  ä xhww  ÿÀŽ äxhhˆˆˆˆÿÀŽ â $~“ÿ   ÿ Ž â r�_’   �    ä |_       Ä |—       ä x^^ww  ÿÀŽ ä xhh22ÿø� ä |™       ä t’_    ät““üúøô    Ä t”—^    äxhTv  ÿÀŽ ä t—™üúøô^    ä xhÿ   h  ä x_� d  •    äx^�1u  •    Ä r—”h   —    ä x�ÿ   e  ä xheww  ÿÀŽ är’   “    ä xeeˆˆˆˆÿÀŽ ä |“       Ä |•       ä |™       ä |›       ä x’�ww  ÿÀŽ ä xee22ÿø� ä t�“h    Ä t“•üúøôh    ä t”™’    ät•›üúøô’    ä x’eTv  ÿÀŽ ä x–ÿ   e  ä xe d  —    Ä xh1u  —    ä r��’   “    ä x’ÿ   ]  ä x]ww  ÿÀŽ äx]]ˆˆˆˆÿÀŽ ä |“       Ä |—       ä x]]22ÿø� ä x’’ww  ÿÀŽ ä r•”–   •    ä t““    ät—üúøô    Ä x”]Tv  ÿÀŽ ä t–™’    ät—›üúøô’    ä x]ÿ   ]  ä r“”       ä r]–]   —    Ä x’� d  •    ä x“ d  ]    äx]1u  ]    ä x��1u  •    ä xaˆˆˆˆÿÀŽ ä |•       Ä x”aww  ÿÀŽ ä x˜\   ÿÀŽ ä |—       ä xaÿ   a  ä x22ÿø� ä t••”    Ä x ˜   pRô ä t”—üúøô”    ä |—       ä x–aww  ÿÀŽ ä |™       ä xaTv  ÿÀŽ Ä r\\ÿ       ä t——–    ät–™üúøô–    ä r”•a   ”    â &ra`   ÿ   âxÿ     â &r•e`   ÿ   â x\\ÿ   ÿÀŽ â &rab   a   â r——   –    â &r•hb   •   â >rÿ\   ÿ  â &racd   a   Ä &r•’d   •   ä 0rdÿ    A   â x`” d  —    äx\”1u  —    â &rbfg   a   âx”Y   ÿÀŽ â 0rÿ     A   â  xad   ?  A  â xdT   ÿÀŽ â &rg�g   •   â x ”   pRô â  r a     A  â Er• b       Ä x d   pRö ä rYYÿ       â #rb •   ÿ   â x–W   ÿÀŽ â &r ‘X   ÿ   ârTTÿ     € ä xYYÿ   ÿÀŽ â &rX“X   ÿ   â x –   pRø â &r !U       â xTTÿ   ÿÀŽ Ä >rYÿY   ÿ  â &r•]U   X   â rdWÿ       â &rU_V       â >rWÿT   ÿ  â 0rTÿY    A   ä &r ^Z   U   ä 0rWÿW    A   â  xTT   ?  A  â 0r[ÿ[    A   Ä &r•`V   •   â  xUW   ?  A  â ErV         â  rWT[     A  â xddÿ   ÿÀŽ Æ #rVWV   b   â >rdÿd   ÿ  â &rWM   ÿ   â  r U     A  â &rYeM   ÿ   ä 0rdÿd    A   ä &rP   W   â Erg g       â &r•\Z   •   â  xMd   ?  A  â &rcO      Ä &rYhP   Y   ä &rP‘N   ÿ   â Er• •       â &rfQ      â #r  g   ÿ   â  r[[M     A  â &rY’O   Y   ä 0rRÿR    A   ä &rP!J   P   â #r[[•       â &rY�Q   Y   â ErO        â  r aR     A  â &rQ_L   P   Ä &rN“N   ÿ   â #rO O   ÿ   â ErY Y       â &rQ^H   Q   â  r UR     A  â &rW]J   N   ä 0rSÿS    A   â #rY Y   ÿ   â &rW`L   W   â ErQ Q       Ä  r TS     A  â &rW\H   W   ä &rHi   ÿ   â #r  Q   O   â &riei   ÿ   ä &rHm   H   â!r) )       â &r ‘t   ÿ   ä &rihm   i   Ä &rOcl   H   ä &rt“t   ÿ   ä &rH!"       ä &ri’l   i   ä &rt]"   t   ä &rOfj   O   Ä &rH_k   H   ä 0rnÿn    A   â ErW W       â &ri�j   i   â  rMS     A  â &rt`k   t   â ErO O       â &rH^#   H   â  r an     A  â &r"r   ÿ   Ä 0rpÿp    A   â #rWW   Y   â &rt\#   t   â Eri i       â  rUn     A  â &r"s   "   â ErH H       â #r  O   ÿ   â &rrer   ÿ   â  r#Tp     A  â Ert t       â #ri   ÿ   â  rOMp     A  â &r"cq   "   â #r #H       â &rrhs   r   Ä &rH‘}   ÿ   â #rOOt      â &r"fo   "   ä &rr’q   r   ä 0rÿy    A   ä &rH!u   H   â !r' '       â &rr�o   r   â Er# "       â  r a     A  â &rH_v   H   â !r&O&       â ErO r       â #r# #   ÿ   â &rH^   H   â  r U     A  â 0rzÿz    A   Ä &r|   ÿ   â #rO O   ÿ   â ErH H       â &rJ“}   ÿ   â  r Tz     A  â &r{      ä &ru]u   J   â #r  H   #   â &rcx      ä &ru`v   u   â !r% %       â &rfw      Ä &r|e|   ÿ   ä &r ‘„   ÿ   ä 0r€ÿ€    A   ä &ru\   u   â Er        â &r{h{   |   ä &r„“„   ÿ   Ä &r"!       â  r a€     A  â &r{’x   {   ä &r#]   „   â #r     ÿ   â &r"_   "   ä &r   ÿ   ä &r{�w   {   Ä &r#`   #   ä &r"^   "   ä 0r~ÿ~    A   ä &re   ÿ   ä &rƒ      â Er{ {       â &r#\   #   â Er" "       â &rhƒ      â  rU€     A  â  rT~     A  â&rc�      â Er# #       â &r’�      â #r{   ÿ   â  rM~     A  â #r"       â &rf      Ä &r�      â #r#      â 0rÿ‹    A   â !rFF       â Er        â &r‘   ÿ   â !r33       â Er        â  r a     A  â rU     A  â &r!      Ä &rˆ   ÿ   â #r     ÿ   â #r   ÿ   â &r“   ÿ   ä &reˆ   ÿ   ä &r_      ä &rŠ      ä &r]      Ä &rŠhŠ      ä &r^      ä 0r‡ÿ‡    A   ä &rc‰      â Er        â &r`      â  rT‡     A  â &rf      Ä 0rŽÿŽ    A   ä &r\      â #r       â &r‰’‰   Š   â Er        â  r aŽ     A  â &r   ÿ   âO Er        â &r‰�   ‰   â  r‡M‡     A  â &re   ÿ   â #r    ÿ   â &r�      Ä &r ‘   ÿ   â #r‡      â &r‘‘    ÿ   â Er‰ ‰       â &r“   ÿ   â  rUŽ     A  â &r�h�      ä &r “    ÿ   ä &rcŒ      Ä &r !†       ä&r!   ‘   â #r‰‰   ÿ   â &r]†      ä&r�’Œ   �   ä &r ]       ä &rf      â x..    ÿàÿ â &r _…       Ä 0rÿ�    A   ä &r__      â !rDD       â &r`…      ä &r��   �   ä &r`       â Er        â &r^       â r a     A  â 0rÿ    A   â | .   pbô â &r\      â Er� �       â 0rÿ    A   Ä &r_^   _   ä�&r\      â  rU     A  â #r    ÿ   â  r T     A  âEru u       â  rTT     A  â Er        â  rzMz     A  âEr        â #r�   ÿ   â Er_ _       â  rM     A  âEr        â  rM     A  â #rOzu   O   â #r       â #r   ‰   â #rT_      â #r      â xCC€  ÿà÷ â !r+V+       â !r*[*       â !r(W(       â !r$O$       â !r22       â !rBB       â !r11       â !r00       â !r//       â x55@   ÿàÿ Ä x??@   ÿàÿ ä xAA@   ÿàÿ ä x==@   ÿàÿ ä xEE@   ÿàÿ ä xGG@   ÿàÿ ä xII@   ÿàÿ Ä xKK@   ÿàÿ ä r;ÿ;   ÿäÿ â D)       À â Gy  P×ÿÿÿÿƒ ê Ay        € ê Ey  0    € â G‰      € ê r,-   ÿHŽ Ê ˆs + øÿ H   è ˆs *€øÿ H   è ˆs ) ùÿ H   è ˆs (€ùÿ H   è ˆs ' úÿ H   è ˆs &€úÿ H   è ˆs % ûÿ H   è ˆs $€ûÿ H   è ˆs F üÿ H   è ˆs 3€üÿ H   è ˆs D ýÿ H   è ˆs 2€ýÿ H   è ˆs B þÿ H   è ˆs 1€þÿ H   è ˆs 0 ÿÿ H   è ˆs /€ÿÿ H   ä Ay        € ê {          ì M        € ê „y -     H   b Ey  @    € æ „y- €   H   ( „y-    H   ¢ !r  +       â/ !r**       È ‰       " !r))       ÆO ‰*      h ‰)      ¢ !r        Æ y       %   " !r*       Æ/ ‰      b !r)       ÆO ‰      ¨ ‰      â $x    ÿ Ž â !r       È/ z  r  p`ð â !r       ÆO r -ÿ   pVp  â ‰ €    " !r	       Æ� ‰ €    h ‰
	 €    ¨ ‰      '   â !r       â x   ÿàÿ â !r       Æ/ ‰  @    " z  r  p`ò â !r
	
       äO ‰ @    b x -   p"ò ä ˆ	        â ‰      &   ¨ ‰
 @    ,       &   b         Æ       '   b !r         â !r       È/ ‰       " $Š x  Ž ÆO ‰      b $Š ~  Ž Æ� „y- €  H   â !r
       â$ x  Ž È‰      ¢ $ ~  Ž â !r        â %† j  	 Ž â !r	       Æ/ % j   Ž â †‰     è † 	   â !r       âO G	  �     € ê y       &   h y      '   " $z   x  Ž È/ $z  ~   Ž Ê z   r  ÿàõ È rÿÿ   ÿä ä z  j  ÿ„ È z  k   Ê †y     è Ay        € ê !r(       â „y-    H   â Ey  �    € æ ‰       h „y	- €  H   è „y-    H   â !r         Ê/ ‰       $ !r        Ê ‰ €    $ !r       Ê ‰ @    $ !r       Ê ‰      $ !r       â G™  °     € ê y       &   " ¹z  r      ä �x   ?àÿ â y      '   b $z   x  Ž È $z  ~   Ž Ê/ |     ÿàõ È rÿÿ   ÿä ä z  j  ÿ„ È z  k   Ê †y     è Ay        € ê „y- €  H   b !r		&       â !r'       â !r%       â „y-    H   ¢ z+  r      Æ „y- €  H   $ $x!+   ÿŽ ä„y-    H   â $x+   ÿŽ Æ „y- €  H   ( „y-    H   è „y- €  H   è „y-    H   è ‰ 	      h ‰      h ‰      b !r$       Æ/ „y- €  H   â !rF       âO !r33       ä ‰
      " !rD       Æ� ‰      b !r2       â!r	        ä‰3      ¢ !rB       â !r       ä ‰      â !r11       â !r       Ä ‰      " !r0       Æ ‰      b !r

       Æ ‰      " !r       Æ/ ‰1      h ‰      b !r33       âO !r//       ä ‰      ¢ !r       Æ� ‰      â !r       Æy       &   " !r       Æy	      '   b !r       â !r11       ä/ ‰
      " !r       Æ ‰      b !r       ÆO ‰3      ¢ !r       Æ� ‰      â $z   x  Ž Æ‰      " $z 	 ~   Ž Æ‰ €    h ‰      b r+    ÿ„ â !r
       â ‚ !   ÿà÷ â ‰1      " rÿÿ   ÿä â !r       â/ z j  ÿ„ â ‰      b !r33       âO ‚ÿÿ   ÿäÿ Ä ‰ /      ¢ !r       â� z k   ä ‰ €    â !r       âŠ
 j  ÿ† ä‰	 €    " !r       âŠ k  � â !r       â ‰ €    b ‚    ÿàù â !r11       Ä ‰ €    " ‚*ÿÿ   ÿä â !r       ä/ ‰3 €    b Š j  ÿˆ â !r//        äO ‰ €    ¢ Š k  * â !r       Ä� ‰  €    â $x+   ÿŽ â !!   ÿàÿ â !r	       â‰ @    " r	+    ÿŠ â$x++   ÿŽ â r    ÿàõ â ‰" €    " !r       ârÿÿ   ÿäÿ Ä ‰$1 €    b rÿÿ   ÿä â !r       â z j  ÿ„ â ‰& €    " !r3       â/ z	 j  ÿŠ ä ‰(/      b !r       âO z k   Ä ‰ @    ¢ !r         â� z		 k  � ä ‰ @    â !r       â‚ +   ÿà÷ â !r""       â ‰ @    " !r$1$       Æ‰ @    b !r&&       Æ ‰ @    " !r(/(       Æ/ ‰ @    b !r       âO    ÿàÿ ä ‰#" @    ¢ !r       â�     ÿàõ ä ‰  @    â ‚ÿÿ   ÿäÿ ä Š j  ÿ† â ‰%$ @    â *ÿÿ   ÿä â !r       â j  ÿ„ â‰)( €    " Š k  � â !r       â k  * â ‰'& @    b ! !   ÿàõ â !r       â !r       â/ ‰      " *ÿÿ   ÿä Ä ! j  ÿ„ â‰      b !r#"#       äO ! k  * â !r         â� ‰      ¢ !r%$%       Æ ‰      â !r)()       Æ‰      " !r'&'       Æ‰      b !r!       Æ ‰      " !r       Æ/ ‰       h ‰#      b !r       ÆO ‰%      ¢ !r       Æ� ‰$) @    â !r       Æ‰"'      " !r       â!r       ä †‰ !    â !r        Æ/ †    â !r##       Æ †‰     â !r%%       ÆO †     â !r$)$       Æ� †‰     â !r''"       Æ†    è †‰ 
    è †     è †‰ #    è † %   è ‰$      h  †‰ '    â M™        € ê x+   ÿàÿ â !r$       Æ/ r    ÿàñ È r ÿÿ   ÿä  ä z j  ÿ€ È z k     Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y(      "   " ¹z  l      â Ey   )    € â ™x?    â y)      !   b ¹z  F   
   â r+ ÿ       â x,      ÿ  â y/      %   ¢ r? ÿ       â x      ÿ  â x       ÿ  â x"      ÿ  â x$      ÿ  â x&      ÿ  â r (ÿ   pRð Ä r (ÿ   pRò är.()   ÿ(Ž È/ r ..   ÿ� â $x//   ÿ Ž ÆO x*ÿ      È | *   pbô Ú G)  P(    € ê y      '   " r ÿ       â $v9ÿ X  ÿ Ž â x þÿÿÿÿÀŽ â y      &   b r ÿ       ä r..  €ÿàÿ ä xB*   ÿ   ä x.   ÿ   â$x4.   ÿ Ž â z7  Y      Ä x;   ÿàÿ ä x5ÿ   4  ä r+ ÿ       ä x=ÿ   .  ä z@B q  ÿàÿ â 'z  y   Ž È 'z s   Ž â/ zÿ z     Æ $z / p  *Ž â zÿ t    â $z	 |  *Ž ä$z |   Ž ä $z w  ÿŽ â x/   ÿàÿ â $z v  Ž ä$z  v  	Ž â xÿ   ;  â%x$   ÿŽ â x;ÿ   ;  Æ $z1 }  ÿŽ â xÿ     â $z p   Ž â r24   ÿà÷ är3   ÿàù â %x
$   ÿŽ âr11   ÿàõ ä x8ÿ   ;  â %x$   4Ž â r4	5   ÿäÿ Ä r5	   ÿä â %x$   Ž â z  q      Æ %x   Ž â r0B   ÿŽ är6B   ÿŽ â %x
   
Ž È $x:   BŽ ä$x<   BŽ ä$x>   BŽ ä $r€ÿÿ   ÿ ä r9   ÿàõ È r	7   ÿä Ê �yw   ™ ¨  �yv   ™ è  �yu   ™ ( �yt   ™ b r
9   ÿàõ È r7   ÿä Ê �ys   ™ ¨ �yU   ™ ¢ r92   ÿàõ Æ �yQ   ™ â r74   ÿä ä rB1   ÿàõ â �y]   ™ æ rÿ€   ÿä â $t^ÿ$   ÿ Ž â r.   ÿàõ â �yP    ‘ æ  r =   ÿä â %v Z  ^ Ž È/ $x $   ÿŽ È $x   Ž â r9   ÿàù È �yR   ™ â  r93   ÿà÷ ä r9   ÿàõ â �yS   ™ â  r7   ÿä ä r@1   ÿàû â �yT   ™ â  r;   ÿàù ä r75   ÿäÿ â �y     ‘ â r7   ÿä Ä r.   ÿà÷ â�y\   ™ â  r;   ÿàõ ä rÿ€   ÿäÿ â �yN    ‘ â  r8   ÿä Æ $x8    â/ �yV    • è �yO    ‘ â r=   ÿäÿ â %v Z  ^ Ž È $x$   ÿŽ Ê r   ÿàÿ Ê �yL   ™ " rE01   ÿàý â %v Z  ^ Ž ä �yM   ™ " rDÿ€   ÿä â $x$   ÿŽ â r.E   ÿàû â/ %v Z  ^ Ž â �yK   ™ & $x$   ÿŽ â r=D   ÿäÿ â �yJ    ™ " r   ÿàÿ Ä r   ÿàÿ â �yC    • " r:1   ÿàû är61   ÿàý ä rE;E   ÿàù â %v Z  ^ Ž â rÿ€   ÿäÿ â�yI   ™ b rÿ€   ÿä â $x$   ÿŽ â r.   ÿàû â �yH   ™ b rD8D   ÿä Ä r;   ÿàù â $x   Ž â r.   ÿàõ â�y[   ™ b r=   ÿäÿ är8   ÿä â %v Z  ^ Ž â rA=   ÿä â �yF   ™ b r;   ÿà÷ â %v Z  ^ Ž Ä �yW    • d $x$   ÿŽ ä $xG$   ÿŽ â �yX    • b %v Z  ^ Ž Æ �yZ   ™ b $xAA$   ÿŽ â r   ÿàÿ â $x8   Ž â rG   ÿàÿ â %vE Z  ^ Ž ârA   ÿàÿ â �yG   ™ b r<1   ÿàù â$xD$   ÿŽ â r>1   ÿàû â �yA   ™ b $xE$   ÿŽ â rY.   ÿà÷ â %v Z  ^ Ž â �yD   ™ b r   ÿàÿ Ä rÿ€   ÿä â$x   EŽ â r;   ÿàõ â�yE    ™ b r_ÿ€   ÿäÿ ä r.   ÿàù â�yb   ™ b r;   ÿàû ä/ r8   ÿä â�yd   ™ b r`=_   ÿä â%vY Z  ^ Ž â r_8_   ÿäÿ â �yY   ™ d $x=   Ž Ä $xc$   ÿŽ â �yi    • b $xa$   ÿŽ ä %v Z  ^ Ž â�yk   ™ f $x__$   ÿŽ â �yj   ™ b %v Z  ^ Ž Æ �yr   ™ b %v Z  ^ Ž Æ �y^   ™ b ra   ÿàÿ â $xa`$   ÿŽ â r_   ÿàÿ â $x   cŽ â �y_   ™ h �y`    • h �yc   ™ b ra   ÿàÿ Æ �ya    ™ h �yh   ™ h �y   ™ h �yf   ™ h �yg   ™ h �yo    • h �y   ™ h# �ye    ™ h �ym    • h �yq   ™ h �yp   ™ h �yn   ™ h �y~    • h �y   ™ h �yl   ™ h �y{   ™ h �y   ™ h �y}   ™ h �y    ™ h �yz    • h �y|   ™ h �y   ™ h �y‚    • b ‚x       Æ �y   ™ b! ‚x  ÿþý     â |ƒ       ä �y   ™ b xÿ   w  ÆO �yy   ™ ¢ xwww  ÿÀŽ ä |…       â �yx   ™ ¢ xww  ÿÀŽ Æ �y    ™ ¨ �y    • ¢ tƒƒ    Æ �y�   ™ ¢  t…üúøô    Æ �y   ™ ¢  $~ÿ   ÿ Ž Æ/ �y   ™ ¨  �y    • ¢  t…    ä|†       â �y   ™ ¨ �y    ™ ¨ �y   ™ ¢ t††üúøô    Æ �y   ™ ¨ �y    • ¢ xwwˆˆˆˆÿÀŽ È xww22ÿø� È x„wTv  ÿÀŽ ä xwÿ   w  ä rƒ„       ä x„ÿ   v  ä� xƒvww  ÿÀŽ äxvvˆˆˆˆÿÀŽ Ä rw…w   †    ä |†       ä |ˆ       ä xvv22ÿø� â $~Œÿ   ÿ Ž â |Š       ä x…„ww  ÿÀŽ Ä t†ƒ    ä/t„ˆüúøôƒ    ä xƒvTv  ÿÀŽ ä t†Š…    ät…Œüúøô…    ä x‡ÿ   v  Ä x d  w    äxv1u  w    ä rƒƒ   „    ä xÿ   u  äxuww  ÿÀŽ äxuuˆˆˆˆÿÀŽ Ä r††‡   …    ä |w       ä |…       ä |„       ä xww  ÿÀŽ ä xuu22ÿø� Ä tww    ät…üúøô    ä t…„    äx„uTv  ÿÀŽ ä tˆˆüúøô    ä x‡ÿ   u  Ä xƒ d  †    äxuƒ1u  †    ä rw„       ä xƒÿ   t  äxwtww  ÿÀŽ är……‡   ˆ    â $~ˆÿ   ÿ Ž â xttˆˆˆˆÿÀŽ Ä |„       ä |‡       ä |‰       ä x†ƒww  ÿÀŽ ä xtt22ÿø� ä tƒ„w    Ä t„ˆüúøôw    ä t‡‡†    ätˆ‰üúøô†    ä x†tTv  ÿÀŽ ä x‰ÿ   t  ä xw d  …    Ä xt1u  …    ä r„ƒ†   „    ä xƒÿ   s  ä xsww  ÿÀŽ äxssˆˆˆˆÿÀŽ â $~†ÿ   ÿ Ž â r‡‡‰   ˆ    Ä |‰       ä |ˆ       ä x…ƒww  ÿÀŽ ä |‹       ä xss22ÿø� ä tƒ‰    Ä t††…    ät‰ˆüúøô…    ä t‹üúøô    ä x…sTv  ÿÀŽ ä xˆÿ   s  ä rƒ…       Ä x…ÿ   U  ä xƒUww  ÿÀŽ äxUUˆˆˆˆÿÀŽ ä r‰†ˆ   ‰    ä |†       ä xs„ d  ‡    Ä |ˆ       ä x„„1u  ‡    ä xUU22ÿø� ä x‡…ww  ÿÀŽ ä t…†ƒ    ät†ˆüúøôƒ    Ä xƒUTv  ÿÀŽ ä t‹Š‡    ätŒŒüúøô‡    ä xUÿ   U  ä xˆ d  ‰    äxŠ1u  ‰    Ä rŒ‹U   Œ    ä xUÿ   Q  ä xQww  ÿÀŽ äxQQˆˆˆˆÿÀŽ â $~Žÿ   ÿ Ž â r……ƒ   †    ä |ƒ       Ä |‡       ä xUUww  ÿÀŽ ä xQQ22ÿø� ä |�       ä t†ƒ    ät‡‡üúøô    Ä t‰ŽU    äxQTv  ÿÀŽ ä tŽ�üúøôU    ä xQÿ   Q  ä r††   ‡    ä r‰‰Q   Ž    Ä xƒ… d  Œ    äxU…1u  Œ    ä x]ˆˆˆˆÿÀŽ ä|Œ       ä x…]ww  ÿÀŽ ä |Ž       Ä xQ† d  ‰    äx]ÿ   ]  ä x††1u  ‰    ä x22ÿø� ä t‡Œ…    ä x‰P   ÿÀŽ Ä tŒŽüúøô…    ä x…]ww  ÿÀŽ ä x]Tv  ÿÀŽ ä |Ž       ä x ‰   pRô ä r‡‡]   Œ    Ä t‰Ž…    ärPPÿ       ä x]ÿ     â &rR   ÿ   âtŽ�üúøô…    â &rRˆR   ÿ   â xPPÿ   ÿÀŽ â &rvS      â rŽ‰]   Ž    â &r]ŠS   R   â >rRÿP   ÿ  â &rT      Ä &r]ƒT   ]   â xT    ÿÀŽ â 0rRÿR    A   ä &ru\      âx T   pRô â 0rVÿV    A   â  xRR   ?  A  â &r\U\   ]   â rT ÿ       ä x…O   ÿÀŽ â  r RV     A  â Er]        Ä xN   ÿÀŽ ä xTTÿ   ÿÀŽ â #r] ]   ÿ   â x …   pRô â &r wL   ÿ   âx    pRö â &rLQL   ÿ   â >rTÿT   ÿ  ä rOOÿ       Ä rNNÿ     € â &r tM       âxS‡ d  Ž    â &rL†M   L   â xOOÿ   ÿÀŽ â 0rTÿT    A   â xNNÿ   ÿÀŽ â &rMsK       âxP‡1u  Ž    â &r‡SK   L   â >rOÿO   ÿ  â  xKT   ?  A  â >rNÿN   ÿ  Ä Er… \       â  r VK     A  â 0rOÿO    A   ä &rM„J   M   ä&r‡PJ   ‡   ä 0rJÿN    A   â #r… …   ÿ   â 0r ÿC    A   â  xCO   ?  A  Ä  xJJ   ?  A  â Er‡ ‡       â  rC      A  â &rOI   ÿ   â r  J     A  â &rIˆI   ÿ   ä &rLvH   O   â#r… ‡   …   â &rHŠH   I   ä &r w[   ÿ   Ä &r[Q[   ÿ   â ErM M       â &rLG   L   ä&rHƒG   H   ä &r[†A   [   ä &rLuD   L   ä&rHUD   H   â #rM   ]   â &rDtA       Ä &r[SF   [   ä0rWÿW    A   â !r''       â &rDsF   D   â ErI H       â &r[PE   [   â  rKW     A  â &rAY   ÿ   ä 0rXÿX    A   â ErG L       â &rD„E   D   â Er[ [       â &rFvZ   A   â  r RW     A  â #rI   ÿ   â &rYˆY   ÿ   â  rEJX     A  â ErD D       â &rF^   F   â #r  G   ÿ   â  rACX     A  â &rYŠZ   Y   â #rEE[      â &rwb   ÿ   â #r AD       â &rFu_   F   Ä &r^ƒ^   Y   ä 0r`ÿ`    A   ä &rtc      â !r% %       â &r^U_   ^   â ErA F       â &rbQb   ÿ   â  r R`     A  â &rsd      â !r$E$       â &rb†c   b   â #rA A   ÿ   â ErE ^       â &r„a      â  r K`     A  â 0riÿi    A   Ä &rbSd   b   â #rE E   ÿ   â Er        â  r Ci     A  â &rbPa   b   ä &rGk   ÿ   â #r     A   â  rDJi     A  â Erb b       â &rGvj   G   â !r# #       â &r wr   ÿ   Ä #rEDb   E   â &rGh   G   ä &rrQr   ÿ   ä &rDt       ä&rGuf   G   ä &rr†   r   ä &rDsg   D   Ä 0roÿo    A   ä &r   ÿ   ä&rAˆ   ÿ   â ErG G       â &rD„e   D   â  r Ro     A  â 0rmÿm    A   ä &rvq      Ä &rAŠq   A   â ErD D       â &rp      â#r  G   ÿ   â &rAƒp   A   â  rCm     A  â &run      ä&rAUn   A   â #r D       â 0r~ÿ~    A   â !r"E"       â Er        â !r! !       â ErE A       â &rkˆk   ÿ   â  r R~     A  â &rAw   ÿ   Ä &rkŠj   k   â #r     ÿ   â &rQ   ÿ   ä &rAtl   A   ä &r{   ÿ   ä &rkƒh   k   ä &r†l      Ä &rAs   A   ä &rv}      ä &rkUf   k   ä &rrSg   r   ä&rS      ä &rA„   A   Ä 0rzÿz    A   ä &r|      â Erk k       â &rrPe   r   âErA A       â &rP      â  rKo     A  â &ru      â  rCz     A  â 0r‚ÿ‚    A   â Err r       â &r{ˆ{   ÿ   â #rk   ÿ   â  rmJm     A  â #rA       â Er        â  r R‚     A  â &r   ÿ   Ä &r{Š}   {   â #rmmr      â &rˆ   ÿ   â  rK~     A  â #r    ÿ   â &rv�      äO &r{ƒ|   {   ä &r w   ÿ   â Er        â &rQ   ÿ   Ä &rŠ�      ä &rww   ÿ   â#rE   ÿ   â &rQQ   ÿ   â  rJz     A  â &r      ä &r{U   {   ä &r ty       Ä &r†y      ä &rƒ      ä &rtt   w   â #r      â &r†   Q   ä &ru      â Er{ {       â &r sx       â x**    ÿàÿ â 0rÿ    A   â  rK‚     A  â &rSx      Ä &rU      â !r       â &rts   t   ä&rS      â Er        â &r„       â r R     A  â &rP      â | *   pbô â 0rÿ    A   â #r{{   ÿ   â Er        â &rt„   t   â rK     A  â &rP      Ä 0rÿ    A   â #r    ÿ   â  r C     A  â#r   ÿ   â  rJ     A  âEr        â  rC     A  â Er        â  rJJ     A  â Ert t       ä Er        â #r        â #r   {   â #rt      â x99€  ÿà÷ â #rJ      â !r&…&       â !r m        â !r       â !r? ?       â !r--       â !r,,       â !r++       â x00@   ÿàÿ Ä x::@   ÿàÿ ä x66@   ÿàÿ ä x<<@   ÿàÿ ä x>>@   ÿàÿ ä x@@@   ÿàÿ ä xBB@   ÿàÿ Ä r7ÿ7   ÿäÿ â D)       À â Gy  àÚÿÿÿÿƒ ê Ay        € ê Ey      € â G‰  ð     € ê $x((À  )Ž Ê ˆs (' ùÿ H   è ˆs (&€ùÿ H   è ˆs (% úÿ H   è ˆs ($€úÿ H   è ˆs (# ûÿ H   è ˆs ("€ûÿ H   è ˆs (! üÿ H   è ˆs ( €üÿ H   è ˆs ( ýÿ H   è ˆs (€ýÿ H   è ˆs (? þÿ H   è ˆs (-€þÿ H   è ˆs (, ÿÿ H   è ˆs (+€ÿÿ H   ä Ay        € ê {          ì M        € ê „y )     H   b z / r  p`ð Ä x/   ÿàÿ â „y) €   H   ¢ r )ÿ   p"p äz  r  p`ò â „y)    H   æ x )   p"ò â „y)    H   ( „y	) €  H   h „y) €  H   ( „y)    H   " 5        Æ „y) €  H   ( „y)    H   è „y) €  H   â !r  '       Æ/ „y)    H   â $v'ÿ r  ÿ Ž â !r&       äO ‰       " !r##       Æ� „y) €  H   â !r%%       Æ„y)    H   â !r		"       Æ„y) €  H   â !r$       Æ ‰      ¢ !r!!       Æ ‰#      â $x)'   ÿŽ â/ !r         â !r        â ‰%      " !r       Æ y       &   b !r       Æ ‰	      " !r??       Æ y      '   b !r
       ÆO ‰      ¢ !r--       Æ ‰!      ¢ !r##       â� !r,,       ä ‰      â !r%%       â !r++       â $z   x  /Ž â/ ‰       " $x/'   ÿŽ â !r	       Ä‰      b $z  ~   Ž Æ‰      " !r       ÆO ‰
      ¢ r'    ÿ† â !r!!       ä ‰?      b rÿÿ   ÿäÿ â !r       ä� ‰"-      â 
  r  ÿàõ ä z j  ÿ† â ‰#      â &ÿÿ   ÿä â !r        â 
 j  ÿ„ â ‰,      " z k  � â !r       â/  /   ÿàù â ‰%      b !r       â
 k  & Ä ‰$+      " !r
       âO ÿÿ   ÿä â $x'   ÿŽ â ‰      ¢ !r??       â
 j  ÿˆ ä ‰!      b !r--"       â� 
 k   Ä ‰	 €    â !r##       â r    ÿàõ ä ‰      â !r,       ä $rÿÿ   ÿ â !r%%       â/ ‰"      " !r++$       Æ‰&      b !r       ÆO ‰$      ¢ r'    ÿˆ â !r!!       ä‰ €    " r0ÿÿ   ÿä â !r	       ä� ‰(?      â 	'   ÿàÿ â !r       Ä ‰*-      b  )   ÿàû ä' 	   ÿà÷ â ‰# €    b ÿÿ   ÿäÿ â !r""       â 
 j  ÿŠ â ‰,      " z j  ÿŒ â !r&&       â/ 2ÿÿ   ÿäÿ â ‰% €    b 
' j  ÿ„ â !r$$       âO 
 k  � â ‰ €    ¢ !r       âz	 k   Ä ‰.+      " !r(?(       â� ' k  2 ä ‰ @    â !r*-*       â)   ÿàÿ ä ‰ €    b !r#       â     ÿà÷ Ä ‰ ! €    b !r,,       â z j  ÿˆ â !r%       â/ ‰#" €    $ z k  0 â !r       âO ‰%$ €    b 0ÿÿ   ÿäÿ â !r.+.       Ä‰'& €    ¢  j  ÿ† â !r       ä� ‰ @    â /   ÿàÿ â !r       ä‰)( €    "     ÿàõ â !r !        Ä ‰+* €    b 2ÿÿ   ÿä ä  j  ÿ„ â ‰-, €    b !r#"#       â  k  0� ä ‰ @    " !r%$%       â/  k  2 Ä ‰/. €    b !r'&'       ÆO ‰ @    ¢ !r       Æ� ‰!  @    â !r)()       Æ‰      " !r+*+       Æ‰ @    b !r-,-       Æ ‰ @    b !r       Æ ‰"# @    " !r/./       Æ/ ‰$% @    b !r       ÆO ‰&' @    ¢ !r! !       Æ� ‰      â !r1       â        ä ‰() @    " !r       ä%  j   Ž â ‰*+ @    b !r       Æ ‰,- @    b !r"#"       Æ ‰      " !r$%$       Æ/ ‰ / @    b !r&'&       ÆO ‰      ¢ !r3       â� %  j  5 Ž â !r()(       â‰      â !r*+*       Æ‰      " !r,-,       Æ ‰ !      b !r)       Æ ‰#"      " !r /        Æ/ ‰%$      b !r+       ÆO ‰'&      ¨ †	 1    â !r       Æ� ‰(      â !r       Æ† 3   â !r!!        Æ‰*      " !r#"#       Æ ‰,      " !r%$%       Æ/ ‰       b !r'&'       ÆO †	 )    è † 
+    â !r(       Æ� †	     è †    â !r*       Æ†	 !    â !r,       Æ † #    â !r        Æ/ †	 %    è † '   è †	     è †     è †	     â M™        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      "   " ¹z  l      â Ey  P&    € â ™x?    â y      !   b ¹z  F   
   â r ÿ       â x      ÿ  â y      %   ¢ x      ÿ  â x      ÿ  â x
      ÿ  â x      ÿ  â $rÿÿ   ÿ Ž â r ÿ   pRð Ä r ÿ   pRò â$x    Ž â/ x   ÿ   ÈO r    ÿ� È xÿ      È |    pbô Ú G)  %    € ê y      '   " $rÿÿ   ÿ Ž â x þÿÿÿÿÀŽ â $rÿÿ   ÿ Ž â y      &   d r  €ÿàÿ â $rÿÿ   ÿ Ž Æ x   ÿ   È x   ÿàÿ È xÿ     â 'z  y   Ž È 'z s   Ž â/ zÿ z     ä x   ÿ   â $z  p  Ž â zÿ t    â $z |  Ž ä$z |   Ž ä $z  v  Ž âxÿ     â $z> v  Ž â x   ÿàÿ â $z }  ÿŽ Ä $z w  ÿŽ â xÿ     â %x>>$   ÿŽ Æ r   ÿàõ â $z p   Ž â r>   ÿà÷ är >   ÿàù â %x@$   ÿŽ È %x~$   Ž â r?   ÿä Æ %x€$   Ž â z  Y      Æ $x?   Ž â z  X      â %x>   >Ž È %x@   @Ž â z  q      Æ $x   ÿ Ž ä $rÿÿ   ÿ ä $x   Ž âr   ÿŽ â$x    Ž âr   ÿŽ ä xÿ     ä xÿ     Ä z" q  ÿàÿ ä x!ÿ     ä r$>   ÿàõ È r%?   ÿä Ê �yn$   ™ ¨  �yl$   ™ è  �yj$   ™ ( �yk$   ™ b r&@   ÿàõ È r'A   ÿä Ê �ym&   ™ è �ys&   ™ â r*   ÿàõ Æ �yu&   ™ â r(   ÿà÷ â $r6ÿÿ    â r-*   ÿàõ â �yz&   ™ â r)   ÿäÿ ä xh $       â $x#   6 È $x%#$   ÿŽ ä %v,- Z  h Ž â �y#(    ‘ â  r2~   ÿà÷ ä r0    ÿàõ ä r--%   ÿàÿ â $x3   Ž â r1   ÿä Æ �y$,   ™ è  �y',   ™ "# r9*   ÿà÷ Æ �y(,   ™ " r.€   ÿàù Æ �y&2    ‘ " r4"   ÿàõ ä r+!6   ÿäÿ â �y),   ™ " r/�   ÿä Æ �y*,    • " r74   ÿàù Æ �y%0    ‘ " rG4   ÿà÷ â $x3+$   ÿŽ ä %v89 Z  h Ž â �y+.    ‘ & $r0ÿÿ    ä/ $x99   3Ž Æ r10   ÿä är0!0   ÿäÿ â %vFG Z  h Ž â�y,8   ™ f $x-0$   ÿŽ â �y.8   ™ h rGG-   ÿàÿ â %v67 Z  h Ž â �y-8   ™ f $x11$   ÿŽ â r<   ÿàù â�y/8    ™ b r:   ÿàû Ä r771   ÿàÿ â �y08    • b rE<   ÿàõ ärQ<   ÿà÷ â�y16   ™ b r<ÿ   ÿäÿ ä r_:   ÿàû â�y26   ™ b $rBÿÿ    â rY:   ÿàù Ä �y36   ™ h �y46   ™ h �y56    • b r8<   ÿä â/$x9!   BŽ âr7!<   ÿäÿ â �y6F   ™ b r:B   ÿä Æ $xC7$   ÿŽ ä �y7F   ™ b %vPQ Z  h Ž È $x99$   ÿŽ ä $x=8$   ÿŽ ä �y8F   ™ b $x;:$   ÿŽ â rQQ9   ÿàÿ ä �y:F    ™ b %vDE Z  h Ž Æ �y9F    • b $xEE   ;Ž â rJ   ÿàù â%vXY Z  h Ž ârH    ÿàû ä �y;D   ™ b %v^_ Z  h Ž â rgJ   ÿàõ äreJ   ÿà÷ â �y<D   ™ b rJÿ   ÿä Ä rFÿ   ÿäÿ â $x__   CŽ â rcH   ÿàù â�yBD   ™ b riH   ÿàû ä rYY=   ÿàÿ â �yCD    • b rGF   ÿä Æ �y=D   ™ b $xF!   FŽ ârH!J   ÿäÿ â $xI   J â �yN^   ™ b %vfg Z  h Ž Æ �yJX   ™ h �yDP   ™ b# $xOF$   ÿŽ Æ �yEP   ™ b $xII$   ÿŽ Æ �yFP   ™ b $xKH$   ÿŽ ä $xMG$   ÿŽ â �yHP    ™ b rggI   ÿàÿ Æ �yGP    • b %vde Z  h Ž Æ �yIX   ™ b %vbc Z  h Ž Æ �yLX   ™ b %vhi Z  h Ž â reeK   ÿàÿ ä �yKX   ™ d riiO   ÿàÿ ä �yO^   ™ b $xcc   MŽ Æ �yMX    • h �yP^   ™ h# �yQ^    ™ h �yR^    • h �ySf   ™ h �yTf   ™ h �yXd   ™ h �yUf   ™ h �yYd   ™ h �y]b   ™ b ‚x  ÿþý     Æ �yVf   ™ b ‚x       â $~rÿ   ÿ Ž ä �yWf    • b |p       Æ �yZd   ™ h �y^b   ™ h! �y_b   ™ h �y`b   ™ h �yab    • b xfÿ   n  ÆO �y[d    ™ ¨ �y\d    • ¨ �ybh   ™ ¢ |q       Æ �ych   ™ ¢  xenww  ÿÀŽ ä/xdnˆˆˆˆÿÀŽ ä tgpe    ätnrüúøôe    ä xffww  ÿÀŽ ä xed22ÿø� Ä |o       â �ydh   ™ ¢  trqüúøôf    äxpeTv  ÿÀŽ ä xqÿ   e  ä toof    â �yeh    ™ ¨  �yfh    • ¢  rggp   n    Ä rnoq   r    ä xrÿ   l  ä� xqlww  ÿÀŽ äxllˆˆˆˆÿÀŽ â $~pÿ   ÿ Ž â |o       â $~tÿ   ÿ Ž â xrrww  ÿÀŽ Ä |v       ä xhl22ÿø� ä toor    ätrpüúøôr    ä titq    ätlvüúøôq    Ä xphTv  ÿÀŽ ä xqÿ   h  ä riip   l    ä xtÿ   j  äxpjˆˆˆˆÿÀŽ ä rloq   r    â $~oÿ   ÿ Ž â |w       Ä |q       ä xrjww  ÿÀŽ ä xttww  ÿÀŽ ä xpp22ÿø� ä xhg d  n    ä |j       Ä xgg1u  n    ä tnwr    ätqqüúøôr    ä toot    äxrpTv  ÿÀŽ ä ttjüúøôt    Ä xji d  l    äxii1u  l    ä xpÿ   p  ä rlnr   q    ä xrÿ   k  äxykww  ÿÀŽ Ä xkkˆˆˆˆÿÀŽ â $~nÿ   ÿ Ž â roop   t    ä |p       ä |q       ä xrrww  ÿÀŽ ä |t       Ä xkk22ÿø� ä tppr    ätqqüúøôr    ä twty    ätnnüúøôy    ä xrkTv  ÿÀŽ Ä xtÿ   k  ä rwwr   n    ä rnpt   q    â $~qÿ   ÿ Ž â xvÿ   m  ä xrmww  ÿÀŽ äxtmˆˆˆˆÿÀŽ Ä xmvww  ÿÀŽ ä xkl d  o    ä|y       ä tvqr    ä xll1u  o    â $~oÿ   ÿ Ž â xqt22ÿø� Ä |p       ä tyyüúøôr    ä xrqTv  ÿÀŽ ä tppm    ätooüúøôm    ä xmw d  n    Ä xnw1u  n    ä rvvr   y    ä xqÿ   q  ä xyÿ   s  ä xwsww  ÿÀŽ äxssˆˆˆˆÿÀŽ Ä ||       ä ropq   o    â $~pÿ   ÿ Ž â |t       ä xss22ÿø� ä |r       ä xyyww  ÿÀŽ Ä t||w    ätwtüúøôw    ä xtsTv  ÿÀŽ ä tqpy    ätrrüúøôy    ä xsÿ   s  Ä xpv d  o    äxov1u  o    ä rqqs   r    ä xrÿ   u  ä xvuww  ÿÀŽ äxuuˆˆˆˆÿÀŽ Ä r||t   w    â $~wÿ   ÿ Ž â |{       ä xuu22ÿø� ä |y       ä |s       ä xrrww  ÿÀŽ Ä tt{v    ätwwüúøôv    ä xxuTv  ÿÀŽ ä tvyr    ätssüúøôr    ä xuÿ   u  Ä rttx   w    ä xr| d  q    äxxzˆˆˆˆÿÀŽ ä xq|1u  q    â $~|ÿ   ÿ Ž â ruvu   s    ä |v       Ä xwzww  ÿÀŽ ä xzÿ   z  ä xxx22ÿø� ä xy#   ÿÀŽ ä xst d  u    äxtt1u  u    â $~{ÿ   ÿ Ž â tu|w    Ä tvvüúøôw    ä |w       ä xzzww  ÿÀŽ ä x|xTv  ÿÀŽ ä x y   pRô ä twwz    Ä rvu|   v    ä r##ÿ       ä tz{üúøôz    ä xuÿ   x  â &rxh$   ÿ   âx##ÿ   ÿÀŽ â &r$p$   ÿ   â rwwu   z    â &rug'   x   â >r#ÿ#   ÿ  â &ryo'   $   Ä &r$j(   u   âxu&   ÿÀŽ â &rxr(   y   ä 0r(ÿ#    A   â x u   pRô â &r#i)   $   ä 0r*ÿ*    A   â  x((   ?  A  â r&&ÿ       Ä xu%   ÿÀŽ ä x'v d  w    äEr$ #       â  r#(*     A  â xvv1u  w    ä xw+   ÿÀŽ ä x&&ÿ   ÿÀŽ Ä x u   pRô â #r$#$   ÿ   â x w   pRö ä >r#ÿ&   ÿ  ä r&%ÿ       â &r%k,   ÿ   âr++ÿ     € ä x&&ÿ   ÿÀŽ â &r,s,   ÿ   â x++ÿ   ÿÀŽ â &r%l-   %   â >r&ÿ&   ÿ  â 0r#ÿ#    A   Ä &r,t-   ,   â >r-ÿ+   ÿ  â &r)q)   x   ä &r+m.   %   â x%#   ?  A  â 0r&ÿ&    A   â Erx )       â &r+n/   +   â  r#*%     A  â 0r0ÿ0    A   â  x&&   ?  A  â &r.'.   ,   â Er+ +       â 0r,ÿ-    A   â #r-#x   ÿ   â &r*h1   ÿ   â r#&0     A  â &r1p1   ÿ   Ä &r*g2   *   â#r+#+   $   â &r1o2   1   ä &r#k6   ÿ   ä &r1r3   1   ä&r*j3   *   ä &r#l7   #   Ä &r.v/   .   ä &r1q4   1   ä0r5ÿ5    A   ä &r*i4   *   ä &r/m8   #   ä &r6s6   ÿ   â Er1 1       â &r/n:   /   â Er* *       â 0r9ÿ9    A   â  r$%5     A  â &r6t7   6   â  r#(5     A  â Er/ /       â #r1$1   ÿ   â &r3'8   6   â #r##*   ÿ   â  r$&9     A  Ä &r3v:   3   ä #r$$/   #   â &r#h;   ÿ   â x),   ?  A  â &r;p;   ÿ   â Er3 3       â !r	$	       â  r*)9     A  â &r#g<   #   ä&r;o<   ;   Ä &r$kD   ÿ   â Er. .       â  r00)     A  â #r1*3   1   â &r#j=   #   ä&r;r=   ;   ä &rDsD   ÿ   ä &r*lE   $   â #r-0.   -   â !r+       â &r$iB   #   Ä &r;qB   ;   ä &rDtE   D   ä 0rCÿC    A   ä &r+mF   *   â!r-       â &r-'F   D   â Er* $       â &r+nH   +   âEr; ;       â 0rGÿG    A   â  r$%C     A  â r#(C     A  â &r-vH   -   â Er+ +       â #r;$;   ÿ   â &r,hI   ÿ   â #r##*   ÿ   â  r$&G     A  â &rIpI   ÿ   â Er- -       â  r*)G     A  â #r$$+   #   â &rIoJ   I   Ä &r,gJ   ,   ä &r#kN   ÿ   â #r-*-   ;   â &rIrK   I   ä&r,jK   ,   ä &r*lO   #   ä &rIqL   I   Ä &r,iL   ,   ä 0rMÿM    A   ä &r*mP   *   â !r$       â ErI I       â  r$%M     A  â &r*nQ   *   â Er, ,       â 0rRÿR    A   â  r#(M     A  â #rI$I   ÿ   â &r+hS   ÿ   âEr* *       â #r##,   ÿ   â &rSpS   ÿ   â  r$&R     A  â &r+gT   +   Ä &rSoT   S   â #r$$*   #   â &r#kX   ÿ   ä &rSrU   S   ä&r+jU   +   ä &r,lY   #   ä &rhh]   ÿ   Ä &r]p]   ÿ   ä &rSqV   S   ä&r+iV   +   ä 0rWÿW    A   ä &r,mZ   ,   ä &r]o^   ]   Ä &rkkb   ÿ   äO &rNsN   ÿ   ä&rXsX   ÿ   ä &rhg^   h   ä &rbsb   ÿ   â !r$       â ErS S       â &r,n[   ,   â Er* +       â  r$%W     A  â0r\ÿ\    A   â  r#(W     A  â &r]r_   ]   Ä &rNtO   N   ä&rXtY   X   ä &rhj_   h   ä &rklc   k   ä&rbtc   b   â #rS$S   ÿ   â Er, ,       â &r]q`   ]   â x    ÿàÿ â &rN'P   N   â #r##*   ÿ   â &rX'Z   X   â  r$&\     A  â &rhi`   h   Ä 0raÿa    A   ä &rkmd   k   ä&rb'd   b   â |    pbô â &rNvQ   N   â #r$$,   #   â &rXv[   X   â Er* ]       â &rkne   k   âEr+ h       â &rbve   b   â  r#%a     A  â0rfÿf    A   â  r((a     A  â ErN N       â  rR)R     A  â #r*#*   ÿ   â  r\)\     A  âErX X       â #r((+   ÿ   â Erk k       â  r#&f     A  â Erb b       â  r))f     A  â #rIRN   I   â #rS\X   S   â #r(#k   (   â #r))b   *   â x€  ÿà÷ â !r
1
       â !r-       â !rI       â !r$       â !rS       â !r(       â !r)       â x@   ÿàÿ Ä x@   ÿàÿ ä x@   ÿàÿ ä x  @   ÿàÿ ä x""@   ÿàÿ ä x@   ÿàÿ ä rÿ   ÿäÿ â D)       À â Gy  Þÿÿÿÿƒ ê Ay        € ê Ey  ð     € â G‰  Ð     € ê $x€  Ž Ê ˆs  úÿ H   è ˆs €úÿ H   è ˆs 	 ûÿ H   è ˆs 
€ûÿ H   è ˆs  üÿ H   è ˆs €üÿ H   è ˆs  ýÿ H   è ˆs €ýÿ H   è ˆs  þÿ H   è ˆs €þÿ H   è ˆs  ÿÿ H   è ˆs €ÿÿ H   ä Ay        € ê {          ì M        € ê „y     H   " z  r  p`ð Ä z#  r      â „y €   H   b r ÿ   p"p Æ „y    H   ¢ $x'#   ÿŽ ä$x)#   ÿŽ â „y €  H   è „y    H   ( „y €  H   h „y    H   h „y €  H   h „y    H   h „y €  H   è „y    H   â !r       â !r       Ä/ „y €  H   â !r	       ÆO ‰       " !r
       Æ� ‰      b !r       Æ‰	      ¢ !r       Æ‰      â !r       Æ ‰      " !r       Æ ‰      b !r       â !r         ä ‰      " !r       â !r       ä/ ‰       b !r       â !r		       ÄO ‰
      ¢ !r       Æ� ‰      â !r       â !r       ä‰      " !r       Æ‰      b x   ÿàÿ Æ ‰      b !r       â z  r  p`ò â !r        â/ ‰	      $ x    p"ò â !r
       âO y       &   b !r       Æ� ‰      ¢ !r       Æy      '   â !r       Æ‰      " !r       Æ ‰      h ‰      b !r	       Æ ‰      " $z   x  Ž Æ/ ‰      h ‰      b $z  ~   Ž â� !r       äO ‰      ¢ !r       â '   ÿàõ Ä ‰!      â !r       âÿÿ   ÿä ä ‰ €    " !r       â 
 j  ÿ„ ä ‰ €    b !r       â 
 k   â !r       â/ ‰ €    " !r       Æ ‰ €    b !r       ÆO ‰ €    ¢ !r!!       Æ� ‰ €    â !r       Æ‰ €    " 
  r  ÿà÷ â !r
       ä‰ €    b r#    ÿˆ â $ÿÿ   ÿŽ â 
 j  ÿ† â‰ €    b rÿÿ   ÿä ä 
 k  � â ‰  €    b !r       â !r       â/ z j  ÿˆ â ‰" €    " !r       ÄO z k   â ‰$!      b !r       â�  )   ÿàù ä ‰	 @    ¢ !r       â&ÿÿ   ÿä ä ‰
 @    â !r       â!r       Ä ‰ @    " !r         Æ ‰ @    b !r""       Æ ‰ @    " !r$!$       Æ/ ‰ @    b !r	       ÆO ‰ @    ¢ 	#   ÿàÿ â!r
       ä� ‰ @    â r#    ÿ„ ä * 	   ÿà÷ â ‰ @    â !r       â
 j  ÿˆ â $r(ÿÿ   ÿ â ‰!  @    " !r       â+ÿÿ   ÿäÿ Ä ‰#" @    b !r       â * j  ÿ† ä ‰%$ €    " !r       â/ 
	 k  & ä ‰      b !r       âO * k  +� Ä ‰      ¢ !r       â� z
 j  ÿ„ ä‰      â !r       â z k  ( â !r! !       â‰      " !r#"#       Æ‰      b !r%$%       Æ ‰      " !r+       Æ/ ‰      b '   ÿàÿ â $t'ÿ   ÿ Ž â !r-       âO ‰      ¢     ÿàõ â%v  j  ' Ž â !r       â� ‰      â &ÿÿ   ÿä ä  j  ÿ„ â ‰ !      â !r       Ä k  & â ‰$% @    " !r       Æ‰"#      b !r       â !r       ä/ †	 +    â !r       ÆO † -   â !r       Æ� †	     â !r!!        Æ †     â !r$%$       Æ†	     â !r##"       Æ†    è †	     è †     è †	 
    è † 
!   è ‰$      h  †	 #    â M™        € ê x)   ÿàÿ â !r$       Æ/ r    ÿàñ È r ÿÿ   ÿä  ä z j  ÿ€ È z k     Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      "   " ¹z  l      â Ey   "    € â ™x?    â y      !   b ¹z  F   
   â r  ÿ       â $r#ÿÿ   ÿ Ž â y"      %   ¢ r1 ÿ       â $r3ÿÿ   ÿ Ž â r5 ÿ       â y      &   â x      ÿ  â x      ÿ  â $rÿÿ   ÿ Ž â y      '   " r ÿ   pRð Ä r ÿ   pRò â$x$    Ž â/ x""   ÿ   ÈO r $$   ÿ� È x!ÿ      È | !   pbô Ú G)  0!    € ê $rÿÿ   ÿ Ž â�r ÿ       â $z" p  !Ž â x	 þÿÿÿÿÀŽ â 'z y   Ž â z  X      ä r  ÿ       â 'z s   Ž â zÿ z    Æ $x$$   	
Ž â zÿ t    â $z  |  !Ž â x	"   ÿàÿ â $z |  Ž â x$   ÿ   â$z  v   Ž âx/ÿ   $  â $z v  Ž â x-   ÿàÿ â $x$   ÿ Ž Ä %x$   ÿŽ â xÿ   -  ä xÿ     â $z% }  ÿŽ â r&   ÿà÷ â$z w  ÿŽ â xÿ     â $z	 p   Ž â r(   ÿàù â $x4!   ÿ Ž â r%%   ÿàõ â %x$   Ž â r'   ÿäÿ Ä z  q      â %x$   ÿŽ â x-ÿ   -  ä r04   ÿŽ â %x
$   Ž â z24 q  ÿàÿ ä x,ÿ   -  â $x)    ä %x   Ž È %x   Ž È $v+ÿ Y  ÿ Ž ä $x*   4Ž ä$x.   4Ž ä $r`ÿÿ   ÿ ä r   ÿàõ È r+   ÿä Ê �y    ™ ¨  �yM   ™ è  �yJ   ™ ( �yE   ™ b r   ÿàõ Ê $x   + Ê �yO   ™ è �yL   ™ è �yD   ™ â r64%   ÿàõ ä r&   ÿà÷ â �yC   ™ " r7ÿ`   ÿä Ä r$6   ÿàõ â $x+   'Ž ä $t^ÿ$   ÿ Ž â r/7   ÿä ä �yB    ‘ " %v Z  ^ Ž È $x$   ÿŽ â r   ÿàõ È r   ÿàÿ â $x	   + â r-6   ÿàõ Æ/ �yG   ™ " r@(   ÿà÷ Æ �yH   ™ " r
   ÿàù ä r,7   ÿä â �yK   ™ b rA+)   ÿäÿ Æ �yF    ‘ b $x   + ä %v Z  ^ Ž â �yN   ™ f $x$   ÿŽ â �yI    • f $x   Ž â �y@@    ‘ b r2%   ÿàõ Æ �y?    ‘ b rÿ`   ÿä ä r$   ÿàõ â�y<   ™ b r-   ÿà÷ Æ $x/    â/ �y>   ™ b r,   ÿäÿ â %v Z  ^ Ž ä�yA   ™ d $x$   ÿŽ ä %v Z  ^ Ž â �y=    • f $x   Ž â �y:    ™ b $x$   ÿŽ â r*%   ÿàõ Ä �y7   ™ d r   ÿàÿ â $rTÿÿ   ` â �y9   ™ b r$   ÿàõ Æ �yR   ™ b r/T   ÿä Æ �y6   ™ b r-   ÿàõ Æ �yP   ™ b %v Z  ^ Ž Æ �y;   ™ b $xW$   ÿŽ â rT,T   ÿä ä �yQ    • b r.%   ÿà÷ Æ �yS   ™ b %v Z  ^ Ž â rW   ÿàÿ ä �y8    ™ b $rdÿÿ   `Ž Æ �yU    • b r$   ÿà÷ Æ/ �yV   ™ h �yZ   ™ b $xT$   ÿŽ Æ �yT   ™ b $x   Ž ä $x/   dŽ â �y\    • b %v Z  ^ Ž Æ �y[   ™ b $xW$   ÿŽ â ra0%   ÿà÷ ä �y_   ™ b r-   ÿàõ ä rW   ÿàÿ â �yY   ™ b r$a   ÿàù â$rbÿÿ   `Ž â ra-a   ÿàû â �yW   ™ b rd,d   ÿä â $xc/   b Ä �yX    ™ b %v Z  ^ Ž â rb,b   ÿäÿ ä �y]    • b %va Z  ^ Ž â/‚x       ä �yh    • b $xad$   ÿŽ ä %v Z  ^ Ž Ä �y^   ™ b r        â $x   aŽ ä $xcc$   ÿŽ â �y   ™ b $xeb$   ÿŽ Æ �yd   ™ b rc   ÿàÿ â $x   eŽ ä �ye   ™ h �yc   ™ h �yf   ™ b ‚x  ÿþý     Æ �y   ™ b $~mÿ   ÿ Ž Æ �yb   ™ b |o       Æ �yk   ™ h �yj   ™ h �yi   ™ h �yl    • h �ya    ™ h �yg    • h �y   ™ b# |p       Æ �y    • b $~ÿ   ÿ Ž Æ �y   ™ b xÿ      äO x ww  ÿÀŽ ä xww  ÿÀŽ ä tmm    ätoüúøô    ä to    Ä �y   ™ ¢ tppüúøô    Æ �y    ™ ¢ x  ˆˆˆˆÿÀŽ È x  22ÿø� È xn Tv  ÿÀŽ ä x ÿ      ä rmn       ä roo    p    ä xmÿ   M  ä� x Mww  ÿÀŽ Ä xMMˆˆˆˆÿÀŽ â $~qÿ   ÿ Ž â |s       â $~uÿ   ÿ Ž â xMM22ÿø� ä xpmww  ÿÀŽ ä |w       ä tnq     Ä tmsüúøô     ä x MTv  ÿÀŽ ä tqup    ätpwüúøôp    ä xMÿ   M  ä x d  o    Ä/ x1u  o    ä rnn    m    ä xÿ   J  äx Jww  ÿÀŽ äxJJˆˆˆˆÿÀŽ ä rqqM   p    â $~Mÿ   ÿ Ž â |m       â $~oÿ   ÿ Ž â xww  ÿÀŽ Ä xJJ22ÿø� ä tpM     ätmmüúøô     ä too    äx JTv  ÿÀŽ ä tsüúøô    Ä xJÿ   J  ä rpp    m    ä rooJ       ä xÿ   E  äx Eww  ÿÀŽ ä xMn d  q    Ä xEEˆˆˆˆÿÀŽ ä xnn1u  q    â $~qÿ   ÿ Ž â |m       â $~uÿ   ÿ Ž â xEE22ÿø� ä |s       ä xJww  ÿÀŽ Ä tm     ätqqüúøô     ä x ETv  ÿÀŽ ä trsJ    ätsuüúøôJ    ä xJÿ   E  Ä rE    q    ä xÿ   O  ä x Oww  ÿÀŽ äxmp d  o    äxOOˆˆˆˆÿÀŽ â $~qÿ   ÿ Ž â xpp1u  o    Ä |o       ä rrrJ   s    ä xOO22ÿø� ä |s       ä xJww  ÿÀŽ ä to     Ä toqüúøô     ä x OTv  ÿÀŽ ä tqsJ    ättuüúøôJ    ä xOÿ   O  ä r     o    Ä roqO   t    ä xOÿ   L  ä xLww  ÿÀŽ â$~qÿ   ÿ Ž â xLLˆˆˆˆÿÀŽ â $~tÿ   ÿ Ž â |s       ä xJE d  r    Ä xLL22ÿø� ä xEE1u  r    ä xOOww  ÿÀŽ ä |v       ä trq    ätqsüúøô    Ä xLTv  ÿÀŽ ä tstO    ättvüúøôO    ä xuÿ   L  ä xO  d  o    äxL 1u  o    Ä rrr   q    ä xÿ   D  ä x Dww  ÿÀŽ ärosu   t    â $~sÿ   ÿ Ž â xDDˆˆˆˆÿÀŽ â $~tÿ   ÿ Ž â |u       Ä xDD22ÿø� ä xqww  ÿÀŽ ä ts     ätsuüúøô     ä x DTv  ÿÀŽ ä tttq    Ä tuvüúøôq    ä xDÿ   D  ä r     s    ä rutD   u    ä |t       ä xsCww  ÿÀŽ Ä xqr d  o    äxrr1u  o    ä xo  d  u    äxD 1u  u    ä xCˆˆˆˆÿÀŽ ä t ts    Ä tsvüúøôs    ä xvB   ÿÀŽ ä xCÿ   C  ä x22ÿø� â $~uÿ   ÿ Ž â x v   pRô ä xtCww  ÿÀŽ Ä xCTv  ÿÀŽ ä rBBÿ       ä tuut    ättwüúøôt    ä rs C   s    â &rCG   ÿ   âxÿ     â &rGOG   ÿ   â xBBÿ   ÿÀŽ â &r H   C   â rtu   t    â &rHLH   G   â >rÿB   ÿ  â &rGMK       â xBs d  t    Ä xCs1u  t    â &rKqK   H   â xsF   ÿÀŽ â 0rHÿ    A   ä &r nN   G   âx s   pRô â &rNrN   K   â xK@   ÿÀŽ â 0rÿI    A   â  xGH   ?  A  â rFFÿ       Ä xs?   ÿÀŽ ä ErI         â  r G     A  â x K   pRô ä xHFÿ   ÿÀŽ ä x s   pRö â #rF I   ÿ   â r@@ÿ       Ä >rHÿH   ÿ  ä r ?ÿ     € â &r?m<   ÿ   âx@@ÿ   ÿÀŽ â &rIo<   ÿ   â x  ÿ   ÿÀŽ â &r?p>   ?   ä 0rHÿH    A   â >r@ÿ@   ÿ  â &rID>   I   â >r ÿ    ÿ  â &r>JA   ?   â  x?H   ?  A  â 0r<ÿ@    A   Ä 0r@ÿ     A   â  r ?     A  â &rIBA   I   ä &r7   ÿ   ä&r7O7   ÿ   â ErA N       â 0r=ÿ=    A   â  x<<   ?  A  â &rIC:   I   Ä &r9      ä&r>E:   >   â  x:@   ?  A  â &r9L9   7   ä &r7mR   ÿ   â #r  A   ÿ   â &rM6      â rA<=     A  â ErI I       â &r9q6   9   â  r==:     A  â &r7pP   7   Ä &rn;      â #r =I       â &r;r;   9   ä 0r6ÿQ    A   ä &r9JS   7   â !r        â Er7        â 0rUÿU    A   â  r G6     A  â &r9E8   9   â Er; ;       â  r?6     A  â #r  7   ÿ   â  r7<U     A  â Er9 9       â #r;   ÿ   â &r;T   ÿ   Æ #r779       â &r;V   ;   ä &r m[   ÿ   â Er> >       â &r=oR   ÿ   ä &r;MZ   ;   ä &r6p_       â #rAA>   F   â &r=DP   =   Ä &r;nW   ;   ä &r6JY   6   ä 0r\ÿ\    A   â !rA       â &r=BS   =   â Er; ;       â &rAOT   ÿ   ä &r6EX   6   â  r G\     A  â 0r]ÿ]    A   Ä &r=C8   =   â !r7       â &rALV   A   â Er6 6       â &r[o[   ÿ   â #r  ;   ÿ   â &r8^   ÿ   â r7<]     A  â &r^O^   ÿ   â Er= =       â &rAqZ   A   â  rU:U     A  â &r[D_   [   â #r776       â &r8   8   Ä &r^L   ^   ä &r md   ÿ   â #rU=      â &rArW   A   ä &r[BY   [   â !r7       â &rMe   8   ä&reqe   ^   Ä &r7pc       â !r       â &r[CX   [   â ErA A       â &rf   ÿ   â  r?\     A  â &rn      ä&rer   e   ä &rOOf   ÿ   Ä &rJb   7   â Er[ [       â 0rhÿh    A   â #rA   ÿ   â &rk      â  r]:]     A  â &rmm   ÿ   ä &rdod   ÿ   ä &rOLk   O   Ä &ro   ÿ   â Er        â &rEa      â  r Gh     A  â 0rgÿg    A   â #r][      â &rMj      ä &rmp   m   ä &rcDc   d   Ä &rOqj   O   ä &rD      â Er        â #r     ÿ   â  r<g     A  â &rni      â x!!    ÿàÿ â &rJJ   m   âO !r55       â Ere e       â &rbBb   c   â  r?h     A  â 0rlÿl    A   Ä &rOri   O   ä &rB      â #r       â | !   pbô â &rbCa   b   â Er        â 0rÿ    A   â #ree   ÿ   â &rJE   J   â r Gl     A  â&rC      â ErO O       â  r?l     A  â #r    ÿ   â Erb b       â  r<<     A  â ErJ J       â #rO   ÿ   â Er        â  r :g     A  â r:     A  â #r<<J      â x€  ÿà÷ â #r  b   e   â #r      â !r33       â !r#<#       â !r1 1       â !r         â x**@   ÿàÿ Ä x..@   ÿàÿ ä x00@   ÿàÿ ä x22@   ÿàÿ ä x44@   ÿàÿ ä r+ÿ+   ÿäÿ â D)       À â Gy  Àáÿÿÿÿƒ ê Ay        € ê�Ey  Ð     € â G‰  °     € ê $x @  Ž Ê ˆs   ûÿ H   è ˆs  €ûÿ H   è ˆs   üÿ H   è ˆs  €üÿ H   è ˆs   ýÿ H   è ˆs  5€ýÿ H   è ˆs  3 þÿ H   è ˆs  1€þÿ H   è ˆs  # ÿÿ H   è ˆs   €ÿÿ H   ä Ay        € ê {          ì M        € ê „y     H   b z " r  p`ð â $z x  "Ž Ä „y €   H   ¢ r ÿ   p"p â $v!ÿ r  ÿ Ž ä „y	    H   ä $x!   ÿŽ ä „y €  H   ( „y    H   h „y €  H   ( „y    H   ( „y    H   ( „y
 €  H   è „y €  H   â !r       â/ !r       ÈO ‰       " !r	       Æ� ‰	      b !r       â!r       ä‰      ¢ !r5       Æ ‰      â !r#       Æ ‰      " !r3       â !r         ä ‰      " !r

1       Æ ‰       b !r		       â/ !r         ä ‰      h ‰      b !r       ÆO ‰
      ¢ !r       Æ� ‰       â !r       â!r       ä ‰	      " !r        Æ‰      " !r       Æ/ ‰  €    b !r       Æ ‰      b !r
       ÆO ‰      ¢ !r        Æ� ‰      è ‰
      â !r	       Æ ‰      " !r       Æ‰      " !r         Æ/ ‰      b !r       â!r       äO ‰	 €    ¢ !r       Æ� ‰ €    â !r
       Æ ‰ €    b x
"   ÿàÿ â !r       ä ‰  @    " z 
 r  p`ò â $z
 ~  Ž â !r       â‰ €    " x    p"ò â !r       Ä/ ‰ €    b r!
   ÿŠ ä 

 r  ÿàù â ‰ €    b !r	       âO 	
   ÿàõ â $r"ÿÿ   ÿŽ â ‰ €    ¢ !r       â�  ÿÿ   ÿä Ä ‰ €    â !r       âz j  ÿŠ ä ‰ €    b !r         â r!
   ÿ† â !r       â‰ @    " 
	 j  ÿ„ â !r       Ä/ ‰ @    b rÿÿ   ÿäÿ â !r       ä ‰       " z j  ÿ† â!r       äO ‰ @    ¢ z k  � â !r       Ä� ‰ @    â 
		 k    â !r       ä‰ @    b $ÿÿ   ÿä ä z k  "� â ‰ @    b !r       â 
 j  ÿˆ Ä ‰ @    " !r       â/    ÿàÿ ä ‰ @    b !r#        â!   ÿàÿ ä ‰ @    " !r       âO !
   ÿàõ â!r       â� ‰       ¢ 
   ÿà÷ â !r       Ä‰      â "ÿÿ   ÿä â !r       â $tÿ   ÿ Ž â ‰      b ! j  ÿ„ â !r       â %v

 j   Ž â ‰      "  ÿÿ   ÿäÿ â !r       â/ 
 k  $ â ‰      b  j  ÿ† â !r       Ä‰      "  k   � â !r%        äO ‰      ¢ ! k  " â !r       ä� ‰      è ‰      â !r       â!r       Ä †	 
#    â !r       Æ/ † 
%   â !r       Æ†	     â !r       ÆO †     â !r       Æ� †	     â !r       Æ †    è †	     è †     è †	     â M™        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      "   " ¹z  l      â Ey  à    € â ™x?    â y      !   b ¹z  F   
   â r$ ÿ       â $r-ÿÿ   ÿ Ž â y!      %   ¢ $r.ÿÿ   ÿ Ž â x      ÿ  â x      ÿ  â y      &   â r ÿ       Æ y      '   " r ÿ   pRð är ÿ   pRò â$x     Ž ä/ $x!!   ÿ Ž ÆO r      ÿ� È xÿ      È |    pbô Ú G)  �    € ê $rÿÿ   ÿ Ž â�r ÿ       â $z" }  ÿŽ âx	 þÿÿÿÿÀŽ â 'z y   Ž È 'z s   Ž â zÿ z    Æ $z ! p  Ž â zÿ t    â $x     	
Ž ä $z |  Ž ä$z |   Ž â x#    ÿ   â $z v  Ž âx !   ÿàÿ â $z v  Ž â x    ÿ   â $z w  ÿŽ â x#   ÿú� â $z  p  Ž â xÿ     â %x$   ÿŽ â x	ÿ     Ä r""   ÿàõ â %x
$   Ž âr&   ÿà÷ är(   ÿàù â %x$   Ž Æ r,	   ÿä â %x$   ÿŽ È $v ÿ q  ÿ Ž ä $x+   ÿ Ž ä $x*   Ž ä $vÿ X  ÿ Ž â r +   ÿŽ â $v%ÿ Y  ÿ Ž â z)+ q  ÿàÿ â $r$ÿÿ   ÿ Ž Ä $x'    +Ž ä $r_ÿÿ   ÿ ä %x#   Ž È %x#   Ž È r   ÿàõ Ê $x   % Ê �yW   ™ ¨  �yV   ™ è  �yU   ™ ( �yT   ™ b r   ÿàõ Ê $x	   % Ê �yS   ™ è �yR   ™ ( �yQ   ™ " r/+"   ÿàõ ä r&   ÿà÷ â �yM   ™ " r1ÿ_   ÿä Ä xJÿ      ä r /   ÿàõ ä r%*   ÿäÿ â $tXÿ$   ÿ Ž ä $x J   1 ä �yO    ‘ " %v Z  X Ž È $x $   ÿŽ È $x   Ž â r
   ÿàù È �yN   ™ b $x   % â xKÿ   #  ä �yP   ™ b r8   ÿà÷ ä r(   ÿàõ â/ �yB   ™ b r/K/   ÿàù â $x9   %Ž â xLÿ   K  â �y     ‘ b r%,   ÿä Ä r0L1   ÿä â �yA   ™ h �y>    • b %v/ Z  X Ž Æ �y7    ‘ b $x/0$   ÿŽ Æ �y88    ‘ b $x   /Ž â r)"   ÿàõ È/ �y:   ™ b rÿ_   ÿä ä r    ÿàõ â�y6   ™ b rK   ÿà÷ Æ $xJ    â �y;   ™ b rL   ÿäÿ â %v Z  X Ž ä�y9    • d $x$   ÿŽ ä �y5    ™ b %v Z  X Ž È $x   Ž ä $x/$   ÿŽ Æ �y0   ™ d r/   ÿàÿ ä r"   ÿàõ â �y4   ™ h �y2   ™ b $rCÿÿ   _ â r    ÿàõ Ä �y/   ™ h �y3   ™ b rJC   ÿä Æ/ �y@   ™ b rK   ÿàõ Æ �y?   ™ h �y1    • b %v Z  X Ž Æ �y<    ™ b $xE$   ÿŽ ä $xCL   C â �y=    • b %v Z  X Ž â rE   ÿàÿ Æ $xCC$   ÿŽ â r'"   ÿàõ â �yD   ™ d $x   CŽ ä �yC   ™ b rYÿ_   ÿä ä r    ÿàõ â/�yG   ™ h �yE   ™ b $xJ   Y â rK   ÿàõ Ä �yH   ™ d rYLY   ÿä ä �yI   ™ b %v Z  X Ž Æ �yF    • b $x[$   ÿŽ Æ �yJ   ™ b %v Z  X Ž Æ �yK    • b $xYY$   ÿŽ ä $x   [Ž â �yL    ™ d rY   ÿàÿ ä �y   ™ h �y   ™ h �yX   ™ h �y\   ™ h �y[   ™ h �y^   ™ h �yZ   ™ h �y    • h! �yY    ™ h �y]    • b ‚x       Ä ‚x  ÿþý     â $~aÿ   ÿ Ž â |c       â $~eÿ   ÿ Ž â |g       ä |b       â $~dÿ   ÿ Ž â x`ÿ   W  äO xWww  ÿÀŽ Ä xWWˆˆˆˆÿÀŽ ä x``ww  ÿÀŽ ä xWW22ÿø� ä ta    ätcüúøô    ä xWTv  ÿÀŽ Ä te`    ä/tgüúøô`    ä xWÿ   W  ä r       ä xÿ   V  ä� xVww  ÿÀŽ Ä rW       â $~Wÿ   ÿ Ž â xVVˆˆˆˆÿÀŽ ä |a       â $~`ÿ   ÿ Ž â xVV22ÿø� ä xww  ÿÀŽ ä tW    Ä tWaüúøô    ä xVTv  ÿÀŽ ä t``    ätabüúøô    ä x d      äx1u      Ä r   W    ä xÿ   U  äxUww  ÿÀŽ â$~Wÿ   ÿ Ž â xUUˆˆˆˆÿÀŽ ä xVÿ   V  ä xUU22ÿø� Ä ra`V   a    ä xww  ÿÀŽ ä tVW    ätWcüúøô    ä xUTv  ÿÀŽ ä t`e    Ä tcgüúøô    ä xUÿ   U  ä rV   W    ä rc`U   c    ä xVÿ   T  äxUTww  ÿÀŽ Ä xTTˆˆˆˆÿÀŽ â $~eÿ   ÿ Ž â x d  a    äx1u  a    ä |b       ä |a       ä x`Vww  ÿÀŽ Ä xTT22ÿø� ä tVbU    ätaa`    ätbeüúøô`    ä tWdüúøôU    ä x`TTv  ÿÀŽ Ä xdÿ   T  ä xU d  c    äxT1u  c    ä rWV`   W    ä xVÿ   S  ä xSww  ÿÀŽ Ä xSSˆˆˆˆÿÀŽ â $~cÿ   ÿ Ž â rbad   b    ä |a       ä xSS22ÿø� ä |e       ä xVVww  ÿÀŽ Ä t`a    ätacüúøô    ä xSTv  ÿÀŽ ä tceV    ätdgüúøôV    ä xeÿ   S  Ä xVW d  b    äxSW1u  b    ä ra`   a    ä xWÿ   R  ä xRww  ÿÀŽ ärdce   d    â $~cÿ   ÿ Ž â xRRˆˆˆˆÿÀŽ â $~bÿ   ÿ Ž â |e       Ä xRR22ÿø� ä |f       ä xWWww  ÿÀŽ ä t`c    ätceüúøô    ä xRTv  ÿÀŽ Ä tbbW    ätefüúøôW    ä xRÿ   R  ä r`   c    ä rebR   e    ä x`ÿ   Q  Ä xRQww  ÿÀŽ äxQQˆˆˆˆÿÀŽ â $~cÿ   ÿ Ž â |i       â $~kÿ   ÿ Ž â xQQ22ÿø� ä x``ww  ÿÀŽ ä xWa d  d    Ä xaa1u  d    ä tbcR    ätcgüúøôR    ä xdQTv  ÿÀŽ ä tfi`    ätgküúøô`    Ä xQÿ   Q  ä rbbd   c    ä rQfQ   g    â $~gÿ   ÿ Ž â x` d  e    äxR1u  e    ä |e       Ä xdMww  ÿÀŽ äxMˆˆˆˆÿÀŽ ä xcb d  Q    äxMÿ   M  ä xfO   ÿÀŽ ä xQb1u  Q    Ä tbed    â$~iÿ   ÿ Ž â tegüúøôd    ä x22ÿø� ä |g       ä xdMww  ÿÀŽ ä x f   pRô Ä xMTv  ÿÀŽ ä tfgd    ärOOÿ       ä tgiüúøôd    ä xdÿ     â &rN   ÿ   ârbbM   e    â &rNWN   ÿ   â xMOÿ   ÿÀŽ â &rOP      Ä &reaP   N   â >rÿM   ÿ  â &rOB   O   ä&re`B   e   â xB    ÿÀŽ â 0rÿ    A   â rgfd   g    â &rOA   O   âx B   pRô â &reRA   e   â xNb d  g    â0rAÿ>    A   â xMb1u  g    â  x>   ?  A  â xP7   ÿÀŽ Ä xb8   ÿÀŽ ä r ÿ       â  r >A     A  â ErO O       ä x P   pRô ä x b   pRö ä xBÿ   ÿÀŽ â #r O   ÿ   â r 7ÿ       Ä r88ÿ     € â &r7U:   ÿ   âx  ÿ   ÿÀŽ â &r:c:   ÿ   â x88ÿ   ÿÀŽ ä >rBÿB   ÿ  â &r7T6   7   â >r ÿ    ÿ  â &r:Q6   :   â >rOÿ8   ÿ  â 0r6ÿB    A   Ä &r8V;   7   ä&rBN;   :   ä 0r7ÿ     A   ä 0rOÿO    A   â  x66   ?  A  â 0r:ÿ9    A   â  x77   ?  A  â &r8S5   8   Ä &rBM5   B   â  x5O   ?  A  â Ere e       â  r A6     A  â  r97:     A  â  r;:5     A  â ErB B       â &r:0   ÿ   â#r  e   ÿ   â &r0W0   ÿ   ä &r:4   :   â #rB;B       â &r0a4   0   Ä &r U2   ÿ   ä &r;/   :   ä&r/`/   0   ä &r T3       ä &r;@   ;   ä&r@R@   /   Ä &r2c2   ÿ   ä &r/V?       ä 0r1ÿ1    A   â Er8 8       â &r2Q3   2   â Er; ;       â &r/S<   /   â  r >1     A  â 0r0ÿ=    A   â #r998      â &r2N?   2   â  r61     A  â Er/ /       â #r  ;   ÿ   â  r170     A  â Er3 @       â &r2M<   2   Ä #r/1/       â &r C   ÿ   â #r3   ÿ   â Er2 2       â  r350     A  â &r D       â !r/       â &r/UG   ÿ   â #r32      â &r E       ä &r1TH   /   â !r       â &rI       Ä 0r/ÿF    A   ä &rCWC   ÿ   ä &r0VJ   1   â Er1        â 0r2ÿK    A   â  r >/     A  â &rCaD   C   ä &rGcG   ÿ   â  r6/     A  â #r  1   ÿ   â  r/72     A  â r152     A  â &rC`E   C   Ä &rGQH   G   ä &r2   ÿ   ä&rW   ÿ   ä &rCRI   C   ä &rGNJ   G   ä &r   2   Ä &rUUX   ÿ   ä&ra      ä &rXcX   ÿ   â ErC C       â &r0SL   0   ä&rGML   G   ä &r\      Ä &r`\      ä &rUT[   U   ä&rQQ[   X   â x@   ÿàÿ â &r^      â Er0 0       â &rR^      â ErG G       â &rVVZ   U   â#rC   ÿ   â &rNNZ   Q   Ä 0rÿ    A   â |    pbô â &rVSY   V   â#r//0       â &rNMY   N   â #r11G      â Er        â 0rÿ]    A   â Er        â  r >     A  â r6     A  â ErV V       â  r77     A  â ErN N       â #r     ÿ   â #r   ÿ   â  r55     A  â x 	  ÿà÷ â #r 7V       Ä #r55N      â !r9       â !rB       â !r/       â !r.1.       â !r- -       â !r$5$       â x€   ÿàÿ ä x''€   ÿàÿ ä x))€   ÿàÿ Ä x++€   ÿàÿ ä r%ÿ%   ÿäÿ â D)       À â Gy  åÿÿÿÿƒ ê Ay        € ê�Ey  °     € â G‰  �     € ê $x   Ž Ê ˆs  üÿ H   è ˆs €üÿ H   è ˆs  ýÿ H   è ˆs €ýÿ H   è ˆs  þÿ H   è ˆs .€þÿ H   è ˆs - ÿÿ H   è ˆs $€ÿÿ H   ä Ay        € ê {          ì M        € ê „y €  H   b z ! r  p`ð â $z x  !Ž â x!!   ÿàÿ â „y     H   ¢ r ÿ   p"p Ä z ! r  p`ò â „y €  H   æ x    p"ò â „y €   H   ( „y     H   h „y €	  H   ( „y    H   ( „y €  H   ( „y	    H   ( „y    H   ( „y €  H   " !r       Æ/ „y    H   b !r       ÆO „y    H   ¢ !r       Æ� „y €  H   â !r       Æ„y    H   " !r         Æ„y
 €  H   b !r       Æ „y%    H   b !r       Æ „y  	  H   b !r       Æ „y  
  H   b !r  	       Æ „y# €
  H   b !r       Æ „y'    H   " !r.       Æ „y €  H   â !r-       Æ/ „y €  H   â !r       ÆO „y €  H   â !r       â� !r       ä‰       b z  r      â !r

       ä‰	      ¤ $x   ÿŽ â !r%       â ‰      â !r       â !r       È ‰      " !r

#       Æ ‰      b !r'       Æ ‰
      " !r        â/ !r$       ä ‰      b !r		       âO !r       ä ‰       ¢ !r       â� !r       Ä ‰	      â !r       Æ‰      " !r       Æ‰      b !r
       Æ ‰      " !r       Æ/ ‰
      b !r         ÆO ‰      ¢ !r	       Æ� ‰      è ‰  €    â !r       â!r       ä‰	 €    " !r       Æ ‰ €    " !r
       Æ/ ‰ €    b !r       ÆO ‰ €    ¢ !r       Æ� ‰ €    â !r        Æ ‰ €    b $z  ~  Ž Æ ‰      b !r		       â
  r  ÿàõ ä ‰ @    " !r       â r    ÿ† â !r
       â/ ‰	 @    " !r       ÆO ‰ @    b !r       â� $ÿÿ   ÿ ä ‰
 @    ¢ 
 j  ÿ„ â!r       ä‰ @    â 
 k   â !r       Ä ‰ @    b !r       Æ‰ @    " !r		       Æ ‰ €    ( ‰      " !r       â/    ÿàÿ â !r
       âO ‰	      b rÿÿ   ÿäÿ â !r       Ä� ‰      ¢ z j  ÿ† â!r       ä‰      â     ÿàõ â !r       ä‰      " z k  � â $ÿÿ   ÿ â !r       â ‰      "     ÿà÷ â !r       Ä ‰ @    b $tÿ   ÿ Ž â  j  ÿ„ ä ‰      b ÿÿ   ÿäÿ â !r	       â/ 

 j  ÿ† â %v  j   Ž â 	 k   â !r       âO 
 k  � â !r       â� †	     â !r       Æ†    â !r       Æ †	     â !r       Æ†     â !r       Æ †	     è †    è ‰      ( †	 
    â M™        € ê x   ÿàÿ â/ !r       Æ r    ÿàñ È r ÿÿ   ÿä  ä z j  ÿ€ È z k     Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      "   " ¹z  l      â Ey  ð    € â ™x?    â y      !   b ¹z  F   
   â $rÿÿ   ÿ Ž â r( ÿ       â y      %   ¢ $r ÿÿ   ÿ Ž â x      ÿ  â $r&ÿÿ   ÿ Ž â r ÿ   pRð Ä r ÿ   pRò â$x    Ž ä/ $x   ÿ Ž ÆO r    ÿ� È xÿ      â $À   Ž Æ |    pbô Ú G)  À    € ê y      '   " $rÿÿ   ÿ Ž â x þÿÿÿÿÀŽ â $rÿÿ   ÿ Ž â y      &   b $z	 p  Ž â z!  Y      â $x   
Ž ä $x'   ÿ Ž ä $x   ÿ Ž âx   ÿ   â $v)ÿ X  ÿ Ž â x%ÿ     â $rÿÿ   ÿ Ž â x$   ÿàÿ Ä xÿ     ä z#' q  ÿàÿ â 'z  y   Ž È 'z s   Ž â/ z ÿ z     Æ $z }  ÿŽ â zÿ t    â $z	  |  	Ž ä$z  |  Ž â x    ÿàÿ â $z	 v  	Ž ä$z v  Ž â xÿ   $  â%x	$   ÿŽ â x$ÿ   $  Ä xÿ     â $z w  ÿŽ â r   ÿà÷ â$z  p  Ž â r   ÿàù â $v ÿ q  ÿ Ž â r   ÿàõ â %x
$   ÿŽ â x"ÿ   $  Æ %x$   Ž È $x	   Ž ä$x	    ä%x$   Ž È %x   Ž È %x
   
Ž È $x    'Ž ä $rCÿÿ   ÿ ä r)   ÿàõ Ê $x	   ! Ê �y    ™ ¨  �y5   ™ è  �y.   ™ ( �y1   ™ b r
)   ÿàõ Ê $x   ! Ê �y6   ™ è �y4   ™ ( �y3   ™ " r'   ÿàõ ä r*)   ÿà÷ â �y0   ™ $ $r7ÿÿ   C â r+!   ÿäÿ Ä r;   ÿàõ â $tHÿ$   ÿ Ž ä �y-*    ‘ $ $x%   7 ä %v:; Z  H Ž È $x$   ÿŽ È $x;;   Ž â r)   ÿàù È �y/:   ™ b r)   ÿàõ Æ �y2:   ™ b r$   ÿà÷ â/ $x!    ä �y*:   ™ b r"7   ÿäÿ â $x   ! â r?#   ÿàõ â �y,:    • b r)   ÿà÷ â %v Z  H Ž Ä �y    ‘ d $x9$   ÿŽ ä $r@ÿÿ   C â r=?   ÿàõ â �y7    ‘ b $x   9Ž ä $x   !Ž â r>%@   ÿä â �y+:   ™ h �y8    ‘ b %v= Z  H Ž Æ �y9   ™ h �y:   ™ b! $x>$   ÿŽ Æ �y;   ™ b $x   Ž â r$?   ÿàõ ä �y    ™ f $x>"   @ â rI   ÿàõ â �y<    • b %v Z  H Ž Æ �y=   ™ b $rKÿÿ   C ä $xA>$   ÿŽ â �y@   ™ b rI   ÿàõ ä $x   AŽ â �y?   ™ b rB%K   ÿä ä rI$I   ÿàõ â %v Z  H Ž â �yE   ™ f $xGB$   ÿŽ â �y>   ™ b $xK"   K Ä $x   GŽ â �yA    • b $xKK$   ÿŽ Æ �yF   ™ h �yD   ™ b %vI Z  H Ž Æ/ �yJ   ™ b $x   KŽ Æ �yB    ™ h �yI   ™ h �yG    • h �yK   ™ h �yH   ™ h �y   ™ h �yL    • h �yO   ™ h �yM   ™ h �y    • h �yN    ™ b ‚x       Ä $~Rÿ   ÿ Ž â ‚x  ÿþý     ä |T       â $~ÿ   ÿ Ž ä/ $~Vÿ   ÿ Ž â |W       â $~Uÿ   ÿ Ž ä $~Yÿ   ÿ Ž â xP ˆˆˆˆÿÀŽ ÄO xQ ww  ÿÀŽ ä xÿ      ä xPP22ÿø� ä t RQ    â$~Rÿ   ÿ Ž â xww  ÿÀŽ ä tQTüúøôQ    Ä xPTv  ÿÀŽ ä t    ätRüúøô    ä r     Q    ä xPÿ   P  ä x5ww  ÿÀŽ Ä� x5ˆˆˆˆÿÀŽ â $~Rÿ   ÿ Ž â x5ÿ   5  ä rP       ä x22ÿø� ä xQ5ww  ÿÀŽ ä tPR    Ä tTüúøô    ä x5Tv  ÿÀŽ ä tRRQ    äx  d      ä tQTüúøôQ    ä xÿ     Ä x 1u      ä r P5       â $~Pÿ   ÿ Ž â x.ww  ÿÀŽ ärQR   Q    ä x.ˆˆˆˆÿÀŽ ä x.ÿ   .  Ä tPP    ätTüúøô    ä x22ÿø� ä |T       ä xS.ww  ÿÀŽ ä x5  d  Q    Ä x. 1u  Q    ä xRTv  ÿÀŽ ä tQTS    ätTVüúøôS    ä xÿ     ä r PR       Ä rQQ   T    â $~Rÿ   ÿ Ž â x1ww  ÿÀŽ äx1ˆˆˆˆÿÀŽ â $~Pÿ   ÿ Ž â x1ÿ   1  ä x22ÿø� ä xT1ww  ÿÀŽ Ä tRR    ätPüúøô    ä xSTv  ÿÀŽ ä xP  d  Q    äx1 1u  Q    ä t UT    Ä tQWüúøôT    ä xÿ     ä rRS       â $~Tÿ   ÿ Ž â rR    Q    ä x6ˆˆˆˆÿÀŽ äxS6ww  ÿÀŽ Ä xQÿ   6  ä x622ÿø� ä t TS    äxTQww  ÿÀŽ ä tSVüúøôS    ä xQ6Tv  ÿÀŽ Ä x d  R    äx1u  R    ä tRUT    ätUWüúøôT    â $~Tÿ   ÿ Ž â r  Q   S    ä x6ÿ   6  Ä xS4ww  ÿÀŽ äxQ4ˆˆˆˆÿÀŽ ä x4ÿ   4  ä rUR6   U    â $~Wÿ   ÿ Ž â t6TS    ätSVüúøôS    Ä xRQ22ÿø� ä xV4ww  ÿÀŽ ä x4  d  U    äxQ 1u  U    ä xTRTv  ÿÀŽ ä tUWV    Ä tVYüúøôV    ä xRÿ   R  â $~Wÿ   ÿ Ž â r66T   S    ä rUUR   V    ä xR3ˆˆˆˆÿÀŽ ä |S       Ä x 3ww  ÿÀŽ ä x3ÿ   3  ä xRR22ÿø� ä tSS     ät Wüúøô     â $~Wÿ   ÿ Ž â xV3ww  ÿÀŽ Ä xTRTv  ÿÀŽ ä x36 d  U    äx661u  U    ä tUWV    ä rSST        ä tVYüúøôV    Ä x ÿ   R  ä xT0ww  ÿÀŽ ä r U    V    ä xV-   ÿÀŽ ä tRWT    ätUYüúøôT    Ä xTÿ   0  ä x00ˆˆˆˆÿÀŽ ä x V   pRô ä xTTww  ÿÀŽ ä xV022ÿø� ä r0-ÿ       Ä t-WT    ätXYüúøôT    ä xT0ÿ   ÿÀŽ â &r0/   ÿ   â xWVTv  ÿÀŽ ä xVÿ   V  ä >rTÿT   ÿ  â &r02   0   â rURW   U    Ä rV-V   X    â 0rTÿT    A   â xRS d       äxSS1u       â &r 5*   0   â x0U d  V    äx-U1u  V    â &rU4/   ÿ   ä 0r/ÿ,    A   â  x,T   ?  A  â xT   ÿÀŽ â &r2Q2   U   â xV7   ÿÀŽ Ä x T   pRô â &r*3*   2   â x V   pRö ä r2ÿ       â &rP:   ÿ   âxT8   ÿÀŽ â &r:R:   ÿ   â x22ÿ   ÿÀŽ ä r77ÿ     € â &r19      â>r2ÿ2   ÿ  â &r:S9   :   â x T   pRô Ä x77ÿ   ÿÀŽ â &r .+       â r88ÿ       â &r;      ä&rT0;   :   â >r:ÿ7   ÿ  â 0r9ÿ2    A   â ErU         â &r2      â x88ÿ   ÿÀŽ â 0r7ÿ<    A   â  r ,/     A  â  x9   ?  A  â 0r:ÿ:    A   â Er2 2       â &r+6+   *   â >r8ÿ8   ÿ  â &rT-   T   â #r  U   ÿ   â  r97     A  â &r=   ÿ   â  x*:   ?  A  â 0r8ÿ8    A   â #r292       â Er+ +       â &r:@      â  r /*     A  â &r=4=   ÿ   â !r2       â &r5?   :   â #r  +   ÿ   â  x28   ?  A  â &r@Q@   =   Ä &r+PE   ÿ   â ErT T       â &r.>      â  r772     A  â &r?3?   @   ä 0rAÿA    A   ä &r+1F   +   â #r77T       â &r?6>   ?   â Er        â  r ,A     A  â &r+D   +   Ä &r8J   ÿ   ä&r44J   ÿ   â #r    ÿ   â Er? ?       â &r+B   +   â  r *A     A  â &r8I   8   ä 0rGÿG    A   ä &rPPK   ÿ   Ä &rERE   ÿ   ä&r4QI   4   ä &rKRK   ÿ   â #r? ?   ÿ   â &r85H   8   â Er+ +       â &rESF   E   â  r G     A  â &r43H   4   Ä &rP1   P   ä&rKS   K   ä 0rÿL    A   â x@   ÿàÿ â &r8.O   8   â #r +      â &rE0D   E   ä &r46O   4   Ä &rPM   P   ä&rK0M   K   â  r ,     A  â r*     A  â Er 8       â &rE-B   E   â |    pbô â !r((       â 0rÿ    A   â Er 4       â &rPN   P   Ä &rK-N   K   â #r     ÿ   â ErE E       â  r     A  â ErP P       â  r2G     A  â ErK K       â #r   ÿ   â  r2     A  â x)) 	  ÿà÷ â #r?E   ?   â #rP       â #rK      â !r7       â !r&?&       â !r         â !r       â x€   ÿàÿ â $r!ÿÿ   !Ž â x##€   ÿàÿ Ä x''€   ÿàÿ â D)       À â Gy  0éÿÿÿÿƒ ê Ay        € ê ˆ  ýÿ H   è ˆ €ýÿ H   è ˆ ( þÿ H   è ˆ &€þÿ H   è ˆ   ÿÿ H   è ˆ €ÿÿ H   è {          ì M        € ê „y      H   " z  r  p`ð Æ „y    H   b r ÿ   p"p Æ „y €   H   ¨ „y    H   è „y €  H   ( „y    H   h „y    H   h „y €  H   h „y    H   h „y €  H   h „y    H   b !r         Æ „y	 €  H   " !r         Æ/ „y    H   b !r       ÆO „y €  H   ¢ !r(       Æ� „y €  H   â !r&       Æ„y    H   â !r         Æ„y
 €  H   â !r        Æ „y €  H   â !r       â !r       ä ‰       " !r       Æ y      &   b !r       â !r	       â !r       È/ ‰      " !r       ÆO ‰      b !r       Æ� ‰	      ¢ !r        â!r       È ‰       â !r

       Æ ‰      " !r

       Ê ‰
      b !r       â !r       È/ ‰      " !r		       ÆO ‰      h ‰	      ¢ !r         â� !r       È‰  €    è ‰      " !r
       Ê‰
      b !r       â !r       È/ ‰ €    " !r	       ÆO ‰ €    h ‰	 €    ¢ !r        â� !r       È‰  @    è ‰ €    " !r

       Ê‰
 €    b !r       â !r       È/ ‰ @    " !r		       ÆO ‰ @    è ‰	 @    b !r         â� !r       äy      '   ¨ ‰ @    â !r
       âx
   ÿàÿ â $z x  Ž â ‰       " $vÿ r  ÿ Ž â z 
 r  p`ò â !r       â ‰ @    $ x    p"ò Ä ‰
      b !r	       Ê/ ‰      b $z ~  Ž âO !r       Æ 	   ÿàÿ â!r       â� 
 r  ÿàõ â ‰      ¢ r   ÿ† â !r        âÿÿ   ÿä â ‰      â 
 j  ÿ„ â !r       â $tÿ   ÿ Ž Ä 
 k   â $r ÿÿ   ÿŽ â ‰      " 		   ÿàõ â%v j   Ž â z j  ÿ† â !r
       ä$ÿÿ   ÿ â 	 j  ÿ„ â !r       â/ z k   � â †	     â 		 k   â !r       ÄO †    â !r       Æ� †	     è †     â !r       Æ †	     â M™        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      "   " ¹z  l      â Ey  €    € â ™x?    â y      !   b ¹z  F   
   â x      ÿ  â $r!ÿÿ   ÿ Ž â y      %   ¢ $r)ÿÿ   ÿ Ž â r ÿ   pRò Ä r ÿ   pRð â$x    Ž ä/ $x   ÿ Ž ÆO r    ÿ� È x&ÿ      â $€   Ž Æ | &   pbô Ú G)  p    € ê y      '   " $rÿÿ   ÿ Ž â x þÿÿÿÿÀŽ â $rÿÿ   ÿ Ž â y      &   b $x%&   ÿ Ž ä $x   
Ž ä $v'ÿ X  ÿ Ž â z% q  ÿàÿ â $x   ÿ Ž âx$ÿ     â $x   ÿ Ž Ä $v#ÿ Y  ÿ Ž â x"   ÿàÿ â $rÿÿ   ÿ Ž â xÿ     â 'z  y   Ž È 'z s   Ž â/ zÿ z     Æ $z  p  &Ž â zÿ t    â $z	 |  &Ž ä$z |   Ž ä $z w  ÿŽ â x   ÿàÿ â $z v  Ž ä$z  v  	Ž â xÿ   "  â$z }  ÿŽ â x"ÿ   "  â %x$   ÿŽ â xÿ     Ä r   ÿàõ â $z p   Ž â r   ÿà÷ är   ÿàù â %x
$   ÿŽ â x ÿ   "  Æ %x$   Ž È $x	   Ž ä$x	    ä%x$   Ž È %x   Ž È %x
   
Ž È $r?ÿÿ   ÿ ä r'   ÿàõ Ê $x	   # Ê �y7   ™ ¨  �y*   ™ è  �y(   ™ " r
'   ÿàõ Æ �y1   ™ d $x   # Ê �y/   ™ è �y+   ™ è �y4   ™ " r%   ÿàù är'   ÿàõ â �y2   ™ " rC   ÿà÷ â $r-ÿÿ   ? â r   ÿàù â$x#    â rC   ÿàõ â $t@ÿ$   ÿ Ž Ä $x$   - â �y0    ‘ " $rAÿÿ   ?Ž ä %v Z  @ Ž È/ $x$   ÿŽ ä $x $   A ä $x   Ž ä %v Z  @ Ž Æ �y.   ™ b $x3 $   ÿŽ â r"   ÿàõ ä �y    ™ b $x   3Ž â r'   ÿà÷ â $x    - â �y6   ™ h �y,   ™ b $x#   Ž Æ �y-   ™ b %v Z  @ Ž Æ �y3   ™ b $x$   ÿŽ Æ �y5   ™ h �y8    ‘ h �y:    • b $x   Ž â r'   ÿàõ Ä �y9   ™ f $x   # â �y<   ™ b r'   ÿàõ Æ/ �y=    • b rC"C   ÿà÷ Æ �y;   ™ b $x   # Æ �y    ‘ b $xA    AŽ Æ �y>   ™ b %vC Z  @ Ž Æ �y    ‘ b $xAA$   ÿŽ Æ �y    ™ b% $x   AŽ Æ �y    • h �yC   ™ h �yB   ™ h �yA   ™ h �y    ™ h �y@    • b ‚x       Ä ‚x  ÿþý     â $~Dÿ   ÿ Ž ä $~Fÿ   ÿ Ž ä $~ÿ   ÿ Ž ä/ $~Eÿ   ÿ Ž ä $~Gÿ   ÿ Ž ä $~Iÿ   ÿ Ž Ä $~Kÿ   ÿ Ž ä $~Jÿ   ÿ Ž ä $~Lÿ   ÿ Ž â x7ˆˆˆˆÿÀŽ äOx7ww  ÿÀŽ ä x7ÿ   7  ä x22ÿø� Ä tD    ätDFüúøô    ä x77ww  ÿÀŽ ä xTv  ÿÀŽ ä t7    ä r   D    â $~Dÿ   ÿ Ž â t7Füúøô7    Ä xÿ     ä x*ww  ÿÀŽ ä�x*ˆˆˆˆÿÀŽ ä x*ÿ   *  ä r   7    ä tD    Ä x22ÿø� ä xD*ww  ÿÀŽ ä t7Füúøô    ä x d      äx*Tv  ÿÀŽ ä x1u      Ä tED    ätEGüúøôD    ä xÿ     ä r*   7    â $~7ÿ   ÿ Ž â x*(ww  ÿÀŽ ärE   E    Ä x(ˆˆˆˆÿÀŽ ä t7*    ätDGüúøô*    ä x*22ÿø� ä x(ÿ   (  â $~Gÿ   ÿ Ž â x7*Tv  ÿÀŽ Ä xF(ww  ÿÀŽ ä x( d  E    äx1u  E    ä r7   D    ä tGF    ätEIüúøôF    Ä xDÿ   *  â $~*ÿ   ÿ Ž â x71ww  ÿÀŽ ärDD   E    â $~Fÿ   ÿ Ž â x1ˆˆˆˆÿÀŽ ä xEÿ   1  ä x122ÿø� Ä t**7    ät7Füúøô7    ä xFEww  ÿÀŽ ä xE1Tv  ÿÀŽ ä x d  D    ä r**E   7    Ä x1u  D    ä xE/ˆˆˆˆÿÀŽ ä tDGF    ätGIüúøôF    ä x7ÿ   1  â $~1ÿ   ÿ Ž â xF/ww  ÿÀŽ Ä xEE22ÿø� ä x/ÿ   /  ä r7D7   G    ä tDIüúøôF    â$~Iÿ   ÿ Ž â t11F    ä xGETv  ÿÀŽ Ä xH/ww  ÿÀŽ ä x/* d  7    äx**1u  7    ä r71G   D    ä tFIH    ä xG+ˆˆˆˆÿÀŽ Ä tIKüúøôH    â $~Hÿ   ÿ Ž â xEÿ   E  â $~Dÿ   ÿ Ž â x1+ww  ÿÀŽ ä xGG22ÿø� ä rFFE   I    ä x+ÿ   +  Ä tEHüúøô1    ätDD1    ä xHGTv  ÿÀŽ ä xI+ww  ÿÀŽ ä rDDH   E    ä x17 d  F    Ä x+71u  F    ä xE4ˆˆˆˆÿÀŽ ä t7JI    ätFLüúøôI    â $~Iÿ   ÿ Ž â xGÿ   G  ä xH4ww  ÿÀŽ Ä xEE22ÿø� ä x4ÿ   4  ä rG7G   F    ä t7IH    ätFKüúøôH    â $~Hÿ   ÿ Ž â xIETv  ÿÀŽ Ä xK4ww  ÿÀŽ ä x4D d  G    äxDD1u  G    ä r77I   F    â $~Fÿ   ÿ Ž â tGJK    ätHHüúøôK    Ä xEÿ   E  ä xI2ww  ÿÀŽ äxJÿ   2  ä x22ˆˆˆˆÿÀŽ ä rHGE   H    ä tFFI    Ä tELüúøôI    â $~Gÿ   ÿ Ž â xI0   ÿÀŽ â $~Kÿ   ÿ Ž â x2222ÿø� ä xJJww  ÿÀŽ ä x I   pRô ä xI2Tv  ÿÀŽ Ä tGGJ    ätJKüúøôJ    ä xKÿ   2  ä rIFI   E    ä rF0ÿ       â &rE.   ÿ   ârJGK   J    â &rG1.   ÿ   â xFFÿ   ÿÀŽ â &rE    E   Ä &rG+    G   â >rFÿF   ÿ  â &r ,   ÿ   ä &rE(6   E   ä &r -       ä 0rFÿF    A   ä &rE3   E   Ä &r((5       â x 8   ÿÀŽ â 0r:ÿ:    A   â  xFF   ?  A  â Er E       ä x     pRô â  r F:     A  â x27 d  H    â&r,1,   ÿ   â x771u  H    â #r    ÿ   â &r(9   (   Ä &r <   ÿ   ä0r=ÿ=    A   ä &r<2<   ÿ   â x0I d  J    â &r,+-   ,   â Er+ (       â &r;       âr88ÿ       â &r;7;   <   â  r F=     A  â x(   ÿÀŽ â &r/>      âx.I1u  J    â &r;0>   ;   â x88ÿ   ÿÀŽ â #r +   ÿ   â x (   pRô â &r *      âx+   ÿÀŽ â &r.   ;   â >r8ÿ8   ÿ  Ä rÿ       ä x +   pRô â 0r8ÿ8    A   â x(ÿ   ÿÀŽ ä rÿ       â 0rÿ    A   â >r(ÿ(   ÿ  â &rC   ÿ   âxÿ   ÿÀŽ â &r22C   ÿ   â  x8   ?  A  â &rG46   G   âEr         â &r,45   ,   â >rÿ   ÿ  â &rB      â r      A  â &r27B   2   Ä 0r(ÿ(    A   â x&&@   ÿàÿ â &r3D3   G   â #r        â &r9D9   ,   ä &r/A      ä&r20A   2   â  x((   ?  A  â 0rÿ    A   â | &   pbô â &r*      â !r) )       â 0rÿ@    A   â Er3 3       â &r2.   2   â Er9 9       â  r :(     A  â  r(=     A  â  x   ?  A  â  r     A  â Er        â #r  3   ÿ   â Er        â #r9   ÿ   â Er2 2       â  r     A  â  r     A  â x'' 	  ÿà÷ â #r      â #r       â #r2      â x€   ÿàÿ â !r!!       â !r       â !r       â x%%€   ÿàÿ â $r#ÿÿ   #Ž â D)       À â Gy  `ìÿÿÿÿƒ ê Ay        € ê ˆ ) þÿ H   è ˆ €þÿ H   è ˆ ! ÿÿ H   è ˆ €ÿÿ H   è {          ì M	        € ê „y      H   " z  r  p`ð Æ „y €   H   b r ÿ   p"p Æ „y    H   ¨ „y    H   è „y €  H   ( „y    H   h „y    H   h „y	 €  H   h „y    H   h „y €  H   h „y €  H   b !r  )       Æ „y
 €  H   " !r       â/ !r!       äO y      &   b !r         â� !r       â!r       ä!r         â !r		       È ‰       ¢ !r       Æ ‰	      â !r       Æ ‰      " !r       È !r

       Ê ‰
      " !r        âO !r	       È� ‰      ¢ !r       Æ‰      è ‰      " !r
       Ê ‰      " !r       âO !r       È� ‰ €    ¢ !r	       Æ‰ €    è ‰ 	 €    " !r       Ê ‰ €    " !r       ÆO y      '   ¢ !r       â� !r	        È‰ @    è ‰  @    ( ‰ @    b !r       Ê ‰
 @    " !r       â� x   ÿàÿ â $z x  Ž â/ !r         äz  r  p`ò â ‰      b $z ~  Ž âO !r	       âx    p"ò â ‰       ¢ $tÿ   ÿ Ž â 
 r  ÿàõ Ä ‰	      â !r

       ä $ÿÿ   ÿ â 
 j  ÿ„ â !r       Æ/ 
 k   ä ‰
      b  !r        âO %v j   Ž â !r		       È� †	     è †    è †	 	    â M™        € ê ¹z  r      â !r
       â/ �x   ?àÿ Ì |   ÿàñ Ê $r ÿÿ   ÿ  â z j  ÿ€ È z k     Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y
      &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž ä $rÿÿ   
 Ž È % \   Ž Ê �	    ™ b ¹z  l      â Ey  €    € â ™x?    â y      "   b $rÿÿ   ÿ Ž Æ y      !   ¢ $ÿÿ   ÿ Ž ä $ÿÿ    Ž È '
  m   Ž È $‚ÿÿ   ÿ Ž â 
 ÿ n     Ê $ÿÿ    
Ž â r ÿ   pRò Æ/ $
 o  Ž â r ÿ   pRô â$x    Ž äO $‚ÿÿ    Ž Æ r    ÿ� â 'Š s   Ž Æ xÿ      È |    pbö ä Šÿ t    Ö G9       € ê y      '   " $rÿÿ   ÿ Ž â x þÿÿÿÿÀŽ â $z w  ÿŽ â y      %   f $x   
Ž ä 'z y   Ž È $z }  ÿŽ â zÿ z    â $x   ÿ Ž È $z	 |  Ž â xÿ     Æ $z v  	Ž ä$x	   ÿ Ž ä $z p  Ž Æ/ x	   ÿàÿ â %x$   ÿŽ Æ xÿ     ä z X  à‰ â%x	   Ž â x	ÿ     äz X  àë â $x   ÿ Ž â z  X  ÿà÷ ä z	 Y  	Äþ Ä r   ÿàý â $rÿÿ   ÿ Ž â x    ÿàû ä z Y  ~ â $r$ÿÿ   ÿ â zÿ Y  dþ ä xÿ     Æ �y#üÿÿ ™ ¢  r   ÿàñ Æ �y"    ™ â  $tÿ$   ÿ Ž ä $rÿÿ   $  â r   ÿàñ â�y   ™ & r    ÿ  â %v Z   Ž â �y    ‘ f $x $   ÿŽ â r   ÿàñ â �y    ™ ä  $x   Ž â r   ÿ  È �y   ™ â  %v Z   Ž Æ �y   ™ " $x$   ÿŽ Æ �y    ‘ " $x   Ž Æ �y   ™ ( �y   ™ ( �y   ™ ( �y   ™ ( �y   ™ h �y     • h �y    ™ h �y!    • b ‚x       Ä ‚x  ÿþý     â $~(ÿ   ÿ Ž ä $~*ÿ   ÿ Ž ä $~'ÿ   ÿ Ž ä $~)ÿ   ÿ Ž ä $~,ÿ   ÿ Ž ä $~+ÿ   ÿ Ž â x@   ÿàÿ Ä x 	  ÿàû ä x 	  ÿàù ä/ x€   ÿàÿ â $rÿÿ   Ž ä $rÿÿ    â x&ÿ   #  äO x%#ww  ÿÀŽ Ä x##ˆˆˆˆÿÀŽ ä x&&ww  ÿÀŽ ä x##22ÿø� ä t(%    äx#Tv  ÿÀŽ ä t%*üúøô%    Ä t'&    ät)üúøô&    ä x#ÿ   #  ä r   %    ä r#       ä xÿ   "  Ä� x d      äx1u      â $~#ÿ   ÿ Ž â x"ww  ÿÀŽ â$~%ÿ   ÿ Ž â x""ˆˆˆˆÿÀŽ ä xww  ÿÀŽ ä x""22ÿø� Ä t##    ät&%üúøô    ä t(    ät*üúøô    â $~*ÿ   ÿ Ž â x%"Tv  ÿÀŽ ä x(ÿ     Äx'ww  ÿÀŽ â$~)ÿ   ÿ Ž â xˆˆˆˆÿÀŽ ä r##%   &    ä x((ww  ÿÀŽ ä t%*'    ät&,üúøô'    Ä x'22ÿø� ä t)(    ät)+üúøô(    ä x"ÿ   "  ä x('Tv  ÿÀŽ ä x'ÿ   '  Ä r"       â &r   ÿ   â r%%(   &    ä r'   )    ä x# d      äx#1u      â &r      â x"   ÿÀŽ â$~#ÿ   ÿ Ž â x% d      Ä x%1u      â $~%ÿ   ÿ Ž â x ww  ÿÀŽ ä x "   pRð ä x ˆˆˆˆÿÀŽ ä x"ÿ      ä t #    Ä t#%üúøô    â $~%ÿ   ÿ Ž â x&   ÿÀŽ â $~'ÿ   ÿ Ž â x22ÿø� ä x""ww  ÿÀŽ ä x &   pRö ä rÿ        â &r   ÿ   â xTv  ÿÀŽ Ä t%%"    ät"'üúøô"    ä xÿ     ä rÿ     € ä xÿ   ÿÀŽ â &r      â r     #    Ä r%   "    ä xÿ   ÿÀŽ ä >rÿ   ÿ  â &r      â x  d      ä>rÿ   ÿ  â 0rÿ    A   â x 1u      â &r      Ä &r      â  x    ?  A  â 0rÿ    A   ä 0rÿ     A   â |    pbð â &r      â Er        â 0r!ÿ!    A   â  x   ?  A  â  r       A  â Er        Ä  r!!     A  â #r     ÿ   â x 	  ÿà÷ Æ #r !       ä $r	ÿÿ   	Ž ä !r        â G‰  àöÿÿÿÿƒ ê Ay        € ê ˆ €ÿÿ H   è {          ì M)        € ê „y      H   " r ÿ   pRð Æ „y €   H   h „y    H   ¢ !r         È !r         È/ !r         ÊO ‰       $ !r        Ê ‰      $ !r       Ê ‰ €    $ !r       Ê ‰ @    $ !r       Ê ‰      b  M	        € ê y      %   ¨ y      '   â $z

 x  Ž äO $tÿ   ÿ Ž ä $z ~  
Ž â� !r       Æ/ %v j   Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž È % \   Ž Ê �	    ™ b $rÿÿ   ÿ Ž â Ey      € â $t
ÿ  €?ÿ Ž â y      "   b 'Š  s   Ž Æ y	      !   b '
 m   Ž È $rÿÿ   ÿ Ž â 
ÿ n    â $rÿÿ   ÿ Ž ä $tÿ  €?ÿ Ž â rÿ	   ÿü‚ â/ $x    	Ž â Šÿ t     È r    ÿ� Î G       € ê z ÿ ^  pPô ä z ÿ `  pPò ä z ÿ _   Sô ä z ÿ b  p ö ä z ÿ a  Sò ä z ÿ `  pPú Ä z ÿ c  0'ò ä z ÿ d  pPö ä z ÿ a  PSú â )      '   " z ÿ e  0Sö ä "       â)      %   h ™
      '   ¢ $* x  ÿŽ Æ ™      %   ä 2ÿ        â $* ~  ÿŽ Ê "	   àé ä/(ÿ     â $š

 ~  ÿŽ ÆO "ÿÿ   Ä~ ä : d  ÿŒ äz ÿ f  p ø ä : e  ÿ ä "   ÿàý ä z ÿ g  @'ø Ä ’       â $"ÿÿ    â ’	
	   àë â�$t
ÿ  €?ÿ Ž â ˜ÿ     â $š x  ÿŽ Æ ’ÿÿ   Äþ â �9
     b ’		   ÿàû ä * ^  ÿŒ ä Âÿ        ä * _   â $’ÿÿ   Ž â Ê f  ÿŒ Ä š	 b  ÿ† â$rÿÿ   ÿ Ž â Ê g  ÿ â $tÿ  €?ÿ Ž â š		 c  � â $rÿÿ   ÿ Ž ä �)     h �É     h �™     d Ay        € ê ¹z  l      â xÿ      â ™x?    â $ÿÿ   
Ž â Ey  0    € â $rÿÿ   ÿ Ž Ä $
 o  Ž â|    pbò â $‚ÿÿ    Ž ä $rÿÿ   ÿ Ž Ô G  À    € ê y      '   " $rÿÿ   ÿ Ž â z ÿ `  pPð â $z w  ÿŽ â y      %   b $rÿÿ   ÿ Ž â z ÿ a   Sð â $rÿÿ   ÿ Ž ä $x   ÿ Ž ä 'z y   Ž È $z }  ÿŽ â zÿ z    â $tÿ   ÿ Ž Æ r   ÿàó â $z |  Ž È $z v  Ž âx þÿÿÿÿÀŽ â $r+ÿÿ   ÿŽ  ä $z p  Ž ä/ $x   
Ž ä %x$   ÿŽ È $x   ÿ Ž ä$rÿÿ    Ž ä $x   ÿ Ž â x   ÿàÿ â%v X  Ž Æ x ÿ     äz X  ÿàõ ä z  X  ÿà÷ ä xÿ     ä xÿ     ä z Y  ÿä Ä z   Y  ÿ� ä r   ÿàó Ê $x   Ž  Ê �y)   ™ ¨  �y'   ™ è  �y    ™ " r   ÿàó ä r"   ÿàõ â �y&   ™ b $t!ÿ$   ÿ Ž Ä $rÿÿ   +Ž  â r   ÿàó â$x#    â r   ÿàõ ä r	   ÿ�  ä r   ÿàó â�y("    ‘ â %v Z  ! Ž ä r   ÿ�  Ä $x	$   ÿŽ ä $x	     ä %v Z  ! Ž Æ �y    ‘ " $x   Ž ä $x$   ÿŽ Æ �y   ™ " $x   Ž Æ �y   ™ ( �y$   ™ ( �y   ™ ( �y"   ™ h# �y!   ™ h �y#   ™ h �y*    • h �y%    ™ h �y	    • b ‚x       Ä ‚x  ÿþý     â $~/ÿ   ÿ Ž ä $~1ÿ   ÿ Ž ä $~ÿ   ÿ Ž ä/ $~.ÿ   ÿ Ž ä $~0ÿ   ÿ Ž ä $~2ÿ   ÿ Ž Ä $~4ÿ   ÿ Ž â x@   ÿàÿ ä x-ÿ   )  äO x,)ww  ÿÀŽ äx))ˆˆˆˆÿÀŽ ä x--ww  ÿÀŽ ä x))22ÿø� Ä t/,    ät1üúøô,    ä t,üúøô-    äx)Tv  ÿÀŽ ä t.-    ä x)ÿ   )  Ä r       ä r)   ,    ä xÿ   '  â� $~-ÿ   ÿ Ž â x'ww  ÿÀŽ äx''ˆˆˆˆÿÀŽ ä x)ww  ÿÀŽ â $~/ÿ   ÿ Ž â x''22ÿø� Ä t,-    ät-.)    ät.0üúøô)    ä x)'Tv  ÿÀŽ ä t/üúøô    ä x'ÿ   '  Ä x d      äx1u      ä r,,)       ä r-'   .    â $~.ÿ   ÿ Ž â xÿ      äx' ww  ÿÀŽ Ä x  ˆˆˆˆÿÀŽ ä x)ww  ÿÀŽ ä x  22ÿø� ä t.'    ät.0üúøô'    ä t-2)    Ä t04üúøô)    ä x)ÿ      ä x' Tv  ÿÀŽ ä r0-)   0    ä r'   .    â $~'ÿ   ÿ Ž â x-&ˆˆˆˆÿÀŽ â$~)ÿ   ÿ Ž â x.&ww  ÿÀŽ Ä x&ÿ   &  ä x--22ÿø� ä t''.    ät.)üúøô.    ä x/&ww  ÿÀŽ ä x)(   ÿÀŽ Ä x&-Tv  ÿÀŽ ä x , d      äx,,1u      ä x d  0    ä x )   pRò ä r&'&   .    Ä x1u  0    â $~0ÿ   ÿ Ž â x'   ÿÀŽ â $~2ÿ   ÿ Ž â r((ÿ     €  ä x '   pRò ä t)0/    â &r   ÿ   â t02üúøô/    Ä x-ÿ   -  ä x((ÿ   ÿÀŽ ä rÿ     €  â &r      â r))-   0    â &r$   ÿ   â >r(ÿ(   ÿ  ä xÿ   ÿÀŽ â &r        â x& d  )    â&r"      â >rÿ   ÿ  â 0r(ÿ(    A   â x&&1u  )    â &r ,!       Ä &r#      â  x(   ?  A  â 0r*ÿ*    A   ä 0rÿ    A   â Er         â &r&%      â  r *     A  â 0r&ÿ	    A   â  x)   ?  A  ä #r(    ÿ   â Er'        â  r))&     A  â G‰  P    € ê $tÿ   ÿ Ž È %v `  Ž Ê r   ÿàó Ê $x   Ž  Ê �y.    ¨  �y,    è  �y     " z `  à£ ä x   ÿú� â �y-    b z a  Dþ  Ê �y     â x0ÿ     äz `  à£ È z	 a  0Dþ  Ê �y     â $~0ÿ   ÿ Ž ä $~ÿ   ÿ Ž ä $~2ÿ   ÿ Ž ä $~4ÿ   ÿ Ž â x/ÿ   .  äO x.ww  ÿÀŽ Ä/ x..ˆˆˆˆÿÀŽ ä x//ww  ÿÀŽ ä x..22ÿø� ä t0    ä t0üúøô/    ä x.Tv  ÿÀŽ Ä t2üúøô    ä t	4/    ä x.ÿ   .  ä r       ä r		.   0    È x. d  	    äxÿ   ,  â� $~2ÿ   ÿ Ž â x1u  	    â &r.   ÿ   â x	,ww  ÿÀŽ â $~.ÿ   ÿ Ž â xww  ÿÀŽ â $~0ÿ   ÿ Ž â x ww  ÿÀŽ â$~/ÿ   ÿ Ž â t22	    Ä t..    ä t	0üúøô	    ät0üúøô    ä t04    ä x1ÿ      ä x4 ˆˆˆˆÿÀŽ Ä t/üúøô    â $~/ÿ   ÿ Ž â x3,ˆˆˆˆÿÀŽ ä x 1ww  ÿÀŽ â $~1ÿ   ÿ Ž â x4422ÿø� ä x3322ÿø� ä t//     Ä x,4Tv  ÿÀŽ ä t 1üúøô     ä x4ÿ   4  ä x13Tv  ÿÀŽ ä x3ÿ   3  ä r0,       Ä r4/4        â &r      â r	21   	    ä r..3       ä x   ÿÀŽ ä x d  4    â $~,ÿ   ÿ Ž â x	 d  .    Ä x 	1u  .    â $~.ÿ   ÿ Ž â x    pRò â &r$$   ÿ   â x-ww  ÿÀŽ äx-ˆˆˆˆÿÀŽ ä x-ÿ   -  ä t	,    Ä t,.üúøô    â $~.ÿ   ÿ Ž â x--ww  ÿÀŽ â $~0ÿ   ÿ Ž â x22ÿø� ä x/   ÿÀŽ ä t..-    ä xTv  ÿÀŽ Ä t-0üúøô-    ä xÿ     ä x /   pRô ä rÿ     €  ä x1u  4    ä r		   ,    Ä r..   -    ä rÿ       ä >rÿ   ÿ  â &r"   $   â x	 d  .    â &r      â >rÿ   ÿ  â 0rÿ    A   â x.	1u  .    â &r#      Ä &r  !      â  x   ?  A  â 0rÿ    A   ä &r..%      â Er         â  r *     A  â  x   ?  A  ä Er. .       â #r     ÿ   â  r&     A  È #r .       È !r        â |    pbò â #r')'   (   â x 	  ÿàõ ä x€   ÿàÿ â !r'       ä $rÿÿ    Ì D       À â Gy  Pïÿÿÿÿƒ ê Ay        € ê y       "   " Ey        € â r  ÿ   pRð är  ÿ   pRò Ö G‰  `     € ê y      !   " z ÿ `  pPð È z ÿ a   Sð â $x     Ž Ê ˆs €ÿÿ H   î ˆ    H   ä Ay        € ê {          ì M        € ê y      !   b z ÿ `  pPð È z ÿ a   Sð â „y     H   > „‰  €   H   h „‰    H   ¢!r       È !‚         È/ !‚        âO G‰        € ê „y  €  H   ( „y €   H   h „y    H   ¨ „y	    H   è „y €  H   " !r         â !r       Æ/ !r         âO !r	       Æ� !r        ä‰      " z ÿ `  pPð ä r ÿ   pRò ä z ÿ a   Sð â !r       Ê ‰       $ !r         Ê ‰  €    $ !r        Ê ‰ @    $ !r       Ê ‰      $ !r       â G‰        € ê ‰       $ !r         Ê ‰       $ !r        Ê ‰ €    $ !r       Ê ‰ @    $ !r       Ê ‰      $ !r       ä M        € ê #r
      â G‰  @    € ê $v ÿ h  ÿ Ž â #r      È x     pBð Ú G	  P    € ê x     pRð Ú G‰  �     € ê x     pRð Ú G	  ð    € ê  x ;ª¸¿  A  Ì s          $ !t   €?    Ì s        $  r     A  È  r     A  â Gy  P    € ê  x '7=  A  â x*BL?  A  Æ #t   €?    È  r      A  â  x    ?  A  Ê s     $   $ !t  €?    È  r       A  È  r      A  â Gy  °    € ê x     pRð Ú G‰  à     € ê x     pRð Ú G	        € ê 	z  i    � ä 	z i    � Æ  x ;ª¸¿  A  â 	z i €  � Ê s        $ !t  €?    Ì s        $  r       A  È  r      A  â Gy  Ð     € ê  r     A  â Gy  °     € ê 	x   à@  � ä 	x  à@  � Æ  x #ÛÙ?  A  â 	x  àÀ  � Æ  x;ª¸¿  A  ä !t  €?    È s        $ !t  €?    Ì s        $  r       A  È  r      A  ä y       &   ( y      %   ( y      '   b $z   x  Ž ä $tÿ   ÿ Ž Ä $z  ~   Ž È/ %v  j   Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž ä $rÿÿ    Ž È % \   Ž Ê �	    ™ b ¹z  l      â Ey  �!    € â ™x?    â y      "   b $rÿÿ   ÿ Ž â x      ÿ  â $r#ÿÿ   ÿ Ž â y      !   ¢ $ÿÿ   ÿ Ž ä $ÿÿ    Ž â y      %   & '
  m   Ž È $‚ÿÿ   ÿ Ž â 
 ÿ n     Ê $ ÿÿ    
Ž â r ÿ   pRò Æ/ $
  o  Ž â r ÿ   pRô â$x    Ž äO $‚ÿÿ    Ž Æ r    ÿ� â 'Š s   Ž Æ x$ÿ      â $€   Ž ä $x   ÿ Ž â | $   pbö ä Šÿ t    Ö G9  à    € ê y      '   " $rÿÿ   ÿ Ž â x þÿÿÿÿÀŽ â $z  p  $Ž ä $v'ÿ X  ÿ Ž ä $x   
Ž ä $v%ÿ Y  ÿ Ž ä $x   ÿ Ž Ä $rÿÿ   ÿ Ž ä $x$   ÿ Ž â x"   ÿàÿ È x!ÿ   "  äx"ÿ   "  â 'z y   Ž È $z }  ÿŽ â z	ÿ z    Ê $z		 |   Ž ä $z  w  ÿŽ ä $z v  	Ž ä$x	   ÿ Ž â r    ÿàù â %x$   ÿŽ È $r?ÿÿ   ÿ â r	   ÿàñ â%x   Ž â r!   ÿà÷ ä r 	   ÿ  är!!   ÿ� ä r'   ÿàñ Ê $x   %  Ê �y0   ™ ¨  �y3   ™ è  �y    ™ " r	   ÿàñ Æ �y5   ™ b $tCÿ$   ÿ Ž ä $rÿÿ   ?  â r
	   ÿàñ â %zC p  Ž Æ r   ÿ  â %v

 Z  C Ž â r"	   ÿàñ â �y6   ™ ä  $x$   ÿŽ â r"   ÿ  â �y1   ™ ä  $x   Ž ä %v Z  C Ž â �y/   ™ æ  $x	$   ÿŽ â �y,
   ™ â r'   ÿàñ Ä $x   	Ž â �y*
   ™ $ $x%      ä �y+
   ™ ( �y)   ™ ( �y-
   ™ ( �y(   ™ ( �y4    ‘ ( �y2   ™ h �y&   ™ h �y.    ™ b %zC p  Ž Æ �y8
    • h �y7   ™ b r'   ÿàñ â %zC p  Ž ä �y;   ™ d $x	%   !  Ä �y:    ‘ h �y9    ‘ h �y=    • h �y>   ™ b %zC p  Ž Æ �y<   ™ h �y@    ‘ b %zC p  Ž Ê �yA   ™ b %z
C p  Ž â/‚x       ä ‚x  ÿþý     â $~Bÿ   ÿ Ž ä %zC p  
Ž ä �y

    ‘ d $~Dÿ   ÿ Ž Ä �y    ‘ b %zC p  Ž È %zC p  Ž ä �y    ‘ d $~ÿ   ÿ Ž ä $~ÿ   ÿ Ž â �y    ‘ h �y	   ™ b $~Cÿ   ÿ Ž Ä $~Eÿ   ÿ Ž â xÿ   0  ÈO xww  ÿÀŽ È tB    ätDüúøô    ä �y   ™ ¢  x0ww  ÿÀŽ È t    ä/tüúøô    ä �y   ™ ¢  x00ˆˆˆˆÿÀŽ È x0022ÿø� È x0Tv  ÿÀŽ ä x0ÿ   0  ä r       ä xÿ   3  ä� x3ww  ÿÀŽ äx33ˆˆˆˆÿÀŽ Ä r0       â $~ÿ   ÿ Ž â xww  ÿÀŽ ä x3322ÿø� ä t    ä tC    ät0Eüúøô    Ä tüúøô    ä x3Tv  ÿÀŽ ä x3ÿ   3  ä r       ä xÿ      äx ww  ÿÀŽ Ä r3   0    â $~0ÿ   ÿ Ž â x  ˆˆˆˆÿÀŽ ä x d      â$~Bÿ   ÿ Ž â x1u      ä xww  ÿÀŽ ä x  22ÿø� Ä t0    ät    ät0üúøô    ä tBüúøô    ä x Tv  ÿÀŽ ä xÿ      Ä x  d      äx1u      ä r       ä xÿ   5  är0   0    â $~ÿ   ÿ Ž â x5ww  ÿÀŽ â$~Bÿ   ÿ Ž â x55ˆˆˆˆÿÀŽ â $~ÿ   ÿ Ž â xww  ÿÀŽ Ä x5522ÿø� ä tüúøô    ätB    ätBDüúøô    ä t    ä x5Tv  ÿÀŽ Ä x5ÿ   5  ä r       ä xÿ   6  ä x6ww  ÿÀŽ â$~Dÿ   ÿ Ž â x66ˆˆˆˆÿÀŽ â $~3ÿ   ÿ Ž â x d  0    Ä rB5   B    â $~5ÿ   ÿ Ž â x1u  0    â $~Fÿ   ÿ Ž â x0ww  ÿÀŽ â &r,   ÿ   â x6622ÿø� ä tD    â &r*      â t330    Ä tD5üúøô0    ä tFüúøô    ä x06Tv  ÿÀŽ â &r  +      â x6ÿ   6  â &r)   ÿ   â r0       ä x d  B    Ä x0ÿ   1  ä x1u  B    â $~Bÿ   ÿ Ž â x1ww  ÿÀŽ â&r-       â x11ˆˆˆˆÿÀŽ â &r(      â r336   D    â $~Dÿ   ÿ Ž â x60ww  ÿÀŽ Ä x 4   ÿÀŽ ä x1122ÿø� â &r&      â t0B    ätBC6    ä x     pRð ä t5Düúøô    Ä tCEüúøô6    ä x61Tv  ÿÀŽ ä xDÿ   1  ä x d  3    â &r.      â r44ÿ        â $~ÿ   ÿ Ž â x11u  3    â $~ÿ   ÿ Ž â r506   5    â &r,   ÿ   â rBBD   C    Ä x 7ww  ÿÀŽ ä x44ÿ   ÿÀŽ â &r1*      â x05 d  B    ä t     ät üúøô     ä xÿ   7  Ä >r4ÿ4   ÿ  ä x77ˆˆˆˆÿÀŽ â &r0+      â xww  ÿÀŽ â $~ÿ   ÿ Ž â x7722ÿø� â $~0ÿ   ÿ Ž â x6/ww  ÿÀŽ â $~3ÿ   ÿ Ž ä 0r4ÿ4    A   Ä $~Cÿ   ÿ Ž â t    â0r8ÿ8    A   â x7Tv  ÿÀŽ ä t00üúøô    â  x4   ?  A  â x51u  B    â $~Bÿ   ÿ Ž â t336    ä x5ÿ   /  Ä t6Cüúøô6    ä xC2ww  ÿÀŽ â $~Eÿ   ÿ Ž â r        â  r 8     A  â xD5ww  ÿÀŽ â $~Gÿ   ÿ Ž â x9   ÿÀŽ ä tBBC    Ä tCFüúøôC    ä xFÿ   2  ä t5ED    äx    pRð â $~Hÿ   ÿ Ž â xE2ˆˆˆˆÿÀŽ â $~Jÿ   ÿ Ž â tDGüúøôD    Ä x//ˆˆˆˆÿÀŽ ä xGFww  ÿÀŽ ä r99ÿ        ä xFE22ÿø� ä x2/22ÿø� ä t/HG    Ä x7ÿ   7  ä tHJüúøôG    ä Er        ä x99ÿ   ÿÀŽ ä xGFTv  ÿÀŽ ä x:   ÿÀŽ Ä xFÿ   F  â &r-      â r77   0    â $~ÿ   ÿ Ž â >r9ÿ9   ÿ  ä xE2Tv  ÿÀŽ ä x    pRö â #r     ÿ   â x2ÿ   2  Ä rBBG   C    ä r//F   H    ä x;ww  ÿÀŽ â $~ÿ   ÿ Ž â x d  7    äx1u  7    â 0r9ÿ9    A   â xÿ   ;  Ä r33E   6    ä r252   D    ä x1B d  /    äxB1u  /    â $~/ÿ   ÿ Ž â t0üúøô    â$~ÿ   ÿ Ž â xww  ÿÀŽ â 0r=ÿ=    A   â t    â  x9   ?  A  â x63 d  2    Ä x331u  2    ä x;;ˆˆˆˆÿÀŽ ä r::ÿ     € ä t    ät2/üúøô    ä Er        â  r=     A  â x;;22ÿø� â &r/,   ÿ   â x::ÿ   ÿÀŽ â #r       â &r66)   ÿ   â x;Tv  ÿÀŽ â &r*   /   â >r:ÿ:   ÿ  Ä xÿ   >  â &r33(   6   â x;ÿ   ;  â $~4ÿ   ÿ Ž â r   0    â $~ÿ   ÿ Ž â x >ww  ÿÀŽ â$~6ÿ   ÿ Ž â xww  ÿÀŽ â 0r:ÿ:    A   â x>>ˆˆˆˆÿÀŽ â $~ÿ   ÿ Ž â r2;   2    Ä t/4    ät0üúøô     ä x>>22ÿø� ä t46üúøô    â  x:   ?  A  â Er        ä x d  2    Ä t     â  r 8     A  â x>Tv  ÿÀŽ â &r+      â x21u  2    â $~ÿ   ÿ Ž â r   0    â #r    ÿ   â x>ÿ   >  â $~0ÿ   ÿ Ž â x<ww  ÿÀŽ â&r11&   3   â x<ˆˆˆˆÿÀŽ Ä x<ÿ   <  â &r.   1   â r4/>   4    â $~/ÿ   ÿ Ž â t    â$~1ÿ   ÿ Ž â t0üúøô    ä x22ÿø� ä x<<ww  ÿÀŽ Ä x0@   ÿÀŽ ä x  d  4    ä xTv  ÿÀŽ ä t//<    äx 0   pRð ä t<1üúøô<    Ä xÿ     â &r  )   ÿ   â x1u  4    ä r@@ÿ        ä r       ä r</   <    â &r (       â x@@ÿ   ÿÀŽ Ä x d  <    äx1u  <    ä xAˆˆˆˆÿÀŽ â$~/ÿ   ÿ Ž â >r@ÿ@   ÿ  â $~ÿ   ÿ Ž â xAww  ÿÀŽ ä xAÿ   A  Ä x22ÿø� â &r2-      â t0/üúøô    â0r@ÿ@    A   â t    â $~2ÿ   ÿ Ž â xAAww  ÿÀŽ â $~4ÿ   ÿ Ž â x/Tv  ÿÀŽ â &r&       â t12A    â  x@   ?  A  â r/   0    Ä t24üúøôA    ä x/ÿ     â  r 8     A  â Er        ä r21/   2    â &r.      â x0
   ÿÀŽ â #r    ÿ   â $~/ÿ   ÿ Ž â x d  2    Ä x1u  2    ä xww  ÿÀŽ äOxÿ     ä x 0   pRð â $~0ÿ   ÿ Ž â x ˆˆˆˆÿÀŽ â $~2ÿ   ÿ Ž â t/    Ä x/ww  ÿÀŽ â $~1ÿ   ÿ Ž â x  22ÿø� â &r,,   ÿ   â t0/    ät02üúøô/    ä x/   ÿÀŽ ä t1üúøô    Ä x Tv  ÿÀŽ ä x /   pRö ä r       â $~ÿ   ÿ Ž â x/ÿ     ä xww  ÿÀŽ â$~2ÿ   ÿ Ž â xˆˆˆˆÿÀŽ Ä x//ww  ÿÀŽ ä rÿ     € ä t    äx22ÿø� ä t1üúøô    ä t12/    Ä x ÿ      ä xÿ   ÿÀŽ ä t24üúøô/    ä x/Tv  ÿÀŽ ä xÿ     ä r     0    Ä >rÿ   ÿ  ä r/       ä r11   2    â &r*   ,   â x d       â0rÿ    A   â x* d  1    ä x 1u       â &r+      â x   ÿÀŽ â  x   ?  A  â x11u  1    â &r**)   ÿ   â x    pRö â &r -      â  r 8     A  â x	ˆˆˆˆÿÀŽ â &r*1(   *   â xÿ   	  â $~ÿ   ÿ Ž â x(   ÿÀŽ â $~ÿ   ÿ Ž â x	ww  ÿÀŽ Ä x22ÿø� ä xww  ÿÀŽ ä r

ÿ        ä x (   pRð ä t	    ä xTv  ÿÀŽ Ä t    ä tüúøô    ätüúøô    ä xÿ     ä rÿ     € ä rÿ        Ä r	       ä x

ÿ   ÿÀŽ ä r       ä xÿ   ÿÀŽ ä xÿ   ÿÀŽ ä >r
ÿ
   ÿ  Ä x	 d      ä>rÿ   ÿ  ä >rÿ   ÿ  â 0r
ÿ
    A   â x1u      â &r		&   *   â x$$@   ÿàÿ â 0rÿ    A   ä 0rÿ    A   â | $   pbð â &r	.   	   â Er        â  x

   ?  A  â  x   ?  A  â  x   ?  A  â Er        â  r
=
     A  â r=     A  âEr        â #r     ÿ   â Er	 	       â  r==     A  â x'' 	  ÿà÷ â #r

      â #r      â #r =	       â !r##       â !r
       â !r       â !r        â x€   ÿàÿ â $r%ÿÿ   %Ž â D	       À â Gy  àáÿÿÿÿƒ ê Ay        € ê ˆ # þÿ H   è ˆ €þÿ H   è ˆ  ÿÿ H   è ˆ €ÿÿ H   è {          ì M)        € ê „y      H   " z  r  p`ô â $z x  Ž Ä „y €   H   b r ÿ   pVt Æ „y    H   ¨ „y    H   è „y €  H   ( „y    H   („y    H   ( „y €  H   ( „y	 €  H   ( „y    H   ( „y €  H   " !r  #       Æ „y €  H   " !r       â/ !r       âO !r         ä� !r       â!r       â !r         â !r       È ‰       b !r		       â !r       È ‰	      ¢ !r       Æ ‰      â !r       Ê ‰
      " !r        â/ !r	       ÈO ‰      b !r       Æ� ‰      ¨ ‰      â !r

       â $tÿ   ÿ Ž È ‰
      " !r       â/ !r       ÈO ‰ €    b !r	       Æ� ‰ €    ¨ ‰ 	 €    â !r
       â x
   ÿàÿ È ‰ €    " z 
 r  p`ò È x    pVò  â !r       â/ x   ÿàÿ â !r       ÆO z  r  p`ö â !r	        â� ‰ @    d x    pVö ä ‰  @    ¢ !r       Æ ‰ @    ( ‰ @    è y      '   " !r       â/ x   ÿàÿ â !r         ÆO ‰	      b z  r  p`ð â !r       ä ‰       " x    pVp  â !r       ä� ‰      ¢ $z ~  Ž Æ‰      â !r		       â/ !r        â %v j   Ž â !r       ÈO †©     â !r       Æ� †¹ 	   è †‰    â M        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      &   " z ÿ \  pPô â ¹z  F   
   Æ z ÿ ]   Sô Ú $$ÿ   ÿ Ž È %& \   Ž Ê �)	    ™ b $rÿÿ   ÿ Ž â ¹z  l      â $rÿÿ   ÿ Ž â y)      %   b '*  m   Ž â™x?    â Ey  p    € â y7      "   ¢ $t+ÿ  €?ÿ Ž â * ÿ n     â 'ª s   Ž â y(      !   æ $"ÿÿ    
Ž â y      '   " $r,ÿÿ   ÿ Ž Ä $*  o  Ž ä$¢ ÿÿ    Ž ä $r.ÿÿ   ÿ Ž ä $z   w  ÿŽ ä $t-ÿ  €?ÿ Ž â x))   ÿ   ä/ r 7ÿ   pRð ÆO $x
)   (Ž â� x (   pDp  ä r77(   ÿ(Ž â $z }  ÿŽ âz 
 r  pdp  â 'z y   Ž â r
77   ÿ� ä r*    ÿàý ä ª	ÿ t    Î G	  Ð    € ê z ÿ d  pPð â $t+ÿ  €?ÿ Ž â z ÿ `  pPú ä z ÿ e   Sð ä z ÿ ^  pPò ä z ÿ f  p ø ä z ÿ a  PSú Ä z ÿ _  Sò ä z ÿ g  @'ø ä z ÿ b  p ú ä  	ÿ       är	       ä 
  d  ÿ† È 
  e  ÿ� â $  ~  ÿŽ â z ÿ `  pPö Æ �	+     b z ÿ a  0Sö È z ÿ c  P'ú ä  )   (à± â $ x  ÿŽ â  ÿ   )  ä Â	ÿ       ä  ÿ    ÿd~  Ê $Ú ~  ÿŽ ä $Ú x  ÿŽ â Øÿ   )  ä Ò)   (à± ä    ÿàõ ä Òÿ   ÿd~  ä ÿ    ÿä Ä Ò   ÿàñ ä  ^  ÿ„ Æ $Òÿÿ     â  _   ä Ê f  ÿ„ äÚ b  ÿ€ â$t-ÿ  €?ÿ Ž â Ê g  ÿ â $r.ÿÿ   ÿ Ž â Ú c    â $r,ÿÿ   ÿ Ž Ä �É-     h �.     h �Ù,     d Ay        € ê x6ÿ   
  â Ey  p>    € â r/ ÿ       â x0      ÿ  â | 6   pbð â x2      ÿ  â $r>ÿÿ   ÿ Ž â x4      ÿ  â $r[ÿÿ   ÿ Ò G	  à=    € ê x

þÿÿÿÿÀŽ â $z) p  6Ž âzÿ z    â $v9ÿ `  ÿ Ž â x)   ÿàÿ â $x77   

Ž â x)   ÿàÿ â $z |  Ž â z ÿ `  pPð â $x:7   ÿ Ž âx7   ÿ   â $z  |  6Ž â x)   ÿàÿ â %x:   ÿŽ âx:   ÿú� â x0      ÿ  â xÿ     â $z	 v  Ž âr/ ÿ       â $z 	 v   Ž â x2      ÿ  â %x$   Ž â z ÿ a   Sð â x4      ÿ  â x:ÿ   :  â $z p   Ž âx8   ÿàó â $z	 p   Ž â xÿ     â $z p   Ž Ä %x$   Ž È %x$   Ž È %x	$   Ž È %x	$   Ž È %x$   Ž È %x$   Ž È %x
$   Ž È %x	$   Ž È %x$   Ž È %x$   Ž È %x$   Ž È $v;ÿ a  ÿ Ž ä $v=ÿ X  ÿ Ž ä $v?ÿ Y  ÿ Ž ä $r>ÿÿ   ÿ Ž ä $xA6   ÿ Ž ä $r<ÿÿ   Ž  Ä r=   ÿàó Ê $x   ?Ž  Ê �y\   ™ ¨  �y    ™ è  �yI   ™ " r$A*   ÿàó â $tYÿ$   ÿ Ž Ä �yJ   ™ b rÿ[   ÿäÿ  ä r&7$   ÿàó är$:$   ÿàõ är7   ÿ�  är:   ÿ â %v&& Z  Y Ž â r"=8   ÿàó Æ $x$   ÿŽ ä %v$$ Z  Y Ž È $x$   ÿŽ ä $x''   Ž ä $x%%   Ž ä $x#?   <Ž  â �y@&   ™ h �yC$   ™ b %z"Y p  "Ž Æ �yE$   ™ b r=   ÿàó Æ �yB&   ™ h �yG$   ™ h �yR"ôÿÿ ™ b r?   ÿäÿ  Æ �yD&   ™ h �yF&   ™ h �yN    ‘ h �yH$    ™ h �yM"øÿÿ ™ b r =   ÿàó Æ �yV&    • d $x!   ?Ž  ä �yL"üÿÿ ™ h �yK     ‘ b %zY p  Ž Æ �yT"    ™ h �yS    ‘ b %zY p   Ž Æ �y$$    • h# �yU    ‘ b %zY p  "Ž Ê �yWôÿÿ ™ h �yOøÿÿ ™ h �yQüÿÿ ™ h �yP    ™ b %z"Y p  Ž È %zY p  Ž â/�yX"    ‘ h �yZôÿÿ ™ h �y%øÿÿ ™ h �yüÿÿ ™ b %z Y p  Ž Ì %zY p   Ž ä�y      ‘ d %z&Y p  "Ž ä �yY    ™ h �y    ‘ h �y&&    ‘ b ‚x       Ä ‚x  ÿþý     â $~^ÿ   ÿ Ž ä $~`ÿ   ÿ Ž ä $~"ÿ   ÿ Ž â/ |b       ä |'       ä x!ÿ   \  ÄO x\ww  ÿÀŽ äx\\ˆˆˆˆÿÀŽ ä x!!ww  ÿÀŽ ä x\\22ÿø� ä t^    ät`üúøô    Ä t"!    ät"büúøô!    ä x\Tv  ÿÀŽ ä x\ÿ   \  ä x! ww  ÿÀŽ ä� r       Ä r"\   "    â $~\ÿ   ÿ Ž â t^!    â $~^ÿ   ÿ Ž â xÿ   I  äxIww  ÿÀŽ äxIIˆˆˆˆÿÀŽ ä x#ww  ÿÀŽ Ä t\    ät^^üúøô    ä xÿ      ä xII22ÿø� ä t!`üúøô!    â $~`ÿ   ÿ Ž â x  ˆˆˆˆÿÀŽ Ä x]ITv  ÿÀŽ ä xww  ÿÀŽ ä x  22ÿø� ä t_`#    ät`büúøô#    ä xIÿ   I  Ä r]]   ^    ä t#\    äx Tv  ÿÀŽ ä t\büúøô    ä x ÿ      ä r`_I   `    Ä r   !    ä x d  "    ä r##    \    ä xÿ   J  â$~!ÿ   ÿ Ž â x1u  "    ä x JˆˆˆˆÿÀŽ Ä x"] d  `    äx`]1u  `    â $~]ÿ   ÿ Ž â xJJww  ÿÀŽ â $~_ÿ   ÿ Ž â |I       ä x\ww  ÿÀŽ ä x 22ÿø� Ä t!!J    ätJIüúøôJ    ä t]]\    äxITv  ÿÀŽ ä t\_üúøô\    ä xÿ     â &r @   ÿ   â r!!I   J    â &r""C   ÿ   â r\]   \    Ä x d  #    äx#1u  #    â &r"`E   "   â x! d  \    â &r B       â xÿ   R  Æ &rG   "   â xRww  ÿÀŽ â&r D       â x!1u  \    â $~"ÿ   ÿ Ž â |J       â $~\ÿ   ÿ Ž â x!ww  ÿÀŽ â $~^ÿ   ÿ Ž â xRRˆˆˆˆÿÀŽ â &r #F       â t"    Ä t"Jüúøô    ä t#\!    ätJ^üúøô!    ä x!N   ÿÀŽ ä xRR22ÿø� ä x !   pRò Ä xRTv  ÿÀŽ ä xRÿ   R  ä rNNÿ     €  ä r   "    ä rJ#R   J    ä xNNÿ   ÿÀŽ â &rH      â x! d  J    Ä x#1u  J    â $~ÿ   ÿ Ž â xMww  ÿÀŽ äxÿ   M  ä xMMˆˆˆˆÿÀŽ ä >rNÿN   ÿ  â $~"ÿ   ÿ Ž â xww  ÿÀŽ â $~Jÿ   ÿ Ž â xMM22ÿø� â 0rNÿN    A   â t    Ä t'üúøô    ä t'"    äx"MTv  ÿÀŽ ä tJJüúøô    â 0rÿV    A   â  xNN   ?  A  â xMÿ   M  ä r""       Ä Er         â  r N     A  â r''M   J    â &r!!@   ÿ   ä #r    ÿ   â x" d  '    äxIK   ÿÀŽ ä x'"1u  '    â &r!#B   !   â x"Lww  ÿÀŽ Ä x LˆˆˆˆÿÀŽ â $~#ÿ   ÿ Ž â xLÿ   L  â $~Mÿ   ÿ Ž â x I   pRò â $~]ÿ   ÿ Ž â |I       ä x  22ÿø� ä xLLww  ÿÀŽ â &rD   !   â t##"    Ä t"Iüúøô"    ä xI Tv  ÿÀŽ ä t!ML    ätL]üúøôL    ä x ÿ      È r!!    L    ä xLS   ÿÀŽ â $~Jÿ   ÿ Ž â r"#I   "    ä x#Tww  ÿÀŽ ä x L   pRô ä |N       Ä x ÿ   T  ä tIJ#    ärSSÿ       â $~Mÿ   ÿ Ž â tJNüúøô#    â $~#ÿ   ÿ Ž â xTTˆˆˆˆÿÀŽ ä x  ww  ÿÀŽ Ä xSSÿ   ÿÀŽ ä xTT22ÿø� ä tL#     ätMMüúøô     ä x " d  !    ä >rSÿS   ÿ  Ä x#TTv  ÿÀŽ â &r  C   ÿ   â xTÿ   T  â 0rSÿS    A   â x""1u  !    â &r'F      â x!U   ÿÀŽ ä r#I#   J    â &r""E       â rLLT   M    â  x S   ?  A  â x !   pRô Ä xJ# d  L    â  r       A  â Er        ä rKKÿ     €  ä rUUÿ       â &r"JG   "   â #r     ÿ   â $~Jÿ   ÿ Ž â xKKÿ   ÿÀŽ Ä x##1u  L    ä xUUÿ   ÿÀŽ ä xWww  ÿÀŽ ä |L       ä >rKÿK   ÿ  ä >rUÿU   ÿ  Ä t!J    ätJLüúøô    ä xÿ   W  â 0rKÿK    A   â xWWˆˆˆˆÿÀŽ â 0rUÿU    A   â xww  ÿÀŽ â $~Lÿ   ÿ Ž â xWW22ÿø� â &r"#H   "   â  xKK   ?  A  â 0r$ÿ$    A   â t#L    â xUU   ?  A  â tLNüúøô    Ä Er        â  rKK$     A  â xWTv  ÿÀŽ â  rU$U     A  â Er" "       ä xWÿ   W  â #rK      â r!   J    â #rU"       â rL#W   L    â $~#ÿ   ÿ Ž â x Oww  ÿÀŽ â $~Iÿ   ÿ Ž â x'ÿ   O  Ä x! d  L    äx1u  L    â $~Lÿ   ÿ Ž â t"#     â$~Kÿ   ÿ Ž â |J       â $~Mÿ   ÿ Ž â x''ww  ÿÀŽ ä t#Iüúøô     Ä x Qww  ÿÀŽ ä tIJ'    ätJLüúøô'    ä t'K     ätLMüúøô     ä x ÿ   Q  Ä xOOˆˆˆˆÿÀŽ â $~Kÿ   ÿ Ž â |M       ä xQQˆˆˆˆÿÀŽ ä x  ww  ÿÀŽ ä xOO22ÿø� ä xQQ22ÿø� Ä tMM     ätNKüúøô     ä x OTv  ÿÀŽ ä xKQTv  ÿÀŽ ä xQÿ   Q  ä r "    #    Ä x#ÿ   P  ä x"Pww  ÿÀŽ äxPPˆˆˆˆÿÀŽ ä xOÿ   O  ä r''K   L    â $~Kÿ   ÿ Ž â rNMQ   N    â $~Mÿ   ÿ Ž â xPP22ÿø� â $~Qÿ   ÿ Ž â rIIO   J    Ä |O       ä xJ#ww  ÿÀŽ ä t#K"    ät"Müúøô"    ä xKPTv  ÿÀŽ ä tLOJ    Ä tMQüúøôJ    ä xPÿ   P  ä xJ' d  N    ä r##K   "    â &r"!@   ÿ   â xN'1u  N    â &rJJC   ÿ   â rLLP   M    â &rB   "   â x!  d  I    â &rJNE   J   â x"# d  L    Ä xI 1u  I    â &r !D      â x#1u  L    â &r""G   J   â x#ÿ   Z  ä x!Zww  ÿÀŽ â$~Jÿ   ÿ Ž â xZZˆˆˆˆÿÀŽ â $~Lÿ   ÿ Ž â x'#ww  ÿÀŽ Ä xZZ22ÿø� ä t#J!    ätJLüúøô!    â $~Pÿ   ÿ Ž â x!ZTv  ÿÀŽ ä |N       â &rH   "   â r!#!   J    Ä x#ÿ   %  ä x"%ww  ÿÀŽ â &r IF       â tKN'    ätLPüúøô'    â $~'ÿ   ÿ Ž â x%%ˆˆˆˆÿÀŽ â $~Iÿ   ÿ Ž Æ x%%22ÿø� â $~Nÿ   ÿ Ž â t''"    äxZÿ   Z  ä t"Iüúøô"    ä |J       ä x##ww  ÿÀŽ Ä xI%Tv  ÿÀŽ ä rLKZ   L    ä tJJ#    ätKNüúøô#    ä x%ÿ   %  ä r"'I   "    Ä x'ˆˆˆˆÿÀŽ ä xIÿ     â $~Mÿ   ÿ Ž â x#! d  L    â$~Oÿ   ÿ Ž â x!!1u  L    ä r%J%   K    ä |L       Ä xww  ÿÀŽ ä x''22ÿø� ä xKIww  ÿÀŽ ä tIM    ä xJ'Tv  ÿÀŽ ä tLLK    Ä tOüúøô    ä tKNüúøôK    ä x'ÿ   '  ä rIIJ       â &r#@   ÿ   â rLL'   K    ä x#" d  %    Ä x%"1u  %    ä x"I d  L    â &r!B      â x!X   ÿÀŽ ä xLI1u  L    â &r""C   ÿ   â xJ    ÿÀŽ ä x !   pRô â &r#D      â x!YˆˆˆˆÿÀŽ â &rLLE   "   â x J   pRò â $~Jÿ   ÿ Ž â x"ÿ   Y  Ä x#   ÿÀŽ ä xYYww  ÿÀŽ â &r%F      â x!!22ÿø� ä x%"ww  ÿÀŽ ä x #   pRö ä t"JY    Ä tYNüúøôY    ä x#!Tv  ÿÀŽ ä r  ÿ     €  ä r""#   Y    ä x#&   ÿÀŽ ä rÿ     € Ä x  ÿ   ÿÀŽ â $~Pÿ   ÿ Ž â x #   pRò â $~Rÿ   ÿ Ž â xÿ   ÿÀŽ ä >r ÿ    ÿ  ä r#&ÿ     €  ä t'P%    Ä tJRüúøô%    ä x!ÿ   !  ä >rÿ   ÿ  â 0r ÿ     A   â rXXÿ       ä x##ÿ   ÿÀŽ â 0rÿ    A   â r!'!   J    Ä xXXÿ   ÿÀŽ â  x     ?  A  â >r#ÿ#   ÿ  ä x&" d  !    â x   ?  A  â x%"1u  !    ä >rXÿX   ÿ  ä Er!         â  r       A  â0r ÿ#    A   â Er#        â &r&&G   L   â  r     A  â 0rXÿX    A   â #r  !   ÿ   â &r%%H   &   â #r#   ÿ   â  x#    ?  A  â  x!X   ?  A  â Er        Ä Er% %       â  r!$!     A  â r#$#     A  â !r22       â !r11       â #r!       â #r#%      â G‰       € ê r 9   ÿàó È r!;   ÿäÿ  Ê �y     ¨ �yJ     è  �yK     " r"9   ÿàó är&9   ÿàõ â �y     d $x#   ;Ž  Ä $x'   ; Æ �y "     è �y&     " $~Lÿ   ÿ Ž â |N       â $~Pÿ   ÿ Ž ä $~'ÿ   ÿ Ž ä $~&ÿ   ÿ Ž â x%ˆˆˆˆÿÀŽ ÄO x!ÿ     ä xIww  ÿÀŽ ä x%%22ÿø� ä x"!ww  ÿÀŽ ä/ tLI    ät NüúøôI    â $~Iÿ   ÿ Ž â x!%Tv  ÿÀŽ Ä x%ÿ   %  ä r!        ä x JˆˆˆˆÿÀŽ ä�t#I"    â$~Iÿ   ÿ Ž â t"'üúøô"    â $~'ÿ   ÿ Ž â x!Jww  ÿÀŽ Ä |L       ä x  22ÿø� ä xJÿ   J  ä r"#%   "    ä t#&!    ät&Lüúøô!    Ä x% Tv  ÿÀŽ ä xJJww  ÿÀŽ ä x! d  "    äx1u  "    ä r##%   &    â &r!!@   ÿ   â t"'J    Ä x&ÿ   K  â&r!B   !   â x%Kww  ÿÀŽ ä t'IüúøôJ    â $~Jÿ   ÿ Ž â xKKˆˆˆˆÿÀŽ ä |N       ä xI&ww  ÿÀŽ Ä xKK22ÿø� ä t&J%    ä tJNI    â$~Nÿ   ÿ Ž â tMPüúøôI    ä x ÿ      ä xIKTv  ÿÀŽ Ä t%Lüúøô%    â $~Lÿ   ÿ Ž â xKÿ   K  ä r "    '    ä r%&I   %    ä rJJK   M    ä x# d       Ä x&ÿ     äx##1u       â &rD   !   â x % d  J    äxJ%1u  J    â &r#F      â x%ww  ÿÀŽ â &r  C   ÿ   â x"ˆˆˆˆÿÀŽ Ä x'&ww  ÿÀŽ â &r JE       â xK    ÿÀŽ ä tL%    ä x""22ÿø� ä tIL'    ätLNüúøô'    Ä x'   ÿÀŽ ä x K   pRò ä x&"Tv  ÿÀŽ ä t%Nüúøô%    ä x"ÿ   "  ä x '   pRô Ä r  ÿ     €  ä r&   %    ä rII"   L    ä r"ÿ       ä >r ÿ    ÿ  ä x d  I    Ä >r"ÿ"   ÿ  â 0r ÿ     A   â xI1u  I    â &rG       â Er        â 0r"ÿ"    A   â  x     ?  A  â &rIH      Æ  x!"   ?  A  â  r       A  â Er        ä  r!$!     A  â #r     ÿ   È #r !       È !r55        ä G‰       € ê r 9   ÿàó È r!;   ÿäÿ  Ê �y     ¨ �yJ     è  �yK     " r"9   ÿàó är&
9   ÿàõ â �y     d $x#   ;Ž  Ä $x'   ; Æ �y "     è �y&     " $~Lÿ   ÿ Ž â |N       â $~Pÿ   ÿ Ž ä $~'ÿ   ÿ Ž ä $~&ÿ   ÿ Ž â x%ˆˆˆˆÿÀŽ ÄO x!ÿ     ä xIww  ÿÀŽ ä x%%22ÿø� ä x"!ww  ÿÀŽ ä/ tLI    ät NüúøôI    â $~Iÿ   ÿ Ž â x!%Tv  ÿÀŽ Ä x%ÿ   %  ä r!        ä x JˆˆˆˆÿÀŽ ä�t#I"    â$~Iÿ   ÿ Ž â t"'üúøô"    â $~'ÿ   ÿ Ž â x!Jww  ÿÀŽ Ä |L       ä x  22ÿø� ä xJÿ   J  ä r"#%   "    ä t#&!    ät&Lüúøô!    Ä x% Tv  ÿÀŽ ä xJJww  ÿÀŽ ä x! d  "    äx1u  "    ä r##%   &    â &r!!@   ÿ   â t"'J    Ä x&ÿ   K  â&r!B   !   â x%Kww  ÿÀŽ ä t'IüúøôJ    â $~Jÿ   ÿ Ž â xKKˆˆˆˆÿÀŽ ä |N       ä xI&ww  ÿÀŽ Ä xKK22ÿø� ä t&J%    ä tJNI    â$~Nÿ   ÿ Ž â tMPüúøôI    ä x ÿ      ä xIKTv  ÿÀŽ Ä t%Lüúøô%    â $~Lÿ   ÿ Ž â xKÿ   K  ä r "    '    ä r%&I   %    ä rJJK   M    ä x# d       Ä x&ÿ     äx##1u       â &rD   !   â x % d  J    äxJ%1u  J    â &r#F      â x%ww  ÿÀŽ â &r  C   ÿ   â x"ˆˆˆˆÿÀŽ Ä x'&ww  ÿÀŽ â &r JE       â xK    ÿÀŽ ä tL%    ä x""22ÿø� ä tIL'    ätLNüúøô'    Ä x'   ÿÀŽ ä x K   pRò ä x&"Tv  ÿÀŽ ä t%Nüúøô%    ä x"ÿ   "  ä x '   pRô Ä r  ÿ     €  ä r&   %    ä rII"   L    ä r"ÿ       ä >r ÿ    ÿ  ä x d  I    Ä >r"ÿ"   ÿ  â 0r ÿ     A   â xI1u  I    â &rG       â Er        â 0r"ÿ"    A   â  x     ?  A  â &rIH      Æ  x!"   ?  A  â  r       A  â Er        ä  r!$!     A  â #r     ÿ   È #r!!       È !r44!       ä G‰       € ê r 9   ÿàó È r!;   ÿäÿ  Ê �y     ¨ �yJ     è  �yK     " r"9   ÿàó är&9   ÿàõ â �y     d $x#   ;Ž  Ä $x'	   ; Æ �y "     è �y&     " $~Lÿ   ÿ Ž â |N       â $~Pÿ   ÿ Ž ä $~'ÿ   ÿ Ž ä $~&ÿ   ÿ Ž â x%ˆˆˆˆÿÀŽ ÄO x!ÿ     ä xIww  ÿÀŽ ä x%%22ÿø� ä x"!ww  ÿÀŽ ä/ tLI    ät NüúøôI    â $~Iÿ   ÿ Ž â x!%Tv  ÿÀŽ Ä x%ÿ   %  ä r!        ä x JˆˆˆˆÿÀŽ ä�t#I"    â$~Iÿ   ÿ Ž â t"'üúøô"    â $~'ÿ   ÿ Ž â x!Jww  ÿÀŽ Ä |L       ä x  22ÿø� ä xJÿ   J  ä r"#%   "    ä t#&!    ät&Lüúøô!    Ä x% Tv  ÿÀŽ ä xJJww  ÿÀŽ ä x! d  "    äx1u  "    ä r##%   &    â &r!!@   ÿ   â t"'J    Ä x&ÿ   K  â&r!B   !   â x%Kww  ÿÀŽ ä t'IüúøôJ    â $~Jÿ   ÿ Ž â xKKˆˆˆˆÿÀŽ ä |N       ä xI&ww  ÿÀŽ Ä xKK22ÿø� ä t&J%    ä tJNI    â$~Nÿ   ÿ Ž â tMPüúøôI    ä x ÿ      ä xIKTv  ÿÀŽ Ä t%Lüúøô%    â $~Lÿ   ÿ Ž â xKÿ   K  ä r "    '    ä r%&I   %    ä rJJK   M    ä x# d       Ä x&ÿ     äx##1u       â &rD   !   â x % d  J    äxJ%1u  J    â &r#F      â x%ww  ÿÀŽ â &r  C   ÿ   â x"ˆˆˆˆÿÀŽ Ä x'&ww  ÿÀŽ â &r JE       â xK    ÿÀŽ ä tL%    ä x""22ÿø� ä tIL'    ätLNüúøô'    Ä x'   ÿÀŽ ä x K   pRò ä x&"Tv  ÿÀŽ ä t%Nüúøô%    ä x"ÿ   "  ä x '   pRô Ä r  ÿ     €  ä r&   %    ä rII"   L    ä r"ÿ       ä >r ÿ    ÿ  ä x d  I    Ä >r"ÿ"   ÿ  â 0r ÿ     A   â xI1u  I    â &rG       â Er        â 0r"ÿ"    A   â  x     ?  A  â &rIH      Æ  x!"   ?  A  â  r       A  â Er        ä  r!$!     A  â #r     ÿ   È #r!!       È !r>>!       â !r00       â !r//       â G‰       € ê r9   ÿàó È r;   ÿäÿ  Ê �y    ¨ �y'    è  �yI    " r 9   ÿàó Æ �y    d $x!   ;Ž  â r"9   ÿàó È �y       â $x#   ;Ž  Ê �y"     â  $~Jÿ   ÿ Ž â |L       â $~Kÿ   ÿ Ž ä $~!ÿ   ÿ Ž â/ |#       ä x&ÿ     äO x%ww  ÿÀŽ Ä xˆˆˆˆÿÀŽ ä x&&ww  ÿÀŽ ä x22ÿø� ä tJ%    äxTv  ÿÀŽ ä t%Lüúøô%    Ä t K&    ät!!üúøô&    ä xÿ     ä r   %    ä r    !    â $~!ÿ   ÿ Ž â x ÿ   '  Ä� x d      äx1u      ä x'ww  ÿÀŽ â $~%ÿ   ÿ Ž â x  ww  ÿÀŽ â &r@@   ÿ   â t!    â$~&ÿ   ÿ Ž â t#üúøô    â $~Jÿ   ÿ Ž â x#Iww  ÿÀŽ Ät!!     ät"%üúøô     ä x ÿ   I  ä xIIˆˆˆˆÿÀŽ ä t%&#    ät&Jüúøô#    â $~Kÿ   ÿ Ž â |#       Ä x  ww  ÿÀŽ ä x''ˆˆˆˆÿÀŽ ä xII22ÿø� ä tJ#     ä x''22ÿø� ä x#ITv  ÿÀŽ Ä tKKüúøô     ä xIÿ   I  ä x 'Tv  ÿÀŽ ä x'ÿ   '  ä r#%#   &    ä rJJI   K    Ä r        ä r!!'   "    ä x    ÿÀŽ ä x # d  J    â &r@B   @   â x d  !    â$~"ÿ   ÿ Ž â x    pRò â &rC C   ÿ   â x!1u  !    â $~&ÿ   ÿ Ž â xww  ÿÀŽ ÄxˆˆˆˆÿÀŽ ä x ÿ     â $~%ÿ   ÿ Ž â xJ#1u  J    ä t"    ä |#       ä x" ww  ÿÀŽ Ä x22ÿø� ä t&üúøô    ä x&   ÿÀŽ ä t##"    äx Tv  ÿÀŽ ä t"%üúøô"    Ä xÿ     ä x &   pRô ä r  ÿ     €  ä r        ä r##   "    ä rÿ       Ä >r ÿ    ÿ  â &rCJE   C   â x d  #    â&rD   @   â >rÿ   ÿ  â 0r ÿ     A   â x#1u  #    â &rG   C   ä &r!F      â  x     ?  A  â 0rÿ    A   Ä &r#H      â Er        â  r       A  â  x   ?  A  ä Er        â #r     ÿ   â  r$     A  È #r        È !r33        â x66@   ÿàÿ ä x99 	  ÿàõ ä | 6   pbò ä x== 	  ÿà÷ â $r;ÿÿ   ; â xAA€   ÿàÿ Æ $r?ÿÿ   ?Ž Ì D       À â Gy  Åÿÿÿÿƒ ê Ay        € ê y      "   " Ey  0    € â r ÿ   pRð är ÿ   pRò Ö G‰  ð     € ê z ÿ `  pPð ä r (   ÿ8Ž ä z ÿ a   Sð Æ ˆs  2 þÿ H   ô ˆƒ  1€þÿ H   è ˆƒ  0 ÿÿ H   è ˆƒ  /€ÿÿ H   â G‰  p     € ê ˆs  5   H   è ˆs  1€þÿ H   è ˆs  4€  H   è ˆs  0 ÿÿ H   è ˆs  >   H   è ˆs  /€ÿÿ H   è ˆs  3€  H   ä Ay        € ê {          ì M        € ê y       &   "> z ÿ `  pPð â $tÿ   ÿ Ž Ä y      '   b z ÿ a   Sð â $z   x  )Ž È $z  ~   Ž È/ %v  j   Ž È G	  P    € ê „y(     H   " z ) r  p`ð Æ „y( €   H   b r (ÿ   pVp  Æ „y(    H   ¨ „y (    H   è „y	( €  H   („y(    H   ( „y(    H   ( „y( €  H   ( „y(    H   ( „y
( €  H   ( „y( €  H   " !r2       Æ „y( €  H   " !r1       â/ !r0       âO !r         ä� !r	       â!r       â !r         ä !r       Æ ‰       b !r       Æ ‰      ¢ !r

/       Æ ‰      â !r

       È !r
       â !r        â/ !r       ÈO ‰      " !r
       Æ� ‰      h ‰      ¨ ‰	
      â !r       â !r       â/ !r       ÈO ‰      " !r
	       Æ� ‰ €    h ‰  €    ¨ ‰	 €    â !r       â !r       È/ ‰ €    " !r
        ÆO ‰  @    b !r		       â� x)   ÿàÿ ä ‰
 @    ¤ z  r  p`ô ä ‰	 @    ä x (   pVt â !r       â !r         È/ ‰ @    " x)   ÿàÿ â!r
       âO x))   ÿàÿ â ‰       b z  r  p`ò â !r	       ä� ‰      ¢ x (   pVò  Æ ‰      â !r       â !r        È/ ‰       " #‚+   .   â !r	       âO !r       Æ� †‰     â z ) r  p`ð â #¢		+   .   â#’+   .   ä x (   pVp  ä †© 	   è †™    â !r        Ì M	        € ê #r++   .   Ê †y +   â My        € ê $v ÿ h  ÿ Ž Ê x     pBð Ú G	  @    € ê x     pRð Ú G‰  À    € ê x     pRð Ú G	   $    € ê „y	(     H   "z ) r  p`ð â Ey  p    € ä „y (    H   b r (ÿ   pVp  Æ „y(    H   ¨ „y(    H   è „y(    H   ( „y(  
  H   ( „y( €  H   ( „y( €   H   ( „y( €  H   ( „y( €  H   " !r		2       â !r  5       Æ/ !r	       âO !r         Æ� !r       â!r         È ‰      " !r4       Æ ‰       b !r1       È !r       È !r       â !r       â !r        È/ ‰      ( ‰      b !r       â !r	       È/ ‰ €    ( ‰
	 €    h „y( €  H   ¢ !r       Æ „y( €
  H   â !r
	
       Æ/ ‰ @    ( ‰
 @    b !r       âO !r       â !r
       È/ ‰       ( ‰      b !r       â !r	        â !r       â/ G	  �     € ê #r-   ,   â #r 	+   .   Æ  x;ª¸¿  A  Ì s        $ !t  €?    Ì s        $  r     A  È  r      A  Ê †y     ä Ay        € ê ‰       b Ey  Ð    € æ ‰      ¨ „y(    H   è „y(    H   ( „y(    H   â !r         Æ/ „y(    H   â !r       âO x)   ÿàÿ ä ‰       d z  r  p`ð ä ‰      $ x (   pVp  â !r0       â� !r>       â!r        Ä/ !r       â !r	       ä ‰ €    ( ‰
	 €    b !r       Æ „y(  	  H   ¢ !r       Æ „y(    H   â !r
	
       Æ/ ‰ @    ( ‰
 @    b !r       âO !r       â !r
       È/ ‰       ( ‰      b !r       â !r	        â !r       â/ G	  �     € ê #r-   ,   â #r 	+   .   Æ  x;ª¸¿  A  Ì s        $ !t  €?    Ì s        $  r     A  È  r      A  Ê †y    ä Ay        € ê ‰       b Ey       € æ ‰      ¨ „y( €	  H   è „y( €  H   è „y( €  H   è „y( €  H   â !r         Æ/ „y( €  H   â !r       ÆO ‰       h ‰      ( „y( €  H   ¢ !r        â/ !r	       È ‰ €    ( ‰
	 €    b !r3       ÈO !r       â !r       â x)   ÿàÿ ä !r       â !r
	
       â/ ‰ @    " z  r  p`ð Æ ‰
 @    b x (   pVp  â !r       â !r
       È/ ‰       ( ‰      b !r	        â !r       â/ G	  �     € ê #r-   ,   â #r 	+   .   Æ  x;ª¸¿  A  Ì s        $ !t  €?    Ì s        $  r     A  È  r      A  Ê †y    ä Ay        € ê ‰      b !r/       â x))   ÿàÿ Æ !r       â z ) r  p`ð Æ !r       â x (   pVp  È ‰       ¢ !r       Ê/ ‰      " !r         ÊO ‰       b !r	       Ê ‰
	 €    " !r        Ê/ ‰ €    b !r
	
       Ê ‰
 @    " !r       Ê/ ‰ @    b !r
       Ê ‰      " !r       Ê/ ‰       ¢ !r       â M	        € ì #r,-   ,   â/ !r        ÆO  x,;ª¸¿  A  â #r++   .   Ê s        $ !t  €?    Ì s        $  r,,     A  È  r++,     A  Ê †y +   â My        € ê „y(     H   " z ) r  p`ð â Ey       € ä „y (    H   b r (ÿ   pVp  Æ „y(    H   ¨ „y(    H   è „y	(    H   („y(  
  H   ( „y( €   H   ( „y( €  H   ( „y( €  H   ( „y( €
  H   â !r2       â !r  5       Æ/ !r       âO !r         Æ� !r		       â!r        È ‰ 	      " !r1       Æ ‰      b !r       È !r       â !r 	        â !r       È/ ‰       ( ‰      h „y( €  H   ¢ !r        â !r       È/ ‰ €    ( ‰ €    h „y( €  H   â !r4       âO !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       È� !r       â !r       â !r
	
       È/ ‰       ( ‰
      b !r	        â !r
       â/ G	  À     € ê #r -   ,   È  x '7=  A  â x *BL?  A  Æ #t   €?   È  r     A  â  x    ?  A  â #r 	+   .   È s     $   $ !t  €?    È  r     A  È  r      A  Ê †y     ä Ay        € ê ‰       b Ey       € æ ‰      ¨ „y(    H   è „y(    H   ( „y(  	  H   â !r         Æ/ „y(    H   â !r       ÆO ‰       h ‰      ( „y(    H   ¢ !r0       â� !r>       â!r        Æ/ !r       â !r       â ‰ €    ( ‰ €    h „y(    H   â !r       âO !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       È� !r       â !r       â x)   ÿàÿ â !r
	
       Æ/ ‰       " z  r  p`ð Æ ‰
      b x (   pVp  â !r	        â !r
       Ö/ G	  À     € ê #r -   ,   È  x '7=  A  â x *BL?  A  Æ #t   €?   È  r     A  â  x    ?  A  â #r 	+   .   È s     $   $ !t  €?    È  r     A  È  r      A  Ê †y    ä Ay        € ê ‰       b Ey  Ð    € æ ‰      ¨ „y( €  H   è „y( €	  H   ( „y( €  H   â !r         Æ/ „y( €  H   â !r       ÆO ‰       h ‰      ( „y( €  H   ¢ !r3       È� !r       â!r        â/ !r       È ‰ €    ( ‰ €    h „y( €  H   â !r       âO !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       â x)   ÿàÿ â !r
	
       Æ/ ‰       " z  r  p`ð Æ ‰
      b x (   pVp  â !r	        â !r
       Ö/ G	  À     € ê #r -   ,   È  x '7=  A  â x *BL?  A  Æ #t   €?   È  r     A  â  x    ?  A  â #r 	+   .   È s     $   $ !t  €?    È  r     A  È  r      A  Ê †y    ä Ay        € ê ‰      b !r/       â x))   ÿàÿ Æ !r       â z ) r  p`ð Æ !r       â x (   pVp  È ‰       ¢ !r       Ê/ ‰      " !r         ÊO ‰       b !r       Ê ‰ €    " !r        Ê/ ‰ €    b !r	       Ê ‰
	 @    " !r       Ê/ ‰ @    b !r
	
       Ê ‰
      " !r       Ê/ ‰       ¢ !r
       â M	        € ì #r,-   ,   â/ !r        ÆO  x,'7=  A  â x,*BL?  A  â#r++   .   ä #t,  €?   â x,,   ?  A  Æ  r     A  Ì s     $   $ !t  €?    È  r,,     A  È  r++,     A  Ê †y +   â My        € ê x     pRð Ú G‰  P    € ê x     pRð Ú G	  `    € ê „y(     H   " z ) r  p`ð â Ey       € ä „y (    H   b r (ÿ   pVp  Æ „y(    H   ¨ „y(    H   è „y	(    H   („y(  
  H   ( „y( €   H   ( „y( €  H   ( „y( €  H   ( „y( €
  H   â !r2       â !r  5       Æ/ !r       âO !r         Æ� !r		       â!r        È ‰ 	      " !r1       Æ ‰      b !r       È !r       â !r 	        â !r       È/ ‰       ( ‰      h „y( €  H   ¢ !r        â !r       È/ ‰ €    ( ‰ €    h „y( €  H   â !r4       âO !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       È� !r       â !r       â !r
	
       È/ ‰       ( ‰
      b !r	        â !r
       â/ G	  À     € ê #r -   ,   Ê 	z  i    � â #r 	+   .   È  x;ª¸¿  A  â 	z   i    � È 	z   i €  � â s        $ !t  €?    Ì s        $  r     A  È  r      A  Ê †y     ä Ay        € ê ‰       b Ey       € æ ‰      ¨ „y(    H   è „y(    H   ( „y(  	  H   â !r         Æ/ „y(    H   â !r       ÆO ‰       h ‰      ( „y(    H   ¢ !r0       â� !r>       â!r        Æ/ !r       â !r       â ‰ €    ( ‰ €    h „y(    H   â !r       âO !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       È� !r       â !r       â x)   ÿàÿ â !r
	
       Æ/ ‰       " z  r  p`ð Æ ‰
      b x (   pVp  â !r	        â !r
       Ö/ G	  À     € ê #r -   ,   Ê 	z  i    � â #r 	+   .   È  x;ª¸¿  A  â 	z   i    � È 	z   i €  � â s        $ !t  €?    Ì s        $  r     A  È  r      A  Ê †y    ä Ay        € ê ‰       b Ey  Ð    € æ ‰      ¨ „y( €  H   è „y( €	  H   ( „y( €  H   è „y( €  H   â !r         Æ/ „y( €  H   â !r       ÆO ‰       h ‰      " !r3       È� !r       â!r        â/ !r       È ‰ €    ( ‰ €    h „y( €  H   ¢ !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       âO !r       â x)   ÿàÿ â !r
	
       Æ/ ‰       " z  r  p`ð Æ ‰
      b x (   pVp  â !r	        â !r
       Ö/ G	  À     € ê #r -   ,   Ê 	z  i    � â #r 	+   .   È  x;ª¸¿  A  â 	z   i    � È 	z   i €  � â s        $ !t  €?    Ì s        $  r     A  È  r      A  Ê †y    ä Ay        € ê ‰      b !r/       â x))   ÿàÿ Æ !r       â z ) r  p`ð Æ !r       â x (   pVp  È ‰       ¢ !r       Ê/ ‰      " !r         ÊO ‰       b !r       Ê ‰ €    " !r        Ê/ ‰ €    b !r	       Ê ‰
	 @    " !r       Ê/ ‰ @    b !r
	
       Ê ‰
      " !r       Ê/ ‰       ¢ !r
       â M	        € ì #r,-   ,   â/ !r        ÈO 	z,, i    � â #r++   .   È  x,;ª¸¿  A  â 	z++ i    � È 	z++ i €  � â s        $ !t  €?    Ì s        $  r,,     A  È  r++,     A  Ê †y +   â My        € ê „y(     H   " x)   ÿàÿ äx)   ÿàÿ â„y (    H   b z ) r  p`ð ä z  r  p`ò â „y( €   H   ¢ z  r  p`ô Ä r (ÿ   pVp  â„y( €  H   â x (   pVò  äx (   pVt â „y(    H   " x))   ÿàÿ Æ „y(    H   " z ) r  p`ö Æ „y(    H   " x (   pVö Æ „y(    H   ( „y( €  H   ( „y( €  H   ( „y
(    H   " !r2       Æ „y(  	  H   " !r  5       Æ/ „y	(    H   b!r1       ÆO „y(  
  H   ¢ !r44       Æ� „y( €  H   â !r0       Æ„y( €
  H   " !r>       Æ „y(    H   " !r       Æ „y(    H   " !r         Æ „y( €  H   " !r       Æ „y( €  H   " !r4       Æ „y( €  H   " !r

       Æ „y!( €	  H   â !r       Æ „y#( €  H   â !r	       Æ/ „y%( €  H   â !r        âO !r       ä� ‰       " !r       Æ‰      b !r

       Æ ‰      ¢ !r       Æ ‰      â !r/       Æ ‰	
      " !r3       Æ ‰      " !r       â !r         â !r!       â !r       â/ !r##       â !r       äO ‰       " !r%%       â !r       Ä� ‰#      b !r

	       Æ‰%      ¢ !r       Æ ‰      è ‰	      ( ‰      ( ‰
      " !r        Æ ‰      " !r#       â/ !r%       äO ‰  €    b !r       Æ� ‰      ¢ !r		       Æ‰      â !r       Æ ‰ €    " !r
       Æ ‰	 €    " !r       Æ ‰ €    ( ‰
 €    " !r         Æ/ ‰ €    b !r       âO !r       ä� ‰  @    ¢ !r       Æ‰ €    â !r	       Æ ‰ €    " !r       Æ ‰ @    " !r

       Æ ‰	 @    " !r       Æ/ ‰ @    h ‰
 @    b !r        ÆO ‰ @    ¢ !r       â� !r       ä‰       â !r       Æ ‰ @    " !r		       Æ ‰ @    " !r       Æ/ ‰      b !r
       Æ ‰	      b !r       ÆO ‰      ¨ ‰
      ¢ !r        Æ� ‰      â !r       â #‚ +   .   â !r       Æ‰      " !r       Æ/ ‰      b !r		       â #‚-   ,   â !r       äO #’	+   .   â  ‚      A  â !r
       â #’-   ,   Æ †‰     â !r       â� #¢+   .   â  ’     A  ä #¢	-   ,   â !r       ä †™    â  ¢		     A  â !r       È/ †© 	   â M9        € ê #r++   .   â #r,-   ,   È  r++,     A  Ê †y +   â My        € ê „y(     H   " z ) r  p`ð â Ey  À    € ä „y (    H   b r (ÿ   pVp  Æ „y(    H   ¨ „y(    H   è „y	(    H   („y(  
  H   ( „y( €   H   ( „y( €  H   ( „y( €  H   ( „y( €
  H   â !r2       â !r  5       Æ/ !r       âO !r         Æ� !r		       â!r        È ‰ 	      " !r1       Æ ‰      b !r       È !r       â !r 	        â !r       È/ ‰       ( ‰      h „y( €  H   ¢ !r        â !r       È/ ‰ €    ( ‰ €    h „y( €  H   â !r4       âO !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       È� !r       â !r       â !r
	
       È/ ‰       ( ‰
      b !r        â !r
       â/ G	  à     € ê #r -   ,   Ê 	x   à@  � Ê  x #ÛÙ?  A  È  x ;ª¸¿  A  â #r +   .   Ê s        " 	x    à@  � È 	x    àÀ  � Ê !t   €?    â !t  €?    È s	        $  r	     A  È  r     A  Ê †y     ä Ay        € ê ‰       b Ey       € æ ‰      ¨ „y(    H   è „y(    H   ( „y(  	  H   â !r         Æ/ „y(    H   â !r       ÆO ‰       h ‰      ( „y(    H   ¢ !r0       â� !r>       â!r        Æ/ !r       â !r       â ‰ €    ( ‰ €    h „y(    H   â !r       âO !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       È� !r       â !r       â x)   ÿàÿ â !r
	
       Æ/ ‰       " z  r  p`ð Æ ‰
      b x (   pVp  â !r        â !r
       Ö/ G	  à     € ê #r -   ,   Ê 	x   à@  � Ê  x #ÛÙ?  A  È  x ;ª¸¿  A  â #r +   .   Ê s        " 	x    à@  � È 	x    àÀ  � Ê !t   €?    â !t  €?    È s	        $  r	     A  È  r     A  Ê †y    ä Ay        € ê ‰       b Ey  ð    € æ ‰      ¨ „y( €  H   è „y( €	  H   ( „y( €  H   è „y( €  H   â !r         Æ/ „y( €  H   â !r       ÆO ‰       h ‰      ( „y( €  H   ¢ !r3       È� !r       â!r        â/ !r       È ‰ €    ( ‰ €    b !r       âO !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       â x)   ÿàÿ â !r
	
       Æ/ ‰       " z  r  p`ð Æ ‰
      b x (   pVp  â !r        â !r
       Ö/ G	  à     € ê #r -   ,   Ê 	x   à@  � Ê  x #ÛÙ?  A  È  x ;ª¸¿  A  â #r +   .   Ê s        " 	x    à@  � È 	x    àÀ  � Ê !t   €?    â !t  €?    È s	        $  r	     A  È  r     A  Ê †y    ä Ay        € ê ‰      b !r/       â x))   ÿàÿ Æ !r       â z ) r  p`ð Æ !r       â x (   pVp  È ‰       ¢ !r       Ê/ ‰      " !r         ÊO ‰       b !r       Ê ‰ €    " !r        Ê/ ‰ €    b !r	       Ê ‰
	 @    " !r       Ê/ ‰ @    b !r
	
       Ê ‰
      " !r       Ê/ ‰       ¢ !r
       â M	        € ì #r,-   ,   â/ !r        ÈO 	x,,  à@  � â #r++   .   È  x,#ÛÙ?  A  â 	x++  à@  � Æ  x;ª¸¿  A  â 	x++  àÀ  � Ê s        " !t++  €?    â !t  €?    Ì s        $  r,,     A  È  r+,+     A  Ê †y +   â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      "   $ z  w  p`ð Ú M	        € ê y      !   " ¹z  l      â Ey       € â ™x?    â y      %   b ¹z  F   
   â $rÿÿ   ÿ Ž ä $rÿÿ   ÿ Ž â xÿ     È |    p`ð â $x   ÿ Ž Ø/ G	  €    € ê y      &   " $tÿ   ÿ Ž ä $z x  Ž È %v \   Ž Ê �y     ™ ¢  $rÿÿ   ÿ Ž ä $x   ÿ Ž ä 'z m   Ž Æ x   ÿÀŽ â $vÿ X  ÿ Ž â zÿ n    â $z r  ÿŽ â x   ÿüŽ â $vÿ Y  ÿ Ž ä $rÿÿ   
Ž â x	ÿ     â%x$   ÿ Ž â x
ÿ     Æ $z o  Ž ä %x	$    Ž È $z u  ÿŽ ä $x	   ÿ Ž ä %x$    Ž È %x$    Ž â x   ÿÀŽ Æ %x	$   Ž È %x$   Ž â z Z  ÿàñ ä x   ÿàÿ â $x   ÿ Ž â x    ÿàó â %x	$   Ž Æ zÿ [  þ  â $z q  Ž â x   ÿÀŽ â $rÿÿ   ÿ Ž â z Z  ÿàó ä x

ÿÿÿÿÿàñ ä x   ÿàõ Æ $tÿÿÿÿÿÿ  â zÿ [  $~ â $z  t  Ž äO$z  t  Ž ä $z q  Ž ä $rÿÿ   ÿ Ž ä %x$   
Ž È %x$   Ž È %x$   Ž È %x
$   
Ž È r   ÿàñ â�yôÿÿ ™ ¨  $x     â �yüÿÿ ™ è  �y0   ™ ( �y/   ™ h �y.   ™ ¢ r   ÿàñ Æ �y+   ™ ¤ $x     ä �y    ™ ¨  �y#   ™ ¢ $rÿÿ    Ž ä/ $rÿÿ    Ž â r,   ÿàñ â �yøÿÿ ™ ¨  �y%üÿÿ ™ ¨ �y&   ™ ¢ $x-	     Æ �y    ™ ¨ �y    ™ ¨ �y",    ‘ ¨ �y,   ‘ è �y$   ™ è �y!   ™ è �y øÿÿ • â r
   ÿàñ Æ �y'   ™ ä $x     ä �y)àÿÿ • è  �y(    ‘ è  �y*   ‘ â  ‚x       Ä ‚x  ÿþý     â $~-ÿ   ÿ Ž ä/ $~1ÿ   ÿ Ž ä $~2ÿ   ÿ Ž ä $~4ÿ   ÿ Ž ä $~5ÿ   ÿ Ž â x   ÿàÿ Ä x€  ÿàõ Ê $rÿÿ    â x0ˆˆˆˆÿÀŽ äx,0ww  ÿÀŽ ä x0ÿ   0  ä x22ÿø� â $~ÿ   ÿ Ž â t-,    ät-1üúøô,    Ä x00ww  ÿÀŽ ä xTv  ÿÀŽ ä x,/ˆˆˆˆÿÀŽ ät0    är   -    â $~-ÿ   ÿ Ž â t01üúøô0    Ä xÿ     ä x/ww  ÿÀŽ ä x,,22ÿø� ä r   0    ä x/ÿ   /  ä t-    Ä t1üúøô    ä x-,Tv  ÿÀŽ ä x1/ww  ÿÀŽ ä x0 d      äx/1u      â $~ÿ   ÿ Ž â r-       Ä x-.ww  ÿÀŽ äOx.ˆˆˆˆÿÀŽ ä t221    ät14üúøô1    ä x,ÿ   ,  ä xÿ   .  Ä t-    äx22ÿø� ä t-4üúøô-    â $~4ÿ   ÿ Ž â r22,   1    â $~.ÿ   ÿ Ž â x3ww  ÿÀŽ ä x1Tv  ÿÀŽ Ä x, d  2    äx1u  2    ä t443    ät3.üúøô3    ä r1   -    ä xÿ     Ä x-+ˆˆˆˆÿÀŽ â$~1ÿ   ÿ Ž â x.+ww  ÿÀŽ ä r44   3    ä x3ÿ   +  ä x--22ÿø� â $~+ÿ   ÿ Ž â t11.    â$~ÿ   ÿ Ž â t.5üúøô.    Ä x33ww  ÿÀŽ ä x2-Tv  ÿÀŽ ä x d  4    ä t++3    ätüúøô3    ä r.12   .    â &r00%   ÿ   â x1u  4    â $~2ÿ   ÿ Ž â x1#ww  ÿÀŽ â $~4ÿ   ÿ Ž â x3ÿ   #  Ä x##ˆˆˆˆÿÀŽ â $~5ÿ   ÿ Ž â x33ww  ÿÀŽ â &r0/&   0   â x##22ÿø� ä t221    ät/4üúøô1    ä x1ÿ   -  Ä t-53    ät44üúøô3    ä x3#Tv  ÿÀŽ ä x#ÿ   #  â &r,,   0   â r++1       â &r   ÿ   â r/23   /    Ä r4-#   4    â &r    ,   â x. d  +    â&r      â� x,/ d  4    äx"   ÿÀŽ ä x#.1u  +    â &r,,%   ÿ   â x+   ÿÀŽ â &r      â x//1u  4    Ä x    pRð ä xÿ   $  ä x$ww  ÿÀŽ â$~.ÿ   ÿ Ž â x$$ˆˆˆˆÿÀŽ â $~0ÿ   ÿ Ž â x +   pRò â &r&/&   ,   â x,ww  ÿÀŽ â $~+ÿ   ÿ Ž â x$$22ÿø� â $~-ÿ   ÿ Ž â t.    Ä t%0üúøô    ä t+,    äx+$Tv  ÿÀŽ ä t,-üúøô,    ä r""ÿ        ä x-!ww  ÿÀŽ Ä r+   %    ä x+"ÿ   ÿÀŽ ä t".-    ät%0üúøô-    ä x-ÿ   !  ä x!!ˆˆˆˆÿÀŽ Ä x--ww  ÿÀŽ ä >r+ÿ+   ÿ  ä x!!22ÿø� ä x$ÿ   $  ä t..-    ät/0üúøô-    â 0r+ÿ+    A   â x-!Tv  ÿÀŽ Ä x0ÿ   !  ä r!$   ,    â 0rÿ     A   â r%"-   %    ä r..0   /    â  x +   ?  A  â x" d  !    äx!1u  !    Ä x% d  .    â r       A  â Er        â &r""   &   â x%%1u  .    â &r   ÿ   ä #r    ÿ   â xÿ   '  â &r !    "   â x 'ww  ÿÀŽ â $~ÿ   ÿ Ž Ä &r%      â xww  ÿÀŽ â $~"ÿ   ÿ Ž ä $~$ÿ   ÿ Ž â t     â $~!ÿ   ÿ Ž â t"    ät"$üúøô    ä x(   ÿÀŽ Ä x''ˆˆˆˆÿÀŽ ä t !üúøô     ä x!*   ÿÀŽ ä x    pRð ä x''22ÿø� ä rÿ     €  Ä x !   pRò ä r((ÿ        ä x'Tv  ÿÀŽ ä x'ÿ   '  ä x((ÿ   ÿÀŽ ä r**ÿ     €  Ä xÿ   ÿÀŽ ä r         ä r'   "    ä >r(ÿ(   ÿ  ä x**ÿ   ÿÀŽ ä >rÿ   ÿ  Ä x  d      â0r(ÿ(    A   â >r*ÿ*   ÿ  â 0rÿ    A   â x 1u      â &r      â  x (   ?  A  â 0r*ÿ*    A   â |    p`ð â &r#      â  x   ?  A  â 0r)ÿ)    A   â Er         â &r      â  r       A  â  x**   ?  A  â Er        â  r)     A  â Er        â #r     ÿ   â  r))*     A  â #r      â x@  ÿàó Ä x€  ÿà÷ â #r))       â !r       â $rÿÿ   Ž  ä $rÿÿ   Ž â !r)       â G‰  �ðÿÿÿÿƒ ê Ay        € ê ‰      " $x   Ž Æ ‰      d z  p  p`ð È x    pDp  â !r       â !r       È/ ‰       ( ‰      b !r         â !r       È/ ‰  €    ( ‰ €    b !r        â !r       È/ ‰ @    ( ‰ @    b !r       â !r       È/ ‰      h  ‰	      ¢  M	        € ê y       &   â $x   ÿ Ž â !r       â/ $z s  Ž â !r		       âO $tÿ   ÿ Ž â x    p"ò Ä r ÿ   p"ð Ö $ÿÿ   	 Ž ä $z  v  Ž ÈŸ %v j   Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      "   $ z  w  p`ð Ú M	        € ê y      &   " $tÿ   ÿ Ž â ¹z  F   
   â $z x  Ž È %v \   Ž Ê �y    ™ b ¹z  l      â Ey   "    € â ™x?    â y      !   b x
      ÿ  â $rÿÿ   ÿ Ž â x      ÿ  â $rÿÿ   ÿ Ž â xÿ     È/ |    p`ð Ú G	   "    € ê $rÿÿ   ÿ Ž â y      %   " $x    ÿ Ž â Ey  !    € â 'z m   Ž Æ x     ÿÀŽ â $x   ÿ Ž â zÿ n    â $rÿÿ   ÿ Ž â x     ÿüŽ â $rÿÿ   ÿ Ž â x   ÿÀŽ â $rÿÿ   
Ž â x
ÿ      â$z r  ÿŽ â x   ÿÀŽ â $z o  Ž â xÿ      â %x$   ÿ Ž È $z	 u  ÿŽ ä %x$    Ž È %x	$    Ž È %x	$    Ž È %x$   Ž È $x   ÿ Ž ä $x   ÿ Ž ä %x$   Ž È %x$   Ž â z Z  ÿàõ ä x   ÿàÿ â $z t  Ž âz Z  ÿàñ â $rÿÿ   ÿ Ž â x   ÿà÷ â $z q  Ž â x    ÿàó â $x   ÿ Ž â zÿ [  Dþ â %x
$   
Ž â zÿ [  þ  Ä x   ÿÀŽ â $z q  Ž â z ÿ `  pPð â %x$   Ž â x   ÿüŽ ä z ÿ a   Sð â $z t  Ž ä $vÿ X  ÿ Ž ä %x$   Ž È $v!ÿ Y  ÿ Ž ä $v#ÿ `  ÿ Ž ä $v%ÿ a  ÿ Ž ä $rÿÿ   ÿ Ž ä %x$   Ž È r   ÿàó Ê $x   !Ž  Ê �y6   ™ ¨  �y5   ™ â  r   ÿàó Æ �y4   ™ $ $x   !Ž  ä �y*   ™ h �y/ÿÿÿ ‘ è �y+    ‘ â $rÿÿ    Ž Ä $rÿÿ    Ž Ê �y)üÿÿ ™ ( �y(   ™ " r   ÿàó Æ �y'    ™ $ $x	   !Ž  ä �y&   ™ ( �y øÿÿ • ( �y,   ™ " $rÿÿ    Ž Ä/ $rÿÿ    Ž â �y.   ™ ( �y$ôÿÿ ™ h �y"üÿÿ ™ h �y-   ™ b r
   ÿàó Æ �y øÿÿ ™ d $x   !Ž  ä �y    ™ h �y1   ™ h �y0ÿÿÿ ‘ h �y3    ‘ h �y2àÿÿ • b ‚x       Ä ‚x  ÿþý     â $~:ÿ   ÿ Ž ä $~<ÿ   ÿ Ž ä $~9ÿ   ÿ Ž ä $~;ÿ   ÿ Ž ä $~=ÿ   ÿ Ž â x   ÿàÿ Ä x6ˆˆˆˆÿÀŽ äOx76ww  ÿÀŽ ä x8ÿ   6  ä x22ÿø� ä t:7    ät6<üúøô7    â $~7ÿ   ÿ Ž â x88ww  ÿÀŽ Ä xTv  ÿÀŽ ä t78    är   6    ä t79üúøô8    ä x6ÿ     â $~ÿ   ÿ Ž â x5ww  ÿÀŽ â� $~8ÿ   ÿ Ž â r66   7    Ä x5ˆˆˆˆÿÀŽ ä x7ÿ   5  ä t    ät8üúøô    â $~8ÿ   ÿ Ž â x522ÿø� â $~:ÿ   ÿ Ž â x97ww  ÿÀŽ Ä x75Tv  ÿÀŽ ä t889    ät9:üúøô9    ä x5ÿ   5  ä x d  6    ä r7       Ä x1u  6    â $~6ÿ   ÿ Ž â x4ˆˆˆˆÿÀŽ äx74ww  ÿÀŽ ä r885   9    ä x5ÿ   4  ä x422ÿø� Ä t667    â$~9ÿ   ÿ Ž â t7:üúøô7    ä x<5ww  ÿÀŽ ä x:4Tv  ÿÀŽ ä x d  8    äx51u  8    Ä t;<    ä r66:   7    â $~:ÿ   ÿ Ž â t89üúøô<    ä x;/   ÿÀŽ ä x9*ww  ÿÀŽ â$~<ÿ   ÿ Ž â x7ÿ   4  Ä x ;   pRò ä t4:9    ä x:ÿ   *  ä r77   8    ä t<üúøô9    â $~9ÿ   ÿ Ž â r;/ÿ     €  â $~8ÿ   ÿ Ž â x**ˆˆˆˆÿÀŽ Ä x/:ww  ÿÀŽ â &r:)   ÿ   â x;;ÿ   ÿÀŽ ä x**22ÿø� ä t9/    â&r:(   :   â t/8üúøô/    ä >r8ÿ;   ÿ  Ä x9*Tv  ÿÀŽ â &r'   :   â x*ÿ   *  â 0r8ÿ8    A   ä &r5&      â r*   /    â 0r*ÿ     A   â x +   ÿÀŽ â  x58   ?  A  â r49       Ä x     pRò â  r 5*     A  â Er9        â $~ÿ   ÿ Ž â x5,ww  ÿÀŽ â $~4ÿ   ÿ Ž â x/ d      äx1u      â #r 9   ÿ   â x6 d  7    Ä x ,ˆˆˆˆÿÀŽ ä t5    ät54üúøô5    ä x4ÿ   ,  â $~9ÿ   ÿ Ž â x761u  7    â $~;ÿ   ÿ Ž â x, 22ÿø� â &r $   ÿ   â x44ww  ÿÀŽ Ä x,Tv  ÿÀŽ â &r 7"       â t694    ät7;üúøô4    â $~8ÿ   ÿ Ž â x4ÿ   ,  â $~,ÿ   ÿ Ž â r   5    ä x5.ww  ÿÀŽ Ä r464   7    ä x.ˆˆˆˆÿÀŽ ä t,,5    ät78üúøô5    ä x5ÿ   .  ä x.22ÿø� Ä x85ww  ÿÀŽ ä x6.Tv  ÿÀŽ ä x d  4    äx1u  4    ä t4;üúøô8    â$~;ÿ   ÿ Ž â t598    Ä r,,6   7    ä x8-ww  ÿÀŽ äx7-ˆˆˆˆÿÀŽ ä x9ÿ   -  ä t6;8    ät-=üúøô8    â $~<ÿ   ÿ Ž â x8722ÿø� â $~7ÿ   ÿ Ž â x99ww  ÿÀŽ Ä x:ÿ   .  ä x;8Tv  ÿÀŽ ä t.<9    ät77üúøô9    ä x8ÿ   8  ä r55:   4    â &r4)   ÿ   â r-6;   -    Ä r..8   7    â &r /        â x, d  5    â &r(   4   â x/- d  .    â &r       ä &r '      â x-1u  .    â &r//$   ÿ   â x5,1u  5    â $~ÿ   ÿ Ž â x,ÿ   1  â $~.ÿ   ÿ Ž â x-1ww  ÿÀŽ â &r"   /   â x11ˆˆˆˆÿÀŽ â $~/ÿ   ÿ Ž â x40   ÿÀŽ â $~7ÿ   ÿ Ž â x,,ww  ÿÀŽ Ä t-    äx1122ÿø� ä x 4   pRô ä t-.üúøô-    ä x43   ÿÀŽ ä t./,    Ä t/7üúøô,    ä x,1Tv  ÿÀŽ ä r00ÿ       ä x1ÿ   1  ä x 4   pRô ä r,   -    Ä r/.1   /    ä x00ÿ   ÿÀŽ ä r++ÿ     €  ä r33ÿ       ä x, d  /    ä >r0ÿ0   ÿ  Ä x++ÿ   ÿÀŽ ä x33ÿ   ÿÀŽ â &r,       â x/1u  /    â 0r0ÿ0    A   â >r+ÿ+   ÿ  ä >r3ÿ3   ÿ  â &r 5&       ä &r//      â  x0   ?  A  â 0r+ÿ+    A   Ä 0r3ÿ3    A   â Er-         â 0r2ÿ2    A   â  r *     A  â  x++   ?  A  â  x33   ?  A  â Er        â #r  -   ÿ   â Er, /       â  r++2     A  â  r323     A  â |    p`ò Ä #r++      â #r,3,       â G‰  @    € ê x ÿ     Ê $z  q   Ž È $z t   Ž È %x$   Ž Ê r#   ÿàõ Ê $x   % Ê �y1    ¨  �y0    è  �y/    " %x$   Ž Æ �y     d r#   ÿàõ Ê $x   % Ê �y-ÿÿÿ  è �y.     " $~7ÿ   ÿ Ž ä $~9ÿ   ÿ Ž ä $~ÿ   ÿ Ž ä $~8ÿ   ÿ Ž Ä $~:ÿ   ÿ Ž ä $~ÿ   ÿ Ž ä/ $~ÿ   ÿ Ž ä $~;ÿ   ÿ Ž â x31ˆˆˆˆÿÀŽ äOx5ÿ   1  ä x41ww  ÿÀŽ Ä x3322ÿø� ä x55ww  ÿÀŽ ä t174    ät49üúøô4    ä x3Tv  ÿÀŽ ä t5    Ä t5üúøô5    ä x3ÿ   3  ä r1   4    â $~1ÿ   ÿ Ž â r43   5    ä x0ˆˆˆˆÿÀŽ ä�x0ww  ÿÀŽ Ä x0ÿ   0  ä x22ÿø� ä x30ww  ÿÀŽ ä t11    ät6üúøô    ä x0Tv  ÿÀŽ Ä x5ÿ   /  äx d  4    ä r010   6    ä x1u  4    â &r)   ÿ   â x1/ˆˆˆˆÿÀŽ ä t83    Ä t4:üúøô3    â $~3ÿ   ÿ Ž â x85ww  ÿÀŽ ä x6/ww  ÿÀŽ ä x5122ÿø� ä t1;8    ät83üúøô8    Ä t/76    â$~7ÿ   ÿ Ž â x35Tv  ÿÀŽ ä t69üúøô6    ä xÿ     ä x5ÿ   5  ä r   4    â &r(      â r3/3   6    Ä r815   8    â $~5ÿ   ÿ Ž â x0 d      äx/3 d  8    äx131u  8    â $~3ÿ   ÿ Ž â x01u      â &r//$   ÿ   â x0 ww  ÿÀŽ â&r'      â x ˆˆˆˆÿÀŽ â &r11"   /   â x4ÿ      â &r&      â t 30    Ä t35üúøô0    â $~5ÿ   ÿ Ž â x44ww  ÿÀŽ ä x6-   ÿÀŽ ä x22ÿø� ä t554    ät47üúøô4    Ä x7.   ÿÀŽ ä x 6   pRô ä x0Tv  ÿÀŽ ä xÿ     ä x 7   pRö ä r--ÿ       Ä r  0   3    ä r55   4    ä r..ÿ     € ä >r-ÿ-   ÿ  ä x/  d  5    ä>r.ÿ.   ÿ  â 0r-ÿ-    A   â x5 1u  5    â &r//    1   â Er        â 0r.ÿ.    A   â  x--   ?  A  â &r/5   /   Æ  x.   ?  A  â  r *-     A  â Er/ /       ä  r2     A  â #r     ÿ   È #r/       È !r       â !r+       â !r,       â G‰       € ê r#   ÿàõ Ê $x	   % Ê �y/    ¨ �y.    è  �y-    " r
#   ÿàõ Æ �y+    d $x   % Ê �y ÿÿÿ  è �y,     â $~3ÿ   ÿ Ž ä $~5ÿ   ÿ Ž ä $~4ÿ   ÿ Ž ä $~6ÿ   ÿ Ž Ä $~7ÿ   ÿ Ž â x1ÿ   /  äO x0/ww  ÿÀŽ äx//ˆˆˆˆÿÀŽ ä x11ww  ÿÀŽ ä x//22ÿø� ä t30    â$~3ÿ   ÿ Ž â t5üúøô0    â $~5ÿ   ÿ Ž â x/Tv  ÿÀŽ Ä/ t41    â$~4ÿ   ÿ Ž â t06üúøô1    ä x/ÿ   /  ä r       ä r0/   0    â $~/ÿ   ÿ Ž â xÿ   .  Ä� x d  0    äx1u  0    â $~0ÿ   ÿ Ž â x.ww  ÿÀŽ â &r))   ÿ   â xww  ÿÀŽ ä t0    â&r((   )   â t4üúøô    â $~4ÿ   ÿ Ž â x1-ww  ÿÀŽ Ät03üúøô    äx3ÿ   -  ä x..ˆˆˆˆÿÀŽ ä x--ˆˆˆˆÿÀŽ ä t//    ä t41    Ä x43ww  ÿÀŽ ä x..22ÿø� ä x3-22ÿø� ä t16üúøô1    ä t554    ät67üúøô4    Ä x-.Tv  ÿÀŽ ä x43Tv  ÿÀŽ ä x.ÿ   .  ä x3ÿ   3  ä r-       ä r./.   0    â $~/ÿ   ÿ Ž â r4   1    â $~1ÿ   ÿ Ž â r353   6    â $~0ÿ   ÿ Ž â x d  .    Ä x1u  .    â $~.ÿ   ÿ Ž â x) d  3    â&r'   (   â x1u  3    ä x+ww  ÿÀŽ â&r$)$   ÿ   â x-ÿ   +  â &r&      â t)/    â &r$"   $   â x++ˆˆˆˆÿÀŽ Ä t1üúøô    ä x--ww  ÿÀŽ ä x1    ÿÀŽ ä x++22ÿø� ä t/0üúøô-    ä x0,   ÿÀŽ Ä x 1   pRô ä t..-    ä x-+Tv  ÿÀŽ ä x+ÿ   +  ä x 0   pRö ä r  ÿ       Ä r)-       ä r+.+   /    ä r,,ÿ     € ä >r ÿ    ÿ  ä x d  +    ä>r,ÿ,   ÿ  â 0r ÿ     A   â x+1u  +    â &r    $   Ä 0r,ÿ,    A   â  x    ?  A  â &r++      â Er        ä  x',   ?  A  â  r *     A  â Er+ +       ä  r'2'     A  â #r     ÿ   È #r '+       È !r        â x@  ÿàõ ä x##@  ÿà÷ ä x€  ÿàù â $r!ÿÿ   ! â x€  ÿàû â $r%ÿÿ   %Ž ä $rÿÿ    ä $rÿÿ   Ž â D       À â Gy  �áÿÿÿÿƒ ê Ay       € ê $r
ÿÿ    Ž ä $rÿÿ    Ž Ä Ay        € ê z ÿ `  pPð È z ÿ a   Sð Ú G	  P    € ê ‰      ( ‰       b !r       â !r        È/ ‰       ( ‰      b !r         â !r       È/ ‰  €    ( ‰ €    b !r        â !r       È/ ‰ @    ( ‰ @    b !r       â !r       È/ ‰      ( ‰	      b !r       â !r		       â/ Gy  €    € ê ‰      ( ‰       h ‰      ¨ ‰      â !r       â !r        È/ ‰      " !r       ÆO ‰      b !r
       Æ� ‰       ¨ ‰	
      â !r       â !r       È/ ‰ €    " !r         âO !r
	       È� ‰  €    h ‰	 €    ¨ ‰ €    â !r       Ê ‰ @    " !r        â/ !r
	       ÈO ‰ @    b !r       Æ� ‰
 @    ¨ ‰	 @    â !r       â !r       â/ !r 
       ÈO ‰      " !r	       Æ� ‰	      h ‰       ¨ ‰      â !r       â !r
	       â/ !r	        âO !r       â� y      %   $ $x    Ž Ê z   p  p`ð È x    pDp  Ú M	        € ê z ÿ d  pPð â $z v   Ž âz ÿ ^  pPò ä z ÿ e   Sð ä z ÿ _  Sò Ö $ÿ   ÿ Ž ä $ÿ   ÿ Ž ä % d   Ž È % ^   Ž ä �	     ¨ �     â $x   ÿ Ž Ê r ÿ   p"ô äx    p"ö äx    p"ø äx    p"ú Î $"ÿÿ    Ž â z ÿ `  pPô â $2ÿÿ   	 Ž ä $Bÿÿ   
 Ž â z ÿ a   Sô â $Rÿÿ    Ž È       A  ÈO !       È� G©  €    € ê z ÿ f  pPð ä z ÿ b  pPò ä z ÿ g   Sð ä z ÿ c  Sò Ö $ÿ   ÿ Ž ä $ÿ   ÿ Ž ä % f   Ž È % b   Ž ä �	     ¨  �     â $x   ÿ Ž Ê x øÿÿÿp"ô â$vÿ h  ÿ Ž â x üÿÿÿp"ö är ÿ   p"ø äx    p"ú Î $"ÿÿ    Ž â x    pBô â $2ÿÿ   	 Ž ä $Bÿÿ   
 Ž ä $Rÿÿ    Ž È       A  ÈO !       â� G)  P    € ê x    pRð Ú G‰  �     € ê x    pRð Ú G	  ð    € ê  x;ª¸¿  A  Ì s        $ !t  €?    Ì s        $  r     A  È  r     A  â Gy  P    € ê  x'7=  A  â x*BL?  A  Æ #t  €?   â  x   ?  A  Æ  r     A  Ì s     $   $ !t  €?    È  r     A  È  r     A  â Gy  °    € ê x    pRð Ú G‰  à     € ê x    pRð Ú G	        € ê 	z i    � ä 	z i    � Æ  x;ª¸¿  A  â 	z i €  � Ê s        $ !t  €?    Ì s        $  r     A  È  r     A  â Gy  Ð     € ê  r     A  â Gy  °     € ê 	x  à@  � ä 	x  à@  � Æ  x#ÛÙ?  A  â 	x  àÀ  � Æ  x;ª¸¿  A  ä !t  €?    È s        $ !t  €?    Ì s        $  r     A  È  r     A  ä y      "   ( y      &   b $z  s   Ž ä $tÿ   ÿ Ž ä $z  v   Ž È/ %v  j   Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      "   " ¹z  l      â Ey       € â ™x?    â y      !   b ¹z  F   
   â x      ÿ  â r ÿ       â y$      %   ¢ $r.ÿÿ   ÿ Ž â x      ÿ  â x       ÿ  â rE ÿ       â x"      ÿ  â rF ÿ       â x      ÿ  â r  ÿ       â $rÿÿ   ÿ Ž â r ÿ   pRð Ä r ÿ   pRò är   ÿ(Ž È/ r   ÿ� ä x$$   ÿ   äO xÿ     È |    pbô Ú G)  €    € ê y	      '   " r ÿ       â $r
ÿÿ   ÿ Ž â xþÿÿÿÿÀŽ â y      &   b x$   ÿàÿ ä x-        ä r ÿ       ä z% q  ÿàÿ â 'z	 y   Ž â r  €ÿàÿ Æ 'z s  
 Ž â/ zÿ z    ä x   ÿ   â $z$ p  Ž â zÿ t    â $z |  Ž â x9   ÿüŽ â $z |  Ž â xÿ     â $z v  Ž Ä $z v  Ž ä $z, p  Ž ä %v X  -Ž È %x   Ž â x*   ÿà÷ Æ $z(	 }  ÿŽ â z X  ÿàù â $z w  ÿŽ â r+ÿ   ÿäÿ â $x   ÿ Ž â x   ÿà÷ â %v,, X  -Ž â r((   ÿàõ ä x;ÿ     â $vÿ q  ÿ Ž â r5 -       Ä zÿ Y  „þ â $x-   Ž âr7   ÿŽ â$x/   Ž âr&   ÿŽ â$x1   Ž â x;   ;  â $x3   Ž â x=   ÿ   ä rXÿÿ   ÿä Ä �y8ýÿÿ ‘ ¨ �yCþÿÿ ‘ ¨ �y)ÿÿÿ ‘ è �y2    ‘ " r,9   ÿàõ Æ �y   ‘ h �yD   ‘ h �y'   ‘ b $x	   5 Æ �y0   ‘ h �y@ýÿÿ ‘ h �yGþÿÿ ‘ h �y4ÿÿÿ ‘ h �y<    ‘ b r(   ÿàõ Æ �y:   ‘ h �y?   ‘ b xP $       Æ �y>   ‘ b rÿX   ÿä Æ �yA   ‘ b %v Z  P Ž â r%(   ÿà÷ ä �yH*üÿÿ ‘ b $x$   ÿŽ â r
ÿX   ÿäÿ â %v Z  P Ž â r=   ÿàõ ä r   ÿàÿ â $x	
$   ÿŽ È $x   ; â r=   ÿàõ â $rÿÿ   5 Ž â r	   ÿàÿ â �yO    • b r
 ,       Æ �y6   ™ b r	;   ÿä ä r7(   ÿàõ â�yB   ™ b r-(   ÿà÷ Æ �yK
    ‘ h �yL   ™ h �yJ   ™ b rÿX   ÿä är
ÿX   ÿäÿ â �yN   ™ b %v Z  P Ž Æ �yI   ™ b $x$   ÿŽ ä %v Z  P Ž â �yM   ™ f $x
$   ÿŽ â/ r=   ÿàõ â $x   Ž â �yR   ™ d r   ÿàÿ Ä r;   ÿä â �yQ    • b r
=   ÿà÷ Æ �yS   ™ b r	&(   ÿàõ är;   ÿäÿ â �yV   ™ b r/(   ÿà÷ â/ $rÿÿ   X ä �yW    • b rÿX   ÿäÿ Æ �yZ   ™ h �y    • b %v	 Z  P Ž Æ/ �yT
   ™ b $x	$   ÿŽ ä $x$   ÿŽ â r=   ÿàõ â �y   ™ b r	   ÿàÿ â %v Z  P Ž ä �yU
   ™ d $x	   ; â r=   ÿà÷ â �yY
   ™ b r   ÿàÿ Æ �y[
   ™ b r;   ÿäÿ ä r3(   ÿà÷ â�y\   ™ h �y`   ™ b r1(   ÿàõ ä/ rÿX   ÿäÿ â�y^    • b r
ÿX   ÿä Æ �ya    • h �yc   ™ b %v Z  P Ž Æ/ �yd   ™ b $x
$   ÿŽ ä %v Z  P Ž â �y]   ™ f $x$   ÿŽ â r=   ÿàõ â $x   Ž â r
=   ÿà÷ â �y_   ™ b r   ÿàÿ Ä r;   ÿä â�yb   ™ b r;   ÿäÿ Æ �y   ™ h# �y	
   ™ h �yf   ™ h �yP
   ™ h �ye   ™ h �yg   ™ h �yh
   ™ h �y   ™ h �y
   ™ h �y    • h �y    • b ‚x       Ä ‚x  ÿþý     â |i       ä |       ä |       ä/ x8Cv  8    ÈO x))Tp  8    È� x)2T  )    Èxÿ   )  ä x))ˆˆˆˆÿÀŽ â $~Cÿ   ÿ Ž â x
ww  ÿÀŽ ä x))22ÿø� ä x88ww  ÿÀŽ ä ti
    Ä t2üúøô
    ä xDDv      äx
)Tv  ÿÀŽ ä tC8    ätüúøô8    ä x)ÿ   )  Ä x''Tp  D    ä r
       ä r)   2    ä x'0T  '    ä x d      äx1u      Ä x'ˆˆˆˆÿÀŽ ä xGGv  @    ä |       ä xDDww  ÿÀŽ ä x'ÿ   '  ä x
22ÿø� Ä x4Tp  G    ä tD    ätDüúøôD    ä |0       ä x''ww  ÿÀŽ ä x
Tv  ÿÀŽ Ä |2       ä x<T      ä t0'    är   D    ä x??v  :    ä t02üúøô'    â $~2ÿ   ÿ Ž â xÿ     Ä xˆˆˆˆÿÀŽ ä xGGww  ÿÀŽ ä |4       ä x>>Tp  ?    ä |8       ä |<       Ä x'ww  ÿÀŽ ä x22ÿø� ä t2G    äx
ÿ   
  ä t24üúøôG    ä x>AT  >    Ä t)8'    ät4<üúøô'    ä x'Tv  ÿÀŽ ä xÿ     ä r
   0    â $~0ÿ   ÿ Ž â x
ÿ   >  Ä x>>ˆˆˆˆÿÀŽ ä r''   2    ä x??ww  ÿÀŽ ä x   ÿàõ ä r4)   4    ä |2       Ä |)       ä x

ww  ÿÀŽ ä x>>22ÿø� ä t0?    â$rÿÿ    â t02üúøô?    ä t?)üúøô
    Ä x)>Tv  ÿÀŽ ä r Hÿ   pRö äx Hÿ   pRô â$xHH€   ÿ Ž â |       ä r0)   0    â &r)6   ÿ   â t2
    Ä x>ÿ   >  ä xHH€  ÿÀŽ â &r)B   )   â x
 d      ä r?2>   ?    ä xHH@     € ä x1u      â &r)
L   )   â x' d  4    Ä x'1u  4    ä x0 d  ?    äx'01u  ?    â &r0J   ÿ   â xHHÿ      â &r2N   )   ä &r0I   0   â x)ÿT  H    â 0rOÿO    A   Ä &r?
M   0   â x Kÿ   pRø â  x))   ?  A  â r Kÿ   pRú â&r8R   ?   â xKK   ÿ   â  r?)O     A  â Er2 2       ä xKK€  ÿÀŽ â &r66   ÿ   Ä #r?2      â &r2S   ÿ   âxKK@     € â &rSS   ÿ   ä &r6B   6   â xKKÿ      â &r?V   2   ä&rSV   S   â x0ÿT  K    â &r6L   6   Ä &r?
Z   ?   ä&rZZ   S   ä &r6'N   6   â x00   ?  A  â 0rÿ    A   ä &rZ'   Z   âEr6 6       â &r?   ?   â  r4O0     A  â  r)     A  â ErZ Z       â  r0     A  â #r46      â &r4T   ÿ   Ä #r#Z   #   â &rTT   ÿ   ä &r\   ÿ   â Er? ?       â &rTU   T   ä&r`      ä&r4U   4   â #rF?   F   â &rTY   T   Ä &r
c      ä0rQÿQ    A   ä &r4
Y   4   â Er8 8       â &rT'[   T   ä &rd      â  rA)Q     A  â 0rWÿW    A   Ä &r4[   4   ä 0r^ÿ^    A   â ErT T       â #rA8      â Er        â  r20W     A  âEr4 4       â  r)^     A  â rA)W     A  â #r!2T   !   â &r2]   ÿ   â #rE   E   â #r"A4   "   â &r   ÿ   Ä &rJJ   ÿ   ä&r\\   ÿ   ä&r4]   ÿ   ä&r   ÿ   ä &r	   ÿ   ä&r		   ÿ   â x    ÿàÿ â &r?_   2   Ä &rJI   J   ä&r\`   \   ä&r4_   4   ä &rf      ä&rf      ä &rP      Ä &r	P   	   â |    pbô â &r?
b   ?   ä&rJM   J   ä&r\c   \   ä&r4b   4   ä &r
      Ä &r      ä&r
      ä&r	   	   ä &r2e   ?   ä &rJ'R   J   ä&r\'d   \   Ä &re'e   4   ä&rg      ä&r'g      ä0raÿa    A   ä &rh      ä&r''h   	   Ä 0rÿ    A   ä 0rÿ    A   â ErJ J       â  r4)a     A  âEr\ \       â  rQ0Q     A  âEr2 2       â  r?0^     A  âEre e       â  ra0a     A  âEr        â  r)     A  âEr        â  r	0     A  âEr        â  r
)     A  âEr' '       â  r0     A  â x,,   ÿà÷ Ä x**   ÿàù â #r QJ       â #r ?\       â #r42      â #rae      â #r.   .   â #r	      â #r
      â #r'      â x%%    ÿàÿ ä x77    ÿàÿ Ä x--    ÿàÿ ä x&&    ÿàÿ ä x//    ÿàÿ ä x11    ÿàÿ ä x33    ÿàÿ ä r5ÿ5   ÿäÿ Ä r+ÿ+   ÿä â D)       À â Gy  péÿÿÿÿƒ ê Ay        € ê Ey  0    € â G‰      € ê $x   Ž Ê ˆs  øÿ H   è ˆs €øÿ H   è ˆs  ùÿ H   è ˆs  €ùÿ H   è ˆs F úÿ H   è ˆs #€úÿ H   è ˆs " ûÿ H   è ˆs !€ûÿ H   è ˆs E üÿ H   è ˆs  €üÿ H   è ˆs  ýÿ H   è ˆs €ýÿ H   è ˆs . þÿ H   è ˆs €þÿ H   è ˆs  ÿÿ H   è ˆs €ÿÿ H   ä Ay        € ê {          ì M        € ê „y     H   b z $ r  p`ð Ä z1  r      â „y €   H   ¢ r ÿ   p"p Æ „y	 €  H   â $x/1   ÿŽ ä$x71   ÿŽ â „y    H   ( „y    H   h y      &   ( „y    H   h „y €  H   h y      '   h „y €  H   â !r       Æ/ „y    H   â !r       ÆO ‰      " $z x  $Ž â x$$   ÿàÿ â !r	        â� „y& €  H   â !rFF       Æ„y €  H   â !r""       â$z ~  Ž ä ‰F      h ‰      â  /   ÿàù â !r       Ä ‰"      ¢ !r##       â !r       ä „y(    H   â $ÿÿ   ÿ â 
  j  ÿˆ ä‰      $ 
  k   Ä ‰      â !r..       Æ „y*    H   â !rFF       Æ/ „y €  H   â !r&&!       Æ „y,    H   â !r""       âO !r         ä ‰      b !r       â r1   ÿ† ä ‰#      " !r
       â� 
 r  ÿàó Ä ‰	 €    ¢ rÿÿ   ÿäÿ ä z j  ÿ† â ‰
      â ÿÿ   ÿäÿ  â !r((E       â 
 j  ÿ„ â ‰"      " z k  � ä 
 k   â !r       â z $ r  p`ò â !r$       â „y5 €  H   â !r       â/ x    p"ò â !r**       â ‰F      b !r##       â 7   ÿàù â !r+       â ‰      " !r 	       âO $x	1   ÿŽ â!r-,       â ‰&      ¢ !r
       â� r1   ÿ† Ä ‰  @    â r		   ÿàõ â !r""       âÿÿ   ÿä â ‰ €    " rÿÿ   ÿä ä z	 j  ÿ„ â ‰'$      b 
 j  ÿˆ ä z		 k   â ‰(      b !rFF       â/ r,ÿÿ   ÿäÿ Ä ‰!       b z
 j  ÿ† ä 
 k   â ‰      b !r       â z k  ,� â !r&&       âO ‰%*      " !r         â� $x1   ÿŽ â !r55       â ‰).      ¢ !r       â1   ÿàÿ Ä ‰0+      â    ÿàõ â !r,$'       â6   ÿàù â ‰2-      " !r((       â $ÿÿ   ÿä ä ‰#      b !r  !       â/ 
 j  ÿ„ â $9ÿÿ   ÿ â ‰       b !r       â 6 j  ÿˆ Ä ‰ @    b !r**%       â 
 k  $ â !r..)       âO ‰%&      " 6 k  9 â !r0+0       ä� ‰'(      ¢ /   ÿàÿ â !r2-2       Ä‰)       â    ÿà÷ â !r#       ä‰ €    " 4ÿÿ   ÿäÿ â !r9        ä/ ‰+*      b  j  ÿ† â !r        Ä ‰-,      b  k  4� ä 7   ÿàÿ â ‰ €    b !r&&%       â x        ä ‰/.      " !r(('       âO    ÿàõ Ä ‰10      ¢ !r$ )       â�  ÿÿ   ÿä ä ‰32      â !r       â%v j   Ž ä ‰5      " !r**+       Æ/ ‰F €    b !r,,-       Æ‰! €    b !r       Æ ‰#" €    b !r../       â !r001       äO ‰'& €    " !r223       Æ� ‰)( €    ¢ !r45       Æ‰+$ €    â  j  ÿ„ â!rFF       ä/ ‰ @    b  k    â !r!       ä‰-* €    " !r""#       Æ ‰/, €    h ‰ @    b !r'&'       Æ ‰1. €    " !r)()       ÆO ‰30 €    ¢ !r+$+       Æ� ‰52 €    â !r       Æ/ ‰74      b !r-*-       Æ‰       " !r/,/       Æ‰!F @    b !r       Æ ‰# @    b !r1.1       Æ ‰%" @    " !r303       ÆO ‰ ' @    ¢ !r525       â� !r747       ä/ ‰$) @    b !r;        Æ‰&+ @    â !r!F!       Æ‰       â !r##       Æ ‰(- @    " !r%"%       Æ ‰*/ @    " !r '        ÆO ‰      ¨ ‰"1 @    b !r$)$       Æ/ ‰,3 @    b !r&+&       Æ� ‰07 €    è ‰.5 @    â !r(-(       Æ‰!      " !r*/*       Æ ‰#      " !r)       ÆO †	 9    â !r"1"       â!r,3,       ä/ † ;   â !r070       Æ� ‰       b !r9        âO !r.5.       ä ‰ %      ¢ !r+!       Æ‰$      â !r-#       Æ ‰&      ( †	 9    è ‰(      " !r        Æ/ ‰*      h † )    â !r%%        ÆO ‰!"      ¢ !r$       Æ� ‰#,      â !r&       Æ ‰'0 @    ( ‰.      b !r(       Æ†	 +    â	 !r*       Æ/ † -   è	 †	 %    â	 !r!"!       âO !r#,#       ä� †     â	 !r'0'       Æ †	 
    â	 !r.       Æ† 
   è	 †	     è	 †     è	 †	 !    è	 † #   è	 ‰ '      ( †	     â	 M™        € ê x   ÿàÿ â !r''        Ær   ÿàñ È rÿÿ   ÿä  ä z j  ÿ€ È z k    Ê †y '    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y       "   " ¹z  l      â Ey       € â ™x?    â y      !   b ¹z  F   
   â x      ÿ  â $rÿÿ   ÿ Ž â y      %   ¢ r# ÿ       â x      ÿ  â r
 ÿ       â $r;ÿÿ   ÿ Ž â x      ÿ  â r ÿ       â $rÿÿ   ÿ Ž â rF ÿ       Ä r ÿ       ä r  ÿ   pRð är  ÿ   pRò är    ÿ(Ž È/ r   ÿ� ä x   ÿ   äO xÿ     È |    pbô Ú G)  €    € ê y      '   " $rÿÿ   ÿ Ž â xþÿÿÿÿÀŽ â $rÿÿ   ÿ Ž â y      &   b x        ä r  €ÿàÿ ä r        ä x.   ÿ   Ä x   ÿ   â $z p  Ž â x/ÿ   .  Æ $x!   ÿ Ž ä 'z y   Ž È 'z s   Ž â/ zÿ z    Æ $z }  ÿŽ â zÿ t    â $z |  Ž ä$z |  Ž â x   ÿàÿ â $z v  Ž ä$z v  Ž ä $z w  ÿŽ Ä %v X  Ž Æ r   ÿàõ â %x   .Ž â x   ÿà÷ ä xÿ     â $z p  Ž â z, X  ÿàù â $rÿÿ   Ž â z  q      â %v X  Ž â x,,   ÿà÷ Ä x     â $x"   Ž âz-ÿ Y  „þ â $x   Ž âr ÿ       â $x   Ž âr   ÿŽ â$x    Ž â z q  ÿàÿ ä x.   ÿüŽ Ä rÿÿ   ÿä ä �y+,ýÿÿ ‘ ¨ �y<,þÿÿ ‘ ¨ �y$,ÿÿÿ ‘ è �yS,    ‘ ( �y8,   ‘ h �y%,   ‘ h �y(,   ‘ b r&   ÿàõ Æ �y5,   ‘ b r'/   ÿä â $tUÿ$   ÿ Ž ä �y*üÿÿ ‘ h �y9&ýÿÿ ‘ h �y6&þÿÿ ‘ h �y4&ÿÿÿ ‘ h �y:&   ‘ h �y)&   ‘ h �y7&    ‘ h �y&   ‘ b rD   ÿàõ Æ �yT&   ‘ b rÿ   ÿä â %vDD Z  U Ž È $x$   ÿŽ â rB!D   ÿàõ â �y&    ‘ f rEE   ÿàÿ Ê $xCE    â r0"   ÿà÷ â �yPD    • h �y'B   ™ h �yB   ™ b rÿ   ÿäÿ Æ �yRB   ™ b r2   ÿàõ â %v00 Z  U Ž ä �yQB   ™ d $x$   ÿŽ â rÿ   ÿä ä r>!0   ÿà÷ ä r11   ÿàÿ â %v22 Z  U Ž È $x$   ÿŽ â �yJ0    • b $x?1   Ž â r@!2   ÿàõ â $x33   Ž â rB   ÿà÷ ä/ �yI>   ™ d rA3   ÿä ä rÿ   ÿäÿ â �y>   ™ b rD   ÿàõ â %v0B Z  U Ž Ä �yH@   ™ b r=ÿ   ÿä â $x$   ÿŽ ä �yN>   ™ b rV!0   ÿà÷ Æ �yG@   ™ b r11   ÿàÿ â $x==$   ÿŽ ä �yL>   ™ h �yK2    • h �yO@   ™ b $xW1   Ž Æ �yM@   ™ b %v2D Z  U Ž â rZ    ÿà÷ ä �y?V   ™ b! $x33   =Ž â rX!2   ÿàõ ä �yV   ™ d rY3   ÿä Ä �yA2    • b r^   ÿàõ ä r[ÿ   ÿäÿ â�y>X   ™ b r\ÿ   ÿä â %v^^ Z  U Ž ä�y@0    • d %v2Z Z  U Ž Ä/ �yDV   ™ d $x[[$   ÿŽ ä �y=X   ™ b $x1\$   ÿŽ Æ �yBV   ™ b r\!^   ÿàõ â $x__   1Ž â r0!2   ÿà÷ â �yEX   ™ b r33[   ÿàÿ ä r]_   ÿä â �yCX   ™ d $x13   Ž Ä �yU\   ™ h �yV0   ™ h# �yW\   ™ h �yX0   ™ h �yY\   ™ h �yZ0   ™ h �y[^    • h �y\\   ™ h# �y22    • h �y`0   ™ b ‚x       Ä ‚x  ÿþý     â |a       â $~3ÿ   ÿ Ž â |]       ä/ |1       ä x+<v  +    ÈO x<$Tp  +    ä� x$+ww  ÿÀŽ ä x+ST  <    Èx<ÿ   +  ä x0+ˆˆˆˆÿÀŽ ä x+%v  8    äx<<ww  ÿÀŽ ä x0022ÿø� ä x((Tp  +    Ä tSa$    ät$]üúøô$    ä t33<    äx%0Tv  ÿÀŽ ä t<1üúøô<    ä x0ÿ   0  Ä x55T  (    ä r%S%   $    ä r$30   <    ä x<ÿ   5  ä x35ˆˆˆˆÿÀŽ ä |0       Ä x5<ww  ÿÀŽ ä x3322ÿø� â $~(ÿ   ÿ Ž â |8       ä xS+ww  ÿÀŽ ä t005    ät11üúøô5    Ä x96v  9    ä x53Tv  ÿÀŽ ä x3ÿ   3  ä t+8S    äx44Tp  9    ä t((üúøôS    Ä x8% d  $    äx$%1u  $    ä r103   1    ä x:)v  :    ä |%       ä x09ww  ÿÀŽ Ä x77T  4    ä r(+5   (    â $~+ÿ   ÿ Ž â x)Tp  :    ä t%0    ä x%ÿ   7  ä xTTT  )    Ä t0+üúøô0    â $~+ÿ   ÿ Ž â |6       ä |4       ä x%%ww  ÿÀŽ ä x::ww  ÿÀŽ ä |3       Ä x5ÿ   T  ä t66%    ät)4üúøô%    ä x77ˆˆˆˆÿÀŽ ä t4+:    ät%3üúøô:    Ä x:5ww  ÿÀŽ ä |5       ä x7722ÿø� ä x3TˆˆˆˆÿÀŽ ä t++:    ät:5üúøô:    Ä x57Tv  ÿÀŽ ä x3322ÿø� ä x7ÿ   7  ä x93Tv  ÿÀŽ ä r05   0    ä r)67   )    Ä x3ÿ   3  ä r *ÿ   pRö äx *ÿ   pRô â$x**€   ÿ Ž â x( d  1    äx1(1u  1    ä x &ÿ   pRø Ä r &ÿ   pRú ä x(0 d  )    ä x&&   ÿ   ä r:+3   :    â &r+8'   ÿ   âx**€  ÿÀŽ â &r'('   ÿ   â r%49   %    Ä x001u  )    ä x&&€  ÿÀŽ â &r+$   +   âx**@     € â &r'0   '   â x)% d  :    ä x&&@     € â &rR   +   âx**ÿ      â &rR)R   '   â x&&ÿ      â &r'1Q      â x%%1u  :    Ä xÿT  *    ä x&ÿT  &    â &rR%Q   R   ä 0r+ÿP    A   â  x   ?  A  â  x&&   ?  A  â ErR R       ä  r*+     A  â  r++&     A  â Er' '       Æ #r+R      â &r+8I   ÿ   ä&rI(I   ÿ   â #r*'      â &r'8H   ÿ   ä &rI0   I   ä &r*$G   '   ä &rI)N   I   Ä 0r'ÿJ    A   ä &rI%L   I   ä &r*O   *   â  r4'     A  â r'&'     A  â ErI I       â &r*1M   *   ä 0rKÿK    A   Ä #r'I      â &r+$   +   â Er* *       â &r'8?   ÿ   â  rK     A  â &r?(?   ÿ   ä &r'$   '   â#rF*   F   â &r?0   ?   ä &r8>   ÿ   Ä &r+N   +   ä &r?)D   ?   ä&r*$=      ä &r'D   '   ä &r+1L   +   ä 0rÿ@    A   Ä &r?%B   ?   ä&r*E   *   ä &r'1B   '   âEr+ +       â &r*1C   *   â  r6     A  â Er? ?       â 0rAÿA    A   â Er' '       â  r&     A  â #r4+      â Er* *       â  r4A     A  â #r?      â #r	6'   	   â &r8U   ÿ   Ä &rH(H   ÿ   ä&r>(>   ÿ   ä&rU(U   ÿ   ä &r8V   ÿ   ä&r'(V   ÿ   â x,,   ÿàõ â #r;4*   ;   â x    ÿàÿ â &rH0G   H   Ä &r>0=   >   ä &r($W      ä&r*0W   U   ä &r$X      ä&r'0X   '   â r-ÿ-   ÿä â &rH)O   H   â|    pbô â &r>)E   >   Ä &r(Y   (   ä&r*)Y   *   ä &rZ      ä&r')Z   '   ä 0r[ÿ[    A   ä &rH%M   H   Ä &r>%C   >   ä &r(1\   (   ä&r*%\   *   ä 0rÿ2    A   ä &r1`      ä&r'%`   '   â  r$[     A  âErH H       â  r     A  â Er> >       â  rK&K     A  âEr( (       â  rA&A     A  âEr* *       â  r[&[     A  âEr        â  r&     A  â Er' '       Ä x   ÿàù ä x   ÿà÷ â #rKH      â #r
A>   
   â #r#$(   #   â #r[*      â #r      â #r'      â x    ÿàÿ â $rÿÿ    â x""    ÿàÿ Ä x    ÿàÿ ä x    ÿàÿ ä x    ÿàÿ ä x      ÿàÿ ä rÿ   ÿäÿ â D)       À â Gy  `ëÿÿÿÿƒ ê Ay        € ê Ey      € â G‰  ð     € ê $x À  Ž Ê ˆs  ùÿ H   è ˆs €ùÿ H   è ˆs F úÿ H   è ˆs €úÿ H   è ˆs  ûÿ H   è ˆs €ûÿ H   è ˆs ; üÿ H   è ˆs 
€üÿ H   è ˆs 	 ýÿ H   è ˆs €ýÿ H   è ˆs # þÿ H   è ˆs €þÿ H   è ˆs  ÿÿ H   è ˆs €ÿÿ H   ä Ay        € ê {          ì M        € ê „y      H   b z  r  p`ð Ä z-  r      â „y €  H   ¢ r ÿ   p"p Æ „y €  H   â $x+-   ÿŽ Æ „y €   H   ( „y    H   h „y    H   ( „y    H   ( „y €  H   è „y    H   è „y €  H   â !r         Æ/ „y    H   â !r       ÆO ‰       " !r       Æ� „y €  H   â !r       Æ„y"    H   â !r       Æ„y €  H   â !r;;       Æ ‰      b !rF       Æ ‰      ¢ !r        Æ y      '   " !r	       Æ y       &   â !r       Æ ‰      " !r#       Æ ‰      b !r       Æ/ ‰;      b !r       ÆO ‰      ¢ !r
       â !r        â !r""       â ‰      ¢ !r$       â $z   x  Ž ä� ‰      ä $z  ~   Ž â !r       âx   ÿàÿ â ‰!      " !r       Äz  r  p`ò â ‰ €    " !r;       ä/ x    p"ò â ‰      b !r       âO 
  r  ÿàõ ä ‰#      ¢ !r       â ÿÿ   ÿä Ä ‰      b 
 j  ÿ„ â !r       ä� ‰%       â 
 k   Æ ‰'"      â !r!!       Æ ‰	      " !r       âr-    ÿˆ ä ‰)$      " !r       â/ r.ÿÿ   ÿä ä ‰      b !r##       ÆO ‰
      ¢ !r       Æ‰      b !r% %       Æ� ‰      â !r'"'       â !r	       ä ‰       " 	 +   ÿàû â !r)$)       ä‰ @    ä $ÿÿ   ÿŽ â !r       â/ ‰"      b r-    ÿ„ â !r
       ÄO ‰$!      ¢ z
 j  ÿˆ â !r       ä‰&#      " 
	 j  ÿŠ â !r       ä� ‰ €    â $x-   ÿŽ â rÿÿ   ÿä Ä ‰(%      b z j  ÿ„ â !r         â     ÿà÷ â ‰*'      " z k   ä ÿÿ   ÿäÿ â ‰ €    " 
 j  ÿ† â !r""       â/ z k  . â ‰ €    b 
 k  � â !r       â !r$!$       âO ‰,)      ¢ !r&#&       â-   ÿàÿ Ä ‰ €    " !r       â� 
		 k  � ä ‰ €    â !r(%(       â/    ÿàù ä ‰ €    b !r*'*       â +   ÿàÿ â $2ÿÿ   ÿ â !r       â ‰#  €    " / j  ÿˆ â !r       Ä/ ‰%" €    b $x-   ÿŽ â 0    ÿà÷ â !r,),       âO ‰'$ €    ¢ / k  2 â !r       ä‰)& €    " r    ÿàõ â !r       Ä� ‰ @    â r.ÿÿ   ÿä â !r       ä‰+( €    b z j  ÿ„ ä 1ÿÿ   ÿäÿ â ‰-* €    b !r# #       â 0 j  ÿ† Ä ‰! @    " !r%"%       â/ z k  . ä ‰ @    b !r'$'       âO 0 k  1� ä ‰ @    ¢ !r)&)       â   ÿàÿ Ä ‰/, €    " !r       â�     ÿàõ ä ‰ @    â !r+(+       Æ‰      b !r-*-       Æ ‰ @    b !r!!       Æ ‰ # @    " !r       Æ/ ‰% @    b !r       ÆO ‰"' @    ¢ !r/,/       â,ÿÿ   ÿä ä ‰$) @    " !r       Æ� ‰      â !r1       âx        ä ‰&+ @    b !r       â  j  ÿ„ â %v  j   Ž â ‰(- @    b !r #        â  k  , Ä ‰      " !r%       Æ/ ‰*/ @    b !r"'"       ÆO ‰      ¢ !r$)$       Æ‰       " !r3       Æ� ‰      â !r&+&       Æ‰!      b !r(-(       Æ ‰       b !r)       Æ ‰#      " !r*/*       Æ/ ‰%"      b !r+       ÆO ‰$      ¢ !r        Æ†	 1    â !r       Æ� † 3   â !r!!       Æ‰'&      â !r        Æ ‰(      " !r##       Æ ‰*      " !r%"%       Æ/ †	 )    â !r$       ÆO † +    è †	 
    è † 
   â !r'&'       Æ� †	 !    â !r(       Æ†     â !r*       Æ †	 #    è † %   è †	     è † '    è †	     â M™        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y       "   " ¹z  l      â Ey  P    € â ™x?    â y      !   b ¹z  F   
   â x      ÿ  â r ÿ       â y      %   ¢ $r.ÿÿ   ÿ Ž â x      ÿ  â x      ÿ  â r9 ÿ       â x      ÿ  â $r0ÿÿ   ÿ Ž â r  ÿ   pRð Ä r  ÿ   pRò â$x     Ž â/ x   ÿ   ÈO r   ÿ� È xÿ     È |    pbô Ú G)       € ê y      '   " $rÿÿ   ÿ Ž â xþÿÿÿÿÀŽ â $rÿÿ   ÿ Ž â y	      &   b $t&ÿ   ÿ Ž â r ÿ       ä z, q  ÿàÿ â 'z y   Ž â r  €ÿàÿ Æ 'z	 s   Ž â/ zÿ z    ä x   ÿ   â $z p  Ž â zÿ t    â $z |  Ž âx)   ÿüŽ â $z |  Ž â x   ÿàÿ â $z v  Ž âxÿ     â $z v  Ž Ä $z p  Ž ä %v$ X  &Ž È %x   Ž â x#$   ÿà÷ Æ $z! }  ÿŽ â z X  ÿàù â $z
	 w  ÿŽ â r$ÿ%   ÿäÿ â $x   ÿ Ž â x   ÿà÷ â $vÿ q  ÿ Ž â r!
!   ÿàõ â %v& X  &Ž â x+ÿ     ä zÿ Y  „þ â $x    Ž âr   ÿŽ â$x(   Ž âx+   +  â$x*   Ž â x-   ÿ   â $r?ÿÿ   ÿ Ä �y3ýÿÿ ‘ ¨  �y<þÿÿ ‘ ¨  �y%ÿÿÿ ‘ è  �y1    ‘ ( �y   ‘ h �y:   ‘ b r&)   ÿàõ Æ �y"   ‘ b r'   ÿä Æ �y/   ‘ h �y8ýÿÿ ‘ h �y;þÿÿ ‘ h �y2ÿÿÿ ‘ h �y6    ‘ h �y4   ‘ h �y5   ‘ b r!   ÿàõ Æ �y7   ‘ b r
,!   ÿà÷ â $tGÿ$   ÿ Ž â rÿ?   ÿä â �y>   ‘ d $rÿÿ   ?Ž ä %v Z  G Ž È $x$   ÿŽ â r-   ÿàõ â $x$   ÿŽ â/ r #       â $x		   Ž ä $rÿÿ   $ Ž ä %v

 Z  G Ž â r	+   ÿä â �yI    • h �yAüÿÿ ‘ b r-
   ÿàõ Ä r   ÿàÿ â �y=   ™ h �y@   ™ b $rÿÿ   & Ž â r '       â $x   + â r!   ÿàõ â�yE   ™ h �yD    ‘ b rL !   ÿà÷ Æ �yC   ™ b %v Z  G Ž Æ �yH   ™ b rÿ?   ÿä Æ �yB   ™ b $rMÿÿ   ?Ž ä $xO$   ÿŽ â �yF   ™ b %vL Z  G Ž â r-   ÿàõ ä/ rO   ÿàÿ â $xMM$   ÿŽ â r-   ÿà÷ â �yK   ™ b r+   ÿä â $x   MŽ Ä �yJ
    • d $x	   +Ž ä �yL   ™ b r(!   ÿàõ Æ �yP   ™ b r
*!   ÿà÷ ä/ rÿ?   ÿä â �yM   ™ d $rÿÿ   ?Ž ä �yN    • b $x$   ÿŽ Æ �yS   ™ b $x$   ÿŽ Æ �yQ    • b %v Z  G Ž Æ �yO   ™ h �yT   ™ b %v
 Z  G Ž â/ r
-   ÿàõ ä �yR   ™ b r   ÿàÿ â $x   Ž â r-   ÿà÷ â �yU   ™ b r+   ÿä Æ $x   +Ž â �y    • h �y
   ™ h �y   ™ h �yV
   ™ h �yG   ™ h �y	
   ™ h! �y   ™ h �yW
   ™ h �yX   ™ h �y    • b ‚x  ÿþý     Ä ‚x       â $~ÿ   ÿ Ž ä $~ÿ   ÿ Ž â x   ÿàõ È rÿ   ÿä ä x<<v  3    ÈO xZ%Tp  <    È� x11T  Z    Èx
ÿ   1  ä x11ˆˆˆˆÿÀŽ ä |%       ä x<<ww  ÿÀŽ ä |3       ä x
ww  ÿÀŽ Ä/ x1122ÿø� ä t
%<    ätüúøô<    ä t3    ät<üúøô    ä x:v      Äx1Tv  ÿÀŽ ä x""Tp      ä r
       â $~ÿ   ÿ Ž â x1ÿ   1  ä xww  ÿÀŽ ä x"/T  "    Ä r<1   <    ä t    ä|:       ä x"ˆˆˆˆÿÀŽ ä x;;v  8    ä t::üúøô    Ä x22ÿø� ä x2Tp  ;    ä x"ÿ   "  ä xTv  ÿÀŽ ä x6T      â $~%ÿ   ÿ Ž â |       Ä x""ww  ÿÀŽ ä r   :    ä xÿ     ä x45v  4    ä xˆˆˆˆÿÀŽ â $~8ÿ   ÿ Ž â t"    Ä |2       ä x;;ww  ÿÀŽ ä t"%üúøô"    ä |6       ä |:       ä x%ww  ÿÀŽ Ä x77Tp  4    ä x22ÿø� ä t2;    ät26üúøô;    ä xÿ     ä t/8%    Ä t6:üúøô%    ä x7>T  7    ä x%Tv  ÿÀŽ ä xÿ     ä r"   "    ä xÿ   7  Ä r%   2    ä x77ˆˆˆˆÿÀŽ ä r2/   6    â $~ÿ   ÿ Ž â |       ä x44ww  ÿÀŽ ä x7722ÿø� Ä |%       ä |/       ä xww  ÿÀŽ ä t4    ät4üúøô4    ä x7Tv  ÿÀŽ Ä x
 d  <    ät%%    ät6/üúøô    ä x7ÿ   7  ä r Aÿ   pRö äx Aÿ   pRô â$xAA€   ÿ Ž â x1u  <    Ä r6%7   6    â &r%
=   ÿ   â x Dÿ   pRø är Dÿ   pRú äxAA€  ÿÀŽ ä x d  "    äxDD   ÿ   â &r%@   %   â x1u  "    Ä x" d  2    ä xAA@     € ä xDD€  ÿÀŽ â &r%E   %   â r   4    â &r="=   ÿ   â x1u  2    â &r/
C   ÿ   â xAAÿ      Ä xDD@     € â &r1H   %   â x d  6    â &r@@   =   â x%ÿT  A    â &r/B   /   â xDDÿ      â &r@E   @   â x1u  6    â &r4F   /   â x/ÿT  D    â 0r2ÿI    A   â  x%%   ?  A  â &r@H   @   â Er1 1       â  x//   ?  A  â  r3%2     A  â &r4K   4   â Er@ @       â 0rJÿJ    A   â #r031   0   â  r22/     A  â &r1
L   ÿ   â Er4 4       â  r5%J     A  â #r2@      â &r1P   1   Ä &r2
M   ÿ   â#r54      â &r4"M   ÿ   ä &r1S   1   ä &r3O   2   ä&r4O   4   ä &r1T   1   Ä &r2R   3   ä0rNÿN    A   â Er1 1       â &rRR   4   ä &r4U   2   â  r2%N     A  â &rC"C   ÿ   Æ #r921   9   â &r1
   ÿ   ä &rL"L   ÿ   ä&r"   ÿ   ä &r

   ÿ   ä&r""   ÿ   â x    ÿàÿ â &r1V   1   Ä &rCB   C   ä&rLP   L   ä&rVV      ä &rG   
   ä&r"G   "   â |    pbô â &rFF   C   Ä &rLS   L   ä&r	   1   ä&rV	   V   ä&r      ä&r"   "   ä &rFK   F   Ä &rLT   L   ä0r6ÿQ    A   ä &rRU   R   ä 0rÿ    A   ä &rW      ä&rVW   V   Ä &rX      ä&r"X   "   ä 0rÿ    A   â ErF F       â  rJ/J     A  â r3/N     A  âErL L       â  r5/6     A  âEr4 4       â  r	/     A  â ErR R       â  r8%6     A  âEr        â  r%     A  âErV V       â  r%     A  âEr        â  r//     A  â Er" "       Ä x##   ÿàù ä x&&   ÿà÷ â #rJF      â #r3L      â #r84      â #r5R      â #r.   .   â #r	V      â #r      â #r/"      â x    ÿàÿ â $r$ÿÿ   $ â x      ÿàÿ Ä x((    ÿàÿ ä x**    ÿàÿ ä x,,    ÿàÿ ä r'ÿ'   ÿäÿ â D)       À â Gy  Àìÿÿÿÿƒ ê Ay        € ê Ey  ð     € â G‰  Ð     € ê $x €  Ž Ê ˆs 0 úÿ H   è ˆs €úÿ H   è ˆs  ûÿ H   è ˆs €ûÿ H   è ˆs 9 üÿ H   è ˆs €üÿ H   è ˆs  ýÿ H   è ˆs €ýÿ H   è ˆs . þÿ H   è ˆs €þÿ H   è ˆs  ÿÿ H   è ˆs €ÿÿ H   ä Ay        € ê {          ì M        € ê „y     H   " z  r  p`ð Ä z+  r      â „y  €   H   b r ÿ   p"p Æ „y    H   ¢ $x'+   ÿŽ ä$x)+   ÿŽ â „y €  H   è „y    H   ( „y €  H   h „y    H   h „y €  H   h „y	 €  H   h „y    H   è „y    H   â !r0       â !r        Ä/ „y! €  H   â !r       ÆO ‰       " !r       Æ� ‰      b !r9       â!r       ä‰      ¢ !r       â !r
       È ‰      â !r	       â !r         ä ‰	      " !r..       â !r       ä/ ‰       b !r       Æ ‰      " !r!!       Æ ‰      b !r       ÆO ‰      ¢ !r       Æ� ‰
      â x   ÿàÿ Æ ‰      â !r		       â z  r  p`ò ä ‰.      " !r        â/ x    p"ò ä ‰      b !r       Æy       &   " !r       â!r       äO y      '   ¢ !r
       Æ� ‰ €    â !r       Æ ‰      b !r..       Æ ‰	      " !r       Æ/ ‰ €    b $z   x  Ž Æ‰      ( ‰      " $z  ~   Ž ÆO ‰      ¢ !r
       â� r+    ÿˆ ä ‰      â !r       â '   ÿàõ ä ‰      b !r	       â ÿÿ   ÿä Ä ‰.      " !r       â/ 
  r  ÿà÷ ä ‰       b !r       â$rÿÿ   ÿ â z j  ÿˆ â ‰"!      " !r       â ÿÿ   ÿäÿ â !r       âO ‰
 @    ¢ 
 j  ÿ† â !r       Ä� ‰ €    â 
 j  ÿ„ â !r       ä‰ €    b 
 k  � â !r..       ä ‰	 @    " z k   â !r         Ä/ ‰ €    b 
 k   â !r"!"       ä‰ €    ( ‰ €    " !r
       âO r+    ÿ„ Ä ‰! €    ¢ !r       â� r&ÿÿ   ÿä ä ‰ €    â !r       âz
 j  ÿ„ ä‰. €    b !r	       â z k  & Ä ‰#  €    " !r       â/ 	+   ÿàÿ ä ‰%"      b !r       â )   ÿàù â !r       â ‰ @    " ( 	   ÿà÷ ä $ÿÿ   ÿ â !r!!       âO ‰ @    ¢ 
 j  ÿˆ â !r       Ä� ‰      â +ÿÿ   ÿäÿ â !r.       ä‰ @    b ( j  ÿ† â !r# #       ä ‰ @    " 
	 k   â !r%"%       Ä/ ‰ @    b ( k  +� Æ ‰"! @    b !r       â        ä ‰ @    " !r       ÆO ‰  @    ¢ !r+       â� '   ÿàÿ ä ‰      â !r       Æ‰$# @    b !r       Æ ‰&% €    " !r       â/ !r"!"       â $!ÿ   ÿ Ž â ‰      b !r       Æ‰      " !r         ÆO ‰      ¢ !r-       â�     ÿàõ ä‰      â %  j   Ž â!r$#$       â(ÿÿ   ÿä â ‰      b %  j  ! Ž â !r&%&       â  j  ÿ„ â‰      &  k  ( â ‰!       " !r       Æ/ ‰#"      b !r       Æ‰'& @    " !r       ÆO ‰%$      ¢ !r       Æ� †	 +    â !r       Æ† -   â !r       Æ †	     â !r! !       â !r#"#       ä/ †     â !r'&'       Æ†	     â !r%$%       ÆO †    è †	     è †     è †	 
!    è † 
#   è ‰'      ( †	 %    â M™        € ê x)   ÿàÿ â� !r''       Æ r    ÿàñ È r ÿÿ   ÿä  ä z j  ÿ€ È z k     Ê †y '    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y       "   " ¹z  l      â Ey  `    € â ™x?    â y      !   b ¹z  F   
   â r ÿ       â $rÿÿ   ÿ Ž â r ÿ       â $r ÿÿ   ÿ Ž â r4 ÿ       â x      ÿ  â $r<ÿÿ   ÿ Ž â r? ÿ       â $rÿÿ   ÿ Ž â r  ÿ   pRð Ä r  ÿ   pRò â$x     Ž Ê/ r   ÿ� È xÿ     È |    pbô Ú G)       € ê y      '   " xþÿÿÿÿÀŽ â $rÿÿ   ÿ Ž â r ÿ       â y      %   b r
  €ÿàÿ ä r        â y      &   ¢ x        â $x"
   ÿ Ž âx

   ÿ   Ä z q  ÿàÿ ä x#ÿ   "  ä xÿ   
  ä x"   ÿüŽ ä x
     äx
   ÿ   â 'z y   Ž È $x   ÿ Ž â/ zÿ z    â $z }  ÿŽ ä 'z s   Ž ÈO $z p  Ž â zÿ t    â $z |  Ž ä$z |  Ž â x   ÿàÿ â $z v  Ž ä$z v  Ž ä $z w  ÿŽ Ä %v X  Ž Æ r   ÿàù â %x   "Ž â x   ÿàõ Æ $z p  Ž â z$ X  ÿà÷ â $rÿÿ    ä %v X  Ž â x$$   ÿàõ Æ $vÿ q  ÿ Ž â z%ÿ Y  d~ â $rÿÿ   ÿ Ž â r        â $x   Ž âr   ÿŽ â$x   Ž ä $rÿÿ   ÿ ä �y.$ýÿÿ ‘ ¨ �y$þÿÿ ‘ ¨ �y$ÿÿÿ ‘ è �y6$    ‘ ( �y0$   ‘ h �y5$   ‘ h �y2$   ‘ h �y$   ‘ b r&   ÿàõ â $t,ÿ$   ÿ Ž Ä �y/üÿÿ ‘ b r'#   ÿä Ê �y:&ýÿÿ ‘ h �y3&þÿÿ ‘ h �y1&ÿÿÿ ‘ h �y9&    ‘ h �y7&   ‘ h �y>&   ‘ h �y;&   ‘ b r)   ÿàõ Æ �y&   ‘ b r
ÿ   ÿä â %v() Z  , Ž â rG   ÿà÷ Æ $x
$   ÿŽ ä $rÿÿ    Ž â r*(   ÿàõ â $x))   Ž â r
ÿ   ÿäÿ ä �y    ‘ b rK   ÿà÷ ä r+)   ÿä â $x!
$   ÿŽ Ä %vFG Z  , Ž ä �yB*   ™ d $r
ÿÿ   Ž â r-   ÿà÷ â $xGG   !Ž â �y*   ™ b r&F   ÿàõ â %vJK Z  , Ž ä �y*   ™ d! $x=
$   ÿŽ Ä �y8*   ™ h �y!(    • b r'G   ÿä Æ �y
F    • b $r*ÿÿ   Ž ä %v(- Z  , Ž â/ r-   ÿà÷ â �y&   ™ b rLJ   ÿàõ ä rKK=   ÿàÿ â $x+*$   ÿŽ â rHÿ   ÿäÿ â �yA&   ™ b rMK   ÿä â $x))   +Ž â r*(   ÿàõ â �y=&   ™ b %v,- Z  , Ž Æ �y@&   ™ b $x+)    Æ �yCL   ™ h �yI*   ™ b $x'H$   ÿŽ â r&,   ÿàõ ä �yH*   ™ d r--'   ÿàÿ Ä �yEJ    • f $x'-    â �yDL   ™ h �y((    • h �yK&   ™ h �yJ*   ™ h �y)&   ™ h! �yFL   ™ h �yGL   ™ h �y,,    • h �yL*   ™ h# �yM&   ™ h �y&&   ™ b ‚x       Ä ‚x  ÿþý     â |+       ä/ xv  .    ÈO xTp      ä� x*ww  ÿÀŽ ä x66T      â$~ÿ   ÿ Ž â t+*    ä x'ÿ   6  ä x+6ˆˆˆˆÿÀŽ Ä tüúøô*    â $~*ÿ   ÿ Ž â x55v  0    ä|.       ä x''ww  ÿÀŽ ä x++22ÿø� ä x22Tp  5    Ä t..'    ät-*üúøô'    ä x'+Tv  ÿÀŽ ä x+ÿ   +  ä x2T  2    ä r'       Ä r.+   -    ä x0ÿ   2  ä x-2ˆˆˆˆÿÀŽ â $~ÿ   ÿ Ž â |.       ä x00ww  ÿÀŽ ä |6       Ä x55ww  ÿÀŽ ä x--22ÿø� ä t+0    ät..üúøô0    ä t'65    ät**üúøô5    Ä x0-Tv  ÿÀŽ ä x-ÿ   -  ä x d      ä x:3v  :    ä x1u      ä r'0   *    Ä r*+-   .    ä x.1Tp  :    ä x::ww  ÿÀŽ ä |+       ä x99T  .    ä x>>v  7    Ä t.+üúøô:    ä x+ÿ   9  â $~0ÿ   ÿ Ž â x1;Tp  >    ä x++ww  ÿÀŽ ä |2       ä x1T  1    â $~3ÿ   ÿ Ž â t-0+    Ä x>ww  ÿÀŽ ä t02üúøô+    ä |5       ä x2ÿ   1  ä x99ˆˆˆˆÿÀŽ ä x11ˆˆˆˆÿÀŽ â $~'ÿ   ÿ Ž â t+3    Ä t5üúøô    ä x22ww  ÿÀŽ ä x9922ÿø� ä x5122ÿø� â $~7ÿ   ÿ Ž â t132    ä t'':    Ä x39Tv  ÿÀŽ ä x65Tv  ÿÀŽ ä x9ÿ   9  ä t27üúøô2    ä x5ÿ   5  ä r.'3   .    Ä r6+6       ä r--9   0    ä x' d  *    äx+1u  *    â $x*€   ÿ Ž â r /ÿ   pRö äx /ÿ   pRô Ä x//   ÿ   ä r515   2    ä x1. d  -    ä x//€  ÿÀŽ ä r ÿ   pRú ä x**€  ÿÀŽ Ä x ÿ   pRø â &rB   ÿ   âx.1u  -    â &rB1B   ÿ   â x//@     € ä x**@     € â &r      âx-6 d  5    â &rB   B   â x//ÿ      Ä x**ÿ      â &r'      âx61u  5    â &rB-   B   â xÿT  /    ä xÿT  *    â &rB8   B   ä &r*   ÿ   â  x   ?  A  â 0r!ÿ!    A   â  x   ?  A  â &r+8      â ErB B       â &r.A   *   â  r*!     A  â  r!!     A  â Er        â &r.'=   .   Ä #r!B      â &r!H   ÿ   ä &r.+@   .   â #r*      â &r1   ÿ   ä 0r
ÿ
    A   ä &rC   ÿ   ä &r!I   !   â Er. .       â &rC1C   ÿ   Ä &rH1H   ÿ   ä &r*A      â  r
     A  â &rK   ÿ   ä&r11K   ÿ   â x    ÿàÿ â &rD      ä &r!'J   !   Ä &rCD   C   ä&rHI   H   â #r?.   ?   â &r)      ä&r)   1   â |    pbô â &r'F      ä &r!+L   !   Ä &r*-=   *   ä&rF-F   C   ä0r(ÿ(    A   ä &rH-J   H   ä &r'M      ä&r-M      â Er! !       â &r*@   *   â  r(     A  â 0rEÿE    A   Ä &rFG   F   ä&rHL   H   ä &r+G      ä&r+&      ä&r&      ä 0r,ÿ,    A   â  r
     A  â r
E     A  â #r !       â Er* *       â  rEE     A  âEr        â  r(     A  âErF F       â  r,     A  â ErH H       â  r,     A  â Er        Ä Er        ä x   ÿà÷ ä x$$   ÿàý ä x   ÿàù â #r<*   <   â r%ÿ%   ÿä â #r	
   	   â #r4EF   4   â #rH      â #r      â #r      â x    ÿàÿ â $rÿÿ   Ž â x    ÿàÿ Ä x    ÿàÿ ä x    ÿàÿ ä rÿ   ÿä â D)       À â Gy  Ðîÿÿÿÿƒ ê Ay        € ê Ey  Ð     € â G‰  °     € ê $x  @  Ž Ê ˆs   ûÿ H   è ˆs  €ûÿ H   è ˆs  ? üÿ H   è ˆs  <€üÿ H   è ˆs  	 ýÿ H   è ˆs  4€ýÿ H   è ˆs    þÿ H   è ˆs  €þÿ H   è ˆs   ÿÿ H   è ˆs  €ÿÿ H   ä Ay        € ê {          ì M        € ê „y      H   " Ey  @    € æ „y €   H   h „y    H   ¢ !r         â !r       È/ ‰       " !r??       ÆO ‰      h ‰?      ¢ !r        Æ y       %   " !r       Æ/ ‰      b !r
?       ÆO ‰      ¨ ‰
      â $x     ÿ Ž â !r       È/ x    ÿàÿ äz   r  p`ð â !r       âO z  r  p`ò â ‰ €    " r ÿ   pVp  â!r
       â� x    p"ò â ‰ €    h ‰ €    ¬ ‰      &   ä $ÿ   ÿ Ž Ä       &   ( ‰      '   b !r       â !r       È/ ‰ @    " !r       ÆO ‰ @    h ‰ @    ¨       '   b $ x   Ž â!r       â !r       È/ ‰      " !r       ÆO ‰      b $Š x   Ž â� ˆ        â $
 ~  Ž â„y €  H   â $Š ~  Ž Æ ‰      ¢ !r       â %† j   Ž â !r       Æ/ %

 j   Ž â †‰     è † 
   â !r       âO G	  �     € ê y      &   ( y      '   b $z x   Ž È $z ~  Ž Ê/ z r  ÿàõ È rÿÿ   ÿä ä z j  ÿ„ È z k   Ê †y     è Ay        € ê !r<       â „y    H   â Ey  À    € æ ‰      ( „y €  H   è „y    H   è „y €  H   è „y    H   è „y €  H   â !r       Ê ‰      $ !r       Ê ‰ €    $ !r       Ê ‰ @    $ !r       Ê ‰      $ !r       â G™  °     € ê y      &   " ¹z  r      ä �x   ?àÿ â y      '   b $z x   Ž È $z ~  Ž Ê/ |   ÿàõ Ê $rÿÿ   ÿ â z j  ÿ„ È z k   Ê †y     è Ay        € ê !r	       â !r4       â !r        â !r       â !r	       â ‰      " !r       â z  r      ä ‰      h ‰      ¨ ‰
      è ‰	      ( ‰      b !r       â !r       È/ ‰      " !r       ÆO ‰      b !r

       Æ� ‰      ¢ !r	       Æ‰	
      â !r       Æ‰      ( ‰      b !r       Æ y      &   " !r       Æ/ ‰ €    b !r       ÆO ‰ €    ¢ !r	
	       Æ� ‰ €    â !r       Æ‰	 €    " !r       Æ‰
 €    h ‰ €    b $z x   Ž â !r       ä/ y      '   " !r       ÆO ‰ @    b !r       Æ� ‰ @    ¢ !r	       Æ‰ @    â !r
       Æ‰	 @    " !r       Æ ‰ @    h ‰ @    b $z ~  Ž â !r        È/ r   ÿ„ â!r
       âO $x   ÿŽ âr   ÿˆ â !r       ä� ‰
      " rÿÿ   ÿä â !r	       âz j  ÿ„ â ‰	       b rÿÿ   ÿä â !r       âz k   â ‰      ¢ ‚   ÿàõ â !r       â    ÿàÿ â ‰      ä $‚ÿÿ   ÿ â Š j  ÿ„ â ‰      "    ÿà÷ Ä Š k   â ‰      b ÿÿ   ÿäÿ ä  j  ÿ„ â !r
       â z j  ÿˆ â !r 	       â/ 	 k   ä †    â z k   â !r       ÄO †‰     â !r       Æ� †‰     â !r       Æ†     â !r       Æ†‰     â M™        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y       "   " ¹z  l      â Ey  �    € â ™x?    â y      !   b ¹z  F   
   â r ÿ       â $rÿÿ   ÿ Ž â r ÿ       â $rÿÿ   ÿ Ž â r
 ÿ       â $rÿÿ   ÿ Ž Ä $r-ÿÿ   ÿ Ž ä $r5ÿÿ   ÿ Ž â r  ÿ   pRð är  ÿ   pRò â$x     Ž Ê/ r   ÿ� È xÿ     È |    pbô Ú G)  `    € ê y      '   " xþÿÿÿÿÀŽ â $rÿÿ   ÿ Ž â r ÿ       â y      %   b z q  ÿàÿ â $x	   
Ž ä y      &   ¦ x	   ÿ   â$x		   ÿ Ž Æ xÿ     ä x	   ÿ   â 'z y   Ž È $x   ÿ Ž â/ zÿ z    â $z }  ÿŽ ä 'z s   Ž ÈO $z p  Ž â zÿ t    â $z |  Ž ä$z |  Ž â x   ÿàÿ â $z v  Ž ä$tÿ   ÿ Ž ä $z v  Ž Ä $z p  Ž ä %v X  Ž È %x   Ž â x   ÿàõ ä xÿ   	  â $z w  ÿŽ â z  X  ÿà÷ â %v X  Ž â x	     ä r   ÿàù â $rÿÿ    â x     ÿàõ Ä r        â $vÿ q  ÿ Ž â z!ÿ Y  d~ â $rÿÿ   ÿ â r ÿ       â $x   Ž âx   ÿüŽ â $x   Ž ä �y ýÿÿ ‘ ¨  �y, þÿÿ ‘ ¨  �y ÿÿÿ ‘ è  �y4     ‘ ( �y1    ‘ h �y2    ‘ b r$   ÿàõ Æ �y    ‘ b r%   ÿä Æ �y    ‘ h �y*$ýÿÿ ‘ h �y	$þÿÿ ‘ h �y+$ÿÿÿ ‘ h �y3$    ‘ b rA   ÿàõ Ä rC   ÿà÷ â �y$   ‘ b rÿ   ÿä Æ �y0$   ‘ b $t(ÿ$   ÿ Ž ä $x#$   ÿŽ â �y$   ‘ b $rÿÿ   Ž ä %vBC Z  ( Ž â�y>$   ‘ f %v@A Z  ( Ž â �y9üÿÿ ‘ f $xAA   #Ž â r"@   ÿàõ â $x%$   ÿŽ â/ r$        Ä �y8@    • b r#A   ÿä ä rCC%   ÿàÿ â $r%ÿÿ    Ž â rDB   ÿàõ â �y."   ™ h �y$    ‘ b $xEC    â r'   ÿàõ Ä �y/"   ™ b r)   ÿà÷ ä r6ÿ   ÿä â �y:D   ™ b %v$' Z  ( Ž Æ �y="   ™ b $r&ÿÿ   Ž ä $x'6$   ÿŽ â �y7D   ™ b %v() Z  ( Ž Æ �y<"   ™ b r%%'   ÿàÿ Æ �y;D   ™ h �y?D   ™ b $x#&$   ÿŽ â/ r"$   ÿàõ ä �y6B    • b $x))   #Ž â r#%   ÿä Ä �y$$    • b r&(   ÿà÷ Æ �y@"   ™ d $x')   Ž ä �yA"   ™ h �yB&   ™ h# �yC"   ™ h �yD&   ™ h �y%"   ™ h �yE&   ™ h �y((    • h �y&&   ™ b# ‚x       Ä ‚x  ÿþý     â |#       â $~'ÿ   ÿ Ž â/ x  @  ÿàõ È r!ÿ!   ÿä ä x@   ÿàÿ ä x,,v      ÈO xTp  ,    È� x"4T      Èx)ÿ   "  ä x""ˆˆˆˆÿÀŽ ä x4,ww  ÿÀŽ ä x,)ww  ÿÀŽ â $~ÿ   ÿ Ž â x""22ÿø� ä |       Ä t##,    ät,'üúøô,    ä x22v  1    äx'"Tv  ÿÀŽ ä t4    ätüúøô4    Ä x"ÿ   "  ä x4Tp  2    ä r'       â $~ÿ   ÿ Ž â r##"   ,    ä x,2ww  ÿÀŽ ä x4T  4    Ä x d  #    äx1u  #    ä t#,    ä x4ˆˆˆˆÿÀŽ ä x*	v  *    ä |       Ä x22ÿø� ä x2+Tp  *    ä t,üúøô,    ä x	Tv  ÿÀŽ ä x23T  2    ä r	#	   ,    â $~#ÿ   ÿ Ž â x**ww  ÿÀŽ Ä |+       ä x,ÿ   2  ä x22ˆˆˆˆÿÀŽ ä t##*    äx4ÿ   4  ä t*+üúøô*    Ä x+222ÿø� â $~ÿ   ÿ Ž â x20v      ä x44ww  ÿÀŽ ä |'       ä xTp  2    ä t"4    Ä t'üúøô4    â $~'ÿ   ÿ Ž â x+Tv  ÿÀŽ ä x>T      ä x,,ww  ÿÀŽ ä |)       ä x0ÿ     Ä r##   *    ä xÿ     â $~ÿ   ÿ Ž â t'',    äxˆˆˆˆÿÀŽ ä t,)üúøô,    ä x+ÿ   +  Ä r0"0       â $~"ÿ   ÿ Ž â x22ww  ÿÀŽ ä |       ä |*       ä xww  ÿÀŽ ä x22ÿø� Ä r,'+   ,    ä t2    äx ÿ   pRø är ÿ   pRú ät2üúøô2    ä x   ÿ   Ä t'"    ät**üúøô    ä xTv  ÿÀŽ ä xÿ     ä x# d  ,    ä r 9ÿ   pRö Ä x 9ÿ   pRô â$x99€   ÿ Ž â x€  ÿÀŽ ä r2   2    â &r.   ÿ   âx"#1u  ,    â &r..   ÿ   â r'   *    ä x99€  ÿÀŽ Ä x@     € â &r/      âx,	 d  0    â &r."/   .   â x*2 d      äx21u      â &r:   ÿ   â x99@     € ä xÿ      â &r,=      âx		1u  0    â &r.*=   .   â x99ÿ      â &r#7      â xÿT      â &r	<      âxÿT  9    â &r<<   .   Ä 0r8ÿ8    A   â  x   ?  A  â &r0,;   #   â Er< <       â  x   ?  A  â  r#8     A  â Er        â &r0	?   0   â  r.8     A  â 0r6ÿ6    A   â #r
#<   
   â &r#@   ÿ   Ä &r@@   ÿ   â #r5.   5   â Er0 0       â  r.6     A  â &r::   ÿ   ä &r#A   #   ä&r@"A   @   ä &rB   ÿ   Ä &rB   ÿ   â #r-.0   -   â &r:"7   :   ä &r.,C   #   ä&r@*C   @   ä &rD      ä&r"D      â |    pbô â &r:*;   :   Ä &r.	%   .   ä&r@%   @   ä0r$ÿ$    A   ä &r,E      ä&r**E      â Er. .       â &r:?   :   â Er@ @       â &r	&      â  r$     A  â &r*&   *   â  r$     A  â 0r(ÿ(    A   â Er: :       â #r.      â #r@      â Er        â  r6     A  â Er* *       â  r(     A  âx@  ÿà÷ â  r(     A  â x@  ÿàù â #r:      â #r      â #r*      â x@   ÿàÿ â $rÿÿ   Ž â x@   ÿàÿ Ä x@   ÿàÿ ä rÿ   ÿä â G©  pðÿÿÿÿƒ ê Ay        € ê Ey  °     € â G‰  �     € ê $x     Ž Ê ˆs  5 üÿ H   è ˆs  
€üÿ H   è ˆs  - ýÿ H   è ˆs  €ýÿ H   è ˆs   þÿ H   è ˆs  €þÿ H   è ˆs   ÿÿ H   è ˆs  €ÿÿ H   ä Ay        € ê {          ì M        € ê „y    H   h „y    H   ¨ „y      H   ( „y	 €   H   è „y    H   ( „y    H   h „y    H   h „y €  H   h „y €  H   h „y €  H   h „y!    H   b !r       Æ/ „y    H   b !r       ÆO „y    H   ¢ !r  5       Æ „y €  H   " !r		
       Æ� „y €  H   â !r       Æ„y €  H   " !r-       Æ„y#    H   b !r         Æ „y  	  H   b !r       Æ „y €	  H   b !r	       Æ „y  
  H   b !r       Æ „y €
  H   b !r!       Æ „y €  H   â !r       Æ/ „y €  H   â !r         ÆO „y €  H   â !r       â !r       ä� ‰       " !r       Æ‰      b !r#       â!r       È ‰      ¢ !r       Æ ‰	      â !r       Æ ‰      " !r       Æ ‰      b !r        â !r       ä ‰      " !r       â/ !r       ä ‰       d !r       â !r       âO ‰      ¢ !r		       Æ� ‰      â !r       Æ‰	      " !r       Æ‰      b !r       Æ ‰      " !r         Æ/ ‰
      h ‰      b !r       âO !r       ä� ‰  €    ¢ !r	       Æy      %   â !r       Æ‰ €    " !r       Æ ‰	 €    " !r

       Æ/ ‰ €    b !r       Æ ‰ €    h ‰
 €    b !r        ÆO ‰ €    ¢ x   ÿ   Æ� ‰      â !r       âz  r  p`ð ä y      &   " !r		       â r ÿ   pVp  ä y      '   " !r       â/ !r       Ä‰ @    â !r
       Æ ‰ @    â !r       ÆO ‰
	 @    b !r       Æ� ‰ @    ¢ $z  x  Ž Æ‰ @    â x   ÿàÿ â $z  ~   Ž ä ‰ @    " z  r  p`ò Æ ‰ @    " Š  r  ÿàõ ä x    p"ò â ‰ €    d $‚ÿÿ   ÿ â Š j  ÿ„ â !r	
       Æ/ Š k   â !r       â !r       â !r       âO z  r      â !r       â� ‰	      b #        â !r       ä ‰      " r    ÿ„ â$x   ÿŽ â !r       â‰      ¢    ÿàÿ â $rÿÿ   ÿ â !r       â‰      â z j  ÿ„ Ä ‚    ÿà÷ â‰      " z k   ä ‚
ÿÿ   ÿäÿ â ‰      b     ÿàõ â $„ÿ   ÿ Ž â Š j  ÿ† â‰ @    b !r	       â/ $ÿÿ   ÿ â Š k  
� â ‰      b 
 j  ÿ„ â %†  j   Ž â !r!       â !r       âO  k   â %  j  # Ž â †‰     â !r       Æ� † !   â !r       Æ†‰     â !r       Æ† 
    â !r       Æ †‰     â !r       Æ/ †    è ‰      h  †‰     â M™        € ê x   ÿàÿ â !r       Æ/ r    ÿàñ È r ÿÿ   ÿä  ä z j  ÿ€ È z k     Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      "   " ¹z  l      â Ey       € â ™x?    â y       !   b ¹z  F   
   â $rÿÿ   ÿ Ž â r ÿ       â y      &   ¢ $r
ÿÿ   ÿ Ž â r ÿ       â $r.ÿÿ   ÿ Ž â y      '   â $rÿÿ   ÿ Ž â r ÿ   pRð Ä r ÿ   pRò â$x     Ž Ê/ r   ÿ� È xÿ     â $À    Ž Æ |    pbô Ú G)  `    € ê y	      %   "Î $rÿÿ   ÿ Ž â xþÿÿÿÿÀŽ â $rÿÿ   ÿ Ž â z q  ÿàÿ â 'z y   Ž â r  €ÿàÿ Æ 'z s   Ž â zÿ z    Æ $x    ÿ Ž â zÿ t    â $z |  Ž ä $tÿ   ÿ Ž â x!ÿ      â $z v  Ž â x    ÿüŽ â $z w  ÿŽ ä $x		   ÿ Ž È $z	 p  Ž È $z |  Ž â x	   ÿàÿ Æ $z	 v  Ž ä $z p  Ž ä %v	 X  Ž È $z }  ÿŽ â x   ÿàù â %x	    Ž Æ r   ÿàõ â $x   ÿ Ž â z" X  ÿà÷ â $rÿÿ    ä %v X  Ž â xÿ     ä x""   ÿàù â $vÿ q  ÿ Ž â x     â$rÿÿ   ÿ Ž â r        â $rÿÿ   ÿ â z#ÿ Y  	d~ â $x   ÿ Ž â r   ÿŽ Æ �y"ýÿÿ ‘ ¨  �y/"þÿÿ ‘ ¨  �y"ÿÿÿ ‘ è  �y1"    ‘ ( �y"   ‘ h �y0"   ‘ h �y"   ‘ h �y,"   ‘ b r(   ÿàõ Ä r$   ÿà÷ â �y>üÿÿ ‘ d $x)!    Ê �y(ýÿÿ ‘ h �y4(þÿÿ ‘ b r&   ÿàõ Æ �y2(ÿÿÿ ‘ h �y6(   ‘ h �y-(   ‘ b x5 $       â $r%ÿÿ    Ä �y3(    ‘ b $rÿÿ   Ž Æ �y	(   ‘ b $x+%$   ÿŽ â r*   ÿàù â %v$$ Z  5 Ž â�y(   ‘ f $x$   ÿŽ ä %v&& Z  5 Ž Æ r%%   ÿàÿ â $rÿÿ    â rD&   ÿàõ â $x''   +Ž ä %v** Z  5 Ž â rB$   ÿà÷ ä�y$$    • b $x$   ÿŽ ä $xE'    âr(*   ÿàõ â �y&&    • b r++   ÿàÿ â $xC%   Ž Ä �y5D   ™ d $x)+    ä �y7B   ™ b $rÿÿ    Ž Æ �y?(   ™ h �y8D   ™ h �y9B   ™ h �y@(   ™ h �y:D   ™ h �y;B   ™ h �yA(   ™ h �y    ‘ h �y<D   ™ h �y=B   ™ h �y**    • h �yD(   ™ b ‚x       Ä ‚x  ÿþý     â $~Bÿ   ÿ Ž ä/ $~'ÿ   ÿ Ž â |)       â $~+ÿ   ÿ Ž â x""@  ÿàõ È r#ÿ#   ÿä ä x@   ÿàÿ ä x@  ÿàù ä x@   ÿàÿ ä x@   ÿàÿ â $rÿÿ    â x//v      ÈO xTp  /    È� x11T      Èxÿ   1  ä x11ˆˆˆˆÿÀŽ ä x00v      äx%/ww  ÿÀŽ ä x1122ÿø� â $~ÿ   ÿ Ž â xTp  0    Ä x(ww  ÿÀŽ ä tB%    ät%)üúøô%    ä x1Tv  ÿÀŽ ä x,T      ä t(    Ä t''üúøô(    ä x1ÿ   1  ä r%   %    ä xÿ     â $~ÿ   ÿ Ž â xˆˆˆˆÿÀŽ ä r1   '    â $~'ÿ   ÿ Ž â x00ww  ÿÀŽ Ä |)       ä xww  ÿÀŽ ä x22ÿø� ä t0    ät0'üúøô0    ä t()    Ä t)+üúøô    ä x'Tv  ÿÀŽ ä x+ÿ     ä x4v      ä x% d      äx%1u      Ä r'   0    ä r(+   )    â $~(ÿ   ÿ Ž â x22Tp      ä x6-v  6    ä xww  ÿÀŽ ä |0       Ä x33T  2    ä x,	Tp  6    ä t	(    ät(0üúøô    ä xÿ   3  ä x%T  ,    â $~ÿ   ÿ Ž â x0ww  ÿÀŽ â $~'ÿ   ÿ Ž â x33ˆˆˆˆÿÀŽ Ä t0    ä|)       ä x,6ww  ÿÀŽ ä t0'üúøô0    ä x'ÿ   %  ä x%%ˆˆˆˆÿÀŽ Ä x3322ÿø� ä t),    ä x)'ww  ÿÀŽ ä x'%22ÿø� ä x%3Tv  ÿÀŽ â $~+ÿ   ÿ Ž â x3ÿ   3  â $~2ÿ   ÿ Ž â |4       Ä r		%   (    ä r03   0    ä t,+üúøô,    ä x( d      ät+2)    ä x1u      Ä t24üúøô)    ä x	 d  0    â&r5   ÿ   âx)'Tv  ÿÀŽ ä x		1u  0    â &r07   ÿ   âx'ÿ   '  â &r?   ÿ   ä &r8      â r2+'   2    â &r09   0   Ä $x'>€   ÿ Ž ä &r@      â r)   ,    â &r(:      ä&r0(;   0   â r >ÿ   pRö â &r((A      â x''€  ÿÀŽ â $x€   ÿ Ž â x, d  2    â&r55   ÿ   âx1u  2    â &r27   ÿ   â x >ÿ   pRô â &r?   ÿ   â x''@     € Ä r ÿ   pRö ä x€  ÿÀŽ â &r5	8   5   â x''ÿ      â &r2	9   2   â x ÿ   pRô â &r	@      â x	@     € ä x'ÿT  '    â &r%<      â x		ÿ      â &r=   0   â |    pbô â &r5,:   5   â x''   ?  A  â &r2,;   2   âx	ÿT  	    â &r,A      Ä 0r$ÿ$    A   â Er        â &r<<   5   ä&r==   2   ä&rD      â  r'$     A  â 0r&ÿ&    A   â  x		   ?  A  â &r(D   (   Ä 0r*ÿ*    A   â #r.   .   â Er% %       â  r'&     A  â Er( (       â  r&	     A  â Er< <       â  r	$     A  â Er= =       â  r''*     A  â Er        â  r		*     A  â x@  ÿà÷ â #r%      â #r'(      â #r<      â #r
=   
   â #r	      â rÿ   ÿäÿ â G©  @òÿÿÿÿƒ ê Ay        € êÏ ˆ  ýÿ H   è ˆ €ýÿ H   è ˆ . þÿ H   è ˆ 
€þÿ H   è ˆ  ÿÿ H   è ˆ €ÿÿ H   è {          ì M        € ê „y      H   " z  r      Æ „y     H   h „y	  €   H   ¨ „y     H   è „y  €  H   ( „y     H   h „y     H   h „y  €  H   h „y     H   h „y  €  H   h „y  €  H   b !r       Æ „y     H   " !r       Æ/ „y  €  H   b !r		       ÆO „y     H   ¢ !r.       Æ� „y  €  H   â !r
       Æ„y  €  H   â !r       Æ„y     H   â !r       Æ „y  €  H   â !r	       â !r       ä ‰      " !r       â !r       â !r       â !r       â/ !r       ÈO ‰      " !r       Æ� ‰      b !r       â!r       ä ‰	      ¢ !r       Æ ‰      â !r       Æ ‰      ( ‰      b !r       Æ y      %   " !r       Æ/ ‰      b !r		       ÆO ‰      ¢ !r       Æ� ‰	      â !r       Æ‰ €    " !r       Æ‰
      h ‰      b $x   ÿ Ž â !r       Æ/ $z x  Ž â z  r  p`ð â !r       âO ‰ €    $ r  ÿ   pVp  â $z ~  Ž â !r	       â� ‰ €    b !r       Æ‰	 €    ¢ !r

       Æ‰ @    â !r       Æ ‰
 €    ( ‰ €    b !r       â !r       È/ ‰ @    " !r		       ÆO ‰ @    b !r       Æ� ‰	 @    ¢ !r
       âx
   ÿàÿ ä ‰      â !r       âz 
 r  p`ò ä ‰ @    $ x     p"ò Ä ‰ @    b !r       â !r
       È/ ‰       " Š r  ÿàõ â !r	       ÆO    ÿàÿ â $‚ÿÿ   ÿ â Š j  ÿ„ â!r       â� r	   ÿ† â $tÿ   ÿ Ž â Š k   â !r       â   ÿàõ â ‰
      b rÿÿ   ÿäÿ â !r       âz	 j  ÿ† â ‰      ¢ $ÿÿ   ÿ â  j  ÿ„ â %v j   Ž â ‰      â !r        â z	 k  � Ä ‰      " 	 k   Æ †‰     è	 †    â	 !r
       â/ !r       ÈO †‰     â	 !r       Æ� †     â	 !r       Æ †‰     â	 M™        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      "   " ¹z  l      â Ey  Ð    € â ™x?    â y       !   b ¹z  F   
   â $rÿÿ   ÿ Ž ä y      %   ¢ $r	ÿÿ   ÿ Ž ä $rÿÿ   ÿ Ž Ä $rÿÿ   ÿ Ž â r ÿ   pRð är ÿ   pRò â$x     Ž ä/ $x   ÿ Ž ÆO r   ÿ� È xÿ     â $€    Ž Æ |    pbô Ú G)  °    € ê y      '   " $rÿÿ   ÿ Ž â xþÿÿÿÿÀŽ â $r
ÿÿ   ÿ Ž â y      &   b $tÿ   ÿ Ž ä $x   
Ž È $x   ÿ Ž Ê xÿ     â 'z y   Ž È 'z s  
 Ž â/ zÿ z    Æ $z p  Ž â zÿ t    â $z |  Ž ä$z |  Ž â x   ÿàÿ â $z v  Ž ä$z v  Ž â z q  ÿàÿ â $z }  ÿŽ Ä %v X  Ž È %x   Ž â x   ÿàù Æ $z
 w  ÿŽ â z X  ÿà÷ â $z p  Ž ä $x   ÿ Ž â r
   ÿàõ â %v X  Ž Æ xÿ     â $rÿÿ    â x   ÿàù â $rÿÿ    Ž â x     â$rÿÿ   ÿ Ž â zÿ Y  d~ â $x   ÿ Ž â x   ÿüŽ ä r
ÿÿ   ÿä Ä �y"ýÿÿ ‘ ¨  �y)þÿÿ ‘ ¨  �yÿÿÿ ‘ è  �y(    ‘ ( �y!   ‘ h �y,   ‘ h �y   ‘ b r   ÿàõ Æ �y%   ‘ d $x    Ê �y'ýÿÿ ‘ h �y&þÿÿ ‘ h �y*ÿÿÿ ‘ h �y   ‘ h �y    ‘ h �y    ‘ h �y#   ‘ h �y   ‘ b r2   ÿàõ â $t-ÿ$   ÿ Ž È $rÿÿ   
 ä %v22 Z  - Ž È $x$   ÿŽ â r   ÿà÷ ä r42   ÿàõ â $x33   Ž ä $r+ÿÿ   
Ž ä $x53    â �y22    • b $r.ÿÿ    Ž Ä $r/ÿÿ    Ž â �y$4   ™ b %v Z  - Ž Æ �y-4   ™ b $x+$   ÿŽ ä/ $r0ÿÿ    Ž â �y+.üÿÿ ‘ b $r1ÿÿ    Ž â r   ÿàõ â $x   Ž Æ �y00    ‘ b $x    Æ �y.4   ™ h# �y/4   ™ h �y1   ™ h �y3   ™ h �y4   ™ h! �y    • h �y   ™ b! ‚x       Ä ‚x  ÿþý     ä $~ÿ   ÿ Ž â x@  ÿàõ Ê $rÿÿ    â x@   ÿàÿ ä x@  ÿàù ä x@   ÿàÿ Æ $rÿÿ    â x))v  "    ÈO x5Tp  )    È� x(T  5    â$~ÿ   ÿ Ž Æ x"ÿ     ä xˆˆˆˆÿÀŽ ä x""ww  ÿÀŽ ä x!,v  !    â$~5ÿ   ÿ Ž â x22ÿø� ä |(       Ä x))ww  ÿÀŽ ä t"    ätüúøô"    ä xTp  !    ä x"Tv  ÿÀŽ ä xÿ     Ä t(()    ät55üúøô)    ä x%%T      ä r       â $~ÿ   ÿ Ž â r"("   5    ä x(ÿ   %  Ä x)!ww  ÿÀŽ ä x,%ˆˆˆˆÿÀŽ â $~ÿ   ÿ Ž â x((ww  ÿÀŽ â $~%ÿ   ÿ Ž â t!)    â $~ÿ   ÿ Ž â x,22ÿø� ä x'&v  '    Ä t(    ät%%üúøô(    ä tüúøô)    ä x(ÿ     ä x)Tv  ÿÀŽ ä x**Tp  '    Ä x  v      â $~ÿ   ÿ Ž â x" d      ä r%(   %    ä x""1u      ä r!)       â $~!ÿ   ÿ Ž â x'ww  ÿÀŽ Ä xT  *    ä x&#Tp       ä t    äxÿ     ä t!üúøô    ä x!T  &    â $~&ÿ   ÿ Ž â xww  ÿÀŽ â $~(ÿ   ÿ Ž â x'ÿ   !  Ä xˆˆˆˆÿÀŽ â $~#ÿ   ÿ Ž â t&&    ät(üúøô    ä x!!ˆˆˆˆÿÀŽ â $~)ÿ   ÿ Ž â x('ww  ÿÀŽ ä x'22ÿø� â $~,ÿ   ÿ Ž â |*       Ä x ww  ÿÀŽ ä x!!22ÿø� ä t#(    äx#'Tv  ÿÀŽ ä t()üúøô(    ä x'ÿ   '  Ä t *    äx)!Tv  ÿÀŽ ä t,üúøô    ä x!ÿ   !  ä r#       ä r&'       Ä r  )       ä r!!   (    ä x d      äx1u      ä x  d  !    â&r$   ÿ   â x  1u  !    â $x!+€   ÿ Ž Ä &r$$   ÿ   â x d  %    â &r"-      âr +ÿ   pRö â &r$-   $   â x!!€  ÿÀŽ â $x#0€   ÿ Ž â x1u  %    â &r.      âx +ÿ   pRô â &r$.   $   â x!!@     € Ä r 0ÿ   pRö ä x##€  ÿÀŽ â &r/      âx!!ÿ      â &r/ /   $   â x 0ÿ   pRô â &r$1   ÿ   âx##@     € â &r1   ÿ   ä &r$"3   $   âx##ÿ      â &r3      â x!ÿT  !    â &r$4   $   â|    pbô â &r4      â x#ÿT  #    â 0r2ÿ2    A   â  x!!   ?  A  â &r$   $   Ä &r       â  x##   ?  A  â 0rÿ    A   â  r!2     A  âEr        â  r22#     A  â x@  ÿà÷ â  r!     A  â Er$ $       â  r##     A  â Er/ /       Ä Er        â #r      â #r	$   	   â #r2/      â $rÿÿ   Ž â #r#      â G©  Ðóÿÿÿÿƒ ê Ay        € ê ˆ  þÿ H   è ˆ €þÿ H   è ˆ 	 ÿÿ H   è ˆ €ÿÿ H   è {          ì M        € ê „y      H   " z  r  p`ð Æ „y  €   H   b r  ÿ   p"p Æ „y     H   ¨ „y     H   è „y  €  H   ( „y     H   h „y     H   h „y  €  H   h „y     H   h „y  €  H   h „y  €  H   b !r       Æ „y  €  H   " !r       â/ !r	       âO !r       ä� !r       â!r       â!r       ä !r       Æ ‰      b !r       Æ ‰      ¢ !r       Æ ‰
      â !r       È !r       Ê ‰      " !r       â/ !r       ÈO ‰      b !r

       Æ� ‰      ¨ ‰	
      â !r       Æ y      &   ( ‰      " !r       â/ !r       ÈO ‰ €    b !r
	       Æ� ‰	 €    ¨ ‰ €    â !r       Ê‰ €    " !r       Æ/ y      '   b !r		       âO !r
       È� ‰	 @    ¢ x   ÿàÿ â $z x  Ž ä ‰ @    " z  r  p`ò â $tÿ   ÿ Ž ä ‰
 @    â x     p"ò â !r       â$z ~  Ž Ê/ 
  r  ÿàõ â !r	       ÈO $ÿÿ   ÿ â !r       â 
  j  ÿ„ â ‰      " !r
       Æ� ‰      h ‰
 @    ¨ ‰	      â !r       â 
  k   â !r       â/ %v j   Ž â !r

       ÈO †	     â !r	       Æ� †    è ‰	
      h  †	     â M™        € ê ¹z  r      â !r	
	       â/ �x   ?àÿ Ì |   ÿàñ Ê $r ÿÿ   ÿ  â z j  ÿ€ È z k     Ê †y 	    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y       &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž ä $rÿÿ     Ž È % \   Ž Ê �	    ™ b $ÿÿ   ÿ Ž â ¹z  l      â Ey  °    € â y	      "   b '
 m   Ž â ™x?    ä y
      !   ¢ $rÿÿ   ÿ Ž â 
ÿ n    Ê $ÿÿ   
Ž È $
 o  Ž ä $‚ÿÿ   ÿ Ž È 'Š s   Ž â r 	ÿ   pRò ä/r 	ÿ   pRô â$x	    
Ž ÊO r   ÿ� È xÿ     È |    pbö ä Šÿ t    Ö G9  €    € ê y      '   " $rÿÿ   ÿ Ž â xþÿÿÿÿÀŽ â $tÿ   ÿ Ž â y      %   b $rÿÿ   ÿ Ž ä $x   
Ž È $x   ÿ Ž ä$x   ÿ Ž ä 'z y   Ž Ê z	ÿ z    Ê $z	 |  Ž È $z v  Ž âxÿ     Æ $z	 p  Ž ä/ $z w  ÿŽ ä %x	   Ž È %v	 X  Ž â z X  ÿà÷ Æ $z }  ÿŽ â x   ÿàñ ä x   ÿàù ä r   ÿàû â $r	ÿÿ   	  â xÿ     ä zÿ Y  d~ â $rÿÿ   ÿŽ â x     Ä $rÿÿ    Ž â �yøÿÿ ‘ ¨ �yýÿÿ ‘ è �yþÿÿ ‘ è �yÿÿÿ ‘ ( �y    ‘ h �y   ‘ ¨ �y   ‘ ¨ �y   ‘ ¢ r   ÿàñ â $tÿ$   ÿ Ž Ä �y   ‘ ¤ $rÿÿ     ä %v Z   Ž È $x$   ÿŽ â r   ÿ€ Æ $x   Ž È $x     â �y    • ¨  �y   ™ ¨ �y   ™ ¨ �y   ™ ¨ �y   ™ ¢ ‚x       Ä ‚x  ÿþý     ä $~ÿ   ÿ Ž ä $~ ÿ   ÿ Ž â x@   ÿàÿ ä xv      È� xTp      â$~ÿ   ÿ Ž Æ xT      äxww  ÿÀŽ ä xÿ     ä xˆˆˆˆÿÀŽ ä xv      äO t    Ä/ tüúøô    â $~ÿ   ÿ Ž â x22ÿø� ä xww  ÿÀŽ ä xTp      ä xTv  ÿÀŽ ä t    Ä xT      ä t üúøô    ä xÿ     ä r       â $~ÿ   ÿ Ž â xÿ     ä r       â $~ÿ   ÿ Ž â xww  ÿÀŽ â $~ÿ   ÿ Ž â xˆˆˆˆÿÀŽ â $~ÿ   ÿ Ž â xww  ÿÀŽ Ä t    äx22ÿø� ä tüúøô    ä t    ätüúøô    ä xTv  ÿÀŽ Ä xÿ     ä x d      äx1u      â $x€   ÿ Ž âr       ä r ÿ   pRö ä x€  ÿÀŽ â &r   ÿ   â r       Ä x ÿ   pRð ä x@     € â &r      â x d      äxÿ       ä x1u      â &r      â |    pbð Ä xÿT      â &r      ä 0rÿ    A   ä  x   ?  A  â x@  ÿà÷ ä x@  ÿàù â  r     A  â Er        â $r	ÿÿ   	Ž Ä $rÿÿ    ä #r      â G‰  @úÿÿÿÿƒ ê Ay        € ê ˆ €ÿÿ H   è {          ì M)        € ê „y
     H   " r 
ÿ   pRð Æ „y
 €   H   h „y
    H   ¢!r       È !r       È/ !r       ÊO ‰      $ !r       Ê ‰      $ !r       Ê ‰ €    $ !r       Ê ‰ @    $ !r       Ê ‰      b  M	        € ê y      %   ¨ y      '   â $z   x  Ž äO $tÿ   ÿ Ž ä $z  ~   Ž â� !r       Æ/ %v  j   Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž È % \   Ž Ê �	    ™ b Ey  P    € â $rÿÿ   ÿ Ž ä y       "   b $rÿÿ   ÿ Ž Æ y      !   b $ÿÿ   ÿ Ž ä $ÿÿ    Ž È '
 m   Ž È $‚ÿÿ   ÿ Ž â 
ÿ n    â $‚ÿÿ    Ž È $ÿÿ   
Ž ä 'Š s   Ž â rÿ    ÿü‚ Æ/ $
 o  Ž ä $x     Ž Ê r   ÿ� ä Šÿ t    â G       € ê z ÿ `  pPô â $rÿÿ   ÿ Ž â z ÿ ^  pPò ä z ÿ b  p ö ä z ÿ a   Sô ä z ÿ _  Sò Ä z ÿ c  0't Ö       '   "        ä¢       â ©
      '   d $ x  ÿŽ ä 	      %   ¢ $ª x  ÿŽ Æ ©      %   â $ ~  ÿŽ ä $ª

 ~  ÿŽ Æ/    	àë äO	ÿ   	  ä ¢

   àÇ ä�¨ÿ     ä 	ÿÿ   	Äþ ä ¢
   ÿàû Ä ¢ÿÿ   „þ ä    ÿàý ä ª b  ÿˆ â $¢ÿÿ   Ž â  ^  ÿ† â $
ÿÿ   	 Æ ª c   â $rÿÿ   ÿ Ž â  _  
� Ê �     h �©     d Ay        € ê ¹z  l      â xÿ     â ™x?    â Ey  €    € â $rÿÿ   ÿ Ž ä $‚ÿÿ    Ž Ä |    pbò â $rÿÿ   ÿ Ž Ø G       € ê y      '   " $rÿÿ   ÿ Ž â xþÿÿÿÿÀŽ â $tÿ   ÿ Ž â y      %   b $z
 w  ÿŽ â z ÿ `  pPð â $x   
Ž ä $rÿÿ   ÿ Ž â z ÿ a   Sð â $rÿÿ   ÿ Ž Ä $x   ÿ Ž ä 'z y   Ž È $z }  ÿŽ â z	ÿ z    â $x   ÿ Ž Æ r
   ÿàù â $z	 |  Ž È $z v  Ž âxÿ     â $rÿÿ   ÿ ä $z p  Ž È/ %x   Ž È %v X  Ž â z X  ÿàõ Æ %v `  Ž â x
   ÿàó ä x   ÿà÷ Æ $rÿÿ   Ž  â zÿ Y  Dþ Ê �yýÿÿ ‘ ¨ �yþÿÿ ‘ ¨ �yÿÿÿ ‘ è �y    ‘ ( �y   ‘ h �y   ‘ h �y   ‘ b r   ÿàó Æ �y    ‘ b $tÿ$   ÿ Ž ä $rÿÿ   Ž  â �y#
üÿÿ ‘ b %v Z   Ž È $x	$   ÿŽ È $x   	Ž È %x   Ž ä �y    • h �y   ™ h �y   ™ h �y!   ™ h �y"   ™ b ‚x       Ä ‚x  ÿþý     ä $~$ÿ   ÿ Ž ä $~%ÿ   ÿ Ž ä $~	ÿ   ÿ Ž â/ x@   ÿàÿ ä xv      âO $~ÿ   ÿ Ž Æ xTp      ä� xww  ÿÀŽ ä xT      ät    ä xˆˆˆˆÿÀŽ ä xÿ     Ä x22ÿø� ä xv      ät$üúøô    ä xww  ÿÀŽ ä xTv  ÿÀŽ ä xTp      Ä t		    är       ä t%üúøô    ä xÿ     ä x T      ä r	       Ä xˆˆˆˆÿÀŽ â $~ÿ   ÿ Ž â xww  ÿÀŽ â $~ÿ   ÿ Ž â x22ÿø� â $x#€   ÿ Ž â xÿ     â $~ÿ   ÿ Ž â t	    â$~ ÿ   ÿ Ž â tüúøô    Ä xTv  ÿÀŽ ä xww  ÿÀŽ ä r #ÿ   pRô ä x€  ÿÀŽ ä r		       ä t    Ä xÿ     ä t üúøô    ä x d      ä x #ÿ   pRò ä x@       â &r   ÿ   â r       Ä x1u      ä xÿ    €  ä x	 d      â &r      â xÿT      ä x	1u      â 0rÿ    A   ä  x	   ?  A  â &r!      Æ  r		     A  â &r"      â G‰  p    € ê r   ÿàó Ê $x   Ž  Ê �y#    ¨  �y    ¨  �y    è  �y    ( �y    h �y    h �y    h �y     d $rÿÿ    Ž Ä $rÿÿ    Ž Ê �y$     d $~ÿ   ÿ Ž â xv  #    ÈO xTp      ä� x%ww  ÿÀŽ ä x#T      â$~ÿ   ÿ Ž â t%    â $~ÿ   ÿ Ž â xÿ   #  ä x#ˆˆˆˆÿÀŽ â $~#ÿ   ÿ Ž â xww  ÿÀŽ Ä x22ÿø� ä xv      â$~ÿ   ÿ Ž â t    ät#üúøô    ä x#Tv  ÿÀŽ ä tüúøô%    Ä xÿ     ä xTp      ä r#       â $~ÿ   ÿ Ž â r       â $~ÿ   ÿ Ž â x T      ä x d      Ä x ˆˆˆˆÿÀŽ ä x%ÿ     â &r   ÿ   â x1u      â $~ÿ   ÿ Ž â x#ww  ÿÀŽ ä x 22ÿø� â &r      â x%ww  ÿÀŽ Ä t#    ätüúøô#    ä x#Tv  ÿÀŽ ä t    ätüúøô    â $x$€   ÿ Ž â xÿ     Ä r $ÿ   pRô ä r#       ä r       ä x $ÿ   pRò ä x@       ä x d      Ä xÿ    €  ä x1u      â &r!      â xÿT      Æ &r"      ä  x   ?  A  Æ Er        â  r     A  È #r      â |    pbø ä x

@  ÿàõ ä x@  ÿà÷ ä Er        â $rÿÿ    â x@  ÿàó â $rÿÿ   Ž ä #r	      Ä $rÿÿ   Ž  â GÉ  Àõÿÿÿÿƒ ê Ay        € ê r  ÿ   pRð âEy  €     € â r  ÿ   pRò Ö G‰  P     € ê z ÿ `  pPð â $x      Ž Æ z ÿ a   Sð ä ˆs  €ÿÿ H   ö ˆ     H   ä Ay        € ê {          ì M        € ê z ÿ `  pPð â „y     H   fz ÿ a   Sð Ú „‰  €   H   ( „‰    H   ¢ !r       È/ !‚         È !‚        âO G‰        € ê „y  €  H   ( „y €   H   h „y    H   ¨ „y	    H   è „y €  H   " !r         â !r       Æ/ !r         âO !r	       Æ� !r        ä‰      " z ÿ `  pPð ä r ÿ   pRò ä z ÿ a   Sð â !r       Ê ‰       $ !r         Ê ‰  €    $ !r        Ê ‰ @    $ !r       Ê ‰      $ !r       â G‰        € ê ‰       $ !r         Ê ‰       $ !r        Ê ‰ €    $ !r       Ê ‰ @    $ !r       Ê ‰      $ !r       ä M        € ê !r       â G‰  `    € ê $v ÿ h  ÿ Ž â !r       È x     pBð Ú G	  p    € ê x  ÿÿÿÿÿàÿ È x       € Ê $x     ÿ Ž È ‚{   €     $ xÿ     È Iy  ëÿÿÿÿƒ ê  x ;ª¸¿  A  Ì s          $ !t   €?    Ì s        $  r     A  È  r     A  â Gy  P    € ê  x '7=  A  â x*BL?  A  Æ #t   €?    È  r      A  â  x    ?  A  Ê s     $   $ !t  €?    È  r      A  È  r     A  â Gy  °    € ê x     pRð Ú G‰  à     € ê x     pRð Ú G	        € ê 	z  i    � ä 	z i    � Æ  x ;ª¸¿  A  â 	z i €  � Ê s        $ !t  €?    Ì s        $  r      A  È  r     A  â Gy  Ð     € ê  r     A  â Gy  °     € ê 	x   à@  � ä 	x  à@  � Æ  x #ÛÙ?  A  â 	x  àÀ  � Æ  x;ª¸¿  A  ä !t  €?    È s        $ !t  €?    Ì s        $  r       A  È  r      A  ä y       &   ( y      %   ( y      '   b $z   x  Ž ä $tÿ   ÿ Ž Ä $z  ~   Ž È/ %v  j   Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y       &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž ä $rÿÿ     Ž È % \   Ž Ê �	    ™ b ¹z  l      â Ey  ð    € â ™x?    â y      "   b $rÿÿ   ÿ Ž Æ y      !   ¢ $ÿÿ   ÿ Ž ä $ÿÿ    Ž â y	      %   & '
 m   Ž È $‚ÿÿ   ÿ Ž â 
ÿ n    Ê $ÿÿ   
Ž â r ÿ   pRò Æ/ $
 o  Ž â r ÿ   pRô â$x    Ž äO $‚ÿÿ    Ž ä $rÿÿ   ÿ Ž â r   ÿ� â 'Š s   Ž Æ xÿ     â $€   Ž ä $x		   ÿ Ž â |    pbö ä Šÿ t    ä x      ÿ  Ò G9  @    € ê y      '   " $r
ÿÿ   ÿ Ž â xþÿÿÿÿÀŽ â $z w  ÿŽ ä $vÿ X  ÿ Ž ä $x   
Ž ä $vÿ Y  ÿ Ž ä $x    ÿ Ž Ä $x   ÿ Ž Æ x!ÿ      â 'z y  
 Ž È $z
 }  ÿŽ â zÿ z    â $z	 p  Ž Æ r

   ÿàñ â $z |  Ž â xÿ     â $rÿÿ   ÿ Ž ä $z v  Ž âx     â $rÿÿ   ÿ  ä %x     Ž Ê r"    ÿàñ Ê $x#!     Ê �y"   ‘ ¨  �y"   ‘ ¨  �y"   ‘ è  �y"   ‘ ( �y"   ‘ h �y"   ‘ h �y"   ‘ h �y"   ‘ b $v-ÿ p  ÿ Ž È %x$-   "Ž Ê �y$   ‘ h �y$   ‘ h �y$   ‘ h �y$   ‘ h �y$   ‘ h �y$   ‘ b r&
   ÿàñ Æ �y$   ‘ b $t#ÿ$   ÿ Ž ä $r'ÿÿ     â �y.$   ‘ b %v"& Z  # Ž È $x)'$   ÿŽ â r("   ÿ€ â %x&-   $Ž È $x##   )Ž â �y/&   ‘ h �y0&   ‘ b $x)#     Æ �y1&   ‘ h �y2(   ™ h �y4&   ‘ h �y3(   ™ h �y5&   ‘ h �y8&   ‘ h �y7&   ‘ h �y6(   ™ h �y:&   ‘ h �y9(   ™ b %x$-   &Ž Æ/ �y""    • h �y;$   ‘ h �y<$   ‘ h �y=$   ‘ h �y>$   ‘ h �y?$   ‘ h �y@$   ‘ h �yA$   ‘ h �yB$   ‘ d $r$ÿÿ    Ž Ä/ $r%ÿÿ    Ž È %x&   $Ž Ê �yC&    ‘ b %x(-   &Ž Ê �yD(    ‘ b %x*-   (Ž Ê �yE*    ‘ b %x,-   *Ž Ì �y,,    ‘ b ‚x       ä ‚x  ÿþý     â $~&ÿ   ÿ Ž ä $~*ÿ   ÿ Ž â/ xv      ÈO x$Tp      â� $~ÿ   ÿ Ž â xww  ÿÀŽ ä x$T  $    ät    ä xÿ   $  ä x$$ˆˆˆˆÿÀŽ â $~ÿ   ÿ Ž â xww  ÿÀŽ Ä x$$22ÿø� ä tüúøô    ä x(v      ät&    äx$Tv  ÿÀŽ ä t&*üúøô    Ä x$ÿ   $  ä x*Tp  (    ä r       ä r$   &    â $~ÿ   ÿ Ž â xT  *    ä x d      Ä x((ww  ÿÀŽ ä x1u      â $~ÿ   ÿ Ž â xˆˆˆˆÿÀŽ ä t(    äxÿ     ä tüúøô(    Ä x22ÿø� ä xv      ä xTv  ÿÀŽ È r       ä xTp      â $~$ÿ   ÿ Ž â xww  ÿÀŽ â $~&ÿ   ÿ Ž â xT      ä xÿ     ä xÿ     Ä xv      ä t$    äxˆˆˆˆÿÀŽ â $~ÿ   ÿ Ž â t&üúøô    â $~$ÿ   ÿ Ž â xww  ÿÀŽ ä xww  ÿÀŽ Ä xTp      ä x22ÿø� ä r       ä t    ät    ä x.T      Ä t$üúøô    ät$üúøô    ä xTv  ÿÀŽ ä xÿ     ä x d      äx1u      Ä xÿ     ä r       ä xˆˆˆˆÿÀŽ ä r       â $~ÿ   ÿ Ž â xww  ÿÀŽ â $~ÿ   ÿ Ž â x22ÿø� â $~ÿ   ÿ Ž â xww  ÿÀŽ â $~ÿ   ÿ Ž â x00v  /    Ä t    ätüúøô    ä xTv  ÿÀŽ ä t    ätüúøô    ä x11Tp  0    Ä xÿ     ä x d      ä r       â &r2   ÿ   âx14T  1    â &r2   ÿ   â x1u      ä r       â &r3      âxÿ   1  â &r3      â x d      Ä x1u      â $~ÿ   ÿ Ž â x11ˆˆˆˆÿÀŽ â $~ÿ   ÿ Ž â x00ww  ÿÀŽ â $~ÿ   ÿ Ž â xww  ÿÀŽ â $~ÿ   ÿ Ž â x1122ÿø� ä t0    Ä t    äx88v  5    ä tüúøô    ä x1Tv  ÿÀŽ ä t0üúøô0    ä x1ÿ   1  Ä x77Tp  8    â &r6      âr   0    ä r1       ä x7:T  7    â &r9      â x d      â&r6      â x1u      â $~ÿ   ÿ Ž â xÿ   7  â $~ÿ   ÿ Ž â x88ww  ÿÀŽ â $~ÿ   ÿ Ž â x77ˆˆˆˆÿÀŽ â $~ÿ   ÿ Ž â xww  ÿÀŽ Ä t8    äx7722ÿø� ä tüúøô8    ä t    ätüúøô    ä x<<v  ;    Ä x7Tv  ÿÀŽ ä x7ÿ   7  ä x==Tp  <    ä r       ä r7       ä x=>T  =    â &r2   ÿ   â x d      Ä x1u      â $~ÿ   ÿ Ž â xÿ   =  â &r3      â x<<ww  ÿÀŽ â $~ÿ   ÿ Ž â x==ˆˆˆˆÿÀŽ â $~ÿ   ÿ Ž â x@@v  ?    â $~ÿ   ÿ Ž â xww  ÿÀŽ Ä t<    äx==22ÿø� ä xAATp  @    ä t<üúøô<    ä t    ätüúøô    Ä x=Tv  ÿÀŽ ä xABT  A    â &r6      â x=ÿ   =  ä r   <    â $~ÿ   ÿ Ž â xÿ   A  â $~ÿ   ÿ Ž â x@@ww  ÿÀŽ Ä xAAˆˆˆˆÿÀŽ ä r=       ä xww  ÿÀŽ ä t@    äxAA22ÿø� ä t@üúøô@    Ä t    ätüúøô    ä xATv  ÿÀŽ ä x d      äxAÿ   A  ä x1u      â &r2   ÿ   â r   @    Ä rA       â &r9      ä&r9      â x d      â &r3      ä $xC€   ÿ Ž ä &r6      â r Cÿ   pRö â $xD€   ÿ Ž â x€  ÿÀŽ Ä x Cÿ   pRú ä r Dÿ   pRü ä x@     € ä x€  ÿÀŽ ä x Eÿ   pRö är Eÿ   pRø â$xEE€   ÿ Ž â xÿ    € Ä x@       ä x ,ÿ   pRú är ,ÿ   pRü â$x,,€   ÿ Ž â xEE€  ÿÀŽ È x,,€  ÿÀŽ ä x Dÿ   pRð ä xEE@       ä x,,@       ä x@   ÿàÿ ä xÿ       Ä xEEÿ    € ä x,,ÿ    € ä x1u      ä |    pbø ä xÿT      ä xÿT      Ä xEÿT  E    ä x,ÿT  ,    â &r99      â  x   ?  A  â 0rÿ"    A   â  x   ?  A  â  xE   ?  A  â  x,,   ?  A  â x@  ÿàñ â  r     A  â  r     A  â r     A  âEr        â  r,     A  â Er        Ä Er        ä Er9 9       â #r      â #r      â #r      â $rÿÿ     â #r9      â DI       À â Gy   íÿÿÿÿƒ ê Ay        € ê ˆ  þÿ H   è ˆ €þÿ H   è ˆ  ÿÿ H   è ˆ €ÿÿ H   è {          ì M)        € ê „y     H   " z 	 r  p`ð â $z   x  	Ž Ä „y €   H   b r ÿ   pVp  Æ „y    H   ¨ „y    H   è„y €  H   ( „y    H   ( „y
    H   ( „y €  H   ( „y €  H   ( „y    H   ( „y €  H   " !r       Æ „y €  H   " !r       â/ !r       äO y      '   b !r       â� !r       â!r       ä !r
       â !r       È ‰      ¢ !r       â $z  ~   Ž â/ !r       Æ ‰      b !r       Æ ‰      â !r       Ê ‰      " !r       âO !r       â/ !r       È� ‰      h ‰      ¢ !r       Æ ‰      ( ‰      â !r       â/ !r       ÈO ‰
 €    b !r       Æ ‰ €    " !r       Æ� ‰ €    ¨ ‰ €    â !r

       â/ !r       È ‰
 @    " x	   ÿàÿ â !r       äO ‰ @    b x	   ÿàÿ â !r       ä� ‰ @    ¢ z  r  p`ò ä z  r  p`ö â ‰ @    â x    pVò  Ä x    pVö â !r
       â x
	   ÿàÿ â $t	ÿ   ÿ Ž â !r       ä/ ‰      " z 
 r  p`ô â !r       äO ‰      b x    pVt â !r       Ä� ‰      ¨ ‰      â !r       â !r       â/ %v  j  	 Ž â !r       ÈO †‰     â !r       Æ� †™    è †©    â M9        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      &   " z ÿ \  pPò â ¹z  F   
   Æ z ÿ ]  Sò Ú         Ê % \   Ž Ê �    ™ ¢  r ÿ       â Ey       € ä y      "   d '  m   Ž äy      %   ä 'š s   Ž â  ÿ n     â y      !   &  ÿ   ÿáÿ â y	      '   ( $  o  Ž â ’         Ä r ÿ   pRð ä/ x   ÿ   ä� x    pDp  är
   ÿàÿ är   ÿ(Ž ä z 
 r  pdp  Ä šÿ t    ä x      ÿ  ä r     € ÎO G	  °    € ê z ÿ `  pPò ä z ÿ ^  pPð ä z ÿ b  p ô ä z ÿ a  Sò ä z ÿ _   Sð ä z ÿ c   'ò Ö $
	 ~  ÿŽ âÿ     â$š	 ~  ÿŽ ä $
 x  ÿŽ â    àë â$š
 x  ÿŽ â ’   àÇ ä ˜ÿ     ä    ÿàõ Ä ÿ   ÿÄþ ä ’

   ÿàû ä ’ÿ   ÿ„þ ä ÿ   ÿä ä 
 ^  ÿ„ ä ’ÿ   ÿäÿ Ä š
 b  ÿ† ä 
 _   ä x      ÿ  â š
 c  � Ê �	     h �™     d Ay        € ê r   ÿ� â ¹z  l      â $z	 }  ÿŽ â ™x?    â xÿ     â $z   w  ÿŽ â Ey  (    € â $rÿÿ   ÿ Ž â r ÿ       â x
      ÿ  â |    pbð â 'z		 y   Ž â r    ÿàó â x      ÿ  â r ÿ       Ä r ÿ       ä r  ÿ       ä rÿÿ   ÿäÿ  Æ G	  @'    € ê z ÿ z  	  â $z	 p  Ž âxþÿÿÿÿÀŽ â $vÿ Y  ÿ Ž â x   ÿàÿ â$z  |  Ž â r  €ÿàÿ â $z   |  	Ž â x	   ÿàÿ â $z v  Ž â x   ÿàÿ â $z v   Ž â x(   ÿ   â$z		 p  Ž âx   ÿ   â $z% p  Ž â x)ÿ   (  â $z p  Ž â z ÿ `  pPð â x
      ÿ  â xÿ     â %x"	   (Ž âz  `      â x      ÿ  â z  a      â %x$%   (Ž â z  X      Ä r ÿ       â %x&   (Ž âr ÿ       ä r ÿ       â %x(   (Ž â r  ÿ       ä z ÿ a   Sð ä x     Ä r(   ÿàó È r)   ÿäÿ  Ê �y0   ‘ ¨  �y1   ‘ ¨  �yJ   ‘ è  �yK   ‘ ( �yD   ‘ ( �y9   ‘ " z  p      Æ �yI   ‘ ( �y@   ‘ " %x    Ž â r.   ÿàó È �yC    ‘ ( �yH    ‘ " x/ $       â $r	ÿÿ   Ž  ä �yE    ‘ $ %v.. Z  / Ž Ä �yG    ‘ $ $x		$   ÿŽ â r*.   ÿ‚ â �y;    ‘ & r//	   ÿàÿ â �yF    ‘ & r+/   ÿäÿ  â �y8    ‘ ( �y*   ™ ( �y*   ™ " %xL    Ž Æ �yB    ‘ ( �y*   ™ ( �y*   ™ ( �y<L   ‘ ( �yAL   ‘ ( �y>L   ‘ ( �y	L   ‘ ( �y4L   ‘ ( �y?L   ‘ ( �y6L   ‘ ( �y=L   ‘ " %x*   Ž Æ �y7.    • " %x,   LŽ Æ �y3*    ‘ ( �y!,   ‘ (# �y:,   ‘ ( �y,   ‘ ( �y ,   ‘ ( �y,   ‘ ( �y5,   ‘ ( �y2,   ‘ " %x.   *Ž Æ �yL,   ‘ & %x*   .Ž â�yM.    ‘ ( �y.*    ‘ $ %x*   *Ž Ê �y*    ‘ " ‚x       ä ‚x  ÿþý     â |,       ä/ x11v  0    ÈO xJJTp  1    È� xJKT  J    Èx-JˆˆˆˆÿÀŽ ä x11ww  ÿÀŽ ä xD9v  D    ä |0       ä xJÿ   J  ä x*-22ÿø� Ä xIITp  D    ä t,,1    ät+0üúøô1    ä |-       ä xJJww  ÿÀŽ ä |1       Ä x/*Tv  ÿÀŽ ä x@@T  I    ä t--J    ät01üúøôJ    ä r+,/   +    ä x*ÿ   *  Ä |/       ä xDDww  ÿÀŽ ä x,ÿ   @  ä r0-*   0    ä t*/D    ät-1üúøôD    Ä |D       ä x@@ˆˆˆˆÿÀŽ ä x/,ww  ÿÀŽ ä xHHv  C    ä x@@22ÿø� ä t,D/    Ä |J       ä xDETp  H    ä x1@Tv  ÿÀŽ ä t9Jüúøô/    ä x@ÿ   @  ä xDGT  D    Ä x/+ d  0    äx0+1u  0    ä r**1   -    ä r+,@   9    â $~,ÿ   ÿ Ž â x1ÿ   D  ä xDDˆˆˆˆÿÀŽ Ä x-Hww  ÿÀŽ ä |@       ä |H       ä x11ww  ÿÀŽ ä xDD22ÿø� ä t,,-    Ä t-@üúøô-    ä t9H1    ät@@üúøô1    ä x1DTv  ÿÀŽ â &r//   ÿ   â xDÿ   D  ä r,,1   -    â &r/0   /   â r99D   @    Ä x0* d  +    ä x;Fv  ;    ä x**1u  +    ä x+, d  9    ä x-8Tp  ;    ä x,,1u  9    â &r++   ÿ   â x-BT  -    â &r/0   /   â |0       â &r+,   +   â x;;ww  ÿÀŽ Ä |8       â &r**   /   â x,-ˆˆˆˆÿÀŽ ä x1ÿ   -  ä t/0;    äxAAv  <    ä t08üúøô;    Ä x-,22ÿø� ä |8       ä x11ww  ÿÀŽ ä |@       ä x>>Tp  A    ä x,-Tv  ÿÀŽ Ä t881    ät1@üúøô1    ä x-ÿ   -  ä x		T  >    ä r-8-   1    ä x1ÿ   	  Ä |8       ä x11ww  ÿÀŽ ä x??v  4    ä x		ˆˆˆˆÿÀŽ ä r,/,   0    ä t481    Ä |0       ä xAAww  ÿÀŽ ä |<       ä x86Tp  ?    ä x		22ÿø� ä t/0A    Ä t9<üúøô1    äx8=T  8    ä t0<üúøôA    ä x1	Tv  ÿÀŽ ä x6ÿ   	  ä x	, d  -    Ä x,,1u  -    ä x-ÿ   8  ä r//1   0    ä r446   9    ä x1-ww  ÿÀŽ ä x-/ d  4    Ä x4/1u  4    â $~0ÿ   ÿ Ž ä &r--   ÿ   â x88ˆˆˆˆÿÀŽ â &r		   +   â x??ww  ÿÀŽ â &r+4   -   â |6       ä x 3ÿ   pRò Ä r 3ÿ   pRô ä|<       ä |>       ä x-:v  !    ä x33   ÿ   ä x8822ÿø� Ä t00?    ät?6üúøô?    ä t6<1    ät9>üúøô1    ä x/Tp  -    ä x33€  ÿÀŽ Ä x18Tv  ÿÀŽ ä x8ÿ   8  ä x  T  /    ä x33@       ä r001   ?    ä r968   9    Ä |4       ä x--ww  ÿÀŽ ä |6       ä xÿ      ä x33ÿ    €  ä x!0 d  9    Ä t/4-    äx001u  9    ä t46üúøô-    ä |-       ä xww  ÿÀŽ ä |9       Ä x3ÿT  3    â 0r7ÿ7    A   â x  ˆˆˆˆÿÀŽ ä t6-    ät99üúøô    â  x3   ?  A  â x  22ÿø� ä Er* *       â  r-7     A  â x55v      Ä x1 Tv  ÿÀŽ â #r -*       â x ÿ      ä x-2Tp  5    â $~ÿ   ÿ Ž â r//1   4    ä r 6    9    ä |       Ä x55ww  ÿÀŽ â &r+!   +   â x-LT  -    ä x!/ d       ät5    ät5üúøô5    ä xÿ   -  â &r!!   ÿ   â x /1u       Ä x--ˆˆˆˆÿÀŽ ä |/       ä xww  ÿÀŽ ä |1       â &r!    !   â x--22ÿø� ä t /    Ä x Mÿ   pRò är Mÿ   pRô ä t/1üúøô    ä xMM   ÿ   ä x-Tv  ÿÀŽ ä xMM€  ÿÀŽ Ä r   5    ä x .ÿ   pRö är .ÿ   pRø äx   ÿ   ä x..   ÿ   ä xMM@       Ä x-ÿ   -  ä r ÿ   pRô ä x*€  ÿÀŽ ä x..€  ÿÀŽ ä xMMÿ    €  ä r- -   /    Ä x ÿ   pRò ä x**@       ä x..@       ä x d  -    ä xMÿT  M    ä x**ÿ    €  Ä x..ÿ    € â &r	,   	   âx 1u  -    â  xM   ?  A  â x*ÿT  *    â &r   !   â x.ÿT  .    â &r+0   +   â Er	 	       â  r7     A  â x**   ?  A  â &r        â  x..   ?  A  â #r	      â  r7*     A  âEr+ +       â  r	7.     A  â Er*         â G‰  `    € ê r(   ÿàó È r)   ÿäÿ  Ê �y,    ¨ �y-    ¨ �y/    è �y.    ( �y0    ( �y1    ( �y3    ( �y2    " %x    Ž Ì �y       " |5       â $~6ÿ   ÿ Ž â |4       ä x,-v  ,    ÈO x//Tp  ,    ä� x,,ww  ÿÀŽ ä x..T  /    ä|/       ä x-ÿ   .  ä x..ˆˆˆˆÿÀŽ Ä x--ww  ÿÀŽ ä x..22ÿø� ä x01v  0    ä t/,    ät5üúøô,    ä t,4-    Ä x!.Tv  ÿÀŽ ä/ x33Tp  0    ä t-6üúøô-    ä x.ÿ   .  ä x22T  3    ä r!       Ä r-,.   -    ä x!2ˆˆˆˆÿÀŽ ä x d  -    äx1u  -    ä x2ÿ   2  â &r   ÿ   â |-       Ä x00ww  ÿÀŽ â &r      â |/       ä x!!22ÿø� ä |1       ä x22ww  ÿÀŽ ä |3       Ä t,-0    ät-/üúøô0    ä x.!Tv  ÿÀŽ ä t/12    ät23üúøô2    ä x!ÿ   !  Ä r,,.   -    ä r  ÿ   pRô äx-    ÿ   är!/!   2    ä x  ÿ   pRò ä x--@       Ä x , d  !    äx--ÿ    €  ä x,,1u  !    â &r       â x-ÿT  -    Æ &r,      ä  x-   ?  A  Æ Er        â  r7     A  È #r      ä G‰  €    € ê r "   ÿàó È r!#   ÿäÿ  Ê �y.     ¨  �y/     ¨  �y1     è  �y0     ( �y3     ( �y4     ( �y-     " x   ÿàÿ Æ �y,     $ $z p  Ž È %x   Ž Ì �y     " $~5ÿ   ÿ Ž â |!       ä x./v  .    äO |/       ä x11Tp  .    ä� x..ww  ÿÀŽ Ä x00T  1    ä|1       ä x ÿ   0  ä x00ˆˆˆˆÿÀŽ ä x  ww  ÿÀŽ ä x0022ÿø� Ä t!!     ät25üúøô     ä x44v  3    ä t/.    ä/x 0Tv  ÿÀŽ ä t.1üúøô.    Ä x0ÿ   0  ä x--Tp  4    ä r    .    ä r0!0   2    ä x,,T  -    ä x! d  0    Ä x01u  0    ä x,ˆˆˆˆÿÀŽ â &r!!   ÿ   â x,ÿ   ,  ä |-       â &r!0   !   â x44ww  ÿÀŽ ä |/       Ä x22ÿø� ä |1       ä x.,ww  ÿÀŽ ä |3       ä t -4    ät-/üúøô4    Ä x,Tv  ÿÀŽ ä t/1.    ät.3üúøô.    ä xÿ     ä r  ,   -    ä r ÿ   pRô Ä x,   ÿ   är/   .    ä x ÿ   pRò ä x,,@       ä x  d      ä x,,ÿ    €  Ä x  1u      â &r   !   â x,ÿT  ,    Æ &r       ä  x,,   ?  A  Æ Er        â  r7,     A  È #r      ä G‰  €    € ê r &   ÿàó È r!'   ÿäÿ  Ê �y.     ¨  �y/     ¨  �y1     è  �y0     ( �y3     ( �y4     ( �y-     " x   ÿàÿ Æ �y,     $ $z p  Ž È %x   Ž Ì �y     " $~5ÿ   ÿ Ž â |!       ä x./v  .    äO |/       ä x11Tp  .    ä� x..ww  ÿÀŽ Ä x00T  1    ä|1       ä x ÿ   0  ä x00ˆˆˆˆÿÀŽ ä x  ww  ÿÀŽ ä x0022ÿø� Ä t!!     ät25üúøô     ä x44v  3    ä t/.    ä/x 0Tv  ÿÀŽ ä t.1üúøô.    Ä x0ÿ   0  ä x--Tp  4    ä r    .    ä r0!0   2    ä x,,T  -    ä x! d  0    Ä x01u  0    ä x,ˆˆˆˆÿÀŽ â &r!!   ÿ   â x,ÿ   ,  ä |-       â &r!0   !   â x44ww  ÿÀŽ ä |/       Ä x22ÿø� ä |1       ä x.,ww  ÿÀŽ ä |3       ä t -4    ät-/üúøô4    Ä x,Tv  ÿÀŽ ä t/1.    ät.3üúøô.    ä xÿ     ä r  ,   -    ä r ÿ   pRô Ä x,   ÿ   är/   .    ä x ÿ   pRò ä x,,@       ä x  d      ä x,,ÿ    €  Ä x  1u      â &r   !   â x,ÿT  ,    Æ &r       ä  x,,   ?  A  Æ Er        â  r,7,     A  È #r,      â #r	+      â #r*      â G‰  €    € ê r$   ÿàó È r%   ÿäÿ  Ê �y	    ¨ �y    ¨ �y*    è �y+    ( �y,    ( �y-    ( �y/    " x!   ÿàÿ Æ �y.    $ $z!! p  Ž È %x !   Ž Ì �y       " $~4ÿ   ÿ Ž â |2       ä |0       ä x	v  	    äO |       ä x**Tp  	    Ä� x		ww  ÿÀŽ ä x*+T  *    ä|+       ä xÿ   *  ä x**ˆˆˆˆÿÀŽ ä xww  ÿÀŽ Ä x**22ÿø� ä t	    ä t2    ät!4üúøô    ä/ x,-v  ,    ä x*Tv  ÿÀŽ Ä t	0üúøô	    ä x*ÿ   *  ä x//Tp  ,    ä r	   	    ä r*   !    ä x..T  /    Ä x	 d      äx	1u      ä x	.ˆˆˆˆÿÀŽ â &r   ÿ   â x.ÿ   .  ä |!       â &r      â x,,ww  ÿÀŽ Ä x		22ÿø� ä |-       ä x..ww  ÿÀŽ ä |/       ä t!,    ät!+üúøô,    Ä x*	Tv  ÿÀŽ ä t+-.    ät./üúøô.    ä x	ÿ   	  ä r*   !    ä r  ÿ   pRô Ä x!    ÿ   är	+	   .    ä x  ÿ   pRò ä x!!@       ä x d  	    ä x!!ÿ    €  Ä x1u  	    â &r      â x!ÿT  !    Æ &r      ä  x!   ?  A  Æ Er        â  r77     A  È #r
7   
   â x@   ÿàÿ ä x@  ÿàõ ä |    pbò ä x@  ÿà÷ ä rÿ   ÿä ä rÿ   ÿäÿ Î D       À â Gy  ðÚÿÿÿÿƒ ê Ay        € ê y	      "   " Ey  0    € â r 	ÿ   pRð är 	ÿ   pRò Ö G‰  ð     € ê z ÿ `  pPð ä r	   ÿ8Ž ä z ÿ a   Sð Æ ˆs   þÿ H   ô ˆƒ €þÿ H   è ˆƒ  ÿÿ H   è ˆƒ €ÿÿ H   â G‰  p     € ê ˆs    H   è ˆs €þÿ H   è ˆs €  H   è ˆs  ÿÿ H   è ˆs    H   è ˆs €ÿÿ H   è ˆs 
€  H   ä Ay        € ê {          ì M        € ê y      &   "> z ÿ `  pPð Ä x        â y	      '   b z ÿ a   Sð â $z x  Ž È $z	 ~  Ž È/ %v j   Ž È G	  P    € ê „y	     H   " z  r  p`ð Æ „y    H   b r ÿ   pVp  Æ „y    H   ¨„y
    H   è „y    H   ( „y €   H   ( „y    H   ( „y €  H   ( „y €  H   ( „y €  H   ( „y €  H   " !r		        Æ „y €  H   " !r       â/ !r	       ÆO !r

       â� !r       â!r       È ‰	      b !r

       â !r       È ‰
      ¢ !r       â !r       È ‰      â !r       È !r       â !r		       Ê/ ‰ 	      " !r
       ÆO ‰      h ‰      ¢ !r       Ê� ‰      â !r 	        â !r       â/ !r
       ÈO ‰      ( ‰  €    b !r       Æ� ‰
 €    ¨ ‰ €    â !r       â !r        È/ ‰ €    " x    ÿàÿ â!r
       äO ‰ @    b z   r  p`ò â !r       ä� ‰ @    ¢ x   ÿàÿ äx   ÿàÿ â ‰ @    â z  r  p`ô Ä x    pVò  äx    pVt â !r       â !r       È/ ‰ @    " !r       ÆO ‰	      b !r       Æ� ‰      ¨ ‰      â !r       â !r	       È/ ‰       " !‚       â !r       âO !r       Æ� †‰     â z  r  p`ð â !¢       â!’	       ä x    pVp  ä †©    è †™ 	   â !r         Ì M	        € ê !r        Ê †y    â My        € ê z  h      È x    pBð Ú G	  @    € ê x    pRð Ú G‰  À    € ê x    pRð Ú G	   $    € ê „y     H   " z  r  p`ð â Ey  p    € ä „y    H   b r ÿ   pVp  Æ „y    H   ¨ „y	    H   è „y    H   ( „y  
  H   ( „y €   H   ( „y €  H   ( „y €  H   è „y €  H   " !r        â !r       Æ/ !r       âO !r	       Æ� !r       â!r       È ‰       " !r       Æ ‰	      b !r       È !r       â !r         ä „y €  H   â !r	       Æ/ ‰	       ( ‰      b !r	 	       â !r       È/ ‰	 €    ( ‰ €    b !r       Æ „y €
  H   ¢ !r       â !r	       â !r       È/ ‰ @    ( ‰ @    b !r       âO !r       â !r       È/ ‰       ( ‰	      b !r         â !r		       â/ G	  �     € ê !r	       â!r         Æ  x	;ª¸¿  A  Ì s	 	       $ !t	  €?    Ì s        $  r     A  È  r      A  Ê †y     ä Ay        € ê ‰       b Ey  Ð    € æ ‰      ¨ „y    H   è „y    H   ( „y    H   è „y  	  H   â !r         Æ/ „y    H   â !r       ÆO ‰	       h ‰      " !r       â� !r       â!r	 	       ä/ !r       â !r       Ä ‰	 €    " !r       Æ ‰ €    b !r       Æ „y    H   ¢ !r	       â x	   ÿàÿ â !r       Æ/ z 	 r  p`ð â ‰ @    & x    pVp  â ‰ @    b !r       âO !r       â !r       È/ ‰       ( ‰      b !r         â !r       â/ G	  �     € ê !r       â!r         Æ  x;ª¸¿  A  Ì s        $ !t	  €?    Ì s	 	       $  r	     A  È  r      A  Ê †y    ä Ay        € ê ‰       b Ey       € æ ‰	      ¨ „y €  H   è „y €  H   è „y €  H   è „y €  H   â !r         Æ/ „y €  H   â !r	       ÆO ‰	       h ‰      ( „y €	  H   ¢ !r
       â� !r	 	       â/ !r       È ‰	 €    " x   ÿàÿ Æ ‰ €    b z  r  p`ð È x    pVp  â !r       ÈO !r       â !r	       â !r       È/ ‰ @    ( ‰ @    b !r       â !r       È/ ‰       ( ‰      b !r         â !r	       â/ G	  �     € ê !r	       â!r         Æ  x;ª¸¿  A  Ì s        $ !t	  €?    Ì s	 	       $  r	     A  È  r      A  Ê †y    ä Ay        € ê ‰	      b !r       â x   ÿàÿ Æ !r       â z  r  p`ð Æ !r       â x    pVp  È ‰       ¢ !r	       Ê/ ‰	      b !r         ÊO ‰       ¢ !r
	       Ê/ ‰	
 €    b !r        ÊO ‰ €    ¢ !r
	       Ê? ‰ @    " !r       ÊO ‰	 @    b !r       Ê ‰      " !r		       Ê/ ‰ 	      ¢ !r       â M	        € ì !r       â/!r 	        ÆO  x;ª¸¿  A  â !r        Ê s        $ !t  €?    Ì s        $  r     A  È  r     A  Ê †y    â My        € ê „y	     H   " z  r  p`ð â Ey       € ä „y    H   b r ÿ   pVp  Æ „y    H   ¨ „y    H   è „y    H   ( „y  
  H   ( „y €   H   ( „y €  H   ( „y €  H   " !r		        â !r       Æ/ !r		       âO !r       Æ� !r	       â!r       È ‰       " !r       Æ ‰      b !r       È !r       â !r         â !r       È/ ‰	       ( ‰      h „y €  H   ¢ !r	 	       â !r       È/ ‰	 €    ( ‰ €    h „y €  H   â !r       âO !r	       Ä „y	 €
  H   â !r       Æ/ ‰ @    ( ‰ @    b !r       â� !r       â !r       È/ ‰       ( ‰      b !r	       â !r        â !r	       â/ G	  À     € ê !r 	       È x	 '7=  A  â x *BL?  A  Æ #t	   €?	   È  r		     A  â  x    ?  A  â !r        È s	 	    $   $ !t	  €?    È  r     A  È  r      A  Ê †y     ä Ay        € ê ‰       b Ey       € æ ‰      ¨ „y    H   è „y    H   ( „y    H   â !r         Æ/ „y    H   â !r       ÆO ‰	       h ‰      ( „y  	  H   ¢ !r       â� !r       â!r	 	       â/ !r       È ‰	 €    ( ‰ €    h „y    H   â !r       ÈO !r       â !r	       â !r       È/ ‰ @    ( ‰ @    b !r       È� !r       â !r       â x   ÿàÿ â !r       Æ/ ‰       " z  r  p`ð Æ ‰	      b x    pVp  â !r        â !r		       Ö/ G	  À     € ê !r 	       È x	 '7=  A  â x *BL?  A  Æ #t	   €?	   È  r		     A  â  x    ?  A  â !r        È s	 	    $   $ !t	  €?    È  r     A  È  r      A  Ê †y    ä Ay        € ê ‰       b Ey  Ð    € æ ‰      ¨ „y €  H   è „y €	  H   ( „y €  H   è „y €  H   â !r         Æ/ „y €  H   â !r       ÆO ‰	       h ‰      ( „y €  H   â !r
       È� !r       â!r	 	       Æ/ !r       â !r       â ‰	 €    ( ‰ €    b !r	       â !r       È/ ‰ @    ( ‰ @    b !r       â x   ÿàÿ â !r       Æ/ ‰       " z  r  p`ð Æ ‰	      b x    pVp  â !r        â !r		       Ö/ G	  À     € ê !r 	       È x	 '7=  A  â x *BL?  A  Æ #t	   €?	   È  r		     A  â  x    ?  A  â !r        È s	 	    $   $ !t	  €?    È  r     A  È  r      A  Ê †y    ä Ay        € ê ‰      b !r       â x   ÿàÿ Æ !r       â z  r  p`ð Æ !r       â x    pVp  È ‰       ¢ !r       Ê/ ‰	      b !r         ÊO ‰       ¢ !r
	       Ê/ ‰	
 €    b !r        ÊO ‰ €    ¢ !r
	       Ê? ‰ @    " !r       ÊO ‰	 @    b !r       Ê ‰      " !r		       Ê/ ‰ 	      ¢ !r       â M	        € ì !r       â/!r 	        ÆO  x'7=  A  â x*BL?  A  â !r        ä #t  €?   â x   ?  A  Æ  r     A  Ì s     $   $ !t  €?    È  r     A  È  r     A  Ê †y    â My        € ê x    pRð Ú G‰  P    € ê x    pRð Ú G	  `    € ê „y	     H   " z  r  p`ð â Ey       € ä „y    H   b r ÿ   pVp  Æ „y    H   ¨ „y    H   è „y    H   ( „y  
  H   ( „y €   H   ( „y €  H   ( „y €  H   ( „y €
  H   â !r		        â !r       Æ/ !r		       âO !r       Æ� !r	       â!r       È ‰       " !r       Æ ‰      b !r       È !r       â !r         â !r       È/ ‰	       ( ‰      h „y €  H   ¢ !r	 	       â !r       È/ ‰	 €    ( ‰ €    h „y €  H   â !r       âO !r	       â !r       È/ ‰ @    ( ‰ @    b !r       È� !r       â !r       â !r       È/ ‰       ( ‰      b !r        â !r	       â/ G	  À     € ê !r 	       Ê	z  i    � â !r        È  x	;ª¸¿  A  â 	z   i    � È 	z   i €  � â s	 	       $ !t	  €?    Ì s        $  r     A  È  r      A  Ê †y     ä Ay        € ê ‰       b Ey       € æ ‰      ¨ „y    H   è „y    H   ( „y    H   è „y    H   â !r         â/ !r       ÈO ‰	       h ‰      ( „y  	  H   ¢ !r       â� !r       â!r	 	       â/ !r       È ‰	 €    ( ‰ €    h „y    H   â !r       ÈO !r       â !r	       â !r       È/ ‰ @    ( ‰ @    b !r       È� !r       â !r       â x   ÿàÿ â !r       Æ/ ‰       " z  r  p`ð Æ ‰	      b x    pVp  â !r        â !r		       Ö/ G	  À     € ê !r 	       Ê	z  i    � â !r        È  x;ª¸¿  A  â 	z   i    � È 	z   i €  � â s        $ !t	  €?    Ì s	 	       $  r	     A  È  r      A  Ê †y    ä Ay        € ê ‰       b Ey  Ð    € æ ‰      ¨ „y €  H   è „y €	  H   ( „y €  H   è „y €  H   â !r         Æ/ „y €  H   â !r       ÆO ‰	       h ‰      " !r
       È� !r       â!r	 	       â/ !r       È ‰	 €    ( ‰ €    h „y €  H   ¢ !r	       â !r       È/ ‰ @    ( ‰ @    b !r       âO !r       â x   ÿàÿ â !r       Æ/ ‰       " z  r  p`ð Æ ‰	      b x    pVp  â !r
        â !r		       Ö/ G	  À     € ê !r 	       Ê	z  i    � â !r 
       È  x;ª¸¿  A  â 	z   i    � È 	z   i €  � â s        $ !t	  €?    Ì s	 	       $  r	     A  È  r      A  Ê †y    ä Ay        € ê ‰      b !r       â x   ÿàÿ Æ !r       â z  r  p`ð Æ !r       â x    pVp  È ‰       ¢ !r       Ê/ ‰	      b !r         ÊO ‰       ¢ !r
	       Ê/ ‰	
 €    b !r        ÊO ‰ €    ¢ !r
	       Ê? ‰ @    " !r       ÊO ‰	 @    b !r       Ê ‰      " !r		       Ê/ ‰ 	      ¢ !r       â M	        € ì !r       â/!r 	        ÈO 	z i    � â !r        È  x;ª¸¿  A  â 	z i    � È 	z i €  � â s        $ !t  €?    Ì s        $  r     A  È  r     A  Ê †y    â My        € ê „y    H   " z  r  p`ð Æ „y	     H   b r ÿ   pVp  Æ „y €   H   ¨ „y €  H   è „y    H   ( „y    H   ( „y    H   ( „y    H   ( „y €  H   ( „y €  H   ( „y    H   " !r       Æ „y  	  H   " !r		        Æ/ „y  
  H   b !r       ÆO „y    H   ¢ !r       Æ� „y €  H   â !r       Æ„y €
  H   " !r       Æ „y    H   " !r       Æ „y!    H   " !r		       Æ „y  €  H   " !r       Æ „y# €  H   " !r       â x   ÿàÿ ä „y% €  H   " !r       â z  r  p`ò ä „y" €	  H   â !r       â x    pVò  Ä „y' €  H   â !r       Æ/ „y) €  H   â !r	       âO !r       ä� ‰      " !r       Æ‰       b !r       Æ ‰      ¢ !r!       Æ ‰	      â !r         Æ ‰      " !r##
       Æ ‰      " !r  %       â !r       â !r"#"       â !r         â/ !r' '       â !r       ÆO ‰       " !r)")       â !r
	       ä� ‰'      b x   ÿàÿ â !r       ä‰)      ¢ z  r  p`ô â !r       Ä ‰	      â x    pVt äx   ÿàÿ â ‰      & z  r  p`ö â ‰
      & x    pVö â ‰      " !r        Æ ‰      " !r'       â/ !r)       äO ‰  €    b !r		       Æ� ‰      ¢ !r       Æ‰      â !r
       Æ ‰	 €    " !r       Æ ‰ €    " !r       Æ ‰
 €    ( ‰ €    " !r         Æ/ ‰ €    b !r       âO !r       ä� ‰  @    ¢ !r	       Æ‰ €    â !r       Æ ‰ €    " !r

       Æ ‰	 @    " !r       Æ ‰ @    " !r       Æ/ ‰
 @    h ‰ @    b !r        ÆO ‰ @    ¢ !r       â� !r       ä‰       â !r		       Æ ‰ @    " !r       Æ ‰ @    " !r
       Æ/ ‰	      b !r       Æ ‰      b !r       ÆO ‰
      ¨ ‰      ¢ !r         Æ� ‰      â !r       â !‚         â!r       Æ‰      " !r	       Æ/ ‰      b !r       â !‚       â !r
       äO !’       â  ‚      A  â !r       â !’       Æ †‰     â !r	       â� !¢       â  ’     A  ä !¢		       â !r       ä †™    â  ¢		     A  â !r       È/ †© 	   â M9        € ê !r       â !r       È  r     A  Ê †y    â My        € ê „y	     H   " z  r  p`ð â Ey  À    € ä „y    H   b r ÿ   pVp  Æ „y    H   ¨ „y    H   è „y    H   ( „y  
  H   ( „y €   H   ( „y €  H   ( „y €  H   ( „y €
  H   â !r		        â !r       Æ/ !r		       âO !r       Æ� !r	       â!r       È ‰       " !r       Æ ‰      b !r       È !r       â !r         â !r       È/ ‰	       ( ‰      h „y €  H   ¢ !r	 	       â !r       È/ ‰	 €    ( ‰ €    h „y €  H   â !r       âO !r	       â !r       È/ ‰ @    ( ‰ @    b !r       È� !r       â !r       â !r       È/ ‰       ( ‰      b !r        â !r	       â/ G	  à     € ê !r 	       Ê	x   à@  � Ê  x #ÛÙ?  A  È  x	 ;ª¸¿  A  â !r        Ê s	 	       " 	x    à@  � È 	x    àÀ  � Ê !t   €?    â !t	  €?    È s        $  r     A  È  r     A  Ê †y     ä Ay        € ê ‰       b Ey       € æ ‰      ¨ „y    H   è „y    H   ( „y    H   è „y    H   â !r         â/ !r       ÈO ‰	       h ‰      ( „y  	  H   ¢ !r       â� !r       â!r	 	       â/ !r       È ‰	 €    ( ‰ €    h „y    H   â !r       ÈO !r       â !r	       â !r       È/ ‰ @    ( ‰ @    b !r       È� !r       â !r       â x   ÿàÿ â !r       Æ/ ‰       " z  r  p`ð Æ ‰	      b x    pVp  â !r        â !r		       Ö/ G	  à     € ê !r 	       Ê	x   à@  � Ê  x #ÛÙ?  A  È  x ;ª¸¿  A  â !r        Ê s        " 	x    à@  � È 	x    àÀ  � Ê !t   €?    â !t	  €?    Ì s	 	       $  r	     A  È  r     A  Ê †y    ä Ay        € ê ‰       b Ey  ð    € æ ‰      ¨ „y €  H   è „y €	  H   ( „y €  H   è „y €  H   â !r         Æ/ „y €  H   â !r       ÆO ‰	       h ‰      " !r
       È� !r       â!r	 	       â/ !r       È ‰	 €    ( ‰ €    h „y €  H   ¢ !r	       â !r       È/ ‰ @    ( ‰ @    b !r       âO !r       â x   ÿàÿ â !r       Æ/ ‰       " z  r  p`ð Æ ‰	      b x    pVp  â !r
        â !r		       Ö/ G	  à     € ê !r 	       Ê	x   à@  � Ê  x #ÛÙ?  A  È  x ;ª¸¿  A  â !r 
       Ê s        " 	x    à@  � È 	x    àÀ  � Ê !t   €?    â !t	  €?    Ì s	 	       $  r	     A  È  r     A  Ê †y    ä Ay        € ê ‰      b !r       â x   ÿàÿ Æ !r       â z  r  p`ð Æ !r       â x    pVp  È ‰       ¢ !r       Ê/ ‰	      b !r         ÊO ‰       ¢ !r
	       Ê/ ‰	
 €    b !r        ÊO ‰ €    ¢ !r
	       Ê? ‰ @    " !r       ÊO ‰	 @    b !r       Ê ‰      " !r		       Ê/ ‰ 	      ¢ !r       â M	        € ì !r       â/!r 	        ÈO 	x  à@  � â !r        È  x#ÛÙ?  A  â 	x  à@  � Æ  x;ª¸¿  A  â 	x  àÀ  � Ê s        " !t  €?    â !t  €?    Ì s        $  r     A  È  r     A  Ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y       "   $ z   w  p`ð Ú M	        € ê y      !   " ¹z  l      â Ey  p    € â ™x?    â y      %   b ¹z  F   
   â $rÿÿ   ÿ Ž ä $rÿÿ   ÿ Ž â xÿ     È |    p`ð â $x   ÿ Ž Ø/ G	  Ð    € ê y      &   " $tÿ   ÿ Ž ä $z  x  Ž È %v \   Ž Ê �y    ™ ¢  $r
ÿÿ   ÿ Ž â x   ÿ   â $z  r  ÿŽ â x   ÿàÿ â 'z m  
 Ž È %x$   ÿ Ž â zÿ n    Æ $x   ÿ Ž ä $rÿÿ   
Ž ä %x$    Ž â x   ÿÀŽ Æ $z
 o  Ž â z Z  ÿàñ â $rÿÿ   ÿ Ž ä $z

 u  ÿŽ ä $rÿÿ   ÿ Ž â z	 [  ä  â $z q  Ž â r
   à£ â%x

$   ÿ Ž Æ rÿÿ   ÿDþ  â %x$    Ž È $x	$   ÿŽ â z Z  ÿàñ â $rÿÿ   
 Ž ä $rÿÿ    Ž â z [  	ä  â $v
ÿ X  ÿ Ž ä $vÿ Y  ÿ Ž ä $z t  Ž ÄO $z	 t  Ž ä $z q  Ž â x   ÿàñ â %x	   Ž È $rÿÿ   ÿ Ž ä %x   Ž È $rÿÿ     ä r
   ÿàñ Ê $x     Ê �y   ‘ ¨  �y!   ‘ ¨  �y   ‘ è  �y$   ‘ ( �y   ‘ h �y*   ‘ b r
   ÿàñ Æ �y   ‘ d $x     ä �y)   ‘ h �y#   ‘ h �y    ‘ h �y&   ‘ h �y   ‘ h �y'   ‘ h �y(   ‘ h �y   ‘ h �y%   ‘ d $rÿÿ    Ž Ä/ $rÿÿ    Ž Ê �yðÿÿ ™ b %x	   
Ž Æ �y"    ™ b %x   
Ž Æ �y    ‘ h �y    ‘ b r,   ÿàñ Æ �y+   ™ d $x-     ä �yôÿÿ ™ h �y,,    • b ‚x       ä ‚x  ÿþý     Ä $~-ÿ   ÿ Ž â x   ÿàÿ ä x!v      ÈO xTp      â� $~!ÿ   ÿ Ž â xww  ÿÀŽ ä x$$T      â$~ÿ   ÿ Ž â t!!    äxÿ   $  ä tüúøô    Ä x$ˆˆˆˆÿÀŽ â $~$ÿ   ÿ Ž â xww  ÿÀŽ ä x22ÿø� ä x**v      ät$$    ät--üúøô    Ä xTv  ÿÀŽ ä xÿ     ä xTp  *    ä r!       ä r$   -    ä x))T      Ä x d      äx1u      ä x)ˆˆˆˆÿÀŽ ä x)ÿ   )  ä x# v  #    ä x)ww  ÿÀŽ Ä x)&Tp  #    â $~$ÿ   ÿ Ž â x*ww  ÿÀŽ â $~*ÿ   ÿ Ž â x&T  )    â $~!ÿ   ÿ Ž â x22ÿø� â $~-ÿ   ÿ Ž â t$$    ä xÿ   &  Ä t*üúøô    ä x Tv  ÿÀŽ ä t!!    â$~)ÿ   ÿ Ž â t-üúøô    ä x-#ww  ÿÀŽ â $~#ÿ   ÿ Ž â x&&ˆˆˆˆÿÀŽ Ä xww  ÿÀŽ ä r$        â $~ ÿ   ÿ Ž â x((v  '    â $~$ÿ   ÿ Ž â x&&22ÿø� ä t))    ät#üúøô    Ä x#ÿ     ä x*Tp  (    ä t  -    äx&Tv  ÿÀŽ ä t-$üúøô-    ä x&ÿ   &  Ä x%%T  *    ä r$!#       â &r   ÿ   â r    -    ä r&)&       ä x%ˆˆˆˆÿÀŽ ä x d  $    Ä x1u  $    â $~ÿ   ÿ Ž â x  d  &    â $~!ÿ   ÿ Ž â x((ww  ÿÀŽ ä x22ÿø� â &r    ÿ   â t(    ät(!üúøô(    Ä x Tv  ÿÀŽ ä x%ÿ   %  â $~$ÿ   ÿ Ž â x1u  &    â $~&ÿ   ÿ Ž â r    (    â $x €   ÿ Ž â x%%ww  ÿÀŽ â $x!€   ÿ Ž ä &r"      â t$%    Ä t$&üúøô%    ä r ÿ   pRò ä r ÿ   pRö ä x  €  ÿÀŽ ä x!!€  ÿÀŽ ä xÿ     Ä x ÿ   pRð ä x ÿ   pRô ä x  @     €  ä x!!@     € ä r   $    â &r""      â x  ÿ       Ä x!!ÿ      ä x d      â&r      âx1u      ä x ÿT       â &r   "   â x!ÿT  !    ä |    p`ð â 0rÿ,    A   â  x     ?  A  â  x!   ?  A  â &r+      Ä &r+      â  r       A  â  r     A  â x@  ÿàó ä x

  ÿàõ ä x@  ÿà÷ ä Er        ä Er        â $rÿÿ   Ž Ä #r       â $rÿÿ   Ž  â #r      â $rÿÿ    â G‰  Ðõÿÿÿÿƒ ê Ay        € ê ‰      ( ‰      b !r       â $x   Ž â !r	       Æ/ ‰      " z  p  p`ð Æ ‰	      b x    pDp  â !r       â !r
	       È/ ‰ €    ( ‰
 €    b !r       â !r
       È/ ‰ @    ( ‰ @    b !r       â !r       È/ ‰      h  ‰      ¢  M	        € ê y      &   â $x   ÿ Ž â !r       â/ $z   s  Ž â !r       âO $tÿ   ÿ Ž â r ÿ   p"ð Ä x    p"ò Ö $ÿÿ    Ž ä $ÿÿ    Ž ä $z  v   Ž È� %v  j   Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      "   $ z  w  p`ð Ú M	        € ê y      !   ( y	      &   b ¹z  l      â $tÿ   ÿ Ž â ™x?    â Ey       € â ¹z  F   
   â x#ÿ     È | #   p`ð â $z x  	Ž È/ %v \   Ž â x      ÿ  Æ $rÿÿ   ÿ Ž â �y     ™ b $rÿÿ   ÿ Ž ä $rÿÿ   ÿ Ž ä $rÿÿ   ÿ Ž â G	  0    € ê $rÿÿ   ÿ Ž â y      %   " $z
 r  ÿŽ â x   ÿ   â'z	 m   Ž â Ey °    € æ $x   ÿ Ž â zÿ n    â $rÿÿ   ÿ Ž Ä $rÿÿ   ÿ Ž â x   ÿÀŽ â $rÿÿ   
Ž ä $vÿ `  ÿ Ž ä $z o  	Ž ä %x
$   ÿ Ž È $z u  ÿŽ ä %x$    Ž Æ r   
à‘ â%x
$   ÿ Ž â z Z  ÿàõ ä rÿÿ   ÿ$~  â %x$    Ž â z	 [  ä Æ $r	ÿÿ    Ž â z Z  ÿàñ â $x   ÿ Ž ä $x$   ÿŽ â x   ÿàó â $z q  #Ž ä $vÿ X  ÿ Ž â z [  ä  â $z  t  Ž âz ÿ `  pPð â $vÿ Y  ÿ Ž Ä $vÿ a  ÿ Ž â z ÿ a   Sð â $rÿÿ   ÿ Ž ä %x   Ž È $rÿÿ   Ž  È r   ÿàó â �y    ™ ¨ $x   Ž  Ê �y%   ‘ è  �y2   ‘ è  �y(   ‘ ( �y$   ‘ b xÿ     Æ �y-   ‘ ¨  �y0   ‘ ¢  x   ÿàÿ â $z  t  Ž ä �y1   ‘ ¤  $z q  Ž Ä �y/   ‘ ¤  %x   Ž ä �yðÿÿ ™ ¦ r   ÿàó â �yôÿÿ ™ ¨ $x   Ž  â�y   ™ ¨ �y+   ‘ ¨  �y.   ‘ ¨  �y)   ‘ ¨  �y&   ‘ ¨  �y'   ‘ ¨  �y,   ‘ ¨  �y"   ‘ ¢  %x    Ž Æ �y*   ‘ ¨  �y      ‘ ¢ %x   Ž Ê �y3    ‘ ¢  r4
   ÿàó Ê $x5	   Ž  Ê �y44    • ¢ ‚x       ä ‚x  ÿþý     â x##   ÿàÿ ä x%2v  %    È� x((Tp  %    â$~2ÿ   ÿ Ž â x%%ww  ÿÀŽ ä x$$T  (    â$~(ÿ   ÿ Ž â t2%    äx!ÿ   $  ä/ t(üúøô%    Ä x%$ˆˆˆˆÿÀŽ â $~$ÿ   ÿ Ž â x!!ww  ÿÀŽ â $~(ÿ   ÿ Ž â x00v  -    äO x%%22ÿø� ä t-$üúøô!    ä x$1Tp  0    Ä t((!    ä x!%Tv  ÿÀŽ ä x//T  $    ä x%ÿ   %  ä r!       ä x!ÿ   /  Ä r(%   -    â $~$ÿ   ÿ Ž â x-/ˆˆˆˆÿÀŽ â $~1ÿ   ÿ Ž â x/!ww  ÿÀŽ â $~(ÿ   ÿ Ž â x00ww  ÿÀŽ â $~%ÿ   ÿ Ž â x--22ÿø� ä t$$/    Ä t!10    ät/(üúøô/    ä t0%üúøô0    ä x(-Tv  ÿÀŽ ä x-ÿ   -  ä x% d      Ä x1u      ä r!!(   0    â &r%%   ÿ   â r$$-   /    ä x..v  +    â &r%   %   â x! d  $    ä x))Tp  .    Ä x$!1u  $    â &r   %   â x)&T  )    â $~!ÿ   ÿ Ž ä &r&$      â x)ˆˆˆˆÿÀŽ â $~%ÿ   ÿ Ž â x..ww  ÿÀŽ ä x',v  '    Ä x22ÿø� ä t!.    äx)ÿ   )  ä t!%üúøô.    â $~(ÿ   ÿ Ž â x%"Tp  '    â $~,ÿ   ÿ Ž â x$Tv  ÿÀŽ Ä x))ww  ÿÀŽ ä x%*T  %    ä r$   !    â $~$ÿ   ÿ Ž â t"()    â$~(ÿ   ÿ Ž â t),üúøô)    ä xÿ     Ä x''ww  ÿÀŽ ä x!%ˆˆˆˆÿÀŽ ä x%ÿ   %  ä r"   )    â $~)ÿ   ÿ Ž â t$$'    â$~+ÿ   ÿ Ž â t'(üúøô'    Ä x"!22ÿø� ä x(%ww  ÿÀŽ ä x 3ÿ   pRò är 3ÿ   pRô â$x33€   ÿ Ž â x  ÿ   pRö är  ÿ   pRø â$x  €   ÿ Ž â x! d      Ä x1u      ä x%"Tv  ÿÀŽ ä t)(    ät(+üúøô(    ä x"ÿ   "  ä x33€  ÿÀŽ Ä x  €  ÿÀŽ â &r!!   ÿ   â r$$%   '    ä r"   (    ä x33@       ä x  @       â &r!   !   â x$ d      Ä x33ÿ    €  ä x  ÿ    € â &r   !   â x$$1u      ä x3ÿT  3    ä x ÿT       â 0r4ÿ4    A   ä &r$$      â  x%3   ?  A  â  x    ?  A  â | #   p`ò Ä  r%%4     A  â  r"4     A  â Er& &       ä Er$ $       â G‰       € ê xÿ     Ê $z q  Ž È $z  t  Ž È %x   Ž Ê r    ÿàõ Ê $x!    Ê �y+     ¨  �y,     ¨  �y.     è  �y-     ( �y*     h �y)     h �y'     h �y(     b %x   Ž Ì �y     b $~0ÿ   ÿ Ž ä $~/ÿ   ÿ Ž ä $~!ÿ   ÿ Ž â x+,v  +    ÈO x..Tp  +    ä� x++ww  ÿÀŽ ä x--T  .    â$~.ÿ   ÿ Ž â t 0üúøô+    ä x,ÿ   -  ä x--ˆˆˆˆÿÀŽ Ä x,,ww  ÿÀŽ ä x--22ÿø� ä t.+    â/ $~.ÿ   ÿ Ž â x*)v  *    ät+/,    ät,!üúøô,    Ä x!-Tv  ÿÀŽ ä x-ÿ   -  ä x''Tp  *    ä r !        ä r++-   ,    â $~-ÿ   ÿ Ž â x'(T  '    â $~,ÿ   ÿ Ž â x  d  +    Ä x!'ˆˆˆˆÿÀŽ ä x  1u  +    â $~+ÿ   ÿ Ž â x'ÿ   '  â &r   ÿ   â x**ww  ÿÀŽ ä x!!22ÿø� â &r       â x)'ww  ÿÀŽ Ä t'+*    ät*-üúøô*    ä x(!Tv  ÿÀŽ ä t+,)    ät,.üúøô)    ä x!ÿ   !  Ä r''(   *    â $x(€   ÿ Ž âr ÿ   pRö är,+!   ,    ä x ÿ   pRô ä x((@     € ä x' d  ,    Ä x((ÿ      ä x,'1u  ,    â &r      â x(ÿT  (    Æ &r,      ä  x(   ?  A  Æ Er        â  r4     A  È #r      â #r%&      â #r"$      â G‰  @    € ê r   ÿàõ Ê $x    Ê �y$    ¨ �y'    ¨ �y)    è �y&    ( �y%    h �y"    h �y     h �y!    b %x   Ž Ì �y     b $~+ÿ   ÿ Ž ä $~-ÿ   ÿ Ž â x$'v  $    âO $~'ÿ   ÿ Ž Æ x))Tp  $    ä� x$$ww  ÿÀŽ ä x&&T  )    â$~)ÿ   ÿ Ž Æ xÿ   &  ä x&&ˆˆˆˆÿÀŽ ä xww  ÿÀŽ ä x&&22ÿø� ä t+    ät(-üúøô    Ä x%"v  %    ät'$    äx&Tv  ÿÀŽ ä t$)üúøô$    ä x&ÿ   &  ä x  Tp  %    Ä r   $    â $~$ÿ   ÿ Ž â r&   (    â $~&ÿ   ÿ Ž â x !T       ä x d      äx1u      ä x ˆˆˆˆÿÀŽ â &r   ÿ   â x ÿ      Ä x%%ww  ÿÀŽ â &r      â x22ÿø� ä x" ww  ÿÀŽ ä t$%    ä/t &üúøô%    ä x!Tv  ÿÀŽ Ä t$'"    ät%)üúøô"    ä xÿ     ä r!        â $x €   ÿ Ž âr ÿ   pRö ä r$$   %    Ä x ÿ   pRô ä x  @     € ä x d  $    äx  ÿ      ä x$1u  $    â &r      â x ÿT       Æ &r$      ä  x    ?  A  Æ Er        â  r44     A  È #r4      â x

@  ÿàõ ä x  ÿà÷ ä x@  ÿàù â $r	ÿÿ   	 â x  ÿàû â $rÿÿ   Ž ä $rÿÿ    ä $rÿÿ   Ž â D       À â Gy  pìÿÿÿÿƒ ê Ay       € ê $rÿÿ    Ž ä $rÿÿ    Ž Ä Ay        € ê z ÿ `  pPð È z ÿ a   Sð Ú G	  P    € ê ‰      ( ‰      b !r       â !r       È/ ‰      ( ‰      b !r       â !r	       È/ ‰ €    ( ‰
	 €    b !r       â !r
	
       È/ ‰ @    ( ‰
 @    b !r       â !r
       È/ ‰      ( ‰      b !r       â !r       â/ Gy  €    € ê ‰	      ( ‰      h ‰      ¨ ‰
      â !r		       â !r       â/ !r       ÈO ‰      " !r
       Æ� ‰
	      h ‰      ¨ ‰      â !r       â !r
	
       È/ ‰ €    " !r       ÆO ‰
 €    b !r       Æ� ‰ €    ¨ ‰ €    â !r       â !r
       â/ !r       ÈO ‰ @    " !r       Æ� ‰ @    h ‰ @    ¨ ‰ @    â !r       â !r       â/ !r       ÈO ‰	      â !r
       Æ� ‰      ( ‰      h ‰
      ¢ !r       â !r       â/ !r	       â !r
       ÄO y      %   $ $x   Ž Ê z  p  p`ð È x    pDp  Ú M	        € ê z ÿ ^  pPð â $z  v  Ž Æz ÿ _   Sð Ú $ÿ   ÿ Ž È % ^   Ž Ì �	     ¢ $x    ÿ Ž Ê r  ÿ   p"ò äx     p"ô äx     p"ö äx     p"ø Î $ ÿÿ    Ž â z ÿ `  pPò â $" ÿÿ    Ž ä $2 ÿÿ    Ž â z ÿ a  Sò â $B ÿÿ    Ž È !         ÐO G™       € ê z ÿ b  pPð È z ÿ c   Sð Ú $ÿ   ÿ Ž È % b   Ž Ì �	     ¢ $x   ÿ Ž ä $vÿ h  ÿ Ž Æ x øÿÿÿp"ò äx üÿÿÿp"ô är ÿ   p"ö äx    p"ø Î $ÿÿ    Ž â x    pBò â $"ÿÿ    Ž ä $2ÿÿ    Ž ä $Bÿÿ    Ž È !       ÈO G  P    € ê x    pRð Ú G‰  �     € ê x    pRð Ú G	  ð    € ê  x;ª¸¿  A  Ì s        $ !t  €?    Ì s        $  r     A  È  r       A  â Gy  P    € ê  x'7=  A  â x*BL?  A  Æ #t  €?   â  x   ?  A  Æ  r     A  Ì s     $   $ !t  €?    È  r     A  È  r       A  â Gy  °    € ê x    pRð Ú G‰  à     € ê x    pRð Ú G	        € ê 	z i    � ä 	z   i    � Æ  x;ª¸¿  A  â 	z   i €  � Ê s        $ !t  €?    Ì s        $  r     A  È  r       A  â Gy  Ð     € ê  r       A  â Gy  °     € ê 	x  à@  � ä 	x    à@  � Æ  x#ÛÙ?  A  â 	x    àÀ  � Æ  x;ª¸¿  A  ä !t   €?    È s        $ !t  €?    Ì s        $  r      A  È  r       A  ä y      "   ( y      &   b $z s  Ž ä $tÿ   ÿ Ž ä $z v  Ž È/ %v j   Ž Ê †y      â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y       "   " ¹z  l      â Ey      € â ™x?    â y      !   b ¹z  F   
   â x      ÿ  â r ÿ       â y      %   ¢ r% ÿ       â x      ÿ  â r
 ÿ       â x      ÿ  â r4 ÿ       â x      ÿ  â r ÿ       Ä r: ÿ       ä r ÿ       ä r@ ÿ       ä r  ÿ   pRð är  ÿ   pRò är    ÿ(Ž È/ xÿ     ä x   ÿ   äO r   ÿ� È xÿ     È |    pbô Ú G)  p    € ê y      '   " r ÿ       ä r ÿ       â y      &   b x   ÿàÿ ä z  Y      â 'z y   Ž È 'z s   Ž â/ zÿ z    ä xüÿÿÿÿÀŽ â $z p  Ž â zÿ t    â $z |  Ž â r  €ÿàÿ â $z }  ÿŽ â r ÿ       â $z v  Ž â x   ÿ   â $z w  ÿŽ â x(   ÿ   â $z |  Ž â x&ÿ     â $z p  Ž â x#   (áÿ â %x((   ÿŽ â z  q      Ä x*ÿ   #  â $z v  Ž â r   ÿàõ â $x   Ž â x   &  â %x&"   (Ž â x!#   *  ä z  X      â $x   Ž âr$   ÿŽ â$x    Ž âr   ÿŽ â$x"   Ž â z q  ÿàÿ â %x("   (Ž â x   ÿ   Ä rÿÿ   ÿä ä x##   ÿ   ä r6&   ÿàõ â%x,"   Ž âr0(   ÿà÷ ä r7'   ÿä â %x."   Ž â r   ÿàõ â �yE,    • ¢  x* $       Ä r+ÿ   ÿä â �yI6   • â r   ÿàõ Æ �y26   • â %v, Z  * Ž Æ �y6   • " $x+$   ÿŽ â r+ÿ   ÿä ä �yB6   • & $x++$   ÿŽ â r1)   ÿäÿ â �yD.    • b %v6 Z  * Ž â/ r9   ÿàõ Ä �yG0   • b r=$   ÿà÷ ä r77+   ÿàÿ â �y.0   • b rX6   ÿàù är--   ÿàÿ â �yH0   • b rÿ   ÿä ärÿ   ÿäÿ â �y30   • b rN#6   ÿàý Ä rY7   ÿä ä rZ#,   ÿàû â %v89 Z  * Ž â rP,   ÿà÷ â �yAX   ™ b rO7!   ÿä â $x$   ÿŽ â rQ-   ÿäÿ â�y0,    • b! r[-!   ÿäÿ â %v<= Z  * Ž â r99   ÿàÿ â �yCN   ™ d $x$   ÿŽ Ä �y/6    • b r-   ÿàû ä r==   ÿàÿ â �y1P   ™ b rÿ   ÿäÿ ä rT<   ÿàõ â �yFZ   ™ b rK   ÿàû ä rU=   ÿä Ä rL#<   ÿà÷ â %v,- Z  * Ž â r+ÿ   ÿäÿ â �y?T   ™ b r68   ÿàù â/ $x$   ÿŽ â rM=!   ÿäÿ â %vJK Z  * Ž â �y<    • f $x++$   ÿŽ â r--   ÿàÿ â �y>8    • b rP#,   ÿà÷ Ä r79   ÿä â �y,    • b rKK+   ÿàÿ ä rNJ   ÿàù â�y<L   ™ b# rR,   ÿàõ ä rQ-!   ÿäÿ â �y;6   ™ b rOK   ÿä Ä r+"   ÿàù ärM    ÿà÷ â/ �y5N   ™ b rV#8   ÿàý ä rS-   ÿä â %vLM Z  * Ž â r7ÿ   ÿäÿ â�y6P   ™ b r-ÿ   ÿä Ä rW9!   ÿä â�y8J    • b rT#J   ÿàõ â $x77$   ÿŽ ä �y9R   ™ b %v*+ Z  * Ž â rUK!   ÿä ä �y=V   ™ b $x--$   ÿŽ â rMM7   ÿàÿ Ä rP*   ÿà÷ â�y7T   ™ b rRL   ÿàõ ä/r++-   ÿàÿ ä rSM   ÿä ärQ+   ÿäÿ ä rN#L   ÿàõ â�yJR   ™ b r,#*   ÿà÷ Ä rOM!   ÿä â�yKP   ™ b r-+!   ÿäÿ Æ �yLL    • h �yNN   ™ h �y**    • h �y,,   ™ b x   ÿàÿ È |    pbô ä x   ÿà÷ ä x   ÿàÿ ä x$$   ÿàÿ ä x   ÿàÿ ä x   ÿàÿ Ä x   ÿàÿ ä x     ÿàÿ ä x""   ÿàÿ ä rÿ   ÿäÿ ä x2IT  2    â� 0rEÿE    A   âO xT  B    â0rDÿD    A   âx.GT  .    Ä x3HT  3    â &rB2A   ÿ   ä 0r0ÿ0    A   ä &r+C   B   ä 0r/ÿ/    A   â  r-D0     A  â/ &rH.A   ÿ   â  rA0E     A  â Er+ +       â &rG21   ÿ   â  r0D/     A  â &rI.1   ÿ   Ä &r1F   G   â #r0+      â &r+2?   ÿ   ä&r?.?   ÿ   â Er1 1       â 0rÿ    A   È #r@-1   @   â  r-D     A  â rE     A  â &r?3<   ?   Ê Er? ?       â &r+<   +   È #r?      â &r02;   ÿ   âEr+ +       â 0rÿ    A   ä 0r>ÿ>    A   ä &r29   ÿ   ä&r9.9   ÿ   ä &r0=   0   Ä &r936   9   â #r:-+   :   â  r+D     A  âEr0 0       â &r6      â Er9 9       â  rE     A  â  r<D>     A  â &rH3C   H   â Er        â #r
9   
   â #r<0      â &rB.;   ÿ   Ä &r2J   ÿ   ä&r025   ÿ   ä&r<.5   ÿ   ä&rJ.J   ÿ   ä &r22K   ÿ   ä&r..K   ÿ   â #r4+   4   â 0r8ÿ8    A   â ErH H       â 0rLÿL    A   â  r/E/     A  â &rF3F   I   Ä &rB3=   B   ä &r07   0   ä&r<37   <   ä &rN      ä&rJ3N   J   ä 0rÿ*    A   Ä &r2,   2   ä&r.3,   .   â  r6D8     A  â r*DL     A  âErF F       â #r/H      â ErB B       â  rDD     A  â Er0 0       â  r/E>     A  âEr< <       â  r-E8     A  âEr        â  rEL     A  âErJ J       â  rE     A  â Er2 2       Ä Er. .       â #rAF      â #r/B      â #r	60   	   â #r-<      â #r%*   %   â #rJ      â #rD2      â #r.      â G©  Pòÿÿÿÿƒ ê Ay        € ê Ey  0    € â G‰      € ê r    ÿHŽ Ê ˆs @ øÿ H   è ˆs €øÿ H   è ˆs  ùÿ H   è ˆs €ùÿ H   è ˆs : úÿ H   è ˆs €úÿ H   è ˆs  ûÿ H   è ˆs €ûÿ H   è ˆs 4 üÿ H   è ˆs 
€üÿ H   è ˆs 	 ýÿ H   è ˆs €ýÿ H   è ˆs % þÿ H   è ˆs €þÿ H   è ˆs  ÿÿ H   è ˆs €ÿÿ H   ä Ay        € ê {          ì M        € ê „y     H   " z  r  p`ð Ä x*   ÿàÿ â „y €   H   b r ÿ   p"p Æ „y €  H   ¨ „y    H   è „y    H   ( „y    H   h „y €  H   h „y' €  H   è y      '   ( „y#    H   â !r@       Æ „y) €  H   â !r       Æ/ ‰       " !r       ÆO „y! €  H   â !r       Æ� ‰      b !r        Æ‰      ¢ !r:       Æ‰       â z  r      â !r       ä „y$    H   â r   ÿŽ â !r        ä „y(    H   è ‰      " !r       Æ/ ‰      b !r       ÆO y       &   ¢ !r         Æ� „y+ €  H   â !r!       Æ „y,    H   è ‰!      â !r       â !r       È/ ‰      " !r$$	       â $z   x  Ž äO ‰      d $z  ~   Ž ä ‰ €    ¢ !r((%       ä     ÿàù â ‰      " r    ÿ† â !r!       â� ÿÿ   ÿä â ‰%      â 
  r  ÿàó Ä 
 j  ÿˆ â ‰/(      â r"ÿÿ   ÿäÿ â !r       â z j  ÿ† â $x   ÿŽ â &ÿÿ   ÿäÿ  â !r       â/ ‰ €    " 
 j  ÿ„ â !r,,       â 
 k   â !r'
       â ‰      b z k  "� â !r"#4       â 
 k  & â !r&)       â ‰#       b !r       âO !r       âz * r  p`ò â ‰)      ¢ !r*+       â r   ÿŽ â !r!%       â� ‰'"      â r    ÿàù Ä     ÿà÷ â ‰ @    " r.ÿÿ   ÿä â !r       â z j  ÿˆ â „y €  H   â 0ÿÿ   ÿäÿ â !r       ä/ ‰      " 

 j  ÿ† Ä z k  . â ‰      b !r  #       â
 k  0� ä ‰+$      b !r#)       âO x    p"ò ä ‰-&      ¢ !r""'       â� r	    ÿ„ Ä ‰1*      â !r       âr  €ÿŽ ä ‰3,      " r2ÿÿ   ÿä ä z	 j  ÿ„ â ‰ €    " !r       â     ÿàù â !r       â/ ‰) @    " z		 k  2 â !r$$+       Ä‰      b    ÿàÿ â !r%&-       âO !r&(/       â !r       â ‰.#      ¢ !r'*1       â� 2ÿÿ   ÿä ä ‰* €    â !r(,3       â
 j  ÿˆ Ä ‰,!      " !r       â     ÿà÷ ä ‰1"      b 
 k  2 â !r)       â 6ÿÿ   ÿäÿ â ‰+ €    "    ÿàÿ â !r       â/  j  ÿ† â ‰3$      b     ÿàõ Ä  k  6� â ‰/  €    b !r##.       âO 8ÿÿ   ÿä ä ‰0%      ¢ !r*       â�  j  ÿ„ ä ‰5&      â !r,!,       â   ÿàÿ Ä ‰4'      " !r)"1       â k  8 ä ‰7(      b !r+       â     ÿàõ ä ‰      " !r*$3       â/ $ÿÿ   ÿä Ä ‰- €    b !r /       â !r2%0       äO ‰1, €    ¢ !r &5       Æ� ‰.) €    â !r!'4       Æ‰0# €    " !r"(7       Æ‰+ @    b !r%       Æ ‰3* €    "  j  ÿ„ â!r-       â/ x-        â ‰52 €    b  k  $ â !r$,1       äO ‰& @    ¢ %v  j  - Ž â !r.).       Ä� ‰7  €    â !r0#0       Æ‰4! €    " !r+       Æ‰9" €    b !r3*3       Æ ‰6%      ( ‰'      " !r525       Æ/ ‰( @    b !r&       ÆO ‰- @    ¢ !r7 7       Æ� ‰/ @    â !r4!4       Æ‰)$ @    " !r9"9       â!r6%6       ä ‰#. @    " !r1'       Æ ‰+0 @    b !r((       Æ/ ‰'      b !r-       ÆO ‰&3 @    ¢ !r//       Æ� ‰*5 @    â !r)$)       Æ‰      ( ‰ 7 @    " !r#.#       Æ ‰!4 @    " !r+0+       Æ‰%6 €    b !r''       Æ/ ‰"9 @    b !r&3&       ÆO †	     â !r*5*       Æ� ‰(      â !r       Æ‰      ¢ !r 7        Æ ‰/      â !r!4!       Æ † 1   â !r%6%       Æ‰)      " !r"9"       Æ/ ‰#      h ‰+      ( †	 '    â !r       ÆO ‰&      ¢ !r//       Æ� ‰*      â !r'(       Æ†     â !r))       â !r##       ä/ ‰       " !r++       Æ‰!      h ‰% @    " !r&       ÆO ‰"      ¢ !r*       Æ� †	 '    è †    è †	 /    â !r        â !r!!       Ä/ † )    â !r%       Æ†	 #    â !r"       ÆO † +   è †	 
    è †     è †	     è † !   è ‰      ( †	     â M™        € ê x   ÿàÿ â !r       ÆŸ r    ÿàñ È r ÿÿ   ÿä  ä z j  ÿ€ È z k     Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y       "   " ¹z  l      â Ey  °    € â ™x?    â y      !   b ¹z  F   
   â r ÿ       â x&      ÿ  â y      %   ¢ r ÿ       â x2      ÿ  â r( ÿ       Ä r ÿ       ä r1 ÿ       ä r ÿ       ä r ÿ       ä r ÿ       ä r: ÿ       Ä r  ÿ   pRð är  ÿ   pRò är    ÿ(Ž ä/ x   ÿ   äO xÿ     È r   ÿ� ä r ÿ       ä xÿ     È |    pbô Ú G)      € ê y      '   " r ÿ       ä r ÿ       â y      &   b xüÿÿÿÿÀŽ â $z p  Ž â z  X      ä r  €ÿàÿ ä z q  ÿàÿ Ä x   ÿ   äx"   ÿ   È x   "áÿ â %x""   ÿŽ È 'z	 y   Ž â r ÿ       Æ 'z
 s   Ž â/ z	ÿ z  	  Æ $z }  ÿŽ â z
ÿ t  
  â $z	 |  Ž ä $z w  ÿŽ ä $z		 |  Ž â x   ÿàÿ â $z
 v  Ž â r   ÿàõ â $z
 v  	Ž â z	  q      â $z p  Ž â xÿ     â %x "   "Ž â x
ÿ     Ä x     â $x	   Ž â x   
  â $x	   Ž âz  Y      â $x	   Ž âr	   ÿŽ â%x""   "Ž â r	   ÿŽ ä x   ÿ   Ä rÿÿ   ÿä ä x   ÿ   ä r$    ÿàõ är*        ä r%!   ÿä ä rH   ÿàõ ä r+        â �y$   • ¢  r	ÿ   ÿä Ä r
   ÿàõ â %x,"   *Ž âx? $       â �y=$   • â  r4"   ÿà÷ â $x	$   ÿŽ â r	ÿ   ÿä â %x6"   *Ž â r5#   ÿäÿ â �y+$   • ¤  %vHH Z  ? Ž â rB   ÿà÷ â�y.$   • ä  $x		$   ÿŽ â rD   ÿàõ â �y<,    • " rII   ÿàÿ Æ �y@4   • " %v$
 Z  ? Ž â rÿ   ÿäÿ ä �y>4   • " rFH   ÿàû ä r%%	   ÿàÿ â %vBB Z  ? Ž â r	ÿ   ÿä â �y-4   • "+ rJ$   ÿàõ â $x$   ÿŽ â rGI   ÿäÿ â�y/4   • " rK%   ÿä Ä r8$   ÿà÷ â �y;F   ™ " rCC   ÿàÿ â %vDD Z  ? Ž ä �y)H    • " r4H   ÿàý ärLB   ÿàù â �yJ   ™ " r5I   ÿä Ä r9%   ÿäÿ â $x		$   ÿŽ â �y*6    • " rHB   ÿàû ä/ rMC   ÿä â �y
B    • " rJ   ÿà÷ Æ�y,4   ™ " rEE	   ÿàÿ Æ �y6$    • " rB   ÿàù â/ %vJJ Z  ? Ž â rFD   ÿàõ â�y98   ™ " r4D   ÿàý ärIC   ÿäÿ â �y	D    • " r%ÿ   ÿäÿ är$ÿ   ÿä Ä r5E   ÿä â $x%%$   ÿŽ â rGE   ÿä â %vBB Z  ? Ž â �y8L   ™ &! $x$$$   ÿŽ â rKK%   ÿàÿ â �y54   ™ ( �y7H   ™ " rLJ   ÿà÷ Ä r%C$   ÿàÿ â �y0J    • " rMK   ÿäÿ ä r$ B       â�y4F   ™ "# rHB   ÿàõ ärDJ   ÿàù â�yAL   ™ ( �y$$    • " rFB   ÿàû Ä/ rB   ÿà÷ È rCÿ   ÿäÿ ä rEK   ÿä â %vJB Z  ? Ž È $xCC$   ÿŽ â rI%   ÿä â �yDD   ™ " rBJ   ÿàõ ärKKC   ÿàÿ â �yHH   ™ " rLJ   ÿà÷ ärCK   ÿä â �yJJ    • b rG%   ÿäÿ Ä rMK   ÿäÿ â �y%B   ™ h �yFF   ™ h �y?L   ™ b x   ÿàÿ È |    pbô ä x   ÿà÷ ä x   ÿàÿ ä x   ÿàÿ ä x   ÿàÿ ä x   ÿàÿ Ä x   ÿàÿ ä x   ÿàÿ ä rÿ   ÿäÿ ä xT  +    äO x==T  .    â� 0r<ÿ<    A   âx@@T  -    Ä x/>T  /    â &r+;   ÿ   ä&r;@;   ÿ   ä 0r)ÿ)    A   Ê  r-<)     A  â 0r*ÿ*    A   ä &r+=,   +   ä&r;/,   ;   â  r))*     A  ä Er+ +       ä Er; ;       â &r,   ÿ   Ä #r:-+   :   â 0r+ÿ6    A   â #r);      â &r)=9   ,   ä &r.@   ÿ   ä 0rÿ
    A   â Er) )       â &r68   ÿ   ä&r8@8   ÿ   â  r
<+     A  â &r./9   .   Ä &r8/7   8   â #r3
)   3   â &r)5   ÿ   â  r
*     A  â &r5@5   ÿ   â Er8 8       â 0r	ÿ	    A   â Er. .       â &r)=4   )   â  r+*+     A  â #r1
8   1   â &r5/4   5   Ä &r
A   ÿ   â  r,<     A  âEr) )       â #r+.      â  r<	     A  â Er5 5       â 0r+ÿ0    A   â  r	*	     A  â #r2)   2   â &r6=7   6   ä #r	5      â &rA@A   ÿ   â  r	<+     A  â &r
=D   
   Ä &rH   ÿ   Æ Er
 
       â &rH@H   ÿ   â Er6 6       â &rA/D   A   ä &r%   ÿ   ä
&r@@%   ÿ   â #r(	
   (   â &r)=F      ä 0rÿ$    A   Ä &rH/F   H   ä &r=?      ä0r	ÿJ    A   ä &r@/?   @   â  r
<     A  â #r,6      â ErA A       â  r<<	     A  â Er) )       â  r,*+     A  âErH H       â  r*     A  âEr        â  r	*	     A  â Er@ @       â #r,A      â #r'
)   '   â #r&H   &   â #r<      â #r	@      â G©   óÿÿÿÿƒ ê Ay        € ê Ey      € â G‰  ð     € ê $x	 À  Ž Ê ˆs 	: ùÿ H   è ˆs 	€ùÿ H   è ˆs 	3 úÿ H   è ˆs 	€úÿ H   è ˆs 	 ûÿ H   è ˆs 	1€ûÿ H   è ˆs 	2 üÿ H   è ˆs 	€üÿ H   è ˆs 	( ýÿ H   è ˆs 	€ýÿ H   è ˆs 	' þÿ H   è ˆs 	&€þÿ H   è ˆs 	 ÿÿ H   è ˆs 	€ÿÿ H   ä Ay        € ê {          ì M        € ê „y	     H   " z  r  p`ð Æ „y    H   b r ÿ   p"p Æ „y €   H   ¨ „y    H   è „y
 €  H   ( „y €  H   h „y    H   h „y €  H   h „y    H   è „y €  H   â !r		:       Æ „y    H   â !r2       Æ/ ‰ 	      " !r       ÆO „y €  H   â !r33       Æ� „y    H   â !r1
1       Æ„y €  H   â !r       Æ‰      b !r       Æ ‰      ¢ !r       â !r	        ä ‰3      " !r((       Æ y       &   è ‰1      " !r''       Æ y      '   b !r       Æ/ ‰
      b !r       â !r       äO ‰	      ¢ !r&       Æ ‰      ¢ !r3       â !r       â $z   x  Ž â� ‰      " !r1       Æ‰      â $z  ~   Ž â!r       ä ‰      " !r

       â/ 
  r  ÿàõ ä ‰      b !r		       âO x   ÿàÿ Ä ‰      ¢ ÿÿ   ÿä â !r       â z  r  p`ò â ‰(      b 
 j  ÿ„ äx    p"ò â ‰      b 
 k   â !r       â !r"       â� ‰'      " z  r      Æ ‰
      â !r       âr   ÿŽ ä ‰      " !r!       Æ/ ‰      b !r       âO     ÿà÷ ä ‰       ¢ !r((       âr   ÿŽ ä‰&      b !r       â r    ÿˆ Ä ‰	 €    b !r'       â r.ÿÿ   ÿä â !r
       â� ‰ €    " z
 j  ÿˆ â !r       ä‰!      â z k  . â !r       Ä/ ‰ €    b r    ÿ„ â !r$        äO ‰#(      ¢  ÿÿ   ÿäÿ â !r&&       ä‰%"      " 
 j  ÿ† â !r	       Ä ‰ €    b 	    ÿàû ä r*ÿÿ   ÿä â ‰ €    b !r       â z j  ÿ„ ä ‰'      " !r!!       â� 
 k   � Ä ‰ €    â !r       â/ ,ÿÿ   ÿäÿ ä ‰)      b !r(#       âO 
	 j  ÿŠ ä ‰ €    ¢ !r""%       âz k  * Ä ‰+$      " !r       â
		 k  ,� ä ‰-&      b !r       â    ÿàÿ ä ‰ @    b !r '       â 0    ÿà÷ â !r       â� ‰(! €    $ 1ÿÿ   ÿäÿ â !r)       â/ ‰* €    b 0 j  ÿ† â !r       ÄO ‰/  €    ¢    ÿàÿ â !r$+       ä‰ @    â !r&-       Æ‰$ @    " !r       Æ ‰-" €    b $x   ÿŽ â   ÿàÿ ä ‰& @    b !r(!(       â r    ÿàù ä ‰, €    " !r**       â/ r.ÿÿ   ÿä Ä ‰' @    b !r/ /       âO z j  ÿˆ ä‰% €    ¢ !r       â� z k  . ä ‰) @    â !r$       â    ÿàõ Ä ‰+ @    " !r-"-       â0 k  1� ä ‰# €    b !r&       â 2ÿÿ   ÿä â !r,,       â ‰      "  j  ÿ„ â !r'       Ä/ ‰ - @    b  k  2 â !r&%       äO ‰'( @    ¢     ÿàõ â !r)       ä� ‰$/ @    â x        â !r+       Ä‰)* @    " ÿÿ   ÿä â %v  j   Ž â !r.#       â‰"      h ‰%, @    b !r       â  j  ÿ„ Ä ‰+& @    " !r--        â/  k   ä ‰      b !r'('       ÆO ‰1. @    ¢ !r$/$       Æ� ‰!      â !r)*)       Æ‰      " !r3"       Æ‰#      b !r%,%       Æ ‰       b !r+&+       Æ ‰'      " !r       Æ/ ‰ )      b !r1.1       ÆO ‰"-      ¢ !r!!       Æ� ‰$      â !r       Æ†	     â !r##       Æ† 3   â !r        Æ ‰%      " !r''       Æ ‰+      " !r))        Æ/ ‰1      b !r--"       ÆO †	     â !r$       Æ� †     è †	 
!    è † 
#   â !r%%       Æ†	     â !r++       Æ † '    â !r11       Æ/ †	 )    è † -   è †	     è † %    è †	 +    â M™        € ê †y 1   â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y       "   " ¹z  l      â Ey       € â ™x?    â y      !   b ¹z  F   
   â r ÿ       â x      ÿ  â y      %   ¢ r ÿ       â x"      ÿ  â r* ÿ       â x,      ÿ  â r ÿ       Ä r ÿ       ä r/ ÿ       ä r  ÿ   pRð är  ÿ   pRò är    ÿ(Ž È/ xÿ     ä x   ÿ   äO r   ÿ� È xÿ     È |    pbô Ú G)  �    € ê y	      '   " r ÿ       â $z p  Ž â r ÿ       â y      &   b xüÿÿÿÿÀŽ ä z  Y      ä r  €ÿàÿ ä z  X      Ä x   ÿ   äx!   ÿ   È x   !áÿ â %x !   ÿŽ Æ xÿ     â 'z	 y   Ž Æ x     â'z s   Ž â/ zÿ z    ä x   ÿ   â $z	 }  ÿŽ â zÿ t    â $z	 |  Ž ä$z w  ÿŽ ä $z |  Ž â x   ÿàÿ â $z	 v  	Ž âr   ÿàõ â $z v  Ž â z  q      â $z p  	Ž â x	ÿ     â %x"    Ž â r ÿ       Ä x   	  â $x   Ž âr   ÿŽ â$x   Ž âr   ÿŽ â %x "    Ž â z q  ÿàÿ ä x   ÿ   ä rÿÿ   ÿä Ä r   ÿàõ är0    ÿà÷ Ä r	   ÿä är9   ÿàõ ä r$        â �y   • ¢  r%        är1!   ÿäÿ â �y)   • ¢  x. $       â %x2"   $Ž â r;   ÿà÷ â �y7   • â  r4ÿ   ÿä â %v89 Z  . Ž â r?   ÿàõ â �y50   • " r
ÿ   ÿäÿ â $x=4$   ÿŽ Ä %x&"   $Ž â �y(0   • ( �y%2    • â
 r99=   ÿàÿ Æ �y$   • â  rJ8   ÿà÷ ärH8   ÿàý â %v>? Z  . Ž â�y60   • æ %v2; Z  . Ž â�y+0   • b r	ÿ   ÿä Ä $x;
$   ÿŽ â rB2   ÿàû â �y'&    • b $x		$   ÿŽ â rK9   ÿäÿ ä r33;   ÿàÿ â �y18    • b# rI9   ÿä ä r@2   ÿàù â �y
2    • b r??	   ÿàÿ Ä rC3   ÿäÿ â �y&J   ™ b r9   ÿàý ä/r;   ÿàû â �y	>    • b r<ÿ   ÿä ä rA3   ÿä â �y4H   ™ b rF>   ÿàõ Ä rD>   ÿà÷ â �y3B   ™ b %v89 Z  . Ž â r>ÿ   ÿäÿ ä/ �y2@   ™ b $x<<$   ÿŽ â rJ8   ÿàù â %v:; Z  . Ž Æ r=9<   ÿàÿ â $x>>$   ÿŽ â rG?   ÿä ärE?   ÿäÿ ä r< 8       â �y0F   ™ b r?;>   ÿàÿ ä rK=   ÿä â �y<    • b r;   ÿàù Ä rH8   ÿàõ ä �y8D   ™ b r<ÿ   ÿä ä rB:   ÿà÷ â %vD; Z  . Ž â/ r@:   ÿàû Æ $x<<$   ÿŽ â r> :       ä rI=   ÿä â �y:J   ™ b rC?   ÿäÿ ä rFD   ÿàõ â �y=H   ™ b rEE<   ÿàÿ ä rA?   ÿäÿ â �y9>    • b rGE   ÿä Æ �yBB   ™ h �y@@   ™ b r>D   ÿà÷ Æ �y.F   ™ b r?E   ÿäÿ Æ �y;D    • h �y>>   ™ b x   ÿàÿ È |    pbô ä x   ÿà÷ ä x   ÿàÿ ä x   ÿàÿ ä x   ÿàÿ ä x   ÿàÿ Ä x   ÿàÿ ä rÿ   ÿäÿ ä x)T  )    äO x55T  (    äx77T  $    â� 0r%ÿ%    A   â x+6T  +    â0r'ÿ'    A   Ä 0r$ÿ1    A   ä 0r
ÿ
    A   ä &r)&   ÿ   ä&r&5&   ÿ   â  r1$%     A  â &r()3   ÿ   È &r(72   (   ä&r74      ä&r&+4   &   â  r4'$     A  âEr( (       â  r$'
     A  â Er        Æ #r$(      â &r$)0   ÿ   â Er& &       â #r/4   /   â &r050   ÿ   ä 0rÿ	    A   ä &r$78   $   â #r1&      â &r0+8   0   â  r1%
     A  â  r
'     A  â Er$ $       â 0rÿ    A   â Er0 0       â  r%     A  Ä #r
$      â &r&):   ÿ   ä&r:5:   ÿ   ä &r&7=   &   ä&r:+=   :   â #r-0   -   â  r'     A  â &r
)B   ÿ   â  r	%     A  â Er: :       â 0rÿ9    A   â Er& &       â &r
7@   
   Ä &r353   ÿ   ä &r)).   ÿ   ä &rB5B   ÿ   ä&r.5.   ÿ   â #r*	:   *   â Er
 
       â &r3+2   3   â #r#&   #   â  r	'     A  â &rB+@   B   Ä &r)7>   )   ä&r.+>   .   ä 0rÿ;    A   â #r"	
   "   â Er3 3       â  r	%     A  â ErB B       â  r'     A  âEr) )       â  r%%     A  â Er. .       â #r,13   ,   â #r	B      â #r)      Ä #r%.      â G©  õÿÿÿÿƒ ê Ay        € ê Ey  ð     € â G‰  Ð     € ê $x  €  Ž Ê ˆs  / úÿ H   è ˆs  €úÿ H   è ˆs   ûÿ H   è ˆs  ,€ûÿ H   è ˆs   üÿ H   è ˆs  -€üÿ H   è ˆs  # ýÿ H   è ˆs  *€ýÿ H   è ˆs  " þÿ H   è ˆs  €þÿ H   è ˆs   ÿÿ H   è ˆs  €ÿÿ H   ä Ay        € ê {          ì M        € ê „y      H   " z  r  p`ð Æ „y €   H   b r ÿ   p"p Æ „y    H   ¨ „y €  H   è „y    H   ( „y €  H   h „y
    H   h „y €  H   h „y    H   h „y €  H   è „y    H   â !r  /       â !r       Ä/ „y €  H   â !r       ÆO ‰       " !r,       Æ� ‰	      b !r       Æy      &   ¢ !r--       Æ‰      â !r#
#       Æ ‰      " !r*       Æ ‰
      b !r"       â !r        ä ‰      " !r       â !r		       ä/ ‰       b !r       Æ ‰	      ¢ !r       Æ� ‰-      â !r       â !r       ä‰#      " !r

       Æ‰      h ‰      b !r       Æ ‰      " !r        â/ $z  x  Ž âO !r		       â ‰      b !r-       Æ� ‰ €    ¢ !r##       Æ‰      â x   ÿàÿ â !r       ä‰	 €    " z  r  p`ò â !r       ä ‰      b x    p"ò â !r       Ä ‰#      ( ‰
      " !r       Æ/ ‰      b !r       ÆO ‰      ¢ !r       Æ� ‰       â !r	       âz	  r      ä ‰      " !r       ä$x	   ÿŽ â ‰      b !r##       Æ y      '   " !r

       â !r       ä/ ‰ @    b !r       ÆO ‰ €    ¢ !r         Æ� ‰ €    â !r       Æ‰ @    â !r       Æ‰
 €    " $z  ~   Ž Æ ‰ €    ( ‰# €    b     ÿàù â !r       â/ 
$  r  ÿàõ â ‰ €    b ÿÿ   ÿä â !r       âO 
 j  ÿˆ â ‰ €    ¢ r	    ÿ† â$x	   ÿŽ â 
 k   â ‰!       ¢ %ÿÿ   ÿä â !r       â� 
$ j  ÿ„ â ‰ €    â !r
       â!r       â r"ÿÿ   ÿäÿ â ‰ €    " !r       â z j  ÿ† â !r#       â
$ k  % â ‰# @    â z k  "� â !r       â/ 	   ÿàÿ â‰      " r	    ÿ† â !r       âO 	    ÿàù â‰% @    b *    ÿàõ â !r !       â r(ÿÿ   ÿäÿ â ‰! @    ¢ &ÿÿ   ÿä â !r       â� z
 j  ÿ† â ‰ @    â 
	 j  ÿˆ â !r$       â+ÿÿ   ÿä â ‰      " * j  ÿ„ Ä z k  (� â ‰ @    b 
		 k  & ä * k  + â ‰  @    b !r       â !r#       ä ‰" @    " !r%%       Æ/ ‰' @    b !r!       ÆO ‰)$ @    ¢ !r       â�    ÿàÿ ä ‰ €    â !r       â    ÿàõ â !r       â‰!      "         â !r        Ä ‰      b         â !r#"       ä ‰       " ,ÿÿ   ÿä â %  j   Ž â !r''       â/ ‰      b  j  ÿ„ â !r)$)       ÄO ‰      ¢  k  , â !r       ä� ‰"%      â %  j   Ž Æ ‰$#      â !r!!       Æ‰&'      " !r       Æ‰* @    b !r        Æ ‰()      " !r       Æ/ †	     â !r       ÆO †    â !r%%"       Æ� †	 !    â !r##$       â !r''&       ä†     â !r**       Æ†	     â !r))(       Æ †    è †	     è † %    è †	 
#    è † 
'   è ‰*      ( †	 )    â M™        € ê x   ÿàÿ â !r*       Æ? r    ÿàñ È r ÿÿ   ÿä  ä z j  ÿ€ È z k     Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y       "   " ¹z  l      â Ey  @    € â ™x?    â y      !   b ¹z  F   
   â r ÿ       â x      ÿ  â y      %   ¢ r% ÿ       ä r ÿ       Ä r& ÿ       ä r ÿ       ä r/ ÿ       ä r ÿ       ä r0 ÿ       ä r  ÿ   pRò Ä r  ÿ   pRð är    ÿ(Ž È/ xÿ     ä x   ÿ   äO r	   ÿ� È x'ÿ   	  È | '   pbô Ú G)  Ð    € ê y      '   " r ÿ       â $z p  'Ž âr ÿ       â y      &   b x		üÿÿÿÿÀŽ ä x   ÿàÿ ä r		  €ÿàÿ ä z  q      Ä x	   ÿ   â$x	   ÿ Ž â r'   ÿŽ â$x   'Ž â z  Y      ä x$   áÿ â %x   ÿŽ â r'   ÿŽ ä xÿ   $  â 'z y   Ž â z' q  ÿàÿ Ä x$     â'z s   Ž â/ z
ÿ z    ä x$$   ÿ   â $z }  ÿŽ â zÿ t    â $z
 |  'Ž ä$z w  ÿŽ ä $z
 |  Ž â x
ÿ     â $z v  Ž âr   ÿàõ â $z v  Ž â x   
  â $z p  Ž â z  X      â %x"   Ž â r ÿ       Ä x   ÿ   â %x"   Ž â rÿÿ   ÿä Æ r   ÿàõ är        är   ÿä ä r='   ÿàõ â %x("   Ž âr"   ÿà÷ â �y*   • ¢  x  $       â %x,"   Ž â r
ÿ   ÿä â �y)(    • â r#   ÿäÿ â %v<= Z    Ž â r3   ÿà÷ â�y.,    • d	 $x1
$   ÿŽ â r	   ÿàõ â �y8"   • ¢  r!ÿ   ÿäÿ Ä r==1   ÿàÿ â �y("   • ¨  �y-"   • ¨ �y7"   • " $r
ÿÿ    â rD$<   ÿàõ â %v23 Z    Ž â �y   • æ $x!!$   ÿŽ â �y4   • b r"<   ÿàý Æ �y+   • b r#=   ÿä ärE=   ÿä â �y5<    • b r33!   ÿàÿ ä rB2   ÿà÷ â �y,"   ™ b %v	 Z    Ž â/ r@$2   ÿàù Ä �y6D   ™ b $x	
$   ÿŽ â rC3   ÿäÿ ä r>   ÿàû är	   ÿàÿ â �y
B   ™ b r9   ÿàõ ä rA3   ÿä â �y	2    • b r:$   ÿàý Ä r?   ÿäÿ â�y1    • b r!   ÿà÷ ä r=ÿ   ÿä är;   ÿä â �y3@   ™ b! r#ÿ   ÿäÿ â $x==$   ÿŽ ä �y2>   ™ b %v9 Z    Ž Æ �y9:   ™ b %v ! Z    Ž â r@   ÿàõ ä r=   ÿàÿ â $x##$   ÿŽ â r>    ÿàù ä/ r<$   ÿà÷ â$x!!   #Ž â rA   ÿä â �y    • b r"$    ÿàõ Ä r?!   ÿä â �y:@   ™ b r=   ÿäÿ är#!   ÿä â �y;>   ™ h �y<<   ™ h �y""   ™ h �y      • b x''   ÿàÿ È | '   pbô ä x   ÿà÷ ä x   ÿàÿ ä x   ÿàÿ ä x   ÿàÿ ä x   ÿàÿ Ä rÿ   ÿäÿ ä x-8T  -    äO xT  *    ä� x7(T  7    â0r(ÿ)    A   â x+4T  +    â&r*,   ÿ   ä&r,-,   ÿ   Ä 0r5ÿ5    A   ä &r,76   ,   ä&r*+6   *   ä 0r.ÿ.    A   â Er! ,       â/ &r4
   ÿ   â  r,5(     A  â Er* *       â  r.5     A  â 0r	ÿ	    A   â #r,!      Ä #r0*   0   â &r*-
   ÿ   ä &r4+3   4   ä &r,2   ÿ   Æ Er4 4       â  r
.	     A  â 0r1ÿ1    A   ä &r,+9   ,   â #r
4      â &r*73   *   â  r
(	     A  â  r	.1     A  â Er, ,       ä Er* *       Æ #r	,      â &r	:   ÿ   â #r/
*   /   â &r2-2   ÿ   ä &r
;   ÿ   ä &r:-:   ÿ   ä &r	+<   	   ä 0rÿ    A   Ä &r--;   ÿ   ä &r++"   
   â Er
 	       â &r279   2   â r	.     A  â&r:7<   :   â r(     A  â&r-7"   -   ä 0rÿ     A   â Er2 2       â  r1(1     A  â Er: :       Ä Er+ +       â  r..     A  â Er- -       â  r((     A  â #r12      â #r&	
   &   â #r:      â #r%.+   %   â #r(-      â G©  Àöÿÿÿÿƒ ê Ay        € ê Ey  Ð     € â G™  °     € ê $x  @  Ž Ê ˆs  0 ûÿ H   è ˆs  €ûÿ H   è ˆs   üÿ H   è ˆs  /€üÿ H   è ˆs   ýÿ H   è ˆs  €ýÿ H   è ˆs  & þÿ H   è ˆs  €þÿ H   è ˆs  % ÿÿ H   è ˆs  €ÿÿ H   ä Ay        € ê {          ì M	        € ê „y     H   b z  r  p`ð Ä x#        â „y €   H   ¢ r ÿ   p"p Æ „y    H   è „y
 €  H   ( „y	    H   h „y €  H   ( „y    H   ( „y €  H   ( „y    H   è „y €  H   â !r0       Æ/ y      &   b !r       ÆO ‰       " !r       Æ� y      '   ¢ !r

/       Æ‰      â !r	       Æ‰
      " !r       Æ ‰	      b !r&       Æ ‰      b !r       â !r         ä ‰      " !r%       â z%  r      ä ‰       ¢ !r       â !r       Ä� ‰      â $x!%   ÿŽ â !r
       ä‰      " !r		       Æ‰      b !r       Æ/ ‰      b !r       Æ ‰      " !r        ÆO ‰	      ¢ !r       Æ� ‰  €    â !r       Æ‰      " !r       Æ‰      b !r       Æ/ ‰
      h ‰      b !r       Æ ‰      " !r		       ÆO ‰      ¢ !r        â� $z  x  Ž ä ‰      â !r       â$z  ~   Ž â !r       â‰ €    " !r
       Æ/ ‰	 €    b x
   ÿàÿ â !r       ä ‰ @    b z 
 r  p`ò â !r       ä ‰ €    " x    p"ò â !r       ÄO ‰ €    ¢ r%    ÿ† â !r       ä� ‰ €    â rÿÿ   ÿäÿ Æ ‰ €    â !r       âr%    ÿŠ ä ‰ €    " !r	       â/ 
	  r  ÿàù ä‰ €    b !r
       â !   ÿàõ Ä ‰ €    b !r       â  ÿÿ   ÿä â !r       âO ‰
      " 
 j  ÿ„ â !r       ä� ‰ @    ¢ z j  ÿ† â !r       Ä ‰ @    â 
 k    â !r       ä‰ @    " z k  � â !r       ä/ ‰ @    b r"ÿÿ   ÿäÿ â !r       Ä‰ @    b z j  ÿŠ ä $ÿÿ   ÿä â ‰ @    b !r'
       â %   ÿàÿ ä ‰ @    " !r       âO z k  "� Ä ‰ @    ¢ !r       â� 
	 j  ÿˆ ä ‰ @    â !r       â!   ÿàÿ â !r       â/ !    ÿàõ ä‰      b     ÿà÷ â !r       â%v  j  # Ž â ‰      " "ÿÿ   ÿä â !r       â ! j  ÿ„ â ‰      b  ÿÿ   ÿäÿ â !r       â 
		 k  $ â ‰      " 
 j  ÿ† â !r       ÄO ‰      ¢  k   � â !r       ä� ‰      è ‰       â !r       Æ/ ‰      b !r       Æ‰      " !r#       â! k  " â !r       â †	 '    â !r       ÆO † #   â !r       Æ� †	     â !r        Æ †     â !r       Æ/ †	     â !r       Æ†    è †	     è † 
    è †	     â M™        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      "   " ¹z  l      â Ey  €    € â ™x?    â y      !   b ¹z  F   
   â r ÿ       â x      ÿ  â y      &   ¢ x      ÿ  â x      ÿ  â r ÿ       â y      '   â r ÿ   pRð Ä r ÿ   pRò är    ÿ(Ž È/ xÿ      È r    ÿ� È xÿ     È |    pbô Ú G)  P
    € ê y      %   "Î r ÿ       â $z }  ÿŽ â r ÿ       â $z w  ÿŽ â xüÿÿÿÿÀŽ â 'z y   Ž â z
  X      ä r   €ÿàÿ â 'z s   Ž â zÿ z    Ä x    ÿ   ä zÿ t    â $z	 |  Ž â x   ÿ   ä x!ÿ      â $z	 v  	Ž â x    áÿ â %x   ÿŽ â r   ÿàõ Ä x   ÿ   ä x    ÿ   ä r ÿ       â $z p  Ž â z q  ÿàÿ Æ $z |  Ž â x   ÿàÿ ä rÿÿ   ÿä â $z v  Ž â z  q      â $z p  	Ž â xÿ     â %x6"   Ž â r   ÿŽ äx    !  â $x   Ž â z	  Y      â %x"   Ž â x     Ä x    ÿ   ä r/   ÿàõ ä r  
       är! 	       ä r06
   ÿà÷ ä x4 $       â %x$"    Ž â r(ÿ   ÿä Ä r17	   ÿäÿ â %v./ Z  4 Ž â r,
   ÿàù ä r3   ÿà÷ â $x)($   ÿŽ â r-	   ÿä â %x&"    Ž â r+ÿ   ÿäÿ â �y $    • ¢  r//)   ÿàÿ Ä r;   ÿàý â �y!0   • è �y"0   • è �y#0   • ( �y$0   • " %v23 Z  4 Ž â r8 .   ÿàù Ä �y&    • b $x++$   ÿŽ Æ �y%,   • ¢  r0.   ÿà÷ Æ/ �y(,   • ¢ r1/   ÿäÿ Æ �y&,   • ¨ �y',   • ¢ r33+   ÿàÿ â %v:; Z  4 Ž â r9/   ÿä â �y+0   ™ ¨  �y).    • ¢  r-ÿ   ÿä Ä/ r< 2   ÿàõ â�y,8   ™ ¤ $x--$   ÿŽ â r5   ÿàù â �y*2    • ¢ r.2   ÿàû är0:   ÿà÷ ä r;;-   ÿàÿ Ä r=3   ÿä är8 :   ÿàõ ä/ r2ÿ   ÿä ä r/3   ÿäÿ är1;   ÿäÿ â %v45 Z  4 Ž â r9;   ÿä â �y-.   ™ ¤  $x22$   ÿŽ Ê r552   ÿàÿ â �y/0   ™ ¨ �y.<   ™ ¨ �y34    • ¨ �y18   ™ ¢ r< 4   ÿà÷ Æ/ �y0:    • ¢ r84   ÿàõ È r95   ÿä är=5   ÿäÿ Æ �y28   ™ ¨ �y4<   ™ ¢ x    ÿàÿ È |    pbô ä x

@  ÿà÷ ä x    ÿàÿ ä x    ÿàÿ ä x    ÿàÿ ä r	ÿ	   ÿäÿ Ä x"!T  "    ä� x##T  $    â0rÿ    A   âx&%T  &    âO &r!"+   ÿ   ä 0r$ÿ)    A   ä &r%#,   !   ä 0r!ÿ     A   â  r $     A  Ä Er% %       â 0r*ÿ*    A   È #r %      â x''T  (    â &r+&+   ÿ   â  r)$!     A  â  r$*     A  â&r "-   ÿ   ä&r-&-   ÿ   â  r%!*     A  â &r+',   +   ä &r #.       Ä &r.'.   -   ä &r("/   ÿ   âEr          â &r*&/   ÿ   â Er+ +       â &r(#1   (   âEr. .       â #r$       â &r*'1   *   ä 0r0ÿ0    A   Ä 0r ÿ3    A   â #r)+      â #r%.      â Er) (       â &r""2   ÿ   ä&r&&2   ÿ   ä &r"#4   "   ä&r&'4   &   â Er* *       â  r%!0     A  â Er" "       â  r(0     A  âEr& &       â  r$      A  â r!!      A  â #r()      â #r%*      â #r$"      â #r!&      â G©  0øÿÿÿÿƒ ê Ay        € êÏ Ey  °     € â G‰  �     € ê r   ÿ@Ž Ê ˆs  üÿ H   è ˆs €üÿ H   è ˆs  ýÿ H   è ˆs €ýÿ H   è ˆs  þÿ H   è ˆs €þÿ H   è ˆs  ÿÿ H   è ˆs €ÿÿ H   ä Ay        € ê {          ì M        € ê „y €   H   h „y     H   ¨ „y
 €  H   è „y    H   ( „y    H   ( „y €  H   h „y    H   h „y    H   h „y €  H   h „y €  H   h „y	    H   b !r       Æ/ „y    H   b !r       ÆO „y €  H   ¢ !r
       Æ� „y    H   â !r       Æ „y" €  H   " !r       Æ„y  	  H   " !r       Æ„y    H   b !r       Æ „y €	  H   b !r       Æ „y   
  H   b !r       Æ „y €
  H   b !r       Æ „y     H   b !r		       Æ „y €  H   â !r       Æ/ „y €  H   â !r       ÆO „y €  H   â !r       Æ� ‰      b !r""       â !r		       ä‰      " !r       Æ‰	      ¢ !r       â !r        È ‰
      â !r""       Æ ‰      " !r         Æ ‰"      b !r       â/ !r       ä ‰       b !r       â !r       ä ‰      " !r	       âO !r       Ä ‰      ¢ !r

       Æ� ‰      â !r       Æ‰	
      " !r"       Æ‰      b !r        Æ/ ‰      b !r       Æ ‰      " !r       ÆO ‰      ¨ ‰  €    ¢ !r       Æ� ‰ €    â !r	
	       â!r       ä‰ €    " !r       Æ/ ‰	 €    b !r       Æ y      %   " !r       ÆO ‰ €    ¢ !r         Æ ‰ €    b !r       Æ� ‰ €    è ‰      â !r       Æ‰  @    " !r	       Æ/ ‰ @    b x   ÿ   â !r       ÆO z  r  p`ð â ‰ @    " $z x  Ž â !r
       âr ÿ   p"p â ‰ @    ¢ $z ~  Ž â !r       â� x   ÿàÿ â ‰ @    â z  r      â !r       â z  r  p`ò â ‰
 @    b !r         Äx    p"ò â ‰ @    " 
 r  ÿàõ â !r	       â/ r   ÿ† â ‰ €    b ÿÿ   ÿä ä 
 j  ÿ„ â ‰       b !r       â rÿÿ   ÿäÿ Ä z j  ÿ† â !r       âO 
 k   â !r       â� ‰	      " z k  � â !r
       ä‰      ¢ r   ÿŽ â !r       Ä‰      â    ÿàÿ â !r       ä/ ‰      b    ÿà÷ ä   ÿàõ â ‰      " !r        â  ÿÿ   ÿäÿ Ä ‰ @    b 
 j  ÿ† äx        â ‰      b ÿÿ   ÿä ä  j  ÿ„ â %v
 j   Ž â 
 k   � â !r	       â 	 k   â !r       âO !r       â� †	 
    â !r       Æ/ † 
   â !r       Æ†	     â !r       Æ†     â !r       Æ †	     è †    è ‰      h  †	     â M™        € ê x   ÿàÿ â !r       Æ/ r   ÿàñ È r ÿÿ   ÿä  ä z j  ÿ€ È z k     Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      "   " ¹z  l      â Ey   
    € â ™x?    â y       !   b ¹z  F   
   â r ÿ       ä y      %   ¢ r ÿ       ä r ÿ       Ä r ÿ       ä r ÿ       ä r ÿ       ä r ÿ   pRð är ÿ   pRò är    ÿ(Ž È/ xÿ     ä x   ÿ   äO r	   ÿ� â $À    Ž Æ xÿ   	  È |    pbô Ú G)  Ð    € ê y      '   " r ÿ       â $z p  Ž â r
 ÿ       â y      &   b z  X      â 'z y   Ž È $z }  ÿŽ â zÿ z    â 'z
 s  
 Ž â/ x	üÿÿÿÿÀŽ Æ $z	 w  ÿŽ â r  €ÿàÿ â $z |  Ž âz
ÿ t  
  â $z |  Ž â r		   ÿàõ â $x   ÿ Ž âx   ÿ   â $z
 v  Ž â x   ÿàÿ â $z

 v  Ž â x   áÿ â %x   ÿŽ â xÿ     Ä xÿ     â $z p  Ž â z  q      â %x
"   Ž â x     ä x     â %x"   Ž â z  Y      ä r ÿ       Ä r   ÿŽ ä z q  ÿàÿ ä x   ÿ   ä rÿÿ   ÿä ä x   ÿ   ä r5	   ÿàõ â $t(ÿ$   ÿ Ž â r        Ä rÿ   ÿä ä r        â %v45 Z  ( Ž â r+	   ÿàû Æ $x$   ÿŽ ä %x 
"   Ž â rÿ   ÿäÿ Æ %x"   Ž â r55   ÿàÿ â �y#     • ¢  r24   ÿàû â %v*+ Z  ( Ž ä �y    • â r.   ÿà÷ ä r35   ÿäÿ ä r)	   ÿàû Ä r,   ÿàù ä r/   ÿäÿ â$x$   ÿŽ â/ r-   ÿä â %v() Z  ( Ž â r0*   ÿà÷ â �y%.   • " r+   ÿàÿ ä r+ÿ   ÿäÿ â �y .   • ( �y'.   • b $x++$   ÿŽ Æ �y".   • h �y&,   • b r))+   ÿàÿ Æ �y!,   • b r.*   ÿàõ Æ/ �y$,   • b r/   ÿä Æ �y,   • b r *       Æ �y4    • b $x1   Ž Æ �y*2   ™ b r,4   ÿàù Æ �y    • b r4(   ÿàõ ä/ r-5   ÿä â �y00   ™ b r5)   ÿä ä r2(   ÿà÷ â �y,,   ™ f r3)   ÿäÿ â �y4   ™ h �y..   ™ h �y((    • h �y22   ™ b x    ÿàÿ È |    pbô ä x@  ÿà÷ ä x    ÿàÿ ä x    ÿàÿ ä rÿ   ÿäÿ â 0r#ÿ#    A   äO 0rÿ    A   â� x%%T       Äx''T  "    äx!&T  !    â 0r ÿ    A   ä &r%*   ÿ   âx$T      â &r+!*   ÿ   ä 0rÿ    A   ä &r$%0   ÿ   Ä &r0!0   ÿ   ä &r"',      ä &r%%   ÿ   ä&r!   ÿ   â Er" "       â &r$'.   $   â  r#      A  â &r+,   +   Ä &r..   0   ä 0r(ÿ(    A   ä &r%'2   %   ä&r2      â #r"      â  r)      A  â  r#     A  âEr  +       â  r##(     A  â Er$ $       â  r     A  âEr. .       â  r((     A  â Er% %       Ä Er        â #r)       â #r$      â #r.      â #r#%      â #r(      â G©   ùÿÿÿÿƒ ê Ay        € ê ˆ  ýÿ H   è ˆ €ýÿ H   è ˆ  þÿ H   è ˆ €þÿ H   è ˆ  ÿÿ H   è ˆ €ÿÿ H   è {          ì M        € ê „y      H   " z  r  p`ð Æ „y
     H   b r  ÿ   p"p Æ „y  €   H   ¨ „y	     H   è „y  €  H   ( „y     H   h „y     H   h „y  €  H   h „y     H   h „y  €  H   h „y      H   b !r       Æ „y$  €  H   " !r
       Æ/ „y  €  H   b !r       ÆO „y     H   ¢ !r		       Æ� „y  €  H   â !r       Æ„y"     H   â !r       Æ„y  €  H   â !r       Æ „y  €  H   â !r       â !r		       ä ‰      " !r       Æ y      &   b !r        Æ y      '   b !r$$       â !r       â/ !r		       ÈO ‰      " !r       Æ� ‰
	      b !r       â!r"       ä ‰      ¢ !r$       Æ ‰      â !r       Æ ‰      ( ‰      b !r       â !r
	
       È/ ‰      " !r       ÆO ‰
      b !r       Æ� ‰	      ¢ !r       Æ‰ €    â !r       Æ‰      ( ‰      b !r       â !r
       È/ ‰ €    " !r		       ÆO ‰
 €    h ‰	 €    ¢ !r       â� !r       È‰ @    â !r       Æ‰ €    ( ‰ €    b !r       â !r

       È/ ‰ @    " !r	       ÆO ‰	
 @    b !r       â� x   ÿàÿ â $z x  Ž â z  r      â !r       âz  r  p`ò â ‰ @    ¢ $z ~  Ž â !r       âx     p"ò â ‰ @    ä 
 r  ÿàõ Ä ‰ @    " r   ÿ† â !r       â ÿÿ   ÿä â ‰      " 
 j  ÿ„ â !r

	       ä/ ‰       b 	   ÿàÿ ä 
 k   Ä 	   ÿàõ ä x	        â !r       âO ÿÿ   ÿä â ‰
      ¢  j  ÿ„ â !r       â� %v j  	 Ž â ‰      â !r       Æ‰      " !r       â rÿÿ   ÿäÿ ä ‰      " z j  ÿ† â!r        â/  k   â †	     â z k  � Æ †    â !r
       âO !r       È� †	     â !r       Æ†     â !r       Æ †	     â M™        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      "   " ¹z  l      â Ey  @    € â ™x?    â y       !   b ¹z  F   
   â x      ÿ  â r ÿ       â y      %   ¨ y      &   â r ÿ   pRò Ä r ÿ   pRð är    ÿ(Ž È/ xÿ     È r
   ÿ� â $€    Ž â r ÿ       ä xÿ   
  ä x   ÿ   äO |    pbô Ú G)      € ê y	      '   "Ž r ÿ       ä r ÿ       ä x
üÿÿÿÿÀŽ â $z
 w  ÿŽ Æ r  €ÿàÿ ä z q  ÿàÿ ä x   ÿ   â$x   ÿ Ž â z  Y      â 'z	 y   Ž È 'z s   Ž â zÿ z    Æ $z p  Ž â zÿ t    â $z		 }  ÿŽ ä $z |  Ž â x   ÿàÿ â $z |  Ž â r	
	   ÿàõ â $z
 v  Ž â x   áÿ â $z v  Ž â xÿ     â %x   ÿŽ â xÿ     Ä x     â $z p  Ž â x     â %x
"   Ž âz  X      ä r ÿ       â $x   ÿ Ž â rÿÿ   ÿä â %x"   Ž â x   ÿ   Æ r	   ÿàõ â $t#ÿ$   ÿ Ž â r        Ä rÿ   ÿä ä r        â %v Z  # Ž â r(   ÿà÷ ä r	   ÿàõ â $x!$   ÿŽ â r)   ÿäÿ â %x$
"   Ž â r&ÿ   ÿä Ä r,   ÿà÷ â%v Z  # Ž â r!   ÿàÿ â �y $    • ¨  $x-   Ž â �y!(   • è �y"(   • â $x%&$   ÿŽ â r&   ÿàõ Ä �y#(   • $ r%   ÿàÿ ä �y$(   • " r'   ÿä ä r*   ÿàù â �y%,   ™ b %x"   Ž Æ �y    • b r(   ÿà÷ ä/ r+   ÿä â�y&&   ™ b r,   ÿàõ ä r)   ÿäÿ â �y    • d $x-    Ä �y((   ™ h �y'*   ™ h# �y    • h �y),   • h �y*,   • h! �y+,   • h �y,,   • b x    ÿàÿ È |    pbô ä x@  ÿà÷ ä x    ÿàÿ ä rÿ   ÿäÿ ä x"!T  "    â� 0r ÿ     A   âO x##T  $    Æ&r!"%   ÿ   ä0rÿ    A   ä &r!#&   !   ä 0rÿ    A   ä &r""'   ÿ   â x*)T  *    Æ &r"#(   "   ä &r%*%   ÿ   ä&r'*'   ÿ   â x++T  ,    â 0rÿ    A   È &r%+&   %   ä&r'+(   '   â  r      A  â r       A  â Er! !       â  r     A  â Er" "       â  r     A  â Er% %       ä Er' '       â #r!      â #r "      â #r%      Ä #r'      â G©  0ûÿÿÿÿƒ ê Ay        € ê� ˆ  þÿ H   è ˆ €þÿ H   è ˆ  ÿÿ H   è ˆ €ÿÿ H   è {          ì M	        € ê „y      H   " z  r  p`ð Æ „y  €   H   b r  ÿ   p"p Æ „y     H   ¨ „y	     H   è „y  €  H   ( „y     H   h „y     H   h „y  €  H   h „y     H   h „y  €  H   h „y  €  H   b !r       Æ „y  €  H   " !r       â/ !r       âO !r	       ä� !r       â!r       â!r       ä !r       Æ ‰      b !r       Æ ‰      ¢ !r       Æ ‰
      â !r       È !r       Ê ‰      " !r       â/ !r       ÈO ‰      b !r

       Æ� ‰      ¨ ‰	
      â !r       â x        È ‰      " !r       â/ !r       ÈO ‰ €    b !r
	       Æ� ‰	 €    ¨ ‰ €    â !r       Ê ‰ €    " !r       Æ/ y      '   b !r		       âO !r
       È� ‰	 @    ¢ x   ÿàÿ â $z x  Ž ä ‰ @    â z  r  p`ò Æ ‰
 @    " x     p"ò â !r       â $z ~  Ž Ê/ 
  r  ÿàõ â !r	       ÆO ÿÿ   ÿä â !r       â� 
  j  ÿ„ â ‰      " !r
       Æ‰      h ‰
 @    ¨ ‰	      â !r       â 
  k   â !r       â/ %v j   Ž â !r

       ÈO †	     â !r	       Æ� †    è ‰	
      h  †	     â M™        € ê ¹z  r      â !r	
	       â/ �x   ?àÿ Ì |   ÿàñ È r ÿÿ   ÿä  ä z j  ÿ€ È z k     Ê †y 	    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y       &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú         â $rÿÿ     Ž È % \   Ž Ì �	    ™ b $ÿÿ   ÿ Ž â ¹z  l      â $ÿÿ    Ž â y      "   b ™x?    â Ey  à    € â '
 m   Ž â y      !   b r ÿ       Æ y      '   ¢ 
ÿ n    Ê $ÿÿ   
Ž ä y      %   ä $
 o  Ž ä $‚ÿÿ   ÿ Ž ä $z w  ÿŽ ä'Š s   Ž â r   ÿ(Ž È/ x	ÿ     â $z }  ÿŽ ÆO r
	   ÿ� ä r   ÿàõ ä x	ÿ   
  È | 	   pbò ä Šÿ t    â $rÿÿ   ÿ Ô G  €    € ê rÿ	   ÿ3Ž âŸ $rÿÿ   ÿ Ž â x
üÿÿÿÿÀŽ â Ey ð    € â |   ÿàÿ â 'z y   Ž â r  €ÿàÿ ä x   ÿØ� â$rÿÿ   ÿ Ž â x `   p`ð â $x   ÿ Ž â x   ÿÀ‚ Ä zÿ z    Ö G™  @    € ê $z |  	Ž â x "       â $x
   ÿ Ž ä $z v  Ž ä$rÿÿ   ÿ Ž â x   
áÿ â $z p  Ž ä %x

   ÿŽ â xÿ     Æ %v X  Ž â x
   ÿüŽ ä x!     â $rÿÿ    Ž â r        ä r	   ÿàó â $tÿ$   ÿ Ž â r"        â $r#ÿÿ    Ž â rÿ   ÿäÿ  â %v Z   Ž Æ �y""    • ¢ $x$   ÿŽ â r   ÿàõ ä x   ÿàó â $x   Ž ä $x    ä $rÿÿ   Ž  â r   ÿ‚ Ä �y    • â %x   Ž Æ �y%   • â r!   ÿäÿ  Æ �y    • ( �y'   • ( �y    ™ h �y   ™ h �y    • b xÿÿÿÿÿàÿ È r ÿ   pRò ä x@  ÿàõ ä x		    ÿàÿ Æ $rÿÿ    ä 0r"ÿ"    A   âO xT  %    ä� x  T  '    Æ&r   ÿ   È&r       ä 0r#ÿ    A   Æ Er        ä  r""#     A  È #r"      â G   ýÿÿÿÿƒ ê Ay       € ê $vÿ X  ÿ Ž â z  Y      â G‰       € ê $z |  	Ž â x
ÿ     â $x   ÿ Ž ä $z v  Ž âx
   
  äx   áÿ â $z p  Ž â x   ÿ   Ä xÿ     â %x"   ÿŽ Æ x     â$x   ÿ Ž ä %x   Ž È r	   ÿàñ äx	    ÿàÿ ä x6 $       â $rÿÿ     â r   ÿàó Æ $x$   ÿŽ â x	@   ÿàÿ â $rÿÿ   Ž  ä %v Z  6 Ž â r   ÿàõ Æ %v Z  6 Ž â r   ÿàÿ ä r    ÿà÷ â$x#$   ÿŽ â r(   ÿàñ â$rÿÿ    â r&   ÿàõ â %v Z  6 Ž â �y    • ¢  r!   ÿäÿ Ä $x   #Ž â r)
   ÿä  â $x$   ÿŽ â r"   ÿà÷ â $x'   
 â r<   ÿàõ â �y((   ™ â r   ÿàÿ ä x	`   ÿàÿ â $x=    â r2   ÿàñ â �y%    ™ b	 r7   ÿàõ Ä r#   ÿäÿ â �y+<   • ¢ rÿ   ÿä â %v67 Z  6 Ž â r.   ÿàó â �y<   • ¢ r8   ÿà÷ â $x-$   ÿŽ ä �y&&   ™ â  $x3     â r9   ÿäÿ Ä �y)<   • b/ r06   ÿàñ â $x/   
Ž  âr77-   ÿàÿ â �y <   • h�y'<B  • ( �y$<D  • " $x17   
  Æ �y    • ( �y""   ™ ( �y<‚  • ( �y<H  • ( �y#<F  • (/ �y<„  • " r:6   ÿàó Æ �y    • ( �y!8    • ( �y*<†  • ( �y.   ™ ( �y-<ˆ  • " r;7   ÿäÿ  Æ �y8@  • ( �y.<Â  • ( �y/<Ä  • ( �y2   ™ ( �y00   ™ ( �y,8€  • ( �y2<È  • ( �y1<Æ  • (/ �y3:   ™ ( �y48À  • ( �y66    • " x		€   ÿàÿ È | 	   pbð ä x   ÿàó Ê $rÿÿ   Ž  â x++T      ÊO &r(+(   ÿ   â� x )T       â0rÿ    A   â x''T  $    Æ&r  %   (   ä &r&'&   ÿ   Æ Er          â 0rÿ    A   â x##T      ä xT      Æ &r"#"   &   ä 0r!ÿ!    A   Æ Er" "       â 0rÿ    A   â  r!     A  â &r   ÿ   â x-*T  -    â 0rÿ    A   â #r       â x/.T  /    Æ  r     A  â &r-      ä &r//0   ÿ   ä 0r,ÿ,    A   â Er        ä x21T  2    â #r"      â  r,     A  Æ &r/23   /   ä 0r4ÿ4    A   ä 0rÿ6    A   â Er/ /       â #r      Æ  r4     A  È #r/      â G‰  °øÿÿÿÿƒ ê Ay        € êŸ r ÿ   pRð är ÿ   pRò Ö ˆ €ÿÿ H   è {          ì M        € ê „y     H   " r ÿ   pRð â $z  x  Ž Ä „y €   H   h „y	    H   ¢ !r       È !r       È/ !r	       ÊO ‰      $ !r       Ê ‰      $ !r       Ê ‰	 €    $ !r		       Ê ‰	 @    $ !r	       Ê ‰      b  M	        € ê x        â $z ~  Ž â !r       Æ/ %v j   Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y	      &   " z ÿ \  pPò â ¹z  F   
   Æ z ÿ ]  Sò Ú $ÿ   ÿ Ž È %	 \   Ž Ê �    ™ b r ÿ       â Ey  0    € ä y      "   d ' 	 m   Ž äy      !   d 'š	 s   Ž â  ÿ n     â y      '   ¢ r ÿ       Æ $ÿÿ    
Ž È $  o  	Ž ä$’ ÿÿ   	 Ž â rÿ   ÿü€ â/ $x    Ž Ê xÿ     ä šÿ t    ä r ÿ       Æ G	       € ê z ÿ `  pPô äO z ÿ ^  pPð ä z ÿ b  p ö ä z ÿ a   Sô ä z ÿ _   Sð ä z ÿ c  0't Ö 	      '   " 	     € ä¢	     € â ©
      '   d $
 x  ÿŽ ä 	      %   ¨ ©      %   â $
 ~  ÿŽ Ä $ª	
 ~  ÿŽ ä/ $ª
 x  ÿŽ â    àë äOÿ     ä ¢		   àÇ ä�   ÿàó ä ¨ÿ     Ä ÿÿ   Äþ ä ¢

	   ÿàû ä ¢ÿÿ   „þ â $ÿÿ   Ž  â 
 ^  ÿ‚ ä ¢ÿ   ÿäÿ ä ª
 b  ÿ† Ä 
	 _  �  â $rÿÿ   ÿ Ž â r ÿ       ä ª
 c  � Æ �	     h �©     d Ay        € êO r   ÿ� â ¹z  l      â $z	 }  ÿŽ â ™x?    â x
ÿ     â $z   w  ÿŽ â Ey  °    € â $rÿÿ   ÿ Ž â r' ÿ       Ä | 
   pbð ä r	    ÿàó Ê $r	ÿÿ   ÿŽ  Ì G	  @    € ê r ÿ
   ÿ3Ž â Ey €    € â r ÿ       â $z v  ÿŽ â|     ÿàÿ â $rÿÿ   ÿ Ž â xüÿÿÿÿÀŽ â 'z y   Ž â xÿ     ÿÀ€ ä r' ÿ       â $x   
Ž â zÿ z    Ò G	  À    € ê y      %   " r
   ÿàñ ä x $       ä x "       ä x   ÿ   â %v Z   Ž â x   ÿ   È x!   áÿ â $z p  Ž È $z |  Ž ä $rÿÿ   	  â x   ÿàñ â $x
   Ž ä $x$   ÿŽ ä %v X  Ž È $x   Ž ä %x   Ž ä �y    • ¤ $rÿÿ     ä �y   • ä %x   Ž Ä �y   • ä %x!   Ž ä �y   • ( �y#   • ( �y    ™ h �y    • h �y    ™ b z ÿ `  pPð È z ÿ a   Sð ä xT      â� 0r'ÿ    A   âO x#T  #    Æ&r   ÿ   ä0rÿ    A   ä &r#      Æ  r''     A  â r ÿ       ä Er        â G‰       € ê xÿ     â %v `   Ž È $x"   ÿŽ È $x   Ž È %x   Ž ä �y     ¨ �y    è �y    è �y    ( �y    " xT      Ê� &r   ÿ   â xT      Ê&r      ä 0rÿ    A   ÆO Er        ä  r     A  È #r   ÿ   â x

    ÿàÿ â #r''   ÿ   Ê Ay       € ê xÿ àÿÿÿÿÀ€ Ú G‰  €    € ê y      %   " $z  |  
Ž â z ÿ `  pPð â $x   ÿ Ž âx   ÿ   ä r     ÿàÿ â $vÿ X  ÿ Ž â x"   áÿ â$x    ÿ Ž â xÿ     Ä x!ÿ   "  ä z  Y      ä x!"   !  ä z ÿ a   Sð ä z  a      ä x     â $z  p   Ž â x""   ÿ   â $vÿ `  ÿ Ž Ä %x "   ÿŽ Ì %x   Ž È r
   ÿàó â $tÿ$   ÿ Ž â r   ÿàõ ä rÿ	   ÿäÿ  â %v Z   Ž â x 
    ÿàÿ Æ $x    â r    ÿàõ â $x$   ÿŽ â r    ÿàó ä �y$   • ¢  $rÿÿ   	 â r   ÿàÿ â %v Z   Ž â �y%   • ¤  r   ÿäÿ  â $x$   ÿŽ â r,"   ÿàõ â �y#   • ä  $x   Ž Ä �y*   • è  �y+B  • ( �y(D  • ( �y)F  • h �y&H  • b $x-!    Æ �y   ™ b r"   ÿàõ Æ �y ,   ™ b r   ÿà÷ Æ �y    • b r    ÿàó ä/ r   ÿäÿ â $x   ! âr   ÿäÿ  â �y    • h �y,    • h �y   ™ h �y   ™ h �y@  • b x%$T  %    ÄO x##T  *    ä� x++T  (    â&r%%   ÿ   âx))T  &    Æ &r%#    %   ä 0r#ÿ    A   Æ Er% %       ä r   ÿàó â 0r,ÿ,    A   ä &r++   ÿ   Æ  r$,#     A  â 0rÿ    A   ä &r))   +   ä 0rÿ    A   â #r$$%   '   â r   ÿäÿ  ä Er' )       â  r     A  â G‰  ð     € ê r   ÿàó â �y    ¨ �y&    ¢ $x   Ž  Æ �y    è �y(    è �y     " xT  &    ÊO &r   ÿ   â xT  (    Ê� &r       ä 0r ÿ    A   ÆEr        ä  r##      A  È #r#      â #r''   $   â G‰  ð     € ê �y B   ¢ r   ÿàó Æ �yD   ¢ r   ÿäÿ  Æ �yF   è �yH   è �y@   " x T      ÊO &r   ÿ   â xT      Ê� &r      ä 0rÿ    A   ÆEr        ä  r     A  È #r      â x

@   ÿàÿ ä x€  ÿàõ ä | 
   pbò ä x€  ÿà÷ â $rÿÿ    Æ rÿ   ÿäÿ Î G™  Ðùÿÿÿÿƒ ê Ay        € ê r ÿ   pRð âEy  €     € â r ÿ   pRò Ö G‰  P     € ê z ÿ `  pPð â $x    Ž Æ z ÿ a   Sð ä ˆs '€ÿÿ H   ö ˆ    H   ä Ay        € ê {          ì M        € ê z ÿ `  pPð â „y      H   f z ÿ a   Sð Ú „‰ €   H   ¨ „‰    H   " !r' '       È/ !‚'       ÈO !‚       â G‰        € ê „y €  H   ( „y €   H   h „y     H   ¨ „y	    H   è „y €  H   " !r       â !r'       Æ/ !r         âO !r	       Æ� !r        ä‰      " z ÿ `  pPð ä r ÿ   pRò ä z ÿ a   Sð â !r       Ê ‰       $ !r         Ê ‰  €    $ !r        Ê ‰
 @    $ !r

       Ê ‰	
      $ !r
	       âG‰        € ê ‰      $ !r       Ê ‰       $ !r         Ê ‰  €    $ !r        Ê ‰ @    $ !r       Ê ‰	      $ !r	       ä M        € ê !r       â G‰  `    € ê z   h      â !r       Æ x     pBð Ú G	  p    € ê x  ÿÿÿÿÿàÿ È x       € Ê $x     ÿ Ž È ‚{   €     $ xÿ     È Iy Ðêÿÿÿÿƒ ê  x ;ª¸¿  A  Ì s          $ !t   €?    Ì s        $  r     A  È  r     A  â Gy  P    € ê  x '7=  A  â x*BL?  A  Æ #t   €?    È  r      A  â  x    ?  A  Ê s     $   $ !t  €?    È  r       A  È  r      A  â Gy  °    € ê x     pRð Ú G‰  à     € ê x     pRð Ú G	        € ê 	z  i    � ä 	z i    � Æ  x ;ª¸¿  A  â 	z i €  � Ê s        $ !t  €?    Ì s        $  r       A  È  r      A  â Gy  Ð     € ê  r     A  â Gy  °     € ê 	x   à@  � ä 	x  à@  � Æ  x #ÛÙ?  A  â 	x  àÀ  � Æ  x;ª¸¿  A  ä !t  €?    È s        $ !t  €?    Ì s        $  r       A  È  r      A  ä y      &   " x        Æ y       %   $ $z x   Ž È $z ~  Ž È %v j   Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y       &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú         â $rÿÿ     Ž È % \   Ž Ì �	    ™ b $ÿÿ   ÿ Ž â ¹z  l      â $ÿÿ    Ž â y      "   b ™x?    â Ey  à    € â '
 m   Ž â y      !   b r ÿ       Æ y      '   ¢ 
ÿ n    Ê $ÿÿ   
Ž ä y      %   ä $
 o  Ž ä $‚ÿÿ   ÿ Ž ä $z w  ÿŽ ä'Š s   Ž â r   ÿ(Ž È/ x	ÿ     â $z }  ÿŽ ÆO r
	   ÿ� ä r   ÿàõ ä x	ÿ   
  È | 	   pbò ä Šÿ t    â $rÿÿ   ÿ Ô G  €    € ê rÿ	   ÿ3Ž âŸ $rÿÿ   ÿ Ž â x
üÿÿÿÿÀŽ â Ey ð    € â |   ÿàÿ â 'z y   Ž â r  €ÿàÿ ä x   ÿØ� â$rÿÿ   ÿ Ž â x `   p`ð â $x   ÿ Ž â x   ÿÀ‚ Ä zÿ z    Ö G™  @    € ê $z |  	Ž â x "       â $x
   ÿ Ž ä $z v  Ž ä$rÿÿ   ÿ Ž â x   
áÿ â $z p  Ž ä %x

   ÿŽ â xÿ     Æ %v X  Ž â x
   ÿüŽ ä x!     â $rÿÿ    Ž â r        ä r	   ÿàó â $tÿ$   ÿ Ž â r"        â $r#ÿÿ    Ž â rÿ   ÿäÿ  â %v Z   Ž Æ �y""    • ¢ $x$   ÿŽ â r   ÿàõ ä x   ÿàó â $x   Ž ä $x    ä $rÿÿ   Ž  â r   ÿ‚ Ä �y    • â %x   Ž Æ �y%   • â r!   ÿäÿ  Æ �y    • ( �y'   • ( �y    ™ h �y   ™ h �y    • b xÿÿÿÿÿàÿ È r ÿ   pRò ä x@  ÿàõ ä x		    ÿàÿ Æ $rÿÿ    ä 0r"ÿ"    A   âO xT  %    ä� x  T  '    Æ&r   ÿ   È&r       ä 0r#ÿ    A   Æ Er        ä  r""#     A  È #r"      â G   ýÿÿÿÿƒ ê Ay       € ê $vÿ X  ÿ Ž â z  Y      â G‰       € ê $z |  	Ž â x
ÿ     â $x   ÿ Ž ä $z v  Ž âx
   
  äx   áÿ â $z p  Ž â x   ÿ   Ä xÿ     â %x"   ÿŽ Æ x     â$x   ÿ Ž ä %x   Ž È r	   ÿàñ äx	    ÿàÿ ä x6 $       â $rÿÿ     â r   ÿàó Æ $x$   ÿŽ â x	@   ÿàÿ â $rÿÿ   Ž  ä %v Z  6 Ž â r   ÿàõ Æ %v Z  6 Ž â r   ÿàÿ ä r    ÿà÷ â$x#$   ÿŽ â r(   ÿàñ â$rÿÿ    â r&   ÿàõ â %v Z  6 Ž â �y    • ¢  r!   ÿäÿ Ä $x   #Ž â r)
   ÿä  â $x$   ÿŽ â r"   ÿà÷ â $x'   
 â r<   ÿàõ â �y((   ™ â r   ÿàÿ ä x	`   ÿàÿ â $x=    â r2   ÿàñ â �y%    ™ b	 r7   ÿàõ Ä r#   ÿäÿ â �y+<   • ¢ rÿ   ÿä â %v67 Z  6 Ž â r.   ÿàó â �y<   • ¢ r8   ÿà÷ â $x-$   ÿŽ ä �y&&   ™ â  $x3     â r9   ÿäÿ Ä �y)<   • b/ r06   ÿàñ â $x/   
Ž  âr77-   ÿàÿ â �y <   • h�y'<B  • ( �y$<D  • " $x17   
  Æ �y    • ( �y""   ™ ( �y<‚  • ( �y<H  • ( �y#<F  • (/ �y<„  • " r:6   ÿàó Æ �y    • ( �y!8    • ( �y*<†  • ( �y.   ™ ( �y-<ˆ  • " r;7   ÿäÿ  Æ �y8@  • ( �y.<Â  • ( �y/<Ä  • ( �y2   ™ ( �y00   ™ ( �y,8€  • ( �y2<È  • ( �y1<Æ  • (/ �y3:   ™ ( �y48À  • ( �y66    • " x		€   ÿàÿ È | 	   pbð ä x   ÿàó Ê $rÿÿ   Ž  â x++T      ÊO &r(+(   ÿ   â� x )T       â0rÿ    A   â x''T  $    Æ&r  %   (   ä &r&'&   ÿ   Æ Er          â 0rÿ    A   â x##T      ä xT      Æ &r"#"   &   ä 0r!ÿ!    A   Æ Er" "       â 0rÿ    A   â  r!     A  â &r   ÿ   â x-*T  -    â 0rÿ    A   â #r       â x/.T  /    Æ  r     A  â &r-      ä &r//0   ÿ   ä 0r,ÿ,    A   â Er        ä x21T  2    â #r"      â  r,     A  Æ &r/23   /   ä 0r4ÿ4    A   ä 0rÿ6    A   â Er/ /       â #r      Æ  r4     A  È #r/      â G‰  °øÿÿÿÿƒ ê Ay        € êŸ r ÿ   pRð är ÿ   pRò Ö ˆ €ÿÿ H   è {          ì M        € ê „y     H   " r ÿ   pRð â $z  x  Ž Ä „y €   H   h „y	    H   ¢ !r       È !r       È/ !r	       ÊO ‰      $ !r       Ê ‰      $ !r       Ê ‰	 €    $ !r		       Ê ‰	 @    $ !r	       Ê ‰      b  M	        € ê x        â $z ~  Ž â !r       Æ/ %v j   Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y	      &   " z ÿ \  pPò â ¹z  F   
   Æ z ÿ ]  Sò Ú $ÿ   ÿ Ž È %	 \   Ž Ê �    ™ b r ÿ       â Ey  0    € ä y      "   d ' 	 m   Ž äy      !   d 'š	 s   Ž â  ÿ n     â y      '   ¢ r ÿ       Æ $ÿÿ    
Ž È $  o  	Ž ä$’ ÿÿ   	 Ž â rÿ   ÿü€ â/ $x    Ž Ê xÿ     ä šÿ t    ä r ÿ       Æ G	       € ê z ÿ `  pPô äO z ÿ ^  pPð ä z ÿ b  p ö ä z ÿ a   Sô ä z ÿ _   Sð ä z ÿ c  0't Ö 	      '   " 	     € ä¢	     € â ©
      '   d $
 x  ÿŽ ä 	      %   ¨ ©      %   â $
 ~  ÿŽ Ä $ª	
 ~  ÿŽ ä/ $ª
 x  ÿŽ â    àë äOÿ     ä ¢		   àÇ ä�   ÿàó ä ¨ÿ     Ä ÿÿ   Äþ ä ¢

	   ÿàû ä ¢ÿÿ   „þ â $ÿÿ   Ž  â 
 ^  ÿ‚ ä ¢ÿ   ÿäÿ ä ª
 b  ÿ† Ä 
	 _  �  â $rÿÿ   ÿ Ž â r ÿ       ä ª
 c  � Æ �	     h �©     d Ay        € êO r   ÿ� â ¹z  l      â $z	 }  ÿŽ â ™x?    â x
ÿ     â $z   w  ÿŽ â Ey  °    € â $rÿÿ   ÿ Ž â r' ÿ       Ä | 
   pbð ä r	    ÿàó Ê $r	ÿÿ   ÿŽ  Ì G	  @    € ê r ÿ
   ÿ3Ž â Ey €    € â r ÿ       â $z v  ÿŽ â|     ÿàÿ â $rÿÿ   ÿ Ž â xüÿÿÿÿÀŽ â 'z y   Ž â xÿ     ÿÀ€ ä r' ÿ       â $x   
Ž â zÿ z    Ò G	  À    € ê y      %   " r
   ÿàñ ä x $       ä x "       ä x   ÿ   â %v Z   Ž â x   ÿ   È x!   áÿ â $z p  Ž È $z |  Ž ä $rÿÿ   	  â x   ÿàñ â $x
   Ž ä $x$   ÿŽ ä %v X  Ž È $x   Ž ä %x   Ž ä �y    • ¤ $rÿÿ     ä �y   • ä %x   Ž Ä �y   • ä %x!   Ž ä �y   • ( �y#   • ( �y    ™ h �y    • h �y    ™ b z ÿ `  pPð È z ÿ a   Sð ä xT      â� 0r'ÿ    A   âO x#T  #    Æ&r   ÿ   ä0rÿ    A   ä &r#      Æ  r''     A  â r ÿ       ä Er        â G‰       € ê xÿ     â %v `   Ž È $x"   ÿŽ È $x   Ž È %x   Ž ä �y     ¨ �y    è �y    è �y    ( �y    " xT      Ê� &r   ÿ   â xT      Ê&r      ä 0rÿ    A   ÆO Er        ä  r     A  È #r   ÿ   â x

    ÿàÿ â #r''   ÿ   Ê Ay       € ê xÿ àÿÿÿÿÀ€ Ú G‰  €    € ê y      %   " $z  |  
Ž â z ÿ `  pPð â $x   ÿ Ž âx   ÿ   ä r     ÿàÿ â $vÿ X  ÿ Ž â x"   áÿ â$x    ÿ Ž â xÿ     Ä x!ÿ   "  ä z  Y      ä x!"   !  ä z ÿ a   Sð ä z  a      ä x     â $z  p   Ž â x""   ÿ   â $vÿ `  ÿ Ž Ä %x "   ÿŽ Ì %x   Ž È r
   ÿàó â $tÿ$   ÿ Ž â r   ÿàõ ä rÿ	   ÿäÿ  â %v Z   Ž â x 
    ÿàÿ Æ $x    â r    ÿàõ â $x$   ÿŽ â r    ÿàó ä �y$   • ¢  $rÿÿ   	 â r   ÿàÿ â %v Z   Ž â �y%   • ¤  r   ÿäÿ  â $x$   ÿŽ â r,"   ÿàõ â �y#   • ä  $x   Ž Ä �y*   • è  �y+B  • ( �y(D  • ( �y)F  • h �y&H  • b $x-!    Æ �y   ™ b r"   ÿàõ Æ �y ,   ™ b r   ÿà÷ Æ �y    • b r    ÿàó ä/ r   ÿäÿ â $x   ! âr   ÿäÿ  â �y    • h �y,    • h �y   ™ h �y   ™ h �y@  • b x%$T  %    ÄO x##T  *    ä� x++T  (    â&r%%   ÿ   âx))T  &    Æ &r%#    %   ä 0r#ÿ    A   Æ Er% %       ä r   ÿàó â 0r,ÿ,    A   ä &r++   ÿ   Æ  r$,#     A  â 0rÿ    A   ä &r))   +   ä 0rÿ    A   â #r$$%   '   â r   ÿäÿ  ä Er' )       â  r     A  â G‰  ð     € ê r   ÿàó â �y    ¨ �y&    ¢ $x   Ž  Æ �y    è �y(    è �y     " xT  &    ÊO &r   ÿ   â xT  (    Ê� &r       ä 0r ÿ    A   ÆEr        ä  r##      A  È #r#      â #r''   $   â G‰  ð     € ê �y B   ¢ r   ÿàó Æ �yD   ¢ r   ÿäÿ  Æ �yF   è �yH   è �y@   " x T      ÊO &r   ÿ   â xT      Ê� &r      ä 0rÿ    A   ÆEr        ä  r     A  È #r      â x

@   ÿàÿ ä x€  ÿàõ ä | 
   pbò ä x€  ÿà÷ â $rÿÿ    Æ rÿ   ÿäÿ Î G™  Ðùÿÿÿÿƒ ê Ay        € ê r ÿ   pRð âEy  €     € â r ÿ   pRò Ö G‰  P     € ê z ÿ `  pPð â $x    Ž Æ z ÿ a   Sð ä ˆs '€ÿÿ H   ö ˆ    H   ä Ay        € ê {          ì M        € ê z ÿ `  pPð â „y      H   f z ÿ a   Sð Ú „‰ €   H   ¨ „‰    H   " !r' '       È/ !‚'       ÈO !‚       â G‰        € ê „y €  H   ( „y €   H   h „y     H   ¨ „y	    H   è „y €  H   " !r       â !r'       Æ/ !r         âO !r	       Æ� !r        ä‰      " z ÿ `  pPð ä r ÿ   pRò ä z ÿ a   Sð â !r       Ê ‰       $ !r         Ê ‰  €    $ !r        Ê ‰
 @    $ !r

       Ê ‰	
      $ !r
	       âG‰        € ê ‰      $ !r       Ê ‰       $ !r         Ê ‰  €    $ !r        Ê ‰ @    $ !r       Ê ‰	      $ !r	       ä M        € ê !r       â G‰  `    € ê z   h      â !r       Æ x     pBð Ú G	  p    € ê x  ÿÿÿÿÿàÿ È x       € Ê $x     ÿ Ž È ‚{   €     $ xÿ     È Iy Ðêÿÿÿÿƒ ê  x ;ª¸¿  A  Ì s          $ !t   €?    Ì s        $  r     A  È  r     A  â Gy  P    € ê  x '7=  A  â x*BL?  A  Æ #t   €?    È  r      A  â  x    ?  A  Ê s     $   $ !t  €?    È  r       A  È  r      A  â Gy  °    € ê x     pRð Ú G‰  à     € ê x     pRð Ú G	        € ê 	z  i    � ä 	z i    � Æ  x ;ª¸¿  A  â 	z i €  � Ê s        $ !t  €?    Ì s        $  r       A  È  r      A  â Gy  Ð     € ê  r     A  â Gy  °     € ê 	x   à@  � ä 	x  à@  � Æ  x #ÛÙ?  A  â 	x  àÀ  � Æ  x;ª¸¿  A  ä !t  €?    È s        $ !t  €?    Ì s        $  r       A  È  r      A  ä y      &   " x        Æ y       %   $ $z x   Ž È $z ~  Ž È %v j   Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú         ä r        Ê % \   Ž Ê �	    ™ b ¹z  l      â Ey  �    € â ™x?    â y      "   b $rÿÿ   ÿ Ž â r ÿ       ä y
      !   ¢ r ÿ       â $ÿÿ    Ž â  ÿ       â y      %   " r	 ÿ       Æ '
  m   Ž â ‚ ÿ       â y      '   æ 
 ÿ n     È   ÿ   ÿáÿ ä r ÿ   pRò Æ/ $
  o  Ž â r 
   ÿ(Ž ÈO ‚        â $x   ÿ Ž â xÿ      Æ 'Š s   Ž â r    ÿ� ä 
   ÿ8Ž ä xÿ     È |    pbô ä Šÿ t    ä r ÿ   pRð Ò G)  °    € ê r ÿ       â� $z w  ÿŽ â xüÿÿÿÿÀŽ ä r        â 'z y   Ž â r   €ÿàÿ Æ $z  p  Ž â zÿ z    â $x   ÿ Ž âx   ÿ   â $z }  ÿŽ â z  X      â $z |   Ž â x   áÿ ä r   ÿàõ â $z v  Ž âxÿ     Ä xÿ     â %x."   ÿŽ â z   Y      ä r ÿ       ä rÿÿ   ÿä â %x   .Ž â x     ä x     Ä r4   ÿàõ â $t5ÿ$   ÿ Ž â r"   ÿà÷ Ä rÿ   ÿä â %v44 Z  5 Ž â r#    ÿäÿ Æ $x$   ÿŽ â z  p      â �y"   • ¢  r,.   ÿàù ä r(4   ÿ„ â %x&"   "Ž â r55   ÿàÿ â �y"   • â  r-/    ÿä ä r 4   ÿ† â %x*"   &Ž â�y"   • ¦  $x)5    âr!5   ÿäÿ â %x0"   *Ž â�y"   • æ  %x2"   ,Ž â�y&   • ( �y&   • ( �y&   • h �y&   • h �y(   ™ h �y#0   • h �y$0   • h �y'0   • h! �y(0   • h �y     ™ h �y"*   • b %x0"   2Ž Æ �y%*   • h �y!*   • h! �y&*   • h �y),    • h �y*4    • h �y+2    • h �y,0    • b/ %x2"   0Ž Ê �y2    • b x    ÿàÿ È |    pbô ä x@  ÿà÷ È r ÿ    ÿäÿ ä xT      äO xT      ä� xT      â&r   ÿ   âxT      Æ &r   ÿ   â x$#T  $    Ê &r$$   ÿ   â x''T  (    â &r       ä&r       â x"!T  "    ä x%%T  &    Æ &r"   ÿ   ä &r$'    $   ä&r%       ä 0r*ÿ*    A   ä 0r)ÿ)    A   ä 0rÿ,    A   Ä 0r+ÿ+    A   Æ  r*     A  â r)*     A  â  r*+     A  â0rÿ    A   â Er        ä Er        ä Er        â  r**     A  â Er$ $       â #r	   	   â #r      â #r      Ä #r*$      â G©  àúÿÿÿÿƒ ê Ay        € ê� ˆ 	 þÿ H   è ˆ €þÿ H   è ˆ  ÿÿ H   è ˆ €ÿÿ H   è {          ì M	        € ê „y 
     H   " z  r  p`ô â $z x  Ž Ä „y
 €   H   b r 
ÿ   pVt Æ „y
    H   ¨ „y
    H   è „y
 €  H   ( „y
    H   ( „y
    H   ( „y
 €  H   ( „y
 €  H   ( „y
    H   ( „y
 €  H   " !r  	       Æ „y
 €  H   "!r       â/ !r       âO !r         ä� !r       â!r       â !r         â !r       È ‰       b !r       â !r       È ‰      ¢ !r       Æ ‰      â !r       Ê ‰      " !r        â/ !r       ÈO ‰      b !r       Æ� ‰      ¨ ‰	      â !r       Ê ‰      " !r       â/ !r       ÈO ‰ €    b !r		       Æ� ‰  €    ¨ ‰	 €    â !r       Ê ‰ €    " !r       â/ !r        ÈO ‰  @    b !r	       â� x	   ÿàÿ ä ‰ @    ¤ z 	 r  p`ò ä ‰ @    â !r       â x 
   pVò  È ‰ @    " !r         â/ !r       âO x   ÿàÿ Æ ‰       b !r       â� z  r  p`ð ä ‰      ¤ x 
   pVp  ä ‰      â !r       â x   ÿàÿ Ä x        ä ‰      " z  r  p`ö â $z ~  Ž Æ x 
   pVö â !r	        â/ %v j   Ž â !r       ÈO †© 	    â !r       Ê� †¹    è †‰    â !r       â M        € ì †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y	      &   " z ÿ \  pPò â ¹z  F   
   Æ z ÿ ]  Sò Ú         Ê %	 \   Ž Ê �    ™ ¢  r ÿ       â Ey  0    € â r ÿ       â y      "   b r ÿ       â ' 	 m   Ž äy      %   ä 'š	 s   Ž â  ÿ n     â y      !   &  ÿ   ÿáÿ â y      '   " r ÿ   pRô Ä/ x   ÿ   ä� x    pDp är   ÿàÿ är    ÿ(Ž ä z  r  pdp  â $ o  	Ž â xÿ      Ä ’ 	       ä šÿ t    È r	     € ÆO G	  À    € ê z ÿ `  pPò ä z ÿ ^  pPð ä z ÿ b  p ö ä z ÿ a  Sò ä z ÿ _   Sð ä z ÿ c  0'ò Ä r ÿ       ä r ÿ       Î $
 ~  ÿŽ âÿ     â$š	 ~  ÿŽ ä $
 x  ÿŽ â    àë â$š x  ÿŽ â ’
	   àÇ ä ˜	ÿ     ä ÿ   ÿÄþ Ä    ÿàý ä ’

   ÿàû ä ’	ÿ	   ÿ„þ ä ÿ   ÿä ä 
 ^  ÿ† ä ’	ÿ	   ÿäÿ Ä š
 b  ÿˆ ä
 _  � ä š
 c  	 Æ �	     h �™     d Ay        € ê r    ÿ� â ¹z  l      â $z }  ÿŽ â ™x?    â x	ÿ     â $z w  ÿŽ â r ÿ       â Ey  p    € â r ÿ       â x      ÿ  â | 	   pbð â 'z y   Ž â r   ÿàó â x      ÿ  â x      ÿ  â r ÿ       Ä r
ÿÿ   ÿäÿ  Ì G	  À    € ê zÿ z    â $z p  	Ž âxüÿÿÿÿÀŽ â x      ÿ  â x   ÿàÿ â $z |  	Ž â r    €ÿàÿ â $z |  Ž â x   ÿàÿ â $z v  Ž âx    ÿ   â$z v  Ž â x    ÿ   â $z p  Ž â x   ÿàÿ â %x8   ÿŽ â x    áÿ âx      ÿ  â xÿ     â $z p  Ž âz ÿ `  pPð â $z p  Ž â x!ÿ      â %x6"   8Ž â x     â x      ÿ  â z  `      â %x"   8Ž â z  a      Ä z  X      â %x"   8Ž âz  Y      ä r ÿ       â %x8"   8Ž â r ÿ       ä z ÿ a   Sð ä x    !  Ä r$	   ÿàó ä x% $       Ä rÿ
   ÿäÿ  Æ %v$$ Z  % Ž È $x$   ÿŽ â r&$   ÿ‚ ä r"        ä r%%   ÿàÿ ä r#        är'%   ÿäÿ  â �y $    • ¢  r,8   ÿàó â %x:"   "Ž â z/  p      â �y&   ™ â r-9   ÿäÿ  Ä r* $   ÿ† â %x2/"   :Ž ä�y!,   • ä %x0/"   ,Ž â r+%   ÿäÿ â �y",   • ä %x</"   2Ž ä�y%:    • ä %x4/"   0Ž Ä �y'0   • è" �y(0   • â %x:/"   4Ž Æ �y&2    • è  �y#,   • è �y$,   • â %x2/"   <Ž Æ �y)0   • è �y*   ™ è  �y,0   • (�y/:   • ( �y*4   • ( �y+4   • ( �y0:   • (/ �y-4   • ( �y.4   • ( �y22    • ( �y1<    • ( �y4:   • ( �y3:   • "/ 0r ÿ     A   âO x!!T  "    Ê� &r"!   ÿ   â x''T  (    Ê &r''   ÿ   ä 0r!ÿ&    A   â x##T  $    â 0r%ÿ%    A   â x,)T  ,    Æ&r"#   "   ä&r',   '   â x+*T  +    ä x//T  0    Æ &r++   ÿ   âEr' '       â &r//   ÿ   â x(-T  .    â  r! !     A  â Er" "       â  r$%      A  â 0r#ÿ2    A   ä &r((   +   ä 0r1ÿ1    A   â x43T  4    â #r!'      È &r/4   /   â #r$"      â  r& 1     A  â r' #     A  â Er( (       ä Er! /       â G‰       € ê r$8   ÿàó är"        ä r%9   ÿäÿ  Ä r#        Æ �y)$    ¨ �y*$    ¢ %x""   "Ž Æ �y+$    è �y,$    è �y""     " x))T  *    ÊO &r))   ÿ   â x,+T  ,    Ê� &r,,   )   ä 0r)ÿ"    A   ÆEr, ,       ä  r* )     A  È #r*,      ä G‰  @    € ê r$6   ÿàó äx)   ÿàÿ ä r%7   ÿäÿ  är"        â $z+) p  Ž â r#        â �y)$    ¨ �y*$    ¢ %x"+"   "Ž Æ �y+$    è �y,$    è �y""     " x))T  *    ÊO &r))   ÿ   â x,+T  ,    Ê� &r,,   )   ä 0r)ÿ"    A   ÆEr, ,       ä  r) )     A  È #r),      ä G‰  @    € ê r$   ÿàó äx)   ÿàÿ ä r%   ÿäÿ  är"        â $z+) p  Ž â r#        â �y)$    ¨ �y*$    ¢ %x"+"   "Ž Æ �y+$    è �y,$    è �y""     " x))T  *    ÊO &r))   ÿ   â x,+T  ,    Ê� &r,,   )   ä 0r)ÿ"    A   ÆEr, ,       ä  r* )     A  È #r*,      â #r&(      â #r'!      â G‰  @    € ê r$   ÿàó äx!   ÿàÿ ä r%   ÿäÿ  är"        â $z'! p  Ž â r#        â �y!$    ¨ �y&$    ¢ %x"'"   "Ž Æ �y'$    è �y($    è �y""     " x!!T  &    ÊO &r!!   ÿ   â x('T  (    Ê� &r!(   !   ä 0rÿ"    A   ÆEr! !       ä  r      A  È #r!      â x		    ÿàÿ ä x@  ÿà÷ ä | 	   pbò ä x@  ÿàù ä rÿ   ÿäÿ ä rÿ   ÿä Î G™  �õÿÿÿÿƒ ê Ay        € ê y       "   " Ey       € â r  ÿ   pRð Ú G‰  ð     € ê z ÿ `  pPð ä r    ÿ8Ž ä z ÿ a   Sð Æ ˆs  þÿ H   ô ˆƒ €þÿ H   è ˆƒ  ÿÿ H   è ˆƒ €ÿÿ H   â G‰  p     € ê ˆs    H   è ˆs €þÿ H   è ˆs €  H   è ˆs  ÿÿ H   è ˆs    H   è ˆs €ÿÿ H   è ˆs €  H   ä Ay        € ê {          ì M)        € ê y      &   ¢ z ÿ `  pPð Æ y      '   "> z ÿ a   Sð â $z  x  Ž âO x        Æ $z  ~   Ž È %v  j   Ž È G	  P    € ê „y      H   " z  r  p`ð Æ „y	 €   H   b r ÿ   pVp  Æ „y    H   ¨ „y    H   è „y €  H   ( „y    H   ( „y    H   ( „y €  H   ( „y    H   („y €  H   ( „y
 €  H   " !r         Æ „y €  H   " !r		       â/ !r       âO !r         ä� !r	       â!r       â !r         ä !r       Æ ‰       b !r       Æ ‰      ¢ !r       Æ ‰      â !r

       È !r
       â !r        â/ !r       ÈO ‰      " !r
       Æ� ‰      h ‰      ¨ ‰	
      â !r       â !r       â/ !r       ÈO ‰      " !r
	       Æ� ‰ €    h ‰  €    ¨ ‰	 €    â !r       â !r       È/ ‰ €    " !r
        ÆO ‰  @    b !r		       â� x   ÿàÿ ä ‰
 @    ¤ z  r  p`ô ä ‰	 @    ä x    pVt â !r       â !r         È/ ‰ @    " x   ÿàÿ â !r
       âO x   ÿàÿ â ‰       b z  r  p`ò â !r	       ä� ‰      ¢ x    pVò  Æ ‰      â !r       â !r        È/ ‰       " !‚       â !r       âO !r       Æ� †‰     â z  r  p`ð â !¢	       â!’       ä x    pVp  ä †© 	   è †™    â !r         Ì M	        € ê !r        Ê †y    â My        € ê z   h      È x     pBð Ú G	  `    € ê x  ÿÿÿÿÿàÿ È x       € È x     ÿ   È ‚{   €     $ xÿ     È Iy pæÿÿÿÿƒ ê „y     H   " z  r  p`ð â Ey  p    € ä „y     H   b r ÿ   pVp  Æ „y	    H   ¨ „y    H   è „y    H   ( „y  
  H   ( „y €  H   è „y €  H   â !r       Æ „y €  H   " !r         â/ !r	       ÆO !r         â� !r       Æ!r         ä ‰      h ‰       ¢ !r       â !r       ä/ „y €   H   â !r        ÆO ‰      ( ‰      b !r       â !r	       È/ ‰ €    ( ‰
	 €    b !r       Æ „y €  H   ¢ !r       È !r       â !r       ä „y €
  H   â !r
	
       Æ/ ‰ @    ( ‰
 @    b !r       âO !r       â !r
       È/ ‰       ( ‰      b !r       â !r         â !r       â/ G	  �     € ê !r       â!r         Æ  x;ª¸¿  A  Ì s        $ !t  €?    Ì s        $  r     A  È  r      A  Ê †y     ä Ay        € ê ‰       b Ey  Ð    € æ ‰      ¨ „y    H   è „y    H   ( „y  	  H   â !r         Æ/ „y    H   â !r       ÆO ‰       h ‰      ( „y    H   ¢ !r       È� !r       â!r        â/ !r	       È ‰ €    " x   ÿàÿ Æ ‰
	 €    b z  r  p`ð â !r       Æ x    pVp  â !r       ÈO !r       â !r       ä „y    H   â !r
	
       Æ/ ‰ @    ( ‰
 @    b !r       â !r
       È/ ‰       ( ‰      b !r       â !r         â !r       â/ G	  �     € ê !r       â!r         Æ  x;ª¸¿  A  Ì s        $ !t  €?    Ì s        $  r     A  È  r      A  Ê †y    ä Ay        € ê ‰       b Ey       € æ ‰      ¨ „y €  H   è „y €  H   è „y €  H   è „y €  H   â !r         Æ/ „y €  H   â !r       ÆO ‰       h ‰      ( „y €	  H   ¢ !r       â� !r        â/ !r	       È ‰ €    ( ‰
	 €    b !r       ÈO !r       â !r       â x   ÿàÿ â !r
	
       Æ/ ‰ @    " z  r  p`ð Æ ‰
 @    b x    pVp  â !r       â !r
       È/ ‰       ( ‰      b !r         â !r       â/ G	  �     € ê !r       â!r         Æ  x;ª¸¿  A  Ì s        $ !t  €?    Ì s        $  r     A  È  r      A  Ê †y    ä Ay        € ê ‰      b !r       â x   ÿàÿ Æ !r       â z  r  p`ð Æ !r       â x    pVp  È ‰       ¢ !r       Ê/ ‰      " !r         ÊO ‰       b !r	       Ê ‰
	 €    " !r        Ê/ ‰ €    b !r
	
       Ê ‰
 @    " !r       Ê/ ‰ @    b !r
       Ê ‰      " !r       Ê/ ‰       ¢ !r       â M	        € ì !r       â/!r         ÆO  x;ª¸¿  A  â !r        Ê s        $ !t  €?    Ì s        $  r      A  È  r      A  Ê †y    â My        € ê „y      H   " z  r  p`ð â Ey       € ä „y    H   b r ÿ   pVp  Æ „y    H   ¨ „y    H   è „y	    H   ( „y  
  H   ( „y €  H   è „y €  H   è „y €
  H   â !r         Æ „y €  H   " !r       Æ/ „y €   H   â !r         âO !r       Æ� !r	 	       â!r       È ‰ 	      h ‰      ¢ !r       â !r 	        â/ !r       ä !r       â !r       âO ‰       $ !r       Ä ‰      b !r        â !r       È/ ‰ €    ( ‰ €    h „y €  H   ¢ !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       ÈO !r       â !r       â !r
	
       È/ ‰       ( ‰
      b !r        â !r 
       â/ G	  À     € ê !r         È x '7=  A  â x *BL?  A  Æ #t   €?   È  r     A  â  x    ?  A  â !r        È s     $   $ !t  €?    È  r     A  È  r      A  Ê †y     ä Ay        € ê ‰       b Ey       € æ ‰      ¨ „y    H   è „y    H   ( „y    H   è „y    H   â !r         Æ/ „y    H   â !r       ÆO ‰       h ‰      ( „y  	  H   ¢ !r       â� !r       â!r        â/ !r       È ‰ €    ( ‰ €    b !r       È !r       â !r       ÈO !r       â !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       â x   ÿàÿ â !r
	
       Æ/ ‰       " z  r  p`ð Æ ‰
      b x    pVp  â !r        â !r 
       Ö/ G	  À     € ê !r         È x '7=  A  â x *BL?  A  Æ #t   €?   È  r     A  â  x    ?  A  â !r        È s     $   $ !t  €?    È  r     A  È  r      A  Ê †y    ä Ay        € ê ‰       b Ey  Ð    € æ ‰      ¨ „y €  H   è „y €	  H   ( „y €  H   è „y €  H   â !r         Æ/ „y €  H   â !r       ÆO ‰       h ‰      ( „y €  H   â !r       È� !r       â!r        Æ/ !r       â !r       â ‰ €    ( ‰ €    b !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       â x   ÿàÿ â !r
	
       Æ/ ‰       " z  r  p`ð Æ ‰
      b x    pVp  â !r        â !r 
       Ö/ G	  À     € ê !r         È x '7=  A  â x *BL?  A  Æ #t   €?   È  r     A  â  x    ?  A  â !r        È s     $   $ !t  €?    È  r     A  È  r      A  Ê †y    ä Ay        € ê ‰      b !r       â x   ÿàÿ Æ !r       â z  r  p`ð Æ !r       â x    pVp  È ‰       ¢ !r       Ê/ ‰      " !r         ÊO ‰       b !r       Ê ‰ €    " !r        Ê/ ‰ €    b !r	       Ê ‰
	 @    " !r       Ê/ ‰ @    b !r
	
       Ê ‰
      " !r       Ê/ ‰       ¢ !r
       â M	        € ì !r       È/ x'7=  A  â x*BL?  A  Æ #t  €?   â !r        âO  x   ?  A  ä  r     A  â !r       Ê s     $   $ !t   €?    È  r       A  È  r      A  Ê †y    â My        € ê x     pRð Ú G‰  P    € ê x     pRð Ú G	  `    € ê „y      H   " z  r  p`ð â Ey       € ä „y    H   b r ÿ   pVp  Æ „y    H   ¨ „y    H   è „y	    H   ( „y  
  H   ( „y €  H   è „y €
  H   â !r         Æ „y €  H   " !r       Æ/ „y €   H   â !r         âO !r       Æ� !r	 	       â!r       È ‰ 	      h ‰      ¢ !r       â !r 	        â/ !r       ä !r       â !r       âO ‰       $ !r       Ä ‰      h „y €  H   â !r        â !r       È/ ‰ €    ( ‰ €    h „y €  H   ¢ !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       ÈO !r       â !r       â !r
	
       È/ ‰       ( ‰
      b !r        â !r 
       â/ G	  À     € ê !r         Ê	z  i    � â !r        È  x;ª¸¿  A  â 	z   i    � È 	z   i €  � â s        $ !t  €?    Ì s        $  r     A  È  r      A  Ê †y     ä Ay        € ê ‰       b Ey       € æ ‰      ¨ „y    H   è „y    H   ( „y    H   è „y    H   â !r         â/ !r       ÈO ‰       h ‰      ( „y  	  H   ¢ !r       â� !r       â!r        Ä/ !r       â !r       ä ‰ €    ( ‰ €    h „y    H   â !r       ÈO !r       â !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       â� !r       â x   ÿàÿ â !r
	
       Æ/ ‰       " z  r  p`ð Æ ‰
      b x    pVp  â !r        â !r 
       Ö/ G	  À     € ê !r         Ê	z  i    � â !r        È  x;ª¸¿  A  â 	z   i    � È 	z   i €  � â s        $ !t  €?    Ì s        $  r     A  È  r      A  Ê †y    ä Ay        € ê ‰       b Ey  Ð    € æ ‰      ¨ „y €  H   è „y €	  H   ( „y €  H   è „y €  H   â !r         Æ/ „y €  H   â !r       ÆO ‰       h ‰      " !r       È� !r       â!r        â/ !r       È ‰ €    ( ‰ €    h „y €  H   ¢ !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       âO !r       â x   ÿàÿ â !r
	
       Æ/ ‰       " z  r  p`ð Æ ‰
      b x    pVp  â !r        â !r 
       Ö/ G	  À     € ê !r         Ê	z  i    � â !r        È  x;ª¸¿  A  â 	z   i    � È 	z   i €  � â s        $ !t  €?    Ì s        $  r     A  È  r      A  Ê †y    ä Ay        € ê ‰      b !r       â x   ÿàÿ Æ !r       â z  r  p`ð Æ !r       â x    pVp  È ‰       ¢ !r       Ê/ ‰      " !r         ÊO ‰       b !r       Ê ‰ €    " !r        Ê/ ‰ €    b !r	       Ê ‰
	 @    " !r       Ê/ ‰ @    b !r
	
       Ê ‰
      " !r       Ê/ ‰       ¢ !r
       â M	        € ì !r       â/!r         ÈO 	z i    � â !r        È  x;ª¸¿  A  â 	z i    � È 	z i €  � â s        $ !t  €?    Ì s        $  r      A  È  r      A  Ê †y    â My        € ê „y     H   " z  r  p`ð Æ „y    H   b r ÿ   pVp  Æ „y    H   ¨ „y €   H   è „y €  H   ( „y    H   ( „y     H   ( „y"    H   ( „y    H   ( „y €  H   ( „y €  H   " !r       Æ „y%    H   " !r       Æ/ „y'  	  H   b !r       ÆO „y  
  H   ¢ !r       Æ� „y! €  H   â !r       Æ„y# €
  H   " !r       Æ „y
    H   " !r         Æ „y	    H   " !r""       Æ „y €  H   " !r       Æ „y €  H   " !r       Æ „y  €  H   " !r       Æ „y €	  H   â !r% %       Æ „y €  H   â !r""'       Æ/ „y €  H   â !r       ÆO ‰      " !r!       â� !r##       ä‰      b !r%
       Æ ‰
#      ¢ !r"	       Æ ‰	      â !r       Æ ‰      " !r       Æ ‰      " !r        â !r       â !r       â !r       â/ x   ÿàÿ â !r       ä ‰      " !r
#
       âO !r       â z  r  p`ô â ‰      b !r		       Ä� x    pVt â ‰      ¢ !r       â!r       ä ‰       è ‰	      ( ‰
      " !r       Æ ‰      " !r       Æ/ ‰      b !r       ÆO ‰      ¨ ‰      ¢ !r        â� !r	       ä‰  €    â !r
       Æ ‰	 €    " !r       Æ ‰ €    " !r
       Æ/ ‰ €    b !r       ÆO ‰ €    ¢ !r       Æ ‰
 €    ¨ ‰ €    ¢ !r         Æ� ‰ €    â !r		       â!r       ä ‰  @    " !r       Æ/ ‰ @    b x   ÿàÿ â !r       äO ‰	 @    ¢ z  r  p`ò â !r
       ä ‰ @    " x    pVò  â !r       Ä ‰
 @    " x   ÿàÿ â !r       ä� ‰ @    â z  r  p`ö Æ ‰ @    â !r        â x    pVö ä ‰ @    " !r       â/ !r	       äO ‰       b !r       Æ‰      ¢ !r

       Æ ‰	      " !r       Æ� ‰      â !r       Æ ‰
      â !r       Æ ‰      ( ‰      " !r         Æ/ ‰      b !r       âO !‚         â!r	       ä!‚       â !r       â� !’       ä  ‚      A  â !r

       â !’       Æ †‰     â !r       â !¢
       â  ’     A  ä !¢	       â !r       ä †™    â  ¢		     A  â !r        È/ †© 	   â M9        € ê !r       â !r         È  r      A  Ê †y    â My        € ê „y      H   " z  r  p`ð â Ey  À    € ä „y    H   b r ÿ   pVp  Æ „y    H   ¨ „y    H   è „y	    H   ( „y  
  H   ( „y €  H   è „y €
  H   â !r         Æ „y €  H   " !r       Æ/ „y €   H   â !r         âO !r       Æ� !r	 	       â!r       È ‰ 	      h ‰      ¢ !r       â !r 	        â/ !r       ä !r       â !r       âO ‰       $ !r       Ä ‰      h „y €  H   â !r        â !r       È/ ‰ €    ( ‰ €    h „y €  H   ¢ !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       ÈO !r       â !r       â !r
	
       È/ ‰       ( ‰
      b !r        â !r 
       â/ G	  à     € ê !r         Ê	x   à@  � Ê  x #ÛÙ?  A  È  x ;ª¸¿  A  â !r        Ê s        " 	x    à@  � È 	x    àÀ  � Ê !t   €?    â !t  €?    È s	        $  r	     A  È  r     A  Ê †y     ä Ay        € ê ‰       b Ey       € æ ‰      ¨ „y    H   è „y    H   ( „y    H   è „y    H   â !r         â/ !r       ÈO ‰       h ‰      ( „y  	  H   ¢ !r       â� !r       â!r        Ä/ !r       â !r       ä ‰ €    ( ‰ €    h „y    H   â !r       ÈO !r       â !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       â� !r       â x   ÿàÿ â !r
	
       Æ/ ‰       " z  r  p`ð Æ ‰
      b x    pVp  â !r        â !r 
       Ö/ G	  à     € ê !r         Ê	x   à@  � Ê  x #ÛÙ?  A  È  x ;ª¸¿  A  â !r        Ê s        " 	x    à@  � È 	x    àÀ  � Ê !t   €?    â !t  €?    È s	        $  r	     A  È  r     A  Ê †y    ä Ay        € ê ‰       b Ey  ð    € æ ‰      ¨ „y €  H   è „y €	  H   ( „y €  H   è „y €  H   â !r         Æ/ „y €  H   â !r       ÆO ‰       h ‰      " !r       È� !r       â!r        â/ !r       È ‰ €    ( ‰ €    h „y €  H   ¢ !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       âO !r       â x   ÿàÿ â !r
	
       Æ/ ‰       " z  r  p`ð Æ ‰
      b x    pVp  â !r        â !r 
       Ö/ G	  à     € ê !r         Ê	x   à@  � Ê  x #ÛÙ?  A  È  x ;ª¸¿  A  â !r        Ê s        " 	x    à@  � È 	x    àÀ  � Ê !t   €?    â !t  €?    È s	        $  r	     A  È  r     A  Ê †y    ä Ay        € ê ‰      b !r       â x   ÿàÿ Æ !r       â z  r  p`ð Æ !r       â x    pVp  È ‰       ¢ !r       Ê/ ‰      " !r         ÊO ‰       b !r       Ê ‰ €    " !r        Ê/ ‰ €    b !r	       Ê ‰
	 @    " !r       Ê/ ‰ @    b !r
	
       Ê ‰
      " !r       Ê/ ‰       ¢ !r
       â M	        € ì !r       â/!r         ÈO 	x  à@  � â !r        È  x#ÛÙ?  A  â 	x  à@  � Æ  x;ª¸¿  A  â 	x  àÀ  � Ê s        " !t   €?    â !t  €?    Ì s        $  r     A  È  r      A  Ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      "   $ z  w  p`ð Ú M	        € ê y       !   " $v?ÿ l  ÿ Ž â ¹z  F   
   â Ey  @    € â yA      %   b r> ÿ       â $rÿÿ   ÿ Ž â x?ÿ   ?  ä xÿ      È r ?   p`ð ä xAA   ÿ   Ö/ G	  °    € ê y      &   " $tÿ   ÿ Ž ä $z x  Ž È %v \   Ž Ì �y    ™ ¢  r ÿ       â Ey °    € â x	    ÿ   â$r>ÿÿ   ÿ Ž â x $       â 'z m   Ž â xH    ÿÀŽ Æ $rÿÿ    Ž â zÿ n    â %xHH   ÿ Ž È $rÿÿ   
Ž È $z o  Ž ä $z r  ÿŽ ä $z u  ÿŽ Ê r	   à‘ È r	ÿÿ   ÿ$~  â %v Z   Ž â r ?   p`ò Æ $x		$   ÿŽ â xB   ÿàñ â $zA q  Ž Æ rC	   ÿä  ä xA   ÿàÿ ä x      pðð ä z	  Y      â $z t  Ž äO$z t  Ž ä $z q  Ž Ä $vÿ X  ÿ Ž ä %xF"   HŽ È %xD"   HŽ â G™  À    € ê rD   ÿàó ärF   ÿàñ â %x"   Ž â�yBøÿÿ • ¤  rG	   ÿä  â $xE   	Ž  âr
HB   ÿàñ â %x"   Ž â �y    • ä rIC   ÿä  Ä �y   • ( �y   • ( �y   • h �y   • h �y   • è �y   • è �y   • è �y
üÿÿ ™ è �y   • è. �y
    ™ è �y    • â xBB   ÿàó Ä x  ÿàõ ä x      páð â $rCÿÿ   CŽ  â x   ÿàÿ ä r	ÿ	   ÿä â 0rÿ    A   âO xT      äxT      â0rÿ    A   â� xT      â &r   ÿ   Ä &r   ÿ   â xT      â &r      È &r      ä 0rÿ    A   â  r     A  â Er        ä Er        â  r     A  ä #r      ä #r>   >   Ä Ay       € ê $x?   
Ž âr ?   p@ò â Ey À    € æ x    p4ò Ú G  �    € ê x      páð ä x?èÿÿÿÿàÿ ä r<F   ÿàó är.D   ÿà÷ ä r:HB   ÿàõ â %xL"   Ž â r=G	   ÿäÿ  â �y)Bøÿÿ • ¢  r;IC   ÿä â $x/E   	Ž Ä �y
<   • è �y<   • â %xJ"   Ž Æ �y.   • h	 �y.   • h	 �y<   • h �y<   • h �y<  • h �y<  • h �y:üÿÿ ™ h �y.   • h	 �y.   • h	 �y.  • h	 �y.  • h	 �y<  • h �y<  • h �y&<"  • h �y'<$  • h �y:    ™ h �y:  ™ h �y*L    • h �y+J    • h �y.  • h	 �y.  • h	 �y."  • h	 �y.$  • h	 �y,<&  • h �y-<(  • h �y.&  • h	 �y!.(  • h	 �y .2  • h	 �y#.4  • h	 �y":   ™ h �y$.6  • h	 �y%.8  • h	 �y(:<  ™ h �y0L  • h �y/B  • h	�y.:@  ™ h �y1J  • h �y2<2  • h �y3<4  • h �y4B8  • h	 �y5L   • h �y6<6  • h �y7<8  • h �y8J   • h �y9:\  ™ h �y<L0  • h/ �y=J0  • h �y::`  ™ h �y;BX  • b x    ÿàÿ È r    p`ò ä x@  ÿà÷ ä xBB€  ÿàõ är	ÿ	   ÿäÿ Æ $rCÿÿ   C â x

T      â� 0r)ÿ)    A   âO xT      äxT      ä xT      â &r
   ÿ   ä&r   ÿ   â xT      Ä xT      ä xT      ä x'&T  '    â &r      ä&r   ÿ   ä 0r*ÿ*    A   â Er        â &r      Ä 0rÿ+    A   ä &r   ÿ   â  r**)     A  â xT      â  r))     A  â Er        ä xT      â #r*      â x-,T  -    â #r)   >   â x!T  !    â &r"      âx  T  #    â &r"      Æ Er        â &r
'(   ÿ   ä0r0ÿ0    A   ä 0r/ÿ/    A   ä &r(   ÿ   ä &r
-.   
   â r00/     A  â 0rÿ1    A   â x%$T  %    â &r.!.      â Er        Ä x22T  3    â 0r4ÿ4    A   ä 0r5ÿ5    A   â  r//     A  â Er
 
       â #r0      ä  r54     A  â x76T  7    â 0rÿ8    A   â Er. .       â &r229   ÿ   Ä &r  9   ÿ   â #r/      â  r4     A  â #r

      â 0r<ÿ<    A   ä 0rÿ=    A   ä &r27:   2   ä&r %:       â #r.      Ä Er2 2       â 0r;ÿ;    A   â Er          È  r<;     A  â  r;;     A  Æ #r2   
   â #r>;       â G™  �÷ÿÿÿÿƒ ê Ay       € ê $x?   
Ž âr ?   p@ò â Ey ð    € æ x    p4ò Ú G  À    € ê rF   ÿàñ ä rHB   ÿàó ä rG	   ÿä  ärD   ÿàñ â $xI   CŽ  ä �y%   • ¢  rE	   ÿä  Æ �y&   • ¢  %x
"   Ž Æ �y)   • â %x"   Ž Æ �y*   • è �y'   • ( �y(   • ( �y"  • h �y#  • h �y  • h �y$  • h �yüÿÿ ™ h �y    ™ b $rÿÿ   B Ž â r C       Ä �y  ™ h �y   ™ h �y!   • h �y    • h �y  • h �y  • h/ �yøÿÿ • h �y    • h �y
    • h �y  • h �y  • h �y  • h �y  • h �y
  • b x   ÿàõ Ä xBB@  ÿàó ä x      páð â $r	ÿÿ   	 â x   ÿàÿ ä rCÿC   ÿäÿ  ä x%%T  &    äO x))T  *    Ä� x''T  (    äx#"T  #    äx$T  $    â &r%%   ÿ   ä&r))   ÿ   ä &r%'   %   ä &r##   ÿ   Æ Er% %       ä x!!T       ä xT      Æ &r!!   )   ä 0r
ÿ    A   ä &r   ÿ   ä 0rÿ    A   ä 0rÿ    A   Æ  r
     A  â &r#$   #   â xT      â  r
     A  â/ Er! !       â 0rÿ    A   ä &r      ä 0r
ÿ    A   ä 0rÿ    A   â Er        â #r!   >   â Er# #       â #r%      â  r

     A  â  r     A  Æ #r>
      â #r#      ä Ay       € ê r ?   pp  Ú G‰  `    € ê rD   ÿàõ â%x"   Ž ârF   ÿàñ ä r
HB   ÿàó â $xE   	 ârG	   ÿä  â %x"   Ž â rIC   ÿäÿ  â �yBBøÿÿ • ¨ �y   • è �y	   • è �y   • ( �y   • ( �y   • h �y   • h �y   • h �y   • h �y
üÿÿ ™ h �y    • h �y
    ™ h �y    • b xT  	    Ä� xT      â0r	ÿB    A   âO xT      äxT      â &r   ÿ   ä&r   ÿ   ä 0rÿ    A   ä &r      Ä &r      ä 0rÿ    A   â  r	     A  â Er        ä Er        â  r		     A  ä #r      ä #r>	   >   Ä Ay        € ê ‰      " rAA    ÿàÿ Æ ‰>      b z A p  p`ð È x     pDp  â !r       â !r>       È/ ‰      ( ‰      b !r       â !r       È/ ‰ €    ( ‰	 €    b !r       â !r		       È/ ‰ @    ( ‰
	 @    b !r       â !r
	
       È/ ‰      h  ‰
      ¢  M	        € ê y      &   â x     ÿ   â $z s  AŽ â !r       â/ !r

       â_ r  ÿ   p"ð â$tÿ   ÿ Ž â x     p"ò Ö         ä  
       â $z v  Ž È� %v j   Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y'      "   $ z ' w  p`ð Ú M	        € ê y      &   " x        â ¹z  F   
   â $z' x  Ž È %v \   Ž Ê �y%    ™ b ¹z  l      â Ey  `    € â ™x?    â y"      !   b r  ÿ       â $rÿÿ   ÿ Ž â r  ÿ       â y#      %   ¢ r ÿ       â $rÿÿ   ÿ Ž â r ÿ       Ä xÿ   "  È/ |    p`ð ä x##   ÿ   ÖO G	  €    € ê r ÿ       â $tÿ$   ÿ Ž â x!"   ÿÀŽ â $z% t  Ž âEy     € â 'z  m   Ž â z  X      ä z  Y      â $vÿ `  ÿ Ž â z ÿ n     Ä z  a      ä r ÿ   ÿáÿ â $z ' r  ÿŽ â r  ÿ       Æ $z o  Ž È $z u  ÿŽ â x"   ÿ   È r    à‘ È r ÿÿ   ÿ$~  â %v Z   Ž â x#   ÿàÿ Æ $x $   ÿŽ â x   ÿàñ â $z # q  Ž ä $z q  Ž â r   ÿä  â $z % t   Ž â z ÿ `  pPð â %x"   ÿŽ Æ z ÿ a   Sð â %x "   ÿŽ È %x
!    Ž È %x!    Ž Ê r   ÿàó â%x!    Ž â r
   ÿà÷ ä r	   ÿäÿ  är   ÿäÿ â �y üÿÿ ™ ¢  r   ÿàõ är   ÿàó â �y$    ™ è  �y)   • ( �y&   • ( �y,   • h �y   • b $x    â r   ÿäÿ  Ä �y(   • è  �y+   • è  �y*   • è �y-   • è �y    • è �yøÿÿ • è �y    • â x))T  &    Äx,T      Æ&r))    ÿ   äO&r    ÿ   â x+(T  +    ä� x-*T  -    Æ &r)+$   )   ä0rÿ    A   ä &r-$      ä 0rÿ    A   ä 0rÿ    A   Æ  r     A  â Er )       ä Er        â  r     A  â G‰      € ê r   ÿàó är   ÿàõ ä r	   ÿäÿ  Æ $x    ä �y    ¨ �y&    ¨ �y(    è �y)    è �y     " xT  &    ÊO &r    ÿ   â x)(T  )    Ê� &r))$      ä 0rÿ    A   ÆEr) )       ä  r     A  È #r)      â #r      â #r      â G‰      € ê r
   ÿàó är   ÿàõ ä r   ÿäÿ  är   ÿä Æ �y    ¨ �y    ¨ �y    è �y    è �y     " xT      ÊO &r    ÿ   â xT      Ê� &r$      ä 0rÿ    A   ÆEr        ä  r     A  È #r        â x   ÿàÿ ä x  ÿà÷ ä |    p`ò ä x  ÿàù â $rÿÿ   Ž â x   ÿàõ ä rÿ   ÿä Ä rÿ   ÿä Ê G™  Àúÿÿÿÿƒ ê Ay       € ê r         ä r         ä Ay        € ê z ÿ `  pPð È z ÿ a   Sð Ú G	  P    € ê ‰      ( ‰      b !r       â !r	       È/ ‰      ( ‰	      b !r       â !r
	       È/ ‰ €    ( ‰
 €    b !r       â !r
       È/ ‰ @    ( ‰ @    b !r       â !r       È/ ‰      ( ‰	      b !r       â !r	       â/ Gy  €    € ê ‰      ( ‰      h ‰	      ¨ ‰       â !r       â !r       È/ ‰       " !r	       ÆO ‰      b !r        Æ� ‰	      ¨ ‰      â !r         â !r       È/ ‰  €    " !r
	       ÆO ‰	 €    b !r       Æ� ‰
 €    ¨ ‰ €    â !r        â !r		       È/ ‰ @    " !r
       ÆO ‰	 @    b !r       Æ� ‰ @    ¨ ‰ @    â !r       â !r	       â/ !r
       ÆO ‰      " !r       Æ� ‰      h ‰
      ¨ ‰      â !r       â !r        â/ !r
       âO !r       â� r#"   ÿàÿ È z  p  p`ð È x "   pDp  Ú M	        € ê z ÿ ^  pPð â $z%% v  Ž Æz ÿ _   Sð Ú $ÿ   ÿ Ž È %% ^   Ž Ì �	     ¢ x"   ÿ   È r ÿ   p"ò äx    p"ö äx    p"ø äx    p"ô Î         ä z ÿ `  pPò ä 2        ä z ÿ a  Sò ä B         â $"ÿÿ    Ž È !       ÌO G™  @    € ê z ÿ b  pPð È z ÿ c   Sð Ú         Ê %% b   Ž Ì �		     ¢ x""   ÿ   â $v
ÿ h  ÿ Ž Æ x "üÿÿÿp"ô är "ÿ   p"ö äx "   p"ø äx "øÿÿÿp"ò ä x 
   pBò Ê "        ä 2         ä B        È r         Ê !  	       âO G  p    € ê x
ÿÿÿÿÿàÿ È x     € È x   ÿ   È ‚{  €     $ xÿ     È Iy ðïÿÿÿÿƒ ê  x ;ª¸¿  A  Ì s        $ !t  €?    Ì s        $  r       A  È  r      A  â Gy  P    € ê  x '7=  A  â x *BL?  A  Æ #t  €?    â  x     ?  A  Æ  r     A  Ì s     $   $ !t  €?    È  r       A  È  r      A  â Gy  °    € ê x 
   pRð Ú G‰  à     € ê x 
   pRð Ú G	        € ê 	z   i    � ä 	z i    � Æ  x ;ª¸¿  A  â 	z i €  � Ê s        $ !t  €?    Ì s        $  r       A  È  r      A  â Gy  Ð     € ê  r      A  â Gy  °     € ê 	x    à@  � ä 	x  à@  � Æ  x #ÛÙ?  A  â 	x  àÀ  � Æ  x;ª¸¿  A  ä !t  €?    È s        $ !t  €?    Ì s        $  r       A  È  r      A  ä $z'' s  Ž â x        Æ $z v  'Ž È %v j   Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y$      "   " ¹z  l      â Ey      € â ™x?    â y#      !   b ¹z  F   
   â r" ÿ       â x       ÿ  â y      %   ¢ x      ÿ  â x      ÿ  â x      ÿ  â x      ÿ  â x      ÿ  â x      ÿ  â r ÿ       Ä r $ÿ   pRð är $ÿ   pRò är$#   ÿ(Ž È/ r   ÿ� ä x   ÿ   äO x ÿ     È |     pbô Ú G)  °    € ê y      '   " r ÿ       ä r ÿ       â y      &   b x
        ä r" ÿ       ä z2  q  ÿàÿ â 'z y   Ž È 'z s   Ž â/ zÿ z    Æ $z p   Ž â zÿ t    â $z	 |   Ž ä$z |  Ž â x   ÿàÿ â $z	 v  	Ž ä$z v  Ž ä $z p  	Ž â x	þÿÿÿÿÀŽ â $z% }  ÿŽ Ä $z w  ÿŽ â r		  €ÿàÿ â %v X  
Ž âz  q      ä x'	   ÿ   â %v X  
Ž â r%%   ÿàõ ä x   ÿàù â $x(    Ž âx   ÿà÷ â $x,    Ž âx)	   ÿ   â $x.    Ž âxÿ   '  â $x0    Ž â rÿ   ÿä Ä rÿ   ÿäÿ ä r&    ÿŽ är*    ÿŽ ä r\ÿÿ   ÿä ä x+)   ÿàÿ ä x-'   ÿ   Ä x/'     ä �yLüÿÿ ™ ¢ %x'   Ž Æ �yJüÿÿ ™ è �yK    ™ ( �yI   ™ b %x'   Ž â r
 %   ÿàõ Ä xW $       â �y4øÿÿ ™ â rÿ\   ÿä Æ �yF    ™ â %v

 Z  W Ž Æ �yE   ™ â $x$   ÿŽ â r-
   ÿàõ ä �y6øÿÿ ™ ä r   ÿàÿ È r/   ÿä ä r2%   ÿàõ Æ �y?   ™ â  r 
       ä rÿ\   ÿä â/ �y@   ™ â  %v Z  W Ž Æ �y1    ™ â $x$   ÿŽ Æ �y3   ™ â  r&%   ÿà÷ Æ �y5   ™ â  r
-   ÿàõ ä r   ÿàÿ ä rÿ\   ÿäÿ ä/ r/   ÿä â %v Z  W Ž È $x$   ÿŽ â �y7
   ™ â r        ä r-   ÿàõ â �y8
   ™ â r	   ÿàÿ Æ �y9    ™ â r/   ÿä ä r(%   ÿàõ â �y:
   ™ æ r	ÿ\   ÿä â �y;
   ™ â %v Z  W Ž Æ �y<   ™ â  $x		$   ÿŽ Æ �y=   ™ â  r        ä r
-   ÿàõ â/ �y>   ™ â  r		   ÿàÿ Æ �yB   ™ â  r	/   ÿä ä r*%   ÿàõ â �yA    ™ æ rÿ\   ÿä â �yC
   ™ â  r        Æ �yD
   ™ â  $x$   ÿŽ ä %v Z  W Ž â/ �yH
   ™ â r,%   ÿà÷ Æ r-   ÿàõ â �yG    ™ â  r   ÿàÿ Æ �yO
   ™ â r/   ÿä ä rÿ\   ÿäÿ Æ �yM   ™ â %v Z  W Ž â r        ä �yN   ™ â  $x$   ÿŽ â r-   ÿàõ ä �yQ    ™ ä  r	   ÿàÿ Ä �yP   ™ ä  r/   ÿä ä r
.%   ÿàõ â/ �yT   ™ æ rÿ\   ÿä â �yR   ™ â r        â %v
 Z  W Ž â r0%   ÿà÷ â �yS   ™ ä  $x$   ÿŽ Ä �y    ™ â  r-   ÿàõ ä/ r		   ÿàÿ â �yU   ™ â r
ÿ\   ÿäÿ ä r	/   ÿä â %v Z  W Ž â �yV   ™ æ $x
$   ÿŽ â �yW   ™ â r
-   ÿàõ Æ r   ÿàÿ â �y   ™ æ r/   ÿä â �y	    ™ è  �y   ™ è  �y
   ™ è �yY   ™ è  �yZ
   ™ è �y    ™ è  �yX
   ™ è �y[
   ™ â r]ÿ)   L  ÄO rLÿ+   L  ä x^]   ÿ   ä xL   ÿ   ä x^^   ÿÀŽ ä xÿ   ]  ä x`ÿ   L  Ä x^^KøŽ äx   ÿÀŽ ä xKÿ   K  ä x   ÿÀŽ ä x_ÿ   I  äx``   ÿÀŽ Ä x]   ÿ   ä xIIøŽ ä xKKøŽ ä xÿ   ]  ä x_`_øŽ ä xÿ   L  Ä x^   øŽ ä x]   ÿ   ä x
L   ÿ   ä/xK   øŽ ä x_   øŽ ä xKL   ÿ   Ä x   øŽ ä x
I   
øŽ ä x]   ÿ   äx]   ÿ   ä xIL   ÿ   ä x   KøŽ Ä rKÿ)   J  ä� x   øŽ ä x   øŽ ä x

   IøŽ ä x]	   ÿ   ä xL   ÿ   Ä xIL	   ÿ   ä xLK   ÿ   ä x   øŽ ä x   IøŽ ä x
   øŽ ä xIL   ÿÀŽ Ä x
ÿ   K  ä rJÿ+   J  ä xÿ   F  äxIIFøŽ ä x]
   ÿÀŽ ä xFJ   ÿ   Ä x
K   ÿ   ä xFF   ÿÀŽ ä x
I   
øŽ ä x]øŽ ä xIÿ   J  ä x]ÿ   K  Ä xFFEøŽ äxLÿ   E  ä xEK   ÿ   ä x_I   ÿÀŽ ä x   ]øŽ ä xIK   ÿ   Ä x]J   ÿ   ä x

   EøŽ ä xEK   ÿ   äx   IøŽ ä xKK	   ÿ   ä xFF   ]øŽ Ä xL_LøŽ ä x]ÿ   J  ä xIJ   ÿ   ä x

   EøŽ ä xE   KøŽ â &r?   ÿ   â x]L   ]øŽ Ä xIF   IøŽ â &rF
?   ÿ   â xLJ   ÿ   â &r@      Æ x?I   LøŽ â &rLE@   F   â 2rF41        â&r@3      ä&rL?3   L   ä &r@5   @   ä 0r3ÿF  0 A   âx^J   ÿ   â 2rI61        â 0rFÿF    A   â Er1 @       â  x33   ?  A  â x]]   ^øŽ Ä xJJ	   ÿ   â #r1F1   3   â &r37   ÿ   ä x]   JøŽ ä &r38   3   â !r1       ä &rL5   L   â 2r149        â 0r5ÿI  0 A   Ä &r3:   3   ä &r7
7   ÿ   â  x@5   ?  A  â &r3;   3   ä &r7E8   7   ä 0r5ÿ1  0 A   ä&rF<   ÿ   â Er8 3       â &r7?:   7   â  x:5   ?  A  â 0r1ÿ1    A   Ä &rF=   F   ä&rJ
<   ÿ   â #r818   :   â &r1>   F   ä &rJE=   J   ä &r1B   1   â 2r34A        â &r5?>   J   â !r8       Ä Er8 1       â &rBB   5   ä &r1C   ÿ   ä 0r5ÿ3  0 A   ä&r1D   1   ä 0r3ÿ3    A   â  x:5   ?  A  â &r1H   1   Æ #r838   :   â 2r34G        â &r1O   1   ä !r8       ä 0r5ÿ3  0 A   âEr8 1       â &r1M   ÿ   ä 0r3ÿ3    A   â  x:5   ?  A  â &r1N   1   Æ #r838   :   â 2r34Q        â &r1P   1   È &r1T   1   ä 0r5ÿ3  0 A   â!r8       â 0r3ÿ3    A   â Er8 1       ä  x:5   ?  A  È #r838   :   â 2r34        ä2r56        ä2r969        â 0rÿ3  0 A   â ErL L       â 0rIÿI    A   ä &r7;   7   â  x:   ?  A  â 0r;ÿ9  0 A   Ä &rC
C   ÿ   ä &rW   ÿ   â2rA6A        â &rW
W   ÿ   â #rIIL   @   â 0r9ÿ9    A   â Er< 7       â &rCED   C   â  x@;   ?  A  â &r      Ä &rM
M   ÿ   â 2rG6G        â 0r7ÿA  0 A   ä &rWE   W   â 2r4	        â &rC?H   C   â #r99<   @   â &r      â ErB B       â &rMEN   M   Ä &r@R   ÿ   â x<7   ?  A  â 0rAÿA    A   ä &rR
R   ÿ   ä &rW?   W   â x   ÿàõ â &rCO   C   ä &r   ÿ   Ä &rY      ä &r

   ÿ   â 2rQ6Q        â0r7ÿG  0 A   ä0rÿ  0 A   ä&rM?P   M   â #rAAB   <   â &r@S   @   â rÿ   ÿä â &rRES   R   â Er< C       â &rZ      âEr        â &r
EZ   
   â x   ÿàõ â 0rGÿG    A   â  x>7   ?  A  â 0rÿ    A   â  x   ?  A  â &rMT   M   â 2r	6	        â 0r7ÿQ  0 A   â 2r44        â&r1U   @   âx      ÿàÿ â &rU?U   R   â 2r6        â &rX      Ä &r
?X   
   â rÿ   ÿä â #rGG<   >   â #r      â 0rQÿQ    A   â Er< M       â 0rÿ	  0 A   â |     pbô â 0rÿ4  0 A   â  x>7   ?  A  â &r1V   1   Ä &rUV   U   ä&rWY   W   ä &r[      ä&r
[   
   ä 0r7ÿ5  0 A   ä 0rÿ  0 A   â #rQQ<   >   â !r8       â  x   ?  A  â  x   ?  A  â 0r3ÿ3    A   â Er8 1       â 0r5ÿ5    A   â Er< U       â 0r	ÿ	    A   â Er W       â 0r4ÿ4    A   â Er        â 0rÿ    A   â Er
 
       â  x>7   ?  A  â  x   ?  A  â #r838   :   â #r		      â #r55<   >   â #r44      â #r
      â !rI       â !r9       â !rA       â !rG       â !rQ       â !r8       â !r5       â !r       â !r 	        â !r!4!       â !r""       â x&&    ÿàÿ Ä x((    ÿàÿ ä x**    ÿàÿ ä x,,    ÿàÿ ä x..    ÿàÿ ä x00    ÿàÿ ä x22    ÿàÿ â D)       À â Gy   èÿÿÿÿƒ ê Ay        € ê Ey  0    € â G‰      € ê r$#   ÿHŽ Ê ˆs  øÿ H   è ˆs €øÿ H   è ˆs  ùÿ H   è ˆs €ùÿ H   è ˆs  úÿ H   è ˆs €úÿ H   è ˆs  ûÿ H   è ˆs €ûÿ H   è ˆs  üÿ H   è ˆs €üÿ H   è ˆs  ýÿ H   è ˆs €ýÿ H   è ˆs  þÿ H   è ˆs  €þÿ H   è ˆs ! ÿÿ H   è ˆs "€ÿÿ H   ä Ay        € ê {          ì M        € ê „y#     H   " z  r  p`ð Ä z1  r      â „y# €   H   b r #ÿ   p"p Æ „y#    H   ¢ $x/1   ÿŽ ä$x71   ÿŽ â „y	# €  H   è „y#    H   ( „y#    H   h „y# €  H   h y      '   è „y+# €  H   è „y%# €  H   â !r       Æ „y&#    H   â !r       Æ/ ‰       " !r       ÆO „y)# €  H   â !r	       Æ� „y(#    H   â !r       Æ‰      b !r       Æ„y*#    H   â !r       Æ ‰      ¢ !r        Æ „y-# €  H   â !r++       Æ ‰      " !r%%       Æ y       &   â !r       Æ/ ‰      b !r'&       Æ ‰      " !r       ÆO „y,#    H   â !r))       â !r(       ä ‰      â !r       Æ „y5# €  H   â !r*       â $z   x  Ž ä� ‰	 €    " x   ÿàÿ â $z  ~   Ž â !r
       â/ ‰&%      â !r       ä /   ÿàó â ‰
      b r1    ÿ† â !r--        â ÿÿ   ÿäÿ  â ‰      ¢ 
 j  ÿˆ Ä 
  r  ÿàõ â ‰      â 
 k   ä rÿÿ   ÿäÿ â ‰      â z j  ÿ† â !r	       â $x	1   ÿŽ â $ÿÿ   ÿä â !r!,!       â z  r  p`ò â ‰ @    " 
 j  ÿ„ â !r       â z k  � â !r
       â/ r		    ÿà÷ â‰('      b 
 k  $ â !r$       âO r1    ÿ„ â ‰      ¢ rÿÿ   ÿäÿ â !r&%&       â z	 j  ÿ† â ‰*)      " rÿÿ   ÿä â !r       â� z
 j  ÿ„ â ‰,      â z		 k  � â $x1   ÿŽ â x #   p"ò â ‰.+      b z k   â !r       â !r       â ‰0      "  7   ÿàù â !r55"       Ä ‰2-      " ÿÿ   ÿä â !r('(       â/ 
 j  ÿˆ â ‰4!      b 1   ÿàÿ â !r       âO 
 k   â ‰ €    ¢     ÿà÷ â!r*)*       â8    ÿàù â ‰$      " !r,,       â� ÿÿ   ÿäÿ Ä ‰       â !r.+.       â
 j  ÿ† ä ‰      b 9ÿÿ   ÿä â !r00       â 8 j  ÿˆ â !r2-2       â ‰%&      " 
 k  � â !r4!4       Ä/ ‰'(      b 8 k  9 â !r       äO ‰)*      ¢ /   ÿàÿ â !r$$       ä‰+,      "     ÿàõ â !r         Ä� ‰ €    â 6ÿÿ   ÿä â !r9       ä‰-.      b  j  ÿ„ ä 7   ÿàÿ â ‰ €    b !r&&%       â  k  6 Ä ‰/0      " !r(('       â/     ÿàõ ä ‰12      b !r**)       âO "ÿÿ   ÿä ä ‰34      ¢ !r,,+       Æ‰5      " !r       Æ� ‰ @    â !r..-       Æ‰!  €    b !r       Æ ‰ €    b !r00/       Æ ‰#$ €    " !r221       â/ !r443       äO ‰'& €    b !r65       Æ‰)( €    ¢  j  ÿ„ â!r       ä� ‰+* €    â  k  " â !r  !       ä‰ @    " x        â !r       Ä ‰-, €    d %v  j   Ž â !r$$#       â ‰/. €    ( ‰ @    " !r'&'       Æ/ ‰10 €    b !r)()       ÆO ‰32 €    ¢ !r+*+       Æ� ‰54 €    â !r       Æ‰76      " !r-,-       Æ‰      b !r/./       Æ ‰#  @    " !r       Æ ‰! @    " !r101       Æ/ ‰%$ @    b !r323       ÆO ‰"' @    ¢ !r545       â� !r767       ä‰&) @    â !r;       Æ‰(+ @    " !r# #       Æ ‰      â !r!!       Æ ‰*- @    " !r%$%       Æ/ ‰ / @    b !r"'"       ÆO ‰      ¨ ‰$1 @    b !r&)&       Æ� ‰,3 @    â !r(+(       Æ‰07 €    ( ‰.5 @    " !r*-*       Æ ‰!      " !r /        Æ/ ‰#      b !r)       ÆO †	 9    â !r$1$       â!r,3,       ä� † ;   â !r070       Æ‰"      â !r9       âO !r.5.       ä ‰%      ¢ !r+!       Æ ‰&      " !r-#       Æ/ ‰(      h †	 9    è ‰*      " !r"       Æ� ‰       è † )    â !r%%       ÆO ‰!$      ¢ !r&       Æ ‰#,      " !r(       Æ/ ‰'0 @    h ‰.      b !r*       Æ†	 +    â	 !r        Æ� † -   è	 †	 %    â	 !r!$!       âO !r#,#       ä †     â	 !r'0'       Æ/ †	 
    â	 !r.       Æ† 
   è	 †	     è	 †     è	 †	 !    è	 † #   è	 ‰'      ( †	     â	 M™        € ê x   ÿàÿ â !r''       Ær    ÿàñ È r ÿÿ   ÿä  ä z j  ÿ€ È z k     Ê †y '    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      "   " ¹z  l      â Ey  °    € â ™x?    â y      !   b ¹z  F   
   â r ÿ       â x0      ÿ  â y      %   ¢ r ÿ       â x      ÿ  â r ÿ       â x      ÿ  â r/ ÿ       â x      ÿ  â r6 ÿ       Ä r  ÿ       ä r ÿ   pRð är ÿ   pRò är   ÿ(Ž È/ r   ÿ� ä x   ÿ   äO x.ÿ     È | .   pbô Ú G)  @    € ê y      '   " r ÿ       â $rÿÿ   ÿ Ž â r ÿ       â y      &   b x
        ä z,. q  ÿàÿ â 'z y   Ž È 'z s   Ž â/ zÿ z    Æ $z p  .Ž âzÿ t    â $z	 |  .Ž ä$z |  Ž â x   ÿàÿ â $z	 v  	Ž ä$z v  Ž ä $z p  	Ž â x	þÿÿÿÿÀŽ â $z# }  ÿŽ Ä %v X  
Ž â r		  €ÿàÿ Æ $z w  ÿŽ â x%	   ÿ   â %v X  
Ž â z  q      ä r##   ÿàõ ä x   ÿàù â $x$   .Ž âx   ÿà÷ â $x(   .Ž âx'	   ÿ   â $x*   .Ž â xÿ   %  Ä rÿ   ÿä ä rÿ   ÿäÿ ä r.   ÿŽ är&.   ÿŽ ä rCÿÿ   ÿä ä x+'   ÿàÿ Ä x-%   ÿ   äx)%     ä �y4üÿÿ ™ ¢ %x%   Ž Æ �y:üÿÿ ™ è �y8    ™ ( �y3   ™ b %x%   Ž â r
.#   ÿàõ Ä x? $       â �yøÿÿ ™ è �y!    ™ â rÿC   ÿä â %v

 Z  ? Ž ä �y2   ™ ä $x$   ÿŽ â r-
   ÿàõ â �y=øÿÿ ™ æ r   ÿàÿ È r)   ÿä ä r,#   ÿàõ ä r 
       â �y5   ™ â  r
ÿC   ÿä Æ �y7   ™ â  %v Z  ? Ž Æ/ �y     ™ â $x
$   ÿŽ Æ �y"   ™ â  $x		   Ž â r
-   ÿàõ ä �y>   ™ ä  r	)   ÿä ä r#   ÿàõ â/ �y@    ™ æ rÿC   ÿä â �y9
   ™ â %v Z  ? Ž Æ �y;
   ™ â $x$   ÿŽ â r-   ÿàõ ä �y<
   ™ ä  r   ÿàÿ ä �yA
   ™ ä  r)   ÿä Ä r$#   ÿàõ Æ �yB   ™ â  r        ä rÿC   ÿä â �yD   ™ â  %v Z  ? Ž Æ �yF    ™ â  $x$   ÿŽ Æ �yE   ™ â  r-   ÿàõ Æ �yI   ™ â  r   ÿàÿ ä r&#   ÿà÷ ä/ r)   ÿä â �yJ    ™ â r
ÿC   ÿäÿ Æ �yG   ™ â  %v Z  ? Ž Æ �yH   ™ â  $x
$   ÿŽ â r-   ÿàõ ä �yK   ™ ä  r	   ÿàÿ ä �yN   ™ ä  r)   ÿä Ä r
(#   ÿàõ Æ �yL   ™ â  rÿC   ÿä ä r        â �yM   ™ â  %v
 Z  ? Ž â r*#   ÿà÷ ä/ �yO   ™ â $x$   ÿŽ â r-   ÿàõ Ä �y    ™ â  $x		   Ž â r
ÿC   ÿäÿ ä �yP   ™ ä r	)   ÿä â %v Z  ? Ž â �y    ™ æ $x
$   ÿŽ â �y?   ™ â r
-   ÿàõ Æ r   ÿàÿ â �y   ™ æ r)   ÿä â �y   ™ è  �y
   ™ è �yR   ™ è  �yS
   ™ è �y    ™ è  �yQ
   ™ è �yT
   ™ â rUÿ'   4  ÄO r4ÿ+   4  ä xÿ   U  ä xW4   ÿ   ä x	U   ÿ   ä x   ÿÀŽ ä xW   ÿÀŽ Ä xYÿ   4  ä x		   ÿÀŽ ä xVÿ   8  äxXÿ   3  äx33øŽ ä xYY   ÿÀŽ Ä x		8øŽ ä xU   ÿ   ä xVøŽ ä x
ÿ   U  ä/ xXYXøŽ ä x84   ÿ   Ä xÿ   4  ä x	   øŽ ä x
   
øŽ ä x	U   ÿ   äxU   ÿ   ä x33   8øŽ Ä xX   øŽ ä xV4   ÿ   äx84   ÿ   ä x

   	øŽ ä x	U   ÿ   â$xUU   ÿ Ž â x   øŽ Ä x   VøŽ ä x33   8øŽ ä rÿ'   :  ä� xV4   ÿ   ä x8
   UøŽ ä x   ÿ   Ä x
3   VøŽ ä x3ÿ     ä x   	øŽ ä r	ÿ+   :  ä x:   ÿÀŽ ä xÿ   !  Ä xV3   ÿÀŽ ä x3	   ÿ   ä xVVøŽ ä x::!øŽ ä x   ÿ   ä x!ÿ   	  Ä x33   ÿÀŽ ä x:   øŽ ä xX!   ÿÀŽ ä x332øŽ ä x:	   ÿ   ä x!ÿ     Ä x33   :øŽ ä xUÿ   2  ä x!V   !øŽ ä x:   ÿ   äx2   ÿ   ä x!!   :øŽ Ä x   2øŽ ä x:   ÿ   ä xUXUøŽ ä xVÿ   	  ä x2	   ÿ   ä x:   :øŽ â &r5   ÿ   â xUU   VøŽ Ä xV	   ÿ   â &r87      â x33   2øŽ ä x44	   ÿ   ä x2	   ÿ   â $xX	   ÿ Ž â xU   VøŽ â &r
"      â xV	   ÿ   â 2r	         â x   4øŽ â &r5:5   ÿ   â x2!   2øŽ Ä xV3   VøŽ â &r>      ä &r527   5   ä 0r!ÿ	  0 A   â2r3=         â &r5V"   5   â Er         â 0r	ÿ	    A   â  x"!   ?  A  â x   XøŽ â &r9   ÿ   Ä #r 	    "   â 0r	ÿ3  0 A   ä&r5>   5   ä &r8;      â  x4	   ?  A  â !r        â 2r	@        â &r 
<      â Er" 5       â 0r3ÿ3    A   Ä &r9:9   ÿ   ä &r A       â #r33"   4   â &r92;   9   ä 0rÿ	  0 A   â2r@=@        â &r<V<   9   â Er          â 0r	ÿ	    A   â !r3       â  x"   ?  A  â &r3B   ÿ   Ä &r<A   <   â #r		    "   â 0r!ÿ@  0 A   ä&r38D   3   ä &r5:B   ÿ   â Er <       â 0r@ÿ@    A   â !r 	        â  x!!   ?  A  â 2r	F        â &r 
E   3   Ä &r52D   5   â #r@@   !   â &r I       ä &r"VE   5   ä 0rÿ	  0 A   âEr          â &rII   "   ä 0r	ÿ	    A   â  x"   ?  A  â 2rF=F        Æ #r 	    "   â &r	G   ÿ   â2rJ        â 0r!ÿF  0 A   ä&r	8H   	   ä&rG:G   ÿ   ä &r	
K   	   â ErI I       â 0rFÿF    A   â !r        â  x!!   ?  A  â &rG2H   G   Ä 0r ÿ  0 A   ä &r	N   	   â #r!FI   !   â &rGVK   G   â 2rJ=J        â  x"    ?  A  â Er  	       â &r	L   ÿ   â !r!       â &rGN   G   Ä 0rÿ    A   ä 0r!ÿJ  0 A   ä&r	8M   	   â ErG G       â 0rJÿJ    A   â #r    "   â  x!!   ?  A  â &r	
O   	   â 2r =        â!r66       â #r!JG   !   â 2r        â &r	P   	   Ä !r!       â 0r"ÿ   0 A   ä0rÿ  0 A   ä0r!ÿ     A   â Er  	       â &r	?   ÿ   â x   ?  A  â 0rÿ    A   ä &r?:?   ÿ   Ä &r	8   	   â#r       â &r?2   ?   â 2r        â &r	
   	   ä&r?V   ?   ä &r   ÿ   ä &rL:L   ÿ   Ä &r	R   	   ä &r:   ÿ   ä 0rÿ  0 A   ä&rL2M   L   ä &r88S      â Er 	       â &r2S      â x   ÿàõ â 0rÿ    A   â  x   ?  A  â 2r=        â x..    ÿàÿ â 2r        â&rLVO   L   â 2r=        â &r8
Q   8   Ä &rVQ      â rÿ   ÿä â #r      â 0rÿ  0 A   â | .   pbô â &rLP   L   ä&r?R   ?   ä &r8T   8   ä&rT      Ä 0r	ÿ  0 A   ä0r
ÿ  0 A   â  x   ?  A  â ErL L       â 0rÿ    A   â Er? ?       â 0rÿ    A   â  x	   ?  A  â 0rÿ    A   â Er8 8       â  x""   ?  A  â Er        â  x	
   ?  A  â x   ÿà÷ â #r"!L   "   â #r?      â #r8      â #r   	   â rÿ   ÿäÿ â !r@       â !r//       â !r1"1       â !r00       â !r       â !r       â !r       â x    ÿàÿ Ä x$$    ÿàÿ ä x&&    ÿàÿ ä x((    ÿàÿ ä x**    ÿàÿ ä x,,    ÿàÿ â D)       À â Gy  `êÿÿÿÿƒ ê Ay        € ê Ey      € â G‰  ð     € ê $xÀ  Ž Ê ˆs  ùÿ H   è ˆs €ùÿ H   è ˆs   úÿ H   è ˆs €úÿ H   è ˆs  ûÿ H   è ˆs €ûÿ H   è ˆs 6 üÿ H   è ˆs €üÿ H   è ˆs / ýÿ H   è ˆs 1€ýÿ H   è ˆs 0 þÿ H   è ˆs €þÿ H   è ˆs  ÿÿ H   è ˆs €ÿÿ H   ä Ay        € ê {          ì M        € ê „y     H   b z  r  p`ð Ä z)  r      â „y €   H   ¢ r ÿ   p"p Æ „y	 €  H   â $x')   ÿŽ Æ „y    H   ( „y    H   h „y    H   ( „y €  H   ( „y €  H   ( „y    H   è „y €  H   â !r       Æ/ „y!    H   â !r       ÆO ‰      b !r		       Æ� „y €  H   ¢ !r        Æ„y    H   â !r6       Æ„y €  H   â !r       Æ y      &   â !r       Æ ‰      " !r11       â !r       ä/ y      '   " !r00       Æ ‰      b !r        Æ ‰      b !r"!       Æ ‰      b $z  x  Ž Æ� ‰      â x   ÿàÿ â !r       äO ‰	      ¢ !r       âz  r  p`ò â !r//       â ‰
      " $z  ~   Ž â x    p"ò â !r$       â !r       â/ ‰      " 
  r  ÿàõ â !r       Ä‰/      b !r       Æ ‰1      b !r       Æ� ‰ €    â ÿÿ   ÿä â !r	       äO ‰0      ¢ 
 j  ÿ„ â !r

       ä‰      " 
 k   Æ ‰!       " !r       â $x)   ÿŽ ä ‰#"      " !r//       Æ/ ‰	      b !r11       Æ‰      b !r       â�     ÿà÷ ä ‰%$      â !r0       âO    ÿàÿ ä ‰      ¢ !r       âr)    ÿ„ Ä ‰      " !r! !       Æ ‰
      " !r#"#       â !r	       ä/ ‰       " 	 '   ÿàû â !r       ä‰"/      b ÿÿ   ÿäÿ â !r%$%       Ä� ‰&      â 
	 j  ÿŠ â !r       äO ‰$1      ¢ ÿÿ   ÿäÿ â !r       ä‰ @    " rÿÿ   ÿä â !r
       Ä ‰ €    b r)    ÿˆ ä 
 j  ÿ† â ‰(!      b z j  ÿ„ â !r         â r.ÿÿ   ÿä â ‰*#      " 
 k  � â !r"/"       â/ z
 j  ÿˆ â ‰ €    b z k   â !r&&       â� 
		 k  � â ‰ €    â !r$1$       âO z k  . Ä ‰,%      ¢ !r       â)   ÿàÿ ä ‰ €    " !r       â'   ÿàÿ ä ‰ €    b !r(!(       â /    ÿàù Ä ‰ €    b !r*#*       â 2ÿÿ   ÿä â !r       â/ ‰#  €    " $x)   ÿŽ â / j  ÿˆ â !r       â� ‰'$ €    b 0    ÿà÷ â !r,%,       ÄO ‰)& €    ¢ / k  2 â !r       ä‰%" €    â r    ÿàõ â !r       ä‰ @    " r.ÿÿ   ÿä â !r       Ä ‰+( €    b z j  ÿ„ ä1ÿÿ   ÿäÿ â ‰-* €    b !r# #       â 0 j  ÿ† ä ‰! @    " !r'$'       â/ z k  . Ä ‰ @    b !r)&)       âO 0 k  1� ä ‰ @    ¢ !r%"%       â�     ÿàõ ä ‰/, €    â !r       Æ‰ @    " !r+(+       Æ‰      b !r-*-       Æ ‰ @    b !r!!       Æ ‰ # @    " !r       Æ/ ‰% @    b !r       ÆO ‰"' @    ¢ !r/,/       â� ,ÿÿ   ÿä ä ‰$) @    â !r       Æ‰      " !r1       âx        ä ‰&+ @    b !r       â  j  ÿ„ â %v  j   Ž â ‰(- @    b !r #        â  k  , Ä ‰      " !r%       Æ/ ‰*/ @    b !r"'"       ÆO ‰      ¢ !r$)$       Æ� ‰       â !r3       Æ‰      " !r&+&       Æ‰!      b !r(-(       Æ ‰       b !r)       Æ ‰#      " !r*/*       Æ/ ‰%"      b !r+       ÆO ‰$      ¢ !r        Æ� †	 1    â !r       Æ† 3   â !r!!       Æ‰'&      â !r        Æ ‰(      " !r##       Æ ‰*      " !r%"%       Æ/ †	 )    â !r$       ÆO † +    è †	 
    è † 
   â !r'&'       Æ� †	 !    â !r(       Æ†     â !r*       Æ †	 #    è † %   è †	     è † '    è †	     â M™        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y       "   " ¹z  l      â Ey  P    € â ™x?    â y      !   b ¹z  F   
   â x      ÿ  â r' ÿ       â y      %   ¢ r ÿ       â x      ÿ  â r. ÿ       â $r-ÿÿ   ÿ Ž â r* ÿ       â x      ÿ  â r ÿ       Ä r  ÿ   pRð är  ÿ   pRò är    ÿ(Ž È/ r   ÿ� ä x   ÿ   äO x)ÿ     È | )   pbô Ú G)  ð    € ê y      '   " r ÿ       â $rÿÿ   ÿ Ž â x
        â y      &   b $z	 p  )Ž âz"  q      ä r ÿ       ä r")   ÿŽ â$x"   )Ž âr ")   ÿŽ â$x""   )Ž â z$) q  ÿàÿ â 'z y   Ž È 'z s   Ž â/ zÿ z    Æ $z }  ÿŽ â zÿ t    â $z |  )Ž ä$z |  	Ž â x	   ÿàÿ â $z v  Ž ä$z v  Ž ä $z	 p  Ž â x	þÿÿÿÿÀŽ â $z w  ÿŽ Ä %v X  
Ž âr		  €ÿàÿ ä r   ÿàõ â %v X  
Ž â x!	   ÿ   äx   ÿà÷ â $rGÿÿ   ÿ â x   ÿàù ä x&	   ÿ   â $rÿÿ   Ž â xÿ   !  Ä rÿ   ÿä ä x(&   ÿàÿ ä x%!   ÿ   äx#!     ä �y1üÿÿ ™ ¢ %x!   Ž Æ �yüÿÿ ™ è �y    ™ ( �y/   ™ b %x
!   Ž â r)   ÿàõ Ä xD $       â �y+øÿÿ ™ â rÿG   ÿä Æ �y,
    ™ â %v Z  D Ž Æ �y2
   ™ â $x	$   ÿŽ â r%   ÿàõ ä �y7øÿÿ ™ ä r	   ÿàÿ È r	#   ÿä â �y    ™ â r$   ÿàõ Æ �y3   ™ ä $rÿÿ   G ä �y0   ™ â  %v Z  D Ž Æ �y4   ™ â  $x$   ÿŽ â r%   ÿàõ ä �y5   ™ ä  r   ÿàÿ È r#   ÿä â �y:    ™ â  r   ÿàõ Æ/ �y6   ™ â r
ÿG   ÿä â %v Z  D Ž ä �y8   ™ ä $x
$   ÿŽ â r
%   ÿàõ â �y9   ™ æ r	   ÿàÿ â �y<   ™ æ r#   ÿä Ä r	   ÿàõ ä r        â �y;
   ™ ä  $r?ÿÿ   G ä %v	 Z  D Ž â �y=
   ™ æ  $x?$   ÿŽ â �y>    ™ è  r	   ÿàÿ â �y?
   ™ â r	    ÿà÷ Æ �yB
   ™ â rÿG   ÿäÿ â %v	 Z  D Ž â r%   ÿàõ Æ/ $x	$   ÿŽ â r#   ÿä är        â $x   	Ž â r%   ÿàõ â �y@   ™ æ  r	#   ÿä â �yA   ™ â  r"   ÿàõ Æ �y    ™ â rHÿG   ÿä Æ �yC   ™ â  %v
 Z  D Ž Æ �yF   ™ â  $xIH$   ÿŽ Æ �yE   ™ ä rI   ÿàÿ ä �yD   ™ â r%
   ÿàõ Æ �y    ™ â  r#   ÿä Æ �y   ™ è" �yH   ™ è  �yI   ™ è �yK   ™ è  �y

    ™ è �yJ   ™ è �y   ™ â rLÿ&   1  ÄO r1ÿ(   1  ä xML   ÿ   ä xÿ   L  ä/ xMM   ÿÀŽ â $x	1   ÿ Ž â x   ÿÀŽ ä xMMøŽ Äxÿ     ä xL   ÿ   ä xOÿ   1  ä x	   ÿÀŽ ä xøŽ ä xM   øŽ Ä xL   ÿ   ä xNÿ   /  äxOO   ÿÀŽ ä x/øŽ ä x	ÿ   L  ä x1   ÿ   Ä x   øŽ ä xNONøŽ ä x/ÿ   1  ä x	   	øŽ ä x   øŽ ä xL   ÿ   Ä xL   ÿ   ä x1   ÿ   äx/N   /øŽ â $xN1   ÿ Ž â x   øŽ ä x	   øŽ ä x   øŽ Ä xL	   ÿ   ä rÿ&     ä� x	1   ÿ   äxL1	   ÿ   ä x   øŽ ä x//   NøŽ Ä x1   ÿ   ä x   	øŽ ä xÿ     ä x	/   LøŽ ä x11   ÿÀŽ ä rÿ(     Ä x/ÿ   ,  äxL   ÿÀŽ ä x11,øŽ ä x   ÿ   â $xM   ÿ Ž â x/L/øŽ ä x,ÿ     Ä x1   øŽ ä x1ÿ     ä x,/   ,øŽ ä x/   ÿ   ä xMM   ÿÀŽ ä xO1   ÿÀŽ Ä x1   ÿ   äx   /øŽ â $x/   ÿ Ž â xMM2øŽ äxLÿ   2  ä x2   ÿ   ä x,,   1øŽ Ä x1	   ÿ   ä x2M   2øŽ ä x   /øŽ â &r/3   ÿ   âxM   ÿ   ä xLOLøŽ â &r33   ÿ   â x,   1øŽ â &r/0   /   â xOÿ     Ä x22   MøŽ ä x1   ÿ   â&r30   3   â xLL   OøŽ â &r04   /   â xM   ÿ   ä x,2   1øŽ â 2r1+        â xLL   MøŽ â &r0	5   0   â x	   ÿ   Ä 0r/ÿ1  0 A   â2r27        â xL   øŽ â &r4,4   3   â Er0 0       â 0r1ÿ1    A   â  x/   ?  A  â &r45   4   ä 0r/ÿ2  0 A   â#r110      â &r6   ÿ   â Er3 4       â 0r2ÿ2    A   â  x//   ?  A  â &r8      Æ #r223   /   â 2r/+:        â&r9      ä&r56   ÿ   ä &r	<      ä 0r0ÿ/  0 A   ä&r58   5   â Er        â 0r/ÿ/    A   â  x00   ?  A  â &r5,9   5   â 2r:7:        Ä #r/   0   â &r0;   ÿ   â !r1       â &r5<   5   ä 0r1ÿ:  0 A   â!r       â &r0=   0   â Er5 5       â 0r:ÿ:    A   â 2r+>        â  x11   ?  A  â &r0?   0   Ä &r;;   ÿ   â #r1:5   1   â 0r/ÿ  0 A   ä&r0	B   0   â !r1       â &r;=   ;   â  x1/   ?  A  â 0rÿ    A   â Er/ 0       â &r;,?   ;   â 2r>7>        â &r0@   ÿ   Ä &r@@   ÿ   â #r//   1   â !r2       â &r;B   ;   â 2r+        â&r0A   0   ä0r2ÿ>  0 A   â!r*/*       â &r@A   @   â 2r/7        â &r0C   0   â Er; ;       â 0r>ÿ>    A   â  x32   ?  A  â 0rÿ  0 A   Ä &r@,C   @   ä &r0	F   0   â#r>>;   3   â &r@F   @   â  x1   ?  A  â 0r3ÿ/  0 A   ä&rE   ÿ   ä0r2ÿ/    A   â Er/ 0       â &rEE   ÿ   Ä 0rÿ    A   â Er5 @       â  x33   ?  A  â &rED   E   ä#r/   1   â 2r/+        â&rD      â #r325   3   â &r0,   E   â 2r7        â &r2H   ÿ   âx   ÿàõ â &rH   ÿ   Ä &r      ä &r0I   0   â rÿ   ÿä â 0rÿ  0 A   ä&rK   2   âx   ÿàõ â &rKK      â Er 0       â 0rÿ    A   â 2r++
        âx))    ÿàÿ â  x   ?  A  â &rJ      â2r
7
        â &r,,J   K   Ä $rÿÿ    â | )   pbô â !r       â &r	I      â#r      â &r	      ä&r,   ,   ä 0rÿ/  0 A   ä 0rÿ+  0 A   Ä 0rÿ
  0 A   âEr        â 0r/ÿ/    A   â Er        â 0r+ÿ+    A   â Er	 ,       â 0r
ÿ
    A   â  x   ?  A  â  x   ?  A  â  x   ?  A  ä #r/      â #r+      â #r
	      â !r->-       â !r.3.       â !r       â !r''       â !r       â !r       â x    ÿàÿ Ä x    ÿàÿ ä x      ÿàÿ ä x""    ÿàÿ ä x$$    ÿàÿ â D)       À â Gy   ìÿÿÿÿƒ ê Ay        € ê Ey  ð     € â G‰  Ð     € ê $x  €  Ž Ê ˆs   úÿ H   è ˆs  €úÿ H   è ˆs   ûÿ H   è ˆs  €ûÿ H   è ˆs  * üÿ H   è ˆs  -€üÿ H   è ˆs   ýÿ H   è ˆs  .€ýÿ H   è ˆs   þÿ H   è ˆs  '€þÿ H   è ˆs   ÿÿ H   è ˆs  €ÿÿ H   ä Ay        € ê {          ì M        € ê „y  €   H   " z  r  p`ð Ä z+  r      â „y     H   b r ÿ   p"p Æ „y    H   ¢ $x)+   ÿŽ Æ „y	 €  H   è „y €  H   ( „y    H   h „y €  H   h „y    H   h „y €  H   h „y    H   è „y
    H   â !r        â !r       Ä/ „y €  H   â !r       ÆO ‰      " !r.	.       Æ� ‰       b !r       â!r**       È‰	      ¢ !r-       Æ ‰      â !r       Æ ‰*      " !r'       â !r       ä ‰      " $x'+   ÿŽ â !r       â !r         â/ ‰      b !r

       Æ ‰.      b !r		       ÆO ‰       ¢ !r       â !r       ä� ‰      â !r*       Æ‰      ( ‰
      " !r       Æ ‰      " !r       â/ !r..       ä‰      b !r        ÆO y      &   ¢ !r       Æ� ‰ €    â x   ÿàÿ â !r       äy      '   " z  r  p`ò â !r
       ä ‰ €    b x    p"ò â !r       Ä ‰      ( ‰	      " !r       Æ/ ‰      b $z  x  Ž ÆO ‰      ¢ !r
       Æ� ‰      â $z  ~   Ž Æ‰.      " !r       â
  r  ÿàù ä‰      b !r       â r+    ÿ† ä ‰       " !r	       â  '   ÿàõ Ä ‰"      " !r       â/ ÿÿ   ÿä â !r       âO ‰
 @    b rÿÿ   ÿäÿ â !r       ä� ‰ €    ¢ 
 j  ÿˆ â !r..       Ä‰ €    â z j  ÿ† â !r       ä‰	 @    " ÿÿ   ÿä â !r         ä ‰ €    " 
 j  ÿ„ â !r""       Ä ‰ €    b 
 k   ä z k  � â ‰ €    b !r
       â/ 
 k   ä ‰! €    b !r       âO r+    ÿ† Ä ‰. €    ¢ !r       â� r&ÿÿ   ÿäÿ ä ‰ €    â !r	       âz
 j  ÿ† ä ‰#  €    " !r       â z k  &� Ä ‰%"      " !r       â	+   ÿàÿ â !r       â ‰ @    b  )   ÿàù â !r!!       ä/ ‰ @    b ( 	   ÿàõ â $ÿÿ   ÿ â !r.       âO ‰      ¢ 
 j  ÿˆ â !r       Ä� ‰ @    â +ÿÿ   ÿä â !r# #       ä‰ @    " ( j  ÿ„ â !r%"%       ä ‰ @    " 
	 k   ä ( k  + â ‰"! @    " !r       â        Ä ‰ @    b !r       Æ/ ‰  @    b !r+       âO '   ÿàÿ ä ‰      ¢ !r       Æ� ‰$# @    â !r       Æ‰&% €    " !r       â !r"!"       â !        â ‰      " !r       Æ‰      b !r         Æ/ ‰      b !r-       âO     ÿàõ ä‰      ¢ %  j   Ž â!r$#$       â� (ÿÿ   ÿä â ‰      â %  j  ! Ž â !r&%&       â j  ÿ„ â‰      &  k  ( â ‰!       " !r       Æ ‰#"      " !r       Æ‰'& @    b !r       Æ/ ‰%$      b !r       ÆO †	 +    â !r       Æ� † -   â !r       Æ†	     â !r! !       â !r#"#       ä †     â !r'&'       Æ†	     â !r%$%       Æ/ †    è †	     è †     è †	 
!    è † 
#   è ‰'      ( †	 %    â M™        € ê x)   ÿàÿ âO !r''       Æ r    ÿàñ È r ÿÿ   ÿä  ä z j  ÿ€ È z k     Ê †y '    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y       "   " ¹z  l      â Ey  ð    € â ™x?    â y      !   b ¹z  F   
   â x      ÿ  â x&      ÿ  â y      %   ¢ $r+ÿÿ   ÿ Ž â x(      ÿ  â r ÿ       â $r-ÿÿ   ÿ Ž â r ÿ       Ä r  ÿ   pRð är  ÿ   pRò är    ÿ(Ž È/ r   ÿ� ä x   ÿ   äO xÿ     È |    pbô Ú G)  °    € ê y      '   " r ÿ       â $rÿÿ   ÿ Ž â x
        â y      &   b $rÿÿ   ÿ Ž â z" q  ÿàÿ â 'z y   Ž È 'z s   Ž â/ zÿ z    Æ $z p  Ž â zÿ t    â $z	 |  Ž ä$z |  Ž â x   ÿàÿ â $z	 v  	Ž ä$z v  Ž ä $z p  	Ž â x	þÿÿÿÿÀŽ â $z }  ÿŽ Ä %v X  
Ž â r		  €ÿàÿ Æ $z w  ÿŽ â x	   ÿ   â %v X  
Ž â x   ÿàù ä r   ÿàõ â $vÿ q  ÿ Ž â x   ÿà÷ ä x!	   ÿ   â $x   Ž âxÿ     â $x    Ž â rÿ   ÿä Ä rÿ   ÿäÿ ä r   ÿŽ ä r1ÿÿ   ÿä ä x$!   ÿàÿ ä x%   ÿ   äx#     Ä �y,üÿÿ ™ ¢ %x   Ž Æ �y*üÿÿ ™ è  �y   ™ ( �y    ™ b %x   Ž â r   ÿàõ Ä �y/øÿÿ ™ è �y    ™ â  r
ÿ1   ÿä â $t5ÿ$   ÿ Ž ä �y   ™ ä  %v Z  5 Ž â r   ÿà÷ â �y7øÿÿ ™ ä  $x
$   ÿŽ â r%   ÿàõ È/ r	   ÿàÿ È r#   ÿä ä r"   ÿàõ â $r
ÿÿ    Ž ä �y   ™ â  rÿ1   ÿä Æ �y.   ™ â  %v Z  5 Ž Æ �y2
    ™ â $x$   ÿŽ Æ �y0   ™ ä  r   ÿàÿ â $rÿÿ   1Ž â �y4   ™ â  r%   ÿàõ â %v
 Z  5 Ž Æ/ r	#   ÿä â $x$   ÿŽ â r        ä r%
   ÿàõ â �y6   ™ â  r   ÿàÿ Æ �y3   ™ â  r#   ÿä ä r   ÿàõ â �y    ™ è �y8   ™ â  $r<ÿÿ   1 ä %v Z  5 Ž â �y:   ™ æ  $x=<$   ÿŽ â �y;   ™ è r=   ÿàÿ â �y9   ™ â r%   ÿàõ Æ �y<
    ™ â  r	#   ÿä ä r@    ÿàõ â �y   ™ æ" rAÿ1   ÿä â �y>   ™ è �y=   ™ â %v
@ Z  5 Ž Æ �y?   ™ â  $xAA$   ÿŽ â r%
   ÿàõ ä/ �y5   ™ â  $x   AŽ Æ �y    ™ ä r#   ÿä ä �yB   ™ è  �y@   ™ è  �yA   ™ è  �y

    ™ è  �yC   ™ è  �y   ™ â" x   ÿàõ Ä x    ÿàÿ ä rÿ   ÿä ä |    pbô ä rDÿ!   ,  äOr,ÿ$   ,  ä xED   ÿ   Ä xÿ   ,  ä x,   ÿ   ä/ x	E   ÿÀŽ â $xD   ÿ Ž â xÿ     äxEÿ   D  ä x   ÿÀŽ Ä x   ÿÀŽ ä x		øŽ äxÿ     ä xEE   ÿÀŽ ä xøŽ ä xøŽ Ä xÿ   ,  ä x	   øŽ ä x,   ÿ   ä x	D   ÿ   ä xEøŽ ä xEÿ   D  Ä x   øŽ ä x   øŽ â $x,   ÿ Ž âx,   ÿ   ä x   	øŽ ä x	D   ÿ   ä x   EøŽ Ä xED   ÿ   ä x   øŽ â $x,   ÿ Ž â x   øŽ ä x   	øŽ ä r	ÿ!   *  ä� x,,	   ÿ   Ä x   EøŽ ä xED	   ÿ   ä x   øŽ ä r*ÿ$   *  ä x	   ÿ   ä x   ,øŽ Ä xÿ   	  ä x   EøŽ ä xE*   ÿ   ä x,   ÿÀŽ ä xÿ     ä x   ÿÀŽ Ä xEE   ÿÀŽ ä x,øŽ ä x,ÿ   *  ä xøŽ ä xÿ     äxEøŽ Ä xEÿ   	  ä x,,   ÿÀŽ ä x   EøŽ â $xD	   ÿ Ž â x,,øŽ ä xEÿ   *  ä x*   ÿ   Ä x,,   EøŽ â $xE*   ÿ Ž â x   DøŽ ä x   øŽ ä xD	   ÿ   äx	   ÿ   ä x   EøŽ Ä x   DøŽ â &rD   ÿ   â xE	   ÿ   äx   øŽ ä x*   ÿ   ä x			   ÿ   ä x   EøŽ â &rE.   D   â x,   øŽ â $x,*   ÿ Ž â x   	øŽ â 2r	/2        â &rE0   E   Ä &r   ÿ   â x,   ,øŽ â &rE4   E   â x*	   ÿ   â 0r*ÿ	  0 A   ä&r.      â x   øŽ â 0r	ÿ	    A   â Er E       â  x**   ?  A  â &r0,0      â 2r272        Ä #r	   *   â &r04   0   ä 0r	ÿ2  0 A   ä&r*6   ÿ   â !r--       â 0rÿ2    A   â Er0 0       â &rE3   *   â  x	   ?  A  â 2r	/        Ä &rE8   E   â #r0      â 2r7        â 0rÿ	  0 A   ä&rE:   E   ä 0r.ÿ  0 A   ä0r*ÿ    A   â  x   ?  A  â &r0;   ÿ   â Er E       â 0r	ÿ	    A   Ä &r09   0   ä &r;;   ÿ   â #r	      â 2r	/<        â&r0   0   ä &r;9   ;   â !r((       â !r       â &r0>   0   â 2r<7<        â 0rÿ	  0 A   Ä &r=   ÿ   ä &r;,   ;   â Er0 0       â 0r	ÿ	    A   â  x   ?  A  â &r?      ä &r;>   ;   ä 0rÿ<  0 A   â #r	0      â &r==   ÿ   Ä &r5      â Er; ;       â 0r<ÿ<    A   â 2r	/        â x   ?  A  â &r=?   =   ä &rB      â !r))       â #r<<;      â 0rÿ	  0 A   â 2r7        â &r=,5   =   â Er        â &r66   ÿ   Ä &r@   ÿ   ä&r@   ÿ   â  x   ?  A  â 0r	ÿ	    A   ä &r=B   =   ä 0rÿ  0 A   ä &r33   6   Ä &rA      ä&rA      â #r	      â 0rÿ    A   â Er =       â 2r//
        â&r3,8   3   â  x   ?  A  â &rC      â2r
7
        â &r,C      â #r      â &r3:   3   Ä &r      ä&r      ä 0rÿ/  0 A   ä0rÿ
  0 A   âEr3 3       â 0r/ÿ/    A   â Er        â 0r
ÿ
    A   â Er	        â  xE.   ?  A  â  x   ?  A  â  x   ?  A  â x   ÿà÷ â #r**3   E   â #r/      â #r
	      â rÿ   ÿäÿ â !r*       â !r+<+       â !r''       â !r&&       â !r       â !r       â x    ÿàÿ Ä x    ÿàÿ ä x      ÿàÿ ä x""    ÿàÿ â D)       À â Gy  Ðîÿÿÿÿƒ ê Ay        € ê Ey  Ð     € â G‰  °     € ê $x  @  Ž Ê ˆs  - ûÿ H   è ˆs  (€ûÿ H   è ˆs   üÿ H   è ˆs  €üÿ H   è ˆs  ) ýÿ H   è ˆs  +€ýÿ H   è ˆs  ' þÿ H   è ˆs  &€þÿ H   è ˆs   ÿÿ H   è ˆs  €ÿÿ H   ä Ay        € ê {          ì M        € ê „y      H   " z  r  p`ð â $v%ÿ r  ÿ Ž â x#        â „y €   H   b r ÿ   p"p â $x!%   ÿŽ Ä „y    H   ¨ „y €  H   è „y    H   ( „y €  H   h „y
    H   h „y €  H   h „y    H   è „y	 €  H   â !r  -       Æ y      '   " !r(       Æ/ ‰       b !r       ÆO ‰      ¢ !r       Æ� ‰      â !r))       Æ‰      " !r++       Æ‰)      b !r'
'       Æ ‰
+      " !r&       â !r        â/ !r       Æ ‰      b !r       âO !r	       ä ‰       ¢ !r       Æ� ‰'      â !r       Æ‰      " !r)       Æ‰      b !r
+
       Æ ‰      " !r       Æ/ ‰	      b !r         ÆO ‰      ¢ !r'       Æ� ‰  €    â !r       Æ‰      " !r       Æ‰      b !r       Æ ‰
      ( ‰      " !r		       Æ/ ‰      b !r       ÆO ‰      ¢ !r        â� !r       ä‰ €    â !r       Æ‰ €    " !r
       Æ ‰ @    " x
   ÿàÿ â !r       ä ‰	 €    b z 
 r  p`ò â !r       ä/ ‰ €    b x    p"ò â !r       ÄO ‰ €    ¨ ‰ €    ¢ !r       Æ� y       &   â !r       Æ‰ €    " !r
       Æ ‰ €    " !r	       Æ‰ €    b !r       â/ !r       äO ‰ @    b !r       Æ ‰ @    ¢ $z   x  Ž Æ� ‰
      â !r       â$z  ~   Ž ä ‰ @    " !r       Æ ‰ @    "  !   ÿàõ â !r       âr%    ÿ† â ‰ @    b  ÿÿ   ÿä â !r       ä/ ‰ @    b 
 j  ÿ„ â $rÿÿ   ÿŽ â !r       âO ‰ @    ¢ z j  ÿ† â !r'
       Ä� ‰ @    â %   ÿàÿ ä !   ÿàÿ â ‰ @    â !r       â
 k    â !r       â ‰      " z k  � â !r       Ä‰      " !    ÿàõ â !r       ä/ ‰      b     ÿà÷ â%v  j  # Ž â !r       âO ‰      ¢ 
	  r  ÿàù â !r       Ä� ‰      â r%    ÿŠ â !r       ä ‰      b r"ÿÿ   ÿäÿ ä z j  ÿŠ â ‰       b $ÿÿ   ÿä â !r#       â z k  "� â ‰      " 
	 j  ÿˆ â !r       â"ÿÿ   ÿä â ‰      " ! j  ÿ„ â !r       â/  ÿÿ   ÿäÿ â !r       âO 
		 k  $ â †	 '    â 
 j  ÿ† â !r       â� ! k  " â † #   â  k   � â !r       Ä†	     â !r        Æ †     â !r       Æ †	     â !r       Æ†    è †	     è † 
    è †	     â M™        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      "   " ¹z  l      â Ey  p    € â ™x?    â y      !   b ¹z  F   
   â x      ÿ  â $rÿÿ   ÿ Ž â y      %   ¢ r ÿ       â x      ÿ  â r  ÿ       â $rÿÿ   ÿ Ž â r ÿ   pRð Ä r ÿ   pRò är   ÿ(Ž È/ r   ÿ� ä x   ÿ   äO x)ÿ     È | )   pbô Ú G)  @    € ê y      '   " r ÿ       â $rÿÿ   ÿ Ž â x
        â y      &   b $z	 p  )Ž â z) q  ÿàÿ â $rÿÿ   ÿ Ž ä 'z y   Ž È 'z s   Ž â/ zÿ z    Æ $z }  ÿŽ â zÿ t    â $z |  )Ž ä$z |  	Ž â x	   ÿàÿ â $z v  Ž ä$z v  Ž ä $z	 p  Ž â x	þÿÿÿÿÀŽ â $vÿ q  ÿ Ž Ä %v X  
Ž âr		  €ÿàÿ ä r)   ÿŽ â %v X  
Ž â x	   ÿ   ä x   ÿàù â $z w  ÿŽ â x   ÿà÷ â $x   )Ž â xÿ     â $x!	   ÿ Ž â r   ÿàõ Ä rÿ   ÿä ä rÿ   ÿäÿ ä r8ÿÿ   ÿä ä x%!   ÿàÿ ä x'   ÿ   äx#     Ä �y,üÿÿ ™ ¢  %x   Ž Æ �yüÿÿ ™ è �y(    ™ h	 �y+   ™ â r)   ÿàõ â %x   Ž Ä �y$øÿÿ ™ â rÿ8   ÿä â $t6ÿ$   ÿ Ž ä �y     ™ ä  %v
 Z  6 Ž ä �y-   ™ ä  $x	$   ÿŽ â r'
   ÿàõ â �yøÿÿ ™ æ  r	   ÿàÿ Ê $x	   # â r   ÿàõ â�y
    ™ æ rÿ8   ÿä â %v Z  6 Ž â �y&   ™ æ  $x$   ÿŽ â r'   ÿàõ â �y*   ™ æ  r   ÿàÿ â �y"   ™ æ  r#   ÿä â �y/   ™ â  r1   ÿàõ Æ �y0   ™ ä $r2ÿÿ   8 ä %v
1 Z  6 Ž â�y.   ™ & $x32$   ÿŽ â r'
   ÿàõ â �y1    ™ & r3   ÿàÿ â �y2   ™ & r	#   ÿä â �y3   ™ " r7   ÿàõ Æ �y4   ™ " r9ÿ8   ÿä â %v7 Z  6 Ž ä �y5   ™ $ $x99$   ÿŽ â r'   ÿàõ â/ �y
    ™ $ $x   9Ž Ä �y6   ™ & r#   ÿä â �y:   ™ ( �y7   ™ ( �y9   ™ ( �y    ™ ( �y;   ™ ( �y
   ™ "# x   ÿàõ Ä x))@   ÿàÿ ä rÿ   ÿä ä x   ÿàõ È rÿ   ÿä ä | )   pbô ä x@   ÿàÿ ä x@   ÿàÿ ä x@   ÿàÿ ä r<ÿ!   ,  ÄO r,ÿ%   ,  ä x=<   ÿ   ä xÿ   <  ä x	,   ÿ   ä x==   ÿÀŽ ä x?ÿ   ,  Ä x   ÿÀŽ â/ $x<   ÿ Ž â x	   ÿÀŽ ä x==(øŽ äx(ÿ   (  ä x>ÿ   +  ä� x??   ÿÀŽ Ä x+øŽ ä x+,   ÿ   äx((øŽ ä x=   øŽ ä x	ÿ   <  â $x,   ÿ Ž â x>?>øŽ Ä x=ÿ   ,  ä x+   +øŽ ä x<   ÿ   ä x	(   	øŽ ä x=>   =øŽ ä x<   ÿ   Ä x(,   ÿ   ä x++   øŽ ä x   øŽ ä rÿ!     ä x<   ÿ   ä x	   øŽ Ä x==   (øŽ â $x(,   ÿ Ž â x	<	   ÿ   ä x<   ÿ   ä x   øŽ ä x   	øŽ ä x,,	   ÿ   Ä x	+   (øŽ ä x+<   ÿÀŽ ä rÿ%     ä x=   ,øŽ ä x,ÿ     ä x(ÿ      Ä x++ øŽ ä x    ÿ   ä x=,   ÿÀŽ ä x,    ÿÀŽ â $x    ÿ Ž â x>ÿ     ä x(=(øŽ Ä x=ÿ     ä x<ÿ   -  äx?>   ÿÀŽ ä x +    øŽ ä x+   ÿ   ä x,,-øŽ Ä x((   =øŽ ä x-   ÿ   ä x<?<øŽ ä x=   ÿ   ä x?ÿ     ä x+    +øŽ Ä x    ÿ   ä x((   -øŽ ä x,,   =øŽ â $x=   ÿ Ž â x-	   ÿ   ä x<<   ?øŽ ä x?   ÿ   Ä x++    øŽ ä x(   -øŽ â $x-   ÿ Ž â x,,   =øŽ â &r=&   ÿ   âx<<   ?øŽ â &r?+&   ÿ   â x-,   -øŽ â &r=*   =   âxA	   ÿ   â &r?*   ?   Ä &r=	"   =   âx<   AøŽ â &r?-"   ?   â 2r"$        ä2r         â &r?/   ?   ä0r&ÿ"  0 A   ä &r=/   =   Ä 0rÿ   0 A   â Er? ?       â 0r"ÿ"    A   â  x/&   ?  A  â &rA0   ÿ   â Er= =       â 0r ÿ     A   â  x   ?  A  â &rA.   A   â#r""?   /   â &r/+0   ÿ   â #r =      â 2r 1        â &rA	2   A   Ä &r/.   /   â !r       â &rA3   A   â 2r1$1        â 0rÿ   0 A   ä&r/-2   /   â ErA A       â &r&4   ÿ   â!r"       â 0r ÿ     A   â  x   ?  A  â &r4+4   ÿ   Ä &r/3   /   ä 0r"ÿ1  0 A   â#r A      â &r45   4   â Er  /       â 0r1ÿ1    A   â  x""   ?  A  â 2r(        â&r3-6   4   â 2r$        â #r11    "   â &r 7   ÿ   Ä &r3:   3   ä &r++7   ÿ   ä 0rÿ  0 A   ä&r&5   &   ä&r"9       â Er 3       â &r9   +   â  x    ?  A  â 0rÿ    A   â 2r        â&r/	6   &   â2r$        â &r		;   "   Ä &r--;      â #r       â 0rÿ  0 A   â !r       â &r/:   /   ä&r	
   	   ä&r-
   -   ä 0rÿ(  0 A   Ä 0rÿ  0 A   â x   ?  A  â 0r(ÿ(    A   â Er/ /       â 0rÿ    A   â Er	 	       â 0rÿ    A   â Er -       â  x   ?  A  â  x   ?  A  â #r	      Ä #r((/      â #r      â !r 1        â !r       â !r(       â !r       â !r       â G©  0ñÿÿÿÿƒ ê Ay        € ê Ey  °     € â G‰  �     € ê $x   Ž Ê ˆs  üÿ H   è ˆs €üÿ H   è ˆs  ýÿ H   è ˆs  €ýÿ H   è ˆs  þÿ H   è ˆs €þÿ H   è ˆs  ÿÿ H   è ˆs €ÿÿ H   ä Ay        € ê {          ì M        € ê „y
    H   b z  r  p`ð Æ „y     H   ¢ r ÿ   p"p Æ „y €  H   ( „y    H   è „y €   H   ( „y    H   h „y €  H   h „y €  H   h „y    H   h „y €  H   h „y    H   b !r

       Æ/ „y    H   b !r       ÆO „y	 €  H   ¢ !r        Æ „y    H   " !r       Æ� „y €  H   â !r       Æ„y!    H   " !r       Æ„y  	  H   b !r       Æ „y €	  H   b !r       Æ „y  
  H   b !r       Æ „y €
  H   b !r       Æ „y#    H   b !r       Æ „y €  H   â !r       Æ/ „y €  H   â !r	       ÆO „y €  H   â !r

       Æ ‰      " !r       â� !r!       ä‰      b !r       Æy      '   ¢ !r       Æ ‰      â !r

       Æ ‰	      " z  r      â !r       ä ‰
      b !r#       Æ ‰      ¢ !r       â !r       ä ‰      " !r       â/ !r       ä ‰       d !r       â !r       â� ‰      â !r		       Æ‰      " !r
       Æ‰	      b !r       ÆO ‰      ¢ !r       Æ ‰
      " !r         Æ/ ‰      h ‰      b !r       Æ� ‰  €    â !r       Æ‰ €    " !r	       â!r       äO ‰ €    ¢ !r
       Æ ‰	 €    " !r       Æ/ ‰ €    b !r       Æ y      &   b !r        Æ� ‰ €    â !r       Æ‰ €    ( ‰      " !r       ÆO ‰ @    ¢ !r
	       â !r       ä/ ‰ @    " $z  x  Ž âx   ÿàÿ ä ‰	 @    b !r       â� z  r  p`ò Ä ‰
 @    â $z  ~   Ž â !r       âx    p"ò â ‰ @    " $x   ÿŽ â !r       ä ‰ @    b !r       âO 
  r  ÿàõ Ä ‰ @    ¤ ÿÿ   ÿä ä ‰ €    ¢ !r       â 
 j  ÿ„ â!r	       â/ ‰      $ 
 k   â !r
       â� ‰      b r    ÿ† â !r       Ä‰      â 	   ÿàÿ â !r       ä‰      " rÿÿ   ÿäÿ â !r       äO ‰      ¢ z j  ÿ† â !r       Ä ‰      b  	   ÿàõ â$t	ÿ   ÿ Ž â z k  � â ‰ @    b     ÿà÷ â !r       â ÿÿ   ÿä â ‰      " 
 j  ÿ„ â !r       â/ ÿÿ   ÿäÿ â %v  j  	 Ž â 
 j  ÿ† Ä  k   â !r       â� !r       â
 k  � â †	     â !r       ÆO †    â !r       Æ†	     â !r       Æ † 
    â !r       Æ †	     è †    è ‰      ( †	     â M™        € ê x   ÿàÿ â/ !r       Æ r    ÿàñ È r ÿÿ   ÿä  ä z j  ÿ€ È z k     Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      "   " ¹z  l      â Ey  @    € â ™x?    â y      !   b ¹z  F   
   â r ÿ       â x      ÿ  â y      %   ¢ $r ÿÿ   ÿ Ž â r  ÿ       â $rÿÿ   ÿ Ž â y      &   è y      '   " r ÿ   pRð Ä r ÿ   pRò â$x
    Ž â/ x   ÿ   ÈO r

   ÿ� Ê $À   Ž â x"ÿ     È | "   pbô Ú G)  ð    € ê $rÿÿ   ÿ Ž â�r ÿ       â $z p  "Ž â xþÿÿÿÿÀŽ â 'z y   Ž â z  q      ä r
  €ÿàÿ â 'z s   Ž â zÿ z    ä r ÿ       â $tÿ   ÿ Ž â zÿ t    â $z	 |  "Ž â r"   ÿŽ â $z |  Ž â x   ÿàÿ â $z	 v  	Ž â z" q  ÿàÿ â $z v  Ž Ä $z	 p  	Ž ä %v X  Ž È %v	 X  Ž â x   ÿàù Æ $z }  ÿŽ â x   ÿà÷ â $z w  ÿŽ ä $x!   ÿ Ž ä$x#   ÿ Ž â r   ÿàõ â $rÿÿ    â x
ÿ   !  â $r	ÿÿ   	Ž â r3ÿÿ   ÿä â $x)!   ÿ Ž â x'#   ÿàÿ Ä x%!   
  ä �yüÿÿ ™ ¢  %x!   Ž Æ �yüÿÿ ™ è �y+    ™ h	 �y(   ™ â %x!   Ž â r
"   ÿàõ Ä x5 $       â �y0øÿÿ ™ ä  $rÿÿ   3 ä �y&    ™ â  %v

 Z  5 Ž Æ �y$   ™ â  $x$   ÿŽ â r)
   ÿàõ ä �y2øÿÿ ™ ä r   ÿàÿ Ê$x   % â r   ÿàõ È �y*   ™ â r 
       â $r
ÿÿ   3 ä �y,   ™ â %v Z  5 Ž Æ �y-    ™ â  $x
$   ÿŽ Æ �y.   ™ â r   ÿà÷ Æ �y/   ™ â r
)   ÿàõ ä r   ÿàÿ â $r1ÿÿ   3Ž Æ r%   ÿä â %v Z  5 Ž â �y    ™ æ $x1$   ÿŽ â �y
   ™ "r)   ÿàõ ä $x   Ž â �y1
   ™ ( r%   ÿä â �y
   ™ ( �y4   ™ ( �y6
   ™ ( �y7   ™ ( �y    ™ ( �y5   ™ ( �y8   ™ " x   ÿàõ È/ r	ÿ	   ÿä ä x   ÿàõ ä x""@   ÿàÿ Æ $rÿÿ    â | "   pbô ä x@   ÿàÿ ä x@   ÿàÿ ä r9ÿ#     äOrÿ'     Æ $x:9   ÿ Ž âx;ÿ   9  â $x9   ÿ Ž â x<   ÿ   ä x::   ÿÀŽ ä xÿ   +  äx
:+øŽ ä x;   ÿÀŽ Ä x=ÿ     ä x+<   ÿÀŽ ä xøŽ ä x

   øŽ â $x9   ÿ Ž â xÿ   (  ä� x==   ÿÀŽ Ä xÿ   9  ä x(+(øŽ ä x+   ÿ   ä x=øŽ ä x;ÿ     ä x   øŽ Ä x
   øŽ â $x=   ÿ Ž â x((   +øŽ ä x
9   ÿ   ä x+   ÿ   ä x   ;øŽ ä x   
øŽ Ä x;9   ÿ   äx
(   +øŽ ä x+   =øŽ ä x9	   ÿ   â $x9   ÿ Ž â x   ;øŽ ä x	   ÿ   Ä rÿ#     ä x++   øŽ ä xÿ     ä rÿ'     ä x   øŽ ä x   ÿ   Ä x;   ÿÀŽ â $x   ÿ Ž â x

   9øŽ ä xÿ   &  ä x9   ÿÀŽ ä x;øŽ ä x=   ÿÀŽ Ä x9&øŽ ä x;ÿ     ä x9   ÿ   ä x=$øŽ ä x&   ;øŽ ä x=   ÿ   Ä x   9øŽ â $x9   ÿ Ž âx;   ÿ   ä x(ÿ     ä x   =øŽ ä x&&   ;øŽ â $x;   ÿ Ž â x=   ÿ   Ä xÿ   $  ä x((   ÿÀŽ ä x   9øŽ ä x9	   ÿ   ä x   =øŽ â $x=   ÿ Ž â x(øŽ Ä x$ÿ     ä x;   ;øŽ â &r*   ÿ   â x   $øŽ ä x   ÿ   ä x   =øŽ â &r=,      â x9&   9øŽ â &r*;*   ÿ   â x   øŽ â 2r0-        â x	   ÿ   â &r
.   =   Ä &r*9,   *   â x=   øŽ â &r+/      ä0rÿ  0 A   â2r2-        â &r*.   *   â Er-        â 0rÿ    A   â  x   ?  A  â &r*=/   *   Ä 0rÿ  0 A   â#r--      â &r1   ÿ   â	Er/ *       â 0rÿ    A   â  x   ?  A  â &r      ä &r1;1   ÿ   â #r//      â 2r0        â &r
      Ä &r19   1   ä &r4   ÿ   ä &r+6      ä &r4;4   ÿ   ä 0rÿ  0 A   ä&r7      âEr        â &r497   4   â  x   ?  A  â 0rÿ    A   â 2r2        Ä 2r00        â&r1   1   â 2r2        â &r
5      ä&r45   4   â #r      â 0r
ÿ  0 A   ä 0rÿ0  0 A   ä &r1=6   1   Ä &r+8      ä&r4=8   4   ä 0rÿ  0 A   â x
   ?  A  â 0rÿ    A   â  x   ?  A  â 0rÿ    A   â Er
 1       â 0r0ÿ0    A   â Er        â  x+   ?  A  â Er 4       â #r

      Ä #r0      ä #r   +   â !r-       â !r /        â !r         â !r
       â !r       â !r       â G©  Póÿÿÿÿƒ ê Ay        € ê�ˆ  ýÿ H   è ˆ  €ýÿ H   è ˆ   þÿ H   è ˆ €þÿ H   è ˆ  ÿÿ H   è ˆ €ÿÿ H   è {          ì M        € ê „y     H   " z  r  p`ð â $z x  Ž Ä „y    H   b r ÿ   p"p Æ „y €   H   ¨ „y	    H   è „y    H   ( „y €  H   h „y    H   h „y €  H   h „y    H   h „y €  H   h „y €  H   b !r       Æ „y
    H   " !r       Æ/ „y €  H   b !r        ÆO „y    H   ¢ !r	       Æ� „y €  H   â !r        Æ„y# €  H   â !r       Æ„y!    H   â !r       Æ „y% €  H   â !r       Æ ‰      " !r       â !r       â !r       â !r

       â !r       â/ !r       ÈO ‰	      " !r       Æ� ‰      b !r       â!r#       ä ‰      ¢ !r

!       Æ ‰       â !r%       Æ ‰
      ( ‰      b !r		       â !r       È/ ‰	      " !r       ÆO ‰      b !r         Æ� ‰      ¢ !r
       Æ‰  €    â !r       Æ‰      ( ‰      b !r	       â !r       È/ ‰	 €    " !r
       ÆO ‰ €    b !r        Æ� ‰
 €    ¢ !r       Æ‰  @    â !r       Æ‰ €    ( ‰ €    b !r	       â x   ÿàÿ â !r       Æ/ z  r  p`ò â ‰ @    " $z ~  Ž â !r
       âO x    p"ò â ‰ @    b z  r      â !r         â� !r       È‰	       â !r       Æ‰ @    ¢    ÿàÿ Æ ‰ @    è ‰ @    " !r       â 
 r  ÿàõ â !r       Æ/ ‰
      " ÿÿ   ÿä â !r       äO ‰      b !r       â� 
 j  ÿ„ ä ‰      ¢ !r       â
 k   Ä ‰      â r   ÿ† â !r 	       â 	   ÿàõ â‰      " $tÿ   ÿ Ž â r ÿÿ   ÿäÿ â !r
       â ÿÿ   ÿä â %v j   Ž â 	 j  ÿ„ Ä z j  ÿ† ä 		 k   â †	     â !r       â/ z k   � ä †    â !r       ÆO †	     â !r       Æ� †     â !r       Æ†	     â M™        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      "   " ¹z  l      â Ey  À    € â ™x?    â y       !   b ¹z  F   
   â r ÿ       â x      ÿ  â y      %   ¢ $r%ÿÿ   ÿ Ž Æ y      &   â r ÿ   pRð är ÿ   pRò â$x
     Ž â/ x   ÿ   ÈO r

   ÿ� Ê $€    Ž â xÿ     È |    pbô Ú G)       € ê y      '   "Ž $rÿÿ   ÿ Ž â xþÿÿÿÿÀŽ â $rÿÿ   ÿ Ž â x        â $rÿÿ   ÿ Ž â z q  ÿàÿ â 'z s   Ž È $x
   
Ž â zÿ t    Æ $x   ÿ Ž âx   ÿ   Æ $x#   ÿ Ž â x!   ÿàÿ â 'z y   Ž È $z }  ÿŽ â zÿ z    â $z p  Ž È $z	 |  Ž ä$z |  Ž â x   ÿàÿ â $z	 v  	Ž ä$z v  Ž ä $z p  	Ž ä %v X  Ž È %v X  Ž â x   ÿàù Æ $z w  ÿŽ â x   ÿà÷ â $r	ÿÿ   	 Æ r   ÿàõ ä xÿ     ä rÿ   ÿäÿ â $r.ÿÿ   ÿ â x     ä �yüÿÿ ™ ¢  %x   Ž Æ �yüÿÿ ™ è �y"    ™ h	 �y   ™ â %x&   Ž â r   ÿàõ Ä x $       â �y,øÿÿ ™ è �y &    ™ â  $r
ÿÿ   . Æ �y&   ™ â  %v Z   Ž â r   ÿà÷ Æ $x
$   ÿŽ â r
#   ÿàõ Æ $x   Ž â rÿ.   ÿäÿ â�y&øÿÿ ™ $ $x    ä %v Z   Ž â �y    ™ & $x$   ÿŽ â �y
   ™ " r#   ÿàõ Ä $x   Ž â �y$
   ™ ( r   ÿä â �y'
   ™ ( �y(   ™ ( �y*
   ™ ( �y+   ™ ( �y    ™ ( �y)   ™ ( �y-   ™ " x   ÿàõ Ä/ x@   ÿàÿ Æ $rÿÿ    â |    pbô ä x   ÿà÷ ä x@   ÿàÿ ä r	ÿ	   ÿäÿ ä rÿ     ÊO $x/   ÿ Ž â x
ÿ     È x//   ÿÀŽ ä rÿ!     ä x/"øŽ äx"ÿ   "  ä x
   ÿÀŽ â $x   ÿ Ž â x/ÿ     Ä x""øŽ â $x   ÿ Ž â xÿ     ä� x//   ÿÀŽ ä x
   ÿÀŽ ä x   øŽ â $x   ÿ Ž â x//øŽ Ä x
øŽ ä x   ÿ   ä xÿ     ä x   øŽ ä x   øŽ ä x   ÿ   Ä x0ÿ     ä x
"   øŽ â $x   ÿ Ž â x   øŽ â $x"   ÿ Ž â x//   0øŽ â $x   ÿ Ž â rÿ     ä x

   øŽ Ä x	   ÿ   â $x   ÿ Ž â x"/   "øŽ ä x   øŽ â $x   ÿ Ž â x/	   ÿ   ä rÿ!     ä x

   øŽ Ä x   øŽ ä x"   /øŽ â $x"   ÿ Ž â x   ÿÀŽ ä xÿ     ä x øŽ äx ÿ      Ä x   ÿÀŽ ä x/"   ÿÀŽ ä x1ÿ     ä x  øŽ ä x/øŽ ä x   ÿ   Ä x/ÿ     ä x0ÿ     ä x11   ÿÀŽ ä x   øŽ ä x     /øŽ â $x/   ÿ Ž âx   ÿ   Ä x010øŽ ä x3ÿ     â $x1   ÿ Ž â x   øŽ â $x   ÿ Ž â x     /øŽ ä x0   3øŽ ä x/   ÿ   Ä x   1øŽ ä x   øŽ ä x1   ÿ   ä x   /øŽ â &r/   ÿ   âx   1øŽ ä x3   ÿ   â &r/
$   /   Ä $x1   ÿ Ž â x   3øŽ â 2r&        â &r/'   /   ä x    1øŽ â &r1   ÿ   ä &r(   ÿ   ä &r/*   /   Ä &r(   ÿ   ä 0rÿ  0 A   ä$x3   ÿ Ž â Er/ /       â &r1$   1   ä &r
+      ä&r+      â  x   ?  A  â 0rÿ    A   â 2r,        â x   3øŽ â 2r&&        â&r1'   1   â 2r,        â &r)      Ä &r)      â #r/      â 0rÿ  0 A   ä &r1*   1   ä&r-      ä&r-      ä 0rÿ&  0 A   Ä 0rÿ  0 A   â  x   ?  A  â 0rÿ    A   â Er
 1       â 0r&ÿ&    A   â Er        â 0rÿ    A   â Er        â  x   ?  A  â  x   ?  A  â #r

      Ä #r&      â #r      â !r%%       â !r
       â !r       â !r       â G©  �õÿÿÿÿƒ ê Ay        € ê� ˆ % þÿ H   è ˆ €þÿ H   è ˆ  ÿÿ H   è ˆ €ÿÿ H   è {          ì M        € ê „y      H   " z  r  p`ð Æ „y  €   H   b r  ÿ   p"p Æ „y     H   ¨ „y     H   è „y	  €  H   ( „y     H   h „y     H   h „y  €  H   h „y     H   h „y
  €  H   h „y  €  H   b !r%       Æ „y  €  H   " !r       â/ !r       âO !r       ä� !r	       â!r       â!r       ä !r       Æ ‰      b !r       Æ ‰      ¢ !r

       Æ ‰      â !r
       Æ y      '   " !r       Ê ‰      " !r       â/ !r       ÈO ‰      b !r
       Æ� ‰      ¨ ‰	
      â !r       Ê ‰      " !r       â/ !r       ÈO ‰ €    b !r
	       Æ� ‰	 €    ¨ ‰ €    â !r       Ê ‰ €    " !r       â/ $z x  Ž â !r		       âO !r
       È� ‰	 @    h ‰ @    ¨ ‰
 @    â !r       Ê ‰ @    " !r	       â/ x   ÿàÿ ä x	        â !r       âO z  r  p`ò â ‰      â $z ~  Ž â!r
       â� x     p"ò â ‰      d 
  r  ÿàõ Ä ‰
      ¢ !r       ä $ÿÿ   ÿ â 
  j  ÿ„ È 
  k   â !r       â/ %v j  	 Ž â !r	       â !r
       âO ‰      h  †	     è † 	   è †	     â M™        € ê ¹z  r      â !r       â/ �x   ?àÿ Ì |   ÿàñ È rÿÿ   ÿä  ä z j  ÿ€ È z k    Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y       &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž ä $rÿÿ     Ž È % \   Ž Ê �	    ™ b ¹z  l      â Ey  �    € â ™x?    â y      "   b $rÿÿ   ÿ Ž Æ y      !   ¢  ÿ       â $ÿÿ    Ž È '
 m   Ž â ‚ ÿ       È 
ÿ n    Ê $ÿÿ   
Ž â r ÿ   pRò Æ/ $
 o  Ž â r ÿ   pRô â$x    Ž äO $‚ÿÿ    Ž Æ r   ÿ� â 'Š s   Ž ä y      '   " xÿ     È |    pbö ä Šÿ t    Ö G9       € ê y	      %   " $rÿÿ   ÿ Ž â xþÿÿÿÿÀŽ â $z w  ÿŽ ä 'z y   Ž È $x   
Ž â zÿ z    â $z }  ÿŽ ä $x   ÿ Ž âx   ÿ   â $z |  Ž â r   ÿàñ â $t
ÿ   ÿ Ž â xÿ     â $z v  Ž âx        â $rÿÿ   ÿ Ž â x   ÿàÿ â $rÿÿ   ÿ  â x     â $z	 p  Ž Ä %v  € 
Ž È %v X  Ž Ê �y   ™ ¢  r   ÿàñ Ê $x     â r
   ÿàñ ä x $       ä �y    ™ â $rÿÿ     ä %v

 Z   Ž â �y   ™ & $x	$   ÿŽ â r
   ÿ€ Æ $x   	Ž â �y    ™ f! $x	     ä �y
    ™ h �y   ™ h �y   ™ h �y   ™ h �y   ™ b x@   ÿàÿ È |    pbð ä x   ÿà÷ È rÿ   ÿäÿ ä rÿ     äOrÿ     ä x   ÿ   ä x ÿ     ä x   ÿÀŽ â $x"   ÿ Ž â x!    ÿÀŽ Ä xøŽ â�$x
   ÿ Ž â/ xÿ     ä x#"   ÿÀŽ ä x	!øŽ â $x   ÿ Ž â xÿ     äx##øŽ Ä x   
øŽ ä x$ÿ     ä xÿ     â $x!   ÿ Ž â x    ÿ   ä x   øŽ ä x%$   ÿÀŽ Ä x
	   øŽ ä x   ÿ   â $x   ÿ Ž â x	#    øŽ â $x   ÿ Ž â x%øŽ ä x#ÿ     ä x

   !øŽ Ä x   øŽ â $x   ÿ Ž â x   #øŽ ä x		   øŽ â &r   ÿ   âx!   ÿ   äx

   øŽ â $x   ÿ Ž â x   !øŽ Ä x		   øŽ â &r

      â x   øŽ â 2r        ä &r		   
   È &r   	   ä 0r	ÿ  0 A   ä0rÿ    A   â Er        ä  x
	   ?  A  È #r   
   È !r       â G‰  Pûÿÿÿÿƒ ê Ay        € ê ˆ €ÿÿ H   è {          ì M)        € ê „y     H   " r ÿ   pRð Æ „y €   H   h „y    H   ¨ y      %   â !r       È !r       È/ !r       ÊO ‰      $ !r       Ê‰      $ !r       Ê ‰ €    $ !r       Ê ‰ @    $ !r	       â $z  x  Ž È� ‰
	      b  M	        € ê $tÿ   ÿ Ž â !r		
       â? $z ~  Ž È %v j   Ž Ê †y 	    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž È % \   Ž Ê �	    ™ b $rÿÿ   ÿ Ž â Ey  ð    € â $r
ÿÿ   ÿ Ž â y       "   b 'Š s   Ž Æ y      !   b '
 m   Ž È $rÿÿ   ÿ Ž â 
ÿ n    ä rÿ    ÿü‚ â/ $x     Ž â Šÿ t    È r   ÿ� Î G       € ê z ÿ `  pPô ä z ÿ ^  pPò ä z ÿ b  p ö ä z ÿ a   Sô ä z ÿ _  Sò ä z ÿ c  0't Ö       '   "        ä¢       â ©
      '   d $ x  ÿŽ ä 	      %   ¢ $ª x  ÿŽ â r ÿ       Ä ©      %   â $ ~  ÿŽ ä $ª

 ~  ÿŽ Æ/    	àë äO	ÿ   	  ä ¢

   àÇ ä�¨ÿ     ä 	ÿÿ   	Äþ ä    ÿàý Ä ¢
   ÿàû ä ¢ÿÿ   „þ â $
ÿÿ   	 â  ^  ÿ† äª b  ÿˆ â $¢ÿÿ   Ž â  _  
� â $r
ÿÿ   ÿ Ž Æ ª c   â �     h �©
     d Ay        € ê ¹z  l      â xÿ     â ™x?    â $ÿÿ   
Ž â Ey  0    € â $rÿÿ   ÿ Ž Ä $
 o  Ž â|    pbò â $‚ÿÿ    Ž ä $rÿÿ   ÿ Ž Ô G  À
    € ê y      '   " $rÿÿ   ÿ Ž â xþÿÿÿÿÀŽ â $z	 w  ÿŽ â y      %   b $rÿÿ   ÿ Ž ä $x   
Ž ä $rÿÿ   ÿ Ž Æ x   ÿ   â 'z y   Ž È $z }  ÿŽ â zÿ z    È r	   ÿà÷ â $z |  Ž È $z v  Ž ä$rÿÿ   ÿŽ ä $z p  Ž È/ %x   ÿŽ Ê z `  ÿàó â%x   Ž â z X  ÿàõ ä z a  ÿäÿ  äx   ÿàñ ä z Y  ÿä ä z `  ÿàó â$rÿÿ     â z X  ÿàõ È z a  ÿäÿ  äz Y  ÿä È $rÿÿ    Ž â �yüÿÿ ™ ¨  �y"   ™ â r   ÿàñ â $tÿ$   ÿ Ž ä �y    ™ $ $rÿÿ     Ä %v Z   Ž â �y!    ™ f $x	$   ÿŽ Ê r	   ÿàÿ Ê %x   Ž â �y    ™ h �y   ™ ( �y   ™ ( �y   ™ ( �y   ™ " $x   ÿ Ž â z ÿ `  pPð È z ÿ a   Sð ä x   ÿàõ ä x@   ÿàÿ Æ $rÿÿ    â |    pbú ä r ÿ   "  Ê� $x#    ÿ Ž â xÿ     ÈO x$#   ÿÀŽ ä x#ÿ      ä x$$øŽ ä x##   ÿÀŽ ä x   ÿàÿ â $x    ÿ Ž â x#øŽ Ä x	ÿ      ä/ r"ÿ   "  â $x    ÿ Ž â x   	øŽ ä x$   øŽ â $x	"   ÿ Ž ä $x#    ÿ Ž â x   øŽ Ä x	   ÿÀŽ ä x%ÿ   "  ä x   #øŽ â $x#"   ÿ Ž â x$ÿ     äxøŽ ä x	%   ÿÀŽ Ä x    ÿ   â$x%    ÿ Ž â x#   #øŽ â $x"   ÿ Ž âx$	$øŽ ä x	ÿ   "  ä x   øŽ â $x "   ÿ Ž â x	$   	øŽ â $x"   ÿ Ž â x#   øŽ â &r   ÿ   â x   %øŽ â $x""   ÿ Ž â x		    øŽ Ä x   øŽ â &r      â x		   "øŽ â 2r!!        ä&r      â r        â $rÿÿ    Ž ä &r"	      ä 0r ÿ!  0 A   âx   ÿàó â 0r!ÿ!    A   â Er" "       â $rÿÿ    Ž â x   ÿà÷ â $r	ÿÿ    Ž â  x     ?  A  â x   ÿàù â $rÿÿ   Ž  Ä #r!!"       â $rÿÿ   	Ž â G‰  p    € ê �y"    ¨ �y$üÿÿ  è  �y      ( �y#     b rÿ   "  ÄO rÿ   "  ä xÿ     â $x%   ÿ Ž ä$x	   ÿ Ž â/ x   ÿÀŽ â $x"   ÿ Ž â x%   ÿÀŽ ä x%ÿ   $  â�2r#        âx$øŽ Ä x$%øŽ â $x   ÿ Ž â xÿ     ä x   	øŽ â $x	   ÿ Ž â x$$   øŽ ä x%   ÿ   â$x   ÿ Ž â x   	øŽ Ä x$   %øŽ ä x%   ÿÀŽ ä x	ÿ     ä xÿ      äx%% øŽ â $x    ÿ Ž â x		   ÿÀŽ Ä x   "øŽ ä x	øŽ ä x	ÿ     â &r   ÿ   â x %    øŽ â $x%   ÿ Ž âx   øŽ â $x   ÿ Ž âx   	øŽ â $x	   ÿ Ž âx     %øŽ â &r      â x   øŽ â 0rÿ  0 A   âx	   ÿ   â 0rÿ    A   â x	    	øŽ Ä x   øŽ Æ &r	      â  x	   ?  A  Æ &r      Ê Er        Ê #r   	   È !r       â $rÿÿ    â !r!       â GÙ  ÷ÿÿÿÿƒ ê Ay        € ê r  ÿ   pRð âEy  €     € â r  ÿ   pRò Ö G‰  P     € ê z ÿ `  pPð â $x      Ž Æ z ÿ a   Sð ä ˆs  €ÿÿ H   ö ˆ     H   ä Ay        € ê {          ì M        € ê z ÿ `  pPð â „y     H   f z ÿ a   Sð Ú „‰  €   H   ( „‰    H   ¢!r       È/ !‚         È !‚        âO G‰        € ê „y  €  H   ( „y €   H   h „y    H   ¨ „y	    H   è „y €  H   " !r         â !r       Æ/ !r         âO !r	       Æ� !r        ä‰      " z ÿ `  pPð ä r ÿ   pRò ä z ÿ a   Sð â !r       Ê ‰       $ !r         Ê ‰  €    $ !r        Ê ‰ @    $ !r       Ê ‰      $ !r       â G‰        € ê ‰       $ !r         Ê ‰       $ !r        Ê ‰ €    $ !r       Ê ‰ @    $ !r       Ê ‰      $ !r       ä M        € ê !r       â G‰  `    € ê $v ÿ h  ÿ Ž â !r
       È x     pBð Ú G	  p    € ê x  ÿÿÿÿÿàÿ È x       € Ê $x     ÿ Ž È ‚{   €     $ xÿ     È Iy °ìÿÿÿÿƒ ê  x ;ª¸¿  A  Ì s          $ !t   €?    Ì s        $  r     A  È  r     A  â Gy  P    € ê  x '7=  A  â x*BL?  A  Æ #t   €?    È  r      A  â  x    ?  A  Ê s     $   $ !t  €?    È  r      A  È  r     A  â Gy  °    € ê x     pRð Ú G‰  à     € ê x     pRð Ú G	        € ê 	z  i    � ä 	z i    � Æ  x ;ª¸¿  A  â 	z i €  � Ê s        $ !t  €?    Ì s        $  r      A  È  r     A  â Gy  Ð     € ê  r     A  â Gy  °     € ê 	x   à@  � ä 	x  à@  � Æ  x #ÛÙ?  A  â 	x  àÀ  � Æ  x;ª¸¿  A  ä !t  €?    È s        $ !t  €?    Ì s        $  r       A  È  r      A  ä y       &   ( y      %   ( y      '   b $z   x  Ž ä $tÿ   ÿ Ž Ä $z  ~   Ž È/ %v  j   Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y       &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž ä $rÿÿ     Ž È % \   Ž Ê �	    ™ b ¹z  l      â Ey  �    € â ™x?    â y      "   b $rÿÿ   ÿ Ž Æ y      !   ¢  ÿ       â $ÿÿ    Ž È '
 m   Ž â ‚ ÿ       È 
ÿ n    Ê $ÿÿ   
Ž â r ÿ   pRò Æ/ $
 o  Ž â r ÿ   pRô â$x    Ž äO $‚ÿÿ    Ž Æ r   ÿ� â 'Š s   Ž ä y      '   " xÿ     È |    pbö ä Šÿ t    Ö G9       € ê y	      %   " $rÿÿ   ÿ Ž â xþÿÿÿÿÀŽ â $z w  ÿŽ ä 'z y   Ž È $x   
Ž â zÿ z    â $z }  ÿŽ ä $x   ÿ Ž âx   ÿ   â $z |  Ž â r   ÿàñ â $t
ÿ   ÿ Ž â xÿ     â $z v  Ž âx        â $rÿÿ   ÿ Ž â x   ÿàÿ â $rÿÿ   ÿ  â x     â $z	 p  Ž Ä %v  € 
Ž È %v X  Ž Ê �y   ™ ¢  r   ÿàñ Ê $x     â r
   ÿàñ ä x $       ä �y    ™ â $rÿÿ     ä %v

 Z   Ž â �y   ™ & $x	$   ÿŽ â r
   ÿ€ Æ $x   	Ž â �y    ™ f! $x	     ä �y
    ™ h �y   ™ h �y   ™ h �y   ™ h �y   ™ b x@   ÿàÿ È |    pbð ä x   ÿà÷ È rÿ   ÿäÿ ä rÿ     äOrÿ     ä x   ÿ   ä x ÿ     ä x   ÿÀŽ â $x"   ÿ Ž â x!    ÿÀŽ Ä xøŽ â�$x
   ÿ Ž â/ xÿ     ä x#"   ÿÀŽ ä x	!øŽ â $x   ÿ Ž â xÿ     äx##øŽ Ä x   
øŽ ä x$ÿ     ä xÿ     â $x!   ÿ Ž â x    ÿ   ä x   øŽ ä x%$   ÿÀŽ Ä x
	   øŽ ä x   ÿ   â $x   ÿ Ž â x	#    øŽ â $x   ÿ Ž â x%øŽ ä x#ÿ     ä x

   !øŽ Ä x   øŽ â $x   ÿ Ž â x   #øŽ ä x		   øŽ â &r   ÿ   âx!   ÿ   äx

   øŽ â $x   ÿ Ž â x   !øŽ Ä x		   øŽ â &r

      â x   øŽ â 2r        ä &r		   
   È &r   	   ä 0r	ÿ  0 A   ä0rÿ    A   â Er        ä  x
	   ?  A  È #r   
   È !r       â G‰  Pûÿÿÿÿƒ ê Ay        € ê ˆ €ÿÿ H   è {          ì M)        € ê „y     H   " r ÿ   pRð Æ „y €   H   h „y    H   ¨ y      %   â !r       È !r       È/ !r       ÊO ‰      $ !r       Ê‰      $ !r       Ê ‰ €    $ !r       Ê ‰ @    $ !r	       â $z  x  Ž È� ‰
	      b  M	        € ê $tÿ   ÿ Ž â !r		
       â? $z ~  Ž È %v j   Ž Ê †y 	    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž È % \   Ž Ê �	    ™ b $rÿÿ   ÿ Ž â Ey  ð    € â $r
ÿÿ   ÿ Ž â y       "   b 'Š s   Ž Æ y      !   b '
 m   Ž È $rÿÿ   ÿ Ž â 
ÿ n    ä rÿ    ÿü‚ â/ $x     Ž â Šÿ t    È r   ÿ� Î G       € ê z ÿ `  pPô ä z ÿ ^  pPò ä z ÿ b  p ö ä z ÿ a   Sô ä z ÿ _  Sò ä z ÿ c  0't Ö       '   "        ä¢       â ©
      '   d $ x  ÿŽ ä 	      %   ¢ $ª x  ÿŽ â r ÿ       Ä ©      %   â $ ~  ÿŽ ä $ª

 ~  ÿŽ Æ/    	àë äO	ÿ   	  ä ¢

   àÇ ä�¨ÿ     ä 	ÿÿ   	Äþ ä    ÿàý Ä ¢
   ÿàû ä ¢ÿÿ   „þ â $
ÿÿ   	 â  ^  ÿ† äª b  ÿˆ â $¢ÿÿ   Ž â  _  
� â $r
ÿÿ   ÿ Ž Æ ª c   â �     h �©
     d Ay        € ê ¹z  l      â xÿ     â ™x?    â $ÿÿ   
Ž â Ey  0    € â $rÿÿ   ÿ Ž Ä $
 o  Ž â|    pbò â $‚ÿÿ    Ž ä $rÿÿ   ÿ Ž Ô G  À
    € ê y      '   " $rÿÿ   ÿ Ž â xþÿÿÿÿÀŽ â $z	 w  ÿŽ â y      %   b $rÿÿ   ÿ Ž ä $x   
Ž ä $rÿÿ   ÿ Ž Æ x   ÿ   â 'z y   Ž È $z }  ÿŽ â zÿ z    È r	   ÿà÷ â $z |  Ž È $z v  Ž ä$rÿÿ   ÿŽ ä $z p  Ž È/ %x   ÿŽ Ê z `  ÿàó â%x   Ž â z X  ÿàõ ä z a  ÿäÿ  äx   ÿàñ ä z Y  ÿä ä z `  ÿàó â$rÿÿ     â z X  ÿàõ È z a  ÿäÿ  äz Y  ÿä È $rÿÿ    Ž â �yüÿÿ ™ ¨  �y"   ™ â r   ÿàñ â $tÿ$   ÿ Ž ä �y    ™ $ $rÿÿ     Ä %v Z   Ž â �y!    ™ f $x	$   ÿŽ Ê r	   ÿàÿ Ê %x   Ž â �y    ™ h �y   ™ ( �y   ™ ( �y   ™ ( �y   ™ " $x   ÿ Ž â z ÿ `  pPð È z ÿ a   Sð ä x   ÿàõ ä x@   ÿàÿ Æ $rÿÿ    â |    pbú ä r ÿ   "  Ê� $x#    ÿ Ž â xÿ     ÈO x$#   ÿÀŽ ä x#ÿ      ä x$$øŽ ä x##   ÿÀŽ ä x   ÿàÿ â $x    ÿ Ž â x#øŽ Ä x	ÿ      ä/ r"ÿ   "  â $x    ÿ Ž â x   	øŽ ä x$   øŽ â $x	"   ÿ Ž ä $x#    ÿ Ž â x   øŽ Ä x	   ÿÀŽ ä x%ÿ   "  ä x   #øŽ â $x#"   ÿ Ž â x$ÿ     äxøŽ ä x	%   ÿÀŽ Ä x    ÿ   â$x%    ÿ Ž â x#   #øŽ â $x"   ÿ Ž âx$	$øŽ ä x	ÿ   "  ä x   øŽ â $x "   ÿ Ž â x	$   	øŽ â $x"   ÿ Ž â x#   øŽ â &r   ÿ   â x   %øŽ â $x""   ÿ Ž â x		    øŽ Ä x   øŽ â &r      â x		   "øŽ â 2r!!        ä&r      â r        â $rÿÿ    Ž ä &r"	      ä 0r ÿ!  0 A   âx   ÿàó â 0r!ÿ!    A   â Er" "       â $rÿÿ    Ž â x   ÿà÷ â $r	ÿÿ    Ž â  x     ?  A  â x   ÿàù â $rÿÿ   Ž  Ä #r!!"       â $rÿÿ   	Ž â G‰  p    € ê �y"    ¨ �y$üÿÿ  è  �y      ( �y#     b rÿ   "  ÄO rÿ   "  ä xÿ     â $x%   ÿ Ž ä$x	   ÿ Ž â/ x   ÿÀŽ â $x"   ÿ Ž â x%   ÿÀŽ ä x%ÿ   $  â�2r#        âx$øŽ Ä x$%øŽ â $x   ÿ Ž â xÿ     ä x   	øŽ â $x	   ÿ Ž â x$$   øŽ ä x%   ÿ   â$x   ÿ Ž â x   	øŽ Ä x$   %øŽ ä x%   ÿÀŽ ä x	ÿ     ä xÿ      äx%% øŽ â $x    ÿ Ž â x		   ÿÀŽ Ä x   "øŽ ä x	øŽ ä x	ÿ     â &r   ÿ   â x %    øŽ â $x%   ÿ Ž âx   øŽ â $x   ÿ Ž âx   	øŽ â $x	   ÿ Ž âx     %øŽ â &r      â x   øŽ â 0rÿ  0 A   âx	   ÿ   â 0rÿ    A   â x	    	øŽ Ä x   øŽ Æ &r	      â  x	   ?  A  Æ &r      Ê Er        Ê #r   	   È !r       â $rÿÿ    â !r!       â GÙ  ÷ÿÿÿÿƒ ê Ay        € ê r  ÿ   pRð âEy  €     € â r  ÿ   pRò Ö G‰  P     € ê z ÿ `  pPð â $x      Ž Æ z ÿ a   Sð ä ˆs  €ÿÿ H   ö ˆ     H   ä Ay        € ê {          ì M        € ê z ÿ `  pPð â „y     H   f z ÿ a   Sð Ú „‰  €   H   ( „‰    H   ¢!r       È/ !‚         È !‚        âO G‰        € ê „y  €  H   ( „y €   H   h „y    H   ¨ „y	    H   è „y €  H   " !r         â !r       Æ/ !r         âO !r	       Æ� !r        ä‰      " z ÿ `  pPð ä r ÿ   pRò ä z ÿ a   Sð â !r       Ê ‰       $ !r         Ê ‰  €    $ !r        Ê ‰ @    $ !r       Ê ‰      $ !r       â G‰        € ê ‰       $ !r         Ê ‰       $ !r        Ê ‰ €    $ !r       Ê ‰ @    $ !r       Ê ‰      $ !r       ä M        € ê !r       â G‰  `    € ê $v ÿ h  ÿ Ž â !r
       È x     pBð Ú G	  p    € ê x  ÿÿÿÿÿàÿ È x       € Ê $x     ÿ Ž È ‚{   €     $ xÿ     È Iy °ìÿÿÿÿƒ ê  x ;ª¸¿  A  Ì s          $ !t   €?    Ì s        $  r     A  È  r     A  â Gy  P    € ê  x '7=  A  â x*BL?  A  Æ #t   €?    È  r      A  â  x    ?  A  Ê s     $   $ !t  €?    È  r      A  È  r     A  â Gy  °    € ê x     pRð Ú G‰  à     € ê x     pRð Ú G	        € ê 	z  i    � ä 	z i    � Æ  x ;ª¸¿  A  â 	z i €  � Ê s        $ !t  €?    Ì s        $  r      A  È  r     A  â Gy  Ð     € ê  r     A  â Gy  °     € ê 	x   à@  � ä 	x  à@  � Æ  x #ÛÙ?  A  â 	x  àÀ  � Æ  x;ª¸¿  A  ä !t  €?    È s        $ !t  €?    Ì s        $  r       A  È  r      A  ä y       &   ( y      %   ( y      '   b $z   x  Ž ä $tÿ   ÿ Ž Ä $z  ~   Ž È/ %v  j   Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y       &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž ä $rÿÿ     Ž È % \   Ž Ê �	    ™ b ¹z  l      â Ey  `    € â ™x?    â y      "   b x      ÿ  â $rÿÿ   ÿ Ž ä y      !   ¢ $r$ÿÿ   ÿ Ž ä $ÿÿ   ÿ Ž â y      %   " $ÿÿ    Ž È '
 m   Ž È $‚ÿÿ   ÿ Ž â 
ÿ n    Ê $ÿÿ   
Ž â r ÿ   pRô ä/r ÿ   pRò â$
 o  Ž ä $x    Ž äO $‚ÿÿ    Ž ä $x   ÿ Ž â r   ÿ� â 'Š s   Ž Æ xÿ     â $(€   Ž Æ |    pbö ä Šÿ t    Ö G9  °    € ê y      '   " $rÿÿ   ÿ Ž â xþÿÿÿÿÀŽ â $z w  ÿŽ ä $v%ÿ X  ÿ Ž ä $x   
Ž ä $vÿ p  ÿ Ž ä $x   ÿ Ž Ä $x    ÿ Ž ä $vÿ Y  ÿ Ž ä $rÿÿ   ÿ Ž â x"    ÿàÿ â 'z y   Ž È $z }  ÿŽ â z	ÿ z    â $z p  Ž Æ r   ÿàñ â $z		 |  Ž ä %x   ÿŽ È $z v  	Ž âxÿ     â $r,ÿÿ   ÿ  â x!   ÿüŽ â %x   ÿŽ â x#     Ê %x   Ž È r
%   ÿàñ Ê $x     â r%   ÿàñ È �y)
   ™ ¢  $x     â r   ÿàñ â $tÿ$   ÿ Ž â �y1
    ™ è  �y(   ™ " $rÿÿ   ,  Æ �y'   ™ b %v Z   Ž È $x	$   ÿŽ â r   ÿˆ ä x   
à± â $x   	Ž Æ r	ÿ   d~  â $x   # â r   ÿàù â �y    ™ è  �y&   ™ â  $x    Æ �y-    ™ è  �y+   ™ è  �y.   ™ è  �y*   ™ è  �y0   ™ è  �y/   ™ â  r!   ÿàñ Ê/ $x   	  â r   ÿàñ È �y3üÿÿ ™ â $x     Ê �y4   ™ è �y2   ™ â xüÿÿÿà± Æ �yøÿÿ ™ â  xÿÿÿÿ	d~  är
!   ÿàñ â �y	üÿÿ ™ è $x     â r   ÿàñ È �y
    ™ â  $x     Æ �y
üÿÿ ™ è  �y   ™ è  �y   ™ â  x@   ÿàÿ È |    pbð ä x%%   ÿà÷ ä rÿ"   )  äOr)ÿ    )  Æ $x5   ÿ Ž ä $x)   ÿ Ž â/ x
ÿ     ä x55   ÿÀŽ ä xÿ   )  ä x   ÿÀŽ ä x55(øŽ Äx(ÿ   (  ä x
   ÿÀŽ ä x   ÿÀŽ â $x)   ÿ Ž â x
ÿ   '  äx'øŽ ä x(øŽ â $x   ÿ Ž â x

øŽ Ä xÿ   )  ä x   øŽ â $x)   ÿ Ž â x55   øŽ â $x)   ÿ Ž âx

   øŽ â $x)   ÿ Ž â x'ÿ     ä x   øŽ â $x   ÿ Ž â x   'øŽ â $x'   ÿ Ž â x

   øŽ Ä x   øŽ ä x))	   ÿ   â $x   ÿ Ž âx5   øŽ â $x   ÿ Ž â x)
   )øŽ â &r
&   ÿ   â� x   'øŽ ä rÿ"   -  Ä r-ÿ    -  â &r))+   
   â x   øŽ ä x   øŽ â $x   ÿ Ž â xÿ     â $x-   ÿ Ž ä &r
.   )   â xÿ   *  Ä x   ÿÀŽ ä x'   ÿÀŽ ä x   ÿÀŽ ä xøŽ ä xÿ   -  â &r
0   
   â x'*øŽ â $x(   ÿ Ž â x/øŽ Ä xÿ   /  ä x'   ÿÀŽ â $x-   ÿ Ž â x)ÿ     ä x'øŽ â $x'-   ÿ Ž â x   (øŽ â $x(   ÿ Ž â x   øŽ Ä x   )øŽ ä x)ÿ   -  ä x   'øŽ â $x'-   ÿ Ž â x   (øŽ â $x-   ÿ Ž â x   )øŽ â 2r11        â $x*   ÿ Ž Ä x(   'øŽ ä x'   øŽ â 0rÿ1  0 A   âx   *øŽ â $x--   ÿ Ž â Er
 
       â &r'&   ÿ   â  x   ?  A  â $x/   ÿ Ž ä 0rÿ1    A   Ä $x'   ÿ Ž â rÿ"   3  ä x((   -øŽ â #r
      â x   /øŽ â $x
   ÿ Ž â r3ÿ    3  â &r(+      â x'   'øŽ ä xÿ     Ä x)
   ÿÀŽ â $x
3   ÿ Ž â x-   ÿÀŽ â &r.      â xÿ   4  äxÿ   3  ä x)4øŽ ä x)
   ÿÀŽ Ä x
-øŽ ä xÿ   2  äx-   ÿÀŽ â $x3   ÿ Ž â x))2øŽ ä x-øŽ ä x-ÿ   3  â $x/   ÿ Ž â x)   øŽ â $x)3   ÿ Ž âx   -øŽ â $x-3   ÿ Ž â x   /øŽ Ä x/ÿ     ä x   )øŽ ä x)3   ÿ   äx   -øŽ â $x-   ÿ Ž â x

   /øŽ â $x/   ÿ Ž â x)   )øŽ â $x33   ÿ Ž â x   -øŽ â $x-   ÿ Ž â x/
   /øŽ â &r
)&   ÿ   â x   3øŽ â $x   ÿ Ž â rÿ      â &r''0      â x   -øŽ â &r)+   
   â x/   øŽ â $x   ÿ Ž â rÿ"     â &r
.   )   â xÿ     Ä x)   ÿÀŽ â $x   ÿ Ž â x-   ÿÀŽ â $x   ÿ Ž â x))øŽ äxÿ     ä x   ÿÀŽ ä x(ÿ     Ä x)   øŽ ä x-øŽ ä x)ÿ     â $x-   ÿ Ž â xøŽ äxÿ     â $x   ÿ Ž â x/(   ÿÀŽ Ä x   )øŽ â $x)   ÿ Ž âx//øŽ â $x   ÿ Ž â x   -øŽ â $x-   ÿ Ž â x   øŽ ä xÿ     ä x   )øŽ â $x)   ÿ Ž â x   øŽ â 2r		        â x/   øŽ â $x/   ÿ Ž âx   -øŽ â $x-   ÿ Ž â x   )øŽ â &r&&   ÿ   â x/   /øŽ â $x   ÿ Ž â x   -øŽ â 0rÿ	  0 A   Ä &r++   &   â Er '       â 0r	ÿ	    A   â x/   øŽ â 2r        â x   ?  A  â &r+.   +   â 2r        â &r
0   
   â #r		      â &r0   +   Ä 0rÿ  0 A   ä0rÿ  0 A   âEr
 
       â 0rÿ    A   â Er        â 0rÿ    A   â  x   ?  A  â  x   ?  A  Æ #r
      â #r      â !r$$       â !r	       â !r       â !r       â $rÿÿ   Ž â D	       À â Gy  ðïÿÿÿÿƒ ê Ay        € ê ˆ# $ þÿ H   è ˆ# €þÿ H   è ˆ#  ÿÿ H   è ˆ# €ÿÿ H   è {          ì M        € ê „y     H   " z  r  p`ô â $z   x  Ž Ä „y €   H   b r ÿ   pVt Æ „y    H   ¨„y    H   è „y €  H   ( „y    H   ( „y    H   ( „y €  H   ( „y	 €  H   ( „y    H   ( „y €  H   " !r$       Æ „y
 €  H   " !r       â/ !r       âO !r       ä� !r       â!r       â !r       â !r       È ‰      b !r		       â !r       È ‰	      ¢ !r       Æ ‰      â !r

       Ê ‰
      " !r       â/ !r	       ÈO ‰      b !r       Æ� ‰      ¨ ‰      â !r
       Ê ‰	      " !r       â/ !r       ÈO ‰ €    b !r       Æ� ‰ €    ¨ ‰
 €    â !r	       Ê ‰ €    " !r       â/ !r       ÈO ‰ @    b !r

       â� x   ÿàÿ ä ‰ @    ¤ z  r  p`ò ä ‰	
 @    â !r       â $tÿ   ÿ Ž â x    pVò  â y      '   ( ‰ @    " !r       â/ x   ÿàÿ â!r       âO x   ÿàÿ Ä ‰      b z  r  p`ö â !r	
	       â� z  r  p`ð â ‰      ¢ x    pVö â$z  ~   Ž â x    pVp  â ‰	      " !r       Ê‰
      â !r       â/ %v  j   Ž â !r       ÈO †©     â !r		       Æ †¹    è †‰ 	   â !r
       â� M        € ì †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú         Ê % \   Ž Ê �		    ™ ¢  r ÿ       â Ey       € â x&      ÿ  â y      "   d '
  m   Ž äy.      %   ä 'Š s   Ž â 
ÿ n     â y/      !   ( y      '   " r ÿ   pRò Ä/ x..   ÿ   ä� x /   pDò  ä r./   ÿàÿ är /   ÿ(Ž ä z  r  pdò  ä ÿ   ÿáÿ Ä r
     ÿ� Æ $
 o  Ž â Š	ÿ t    È r	       âO G  °    € ê z ÿ `  pPô âx&      ÿ  â z ÿ ^  pPò ä z ÿ b  p ö ä z ÿ a   Sô ä z ÿ _  Sò ä z ÿ c  0't Ö $ ~  ÿŽ âÿ   .  â$ª ~  ÿŽ ä $ x  ÿŽ â.   /àë â$ª x  ÿŽ â ¢.   /àÇ ä ¨ÿ   .  ä ÿ   ÿÄþ Ä    ÿàý ä ¢   ÿàû ä ¢ÿ   ÿ„þ ä ÿ   ÿä ä  ^  ÿ† ä ¢ÿ   ÿäÿ Ä ª b  ÿˆ ä  _  � ä ª c   Æ �&     h �©'     d Ay        € ê¹z  l      â xÿ   
  â ™x?    â ‚        â Ey  0!    € â x      ÿ  â x      ÿ  â x      ÿ  â |    pbð â x      ÿ  Ø G	  À     € ê r ÿ       â $z. p  Ž â x
þÿÿÿÿÀŽ â $z w  ÿŽ â z ÿ `  pPð â 'z y   Ž âr   €ÿàÿ â x      ÿ  â z
  `      â $z }  ÿŽ â zÿ z    â x      ÿ  â x   ÿ   â x      ÿ  â r   ÿàó â $z |  Ž âx   ÿ   â $z |  Ž â z  a      â $z	 v  Ž â z  X      â %x   ÿŽ â z ÿ a   Sð â x      ÿ  â rÿÿ   ÿäÿ  â $z	 v  Ž â z	  Y      â %x   Ž â x   ÿàÿ Æ %x   Ž â r(   ÿàó È �y5   ™ ¢  r)	   ÿäÿ  Ê �y3(   ™ â r   ÿàó ä x $       â �y2(   ™ " r ÿ   ÿäÿ  Æ %v Z   Ž â z7  p      Æ $x $   ÿŽ ä %x 7   Ž ä �y    ™ " r   ÿàÿ Æ �y9    ™ $ %x"   Ž ä �y1    ™ $ %x$7   (Ž ä�y"   ™ ( �y"   ™ ( �y4$   ™ ( �y "   ™ ( �y0"   ™ ( �y6$   ™ " %x(7    Ž Æ/ �y      ™ ( �y;(   ™ " %x,7   $Ž Æ �y$(    ™ ( �y:,   ™ ( �y8,   ™ " %x*7   (Ž Ê �y=*   ™ " %x"7   ,Ž Æ �y<*    ™ ( �y"   ™ ( �y"   ™ " rÿ   5  ÈO x7ÿ     ä x!   ÿ   ä x%ÿ   3  ä� x77   ÿÀŽ ä x,!   ÿÀŽ ä x%7%øŽ Ä x"ÿ     ä x,,3øŽ ä x!   ÿ   ä x%%   "øŽ ä x!,   !øŽ ä x"   ÿ   Ä x(   ÿ   ä/ r5ÿ   5  ä x"!   "øŽ ä x!%   (øŽ ä x#5   ÿ   ä x(ÿ   5  Ä x%#   ÿÀŽ ä x)ÿ   2  äx*(   ÿÀŽ ä x%%2øŽ ä x(5   ÿ   ä x)*)øŽ Ä x,ÿ   5  ä x#   ÿ   ä x%%   (øŽ ä x*5   ÿ   ä x))   ,øŽ ä x(	   ÿ   Ä x,5   ÿ   ä x""   #øŽ ä x%%   *øŽ ä x!!   (øŽ ä x#)   ,øŽ ä x*5	   ÿ   Ä x(5   ÿ   ä rÿ   9  â &r""   ÿ   â x##   *øŽ ä x%%   (øŽ ä x*ÿ     ä x(   ÿ   â &r!!   "   â x"ÿ   4  Ä x+*   ÿÀŽ ä x)(   ÿÀŽ ä r9ÿ   9  â &r!%    !   â x++"øŽ ä x))4øŽ ä x"   ÿ   Ä x%9   ÿ   â &r##0   !   â x(ÿ     ä x")   "øŽ ä x!   ÿ   ä x*ÿ   9  ä x)%   ÿÀŽ Ä x(+   (øŽ ä x%   ÿ   ä x+*   ÿÀŽ ä x!"   !øŽ ä x))6øŽ ä x*9   ÿ   Ä x"   ÿ   ä x6ÿ   6  ä x%(   %øŽ ä x))   *øŽ ä x(	   ÿ   ä x"!   "øŽ Ä x*9   ÿ   ä x++6øŽ ä x,ÿ   9  â 2r1        â x%%   (øŽ â &r"   ÿ   â x))   *øŽ ä x++   ,øŽ Ä x*9   ÿ   äx(9   ÿ   â &r%%      â x*+   *øŽ â 0r!ÿ  0 A   âx99	   ÿ   â 0rÿ    A   â x))   (øŽ ä rÿ   ;  â  x(!   ?  A  â x**   9øŽ â &r%)    %   â Er" #       â 2r  1        â x!   ÿ   â &r%*0   %   Ä #r"   (   â x#!   ÿÀŽ â 0r!ÿ   0 A   âx"ÿ     â 0r ÿ     A   â Er% %       ä x)"   ÿÀŽ â  x!!   ?  A  â x##:øŽ ä x"   ÿ   Ä x:ÿ   :  â #r  %   !   â x"#   "øŽ ä x!   ÿ   ä x)):øŽ ä x(ÿ     ä x!"   !øŽ Ä x()   (øŽ ä x#   ÿ   är;ÿ   ;  ä x"   ÿ   ä x#(   #øŽ ä x%;   ÿ   Ä x	   ÿ   ä x"!   "øŽ ä x!ÿ   ;  ä x%%   ÿÀŽ ä x#   øŽ ä x#ÿ   8  Ä x(!   ÿÀŽ â &r!"   ÿ   â x%%8øŽ ä x";   ÿ   ä x((#øŽ ä x)ÿ   ;  â &r!   !   â x"%   "øŽ Ä x#;   ÿ   äx((   )øŽ ä x%;   ÿ   ärÿ   =  ä x""   #øŽ ä x#;   ÿ   Ä x%(   %øŽ ä x)   ÿ   ä x(;	   ÿ   ä x""   #øŽ ä x*)   ÿÀŽ ä x#%   (øŽ Ä r=ÿ   =  ä x%ÿ     ä x**øŽ äx)=   ÿ   ä xÿ     ä x+ÿ   =  Ä x(%   ÿÀŽ ä x))   ÿÀŽ ä x%ÿ     ä x,+   ÿÀŽ ä x((øŽ ä x   ÿ   Ä x))øŽ ä x+,%øŽ ä x%ÿ     ä x*   øŽ ä x   ÿ   ä x*ÿ   =  Ä x,=   ÿ   ä x%(   %øŽ ä x   øŽ ä x(   ÿ   äx   ÿ   ä x++   *øŽ Ä x))   ,øŽ ä x*=   ÿ   ä x%%   (øŽ ä x   øŽ ä x	   ÿ   ä x))   *øŽ â &r   ÿ   â x*=   ÿ   Ä x(=   ÿ   ä x%   øŽ ä x*+   *øŽ ä x()   (øŽ â &r      â x==	   ÿ   â &r!"    !   â2r$$1        â&r(       â x**   =øŽ â 2r<<1        â &r!#0   !   Ä 0rÿ$  0 A   ä &r**0      ä 0rÿ<  0 A   â Er! !       â 0r$ÿ$    A   â Er# *       â 0r<ÿ<    A   â  x   ?  A  â  x%   ?  A  â !r        â !r       â #r!$!      â #r <#   %   â G‰       € ê %x   
Ž â r
   ÿàó È �y%    ¢ r   ÿäÿ  Æ �y"     è  �y(    ( �y)    " r#ÿ   %  ÈO x*#   ÿ   ä r$ÿ   %  ä x%*   ÿÀŽ ä x*ÿ   (  äx(%(øŽ ä x%$   ÿ   Ä xÿ   $  ä/ x%   ÿÀŽ ä xÿ   )  äx   ÿÀŽ ä x)øŽ ä x$   ÿ   Ä x+ÿ   #  ä xøŽ ä xÿ   $  ä x   øŽ ä x++   ÿÀŽ ä x$   ÿ   Ä x   øŽ ä x*+*øŽ ä x%ÿ   #  ä x$   ÿ   ä x   øŽ ä x+#   ÿ   Ä x$   ÿ   ä x**   %øŽ ä x   øŽ ä x((   +øŽ ä x%#   ÿ   ä x$	   ÿ   Ä x   øŽ ä x%(   %øŽ ä x)#   ÿ   â&r   ÿ   â x   øŽ ä x$#   ÿ   äx)*   )øŽ â &r      â x#	   ÿ   Ä x$%   $øŽ â 2r""1        â� x)   øŽ Æ &r$       È &r0      ä 0rÿ"  0 A   ä0r"ÿ"    A   â Er        ä  x   ?  A  È #r"      È !r       ä G‰  Ð    € ê x.   ÿàÿ Ê $z p  Ž È %x   
Ž È %x   Ž â �y%    ¨ r
   ÿàó â �y"     æ  r   ÿäÿ  Ê �y)    ( �y(    â r#ÿ   %  ÈO x$#   ÿ   È x+$   ÿÀŽ ä r$ÿ   %  È x%$   ÿ   ä xÿ   $  ä/ x%   ÿÀŽ ä x   ÿÀŽ ä xÿ   )  äx)øŽ Ä x$   ÿ   ä xøŽ ä x*#   ÿ   ä x++(øŽ ä� xÿ   $  ä x   øŽ Ä x,ÿ   #  ä x$   ÿ   ä x*+   *øŽ ä x   øŽ ä x+,   ÿÀŽ ä x(ÿ   (  Ä x$   ÿ   ä x   øŽ ä x$   ÿ   ä x(+(øŽ ä x)ÿ   #  ä x   øŽ Ä x%#   ÿ   äx$	   ÿ   ä x   øŽ ä x((   )øŽ ä x%*   %øŽ â &r   ÿ   â x)#   ÿ   Ä x   øŽ ä x$#   ÿ   äx((   )øŽ â &r      â x##	   ÿ   ä x$%   $øŽ â 2r""1        â x#(   #øŽ Æ &r$       ä 0rÿ"  0 A   ä&r#0      ä 0r"ÿ"    A   â  x   ?  A  ä Er        Ê #r""      È !r"       ä G‰  Ð    € ê x.   ÿàÿ Ê $z p  Ž È %x   
Ž È %x   Ž â �y%    ¨ r
   ÿàó â �y"     æ  r   ÿäÿ  Ê �y(    ( �y)    â r#ÿ   %  ÈO x$#   ÿ   È x+$   ÿÀŽ ä r$ÿ   %  ä x,ÿ   #  ä x%$   ÿ   ä x++(øŽ äxÿ   $  Ä/ x,   ÿÀŽ ä x(ÿ   (  ä x%   ÿÀŽ ä x   ÿÀŽ ä xÿ   )  ä� x((øŽ Ä x)øŽ ä x$   ÿ   ä xøŽ ä xÿ   $  ä x   øŽ ä x$   ÿ   Ä x   øŽ ä x$   ÿ   ä x   øŽ ä x*#   ÿ   ä x$   ÿ   ä x)ÿ   #  Ä x   øŽ ä x*+   *øŽ ä x%#   ÿ   ä x$	   ÿ   ä x   øŽ ä x((   )øŽ Ä x%*   %øŽ â &r   ÿ   â x)#   ÿ   äx   øŽ ä x$#   ÿ   äx((   )øŽ â &r      â x##	   ÿ   Ä x$%   $øŽ â 2r""1        â x#(   #øŽ Æ &r$       ä 0rÿ"  0 A   ä&r#0      ä 0r"ÿ"    A   â  x   ?  A  ä Er        Ê #r"      È !r       â !r!       â !r        â G‰  Ð    € ê x.   ÿàÿ Ê $z p  Ž È %x   
Ž È %x   Ž â �y!    ¨ r
   ÿàó â �y      æ  r   ÿäÿ  Ê �y$    ( �y#    â r"ÿ   !  äOr!ÿ   !  ä x%"   ÿ   ä x*!   ÿ   Ä x%%   ÿÀŽ ä x**   ÿÀŽ ä x)ÿ   "  ä x%%$øŽ äx!   ÿ   ä x**#øŽ Ä� x))   ÿÀŽ ä x$ÿ   $  ä xÿ   !  ä x("   ÿ   ä x**   øŽ ä x$)$øŽ Ä x   ÿÀŽ ä xÿ   "  ä x%%   (øŽ ä x"   ÿ   ä x#ÿ   #  ä x$$   øŽ Ä x%   øŽ ä x"   ÿ   äx"   ÿ   ä x#øŽ ä x%ÿ   !  ä x#!   ÿ   Ä x$   øŽ ä x""	   ÿ   ä x   øŽ ä x%   %øŽ ä x#*   #øŽ â &r   ÿ   â x$!   ÿ   Ä x   "øŽ ä x!   ÿ   ä x$%   $øŽ â &r      â x!!	   ÿ   ä x#   øŽ â 2r  1        â x!$   !øŽ Æ &r       ä 0r ÿ   0 A   ä&r!0      ä 0r ÿ     A   â  x    ?  A  ä Er        Ê #r        È !r        â x@   ÿàÿ ä x

   ÿàõ ä |    pbò ä x   ÿà÷ ä rÿ   ÿä ä r	ÿ	   ÿäÿ Î D       À â Gy  áÿÿÿÿƒ ê Ay        € ê y       "   " Ey  0    € â r  ÿ   pRð är  ÿ   pRò Ö G‰  ð     € ê z ÿ `  pPð ä r  /   ÿ8Ž ä z ÿ a   Sð Æ ˆs   þÿ H   ô ˆƒ  €þÿ H   è ˆƒ   ÿÿ H   è ˆƒ  €ÿÿ H   â G‰  p     € ê ˆs     H   è ˆs  €þÿ H   è ˆs  €  H   è ˆs   ÿÿ H   è ˆs     H   è ˆs  €ÿÿ H   è ˆs  €  H   ä Ay        € ê {          ì M        € ê y      &   ¢ z ÿ `  pPð Æ y      '   â z ÿ a   Sð â $z  x  .Ž â x        Æ $z  ~   Ž È� %v  j   Ž È G	  P    € ê „y/     H   " z . r  p`ð Æ „y/ €   H   b r /ÿ   pVp  Æ „y/    H   ¨ „y /    H   è „y	/ €  H   ( „y/    H   ( „y/    H   ( „y/ €  H   ( „y/    H   ( „y
/ €  H   ( „y/ €  H   " !r       Æ „y/ €  H   " !r       â/ !r       âO !r         ä� !r	       â!r       â !r         ä !r       Æ ‰       b !r       Æ ‰      ¢ !r

       Æ ‰      â !r

       È !r
       â !r        â/ !r       ÈO ‰      " !r
       Æ� ‰      h ‰      ¨ ‰	
      â !r       â !r       â/ !r       ÈO ‰      " !r
	       Æ� ‰ €    h ‰  €    ¨ ‰	 €    â !r       â !r       È/ ‰ €    " !r
        ÆO ‰  @    b !r		       â� x.   ÿàÿ ä ‰
 @    ¤ z  r  p`ô ä ‰	 @    ä x /   pVt â !r       â !r         È/ ‰ @    " x.   ÿàÿ â !r
       âO x..   ÿàÿ â ‰       b z  r  p`ò â !r	       ä� ‰      ¢ x /   pVò  Æ ‰      â !r       â !r        È/ ‰       " !‚&       â!r	       âO !r       Æ� †‰     â z . r  p`ð â !¢		&       â!’&       ä x /   pVp  ä †© 	   è †™    â !r        Ì M	        € ê !r&       Ê/ †y    â My        € ê z   h      È x     pBð Ú G	  @    € ê x     pRð Ú G‰  À    € ê x     pRð Ú G	   $    € ê „y	/     H   " z . r  p`ð â Ey  p    € ä „y/    H   b r /ÿ   pVp  Æ „y/    H   ¨ „y /    H   è „y/    H   ( „y/  
  H   ( „y/ €   H   ( „y/ €  H   ( „y/ €  H   ( „y/ €  H   " !r		       Æ „y/ €  H   " !r       â/ !r	       ÆO !r         â� !r       Æ!r         ä ‰      b !r       Æ ‰       ¢ !r       â !r       Æ !r       â !r       â !r       â/ !r        ÈO ‰      ( ‰      b !r       â !r	       È/ ‰ €    ( ‰
	 €    b !r       Æ „y/ €
  H   â !r
	
       Æ/ ‰ @    ( ‰
 @    b !r       â !r
       È/ ‰       ( ‰      b !r       â !r	        â !r       â/ G	  �     € ê !r'       â!r 	&       Æ  x;ª¸¿  A  Ì s        $ !t  €?    Ì s        $  r     A  È  r      A  Ê †y     ä Ay        € ê ‰       b Ey  Ð    € æ ‰      ¨ „y/    H   è „y/    H   ( „y/    H   è „y/    H   â !r         â/ !r       ÈO ‰       h ‰      ( „y/    H   ¢ !r       â� !r       â!r        â/ !r       È ‰ €    ( ‰ €    h „y/  	  H   â !r       ÈO !r       â !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       È� !r       â !r       â x.   ÿàÿ â !r
	
       Æ/ ‰       " z  r  p`ð Æ ‰
      b x /   pVp  â !r	        â !r
       Ö/ G	  �     € ê !r'       â!r 	&       Æ  x;ª¸¿  A  Ì s        $ !t  €?    Ì s        $  r     A  È  r      A  Ê †y    ä Ay        € ê ‰       b Ey       € æ ‰      ¨ „y/ €  H   è „y/ €	  H   ( „y/ €  H   è „y/ €  H   â !r         â/ !r       ÈO ‰       h ‰      ( „y/ €  H   â !r       È� !r       â!r        â/ !r       È ‰ €    ( ‰ €    h „y/ €  H   ¢ !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       âO !r       â x.   ÿàÿ â !r
	
       Æ/ ‰       " z  r  p`ð Æ ‰
      b x /   pVp  â !r	        â !r
       Ö/ G	  �     € ê !r'       â!r 	&       Æ  x;ª¸¿  A  Ì s        $ !t  €?    Ì s        $  r     A  È  r      A  Ê †y    ä Ay        € ê ‰      b !r       â x..   ÿàÿ Æ !r       â z . r  p`ð Æ !r       â x /   pVp  È ‰       ¢ !r       Ê/ ‰      " !r         ÊO ‰       b !r       Ê ‰ €    " !r        Ê/ ‰ €    b !r	       Ê ‰
	 @    " !r       Ê/ ‰ @    b !r
	
       Ê ‰
      " !r       Ê/ ‰       ¢ !r
       â M	        € ì !r''       â/!r        ÆO  x';ª¸¿  A  â !r&&       Ê s        $ !t  €?    Ì s        $  r''     A  È  r'&'     A  Ê †y '   â My        € ê „y/     H   " z . r  p`ð â Ey       € ä „y/    H   b r /ÿ   pVp  Æ „y /    H   ¨ „y/    H   è „y	/    H   ( „y/  
  H   ( „y/ €   H   ( „y/ €  H   ( „y/ €  H   " !r       â !r       Ä/ „y/ €
  H   â !r         âO !r       Æ� !r	 	       â!r       È ‰ 	      " !r       Æ ‰      b !r       È !r       â !r 	        Æ !r       â !r       â/ ‰       ( ‰      h „y/ €  H   ¢ !r        â !r       È/ ‰ €    ( ‰ €    h „y/ €  H   â !r       âO !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       â� !r       â !r
	
       È/ ‰       ( ‰
      b !r	        â !r 
       â/ G	  À     € ê !r  '       È x '7=  A  â x *BL?  A  Æ #t   €?   È  r     A  â  x    ?  A  â !r 	&       È s     $   $ !t  €?    È  r     A  È  r      A  Ê †y     ä Ay        € ê ‰       b Ey       € æ ‰      ¨ „y/  	  H   è „y/    H   è „y/    H   â !r         Æ/ „y/    H   â !r       ÆO ‰       h ‰      ( „y/    H   ¢ !r        â/ !r       È ‰ €    ( ‰ €    b !r       Æ „y/    H   â !r       ÈO !r       â !r       Æ !r       â !r	       ä/ ‰ @    ( ‰
	 @    b !r       È� !r       â !r       â x.   ÿàÿ â !r
	
       Æ/ ‰       " z  r  p`ð Æ ‰
      b x /   pVp  â !r	        â !r 
       Ö/ G	  À     € ê !r  '       È x '7=  A  â x *BL?  A  Æ #t   €?   È  r     A  â  x    ?  A  â !r 	&       È s     $   $ !t  €?    È  r     A  È  r      A  Ê †y    ä Ay        € ê ‰       b Ey  Ð    € æ ‰      ¨ „y/ €  H   è „y/ €	  H   ( „y/ €  H   â !r         Æ/ „y/ €  H   â !r       ÆO ‰       h ‰      ( „y/ €  H   ¢ !r       È� !r       â!r        â/ !r       È ‰ €    ( ‰ €    h „y/ €  H   â !r       âO !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       â x.   ÿàÿ â !r
	
       Æ/ ‰       " z  r  p`ð Æ ‰
      b x /   pVp  â !r	        â !r 
       Ö/ G	  À     € ê !r  '       È x '7=  A  â x *BL?  A  Æ #t   €?   È  r     A  â  x    ?  A  â !r 	&       È s     $   $ !t  €?    È  r     A  È  r      A  Ê †y    ä Ay        € ê ‰      b !r       â x..   ÿàÿ Æ !r       â z . r  p`ð Æ !r       â x /   pVp  È ‰       ¢ !r       Ê/ ‰      " !r         ÊO ‰       b !r       Ê ‰ €    " !r        Ê/ ‰ €    b !r	       Ê ‰
	 @    " !r       Ê/ ‰ @    b !r
	
       Ê ‰
      " !r       Ê/ ‰       ¢ !r
       â M	        € ì !r''       â/!r        ÆO  x''7=  A  â x'*BL?  A  â !r&&       ä #t'  €?   â x''   ?  A  Æ  r     A  Ì s     $   $ !t   €?    È  r''      A  È  r'&'     A  Ê †y '   â My        € ê x     pRð Ú G‰  P    € ê x     pRð Ú G	  `    € ê „y/     H   " z . r  p`ð â Ey       € ä „y/    H   b r /ÿ   pVp  Æ „y /    H   ¨ „y/    H   è „y	/    H   ( „y/  
  H   ( „y/ €   H   ( „y/ €  H   ( „y/ €  H   " !r       â !r       Ä/ „y/ €
  H   â !r         âO !r       Æ� !r	 	       â!r       È ‰ 	      " !r       Æ ‰      b !r       È !r       â !r 	        Æ !r       â !r       â/ ‰       ( ‰      h „y/ €  H   ¢ !r        â !r       È/ ‰ €    ( ‰ €    h „y/ €  H   â !r       âO !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       â� !r       â !r
	
       È/ ‰       ( ‰
      b !r	        â !r 
       â/ G	  À     € ê !r  '       Ê	z  i    � â !r 	&       È  x;ª¸¿  A  â 	z   i    � È 	z   i €  � â s        $ !t  €?    Ì s        $  r     A  È  r      A  Ê †y     ä Ay        € ê ‰       b Ey       € æ ‰      ¨ „y/  	  H   è „y/    H   è „y/    H   â !r         Æ/ „y/    H   â !r       ÆO ‰       h ‰      ( „y/    H   ¢ !r        â/ !r       È ‰ €    ( ‰ €    b !r       Æ „y/    H   â !r       ÈO !r       â !r       Æ !r       â !r	       ä/ ‰ @    ( ‰
	 @    b !r       È� !r       â !r       â x.   ÿàÿ â !r
	
       Æ/ ‰       " z  r  p`ð Æ ‰
      b x /   pVp  â !r	        â !r 
       Ö/ G	  À     € ê !r  '       Ê	z  i    � â !r 	&       È  x;ª¸¿  A  â 	z   i    � È 	z   i €  � â s        $ !t  €?    Ì s        $  r     A  È  r      A  Ê †y    ä Ay        € ê ‰       b Ey  Ð    € æ ‰      ¨ „y/ €  H   è „y/ €	  H   ( „y/ €  H   è „y/ €  H   â !r         Æ/ „y/ €  H   â !r       ÆO ‰       h ‰      " !r       È� !r       â!r        â/ !r       È ‰ €    ( ‰ €    h „y/ €  H   ¢ !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       âO !r       â x.   ÿàÿ â !r
	
       Æ/ ‰       " z  r  p`ð Æ ‰
      b x /   pVp  â !r	        â !r 
       Ö/ G	  À     € ê !r  '       Ê	z  i    � â !r 	&       È  x;ª¸¿  A  â 	z   i    � È 	z   i €  � â s        $ !t  €?    Ì s        $  r     A  È  r      A  Ê †y    ä Ay        € ê ‰      b !r       â x..   ÿàÿ Æ !r       â z . r  p`ð Æ !r       â x /   pVp  È ‰       ¢ !r       Ê/ ‰      " !r         ÊO ‰       b !r       Ê ‰ €    " !r        Ê/ ‰ €    b !r	       Ê ‰
	 @    " !r       Ê/ ‰ @    b !r
	
       Ê ‰
      " !r       Ê/ ‰       ¢ !r
       â M	        € ì !r''       â/!r        ÈO 	z'' i    � â !r&&       È  x';ª¸¿  A  â 	z&& i    � È 	z&& i €  � â s        $ !t  €?    Ì s        $  r''     A  È  r'&'     A  Ê †y '   â My        € ê „y/     H   " z . r  p`ð Æ „y	/    H   b r /ÿ   pVp  Æ „y/ €   H   ¨ „y/ €  H   è „y/    H   ( „y/    H   ( „y/    H   ( „y
/    H   ( „y/ €  H   ( „y/ €  H   ( „y /    H   " !r       Æ „y"/  	  H   " !r		       Æ/ „y/    H   b !r       ÆO „y/  
  H   ¢ !r       Æ� „y/ €  H   â !r       Æ„y/ €
  H   " !r       Æ „y$/    H   " !r       Æ „y(/    H   " !r		
       Æ „y / €  H   " !r       Æ „y/ €  H   " !r       Æ „y/ €  H   " !r        Æ „y/ €	  H   â !r"       Æ „y/ €  H   â !r       Æ/ „y/ €  H   â !r	       âO !r       ä� ‰	      " !r       Æ‰
      b !r$$       Æ ‰      ¢ !r(       Æ ‰      â !r        Æ ‰$      " !r       Æ ‰      " !r       â !r		       â !r       â !r

       Æ/ ‰ 	      " !r       â !r       äO ‰
      â !r       â !r       ä� ‰      b !r$       Æ‰      ¢ !r       Æ ‰      è ‰      ( ‰      " !r 	        Æ ‰      " !r       â/ !r
       ä ‰  €    b !r       ÆO ‰      ¢ !r       Æ� ‰      â !r       Æ‰ €    " !r       Æ ‰	 €    " !r       Æ ‰ €    ( ‰ €    " !r        Æ/ ‰ €    b !r       âO !r       ä� ‰  @    ¢ !r       Æ‰ €    â !r		       Æ ‰ €    " !r       Æ ‰ @    " !r       Æ ‰	 @    " !r       Æ/ ‰ @    h ‰
 @    b !r         ÆO ‰ @    ¢ !r       â� x.   ÿàÿ â !r       â‰       â x.   ÿàÿ â !r       ä ‰ @    " z  r  p`ò â !r	       Ä ‰ @    " z  r  p`ô â !r       ä/ ‰      b x /   pVò  â !r

       ä ‰	      b x /   pVt â !r       ÄO ‰      ¢ x..   ÿàÿ Æ ‰
      ¢ z . r  p`ö â !r        ä� ‰      â !r       â !‚ &       âx /   pVö â !r       ä‰      " !r       Æ/ ‰      b !r		       â !‚'       â !r       äO !’	&       â  ‚      A  â !r
       â !’'       Æ †‰     â !r       â� !¢&       â  ’     A  ä !¢	'       â !r       ä †™    â  ¢		     A  â !r       È/ †© 	   â M9        € ê !r&&       â !r''       È  r'&'     A  Ê †y '   â My        € ê „y/     H   " z . r  p`ð â Ey  À    € ä „y/    H   b r /ÿ   pVp  Æ „y /    H   ¨ „y/    H   è „y	/    H   ( „y/  
  H   ( „y/ €   H   ( „y/ €  H   ( „y/ €  H   ( „y/ €  H   â !r       â !r       Æ/ !r         âO !r       Æ� !r	 	       â!r       È ‰ 	      " !r       Æ ‰      b !r       È !r       â !r 	        â !r       È/ ‰       ( ‰      h „y/ €  H   ¢ !r        â !r       È/ ‰ €    ( ‰ €    h „y/ €
  H   â !r       ÈO !r       â !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       â� !r       â !r
	
       È/ ‰       ( ‰
      b !r        â !r 
       â/ G	  à     € ê !r  '       Ê	x   à@  � Ê  x #ÛÙ?  A  È  x ;ª¸¿  A  â !r &       Ê s        " 	x    à@  � È 	x    àÀ  � Ê !t   €?    â !t  €?    È s	        $  r	     A  È  r     A  Ê †y     ä Ay        € ê ‰       b Ey       € æ ‰      ¨ „y/    H   è „y/  	  H   è „y/    H   â !r         Æ/ „y/    H   â !r       ÆO ‰       h ‰      ( „y/    H   ¢ !r       â� !r        â/ !r       È ‰ €    ( ‰ €    h „y/    H   â !r       ÈO !r       â !r       Æ !r       â !r	       ä/ ‰ @    ( ‰
	 @    b !r       È� !r       â !r       â x.   ÿàÿ â !r
	
       Æ/ ‰       " z  r  p`ð Æ ‰
      b x /   pVp  â !r        â !r 
       Ö/ G	  à     € ê !r  '       Ê	x   à@  � Ê  x #ÛÙ?  A  È  x ;ª¸¿  A  â !r &       Ê s        " 	x    à@  � È 	x    àÀ  � Ê !t   €?    â !t  €?    È s	        $  r	     A  È  r     A  Ê †y    ä Ay        € ê ‰       b Ey  ð    € æ ‰      ¨ „y/ €  H   è „y/ €	  H   ( „y/ €  H   è „y/ €  H   â !r         Æ/ „y/ €  H   â !r       ÆO ‰       h ‰      ( „y/ €  H   ¢ !r       È� !r       â!r        â/ !r       È ‰ €    ( ‰ €    b !r       âO !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       â x.   ÿàÿ â !r
	
       Æ/ ‰       " z  r  p`ð Æ ‰
      b x /   pVp  â !r        â !r 
       Ö/ G	  à     € ê !r  '       Ê	x   à@  � Ê  x #ÛÙ?  A  È  x ;ª¸¿  A  â !r &       Ê s        " 	x    à@  � È 	x    àÀ  � Ê !t   €?    â !t  €?    È s	        $  r	     A  È  r     A  Ê †y    ä Ay        € ê ‰      b !r       â x..   ÿàÿ Æ !r       â z . r  p`ð Æ !r       â x /   pVp  È ‰       ¢ !r       Ê/ ‰      " !r         ÊO ‰       b !r       Ê ‰ €    " !r        Ê/ ‰ €    b !r	       Ê ‰
	 @    " !r       Ê/ ‰ @    b !r
	
       Ê ‰
      " !r       Ê/ ‰       ¢ !r
       â M	        € ì !r''       â/!r        ÈO 	x''  à@  � â !r&&       È  x'#ÛÙ?  A  â 	x&&  à@  � Æ  x;ª¸¿  A  â 	x&&  àÀ  � Ê s        " !t&&  €?    â !t  €?    Ì s         $  r''      A  È  r''&     A  Ê †y '   â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y       "   $ z   w  p`ð Ú M	        € ê y      !   " ¹z  l      â Ey  Ð    € â ™x?    â y      %   b ¹z  F   
   â $rÿÿ   ÿ Ž ä $rÿÿ   ÿ Ž â xÿ     È |    p`ð â $x   ÿ Ž Ø/ G	  0    € ê y      &   " $tÿ   ÿ Ž ä $z  x  Ž È %v \   Ž Ê �y    ™ ¢  $rÿÿ   ÿ Ž â x   ÿ   â $z	  r  ÿŽ â x   ÿàÿ â 'z m   Ž â r ÿ       Æ $rÿÿ   ÿ Ž â zÿ n    â $x   ÿ Ž ä $vÿ X  ÿ Ž â rÿ   ÿáÿ â %x	$   ÿ Ž â x   ÿÀŽ Æ $z o  Ž â x   ÿÀŽ â %x$    Ž â x   ÿüŽ Æ $z u  ÿŽ â z Z  ÿàñ â %x   ÿ Ž Æ r   	à£ â%x$   ÿ Ž â z [  ä  ä rÿÿ   ÿDþ  â %x$    Ž È $x$   ÿŽ â z Z  ÿàñ â $z q  Ž ä $rÿÿ    Ž â z [  ä  â $vÿ Y  ÿ Ž â x   ÿàñ Ê $rÿÿ     ä $z t  Ž äO$z t  Ž ä $z q  Ž ä %x   Ž È %x   Ž È $rÿÿ    Ž â r
   ÿàñ â $r	ÿÿ    Ž È %x   Ž Ê �y   ™ ¢  $x     Ê �y 
   ™ è �y!
   ™ " %x   Ž Æ �y    ™ h �y%   ™ " $r
ÿÿ    Ž â/ r        ä r   ÿàñ Æ �y#
ðÿÿ ™ $ $x     ä �y$
    ™ h �y"   ™ h �y'   ™ b r   ÿàñ Æ �y&
ôÿÿ ™ d $x     ä �y    ™ h �y    ™ h �y(
   ™ b x   ÿàÿ È |    p`ð ä x@  ÿàó ä x€  ÿàõ Æ $rÿÿ   Ž  ä $rÿÿ    â r)ÿ     ÊO $x*)   ÿ Ž â x+ÿ   )  ä rÿ     ä x*   ÿÀŽ ä/ x*ÿ      ä�x++   ÿÀŽ ä x øŽ â $x    ÿ Ž â x+*øŽ â $x)   ÿ Ž â xÿ   )  Ä x     ÿÀŽ ä x   øŽ â $x)   ÿ Ž â x   øŽ ä x !øŽ â$x )   ÿ Ž â x	ÿ   !  ä x   øŽ â $x)   ÿ Ž â x!   ÿ   Ä x    øŽ â $x )   ÿ Ž â x   !øŽ â $x!   ÿ Ž â x   øŽ ä x*ÿ     ä rÿ   %  ä x     øŽ â &r#   ÿ   â x   !øŽ â $x!   ÿ Ž â x**   ÿÀŽ â $x   ÿ Ž â r%ÿ   %  â &r $      âx	*	øŽ Ä x ÿ     ä x*ÿ     ä x   !øŽ ä x!   ÿÀŽ ä x)    ÿÀŽ â $x %   ÿ Ž â x		   *øŽ Ä xÿ   "  äx*   ÿ   ä x"!"øŽ ä x!ÿ   %  ä x		   *øŽ ä x     ÿÀŽ Ä x*ÿ   '  äx+!   ÿÀŽ â $x!   ÿ Ž â x' 'øŽ ä x)øŽ ä x +*øŽ â $x*%   ÿ Ž â x)ÿ     Ä x""   !øŽ ä x!   )øŽ ä x)   ÿ   ä x'   *øŽ â $x*   ÿ Ž âx""   )øŽ â $x'   ÿ Ž â x+ÿ   %  â $x)%   ÿ Ž â x!!   *øŽ â $x*   ÿ Ž â x""   'øŽ Ä x     +øŽ â $x+%   ÿ Ž âx)   )øŽ â &r#"#   ÿ   â x,%   ÿ   ä x!!   *øŽ â $x   ÿ Ž â x     +øŽ â $x%%   ÿ Ž â x))   ,øŽ â &r!!$   #   â x		   øŽ â &r&      âx     %øŽ â &r!)&   !   â 2r        Ä 2r        â &r		(      ä&r  (   !   ä 0rÿ  0 A   ä 0rÿ  0 A   â Er	 	       â 0rÿ    A   â Er         â 0rÿ    A   â  x   ?  A  â  x   ?  A  â x
@  ÿà÷ Ä #r	      â #r      ä $rÿÿ   Ž â !r       â !r       â G‰  €÷ÿÿÿÿƒ ê Ay        € ê ‰      " r   ÿàÿ Æ ‰      b z  p  p`ð È x    pDp  â !r       â !r       È/ ‰      ( ‰      b !r       â !r       È/ ‰ €    ( ‰ €    b !r       â !r       È/ ‰ @    ( ‰	 @    b !r       â !r		       È/ ‰      h  ‰	      ¢  M	        € ê y      &   â $x   ÿ Ž â !r       â/ $z   s  Ž â !r		       â_ $tÿ   ÿ Ž â x    p"ò Ä r ÿ   p"ð Ö $ÿÿ   	 Ž ä $z  v   Ž È� %v  j   Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      "   $ z  w  p`ð Ú M	        € ê y      &   " $tÿ   ÿ Ž â ¹z  F   
   â $z x  Ž È %v \   Ž Ê �y     ™ b ¹z  l      â Ey  P    € â ™x?    â y      !   b x
      ÿ  â $rÿÿ   ÿ Ž â x      ÿ  â $rÿÿ   ÿ Ž â x"ÿ     È/ | "   p`ð Ú G	  °    € ê $rÿÿ   ÿ Ž â y      %   " $z r  ÿŽ â x	   ÿ   â'z m   Ž â Ey 0    € â z  X      ä $rÿÿ   ÿ Ž â zÿ n    â $x   ÿ Ž Ä $z  t  "Ž ârÿ   ÿáÿ â $v!ÿ Y  ÿ Ž â x   ÿÀŽ â $v#ÿ `  ÿ Ž ä $z o  Ž â x   ÿÀŽ â %x$   ÿ Ž È $z u  ÿŽ ä %x   ÿ Ž Æ r	   à‘ â %x	$    Ž Æ r	ÿÿ   ÿ$~  â $x   ÿ Ž â z Z  ÿàñ â %x$    Ž Æ x   ÿàÿ â $x		$   ÿŽ â z Z  ÿàó â $z' q  "Ž â z [  ä  â %x$   ÿ Ž â z [  	äÿ  ä z ÿ `  pPð â $z q  Ž â x   ÿàó â $z'  t  'Ž â z ÿ a   Sð â $rÿÿ    Ž â x   ÿüŽ â $v%ÿ a  ÿ Ž â rÿ   ÿäÿ  â $rÿÿ   ÿ Ž Ä %x   Ž È %x'   Ž È $r ÿÿ    Ž È %x
'    Ž â r   ÿàó È �y+
   ™ ¢  $x	   !Ž  Ê �y$   ™ è �y,   ™ " %x    Ž â r   ÿàó ä �y
    ™ f $x   !Ž  â �y-   ™ " $rÿÿ    Ž â/ r	        Ä �y*   ™ ( �y&ðÿÿ ™ ( �y     ™ ( �y.   ™ " r   ÿàó Æ �y(ôÿÿ ™ $ $x   Ž  ä �y    ™ h �y    ™ h �y)   ™ b x""   ÿàÿ È | "   p`ò ä r/ÿ   +  ÊO $x0/   ÿ Ž â xÿ   /  È/ x10   ÿÀŽ â $x
/   ÿ Ž â x   ÿÀŽ ä x11$øŽ ä�x$ÿ   $  â $x/   ÿ Ž â rÿ   +  ä x$$øŽ Ä x1   
øŽ ä xÿ   /  ä x
   ÿ   ä x   øŽ ä xÿ     ä x$$   øŽ Ä x
   ÿÀŽ ä x+   ÿÀŽ â $x   ÿ Ž âx,øŽ äxÿ   ,  ä x   øŽ â $x   ÿ Ž â x+øŽ Ä x+ÿ     â $x1/   ÿ Ž âx   øŽ â $x   ÿ Ž â x   +øŽ ä x+   ÿ   ä x
$   1øŽ â $x$/   ÿ Ž â x   +øŽ Ä x+   øŽ â $x1   ÿ Ž â rÿ   -  â $x   ÿ Ž â x   $øŽ Æ $x$   ÿ Ž â x   ÿÀŽ â &r++&   ÿ   â x   1øŽ ä x,ÿ     ä x1$   ÿÀŽ ä r-ÿ   -  â &r    +   â xÿ   *  Ä x,,   ÿÀŽ ä x$1*øŽ ä x+-   ÿ   ä x*ÿ   -  ä x,øŽ ä x++   ÿÀŽ Ä x,ÿ   .  äx1*   ÿÀŽ â $x*-   ÿ Ž â x++.øŽ ä x,1,øŽ ä x1ÿ   -  ä x*+   *øŽ â $x+-   ÿ Ž âx,,   1øŽ â $x1-   ÿ Ž Ä $x3   ÿ Ž â x**   +øŽ ä x+-   ÿ   äx,,   1øŽ â $x1   ÿ Ž â x$$   3øŽ â $x--   ÿ Ž â x.ÿ     ä x+*   +øŽ Ä x   .øŽ â $x.   ÿ Ž â x$$   1øŽ â $x1   ÿ Ž â x,,   -øŽ â &r++&   ÿ   â x   .øŽ â $x   ÿ Ž â x//	   ÿ   â &r+,    +   â x$$   1øŽ â &r(      âx

   /øŽ â 2r        â
x   øŽ â &r$$(   +   â 2r        â &r

)      Ä &r$)   $   ä 0rÿ  0 A   ä 0rÿ  0 A   â Er
 
       â 0rÿ    A   â Er$ $       â 0rÿ    A   â  x   ?  A  â  x   ?  A  Æ #r
      â #r$      â G‰      € ê xÿ     â $r$ÿÿ   # Ž È $z' q  Ž È $z'  t  'Ž È %x
'   $Ž È %x'   Ž â �y-
    ¨ r#   ÿàõ â �y
     è  $x   % Ê �y,    ( �y+    b x.   ÿ   ä r$ÿ   -  ÊO $x*$   ÿ Ž Ê x/*   ÿÀŽ ä x*.   ÿÀŽ È r*ÿ*   -  ä x-ÿ   $  â $x.$   ÿ Ž ä $x
*   ÿ Ž â x//,øŽ äxÿ   *  ä x0-   ÿÀŽ Ä x-ÿ   ,  ä x,
   ÿÀŽ ä x/   .øŽ ä/ x.   ÿÀŽ â $x*   ÿ Ž â xÿ   +  äx,,+øŽ Ä x.øŽ ä x,   øŽ ä x.ÿ   *  ä x,*   ÿ   â $x+$   ÿ Ž â x   .øŽ â $x.*   ÿ Ž âx   ,øŽ â $x,*   ÿ Ž â x   +øŽ â $x**   ÿ Ž â x
0-øŽ Ä x+ÿ   $  â $x-$   ÿ Ž â x   .øŽ ä x   ,øŽ ä x

   +øŽ ä x   *øŽ â &r&   ÿ   â x+$   ÿ   Ä x

   -øŽ â $x-$   ÿ Ž â x   +øŽ â &r       â 2r        ä� x

   -øŽ â &r(      ä 0rÿ  0 A   ä &r

)      Ä 0rÿ    A   â  x   ?  A  ä Er
 
       Ê #r
      È !r       â !r       â !r       â G‰  À    € ê $r
ÿÿ   # Ž ä $rÿÿ   % Ž È %x
   
Ž Ê �y+
    ¢ r#   ÿàõ Æ �y
     ä  $x   % Ê �y*    ( �y$    b rÿ   +  ÈO x   ÿ   È x-   ÿÀŽ ä rÿ   +  ä x+ÿ     Æ $x,   ÿ Ž â x--*øŽ äx*ÿ   *  ä x,   ÿÀŽ ä x.ÿ     ä x++   ÿÀŽ ä xÿ   $  Äx$$øŽ â $x
   ÿ Ž â x.   ÿÀŽ ä x*+*øŽ ä xÿ     ä xøŽ â $x   ÿ Ž â x**   øŽ â $x   ÿ Ž â x
-   
øŽ Ä x+ÿ     ä x$$   øŽ ä x

   øŽ â $x   ÿ Ž âx   ÿ   äx,   +øŽ â $x   ÿ Ž â x*   øŽ â $x+   ÿ Ž â x
   øŽ â $x   ÿ Ž â x   øŽ Ä x$$   +øŽ â &r&   ÿ   â x   ÿ   â$x   ÿ Ž â x,   øŽ â &r       â x$   øŽ â 2r        â� x   øŽ Æ &r((      ä 0r
ÿ  0 A   ä&r)   (   ä 0rÿ    A   â  x

   ?  A  ä Er        Ê #r
   
   È !r
       â x@  ÿàõ ä x€  ÿà÷ ä x@  ÿàù â $rÿÿ    â x##€  ÿàû â $r!ÿÿ   !Ž ä $rÿÿ   	 â r%ÿ%   ÿäÿ â D       À â Gy  Pïÿÿÿÿƒ ê Ay       € ê $r
ÿÿ    Ž ä $rÿÿ    Ž Ä Ay        € ê z ÿ `  pPð È z ÿ a   Sð Ú G	  P    € ê ‰      ( ‰      b !r       â !r       È/ ‰      ( ‰      b !r       â !r       È/ ‰ €    ( ‰ €    b !r       â !r       È/ ‰ @    ( ‰	 @    b !r       â !r		       È/ ‰      ( ‰	      b !r       â !r	       â/ Gy  €    € ê ‰      ( ‰      h ‰      ¨ ‰      â !r       â !r       È/ ‰      " !r       ÆO ‰      b !r       Æ� ‰      ¨ ‰
      â !r	       â !r       â/ !r       ÈO ‰ €    " !r
       Æ� ‰
	 €    h ‰ €    ¨ ‰ €    â !r       â !r
	
       È/ ‰ @    " !r       âO !r       È� ‰ @    h ‰
 @    ¨ ‰ @    â !r       â !r       â/ !r
       ÈO ‰	      " !r       Æ� ‰      h ‰      ¨ ‰      â !r
	       â !r       â/ !r       âO !r       â� y      %   $ $x   Ž Ê z  p  p`ð È x    pDp  Ú M	        € ê z ÿ ^  pPð â $z  v  Ž Æz ÿ _   Sð Ú $ÿ   ÿ Ž È % ^   Ž Ì �	     ¢ x    ÿ   È r  ÿ   p"ò äx     p"ô äx     p"ö äx     p"ø Î $ ÿÿ    Ž â z ÿ `  pPò â $" ÿÿ    Ž ä $2 ÿÿ   
 Ž â z ÿ a  Sò â $B ÿÿ    Ž È !         ÐO G™       € ê z ÿ b  pPð È z ÿ c   Sð Ú         Ê % b   Ž Ì �	     ¢ $x   ÿ Ž Ê x øÿÿÿp"ò äx üÿÿÿp"ô är ÿ   p"ö äx    p"ø Î $ÿÿ    Ž ä $"ÿÿ    Ž â z  h      â $2ÿÿ   
 Ž ä $Bÿÿ    Ž â x    pBò Æ !       ÔO G  P    € ê x    pRð Ú G‰  �     € ê x    pRð Ú G	  ð    € ê  x;ª¸¿  A  Ì s        $ !t  €?    Ì s        $  r     A  È  r       A  â Gy  P    € ê  x'7=  A  â x*BL?  A  Æ #t  €?   â  x   ?  A  Æ  r     A  Ì s     $   $ !t  €?    È  r     A  È  r       A  â Gy  °    € ê x    pRð Ú G‰  à     € ê x    pRð Ú G	        € ê 	z i    � ä 	z   i    � Æ  x;ª¸¿  A  â 	z   i €  � Ê s        $ !t  €?    Ì s        $  r     A  È  r       A  â Gy  Ð     € ê  r       A  â Gy  °     € ê 	x  à@  � ä 	x    à@  � Æ  x#ÛÙ?  A  â 	x    àÀ  � Æ  x;ª¸¿  A  ä !t   €?    È s        $ !t  €?    Ì s        $  r      A  È  r       A  ä y      "   ( y      &   b $z s  Ž ä $tÿ   ÿ Ž ä $z v  Ž È/ %v j   Ž Ê †y      â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      "   " ¹z  l      â Ey  À    € â ™x?    â y      !   b ¹z  F   
   â x      ÿ  â x      ÿ  â y&      %   ¢ r, ÿ       â x       ÿ  â r4 ÿ       â $rHÿÿ   ÿ Ž â rB ÿ       â $r ÿÿ   ÿ Ž â rD ÿ       Ä r ÿ       ä r$ ÿ       ä rF ÿ       ä r ÿ       ä r ÿ   pRò är ÿ   pRð Ä r   ÿ(Ž ä/ x&&   ÿ   äO r   ÿ� È xÿ     È |    pbô Ú G)       € ê y      '   " r ÿ       ä r ÿ       â y      &   b xþÿÿÿÿÀŽ ä z/  X      ä z9  Y      â $x   
Ž â r ÿ       Ä z7 q  ÿàÿ ä x=   ÿ   äx*   ÿ   â 'z y   Ž È 'z s   Ž â/ zÿ z    Æ $z( }  ÿŽ â zÿ t    â $z& p  Ž ä$z	 |  Ž ä$z |  Ž â x&   ÿàÿ â $z	 v  	Ž ä$z' v  Ž â x   ÿ   â $z w  ÿŽ Ä $z) p  	Ž â z  q      ä xA=   áÿ ä r((   ÿàõ â $x-   Ž âr%   ÿŽ â$x+   Ž âxÿ   A  â $x1   Ž ârSÿÿ   ÿä â $x3   Ž âx?A     â $x5   Ž â xÿ   =  â %x   ÿŽ Æ x;=     â $xAA   ÿ Ž â x==   ÿ   â %x'   Ž È %x)   Ž È r8 /       är/   ÿàõ Æ %x'   8Ž â r9   ÿä È �y0   • ¨  �yC   • ¨  �y   • è �yE   • è �yG
   • ( �y6   • " %x)   8Ž â r/   ÿàõ Ä xU $       â �y"    • b r	9   ÿä Æ �yK   • h �y@   • h �yI   • h �y>   • b r(   ÿàõ Æ �yM
   • b rÿS   ÿä Æ/ �y2   • b %v Z  U Ž â r7(   ÿà÷ ä �y#    • b $x$   ÿŽ â r=   ÿàõ Æ $x   Ž â rÿS   ÿäÿ â %v Z  U Ž Æ r;   ÿä â $x$   ÿŽ â r
A   ÿà÷ â �y.    ™ b r=   ÿàõ Æ/ �y8   ™ b r		   ÿàÿ ä r?   ÿäÿ â �y:   ™ b r%(   ÿàù ä r	;   ÿä â �y<
   ™ b rA   ÿà÷ â $rÿÿ   S Ä �yN   ™ b r	?   ÿäÿ Æ �yL   ™ b r
-(   ÿàõ â %v Z  U Ž ä �yO   ™ b rÿS   ÿä â $x$   ÿŽ ä �yJ    ™ b %v

 Z  U Ž â r=   ÿà÷ Ä/ �yP   ™ b r   ÿàÿ Æ �yQ   ™ b $x	$   ÿŽ â r;   ÿäÿ Æ $x   	Ž â rA   ÿàõ â �yT   ™ b r=
   ÿàù ä/ r        â �yW   ™ b r;   ÿä ä r	?   ÿä â �yR    ™ b r+(   ÿàõ Ä rA
   ÿà÷ â �yY   ™ f r?   ÿäÿ â �yX   ™ b rÿS   ÿä ä/ r1(   ÿà÷ â �yZ   ™ b %v Z  U Ž Æ �y[   ™ b $x	$   ÿŽ Æ �yV
    ™ b $x   	Ž Æ �y\   ™ b %v Z  U Ž Æ �y]   ™ b r
ÿS   ÿäÿ ä r=   ÿà÷ â �y_    ™ d $x
$   ÿŽ â r	;   ÿäÿ ä r=   ÿàõ ä/ r   ÿàÿ â �y^   ™ b r
A   ÿà÷ Ä r;   ÿä â �ya   ™ b r?   ÿäÿ ä rf3(   ÿàõ â�y`   ™ b r5(   ÿà÷ ä rA   ÿàù â �yb
   ™ b $rÿÿ   S â rÿS   ÿäÿ Ä r?   ÿä â �yd   ™ b r        â %vf Z  U Ž ä�yc   ™ d $x$   ÿŽ ä/ %v Z  U Ž â �ye   ™ f $x$   ÿŽ â r   ÿàÿ â �yf   ™ b r=   ÿàõ Ä r
=   ÿà÷ â �y    ™ b r   ÿàÿ ä r	;   ÿä Æ $x   ;Ž â rA   ÿàõ â/�y   ™ b rA   ÿà÷ ä r?   ÿä â�y
   ™ b r?   ÿäÿ Æ �yU   ™ h �y
   ™ h �y   ™ h �y   ™ h �yg   ™ h �yh
   ™ h �y    ™ h �y    ™ b xC0T  C    ÈO rÿA   C  ä/r0ÿ*   C  ä x   ÿ   ä x0   ÿ   ä xT  E    ä� xÿ   0  Ä x6GT  6    äx   ÿÀŽ ä x   ÿÀŽ ä x	ÿ     ä x
   ÿÀŽ ä xÿ     Ä xÿ   6  äx6øŽ â $x6   ÿ Ž â xøŽ ä x	
	øŽ ä xj   ÿÀŽ ä xÿ   0  Ä x
0   ÿ   ä x6   6øŽ ä xjøŽ ä x	   øŽ ä x0   ÿ   ä xjÿ     Ä x
   
øŽ ä x	0   ÿ   ä x   ÿ   ä x   øŽ ä x   jøŽ ä x   ÿ   Ä x	
   	øŽ ä x
0   ÿ   â$x00   ÿ Ž â x@KT  @    äx66   øŽ ä x   øŽ ä x   ÿ   Ä rÿ*   @  ä x		   
øŽ ä x
   0øŽ ä x6   øŽ ä x>IT  >    ä rÿA   @  Ä x0ÿ     ä x   ÿ   ä x6   ÿ   ä xÿ   >  ä xC0   ÿÀŽ ä x   ÿÀŽ Ä x2MT  2    ä xE6   ÿÀŽ ä x0ÿ     ä xCCøŽ ä x>øŽ ä x   ÿ   Ä xEE2øŽ äx6ÿ   2  ä xG0   ÿÀŽ ä x0ÿ     ä x2   ÿ   ä x   øŽ â $x   ÿ Ž â x0C   0øŽ Ä x2E   2øŽ ä xC   ÿ   ä xE   ÿ   ä x   øŽ ä x   ÿ   ä x6G6øŽ Ä x00   CøŽ ä x22   EøŽ â $xE   ÿ Ž â xGÿ     ä xC	   ÿ   ä x   øŽ ä x66   GøŽ Ä xG   ÿ   ä x0   CøŽ â &rC	8   ÿ   âx2   EøŽ â &rE8   ÿ   â x	   ÿ   â &rC
:   C   âx66   GøŽ â &rE:   E   â x	   ÿ   Ä x   øŽ â 0rÿ.  0 A   â x06   øŽ â &rC<   C   ä&rE<   E   â  x   A  A  â &r	N   ÿ   ä &rCL   C   ä&rE0L   E   Ä &rNN   ÿ   â ErC C       â &r6
O      â Er2 E       â &rNO   N   ä 0r.ÿ.    A   ä &r6P   6   ä&rNP   N   â #rC.C     â0rÿ"    A   â #r2.2     â 0rÿJ  0 A   Ä &r6Q   6   ä&rN0Q   N   â #rC      â  xC   A  A  â 0rJÿJ    A   â Er 6       â &r"	T   ÿ   â ErN N       â &rTT   ÿ   ä &r"
W   "   â #rJ   C  â#rNJN   C  â &rTW   T   Ä &rC	Y   ÿ   ä &r"X   "   â#rF   F   â &rTX   T   ä &r.
Z   C   ä 0rÿR  0 A   ä &r"[   "   Ä 0r#ÿ#    A   ä &rT0[   T   ä &rYY   ÿ   ä &r.\   .   â  xC   A  A  â 0rRÿR    A   â Er "       â #r #2       â &rYZ   Y   â ErT T       â &r.]   .   Ä 0r2ÿV  0 A   â#r"R   C  â&rY\   Y   â #rTRT   C  â 0rVÿV    A   â Er .       â  xC2   A  A  â &rY0]   Y   ä &r.	^   ÿ   â#rV   C  â &r^^   ÿ   â #r"      âEr" Y       â #rH   H   â &r^a   ^   Ä &r	`   ÿ   â #r"V"   C  â &r^b   ^   ä &rC`   ÿ   ä &r2
d      ä 0rÿ_  0 A   ä &r^0c   ^   Ä &rdd   C   ä &rCe   2   â  x2   A  A  â 0r_ÿ_    A   â Er ^       Ê #r_   2  â &r.
a   .   Æ #r #       â &r	   ÿ   ä&r   ÿ   ä &r		   ÿ   ä&r   ÿ   ä &r
U      ä &r.b   .   Ä &rU      ä &r	
   	   ä&r      ä &r      âx    ÿàÿ â &r.c   .   ä &r      Ä &ree   d   ä &r	   	   ä&r      ä &rg      âEr. .       â &rg0g      â |    pbô â &rCf   C   Ä &re0f   e   ä &r	h   	   ä0rÿ  0 A   ä 0rÿ  0 A   ä &r00h      ä 0r6ÿ  0 A   â #r._.   2  â 0rÿ    A   â  x   A  A  â Er	 	       â 0rÿ    A   â ErC C       â 0rÿ    A   â Er2 e       â  x   A  A  â Er        â  xE6   A  A  â Er g       Ä Er0 0       â #rB#"   B   â #r	     â #r"C   E  â#r22   E  â #r     â#r     â #r0     â x//À  ÿà÷ â #r$#N   $   â#rD#T   D   â #r!.   !   â#r"      â #r4#2   4   â #r,   ,   â #r#      â #r      â #r#      â x%%    ÿàÿ Ä x--    ÿàÿ ä x++    ÿàÿ ä x11    ÿàÿ ä x33    ÿàÿ ä x55    ÿàÿ ä x77    ÿàÿ Ä r9ÿ9   ÿäÿ â D)       À â Gy  °èÿÿÿÿƒ ê Ay        € ê Ey  0    € â G™      € ê r   ÿHŽ Ê ˆs  øÿ H   è ˆs  €øÿ H   è ˆs F ùÿ H   è ˆs $€ùÿ H   è ˆs  úÿ H   è ˆs D€úÿ H   è ˆs H ûÿ H   è ˆs B€ûÿ H   è ˆs ! üÿ H   è ˆs  €üÿ H   è ˆs  ýÿ H   è ˆs 4€ýÿ H   è ˆs , þÿ H   è ˆs €þÿ H   è ˆs  ÿÿ H   è ˆs €ÿÿ H   ä Ay        € ê {          ì M	        € ê „y     H   b z & r  p`ð Ä z/  r      â „y €   H   ¢ r ÿ   p"p Æ „y	 €  H   â $x-/   ÿŽ ä$x7/   ÿŽ â „y €  H   ( „y    H   h „y    H   ( y      &   è „y    H   ( y      '   h „y €  H   â !r       Æ/ „y €  H   â !r        ÆO ‰      " $z  x  &Ž â� !r$	$       ä „y    H   â x&&   ÿàÿ â !rDD       ä„y    H   â $z  ~   Ž â!rFF       Ä ‰$      b !rHH       â  -   ÿàó ä „y €  H   ä ÿÿ   ÿäÿ  ä „y#    H   â z & r  p`ò â !r       â !r       â „y% €  H   â 
 j  ÿˆ Ä x    p"ò â ‰      " 
 k   â !rBB       ä ‰      ¢ !r         Æ „y    H   â !r$$       Æ/ ‰D      h ‰F      â !r!!       Æ ‰H      " !r"       â !r       ä ‰      " r/    ÿ† â !r
       äO ‰	 €    ¢ rÿÿ   ÿäÿ â !r44       â z j  ÿ† â ‰
      b !r,#,       â 
  r  ÿàõ â !rDD       â/ ‰$      b z k  � â !r&%       â !rFF       â� „y5 €  H   â $ÿÿ   ÿ â 
 j  ÿ„ â !rHH       â‰D      â  7   ÿàù â !r(       Ä ‰F      " !r       â 
 k   â !r	       âO $x	/   ÿŽ â‰B      " ÿÿ   ÿä â !r
       âr/    ÿ„ â‰ @    ¢ r		    ÿà÷ â !r$$       â/ 
 j  ÿˆ â ‰!      b rÿÿ   ÿäÿ Ä z	 j  ÿ† â‰       b rÿÿ   ÿä â !rDD       â� z		 k  � â ‰#"      â $x/   ÿŽ â z
 j  ÿ„ â !rFF       â‰      " 
 k   Ä     ÿà÷ â ‰%4      " z k   â !rBB       â !r55       â ‰',      " !r       âO /   ÿàÿ ä ‰)&      ¢ !r!       â/ .    ÿàù â $ÿÿ   ÿŽ â ‰+(      b !r         â
 j  ÿ† Ä ‰ €    b !r""#       â� 9ÿÿ   ÿä ä ‰      â !r       â. j  ÿˆ ä ‰H      " !r44%       â 
 k  � â !r,,'       â ‰%B      " . k  9 â !r&&)       ÄO ‰'      ¢ !r((+       Æ/ ‰)       b !r       â-   ÿàÿ ä ‰F €    b !r9       â�     ÿàõ ä ‰+"      â 7   ÿàÿ â !rHH       â*ÿÿ   ÿä â ‰-4      "  j  ÿ„ â !rBB%       Ä ‰$ €    "  k  * â !r'       äO ‰/,      ¢     ÿàõ â !r )       ä/ ‰1&      b  ÿÿ   ÿä Æ ‰3(      b !rFF       Æ‰5      b !r""+       Æ� ‰ @    â !r44-       Æ‰ €    " !r$$       Æ ‰!D €    " !r,,/       ÆO ‰#H €    ¢ !r*&1       â/ !r.(3       ä ‰'B €    b !r05       Æ‰) €    b  j  ÿ„ â!r       ä� ‰+ €    â  k    â !r       ä‰F @    " x        â !rDD!       Ä ‰-" €    $ %v  j   Ž â !rHH#       âO ‰/4 €    ¨ ‰$ @    ¢ !r'B'       Æ/ ‰1, €    b !r))       Æ‰3* €    b !r++       Æ� ‰5. €    â !rF       Æ‰70      " !r-"-       Æ ‰      " !r/4/       ÆO ‰! @    ¢ !r$       Æ ‰#D @    ¢ !r1,1       Æ/ ‰%H @    b !r3*3       Æ‰ ' @    b !r5.5       â� !r707       ä‰") @    â !r;       Æ ‰$+ @    " !r!!       ÆO ‰      â !r#D#       Æ ‰&- @    ¢ !r%H%       Æ/ ‰(/ @    b !r '        Æ‰      ( ‰*1 @    b !r")"       Æ� ‰,3 @    â !r$+$       Æ ‰07 €    ( ‰.5 @    " !r&-&       ÆO ‰!      ¢ !r(/(       Æ/ ‰#      b !r)       Æ†	 9    â	 !r*1*       â!r,3,       ä� † ;   â !r070       Æ ‰       " !r9       â!r.5.       ä ‰%      â !r+!       ÆO ‰"      ¢ !r-#       Æ/ ‰$      h †	 9    è ‰&      " !r        Æ ‰(      ( † )    â !r%%       Æ� ‰!*      â !r"       ÆO ‰#,      ¢ !r$       Æ/ ‰'0 @    h ‰.      b !r&       Æ†	 +    â	 !r(       Æ † -   è	 †	 %    â	 !r!*!       â� !r#,#       äO †     â	 !r'0'       Æ/ †	 
    â	 !r.       Æ† 
   è	 †	     è	 †     è	 †	 !    è	 † #   è	 ‰'      ( †	     â	 M™        € ê x   ÿàÿ â !r''       Ær    ÿàñ È r ÿÿ   ÿä  ä z j  ÿ€ È z k     Ê †y '    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y
      "   " ¹z  l      â Ey  p    € â ™x?    â y      !   b ¹z  F   
   â r ÿ       â $rÿÿ   ÿ Ž â y      %   ¢ r9 ÿ       â $r ÿÿ   ÿ Ž â r; ÿ       â x      ÿ  â r= ÿ       â x      ÿ  â x      ÿ  â x      ÿ  â r 
ÿ   pRð Ä r 
ÿ   pRò är
   ÿ(Ž È/ r   ÿ� ä x   ÿ   äO xÿ     È |    pbô Ú G)       € ê y      '   " r ÿ       â $rÿÿ   ÿ Ž â r ÿ       â y      &   b xþÿÿÿÿÀŽ ä x   ÿàÿ ä z  X      â $x   
Ž â z'  Y      Ä z% q  ÿàÿ â $x+   ÿ Ž âx$   ÿ   äx    ÿ   ä x6$   +áÿ â %x*+   ÿŽ Æ x7ÿ   6  â 'z y   Ž Æ x76   7  â 'z s   Ž â/ zÿ z    Æ $z }  ÿŽ â zÿ t    â $z p  Ž ä$z |  Ž ä$z |  Ž â z  q      â $z w  ÿŽ ä $z v  Ž â r   ÿŽ â $z v  Ž â xÿ   $  â $z p  Ž â r   ÿàõ â %x(   *Ž â x"$     Ä r   ÿŽ â$x   Ž âx$$   ÿ   â $x!   Ž âr&ÿÿ   ÿä â $x#   Ž ä $x66   ÿ Ž ä %x*   *Ž È r        ä r '       Ä r(   ÿàõ Æ %x   Ž Ê �y\   • ¨  �y>   • ¢  r)'   ÿä â %xP   Ž È $tÿ$   ÿ Ž â �y
   • è �yF   • è �y:   • ( �y?   • " r*   ÿàõ Æ �yBP   • ( �yP   • " r+'   ÿä ä r1%   ÿà÷ â�yX    • ( �y]   • ( �y@   • ( �y[
   • ( �yZ   • " r-   ÿàõ Æ �yPP    • "rÿ&   ÿä â/ %v,- Z   Ž È $x$   ÿŽ â r$,   ÿàõ È r--   ÿàÿ È r-"   ÿä â $r2ÿÿ   &Ž â r6,   ÿàõ â %v01 Z   Ž â �yV,    ™ $ r-7   ÿä ä �yD   ™ " r/   ÿàõ Æ �yA   ™ " r.ÿ&   ÿä â $x32$   ÿŽ ä �yY   ™ " %v/ Z   Ž Æ �yW   ™ " $x/.$   ÿŽ â r,$0   ÿàõ är113   ÿàÿ ä r.$   ÿàù ä r/   ÿàÿ â �y80    ™ " r-1"   ÿä Ä r460   ÿà÷ â$x/   " â �yR    ™ " r26   ÿàõ Æ �yQ,   ™ "r517   ÿäÿ Æ �y.   ™ "# r37   ÿä Æ �yO,   ™ " rC   ÿà÷ Æ �y.   ™ " rÿ&   ÿäÿ Æ �yU4   ™ " rE   ÿàõ Æ �yT2   ™ ( �yS,   ™ " %v4C Z   Ž â r<ÿ&   ÿä ä �y.   ™ " $x3$   ÿŽ â/ r^$4   ÿàù â %v0E Z   Ž Æ r553   ÿàÿ â $xE<$   ÿŽ Æ r_5"   ÿä â $x11   EŽ â r,$0   ÿàõ â�yK4    ™ b r264   ÿàû Æ �y^   ™ b r`60   ÿà÷ â $x35   7Ž â r-1"   ÿä â �y^   ™ b r/!   ÿàõ ära17   ÿäÿ â �yM2   ™ b rC#   ÿà÷ ä rEÿ&   ÿä â�y^   ™ b r<ÿ&   ÿäÿ â %v./ Z   Ž Ä �yJ,   ™ d $xEE$   ÿŽ ä %v2C Z   Ž â �yH,   ™ d r//E   ÿàÿ â $xC<$   ÿŽ â �yL,   ™ h �yI0    ™ b r33C   ÿàÿ Æ �yN`   ™ b r,$.   ÿàõ ä/r0$2   ÿàù Æ $x-/   " â r^6.   ÿà÷ är13"   ÿä â �y..    ™ b r462   ÿàõ Æ �y,   ™ b r_/7   ÿäÿ Æ �y<0   ™ b r537   ÿä Æ �yG,   ™ h �yE0   ™ h �yC^   ™ h �y44   ™ h �y22    ™ h �y`0   ™ h �y5,   ™ d x-\T  >    ÈO r>ÿ    -  È x\>   ÿ   ä xFT  F    ä� rÿ6   -  ä x_\   ÿÀŽ ä x::T  ?    â$x3   ÿ Ž â/ x,ÿ   >  Ä x0_:øŽ äx-ÿ   :  ä x:ÿ     ä x,,   ÿÀŽ ä x1>   ÿ   ä x/ÿ   F  Ä x33   ÿÀŽ ä x::   ÿÀŽ ä x-,-øŽ ä x00   1øŽ ä x,3FøŽ ä x/:/øŽ Ä x1>   ÿ   ä xFÿ   >  ä x:ÿ     ä x10   1øŽ ä x--   FøŽ ä x//   :øŽ Ä x0>   ÿ   äx:>   ÿ   ä x0-   0øŽ ä x-1   :øŽ ä x3   ÿ   ä x1BT      Ä xF   ÿ   äx,,   3øŽ â $x3   ÿ Ž ârÿ    1  â $x:   ÿ Ž â x//   FøŽ ä x?   ÿ   ä x   ÿ   Ä x,,   3øŽ ä x3>	   ÿ   ä x:/   :øŽ ä x@]T  @    ä x/   ÿÀŽ ä rÿ6   1  Ä x00   3øŽ ä x1ÿ     ä x//@øŽ äx3ÿ   @  ä xZ[T  Z    ä x@ÿ     Ä x>1   ÿÀŽ ä x1ÿ   Z  ä x@@   ÿÀŽ ä x3>3øŽ ä x>   ÿ   ä x1@1øŽ Ä x,,   ?øŽ ä x@ÿ     ä x?   ÿ   ä x>/   >øŽ ä x/3   @øŽ â $x3   ÿ Ž â x??   ÿÀŽ Ä xBÿ     ä xZ?ZøŽ ä x11   BøŽ ä xF   ÿ   ä x?   ÿ   ä x3>   3øŽ Ä xB   ÿ   ä x>/   FøŽ ä x@Z   ?øŽ ä x/3   BøŽ â $x3   ÿ Ž â x?   ÿ   â &r-D   ÿ   â xF   ÿ   Ä x@@   ?øŽ â &rD/D   ÿ   â x?   ÿ   ä x>>   3øŽ â &r0A      âx11   FøŽ ä x@@   ?øŽ â &rD>A   D   â xB	   ÿ   â &r,Y      Ä &rD@Y   D   â xB1   BøŽ â &r:W      ä0rÿV  0 A   ä&rDBW   D   â Er        â 0rVÿV    A   â  x   A  A  ä ErD D       â 0rXÿX    A   â #r1V     Æ #rVD     â &r-Q   ÿ   â #r	X1   	   â &rQ/Q   ÿ   ä &r1-   ÿ   ä&r/   ÿ   ä &r0O      ä&rQ>O   Q   Ä &r>      ä &r,U      ä &r10   1   ä &rQ@U   Q   ä&r@T      ä &r:S      Ä &r1,T   1   ä &rQBS   Q   ä 0rÿ8  0 A   âEr        â &r1:   1   ä&r3B      â  x   A  A  â 0r8ÿ8    A   â Er Q       â 0rÿR  0 A   â Er1 1       â 0rPÿP    A   â #r8     â#r8     â 0rRÿR    A   â Er8 3       â  x   A  A  â #rP      Æ #rR1     â#rR8     â &r-   ÿ   ä
&r1/   ÿ   ä &r0      ä&r1>   1   ä &r,M      Ä &r1@M   1   â #rX      â &r:      ä&r1B   1   ä 0rÿK  0 A   ä&r-J   ÿ   â #rX      â 0rKÿK    A   â Er        â &r0H      â  x8   A  A  â #rP      Ä &r,N      â #rK   8  â 0rÿI  0 A   ä &r:L      â #r X       â &r-   ÿ   â#rP      â &r/   ÿ   â x   A  A  â &r/J   ÿ   â Er        â &r--<   ÿ   Ä &r//<   ÿ   ä 0rIÿI    A   ä &r0G      ä&r>G      ä&r>H      ä &r-0E   -   Ä &r/>E   /   â #rI     â &r,C      âx    ÿàÿ â &r@C      ä&r@N      ä &r-,4   -   ä&r/@4   /   â #rX      â &r:5      â|    pbô â &r5B5      Ä 0rÿ.  0 A   ä 0rÿ2  0 A   ä &rBL      ä &r-:`   -   ä&r/B`   /   â  x   A  A  â  x   A  A  â Er        â 0r.ÿ.    A   â Er1 1       â 0r2ÿ2    A   â Er        Ä Er 5       ä Er- -       ä Er /       â #rI     â #r1K1   8  â #r.     â#r.     â #r-2-     â#r2     â xÀ  ÿà÷ â #rP      â#r=P1   =   â #r;X   ;   â #r9P   9   â #rX-      â #rP      â x    ÿàÿ Ä x    ÿàÿ ä x    ÿàÿ ä x!!    ÿàÿ ä x##    ÿàÿ ä x%%    ÿàÿ ä r'ÿ'   ÿäÿ â D)       À â Gy  Àêÿÿÿÿƒ ê Ay        € ê Ey      € â G‰  ð     € ê $x

À  Ž Ê ˆs 
	 ùÿ H   è ˆs 
€ùÿ H   è ˆs 
 úÿ H   è ˆs 
€úÿ H   è ˆs 
 ûÿ H   è ˆs 
€ûÿ H   è ˆs 
 üÿ H   è ˆs 
€üÿ H   è ˆs 
  ýÿ H   è ˆs 
=€ýÿ H   è ˆs 
; þÿ H   è ˆs 
9€þÿ H   è ˆs 
 ÿÿ H   è ˆs 
€ÿÿ H   ä Ay        € ê {          ì M        € ê „y
     H   " z  r  p`ð â $v'ÿ r  ÿ Ž Ä „y €   H   b r ÿ   p"p â $x%'   ÿŽ ä „y    H   ¨ „y    H   è „y €  H   ( „y    H   h „y €  H   h „y €  H   h „y    H   è „y €  H   â !r

	       Æ „y    H   â !r       Æ/ ‰	
      " !r       ÆO „y  €  H   â !r       Æ� „y    H   â !r       Æ„y" €  H   â !r       Æy      &   b !r       Æ ‰      ¢ !r       â !r	
	       ä ‰      " x   ÿàÿ â !r        ä ‰	      â z  r  p`ò â !r==       Ä ‰      " x    p"ò â !r;;       ä ‰
      b $z x  Ž Æ/ y      '   b !r       ÆO ‰      ¢ !r9 9       â !r!       ä ‰      ¢ !r       â !r#"       â !r	       â� ‰      " !r       Æ‰=      â !r

       Æ‰      " $z ~  Ž Æ/ ‰ ;      b !r       ÆO ‰      ¢ 
  r  ÿàõ â !r       ä ‰ €    b ÿÿ   ÿä ä 
  j  ÿ„ â ‰"9      b !r       â r'   ÿˆ Ä ‰	      " 
  k   â !r==       â� r.ÿÿ   ÿä â ‰$!      â !r       Æ‰&#      " !r;;        Æ/ ‰      b !r       âO r'   ÿ„ ä ‰      ¢ !r        Æ‰      b !r99"       Æ ‰
      b !r	       â 	%   ÿàû â !r!!$       â� ‰       $ ÿÿ   ÿäÿ â !r##&       â‰"      â 
	 j  ÿŠ â !r       Ä/ ‰$=      b 
		 k  � â !r       äO ‰&;      ¢ rÿÿ   ÿä â !r       ä‰  @    " z j  ÿ„ â !r
       Ä ‰ €    b $x'   ÿŽ â z
 j  ÿˆ ä ‰(9      b !r         â    ÿà÷ ä ‰ €    " !r""       â� ÿÿ   ÿäÿ Ä ‰*!      â 
 j  ÿ† â !r$=$       â/ z k   â ‰,#      b 
 k  � â !r&;&       âO z k  . â ‰ €    ¢ !r         â'   ÿàÿ Ä ‰ €    " !r       â%   ÿàÿ ä ‰ €    b !r(9(       â /   ÿàù ä ‰ €    b !r       â $x'   ÿŽ â 2ÿÿ   ÿä â !r*!*       â� ‰%" €    " / j  ÿˆ â !r,#,       Ä/ ‰'$ €    b 0   ÿà÷ â !r       äO ‰#  €    ¢ / k  2 â !r       ä‰)& €    â r   ÿàõ â !r       Ä‰ @    " 1ÿÿ   ÿäÿ â $r.ÿÿ   ÿ â !r       â ‰+( €    b z j  ÿ„ ä0 j  ÿ† â ‰-* €    b !r%"%       â z k  . Ä ‰! @    " !r'$'       â/ 0 k  1� ä ‰ @    b !r# #       âO    ÿàÿ ä ‰       ¢ !r)&)       Æ� ‰/, €    â !r       Æ‰ @    " !r+(+       Æ‰ @    b !r-*-       Æ ‰ @    b !r!!       Æ ‰ # @    " !r       Æ/ ‰% @    b !r1        âO    ÿàõ ä ‰"' @    ¢ !r/,/       â� x        ä ‰$) @    â !r       â,ÿÿ   ÿä â %v j   Ž â ‰       " !r       â j  ÿ„ Ä ‰&+ @    b !r       â  k  , ä ‰(- @    b !r #        Æ ‰      " !r%       Æ/ ‰*/ @    b !r"'"       ÆO ‰      ¢ !r$)$       Æ� ‰      â !r3        Æ‰      " !r&+&       Æ‰ !      b !r(-(       Æ ‰       b !r)       Æ ‰#      " !r*/*       Æ/ ‰%"      b !r+       ÆO ‰$      ¢ !r       Æ� †	 1    â !r       Æ† 3   â !r!!        Æ‰'&      â !r        Æ ‰(      " !r##       Æ ‰*      " !r%"%       Æ/ †	 )    â !r$       ÆO † +    è †	 
    è † 
   â !r'&'       Æ� †	 !    â !r(       Æ†     â !r*       Æ †	 #    è † %   è †	     è † '    è †	     â M™        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y       "   " ¹z  l      â Ey  p    € â ™x?    â y      !   b ¹z  F   
   â $rÿÿ   ÿ Ž â r ÿ       â y      %   ¢ x      ÿ  â r! ÿ       â $r"ÿÿ   ÿ Ž â r ÿ       â $rCÿÿ   ÿ Ž â r	 ÿ       â $r$ÿÿ   ÿ Ž â rF ÿ       Ä r ÿ       ä r  ÿ   pRð är  ÿ   pRò är    ÿ(Ž È/ r   ÿ� ä x   ÿ   äO xÿ     È |    pbô Ú G)  ð    € ê y      '   " r ÿ       â $rÿÿ   ÿ Ž â xþÿÿÿÿÀŽ â y      &   b z  Y      ä z q  ÿàÿ â 'z y   Ž â r  €ÿàÿ Æ 'z s   Ž â/ zÿ z    ä x   ÿ   â $z }  ÿŽ â zÿ t    â $z p  Ž âx)   ÿ   â $z |  Ž âxÿ     â $z |  Ž â x   ÿàÿ â $z w  ÿŽ â x   )áÿ â $z v  Ž â x     â %x()   ÿŽ â x#ÿ     Ä r   ÿàõ â $z v  Ž â z  q      â $z p  Ž â x   #  â %x*   (Ž â r ÿ       ä x   ÿ   â $vÿ X  ÿ Ž â rÿÿ   ÿä â $x   Ž Ä $x   Ž ä$x   Ž ä$x   Ž ä $x   ÿ Ž ä %x(   (Ž È r        är,*   ÿàõ Æ %x4   Ž â r-+   ÿä È �y%4   • ¨  �yB4   • ¨  �y ,   • è �y#,   • è �yA,
   • ( �y&,   • " %x2   Ž â r0(   ÿàõ Ä xM $       â �y4    • d $x1)    ä �y2   • ( �yD2   • b r.   ÿàõ Æ �y@0   • h �y?0   • h �yH0
   • h �y'0   • b rÿ   ÿä â %v.. Z  M Ž ârJ   ÿà÷ â �y
2    • d $x-$   ÿŽ â/ r,.   ÿàõ â %vJJ Z  M Ž â rÿ   ÿäÿ Æ $x//   -Ž ä $x1$   ÿŽ Æ r-/   ÿä â �yE.    ™ b r0.   ÿàõ â $xKK   1Ž ä �y<,   ™ b r8J   ÿà÷ ä r1/   ÿä â �y>,   ™ b r9K   ÿäÿ Ä r4J   ÿàõ â �y=0   ™ b r6   ÿà÷ Æ �y8   ™ b $x5K    â r2   ÿàõ ä �yG,   ™ b r0ÿ   ÿäÿ Æ �yI8   ™ b r.ÿ   ÿä â %v66 Z  M Ž ä�yL4   ™ d $x/0$   ÿŽ â r06   ÿàõ â%v,2 Z  M Ž â/ �yN8   ™ f $x;.$   ÿŽ â �yJJ    ™ b $x77   /Ž â r.6   ÿà÷ Ä r;-;   ÿàÿ ä r4,   ÿàû â �y66    ™ b r17   ÿä är2,   ÿàù ä r/7   ÿäÿ ä r: ,       â $x5;   Ž â r8   ÿàõ â/�yR.   ™ b r,   ÿà÷ Ä r-ÿ   ÿä â�yS4   ™ b rVÿ   ÿäÿ â $x3;    ä �yQ0   ™ b %v88 Z  M Ž Æ �yP0   ™ b $x/-$   ÿŽ ä/ %v,, Z  M Ž â �yT0   ™ f $x5V$   ÿŽ â �yK2   ™ b $x99   /Ž â r.8   ÿàõ Ä r--5   ÿàÿ â �yO2   ™ b r0,   ÿàù ä/ r/9   ÿä â �yU2   ™ d $x1-    â r4,   ÿàõ â�y::    ™ h �y7.   ™ b r28   ÿà÷ Ä r5-   ÿä â�yV.   ™ b r39   ÿäÿ Æ �y;0   ™ h# �yM0   ™ h �y22   ™ h �y44   ™ h �yW.   ™ h �yX0   ™ h �y,,    ™ h �y88    ™ b $xY   ÿ Ž â xB%T  B    ÈO r%ÿY   B  È x3%   ÿ   ä x# T  #    ä� x.3   ÿÀŽ ä r ÿ   B  ä x5ÿ   %  ä x/ÿ   #  Ä x..#øŽ ä x-    ÿ   ä x#ÿ      ä x05   ÿÀŽ ä/ x&AT  &    äx--   ÿÀŽ Ä x1#   ÿÀŽ â $x#%   ÿ Ž â x/0/øŽ ä x0ÿ   &  äx--&øŽ ä x&ÿ   %  ä x010øŽ Ä x#.   #øŽ ä x.    ÿ   ä x1ÿ      ä x//   &øŽ ä x&%   ÿ   ä x--   .øŽ â $x.%   ÿ Ž â x10   1øŽ Ä x0    ÿ   ä x&#   &øŽ â $x#%   ÿ Ž â x0-   0øŽ ä x./   .øŽ ä x-%	   ÿ   ä xDT  D    Äx&&   #øŽ ä x#.   -øŽ ä xB    ÿ   är-ÿ   D  äx/    ÿ   ä rYÿY   D  Ä x%1   BøŽ â $x1-   ÿ Ž â x0   /øŽ ä x?@T  ?    ä x.Y   ÿ   ä x0ÿ   Y  ä x'HT  '    Ä x@1   ÿÀŽ ä x..   ÿÀŽ ä x/ÿ   ?  ä x00   ÿÀŽ ä x@@'øŽ äx3ÿ   '  Ä x..?øŽ ä x/0/øŽ ä x'Y   ÿ   ä x0ÿ   Y  ä x1ÿ   -  ä x'.   'øŽ Ä x//   0øŽ â $x0Y   ÿ Ž âx.Y   ÿ   ä xB1   ÿÀŽ ä x1-   ÿ   ä x.'   .øŽ ä x'Y   ÿ   Ä x1@   1øŽ ä x//   0øŽ â $x0Y   ÿ Ž â x3B3øŽ ä x@-   ÿ   ä xBÿ   -  ä x..   'øŽ Ä x11   @øŽ ä x33   BøŽ ä x@-   ÿ   ä x//   0øŽ â &r0&<   ÿ   âxB-   ÿ   â &r<.<   ÿ   â x'1   @øŽ â &r0#>   0   âx  	   ÿ   â &r</>   <   â x33   BøŽ Ä x@-	   ÿ   â &r0=   0   âx%%    øŽ â &r<'=   <   â x 3   @øŽ â &r>&   ÿ   ä0r-ÿE  0 A   ä &r.   ÿ   ä &r0%G   0   Ä &r< G   <   ä &r>#I   >   âEr0 0       â &rI/I      â  x-   A  A  â 0rEÿE    A   â Er- <       â &r>L   >   Æ #r1E0     â#rE-     â &r>%N   >   ä 0r-ÿJ  0 A   ä&rI'L   I   â Er3 >       â 0rJÿJ    A   â  x5-   A  A  â &rI N   I   Ä 0rÿ    A   ä 0r-ÿ
    A   â #r
J3   5  â Er0 I       Æ #r
      â &r
&Q   ÿ   ä&rQ.Q   ÿ   â #r-      â #rJ0   5  â &r
#P   
   ä&r0&K   ÿ   ä &rQ/P   Q   Ä &r<.K   ÿ   ä &r
R   
   ä&r0#O   0   ä&rQ'R   Q   â #rF-   F   â &r>/O   <   ä &r
%T   
   Ä &r<S   0   ä &rQ T   Q   ä 0r0ÿ6  0 A   ä&r&7   ÿ   â #r$1   $   â &r7.7   ÿ   ä &r&&;   ÿ   âEr1 
       â &r..;   ÿ   â Er
 Q       â 0r6ÿ6    A   â  x30   A  A  â &r#V      Ä &r7/V   7   ä &r&#M   &   ä&r./M   .   â #r
6
   3  â &r>'S   >   â x    ÿàÿ â &r2      ä&r7'2   7   Ä &r&4   &   ä&r''4   .   â #r	-
   	   â &r> U   >   â|    pbô â 0r5ÿ:  0 A   ä &r%W      ä&r7 W   7   Ä &r&%X   &   ä&r<%U   <   ä &r  X   '   ä 0r
ÿ,  0 A   ä 0rÿ8  0 A   â  x95   A  A  â 0r,ÿ,    A   â Er#        â 0r:ÿ:    A   â Er% &       â 0r8ÿ8    A   â Er5 <       â  x'
   A  A  â Er> >       â  x   A  A  â Er 7       Ä Er          â #r%,%   '  â xÀ  ÿà÷ â #r161   3  â #r5:5   9  â#r>:>   9  â #r#8#     â#r8     â #r',    '  â #rC1   C   â#r5      â#r"->   "   â#r!#   !   â#r-      â#r%      â #r-'      â x    ÿàÿ â $rÿÿ   Ž â x    ÿàÿ Ä x    ÿàÿ ä x    ÿàÿ ä x    ÿàÿ â D)       À â Gy  °ìÿÿÿÿƒ ê Ay        € ê Ey  ð     € â G‰  Ð     € ê $x  €  Ž Ê ˆs  $ úÿ H   è ˆs  €úÿ H   è ˆs   ûÿ H   è ˆs  F€ûÿ H   è ˆs  C üÿ H   è ˆs  	€üÿ H   è ˆs   ýÿ H   è ˆs  "€ýÿ H   è ˆs  ! þÿ H   è ˆs  €þÿ H   è ˆs   ÿÿ H   è ˆs  €ÿÿ H   ä Ay        € ê {          ì M        € ê „y  €   H   " z  r  p`ð Ä z+  r      â „y     H   b r ÿ   p"p Æ „y    H   ¢ $x'+   ÿŽ ä$x)+   ÿŽ â „y    H   è „y €  H   ( „y
    H   h „y €  H   h „y €  H   h „y €  H   h „y    H   è „y    H   â !r        â !r$       Ä/ „y €  H   â !r       ÆO ‰      " !r       Æ� ‰       b !rFF       â!r

C       â!r	       ä ‰	      ¢ !r       Æ ‰      â !r""       â !r       ä ‰      " !r!       â !r         ä/ ‰      b !r       Æ ‰F      " !r       Æ ‰       h ‰
      b !r		       âO x   ÿàÿ ä ‰      ¢ !r       â� z  r  p`ò Ä ‰"      â !r       â x    p"ò ä ‰      " !r       Æ/ ‰      b !rF       Æy      &   " !r        â!r
       ä y      '   b !r       ÆO ‰ €    ¢ !r""       Æ� ‰	      â !r       Æ ‰      " !r       Æ/ ‰ €    b $z  x  Ž Æ‰      ( ‰      " $z  ~   Ž Æ‰      b !r
       âO  '   ÿàõ ä ‰      ¢ !r	       Æ� ‰"      â !r       â $ÿÿ   ÿ â 
 j  ÿ„ â ‰      " !r       â/ r+    ÿˆ ä ‰      b !r       ârÿÿ   ÿä Ä ‰$      " !r       â 
  r  ÿà÷ â !r       â‰
 @    d ÿÿ   ÿäÿ â !r       âO ‰ €    ¢ 
 j  ÿ† â !r""       Ä� ‰ €    â z j  ÿˆ â!r       ä ‰	 @    " 
 k  � â !r        ä/ ‰ €    b z k   â !r$$       Ä‰ €    " 
 k   Æ ‰ €    " !r
       âr+    ÿ„ ä ‰! €    b !r       âO r&ÿÿ   ÿä ä ‰" €    ¢ !r       â� z
 j  ÿ„ Ä ‰ €    â !r	       â z k  & ä ‰#  €    " !r       â/ 	+   ÿàÿ ä ‰%$      b !r       â )   ÿàù â !r       â ‰ @    " ( 	   ÿà÷ â !r!!       Ä‰ @    b ÿÿ   ÿä â !r"       äO ‰      ¢ 
 j  ÿˆ â !r       ä� ‰ @    â +ÿÿ   ÿäÿ â !r# #       Ä ‰ @    " ( j  ÿ† â !r%$%       ä/ ‰ @    b 
	 k   ä ( k  +� â ‰"! @    b !r       â$ÿ   ÿ Ž Ä ‰ @    " !r       Æ‰  @    b !r+       âO '   ÿàÿ ä ‰      ¢ !r       Æ� ‰$# @    â !r       Æ ‰&% €    " !r       â/ !r"!"       â !        â ‰      b !r       Æ‰      " !r         Æ‰      b !r-       âO     ÿàõ ä‰      ¢ %  j   Ž â!r$#$       â� (ÿÿ   ÿä â ‰      â %  j  ! Ž â !r&%&       â  j  ÿ„ â‰      &  k  ( â ‰!       " !r       Æ/ ‰#"      b !r       Æ‰'& @    " !r       Æ‰%$      b !r       ÆO †	 +    â !r       Æ� † -   â !r       Æ †	     â !r! !       â !r#"#       ä/ †     â !r'&'       Æ†	     â !r%$%       Æ†    è †	     è †     è †	 
!    è † 
#   è ‰'      ( †	 %    â M™        € ê x)   ÿàÿ âO !r''       Æ r    ÿàñ È r ÿÿ   ÿä  ä z j  ÿ€ È z k     Ê †y '    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y       "   " ¹z  l      â Ey  @    € â ™x?    â y      !   b ¹z  F   
   â r ÿ       â $rÿÿ   ÿ Ž â y%      &   ¢ r ÿ       â $r ÿÿ   ÿ Ž â r ÿ       â y'      '   â x      ÿ  â $r5ÿÿ   ÿ Ž â r: ÿ       â $r9ÿÿ   ÿ Ž â r  ÿ   pRð Ä r  ÿ   pRò â$x     Ž Ê/ r   ÿ� È xÿ     È |    pbô Ú G)  à    € ê y      %   "Î r& ÿ       â $r$ÿÿ   ÿ Ž â xþÿÿÿÿÀŽ â $z' }  ÿŽ âz  q      â 'z
' y  & Ž â z q  ÿàÿ Æ 'z% s  $ Ž â z
ÿ z  
  Æ $x   
Ž â zÿ t    â $z
 |  Ž â r   ÿŽ â $x   ÿ Ž âx)   ÿ   â $z% w  ÿŽ â x   ÿ   â $z v  Ž â x   )áÿ â %x()   ÿŽ â x   ÿ   Ä r   ÿàõ â $x   Ž âxÿ     â $z p  Ž â xÿ     â $rÿÿ   ÿ Ž â x     â $z
 |  Ž â x   ÿàÿ â $x   Ž â x     â $z v  Ž â x   ÿ   â $z p  Ž â z  Y      â %xN   (Ž È $vÿ X  ÿ Ž ä $rÿÿ   ÿ ä $x   ÿ Ž ä %x(   (Ž È r
        â$rÿÿ    Ž â r0N   ÿàõ Æ %x"   
Ž â r1O   ÿä È �y*"   • ¨  �y"   • ¨  �yK0   • è �yJ0   • è �yL0
   • ( �y/0   • " %x
   
Ž â r,(   ÿàõ Ä xP $       â �yB"    • h �yG
   • h �y
   • b $x-)    â r3   ÿà÷ Ä �y8
    • h �yF,   • h �y+,   • h �yE,
   • h �yD,   • b r?   ÿàõ Ê $rÿÿ    ä %v>? Z  P Ž È $x$   ÿŽ â r6>   ÿàõ È r??   ÿàÿ È r7?   ÿä â %v23 Z  P Ž â r<>   ÿàõ â �y@>    ™ h $x=?    â r   ÿàõ â �y16   ™ f# rÿ   ÿä â�y-6   ™ b rÿ   ÿäÿ â %v" Z  P Ž Ä �yC<   ™ d $x$   ÿŽ ä $x$   ÿŽ â �yA6   ™ b rH"   ÿàù är##   ÿàÿ â $x33   Ž â rT2   ÿàõ Æ $xI#    â rU3   ÿä â �y;"    ™ b rV"   ÿàõ ärX2   ÿà÷ â �y7H   ™ b# rW#   ÿä Æ �y6T   ™ b rY3   ÿäÿ Æ �yH   ™ b rS   ÿàõ Æ �y
T   ™ b rQ   ÿà÷ Æ �y>V   ™ b $rÿÿ    â rÿ   ÿäÿ ä �y?X   ™ h �y<H   ™ b %vRS Z  P Ž Æ �y=T   ™ b $x$   ÿŽ Æ �y2    ™ b %vPQ Z  P Ž È $x!$   ÿŽ â rVR   ÿàõ ä/rSS   ÿàÿ â $xQQ   !Ž â rTP   ÿàù ä rWS   ÿä â�y!R    ™ b rUQ   ÿä ä rHR   ÿà÷ â�y"V   ™ b rXP   ÿàõ Æ �yT   ™ b $xIS   Ž â rYQ   ÿä ä �y4V   ™ h �y2T   ™ h �y0H   ™ h �y.X   ™ h �y,V   ™ h �yT   ™ h �y#P    ™ b x**T      ÈO rÿ   *  är*ÿ   *  ä x3   ÿ   ä xKKT  J    â� $xI*   ÿ Ž â xJÿ   *  ä xH3   ÿÀŽ Ä xLLT  /    äxJJ   ÿÀŽ ä x3ÿ   L  ä xII   ÿÀŽ ä xHHKøŽ ä x/   ÿ   Ä xMÿ     ä x3J3øŽ ä xIILøŽ ä xJ*   ÿ   ä xPH   /øŽ ä/ xKÿ   K  Ä xMM   ÿÀŽ ä xHÿ   *  â $x/   ÿ Ž â xII   JøŽ â $xJ*   ÿ Ž â xKMKøŽ ä xLÿ     ä x33   HøŽ Ä xGT      äxKK   LøŽ ä xPP   /øŽ ä x33   JøŽ ä xL   ÿ   ä xH*   ÿ   Ä x/   ÿ   äxJ	   ÿ   ä rÿ     ä xHI   HøŽ ä xKK   LøŽ ä xI   ÿ   Ä rÿ     â $xG*   ÿ Ž âx**	   ÿ   ä xFFT  +    ä xKK   JøŽ â $xJ   ÿ Ž â x+I   ÿÀŽ ä xEET  D    Ä xDÿ     ä x*3   *øŽ ä x++FøŽ äx3ÿ   F  ä xFD   ÿÀŽ ä xDJ   ÿÀŽ Ä xGH   GøŽ ä xHÿ   E  äxEDEøŽ ä x3F3øŽ ä xDÿ     ä xF   ÿ   Ä x33   DøŽ â $xD   ÿ Ž â xIÿ     ä x++   FøŽ ä xF   ÿ   ä xII   ÿÀŽ ä xD+   DøŽ â $x+   ÿ Ž â xEE   FøŽ Ä xJ   ÿ   ä xF   ÿ   ä xHIHøŽ ä x/P   /øŽ ä xIÿ     ä xEE   JøŽ Ä x33   FøŽ ä x+D   +øŽ ä xJ	   ÿ   â $xD   ÿ Ž â xHH   IøŽ â &rF/1   ÿ   âxI   ÿ   ä x33   JøŽ â &rJ+1   ÿ   â xHH   IøŽ â &rFK-   F   âxEE   DøŽ â &rJ3-   J   â x	   ÿ   â &rFGC   F   Ä &rJEC   J   â xHH   øŽ â &rF*A   F   ä0rÿ@  0 A   ä&rJHA   J   â ErF F       â 0r-ÿ@    A   â  x   A  A  ä Er1 J       â 0rÿB    A   â #rF-F     â &r@/7   ÿ   Ä #r--1     â &rB+7   ÿ   ä &r/6   ÿ   ä&r6+6   ÿ   ä &r@K   @   ä&rB3   B   ä &rK
      Ä &r63
   6   ä &rG>   @   ä&r1E>   B   ä &rG?      ä&r6E?   6   ä &r7*<      Ä &r<H<   1   ä &r*=      ä0r1ÿ;  0 A   ä &r6H=   6   ä 0r
ÿ  0 A   âEr        â 0rÿ8    A   â  x81   A  A  â 0rÿ    A   â  x

   A  A  â Er1 6       Æ #r   
  ä#r
1   
  â 0r;ÿ;    A   â Er6 7       ä #r	
   	   â &r
/"   ÿ   ä&r//   ÿ   ä &r"+"   ÿ   Ä &r+   ÿ   â #r;6   8  â &r
K4   
   ä&r/K2   /   ä &r"34   "   ä&r32      â #r      â &rG0   
   â x    ÿàÿ â &r/G.   /   Ä &r+E0   "   ä&r.E.      â |    pbô â &r*,      â #r5   5   â &r+H,   +   ä &r/*   /   ä&r.H   .   Ä 0rÿ!  0 A   ä 0rÿ#  0 A   â Er
        â 0r!ÿ!    A   â Er7 <       â 0r#ÿ#    A   â Er+ +       â  x   A  A  â Er" /       â  x   A  A  â Er .       â #r8;7   8  â #r
!
     â#r+!+     â #r"#"     â#r#     â xÀ  ÿà÷ â #r9F   9   â #r:-   :   â#r8      â #r 
       â #r+      â #r"      â #r      â x    ÿàÿ Ä x    ÿàÿ ä x    ÿàÿ ä x    ÿàÿ ä rÿ   ÿäÿ â D)       À â Gy  Àîÿÿÿÿƒ ê Ay        € êÏ Ey  Ð     € â G‰  °     € ê $x @  Ž Ê ˆs 9 ûÿ H   è ˆs :€ûÿ H   è ˆs 5 üÿ H   è ˆs 	€üÿ H   è ˆs  ýÿ H   è ˆs €ýÿ H   è ˆs   þÿ H   è ˆs €þÿ H   è ˆs  ÿÿ H   è ˆs €ÿÿ H   ä Ay        € ê {          ì M        € ê „y      H   b z!  r      Æ „y €   H   ¤ $x!   ÿŽ ä „y    H   è „y €  H   ( „y    H   h „y €  H   ( „y    H   ( „y    H   ( „y €  H   è „y €  H   â !r  9       Æ/ y      %   b !r:       ÆO ‰       " !r55       Æ� ‰      ¢ !r		       Æ‰5      â !r       â!r       È ‰      " !r        Æ ‰	      b $x   ÿ Ž â/ !r       â !r        â ‰
      " $z&% x  Ž â !r       â z  r  p`ð â ‰       b !r       âO !r       â r ÿ   p"p â ‰      ¢ !r5       â� $z&' ~  &Ž Ä ‰      â $t'ÿ   ÿ Ž â !r       ä‰      " !r	       Æ‰      b !r

       Æ ‰	      " !r         Æ/ ‰      b !r       ÆO ‰      ¢ !r       Æ� ‰      â !r       Æ‰
      " !r       Æ‰  €    h ‰      b !r		       Æ ‰      " !r       Æ/ ‰      b !r       ÆO ‰      ¢ !r       â� !r
       ä‰ €    â x
   ÿàÿ â !r        ä‰	 €    " z 
 r  p`ò â !r       Ä ‰ €    b x    p"ò â !r       ä ‰ €    " r!&   ÿ† â !r       ä/ ‰  @    b !r       ÆO ‰ €    ¨ ‰ €    ¢ !r
       â� r!&   ÿŠ ä ‰ €    â !r	       â
	& r  ÿàù â $r ÿÿ   ÿŽ â ‰ €    " !r       âz j  ÿŠ Ä ‰ €    b !r       â "ÿÿ   ÿä â !r         â/ ‰
 @    " &   ÿàõ â !r       äO ‰       b ÿÿ   ÿä â !r       Ä ‰ @    ¢ 
 j  ÿ„ â !r       ä� ‰ @    â 
 k   â !r       ä‰ @    " z k   � â !r       Ä‰ @    b rÿÿ   ÿäÿ ä z j  ÿ† â ‰ @    b !r
       â 
	 j  ÿˆ ä ‰ @    " z k  � â !r#        â/ !   ÿàÿ â ‰ @    b !r       âO 
		 k  " Ä ‰ @    ¢ !r       â� &   ÿàõ â !r       â‰       ä  ÿÿ   ÿä â !r       â‰      " 
 j  ÿ„ â !r       Ä ‰      b  k    â !r       â    ÿàÿ â ‰      " !r       ä/ &   ÿà÷ â‰      b %v&& j  ' Ž â !r       âO ÿÿ   ÿäÿ â ‰      ¢  j  ÿ† â !r!        Ä� ‰      â  k  � â !r       ä‰      " !r       Æ‰      b !r       Æ †	 &#    â !r       Æ/ † &!   â !r       ÆO †	     â !r       Æ� † 
    â !r       Æ†	     â !r       Æ†    è †	     è †     è †	     â M™        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y       "   " ¹z  l      â Ey  Ð    € â ™x?    â y      !   b ¹z  F   
   â $rÿÿ   ÿ Ž â r ÿ       â y      %   ¢ r ÿ       â $rÿÿ   ÿ Ž â r ÿ       â y!      &   â r ÿ       â $r
ÿÿ   ÿ Ž â r5 ÿ       â y#      '   " r  ÿ   pRð Ä r  ÿ   pRò är    ÿ(Ž È/ r   ÿ� ä x   ÿ   äO xÿ     È |    pbô Ú G)  `    € ê $r"ÿÿ   ÿ Ž â�r  ÿ       â $z# }  ÿŽ â xþÿÿÿÿÀŽ â 'z# y  " Ž â z  Y      ä r  €ÿàÿ â 'z! s    Ž â zÿ z    ä x$   ÿ   â $z p  Ž â zÿ t    â $z |  Ž Ä $x   ÿ Ž ä $z |  Ž â x   ÿàÿ â $z v  Ž â x   $áÿ â $z	! w  ÿŽ ä $z p  Ž â z  q      â %x$$   ÿŽ â xÿ     Ä r	   ÿàõ â $z v  Ž â x     â $x   Ž âr ÿ       â $x   Ž â z q  ÿàÿ â %xF   $Ž â rÿÿ   ÿä ä x   ÿ   â $vÿ X  ÿ Ž Ä %x$   $Ž È $rÿÿ    Ž â r*F   ÿàõ Æ %xL   Ž â r+G   ÿä È �y)L   • ¨  �y?L   • ¨  �yD*   • è �yA*   • è �y'*
   • ( �y=*   • " r$   ÿàõ â %xB   Ž Ä �y	L    • b r%   ÿä Æ �y@B   • h �y0B   • h �y>   • h �y-   • h �y9
   • h �y<   • b r;   ÿàõ â $tJÿ$   ÿ Ž Æ rÿ   ÿä â %v:; Z  J Ž â r/   ÿàõ Æ $x3$   ÿŽ â rÿ   ÿä â %v./ Z  J Ž â r2 :       är::   ÿàõ â $x3;   3Ž ä $x$   ÿŽ ä %x6   2Ž â rH .       â �y42    ™ b rI/   ÿàÿ â $x;3    Ä �y*6   ™ d# %x   HŽ ä �y+6   ™ b r..   ÿà÷ ä r   ÿàõ â �y;:   ™ b r/I   ÿäÿ Æ �y3   ™ b rK   ÿà÷ â $rÿÿ    ä �y86   ™ h �y:.   ™ b $x$   ÿŽ Æ �yH    ™ h �y7   ™ b# %v. Z  J Ž â rÿ   ÿäÿ ä �y6   ™ b %vJK Z  J Ž È $x$   ÿŽ â r/   ÿàÿ ä/ r .       â�yB    • b r..   ÿàõ Æ %xL   Ž â rB J       Æ $xCK   Ž â �y2L   ™ f %xH   BŽ â �yL   ™ f $x/    â rJJ   ÿàõ â �y&H   ™ f rKC   ÿä â �y1.   ™ h �y    ™ h �y,J   ™ h �y/H   ™ h �y(L   ™ h �yH   ™ h/ �yB    ™ b x.   ÿ   Ä x@   ÿàÿ È |    pbô ä x))T  ?    ÈO r?ÿ.   )  ä xDDT  A    Æ� $xB?   ÿ Ž â xAÿ   ?  ä r)ÿ   )  ä xCÿ   D  ä xAA   ÿÀŽ ä xBB   ÿÀŽ ä x''T  =    Äx=)   ÿ   ä xEÿ   )  ä xAACøŽ ä xBBDøŽ ä xC?   ÿ   ä x==   ÿÀŽ Ä xDÿ   '  ä xEE   ÿÀŽ ä xCB   CøŽ â $xB)   ÿ Ž â x=='øŽ ä xDEDøŽ ä x'ÿ   ?  Ä xE?   ÿ   ä x==   BøŽ ä xAA   'øŽ ä xB?   ÿ   ä xEC   EøŽ â $xC)   ÿ Ž â x'ÿ   )  Ä xBA   BøŽ ä x'D   'øŽ ä xA)   ÿ   ä x==   CøŽ ä xC@T  0    äx''   AøŽ Ä xA)   ÿ   ä x>>T  -    ä r.ÿ.   C  är-ÿ   C  ä x==   AøŽ ä x<9T  <    â $x9-   ÿ Ž â xAÿ   .  Ä xC.   ÿ   ä xD?   ÿ   â$x??   ÿ Ž â x0ÿ   >  ä xAA   ÿÀŽ ä xCC   ÿÀŽ ä x99   ÿÀŽ Ä x0A0øŽ ä x?B   ?øŽ ä xAÿ   .  ä xCC>øŽ ä x99<øŽ ä x@.   ÿ   Ä xB-   ÿ   ä x>ÿ   -  ä x00   AøŽ ä x@C   @øŽ â $xC.   ÿ Ž â xB9   BøŽ ä xA-   ÿ   Ä x<ÿ   <  ä x>>   ÿÀŽ ä xBB   AøŽ â $xA.   ÿ Ž â x<><øŽ ä x9ÿ   -  ä x@@   CøŽ Ä xDE   DøŽ ä x>.   ÿ   ä x9<   9øŽ ä x<-   ÿ   ä x@@   AøŽ ä xE.	   ÿ   â &r.D*   ÿ   â x00   >øŽ Ä x99   <øŽ â &r<@*   ÿ   â x*)	   ÿ   â $xC-   ÿ Ž â x00   EøŽ â &r.?+   .   â x''   *øŽ â &r*D3   ÿ   â xBB   CøŽ â &r.=;   .   â x>-	   ÿ   â &r<0+   <   Ä &r3@3   ÿ   â x99   >øŽ â 0r)ÿ4  0 A   ä&r.'8   .   ä&r<B;   <   ä &r*?7   *   ä&r307   3   â  x+)   A  A  â &r<98   <   â Er) .       â 0r4ÿ4    A   Ä &r-=:   *   ä&r3B:   3   â Er< <       â #r*4)   +  â&r-'6   -   ä&r396   3   ä 0r)ÿ  0 A   â#r<4<   +  â 0rÿ    A   â Er. -       Ä Er3 3       â  x4)   A  A  â 0r	ÿ	    A   Æ #r..   4  â#r33   4  â &rD   ÿ   ä&r@   ÿ   â #r5	*   5   â &rDD&   ÿ   ä&r&@&   ÿ   ä &r*?2      Ä &r02      ä&rD?/   D   ä&r/0/   &   ä &r*=1   *   ä&r1B1      ä 0rÿ  0 A   Ä &rD=,   D   ä&r/B,   /   â  x&   A  A  â &r*'(   *   ä&r19(   1   ä &rD'   D   ä&r/9   /   Ä 0rÿ  0 A   â Er* *       â 0rÿ    A   â Er1 1       â 0rÿ    A   â ErD D       â  x   A  A  â Er/ /       â 0rÿ    A   â x€  ÿà÷ â #r**   &  â#r11   &  â #rDD     â#r//     â #r
<   
   â #r	.      â #r3      â #r	*      â #r1      â #r	D      â #r/      â x@   ÿàÿ â $rÿÿ   Ž â x@   ÿàÿ Ä x@   ÿàÿ â G©  Ððÿÿÿÿƒ ê Ay        € ê�Ey  °     € â G‰  �     € ê r    ÿ@Ž Ê ˆs 5 üÿ H   è ˆs 
€üÿ H   è ˆs  ýÿ H   è ˆs €ýÿ H   è ˆs  þÿ H   è ˆs €þÿ H   è ˆs  ÿÿ H   è ˆs €ÿÿ H   ä Ay        € ê {          ì M        € ê „y      H   b z  r  p`ð â $z!! x  Ž â x   ÿàÿ â „y	    H   ¢ r ÿ   p"p â$z!# ~  !Ž â z  r  p`ò â „y    H   & x    p"ò â „y    H   è „y €   H   ( „y    H   h „y €  H   h „y €  H   h „y    H   h „y €  H   h „y    H   b !r  5       Æ/ „y €  H   b !r		       ÆO „y €  H   ¢ !r         Æ „y    H   " !r		       Æ� „y" €  H   â !r
       Æ„y&    H   " !r         Æ„y  	  H   b !r       Æ „y €	  H   b !r       Æ „y   
  H   b !r       Æ „y$ €
  H   b !r       Æ „y(    H   b !r       Æ „y €  H   â !r       Æ/ „y €  H   â !r       âO z  r      ä „y €  H   â !r       Æ ‰       " !r"       â� !r&       ä‰      b !r		       â!r       È ‰	      ¢ $x   ÿŽ â !r        ä ‰
      â !r$       Æ ‰      " !r(       Æ ‰      b !r        â !r       ä ‰      " !r       â/ !r       ä ‰       d !r       â !r	       âO ‰      ¢ !r

       Æ� ‰      â !r       Æ‰	
      " !r       Æ‰      b !r       Æ ‰      " !r         Æ/ ‰      h ‰      b !r       ÆO ‰  €    ¢ !r       Æ� ‰ €    â !r	
	       â!r       ä‰ €    " !r       Æ ‰	 €    " !r       Æ/ ‰
 €    b !r       Æ ‰ €    b !r        ÆO ‰ €    ¢ !r       Æ� ‰      è ‰  @    â !r       Æ‰ @    " !r	       â !r

       ä/ ‰	 @    " !r       Æ‰ @    b !r       ÆO ‰
 @    ¢ !r       Æ� ‰ @    â !r         â 
! r  ÿà÷ ä ‰ @    b !r       âr!   ÿ„ ä ‰ €    " $ÿÿ   ÿŽ â 
 j  ÿ† â !r	       â ‰       " rÿÿ   ÿä Ä z j  ÿ„ â !r       â/ 
 k  � â ‰      â z k   â !r
       âO    ÿàÿ â ‰      b !   ÿàõ â!r       â� !   ÿà÷ â ‰      ¢ x        â !r       âÿÿ   ÿäÿ â ‰      â 
 j  ÿ† â !r       â%v! j   Ž â ‰      " !r        Ä $ ÿÿ   ÿ â ‰ @    " 
 j  ÿ„ Æ ‰      b 
 k    â !r       â/  k  � â !r       â †	     â !r       ÆO †    â !r       Æ� †	     â !r       Æ† 
    â !r       Æ †	     â !r       Æ†    è ‰      ( †	     â M™        € ê x    ÿàÿ â !r       Æ r !    ÿàñ È rÿÿ   ÿä  ä/ z  j  ÿ€ È z  k    Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      "   " ¹z  l      â Ey  €    € â ™x?    â y       !   b ¹z  F   
   â r ÿ       â $rÿÿ   ÿ Ž â y      %   ¢ r ÿ       â $rÿÿ   ÿ Ž â r ÿ       â y      &   â $r0ÿÿ   ÿ Ž Æ y      '   " r ÿ   pRð är ÿ   pRò â$x
     Ž â/ x   ÿ   ÈO r

   ÿ� Ê $À    Ž â xÿ     È |    pbô Ú G)       € ê $rÿÿ   ÿ Ž â�r ÿ       â $z }  ÿŽ âxþÿÿÿÿÀŽ â 'z y   Ž â z  Y      Æ 'z	 s   Ž â zÿ z    Æ $x

   
Ž â z	ÿ t  	  â $z p  Ž ä$x 
   ÿ Ž âx
   ÿ   â $z |  Ž ä$z w  ÿŽ â x    áÿ â $z |  Ž â x   ÿàÿ â $z	 v  Ž â r   ÿàõ â %x     ÿŽ â xÿ     Ä z  q      â $z	 v  Ž â x     â $z p  Ž â r   ÿŽ â %x>    Ž â z q  ÿàÿ ä x   ÿ   â $vÿ X  ÿ Ž ä $rÿÿ   ÿ Ž Ä $rÿÿ   ÿ ä %x     Ž È $rÿÿ    Ž â r">   ÿàõ Æ %xL   Ž â r#?   ÿä È �y$L   • ¨ �y7L   • ¨ �y<"   • è  �y5"   • â  r&    ÿàõ Æ �y@"
   • ( �y"   • " %x*   Ž â xB $       ä �y2L    • b $x'!    Æ �y8*   • h �y-*   • h �y6&   • h �y)&   • h �y1&
   • h �y4&   • b rG   ÿàõ Ê $r	ÿÿ    ä %vFG Z  B Ž È $x		$   ÿŽ â r:F   ÿà÷ È rGG	   ÿàÿ Ê %x"   FŽ â rE   ÿàõ â�yF    ™ d $x;G   Ž ä �y%"   ™ b rC   ÿà÷ ä r	ÿ   ÿä â �y'"   ™ b %vDE Z  B Ž Æ �y3:   ™ b $rÿÿ   Ž ä $x		$   ÿŽ â �y/"   ™ b %vBC Z  B Ž È $x$   ÿŽ â rEE	   ÿàÿ Æ $xCC   Ž ä %xF   DŽ ârHB   ÿà÷ â�y*    • d %x:   BŽ â rJD   ÿàõ â �y#F   ™ h �y:   ™ b rKE   ÿä â $xIC   Ž Ä �y.F   ™ h �y,:   ™ h �y*J   ™ h/ �y(H   ™ h �y&F   ™ h �y":   ™ h �yB    ™ h �y	D    ™ b x+
   ÿ   Ä x@   ÿàÿ È |    pbô ä x€  ÿà÷ ä x@   ÿàÿ ä x@   ÿàÿ â $rÿÿ   Ž â x$$T  7    ÈO r7ÿ+   $  Ê $x97   ÿ Ž â r$ÿ   $  ä x<<T  5    ä� x99   ÿÀŽ ä x@@T      äx:ÿ   7  ä x5$   ÿ   Ä xÿ   <  äx99<øŽ â $x<7   ÿ Ž â x::   ÿÀŽ ä x55   ÿÀŽ ä x=ÿ   $  ä x99   <øŽ Ä x::øŽ ä x55@øŽ äx<$   ÿ   ä xÿ   7  ä x;ÿ   @  ä x==   ÿÀŽ Ä x-8T  -    â$x@7   ÿ Ž â x55   <øŽ ä x:   øŽ ä x<7   ÿ   ä x66T  )    ä x;=;øŽ Ä x:ÿ   $  ä r)ÿ   -  ä x<   <øŽ â $xB7   ÿ Ž â/ x:;   :øŽ ä x99   @øŽ â $x@$   ÿ Ž âx$   ÿ   Ä r+ÿ+   -  ä x41T  4    â $x1)   ÿ Ž â x:   øŽ â $x:$   ÿ Ž â x;+   ÿ   ä x55   @øŽ ä xB9   BøŽ Ä x8ÿ   +  ä x11   ÿÀŽ ä x9ÿ   )  ä x;;   ÿÀŽ ä x55   :øŽ â $x:)   ÿ Ž â x-ÿ   6  Ä x88   ÿÀŽ ä x114øŽ äx4ÿ   4  ä x99   ÿÀŽ ä x;;6øŽ ä x6+   ÿ   Ä x-8-øŽ ä x8ÿ   +  ä x494øŽ ä x:1   :øŽ ä x9)   ÿ   ä x6;   6øŽ Ä x;+   ÿ   ä x--   8øŽ â $x8+   ÿ Ž âx::   9øŽ â $x9+   ÿ Ž â x1ÿ   )  ä x66   ;øŽ ä x14   1øŽ Ä x--   8øŽ ä x77	   ÿ   ä x4)   ÿ   ä x8+	   ÿ   ä x66   9øŽ â $x+)   ÿ Ž â x7<   7øŽ Ä x11   4øŽ â &r4B%   ÿ   âx--   8øŽ â &r86%   ÿ   â x+:   +øŽ â &r47'   4   âx$$	   ÿ   â &r8-'   8   â x:)	   ÿ   â &r453   4   âx   $øŽ â &r8+3   8   â x11   :øŽ â 0r%ÿ  0 A   Ä &r4/   4   ä&r81/   8   â  x'%   A  A  â 0r$ÿ    A   â Er% 4       ä Er8 8       Æ #r%$%   '  ä#r8$8   '  â &r$B#   ÿ   ä&rBB   ÿ   ä&r#6#   ÿ   ä&r6   ÿ   ä &r$7.   $   Ä &rB7,   B   ä &r#-.   #   ä&r-,      ä &r$5*   $   ä&rB5(   B   ä &r#+*   #   Ä &r(+(      ä &r$&   $   ä&rB"   B   ä 0rÿ  0 A   ä &r#1&   #   ä&r(1"   (   Ä 0rÿ	  0 A   â  x"   A  A  â 0rÿ    A   â ErB B       â 0r	ÿ	    A   â Er$ $       â  x   A  A  â Er# #       ä Er (       â 0rÿ2    A   â #rBB   "  â 0rÿ    A   â #r$	$     â#r#	#     â #r"   "  â #r0%   0   â #r8      â #r$      â #r#      â #rB      â #r"      â G©   óÿÿÿÿƒ ê Ay        € ê�ˆ 0 ýÿ H   è ˆ €ýÿ H   è ˆ  þÿ H   è ˆ €þÿ H   è ˆ  ÿÿ H   è ˆ €ÿÿ H   è {          ì M        € ê „y      H   " z  r  p`ð â $z x  Ž Ä „y     H   b r  ÿ   p"p â $z ~  Ž ä „y	  €   H   ¨ „y     H   è „y  €  H   ( „y     H   h „y     H   h „y
  €  H   h „y     H   h „y  €  H   h „y  €  H   b !r0       Æ „y     H   " !r       Æ/ „y  €  H   b !r		       ÆO „y     H   ¢ !r       Æ� „y  €  H   â !r       Æ„y  €  H   â !r       Æ„y     H   â !r       Æ „y   €  H   â !r		
       â !r       ä ‰      " !r       â !r       â !r       â !r		       â/ !r       ÈO ‰	      " z  r      â !r       ä� ‰      b !r       â!r       ä ‰
      ¢ !r       Æ ‰      â !r        Æ ‰      ( ‰      b !r	       â !r       È/ ‰      " !r

       ÆO ‰	      b !r       Æ� ‰
      ¢ !r       Æ‰ €    â !r       Æ‰      ( ‰      b !r       â !r		       È/ ‰ €    " !r
       ÆO ‰	 €    b !r       Æ� ‰
 €    ¢ !r       Æ‰ @    â !r       Æ‰ €    ( ‰ €    b !r       â !r	       È/ ‰ @    " !r
       ÆO ‰ @    b x
   ÿàÿ â !r       ä� ‰	 @    ¢ z 
 r  p`ò â !r       ä‰      â x     p"ò â !r       Ä‰ @    ( ‰ @    b !r       Ê    ÿàÿ â !r
       â/ ‰       " 
 r  ÿàõ â!r	       âO    ÿà÷ ä ÿÿ   ÿä ä 
 j  ÿ„ â!r       â� $tÿ   ÿ Ž â ÿÿ   ÿäÿ â !r       â
 k   â ‰
      b r	   ÿ„ â !r       â j  ÿ† â ‰      ¢ rÿÿ   ÿä â %v j   Ž â z	 j  ÿ„ â ‰      â !r        Ä z	 k   â ‰      " 	 k  � Æ †	     è	 †    â	 !r
       â/ !r       ÈO †	     â	 !r       Æ� †     â	 !r       Æ †	     â	 M™        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      "   " ¹z  l      â Ey       € â ™x?    â y       !   b ¹z  F   
   â $rÿÿ   ÿ Ž â x      ÿ  â y      %   ¢ r ÿ       ä r ÿ   pRð Ä r ÿ   pRò â$x
     Ž Ê/ r

   ÿ� È x&ÿ     â $€    Ž ä $x   ÿ Ž âO | &   pbô Ú G)  �    € ê y      '   " $rÿÿ   ÿ Ž â r ÿ       â $v"ÿ X  ÿ Ž â y	      &   b xþÿÿÿÿÀŽ â $z p  &Ž È $x
   
Ž È $x   ÿ Ž âx   ÿ   È x   áÿ â%x   ÿŽ â xÿ     ä xÿ     â 'z y   Ž â x     ä x     â 'z	 s   Ž â/ zÿ z    Æ $z }  ÿŽ â zÿ t    â $z |  &Ž ä $z		 w  ÿŽ ä $z |  Ž â x   ÿàÿ â $z v  Ž âr		   ÿàõ â $z v  Ž â z& q  ÿàÿ â $z
 p  Ž Ä $x   ÿ Ž â x   ÿ   â %x   Ž È $vÿ Y  ÿ Ž ä $rÿÿ   ÿ Ž ä $rÿÿ   ÿ ä $x   ÿ Ž ä %x
   Ž È r "       â$rÿÿ    Ž â r"   ÿàõ Æ %x   Ž Ê �y5   • ¨  �y#   • ¢  $x    Ê �y   • è �y0   • è �y,
   • ( �y-   • " %x
   Ž â r"   ÿàõ Ä �y    • h �y3   • h �y'   • b r   ÿä Ê �y*   • h �y+   • b r8&	   ÿàõ Æ �y4
   • h �y1   • b $t%ÿ$   ÿ Ž â r	   ÿà÷ â/ $rÿÿ    â �y    • b %v88 Z  % Ž â r ÿ   ÿäÿ Æ $x!$   ÿŽ â r68   ÿà÷ â %v Z  % Ž È $x99   !Ž â r   ÿàõ â $x% $   ÿŽ Æ r79   ÿäÿ â $x   %Ž â r:8   ÿà÷ â�y%8    ™ d $x    ä �y)6   ™ b $x;9   Ž â r    ÿàõ ä �y26   ™ h �y$   ™ b r!   ÿä Æ �y/:   ™ h �y.   ™ h �y(6   ™ h �y     ™ h �y<   ™ h �y    ™ b x&&@   ÿàÿ È | &   pbô ä x""€  ÿà÷ ä x@   ÿàÿ Æ $rÿÿ   Ž â x55T  #    ÈO r#ÿ   5  Ê $x:#   ÿ Ž â x0T  0    ä� rÿ   5  ä x9:   ÿÀŽ Æ/ $x6   ÿ Ž â x90øŽ äx!ÿ   0  ä x0ÿ   #  ä x-,T  -    äx66   ÿÀŽ ä x00   ÿÀŽ Ä xÿ     ä x6-øŽ â $x6   ÿ Ž â x,#   ÿ   ä x0!øŽ ä x0ÿ   #  ä xÿ   -  Ä x   ÿÀŽ ä x   ,øŽ â $x,#   ÿ Ž â x   6øŽ â $x6   ÿ Ž â x   0øŽ ä xøŽ ä x0#   ÿ   Ä xÿ     ä x'3T  '    äx,   ,øŽ â $x   ÿ Ž â x   0øŽ â $x0   ÿ Ž â x   6øŽ ä x   øŽ Ä x6	   ÿ   ä x#   ÿ   ä rÿ   '  ä x   øŽ ä x,   øŽ â $x   ÿ Ž ä $x,#   ÿ Ž â x**T  +    Ä x!ÿ     ä x#   ÿÀŽ ä rÿ   '  ä x   ,øŽ ä x'ÿ   *  ä x,!   ÿÀŽ Ä x##*øŽ â $x!   ÿ Ž â x14T  1    ä x*ÿ     ä x','øŽ ä x+ÿ   1  ä x4*   ÿÀŽ Ä x,ÿ     ä x   0øŽ ä x0!   ÿÀŽ ä x*   ÿ   ä x!4+øŽ ä x,'   ,øŽ â $x'   ÿ Ž â x4ÿ     Ä x+   ÿ   ä x101øŽ â $x0   ÿ Ž â x*#   *øŽ ä x!!   4øŽ â $x4   ÿ Ž â x,,   +øŽ â $x-   ÿ Ž â x**   'øŽ Ä x+   ÿ   ä x#1   0øŽ â $x0   ÿ Ž â x*   +øŽ ä x'!   4øŽ ä x!,   -øŽ â $x,   ÿ Ž â x##   0øŽ â &r*)   ÿ   Ä &r))   ÿ   â x0	   ÿ   â &r*2   *   âx##   ,øŽ â &r2!2   )   â x   6øŽ â &r$   ÿ   âx'   0øŽ â &r$   ÿ   ä &r*/   *   Ä &r2#/   2   ä &r.      ä&r!!.      ä 0rÿ%  0 A   ä &r*(   *   ä&r2(   2   Ä &r       ä&r!#    !   â  x   A  A  â 0r%ÿ%    A   â Er* *       â &r<      â Er 2       â &r!<   !   ä 0rÿ  0 A   â#r*%*     â#r%     â 0rÿ    A   â Er        â  x   A  A  â Er !       â 0rÿ    A   Ä 0rÿ    A   â #r     â#r     â #r*      ä #r      â #r      â #r      â G©  Ðôÿÿÿÿƒ ê Ay        € ê ˆ  þÿ H   è ˆ €þÿ H   è ˆ  ÿÿ H   è ˆ €ÿÿ H   è {          ì M        € ê „y      H   " z  r  p`ð Æ „y  €   H   b r  ÿ   p"p Æ „y     H   ¨ „y	     H   è „y  €  H   ( „y     H   h „y     H   h „y  €  H   h „y     H   h „y
  €  H   h „y  €  H   b !r       Æ „y  €  H   " !r       â/ !r       äO y      &   b !r	       â� !r       â!r       ä!r       â !r       È ‰      ¢ !r       Æ ‰	      â !r

       Æ ‰      " !r
       È !r       â !r       ÈO ‰      " !r		       â� !r
       È‰	      ¨ ‰      è ‰
      " !r       â !r	       âO !r       È� ‰ €    " !r
       Æ‰      ¨ ‰ €    è ‰ €    " !r       â !r	       ÈO ‰ @    " !r       Æ� ‰	 €    ¢ !r       Æ‰ @    è ‰
 @    ( y      '   b !r       â !r	       ÈO ‰      " x   ÿàÿ â $z x  Ž â/ !r       â� ‰ @    b z  r  p`ò â !r

       ä‰      ¢ x     p"ò â $z ~  Ž Ä‰	
      â $tÿ   ÿ Ž ä 
  r  ÿàõ È ÿÿ   ÿä ä 
  j  ÿ„ â !r       Æ 
  k   â !r       â/ !r       âO %v j   Ž Æ ‰       b  !r	
	       Æ� †	     è †    è †	 	    â M™        € ê ¹z  r      â !r        â? �x   ?àÿ Ì |   ÿàñ Ê $rÿÿ   ÿ  â z j  ÿ€ È z k    Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž È % \   Ž Ì �	    ™ b ¹z  l      â $rÿÿ   ÿ Ž â ™x?    â y      "   b Ey  @    € â '
 m   Ž äy       !   ¤ 'Š	 s   Ž â 
ÿ n    Æ $rÿÿ   ÿ Ž â ÿ   ÿáÿ Ê $
 o  Ž ä$‚ÿÿ    Ž â r ÿ   pRò ä/r ÿ   pRô â$x     Ž äO y      '   & r   ÿ� È xÿ     È |    pbö ä Šÿ t  	  Ö G9       € ê $rÿÿ   ÿ Ž â y      %   " $tÿ   ÿ Ž ä 'z	 y   Ž È $z }  ÿŽ â z
ÿ z  	  â $rÿÿ   ÿ Ž È $z	
 |  Ž È $z v  	Ž âx	þÿÿÿÿÀŽ È r		  €ÿàÿ â $z p  Ž ä $z w  ÿŽ ä $x	   ÿ Ž ä$x	   ÿ Ž â r   ÿàñ â %v X  Ž â x
ÿ     Ä x   áÿ â $rÿÿ    Ž â x        â $rÿÿ    Ž â xÿ     â $x	   ÿ Ž â x   
  â $r ÿÿ   ÿ  â x     â%v  € Ž â x   ÿ   Æ $rÿÿ    Ž ä $r	ÿÿ    Ž Æ r   ÿàñ ä �y   • ¦ $x   	  â �y   • ¨ �yþÿÿ • è  �y!    • â  r   ÿàñ Æ �y   • " x $       Æ �y#   • " $r
ÿÿ      ä %v Z   Ž â �y"    • f $x
$   ÿŽ â r
   ÿ€ Æ $x   Ž È $x     â r   ÿàñ ä �y    ™ h �y
   ™ b r   ÿä  Æ �y
   ™ h �y   ™ h �y
   ™ b x@   ÿàÿ È |    pbð ä xT      ÈO rÿ     Ê $x$   ÿ Ž â xT  !    ä� rÿ     ä x!$   ÿÀŽ ä xÿ     ä x!øŽ â/ $x!   ÿ Ž Ä $x
   ÿ Ž â xT  #    äx$   ÿÀŽ ä xÿ     ä x!!   ÿÀŽ ä xÿ     ä x   
øŽ Ä x$øŽ ä x
   ÿ   ä x!øŽ äx$   ÿÀŽ ä x!ÿ     â $x   ÿ Ž â xÿ     Ä x   
øŽ ä x
$!øŽ â $x$   ÿ Ž â x   øŽ â $x   ÿ Ž âx   øŽ ä x!   ÿ   äx   $øŽ â $x   ÿ Ž â x#ÿ     Ä x   øŽ ä x   !øŽ â $x   ÿ Ž âx#
   #øŽ â $x
   ÿ Ž â x   øŽ â &r   ÿ   âx#   øŽ ä x
   
øŽ â &r      â x	   ÿ   Æ &r

      â x   øŽ â 0rÿ  0 A   È &r
   
   â  x   Á  A  â 0rÿ    A   Æ Er 
       â 0r"ÿ"    A   â x€  ÿà÷ Æ #r      ä $rÿÿ   	Ž ä #r"      â G‰  °úÿÿÿÿƒ ê Ay        € ê ˆ €ÿÿ H   è {          ì M)        € ê „y      H   "r  ÿ   pRð Æ „y  €   H   h „y	     H   ¨ y      %   ä $z x  Ž â� !r       È !r       È/ !r	       ÊO ‰      $ !r       Ê ‰      $ !r       Ê ‰	 €    $ !r		       Ê ‰	 @    $ !r	       Ê ‰      b  M	        € ê $tÿ   ÿ Ž â !r       â/ $z ~  Ž È %v j   Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž È %
 \   Ž Ê �		
    ™ ¢  $rÿÿ   ÿ Ž â Ey  °    € â x      ÿ  â y       "   b 'Š s   Ž Æ y      !   b '
 m   Ž Ê 
ÿ n    ä rÿ    ÿü‚ â/ $x     Ž Ê r
   ÿ� ä Š	ÿ t    ä y      %   $ r	       âO G  Ð    € ê z ÿ `  pPô ä z ÿ ^  pPò ä z ÿ b  p ö Ä z ÿ a   Sô ä z ÿ _  Sò ä z ÿ c  0't Ö       '   b $ x  ÿŽ ä$ª x  ÿŽ â ©      '   ¢ ¨ÿ     â$ ~  ÿŽ ä/ $ª ~  ÿŽ âO ÿ     ä   àë Ä ¢   àÇ ä ÿÿ   Äþ ä    ÿàý ä ¢   ÿàû ä ¢ÿÿ   „þ â $ÿÿ    â  ^  ÿ† Ä ª b  ÿˆ â $¢ÿÿ   Ž â  _  � ä x      ÿ  ä ª c   È �     h �©     d Ay        € ê ¹z  l      â xÿ   
  â ™x?    â $ÿÿ   
Ž â Ey  €    € æ $
 o  Ž â|    pbò â $‚ÿÿ    Ž Ä x      ÿ  Ô G       € ê y      '   b $rÿÿ   ÿ Ž ä $tÿ   ÿ Ž ä 'z y   Ž Ê/ zÿ z    Ê $z |  Ž â x
þÿÿÿÿÀŽ Æ $z		 v  Ž ä $x   
Ž ä $z p  	Ž ä $x	   ÿ Ž ä$x
   ÿ Ž ä $z w  ÿŽ Ä $z }  ÿŽ â x	   
áÿ â %x   ÿŽ Æ r   ÿàñ â $rÿÿ   ÿ Ž â xÿ     â %v
 `  Ž â r ÿ       ä x     â$rÿÿ   ÿ  ä $x   ÿ Ž ä z X  ÿàñ â $tÿ   ÿ Ž Æ z Y  ÿä  â %v
 X  Ž È �y   • ¢ r"   ÿàñ Æ �y   • ¤ $x#     Ê �y"   • è  �y"   • â  r   ÿàñ Æ �y "
   • ( �y!"   • " $tÿ$   ÿ Ž ä $rÿÿ     ä %v Z   Ž È $x$   ÿŽ â x   ÿàñ Æ $x   Ž È $rÿÿ     â r   ÿàñ Æ %x	   Ž Ê �y#    ™ " $x     Æ �y"   ™ ( �y   ™ ( �y$   ™ ( �y    ™ " $x%   ÿ Ž Æ �y    • b%z ÿ `  pPð È z ÿ a   Sð ä x@   ÿàÿ ä xT      ÈO rÿ%     ärÿ     ä x   ÿ   ä xT      ä� x   ÿÀŽ ä xÿ     â $x   ÿ Ž â xøŽ Ä x  T  !    äxÿ     ä xÿ     ä x   ÿÀŽ ä x!   ÿÀŽ ä xÿ      Ä x   ÿÀŽ ä xøŽ â $x   ÿ Ž â xøŽ â $x   ÿ Ž â x! øŽ ä x   øŽ â $x   ÿ Ž â x!ÿ     Ä x   øŽ â $x   ÿ Ž â x   !øŽ â $x!   ÿ Ž â x   øŽ â $x   ÿ Ž â xÿ     ä x   øŽ â $x   ÿ Ž â x   !øŽ â $x   ÿ Ž â x   øŽ â $x!   ÿ Ž â x   øŽ â &r#   ÿ   â x   øŽ Ä x   øŽ ä x   !øŽ â &r"      â x	   ÿ   Æ &r      â x   øŽ â 0rÿ  0 A   È &r$      â  x   Á  A  â 0rÿ    A   Æ Er        â G‰  à    € ê z `  ÿàñ ä r   ÿàó ä z a  ÿä  Æ $x   Ž  ä �y!    ¨  �y    ¨  �y    è �y    è �y 
    ( �y     ( �y    $% x!T      ÈO r%ÿ%     ärÿ     Æ $x!%   ÿ Ž â xT      È� x!!   ÿÀŽ ä xÿ     äx!!øŽ ä x T      â$x    ÿ Ž â xÿ     È x     ÿÀŽ ä xÿ   %  ä xÿ     äx   ÿÀŽ ä x øŽ â $x %   ÿ Ž â x   ÿÀŽ Ä xøŽ â $x   ÿ Ž â x!!    øŽ â $x %   ÿ Ž â xøŽ ä xÿ   %  ä x   øŽ ä xÿ     Ä x   øŽ â $x%   ÿ Ž â x !    øŽ â $x!%   ÿ Ž â x   øŽ â $x   ÿ Ž â x     øŽ â $x%%   ÿ Ž â x   !øŽ ä x   øŽ â $x   ÿ Ž âx   ÿ   â&r  #   ÿ   â x%   %øŽ â $x   ÿ Ž â x   øŽ Ä x   øŽ â &r%%"       â x   øŽ Æ &r   %   È &r$      Ê Er        â 0rÿ    A   È #r      È #r      â |    pbò â 0rÿ    A   âx€  ÿàñ â #r      È $rÿÿ     â #r      Ê G™  `öÿÿÿÿƒ ê Ay        € ê r  ÿ   pRð âEy  �     € â r  ÿ   pRò â $rÿÿ    Ž Ô G‰  P     € ê z ÿ `  pPð â $x      Ž Æ z ÿ a   Sð ä ˆs  €ÿÿ H   ö ˆ     H   ä Ay        € ê {          ì M        € ê z ÿ `  pPð â „y      H   f. z ÿ a   Sð Ú „‰ €   H   ¨ „‰	    H   â !r        È? !‚       ÈO !‚	       â� G‰        € ê „y  €  H   ( „y €   H   h „y    H   ¨ „y    H   è „y	 €  H   " !r         â !r       Æ/ !r         âO !r       Æ� !r 	       ä‰      " z ÿ `  pPð ä r ÿ   pRò ä z ÿ a   Sð â !r       Ê ‰       $ !r         Ê ‰	  €    $ !r	 	       Ê ‰	 @    $ !r	       Ê ‰      $ !r
       â G‰        € ê ‰       $ !r         Ê ‰       $ !r        Ê ‰ €    $ !r       Ê ‰ @    $ !r       Ê ‰      $ !r       ä M        € ê !r
       âG‰  �    € ê $t ÿ   ÿ Ž â !r       È z   h  pð Ú G	  �    € ê ¹z  h      ä �xÿÿÿÿ?àÿ È Œx    pð È ‡x       È ™x   ?  Ø »z  €    ä ™x?    Ô Xy pìÿÿÿÿƒ ê  x ;ª¸¿  A  Ì s          $ !t   €?    Ì s        $  r     A  È  r     A  â Gy  `    € ê  x '7=  A  â x*BL?  A  Æ #t   €?    È  r      A  â  x    ?  A  Ê s     $   $ !t  €?    È  r       A  È  r      A  â Gy  À    € ê $v ÿ h  ÿ Ž Ê x     pRð Ú G‰  à     € ê x     pRð Ú G	        € ê 	z  i    � ä 	z i    � Æ  x ;ª¸¿  A  â 	z i €  � Ê s        $ !t  €?    Ì s        $  r       A  È  r      A  â Gy  Ð     € ê  r     A  â Gy  °     € ê 	x   à@  � ä 	x  à@  � Æ  x #ÛÙ?  A  â 	x  àÀ  � Æ  x;ª¸¿  A  ä !t  €?    È s        $ !t  €?    Ì s        $  r       A  È  r      A  ä y      &   ( y      '   b $z x  Ž ä $tÿ   ÿ Ž ä $z ~  Ž È/ %v j   Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž È % \   Ž Ì �	    ™ b ¹z  l      â $rÿÿ   ÿ Ž â ™x?    â y      "   b Ey  @    € â '
 m   Ž äy       !   ¤ 'Š	 s   Ž â 
ÿ n    Æ $rÿÿ   ÿ Ž â ÿ   ÿáÿ Ê $
 o  Ž ä$‚ÿÿ    Ž â r ÿ   pRò ä/r ÿ   pRô â$x     Ž äO y      '   & r   ÿ� È xÿ     È |    pbö ä Šÿ t  	  Ö G9       € ê $rÿÿ   ÿ Ž â y      %   " $tÿ   ÿ Ž ä 'z	 y   Ž È $z }  ÿŽ â z
ÿ z  	  â $rÿÿ   ÿ Ž È $z	
 |  Ž È $z v  	Ž âx	þÿÿÿÿÀŽ È r		  €ÿàÿ â $z p  Ž ä $z w  ÿŽ ä $x	   ÿ Ž ä$x	   ÿ Ž â r   ÿàñ â %v X  Ž â x
ÿ     Ä x   áÿ â $rÿÿ    Ž â x        â $rÿÿ    Ž â xÿ     â $x	   ÿ Ž â x   
  â $r ÿÿ   ÿ  â x     â%v  € Ž â x   ÿ   Æ $rÿÿ    Ž ä $r	ÿÿ    Ž Æ r   ÿàñ ä �y   • ¦ $x   	  â �y   • ¨ �yþÿÿ • è  �y!    • â  r   ÿàñ Æ �y   • " x $       Æ �y#   • " $r
ÿÿ      ä %v Z   Ž â �y"    • f $x
$   ÿŽ â r
   ÿ€ Æ $x   Ž È $x     â r   ÿàñ ä �y    ™ h �y
   ™ b r   ÿä  Æ �y
   ™ h �y   ™ h �y
   ™ b x@   ÿàÿ È |    pbð ä xT      ÈO rÿ     Ê $x$   ÿ Ž â xT  !    ä� rÿ     ä x!$   ÿÀŽ ä xÿ     ä x!øŽ â/ $x!   ÿ Ž Ä $x
   ÿ Ž â xT  #    äx$   ÿÀŽ ä xÿ     ä x!!   ÿÀŽ ä xÿ     ä x   
øŽ Ä x$øŽ ä x
   ÿ   ä x!øŽ äx$   ÿÀŽ ä x!ÿ     â $x   ÿ Ž â xÿ     Ä x   
øŽ ä x
$!øŽ â $x$   ÿ Ž â x   øŽ â $x   ÿ Ž âx   øŽ ä x!   ÿ   äx   $øŽ â $x   ÿ Ž â x#ÿ     Ä x   øŽ ä x   !øŽ â $x   ÿ Ž âx#
   #øŽ â $x
   ÿ Ž â x   øŽ â &r   ÿ   âx#   øŽ ä x
   
øŽ â &r      â x	   ÿ   Æ &r

      â x   øŽ â 0rÿ  0 A   È &r
   
   â  x   Á  A  â 0rÿ    A   Æ Er 
       â 0r"ÿ"    A   â x€  ÿà÷ Æ #r      ä $rÿÿ   	Ž ä #r"      â G‰  °úÿÿÿÿƒ ê Ay        € ê ˆ €ÿÿ H   è {          ì M)        € ê „y      H   "r  ÿ   pRð Æ „y  €   H   h „y	     H   ¨ y      %   ä $z x  Ž â� !r       È !r       È/ !r	       ÊO ‰      $ !r       Ê ‰      $ !r       Ê ‰	 €    $ !r		       Ê ‰	 @    $ !r	       Ê ‰      b  M	        € ê $tÿ   ÿ Ž â !r       â/ $z ~  Ž È %v j   Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž È %
 \   Ž Ê �		
    ™ ¢  $rÿÿ   ÿ Ž â Ey  °    € â x      ÿ  â y       "   b 'Š s   Ž Æ y      !   b '
 m   Ž Ê 
ÿ n    ä rÿ    ÿü‚ â/ $x     Ž Ê r
   ÿ� ä Š	ÿ t    ä y      %   $ r	       âO G  Ð    € ê z ÿ `  pPô ä z ÿ ^  pPò ä z ÿ b  p ö Ä z ÿ a   Sô ä z ÿ _  Sò ä z ÿ c  0't Ö       '   b $ x  ÿŽ ä$ª x  ÿŽ â ©      '   ¢ ¨ÿ     â$ ~  ÿŽ ä/ $ª ~  ÿŽ âO ÿ     ä   àë Ä ¢   àÇ ä ÿÿ   Äþ ä    ÿàý ä ¢   ÿàû ä ¢ÿÿ   „þ â $ÿÿ    â  ^  ÿ† Ä ª b  ÿˆ â $¢ÿÿ   Ž â  _  � ä x      ÿ  ä ª c   È �     h �©     d Ay        € ê ¹z  l      â xÿ   
  â ™x?    â $ÿÿ   
Ž â Ey  €    € æ $
 o  Ž â|    pbò â $‚ÿÿ    Ž Ä x      ÿ  Ô G       € ê y      '   b $rÿÿ   ÿ Ž ä $tÿ   ÿ Ž ä 'z y   Ž Ê/ zÿ z    Ê $z |  Ž â x
þÿÿÿÿÀŽ Æ $z		 v  Ž ä $x   
Ž ä $z p  	Ž ä $x	   ÿ Ž ä$x
   ÿ Ž ä $z w  ÿŽ Ä $z }  ÿŽ â x	   
áÿ â %x   ÿŽ Æ r   ÿàñ â $rÿÿ   ÿ Ž â xÿ     â %v
 `  Ž â r ÿ       ä x     â$rÿÿ   ÿ  ä $x   ÿ Ž ä z X  ÿàñ â $tÿ   ÿ Ž Æ z Y  ÿä  â %v
 X  Ž È �y   • ¢ r"   ÿàñ Æ �y   • ¤ $x#     Ê �y"   • è  �y"   • â  r   ÿàñ Æ �y "
   • ( �y!"   • " $tÿ$   ÿ Ž ä $rÿÿ     ä %v Z   Ž È $x$   ÿŽ â x   ÿàñ Æ $x   Ž È $rÿÿ     â r   ÿàñ Æ %x	   Ž Ê �y#    ™ " $x     Æ �y"   ™ ( �y   ™ ( �y$   ™ ( �y    ™ " $x%   ÿ Ž Æ �y    • b%z ÿ `  pPð È z ÿ a   Sð ä x@   ÿàÿ ä xT      ÈO rÿ%     ärÿ     ä x   ÿ   ä xT      ä� x   ÿÀŽ ä xÿ     â $x   ÿ Ž â xøŽ Ä x  T  !    äxÿ     ä xÿ     ä x   ÿÀŽ ä x!   ÿÀŽ ä xÿ      Ä x   ÿÀŽ ä xøŽ â $x   ÿ Ž â xøŽ â $x   ÿ Ž â x! øŽ ä x   øŽ â $x   ÿ Ž â x!ÿ     Ä x   øŽ â $x   ÿ Ž â x   !øŽ â $x!   ÿ Ž â x   øŽ â $x   ÿ Ž â xÿ     ä x   øŽ â $x   ÿ Ž â x   !øŽ â $x   ÿ Ž â x   øŽ â $x!   ÿ Ž â x   øŽ â &r#   ÿ   â x   øŽ Ä x   øŽ ä x   !øŽ â &r"      â x	   ÿ   Æ &r      â x   øŽ â 0rÿ  0 A   È &r$      â  x   Á  A  â 0rÿ    A   Æ Er        â G‰  à    € ê z `  ÿàñ ä r   ÿàó ä z a  ÿä  Æ $x   Ž  ä �y!    ¨  �y    ¨  �y    è �y    è �y 
    ( �y     ( �y    $% x!T      ÈO r%ÿ%     ärÿ     Æ $x!%   ÿ Ž â xT      È� x!!   ÿÀŽ ä xÿ     äx!!øŽ ä x T      â$x    ÿ Ž â xÿ     È x     ÿÀŽ ä xÿ   %  ä xÿ     äx   ÿÀŽ ä x øŽ â $x %   ÿ Ž â x   ÿÀŽ Ä xøŽ â $x   ÿ Ž â x!!    øŽ â $x %   ÿ Ž â xøŽ ä xÿ   %  ä x   øŽ ä xÿ     Ä x   øŽ â $x%   ÿ Ž â x !    øŽ â $x!%   ÿ Ž â x   øŽ â $x   ÿ Ž â x     øŽ â $x%%   ÿ Ž â x   !øŽ ä x   øŽ â $x   ÿ Ž âx   ÿ   â&r  #   ÿ   â x%   %øŽ â $x   ÿ Ž â x   øŽ Ä x   øŽ â &r%%"       â x   øŽ Æ &r   %   È &r$      Ê Er        â 0rÿ    A   È #r      È #r      â |    pbò â 0rÿ    A   âx€  ÿàñ â #r      È $rÿÿ     â #r      Ê G™  `öÿÿÿÿƒ ê Ay        € ê r  ÿ   pRð âEy  �     € â r  ÿ   pRò â $rÿÿ    Ž Ô G‰  P     € ê z ÿ `  pPð â $x      Ž Æ z ÿ a   Sð ä ˆs  €ÿÿ H   ö ˆ     H   ä Ay        € ê {          ì M        € ê z ÿ `  pPð â „y      H   f. z ÿ a   Sð Ú „‰ €   H   ¨ „‰	    H   â !r        È? !‚       ÈO !‚	       â� G‰        € ê „y  €  H   ( „y €   H   h „y    H   ¨ „y    H   è „y	 €  H   " !r         â !r       Æ/ !r         âO !r       Æ� !r 	       ä‰      " z ÿ `  pPð ä r ÿ   pRò ä z ÿ a   Sð â !r       Ê ‰       $ !r         Ê ‰	  €    $ !r	 	       Ê ‰	 @    $ !r	       Ê ‰      $ !r
       â G‰        € ê ‰       $ !r         Ê ‰       $ !r        Ê ‰ €    $ !r       Ê ‰ @    $ !r       Ê ‰      $ !r       ä M        € ê !r
       âG‰  �    € ê $t ÿ   ÿ Ž â !r       È z   h  pð Ú G	  �    € ê ¹z  h      ä �xÿÿÿÿ?àÿ È Œx    pð È ‡x       È ™x   ?  Ø »z  €    ä ™x?    Ô Xy pìÿÿÿÿƒ ê  x ;ª¸¿  A  Ì s          $ !t   €?    Ì s        $  r     A  È  r     A  â Gy  `    € ê  x '7=  A  â x*BL?  A  Æ #t   €?    È  r      A  â  x    ?  A  Ê s     $   $ !t  €?    È  r       A  È  r      A  â Gy  À    € ê $v ÿ h  ÿ Ž Ê x     pRð Ú G‰  à     € ê x     pRð Ú G	        € ê 	z  i    � ä 	z i    � Æ  x ;ª¸¿  A  â 	z i €  � Ê s        $ !t  €?    Ì s        $  r       A  È  r      A  â Gy  Ð     € ê  r     A  â Gy  °     € ê 	x   à@  � ä 	x  à@  � Æ  x #ÛÙ?  A  â 	x  àÀ  � Æ  x;ª¸¿  A  ä !t  €?    È s        $ !t  €?    Ì s        $  r       A  È  r      A  ä y      &   ( y      '   b $z x  Ž ä $tÿ   ÿ Ž ä $z ~  Ž È/ %v j   Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž ä $rÿÿ    Ž È % \   Ž Ê �	
    ™ b ¹z  l      â Ey  °    € â ™x?    â y	      "   b $r6ÿÿ   ÿ Ž Æ y       !   ¢ $ÿÿ   ÿ Ž ä $ÿÿ    Ž â y7      %   & '
 m   Ž È $‚ÿÿ   ÿ Ž â 
ÿ n    Ê $ÿÿ   
Ž â r 	ÿ   pRò Æ/ $
 o  Ž â r 	ÿ   pRô â$x	     Ž äO $‚ÿÿ    Ž Æ r   ÿ� â 'Š s   Ž ä x      ÿ  â x8ÿ     â $	€    Ž ä $x77   ÿ Ž â | 8   pbö ä Š
ÿ t    â $rÿÿ   ÿ Ž Ô G9       € ê y	      '   " $rÿÿ   ÿ Ž â xþÿÿÿÿÀŽ â $z w  ÿŽ ä $v>ÿ p  ÿ Ž ä $x   
Ž ä $vÿ Y  ÿ Ž ä $x;   ÿ Ž Ä $r6ÿÿ   ÿ Ž ä 'z	 y   Ž È $z		 }  ÿŽ â zÿ z    â $z7 p  8Ž Æ r	   ÿàñ â $z |  Ž ä $x   ÿ Ž ä $z
 v  Ž ä$t	ÿ   ÿ Ž â x   ;áÿ â %x<   ÿŽ â xÿ     Ä xÿ     â $v
ÿ X  ÿ Ž â x     â %v>>  € 	Ž â x     Æ $x   ÿ Ž ä $rÿÿ   ÿ  ä $x   ÿ Ž ä %x:;   <Ž È r(<
   ÿàñ Ê $x)=     â r:
   ÿàñ È �y,(   • ¨  �y-(   • ¢  $x;     â r*8   ÿàñ È �y.   • è �y3   • â $rÿÿ     Æ �y2
   • " %zD	 p  (Ž Æ �y1   • " $x$   ÿŽ ä $t+ÿ$   ÿ Ž â �yD   • b %z$	 p  Ž Æ �yD   • b %v** Z  + Ž â x (   ÿàñ ä �y$   • b# $x++   Ž ä $r!ÿÿ   )  â �y$   • b r"*   ÿàñ ä%z 	 p   Ž â r4*   ÿ† â �y$
   • d $x#+     ârB>    ÿàñ â $x5+   Ž â �y$   • f $xC?   !  â �y4   ™ h �yBþÿÿ • h �yB    • b %z&	 p  $Ž Æ �y4   ™ h �y&   • h �y&   • h �y"   ™ h �y&
   • h �y4   ™ b x" þÿÿÿÿàñ Æ �y &   • b x#!ÿÿÿÿÿä  Æ �y((    • d %z"	 p  "Ž Ê r@>"   ÿàñ â %z4	 p  &Ž â/ �y)D    • f $xA?   #  â �y$4   • h �y!@    • h �y"@   • h �y#4   • h �y%4
   • h �y&4   • h �y'*    ™ h �y*Büÿÿ • h/ �y+@þÿÿ • b x88@   ÿàÿ Ä x-,T  -    ÈO r,ÿ   -  är-ÿ   -  Æ $x4,   ÿ Ž â x3.T  3    â� $x0-   ÿ Ž â x.ÿ   ,  ä x44   ÿÀŽ â $x/,   ÿ Ž â x..   ÿÀŽ ä x443øŽ äx3ÿ   3  Ä x00   ÿÀŽ ä x/4   /øŽ ä x12T  1    äx4ÿ   -  â $x5-   ÿ Ž â x..3øŽ â $x2,   ÿ Ž â x3ÿ   ,  Ä x001øŽ äx1ÿ   1  ä x44   ÿÀŽ ä x..   3øŽ â $x3,   ÿ Ž â xT      äx50   5øŽ Ä x//   2øŽ â $x2-   ÿ Ž â x141øŽ ä x0ÿ   -  ä rÿ     ä x..   3øŽ â $x3,   ÿ Ž âx01   0øŽ â $x1-   ÿ Ž â x25   2øŽ â $x4,   ÿ Ž Ä $x,   ÿ Ž â x3.   3øŽ ä x22   1øŽ ä x.ÿ     ä x1T      ä x,,   ÿÀŽ ä xÿ   1  Ä x..   ÿÀŽ ä x,,1øŽ â $x1   ÿ Ž â x4/   4øŽ â $x/-   ÿ Ž âx..øŽ â $x-   ÿ Ž â xÿ     â $x-   ÿ Ž â x,,   1øŽ Ä x..   øŽ â $x   ÿ Ž â x,,   -øŽ â $x-   ÿ Ž ârÿ     â $x   ÿ Ž â x.   øŽ ä xT      ä x/0   /øŽ Ä x,,   -øŽ ä xT      â $x   ÿ Ž â x   øŽ ä rÿ     ä x//   øŽ â &r,   ÿ   â x-ÿ     Ä x,   ÿÀŽ â $x   ÿ Ž â xÿ     äx--   ÿÀŽ â &r      â x,øŽ â $x,   ÿ Ž â x-øŽ ä xT      Ä x-   ÿÀŽ ä xÿ     ä x   ,øŽ â $x,   ÿ Ž â xÿ     ä x-øŽ äxÿ     Ä x   ÿÀŽ â $x-   ÿ Ž âx   øŽ â $x   ÿ Ž â xøŽ ä x   ,øŽ ä xÿ     ä r,ÿ     â $x   ÿ Ž â x   -øŽ â $x-   ÿ Ž â x   øŽ â $x,   ÿ Ž â x   øŽ â $x   ÿ Ž â x   -øŽ â $x.   ÿ Ž âx   øŽ â $x   ÿ Ž â x   ÿÀŽ Ä xT       ä xÿ   ,  â $x   ÿ Ž â x   øŽ ä xøŽ äx    ÿÀŽ â $x,   ÿ Ž â xÿ     Ä x   .øŽ ä x   øŽ â &r   ÿ   â x   øŽ â $x,   ÿ Ž â x øŽ ä xÿ   ,  ä x   øŽ â &r      â x!!T  "    Ä x   øŽ â $x,   ÿ Ž âx   øŽ â $x,   ÿ Ž â rÿ   !  â&r      â x   øŽ ä rÿ   !  â $x   ÿ Ž â x   øŽ Ä x##T  $    ä xÿ     â $x   ÿ Ž â x   ÿÀŽ ä xÿ   #  ä x   ÿÀŽ ä x%%T  &    Ä x##øŽ ä x!ÿ     ä xøŽ â $x   ÿ Ž â x    ÿÀŽ ä xÿ   %  ä x"!   ÿÀŽ Ä x% %øŽ â $x    ÿ Ž âxÿ     â $x!   ÿ Ž â x##   øŽ ä x"øŽ â $x"   ÿ Ž â x   øŽ â $x   ÿ Ž â x #    øŽ Ä x   !øŽ â $x!   ÿ Ž â x%   "øŽ ä x     øŽ â $x   ÿ Ž â x$ÿ     â $x"   ÿ Ž âx   !øŽ â $x!   ÿ Ž â x   $øŽ â &r44   ÿ   âx   øŽ â &r    ÿ   â x   "øŽ â $x,,   ÿ Ž â x   !øŽ â $x   ÿ Ž Ä &r33   4   ä&r      â x   ,øŽ â &r22   3   âx   øŽ â &r      ä&r      â | 8   pbð â &r//   2   Ä &r      ä&r      ä&r      ä 0rÿ'  0 A   âEr /       â 0r'ÿ'    A   â Er        ä Er        â  x   A  A  â Er        â 0r(ÿ(    A   â x

€  ÿà÷ â 0r)ÿ)    A   â #r'     â0r*ÿ*    A   â #r'     â0r+ÿ+    A   â #r'     â#r'     â #r(      â #r)      â #r*      â #r6+   6   â $rÿÿ   Ž â D	       À â Gy  àîÿÿÿÿƒ ê Ay        € ê ˆ  þÿ H   è ˆ €þÿ H   è ˆ  ÿÿ H   è ˆ 6€ÿÿ H   è {          ì M)        € ê „y      H   " z 7 r  p`ô â $z x  7Ž Ä „y  €   H   b r  ÿ   pVt Æ „y
     H   ¨„y     H   è „y  €  H   ( „y     H   ( „y	     H   ( „y  €  H   ( „y  €  H   ( „y     H   ( „y  €  H   " !r       Æ „y  €  H   " !r       â/ !r

       âO !r       ä� !r       â!r

       â !r	       â !r6       È ‰      b !r       â !r
       È ‰      ¢ !r       Æ ‰      â !r       Ê ‰      " !r       â/ !r       âO !r       È� ‰
      h ‰      ¢ !r       Æ ‰      ( ‰      â !r

       â/ !r       ÈO ‰	
 €    b !r       Æ ‰ €    " !r       Æ� ‰ €    ¨ ‰ €    â !r	
	       â/ !r       È ‰	 @    " x7   ÿàÿ â!r       äO ‰ @    b x7   ÿàÿ â !r       ä� ‰ @    ¢ z  r  p`ö ä z  r  p`ò â ‰ @    â x     pVö Ä x     pVò  â y      '   " !r	       â x	7   ÿàÿ â !r       Æ/ ‰      " z 	 r  p`ð â !r       äO ‰      b x     pVp  â $tÿ   ÿ Ž â !r       â� ‰      ¢ $z ~  Ž Æ‰
      â !r       â !r	       â/ %v j   Ž â !r       ÈO †© 	    â !r
       Æ� †¹    è †‰    â M        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú         Ê % \   Ž Ê �	    ™ ¢  r ÿ       â Ey       € ä y	      "   d '
 m   Ž äy>      %   ä 'Š s   Ž â 
ÿ n    â y       !   " x      ÿ  Æ y      '   " r 	ÿ   pRò ä/x>>   ÿ   ä� x     pDò  är>    ÿàÿ är	    ÿ(Ž ä z  r  pdò  Ä r   ÿ� ä Šÿ t    ä ÿ   ÿáÿ ä r       ÆO $
 o  Ž ä G  °    € ê z ÿ `  pPô ä z ÿ ^  pPò ä z ÿ b  p ö ä z ÿ a   Sô ä z ÿ _  Sò Ä z ÿ c  0't Ö $ ~  ÿŽ âÿ   >  â $ª ~  ÿŽ ä $
 x  ÿŽ â>    àë â$ª	 x  ÿŽ â ¢>    àÇ ä ¨ÿ   >  ä ÿ   ÿÄþ Ä 
   ÿàý ä ¢	   ÿàû ä ¢ÿ   ÿ„þ ä ÿ   ÿä ä  ^  ÿ† ä ¢	ÿ   ÿäÿ Ä ª
 b  ÿˆ ä  _  � ä x      ÿ  â ª c  	 Ê �     h �©
     d Ay        € ê ¹z  l      â xÿ     â ™x?    â ‚        â Ey  @#    € â x      ÿ  â x      ÿ  â x      ÿ  â |    pbð â x
      ÿ  Ø G	  Ð"    € ê r ÿ       â $z w  ÿŽ â xþÿÿÿÿÀŽ â x      ÿ  â z ÿ `  pPð â 'z y   Ž âr        â x
      ÿ  â r  €ÿàÿ â $z }  ÿŽ â zÿ z    â $z> p  Ž â x   ÿ   Ä x   ÿ   â $z |  Ž â r   ÿàó â %xB   ÿŽ â x   áÿ ä z  `      â $z |  Ž â z  a      â $z v  Ž âz  X      â $z v  Ž â z  Y      â x      ÿ  â x      ÿ  â z ÿ a   Sð â %x@   BŽ â x   ÿ   Ä rÿÿ   ÿäÿ  ä x   ÿ   ä %xH   Ž â r@   ÿàó È �y4H   • ¨ �y5H   • ¢ rA   ÿäÿ  Ê �y=   • è  �y6   • â  z)  p      Æ �y8
   • $ %xF)   HŽ ä�y;   • " rJ   ÿàó ä xK $       â �yF   • " %x*)   Ž â rÿ   ÿäÿ  Ä �yF   • " %vJJ Z  K Ž Æ �y*   • " $x$   ÿŽ Æ �y *   • " x.J   ÿàó ä rKK   ÿàÿ â %x2)   FŽ â �y!*
   • $ r/ÿK   ÿäÿ  ä �y#2   • & %x,   .Ž â�y$2   • ( �y,    ™ " %x.   .Ž Æ �y,   ™ ( �y.    ™ " %x0)   *Ž Æ �y,   ™ ( �y"*   • ( �y%0   • ( �y&0   • ( �y'0
   • ( �y(0   • " %xD)   2Ž Æ �y/J    ™ " %xL)   0Ž Æ �y)D   • ( �y*D   • (/ �y+L   • ( �y,L   • ( �y-L
   • ( �y.L   • ( �y0H    • ( �y1F    • ( �y22    • ( �y3D    • " x54T  5    ÈO r4ÿ   5  È x74   ÿ   äx6=T  6    ä� x77   ÿÀŽ ä x<4   ÿ   ä x776øŽ ä r5ÿ   5  Ä x9ÿ   4  ä x<7   <øŽ ä x74   ÿ   ä x:5   ÿ   ä x6ÿ   6  ä x99   ÿÀŽ Ä x7<   7øŽ ä x;8T  ;    äx::   ÿÀŽ ä x<ÿ   5  ä x996øŽ ä x6ÿ   4  Ä x84   ÿ   ä x::;øŽ äx;ÿ   ;  ä x<<   ÿÀŽ ä xT      ä x69   6øŽ Ä x87   8øŽ ä x94   ÿ   äx74	   ÿ   ä x=5   ÿ   ä x;<;øŽ ä x4ÿ   5  Ä rÿ     ä x66   9øŽ ä x::   =øŽ ä x95   ÿ   ä x;;   4øŽ ä x4   ÿ   Ä x::   9øŽ ä x T       ä x94   ÿÀŽ ä xÿ     ä x4ÿ      äx99 øŽ Ä x   ÿÀŽ ä x    ÿ   äx44øŽ ä x99    øŽ ä xÿ     ä x    ÿ   Ä x76   7øŽ ä x65   ÿ   ä x4   øŽ ä x99    øŽ ä x    ÿ   äx4   ÿ   Ä x6;   6øŽ â &r88   ÿ   â x;5   ÿ   ä x    øŽ ä x49   4øŽ ä x	   ÿ   â &r77   8   â x::   ;øŽ Ä x55	   ÿ   ä x    øŽ â &r4   ÿ   â rÿ     â &r7:   7   â x56   5øŽ â &r        â xÿ     ä x##T  $    â &r5   7   â x4   ÿ   Ä x5   ÿÀŽ ä rÿ   #  ä x!!T  "    ä x44   ÿÀŽ ä x&%T  &    ä x%ÿ     Ä x"ÿ   !  äx!4!øŽ ä x4   ÿ   ä x7   ÿ   ä x$ÿ   &  ä x%%   ÿÀŽ Ä x"5"øŽ ä x5ÿ     ä x77   ÿÀŽ ä x!!   4øŽ ä x%%$øŽ ä x4ÿ     Ä r#ÿ   #  ä x""   5øŽ ä x$   ÿ   ä x&7&øŽ ä x5   ÿ   ä x7   ÿ   Ä x%%   4øŽ ä x4#   ÿ   ä x$!   $øŽ ä x!"   5øŽ ä x&&   7øŽ ä x"   ÿ   Ä x7   ÿ   ä x''T  (    ä x44   ÿÀŽ ä x(ÿ   #  ä x%%   "øŽ ä x5   ÿ   Ä x&&   7øŽ ä x"ÿ   '  äx44'øŽ ä x7(   ÿÀŽ ä x'#   ÿ   ä x$$   5øŽ Ä x5   ÿ   ä x77"øŽ ä x'4   'øŽ ä x(#   ÿ   ä x"	   ÿ   ä x6ÿ   #  Ä x&&   5øŽ ä x''   (øŽ ä x%%   "øŽ ä x(#   ÿ   ä x))T  *    ä x7   6øŽ Ä x4#   ÿ   ä x"	   ÿ   â &r$       â x''   (øŽ â &r&&   ÿ   â r ÿ   )  ä x   4øŽ ä x(#	   ÿ   Ä x!!   "øŽ â &r&%   &   â x++T  ,    ä x$ÿ      ä x(   (øŽ â &r!      â x"    ÿ   â &r''   &   â r!ÿ   )  Ä x#ÿ   +  ä x$$   ÿÀŽ ä x""   ÿÀŽ â &r(   '   â x%!   ÿ   ä x$$#øŽ ä x--T  .    Ä x""+øŽ ä x(ÿ   !  ä x#    ÿ   ä x&%   ÿÀŽ ä x%ÿ      ä x'ÿ   -  Ä x((   ÿÀŽ ä x""   #øŽ ä x#    ÿ   ä x$$   %øŽ ä x(('øŽ ä x%    ÿ   Ä x&&-øŽ ä x'!   ÿ   ä x""   #øŽ ä x#    ÿ   ä x$$   %øŽ ä x)ÿ   !  Ä x&&   'øŽ ä x% 	   ÿ   ä x'!   ÿ   ä x""   #øŽ ä x((   )øŽ ä x$$   %øŽ Ä x)!   ÿ   äx&&   'øŽ ä x#!   ÿ   äx%!	   ÿ   â &r!"   ÿ   â x((   )øŽ ä x#&   #øŽ â &r$$   !   â x%(   %øŽ â 0r"ÿ/  0 A   Ä &r##   $   â Er        â 0r/ÿ/    A   â Er        â &r#%   #   â  x""   Á  A  â Er        â 0r0ÿ0    A   ä 0r1ÿ1    A   â Er! #       â #r/   "   â#r/   "   â#r#/   "   â 0r2ÿ2    A   â #r0      â #r1      â 0r3ÿ3    A   â #r/!   "   â G‰  Ð    € ê %x    Ž â r@   ÿàó È �y%     ¢ rA   ÿäÿ  Æ �y&     ¨ �y'    è �y(    è �y)
    ( �y*    ( �y$      " x%T  &    ÈO r%ÿ     är&ÿ     ä x+%   ÿ   ä x''T  (    ä� x(+   ÿÀŽ ä x+&   ÿ   Ä xÿ   &  ä x))T  *    äx*ÿ   %  ä x+   ÿÀŽ ä x!   ÿÀŽ ä x(('øŽ Ä x ÿ   )  ä x%   ÿ   ä x'ÿ   '  ä x**   ÿÀŽ ä x))øŽ ä x+! øŽ Ä x(   øŽ ä x%   ÿ   ä x'*'øŽ ä x ÿ   %  ä x(ÿ   &  ä x   øŽ Ä x*&   ÿ   ä x '    øŽ ä x!%   ÿ   äx%   ÿ   ä x++   (øŽ ä x))   *øŽ Ä x(&   ÿ   ä x     !øŽ ä x   øŽ ä x%%	   ÿ   ä x()   (øŽ â &r   ÿ   â x*&   ÿ   Ä x!&   ÿ   ä x     %øŽ ä x*+   *øŽ ä x!(   !øŽ â &r        â x%&	   ÿ   Æ &r!!       â x%*   %øŽ Ê &r!%   !   Ê Er! !       â 0r$ÿ$    A   È #r!/!   "   È #r$!      ä G‰       € ê x>   ÿàÿ Ê $z! p  Ž È %x!   BŽ È %x !   Ž â r   ÿàó È �y%     ¢ r   ÿäÿ  Æ �y&     ¨ �y'    è �y(    è �y)
    ( �y*    ( �y$      " x%T  &    ÈO r%ÿ     È x&%   ÿ   ä x''T  (    ä� x(&   ÿÀŽ ä r&ÿ     ä x))T  *    äx&   ÿ   Ä x*ÿ   %  ä x   ÿÀŽ ä xÿ   &  ä x+%   ÿ   ä x(('øŽ ä x!&   ÿ   Ä x)øŽ ä x**   ÿÀŽ ä x'ÿ   '  ä x    ÿÀŽ ä x((   +øŽ ä x%   ÿ   Ä x!   !øŽ ä x'*'øŽ ä x)ÿ   )  ä xÿ   %  ä x(   øŽ ä x(&   ÿ   Ä x) )øŽ ä x''   øŽ ä x%   ÿ   äx %   ÿ   ä x!!   (øŽ ä x(ÿ   &  Ä x   øŽ ä x '    øŽ ä x%%	   ÿ   â &r   ÿ   â x))   (øŽ ä x(&   ÿ   äx*&   ÿ   Ä x     %øŽ ä x!!   (øŽ ä x))   *øŽ â &r        â x&&	   ÿ   â 0rÿ$    A   ä &r!!       â x&)   &øŽ Ê &r!&   !   Ê Er! !       Ê #r!/!   "   È #r
!   
   ä G‰       € ê x>   ÿàÿ Ê $z! p  Ž È %x!   BŽ È %x !   Ž â r   ÿàó È �y%     ¢ r   ÿäÿ  Æ �y&     ¨ �y'    è �y(    è �y)
    ( �y*    ( �y$      " x%T  &    ÈO r%ÿ     È x&%   ÿ   ä x''T  (    ä� x(&   ÿÀŽ ä r&ÿ     ä x))T  *    äx&   ÿ   Ä x*ÿ   %  â 0r$ÿ$    A   â x   ÿÀŽ ä xÿ   &  ä x+%   ÿ   ä x(('øŽ ä x!&   ÿ   Ä x)øŽ ä x**   ÿÀŽ ä x'ÿ   '  ä x    ÿÀŽ ä x((   +øŽ ä x%   ÿ   Ä x!   !øŽ ä x'*'øŽ ä x)ÿ   )  ä xÿ   %  ä x(   øŽ ä x(&   ÿ   Ä x) )øŽ ä x''   øŽ ä x%   ÿ   äx %   ÿ   ä x!!   (øŽ ä x(ÿ   &  Ä x   øŽ ä x '    øŽ ä x%%	   ÿ   â &r   ÿ   â x))   (øŽ ä x(&   ÿ   äx*&   ÿ   Ä x     %øŽ ä x!!   (øŽ ä x))   *øŽ â &r        â x&&	   ÿ   Æ &r!!       â x&)   &øŽ Ê &r!&   !   Ê Er! !       Ê #r!/!   "   È #r	$!   	   â #r2#      â #r3      â G‰       € ê x>   ÿàÿ Ê $z p  Ž È %x    Ž È %x   BŽ â �y$     ¨ �y%     ¢ r   ÿàó Æ �y      â  r   ÿäÿ  Ê �y'    ( �y(    ( �y#
    è �y&    â x%$T  %    ÈO r$ÿ   %  ärÿ   %  ä/ x)$   ÿ   ä x+ÿ   $  ä x*)   ÿÀŽ ä x''T  (    Äx   ÿ   ä x**'øŽ äx)$   ÿ   ä x##T  &    ä� x(+   ÿÀŽ ä x'ÿ   '  Ä x&   ÿÀŽ ä x)*   )øŽ ä x'('øŽ ä x $   ÿ   ä x%ÿ     ä x!   ÿ   Ä x&&#øŽ äx(ÿ   $  ä x )    øŽ ä x$   ÿ   ä x*%   ÿÀŽ ä x#ÿ   #  Ä x!&   !øŽ ä x''   (øŽ ä x&$   ÿ   ä x     øŽ ä x(   ÿ   ä x#*#øŽ Ä x*ÿ     ä x&'   &øŽ ä x$	   ÿ   ä x!!   (øŽ â &r    ÿ   â x##   *øŽ ä x$   ÿ   Ä x(   ÿ   ä x&   øŽ ä x!!   $øŽ ä x##   (øŽ â &r      â x	   ÿ   Æ &r!!      â x#   øŽ Ê &r   !   Ê Er        â 0rÿ    A   È #r/   "   È #r      â x@   ÿàÿ ä x€  ÿàõ ä |    pbò ä x€  ÿà÷ ä rÿ   ÿä ä rÿ   ÿäÿ Î D       À â Gy  0ßÿÿÿÿƒ ê Ay        € ê y      "   " Ey  0    € â r ÿ   pRð är ÿ   pRò Ö G‰  ð     € ê z ÿ `  pPð ä r    ÿ8Ž ä z ÿ a   Sð Æ ˆs  þÿ H   ô ˆƒ €þÿ H   è ˆƒ  ÿÿ H   è ˆƒ €ÿÿ H   â G‰  p     € ê ˆs    H   è ˆs €þÿ H   è ˆs 
€  H   è ˆs  ÿÿ H   è ˆs 	   H   è ˆs €ÿÿ H   è ˆs €  H   ä Ay        € ê {          ì M        € ê y      &   ¢ z ÿ `  pPð Æ y      '   "> z ÿ a   Sð â $z x  >Ž âO x        Æ $z ~  Ž È %v j   Ž È G	  P    € ê „y      H   "z > r  p`ð Æ „y  €   H   b r  ÿ   pVp  Æ „y
     H   ¨ „y	     H   è „y  €  H   ( „y     H   ( „y     H   ( „y  €  H   ( „y     H   ( „y  €  H   ( „y  €  H   " !r       Æ „y  €  H   " !r       â/ !r

       âO !r	       ä� !r       â!r

       â !r       ä !r       Æ ‰      b !r

       Æ ‰      ¢ !r       Æ ‰	
      â !r       È !r       â !r       â/ !r       ÈO ‰      " !r
	       Æ� ‰      h ‰	      ¨ ‰      â !r       â !r       È/ ‰ €    " !r	       ÆO ‰	 €    b !r
       Æ� ‰      ¨ ‰
 €    â !r       â x>   ÿàÿ â !r		       Æ/ ‰ @    " z  r  p`ò â !r       äO ‰	 @    b x     pVò  â !r
       ä� ‰ €    ¨ ‰ @    â !r       â x>   ÿàÿ Ä x>>   ÿàÿ â !r	       â/ ‰      " z  r  p`ô â !r       äO ‰      b x     pVt â !r       ä� ‰ @    ¨ ‰      â !r       È !‚       â !r       â/ !r       ÆO †‰     â z > r  p`ð â !’       â !r       â� ‰      " x     pVp  ä !¢	       â †™    è †© 	   â !r       Ê M	        € ê !r       Ê/ †y    â My        € ê z  h      È x    pBð Ú G	  @    € ê x    pRð Ú G‰  À    € ê x    pRð Ú G	   $    € ê „y      H   " z > r  p`ð â Ey  p    € ä „y     H   b r  ÿ   pVp  Æ „y     H   ¨ „y     H   è „y     H   ( „y   
  H   ( „y  €  H   è „y  €  H   è „y  €  H   â !r       â !r       Æ/ !r       äO „y  €  H   â !r       â� !r       ä„y  €   H   â !r       Æ ‰      ( ‰      b !r       â !r       È/ ‰      " !r
       Æ ‰      b !r       â !r       È !r       È !r       â !r       â !r       È/ ‰ €    ( ‰ €    b !r       Æ „y  €
  H   â !r       Æ/ ‰ @    ( ‰ @    b !r       â !r       È/ ‰      ( ‰      b !r       â !r       â !r       â/ G	  �     € ê !r       â!r       Æ  x
;ª¸¿  A  Ì s
 
       $ !t
  €?    Ì s        $  r     A  È  r     A  Ê †y     ä Ay        € ê ‰      b Ey  Ð    € æ ‰      ( „y     H   ¨ „y   	  H   è „y     H   â !r       Æ/ „y     H   â !r       Æ ‰      ( ‰      h „y     H   â !r       âO !r       â !r       È/ ‰
 €    " x>   ÿàÿ Æ ‰ €    b z  r  p`ð â !r       Æ x     pVp  â !r       â !r	       È� !r       â !r

       ä „y     H   â !r       Æ/ ‰
 @    ( ‰ @    b !r
       â !r       È/ ‰      ( ‰
      b !r       â !r       â !r
       â/ G	  �     € ê !r       â!r       Æ  x;ª¸¿  A  Ì s        $ !t	  €?    Ì s	 	       $  r	     A  È  r     A  Ê †y    ä Ay        € ê ‰      " Ey       € æ ‰      h „y  €  H   ¨ „y  €  H   è „y  €  H   è „y  €  H   â !r       â !r	       È/ ‰      ( ‰
	      h „y  €	  H   â !r       âO !r       â !r	
       È/ ‰ €    ( ‰
 €    h „y	  €  H   ¢ !r       â� !r       â !r
       È/ ‰
 @    ( ‰ @    b !r		       âO !r

       â x>   ÿàÿ â !r       Æ/ ‰
      " z  r  p`ð Æ ‰      b x     pVp  â !r
       â !r       Ö/ G	  �     € ê !r       â!r       Æ  x;ª¸¿  A  Ì s        $ !t  €?    Ì s        $  r     A  È  r     A  Ê †y    ä Ay        € ê ‰	      b !r       â x>>   ÿàÿ Æ !r       â z > r  p`ð Æ !r       â x     pVp  È ‰      ¢ !r	       Ê/ ‰      b !r       ÊO ‰      " !r
       Ê/ ‰
 €    b !r       Ê ‰ €    " !r	
       Ê/ ‰	 @    b !r       Ê ‰ @    " !r	       Ê/ ‰      b !r       Ê ‰       ¢  !r       â/ M	        € ì !r       â!r         ÆO  x;ª¸¿  A  â !r        Ê s        $ !t  €?    Ì s        $  r     A  È  r     A  Ê †y    â My        € ê „y      H   " z > r  p`ð â Ey       € ä „y     H   b r  ÿ   pVp  Æ „y     H   ¨ „y     H   è „y     H   ( „y   
  H   ( „y  €  H   ( „y  €  H   ( „y  €
  H   " !r       â !r       Æ/ !r       âO !r       Æ� !r       â!r       È ‰      " !r
       Æ ‰      b !r       Æ „y  €   H   ¢ !r       â !r       â !r       È/ ‰      ( ‰      h „y  €  H   â !r       âO !r       â !r       È/ ‰ €    ( ‰ €    h „y  €  H   ¢ !r       â� !r       â !r       È/ ‰ @    ( ‰ @    b !r       âO !r       â !r       È/ ‰      ( ‰      b !r       â !r       â/ G	  À     € ê !r       È x'7=  A  â x*BL?  A  Æ #t
  €?   È  r

     A  â  x   ?  A  â !r       È s
 
    $   $ !t
  €?    È  r     A  È  r     A  Ê †y     ä Ay        € ê ‰      b Ey       € æ ‰      ( „y     H   ¨ „y     H   è „y     H   ( „y   	  H   â !r       Æ/ „y     H   â !r       Æ „y     H   è ‰      ( ‰      b !r       âO !r	       Æ� !r       â!r       â !r       â !r       Æ/ ‰
 €    " !r       Æ ‰ €    b !r       â !r

       â x>   ÿàÿ â !r       Æ/ ‰
 @    " z  r  p`ð Æ ‰ @    b x     pVp  â !r
       â !r       È/ ‰      ( ‰
      b !r       â !r
       â/ G	  À     € ê !r       È x
'7=  A  â x*BL?  A  Æ #t  €?
   È  r     A  â  x   ?  A  â !r       È s     $   $ !t	  €?    È  r	     A  È  r     A  Ê †y    ä Ay        € ê ‰      " Ey  Ð    € æ ‰      h „y  €  H   ¨ „y  €  H   è „y  €  H   è „y  €  H   â !r       Æ „y  €	  H   â !r	       Æ/ ‰      ( ‰
	      b !r       âO !r       â !r	
       È/ ‰ €    ( ‰
 €    b !r       Æ „y	  €  H   ¢ !r       â !r
       È/ ‰
 @    ( ‰ @    b !r	       âO !r

       â x>   ÿàÿ â !r       Æ/ ‰
      " z  r  p`ð Æ ‰      b x     pVp  â !r
       â !r       Ö/ G	  À     € ê !r       È x'7=  A  â x*BL?  A  Æ #t  €?   È  r     A  â  x   ?  A  â !r       È s     $   $ !t	  €?    È  r	     A  È  r     A  Ê †y    ä Ay        € ê ‰      " !r       â x>>   ÿàÿ Æ !r       â z > r  p`ð Æ !r       â x     pVp  È ‰      b !r       Ê ‰      " !r       Ê/ ‰      b !r	       Ê ‰
	 €    " !r       Ê/ ‰ €    b !r
	
       Ê ‰
 @    " !r       Ê/ ‰ @    b !r
       Ê ‰      " !r       Ê/ ‰       ¢ !r       â M	        € ì !r       â/!r         ÆO  x'7=  A  â x*BL?  A  â !r        ä #t  €?   â x   ?  A  Æ  r     A  Ì s     $   $ !t  €?    È  r     A  È  r     A  Ê †y    â My        € ê x    pRð Ú G‰  P    € ê x    pRð Ú G	  `    € ê „y      H   " z > r  p`ð â Ey       € ä „y     H   b r  ÿ   pVp  Æ „y     H   ¨ „y     H   è „y     H   ( „y   
  H   ( „y  €  H   ( „y  €  H   ( „y  €
  H   " !r       â !r       Æ/ !r       âO !r       Æ� !r       â!r       È ‰      " !r
       Æ ‰      b !r       Æ „y  €   H   ¢ !r       â !r       â !r       È/ ‰      ( ‰      h „y  €  H   â !r       âO !r       â !r       È/ ‰ €    ( ‰ €    h „y  €  H   ¢ !r       â� !r       â !r       È/ ‰ @    ( ‰ @    b !r       âO !r       â !r       È/ ‰      ( ‰      b !r       â !r       â/ G	  À     € ê !r       Ê	z i    � â !r       È  x
;ª¸¿  A  â 	z i    � È 	z i €  � â s
 
       $ !t
  €?    Ì s        $  r     A  È  r     A  Ê †y     ä Ay        € ê ‰      b Ey       € æ ‰      ( „y     H   ¨ „y     H   è „y     H   ( „y   	  H   â !r       Æ/ „y     H   â !r       Æ „y     H   è ‰      ( ‰      b !r       âO !r	       Æ� !r       â!r       â !r       â !r       Æ/ ‰
 €    " !r       Æ ‰ €    b !r       â !r

       â x>   ÿàÿ â !r       Æ/ ‰
 @    " z  r  p`ð Æ ‰ @    b x     pVp  â !r
       â !r       È/ ‰      ( ‰
      b !r       â !r
       â/ G	  À     € ê !r       Ê	z i    � â !r       È  x;ª¸¿  A  â 	z i    � È 	z i €  � â s        $ !t	  €?    Ì s	 	       $  r	     A  È  r     A  Ê †y    ä Ay        € ê ‰      " Ey  Ð    € æ ‰      h „y  €  H   ¨ „y  €  H   è „y  €  H   è „y  €  H   â !r       Æ „y  €	  H   â !r	       Æ/ ‰      ( ‰
	      b !r       âO !r       â !r	
       È/ ‰ €    ( ‰
 €    b !r       Æ „y	  €  H   ¢ !r       â !r
       È/ ‰
 @    ( ‰ @    b !r	       âO !r

       â x>   ÿàÿ â !r       Æ/ ‰
      " z  r  p`ð Æ ‰      b x     pVp  â !r
       â !r       Ö/ G	  À     € ê !r       Ê	z i    � â !r       È  x;ª¸¿  A  â 	z i    � È 	z i €  � â s        $ !t  €?    Ì s	        $  r	     A  È  r     A  Ê †y    ä Ay        € ê ‰      " !r       â x>>   ÿàÿ Æ !r       â z > r  p`ð Æ !r       â x     pVp  È ‰      b !r       Ê ‰      " !r       Ê/ ‰      b !r	       Ê ‰
	 €    " !r       Ê/ ‰ €    b !r
	
       Ê ‰
 @    " !r       Ê/ ‰ @    b !r
       Ê ‰      " !r       Ê/ ‰       ¢ !r       â M	        € ì !r       â/!r         ÈO 	z i    � â !r        È  x;ª¸¿  A  â 	z i    � È 	z i €  � â s        $ !t  €?    Ì s        $  r     A  È  r     A  Ê †y    â My        € ê „y      H   " z > r  p`ð Æ „y     H   b r  ÿ   pVp  Æ „y  €   H   ¨ „y  €  H   è „y"     H   ( „y$     H   ( „y     H   ( „y     H   ( „y  €  H   ( „y  €  H   ( „y     H   " !r       Æ „y   	  H   " !r       Æ/ „y     H   b !r       ÆO „y   
  H   ¢ !r
       Æ� „y  €  H   â !r""       Æ„y   €
  H   " !r$$	       Æ „y&     H   " !r       â x>   ÿàÿ ä „y(     H   " !r       â z  r  p`ò ä „y  €  H   " !r       â x     pVò  Ä „y  €  H   " !r       Æ „y  €  H   " !r"       Æ „y  €	  H   â !r$       Æ „y  €  H   â !r       Æ/ „y  €  H   â !r       âO !r       ä� ‰      " !r        Æ‰      b !r&&       Æ ‰	      ¢ !r((       Æ ‰
      â !r       Æ ‰&      " !r       Æ ‰(      " !r       â !r       â !r       â !r       Æ/ ‰      " !r       â !r		       âO !r       Æ ‰      b !r

       Æ� ‰      ¢ !r&       â!r(       ä ‰      è ‰	      ( ‰
      " !r       Æ ‰      " !r       Æ/ ‰      b !r       ÆO ‰ €    ¨ ‰      ¢ !r       Æ� ‰      â !r	       â!r
       ä ‰ €    " !r       Æ ‰	 €    " !r       Æ/ ‰
 €    b !r       ÆO ‰ €    ¢ !r       Æ ‰ €    ¢ !r       Æ� ‰ €    è ‰ €    â !r       â!r		       ä ‰ @    " !r

       Æ/ ‰ @    b !r       ÆO ‰	 @    ¢ !r       Æ ‰
 @    " !r       Æ� ‰ @    â !r       Æ ‰ @    â x>   ÿàÿ äx>>   ÿàÿ â ‰ @    â !r       â z  r  p`ô ä ‰ @    " !r       â/ x     pVt â !r	       âO ‰      b z > r  p`ö â !r
       Ä‰      ¢ x     pVö â !r       ä� ‰	      â !r       Æ ‰
      " !r       Æ ‰      " !r       Æ ‰      ( ‰      " !r       Æ/ ‰      b !r       âO !‚       â!r	       ä� !‚       â !r
       â!’       ä  ‚     A  â !r       â !’       Æ †‰     â !r	       â !¢       â  ’     A  ä !¢		       â !r       ä †™    â  ¢		     A  â !r       È/ †© 	   â M9        € ê !r       â !r       È  r     A  Ê †y    â My        € ê „y      H   " z > r  p`ð â Ey  À    € ä „y     H   b r  ÿ   pVp  Æ „y     H   ¨ „y     H   è „y     H   ( „y   
  H   ( „y  €  H   ( „y  €  H   ( „y  €
  H   " !r       â !r       Æ/ !r       âO !r       Æ� !r       â!r       È ‰      " !r
       Æ ‰      b !r       Æ „y  €   H   ¢ !r       â !r       â !r       È/ ‰      ( ‰      h „y  €  H   â !r       âO !r       â !r       È/ ‰ €    ( ‰ €    h „y  €  H   ¢ !r       â� !r       â !r       È/ ‰ @    ( ‰ @    b !r       âO !r       â !r       È/ ‰      ( ‰      b !r       â !r       â/ G	  à     € ê !r       Ê	x  à@  � Ê  x#ÛÙ?  A  È  x
;ª¸¿  A  â !r       Ê s
 
       " 	x  à@  � È 	x  àÀ  � Ê !t  €?    â !t
  €?    È s        $  r     A  È  r     A  Ê †y     ä Ay        € ê ‰      b Ey       € æ ‰      ( „y     H   ¨ „y     H   è „y     H   ( „y   	  H   â !r       Æ/ „y     H   â !r       Æ „y     H   è ‰      ( ‰      b !r       âO !r	       Æ� !r       â!r       â !r       â !r       Æ/ ‰
 €    " !r       Æ ‰ €    b !r       â !r

       â x>   ÿàÿ â !r       Æ/ ‰
 @    " z  r  p`ð Æ ‰ @    b x     pVp  â !r
       â !r       È/ ‰      ( ‰
      b !r       â !r
       â/ G	  à     € ê !r       Ê	x  à@  � Ê  x#ÛÙ?  A  È  x;ª¸¿  A  â !r       Ê s        " 	x  à@  � È 	x  àÀ  � Ê !t  €?    â !t	  €?    Ì s	 	       $  r	     A  È  r     A  Ê †y    ä Ay        € ê ‰      " Ey  ð    € æ ‰      h „y  €  H   ¨ „y  €  H   è „y  €  H   è „y  €  H   â !r       Æ „y  €	  H   â !r	       Æ/ ‰      ( ‰
	      b !r       âO !r       â !r	
       È/ ‰ €    ( ‰
 €    b !r       Æ „y	  €  H   ¢ !r       â !r
       È/ ‰
 @    ( ‰ @    b !r	       âO !r

       â x>   ÿàÿ â !r       Æ/ ‰
      " z  r  p`ð Æ ‰      b x     pVp  â !r
       â !r       Ö/ G	  à     € ê !r       Ê	x  à@  � Ê  x#ÛÙ?  A  È  x;ª¸¿  A  â !r       Ê s        " 	x  à@  � È 	x  àÀ  � Ê !t  €?    â !t  €?    È s	        $  r	     A  È  r     A  Ê †y    ä Ay        € ê ‰      " !r       â x>>   ÿàÿ Æ !r       â z > r  p`ð Æ !r       â x     pVp  È ‰      b !r       Ê ‰      " !r       Ê/ ‰      b !r	       Ê ‰
	 €    " !r       Ê/ ‰ €    b !r
	
       Ê ‰
 @    " !r       Ê/ ‰ @    b !r
       Ê ‰      " !r       Ê/ ‰       ¢ !r       â M	        € ì !r       â/!r         ÈO 	x  à@  � â !r        È  x#ÛÙ?  A  â 	x  à@  � Æ  x;ª¸¿  A  â 	x  àÀ  � Ê s        " !t  €?    â !t  €?    Ì s        $  r     A  È  r     A  Ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      "   $ z  w  p`ð Ú M	        € ê y      !   " ¹z  l      â Ey  Ð    € â ™x?    â $rÿÿ   ÿ Ž â ¹z  F   
   â $rÿÿ   ÿ Ž â xÿ     È |    p`ð Ú G	  P    € ê y      &   " $tÿ   ÿ Ž ä $z x  Ž È %v \   Ž Ê �y    ™ ¢  $rÿÿ   ÿ Ž ä $z r  ÿŽ â y	      %   b 'z  m   Ž È %x$   ÿ Ž â z ÿ n     Æ $x   ÿ Ž ä$rÿÿ    
Ž ä $x   ÿ Ž â x   ÿÀŽ â $z o  Ž â x   ÿ   â $rÿÿ   ÿ Ž â x   ÿÀŽ â $z  u  ÿŽ â x   ÿÀŽ â %x$    Ž â x   ÿüŽ Ä r    à£ â %x   ÿ Ž â z Z  ÿàñ ä x		   ÿ   â/ $x   
Ž â r
ÿÿ   ÿDþ  â %x$    Ž â z [  ä  ä x	   ÿàÿ â$z	 q  Ž â z( Z  ÿàñ â $x	
$   ÿŽ Ä %x $   ÿ Ž Æ z [  	ä  â $x   ÿ Ž â x((   ÿàñ â $r ÿÿ    Ž ä $rÿÿ    Ž â x   ÿÀŽ â $vÿ X  ÿ Ž â r)ÿ   ÿä  â $vÿ Y  ÿ Ž ä $rÿÿ   ÿ Ž Ä $x   ÿ Ž ä $z t  Ž äO$z t  Ž ä $z q  Ž ä %x   Ž È %x   Ž È $rÿÿ    Ž â r   ÿàñ â $r	ÿÿ    Ž È %x   Ž È $x     â �y   • ¨  �y   • ¨  �y   • è �y   • è �y 
   • ( �y!   • " %x   Ž â r
   ÿàñ Ä �y    • f $x     â �y%   • h �y$   • h �y&
   • h �y#
   • h �y"

   • h �y'
   • b r    ÿàñ Æ/ �y    • b $r
ÿÿ   ( Ž ä $rÿÿ   ) Ž Ê �y(
ðÿÿ ™ h �y*
    ™ b $x     Æ �y)
ôÿÿ ™ h �y+
   ™ h �y    ™ b x   ÿàÿ È |    p`ð ä x  @  ÿàó ä x`  ÿàõ Æ $rÿÿ   Ž  ä $rÿÿ    â xT      ÈO rÿ     äxT      ä� x   ÿ   ä rÿ     ä x  T  !    äx!ÿ     Ä x   ÿÀŽ ä/ xÿ     äx,!   ÿÀŽ â $x   ÿ Ž â xøŽ ä x,øŽ â $x,   ÿ Ž â x!ÿ     Ä x	   ÿÀŽ ä x   ,øŽ â $x,   ÿ Ž â xÿ     ä xÿ      äx!!   ÿÀŽ ä x		 øŽ Ä x   øŽ â $x   ÿ Ž â x!øŽ â $x    ÿ Ž â x		   ,øŽ â $x,   ÿ Ž â x!ÿ     ä x   øŽ ä x   !øŽ â $x!   ÿ Ž â x    øŽ Ä x	   ,øŽ â $x,   ÿ Ž âx    ÿ   ä x$%T  $    äx   !øŽ â $x!   ÿ Ž â x	    øŽ ä rÿ   $  Ä x   ,øŽ ä rÿ   $  â $x    ÿ Ž â x   !øŽ ä x#&T  #    â $x&   ÿ Ž â x!ÿ     ä x     ÿÀŽ Ä x""T  '    ä x$ÿ   #  äx%!   ÿÀŽ ä x'&   ÿÀŽ ä x! #øŽ ä x&ÿ     Ä x#%$øŽ ä x$'"øŽ äx ÿ   "  â $x"   ÿ Ž â x'&   ÿÀŽ â $x%   ÿ Ž â x&ÿ     ä x!!   "øŽ Ä x"#   &øŽ â $x#   ÿ Ž â x$$   %øŽ â $x&   ÿ Ž âx%   ÿ   äx#$   #øŽ â $x$   ÿ Ž â x!!   &øŽ â $x   ÿ Ž â x ' øŽ Ä x'ÿ     ä x""   %øŽ ä x!!   $øŽ â $x%   ÿ Ž âx     'øŽ â $x&   ÿ Ž â x"   øŽ â &r		(   ÿ   ä&r!!(   ÿ   â x     %øŽ â $x   ÿ Ž â x"#   &øŽ â $x   ÿ Ž Ä &r	*   	   ä&r*   !   â x   øŽ â &r)   	   âx    øŽ â &r"")      ä &r+      ä&r+   "   Ä 0r	ÿ  0 A   âEr        â 0rÿ    A   â Er        ä  x		   A  A  â x(
@  ÿà÷ Æ #r   	  â#r   	  â 0rÿ    A   ä 0r	ÿ    A   â r)ÿ   ÿäÿ ä #r      ä #r	      â G‰  ðöÿÿÿÿƒ ê Ay        € ê ‰      ( ‰      h y	      %   ¢ !r       â !r       È/ ‰       " $x
	   Ž ÆO ‰      d z 
 p  p`ð È x    pDp  â !r         â !r       È/ ‰  €    ( ‰ €    b !r        â !r       È/ ‰ @    ( ‰ @    b !r       â !r       È/ ‰      h  ‰	      ¢  M	        € ê y      &   â $x   ÿ Ž â !r       â/ $z
 s  
Ž â !r		       âO $tÿ   ÿ Ž â x    p"ò Ä r ÿ   p"ð Ö $ÿÿ   	 Ž ä $z v  
Ž ÈŸ %v j   Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      "   $ z  w  p`ð Ú M	        € ê y      &   " $t%ÿ   ÿ Ž â ¹z  F   
   â $z$ x  Ž È %v$$ \  % Ž Ì �y$$    ™ b ¹z  l      â Ey  À    € â ™x?    â y#      !   b $rÿÿ   ÿ Ž â x       ÿ  â $r
ÿÿ   ÿ Ž ä $r"ÿÿ   ÿ Ž ä $rÿÿ   ÿ Ž â xÿ   #  È/ |    p`ð Ú G	      € ê $rÿÿ   ÿ Ž â y      %   " $x#   ÿ Ž âx#   ÿ   â 'z  m   Ž â x#   ÿ   â Ey €    € â x	   ÿÀŽ â $z r  ÿŽ â z ÿ n     â $rÿÿ   ÿ Ž â x   ÿÀŽ â %x$   ÿ Ž â r ÿ   ÿáÿ Æ $x #   ÿ Ž ä$z o  Ž â x#   ÿÀŽ â $z
$ t  Ž âx     ÿÀŽ â $vÿ X  ÿ Ž ä $vÿ Y  ÿ Ž â x    ÿüŽ â $z  u  ÿŽ Ä %x   ÿ Ž Æ r    à‘ â $x   	
Ž ä $x   ÿ Ž â rÿÿ   ÿ$~  â %x$    Ž Æ x   ÿàÿ â %x$    Ž â z Z  ÿàó Æ $x$   ÿŽ â z Z  ÿàñ â $z q  Ž â z	 [  äÿ  â %x $   ÿ Ž â z& [  ä  ä z ÿ `  pPð â $z q  
Ž â x   ÿàó â $z$ t  Ž â z ÿ a   Sð â $rÿÿ    Ž â r&ÿ&   ÿäÿ  â $rÿÿ    Ž Ä $vÿ `  ÿ Ž ä $vÿ a  ÿ Ž ä $r"ÿÿ   ÿ Ž ä $x    ÿ Ž ä %x   Ž È %x   Ž È $rÿÿ    Ž â r   ÿàó â $r	ÿÿ    Ž È %x
   Ž È $x   Ž  â �y+
   • ¨  �y*
   • ¨  �y)   • è �y,   • è �y-
   • ( �y.   • " %x   Ž â r   ÿàó Ä �y

    • b $rÿÿ    Ž ä $r	ÿÿ   & Ž â �y4   • b $x   Ž  Æ �y1   • h �y&ðÿÿ ™ h �y%    ™ h �y0
   • h �y/   • h �y2   • h �y3   • b r   ÿàó Æ/ �y'ôÿÿ ™ b r   ÿäÿ  Æ �y(   ™ h �y    ™ h �y    • b x   ÿàÿ È |    p`ò ä x++T  *    ÈO rÿ   +  Ê $x   ÿ Ž â x))T  ,    ä� rÿ    +  ä x*   ÿÀŽ ä x+ÿ     â $x   ÿ Ž â x*)øŽ â $x*   ÿ Ž â x--T  .    Äxÿ   )  ä x,+   ÿÀŽ ä x.   ÿÀŽ ä x)ÿ     ä x   *øŽ â $x*   ÿ Ž â x,øŽ Ä x.-øŽ â/$x.   ÿ Ž â xÿ   -  ä x))   ÿÀŽ ä x,ÿ     ä x   *øŽ â $x*   ÿ Ž â x)øŽ Ä x   ,øŽ â $x,   ÿ Ž â x)ÿ     ä x   .øŽ ä x.   ÿ   äx   *øŽ â $x*   ÿ Ž â x   )øŽ â $x   ÿ Ž â x   ,øŽ â $x)   ÿ Ž â x.   .øŽ Ä x   *øŽ ä x   øŽ â &r&   ÿ   âx*.   )øŽ â $x)#   ÿ Ž â x14T  1    â &r%      Æ rÿ    1  ä x)   ÿÀŽ ä x/0T  /    â $x)   ÿ Ž â x+ÿ     ä rÿ   1  ä x-ÿ   /  Ä x0+   ÿÀŽ ä x.)   ÿÀŽ ä x,   ÿ   ä x)0-øŽ ä x22T  3    ä x+./øŽ Ä x-,   ÿÀŽ ä x/ÿ     â $x,   ÿ Ž â x--2øŽ â$x0   ÿ Ž â x.ÿ   2  ä x//   ÿÀŽ ä x,-   ,øŽ â $x-   ÿ Ž â x./.øŽ Ä x/ÿ     ä x++   0øŽ ä x0ÿ     ä x..   /øŽ â $x/   ÿ Ž âx,,   -øŽ â $x-   ÿ Ž â x))   0øŽ â $x0   ÿ Ž â x..   /øŽ â $x   ÿ Ž â x-,   -øŽ Ä x++   0øŽ â $x0   ÿ Ž âx2   ÿ   â &r,-&   ÿ   â x.   øŽ â $x   ÿ Ž â x))   2øŽ â $x.   ÿ Ž â x++   0øŽ â &r,%   ,   â x   øŽ â &r*'      âx))   .øŽ â &r++'   ,   Ä 0rÿ  0 A   ä&r(      ä&r))(   +   â  x   Á  A  â 0rÿ    A   â Er        ä Er )       â 0r)ÿ    A   Ä 0r*ÿ
    A   â #r      â#r      â G‰  0    € ê xÿ   #  Ê $z q  Ž È $z$ t  Ž È %x
   Ž È %x   Ž â �y,
    ¨ �y1
    ¢ r   ÿàõ Æ �y+
     ä  $x    Ê �y.    ( �y/    ( �y-
    ( �y0    " $x2#   ÿ Ž Ê x22   ÿÀŽ ä x1,T  1    ÈO r,ÿ2   1  È x2,   ÿ   ä xÿ   ,  ä/ x..T  /    äx/2   ÿÀŽ ä x
-T  0    ä r-ÿ    1  Ä x/.øŽ â$x,   ÿ Ž ä $x/-   ÿ Ž â x   ÿÀŽ ä x.ÿ   .  ä x//   ÿÀŽ ä x   øŽ Ä x.øŽ â $x.,   ÿ Ž â x/
øŽ ä x/ÿ   ,  ä x1ÿ   -  â $x0-   ÿ Ž â x   .øŽ ä x/   /øŽ â $x,   ÿ Ž âx.,   ÿ   Ä x11   ÿÀŽ ä x
ÿ   
  ä x   0øŽ â $x0-   ÿ Ž â x./   .øŽ â $x/,   ÿ Ž â x
1
øŽ ä x1ÿ   -  Ä x   øŽ â $x-   ÿ Ž â x   0øŽ â $x,-   ÿ Ž â x1
   1øŽ â &r
&   ÿ   â x..   /øŽ â $x--   ÿ Ž â x   øŽ ä x,1   ,øŽ â &r
.%   
   Æ x,,   -øŽ â &r'   
   È &r,(      Ê Er        â 0r
ÿ+    A   È� #r      È #r!
   !   â #r*      â #r )       â G‰  ð    € ê $rÿÿ    Ž â r
   ÿàõ â $rÿÿ    Ž Æ r   ÿä â %x   Ž È �y+
    ¨  �y    è �y*    è �y,
    ¨  �y-

    ( �y.
    ( �y     b x)T  *    È� rÿ   )  är)ÿ    )  Æ $x*   ÿ Ž â xÿ   )  ä x++T  ,    äO x**   ÿÀŽ ä x--T  .    â$x.)   ÿ Ž â x,ÿ     ä x   ÿÀŽ â/ $x   ÿ Ž â x
*+øŽ Ä xÿ   -  ä x+ÿ   +  ä x,,   ÿÀŽ ä x..   ÿÀŽ ä x*øŽ â $x)   ÿ Ž â x+,+øŽ Ä xÿ     ä x

   øŽ â $x   ÿ Ž â x..-øŽ ä x+   øŽ ä x..   øŽ â $x   ÿ Ž â x+ÿ   )  Ä x

   øŽ ä x   ÿ   â$x   ÿ Ž â x**   +øŽ â $x+)   ÿ Ž âx   øŽ â $x)   ÿ Ž âx
   øŽ â $x
)   ÿ Ž â x+.   +øŽ Ä x   øŽ â &r&&   ÿ   â x*   øŽ â $x*)   ÿ Ž â x
+   
øŽ â &r%   &   Æ x   *øŽ â &r'
'      È &r(   '   Ê Er        â 0rÿ    A   È#r      È #r"   "   â x`  ÿà÷ ä x@  ÿàù ä x`  ÿàû â $rÿÿ   Ž â x@  ÿàõ â $r&ÿÿ   	 ä $rÿÿ   Ž â rÿ   ÿä â D       À â Gy  `îÿÿÿÿƒ ê Ay       € ê $rÿÿ   ! Ž ä $r
ÿÿ   " Ž Ä Ay        € ê z ÿ `  pPð È z ÿ a   Sð Ú G	  P    € ê ‰       ( ‰       b !r         â !r        È/ ‰       ( ‰      b !r        â !r       È/ ‰ €    ( ‰ €    b !r       â !r       È/ ‰ @    ( ‰ @    b !r       â !r       È/ ‰       ( ‰	      b !r        â !r		       â/ Gy  €    € ê ‰       ( ‰!      h ‰       ¨ ‰"      â !r         â !r!       â/ !r        ÆO ‰       " !r
"       Æ� ‰      h ‰      ¨ ‰	
      â !r        â !r       â/ !r       ÈO ‰ €    " !r
	       Æ� ‰ €    h ‰	 €    ¨ ‰  €    â !r       â !r       È/ ‰ @    " !r		       ÆO ‰ @    b !r
        Æ� ‰ 	 @    ¨ ‰
 @    â !r       â !r       È/ ‰      " !r	        ÆO ‰       b !r
       Æ� ‰      ¨ ‰      â !r       â !r        â/ !r	       âO !r
       â� y       %   $ $x    #Ž Ê z  p  p`ð È x #   pDp  Ú M	        € ê z ÿ ^  pPð â $z$$ v  Ž Æz ÿ _   Sð Ú $ÿ   ÿ Ž È %$ ^   Ž Ì �	     ¢ x #   ÿ   È r  ÿ   p"ò äx     p"ô äx     p"ö äx     p"ø Î $ ÿÿ    Ž â z ÿ `  pPò â $" ÿÿ   	 Ž ä $2 ÿÿ    Ž â z ÿ a  Sò â $B ÿÿ   
 Ž È !         ÐO G™       € ê z ÿ b  pPð È z ÿ c   Sð Ú $ÿ   ÿ Ž È %$ b   Ž Ì �	     ¢ $x##   ÿ Ž ä $vÿ h  ÿ Ž Æ x #øÿÿÿp"ò äx #üÿÿÿp"ô är #ÿ   p"ö äx #   p"ø Î         ä x    pBò â $"ÿÿ   	 Ž ä $2ÿÿ    Ž ä $Bÿÿ   
 Ž È !       ÈO G  P    € ê x    pRð Ú G‰  �     € ê x    pRð Ú G	  ð    € ê  x;ª¸¿  A  Ì s        $ !t  €?    Ì s        $  r     A  È  r       A  â Gy  P    € ê  x'7=  A  â x*BL?  A  Æ #t  €?   â  x   ?  A  Æ  r     A  Ì s     $   $ !t  €?    È  r     A  È  r       A  â Gy  °    € ê x    pRð Ú G‰  à     € ê x    pRð Ú G	        € ê 	z i    � ä 	z   i    � Æ  x;ª¸¿  A  â 	z   i €  � Ê s        $ !t  €?    Ì s        $  r     A  È  r       A  â Gy  Ð     € ê  r       A  â Gy  °     € ê 	x  à@  � ä 	x    à@  � Æ  x#ÛÙ?  A  â 	x    àÀ  � Æ  x;ª¸¿  A  ä !t   €?    È s        $ !t  €?    Ì s        $  r      A  È  r       A  ä y      "   ( y      &   b $z s  Ž ä $tÿ   ÿ Ž ä $z v  Ž È/ %v j   Ž Ê †y      â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      "   " ¹z  l      â Ey  €    € â ™x?    â y      !   b ¹z  F   
   â r ÿ       â x      ÿ  â y$      %   ¢ r# ÿ       â x      ÿ  â r* ÿ       â x       ÿ  â r. ÿ       Ä r0 ÿ       ä r, ÿ       ä r ÿ       ä r< ÿ       ä r  ÿ       ä r ÿ   pRð Ä r ÿ   pRò är   ÿ(Ž ä/ r ÿ       ä r   ÿ� ä x$$   ÿ   äO xÿ     È |    pbô Ú G)  à    € ê y      '   " r ÿ       â $z	$ p  Ž â r ÿ       â y      &   b x
        ä r ÿ       ä z5 q  ÿàÿ â 'z y   Ž È 'z s   Ž â/ zÿ z    Æ $z( }  ÿŽ â zÿ t    â $z |  Ž ä$z |  	Ž â x	$   ÿàÿ â $z v  Ž ä$z v  Ž ä $z	 p  Ž â x	þÿÿÿÀŽ â %v X  
Ž â z  q      Ä r		  €ÿàÿ â %v X  
Ž â r   ÿŽ äx/	   ÿ   â $z w  ÿŽ â r+   ÿŽ â$x)   Ž âx   ÿàù â $x-   Ž âx   ÿà÷ â $x1   Ž âr((   ÿàõ â $x3   Ž â xÿ   /  Ä rÿ   ÿä ä rÿ   ÿäÿ ä r@ÿÿ   ÿä ä x9/   ÿ   äx7/     ä r	(   ÿàõ â %x
/   Ž âxD $       â �yøÿÿ ™ ¢ rÿ@   ÿä â %x/   Ž Ä �y4
üÿÿ ™ ä  %v	 Z  D Ž ä �y:üÿÿ ™ $ $x$   ÿŽ â r9   ÿàõ â �y=
    ™ f r		   ÿàÿ â �y&    ™ ¦ r	7   ÿä â �y;    ™ ¢  r5(   ÿàõ Æ �y2   ™ ¢  rÿ@   ÿä Æ �y>   ™ ¢  r(   ÿà÷ â/ %v Z  D Ž ä �y'   ™ ¤ $x$   ÿŽ â rÿ@   ÿäÿ â �y8   ™ ¢ r
9   ÿàõ ä r   ÿàÿ â %v Z  D Ž â �y6øÿÿ ™ ¤ r7   ÿä â $x	$   ÿŽ â �yA    ™ ¢  r9   ÿàõ Æ/ r	   ÿàÿ â �y"
   ™ ¦ r7   ÿä â �y%
   ™ ¢ r)(   ÿàõ Æ �y?
   ™ ¢ rÿ@   ÿä â %v Z  D Ž ä �yB
   ™ ¤ $x	$   ÿŽ ä �yE   ™ ¢ r9   ÿàõ ä r	   ÿàÿ â �yC   ™ ¦ r	7   ÿä â �yF   ™ ¢ r+(   ÿàõ Æ �yG    ™ ¢  r
ÿ@   ÿä Æ/ �yH   ™ ¢ %v Z  D Ž Æ �yJ   ™ ¢  $x
$   ÿŽ Æ �yI   ™ ¢  r
9   ÿàõ ä r   ÿàÿ â/ �yK   ™ ¨ �y    ™ ¢  r7   ÿä ä r-(   ÿàõ â �yM   ™ ¦ rÿ@   ÿä â �yL
   ™ ¨ $x$   ÿŽ â �yN
   ™ ¢  r        â %v Z  D Ž Ä �y
   ™ ¨  �yO    ™ ¢  r9   ÿàõ ä r   ÿàÿ â �yQ
   ™ ¦  r7   ÿä Ä r	1(   ÿàõ â/�y    ™ ¢ r3(   ÿà÷ ä rÿ@   ÿä â �yP   ™ ¢ %v	 Z  D Ž â r
ÿ@   ÿäÿ ä �yR   ™ ¢  $xU$   ÿŽ â r9   ÿàõ Ä �yS   ™ ¢  %v Z  D Ž â r		U   ÿàÿ ä �yT   ™ ¢  $x
$   ÿŽ â r	7   ÿä ä r
9   ÿàõ â �y	    ™ ¢  r   ÿàÿ Æ �yD   ™ ¢ r7   ÿä Æ �y   ™ ¨$ �y
   ™ ¨ �y   ™ ¨ �yW
   ™ ¨
 �yV   ™ ¨ �y    ™ ¨  �yU
   ™ ¨
 �yX
   ™ ¢
 xY4ÿÀŽ Ä� x4ÿ   4  ä x[:ÿÀŽ äx4ÿÀŽ ä x
=ÿÀŽ äxZÿ   :  â &rY2   ÿ   äO&r][2   ÿ   Ä &r2>      â xÿ   =  â 2r;        ä&r4
'   2   âxÿÀŽ ä xZÿÀŽ â 0r2ÿ  0 A   ä&r48   4   â x&ÿÀŽ â &r]>   ]   â x=ÿ   &  â 0rÿ    A   â Er4 4       â  x22   ?  A  â x==ÿÀŽ â &r''   ]   â 2r;6;        â #r4   2   â &r2Y"   ÿ   Ä &r'=8   '   ä 0r&ÿ;  0 A   ä&r2%   2   â !r       â &r4["   ÿ   â Er" '       â 0r;ÿ;    A   â  x&&   ?  A  â &r2
?   2   â 2rA        â &r4%   4   â #r;;"   &   â &r2B   2   Ä 0r"ÿ  0 A   ä&r'YE   ÿ   â 2rA6A        â &r4?   4   â Er2 2       â 0rÿ    A   â  x""   ?  A  â &r'C   '   ä &r4=B   4   â #r"2   "   â 0r%ÿA  0 A   â2rG        â &r'
F   '   â Er4 4       â 0rAÿA    A   â  x&%   ?  A  â !r!"!       â &r'H   '   Ä 0r"ÿ  0 A   â#rAA4   &   â &r8YJ   ÿ   â Er& '       â 0rÿ    A   â  x""   ?  A  â &r8I   8   â 2rG6G        ä #r"&   "   â &r8
K   8   â 2r        â0r%ÿG  0 A   â !r"       â 2r"6        â &r8M   8   Ä &rJ[J   ÿ   ä 0rÿ  0 A   â x4%   ?  A  â &rJI   J   â Er8 8       â 0rÿ    A   ä 0r&ÿ"  0 A   ä0r%ÿ"    A   â  x"   ?  A  â &r'YL   ÿ   Ä &r[L   ÿ   â #r"8   "   â &rJK   J   ä &r'N   '   ä&rN      â !r"       â &rJ=M   J   â 2rO        â &r"
   '   Ä &r      â ErJ J       â &rE[E   ÿ   â  x&&   ?  A  â &r"Q   "   ä&rQ=Q      ä &rEC   E   ä 0rÿ  0 A   â#r%%J   &   â &rEF   E   â Er" "       â 0rÿ    A   â  x   ?  A  â !r%       â &r%YP   ÿ   â 2rO6O        â &rE=H   E   â #r"      â &r%R   %   Ä 0rGÿG    A   â Er2 E       â !r<<       â 0rÿO  0 A   â2r        â&r%
S   %   â 2r6        â #rGG2   4   â 0rOÿO    A   â Er& Q       â  x2   ?  A  â &r%T   %   Ä 0rÿ  0 A   ä 0rÿ  0 A   â #r&O&   2   â 0rÿ    A   â Er" %       â  x   ?  A  â  x2   ?  A  â &rYD   ÿ   ä #r"      â &r%[D   ÿ   Ä &r      â !r,,       â &rYY   ÿ   â2r	        â &r'[P   ÿ   ä&r[   ÿ   ä &r%   %   ä &r
      Ä &r'R   '   ä &rYW   Y   ä&rW      â x€  ÿàõ â &r%   %   â2r	6	        â &rV      â 2r        â0rÿ  0 A   â x    ÿàÿ â &r'S   '   â 2r6        â &rY
U   Y   Ä &rU      â rÿ   ÿä â 0rÿ    A   â Er        â  x   ?  A  â 0rÿ	  0 A   â |    pbô â &r'=T   '   ä&r%=V   %   ä &rYX   Y   Ä &r=X      ä 0rÿ  0 A   ä0r
ÿ  0 A   â #r      â !r&       â  x   ?  A  â 0rÿ    A   â Er& '       â 0r	ÿ	    A   â Er %       â 0rÿ    A   â ErY Y       â 0rÿ    A   â Er        â  x   ?  A  â  x

   ?  A  â x€  ÿà÷ â #r&   2   â #r		      â #rY      â #r   
   â rÿ   ÿäÿ â !r ;        â !rA       â !r G        â !r00       â !r..       â !r*	*       â !r##       â !r       â x    ÿàÿ Ä x))    ÿàÿ ä x++    ÿàÿ ä x--    ÿàÿ ä x11    ÿàÿ ä x33    ÿàÿ ä x55    ÿàÿ â D)       À â Gy  °ìÿÿÿÿƒ ê Ay        € ê Ey  0    € â G‰      € ê r   ÿHŽ Ê ˆs  øÿ H   è ˆs  €øÿ H   è ˆs ! ùÿ H   è ˆs €ùÿ H   è ˆs  úÿ H   è ˆs  €úÿ H   è ˆs  ûÿ H   è ˆs €ûÿ H   è ˆs < üÿ H   è ˆs €üÿ H   è ˆs , ýÿ H   è ˆs 0€ýÿ H   è ˆs . þÿ H   è ˆs *€þÿ H   è ˆs # ÿÿ H   è ˆs €ÿÿ H   ä Ay        € ê {          ì M        € ê „y     H   b z $ r  p`ð Ä z  r      â „y €   H   ¢ r ÿ   p"p ä r   ÿŽ â „y	 €  H   è „y €  H   ( „y    H   h „y    H   ( y      &   è y      '   ( „y    H   è „y    H   â !r       Æ/ „y    H   â !r        ÆO ‰      b !r 	        âŸ $z  x  $Ž ä „y €  H   â x$$   ÿàÿ â !r       â$z  ~   Ž â ‰       â !r       Æ    ÿàó â „y €  H   â !r!!       ä ÿÿ   ÿäÿ  â ‰      " 
 j  ÿˆ äz $ r  p`ò â „y' €  H   â 
 k   â !r       â/ x    p"ò â ‰      b !r       Æ ‰      ¢ !r<<       Æ „y)    H   â !r,,       Æ „y+ €  H   â    ÿàÿ Æ „y&    H   â !r$       â !r%       ä ‰      " r   ÿŽ â !r
       â/ 
  r  ÿàõ â ‰!      b !r       äO "ÿÿ   ÿä â ‰
      ¢ 
 j  ÿ„ â !r(       â r    ÿ† â ‰	 €    â 
 k  " â !r"        â rÿÿ   ÿäÿ â ‰      " z j  ÿ† â !r0'0       Ä ‰'<      b z k  � â !r       â     ÿàù â ‰-0      "    ÿàÿ â !r.).       â ÿÿ   ÿä â ‰      " !r!!       â/ !r*+*       â 
 j  ÿˆ â!r
       âO ‰)(      b !r2&#       â 
 k   â !r	       â� $x	   ÿŽ â‰ €    ¢ r    ÿ„ â !r       Ä‰ @    â r		    ÿà÷ â !r<<'       âr ÿÿ   ÿä â ‰&%      " rÿÿ   ÿäÿ â !r00-       â z	 j  ÿ† â ‰!      " z
 j  ÿ„ â !r       â z		 k  � â ‰+,      b r  €ÿŽ Ä    ÿàÿ â ‰#$      b !r(()       â/ z k    ä ‰/.      b !r       âO     ÿàù â !r       â� ‰1*      ¤ 4ÿÿ   ÿä â !r&%&       â„y €  H   â  j  ÿˆ â !r!!       Ä ‰32      "  k  4 â !r,,+       ä‰      â     ÿà÷ â !r$$#       ä ‰"      " !r../       Æ/ ‰      b !r**1       ÆO ‰)&      ¨ ‰+<      b !r223       Æ ‰-(      " !r        Æ� ‰! €    â     ÿàõ â !r""       ä‰/,      " 6ÿÿ   ÿä â !r4       ä ‰10      " !r;       â/ ÿÿ   ÿäÿ Ä ‰ €    b  j  ÿ„ â!r)&)       âO 
 j  ÿ† â ‰'$ €    ¢ !r+<+       â k  6 ä ‰3.      b !r-(-       â 
 k  � Ä ‰5*      " !r!!       â�     ÿàõ ä ‰ @    â !r/,/       â6ÿÿ   ÿä ä ‰72      " !r101       â  j  ÿ„ Ä ‰4      " !r       â/  k  6 ä ‰#  €    b !r$'       ÆO ‰%" €    ¢ !r3.3       â!r5*5       ä ‰$) €    " !r       Æ� ‰&+ €    â !r727       Æ‰(- €    " !r44       Æ ‰! @    b !r #       â/ x#        ä ‰*/ €    b !r"%       ÆO ‰,1 €    ¢ %v  j  # Ž â !r$)$       ä ‰  @    " !r&+&       Æ� ‰.3 €    â !r(-(       Æ‰05 €    " !r!       Æ‰27 €    h ‰94      b !r*/*       Æ/ ‰      b !r,1,       ÆO ‰" @    ¢ !r         Æ ‰# @    " !r.3.       Æ� ‰% @    â !r050       Æ‰!$ @    " !r272       â!r949       ä ‰'& @    b !r=       Æ/ ‰)( @    b !r""       ÆO ‰      ¢ !r##       Æ ‰+* @    " !r%%       Æ� ‰-, @    â !r!$!       Æ‰       ( ‰/. @    " !r'&'       Æ‰10 @    b !r)()       Æ/ ‰49 €    b !r       ÆO ‰32 @    ¢ !r+*+       Æ ‰"      " !r-,-       Æ� ‰#      â !r        Æ‰%      " !r/./       â !r101       ä†	 ;    â !r494       Æ/ † =   â !r323       ÆO ‰!      b !r"       Æ ‰'      " !r##       Æ� ‰)      ¢ !r%%       Æ†	     è ‰ +      è ‰-      ( †     â !r!!       Æ/ ‰/      b !r''       Æ ‰"1      " !r))       ÆO ‰4 @    ¨ ‰3      b !r++        Æ� †	     â !r--       Æ† #   è †	 %    â !r//       â/ !r11"       ä † !    â !r44       ÆO †	 
'    â !r33       Æ† 
)   è †	 +    è † -    è †	 /    è † 1   è ‰4      ( †	 3    â M™        € ê x   ÿàÿ â !r4       ÆŸ r    ÿàñ È r ÿÿ   ÿä  ä z j  ÿ€ È z k     Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      "   " ¹z  l      â Ey       € â ™x?    â y      !   b ¹z  F   
   â r ÿ       â x.      ÿ  â y      %   ¢ r ÿ       â x      ÿ  â r ÿ       â x      ÿ  â r- ÿ       Ä r ÿ       ä r5 ÿ       ä r9 ÿ       ä r  ÿ       ä r ÿ   pRð är ÿ   pRò Ä r   ÿ(Ž È/ r   ÿ� ä x   ÿ   äO x+ÿ     È | +   pbô Ú G)       € ê y      '   " r ÿ       â $z	 p  +Ž âr ÿ       â y      &   b x
        ä z*  q      ä r ÿ       ä r*+   ÿŽ â$x$*   +Ž âr&*+   ÿŽ â$x(*   +Ž âz,+ q  ÿàÿ â $x**   +Ž Ä 'z y   Ž È 'z s   Ž â/ zÿ z    Æ $z# }  ÿŽ â zÿ t    â $z |  +Ž ä$z |  	Ž â x	   ÿàÿ â $z v  Ž ä$z v  Ž ä $z	 p  Ž â x	þÿÿÿÀŽ â $z w  ÿŽ Ä %v X  
Ž âr		  €ÿàÿ ä r##   ÿàõ â %v X  
Ž â x%	   ÿ   ä x   ÿàù ä x   ÿà÷ ä xÿ   %  Ä rÿ   ÿä ä rÿ   ÿäÿ ä rEÿÿ   ÿä ä x)%   ÿ   äx'%     ä r	+#   ÿàõ â %x
%   Ž âx> $       â �yøÿÿ ™ ¢ rÿE   ÿä â %x%   Ž Ä �y1
üÿÿ ™ ä  %v	 Z  > Ž ä �y8üÿÿ ™ $ $x$   ÿŽ â r)   ÿàõ â �y6
    ™ f r		   ÿàÿ â �y!    ™ ¦ r	'   ÿä â �y3    ™ ¢  r,#   ÿàõ Æ �y0   ™ ¢  rÿE   ÿä Æ �y7   ™ ¢  %v Z  > Ž Æ �y"   ™ ¢ $x$   ÿŽ â r)   ÿàõ ä �y4   ™ ¤ r   ÿàÿ ä �y2øÿÿ ™ ¢ r#   ÿà÷ ä r'   ÿä â �y;    ™ ¢  rÿE   ÿäÿ Æ �y   ™ ¢  %v
 Z  > Ž Æ �y    ™ ¢  $x	$   ÿŽ â r)
   ÿàõ ä/ �y:   ™ ¤  r	   ÿàÿ ä �y<   ™ ¤  r'   ÿä Ä r$#   ÿàõ â �yA
    ™ ¦ rÿE   ÿä â �y?   ™ ¢  %v Z  > Ž Æ �y=   ™ ¢ $x	$   ÿŽ â r)   ÿàõ ä �y@   ™ ¤ r	   ÿàÿ ä �yC   ™ ¤ r	'   ÿä Ä r&#   ÿàõ Æ �yB   ™ ¢  rÿE   ÿä â %v Z  > Ž ä �yD   ™ ¤  $x$   ÿŽ â r        â �yF   ™ ¢  r)   ÿàõ Æ �yG    ™ ¢  r   ÿàÿ Æ �yI   ™ ¢  r'   ÿä ä r(#   ÿàõ â/ �y    ™ ¦ r
ÿE   ÿä â �yH   ™ ¢ %v Z  > Ž â r*#   ÿà÷ Ä �yJ   ™ ¢  $x
$   ÿŽ â r)   ÿàõ ä �yK   ™ ¢  r
ÿE   ÿäÿ ä r   ÿàÿ â �yL   ™ ¢  %v Z  > Ž ä r	'   ÿä â �y    ™ ¢  $x
$   ÿŽ â r
)   ÿàõ Ä �y>   ™ ¤ r   ÿàÿ ä �y   ™ ¤" r'   ÿä ä �y   ™ ¨ �y
   ™ ¨  �yN   ™ ¨ �yO
   ™ ¨  �y    ™ ¨
 �yM
   ™ ¨  �yP
   ™ ¢  x€  ÿàõ Ä x++    ÿàÿ ä rÿ   ÿä ä | +   pbô ä xQ1ÿÀŽ ä� x1ÿ   1  ä xRÿ   8  Äx8ÿÀŽ ä x81ÿÀŽ ä xRÿÀŽ ä/ xÿ   6  â&r	Q0   ÿ   äO&r10   ÿ   ä &r087   	   âx	6ÿÀŽ â &r17   1   â x!ÿÀŽ â 2r
3        â xÿÀŽ â &r0	"   0   Ä &r6"   1   ä &r04   0   ä 0r"ÿ
  0 A   â2r723        â 0r
ÿ
    A   â Er1 0       ä  x3"   ?  A  â x!ÿ   !  Æ #r

1   3   â &r1Q   ÿ   ä&r   ÿ   â x!!ÿÀŽ â &r18    1   ä&r       â !r
       â &r1	:   1   â 2r
;        â &r6!4   6   Ä &r:      ä 0r"ÿ7  0 A   ä&r1<   1   â Er6 6       â &r<!<      â  x""   ?  A  â 0r7ÿ7    A   ä 0rÿ
  0 A   âEr1 1       â 0r
ÿ
    A   â #r776   "   â &r"Q?   ÿ   â x   ?  A  â 2r;2;        â &r??   ÿ   Ä #r

1      â &r"8=   "   ä0r ÿ;  0 A   â!r
       â &r"	@   "   â 2r
A        â&r?=   ?   â Er< <       â 0r;ÿ;    A   â  x     ?  A  â &r"C   "   â 2rA2A        â &r@@   ?   Ä 0rÿ
  0 A   â#r;;<       â &r@!C   @   â Er1 "       â 0r
ÿ
    A   â  x   ?  A  â 0r ÿA  0 A   ä&r3QB   ÿ   â Er@ @       â 0rAÿA    A   â #r
1      â  x     ?  A  â &r38D   3   Ä !r         â #r A@       â &r
	F   3   â 2rG        â!r7       â &r7B   ÿ   â !r9 9       â &r
I   
   ä 0r ÿ  0 A   ä&r7D   7   â Er
 
       â 0rÿ    A   â  x     ?  A  â &r1QH   ÿ   â 2rG2G        â &rFF   7   â #r

       â &r 8J   1   Ä &rF!I   F   â !r5
5       â 0r"ÿG  0 A   â2r
        â&r 	K       â ErF F       â 0rGÿG    A   â  x""   ?  A  â 2r2        â &r L       ä 0rÿ
  0 A   â #rGGF   "   â 0r0ÿ  0 A   Ä 0r"ÿ    A   â Er         â 0r
ÿ
    A   â  x1   ?  A  â &rQ>   ÿ   ä&r>>   ÿ   â #r

   1   â &r8      ä&r>   >   â !r-
-       â 2r
        â &r	      Ä &r3H   ÿ   ä &rQQ   ÿ   ä&r   ÿ   ä &r>   >   ä &rN      ä &rJJ   3   Ä 0rÿ
  0 A   ä&r88O   Q   ä&rO      â Er        â &rJK   J   â 2r        â 0r
ÿ
    A   â  x   ?  A  â &r8	M   8   â2r2        â&rM      â 2r2        â &rJ!L   J   â#r
      â &r>!N   >   Ä &r8P   8   ä&r!!P      ä 0rÿ  0 A   ä 0r
ÿ  0 A   ä0rÿ  0 A   â Er3 J       â 0rÿ    A   â Er	 >       â 0rÿ    A   â Er8 8       â 0rÿ    A   â Er! !       â  x70   ?  A  â  x   ?  A  â  x

   ?  A  â  x   ?  A  â x€  ÿà÷ â #r""3   7   â #r	      â #r8   
   â #r!      â rÿ   ÿäÿ â !r;       â !rG       â !r/"/       â !r..       â !r       â !r       â !r       â x    ÿàÿ Ä x$$    ÿàÿ ä x&&    ÿàÿ ä x((    ÿàÿ ä x**    ÿàÿ ä x,,    ÿàÿ â D)       À â Gy  àîÿÿÿÿƒ ê Ay        € ê Ey      € â G‰  ð     € ê $xÀ  Ž Ê ˆs  ùÿ H   è ˆs €ùÿ H   è ˆs  úÿ H   è ˆs €úÿ H   è ˆs   ûÿ H   è ˆs 9€ûÿ H   è ˆs 5 üÿ H   è ˆs €üÿ H   è ˆs - ýÿ H   è ˆs /€ýÿ H   è ˆs . þÿ H   è ˆs €þÿ H   è ˆs  ÿÿ H   è ˆs €ÿÿ H   ä Ay        € ê {          ì M        € ê „y     H   b z  r  p`ð Ä z)  r      â „y €   H   ¢ r ÿ   p"p ä r'))   ÿŽ â „y €  H   è „y    H   ( „y    H   h „y	    H   ( „y €  H   ( „y €  H   ( „y €  H   è „y    H   â !r       Æ/ „y €  H   â !r       ÆO ‰      " !r       Æ� „y €  H   b !r       Æ„y    H   â !r55       Æ„y    H   â !r		        Æ y      &   ¢ !r99       Æ ‰      â !r       â !r       ä ‰      " !r        Æ ‰      " !r"       Æ y      '   b !r$       Æ ‰5      b $z  x  Ž ÆO ‰	      ¢ x   ÿàÿ â !r//       ä/ ‰
      b !r       â� z  r  p`ò â !r--       â ‰9      â !r       â x    p"ò â !r..       â !r       â‰      " $z  ~   Ž Æ‰-      " !r5       Æ ‰/      b 
  r  ÿàõ â !r	       äO ‰ €    ¢ ÿÿ   ÿä â !r

       ä/ ‰.      b 
 j  ÿ„ â !r9       Ä� ‰      â 
 k   Æ ‰!       â !r       â r))   ÿŽ ä ‰#"      " !r--       Æ‰	      " !r//       Æ‰      b !r       âO     ÿà÷ ä ‰%$      ¢ !r.       â/    ÿàÿ ä ‰      b !r       â� r)    ÿ„ Ä ‰      â !r! !       Æ ‰
      â !r#"#       â !r	       ä‰       " 	 '   ÿàû â !r       ä‰"-      " ÿÿ   ÿäÿ â !r%$%       ÄO ‰&      ¢ 
	 j  ÿŠ â !r       ä/ ‰$/      b ÿÿ   ÿäÿ â !r       ä� ‰ @    â rÿÿ   ÿä â !r
       Ä ‰ €    b r)    ÿˆ ä 
 j  ÿ† â ‰(!      b z j  ÿ„ â !r         â r.ÿÿ   ÿä â ‰*#      " 
 k  � â !r"-"       âz
 j  ÿˆ â ‰ €    " z k   â !r&&       âO 
		 k  � â ‰ €    ¢ !r$/$       â/ z k  . Ä ‰,%      b !r       â� )   ÿàÿ ä ‰ €    â !r       â'   ÿàÿ ä ‰ €    b !r(!(       â /    ÿàù Ä ‰ €    b !r*#*       â 2ÿÿ   ÿä â !r       â‰#  €    " $x)   ÿŽ â / j  ÿˆ â !r       âO ‰'$ €    ¢ 0    ÿà÷ â !r,%,       Ä/ ‰)& €    b / k  2 â !r       ä� ‰%" €    â r    ÿàõ â !r       ä‰ @    " r.ÿÿ   ÿä â !r       Ä ‰+( €    b z j  ÿ„ ä1ÿÿ   ÿäÿ â ‰-* €    b !r# #       â 0 j  ÿ† ä ‰! @    " !r'$'       âO z k  . Ä ‰ @    ¢ !r)&)       â/ 0 k  1� ä ‰ @    b !r%"%       â�     ÿàõ ä ‰/, €    â !r       Æ‰ @    " !r+(+       Æ‰      b !r-*-       Æ ‰ @    b !r!!       Æ ‰ # @    " !r       ÆO ‰% @    ¢ !r       Æ/ ‰"' @    b !r/,/       â� ,ÿÿ   ÿä ä ‰$) @    â !r       Æ‰      " !r1       âx        ä ‰&+ @    b !r       â  j  ÿ„ â %v  j   Ž â ‰(- @    b !r #        â  k  , Ä ‰      " !r%       ÆO ‰*/ @    ¢ !r"'"       Æ/ ‰      b !r$)$       Æ� ‰       â !r3       Æ‰      " !r&+&       Æ‰!      b !r(-(       Æ ‰       b !r)       Æ ‰#      " !r*/*       ÆO ‰%"      ¢ !r+       Æ/ ‰$      b !r        Æ� †	 1    â !r       Æ† 3   â !r!!       Æ‰'&      â !r        Æ ‰(      " !r##       Æ ‰*      " !r%"%       ÆO †	 )    â !r$       Æ/ † +    è †	 
    è † 
   â !r'&'       Æ� †	 !    â !r(       Æ†     â !r*       Æ †	 #    è † %   è †	     è † '    è †	     â M™        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      "   " ¹z  l      â Ey  °    € â ™x?    â y      !   b ¹z  F   
   â x      ÿ  â r  ÿ       â y      %   ¢ r ÿ       â x      ÿ  â r$ ÿ       Ä r, ÿ       ä r( ÿ       ä r ÿ       ä r  ÿ       ä r ÿ       ä r ÿ   pRð Ä r ÿ   pRò är   ÿ(Ž È/ r   ÿ� ä x   ÿ   äO x"ÿ     È | "   pbô Ú G)  @    € ê y      '   " r ÿ       ä r ÿ       â y      &   b x
        ä z'  q      ä r ÿ       ä r'"   ÿŽ â$x!'   "Ž âr%'"   ÿŽ â$x''   "Ž â z)" q  ÿàÿ â 'z y   Ž È 'z s   Ž â/ zÿ z    Æ $z p  "Ž â zÿ t    â $z	 |  "Ž ä$z |  Ž â x   ÿàÿ â $z	 v  	Ž ä$z v  Ž ä $z p  	Ž â x	þÿÿÿÀŽ â $z }  ÿŽ Ä %v X  
Ž â r		  €ÿàÿ Æ %v X  
Ž â x#	   ÿ   ä x   ÿàù â $z w  ÿŽ â x   ÿà÷ ä xÿ   #  ä r   ÿàõ ä rÿ   ÿä Ä rÿ   ÿäÿ ä r9ÿÿ   ÿä ä x-#   ÿ   äx+#     ä r	"   ÿàõ â %x
#   Ž âx8 $       â �y&øÿÿ ™ ¢  rÿ9   ÿä â %x#   Ž Ä �y.
üÿÿ ™ ä %v	 Z  8 Ž ä �y2üÿÿ ™ d	 $x$   ÿŽ â r-   ÿàõ â �y0
    ™ ¦ r		   ÿàÿ â �y    ™ ¦ r	+   ÿä â �y/    ™ ¨  �y*   ™ ¢  r)   ÿàõ Æ �y1   ™ ¢  rÿ9   ÿä Æ �y   ™ ¢  %v Z  8 Ž Æ �y5   ™ ¢  $x$   ÿŽ â r
-   ÿàõ ä/ r   ÿà÷ â�y6øÿÿ ™ " r   ÿàÿ ä rÿ9   ÿäÿ ä r+   ÿä â %v Z  8 Ž â �y;    ™ & $x	$   ÿŽ â �y3
   ™ " r-   ÿàõ Æ �y4
   ™ " r	   ÿàÿ Æ �y7
   ™ " r+   ÿä ä r!   ÿàõ â �y=
   ™ & rÿ9   ÿä â �y<   ™ " %v Z  8 Ž Æ �y:   ™ " $x	$   ÿŽ â r-   ÿàõ ä r%   ÿà÷ â �y>   ™ " r	   ÿàÿ Æ �y    ™ " r	+   ÿä ä r
ÿ9   ÿäÿ â �yA   ™ " %v Z  8 Ž Æ �y?   ™ " $x
$   ÿŽ â r
-   ÿàõ ä �y@   ™ $ r   ÿàÿ ä �y   ™ $ r+   ÿä Ä �y    ™ " r'   ÿàõ Æ/ �yC   ™ " rDÿ9   ÿä â %v Z  8 Ž ä �yB
   ™ $ $xED$   ÿŽ ä �y8
   ™ " r-   ÿàõ Æ �y    ™ " rE   ÿàÿ Æ �y
   ™ " r	+   ÿä Æ �yE
   ™ ( �yD   ™ ( �yG   ™ ( �y    ™ ( �yF   ™ ( �y   ™ " x€  ÿàõ Ä x""    ÿàÿ ä rÿ   ÿä ä | "   pbô ä x€  ÿà÷ ä x    ÿàÿ ä rÿ   ÿäÿ Ä x!!    ÿàÿ ä x%%    ÿàÿ ä x''    ÿàÿ ä x))    ÿàÿ ä xI.ÿÀŽ ä� xKÿ   .  Ä x.ÿ   2  äxM2ÿÀŽ ä xKKÿÀŽ ä x.ÿÀŽ ä/ x	0ÿÀŽ äO xÿ   0  â &r
I*   ÿ   Ä &r*M*   ÿ   ä &r
K1   
   ä&r21   *   â x1ÿÀŽ â 2r.&/        â xÿÀŽ â &r*	   
   ä&r1   2   Ä 0rÿ.  0 A   ä0r
ÿ.    A   â x.ÿ     â &r5   *   â 2r/6/        â x   ?  A  â x*.ÿÀŽ ä Er        Æ &r*5      ä #r
      â 0r
ÿ/  0 A   ä&rI3   ÿ   â Er        â 0r/ÿ/    A   â  x

   ?  A  â &r.K4      Æ #r//   
   â &r.	7   .   â 2r&;        â&r3M3   ÿ   ä &r.=   .   ä &r44   3   ä 0r
ÿ  0 A   â!r       â &r417   4   â 2r;6;        â 0rÿ    A   â Er .       â  x
   ?  A  â &r0I<   ÿ   â !r /        â &r4*=   4   â #r      â 0r/ÿ;  0 A   Ä &r0K:   0   â Er4 4       â 0r;ÿ;    A   â !r       â  x
/   ?  A  â &r0	>   0   â 2r&        ä2r6        â #r;;4   
   â &r0A   0   Ä 0r
ÿ  0 A   ä&r/I?   ÿ   ä 0r.ÿ  0 A   ä0rÿ    A   â Er 0       â 0rÿ    A   â  x
   ?  A  â &r/K@   /   Ä &r?M?   ÿ   â #r      â &r/	   /   â 2r&        â&r?@   ?   â 2r6        â &r/C   /   ä&r?1   ?   ä 0r
ÿ  0 A   â!r((       â &r?*C   ?   â Er/ /       â 0rÿ    A   â  x
   ?  A  â 0rÿ  0 A   Ä &rIB   ÿ   â Er
 ?       â 0rÿ    A   â #r/      â  x   ?  A  â &rK8      â 2r&        ä #r
      â &r	      ä 0r
ÿ  0 A   Ä &rBMB   ÿ   â !r       â &rE      â  x
   ?  A  â &rB8   B   ä &r<M<   ÿ   ä &r
ID   ÿ   âEr        â &rDMD   ÿ   â 2r6        â 0rÿ    A   Ä &rB1   B   ä &r<:   <   ä &r
KG   
   ä&rDG   D   â #r      â &rB*E   B   â 2r&&        â&r<1>   <   â 2r6        â 0rÿ  0 A   Ä &r		F   
   ä&r11F   D   â ErB B       â 0rÿ    A   â  x   ?  A  â &r<*A   <   ä &r	   	   ä&r1*   1   Ä 0rÿ&  0 A   ä0r
ÿ  0 A   â#rB      â Er< <       â 0r&ÿ&    A   â Er	 	       â 0rÿ    A   â Er1 1       â  x..   ?  A  â  x   ?  A  â  x
   ?  A  Ä #r<   .   â #r&&	      â #r1      â !r;       â !r,,       â !r$$       â !r       â !r         â !r&       â !r       â G©  0ñÿÿÿÿƒ ê Ay        € ê Ey  ð     € â G‰  Ð     € ê $x€  Ž Ê ˆs  úÿ H   è ˆs  €úÿ H   è ˆs  ûÿ H   è ˆs €ûÿ H   è ˆs ( üÿ H   è ˆs ,€üÿ H   è ˆs  ýÿ H   è ˆs $€ýÿ H   è ˆs  þÿ H   è ˆs  €þÿ H   è ˆs  ÿÿ H   è ˆs €ÿÿ H   ä Ay        € ê {          ì M        € ê „y     H   b z  r  p`ð Ä x   ÿàÿ â „y €   H   ¢ r ÿ   p"p ä z  r  p`ò â „y    H   â z%  r      ä x    p"ò â „y €  H   ( „y	    H   h „y €  H   ( „y    H   ( „y €  H   ( „y    H   è „y    H   è „y €  H   â !r       â/ !r        ÈO ‰       " !r       Æ� ‰      b !r       Æ„y €  H   â !r(	(       Æ‰      ¢ !r,,       Æ ‰	      â !r       Æ ‰(      " !r$$       â !r         ä ‰,      " !r       â !r       ä/ ‰       b !r       Æ ‰      b !r         â !r       äO ‰      ¢ !r		       Æ� ‰$      â !r(       Æ‰      " !r       Æ ‰      " !r,       Æ ‰       " !r        Æ/ y       &   b !r       â!r       äO ‰	      â !r$$       Æ� ‰      ¢ !r       Æy      '   â !r       Æ ‰      " !r         Æ ‰      " $z   x  Ž Æ/ ‰      h ‰ €    h ‰$      b !r       ÆO ‰      ¢ $z  ~   Ž â� !r	       ä ‰       â !r       â
  r  ÿàù ä ‰!      " !r       â ÿÿ   ÿä Ä ‰#      " !r       â/ r%%   ÿŽ ä ‰ €    b !r
       â
 j  ÿˆ â !r$       â ‰ €    b     ÿàõ â !r       ÄO ‰ €    ¢ 
 k   â !r        ä� ‰ €    â ÿÿ   ÿä â !r!!       ä‰
 @    " 
 j  ÿ„ â !r##       Ä ‰ €    " 
 k   â !r       ä/ ‰ €    b r%    ÿ† ä    ÿàÿ â ‰ €    b rÿÿ   ÿäÿ â !r       âz j  ÿ† â ‰  €    b !r       âO r%%   ÿŽ Ä ‰" €    ¢ z k  � â !r       â�     ÿàù â ‰$! €    â !r
       âr%    ÿ† ä ‰&#      " !r       â ÿÿ   ÿä Ä ‰	 @    " !r       â/ r(ÿÿ   ÿäÿ â !r       â ‰ @    b z
 j  ÿ† â !r         ä‰ @    b z k  (� â !r""       ÄO ‰ @    ¢ !r$!$       Æ� ‰ @    â !r&#&       Æ‰ @    " !r	       â 	%   ÿàÿ ä ‰      " 
 j  ÿˆ â !r       ä/ ‰! @    b * 	   ÿàõ â !r       Ä‰#  @    b +ÿÿ   ÿä â !r       äO ‰%" @    ¢ * j  ÿ„ â !r       ä� ‰'$ @    â 
	 k   â !r       Ä‰)& €    " * k  + ä     ÿàõ â ‰      " !r-       â         â !r!!       â/ ‰      " (ÿÿ   ÿä ä %  j   Ž â !r# #       â‰      b !r%"%       ÆO ‰      ¢ !r'$'       Æ� ‰      â !r)&)       Æ‰      " !r+       Æ ‰!      b         ä  j  ÿ„ â ‰ #      b !r       â %  j   Ž â  k  ( â ‰"%      " !r       Æ/ ‰&) @    b !r       ÆO ‰$'      ¢ !r       Æ� †	 +    â !r       Æ† -   â !r!!       Æ†	     â !r##        â !r%%"       ä †     â !r&)&       Æ/ †	     â !r''$       ÆO †    è †	     è † !    è †	 
#    è † 
%   è ‰&      ( †	 '    â M™        € ê x   ÿàÿ â !r&       ÆŸ r    ÿàñ È r ÿÿ   ÿä  ä z j  ÿ€ È z k     Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      "   " ¹z  l      â Ey  `    € â ™x?    â y      !   b ¹z  F   
   â r ÿ       â x&      ÿ  â y      %   ¢ r ÿ       â x      ÿ  â r* ÿ       Ä r ÿ       ä r( ÿ       ä r  ÿ       ä r ÿ   pRð är ÿ   pRò är   ÿ(Ž È/ r   ÿ� ä x   ÿ   äO xÿ     È |    pbô Ú G)      € ê y      '   " r ÿ       â $z	 p  Ž âr ÿ       â y      &   b x        ä z"  q      ä r ÿ       ä r"   ÿŽ â$x "   Ž âr""   ÿŽ Ä z$ q  ÿàÿ â 'z y   Ž È 'z s   Ž â/ zÿ z    Æ $z }  ÿŽ â zÿ t    â $z |  Ž ä$z	 |  	Ž â x   ÿàÿ â $z v  Ž ä$z v  	Ž â x	þÿÿÿÀŽ â $z p  Ž Ä %v X  Ž â r		  €ÿàÿ Æ $z w  ÿŽ â x!	   ÿ   â %v X  Ž â x   ÿàù ä r   ÿàõ ä x   ÿà÷ ä xÿ   !  ä rÿ   ÿä Ä rÿ   ÿäÿ ä r5ÿÿ   ÿä ä x%!   ÿ   äx#!     ä r   ÿàõ â %x!   Ž âx4 $       â �y)øÿÿ ™ ¢  rÿ5   ÿä â %x!   Ž Ä �y-üÿÿ ™ ä %v
 Z  4 Ž ä �y.üÿÿ ™ d	 $x$   ÿŽ â r%
   ÿàõ â �y    ™ ¦ r   ÿàÿ â �y2    ™ ¦ r#   ÿä â �y,
    ™ ¢  r$   ÿàõ Æ �y   ™ ¢ rÿ5   ÿä â %v Z  4 Ž ä�y+   ™ ¤ $x$   ÿŽ â r   ÿà÷ â �y   ™ ¦  r   ÿàÿ â �y/   ™ ¢  rÿ5   ÿäÿ â %v
 Z  4 Ž â r%   ÿàõ â/ �y1øÿÿ ™ ¤ $x$   ÿŽ â r	#   ÿä Ä r        ä r%
   ÿàõ â �y0   ™ ¢ r   ÿàÿ Æ �y3   ™ ¢ r#   ÿä ä r    ÿàõ â�y   ™ & r8ÿ5   ÿä â %v Z  4 Ž â �y    ™ & $x98$   ÿŽ â �y6   ™ ( r9   ÿàÿ â �y7   ™ ( �y   ™ " r%   ÿàõ Æ/ �y8
    ™ " r	#   ÿä ä r="   ÿàõ â �y   ™ & r<ÿ5   ÿä â �y:   ™ " %v
= Z  4 Ž Æ/ �y9   ™ " $x=<$   ÿŽ Æ �y;   ™ " r%
   ÿàõ ä r=   ÿàÿ â �y    ™ & r#   ÿä â �y4   ™ ( �y<   ™ ( �y>   ™ ( �y=   ™ ( �y

    ™ ( �y?   ™ ( �y   ™ " x€  ÿàõ Ä x    ÿàÿ ä rÿ   ÿä ä x€  ÿàõ È rÿ   ÿä ä |    pbô ä x    ÿàÿ ä x      ÿàÿ ä x""    ÿàÿ ä x$$    ÿàÿ Ä x@-ÿÀŽ ä� x-ÿ   -  ä xBÿ   .  äx-ÿÀŽ ä/ x..ÿÀŽ ä x	ÿÀŽ ÄO xÿ     ä xBBÿÀŽ â &r-@   ÿ   â2r),        â &r.   ÿ   ä &r-+   -   âxÿÀŽ â &r+B+      ä &rD	   -   Ä 0rÿ  0 A   âxÿ   2  â &rD/   D   ä 0rÿ    A   â x2ÿÀŽ â  x2   ?  A  â ErD D       â 2r,1,        â xÿÀŽ â &r   +   Ä &r+@0   ÿ   ä&r0.0   ÿ   â #rD   2   â &r/      ä 0rÿ,  0 A   ä&r+3   +   ä&r3B3   0   â !r         â 0rÿ,    A   â Er        â  x,   ?  A  â &r+	   +   Ä&r3   3   â 2r)        â#r   ,   â 2r1        â &r+6   +   ä 0rÿ  0 A   ä&r/@7   ÿ   â Er+ +       â 0rÿ    A   â  x   ?  A  â &r   /   â !r((       â 2r)8        â #r+      â &r	      Ä &r7.7   ÿ   â !r       â &r:      ä 0rÿ  0 A   ä&r7B   7   ä &r@9   ÿ   â Er        â 0rÿ    A   â  x   ?  A  â &r;      Æ #r      â 2r)        â &r9.9   ÿ   ä &r	4      â !r       â &r@@<   ÿ   ä&r..<   ÿ   ä &r>      Ä 0rÿ  0 A   ä&r9B;   9   ä &r@=   @   ä&r.B=   .   â Er        â 0rÿ    A   â 2r))
        â &r7   7   â2r818        â&r94   9   â2r1        â x   ?  A  â &r@	?   @   â2r
1
        â &r.?   .   â #r      â &r36   3   Ä &r7:   7   ä&r9>   9   ä &r@   @   ä&r.   .   ä 0r,ÿ  0 A   ä 0rÿ8  0 A   Ä 0rÿ  0 A   ä 0rÿ)  0 A   ä0rÿ
  0 A   â Er3 3       â 0rÿ    A   â Er7 7       â 0r8ÿ8    A   â Er9 9       â 0rÿ    A   â Er@ @       â 0r)ÿ)    A   â Er	 .       â 0r
ÿ
    A   â  x-,   ?  A  â  x   ?  A  â  x   ?  A  â  x   ?  A  â  x   ?  A  â #r-3   -   â #r87      â #r9      â #r))@      â #r

	      â !r-       â !r**       â !r''       â !r&&       â !r)       â !r
       â G©  Póÿÿÿÿƒ ê Ay        € ê Ey  Ð     € â G‰  °     € ê $x@  Ž Ê ˆs   ûÿ H   è ˆs (€ûÿ H   è ˆs  üÿ H   è ˆs €üÿ H   è ˆs  ýÿ H   è ˆs *€ýÿ H   è ˆs ' þÿ H   è ˆs &€þÿ H   è ˆs  ÿÿ H   è ˆs €ÿÿ H   ä Ay        € ê {          ì M        € ê „y     H   " z  r  p`ð Ä z%  r      â „y €   H   b r ÿ   p"p ä r!%%   ÿŽ â „y    H   ¢ x#        Æ „y €  H   è „y	    H   ( „y €  H   h „y    H   h „y €  H   h „y    H   è „y €  H   â !r        Æ y      &   " !r(       Æ/ ‰       b !r       ÆO ‰      ¢ !r       â� !r		       È‰      â !r*       Æ‰	      " !r''       Æ ‰
      b !r&       â !r         ä/ ‰      b !r       Æ ‰       " !r       âO !r       ä ‰'      ¢ !r       Æ� ‰      â !r	       Æ‰      " !r

       Æ‰      b !r       Æ/ ‰      b !r        Æ ‰      " !r'       ÆO ‰  €    ¢ !r       Æ� ‰	      â !r       Æ‰      " !r       Æ‰
      h ‰      b !r       Æ/ ‰      b !r       Æ ‰      " !r        âO $z  x  Ž ä ‰      ¢ !r		       â� !r       ä‰ €    â !r
       Æ‰ €    " x
   ÿàÿ â !r       ä ‰ @    b z 
 r  p`ò â !r       ä/ ‰	 €    b x    p"ò â !r       Ä ‰ €    " !r       ÆO ‰ €    ¨ ‰ €    ¢ !r       Æ� ‰ €    â !r       Æ‰ €    " !r
       Æ‰ €    b !r	       Æ/ y      '   b !r       â !r       äO ‰ @    " !r       Æ ‰ @    ¢ !r       Æ� ‰
      â !r       Æ‰ @    " !r       Æ‰ @    b $z  ~   Ž Æ/ ‰ @    h ‰ @    b  !   ÿàõ â !r       â r%    ÿ† â‰ @    "  ÿÿ   ÿä â !r       âO 
 j  ÿ„ â ‰ @    ¢ rÿÿ   ÿäÿ â !r'
       â� z j  ÿ† â ‰ @    â !r       â%   ÿàÿ â !r       â!   ÿàÿ â ‰      " 
 k    â !r       â/ z k  � â ‰      b !    ÿàõ â !r       â     ÿà÷ â‰      b 
	  r  ÿàù â !r       â r%    ÿŠ â ‰      " %v  j  # Ž â !r       âO r"ÿÿ   ÿäÿ â ‰      ¢ z j  ÿŠ â!r       â� $ÿÿ   ÿä â ‰      â z k  "� Ä 
	 j  ÿˆ â ‰       â "ÿÿ   ÿä â !r#       â! j  ÿ„ â ‰      "  ÿÿ   ÿäÿ â !r       â/ 
		 k  $ â ‰      b 
 j  ÿ† â !r       â! k  " â !r       â  k   � â †	 '    â !r       ÆO † #   â !r       Æ� †	     â !r        Æ †     â !r       Æ†	     â !r       Æ/ †    è †	     è † 
    è †	     â M™        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      "   " ¹z  l      â Ey  ð    € â ™x?    â y      !   b ¹z  F   
   â x      ÿ  â r ÿ       â y      %   ¢ r ÿ       â x      ÿ  â r  ÿ       Ä r ÿ       ä r ÿ   pRð är ÿ   pRò är   ÿ(Ž È/ r   ÿ� ä x   ÿ   äO x'ÿ     È | '   pbô Ú G)  À    € ê y      '   " r ÿ       â $z	 p  'Ž â r ÿ       â y      &   b x        ä r ÿ       ä z!' q  ÿàÿ â 'z y   Ž È 'z s   Ž â/ zÿ z    Æ $z }  ÿŽ â zÿ t    â $z |  'Ž ä$z	 |  	Ž â x   ÿàÿ â $z v  Ž ä$z v  	Ž â x	þÿÿÿÀŽ â $z p  Ž Ä %v X  Ž â r		  €ÿàÿ Æ %v X  Ž â x	   ÿ   ä x   ÿàù â $z w  ÿŽ â x   ÿà÷ ä z  q      ä r   ÿàõ ä r'   ÿŽ â$x   'Ž â xÿ     Ä rÿ   ÿä ä rÿ   ÿäÿ ä r,ÿÿ   ÿä ä x%   ÿ   äx#     ä r'   ÿàõ â %x   Ž âx+ $       â �y-øÿÿ ™ ¢  rÿ,   ÿä â %x   Ž Ä �y"üÿÿ ™ ä %v Z  + Ž ä �y üÿÿ ™ d	 $x$   ÿŽ â r%   ÿàõ â �y(    ™ ¦ r   ÿàÿ â �y&    ™ ¦ r#   ÿä â �y    ™ ¢  r
!   ÿàõ Æ �y$   ™ ¢  rÿ,   ÿä â %v

 Z  + Ž ä �y*   ™ ¤  $x	$   ÿŽ â r%
   ÿàõ â �y)   ™ ¦  r	   ÿàÿ â �yøÿÿ ™ ¦" r	#   ÿä â �y   ™ ¢  r   ÿàõ Æ�y.   ™ ¢ r/ÿ,   ÿä â %v Z  + Ž ä �y0
    ™ ¤ $x1/$   ÿŽ â r%   ÿàõ â �y   ™ ¦ r1   ÿàÿ â �y/   ™ ¦ r#   ÿä â �y2   ™ ¢ r4   ÿàõ Æ �y1   ™ ¢  r5ÿ,   ÿä â %v
4 Z  + Ž ä/ �y3   ™ ¤  $x55$   ÿŽ â r%
   ÿàõ â�y+   ™ & r5   ÿàÿ â �y    ™ & r	#   ÿä â �y6   ™ ( �y4   ™ ( �y5   ™ ( �y

    ™ ( �y7   ™ ( �y   ™ "% x   ÿàõ Ä x''@   ÿàÿ ä rÿ   ÿä ä x   ÿàõ È rÿ   ÿä ä | '   pbô ä x@   ÿàÿ ä x@   ÿàÿ ä x!!@   ÿàÿ ä x9"ÿÀŽ Ä� x"ÿ   "  ä x8ÿ      äx ÿÀŽ ä x "ÿÀŽ ä x8ÿÀŽ ä x	ÿ   &  ÄO x"(ÿÀŽ ä xÿ   (  â &r;9$   ÿ   ä&r=$   ÿ   â x$&ÿÀŽ â &r; *   ;   ä&r=*   =   â x		ÿÀŽ â &r;$)   ;   âxÿÀŽ â 2r-        â&r=")   =   â 2r        Æ 0r&ÿ  0 A   ä &r;	   ;   ä&r=   =   ä 0r(ÿ  0 A   â Er; ;       â 0rÿ    A   â  x)&   ?  A  â 0rÿ    A   â Er= =       â  x((   ?  A  â &r9.   ÿ   â #r);   )   â 2r-0        â #r=   (   â &r       Æ !r       â 0rÿ  0 A   ä &r$/      â !r)       â &r..   ÿ   â  x)   ?  A  â &r	2      â 2r00        â &r91   ÿ   ä &r.   .   â Er        â &r 3      Ä 0rÿ0  0 A   ä &r11   ÿ   ä 0rÿ    A   ä &r$+      ä&r994   ÿ   ä&r44   ÿ   Ä &r."/   .   â  x/   ?  A  â &r13   1   ä &r9 5   9   ä&r45   4   â 2r-        â&r	6      â 2r--
        â #r   )   â &r1"+   1   â 2r        â&r$$7   9   â2r

        â &r4"7   4   â !r       â &r.2   .   âEr+        â &r16   1   Ä &r$	   $   ä&r4   4   ä 0rÿ  0 A   ä 0rÿ  0 A   ä 0rÿ-  0 A   ä0r	ÿ
  0 A   â Er) .       â 0r0ÿ0    A   â Er1 1       â 0rÿ    A   â Er$ $       â 0rÿ    A   â Er 4       â 0r-ÿ-    A   â  x   ?  A  â 0r
ÿ
    A   â  x   ?  A  â  x   ?  A  â  x		   ?  A  â #r)0)   /   â #r+      â #r1      â #r-$      â #r	
   	   â !r )        â !r       â !r       â !r       â !r	       â G©  �õÿÿÿÿƒ ê Ay        € ê Ey  °     € â G‰  �     € ê r   ÿ@Ž Ê ˆs  üÿ H   è ˆs €üÿ H   è ˆs  ýÿ H   è ˆs  €ýÿ H   è ˆs  þÿ H   è ˆs €þÿ H   è ˆs  ÿÿ H   è ˆs €ÿÿ H   ä Ay        € ê {          ì M        € ê „y €   H   b z  r  p`ð Æ „y €  H   ¢ r ÿ   p"p Æ „y     H   è „y €  H   ( „y    H   h „y €	  H   ( „y    H   ( „y €  H   ( „y    H   ( „y
    H   ( „y €  H   " !r       Æ/ „y    H   b !r        ÆO „y    H   ¢ !r       Æ� „y	 €  H   â !r       Æ„y    H   " !r       Æ„y €  H   " !r       Æ „y!    H   b !r       Æ „y  	  H   b !r       Æ „y  
  H   b !r       Æ „y €
  H   b !r

       â z  r      ä „y#    H   b !r       Æ „y €  H   â !r       Æ/ „y €  H   â !r       ÆO „y €  H   â !r	       Æ� ‰      b !r

       â !r       ä‰      " !r!       Æ‰	      ¢ !r       Æ y      '   â !r

       Æ ‰      " !r       Æ ‰
      b !r#       Æ ‰      â !r       â/ !r       ä ‰      b !r       â !r       ä ‰       " !r		       âO !r       Ä ‰      ¢ !r       Æ‰      " !r
       Æ‰      b !r       Æ� ‰	      â !r       Æ/ ‰      b !r         Æ ‰
      " !r       ÆO ‰      ¨ ‰  €    ¢ !r       â!r       ä‰ €    " !r	       Æ� ‰ €    â !r       Æ/ ‰	 €    b !r
       Æ ‰ €    " !r       ÆO ‰ €    ¢ !r        Æ ‰ €    h ‰      b !r       Æy      &   " !r       Æ� ‰ @    â !r
	       â/ !r       ä ‰ @    " !r       ÆO ‰	 @    b !r       Æ‰ @    ¢ !r       Æ ‰ @    b $z  x  Ž âx   ÿàÿ ä ‰
 @    " $z  ~   Ž â !r       ä� ‰ @    â z  r  p`ò ä 
  r  ÿàõ â ‰ €    â !r       â x    p"ò â !r	       â/ ÿÿ   ÿä â ‰      " 
 j  ÿ„ â !r       ÄO ‰      â 
 k   â !r       ä‰      b r    ÿ† â!r
       ä‰      ¢ 	   ÿàÿ â !r       Ä� ‰      â rÿÿ   ÿäÿ â !r       ä ‰      " z j  ÿ† ä r   ÿŽ â ‰ @    b  	   ÿàõ â !r       â z k  � â ‰      "     ÿà÷ Ä x	        â !r       â/ ÿÿ   ÿä ä 
 j  ÿ„ â %v  j  	 Ž â ÿÿ   ÿäÿ â !r       âO 
 j  ÿ† â!r       â�  k   â !r       â 
 k  � â !r       â†	     â !r       Æ†    â !r       Æ †	     è † 
    è †	     è †    è ‰      ( †	     â M™        € ê x   ÿàÿ â/ !r       Æ r    ÿàñ È r ÿÿ   ÿä  ä z j  ÿ€ È z k     Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      "   " ¹z  l      â Ey  °    € â ™x?    â y      !   b ¹z  F   
   â r ÿ       â x      ÿ  â y      &   ¢ r ÿ       ä r  ÿ       â y      '   â r ÿ       Ä r ÿ   pRð är ÿ   pRò är
   ÿ(Ž È/ r

   ÿ� È xÿ     â $À   Ž Æ |    pbô Ú G)  €
    € ê y	      %   "Î r ÿ       â $z }  ÿŽ âr ÿ       ä x        â 'z y   Ž â r ÿ       ä z q  ÿàÿ â 'z s   Ž â zÿ z    È zÿ t    â $z |  Ž È $z v  Ž â x		   ÿ   Ê $z	 p  Ž È $z |  Ž â x	   ÿàÿ Æ $z v  Ž ä $z p  Ž â xþÿÿÿÀŽ â %v X  Ž Æ r
  €ÿàÿ â $z w  ÿŽ â x   ÿàù â %v X  Ž â x   ÿ   ä r   ÿàõ ä x   ÿà÷ ä z  q      Ä x
ÿ     ä r	ÿ	   ÿä ä rÿ   ÿäÿ ä r   ÿŽ ä r)ÿÿ   ÿä ä x#   ÿ   Ä x!   
  ä r   ÿàõ â %x   Ž âx- $       â �y*øÿÿ ™ ¢  r
ÿ)   ÿä â %x   Ž ä �y üÿÿ ™ ä %v Z  - Ž ä �y"üÿÿ ™ d	 $x
$   ÿŽ â r
#   ÿàõ â �y$    ™ ¦ r   ÿàÿ â �y&    ™ ¦ r!   ÿä â �y+    ™ ¢  r   ÿàõ Æ �y%
   ™ ¢  r,ÿ)   ÿä Æ �y'
   ™ ¢  %v Z  - Ž Æ �y(
   ™ ¢  $x,$   ÿŽ â/ r   ÿà÷ ä �y,
   ™ ¢ r#   ÿàõ är   ÿàÿ â �y.øÿÿ ™ ¢ r/ÿ)   ÿäÿ ä r!   ÿä â %v Z  - Ž â �y    ™ ¦ $x//$   ÿŽ â �y-   ™ ¢ r#   ÿàõ Æ r/   ÿàÿ â �y
   ™ ¦  r!   ÿä â �y   ™ ¨  �y/   ™ ¨ �y1   ™ ¨  �y2   ™ ¨ �y    ™ ¨  �y0   ™ ¨ �y3   ™ ¢ x   ÿàõ Äx@   ÿàÿ ä rÿ   ÿä ä x   ÿàõ È r	ÿ	   ÿä ä |    pbô ä x@   ÿàÿ ä x@   ÿàÿ ä x ÿÀŽ ä� x ÿ      Ä x5ÿ   "  äx""ÿÀŽ ä x ÿÀŽ ä x$ÿÀŽ äO x55ÿÀŽ ä x$ÿ   $  â &r%   ÿ   Ä/ &r "%   ÿ   ä &r'      âx$ÿÀŽ â &r%5'       â 2r *+        â &r'(      â x&ÿÀŽ ä x$ÿ   &  â &r',   '   Ä 0rÿ   0 A   âx$$ÿÀŽ â &r%(   %   â 2r+.+        â 0r ÿ     A   â Er' '       â  x   ?  A  â &r,$,   %   ä 0r%ÿ+  0 A   â#r' '      â &r -   ÿ   âEr, ,       â 0r+ÿ+    A   â  x%   ?  A  â &r-"-   ÿ   Ä &r 
       â #r++,      â &r5
   -   â 2r
*        â &r        ä&r      ä &r%/   ÿ   ä&r""/   ÿ   Ä &r 1       ä 0rÿ
  0 A   ä&r2   %   â Er%         â &r"52   "   â  x   ?  A  â 0r
ÿ
    A   â 2r.        ä 2r**        â&r0      â2r.        â &r"0   "   â #r
%      â 0r
ÿ  0 A   Ä 0rÿ*  0 A   ä&r1$1      ä &r3      ä&r"$3   "   ä 0rÿ  0 A   â  x
   ?  A  â  x   ?  A  â 0rÿ    A   â Er
 1       â 0r*ÿ*    A   â Er        â 0rÿ    A   â Er" "       â  x   ?  A  â #r

      â #r*      Ä #r"      â !r'       â !r +        â !r       â !r
       â !r       â !r       â G©  À÷ÿÿÿÿƒ ê Ay        € êÏ ˆ  ýÿ H   è ˆ  €ýÿ H   è ˆ  þÿ H   è ˆ €þÿ H   è ˆ  ÿÿ H   è ˆ €ÿÿ H   è {          ì M        € ê „y     H   ( „y €   H   h „y    H   ¨ „y    H   è „y €  H   ( „y    H   h „y €  H   h „y	    H   h „y    H   h „y €  H   h „y €  H   b !r       Æ „y
    H   " !r        Æ/ „y €  H   b !r       ÆO „y    H   ¢ !r       Æ� „y €  H   â !r       Æ„y! €  H   â !r       Æ„y    H   â !r       Æ „y# €  H   â !r	       â !r       È ‰      " !r       â !r       â !r

       â !r       â/ !r       ÈO ‰	      " !r       Æ� ‰      b !r       â!r!       ä ‰      ¢ !r

       Æ ‰       â !r#       Æ ‰
      ( ‰      b !r		       â !r       È/ ‰	      " !r       ÆO ‰      b !r         Æ� ‰      ¢ !r
       Æ‰  €    â !r       Æ‰      ( ‰
      b !r	       Æ y      %   " !r       Æ/ ‰	 €    b !r       ÆO ‰ €    ¢ !r        Æ� ‰ €    â !r       Æ‰  @    " !r

       Æ‰ €    h ‰
 €    b x   ÿ   â !r		       Æ/ z  r  p`ð â $z x  Ž â !r       âO ‰	 @    " r ÿ   p"p â $z ~  Ž â !r       â� ‰ @    b !r         Æ‰ @    ¢ !r       Æ‰       ä 
 r  ÿàõ â !r
       â ‰ @    " x
   ÿàÿ Æ ‰ @    b z 
 r  p`ò â !r		       Æ x    p"ò â !r       â/ ‰
	      " !r       ÆO ‰      b ÿÿ   ÿä ä 
 j  ÿ„ â !r        â� !r       âz  r      ä 
 k   â !r       â‰      ¢    ÿàÿ Ä r   ÿ† â ‰      â    ÿàõ ä x        â ‰      " ÿÿ   ÿä â !r	
       â r ÿÿ   ÿäÿ â %v j   Ž â  j  ÿ„ â !r       â/ z j  ÿ† Ä 	 k   â †	     â z k   � Æ †    â !r       ÆO †	     â !r       â� !r       È†     è †	     â M™        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      "   " ¹z  l      â Ey  @	    € â ™x?    â y       !   b ¹z  F   
   â r ÿ       â x      ÿ  â y      %   ¢ r! ÿ       ä r ÿ   pRð Ä r ÿ   pRò är    ÿ(Ž È/ r   ÿ� ä x   ÿ   äO xÿ     ä     ÿ8Ž ä |    pbô Ú G)  0    € ê y      '   " r ÿ       ä r ÿ       â y      &   b x
        ä r ÿ       ä z q  ÿàÿ â 'z y   Ž È 'z s   Ž â/ zÿ z    Æ $z p  Ž â zÿ t    â $z	 |  Ž ä$z |  Ž â x   ÿàÿ â $z	 v  	Ž ä$z v  Ž ä $z p  	Ž â x	þÿÿÿÀŽ â $z }  ÿŽ Ä %v X  
Ž â r		  €ÿàÿ Æ %v X  
Ž â x	   ÿ   ä x   ÿà÷ â $z
 w  ÿŽ â x   ÿàõ ä xÿ     ä r
   ÿàù ä rÿ   ÿäÿ Ä rÿ   ÿä ä r&ÿÿ   ÿä ä x   ÿ   äx     ä r
   ÿàõ â %x   Ž â x $       â �y%øÿÿ ™ ¢  rÿ&   ÿä Ä r   ÿà÷ â %v

 Z   Ž â �yüÿÿ ™ ä rÿ&   ÿäÿ â $x	$   ÿŽ â r
   ÿàõ â %x   Ž â �y    ™ $ r	   ÿàÿ â %v Z   Ž â �yüÿÿ ™ ¤
 r	   ÿä â $x$   ÿŽ â r   ÿàõ â �y

    ™ & r   ÿàÿ â �y    ™ & r   ÿä â �y#   ™ ( �y$   ™ ( �y   ™ (# �y"    ™ ( �y   ™ ( �y'   ™ ( �y(   ™ ( �yøÿÿ ™ (�y    ™ (	 �y   ™ " x   ÿàõ Äx@   ÿàÿ ä rÿ   ÿä ä x   ÿàõ È rÿ   ÿä ä |    pbô ä x@   ÿàÿ ä xÿÀŽ ä� xÿ     ä xÿ     ÄO x)ÿÀŽ ä xÿÀŽ ä xÿÀŽ â&r    ÿ   â/ xÿ     Æ &r#      â xÿÀŽ â &r )    ÿ   ä &r$      â xÿÀŽ â 2r%
        â &r   ÿ   ä&r))   ÿ   â x	"ÿÀŽ Ä xÿ   "  â 0rÿ  0 A   ä &r      ä &r'      ä&r)'   )   âxÿÀŽ â &r #       â Er        â &r(      Ä &r)	(   )   â2r

        â 0rÿ    A   â  x#   ?  A  â &r$	$       â 2r%%        ä2r        â &r      ä&r)   )   â #r   #   â 0rÿ
  0 A   Ä &r$   $   ä 0rÿ%  0 A   ä 0rÿ  0 A   â x	   ?  A  â 0rÿ    A   â Er $       â 0r
ÿ
    A   â Er        â 0r%ÿ%    A   â Er) )       â  x   ?  A  â  x   ?  A  â #r

   	   Ä #r%%      â #r)      â !r!!       â !r
       â !r%       â !r       â G©  ðùÿÿÿÿƒ ê Ay        € ê ˆ ! þÿ H   è ˆ €þÿ H   è ˆ  ÿÿ H   è ˆ €ÿÿ H   è {          ì M        € ê „y      H   " z  r  p`ð Æ „y  €   H   b r  ÿ   p"p Æ „y     H   ¨ „y     H   è „y  €  H   ( „y     H   h „y     H   h „y	  €  H   h „y     H   h „y
  €  H   h „y  €  H   b !r!       Æ „y  €  H   " !r       â/ !r       äO y      &   b !r       â� !r       â!r       ä!r       â !r		       È ‰      ¢ !r       Æ ‰	      â !r

       Æ ‰      " !r

       È !r
       â !r       ÈO ‰      " !r	       Æ� ‰      ¢ !r       Æ‰      è ‰      " !r       â !r       ÈO ‰      " !r       Æ� ‰ €    ¢ !r       Æ‰ €    è ‰
 €    " !r       â x        â !r       ÆO ‰ €    " !r       Æ� y      '   ¢ !r

       Æ‰ @    è ‰ @    ( ‰	
 @    b !r       â !r       â� x   ÿàÿ â $z x  Ž â/ !r       Äz  r  p`ò â ‰      " $z ~  Ž âO !r	
	       âx     p"ò â ‰      d 
  r  ÿàõ ä ‰
 @    ¤ ÿÿ   ÿä Ä ‰	      â 
  j  ÿ„ È 
  k   â !r       â %v j   Ž â !r       È/ †	     â !r

       ÆO †    â !r		       Æ� ‰
      h  †	 	    â M™        € ê ¹z  r      â !r
       â/ �x   ?àÿ Ì |   ÿàñ È r ÿÿ   ÿä  ä z j  ÿ€ È z k     Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $	ÿ   ÿ Ž È % \  	 Ž Ì �	    ™ b r ÿ       â ¹z  l      â Ey  P
    € â y       "   b ™x?    â '
 m   Ž äy      !   d 'Š s   Ž â 
ÿ n    â y      '   ¨ $ÿÿ   
Ž È $
 o  Ž äy      %   â $‚ÿÿ    Ž È $z w  ÿŽ â r    ÿ(Ž È/ r	   ÿ� â $z }  ÿŽ ÆO xÿ   	  ä r   ÿàõ ä |    pbò Æ $rÿÿ   ÿ â Šÿ t    ä r ÿ       Î G  à    € ê rÿ   ÿ3Ž â� Ey @    € â r ÿ       ä |   ÿàÿ ä x		þÿÿÿÀŽ â 'z y   Ž â xÿ@   ÿÀ€ äxÿÀÿÿÿÿÀ‚ â $x   	
Ž â r ÿ       â $z	 p  ÿŽ â zÿ z    â $x   ÿ Ž Ì G	  p    € ê $z
 v  	Ž âr   ÿàñ â $tÿ$   ÿ Ž â x        â $z |  
Ž â r
ÿ   ÿä  â %v Z   Ž È $x   Ž ä $x
$   ÿŽ ä %v
 X  Ž Æ r   ÿàÿ Æ %x   
Ž ä�y

    ™ ¤ %x   Ž ä �y   ™ è �y   ™ ( �y   ™ ¨ �y   ™ ¨ �y   ™ ¨ �y    ™ ¨ �y   ™ ¢ x@   ÿàÿ Ä xÿÀŽ ä� xÿ     Æ &r   ÿ   âxÿÀŽ ä xÿÀŽ äO xÿ     â &r      Æ xÿÀŽ â &r      â 2r

        Æ &r      ä 0rÿ
  0 A   ä0r
ÿ
    A   â Er        ä  x   ?  A  È #r
      È !rÿ       ä Ay       € ê G™  p    € ê $z |  Ž â x
ÿ     â $tÿ   ÿ Ž ä $z v  Ž âx
   
  â $tÿ   ÿ Ž Æ r	   ÿàÿ Ê %v X  Ž Ê r        â $rÿÿ   	 Ž â x	   ÿ   â%v  € Ž È r    ÿàñ â$tÿ$   ÿ Ž â x@   ÿàÿ ä rÿ   ÿä  â %v   Z   Ž â r   ÿàñ ä r$   ÿàó â $x$   ÿŽ Ä $rÿÿ     â r%   ÿäÿ  ä r"	    ÿàñ ä r!!   ÿàÿ â %v Z   Ž â �y$   ™ ¦  $x$   ÿŽ â r#!
   ÿä  â �y$    ™ â  r	   ÿàñ â $x   Ž Ä �y"   ™ & r
   ÿä  â �y"   ™ h �y$   ™ h �y$  ™ h �y   ™ h �y     ™ b r$        â $r%ÿÿ    Ž Ä �y"   ™ h �y    ™ h �y    ™ h# �y!$    ™ h �y""   ™ h �y   ™ h �y$$   ™ h' �y#   ™ b x€   ÿàÿ È |    pbð ä x 
  ÿàó È rÿ   ÿäÿ  ä x%ÿÀŽ ä� xÿ     Æ &r%%   ÿ   âxÿÀŽ ä xÿÀŽ äO xÿ     â &r   %   âxÿÀŽ ä xÿ     ä xÿÀŽ â &r   ÿ   â xÿÀŽ Ä xÿÀŽ ä xÿ     â &r      Æ xÿÀŽ â &r       â 2r!        â &r"      â 2r$        â &r##      ä 0rÿ  0 A   âEr        â 0rÿ    A   ä &r   #   â  x   ?  A  â 0rÿ  0 A   Ä 0rÿ    A   â #r      â Er        â  x   ?  A  ä !r       ä #r      È !r       â G‰  Pûÿÿÿÿƒ ê Ay        € ê� r  ÿ   pRð är  ÿ   pRò Ö ˆ €ÿÿ H   è {          ì M        € ê „y      H   " r ÿ   pRð â $z x  Ž Ä „y €   H   h „y	    H   ¢ !r         È !r         È/ !r  	       ÊO ‰       $ !r        Ê ‰      $ !r       Ê ‰ €    $ !r       Ê ‰ @    $ !r       Ê ‰	      b  M	        € ê x        â $z ~  Ž â !r		       Æ/ %v j   Ž Ê †y 	    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú         â $ÿÿ    Ž È % \   Ž Ì �	    ™ b $ÿÿ   ÿ Ž â Ey       € â x
      ÿ  â y      "   b '
  m   Ž Æ y	      !   d 
 ÿ n     È   ÿ   ÿáÿ Ê $
  o  Ž ä $‚ÿÿ    Ž â ‚ ÿ       Æ $‚ÿÿ    Ž È 'Š  s   Ž â rÿ	   ÿü‚ ä/ r	   ÿ(Ž È r   ÿ� ä Šÿ t     Ê G      € ê z ÿ `  pPô ä z ÿ ^  pPò ä z ÿ b  p ö ä z ÿ a   Sô ä z ÿ _  Sò ä z ÿ c  0't Ö        '   "        ä¢       â ©      '   h 
      %   ¢ $ª x  ÿŽ Æ ©      %   â $   ~  ÿŽ ä $ª ~  ÿŽ ä/ $ x  ÿŽ â   	   
àë äO
ÿ   
  ä ¢	   àÇ Ä� ¨ÿ     ä ÿÿ   
Äþ ä ¢
   ÿàû ä ¢ÿÿ   „þ ä      ÿàý ä ¢ÿ   ÿäÿ Ä ª
 b  ÿˆ â $ÿÿ    â   ^  ÿ† ä ª
 c   ä x
      ÿ  â   _  � Ê �     h �©
     d Ay        € ê ¹z  l      â x ÿ     â ™x?    â Ey  à    € â x      ÿ  â ‚        Æ |     pbò Ú G  �    € ê y      '   " $rÿÿ   ÿ Ž â x        â $z w  ÿŽ â y      %   b xþÿÿÿÀŽ ä z ÿ `  pPð ä r  €ÿàÿ ä z ÿ a   Sð Æ $x   ÿ Ž ä 'z y   Ž â r ÿ       Æ $z }  ÿŽ â zÿ z    ä r ÿ       ä r   ÿàó â $z |   Ž Æ rÿÿ   ÿäÿ  â $z v  Ž È$z p  Ž È/ %v `  Ž È %v X  Ž È $rÿÿ    Ž ä r    ÿàó ä x $       ä r        â $rÿÿ   Ž  ä %v Z   Ž È $x$   ÿŽ â r        Æ $x   Ž ä %x   Ž â�y    ™ ¦ %x   Ž â �y   ™ è �y   ™ ( �y   ™ h �y   ™ ¨  �y    ™ ¨ �y   ™ ¨  �y   ™ ¢  x  @   ÿàÿ Ä r        ä r        ä x   ÿàó ä |     pbö Æ $rÿÿ   Ž  â xÿÀŽ ä� xÿ     Æ &r   ÿ   âxÿÀŽ ä xÿÀŽ äxÿ     â &r      ÆO xÿÀŽ â 2r        â &r      È 0rÿ  0 A   ä&r      ä 0rÿ    A   â  x   ?  A  ä Er        ä x   ÿàõ Æ #r      â G‰  P    € ê %x   Ž ä �y     ¨ �y    è �y    " xÿÀŽ Ä� xÿ     Æ &r   ÿ   â xÿÀŽ ä xÿÀŽ äxÿ     â &r      â 2r        äO xÿÀŽ â &r      Ä 0rÿ  0 A   ä&r      ä 0rÿ    A   â  x   ?  A  ä Er        Ê #r      È !r       â rÿ   ÿä â !r       â G¹  àûÿÿÿÿƒ ê Ay        € ê r ÿ   pRð âEy  €     € â r ÿ   pRò Ö G‰  P     € ê z ÿ `  pPð ä r	   ÿ(Ž ä z ÿ a   Sð Æ ˆs €ÿÿ H   ô ˆ    H   ä Ay        € ê {          ì M        € ê z ÿ `  pPð â „y 	     H   f z ÿ a   Sð Ú „‰	 €   H   ¨„‰	    H   â !r        È? !‚       ÈO !‚       â� G‰        € ê „y 	 €  H   ( „y	 €   H   h „y	    H   ¨ „y	    H   è „y	 €  H   " !r         â !r       Æ/ !r         âO !r       Æ� !r        ä‰      " z ÿ `  pPð ä r 	ÿ   pRò ä z ÿ a   Sð â !r       Ê ‰       $ !r         Ê ‰  €    $ !r        Ê ‰ @    $ !r       Ê ‰      $ !r       â G‰        € ê ‰       $ !r         Ê ‰       $ !r        Ê ‰ €    $ !r       Ê ‰ @    $ !r       Ê ‰      $ !r       ä M        € ê !r       â G‰  `    € ê $v ÿ h  ÿ Ž â !r
       È x     pBð Ú G	  p    € ê x  ÿÿÿÿÿàÿ È x       € È x     ÿ   È ‚{   €     $ xÿ     È Iy Ðñÿÿÿÿƒ ê  x ;ª¸¿  A  Ì s          $ !t   €?    Ì s        $  r     A  È  r     A  â Gy  P    € ê  x '7=  A  â x*BL?  A  Æ #t   €?    È  r      A  â  x    ?  A  Ê s     $   $ !t  €?    È  r       A  È  r      A  â Gy  °    € ê x     pRð Ú G‰  à     € ê x     pRð Ú G	        € ê 	z  i    � ä 	z i    � Æ  x ;ª¸¿  A  â 	z i €  � Ê s        $ !t  €?    Ì s        $  r       A  È  r      A  â Gy  Ð     € ê  r     A  â Gy  °     € ê 	x   à@  � ä 	x  à@  � Æ  x #ÛÙ?  A  â 	x  àÀ  � Æ  x;ª¸¿  A  ä !t  €?    È s        $ !t  €?    Ì s        $  r       A  È  r      A  ä y       &   ( y      %   ( y      '   b $z   x  Ž â x        Æ $z  ~   Ž È/ %v  j   Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $	ÿ   ÿ Ž È % \  	 Ž Ì �	    ™ b r ÿ       â ¹z  l      â Ey  P
    € â y       "   b ™x?    â '
 m   Ž äy      !   d 'Š s   Ž â 
ÿ n    â y      '   ¨ $ÿÿ   
Ž È $
 o  Ž äy      %   â $‚ÿÿ    Ž È $z w  ÿŽ â r    ÿ(Ž È/ r	   ÿ� â $z }  ÿŽ ÆO xÿ   	  ä r   ÿàõ ä |    pbò Æ $rÿÿ   ÿ â Šÿ t    ä r ÿ       Î G  à    € ê rÿ   ÿ3Ž â� Ey @    € â r ÿ       ä |   ÿàÿ ä x		þÿÿÿÀŽ â 'z y   Ž â xÿ@   ÿÀ€ äxÿÀÿÿÿÿÀ‚ â $x   	
Ž â r ÿ       â $z	 p  ÿŽ â zÿ z    â $x   ÿ Ž Ì G	  p    € ê $z
 v  	Ž âr   ÿàñ â $tÿ$   ÿ Ž â x        â $z |  
Ž â r
ÿ   ÿä  â %v Z   Ž È $x   Ž ä $x
$   ÿŽ ä %v
 X  Ž Æ r   ÿàÿ Æ %x   
Ž ä�y

    ™ ¤ %x   Ž ä �y   ™ è �y   ™ ( �y   ™ ¨ �y   ™ ¨ �y   ™ ¨ �y    ™ ¨ �y   ™ ¢ x@   ÿàÿ Ä xÿÀŽ ä� xÿ     Æ &r   ÿ   âxÿÀŽ ä xÿÀŽ äO xÿ     â &r      Æ xÿÀŽ â &r      â 2r

        Æ &r      ä 0rÿ
  0 A   ä0r
ÿ
    A   â Er        ä  x   ?  A  È #r
      È !rÿ       ä Ay       € ê G™  p    € ê $z |  Ž â x
ÿ     â $tÿ   ÿ Ž ä $z v  Ž âx
   
  â $tÿ   ÿ Ž Æ r	   ÿàÿ Ê %v X  Ž Ê r        â $rÿÿ   	 Ž â x	   ÿ   â%v  € Ž È r    ÿàñ â$tÿ$   ÿ Ž â x@   ÿàÿ ä rÿ   ÿä  â %v   Z   Ž â r   ÿàñ ä r$   ÿàó â $x$   ÿŽ Ä $rÿÿ     â r%   ÿäÿ  ä r"	    ÿàñ ä r!!   ÿàÿ â %v Z   Ž â �y$   ™ ¦  $x$   ÿŽ â r#!
   ÿä  â �y$    ™ â  r	   ÿàñ â $x   Ž Ä �y"   ™ & r
   ÿä  â �y"   ™ h �y$   ™ h �y$  ™ h �y   ™ h �y     ™ b r$        â $r%ÿÿ    Ž Ä �y"   ™ h �y    ™ h �y    ™ h# �y!$    ™ h �y""   ™ h �y   ™ h �y$$   ™ h' �y#   ™ b x€   ÿàÿ È |    pbð ä x 
  ÿàó È rÿ   ÿäÿ  ä x%ÿÀŽ ä� xÿ     Æ &r%%   ÿ   âxÿÀŽ ä xÿÀŽ äO xÿ     â &r   %   âxÿÀŽ ä xÿ     ä xÿÀŽ â &r   ÿ   â xÿÀŽ Ä xÿÀŽ ä xÿ     â &r      Æ xÿÀŽ â &r       â 2r!        â &r"      â 2r$        â &r##      ä 0rÿ  0 A   âEr        â 0rÿ    A   ä &r   #   â  x   ?  A  â 0rÿ  0 A   Ä 0rÿ    A   â #r      â Er        â  x   ?  A  ä !r       ä #r      È !r       â G‰  Pûÿÿÿÿƒ ê Ay        € ê� r  ÿ   pRð är  ÿ   pRò Ö ˆ €ÿÿ H   è {          ì M        € ê „y      H   " r ÿ   pRð â $z x  Ž Ä „y €   H   h „y	    H   ¢ !r         È !r         È/ !r  	       ÊO ‰       $ !r        Ê ‰      $ !r       Ê ‰ €    $ !r       Ê ‰ @    $ !r       Ê ‰	      b  M	        € ê x        â $z ~  Ž â !r		       Æ/ %v j   Ž Ê †y 	    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú         â $ÿÿ    Ž È % \   Ž Ì �	    ™ b $ÿÿ   ÿ Ž â Ey       € â x
      ÿ  â y      "   b '
  m   Ž Æ y	      !   d 
 ÿ n     È   ÿ   ÿáÿ Ê $
  o  Ž ä $‚ÿÿ    Ž â ‚ ÿ       Æ $‚ÿÿ    Ž È 'Š  s   Ž â rÿ	   ÿü‚ ä/ r	   ÿ(Ž È r   ÿ� ä Šÿ t     Ê G      € ê z ÿ `  pPô ä z ÿ ^  pPò ä z ÿ b  p ö ä z ÿ a   Sô ä z ÿ _  Sò ä z ÿ c  0't Ö        '   "        ä¢       â ©      '   h 
      %   ¢ $ª x  ÿŽ Æ ©      %   â $   ~  ÿŽ ä $ª ~  ÿŽ ä/ $ x  ÿŽ â   	   
àë äO
ÿ   
  ä ¢	   àÇ Ä� ¨ÿ     ä ÿÿ   
Äþ ä ¢
   ÿàû ä ¢ÿÿ   „þ ä      ÿàý ä ¢ÿ   ÿäÿ Ä ª
 b  ÿˆ â $ÿÿ    â   ^  ÿ† ä ª
 c   ä x
      ÿ  â   _  � Ê �     h �©
     d Ay        € ê ¹z  l      â x ÿ     â ™x?    â Ey  à    € â x      ÿ  â ‚        Æ |     pbò Ú G  �    € ê y      '   " $rÿÿ   ÿ Ž â x        â $z w  ÿŽ â y      %   b xþÿÿÿÀŽ ä z ÿ `  pPð ä r  €ÿàÿ ä z ÿ a   Sð Æ $x   ÿ Ž ä 'z y   Ž â r ÿ       Æ $z }  ÿŽ â zÿ z    ä r ÿ       ä r   ÿàó â $z |   Ž Æ rÿÿ   ÿäÿ  â $z v  Ž È$z p  Ž È/ %v `  Ž È %v X  Ž È $rÿÿ    Ž ä r    ÿàó ä x $       ä r        â $rÿÿ   Ž  ä %v Z   Ž È $x$   ÿŽ â r        Æ $x   Ž ä %x   Ž â�y    ™ ¦ %x   Ž â �y   ™ è �y   ™ ( �y   ™ h �y   ™ ¨  �y    ™ ¨ �y   ™ ¨  �y   ™ ¢  x  @   ÿàÿ Ä r        ä r        ä x   ÿàó ä |     pbö Æ $rÿÿ   Ž  â xÿÀŽ ä� xÿ     Æ &r   ÿ   âxÿÀŽ ä xÿÀŽ äxÿ     â &r      ÆO xÿÀŽ â 2r        â &r      È 0rÿ  0 A   ä&r      ä 0rÿ    A   â  x   ?  A  ä Er        ä x   ÿàõ Æ #r      â G‰  P    € ê %x   Ž ä �y     ¨ �y    è �y    " xÿÀŽ Ä� xÿ     Æ &r   ÿ   â xÿÀŽ ä xÿÀŽ äxÿ     â &r      â 2r        äO xÿÀŽ â &r      Ä 0rÿ  0 A   ä&r      ä 0rÿ    A   â  x   ?  A  ä Er        Ê #r      È !r       â rÿ   ÿä â !r       â G¹  àûÿÿÿÿƒ ê Ay        € ê r ÿ   pRð âEy  €     € â r ÿ   pRò Ö G‰  P     € ê z ÿ `  pPð ä r	   ÿ(Ž ä z ÿ a   Sð Æ ˆs €ÿÿ H   ô ˆ    H   ä Ay        € ê {          ì M        € ê z ÿ `  pPð â „y 	     H   f z ÿ a   Sð Ú „‰	 €   H   ¨„‰	    H   â !r        È? !‚       ÈO !‚       â� G‰        € ê „y 	 €  H   ( „y	 €   H   h „y	    H   ¨ „y	    H   è „y	 €  H   " !r         â !r       Æ/ !r         âO !r       Æ� !r        ä‰      " z ÿ `  pPð ä r 	ÿ   pRò ä z ÿ a   Sð â !r       Ê ‰       $ !r         Ê ‰  €    $ !r        Ê ‰ @    $ !r       Ê ‰      $ !r       â G‰        € ê ‰       $ !r         Ê ‰       $ !r        Ê ‰ €    $ !r       Ê ‰ @    $ !r       Ê ‰      $ !r       ä M        € ê !r       â G‰  `    € ê $v ÿ h  ÿ Ž â !r
       È x     pBð Ú G	  p    € ê x  ÿÿÿÿÿàÿ È x       € È x     ÿ   È ‚{   €     $ xÿ     È Iy Ðñÿÿÿÿƒ ê  x ;ª¸¿  A  Ì s          $ !t   €?    Ì s        $  r     A  È  r     A  â Gy  P    € ê  x '7=  A  â x*BL?  A  Æ #t   €?    È  r      A  â  x    ?  A  Ê s     $   $ !t  €?    È  r       A  È  r      A  â Gy  °    € ê x     pRð Ú G‰  à     € ê x     pRð Ú G	        € ê 	z  i    � ä 	z i    � Æ  x ;ª¸¿  A  â 	z i €  � Ê s        $ !t  €?    Ì s        $  r       A  È  r      A  â Gy  Ð     € ê  r     A  â Gy  °     € ê 	x   à@  � ä 	x  à@  � Æ  x #ÛÙ?  A  â 	x  àÀ  � Æ  x;ª¸¿  A  ä !t  €?    È s        $ !t  €?    Ì s        $  r       A  È  r      A  ä y       &   ( y      %   ( y      '   b $z   x  Ž â x        Æ $z  ~   Ž È/ %v  j   Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y       &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú         ä r         Ê % \   Ž Ê �	    ™ b ¹z  l      â Ey  Ð	    € â ™x?    â y      "   b r ÿ       Æ y      !   ¢  ÿ       ä         â y	      %   ( '
 m   Ž â y      '   â ‚ ÿ       Æ 
ÿ n    È ÿ   ÿáÿ ä r ÿ   pRò Æ/ $
 o  Ž â r   ÿ(Ž äO r ÿ       â $‚ÿÿ    Ž â r
   ÿ� ä x		   ÿ   â 'Š s   Ž â xÿ   
  È |    pbô ä Šÿ t    ä r ÿ   pRð âx      ÿ  â    ÿ8Ž Î G)      € ê r ÿ       â� $z w  ÿŽ â x

þÿÿÿÀŽ â $z }  ÿŽ âz  X      â 'z y   Ž â r

  €ÿàÿ ä r   ÿàõ ä zÿ z    â $z	 p  Ž â z  Y      â $x

   ÿ Ž Ä $z |  Ž â r ÿ       ä xÿ   
  â $z v  Ž ârÿÿ   ÿä ä x
     â %x   ÿŽ Ì %x
   Ž È r   ÿàõ ä x $       ä rÿ   ÿä ä r   ÿàõ â %v Z   Ž â z+  p      Æ $x$   ÿŽ ä $x    â r,
   ÿ„ ä r   ÿàÿ â %x +   Ž â �y   ™ ¤  r-   ÿä ä �y"    ™ è �y(,   ™ b	 %x+    Ž Æ �y%   ™ â  r   ÿàõ Æ �y),   ™ â r   ÿä Æ �y#    ™ â %x+   Ž Æ �y&,   ™ è �y   ™ è  �y'   ™ è  �y    ™ è  �y*    ™ â %x +   Ž Æ/ �y$,   ™ è �y   ™ è �y   ™ "%x,+    Ž Æ/ �y     ™ & %x+   ,Ž ä �y-,    ™ ( �y+    ™ " x@   ÿàÿ È |    pbô ä x   ÿà÷ È rÿ   ÿäÿ ä xÿÀŽ äO xÿ     È xÿÀŽ â &r(   ÿ   â
x"ÿÀŽ ä� x"ÿ   "  ä xÿ   %  â &r(   ÿ   â x%%ÿÀŽ â &r)      â x""ÿÀŽ Ä x!ÿÀŽ ä/ x#ÿÀŽ â &r")      ä &r%&      âxÿÀŽ ä xÿ     â &r&      â x'ÿÀŽ â &r(   ÿ   â xÿÀŽ Ä x'ÿ   '  â 2r**        ä &r)      âx 'ÿÀŽ â &r(   ÿ   ä 0rÿ*  0 A   â x#ÿ   #  â &r!$      ä &r) )      â  x   ?  A  â x##ÿÀŽ Ä Er        ä xÿÀŽ â2r        â0r*ÿ*    A   â x ÿÀŽ â &r#$      â xÿ     â &r&      âxÿ     â 0rÿ  0 A   â#r*      â 0rÿ    A   â Er        â 2r--        âxÿÀŽ â &r  &   )   â xÿÀŽ â  x   ?  A  â 2r+        â &r$      Ä #r      â &r$       ä 0rÿ-  0 A   ä0rÿ  0 A   âEr        â 0r-ÿ-    A   â Er        â 0rÿ    A   â  x   ?  A  â  x   ?  A  Æ #r-      â #r      â !r       â !r       â !r       â !r       â G©  0ùÿÿÿÿƒ ê Ay        € ê� ˆ  þÿ H   è ˆ €þÿ H   è ˆ  ÿÿ H   è ˆ €ÿÿ H   è {          ì M	        € ê „y     H   " z 	 r  p`ô Æ „y €   H   b r ÿ   pVt Æ „y    H   ¨ „y    H   è„y €  H   ( „y    H   ( „y
    H   ( „y €  H   ( „y €  H   ( „y    H   ( „y €  H   " !r       Æ „y €  H   " !r       â/ !r       âO !r       ä� !r       â!r       â !r
       â !r       È ‰      b !r       â !r       È ‰      ¢ !r       Æ ‰      â !r       Ê ‰      " !r       â/ !r       âO !r       È� ‰      h ‰      ¢ !r       Æ ‰      ( ‰      â !r       â/ !r       ÈO ‰
 €    b !r       Æ ‰ €    " !r       Æ� ‰ €    ¨ ‰ €    â !r

       â/ !r       È ‰
 @    " !r       ÆO ‰ @    b !r       Æ� ‰ @    ¨ ‰ @    â !r
       â x
	   ÿàÿ â !r       Æ/ ‰      " $z  x  	Ž â x 	   ÿàÿ â!r       âO ‰      b x	   ÿàÿ â $z ~  Ž â !r       â� ‰      ¢ z   r  p`ö ä z  r  p`ð â ‰      â z 
 r  p`ò Ä x    pVö äx    pVp  äx    pVò  ä x         â !r       È %v j    Ž â !r       È/ †¹    â !r       ÆO †©     â !r       Æ� †‰    â M        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y	      &   " z ÿ \  pPò â ¹z  F   
   Æ z ÿ ]  Sò Ú         Ê %	 \   Ž Ê �    ™ ¢  r ÿ       â Ey       € ä y      "   d '	 m   Ž äy      %   æ ÿ n    â y       !   & ÿ   ÿáÿ â y      '   " 'š	 s   Ž È $ o  	Ž â ’ 	       ä r ÿ   pRð ä/ x   ÿ   ä� x     pDp  är    ÿàÿ är    ÿ(Ž Ä z  r  pdp  ä r ÿ       ä šÿ t    ä r ÿ       ä r	     € ÊO G	  À    € ê z ÿ `  pPò ä z ÿ ^  pPð ä z ÿ b  p ô ä z ÿ a  Sò ä z ÿ _   Sð ä z ÿ c   'ò Ö $
 ~  ÿŽ âÿ     â$š	 ~  ÿŽ ä $
 x  ÿŽ â     àë â$š
 x  ÿŽ â ’		    àÇ ä ˜ÿ     ä    ÿàõ Ä ÿ   ÿÄþ ä ’
	   ÿàû ä ’
ÿ   ÿ„þ ä 	ÿ   ÿä ä 
 ^  ÿ„ ä ’
ÿ
   ÿäÿ Ä š b  ÿ† ä
	 _  	 ä r ÿ       ä r ÿ       ä š c  
� ä �	     h �™     d Ay        € ê r   ÿ� â ¹z  l      â $z
 }  ÿŽ â ™x?    â xÿ     â $z w  ÿŽ â r ÿ       â Ey       € â r	 ÿ       â x      ÿ  â |    pbð â 'z y   Ž â r
   ÿàó â x      ÿ  â r- ÿ       â x
      ÿ  â rÿÿ   ÿäÿ  Î G	  P    € ê zÿ z    â $z p  Ž âxþÿÿÿÀŽ â x
      ÿ  â x   ÿàÿ â $z |  Ž â r&  €ÿàÿ â $z |  Ž â x   ÿàÿ â $z v  Ž âx&&   ÿ   â $z v  Ž â x   ÿàÿ â $z p  Ž âz ÿ `  pPð â %x&   ÿŽ â x'ÿ   &  â x      ÿ  â z  `      â $z p  Ž âz  a      â $z p  Ž â z  X      â %x   Ž âz  Y      â x      ÿ  â r	 ÿ       â %x   Ž â r- ÿ       Ä z ÿ a   Sð â %x   Ž â x'&   '  Æ %x   Ž È r    ÿàó ä x! $       ä rÿ   ÿäÿ  ä r0   ÿàó â %v   Z  ! Ž Æ r1   ÿäÿ  â $x$   ÿŽ â z  p      ä r$&    ÿ‚ â �y.0   ™ ¢  r!!   ÿàÿ â %x"   0Ž ä �y50   ™ â  r%!'   ÿäÿ  Æ �y2"   ™ ( �y($   ™ è  �y3"   ™ è �y)$   ™ â  %x   Ž Æ �y*$   ™ â  %x"   "Ž Æ/ �y+     ™ è �y1"   ™ è �y,$   ™ â  %x    "Ž Æ/ �y4    ™ è �y/"   ™ â  %x   Ž Æ/ �y"    ™ æ %x$   Ž â�y0    ™ è �y#    ™ â  %x   $Ž Æ/ �y$    ™ è �y    ™ ä x.ÿÀŽ äO x$ÿ   .  ä/ x.2ÿÀŽ Äx2ÿ   2  ä x$$ÿÀŽ â &r(   ÿ   â�x22ÿÀŽ â &r..(   ÿ   â x 5ÿÀŽ ä x5ÿ   5  â &r!$)      âx$3ÿÀŽ â &r2)   .   â x.ÿ   3  â &r! *   !   âx55ÿÀŽ â &r$*      â x..ÿÀŽ Ä x31ÿÀŽ ä x1ÿ   1  â &r!5,   !   Æ x 1ÿÀŽ â &r33(   ÿ   ä &r..,      â Er1 !       â &r )   3   Æ Er3 .       ä x /ÿÀŽ â 2r%4+        âx!"ÿÀŽ ä x.ÿ   "  â &r *      ä &r!!(   ÿ   â x..ÿÀŽ â 2r$0+        âx #ÿÀŽ â 0r0ÿ%  0 A   âx/ÿ   /  â &r!.)   !   â x#ÿ   #  â 0r%ÿ%    A   â x//ÿÀŽ â &r  *   !   â 2r+        âx##ÿÀŽ â  x00   ?  A  â 0r2ÿ$  0 A   â 2r!+        â &r/,      â #r"%1   0   â 0rÿ  0 A   Ä &r #,       â  x52   ?  A  â 0r$ÿ$    A   ä 0r%ÿ!  0 A   âEr        â 0r#ÿ!    A   â  x!   ?  A  â 0rÿ    A   â Er          â #r$$3   5   â  x%   ?  A  â !r-"-       â #r   !   â !r	$	       â #r##       â G‰  p    € ê r   ÿàó È r   ÿäÿ  Ê �y$    ¢ %x    Ž Æ �y"    è �y       " x%$ÿÀŽ äO x$ÿ   $  ä x."ÿÀŽ â� &r%%(   ÿ   â x$$ÿÀŽ Ä x"ÿ   "  â 2r +        ä&r%$)   %   â x/"ÿÀŽ ä 0rÿ  0 A   ä&r%.*   %   ä 0rÿ    A   â  x   ?  A  â &r%/,   %   Ê Er  %       Ê #r       È !r       ä G‰  �    € ê r   ÿàó ä x    ÿàÿ ä r   ÿäÿ  Ê �y$    ¢ $z!  p  Ž Æ �y"    â %x !   Ž Ì �y       " x%$ÿÀŽ äO x$ÿ   $  Æ &r%%(   ÿ   â x$$ÿÀŽ ä x."ÿÀŽ ä� x"ÿ   "  â &r%$)   %   â 2r +        äx/"ÿÀŽ â &r%.*   %   Ä 0rÿ  0 A   ä&r%/,   %   ä 0rÿ    A   â  x   ?  A  ä Er  %       Ê #r       È !r       ä G‰  �    € ê r   ÿàó ä x    ÿàÿ ä r   ÿäÿ  Ê �y$    ¢ $z!  p  Ž Æ �y"    â %x !   Ž Ì �y       " x%$ÿÀŽ äO x$ÿ   $  Æ &r%%(   ÿ   â x$$ÿÀŽ ä x."ÿÀŽ ä� x"ÿ   "  â &r%$)   %   â 2r +        äx/"ÿÀŽ â &r%.*   %   Ä 0rÿ  0 A   ä&r%/,   %   ä 0rÿ    A   â  x   ?  A  ä Er  %       Ê #r       È !r       â !r       â !r#       â G‰  �    € ê r   ÿàó ä x   ÿàÿ ä r   ÿäÿ  Ê �y"    ¢ $z! p  Ž Æ �y    â %x !   Ž Ì �y       " x#"ÿÀŽ äO x"ÿ   "  Æ &r##(   ÿ   â x""ÿÀŽ ä x$ÿÀŽ ä� xÿ     â &r#")   #   â 2r+ +        äx%ÿÀŽ â &r#$*   #   Ä 0rÿ+  0 A   ä&r#%,   #   ä 0r+ÿ+    A   â  x   ?  A  ä Er #       Ê #r++      È !r

+       â x@   ÿàÿ ä x   ÿàõ ä |    pbò ä x   ÿà÷ ä rÿ   ÿä ä rÿ   ÿäÿ Î G™  Àòÿÿÿÿƒ ê Ay        € ê y      "   " Ey  0    € â r ÿ   pRð är ÿ   pRò Ö G‰  ð     € ê z ÿ `  pPð ä r    ÿ8Ž ä z ÿ a   Sð Æ ˆs - þÿ H   ô ˆƒ 	€þÿ H   è ˆƒ  ÿÿ H   è ˆƒ €ÿÿ H   â G‰  p     € ê ˆs    H   è ˆs 	€þÿ H   è ˆs €  H   è ˆs  ÿÿ H   è ˆs    H   è ˆs €ÿÿ H   è ˆs 
€  H   ä Ay        € ê {          ì M        € ê y      &   ¢ z ÿ `  pPð Æ y      '   "> z ÿ a   Sð â $z x  Ž âO x        Æ $z ~  Ž È %v j   Ž È G	  P    € ê „y      H   "z  r  p`ð Æ „y
     H   b r  ÿ   pVp  Æ „y     H   ¨ „y  €   H   è „y     H   ( „y     H   ( „y  €  H   ( „y     H   ( „y  €  H   ( „y  €  H   ( „y  €  H   " !r-       Æ „y  €  H   " !r

       â/ !r       âO !r	       ä� !r

       â!r       â !r       È ‰      b !r

       â !r       È ‰
      ¢ !r       Æ ‰      â !r       È !r       â !r       È/ ‰      ( ‰      b !r
       âO !r	       È� ‰      ¨ ‰	      â !r       â !r       È/ ‰      " !r       ÆO ‰ €    b !r	       Æ� ‰ €    ¨ ‰ €    â !r       Ê ‰ €    " !r       â/ !r       ÈO ‰ @    b !r       â� x   ÿàÿ ä ‰ @    ¤ z  r  p`ô ä ‰
 @    ä x     pVt â !r       Ê ‰ @    " !r       â/ x   ÿàÿ äx   ÿàÿ â !r       âO z  r  p`ò â ‰      b !r

       ä� x     pVò  â ‰      è ‰	
      ¢ !r       â !r       È/ ‰      " !‚       â !r
	       ÈO †‰     â z  r  p`ð â !r	       â !’       äx     pVp  â !¢		       ä †™    è †© 	   â !r       Ê M	        € ê !r       Ê †y    â My        € ê z  h      È x    pBð Ú G	  @    € ê x    pRð Ú G‰  À    € ê x    pRð Ú G	   $    € ê „y      H   " z  r  p`ð â Ey  p    € ä „y     H   b r  ÿ   pVp  Æ „y     H   ¨ „y     H   è „y     H   ( „y   
  H   ( „y  €   H   ( „y  €  H   è „y  €  H   è „y  €
  H   â !r-       â !r       Æ/ !r       âO !r       Æ� !r       â!r       È ‰      " !r	       Æ ‰      b !r       Æ „y  €  H   â !r       Æ/ ‰      ( ‰      b !r       â !r       È/ ‰ €    ( ‰ €    b !r       Æ „y  €  H   ¢ !r       È !r       â !r       â !r       È/ ‰ @    ( ‰ @    b !r       ÈO !r       â !r       â !r       È/ ‰      ( ‰      b !r       â !r       â/ G	  �     € ê !r       â!r       Æ  x	;ª¸¿  A  Ì s	 	       $ !t	  €?    Ì s        $  r     A  È  r     A  Ê †y     ä Ay        € ê ‰      " Ey  Ð    € æ ‰      h „y     H   ¨ „y     H   è „y     H   è „y     H   â !r       â !r       È/ ‰      ( ‰      b !r       âO !r       â� !r       ä !r       â !r       Ä/ ‰	 €    " !r       Æ ‰ €    h „y   	  H   ¢ !r		       Æ „y     H   â !r       â/ x   ÿàÿ ä ‰	 @    $ z  r  p`ð ä ‰ @    d x     pVp  â !r       âO !r	       â !r       È/ ‰      ( ‰	      b !r       â !r       â !r	       â/ G	  �     € ê !r       â!r       Æ  x;ª¸¿  A  Ì s        $ !t	  €?    Ì s 	       $  r     A  È  r     A  Ê †y    ä Ay        € ê ‰      " Ey       € æ ‰      h „y  €  H   ¨ „y  €	  H   è „y  €  H   ( „y  €  H   â !r       Æ „y  €  H   â !r	       Æ/ „y  €  H   è ‰      ( ‰	      b !r
       ÈO !r       È� !r       â!r       â !r	       È/ ‰ €    ( ‰ €    b !r       â !r       È/ ‰ @    ( ‰ @    b !r       â x   ÿàÿ â !r       Æ/ ‰      " z  r  p`ð Æ ‰      b x     pVp  â !r       â !r       Ö/ G	  �     € ê !r       â!r       Æ  x;ª¸¿  A  Ì s        $ !t	  €?    Ì s
 	       $  r

     A  È  r
     A  Ê †y    ä Ay        € ê ‰      b !r       â x   ÿàÿ Æ !r       â z  r  p`ð Æ !r       â x     pVp  È ‰      " !r	       Ê/ ‰	      b !r       Ê ‰      " !r
	       Ê/ ‰
 €    b !r       Ê ‰ €    " !r
       Ê/ ‰ @    b !r       Ê ‰ @    " !r       Ê/ ‰	      b !r       Ê ‰      ¢  !r		       â/ M	        € ì !r	       â!r       ÆO  x ;ª¸¿  A  â !r       Ê s          $ !t   €?    Ì s        $  r     A  È  r     A  Ê †y    â My        € ê „y      H   " z  r  p`ð â Ey       € ä „y     H   b r  ÿ   pVp  Æ „y     H   ¨ „y     H   è „y     H   ( „y   
  H   ( „y  €   H   ( „y  €  H   è „y  €  H   â !r-       â !r       Æ/ !r       âO !r       Æ� !r       â!r       È ‰      " !r	       Æ ‰      b !r       â !r       È/ ‰      ( ‰      h „y  €  H   ¢ !r       â !r       È/ ‰ €    ( ‰ €    h „y  €  H   â !r       ÈO !r       â !r       ä „y  €
  H   â !r       Æ/ ‰ @    ( ‰ @    b !r       È� !r       â !r       â !r       È/ ‰      ( ‰      b !r       â !r       â !r       â/ G	  À     € ê !r       È x'7=  A  â x*BL?  A  Æ #t	  €?   È  r		     A  â  x   ?  A  â !r       È s	 	    $   $ !t	  €?    È  r     A  È  r     A  Ê †y     ä Ay        € ê ‰      " Ey       € æ ‰      h „y     H   ¨ „y   	  H   è „y     H   è „y     H   â !r       Æ „y     H   â !r       Æ/ ‰      ( ‰      b !r       âO !r       â !r       È/ ‰	 €    ( ‰ €    b !r       Æ „y     H   ¢ !r       È !r       â !r		       â x   ÿàÿ â !r       Æ/ ‰	 @    " z  r  p`ð Æ ‰ @    b x     pVp  â !r       ÈO !r       â !r	       â !r       È/ ‰      ( ‰	      b !r       â !r	       â/ G	  À     € ê !r       È x'7=  A  â x*BL?  A  Æ #t  €?   È  r     A  â  x   ?  A  â !r       È s     $   $ !t  €?    È  r     A  È  r     A  Ê †y    ä Ay        € ê ‰      " Ey  Ð    € æ ‰      h „y  €  H   ¨ „y  €	  H   è „y  €  H   è „y  €  H   â !r       Æ „y  €  H   â !r	       Æ/ ‰      ( ‰	      h „y  €  H   " !r
       ÈO !r       â� !r       â !r	       È/ ‰ €    ( ‰ €    b !r       â!r       â x   ÿàÿ â !r       Æ/ z  r  p`ð â ‰ @    & x     pVp  â ‰ @    b !r       â !r       È/ ‰      ( ‰      b !r	       â !r       â/ G	  À     € ê !r       È x'7=  A  â x*BL?  A  Æ #t  €?   È  r     A  â  x   ?  A  â !r	       È s     $   $ !t
  €?    È  r

     A  È  r
     A  Ê †y    ä Ay        € ê ‰      b !r       â x   ÿàÿ Æ !r       â z  r  p`ð Æ !r       â x     pVp  È ‰      " !r	       Ê/ ‰	      b !r       Ê ‰      " !r
	       Ê/ ‰
 €    b !r       Ê ‰ €    " !r
       Ê/ ‰ @    b !r       Ê ‰ @    " !r       Ê/ ‰	      b !r       Ê ‰      ¢  !r		       â/ M	        € ì !r	       È x'7=  A  â x *BL?  A  Æ #t  €?   â x   ?  A  Æ  r       A  â !r       ÈO !r       â s       $   $ !t   €?    È  r     A  È  r     A  Ê †y    â My        € ê x    pRð Ú G‰  P    € ê x    pRð Ú G	  `    € ê „y      H   " z  r  p`ð â Ey       € ä „y     H   b r  ÿ   pVp  Æ „y     H   ¨ „y     H   è „y     H   ( „y   
  H   ( „y  €   H   ( „y  €  H   è „y  €  H   â !r-       â !r       Æ/ !r       âO !r       Æ� !r       â!r       È ‰      " !r	       Æ ‰      b !r       â !r       È/ ‰      ( ‰      h „y  €  H   ¢ !r       â !r       È/ ‰ €    ( ‰ €    h „y  €  H   â !r       ÈO !r       â !r       ä „y  €
  H   â !r       Æ/ ‰ @    ( ‰ @    b !r       È� !r       â !r       â !r       È/ ‰      ( ‰      b !r       â !r       â !r       â/ G	  À     € ê !r       Ê	z i    � â !r       È  x	;ª¸¿  A  â 	z i    � È 	z i €  � â s	 	       $ !t	  €?    Ì s        $  r     A  È  r     A  Ê †y     ä Ay        € ê ‰      " Ey       € æ ‰      h „y     H   ¨ „y   	  H   è „y     H   è „y     H   â !r       Æ „y     H   â !r       Æ/ ‰      ( ‰      b !r       âO !r       â !r       È/ ‰	 €    ( ‰ €    b !r       Æ „y     H   ¢ !r       È !r       â !r		       â x   ÿàÿ â !r       Æ/ ‰	 @    " z  r  p`ð Æ ‰ @    b x     pVp  â !r       ÈO !r       â !r	       â !r       È/ ‰      ( ‰	      b !r       â !r	       â/ G	  À     € ê !r       Ê	z i    � â !r       È  x;ª¸¿  A  â 	z i    � È 	z i €  � â s        $ !t	  €?    Ì s 	       $  r     A  È  r     A  Ê †y    ä Ay        € ê ‰      " Ey  Ð    € æ ‰      h „y  €  H   ¨ „y  €	  H   è „y  €  H   è „y  €  H   â !r       Æ „y  €  H   â !r	       Æ/ ‰      ( ‰	      h „y  €  H   " !r
       ÈO !r       â� !r       â !r	       È/ ‰ €    ( ‰ €    b !r       â!r       â x   ÿàÿ â !r       Æ/ z  r  p`ð â ‰ @    & x     pVp  â ‰ @    b !r       â !r       È/ ‰      ( ‰      b !r       â !r       â/ G	  À     € ê !r       Ê	z i    � â !r       È  x;ª¸¿  A  â 	z i    � È 	z i €  � â s        $ !t	  €?    Ì s
 	       $  r

     A  È  r
     A  Ê †y    ä Ay        € ê ‰      b !r       â x   ÿàÿ Æ !r       â z  r  p`ð Æ !r       â x     pVp  È ‰      " !r	       Ê/ ‰	      b !r       Ê ‰      " !r
	       Ê/ ‰
 €    b !r       Ê ‰ €    " !r
       Ê/ ‰ @    b !r       Ê ‰ @    " !r       Ê/ ‰	      b !r       Ê ‰      ¢  !r		       â/ M	        € ì !r	       â!r       ÈO 	z i    � â !r       È  x ;ª¸¿  A  â 	z i    � È 	z i €  � â s          $ !t   €?    Ì s        $  r     A  È  r     A  Ê †y    â My        € ê „y      H   " z  r  p`ð Æ „y     H   b r  ÿ   pVp  Æ „y  €   H   ¨ „y  €  H   è „y     H   ( „y      H   ( „y     H   ( „y     H   ( „y  €  H   ( „y  €  H   ( „y%     H   " !r--       Æ „y'   	  H   " !r       Æ/ „y     H   b !r	       ÆO „y   
  H   ¢ !r       Æ� „y!  €  H   â !r       Æ„y#  €
  H   " !r         Æ „y     H   " !r-       Æ „y     H   " !r       Æ „y  €  H   " !r       Æ „y  €  H   " !r       Æ „y  €  H   " !r%       Æ „y  €	  H   â !r' '       Æ „y  €  H   â !r	       Æ/ „y  €  H   â !r       âO !r!       ä� ‰	      " !r#       Æ‰      b !r       Æ ‰      ¢ !r'       Æ ‰      â !r       Æ ‰      " !r
       Æ ‰      " !r       â !r	       â !r       â !r       Æ/ ‰      " !r       â !r       äO ‰      â !r       â !r       ä� ‰      b !r       Æ‰      ¢ !r       Æ ‰
      è ‰      ( ‰	      " !r       Æ ‰      " !r       â/ !r       ä ‰ €    b !r       ÆO ‰      ¢ !r

       Æ� ‰      â !r       Æ‰ €    " !r	       Æ ‰	
 €    " !r       Æ ‰ €    ( ‰ €    " !r       Æ/ ‰ €    b !r       âO !r       ä� ‰ @    ¢ x   ÿàÿ â !r       ä‰ €    â z  r  p`ò â !r	
	       Ä ‰ €    " x     pVò  â !r       ä ‰ @    " !r       Æ ‰
	 @    " !r       Æ/ ‰ @    h ‰ @    b !r       ÆO ‰ @    ¢ !r       â� !r       ä‰      â x   ÿàÿ â !r       ä ‰ @    " z  r  p`ô â !r
	
       Ä ‰ @    " x     pVt â !r       ä/ ‰      b x   ÿàÿ â !r       ä ‰	
      b z  r  p`ö â !r       ÄO ‰      ¢ x     pVö Æ ‰      ¢ !r       Æ� ‰      â !r       â !‚       â!r       Æ‰      " !r       Æ/ ‰      b !r
	       â !‚       â !r       äO !’       â  ‚     A  â !r       â !’       Æ †‰     â !r       â� !¢       â  ’     A  ä !¢
       â !r       ä †™    â  ¢
     A  â !r       È/ †©    â M9        € ê !r       â !r       ÈO  r     A  Ê †y    â My        € ê „y      H   " z  r  p`ð â Ey  À    € ä „y     H   b r  ÿ   pVp  Æ „y     H   ¨ „y     H   è „y     H   ( „y   
  H   ( „y  €   H   ( „y  €  H   è „y  €  H   â !r-       â !r       Æ/ !r       âO !r       Æ� !r       â!r       È ‰      " !r	       Æ ‰      b !r       â !r       È/ ‰      ( ‰      h „y  €  H   ¢ !r       â !r       È/ ‰ €    ( ‰ €    h „y  €  H   â !r       ÈO !r       â !r       ä „y  €
  H   â !r       Æ/ ‰ @    ( ‰ @    b !r       È� !r       â !r       â !r       È/ ‰      ( ‰      b !r       â !r       â !r       â/ G	  à     € ê !r       Ê	x  à@  � Ê  x#ÛÙ?  A  È  x	;ª¸¿  A  â !r       Ê s	 	       " 	x  à@  � È 	x  àÀ  � Ê !t  €?    â !t	  €?    È s        $  r     A  È  r     A  Ê †y     ä Ay        € ê ‰      " Ey       € æ ‰      h „y     H   ¨ „y   	  H   è „y     H   è „y     H   â !r       Æ „y     H   â !r       Æ/ ‰      ( ‰      b !r       âO !r       â !r       È/ ‰	 €    ( ‰ €    b !r       Æ „y     H   ¢ !r       È !r       â !r		       â x   ÿàÿ â !r       Æ/ ‰	 @    " z  r  p`ð Æ ‰ @    b x     pVp  â !r       ÈO !r       â !r	       â !r       È/ ‰      ( ‰	      b !r       â !r	       â/ G	  à     € ê !r       Ê	x  à@  � Ê  x#ÛÙ?  A  È  x;ª¸¿  A  â !r       Ê s        " 	x  à@  � È 	x  àÀ  � Ê !t  €?    â !t	  €?    È s 	       $  r     A  È  r     A  Ê †y    ä Ay        € ê ‰      b Ey  ð    € æ ‰      ( „y  €  H   ¨ „y  €	  H   è „y  €  H   è „y  €  H   â !r       Æ/ „y  €  H   â !r	       Æ ‰      ( ‰	      h „y  €  H   " !r
       ÈO !r       â� !r       â !r	       È/ ‰ €    ( ‰ €    b !r       â!r       â x   ÿàÿ â !r       Æ/ z  r  p`ð â ‰ @    & x     pVp  â ‰ @    b !r       â !r       È/ ‰      ( ‰      b !r       â !r       â/ G	  à     € ê !r       Ê	x  à@  � Ê  x#ÛÙ?  A  È  x;ª¸¿  A  â !r       Ê s        " 	x  à@  � È 	x  àÀ  � Ê !t
  €?    â !t	  €?    È s 	       $  r     A  È  r
     A  Ê †y    ä Ay        € ê ‰      " !r       â x   ÿàÿ Æ !r       â z  r  p`ð Æ !r       â x     pVp  È ‰      b !r	       Ê ‰	      " !r       Ê/ ‰      b !r
	       Ê ‰
 €    " !r       Ê/ ‰ €    b !r
       Ê ‰ @    " !r       Ê/ ‰ @    b !r       Ê ‰	      " !r       Ê/ ‰      ¢ !r		       â M	        € ì !r	       â/!r       ÈO 	x  à@  � â !r       È  x #ÛÙ?  A  â 	x  à@  � Æ  x  ;ª¸¿  A  â 	x  àÀ  � Ê s          " !t  €?    â !t   €?    Ì s        $  r     A  È  r     A  Ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y       "   $ z   w  p`ð Ú M	        € ê y      !   " ¹z  l      â Ey       € â ™x?    â y      %   b ¹z  F   
   â $rÿÿ   ÿ Ž â r ÿ       ä xÿ     È |    p`ð â $x   ÿ Ž Ø/ G	  €    € ê y      &   " x        â $z  x  Ž È %v \   Ž Ê �y    ™ ¢  $rÿÿ   ÿ Ž â x   ÿ   â $z	  r  ÿŽ â z  X      â 'z m   Ž â r ÿ       Æ $rÿÿ   ÿ Ž â zÿ n    â $vÿ Y  ÿ Ž Æ rÿ   ÿáÿ â %x	$   ÿ Ž È $z o  Ž ä %x$    Ž â x   ÿÀŽ Æ $z u  ÿŽ â z Z  ÿàñ â %x   ÿ Ž Æ r   	à£ â%x$   ÿ Ž â z [  ä  ä r	ÿÿ   ÿDþ  â %x$    Ž â x   ÿàÿ ä r        â $x		$   ÿŽ â z Z  ÿàñ â $z q  Ž Æ z	 [  	ä  â $z t  Ž äO$z t  Ž â x   ÿàñ â $z q  Ž ä %x   Ž â rÿ	   ÿä  Æ %x   Ž È r   ÿàñ â�yðÿÿ ™ ¢  r   ÿàó Æ $x     â r   ÿäÿ  â �y"    ™ è  �y!   ™ ( �y    ™ ¢
 r   ÿàñ Æ �y$   ™ â
 r	   ÿä  Æ �y#   ™ è �y%ôÿÿ ™ â  r        â$rÿÿ    Ž ä �y	    ™ ä
 %x
   Ž Ä �y   ™ ä  %x   Ž ä �y

    ™ è �y    ™ â x   ÿàÿ È |    p`ð ä x@  ÿàó ä x@  ÿà÷ ä x@  ÿàõ â $rÿÿ   Ž  â rÿ   ÿäÿ ä rÿ   ÿä Ä x!ÿÀŽ äx!ÿ   !  ä x ÿÀŽ äO x ÿ      â &r   ÿ   âx!!ÿÀŽ â &r   ÿ   â x  ÿÀŽ Ä x#ÿÀŽ â� &r!"      â
x$ÿÀŽ â &r  "      â x$ÿ   $  ä x#ÿ   #  â &r%      âx$ÿÀŽ â &r%       â x##ÿÀŽ Æ &r      â2r

	        ä2r		        â &r#      ä 0rÿ
  0 A   ä0rÿ
    A   ä 0r
ÿ	  0 A   âEr        â 0r	ÿ	    A   â Er        â  x   ?  A  â  x

   ?  A  Æ #r      â #r	   
   Æ !r       â !r       â G‰   üÿÿÿÿƒ ê Ay        € ê ‰      " r   ÿàÿ Æ ‰      b z  p  p`ð È x    pDp  â !r       â !r       È/ ‰      ( ‰      b !r       â !r       È/ ‰ €    ( ‰ €    b !r       â !r       È/ ‰ @    ( ‰	 @    b !r       â !r		       È/ ‰      h  ‰	      ¢  M	        € ê y      &   â $x   ÿ Ž â !r	       âO $z   s  Ž â x        â !r       â? x    p"ò Ä r ÿ   p"ð Ö         â $z  v   Ž È� %v  j   Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      "   $ z  w  p`ð Ú M	        € ê y      &   " x        â ¹z  F   
   â $z x  Ž È %v \   Ž Ê �y     ™ b ¹z  l      â Ey  À
    € â ™x?    â y      !   b x      ÿ  â r ÿ       â $rÿÿ   ÿ Ž â x      ÿ  â x	ÿ     È/ | 	   p`ð Ú G	   
    € ê r
 ÿ       â y      %   " $z r  ÿŽ â x   ÿÀŽ âEy °	    € â 'z m  
 Ž â x   ÿ   ä r 	       â %x   ÿ Ž â zÿ n    Ä r ÿ       â $z  t  Ž ârÿ   ÿáÿ â $vÿ Y  ÿ Ž â z  X      ä z  a      â $z o  Ž ä %x
$   ÿ Ž È $z u  ÿŽ ä %x
$   
 Ž Æ r   à‘ â$x   ÿ Ž ä %x$   ÿ Ž â rÿÿ   ÿ$~  ä x	   ÿàÿ â %x$    Ž â z
 Z  ÿàñ Æ $x$   ÿŽ â z Z  ÿàó â $z q  Ž â z
 [  ä  â $z		 q  Ž â z [  äÿ  â $z  t  Ž â x   ÿàó ä z ÿ `  pPð â %x   Ž â r        Ä z ÿ a   Sð â %x	   Ž â z  `      ä r ÿ       â $rÿÿ   Ž  È r   ÿàó â�yðÿÿ ™ ¦ r   ÿäÿ  â �y    ™ â r   ÿàó Æ �y!   ™ " r   ÿäÿ  Æ �y"   ™ h �y#   ™ â r   ÿàó Æ �y ôÿÿ ™ â %x   Ž âr
   ÿäÿ  ä �y$   ™ è �y%    ™ è  �y    ™ â %x	   Ž Æ �y   ™ è. �y    ™ â  x!ÿÀŽ äx!ÿ   !  Æ &r   ÿ   âO x!!ÿÀŽ ä x"ÿÀŽ Æ&r!      â� x!#ÿÀŽ ä x#ÿ   #  â &r       ä &r!   ÿ   â x#ÿÀŽ ä x"ÿ   "  ä x!$ÿÀŽ â &r      â x$ÿ   $  â 2r%%        â x""ÿÀŽ â &r!!       â x$$ÿÀŽ Ä 0rÿ%  0 A   ä &r"      â2r        â &r$$   !   Æ Er        â 0rÿ  0 A   â  x"   ?  A  â 0r!ÿ%    A   â Er$ $       â 0rÿ    A   â  x   ?  A  ä #r!!   "   ä #r"$      â G‰  �    € ê r   ÿàó Ê $x   Ž  â r        È �y#    ¢ r        Æ �y$    ä %x   Ž Ì �y     " x%#ÿÀŽ äO x#ÿ   #  Æ &r%%   ÿ   â x#ÿÀŽ ä x#$ÿÀŽ ä� x$ÿ   $  â &r%   %   â 2r        äx$$ÿÀŽ â &r##    %   Ä 0rÿ  0 A   ä&r#$   #   ä 0rÿ    A   â  x   ?  A  ä Er #       Ê #r      È !r       â !r!       â !r"       â G‰  �    € ê r   ÿàó È r   ÿäÿ  â$rÿÿ    Ž â r        Æ �y!    ¤ %x	   Ž ä �y"    è �y     " x#!ÿÀŽ äO x!ÿ   !  Æ &r##   ÿ   â x!!ÿÀŽ ä x"ÿÀŽ â� 2r        âx"ÿ   "  â &r!!   #   Æ x""ÿÀŽ â &r    !   ä 0rÿ  0 A   ä&r"      ä 0rÿ    A   â  x   ?  A  ä Er        Ê #r      È !r       â x   ÿàÿ ä x@  ÿà÷ ä |    p`ò ä x@  ÿàõ â $rÿÿ   Ž â x@  ÿàû ä x@  ÿàù Ä rÿ   ÿä ä rÿ   ÿä ä rÿ   ÿäÿ â G™  Àøÿÿÿÿƒ ê Ay       € ê r        ä r        ä Ay        € ê z ÿ `  pPð È z ÿ a   Sð Ú G	  P    € ê ‰      ( ‰      b !r       â !r
       È/ ‰      ( ‰	
      b !r       â !r
	       È/ ‰	 €    ( ‰ €    b !r		       â !r       È/ ‰	 @    ( ‰ @    b !r	       â !r       È/ ‰      ( ‰
      b !r       â !r
       â/ Gy  €    € ê ‰	      ( ‰      h ‰      ¨ ‰
      â !r		       â !r       â/ !r       ÈO ‰      " !r
       Æ� ‰
	      h ‰      ¨ ‰      â !r       â !r
	
       È/ ‰ €    " !r       ÆO ‰
 €    b !r       Æ� ‰ €    ¨ ‰ €    â !r       â !r
       â/ !r       ÈO ‰ @    " !r       Æ� ‰ @    h ‰ @    ¨ ‰ @    â !r	       â !r       È/ ‰	      " !r       ÆO ‰      b !r
       Æ� ‰      ¨ ‰
      â !r	       â !r       â/ !r       âO !r
       ä� y      %   $ r   ÿŽ È z  p  p`ð È x    pDp  Ú M	        € ê z ÿ ^  pPð â $z  v  Ž Æz ÿ _   Sð Ú $ÿ   ÿ Ž È % ^   Ž Ì �	     ¢ x    ÿ   È r  ÿ   p"ò äx     p"ô äx     p"ö äx     p"ø Î          ä z ÿ `  pPò ä "         ä z ÿ a  Sò ä 2         â $B ÿÿ    Ž È !         ÌO G™  @    € ê z ÿ b  pPð È z ÿ c   Sð Ú         Ê % b   Ž Ì �	     ¢ x   ÿ   ä z  h      ä x øÿÿÿp"ò äx üÿÿÿp"ô är ÿ   p"ö äx    p"ø Î         ä x    pBò ä "        â $2ÿÿ    Ž â B        Ê !       ÈO G  p    € ê xÿÿÿÿÿàÿ È x     € È x   ÿ   È ‚{  €     $ xÿ     È Iy €íÿÿÿÿƒ ê  x;ª¸¿  A  Ì s        $ !t  €?    Ì s        $  r     A  È  r       A  â Gy  P    € ê  x'7=  A  â x*BL?  A  Æ #t  €?   â  x   ?  A  Æ  r     A  Ì s     $   $ !t  €?    È  r     A  È  r       A  â Gy  °    € ê x    pRð Ú G‰  à     € ê x    pRð Ú G	        € ê 	z i    � ä 	z   i    � Æ  x;ª¸¿  A  â 	z   i €  � Ê s        $ !t  €?    Ì s        $  r     A  È  r       A  â Gy  Ð     € ê  r       A  â Gy  °     € ê 	x  à@  � ä 	x    à@  � Æ  x#ÛÙ?  A  â 	x    àÀ  � Æ  x;ª¸¿  A  ä !t   €?    È s        $ !t  €?    Ì s        $  r      A  È  r       A  ä y      "   ( y      &   b $z s  Ž â x        Æ $z v  Ž È/ %v j   Ž Ê †y      â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y       "   " ¹z  l      â Ey  à    € â ™x?    â y      !   b ¹z  F   
   â x      ÿ  â x      ÿ  â y      %   ¢ r& ÿ       â x
      ÿ  â r) ÿ       Ä r= ÿ       ä r@ ÿ       ä rF ÿ       ä r	 ÿ       ä r ÿ       ä r ÿ       Ä r  ÿ   pRð är  ÿ   pRò är    ÿ(Ž ä/ rI ÿ       ä r   ÿ� ä r ÿ       Ä xÿ     ä x   ÿ   äO |    pbô Ú G)  @    € ê y      '   " r ÿ       ä r ÿ       â y      &   b xþÿÿÿÿÀŽ ä z  X      ä r  €ÿàÿ ä z  Y      Ä x   ÿ   äx,   ÿ   ä xÿ     ä x#   ,áÿ â%x,,   ÿŽ â x     ä z$ q  ÿàÿ â 'z y   Ž â x   ÿ   Æ 'z s   Ž â/ zÿ z    Æ $z p  Ž â zÿ t    â $z }  ÿŽ ä $z |  Ž ä$z |  Ž â x   ÿàÿ â $z w  ÿŽ ä $z v  Ž Ä $z v  Ž â r   ÿàõ â $z p  Ž â z  q      â %x.   ,Ž â xÿ   #  ä r ÿ       â $x   Ž â x!#     â $x   Ž âr   ÿŽ â$x    Ž âr   ÿŽ â$x"   Ž â rÿÿ   ÿä â %x,   ,Ž â x##   ÿ   Æ r   ÿàõ â %x*   Ž â x2 $       Ä rÿ   ÿä ä rN,   ÿàù är>.   ÿà÷ ä rC$   ÿàû â %v Z  2 Ž â rO-   ÿä â�y%*    • ¢  r?/   ÿäÿ â $x'$   ÿŽ â rÿ   ÿäÿ â �y(N   • â rJ   ÿàõ â %vBC Z  2 Ž â r'   ÿàÿ â �y9>   • d	 $x1$   ÿŽ Ä �y+N   • â rK   ÿä Æ �y<>   • b	 rDB   ÿà÷ ä rCC1   ÿàÿ â �y6J   ™ ¢  rL#   ÿàõ ä rEC   ÿäÿ â �y:N   • ¢ rM!   ÿä Æ �y5N   • ¢ r;   ÿàõ Æ �y8>   • ¨ �yQ>   • ¨ �yJ   ™ ¢  r0#B   ÿà÷ Æ �y*D   ™ ¢  rÿ   ÿä ä r1C!   ÿäÿ â �yRL   ™ ¢  r3   ÿàõ Æ �y'    ™ ¨  �yOD   ™ ¨" �yUJ   ™ ¢ %v; Z  2 Ž Æ �y70   ™ ¢  $x;$   ÿŽ â rÿ   ÿä ä �ySD   ™ ¢ %x>   Ž âr;   ÿàÿ ä �yTB    ™ ¢ %v03 Z  2 Ž â rV#   ÿà÷ Ä �yN>    • ¢  rD   ÿàõ â/ $x3$   ÿŽ ä �yH    ™ ¢ rE   ÿä ä r113   ÿàÿ ä r>0   ÿàõ â �yD   ™ ¢  rW!   ÿäÿ Ä r?1   ÿä â �yPD   ™ ¢  r3   ÿàõ ä rB#0   ÿàù â�yMV   ™ ¢ r4ÿ   ÿä ä rC1!   ÿä â �y0    ™ ¢  rY   ÿà÷ Æ �yG>   ™ ¨ �yKD   ™ ¢ %v03 Z  2 Ž â rÿ   ÿäÿ ä/ �y>   ™ ¢  $x34$   ÿŽ â r\0   ÿàõ Ä �yLB   ™ ¤ r113   ÿàÿ â %vXY Z  2 Ž â �yJ>   ™ ¤  r]1   ÿä â $x;$   ÿŽ â rVX   ÿà÷ â�y40    ™ ¢ r^#0   ÿàõ Ä rYY;   ÿàÿ â �y\   ™ ¢  r_1!   ÿä ä rWY   ÿäÿ â �yE\   ™ ¢ ra    ÿàõ är3"   ÿà÷ â �yA^   ™ ¢ rZ#X   ÿàù Ä rBÿ   ÿä â/�y?V   ™ ¢ r`ÿ   ÿäÿ â %v0a Z  2 Ž â	r[Y!   ÿä â �y;\   ™ $ $x_B$   ÿŽ ä �yDV   ™ " %v23 Z  2 Ž Æ �yCZ   ™ " $x]`$   ÿŽ â/ r`0   ÿàõ ä �y>X    ™ " r11_   ÿàÿ ä r^2   ÿà÷ â �yBV   ™ " r33]   ÿàÿ ä ra1   ÿä Ä r_3   ÿäÿ ä rd#0   ÿàõ â�yX`   ™ "/ rb#2   ÿà÷ Æ �yV`   ™ " re1!   ÿä Æ �yW^   ™ " rc3!   ÿäÿ Æ �yY^   ™ ( �yZd   ™ ( �y[b   ™ ( �y\`   ™ ( �y]^   ™ ( �y22    ™ ( �y00    ™ " x    ÿàÿ Ä x((T  +    ä� x99T  <    äxg(ÿÀŽ ä x(ÿ   (  ä x<ÿ   9  ä x+9ÿÀŽ â &r_g6   ÿ   âO x((ÿÀŽ Ä x9<ÿÀŽ â &r6+6   ÿ   â x5:T  5    ä x88T  Q    â &r_(   _   ä&r19   6   â/ x5ÿÀŽ â &r6+*   ÿ   âx5ÿ   5  â &r3g*   ÿ   â x*8ÿÀŽ Ä x8ÿ   8  â &r_R   _   âx:5ÿÀŽ â &rR*R   1   â x88ÿÀŽ â &r69O   6   ä&r3(O   3   ä &r1:U   _   ä&rU8U   R   Ä &r6*7   6   ä 0r5ÿ'  0 A   âEr< 1       â &r77   3   â ErU U       â &r38S   6   â  x65  €@  A  â 0r'ÿ'    A   Ê #r<'<   6  â#rU'U   6  â 0r6ÿ%    A   ä &rS:S   7   ä &r%+   ÿ   ä&r'g   ÿ   ä &r%9P   %   Ä 0rNÿN    A   ä &r'(P   '   ä 0r7ÿT  0 A   ä&r%*M   %   â Er3 3       â &r'M   '   â ErR S       â 0rTÿT    A   â #rIN<   I   â  x17  €@  A  â &r<+G   ÿ   Ä 0rÿH  0 A   ä&rGgG   ÿ   ä &r%8K   %   ä&r':K   '   â #r3T3   1  â#rRTR   1  â &r19   <   â Er% %       â &rG(   G   â  x  €@  A  â 0rHÿH    A   â Er '       â &r1*L   1   Ä &rGL   G   â #r%H%     â#rH     â &r18J   1   ä&rG:J   G   â #rFN   F   â 0r'ÿ  0 A   ä&r<+   ÿ   ä&rg   ÿ   â Er1 1       â 0rÿ    A   â Er G       â  x''  €@  A  â &r<9E   <   Ä &r(E      â #r11   '  â#r   '  â &r<*A   <   ä &r+?   ÿ   ä &rAA      ä 0rÿ4  0 A   ä &r<8;   <   Ä&r9D      ä &rA:;   A   â #r	6%   	   â  x%  €@  A  â &r'*C      â Er <       â 0r4ÿ4    A   â Er A       â 0rÿ>  0 A   Æ #r4   %  â#r%4   %  â #r@61   @   â  x1  €@  A  â #r=N   =   â#r
N%   
   â &r+V   ÿ   ä &r?g?   ÿ   ä &r+W   ÿ   Ä &r%gV   ÿ   ä&rWgW   ÿ   â #r6      â &r'8B   '   ä &r9X      ä&r9Y      ä &r?(D   ?   Ä &r%(X   %   ä&rW(Y   W   â Er' '       â 0r>ÿ>    A   ä &r*Z      ä&r*[      ä &r?C   ?   Ä &r%Z   %   ä&rW[   W   â #r'>'   1  â &r8\      â |    pbô â &r8]      ä 0rÿ2  0 A   ä &r?:B   ?   Ä &r\:\   %   ä0rÿ0  0 A   ä&rW:]   W   â #r6'      â 0r0ÿ0    A   â  x+  €@  A  â 0r2ÿ2    A   â Er        â  x%  €@  A  â Er'        Ä Er ?       ä Er\ \       ä Er W       â #r0   %  â #r2'   +  â #r>   1  â #r%0\   %  â #r+2   +  â x@  ÿà÷ â #r6U      â#r63      â #rNR      â#r)N   )   â #r&6   &   â #rN%      â #r6      â #rN+      â x    ÿàÿ Ä x    ÿàÿ ä x    ÿàÿ ä x    ÿàÿ ä x      ÿàÿ ä x""    ÿàÿ ä x$$    ÿàÿ Ä rÿ   ÿäÿ â D)       À â Gy  €íÿÿÿÿƒ ê Ay        € ê Ey  0    € â G‰      € ê r     ÿHŽ Ê ˆs   øÿ H   è ˆs  I€øÿ H   è ˆs   ùÿ H   è ˆs  €ùÿ H   è ˆs  	 úÿ H   è ˆs  F€úÿ H   è ˆs  @ ûÿ H   è ˆs  =€ûÿ H   è ˆs   üÿ H   è ˆs  
€üÿ H   è ˆs   ýÿ H   è ˆs  )€ýÿ H   è ˆs  & þÿ H   è ˆs  €þÿ H   è ˆs   ÿÿ H   è ˆs  €ÿÿ H   ä Ay        € ê {          ì M        € ê „y      H   " z  r  p`ð Æ „y €   H   b r ÿ   p"p Æ „y €  H   ¨ „y    H   è „y €  H   ( „y    H   h y      &   ¨ y      '   è „y €  H   h „y    H   â !r         Æ „y% €  H   â !rII       Æ/ „y    H   â !r       ÆO ‰       " !r	       Æ� ‰I      b !rFF       Æ„y    H   ¢ !r@@       Æ„y  €  H   â z  r      Æ „y'    H   â r   ÿŽ Æ „y+ €  H   â !r==       â !r        ä „y.    H   â $z  x  Ž â x   ÿàÿ â !rI       â/ ‰      " $z  ~   Ž Æ ‰      d 
	  r  ÿàó â!r       â ‰      "     ÿàù ä $ÿÿ   ÿäÿ  â ‰      b 
	 j  ÿ„ äÿÿ   ÿä â ‰      b 
		 k  $ â !r$%
       â !r(       âO ‰!F      ¢ 
 j  ÿˆ â $x   ÿŽ â !r* )       â� ‰#@      â !r       â r    ÿ† â !r&'&       â „y5 €  H   â !r       â/ r"ÿÿ   ÿäÿ â !r,+       â z j  ÿ† â ‰ €    " !r       â
 k   â !r       âz k  "� â !r"       â !r..       â ‰ =      b !r       â r   ÿŽ Ä ‰'"      " !rFF!       âO r    ÿ† ä ‰)$      ¢ !r@@#       â� rÿÿ   ÿäÿ ä ‰      â z
 j  ÿ† ä z  r  p`ò â ‰+(      b !r       â     ÿàù Ä ‰-*      " z k  � ä ÿÿ   ÿä â ‰      " !r==        â/ 
 j  ÿˆ ä ‰/&      b !r'"'       âx    p"ò â !r55       â ‰1,      " !r)$)       âO 
 k   Ä ‰3.      ¢ !r       â� r    ÿàõ ä ‰ €    â !r+(+       ârÿÿ   ÿä ä ‰!      b !r-*-       â    ÿàÿ Ä ‰#F      " !r       â z j  ÿ„ ä‰%@      " !r/&/       â/     ÿà÷ ä ‰ @    b !r1,1       âz k   â !r3.3       âO ‰(=      ¢    ÿàÿ â !r       Ä� ‰*'      â r  €ÿŽ â!r!!       ä‰,)      "    ÿàÿ â !r#F#       ä ‰ €    "     ÿàõ â !r%@%       Ä ‰.+      b ÿÿ   ÿäÿ ä  j  ÿ† â ‰0-      b !r       â/     ÿàù ä ‰  €    b ÿÿ   ÿä â !r(=(       âO 
 j  ÿ„ â ‰2/      ¢  k  � â !r*'*       â� 8ÿÿ   ÿä â ‰41      â !r,),       â j  ÿˆ Ä ‰63      " !r       â 
 k   ä ‰5      " !r.+.       â k  8 ä ‰ @    b !r0-0       â     ÿàõ Ä ‰"! €    b !r         Æ/ ‰$# €    b !r2/2       ÆO ‰&% €    ¢ !r414       â� !r636       ä‰'( €    â !r75       Æ ‰)* €    " !r       Æ‰+, €    " !r!"       Æ ‰ @    b !r#$       Æ/ ‰-. €    b $ÿÿ   ÿä â !r%&       äO ‰/0 €    ¢  j  ÿ„ Æ ‰!  @    ¢ !r'('       â�  k  $ ä ‰12 €    â !r)*)       Æ ‰34 €    " !r+,+       Æ‰56 €    " !r       Æ‰87      b !r-.-       Æ/ ‰      b !r/0/       ÆO ‰" @    ¢ !r! !       Æ ‰# @    ¢ !r121       Æ� ‰% @    â !r343       Æ ‰      " !r565       Æ‰ ' @    " !r878       â!r;       ä/ ‰$) @    b !r""       ÆO ‰&+ @    ¢ !r##       Æ ‰      â !r%%       Æ� ‰(- @    â !r9       Æ ‰*/ @    " x        â !r '        ä‰!      $ %v  j   Ž ä ‰,1 @    b !r$)$       Æ/ ‰.3 @    b !r&+&       ÆO ‰78 €    ¨ ‰05 @    ¢ !r(-(       Æ� ‰"      â !r*/*       Æ ‰#      " !r'!       Æ†	 9    â	 !r,1,       Æ‰%      b !r.3.       Æ/ † ;   â !r787       âO !r9       ä‰       b !r050       â !r)"       ä� ‰$      ¢ !r+#       Æ ‰&      ( †	 9    â !r%%       Æ‰!(      è ‰	*      ( † '    â	 !r        Æ/ ‰,      b !r$       ÆO ‰#.      ¢ !r&       Æ ‰7 @    ( ‰0      b !r!(!       Æ� †	 )    â !r'*	       Æ† +   è †	 %    â !r,       â/ !r#.#       äO †     â !r7       Æ †	 
    â !r0       Æ† 
   è †	 !    è † '    è †	     è † #   è ‰	      ( †	     â M™        € ê x   ÿàÿ â !r		       Æ r    ÿàñ È r ÿÿ   ÿä  ä z j  ÿ€ È� z k     Ê †y 	    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y       "   " ¹z  l      â Ey  Ð    € â ™x?    â y      !   b ¹z  F   
   â r ÿ       â x
      ÿ  â y      %   ¢ r ÿ       ä r! ÿ       Ä r# ÿ       ä r' ÿ       ä r$ ÿ       ä rC ÿ       ä rE ÿ       ä rH ÿ       Ä r  ÿ   pRð är  ÿ   pRò är    ÿ(Ž ä/ r ÿ       ä r   ÿ� ä rG ÿ       Ä xÿ     ä rK ÿ       ä |    pbô ä x   ÿ   ÖO G)  0    € ê y      '   " r ÿ       ä r ÿ       â y	      &   b xþÿÿÿÿÀŽ ä z  X      ä r  €ÿàÿ ä z  Y      Ä x   ÿ   äx(   ÿ   ä r ÿ       ä x   (áÿ â %x((   ÿŽ â z q  ÿàÿ Æ 'z y   Ž È 'z	 s   Ž â/ zÿ z    Æ $z p  Ž â zÿ t    â $z }  ÿŽ ä $z |  Ž ä$z |  Ž â x   ÿàÿ â $z v  Ž ä $z	 w  ÿŽ Ä $z v  Ž â xÿ     â $z p  Ž â z  q      â %x*   (Ž â xÿ     ä r   ÿàõ â $x   Ž â x     â $x   Ž â x     â $x   Ž âr   ÿŽ â%x(   (Ž â r   ÿŽ Ä x   ÿ   ä rÿÿ   ÿä ä x   ÿ   ä r,(   ÿà÷ ä rN   ÿàõ ä r-)   ÿäÿ ä xS $       Ä rÿ   ÿä â �y,   • ¢  r2*   ÿà÷ â %vNN Z  S Ž ä �y8,   • ¢  r3+   ÿäÿ â $x	$   ÿŽ â r0   ÿà÷ â �yB,   • â  r<N   ÿàõ Ä rOO	   ÿàÿ â �y?2   • " rÿ   ÿäÿ ä r   ÿàù â %v00 Z  S Ž â r=O   ÿä â �y2   • " r	ÿ   ÿä â $x%$   ÿŽ â r.N   ÿàõ â �y9,   • â  rL0   ÿà÷ â $x		$   ÿŽ â r/O   ÿä â �yA2   • b r11%   ÿàÿ Ä rP0   ÿàõ â �y2   • b %v, Z  S Ž â rM1   ÿäÿ ä�y:<   ™ b rQ1   ÿä ä r--	   ÿàÿ â �y<   ™ b r2,   ÿàõ Ä/ r6,   ÿà÷ â �yDL   ™ b r3-   ÿä Æ �y	N    ™ b r7-   ÿäÿ Æ �y;.   ™ h �y>L   ™ b rN   ÿàõ Æ �yJL   ™ b rOÿ   ÿä Æ �yF6   ™ b %x4   Ž Æ �y=<   ™ b rL   ÿà÷ â/ %x.   Ž ä �y@P   ™ b rMÿ   ÿäÿ â $xOO$   ÿŽ ä �y&0    ™ h �y%6   ™ b $xMM$   ÿŽ Æ �y4    • b %v0N Z  S Ž Æ �y".    • d r11O   ÿàÿ ä �y<2   ™ b %v4L Z  S Ž Æ/ �yI6   ™ b r.0   ÿàõ ä r55M   ÿàÿ â �y ,    ™ b r/1   ÿä är24   ÿàù â �yU0    ™ b rP0   ÿà÷ ä r35   ÿä â �yL.   ™ b rQ1   ÿäÿ Ä r,4   ÿàû â �yN.   ™ b r6   ÿàõ ä/rX   ÿà÷ â �yM2   ™ b r-5   ÿäÿ ä rZÿ   ÿä â�yPP   ™ b rYÿ   ÿäÿ Æ �yO2   ™ b %v66 Z  S Ž Æ �yR,   ™ b $xYY$   ÿŽ ä $xQZ$   ÿŽ â �yT.   ™ h �yV2   ™ b %v,X Z  S Ž â/ r77Q   ÿàÿ Ä �yW4    ™ b r.6   ÿàõ är0,   ÿàù är--Y   ÿàÿ ä r/7   ÿä är1-   ÿä ä r4,   ÿàõ â/�yQ.   ™ b r26   ÿà÷ Æ �yS0   ™ b r5-   ÿä är37   ÿäÿ â �yX0   ™ h �yY.   ™ h �y44   ™ h �y22   ™ h �y[0   ™ h �y66    ™ h �yZ.   ™ h �y,,    ™ b x    ÿàÿ È |    pbô ä x@  ÿà÷ ä x    ÿàÿ ä x    ÿàÿ ä x    ÿàÿ ä x    ÿàÿ Ä x    ÿàÿ ä x    ÿàÿ ä rÿ   ÿäÿ ä x8T  8    ÈO x8ÿÀŽ ä x3ÿ   8  ä x?T      Èx5ÿ     ä x/ÿÀŽ ä x33ÿÀŽ ä xBBT  9    ä� xAT      äx15ÿÀŽ â/ &r:   ÿ   Ä &r:/:   ÿ   â x5ÿ   B  â &r3      â x9ÿ     ä x-BÿÀŽ ä x7ÿÀŽ â &r/D   ÿ   ä &r:1   :   â x55ÿÀŽ Ä x99ÿÀŽ â &r-;      ä&r:7;   :   ä &r1>      ä 0rÿ	  0 A   ä &r0/F   ÿ   ä&rFF   ÿ   Ä &r5=      ä&r=9=   :   ä &r7@      ä 0r	ÿ	    A   ä &r.9J      â  x  €@  A  â Er =       â &r01%   0   Ä &rF3%   F   ä 0r%ÿ&  0 A   â Er        â 0rÿ    A   â #r	     â 0r&ÿ&    A   â  x;%  €@  A  â Er% .       â &r07<   0   ä&rF-<   F   â #r	     â &r09I   0   Ä &rF5I   F   â #r      â 0r	ÿ   0 A   â #r&%   ;  â 0r"ÿ"    A   â Er% 0       â 0r ÿ     A   â ErF F       â &r/L   ÿ   â  x		  €@  A  â #rG   G   Ä &r1N      â #rK"   K   â #r% %   	  â&r/M   ÿ   â#r F   	  â &r M   ÿ   ä &r7P      ä &r1O      ä &rDD   ÿ   Ä &rLL   ÿ   ä &r 3O       ä &r9T      ä &r7R      ä 0r	ÿU  0 A   â Er        â &rD3>   D   Ä &rL3N   L   ä &rR-R       ä &r 9V      â  x	  €@  A  â 0rUÿU    A   ä &rD-@   D   ä&rL-P   L   â #rU     â &rD5J   D   â#rE"   E   â &rL5T   L   â #rC   C   â &r/Q   ÿ   Ä &rQ   ÿ   ä &r//S   ÿ   âErD D       â &rS   ÿ   â #rH%   H   â Er	 L       â 0r%ÿW  0 A   ä &r/1X   /   ä&r1Y      Ä &r3Y      ä&r3X      â #rD&D   ;  â 0rWÿW    A   â Er          â #r	U	     â  x&%  €@  A  â &r74   /   ä&r72      Ä &r-2      ä&r4-4      â #r%W    &  â &r9[      ä 0rÿ6  0 A   ä &rR5V   R   ä &r9Z      Ä &r5Z      ä&r45[   4   ä 0r ÿ,  0 A   â#r
"	   
   â #r$%   $   â 0r,ÿ,    A   â  x	  €@  A  â Er%        â 0r6ÿ6    A   â Er; R       â  x-   €@  A  â Er        Ä Er        ä Er4 4       â #r,%   -  â #r;W;   &  â #r6   	  â#r6   	  â #r-,4   -  â #r"D      â#r'";   '   â #r#   #   â #r!"   !   â #r      â #r"-      â D)       À â Gy  €ïÿÿÿÿƒ ê Ay        € ê Ey      € â G‰  ð     € ê $x  À  Ž Ê ˆs   ùÿ H   è ˆs  K€ùÿ H   è ˆs  G úÿ H   è ˆs  €úÿ H   è ˆs  H ûÿ H   è ˆs  E€ûÿ H   è ˆs  C üÿ H   è ˆs  
€üÿ H   è ˆs  $ ýÿ H   è ˆs  '€ýÿ H   è ˆs  # þÿ H   è ˆs  !€þÿ H   è ˆs   ÿÿ H   è ˆs  €ÿÿ H   ä Ay        € ê {          ì M        € ê „y      H   " z  r  p`ð Ä z-  r      â „y €   H   b r ÿ   p"p ä r+--   ÿŽ â „y €  H   ¨ „y    H   è „y    H   ( „y €  H   h „y	    H   h „y €  H   h „y    H   è „y €  H   â !r         Æ „y    H   â !rKK       Æ/ ‰       " !r       ÆO „y €  H   â !rCC       Æ� „y    H   â !rGG       Æ„y €  H   â !rEE       Æ‰K      b !r		H       Æ y      '   ¢ !r
       â !r        ä ‰      â !r$$       Æ y       &   " !r''       Æ ‰      â !r##       Æ ‰C      " !rK       Æ/ ‰G      b !r!!       Æ ‰E      b !r       Æ ‰	      b !r"       â $z   x  Ž ä ‰      " !r       â� $z  ~   Ž âO !r       â x   ÿàÿ â !rC       â‰$      ¢ 
  r  ÿàõ â !rG       Ä/ ‰'      b z  r  p`ò â !rE       ä‰ €    â ÿÿ   ÿä â !r	       ä ‰#      " 
 j  ÿ„ ä x    p"ò â ‰      b !r       â 
 k   Ä ‰!      ( ‰       " !r$$       ÆO ‰      ¢ !r''       Æ/ ‰)"      b !r       Æ� ‰	      â !r##       Æ‰      " !r       âr-    ÿ„ ä ‰      b !r!!       Æ ‰
      " !r%        â !r       äO ‰       ¢ r--   ÿŽ â !r)")       ä/ ‰$      b     ÿà÷ â !r	       Ä� ‰"'      â 	 +   ÿàû â !r       ä‰&#      " rÿÿ   ÿä â !r       ä‰ @    b ÿÿ   ÿäÿ â !r
       Ä ‰ €    " r-    ÿˆ ä 
 j  ÿ† â ‰(!      " z j  ÿ„ â !r         âO ÿÿ   ÿäÿ â ‰*%      ¢ r.ÿÿ   ÿä â !r$$       â/ 
	 j  ÿŠ â ‰ €    b 
 k  � â !r"'"       â� z
 j  ÿˆ â ‰ €    â z k   â !r&#&       â
		 k  � â ‰,)      " !r       â-   ÿàÿ Ä ‰ €    b !r       â z k  . ä ‰ €    " !r(!(       â +   ÿàÿ ä ‰ €    " !r*%*       âO /    ÿàù â !r       â/ ‰#  €    d 2ÿÿ   ÿä â !r       â� ‰%$ €    ¢ $x-   ÿŽ â / j  ÿˆ â !r,),       â‰'" €    â 0    ÿà÷ â !r       Ä‰)& €    " / k  2 â !r       ä ‰ @    " r    ÿàõ â !r       ä ‰+( €    b r.ÿÿ   ÿä ä z j  ÿ„ â ‰-* €    b !r# #       â/ 1ÿÿ   ÿäÿ Ä ‰! @    b !r%$%       âO 0 j  ÿ† ä ‰ @    ¢ !r'"'       â� z k  . ä ‰ @    â !r)&)       â0 k  1� Ä ‰/, €    " !r       â    ÿàÿ ä ‰ @    " !r+(+       â    ÿàõ ä ‰      b !r-*-       Æ ‰ @    b !r!!       Æ/ ‰ # @    b !r       ÆO ‰% @    ¢ !r       Æ� ‰"' @    â !r/,/       â,ÿÿ   ÿä ä ‰$) @    " !r       Æ ‰      " !r1       âx        ä ‰&+ @    b !r       â  j  ÿ„ â %v  j   Ž â ‰(- @    b !r #        â/  k  , Ä ‰      b !r%       ÆO ‰*/ @    ¢ !r"'"       Æ� ‰      â !r$)$       Æ‰       " !r3       Æ ‰      " !r&+&       Æ‰!      b !r(-(       Æ ‰       b !r)       Æ/ ‰#      b !r*/*       ÆO ‰%"      ¢ !r+       Æ� ‰$      â !r        Æ†	 1    â !r       Æ † 3   â !r!!       Æ‰'&      " !r        Æ ‰(      " !r##       Æ/ ‰*      b !r%"%       ÆO †	 )    â !r$       Æ� † +    è †	 
    è † 
   â !r'&'       Æ †	 !    â !r(       Æ†     â !r*       Æ/ †	 #    è † %   è †	     è † '    è †	     â M™        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y       "   " ¹z  l      â Ey  °    € â ™x?    â y      !   b ¹z  F   
   â r ÿ       â x      ÿ  â y      %   ¢ r  ÿ       ä r# ÿ       Ä r- ÿ       ä r ÿ       ä r
 ÿ       ä r9 ÿ       ä r; ÿ       ä r ÿ       Ä r  ÿ   pRð är  ÿ   pRò är    ÿ(Ž ä/ r> ÿ       ä r   ÿ� ä x   ÿ   ÄO xÿ     È |    pbô Ú G)  0    € ê y	      '   " r ÿ       â $z p  Ž â r ÿ       â y      &   b xþÿÿÿÿÀŽ ä z  Y      ä r  €ÿàÿ ä z  X      Ä x   ÿ   äx   ÿ   ä z q  ÿàÿ ä x   áÿ â %x&   ÿŽ Æ xÿ     â 'z	 y   Ž Æ x     â'z s   Ž â/ zÿ z    ä x   ÿ   â $z	 }  ÿŽ â zÿ t    â $z	 |  Ž ä$z w  ÿŽ ä $z |  Ž â x   ÿàÿ â $z	 v  	Ž âr   ÿàõ â $z v  Ž â z  q      â $z p  	Ž â x	ÿ     â %x$   &Ž â r ÿ       Ä x   	  â $x   Ž âr   ÿŽ â$x   Ž âr   ÿŽ â %x&   &Ž â x   ÿ   ä rÿÿ   ÿä ä rK   ÿàõ ä x* $       Ä rÿ   ÿä ä r=   ÿàû â %vJK Z  * Ž â rN&   ÿàù ärL$   ÿà÷ â $x	$   ÿŽ â rÿ   ÿäÿ ä rO'   ÿä â%v<= Z  * Ž â rM%   ÿäÿ Ä rHJ   ÿàõ â$x$   ÿŽ â rKK	   ÿàÿ â �y,N   • ¢  r(        ä rIK   ÿä â �y/N   • ¢  r@<   ÿàõ ä r==   ÿàÿ â �y5L   • â r)        Æ �yL   • â rA=   ÿä â %xD   (Ž â rBJ   ÿàõ â�y8N   • " rF<   ÿà÷ Æ �y1N   • ( �y7L   • h �yL   • b rCK   ÿä Æ �y3H   ™ b r   ÿàõ Æ �y.@   ™ b rG=   ÿäÿ Æ �y"H   ™ b r+ÿ   ÿä Æ �yDD    • h �yCB   ™ h �y4F   ™ h �yE@   ™ b $x++$   ÿŽ Æ �y!J    ™ h �y6H   ™ b r	   ÿà÷ Æ �y0@   ™ b %vJ Z  * Ž Æ �y<    ™ b rÿ   ÿäÿ ä rLJ   ÿàõ ä/ rKK+   ÿàÿ â %x(   (Ž Æ rMK   ÿä â %vF	 Z  * Ž â rPJ   ÿà÷ â �y2(    • d $x$   ÿŽ ä �y=L   ™ b rHF   ÿàù ä rGG   ÿàÿ â �y:L   ™ b rQK   ÿäÿ Ä r)   ÿàõ â/ �y@L   ™ b rNF   ÿàû ä r+   ÿà÷ â �yBP   ™ b rIG   ÿä ärSÿ   ÿä â �y<J    ™ b rOG   ÿäÿ Ä rRÿ   ÿäÿ â �y	H   ™ b %v() Z  * Ž Æ �yAN   ™ b $xSS$   ÿŽ ä %v*+ Z  * Ž â �yH   ™ b rP(   ÿàõ ä/ r))S   ÿàÿ â �yH   ™ b $xOR$   ÿŽ â rN*   ÿàù Ä �y?F    ™ b rQ)   ÿä ä r++O   ÿàÿ ä rT(   ÿà÷ â�yHP   ™ b# rO+   ÿä ä rR*   ÿàõ â �yFP   ™ b rU)   ÿäÿ Æ �yGN   ™ b rS+   ÿä Æ �yIN   ™ h �yJT   ™ h �yKR   ™ h �yLP   ™ h �yMN   ™ h �y**    ™ h �y((    ™ b x    ÿàÿ È |    pbô ä x@  ÿà÷ ä x    ÿàÿ ä x    ÿàÿ ä x    ÿàÿ ä x    ÿàÿ Ä x    ÿàÿ ä rÿ   ÿäÿ ä x,,T  /    ÈO x/ÿ   ,  ä x55T      ä� x,ÿÀŽ ä xRÿ   5  ä x,5ÿÀŽ ä x18T  1    Äx//ÿÀŽ ä x5RÿÀŽ ä x7T      äx)1ÿÀŽ â &rN3   ÿ   âx+ÿÀŽ â &r3,3   ÿ   âx1ÿ   1  â &r8,.   ÿ   âxÿ     â &r..   ÿ   Ä &rN/"   N   ä&r"5"   3   â x11ÿÀŽ ä x3ÿÀŽ â &r)C   N   ä&r"+C   "   ä &r85E   8   Ä &r./E   .   ä &r8+4   8   ä&r.)4   .   ä &r16      ä&r636   "   ä&r"30   8   Ä &r010   .   ä 0r.ÿ  0 A   âEr" "       â 0rÿ    A   â ErQ 0       ä/  x..  €@  A  È #rO"   .  â#r.Q   .  â &r,=   ÿ   È &r5:      â ErE        Æ &r+B      È &r3@      ä &r,	   ÿ   ä&r"	   ÿ   ä 0r	ÿ<  0 A   âEr        â 0r4ÿ!  0 A   ä 0r<ÿ<    A   â  x		  €@  A  â &r5      Ä &r"/   "   â Er6 6       â 0r!ÿ!    A   â  x44  €@  A  â &r==   ÿ   â #r<   	  â 0rDÿD    A   ä &r+A      ä&r")A   "   â #rC!6   4  â&r=/:   =   â #rE!E   4  â &r3      â #r9D   9   â &r!1   "   Ä 0rÿ?  0 A   ä&r,F   ÿ   ä &r=)B   =   ä &rFF   ÿ   ä&rG   ÿ   â Er"        â &r,,G   ÿ   â  x0  €@  A  â 0r?ÿ?    A   Ä &r=1@   =   ä &rF/H   F   ä&r/I      ä&r,5I   ,   â #r?"   0  â &r5H      â Er =       â &rF)J   F   Ä &r)K      ä&r,+K   ,   â#rD      â &r+J      ä &rF1L   F   â#r	<   	  â 0r7ÿ2    A   ä &r11M      Ä 0rÿ*  0 A   ä &r,3M   ,   ä0rÿ(  0 A   ä &r3L      â #r
7	   
   â  x  €@  A  â 0r*ÿ*    A   â  x	  €@  A  â Er ,       â 0r(ÿ(    A   â Er! !       Ä Er        ä ErF F       ä Er 1       â #r*     â #r?!   0  â #r(   	  â#rF(F   	  â #r*     â #r>DC   >   â#rDO      â#r7E      â#r;7.   ;   â#r-7   -   â#r#D   #   â#r 7F       â#rD      â #r7      â G©  pñÿÿÿÿƒ ê Ay        € ê Ey  ð     € â G‰  Ð     € ê $x €  Ž Ê ˆs > úÿ H   è ˆs €úÿ H   è ˆs  ûÿ H   è ˆs ;€ûÿ H   è ˆs 9 üÿ H   è ˆs 
€üÿ H   è ˆs  ýÿ H   è ˆs -€ýÿ H   è ˆs # þÿ H   è ˆs  €þÿ H   è ˆs  ÿÿ H   è ˆs €ÿÿ H   ä Ay        € ê {          ì M        € ê „y €   H   b z  r  p`ð Ä z%  r      â „y     H   " r ÿ   p"p Æ „y	    H   ¨ „y €  H   è „y €  H   ( „y    H   h „y €  H   h „y    H   h „y    H   h „y €  H   è „y    H   â !r       â/ !r>       Ä „y €  H   â !r		       ÆO ‰      " !r-       Æ� ‰       b !r;       Æ‰      ¢ !r9       â!r

       ä ‰      â !r       Æ ‰	      " !r##       â !r       ä ‰      " !r         â !r         ä/ ‰      b !r       â !r       ÄO ‰
      ¢ !r       Æ ‰      h ‰#      b !r       â� x   ÿàÿ ä ‰       â !r		       âz  r  p`ò Ä ‰      " !r       â x    p"ò ä ‰       " !r       Æ/ y      &   b !r
       âO !r       äy      '   ¢ !r##       Æ ‰	      b !r         Æ� ‰      â !r       Æ‰      " !r        Æ ‰      " $z  x  Ž Æ/ ‰      h ‰ €    b $z  ~   Ž ÆO ‰      ¢ !r	       âr%    ÿˆ ä ‰#      b !r       â� r!ÿÿ   ÿä ä ‰       â !r       Æ‰      " !r       â r%%   ÿŽ ä ‰      " !r       Æ/ ‰ €    b !r
       â 
  r  ÿà÷ â !r       âO ‰ €    ¤ ÿÿ   ÿäÿ â !r##       â‰ €    b 
 j  ÿ† â !r        Ä� ‰ €    â 
 k  � â !r       ä‰ €    " !r       Æ ‰ €    " !r       Æ/ ‰
 @    b     ÿàõ ä z j  ÿˆ â ‰ €    b !r       âO ÿÿ   ÿä ä ‰# €    ¢ !r       â
 j  ÿ„ Ä ‰  €    b !r       â� z k  ! ä ‰" €    â !r       â
 k   ä ‰$      " !r       â r%%   ÿŽ Ä ‰	 @    " !r
       â/ r%    ÿ„ â !r       â ‰ @    b     ÿàù â !r#       äO ‰ @    ¢ &ÿÿ   ÿä â !r         Ä‰      b r(ÿÿ   ÿä â !r""       ä� ‰ @    â z
 j  ÿ„ â !r$$       ä‰ @    "    ÿàÿ â !r	       Ä ‰ @    " 	%   ÿàÿ ä 
 j  ÿˆ â‰! @    " !r       â/ * 	   ÿà÷ ä ‰# @    b !r       âO 
	 k  & Ä ‰%  @    ¢ !r-       â+ÿÿ   ÿäÿ ä ‰'" @    b !r       â� * j  ÿ† ä ‰)$ €    â !r       â* k  +� Ä ‰      " !r       â z k  ( â !r!!       â ‰      "         â !r##       ä/ ‰      b     ÿàõ â%  j   Ž â !r% %       âO ‰      ¢ (ÿÿ   ÿä â !r'"'       Ä‰      b !r)$)       Æ� ‰      â !r+       Æ‰!      "         ä  j  ÿ„ â ‰ #      " !r       â %  j   Ž â  k  ( â ‰"%      " !r       Æ/ ‰&) @    b !r       ÆO ‰$'      ¢ !r       Æ†	 +    â !r       Æ� † -   â !r!!       Æ†	     â !r##        â !r%%"       ä †     â !r&)&       Æ/ †	     â !r''$       ÆO †    è †	     è † !    è †	 
#    è † 
%   è ‰&      ( †	 '    â M™        € ê x   ÿàÿ â !r&       ÆŸ r    ÿàñ È r ÿÿ   ÿä  ä z j  ÿ€ È z k     Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y       "   " ¹z  l      â Ey  p    € â ™x?    â y      !   b ¹z  F   
   â r ÿ       â x      ÿ  â y      %   ¢ r ÿ       â x
      ÿ  â r ÿ       Ä r ÿ       ä r2 ÿ       ä r ÿ       ä r  ÿ   pRð är  ÿ   pRò är    ÿ(Ž È/ r   ÿ� ä x   ÿ   äO xÿ     È |    pbô Ú G)       € ê y      '   " r ÿ       ä r ÿ       â y      &   b xþÿÿÿÿÀŽ ä z  Y      ä r  €ÿàÿ ä z q  ÿàÿ Ä x   ÿ   äx   ÿ   È x   áÿ â %x    ÿŽ Æ xÿ     â 'z y   Ž Æ x     â'z	 s   Ž â/ zÿ z    ä x   ÿ   â $z }  ÿŽ â z	ÿ t  	  â $z p  Ž ä$z |  Ž ä $z w  ÿŽ â x   ÿàÿ â $z |  Ž Ä $z	 v  Ž â r   ÿàõ â $z	 v  Ž â z  q      â $z p  Ž â x	ÿ     â %x"    Ž â z  X      ä x   	  â $x   Ž â r ÿ       â %x     Ž â r   ÿŽ Ä r   ÿŽ ä x   ÿ   ä rÿÿ   ÿä ä r$   ÿàõ ä r,"   ÿà÷ är*    ÿàù ä x; $       Ä rÿ   ÿä ä r-#   ÿäÿ â %v$$ Z  ; Ž â r+!   ÿä ä r>   ÿà÷ â $x	$   ÿŽ â rJ$   ÿàõ â �y'*   • ¢  rÿ   ÿäÿ Ä r%%	   ÿàÿ â �y(*   • ¢  %v>> Z  ; Ž Æ �y/,   • â $x$   ÿŽ Æ �y,   • â r<   ÿàù ä rK%   ÿä â �y1*   • " r	ÿ   ÿä ä rF>   ÿà÷ â �y8*   • " r??   ÿàÿ Æ �y)J   ™ " rG?   ÿäÿ â %v<< Z  ; Ž ä �y.,   • " r*$   ÿàõ Æ �y9,   • " $x		$   ÿŽ â r+%   ÿä ä �yJ   ™ " rH>   ÿàõ ä r@<   ÿà÷ â �y5F   ™ " r==	   ÿàÿ ä rI?   ÿä â �y**   ™ " rA=   ÿäÿ Æ �y-F   ™ "! rB<   ÿàõ Æ �y,J   ™ ( �y0H   ™ " rC=   ÿä Æ �y6$    ™ ( �y:@   ™ " rJ   ÿàõ Æ�y7F   ™ " r$        ä r%        â �y3@   ™ ( �y&>    ™ " rF   ÿà÷ â%xD   $Ž â rGÿ   ÿä â�y+B   ™ " rHÿ   ÿäÿ Æ/ �y	D    • " $xGG$   ÿŽ ä %x>   $Ž â �y<    ™ & %v$F Z  ; Ž â�y4@   ™ & %vDJ Z  ; Ž È/ $x=H$   ÿŽ â rHD   ÿàõ â�y;>    • " rEEG   ÿàÿ ä rF$   ÿàù ä r%%=   ÿàÿ ä rIE   ÿä ärG%   ÿä Ä rLD   ÿà÷ â�y<H   ™ " rJ$   ÿàõ Æ �y=F   ™ " rME   ÿäÿ ärK%   ÿä â �y>H   ™ ( �y?F   ™ ( �y@L   ™ (�yAJ   ™ ( �yBH   ™ ( �yCF   ™ ( �yDD    ™ ( �y$$    ™ " x    ÿàÿ È |    pbô ä x@  ÿà÷ ä x    ÿàÿ ä x    ÿàÿ ä x    ÿàÿ ä x    ÿàÿ Ä rÿ   ÿäÿ ä x''T  (    ÈO xLÿ   '  äx//T      ä� x'ÿÀŽ ä x'ÿ   /  ä x(/ÿÀŽ ä xHLÿÀŽ Ä/ xF'ÿÀŽ â &r')   ÿ   â	x11T  8    â &r)()   ÿ   â x9.T  9    ä x.1ÿÀŽ â &r%H   '   âx'ÿ   1  â &rF   )   â x)ÿ   9  Ä x/9ÿÀŽ â &r1(5   ÿ   âx''ÿÀŽ â &r55   ÿ   â x))ÿÀŽ â &r%.*   %   ä&r*/*      ä &r1F-   1   ä&r5H-   5   Ä &r',   %   ä&r,),   *   ä &r*/0   1   ä&r0.0   5   ä 0r-ÿ6  0 A   ä &r1(:   ÿ   âEr        â &r::   ÿ   Ä &r%)7   *   â x--  €@  A  â &r7'7   0   â Er0 ,       â 0r6ÿ6    A   ä &r*F3   1   ä 0r1ÿ&  0 A   â#r6   -  â&r3H3   :   â Er% %       â 0r&ÿ&    A   â #r660   -  â  x11  €@  A  â &r-/+   *   Ä &r3.+   3   â Er, 7       â #r*&%   1  â0r%ÿ  0 A   ä&r-)4   -   ä&r3'4   3   â #r1&,   1  â 0rÿ    A   â Er& -       â  x,%  €@  A  â Er3 3       â 0r;ÿ;    A   Ä #r&&   ,  ä#r,3   ,  â #r2;   2   â &r(<   ÿ   ä&r<   ÿ   ä &r((=   ÿ   ä&r=   ÿ   Ä &rF>      ä&rH>      ä &r(F?   (   ä&rH?      ä &r/@      ä&r.@      Ä &r(/A   (   ä&r..A      ä &r)B      ä&rB'B      ä &r()C   (   ä&r.'C   .   Ä 0rÿD  0 A   ä 0r%ÿ$  0 A   â Er        â 0rDÿD    A   â ErB B       â 0r$ÿ$    A   â Er' (       â  x  €@  A  â Er. .       â  x%%  €@  A  â 0r	ÿ	    A   â #rD     â#rBDB     â #r'$'   %  â#r.$.   %  â #r	6      â#r	*      â #r
;1   
   â #r	&      â #r;,      â #r	      â #r;B      â #r	'      â #r;.      â G©  póÿÿÿÿƒ ê Ay        € ê Ey  Ð     € â G‰  °     € ê $x  @  Ž Ê ˆs   ûÿ H   è ˆs  2€ûÿ H   è ˆs   üÿ H   è ˆs  
€üÿ H   è ˆs   ýÿ H   è ˆs  €ýÿ H   è ˆs   þÿ H   è ˆs  €þÿ H   è ˆs   ÿÿ H   è ˆs  €ÿÿ H   ä Ay        € ê {          ì M        € ê „y      H   " z  r  p`ð Ä z%  r      â „y	 €   H   b r ÿ   p"p ä r!%%   ÿŽ â „y    H   ¢ x#        Æ „y €  H   è „y    H   ( „y €  H   h „y    H   h „y €  H   h „y    H   è „y €  H   â !r         Æ y      '   " !r		2       Æ/ ‰       b !r       ÆO y      &   ¢ !r
       Æ� ‰	      â !r       â!r       È‰      " !r       Æ ‰      b !r       â !r        ä/ ‰      b !r       Æ ‰       " !r       â !r	       ä� ‰
      è ‰      ¢ !r       Æ‰      " !r       Æ‰      h ‰      b !r       â/ !r         ä ‰	      " !r

       Æ� ‰  €    b !r       ÆO ‰      ¢ !r       Æ‰      â !r       Æ‰      " !r       Æ ‰
      h ‰      b !r		       Æ ‰      " !r        â/ $z  x  Ž ä ‰      b !r       âO $z  ~   Ž ä ‰      ¢ !r       â� !r       Ä‰ €    â !r
       Æ‰	 €    " x
   ÿàÿ â !r       ä ‰ @    b z 
 r  p`ò â !r       ä ‰ €    " x    p"ò â !r       Ä/ ‰ €    b r%    ÿ† â !r       äO ‰ €    ¢ rÿÿ   ÿäÿ Æ ‰ €    ¢ !r       â� r%    ÿŠ ä ‰ €    â !r	       â
	  r  ÿàù ä‰ €    " !r
       â !   ÿàõ Ä ‰ €    b !r       â  ÿÿ   ÿä â !r       â/ ‰
      " 
 j  ÿ„ â !r       äO ‰ @    b z j  ÿ† â !r       Ä ‰ @    ¢ 
 k    â !r       ä� ‰ @    â z k  � â !r       ä‰ @    " r"ÿÿ   ÿäÿ â !r       Ä‰ @    b z j  ÿŠ ä $ÿÿ   ÿä â ‰ @    b !r'
       â %   ÿàÿ ä ‰ @    " !r       â/ z k  "� Ä ‰ @    b !r       âO 
	 j  ÿˆ ä ‰ @    ¢ !r       â� !   ÿàÿ â !r       â!    ÿàõ â‰      â     ÿà÷ â !r       â%v  j  # Ž â ‰      " "ÿÿ   ÿä â !r       â ! j  ÿ„ â ‰      b  ÿÿ   ÿäÿ â !r       â 
		 k  $ â ‰      " 
 j  ÿ† â !r       Ä/ ‰      b  k   � â !r       äO ‰      ¨ ‰       ¢ !r#       â� ! k  " Ä ‰      â !r       Æ‰      " !r       â!r       ä †	 '    â !r       Æ/ † #   â !r       ÆO †	     â !r        Æ †     â !r       Æ� †	     â !r       Æ†    è †	     è † 
    è †	     â M™        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y       "   " ¹z  l      â Ey  P    € â ™x?    â y      !   b ¹z  F   
   â r ÿ       â x
      ÿ  â y      %   ¢ r ÿ       ä r ÿ       â y      &   â r/ ÿ       Ä r	 ÿ       â y      '   " r ÿ       ä r  ÿ   pRð är  ÿ   pRò är    ÿ(Ž È/ r   ÿ� ä x   ÿ   äO xÿ     È |    pbô Ú G)  ð    € ê r ÿ       â�$z p  Ž â r ÿ       â $z }  ÿŽ â xþÿÿÿÿÀŽ â 'z y   Ž â z  q      ä r  €ÿàÿ â 'z s   Ž â zÿ z    Ä x   ÿ   â $z w  ÿŽ â zÿ t    â $z |  Ž âx   ÿ   â $z |  Ž â x   ÿàÿ â $z v  Ž â x   áÿ â %x    ÿŽ â r   ÿàõ Ä xÿ     â $z v  Ž â xÿ     â $z p  Ž â r   ÿŽ â %x    Ž â x     ä x     â $x   Ž â z,  X      â %x     Ž â z  Y      Ä r ÿ       ä z q  ÿàÿ ä rÿÿ   ÿä ä x   ÿ   ä x   ÿ   ä r?   ÿà÷ ä r" ,   ÿàù Ä rÿ   ÿäÿ ä x* $       ä r)   ÿà÷ â $x$   ÿŽ â r6,   ÿàõ â %v>? Z  * Ž â r#!   ÿä ä r$ÿ   ÿäÿ Ä r7   ÿä â �y3"   • ¢  %v() Z  * Ž â r&>   ÿàõ ä �y0"   • ¢  $x%$$   ÿŽ â r??   ÿàÿ ä �y6   • â rB(   ÿà÷ Æ �y<6   • â r'?   ÿä ä r))%   ÿàÿ â �y:"   • " r$>   ÿàõ ä rC)   ÿäÿ â �y-"   • " r%?   ÿä Æ �y4&   ™ ( �y96   • " r"(   ÿàù Æ �y6   • ( �y2&   ™ " r#)   ÿä Æ �y=B   ™ " r7   ÿàõ Æ/ �y8B   ™ " rDÿ   ÿä Æ �y5$   ™ " r+   ÿà÷ â $xED$   ÿŽ ä �y1"   ™ ( �y>>    ™ " %v$7 Z  * Ž Æ/ �y6&   ™ ( �y;(    ™ " r?ÿ   ÿäÿ Æ �y.B   ™ " r& ,       är'        â $x??$   ÿŽ â r"$   ÿàõ â %v(+ Z  * Ž â/ r%%E   ÿàÿ Æ %x@   &Ž âr#%   ÿä ä r))?   ÿàÿ â %x*   &Ž â r&(   ÿà÷ â �y7@    • b rD$   ÿàõ ä r')   ÿäÿ â �y?*    • b rE%   ÿä Æ �yB&   ™ h �y@"   ™ b# r*(   ÿàõ Æ �yA"   ™ b r+)   ÿä Æ �yCD   ™ h �y$$    ™ h �y**   ™ h �yD&   ™ h �y%"   ™ h# �y+&   ™ h �y((    ™ b x@   ÿàÿ È |    pbô ä x,,€  ÿà÷ ä x@   ÿàÿ ä x@   ÿàÿ ä x@   ÿàÿ ä rÿ   ÿäÿ Ä x03T  0    ÈO xE0ÿÀŽ ä xT  <    ä� x0ÿ   0  ä x3ÿ     ä xÿÀŽ ä x#0ÿÀŽ Ä/ x-:T  -    â&r"E4   ÿ   âx33ÿÀŽ â &r44   ÿ   â x&ÿ   -  ä x99T      ä x-ÿÀŽ â &r)#2   "   âx'ÿ   9  â &r232   4   â x"9ÿÀŽ â &r-=   ÿ   âx&&ÿÀŽ â &r=E=   ÿ   â x''ÿÀŽ â &r-38   -   Ä &r=#8   =   ä &r)5   )   ä&r2"5   2   ä&r0"1   -   ä&r=1   =   ä 0r-ÿ>  0 A   Ä &r)&6   )   ä&r6'6   2   ä0r2ÿ;  0 A   ä&r1'.   0   â Er0 )       â 0r>ÿ>    A   â  x42  €@  A  â Er) 6       â 0r;ÿ;    A   â  x--  €@  A  â Er2 1       Æ #r0>0   -  â#r)>)   -  â #r-;2   4  â &r2@   ÿ   ä
&r@E@   ÿ   ä &r23A   2   ä&r@#A   @   ä &rB   ÿ   Ä &rBEB   ÿ   ä &r2"C   2   ä&r@C   @   ä &r33D      ä&r##D   B   ä 0rÿ$  0 A   Ä &r2'%   2   ä&r@&%   @   ä &r""*   3   ä&r**   #   â  x#  €@  A  â 0r$ÿ$    A   â Er 2       â &r=&.   =   â Er@ @       â &r"'+   "   Ä &r*&+   *   ä 0rÿ(  0 A   â#r$   #  â#r@$@   #  â Er= =       â 0r(ÿ(    A   â Er# "       â  x%  €@  A  â Er* *       â 0r.ÿ7    A   â #r4;=   4  â 0r?ÿ?    A   â #r#(#   %  â#r*(*   %  â #r.)      â#r
.-   
   â#r?0      â#r	?4   	   â#r/.   /   â#r?@      â#r.#      â #r?*      â G©  põÿÿÿÿƒ ê Ay        € ê�Ey  °     € â G‰  �     € ê r     ÿ@Ž Ê ˆs   üÿ H   è ˆs  €üÿ H   è ˆs  
 ýÿ H   è ˆs  	€ýÿ H   è ˆs  / þÿ H   è ˆs  €þÿ H   è ˆs   ÿÿ H   è ˆs  €ÿÿ H   ä Ay        € ê {          ì M        € ê „y    H   b z  r  p`ð â$z x  Ž â x   ÿàÿ â „y    H   ¢ r ÿ   p"p â $z ~  Ž â z  r  p`ò â „y      H   & x    p"ò â „y  	  H   è „y €   H   ( „y    H   h „y# €  H   h „y €  H   h „y    H   h „y €  H   h „y    H   b !r
       Æ/ „y €  H   b !r       ÆO „y'    H   ¢ !r         Æ „y €  H   " !r       Æ� „y    H   â !r       Æ„y €  H   " !r         Æ„y    H   b !r##       Æ „y €	  H   b !r       Æ „y!  
  H   b !r         Æ „y% €
  H   b !r	       Æ „y)    H   b !r/       Æ „y €  H   â !r       Æ/ „y €  H   â !r''       ÆO „y €  H   â !r       Æ ‰       " !r       â� !r#       ä‰      ¢ 
 r  ÿàõ â/ !r'       ä‰	      b ÿÿ   ÿä â !r       â !r!       È ‰      â !r%       Æ ‰      " !r)       Æ ‰      b !r        â !r       ä ‰      " !r       âO !r       ä ‰       ¢ !r		       â/ !r       Ä ‰      b !r       Æ� ‰	      â !r       Æ‰      " !r       Æ‰      b !r       Æ ‰
      " z  r      â !r         äO ‰      ¢ r   ÿŽ â !r       ä/ ‰      h ‰  €    b !r	       Æ� ‰ €    â !r       â!r       ä‰	 €    " !r

       Æ ‰ €    " !r       ÆO ‰ €    ¢ !r       Æ/ ‰
 €    b !r        Æ ‰ €    b !r       Æ� ‰      è ‰  @    â !r		       Æ‰ @    " !r       â !r       äO ‰	 @    " !r

       Æ/ ‰ @    b !r       Æ‰ @    ¢ !r       Æ� ‰
 @    â !r         â r   ÿ† ä ‰ @    b !r       â
 j  ÿ„ ä ‰ €    " rÿÿ   ÿäÿ â !r	       Ä ‰       " z j  ÿ† ä 
 k   â !r       â/ ‰      b z k  � â !r       äO ‰      ¢    ÿàÿ â !r
       Ä� ‰      â    ÿà÷ â!r       â   ÿàõ â ‰      b x        â !r       âÿÿ   ÿä â ‰      " 
 j  ÿ„ â !r        â  ÿÿ   ÿäÿ â ‰ @    " 
 j  ÿ† â %v j   Ž Ä ‰      " 
 k   � â !r       â/  k   â !r       âO †	     â !r       Æ� †    â !r       Æ†	     â !r       Æ† 
    â !r       Æ †	     â !r       Æ †    è ‰      ( †	     â M™        € ê x    ÿàÿ â !r       Æ r     ÿàñ È rÿÿ   ÿä  ä/ z  j  ÿ€ È z  k    Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y	      "   " ¹z  l      â Ey       € â ™x?    â y       !   b ¹z  F   
   â r ÿ       â x      ÿ  â y
      %   ¢ r ÿ       ä r+ ÿ       â y      &   â r' ÿ       Æ y      '   " r 	ÿ   pRð är 	ÿ   pRò är	    ÿ(Ž È/ r   ÿ� È xÿ     â $	À    Ž â x	
   ÿ   äO |    pbô Ú G)  Ð
    € ê r ÿ       â�$z	 p  Ž â r ÿ       â $z }  ÿŽ â xþÿÿÿÿÀŽ â 'z y   Ž â z  q      ä r  €ÿàÿ â 'z
 s   Ž â zÿ z    Ä x   ÿ   â $z w  ÿŽ â z
ÿ t  
  â $z |  Ž â x   ÿ   â $z |  Ž â r   ÿŽ â $z
 v  Ž âx	   ÿàÿ â $z
 v  Ž â x   áÿ â %x   ÿŽ â xÿ     Ä xÿ     â $z p  Ž â r   ÿàõ â %x   Ž â x     ä x     â %x   Ž â z  X      ä z  Y      Ä r ÿ       ä z q  ÿàÿ ä rÿÿ   ÿä ä x   ÿ   ä x   ÿ   ä r    ÿàõ ä r$   ÿà÷ Ä x $       ä rÿ   ÿä ä r%   ÿäÿ â %v   Z   Ž â r(   ÿàõ ä r4   ÿà÷ â $x#$   ÿŽ â rÿ   ÿä â �y;$   • ¢  r5   ÿäÿ Ä r!!#   ÿàÿ â �y
$   • ¨  �y9$   • è  �y6$   • â  %v(( Z   Ž â r"    ÿàõ Ä �y84   • " $x$   ÿŽ Æ �y34   • " r$    ÿà÷ Æ �y-4   • b r#!   ÿä är%!   ÿäÿ â �y&4   • b r.   ÿà÷ ä r))   ÿàÿ â �y,$   ™ b r<(   ÿàõ ä rÿ   ÿäÿ â �y1     ™ b r=)   ÿä Ä r4(   ÿàõ â/ $x$   ÿŽ â �y0"   ™ b %v$. Z   Ž â r5)   ÿä ä �y/"   ™ d r $   ÿàõ ä r%%   ÿàÿ â �y7"   ™ f r!%   ÿä â �y*<   ™ b r        Æ �y2    ™ b r"$   ÿàõ Æ/ �y<   ™ b r#%   ÿä Æ �y54   ™ h �y.<   ™ h �y""   ™ h �y4    ™ b# %x<   Ž Æ �y((    ™ h �y$$    ™ h �y)<    • d %x<   Ž Ä �y    ™ h �y:<    • b x@   ÿàÿ È |    pbô ä x€  ÿà÷ ä x@   ÿàÿ ä x@   ÿàÿ ä rÿ   ÿäÿ ä x
;T  
    ÄO x69T  6    ä� x88T  3    äx3
ÿÀŽ ä x
ÿ   
  ä x9ÿ   8  ä x#8ÿÀŽ Ä x&-T  &    äx

ÿÀŽ ä x 9ÿÀŽ â/ &r-30   ÿ   ä&r0#0   ÿ   ä &r8
/   -   ä&r0 /   0   â x!6ÿÀŽ Ä x%&ÿÀŽ â &r/#*   ÿ   ä&r*3*   ÿ   ä &r##2   ÿ   ä&r332   ÿ   â x-ÿ   &  â &r<    /   âx6ÿ   6  â &r/
   *   Ä &r8!,   8   ä&r0%,   0   â x,-ÿÀŽ â 0r-ÿ1  0 A   â x&6ÿÀŽ â &r%5   <   ä&r/!5   /   ä &r# 4   #   Ä &r3
4   3   ä &r#%"   #   ä&r3!"   3   â  x6-  €@  A  â 0r ÿ$  0 A   ä &r8&7   8   ä&r0,7   0   Ä &r,.      ä&r/&.   /   ä 0r-ÿ(  0 A   âEr* 8       â 0r1ÿ1    A   â Er0 0       â &r#,   #   ä&r3&   3   â  x   €@  A  â 0r(ÿ(    A   â Er        â 0r$ÿ$    A   â Er. /       â  x--  €@  A  â Er# #       Ä Er  3       â 0r
ÿ)    A   â #r*1*   6  â0r:ÿ:    A   â #r010   6  â #r(   -  â#r.(.   -  â #r#$#     â#r $      â #r'
0   '   â#r+
   +   â #r:*      â#r:.      â #r
#      â #r:       â G©  €÷ÿÿÿÿƒ ê Ay        € ê�ˆ ' ýÿ H   è ˆ €ýÿ H   è ˆ + þÿ H   è ˆ €þÿ H   è ˆ  ÿÿ H   è ˆ €ÿÿ H   è {          ì M        € ê „y      H   " z 	 r  p`ð Ä z  r      â „y     H   b r  ÿ   p"p Æ „y  €   H   ¨ „y     H   è „y     H   ( „y
  €  H   h „y     H   h „y  €  H   h „y     H   h „y  €  H   h „y  €  H   b !r'       Æ „y     H   " !r       Æ/ „y  €  H   b !r       ÆO „y     H   ¢ !r+       Æ� „y  €  H   â !r       Æ„y#  €  H   â !r

       Æ„y!     H   â !r       Æ „y%  €  H   â !r       â !r       ä ‰      " !r

       â !r       â !r       â !r       â/ !r       ÈO ‰      " !r

       Æ� ‰      b !r       â!r#       ä ‰
      ¢ !r!!       Æ ‰      â !r%%       Æ ‰!      ( ‰%      b !r       â !r       È/ ‰      " !r
       ÆO ‰      b !r       Æ� ‰      ¢ !r!       Æ‰ €    â !r%       Æ‰      ( ‰      b !r       â !r       È/ ‰ €    " !r       ÆO ‰ €    b !r       Æ� ‰ €    ¢ !r       Æ‰ @    â !r       Æ‰
 €    ( ‰ €    b !r       â !r       È/ ‰ @    " !r       ÆO ‰ @    b !r       Æ� ‰ @    ¢ !r
       âx
	   ÿàÿ ä ‰      â $z	 x  	Ž â !r       âz 
 r  p`ò â ‰ @    " $z	 ~  	Ž ä x     p"ò â ‰ @    b !r       â !r
       È/ ‰       " 
	 r  ÿàõ â !r       ÈO    ÿàÿ â ‰
      b ÿÿ   ÿä ä 
 j  ÿ„ â !r       â� 	   ÿà÷ â !r       â
 k   â ‰      ¢ r	   ÿ„ â !r       âx        â ‰      â ÿÿ   ÿäÿ Ä rÿÿ   ÿä â ‰      "  j  ÿ† â %v	 j   Ž â z j  ÿ„ â !r        â 	 k  � â !r
       â/ z k   â †	     è †    â !r       ÆO †	     â !r       Æ� †     â !r       Æ†	     â M™        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      "   " ¹z  l      â Ey  ð	    € â ™x?    â y       !   b ¹z  F   
   â x      ÿ  â r ÿ       â y	      %   ¢ r ÿ       Æ y      &   è y      '   " r ÿ   pRð är ÿ   pRò är    ÿ(Ž È/ r   ÿ� È x*ÿ     â $€    Ž â x	   ÿ   äO | *   pbô Ú G)  À    € ê r ÿ       â�$z	 }  ÿŽ âr ÿ       ä xþÿÿÿÿÀŽ â 'z y   Ž â z&  X      ä r  €ÿàÿ â 'z
 s   Ž â zÿ z    Ä z* q  ÿàÿ â $z p  *Ž â z
ÿ t  
  â $z |  *Ž ä$z |  Ž â x   ÿàÿ â $z
 v  Ž ä$z

 v  Ž â x   ÿ   â$x   ÿ Ž â z  Y      â $z w  ÿŽ Ä $z p  Ž â x'   áÿ â %x   ÿŽ â r		   ÿàõ ä xÿ     ä xÿ   '  â %x
   Ž â x     äx'     â $rÿÿ   ÿ â r ÿ       â %x   Ž â x   ÿ   Ä x''   ÿ   ä r*	   ÿàõ är$	   ÿàû ä r"&   ÿàù â$rÿÿ    â x% $       ä r &   ÿà÷ â $x$   ÿŽ â rÿ   ÿäÿ â %v Z  % Ž â r#   ÿä Ä r!   ÿäÿ â %v$$ Z  % Ž â r   ÿàõ â �y)"   • ¢  r   ÿàÿ â $x$   ÿŽ â r$   ÿà÷ â �y"   • ¨  �y-    • â r%%   ÿàÿ Æ �y4    • â $x    â r%   ÿäÿ ä �y2"   • ( �y1   ™ ( �y/"   • ( �y(    • ( �y.   ™ " r"'$   ÿà÷ Æ �y+    • " r#%   ÿäÿ Æ �y0   ™ ( �y,   ™ " r '   ÿàõ Æ/ �y""   ™ " r!   ÿä Æ �y3   ™ ( �y     ™ ( �y5   ™ ( �y    ™ (�y$$    ™ " r &       â$rÿÿ    Ž È %x
   Ž È %x   Ž â �y!    • h �y#    • b# x**@   ÿàÿ È | *   pbô ä x&&€  ÿà÷ ä x@   ÿàÿ ä rÿ   ÿäÿ ä x)T  )    ÈO x)ÿÀŽ ä x--T  4    ä� xÿ   )  ä x)ÿ   -  ä x--ÿÀŽ ä xÿÀŽ Ä x%)ÿÀŽ â &r)1   ÿ   â	x/2T  /    â &r-1   ÿ   ä&r--.   ÿ   ä&r..   ÿ   â x((T  +    â &r0   )   â/x)/ÿÀŽ â &r%0      âx/ÿ   /  â &r-%,   -   â xÿ   (  Ä x%(ÿÀŽ â &r.,   .   â x/ÿÀŽ ä xÿÀŽ â &r-%"   -   ä&r.)"   .   ä&r)       ä&r%       Ä &r3      ä&r3      ä&r-5   -   ä0rÿ  0 A   ä &r.5   .   ä 0rÿ$  0 A   â Er        â 0rÿ    A   â Er        â 0r$ÿ$    A   â  x  €@  A  â Er- -       â  x  €@  A  â Er. .       â #r     â#r     â #r$-     Ä #r$.     â 0rÿ!    A   ä0rÿ#    A   Æ #r      â#r      â #r      â#r      â G©  pùÿÿÿÿƒ ê Ay        € ê�ˆ  þÿ H   è ˆ €þÿ H   è ˆ  ÿÿ H   è ˆ €ÿÿ H   è {          ì M        € ê „y      H   " z  r  p`ð â $z x  Ž â x   ÿàÿ â „y  €   H   b r  ÿ   p"p â$z ~  Ž â z  r  p`ò â „y     H   ¦ x     p"ò â „y	     H   è „y  €  H   ( „y     H   h „y     H   h „y  €  H   h „y     H   h „y  €  H   h „y  €  H   b !r       Æ „y  €  H   b !r       â/ !r       âO !r	       â� 
  r  ÿàõ â !r       â!r       â!r       ä !r       Æ ‰      " !r       Æ ‰      b !r       Æ ‰
      ¢ !r       È !r       Ê ‰      â !r       â !r       È/ ‰      " !r

       ÆO ‰      h ‰	
      ¢ !r       Ê� ‰      â !r       â !r       È/ ‰ €    " !r
	       ÆO ‰	 €    h ‰ €    ¢ !r       Ê� ‰ €    â !r       â !r		       â/ !r
       ÈO ‰	 @    ( ‰ @    h ‰
 @    ¢ !r       Ê� ‰ @    â !r	       â !r       È/ ‰      " ÿÿ   ÿä â !r
       âO x
        â ‰      b 
  j  ÿ„ Æ ‰	      ¢ 
  k   â !r       Ê� ‰      " !r       â !r       â/ %v j  
 Ž â !r		       ÈO †	     è †    è †	 	    â M™        € ê ¹z  r      â!r       â �x   ?àÿ Ì |   ÿàñ È r ÿÿ   ÿä  ä z j  ÿ€ È� z k     Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž È % \   Ž Ê �	    ™ b r ÿ       â ¹z  l      â Ey  Ð    € â y       "   b ™x?    â '
 m   Ž ä y      !   d $rÿÿ   ÿ Ž â 
ÿ n    â y      '   ¨ $	ÿÿ   
Ž Ä y      %   ä $
	 o  Ž â ‚        Ê $z w  ÿŽ ä 'Š s   Ž È $x     Ž ä/ $z
 }  ÿŽ ÆO r   ÿ� È x"ÿ     ä r
   ÿàõ ä | "   pbò ä Šÿ t    â $rÿÿ   ÿ Ô G  `
    € ê r	ÿ"   ÿ3Ž â� Ey Ð    € â r ÿ       â $z p  ÿŽ â |
	   ÿàÿ â $rÿÿ   ÿ Ž â x	þÿÿÿÿÀŽ â 'z y   Ž â xÿ
@   ÿÀ‚ äxÿ
ÀÿÿÿÿÀ€ â $x   	
Ž â zÿ z    È x   ÿ   Ê G       € ê $z v  Ž âr"   ÿàó â$tÿ   ÿ Ž ä $z	 |  Ž â rÿ   ÿäÿ  â $tÿ$   ÿ Ž ä $x
"   	Ž ä %v Z   Ž È %v

 X  Ž â x   ÿàó Æ $x$   ÿŽ ä $x   ÿ Ž Æ r		   ÿàÿ â %x   
Ž â x   áÿ â�y

    • ¤  $rÿÿ   	Ž  ä �y   • è �y   • â %x   Ž Æ �y   • ( �y!   • " %x   Ž Æ �y    ™ ¨ �y   ™ ¨ �y    ™ ¨ �y   ™ ¨ �y    ™ ¢ x""@   ÿàÿ Ä xT      È� xÿÀŽ ä xÿ     ä xT  !    äxÿÀŽ â &r   ÿ   âO xÿÀŽ ä/ xÿ     â &r      Æ xÿÀŽ â &r      È &r      ä 0r	ÿ  0 A   Æ Er        â 0rÿ    A   â  x		  €À  A  â 0rÿ
    A   Æ #r   	   È #r   ÿ   ä Ay       € ê z!  X      â $vÿ Y  ÿ Ž â G‰  @    € ê $z
 |  "Ž â x   ÿ   Æ $z v  
Ž âx   áÿ äx
ÿ     â $x   Ž â xÿ     ä x	   
  â %x   ÿŽ â x     äx
   ÿ   â $x   ÿ Ž Ä %x   Ž È r"   ÿàñ äx"@   ÿàÿ ä x $       â $rÿÿ     â r   ÿàñ ä r0!   ÿàó â $x$   ÿŽ ä %v Z   Ž â r1   ÿäÿ  Æ $rÿÿ     â r   ÿàÿ â %v Z   Ž â r
   ÿàñ â�y 0   • ¤  $x$   ÿŽ ä �y0   • ¢  $x   	  â r
   ÿàñ Ä r   ÿàÿ â �y-0‚  • è  �y)0„  • â  $x   	  â r.   ÿàñ ä�y,0   • ( �y#0   • " r   ÿàó Æ �y+   ™ b r/   ÿä  Æ �y&   ™ h �y*0†  • h �y'0ˆ  • b $x   Ž  Æ �y%   ™ h �y$   ™ h �y(.   ™ b r0!   ÿàñ Æ �y   ™ h �y   ™ b! r1   ÿä  Æ �y    ™ h �y2   ™ h �y    ™ h �y30    • h �y40€  • b x""€   ÿàÿ È | "   pbð ä x!! 	  ÿàó Ê $rÿÿ   Ž  â x T      ÈO x ÿÀŽ ä xÿ     ä x)-T  )    ä� x.ÿÀŽ ä x)ÿÀŽ ä x)ÿ   )  Ä x#,T  #    â&r  +   ÿ   âx))ÿÀŽ â &r&&   ÿ   â x#ÿÀŽ ä x#ÿ   #  ä x'*T  '    â &r%.%       Æ x'ÿÀŽ ä x'ÿ   '  â &r))$   &   â x##ÿÀŽ â &r(   %   â x''ÿÀŽ â &r   )   ä &r#      ä 0rÿ  0 A   Æ Er        â 0rÿ    A   ä &r''2      â  x  €@  A  â 0rÿ  0 A   Æ Er' '       â 0rÿ    A   â #r     â 0rÿ3    A   â  x  €@  A  â 0r4ÿ4    A   Æ #r'     â #r      È #r4      â G‰  �úÿÿÿÿƒ ê Ay        € ê� r  ÿ   pRð är  ÿ   pRò Ö ˆ €ÿÿ H   è {          ì M        € ê „y      H   " r ÿ   pRð â $z x  Ž Ä „y €   H   h „y    H   ¢!r         È !r         È/ !r         ÊO ‰       $ !r        Ê ‰      $ !r       Ê ‰ €    $ !r       Ê ‰ @    $ !r       Ê ‰	      b  M	        € ê x        â $z ~  Ž â !r		       Æ/ %v j   Ž Ê †y 	    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y	      &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž È %	 \   Ž Ì �	    ™ ¢  $rÿÿ   ÿ Ž â ¹z  l      â Ey  0    € â y       "   b '
	 m   Ž â ™x?    â r ÿ       â y      !   b $rÿÿ   ÿ Ž â 
ÿ n    Ä y      '   ä ÿ   ÿáÿ â 'Š	 s   Ž È $
 o  	Ž ä$‚ÿÿ   	 Ž â rÿ    ÿü‚ ä/ r    ÿ(Ž È r   ÿ� È xÿ     È |    pbô ä Šÿ t    È r
	       âO G  à    € ê z ÿ `  pPò â� $rÿÿ   ÿ Ž â z ÿ ^  pPð ä z ÿ b  p ö ä z ÿ a  Sò ä z ÿ _   Sð ä z ÿ c  0'ò Ä r ÿ       Ò 		      %   " $
 ~  ÿŽ ä$š ~  ÿŽ â ™      %   b $š
 x  ÿŽ ä$

 x  ÿŽ â    	àë ä	ÿ   	  Ä ’   àÇ ä/˜ÿ     ä 
ÿÿ   	Äþ ä ’   ÿàû ä ’ÿÿ   „þ ä 	   ÿàý Ä š b  ÿˆ â$’ÿÿ   Ž â 
ÿ
   ÿä ä 
	 ^  ÿ† äš c   ä 
		 _  
� Æ �™     h �	     d Ay        € ê� Ey  p    € â r ÿ       â $rÿÿ   ÿ Ž â G)  0    € ê y
      %   b r ÿ       â $z }  ÿŽ â xþÿÿÿÿÀŽ Ä z ÿ `  pPð â 'z y   Ž â r ÿ       ä z ÿ a   Sð â $x   
Ž â z	ÿ z    Æ $x   ÿ Ž ä $z		 |  Ž ä $x   ÿ Ž ä $z	 v  	Ž ä $z w  ÿŽ â x   áÿ â $z	
 p  	Ž Ä/ $tÿ   ÿ Ž â r   ÿàó â %x	   ÿŽ Æ rÿÿ   ÿäÿ  â %v X  Ž âr        ä r ÿ       â %v `  Ž È r
   ÿàó â $tÿ$   ÿ Ž â r   ÿàõ ä rÿ   ÿäÿ  â %v

 Z   Ž â r   ÿä Æ $x	$   ÿŽ â x
   ÿàó â �y   • ¤ $x   	Ž ä �y!   • ¤ $rÿÿ   Ž  ä �y    • ä  %x   Ž Ä �y#   • ä  %x   Ž ä �y    ™ ( �y"   ™ ( �y    ™ " z X  ÿàó Æ �y$   ™ ( �y

    ™ " z Y  ÿäÿ  Ê �y    • bx@   ÿàÿ ä xT  !    ÈO xÿÀŽ ä xÿ     ä x  T  #    ä� x!ÿÀŽ â &r   ÿ   âx ÿÀŽ ä/ x ÿ      â &r!"      Æ x	 ÿÀŽ â &r      È &r	$      ä 0rÿ
  0 A   Æ Er        â 0rÿ
    A   â  x  €À  A  â G‰  �    € ê r
   ÿàó È r   ÿäÿ  Ê �y 
    ¨ �y!
    ¨ �y
    è �y
    â z `  ÿàó È z	 a  ÿäÿ  Ê �y     " x  T  !    ÈO x! ÿÀŽ ä x ÿ      ä xT      â� &r!!   ÿ   â x# ÿÀŽ ä xÿÀŽ ä xÿ     â &r!#"   !   Æ xÿÀŽ â &r   !   ä 0r	ÿ    A   ä&r$      Ê Er        Ê #r      È #r	      â |    pbô â 0rÿ    A   âx€  ÿàó â #r      È $rÿÿ   Ž  â #r      Ê G©  `ûÿÿÿÿƒ ê Ay        € ê r  ÿ   pRð âEy  €     € â r  ÿ   pRò Ö G‰  P     € ê z ÿ `  pPð ä r     ÿ(Ž ä z ÿ a   Sð Æ ˆs  €ÿÿ H   ô ˆ     H   ä Ay        € ê {          ì M        € ê z ÿ `  pPð â „y      H   f. z ÿ a   Sð Ú „‰ €   H   ¨ „‰    H   â !r        È/ !‚       ÈO !‚       â� G‰        € ê „y  €  H   h „y €   H   ¨ „y    H   è „y	    H   ( „y €  H   " !r         â/ !r       ÆO !r         â� !r	       Æ !r        ä‰      b z ÿ `  pPð ä r ÿ   pRò ä z ÿ a   Sð â !r       Ê/ ‰       d !r         Ê/ ‰  €    d !r        Ê/ ‰ @    d !r       Ê/ ‰	      $ !r		       â G‰        € ê ‰       $ !r         Ê ‰       $ !r        Ê ‰ €    $ !r       Ê ‰ @    $ !r       Ê ‰      $ !r       ä M        € ê !r	       âG‰  �    € ê $t ÿ   ÿ Ž â !r       È z   h  pð Ú G	  �    € ê ¹z  h      ä �xÿÿÿÿ?àÿ È Œx    pð È ‡x       È ™x   ?  Ø »z  €    ä ™x?    Ô Xy @ñÿÿÿÿƒ ê  x ;ª¸¿  A  Ì s          $ !t   €?    Ì s        $  r     A  È  r     A  â Gy  `    € ê  x '7=  A  â x*BL?  A  Æ #t   €?    È  r      A  â  x    ?  A  Ê s     $   $ !t  €?    È  r      A  È  r     A  â Gy  À    € ê z   h      È x     pRð Ú G‰  à     € ê x     pRð Ú G	        € ê 	z  i    � ä 	z i    � Æ  x ;ª¸¿  A  â 	z i €  � Ê s        $ !t  €?    Ì s        $  r      A  È  r     A  â Gy  Ð     € ê  r     A  â Gy  °     € ê 	x   à@  � ä 	x  à@  � Æ  x #ÛÙ?  A  â 	x  àÀ  � Æ  x;ª¸¿  A  ä !t  €?    È s        $ !t  €?    Ì s        $  r       A  È  r      A  ä y      &   ( y      %   $ $z x  Ž â x        Æ $z ~  Ž È %v j   Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž È % \   Ž Ê �	    ™ b r ÿ       â ¹z  l      â Ey  Ð    € â y       "   b ™x?    â '
 m   Ž ä y      !   d $rÿÿ   ÿ Ž â 
ÿ n    â y      '   ¨ $	ÿÿ   
Ž Ä y      %   ä $
	 o  Ž â ‚        Ê $z w  ÿŽ ä 'Š s   Ž È $x     Ž ä/ $z
 }  ÿŽ ÆO r   ÿ� È x"ÿ     ä r
   ÿàõ ä | "   pbò ä Šÿ t    â $rÿÿ   ÿ Ô G  `
    € ê r	ÿ"   ÿ3Ž â� Ey Ð    € â r ÿ       â $z p  ÿŽ â |
	   ÿàÿ â $rÿÿ   ÿ Ž â x	þÿÿÿÿÀŽ â 'z y   Ž â xÿ
@   ÿÀ‚ äxÿ
ÀÿÿÿÿÀ€ â $x   	
Ž â zÿ z    È x   ÿ   Ê G       € ê $z v  Ž âr"   ÿàó â$tÿ   ÿ Ž ä $z	 |  Ž â rÿ   ÿäÿ  â $tÿ$   ÿ Ž ä $x
"   	Ž ä %v Z   Ž È %v

 X  Ž â x   ÿàó Æ $x$   ÿŽ ä $x   ÿ Ž Æ r		   ÿàÿ â %x   
Ž â x   áÿ â�y

    • ¤  $rÿÿ   	Ž  ä �y   • è �y   • â %x   Ž Æ �y   • ( �y!   • " %x   Ž Æ �y    ™ ¨ �y   ™ ¨ �y    ™ ¨ �y   ™ ¨ �y    ™ ¢ x""@   ÿàÿ Ä xT      È� xÿÀŽ ä xÿ     ä xT  !    äxÿÀŽ â &r   ÿ   âO xÿÀŽ ä/ xÿ     â &r      Æ xÿÀŽ â &r      È &r      ä 0r	ÿ  0 A   Æ Er        â 0rÿ    A   â  x		  €À  A  â 0rÿ
    A   Æ #r   	   È #r   ÿ   ä Ay       € ê z!  X      â $vÿ Y  ÿ Ž â G‰  @    € ê $z
 |  "Ž â x   ÿ   Æ $z v  
Ž âx   áÿ äx
ÿ     â $x   Ž â xÿ     ä x	   
  â %x   ÿŽ â x     äx
   ÿ   â $x   ÿ Ž Ä %x   Ž È r"   ÿàñ äx"@   ÿàÿ ä x $       â $rÿÿ     â r   ÿàñ ä r0!   ÿàó â $x$   ÿŽ ä %v Z   Ž â r1   ÿäÿ  Æ $rÿÿ     â r   ÿàÿ â %v Z   Ž â r
   ÿàñ â�y 0   • ¤  $x$   ÿŽ ä �y0   • ¢  $x   	  â r
   ÿàñ Ä r   ÿàÿ â �y-0‚  • è  �y)0„  • â  $x   	  â r.   ÿàñ ä�y,0   • ( �y#0   • " r   ÿàó Æ �y+   ™ b r/   ÿä  Æ �y&   ™ h �y*0†  • h �y'0ˆ  • b $x   Ž  Æ �y%   ™ h �y$   ™ h �y(.   ™ b r0!   ÿàñ Æ �y   ™ h �y   ™ b! r1   ÿä  Æ �y    ™ h �y2   ™ h �y    ™ h �y30    • h �y40€  • b x""€   ÿàÿ È | "   pbð ä x!! 	  ÿàó Ê $rÿÿ   Ž  â x T      ÈO x ÿÀŽ ä xÿ     ä x)-T  )    ä� x.ÿÀŽ ä x)ÿÀŽ ä x)ÿ   )  Ä x#,T  #    â&r  +   ÿ   âx))ÿÀŽ â &r&&   ÿ   â x#ÿÀŽ ä x#ÿ   #  ä x'*T  '    â &r%.%       Æ x'ÿÀŽ ä x'ÿ   '  â &r))$   &   â x##ÿÀŽ â &r(   %   â x''ÿÀŽ â &r   )   ä &r#      ä 0rÿ  0 A   Æ Er        â 0rÿ    A   ä &r''2      â  x  €@  A  â 0rÿ  0 A   Æ Er' '       â 0rÿ    A   â #r     â 0rÿ3    A   â  x  €@  A  â 0r4ÿ4    A   Æ #r'     â #r      È #r4      â G‰  �úÿÿÿÿƒ ê Ay        € ê� r  ÿ   pRð är  ÿ   pRò Ö ˆ €ÿÿ H   è {          ì M        € ê „y      H   " r ÿ   pRð â $z x  Ž Ä „y €   H   h „y    H   ¢!r         È !r         È/ !r         ÊO ‰       $ !r        Ê ‰      $ !r       Ê ‰ €    $ !r       Ê ‰ @    $ !r       Ê ‰	      b  M	        € ê x        â $z ~  Ž â !r		       Æ/ %v j   Ž Ê †y 	    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y	      &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž È %	 \   Ž Ì �	    ™ ¢  $rÿÿ   ÿ Ž â ¹z  l      â Ey  0    € â y       "   b '
	 m   Ž â ™x?    â r ÿ       â y      !   b $rÿÿ   ÿ Ž â 
ÿ n    Ä y      '   ä ÿ   ÿáÿ â 'Š	 s   Ž È $
 o  	Ž ä$‚ÿÿ   	 Ž â rÿ    ÿü‚ ä/ r    ÿ(Ž È r   ÿ� È xÿ     È |    pbô ä Šÿ t    È r
	       âO G  à    € ê z ÿ `  pPò â� $rÿÿ   ÿ Ž â z ÿ ^  pPð ä z ÿ b  p ö ä z ÿ a  Sò ä z ÿ _   Sð ä z ÿ c  0'ò Ä r ÿ       Ò 		      %   " $
 ~  ÿŽ ä$š ~  ÿŽ â ™      %   b $š
 x  ÿŽ ä$

 x  ÿŽ â    	àë ä	ÿ   	  Ä ’   àÇ ä/˜ÿ     ä 
ÿÿ   	Äþ ä ’   ÿàû ä ’ÿÿ   „þ ä 	   ÿàý Ä š b  ÿˆ â$’ÿÿ   Ž â 
ÿ
   ÿä ä 
	 ^  ÿ† äš c   ä 
		 _  
� Æ �™     h �	     d Ay        € ê� Ey  p    € â r ÿ       â $rÿÿ   ÿ Ž â G)  0    € ê y
      %   b r ÿ       â $z }  ÿŽ â xþÿÿÿÿÀŽ Ä z ÿ `  pPð â 'z y   Ž â r ÿ       ä z ÿ a   Sð â $x   
Ž â z	ÿ z    Æ $x   ÿ Ž ä $z		 |  Ž ä $x   ÿ Ž ä $z	 v  	Ž ä $z w  ÿŽ â x   áÿ â $z	
 p  	Ž Ä/ $tÿ   ÿ Ž â r   ÿàó â %x	   ÿŽ Æ rÿÿ   ÿäÿ  â %v X  Ž âr        ä r ÿ       â %v `  Ž È r
   ÿàó â $tÿ$   ÿ Ž â r   ÿàõ ä rÿ   ÿäÿ  â %v

 Z   Ž â r   ÿä Æ $x	$   ÿŽ â x
   ÿàó â �y   • ¤ $x   	Ž ä �y!   • ¤ $rÿÿ   Ž  ä �y    • ä  %x   Ž Ä �y#   • ä  %x   Ž ä �y    ™ ( �y"   ™ ( �y    ™ " z X  ÿàó Æ �y$   ™ ( �y

    ™ " z Y  ÿäÿ  Ê �y    • bx@   ÿàÿ ä xT  !    ÈO xÿÀŽ ä xÿ     ä x  T  #    ä� x!ÿÀŽ â &r   ÿ   âx ÿÀŽ ä/ x ÿ      â &r!"      Æ x	 ÿÀŽ â &r      È &r	$      ä 0rÿ
  0 A   Æ Er        â 0rÿ
    A   â  x  €À  A  â G‰  �    € ê r
   ÿàó È r   ÿäÿ  Ê �y 
    ¨ �y!
    ¨ �y
    è �y
    â z `  ÿàó È z	 a  ÿäÿ  Ê �y     " x  T  !    ÈO x! ÿÀŽ ä x ÿ      ä xT      â� &r!!   ÿ   â x# ÿÀŽ ä xÿÀŽ ä xÿ     â &r!#"   !   Æ xÿÀŽ â &r   !   ä 0r	ÿ    A   ä&r$      Ê Er        Ê #r      È #r	      â |    pbô â 0rÿ    A   âx€  ÿàó â #r      È $rÿÿ   Ž  â #r      Ê G©  `ûÿÿÿÿƒ ê Ay        € ê r  ÿ   pRð âEy  €     € â r  ÿ   pRò Ö G‰  P     € ê z ÿ `  pPð ä r     ÿ(Ž ä z ÿ a   Sð Æ ˆs  €ÿÿ H   ô ˆ     H   ä Ay        € ê {          ì M        € ê z ÿ `  pPð â „y      H   f. z ÿ a   Sð Ú „‰ €   H   ¨ „‰    H   â !r        È/ !‚       ÈO !‚       â� G‰        € ê „y  €  H   h „y €   H   ¨ „y    H   è „y	    H   ( „y €  H   " !r         â/ !r       ÆO !r         â� !r	       Æ !r        ä‰      b z ÿ `  pPð ä r ÿ   pRò ä z ÿ a   Sð â !r       Ê/ ‰       d !r         Ê/ ‰  €    d !r        Ê/ ‰ @    d !r       Ê/ ‰	      $ !r		       â G‰        € ê ‰       $ !r         Ê ‰       $ !r        Ê ‰ €    $ !r       Ê ‰ @    $ !r       Ê ‰      $ !r       ä M        € ê !r	       âG‰  �    € ê $t ÿ   ÿ Ž â !r       È z   h  pð Ú G	  �    € ê ¹z  h      ä �xÿÿÿÿ?àÿ È Œx    pð È ‡x       È ™x   ?  Ø »z  €    ä ™x?    Ô Xy @ñÿÿÿÿƒ ê  x ;ª¸¿  A  Ì s          $ !t   €?    Ì s        $  r     A  È  r     A  â Gy  `    € ê  x '7=  A  â x*BL?  A  Æ #t   €?    È  r      A  â  x    ?  A  Ê s     $   $ !t  €?    È  r      A  È  r     A  â Gy  À    € ê z   h      È x     pRð Ú G‰  à     € ê x     pRð Ú G	        € ê 	z  i    � ä 	z i    � Æ  x ;ª¸¿  A  â 	z i €  � Ê s        $ !t  €?    Ì s        $  r      A  È  r     A  â Gy  Ð     € ê  r     A  â Gy  °     € ê 	x   à@  � ä 	x  à@  � Æ  x #ÛÙ?  A  â 	x  àÀ  � Æ  x;ª¸¿  A  ä !t  €?    È s        $ !t  €?    Ì s        $  r       A  È  r      A  ä y      &   ( y      %   $ $z x  Ž â x        Æ $z ~  Ž È %v j   Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú         â $rÿÿ    Ž È % \   Ž Ê �	    ™ b ¹z  l      â Ey  °
    € â ™x?    â y      "   b $rÿÿ   ÿ Ž â r ÿ       â $rÿÿ   ÿ Ž â y      !   ¢ r ÿ       â $ÿÿ   ÿ Ž â         â y	      %   ( '
  m   Ž â y      '   æ $‚ÿÿ   ÿ Ž â 
 ÿ n     Ê $ ÿÿ    
Ž â r ÿ   pRò Æ/ $
  o  Ž â r    ÿ(Ž ÈO ‚        ä x		   ÿ   Æ 'Š s   Ž â r     ÿ� ä    ÿ8Ž ä x
ÿ     È | 
   pbô ä Šÿ t    ä r ÿ   pRð Ò G)  à    € ê $rÿÿ   ÿ Ž â� xþÿÿÿÿÀŽ â $z w  ÿŽ â z  X      â 'z y   Ž â r   €ÿàÿ Æ $z 	 p  
Ž â zÿ z    â $x/   ÿ Ž âx   ÿ   â $z }  ÿŽ ä $z |   Ž ä $rÿÿ   ÿ Ž â r   ÿàõ â $z0 v  Ž âx   /áÿ Ä xÿ     â %x00   ÿŽ â x ÿ     ä z  Y      ä rÿÿ   ÿä â %x./   0Ž â x     ä x       Ä r,
   ÿàõ ä x- $       Æ $rÿÿ    â r".   ÿàõ â %v,, Z  - Ž Æ r#/   ÿä â $x$   ÿŽ â r(,   ÿ„ ä z  p      â $x--   Ž â �y"   • ¨  �y"   • ¢  r)-   ÿä â %x*   "Ž â r&,   ÿ„ â �y"   • è  �y(   ™ ( �y"   • â  r'-    ÿä Æ �y*   • â
 %x2   *Ž Æ �y*   • è
 �y(   ™ â %x4   2Ž Æ �y&   ™ è �y!2   • è �y"2   • è �y (   ™ è  �y#4   • è  �y$4   • è  �y*   • è
 �y*   • è
 �y%2   • è  �y&2   • â" r*0   ÿàõ Æ�y'4   • è
 �y(4   • â $x+1    Æ �y),    ™ â  %x4   *Ž Æ�y**    • æ
 %x2   4Ž Ê/ �y,2    • è �y+4    • b%x2   2Ž Ê �y2    • b x

@   ÿàÿ È | 
   pbô ä x€  ÿà÷ È rÿ   ÿäÿ ä xT      ÈO xÿÀŽ ä xÿ     Æ &r   ÿ   âxÿÀŽ ä xT      ä� xT      â &r      â xÿÀŽ ä xÿÀŽ ä xÿ     Ä xÿ     â &r      â xÿÀŽ â &r   ÿ   â xÿÀŽ ä x!!T  "    â &r       Æ xÿ   !  â &r      â x!ÿÀŽ ä x##T  $    Æ &r   ÿ   âxÿÀŽ ä x"#ÿÀŽ ä xÿ   #  ä xT      â &r""   ÿ   â xÿÀŽ â &r      â xÿÀŽ Ä x%%T  &    â &r   "   â xÿ     ä x%ÿÀŽ ä x''T  (    ä x%ÿ   %  ä x'ÿÀŽ Ä x'ÿ   '  â &r      âxÿÀŽ â &r      âx%%ÿÀŽ â &r      â x''ÿÀŽ â &r       ä&r%%       ä&r''       Ä 0rÿ)  0 A   âEr        â 0r)ÿ)    A   â Er        ä Er% %       â  x  €@  A  â Er' '       â 0rÿ*    A   ä 0r+ÿ+    A   â#r)     â0rÿ,    A   â #r)     â#r%)%     â#r')'     â 0rÿ    A   â #r      â #r+      â #r%      Ä #r'      â G©   øÿÿÿÿƒ ê Ay        € ê� ˆ  þÿ H   è ˆ €þÿ H   è ˆ  ÿÿ H   è ˆ €ÿÿ H   è {          ì M	        € ê „y     H   " z 	 r  p`ô Æ „y €   H   b r ÿ   pVt Æ „y    H   ¨ „y     H   è „y €  H   ( „y    H   („y    H   ( „y €  H   ( „y €  H   ( „y    H   ( „y €  H   " !r       Æ „y €  H   " !r       â/ !r       âO !r         ä� !r       â!r       â !r         â !r       È ‰       b !r       â !r       È ‰      ¢ !r       Æ ‰      â !r       Ê ‰
      " !r        â/ !r       ÈO ‰      b !r       Æ� ‰      ¨ ‰      â !r 
       Ê ‰       " !r       â/ !r       ÈO ‰ €    b !r
       Æ� ‰ €    ¨ ‰
 €    â !r        Ê ‰ €    " !r       â/ x	   ÿàÿ â !r       ÆO ‰  @    b z  r  p`ö â !r
       ä� ‰ @    ¢ x
	   ÿàÿ ä x    pVö â ‰ @    â !r       â z 
 r  p`ò Ä x        ä ‰ @    " x    pVò  â !r         â/ x	   ÿàÿ â !r       ÆO z  r  p`ð â $z x  	Ž â !r       â� ‰      b x    pVp  â $z ~  Ž ä ‰       ¢ !r       â %v j   Ž ä ‰      ( ‰      â !r       â/ !r        ÈO †¹    â !r       Æ †©     â !r       Æ� †‰    â M        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y	      &   " z ÿ \  pPò â ¹z  F   
   Æ z ÿ ]  Sò Ú         Ê %	 \   Ž Ê �    ™ ¢  r ÿ       â Ey       € â r ÿ       â y      "   b r ÿ       â ' 	 m   Ž äy      %   æ  ÿ n     â y      !   &  ÿ   ÿáÿ â y      '   " 'š 	 s   Ž È $ o  	Ž â ’ 	       ä r ÿ   pRð ä/ x   ÿ   ä� x    pDp  är   ÿàÿ È z  r  pdp  ä šÿ t     ä r    ÿ(Ž ä r	     € ÎO G	  À    € ê z ÿ `  pPò ä z ÿ ^  pPð ä z ÿ b  p ô ä z ÿ a  Sò ä z ÿ _   Sð ä z ÿ c   'ò Ä r ÿ       ä r ÿ       Î $
 ~  ÿŽ âÿ     â$š	 ~  ÿŽ ä $
 x  ÿŽ â    àë â$š x  ÿŽ â ’
	   àÇ ä ˜	ÿ     ä    ÿàõ Ä ÿ   ÿÄþ ä ’

   ÿàû ä ’	ÿ	   ÿ„þ ä ÿ   ÿä ä 
 ^  ÿ„ ä ’	ÿ	   ÿäÿ Ä š
 b  ÿ† ä 
 _   ä š
 c  	� Æ �	     h �™     d Ay        € ê r     ÿ� â ¹z  l      â $z }  ÿŽ â ™x?    â x	ÿ     â $z w  ÿŽ â r ÿ       â Ey  °    € â r ÿ       â x      ÿ  â | 	   pbð â 'z y   Ž â r   ÿàó â x      ÿ  â x      ÿ  â r ÿ       Ä r
ÿÿ   ÿäÿ  Ì G	       € ê xþÿÿÿÿÀŽ â $z p  	Ž âzÿ z    â x      ÿ  â r    €ÿàÿ â x      ÿ  â x   ÿàÿ â $z |  	Ž â x    ÿ   â $z |  Ž â x   ÿàÿ â$z v  Ž âx   ÿàÿ â $z v  Ž â x    ÿ   â %x:   ÿŽ â z ÿ `  pPð â x      ÿ  â x    áÿ â$z p  Ž â xÿ     â $z p  Ž âz  `      â $z p  Ž â x     â %x6   :Ž â xÿ      Ä z  a      â %x8   :Ž âz  X      ä z  Y      â %x   :Ž âr ÿ       ä r ÿ       â %x:   :Ž â z ÿ a   Sð Ä x      ä r(:   ÿàõ ä z/  p      ä r);   ÿä ä r*	   ÿàó ä x+ $       â %x@/   (Ž â rÿ
   ÿäÿ  â �y3(   • ¤  %v** Z  + Ž Ä �y2(   • ¤  $x$   ÿŽ â r>*   ÿ‚ â %x,/   @Ž â �y (   • ä  r++   ÿàÿ ä �y",   • ( �y#,   • " r?+   ÿäÿ  Æ �y!(   • â  r< *   ÿ‚ â %xD/   ,Ž ä �y4@   • è  �y1@   • è  �y&,   • è �y',   • â r=+   ÿäÿ  Æ �y>   ™ è  �y>   ™ â  r,        Æ/ �y(D   • â r-        Æ �y)D   • è �y$@   • è  �y%@   • è  �y<   ™ â %xB   ,Ž Æ �y**    ™ è  �y,D   • è �y>   ™ è �y+D   • â %x@/   BŽ Æ �y-B    • æ %x</   @Ž â/�y.@    • ê %x>/   <Ž ä �y/<    • è �y0>    • â x23T  2    ÈO x32ÿÀŽ ä x2ÿ   2  ä x""T  #    äx  T  !    ä� x#"ÿÀŽ ä x!ÿ   "  Ä x14T  1    ä x"!ÿÀŽ ä x22ÿÀŽ ä x&&T  '    ä x51ÿÀŽ â &r##   ÿ   âx1ÿ   1  â &r33   ÿ   â x!&ÿÀŽ â &r""   #   Ä &r32   3   â x21ÿÀŽ â &r55   ÿ   â x((T  )    Æ &r52   5   â x' ÿÀŽ ä x$$T  %    â &r"!   "   â x!(ÿÀŽ ä x(ÿ   (  ä x%$ÿÀŽ ä x ÿ      Ä x$ÿ   $  ä x&ÿ   &  â &r!!   ÿ   â x((ÿÀŽ ä x++T  ,    â &r4%   5   âx  ÿÀŽ â &r2'   3   â x$$ÿÀŽ Ä x&&ÿÀŽ ä x%+ÿÀŽ â &r!(   !   â x+ÿ   +  â &r     2   ä&r$$   4   âx(+ÿÀŽ â &r&&   "   ä 0r"ÿ*  0 A   Ä &r!%   !   â Er'         â 0r#ÿ*    A   â Er% $       â &r((   !   â  x""  €À  A  â 0r ÿ-    A   â Er& &       â 0r!ÿ.    A   â #r'#'   "   â#r%#%   "   â Er) (       â 0r$ÿ/    A   â #r '      â #r!%      â 0r'ÿ0    A   â #r&#&   "   â#r%#)   "   â G‰       € ê r :   ÿàó È r!;   ÿäÿ  Ê �y+     ¨  �y,     ¨  �y*     è  �y-     ä  r         Ä r!        Ê %x(    Ž Ì �y((     " x++T  ,    ÈO x/+ÿÀŽ ä x+ÿ   +  ä x**T  -    â� &r//   ÿ   â x,+ÿÀŽ ä x+*ÿÀŽ ä x*ÿ   *  â &r/,   /   Æ x**ÿÀŽ â &r++   /   È &r**   +   Ê Er* *       â 0r(ÿ(    A   È#r*#*   "   È #r(*      ä G‰  À    € ê r 8   ÿàó È r!9   ÿäÿ  Ê �y+     ¨  �y,     ¨  �y*     è  �y-     â  x)   ÿàÿ Ê $z)) p  Ž â r         ä r!        Ê %x()    Ž Ì �y((     " x++T  ,    ÈO x/+ÿÀŽ ä x+ÿ   +  ä x**T  -    â� &r//   ÿ   â x,+ÿÀŽ ä x *ÿÀŽ ä x*ÿ   *  â &r/,   /   Æ x**ÿÀŽ â &r     /   È &r *       Ê Er          â 0r!ÿ(    A   È#r #    "   È #r!       ä G‰  À    € ê r 6   ÿàó È r!7   ÿäÿ  Ê �y+     ¨  �y,     ¨  �y*     è  �y-     â  x)   ÿàÿ Ê $z)) p  Ž â r         ä r!        Ê %x()    Ž Ì �y((     " x++T  ,    ÈO x/+ÿÀŽ ä x+ÿ   +  ä x**T  -    â� &r//   ÿ   â x,+ÿÀŽ ä x *ÿÀŽ ä x*ÿ   *  â &r/,   /   Æ x**ÿÀŽ â &r     /   È &r *       Ê Er          â 0r(ÿ(    A   È#r #    "   È #r(       â #r$&      â #r'%      â G‰  À    € ê r$   ÿàó È r%   ÿäÿ  Ê �y($    ¨ �y)$    ¨ �y&$    è �y'$    â x!   ÿàÿ Ä r         Æ $z+! p  Ž â r!        Ê %x +    Ž Ì �y       " x((T  )    ÈO x)(ÿÀŽ ä x(ÿ   (  ä x&&T  '    â� &r))   ÿ   â x((ÿÀŽ ä x&ÿÀŽ ä x&ÿ   &  â &r)(   )   Æ x&&ÿÀŽ â &r   )   È &r&      ä 0rÿ     A   ÆEr        Ê #r#   "   È #r      â x		@   ÿàÿ ä x€  ÿàõ ä | 	   pbò ä x€  ÿà÷ ä rÿ   ÿä ä rÿ   ÿäÿ Î G™  Pñÿÿÿÿƒ ê Ay        € ê y       "   " Ey  0    € â r  ÿ   pRð är  ÿ   pRò Ö G‰  ð     € ê z ÿ `  pPð ä r    ÿ8Ž ä z ÿ a   Sð Æ ˆs  þÿ H   ô ˆƒ €þÿ H   è ˆƒ  ÿÿ H   è ˆƒ €ÿÿ H   â G‰  p     € ê ˆs    H   è ˆs €þÿ H   è ˆs €  H   è ˆs  ÿÿ H   è ˆs    H   è ˆs €ÿÿ H   è ˆs €  H   ä Ay        € ê {          ì M        € ê y      &   ¢ z ÿ `  pPð Æ y      '   "> z ÿ a   Sð â $z  x  Ž âO x        Æ $z  ~   Ž È %v  j   Ž È G	  P    € ê „y      H   " z  r  p`ð Æ „y	 €   H   b r ÿ   pVp  Æ „y    H   ¨ „y    H   è „y €  H   ( „y    H   ( „y    H   ( „y €  H   ( „y    H   („y €  H   ( „y
 €  H   " !r         Æ „y €  H   " !r		       â/ !r       âO !r         ä� !r	       â!r       â !r         ä !r       Æ ‰       b !r       Æ ‰      ¢ !r       Æ ‰      â !r

       È !r
       â !r        â/ !r       ÈO ‰      " !r
       Æ� ‰      h ‰      ¨ ‰	
      â !r       â !r       â/ !r       ÈO ‰      " !r
	       Æ� ‰ €    h ‰  €    ¨ ‰	 €    â !r       â !r       È/ ‰ €    " !r
        ÆO ‰  @    b !r		       â� x   ÿàÿ ä ‰
 @    ¤ z  r  p`ô ä ‰	 @    ä x    pVt â !r       â !r         È/ ‰ @    " x   ÿàÿ â !r
       âO x   ÿàÿ â ‰       b z  r  p`ò â !r	       ä� ‰      ¢ x    pVò  Æ ‰      â !r       â !r        È/ ‰       " !‚       â !r       âO !r       Æ� †‰     â z  r  p`ð â !¢	       â!’       ä x    pVp  ä †© 	   è †™    â !r         Ì M	        € ê !r        Ê †y    â My        € ê z   h      È x     pBð Ú G	  @    € ê x     pRð Ú G‰  À    € ê x     pRð Ú G	   $    € ê „y     H   " z  r  p`ð â Ey  p    € ä „y     H   b r ÿ   pVp  Æ „y	    H   ¨ „y    H   è „y    H   ( „y  
  H   ( „y €  H   è „y €  H   â !r       Æ „y €  H   " !r         â/ !r	       ÆO !r         â� !r       Æ!r         ä ‰      h ‰       ¢ !r       â !r       ä/ „y €   H   â !r        ÆO ‰      ( ‰      b !r       â !r	       È/ ‰ €    ( ‰
	 €    b !r       Æ „y €  H   ¢ !r       È !r       â !r       ä „y €
  H   â !r
	
       Æ/ ‰ @    ( ‰
 @    b !r       âO !r       â !r
       È/ ‰       ( ‰      b !r       â !r         â !r       â/ G	  �     € ê !r       â!r         Æ  x;ª¸¿  A  Ì s        $ !t  €?    Ì s        $  r     A  È  r      A  Ê †y     ä Ay        € ê ‰       b Ey  Ð    € æ ‰      ¨ „y    H   è „y    H   ( „y  	  H   â !r         Æ/ „y    H   â !r       ÆO ‰       h ‰      ( „y    H   ¢ !r       È� !r       â!r        â/ !r	       È ‰ €    " x   ÿàÿ Æ ‰
	 €    b z  r  p`ð â !r       Æ x    pVp  â !r       ÈO !r       â !r       ä „y    H   â !r
	
       Æ/ ‰ @    ( ‰
 @    b !r       â !r
       È/ ‰       ( ‰      b !r       â !r         â !r       â/ G	  �     € ê !r       â!r         Æ  x;ª¸¿  A  Ì s        $ !t  €?    Ì s        $  r     A  È  r      A  Ê †y    ä Ay        € ê ‰       b Ey       € æ ‰      ¨ „y €  H   è „y €  H   è „y €  H   è „y €  H   â !r         Æ/ „y €  H   â !r       ÆO ‰       h ‰      ( „y €	  H   ¢ !r       â� !r        â/ !r	       È ‰ €    ( ‰
	 €    b !r       ÈO !r       â !r       â x   ÿàÿ â !r
	
       Æ/ ‰ @    " z  r  p`ð Æ ‰
 @    b x    pVp  â !r       â !r
       È/ ‰       ( ‰      b !r         â !r       â/ G	  �     € ê !r       â!r         Æ  x;ª¸¿  A  Ì s        $ !t  €?    Ì s        $  r     A  È  r      A  Ê †y    ä Ay        € ê ‰      b !r       â x   ÿàÿ Æ !r       â z  r  p`ð Æ !r       â x    pVp  È ‰       ¢ !r       Ê/ ‰      " !r         ÊO ‰       b !r	       Ê ‰
	 €    " !r        Ê/ ‰ €    b !r
	
       Ê ‰
 @    " !r       Ê/ ‰ @    b !r
       Ê ‰      " !r       Ê/ ‰       ¢ !r       â M	        € ì !r       â/!r         ÆO  x;ª¸¿  A  â !r        Ê s        $ !t  €?    Ì s        $  r      A  È  r      A  Ê †y    â My        € ê „y      H   " z  r  p`ð â Ey       € ä „y    H   b r ÿ   pVp  Æ „y    H   ¨ „y    H   è „y	    H   ( „y  
  H   ( „y €  H   è „y €  H   è „y €
  H   â !r         Æ „y €  H   " !r       Æ/ „y €   H   â !r         âO !r       Æ� !r	 	       â!r       È ‰ 	      h ‰      ¢ !r       â !r 	        â/ !r       ä !r       â !r       âO ‰       $ !r       Ä ‰      b !r        â !r       È/ ‰ €    ( ‰ €    h „y €  H   ¢ !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       ÈO !r       â !r       â !r
	
       È/ ‰       ( ‰
      b !r        â !r 
       â/ G	  À     € ê !r         È x '7=  A  â x *BL?  A  Æ #t   €?   È  r     A  â  x    ?  A  â !r        È s     $   $ !t  €?    È  r     A  È  r      A  Ê †y     ä Ay        € ê ‰       b Ey       € æ ‰      ¨ „y    H   è „y    H   ( „y    H   è „y    H   â !r         Æ/ „y    H   â !r       ÆO ‰       h ‰      ( „y  	  H   ¢ !r       â� !r       â!r        â/ !r       È ‰ €    ( ‰ €    b !r       È !r       â !r       ÈO !r       â !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       â x   ÿàÿ â !r
	
       Æ/ ‰       " z  r  p`ð Æ ‰
      b x    pVp  â !r        â !r 
       Ö/ G	  À     € ê !r         È x '7=  A  â x *BL?  A  Æ #t   €?   È  r     A  â  x    ?  A  â !r        È s     $   $ !t  €?    È  r     A  È  r      A  Ê †y    ä Ay        € ê ‰       b Ey  Ð    € æ ‰      ¨ „y €  H   è „y €	  H   ( „y €  H   è „y €  H   â !r         Æ/ „y €  H   â !r       ÆO ‰       h ‰      ( „y €  H   â !r       È� !r       â!r        Æ/ !r       â !r       â ‰ €    ( ‰ €    b !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       â x   ÿàÿ â !r
	
       Æ/ ‰       " z  r  p`ð Æ ‰
      b x    pVp  â !r        â !r 
       Ö/ G	  À     € ê !r         È x '7=  A  â x *BL?  A  Æ #t   €?   È  r     A  â  x    ?  A  â !r        È s     $   $ !t  €?    È  r     A  È  r      A  Ê †y    ä Ay        € ê ‰      b !r       â x   ÿàÿ Æ !r       â z  r  p`ð Æ !r       â x    pVp  È ‰       ¢ !r       Ê/ ‰      " !r         ÊO ‰       b !r       Ê ‰ €    " !r        Ê/ ‰ €    b !r	       Ê ‰
	 @    " !r       Ê/ ‰ @    b !r
	
       Ê ‰
      " !r       Ê/ ‰       ¢ !r
       â M	        € ì !r       È/ x'7=  A  â x*BL?  A  Æ #t  €?   â !r        âO  x   ?  A  ä  r     A  â !r       Ê s     $   $ !t   €?    È  r       A  È  r      A  Ê †y    â My        € ê x     pRð Ú G‰  P    € ê x     pRð Ú G	  `    € ê „y      H   " z  r  p`ð â Ey       € ä „y    H   b r ÿ   pVp  Æ „y    H   ¨ „y    H   è „y	    H   ( „y  
  H   ( „y €  H   è „y €
  H   â !r         Æ „y €  H   " !r       Æ/ „y €   H   â !r         âO !r       Æ� !r	 	       â!r       È ‰ 	      h ‰      ¢ !r       â !r 	        â/ !r       ä !r       â !r       âO ‰       $ !r       Ä ‰      h „y €  H   â !r        â !r       È/ ‰ €    ( ‰ €    h „y €  H   ¢ !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       ÈO !r       â !r       â !r
	
       È/ ‰       ( ‰
      b !r        â !r 
       â/ G	  À     € ê !r         Ê	z  i    � â !r        È  x;ª¸¿  A  â 	z   i    � È 	z   i €  � â s        $ !t  €?    Ì s        $  r     A  È  r      A  Ê †y     ä Ay        € ê ‰       b Ey       € æ ‰      ¨ „y    H   è „y    H   ( „y    H   è „y    H   â !r         â/ !r       ÈO ‰       h ‰      ( „y  	  H   ¢ !r       â� !r       â!r        Ä/ !r       â !r       ä ‰ €    ( ‰ €    h „y    H   â !r       ÈO !r       â !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       â� !r       â x   ÿàÿ â !r
	
       Æ/ ‰       " z  r  p`ð Æ ‰
      b x    pVp  â !r        â !r 
       Ö/ G	  À     € ê !r         Ê	z  i    � â !r        È  x;ª¸¿  A  â 	z   i    � È 	z   i €  � â s        $ !t  €?    Ì s        $  r     A  È  r      A  Ê †y    ä Ay        € ê ‰       b Ey  Ð    € æ ‰      ¨ „y €  H   è „y €	  H   ( „y €  H   è „y €  H   â !r         Æ/ „y €  H   â !r       ÆO ‰       h ‰      " !r       È� !r       â!r        â/ !r       È ‰ €    ( ‰ €    h „y €  H   ¢ !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       âO !r       â x   ÿàÿ â !r
	
       Æ/ ‰       " z  r  p`ð Æ ‰
      b x    pVp  â !r        â !r 
       Ö/ G	  À     € ê !r         Ê	z  i    � â !r        È  x;ª¸¿  A  â 	z   i    � È 	z   i €  � â s        $ !t  €?    Ì s        $  r     A  È  r      A  Ê †y    ä Ay        € ê ‰      b !r       â x   ÿàÿ Æ !r       â z  r  p`ð Æ !r       â x    pVp  È ‰       ¢ !r       Ê/ ‰      " !r         ÊO ‰       b !r       Ê ‰ €    " !r        Ê/ ‰ €    b !r	       Ê ‰
	 @    " !r       Ê/ ‰ @    b !r
	
       Ê ‰
      " !r       Ê/ ‰       ¢ !r
       â M	        € ì !r       â/!r         ÈO 	z i    � â !r        È  x;ª¸¿  A  â 	z i    � È 	z i €  � â s        $ !t  €?    Ì s        $  r      A  È  r      A  Ê †y    â My        € ê „y     H   " z  r  p`ð Æ „y    H   b r ÿ   pVp  Æ „y    H   ¨ „y €   H   è „y €  H   ( „y    H   ( „y     H   ( „y"    H   ( „y    H   ( „y €  H   ( „y €  H   " !r       Æ „y%    H   " !r       Æ/ „y'  	  H   b !r       ÆO „y  
  H   ¢ !r       Æ� „y! €  H   â !r       Æ„y# €
  H   " !r       Æ „y
    H   " !r         Æ „y	    H   " !r""       Æ „y €  H   " !r       Æ „y €  H   " !r       Æ „y  €  H   " !r       Æ „y €	  H   â !r% %       Æ „y €  H   â !r""'       Æ/ „y €  H   â !r       ÆO ‰      " !r!       â� !r##       ä‰      b !r%
       Æ ‰
#      ¢ !r"	       Æ ‰	      â !r       Æ ‰      " !r       Æ ‰      " !r        â !r       â !r       â !r       â/ x   ÿàÿ â !r       ä ‰      " !r
#
       âO !r       â z  r  p`ô â ‰      b !r		       Ä� x    pVt â ‰      ¢ !r       â!r       ä ‰       è ‰	      ( ‰
      " !r       Æ ‰      " !r       Æ/ ‰      b !r       ÆO ‰      ¨ ‰      ¢ !r        â� !r	       ä‰  €    â !r
       Æ ‰	 €    " !r       Æ ‰ €    " !r
       Æ/ ‰ €    b !r       ÆO ‰ €    ¢ !r       Æ ‰
 €    ¨ ‰ €    ¢ !r         Æ� ‰ €    â !r		       â!r       ä ‰  @    " !r       Æ/ ‰ @    b x   ÿàÿ â !r       äO ‰	 @    ¢ z  r  p`ò â !r
       ä ‰ @    " x    pVò  â !r       Ä ‰
 @    " x   ÿàÿ â !r       ä� ‰ @    â z  r  p`ö Æ ‰ @    â !r        â x    pVö ä ‰ @    " !r       â/ !r	       äO ‰       b !r       Æ‰      ¢ !r

       Æ ‰	      " !r       Æ� ‰      â !r       Æ ‰
      â !r       Æ ‰      ( ‰      " !r         Æ/ ‰      b !r       âO !‚         â!r	       ä!‚       â !r       â� !’       ä  ‚      A  â !r

       â !’       Æ †‰     â !r       â !¢
       â  ’     A  ä !¢	       â !r       ä †™    â  ¢		     A  â !r        È/ †© 	   â M9        € ê !r       â !r         È  r      A  Ê †y    â My        € ê „y      H   " z  r  p`ð â Ey  À    € ä „y    H   b r ÿ   pVp  Æ „y    H   ¨ „y    H   è „y	    H   ( „y  
  H   ( „y €  H   è „y €
  H   â !r         Æ „y €  H   " !r       Æ/ „y €   H   â !r         âO !r       Æ� !r	 	       â!r       È ‰ 	      h ‰      ¢ !r       â !r 	        â/ !r       ä !r       â !r       âO ‰       $ !r       Ä ‰      h „y €  H   â !r        â !r       È/ ‰ €    ( ‰ €    h „y €  H   ¢ !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       ÈO !r       â !r       â !r
	
       È/ ‰       ( ‰
      b !r        â !r 
       â/ G	  à     € ê !r         Ê	x   à@  � Ê  x #ÛÙ?  A  È  x ;ª¸¿  A  â !r        Ê s        " 	x    à@  � È 	x    àÀ  � Ê !t   €?    â !t  €?    È s	        $  r	     A  È  r     A  Ê †y     ä Ay        € ê ‰       b Ey       € æ ‰      ¨ „y    H   è „y    H   ( „y    H   è „y    H   â !r         â/ !r       ÈO ‰       h ‰      ( „y  	  H   ¢ !r       â� !r       â!r        Ä/ !r       â !r       ä ‰ €    ( ‰ €    h „y    H   â !r       ÈO !r       â !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       â� !r       â x   ÿàÿ â !r
	
       Æ/ ‰       " z  r  p`ð Æ ‰
      b x    pVp  â !r        â !r 
       Ö/ G	  à     € ê !r         Ê	x   à@  � Ê  x #ÛÙ?  A  È  x ;ª¸¿  A  â !r        Ê s        " 	x    à@  � È 	x    àÀ  � Ê !t   €?    â !t  €?    È s	        $  r	     A  È  r     A  Ê †y    ä Ay        € ê ‰       b Ey  ð    € æ ‰      ¨ „y €  H   è „y €	  H   ( „y €  H   è „y €  H   â !r         Æ/ „y €  H   â !r       ÆO ‰       h ‰      " !r       È� !r       â!r        â/ !r       È ‰ €    ( ‰ €    h „y €  H   ¢ !r       â !r	       È/ ‰ @    ( ‰
	 @    b !r       âO !r       â x   ÿàÿ â !r
	
       Æ/ ‰       " z  r  p`ð Æ ‰
      b x    pVp  â !r        â !r 
       Ö/ G	  à     € ê !r         Ê	x   à@  � Ê  x #ÛÙ?  A  È  x ;ª¸¿  A  â !r        Ê s        " 	x    à@  � È 	x    àÀ  � Ê !t   €?    â !t  €?    È s	        $  r	     A  È  r     A  Ê †y    ä Ay        € ê ‰      b !r       â x   ÿàÿ Æ !r       â z  r  p`ð Æ !r       â x    pVp  È ‰       ¢ !r       Ê/ ‰      " !r         ÊO ‰       b !r       Ê ‰ €    " !r        Ê/ ‰ €    b !r	       Ê ‰
	 @    " !r       Ê/ ‰ @    b !r
	
       Ê ‰
      " !r       Ê/ ‰       ¢ !r
       â M	        € ì !r       â/!r         ÈO 	x  à@  � â !r        È  x#ÛÙ?  A  â 	x  à@  � Æ  x;ª¸¿  A  â 	x  àÀ  � Ê s        " !t   €?    â !t  €?    Ì s        $  r     A  È  r      A  Ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y       "   $ z   w  p`ð Ú M	        € ê y      !   " ¹z  l      â Ey  €    € â ™x?    â y      %   b ¹z  F   
   â $rÿÿ   ÿ Ž â r ÿ       ä xÿ     È |    p`ð â $x   ÿ Ž Ø/ G	  à    € ê y      &   " x        â $z  x  Ž È %v \   Ž Ê �y    ™ ¢  $r
ÿÿ   ÿ Ž â x   ÿÀŽ â $z  r  ÿŽ â x   ÿàÿ â 'z m  
 Ž È %x$   ÿ Ž â zÿ n    Æ $rÿÿ   ÿ Ž â r
ÿ   ÿáÿ â %x   ÿ Ž â x   ÿ   Æ $z

 o  Ž ä %x$    Ž È $z

 u  ÿŽ â z Z  ÿàñ È r
   à£ â%x

$   ÿ Ž â z	 [  ä  ä rÿÿ   ÿDþ  â %x$    Ž È $x$   ÿŽ â z Z  ÿàñ â $z q  Ž â r 
       â $rÿÿ    Ž â z	 [  ä  â $v
ÿ Y  ÿ Ž â x   ÿàñ ä z	  X      Æ $rÿÿ     ä $z t  Ž äO$z t  Ž â r ÿ       â $z q  Ž ä %x   Ž È %x   Ž Ê r	   ÿàó â�yðÿÿ ™ ¢  r	   ÿàñ Æ $x   
Ž  â r
   ÿä  â �y    ™ è  �y%   • ( �y   • ( �y$   • ¨
 �y#   • ¨
 �y!   • è �y   • è �y"   • è
 �y   • â
 r   ÿàñ Æ/ �y ôÿÿ ™ â  r   ÿä  Æ �y   ™ è  �y    ™ ä r 
       â/ $rÿÿ   	 Ž È %x   Ž È%x   Ž ä �y    • h �y    • b x   ÿàÿ È |    p`ð ä x		   ÿàõ ä x@  ÿà÷ ä x@  ÿàó â $r
ÿÿ   
 ä $rÿÿ   Ž â rÿ   ÿäÿ  Ä x%%T      Èx%ÿÀŽ ä x#$T  #    äO x%ÿ   %  ä x$#ÿÀŽ ä x#ÿ   #  ä x!T      â� &r   ÿ   âx%%ÿÀŽ â &r$$   ÿ   â x##ÿÀŽ Ä x""T      ä xÿÀŽ ä/ x"ÿÀŽ â &r%%      âxÿ     â &r##   $   â x"ÿ   "  â &r    %   âxÿÀŽ â &r#    #   â x"ÿÀŽ â 0rÿ  0 A   Ä &r      ä&r   #   â  x  €@  A  â 0rÿ    A   â Er        ä Er        Æ #r     ä#r     â 0rÿ    A   ä0rÿ    A   Æ #r      ä #r      â G‰  °ûÿÿÿÿƒ ê Ay        € ê ‰      " r   ÿàÿ Æ ‰      b z  p  p`ð È x    pDp  â !r       â !r       È/ ‰      ( ‰      b !r       â !r	       È/ ‰ €    ( ‰
	 €    b !r       â !r
	
       È/ ‰ @    ( ‰
 @    b !r       â !r
       È/ ‰      h  ‰      ¢  M	        € ê y      &   â x   ÿ   â $z   s  Ž â !r       â/ !r       âO r ÿ   p"ð â$t	ÿ   ÿ Ž â x    p"ò Ö         ä         â $z  v   Ž È� %v  j  	 Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      "   $ z  w  p`ð Ú M	        € ê y      &   " x	        â ¹z  F   
   â $z x  Ž È %v \  	 Ž Ê �y     ™ b ¹z  l      â Ey  `    € â ™x?    â y      !   b x      ÿ  â $rÿÿ   ÿ Ž â r  ÿ       â x      ÿ  â xÿ     È/ |    p`ð Ú G	  À
    € ê r
 ÿ       â y      %   " $z r  ÿŽ â x   ÿÀŽ â$rÿÿ   ÿ Ž â x   ÿ   â 'z m  
 Ž â r        â Ey 
    € ä %x   ÿ Ž â zÿ n    Æ $z  t  Ž ä$rÿÿ   
Ž È $z o  Ž ä %x
$   ÿ Ž â x   ÿ   Æ $z u  ÿŽ ä %x
$   
 Ž â x	   ÿàÿ ä r   à‘ â%x$   ÿ Ž Æ rÿÿ   ÿ$~  â %x$    Ž â z
 Z  ÿàñ Æ $x$   ÿŽ â z Z  ÿàó â $z q  Ž â z
 [  ä  â $z		 q  Ž â z [  äÿ  â $z  t  Ž â z ÿ `  pPð â $vÿ Y  ÿ Ž â x   ÿàó â %x   Ž â r        Ä z  X      â $rÿÿ   ÿ Ž â z ÿ a   Sð â %x	   Ž â z  `      ä z  a      ä rÿ   ÿäÿ  ä r   ÿàó È r   ÿäÿ  ä r$   ÿàó Æ �y&   • ¤  $x%   Ž  ä �y'   • ¨  �y)$   • è �y($   • è �y!   • ( �y*   • ( �y+$   • h �y,$   • b r        Ä r        ä r"   ÿàó Æ �yðÿÿ ™ h �y    ™ b $x#   
Ž  â r        ä �yôÿÿ ™ b r        Æ �y   ™ d %x   Ž ä�y""    ™ d %x	   Ž ä �y    • h �y    • b x&&T  '    ÈO x%&ÿÀŽ ä/ x()T  (    ä� x&ÿ   &  ä x#(ÿÀŽ ä x(ÿ   (  ä x!!T  *    Äx&&ÿÀŽ ä x((ÿÀŽ ä x++T  ,    äx!ÿÀŽ ä x+ÿÀŽ ä x!ÿ   !  â &r%%   ÿ   Ä &r##   ÿ   â x+ÿ   +  â &r%&   %   ä&r#(   #   â x$!ÿÀŽ â &r   %   âx&+ÿÀŽ â &r   #   ä &r$      Ä &r&      ä 0r!ÿ"  0 A   âEr        â 0r"ÿ"    A   â Er(        ä  x!!  €À  A  â 0r$ÿ    A   ä 0r#ÿ    A   â #r&"   !   â#r%"(   !   â G‰  °    € ê r   ÿàó Ê $x   Ž  Ê �y)    ¨ �y(    ¨ �y'    è �y*    â xÿ     Ê $z q  Ž È $z  t  Ž È %x   Ž Ì �y     " x()T  (    ÈO x)(ÿÀŽ ä x(ÿ   (  ä x''T  *    â� &r))   ÿ   â x((ÿÀŽ ä x'ÿÀŽ ä x'ÿ   '  â &r)(   )   Æ x('ÿÀŽ â &r   )   È &r(      Ê Er        â 0rÿ    A   È#r"   !   È #r        â #r$&      â #r#%      â G‰  €    € ê r   ÿàó È r   ÿäÿ  Ê �y$    ¨ �y%    ¨ �y#    è �y&    â %x	   Ž Ì �y     " x$$T  %    ÈO x%$ÿÀŽ ä x$ÿ   $  ä x##T  &    â� &r%%   ÿ   â x$$ÿÀŽ ä x#ÿÀŽ ä x#ÿ   #  â &r%$   %   Ä 0rÿ    A   âx#ÿÀŽ â &r   %   È &r      Ê Er        Ê #r"   !   È #r      â x   ÿàÿ ä x@  ÿàõ ä |    p`ò ä x   ÿà÷ â $rÿÿ    â x   ÿàû ä x@  ÿàù Ä rÿ   ÿäÿ ä rÿ   ÿä ä rÿ   ÿäÿ â G™   øÿÿÿÿƒ ê Ay       € ê $rÿÿ     Ž â r        ä Ay        € ê z ÿ `  pPð È z ÿ a   Sð Ú G	  P    € ê ‰      ( ‰      b !r       â !r	       È/ ‰      ( ‰
	      b !r       â !r	
       È/ ‰
 €    ( ‰ €    b !r

       â !r       È/ ‰
 @    ( ‰ @    b !r
       â !r       È/ ‰      ( ‰      b !r       â !r       â/ Gy  €    € ê ‰      ( ‰       h ‰	      ¨ ‰
      â !r       â !r        È/ ‰      " !r	       ÆO ‰	      b !r
       Æ� ‰
      ¨ ‰      â !r       â !r		       È/ ‰ €    " !r
       ÆO ‰
	 €    b !r       Æ� ‰ €    ¨ ‰ €    â !r       â !r
	
       â/ !r       ÆO ‰ @    " !r       Æ� ‰
 @    h ‰ @    ¨ ‰ @    â !r       â !r
       â/ !r	       ÆO ‰      " !r       Æ� ‰      h ‰	      ¨ ‰      â !r       â !r       â/ !r	       âO !r       â� y      %   $ r   ÿŽ È z  p  p`ð È x    pDp  Ú M	        € ê z ÿ ^  pPð â $z  v  Ž Æz ÿ _   Sð Ú $ÿ   ÿ Ž È % ^   Ž Ì �	     ¢ x    ÿ   È r  ÿ   p"ò äx     p"ô äx     p"ö äx     p"ø Î          ä z ÿ `  pPò â $" ÿÿ    Ž â 2         ä z ÿ a  Sò ä B         Ê !         ÌO G™  @    € ê z ÿ b  pPð È z ÿ c   Sð Ú $ÿ   ÿ Ž È % b   Ž Ì �	     ¢ x   ÿ   È x øÿÿÿp"ò äx üÿÿÿp"ô är ÿ   p"ö äx    p"ø Î         ä $"ÿÿ    Ž ä 2        ä z  h      â $Bÿÿ    Ž Æ x    pBò â !       ØO G  p    € ê xÿÿÿÿÿàÿ È x     € È x   ÿ   È ‚{  €     $ xÿ     È Iy àìÿÿÿÿƒ ê  x;ª¸¿  A  Ì s        $ !t  €?    Ì s        $  r     A  È  r       A  â Gy  P    € ê  x'7=  A  â x*BL?  A  Æ #t  €?   â  x   ?  A  Æ  r     A  Ì s     $   $ !t  €?    È  r     A  È  r       A  â Gy  °    € ê x    pRð Ú G‰  à     € ê x    pRð Ú G	        € ê 	z i    � ä 	z   i    � Æ  x;ª¸¿  A  â 	z   i €  � Ê s        $ !t  €?    Ì s        $  r     A  È  r       A  â Gy  Ð     € ê  r       A  â Gy  °     € ê 	x  à@  � ä 	x    à@  � Æ  x#ÛÙ?  A  â 	x    àÀ  � Æ  x;ª¸¿  A  ä !t   €?    È s        $ !t  €?    Ì s        $  r      A  È  r       A  ä y      "   ( y      &   b $z s  Ž â x        Æ $z v  Ž È/ %v j   Ž Ê †y      â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      "   " ¹z  l      â Ey  Ð    € â ™x?    â y      !   b ¹z  F   
   â r  ÿ       â x      ÿ  â y      %   ¢ r ÿ       â x      ÿ  â r ÿ       â x      ÿ  â x      ÿ  â x      ÿ  â x      ÿ  â r ÿ       Ä r ÿ   pRð är ÿ   pRò är   ÿ(Ž È/ r   ÿ� ä x   ÿ   äO x ÿ     È |     pbô Ú G)  `    € ê y      '   " r ÿ       â $z p   Ž â xþÿÿÿÿÀŽ â y	      &   b r ÿ       ä r
  €ÿàÿ ä x$        ä x
   ÿ   Ä x&    ÿ   ä x'ÿ   
  ä z*& q  ÿàÿ â 'z  y   Ž È 'z	 s   Ž â/ zÿ z     Æ $z!	 w  ÿŽ â z ÿ t    â $z |   Ž âz	  q      â $z |  Ž â x   ÿàÿ â %x   ÿŽ â r4	&   ÿŽ är(	&   ÿŽ â $z  v  Ž Ä $z  v  Ž ä $z p  Ž ä %x   Ž È $z  }  ÿŽ â z X  ÿàû â %x   Ž Æ x   ÿàý â %v" X  $Ž âz X  ÿàù ä zÿ Y  ¤~ â %v$ X  $Ž â r! !   
àµ ä x   ÿàû â $x,	   &Ž âr  ÿ       â $x.	   &Ž âzÿ Y  „þ â $x0	   &Ž âr'ÿÿ   'd~ â $x2	   &Ž Ä �y|úÿÿ • ¨ �y{üÿÿ • è �yWþÿÿ • ( �yI    • h �yFþÿÿ • ¨ �yE    • ¢ r&!   ÿàõ Æ �y>   • ¢ xq $       ä r	ÿ'   ÿä â �y=   • ¢ r*!   ÿàõ â %v Z  q Ž ä �y;"    • â r
ÿ'   ÿä â $x	$   ÿŽ Ä %v Z  q Ž â �yB$    • ä r   ÿàÿ â $x
$   ÿŽ È �y:   ™ " r		   ÿàÿ Æ �y6   ™ ( �y)   ™ ( �y5   ™ h �y3   ™ h �y7   ™ h �y+   ™ h �y1   ™ b r4!   ÿàõ Æ �y-   ™ h �y/   ™ b r
,!   ÿà÷ Æ �y8   ™ b rÿ'   ÿä Æ �y9   ™ b rGÿ'   ÿäÿ Æ �y<   ™ b %v Z  q Ž Æ �y?    ™ b $xK$   ÿŽ Æ �y@   ™ h �yD    • b %v

 Z  q Ž â rK   ÿàÿ ä �yA    ™ b $xGG$   ÿŽ Æ �yC    • d rG   ÿàÿ ä �yG   ™ h �yK   ™ h �yV   ™ h �yJ
   ™ h �yT   ™ h �yH
   ™ h �yR   ™ b r(!   ÿàõ Æ/ �yU
   ™ h �yP   ™ b r.!   ÿà÷ ä rÿ'   ÿä â�yS
   ™ h �yN   ™ b r[ÿ'   ÿäÿ â %v Z  q Ž Ä �yQ
   ™ d $x]$   ÿŽ ä �yL    ™ h �yX    • b %v Z  q Ž â r		]   ÿàÿ Ä �yO
   ™ b $x[[$   ÿŽ Æ �yM
   ™ d r[   ÿàÿ ä/ �y[   ™ h �yY
    ™ h �yZ
    • h �y^   ™ h �yi   ™ h �yg   ™ h �ye   ™ b r        Ä r
0!   ÿàõ â�yc   ™ b r2!   ÿà÷ ä rÿ'   ÿä â�ya   ™ b rmÿ'   ÿäÿ â %v

 Z  q Ž ä �y_    ™ d $xo$   ÿŽ Ä �yj    • b %v Z  q Ž Æ �y]   ™ b $xmm$   ÿŽ â ro   ÿàÿ ä �y\   ™ d rm   ÿàÿ ä �y
   ™ h# �y	   ™ h �yw
   ™ h �yv   ™ h �yh   ™ h �yu
   ™ h �yt   ™ h �yf   ™ h �ys
   ™ h �yr   ™ h �yd   ™ h �yq
   ™ h �yp   ™ h �yb   ™ h �yo
   ™ h �yn   ™ h �y`   ™ h �yk    ™ h �yl    • h �ym
   ™ h �yy
    ™ h �y   ™ h �yz    ™ h �y
    • h �yx    • b x||™  ÿ    âO ‚x ÿ      Æ x}ÿ   |  ä |       ä |~       ä x{™  ÿ    ä� x||ww  ÿÀŽ ä x}}ww  ÿÀŽ Ä xÿ     ä t|ÿ |    ä t}~ÿ }    ä |{       ä x
ww  ÿÀŽ ä/ xWW™  ÿ    Äxww  ÿÀŽ ä x|@Q  }    äx|bs  }    ä t|{ÿ 
    ä x
ÿ   W  ä t~ÿ     Ä x{I™  ÿ    äxI
ww  ÿÀŽ ä x
|@Q      äxWWww  ÿÀŽ ä x||bs      ä |€       Ä r {       ä tW~ÿ W    ä t~€ÿ I    ä x{ÿ     ä x}F™  ÿ    ä xIww  ÿÀŽ Ä |‚       ä x{{ww  ÿÀŽ ä xW@Q  ~    äxFWbs  ~    ä rW }       ä tI€ÿ I    Ä t€‚ÿ {    ä x{ÿ   W  ä xEE™  ÿ    ä xWWww  ÿÀŽ ä |„       ä x{{ww  ÿÀŽ Ä x~I@Q  €    äx€Ibs  €    ä rI E       ä tW‚ÿ W    ä t‚„ÿ {    ä x}>™  ÿ    Ä x{Iww  ÿÀŽ ä xIÿ   I  ä x>W@Q  ‚    äxEWbs  ‚    ä rW }       ä xIIww  ÿÀŽ Ä x==™  ÿ    ä x}ÿ   W  ä t{„ÿ {    ät„„ÿ I    ä xIÿ   =  ä |‚       Ä |†       ä xWWww  ÿÀŽ ä x}}ww  ÿÀŽ ä |ˆ       ä x==ww  ÿÀŽ ä |Š       Ä xIIww  ÿÀŽ ä tW‚ÿ W    ä t††ÿ }    ä t}ˆÿ =    ä x‚{@Q  „    ä tˆŠÿ I    Ä x„{bs  „    â &r{:>   ÿ   âx=W@Q  †    äxIWbs  †    â &rW:   ÿ   ä&r{6E   {   ä&rW6   W   ä &r6)   ÿ   â x:}@Q  ˆ    â&rW5
   W   â x†}bs  ˆ    â &rˆ)>   ÿ   Ä &r)3   6   ä &r67|   W   ä &r)+
   )   ä &r61   6   ä &r)-|   )   ä &r6/F   6   Ä &r)8   )   ä &r69~   6   ä &r)<F   )   ä &r33E   ˆ   ä &r6?€   6   ä &r)@~   )   Ä 0r;ÿ;    A   ä 0rDÿD    A   â Er6 6       â &r3+‚   3   ä &r+A€   )   â  r);D     A  â 0rCÿC    A   Æ #r)6      â &r6G   ÿ   â Er+ +       â &r3-„   3   ä &r)K   6   â  r-;C     A  Æ &r)V
   )   â #r-+      â &r+J   ÿ   ä &r)T|   )   ä &r+H   +   ä &r)R   )   ä &r+U
   +   Ä &r)PF   )   ä &r6S|   +   ä &r)N~   )   ä &r38=   3   ä &r6Q   6   ä &r)L€   )   Ä 0rXÿX    A   ä &r3<I   3   â Er) )       â &r6OF   6   â  r+;X     A  â &r3@:   3   ä &r6M~   6   â #r+)      â &r)[   ÿ   Ä &r3A†   3   ä &r6Y€   6   ä 0rBÿB    A   ä 0rZÿZ    A   ä &r)^   )   â Er3 3       â &r[[>   ÿ   â Er6 6       â  r8BC     A  â  r-;Z     A  â &r)i
   )   Ä &r[^E   [   â #r83      â #r-6      â &r8G>   ÿ   ä &r6g|   )   ä &r[i‚   [   ä &rKKE   8   ä &r6e   6   Ä &r8g„   [   ä &r6cF   6   ä &r8e=   8   ä &r6a~   6   ä &r8cI   8   ä &r6_€   6   Ä &r8a:   8   ä 0rjÿj    A   â Er6 6       â &r__†   8   Æ  r8;j     A  â &r+]   ÿ   Æ #r86      â &r6   ÿ   ä&r	   ÿ   ä &rJJ>   ÿ   ä&r]]>   ÿ   ä&r>   ÿ   ä&r		>   ÿ   Ä &r+\   +   ä&r6w   6   ä&rv      ä &rJHE   J   ä&r]\E   ]   ä&rwE      Ä &r	vE   	   ä &r+h
   +   ä&r6u
   6   ä&rt
      ä &r{5‚   {   ä&rKV‚   K   Ä &rJU‚   J   ä&r]h‚   ]   ä&ru‚      ä&r	t‚   	   ä &r+f|   +   ä&r6s|   6   Ä &rr|      â x@  ÿàõ â &rˆ7„   {   ä&rKT„   K   ä&rJS„   J   ä&r]f„   ]   ä&rs„      Ä &r	r„   	   ä &r+d   +   ä&r6q   6   ä&rp      â rÿ   ÿä â &rˆ1=   ˆ   ä&rKR=   K   Ä &rJQ=   J   ä&r]d=   ]   ä&rq=      ä&r	p=   	   â x@  ÿàõ â &r+bF   +   âx      ÿàÿ â &r6oF   6   Ä &rnF      ä &rˆ/I   ˆ   ä&rKPI   K   ä&rJOI   J   ä&r]bI   ]   ä&roI      Ä &r	nI   	   â rÿ   ÿä â &r+`~   +   â|     pbô â &r6m~   6   ä&r~      ä &rˆ9:   ˆ   ä&rKN:   K   Ä &rJM:   J   ä&r]`:   ]   ä&rm:      ä&r	:   	   ä &r+k€   +   ä&r6y€   6   Ä &r€z€      ä &rˆ?†   ˆ   ä&rKL†   K   ä&rJY†   J   ä&r]k†   ]   ä&ry†      Ä &rzz†   	   ä 0rlÿl    A   ä 0rÿ    A   ä 0rxÿx    A   â Erˆ ˆ       â  rDDB     A  â ErK K       â  rXBX     A  âErJ J       â  rZBZ     A  âEr_ _       â  r)Bj     A  âEr+ +       â  r<;l     A  âEr] ]       â  r-Bl     A  âEr6 6       â  r;     A  âEr        â  r	B     A  âEr€ €       â  r;x     A  â Erz z       â  rBx     A  â x""@  ÿà÷ Ä x$$@  ÿàù â #rDˆ      â #rXK      â #rZJ      â #r)_      â #r<+      â #r-]      â #r6      â #r	      â #r€      â #r z       â x&&@   ÿàÿ Ä x**@   ÿàÿ ä x44@   ÿàÿ ä x,,@   ÿàÿ ä x((@   ÿàÿ ä x..@   ÿàÿ ä x00@   ÿàÿ Ä x22@   ÿàÿ ä r#ÿ#   ÿäÿ ä r%ÿ%   ÿä â D)       À â Gy  pæÿÿÿÿƒ ê Ay        € ê Ey  0    € â G‰      € ê r   ÿHŽ Ê ˆs  øÿ H   è ˆs €øÿ H   è ˆs  ùÿ H   è ˆs €ùÿ H   è ˆs  úÿ H   è ˆs €úÿ H   è ˆs  ûÿ H   è ˆs €ûÿ H   è ˆs  üÿ H   è ˆs €üÿ H   è ˆs  ýÿ H   è ˆs €ýÿ H   è ˆs  þÿ H   è ˆs €þÿ H   è ˆs  ÿÿ H   è ˆs  €ÿÿ H   ä Ay        € ê {          ì M        € ê „y     H   b z  r  p`ð Ä z8  r      â „y €   H   ¢ r ÿ   p"p ä r488   ÿŽ â„y	    H   â r688   ÿŽ Æ y
      &   ( y      '   h „y    H   ( „y    H   ( „y €  H   ( „y €  H   è „y" €  H   â $z
 x  Ž âx   ÿàÿ â !r       â/ „y$    H   â $z ~  Ž â!r       ÄO ‰      " !r	       Æ� „y(    H   h „y&    H   è „y% €  H   ¢ !r       Æ „y' €  H   â 4   ÿàù â !r        ä „y)    H   â ÿÿ   ÿä â !r       â !r       â „y+ €  H   â 
 j  ÿˆ â !r       Ä ‰      " 
 k   â !r""       ä ‰      â !r$       Æ ‰      ( ‰      b !r-(       Æ/ ‰!      h ‰      b !r%%       âO !r       ä ‰#       " r8   ÿ† â !r''       â !r
       â� ‰	 €    ¢ 
 r  ÿàó â !r))       â r*ÿÿ   ÿäÿ â ‰
      â ,ÿÿ   ÿäÿ  â !r       âz  r  p`ò â !r&       â „y5 €  H   â z j  ÿ† â !r       â
 j  ÿ„ â ‰&      " !r++       â z k  *� â !r!!       â/ ‰"      b 
 k  , â !r       â !r# #       â ‰(%      " 6   ÿà÷ â !r	       ÄO ‰      ¢ $x8   ÿŽ â ÿÿ   ÿäÿ â !r
       â� ‰ @    â r8   ÿ„ är	   ÿàù â ‰*      b r$ÿÿ   ÿä ä z
 j  ÿ„ â ‰,'      b !r&&       ârÿÿ   ÿä Ä ‰      " z	 j  ÿˆ ä z k  $ â ‰.)      " !r$"       â/ !r(%(       â 
 j  ÿ† â ‰0+      " !r       âO z		 k   â !r55        â !r       â� ‰2-      b 
 k  � â !r**       Ä‰ €    ¢ x    p"ò â !r,',       ä ‰!      â !r       Æ‰      " !r.).       Æ ‰ #      d 8   ÿàÿ â !r0+0       â ‰      $    ÿàõ â !r2-2       â/ ‰%$      d :ÿÿ   ÿä â !r       âO ‰'&      ¢ 4   ÿàÿ â !r!!       Ä� ‰)(      â r88  €ÿŽ â !r       ä‰ €    "    ÿàù â !r##        ä‰+*      b 4ÿÿ   ÿä â !r9       Ä ‰-,      " 
 j  ÿˆ ä    ÿà÷ â ‰ €    " !r$$%       â/ 
 k  4 ä ‰/.      b !r&&'       âO 8ÿÿ   ÿäÿ Ä ‰10      ¢ !r(()       â�  j  ÿ† ä‰32      â !r       â k  8� ä ‰5      " !r**+       âx8        Ä ‰ @    b !r,,-       â  j  ÿ„ ä‰ ! €    " !r       â  k  : ä ‰ €    " !r../       Æ/ ‰"# €    b !r001       âO !r223       ä� ‰'$ €    ¢ !r45       Æ‰)& €    â !r        Æ‰+( €    " %v j  8 Ž â 6   ÿàÿ â !r !        â ‰ @    $    ÿàõ â !r       â ‰-* €    d 6ÿÿ   ÿä â !r"#"       â/ ‰/, €    b  j  ÿ„ Æ ‰ @    b !r'$'       âO  k  6 ä ‰1. €    ¢ !r)&)       Æ� ‰30 €    â !r+(+       Æ‰52 €    " !r       Æ ‰74      " !r-*-       Æ‰       b !r/,/       Æ/ ‰#  @    b !r       Æ ‰! @    b !r1.1       ÆO ‰%" @    ¢ !r303       Æ� ‰$' @    â !r525       â!r747       ä ‰&) @    " !r;        Æ‰(+ @    " !r# #       Æ/ ‰       â !r!!       Æ ‰*- @    b !r%"%       ÆO ‰ / @    ¢ !r$'$       Æ� ‰      è ‰"1 @    b !r&)&       Æ ‰,3 @    " !r(+(       Æ‰07 €    ( ‰.5 @    " !r*-*       Æ/ ‰!      b !r /        ÆO ‰#      ¢ !r)       Æ� †	 9    â !r"1"       â!r,3,       ä † ;   â !r070       Æ‰$      " !r9        â� !r.5.       ä ‰ %      â !r+!       Æ/ ‰&      b !r-#       ÆO ‰(      ¨ †	 9    è ‰*      " !r$       Æ ‰       ( † )    â !r%%        Æ� ‰!"      â !r&       Æ/ ‰#,      b !r(       ÆO ‰'0 @    ¨ ‰.      b !r*       Æ†	 +    â	 !r        Æ † -   è	 †	 %    â	 !r!"!       â� !r#,#       ä/ †     â	 !r'0'       ÆO †	 
    â	 !r.       Æ† 
   è	 †	     è	 †     è	 †	 !    è	 † #   è	 ‰ '      ( †	     â	 M™        € ê x   ÿàÿ â !r''        Ær   ÿàñ È rÿÿ   ÿä  ä z j  ÿ€ È z k    Ê †y '    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      "   " ¹z  l      â Ey  €    € â ™x?    â y      !   b ¹z  F   
   â x      ÿ  â r ÿ       â y       %   ¢ r, ÿ       â x      ÿ  â r  ÿ       â x      ÿ  â x      ÿ  â x      ÿ  â r ÿ       Ä r ÿ   pRð är ÿ   pRò är   ÿ(Ž È/ r   ÿ� ä x     ÿ   äO xÿ     È |    pbô Ú G)       € ê y      '   " r ÿ       ä xþÿÿÿÿÀŽ â y	      &   b r ÿ       ä r  €ÿàÿ â $z  p  Ž â x$        ä x3   ÿ   Ä r ÿ       ä z13 q  ÿàÿ â 'z y   Ž È 'z	 s   Ž â/ zÿ z    ä x   ÿ   â $z"	 w  ÿŽ â zÿ t    â $z |  Ž ä$z |  Ž â x    ÿàÿ â %x   ÿŽ È $z v  Ž ä$z v  Ž ä $z% p  Ž ä %x   Ž È %x%   Ž â z X  ÿàû Æ $z }  ÿŽ â x   ÿàý â %v( X  $Ž â z X  ÿàù ä r""   àµ â %v$% X  $Ž â r* (       ä zÿ Y  ¤~ ä z  q      Ä x   ÿàû ä x(ÿ     â $x+   3Ž âzÿ Y  „þ â $x'   3Ž âr!3   ÿŽ â$x/   3Ž âr-3   ÿŽ ä r(ÿÿ   (d~ Ä �y4úÿÿ • ¨ �yBüÿÿ • è �y>þÿÿ • ( �y&    • h �y:þÿÿ • " r3"   ÿàõ Æ �y5    • " xc $       ä rÿ(   ÿä â �y=   • $ %v Z  c Ž ä �y9   • d $x	$   ÿŽ â r!"   ÿà÷ â �y#$    • f r	   ÿàÿ Ê �y;   ™ h �y6   ™ h �y?   ™ h �y7   ™ h �y.   ™ b r
1"   ÿàõ Æ �y8   ™ b r *       ä r	 )       â �y<   ™ b rÿ(   ÿä ärÿ(   ÿäÿ â �y@    ™ b %v

 Z  c Ž Æ �y0    • h �y2    • b $x$   ÿŽ ä $x	$   ÿŽ ä/ %v Z  c Ž Æ r	   ÿàÿ ä r   ÿàÿ Æ �yE
   ™ h �yC   ™ h �yD
   ™ h �y   ™ h �yA
   ™ h �yN
   ™ h �yL
   ™ b r+"   ÿàõ Æ �yJ
   ™ b r-"   ÿà÷ ä rÿ(   ÿä â�yH
   ™ b rÿ(   ÿäÿ â %v Z  c Ž ä �yF
    ™ d $x$   ÿŽ Ä �yP
    • b %v Z  c Ž ä r		   ÿàÿ â �yO   ™ b $xS$   ÿŽ â r        ä �yM   ™ d rS   ÿàÿ Ä �yS   ™ h �yK   ™ h �yI   ™ h �yG   ™ h �yQ    ™ h �yR    • h �yT   ™ h �y_   ™ h �y   ™ h �y   ™ h �y]   ™ h �y^   ™ b r
'"   ÿàõ Æ/ �y[   ™ h �y\   ™ b r/"   ÿà÷ Æ �yY   ™ b rÿ(   ÿä Æ �yZ   ™ b rdÿ(   ÿäÿ Æ �yW   ™ b $xe$   ÿŽ Æ �yX   ™ b %v

 Z  c Ž Æ �yU    ™ b %v Z  c Ž Æ �yV   ™ b $xcd$   ÿŽ Æ �y`    • b re   ÿàÿ Æ �ya    ™ b rc   ÿàÿ Æ �ym
   ™ h �y
   ™ h �y	   ™ h �yl   ™ h �yk
   ™ h �yj   ™ h �yi
   ™ h �yh   ™ h �yg
   ™ h �yf   ™ h �ye
   ™ h �yd   ™ h �yb    • h �yc
   ™ h �yo
    ™ h �y   ™ h# �y
    • h �yp    ™ h �yn    • b x44™  ÿ    ÄO xBB™  ÿ    ä� xqÿ   4  â ‚x ÿ      â r
 B       ä |s       ä |r       ä x44ww  ÿÀŽ Ä xqqww  ÿÀŽ ä xÿ   
  ä/ t4sÿ 4    ä tqrÿ q    ä xB>™  ÿ    äx>ww  ÿÀŽ Ä |       ä x

ww  ÿÀŽ ä x4@Q  q    äx4bs  q    ä r4 B       ä tÿ 
    Ä x
ÿ   4  ä t>sÿ >    ä xr&™  ÿ    ä|q       ä x44ww  ÿÀŽ ä xB
ww  ÿÀŽ Ä x
@Q  >    äx&bs  >    ä r> r       ä t4qÿ 4    ä tqsÿ B    ä xB>ww  ÿÀŽ Ä x::™  ÿ    ä x>ÿ   >  ä x4@Q  q    äxq4bs  q    ä r4 :       ä x>>ww  ÿÀŽ Ä x55™  ÿ    ä tBsÿ B    äx:ÿ   4  ä |u       ä x44ww  ÿÀŽ ä tssÿ >    Ä r> 5       ä |w       ä x::ww  ÿÀŽ ä t5uÿ 4    ä x4ÿ   >  ä t:wÿ :    Ä xr=™  ÿ    ä |t       ä x>>ww  ÿÀŽ ä x=4ww  ÿÀŽ ä x45@Q  :    äx55bs  :    Ä t:uÿ >    ä t=tÿ =    ä xtrww  ÿÀŽ ä xrÿ   r  ä x>:@Q  =    äx::bs  =    â &r=;   ÿ   âxrrww  ÿÀŽ â &r;;4   ÿ   â xv9™  ÿ    â &r=6   =   ât9uÿ r    â &r665   ;   â rr v       â &r=?
   =   â ttuÿ t    â &r6?>   6   â x?ÿ   r  Ä x;t@Q  9    â&r=7&   =   âx9tbs  9    â &r67:   6   â |t       ä xrrww  ÿÀŽ ä x??ww  ÿÀŽ â &r=.   =   âtruÿ r    â &r7.;   6   â t?tÿ ?    â &r=8q   =   â x6B@Q  s    â &r789   7   â x.r@Q  ?    Ä xBBbs  s    â &r=<6   =   â x8rbs  ?    â &r7<.   7   ä &r=@B   =   ä0r0ÿ0    A   ä 0r2ÿ2    A   ä &r@@8   7   Ä 0r#ÿ#    A   â  r702     A  ä Er@ @       ä  r22#     A  È #r2@      â &r2E   ÿ   ä&r<C   ÿ   ä&r@C4   ÿ   ä &r2D   2   ä&r<   <   ä&r@5   @   Ä &rA
   2   È &rN&      È &rL      È &rJq      È &rH6      â Er= =       â 0rPÿP    A   ä &rFB      ä #r7=      â  r70P     A  ä Er        Ê #r7      â &rS   ÿ   ä&r2   ÿ   ä&rT      ä&r74   ÿ   ä &r_
      ä &r   2   Ä &r75   7   ä &r]&      ä &r^
      ä &r[      ä &r\&      ä &rYq      Ä &rZ      ä &rW6      ä &rXq      ä &rUB      ä &rV6      ä 0r`ÿ`    A   â Er        â &raB      Æ  r0`     A  â &rEE4   ÿ   Æ #r      â &r   ÿ   ä&rSS4   ÿ   ä&r4   ÿ   ä &r	   ÿ   ä&r		4   ÿ   ä &r<O
   <   Ä &rED5   E   ä&rST5   S   ä &rm      ä&rm5      ä &rl      ä&r	l5   	   Ä &r<M&   <   ä &rEA>   E   ä&r@O>   @   ä&rS_>   S   ä&r7^>   7   ä &rk
      Ä &rk>      ä &rj
      ä&r	j>   	   ä &r<K   <   ä &rEN:   E   ä&r@M:   @   Ä &r2]:   S   ä&r7\:   7   ä &ri&      ä&ri:      ä &rh&      ä&r	h:   	   Ä &r<Iq   <   ä &rEL;   E   ä&r@K;   @   ä&r2[;   2   ä&r7Z;   7   ä &rg      Ä &rg;      ä &rf      ä&r	f;   	   â x@  ÿàõ â &r<G6   <   â x    ÿàÿ â &rEJ9   E   ä&r@I9   @   Ä &r2Y9   2   ä&r7X9   7   ä &req      ä&re9      ä &rdq      ä&r	d9   	   â rÿ   ÿä â &r<QB   <   â |    pbô â &rEH.   E   Ä &r@G.   @   ä0rRÿR    A   ä &r2W.   2   ä&r7V.   7   ä &rc6      ä&rc.      Ä &r6      ä&r	.   	   â Er< <       â &rEF8   E   â  r=0R     A  â &r@Q8   @   ä&r2U8   2   ä&raa8   7   Ä &roB      ä&ro8      â x@  ÿà÷ â 0rbÿb    A   ä 0rÿ    A   ä &rBpB      ä&r8p8   	   Ä 0rnÿn    A   â #r=<      â rÿ   ÿäÿ â  rP#P     A  âErE E       â  rR#R     A  âEr@ @       â  r`#`     A  â r70b     A  âEr2 2       â  r=#b     A  âEr        â  r0     A  âEra a       â  r	#     A  âEr        â  r00n     A  â Er        â  r##n     A  â ErB B       Ä Er8 8       ä x**@  ÿà÷ ä x$$@  ÿàù â #rPE      â #rR@      â #r`2      â #r7      â #r =a       â #r,   ,   â #r	      â #r0B      â #r#8      â x!!@   ÿàÿ Ä x++@   ÿàÿ ä x--@   ÿàÿ ä x''@   ÿàÿ ä x//@   ÿàÿ ä x11@   ÿàÿ ä x33@   ÿàÿ Ä r)ÿ)   ÿäÿ ä r%ÿ%   ÿä â D)       À â Gy  °èÿÿÿÿƒ ê Ay        € ê Ey      € â G‰  ð     € ê $xÀ  Ž Ê ˆs  ùÿ H   è ˆs €ùÿ H   è ˆs  úÿ H   è ˆs €úÿ H   è ˆs  ûÿ H   è ˆs €ûÿ H   è ˆs  üÿ H   è ˆs €üÿ H   è ˆs  ýÿ H   è ˆs  €ýÿ H   è ˆs , þÿ H   è ˆs €þÿ H   è ˆs  ÿÿ H   è ˆs €ÿÿ H   ä Ay        € ê {          ì M        € ê „y     H   b z   r  p`ð Ä z+  r      â „y    H   ¢ r ÿ   p"p ä r/++   ÿŽ â„y €  H   â r)++   ÿŽ Æ „y €   H   ( „y	    H   h „y    H   ( „y €  H   ( „y    H   ( „y €  H   è „y    H   â !r       Æ/ „y# €  H   â !r       ÆO ‰      " !r       Æ� „y €  H   â !r       Æ„y"    H   â !r		       Æ„y% €  H   â !r       Æ y      &   b !r       Æ ‰      ¢ !r       â !r       ä ‰      " !r!        Æ ‰      â !r,,       Æ y      '   " !r##       Æ ‰      b $z x   Ž Æ/ ‰      b !r       ÆO ‰
      ¢ !r       â !r$"       ä ‰	      ¢ !r       â !r&%       â !r       â� ‰      " x    ÿàÿ â $z ~  Ž Ä‰      â z  r  p`ò â !r       ä‰ !      " 
  r  ÿàõ â !r       ä/ ‰ €    b ÿÿ   ÿä â !r

       ÄO ‰,      ¢ 
  j  ÿ„ â!r	       ä ‰      b 
  k   ä x    p"ò â ‰"#      b !r       Æ ‰%$      " !r       Æ� ‰	      â !r!!        Æ‰      " !r        â/ r+   ÿˆ ä ‰'&      b !r,       âO r.ÿÿ   ÿä ä ‰
      ¢ !r       â)   ÿàû Ä ‰      b !r##"       â z j  ÿˆ ä‰      b !r%$%       â z k  . ä ‰       " !r	       â� r	+   ÿ„ â !r       â‰"      â !r'&'       Æ/ ‰$!      b !r
       ÆO ‰  @    ¢ /   ÿà÷ â !r       ä‰&      " ÿÿ   ÿäÿ â !r       ä ‰ €    b 

 j  ÿ† ä rÿÿ   ÿä â ‰(#      b !r         â ÿÿ   ÿäÿ Ä ‰*%      " z	 j  ÿ„ â !r""       â� 
 j  ÿŠ â ‰ €    â !r$!$       â/ 
 k  � ä ‰,'      b !r        âO z		 k   Ä ‰ €    ¢ !r&&       â
 k  � ä ‰ €    " !r       â+   ÿàÿ ä ‰ €    b !r(#(       â    ÿàõ Ä ‰ €    b !r*%*       â 2ÿÿ   ÿä ä ‰!  €    " !r       â� )   ÿàÿ â !r,',       â/ ‰#" €    d 0   ÿà÷ â !r       âO ‰%$ €    ¢ $x+   ÿŽ â 1ÿÿ   ÿäÿ â !r       â‰       â 0 j  ÿ† â !r       Ä‰'& €    " r   ÿàù â !r       ä ‰ @    b r.ÿÿ   ÿä ä z j  ÿˆ â ‰)( €    b !r! !       â  j  ÿ„ Ä ‰+* €    " z k  . â !r#"#       â/ 0 k  1� â ‰ @    b !r%$%       âO  k  2 ä ‰ @    ¢ !r1        â� /   ÿàÿ Ä ‰-, €    â !r'&'       Æ‰ @    " !r        â   ÿàõ ä ‰ @    b !r)()       â x        ä ‰ @    b !r+*+       â .ÿÿ   ÿä â %v j   Ž â ‰ ! @    " !r       â/  j  ÿ„ Ä ‰"# @    b !r       âO  k  . ä ‰$% @    ¢ !r-,-       Æ� ‰&' @    â !r       Æ‰       " !r       Æ‰() @    b !r       Æ ‰*+ @    b !r !        Æ ‰      " !r"#"       Æ/ ‰,- @    b !r$%$       ÆO ‰      ¢ !r&'&       Æ� ‰      â !r/        Æ‰      " !r()(       Æ‰      b !r*+*       Æ ‰!       b !r3       Æ ‰#"      " !r,-,       Æ/ ‰%$      b !r)       ÆO ‰&      ¢ !r       Æ� †	 1    â !r       Æ† /   â !r       Æ‰'(      â !r! !       Æ ‰*      " !r#"#       Æ ‰,      " !r%$%       Æ/ †	 3    â !r&       ÆO † )    è †	     è †    â !r'('       Æ� †	     â !r*       Æ† !    â !r,       Æ †	 #    è † %   è †	 
    è † '    è †	     â M™        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      "   " ¹z  l      â Ey       € â ™x?    â y      !   b ¹z  F   
   â x      ÿ  â r ÿ       â x      ÿ  â r% ÿ       â x      ÿ  â r  ÿ       â x      ÿ  â r ÿ       Ä r ÿ   pRð är ÿ   pRò är   ÿ(Ž È/ r   ÿ� È xÿ     È |    pbô Ú G)  p    € ê y      '   " xþÿÿÿÿÀŽ ä r ÿ       â y      %   b r
  €ÿàÿ ä r ÿ       â y	      &   ¢ x
   ÿ   ä x        Ä x*   ÿ   ä x!ÿ   
  ä r ÿ       ä z(* q  ÿàÿ â 'z y   Ž â x   ÿ   È/ zÿ z    â 'z	 s   Ž ÈO $z p  Ž â zÿ t    â $z |  Ž â x   ÿàÿ â $z |  Ž ä %x   ÿŽ È $z v  Ž ä$z v  Ž ä %x   Ž È $z p  Ž â z X  ÿàû â $z }  ÿŽ â z  q      â %x   Ž â x   ÿàý ä r*   ÿŽ â $z	 w  ÿŽ â z X  ÿàù â %v" X  Ž â zÿ Y  ¤~ Ä r   
àµ â %v X  Ž â x   ÿàû ä r$ "       â $x"   *Ž âzÿ Y  „þ â $x&   *Ž âr *   ÿŽ ä r!ÿÿ   !d~ Ä �ycúÿÿ • ¨ �yRüÿÿ • è �ySþÿÿ • ( �yQ    • h �yPþÿÿ • ( �yO    • h �yN   • b r*   ÿàõ Æ �yM   • b xY $       ä rÿ!   ÿä â �y:    • b r(   ÿàõ â %v Z  Y Ž Æ r
ÿ!   ÿä â $x$   ÿŽ ä %v Z  Y Ž Æ r		   ÿàÿ â $x
$   ÿŽ È �yL   ™ b r   ÿàÿ Æ �yI   ™ h �y9   ™ h �y;   ™ h �yH   ™ h �yG   ™ h �yF   ™ h �yE   ™ b r   ÿàõ Æ �yD   ™ b r
 $       Æ �yC   ™ b r #       ä r+ÿ!   ÿä â �yB   ™ h �yA   ™ b $x++$   ÿŽ Æ �y@   ™ h �y<
    • h �y?    ™ h �y)    • b %v
 Z  Y Ž Æ �y>   ™ d r+   ÿàÿ ä �y'    • h �y-
   ™ b r"   ÿà÷ Æ �y,
   ™ b rÿ!   ÿäÿ Æ �y8
   ™ b %v Z  Y Ž Æ �y=    ™ b $x$   ÿŽ Æ �y6
   ™ h �y4
   ™ b r   ÿàÿ ä r    ÿàõ â/�y2
   ™ b r&   ÿà÷ ä rÿ!   ÿä â�y+   ™ b rUÿ!   ÿäÿ Æ �y0
   ™ b %v Z  Y Ž Æ �y   ™ b $xW$   ÿŽ Æ �y.
    ™ b %v Z  Y Ž Æ �yJ
    • b $xUU$   ÿŽ â r		W   ÿàÿ ä �y7   ™ d rU   ÿàÿ ä �y_   ™ h �y
   ™ h# �y   ™ h �y^   ™ h �y\   ™ h �y]   ™ h �yZ   ™ h �y5   ™ h �y[   ™ h �yX   ™ h �y3   ™ h �yY   ™ h �yV   ™ h �y1   ™ h �yW   ™ h �y/   ™ h �yK    ™ h �yT    • h �yU   ™ h �ya    ™ h �y   ™ h �y    • h �y`    • h �yb    ™ b ‚x ÿ      Ä |f       ä |e       ä |	       ä/ xcc™  ÿ    ÈO xdÿ   c  ä xR™  ÿ    ä� xccww  ÿÀŽ ä xddww  ÿÀŽ ä xÿ     ä xww  ÿÀŽ Ä xSS™  ÿ    ätcfÿ c    ä tdeÿ d    ä xRww  ÿÀŽ ä t		ÿ     ä xÿ   S  Ä xc@Q  d    äxcbs  d    ä tReÿ R    ä xcQ™  ÿ    äxQww  ÿÀŽ ä |d       Ä xSSww  ÿÀŽ ä x	@Q  R    äx		bs  R    ä tRfÿ Q    ä xQÿ   c  ä tSdÿ S    Ä xgP™  ÿ    ä xdQww  ÿÀŽ ä xPS@Q  R    äxQSbs  R    ä rR g       ä xccww  ÿÀŽ Ä xSÿ   R  ä tdeÿ d    ä tcfÿ c    ä xeO™  ÿ    ä xORww  ÿÀŽ ä |h       Ä xSSww  ÿÀŽ ä xRc@Q  d    äxdcbs  d    ä tOfÿ O    ä rc e       ä tfhÿ S    Ä xeN™  ÿ    ä xScww  ÿÀŽ ä xNO@Q  f    äxOObs  f    ä xcÿ   c  ä xfÿ   e  Ä xgM™  ÿ    ä xccww  ÿÀŽ ä |l       ä xMfww  ÿÀŽ ä tShÿ S    ätfhÿ c    Ä thlÿ M    ä xMÿ   g  ä |j       ä xeeww  ÿÀŽ ä xcMww  ÿÀŽ ä tejÿ e    Ä xMS@Q  f    äxSSbs  f    ä xggww  ÿÀŽ ä tflÿ c    ä xce@Q  h    äxeebs  h    â &rhL   ÿ   â tgjÿ g    â &rjLN   ÿ   Ä &rhI   h   ä&rjIO   j   ä &rI9   ÿ   ä&r99N   ÿ   ä &rh;   h   ä &rIH   I   Ä &r9HO   9   ä &rhG	   h   ä &rIF   I   ä&r9FM   9   ä &rhEP   h   ä &rID	   I   Ä &r9DS   9   ä &rhCQ   h   ä &rIBP   I   ä&r9Bc   9   ä &rhAR   h   â xLg@Q  f    â &rI@Q   I   Ä &r9@e   9   ä &rh?d   h   ä 0r<ÿ<    A   ä 0r)ÿ)    A   ä &rI>R   I   âErh h       â &r>>L   9   Ä &rj;M   j   ä 0r9ÿ'    A   â  r'<)     A  â &rjGS   j   Æ #r'h      â &r'-   ÿ   ä&r--N   ÿ   ä &rjEc   j   ä &r-,O   -   ä &rjCe   j   â xfgbs  f    â &r-8M   -   Ä &rjAL   j   ä &rI=d   I   ä &r-6S   -   ä &rj?f   j   ä&r==f   >   ä 0r:ÿ:    A   Ä &r-4c   -   â Erj j       ä Er= =       â  r>):     A  â  r@:9     A  â &r-2e   -   ä #r>j      â #r@=      â &r>+   ÿ   Ä &r@+N   ÿ   ä &r-0L   -   ä &r>   >   ä&r@O   @   ä &r-.f   -   ä 0rÿJ    A   Ä &r',   '   â Er- -       ä  r,:     A  È #r,-      â &r,
   ÿ   ä&r   ÿ   ä&r

N   ÿ   ä&rN   ÿ   ä &r,_   ,   ä&r^      Ä &r
_O   
   ä&r^O      ä &r\      ä&r'8   '   ä&r>7   >   ä&r@7M   @   Ä &r,]   ,   ä&r
]M   
   ä&r\M      ä &rZ	      âx@  ÿàõ â &r'6	   '   ä&r>5	   >   Ä &r@5S   @   ä &r,[	   ,   ä&r
[S   
   ä&rZS      ä &rXP      ârÿ   ÿä â &r'4P   '   Ä &r>3P   >   ä&r@3c   @   ä&r,YP   ,   ä&r
Yc   
   ä&rXc      â x@  ÿàõ â &rVQ      âx    ÿàÿ â &r'2Q   '   Ä &r>1Q   >   ä&r@1e   @   ä &r,WQ   ,   ä&r
We   
   ä&rVe      â rÿ   ÿä â &r	R      â|    pbô â &r'0R   '   âErI I       â &r>/R   >   Ä &r@/L   @   ä &r,UR   ,   ä&r
UL   
   ä&rL      â  r)<9     A  â 0r+ÿT    A   ä 0rÿ    A   Ä &r'.d   '   ä&r>Kd   >   ä&r@Kf   @   ä &r,ad   ,   ä&r
af   
   ä 0rÿ`    A   Ä &r	bd   	   ä&rbf      â #r)I      â  r)<     A  â r<+     A  â r<     A  âEr' '       â  r<<     A  â Er> >       â  r+:+     A  âEr@ @       â  r:     A  âEr, ,       â  r:     A  â Er
 
       Ä Er	 	       ä Er        ä x$$@  ÿà÷ ä x@  ÿàù â #r)'      â #r>      â #r +@       â #r%,   %   â #r
      â #r<	      â #r      â x@   ÿàÿ Ä x""@   ÿàÿ ä x  @   ÿàÿ ä x&&@   ÿàÿ ä x((@   ÿàÿ ä x**@   ÿàÿ ä r#ÿ#   ÿäÿ Ä rÿ   ÿä â D)       À â Gy  pëÿÿÿÿƒ ê Ay        € ê Ey  ð     € â G‰  Ð     € ê $x€  Ž Ê ˆs  úÿ H   è ˆs €úÿ H   è ˆs  ûÿ H   è ˆs €ûÿ H   è ˆs  üÿ H   è ˆs €üÿ H   è ˆs  ýÿ H   è ˆs  €ýÿ H   è ˆs % þÿ H   è ˆs €þÿ H   è ˆs  ÿÿ H   è ˆs €ÿÿ H   ä Ay        € ê {          ì M        € ê „y     H   b Ey  @    € æ „y €   H   ¨ „y    H   â !r       â/ !r       È_ ‰      " !r       Æ� ‰      h ‰      ¢ !r       Æ y      %   " !r       Æ/ ‰      b !r	       ÆO ‰      ¨ ‰	      â !r       â/ x   ÿ   â !r       ÆO ‰ €    " z  r  p`ð â !r
	       ä� ‰ €    b r ÿ   pVp  Æ ‰
 €    ´ ‰      '   â ˆ        â !r       â x   ÿàÿ â !r       Æ/ ‰ @    " z  r  p`ò â !r
       äO ‰	 @    b x    p"ò Æ ‰ @    ¨ ‰      &   ( „y
 €  H   è       &   b !r       Æ       '   " !r		       Æ/ ‰      b !r       ÆO ‰	      ¢ $Š x  Ž Æ‰      " $Š ~  Ž â�         â $ x  Ž È$ ~  Ž â !r       â/ %† j   Ž â !r	       ÆO % j   Ž â †‰     â !r       Æ†    â G	  �     € ê y      &   ( y      '   b $z x  Ž È $z ~  Ž Ê/ z r  ÿàõ È rÿÿ   ÿä ä z j  ÿ„ È z k   Ê †y     è Ay        € ê !r

       â „y    H   â Ey  Ð    € æ ‰
      ( „y €  H   è „y	    H   è „y €  H   è „y €  H   è „y    H   â !r
       Æ „y
    H   è ‰      $ !r       Ê ‰ €    $ !r       Ê ‰ @    $ !r       Ê ‰      $ !r       â G™  °     € ê y      &   " ¹z  r      ä �x   ?àÿ â y      '   b $z x  Ž È $z ~  Ž Ê/ |   ÿàõ È rÿÿ   ÿä ä z j  ÿ„ È z k   Ê †y     è Ay        € ê „y €  H   â !r       â !r       â !r		       â !r        â !r%
%       â !r       â !r       â ‰       " z  r      Æ ‰      b r   ÿŽ Æ ‰	      ¨ ‰
      è ‰%      ( ‰      h ‰      b !r         â !r       Ä/ y      &   " !r	       âO !r       ä ‰       b !r

       Æ� ‰      ¢ !r%       Æ‰      â !r       Æ‰	
      " !r       Æ ‰      b $z x  Ž Æ ‰      ( ‰      " !r        Æ/ ‰      b !r       âO !r       ä� ‰  €    ¢ !r	
	       Æ‰ €    â !r       Æ‰ €    " !r       Æ ‰
	 €    " !r       Æ ‰ €    b !r       Æ/ ‰ €    h ‰ €    b !r         ÆO ‰      ¢ !r       Æ� y      '   â !r       â!r
	
       ä ‰  @    " !r       Æ‰	 @    â !r       Æ/ ‰ @    â !r       Æ ‰
 @    b !r       ÆO ‰ @    ¢ $z ~  Ž Æ� ‰ @    â r   ÿŽ Æ ‰ @    " ‚   ÿà÷ â !r         â r   ÿ„ â‰ €    " ‚ÿÿ   ÿäÿ ä rÿÿ   ÿä ä Š j  ÿ† ä z j  ÿ„ â !r
       â/ Š k  � â !r	       â z k   â !r       â !r       âO ‰	       b r   ÿ„ â !r       â�    ÿàÿ â ‰      ¢ rÿÿ   ÿä â !r       â   ÿàõ â‰      â ‚   ÿàù â !r       â  ÿÿ   ÿä â ‰      "  j  ÿ„ Ä z j  ÿ† â ‰      " ‚ÿÿ   ÿä ä Š
 j  ÿˆ â ‰      b z k  � ä Š k   â ‰ @    b !r 	       â/ 	 k    Ä ‰      b !r       âO !r       ä� †‰     â !r       Æ †    â !r       Æ†‰     â !r       Æ†     â !r       Æ †‰     â !r       Æ/ †    è ‰       ( †‰ 
    â M™        € ê x   ÿàÿ âO !r        Æ r   ÿàñ È rÿÿ   ÿä  ä z j  ÿ€ È z k    Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      "   " ¹z  l      â Ey  À    € â ™x?    â y      !   b ¹z  F   
   â x      ÿ  â r ÿ       â x      ÿ  â r$ ÿ       â x      ÿ  â r  ÿ       Ä r ÿ       ä r ÿ   pRð är ÿ   pRò är   ÿ(Ž È/ r   ÿ� È xÿ     È |    pbô Ú G)       € ê y      '   " xþÿÿÿÿÀŽ ä r ÿ       â y      %   b r
  €ÿàÿ ä r ÿ       â y	      &   ¢ x
   ÿ   ä x        Ä x'   ÿ   ä r ÿ       ä z%' q  ÿàÿ â 'z y   Ž â x   ÿ   Æ/ $z }  ÿŽ â zÿ z    â 'z	 s   Ž ÈO $z p  Ž â zÿ t    â $z |  Ž â x   ÿàÿ â $z |  Ž ä %x   ÿŽ È $z v  Ž ä$z v  Ž ä %x   Ž È $z p  Ž â z X  ÿàû â %v  X  Ž â z  q      ä x   ÿàý â %x   Ž â r"         ä zÿ Y  ¤~ â $z	 w  ÿŽ â z X  ÿàù â %v X  Ž â x ÿ   
  Ä r   
àµ â $x   'Ž â x   ÿàû ä r'   ÿŽ äzÿ Y  „þ ä r#'   ÿŽ ä r ÿÿ    d~ Ä �y*úÿÿ • ¨ �y7üÿÿ • è �y4þÿÿ • ( �y    • h �y0þÿÿ • " r'   ÿàõ Æ �y+    • b xJ $       ä rÿ    ÿä â �y3   • d %v
 Z  J Ž ä �y/   • d $x$   ÿŽ â r   ÿà÷ â �y    • f r   ÿàÿ Ê �y1
   ™ h �y,
   ™ h �y5
   ™ h �y-
   ™ h �y&
   ™ b r	%   ÿàõ Æ �y.
   ™ b r "       ä r !       â �y2
   ™ b r8ÿ    ÿä ärÿ    ÿäÿ â �y6
    ™ b %v	 Z  J Ž Æ �y(    • b $x98$   ÿŽ Æ �y)
    • d r		9   ÿàÿ â %v Z  J Ž È/ �y8   ™ b $x$   ÿŽ â r        ä �y<   ™ d r   ÿàÿ ä �y:   ™ h �y;   ™ h �y9   ™ h �yF   ™ h �yE   ™ h �yC   ™ b r   ÿàõ Æ �yA   ™ b r#   ÿà÷ ä rÿ    ÿä â�y?   ™ b rKÿ    ÿäÿ â %v
 Z  J Ž ä �y=    ™ d $xM$   ÿŽ Ä �yG    • b %v Z  J Ž Æ �yD   ™ b $xKK$   ÿŽ â rM   ÿàÿ ä �yB   ™ d rK   ÿàÿ ä �y
   ™ h# �y	   ™ h �yT
   ™ h �yS   ™ h �yR
   ™ h �yQ   ™ h �yP
   ™ h �yO   ™ h �yN
   ™ h �yM   ™ h �y@   ™ h �yL
   ™ h �yK   ™ h �y>   ™ h �yH    ™ h �yI    • h �yJ
   ™ h �yV
    ™ h �y   ™ h �y
    • h �yU    • h �yW    ™ b ‚x ÿ      Ä |[       ä |Y       ä |
       ä/ |]       ä x**™  ÿ    äO x77™  ÿ    Ä� xXÿ   *  ä r 7       ä xXXww  ÿÀŽ ä xÿ     ä t[ÿ X    ä xX4™  ÿ    Äx**ww  ÿÀŽ ä x4ww  ÿÀŽ ä xww  ÿÀŽ ä |7       ä t*Yÿ *    ä tY
ÿ     Ä x
ÿ   X  ä t47ÿ 4    ä x™  ÿ    äx*@Q      äx*bs      ä x[
ww  ÿÀŽ Ä |*       ä xXXww  ÿÀŽ ä x4@Q  Y    äx
4bs  Y    ä r4        ä t77ÿ X    Ä t**ÿ [    ä |[       ä x4ww  ÿÀŽ ä xZ0™  ÿ    ä x0ÿ   4  ä xY7@Q  *    Ä x77bs  *    ä t[ÿ     ä |*       ä x[ÿ   Z  ä x00ww  ÿÀŽ ä x4+™  ÿ    Ä |X       ä xZZww  ÿÀŽ ä x[[ww  ÿÀŽ ä t+*ÿ 0    ä r* 4       ä t]]ÿ Z    Ä tXXÿ [    ä x[ÿ   *  ä |4       ä x33™  ÿ    ä xZ*ww  ÿÀŽ ä x*]@Q  X    Ä |0       ä x[[ww  ÿÀŽ ä x]]bs  X    ä t44ÿ Z    ä xX3ww  ÿÀŽ ä xZÿ   3  Ä t\0ÿ [    ät00ÿ X    â &r31   ÿ   â|X       ä xZZww  ÿÀŽ â &r11*   ÿ   â x//™  ÿ    ä tXXÿ Z    â &rZ,   3   â x[4@Q  \    â &r,,]   1   â r1 /       â &rZ5   Z   âx44bs  \    â &r,5[   ,   â x\ÿ   1  â &rZ-
   Z   â x/0@Q  X    Ä x30bs  X    ä xX1ww  ÿÀŽ â &r,-4   ,   â |5       ä |0       ä x1\ww  ÿÀŽ â &rZ&Y   Z   ât55ÿ X    â &r,&/   ,   â t00ÿ 1    â &rZ.7   Z   â x1@Q  +    â &r,.3   ,   â x-5@Q  0    Ä x+bs  +    â &rZ21   Z   âx55bs  0    â &r,2-   ,   ä &rZ6+   Z   ä0r(ÿ(    A   ä 0r)ÿ)    A   ä &r,65   ,   â ErZ Z       â 0rÿ    A   â  r()     A  Ä Er, ,       ä  r&)     A  â #rZ      â &r8   ÿ   ä #r&,      â &r&;   ÿ   ä &r<      ä&r&9   &   Ä &r:      ä&r)F   &   ä &r&E
      È &r&CY   &   È &r&A7   &   È &r&?1   &   ä 0rGÿG    A   ä &r&=+   &   Æ  r(G     A  ä Er& &       â &r88*   ÿ   ä&r,;*   ÿ   ä #r&      â &r   ÿ   ä&r*   ÿ   Ä &r	   ÿ   ä&r*	*   ÿ   ä &rT      ä &r8<]   8   ä&r,9]   ,   ä&rT]      Ä &rS      ä&r*S]   *   ä &rR      ä &r8:[   8   ä&r9F[   ,   ä&rR[      Ä &rQ      ä&r*Q[   *   ä &r)D
   )   ä&rP
      ä &r8E4   8   ä&r9D4   9   Ä &rP4      ä &rO
      ä&r*O4   *   ä &r)BY   )   ä&rNY      ä &r8C/   8   Ä &r9B/   9   ä&rN/      ä &rMY      ä&r*M/   *   â x@  ÿàõ â &r)@7   )   âx    ÿàÿ â &rL7      Ä &r8A3   8   ä&r9@3   9   ä&rL3      ä &rK7      ä&r*K3   *   â rÿ   ÿä â &r)>1   )   â|    pbô â &rJ1      Ä &r8?-   8   ä&r9>-   9   ä&rJJ-      ä &r1      ä&r*-   *   ä 0rIÿI    A   Ä &r)H+   )   ä 0rÿ    A   â x@  ÿà÷ â &r8=5   8   ä&rHH5   9   ä &rV+      ä&rJV5   J   Ä 0rUÿU    A   ä &rW+      ä&r*W5   *   â  r9(I     A  â r	(     A  ârÿ   ÿäÿ â  r((U     A  â Er8 8       â  r,G     A  âEr) )       â  rII     A  âErH H       â  r     A  âEr        â  rUU     A  â ErJ J       Ä Er        ä Er* *       ä x""@  ÿà÷ ä x@  ÿàù â #r,8      â #r9)      â #r IH       â #r$	   $   â #rJ      â #r(      â #rU*      â x@   ÿàÿ Ä x@   ÿàÿ ä x##@   ÿàÿ ä x%%@   ÿàÿ ä x''@   ÿàÿ ä r!ÿ!   ÿäÿ ä rÿ   ÿä â D)       À â Gy  0íÿÿÿÿƒ ê Ay        € ê Ey  Ð     € â G‰  °     € ê $x@  Ž Ê ˆs  ûÿ H   è ˆs €ûÿ H   è ˆs  üÿ H   è ˆs €üÿ H   è ˆs  ýÿ H   è ˆs  €ýÿ H   è ˆs $ þÿ H   è ˆs €þÿ H   è ˆs  ÿÿ H   è ˆs €ÿÿ H   ä Ay        € ê {          ì M        € ê „y     H   b Ey  @    € æ „y €   H   ¨ „y    H   â !r       â/ !r       È_ ‰      " !r       Æ� ‰      h ‰      ¢ !r       Æ y      %   " !r       Æ/ ‰      b !r	       ÆO ‰      ¨ ‰	      â !r       â/ x   ÿ   â !r       ÆO ‰ €    " z  r  p`ð â !r
	       ä� ‰ €    b r ÿ   pVp  Æ ‰
 €    ´ ‰      '   â ˆ        â !r       â x   ÿàÿ â !r       Æ/ ‰ @    " z  r  p`ò â !r
       äO ‰	 @    b x    p"ò Æ ‰ @    ¨ ‰      &   ( „y
 €  H   è       &   b !r       Æ       '   " !r		       Æ/ ‰      b !r       ÆO ‰	      ¢ $Š x  Ž Æ‰      " $Š ~  Ž â�         â $ x  Ž È$ ~  Ž â !r       â/ %† j   Ž â !r	       ÆO % j   Ž â †‰     â !r       Æ†    â G	  �     € ê y      &   ( y      '   b $z x  Ž È $z ~  Ž Ê/ z r  ÿàõ È rÿÿ   ÿä ä z j  ÿ„ È z k   Ê †y     è Ay        € ê !r

       â „y    H   â Ey  À    € æ ‰
      ( „y	 €  H   è „y    H   è „y €  H   è „y    H   è „y €  H   â !r
       Ê ‰      $ !r       Ê ‰ €    $ !r       Ê ‰ @    $ !r       Ê ‰      $ !r       â G™  °     € ê y      &   " ¹z  r      ä �x   ?àÿ â y      '   b $z x  Ž È $z ~  Ž Ê/ |   ÿàõ È rÿÿ   ÿä ä z j  ÿ„ È z k   Ê †y     è Ay        € ê !r       â !r		        â !r$       â !r       â !r       â ‰       b !r       â z  r      ä ‰	      ( ‰      ¨ ‰      è ‰
      ( ‰      b !r         Æ/ y      &   b !r	       Æ ‰       " !r       ÆO ‰      ¢ !r       Æ� ‰      â !r

       Æ‰	      " !r       Æ‰
      h ‰      b $z x  Ž â/ !r        ä y      '   " !r       ÆO ‰  €    b !r       Æ� ‰ €    ¢ !r		       Æ‰ €    â !r
       Æ‰	 €    " !r       Æ ‰
 €    h ‰ €    b $z ~  Ž â !r         â/ !r       ÈO ‰  @    " !r       Æ� ‰ @    b !r	       Æ‰ @    ¢ !r
       Æ‰ @    â !r       Æ ‰ @    ( ‰ @    b !r	        â r   ÿ„ â !r
       Æ/ ‰ 	      " r   ÿŽ â !r       âO rÿÿ   ÿä ä z j  ÿ„ â!r       â� r   ÿˆ â ‰
      b z k   â !r       â‚   ÿàõ â ‰      ¢    ÿàÿ â !r       â‚ÿÿ   ÿä â ‰      â Š j  ÿ„ Ä    ÿà÷ â ‰      " Š k   â !r	        â rÿÿ   ÿä â ‰      " ÿÿ   ÿäÿ ä  j  ÿ„ â †‰     â z j  ÿˆ Ä 	 k   â !r
       â/ z k   â !r       ÆO †    â !r       Æ� †‰     â !r       Æ†     â !r       Æ †‰     â M™        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      "   " ¹z  l      â Ey      € â ™x?    â y      !   b ¹z  F   
   â r ÿ       â x      ÿ  â r ÿ       ä r ÿ       Ä r  ÿ       ä r  ÿ       ä r ÿ       ä r ÿ   pRð är ÿ   pRò är   ÿ(Ž È/ r   ÿ� È xÿ     È |    pbô Ú G)  ð    € ê y      '   " xþÿÿÿÿÀŽ ä r ÿ       â y      %   b r
  €ÿàÿ ä r ÿ       â y	      &   ¢ x
   ÿ   ä x        Ä x#   ÿ   ä r ÿ       ä z!# q  ÿàÿ â 'z y   Ž â x   ÿ   È/ zÿ z    â 'z	 s   Ž ÈO $z p  Ž â zÿ t    â $z |  Ž â x   ÿàÿ â $z |  Ž ä %x   ÿŽ È $z v  Ž ä$z v  Ž ä %x   Ž È $z p  Ž â z X  ÿàû â $z }  ÿŽ ä %x   Ž â x   ÿàý Æ %v X  Ž â z X  ÿàù ä zÿ Y  ¤~ â $z	 w  ÿŽ â r        â %v X  Ž â z  q      ä r   
àµ äx   ÿàû â $x   #Ž â xÿ   
  Ä zÿ Y  „þ ä r#   ÿŽ ä rÿÿ   d~ ä �yKúÿÿ • ¨ �yJüÿÿ • è �yIþÿÿ • ( �yH    • h �yGþÿÿ • ( �y&    • b r#   ÿàõ Æ �y+   • b x= $       ä rÿ   ÿä â �y*   • b r!   ÿàõ â %v Z  = Ž Æ rÿ   ÿä â $x$   ÿŽ È $x	$   ÿŽ â r   ÿàÿ â %v Z  = Ž È �y(   ™ b r	   ÿàÿ Æ �y'   ™ h �y)   ™ h �y$   ™ h �y-   ™ h �y,   ™ b r6   ÿàõ Æ �y"   ™ h �y%   ™ b r        Æ �y2   ™ b r	        â $r8ÿÿ    ä �y1   ™ b r
        ä r        â �y0   ™ b r9   ÿàõ Æ �y/   ™ b r:ÿ   ÿä â $x;8$   ÿŽ ä �y    • h �y.    ™ h �y3
    • b %v6 Z  = Ž Æ �y4   ™ h �y5    • b %v
9 Z  = Ž â/ r		;   ÿàÿ ä �y7   ™ b $x9:$   ÿŽ Æ �y?   ™ d r9   ÿàÿ ä �y   ™ h �y9   ™ h �yA
   ™ h �y   ™ h �y@
   ™ h �y>
   ™ h �y=   ™ h �y<
   ™ h �y;   ™ h �y:
   ™ h �y6    • h �y8    ™ h �yE
   ™ h �yD
   ™ h �y   ™ h! �y    ™ h �yB    • h �yC
    ™ h �yF
    • b ‚x ÿ      Ä |L       ä |M       ä |N       ä |P       ä |Q       ä |T       Ä |V       ä |S       ä xKK™  ÿ    ÈO xÿ   K  ä xKKww  ÿÀŽ ä xww  ÿÀŽ ä x	J™  ÿ    ä� tKLÿ K    ä tMÿ     Ä xLI™  ÿ    äxÿ   	  ä/ x
	ww  ÿÀŽ ä x	K@Q      äxKbs      ä rK L       Ä xIww  ÿÀŽ ä |J       ä tMÿ 
    ä x
ÿ   K  ä tJJÿ I    ä xKKww  ÿÀŽ Ä xI
ww  ÿÀŽ ä xLH™  ÿ    äx
@Q  J    äxHbs  J    ä tKNÿ K    ä tJPÿ I    Ä xILww  ÿÀŽ ä xOG™  ÿ    ä xK@Q  J    äxGKbs  J    ä tKNÿ I    ä rI O       Ä xLÿ   L  ä xJÿ   I  ä xN&™  ÿ    ä xLLww  ÿÀŽ ä |O       ä x&Jww  ÿÀŽ Ä tLMÿ L    ä xIIww  ÿÀŽ ä tMOÿ &    ä x&ÿ   N  ä tJPÿ I    ä xNNww  ÿÀŽ Ä xI&ww  ÿÀŽ ä x&K@Q  L    äxKKbs  L    ä tLOÿ N    ä x++™  ÿ    ä tOPÿ I    Ä xIJ@Q  M    äxJJbs  M    ä xM*™  ÿ    â $r*ÿÿ   + Ž â xNL@Q  O    È x+ÿ   *  ä xLLbs  O    ä xPÿ   M  ä x**ww  ÿÀŽ ä x++ww  ÿÀŽ ä xOMww  ÿÀŽ Ä xRPww  ÿÀŽ ä tMQÿ *    ä tPTÿ +    â &r+(	   ÿ   â tQVÿ O    â &rO(I   ÿ   â x*M@Q  P    äx(Mbs  P    â &rM'   +   Ä &rO'J   O   ä &rM)
   M   ä&rO)N   O   ä &r)$	   ÿ   ä &rO-L   O   ä &r),   )   â tRSÿ R    â &rM-H   M   Ä &rO"*   O   ä&r)%
   )   â x+Q@Q  R    â&rM"   M   ä &rO2(   O   ä &r"1H   )   â x'Qbs  R    â &rO0+   O   Ä &r"/   "   ä &rO.'   O   ä &r"4G   "   ä 0r5ÿ5    A   ä 0r3ÿ3    A   â ErO O       â &rQ$I   ÿ   Ä &r$7&   "   â  r"53     A  È #r"O      â &r"	   ÿ   ä&r	A	   ÿ   ä &r"9   "   ä &rI   ÿ   ä&rAAI   ÿ   ä &r"
   "   Ä &r	@   	   ä &rQ,J   Q   ä&r9J      ä&rA@J   A   ä &r"?H   "   ä &r	>
   	   Ä &rQ%N   Q   ä&rN      ä&rA>N   A   ä &r"=   "   ä &r	<H   	   â x€  ÿàõ â &rQ1L   Q   Ä &r?L      ä&rA<L   A   ä &rM2G   M   ä&r";G   "   ä &r:   	   â rÿ   ÿä â &rQ/*   Q   Ä &r=*      ä&rA:*   A   â x€  ÿàõ â &rM0&   M   âx@   ÿàÿ â &r"&   "   ä &rGEG      ä &r44(   Q   Ä &r;(      ä&rAE(   A   â rÿ   ÿä â &rM.K   M   â|    pbô â 0rÿ    A   ä &r"K   "   ä &r&D&   G   Ä &r47+   4   ä0rBÿB    A   ä &r+      ä&rDD+   A   â ErM M       â  r%5     A  â 0r6ÿ6    A   â Er" "       â &r$8K   $   â  rB     A  â &r48'   4   Ä &r'      ä &r&CK   &   ä0rFÿF    A   ä &rDC'   D   â #r%M      â  r%6     A  âEr$ $       â #r "       â Er4 4       â  rF     A  â Er        â  r)36     A  âEr& &       â  r3B     A  âErD D       â  rF3F     A  â x€  ÿà÷ Ä x€  ÿàù â #r%$      â #r )4       â #r      â #r&      â #rFD      â x€   ÿàÿ ä x€   ÿàÿ ä x!!€   ÿàÿ Ä x##€   ÿàÿ ä rÿ   ÿäÿ ä rÿ   ÿä â D)       À â Gy  Ðïÿÿÿÿƒ ê Ay        € ê Ey  °     € â G‰  �     € ê r   ÿ@Ž Ê ˆs  üÿ H   è ˆs €üÿ H   è ˆs  ýÿ H   è ˆs  €ýÿ H   è ˆs   þÿ H   è ˆs €þÿ H   è ˆs  ÿÿ H   è ˆs €ÿÿ H   ä Ay        € ê {          ì M        € ê „y €   H   b Ey       € æ „y     H   ¨ „y €  H   è „y    H   ( „y	 €  H   h „y    H   ( „y    H   ( „y    H   ( „y  	  H   ( „y €  H   â !r       â/ !r       ÆO !r       â� !r       Æ!r		       â!r       È ‰	      " !r       Æ ‰      b !r       Æ „y €	  H   â !r       Ê ‰      ¢ !r	       Æ y	      %   " !r       Æ/ ‰      h ‰      â !r       ÊO ‰      ¢ !r       â/ !r       È� ‰ €    h ‰ €    â !r       ÊO ‰
 €    ¢ !r       â/ x	   ÿ   â !r       Æ� ‰ @    " x   ÿàÿ äz  r  p`ð â ‰ @    b z  r  p`ò ä r ÿ   pVp  äx    pVò  â !r

       ÄO „y €  H   è ‰
 @    ¨ ‰      &   è ™      &   " !r       â !r       Ä/ ‰      '   ( ™      '   h ‰      b !r

       ÆO ‰      ¢ ˆ        â $Š x  Ž ä�‰	
      â $š x  Ž ä$Š ~  Ž â ˜        â $š ~  Ž â/ !r       â%† j   Ž â !r       ÆO %– j   Ž â !r	
	       â� †‰     è †™    â G	  �     € ê y      &   ( y      '   b $z x  Ž È $z ~  Ž Ê/ z r  ÿàõ È rÿÿ   ÿä ä z j  ÿ„ È z k   Ê †y 	    è Ay        € ê !r        â „y    H   â Ey  0    € ä !r       â „y    H   æ !r       â „y	  
  H   è ‰       ( „y €  H   è „y €  H   è „y
    H   è „y    H   è „y €  H   â !r         Æ „y €  H   è ‰       ( „y €
  H   â !r        Ê ‰ €    $ !r       ä „y    H   è ‰ @    $ !r       Ê ‰       $ !r        â G  °     € ê y      &   " ¹z  r      ä �x   ?àÿ â y       '   b $z x  Ž È $z   ~  Ž Ê/ |     ÿàõ È rÿÿ   ÿä ä z  j  ÿ„ È z  k   Ê †y     è Ay        € ê „y €  H   â !r        â !r       â !r

       â !r       â !r       â !r       â !r

       â !r       â !r	       â !r       â !r

       â y      '   ( ‰      h ‰      ¨ ‰
      â !r       â/ !r       â !r       ÆO ‰       b !r	
       Æ� ‰      ¨ ‰      è ‰	      " !r         â/ !r       ÈO ‰  €    b !r       Æ� ‰ €    ¢ !r
	       Æ‰      è ‰
 €    ( y	      &   b !r        â/ !r       ÈO ‰ @    b !r       Æ� ‰ @    ¢ !r
       Æ‰ €    â $z	 x  Ž Æ‰ @    " $z ~  Ž â !r        â/ !r       âO z  r      â !r       Æ� ‰	      " $x
   ÿŽ âr   ÿ„ â !r       â‰       d ‚
   ÿà÷ ä ‰ @    ¢ rÿÿ   ÿä ä z j  ÿ„ â ‰      â ‚ÿÿ   ÿäÿ Ä Š j  ÿ† ä z k   ä Š k  � â !r		       â !r        È/ †™ 	   â !r       ÆO †‰     â !r       Æ� ‰       h  †‰     â M        € ê x
   ÿàÿ â !r        Æ/ r   ÿàñ È rÿÿ   ÿä  ä z j  ÿ€ È z k    Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      "   " ¹z  l      â Ey  �    € â ™x?    â y      !   b ¹z  F   
   â r ÿ       â $r ÿÿ   ÿ Ž â r  ÿ       â x      ÿ  â r ÿ       Ä r ÿ   pRð är ÿ   pRò är   ÿ(Ž È/ r   ÿ� È xÿ     â $À   Ž Æ |    pbô Ú G)  €    € ê y      '   " xþÿÿÿÿÀŽ â $rÿÿ   ÿ Ž â r ÿ       â y      %   b x        â $x
   
Ž â x   ÿ   â y	      &   ¦ xÿ   
  Ä z q  ÿàÿ â 'z y   Ž â x   ÿ   È/ zÿ z    â 'z	 s   Ž âO x
   ÿ   Æ $z p  Ž â zÿ t    â $z |  Ž ä$z |  Ž â x   ÿàÿ â %x   ÿŽ È $z v  Ž ä$z v  Ž ä $z p  Ž ä %x   Ž È $z }  ÿŽ â z X  ÿàû â %x   Ž Æ x   ÿàý â $z	 w  ÿŽ â z X  ÿàù â %v X  Ž âzÿ Y  ¤~ ä r   
àµ â $v	ÿ q  ÿ Ž â x   ÿàû â %v X  Ž â r        Ä zÿ Y  „þ ä r ÿ       ä r	   ÿŽ ä rÿÿ   d~ ä �yúÿÿ • ¨  �yüÿÿ • è  �y'þÿÿ • ( �y)    • h �y$þÿÿ • ( �y    • b r   ÿàõ â $t5ÿ$   ÿ Ž Ä �y&   • b rÿ   ÿä â %v Z  5 Ž ä �y#   • d $x	$   ÿŽ â r
        â $rÿÿ    Ž â �y,    • d r	   ÿàÿ Ä r   ÿàõ â �y*
    • h �y%   ™ h �y   ™ b r.ÿ   ÿä Æ �y(   ™ b r0   ÿàõ Æ �y"   ™ b %v Z  5 Ž â r1ÿ   ÿä ä �y!   ™ b $x3.$   ÿŽ Æ �y+   ™ d r		3   ÿàÿ â %v
0 Z  5 Ž â �y-   ™ f $x11$   ÿŽ â �y0   ™ f $x   1Ž â �y.    ™ h �y/    • h �y1   ™ h �y;
   ™ h �y:   ™ h �y9
   ™ h �y8   ™ h �y7
   ™ h �y6   ™ h �y5
   ™ h �y4   ™ h �y3
   ™ h �y2   ™ h �y
   ™ h# �y    • h �y>    ™ h �y=
   ™ h �y<
    ™ h �y?
    • b ‚x ÿ      Ä |A       ä |C       â $~	ÿ   ÿ Ž â x€  ÿàõ ä x@   ÿàÿ ä rÿ   ÿä ä x€  ÿàõ È rÿ   ÿä ä |    pbô ä x€  ÿàù ä x€  ÿà÷ ä x™  ÿ    ÈO x@ÿ     ä x™  ÿ    ä� xww  ÿÀŽ ä x@@ww  ÿÀŽ ä x
ÿ     ä/ tAÿ     Ä tCÿ @    ä x''™  ÿ    äxww  ÿÀŽ ä x@
ww  ÿÀŽ ä x@Q      äx
bs      Ä r '       ä tAÿ     ä xAÿ     ä t		ÿ @    ä |'       ä |       Ä x@ww  ÿÀŽ ä xAAww  ÿÀŽ ä x))™  ÿ    ät''ÿ @    ä tÿ A    ä xDÿ   )  â $~Aÿ   ÿ Ž â x@Q  	    Ä x	bs  	    ä xB$™  ÿ    ä x@)ww  ÿÀŽ ä x$'@Q      äx'bs      ä |'       Ä xDDww  ÿÀŽ ä x)™  ÿ    ä tAÿ @    ä rA B       ä t'ÿ D    â $r'ÿÿ   ) Ž â x@ÿ   A  Ä |B       ä xAAww  ÿÀŽ ä xDÿ   '  ä x@@ww  ÿÀŽ ä tBBÿ A    ä |A       Ä |)       ä x''ww  ÿÀŽ ä xDDww  ÿÀŽ ä tCCÿ @    ä x&&™  ÿ    ä t'Aÿ '    Ä t))ÿ D    ä x@B@Q  C    äxDÿ   &  ä xA'@Q  )    äxBBbs  C    ä x)'bs  )    â &r'%   ÿ   â|C       â &r%%@   ÿ   â xE&ww  ÿÀŽ Ä xDDww  ÿÀŽ â &r%B   %   â t&Cÿ E    ä x##™  ÿ    â &r'
   '   â tCCÿ D    â &r%(A   %   âxEÿ   #  â &r'(   '   â xD&@Q  C    â$~(ÿ   ÿ Ž â x&&bs  C    Ä xC#ww  ÿÀŽ â &r%")   %   â |       ä x#Eww  ÿÀŽ â &r%!D   %   â t((ÿ C    ä t#ÿ #    â &r'"	   '   ä &r%+&   %   â x"(@Q  #    â&r'!$   '   â x#(bs  #    â &r(0   ÿ   â x@Q      â &r%-"   %   Ä &r'+   '   ä &r00@   ÿ   â xbs      â &r%.#   %   ä 0r/ÿ/    A   ä 0r,ÿ,    A   ä &r'-   '   Ä &r01B   0   â Er% %       â &r;   ÿ   â  r/,     A  â &r;;@   ÿ   ä &r'.   '   ä 0r*ÿ*    A   ä &r0:A   0   â #r%      â &r(1
   (   Ä &r9
      ä&rB9B   ;   â Er' '       â &r%8)   0   â  r!*/     A  â &r(:   (   ä&r7      ä&rB7A   B   â #r!'      â &r%6D   %   Ä &r!8	   (   ä&r5	      ä&rB5)   B   ä &r%4&   %   ä &r!6$   !   ä&r3$      Ä &rB3D   B   ä &r%2"   %   ä &r!4   !   ä&r      ä&rB&   B   ä 0rÿ    A   Ä &r%>#   %   ä &r!2   !   ä&r=      ä&rB="   B   â  r	*     A  â Er% %       â &r!>   !   â  r,     A  â &r<      Ä 0r?ÿ?    A   ä &rB<#   B   â #r %       â Er! !       â  r*?     A  âEr        â  r?,?     A  â ErB B       â #r	!      â x€   ÿàÿ â #r        â x€   ÿàÿ â #r?B      â x€   ÿàÿ â $rÿÿ    â rÿ   ÿäÿ â G©  0òÿÿÿÿƒ ê Ay        € ê ˆ  ýÿ H   è ˆ €ýÿ H   è ˆ  þÿ H   è ˆ  €þÿ H   è ˆ   ÿÿ H   è ˆ €ÿÿ H   è {          ì M        € ê „y €   H   " Ey       € æ „y     H   h „y €  H   ¨ „y    H   è „y	 €  H   ( „y    H   h „y    H   h „y    H   h „y    H   h „y €  H   â !r       â !r       Æ/ !r       âO !r       Æ� !r		       â!r       È‰	      " !r       Æ ‰      b !r       Æ „y €  H   â !r       Ê ‰      ¢ !r	       Æ y	      %   " !r       Æ/ ‰      h ‰      â !r       ÊO ‰      ¢ !r       â/ !r       È� ‰ €    h ‰ €    â !r       ÊO ‰
 €    ¢ !r       â/ x	   ÿ   â !r       Æ� ‰ @    " x   ÿàÿ äz  r  p`ð â ‰ @    b z  r  p`ò ä r ÿ   pVp  äx    p"ò â !r

       ÄO „y €  H   è ‰
 @    ¨ ‰      &   è       &   " !r       â !r       Ä/ ‰      '   (       '   h ‰      b !r

       ÆO ‰      ¢ ˆ        â $Š x  Ž ä�‰	
      â $ x  Ž ä$Š ~  Ž â         â $ ~  Ž â/ !r       â%† j   Ž â !r       ÆO % j   Ž â !r	
	       â� †‰     è †    â G	  �     € ê y      &   ( y      '   b $z x  Ž È $z ~  Ž Ê/ z r  ÿàõ Ê $rÿÿ   ÿ â z j  ÿ„ È z k   Ê †y 	    è Ay        € ê !r        â „y    H   â Ey  à    € ä !r       â „y    H   æ !r       â „y	    H   è ‰       ( „y €  H   è „y €  H   â !r         Æ „y €  H   è ‰       $ !r        Ê ‰ €    $ !r       Ê ‰ @    $ !r       Ê ‰       $ !r        â G™  °     € ê y      &   " ¹z  r      ä �x   ?àÿ â y       '   b $z x  Ž È $z   ~  Ž Ê/ |     ÿàõ È rÿÿ   ÿä ä z  j  ÿ„ È z  k   Ê †y     è Ay        € ê !r        â !r       Æ !r       â !r       ä y      &   b !r	       â !r       È ‰      ¨ ‰      " $z x  Ž â/ !r       âO !r       È ‰       ( ‰      h y      '   ¢ !r         â z  r      â !r	       Æ/ ‰  €    " $z ~  Ž ÆO ‰	 €    d r   ÿ„ â !r        â !r	       È/ ‰ @    ( ‰ @    b !r       â !r       È/ ‰      " rÿÿ   ÿä ä z j  ÿ„ â‰       f z k   â !r       â !r        È/ †‰     â M™        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      "   " ¹z  l      â Ey  à    € â ™x?    â y       !   b ¹z  F   
   â r ÿ       â $rÿÿ   ÿ Ž â y      %   ¢ $rÿÿ   ÿ Ž â r ÿ       Ä r ÿ   pRð är ÿ   pRò â$x     Ž ä/ $x   ÿ Ž ÆO r   ÿ� È xÿ     â $€    Ž Æ |    pbô Ú G)  À    € ê y      '   " r ÿ       â $rÿÿ   ÿ Ž â xþÿÿÿÿÀŽ â y	      &   b $tÿ   ÿ Ž ä $x   ÿ Ž ä $x   
Ž ä $z p  Ž Æ xÿ     â 'z y   Ž È 'z	 s   Ž â/ zÿ z    ä x   ÿ   â $z }  ÿŽ â zÿ t    â $z |  Ž ä$z |  Ž â x   ÿàÿ â %x   ÿŽ È $z v  Ž âr ÿ       â $z v  Ž ä $z p  Ž ä %x   Ž È %x   Ž â z X  ÿàû Æ $z	 w  ÿŽ â x   ÿàý â %v
 X  Ž âz X  ÿàù ä zÿ Y  ¤~ â %v X  Ž â r   àµ ä x   ÿàû ä z q  ÿàÿ Ä zÿ Y  „þ ä rÿÿ   d~ ä �y3úÿÿ • ¨  �y4þÿÿ • â r   ÿàõ Æ �y0üÿÿ • " $t#ÿ$   ÿ Ž ä $rÿÿ    â �y     • b %v Z  # Ž Æ �yþÿÿ • b $x	$   ÿŽ Æ �y   • d r	   ÿàÿ ä �y    • h �y   ™ h �y   ™ h �y!   ™ b r   ÿàõ Æ �y   ™ d $rÿÿ    ä %v Z  # Ž â �y"   ™ f $x$   ÿŽ â �y$   • f $x		   Ž â �y'   ™ h �y   ™ h �y%   ™ h �y+   ™ h �y*   ™ h �y)   ™ h �y(   ™ h �y#   ™ h �y&   ™ h �y,    • h �y/    ™ h �y-
    • h �y.    ™ b r        â $rÿÿ    Ž Ä �y1    • h �y2    • b ‚x ÿ      ä $~8ÿ   ÿ Ž â |6       ä |7       Ä x€  ÿàõ ä x@   ÿàÿ Æ $rÿÿ    â x€  ÿàõ Ê/ $rÿÿ    â |    pbô ä x€  ÿàù ä x

€  ÿà÷ ä x€   ÿàÿ â $rÿÿ    â x€   ÿàÿ ä rÿ   ÿäÿ Ä x33™  ÿ    ÈO x5ÿ   3  ä x4™  ÿ    ä� x55ww  ÿÀŽ ä x	ÿ     ä t8ÿ 5    ä x50™  ÿ    Äxww  ÿÀŽ ä x4	ww  ÿÀŽ â $r	ÿÿ   5 Ž â x33ww  ÿÀŽ ä t06ÿ     ä x ™  ÿ    ät36ÿ 3    Ä x ÿ   	  ä t57ÿ 4    ä x3@Q      äx4 ww  ÿÀŽ ä x3bs      ä x 0@Q  5    Ä x3	ww  ÿÀŽ ä x00bs  5    ä x5ÿ     ä t36ÿ 3    ä x	ww  ÿÀŽ â &r   ÿ   â x65ww  ÿÀŽ â &r    ÿ   â t47ÿ 4    Ä t:7ÿ 6    â &r6      ât58ÿ 	    â &r80      â x3@Q  4    äx3bs  4    ä x3™  ÿ    ä x	5@Q  :    â &r!   6   â r6 3       Æ &r!!	   8   â x8™  ÿ    ä x7ÿ   6  â $~4ÿ   ÿ Ž â x5bs  :    â $~5ÿ   ÿ Ž â x9ÿ   8  ä x66ww  ÿÀŽ ä x77ww  ÿÀŽ â $~3ÿ   ÿ Ž â |       Ä x88ww  ÿÀŽ ä x:9ww  ÿÀŽ ä t55ÿ 6    ä t44ÿ 7    â &r7   !   â t33ÿ 8    ä t:ÿ :    Ä x!™  ÿ    â &r6      â x5@Q  4    ä x3@Q  :    äx3bs  :    â $r3ÿÿ   ! Ž â x5bs  4    â &r!   6   È &r4"   !   ä &r!%   ÿ   ä&r %    ÿ   ä &r!+   !   âx$$™  ÿ    â &r +0       ä &r*   !   Ä &r7   7   â x3ww  ÿÀŽ â &r *	       â x3ÿ   3  â &r)      ä $rÿÿ   $ Ž â |0       â $~6ÿ   ÿ Ž â x33ww  ÿÀŽ â &r)       â x	ÿ     Ä t6ÿ     â &r(      â t30ÿ 3    â $~ÿ   ÿ Ž â |       ä xww  ÿÀŽ ä x		ww  ÿÀŽ â &r(      â x@Q  3    Ä tÿ     ä t	ÿ 	    â &r#      â xbs  3    â &r4'   4   âx@Q  	    â &r""   7   ä&r#      â xbs  	    â &r&      Ä &r4/   4   ä 0r-ÿ-    A   ä 0r,ÿ,    A   ä &r"'   "   ä&r&      â Er4 4       â &r.      â  r-,     A  â 0rÿ1    A   Ä &r"/   "   ä&r.      ä 0r	ÿ2    A   â Er        â  r--     A  â #r4      â Er" "       â  r,	     A  â Er        â  r	     A  â #r-      â #r"      Ä #r      â G©   ôÿÿÿÿƒ ê Ay        € ê ˆ  þÿ H   è ˆ €þÿ H   è ˆ  ÿÿ H   è ˆ €ÿÿ H   è {          ì M        € ê „y      H   " z  r  p`ð Æ „y  €   H   b r  ÿ   pVp  Æ „y     H   ¨ „y     H   è „y  €  H   ( „y     H   h „y     H   h „y	  €  H   h „y     H   h „y
  €  H   h „y  €  H   b !r       Æ „y  €  H   " !r       â/ !r       âO !r       ä� !r       â!r       â!r       ä !r		       Æ ‰      b !r       Æ ‰	      ¢ !r

       Æ ‰      â !r

       È !r
       â !r       È/ ‰      " !r	       ÆO ‰      b !r       Æ� ‰      ¨ ‰      è y      &   " !r       â !r       È/ ‰      " !r       ÆO ‰ €    b !r       Æ� ‰ €    ¨ ‰
 €    â !r       â !r       È/ ‰ €    " x   ÿàÿ â $z x  Ž â!r       äO ‰ @    b z  r  p`ò â !r

       ä� ‰ @    ¢ x     p"ò Æ ‰	
 @    è y      '   " !r       â !r       â/ !r       ÈO ‰      " !r	
	       Æ� ‰      b $z ~  Ž âx        ä ‰
 @    ¤ Š  r  ÿàõ ä ‰	      æ $‚ÿÿ   ÿ â Š  j  ÿ„ È Š  k   â !r       â %v j   Ž â !r       È/ †‰     â !r

       ÆO †    â !r		       Æ� ‰
      h  †‰ 	    â M™        € ê ¹z  r      â !r
       â/ �x   ?àÿ Ì |   ÿàñ È rÿÿ   ÿä  ä z j  ÿ€ È z k    Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y
      &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž ä $rÿÿ   
 Ž È % \   Ž Ì �	    ™ b $ÿÿ   ÿ Ž â ¹z  l      â $ÿÿ    Ž â y      "   b ™x?    â $‚ÿÿ   ÿ Ž â Ey  °    € â y      !   b '
  m   Ž Æ y      '   ¤ 
 ÿ n     ä y      %   æ $ ÿÿ    
Ž È $
  o  Ž È 'Š  s   Ž È $x    Ž Ê/ r   ÿ� È xÿ     È |    pbò ä Šÿ t     â $r ÿÿ   ÿ Ž Ô G  °    € ê x þÿÿÿÿÀŽ âÏ $z w  ÿŽ â rÿ   ÿ3Ž â $z }  ÿŽ â Ey      € â $x    
Ž â |    ÿàÿ â $rÿÿ   ÿ Ž ä $x   ÿ Ž â xÿ @   ÿÀ‚ â 'z y   Ž â r   àµ Ä xÿ     â $z p  ÿŽ â xÿ ÀÿÿÿÿÀ€ â $r ÿÿ   ÿ Ž â zÿ z    ä rÿÿ   d~ ä xÿ     â G       € ê $z v  Ž ä$t	ÿ   ÿ Ž Ä $z  |  Ž È $x     Ž È %v  X  	Ž Ê r   ÿ‚ â$tÿ$   ÿ Ž â �y    • ¤  r	   �  Ê �y    • â r   ÿ‚ Æ �y"   • $ $r ÿÿ   Ž  ä %v Z   Ž â �y   • & $x $   ÿŽ â �y   • ¦ $x   Ž Ê �y!   ™ ¨  �y   ™ ¨  �y   ™ ¨  �y   ™ ¨  �y   ™ ¨  �y   ™ ¨  �y    ™ ¨  �y    ™ ¨  �y    • ¢  ‚x ÿ      Ä $~ÿ   ÿ Ž ä/ $~#ÿ   ÿ Ž â x@   ÿàÿ ä x  ™  ÿ    È� xÿ      ä x  ww  ÿÀŽ ä xww  ÿÀŽ ä x	"™  ÿ    ä t ÿ      ä t#ÿ     Ä xÿ   	  ä x @Q      äx bs      â $~ ÿ   ÿ Ž â x		ww  ÿÀŽ ä xww  ÿÀŽ ä x™  ÿ    â&r!   ÿ   âO t	 ÿ 	    Ä tÿ     â $rÿÿ    Ž ä &r      â x	@Q      äx		bs      ä xÿ     â $~ÿ   ÿ Ž â x™  ÿ    â $~ÿ   ÿ Ž â xww  ÿÀŽ Ä xww  ÿÀŽ â &r      â xÿ     ä tÿ     ä tÿ     ä xww  ÿÀŽ â &r		      â xww  ÿÀŽ Ä x@Q      äxbs      ä tÿ     â &r   	   â tÿ     Æ &r      â x@Q      äxbs      Æ &r       ä 0r ÿ    A   ä &r      ä 0rÿ    A   Æ Er        ä  r       A  È #r     ÿ   ä Ay       € ê G‰  P
    € ê $z |  Ž â x     â $tÿ   ÿ Ž ä $z v  Ž È$x   Ž â x€   ÿŽ Æ %v X  Ž Ê r   ÿ€ Ê $x     â x€ÿÿÿÿàÿ È �y"   • ¢  r   ÿàñ â $t$ÿ$   ÿ Ž ä �y   • ä  $rÿÿ     ä %v Z  $ Ž â �y!   • & $x	$   ÿŽ â �y   • f $x   	Ž â �y‚  • h �y   ™ h �y   ™ h �y    ™ h �y   ™ h �y   ™ h �y   ™ b r   ÿàñ Æ �y„  • d $r#ÿÿ     ä %v Z  $ Ž â �y&   ™ f $x##$   ÿŽ â �y$†  • f $x		   #Ž â �y)ˆ  • h �y#   ™ h �y%   ™ h �y,   ™ h �y+   ™ h �y'   ™ b $rÿÿ    Ž Æ �y/    ™ b $rÿÿ    Ž Æ �y.   ™ h �y-   ™ h �y(    • h �y*    • h �y2    ™ h �y1€  • h �y0    • b ‚x ÿ      Ä $~5ÿ   ÿ Ž ä $~ÿ   ÿ Ž ä/ $~	ÿ   ÿ Ž â x€   ÿàÿ È |    pbð ä x 	  ÿàó ä x   ÿàÿ Æ $rÿÿ   Ž  â x""™  ÿ    ÈO x3ÿ   "  ä x""ww  ÿÀŽ ä x33ww  ÿÀŽ ä x™  ÿ    ä� t"5ÿ "    ä t3ÿ 3    Ä x!!™  ÿ    äxÿ     â $~ÿ   ÿ Ž â x"@Q  3    ä xww  ÿÀŽ â $rÿÿ   ! Ž â xww  ÿÀŽ â &r   ÿ   âtÿ     Ä x"bs  3    ä t		ÿ     ä xÿ     â $~!ÿ   ÿ Ž â xww  ÿÀŽ â $~ÿ   ÿ Ž â xww  ÿÀŽ â &r      â x@Q  	    Ä xbs  	    ä t!ÿ     â &r       â tÿ     ä x™  ÿ    â &r      â x@Q      äxbs      â $rÿÿ    Ž â x™  ÿ    â &r      Ä $~	ÿ   ÿ Ž â xww  ÿÀŽ â &r      ä $rÿÿ    Ž â t	ÿ     ä xÿ     ä x	ÿ     â $~ÿ   ÿ Ž â xww  ÿÀŽ â $~ÿ   ÿ Ž â xww  ÿÀŽ â $~ÿ   ÿ Ž â x		ww  ÿÀŽ Ä tÿ     ä tÿ     ä t	ÿ 	    ä x™  ÿ    ä x@Q      äxbs      Ä x@Q  	    äxÿ     â $~ÿ   ÿ Ž â xbs  	    â $~	ÿ   ÿ Ž â xww  ÿÀŽ â &r#   ÿ   â xww  ÿÀŽ ä x$$™  ÿ    â &r%%      â t	ÿ     Ä tÿ     â $rÿÿ   $ Ž ä &r&      â x@Q      äxbs      ä xÿ     â &r,   %   â xww  ÿÀŽ Ä xww  ÿÀŽ ä x)™  ÿ    â &r+      â t	ÿ     ä tÿ     ä xÿ     â $~ÿ   ÿ Ž â x	@Q      Ä xww  ÿÀŽ ä xww  ÿÀŽ â &r'	      â xbs      ä tÿ     ä t	ÿ     â &r/      ä &r.      â x@Q  	    Ä x	bs  	    â 0r(ÿ(    A   ä &r-      ä 0rÿ*    A   â Er        â &r2	      ä 0r1ÿ1    A   â  r(     A  â 0r0ÿ0    A   â Er        Ä #r        ä  r110     A  È #r 1       â G‰   öÿÿÿÿƒ ê Ay        € êÏ r ÿ   pRð är ÿ   pRò Ö ˆ  €ÿÿ H   è {          ì M        € ê „y     H   " r ÿ   pRð â $z
 x  Ž Ä „y €   H   h „y    H   ¢ !r        È !r       È/ !r       ÊO ‰      $ !r       Ê ‰       $ !r         Ê ‰  €    $ !r        Ê ‰ @    $!r       Ê ‰	      b  M	        € ê $tÿ   ÿ Ž â !r		       â/ $z ~  Ž È %v j   Ž Ê †y 	    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž È %
 \   Ž Ê �	
    ™ b $rÿÿ   ÿ Ž â Ey  ð    € â $rÿÿ   ÿ Ž â y       "   b 'Š s   Ž Æ y	      !   b '
 m   Ž È $r
ÿÿ   ÿ Ž â 
ÿ n    ä rÿ	    ÿü‚ â/ $x     	Ž â Šÿ t    È r   ÿ� Î G       € ê z ÿ `  pPô ä z ÿ ^  pPò ä z ÿ b  p ö ä z ÿ a   Sô ä z ÿ _  Sò ä z ÿ c  0't Ö       '   " 
       ä¢       â ©      '   d $
 x  ÿŽ ä       %   ¢ $ª x  ÿŽ ä $r
ÿÿ   ÿ Ž â ©      %   â $ ~  ÿŽ Ä $ª ~  ÿŽ Æ/ 	   àë äOÿ     ä ¢	   àÇ ä�¨ÿ     ä ÿÿ   Äþ ä    ÿàý Ä ¢   ÿàû ä ¢ÿÿ   „þ â $ÿÿ    â  ^  ÿ† äª b  ÿˆ â $¢ÿÿ   Ž â  _  � â $rÿÿ   ÿ Ž Æ ª c   â �
     h �©     d Ay        € ê ¹z  l      â xÿ     â ™x?    â $ÿÿ   
Ž â Ey  P    € â $rÿÿ   ÿ Ž Ä $
 o  Ž â|    pbò â $‚ÿÿ    Ž ä $rÿÿ   ÿ Ž Ô G  à
    € ê y      '   " $rÿÿ   ÿ Ž ä $z w  ÿŽ â y      %   b $rÿÿ   ÿ Ž ä $x   ÿ Ž ä 'z y   Ž È $z }  ÿŽ â zÿ z    Ê $z |  Ž È $z v  Ž âxþÿÿÿÿÀŽ Æ $z p  Ž ä/ $x   
Ž ä %x   ÿŽ Æ xÿ     â $x   ÿ Ž â z `  ÿà÷ äz X  ÿàõ â %x   Ž â z a  ÿäÿ ä r   àÇ â $rÿÿ   ÿ Ž â z `  ÿàñ Ä z X  ÿàó ä x   ÿàû ä x   ÿàý ä z Y  ÿä ä zÿ a  þ äzÿ Y  $~ Ä rÿÿ   „þ Æ �y#úÿÿ • ¢ r   ÿàñ â $tÿ$   ÿ Ž ä �y"üÿÿ • ä $rÿÿ     ä %v Z   Ž â �y!þÿÿ • & $x$   ÿŽ â �y     • & $x   Ž Ê �y   ™ ( �y   ™ ( �y   ™ ( �y   ™ ( �y   ™ " $rÿÿ    Ž Ä $rÿÿ    Ž â �y   ™ ( �y   ™ ( �y    • ( �y    • " ‚x ÿ      Ä $~&ÿ   ÿ Ž â �y    ™ b $~%ÿ   ÿ Ž â z ÿ `  pPð È z ÿ a   Sð ä x##™  ÿ    ÈO x$ÿ   #  ä x"™  ÿ    ä� x##ww  ÿÀŽ ä x$$ww  ÿÀŽ ä t#&ÿ #    ä t$%ÿ $    Ä xÿ     â $~"ÿ   ÿ Ž â xww  ÿÀŽ ä xww  ÿÀŽ ä x#@Q  $    äx$#bs  $    ä t"ÿ     â &r   ÿ   ât%ÿ     Ä x"!™  ÿ    â &r$      â x!@Q      äxbs      ä xÿ   "  â &r!      â x  ™  ÿ    â $~!ÿ   ÿ Ž â x""ww  ÿÀŽ â $~$ÿ   ÿ Ž â xww  ÿÀŽ â &r      â xÿ      Ä t"!ÿ "    ä t$ÿ     â $~#ÿ   ÿ Ž â x!ww  ÿÀŽ ä x  ww  ÿÀŽ ä x"@Q      äx""bs      Ä t #ÿ      â &r      â t$ÿ !    Æ &r!"      â x @Q      â$rÿÿ    Ž â x  bs      â 0rÿ    A   ä &r!   !   ä $rÿÿ    Ž ä 0rÿ    A   â G‰  @    € ê �y$þÿÿ  ¨ �y#     è �y"    ( �y    â $rÿÿ    Ž Ä $rÿÿ    Ž Ê �y     " $~'ÿ   ÿ Ž ä $~&ÿ   ÿ Ž ä $~(ÿ   ÿ Ž â x%ÿ   $  äO x$$ww  ÿÀŽ ä x%%ww  ÿÀŽ Ä t$'ÿ $    ä t%&ÿ %    È x'$@Q  %    äx$$bs  %    ä x%ÿ   #  â� &r''   ÿ   â x##ww  ÿÀŽ ä x%%ww  ÿÀŽ â &r'$   '   â t#&ÿ #    â $~$ÿ   ÿ Ž â t(ÿ %    â $~%ÿ   ÿ Ž â xÿ   "  â0rÿ    A   â x#@Q      Ä x""ww  ÿÀŽ ä xww  ÿÀŽ â &r   '   â xÿ     â  r     A  â x#bs      ä t"%ÿ "    ä t$ÿ     â &r      â x#ww  ÿÀŽ Ä xww  ÿÀŽ ä x"@Q      äx""bs      ä t%ÿ     â &r      â t#$ÿ #    Æ &r"      â x@Q  #    äxbs  #    Æ &r      È &r      ÊEr        Ê #r      â x@   ÿàÿ â &r     !   â r     A  â x€  ÿàù ä |    pbð ä x€  ÿàó â $rÿÿ    â x€  ÿàõ ä Er          â $rÿÿ   Ž  â x€  ÿà÷ â $rÿÿ    â x€   ÿàÿ â #r       Ä $rÿÿ   Ž â G‰  ÷ÿÿÿÿƒ ê Ay        € ê r  ÿ   pRð âEy  €     € â r  ÿ   pRò Ö G‰  P     € ê z ÿ `  pPð â $x      	Ž Æ z ÿ a   Sð ä ˆs  €ÿÿ H   ö ˆ     H   ä Ay        € ê {          ì M        € ê z ÿ `  pPð â „y	     H   f z ÿ a   Sð Ú „‰ 	 €   H   ( „‰	    H   ¢!r       È/ !‚         È !‚        âO G‰        € ê „y 	 €  H   ( „y	 €   H   h „y	    H   ¨ „y	    H   è „y	 €  H   " !r         â !r       Æ/ !r         âO !r       Æ� !r        ä‰      " z ÿ `  pPð ä r 	ÿ   pRò ä z ÿ a   Sð â !r       Ê ‰       $ !r         Ê ‰  €    $ !r        Ê ‰ @    $ !r       Ê ‰      $ !r       â G‰        € ê ‰       $ !r         Ê ‰       $ !r        Ê ‰ €    $ !r       Ê ‰ @    $ !r       Ê ‰      $ !r       ä M        € ê !r
       â G‰  `    € ê $v ÿ h  ÿ Ž â !r       È x     pBð Ú G	  p    € ê x  ÿÿÿÿÿàÿ È x       € Ê $x     ÿ Ž È ‚{   €     $ xÿ     È Iy �ìÿÿÿÿƒ ê  x ;ª¸¿  A  Ì s          $ !t   €?    Ì s        $  r     A  È  r     A  â Gy  P    € ê  x '7=  A  â x*BL?  A  Æ #t   €?    È  r      A  â  x    ?  A  Ê s     $   $ !t  €?    È  r       A  È  r      A  â Gy  °    € ê x     pRð Ú G‰  à     € ê x     pRð Ú G	        € ê 	z  i    � ä 	z i    � Æ  x ;ª¸¿  A  â 	z i €  � Ê s        $ !t  €?    Ì s        $  r       A  È  r      A  â Gy  Ð     € ê  r     A  â Gy  °     € ê 	x   à@  � ä 	x  à@  � Æ  x #ÛÙ?  A  â 	x  àÀ  � Æ  x;ª¸¿  A  ä !t  €?    È s        $ !t  €?    Ì s        $  r       A  È  r      A  ä y       &   ( y      %   ( y      '   b $z   x  Ž ä $tÿ   ÿ Ž Ä $z  ~   Ž È/ %v  j   Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y       &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž ä $rÿÿ     Ž È % \   Ž Ê �	    ™ b ¹z  l      â Ey  �    € â ™x?    â y      "   b $rÿÿ   ÿ Ž Æ y      !   ¢ $ÿÿ   ÿ Ž ä $ÿÿ    Ž â y	      %   & '
 m   Ž È $‚ÿÿ   ÿ Ž â 
ÿ n    Ê $ÿÿ   
Ž â r ÿ   pRò Æ/ $
 o  Ž â r ÿ   pRô â$x    Ž äO $‚ÿÿ    Ž ä $rÿÿ   ÿ Ž â r   ÿ� â 'Š s   Ž Æ xÿ     â $€   Ž ä $x		   ÿ Ž â |    pbö ä Šÿ t    ä x      ÿ  Ò G9  à    € ê y      '   " $r
ÿÿ   ÿ Ž â xþÿÿÿÿÀŽ Ê $x   
Ž È $x   ÿ Ž â xÿ     â 'z y  
 Ž È $z
 w  ÿŽ â zÿ z    â $z	 p  Ž ä $z }  ÿŽ ä $z |  Ž ä $rÿÿ   ÿ Ž â r

   à± â $z v  Ž Ä$vÿ X  ÿ Ž â rÿÿ   d~  â %x   ÿŽ È $vÿ Y  ÿ Ž ä $x   ÿ Ž ä %x   Ž Ê r   ÿàñ Ê $x     Ê �y   • ¢  r
   ÿàñ Æ �y   • â  $tÿ$   ÿ Ž ä $rÿÿ     â �y"   • " %v Z   Ž Æ �y#   • b $x$   ÿŽ È $x   Ž Ê �y   ™ ( �y   ™ b $v!ÿ p  ÿ Ž Æ �y   ™ b %x!   Ž Æ �y&   ™ h �y'   ™ h �y(   • h �y)   ™ h �y*   • h �y+   • b %x!   Ž Æ �y,   ™ h �y-   • h �y.    ™ h �y/   • h �y0   • h �y1   • b %x$!   Ž Æ �y2   • h �y3$   • h �y4$   • h �y5$   • b r   ÿàñ Æ �y    • d! $x     ä �y$$   • d %x!   Ž ä�y    • h %x!   Ž Ä �y    • h %x !   Ž ä �y    • h	 �y      • b ‚x ÿ      Ä $~%ÿ   ÿ Ž â/ x™  ÿ    ÊO $rÿÿ    Ž â x™  ÿ    È� xÿ     â $~ÿ   ÿ Ž Æ xww  ÿÀŽ â $rÿÿ    Ž â xww  ÿÀŽ ä t%ÿ     ä x!ÿ     ä tÿ     ä x"™  ÿ    â$~ÿ   ÿ Ž â x"ww  ÿÀŽ â $~ÿ   ÿ Ž â x!ww  ÿÀŽ Ä x!@Q      äxbs      ä t"ÿ "    â &r!!   ÿ   â tÿ     ä xÿ     â $~ÿ   ÿ Ž â x%"@Q      â &r   !   âxww  ÿÀŽ Ä xww  ÿÀŽ â &r%      â x"bs      ä tÿ     ä t"ÿ     ä x##™  ÿ    â &r&      â x@Q  "    Æ $rÿÿ   # Ž â x"bs  "    â &r'      â x(™  ÿ    ä xÿ     â $~ÿ   ÿ Ž â x(ww  ÿÀŽ â &r)"      â xww  ÿÀŽ â $rÿÿ    Ž Ä $~ÿ   ÿ Ž â t"ÿ     ä xÿ     ä tÿ (    ä x!*™  ÿ    ä x(ww  ÿÀŽ ä x*ww  ÿÀŽ Ä x@Q  "    äxbs  "    â $rÿÿ   ! Ž ä $~ÿ   ÿ Ž ä $~ÿ   ÿ Ž â x"ÿ     ä tÿ (    Ä t*ÿ *    â $~(ÿ   ÿ Ž â x!ww  ÿÀŽ ä x"ww  ÿÀŽ ä x"@Q  *    äx#+™  ÿ    ä xbs  *    â &r%"   ÿ   â tÿ !    Ä t((ÿ     ä x*ÿ   #  â $~!ÿ   ÿ Ž â x"@Q  (    â$~ÿ   ÿ Ž â x##ww  ÿÀŽ â &r   %   â x**ww  ÿÀŽ ä xbs  (    â &r"      â t!!ÿ #    Ä t*ÿ *    ä x--™  ÿ    â &r,      â x!@Q  *    â &r"&      ä $rÿÿ   - Ž ä &r.      ä &r'   "   â xÿ     â $~"ÿ   ÿ Ž â x*!bs  *    â $~(ÿ   ÿ Ž â xww  ÿÀŽ Ä xww  ÿÀŽ ä x!/™  ÿ    ä t"ÿ     ä t((ÿ     ä x!ww  ÿÀŽ ä x!ÿ   !  â &r)*      â x00™  ÿ    â $~*ÿ   ÿ Ž â x"@Q  (    Ä xbs  (    â $~(ÿ   ÿ Ž â x!!ww  ÿÀŽ â $rÿÿ   0 Ž â x#1™  ÿ    â &r""   ÿ   â t*ÿ     ä t((ÿ !    ä x%ÿ   #  Ä xww  ÿÀŽ â &r"   "   â x!@Q  (    â$~0ÿ   ÿ Ž â xÿ     ä x##ww  ÿÀŽ ä x%%ww  ÿÀŽ ä xbs  (    â &r!!   "   â x(ww  ÿÀŽ â $~ÿ   ÿ Ž â t*ÿ     Ä t*ÿ #    ä t*0ÿ %    â &r"&   !   â t(ÿ (    ä x!@Q  *    äx2™  ÿ    ä x*bs  *    â &r'!   "   â x@Q  (    â $~"ÿ   ÿ Ž â x!ww  ÿÀŽ Ä x33™  ÿ    â &r,      â t"ÿ !    Æ $r!ÿÿ   3 Ž â xÿ     ä xbs  (    ä x"ÿ   !  â $~(ÿ   ÿ Ž â xww  ÿÀŽ â &r)*      â x%"ww  ÿÀŽ â $~*ÿ   ÿ Ž â x44™  ÿ    Ä x!!ww  ÿÀŽ ä t"(ÿ     ä t(0ÿ %    â $r%ÿÿ   4 Ž â t#*ÿ !    ä x!@Q  "    äxbs  "    Ä x"#@Q  (    äxÿ   %  ä x##bs  (    â $~(ÿ   ÿ Ž â x%%ww  ÿÀŽ ä xww  ÿÀŽ â &r""   ÿ   â t%(ÿ %    Ä x5™  ÿ    ä t(*ÿ     â &r"#   "   â xÿ     ä x%@Q  (    äx%%bs  (    â $~(ÿ   ÿ Ž â xww  ÿÀŽ â &r,!      â xww  ÿÀŽ â &r   "   â x!$™  ÿ    Ä t(ÿ     ä t"*ÿ     ä xÿ   !  â &r&%      â x@Q  "    â $~$ÿ   ÿ Ž â x!!ww  ÿÀŽ â $~&ÿ   ÿ Ž â xww  ÿÀŽ â &r'      â x"bs  "    Ä t!$ÿ !    ä t&ÿ     â &r)"      â x@   ÿàÿ ä x!@Q      äx!!bs      ä |    pbð â &r,,      Ä &r.      ä&r.      ä 0rÿ    A   ä 0rÿ    A   ä 0rÿ    A   ä 0rÿ    A   Ä &r!.!   ,   ä 0r ÿ     A   â  r     A  â  r     A  â r     A  âx€  ÿà÷ â  r      A  â Er        ä Er        Ä Er        ä Er! !       â #r      â #r      â #r      â x€   ÿàÿ â #r!      â $rÿÿ   Ž â D	       À â Gy  pïÿÿÿÿƒ ê Ay        € ê ˆ  þÿ H   è ˆ €þÿ H   è ˆ  ÿÿ H   è ˆ €ÿÿ H   è {          ì M)        € ê „y     H   " z 	 r  p`ô â $z   x  	Ž Ä „y €   H   b r ÿ   pVt Æ „y    H   ¨ „y    H   è„y €  H   ( „y    H   ( „y
    H   ( „y €  H   ( „y €  H   ( „y    H   ( „y €  H   " !r       Æ „y €  H   " !r       â/ !r       äO y      '   b !r       â� !r       â!r       ä !r
       â !r       È ‰      ¢ !r       â $z  ~   Ž â/ !r       Æ ‰      b !r       Æ ‰      â !r       Ê ‰      " !r       âO !r       â/ !r       È� ‰      h ‰      ¢ !r       Æ ‰      ( ‰      â !r       â/ !r       ÈO ‰
 €    b !r       Æ ‰ €    " !r       Æ� ‰ €    ¨ ‰ €    â !r

       â/ !r       È ‰
 @    " x	   ÿàÿ â !r       äO ‰ @    b x	   ÿàÿ â !r       ä� ‰ @    ¢ z  r  p`ö ä z  r  p`ò â ‰ @    â x    pVö Ä x    pVò  â !r
       â x
	   ÿàÿ â $t	ÿ   ÿ Ž â !r       ä/ ‰      " z 
 r  p`ð â !r       äO ‰      b x    pVp  â !r       Ä� ‰      ¨ ‰      â !r       â !r       â/ %v  j  	 Ž â !r       ÈO †©     â !r       Æ� †¹    è †‰    â M        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú         Ê % \   Ž Ê �	    ™ ¢  r ÿ       â Ey       € ä y      "   d '
 m   Ž äy      %   ä 'Š s   Ž â 
ÿ n    â y       !   " r ÿ       Ä ÿ   ÿáÿ â y      '   " r ÿ       Æ $
 o  Ž â r ÿ   pRò ä/ x   ÿ   ä� x     pDò  är	    ÿàÿ är    ÿ(Ž ä z 	 r  pdò  Ä r   ÿ� ä Šÿ t    È r       ÊO G  À    € ê z ÿ `  pPô ä z ÿ ^  pPò ä z ÿ b  p ö ä z ÿ a   Sô ä z ÿ _  Sò ä z ÿ c  0't Ö $ ~  ÿŽ âÿ     â$ª ~  ÿŽ ä $	 x  ÿŽ â    àë â$ª x  ÿŽ â ¢    àÇ ä ¨ÿ     ä ÿ   ÿÄþ Ä 	   ÿàý ä ¢
   ÿàû ä ¢	ÿ   ÿ„þ ä ÿ   ÿä ä  ^  ÿ† ä ¢ÿ	   ÿäÿ Ä ª
 b  ÿˆ ä 	 _  � ä r ÿ       ä r ÿ       ä ª
 c   ä �     h �©     d Ay        € ê ¹z  l      â xÿ     â ™x?    â ‚        â Ey   !    € â r ÿ       â x      ÿ  â x      ÿ  â |    pbð â x
      ÿ  â r2 ÿ       Ö G	        € ê r ÿ       â $z w  ÿŽ â xþÿÿÿÿÀŽ â x
      ÿ  â x   ÿàÿ â 'z y   Ž âr        ä r  €ÿàÿ â $z }  ÿŽ â zÿ z    â $z p  Ž â x$   ÿ   Ä x   ÿàÿ â $z |  Ž â x   ÿàÿ â %x$$   ÿŽ â r   à£ äxÿ     â $z |  Ž â z ÿ `  pPð â $z v  Ž âz  `      â $z v  Ž â z  a      â %x   $Ž â z  X      â x      ÿ  â z  Y      â $z! p  Ž âx   ÿà÷ â $z# p  Ž âr ÿ       â $z p  Ž â x      ÿ  â %x !   $Ž â r2 ÿ       Ä z ÿ a   Sð â %x"#   $Ž â x   ÿ   ä rÿÿ   Dþ  â %x$   $Ž â rÿ   ÿäÿ Æ r(   ÿàó È r)   ÿäÿ  ä r   ÿàó Æ �y5(   • ¢  x $       ä rÿ   ÿäÿ  â �y7(   • ä  %v Z   Ž ä �y4(   • $ $x$   ÿŽ Ä �y6(   • & r   ÿàÿ Ê �y   ™ ( �y   ™ " r&   ÿàó Æ �y   ™ " r'   ÿäÿ  ä z9  p      â �y.   ™ ( %x&9   &Ž â�y/   ™ ( �y:&úÿÿ • ( �y8&üÿÿ • ( �y;&þÿÿ • ( �y0   ™ ( �y1   ™ " %x*9   &Ž Æ �y3    ™ " %x(   Ž Æ �y?*úÿÿ • ( �y=(    • ( �y>    • ( �y<&    • ( �yB*üÿÿ • ( �yA*þÿÿ • " %x,9   *Ž Æ �yC*    • ( �y@,úÿÿ • ( �yD,üÿÿ • ( �yE,þÿÿ • " %x(9   (Ž Æ �y,,    • &! %x9   (Ž ä�y((    • ( %x&9   Ž ä �y    • ( �y&&    • " ‚x ÿ      Ä |-       ä |       ä/ |'       ä x*5™  ÿ    ÈO x+ÿ   *  ä x77™  ÿ    ä� |5       ä x)*ww  ÿÀŽ ä x++ww  ÿÀŽ ä r* 7       Ä t)5ÿ )    ä t+-ÿ +    ä x5ÿ   *  ä x-*ww  ÿÀŽ ä x*)@Q  +    ä x55ww  ÿÀŽ Ä x))bs  +    â &r**   ÿ   ât''ÿ -    ä tÿ 5    ä x-4™  ÿ    â &r4)   *   â x+'@Q      ä r* -       Ä x)'bs      â &r++   4   â x-ÿ   *  ä x56™  ÿ    ä x4*ww  ÿÀŽ â &r*.)   +   â |       ä x6-ww  ÿÀŽ Ä r+ 5       ä |'       ä t6ÿ 6    ä xÿ   +  ä t''ÿ 4    ä |-       Ä x4+ww  ÿÀŽ ä |)       ä x+ww  ÿÀŽ ä x::™  ÿ    ä t--ÿ 4    ä t4)ÿ +    Ä x+ÿ   :  ä x58™  ÿ    ä |8       ä x++ww  ÿÀŽ ä x'@Q  6    ä x::ww  ÿÀŽ Ä x9ÿ   5  ä x''bs  6    ä x6-@Q  4    äx4-bs  4    ä t88ÿ +    ä x75ww  ÿÀŽ Ä t-)ÿ :    ä |+       ä x59ww  ÿÀŽ â &r/   *   â x:-@Q  8    ät++ÿ 7    ä t*)ÿ 5    â &r0'      â x);™  ÿ    â &r::   ÿ   â x--bs  8    Ä x'+@Q  *    äx++bs  *    ä x*ÿ   )  â &r16      â |6       â &r:-   :   â x))ww  ÿÀŽ ä |8       Ä x-*ww  ÿÀŽ ä t)6ÿ )    â &r''   :   â t68ÿ -    â &r434      ä &r*.+   '   â x)@Q  6    äx))bs  6    Ä x??™  ÿ    â &r*/   *   â x'<™  ÿ    â 0r=ÿ=    A   ä &r*0)   *   â r) ?       â 0r>ÿ>    A   â Er4 4       ä x+ÿ   )  Ä  r==>     A  â x'ww  ÿÀŽ ä x'ÿ   '  ä |6       ä x))ww  ÿÀŽ â #r=4      â x++ww  ÿÀŽ ä |4       Ä x''ww  ÿÀŽ ä t6ÿ     ä x-B™  ÿ    ä t)6ÿ )    ä t68ÿ +    ä t44ÿ '    Ä x'ÿ   -  ä x8)@Q  6    ä|:       ä |<       ä x--ww  ÿÀŽ ä x''ww  ÿÀŽ â &r88   ÿ   â x))bs  6    Ä x+A™  ÿ    ä t-:ÿ -    ä t6<ÿ '    â &r8)   8   â x'@Q  4    äx4bs  4    ä x)ÿ   +  Ä x-@Q  6    äx--bs  6    ä |6       ä x++ww  ÿÀŽ ä x))ww  ÿÀŽ ä t+6ÿ +    â &r   8   â t6:ÿ )    Ä x5@™  ÿ    â &r'1'   *   â x)+@Q  6    â &r*.-      â x-5ww  ÿÀŽ â &r434   '   â x5ÿ   5  â &r'/)   *   â x+bs  6    ä x*C™  ÿ    Ä |8       ä |6       ä x55ww  ÿÀŽ ä x+D™  ÿ    ä x)*ww  ÿÀŽ ä t-8ÿ -    Ä t66ÿ 5    ä x5+ww  ÿÀŽ ä t)8ÿ )    ä x8-@Q  6    äx+ÿ   +  ä x--bs  6    â &r88   ÿ   â |6       Ä x++ww  ÿÀŽ ä xEE™  ÿ    â &r8-   8   â t5:ÿ 5    ä t66ÿ +    ä x*ÿ   *  ä r- E       Ä x+5@Q  6    ä|7       ä x**ww  ÿÀŽ ä x55bs  6    ä x6ÿ   -  â &r++   8   â t*7ÿ *    Ä |9       ä x--ww  ÿÀŽ ä x7,™  ÿ    ä x66ww  ÿÀŽ â &r,.5   +   â t-:ÿ -    ä t69ÿ 6    Ä x5ÿ   7  ä x+-@Q  6    ä x77ww  ÿÀŽ ä x85ww  ÿÀŽ â &r5/+   ,   â x--bs  6    ä t7:ÿ 7    Ä t89ÿ 8    â &r,0   '   âx+)@Q  *    â &r60-   5   â x'7@Q  8    â 0rÿ(    A   â x*)bs  *    â &r+1+   ,   â x87bs  8    â &r'1'   6   ä 0r)ÿ    A   â  r>     A  â Er4 4       â 0rÿ&    A   Ä &r+3*   +   ä&r'38   '   â #r4      â  r)>)     A  â r*>     A  â Er+ +       ä Er( '       â G‰  `    € ê r   ÿàó È r   ÿäÿ  Ê �y4    ¨ �y6    è �y,    ( �y-    " %x&   Ž Ì �y&&     " |7       ä |8       ä |'       ä x5ÿ   4  äO x44ww  ÿÀŽ ä x55ww  ÿÀŽ Ä t47ÿ 4    ä t58ÿ 5    ä xÿ   6  ä� x4@Q  5    äx54bs  5    ä |4       â &r   ÿ   â x66ww  ÿÀŽ Ä xww  ÿÀŽ â &r5      â t6'ÿ 6    ä t4ÿ     ä x4ÿ   ,  äx'6@Q      ä |5       Ä x,,ww  ÿÀŽ â &r'      â x44ww  ÿÀŽ ä t,5ÿ ,    ä x6bs      ä t'7ÿ 4    ä x5ÿ   -  â &r.      â x4-ww  ÿÀŽ Ä |6       ä x55ww  ÿÀŽ ä x-,@Q  '    äx',bs  '    ä t47ÿ 4    â &r/-      â t56ÿ 5    Æ &r0'      â x4@Q  5    äx44bs  5    Æ &r1      ä 0rÿ&    A   ä &r34      Ê Er&        â  r>     A  È #r&      ä G‰  `    € ê r    ÿàó È r!   ÿäÿ  Ê �y5    ¨ �y4    è �y,    ( �y-    " x&   ÿàÿ Ê $z'& p  Ž È %x&'   Ž Ì �y&&     " |8       ä |7       ä |'       ä x6ÿ   5  äO x55ww  ÿÀŽ ä x66ww  ÿÀŽ Ä t58ÿ 5    ä t67ÿ 6    ä xÿ   4  ä� x5@Q  6    äx55bs  6    ä |6       â &r   ÿ   â x44ww  ÿÀŽ Ä xww  ÿÀŽ â &r5      â t4'ÿ 4    ä t6ÿ     ä x5ÿ   ,  äx'4@Q      ä x,,ww  ÿÀŽ Ä x55ww  ÿÀŽ â &r'      â x'ÿ   -  ä x4bs      ä t,7ÿ ,    ä t56ÿ 5    â &r.      â x4'ww  ÿÀŽ Ä x--ww  ÿÀŽ ä x',@Q  5    äx5,bs  5    ä t-6ÿ -    â &r/'      â t47ÿ 4    Æ &r05      â x-@Q  4    äx4-bs  4    Æ &r1      ä 0rÿ&    A   ä &r34      Æ  r>     A  ä Er        Ê #r
   
   ä G‰  `    € ê r"   ÿàó È r#   ÿäÿ  Ê �y5    ¨ �y4    è �y,    ( �y-    " x&   ÿàÿ Ê $z'& p  Ž È %x&'   Ž Ì �y&&     " |8       ä |7       ä |'       ä x6ÿ   5  äO x55ww  ÿÀŽ ä x66ww  ÿÀŽ Ä t58ÿ 5    ä t67ÿ 6    ä xÿ   4  ä� x5@Q  6    äx55bs  6    ä |6       â &r   ÿ   â x44ww  ÿÀŽ Ä xww  ÿÀŽ â &r5      â t4'ÿ 4    ä t6ÿ     ä x5ÿ   ,  äx'4@Q      ä x,,ww  ÿÀŽ Ä x55ww  ÿÀŽ â &r'      â x'ÿ   -  ä x4bs      ä t,7ÿ ,    ä t56ÿ 5    â &r.      â x4'ww  ÿÀŽ Ä x--ww  ÿÀŽ ä x',@Q  5    äx5,bs  5    ä t-6ÿ -    â &r/'      â t47ÿ 4    Æ &r05      â x-@Q  4    äx4-bs  4    Æ &r1      ä 0rÿ&    A   ä &r34      Æ  r>     A  ä Er        Ê #r2   2   â #r)+      â #r*(      â G‰  P    € ê r$   ÿàó È r%   ÿäÿ  Ê �y*    ¨ �y,    è �y)    ( �y(    " %x&   Ž Ì �y&&     " |-       ä |4       ä |'       ä x+ÿ   *  äO x**ww  ÿÀŽ ä x++ww  ÿÀŽ Ä t*-ÿ *    ä t+4ÿ +    ä xÿ   ,  ä� x*@Q  +    äx+*bs  +    ä |*       â &r   ÿ   â x,,ww  ÿÀŽ Ä xww  ÿÀŽ â &r+      â t,'ÿ ,    ä t*ÿ     ä xÿ   )  äx,@Q      ä x))ww  ÿÀŽ Ä xww  ÿÀŽ â &r      â xÿ   (  ä x,bs      ä t)*ÿ )    ä t'ÿ     â &r..      â xww  ÿÀŽ Ä |+       ä x((ww  ÿÀŽ ä x)@Q      äx))bs      ä t('ÿ (    â &r/   .   â t+ÿ     Æ &r00)      â x(@Q      äx((bs      â 0rÿ&    A   ä &r1   0   Æ  r>>     A  â &r3(      Ê Er        Ê #r	>   	   â x@   ÿàÿ ä x€  ÿàõ ä |    pbò ä x€  ÿà÷ ä x€   ÿàÿ ä rÿ   ÿä Ä rÿ   ÿäÿ Ê D       À â Gy   âÿÿÿÿƒ ê Ay        € ê y      "   " Ey  0    € â r ÿ   pRð är ÿ   pRò Ö G‰  ð     € ê z ÿ `  pPð ä r    ÿ8Ž ä z ÿ a   Sð Æ ˆs  þÿ H   ô ˆƒ €þÿ H   è ˆƒ  ÿÿ H   è ˆƒ €ÿÿ H   â G‰  p     € ê ˆs    H   è ˆs €þÿ H   è ˆs 
€  H   è ˆs  ÿÿ H   è ˆs 2   H   è ˆs €ÿÿ H   è ˆs 	€  H   ä Ay        € ê {          ì M        € ê y      &   "> z ÿ `  pPð Ä x        â y      '   b z ÿ a   Sð â $z x  Ž È $z ~  Ž È/ %v j   Ž È G	  P    € ê „y	      H   " z  r  p`ð Æ „y     H   br  ÿ   pVp  Æ „y     H   ¨ „y
  €   H   è „y     H   ( „y  €  H   ( „y     H   ( „y  €  H   ( „y     H   ( „y  €  H   ( „y  €  H   â !r		       Æ „y  €  H   â !r	       È/ !r       âO !r

       È� ‰      " !r       â!r

       Æ !r       â !r

       Æ !r       ä ‰	
      b !r       Æ ‰      ¢ !r       â !r       È ‰      " !r       Ê ‰      â !r	
	       â/ !r       ÈO ‰	      h ‰
      ¢ !r       â !r       È� ‰ €    ( ‰      â !r	       â/ !r
       ÈO ‰ €    h ‰ €    ¢ !r       â x   ÿàÿ â !r       Æ� ‰ @    " z  r  p`ô Æ ‰ €    â x     pVt â !r       â/ !r       ÈO ‰
 @    h ‰ @    ¢ !r       â x   ÿàÿ äx   ÿàÿ â !r       â� ‰      " z  r  p`ò Æ ‰ @    â x     pVò  â !r

       â/ !r       ÈO ‰	
      h ‰      ¢ !r       â !r       Æ� !‚       ä ‰      ( †‰     â z  r  p`ð â !r
	       Æ/ x     pVp  â !r	       âO !’       Æ !¢		       ä †™    è †© 	   â !r       â M	        € ì !r       Ê †y    â My        € ê z  h      È x    pBð Ú G	  @    € ê x    pRð Ú G‰  À    € ê x    pRð Ú G	   $    € ê „y      H   " z  r  p`ð â Ey  p    € ä „y     H   b r  ÿ   pVp  Æ „y     H   ¨ „y     H   è „y     H   ( „y   
  H   ( „y  €   H   ( „y  €  H   ( „y  €  H   è „y  €  H   " !r       â !r       Æ/ !r       âO !r       Æ� !r       â!r       È ‰      " !r       Æ ‰      b !r       È !r       â !r       ä „y  €  H   â !r       Æ/ ‰      ( ‰      b !r       â !r       È/ ‰ €    ( ‰ €    b !r
       Æ „y  €
  H   ¢ !r       â !r       â !r       È/ ‰ @    ( ‰ @    b !r       âO !r       â !r       È/ ‰      ( ‰      b !r       â !r       â/ G	  �     € ê !r       â!r       Æ  x
;ª¸¿  A  Ì s
 
       $ !t
  €?    Ì s        $  r     A  È  r     A  Ê †y     ä Ay        € ê ‰      " Ey  Ð    € æ ‰      h „y     H   ¨ „y     H   è „y     H   è „y     H   â !r       Æ „y     H   â !r       Æ/ ‰      ( ‰      b !r       âO !r2       â� !r       ä !r       â !r       Ä/ ‰ €    " !r       Æ ‰
 €    h „y   	  H   ¢ !r       â !r
       È/ ‰
 @    ( ‰ @    b !r       ÈO !r       â !r

       â x   ÿàÿ â !r       Æ/ ‰
      " z  r  p`ð Æ ‰      b x     pVp  â !r
       â !r       Ö/ G	  �     € ê !r       â!r       Æ  x;ª¸¿  A  Ì s        $ !t
  €?    Ì s
 
       $  r
     A  È  r     A  Ê †y    ä Ay        € ê ‰      b Ey       € æ ‰      ( „y  €	  H   è „y  €  H   è „y  €  H   è „y  €  H   â !r       Æ/ „y  €  H   â !r
       Æ ‰      ( ‰
      h „y  €  H   ¢ !r       â !r
       È/ ‰ €    ( ‰ €    b !r	       ÈO !r       â !r       â x   ÿàÿ ä !r       â !r       â/ ‰ @    " z  r  p`ð Æ ‰ @    b x     pVp  â !r       â !r       È/ ‰      ( ‰      b !r       â !r       â/ G	  �     € ê !r       â!r       Æ  x;ª¸¿  A  Ì s        $ !t	  €?    Ì s
 	       $  r
     A  È  r     A  Ê †y    ä Ay        € ê ‰      " !r       â x   ÿàÿ Æ !r       â z  r  p`ð Æ !r       â x     pVp  È ‰      b !r	       Ê ‰	      " !r       Ê/ ‰      b !r
	       Ê ‰
 €    " !r       Ê/ ‰ €    b !r
       Ê ‰ @    " !r       Ê/ ‰ @    b !r       Ê ‰	      " !r       Ê/ ‰      ¢ !r		       â M	        € ì !r	       â/!r       ÆO  x ;ª¸¿  A  â !r       Ê s          $ !t   €?    Ì s        $  r     A  È  r     A  Ê †y    â My        € ê „y      H   " z  r  p`ð â Ey       € ä „y     H   b r  ÿ   pVp  Æ „y     H   ¨ „y     H   è „y     H   ( „y   
  H   ( „y  €  H   è „y  €  H   â !r       â !r       Æ/ !r       âO !r       Æ� !r       â!r       È ‰      ( ‰      h „y  €  H   ¢ !r       â !r       È/ ‰      ( ‰      h „y  €   H   â !r
       âO !r       â !r       È/ ‰ €    ( ‰ €    h „y  €  H   ¢ !r       È� !r       â !r       ä „y  €
  H   â !r       â !r       ä/ ‰ @    ( ‰ @    b !r       âO !r       â !r       È/ ‰      ( ‰      b !r       â !r       â !r       â/ G	  À     € ê !r       È x
'7=  A  â x*BL?  A  Æ #t  €?
   È  r     A  â  x   ?  A  â !r       È s     $   $ !t  €?    È  r     A  È  r     A  Ê †y     ä Ay        € ê ‰      " Ey       € æ ‰      h „y     H   ¨ „y   	  H   è „y     H   è „y     H   â !r       â !r       È/ ‰      ( ‰      h „y     H   â !r       âO !r       â !r       È/ ‰ €    ( ‰
 €    h „y     H   ¢ !r2       È� !r       â !r       Æ !r       â !r
       ä/ ‰
 @    ( ‰ @    b !r       ÈO !r       â !r

       â x   ÿàÿ â !r       Æ/ ‰
      " z  r  p`ð Æ ‰      b x     pVp  â !r
       â !r       Ö/ G	  À     € ê !r       È x'7=  A  â x*BL?  A  Æ #t  €?   È  r     A  â  x   ?  A  â !r       È s     $   $ !t
  €?    È  r
     A  È  r     A  Ê †y    ä Ay        € ê ‰      b Ey  Ð    € æ ‰      ( „y  €  H   ¨ „y  €	  H   è „y  €  H   è „y  €  H   â !r       â/ !r
       È ‰      ( ‰
      h „y  €  H   â !r	       ÈO !r       â� !r       â !r
       È/ ‰ €    ( ‰ €    h „y
  €  H   ¢ !r       â !r       È/ ‰ @    ( ‰ @    b !r
       âO !r       â x   ÿàÿ â !r       Æ/ ‰      " z  r  p`ð Æ ‰      b x     pVp  â !r       â !r       Ö/ G	  À     € ê !r       È x	'7=  A  â x*BL?  A  Æ #t  €?	   È  r     A  â  x   ?  A  â !r       È s     $   $ !t
  €?    È  r
     A  È  r     A  Ê †y    ä Ay        € ê ‰      b !r       â x   ÿàÿ Æ !r       â z  r  p`ð Æ !r       â x     pVp  È ‰      ¢ !r	       Ê/ ‰	      b !r       ÊO ‰      ¢ !r
	       Ê/ ‰
 €    b !r       ÊO ‰ €    " !r
       Ê/ ‰ @    b !r       Ê ‰ @    " !r       Ê/ ‰	      b !r       Ê ‰      ¢  !r		       â/ M	        € ì !r	       È x'7=  A  â x *BL?  A  Æ #t  €?   â x   ?  A  Æ  r       A  â !r       ÈO !r       â s       $   $ !t   €?    È  r     A  È  r     A  Ê †y    â My        € ê x    pRð Ú G‰  P    € ê x    pRð Ú G	  `    € ê „y      H   " z  r  p`ð â Ey       € ä „y     H   b r  ÿ   pVp  Æ „y     H   ¨ „y     H   è „y     H   ( „y   
  H   ( „y  €  H   è „y  €  H   â !r       â !r       Æ/ !r       âO !r       Æ� !r       â!r       È ‰      ( ‰      h „y  €  H   ¢ !r       â !r       È/ ‰      ( ‰      h „y  €   H   â !r
       âO !r       â !r       È/ ‰ €    ( ‰ €    h „y  €  H   ¢ !r       È� !r       â !r       ä „y  €
  H   â !r       â !r       ä/ ‰ @    ( ‰ @    b !r       âO !r       â !r       È/ ‰      ( ‰      b !r       â !r       â !r       â/ G	  À     € ê !r       Ê	z i    � â !r       È  x
;ª¸¿  A  â 	z i    � È 	z i €  � â s
 
       $ !t
  €?    Ì s        $  r     A  È  r     A  Ê †y     ä Ay        € ê ‰      " Ey       € æ ‰      h „y     H   ¨ „y   	  H   è „y     H   è „y     H   â !r       â !r       È/ ‰      ( ‰      h „y     H   â !r       âO !r       â !r       È/ ‰ €    ( ‰
 €    h „y     H   ¢ !r2       È� !r       â !r       Æ !r       â !r
       ä/ ‰
 @    ( ‰ @    b !r       ÈO !r       â !r

       â x   ÿàÿ â !r       Æ/ ‰
      " z  r  p`ð Æ ‰      b x     pVp  â !r
       â !r       Ö/ G	  À     € ê !r       Ê	z i    � â !r       È  x;ª¸¿  A  â 	z i    � È 	z i €  � â s        $ !t
  €?    Ì s
 
       $  r
     A  È  r     A  Ê †y    ä Ay        € ê ‰      b Ey  Ð    € æ ‰      ( „y  €  H   ¨ „y  €	  H   è „y  €  H   è „y  €  H   â !r       â/ !r
       È ‰      ( ‰
      h „y  €  H   â !r	       ÈO !r       â� !r       â !r
       È/ ‰ €    ( ‰ €    h „y
  €  H   ¢ !r       â !r       È/ ‰ @    ( ‰ @    b !r
       âO !r       â x   ÿàÿ â !r       Æ/ ‰      " z  r  p`ð Æ ‰      b x     pVp  â !r       â !r       Ö/ G	  À     € ê !r       Ê	z i    � â !r       È  x;ª¸¿  A  â 	z i    � È 	z i €  � â s        $ !t	  €?    Ì s
 	       $  r
     A  È  r     A  Ê †y    ä Ay        € ê ‰      b !r       â x   ÿàÿ Æ !r       â z  r  p`ð Æ !r       â x     pVp  È ‰      ¢ !r	       Ê/ ‰	      b !r       ÊO ‰      ¢ !r
	       Ê/ ‰
 €    b !r       ÊO ‰ €    " !r
       Ê/ ‰ @    b !r       Ê ‰ @    " !r       Ê/ ‰	      b !r       Ê ‰      ¢  !r		       â/ M	        € ì !r	       â!r       ÈO 	z i    � â !r       È  x ;ª¸¿  A  â 	z i    � È 	z i €  � â s          $ !t   €?    Ì s        $  r     A  È  r     A  Ê †y    â My        € ê „y      H   " z  r  p`ð Æ „y     H   b r  ÿ   pVp  Æ „y  €   H   ¨ „y  €  H   è „y     H   ( „y     H   ( „y     H   ( „y     H   ( „y  €  H   ( „y  €  H   ( „y     H   " !r       Æ „y   	  H   " !r       Æ/ „y     H   b !r       ÆO „y   
  H   ¢ !r
       Æ� „y  €  H   â !r       Æ„y  €
  H   " !r2       Æ „y     H   " !r       Æ „y!     H   " !r       Æ „y  €  H   " !r       â x   ÿàÿ ä „y   €  H   " !r       â z  r  p`ò ä „y#  €  H   " !r       â x     pVò  Ä „y%  €	  H   â !r       Æ „y'  €  H   â !r       Æ/ „y)  €  H   â !r       âO !r       ä� ‰      " !r       Æ‰      b x   ÿàÿ â !r       ä ‰      ¢ z  r  p`ô â !r!!       ä ‰
      â x     pVt â !r       Ä ‰      " x   ÿàÿ â !r  	       ä ‰!      " z  r  p`ö â !r#       â !r       ä x     pVö â !r  %       â !r       â/ ‰      " !r'       â !r	       ÄO ‰      b !r) )       â !r
       ä� ‰      ¢ !r       Æ‰)      â !r!       Æ ‰
	      ( ‰      " !r       Æ ‰      ( ‰      " !r       â/ !r       äO ‰ €    b !r)       Æ� ‰ €    ¢ !r
	
       Æ‰      â !r       Æ ‰      " !r       Æ ‰	
 €    " !r       Æ ‰ €    ( ‰ €    " !r       Æ/ ‰ €    b !r       âO !r       ä� ‰ @    ¢ !r       Æ‰ €    â !r	
	       Æ ‰ €    " !r       Æ ‰ @    " !r       Æ ‰
	 @    " !r       Æ/ ‰ @    h ‰ @    b !r       ÆO ‰ @    ¢ !r       â� !r       ä ‰      " !r       Æ‰ @    â !r
	
       Æ ‰ @    " !r       Æ/ ‰      b !r       Æ ‰	
      b !r       ÆO ‰      ¨ ‰      ¢ !r       Æ ‰      " !r       â� !‚       â!r       Æ‰      â !r       Æ/ ‰      b !r
	       â !‚       â !r       äO !’       â  ‚     A  â !r       â !’       Æ †‰     â !r       â !¢       â  ’     A  ä !¢
       â !r       ä� †™    â  ¢
     A  â !r       È/ †©    â M9        € ê !r       â !r       ÈO  r     A  Ê †y    â My        € ê „y      H   " z  r  p`ð â Ey  À    € ä „y     H   b r  ÿ   pVp  Æ „y     H   ¨ „y     H   è „y     H   ( „y   
  H   ( „y  €  H   è „y  €  H   â !r       â !r       Æ/ !r       âO !r       Æ� !r       â!r       È ‰      ( ‰      h „y  €  H   ¢ !r       â !r       È/ ‰      ( ‰      h „y  €   H   â !r
       âO !r       â !r       È/ ‰ €    ( ‰ €    h „y  €  H   ¢ !r       È� !r       â !r       ä „y  €
  H   â !r       â !r       ä/ ‰ @    ( ‰ @    b !r       âO !r       â !r       È/ ‰      ( ‰      b !r       â !r       â !r       â/ G	  à     € ê !r       Ê	x  à@  � Ê  x#ÛÙ?  A  È  x
;ª¸¿  A  â !r       Ê s
 
       " 	x  à@  � È 	x  àÀ  � Ê !t  €?    â !t
  €?    Ì s        $  r     A  È  r     A  Ê †y     ä Ay        € ê ‰      " Ey       € æ ‰      h „y     H   ¨ „y   	  H   è „y     H   è „y     H   â !r       â !r       È/ ‰      ( ‰      h „y     H   â !r       âO !r       â !r       È/ ‰ €    ( ‰
 €    h „y     H   ¢ !r2       È� !r       â !r       Æ !r       â !r
       ä/ ‰
 @    ( ‰ @    b !r       ÈO !r       â !r

       â x   ÿàÿ â !r       Æ/ ‰
      " z  r  p`ð Æ ‰      b x     pVp  â !r
       â !r       Ö/ G	  à     € ê !r       Ê	x  à@  � Ê  x#ÛÙ?  A  È  x;ª¸¿  A  â !r       Ê s        " 	x  à@  � È 	x  àÀ  � Ê !t  €?    â !t
  €?    Ì s
 
       $  r
     A  È  r     A  Ê †y    ä Ay        € ê ‰      b Ey  ð    € æ ‰      ( „y  €  H   ¨ „y  €	  H   è „y  €  H   è „y  €  H   â !r       â/ !r
       È ‰      ( ‰
      h „y  €  H   â !r	       ÈO !r       â� !r       â !r
       È/ ‰ €    ( ‰ €    h „y
  €  H   ¢ !r       â !r       È/ ‰ @    ( ‰ @    b !r
       âO !r       â x   ÿàÿ â !r       Æ/ ‰      " z  r  p`ð Æ ‰      b x     pVp  â !r       â !r       Ö/ G	  à     € ê !r       Ê	x  à@  � Ê  x#ÛÙ?  A  È  x;ª¸¿  A  â !r       Ê s        " 	x  à@  � È 	x  àÀ  � Ê !t  €?    â !t	  €?    È s
 	       $  r
     A  È  r     A  Ê †y    ä Ay        € ê ‰      b !r       â x   ÿàÿ Æ !r       â z  r  p`ð Æ !r       â x     pVp  È ‰      ¢ !r	       Ê/ ‰	      b !r       ÊO ‰      ¢ !r
	       Ê/ ‰
 €    b !r       ÊO ‰ €    " !r
       Ê/ ‰ @    b !r       Ê ‰ @    " !r       Ê/ ‰	      b !r       Ê ‰      ¢  !r		       â/ M	        € ì !r	       â!r       ÈO 	x  à@  � â !r       È  x #ÛÙ?  A  â 	x  à@  � Æ  x  ;ª¸¿  A  â 	x  àÀ  � Ê s          " !t  €?    â !t   €?    Ì s        $  r     A  È  r     A  Ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      "   $ z  w  p`ð Ú M	        € ê y      !   " ¹z  l      â Ey  0    € â ™x?    â y      %   b ¹z  F   
   â x      ÿ  â xÿ     È |    p`ð â $x   ÿ Ž Ø/ G	       € ê y      &   " $tÿ   ÿ Ž ä $z x  Ž È %v \   Ž Ì �y    ™ ¢  $rÿÿ   ÿ Ž â x	   ÿàÿ â $z r  ÿŽ ä 'z  m   Ž È $t
ÿ   ÿ Ž â z ÿ n     â $x   ÿ Ž ä $rÿÿ   ÿ Ž ä $rÿÿ    
Ž â x   ÿÀŽ â $z  q  Ž ä $z o  Ž È $z u  ÿŽ È %x$   ÿ Ž Ì %x$    Ž â x   ÿÀŽ Ê %x$   Ž È $x   ÿ Ž È %x$   Ž Ê z Z  ÿàõ È x   ÿà÷ È zÿ [  Dþ â $z t  Ž äO$z t   Ž ä $rÿÿ   ÿ Ž ä $z		 q  Ž ä %x   Ž È %x	   Ž â z X  ÿàñ Æ %v	 X  
Ž âz X  ÿàó ä x   ÿàõ â %v
 X  
Ž â x    ÿà÷ ä x   ÿàù ä zÿ Y  ~ â $rÿÿ   	Ž â zÿ Y  $~ Æ �y%úÿÿ • ¨  �y$üÿÿ • è  �y#þÿÿ • ( �y ôÿÿ ™ h �yøÿÿ ™ ( �yüÿÿ ™ h �y"úÿÿ • h �y    ™ h �y!    • h �y   ™ h �yüÿÿ • h �y   ™ h �yþÿÿ • h �y    • h �y   ™ b $rÿÿ   
 Ž Ä $r	ÿÿ    Ž â �y   ™ h �y    • h �yðÿÿ • b $rÿÿ     Ž Ä $r	ÿÿ    Ž Ê �yüÿÿ • b ‚x ÿ      ä $~(ÿ   ÿ Ž ä $~)ÿ   ÿ Ž â x   ÿàñ ä x   ÿàÿ Æ $rÿÿ     â |    p`ð ä x

   ÿàó ä x   ÿà÷ ä x€  ÿàù ä/ x     ÿàõ â $rÿÿ   Ž Ä $rÿÿ    ä $rÿÿ   Ž  ä $rÿÿ    â x%%™  ÿ    ÈO x&ÿ   %  ä x%%ww  ÿÀŽ ä x'$™  ÿ    ä� x&&ww  ÿÀŽ ä t	(ÿ %    ä x%ÿ   '  Ä t&)ÿ &    ä x(#™  ÿ    äx#%ww  ÿÀŽ ä x%	@Q  &    â$~$ÿ   ÿ Ž â x		bs  &    â $r&ÿÿ   ( Ž â x'ww  ÿÀŽ â &r% %   ÿ   ât'$ÿ #    Ä t)ÿ     ä x(ÿ   &  â $~#ÿ   ÿ Ž â x)&ww  ÿÀŽ â &r		   %   â x((ww  ÿÀŽ ä x&@Q  '    äxbs  '    Ä t$$ÿ )    â &r	&   	   â t##ÿ (    ä x%"™  ÿ    â &r	   	   â x"$@Q  #    â $~'ÿ   ÿ Ž â x%ww  ÿÀŽ ä x!!™  ÿ    â &r	"   	   â x#$bs  #    Ä x%ÿ   %  â $~$ÿ   ÿ Ž â t'ÿ     â $~&ÿ   ÿ Ž â x"ÿ   !  ä x!!ww  ÿÀŽ ä x'™  ÿ    â &r	#   	   â x%%ww  ÿÀŽ â $~#ÿ   ÿ Ž â x""ww  ÿÀŽ Ä t$ÿ !    â $r$ÿÿ   ' Ž â t%&ÿ %    ä t"#ÿ "    ä x#@Q  %    äx!bs  %    ä xÿ   $  â $~%ÿ   ÿ Ž â x$$ww  ÿÀŽ â $~'ÿ   ÿ Ž â xww  ÿÀŽ â &r# #   ÿ   â x ™  ÿ    Ä t$%ÿ $    ä t%'ÿ     ä x@Q  "    äxbs  "    â $rÿÿ     Ž ä &r!!   #   â x$@Q  %    Ä x ÿ     â $~"ÿ   ÿ Ž â x$$bs  %    â $~#ÿ   ÿ Ž â xww  ÿÀŽ â &r   !   â x  ww  ÿÀŽ ä x™  ÿ    â &r$      â t"ÿ     Ä t #ÿ      ä xÿ     â &r	   	   â x@Q       â$~ÿ   ÿ Ž â xww  ÿÀŽ â $~!ÿ   ÿ Ž â xww  ÿÀŽ â &r      â xbs       Ä tÿ     ä t!ÿ     â &r      Æ x@Q      äxbs      Æ &r      ä &r   	   ä0rÿ    A   ä 0rÿ    A   ä &r      ä 0r	ÿ    A   â  r     A  â Er        Ä Er        â  r	     A  ä #r      ä #r      â G‰  ÷ÿÿÿÿƒ ê Ay        € ê ‰      " $x
   Ž Æ ‰       d z 
 p  p`ð È x    pDp  â !r       â !r        È/ ‰       ( ‰      b !r         â !r       È/ ‰  €    ( ‰ €    b !r        â !r       È/ ‰ @    ( ‰ @    b !r       â !r       È/ ‰      h  ‰	      ¢  M	        € ê y      &   â $x   ÿ Ž â !r       â/ $z
 s  
Ž â !r		       âO $tÿ   ÿ Ž â x    p"ò Ä r ÿ   p"ð Ö $ÿÿ   	 Ž ä $z v  
Ž ÈŸ %v j   Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      "   $ z  w  p`ð Ú M	        € ê y	      &   " $tÿ   ÿ Ž â ¹z  F   
   â $z x  	Ž È %v \   Ž Ê �y     ™ b ¹z  l      â Ey       € â ™x?    â y      !   b $rÿÿ   ÿ Ž â x      ÿ  â $rÿÿ   ÿ Ž ä $rÿÿ   ÿ Ž ä $rÿÿ   ÿ Ž â xÿ     È/ |    p`ð Ú G	  p    € ê $rÿÿ   ÿ Ž â y      %   " $z r  ÿŽ â z ÿ `  pPð â 'z	 m   Ž â x   ÿ   â Ey à    € â z ÿ a   Sð â $rÿÿ   ÿ Ž â zÿ n    ä x   ÿÀŽ Æ $rÿÿ   
Ž â x   ÿÀŽ Æ $z o  	Ž È $z u  ÿŽ ä $x   ÿ Ž Æ r   à£ È rÿÿ   ÿDþ  ä $z!  t  Ž ä$tÿ   ÿ Ž ä $z! q  !Ž ä $x   ÿ Ž ä %v! X  Ž Ì %x
    Ž â r   ÿàó â�y    • ¨  �y$
   • â r   ÿ�  â $tÿ$   ÿ Ž ä �y&
   • " z! q  ÿàÿ â %v Z   Ž Ä �y(
   • $ $x$   ÿŽ ä �y)
   • ¤ $x   Ž Ê �y   ™ ¢ %v X  Ž Æ �y   ™ ¦ %x    Ž â �y   ™ ( �y'   • ( �y   ™ ( �y%   • ( �y   ™ ( �y    • ( �y   ™ ( �y"   • ( �y   ™ ( �y    ™ ( �y#    • ( �y    • " ‚x ÿ      Ä $~+ÿ   ÿ Ž â |       â/ $~	ÿ   ÿ Ž â x   ÿàÿ È |    p`ò ä x
$™  ÿ    È� xÿ   
  ä x

ww  ÿÀŽ ä xww  ÿÀŽ ä x&™  ÿ    ä t
+ÿ 
    ä tÿ     Ä xÿ     ä xww  ÿÀŽ ä x
@Q      äxww  ÿÀŽ ä x
bs      â &r   ÿ   âO t	ÿ     Ä t+ÿ     â &r      â x(™  ÿ    äx@Q      äxbs      ä xÿ     â $~
ÿ   ÿ Ž â xww  ÿÀŽ Ä xww  ÿÀŽ ä t	ÿ     ä t
ÿ     ä x''™  ÿ    â &r      â x@Q      äxbs      â $rÿÿ   ' Ž â x	)™  ÿ    â &r      Æ x
ÿ     â $~ÿ   ÿ Ž â x	ww  ÿÀŽ â $~'ÿ   ÿ Ž â x$%™  ÿ    â $~)ÿ   ÿ Ž â xww  ÿÀŽ ä x

ww  ÿÀŽ ä tÿ     Ä x%ÿ   $  ä t'ÿ     ä t)ÿ 
    ä x'%ww  ÿÀŽ ä |
       ä x$$ww  ÿÀŽ Ä x%@Q      â&r      â xbs      ä t$)ÿ $    â &r%%   ÿ   â t'
ÿ '    ä x  ™  ÿ    â &r      â x$@Q  '    â &r   %   Ä $rÿÿ     Ž ä &r      Æ xÿ     â $~ ÿ   ÿ Ž â x	ÿ   	  ä xww  ÿÀŽ ä xww  ÿÀŽ â $~&ÿ   ÿ Ž â x"™  ÿ    ä x		ww  ÿÀŽ Ä t
ÿ     ä t
 ÿ     ä x$$bs  '    ä xÿ     â $~%ÿ   ÿ Ž â t	&ÿ 	    â $~'ÿ   ÿ Ž â x ww  ÿÀŽ â &r$      â x"ww  ÿÀŽ Ä x@Q  
    ä x@Q  	    ä t %ÿ      â &r      â xbs  
    ä t%'ÿ "    â &r      ä &r      â x @Q  %    Ä x	bs  	    ä x% bs  %    â &r      ä &r		      ä 0rÿ    A   ä 0rÿ#    A   ä 0rÿ    A   Ä &r%      â Er
 	       â  r     A  â  r     A  ä Er	        â G‰  €    € ê xÿ   !  â %v! `   Ž È $x   ÿŽ È $x   Ž È %x    Ž ä �y     ¨  �y"    è �y!    ( �y    h	 �y     ¢ $~%ÿ   ÿ Ž Ä $~$ÿ   ÿ Ž ä $~ÿ   ÿ Ž ä $~&ÿ   ÿ Ž â x#ÿ   "  ä� x""ww  ÿÀŽ ä x##ww  ÿÀŽ ä t"%ÿ "    Ä t#$ÿ #    ä xÿ   !  äx%"@Q  #    äx#"bs  #    ä |"       ä x!!ww  ÿÀŽ Ä xww  ÿÀŽ ä t!"ÿ !    â &r%%   ÿ   â t"ÿ     ä xÿ     â&r#   %   â x!@Q  "    ä xww  ÿÀŽ Ä xww  ÿÀŽ â &r      â t&ÿ     ä xÿ      äO x"!bs  "    ä t$ÿ     â $~#ÿ   ÿ Ž â x!ww  ÿÀŽ â &r""      â x  ww  ÿÀŽ Ä x@Q      äxbs      ä t #ÿ      â &r   "   â t!$ÿ !    Æ &r      â x @Q  !    äx! bs  !    Æ &r      È &r!!      ä 0rÿ    A   Æ Er! !       ä  r     A  È #r!      â #r
      â #r	      â G‰  �    € ê xÿ     â %v `   Ž È $x   ÿŽ È $x   Ž È %x    Ž ä �y     ¨  �y	    è �y    ( �y    h �y    ¢ $~ ÿ   ÿ Ž Ä $~ÿ   ÿ Ž ä $~ÿ   ÿ Ž â |       ä/ x
ÿ   	  ä� x		ww  ÿÀŽ ä x

ww  ÿÀŽ ä t	 ÿ 	    Ä t
ÿ 
    ä xÿ     äx	@Q  
    äxww  ÿÀŽ ä xww  ÿÀŽ â &r   ÿ   â x		bs  
    Ä tÿ     ä tÿ     ä xÿ     â&r	      â x@Q      â $~	ÿ   ÿ Ž â xww  ÿÀŽ â $~
ÿ   ÿ Ž â xww  ÿÀŽ Ä xbs      ä xÿ     âO &r      â t	ÿ     ä t
ÿ     â $~ÿ   ÿ Ž â x	ww  ÿÀŽ â &r      â xww  ÿÀŽ Ä x@Q      äxbs      ä t
ÿ     â &r      â tÿ 	    Æ &r      â x@Q      äxbs      Æ &r      ä 0rÿ    A   ä &r      Æ  r     A  ä Er        Ê #r      ä D       À â Gy  ðïÿÿÿÿƒ ê y           È Ay       € ê $rÿÿ    Ž ä $rÿÿ    Ž ä Ay        € ê z ÿ `  pPð È z ÿ a   Sð Ú G	  P    € ê ‰      ( ‰      b !r       â !r       È/ ‰      ( ‰      b !r       â !r	       È/ ‰ €    ( ‰	 €    b !r       â !r
	       È/ ‰ @    ( ‰
 @    b !r       â !r
       È/ ‰      ( ‰      b !r       â !r       â/ Gy  €    € ê ‰      ( ‰      h ‰      ¨ ‰      â !r       â !r       È/ ‰      " !r       âO !r       È� ‰      h ‰      ¨ ‰
      â !r       â !r       È/ ‰ €    " !r	       ÆO ‰ €    b !r
       Æ� ‰
	 €    ¨ ‰ €    â !r       â !r       È/ ‰ @    " !r
	
       ÆO ‰ @    b !r       Æ� ‰
 @    ¨ ‰ @    â !r       â !r       È/ ‰      " !r
       ÆO ‰	      b !r       Æ� ‰      ¨ ‰      â !r       â !r	       â/ !r       âO !r       â� y      %   $ $x   Ž Ê z  p  p`ð È x    pDp  Ú M	        € ê z ÿ ^  pPð â $z  v  Ž Æz ÿ _   Sð Ú $ÿ   ÿ Ž È % ^   Ž Ì �	     ¢ x    ÿ   È r  ÿ   p"ò äx     p"ö äx     p"ø äx     p"ô Î $ ÿÿ    Ž â z ÿ `  pPò â $2 ÿÿ    Ž ä $B ÿÿ    Ž â z ÿ a  Sò â $" ÿÿ    Ž È !         ÐO G™       € ê z ÿ b  pPð È z ÿ c   Sð Ú $ÿ   ÿ Ž È % b   Ž Ì �	     ¢ $x   ÿ Ž Ê x øÿÿÿp"ò äx üÿÿÿp"ô är ÿ   p"ö äx    p"ø Î         ä $"ÿÿ    Ž ä $2ÿÿ    Ž ä $vÿ h  ÿ Ž ä $Bÿÿ    Ž Æ x    pBò â !       ØO G  P    € ê x    pRð Ú G‰  �     € ê x    pRð Ú G	  ð    € ê  x;ª¸¿  A  Ì s        $ !t  €?    Ì s        $  r     A  È  r       A  â Gy  P    € ê  x'7=  A  â x*BL?  A  Æ #t  €?   â  x   ?  A  Æ  r     A  Ì s     $   $ !t  €?    È  r     A  È  r       A  â Gy  °    € ê x    pRð Ú G‰  à     € ê x    pRð Ú G	        € ê 	z i    � ä 	z   i    � Æ  x;ª¸¿  A  â 	z   i €  � Ê s        $ !t  €?    Ì s        $  r     A  È  r       A  â Gy  Ð     € ê  r       A  â Gy  °     € ê 	x  à@  � ä 	x    à@  � Æ  x#ÛÙ?  A  â 	x    àÀ  � Æ  x;ª¸¿  A  ä !t   €?    È s        $ !t  €?    Ì s        $  r      A  È  r       A  ä y      "   ( y      &   b $z s  Ž ä $tÿ   ÿ Ž ä $z v  Ž È/ %v j   Ž Ê †y      â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      "   " ¹z  l      â Ey       € â ™x?    â y      !   b ¹z  F   
   â r  ÿ       â x      ÿ  â y      %   ¢ r ÿ       â x      ÿ  â x      ÿ  â r ÿ       â x      ÿ  â x      ÿ  â x      ÿ  â r ÿ       Ä r ÿ   pRð är ÿ   pRò är   ÿ(Ž È/ xÿ     ä x   ÿ   äO r   ÿ� È x ÿ     È |     pbô Ú G)       € ê y      '   " r ÿ       â $z p   Ž â xüÿÿÿÿÀŽ â y	      &   b r ÿ       ä r
  €ÿàÿ ä x$        ä x
   ÿ   Ä x&    ÿ   äz(  q  ÿàÿ ä x'ÿ   
  ä x((   ÿ   ä z*& q  ÿàÿ â 'z  y   Ž â x   ÿàÿ Æ 'z	 s   Ž â/ zÿ z     Æ $z!	 w  ÿŽ â z ÿ t    â $z |  Ž ä$z |   Ž ä %x   ÿŽ È $z  v  Ž ä$z  v  Ž â r  ÿ       â %x   Ž È $z p  Ž â z X  ÿàû â $z }  ÿŽ â z  q      â %x   Ž â x   ÿàý ä r!!   
àµ â %v" X  $Ž âz X  ÿàù ä zÿ Y  ¤~ â %v$ X  $Ž â x   ÿàû Ä r4&   ÿŽ â$x,   &Ž âzÿ Y  „þ â $x2   &Ž âr'ÿÿ   'd~ â $x.   &Ž ä$x0   &Ž ä �yxþÿÿ • ¨ �yV    • è �yDþÿÿ • ( �y>    • b r&!   ÿàõ Ä xm $       â �y:"    • ¢ r	ÿ'   ÿä ä r*!   ÿàõ â %v Z  m Ž â �y=$    • ä r
ÿ'   ÿä â $x	$   ÿŽ ä %v Z  m Ž Æ r   ÿàÿ â $x
$   ÿŽ È �y/   ™ â  r		   ÿàÿ Æ �y)   ™ è  �y6   ™ è �y8   ™ è �y+   ™ ( �y7   ™ ( �y9   ™ ( �y-   ™ ( �y1   ™ " rX4!   ÿàõ Æ �y3   ™ " r
,!   ÿà÷ Æ �y5   ™ " rÿ'   ÿä Æ �y<   ™ " rÿ'   ÿäÿ Æ �y;   ™ " %vXX Z  m Ž Æ �y?   ™ " $x$   ÿŽ Æ �yA    ™ " %v

 Z  m Ž Æ �yB    • " $xE$   ÿŽ Æ �y@    ™ b rYY   ÿàÿ Æ �yC    • b rE   ÿàÿ ä r 
       â �yEX   ™ h �yG   ™ h �yHX   ™ h �yF   ™ h �ySX   ™ h �yR   ™ h �yQX   ™ h �yP   ™ h �yOX   ™ b r(!   ÿàõ Æ �yN   ™ b r
2!   ÿà÷ Æ �yMX   ™ b r	ÿ'   ÿä Æ/ �yL   ™ b rÿ'   ÿäÿ Æ �yKX   ™ b %v Z  m Ž Æ �yJ   ™ b $x		$   ÿŽ Æ �yIX    ™ b %v

 Z  m Ž Æ �yTX    • b $x[$   ÿŽ Æ �yU    ™ b r	   ÿàÿ Æ �yW    • b r[   ÿàÿ Æ �yX   ™ h �yZ
   ™ h �y[   ™ h �yY
   ™ h �yf   ™ h �ye
   ™ h �yd   ™ h �yc
   ™ b r.!   ÿàõ Æ �yb   ™ b r0!   ÿà÷ Æ/ �ya
   ™ b rÿ'   ÿä Æ �y`   ™ b rjÿ'   ÿäÿ Æ �y_
   ™ h �y^   ™ b $xk$   ÿŽ Æ �y]
   ™ b %v Z  m Ž Æ �y\    ™ b %v Z  m Ž Æ �yg    • b $xmj$   ÿŽ Æ �yh
    ™ b r		k   ÿàÿ Æ �yi
    • b rm   ÿàÿ Æ �y   ™ h �yt   ™ h �y
   ™ h# �ys   ™ h �yr   ™ h �yq   ™ h �yp   ™ h �yo   ™ h �yn   ™ h �ym   ™ h �yl   ™ h �yk   ™ h �yj   ™ h �y   ™ h �y    • h �yv    ™ h �yw    ™ h �yu    • b xxx™  ÿ    âO ‚x       Æ xyxww  ÿÀŽ ä xxÿ   x  ä |z       ä |{       ä xxxww  ÿÀŽ ä tyz y    Ä tx{ x    â ‚x ÿ       â xyww  ÿÀŽ ä/ |	       ä xÿ   x  ä xxxww  ÿÀŽ ä xV™  ÿ    Ä� t	ÿ      ä ||       ä xww  ÿÀŽ ä xyÿ   y  ä t		ÿ  x    ä xxww  ÿÀŽ Ä xÿ     ä tV|ÿ      ä |z       ä xyyww  ÿÀŽ ä xT  	    äx2v  	    Ä ||       ä x	ww  ÿÀŽ ä tyzÿ  y    ä tx{ x    ä t	| 	    ä |{       Ä xzxww  ÿÀŽ ä xyT  V    äxVy2v  V    ä xyÿ   	  ä tz{ÿ  z    ä ||       Ä x{D™  ÿ    äx		ww  ÿÀŽ ä |       ä xxÿ   x  ä xDyww  ÿÀŽ ä t	|ÿ  	    Ä |}       ä xxxww  ÿÀŽ ä t|ÿ  D    ä xDÿ   {  ä ty}ÿ  x    ä |}       Ä xxDww  ÿÀŽ ä xDzT  	    ä|~       ä x{{ww  ÿÀŽ ä x	z2v  	    ä tz} x    Ä t{~ {    ä x~zww  ÿÀŽ ä x>>™  ÿ    äxzÿ   z  ä xxyT  |    ä |€       Ä x}{ww  ÿÀŽ ä xyy2v  |    ä x|ÿ   >  ä xzzww  ÿÀŽ ä |�       ä x>>ww  ÿÀŽ Ä t}€ÿ  }    ä x{ÿ   {  ä t~ÿ  ~    ä|ƒ       ä x€|ww  ÿÀŽ ä tÿ  z    Ä tz� >    ä |‚       ä x{{ww  ÿÀŽ ä t€ƒ €    ä x�zww  ÿÀŽ ä t|‚ÿ  {    Ä xzÿ   z  ä |„       ä |ƒ       ä x‚€ww  ÿÀŽ ä x{}T  ~    äx>}2v  ~    Ä |~       ä x}zww  ÿÀŽ ä t�„ÿ  �    ä t‚ƒÿ  ‚    ä xz|T      ä tƒ~ÿ  }    Ä x||2v      â &r/   ÿ   âx~�T  ‚    äx}�2v  ‚    â &r�{/   ÿ   â x€ÿ   €  â &r)      ä&r�>)   �   ä &r)6   ÿ   â |…       Ä x€€ww  ÿÀŽ â &r)8   )   ä &r+      ât„…ÿ  €    â &r)7   )   ä &rV9      âx€ƒT  „    â&r)V-   )   â x/ƒ2v  „    â &rƒ{6   ÿ   Ä &r6D1      ä&r)D3   )   ä &r6	5   6   ä&r)	<   )   ä &r6x;   6   ä&r)x?   )   Ä &rƒ>8   ƒ   ä &r6yA   6   ä&r�z+   �   ä 0r:ÿ:    A   ä 0rBÿB    A   ä &r+y@   )   Ä &rƒz7   ƒ   ä 0rCÿC    A   â Er6 6       â  r):B     A  â &rƒ|-   ƒ   â Er+ +       â  r-:C     A  â #r)6      â &r)E   ÿ   ä #r-+      â &r+G   ÿ   Ä &rƒ~3   ƒ   ä &r)H   )   ä&r+F   +   ä &rƒ}<   ƒ   ä &r)S   )   ä&r+R   +   Ä &rƒ€?   ƒ   ä &r)VQ   )   ä&r+VP   +   ä &rƒ/@   ƒ   ä 0r=ÿ=    A   ä &r6DO   )   Ä &r+DN   +   â Erƒ ƒ       â &r6	M   6   â  r<=C     A  â &r+	L   +   ä &r6xK   6   â #r<ƒ      â &r�|9   �   ä &r<xJ   +   Ä &r6yI   6   ä&r8~1   �   ä 0rTÿT    A   ä &r<yU   <   ä 0rWÿW    A   â Er6 6       â &r8}5   8   â  r):T     A  âEr< <       â &rE{E   ÿ   â  r+:W     A  â &r8€;   8   â #r)6      â &r)X   ÿ   â #r+<      â &rE>H   E   Ä &r8/A   8   ä &r+Z   ÿ   ä &r)[   )   ä &rEzS   E   â Er8 8       â &r+Y   +   â  rBB=     A  â &r)f   )   Ä &rE|Q   E   ä &r+e   +   â #rB8      â &r)Vd   )   ä &r8~O   E   ä &r+Vc   +   ä &r)Db   )   Ä &r8}M   8   ä &r6Da   +   ä &r)	`   )   ä&r8€K   8   ä &r6	_   6   ä &r)x^   )   Ä &r8/I   8   ä &r6x]   6   ä &r)y\   )   â Er8 8       â 0rgÿg    A   â  rT=T     A  â &r+yh   6   ä 0riÿi    A   â Er) )       â  r6:g     A  â #rT8      â Er+ +       â  r8:i     A  â #r6)      â &r)   ÿ   Ä #r8+      â &r
   ÿ   ä &rG{G   ÿ   ä&rX{X   ÿ   ä&rZ{Z   ÿ   ä&r+{   ÿ   ä&r
{
   ÿ   Ä &r)t   )   ä &rG>F   G   ä&r[>[   X   ä&rZ>Y   Z   ä&r+>t   +   ä &rs      Ä &r
>s   
   ä &r)r   )   ä &rGzR   G   ä&r[zf   [   ä&rZze   Z   ä&r+zr   +   Ä &rq      ä&r
zq   
   ä &r)Vp   )   ä &rG|P   G   ä&r[|d   [   ä&rZ|c   Z   Ä &r+|p   +   ä &rVo      ä&r
|o   
   ä &r)Dn   )   ä &rG~N   G   ä&r[~b   [   Ä &rZ~a   Z   ä&r+~n   +   ä &rDDm      ä&r
~m   
   â x   ÿàõ â &rG}L   G   âx     ÿàÿ â &r[}`   [   Ä &rZ}_   Z   ä &r)	l   )   ä&r+}l   +   ä &rD	k   D   ä&r
}k   
   â rÿ   ÿä â &rJ€J   G   â |     pbô â &r[€^   [   Ä &rZ€]   Z   ä &r)xj   )   ä&r+€j   +   ä &rDx   D   ä&r
€   
   ä 0rÿ    A   â x   ÿà÷ â &rJ/U   J   Ä &r[/\   [   ä&rZ/h   Z   ä &r)yv   )   ä&r+/v   +   ä &rDyw   D   ä&r
/w   
   Ä 0ruÿu    A   â  r:     A  ârÿ   ÿäÿ â  rW=W     A  âErJ J       â  rg=g     A  âEr[ [       â  ri=i     A  âErZ Z       â  r=     A  âEr) )       â  r::u     A  â Er+ +       â  r==u     A  â ErD D       Ä Er
 
       ä x$$   ÿà÷ ä x""   ÿàù â #rWJ      â #rg[      â #riZ      â #r)      â #r+      â #r:D      â #r =
       â x&&@   ÿàÿ Ä x**@   ÿàÿ ä x44@   ÿàÿ ä x,,@   ÿàÿ ä x((@   ÿàÿ ä x22@   ÿàÿ ä x..@   ÿàÿ Ä x00@   ÿàÿ ä r%ÿ%   ÿäÿ ä r#ÿ#   ÿä â D)       À â Gy  @åÿÿÿÿƒ ê Ay        € ê Ey  0    € â G‰      € ê r   ÿHŽ Ê ˆs  øÿ H   è ˆs €øÿ H   è ˆs  ùÿ H   è ˆs €ùÿ H   è ˆs  úÿ H   è ˆs €úÿ H   è ˆs  ûÿ H   è ˆs €ûÿ H   è ˆs  üÿ H   è ˆs €üÿ H   è ˆs  ýÿ H   è ˆs €ýÿ H   è ˆs  þÿ H   è ˆs €þÿ H   è ˆs  ÿÿ H   è ˆs  €ÿÿ H   ä Ay        € ê {          ì M        € ê „y     H   b z  r  p`ð Ä z8  r      â „y €   H   ¢ r ÿ   p"p ä r488   ÿŽ â„y	    H   â r688   ÿŽ Æ y
      &   ( y      '   h „y    H   ( „y    H   ( „y €  H   ( „y €  H   è „y"    H   â $z
 x  Ž âx   ÿàÿ â !r       â/ „y(    H   â $z ~  Ž â!r       ÄO ‰      " !r	       Æ� „y$    H   h „y% €  H   ¨ „y' €  H   â !r       Æ „y) €  H   â 4   ÿàù â !r        ä „y&    H   â ÿÿ   ÿä â !r       â !r       â „y+ €  H   â 
 j  ÿˆ â !r       Ä ‰      " 
 k   â !r"       ä ‰      " !r-(       Æ ‰      h ‰      b !r$       Æ/ ‰!      b !r%%       ÆO ‰      ¢ !r''       â� !r       ä ‰#       " r8   ÿ† â!r))       â !r
       â‰	 €    â 
 r  ÿàó ä r*ÿÿ   ÿäÿ â ‰
      " ,ÿÿ   ÿäÿ  â !r       â
 j  ÿ„ â ‰$      b z j  ÿ† â !r       â 
 k  , â „y5 €  H   â z  r  p`ò â !r&       â !r++       â ‰"%      b z k  *� â !r!!       â/ !r       âO ‰('      b !r# #       â 6   ÿà÷ â !r	       â� ‰      " $x8   ÿŽ â ÿÿ   ÿäÿ â !r
       â‰ @    ¢ r8   ÿ„ Ä r	   ÿàù â ‰*      â r&ÿÿ   ÿä ä z
 j  ÿ„ â‰,)      " rÿÿ   ÿä ä z k  & â ‰      " !r&$       âz	 j  ÿˆ â !r%%"       â ‰.      b !r('(       â/ 
 j  ÿ† Ä ‰0+      b !r       â z		 k   â !r55        â !r       âO ‰2-      " 
 k  � â !r**       ä� ‰ €    ¢ x    p"ò â !r,),       Ä‰!      â !r       Æ ‰      " !r..       Æ‰ #      d 8   ÿàÿ â !r0+0       â/ ‰      d    ÿàõ â !r2-2       â ‰$%      $ :ÿÿ   ÿä â !r       âO ‰'&      ¢ 4   ÿàÿ â !r!!       Ä� ‰)(      â r88  €ÿŽ â !r       ä‰ €    "    ÿàù â !r##        ä‰+*      b 4ÿÿ   ÿä â !r9       Ä/ ‰-,      b 
 j  ÿˆ ä    ÿà÷ â ‰ €    b !r$%$       â 
 k  4 ä ‰/.      " !r&&'       âO 8ÿÿ   ÿäÿ Ä ‰10      ¢ !r(()       â�  j  ÿ† ä‰32      â !r       â k  8� ä ‰5      " !r**+       âx8        Ä ‰ @    b !r,,-       â/  j  ÿ„ ä‰ ! €    b !r       â  k  : ä ‰ €    b !r../       Æ ‰"# €    " !r001       âO !r223       ä� ‰'$ €    ¢ !r45       Æ‰)& €    â !r        Æ‰+( €    " %v j  8 Ž â 6   ÿàÿ â !r !        â/ ‰ @    d    ÿàõ â !r       â ‰-* €    d 6ÿÿ   ÿä â !r"#"       â ‰/, €    "  j  ÿ„ Æ ‰ @    " !r'$'       âO  k  6 ä ‰1. €    ¢ !r)&)       Æ� ‰30 €    â !r+(+       Æ‰52 €    " !r       Æ/ ‰74      b !r-*-       Æ‰       b !r/,/       Æ ‰#  @    " !r       Æ ‰! @    " !r1.1       ÆO ‰%" @    ¢ !r303       Æ� ‰$' @    â !r525       â!r747       ä/ ‰&) @    b !r;        Æ‰(+ @    " !r# #       Æ ‰       â !r!!       Æ ‰*- @    " !r%"%       ÆO ‰ / @    ¢ !r$'$       Æ� ‰      è ‰"1 @    b !r&)&       Æ/ ‰,3 @    b !r(+(       Æ‰07 €    ( ‰.5 @    " !r*-*       Æ ‰!      " !r /        ÆO ‰#      ¢ !r)       Æ� †	 9    â !r"1"       â!r,3,       ä/ † ;   â !r070       Æ‰$      b !r9        â� !r.5.       ä ‰ %      â !r+!       Æ ‰&      " !r-#       ÆO ‰(      ¨ †	 9    è ‰*      " !r$       Æ/ ‰       h † )    â !r%%        Æ� ‰!"      â !r&       Æ ‰#,      " !r(       ÆO ‰'0 @    ¨ ‰.      b !r*       Æ†	 +    â	 !r        Æ/ † -   è	 †	 %    â	 !r!"!       â� !r#,#       ä †     â	 !r'0'       ÆO †	 
    â	 !r.       Æ† 
   è	 †	     è	 †     è	 †	 !    è	 † #   è	 ‰ '      ( †	     â	 M™        € ê x   ÿàÿ â !r''        Ær   ÿàñ È rÿÿ   ÿä  ä z j  ÿ€ È z k    Ê †y '    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      "   " ¹z  l      â Ey  °    € â ™x?    â y      !   b ¹z  F   
   â x      ÿ  â r ÿ       â y       %   ¢ r* ÿ       â x      ÿ  â r  ÿ       â x      ÿ  â x      ÿ  â x      ÿ  â r ÿ       Ä r ÿ   pRð är ÿ   pRò är   ÿ(Ž È/ xÿ     ä x     ÿ   äO r   ÿ� È xÿ     È |    pbô Ú G)  @    € ê y      '   " r ÿ       â $z  p  Ž â xüÿÿÿÿÀŽ â y	      &   b r ÿ       ä r  €ÿàÿ ä x(        ä x3   ÿ   Ä z- q  ÿàÿ ä x%ÿ     ä r ÿ       ä x--   ÿ   ä z13 q  ÿàÿ â 'z y   Ž È 'z	 s   Ž â/ zÿ z    ä x   ÿ   â $z"	 w  ÿŽ â zÿ t    â $z |  Ž ä$z |  Ž ä %x   ÿŽ È $z' v  Ž âx    ÿàÿ â $z v  Ž ä %x'   Ž È $z p  Ž â z X  ÿàû â $z }  ÿŽ ä %x   Ž â x   ÿàý ä r""   àµ â %v&' X  (Ž âz X  ÿàù ä zÿ Y  ¤~ â %v( X  (Ž â z  q      Ä x   ÿàû ä r3   ÿŽ â$x+   3Ž âzÿ Y  „þ â $x#   3Ž âr%ÿÿ   %d~ â $x/   3Ž ä �ykþÿÿ • ¨ �yl    • è �y6þÿÿ • " r3"   ÿàõ Æ �y9    • b x] $       ä rÿ%   ÿä â �y,&    • ä %v Z  ] Ž â r
"   ÿà÷ â �y!(    • ä $x	$   ÿŽ Ê r	   ÿàÿ Ê �y0   ™ ( �y5   ™ ( �y7   ™ ( �y2   ™ ( �y:   ™ ( �y4   ™ " rÿ%   ÿäÿ Æ �y   ™ " %v

 Z  ] Ž â r1"   ÿàõ ä �y.    • " $x;$   ÿŽ Æ �y8    ™ " r	ÿ%   ÿä ä r;   ÿàÿ â %v Z  ] Ž È $x		$   ÿŽ â �y;
   ™ ( r	   ÿàÿ â �y
   ™ ( �yG
   ™ ( �y=   ™ h �yE
   ™ h �y<   ™ h �yC
   ™ b r+"   ÿàõ Æ �y$   ™ h �yA
   ™ b r-"   ÿà÷ Æ �yF   ™ b rÿ%   ÿä Æ �y?
   ™ b rKÿ%   ÿäÿ Æ �yD   ™ b %v Z  ] Ž Æ �yI
    ™ b $xM$   ÿŽ Æ �yJ
    • h �yB   ™ b %v Z  ] Ž â r		M   ÿàÿ ä �y@   ™ b $xKK$   ÿŽ Æ �y>    ™ h �yH    • h �yM   ™ h �yO   ™ b rK   ÿàÿ Æ/ �yK   ™ b r        Æ �yQ   ™ h �yL   ™ h �yN   ™ b r
#"   ÿàõ Æ �yS   ™ h �yP   ™ b r/"   ÿà÷ Æ �yU   ™ b rÿ%   ÿä Æ �yR   ™ b r^ÿ%   ÿäÿ Æ �yW   ™ b $x_$   ÿŽ Æ �yT   ™ b %v

 Z  ] Ž Æ �yX    ™ b %v Z  ] Ž Æ �yV   ™ b $x]^$   ÿŽ Æ �yZ    • b r_   ÿàÿ Æ �yY   ™ b r]   ÿàÿ Æ �yg
   ™ h �y
   ™ h# �y	   ™ h �yf   ™ h �yd   ™ h �ye
   ™ h �yb   ™ h �yc
   ™ h �y`   ™ h �ya
   ™ h �y^   ™ h �y_
   ™ h �y[    ™ h �y\    • h �y]
   ™ h �yj    ™ h �y   ™ h �y
    • h �yi
    ™ h �yh    • b ‚x       Ä |n       â ‚x ÿ       â xkk™  ÿ    ÈO xmkww  ÿÀŽ ä xkÿ   k  ä tmn m    ä |
       ä/ xkkww  ÿÀŽ ä |       Ä xmww  ÿÀŽ ä tk
 k    ä tÿ      ä xÿ   k  ä |o       ä xmÿ   m  Ä xww  ÿÀŽ ä xl™  ÿ    ä� |n       ä xmmww  ÿÀŽ ä tloÿ      ä xÿ     Ä |
       ä xkkww  ÿÀŽ ä tmnÿ  m    ä |p       ä |o       ä xww  ÿÀŽ Ä xnww  ÿÀŽ ä t

ÿ  k    ä tkp     ä tno n    ä |q       ä xokww  ÿÀŽ Ä xpnww  ÿÀŽ ä xT  
    äxkÿ   k  ä x2v  
    ä x
mT  l    äxm2v  l    Ä |r       ä tlqÿ  p    ä |p       ä x66™  ÿ    äxkkww  ÿÀŽ ä tmrÿ  o    Ä |r       ä xo6ww  ÿÀŽ ä tqpÿ  k    ä xkÿ   6  ä t6r o    ä xokww  ÿÀŽ Ä xkmT  l    ä|s       ä xp6ww  ÿÀŽ ä xmm2v  l    ä tlr o    ä xnÿ   n  Ä tosÿ  p    ä xpÿ   l  ä xnnww  ÿÀŽ ä xllww  ÿÀŽ ä xr9™  ÿ    ätnsÿ  n    Ä x6ÿ   6  ä tlsÿ  l    ä |t       ä xsrww  ÿÀŽ ä |w       ä xppww  ÿÀŽ Ä |u       ä x66ww  ÿÀŽ ä xrÿ   r  ä tst s    ä tpwÿ  p    ä t9uÿ  6    Ä |w       ä xrrww  ÿÀŽ ä x6oT  l    ä |v       ä xusww  ÿÀŽ ä ttw r    Ä xoo2v  l    ä trvÿ  u    â &rv60   ÿ   âxl9T  p    äx992v  p    â &rp0   ÿ   â xuÿ   s  â &rvo5   v   â |x       Ä xstww  ÿÀŽ â &r05   p   â |z       ä xuuww  ÿÀŽ â &rvl7   v   â tsxÿ  s    ä xtÿ   t  â &r0
7   0   â tpzÿ  u    â &rv92   v   â |u       Ä xttww  ÿÀŽ ä x5rT  s    â&r02   0   â tuuÿ  t    ä x7r2v  s    â &rv5:   v   ä&rrk:   0   â x0pT  u    â&rv74   v   â xsp2v  u    Æ &rv0   v   âx2qT  n    â &rrm4   r   ä 0r,ÿ,    A   ä 0r.ÿ.    A   ä &rvs8   v   ä 0r!ÿ!    A   â xqq2v  n    â &rr2   r   â  r,.     A  â Erv v       â  r..!     A  â &rrq8   r   Æ #r.v      â &r.;   ÿ   ä&r;6;   ÿ   â Err r       Æ &r;o   ;   ä&r.   .   â #rr      â &rlG   ;   ä &r=   ÿ   ä &r9E      ä &r<      Ä &r.
G   .   ä&r5C      ä &r
$      ä &r.E   .   ä &r7A      ä &rF      Ä &r.kC   .   ä&r0?      ä &rkD      ä &r.mA   .   ä 0rJÿJ    A   ä &rsI      Ä &rmB      ä &r.2?   .   â  r?,J     A  â &r2@      â Er        â  rJ!J     A  â 0rHÿH    A   ä &rq>      â #rJ      â &rK   ÿ   â  r;,H     A  Ä Er        â &rM      È #r;      â &r
O      ä &rL   ÿ   ä &rQ      ä &rN      ä &rkS      ä &r
P      Ä &rmU      ä &rR      ä &r.qI   .   ä &r2W      ä &rkT      â Er. .       â &rqX      Ä &rmV      ä 0rZÿZ    A   â #r?.      â &r=6=   ÿ   â Er        â &r.2Y      â  r,Z     A  â &r=o<   =   Æ #r      â &r   ÿ   ä &r=l$   =   ä &r	   ÿ   ä &r$6K   ÿ   ä&rL6L   ÿ   ä&r6   ÿ   Ä &r	6	   ÿ   ä &rg      ä&rf      ä &r$oM   $   ä&r;oN   L   ä&rog      Ä &r	of   	   ä &r=9F   =   ä &r
d      ä &r$lO   $   ä&r;lP   ;   ä &r
e      Ä &rle      ä&r	ld   	   ä &r=5D   =   ä &rb      ä &r$9Q   $   ä&r;9R   ;   Ä &rc      ä&r9c      ä&r	9b   	   ä &r=7B   =   ä &rk`      ä &r$5S   $   Ä &r;5T   ;   ä &rka      ä&r5a      ä&r	5`   	   â x   ÿàõ â &r=0@   =   â x   ÿàÿ â &rm^      Ä &r$7U   $   ä&r;7V   ;   ä &rm_      ä&r7_      ä&r	7^   	   â rÿ   ÿä â &r=s>   =   â |    pbô â &r2      Ä &r$0W   $   ä&r40Y   ;   ä &r2]      ä&r0]      ä&r	0   	   â Er= =       â &rqj      â  rH!H     A  â &r$sX   $   â x   ÿà÷ â 0r\ÿ\    A   Ä &r.q[   .   ä&r4s[   4   ä 0rÿ    A   ä &rqi      ä&rsi      ä&rjsj   	   Ä 0rhÿh    A   â #rH=      â rÿ   ÿäÿ â  rZ!Z     A  â r;,\     A  âEr$ $       â  r=!\     A  âEr. .       â  r,     A  âEr4 4       â  r	!     A  âEr        â  r,,h     A  â Er        â  r!!h     A  â Er        Ä Erj j       ä x((   ÿà÷ ä x&&   ÿàù â #rZ$      â #r;.      â #r =4       â #r*   *   â #r	      â #r,      â #r!j      â x@   ÿàÿ Ä x++@   ÿàÿ ä x--@   ÿàÿ ä x##@   ÿàÿ ä x//@   ÿàÿ ä x11@   ÿàÿ ä x33@   ÿàÿ Ä r)ÿ)   ÿäÿ ä r'ÿ'   ÿä â D)       À â Gy  �çÿÿÿÿƒ ê Ay        € ê Ey      € â G‰  ð     € ê $xÀ  Ž Ê ˆs  ùÿ H   è ˆs €ùÿ H   è ˆs  úÿ H   è ˆs €úÿ H   è ˆs  ûÿ H   è ˆs €ûÿ H   è ˆs  üÿ H   è ˆs €üÿ H   è ˆs  ýÿ H   è ˆs  €ýÿ H   è ˆs * þÿ H   è ˆs €þÿ H   è ˆs  ÿÿ H   è ˆs €ÿÿ H   ä Ay        € ê {          ì M        € ê „y     H   b z   r  p`ð Ä z+  r      â „y €  H   ¢ r ÿ   p"p ä r/++   ÿŽ â„y €  H   â r)++   ÿŽ Æ „y €   H   ( „y    H   h „y    H   ( „y	    H   ( „y    H   è „y €  H   è „y    H   â !r       Æ/ „y! €  H   â !r       ÆO ‰      " !r       Æ� „y €  H   â !r       Æ„y    H   â !r       Æ„y# €  H   â !r       Æ ‰
      b !r		       Æ ‰      ¢ !r       Æ y      &   " !r        Æ ‰      â !r**       Æ ‰      " !r"!       Æ y      '   b !r

       Æ/ ‰      b !r       ÆO ‰      ¢ !r       â !r       â !r$       â ‰	      ¢ $z x   Ž â !r&#       â !r       â� ‰      " x    ÿàÿ â !r       Ä‰      â z  r  p`ò â $z ~  Ž ä‰       " x    p"ò â !r       ä/ ‰!*      b 
  r  ÿàõ â !r       ÄO ‰ €    ¢ ÿÿ   ÿä â !r	       ä ‰      b 
  j  ÿ„ Æ ‰#"      b !r       â 
  k   ä ‰%$      " !r       Æ� ‰	      â !r        Æ‰      " !r!*!       Æ/ ‰'&      b !r        âO r+   ÿˆ ä ‰
      ¢ !r       â)   ÿàû ä ‰      b !r#"#       â ÿÿ   ÿäÿ Ä ‰      b !r%$%       â 
 j  ÿŠ ä ‰       " !r	       â� r	+   ÿ„ â !r       â‰"      â 
 k  � â !r'&'       Ä/ ‰$      b r.ÿÿ   ÿä â !r
       äO ‰  @    ¢ /   ÿà÷ â !r       ä‰&!      " ÿÿ   ÿäÿ â !r       Ä ‰ €    b 

 j  ÿ† ä rÿÿ   ÿä â ‰(#      b !r         â z	 j  ÿ„ ä ‰*%      " !r""       â� 
 k  � Ä ‰ €    â !r$$       â/ z		 k   ä ‰,'      b !r        âO z j  ÿˆ ä ‰ €    ¢ !r&!&       âz k  . Ä ‰ €    " !r       â+   ÿàÿ ä ‰ €    b !r(#(       â    ÿàõ ä ‰ €    b !r*%*       â 2ÿÿ   ÿä Ä ‰!  €    " !r       â� )   ÿàÿ â !r,',       â/ ‰#" €    d 0   ÿà÷ â !r       âO ‰%$ €    ¢ $x+   ÿŽ â 1ÿÿ   ÿäÿ â !r       â‰       â 0 j  ÿ† â !r       Ä‰'& €    " r   ÿàù â !r       ä ‰ @    b r.ÿÿ   ÿä ä z j  ÿˆ â ‰)( €    b !r! !       â  j  ÿ„ Ä ‰+* €    " z k  . â !r#"#       â/ 0 k  1� â ‰ @    b !r%$%       âO  k  2 ä ‰ @    ¢ !r1        â� /   ÿàÿ Ä ‰-, €    â !r'&'       Æ‰ @    " !r        â   ÿàõ ä ‰ @    b !r)()       â x        ä ‰ @    b !r+*+       â .ÿÿ   ÿä â %v j   Ž â ‰ ! @    " !r       â/  j  ÿ„ Ä ‰"# @    b !r       âO  k  . ä ‰$% @    ¢ !r-,-       Æ� ‰&' @    â !r       Æ‰       " !r       Æ‰() @    b !r       Æ ‰*+ @    b !r !        Æ ‰      " !r"#"       Æ/ ‰,- @    b !r$%$       ÆO ‰      ¢ !r&'&       Æ� ‰      â !r/        Æ‰      " !r()(       Æ‰      b !r*+*       Æ ‰!       b !r3       Æ ‰#"      " !r,-,       Æ/ ‰%$      b !r)       ÆO ‰&      ¢ !r       Æ� †	 1    â !r       Æ† /   â !r       Æ‰'(      â !r! !       Æ ‰*      " !r#"#       Æ ‰,      " !r%$%       Æ/ †	 3    â !r&       ÆO † )    è †	     è †    â !r'('       Æ� †	     â !r*       Æ† !    â !r,       Æ †	 #    è † %   è †	 
    è † '    è †	     â M™        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      "   " ¹z  l      â Ey       € â ™x?    â y      !   b ¹z  F   
   â x      ÿ  â x      ÿ  â x      ÿ  â x      ÿ  â r  ÿ       â x      ÿ  â r ÿ       Ä r ÿ   pRð är ÿ   pRò är   ÿ(Ž È/ xÿ     È r   ÿ� È xÿ     È |    pbô Ú G)  p    € ê y      '   " xüÿÿÿÿÀŽ ä r ÿ       â y      %   b r
  €ÿàÿ ä r ÿ       â y	      &   ¢ x+   ÿ   äz# q  ÿàÿ Ä x"ÿ   
  ä r ÿ       ä x##   ÿ   ä z'+ q  ÿàÿ â 'z y   Ž â x   ÿ   Æ/ $z }  ÿŽ â zÿ z    â 'z	 s   Ž âO x
   ÿ   Æ $z p  Ž â zÿ t    â %x   ÿŽ â x   ÿàÿ ä x        â $z |  Ž ä$z |  Ž ä $z v  Ž Ä $z v  Ž ä %x   Ž È $z p  Ž â z X  ÿàû â $z	 w  ÿŽ ä %x   Ž â x   ÿàý ä r   
àµ â %v X  Ž âz X  ÿàù ä zÿ Y  ¤~ â %v X  Ž â r         Ä x   ÿàû ä r        ä z  q      ä r        ä zÿ Y  „þ â $x%   +Ž âr!+   ÿŽ â$x)   +Ž â r"ÿÿ   "d~ Ä �y_þÿÿ • ¨ �yM    • è �y=þÿÿ • " r+   ÿàõ Æ �y;    • b xU $       ä r	ÿ"   ÿä ä r'   ÿàõ â %v Z  U Ž Æ r
ÿ"   ÿä â $x	$   ÿŽ ä %v Z  U Ž Æ r   ÿàÿ â $x
$   ÿŽ È �y:   ™ " r		   ÿàÿ Æ �y&   ™ ( �y(   ™ ( �y,   ™ ( �y.   ™ ( �y*   ™ ( �y0   ™ ( �y-   ™ ( �y2   ™ " r!   ÿàõ Æ �y/   ™ " r
         Æ �y5   ™ " r        ä r?ÿ"   ÿä â �y1   ™ " r@%   ÿà÷ Æ �y7   ™ b r>ÿ"   ÿäÿ â %v Z  U Ž ä �y3
    • d $x??$   ÿŽ ä �y9    ™ h �y<    • b r
        Ä r        â �y4   ™ h �y6    ™ b r#   ÿàõ Æ/ �y8    • h �y$
    • b r)   ÿà÷ ä rÿ"   ÿä ä r?   ÿàÿ ä/ r
        Ä rQÿ"   ÿäÿ â %v@@ Z  U Ž ä�yL
   ™ d $xC>$   ÿŽ ä %v Z  U Ž â�y>
   ™ f $xS$   ÿŽ â �yJ
   ™ b %v Z  U Ž â rAC   ÿàÿ Ä �yH
   ™ b $xQQ$   ÿŽ â r @       ä �yA
   ™ b r		S   ÿàÿ ä rQ   ÿàÿ â �yF
   ™ h �yD
   ™ h �yB
    ™ h �yN
    • h �y@   ™ h �y?   ™ h �y
   ™ h �y   ™ h �y[   ™ h �yZ   ™ h �yK   ™ h �yY   ™ h �yX   ™ h �yI   ™ h �yW   ™ h �yV   ™ h �yG   ™ h �yU   ™ h �yT   ™ h �yE   ™ h �yS   ™ h �yR   ™ h �yC   ™ h �yO    ™ h �yP    • h �yQ   ™ h �y]    ™ h �y   ™ h# �y    • h �y^    ™ h �y\    • b ‚x       Ä |a       ä |b       â ‚x ÿ       ä |       ä |	       ä x__™  ÿ    ÈO x`_ww  ÿÀŽ ä x_ÿ   _  È x__ww  ÿÀŽ ä t`a `    ä t_b _    ä x`ww  ÿÀŽ ä/ xÿ   _  ä x`ÿ   `  Ä x__ww  ÿÀŽ ä xMM™  ÿ    ä� x``ww  ÿÀŽ ä t_ÿ  _    ä r M       ä |b       Ä xww  ÿÀŽ ä t	ÿ      ät`	ÿ  `    ä x	ww  ÿÀŽ ä xÿ     ä tMbÿ      Ä |b       ä xT  _    ä|d       ä xaww  ÿÀŽ ä x2v  _    ä t_b 	    Ä tad a    ä xb_ww  ÿÀŽ ä x	`T  M    äx`2v  M    ä x_ÿ   _  ä xMÿ   a  Ä |`       ä xaaww  ÿÀŽ ä |d       ä x==™  ÿ    äx__ww  ÿÀŽ ä |c       Ä |f       ä xMMww  ÿÀŽ ä ta`ÿ  a    ä t_dÿ  _    ä x`ÿ   =  ä tbcÿ  b    Ä tdfÿ  M    ä |f       ä xM=ww  ÿÀŽ ä xc`ww  ÿÀŽ ä |h       ä x=bT  a    Ä x`b2v  a    ä taf M    ä tch c    ä xb_T  d    äxM_2v  d    ä xeaww  ÿÀŽ Ä x_;™  ÿ    äxaÿ   a  ä |g       ä xfcww  ÿÀŽ ä xcÿ   c  ä xdÿ   _  Ä |h       ä xaaww  ÿÀŽ ä |l       ä x__ww  ÿÀŽ ä tfgÿ  f    ä |n       Ä x;cww  ÿÀŽ ä xgdww  ÿÀŽ ä tchÿ  a    ä tal _    ä tgn g    ä tehÿ  e    Ä |j       ä xhaww  ÿÀŽ ä xaÿ   a  ä |k       ä xigww  ÿÀŽ ä tdjÿ  ;    Ä x;eT  f    äx_e2v  f    ä xgÿ   g  ä |f       ä xeaww  ÿÀŽ ä thkÿ  h    Ä tijÿ  i    ä xacT  d    äxggww  ÿÀŽ ä xcc2v  d    ä tefÿ  e    ä xdhT  i    Ä xfh2v  i    â &ri(   ÿ   âtjjÿ  g    â &rg:   ÿ   ä&rh;:   ÿ   ä &ri,   i   ä&rg&   g   ä&rh_&   h   Ä &ri	.   i   ä &r(;(   ÿ   ä &rg	*   g   ä&rha*   h   ä &ri0   i   ä &r(_,   (   Ä &r&-   g   ä&rhc-   h   ä &ri=2   i   ä &r(a.   (   ä &r&=/   &   ä&rhd/   h   Ä &ri`5   i   ä&r/c0   (   ä &r-`1   &   ä &r&b7   i   ä &r/d2   /   â x:eT  j    â &r&M9   &   Ä &r/f5   /   ä 0r3ÿ3    A   ä 0r<ÿ<    A   â xee2v  j    â &r-b4   -   â Er& &       â &r(:7   /   â  r53<     A  â &r-M6   -   Ä 0r8ÿ8    A   â #r5&      â &r(e9   (   ä 0r$ÿ$    A   ä &r&>   ÿ   â Er- -       â  r/38     A  â Er( (       â  r7$<     A  â &r&A   &   â #r/-      Ä #r 7(       â &r-	L   &   ä &r(@   ÿ   ä&r&
   ÿ   ä&r   ÿ   ä &r>;>   ÿ   ä&r@;@   ÿ   Ä &r
;
   ÿ   ä&r;   ÿ   ä &r(?   (   ä&r&[   &   ä &r>_A   >   ä&r@_?   @   Ä &r
_[   
   ä&rZ      ä&r_Z      ä &r(	K   (   ä&r&	Y   &   ä &r>aL   >   Ä &r@aK   @   ä&r
aY   
   ä &r	X      ä&raX      ä &r-J   -   ä&r(I   (   Ä &r&W   &   ä&r/cJ   >   ä&r@cI   @   ä&r
cW   
   ä&rV      ä&rcV      Ä &r1f1   h   ä &r-=H   -   ä&r(=G   (   ä&r&=U   &   ä &r/dH   /   ä&r@dG   @   Ä &r
dU   
   ä&r=T      ä&rdT      â x   ÿàõ â &r1:4   1   â x   ÿàÿ â &r-`F   -   ä&r(`E   (   Ä &r&`S   &   ä &r/fF   /   ä&r@fE   @   ä&r
fS   
   ä &r`R      ä&rfR      â rÿ   ÿä â &r1e6   1   â |    pbô â &r-bD   -   Ä &r(bC   (   ä&r&bQ   &   ä &r/:D   /   ä&r@:C   @   ä&r
:Q   
   ä &rb      Ä &r:      â Er1 1       â  r88$     A  â 0rNÿN    A   ä &r-MB   -   ä0rPÿP    A   ä &r(MO   (   ä 0rÿ    A   Ä &r&M]   &   â x   ÿà÷ â &rBeB   /   ä&r@eO   @   ä&r
e]   
   ä &rMM^      ä&re^      Ä 0r\ÿ\    A   â #r81      â  r/3N     A  â r13P     A  â r3     A  ârÿ   ÿäÿ â  rN$N     A  âEr- -       â  rP$P     A  âErB B       â  r$     A  âEr( (       â  r33\     A  â Er@ @       â  r$$\     A  â Er& &       Ä Er
 
       ä ErM M       ä Er        ä x   ÿà÷ ä x     ÿàù â #r/-      â #rNB      â #r1(      â #rP@      â #r&      â #r
      â #r3M      â #r$      â x!!@   ÿàÿ Ä x%%@   ÿàÿ ä x##@   ÿàÿ ä x))@   ÿàÿ ä x''@   ÿàÿ ä x++@   ÿàÿ ä rÿ   ÿäÿ Ä rÿ   ÿä â D)       À â Gy   éÿÿÿÿƒ ê Ay        € ê Ey  ð     € â G‰  Ð     € ê $x€  Ž Ê ˆs  úÿ H   è ˆs €úÿ H   è ˆs  ûÿ H   è ˆs  €ûÿ H   è ˆs  üÿ H   è ˆs €üÿ H   è ˆs  ýÿ H   è ˆs €ýÿ H   è ˆs  þÿ H   è ˆs €þÿ H   è ˆs  ÿÿ H   è ˆs €ÿÿ H   ä Ay        € ê {          ì M        € ê „y     H   b Ey  @    € æ „y €   H   ¨ „y    H   è „y €  H   â !r       â/ !r       È_ ‰      " !r       Æ� ‰      h ‰      ¢ !r       Æ y      %   " !r       Æ/ ‰      b !r	       ÆO ‰      ¨ ‰	      â !r       â/ x   ÿ   â !r       ÆO ‰ €    " z  r  p`ð â !r
	       ä� ‰ €    b r ÿ   pVp  Æ ‰
 €    ´ ‰      '   â !r       â x   ÿàÿ â !r       Æ/ ‰ @    " z  r  p`ò â !r
       äO ‰	 @    b x    p"ò ä ˆ
        â ‰ @    ¨ ‰      &   ,       &   h       '   b !r       â !r		       È/ ‰      " !r       ÆO ‰	      b $Š x  Ž Æ‰      ¢ $Š ~  Ž â�         â $ x  Ž È$ ~  Ž â !r       â %† j  
 Ž â !r	       Æ/ % j   Ž â †‰     â !r       ÆO †    â G	  �     € ê y      &   ( y      '   b $z x  Ž È $z ~  Ž Ê/ z r  ÿàõ È rÿÿ   ÿä ä z j  ÿ„ È z k   Ê †y     è Ay        € ê !r        â „y    H   â Ey  Ð    € æ ‰       ( „y €  H   è „y	    H   è „y €  H   è „y    H   è „y
 €  H   â !r         Æ „y    H   è ‰       $ !r        Ê ‰ €    $ !r       Ê ‰ @    $ !r       Ê ‰       $ !r        â G™  °     € ê y      &   " ¹z  r      ä �x   ?àÿ â y       '   b $z x  Ž È $z   ~  Ž Ê/ |     ÿàõ È rÿÿ   ÿä ä z  j  ÿ„ È z  k   Ê †y     è Ay        € ê „y €  H   â !r       â !r       â !r		       â !r       â !r       â !r
       â !r       â ‰       b z  r      Æ ‰      ¤ $x   ÿŽ ä ‰	      è ‰
      ( ‰      h ‰      ( ‰      " !r         â/ !r       Ä_ y      &   " !r	       â� !r       ä ‰       b !r

       Æ‰      ¢ !r       Æ‰      â !r       Æ ‰	
      " !r       Æ ‰      b $z x  Ž Æ ‰      ( ‰      " !r        Æ/ ‰      b !r       âO !r       ä� ‰  €    ¢ !r	
	       Æ‰ €    â !r       Æ‰ €    " !r       Æ ‰
	 €    " !r       Æ ‰ €    b !r       Æ/ ‰ €    h ‰ €    b !r         ÆO ‰      ¢ !r       Æ� y      '   â !r       â!r
	
       ä ‰  @    " !r       Æ‰	 @    â !r       Æ/ ‰ @    â !r       Æ ‰
 @    b !r       ÆO ‰ @    ¢ $z ~  Ž Æ� ‰ @    â $x   ÿŽ Æ ‰ @    " r   ÿ„ â !r         â ‚   ÿà÷ â ‰ €    " rÿÿ   ÿä ä ‚ÿÿ   ÿäÿ ä Š j  ÿ† ä z j  ÿ„ â !r
       â/ Š k  � â !r	       â z k   â !r       â !r       âO ‰	       b r   ÿ„ â !r       â�    ÿàÿ â ‰      ¢ rÿÿ   ÿä â !r       â   ÿàõ â‰      â ‚   ÿàù â !r       â  ÿÿ   ÿä â ‰      "  j  ÿ„ Ä z j  ÿ† â ‰      " ‚ÿÿ   ÿä ä Š
 j  ÿˆ â ‰      b z k  � ä Š k   â ‰ @    b !r 	       â/ 	 k    Ä ‰      b !r       âO !r       ä� †‰     â !r       Æ †    â !r       Æ†‰     â !r       Æ†     â !r       Æ †‰     â !r       Æ/ †    è ‰       ( †‰ 
    â M™        € ê x   ÿàÿ âO !r        Æ r   ÿàñ È rÿÿ   ÿä  ä z j  ÿ€ È z k    Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y       "   " ¹z  l      â Ey  à    € â ™x?    â y      !   b ¹z  F   
   â x      ÿ  â x      ÿ  â r ÿ       â x      ÿ  â r ÿ       Ä r
 ÿ       ä r ÿ       ä r  ÿ   pRð är  ÿ   pRò är    ÿ(Ž È/ xÿ     È r   ÿ� È xÿ     È |    pbô Ú G)  °    € ê y      '   " r ÿ       ä xüÿÿÿÿÀŽ â y      %   b r ÿ       ä r  €ÿàÿ â y      &   ¦ x   ÿ   Ê %x   ÿŽ È 'z y   Ž È $x   ÿ Ž â/ zÿ z    â 'z s   Ž ÆO x   ÿàÿ äx        â $z p  Ž â zÿ t    Æ $z |  Ž âxÿ     â $z |  Ž ä $z v  Ž ä$z v  Ž ä %x   Ž È $z p  Ž â z^ X  ÿàû â $z }  ÿŽ â r ÿ       â %x   Ž â x^^   ÿàý Æ $z w  ÿŽ â z. X  ÿàù â %v X  Ž âz_ÿ Y  ¤~ ä x..   ÿàû â %v X  Ž â z  q      ä z/ÿ Y  „þ â $rÿÿ    Ž â x   ÿ   Ä z q  ÿàÿ ä r   àµ â $x   Ž âr        â $x   ÿ Ž â r   ÿŽ ä z q  ÿàÿ ä rÿÿ   d~ Ä �yO^þÿÿ • ¨ �yI^    • â r'   ÿàõ Æ �yP.þÿÿ • " xJ $       ä rÿ   ÿä â �yN.    • d %v&' Z  J Ž â r9   ÿà÷ â �y    • ä $x$   ÿŽ Ê r''   ÿàÿ ä r;   ÿàõ â �y    • ( �yZ&   ™ ( �y&   ™ ( �y&   ™ ( �y&   ™ ( �y&   ™ " r$ÿ   ÿä Æ �y&   ™ " r#ÿ   ÿäÿ Æ �y &   ™ " %v:; Z  J Ž Æ �y"&    ™ " $x%$$   ÿŽ Æ �y!&    • " $x##$   ÿŽ â r;;%   ÿàÿ â %v89 Z  J Ž È $x99   #Ž â �y%:   ™ ( �y#:   ™ h �y':   ™ h �y):   ™ h �y+:   ™ b rM   ÿàõ Ä rK   ÿà÷ â �y-:   ™ b r<ÿ   ÿä är7ÿ   ÿäÿ â �y1:   ™ b %vLM Z  J Ž Æ �y2:    ™ b $x=<$   ÿŽ ä %vJK Z  J Ž â �y4:    • f $x77$   ÿŽ â �y$8   ™ b rMM=   ÿàÿ Æ �y&8   ™ b rKK7   ÿàÿ Æ �y(8   ™ h �y*8   ™ h �y,8   ™ h �y08   ™ h �y38   ™ h �y58    ™ h �y68    • h �y7L   ™ h �y:J   ™ h# �y8J   ™ h �y9L   ™ h �y;L   ™ h �y<J   ™ h �y=L   ™ h �y>J   ™ h �y?L   ™ h �y@J   ™ h �yAL   ™ h �yBJ   ™ h �yCL   ™ h �yDJ   ™ h �yEL    • h �yFL    ™ h �yHJ    ™ h �yGJ    • b ‚x       Ä |L       ä |R       â ‚x ÿ       ä |M       ä |T       ä |W       ä |S       Ä |]       ä |[       ä xOO™  ÿ    ÈO xQOww  ÿÀŽ ä xOÿ   O  È xOOww  ÿÀŽ ä tQR Q    ä tOL O    ä xKQww  ÿÀŽ ä/ xJÿ   O  ä xQÿ   Q  Ä xJJww  ÿÀŽ â $~Rÿ   ÿ Ž â xII™  ÿ    ä� |L       ä xOOww  ÿÀŽ ä xQQww  ÿÀŽ ä tJMÿ  J    Ä xMÿ   I  ä tLLÿ  O    ä tKRÿ  K    ätORÿ  Q    ä |R       ä xQIww  ÿÀŽ Ä xMMww  ÿÀŽ ä tQR Q    ä tRT M    ä xIKT  L    äxKK2v  L    ä xUQww  ÿÀŽ Ä xLRww  ÿÀŽ ä xQÿ   Q  ä xMOT  J    äxRÿ   R  ä xOO2v  J    â $~Jÿ   ÿ Ž â xQQww  ÿÀŽ Ä xRRww  ÿÀŽ ä xPP™  ÿ    ätJJÿ  Q    ä |T       ä tQWÿ  R    ä xRÿ   P  Ä tUTÿ  U    ä tLSÿ  L    ä |T       ä xSPww  ÿÀŽ ä xRRww  ÿÀŽ ä tPT S    Ä t]] R    ä xSUT  L    ä|T       ä xWPww  ÿÀŽ ä xUU2v  L    ä xYÿ   ]  Ä |L       ä x]]ww  ÿÀŽ â &rXIZ   ÿ   â tWTÿ  W    â $~Rÿ   ÿ Ž â xPÿ   P  ä tTLÿ  ]    ä xYYww  ÿÀŽ â &rXK   X   â xNN™  ÿ    ÄxPPww  ÿÀŽ ä xLWT  T    ä tRRÿ  Y    â &rXM   X   â |Y       ä xVNww  ÿÀŽ ä xaÿ   N  â &rZLZ   ÿ   â t[[ÿ  P    Ä xWW2v  T    â &rXO   X   â tVY V    ä |N       ä xaaww  ÿÀŽ â &rZW   Z   â xY[T  R    â &rXS   X   â |P       Ä x]Vww  ÿÀŽ ä tNN a    â &rZY   Z   â x[[2v  R    â &reU   X   â tPPÿ  ]    ä xJT  Q    ä |R       Ä x]Nww  ÿÀŽ â &re    e   â xQJ2v  Q    ä t]Rÿ  ]    â &rZ[   Z   ä &reQ"   e   â xaPT  ]    â 0rÿ    A   ä 0rÿ!    A   â Ere e       â 0rÿ    A   Ä &rZa   Z   â  r     A  ä  r!     A  â &rI#   ÿ   â #re      â &rL#   ÿ   â xVÿ   V  ä xNÿ   N  â &rK%      â|T       â &rW%      â xVVww  ÿÀŽ Ä |       ä xcNww  ÿÀŽ â &rM'      âtTTÿ  V    â &rY'      â tcÿ  c    ä x]P2v  ]    â &rO)      âxTT  c    â &r[)      Ä &rg]   Z   ä &rS+      âxcT2v  c    â &ra+      ä &rg    g   ä &rU-      ä&r]-      Ä &rgc"   g   ä &r1      ä&r1      â Erg g       â &rQ2      ä0r4ÿ4    A   ä &rc2      Ä &rI$   ÿ   â #r
!g   
   â Er        â  r!4     A  â Er        â &rK&      â  r44     A  â #r!      â &rI7   ÿ   ä &rM(      â #r	4   	   â &r$L$   ÿ   Ä &rL7   ÿ   ä&rII8   ÿ   ä&r8L8   ÿ   ä &rO*      ä &r$W&   $   ä&rW9      Ä &rK9      ä&rIK:   I   ä&r8W:   8   ä &rS,      ä &r$Y(   $   ä&rY;      Ä &rM;      ä&rIM<   I   ä&r8Y<   8   ä &rU0      ä &r$[*   $   ä&r[=      Ä &rO=      ä&rIO>   I   ä&r8[>   8   ä &r3      ä &r$a,   $   ä&ra?      âx^^   ÿàõ â &rS?      Ä &rIS@   I   ä&r8a@   8   â x   ÿàÿ â &rQ5      ä &r$]0   $   ä0r6ÿ6    A   ä &r]A      Ä &rUA      ä&rIUB   I   ä&r]]B   8   â Er        â $r_ÿÿ   _ â |    pbô â &r$3   $   â  r#6     A  â &rC      Ä &rC      ä &rID   I   ä&r]D   ]   â #r#      â &r$c5   $   â x..   ÿà÷ â 0rÿE    A   ä &rQF      Ä &rcF      ä &rIQH   I   ä&r]cH   ]   ä 0rÿG    A   â r/ÿ/   ÿäÿ â  r66     A  â Er$ $       â  r     A  â r!     A  â Er        â  r     A  â Er        â  r     A  â ErI I       Ä Er] ]       ä x   ÿà÷ ä x   ÿàù â #r6$      â #r      â #r!      â #rI      â #r]      â x@   ÿàÿ Ä x@   ÿàÿ ä x@   ÿàÿ ä x@   ÿàÿ ä x@   ÿàÿ ä rÿ   ÿäÿ ä rÿ   ÿä â D)       À â Gy  @ìÿÿÿÿƒ ê Ay        € ê Ey  Ð     € â G‰  °     € ê $x  @  Ž Ê ˆs   ûÿ H   è ˆs  
€ûÿ H   è ˆs   üÿ H   è ˆs  	€üÿ H   è ˆs   ýÿ H   è ˆs  €ýÿ H   è ˆs   þÿ H   è ˆs  €þÿ H   è ˆs   ÿÿ H   è ˆs  €ÿÿ H   ä Ay        € ê {          ì M        € ê „y      H   " Ey  @    € æ „y €   H   h „y    H   ¢ !r         â !r

       È/ ‰       " !r       ÆO ‰
      h ‰      ¢ !r        Æ y       %   " !r
       Æ/ ‰      b !r       ÆO ‰      ¨ ‰      â x     ÿ   â !r       Æ/ z   r  p`ð â !r       ÆO r ÿ   pVp  â ‰ €    " !r       Æ� ‰ €    h ‰ €    ¨ ‰      '   è „y €  H   â !r       â x    ÿàÿ â !r       Æ/ ‰
 @    " z  r  p`ò â !r       äO ‰ @    b x    p"ò Æ ‰      &   ¨ ‰ @    ,       &   h       '   b !r

       â !r       È/ ‰
      " ˆ        â $Š x   Ž äO ‰      d $Š ~  Ž â�         â !r       â$ x   Ž È‰      ¢ $ ~  Ž â !r
       â %†
 j   Ž â !r       Æ/ % j   Ž â †‰ 
    è †    â !r       âO G	  �     € ê y      &   ( y      '   b $z x   Ž È $z ~  Ž Ê/ z r  ÿàõ Ê $rÿÿ   ÿ â z
 j  ÿ„ È z k   Ê †y 
    è Ay        € ê !r	       â „y    H   â Ey  À    € æ ‰	      ( „y €  H   è „y    H   è „y €  H   è „y    H   è „y €  H   â !r		       Ê ‰	      $ !r	       Ê ‰ €    $ !r       Ê ‰
 @    $ !r

       Ê ‰	
      $ !r	
	       â G™  °     € ê y      &   " ¹z  r      ä �x   ?àÿ â y      '   b $z x   Ž È $z ~  Ž Ê/ |   ÿàõ È r
ÿÿ   ÿä ä z j  ÿ„ È z k  
 Ê †y 	    è Ay        € ê !r       â !r       â !r       â !r       â !r       â ‰      " !r       â z  r      ä ‰      h ‰      ¨ ‰      è ‰      ( ‰      b !r       â !r       È/ ‰      " !r       ÆO ‰      b !r       Æ� y      '   ¢ !r
       Æ‰      â !r       Æ‰	      ( ‰
      h ‰      ¢ !r       â !r       Ä/ y      &   ( ‰ €    b !r       Æ� ‰ €    â !r		       Æ‰ €    " !r
       Æ‰	 €    b !r       ÆO ‰
 €    ¨ ‰ €    ¢ $z x   Ž â !r       Æ/ $z ~  Ž â !r       â� ‰ @    " !r       Æ‰ @    b !r	       Æ‰ @    â !r
       ÆO ‰	 @    ¢ !r       Æ ‰ @    ( ‰ @    b !r        â r   ÿ„ â!r
       â/ $x   ÿŽ âr   ÿˆ â !r       Ä� ‰
      " rÿÿ   ÿä â !r	       âO z j  ÿ„ â ‰	       b rÿÿ   ÿä â !r       âz k   â ‰      ¢ ‚   ÿàõ â !r       â   ÿàÿ â ‰      â ‚ÿÿ   ÿä Ä Š j  ÿ„ â ‰      "    ÿà÷ ä Š k   â ‰      b ÿÿ   ÿäÿ ä  j  ÿ„ â !r
       â z j  ÿˆ â !r 	       â/ 	 k   Ä †    â z k   â !r       äO †‰     â !r       Æ� †‰     â !r       Æ†     â !r       Æ†‰     â M™        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      "   " ¹z  l      â Ey  �    € â ™x?    â y      !   b ¹z  F   
   â r ÿ       â $rÿÿ   ÿ Ž â r ÿ       â x      ÿ  â r ÿ       â $rÿÿ   ÿ Ž â r  ÿ       Ä r ÿ   pRð är ÿ   pRò är   ÿ(Ž È/ xÿ     È r   ÿ� È xÿ     È |    pbô Ú G)  `    € ê y      '   " xüÿÿÿÿÀŽ â $rÿÿ   ÿ Ž â r ÿ       â y      %   b r  €ÿàÿ ä x        â y	      &   ¢ x   ÿ   ä x!   ÿ   È z! q  ÿàÿ â 'z y   Ž â x   ÿ   È/ zÿ z    â 'z	 s   Ž ÈO $z p  Ž â zÿ t    â $z |  Ž â x   ÿàÿ â $z |  Ž ä %x   ÿŽ È $z v  Ž ä$z v  Ž ä %x   Ž È $z p  Ž â z X  ÿàû â $z }  ÿŽ ä %x   Ž â x   ÿàý Æ $z	 w  ÿŽ â z X  ÿàù â %v X  Ž â zÿ Y  ¤~ ä r   àµ â %v X  Ž â x   ÿàû ä r ÿ       â $vÿ q  ÿ Ž â r        Ä xÿ     â $x   !Ž âzÿ Y  „þ ä r!   ÿŽ ä rÿÿ   d~ ä �yGþÿÿ • ¨ �yE    • è �y8þÿÿ • ( �y7    • b r4!   ÿàõ â $t;ÿ$   ÿ Ž Ä �y    • â rÿ   ÿä â %v44 Z  ; Ž â r
   ÿàõ Æ $x$   ÿŽ â rÿ   ÿä â %v

 Z  ; Ž Æ r55   ÿàÿ â $x$   ÿŽ È �y$4   ™ " $x   Ž Æ �y#4   ™ ( �y"4   ™ ( �y 
   ™ ( �y'4   ™ ( �y&
   ™ " r   ÿàõ Æ �y/4   ™ ( �y.
   ™ " r1ÿ   ÿä Æ �y-4   ™ " r6   ÿàõ Æ �y,
   ™ " r        ä r	        â �y+4   ™ b $r9ÿÿ    Æ �y*
   ™ b %v Z  ; Ž Æ �y)4    ™ b $x31$   ÿŽ Æ �y(
   ™ b $x99$   ÿŽ Æ �y04    • h �y%    • b r3   ÿàÿ Æ �y2
   ™ h �y4   ™ b %v6 Z  ; Ž Æ/ �y6   ™ d r		9   ÿàÿ ä �y5   ™ h �y9   ™ h �y@   ™ h �y>   ™ h �y?   ™ h �y<   ™ h �y=   ™ h �y:   ™ h �y;   ™ h �yD   ™ h �y1
    • h �y3
    ™ h �yC   ™ h �yA    • h �y   ™ h �y
    ™ h �yB    ™ h �yF    • b ‚x       Ä |I       â $~Jÿ   ÿ Ž â ‚x ÿ       ä |       â/ $~ÿ   ÿ Ž â |	       ä |L       ä |K       â $~Mÿ   ÿ Ž â |N       â $~Pÿ   ÿ Ž â |O       Ä |Q       â $~Sÿ   ÿ Ž â |T       ä |U       ä xGG™  ÿ    ÈO xHGww  ÿÀŽ ä xGÿ   G  ä tHI H    ä xGGww  ÿÀŽ ä xHww  ÿÀŽ ä tGJ G    Ä tÿ      ä xHÿ   H  ä xÿ   G  ä xGGww  ÿÀŽ ä xEE™  ÿ    ä� |J       Ä xHHww  ÿÀŽ ä xww  ÿÀŽ ä tGÿ  G    ä r E       ä tH	ÿ  H    ä tIJÿ      Ä xT  G    äx	2v  G    â $~Jÿ   ÿ Ž â xÿ     ä xEww  ÿÀŽ ä xGww  ÿÀŽ ä tJ E    Ä tGL G    ä xJww  ÿÀŽ ä xEHT  I    äxH2v  I    ä xÿ     ä xHÿ   G  Ä tJKÿ  J    ä |I       ä xK8™  ÿ    äxww  ÿÀŽ ä x8Hww  ÿÀŽ ä |L       Ä xGGww  ÿÀŽ ä tHIÿ      ä tIMÿ  8    ä x8Kww  ÿÀŽ ä tGLÿ  G    ä xKÿ   K  Ä |M       ä xJT  G    äxKKww  ÿÀŽ ä xGJ2v  G    ä tJM 8    ä tKN K    Ä xLJww  ÿÀŽ ä x8HT  I    äxHH2v  I    ä xJÿ   J  ä xIÿ   K  ä |M       Ä |N       ä xKKww  ÿÀŽ ä x77™  ÿ    äxJJww  ÿÀŽ ä xIIww  ÿÀŽ ä tLMÿ  L    Ä tKNÿ  K    ä rN 7       ä tJMÿ  J    ä tMPÿ  I    ä x7LT  K    äxIL2v  K    Ä xKÿ   N  ä xNNww  ÿÀŽ ä xLKww  ÿÀŽ ä tNO N    ä tLQ L    ä xKJT  M    Ä xJJ2v  M    ä xMÿ   N  ä xPÿ   L  ä xOMww  ÿÀŽ ä |Q       ä xNNww  ÿÀŽ Ä xLLww  ÿÀŽ ä xPPww  ÿÀŽ ä tOTÿ  O    ä tMQÿ  N    â &rN$   ÿ   â tRSÿ  L    ä tTUÿ  P    â &rP7$   ÿ   â xLMT  R    Ä x$M2v  R    â &rM	#   N   ä&rPI#   P   ä &rME"   M   ä&rPK"   P   ä &r"    ÿ   ä&r 7    ÿ   Ä &rM'   M   ä&r"	&   "   ä&r&I&       ä &rPJ'   P   ä &r /   M   ä &r'E.   "   Ä &r G-       ä&r',   '   ä &r 8+       ä &r'*   '   ä &r H)       ä &r'G(   '   Ä 0rÿ    A   ä 0r0ÿ0    A   â Er          â &r"82   '   Æ  r'0     A  È #r'       â &r 4   ÿ   ä&r474   ÿ   ä &r9   ÿ   ä&r779   ÿ   ä &r 	6       ä&r4I6   4   Ä &r	@      ä&r7I@   7   ä &r E5       ä &rPL/   P   ä &r/K.   &   ä&r4K5   4   Ä &rE>      ä&r7K>   7   ä &r ?       ä&r/J,   /   ä&r4J?   4   ä &r	<      Ä &r7J<   7   ä &r =       ä&r/L*   /   ä&r4L=   4   ä &r	:   	   ä&r7L:   7   â x@  ÿàõ â &r G;       â xNOT  T    â &rP$-   P   âx    ÿàÿ â &r/$(   /   Ä &r4$;   4   ä&r	GD   	   ä&r7$D   7   â rÿ   ÿä â &r8       â x#O2v  T    â &rPN+   P   â|    pbô â &r2N2   /   Ä &r4N   4   ä &r	8C   	   ä&r7NC   7   ä &rH
      ä 0r1ÿ1    A   ä 0rAÿA    A   â x@  ÿà÷ â &rP#)   P   Ä &r"H3   "   ä&r2#3   2   ä&r4#
   4   ä&r	HB   	   ä0rFÿF    A   ä &r7#B   7   Ä 0r%ÿ%    A   â Er        â  r'1     A  â rA     A  ârÿ   ÿäÿ â  rF     A  â ErP P       â  r00%     A  â Er" "       â  r1%1     A  âEr2 2       â  rA%A     A  âEr4 4       â  rF%F     A  â Er	 	       Ä Er7 7       ä x@  ÿà÷ ä x@  ÿàù â #r0P      â #r'"      â #r 12       â #r      â #rA4      â #r	      â #rF7      â x€   ÿàÿ Ä x€   ÿàÿ ä x€   ÿàÿ ä x!!€   ÿàÿ ä rÿ   ÿäÿ ä rÿ   ÿä â D)       À â Gy  `îÿÿÿÿƒ ê Ay        € ê Ey  °     € â G‰  �     € ê $x   Ž Ê ˆs  üÿ H   è ˆs €üÿ H   è ˆs  ýÿ H   è ˆs  €ýÿ H   è ˆs  þÿ H   è ˆs €þÿ H   è ˆs  ÿÿ H   è ˆs €ÿÿ H   ä Ay        € ê {          ì M        € ê „y €   H   b Ey       € æ „y     H   ¨ „y €  H   è „y    H   ( „y	 €  H   h „y    H   ( „y    H   ( „y    H   ( „y  	  H   " !r       â/ !r       ÄO „y €	  H   â !r       â� !r       Æ!r		       â!r       È ‰	      b !r       Æ ‰      ¢ !r       Æ „y €  H   â !r       Ê ‰      " !r	       Æ/ y	      %   b !r       ÆO ‰      ¨ ‰      â !r       Ê ‰      " !r       âO !r       È� ‰ €    ¨ ‰ €    â !r       Ê ‰
 €    " !r       âO x	   ÿ   â/ !r       Æ� ‰ @    b x   ÿàÿ äz  r  p`ð â ‰ @    ¢ z  r  p`ò ä r ÿ   pVp  äx    pVò  â !r

       Ä „y €  H   è ‰
 @    ( ‰      &   ä ˜        Ä ™      &   " !r       â/ !r       äO ‰      '   h ™      '   ¨ ‰      b !r

       Æ ‰      " ˆ        â $Š x  Ž ä�‰	
      â $š x  Ž ä$Š ~  Ž ä/ $š ~  Ž âO !r       â%† j   Ž â !r       Æ %– j   Ž â !r	
	       â� †‰     è †™    â G	  �     € ê y      &   ( y      '   b $z x  Ž È $z ~  Ž Ê/ z r  ÿàõ È rÿÿ   ÿä ä z j  ÿ„ È z k   Ê †y 	    è Ay        € ê !r        â „y    H   â Ey  0    € ä !r       â „y    H   æ !r       â „y	  
  H   è ‰       ( „y €  H   è „y €  H   è „y
    H   è „y    H   è „y €  H   â !r         Æ „y €  H   è ‰       ( „y €
  H   â !r        Ê ‰ €    $ !r       ä „y    H   è ‰ @    $ !r       Ê ‰       $ !r        â G  °     € ê y      &   " ¹z  r      ä �x   ?àÿ â y       '   b $z x  Ž È $z   ~  Ž Ê/ |     ÿàõ Ê $rÿÿ   ÿ â z  j  ÿ„ È z  k   Ê †y     è Ay        € ê „y €  H   â !r       â !r       â !r

       â !r       â !r       â !r       â !r

       â !r       â !r	       â !r       â !r

       â y      '   ( ‰      h ‰      ¨ ‰
      â !r       â/ !r       â !r       ÆO ‰       b !r	
       Æ� ‰      ¨ ‰      è ‰	      " !r         â/ !r       ÈO ‰  €    b !r       Æ� ‰ €    ¢ !r
	       Æ‰      è ‰
 €    ( y	      &   b !r        â/ !r       ÈO ‰ @    b !r       Æ� ‰ @    ¢ !r
       Æ‰ €    â $z	 x  Ž Æ‰ @    " $z ~  Ž â !r        â/ !r       âO z  r      â !r       Æ� ‰	      " $x
   ÿŽ âr   ÿ„ â !r       â‰       d ‚
   ÿà÷ ä ‰ @    ¢ rÿÿ   ÿä ä z j  ÿ„ â ‰      â ‚ÿÿ   ÿäÿ Ä Š j  ÿ† ä z k   ä Š k  � â !r		       â !r        È/ †™ 	   â !r       ÆO †‰     â !r       Æ� ‰       h  †‰     â M        € ê x
   ÿàÿ â !r        Æ/ r   ÿàñ È rÿÿ   ÿä  ä z j  ÿ€ È z k    Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      "   " ¹z  l      â Ey       € â ™x?    â y       !   b ¹z  F   
   â r ÿ       â x      ÿ  â $rÿÿ   ÿ Ž â r ÿ       ä r ÿ   pRð Ä r ÿ   pRò â$x     Ž Ê/ xÿ     È r   ÿ� â $
À    Ž ä $rÿÿ   ÿ Ž â xÿ     È |    pbô Ú G)  à    € ê y      '   " xüÿÿÿÿÀŽ â $rÿÿ   ÿ Ž ä y      %   b r  €ÿàÿ â $rÿÿ   ÿ Ž ä y      &   ¢ 'z y   Ž â x   ÿ   È/ zÿ z    â 'z s   Ž âO x   ÿ   Æ $z p  Ž â zÿ t    â $z |  Ž ä$z |  Ž â x   ÿàÿ â %x   ÿŽ È $z v  Ž ä$z v  Ž ä $z p  Ž ä %x   Ž È $tÿ   ÿ Ž â z X  ÿàû â %x   Ž Æ x   ÿàý â $z }  ÿŽ â zF X  ÿàù â $z w  ÿŽ â zÿ Y  ¤~ â %v X  Ž âz  q      ä r   àµ â %v X  Ž â xFF   ÿàû Ä xÿ     â $rÿÿ    Ž â r        â $x   ÿ Ž â zGÿ Y  „þ ä r ÿ       â $x   Ž â z q  ÿàÿ ä rÿÿ   d~ Ä �y6þÿÿ • ¨  �y1    • è  �y7Fþÿÿ • " r$   ÿàõ Ä x3 $       â �y9F    • d $rÿÿ    ä %v$$ Z  3 Ž È $x$   ÿŽ â r4   ÿàõ È r%%   ÿàÿ â $rÿÿ    â r2   ÿàõ Æ �yC$   ™ " $rÿÿ    Ž â r        ä �y$   ™ " r"ÿ   ÿä â %v44 Z  3 Ž ä �y$   ™ $ $x#$   ÿŽ Ä �y    • " %v22 Z  3 Ž â r!        ä �y$   ™ " $r ÿÿ    Ž Æ �y$   ™ " $x"$   ÿŽ ä $x55   #Ž â �y     • $ r33   ÿàÿ ä �y4   ™ h �y#2   ™ h �y!2   ™ h �y 4   ™ h �y$   ™ h �y$   ™ h �y$    ™ h �y$    • h �y"4   ™ h �y'2   ™ h �y%2   ™ h# �y$4   ™ h �y&4   ™ h �y)2   ™ h �y(4   ™ h �y+2   ™ h �y*4   ™ h �y.2   ™ h �y,4    • h �y-4    ™ h �y/2    ™ h �y02    • b ‚x       Ä |5       â $~;ÿ   ÿ Ž â ‚x ÿ       ä |=       â $~3ÿ   ÿ Ž ä/ $~:ÿ   ÿ Ž ä $~<ÿ   ÿ Ž â |@       â $~>ÿ   ÿ Ž Ä $~?ÿ   ÿ Ž â |A       ä x@  ÿàõ ä x    ÿàÿ ä rÿ   ÿä ä xFF@  ÿàõ È rGÿG   ÿä ä |    pbô ä x66™  ÿ    ÈO x86ww  ÿÀŽ ä x6ÿ   6  È x66ww  ÿÀŽ È t65 6    ä t8; 8    ä x2ÿ   6  ä x48ww  ÿÀŽ ä x66ww  ÿÀŽ ä x8ÿ   8  Ä x22ww  ÿÀŽ ä x51™  ÿ    â� $~;ÿ   ÿ Ž â t13ÿ  6    ä x88ww  ÿÀŽ ä t3=ÿ  2    ä x2ÿ   5  Ä t4;ÿ  4    ät6;ÿ  8    ä |;       ä x82ww  ÿÀŽ ä x55ww  ÿÀŽ ä t8; 8    Ä t5: 5    ä x;8ww  ÿÀŽ ä x:5ww  ÿÀŽ ä t;<ÿ  ;    ä x<7™  ÿ    äx24T  1    Ä x5ÿ   5  ä x442v  1    ä t:=ÿ  :    ä x7ÿ   8  ä x16T  3    ä x=<ww  ÿÀŽ Ä x662v  3    ä |3       ä x85ww  ÿÀŽ ä x<ÿ   <  ä x77ww  ÿÀŽ ä t=@ =    Ä t33ÿ  8    ä x<<ww  ÿÀŽ ä t8>ÿ  7    ä x>=ww  ÿÀŽ ä x5:T  ;    äx7:2v  ;    Ä t;? <    â $~:ÿ   ÿ Ž â t<Aÿ  >    ä xAÿ   =  ä x@ÿ   ;  ä x=;ww  ÿÀŽ ä x99™  ÿ    â$~?ÿ   ÿ Ž â |>       Ä x;Aww  ÿÀŽ ä x@@ww  ÿÀŽ ä t=:ÿ  =    ä xBÿ   9  ä t>>ÿ  ;    â $~;ÿ   ÿ Ž â t??ÿ  @    Ä x:<T  =    ä|@       ä xE9ww  ÿÀŽ ä xBBww  ÿÀŽ â &rA2C   ÿ   âx<<2v  =    â &rC:C   ÿ   â tE@ E    Ä t;; B    â &rA4   A   âx9>T  ?    â &rC<   C   â |=       ä x@Eww  ÿÀŽ ä xB;ww  ÿÀŽ ä xEÿ   E  Ä x;ÿ   ;  â &rA1   A   â t==ÿ  @    â &rC9   C   â |@       â $~ÿ   ÿ Ž â xEEww  ÿÀŽ ä x;ww  ÿÀŽ ä x>>2v  ?    Ä t@@ÿ  E    ä tÿ      â &rA6   A   ä&rC>   C   â x@T      äx@@2v      â &r2   ÿ   ä&r:   ÿ   Ä &r22!   ÿ   ä&r!:!   ÿ   ä $~?ÿ   ÿ Ž ä &r<       ä&r4       ä&r24#   2   Ä &r<<#   !   â tB?ÿ  B    â &r1"      ä&r9"      ä &r21%   2   ä&r<9%   <   ä &rA5   A   â x;=T  B    â &r6$      âx?3T  8    â &r>$      Ä &r26'   2   ä&r<>'   <   ä &rA7   A   â x==2v  B    â &rC;   C   â x832v  8    â &r5&      ä&r;&      Ä &r25)   2   ä&r<;)   <   ä &rA?   A   ä &rC=   C   ä &r7(      ä&r=(      Ä &r27+   2   ä&r<=+   <   ä &rA8   A   ä 0rÿ    A   ä 0rÿ    A   ä &rC   C   Ä &r?*      ä&r*      ä &r2?.   2   ä&r<.   <   â ErA A       â  r     A  â 0r,ÿ,    A   ä &rC@   C   Ä &r8-      ä&r@-      ä &r28/   2   ä0r0ÿ0    A   ä &r<@/   <   ä 0rÿ    A   â #rA      â  r,     A  âErC C       â  r0     A  â Er        â  r     A  â Er        â  r,,     A  âEr2 2       â  r00     A  â Er< <       Ä x@  ÿà÷ ä x@  ÿàù â #r	C   	   â #r      â #r,      â #r2      â #r0<      â x€   ÿàÿ â $rÿÿ   Ž â x€   ÿàÿ Ä x€   ÿàÿ ä rÿ   ÿä â G©  àðÿÿÿÿƒ ê Ay        € ê ˆ 
 ýÿ H   è ˆ 
	€ýÿ H   è ˆ 
 þÿ H   è ˆ 
€þÿ H   è ˆ 
 ÿÿ H   è ˆ 
€ÿÿ H   è {          ì M        € ê „y      H   ( „y     H   h „y  €   H   ¨ „y     H   è „y  €  H   ( „y  €  H   h „y     H   h „y
  €  H   h „y     H   h „y     H   h „y  €  H   b !r       Æ „y  €  H   " !r       Æ/ „y     H   b !r	       ÆO „y     H   ¢ !r       Æ� „y  €  H   â !r       Æ„y!  €  H   â !r       Æ„y     H   â !r       Æ „y#  €  H   â !r

       Æ ‰      " !r       Æ ‰      b !r       â !r

       ä y      &   b !r       Æ y      '   " !r       â/ !r       âO !r

       È� ‰      b !r       â!r!       ä ‰	
      ¢ !r       â!r       ä ‰      â !r#       Æ ‰      ( ‰      h ‰      " !r       â/ !r	
	       ÈO ‰      b !r       Æ� ‰	      ¢ !r       Æ‰ €    â !r       Æ‰
      " !r       Æ ‰      ( ‰ €    b !r       Æ/ y      %   b !r	       ÆO ‰ €    ¢ !r       Æ� ‰	 €    â !r

       Æ‰ @    " !r       Æ ‰
 €    " !r       Æ‰ €    h ‰ @    b !r       âO !r		       È� ‰ @    ¢ !r       â!r
       â $x
   ÿ Ž ä/ ‰	 @    " !r       â$z x  
Ž â z 
 r  p`ð â‰ @    b !r       â x
   ÿàÿ â $z
 ~  Ž â ‰ @    â r  ÿ   pVp  Ä z  r  p`ò â ‰      " !r       äO x     p"ò â ‰      ¨ ‰      d Š 
 r  ÿàõ â !r	       â $v	ÿ r  ÿ Ž Ä ‚ÿÿ   ÿä â !r       â/ ‰      " Š  j  ÿ„ â$ÿ   ÿ Ž â !r       â� ˆ        â ‰      b Š  k   â !r       â	   ÿàÿ â‰      â r 	
   ÿ† â !r       âO 	
   ÿàõ â%†
 j   Ž â rÿÿ   ÿäÿ â !r       âÿÿ   ÿä â %

 j   Ž â 	 j  ÿ„ â †‰     â z  j  ÿ† Ä 		 k   â † 
   â !r       â z  k  � ä †‰     â !r       Æ/ †     â !r       Æ� †‰     â M™        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      "   " ¹z  l      â Ey       € â ™x?    â y       !   b ¹z  F   
   â $rÿÿ   ÿ Ž â r ÿ       â y      %   ¢ $rÿÿ   ÿ Ž ä $rÿÿ   ÿ Ž â r ÿ   pRð Ä r ÿ   pRò â$x     Ž ä/ $x   ÿ Ž ÆO xÿ     È r   ÿ� â $€    Ž Æ xÿ     È |    pbô Ú G)  Ð    € ê y      '   " $rÿÿ   ÿ Ž â xüÿÿÿÿÀŽ â $rÿÿ   ÿ Ž â y	      &   b $tÿ   ÿ Ž â r  €ÿàÿ â $x   ÿ Ž Æ xÿ     â $x   ÿ Ž ä 'z y   Ž È 'z	 s   Ž â/ zÿ z    Æ $z p  Ž â zÿ t    â $z |  Ž ä$z |  Ž â x   ÿàÿ â %x   ÿŽ È $z v  Ž ä$z v  Ž ä %x   Ž È $z p  Ž â z X  ÿàû â $z }  ÿŽ ä %x   Ž â x   ÿàý Æ $z	 w  ÿŽ â z X  ÿàù â %v
 X  Ž âzÿ Y  ¤~ ä r   àµ â %v X  Ž â x   ÿàû ä r ÿ       ä zÿ Y  „þ Ä z q  ÿàÿ ä rÿÿ   d~ ä �y0þÿÿ • ¨  �y/þÿÿ • â r   ÿàõ â $tÿ$   ÿ Ž ä �y.    • $ $rÿÿ    Ä %v Z   Ž â �y     • f $x$   ÿŽ â �y&
    • æ  $x		   Ž Ê �y-   ™ è  �y(   ™ ( �y%   ™ ( �y$   ™ " r   ÿàõ Æ �y!   ™ $ $rÿÿ    ä %v Z   Ž â �y   ™ & $x$   ÿŽ â �y'    ™ ( r   ÿàÿ â �y)    • ( �y   ™ ( �y   ™ ( �y#   ™ ( �y"   ™ ( �y   ™ ( �y   ™ ( �y   ™ h �y*    ™ h �y+    • h �y   ™ b $rÿÿ    Ž Ä $rÿÿ    Ž Ê �y,    • b ‚x       ä $~2ÿ   ÿ Ž ä $~	ÿ   ÿ Ž â/ ‚x ÿ       ä |       ä |5       â $~3ÿ   ÿ Ž Ä $~4ÿ   ÿ Ž ä $~6ÿ   ÿ Ž â |8       â $~:ÿ   ÿ Ž ä $~9ÿ   ÿ Ž â x@  ÿàõ ä x    ÿàÿ Æ $rÿÿ    â |    pbô ä x@  ÿà÷ ä x

@  ÿàù Æ $rÿÿ   Ž â x@  ÿà÷ â $rÿÿ    â x€   ÿàÿ ä x€   ÿàÿ â $rÿÿ   Ž â x00™  ÿ    ÈO x10ww  ÿÀŽ ä x0ÿ   0  ä t12 1    ä x00ww  ÿÀŽ ä x1ww  ÿÀŽ ä t0	 0    Ä x1ÿ   1  ä xÿ   0  â $~2ÿ   ÿ Ž â x	1ww  ÿÀŽ ä tÿ      ä xww  ÿÀŽ ä x/™  ÿ    Ä� x00ww  ÿÀŽ ä t	2ÿ  	    ä t25ÿ      ä xÿ     ä t03ÿ  0    â $~3ÿ   ÿ Ž â xww  ÿÀŽ Ä x1ww  ÿÀŽ ä xT  0    äx/2v  0    ä t03     ä t14 1    ä x30ww  ÿÀŽ Ä x41ww  ÿÀŽ ä x0ÿ   0  ä x1ÿ   1  ä x	T  2    äx	2v  2    ä t36ÿ  3    Ä t45ÿ  4    ä |2       ä x	0ww  ÿÀŽ ä x11ww  ÿÀŽ ä x03T  4    ä t52ÿ  	    Ä x	32v  4    ä t46ÿ  1    â &r1-   ÿ   ä&r-0-   ÿ   â x2.™  ÿ    â&r1/(   1   âx.5T  4    â &r3	(   -   ä &r(%   1   Ä &r3.%   3   â x%ÿ   2  â $~1ÿ   ÿ Ž â x22ww  ÿÀŽ ä x-52v  4    â $~4ÿ   ÿ Ž â x%%ww  ÿÀŽ â &r($   (   â t21 2    â &r$-$   3   â t34 %    Ä x1ÿ   2  â $~5ÿ   ÿ Ž â x  ™  ÿ    äx43ww  ÿÀŽ ä x%2ww  ÿÀŽ ä x11ww  ÿÀŽ ä x7ÿ      Ä t45ÿ  4    â $~5ÿ   ÿ Ž â x3ÿ   3  ä t%6ÿ  %    ä t28ÿ  1    â $~1ÿ   ÿ Ž â x6 ww  ÿÀŽ ä x87ww  ÿÀŽ Ä x33ww  ÿÀŽ ä t55 6    ä t81 8    ä x %T  4    ät3:ÿ  3    ä x%%2v  4    Ä |7       ä x45ww  ÿÀŽ ä x68ww  ÿÀŽ ä x12T  3    äx222v  3    ä t37ÿ  4    Ä t49ÿ  6    â &r6   ÿ   ä&r0   ÿ   ä &r6/   6   ä&r		      â x5ÿ   5  â &r6#   6   âx8ÿ   8  â &r	.#   	   â x55ww  ÿÀŽ â &r6"   6   âx88ww  ÿÀŽ â &r	-"   	   â x3T  4    â &r( !   (   â t:ÿ  5    â &r6    6   â t57ÿ  8    â &r$!   $   âx332v  4    â &r	   	   â xT  5    â &r(%   (   Ä &r$3   $   ä &r6%   6   ä&r	3   	   â x2v  5    â &r(1   (   ä&r$   $   ä &r11   6   Ä &r	   	   ä 0r&ÿ&    A   ä &r(2'   (   ä0r)ÿ)    A   ä &r12*   1   ä 0r,ÿ,    A   Ä &r$'   $   ä&r	*   	   ä 0r+ÿ+    A   â Er( (       â  r&)     A  âEr1 1       â  r),     A  â Er$ $       â  r&&+     A  â Er	 	       â  r,,+     A  â #r(      â #r&1      â #r$      â #r,	      â G©  �óÿÿÿÿƒ ê Ay        € ê ˆ  þÿ H   è ˆ €þÿ H   è ˆ  ÿÿ H   è ˆ €ÿÿ H   è {          ì M        € ê „y      H   " z  r  p`ð Æ „y  €   H   b r  ÿ   p"p Æ „y     H   ¨ „y     H   è „y  €  H   ( „y     H   h „y     H   h „y	  €  H   h „y     H   h „y
  €  H   h „y  €  H   b !r       Æ „y  €  H   " !r       â/ !r       äO y      &   b !r       â� !r       â!r       ä!r       â !r		       È ‰      ¢ !r       Æ ‰	      â !r

       Æ ‰      " !r

       È !r
       â !r       ÈO ‰      " !r	       Æ� ‰      ¢ !r       Æ‰      è ‰      " !r       â !r       ÈO ‰      " !r       Æ� ‰ €    ¢ !r       Æ‰ €    è ‰
 €    " !r       â x        â !r       ÆO ‰ €    " !r       Æ� y      '   ¢ !r

       Æ‰ @    è ‰ @    ( ‰	
 @    b !r       â !r       â� x   ÿàÿ â $z x  Ž â/ !r       Äz  r  p`ò â ‰      " $z ~  Ž âO !r	
	       âx     p"ò â ‰      d 
  r  ÿàõ ä ‰
 @    ¦ $ÿÿ   ÿ â ‰	      â 
  j  ÿ„ È 
  k   â !r       â %v j   Ž â !r       È/ †	     â !r

       ÆO †    â !r		       Æ� ‰
      h  †	 	    â M™        € ê ¹z  r      â !r
       â/ �x   ?àÿ Ì |   ÿàñ Ê $r ÿÿ   ÿ  â z j  ÿ€ È z k     Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y/      &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž È %/ \   Ž Ì �	    ™ b ¹z  l      â $r.ÿÿ   ÿ Ž â ™x?    â y       "   b Ey  °    € â '
/ m  . Ž äy      !   d 'Š	/ s  . Ž â 
ÿ n    â y      %   ¨ y      '   â $ÿÿ   
Ž Ä $x     Ž Ê/ xÿ     È r   ÿ� â $
 o  /Ž ä$rÿÿ   ÿ Ž â xÿ     â $‚ÿÿ   / Ž Æ |    pbò ä Šÿ t  	  Ö G  �    € ê r
ÿ   ÿ3Ž âß $z w  ÿŽ â xüÿÿÿÿÀŽ â $z	 }  ÿŽ â|

   ÿàÿ â $rÿÿ   ÿ Ž â Ey °    € â $x   
Ž â xÿ
    ÿÀ‚ â'z y   Ž â xÿ
àÿÿÿÿÀ€ Ä r	   àµ â$z p  ÿŽ â x	ÿ     â $x   ÿ Ž â zÿ z    â $rÿÿ   ÿ Ž â r	ÿÿ   	d~ ä xÿ     â G  Ð    € ê $z
 v  Ž ä$tÿ   ÿ Ž Ä $z |  
Ž È $x   Ž È %v X  Ž Ê r   ÿ‚ â$tÿ$   ÿ Ž â �y    • ¤  r   �  Ê �y   • â r
   ÿ‚ Ê $rÿÿ   	Ž  ä %v

 Z   Ž È $x$   ÿŽ È $x   Ž ä �y   • ( �y
   ™ ¨  �y
   ™ ¨  �y
   ™ ¨  �y
   ™ ¨  �y
   ™ ¨  �y
   ™ ¨  �y
   ™ ¨  �y
    ™ ¨  �y
    • ¢ ‚x       Ä $~ ÿ   ÿ Ž ä $~ÿ   ÿ Ž â/ ‚x ÿ       ä $~ÿ   ÿ Ž ä $~"ÿ   ÿ Ž â x    ÿàÿ ä x™  ÿ    È� xÿ     ä xww  ÿÀŽ ä xww  ÿÀŽ ä t      ä t     â $~ ÿ   ÿ Ž â x
ww  ÿÀŽ Ä xww  ÿÀŽ ä t
ÿ  
    ä t ÿ      ä x™  ÿ    äxÿ     ä x
T      Ä x

2v      â $rÿÿ    Ž â xww  ÿÀŽ â &r   ÿ   âO xÿ     ä t ÿ      â $~ÿ   ÿ Ž â xÿ     â $~ ÿ   ÿ Ž â xww  ÿÀŽ â $~ÿ   ÿ Ž â xww  ÿÀŽ Ä xww  ÿÀŽ â &r

      â tÿ      ä t      ä t     â $~ ÿ   ÿ Ž â xT      ä xww  ÿÀŽ Ä xww  ÿÀŽ â &r
   
   â x2v      ä t ÿ      ä t"ÿ      ä xÿ     ä xÿ     â $~ÿ   ÿ Ž â xww  ÿÀŽ â $~ÿ   ÿ Ž â xww  ÿÀŽ â &r
   
   â xT      Ä tÿ      ä x2v      â &r
   
   â tÿ      Æ &r
   
   â xT      äx2v      Æ &r
   
   ä 0rÿ    A   ä &r
   
   ä 0rÿ    A   Æ Er
 
       ä  r     A  È #r
   ÿ   ä Ay       € ê G‰  `    € ê $z |  Ž â x
     â $tÿ   ÿ Ž ä $z v  Ž È$x   Ž â x€   ÿŽ Æ %v X  Ž Ê r"   ÿ€ ä x€ÿÿÿÿàÿ â $t)ÿ$   ÿ Ž â �y    • ¢  $x#   
  Æ �y$@  • è  �y'"   • " r   ÿàñ Æ �y"B  • d $rÿÿ   	  ä %v* Z  ) Ž È $x$   ÿŽ È $x++   Ž ä �y"   • ¨ �y*   ™ ¨  �y*   ™ ¨  �y*   ™ è  �y*   ™ â  r(   ÿàñ Æ �y*   ™ ä  $rÿÿ   	  ä %v(( Z  ) Ž â �y*   ™ æ  $x$   ÿŽ â �y!*    ™ æ  $x))   Ž â �y*   ™ è  �y"D  • h �y(   ™ h �y(   ™ h �y(   ™ h �y(   ™ h �y(   ™ h �y"(   ™ h# �y#(   ™ h �y *    • h �y&(    ™ h �y%(    • b ‚x       Ä $~,ÿ   ÿ Ž ä $~*ÿ   ÿ Ž â ‚x ÿ       ä $~(ÿ   ÿ Ž ä/ $~+ÿ   ÿ Ž â x@   ÿàÿ È |    pbð ä x€  ÿàó ä x   ÿàÿ Æ $rÿÿ   Ž  â x''™  ÿ    Èx-ÿ   '  ä x''ww  ÿÀŽ ä x--ww  ÿÀŽ ä t,, '    ä t** -    â $~'ÿ   ÿ Ž â x0,ww  ÿÀŽ Ä x)*ww  ÿÀŽ ä t''ÿ  0    ä t0(ÿ  )    ä x*ÿ   *  ä x('T  0    äx,ÿ   ,  â $~)ÿ   ÿ Ž â x**ww  ÿÀŽ â &r(   ÿ   âO x''2v  0    Ä x™  ÿ    ä x,,ww  ÿÀŽ ä t*+ÿ  *    â &r'      â x+ÿ     â $~'ÿ   ÿ Ž â t))ÿ  ,    â $~ÿ   ÿ Ž â x,ww  ÿÀŽ Ä x0+ww  ÿÀŽ ä x()T  *    ä t'' ,    ä t 0    â &r(      â� x))2v  *    â $~*ÿ   ÿ Ž â xÿ   '  â $~(ÿ   ÿ Ž â x+'ww  ÿÀŽ Ä x'ww  ÿÀŽ â $~ÿ   ÿ Ž â x,ww  ÿÀŽ â &r))      â x™  ÿ    ät**ÿ  +    ä t'(ÿ  '    ä xÿ     Ä t,ÿ  ,    ä x(ÿ     ä x*T  '    äx*2v  '    â $~*ÿ   ÿ Ž â x'ww  ÿÀŽ â $~ÿ   ÿ Ž â xww  ÿÀŽ â $~ÿ   ÿ Ž â x((ww  ÿÀŽ Ä t'*ÿ  '    ä t     ä t (    â $~(ÿ   ÿ Ž â x,T  '    äx,2v  '    ä x+ww  ÿÀŽ Ä x'ww  ÿÀŽ â &r   )   â t**ÿ  +    ä t'(ÿ  '    â &r      â x™  ÿ    ä x*T  '    ä xÿ     â $~ÿ   ÿ Ž Ä &r   ÿ   â xww  ÿÀŽ â $rÿÿ    Ž â xÿ     ä tÿ      â &r      â xÿ     â $~ÿ   ÿ Ž â xww  ÿÀŽ â $~ÿ   ÿ Ž â xww  ÿÀŽ â $~ÿ   ÿ Ž â xww  ÿÀŽ Ä x**2v  '    ä tÿ      ä t     ä t     â $~ÿ   ÿ Ž â xT      â &r*      â xww  ÿÀŽ Ä xww  ÿÀŽ ä x2v      â &r      â t(ÿ      ä tÿ      ä xÿ     ä xÿ     â $~ÿ   ÿ Ž â xww  ÿÀŽ â $~ÿ   ÿ Ž â xww  ÿÀŽ â &r      â xT      Ä tÿ      ä x2v      â &r      â tÿ      Æ &r"      â xT      â&r!      â x2v      â 0rÿ    A   ä &r#      ä 0r ÿ     A   â Er        â &r&      Ä 0r$ÿ$    A   â  r       A  â 0r%ÿ%    A   â Er        ä #r       ä  r$$%     A  È #r$      â G‰  õÿÿÿÿƒ ê Ay        € êß r  ÿ   pRð är  ÿ   pRò Ö ˆ €ÿÿ H   è {          ì M        € ê „y      H   " r ÿ   pRð â $z// x  Ž Ä „y €   H   h „y	    H   ¢ !r         È !r         È/ !r  	       ÊO ‰       $ !r        Ê ‰      $ !r       Ê ‰ €    $ !r       Ê ‰ @    $ !r       Ê ‰	      b  M	        € ê $tÿ   ÿ Ž â !r		       â/ $z ~  /Ž È %v j   Ž Ê †y 	    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž È % \   Ž Ê �	    ™ b $rÿÿ   ÿ Ž â Ey       € â $r
ÿÿ   ÿ Ž â y      "   b 'Š s   Ž Æ y      !   b '
 m   Ž â rÿ   ÿü‚ Æ/ $x     Ž Ê x	ÿ      ä Šÿ t    ä 
ÿ n    ä r	    ÿ� â $r	ÿÿ   ÿ Ž â G       € ê z ÿ `  pPô ä z ÿ ^  pPò Ä z ÿ b  p ö ä z ÿ a   Sô ä z ÿ _  Sò ä z ÿ c  0't Ö       '   "        ä¢       â ©
      '   d $ x  ÿŽ ä 	      %   ¢ $ª x  ÿŽ Æ ©      %   â $ ~  ÿŽ ä $ª

 ~  ÿŽ Æ/    	àë äO	ÿ   	  ä ¢

   àÇ ä�¨ÿ     ä 	ÿÿ   	Äþ ä    ÿàý Ä ¢
   ÿàû ä ¢ÿÿ   „þ â $
ÿÿ   	 â  ^  ÿ† â $r	ÿÿ   ÿ Ž â ª b  ÿˆ â $¢ÿÿ   Ž â  _  
� â $r
ÿÿ   ÿ Ž Æ ª c   â �	     h �©
     d Ay        € ê ¹z  l      â xÿ     â ™x?    â $ÿÿ   
Ž â Ey       € â $rÿÿ   ÿ Ž Ä $
 o  Ž â|    pbò â $‚ÿÿ    Ž ä $rÿÿ   ÿ Ž Ô G  �    € ê y      '   " $rÿÿ   ÿ Ž ä $z w  ÿŽ â y      %   b $tÿ   ÿ Ž ä 'z y   Ž Ê zÿ z    Ê $z |  Ž ä $z }  ÿŽ ä $z v  Ž âxüÿÿÿÿÀŽ Æ $z p  Ž ä/ $x    
Ž ä %v `  Ž Æ r    à‘ â$x&   ÿ Ž â xÿ     â %v X  Ž Æ rÿÿ   $~  â %x&&   ÿŽ È $rÿÿ   ÿ Ž â x&   ÿüŽ â $rÿÿ   ÿ Ž ä $x   ÿ Ž ä r   ÿàñ â $tÿ$   ÿ Ž â �y    • ¦ $x'     Ê �y    • â  r    ÿàñ Æ �y   • $ $rÿÿ     ä %v Z   Ž È $x$   ÿŽ È $x   Ž Ê �y   ™ h �y   ™ ¨ �y   ™ ¨ �y   ™ ¨ �y   ™ ¨ �y   ™ ¨ �y   ™ ¨ �y    ™ ¨ �y    • ¢ ‚x       Ä $~ÿ   ÿ Ž â ‚x ÿ       ä $~!ÿ   ÿ Ž ä $~ÿ   ÿ Ž â/ z ÿ `  pPð È z ÿ a   Sð ä x™  ÿ    È� xww  ÿÀŽ ä xÿ     ä t     äxww  ÿÀŽ ä xww  ÿÀŽ ä t      Ä t!ÿ      ä x!ÿ     â $~ÿ   ÿ Ž â xÿ      ä x™  ÿ    äx  ww  ÿÀŽ â $~ÿ   ÿ Ž â xww  ÿÀŽ Ä x!!ww  ÿÀŽ ä tÿ       ä x#ÿ     â $~ ÿ   ÿ Ž â tÿ      â $~ÿ   ÿ Ž â tÿ  !    ä xww  ÿÀŽ Ä x##ww  ÿÀŽ ä x!T      ät      ä x2v      â &r!!   ÿ   ât #    â $~ÿ   ÿ Ž â x"ww  ÿÀŽ â $~ ÿ   ÿ Ž â x#ww  ÿÀŽ â &r   !   âO x%T      Ä x2v      ä tÿ  "    â &r%      â t  ÿ  #    ä x"ÿ     ä xÿ     â $~!ÿ   ÿ Ž â x""ww  ÿÀŽ â $~ÿ   ÿ Ž â xww  ÿÀŽ â &r      â xT       Ä t!!ÿ  "    ä x 2v       â &r      â tÿ      Æ &r        â x!T      äx!2v      Æ &r       ä 0rÿ    A   ä &r      ä 0rÿ    A   â G‰       € ê r   ÿàñ â �y     ¨ $x'     Ê �y     è  �y    " $~"ÿ   ÿ Ž ä $~$ÿ   ÿ Ž ä $~#ÿ   ÿ Ž ä $~ÿ   ÿ Ž â x ÿ     Ä� x!ww  ÿÀŽ ä x ww  ÿÀŽ ä t!" !    ä t$     ä x !ww  ÿÀŽ ä xww  ÿÀŽ Ä x!ÿ   !  ä t #ÿ       ä tÿ      ä x!ww  ÿÀŽ ä x! T      äx  2v      Ä x"ÿ     â&r!   ÿ   â xÿ     â $~!ÿ   ÿ Ž â x$ww  ÿÀŽ â $~ÿ   ÿ Ž â xww  ÿÀŽ ä x""ww  ÿÀŽ ä t#ÿ      Ä t##ÿ      ä t!! $    ä t "    â &r       â x"T  #    â $~ÿ   ÿ Ž â x!ww  ÿÀŽ â $~ ÿ   ÿ Ž â xww  ÿÀŽ Ä x!ÿ   !  ä xÿ     â &r"      â x2v  #    ä tÿ      ä t ÿ      â $~ÿ   ÿ Ž â x !ww  ÿÀŽ Ä x!ww  ÿÀŽ â &r      â xT      ätÿ       ä x2v      â &r      â tÿ  !    Æ &r      â xT      äx2v      Æ &r      ä 0rÿ    A   äO &r      Æ  r     A  ä Er        Ê #r      â x    ÿàÿ â  r     A  â x@  ÿàõ ä |    pbð ä Er        â $rÿÿ    â x@  ÿàó ä x€   ÿàÿ â #r      Ä $rÿÿ   Ž  È G‰  ðõÿÿÿÿƒ ê Ay        € ê r ÿ   pRð âEy  €     € â r ÿ   pRò Ö G‰  P     € ê z ÿ `  pPð â $x    Ž Æ z ÿ a   Sð ä ˆs €ÿÿ H   ö ˆ    H   ä Ay        € ê {          ì M        € ê z ÿ `  pPð â „y      H   f z ÿ a   Sð Ú „‰ €   H   ¨ „‰    H   â !r        È? !‚       ÈO !‚       â� G‰        € ê „y  €  H   ( „y €   H   h „y    H   ¨ „y    H   è „y €  H   "!r         â !r       Æ/ !r         âO !r       Æ� !r        ä‰      " z ÿ `  pPð ä r ÿ   pRò ä z ÿ a   Sð â !r       Ê ‰       $ !r         Ê ‰  €    $!r        Ê ‰ @    $ !r       Ê ‰      $ !r       â G‰        € ê ‰       $ !r         Ê ‰       $ !r        Ê ‰ €    $ !r       Ê ‰ @    $ !r       Ê ‰      $ !r       ä M        € ê !r		       â G‰  `    € ê $v ÿ h  ÿ Ž â !r
       È x     pBð Ú G	  p    € ê x  ÿÿÿÿÿàÿ È x       € Ê $x     ÿ Ž È ‚{   €     $ xÿ     È Iy Ðëÿÿÿÿƒ ê  x ;ª¸¿  A  Ì s          $ !t   €?    Ì s        $  r     A  È  r		     A  â Gy  P    € ê  x '7=  A  â x*BL?  A  Æ #t   €?    È  r      A  â  x    ?  A  Ê s     $   $ !t  €?    È  r       A  È  r		      A  â Gy  °    € ê x     pRð Ú G‰  à     € ê x     pRð Ú G	        € ê 	z  i    � ä 	z		 i    � Æ  x ;ª¸¿  A  â 	z		 i €  � Ê s        $ !t  €?    Ì s        $  r       A  È  r		      A  â Gy  Ð     € ê  r		     A  â Gy  °     € ê 	x   à@  � ä 	x		  à@  � Æ  x #ÛÙ?  A  â 	x		  àÀ  � Æ  x;ª¸¿  A  ä !t		  €?    È s        $ !t  €?    Ì s        $  r       A  È  r	 	     A  ä y       &   ( y      %   ( y      '   b $z   x  Ž ä $tÿ   ÿ Ž Ä $z  ~   Ž È/ %v  j   Ž Ê †y 	    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y	      &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú $ÿ   ÿ Ž ä $rÿÿ   	 Ž È % \   Ž Ê �	    ™ b ¹z  l      â Ey  `    € â ™x?    â y      "   b $rÿÿ   ÿ Ž ä $r
ÿÿ   ÿ Ž â y      !   ¢ $rÿÿ   ÿ Ž ä $ÿÿ   ÿ Ž â y      %   " $ÿÿ    Ž È '
  m   Ž È $‚ÿÿ   ÿ Ž â 
 ÿ n     Ê $ÿÿ    
Ž â r ÿ   pRò ä/r ÿ   pRô â $
 o  Ž ä $x    Ž äO $‚ÿÿ    Ž ä $rÿÿ   ÿ Ž â x ÿ     â 'Š s   Ž Æ r     ÿ� â $€   Ž ä $x   ÿ Ž â xÿ      È |    pbö ä Šÿ t    Ö G9  �    € ê y      '   " $rÿÿ   ÿ Ž ä $rÿÿ    Ž ä 'z y   Ž È $z p  Ž â zÿ z    â $z }  ÿŽ â x üÿÿÿÿÀŽ â $z  w  ÿŽ ä $z |  Ž ä $x   
Ž ä $z v  Ž Ä$x   ÿ Ž â r     à± â%x.   ÿŽ â xÿ     Æ $vÿ X  ÿ Ž â rÿÿ   d~  â $vÿ Y  ÿ Ž ä $rÿÿ   ÿ Ž ä $x   ÿ Ž ä %x   .Ž Ê r   ÿàñ Ê $x     Ê �y%   • ¢  r,    ÿàñ â $t-ÿ$   ÿ Ž ä �y#   • ä  $rÿÿ     ä %v,, Z  - Ž È $x$   ÿŽ È $x--   Ž Ê �y,   ™ ( �y,   ™ b $v'ÿ p  ÿ Ž Æ �y,   ™ " %x'   Ž Æ �y,   ™ ( �y"   • ( �y,   ™ ( �y,   ™ h �y,   ™ h �y$   • b %x '   Ž Æ �y,    ™ h �y    • h �y    • b %x2'    Ž â r0.   ÿàñ Ä �y,    • h# �y2   • h �y2   • b $x1/     È %x*'   0Ž â�y0    • h �y*    • b %x('   *Ž Ê �y (    • b %x&'   (Ž Ê �y!&    • b ‚x       ä $~4ÿ   ÿ Ž ä $~1ÿ   ÿ Ž â ‚x ÿ       ä $~'ÿ   ÿ Ž ä $~0ÿ   ÿ Ž Ä $~+ÿ   ÿ Ž â/ x%%™  ÿ    ÈO x2ÿ   %  ä x%%ww  ÿÀŽ ä x,2ww  ÿÀŽ ä t*4 %    ä t,1 ,    â $~%ÿ   ÿ Ž â x(*ww  ÿÀŽ Ä x&,ww  ÿÀŽ ä t''ÿ  (    ä t(%ÿ  &    ä x)ÿ   *  ä x*#™  ÿ    ä� x&'T  (    Ä x,ÿ   ,  ä x''2v  (    ä x(ÿ   *  â $~#ÿ   ÿ Ž â x))ww  ÿÀŽ â $~%ÿ   ÿ Ž â x,,ww  ÿÀŽ ä x**ww  ÿÀŽ Ä x((ww  ÿÀŽ â &r&&   ÿ   ât00ÿ  )    ä t++ÿ  ,    ä t## *    ä t%% (    â &r&'   &   âx(0T  +    â$~'ÿ   ÿ Ž â x*#ww  ÿÀŽ â $~)ÿ   ÿ Ž â x,%ww  ÿÀŽ â &r&(   &   â x+02v  +    Ä t''ÿ  *    ä t,)ÿ  ,    â &r++   &   â x*ÿ   #  ä x&'T  ,    äx)ÿ   %  â $~#ÿ   ÿ Ž â x("™  ÿ    â $~%ÿ   ÿ Ž â x''2v  ,    â &r&&   +   â x**ww  ÿÀŽ Ä x")ww  ÿÀŽ ä x+ÿ   (  â &r&'   &   â t##ÿ  *    ä t"%ÿ  "    â $~%ÿ   ÿ Ž â x*(ww  ÿÀŽ â $~'ÿ   ÿ Ž â x(+ww  ÿÀŽ Ä x)#T  "    ät*% *    ä t(' (    ä x"#2v  "    â &r#)   &   â x&*ww  ÿÀŽ â $~)ÿ   ÿ Ž â x'ÿ   *  â $~%ÿ   ÿ Ž â x*(ww  ÿÀŽ â $~,ÿ   ÿ Ž â x+$™  ÿ    Ä x''ww  ÿÀŽ ä t&)ÿ  &    ä t%%ÿ  *    ä x(ÿ   (  ä x)ÿ   +  ä t,,ÿ  '    â $~'ÿ   ÿ Ž â x$(ww  ÿÀŽ â $~(ÿ   ÿ Ž â x*&T  %    Ä x%&2v  %    â $~&ÿ   ÿ Ž â x))ww  ÿÀŽ ä x++ww  ÿÀŽ ä t''ÿ  $    â &r**   ÿ   â t(( )    ä t&& +    â $~+ÿ   ÿ Ž â x$,T  '    â$~)ÿ   ÿ Ž â x',2v  '    â &r%%   *   â x-(ww  ÿÀŽ Ä x,&ww  ÿÀŽ ä t*+ÿ  -    â &r+$   %   â t))ÿ  ,    ä x&ÿ   &  â $~%ÿ   ÿ Ž â x$)T  *    â &r+'   +   â x&&ww  ÿÀŽ Ä x'™  ÿ    ä t%ÿ  &    â &r%$   +   ä $r$ÿÿ   ' Ž â x'ÿ   (  ä x&)2v  *    â $~*ÿ   ÿ Ž â x+ÿ   $  â $~)ÿ   ÿ Ž â x''ww  ÿÀŽ â $~(ÿ   ÿ Ž â x$$ww  ÿÀŽ Ä x++ww  ÿÀŽ ä t'*ÿ  '    ä t$) $    ä t(( +    â &r&&   %   â x+T  '    â$~%ÿ   ÿ Ž â x*$ww  ÿÀŽ â $~)ÿ   ÿ Ž â x,(ww  ÿÀŽ â &r&+   &   â x2v  '    Ä t%%ÿ  *    ä t,)ÿ  ,    â &r   &   âx(ÿ   (  ä x&%T  ,    äx*™  ÿ    â &r""   #   â x'ÿ   $  â $~#ÿ   ÿ Ž â x%%2v  ,    â $~$ÿ   ÿ Ž â x((ww  ÿÀŽ â &r)&   ÿ   â x''ww  ÿÀŽ â $r&ÿÿ   * Ž â t##ÿ  (    â &r(%   )   â t$ÿ  '    â $~%ÿ   ÿ Ž â x'ÿ   &  Ä x&&ww  ÿÀŽ â $~$ÿ   ÿ Ž â x''ww  ÿÀŽ ä t%% &    ä x)T  #    ä t$$ '    ä x+ÿ   %  Ä x2v  #    â &r#)   (   â x*%ww  ÿÀŽ â $~)ÿ   ÿ Ž â x($ww  ÿÀŽ â $~'ÿ   ÿ Ž â x%+ww  ÿÀŽ â $~&ÿ   ÿ Ž â x$ÿ   $  â &r   #   â t))ÿ  *    Ä t('ÿ  (    ä t#&ÿ  %    ä x&™  ÿ    â $~ÿ   ÿ Ž â x$$ww  ÿÀŽ ä x+)T  (    äx%)2v  (    Ä x(ÿ   &  â $~'ÿ   ÿ Ž â t$ÿ  $    â $~ÿ   ÿ Ž â x&&ww  ÿÀŽ ä x((ww  ÿÀŽ ä t'' &    ä t (    â $~)ÿ   ÿ Ž â x&#T  $    â$~(ÿ   ÿ Ž â x$#2v  $    Ä x*'ww  ÿÀŽ ä x#ww  ÿÀŽ ä t))ÿ  *    ä t((ÿ  #    ä x™  ÿ    ä x*)T  (    Ä x))2v  (    â $r(ÿÿ    Ž ä &r+      â x'ÿ   '  ä xÿ     â &r%%      â x#ÿ   (  â &r**   ÿ   â x''ww  ÿÀŽ â $~ÿ   ÿ Ž â x,ww  ÿÀŽ â $~ÿ   ÿ Ž â x((ww  ÿÀŽ â $~ÿ   ÿ Ž â x##ww  ÿÀŽ â $~ÿ   ÿ Ž â tÿ  '    â &r*)   *   â tÿ  ,    Ä t (    ä t #    â &r&&   %   â x#T      â$~,ÿ   ÿ Ž â x%ww  ÿÀŽ â $~ÿ   ÿ Ž â x(ww  ÿÀŽ â &r*#   *   â x2v      Ä t,ÿ  %    ä tÿ  (    ä xÿ     ä xÿ     â $~#ÿ   ÿ Ž â xww  ÿÀŽ â $~%ÿ   ÿ Ž â xww  ÿÀŽ â &r*   *   â xT      Ä t#ÿ      ä t%ÿ      â &r   *   â x2v      ä xT      äx    ÿàÿ â &r      â x2v      Ä |    pbð â &r      ä &r$$   &   ä&r      ä 0rÿ    A   ä 0rÿ    A   ä 0rÿ    A   Ä 0rÿ     A   ä 0r!ÿ!    A   â x@  ÿà÷ â  r     A  â Er" "       â  r     A  âEr        â  r     A  âEr$ $       â  r!!     A  â Er        â #r"      â #r      â #r
$   
   â x€   ÿàÿ â #r!      â $rÿÿ   Ž â D	       À â Gy  Ðìÿÿÿÿƒ ê Ay        € ê ˆ  þÿ H   è ˆ €þÿ H   è ˆ 
 ÿÿ H   è ˆ €ÿÿ H   è {          ì M)        € ê „y     H   " z  r  p`ô â $z		 x  Ž Ä „y €   H   b r ÿ   pVt Æ „y    H   ¨ „y     H   è „y €  H   ( „y    H   („y    H   ( „y €  H   ( „y €  H   ( „y    H   ( „y €  H   " !r       Æ „y €  H   " !r       â/ !r
       âO !r         ä� !r       â!r       â !r         â !r       È ‰       b !r       â !r       È ‰      ¢ !r       Æ ‰      â !r       Ê ‰      " !r        â/ !r       ÈO ‰      b !r
       Æ� ‰      ¨ ‰
      â !r       Æ y      '   ( ‰       " !r       â/ !r       ÈO ‰ €    b !r
       Æ� ‰ €    ¨ ‰ €    â !r        â$tÿ   ÿ Ž È ‰ €    " !r       â/ !r       ÈO ‰  @    b !r
       Æ� ‰ @    ¨ ‰
 @    â !r       âx   ÿàÿ È ‰ @    " z  r  p`ò È x    pVò  â !r         â/ !r       âO !r
       È� ‰      b x
   ÿàÿ Æ ‰       ¢ z 
 r  p`ð â !r       âx   ÿàÿ â ‰      â x    pVp  äz  r  p`ö â ‰      " $z ~  	Ž ä x    pVö â !r       â/ !r        âO %v j   Ž â !r       È� †©     â !r       È†¹    è †‰    â M        € ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À z  
      ä y      &   " z ÿ \  pPð â ¹z  F   
   Æ z ÿ ]   Sð Ú         Ê % \   Ž Ê �	    ™ ¢  r ÿ       â Ey  0    € ä y      "   d '
 m   Ž äy      %   ä 'Š s   Ž â 
ÿ n    â y       !   " r ÿ       Ä ÿ   ÿáÿ â y      '   " r ÿ       Æ $
 o  Ž â r ÿ   pRò ä/ x   ÿ   ä� x     pDò  är	    ÿàÿ är    ÿ(Ž ä z 	 r  pdò  Ä Šÿ t    ä xÿ     ä r       äO r   ÿ� Ê G  À    € ê z ÿ `  pPô ä z ÿ ^  pPò ä z ÿ b  p ö ä z ÿ a   Sô ä z ÿ _  Sò ä z ÿ c  0't Ö $ ~  ÿŽ âÿ     â$ª ~  ÿŽ ä $	 x  ÿŽ â    àë â$ª x  ÿŽ â ¢    àÇ ä ¨ÿ     ä ÿ   ÿÄþ Ä 	   ÿàý ä ¢
   ÿàû ä ¢	ÿ   ÿ„þ ä ÿ   ÿä ä  ^  ÿ† ä ¢ÿ	   ÿäÿ Ä ª
 b  ÿˆ ä 	 _  � ä r ÿ       ä r ÿ       ä ª
 c   ä �     h �©     d Ay        € ê ¹z  l      â xÿ     â ™x?    â ‚        â Ey  €'    € â r ÿ       â x      ÿ  â r7 ÿ       â x      ÿ  â |    pbð Ä r9 ÿ       ä r
 ÿ       Ò G	  ð&    € ê r ÿ       â $z w  ÿŽ â xüÿÿÿÿÀŽ ä x   ÿàÿ â 'z y   Ž â r        ä r	  €ÿàÿ â $z }  ÿŽ â zÿ z    â $z p  Ž â x 	   ÿ   Ä r   	à£ â $z |  Ž â x   ÿàÿ â $z |  Ž â z ÿ `  pPð â $z v  Ž âz  `      â $z v  Ž â x   ÿàÿ â %x     ÿŽ â z  a      Ä z  X      â $z p  Ž âz  Y      â $z p  Ž âr ÿ       â $z p  Ž â r7 ÿ       â %x    Ž âr9 ÿ       â x      ÿ  â z ÿ a   Sð â %x    Ž â r
 ÿ       Ä x   ÿ   â %x    Ž â xÿ   	  ä x      ÿ  â %x     Ž â rÿÿ   Dþ  È x    ÿà÷ È rÿ!   ÿäÿ ä r$    ÿàó È r%!   ÿäÿ  Ê �y8$   • ¢  r"   ÿàó ä x# $       â �y6$   • â  rÿ   ÿäÿ  ä r   ÿàó â %v"" Z  # Ž È $x$   ÿŽ Ê r##   ÿàÿ ä r   ÿäÿ  ä z5  p      â �y"   ™ ( �y&"   ™ â %x5   Ž Æ �y'"   ™ ( �y:þÿÿ • ( �y0"   ™ ( �y1"   ™ ( �y2"   ™ ( �y3"   ™ ( �y4"    ™ " %x$   Ž Æ �y;    • ( �y<$    • ( �y="    • " %x(5   Ž Ê �y"(þÿÿ • (# �y#(    • " %x*5   (Ž Ê �y(*þÿÿ • (# �y**    • "# %x,5   $Ž Ì %x.5   ,Ž ä�y,,    • ( %x5   .Ž ä �y..    • ( �y    • " ‚x       Ä |)       â ‚x ÿ       ä |-       ä/ x%8™  ÿ    ÈO x$%ww  ÿÀŽ ä x%ÿ   %  ä |8       ä x%%ww  ÿÀŽ ä t$) $    ä t/8 %    Ä |)       ä x$ww  ÿÀŽ ä x+ÿ   /  ä t)ÿ      ä x$ÿ   $  ä x//ww  ÿÀŽ Ä x)6™  ÿ    ä� |%       ä |8       ä x$$ww  ÿÀŽ ä x++ww  ÿÀŽ ä t6-ÿ  /    Ä x/ÿ   )  ä t%%ÿ  $    ä t88ÿ  +    ä x$T  6    ä |-       ä |+       Ä x))ww  ÿÀŽ ä x//ww  ÿÀŽ â &r5$   ÿ   âx2v  6    ä t)- )    ä t++ /    ä x6%T  8    Ä x%%2v  8    â &r&   5   â |-       ä x8)ww  ÿÀŽ ä |$       ä x/+ww  ÿÀŽ ä x)ÿ   )  Ä t--ÿ  8    ä t$$ÿ  /    â &r/6'      â |       ä x8)ww  ÿÀŽ ä x::™  ÿ    â &r%%0   /   â x6-T  $    Ä x+ÿ   +  ä tÿ  8    ä x8ÿ   :  â &r661   %   â x--2v  $    ä |$       ä x++ww  ÿÀŽ Ä |%       ä x::ww  ÿÀŽ ä |)       ä x88ww  ÿÀŽ ä t$$ÿ  +    ä t%% :    Ä t)) 8    â &r6-2   6   â x/T  $    ä |-       ä x:%ww  ÿÀŽ ä |+       ä x8)ww  ÿÀŽ â &r6/3   6   â x2v  $    Ä t$-ÿ  :    ä t++ÿ  8    ä x-ÿ   %  ä x/ÿ   )  ä |)       ä |%       Ä x:-ww  ÿÀŽ ä x//ww  ÿÀŽ ä x8$T  +    â&r4   6   â t))ÿ  :    ä t6%ÿ  /    â &r88   ÿ   â x+$2v  +    Ä x;;™  ÿ    ä x$)T  6    â &r++&   8   â x%)2v  6    ä r) ;       â 0r<ÿ<    A   ä 0r=ÿ=    A   ä &r$$'   +   â x+ÿ   )  Ä |6       â  r<<=     A  â Er        ä x))ww  ÿÀŽ ä |8       ä x++ww  ÿÀŽ â #r9<   9   â t)6 )    Ä t8 +    â &r$%0   $   â x+ÿ   )  ä x/"™  ÿ    ä |8       ä |-       ä x%)ww  ÿÀŽ Ä x6ww  ÿÀŽ ä |:       ä x)+ww  ÿÀŽ ä xÿ     ä x5ÿ   /  ä t%8ÿ  %    Ä t6-ÿ  6    ä t):ÿ  )    ä x8ww  ÿÀŽ ä |+       ä x:5ww  ÿÀŽ ä |"       Ä x//ww  ÿÀŽ ä x%T  6    äx%%2v  6    ä t6-ÿ  8    ä t-+ :    ä t/" /    Ä |:       ä x5-ww  ÿÀŽ ä |8       ä x+/ww  ÿÀŽ ä x")T  6    äx))2v  6    Ä t6:ÿ  5    ä t+8ÿ  +    ä x5ÿ   /  ä x:ÿ   -  â &r1   $   â |/       ä |-       Ä x<5ww  ÿÀŽ ä x::ww  ÿÀŽ ä x8+T  6    â&r%2      â t//ÿ  <    ä x++2v  6    â &r88   ÿ   â t:-ÿ  :    Ä x##™  ÿ    â &r++&   8   â x$/T  :    â &r"3      â x"ÿ   #  ä |6       ä x##ww  ÿÀŽ â &r$$'   +   â |+       Ä x""ww  ÿÀŽ ä t#6 #    ä t++ "    ä x//2v  :    ä x"ÿ   #  â &r)4      â x-(™  ÿ    â &r$/0   $   â |)       Ä x""ww  ÿÀŽ ä x/ÿ   -  ä |8       ä x##ww  ÿÀŽ ä t")ÿ  "    ä x(ÿ   +  Ä x6+ww  ÿÀŽ ä |)       ä x:-ww  ÿÀŽ ä |+       ä x</ww  ÿÀŽ ä |%       Ä t#8ÿ  #    ä x8(ww  ÿÀŽ ä t)) :    ä t++ <    ä t6%ÿ  6    ät%%ÿ  8    Ä |8       ä x5ÿ   )  ä x-)ww  ÿÀŽ ä |:       ä x/+ww  ÿÀŽ ä x(#T  6    Ä x##2v  6    ä |)       ä x65ww  ÿÀŽ ä t-8ÿ  -    ä t8:ÿ  /    ä x5*™  ÿ    â &r$(1   $   â x+ÿ   +  Ä t()ÿ  6    ä x6-T  8    äx;ÿ   5  ä x--2v  8    ä |/       ä x8+ww  ÿÀŽ Ä |)       ä |*       ä x:5ww  ÿÀŽ ä x;;ww  ÿÀŽ â &r66   ÿ   â t+/ÿ  8    ä t)) :    Ä t** ;    â &r6-&   6   â x/(T  +    ä|:       ä |-       ä x5)ww  ÿÀŽ ä x8*ww  ÿÀŽ â &r6/'   6   â x+(2v  +    Ä t(:ÿ  5    ä t--ÿ  8    ä x)ÿ   )  ä x/ÿ   *  â &r++0   6   â |*       ä x))ww  ÿÀŽ Ä x//ww  ÿÀŽ ä x8(T  -    ät):ÿ  )    ä x-(2v  -    â &r+81   +   â t**ÿ  /    ä x("T  %    Ä x%"2v  %    â &r##2   $   âx")T  *    â&r+-2   +   â x))2v  *    â &r#(3   #   ä0r,ÿ,    A   ä &r""3   +   â Er        â 0r.ÿ.    A   â  r,=,     A  â 0rÿ    A   Ä &r#%4   #   ä&r")4   "   â #r7,   7   â  r%=.     A  â r(=     A  â Er) #       ä Er$ "       â G‰  0    € ê r    ÿàó È r!   ÿäÿ  Ê �y+    ¨  �y*    â  %x"   Ž Ì �y""     " |.       ä |6       ä |       ä |8       ä x,ÿ   +  äO x++ww  ÿÀŽ Ä x-,ww  ÿÀŽ ä t,. +    ä t+6 -    ä x,ww  ÿÀŽ â 0r"ÿ"    A   âx,ÿ   ,  ä |.       Ä x#+ww  ÿÀŽ ä tÿ      ä x,ww  ÿÀŽ ä t,.ÿ  #    ä |-       ä x.T  ,    Ä x#2v  ,    ä x+ÿ   +  â &r..   ÿ   â x,ÿ   *  ä� t-ÿ      â &r.#&   .   â |6       ä |-       Ä |       ä x++ww  ÿÀŽ ä x**ww  ÿÀŽ ä x,,ww  ÿÀŽ ä t+6ÿ  +    ä t#- *    Ä t, ,    ä x*T  +    äx-#ww  ÿÀŽ ä x/,ww  ÿÀŽ â &r**'   .   â x2v  +    ä t6ÿ  -    Ä t+8ÿ  /    â &r*0   *   â x#ÿ   #  â  r="     A  â x,ÿ   ,  ä |.       ä |/       ä x##ww  ÿÀŽ Ä x,,ww  ÿÀŽ ä x-T  +    ät#.ÿ  #    ä t,/ÿ  ,    â &r*-1   *   â x+2v  +    ä x#T  ,    Ä x##2v  ,    â &r*+2   *   È &r3   *   È &r#4      Ê Er        Ê #r
   
   ä G‰  @    € ê r   ÿàó È r   ÿäÿ  Ê �y+    ¨ �y*    â  x#   ÿàÿ Ê $z## p  Ž È %x"#   Ž Ì �y""     " |.       ä |-       ä |6       ä |/       ä |5       ä x,ÿ   +  ÄO x++ww  ÿÀŽ ä x,,ww  ÿÀŽ ä t+. +    ä t- ,    ä |-       ä x+ww  ÿÀŽ Ä x,ÿ   +  ä |.       ä x+ww  ÿÀŽ ä t#-ÿ      â/ 0r"ÿ"    A   âx,ww  ÿÀŽ ä t,.ÿ  +    Ä t-ÿ      â  r"="     A  â x.#T  ,    äx-#2v  ,    ä x#ÿ     â &r..   ÿ   â x+ÿ   *  ä� x##ww  ÿÀŽ â &r--&   .   â |       Ä x**ww  ÿÀŽ ä x,+ww  ÿÀŽ ä t+6ÿ  #    ä t#/ *    ä t, ,    ä x*T  +    Ä x.#ww  ÿÀŽ ä x/,ww  ÿÀŽ â &r**'   -   â x2v  +    ä t5ÿ  .    ä t+6ÿ  /    â &r*0   *   â x#ÿ   #  Ä x,ÿ   ,  ä |.       ä |/       ä x##ww  ÿÀŽ ä x,,ww  ÿÀŽ ä x-T  +    Ä t#.ÿ  #    ä t,/ÿ  ,    â &r*-1   *   â x+2v  +    ä x#T  ,    äx##2v  ,    â &r*+2   *   È &r3   *   È &r#4      Ê Er        Ê #r	"   	   ä G‰  @    € ê r   ÿàó È r   ÿäÿ  Ê �y+    ¨ �y*    â  x#   ÿàÿ Ê $z## p  Ž È %x"#   Ž Ì �y""     " |.       ä |-       ä |6       ä |/       ä |5       ä x,ÿ   +  ÄO x++ww  ÿÀŽ ä x,,ww  ÿÀŽ ä t+. +    ä t- ,    ä |-       ä x+ww  ÿÀŽ Ä x,ÿ   +  ä |.       ä x+ww  ÿÀŽ ä t#-ÿ      â/ 0r"ÿ"    A   âx,ww  ÿÀŽ ä t,.ÿ  +    Ä t-ÿ      ä x.#T  ,    äx-#2v  ,    ä x#ÿ     â &r..   ÿ   â x+ÿ   *  ä� x##ww  ÿÀŽ â &r--&   .   â |       Ä x**ww  ÿÀŽ ä x,+ww  ÿÀŽ ä t+6ÿ  #    ä t#/ *    ä t, ,    ä x*T  +    Ä x.#ww  ÿÀŽ ä x/,ww  ÿÀŽ â &r**'   -   â x2v  +    ä t5ÿ  .    ä t+6ÿ  /    â &r*0   *   â x#ÿ   #  â  r="     A  â x,ÿ   ,  Ä |.       ä |/       ä x##ww  ÿÀŽ ä x,,ww  ÿÀŽ ä x-T  +    ä t#.ÿ  #    Ä t,/ÿ  ,    â &r*-1   *   â x+2v  +    ä x#T  ,    äx##2v  ,    â &r*+2   *   È &r3   *   È &r#4      Ê Er        Ê #r      â #r%)      â #r($      â G‰  @    € ê r   ÿàó È r   ÿäÿ  Ê �y%    ¨  �y$    â  x#   ÿàÿ Ê $z## p  Ž È %x"#   Ž Ì �y""     " |*       ä |)       ä |       ä |+       ä x(ÿ   %  äO x%%ww  ÿÀŽ Ä x((ww  ÿÀŽ ä t%* %    ä t() (    ä x%ww  ÿÀŽ ä |*       ä x#(ww  ÿÀŽ Ä/ tÿ      ä t**ÿ  #    ä x%ÿ   %  â 0r"ÿ"    A   âx,T  *    ä x(ÿ   (  ä |#       â &r,,   ÿ   â xÿ   $  â�  r"="     A  â x%ww  ÿÀŽ Ä x((ww  ÿÀŽ ä x2v  *    ä |*       ä x$$ww  ÿÀŽ â &r&   ,   â x%ww  ÿÀŽ ä t#ÿ      Ä t#ÿ  (    ä t#) $    ä t%* %    ä x$T      ä |)       ä x&#ww  ÿÀŽ â &r$'      â x(%ww  ÿÀŽ Ä x#ÿ   #  ä x2v      ä t&)ÿ  &    ä t'+ÿ  (    â &r00      â x%ÿ   %  ä |       Ä x##ww  ÿÀŽ ä |(       ä x%%ww  ÿÀŽ ä x$&T  '    ät#ÿ  #    ä x'&2v  '    â &r0$1   0   â t(ÿ  %    Æ &r''2   0   â x$#T      äx##2v      Æ &r$$3   '   È &r##4   $   Ê Er# #       Ê #r"#      â x    ÿàÿ ä x@  ÿàõ ä |    pbò ä x@  ÿà÷ ä x€   ÿàÿ ä rÿ   ÿä Ä rÿ   ÿäÿ Ê D       À â Gy  ÀÛÿÿÿÿƒ ê Ay        € ê y      "   " Ey  0    € â r ÿ   pRð är ÿ   pRò Ö G‰  ð     € ê z ÿ `  pPð ä r    ÿ8Ž ä z ÿ a   Sð Æ ˆs 9 þÿ H   ô ˆƒ 7€þÿ H   è ˆƒ  ÿÿ H   è ˆƒ €ÿÿ H   â G‰  p     € ê ˆs 
   H   è ˆs 7€þÿ H   è ˆs 	€  H   è ˆs  ÿÿ H   è ˆs    H   è ˆs €ÿÿ H   è ˆs €  H   ä Ay        € ê {          ì M        € ê y      &   ¢ z ÿ `  pPð Æ y      '   "> z ÿ a   Sð â $z x  Ž âO x        Æ $z ~  Ž È %v j   Ž È G	  P    € ê „y      H   "z  r  p`ð Æ „y     H   b r  ÿ   pVp  Æ „y     H   ¨ „y  €   H   è „y
     H   ( „y	     H   ( „y  €  H   ( „y     H   ( „y  €  H   ( „y  €  H   ( „y  €  H   " !r9       Æ „y  €  H   " !r       â/ !r       âO !r7       ä� !r

       â!r	       â !r       È ‰      b !r

       â !r       È ‰
      ¢ !r       Æ ‰	      â !r       È !r       â !r       È/ ‰      ( ‰      b !r
       âO !r		       È� ‰      ¨ ‰	      â !r       â !r       È/ ‰      " !r       ÆO ‰ €    b !r	       Æ� ‰ €    ¨ ‰ €    â !r       Ê ‰ €    " !r       â/ !r       ÈO ‰ @    b !r       â� x   ÿàÿ ä ‰ @    ¤ z  r  p`ô ä ‰
 @    ä x     pVt â !r       Ê ‰ @    " !r       â/ x   ÿàÿ äx   ÿàÿ â !r       âO z  r  p`ò â ‰      b !r

       ä� x     pVò  â ‰      è ‰	
      ¢ !r       â !r       È/ ‰      " !‚       â !r
	       ÈO †‰     â z  r  p`ð â !r	       â !’       äx     pVp  â !¢		       ä †™    è †© 	   â !r       Ê M	        € ê !r       Ê †y    â My        € ê z  h      È x    pBð Ú G	  @    € ê x    pRð Ú G‰  À    € ê x    pRð Ú G	   $    € ê „y      H   " z  r  p`ð â Ey  p    € ä „y     H   b r  ÿ   pVp  Æ „y     H   ¨ „y     H   è „y     H   ( „y   
  H   ( „y  €   H   ( „y  €  H   ( „y  €  H   è „y  €  H   " !r9       â !r
       Æ/ !r       âO !r       Æ� !r       ä„y  €  H   â !r       Æ ‰
      " !r7       Æ ‰      b !r       È !r       â !r

       â !r       È/ ‰
      ( ‰      b !r	       È !r       â !r
       â !r       È/ ‰ €    ( ‰ €    b !r       Æ „y  €
  H   â !r       Æ/ ‰ @    ( ‰ @    b !r       â !r       È/ ‰
      ( ‰      b !r       â !r	
       â !r       â/ G	  �     € ê !r
       â!r		       Æ  x
;ª¸¿  A  Ì s        $ !t  €?    Ì s        $  r

     A  È  r		
     A  Ê †y 	    ä Ay        € ê ‰	      " Ey  Ð    € æ ‰
      h „y     H   ¨ „y     H   è „y     H   è „y     H   â !r		       Æ „y     H   â !r
       Æ/ ‰
	      ( ‰      b !r       âO !r       â� !r
	
       ä !r       â !r       Ä/ ‰
 €    " !r       Æ ‰ €    h „y   	  H   ¢ !r
       â !r       È/ ‰ @    ( ‰ @    b !r       ÈO !r       â !r       â x   ÿàÿ â !r       Æ/ ‰	      " z  r  p`ð Æ ‰
      b x     pVp  â !r	       â !r
       Ö/ G	  �     € ê !r       â!r       Æ  x	;ª¸¿  A  Ì s	 	       $ !t
	  €?    Ì s 
       $  r     A  È  r     A  Ê †y    ä Ay        € ê ‰      b Ey       € æ ‰	      ¨ „y  €  H   è „y  €	  H   ( „y  €  H   è „y  €  H   â !r       Æ/ „y  €  H   â !r
	       ÆO ‰	      h ‰
      " !r       È� !r       â!r		       â/ !r
       È ‰	 €    ( ‰ €    h „y
  €  H   ¢ !r	       â !r       È/ ‰ @    ( ‰ @    b !r

       âO !r       â x   ÿàÿ â !r       Æ/ ‰      " z  r  p`ð Æ ‰	      b x     pVp  â !r       â !r	       Ö/ G	  �     € ê !r       â!r       Æ  x;ª¸¿  A  Ì s        $ !t	  €?    Ì s 	       $  r     A  È  r     A  Ê †y    ä Ay        € ê ‰
      " !r       â x   ÿàÿ Æ !r       â z  r  p`ð Æ !r       â x     pVp  È ‰      b !r	
       Ê ‰	      " !r       Ê/ ‰      b !r	       Ê ‰ €    " !r       Ê/ ‰ €    b !r
       Ê ‰
 @    " !r       Ê/ ‰ @    b !r
       Ê ‰      " !r       Ê/ ‰      ¢ !r       â M	        € ì !r       â/!r       ÆO  x ;ª¸¿  A  â !r       Ê s          $ !t   €?    Ì s        $  r     A  È  r     A  Ê †y    â My        € ê „y      H   " z  r  p`ð â Ey       € ä „y     H   b r  ÿ   pVp  Æ „y     H   ¨ „y     H   è „y     H   ( „y   
  H   ( „y  €  H   ( „y  €  H   è „y  €  H   ( „y  €  H   â !r9       â !r
       Æ/ !r       âO !r       Æ� !r       â!r       È ‰
      " !r	       Æ ‰      b !r       â !r

       â !r       È/ ‰
      ( ‰      h „y  €   H   ¢ !r
       â !r       È/ ‰ €    ( ‰ €    b !r7       ÈO !r       â !r       ä „y  €
  H   â !r       â !r       ä/ ‰ @    ( ‰ @    b !r       â !r       È/ ‰
      ( ‰      b !r       â !r
       â !r	       â/ G	  À     € ê !r		       È x	'7=  A  â x
	*BL?  A  Æ #t	  €?   È  r
     A  â  x
	   ?  A  â !r	       È s     $   $ !t  €?    È  r

     A  È  r		
     A  Ê †y 	    ä Ay        € ê ‰	      " Ey       € æ ‰
      h „y     H   ¨ „y   	  H   è „y     H   è „y     H   â !r		       Æ „y     H   â !r
       Æ/ ‰
	      ( ‰      h „y     H   â !r       âO !r
	
       â !r       È/ ‰
 €    ( ‰ €    b !r       È !r       â !r       È� !r       â !r
       â x
   ÿàÿ ä !r       â !r       â/ ‰ @    " z 
 r  p`ð Æ ‰ @    b x     pVp  â !r       â !r       È/ ‰	      ( ‰      b !r	       â !r       â/ G	  À     € ê !r       È x	'7=  A  â x*BL?  A  Æ #t	  €?	   È  r		     A  â  x   ?  A  â !r       È s	 	    $   $ !t	  €?    È  r     A  È  r     A  Ê †y    ä Ay        € ê ‰      b Ey  Ð    € æ ‰      ¨ „y  €  H   è „y  €  H   è „y  €  H   è „y  €  H   â !r       Æ/ „y  €	  H   â !r
       ÆO ‰	      h ‰
      " !r       â� !r		       â/ !r
       È ‰	 €    ( ‰ €    b !r       Æ „y
  €  H   ¢ !r	       â !r       È/ ‰ @    ( ‰ @    b !r
       âO !r       â x   ÿàÿ â !r       Æ/ ‰      " z  r  p`ð Æ ‰	      b x     pVp  â !r       â !r		       Ö/ G	  À     € ê !r	       È x	'7=  A  â x*BL?  A  Æ #t  €?	   È  r     A  â  x   ?  A  â !r       È s     $   $ !t
  €?    È  r
     A  È  r     A  Ê †y    ä Ay        € ê ‰      b !r       â x   ÿàÿ Æ !r       â z  r  p`ð Æ !r       â x     pVp  È ‰      ¢ !r	       Ê/ ‰	      b !r       ÊO ‰      ¢ !r
	       Ê/ ‰
 €    b !r       ÊO ‰ €    " !r
       Ê/ ‰ @    b !r       Ê ‰ @    " !r       Ê/ ‰	      b !r       Ê ‰      ¢  !r		       â/ M	        € ì !r	       È x'7=  A  â x *BL?  A  Æ #t  €?   â x   ?  A  Æ  r       A  â !r       ÈO !r       â s       $   $ !t   €?    È  r     A  È  r     A  Ê †y    â My        € ê x    pRð Ú G‰  P    € ê x    pRð Ú G	  `    € ê „y      H   " z  r  p`ð â Ey       € ä „y     H   b r  ÿ   pVp  Æ „y     H   ¨ „y     H   è „y     H   ( „y   
  H   ( „y  €  H   ( „y  €  H   è „y  €  H   ( „y  €  H   â !r9       â !r
       Æ/ !r       âO !r       Æ� !r       â!r       È ‰
      " !r	       Æ ‰      b !r       â !r

       â !r       È/ ‰
      ( ‰      h „y  €   H   ¢ !r
       â !r       È/ ‰ €    ( ‰ €    b !r7       ÈO !r       â !r       ä „y  €
  H   â !r       â !r       ä/ ‰ @    ( ‰ @    b !r       â !r       È/ ‰
      ( ‰      b !r       â !r
       â !r	       â/ G	  À     € ê !r		       Ê	z
	 i    � â !r	       È  x
;ª¸¿  A  â 	z		 i    � È 	z		 i €  � â s        $ !t  €?    Ì s        $  r

     A  È  r		
     A  Ê †y 	    ä Ay        € ê ‰	      " Ey       € æ ‰
      h „y     H   ¨ „y   	  H   è „y     H   è „y     H   â !r		       Æ „y     H   â !r
       Æ/ ‰
	      ( ‰      h „y     H   â !r       âO !r
	
       â !r       È/ ‰
 €    ( ‰ €    b !r       È !r       â !r       È� !r       â !r
       â x
   ÿàÿ ä !r       â !r       â/ ‰ @    " z 
 r  p`ð Æ ‰ @    b x     pVp  â !r       â !r       È/ ‰	      ( ‰      b !r	       â !r       â/ G	  À     € ê !r       Ê	z i    � â !r       È  x	;ª¸¿  A  â 	z i    � È 	z i €  � â s	 	       $ !t
	  €?    Ì s 
       $  r     A  È  r     A  Ê †y    ä Ay        € ê ‰      b Ey  Ð    € æ ‰      ¨ „y  €  H   è „y  €  H   è „y  €  H   è „y  €  H   â !r       Æ/ „y  €	  H   â !r
       ÆO ‰	      h ‰
      " !r       â� !r		       â/ !r
       È ‰	 €    ( ‰ €    b !r       Æ „y
  €  H   ¢ !r	       â !r       È/ ‰ @    ( ‰ @    b !r
       âO !r       â x   ÿàÿ â !r       Æ/ ‰      " z  r  p`ð Æ ‰	      b x     pVp  â !r       â !r		       Ö/ G	  À     € ê !r	       Ê	z i    � â !r       È  x;ª¸¿  A  â 	z i    � È 	z i €  � â s        $ !t	  €?    Ì s
 	       $  r
     A  È  r     A  Ê †y    ä Ay        € ê ‰      b !r       â x   ÿàÿ Æ !r       â z  r  p`ð Æ !r       â x     pVp  È ‰      ¢ !r	       Ê/ ‰	      b !r       ÊO ‰      ¢ !r
	       Ê/ ‰
 €    b !r       ÊO ‰ €    " !r
       Ê/ ‰ @    b !r       Ê ‰ @    " !r       Ê/ ‰	      b !r       Ê ‰      ¢  !r		       â/ M	        € ì !r	       â!r       ÈO 	z i    � â !r       È  x ;ª¸¿  A  â 	z i    � È 	z i €  � â s          $ !t   €?    Ì s        $  r     A  È  r     A  Ê †y    â My        € ê „y      H   " z  r  p`ð Æ „y     H   b r  ÿ   pVp  Æ „y  €   H   ¨ „y  €  H   è „y     H   ( „y     H   ( „y     H   ( „y     H   ( „y  €  H   ( „y  €  H   ( „y"     H   " !r99       Æ „y$   	  H   " !r
       Æ/ „y     H   b !r7       ÆO „y   
  H   ¢ !r	       Æ� „y  €  H   â !r       Æ„y   €
  H   " !r       Æ „y&     H   " !r9       Æ „y(     H   " !r       Æ „y  €  H   " !r       Æ „y  €  H   " !r       Æ „y  €  H   " !r"       Æ „y  €	  H   â !r$       Æ „y  €  H   â !r       Æ/ „y  €  H   â !r       âO !r       ä� ‰      " !r        Æ‰      b !r&&       Æ ‰	      ¢ !r(       Æ ‰
      â !r       Æ ‰&      " !r       Æ ‰      " !r       â !r       â !r       â !r       Æ/ ‰      " !r       â !r		       äO ‰      b !r       â !r
       ä� ‰      ¢ !r&       Æ‰      â !r       Æ ‰
	      ( ‰      ( ‰      " !r       Æ ‰      " !r       â/ !r       äO ‰ €    b !r       Æ� ‰      ¢ !r
	
       Æ‰      â !r       Æ ‰ €    " !r       Æ ‰	
 €    " !r       Æ ‰ €    ( ‰ €    " !r       Æ/ ‰ €    b !r       âO !r       ä� ‰ @    ¢ x   ÿàÿ â !r       ä‰ €    â z  r  p`ò â !r	
	       Ä ‰ €    " x     pVò  â !r       ä ‰ @    " !r       Æ ‰
	 @    " !r       Æ/ ‰ @    h ‰ @    b !r       ÆO ‰ @    ¢ !r       â� !r       ä‰      â x   ÿàÿ â !r       ä ‰ @    " z  r  p`ô â !r
	
       Ä ‰ @    " x     pVt â !r       ä/ ‰      b x   ÿàÿ â !r       ä ‰	
      b z  r  p`ö â !r       ÄO ‰      ¢ x     pVö Æ ‰      ¢ !r       Æ� ‰      â !r       â !‚       â!r       Æ‰      " !r       Æ/ ‰      b !r
	       â !‚       â !r       äO !’       â  ‚     A  â !r       â !’       Æ †‰     â !r       â� !¢       â  ’     A  ä !¢
       â !r       ä †™    â  ¢
     A  â !r       È/ †©    â M9        € ê !r       â !r       ÈO  r     A  Ê †y    â My        € ê „y      H   " z  r  p`ð â Ey  À    € ä „y     H   b r  ÿ   pVp  Æ „y     H   ¨ „y     H   è „y     H   ( „y   
  H   ( „y  €  H   ( „y  €  H   è „y  €  H   ( „y  €  H   â !r9       â !r
       Æ/ !r       âO !r       Æ� !r       â!r       È ‰
      " !r	       Æ ‰      b !r       â !r

       â !r       È/ ‰
      ( ‰      h „y  €   H   ¢ !r
       â !r       È/ ‰ €    ( ‰ €    b !r7       ÈO !r       â !r       ä „y  €
  H   â !r       â !r       ä/ ‰ @    ( ‰ @    b !r       â !r       È/ ‰
      ( ‰      b !r       â !r
       â !r	       â/ G	  à     € ê !r		       Ê	x
	  à@  � Ê  x	
#ÛÙ?  A  È  x	;ª¸¿  A  â !r	       Ê s        " 	x		  à@  � È 	x		  àÀ  � Ê !t		  €?    â !t  €?    È s        $  r

     A  È  r	
	     A  Ê †y 	    ä Ay        € ê ‰	      " Ey       € æ ‰
      h „y     H   ¨ „y   	  H   è „y     H   è „y     H   â !r		       Æ „y     H   â !r
       Æ/ ‰
	      ( ‰      h „y     H   â !r       âO !r
	
       â !r       È/ ‰
 €    ( ‰ €    b !r       È !r       â !r       È� !r       â !r
       â x
   ÿàÿ ä !r       â !r       â/ ‰ @    " z 
 r  p`ð Æ ‰ @    b x     pVp  â !r       â !r       È/ ‰	      ( ‰      b !r	       â !r       â/ G	  à     € ê !r       Ê	x  à@  � Ê  x#ÛÙ?  A  È  x	;ª¸¿  A  â !r       Ê s	 	       " 	x  à@  � È 	x  àÀ  � Ê !t  €?    â !t
	  €?    È s 
       $  r     A  È  r     A  Ê †y    ä Ay        € ê ‰      b Ey  ð    € æ ‰      ¨ „y  €  H   è „y  €  H   è „y  €  H   è „y  €  H   â !r       Æ/ „y  €	  H   â !r
       ÆO ‰	      h ‰
      " !r       â� !r		       â/ !r
       È ‰	 €    ( ‰ €    b !r       Æ „y
  €  H   ¢ !r	       â !r       È/ ‰ @    ( ‰ @    b !r
       âO !r       â x   ÿàÿ â !r       Æ/ ‰      " z  r  p`ð Æ ‰	      b x     pVp  â !r       â !r		       Ö/ G	  à     € ê !r	       Ê	x  à@  � Ê  x#ÛÙ?  A  È  x;ª¸¿  A  â !r       Ê s        " 	x  à@  � È 	x  àÀ  � Ê !t  €?    â !t	  €?    È s
 	       $  r
     A  È  r     A  Ê †y    ä Ay        € ê ‰      b !r       â x   ÿàÿ Æ !r       â z  r  p`ð Æ !r       â x     pVp  È ‰      ¢ !r	       Ê/ ‰	      b !r       ÊO ‰      ¢ !r
	       Ê/ ‰
 €    b !r       ÊO ‰ €    " !r
       Ê/ ‰ @    b !r       Ê ‰ @    " !r       Ê/ ‰	      b !r       Ê ‰      ¢  !r		       â/ M	        € ì !r	       â!r       ÈO 	x  à@  � â !r       È  x #ÛÙ?  A  â 	x  à@  � Æ  x  ;ª¸¿  A  â 	x  àÀ  � Ê s          " !t  €?    â !t   €?    Ì s        $  r     A  È  r     A  Ê †y    â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y       "   $ z   w  p`ð Ú M	        € ê y
      !   " ¹z  l      â Ey  P    € â ™x?    â y      %   b ¹z  F   
   â $rÿÿ   ÿ Ž ä $rÿÿ   ÿ Ž â xÿ   
  È |    p`ð â $x   ÿ Ž Ø/ G	  °    € ê y      &   " $tÿ   ÿ Ž ä $z  x  Ž È %v \   Ž Ì �y    ™ ¢ $rÿÿ   ÿ Ž â x	   ÿàÿ â $z  r  ÿŽ â x
   ÿÀŽ â 'z m   Ž È $z q  Ž â zÿ n    â $z t  Ž äO$z t  Ž ä $z	 q  Ž ä $rÿÿ   
Ž ä %x   ÿŽ È %x   ÿŽ â z X  ÿàý Æ $z o  Ž â z X  ÿàõ â %x    Ž â z Y  ÿä ä z	 Y  ÿä â %x    Ž â z X  ÿàñ Æ $z u  ÿŽ â z X  ÿàó ä x   ÿàù ä x   ÿàû ä r   àµ â $rÿÿ   ÿ Ž â zÿ Y  ~ â $x   ÿ Ž â z	ÿ Y  $þ Ä rÿÿ   ÿd~ ä $rÿÿ    Ž Ê �y$þÿÿ • ¢  r   ÿàñ â$tÿ$   ÿ Ž ä �y#    • â  r   ÿ  â %v Z   Ž È $x$   ÿŽ È $x   Ž Ê �y   ™ ( �y   ™ h �y   ™ â $rÿÿ   	 Ž Æ �y   ™ è �y"þÿÿ • è  �y   ™ h �y    • b $rÿÿ    Ž Ä $r	ÿÿ    Ž â �y   ™ h �y    • h �y   ™ h �y    • b $rÿÿ    Ž Ä $r	ÿÿ    Ž â �y!    ™ h �y     • b ‚x       ä $~'ÿ   ÿ Ž ä $~&ÿ   ÿ Ž â ‚x ÿ       Ä $~	ÿ   ÿ Ž ä $~ÿ   ÿ Ž ä $~(ÿ   ÿ Ž ä $~)ÿ   ÿ Ž ä $~+ÿ   ÿ Ž â x�   ÿàñ ä x   ÿàÿ Ä x�   ÿà÷ ä x�   ÿàó ä x�   ÿàõ â $rÿÿ   Ž â x    ÿàÿ Æ $rÿÿ    â x$$™  ÿ    ÈO x%ÿ   $  ä x$$ww  ÿÀŽ ä x%%ww  ÿÀŽ ä t$' $    ä t%& %    ä x$ww  ÿÀŽ Ä/ x%ww  ÿÀŽ ä t	ÿ      ä tÿ      ä x$ÿ   $  ä x%ÿ   %  â $~&ÿ   ÿ Ž â xT      Ä x$$ww  ÿÀŽ ä x%%ww  ÿÀŽ ä x2v      â &r   ÿ   ât$	ÿ  $    ä t%&ÿ  %    ä x##™  ÿ    â� &r      âx$T  %    Æ $rÿÿ   # Ž ä &r      Æ xÿ     â $~&ÿ   ÿ Ž â xww  ÿÀŽ â $~#ÿ   ÿ Ž â xww  ÿÀŽ ä x$$2v  %    ä t&     ä t#     â &r	$      â x$"™  ÿ    â $~%ÿ   ÿ Ž â xÿ     â $~&ÿ   ÿ Ž â x"ww  ÿÀŽ Ä xww  ÿÀŽ ä x#ww  ÿÀŽ ä t%%ÿ  "    ä t&ÿ      ä x"ÿ   $  ä x$$ww  ÿÀŽ Ä t(ÿ  #    ä x""ww  ÿÀŽ ä x#T  %    äx2v  %    ä t%' $    ä t") "    â $~)ÿ   ÿ Ž â x$%ww  ÿÀŽ Ä xÿ     ä x("ww  ÿÀŽ ä t$)ÿ  $    ä x'ww  ÿÀŽ ä x)ÿ   %  ä t%+ÿ  (    Ä t&ÿ  '    â &r	#   	   â x&$T  %    â $~(ÿ   ÿ Ž â x#™  ÿ    ä x))ww  ÿÀŽ â &r&   ÿ   â x"ÿ   "  â $~'ÿ   ÿ Ž â x&ÿ   #  Ä t(ÿ  )    â $~(ÿ   ÿ Ž â x$$2v  %    â $~)ÿ   ÿ Ž â x""ww  ÿÀŽ ä x%#ww  ÿÀŽ ä x&&ww  ÿÀŽ â &r$      â t#'ÿ  "    Ä t( %    ä t&) &    â $~(ÿ   ÿ Ž â x"T  #    ä x$ww  ÿÀŽ ä x%&ww  ÿÀŽ â &r"      â x2v  #    Ä t'ÿ  $    ä t"(ÿ  %    ä xÿ     ä x&ÿ   &  â $~$ÿ   ÿ Ž â xww  ÿÀŽ â $~%ÿ   ÿ Ž â x&&ww  ÿÀŽ â &r      â x#T  "    Ä t$ÿ      ä x"2v  "    â &r#      â t&%ÿ  &    â &r   	   âxT      â &r"      â xT  &    ä x2v      â $rÿÿ     â x&2v  &    â &r      â|    p`ð â &r      Ä 0rÿ    A   ä 0rÿ    A   ä &r!      ä&r&!      ä 0r ÿ     A   â  r     A  â Er        Ä Er        â  r      A  â $r	ÿÿ   Ž  â #r      ä #r      â G‰  �õÿÿÿÿƒ ê Ay        € ê ‰      " $x   
Ž Æ ‰      d z  p  p`ð È x 
   pDp  â !r       â !r       È/ ‰      ( ‰      b !r       â !r       È/ ‰ €    ( ‰ €    b !r       â !r       È/ ‰ @    ( ‰	 @    b !r       â !r		       È/ ‰      h  ‰	      ¢  M	        € ê y      &   â $x

   ÿ Ž â !r       â/ $z   s  Ž â !r		       â_ $tÿ   ÿ Ž â x 
   p"ò Ä r 
ÿ   p"ð Ö $ÿÿ   	 Ž ä $z  v   Ž È� %v  j   Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À $vÿ 
  ÿ Ž ä y      "   $ z  w  p`ð Ú M	        € ê y	      &   " $t'ÿ   ÿ Ž â ¹z  F   
   â $z& x  	Ž È %v&& \  ' Ž Ì �y&&    ™ b ¹z  l      â Ey  Ð    € â ™x?    â y       !   b x      ÿ  â x      ÿ  â x      ÿ  â x
ÿ      È/ | 
   p`ð Ú G	  @    € ê $rÿÿ   ÿ Ž â y      %   " $z r  ÿŽ â x    ÿÀŽ â 'z	 m   Ž â z ÿ `  pPð â Ey °    € æ zÿ n    ä z ÿ a   Sð Æ $rÿÿ   
Ž ä $rÿÿ   ÿ Ž ä $z o  	Ž ä $x	    ÿ Ž ä $z u  ÿŽ Æ x		   ÿÀŽ ä r   à£ â $rÿÿ   
 Ž ä $x   ÿ Ž â r
ÿÿ   ÿDþ  ä $z& t  Ž ä$tÿ   ÿ Ž ä $z q  Ž È %v X  Ž È $x   ÿ Ž ä $tÿ$   ÿ Ž â z q  ÿàÿ â %x	    Ž â �y    • ¨  �y"   • â r   ÿàó â %v X  Ž Ä �y#   • " r
   ÿ�  â %v Z   Ž ä �y    • ¤  $x$   ÿŽ È $x   Ž ä %x 	    Ž Æ �y   ™ ¨  �y   ™ ¨  �y$    • ¨  �y   ™ ¨" �y   ™ ¨ �y   ™ ¨ �y   ™ ¨ �y    • ¨  �y   ™ ¨ �y    ™ ¨ �y    • ¢ ‚x       Ä $~(ÿ   ÿ Ž ä $~ÿ   ÿ Ž â ‚x ÿ       ä $~ÿ   ÿ Ž ä/ $~ÿ   ÿ Ž ä $~ ÿ   ÿ Ž â x   ÿàÿ È |    p`ò ä x""™  ÿ    È� x%"ww  ÿÀŽ ä x"ÿ   "  ä t!( %    ä x""ww  ÿÀŽ ä x!ww  ÿÀŽ ä t" "    Ä tÿ      ä xÿ   "  ä x""ww  ÿÀŽ â $~ÿ   ÿ Ž â xÿ   !  ä xww  ÿÀŽ ä tÿ  "    Ä xww  ÿÀŽ ä x#™  ÿ    ätÿ      ä xT      â $~%ÿ   ÿ Ž â t  ÿ      ä xÿ     Ä xww  ÿÀŽ â &r   ÿ   âO x2v      ä t%%     â $~"ÿ   ÿ Ž â x T      â&r!      â xww  ÿÀŽ â $~ÿ   ÿ Ž â x  2v      Ä x%ww  ÿÀŽ ä x$$™  ÿ    ä t""     ä tÿ      â $rÿÿ   $ Ž â x"ww  ÿÀŽ â $~ÿ   ÿ Ž ä &r!   !   â xÿ     Ä tÿ      â $~ÿ   ÿ Ž â xww  ÿÀŽ â $~ÿ   ÿ Ž â xww  ÿÀŽ â &r!    !   â x T      ä t     ä t     Ä x2v      â &r    !   â x$ÿ   "  â $~ ÿ   ÿ Ž â x#ww  ÿÀŽ â $~!ÿ   ÿ Ž â x"ww  ÿÀŽ ä t  ÿ  #    ä t!!ÿ  "    â &r      â x"ÿ     Ä x T  !    äx  2v  !    â $~!ÿ   ÿ Ž â xÿ     ä x%ÿ   %  â $~ÿ   ÿ Ž â x""ww  ÿÀŽ â $~ÿ   ÿ Ž â x#™  ÿ    â &r   ÿ   â xww  ÿÀŽ Ä x%%ww  ÿÀŽ ä t!ÿ  "    â &r!       â x"ÿ   #  â $~ÿ   ÿ Ž â tÿ      â $~ ÿ   ÿ Ž â tÿ  %    ä x%#ww  ÿÀŽ Ä x#"ww  ÿÀŽ ä xT      ät %    ä t   #    â $~#ÿ   ÿ Ž â x2v      â &r   !   â x"ww  ÿÀŽ â $~!ÿ   ÿ Ž â x ww  ÿÀŽ â &r      â t#ÿ  "    Ä xÿ     ä t!!ÿ      ä x"ÿ      â $~ÿ   ÿ Ž â x%ww  ÿÀŽ â $~ÿ   ÿ Ž â x$$ww  ÿÀŽ ä x""ww  ÿÀŽ Ä xT  !    ä tÿ  $    ä tÿ  %    â &r      â t##ÿ  "    ä x2v  !    ä x T      Ä xT  #    â &r      â x2v      â &r       âx2v  #    â &r      ä &r      ä 0rÿ    A   ä 0rÿ    A   Ä 0rÿ    A   ä &r      â  r     A  â Er        â  r     A  ä Er        â G‰  @    € ê xÿ     â %v `   Ž È $x   ÿŽ È $x   Ž È %x	    Ž ä �y     ¨  �y    è �y!    " $~#ÿ   ÿ Ž Ä $~"ÿ   ÿ Ž ä $~ÿ   ÿ Ž ä $~ÿ   ÿ Ž â/ xÿ     ä� xww  ÿÀŽ ä xww  ÿÀŽ ä t #     Ä t"     ä x" ww  ÿÀŽ â $~#ÿ   ÿ Ž â x ÿ      â $~ÿ   ÿ Ž â xww  ÿÀŽ ä x  ww  ÿÀŽ ä t"#ÿ  "    Ä tÿ      ä tÿ       ä x$ÿ     ä x ÿ   !  äx("T      â$~ÿ   ÿ Ž â x""2v      â $~ÿ   ÿ Ž â x!!ww  ÿÀŽ Ä x$$ww  ÿÀŽ ä x  ww  ÿÀŽ â &r(   ÿ   â t !    ä tÿ  $    ä t      â &r"      â x ww  ÿÀŽ â $~"ÿ   ÿ Ž â x$T      Ä x!ww  ÿÀŽ ä t #ÿ       â &r$      â x2v      ä t!"ÿ  !    ä x#ÿ     ä xÿ     â $~ÿ   ÿ Ž â x##ww  ÿÀŽ â &r      â xww  ÿÀŽ Ä x T  !    ät""ÿ  #    ä tÿ      â &r      â x  2v  !    ä x"T      Æ &r       â x""2v      Æ &r      ä 0rÿ    A   äO &r"      Ê Er        â  r     A  È #r      â #r      â #r      â G‰       € ê %v `  Ž Ì %x	    Ž ä �y     ¨  �y    è �y    " $~ÿ   ÿ Ž Ä $~ÿ   ÿ Ž ä $~ÿ   ÿ Ž ä $~ÿ   ÿ Ž ä/ $~ÿ   ÿ Ž ä $~ÿ   ÿ Ž ä $~ ÿ   ÿ Ž â xÿ     Ä� xww  ÿÀŽ ä xww  ÿÀŽ ä t     ä t     ä xww  ÿÀŽ ä xww  ÿÀŽ Ä xÿ     ä tÿ      ä tÿ      ä xww  ÿÀŽ ä xT      ä xÿ     Ä tÿ      ä xÿ     â&r   ÿ   â xww  ÿÀŽ ä xww  ÿÀŽ ä xww  ÿÀŽ ä x2v      Ä tÿ      ä t     ä t      â &r      â xT      â$~ÿ   ÿ Ž â xww  ÿÀŽ â $~ÿ   ÿ Ž â xww  ÿÀŽ â &r      â x2v      Ä tÿ      ä tÿ      ä xÿ     ä xÿ     â $~ÿ   ÿ Ž â xww  ÿÀŽ â $~ÿ   ÿ Ž â xww  ÿÀŽ â &r      â xT      Ä tÿ      ä x2v      â &r      â tÿ      Æ &r      â xT      äx2v      Æ &r      È &r      ä 0rÿ    A   ÆO Er        ä  r     A  È #r      ä D       À â Gy  0íÿÿÿÿƒ ê y           È Ay       € ê $rÿÿ    Ž ä $rÿÿ    Ž ä Ay        € ê z ÿ `  pPð È z ÿ a   Sð Ú G	  P    € ê ‰      ( ‰      b !r       â !r	       È/ ‰      ( ‰	      b !r       â !r
	       È/ ‰ €    ( ‰
 €    b !r       â !r
       È/ ‰ @    ( ‰ @    b !r       â !r       È/ ‰      ( ‰	      b !r       â !r	       â/ Gy  €    € ê ‰      ( ‰      h ‰      ¨ ‰	      â !r       â !r       â/ !r
       ÈO ‰      " !r	       Æ� ‰	      h ‰
      ¨ ‰      â !r       â !r		       È/ ‰ €    " !r
       ÆO ‰	 €    b !r       Æ� ‰ €    ¨ ‰ €    â !r       â !r	       â/ !r       ÈO ‰ @    " !r       Æ� ‰ @    h ‰ @    ¨ ‰ @    â !r       â !r       È/ ‰      " !r       ÆO ‰
      b !r	       Æ� ‰      ¨ ‰	      â !r       â !r
       â/ !r       âO !r	       â� y      %   $ $x    Ž Ê z  p  p`ð È x     pDp  Ú M	        € ê z ÿ ^  pPð â $z&& v  Ž Æz ÿ _   Sð Ú $ÿ   ÿ Ž È %& ^   Ž Ì �	     ¢ $x    ÿ Ž Ê r ÿ   p"ò äx    p"ö äx    p"ø äx    p"ô Î $ÿÿ    Ž â z ÿ `  pPò â $2ÿÿ    Ž ä $Bÿÿ    Ž â z ÿ a  Sò â $"ÿÿ    Ž È !       ÐO G™       € ê z ÿ b  pPð È z ÿ c   Sð Ú $ÿ   ÿ Ž È %& b   Ž Ì �	     ¢ $x     ÿ Ž Ê x  øÿÿÿp"ò äx  üÿÿÿp"ô är  ÿ   p"ö äx     p"ø Î $ ÿÿ    Ž ä $" ÿÿ    Ž ä $2 ÿÿ    Ž ä $vÿ h  ÿ Ž ä $B ÿÿ    Ž Æ x    pBò â !         ØO G  P    € ê x    pRð Ú G‰  �     € ê x    pRð Ú G	  ð    € ê  x ;ª¸¿  A  Ì s        $ !t  €?    Ì s        $  r      A  È  r     A  â Gy  P    € ê  x '7=  A  â x *BL?  A  Æ #t  €?    â  x     ?  A  Æ  r     A  Ì s     $   $ !t  €?    È  r      A  È  r     A  â Gy  °    € ê x    pRð Ú G‰  à     € ê x    pRð Ú G	        € ê 	z   i    � ä 	z i    � Æ  x ;ª¸¿  A  â 	z i €  � Ê s        $ !t  €?    Ì s        $  r      A  È  r     A  â Gy  Ð     € ê  r      A  â Gy  °     € ê 	x    à@  � ä 	x  à@  � Æ  x #ÛÙ?  A  â 	x  àÀ  � Æ  x;ª¸¿  A  ä !t  €?    È s        $ !t  €?    Ì s        $  r       A  È  r      A  ä y      "   ( y       &   b $z s  Ž ä $tÿ   ÿ Ž ä $z  v  Ž È/ %v j   Ž Ê †y     â My        € ê Gy  ðÿÿÿÿÿƒ À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À y           À         ÿ      ÿ ÿ     ÿÿÿ        ÿ    ÿÿ ÿ      ÿÿ     ÿÿÿ    ÿ   ÿ   ÿÿ  ÿ     ÿ ÿ    ÿÿ ÿ   ÿ  ÿÿ    ÿ ÿÿ   ÿ ÿÿÿ   ÿÿÿÿÿ        ÿ  ÿÿ   ÿ    ÿ  ÿ   ÿÿ  ÿ  ÿ  ÿ ÿ   ÿ ÿ ÿ  ÿ ÿÿ ÿ  ÿÿÿÿ ÿ      ÿÿ   ÿ  ÿÿ  ÿ ÿ ÿÿ  ÿÿÿ ÿÿ     ÿÿÿ  ÿÿ ÿÿÿ    ÿÿÿÿ   ÿÿÿÿÿ  ÿ     ÿ ÿÿ    ÿ   ÿ   ÿ  ÿÿ   ÿ ÿ  ÿ  ÿ  ÿ ÿ  ÿ ÿ ÿÿ  ÿ ÿÿÿÿ  ÿ     ÿ ÿ  ÿ  ÿ ÿ ÿ ÿ ÿ ÿ ÿÿÿ ÿ ÿ    ÿÿ ÿ ÿÿ ÿÿ ÿ   ÿÿÿ ÿ  ÿÿÿÿ ÿ ÿ    ÿÿ  ÿ   ÿÿ ÿ ÿ  ÿÿ ÿÿÿ  ÿÿ    ÿ ÿÿ ÿÿ ÿ ÿÿ   ÿÿ ÿÿ  ÿÿÿ ÿÿ ÿ   ÿÿÿ ÿÿ  ÿÿÿ   ÿ ÿÿÿ  ÿÿ ÿÿÿ ÿ  ÿÿÿÿ  ÿ ÿÿÿÿ ÿ ÿÿÿÿÿ ÿÿÿÿÿÿÿ        ÿÿÿ     ÿ  ÿ    ÿ ÿÿ    ÿÿ  ÿ   ÿ ÿ ÿ   ÿÿ ÿÿ   ÿÿÿÿÿ   ÿ    ÿ  ÿ ÿ  ÿ  ÿÿ ÿ ÿ  ÿÿÿÿ ÿ  ÿ   ÿÿ  ÿÿÿ ÿÿ  ÿ  ÿÿÿ  ÿ ÿÿÿÿ  ÿÿ    ÿ ÿ ÿ   ÿ ÿÿ ÿ  ÿ ÿÿÿÿ  ÿ ÿ   ÿ ÿ ÿÿÿ ÿ ÿ ÿ  ÿÿ ÿ ÿ ÿÿÿ ÿ ÿÿ   ÿÿ ÿÿÿ  ÿÿ ÿ  ÿ ÿÿ ÿ ÿÿ ÿÿ ÿÿ  ÿÿÿ ÿ ÿ ÿÿÿ ÿÿ ÿÿÿÿ ÿÿÿÿÿÿÿ ÿ      ÿÿ ÿ    ÿÿÿ ÿ   ÿÿÿÿÿ   ÿÿ   ÿ  ÿÿÿÿ ÿ  ÿÿ  ÿÿ  ÿÿ ÿÿÿ  ÿÿÿ   ÿ ÿÿÿÿ  ÿ ÿÿ  ÿ ÿ ÿÿ ÿÿ ÿ ÿÿÿ  ÿÿ ÿÿ ÿ ÿÿ ÿÿÿ ÿÿÿ ÿÿÿÿÿÿÿ ÿÿ     ÿÿÿÿÿ   ÿÿÿ  ÿ  ÿÿÿ ÿÿ  ÿÿÿÿ  ÿ ÿÿÿ ÿ ÿ ÿÿÿÿ ÿÿ ÿÿÿÿÿÿÿ ÿÿÿ    ÿÿÿÿ ÿ  ÿÿÿÿÿ ÿ ÿÿÿÿÿÿÿ ÿÿÿÿ   ÿÿÿÿÿÿÿ ÿÿÿÿÿ  ÿÿÿÿÿÿ ÿÿÿÿÿÿÿÿ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++$>$>,>,$$,>$>,>,,>>,>$>$$$$>,$>$,>$,>,,,,,4$44>$>4>$>4>$>$,>4$44$,$,$,$,4$,>,,4>$>4$,4>4>>>$,,$>$>,$,$,>,$,$>,>>4>,$>,,4>$$4$,,>$>$$>$,4,,,,,$44444>4>$$>$,$>$,$>,$$>$$>$$$4$44$>$,$$$$$,$$$4$$,,$$,$>,$,>$>$>$>$,$,>,,4,4,,,,$,,>,,,,$,$,>$,>$,,,,,,4,$>,$>,$4$$44$4$44444>4$44444$$$4,,4,,444>4>4>,>>>>>,>4>$>>,$>,>>,>>4>$>$$>,>,>$,>4>																																																																																																																																																																																																																															                                                               "      "  !    "    " ! ! ! !   "    " ! !!  "  "                          ! !    "!  !"!! ! !        "  "  !       " " !  !!      " " !    ""!    ! " ! !     ! !" !!"!    "    "  ""  !!    "   " "" !! ! !"! !!   "  !!  " "" !!  "  """                                          !    " !   "!" ! " "!!  "!                                !   ! " ! "   " ! "  " ! "!!" !! ! !""""! !  ! !   !!"  ! ""    !!"!  !!    ! "   !  !!" !!!"" !""  !" !"!" "!  ! "!  !"!!"  " !! "! "!!"!!! "!""            !   "  " "  " "   !! ""        !  " "" ""   ! ! ! "     !!! """"  !!!"!      "  ! ! !!   " " "  " "" ! ! !"!!!!"!    " ! !!!  "" " """   ! ! ! "      !! !!" ! !!""  ! " "    !!!!""" !!!!!""""!   ! "! !! !" ! ! !! "!"!"  "  !"!        "  "  !!   "  "  "" "" !  ! !!!"      " "   " " """""   ! !! "! !" ! "  !!! "!"!"""!  !!!"!!"!"    "    "  ""    " " "  "" """ !! ! !"!!!"!!!"   "   " "" !!!  "" " """""                               ! !  "!!  " !" !!  !                         ! ! !  ! "! !  " !    !"!"!"! ! ! "  !! "!  !"!  "   ! "  !!"!   !!"!  ! ! "  !!" !!  ! "   !!!"  !!"! "  !!" !!""! ! "! !" !! """ !!" !!!                         ! "" !"!  " !"!  !"!" !!  " !!"               !" " !! ! ! !" !" !!!" !" !" !" !"! !!"!" !"!!""!" ! !!!"  ! " "! ""  !!! ""    !! "   ! " "  !" !!!"! !!!"   !!!""" !!!" !" !!!""! ! "!!""  !!"!!"!! "!"! !!"! "!"" !! """! """ !!"! """!   !!!"    !!! !! " """  !"      !! " """   !!!!"""   !! """! !"!!!   !  " !!"! !  "! ! !"!!"!!!!!"""       ! !!!!! """     !! !!! !!" ""   !!!!    !!!!!!"""""   !!!!!""""" ! !! """ "!"  !   " !! !!!"! !"! "!" "  " ! "  !"! !!! " """""!!"!!"   !!!"! """!  ! "!!"  !  !!! "!"!"""    ! ! "  !!!!!!!"! "!"""  ! " !! "!"!"""! !"!! "!!   ! !"   !!!! "!""! !"!!"!"""!  !! "!""!"!"                     !  ! ! !!!!   "  "  " ""  " " ""          ! ! ! ! !!!!!!!! " """""    "      "  !    "    " ! ! !!!!!   ""  "  "" "!"  "" " """    !  !!!!! !! ""          ! ! ! !!!!!!!!!!! " " " """"""! " ! ! ! !  !! !" ! !!!!!"!!!"!"!! "!""" "        "  "  !       " "   !! ! ! !!"!!!"!   "  " " "" "!"  " " """"!  !    ! ! !   !! !! !" ! ! !!!"! !!!! "!"""!"    "    "  ""  !!    "   " "" !! !! ! !!"!!   ""  " " """ "!!"  "" " """""         !!!! !!!!" " """"                !! ! !!!!! !!"" """ """"! !    ! !  !"!!!  !" !!! !"!!!!" """!"           !!! !!! !! !!!! """ "" """              !!!!!!!!!!!!!!!!!!!"""""""""""""""!     ! !   " ! !! !!!!!"!!!"!!!!! !"" "!""""" "!" """!"! ! " !  !!!  ! !!!!!"! !!!!!" "!"""!"! ! "     ! " "     ! !  ! !!!"!"! !!!!!!!"! !!!!!"! "!" " "!""" "!"!! !! !!  !!"!!!"!  !!!! "!! !!!"! " !"!!""!"            !   "  "  " "  ! !! !!!!!!   "  "  " "!"  "" "" """"     ! ! ! ! !  ! ! !!!!!!!! "!"! " !" !"!"!""      "    ! !! " "  " "" ! !! !!"!!!!!!"!   ""  "  "" "!!"  """ "" """""! "  !  ! !! !!!! !!!!"!"!"! "!"! "    ! ! ! ! " "  ! !!!!!!!!!!!"!"! " " " " "!"!"!"!  !! "! "! !" !"   !" !" !"!!!!!"!! !! "!""!" " !"!!""!"!!"!"" """""        "  "    "  "  "" ""  !! !!!!!!"!!   "  " " "" "!!"  "" "" """"""!  !! "!  " !"   ! !!!!!"!! "!!"!""! !"!!"    "    "  ""    " " "  "" """ ! !! !!!"!!!   ""  " " """ "!!!"  """ "" """"""" @€�˜­¿ÏÝêö&5EYq  ÿþýüúøô �‚„‡ˆ	
‹�Ž�“•–™šœŸ !"£$¥¦'(©ª+¬-.¯0±²3´56·¸9:»<½¾?ÀABÃDÅÆGHÉÊKÌMNÏPÑÒSÔUV×ØYZÛ\ÝÞ_`áâcäefçèijëlíîoðqrótõöwxùú{ü}~ÿ                                                                                      @       €J                                                  ÀJ     ÚÚ                                                   %      }         <                J                    @£     °Š                             )                     ð-     ¤                              9                     ”.                                   I      p                ´.     ”,                            Î      p@               H[              ê                 ø     p@               X_     À         ë                      p@               c     p         ì                 
     p@               ˆf               í                      p@               ¨i     Ð         î                 (     p@               xl     €         ï                 4     p@               øn     0         ð                 @     p@               (q     ¸         ñ                 L     p@               àr     à         ò                 Ô     p@               Àt     0         ó                 \     p@               ðv     ˜         ô                 ]!     p@               ˆ     ¨         õ                 ®#     p@               0�     H         ö                 
&     p@               xƒ              ÷                 ’(     p@               ˆ‡     À         ø                 +     p@               H‹     p         ù                 ¢-     p@               ¸Ž               ú                 *0     p@               Ø‘     Ð         û                 ²2     p@               ¨”     €         ü                 :5     p@               (—     0         ý                 Â7     p@               X™     ¸         þ                 J:     p@               ›     à         ÿ                 Ò<     p@               ðœ     0                           Z?     p@                Ÿ     ˜                          ×A     p@               ¸§     ¨                          (D     p@               `©     H                          „F     p@               ¨«                               I     p@               ¸¯     À                          ”K     p@               x³     p                          N     p@               è¶                                ¤P     p@               º     Ð                          ,S     p@               Ø¼     €         	                 ´U     p@               X¿     0         
                 <X     p@               ˆÁ     ¸                          ÄZ     p@               @Ã     ü                          Ð]     p@               <Å     ¸                          X`     p@               ôÆ     ü                          dc     p@               ðÈ     0                          ìe     p@                Ë     ˜                          ih     p@               ¸Ó     ¨                          ºj     p@               `Õ     H                          m     p@               ¨×                               "p     p@               ¸Û     À                          .s     p@               xß     p                          :v     p@               èâ                                Fy     p@               æ     Ð                          R|     p@               Øè     €                          ^     p@               Xë     0                          j‚     p@               ˆí     ¸                          v…     p@               @ï     ü                          ‚ˆ     p@               <ñ     0                          Ž‹     p@               ló     ˜                          �Ž     p@               ü     ¨                          Y‘     p@               ¬ý     H                          .”     p@               ôÿ                                ¶–     p@               	     À         !                 >™     p@               Ä	     p         "                 Æ›     p@               4	               #                 Nž     p@               T	     Ð         $                 Ö      p@               $	     €         %                 ^£     p@               ¤	     0         &                 æ¥     p@               Ô	     ¸         '                 ò¨     p@               Œ	     ü         (                 þ«     p@               ˆ	     0         )                 †®     p@               ¸	     ˜         *                 ±     p@               P$	     ¨         +                 T³     p@               ø%	     H         ,                 °µ     p@               @(	              -                 ¼¸     p@               P,	     À         .                 È»     p@               0	     p         /                 Ô¾     p@               €3	               0                 àÁ     p@                6	     Ð         1                 ìÄ     p@               p9	     €         2                 øÇ     p@               ð;	     0         3                 Ë     p@                >	     ¸         4                 Î     p@               Ø?	     à         5                 ˜Ð     p@               ¸A	     0         6                  Ó     p@               èC	     ˜         7                 !Ö     p@               €L	     ¨         8                 rØ     p@               (N	     H         9                 ÎÚ     p@               pP	              :                 VÝ     p@               €T	     À         ;                 Þß     p@               @X	     p         <                 fâ     p@               °[	               =                 îä     p@               Ð^	     Ð         >                 vç     p@                a	     €         ?                 þé     p@                d	     0         @                 †ì     p@               Pf	     ¸         A                 ’ï     p@               h	     ü         B                 žò     p@               j	     0         C                 &õ     p@               4l	     ˜         D                 £÷     p@               Ìt	     ¨         E                 ôù     p@               tv	     H         F                 Pü     p@               ¼x	              G                 Øþ     p@               Ì|	     À         H                 `    p@               Œ€	     p         I                 è    p@               üƒ	               J                 p    p@               ‡	     Ð         K                 ø    p@               ì‰	     €         L                 €    p@               lŒ	     0         M                     p@               œŽ	     ¸         N                 �    p@               T�	     à         O                     p@               4’	     0         P                      p@               d”	     ˜         Q                     p@               üœ	     ¨         R                 n    p@               ¤ž	     H         S                 Ê    p@               ì 	              T                 R    p@               ü¤	     À         U                 Ú!    p@               ¼¨	     p         V                 b$    p@               ,¬	               W                 ê&    p@               L¯	     Ð         X                 r)    p@               ²	     €         Y                 ú+    p@               œ´	     0         Z                 ‚.    p@               Ì¶	     ¸         [                 
1    p@               „¸	     ü         \                 4    p@               €º	     0         ]                 ž6    p@               °¼	     ˜         ^                 9    p@               HÅ	     ¨         _                 l;    p@               ðÆ	     H         `                 È=    p@               8É	              a                 P@    p@               HÍ	     À         b                 ØB    p@               Ñ	     p         c                 `E    p@               xÔ	               d                 èG    p@               ˜×	     Ð         e                 pJ    p@               hÚ	     €         f                 øL    p@               èÜ	     0         g                 €O    p@               ß	     ¸         h                 ŒR    p@               Ðà	     ü         i                 ˜U    p@               Ìâ	     ¸         j                 ¤X    p@               „ä	     ü         k                 °[    p@               €æ	     0         l                 8^    p@               °è	     ˜         m                 µ`    p@               Hñ	     ¨         n                 c    p@               ðò	     H         o                 be    p@               8õ	              p                 nh    p@               Hù	     À         q                 zk    p@               ý	     p         r                 †n    p@               x 
               s                 ’q    p@               ˜
     Ð         t                 žt    p@               h
     €         u                 ªw    p@               è
     0         v                 ¶z    p@               
     ¸         w                 Â}    p@               Ð
     ü         x                 Î€    p@               Ì
     ¸         y                 Úƒ    p@               „
     ü         z                 æ†    p@               €
     0         {                 ò‰    p@               °
     ˜         |                 óŒ    p@               H
     ¨         }                 ½�    p@               ð
     H         ~                 ’’    p@               8!
                               ž•    p@               H%
     À         €                 ª˜    p@               )
     p         �                 ¶›    p@               x,
               ‚                 Âž    p@               ˜/
     Ð         ƒ                 Î¡    p@               h2
     €         „                 Ú¤    p@               è4
     0         …                 æ§    p@               7
     ¸         †                 òª    p@               Ð8
     ü         ‡                 þ­    p@               Ì:
     ¸         ˆ                 
±    p@               „<
     ü         ‰                 ´    p@               €>
     0         Š                 "·    p@               °@
     ˜         ‹                 #º    p@               HI
     ¨         Œ                 í¼    p@               ðJ
     H         �                 Â¿    p@               8M
              Ž                 JÂ    p@               HQ
     À         �                 ÒÄ    p@               U
     p         �                 ZÇ    p@               xX
               ‘                 âÉ    p@               ˜[
     Ð         ’                 jÌ    p@               h^
     €         “                 òÎ    p@               è`
     0         ”                 zÑ    p@               c
     ¸         •                 Ô    p@               Ðd
     à         –                 ŠÖ    p@               °f
     0         —                 –Ù    p@               àh
     ˜         ˜                 —Ü    p@               xq
     ¨         ™                 èÞ    p@                s
     H         š                 Dá    p@               hu
              ›                 Ìã    p@               xy
     À         œ                 Tæ    p@               8}
     p         �                 Üè    p@               ¨€
               ž                 dë    p@               Èƒ
     Ð         Ÿ                 ìí    p@               ˜†
     €                           tð    p@               ‰
     0         ¡                 üò    p@               H‹
     ¸         ¢                 „õ    p@                �
     ü         £                 �ø    p@               üŽ
     0         ¤                 û    p@               ,‘
     ˜         ¥                 •ý    p@               Ä™
     ¨         ¦                 æÿ    p@               l›
     H         §                 B    p@               ´�
              ¨                 Ê    p@               Ä¡
     À         ©                 R    p@               „¥
     p         ª                 Ú	    p@               ô¨
               «                 b    p@               ¬
     Ð         ¬                 ê    p@               ä®
     €         ­                 r    p@               d±
     0         ®                 ú    p@               ”³
     ¸         ¯                 ‚    p@               Lµ
     à         °                 
    p@               ,·
     0         ±                 ’    p@               \¹
     ˜         ²                     p@               ôÁ
     ¨         ³                 `     p@               œÃ
     H         ´                 ¼"    p@               äÅ
              µ                 D%    p@               ôÉ
     À         ¶                 Ì'    p@               ´Í
     p         ·                 T*    p@               $Ñ
               ¸                 Ü,    p@               DÔ
     Ð         ¹                 d/    p@               ×
     €         º                 ì1    p@               ”Ù
     0         »                 t4    p@               ÄÛ
     ¸         ¼                 ü6    p@               |Ý
     ü         ½                 :    p@               xß
     0         ¾                 �<    p@               ¨á
     ˜         ¿                 ?    p@               @ê
     ¨         À                 ^A    p@               èë
     H         Á                 ¹C    p@               0î
              Â                 <F    p@               @ò
     À         Ã                 ¿H    p@                ö
     p         Ä                 BK    p@               pù
               Å                 ÅM    p@               �ü
     Ð         Æ                 HP    p@               `ÿ
     €         Ç                 ËR    p@               à     0         È                 NU    p@                    ¸         É                 ÑW    p@               È     ü         Ê                 ×Z    p@               Ä     ¸         Ë                 Z]    p@               |	     ü         Ì                 ``    p@               x     0         Í                 ãb    p@               ¨     ´         Î                 Þe    p@               \     ¨         Ï                 *h    p@                    d         Ð                 ùj    p@               h              Ñ                 |m    p@               x     À         Ò                 ÿo    p@               8"     p         Ó                 ‚r    p@               ¨%               Ô                 u    p@               È(     Ð         Õ                 ˆw    p@               ˜+     €         Ö                 z    p@               .     0         ×                 Ž|    p@               H0     ¸         Ø                 ”    p@                2     ü         Ù                 š‚    p@               ü3     ¸         Ú                  …    p@               ´5     ü         Û                 ¦ˆ    p@               °7     0         Ü                 )‹    p@               à9     ˜         Ý                 ¡�    p@               xB     ¨         Þ                 í�    p@                D     H         ß                 D’    p@               hF              à                 Ç”    p@               xJ     À         á                 J—    p@               8N     p         â                 Í™    p@               ¨Q               ã                 Pœ    p@               ÈT     Ð         ä                 Óž    p@               ˜W     €         å                 V¡    p@               Z     0         æ                 Ù£    p@               H\     ¸         ç                 ß¦    p@                ^     ü         è                 å©    p@               ü_     ¸         é                 ë¬    p@               ´a     ü         ê                 ñ¯    p@               °c     0         ë                 ÷²    p@               àe     ˜         ì                 oµ    p@               xn     ¨         í                 »·    p@                p     H         î                 º    p@               hr              ï                 •¼    p@               xv     À         ð                 ¿    p@               8z     p         ñ                 ›Á    p@               ¨}               ò                 Ä    p@               È€     Ð         ó                 ¡Æ    p@               ˜ƒ     €         ô                 $É    p@               †     0         õ                 §Ë    p@               Hˆ     ¸         ö                 ­Î    p@                Š     ü         ÷                 ³Ñ    p@               ü‹     ¸         ø                 ¹Ô    p@               ´�     ü         ù                 ¿×    p@               °�     0         ú                 BÚ    p@               à‘     ˜         û                 ºÜ    p@               xš     ¨         ü                 ß    p@                œ     d         ý                 Õá    p@               „ž              þ                 Xä    p@               ”¢     À         ÿ                 Ûæ    p@               T¦     p                           ^é    p@               Ä©                                áë    p@               ä¬     Ð                          dî    p@               ´¯     €                          çð    p@               4²     0                          jó    p@               d´     ¸                          íõ    p@               ¶     ü                          óø    p@               ¸     ¸                          vû    p@               Ð¹     ü                          |þ    p@               Ì»     0         	                 ÿ     p@               ü½     ˜         
                 w    p@               ”Æ     ¨                          Ã    p@               <È     d                          “    p@                Ê                                   p@               °Î     À                          £    p@               pÒ     p                          +    p@               àÕ                                ³    p@                Ù     Ð                          ;    p@               ÐÛ     €                          Ã    p@               PÞ     0                          K    p@               €à     ¸                          Ó    p@               8â     ü                          ß    p@               4ä     0                          g"    p@               dæ     ˜                          ä$    p@               üî     ¨                          5'    p@               ¤ð     H                          ‘)    p@               ìò                               ,    p@               üö     À                          ¡.    p@               ¼ú     p                          )1    p@               ,þ                                ±3    p@               L     Ð                          96    p@                    €                          Á8    p@               œ     0                           I;    p@               Ì     ¸         !                 Ñ=    p@               „
     ü         "                 Ý@    p@               €     0         #                 eC    p@               °     ˜         $                 âE    p@               H     ¨         %                 3H    p@               ð     H         &                 UJ   p                8                                  qJ   p                X                                   J 	   @               h     °          I                2J 	   @                    Ð                          Þ                    è/     X                              ì     B               @0                ê                 p     B               H0                ê                 ø     B               H2                ë                 |     B               P2                ë                      B               P4                ì                 ˆ     B               X4                ì                      B               X6                í                 ”     B               `6                í                      B               `8                î                       B               h8                î                 (     B               h:                ï                 ¬     B               p:                ï                 4     B               p<                ð                 ¸     B               x<                ð                 @     B               x>                ñ                 Ä     B               €>                ñ                 L     B               €@                ò                 Ô     B               €B                ó                 \     B               €D                ô                 à     B               ˆD                ô                 G"     B               ˆF     ä          õ                 ˜$     B               lH     ä          ö                 
'     B               PJ                ÷                 ’)     B               PL                ø                 ,     B               PN                ù                 ¢.     B               PP                ú                 *1     B               PR                û                 ²3     B               PT                ü                 :6     B               PV                ý                 Â8     B               PX                þ                 J;     B               PZ                ÿ                 Ò=     B               P\                                  Z@     B               P^                                 ÁB     B               P`     ä                           E     B               4b     ä                           „G     B               d                                 J     B               f                                 ”L     B               h                                 O     B               j                                 ¤Q     B               l                                 ,T     B               n                	                 ´V     B               p                
                 <Y     B               r                                 Ä[     B               t                                 H\     B               $t                                 Ð^     B               $v                                 Xa     B               $x                                 Üa     B               0x                                 dd     B               0z                                 ìf     B               0|                                 Si     B               0~     ä                           ¤k     B               €     ä                           n     B               ø�                                 šn     B               ü�                                 "q     B               üƒ                                 ¦q     B                „                                 .t     B                †                                 ²t     B               †                                 :w     B               ˆ                                 ¾w     B               ˆ                                 Fz     B               Š                                 Êz     B               Š                                 R}     B               Œ                                 Ö}     B               Œ                                 ^€     B               Ž                                 â€     B               Ž                                 jƒ     B               �                                 îƒ     B                �                                 v†     B                ’                                 ú†     B               0’                                 ‚‰     B               0”                                 Š     B               4”                                 ŽŒ     B               4–                                 �     B               8–                                 y�     B               8˜                                 ò�     B               <˜     ä                           C’     B                š                                 ¼’     B               $š     ä                           .•     B               œ                                  ¶—     B               ž                !                 >š     B                                "                 Æœ     B               ¢                #                 NŸ     B               ¤                $                 Ö¡     B               ¦                %                 ^¤     B               ¨                &                 æ¦     B               ª                '                 j§     B               ª                '                 ò©     B               ¬                (                 vª     B               ¬                (                 þ¬     B               ®                )                 †¯     B               °                *                 í±     B               ²     ä          +                 >´     B                ´     ä          ,                 °¶     B               äµ                -                 4·     B               ìµ                -                 ¼¹     B               ì·                .                 @º     B               ô·                .                 È¼     B               ô¹                /                 L½     B               ü¹                /                 Ô¿     B               ü»                0                 XÀ     B               ¼                0                 àÂ     B               ¾                1                 dÃ     B               ¾                1                 ìÅ     B               À                2                 pÆ     B               À                2                 øÈ     B               Â                3                 |É     B               Â                3                 Ì     B               Ä                4                 ˆÌ     B               $Ä                4                 Ï     B               $Æ                5                 ˜Ñ     B               $È                6                  Ô     B               $Ê                7                 ¤Ô     B               ,Ê                7                 ×     B               ,Ì     ä          8                 \Ù     B               Î     ä          9                 ÎÛ     B               ôÏ                :                 VÞ     B               ôÑ                ;                 Þà     B               ôÓ                <                 fã     B               ôÕ                =                 îå     B               ô×                >                 vè     B               ôÙ                ?                 þê     B               ôÛ                @                 †í     B               ôÝ                A                 
î     B               üÝ                A                 ’ð     B               üß                B                 ñ     B               à                B                 žó     B               â                C                 &ö     B               ä                D                 �ø     B               æ     ä          E                 Þú     B               ìç     ä          F                 Pý     B               Ðé                G                 Øÿ     B               Ðë                H                 `    B               Ðí                I                 è    B               Ðï                J                 p    B               Ðñ                K                 ø	    B               Ðó                L                 €    B               Ðõ                M                     B               Ð÷                N                 �    B               Ðù                O                     B               Ðû                P                      B               Ðý                Q                     B               Ðÿ     ä          R                 X    B               ´     ä          S                 Ê    B               ˜                T                 R     B               ˜                U                 Ú"    B               ˜                V                 b%    B               ˜	                W                 ê'    B               ˜                X                 r*    B               ˜                Y                 ú,    B               ˜                Z                 ‚/    B               ˜                [                 
2    B               ˜                \                 Ž2    B               ¤                \                 5    B               ¤                ]                 ž7    B               ¤                ^                 :    B               ¤     ä          _                 V<    B               ˆ     ä          `                 È>    B               l                a                 PA    B               l                b                 ØC    B               l!                c                 `F    B               l#                d                 èH    B               l%                e                 pK    B               l'                f                 øM    B               l)                g                 €P    B               l+                h                 Q    B               t+                h                 ŒS    B               t-                i                 T    B               €-                i                 ˜V    B               €/                j                 W    B               ˆ/                j                 ¤Y    B               ˆ1                k                 (Z    B               ”1                k                 °\    B               ”3                l                 8_    B               ”5                m                 Ÿa    B               ”7     ä          n                 ðc    B               x9     ä          o                 bf    B               \;                p                 æf    B               `;                p                 ni    B               `=                q                 òi    B               d=                q                 zl    B               d?                r                 þl    B               h?                r                 †o    B               hA                s                 
p    B               lA                s                 ’r    B               lC                t                 s    B               pC                t                 žu    B               pE                u                 "v    B               tE                u                 ªx    B               tG                v                 .y    B               xG                v                 ¶{    B               xI                w                 :|    B               |I                w                 Â~    B               |K                x                 F    B               ŒK                x                 Î�    B               ŒM                y                 R‚    B               �M                y                 Ú„    B               �O                z                 ^…    B                O                z                 æ‡    B                Q                {                 jˆ    B               ¤Q                {                 òŠ    B               ¤S                |                 v‹    B               ¨S                |                 Ý�    B               ¨U                }                 VŽ    B               ¬U     ä          }                 §�    B               �W                ~                  ‘    B               ”W     ä          ~                 ’“    B               xY                                 ”    B               |Y                                 ž–    B               |[                €                 "—    B               €[                €                 ª™    B               €]                �                 .š    B               „]                �                 ¶œ    B               „_                ‚                 :�    B               ˆ_                ‚                 ÂŸ    B               ˆa                ƒ                 F     B               Œa                ƒ                 Î¢    B               Œc                „                 R£    B               �c                „                 Ú¥    B               �e                …                 ^¦    B               ”e                …                 æ¨    B               ”g                †                 j©    B               ˜g                †                 ò«    B               ˜i                ‡                 v¬    B               ¨i                ‡                 þ®    B               ¨k                ˆ                 ‚¯    B               ¬k                ˆ                 
²    B               ¬m                ‰                 Ž²    B               ¼m                ‰                 µ    B               ¼o                Š                 šµ    B               Ào                Š                 "¸    B               Àq                ‹                 ¦¸    B               Äq                ‹                 »    B               Äs                Œ                 †»    B               Ès     ä          Œ                 ×½    B               ¬u                �                 P¾    B               °u     ä          �                 ÂÀ    B               ”w                Ž                 JÃ    B               ”y                �                 ÒÅ    B               ”{                �                 ZÈ    B               ”}                ‘                 âÊ    B               ”                ’                 jÍ    B               ”�                “                 òÏ    B               ”ƒ                ”                 zÒ    B               ”…                •                 Õ    B               ”‡                –                 Š×    B               ”‰                —                 Ø    B               œ‰                —                 –Ú    B               œ‹                ˜                 Û    B               ¤‹                ˜                 �Ý    B               ¤�     ä          ™                 Òß    B               ˆ�     ä          š                 Dâ    B               l‘                ›                 Ìä    B               l“                œ                 Tç    B               l•                �                 Üé    B               l—                ž                 dì    B               l™                Ÿ                 ìî    B               l›                                  tñ    B               l�                ¡                 üó    B               lŸ                ¢                 „ö    B               l¡                £                 ÷    B               x¡                £                 �ù    B               x£                ¤                 ü    B               x¥                ¥                 þ    B               x§     ä          ¦                 Ð     B               \©     ä          §                 B    B               @«                ¨                 Ê    B               @­                ©                 R    B               @¯                ª                 Ú
    B               @±                «                 b    B               @³                ¬                 ê    B               @µ                ­                 r    B               @·                ®                 ú    B               @¹                ¯                 ‚    B               @»                °                 
    B               @½                ±                 ’    B               @¿                ²                 ù    B               @Á     ä          ³                 J!    B               $Ã     ä          ´                 ¼#    B               Å                µ                 D&    B               Ç                ¶                 Ì(    B               É                ·                 T+    B               Ë                ¸                 Ü-    B               Í                ¹                 d0    B               Ï                º                 ì2    B               Ñ                »                 t5    B               Ó                ¼                 ü7    B               Õ                ½                 €8    B               Õ                ½                 ;    B               ×                ¾                 �=    B               Ù                ¿                 ÷?    B               Û     ä          À                 HB    B               øÜ     ä          Á                 ·D    B               ÜÞ                Â                 :G    B               Üà                Ã                 ½I    B               Üâ                Ä                 @L    B               Üä                Å                 ÃN    B               Üæ                Æ                 FQ    B               Üè                Ç                 ÉS    B               Üê                È                 LV    B               Üì                É                 ÏX    B               Üî                Ê                 RY    B               èî                Ê                 Õ[    B               èð                Ë                 X^    B               èò                Ì                 Û^    B               ôò                Ì                 ^a    B               ôô                Í                 ác    B               ôö                Î                 dd    B                ÷                Î                 Æf    B                ù     ä          Ï                 i    B               äú                Ð                 Ši    B               ðú     ä          Ð                 ÷k    B               Ôü                Ñ                 zn    B               Ôþ                Ò                 ýp    B               Ô                 Ó                 €s    B               Ô                Ô                 v    B               Ô                Õ                 †x    B               Ô                Ö                 	{    B               Ô                ×                 Œ}    B               Ô
                Ø                 ~    B               Ü
                Ø                 ’€    B               Ü                Ù                 �    B               è                Ù                 ˜ƒ    B               è                Ú                 „    B               ð                Ú                 ž†    B               ð                Û                 !‡    B               ü                Û                 ¤‰    B               ü                Ü                 'Œ    B               ü                Ý                 ‰Ž    B               ü     ä          Þ                 Õ�    B               à     ä          ß                 B“    B               Ä                à                 Å•    B               Ä                á                 H˜    B               Ä                â                 Ëš    B               Ä                 ã                 N�    B               Ä"                ä                 ÑŸ    B               Ä$                å                 T¢    B               Ä&                æ                 ×¤    B               Ä(                ç                 Z¥    B               Ì(                ç                 Ý§    B               Ì*                è                 `¨    B               Ø*                è                 ãª    B               Ø,                é                 f«    B               à,                é                 é­    B               à.                ê                 l®    B               ì.                ê                 ï°    B               ì0                ë                 r±    B               ô0                ë                 õ³    B               ô2                ì                 W¶    B               ô4     ä          í                 £¸    B               Ø6     ä          î                 »    B               ¼8                ï                 “½    B               ¼:                ð                 À    B               ¼<                ñ                 ™Â    B               ¼>                ò                 Å    B               ¼@                ó                 ŸÇ    B               ¼B                ô                 "Ê    B               ¼D                õ                 ¥Ì    B               ¼F                ö                 (Í    B               ÄF                ö                 «Ï    B               ÄH                ÷                 .Ð    B               ÐH                ÷                 ±Ò    B               ÐJ                ø                 4Ó    B               ØJ                ø                 ·Õ    B               ØL                ù                 :Ö    B               äL                ù                 ½Ø    B               äN                ú                 @Û    B               äP                û                 ¢Ý    B               äR     ä          ü                 îß    B               ÈT                ý                 fà    B               ÔT     ä          ý                 Óâ    B               ¸V                þ                 Vå    B               ¸X                ÿ                 Ùç    B               ¸Z                                  \ê    B               ¸\                                 ßì    B               ¸^                                 bï    B               ¸`                                 åñ    B               ¸b                                 hô    B               ¸d                                 ëö    B               ¸f                                 n÷    B               Äf                                 ñù    B               Äh                                 tü    B               Äj                                 ÷ü    B               Ðj                                 zÿ    B               Ðl                	                 ý    B               Ðn                
                 _    B               Ðp     ä                           «    B               ´r                                 #    B               Àr     ä                           “	    B               ¤t                                     B               ¤v                                 £    B               ¤x                                 +    B               ¤z                                 ³    B               ¤|                                 ;    B               ¤~                                 Ã    B               ¤€                                 K    B               ¤‚                                 Ó    B               ¤„                                 W    B               °„                                 ß     B               °†                                 g#    B               °ˆ                                 Î%    B               °Š     ä                           (    B               ”Œ     ä                           ‘*    B               xŽ                                 -    B               x�                                 ¡/    B               x’                                 )2    B               x”                                 ±4    B               x–                                 97    B               x˜                                 Á9    B               xš                                  I<    B               xœ                !                 Ñ>    B               xž                "                 U?    B               „ž                "                 ÝA    B               „                 #                 eD    B               „¢                $                 ÌF    B               „¤     ä          %                 I    B               h¦     ä          &                 R                     €¨     €@           ¨€               |                     é     €;           š€               ˆ                    €$      7           �€               ”	                    €[      2           €€                                    €�     €/         $  v€               ¬                     ½      *         )  o€               ¸                     ç     €$         .  _€               Ä                    €     €         3  7€               Ð                          €%         8  7€               X                    €C      5         <  k€               à                    €x      ¢         @  ‹€               ì                     €               E  N€               =#                    €7      :         H  P€               Ž%                    €q      :         K  ™€               (                    €«     €3         O  Š€               ž*                     ß     €.         S  ‚€               &-                    €      )         W  u€               ®/                    €6     €&         [  k€               62                     ]      !         _  _€               ¾4                     ~               c  S€               F7                     š               g  /€               Î9                     ©     €         k  0€               V<                    €Æ      '         o  R€               Þ>                    €í      ƒ         s  d€               fA                    €p               w  =€               ·C                    €…     €+         z  @€               F                     ±     €,         }  n€               �H                    €Ý     €(         �  _€               K                          €$         …  [€                M                    €*     €!         ‰  N€               (P                     L               �  G€               °R                     j               ‘  @€               8U                     ƒ     €         •  7€               ÀW                    €—               ™  &€               HZ                    €¢               �  (€               T]                    €¹               ¢  &€               Ü_                    €Ä               ¦  (€               èb                    €Û               «  8€               pe                    €ô     €g         ¯  H€               øg                     \     €         ³  0€               Ij                    €j     €         ¶  5€               šl                     ‰     €E         ¹  £€               ¦o                    €Î     €>         ¾  ™€               ²r                           8         Ã  “€               ¾u                     E      1         È  €€               Êx                     v      -         Í  x€               Ö{                     £     €%         Ò  m€               â~                    €È     €         ×  _€               î�                     ç     €         Ü  0€               ú„                    €ö               á  ?€               ˆ                    €      &         æ  `€               ‹                    €:     €�         ë  €€               Ž                     ¼               ð  G€               è�                     Ò      +         ô  H€               ²“                     ý      5         ø  �€               :–                     2      0         ü  €€               Â˜                     b     €)           w€               J›                    €‹     €$          q€               Ò�                     °     €!          `€               Z                     €Ñ     €          U€               â¢                     í     €          H€               j¥                    €     €          (€               v¨                                     0€               ‚«                     '                H€               
®                     B     €l         " ]€               ’°                    €®     €         & 6€               ã²                     ¾     €          ) >€               4µ                    €Þ      D         , �€               @¸                    €"     €>         1 Ž€               L»                     a      8         6 …€               X¾                     ™      2         ; x€               dÁ                     Ë      /         @ l€               pÄ                     ú      )         E `€               |Ç                     #      #         J R€               ˆÊ                     F     €         O 7€               ”Í                    €W      $         T 0€               Ð                    €{     €3         X [€               ¤Ò                     ¯      �         \ r€               °Õ                     L     €         a G€               Ø                    €g      8         d G€               RÚ                    €Ÿ     €D         g ›€               ÚÜ                     ä      =         k �€               bß                     !     €7         o ‚€               êá                    €X     €1         s y€               rä                     Š      .         w l€               úæ                     ¸     €'         { `€               ‚é                    €ß     €!          U€               
ì                          €         ƒ /€               ï                    €      "         ˆ 5€               "ò                    €3     €/         � ^€               ªô                     c      •         ‘ q€               2÷                     ø     €         • G€               ƒù                    €      4         ˜ H€               Ôû                    €E      E         › ž€               \þ                    €Š      ?         Ÿ ’€               ä                    €É      8         £ Š€               l                   €      €1         § {€               ô                    3      €.         « o€               |                   €a      €'         ¯ `€                                   ‰      €!         ³ W€               Œ                   €ª      €         · 0€                                   ¼      €"         » 3€               œ                   €Þ       0         ¿ `€               $                   €!     €•         Ã s€               ¬                    ¤!               Ç G€               ý                    ¾!     €4         Ê H€               N                   €ò!      ?         Í š€               Ö                   €1"     €;         Ñ Ž€               ^!                    m"     €6         Õ €€               æ#                   €£"      1         Ù t€               n&                   €Ô"      .         Ý h€               ö(                   €#      &         á [€               ~+                   €(#     €          å P€               .                    I#     €         é .€               Ž0                   €Y#      !         í 3€               š3                   €z#     €.         ò V€               "6                    ©#     €’         ö d€               ª8                   €;$               ú @€               û:                   €T$      3         ý D€               L=                   €‡$     €/           p€               Ô?                    ·$      +          j€               \B                    â$     €&          `€               äD                   €%      "          \€               lG                   €*%     €          P€               ôI                    J%     €          H€               |L                   €d%     €          H€               O                    z%                (€               R                    †%     €         ! 0€               U                   €�%               & (€               (X                   €©%     €         + 0€               4[                    Á%     €         0 H€               ¼]                   €Ú%     €h         4 `€               D`                    C&     €         8 8€               •b                   €R&               ; >€               æd                   €q&     €<         > €€               òg                    ®&     €6         C v€               þj                   €ä&     €0         H s€               
n                    '     €*         M h€               q                   €?'     €&         R ^€               "t                    f'                W P€               .w                    †'     €         \ H€               :z                   €Ÿ'     €         a (€               F}                    ¬'               f 8€               R€                    Å'     €         k (€               ^ƒ                   €Ñ'               p 8€               j†                   €ê'               u H€               v‰                   €(      q         z n€               ‚Œ                   €x(                8€               L�                   €‰(     €"         ƒ ;€               ’                    ¬(     €:         ‡ €               "•                   €æ(     €4         Œ q€               .˜                    )     €.         ‘ h€               :›                   €I)     €(         – `€               Fž                    r)     €$         › Y€               R¡                   €–)     €           O€               ^¤                    ´)     €         ¥ G€               j§                   €Ë)     €         ª 8€               vª                    á)               ¯ 0€               ‚­                    ø)     €         ´ 8€               Ž°                   €*               ¹ 0€               š³                   €$*     €         ¾ @€               ¦¶                    >*      i         Ã [€               ²¹                    §*               È 0€               |¼                    ¶*               Ì 8€               F¿                    Õ*      G         Ð À€               ÎÁ                    +     €A         Ô ¸€               VÄ                   €]+     €;         Ø §€               ÞÆ                    ™+      6         Ü �€               fÉ                    Ï+     €2         à •€               îË                   €,      ,         ä Š€               vÎ                   €-,     €%         è ~€               þÐ                    S,               ì @€               †Ó                    f,     €$         ð O€               Ö                   €Š,     €0         ô ~€               Ù                    »,      ”         ù Ë€               &Ü                    O-     €         þ k€               wÞ                   €l-     €8          „€               Èà                    ¥-      G          •€               Pã                    ì-     €?          Š€               Øå                   €+.     €8          €               `è                    d.     €0          v€               èê                   €”.      +          i€               pí                   €¿.      #          _€               øï                   €â.                P€               €ò                   €ý.     €           (€               õ                    /     €         $ 0€               ø                   €$/               ) F€               œú                   €B/      q         - ]€               $ý                   €³/               1 0€               uÿ                   €Ä/      #         4 8€               Æ                   €ç/     €B         7 ž€               N                    *0      ;         ; “€               Ö                    e0     €5         ? „€               ^	                   €š0      /         C z€               æ                   €É0     €+         G n€               n                    õ0     €$         K \€               ö                   €1     €         O O€               ~                    81               S /€                                   H1     €          W 7€               Ž                   €h1      )         [ M€                                  €‘1      ˆ         _ e€               ž                   €2     €         c 8€               ï                    02      /         f @€               @"                    _2     €.         i l€               È$                   €�2      *         m d€               P'                   €·2     €&         q ]€               Ø)                    Þ2      $         u P€               `,                    3     €          y H€               è.                   €"3               } H€               p1                   €=3               � 8€               ø3                   €T3               … #€               €6                   €`3               ‰ (€               Œ9                   €y3               Ž H€               <                   €–3      p         ’ P€               œ>                   €4               – 0€               í@                   €4     €"         ™ 8€               >C                    :4     €%         œ ^€               ÁE                   €_4      "           P€               DH                   €�4               ¤ N€               ÇJ                   €Ÿ4               ¨ H€               JM                   €¹4               ¬ @€               ÍO                   €Ñ4     €         ° 8€               PR                    å4     €         ´ 0€               ÓT                   €ô4     €         ¸ @€               VW                    5               ¼ 0€               \Z                    5     €         Á @€               ß\                   €05               Å 0€               å_                   €I5     €         Ê 8€               hb                    Y5     €U         Î @€               ne                   €®5               Ó P€               ºg                   €È5               Ö 0€               ~j                   €Ü5     €/         Ú c€               m                    6      +         Þ \€               „o                    76     €&         â R€               r                   €]6     €!         æ H€               Št                    6               ê D€               w                    ž6     €         î @€               �y                   €·6               ò 6€               |                   €Ì6               ö (€                                  €×6     €         û (€               ‚                    ï6                 (€               %…                    ú6     €          (€               +ˆ                   €7     €         
 8€               ®Š                    ,7      i          @€               1�                    •7                /€               }�                    ¤7                 6€               É‘                    Ä7     €/          m€               L”                   €ó7     €*          d€               Ï–                    8     €&           \€               R™                   €D8      "         $ \€               Õ›                   €f8     €         ( P€               Xž                    †8               , P€               Û                      8     €         0 ?€               ^£                   €µ8     €         4 '€               d¦                    Á8     €         9 (€               j©                   €Ø8     €         > '€               p¬                    ä8     €         C (€               v¯                   €û8               H H€               |²                   €9      k         M P€               ÿ´                   €‚9     €         Q /€               K·                    “9     €!         T 7€               —¹                   €´9      +         W `€               ¼                   €ß9     €&         [ U€               �¾                    :      "         _ P€                Á                    (:               c G€               £Ã                    E:     €         g @€               &Æ                   €_:     €         k 8€               ©È                    u:     €         o ,€               ,Ë                   €…:     €         s (€               2Î                    “:     €         x "€               8Ñ                   €¥:     €         } (€               >Ô                    ³:     €         ‚ "€               D×                   €Å:               ‡ 0€               ÇÙ                   €Ö:      X         ‹ 8€               JÜ                   €.;     €
         � (€               –Þ                    9;     €         ’ (€               Zá                   €O;     €*         – j€               Ýã                    z;      &         š ^€               `æ                     ;      "         ž X€               ãè                    Â;               ¢ P€               fë                    ß;               ¦ H€               éí                    ú;     €         ª @€               lð                   €<               ® 8€               ïò                   € <               ² 7€               rõ                   €/<               ¶ '€               xø                   €B<               » 7€               ûú                   €Q<               ¿ '€               þ                   €d<               Ä 8€               „                    €v<      Z         È H€                                  €Ð<               Ì (€               S                   €Û<               Ï /€                                  €ò<     €1         Ó �€               Ÿ
                    $=     €,         × z€               '                   €P=     €)         Û o€               ¯                    z=      %         ß `€               7                    Ÿ=     €"         ã Y€               ¿                   €Á=               ç H€               G                   €Þ=               ë =€               Ï                   €ô=     €         ï 8€               W                    
>     €         ó +€               c                   €!>               ø 8€               ë!                   €<>      i         ü H€               s$                   €¥>     €           ,€               Ä&                    µ>                .€               )                    Ó>     €2          ˆ€               �+                   €?      .         
 }€               %.                   €3?     €+          q€               ­0                    _?     €&          j€               53                   €…?      $          X€               ½5                   €©?                J€               E8                   €Æ?                =€               Í:                   €Ý?               " 3€               U=                   €ô?     €         & *€               a@                    @     €         + 7€               éB                   €*@     €o         / @€               qE                    š@     €         3 .€               ÂG                   €ª@     €          6 +€               Î                     Ë@     ¨h                             M     C               ¨3A                ê                 w     C               ¨3A                ë                 ƒ     C               ¨3A                ì                 �
     C               ¨3A                í                 ›     C               ¨3A                î                 §     C               ¨3A      	          ï                 ³     C               ¨3A                ð                 ¿     C               ¨3A     €          ñ                 Ë     C               ¨3A                ò                 S     C               ¨3A                ó                 Û     C               ¨3A                ô                 ‰&     C               ¨3A                ÷                 )     C               ¨3A                ø                 ™+     C               ¨3A                ù                 !.     C               ¨3A                ú                 ©0     C               ¨3A                û                 13     C               ¨3A      	          ü                 ¹5     C               ¨3A                ý                 A8     C               ¨3A     €          þ                 É:     C               ¨3A                ÿ                 Q=     C               ¨3A                                  Ù?     C               ¨3A                                 G     C               ¨3A                                 ‹I     C               ¨3A                                 L     C               ¨3A                                 ›N     C               ¨3A                                 #Q     C               ¨3A                                 «S     C               ¨3A      	          	                 3V     C               ¨3A                
                 »X     C               ¨3A     €                           C[     C               ¨3A                                 O^     C               ¨3A     €                           ×`     C               ¨3A                                 ãc     C               ¨3A                                 kf     C               ¨3A                                 •m     C               ¨3A                                 ¡p     C               ¨3A                                 ­s     C               ¨3A                                 ¹v     C               ¨3A                                 Åy     C               ¨3A                                 Ñ|     C               ¨3A      	                           Ý     C               ¨3A                                 é‚     C               ¨3A     €                           õ…     C               ¨3A                                 ‰     C               ¨3A                                 Œ     C               ¨3A                                 ­”     C               ¨3A                                  5—     C               ¨3A                !                 ½™     C               ¨3A                "                 Eœ     C               ¨3A                #                 Íž     C               ¨3A                $                 U¡     C               ¨3A      	          %                 Ý£     C               ¨3A                &                 e¦     C               ¨3A     €          '                 q©     C               ¨3A                (                 }¬     C               ¨3A                )                 ¯     C               ¨3A                *                 /¶     C               ¨3A                -                 ;¹     C               ¨3A                .                 G¼     C               ¨3A                /                 S¿     C               ¨3A                0                 _Â     C               ¨3A                1                 kÅ     C               ¨3A      	          2                 wÈ     C               ¨3A                3                 ƒË     C               ¨3A     €          4                 �Î     C               ¨3A                5                 Ñ     C               ¨3A                6                 ŸÓ     C               ¨3A                7                 MÛ     C               ¨3A                :                 ÕÝ     C               ¨3A                ;                 ]à     C               ¨3A                <                 åâ     C               ¨3A                =                 må     C               ¨3A                >                 õç     C               ¨3A      	          ?                 }ê     C               ¨3A                @                 í     C               ¨3A     €          A                 ð     C               ¨3A                B                 ó     C               ¨3A                C                 ¥õ     C               ¨3A                D                 Ïü     C               ¨3A                G                 Wÿ     C               ¨3A                H                 ß    C               ¨3A                I                 g    C               ¨3A                J                 ï    C               ¨3A                K                 w	    C               ¨3A      	          L                 ÿ    C               ¨3A                M                 ‡    C               ¨3A     €          N                     C               ¨3A                O                 —    C               ¨3A                P                     C               ¨3A                Q                 I    C               ¨3A                T                 Ñ    C               ¨3A                U                 Y"    C               ¨3A                V                 á$    C               ¨3A                W                 i'    C               ¨3A                X                 ñ)    C               ¨3A      	          Y                 y,    C               ¨3A                Z                 /    C               ¨3A     €          [                 ‰1    C               ¨3A                \                 •4    C               ¨3A                ]                 7    C               ¨3A                ^                 G>    C               ¨3A                a                 Ï@    C               ¨3A                b                 WC    C               ¨3A                c                 ßE    C               ¨3A                d                 gH    C               ¨3A                e                 ïJ    C               ¨3A      	          f                 wM    C               ¨3A                g                 ÿO    C               ¨3A     €          h                 S    C               ¨3A                i                 V    C               ¨3A     €          j                 #Y    C               ¨3A                k                 /\    C               ¨3A                l                 ·^    C               ¨3A                m                 áe    C               ¨3A                p                 íh    C               ¨3A                q                 ùk    C               ¨3A                r                 o    C               ¨3A                s                 r    C               ¨3A                t                 u    C               ¨3A      	          u                 )x    C               ¨3A                v                 5{    C               ¨3A     €          w                 A~    C               ¨3A                x                 M�    C               ¨3A     €          y                 Y„    C               ¨3A                z                 e‡    C               ¨3A                {                 qŠ    C               ¨3A                |                 “    C               ¨3A                                 –    C               ¨3A                €                 )™    C               ¨3A                �                 5œ    C               ¨3A                ‚                 AŸ    C               ¨3A                ƒ                 M¢    C               ¨3A      	          „                 Y¥    C               ¨3A                …                 e¨    C               ¨3A     €          †                 q«    C               ¨3A                ‡                 }®    C               ¨3A     €          ˆ                 ‰±    C               ¨3A                ‰                 •´    C               ¨3A                Š                 ¡·    C               ¨3A                ‹                 AÀ    C               ¨3A                Ž                 ÉÂ    C               ¨3A                �                 QÅ    C               ¨3A                �                 ÙÇ    C               ¨3A                ‘                 aÊ    C               ¨3A                ’                 éÌ    C               ¨3A      	          “                 qÏ    C               ¨3A                ”                 ùÑ    C               ¨3A     €          •                 �Ô    C               ¨3A                –                 	×    C               ¨3A                —                 Ú    C               ¨3A                ˜                 Ãá    C               ¨3A                ›                 Kä    C               ¨3A                œ                 Óæ    C               ¨3A                �                 [é    C               ¨3A                ž                 ãë    C               ¨3A                Ÿ                 kî    C               ¨3A      	                            óð    C               ¨3A                ¡                 {ó    C               ¨3A     €          ¢                 ö    C               ¨3A                £                 ù    C               ¨3A                ¤                 —û    C               ¨3A                ¥                 Á    C               ¨3A                ¨                 I    C               ¨3A                ©                 Ñ    C               ¨3A                ª                 Y
    C               ¨3A                «                 á    C               ¨3A                ¬                 i    C               ¨3A      	          ­                 ñ    C               ¨3A                ®                 y    C               ¨3A     €          ¯                     C               ¨3A                °                 ‰    C               ¨3A                ±                     C               ¨3A                ²                 ;#    C               ¨3A                µ                 Ã%    C               ¨3A                ¶                 K(    C               ¨3A                ·                 Ó*    C               ¨3A                ¸                 [-    C               ¨3A                ¹                 ã/    C               ¨3A      	          º                 k2    C               ¨3A                »                 ó4    C               ¨3A     €          ¼                 {7    C               ¨3A                ½                 ‡:    C               ¨3A                ¾                 =    C               ¨3A                ¿                 7D    C               ¨3A                Â                 ºF    C               ¨3A                Ã                 =I    C               ¨3A                Ä                 ÀK    C               ¨3A                Å                 CN    C               ¨3A                Æ                 ÆP    C               ¨3A      	          Ç                 IS    C               ¨3A                È                 ÌU    C               ¨3A     €          É                 OX    C               ¨3A                Ê                 U[    C               ¨3A     €          Ë                 Ø]    C               ¨3A                Ì                 Þ`    C               ¨3A                Í                 ac    C               ¨3A                Î                 wk    C               ¨3A                Ñ                 úm    C               ¨3A                Ò                 }p    C               ¨3A                Ó                  s    C               ¨3A                Ô                 ƒu    C               ¨3A                Õ                 x    C               ¨3A      	          Ö                 ‰z    C               ¨3A                ×                 }    C               ¨3A     €          Ø                 €    C               ¨3A                Ù                 ƒ    C               ¨3A     €          Ú                 †    C               ¨3A                Û                 $‰    C               ¨3A                Ü                 §‹    C               ¨3A                Ý                 Â’    C               ¨3A                à                 E•    C               ¨3A                á                 È—    C               ¨3A                â                 Kš    C               ¨3A                ã                 Îœ    C               ¨3A                ä                 QŸ    C               ¨3A      	          å                 Ô¡    C               ¨3A                æ                 W¤    C               ¨3A     €          ç                 ]§    C               ¨3A                è                 cª    C               ¨3A     €          é                 i­    C               ¨3A                ê                 o°    C               ¨3A                ë                 u³    C               ¨3A                ì                 �º    C               ¨3A                ï                 ½    C               ¨3A                ð                 –¿    C               ¨3A                ñ                 Â    C               ¨3A                ò                 œÄ    C               ¨3A                ó                 Ç    C               ¨3A      	          ô                 ¢É    C               ¨3A                õ                 %Ì    C               ¨3A     €          ö                 +Ï    C               ¨3A                ÷                 1Ò    C               ¨3A     €          ø                 7Õ    C               ¨3A                ù                 =Ø    C               ¨3A                ú                 ÀÚ    C               ¨3A                û                 Sâ    C               ¨3A                þ                 Öä    C               ¨3A                ÿ                 Yç    C               ¨3A                                  Üé    C               ¨3A                                 _ì    C               ¨3A                                 âî    C               ¨3A      	                           eñ    C               ¨3A                                 èó    C               ¨3A     €                           kö    C               ¨3A                                 qù    C               ¨3A     €                           ôû    C               ¨3A                                 úþ    C               ¨3A                	                 }    C               ¨3A                
                 	    C               ¨3A                                 š    C               ¨3A                                 "    C               ¨3A                                 ª    C               ¨3A                                 2    C               ¨3A                                 º    C               ¨3A      	                           B    C               ¨3A                                 Ê    C               ¨3A     €                           R    C               ¨3A                                 ^     C               ¨3A                                 æ"    C               ¨3A                                 *    C               ¨3A                                 ˜,    C               ¨3A                                  /    C               ¨3A                                 ¨1    C               ¨3A                                 04    C               ¨3A                                 ¸6    C               ¨3A      	                           @9    C               ¨3A                                  È;    C               ¨3A     €          !                 P>    C               ¨3A                "                 \A    C               ¨3A                #                 äC    C               ¨3A                $                       h�B                     à       à                    è/                     ›4     ›4                   Ë@                     ¨h      ¨Y                  h�B                     à       à              